diff --git a/.gitattributes b/.gitattributes index 30de84c85b970f4d236e1212ce373dd5f2331e0a..ebda80e88197de82a21585b56880e1134138b32d 100644 --- a/.gitattributes +++ b/.gitattributes @@ -6889,3 +6889,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/MOTRv2_[[:space:]]Bootstrapping[[:space:]]End-to-End[[:space:]]Multi-Object[[:space:]]Tracking[[:space:]]by[[:space:]]Pretrained[[:space:]]Object[[:space:]]Detectors/02a94155-5f1f-4c5a-b31e-6e19a4b86117_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/MOT_[[:space:]]Masked[[:space:]]Optimal[[:space:]]Transport[[:space:]]for[[:space:]]Partial[[:space:]]Domain[[:space:]]Adaptation/fe7a584b-6d50-46aa-831e-cebc57fd2083_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/MOVES_[[:space:]]Manipulated[[:space:]]Objects[[:space:]]in[[:space:]]Video[[:space:]]Enable[[:space:]]Segmentation/dfff92a5-3ca1-43fe-8e97-b9057b1c72e6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MP-Former_[[:space:]]Mask-Piloted[[:space:]]Transformer[[:space:]]for[[:space:]]Image[[:space:]]Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MSF_[[:space:]]Motion-Guided[[:space:]]Sequential[[:space:]]Fusion[[:space:]]for[[:space:]]Efficient[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]From[[:space:]]Point[[:space:]]Cloud[[:space:]]Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MSINet_[[:space:]]Twins[[:space:]]Contrastive[[:space:]]Search[[:space:]]of[[:space:]]Multi-Scale[[:space:]]Interaction[[:space:]]for[[:space:]]Object[[:space:]]ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MSMDFusion_[[:space:]]Fusing[[:space:]]LiDAR[[:space:]]and[[:space:]]Camera[[:space:]]at[[:space:]]Multiple[[:space:]]Scales[[:space:]]With[[:space:]]Multi-Depth[[:space:]]Seeds[[:space:]]for[[:space:]]3D[[:space:]]Object[[:space:]]Detection/2069ca18-1048-442b-b642-22315aec0f29_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MSeg3D_[[:space:]]Multi-Modal[[:space:]]3D[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MV-JAR_[[:space:]]Masked[[:space:]]Voxel[[:space:]]Jigsaw[[:space:]]and[[:space:]]Reconstruction[[:space:]]for[[:space:]]LiDAR-Based[[:space:]]Self-Supervised[[:space:]]Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MVImgNet_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Dataset[[:space:]]of[[:space:]]Multi-View[[:space:]]Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MaLP_[[:space:]]Manipulation[[:space:]]Localization[[:space:]]Using[[:space:]]a[[:space:]]Proactive[[:space:]]Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MaPLe_[[:space:]]Multi-Modal[[:space:]]Prompt[[:space:]]Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Magic3D_[[:space:]]High-Resolution[[:space:]]Text-to-3D[[:space:]]Content[[:space:]]Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MagicNet_[[:space:]]Semi-Supervised[[:space:]]Multi-Organ[[:space:]]Segmentation[[:space:]]via[[:space:]]Magic-Cube[[:space:]]Partition[[:space:]]and[[:space:]]Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MagicPony_[[:space:]]Learning[[:space:]]Articulated[[:space:]]3D[[:space:]]Animals[[:space:]]in[[:space:]]the[[:space:]]Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Make[[:space:]]Landscape[[:space:]]Flatter[[:space:]]in[[:space:]]Differentially[[:space:]]Private[[:space:]]Federated[[:space:]]Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Make-a-Story_[[:space:]]Visual[[:space:]]Memory[[:space:]]Conditioned[[:space:]]Consistent[[:space:]]Story[[:space:]]Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Making[[:space:]]Vision[[:space:]]Transformers[[:space:]]Efficient[[:space:]]From[[:space:]]a[[:space:]]Token[[:space:]]Sparsification[[:space:]]View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MammalNet_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Video[[:space:]]Benchmark[[:space:]]for[[:space:]]Mammal[[:space:]]Recognition[[:space:]]and[[:space:]]Behavior[[:space:]]Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Manipulating[[:space:]]Transfer[[:space:]]Learning[[:space:]]for[[:space:]]Property[[:space:]]Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Mapping[[:space:]]Degeneration[[:space:]]Meets[[:space:]]Label[[:space:]]Evolution_[[:space:]]Learning[[:space:]]Infrared[[:space:]]Small[[:space:]]Target[[:space:]]Detection[[:space:]]With[[:space:]]Single[[:space:]]Point[[:space:]]Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MarS3D_[[:space:]]A[[:space:]]Plug-and-Play[[:space:]]Motion-Aware[[:space:]]Model[[:space:]]for[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]on[[:space:]]Multi-Scan[[:space:]]3D[[:space:]]Point[[:space:]]Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Marching-Primitives_[[:space:]]Shape[[:space:]]Abstraction[[:space:]]From[[:space:]]Signed[[:space:]]Distance[[:space:]]Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MarginMatch_[[:space:]]Improving[[:space:]]Semi-Supervised[[:space:]]Learning[[:space:]]with[[:space:]]Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Markerless[[:space:]]Camera-to-Robot[[:space:]]Pose[[:space:]]Estimation[[:space:]]via[[:space:]]Self-Supervised[[:space:]]Sim-to-Real[[:space:]]Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Mask[[:space:]]DINO_[[:space:]]Towards[[:space:]]a[[:space:]]Unified[[:space:]]Transformer-Based[[:space:]]Framework[[:space:]]for[[:space:]]Object[[:space:]]Detection[[:space:]]and[[:space:]]Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Mask-Free[[:space:]]OVIS_[[:space:]]Open-Vocabulary[[:space:]]Instance[[:space:]]Segmentation[[:space:]]Without[[:space:]]Manual[[:space:]]Mask[[:space:]]Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Mask-Free[[:space:]]Video[[:space:]]Instance[[:space:]]Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Mask-Guided[[:space:]]Matting[[:space:]]in[[:space:]]the[[:space:]]Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Mask3D_[[:space:]]Pre-Training[[:space:]]2D[[:space:]]Vision[[:space:]]Transformers[[:space:]]by[[:space:]]Learning[[:space:]]Masked[[:space:]]3D[[:space:]]Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MaskCLIP_[[:space:]]Masked[[:space:]]Self-Distillation[[:space:]]Advances[[:space:]]Contrastive[[:space:]]Language-Image[[:space:]]Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MaskCon_[[:space:]]Masked[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Coarse-Labelled[[:space:]]Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MaskSketch_[[:space:]]Unpaired[[:space:]]Structure-Guided[[:space:]]Masked[[:space:]]Image[[:space:]]Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Auto-Encoders[[:space:]]Meet[[:space:]]Generative[[:space:]]Adversarial[[:space:]]Networks[[:space:]]and[[:space:]]Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Autoencoders[[:space:]]Enable[[:space:]]Efficient[[:space:]]Knowledge[[:space:]]Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Autoencoding[[:space:]]Does[[:space:]]Not[[:space:]]Help[[:space:]]Natural[[:space:]]Language[[:space:]]Supervision[[:space:]]at[[:space:]]Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Image[[:space:]]Modeling[[:space:]]With[[:space:]]Local[[:space:]]Multi-Scale[[:space:]]Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Image[[:space:]]Training[[:space:]]for[[:space:]]Generalizable[[:space:]]Deep[[:space:]]Image[[:space:]]Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Images[[:space:]]Are[[:space:]]Counterfactual[[:space:]]Samples[[:space:]]for[[:space:]]Robust[[:space:]]Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Jigsaw[[:space:]]Puzzle_[[:space:]]A[[:space:]]Versatile[[:space:]]Position[[:space:]]Embedding[[:space:]]for[[:space:]]Vision[[:space:]]Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Motion[[:space:]]Encoding[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Video[[:space:]]Representation[[:space:]]Learning/ffb96eff-92ef-423e-9c85-df5684067148_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Domain[[:space:]]Generalized[[:space:]]Stereo[[:space:]]Matching/33890d9e-3bb5-4cca-822b-e87656eea803_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Scene[[:space:]]Contrast_[[:space:]]A[[:space:]]Scalable[[:space:]]Framework[[:space:]]for[[:space:]]Unsupervised[[:space:]]3D[[:space:]]Representation[[:space:]]Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Video[[:space:]]Distillation_[[:space:]]Rethinking[[:space:]]Masked[[:space:]]Feature[[:space:]]Modeling[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Video[[:space:]]Representation[[:space:]]Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]Wavelet[[:space:]]Representation[[:space:]]for[[:space:]]Compact[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/45d0202b-5376-4366-b728-470ed1f9b25b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Masked[[:space:]]and[[:space:]]Adaptive[[:space:]]Transformer[[:space:]]for[[:space:]]Exemplar[[:space:]]Based[[:space:]]Image[[:space:]]Translation/122f9675-10c2-4e31-827b-15924283ce1b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Master_[[:space:]]Meta[[:space:]]Style[[:space:]]Transformer[[:space:]]for[[:space:]]Controllable[[:space:]]Zero-Shot[[:space:]]and[[:space:]]Few-Shot[[:space:]]Artistic[[:space:]]Style[[:space:]]Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Matching[[:space:]]Is[[:space:]]Not[[:space:]]Enough_[[:space:]]A[[:space:]]Two-Stage[[:space:]]Framework[[:space:]]for[[:space:]]Category-Agnostic[[:space:]]Pose[[:space:]]Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MeMaHand_[[:space:]]Exploiting[[:space:]]Mesh-Mano[[:space:]]Interaction[[:space:]]for[[:space:]]Single[[:space:]]Image[[:space:]]Two-Hand[[:space:]]Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Megahertz[[:space:]]Light[[:space:]]Steering[[:space:]]Without[[:space:]]Moving[[:space:]]Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Memory-Friendly[[:space:]]Scalable[[:space:]]Super-Resolution[[:space:]]via[[:space:]]Rewinding[[:space:]]Lottery[[:space:]]Ticket[[:space:]]Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Meta[[:space:]]Architecture[[:space:]]for[[:space:]]Point[[:space:]]Cloud[[:space:]]Analysis/ef212f38-9097-46a0-9adc-735684e16116_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Meta[[:space:]]Compositional[[:space:]]Referring[[:space:]]Expression[[:space:]]Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Meta[[:space:]]Omnium_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]General-Purpose[[:space:]]Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Meta-Causal[[:space:]]Learning[[:space:]]for[[:space:]]Single[[:space:]]Domain[[:space:]]Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Meta-Explore_[[:space:]]Exploratory[[:space:]]Hierarchical[[:space:]]Vision-and-Language[[:space:]]Navigation[[:space:]]Using[[:space:]]Scene[[:space:]]Object[[:space:]]Spectrum[[:space:]]Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Meta-Learning[[:space:]]With[[:space:]]a[[:space:]]Geometry-Adaptive[[:space:]]Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Meta-Personalizing[[:space:]]Vision-Language[[:space:]]Models[[:space:]]To[[:space:]]Find[[:space:]]Named[[:space:]]Instances[[:space:]]in[[:space:]]Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Meta-Tuning[[:space:]]Loss[[:space:]]Functions[[:space:]]and[[:space:]]Data[[:space:]]Augmentation[[:space:]]for[[:space:]]Few-Shot[[:space:]]Object[[:space:]]Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MetaCLUE_[[:space:]]Towards[[:space:]]Comprehensive[[:space:]]Visual[[:space:]]Metaphors[[:space:]]Research/5a097123-23c7-455f-934d-5ab8f12c4f80_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MetaFusion_[[:space:]]Infrared[[:space:]]and[[:space:]]Visible[[:space:]]Image[[:space:]]Fusion[[:space:]]via[[:space:]]Meta-Feature[[:space:]]Embedding[[:space:]]From[[:space:]]Object[[:space:]]Detection/70dfb2c3-4648-4819-a764-7cd79a227474_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MetaMix_[[:space:]]Towards[[:space:]]Corruption-Robust[[:space:]]Continual[[:space:]]Learning[[:space:]]With[[:space:]]Temporally[[:space:]]Self-Adaptive[[:space:]]Data[[:space:]]Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MetaPortrait_[[:space:]]Identity-Preserving[[:space:]]Talking[[:space:]]Head[[:space:]]Generation[[:space:]]With[[:space:]]Fast[[:space:]]Personalized[[:space:]]Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MetaViewer_[[:space:]]Towards[[:space:]]a[[:space:]]Unified[[:space:]]Multi-View[[:space:]]Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Metadata-Based[[:space:]]RAW[[:space:]]Reconstruction[[:space:]]via[[:space:]]Implicit[[:space:]]Neural[[:space:]]Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/MethaneMapper_[[:space:]]Spectral[[:space:]]Absorption[[:space:]]Aware[[:space:]]Hyperspectral[[:space:]]Transformer[[:space:]]for[[:space:]]Methane[[:space:]]Detection/380575b8-d276-40f4-954b-81a84c333f47_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Micron-BERT_[[:space:]]BERT-Based[[:space:]]Facial[[:space:]]Micro-Expression[[:space:]]Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_content_list.json b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a0ec1f67e9a9a5c500fca125ad5c47a884fca58d --- /dev/null +++ b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_content_list.json @@ -0,0 +1,1488 @@ +[ + { + "type": "text", + "text": "MP-Former: Mask-Piloted Transformer for Image Segmentation", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Zhang $^{1,4*†}$ , Feng Li $^{1,4*†}$ , Huaizhe Xu $^{1,4}$ , Shijia Huang $^{2,4}$ , Shilong Liu $^{3,4}$ , Lionel M. Ni $^{5,1}$ , Lei Zhang $^{4‡}$", + "bbox": [ + 222, + 178, + 743, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Hong Kong University of Science and Technology.", + "bbox": [ + 263, + 217, + 705, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2The Chinese University of Hong Kong.", + "bbox": [ + 326, + 233, + 643, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Dept. of CST., BNRist Center, Institute for AI, Tsinghua University.", + "bbox": [ + 210, + 251, + 754, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4International Digital Economy Academy (IDEA).", + "bbox": [ + 285, + 268, + 683, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5The Hong Kong University of Science and Technology (Guangzhou).", + "bbox": [ + 207, + 286, + 759, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 312, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a mask-piloted Transformer which improves masked-attention in Mask2Former for image segmentation. The improvement is based on our observation that Mask2Former suffers from inconsistent mask predictions between consecutive decoder layers, which leads to inconsistent optimization goals and low utilization of decoder queries. To address this problem, we propose a mask-piloted training approach, which additionally feeds noised ground-truth masks in masked-attention and trains the model to reconstruct the original ones. Compared with the predicted masks used in mask-attention, the ground-truth masks serve as a pilot and effectively alleviate the negative impact of inaccurate mask predictions in Mask2Former. Based on this technique, our MP-Former achieves a remarkable performance improvement on all three image segmentation tasks (instance, panoptic, and semantic), yielding $+2.3AP$ and $+1.6mIoU$ on the Cityscapes instance and semantic segmentation tasks with a ResNet-50 backbone. Our method also significantly speeds up the training, outperforming Mask2Former with half of the number of training epochs on ADE20K with both a ResNet-50 and a Swin-L backbones. Moreover, our method only introduces little computation during training and no extra computation during inference. Our code will be released at https://github.com/IDEA-Research/MP-Former.", + "bbox": [ + 76, + 371, + 473, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 781, + 207, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image segmentation is a fundamental problem in computer vision which includes semantic, instance, and panoptic", + "bbox": [ + 76, + 806, + 470, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "segmentation. Early works design specialized architectures for different tasks, such as mask prediction models [2, 15] for instance segmentation and per-pixel prediction models [25, 28] for semantic segmentation. Panoptic segmentation [18] is a unified task of instance and semantic segmentation. But universal models for panoptic segmentation such as Panoptic FPN [18] and K-net [34] are usually suboptimal on instance and semantic segmentation compared with specialized models.", + "bbox": [ + 500, + 340, + 893, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Vision Transformers [29] has achieved tremendous success in many computer vision tasks, such as image classification [11, 24] and object detection [1]. Inspired by DETR's set prediction mechanism, Mask2Former [7] proposes a unified framework and achieves a remarkable performance on all three segmentation tasks. Similar to DETR, it uses learnable queries in Transformer decoders to probe features and adopts bipartite matching to assign predictions to ground truth (GT) masks. Mask2Former also uses predicted masks in a decoder layer as attention masks for the next decoder layer. The attention masks serve as a guidance to help next layer's prediction and greatly ease training.", + "bbox": [ + 496, + 477, + 893, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its great success, Mask2Former is still an initial attempt which is mainly based on a vanilla DETR model that has a sub-optimal performance compared to its later variants. For example, in each decoder layer, the predicted masks are built from scratch by dot-producting queries and feature map without performing layer-wise refinement as proposed in deformable DETR [37]. Because each layer build masks from scratch, the masks predicted by a query in different layers may change dramatically. To show this inconsistent predictions among layers, we build two metrics mIoU- $L^i$ and $Util^i$ to quantify this problem and analyze its consequences in Section 3. As a result, this problem leads to a low utilization rate of decoder queries, which is especially severe in the first few decoder layers. As the attention masks predicted from an early layer are usually inaccurate, when serving as a guidance for its next layer, they will lead to", + "bbox": [ + 496, + 659, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 849, + 205, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was done when Hao Zhang and Feng Li were interns at IDEA.", + "bbox": [ + 78, + 863, + 468, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18074", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sub-optimal predictions in the next layer.", + "bbox": [ + 76, + 90, + 351, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since detection is a similar task to segmentation, both aiming to locate instances in images, we seek inspirations from detection to improve segmentation. For example, the recently proposed denoising training [19] for detection has shown very effective to stabilize bipartite matching between training epochs. DINO [33] achieves a new SOTA on COCO detection built upon an improved denoising training method. The key idea in denoising training is to feed noised GT boxes in parallel with learnable queries into Transformer decoders and train the model to denoise and recover the GT boxes.", + "bbox": [ + 75, + 106, + 470, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by DN-DETR [19], we propose a mask-piloted (MP) training approach. We divide the decoder queries into two parts, an MP part and a matching part. The matching part is the same as in Mask2Former, feeding learnable queries and adopting bipartite matching to assign predictions to GT instances. In the MP part, we feed class embeddings of GT categories as queries and GT masks as attention masks into a Transformer decoder layer and let the model to reconstruct GT masks and labels. To reinforce the effectiveness of our method, we propose to apply MP training for all decoder layers. Moreover, we also add point noises to GT masks and flipping noises to GT class embeddings to force the model to recover GT masks and labels from inaccurate ones.", + "bbox": [ + 75, + 257, + 470, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows.", + "bbox": [ + 96, + 454, + 387, + 467 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a mask-piloted training approach to improving masked-attention in Mask2Former, which suffers from inconsistent and inaccurate mask predictions among layers. We also develop techniques including multi-layer mask-piloted training with point noises and label-guided training to further improve the segmentation performance. Our method is computationally efficient, introducing no extra computation during inference and little computation during training.", + "2. Through analyzing the predictions of Mask2Former, we discover a failure mode of Mask2Former—inconsistent predictions between consecutive layers. We build two novel metrics to measure this failure, which are layer-wise query utilization and mean IoU. We also show that our method improves the two metrics by a large margin, which validates its effectiveness on alleviating inconsistent predictions.", + "3. When evaluating on three datasets, including ADE20k, Cityscapes, and MS COCO, our MP-Former outperforms Mask2Former by a large margin. Besides improved performance, our method also speeds up training. Our model exceeds Mask2Former on all three segmentation tasks with only half of the number of training steps on ADE20k. Also, our model trained for 36 epochs is comparable with Mask2Former trained for 50 epochs on instance and panoptic segmentation on MS COCO." + ], + "bbox": [ + 89, + 474, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generally, image segmentation can be divided into three tasks including instance segmentation, semantic segmentation, and panoptic segmentation with respect to different semantics.", + "bbox": [ + 498, + 114, + 893, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Segmentation models Traditionally, researchers develop specialized models and optimization objectives for each task. Instance segmentation is to predict a set of binary masks and their associated categories. Previous methods often predict masks based on bounding boxes produced by detection models. Mask R-CNN [15] builds segmentation upon detection model Faster R-CNN [27] by adding a mask branch in parallel with the detection branch. HTC [2] further proposes to interleave the two branches and add mask information flow to improve segmentation performance. Semantic segmentation focus on category-level semantics without distinguishing instances. Previous models widely formulate it into a per-pixel classification problem. The pioneering work, FCN [25] generates a label for each pixel to solve this problem. Many follow-up works continue with this idea and design more precise pixel-level classification models [3, 4]. Panoptic segmentation [18, 30] is a combination of the two segmentation tasks above to segment both the foreground instances (\"thing\") and background semantics (\"stuff\").", + "bbox": [ + 498, + 176, + 893, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers for segmentation Transformer-based segmentation models emerge with DETR (DEtection Transformer) [1]. Since that, Transformer [29] has been widely used in many detection [1, 19, 23, 33] and segmentation models [7, 8, 32]. Generally, these models adopt a set prediction objective with bipartite matching to query features of interest. Inspiring progress has been made in many specialized query-based architectures for instance segmentation [13], semantic segmentation [32, 35], panoptic segmentation [21]. Notably, some recent works unify the three segmentation tasks into one model [7, 8, 20, 34]. Among them, Mask2Former [7] proposes masked attention and achieves remarkable performance on all three segmentation tasks. Following DETR, Mask2Former is an end-to-end Transformer for image segmentation. It is composed of a backbone, a pixel decoder, and a Transformer decoder. The pixel decoder is an encoder-only deformable Transformer that takes multi-scale image features as input and output processed multi-scale features. The Transformer decoder takes learnable queries as input to attend to the feature maps output by the pixel decoder. Each decoder layer attends to a feature map of one scale following a coarse-to-fine manner. The predicted masks of one decoder layer will feed to the next layer as the attention masks, which enables the following decoder to focus on more meaningful regions. The design of attention masks eases the training and improves performance.", + "bbox": [ + 498, + 478, + 893, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ease Training for Vision transformers Vision transformers have an advantage over the convolution-based methods due", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18075", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c3af7d9e6634f3437a752fb5eb51768471217a049a4fd52abbac09e33aaf446a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 186, + 87, + 279, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b7e32641f7349b6af0392d8ad474c1a5a56b598144954c2924fbba8166c17d52.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 339, + 87, + 431, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/948410ed82d867ea03ec35c812a6c12d76f33491182ce16394a9a9a00f2c91db.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 192, + 156, + 285, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7ac3c9784191bf7037977df8cdc3e185d93692f0a9eafd4691aab062dc6c28f5.jpg", + "image_caption": [ + "(f)", + "Figure 1. We visualize layer-wise predictions of Mask2Former and show 4 pairs of failure cases. Each pair is the predictions of the same query in adjacent decoder layers. The red regions are the predicted masks. These cases show that the predictions of a query may change dramatically between consecutive layers." + ], + "image_footnote": [], + "bbox": [ + 344, + 156, + 439, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d8207ffc706fa8b6922a1cb7477a4af8a20f9935dde61c27e5d94037bff68c67.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 540, + 94, + 632, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3337589aaf9d9945a6cce969bc392970f60349babb89913d006c85df7e630414.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 694, + 94, + 785, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dc63c027b28d3a20e6f82405d8fd905b288b4665509e11f6ef82e5662a537e47.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 537, + 156, + 629, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/135dcd337bf1b9fc4bcf3b662327ef035cb7c9d9d302830e8037f05b808509cb.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 687, + 156, + 779, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to their global attention—they can see all parts of the image. However, it is hard to train because globally searching for an object is a hard task. This phenomenon exists in both detection and segmentation. In detection, DETR suffers from slow convergence requiring 500 training epochs for convergence. Recently, researchers have dived into the meaning of the learnable queries [23, 26, 31, 37]. They either express the queries as reference points or anchor boxes. [19, 33] proposed to add noised ground truth boxes as positional queries for denoising training and they speed up detection greatly. In segmentation, Mask2Former proposed mask attention which makes training easier and speeds up convergence when compared with MaskFormer. We have found some common points for these works that speed up training for vision Transformers. Firstly, they usually give clearer meaning to learnable queries to reduce ambiguity. Secondly, they give local limitations to cross attention and reduce the scope for the model to search objects. Therefore, we believe that giving guidance to cross attention leads to good segmentation performance. Our method is similar to DN-DETR [19] in introducing GT annotations to the Transformer decoder but ours is totally different from it in both the problems we solve and the method we take. Firstly, our method aims to address the inconsistent prediction problem which leads to the low utilization of queries, while DN-DETR aims to solve unstable matching among epochs. Secondly, we use GT masks as the attention masks, while DN-DETR feeds noised GT boxes as decoder queries. Thirdly, noises are optional in our method, but DN-DETR does not work without noises.", + "bbox": [ + 76, + 334, + 472, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Failure Cases and Solution", + "text_level": 1, + "bbox": [ + 76, + 784, + 326, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Failure cases: We view the mask prediction in Mask2Former as a mask refinement process through Transformer decoder layers. By taking predictions from a layer as the attention masks of the next layer, we expected to obtain refined predictions from the next layer. Generally, the predicted masks are assigned to the GT masks by bipartite matching and", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "refined in a coarse-to-fine procedure. Therefore, keeping refinement consistency in different layers is the key to refining the predicted masks consistently. However, when we visualize the predictions of Mask2Former layer by layer, we found a severe inconsistency problem among the predictions from the same query between consecutive decoder layers. As shown in Fig. 1, the predictions of adjacent layers differ significantly. We define a metric named layer-wise mean Intersection-over-Union (IoU) denoted as $\\mathrm{mIoU - L}^i$ to quantify this inconsistency. We denote the predicted masks from the $i$ -th decoder layer as $\\mathbf{M}^{\\mathbf{i}} = \\left\\{M_0^i,M_1^i,\\dots,M_{N - 1}^i\\right\\}$ , where $N$ is the number of decoder queries. Note that $M^0$ means the predictions before queries enter the first Transformer decoder layer. We have", + "bbox": [ + 496, + 334, + 895, + 546 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {m I o U - L} ^ {i} = \\sum_ {n = 0} ^ {N - 1} \\mathrm {I o U} \\left(M _ {n} ^ {i - 1}, M _ {n} ^ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 556, + 892, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Table 1, the mIoU-L is quite low, especially for the first few layers, which indicates inconsistent mask prediction results between adjacent layers. Such an inconsistent result means a query may be assigned to different instances (or background for panoptic and semantic segmentation) during layer-by-layer updates. To evaluate the inconsistent matching problem, we define another metric named as layer-wise query utilization denoted as $\\text{Util}^i$ for the $i$ -th layer, which means the ratio of the matched queries in the $i$ -th layer that matches exactly the same instance as in the last layer. We denote the ground truth objects as $\\mathbf{T} = \\{T_0, T_1, T_2, \\dots, T_{O-1}\\}$ where $O$ is the number of ground truth objects. After bipartite matching, we compute an index vector $\\mathbf{V}^i = \\{V_0^i, V_1^i, \\dots, V_{N-1}^i\\}$ to store the matching result of layer $i$ as follows.", + "bbox": [ + 496, + 606, + 893, + 818 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n} ^ {i} = \\left\\{ \\begin{array}{l l} o, & \\text {i f} M _ {n} ^ {i} \\text {m a t c h e s} T _ {o} \\\\ - 1, & \\text {i f} M _ {n} ^ {i} \\text {m a t c h e s n o t h i n g} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 828, + 892, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define the query utilization of layer $i$ as the proportion of GT masks that match with the same query in layer $i$ and", + "bbox": [ + 496, + 869, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18076", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/36467bc75245240489b7851b8cde2d0f16b8b776db05d51ceff2438f86061aab.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. The mIoU-L $^i$ and Util $^i$ of Mask2Former and the proposed methods. The mIoU-L $^i$ for a query is taken as the IoU of the predicted masks in the $(i-1)$ -th and $i$ -th decoder layers. The mIoU-L $^i$ takes the average over all queries. Note that the index of decoder layers starts from 1, so \"prediction of the 0-th decoder\" means the prediction before entering Transformer decoder. The Util $^i$ is the ratio of GT masks that match with the same query in $i$ -th layer and the last layer. * means that We test Util $^i$ with bipartite matching to assign predictions in MP part to GT masks. Note that when directly assigning predictions in MP part to corresponding GT masks with hard assignment, Util $^i$ is always 100%." + ], + "table_body": "
Layer Number123456789
Mask2FormermIoU-Li(%)23.050.359.569.770.169.978.676.376.0
Utili(%)38.165.570.873.578.581.183.283.2100
MP trainingmIoU-Li(%)51.583.286.696.296.093.897.697.996.3
Utili* (%)94.093.894.693.294.094.894.094.0100
Multi-layer MPmIoU-Li(%)51.387.888.897.597.596.299.098.898.2
Utili* (%)94.295.497.397.697.397.098.298.2100
", + "bbox": [ + 256, + 90, + 715, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the last layer, which is calculated as", + "bbox": [ + 76, + 310, + 316, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {U t i l} ^ {i} = \\frac {1}{O} \\sum_ {j = 0} ^ {N} \\mathbb {1} \\left(V _ {n} ^ {i} = V _ {n} ^ {l} \\text {a n d} V _ {n} ^ {i} \\neq - 1\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 334, + 470, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is the indicator function. $\\mathbb{1}(x) = 1$ if $x$ is true and 0 otherwise. The query utilization of layer $i$ for the whole data set is averaged over the utilization numbers for all images. Table 1 shows that the query utilizations of the first few layers are too low, which indicate many matched queries in early layers are ended up not being used in the last layer. The low mIoU- $L^i$ and $\\mathrm{Util}^i$ have potential consequences such as inconstant optimization goals and low training efficiency.", + "bbox": [ + 75, + 386, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our observations based on these two metrics motivate us to develop a more effective training method to make the matching more consistent among layers.", + "bbox": [ + 75, + 508, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Solution: To address the inconsistent prediction problem, inspired by DN-DETR, we proposed a mask-piloted training method. We feed GT masks as attention masks and expect these queries with GT masks can focus on their corresponding GT masks and not be distracted by other instances. As shown in row 2 and 3, Table 1, both mIoU- $L^i$ and $\\text{Util}^i$ are improved by a large margin when adopting mask-piloted training. We also find that only applying mask-piloted (MP) training in the first layer is not enough. Fig 3(a)(b) shows that inconsistent prediction exists even when given GT masks. Therefore, we add a GT mask to all decoder layers to reinforce MP training. As shown in Table 1, multi-layer MP training further improves both mIoU- $L^i$ and $\\text{Util}^i$ especially in latter decoder layers.", + "bbox": [ + 75, + 551, + 470, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. MP-Former", + "text_level": 1, + "bbox": [ + 76, + 777, + 204, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 801, + 187, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method is a training method for improving Mask2Former. Our MP-Former only introduces minor changes in the Mask2Former pipeline during training as shown in the red-line part in Fig. 2. Same as Mask2Former, our model is composed of a backbone, a pixel decoder, and", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a Transformer decoder. Our improvements are on the Transformer decoder. We divide decoder queries into two parts, a mask-piloted (MP) part and a matching part. The matching part is the same as Mask2Former. The MP part feeds GT masks and GT class embeddings as input and the predictions are directly assigned to the corresponding GT masks without bipartite matching. Our method includes three components which are multi-layer MP training, mask noises, and label-guided training.", + "bbox": [ + 496, + 310, + 893, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Multi-layer MP training", + "text_level": 1, + "bbox": [ + 498, + 453, + 723, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to address the inconsistent prediction problem, we propose a multi-layer mask-piloted (MP) training method. In the MP part, we feed an additional set of queries and masks into Transformer where the queries and masks are GT class embeddings and GT masks. On the loss side, we independently assign the outputs of the MP part and the matching part with the GT instances. The predictions of the MP part are directly assigned to their corresponding GT masks and those in the matching part are assigned with bipartite graph matching. The losses for both parts are the same, following the loss design in Mask2Former. We also propose to add GT masks in multiple layers and show that this can further improve the performance with experiments in Table 8. Since we follow Mask2Fomer to use feature maps of different resolutions in different decoder layers, we interpolate GT masks into different resolutions when applied to different layers.", + "bbox": [ + 496, + 477, + 893, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Noised masks", + "text_level": 1, + "bbox": [ + 498, + 741, + 640, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides feeding GT masks to alleviate the inconsistency problem, we further propose to use noised GT masks to make mask refinement more robust in each decoder. The intuition is simple, as Mask2former refines the mask from an inaccurate mask predicted from the previous decoder. GT masks without noise may be too easy for the task which prevents further refinement. Therefore, we feed noised masks to the decoder and train the model to reconstruct the original ones, which is similar to DN-DETR [19] for box denoising", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18077", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ef37a30044fec179f93f0049192a0895496c5a34aca5ffc34c351966353ad34.jpg", + "image_caption": [ + "Figure 2. The architecture of our method is the same as Mask2Former (the blue-shaded part), which consists of a backbone, a pixel decoder, and a Transformer decoder. The difference is that we feed extra queries and attention masks which are called the MP part to the Transformer decoder (red-line part in the figure). The MP part contains GT masks as attention masks and GT class embeddings as queries. We feed GT masks into the MP part of all decoder layers. We also add point noises to GT masks and flipping noises to class embeddings which can further improve the performance. Note that this architecture is just for training. In the inference time, the red-line part does not exist, and thus, our pipeline is exactly the same as Mask2Former." + ], + "image_footnote": [], + "bbox": [ + 183, + 75, + 787, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ffb0406ad8ecb8c0d139e417bf7dc7a71ce5409276ee7822ba95b63aedbdb58d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 153, + 470, + 285, + 604 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d8dabe7dcf632b3457faf852f0307df0e9a4e540580e91d75365d2bc03571e91.jpg", + "image_caption": [ + "(b)", + "Figure 3. (a) The red region is a given GT mask in first decoder layer which covers the man riding the black horse. (b) The predicted mask of the first decoder layer which only covers the head of the man. The predicted mask will be used as the attention mask of the second layer and may mislead the second layer. (c)(d)(e) are demonstrations of point, shift and scale noises. The blue regions are noised masks and the red regions are the GT masks." + ], + "image_footnote": [], + "bbox": [ + 287, + 470, + 419, + 604 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/021dd0923be2447f4d3feb4a6d2a45488a564226801cc11d865edf0411c1ce4a.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 421, + 470, + 552, + 604 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6c15f3fec374963edf72c391848f3981debf8374fda855e754995b7580752d09.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 553, + 470, + 686, + 604 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3102a2f869c631bf2502811a4bc99b99e4674d7779db3e4fc5de5b235b5fdeff.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 687, + 470, + 820, + 604 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d1542bc0fcda542308452306a9e534b1c6ed9010c16cf256292329df0a4b8a39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodbackbonePQpanoptic APth panmIoUpaninstancesemantic mIoU
APAPSAPMAPL
MaskFormer [8]R5034.7-------
Panoptic-DeepLab [6]SWideRNet [5]37.9*-50.0*-----
Mask2Former [7]R5039.726.546.126.410.428.943.147.2
MP-FormerR5040.8(+1.1)27.148.3(+2.2)28.0(+1.6)10.530.7(+1.8)44.648.1
MaskFormer [8]Swin-L†-------54.1
FaPN-MaskFormer [8, 17]Swin-L†-------55.2
Mask2Former [7]Swin-L†48.134.254.534.916.340.054.756.1
MP-FormerSwin-L†49.4(+1.3)35.256.035.316.340.755.556.9
", + "bbox": [ + 174, + 686, + 795, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Image segmentation results on ADE20K va1. MP-Former improves the performance on all three segmentation tasks with R50 and Swin-L as backbone. All metrics are evaluated with single-scale inference. Backbones pre-trained on ImageNet-22K are marked with ${}^{ \\dagger }$ . - means the results are not reported for the methods.", + "bbox": [ + 75, + 818, + 893, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training. We have tried three types of noise which are point noise, shift noise, and scale noise as shown in Fig. 3. For", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "point noise, we randomly sample some points on the mask and flip them from 1 to 0 or from 0 to 1. The number of", + "bbox": [ + 500, + 869, + 892, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18078", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f85ec653f25d7039640affb01efa4e4c17a1e4fbf89e48d956b812cf286be51d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodbackbonePQpanoptic APTh panmIoUpaninstancesemantic mIoU
APAP50
Panoptic-DeepLab [6]R5060.332.178.7---
Mask2Former [7]R5062.137.377.537.461.979.4
MP-FormerR5062.739.3(+2.0)79.6(+2.1)39.7(+2.3)65.4(+3.5)81.0(+1.6)
Mask2Former [7]Swin-L†66.643.682.943.771.483.3
MP-FormerSwin-L†67.545.8(+2.2)83.544.9(+1.2)72.483.9
", + "bbox": [ + 212, + 88, + 756, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Image segmentation results on Cityscapes va1. We report results on all three segmentation tasks with R50 or Swin-L as backbone. All metrics are evaluated with single-scale inference. Backbones pre-trained on ImageNet-22K are marked with", + "bbox": [ + 75, + 186, + 893, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sampled points is proportional to the area of a GT mask. For shift noise, we randomly shift a mask instance vertically and horizontally while keeping the center of the instance within its GT bounding box. For the scaling noise, we enlarge or shrink the instance by a random ratio. A comparison of three types of noises is shown in Table 8. Through experiments, we find point noises improve the performance while shifting and scaling noises do not work for our method. Therefore, we adopt point noises in our implementation. We use a noise ratio $\\lambda_{p}$ to control how many noise points are added. Denote a binary GT mask as $M = [m_{i,j}]_{H\\times W}$ where $m_{i,j} = 1$ means the pixel belongs to the mask. The number of noise points is randomly sampled from 0 to $\\lambda_{p}\\times \\sum_{i,j}m_{i,j}$ . Since we apply mask-piloted training to all decoder layers, we independently add different noises to GT masks in different layers.", + "bbox": [ + 75, + 234, + 472, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Label-guided training", + "text_level": 1, + "bbox": [ + 76, + 488, + 282, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the class embeddings of GT categories as queries in the MP part. The reason why we use class embeddings is that queries will dot-product with image features and an intuitive way to distinguish instances is to use their categories. Note that we only use class embeddings in the first layer and do not replace predicted queries with class embeddings in subsequent layers. We also have a classification loss in the MP part. Since the classification task becomes very easy given GT class embeddings, we add flipping noises in class embeddings. We randomly sample GT class embeddings and let them randomly flip to the class embeddings of other categories. For example, $E_0, E_1, \\dots, E_K$ denotes the class embeddings of $K$ categories. An image contains $N$ instances which belong to Categories $C^0, C^1, \\dots, C^N$ . So their default class embeddings should be $E_{C^0}, E_{C^1}, \\dots, E_{C^N}$ . To add flipping noise, we enumerate all $N$ instances. For $i$ -th instance, we generate a random number $r_i$ in [0, 1] to decide whether to flip the class embedding $E_{C^i}$ or not. If $r_i < \\lambda_l$ , we randomly sample another class embedding in $E_0, E_1, \\dots, E_K$ to replace $E_{C^i}$ . $\\lambda_l$ denotes the proportion of the sampled class embeddings.", + "bbox": [ + 75, + 512, + 472, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments to compare our method with Mask2Former and other segmentation methods", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on all three segmentation tasks on three datasets to show the effectiveness of our method. We also conduct several ablation studies to show the effectiveness of each component in our method.", + "bbox": [ + 496, + 234, + 893, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation details", + "text_level": 1, + "bbox": [ + 498, + 305, + 714, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model: Since our method is an improved training method based on Mask2Former, we adopt an identical model setting as Mask2Former except for the mask-piloted (MP) part. In the MP part, we adopt 100 GT masks with point noises. We adopt noise ratios $\\lambda_{p} = 0.2$ and $\\lambda_{l} = 0.2$ for point noises on masks and flipping noises on class embeddings. We adopt ResNet50 [16] and Swin-L [24] when evaluating our method on ADE20K and Cityscapes. On COCO2017, we use ResNet50 and ResNet101 as the backbone.", + "bbox": [ + 496, + 329, + 893, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and metrics: We evaluate MP-Former on three challenging datasets: Cityscapes [9, 10] and ADE20K [36] for all three segmentation tasks and COCO 2017 [22] dataset for panoptic segmentation and instance segmentation. For instance segmentation, we evaluate the mask AP [22] on instances that are denoted as \"thing\" categories in datasets. For semantic segmentation, we evaluate mean Intersection-over-Union (mIOU) over all categories including both foreground and background. For panoptic segmentation, we evaluate the results with the panoptic quality (PQ) metric [18].", + "bbox": [ + 496, + 465, + 893, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 498, + 628, + 638, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We trained on three challenging datasets ADE20K, Cityscapes, and MS COCO with different backbones. Our method outperforms Mask2Former on all three segmentation tasks (instance, panoptic, and semantic) under the same setting, which demonstrates its effectiveness and generalization ability to different datasets.", + "bbox": [ + 496, + 643, + 893, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on ADE20K. On ADE20K, we report our performance trained for $160K$ steps following Mask2Former. As shown in Table 2, our method exceeds Mask2Former and other baselines on all metrics with both R50 and Swin-L. Notably, MP-Former achieves $+1.6\\mathrm{AP}$ on instance segmentation with a R50 backbone and $+1.3\\mathrm{PQ}$ on panoptic segmentation with a Swin-L backbone.", + "bbox": [ + 496, + 734, + 893, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on Cityscapes. On Cityscapes, we train the model with 90K training steps following Mask2Former. As shown in Table 3, our method consistently outperforms Mask2Former on all three segmentation tasks across differ-", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18079", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/faaa9133255d1fcdf4bedade27e8e20c29b2a1c11aabb11831de73ab5f6302c0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodbackbonequery typeepochsAP\\( AP^S \\)\\( AP^M \\)\\( AP^L \\)#params.FLOPsfps
MaskFormer [8]R50100 queries30034.016.437.854.245M181G19.2
Mask R-CNN [15]R50dense anchors3637.218.639.553.344M201G15.2
Mask R-CNN [12, 14, 15]R50dense anchors40042.523.845.060.046M358G10.3
Mask2Former [7]R50100 queries5043.723.447.264.844M226G9.7
MP-FormerR50100 queries5044.8(+1.1)24.748.165.544M226G9.7
Mask R-CNN [15]R101dense anchors3638.619.541.355.363M266G10.8
Mask R-CNN [12, 14, 15]R101dense anchors40043.724.646.461.865M423G8.6
Mask2Former [7]R101100 queries5044.223.847.766.763M293G7.8
MP-FormerR101100 queries5045.1(+0.9)24.648.766.263M293G7.8
QueryInst [13]Swin-L†300 queries5048.930.852.668.3--3.3
Swin-HTC++ [2, 24]Swin-L†dense anchors7249.531.052.467.2284M1470G-
Mask2Former [7]Swin-L†200 queries10050.129.953.972.1216M868G4.0
MP-Former (ours)Swin-L†200 queries7250.8(+0.7)30.354.572.6216M868G4.0
", + "bbox": [ + 161, + 87, + 803, + 262 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45a17be18ebfcb4cde6a358565274057a4ebb76f2c906bbda1cf1ce448f7e455.jpg", + "table_caption": [ + "Table 4. Instance segmentation on COCO va12017 with 80 categories. MP-Former outperforms strong baseline Mask2Former without extra computation cost in inference time. For a fair comparison, we only consider single-scale inference and models trained using only COCO train2017 set data. Backbones pre-trained on ImageNet-22K are marked with" + ], + "table_footnote": [], + "table_body": "
methodbackbonequery typeepochsPQ\\( PQ^{Th} \\)\\( PQ^{St} \\)\\( AP^{Th}_{pan} \\)\\( mIoU_{pan} \\)#params.FLOPsfps
DETR [1]R50100 queries500+2543.448.236.331.1----
MaskFormer [8]R50100 queries30046.551.039.833.057.845M181G17.6
Mask2Former [7]R50100 queries5051.4*57.242.741.961.544M226G8.6
MP-Former (ours)R50100 queries5052.2(+0.8)58.342.942.561.644M226G8.6
DETR [1]R101100 queries500+2545.150.537.033.0----
MaskFormer [8]R101100 queries30047.652.540.334.159.364M248G14.0
Mask2Former [7]R101100 queries5052.658.543.742.662.463M293G7.2
MP-Former (ours)R101100 queries5052.9(+0.3)58.744.043.061.863M293G7.2
Max-DeepLab [30]Max-L128 queries21651.157.042.2--451M3692G-
MaskFormer [8]Swin-L†100 queries30052.758.544.040.164.8212M792G5.2
K-Net [34]Swin-L†100 queries3654.660.246.0-----
Mask2FormerSwin-L†200 queries10057.864.248.148.667.4216M868G4.0
MP-FormerSwin-L†200 queries7258.1(+0.3)64.448.448.967.6216M868G4.0
", + "bbox": [ + 171, + 321, + 797, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Panoptic segmentation on COCO panoptic va12017 with 133 categories. MP-Former outperforms Mask2Former on COCO panoptic segmentation with both R50 and R101 as backbone. Results that are implemented are marked with * .", + "bbox": [ + 75, + 503, + 892, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ent backbones. MP-Former also scales well when adopting Swin-L as the backbone. Especially, our method outperforms Mask2Former by $+2.3\\mathrm{AP}$ on instance segmentation and $+1.6\\mathrm{mIoU}$ on semantic segmentation with R50.", + "bbox": [ + 75, + 550, + 470, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on COCO. On COCO val2017, we follow Mask2Former's setting to train for 50epochs. As shown in Table 4 and Table 5, MP-Former outperforms both Mask2Former and classical methods such as Mask R-CNN [15] on instance segmentation and panoptic segmentation. For instance segmentation, our method achieves +1.1AP with R50, +0.9AP with R101 and +0.7 with Swin-L [24]. Note that we only train MP-Former for 72 epochs with Swin-L. On panoptic segmentation, our method achieves +0.8PQ with R50 and +0.3PQ with R101. Notably, our method is only a pluggable training method that will not affect the model architecture or inference speed. In addition, it only introduces negligible parameters (class embeddings) during training.", + "bbox": [ + 75, + 611, + 470, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Speed up training", + "text_level": 1, + "bbox": [ + 76, + 832, + 251, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 6, we show our MP-Former can significantly speed up training on ADE20k and COCO. We do not report the results on Cityscapes because this dataset is relatively", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "small and only needs 90k steps to converge in Mask2Former, which is not necessary for additional speedup. The results show that our method trained with 36 epochs achieved comparable results with Mask2Former trained with 50 epochs with both R50 and R101 as the backbone on both panoptic and instance segmentation. Moreover, our method trained with half number of training steps (80K steps) achieves comparable results with Mask2Former trained with 160K steps. Note that we use the same batch size as Mask2Former. We also report the clock time. The results show that our method reduces the actual training time by large. For a fair comparison, we use the same number of GPUs to run MP-Former and Mask2Former. Because there may exist differences in speed for different machines, we run our method and Mask2Former on the same machine to test the running time.", + "bbox": [ + 496, + 550, + 893, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 786, + 653, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study on the COCO instance segmentation task with R50 as backbone.", + "bbox": [ + 496, + 810, + 890, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of each component: We show the effectiveness of each version of our method trained for 12 epochs and 50 epochs in Table 7, where we show the result of raw Mask2Former and the result after each version of our method", + "bbox": [ + 496, + 840, + 893, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18080", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b4144989caf7c37b9f2b843192672d337991c9679e1697391e41edc1987a2c5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
datasetmethodbackboneepochs*PQpanopticAPAPSinstancesemantic
APthpanmIoUpanclocktime/hAPMAPLclocktime/hmIoUclocktime/h
ADE20kMask2FormerR50160k39.726.546.152.826.410.428.943.140.347.221.3
MP-FormerR5080k39.727.048.434.227.410.230.044.525.547.412.3
Mask2FormerSwin-L†160k48.134.254.563.234.916.340.054.751.456.139.8
MP-FormerSwin-L†80k48.534.756.145.735.215.839.655.626.756.921.5
COCOMask2FormerR505051.4*41.961.577.943.723.447.264.862.5--
MP-FormerR503651.541.861.659.544.124.147.465.350.8--
Mask2FormerR1015052.642.662.4113.144.223.847.766.786.0--
MP-FormerR1013652.542.662.084.444.223.247.966.671.5--
", + "bbox": [ + 151, + 87, + 823, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/36d2f876229b73493d672388176fd90619c5225bebde994e15cd5135b2369dc0.jpg", + "table_caption": [ + "Table 6. Image segmentation results on ADE20K va1. and COCO va12017 MP-Former speeds up training by large. On ADE20K, MP-Former trained for 80K steps can achieve comparable result with Mask2Former trained for 160K steps with both R50 and Swin-L as backbone. On COCO, MP-Former trained for 36 epochs can achieve comparable result with Mask2Former trained for 50 epochs with both R50 and R101 as backbone. * means that we adopt number of steps instead of epochs on ADE20k following Mask2Former. We report only single-scale inference results. Backbones pre-trained on ImageNet-22K are marked with †" + ], + "table_footnote": [ + "Table 7. The effectiveness of each component of our method when trained for 12 epochs and 50 epochs. The table shows that each component of our model effectively improves the results. Label noise does not work when trained for 50 epochs. Especially, our method without noises can improve the results by $+0.9\\mathrm{AP}$ . With multi-layer MP training, our method achieves $+1.5\\mathrm{AP}$ on COCO instance segmentation with R50 backbone trained for 12 epochs. \"ep\" is short for epochs." + ], + "table_body": "
#Row12 ep (AP mask)50 ep (AP mask)
1. Mask2Former [7]38.743.7
2. Row1+MP training on 1st layer39.444.2
3. Row2+Label noises39.644.2
4. Row3+Point noises39.944.6
5. Ours (Row4+All-layer MP)40.244.8
", + "bbox": [ + 80, + 338, + 405, + 404 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e101018c147e6735c3ccb1d19edb501526673d457d1bc8225a4cbcae8999cc12.jpg", + "table_caption": [], + "table_footnote": [ + "Table 8. Top 3 lines shows the effectiveness of multi-layer Mask-Piloted (MP) training. The bottom lines of the table shows that shifting and scaling noise do not work for our method. \"ep\" is short for epochs." + ], + "table_body": "
Layers for MP trainingepAP mask
First layer1239.4
First three layers1239.6(+0.2)
All layers1240.2(+0.8)
Noise Type
No noise1239.6
Point noise1239.9
Shifting noise1239.6
Scaling noise1239.4
", + "bbox": [ + 413, + 335, + 616, + 431 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/92720e287f54cbdb99c79ed8eb395d8e74b1cba5f458c14af84bea7d7fd9a4b7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 9. Some hyperparemeter tuning:1.coarse-to-fine manner works better for our method than all-large feature maps. 2.Hard assignment lead to better results than matching for MP part queries. \"ep\" is short for epochs." + ], + "table_body": "
Feature mapepAP mask
Coarse to fine1240.2
All large1240.0
Assignment in Mask-Piloted part
Match1239.4
Hard assignment1239.6
", + "bbox": [ + 640, + 335, + 885, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is adopted. Without adding noises, our method achieves an improvement of $+0.9AP$ . With multi-layer MP-training, our method achieves $+1.5AP$ . We also find that label noise does not work when trained for 50 epochs. Note that except for the last line of the table, other lines of our method apply MP training on the first 3 layers.", + "bbox": [ + 75, + 502, + 468, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Apply MP training on multiple layers: In Table 8, we show the effectiveness of applying GT masks on multiple layers. The results indicate that more layers of GT masks can lead to better performance. By adding noised GT masks in all layers, MP-Former achieves $+0.8\\mathrm{AP}$ .", + "bbox": [ + 75, + 595, + 468, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Add noises to GT masks: In the last four rows of Table 8, we show the effectiveness of adding different types of noises. Note that without any noise, MP-Former achieves 39.6AP which is corresponding to the 4th row in Table 7. Shifting noise and scaling noise do not work in our method. Point noise achieved $+0.3$ extra growth.", + "bbox": [ + 75, + 671, + 470, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Some hyper-parameter tuning: Mask2Former adopts coarse-to-fine feature maps in decoder layers. We try to adopt the largest feature map (1/8 feature map) in all decoder layers. Table 9 shows that a coarse-to-fine manner is better than using the largest feature map in all layers for our MP-Former. We also try to use matching to assign predictions to GT instances in the MP part. The result shows that matching can also improve the result compared with raw Mask2Former but the hard assignment is better.", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 501, + 619, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have analyzed failure cases of Mask2Former and found a problem of inconsistent predictions between adjacent layers which potentially leads to inconstant optimization goals and low utilization of queries and therefore limits the performance of Mask2Former. In order to address the problem, we developed a new MP-Former, whose key components include multi-layer mask-piloted training with point noises and label-guided training. Our experimental results showed that the proposed training method effectively mitigates the inconsistent prediction problem. When evaluating on three challenging datasets (ADE20K, Citescapes, and MS COCO), our method achieved improvement by a large margin on all three image segmentation tasks (semantic, instance, and panoptic). In addition, our method only introduces little computation during training and no extra computation in inference time. Finally, our method can also significantly speed up training. Especially, when trained with half number of steps, our method exceeds Mask2Former on all three segmentation tasks on ADE20K with both R50 and Swin-L as the backbones.", + "bbox": [ + 496, + 553, + 893, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: We evaluated MP-Former on Mask2Former, but our method may be useful to some other models such as MaskFormer and HTC. We will take it as our future work.", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18081", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 7", + "[2] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 1, 2, 7", + "[3] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2", + "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2", + "[5] Liang-Chieh Chen, Huiyu Wang, and Siyuan Qiao. Scaling wide residual networks for panoptic segmentation. arXiv preprint arXiv:2011.11675, 2020. 5", + "[6] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12475-12485, 2020. 5, 6", + "[7] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 1, 2, 5, 6, 7, 8", + "[8] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. 2021. 2, 5, 7", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6", + "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Scharwächter, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset. In CVPR Workshop on The Future of Datasets in Vision, 2015. 6", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[12] Xianzhi Du, Barret Zoph, Wei-Chih Hung, and Tsung-Yi Lin. Simple training strategies and model scaling for object detection. arXiv preprint arXiv:2107.00057, 2021. 7" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 2, 7", + "[14] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 7", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2, 7", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[17] Shihua Huang, Zhichao Lu, Ran Cheng, and Cheng He. Fapn: Feature-aligned pyramid network for dense image prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 864-873, 2021. 5", + "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 2, 6", + "[19] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. arXiv preprint arXiv:2203.01305, 2022. 2, 3, 4", + "[20] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 2", + "[21] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Tong Lu, and Ping Luo. Panoptic segformer. arXiv preprint arXiv:2109.03814, 2021. 2", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6", + "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 6, 7", + "[25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 2", + "[26] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Con" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18082", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ditional detr for fast training convergence. arXiv preprint arXiv:2108.06152, 2021. 3", + "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2", + "[28] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 1", + "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1, 2", + "[30] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2, 7", + "[31] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. arXiv preprint arXiv:2109.07107, 2021. 3", + "[32] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[33] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2, 3", + "[34] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 7", + "[35] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2", + "[36] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6", + "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR 2021: The Ninth International Conference on Learning Representations, 2021. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 796 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18083", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_model.json b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_model.json new file mode 100644 index 0000000000000000000000000000000000000000..823c54e23a3db90dfad013b699a35b01bf4f1241 --- /dev/null +++ b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.131, + 0.812, + 0.152 + ], + "angle": 0, + "content": "MP-Former: Mask-Piloted Transformer for Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.179, + 0.744, + 0.217 + ], + "angle": 0, + "content": "Hao Zhang\\(^{1,4*†}\\), Feng Li\\(^{1,4*†}\\), Huaizhe Xu\\(^{1,4}\\), Shijia Huang\\(^{2,4}\\), Shilong Liu\\(^{3,4}\\), Lionel M. Ni\\(^{5,1}\\), Lei Zhang\\(^{4‡}\\)" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.218, + 0.706, + 0.234 + ], + "angle": 0, + "content": "1The Hong Kong University of Science and Technology." + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.234, + 0.644, + 0.252 + ], + "angle": 0, + "content": "2The Chinese University of Hong Kong." + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.252, + 0.756, + 0.27 + ], + "angle": 0, + "content": "\\(^{3}\\)Dept. of CST., BNRist Center, Institute for AI, Tsinghua University." + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.27, + 0.684, + 0.287 + ], + "angle": 0, + "content": "4International Digital Economy Academy (IDEA)." + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.287, + 0.761, + 0.305 + ], + "angle": 0, + "content": "5The Hong Kong University of Science and Technology (Guangzhou)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.313, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.372, + 0.474, + 0.751 + ], + "angle": 0, + "content": "We present a mask-piloted Transformer which improves masked-attention in Mask2Former for image segmentation. The improvement is based on our observation that Mask2Former suffers from inconsistent mask predictions between consecutive decoder layers, which leads to inconsistent optimization goals and low utilization of decoder queries. To address this problem, we propose a mask-piloted training approach, which additionally feeds noised ground-truth masks in masked-attention and trains the model to reconstruct the original ones. Compared with the predicted masks used in mask-attention, the ground-truth masks serve as a pilot and effectively alleviate the negative impact of inaccurate mask predictions in Mask2Former. Based on this technique, our MP-Former achieves a remarkable performance improvement on all three image segmentation tasks (instance, panoptic, and semantic), yielding \\(+2.3AP\\) and \\(+1.6mIoU\\) on the Cityscapes instance and semantic segmentation tasks with a ResNet-50 backbone. Our method also significantly speeds up the training, outperforming Mask2Former with half of the number of training epochs on ADE20K with both a ResNet-50 and a Swin-L backbones. Moreover, our method only introduces little computation during training and no extra computation during inference. Our code will be released at https://github.com/IDEA-Research/MP-Former." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.782, + 0.208, + 0.797 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.808, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Image segmentation is a fundamental problem in computer vision which includes semantic, instance, and panoptic" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.341, + 0.895, + 0.476 + ], + "angle": 0, + "content": "segmentation. Early works design specialized architectures for different tasks, such as mask prediction models [2, 15] for instance segmentation and per-pixel prediction models [25, 28] for semantic segmentation. Panoptic segmentation [18] is a unified task of instance and semantic segmentation. But universal models for panoptic segmentation such as Panoptic FPN [18] and K-net [34] are usually suboptimal on instance and semantic segmentation compared with specialized models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.659 + ], + "angle": 0, + "content": "Recently, Vision Transformers [29] has achieved tremendous success in many computer vision tasks, such as image classification [11, 24] and object detection [1]. Inspired by DETR's set prediction mechanism, Mask2Former [7] proposes a unified framework and achieves a remarkable performance on all three segmentation tasks. Similar to DETR, it uses learnable queries in Transformer decoders to probe features and adopts bipartite matching to assign predictions to ground truth (GT) masks. Mask2Former also uses predicted masks in a decoder layer as attention masks for the next decoder layer. The attention masks serve as a guidance to help next layer's prediction and greatly ease training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Despite its great success, Mask2Former is still an initial attempt which is mainly based on a vanilla DETR model that has a sub-optimal performance compared to its later variants. For example, in each decoder layer, the predicted masks are built from scratch by dot-producting queries and feature map without performing layer-wise refinement as proposed in deformable DETR [37]. Because each layer build masks from scratch, the masks predicted by a query in different layers may change dramatically. To show this inconsistent predictions among layers, we build two metrics mIoU-\\(L^i\\) and \\(Util^i\\) to quantify this problem and analyze its consequences in Section 3. As a result, this problem leads to a low utilization rate of decoder queries, which is especially severe in the first few decoder layers. As the attention masks predicted from an early layer are usually inaccurate, when serving as a guidance for its next layer, they will lead to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.851, + 0.207, + 0.864 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.079, + 0.864, + 0.469, + 0.887 + ], + "angle": 0, + "content": "This work was done when Hao Zhang and Feng Li were interns at IDEA." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.851, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.352, + 0.107 + ], + "angle": 0, + "content": "sub-optimal predictions in the next layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.471, + 0.257 + ], + "angle": 0, + "content": "Since detection is a similar task to segmentation, both aiming to locate instances in images, we seek inspirations from detection to improve segmentation. For example, the recently proposed denoising training [19] for detection has shown very effective to stabilize bipartite matching between training epochs. DINO [33] achieves a new SOTA on COCO detection built upon an improved denoising training method. The key idea in denoising training is to feed noised GT boxes in parallel with learnable queries into Transformer decoders and train the model to denoise and recover the GT boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.471, + 0.452 + ], + "angle": 0, + "content": "Inspired by DN-DETR [19], we propose a mask-piloted (MP) training approach. We divide the decoder queries into two parts, an MP part and a matching part. The matching part is the same as in Mask2Former, feeding learnable queries and adopting bipartite matching to assign predictions to GT instances. In the MP part, we feed class embeddings of GT categories as queries and GT masks as attention masks into a Transformer decoder layer and let the model to reconstruct GT masks and labels. To reinforce the effectiveness of our method, we propose to apply MP training for all decoder layers. Moreover, we also add point noises to GT masks and flipping noises to GT class embeddings to force the model to recover GT masks and labels from inaccurate ones." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.455, + 0.388, + 0.468 + ], + "angle": 0, + "content": "We summarize our contributions as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.476, + 0.472, + 0.613 + ], + "angle": 0, + "content": "1. We propose a mask-piloted training approach to improving masked-attention in Mask2Former, which suffers from inconsistent and inaccurate mask predictions among layers. We also develop techniques including multi-layer mask-piloted training with point noises and label-guided training to further improve the segmentation performance. Our method is computationally efficient, introducing no extra computation during inference and little computation during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.471, + 0.743 + ], + "angle": 0, + "content": "2. Through analyzing the predictions of Mask2Former, we discover a failure mode of Mask2Former—inconsistent predictions between consecutive layers. We build two novel metrics to measure this failure, which are layer-wise query utilization and mean IoU. We also show that our method improves the two metrics by a large margin, which validates its effectiveness on alleviating inconsistent predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.472, + 0.901 + ], + "angle": 0, + "content": "3. When evaluating on three datasets, including ADE20k, Cityscapes, and MS COCO, our MP-Former outperforms Mask2Former by a large margin. Besides improved performance, our method also speeds up training. Our model exceeds Mask2Former on all three segmentation tasks with only half of the number of training steps on ADE20k. Also, our model trained for 36 epochs is comparable with Mask2Former trained for 50 epochs on instance and panoptic segmentation on MS COCO." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.476, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.895, + 0.176 + ], + "angle": 0, + "content": "Generally, image segmentation can be divided into three tasks including instance segmentation, semantic segmentation, and panoptic segmentation with respect to different semantics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.177, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Traditional Segmentation models Traditionally, researchers develop specialized models and optimization objectives for each task. Instance segmentation is to predict a set of binary masks and their associated categories. Previous methods often predict masks based on bounding boxes produced by detection models. Mask R-CNN [15] builds segmentation upon detection model Faster R-CNN [27] by adding a mask branch in parallel with the detection branch. HTC [2] further proposes to interleave the two branches and add mask information flow to improve segmentation performance. Semantic segmentation focus on category-level semantics without distinguishing instances. Previous models widely formulate it into a per-pixel classification problem. The pioneering work, FCN [25] generates a label for each pixel to solve this problem. Many follow-up works continue with this idea and design more precise pixel-level classification models [3, 4]. Panoptic segmentation [18, 30] is a combination of the two segmentation tasks above to segment both the foreground instances (\"thing\") and background semantics (\"stuff\")." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.479, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Vision Transformers for segmentation Transformer-based segmentation models emerge with DETR (DEtection Transformer) [1]. Since that, Transformer [29] has been widely used in many detection [1, 19, 23, 33] and segmentation models [7, 8, 32]. Generally, these models adopt a set prediction objective with bipartite matching to query features of interest. Inspiring progress has been made in many specialized query-based architectures for instance segmentation [13], semantic segmentation [32, 35], panoptic segmentation [21]. Notably, some recent works unify the three segmentation tasks into one model [7, 8, 20, 34]. Among them, Mask2Former [7] proposes masked attention and achieves remarkable performance on all three segmentation tasks. Following DETR, Mask2Former is an end-to-end Transformer for image segmentation. It is composed of a backbone, a pixel decoder, and a Transformer decoder. The pixel decoder is an encoder-only deformable Transformer that takes multi-scale image features as input and output processed multi-scale features. The Transformer decoder takes learnable queries as input to attend to the feature maps output by the pixel decoder. Each decoder layer attends to a feature map of one scale following a coarse-to-fine manner. The predicted masks of one decoder layer will feed to the next layer as the attention masks, which enables the following decoder to focus on more meaningful regions. The design of attention masks eases the training and improves performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Ease Training for Vision transformers Vision transformers have an advantage over the convolution-based methods due" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18075" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.187, + 0.088, + 0.28, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.146, + 0.238, + 0.156 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.088, + 0.432, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.145, + 0.391, + 0.156 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.157, + 0.286, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.254, + 0.245, + 0.264 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.157, + 0.44, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.254, + 0.398, + 0.265 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.095, + 0.633, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.145, + 0.591, + 0.156 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.095, + 0.786, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.145, + 0.746, + 0.155 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.157, + 0.63, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.254, + 0.589, + 0.265 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.157, + 0.781, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.254, + 0.741, + 0.265 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.895, + 0.319 + ], + "angle": 0, + "content": "Figure 1. We visualize layer-wise predictions of Mask2Former and show 4 pairs of failure cases. Each pair is the predictions of the same query in adjacent decoder layers. The red regions are the predicted masks. These cases show that the predictions of a query may change dramatically between consecutive layers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.335, + 0.473, + 0.773 + ], + "angle": 0, + "content": "to their global attention—they can see all parts of the image. However, it is hard to train because globally searching for an object is a hard task. This phenomenon exists in both detection and segmentation. In detection, DETR suffers from slow convergence requiring 500 training epochs for convergence. Recently, researchers have dived into the meaning of the learnable queries [23, 26, 31, 37]. They either express the queries as reference points or anchor boxes. [19, 33] proposed to add noised ground truth boxes as positional queries for denoising training and they speed up detection greatly. In segmentation, Mask2Former proposed mask attention which makes training easier and speeds up convergence when compared with MaskFormer. We have found some common points for these works that speed up training for vision Transformers. Firstly, they usually give clearer meaning to learnable queries to reduce ambiguity. Secondly, they give local limitations to cross attention and reduce the scope for the model to search objects. Therefore, we believe that giving guidance to cross attention leads to good segmentation performance. Our method is similar to DN-DETR [19] in introducing GT annotations to the Transformer decoder but ours is totally different from it in both the problems we solve and the method we take. Firstly, our method aims to address the inconsistent prediction problem which leads to the low utilization of queries, while DN-DETR aims to solve unstable matching among epochs. Secondly, we use GT masks as the attention masks, while DN-DETR feeds noised GT boxes as decoder queries. Thirdly, noises are optional in our method, but DN-DETR does not work without noises." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.327, + 0.801 + ], + "angle": 0, + "content": "3. Failure Cases and Solution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Failure cases: We view the mask prediction in Mask2Former as a mask refinement process through Transformer decoder layers. By taking predictions from a layer as the attention masks of the next layer, we expected to obtain refined predictions from the next layer. Generally, the predicted masks are assigned to the GT masks by bipartite matching and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.896, + 0.547 + ], + "angle": 0, + "content": "refined in a coarse-to-fine procedure. Therefore, keeping refinement consistency in different layers is the key to refining the predicted masks consistently. However, when we visualize the predictions of Mask2Former layer by layer, we found a severe inconsistency problem among the predictions from the same query between consecutive decoder layers. As shown in Fig. 1, the predictions of adjacent layers differ significantly. We define a metric named layer-wise mean Intersection-over-Union (IoU) denoted as \\(\\mathrm{mIoU - L}^i\\) to quantify this inconsistency. We denote the predicted masks from the \\(i\\)-th decoder layer as \\(\\mathbf{M}^{\\mathbf{i}} = \\left\\{M_0^i,M_1^i,\\dots,M_{N - 1}^i\\right\\}\\), where \\(N\\) is the number of decoder queries. Note that \\(M^0\\) means the predictions before queries enter the first Transformer decoder layer. We have" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.557, + 0.893, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mathrm {m I o U - L} ^ {i} = \\sum_ {n = 0} ^ {N - 1} \\mathrm {I o U} \\left(M _ {n} ^ {i - 1}, M _ {n} ^ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.895, + 0.819 + ], + "angle": 0, + "content": "As shown in Table 1, the mIoU-L is quite low, especially for the first few layers, which indicates inconsistent mask prediction results between adjacent layers. Such an inconsistent result means a query may be assigned to different instances (or background for panoptic and semantic segmentation) during layer-by-layer updates. To evaluate the inconsistent matching problem, we define another metric named as layer-wise query utilization denoted as \\( \\text{Util}^i \\) for the \\( i \\)-th layer, which means the ratio of the matched queries in the \\( i \\)-th layer that matches exactly the same instance as in the last layer. We denote the ground truth objects as \\( \\mathbf{T} = \\{T_0, T_1, T_2, \\dots, T_{O-1}\\} \\) where \\( O \\) is the number of ground truth objects. After bipartite matching, we compute an index vector \\( \\mathbf{V}^i = \\{V_0^i, V_1^i, \\dots, V_{N-1}^i\\} \\) to store the matching result of layer \\( i \\) as follows." + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.829, + 0.893, + 0.862 + ], + "angle": 0, + "content": "\\[\nV _ {n} ^ {i} = \\left\\{ \\begin{array}{l l} o, & \\text {i f} M _ {n} ^ {i} \\text {m a t c h e s} T _ {o} \\\\ - 1, & \\text {i f} M _ {n} ^ {i} \\text {m a t c h e s n o t h i n g} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We define the query utilization of layer \\(i\\) as the proportion of GT masks that match with the same query in layer \\(i\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18076" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.257, + 0.091, + 0.716, + 0.205 + ], + "angle": 0, + "content": "
Layer Number123456789
Mask2FormermIoU-Li(%)23.050.359.569.770.169.978.676.376.0
Utili(%)38.165.570.873.578.581.183.283.2100
MP trainingmIoU-Li(%)51.583.286.696.296.093.897.697.996.3
Utili* (%)94.093.894.693.294.094.894.094.0100
Multi-layer MPmIoU-Li(%)51.387.888.897.597.596.299.098.898.2
Utili* (%)94.295.497.397.697.397.098.298.2100
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.218, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Table 1. The mIoU-L\\(^i\\) and Util\\(^i\\) of Mask2Former and the proposed methods. The mIoU-L\\(^i\\) for a query is taken as the IoU of the predicted masks in the \\((i-1)\\)-th and \\(i\\)-th decoder layers. The mIoU-L\\(^i\\) takes the average over all queries. Note that the index of decoder layers starts from 1, so \"prediction of the 0-th decoder\" means the prediction before entering Transformer decoder. The Util\\(^i\\) is the ratio of GT masks that match with the same query in \\(i\\)-th layer and the last layer. * means that We test Util\\(^i\\) with bipartite matching to assign predictions in MP part to GT masks. Note that when directly assigning predictions in MP part to corresponding GT masks with hard assignment, Util\\(^i\\) is always 100%." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.311, + 0.318, + 0.326 + ], + "angle": 0, + "content": "the last layer, which is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.335, + 0.471, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\operatorname {U t i l} ^ {i} = \\frac {1}{O} \\sum_ {j = 0} ^ {N} \\mathbb {1} \\left(V _ {n} ^ {i} = V _ {n} ^ {l} \\text {a n d} V _ {n} ^ {i} \\neq - 1\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.509 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is the indicator function. \\(\\mathbb{1}(x) = 1\\) if \\(x\\) is true and 0 otherwise. The query utilization of layer \\(i\\) for the whole data set is averaged over the utilization numbers for all images. Table 1 shows that the query utilizations of the first few layers are too low, which indicate many matched queries in early layers are ended up not being used in the last layer. The low mIoU-\\(L^i\\) and \\(\\mathrm{Util}^i\\) have potential consequences such as inconstant optimization goals and low training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Our observations based on these two metrics motivate us to develop a more effective training method to make the matching more consistent among layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.471, + 0.765 + ], + "angle": 0, + "content": "Solution: To address the inconsistent prediction problem, inspired by DN-DETR, we proposed a mask-piloted training method. We feed GT masks as attention masks and expect these queries with GT masks can focus on their corresponding GT masks and not be distracted by other instances. As shown in row 2 and 3, Table 1, both mIoU-\\(L^i\\) and \\(\\text{Util}^i\\) are improved by a large margin when adopting mask-piloted training. We also find that only applying mask-piloted (MP) training in the first layer is not enough. Fig 3(a)(b) shows that inconsistent prediction exists even when given GT masks. Therefore, we add a GT mask to all decoder layers to reinforce MP training. As shown in Table 1, multi-layer MP training further improves both mIoU-\\(L^i\\) and \\(\\text{Util}^i\\) especially in latter decoder layers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.778, + 0.205, + 0.793 + ], + "angle": 0, + "content": "4. MP-Former" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.188, + 0.817 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our method is a training method for improving Mask2Former. Our MP-Former only introduces minor changes in the Mask2Former pipeline during training as shown in the red-line part in Fig. 2. Same as Mask2Former, our model is composed of a backbone, a pixel decoder, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.448 + ], + "angle": 0, + "content": "a Transformer decoder. Our improvements are on the Transformer decoder. We divide decoder queries into two parts, a mask-piloted (MP) part and a matching part. The matching part is the same as Mask2Former. The MP part feeds GT masks and GT class embeddings as input and the predictions are directly assigned to the corresponding GT masks without bipartite matching. Our method includes three components which are multi-layer MP training, mask noises, and label-guided training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.454, + 0.724, + 0.471 + ], + "angle": 0, + "content": "4.2. Multi-layer MP training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.735 + ], + "angle": 0, + "content": "In order to address the inconsistent prediction problem, we propose a multi-layer mask-piloted (MP) training method. In the MP part, we feed an additional set of queries and masks into Transformer where the queries and masks are GT class embeddings and GT masks. On the loss side, we independently assign the outputs of the MP part and the matching part with the GT instances. The predictions of the MP part are directly assigned to their corresponding GT masks and those in the matching part are assigned with bipartite graph matching. The losses for both parts are the same, following the loss design in Mask2Former. We also propose to add GT masks in multiple layers and show that this can further improve the performance with experiments in Table 8. Since we follow Mask2Fomer to use feature maps of different resolutions in different decoder layers, we interpolate GT masks into different resolutions when applied to different layers." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.742, + 0.642, + 0.757 + ], + "angle": 0, + "content": "4.3. Noised masks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Besides feeding GT masks to alleviate the inconsistency problem, we further propose to use noised GT masks to make mask refinement more robust in each decoder. The intuition is simple, as Mask2former refines the mask from an inaccurate mask predicted from the previous decoder. GT masks without noise may be too easy for the task which prevents further refinement. Therefore, we feed noised masks to the decoder and train the model to reconstruct the original ones, which is similar to DN-DETR [19] for box denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18077" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.184, + 0.076, + 0.788, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.381, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Figure 2. The architecture of our method is the same as Mask2Former (the blue-shaded part), which consists of a backbone, a pixel decoder, and a Transformer decoder. The difference is that we feed extra queries and attention masks which are called the MP part to the Transformer decoder (red-line part in the figure). The MP part contains GT masks as attention masks and GT class embeddings as queries. We feed GT masks into the MP part of all decoder layers. We also add point noises to GT masks and flipping noises to class embeddings which can further improve the performance. Note that this architecture is just for training. In the inference time, the red-line part does not exist, and thus, our pipeline is exactly the same as Mask2Former." + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.472, + 0.287, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.608, + 0.225, + 0.62 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.472, + 0.42, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.608, + 0.359, + 0.619 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.472, + 0.553, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.608, + 0.492, + 0.619 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.472, + 0.687, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.608, + 0.626, + 0.619 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.472, + 0.821, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.608, + 0.759, + 0.619 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.632, + 0.892, + 0.686 + ], + "angle": 0, + "content": "Figure 3. (a) The red region is a given GT mask in first decoder layer which covers the man riding the black horse. (b) The predicted mask of the first decoder layer which only covers the head of the man. The predicted mask will be used as the attention mask of the second layer and may mislead the second layer. (c)(d)(e) are demonstrations of point, shift and scale noises. The blue regions are noised masks and the red regions are the GT masks." + }, + { + "type": "table", + "bbox": [ + 0.175, + 0.687, + 0.797, + 0.81 + ], + "angle": 0, + "content": "
methodbackbonePQpanoptic APth panmIoUpaninstancesemantic mIoU
APAPSAPMAPL
MaskFormer [8]R5034.7-------
Panoptic-DeepLab [6]SWideRNet [5]37.9*-50.0*-----
Mask2Former [7]R5039.726.546.126.410.428.943.147.2
MP-FormerR5040.8(+1.1)27.148.3(+2.2)28.0(+1.6)10.530.7(+1.8)44.648.1
MaskFormer [8]Swin-L†-------54.1
FaPN-MaskFormer [8, 17]Swin-L†-------55.2
Mask2Former [7]Swin-L†48.134.254.534.916.340.054.756.1
MP-FormerSwin-L†49.4(+1.3)35.256.035.316.340.755.556.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.819, + 0.894, + 0.861 + ], + "angle": 0, + "content": "Table 2. Image segmentation results on ADE20K va1. MP-Former improves the performance on all three segmentation tasks with R50 and Swin-L as backbone. All metrics are evaluated with single-scale inference. Backbones pre-trained on ImageNet-22K are marked with \\( {}^{ \\dagger } \\) . - means the results are not reported for the methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.899 + ], + "angle": 0, + "content": "training. We have tried three types of noise which are point noise, shift noise, and scale noise as shown in Fig. 3. For" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.893, + 0.899 + ], + "angle": 0, + "content": "point noise, we randomly sample some points on the mask and flip them from 1 to 0 or from 0 to 1. The number of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18078" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.214, + 0.089, + 0.758, + 0.179 + ], + "angle": 0, + "content": "
methodbackbonePQpanoptic APTh panmIoUpaninstancesemantic mIoU
APAP50
Panoptic-DeepLab [6]R5060.332.178.7---
Mask2Former [7]R5062.137.377.537.461.979.4
MP-FormerR5062.739.3(+2.0)79.6(+2.1)39.7(+2.3)65.4(+3.5)81.0(+1.6)
Mask2Former [7]Swin-L†66.643.682.943.771.483.3
MP-FormerSwin-L†67.545.8(+2.2)83.544.9(+1.2)72.483.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.188, + 0.894, + 0.217 + ], + "angle": 0, + "content": "Table 3. Image segmentation results on Cityscapes va1. We report results on all three segmentation tasks with R50 or Swin-L as backbone. All metrics are evaluated with single-scale inference. Backbones pre-trained on ImageNet-22K are marked with" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.473, + 0.478 + ], + "angle": 0, + "content": "sampled points is proportional to the area of a GT mask. For shift noise, we randomly shift a mask instance vertically and horizontally while keeping the center of the instance within its GT bounding box. For the scaling noise, we enlarge or shrink the instance by a random ratio. A comparison of three types of noises is shown in Table 8. Through experiments, we find point noises improve the performance while shifting and scaling noises do not work for our method. Therefore, we adopt point noises in our implementation. We use a noise ratio \\(\\lambda_{p}\\) to control how many noise points are added. Denote a binary GT mask as \\(M = [m_{i,j}]_{H\\times W}\\) where \\(m_{i,j} = 1\\) means the pixel belongs to the mask. The number of noise points is randomly sampled from 0 to \\(\\lambda_{p}\\times \\sum_{i,j}m_{i,j}\\). Since we apply mask-piloted training to all decoder layers, we independently add different noises to GT masks in different layers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.489, + 0.283, + 0.505 + ], + "angle": 0, + "content": "4.4. Label-guided training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.473, + 0.83 + ], + "angle": 0, + "content": "We use the class embeddings of GT categories as queries in the MP part. The reason why we use class embeddings is that queries will dot-product with image features and an intuitive way to distinguish instances is to use their categories. Note that we only use class embeddings in the first layer and do not replace predicted queries with class embeddings in subsequent layers. We also have a classification loss in the MP part. Since the classification task becomes very easy given GT class embeddings, we add flipping noises in class embeddings. We randomly sample GT class embeddings and let them randomly flip to the class embeddings of other categories. For example, \\( E_0, E_1, \\dots, E_K \\) denotes the class embeddings of \\( K \\) categories. An image contains \\( N \\) instances which belong to Categories \\( C^0, C^1, \\dots, C^N \\). So their default class embeddings should be \\( E_{C^0}, E_{C^1}, \\dots, E_{C^N} \\). To add flipping noise, we enumerate all \\( N \\) instances. For \\( i \\)-th instance, we generate a random number \\( r_i \\) in [0, 1] to decide whether to flip the class embedding \\( E_{C^i} \\) or not. If \\( r_i < \\lambda_l \\), we randomly sample another class embedding in \\( E_0, E_1, \\dots, E_K \\) to replace \\( E_{C^i} \\). \\( \\lambda_l \\) denotes the proportion of the sampled class embeddings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.21, + 0.862 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We conduct extensive experiments to compare our method with Mask2Former and other segmentation methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.894, + 0.296 + ], + "angle": 0, + "content": "on all three segmentation tasks on three datasets to show the effectiveness of our method. We also conduct several ablation studies to show the effectiveness of each component in our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.306, + 0.715, + 0.322 + ], + "angle": 0, + "content": "5.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.895, + 0.465 + ], + "angle": 0, + "content": "Model: Since our method is an improved training method based on Mask2Former, we adopt an identical model setting as Mask2Former except for the mask-piloted (MP) part. In the MP part, we adopt 100 GT masks with point noises. We adopt noise ratios \\(\\lambda_{p} = 0.2\\) and \\(\\lambda_{l} = 0.2\\) for point noises on masks and flipping noises on class embeddings. We adopt ResNet50 [16] and Swin-L [24] when evaluating our method on ADE20K and Cityscapes. On COCO2017, we use ResNet50 and ResNet101 as the backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.895, + 0.619 + ], + "angle": 0, + "content": "Dataset and metrics: We evaluate MP-Former on three challenging datasets: Cityscapes [9, 10] and ADE20K [36] for all three segmentation tasks and COCO 2017 [22] dataset for panoptic segmentation and instance segmentation. For instance segmentation, we evaluate the mask AP [22] on instances that are denoted as \"thing\" categories in datasets. For semantic segmentation, we evaluate mean Intersection-over-Union (mIOU) over all categories including both foreground and background. For panoptic segmentation, we evaluate the results with the panoptic quality (PQ) metric [18]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.63, + 0.64, + 0.643 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.734 + ], + "angle": 0, + "content": "We trained on three challenging datasets ADE20K, Cityscapes, and MS COCO with different backbones. Our method outperforms Mask2Former on all three segmentation tasks (instance, panoptic, and semantic) under the same setting, which demonstrates its effectiveness and generalization ability to different datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Performance on ADE20K. On ADE20K, we report our performance trained for \\(160K\\) steps following Mask2Former. As shown in Table 2, our method exceeds Mask2Former and other baselines on all metrics with both R50 and Swin-L. Notably, MP-Former achieves \\(+1.6\\mathrm{AP}\\) on instance segmentation with a R50 backbone and \\(+1.3\\mathrm{PQ}\\) on panoptic segmentation with a Swin-L backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Performance on Cityscapes. On Cityscapes, we train the model with 90K training steps following Mask2Former. As shown in Table 3, our method consistently outperforms Mask2Former on all three segmentation tasks across differ-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18079" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.163, + 0.088, + 0.805, + 0.263 + ], + "angle": 0, + "content": "
methodbackbonequery typeepochsAP\\( AP^S \\)\\( AP^M \\)\\( AP^L \\)#params.FLOPsfps
MaskFormer [8]R50100 queries30034.016.437.854.245M181G19.2
Mask R-CNN [15]R50dense anchors3637.218.639.553.344M201G15.2
Mask R-CNN [12, 14, 15]R50dense anchors40042.523.845.060.046M358G10.3
Mask2Former [7]R50100 queries5043.723.447.264.844M226G9.7
MP-FormerR50100 queries5044.8(+1.1)24.748.165.544M226G9.7
Mask R-CNN [15]R101dense anchors3638.619.541.355.363M266G10.8
Mask R-CNN [12, 14, 15]R101dense anchors40043.724.646.461.865M423G8.6
Mask2Former [7]R101100 queries5044.223.847.766.763M293G7.8
MP-FormerR101100 queries5045.1(+0.9)24.648.766.263M293G7.8
QueryInst [13]Swin-L†300 queries5048.930.852.668.3--3.3
Swin-HTC++ [2, 24]Swin-L†dense anchors7249.531.052.467.2284M1470G-
Mask2Former [7]Swin-L†200 queries10050.129.953.972.1216M868G4.0
MP-Former (ours)Swin-L†200 queries7250.8(+0.7)30.354.572.6216M868G4.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.27, + 0.895, + 0.313 + ], + "angle": 0, + "content": "Table 4. Instance segmentation on COCO va12017 with 80 categories. MP-Former outperforms strong baseline Mask2Former without extra computation cost in inference time. For a fair comparison, we only consider single-scale inference and models trained using only COCO train2017 set data. Backbones pre-trained on ImageNet-22K are marked with" + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.322, + 0.798, + 0.497 + ], + "angle": 0, + "content": "
methodbackbonequery typeepochsPQ\\( PQ^{Th} \\)\\( PQ^{St} \\)\\( AP^{Th}_{pan} \\)\\( mIoU_{pan} \\)#params.FLOPsfps
DETR [1]R50100 queries500+2543.448.236.331.1----
MaskFormer [8]R50100 queries30046.551.039.833.057.845M181G17.6
Mask2Former [7]R50100 queries5051.4*57.242.741.961.544M226G8.6
MP-Former (ours)R50100 queries5052.2(+0.8)58.342.942.561.644M226G8.6
DETR [1]R101100 queries500+2545.150.537.033.0----
MaskFormer [8]R101100 queries30047.652.540.334.159.364M248G14.0
Mask2Former [7]R101100 queries5052.658.543.742.662.463M293G7.2
MP-Former (ours)R101100 queries5052.9(+0.3)58.744.043.061.863M293G7.2
Max-DeepLab [30]Max-L128 queries21651.157.042.2--451M3692G-
MaskFormer [8]Swin-L†100 queries30052.758.544.040.164.8212M792G5.2
K-Net [34]Swin-L†100 queries3654.660.246.0-----
Mask2FormerSwin-L†200 queries10057.864.248.148.667.4216M868G4.0
MP-FormerSwin-L†200 queries7258.1(+0.3)64.448.448.967.6216M868G4.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.504, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Table 5. Panoptic segmentation on COCO panoptic va12017 with 133 categories. MP-Former outperforms Mask2Former on COCO panoptic segmentation with both R50 and R101 as backbone. Results that are implemented are marked with * ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.471, + 0.611 + ], + "angle": 0, + "content": "ent backbones. MP-Former also scales well when adopting Swin-L as the backbone. Especially, our method outperforms Mask2Former by \\(+2.3\\mathrm{AP}\\) on instance segmentation and \\(+1.6\\mathrm{mIoU}\\) on semantic segmentation with R50." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.472, + 0.822 + ], + "angle": 0, + "content": "Performance on COCO. On COCO val2017, we follow Mask2Former's setting to train for 50epochs. As shown in Table 4 and Table 5, MP-Former outperforms both Mask2Former and classical methods such as Mask R-CNN [15] on instance segmentation and panoptic segmentation. For instance segmentation, our method achieves +1.1AP with R50, +0.9AP with R101 and +0.7 with Swin-L [24]. Note that we only train MP-Former for 72 epochs with Swin-L. On panoptic segmentation, our method achieves +0.8PQ with R50 and +0.3PQ with R101. Notably, our method is only a pluggable training method that will not affect the model architecture or inference speed. In addition, it only introduces negligible parameters (class embeddings) during training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.253, + 0.849 + ], + "angle": 0, + "content": "5.3. Speed up training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Table 6, we show our MP-Former can significantly speed up training on ADE20k and COCO. We do not report the results on Cityscapes because this dataset is relatively" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.895, + 0.779 + ], + "angle": 0, + "content": "small and only needs 90k steps to converge in Mask2Former, which is not necessary for additional speedup. The results show that our method trained with 36 epochs achieved comparable results with Mask2Former trained with 50 epochs with both R50 and R101 as the backbone on both panoptic and instance segmentation. Moreover, our method trained with half number of training steps (80K steps) achieves comparable results with Mask2Former trained with 160K steps. Note that we use the same batch size as Mask2Former. We also report the clock time. The results show that our method reduces the actual training time by large. For a fair comparison, we use the same number of GPUs to run MP-Former and Mask2Former. Because there may exist differences in speed for different machines, we run our method and Mask2Former on the same machine to test the running time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.655, + 0.803 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.841 + ], + "angle": 0, + "content": "We conduct an ablation study on the COCO instance segmentation task with R50 as backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Effectiveness of each component: We show the effectiveness of each version of our method trained for 12 epochs and 50 epochs in Table 7, where we show the result of raw Mask2Former and the result after each version of our method" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18080" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.152, + 0.088, + 0.825, + 0.228 + ], + "angle": 0, + "content": "
datasetmethodbackboneepochs*PQpanopticAPAPSinstancesemantic
APthpanmIoUpanclocktime/hAPMAPLclocktime/hmIoUclocktime/h
ADE20kMask2FormerR50160k39.726.546.152.826.410.428.943.140.347.221.3
MP-FormerR5080k39.727.048.434.227.410.230.044.525.547.412.3
Mask2FormerSwin-L†160k48.134.254.563.234.916.340.054.751.456.139.8
MP-FormerSwin-L†80k48.534.756.145.735.215.839.655.626.756.921.5
COCOMask2FormerR505051.4*41.961.577.943.723.447.264.862.5--
MP-FormerR503651.541.861.659.544.124.147.465.350.8--
Mask2FormerR1015052.642.662.4113.144.223.847.766.786.0--
MP-FormerR1013652.542.662.084.444.223.247.966.671.5--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.238, + 0.895, + 0.309 + ], + "angle": 0, + "content": "Table 6. Image segmentation results on ADE20K va1. and COCO va12017 MP-Former speeds up training by large. On ADE20K, MP-Former trained for 80K steps can achieve comparable result with Mask2Former trained for 160K steps with both R50 and Swin-L as backbone. On COCO, MP-Former trained for 36 epochs can achieve comparable result with Mask2Former trained for 50 epochs with both R50 and R101 as backbone. * means that we adopt number of steps instead of epochs on ADE20k following Mask2Former. We report only single-scale inference results. Backbones pre-trained on ImageNet-22K are marked with †" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.339, + 0.406, + 0.405 + ], + "angle": 0, + "content": "
#Row12 ep (AP mask)50 ep (AP mask)
1. Mask2Former [7]38.743.7
2. Row1+MP training on 1st layer39.444.2
3. Row2+Label noises39.644.2
4. Row3+Point noises39.944.6
5. Ours (Row4+All-layer MP)40.244.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.41, + 0.376, + 0.476 + ], + "angle": 0, + "content": "Table 7. The effectiveness of each component of our method when trained for 12 epochs and 50 epochs. The table shows that each component of our model effectively improves the results. Label noise does not work when trained for 50 epochs. Especially, our method without noises can improve the results by \\(+0.9\\mathrm{AP}\\). With multi-layer MP training, our method achieves \\(+1.5\\mathrm{AP}\\) on COCO instance segmentation with R50 backbone trained for 12 epochs. \"ep\" is short for epochs." + }, + { + "type": "table", + "bbox": [ + 0.414, + 0.336, + 0.617, + 0.433 + ], + "angle": 0, + "content": "
Layers for MP trainingepAP mask
First layer1239.4
First three layers1239.6(+0.2)
All layers1240.2(+0.8)
Noise Type
No noise1239.6
Point noise1239.9
Shifting noise1239.6
Scaling noise1239.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.4, + 0.439, + 0.629, + 0.478 + ], + "angle": 0, + "content": "Table 8. Top 3 lines shows the effectiveness of multi-layer Mask-Piloted (MP) training. The bottom lines of the table shows that shifting and scaling noise do not work for our method. \"ep\" is short for epochs." + }, + { + "type": "table", + "bbox": [ + 0.642, + 0.336, + 0.886, + 0.405 + ], + "angle": 0, + "content": "
Feature mapepAP mask
Coarse to fine1240.2
All large1240.0
Assignment in Mask-Piloted part
Match1239.4
Hard assignment1239.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.633, + 0.41, + 0.893, + 0.448 + ], + "angle": 0, + "content": "Table 9. Some hyperparemeter tuning:1.coarse-to-fine manner works better for our method than all-large feature maps. 2.Hard assignment lead to better results than matching for MP part queries. \"ep\" is short for epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.594 + ], + "angle": 0, + "content": "is adopted. Without adding noises, our method achieves an improvement of \\(+0.9AP\\). With multi-layer MP-training, our method achieves \\(+1.5AP\\). We also find that label noise does not work when trained for 50 epochs. Note that except for the last line of the table, other lines of our method apply MP training on the first 3 layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Apply MP training on multiple layers: In Table 8, we show the effectiveness of applying GT masks on multiple layers. The results indicate that more layers of GT masks can lead to better performance. By adding noised GT masks in all layers, MP-Former achieves \\(+0.8\\mathrm{AP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.471, + 0.763 + ], + "angle": 0, + "content": "Add noises to GT masks: In the last four rows of Table 8, we show the effectiveness of adding different types of noises. Note that without any noise, MP-Former achieves 39.6AP which is corresponding to the 4th row in Table 7. Shifting noise and scaling noise do not work in our method. Point noise achieved \\(+0.3\\) extra growth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Some hyper-parameter tuning: Mask2Former adopts coarse-to-fine feature maps in decoder layers. We try to adopt the largest feature map (1/8 feature map) in all decoder layers. Table 9 shows that a coarse-to-fine manner is better than using the largest feature map in all layers for our MP-Former. We also try to use matching to assign predictions to GT instances in the MP part. The result shows that matching can also improve the result compared with raw Mask2Former but the hard assignment is better." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.502, + 0.62, + 0.518 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.854 + ], + "angle": 0, + "content": "In this paper, we have analyzed failure cases of Mask2Former and found a problem of inconsistent predictions between adjacent layers which potentially leads to inconstant optimization goals and low utilization of queries and therefore limits the performance of Mask2Former. In order to address the problem, we developed a new MP-Former, whose key components include multi-layer mask-piloted training with point noises and label-guided training. Our experimental results showed that the proposed training method effectively mitigates the inconsistent prediction problem. When evaluating on three challenging datasets (ADE20K, Citescapes, and MS COCO), our method achieved improvement by a large margin on all three image segmentation tasks (semantic, instance, and panoptic). In addition, our method only introduces little computation during training and no extra computation in inference time. Finally, our method can also significantly speed up training. Especially, when trained with half number of steps, our method exceeds Mask2Former on all three segmentation tasks on ADE20K with both R50 and Swin-L as the backbones." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Limitations: We evaluated MP-Former on Mask2Former, but our method may be useful to some other models such as MaskFormer and HTC. We will take it as our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18081" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.472, + 0.339 + ], + "angle": 0, + "content": "[3] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.472, + 0.394 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.472, + 0.437 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, Huiyu Wang, and Siyuan Qiao. Scaling wide residual networks for panoptic segmentation. arXiv preprint arXiv:2011.11675, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.472, + 0.521 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12475-12485, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.472, + 0.576 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.578, + 0.472, + 0.62 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.472, + 0.703 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Scharwächter, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset. In CVPR Workshop on The Future of Datasets in Vision, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.858, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[12] Xianzhi Du, Barret Zoph, Wei-Chih Hung, and Tsung-Yi Lin. Simple training strategies and model scaling for object detection. arXiv preprint arXiv:2107.00057, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[13] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.895, + 0.232 + ], + "angle": 0, + "content": "[14] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.895, + 0.287 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.894, + 0.346 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[17] Shihua Huang, Zhichao Lu, Ran Cheng, and Cheng He. Fapn: Feature-aligned pyramid network for dense image prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 864-873, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.894, + 0.46 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.894, + 0.516 + ], + "angle": 0, + "content": "[19] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. arXiv preprint arXiv:2203.01305, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.894, + 0.573 + ], + "angle": 0, + "content": "[20] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[21] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Tong Lu, and Ping Luo. Panoptic segformer. arXiv preprint arXiv:2109.03814, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.894, + 0.743 + ], + "angle": 0, + "content": "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[26] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Con" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.956 + ], + "angle": 0, + "content": "18082" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "ditional detr for fast training convergence. arXiv preprint arXiv:2108.06152, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.472, + 0.246 + ], + "angle": 0, + "content": "[28] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[30] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[31] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. arXiv preprint arXiv:2109.07107, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.472, + 0.485 + ], + "angle": 0, + "content": "[32] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.472, + 0.544 + ], + "angle": 0, + "content": "[33] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[34] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.472, + 0.671 + ], + "angle": 0, + "content": "[35] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.472, + 0.728 + ], + "angle": 0, + "content": "[36] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.472, + 0.797 + ], + "angle": 0, + "content": "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR 2021: The Ninth International Conference on Learning Representations, 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18083" + } + ] +] \ No newline at end of file diff --git a/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_origin.pdf b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c82d4d055d5f94a4446fce85fa7964abbaad204b --- /dev/null +++ b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/c436cdaa-0279-4efb-be23-f53eeb77c336_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bb9cbb2fec8f3010626d0d603a08e370b7ad79008935b912d0bfbdeece6e1db +size 7271143 diff --git a/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/full.md b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a11d8300cce34701a7cd8dad0e6380f1c40a65cf --- /dev/null +++ b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/full.md @@ -0,0 +1,280 @@ +# MP-Former: Mask-Piloted Transformer for Image Segmentation + +Hao Zhang $^{1,4*†}$ , Feng Li $^{1,4*†}$ , Huaizhe Xu $^{1,4}$ , Shijia Huang $^{2,4}$ , Shilong Liu $^{3,4}$ , Lionel M. Ni $^{5,1}$ , Lei Zhang $^{4‡}$ + +1The Hong Kong University of Science and Technology. + +2The Chinese University of Hong Kong. + +$^{3}$ Dept. of CST., BNRist Center, Institute for AI, Tsinghua University. + +4International Digital Economy Academy (IDEA). + +5The Hong Kong University of Science and Technology (Guangzhou). + +# Abstract + +We present a mask-piloted Transformer which improves masked-attention in Mask2Former for image segmentation. The improvement is based on our observation that Mask2Former suffers from inconsistent mask predictions between consecutive decoder layers, which leads to inconsistent optimization goals and low utilization of decoder queries. To address this problem, we propose a mask-piloted training approach, which additionally feeds noised ground-truth masks in masked-attention and trains the model to reconstruct the original ones. Compared with the predicted masks used in mask-attention, the ground-truth masks serve as a pilot and effectively alleviate the negative impact of inaccurate mask predictions in Mask2Former. Based on this technique, our MP-Former achieves a remarkable performance improvement on all three image segmentation tasks (instance, panoptic, and semantic), yielding $+2.3AP$ and $+1.6mIoU$ on the Cityscapes instance and semantic segmentation tasks with a ResNet-50 backbone. Our method also significantly speeds up the training, outperforming Mask2Former with half of the number of training epochs on ADE20K with both a ResNet-50 and a Swin-L backbones. Moreover, our method only introduces little computation during training and no extra computation during inference. Our code will be released at https://github.com/IDEA-Research/MP-Former. + +# 1. Introduction + +Image segmentation is a fundamental problem in computer vision which includes semantic, instance, and panoptic + +segmentation. Early works design specialized architectures for different tasks, such as mask prediction models [2, 15] for instance segmentation and per-pixel prediction models [25, 28] for semantic segmentation. Panoptic segmentation [18] is a unified task of instance and semantic segmentation. But universal models for panoptic segmentation such as Panoptic FPN [18] and K-net [34] are usually suboptimal on instance and semantic segmentation compared with specialized models. + +Recently, Vision Transformers [29] has achieved tremendous success in many computer vision tasks, such as image classification [11, 24] and object detection [1]. Inspired by DETR's set prediction mechanism, Mask2Former [7] proposes a unified framework and achieves a remarkable performance on all three segmentation tasks. Similar to DETR, it uses learnable queries in Transformer decoders to probe features and adopts bipartite matching to assign predictions to ground truth (GT) masks. Mask2Former also uses predicted masks in a decoder layer as attention masks for the next decoder layer. The attention masks serve as a guidance to help next layer's prediction and greatly ease training. + +Despite its great success, Mask2Former is still an initial attempt which is mainly based on a vanilla DETR model that has a sub-optimal performance compared to its later variants. For example, in each decoder layer, the predicted masks are built from scratch by dot-producting queries and feature map without performing layer-wise refinement as proposed in deformable DETR [37]. Because each layer build masks from scratch, the masks predicted by a query in different layers may change dramatically. To show this inconsistent predictions among layers, we build two metrics mIoU- $L^i$ and $Util^i$ to quantify this problem and analyze its consequences in Section 3. As a result, this problem leads to a low utilization rate of decoder queries, which is especially severe in the first few decoder layers. As the attention masks predicted from an early layer are usually inaccurate, when serving as a guidance for its next layer, they will lead to + +sub-optimal predictions in the next layer. + +Since detection is a similar task to segmentation, both aiming to locate instances in images, we seek inspirations from detection to improve segmentation. For example, the recently proposed denoising training [19] for detection has shown very effective to stabilize bipartite matching between training epochs. DINO [33] achieves a new SOTA on COCO detection built upon an improved denoising training method. The key idea in denoising training is to feed noised GT boxes in parallel with learnable queries into Transformer decoders and train the model to denoise and recover the GT boxes. + +Inspired by DN-DETR [19], we propose a mask-piloted (MP) training approach. We divide the decoder queries into two parts, an MP part and a matching part. The matching part is the same as in Mask2Former, feeding learnable queries and adopting bipartite matching to assign predictions to GT instances. In the MP part, we feed class embeddings of GT categories as queries and GT masks as attention masks into a Transformer decoder layer and let the model to reconstruct GT masks and labels. To reinforce the effectiveness of our method, we propose to apply MP training for all decoder layers. Moreover, we also add point noises to GT masks and flipping noises to GT class embeddings to force the model to recover GT masks and labels from inaccurate ones. + +We summarize our contributions as follows. + +1. We propose a mask-piloted training approach to improving masked-attention in Mask2Former, which suffers from inconsistent and inaccurate mask predictions among layers. We also develop techniques including multi-layer mask-piloted training with point noises and label-guided training to further improve the segmentation performance. Our method is computationally efficient, introducing no extra computation during inference and little computation during training. +2. Through analyzing the predictions of Mask2Former, we discover a failure mode of Mask2Former—inconsistent predictions between consecutive layers. We build two novel metrics to measure this failure, which are layer-wise query utilization and mean IoU. We also show that our method improves the two metrics by a large margin, which validates its effectiveness on alleviating inconsistent predictions. +3. When evaluating on three datasets, including ADE20k, Cityscapes, and MS COCO, our MP-Former outperforms Mask2Former by a large margin. Besides improved performance, our method also speeds up training. Our model exceeds Mask2Former on all three segmentation tasks with only half of the number of training steps on ADE20k. Also, our model trained for 36 epochs is comparable with Mask2Former trained for 50 epochs on instance and panoptic segmentation on MS COCO. + +# 2. Related Work + +Generally, image segmentation can be divided into three tasks including instance segmentation, semantic segmentation, and panoptic segmentation with respect to different semantics. + +Traditional Segmentation models Traditionally, researchers develop specialized models and optimization objectives for each task. Instance segmentation is to predict a set of binary masks and their associated categories. Previous methods often predict masks based on bounding boxes produced by detection models. Mask R-CNN [15] builds segmentation upon detection model Faster R-CNN [27] by adding a mask branch in parallel with the detection branch. HTC [2] further proposes to interleave the two branches and add mask information flow to improve segmentation performance. Semantic segmentation focus on category-level semantics without distinguishing instances. Previous models widely formulate it into a per-pixel classification problem. The pioneering work, FCN [25] generates a label for each pixel to solve this problem. Many follow-up works continue with this idea and design more precise pixel-level classification models [3, 4]. Panoptic segmentation [18, 30] is a combination of the two segmentation tasks above to segment both the foreground instances ("thing") and background semantics ("stuff"). + +Vision Transformers for segmentation Transformer-based segmentation models emerge with DETR (DEtection Transformer) [1]. Since that, Transformer [29] has been widely used in many detection [1, 19, 23, 33] and segmentation models [7, 8, 32]. Generally, these models adopt a set prediction objective with bipartite matching to query features of interest. Inspiring progress has been made in many specialized query-based architectures for instance segmentation [13], semantic segmentation [32, 35], panoptic segmentation [21]. Notably, some recent works unify the three segmentation tasks into one model [7, 8, 20, 34]. Among them, Mask2Former [7] proposes masked attention and achieves remarkable performance on all three segmentation tasks. Following DETR, Mask2Former is an end-to-end Transformer for image segmentation. It is composed of a backbone, a pixel decoder, and a Transformer decoder. The pixel decoder is an encoder-only deformable Transformer that takes multi-scale image features as input and output processed multi-scale features. The Transformer decoder takes learnable queries as input to attend to the feature maps output by the pixel decoder. Each decoder layer attends to a feature map of one scale following a coarse-to-fine manner. The predicted masks of one decoder layer will feed to the next layer as the attention masks, which enables the following decoder to focus on more meaningful regions. The design of attention masks eases the training and improves performance. + +Ease Training for Vision transformers Vision transformers have an advantage over the convolution-based methods due + +![](images/c3af7d9e6634f3437a752fb5eb51768471217a049a4fd52abbac09e33aaf446a.jpg) +(a) + +![](images/b7e32641f7349b6af0392d8ad474c1a5a56b598144954c2924fbba8166c17d52.jpg) +(b) + +![](images/948410ed82d867ea03ec35c812a6c12d76f33491182ce16394a9a9a00f2c91db.jpg) +(e) + +![](images/7ac3c9784191bf7037977df8cdc3e185d93692f0a9eafd4691aab062dc6c28f5.jpg) +(f) +Figure 1. We visualize layer-wise predictions of Mask2Former and show 4 pairs of failure cases. Each pair is the predictions of the same query in adjacent decoder layers. The red regions are the predicted masks. These cases show that the predictions of a query may change dramatically between consecutive layers. + +![](images/d8207ffc706fa8b6922a1cb7477a4af8a20f9935dde61c27e5d94037bff68c67.jpg) +(c) + +![](images/3337589aaf9d9945a6cce969bc392970f60349babb89913d006c85df7e630414.jpg) +(d) + +![](images/dc63c027b28d3a20e6f82405d8fd905b288b4665509e11f6ef82e5662a537e47.jpg) +(g) + +![](images/135dcd337bf1b9fc4bcf3b662327ef035cb7c9d9d302830e8037f05b808509cb.jpg) +(h) + +to their global attention—they can see all parts of the image. However, it is hard to train because globally searching for an object is a hard task. This phenomenon exists in both detection and segmentation. In detection, DETR suffers from slow convergence requiring 500 training epochs for convergence. Recently, researchers have dived into the meaning of the learnable queries [23, 26, 31, 37]. They either express the queries as reference points or anchor boxes. [19, 33] proposed to add noised ground truth boxes as positional queries for denoising training and they speed up detection greatly. In segmentation, Mask2Former proposed mask attention which makes training easier and speeds up convergence when compared with MaskFormer. We have found some common points for these works that speed up training for vision Transformers. Firstly, they usually give clearer meaning to learnable queries to reduce ambiguity. Secondly, they give local limitations to cross attention and reduce the scope for the model to search objects. Therefore, we believe that giving guidance to cross attention leads to good segmentation performance. Our method is similar to DN-DETR [19] in introducing GT annotations to the Transformer decoder but ours is totally different from it in both the problems we solve and the method we take. Firstly, our method aims to address the inconsistent prediction problem which leads to the low utilization of queries, while DN-DETR aims to solve unstable matching among epochs. Secondly, we use GT masks as the attention masks, while DN-DETR feeds noised GT boxes as decoder queries. Thirdly, noises are optional in our method, but DN-DETR does not work without noises. + +# 3. Failure Cases and Solution + +Failure cases: We view the mask prediction in Mask2Former as a mask refinement process through Transformer decoder layers. By taking predictions from a layer as the attention masks of the next layer, we expected to obtain refined predictions from the next layer. Generally, the predicted masks are assigned to the GT masks by bipartite matching and + +refined in a coarse-to-fine procedure. Therefore, keeping refinement consistency in different layers is the key to refining the predicted masks consistently. However, when we visualize the predictions of Mask2Former layer by layer, we found a severe inconsistency problem among the predictions from the same query between consecutive decoder layers. As shown in Fig. 1, the predictions of adjacent layers differ significantly. We define a metric named layer-wise mean Intersection-over-Union (IoU) denoted as $\mathrm{mIoU - L}^i$ to quantify this inconsistency. We denote the predicted masks from the $i$ -th decoder layer as $\mathbf{M}^{\mathbf{i}} = \left\{M_0^i,M_1^i,\dots,M_{N - 1}^i\right\}$ , where $N$ is the number of decoder queries. Note that $M^0$ means the predictions before queries enter the first Transformer decoder layer. We have + +$$ +\mathrm {m I o U - L} ^ {i} = \sum_ {n = 0} ^ {N - 1} \mathrm {I o U} \left(M _ {n} ^ {i - 1}, M _ {n} ^ {i}\right) \tag {1} +$$ + +As shown in Table 1, the mIoU-L is quite low, especially for the first few layers, which indicates inconsistent mask prediction results between adjacent layers. Such an inconsistent result means a query may be assigned to different instances (or background for panoptic and semantic segmentation) during layer-by-layer updates. To evaluate the inconsistent matching problem, we define another metric named as layer-wise query utilization denoted as $\text{Util}^i$ for the $i$ -th layer, which means the ratio of the matched queries in the $i$ -th layer that matches exactly the same instance as in the last layer. We denote the ground truth objects as $\mathbf{T} = \{T_0, T_1, T_2, \dots, T_{O-1}\}$ where $O$ is the number of ground truth objects. After bipartite matching, we compute an index vector $\mathbf{V}^i = \{V_0^i, V_1^i, \dots, V_{N-1}^i\}$ to store the matching result of layer $i$ as follows. + +$$ +V _ {n} ^ {i} = \left\{ \begin{array}{l l} o, & \text {i f} M _ {n} ^ {i} \text {m a t c h e s} T _ {o} \\ - 1, & \text {i f} M _ {n} ^ {i} \text {m a t c h e s n o t h i n g} \end{array} \right. \tag {2} +$$ + +We define the query utilization of layer $i$ as the proportion of GT masks that match with the same query in layer $i$ and + +
Layer Number123456789
Mask2FormermIoU-Li(%)23.050.359.569.770.169.978.676.376.0
Utili(%)38.165.570.873.578.581.183.283.2100
MP trainingmIoU-Li(%)51.583.286.696.296.093.897.697.996.3
Utili* (%)94.093.894.693.294.094.894.094.0100
Multi-layer MPmIoU-Li(%)51.387.888.897.597.596.299.098.898.2
Utili* (%)94.295.497.397.697.397.098.298.2100
+ +Table 1. The mIoU-L $^i$ and Util $^i$ of Mask2Former and the proposed methods. The mIoU-L $^i$ for a query is taken as the IoU of the predicted masks in the $(i-1)$ -th and $i$ -th decoder layers. The mIoU-L $^i$ takes the average over all queries. Note that the index of decoder layers starts from 1, so "prediction of the 0-th decoder" means the prediction before entering Transformer decoder. The Util $^i$ is the ratio of GT masks that match with the same query in $i$ -th layer and the last layer. * means that We test Util $^i$ with bipartite matching to assign predictions in MP part to GT masks. Note that when directly assigning predictions in MP part to corresponding GT masks with hard assignment, Util $^i$ is always 100%. + +the last layer, which is calculated as + +$$ +\operatorname {U t i l} ^ {i} = \frac {1}{O} \sum_ {j = 0} ^ {N} \mathbb {1} \left(V _ {n} ^ {i} = V _ {n} ^ {l} \text {a n d} V _ {n} ^ {i} \neq - 1\right), \tag {3} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function. $\mathbb{1}(x) = 1$ if $x$ is true and 0 otherwise. The query utilization of layer $i$ for the whole data set is averaged over the utilization numbers for all images. Table 1 shows that the query utilizations of the first few layers are too low, which indicate many matched queries in early layers are ended up not being used in the last layer. The low mIoU- $L^i$ and $\mathrm{Util}^i$ have potential consequences such as inconstant optimization goals and low training efficiency. + +Our observations based on these two metrics motivate us to develop a more effective training method to make the matching more consistent among layers. + +Solution: To address the inconsistent prediction problem, inspired by DN-DETR, we proposed a mask-piloted training method. We feed GT masks as attention masks and expect these queries with GT masks can focus on their corresponding GT masks and not be distracted by other instances. As shown in row 2 and 3, Table 1, both mIoU- $L^i$ and $\text{Util}^i$ are improved by a large margin when adopting mask-piloted training. We also find that only applying mask-piloted (MP) training in the first layer is not enough. Fig 3(a)(b) shows that inconsistent prediction exists even when given GT masks. Therefore, we add a GT mask to all decoder layers to reinforce MP training. As shown in Table 1, multi-layer MP training further improves both mIoU- $L^i$ and $\text{Util}^i$ especially in latter decoder layers. + +# 4. MP-Former + +# 4.1. Overview + +Our method is a training method for improving Mask2Former. Our MP-Former only introduces minor changes in the Mask2Former pipeline during training as shown in the red-line part in Fig. 2. Same as Mask2Former, our model is composed of a backbone, a pixel decoder, and + +a Transformer decoder. Our improvements are on the Transformer decoder. We divide decoder queries into two parts, a mask-piloted (MP) part and a matching part. The matching part is the same as Mask2Former. The MP part feeds GT masks and GT class embeddings as input and the predictions are directly assigned to the corresponding GT masks without bipartite matching. Our method includes three components which are multi-layer MP training, mask noises, and label-guided training. + +# 4.2. Multi-layer MP training + +In order to address the inconsistent prediction problem, we propose a multi-layer mask-piloted (MP) training method. In the MP part, we feed an additional set of queries and masks into Transformer where the queries and masks are GT class embeddings and GT masks. On the loss side, we independently assign the outputs of the MP part and the matching part with the GT instances. The predictions of the MP part are directly assigned to their corresponding GT masks and those in the matching part are assigned with bipartite graph matching. The losses for both parts are the same, following the loss design in Mask2Former. We also propose to add GT masks in multiple layers and show that this can further improve the performance with experiments in Table 8. Since we follow Mask2Fomer to use feature maps of different resolutions in different decoder layers, we interpolate GT masks into different resolutions when applied to different layers. + +# 4.3. Noised masks + +Besides feeding GT masks to alleviate the inconsistency problem, we further propose to use noised GT masks to make mask refinement more robust in each decoder. The intuition is simple, as Mask2former refines the mask from an inaccurate mask predicted from the previous decoder. GT masks without noise may be too easy for the task which prevents further refinement. Therefore, we feed noised masks to the decoder and train the model to reconstruct the original ones, which is similar to DN-DETR [19] for box denoising + +![](images/0ef37a30044fec179f93f0049192a0895496c5a34aca5ffc34c351966353ad34.jpg) +Figure 2. The architecture of our method is the same as Mask2Former (the blue-shaded part), which consists of a backbone, a pixel decoder, and a Transformer decoder. The difference is that we feed extra queries and attention masks which are called the MP part to the Transformer decoder (red-line part in the figure). The MP part contains GT masks as attention masks and GT class embeddings as queries. We feed GT masks into the MP part of all decoder layers. We also add point noises to GT masks and flipping noises to class embeddings which can further improve the performance. Note that this architecture is just for training. In the inference time, the red-line part does not exist, and thus, our pipeline is exactly the same as Mask2Former. + +![](images/ffb0406ad8ecb8c0d139e417bf7dc7a71ce5409276ee7822ba95b63aedbdb58d.jpg) +(a) + +![](images/d8dabe7dcf632b3457faf852f0307df0e9a4e540580e91d75365d2bc03571e91.jpg) +(b) +Figure 3. (a) The red region is a given GT mask in first decoder layer which covers the man riding the black horse. (b) The predicted mask of the first decoder layer which only covers the head of the man. The predicted mask will be used as the attention mask of the second layer and may mislead the second layer. (c)(d)(e) are demonstrations of point, shift and scale noises. The blue regions are noised masks and the red regions are the GT masks. + +![](images/021dd0923be2447f4d3feb4a6d2a45488a564226801cc11d865edf0411c1ce4a.jpg) +(c) + +![](images/6c15f3fec374963edf72c391848f3981debf8374fda855e754995b7580752d09.jpg) +(d) + +![](images/3102a2f869c631bf2502811a4bc99b99e4674d7779db3e4fc5de5b235b5fdeff.jpg) +(e) + +
methodbackbonePQpanoptic APth panmIoUpaninstancesemantic mIoU
APAPSAPMAPL
MaskFormer [8]R5034.7-------
Panoptic-DeepLab [6]SWideRNet [5]37.9*-50.0*-----
Mask2Former [7]R5039.726.546.126.410.428.943.147.2
MP-FormerR5040.8(+1.1)27.148.3(+2.2)28.0(+1.6)10.530.7(+1.8)44.648.1
MaskFormer [8]Swin-L†-------54.1
FaPN-MaskFormer [8, 17]Swin-L†-------55.2
Mask2Former [7]Swin-L†48.134.254.534.916.340.054.756.1
MP-FormerSwin-L†49.4(+1.3)35.256.035.316.340.755.556.9
+ +Table 2. Image segmentation results on ADE20K va1. MP-Former improves the performance on all three segmentation tasks with R50 and Swin-L as backbone. All metrics are evaluated with single-scale inference. Backbones pre-trained on ImageNet-22K are marked with ${}^{ \dagger }$ . - means the results are not reported for the methods. + +training. We have tried three types of noise which are point noise, shift noise, and scale noise as shown in Fig. 3. For + +point noise, we randomly sample some points on the mask and flip them from 1 to 0 or from 0 to 1. The number of + +
methodbackbonePQpanoptic APTh panmIoUpaninstancesemantic mIoU
APAP50
Panoptic-DeepLab [6]R5060.332.178.7---
Mask2Former [7]R5062.137.377.537.461.979.4
MP-FormerR5062.739.3(+2.0)79.6(+2.1)39.7(+2.3)65.4(+3.5)81.0(+1.6)
Mask2Former [7]Swin-L†66.643.682.943.771.483.3
MP-FormerSwin-L†67.545.8(+2.2)83.544.9(+1.2)72.483.9
+ +Table 3. Image segmentation results on Cityscapes va1. We report results on all three segmentation tasks with R50 or Swin-L as backbone. All metrics are evaluated with single-scale inference. Backbones pre-trained on ImageNet-22K are marked with + +sampled points is proportional to the area of a GT mask. For shift noise, we randomly shift a mask instance vertically and horizontally while keeping the center of the instance within its GT bounding box. For the scaling noise, we enlarge or shrink the instance by a random ratio. A comparison of three types of noises is shown in Table 8. Through experiments, we find point noises improve the performance while shifting and scaling noises do not work for our method. Therefore, we adopt point noises in our implementation. We use a noise ratio $\lambda_{p}$ to control how many noise points are added. Denote a binary GT mask as $M = [m_{i,j}]_{H\times W}$ where $m_{i,j} = 1$ means the pixel belongs to the mask. The number of noise points is randomly sampled from 0 to $\lambda_{p}\times \sum_{i,j}m_{i,j}$ . Since we apply mask-piloted training to all decoder layers, we independently add different noises to GT masks in different layers. + +# 4.4. Label-guided training + +We use the class embeddings of GT categories as queries in the MP part. The reason why we use class embeddings is that queries will dot-product with image features and an intuitive way to distinguish instances is to use their categories. Note that we only use class embeddings in the first layer and do not replace predicted queries with class embeddings in subsequent layers. We also have a classification loss in the MP part. Since the classification task becomes very easy given GT class embeddings, we add flipping noises in class embeddings. We randomly sample GT class embeddings and let them randomly flip to the class embeddings of other categories. For example, $E_0, E_1, \dots, E_K$ denotes the class embeddings of $K$ categories. An image contains $N$ instances which belong to Categories $C^0, C^1, \dots, C^N$ . So their default class embeddings should be $E_{C^0}, E_{C^1}, \dots, E_{C^N}$ . To add flipping noise, we enumerate all $N$ instances. For $i$ -th instance, we generate a random number $r_i$ in [0, 1] to decide whether to flip the class embedding $E_{C^i}$ or not. If $r_i < \lambda_l$ , we randomly sample another class embedding in $E_0, E_1, \dots, E_K$ to replace $E_{C^i}$ . $\lambda_l$ denotes the proportion of the sampled class embeddings. + +# 5. Experiments + +We conduct extensive experiments to compare our method with Mask2Former and other segmentation methods + +on all three segmentation tasks on three datasets to show the effectiveness of our method. We also conduct several ablation studies to show the effectiveness of each component in our method. + +# 5.1. Implementation details + +Model: Since our method is an improved training method based on Mask2Former, we adopt an identical model setting as Mask2Former except for the mask-piloted (MP) part. In the MP part, we adopt 100 GT masks with point noises. We adopt noise ratios $\lambda_{p} = 0.2$ and $\lambda_{l} = 0.2$ for point noises on masks and flipping noises on class embeddings. We adopt ResNet50 [16] and Swin-L [24] when evaluating our method on ADE20K and Cityscapes. On COCO2017, we use ResNet50 and ResNet101 as the backbone. + +Dataset and metrics: We evaluate MP-Former on three challenging datasets: Cityscapes [9, 10] and ADE20K [36] for all three segmentation tasks and COCO 2017 [22] dataset for panoptic segmentation and instance segmentation. For instance segmentation, we evaluate the mask AP [22] on instances that are denoted as "thing" categories in datasets. For semantic segmentation, we evaluate mean Intersection-over-Union (mIOU) over all categories including both foreground and background. For panoptic segmentation, we evaluate the results with the panoptic quality (PQ) metric [18]. + +# 5.2. Main Results + +We trained on three challenging datasets ADE20K, Cityscapes, and MS COCO with different backbones. Our method outperforms Mask2Former on all three segmentation tasks (instance, panoptic, and semantic) under the same setting, which demonstrates its effectiveness and generalization ability to different datasets. + +Performance on ADE20K. On ADE20K, we report our performance trained for $160K$ steps following Mask2Former. As shown in Table 2, our method exceeds Mask2Former and other baselines on all metrics with both R50 and Swin-L. Notably, MP-Former achieves $+1.6\mathrm{AP}$ on instance segmentation with a R50 backbone and $+1.3\mathrm{PQ}$ on panoptic segmentation with a Swin-L backbone. + +Performance on Cityscapes. On Cityscapes, we train the model with 90K training steps following Mask2Former. As shown in Table 3, our method consistently outperforms Mask2Former on all three segmentation tasks across differ- + +
methodbackbonequery typeepochsAP\( AP^S \)\( AP^M \)\( AP^L \)#params.FLOPsfps
MaskFormer [8]R50100 queries30034.016.437.854.245M181G19.2
Mask R-CNN [15]R50dense anchors3637.218.639.553.344M201G15.2
Mask R-CNN [12, 14, 15]R50dense anchors40042.523.845.060.046M358G10.3
Mask2Former [7]R50100 queries5043.723.447.264.844M226G9.7
MP-FormerR50100 queries5044.8(+1.1)24.748.165.544M226G9.7
Mask R-CNN [15]R101dense anchors3638.619.541.355.363M266G10.8
Mask R-CNN [12, 14, 15]R101dense anchors40043.724.646.461.865M423G8.6
Mask2Former [7]R101100 queries5044.223.847.766.763M293G7.8
MP-FormerR101100 queries5045.1(+0.9)24.648.766.263M293G7.8
QueryInst [13]Swin-L†300 queries5048.930.852.668.3--3.3
Swin-HTC++ [2, 24]Swin-L†dense anchors7249.531.052.467.2284M1470G-
Mask2Former [7]Swin-L†200 queries10050.129.953.972.1216M868G4.0
MP-Former (ours)Swin-L†200 queries7250.8(+0.7)30.354.572.6216M868G4.0
+ +Table 4. Instance segmentation on COCO va12017 with 80 categories. MP-Former outperforms strong baseline Mask2Former without extra computation cost in inference time. For a fair comparison, we only consider single-scale inference and models trained using only COCO train2017 set data. Backbones pre-trained on ImageNet-22K are marked with + +
methodbackbonequery typeepochsPQ\( PQ^{Th} \)\( PQ^{St} \)\( AP^{Th}_{pan} \)\( mIoU_{pan} \)#params.FLOPsfps
DETR [1]R50100 queries500+2543.448.236.331.1----
MaskFormer [8]R50100 queries30046.551.039.833.057.845M181G17.6
Mask2Former [7]R50100 queries5051.4*57.242.741.961.544M226G8.6
MP-Former (ours)R50100 queries5052.2(+0.8)58.342.942.561.644M226G8.6
DETR [1]R101100 queries500+2545.150.537.033.0----
MaskFormer [8]R101100 queries30047.652.540.334.159.364M248G14.0
Mask2Former [7]R101100 queries5052.658.543.742.662.463M293G7.2
MP-Former (ours)R101100 queries5052.9(+0.3)58.744.043.061.863M293G7.2
Max-DeepLab [30]Max-L128 queries21651.157.042.2--451M3692G-
MaskFormer [8]Swin-L†100 queries30052.758.544.040.164.8212M792G5.2
K-Net [34]Swin-L†100 queries3654.660.246.0-----
Mask2FormerSwin-L†200 queries10057.864.248.148.667.4216M868G4.0
MP-FormerSwin-L†200 queries7258.1(+0.3)64.448.448.967.6216M868G4.0
+ +Table 5. Panoptic segmentation on COCO panoptic va12017 with 133 categories. MP-Former outperforms Mask2Former on COCO panoptic segmentation with both R50 and R101 as backbone. Results that are implemented are marked with * . + +ent backbones. MP-Former also scales well when adopting Swin-L as the backbone. Especially, our method outperforms Mask2Former by $+2.3\mathrm{AP}$ on instance segmentation and $+1.6\mathrm{mIoU}$ on semantic segmentation with R50. + +Performance on COCO. On COCO val2017, we follow Mask2Former's setting to train for 50epochs. As shown in Table 4 and Table 5, MP-Former outperforms both Mask2Former and classical methods such as Mask R-CNN [15] on instance segmentation and panoptic segmentation. For instance segmentation, our method achieves +1.1AP with R50, +0.9AP with R101 and +0.7 with Swin-L [24]. Note that we only train MP-Former for 72 epochs with Swin-L. On panoptic segmentation, our method achieves +0.8PQ with R50 and +0.3PQ with R101. Notably, our method is only a pluggable training method that will not affect the model architecture or inference speed. In addition, it only introduces negligible parameters (class embeddings) during training. + +# 5.3. Speed up training + +In Table 6, we show our MP-Former can significantly speed up training on ADE20k and COCO. We do not report the results on Cityscapes because this dataset is relatively + +small and only needs 90k steps to converge in Mask2Former, which is not necessary for additional speedup. The results show that our method trained with 36 epochs achieved comparable results with Mask2Former trained with 50 epochs with both R50 and R101 as the backbone on both panoptic and instance segmentation. Moreover, our method trained with half number of training steps (80K steps) achieves comparable results with Mask2Former trained with 160K steps. Note that we use the same batch size as Mask2Former. We also report the clock time. The results show that our method reduces the actual training time by large. For a fair comparison, we use the same number of GPUs to run MP-Former and Mask2Former. Because there may exist differences in speed for different machines, we run our method and Mask2Former on the same machine to test the running time. + +# 5.4. Ablation Study + +We conduct an ablation study on the COCO instance segmentation task with R50 as backbone. + +Effectiveness of each component: We show the effectiveness of each version of our method trained for 12 epochs and 50 epochs in Table 7, where we show the result of raw Mask2Former and the result after each version of our method + +
datasetmethodbackboneepochs*PQpanopticAPAPSinstancesemantic
APthpanmIoUpanclocktime/hAPMAPLclocktime/hmIoUclocktime/h
ADE20kMask2FormerR50160k39.726.546.152.826.410.428.943.140.347.221.3
MP-FormerR5080k39.727.048.434.227.410.230.044.525.547.412.3
Mask2FormerSwin-L†160k48.134.254.563.234.916.340.054.751.456.139.8
MP-FormerSwin-L†80k48.534.756.145.735.215.839.655.626.756.921.5
COCOMask2FormerR505051.4*41.961.577.943.723.447.264.862.5--
MP-FormerR503651.541.861.659.544.124.147.465.350.8--
Mask2FormerR1015052.642.662.4113.144.223.847.766.786.0--
MP-FormerR1013652.542.662.084.444.223.247.966.671.5--
+ +Table 6. Image segmentation results on ADE20K va1. and COCO va12017 MP-Former speeds up training by large. On ADE20K, MP-Former trained for 80K steps can achieve comparable result with Mask2Former trained for 160K steps with both R50 and Swin-L as backbone. On COCO, MP-Former trained for 36 epochs can achieve comparable result with Mask2Former trained for 50 epochs with both R50 and R101 as backbone. * means that we adopt number of steps instead of epochs on ADE20k following Mask2Former. We report only single-scale inference results. Backbones pre-trained on ImageNet-22K are marked with † + +
#Row12 ep (AP mask)50 ep (AP mask)
1. Mask2Former [7]38.743.7
2. Row1+MP training on 1st layer39.444.2
3. Row2+Label noises39.644.2
4. Row3+Point noises39.944.6
5. Ours (Row4+All-layer MP)40.244.8
+ +Table 7. The effectiveness of each component of our method when trained for 12 epochs and 50 epochs. The table shows that each component of our model effectively improves the results. Label noise does not work when trained for 50 epochs. Especially, our method without noises can improve the results by $+0.9\mathrm{AP}$ . With multi-layer MP training, our method achieves $+1.5\mathrm{AP}$ on COCO instance segmentation with R50 backbone trained for 12 epochs. "ep" is short for epochs. + +
Layers for MP trainingepAP mask
First layer1239.4
First three layers1239.6(+0.2)
All layers1240.2(+0.8)
Noise Type
No noise1239.6
Point noise1239.9
Shifting noise1239.6
Scaling noise1239.4
+ +Table 8. Top 3 lines shows the effectiveness of multi-layer Mask-Piloted (MP) training. The bottom lines of the table shows that shifting and scaling noise do not work for our method. "ep" is short for epochs. + +
Feature mapepAP mask
Coarse to fine1240.2
All large1240.0
Assignment in Mask-Piloted part
Match1239.4
Hard assignment1239.6
+ +Table 9. Some hyperparemeter tuning:1.coarse-to-fine manner works better for our method than all-large feature maps. 2.Hard assignment lead to better results than matching for MP part queries. "ep" is short for epochs. + +is adopted. Without adding noises, our method achieves an improvement of $+0.9AP$ . With multi-layer MP-training, our method achieves $+1.5AP$ . We also find that label noise does not work when trained for 50 epochs. Note that except for the last line of the table, other lines of our method apply MP training on the first 3 layers. + +Apply MP training on multiple layers: In Table 8, we show the effectiveness of applying GT masks on multiple layers. The results indicate that more layers of GT masks can lead to better performance. By adding noised GT masks in all layers, MP-Former achieves $+0.8\mathrm{AP}$ . + +Add noises to GT masks: In the last four rows of Table 8, we show the effectiveness of adding different types of noises. Note that without any noise, MP-Former achieves 39.6AP which is corresponding to the 4th row in Table 7. Shifting noise and scaling noise do not work in our method. Point noise achieved $+0.3$ extra growth. + +Some hyper-parameter tuning: Mask2Former adopts coarse-to-fine feature maps in decoder layers. We try to adopt the largest feature map (1/8 feature map) in all decoder layers. Table 9 shows that a coarse-to-fine manner is better than using the largest feature map in all layers for our MP-Former. We also try to use matching to assign predictions to GT instances in the MP part. The result shows that matching can also improve the result compared with raw Mask2Former but the hard assignment is better. + +# 6. Conclusion + +In this paper, we have analyzed failure cases of Mask2Former and found a problem of inconsistent predictions between adjacent layers which potentially leads to inconstant optimization goals and low utilization of queries and therefore limits the performance of Mask2Former. In order to address the problem, we developed a new MP-Former, whose key components include multi-layer mask-piloted training with point noises and label-guided training. Our experimental results showed that the proposed training method effectively mitigates the inconsistent prediction problem. When evaluating on three challenging datasets (ADE20K, Citescapes, and MS COCO), our method achieved improvement by a large margin on all three image segmentation tasks (semantic, instance, and panoptic). In addition, our method only introduces little computation during training and no extra computation in inference time. Finally, our method can also significantly speed up training. Especially, when trained with half number of steps, our method exceeds Mask2Former on all three segmentation tasks on ADE20K with both R50 and Swin-L as the backbones. + +Limitations: We evaluated MP-Former on Mask2Former, but our method may be useful to some other models such as MaskFormer and HTC. We will take it as our future work. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 7 +[2] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 1, 2, 7 +[3] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2 +[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2 +[5] Liang-Chieh Chen, Huiyu Wang, and Siyuan Qiao. Scaling wide residual networks for panoptic segmentation. arXiv preprint arXiv:2011.11675, 2020. 5 +[6] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12475-12485, 2020. 5, 6 +[7] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 1, 2, 5, 6, 7, 8 +[8] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. 2021. 2, 5, 7 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6 +[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Scharwächter, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset. In CVPR Workshop on The Future of Datasets in Vision, 2015. 6 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[12] Xianzhi Du, Barret Zoph, Wei-Chih Hung, and Tsung-Yi Lin. Simple training strategies and model scaling for object detection. arXiv preprint arXiv:2107.00057, 2021. 7 + +[13] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 2, 7 +[14] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 7 +[15] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2, 7 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[17] Shihua Huang, Zhichao Lu, Ran Cheng, and Cheng He. Fapn: Feature-aligned pyramid network for dense image prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 864-873, 2021. 5 +[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 2, 6 +[19] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. arXiv preprint arXiv:2203.01305, 2022. 2, 3, 4 +[20] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 2 +[21] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Tong Lu, and Ping Luo. Panoptic segformer. arXiv preprint arXiv:2109.03814, 2021. 2 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6 +[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 6, 7 +[25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 2 +[26] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Con + +ditional detr for fast training convergence. arXiv preprint arXiv:2108.06152, 2021. 3 +[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2 +[28] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 1 +[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1, 2 +[30] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2, 7 +[31] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. arXiv preprint arXiv:2109.07107, 2021. 3 +[32] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 2 +[33] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2, 3 +[34] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 7 +[35] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2 +[36] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6 +[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR 2021: The Ninth International Conference on Learning Representations, 2021. 1, 3 \ No newline at end of file diff --git a/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/images.zip b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fe8231e46ac5fe797866b3479b427fddf14faf96 --- /dev/null +++ b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d45f8b7d559340561d0bc26b3422c00ed04c7ec4fc10b47291241b6839d2192b +size 580559 diff --git a/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/layout.json b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5570e7321da11fbea64baac71780b399fcb3b5e --- /dev/null +++ b/2023/MP-Former_ Mask-Piloted Transformer for Image Segmentation/layout.json @@ -0,0 +1,7897 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 120 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 120 + ], + "type": "text", + "content": "MP-Former: Mask-Piloted Transformer for Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "spans": [ + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "text", + "content": "Hao Zhang" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "inline_equation", + "content": "^{1,4*†}" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "text", + "content": ", Feng Li" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "inline_equation", + "content": "^{1,4*†}" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "text", + "content": ", Huaizhe Xu" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "text", + "content": ", Shijia Huang" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "text", + "content": ", Shilong Liu" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "text", + "content": ", Lionel M. Ni" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "inline_equation", + "content": "^{5,1}" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "text", + "content": ", Lei Zhang" + }, + { + "bbox": [ + 136, + 141, + 455, + 171 + ], + "type": "inline_equation", + "content": "^{4‡}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 172, + 432, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 172, + 432, + 185 + ], + "spans": [ + { + "bbox": [ + 161, + 172, + 432, + 185 + ], + "type": "text", + "content": "1The Hong Kong University of Science and Technology." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 185, + 394, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 185, + 394, + 199 + ], + "spans": [ + { + "bbox": [ + 200, + 185, + 394, + 199 + ], + "type": "text", + "content": "2The Chinese University of Hong Kong." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 199, + 462, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 199, + 462, + 213 + ], + "spans": [ + { + "bbox": [ + 129, + 199, + 462, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 129, + 199, + 462, + 213 + ], + "type": "text", + "content": "Dept. of CST., BNRist Center, Institute for AI, Tsinghua University." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 175, + 213, + 418, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 213, + 418, + 227 + ], + "spans": [ + { + "bbox": [ + 175, + 213, + 418, + 227 + ], + "type": "text", + "content": "4International Digital Economy Academy (IDEA)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 227, + 465, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 227, + 465, + 241 + ], + "spans": [ + { + "bbox": [ + 127, + 227, + 465, + 241 + ], + "type": "text", + "content": "5The Hong Kong University of Science and Technology (Guangzhou)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 269, + 191, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 191, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 191, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 294, + 290, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 290, + 594 + ], + "type": "text", + "content": "We present a mask-piloted Transformer which improves masked-attention in Mask2Former for image segmentation. The improvement is based on our observation that Mask2Former suffers from inconsistent mask predictions between consecutive decoder layers, which leads to inconsistent optimization goals and low utilization of decoder queries. To address this problem, we propose a mask-piloted training approach, which additionally feeds noised ground-truth masks in masked-attention and trains the model to reconstruct the original ones. Compared with the predicted masks used in mask-attention, the ground-truth masks serve as a pilot and effectively alleviate the negative impact of inaccurate mask predictions in Mask2Former. Based on this technique, our MP-Former achieves a remarkable performance improvement on all three image segmentation tasks (instance, panoptic, and semantic), yielding " + }, + { + "bbox": [ + 47, + 294, + 290, + 594 + ], + "type": "inline_equation", + "content": "+2.3AP" + }, + { + "bbox": [ + 47, + 294, + 290, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 294, + 290, + 594 + ], + "type": "inline_equation", + "content": "+1.6mIoU" + }, + { + "bbox": [ + 47, + 294, + 290, + 594 + ], + "type": "text", + "content": " on the Cityscapes instance and semantic segmentation tasks with a ResNet-50 backbone. Our method also significantly speeds up the training, outperforming Mask2Former with half of the number of training epochs on ADE20K with both a ResNet-50 and a Swin-L backbones. Moreover, our method only introduces little computation during training and no extra computation during inference. Our code will be released at https://github.com/IDEA-Research/MP-Former." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 619, + 127, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 127, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 127, + 631 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 639, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 288, + 664 + ], + "type": "text", + "content": "Image segmentation is a fundamental problem in computer vision which includes semantic, instance, and panoptic" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 270, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 547, + 376 + ], + "type": "text", + "content": "segmentation. Early works design specialized architectures for different tasks, such as mask prediction models [2, 15] for instance segmentation and per-pixel prediction models [25, 28] for semantic segmentation. Panoptic segmentation [18] is a unified task of instance and semantic segmentation. But universal models for panoptic segmentation such as Panoptic FPN [18] and K-net [34] are usually suboptimal on instance and semantic segmentation compared with specialized models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 378, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 521 + ], + "type": "text", + "content": "Recently, Vision Transformers [29] has achieved tremendous success in many computer vision tasks, such as image classification [11, 24] and object detection [1]. Inspired by DETR's set prediction mechanism, Mask2Former [7] proposes a unified framework and achieves a remarkable performance on all three segmentation tasks. Similar to DETR, it uses learnable queries in Transformer decoders to probe features and adopts bipartite matching to assign predictions to ground truth (GT) masks. Mask2Former also uses predicted masks in a decoder layer as attention masks for the next decoder layer. The attention masks serve as a guidance to help next layer's prediction and greatly ease training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": "Despite its great success, Mask2Former is still an initial attempt which is mainly based on a vanilla DETR model that has a sub-optimal performance compared to its later variants. For example, in each decoder layer, the predicted masks are built from scratch by dot-producting queries and feature map without performing layer-wise refinement as proposed in deformable DETR [37]. Because each layer build masks from scratch, the masks predicted by a query in different layers may change dramatically. To show this inconsistent predictions among layers, we build two metrics mIoU-" + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "inline_equation", + "content": "L^i" + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "inline_equation", + "content": "Util^i" + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": " to quantify this problem and analyze its consequences in Section 3. As a result, this problem leads to a low utilization rate of decoder queries, which is especially severe in the first few decoder layers. As the attention masks predicted from an early layer are usually inaccurate, when serving as a guidance for its next layer, they will lead to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 673, + 126, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 673, + 126, + 684 + ], + "spans": [ + { + "bbox": [ + 58, + 673, + 126, + 684 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 684, + 287, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 684, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 48, + 684, + 287, + 702 + ], + "type": "text", + "content": "This work was done when Hao Zhang and Feng Li were interns at IDEA." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18074" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 215, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 215, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 215, + 84 + ], + "type": "text", + "content": "sub-optimal predictions in the next layer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 288, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 288, + 203 + ], + "type": "text", + "content": "Since detection is a similar task to segmentation, both aiming to locate instances in images, we seek inspirations from detection to improve segmentation. For example, the recently proposed denoising training [19] for detection has shown very effective to stabilize bipartite matching between training epochs. DINO [33] achieves a new SOTA on COCO detection built upon an improved denoising training method. The key idea in denoising training is to feed noised GT boxes in parallel with learnable queries into Transformer decoders and train the model to denoise and recover the GT boxes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 288, + 357 + ], + "type": "text", + "content": "Inspired by DN-DETR [19], we propose a mask-piloted (MP) training approach. We divide the decoder queries into two parts, an MP part and a matching part. The matching part is the same as in Mask2Former, feeding learnable queries and adopting bipartite matching to assign predictions to GT instances. In the MP part, we feed class embeddings of GT categories as queries and GT masks as attention masks into a Transformer decoder layer and let the model to reconstruct GT masks and labels. To reinforce the effectiveness of our method, we propose to apply MP training for all decoder layers. Moreover, we also add point noises to GT masks and flipping noises to GT class embeddings to force the model to recover GT masks and labels from inaccurate ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 360, + 237, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 360, + 237, + 370 + ], + "spans": [ + { + "bbox": [ + 59, + 360, + 237, + 370 + ], + "type": "text", + "content": "We summarize our contributions as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 376, + 288, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 288, + 485 + ], + "type": "text", + "content": "1. We propose a mask-piloted training approach to improving masked-attention in Mask2Former, which suffers from inconsistent and inaccurate mask predictions among layers. We also develop techniques including multi-layer mask-piloted training with point noises and label-guided training to further improve the segmentation performance. Our method is computationally efficient, introducing no extra computation during inference and little computation during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 288, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 288, + 588 + ], + "type": "text", + "content": "2. Through analyzing the predictions of Mask2Former, we discover a failure mode of Mask2Former—inconsistent predictions between consecutive layers. We build two novel metrics to measure this failure, which are layer-wise query utilization and mean IoU. We also show that our method improves the two metrics by a large margin, which validates its effectiveness on alleviating inconsistent predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 288, + 713 + ], + "type": "text", + "content": "3. When evaluating on three datasets, including ADE20k, Cityscapes, and MS COCO, our MP-Former outperforms Mask2Former by a large margin. Besides improved performance, our method also speeds up training. Our model exceeds Mask2Former on all three segmentation tasks with only half of the number of training steps on ADE20k. Also, our model trained for 36 epochs is comparable with Mask2Former trained for 50 epochs on instance and panoptic segmentation on MS COCO." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 91, + 547, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 547, + 139 + ], + "type": "text", + "content": "Generally, image segmentation can be divided into three tasks including instance segmentation, semantic segmentation, and panoptic segmentation with respect to different semantics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 140, + 547, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 547, + 378 + ], + "type": "text", + "content": "Traditional Segmentation models Traditionally, researchers develop specialized models and optimization objectives for each task. Instance segmentation is to predict a set of binary masks and their associated categories. Previous methods often predict masks based on bounding boxes produced by detection models. Mask R-CNN [15] builds segmentation upon detection model Faster R-CNN [27] by adding a mask branch in parallel with the detection branch. HTC [2] further proposes to interleave the two branches and add mask information flow to improve segmentation performance. Semantic segmentation focus on category-level semantics without distinguishing instances. Previous models widely formulate it into a per-pixel classification problem. The pioneering work, FCN [25] generates a label for each pixel to solve this problem. Many follow-up works continue with this idea and design more precise pixel-level classification models [3, 4]. Panoptic segmentation [18, 30] is a combination of the two segmentation tasks above to segment both the foreground instances (\"thing\") and background semantics (\"stuff\")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 379, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 547, + 689 + ], + "type": "text", + "content": "Vision Transformers for segmentation Transformer-based segmentation models emerge with DETR (DEtection Transformer) [1]. Since that, Transformer [29] has been widely used in many detection [1, 19, 23, 33] and segmentation models [7, 8, 32]. Generally, these models adopt a set prediction objective with bipartite matching to query features of interest. Inspiring progress has been made in many specialized query-based architectures for instance segmentation [13], semantic segmentation [32, 35], panoptic segmentation [21]. Notably, some recent works unify the three segmentation tasks into one model [7, 8, 20, 34]. Among them, Mask2Former [7] proposes masked attention and achieves remarkable performance on all three segmentation tasks. Following DETR, Mask2Former is an end-to-end Transformer for image segmentation. It is composed of a backbone, a pixel decoder, and a Transformer decoder. The pixel decoder is an encoder-only deformable Transformer that takes multi-scale image features as input and output processed multi-scale features. The Transformer decoder takes learnable queries as input to attend to the feature maps output by the pixel decoder. Each decoder layer attends to a feature map of one scale following a coarse-to-fine manner. The predicted masks of one decoder layer will feed to the next layer as the attention masks, which enables the following decoder to focus on more meaningful regions. The design of attention masks eases the training and improves performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "Ease Training for Vision transformers Vision transformers have an advantage over the convolution-based methods due" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18075" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 69, + 171, + 113 + ], + "blocks": [ + { + "bbox": [ + 114, + 69, + 171, + 113 + ], + "lines": [ + { + "bbox": [ + 114, + 69, + 171, + 113 + ], + "spans": [ + { + "bbox": [ + 114, + 69, + 171, + 113 + ], + "type": "image", + "image_path": "c3af7d9e6634f3437a752fb5eb51768471217a049a4fd52abbac09e33aaf446a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 115, + 145, + 123 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 145, + 123 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 145, + 123 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 208, + 69, + 264, + 113 + ], + "blocks": [ + { + "bbox": [ + 208, + 69, + 264, + 113 + ], + "lines": [ + { + "bbox": [ + 208, + 69, + 264, + 113 + ], + "spans": [ + { + "bbox": [ + 208, + 69, + 264, + 113 + ], + "type": "image", + "image_path": "b7e32641f7349b6af0392d8ad474c1a5a56b598144954c2924fbba8166c17d52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 114, + 239, + 123 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 239, + 123 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 239, + 123 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 118, + 124, + 175, + 198 + ], + "blocks": [ + { + "bbox": [ + 118, + 124, + 175, + 198 + ], + "lines": [ + { + "bbox": [ + 118, + 124, + 175, + 198 + ], + "spans": [ + { + "bbox": [ + 118, + 124, + 175, + 198 + ], + "type": "image", + "image_path": "948410ed82d867ea03ec35c812a6c12d76f33491182ce16394a9a9a00f2c91db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 201, + 149, + 209 + ], + "lines": [ + { + "bbox": [ + 138, + 201, + 149, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 149, + 209 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 211, + 124, + 269, + 198 + ], + "blocks": [ + { + "bbox": [ + 211, + 124, + 269, + 198 + ], + "lines": [ + { + "bbox": [ + 211, + 124, + 269, + 198 + ], + "spans": [ + { + "bbox": [ + 211, + 124, + 269, + 198 + ], + "type": "image", + "image_path": "7ac3c9784191bf7037977df8cdc3e185d93692f0a9eafd4691aab062dc6c28f5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 201, + 243, + 209 + ], + "lines": [ + { + "bbox": [ + 233, + 201, + 243, + 209 + ], + "spans": [ + { + "bbox": [ + 233, + 201, + 243, + 209 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 219, + 547, + 252 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 547, + 252 + ], + "type": "text", + "content": "Figure 1. We visualize layer-wise predictions of Mask2Former and show 4 pairs of failure cases. Each pair is the predictions of the same query in adjacent decoder layers. The red regions are the predicted masks. These cases show that the predictions of a query may change dramatically between consecutive layers." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 75, + 387, + 113 + ], + "blocks": [ + { + "bbox": [ + 331, + 75, + 387, + 113 + ], + "lines": [ + { + "bbox": [ + 331, + 75, + 387, + 113 + ], + "spans": [ + { + "bbox": [ + 331, + 75, + 387, + 113 + ], + "type": "image", + "image_path": "d8207ffc706fa8b6922a1cb7477a4af8a20f9935dde61c27e5d94037bff68c67.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 114, + 361, + 123 + ], + "lines": [ + { + "bbox": [ + 351, + 114, + 361, + 123 + ], + "spans": [ + { + "bbox": [ + 351, + 114, + 361, + 123 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 75, + 481, + 113 + ], + "blocks": [ + { + "bbox": [ + 425, + 75, + 481, + 113 + ], + "lines": [ + { + "bbox": [ + 425, + 75, + 481, + 113 + ], + "spans": [ + { + "bbox": [ + 425, + 75, + 481, + 113 + ], + "type": "image", + "image_path": "3337589aaf9d9945a6cce969bc392970f60349babb89913d006c85df7e630414.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 114, + 456, + 122 + ], + "lines": [ + { + "bbox": [ + 446, + 114, + 456, + 122 + ], + "spans": [ + { + "bbox": [ + 446, + 114, + 456, + 122 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 329, + 124, + 385, + 198 + ], + "blocks": [ + { + "bbox": [ + 329, + 124, + 385, + 198 + ], + "lines": [ + { + "bbox": [ + 329, + 124, + 385, + 198 + ], + "spans": [ + { + "bbox": [ + 329, + 124, + 385, + 198 + ], + "type": "image", + "image_path": "dc63c027b28d3a20e6f82405d8fd905b288b4665509e11f6ef82e5662a537e47.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 201, + 360, + 209 + ], + "lines": [ + { + "bbox": [ + 349, + 201, + 360, + 209 + ], + "spans": [ + { + "bbox": [ + 349, + 201, + 360, + 209 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 421, + 124, + 477, + 198 + ], + "blocks": [ + { + "bbox": [ + 421, + 124, + 477, + 198 + ], + "lines": [ + { + "bbox": [ + 421, + 124, + 477, + 198 + ], + "spans": [ + { + "bbox": [ + 421, + 124, + 477, + 198 + ], + "type": "image", + "image_path": "135dcd337bf1b9fc4bcf3b662327ef035cb7c9d9d302830e8037f05b808509cb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 201, + 453, + 209 + ], + "lines": [ + { + "bbox": [ + 443, + 201, + 453, + 209 + ], + "spans": [ + { + "bbox": [ + 443, + 201, + 453, + 209 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 265, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 289, + 612 + ], + "type": "text", + "content": "to their global attention—they can see all parts of the image. However, it is hard to train because globally searching for an object is a hard task. This phenomenon exists in both detection and segmentation. In detection, DETR suffers from slow convergence requiring 500 training epochs for convergence. Recently, researchers have dived into the meaning of the learnable queries [23, 26, 31, 37]. They either express the queries as reference points or anchor boxes. [19, 33] proposed to add noised ground truth boxes as positional queries for denoising training and they speed up detection greatly. In segmentation, Mask2Former proposed mask attention which makes training easier and speeds up convergence when compared with MaskFormer. We have found some common points for these works that speed up training for vision Transformers. Firstly, they usually give clearer meaning to learnable queries to reduce ambiguity. Secondly, they give local limitations to cross attention and reduce the scope for the model to search objects. Therefore, we believe that giving guidance to cross attention leads to good segmentation performance. Our method is similar to DN-DETR [19] in introducing GT annotations to the Transformer decoder but ours is totally different from it in both the problems we solve and the method we take. Firstly, our method aims to address the inconsistent prediction problem which leads to the low utilization of queries, while DN-DETR aims to solve unstable matching among epochs. Secondly, we use GT masks as the attention masks, while DN-DETR feeds noised GT boxes as decoder queries. Thirdly, noises are optional in our method, but DN-DETR does not work without noises." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 621, + 200, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 200, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 200, + 634 + ], + "type": "text", + "content": "3. Failure Cases and Solution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Failure cases: We view the mask prediction in Mask2Former as a mask refinement process through Transformer decoder layers. By taking predictions from a layer as the attention masks of the next layer, we expected to obtain refined predictions from the next layer. Generally, the predicted masks are assigned to the GT masks by bipartite matching and" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "text", + "content": "refined in a coarse-to-fine procedure. Therefore, keeping refinement consistency in different layers is the key to refining the predicted masks consistently. However, when we visualize the predictions of Mask2Former layer by layer, we found a severe inconsistency problem among the predictions from the same query between consecutive decoder layers. As shown in Fig. 1, the predictions of adjacent layers differ significantly. We define a metric named layer-wise mean Intersection-over-Union (IoU) denoted as " + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU - L}^i" + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "text", + "content": " to quantify this inconsistency. We denote the predicted masks from the " + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "text", + "content": "-th decoder layer as " + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathbf{i}} = \\left\\{M_0^i,M_1^i,\\dots,M_{N - 1}^i\\right\\}" + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "text", + "content": " is the number of decoder queries. Note that " + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "inline_equation", + "content": "M^0" + }, + { + "bbox": [ + 304, + 265, + 548, + 433 + ], + "type": "text", + "content": " means the predictions before queries enter the first Transformer decoder layer. We have" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 356, + 441, + 546, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 441, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 356, + 441, + 546, + 473 + ], + "type": "interline_equation", + "content": "\\mathrm {m I o U - L} ^ {i} = \\sum_ {n = 0} ^ {N - 1} \\mathrm {I o U} \\left(M _ {n} ^ {i - 1}, M _ {n} ^ {i}\\right) \\tag {1}", + "image_path": "c7e667da8fc83a975d9f40948a385e5e7a8ee56efcb78de2528a1af5db51e153.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "content": "As shown in Table 1, the mIoU-L is quite low, especially for the first few layers, which indicates inconsistent mask prediction results between adjacent layers. Such an inconsistent result means a query may be assigned to different instances (or background for panoptic and semantic segmentation) during layer-by-layer updates. To evaluate the inconsistent matching problem, we define another metric named as layer-wise query utilization denoted as " + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "inline_equation", + "content": "\\text{Util}^i" + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "content": "-th layer, which means the ratio of the matched queries in the " + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "content": "-th layer that matches exactly the same instance as in the last layer. We denote the ground truth objects as " + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = \\{T_0, T_1, T_2, \\dots, T_{O-1}\\}" + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "content": " is the number of ground truth objects. After bipartite matching, we compute an index vector " + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i = \\{V_0^i, V_1^i, \\dots, V_{N-1}^i\\}" + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "content": " to store the matching result of layer " + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 480, + 547, + 648 + ], + "type": "text", + "content": " as follows." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 343, + 656, + 546, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 656, + 546, + 682 + ], + "spans": [ + { + "bbox": [ + 343, + 656, + 546, + 682 + ], + "type": "interline_equation", + "content": "V _ {n} ^ {i} = \\left\\{ \\begin{array}{l l} o, & \\text {i f} M _ {n} ^ {i} \\text {m a t c h e s} T _ {o} \\\\ - 1, & \\text {i f} M _ {n} ^ {i} \\text {m a t c h e s n o t h i n g} \\end{array} \\right. \\tag {2}", + "image_path": "83a2e1f1d0cf511315176693faa34e46bf38b18436b96bd7e283af5d7ebc7aef.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "content": "We define the query utilization of layer " + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "content": " as the proportion of GT masks that match with the same query in layer " + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18076" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 72, + 438, + 162 + ], + "blocks": [ + { + "bbox": [ + 157, + 72, + 438, + 162 + ], + "lines": [ + { + "bbox": [ + 157, + 72, + 438, + 162 + ], + "spans": [ + { + "bbox": [ + 157, + 72, + 438, + 162 + ], + "type": "table", + "html": "
Layer Number123456789
Mask2FormermIoU-Li(%)23.050.359.569.770.169.978.676.376.0
Utili(%)38.165.570.873.578.581.183.283.2100
MP trainingmIoU-Li(%)51.583.286.696.296.093.897.697.996.3
Utili* (%)94.093.894.693.294.094.894.094.0100
Multi-layer MPmIoU-Li(%)51.387.888.897.597.596.299.098.898.2
Utili* (%)94.295.497.397.697.397.098.298.2100
", + "image_path": "36467bc75245240489b7851b8cde2d0f16b8b776db05d51ceff2438f86061aab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": "Table 1. The mIoU-L" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": " and Util" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": " of Mask2Former and the proposed methods. The mIoU-L" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": " for a query is taken as the IoU of the predicted masks in the " + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "(i-1)" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": "-th decoder layers. The mIoU-L" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": " takes the average over all queries. Note that the index of decoder layers starts from 1, so \"prediction of the 0-th decoder\" means the prediction before entering Transformer decoder. The Util" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": " is the ratio of GT masks that match with the same query in " + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": "-th layer and the last layer. * means that We test Util" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": " with bipartite matching to assign predictions in MP part to GT masks. Note that when directly assigning predictions in MP part to corresponding GT masks with hard assignment, Util" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 172, + 547, + 240 + ], + "type": "text", + "content": " is always 100%." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 246, + 194, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 246, + 194, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 194, + 258 + ], + "type": "text", + "content": "the last layer, which is calculated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 265, + 288, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 265, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 80, + 265, + 288, + 299 + ], + "type": "interline_equation", + "content": "\\operatorname {U t i l} ^ {i} = \\frac {1}{O} \\sum_ {j = 0} ^ {N} \\mathbb {1} \\left(V _ {n} ^ {i} = V _ {n} ^ {l} \\text {a n d} V _ {n} ^ {i} \\neq - 1\\right), \\tag {3}", + "image_path": "8c2fadffbe582ecf4754b1bf758f04ac14269df5f7f33fa1e84460b3747fb6a2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": " is the indicator function. " + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(x) = 1" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": " is true and 0 otherwise. The query utilization of layer " + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": " for the whole data set is averaged over the utilization numbers for all images. Table 1 shows that the query utilizations of the first few layers are too low, which indicate many matched queries in early layers are ended up not being used in the last layer. The low mIoU-" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "inline_equation", + "content": "L^i" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathrm{Util}^i" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": " have potential consequences such as inconstant optimization goals and low training efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 437 + ], + "type": "text", + "content": "Our observations based on these two metrics motivate us to develop a more effective training method to make the matching more consistent among layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "text", + "content": "Solution: To address the inconsistent prediction problem, inspired by DN-DETR, we proposed a mask-piloted training method. We feed GT masks as attention masks and expect these queries with GT masks can focus on their corresponding GT masks and not be distracted by other instances. As shown in row 2 and 3, Table 1, both mIoU-" + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "inline_equation", + "content": "L^i" + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "inline_equation", + "content": "\\text{Util}^i" + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "text", + "content": " are improved by a large margin when adopting mask-piloted training. We also find that only applying mask-piloted (MP) training in the first layer is not enough. Fig 3(a)(b) shows that inconsistent prediction exists even when given GT masks. Therefore, we add a GT mask to all decoder layers to reinforce MP training. As shown in Table 1, multi-layer MP training further improves both mIoU-" + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "inline_equation", + "content": "L^i" + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "inline_equation", + "content": "\\text{Util}^i" + }, + { + "bbox": [ + 46, + 437, + 288, + 605 + ], + "type": "text", + "content": " especially in latter decoder layers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 616, + 125, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 125, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 125, + 628 + ], + "type": "text", + "content": "4. MP-Former" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 115, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 115, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 115, + 647 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Our method is a training method for improving Mask2Former. Our MP-Former only introduces minor changes in the Mask2Former pipeline during training as shown in the red-line part in Fig. 2. Same as Mask2Former, our model is composed of a backbone, a pixel decoder, and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "text", + "content": "a Transformer decoder. Our improvements are on the Transformer decoder. We divide decoder queries into two parts, a mask-piloted (MP) part and a matching part. The matching part is the same as Mask2Former. The MP part feeds GT masks and GT class embeddings as input and the predictions are directly assigned to the corresponding GT masks without bipartite matching. Our method includes three components which are multi-layer MP training, mask noises, and label-guided training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 359, + 443, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 443, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 443, + 373 + ], + "type": "text", + "content": "4.2. Multi-layer MP training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 378, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 582 + ], + "type": "text", + "content": "In order to address the inconsistent prediction problem, we propose a multi-layer mask-piloted (MP) training method. In the MP part, we feed an additional set of queries and masks into Transformer where the queries and masks are GT class embeddings and GT masks. On the loss side, we independently assign the outputs of the MP part and the matching part with the GT instances. The predictions of the MP part are directly assigned to their corresponding GT masks and those in the matching part are assigned with bipartite graph matching. The losses for both parts are the same, following the loss design in Mask2Former. We also propose to add GT masks in multiple layers and show that this can further improve the performance with experiments in Table 8. Since we follow Mask2Fomer to use feature maps of different resolutions in different decoder layers, we interpolate GT masks into different resolutions when applied to different layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 587, + 392, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 392, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 392, + 599 + ], + "type": "text", + "content": "4.3. Noised masks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Besides feeding GT masks to alleviate the inconsistency problem, we further propose to use noised GT masks to make mask refinement more robust in each decoder. The intuition is simple, as Mask2former refines the mask from an inaccurate mask predicted from the previous decoder. GT masks without noise may be too easy for the task which prevents further refinement. Therefore, we feed noised masks to the decoder and train the model to reconstruct the original ones, which is similar to DN-DETR [19] for box denoising" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18077" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 60, + 482, + 289 + ], + "blocks": [ + { + "bbox": [ + 112, + 60, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 112, + 60, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 112, + 60, + 482, + 289 + ], + "type": "image", + "image_path": "0ef37a30044fec179f93f0049192a0895496c5a34aca5ffc34c351966353ad34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 301, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 545, + 367 + ], + "type": "text", + "content": "Figure 2. The architecture of our method is the same as Mask2Former (the blue-shaded part), which consists of a backbone, a pixel decoder, and a Transformer decoder. The difference is that we feed extra queries and attention masks which are called the MP part to the Transformer decoder (red-line part in the figure). The MP part contains GT masks as attention masks and GT class embeddings as queries. We feed GT masks into the MP part of all decoder layers. We also add point noises to GT masks and flipping noises to class embeddings which can further improve the performance. Note that this architecture is just for training. In the inference time, the red-line part does not exist, and thus, our pipeline is exactly the same as Mask2Former." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 94, + 373, + 175, + 479 + ], + "blocks": [ + { + "bbox": [ + 94, + 373, + 175, + 479 + ], + "lines": [ + { + "bbox": [ + 94, + 373, + 175, + 479 + ], + "spans": [ + { + "bbox": [ + 94, + 373, + 175, + 479 + ], + "type": "image", + "image_path": "ffb0406ad8ecb8c0d139e417bf7dc7a71ce5409276ee7822ba95b63aedbdb58d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 481, + 137, + 491 + ], + "lines": [ + { + "bbox": [ + 127, + 481, + 137, + 491 + ], + "spans": [ + { + "bbox": [ + 127, + 481, + 137, + 491 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 373, + 257, + 479 + ], + "blocks": [ + { + "bbox": [ + 176, + 373, + 257, + 479 + ], + "lines": [ + { + "bbox": [ + 176, + 373, + 257, + 479 + ], + "spans": [ + { + "bbox": [ + 176, + 373, + 257, + 479 + ], + "type": "image", + "image_path": "d8dabe7dcf632b3457faf852f0307df0e9a4e540580e91d75365d2bc03571e91.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 481, + 219, + 490 + ], + "lines": [ + { + "bbox": [ + 209, + 481, + 219, + 490 + ], + "spans": [ + { + "bbox": [ + 209, + 481, + 219, + 490 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 500, + 545, + 543 + ], + "lines": [ + { + "bbox": [ + 46, + 500, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 545, + 543 + ], + "type": "text", + "content": "Figure 3. (a) The red region is a given GT mask in first decoder layer which covers the man riding the black horse. (b) The predicted mask of the first decoder layer which only covers the head of the man. The predicted mask will be used as the attention mask of the second layer and may mislead the second layer. (c)(d)(e) are demonstrations of point, shift and scale noises. The blue regions are noised masks and the red regions are the GT masks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 258, + 373, + 338, + 479 + ], + "blocks": [ + { + "bbox": [ + 258, + 373, + 338, + 479 + ], + "lines": [ + { + "bbox": [ + 258, + 373, + 338, + 479 + ], + "spans": [ + { + "bbox": [ + 258, + 373, + 338, + 479 + ], + "type": "image", + "image_path": "021dd0923be2447f4d3feb4a6d2a45488a564226801cc11d865edf0411c1ce4a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 481, + 301, + 490 + ], + "lines": [ + { + "bbox": [ + 290, + 481, + 301, + 490 + ], + "spans": [ + { + "bbox": [ + 290, + 481, + 301, + 490 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 339, + 373, + 420, + 479 + ], + "blocks": [ + { + "bbox": [ + 339, + 373, + 420, + 479 + ], + "lines": [ + { + "bbox": [ + 339, + 373, + 420, + 479 + ], + "spans": [ + { + "bbox": [ + 339, + 373, + 420, + 479 + ], + "type": "image", + "image_path": "6c15f3fec374963edf72c391848f3981debf8374fda855e754995b7580752d09.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 481, + 383, + 490 + ], + "lines": [ + { + "bbox": [ + 372, + 481, + 383, + 490 + ], + "spans": [ + { + "bbox": [ + 372, + 481, + 383, + 490 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 421, + 373, + 502, + 479 + ], + "blocks": [ + { + "bbox": [ + 421, + 373, + 502, + 479 + ], + "lines": [ + { + "bbox": [ + 421, + 373, + 502, + 479 + ], + "spans": [ + { + "bbox": [ + 421, + 373, + 502, + 479 + ], + "type": "image", + "image_path": "3102a2f869c631bf2502811a4bc99b99e4674d7779db3e4fc5de5b235b5fdeff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 481, + 464, + 490 + ], + "lines": [ + { + "bbox": [ + 454, + 481, + 464, + 490 + ], + "spans": [ + { + "bbox": [ + 454, + 481, + 464, + 490 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 107, + 544, + 487, + 641 + ], + "blocks": [ + { + "bbox": [ + 107, + 544, + 487, + 641 + ], + "lines": [ + { + "bbox": [ + 107, + 544, + 487, + 641 + ], + "spans": [ + { + "bbox": [ + 107, + 544, + 487, + 641 + ], + "type": "table", + "html": "
methodbackbonePQpanoptic APth panmIoUpaninstancesemantic mIoU
APAPSAPMAPL
MaskFormer [8]R5034.7-------
Panoptic-DeepLab [6]SWideRNet [5]37.9*-50.0*-----
Mask2Former [7]R5039.726.546.126.410.428.943.147.2
MP-FormerR5040.8(+1.1)27.148.3(+2.2)28.0(+1.6)10.530.7(+1.8)44.648.1
MaskFormer [8]Swin-L†-------54.1
FaPN-MaskFormer [8, 17]Swin-L†-------55.2
Mask2Former [7]Swin-L†48.134.254.534.916.340.054.756.1
MP-FormerSwin-L†49.4(+1.3)35.256.035.316.340.755.556.9
", + "image_path": "d1542bc0fcda542308452306a9e534b1c6ed9010c16cf256292329df0a4b8a39.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 648, + 547, + 681 + ], + "lines": [ + { + "bbox": [ + 46, + 648, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 547, + 681 + ], + "type": "text", + "content": "Table 2. Image segmentation results on ADE20K va1. MP-Former improves the performance on all three segmentation tasks with R50 and Swin-L as backbone. All metrics are evaluated with single-scale inference. Backbones pre-trained on ImageNet-22K are marked with " + }, + { + "bbox": [ + 46, + 648, + 547, + 681 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 46, + 648, + 547, + 681 + ], + "type": "text", + "content": " . - means the results are not reported for the methods." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "training. We have tried three types of noise which are point noise, shift noise, and scale noise as shown in Fig. 3. For" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 712 + ], + "type": "text", + "content": "point noise, we randomly sample some points on the mask and flip them from 1 to 0 or from 0 to 1. The number of" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18078" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 70, + 463, + 141 + ], + "blocks": [ + { + "bbox": [ + 130, + 70, + 463, + 141 + ], + "lines": [ + { + "bbox": [ + 130, + 70, + 463, + 141 + ], + "spans": [ + { + "bbox": [ + 130, + 70, + 463, + 141 + ], + "type": "table", + "html": "
methodbackbonePQpanoptic APTh panmIoUpaninstancesemantic mIoU
APAP50
Panoptic-DeepLab [6]R5060.332.178.7---
Mask2Former [7]R5062.137.377.537.461.979.4
MP-FormerR5062.739.3(+2.0)79.6(+2.1)39.7(+2.3)65.4(+3.5)81.0(+1.6)
Mask2Former [7]Swin-L†66.643.682.943.771.483.3
MP-FormerSwin-L†67.545.8(+2.2)83.544.9(+1.2)72.483.9
", + "image_path": "f85ec653f25d7039640affb01efa4e4c17a1e4fbf89e48d956b812cf286be51d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 148, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 547, + 171 + ], + "type": "text", + "content": "Table 3. Image segmentation results on Cityscapes va1. We report results on all three segmentation tasks with R50 or Swin-L as backbone. All metrics are evaluated with single-scale inference. Backbones pre-trained on ImageNet-22K are marked with" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "text", + "content": "sampled points is proportional to the area of a GT mask. For shift noise, we randomly shift a mask instance vertically and horizontally while keeping the center of the instance within its GT bounding box. For the scaling noise, we enlarge or shrink the instance by a random ratio. A comparison of three types of noises is shown in Table 8. Through experiments, we find point noises improve the performance while shifting and scaling noises do not work for our method. Therefore, we adopt point noises in our implementation. We use a noise ratio " + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "inline_equation", + "content": "\\lambda_{p}" + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "text", + "content": " to control how many noise points are added. Denote a binary GT mask as " + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "inline_equation", + "content": "M = [m_{i,j}]_{H\\times W}" + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "inline_equation", + "content": "m_{i,j} = 1" + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "text", + "content": " means the pixel belongs to the mask. The number of noise points is randomly sampled from 0 to " + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "inline_equation", + "content": "\\lambda_{p}\\times \\sum_{i,j}m_{i,j}" + }, + { + "bbox": [ + 46, + 186, + 289, + 378 + ], + "type": "text", + "content": ". Since we apply mask-piloted training to all decoder layers, we independently add different noises to GT masks in different layers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 387, + 173, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 173, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 173, + 399 + ], + "type": "text", + "content": "4.4. Label-guided training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": "We use the class embeddings of GT categories as queries in the MP part. The reason why we use class embeddings is that queries will dot-product with image features and an intuitive way to distinguish instances is to use their categories. Note that we only use class embeddings in the first layer and do not replace predicted queries with class embeddings in subsequent layers. We also have a classification loss in the MP part. Since the classification task becomes very easy given GT class embeddings, we add flipping noises in class embeddings. We randomly sample GT class embeddings and let them randomly flip to the class embeddings of other categories. For example, " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "E_0, E_1, \\dots, E_K" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": " denotes the class embeddings of " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": " categories. An image contains " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": " instances which belong to Categories " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "C^0, C^1, \\dots, C^N" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": ". So their default class embeddings should be " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "E_{C^0}, E_{C^1}, \\dots, E_{C^N}" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": ". To add flipping noise, we enumerate all " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": " instances. For " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": "-th instance, we generate a random number " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": " in [0, 1] to decide whether to flip the class embedding " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "E_{C^i}" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": " or not. If " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "r_i < \\lambda_l" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": ", we randomly sample another class embedding in " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "E_0, E_1, \\dots, E_K" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": " to replace " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "E_{C^i}" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "inline_equation", + "content": "\\lambda_l" + }, + { + "bbox": [ + 46, + 406, + 289, + 657 + ], + "type": "text", + "content": " denotes the proportion of the sampled class embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "We conduct extensive experiments to compare our method with Mask2Former and other segmentation methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 186, + 547, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 547, + 234 + ], + "type": "text", + "content": "on all three segmentation tasks on three datasets to show the effectiveness of our method. We also conduct several ablation studies to show the effectiveness of each component in our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 242, + 437, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 242, + 437, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 437, + 255 + ], + "type": "text", + "content": "5.1. Implementation details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 261, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 547, + 368 + ], + "type": "text", + "content": "Model: Since our method is an improved training method based on Mask2Former, we adopt an identical model setting as Mask2Former except for the mask-piloted (MP) part. In the MP part, we adopt 100 GT masks with point noises. We adopt noise ratios " + }, + { + "bbox": [ + 304, + 261, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\lambda_{p} = 0.2" + }, + { + "bbox": [ + 304, + 261, + 547, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 261, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\lambda_{l} = 0.2" + }, + { + "bbox": [ + 304, + 261, + 547, + 368 + ], + "type": "text", + "content": " for point noises on masks and flipping noises on class embeddings. We adopt ResNet50 [16] and Swin-L [24] when evaluating our method on ADE20K and Cityscapes. On COCO2017, we use ResNet50 and ResNet101 as the backbone." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": "Dataset and metrics: We evaluate MP-Former on three challenging datasets: Cityscapes [9, 10] and ADE20K [36] for all three segmentation tasks and COCO 2017 [22] dataset for panoptic segmentation and instance segmentation. For instance segmentation, we evaluate the mask AP [22] on instances that are denoted as \"thing\" categories in datasets. For semantic segmentation, we evaluate mean Intersection-over-Union (mIOU) over all categories including both foreground and background. For panoptic segmentation, we evaluate the results with the panoptic quality (PQ) metric [18]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 498, + 391, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 391, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 391, + 509 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 581 + ], + "type": "text", + "content": "We trained on three challenging datasets ADE20K, Cityscapes, and MS COCO with different backbones. Our method outperforms Mask2Former on all three segmentation tasks (instance, panoptic, and semantic) under the same setting, which demonstrates its effectiveness and generalization ability to different datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": "Performance on ADE20K. On ADE20K, we report our performance trained for " + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "inline_equation", + "content": "160K" + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": " steps following Mask2Former. As shown in Table 2, our method exceeds Mask2Former and other baselines on all metrics with both R50 and Swin-L. Notably, MP-Former achieves " + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "inline_equation", + "content": "+1.6\\mathrm{AP}" + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": " on instance segmentation with a R50 backbone and " + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "inline_equation", + "content": "+1.3\\mathrm{PQ}" + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": " on panoptic segmentation with a Swin-L backbone." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Performance on Cityscapes. On Cityscapes, we train the model with 90K training steps following Mask2Former. As shown in Table 3, our method consistently outperforms Mask2Former on all three segmentation tasks across differ-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18079" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 69, + 492, + 208 + ], + "blocks": [ + { + "bbox": [ + 99, + 69, + 492, + 208 + ], + "lines": [ + { + "bbox": [ + 99, + 69, + 492, + 208 + ], + "spans": [ + { + "bbox": [ + 99, + 69, + 492, + 208 + ], + "type": "table", + "html": "
methodbackbonequery typeepochsAP\\( AP^S \\)\\( AP^M \\)\\( AP^L \\)#params.FLOPsfps
MaskFormer [8]R50100 queries30034.016.437.854.245M181G19.2
Mask R-CNN [15]R50dense anchors3637.218.639.553.344M201G15.2
Mask R-CNN [12, 14, 15]R50dense anchors40042.523.845.060.046M358G10.3
Mask2Former [7]R50100 queries5043.723.447.264.844M226G9.7
MP-FormerR50100 queries5044.8(+1.1)24.748.165.544M226G9.7
Mask R-CNN [15]R101dense anchors3638.619.541.355.363M266G10.8
Mask R-CNN [12, 14, 15]R101dense anchors40043.724.646.461.865M423G8.6
Mask2Former [7]R101100 queries5044.223.847.766.763M293G7.8
MP-FormerR101100 queries5045.1(+0.9)24.648.766.263M293G7.8
QueryInst [13]Swin-L†300 queries5048.930.852.668.3--3.3
Swin-HTC++ [2, 24]Swin-L†dense anchors7249.531.052.467.2284M1470G-
Mask2Former [7]Swin-L†200 queries10050.129.953.972.1216M868G4.0
MP-Former (ours)Swin-L†200 queries7250.8(+0.7)30.354.572.6216M868G4.0
", + "image_path": "faaa9133255d1fcdf4bedade27e8e20c29b2a1c11aabb11831de73ab5f6302c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 105, + 255, + 488, + 393 + ], + "blocks": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": "Table 4. Instance segmentation on COCO va12017 with 80 categories. MP-Former outperforms strong baseline Mask2Former without extra computation cost in inference time. For a fair comparison, we only consider single-scale inference and models trained using only COCO train2017 set data. Backbones pre-trained on ImageNet-22K are marked with" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 255, + 488, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 488, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 488, + 393 + ], + "type": "table", + "html": "
methodbackbonequery typeepochsPQ\\( PQ^{Th} \\)\\( PQ^{St} \\)\\( AP^{Th}_{pan} \\)\\( mIoU_{pan} \\)#params.FLOPsfps
DETR [1]R50100 queries500+2543.448.236.331.1----
MaskFormer [8]R50100 queries30046.551.039.833.057.845M181G17.6
Mask2Former [7]R50100 queries5051.4*57.242.741.961.544M226G8.6
MP-Former (ours)R50100 queries5052.2(+0.8)58.342.942.561.644M226G8.6
DETR [1]R101100 queries500+2545.150.537.033.0----
MaskFormer [8]R101100 queries30047.652.540.334.159.364M248G14.0
Mask2Former [7]R101100 queries5052.658.543.742.662.463M293G7.2
MP-Former (ours)R101100 queries5052.9(+0.3)58.744.043.061.863M293G7.2
Max-DeepLab [30]Max-L128 queries21651.157.042.2--451M3692G-
MaskFormer [8]Swin-L†100 queries30052.758.544.040.164.8212M792G5.2
K-Net [34]Swin-L†100 queries3654.660.246.0-----
Mask2FormerSwin-L†200 queries10057.864.248.148.667.4216M868G4.0
MP-FormerSwin-L†200 queries7258.1(+0.3)64.448.448.967.6216M868G4.0
", + "image_path": "45a17be18ebfcb4cde6a358565274057a4ebb76f2c906bbda1cf1ce448f7e455.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 399, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 546, + 422 + ], + "type": "text", + "content": "Table 5. Panoptic segmentation on COCO panoptic va12017 with 133 categories. MP-Former outperforms Mask2Former on COCO panoptic segmentation with both R50 and R101 as backbone. Results that are implemented are marked with * ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 436, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 288, + 483 + ], + "type": "text", + "content": "ent backbones. MP-Former also scales well when adopting Swin-L as the backbone. Especially, our method outperforms Mask2Former by " + }, + { + "bbox": [ + 46, + 436, + 288, + 483 + ], + "type": "inline_equation", + "content": "+2.3\\mathrm{AP}" + }, + { + "bbox": [ + 46, + 436, + 288, + 483 + ], + "type": "text", + "content": " on instance segmentation and " + }, + { + "bbox": [ + 46, + 436, + 288, + 483 + ], + "type": "inline_equation", + "content": "+1.6\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 436, + 288, + 483 + ], + "type": "text", + "content": " on semantic segmentation with R50." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 484, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 651 + ], + "type": "text", + "content": "Performance on COCO. On COCO val2017, we follow Mask2Former's setting to train for 50epochs. As shown in Table 4 and Table 5, MP-Former outperforms both Mask2Former and classical methods such as Mask R-CNN [15] on instance segmentation and panoptic segmentation. For instance segmentation, our method achieves +1.1AP with R50, +0.9AP with R101 and +0.7 with Swin-L [24]. Note that we only train MP-Former for 72 epochs with Swin-L. On panoptic segmentation, our method achieves +0.8PQ with R50 and +0.3PQ with R101. Notably, our method is only a pluggable training method that will not affect the model architecture or inference speed. In addition, it only introduces negligible parameters (class embeddings) during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 154, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 154, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 154, + 672 + ], + "type": "text", + "content": "5.3. Speed up training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "In Table 6, we show our MP-Former can significantly speed up training on ADE20k and COCO. We do not report the results on Cityscapes because this dataset is relatively" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 436, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 616 + ], + "type": "text", + "content": "small and only needs 90k steps to converge in Mask2Former, which is not necessary for additional speedup. The results show that our method trained with 36 epochs achieved comparable results with Mask2Former trained with 50 epochs with both R50 and R101 as the backbone on both panoptic and instance segmentation. Moreover, our method trained with half number of training steps (80K steps) achieves comparable results with Mask2Former trained with 160K steps. Note that we use the same batch size as Mask2Former. We also report the clock time. The results show that our method reduces the actual training time by large. For a fair comparison, we use the same number of GPUs to run MP-Former and Mask2Former. Because there may exist differences in speed for different machines, we run our method and Mask2Former on the same machine to test the running time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "text", + "content": "We conduct an ablation study on the COCO instance segmentation task with R50 as backbone." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 666, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 712 + ], + "type": "text", + "content": "Effectiveness of each component: We show the effectiveness of each version of our method trained for 12 epochs and 50 epochs in Table 7, where we show the result of raw Mask2Former and the result after each version of our method" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18080" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 69, + 504, + 180 + ], + "blocks": [ + { + "bbox": [ + 93, + 69, + 504, + 180 + ], + "lines": [ + { + "bbox": [ + 93, + 69, + 504, + 180 + ], + "spans": [ + { + "bbox": [ + 93, + 69, + 504, + 180 + ], + "type": "table", + "html": "
datasetmethodbackboneepochs*PQpanopticAPAPSinstancesemantic
APthpanmIoUpanclocktime/hAPMAPLclocktime/hmIoUclocktime/h
ADE20kMask2FormerR50160k39.726.546.152.826.410.428.943.140.347.221.3
MP-FormerR5080k39.727.048.434.227.410.230.044.525.547.412.3
Mask2FormerSwin-L†160k48.134.254.563.234.916.340.054.751.456.139.8
MP-FormerSwin-L†80k48.534.756.145.735.215.839.655.626.756.921.5
COCOMask2FormerR505051.4*41.961.577.943.723.447.264.862.5--
MP-FormerR503651.541.861.659.544.124.147.465.350.8--
Mask2FormerR1015052.642.662.4113.144.223.847.766.786.0--
MP-FormerR1013652.542.662.084.444.223.247.966.671.5--
", + "image_path": "5b4144989caf7c37b9f2b843192672d337991c9679e1697391e41edc1987a2c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 268, + 248, + 320 + ], + "blocks": [ + { + "bbox": [ + 46, + 188, + 547, + 244 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 244 + ], + "type": "text", + "content": "Table 6. Image segmentation results on ADE20K va1. and COCO va12017 MP-Former speeds up training by large. On ADE20K, MP-Former trained for 80K steps can achieve comparable result with Mask2Former trained for 160K steps with both R50 and Swin-L as backbone. On COCO, MP-Former trained for 36 epochs can achieve comparable result with Mask2Former trained for 50 epochs with both R50 and R101 as backbone. * means that we adopt number of steps instead of epochs on ADE20k following Mask2Former. We report only single-scale inference results. Backbones pre-trained on ImageNet-22K are marked with †" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 268, + 248, + 320 + ], + "lines": [ + { + "bbox": [ + 49, + 268, + 248, + 320 + ], + "spans": [ + { + "bbox": [ + 49, + 268, + 248, + 320 + ], + "type": "table", + "html": "
#Row12 ep (AP mask)50 ep (AP mask)
1. Mask2Former [7]38.743.7
2. Row1+MP training on 1st layer39.444.2
3. Row2+Label noises39.644.2
4. Row3+Point noises39.944.6
5. Ours (Row4+All-layer MP)40.244.8
", + "image_path": "36d2f876229b73493d672388176fd90619c5225bebde994e15cd5135b2369dc0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 324, + 230, + 376 + ], + "lines": [ + { + "bbox": [ + 47, + 324, + 230, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 230, + 376 + ], + "type": "text", + "content": "Table 7. The effectiveness of each component of our method when trained for 12 epochs and 50 epochs. The table shows that each component of our model effectively improves the results. Label noise does not work when trained for 50 epochs. Especially, our method without noises can improve the results by " + }, + { + "bbox": [ + 47, + 324, + 230, + 376 + ], + "type": "inline_equation", + "content": "+0.9\\mathrm{AP}" + }, + { + "bbox": [ + 47, + 324, + 230, + 376 + ], + "type": "text", + "content": ". With multi-layer MP training, our method achieves " + }, + { + "bbox": [ + 47, + 324, + 230, + 376 + ], + "type": "inline_equation", + "content": "+1.5\\mathrm{AP}" + }, + { + "bbox": [ + 47, + 324, + 230, + 376 + ], + "type": "text", + "content": " on COCO instance segmentation with R50 backbone trained for 12 epochs. \"ep\" is short for epochs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 253, + 266, + 377, + 342 + ], + "blocks": [ + { + "bbox": [ + 253, + 266, + 377, + 342 + ], + "lines": [ + { + "bbox": [ + 253, + 266, + 377, + 342 + ], + "spans": [ + { + "bbox": [ + 253, + 266, + 377, + 342 + ], + "type": "table", + "html": "
Layers for MP trainingepAP mask
First layer1239.4
First three layers1239.6(+0.2)
All layers1240.2(+0.8)
Noise Type
No noise1239.6
Point noise1239.9
Shifting noise1239.6
Scaling noise1239.4
", + "image_path": "e101018c147e6735c3ccb1d19edb501526673d457d1bc8225a4cbcae8999cc12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 244, + 347, + 384, + 378 + ], + "lines": [ + { + "bbox": [ + 244, + 347, + 384, + 378 + ], + "spans": [ + { + "bbox": [ + 244, + 347, + 384, + 378 + ], + "type": "text", + "content": "Table 8. Top 3 lines shows the effectiveness of multi-layer Mask-Piloted (MP) training. The bottom lines of the table shows that shifting and scaling noise do not work for our method. \"ep\" is short for epochs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 392, + 266, + 542, + 320 + ], + "blocks": [ + { + "bbox": [ + 392, + 266, + 542, + 320 + ], + "lines": [ + { + "bbox": [ + 392, + 266, + 542, + 320 + ], + "spans": [ + { + "bbox": [ + 392, + 266, + 542, + 320 + ], + "type": "table", + "html": "
Feature mapepAP mask
Coarse to fine1240.2
All large1240.0
Assignment in Mask-Piloted part
Match1239.4
Hard assignment1239.6
", + "image_path": "92720e287f54cbdb99c79ed8eb395d8e74b1cba5f458c14af84bea7d7fd9a4b7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 387, + 324, + 546, + 354 + ], + "lines": [ + { + "bbox": [ + 387, + 324, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 387, + 324, + 546, + 354 + ], + "type": "text", + "content": "Table 9. Some hyperparemeter tuning:1.coarse-to-fine manner works better for our method than all-large feature maps. 2.Hard assignment lead to better results than matching for MP part queries. \"ep\" is short for epochs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "content": "is adopted. Without adding noises, our method achieves an improvement of " + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "inline_equation", + "content": "+0.9AP" + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "content": ". With multi-layer MP-training, our method achieves " + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "inline_equation", + "content": "+1.5AP" + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "content": ". We also find that label noise does not work when trained for 50 epochs. Note that except for the last line of the table, other lines of our method apply MP training on the first 3 layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 472, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 531 + ], + "type": "text", + "content": "Apply MP training on multiple layers: In Table 8, we show the effectiveness of applying GT masks on multiple layers. The results indicate that more layers of GT masks can lead to better performance. By adding noised GT masks in all layers, MP-Former achieves " + }, + { + "bbox": [ + 46, + 472, + 287, + 531 + ], + "type": "inline_equation", + "content": "+0.8\\mathrm{AP}" + }, + { + "bbox": [ + 46, + 472, + 287, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 532, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 288, + 604 + ], + "type": "text", + "content": "Add noises to GT masks: In the last four rows of Table 8, we show the effectiveness of adding different types of noises. Note that without any noise, MP-Former achieves 39.6AP which is corresponding to the 4th row in Table 7. Shifting noise and scaling noise do not work in our method. Point noise achieved " + }, + { + "bbox": [ + 46, + 532, + 288, + 604 + ], + "type": "inline_equation", + "content": "+0.3" + }, + { + "bbox": [ + 46, + 532, + 288, + 604 + ], + "type": "text", + "content": " extra growth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Some hyper-parameter tuning: Mask2Former adopts coarse-to-fine feature maps in decoder layers. We try to adopt the largest feature map (1/8 feature map) in all decoder layers. Table 9 shows that a coarse-to-fine manner is better than using the largest feature map in all layers for our MP-Former. We also try to use matching to assign predictions to GT instances in the MP part. The result shows that matching can also improve the result compared with raw Mask2Former but the hard assignment is better." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 397, + 379, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 379, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 379, + 410 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 438, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 676 + ], + "type": "text", + "content": "In this paper, we have analyzed failure cases of Mask2Former and found a problem of inconsistent predictions between adjacent layers which potentially leads to inconstant optimization goals and low utilization of queries and therefore limits the performance of Mask2Former. In order to address the problem, we developed a new MP-Former, whose key components include multi-layer mask-piloted training with point noises and label-guided training. Our experimental results showed that the proposed training method effectively mitigates the inconsistent prediction problem. When evaluating on three challenging datasets (ADE20K, Citescapes, and MS COCO), our method achieved improvement by a large margin on all three image segmentation tasks (semantic, instance, and panoptic). In addition, our method only introduces little computation during training and no extra computation in inference time. Finally, our method can also significantly speed up training. Especially, when trained with half number of steps, our method exceeds Mask2Former on all three segmentation tasks on ADE20K with both R50 and Swin-L as the backbones." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Limitations: We evaluated MP-Former on Mask2Former, but our method may be useful to some other models such as MaskFormer and HTC. We will take it as our future work." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18081" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 212 + ], + "type": "text", + "content": "[2] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "type": "text", + "content": "[3] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, Huiyu Wang, and Siyuan Qiao. Scaling wide residual networks for panoptic segmentation. arXiv preprint arXiv:2011.11675, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 412 + ], + "type": "text", + "content": "[6] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12475-12485, 2020. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 457, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 457, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 457, + 288, + 491 + ], + "type": "text", + "content": "[8] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. 2021. 2, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 556 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "type": "text", + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Scharwächter, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset. In CVPR Workshop on The Future of Datasets in Vision, 2015. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 679, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 679, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 679, + 288, + 713 + ], + "type": "text", + "content": "[12] Xianzhi Du, Barret Zoph, Wei-Chih Hung, and Tsung-Yi Lin. Simple training strategies and model scaling for object detection. arXiv preprint arXiv:2107.00057, 2021. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[13] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "type": "text", + "content": "[14] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 229, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 547, + 274 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 274, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 547, + 319 + ], + "type": "text", + "content": "[17] Shihua Huang, Zhichao Lu, Ran Cheng, and Cheng He. Fapn: Feature-aligned pyramid network for dense image prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 864-873, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 319, + 547, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 547, + 364 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 547, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 547, + 408 + ], + "type": "text", + "content": "[19] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. arXiv preprint arXiv:2203.01305, 2022. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 410, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 547, + 453 + ], + "type": "text", + "content": "[20] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 455, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 547, + 488 + ], + "type": "text", + "content": "[21] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Tong Lu, and Ping Luo. Panoptic segformer. arXiv preprint arXiv:2109.03814, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 489, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 547, + 544 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 547, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 547, + 588 + ], + "type": "text", + "content": "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 646, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 547, + 690 + ], + "type": "text", + "content": "[25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[26] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Con" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18082" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 631 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "ditional detr for fast training convergence. arXiv preprint arXiv:2108.06152, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 194 + ], + "type": "text", + "content": "[28] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "type": "text", + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "type": "text", + "content": "[30] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "type": "text", + "content": "[31] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. arXiv preprint arXiv:2109.07107, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 384 + ], + "type": "text", + "content": "[32] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 288, + 430 + ], + "type": "text", + "content": "[33] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 432, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 288, + 464 + ], + "type": "text", + "content": "[34] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 466, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 531 + ], + "type": "text", + "content": "[35] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 576 + ], + "type": "text", + "content": "[36] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 631 + ], + "type": "text", + "content": "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR 2021: The Ninth International Conference on Learning Representations, 2021. 1, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18083" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_content_list.json b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..539007c6a4646d681d57f552d07117d220ea2913 --- /dev/null +++ b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_content_list.json @@ -0,0 +1,1383 @@ +[ + { + "type": "text", + "text": "MSF: Motion-guided Sequential Fusion for Efficient 3D Object Detection from Point Cloud Sequences", + "text_level": 1, + "bbox": [ + 204, + 128, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenhang He* Ruihuang Li* Yabin Zhang Shuai Li Lei Zhang† The Hong Kong Polytechnic University", + "bbox": [ + 214, + 203, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{csche, csrhli, csybhang, cssli, cslzhang}@comp.polyu.edu.hk", + "bbox": [ + 214, + 241, + 754, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud sequences are commonly used to accurately detect 3D objects in applications such as autonomous driving. Current top-performing multi-frame detectors mostly follow a Detect-and-Fuse framework, which extracts features from each frame of the sequence and fuses them to detect the objects in the current frame. However, this inevitably leads to redundant computation since adjacent frames are highly correlated. In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, which exploits the continuity of object motion to mine useful sequential contexts for object detection in the current frame. We first generate 3D proposals on the current frame and propagate them to preceding frames based on the estimated velocities. The points-of-interest are then pooled from the sequence and encoded as proposal features. A novel Bidirectional Feature Aggregation (BiFA) module is further proposed to facilitate the interactions of proposal features across frames. Besides, we optimize the point cloud pooling by a voxel-based sampling technique so that millions of points can be processed in several milliseconds. The proposed MSF method achieves not only better efficiency than other multi-frame detectors but also leading accuracy, with $83.12\\%$ and $78.30\\%$ mAP on the LEVEL1 and LEVEL2 test sets of Waymo Open Dataset, respectively. Codes can be found at https://github.com/skyhehel23/MSF.", + "bbox": [ + 75, + 323, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection [1,2,6,7,9,14,21,27-29,36] is one of the key technologies in autonomous driving, which helps the vehicle to better understand the surrounding environment and make critical decisions in the downstream tasks. As an indispensable sensing device in autonomous driving systems, LiDAR collects 3D measurements of the scene in", + "bbox": [ + 75, + 772, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16385f200141b776c4a846848ac3ffa1b9ab7eeb7285b35ce2bb067b73d9d325.jpg", + "image_caption": [ + "Figure 1. (a) The \"Detect-and-Fuse\" framework extracts features from each frame of the sequence and then fuses them, while (b) our proposed \"Motion-guided Sequential Fusion\" (MSF) method generates proposals on the current frame and propagates them to preceding frames to explore useful contexts in the sequence." + ], + "image_footnote": [], + "bbox": [ + 509, + 289, + 746, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ae0c5cec61df169bef6a172848d8978ddc21cf119d375f5cb6a2f001b3a1c5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 287, + 888, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the form of point clouds. However, LiDAR can only produce partial view of the scene at a time, and the sparse and incomplete representation of point clouds brings considerable challenges to the 3D object detection task. In practice, the LiDAR sensor will continuously sense the environment and produce a sequence of point cloud frames over time. The multi-frame data can provide a denser representation of the scene as the vehicle moves. Therefore, how to fuse these multi-frame point cloud data for more accurate object detection is worth deep investigation.", + "bbox": [ + 496, + 566, + 893, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works mainly focus on deep feature fusion with multi-frame point clouds, for example, aggregating dense birds-eye-view features via Transformer models [31, 37], passing the voxel features to LSTM [8] or GRU [32] modules for temporal modeling. Some top-performing detectors [2, 18] focus on fusing proposal features, where a tracker is employed to associate the 3D proposals across frames, and a region-based network is applied to refine the current proposals by incorporating contextual features from the proposal trajectories. These approaches generally follow a \"Detect-and-Fuse\" framework, as shown in Fig. 1(a), where the model requires to process each frame of the sequence,", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5196", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and the predictions on the current frame rely on the results of preceding frames. Since online detection is a causal system, such a detection framework might cause significant delay if the network is still processing a preceding frame when the current frame is loaded.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, as shown in Fig. 1(b), which leverages the continuity of object motion to extract useful contexts from point cloud sequences and improve the detection of current frame. Specifically, considering that the motions of objects are relatively smooth in a short sequence, we propagate the proposals generated on current frame to preceding frames based on the velocities of objects, and sample reliable points-of-interest from the sequence. In this way, we bypass extracting features on each frame of the sequence, which reduces the redundant computation and reliance on the results of preceding frames. The sampled points are then transformed to proposal features via two encoding schemes and passed to a region-based network for further refinement. Specifically, a self-attention module is employed to enhance the interaction of point features within proposals, while a novel Bidirectional Feature Aggregation (BiFA) module is proposed to enforce the information exchange between proposals across frames. The refined proposal features consequently capture both spatial details and long-term dependencies over the sequence, leading to more accurate bounding-box prediction.", + "bbox": [ + 75, + 167, + 470, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is found that the existing point cloud pooling methods [2, 19, 23, 30] are inefficient, taking more than 40 milliseconds when processing millions of points from sequential point clouds. We find that the major bottleneck lies in the heavy computation of pair-wise distances between $n$ points and $m$ proposals, which costs $\\mathcal{O}(nm)$ complexity. To further improve the efficiency, we optimize the point cloud pooling with a voxel sampling technique. The improved pooling operation is of linear complexity and can process millions of points in several milliseconds, more than eight times faster than the original method.", + "bbox": [ + 75, + 500, + 468, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions can be summarized as follows.", + "bbox": [ + 96, + 665, + 467, + 679 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An efficient Motion-guided Sequential Fusion (MSF) method is proposed to fuse multi-frame point clouds at region level by propagating the proposals of current frame to preceding frames based on the object motions.", + "- A novel Bidirectional Feature Aggregation (BiFA) module is introduced to facilitate the interactions of proposal features across frames.", + "- The point cloud pooling method is optimized with a voxel-based sampling technique, significantly reducing the runtime on large-scale point cloud sequence." + ], + "bbox": [ + 93, + 686, + 467, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed MSF method is validated on the challenging Waymo Open Dataset, and it achieves leading accuracy on the LEVEL1 and LEVEL2 test sets with fast speed.", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-frame 3D object detection. Recent research on single-frame 3D object detection is mainly focused on representation learning on point clouds. Voxel-based detectors [28, 33, 36] rasterize the point cloud into volumetric representation, followed by 3D CNN to extract dense features. Some works convert point clouds into 2D birds-eye-view [9] or range view [4, 11] representations, and process them with more efficient 2D CNN. Following PointNet++ [17], point-based methods [16, 23, 29, 30, 34] directly process point clouds in continuous space, and extract highly-semantic features through a series of downsampling and set abstraction layers. Voxel-point approaches [7, 12, 21] employ a hybrid representation, where the flexible conversion between voxel-based and point-based representations are explored, leading to better balance between efficiency and performance. Our method employs a high-quality voxel-based detector CenterPoint [33] as the proposal generation network to predict 3D proposals of current frame and their motions. We then employ an efficient region-based network to further refine these proposals by mining sequential points from point cloud sequence.", + "bbox": [ + 496, + 119, + 890, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D object detection from point cloud sequence. Multi-frame point clouds provide richer 3D information of the environment. While some single-frame detectors [3, 33] can be adapted to point cloud sequence by simply concatenating multi-frame point cloud as the input, the improvements are typically marginal and the performance can be even worse when encountering moving objects. Fast-and-furious [13] explores an intermediate fusion to align multi-frame point cloud by concatenating the hidden feature maps of the backbone network. However, it still suffers from the misalignment brought by the fast-moving objects in long sequence. Recent approaches [8, 32] demonstrate that an in-depth fusion can be achieved with recurrent networks. Unfortunately, the use of a single memory to store and update features across frames builds a potential bottleneck. To resolve such limitations, 3D-MAN [31] first attempts to employ the attention mechanism to align different views of 3D objects and then exploits a memory bank to store and aggregate multi-frame features for long sequence. Recently, Offboard3D [18] and MPPNet [2] improve much the detection performance, where they associate the detected boxes from each frame of the sequence as proposal trajectories, and extract high-quality proposal features by sampling sequential point cloud on the trajectories.", + "bbox": [ + 496, + 441, + 890, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our MSF method also samples points from the sequence, but it differs from those methods with proposal trajectories [2, 18] in that we only generate proposals on the current frame and propagate them to explore features in preceding frames. This makes our method much more efficient and favorable to online detection systems.", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5197", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4f2c2d8d78724a94d855046ec0cf9856ee549fcf4929dfbc796aa31204874c53.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our proposed Motion-guided Sequential Fusion (MSF) approach. By taking a point cloud sequence as input, MSF employs a region proposal network to generate proposals on the current frame and sample points-of-interest from the sequence by using motion-guided sequential pooling. The sampled points are encoded as high-dimensional proposal features and passed to a region-based network, where three learning blocks are consequently applied to refine the proposal features. A Bidirectional Feature Aggregation (BiFA) module is introduced in the region-based network to facilitate the interactions of proposal features across frames. The red and blue cubes represent single-point features from the current frame and preceding frame, respectively." + ], + "image_footnote": [], + "bbox": [ + 80, + 85, + 478, + 358 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/46252ebc4e81f5e1dc003f9b119be1c35fd45f78eefa90de399916842dc056cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 85, + 890, + 357 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/7b7974c2362f8ff564fcbb2c6a4fc8e678cea421762f11f594d66445f3c25f50.jpg", + "table_caption": [ + "Table 1. Recall rates of foreground points by using per-frame detection based proposal trajectory method [2] and our motion guided proposal generation method. We employ the CenterPoint [33] as proposal generator and evaluate on Waymo validation split." + ], + "table_footnote": [], + "table_body": "
4-frame8-frame16-frame
Trajectory [2]93.2%92.8%90.5%
Ours (γ = 1.0)92.3%87.5%78.3%
Ours (γ = 1.1)93.5%91.7%87.3%
", + "bbox": [ + 94, + 542, + 447, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Motion-guided Sequential Fusion", + "text_level": 1, + "bbox": [ + 76, + 622, + 380, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents our Motion-guided Sequential Fusion (MSF) approach for efficient 3D object detection on point cloud sequences. The overall architecture of MSF is illustrated in Fig. 2. In Sec. 3.1, we describe the details of motion-guided sequential pooling, which effectively mines reliable sequential points-of-interest based on the proposals of current frame. In Sec. 3.2, we present the region-based network, including the formulation of proposal features and a novel bidirectional feature aggregation module. In Sec. 3.3, we demonstrate a voxel-based sampling technique to accelerate the current point cloud pooling method.", + "bbox": [ + 75, + 648, + 468, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motion-guided Sequential Pooling", + "text_level": 1, + "bbox": [ + 76, + 830, + 375, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current multi-frame detection methods [2,18] mostly explore proposal trajectories to generate high-quality point cloud representations. However, such a scheme relies on", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frame-by-frame proposal generation, which is not suitable for online detection systems. We observe that in a point cloud sequence, although objects move at different speeds, their motions are relatively smooth. That is to say, we can estimate their motion displacements and roughly localize their positions in preceding frames. To this end, given a point cloud sequence $\\{I^t\\}_{t=1}^T$ , we propose to propagate the proposals generated on the current frame $I^T$ to preceding frames $\\{I^t\\}_{t=1}^{T-1}$ based on their estimated velocities. Since moving objects may slightly deviate from the estimated positions in the preceding frames, we sample the points-of-interest in a cylindrical region of each proposal and gradually increase the diameter of the region by a factor $\\gamma$ as the proposal propagates. Let's denote a proposal of current frame as $(p_x, p_y, p_z, w, l, h, \\theta)$ , where $(p_x, p_y, p_z)$ denotes its center location, $w, l, h$ and $\\theta$ denote its width, length, height and yaw angle, respectively. Suppose that the object has a unit-time velocity $\\vec{v} = (v_x, v_y)$ . The corresponding points-of-interest $(x^t, y^t, z^t)$ sampled from frame $t$ will satisfy the following condition:", + "bbox": [ + 496, + 479, + 890, + 781 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(x ^ {t} - p _ {x} + v _ {x} \\cdot \\Delta t\\right) ^ {2} + \\left(y ^ {t} - p _ {y} + v _ {y} \\cdot \\Delta t\\right) ^ {2} < \\left(\\frac {d ^ {t}}{2}\\right) ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 787, + 890, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Delta t = T - t$ is the time offset of frame $t$ and $d^t = \\sqrt{(w^2 + l^2)}\\cdot \\gamma^{\\Delta t + 1}$ is the diameter of cylindrical region.", + "bbox": [ + 500, + 824, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our preliminary study, we compare the overall recall rates of foreground points between our method and the trajectory-based method [2]. As can be seen in Tab. 1,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5198", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "our method can achieve very close result to the proposal trajectory method on 4-frame sequences. For 8-frame sequences, since it is difficult to estimate the positions of fast-moving objects on distant frames, we use $\\gamma = 1.1$ to recall more points and obtain comparable results to the proposal trajectory method. The recall rates only drop slightly even when 16-frame sequences are used. Interestingly, we find that setting $\\gamma = 1.1$ is not only beneficial for detecting fast-moving objects but also beneficial for improving the performance on slow and stationary objects. We believe this is because the points sampled from regions of different sizes contain multi-level contextual information about the objects, which will be further discussed in Sec. 4.3.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Region-based Network", + "text_level": 1, + "bbox": [ + 76, + 297, + 289, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposal feature encoding. After sampling $K$ points in each proposal, we adopt two encoding schemes to generate proposal features. We first follow [19] to calculate the relative offsets between each point-of-interest $l_i^t \\in I^t$ and the nine key-points (eight corner points plus one center point) of the proposal box $\\{b_j^t \\in P^t : j = 0, \\dots, 8\\}$ . The resulted offsets are then converted to spherical coordinates and transformed, via a Multi-layer Perceptron (MLP), to a geometric embedding $g^t \\in \\mathbb{R}^{K \\times D}$ that encodes the spatial correlation between the sampled point and the proposal box. The encoding scheme can be formulated as:", + "bbox": [ + 75, + 320, + 473, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} ^ {t} = \\operatorname {M L P} \\left(S \\left(\\left\\{l _ {i} ^ {t} - b _ {j} ^ {t} \\right\\} _ {j = 0} ^ {8}\\right)\\right), \\text {f o r} i = 1, \\dots , K, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 498, + 468, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S:(x,y,z)\\to (r,\\theta ,\\phi)$ denotes the spherical transform where $r = \\sqrt{x^2 + y^2 + z^2}$ , $\\theta = \\arcsin (z / r)$ and $\\phi = \\arctan (y / x)$ respectively. The second scheme produces a motion embedding $m^t\\in \\mathbb{R}^{K\\times D}$ , which encodes the displacements of the points-of-interest at each frame $I^t$ relative to the key-points of the proposal boxes $b^{0}$ in the first frame. The time offsets $\\Delta t$ are also concatenated, resulting in the motion embedding at frame $t$ as:", + "bbox": [ + 75, + 527, + 472, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} ^ {t} = \\operatorname {M L P} \\left(\\operatorname {C o n c a t} \\left(\\left\\{l _ {i} ^ {t} - b _ {j} ^ {0} \\right\\} _ {j = 0} ^ {8}, \\Delta t\\right)\\right), \\text {f o r} i = 1, \\dots , K. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 660, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposal feature $f^t \\in \\mathbb{R}^{K \\times D}$ can be formulated as the summation of geometric and motion embeddings:", + "bbox": [ + 75, + 691, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {t} = g ^ {t} + m ^ {t}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 734, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bidirectional feature aggregation. MSF employs a region-based network to explore spatial-temporal dependencies among proposal features and fuse them into global representations for final bounding-box refinement. As shown in Fig. 2, the region-based network is composed of three learning blocks. Each block consists of a traditional Multi-Head Self-Attention (MHSA) layer, followed by a Feed-Forward Network (FFN) with residual connection, and our proposed Bidirectional Feature Aggregation", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(BiFA) module. The MHSA layer aims to encode rich spatial relationships and point dependencies in the proposal for refining point features, while the proposed BiFA module aims to encode temporal information by facilitating information exchange between proposals across frames.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, BiFA involves a forward path and a backward path, representing two ways of information flow along the sequence. Since proposal features have an unordered representation, we leverage the Max-pool&Repeat [36] tensor manipulation to obtain a summarized contextual features. In the forward path, aside from the first frame in the sequence, which is concatenated with its own contextual features, each of the other frames is concatenated with the contextual features of its preceding frame along channel dimension. A point-wise convolutional layer is thereby employed to refine the concatenated feature and halve their channels. Given proposal features $f^t$ and $f^{t-1}$ from the current frame and its preceding frame, the forward output of frame $t$ , denote by $h_F^t$ , can be obtained by:", + "bbox": [ + 496, + 167, + 892, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {F} ^ {t} = \\operatorname {C o n v} \\left(\\operatorname {C o n c a t} \\left(f ^ {t}, \\text {R e p e a t} \\circ \\operatorname {M a x - p o o l} \\left(f ^ {t - 1}\\right)\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 383, + 892, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unfortunately, introducing the forward path only will lead to information imbalance among different frames, i.e., the current frame receives information from other frames, whereas the last preceding frame receives no information from the sequence. To overcome this limitation, we augment the backward path, where the features of frame $t$ are aggregated with the contextual features of frame $t + 1$ , resulting in the backward output $h_B^t$ as follows:", + "bbox": [ + 496, + 407, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {B} ^ {t} = \\operatorname {C o n v} \\left(\\operatorname {C o n c a t} \\left(h _ {F} ^ {t}, \\text {R e p e a t} \\circ \\operatorname {M a x - p o o l} \\left(h _ {F} ^ {t + 1}\\right)\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 536, + 890, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, each frame can simultaneously exchange information with its two adjacent frames from both directions, and the information can be propagated to more distant frames after three learning blocks. The resulted proposal features can capture long-term dependencies over the point cloud sequence. It is worthy noting that, the parameters of the convolutional layer in the same path can be shared and all the tensor operations can be performed in parallel. This makes our BiFA module lightweight and much more efficient than other multi-frame fusion modules.", + "bbox": [ + 496, + 561, + 892, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Outputs and loss function. Finally, we aggregate the point-wise features of each proposal into a global representation through a query-based transformer decoder layer, where a learnable query feature $q \\in \\mathbb{R}^{D}$ attends to each point vector of proposal features $h^t \\in \\mathbb{R}^{K \\times D}$ through a cross-attention module. The decoder layer can be depicted as follows:", + "bbox": [ + 496, + 712, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {e} ^ {t} = \\operatorname {A t t e n t i o n} \\left(q, h ^ {t}, h ^ {t}\\right) + h ^ {t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 823, + 890, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {t} = \\operatorname {F F N} \\left(\\hat {e} ^ {t}\\right) + \\hat {e} ^ {t}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 844, + 890, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The decoded outputs across frames $\\{e^t\\}_{t=0}^T$ are concatenated and passed to a group of detection heads for", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5199", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7d5bfb18b35dc3d2f27ac221e15715f6fbe352e33ef0c869c67ca19a376ca798.jpg", + "image_caption": [ + "Figure 3. Illustration of our optimized point cloud pooling method. We first perform intra-voxel sampling to keep a fixed number of points in each voxel. Then we query $n \\times n$ voxels fields for each proposal and uniformly draw points from the non-empty voxels within." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 885, + 234 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2038d39f89e82fe8918bdc191e9d0f409aed93499821c519243efa5781ae5ecd.jpg", + "table_caption": [ + "Table 2. The latency of point cloud pooling on 1-frame, 4-frames and 8-frames sequences." + ], + "table_footnote": [], + "table_body": "
N=168kN=674kN=1382k
Cylindrical Pooling8.2ms25.2ms40.1ms
Our Optimized2.3ms3.4ms5.0ms
", + "bbox": [ + 93, + 330, + 444, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "boundong-box refinement. The overall loss function $\\mathcal{L}_{\\mathrm{total}}$ is the summation of the confidence prediction loss $\\mathcal{L}_{\\mathrm{conf}}$ and the box regression loss $\\mathcal{L}_{\\mathrm{reg}}$ as follows:", + "bbox": [ + 75, + 391, + 468, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n f}} + \\alpha \\mathcal {L} _ {\\text {r e g}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 449, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a balancing hyper-parameter. We adopt the same binary cross entropy loss and box regression loss employed in CT3D [19] as our $\\mathcal{L}_{\\mathrm{conf}}$ and $\\mathcal{L}_{\\mathrm{reg}}$ .", + "bbox": [ + 75, + 476, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Efficient Point Cloud Pooling", + "text_level": 1, + "bbox": [ + 76, + 530, + 339, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this subsection, we optimize the point cloud pooling method to sample a fixed number of points more efficiently from the cylindrical region of each proposal. As shown in Fig. 3, we perform proposal-based point sampling in two steps, i.e., the intra-voxel sampling and the voxel-field sampling. In the first step, the input space is discretized into a voxel grid with voxel size $v$ , and each point corresponds to a voxel coordinates $(\\lfloor \\frac{x}{v} \\rfloor, \\lfloor \\frac{y}{v} \\rfloor)$ . Here the voxel partition in $z$ -axis is omitted considering that the cylindrical regions have unlimited height. Then, up to $k$ points are kept in each voxel and padding points are inserted if the voxel has less than $k$ points. Due to the high memory consumption required to store all voxels in a dense grid, we follow [14] to store only non-empty voxels in a contiguous memory and use a hash table to store the mappings between the coordinates of non-empty voxels and their indices in memory space.", + "bbox": [ + 75, + 553, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the second step, we first query $n$ -by- $n$ voxel-field for each proposal and calculate the coordinates of the voxels within. Next, we convert the coordinates of voxel into hashed keys and look up the table for querying the sampled points generated from the first hierarchy. The hash-table will return \"-1\" if no keys are found, which means it is an empty voxel. The queried points are then drawn from these", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "voxels and stored in an output buffer if the following conditions are met: 1) it is a valid point and 2) it falls into the cylindrical region, i.e. satisfying Eq. 1.", + "bbox": [ + 496, + 301, + 890, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Complexity and efficiency We perform an in-depth analysis of our optimized pooling method and the previous cylindrical pooling methods [2, 19, 30] in continuous space. Given $N$ points, $M$ proposals and the requirement of sampling $K$ points in each proposal, the original method costs $\\mathcal{O}(NM) + \\mathcal{O}(MK)$ complexity due to the calculation of pair-wise distances between points and proposals. In contrast, our optimized version costs $\\mathcal{O}(N) + \\mathcal{O}(M) + \\mathcal{O}(MK)$ complexity, where the first term is the cost of intra-voxel sampling, the second term is the cost of querying voxel-field for each proposal and the third term is the cost of drawing points from the queried voxels. We evaluate the latency of point cloud pooling on sequences with different lengths. As shown in Tab. 2, our optimized pooling method can achieve an $8\\times$ speedup over the original pooling method.", + "bbox": [ + 496, + 348, + 892, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 592, + 624, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 617, + 812, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We evaluate our MSF on Waymo Open Dataset (WOD), which is a large-scale multi-modality dataset for autonomous driving. WOD contains 1,150 sequences, which are divided into 798 training, 202 validation, and 150 testing sequences, respectively. Each sequence is 20s long, captured by a 64-line LiDAR sensor at $10\\mathrm{Hz}$ frequency. The evaluation metrics used in WOD are mean average precision (mAP) and mAP weighted by heading accuracy (mAPH). Three object classes, \"Vehicle\", \"Pedestrian\" and \"Cyclist\", are evaluated. Each object class is further categorized into two levels of difficulties, LEVEL1 and LEVEL2. The former refers to objects with more than 5 points and the latter refers to objects with less than 5 but at least 1 point.", + "bbox": [ + 496, + 642, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation. We employ the traditional Center-Point [28, 33, 36] as our region proposal network (RPN). To include motion information, we concatenate four adjacent frames at input and add one additional head for pre", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5200", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/890ed456711b4a397346682fb3913f541f92eec4b3fd0faf19440cfebbc9e9bf.jpg", + "table_caption": [ + "Table 3. Performance comparison on the validation set of Waymo Open Dataset." + ], + "table_footnote": [], + "table_body": "
MethodFramesALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
SECOND [28]163.0557.2372.27/71.6963.85/63.3368.70/58.1860.72/51.3160.62/59.2858.34/57.05
PointPillar [9]163.3357.5371.60/71.0063.10/62.5070.60/56.7062.90/50.2064.40/62.3061.90/59.90
IA-SSD [35]164.4858.0870.53/69.6761.55/60.8069.38/58.4760.30/50.7367.67/65.3064.98/62.71
LiDAR R-CNN [10]166.2060.1073.50/73.0064.70/64.2071.20/58.7063.10/51.7068.60/66.9066.10/64.40
RSN [26]1--75.10/74.6066.00/65.5077.80/72.7068.30/63.70--
PV-RCNN [21]169.6363.3377.51/76.8968.98/68.4175.01/65.6566.04/57.6167.81/66.3565.39/63.98
Part-A2 [24]170.2563.8477.05/76.5168.47/67.9775.24/66.8766.18/58.6268.60/67.3666.13/64.93
Centerpoint [33]1-65.50--/66.20--/62.60--/67.60
VoTR [14]1--74.95/74.2565.91/65.29----
VoxSeT [6]172.2466.2274.50/74.0365.99/65.5680.03/72.4272.45/65.3971.56/70.2968.95/67.73
SST-1f [3]1--76.22/75.7968.04/67.6481.39/74.0572.82/65.93--
SWFormer-1f [25]1--77.8/77.369.2/68.880.9/72.772.5/64.9--
PillarNet [20]174.6068.4379.09/78.5970.92/70.4680.59/74.0172.28/66.1772.29/71.2169.72/68.67
PV-RCNN++ [22]175.2168.6179.10/78.6370.34/69.9180.62/74.6271.86/66.3073.49/72.3870.70/69.62
3D-MAN [31]16--74.53/74.0367.61/67.1471.7/67.762.6/59.0--
SST-3f [3]3--78.66/78.2169.98/69.5783.81/80.1475.94/72.37--
SWFormer-3f [25]3--79.4/78.971.1/70.682.9/79.074.8/71.1--
CenterFormer [37]477.073.278.1/77.673.4/72.981.7/78.677.2/74.275.6/74.873.4/72.6
CenterFormer [37]877.373.778.8/78.374.3/73.882.1/79.377.8/75.075.2/74.473.2/72.3
MPPNet [2]479.8374.2281.54/81.0674.07/73.6184.56/81.9477.20/74.6777.15/76.5075.01/74.38
MPPNet [2]1680.4074.8582.74/82.2875.41/74.9684.69/82.2577.43/75.0677.28/76.6675.13/74.52
MSF (ours)480.2074.6281.36/80.8773.81/73.3585.05/82.1077.92/75.1178.40/77.6176.17/75.40
MSF (ours)880.6575.4682.83/82.0175.76/75.3185.24/82.2178.32/75.6178.52/77.7476.32/75.47
", + "bbox": [ + 76, + 114, + 893, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dicting the velocity of objects. In our experiments, we first train RPN using the official settings of OpenPCDet $^1$ , and use it to generate proposals of WOD. Based on these proposals, we then train our region-based network for 6 epochs by using the ADAM optimizer with an initial learning rate of 0.003 and a batch size of 16. The learning rate is decayed with One-Cycle policy and the momentum is set between $[85\\%, 95\\%]$ . During the training, we sample region proposals with IoU>0.5 and conduct proposal augmentation following PointRCNN [23]. A number of 128 raw Li-DAR points are randomly sampled for each proposal. The voxel size $v$ and the points-per-voxel $k$ used in voxel-based sampling are set to 0.4 and 32, respectively. The feature dimension of each point in the learning block is set to 256. At the training stage, we use the intermediate supervision by adding loss to the output of each learning block and sum all the intermediate losses to train the model. At the test stage, we only use the bounding boxes and the confidence scores predicted from the last learning block.", + "bbox": [ + 75, + 476, + 472, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on WOD", + "text_level": 1, + "bbox": [ + 76, + 773, + 243, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Waymo Validation Set. Tab. 3 compares the performance of MSF and the current state-of-the-art methods. As can be seen, multi-frame detectors generally outperform single-frame methods. The best two multi-frame methods by far are MPPNet [2] based on proposal trajectory and", + "bbox": [ + 75, + 799, + 468, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CenterFormer [37] based on transformer model. Using the same number of frames, our MSF method significantly outperforms CenterFormer by $3.4\\%$ APH and $1.2\\%$ APH on LEVEL1 and LEVEL2, respectively. This demonstrates that fusing multiple frame feature at the region level is more effective than the fusion with convolutional features. Compared with MPPNet [2], which is also based on region-level fusion, our method outperforms it on almost all cases, except for the APH of Vehicle LEVEL1. It should be noted that our MSF method only use 8 frames, while MPPNet uses 16 frames. MSF uses different pooling sizes in different frames, therefore generating multi-level contextual features for each object. Specifically, MSF models with 4- and 8-frames achieve the mAPH of $74.62\\%$ and $75.13\\%$ , respectively, recording new state-of-the-arts. In addition, both CenterFormer and MPPNet extract features on each frame of the sequence with the need of a memory bank to store the intermediate results of preceding frames, while our method performs proposal generation only on the current frame, which is more practical to online inference.", + "bbox": [ + 496, + 476, + 893, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Waymo Test Set. In Tab. 4, we show the results of our 8-frame model by submitting the prediction result to the online server for evaluation. Our method outperforms all the previously published methods. In particular, the improvements on Vehicle and Pedestrian classes are significant, outperforming the best competitor, i.e., CenterFormer, by $1.91\\%$ APH and $1.89\\%$ APH on LEVEL2, respectively.", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/open-mmlab/OpenPCDet/", + "bbox": [ + 94, + 886, + 339, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5201", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7e9729a2a56844773312ceddd079113b8d2d33366a1218bf200d7cdecbcf44b2.jpg", + "table_caption": [ + "Table 4. Performance comparison on the test set of Waymo Open Dataset." + ], + "table_footnote": [], + "table_body": "
MethodALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
PointPillar [9]--68.1060.1068.00/55.5061.40/50.10--
StarNet [15]--61.0054.5067.80/59.9061.10/54.00--
M3DETR [5]67.161.977.7/77.170.5/70.068.2/58.560.6/52.067.3/65.765.3/63.8
3D-MAN [31]--78.2869.9869.97/65.9863.98/60.26--
PV-RCNN++ [22]75.770.281.6/81.273.9/73.580.4/75.074.1/69.071.9/70.869.3/68.2
CenterPoint [33]77.271.981.1/80.673.4/73.080.5/77.374.6/71.574.6/73.772.2/71.3
RSN [26]--80.3071.6078.90/75.6070.70/67.80--
SST-3f [3]78.372.881.0/80.673.1/72.783.3/79.776.9/73.575.7/74.673.2/72.2
MPPNet [2]80.5975.6784.27/83.8877.29/76.9184.12/81.5278.44/75.9377.11/76.3674.91/74.18
CenterFormer [37]80.9176.2985.36/84.9478.68/78.2885.22/ 82.4880.09/77.4276.21/75.3274.04/73.17
MSF (ours)81.7476.9686.07/85.6779.20/78.8285.99/83.1080.61/77.8277.29/76.4475.09/74.25
", + "bbox": [ + 78, + 114, + 895, + 321 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2fd076674a619af93cd622c1e62a2edaf9c9087df8253340d856a7fd6c13b41.jpg", + "table_caption": [ + "Table 5. Ablation experiments on the WOD validation set. \"ME\" refers to using Motion Embedding and \"SA\" refers to using Self-Attention modules. \"Q\" means using query-based decoder layer to generate global representation of proposal features and \"M\" means using single max-pooling layer for global feature generation. APH scores on LEVEL1 and LEVEL2 are reported." + ], + "table_footnote": [], + "table_body": "
MESABiFADecoderL1L2
✓✓✓Q80.2074.62
X✓✓Q80.03 (-0.17)74.50 (-0.12)
✓X✓Q76.51 (-3.69)71.91 (-2.71)
✓✓XQ78.25 (-1.95)73.11 (-1.51)
✓✓✓M79.54 (-0.66)74.08 (-0.54)
", + "bbox": [ + 78, + 441, + 468, + 532 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b47319960727f38a1a4f1e94b630524d397c8737f75875a692e1e14e65162fd6.jpg", + "table_caption": [ + "Table 6. Bidirectional feature aggregation vs. unidirectional feature aggregation." + ], + "table_footnote": [], + "table_body": "
ForwardBackwardL1L2
✓✓80.2074.62
✓×79.64 (-0.56)74.35 (-0.27)
×✓78.97 (-1.23)73.77 (-0.85)
", + "bbox": [ + 94, + 585, + 447, + 650 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5490bcc10e3be1fc93184d349ed145b276f094490e0287315e10885e142f3e8e.jpg", + "table_caption": [ + "Table 7. The performance by integrating the proxy points and MLP mixer from MPPNet. \"SA\" and \"Mixer\" denote the self-attention and MLP Mixer [2], respectively. The APH scores of LEVEL2 on three object classes are reported." + ], + "table_footnote": [], + "table_body": "
ConfigVehicle.Pedestrian.Cyclist
Raw + SA73.3575.1175.40
Proxy + SA73.12 (-0.23)74.20 (-0.91)74.32 (-1.08)
Proxy + Mixer73.45(+0.10)74.13 (-0.98)74.39 (-1.01)
", + "bbox": [ + 78, + 729, + 468, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 816, + 240, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct in-depth analysis on MSF by evaluating the effectiveness of each individual component of it. We report the APH metric of our 4-frame model on the WOD validation set. The ablation study results are shown", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Tab. 5 to Tab. 8, respectively.", + "bbox": [ + 498, + 349, + 710, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion embedding. As can be seen from the $1^{\\mathrm{st}}$ and $2^{\\mathrm{nd}}$ rows of Tab. 5, without motion encoding, the performance of our proposed MSF will drop by $0.17\\%$ on LEVEL1 and $0.12\\%$ in LEVEL2. This indicates that the motion information is beneficial to infer the object's geometry information in point cloud sequences.", + "bbox": [ + 496, + 366, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self-attention. As can be seen from the $1^{\\text{st}}$ and $3^{\\text{rd}}$ rows of Tab. 5, without using self-attention for spatial interactions, MSF will suffer from significant performance drop, $3.69\\%$ on LEVEL1 and $2.71\\%$ on LEVEL2. This indicates that the intra-frame interaction is vital for learning internal geometry information of the proposals.", + "bbox": [ + 496, + 458, + 890, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bidirectional feature aggregation. From the $1^{\\text{st}}$ and $4^{\\text{th}}$ rows of Tab. 5, we can find that enforcing temporal interactions among hidden features of proposals by using the proposed BiFA module can bring an improvement of $1.95\\%$ on LEVEL1 and $1.51\\%$ on LEVEL2, which demonstrates the benefits of exploring long-term dependencies in the point cloud sequence. We also conduct experiments by performing unidirectional feature aggregation with either forward or backward path. The results are shown in Tab. 6, from which we see that the unidirectional model obtains lower APH scores than that of the bidirectional model.", + "bbox": [ + 496, + 550, + 890, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query-based decoder layer. As shown in the last row of Tab. 5, by replacing the final query-based decoder layer with a single max-pooling layer, the performance will drop slightly by $0.66\\%$ and $0.54\\%$ . This is because the final proposal representation after decoder can be regarded as a weighted sum of all point features, and the decoding weights of different points can intrinsically introduce more dynamics for feature representation.", + "bbox": [ + 496, + 718, + 890, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial modeling with proxy points and MLP mixer. We perform an analysis by integrating the proxy points and MLP mixer developed in our best competitor MPPNet [2] into MSP. We first follow MPPNet to generate proxy point", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5202", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fad166e64ec156533aec0dfff0dacede22b04e399a4556d659cb2ca74aca14b6.jpg", + "table_caption": [ + "Table 8. The APH (L2) scores on objects with different velocities using the 8-frame model. The objects are categorized into stationary $(<0.2\\mathrm{m/s})$ , slow $(0.2\\sim1\\mathrm{m/s})$ , medium $(1\\sim6\\mathrm{m/s})$ and fast $(>6\\mathrm{m/s})$ groups." + ], + "table_footnote": [], + "table_body": "
StationarySlowMediumFast
γ=1.070.3270.3075.6080.32
γ=1.170.4371.1777.5682.48
γ=1.270.3571.2377.4482.31
", + "bbox": [ + 93, + 155, + 444, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of each proposal and apply a set abstraction [17] to aggregate the point-wise features to every proxy point. As shown in the $1^{\\mathrm{st}}$ and $2^{\\mathrm{nd}}$ rows of Tab. 7, using features of proxy points will degrade the performance on Vehicle and more significantly on Pedestrian and Cyclist classes. After replacing the self-attention module with the MLP mixer, the performance degradation on Pedestrian and Cyclist classes still exists, as shown in the $3^{\\mathrm{rd}}$ row of Tab. 7. This phenomenon is contradictory to the conclusion in MPPNet that using proxy points to formulate proposal features is more effective than using raw points. We believe this is because MPPNet applies per-frame detection, and thus the proposals across frames may have different sizes, while proxy points can provide consistent representations for different frames. In contrast, our method uses propagated proposals with the same size over the sequence, and hence our proposal features are naturally on the same scale. For small objects such as pedestrian and cyclist, the features from raw points can provide more fine-grained details than proxy points.", + "bbox": [ + 75, + 233, + 468, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of $\\gamma$ . We evaluate how $\\gamma$ affects the performance on the objects with different speeds. As shown in Tab. 8, gradually expanding the proposal region with $\\gamma = 1.1$ will bring substantial improvements on the moving objects. This demonstrates that our model is able to capture the information of fast moving objects, even in the distant frames. Interestingly, we also see improvements on slow and stationary objects. This is because the different pooling sizes could provide multi-level contextual information, which is beneficial for detecting the objects. As shown in the $3^{\\mathrm{rd}}$ row of Tab. 8, no further improvements can be made by increasing the value of $\\gamma$ to 1.2.", + "bbox": [ + 75, + 520, + 468, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Runtime Analysis", + "text_level": 1, + "bbox": [ + 76, + 710, + 250, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 illustrates the latency decomposition of different methods. Here the latency is computed as the average runtime over 100 samples that are randomly drawn from WOD validation set, which is measured by a single Nvidia GeForce RTX A6000 GPU. Since all the methods employ the same backbone network, we exclude the runtime of the backbone from the latency for better comparison. For MPPNet and CenterFormer, we assume that the features from the preceding frames are already available in the memory bank. As can be seen, MPPNet costs much time in the feature encoding stage because it performs set-abstraction [17] to", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/645401f9ba61a5b294e407b6e118c4ac15d7e17b14799607950fd91c1aa03ee6.jpg", + "image_caption": [ + "Figure 4. Comparison of the runtime of different methods." + ], + "image_footnote": [], + "bbox": [ + 498, + 89, + 890, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e596edc04c94dea7b0d5743ab0512942c363708efef1d4a1718add05ddd9e15.jpg", + "table_caption": [ + "Table 9. The runtime decomposition of MPPNet and MSF." + ], + "table_footnote": [], + "table_body": "
MPPNetMSF
MLP-Mixer 75 msCross-Attention 24 msSelf-Attention 36 msBiFA 12 ms
", + "bbox": [ + 501, + 306, + 890, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "group raw points on the proxy points, which requires manipulating point clouds in continuous space. In comparison, our method takes only 4 milliseconds to encode raw points as proposal features. The overall latency of MPP-Net, CenterFormer and MSF are 99, 80 and 50 milliseconds, respectively. Tab. 9 further decomposes the network latency regarding the spatial and temporal modules in MPP-Net and MSF. As can be seen, the self-attention module and BiFA module are much more efficient than MLP-Mixer and cross-attention module, respectively. Thanks to our optimized point cloud pooling, MSF achieves higher efficiency even compared with CenterFormer, which performs feature fusion directly on the convolutional features.", + "bbox": [ + 496, + 375, + 890, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 587, + 617, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a novel motion-guided sequential fusion method, namely MSF, for 3D object detection from point cloud sequence. Unlike previous multi-frame detectors that performed feature extraction on each frame of the sequence, MSF adopted a proposal propagation strategy to mine points-of-interest based on the proposals generated on the current frame, therefore reducing the redundant computations and relieving reliance on the preceding results. A bidirectional feature aggregation module was proposed to enable cross-frame interaction between proposal features, facilitating MSF to capture long-term dependencies over the sequence. Besides, we optimized the point cloud pooling process, allowing MSF to process large-scale point cloud sequences in milliseconds. The proposed MSF achieved state-of-the-art performance on the Waymo Open Datasets and it was more efficient than other multi-frame detectors. In future research, we plan to extend MSF to generate detection priors on the future frames to further reduce the overall computation of multi-frame detection.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5203", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1", + "[2] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2, 3, 5, 6, 7", + "[3] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8458-8468, June 2022. 2, 6, 7", + "[4] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and ZhaoXiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2918-2927, October 2021. 2", + "[5] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (CVPR), pages 772-782, 2022. 7", + "[6] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8417-8427, June 2022. 1, 6", + "[7] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11873-11882, 2020. 1, 2", + "[8] Rui Huang, Wanyue Zhang, Abhijit Kundu, Caroline Pantofoaru, David A Ross, Thomas Funkhouser, and Alireza Fathi. An lstm approach to temporal 3d object detection in lidar point clouds. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2020. 1, 2", + "[9] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 1, 2, 6, 7", + "[10] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6", + "[11] Zhidong Liang, Zehan Zhang, Ming Zhang, Xian Zhao, and Shiliang Pu. Rangeoudet: Range image based real-time 3d object detector optimized by intersection over union. In Proceedings of the IEEE/CVF Conference on Computer Vision" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Pattern Recognition (CVPR), pages 7140-7149, June 2021. 2", + "[12] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[13] Wenjie Luo, Bin Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3569-3577, 2018. 2", + "[14] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3164-3173, October 2021. 1, 5, 6", + "[15] Jiquan Ngiam, Benjamin Caine, Wei Han, Brandon Yang, Yuning Chai, Pei Sun, Yin Zhou, Xi Yi, Ouais Alsharif, Patrick Nguyen, et al. Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069, 2019. 7", + "[16] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3d object detection in point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2", + "[17] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 2, 8", + "[18] Charles R. Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6134-6144, June 2021. 1, 2, 3", + "[19] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2743-2752, October 2021. 2, 4, 5", + "[20] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: High-performance pillar-based 3d object detection. 2022. 6", + "[21] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2, 6", + "[22] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. arXiv preprint arXiv:2102.00463, 2021. 6, 7", + "[23] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincarnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-779, 2019. 2, 6", + "[24] Shaoshuai Shi, Zhe Wang, Xiaogang Wang, and Hongsheng Li. Part-a $^2$ net: 3d part-aware and aggregation neural net" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5204", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "work for object detection from point cloud. arXiv preprint arXiv:1907.03670, 2019. 6", + "[25] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. 2022. 6", + "[26] Pei Sun, Weiyue Wang, Yuning Chai, Gamaleldin El-sayed, Alex Bewley, Xiao Zhang, Cristian Sminchisescu, and Dragomir Anguelov. Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR, pages 5725-5734, 2021. 6, 7", + "[27] Dominic Zeng Wang and Ingmar Posner. Voting for voting in online point cloud object detection. Robotics: Science and Systems, 1(3):10-15607, 2015. 1", + "[28] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 2, 5, 6", + "[29] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2", + "[30] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE international conference on computer vision (ICCV), 2019. 2, 5", + "[31] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1863-1872, June 2021. 1, 2, 6, 7", + "[32] Junbo Yin, Jianbing Shen, Chenye Guan, Dingfu Zhou, and Ruigang Yang. Lidar-based online 3d video object detection with graph-based message passing and spatiotemporal transformer attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2", + "[33] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11784-11793, June 2021. 2, 3, 5, 6, 7", + "[34] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18953-18962, June 2022. 2", + "[35] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18953-18962, 2022. 6", + "[36] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2018. 1, 2, 4, 5" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[37] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2022. 1, 6, 7", + "bbox": [ + 501, + 90, + 893, + 159 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5205", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_model.json b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_model.json new file mode 100644 index 0000000000000000000000000000000000000000..abb3778f4d98488c919cb22da358363a5518f3dc --- /dev/null +++ b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_model.json @@ -0,0 +1,1870 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.13, + 0.768, + 0.177 + ], + "angle": 0, + "content": "MSF: Motion-guided Sequential Fusion for Efficient 3D Object Detection from Point Cloud Sequences" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.204, + 0.761, + 0.24 + ], + "angle": 0, + "content": "Chenhang He* Ruihuang Li* Yabin Zhang Shuai Li Lei Zhang† The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.242, + 0.755, + 0.259 + ], + "angle": 0, + "content": "{csche, csrhli, csybhang, cssli, cslzhang}@comp.polyu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Point cloud sequences are commonly used to accurately detect 3D objects in applications such as autonomous driving. Current top-performing multi-frame detectors mostly follow a Detect-and-Fuse framework, which extracts features from each frame of the sequence and fuses them to detect the objects in the current frame. However, this inevitably leads to redundant computation since adjacent frames are highly correlated. In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, which exploits the continuity of object motion to mine useful sequential contexts for object detection in the current frame. We first generate 3D proposals on the current frame and propagate them to preceding frames based on the estimated velocities. The points-of-interest are then pooled from the sequence and encoded as proposal features. A novel Bidirectional Feature Aggregation (BiFA) module is further proposed to facilitate the interactions of proposal features across frames. Besides, we optimize the point cloud pooling by a voxel-based sampling technique so that millions of points can be processed in several milliseconds. The proposed MSF method achieves not only better efficiency than other multi-frame detectors but also leading accuracy, with \\(83.12\\%\\) and \\(78.30\\%\\) mAP on the LEVEL1 and LEVEL2 test sets of Waymo Open Dataset, respectively. Codes can be found at https://github.com/skyhehel23/MSF." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.47, + 0.866 + ], + "angle": 0, + "content": "3D object detection [1,2,6,7,9,14,21,27-29,36] is one of the key technologies in autonomous driving, which helps the vehicle to better understand the surrounding environment and make critical decisions in the downstream tasks. As an indispensable sensing device in autonomous driving systems, LiDAR collects 3D measurements of the scene in" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.29, + 0.747, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.289, + 0.889, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.49, + 0.895, + 0.561 + ], + "angle": 0, + "content": "Figure 1. (a) The \"Detect-and-Fuse\" framework extracts features from each frame of the sequence and then fuses them, while (b) our proposed \"Motion-guided Sequential Fusion\" (MSF) method generates proposals on the current frame and propagates them to preceding frames to explore useful contexts in the sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.719 + ], + "angle": 0, + "content": "the form of point clouds. However, LiDAR can only produce partial view of the scene at a time, and the sparse and incomplete representation of point clouds brings considerable challenges to the 3D object detection task. In practice, the LiDAR sensor will continuously sense the environment and produce a sequence of point cloud frames over time. The multi-frame data can provide a denser representation of the scene as the vehicle moves. Therefore, how to fuse these multi-frame point cloud data for more accurate object detection is worth deep investigation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Recent works mainly focus on deep feature fusion with multi-frame point clouds, for example, aggregating dense birds-eye-view features via Transformer models [31, 37], passing the voxel features to LSTM [8] or GRU [32] modules for temporal modeling. Some top-performing detectors [2, 18] focus on fusing proposal features, where a tracker is employed to associate the 3D proposals across frames, and a region-based network is applied to refine the current proposals by incorporating contextual features from the proposal trajectories. These approaches generally follow a \"Detect-and-Fuse\" framework, as shown in Fig. 1(a), where the model requires to process each frame of the sequence," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5196" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "and the predictions on the current frame rely on the results of preceding frames. Since online detection is a causal system, such a detection framework might cause significant delay if the network is still processing a preceding frame when the current frame is loaded." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.499 + ], + "angle": 0, + "content": "In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, as shown in Fig. 1(b), which leverages the continuity of object motion to extract useful contexts from point cloud sequences and improve the detection of current frame. Specifically, considering that the motions of objects are relatively smooth in a short sequence, we propagate the proposals generated on current frame to preceding frames based on the velocities of objects, and sample reliable points-of-interest from the sequence. In this way, we bypass extracting features on each frame of the sequence, which reduces the redundant computation and reliance on the results of preceding frames. The sampled points are then transformed to proposal features via two encoding schemes and passed to a region-based network for further refinement. Specifically, a self-attention module is employed to enhance the interaction of point features within proposals, while a novel Bidirectional Feature Aggregation (BiFA) module is proposed to enforce the information exchange between proposals across frames. The refined proposal features consequently capture both spatial details and long-term dependencies over the sequence, leading to more accurate bounding-box prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.664 + ], + "angle": 0, + "content": "It is found that the existing point cloud pooling methods [2, 19, 23, 30] are inefficient, taking more than 40 milliseconds when processing millions of points from sequential point clouds. We find that the major bottleneck lies in the heavy computation of pair-wise distances between \\( n \\) points and \\( m \\) proposals, which costs \\( \\mathcal{O}(nm) \\) complexity. To further improve the efficiency, we optimize the point cloud pooling with a voxel sampling technique. The improved pooling operation is of linear complexity and can process millions of points in several milliseconds, more than eight times faster than the original method." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.666, + 0.468, + 0.68 + ], + "angle": 0, + "content": "Overall, our contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.687, + 0.468, + 0.745 + ], + "angle": 0, + "content": "- An efficient Motion-guided Sequential Fusion (MSF) method is proposed to fuse multi-frame point clouds at region level by propagating the proposals of current frame to preceding frames based on the object motions." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.752, + 0.468, + 0.795 + ], + "angle": 0, + "content": "- A novel Bidirectional Feature Aggregation (BiFA) module is introduced to facilitate the interactions of proposal features across frames." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.805, + 0.468, + 0.849 + ], + "angle": 0, + "content": "- The point cloud pooling method is optimized with a voxel-based sampling technique, significantly reducing the runtime on large-scale point cloud sequence." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.687, + 0.468, + 0.849 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The proposed MSF method is validated on the challenging Waymo Open Dataset, and it achieves leading accuracy on the LEVEL1 and LEVEL2 test sets with fast speed." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.438 + ], + "angle": 0, + "content": "Single-frame 3D object detection. Recent research on single-frame 3D object detection is mainly focused on representation learning on point clouds. Voxel-based detectors [28, 33, 36] rasterize the point cloud into volumetric representation, followed by 3D CNN to extract dense features. Some works convert point clouds into 2D birds-eye-view [9] or range view [4, 11] representations, and process them with more efficient 2D CNN. Following PointNet++ [17], point-based methods [16, 23, 29, 30, 34] directly process point clouds in continuous space, and extract highly-semantic features through a series of downsampling and set abstraction layers. Voxel-point approaches [7, 12, 21] employ a hybrid representation, where the flexible conversion between voxel-based and point-based representations are explored, leading to better balance between efficiency and performance. Our method employs a high-quality voxel-based detector CenterPoint [33] as the proposal generation network to predict 3D proposals of current frame and their motions. We then employ an efficient region-based network to further refine these proposals by mining sequential points from point cloud sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.806 + ], + "angle": 0, + "content": "3D object detection from point cloud sequence. Multi-frame point clouds provide richer 3D information of the environment. While some single-frame detectors [3, 33] can be adapted to point cloud sequence by simply concatenating multi-frame point cloud as the input, the improvements are typically marginal and the performance can be even worse when encountering moving objects. Fast-and-furious [13] explores an intermediate fusion to align multi-frame point cloud by concatenating the hidden feature maps of the backbone network. However, it still suffers from the misalignment brought by the fast-moving objects in long sequence. Recent approaches [8, 32] demonstrate that an in-depth fusion can be achieved with recurrent networks. Unfortunately, the use of a single memory to store and update features across frames builds a potential bottleneck. To resolve such limitations, 3D-MAN [31] first attempts to employ the attention mechanism to align different views of 3D objects and then exploits a memory bank to store and aggregate multi-frame features for long sequence. Recently, Offboard3D [18] and MPPNet [2] improve much the detection performance, where they associate the detected boxes from each frame of the sequence as proposal trajectories, and extract high-quality proposal features by sampling sequential point cloud on the trajectories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Our MSF method also samples points from the sequence, but it differs from those methods with proposal trajectories [2, 18] in that we only generate proposals on the current frame and propagate them to explore features in preceding frames. This makes our method much more efficient and favorable to online detection systems." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5197" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.086, + 0.48, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.087, + 0.891, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.893, + 0.455 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our proposed Motion-guided Sequential Fusion (MSF) approach. By taking a point cloud sequence as input, MSF employs a region proposal network to generate proposals on the current frame and sample points-of-interest from the sequence by using motion-guided sequential pooling. The sampled points are encoded as high-dimensional proposal features and passed to a region-based network, where three learning blocks are consequently applied to refine the proposal features. A Bidirectional Feature Aggregation (BiFA) module is introduced in the region-based network to facilitate the interactions of proposal features across frames. The red and blue cubes represent single-point features from the current frame and preceding frame, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.478, + 0.471, + 0.534 + ], + "angle": 0, + "content": "Table 1. Recall rates of foreground points by using per-frame detection based proposal trajectory method [2] and our motion guided proposal generation method. We employ the CenterPoint [33] as proposal generator and evaluate on Waymo validation split." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.544, + 0.449, + 0.615 + ], + "angle": 0, + "content": "
4-frame8-frame16-frame
Trajectory [2]93.2%92.8%90.5%
Ours (γ = 1.0)92.3%87.5%78.3%
Ours (γ = 1.1)93.5%91.7%87.3%
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.623, + 0.382, + 0.641 + ], + "angle": 0, + "content": "3. Motion-guided Sequential Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.817 + ], + "angle": 0, + "content": "This section presents our Motion-guided Sequential Fusion (MSF) approach for efficient 3D object detection on point cloud sequences. The overall architecture of MSF is illustrated in Fig. 2. In Sec. 3.1, we describe the details of motion-guided sequential pooling, which effectively mines reliable sequential points-of-interest based on the proposals of current frame. In Sec. 3.2, we present the region-based network, including the formulation of proposal features and a novel bidirectional feature aggregation module. In Sec. 3.3, we demonstrate a voxel-based sampling technique to accelerate the current point cloud pooling method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.376, + 0.848 + ], + "angle": 0, + "content": "3.1. Motion-guided Sequential Pooling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Current multi-frame detection methods [2,18] mostly explore proposal trajectories to generate high-quality point cloud representations. However, such a scheme relies on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.782 + ], + "angle": 0, + "content": "frame-by-frame proposal generation, which is not suitable for online detection systems. We observe that in a point cloud sequence, although objects move at different speeds, their motions are relatively smooth. That is to say, we can estimate their motion displacements and roughly localize their positions in preceding frames. To this end, given a point cloud sequence \\(\\{I^t\\}_{t=1}^T\\), we propose to propagate the proposals generated on the current frame \\(I^T\\) to preceding frames \\(\\{I^t\\}_{t=1}^{T-1}\\) based on their estimated velocities. Since moving objects may slightly deviate from the estimated positions in the preceding frames, we sample the points-of-interest in a cylindrical region of each proposal and gradually increase the diameter of the region by a factor \\(\\gamma\\) as the proposal propagates. Let's denote a proposal of current frame as \\((p_x, p_y, p_z, w, l, h, \\theta)\\), where \\((p_x, p_y, p_z)\\) denotes its center location, \\(w, l, h\\) and \\(\\theta\\) denote its width, length, height and yaw angle, respectively. Suppose that the object has a unit-time velocity \\(\\vec{v} = (v_x, v_y)\\). The corresponding points-of-interest \\((x^t, y^t, z^t)\\) sampled from frame \\(t\\) will satisfy the following condition:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.789, + 0.892, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\left(x ^ {t} - p _ {x} + v _ {x} \\cdot \\Delta t\\right) ^ {2} + \\left(y ^ {t} - p _ {y} + v _ {y} \\cdot \\Delta t\\right) ^ {2} < \\left(\\frac {d ^ {t}}{2}\\right) ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.825, + 0.892, + 0.856 + ], + "angle": 0, + "content": "where \\(\\Delta t = T - t\\) is the time offset of frame \\(t\\) and \\(d^t = \\sqrt{(w^2 + l^2)}\\cdot \\gamma^{\\Delta t + 1}\\) is the diameter of cylindrical region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our preliminary study, we compare the overall recall rates of foreground points between our method and the trajectory-based method [2]. As can be seen in Tab. 1," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "our method can achieve very close result to the proposal trajectory method on 4-frame sequences. For 8-frame sequences, since it is difficult to estimate the positions of fast-moving objects on distant frames, we use \\(\\gamma = 1.1\\) to recall more points and obtain comparable results to the proposal trajectory method. The recall rates only drop slightly even when 16-frame sequences are used. Interestingly, we find that setting \\(\\gamma = 1.1\\) is not only beneficial for detecting fast-moving objects but also beneficial for improving the performance on slow and stationary objects. We believe this is because the points sampled from regions of different sizes contain multi-level contextual information about the objects, which will be further discussed in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.298, + 0.29, + 0.314 + ], + "angle": 0, + "content": "3.2. Region-based Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.474, + 0.488 + ], + "angle": 0, + "content": "Proposal feature encoding. After sampling \\(K\\) points in each proposal, we adopt two encoding schemes to generate proposal features. We first follow [19] to calculate the relative offsets between each point-of-interest \\(l_i^t \\in I^t\\) and the nine key-points (eight corner points plus one center point) of the proposal box \\(\\{b_j^t \\in P^t : j = 0, \\dots, 8\\}\\). The resulted offsets are then converted to spherical coordinates and transformed, via a Multi-layer Perceptron (MLP), to a geometric embedding \\(g^t \\in \\mathbb{R}^{K \\times D}\\) that encodes the spatial correlation between the sampled point and the proposal box. The encoding scheme can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.499, + 0.47, + 0.518 + ], + "angle": 0, + "content": "\\[\ng _ {i} ^ {t} = \\operatorname {M L P} \\left(S \\left(\\left\\{l _ {i} ^ {t} - b _ {j} ^ {t} \\right\\} _ {j = 0} ^ {8}\\right)\\right), \\text {f o r} i = 1, \\dots , K, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.473, + 0.651 + ], + "angle": 0, + "content": "where \\(S:(x,y,z)\\to (r,\\theta ,\\phi)\\) denotes the spherical transform where \\(r = \\sqrt{x^2 + y^2 + z^2}\\), \\(\\theta = \\arcsin (z / r)\\) and \\(\\phi = \\arctan (y / x)\\) respectively. The second scheme produces a motion embedding \\(m^t\\in \\mathbb{R}^{K\\times D}\\), which encodes the displacements of the points-of-interest at each frame \\(I^t\\) relative to the key-points of the proposal boxes \\(b^{0}\\) in the first frame. The time offsets \\(\\Delta t\\) are also concatenated, resulting in the motion embedding at frame \\(t\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.661, + 0.469, + 0.693 + ], + "angle": 0, + "content": "\\[\nm _ {i} ^ {t} = \\operatorname {M L P} \\left(\\operatorname {C o n c a t} \\left(\\left\\{l _ {i} ^ {t} - b _ {j} ^ {0} \\right\\} _ {j = 0} ^ {8}, \\Delta t\\right)\\right), \\text {f o r} i = 1, \\dots , K. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.47, + 0.724 + ], + "angle": 0, + "content": "The proposal feature \\( f^t \\in \\mathbb{R}^{K \\times D} \\) can be formulated as the summation of geometric and motion embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.735, + 0.47, + 0.754 + ], + "angle": 0, + "content": "\\[\nf ^ {t} = g ^ {t} + m ^ {t}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Bidirectional feature aggregation. MSF employs a region-based network to explore spatial-temporal dependencies among proposal features and fuse them into global representations for final bounding-box refinement. As shown in Fig. 2, the region-based network is composed of three learning blocks. Each block consists of a traditional Multi-Head Self-Attention (MHSA) layer, followed by a Feed-Forward Network (FFN) with residual connection, and our proposed Bidirectional Feature Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "(BiFA) module. The MHSA layer aims to encode rich spatial relationships and point dependencies in the proposal for refining point features, while the proposed BiFA module aims to encode temporal information by facilitating information exchange between proposals across frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.38 + ], + "angle": 0, + "content": "Specifically, BiFA involves a forward path and a backward path, representing two ways of information flow along the sequence. Since proposal features have an unordered representation, we leverage the Max-pool&Repeat [36] tensor manipulation to obtain a summarized contextual features. In the forward path, aside from the first frame in the sequence, which is concatenated with its own contextual features, each of the other frames is concatenated with the contextual features of its preceding frame along channel dimension. A point-wise convolutional layer is thereby employed to refine the concatenated feature and halve their channels. Given proposal features \\( f^t \\) and \\( f^{t-1} \\) from the current frame and its preceding frame, the forward output of frame \\( t \\), denote by \\( h_F^t \\), can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.385, + 0.893, + 0.404 + ], + "angle": 0, + "content": "\\[\nh _ {F} ^ {t} = \\operatorname {C o n v} \\left(\\operatorname {C o n c a t} \\left(f ^ {t}, \\text {R e p e a t} \\circ \\operatorname {M a x - p o o l} \\left(f ^ {t - 1}\\right)\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Unfortunately, introducing the forward path only will lead to information imbalance among different frames, i.e., the current frame receives information from other frames, whereas the last preceding frame receives no information from the sequence. To overcome this limitation, we augment the backward path, where the features of frame \\( t \\) are aggregated with the contextual features of frame \\( t + 1 \\), resulting in the backward output \\( h_B^t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.537, + 0.892, + 0.556 + ], + "angle": 0, + "content": "\\[\nh _ {B} ^ {t} = \\operatorname {C o n v} \\left(\\operatorname {C o n c a t} \\left(h _ {F} ^ {t}, \\text {R e p e a t} \\circ \\operatorname {M a x - p o o l} \\left(h _ {F} ^ {t + 1}\\right)\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.893, + 0.711 + ], + "angle": 0, + "content": "In this way, each frame can simultaneously exchange information with its two adjacent frames from both directions, and the information can be propagated to more distant frames after three learning blocks. The resulted proposal features can capture long-term dependencies over the point cloud sequence. It is worthy noting that, the parameters of the convolutional layer in the same path can be shared and all the tensor operations can be performed in parallel. This makes our BiFA module lightweight and much more efficient than other multi-frame fusion modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Outputs and loss function. Finally, we aggregate the point-wise features of each proposal into a global representation through a query-based transformer decoder layer, where a learnable query feature \\( q \\in \\mathbb{R}^{D} \\) attends to each point vector of proposal features \\( h^t \\in \\mathbb{R}^{K \\times D} \\) through a cross-attention module. The decoder layer can be depicted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.824, + 0.891, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\hat {e} ^ {t} = \\operatorname {A t t e n t i o n} \\left(q, h ^ {t}, h ^ {t}\\right) + h ^ {t}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.845, + 0.891, + 0.863 + ], + "angle": 0, + "content": "\\[\ne ^ {t} = \\operatorname {F F N} \\left(\\hat {e} ^ {t}\\right) + \\hat {e} ^ {t}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.87, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The decoded outputs across frames \\(\\{e^t\\}_{t=0}^T\\) are concatenated and passed to a group of detection heads for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5199" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.887, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.247, + 0.895, + 0.278 + ], + "angle": 0, + "content": "Figure 3. Illustration of our optimized point cloud pooling method. We first perform intra-voxel sampling to keep a fixed number of points in each voxel. Then we query \\( n \\times n \\) voxels fields for each proposal and uniformly draw points from the non-empty voxels within." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.3, + 0.471, + 0.329 + ], + "angle": 0, + "content": "Table 2. The latency of point cloud pooling on 1-frame, 4-frames and 8-frames sequences." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.332, + 0.446, + 0.379 + ], + "angle": 0, + "content": "
N=168kN=674kN=1382k
Cylindrical Pooling8.2ms25.2ms40.1ms
Our Optimized2.3ms3.4ms5.0ms
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.47, + 0.44 + ], + "angle": 0, + "content": "boundong-box refinement. The overall loss function \\(\\mathcal{L}_{\\mathrm{total}}\\) is the summation of the confidence prediction loss \\(\\mathcal{L}_{\\mathrm{conf}}\\) and the box regression loss \\(\\mathcal{L}_{\\mathrm{reg}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.45, + 0.47, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n f}} + \\alpha \\mathcal {L} _ {\\text {r e g}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.524 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a balancing hyper-parameter. We adopt the same binary cross entropy loss and box regression loss employed in CT3D [19] as our \\(\\mathcal{L}_{\\mathrm{conf}}\\) and \\(\\mathcal{L}_{\\mathrm{reg}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.34, + 0.548 + ], + "angle": 0, + "content": "3.3. Efficient Point Cloud Pooling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.795 + ], + "angle": 0, + "content": "In this subsection, we optimize the point cloud pooling method to sample a fixed number of points more efficiently from the cylindrical region of each proposal. As shown in Fig. 3, we perform proposal-based point sampling in two steps, i.e., the intra-voxel sampling and the voxel-field sampling. In the first step, the input space is discretized into a voxel grid with voxel size \\( v \\), and each point corresponds to a voxel coordinates \\( (\\lfloor \\frac{x}{v} \\rfloor, \\lfloor \\frac{y}{v} \\rfloor) \\). Here the voxel partition in \\( z \\)-axis is omitted considering that the cylindrical regions have unlimited height. Then, up to \\( k \\) points are kept in each voxel and padding points are inserted if the voxel has less than \\( k \\) points. Due to the high memory consumption required to store all voxels in a dense grid, we follow [14] to store only non-empty voxels in a contiguous memory and use a hash table to store the mappings between the coordinates of non-empty voxels and their indices in memory space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In the second step, we first query \\( n \\)-by-\\( n \\) voxel-field for each proposal and calculate the coordinates of the voxels within. Next, we convert the coordinates of voxel into hashed keys and look up the table for querying the sampled points generated from the first hierarchy. The hash-table will return \"-1\" if no keys are found, which means it is an empty voxel. The queried points are then drawn from these" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.892, + 0.348 + ], + "angle": 0, + "content": "voxels and stored in an output buffer if the following conditions are met: 1) it is a valid point and 2) it falls into the cylindrical region, i.e. satisfying Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.893, + 0.576 + ], + "angle": 0, + "content": "Complexity and efficiency We perform an in-depth analysis of our optimized pooling method and the previous cylindrical pooling methods [2, 19, 30] in continuous space. Given \\(N\\) points, \\(M\\) proposals and the requirement of sampling \\(K\\) points in each proposal, the original method costs \\(\\mathcal{O}(NM) + \\mathcal{O}(MK)\\) complexity due to the calculation of pair-wise distances between points and proposals. In contrast, our optimized version costs \\(\\mathcal{O}(N) + \\mathcal{O}(M) + \\mathcal{O}(MK)\\) complexity, where the first term is the cost of intra-voxel sampling, the second term is the cost of querying voxel-field for each proposal and the third term is the cost of drawing points from the queried voxels. We evaluate the latency of point cloud pooling on sequences with different lengths. As shown in Tab. 2, our optimized pooling method can achieve an \\(8\\times\\) speedup over the original pooling method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.593, + 0.625, + 0.61 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.813, + 0.635 + ], + "angle": 0, + "content": "4.1. Dataset and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Dataset. We evaluate our MSF on Waymo Open Dataset (WOD), which is a large-scale multi-modality dataset for autonomous driving. WOD contains 1,150 sequences, which are divided into 798 training, 202 validation, and 150 testing sequences, respectively. Each sequence is 20s long, captured by a 64-line LiDAR sensor at \\(10\\mathrm{Hz}\\) frequency. The evaluation metrics used in WOD are mean average precision (mAP) and mAP weighted by heading accuracy (mAPH). Three object classes, \"Vehicle\", \"Pedestrian\" and \"Cyclist\", are evaluated. Each object class is further categorized into two levels of difficulties, LEVEL1 and LEVEL2. The former refers to objects with more than 5 points and the latter refers to objects with less than 5 but at least 1 point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation. We employ the traditional Center-Point [28, 33, 36] as our region proposal network (RPN). To include motion information, we concatenate four adjacent frames at input and add one additional head for pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5200" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.246, + 0.09, + 0.723, + 0.104 + ], + "angle": 0, + "content": "Table 3. Performance comparison on the validation set of Waymo Open Dataset." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.116, + 0.895, + 0.46 + ], + "angle": 0, + "content": "
MethodFramesALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
SECOND [28]163.0557.2372.27/71.6963.85/63.3368.70/58.1860.72/51.3160.62/59.2858.34/57.05
PointPillar [9]163.3357.5371.60/71.0063.10/62.5070.60/56.7062.90/50.2064.40/62.3061.90/59.90
IA-SSD [35]164.4858.0870.53/69.6761.55/60.8069.38/58.4760.30/50.7367.67/65.3064.98/62.71
LiDAR R-CNN [10]166.2060.1073.50/73.0064.70/64.2071.20/58.7063.10/51.7068.60/66.9066.10/64.40
RSN [26]1--75.10/74.6066.00/65.5077.80/72.7068.30/63.70--
PV-RCNN [21]169.6363.3377.51/76.8968.98/68.4175.01/65.6566.04/57.6167.81/66.3565.39/63.98
Part-A2 [24]170.2563.8477.05/76.5168.47/67.9775.24/66.8766.18/58.6268.60/67.3666.13/64.93
Centerpoint [33]1-65.50--/66.20--/62.60--/67.60
VoTR [14]1--74.95/74.2565.91/65.29----
VoxSeT [6]172.2466.2274.50/74.0365.99/65.5680.03/72.4272.45/65.3971.56/70.2968.95/67.73
SST-1f [3]1--76.22/75.7968.04/67.6481.39/74.0572.82/65.93--
SWFormer-1f [25]1--77.8/77.369.2/68.880.9/72.772.5/64.9--
PillarNet [20]174.6068.4379.09/78.5970.92/70.4680.59/74.0172.28/66.1772.29/71.2169.72/68.67
PV-RCNN++ [22]175.2168.6179.10/78.6370.34/69.9180.62/74.6271.86/66.3073.49/72.3870.70/69.62
3D-MAN [31]16--74.53/74.0367.61/67.1471.7/67.762.6/59.0--
SST-3f [3]3--78.66/78.2169.98/69.5783.81/80.1475.94/72.37--
SWFormer-3f [25]3--79.4/78.971.1/70.682.9/79.074.8/71.1--
CenterFormer [37]477.073.278.1/77.673.4/72.981.7/78.677.2/74.275.6/74.873.4/72.6
CenterFormer [37]877.373.778.8/78.374.3/73.882.1/79.377.8/75.075.2/74.473.2/72.3
MPPNet [2]479.8374.2281.54/81.0674.07/73.6184.56/81.9477.20/74.6777.15/76.5075.01/74.38
MPPNet [2]1680.4074.8582.74/82.2875.41/74.9684.69/82.2577.43/75.0677.28/76.6675.13/74.52
MSF (ours)480.2074.6281.36/80.8773.81/73.3585.05/82.1077.92/75.1178.40/77.6176.17/75.40
MSF (ours)880.6575.4682.83/82.0175.76/75.3185.24/82.2178.32/75.6178.52/77.7476.32/75.47
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.473, + 0.765 + ], + "angle": 0, + "content": "dicting the velocity of objects. In our experiments, we first train RPN using the official settings of OpenPCDet\\(^1\\), and use it to generate proposals of WOD. Based on these proposals, we then train our region-based network for 6 epochs by using the ADAM optimizer with an initial learning rate of 0.003 and a batch size of 16. The learning rate is decayed with One-Cycle policy and the momentum is set between \\([85\\%, 95\\%]\\). During the training, we sample region proposals with IoU>0.5 and conduct proposal augmentation following PointRCNN [23]. A number of 128 raw Li-DAR points are randomly sampled for each proposal. The voxel size \\(v\\) and the points-per-voxel \\(k\\) used in voxel-based sampling are set to 0.4 and 32, respectively. The feature dimension of each point in the learning block is set to 256. At the training stage, we use the intermediate supervision by adding loss to the output of each learning block and sum all the intermediate losses to train the model. At the test stage, we only use the bounding boxes and the confidence scores predicted from the last learning block." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.775, + 0.245, + 0.79 + ], + "angle": 0, + "content": "4.2. Results on WOD" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.8, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Waymo Validation Set. Tab. 3 compares the performance of MSF and the current state-of-the-art methods. As can be seen, multi-frame detectors generally outperform single-frame methods. The best two multi-frame methods by far are MPPNet [2] based on proposal trajectory and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.895, + 0.78 + ], + "angle": 0, + "content": "CenterFormer [37] based on transformer model. Using the same number of frames, our MSF method significantly outperforms CenterFormer by \\(3.4\\%\\) APH and \\(1.2\\%\\) APH on LEVEL1 and LEVEL2, respectively. This demonstrates that fusing multiple frame feature at the region level is more effective than the fusion with convolutional features. Compared with MPPNet [2], which is also based on region-level fusion, our method outperforms it on almost all cases, except for the APH of Vehicle LEVEL1. It should be noted that our MSF method only use 8 frames, while MPPNet uses 16 frames. MSF uses different pooling sizes in different frames, therefore generating multi-level contextual features for each object. Specifically, MSF models with 4- and 8-frames achieve the mAPH of \\(74.62\\%\\) and \\(75.13\\%\\), respectively, recording new state-of-the-arts. In addition, both CenterFormer and MPPNet extract features on each frame of the sequence with the need of a memory bank to store the intermediate results of preceding frames, while our method performs proposal generation only on the current frame, which is more practical to online inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Waymo Test Set. In Tab. 4, we show the results of our 8-frame model by submitting the prediction result to the online server for evaluation. Our method outperforms all the previously published methods. In particular, the improvements on Vehicle and Pedestrian classes are significant, outperforming the best competitor, i.e., CenterFormer, by \\(1.91\\%\\) APH and \\(1.89\\%\\) APH on LEVEL2, respectively." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.341, + 0.901 + ], + "angle": 0, + "content": "1 https://github.com/open-mmlab/OpenPCDet/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5201" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.265, + 0.09, + 0.704, + 0.104 + ], + "angle": 0, + "content": "Table 4. Performance comparison on the test set of Waymo Open Dataset." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.116, + 0.896, + 0.323 + ], + "angle": 0, + "content": "
MethodALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
PointPillar [9]--68.1060.1068.00/55.5061.40/50.10--
StarNet [15]--61.0054.5067.80/59.9061.10/54.00--
M3DETR [5]67.161.977.7/77.170.5/70.068.2/58.560.6/52.067.3/65.765.3/63.8
3D-MAN [31]--78.2869.9869.97/65.9863.98/60.26--
PV-RCNN++ [22]75.770.281.6/81.273.9/73.580.4/75.074.1/69.071.9/70.869.3/68.2
CenterPoint [33]77.271.981.1/80.673.4/73.080.5/77.374.6/71.574.6/73.772.2/71.3
RSN [26]--80.3071.6078.90/75.6070.70/67.80--
SST-3f [3]78.372.881.0/80.673.1/72.783.3/79.776.9/73.575.7/74.673.2/72.2
MPPNet [2]80.5975.6784.27/83.8877.29/76.9184.12/81.5278.44/75.9377.11/76.3674.91/74.18
CenterFormer [37]80.9176.2985.36/84.9478.68/78.2885.22/ 82.4880.09/77.4276.21/75.3274.04/73.17
MSF (ours)81.7476.9686.07/85.6779.20/78.8285.99/83.1080.61/77.8277.29/76.4475.09/74.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.348, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Table 5. Ablation experiments on the WOD validation set. \"ME\" refers to using Motion Embedding and \"SA\" refers to using Self-Attention modules. \"Q\" means using query-based decoder layer to generate global representation of proposal features and \"M\" means using single max-pooling layer for global feature generation. APH scores on LEVEL1 and LEVEL2 are reported." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.443, + 0.47, + 0.533 + ], + "angle": 0, + "content": "
MESABiFADecoderL1L2
✓✓✓Q80.2074.62
X✓✓Q80.03 (-0.17)74.50 (-0.12)
✓X✓Q76.51 (-3.69)71.91 (-2.71)
✓✓XQ78.25 (-1.95)73.11 (-1.51)
✓✓✓M79.54 (-0.66)74.08 (-0.54)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.547, + 0.469, + 0.576 + ], + "angle": 0, + "content": "Table 6. Bidirectional feature aggregation vs. unidirectional feature aggregation." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.586, + 0.449, + 0.651 + ], + "angle": 0, + "content": "
ForwardBackwardL1L2
✓✓80.2074.62
✓×79.64 (-0.56)74.35 (-0.27)
×✓78.97 (-1.23)73.77 (-0.85)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.72 + ], + "angle": 0, + "content": "Table 7. The performance by integrating the proxy points and MLP mixer from MPPNet. \"SA\" and \"Mixer\" denote the self-attention and MLP Mixer [2], respectively. The APH scores of LEVEL2 on three object classes are reported." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.792 + ], + "angle": 0, + "content": "
ConfigVehicle.Pedestrian.Cyclist
Raw + SA73.3575.1175.40
Proxy + SA73.12 (-0.23)74.20 (-0.91)74.32 (-1.08)
Proxy + Mixer73.45(+0.10)74.13 (-0.98)74.39 (-1.01)
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.241, + 0.832 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct in-depth analysis on MSF by evaluating the effectiveness of each individual component of it. We report the APH metric of our 4-frame model on the WOD validation set. The ablation study results are shown" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.351, + 0.712, + 0.366 + ], + "angle": 0, + "content": "in Tab. 5 to Tab. 8, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Motion embedding. As can be seen from the \\(1^{\\mathrm{st}}\\) and \\(2^{\\mathrm{nd}}\\) rows of Tab. 5, without motion encoding, the performance of our proposed MSF will drop by \\(0.17\\%\\) on LEVEL1 and \\(0.12\\%\\) in LEVEL2. This indicates that the motion information is beneficial to infer the object's geometry information in point cloud sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Self-attention. As can be seen from the \\(1^{\\text{st}}\\) and \\(3^{\\text{rd}}\\) rows of Tab. 5, without using self-attention for spatial interactions, MSF will suffer from significant performance drop, \\(3.69\\%\\) on LEVEL1 and \\(2.71\\%\\) on LEVEL2. This indicates that the intra-frame interaction is vital for learning internal geometry information of the proposals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Bidirectional feature aggregation. From the \\(1^{\\text{st}}\\) and \\(4^{\\text{th}}\\) rows of Tab. 5, we can find that enforcing temporal interactions among hidden features of proposals by using the proposed BiFA module can bring an improvement of \\(1.95\\%\\) on LEVEL1 and \\(1.51\\%\\) on LEVEL2, which demonstrates the benefits of exploring long-term dependencies in the point cloud sequence. We also conduct experiments by performing unidirectional feature aggregation with either forward or backward path. The results are shown in Tab. 6, from which we see that the unidirectional model obtains lower APH scores than that of the bidirectional model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Query-based decoder layer. As shown in the last row of Tab. 5, by replacing the final query-based decoder layer with a single max-pooling layer, the performance will drop slightly by \\(0.66\\%\\) and \\(0.54\\%\\). This is because the final proposal representation after decoder can be regarded as a weighted sum of all point features, and the decoding weights of different points can intrinsically introduce more dynamics for feature representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Spatial modeling with proxy points and MLP mixer. We perform an analysis by integrating the proxy points and MLP mixer developed in our best competitor MPPNet [2] into MSP. We first follow MPPNet to generate proxy point" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5202" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 8. The APH (L2) scores on objects with different velocities using the 8-frame model. The objects are categorized into stationary \\((<0.2\\mathrm{m/s})\\), slow \\((0.2\\sim1\\mathrm{m/s})\\), medium \\((1\\sim6\\mathrm{m/s})\\) and fast \\((>6\\mathrm{m/s})\\) groups." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.156, + 0.446, + 0.226 + ], + "angle": 0, + "content": "
StationarySlowMediumFast
γ=1.070.3270.3075.6080.32
γ=1.170.4371.1777.5682.48
γ=1.270.3571.2377.4482.31
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.52 + ], + "angle": 0, + "content": "of each proposal and apply a set abstraction [17] to aggregate the point-wise features to every proxy point. As shown in the \\( 1^{\\mathrm{st}} \\) and \\( 2^{\\mathrm{nd}} \\) rows of Tab. 7, using features of proxy points will degrade the performance on Vehicle and more significantly on Pedestrian and Cyclist classes. After replacing the self-attention module with the MLP mixer, the performance degradation on Pedestrian and Cyclist classes still exists, as shown in the \\( 3^{\\mathrm{rd}} \\) row of Tab. 7. This phenomenon is contradictory to the conclusion in MPPNet that using proxy points to formulate proposal features is more effective than using raw points. We believe this is because MPPNet applies per-frame detection, and thus the proposals across frames may have different sizes, while proxy points can provide consistent representations for different frames. In contrast, our method uses propagated proposals with the same size over the sequence, and hence our proposal features are naturally on the same scale. For small objects such as pedestrian and cyclist, the features from raw points can provide more fine-grained details than proxy points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Effects of \\(\\gamma\\). We evaluate how \\(\\gamma\\) affects the performance on the objects with different speeds. As shown in Tab. 8, gradually expanding the proposal region with \\(\\gamma = 1.1\\) will bring substantial improvements on the moving objects. This demonstrates that our model is able to capture the information of fast moving objects, even in the distant frames. Interestingly, we also see improvements on slow and stationary objects. This is because the different pooling sizes could provide multi-level contextual information, which is beneficial for detecting the objects. As shown in the \\(3^{\\mathrm{rd}}\\) row of Tab. 8, no further improvements can be made by increasing the value of \\(\\gamma\\) to 1.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.25, + 0.727 + ], + "angle": 0, + "content": "4.4. Runtime Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Fig. 4 illustrates the latency decomposition of different methods. Here the latency is computed as the average runtime over 100 samples that are randomly drawn from WOD validation set, which is measured by a single Nvidia GeForce RTX A6000 GPU. Since all the methods employ the same backbone network, we exclude the runtime of the backbone from the latency for better comparison. For MPPNet and CenterFormer, we assume that the features from the preceding frames are already available in the memory bank. As can be seen, MPPNet costs much time in the feature encoding stage because it performs set-abstraction [17] to" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.09, + 0.891, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.259, + 0.868, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Comparison of the runtime of different methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.29, + 0.87, + 0.304 + ], + "angle": 0, + "content": "Table 9. The runtime decomposition of MPPNet and MSF." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.307, + 0.892, + 0.357 + ], + "angle": 0, + "content": "
MPPNetMSF
MLP-Mixer 75 msCross-Attention 24 msSelf-Attention 36 msBiFA 12 ms
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.571 + ], + "angle": 0, + "content": "group raw points on the proxy points, which requires manipulating point clouds in continuous space. In comparison, our method takes only 4 milliseconds to encode raw points as proposal features. The overall latency of MPP-Net, CenterFormer and MSF are 99, 80 and 50 milliseconds, respectively. Tab. 9 further decomposes the network latency regarding the spatial and temporal modules in MPP-Net and MSF. As can be seen, the self-attention module and BiFA module are much more efficient than MLP-Mixer and cross-attention module, respectively. Thanks to our optimized point cloud pooling, MSF achieves higher efficiency even compared with CenterFormer, which performs feature fusion directly on the convolutional features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.619, + 0.603 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We presented a novel motion-guided sequential fusion method, namely MSF, for 3D object detection from point cloud sequence. Unlike previous multi-frame detectors that performed feature extraction on each frame of the sequence, MSF adopted a proposal propagation strategy to mine points-of-interest based on the proposals generated on the current frame, therefore reducing the redundant computations and relieving reliance on the preceding results. A bidirectional feature aggregation module was proposed to enable cross-frame interaction between proposal features, facilitating MSF to capture long-term dependencies over the sequence. Besides, we optimized the point cloud pooling process, allowing MSF to process large-scale point cloud sequences in milliseconds. The proposed MSF achieved state-of-the-art performance on the Waymo Open Datasets and it was more efficient than other multi-frame detectors. In future research, we plan to extend MSF to generate detection priors on the future frames to further reduce the overall computation of multi-frame detection." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5203" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.472, + 0.258 + ], + "angle": 0, + "content": "[2] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[3] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8458-8468, June 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[4] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and ZhaoXiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2918-2927, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[5] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (CVPR), pages 772-782, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[6] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8417-8427, June 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[7] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11873-11882, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[8] Rui Huang, Wanyue Zhang, Abhijit Kundu, Caroline Pantofoaru, David A Ross, Thomas Funkhouser, and Alireza Fathi. An lstm approach to temporal 3d object detection in lidar point clouds. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.717, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[9] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.788, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[10] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[11] Zhidong Liang, Zehan Zhang, Ming Zhang, Xian Zhao, and Shiliang Pu. Rangeoudet: Range image based real-time 3d object detector optimized by intersection over union. In Proceedings of the IEEE/CVF Conference on Computer Vision" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "and Pattern Recognition (CVPR), pages 7140-7149, June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.164 + ], + "angle": 0, + "content": "[12] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[13] Wenjie Luo, Bin Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3569-3577, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.305 + ], + "angle": 0, + "content": "[14] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3164-3173, October 2021. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[15] Jiquan Ngiam, Benjamin Caine, Wei Han, Brandon Yang, Yuning Chai, Pei Sun, Yin Zhou, Xi Yi, Ouais Alsharif, Patrick Nguyen, et al. Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.894, + 0.433 + ], + "angle": 0, + "content": "[16] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3d object detection in point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.894, + 0.49 + ], + "angle": 0, + "content": "[17] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[18] Charles R. Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6134-6144, June 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.894, + 0.643 + ], + "angle": 0, + "content": "[19] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2743-2752, October 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.894, + 0.674 + ], + "angle": 0, + "content": "[20] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: High-performance pillar-based 3d object detection. 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[21] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[22] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. arXiv preprint arXiv:2102.00463, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.873 + ], + "angle": 0, + "content": "[23] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincarnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-779, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[24] Shaoshuai Shi, Zhe Wang, Xiaogang Wang, and Hongsheng Li. Part-a\\(^2\\) net: 3d part-aware and aggregation neural net" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5204" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "work for object detection from point cloud. arXiv preprint arXiv:1907.03670, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[25] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[26] Pei Sun, Weiyue Wang, Yuning Chai, Gamaleldin El-sayed, Alex Bewley, Xiao Zhang, Cristian Sminchisescu, and Dragomir Anguelov. Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR, pages 5725-5734, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[27] Dominic Zeng Wang and Ingmar Posner. Voting for voting in online point cloud object detection. Robotics: Science and Systems, 1(3):10-15607, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[28] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.345, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[29] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[30] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE international conference on computer vision (ICCV), 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[31] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1863-1872, June 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[32] Junbo Yin, Jianbing Shen, Chenye Guan, Dingfu Zhou, and Ruigang Yang. Lidar-based online 3d video object detection with graph-based message passing and spatiotemporal transformer attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.471, + 0.677 + ], + "angle": 0, + "content": "[33] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11784-11793, June 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[34] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18953-18962, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[35] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18953-18962, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[36] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2018. 1, 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[37] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2022. 1, 6, 7" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5205" + } + ] +] \ No newline at end of file diff --git a/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_origin.pdf b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5107028833a80c47fcdae6cf9983250bfa058ce --- /dev/null +++ b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/e0f58d96-cf61-42b5-9006-0e23915bae11_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:191e90a5512a93a08b5e11093a335d5b6fe30414f8ee6910fd88829ba960ba20 +size 1270023 diff --git a/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/full.md b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dd7e89fb9d8a4ae5340e3d882ee60787d79228d6 --- /dev/null +++ b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/full.md @@ -0,0 +1,272 @@ +# MSF: Motion-guided Sequential Fusion for Efficient 3D Object Detection from Point Cloud Sequences + +Chenhang He* Ruihuang Li* Yabin Zhang Shuai Li Lei Zhang† The Hong Kong Polytechnic University + +{csche, csrhli, csybhang, cssli, cslzhang}@comp.polyu.edu.hk + +# Abstract + +Point cloud sequences are commonly used to accurately detect 3D objects in applications such as autonomous driving. Current top-performing multi-frame detectors mostly follow a Detect-and-Fuse framework, which extracts features from each frame of the sequence and fuses them to detect the objects in the current frame. However, this inevitably leads to redundant computation since adjacent frames are highly correlated. In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, which exploits the continuity of object motion to mine useful sequential contexts for object detection in the current frame. We first generate 3D proposals on the current frame and propagate them to preceding frames based on the estimated velocities. The points-of-interest are then pooled from the sequence and encoded as proposal features. A novel Bidirectional Feature Aggregation (BiFA) module is further proposed to facilitate the interactions of proposal features across frames. Besides, we optimize the point cloud pooling by a voxel-based sampling technique so that millions of points can be processed in several milliseconds. The proposed MSF method achieves not only better efficiency than other multi-frame detectors but also leading accuracy, with $83.12\%$ and $78.30\%$ mAP on the LEVEL1 and LEVEL2 test sets of Waymo Open Dataset, respectively. Codes can be found at https://github.com/skyhehel23/MSF. + +# 1. Introduction + +3D object detection [1,2,6,7,9,14,21,27-29,36] is one of the key technologies in autonomous driving, which helps the vehicle to better understand the surrounding environment and make critical decisions in the downstream tasks. As an indispensable sensing device in autonomous driving systems, LiDAR collects 3D measurements of the scene in + +![](images/16385f200141b776c4a846848ac3ffa1b9ab7eeb7285b35ce2bb067b73d9d325.jpg) +Figure 1. (a) The "Detect-and-Fuse" framework extracts features from each frame of the sequence and then fuses them, while (b) our proposed "Motion-guided Sequential Fusion" (MSF) method generates proposals on the current frame and propagates them to preceding frames to explore useful contexts in the sequence. + +![](images/8ae0c5cec61df169bef6a172848d8978ddc21cf119d375f5cb6a2f001b3a1c5b.jpg) + +the form of point clouds. However, LiDAR can only produce partial view of the scene at a time, and the sparse and incomplete representation of point clouds brings considerable challenges to the 3D object detection task. In practice, the LiDAR sensor will continuously sense the environment and produce a sequence of point cloud frames over time. The multi-frame data can provide a denser representation of the scene as the vehicle moves. Therefore, how to fuse these multi-frame point cloud data for more accurate object detection is worth deep investigation. + +Recent works mainly focus on deep feature fusion with multi-frame point clouds, for example, aggregating dense birds-eye-view features via Transformer models [31, 37], passing the voxel features to LSTM [8] or GRU [32] modules for temporal modeling. Some top-performing detectors [2, 18] focus on fusing proposal features, where a tracker is employed to associate the 3D proposals across frames, and a region-based network is applied to refine the current proposals by incorporating contextual features from the proposal trajectories. These approaches generally follow a "Detect-and-Fuse" framework, as shown in Fig. 1(a), where the model requires to process each frame of the sequence, + +and the predictions on the current frame rely on the results of preceding frames. Since online detection is a causal system, such a detection framework might cause significant delay if the network is still processing a preceding frame when the current frame is loaded. + +In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, as shown in Fig. 1(b), which leverages the continuity of object motion to extract useful contexts from point cloud sequences and improve the detection of current frame. Specifically, considering that the motions of objects are relatively smooth in a short sequence, we propagate the proposals generated on current frame to preceding frames based on the velocities of objects, and sample reliable points-of-interest from the sequence. In this way, we bypass extracting features on each frame of the sequence, which reduces the redundant computation and reliance on the results of preceding frames. The sampled points are then transformed to proposal features via two encoding schemes and passed to a region-based network for further refinement. Specifically, a self-attention module is employed to enhance the interaction of point features within proposals, while a novel Bidirectional Feature Aggregation (BiFA) module is proposed to enforce the information exchange between proposals across frames. The refined proposal features consequently capture both spatial details and long-term dependencies over the sequence, leading to more accurate bounding-box prediction. + +It is found that the existing point cloud pooling methods [2, 19, 23, 30] are inefficient, taking more than 40 milliseconds when processing millions of points from sequential point clouds. We find that the major bottleneck lies in the heavy computation of pair-wise distances between $n$ points and $m$ proposals, which costs $\mathcal{O}(nm)$ complexity. To further improve the efficiency, we optimize the point cloud pooling with a voxel sampling technique. The improved pooling operation is of linear complexity and can process millions of points in several milliseconds, more than eight times faster than the original method. + +Overall, our contributions can be summarized as follows. + +- An efficient Motion-guided Sequential Fusion (MSF) method is proposed to fuse multi-frame point clouds at region level by propagating the proposals of current frame to preceding frames based on the object motions. +- A novel Bidirectional Feature Aggregation (BiFA) module is introduced to facilitate the interactions of proposal features across frames. +- The point cloud pooling method is optimized with a voxel-based sampling technique, significantly reducing the runtime on large-scale point cloud sequence. + +The proposed MSF method is validated on the challenging Waymo Open Dataset, and it achieves leading accuracy on the LEVEL1 and LEVEL2 test sets with fast speed. + +# 2. Related Work + +Single-frame 3D object detection. Recent research on single-frame 3D object detection is mainly focused on representation learning on point clouds. Voxel-based detectors [28, 33, 36] rasterize the point cloud into volumetric representation, followed by 3D CNN to extract dense features. Some works convert point clouds into 2D birds-eye-view [9] or range view [4, 11] representations, and process them with more efficient 2D CNN. Following PointNet++ [17], point-based methods [16, 23, 29, 30, 34] directly process point clouds in continuous space, and extract highly-semantic features through a series of downsampling and set abstraction layers. Voxel-point approaches [7, 12, 21] employ a hybrid representation, where the flexible conversion between voxel-based and point-based representations are explored, leading to better balance between efficiency and performance. Our method employs a high-quality voxel-based detector CenterPoint [33] as the proposal generation network to predict 3D proposals of current frame and their motions. We then employ an efficient region-based network to further refine these proposals by mining sequential points from point cloud sequence. + +3D object detection from point cloud sequence. Multi-frame point clouds provide richer 3D information of the environment. While some single-frame detectors [3, 33] can be adapted to point cloud sequence by simply concatenating multi-frame point cloud as the input, the improvements are typically marginal and the performance can be even worse when encountering moving objects. Fast-and-furious [13] explores an intermediate fusion to align multi-frame point cloud by concatenating the hidden feature maps of the backbone network. However, it still suffers from the misalignment brought by the fast-moving objects in long sequence. Recent approaches [8, 32] demonstrate that an in-depth fusion can be achieved with recurrent networks. Unfortunately, the use of a single memory to store and update features across frames builds a potential bottleneck. To resolve such limitations, 3D-MAN [31] first attempts to employ the attention mechanism to align different views of 3D objects and then exploits a memory bank to store and aggregate multi-frame features for long sequence. Recently, Offboard3D [18] and MPPNet [2] improve much the detection performance, where they associate the detected boxes from each frame of the sequence as proposal trajectories, and extract high-quality proposal features by sampling sequential point cloud on the trajectories. + +Our MSF method also samples points from the sequence, but it differs from those methods with proposal trajectories [2, 18] in that we only generate proposals on the current frame and propagate them to explore features in preceding frames. This makes our method much more efficient and favorable to online detection systems. + +![](images/4f2c2d8d78724a94d855046ec0cf9856ee549fcf4929dfbc796aa31204874c53.jpg) +Figure 2. The overall architecture of our proposed Motion-guided Sequential Fusion (MSF) approach. By taking a point cloud sequence as input, MSF employs a region proposal network to generate proposals on the current frame and sample points-of-interest from the sequence by using motion-guided sequential pooling. The sampled points are encoded as high-dimensional proposal features and passed to a region-based network, where three learning blocks are consequently applied to refine the proposal features. A Bidirectional Feature Aggregation (BiFA) module is introduced in the region-based network to facilitate the interactions of proposal features across frames. The red and blue cubes represent single-point features from the current frame and preceding frame, respectively. + +![](images/46252ebc4e81f5e1dc003f9b119be1c35fd45f78eefa90de399916842dc056cd.jpg) + +Table 1. Recall rates of foreground points by using per-frame detection based proposal trajectory method [2] and our motion guided proposal generation method. We employ the CenterPoint [33] as proposal generator and evaluate on Waymo validation split. + +
4-frame8-frame16-frame
Trajectory [2]93.2%92.8%90.5%
Ours (γ = 1.0)92.3%87.5%78.3%
Ours (γ = 1.1)93.5%91.7%87.3%
+ +# 3. Motion-guided Sequential Fusion + +This section presents our Motion-guided Sequential Fusion (MSF) approach for efficient 3D object detection on point cloud sequences. The overall architecture of MSF is illustrated in Fig. 2. In Sec. 3.1, we describe the details of motion-guided sequential pooling, which effectively mines reliable sequential points-of-interest based on the proposals of current frame. In Sec. 3.2, we present the region-based network, including the formulation of proposal features and a novel bidirectional feature aggregation module. In Sec. 3.3, we demonstrate a voxel-based sampling technique to accelerate the current point cloud pooling method. + +# 3.1. Motion-guided Sequential Pooling + +Current multi-frame detection methods [2,18] mostly explore proposal trajectories to generate high-quality point cloud representations. However, such a scheme relies on + +frame-by-frame proposal generation, which is not suitable for online detection systems. We observe that in a point cloud sequence, although objects move at different speeds, their motions are relatively smooth. That is to say, we can estimate their motion displacements and roughly localize their positions in preceding frames. To this end, given a point cloud sequence $\{I^t\}_{t=1}^T$ , we propose to propagate the proposals generated on the current frame $I^T$ to preceding frames $\{I^t\}_{t=1}^{T-1}$ based on their estimated velocities. Since moving objects may slightly deviate from the estimated positions in the preceding frames, we sample the points-of-interest in a cylindrical region of each proposal and gradually increase the diameter of the region by a factor $\gamma$ as the proposal propagates. Let's denote a proposal of current frame as $(p_x, p_y, p_z, w, l, h, \theta)$ , where $(p_x, p_y, p_z)$ denotes its center location, $w, l, h$ and $\theta$ denote its width, length, height and yaw angle, respectively. Suppose that the object has a unit-time velocity $\vec{v} = (v_x, v_y)$ . The corresponding points-of-interest $(x^t, y^t, z^t)$ sampled from frame $t$ will satisfy the following condition: + +$$ +\left(x ^ {t} - p _ {x} + v _ {x} \cdot \Delta t\right) ^ {2} + \left(y ^ {t} - p _ {y} + v _ {y} \cdot \Delta t\right) ^ {2} < \left(\frac {d ^ {t}}{2}\right) ^ {2}, \tag {1} +$$ + +where $\Delta t = T - t$ is the time offset of frame $t$ and $d^t = \sqrt{(w^2 + l^2)}\cdot \gamma^{\Delta t + 1}$ is the diameter of cylindrical region. + +In our preliminary study, we compare the overall recall rates of foreground points between our method and the trajectory-based method [2]. As can be seen in Tab. 1, + +our method can achieve very close result to the proposal trajectory method on 4-frame sequences. For 8-frame sequences, since it is difficult to estimate the positions of fast-moving objects on distant frames, we use $\gamma = 1.1$ to recall more points and obtain comparable results to the proposal trajectory method. The recall rates only drop slightly even when 16-frame sequences are used. Interestingly, we find that setting $\gamma = 1.1$ is not only beneficial for detecting fast-moving objects but also beneficial for improving the performance on slow and stationary objects. We believe this is because the points sampled from regions of different sizes contain multi-level contextual information about the objects, which will be further discussed in Sec. 4.3. + +# 3.2. Region-based Network + +Proposal feature encoding. After sampling $K$ points in each proposal, we adopt two encoding schemes to generate proposal features. We first follow [19] to calculate the relative offsets between each point-of-interest $l_i^t \in I^t$ and the nine key-points (eight corner points plus one center point) of the proposal box $\{b_j^t \in P^t : j = 0, \dots, 8\}$ . The resulted offsets are then converted to spherical coordinates and transformed, via a Multi-layer Perceptron (MLP), to a geometric embedding $g^t \in \mathbb{R}^{K \times D}$ that encodes the spatial correlation between the sampled point and the proposal box. The encoding scheme can be formulated as: + +$$ +g _ {i} ^ {t} = \operatorname {M L P} \left(S \left(\left\{l _ {i} ^ {t} - b _ {j} ^ {t} \right\} _ {j = 0} ^ {8}\right)\right), \text {f o r} i = 1, \dots , K, \tag {2} +$$ + +where $S:(x,y,z)\to (r,\theta ,\phi)$ denotes the spherical transform where $r = \sqrt{x^2 + y^2 + z^2}$ , $\theta = \arcsin (z / r)$ and $\phi = \arctan (y / x)$ respectively. The second scheme produces a motion embedding $m^t\in \mathbb{R}^{K\times D}$ , which encodes the displacements of the points-of-interest at each frame $I^t$ relative to the key-points of the proposal boxes $b^{0}$ in the first frame. The time offsets $\Delta t$ are also concatenated, resulting in the motion embedding at frame $t$ as: + +$$ +m _ {i} ^ {t} = \operatorname {M L P} \left(\operatorname {C o n c a t} \left(\left\{l _ {i} ^ {t} - b _ {j} ^ {0} \right\} _ {j = 0} ^ {8}, \Delta t\right)\right), \text {f o r} i = 1, \dots , K. \tag {3} +$$ + +The proposal feature $f^t \in \mathbb{R}^{K \times D}$ can be formulated as the summation of geometric and motion embeddings: + +$$ +f ^ {t} = g ^ {t} + m ^ {t}. \tag {4} +$$ + +Bidirectional feature aggregation. MSF employs a region-based network to explore spatial-temporal dependencies among proposal features and fuse them into global representations for final bounding-box refinement. As shown in Fig. 2, the region-based network is composed of three learning blocks. Each block consists of a traditional Multi-Head Self-Attention (MHSA) layer, followed by a Feed-Forward Network (FFN) with residual connection, and our proposed Bidirectional Feature Aggregation + +(BiFA) module. The MHSA layer aims to encode rich spatial relationships and point dependencies in the proposal for refining point features, while the proposed BiFA module aims to encode temporal information by facilitating information exchange between proposals across frames. + +Specifically, BiFA involves a forward path and a backward path, representing two ways of information flow along the sequence. Since proposal features have an unordered representation, we leverage the Max-pool&Repeat [36] tensor manipulation to obtain a summarized contextual features. In the forward path, aside from the first frame in the sequence, which is concatenated with its own contextual features, each of the other frames is concatenated with the contextual features of its preceding frame along channel dimension. A point-wise convolutional layer is thereby employed to refine the concatenated feature and halve their channels. Given proposal features $f^t$ and $f^{t-1}$ from the current frame and its preceding frame, the forward output of frame $t$ , denote by $h_F^t$ , can be obtained by: + +$$ +h _ {F} ^ {t} = \operatorname {C o n v} \left(\operatorname {C o n c a t} \left(f ^ {t}, \text {R e p e a t} \circ \operatorname {M a x - p o o l} \left(f ^ {t - 1}\right)\right)\right). \tag {5} +$$ + +Unfortunately, introducing the forward path only will lead to information imbalance among different frames, i.e., the current frame receives information from other frames, whereas the last preceding frame receives no information from the sequence. To overcome this limitation, we augment the backward path, where the features of frame $t$ are aggregated with the contextual features of frame $t + 1$ , resulting in the backward output $h_B^t$ as follows: + +$$ +h _ {B} ^ {t} = \operatorname {C o n v} \left(\operatorname {C o n c a t} \left(h _ {F} ^ {t}, \text {R e p e a t} \circ \operatorname {M a x - p o o l} \left(h _ {F} ^ {t + 1}\right)\right)\right). \tag {6} +$$ + +In this way, each frame can simultaneously exchange information with its two adjacent frames from both directions, and the information can be propagated to more distant frames after three learning blocks. The resulted proposal features can capture long-term dependencies over the point cloud sequence. It is worthy noting that, the parameters of the convolutional layer in the same path can be shared and all the tensor operations can be performed in parallel. This makes our BiFA module lightweight and much more efficient than other multi-frame fusion modules. + +Outputs and loss function. Finally, we aggregate the point-wise features of each proposal into a global representation through a query-based transformer decoder layer, where a learnable query feature $q \in \mathbb{R}^{D}$ attends to each point vector of proposal features $h^t \in \mathbb{R}^{K \times D}$ through a cross-attention module. The decoder layer can be depicted as follows: + +$$ +\hat {e} ^ {t} = \operatorname {A t t e n t i o n} \left(q, h ^ {t}, h ^ {t}\right) + h ^ {t}, \tag {7} +$$ + +$$ +e ^ {t} = \operatorname {F F N} \left(\hat {e} ^ {t}\right) + \hat {e} ^ {t}. \tag {8} +$$ + +The decoded outputs across frames $\{e^t\}_{t=0}^T$ are concatenated and passed to a group of detection heads for + +![](images/7d5bfb18b35dc3d2f27ac221e15715f6fbe352e33ef0c869c67ca19a376ca798.jpg) +Figure 3. Illustration of our optimized point cloud pooling method. We first perform intra-voxel sampling to keep a fixed number of points in each voxel. Then we query $n \times n$ voxels fields for each proposal and uniformly draw points from the non-empty voxels within. + +Table 2. The latency of point cloud pooling on 1-frame, 4-frames and 8-frames sequences. + +
N=168kN=674kN=1382k
Cylindrical Pooling8.2ms25.2ms40.1ms
Our Optimized2.3ms3.4ms5.0ms
+ +boundong-box refinement. The overall loss function $\mathcal{L}_{\mathrm{total}}$ is the summation of the confidence prediction loss $\mathcal{L}_{\mathrm{conf}}$ and the box regression loss $\mathcal{L}_{\mathrm{reg}}$ as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c o n f}} + \alpha \mathcal {L} _ {\text {r e g}}, \tag {9} +$$ + +where $\alpha$ is a balancing hyper-parameter. We adopt the same binary cross entropy loss and box regression loss employed in CT3D [19] as our $\mathcal{L}_{\mathrm{conf}}$ and $\mathcal{L}_{\mathrm{reg}}$ . + +# 3.3. Efficient Point Cloud Pooling + +In this subsection, we optimize the point cloud pooling method to sample a fixed number of points more efficiently from the cylindrical region of each proposal. As shown in Fig. 3, we perform proposal-based point sampling in two steps, i.e., the intra-voxel sampling and the voxel-field sampling. In the first step, the input space is discretized into a voxel grid with voxel size $v$ , and each point corresponds to a voxel coordinates $(\lfloor \frac{x}{v} \rfloor, \lfloor \frac{y}{v} \rfloor)$ . Here the voxel partition in $z$ -axis is omitted considering that the cylindrical regions have unlimited height. Then, up to $k$ points are kept in each voxel and padding points are inserted if the voxel has less than $k$ points. Due to the high memory consumption required to store all voxels in a dense grid, we follow [14] to store only non-empty voxels in a contiguous memory and use a hash table to store the mappings between the coordinates of non-empty voxels and their indices in memory space. + +In the second step, we first query $n$ -by- $n$ voxel-field for each proposal and calculate the coordinates of the voxels within. Next, we convert the coordinates of voxel into hashed keys and look up the table for querying the sampled points generated from the first hierarchy. The hash-table will return "-1" if no keys are found, which means it is an empty voxel. The queried points are then drawn from these + +voxels and stored in an output buffer if the following conditions are met: 1) it is a valid point and 2) it falls into the cylindrical region, i.e. satisfying Eq. 1. + +Complexity and efficiency We perform an in-depth analysis of our optimized pooling method and the previous cylindrical pooling methods [2, 19, 30] in continuous space. Given $N$ points, $M$ proposals and the requirement of sampling $K$ points in each proposal, the original method costs $\mathcal{O}(NM) + \mathcal{O}(MK)$ complexity due to the calculation of pair-wise distances between points and proposals. In contrast, our optimized version costs $\mathcal{O}(N) + \mathcal{O}(M) + \mathcal{O}(MK)$ complexity, where the first term is the cost of intra-voxel sampling, the second term is the cost of querying voxel-field for each proposal and the third term is the cost of drawing points from the queried voxels. We evaluate the latency of point cloud pooling on sequences with different lengths. As shown in Tab. 2, our optimized pooling method can achieve an $8\times$ speedup over the original pooling method. + +# 4. Experiment + +# 4.1. Dataset and Implementation Details + +Dataset. We evaluate our MSF on Waymo Open Dataset (WOD), which is a large-scale multi-modality dataset for autonomous driving. WOD contains 1,150 sequences, which are divided into 798 training, 202 validation, and 150 testing sequences, respectively. Each sequence is 20s long, captured by a 64-line LiDAR sensor at $10\mathrm{Hz}$ frequency. The evaluation metrics used in WOD are mean average precision (mAP) and mAP weighted by heading accuracy (mAPH). Three object classes, "Vehicle", "Pedestrian" and "Cyclist", are evaluated. Each object class is further categorized into two levels of difficulties, LEVEL1 and LEVEL2. The former refers to objects with more than 5 points and the latter refers to objects with less than 5 but at least 1 point. + +Implementation. We employ the traditional Center-Point [28, 33, 36] as our region proposal network (RPN). To include motion information, we concatenate four adjacent frames at input and add one additional head for pre + +Table 3. Performance comparison on the validation set of Waymo Open Dataset. + +
MethodFramesALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
SECOND [28]163.0557.2372.27/71.6963.85/63.3368.70/58.1860.72/51.3160.62/59.2858.34/57.05
PointPillar [9]163.3357.5371.60/71.0063.10/62.5070.60/56.7062.90/50.2064.40/62.3061.90/59.90
IA-SSD [35]164.4858.0870.53/69.6761.55/60.8069.38/58.4760.30/50.7367.67/65.3064.98/62.71
LiDAR R-CNN [10]166.2060.1073.50/73.0064.70/64.2071.20/58.7063.10/51.7068.60/66.9066.10/64.40
RSN [26]1--75.10/74.6066.00/65.5077.80/72.7068.30/63.70--
PV-RCNN [21]169.6363.3377.51/76.8968.98/68.4175.01/65.6566.04/57.6167.81/66.3565.39/63.98
Part-A2 [24]170.2563.8477.05/76.5168.47/67.9775.24/66.8766.18/58.6268.60/67.3666.13/64.93
Centerpoint [33]1-65.50--/66.20--/62.60--/67.60
VoTR [14]1--74.95/74.2565.91/65.29----
VoxSeT [6]172.2466.2274.50/74.0365.99/65.5680.03/72.4272.45/65.3971.56/70.2968.95/67.73
SST-1f [3]1--76.22/75.7968.04/67.6481.39/74.0572.82/65.93--
SWFormer-1f [25]1--77.8/77.369.2/68.880.9/72.772.5/64.9--
PillarNet [20]174.6068.4379.09/78.5970.92/70.4680.59/74.0172.28/66.1772.29/71.2169.72/68.67
PV-RCNN++ [22]175.2168.6179.10/78.6370.34/69.9180.62/74.6271.86/66.3073.49/72.3870.70/69.62
3D-MAN [31]16--74.53/74.0367.61/67.1471.7/67.762.6/59.0--
SST-3f [3]3--78.66/78.2169.98/69.5783.81/80.1475.94/72.37--
SWFormer-3f [25]3--79.4/78.971.1/70.682.9/79.074.8/71.1--
CenterFormer [37]477.073.278.1/77.673.4/72.981.7/78.677.2/74.275.6/74.873.4/72.6
CenterFormer [37]877.373.778.8/78.374.3/73.882.1/79.377.8/75.075.2/74.473.2/72.3
MPPNet [2]479.8374.2281.54/81.0674.07/73.6184.56/81.9477.20/74.6777.15/76.5075.01/74.38
MPPNet [2]1680.4074.8582.74/82.2875.41/74.9684.69/82.2577.43/75.0677.28/76.6675.13/74.52
MSF (ours)480.2074.6281.36/80.8773.81/73.3585.05/82.1077.92/75.1178.40/77.6176.17/75.40
MSF (ours)880.6575.4682.83/82.0175.76/75.3185.24/82.2178.32/75.6178.52/77.7476.32/75.47
+ +dicting the velocity of objects. In our experiments, we first train RPN using the official settings of OpenPCDet $^1$ , and use it to generate proposals of WOD. Based on these proposals, we then train our region-based network for 6 epochs by using the ADAM optimizer with an initial learning rate of 0.003 and a batch size of 16. The learning rate is decayed with One-Cycle policy and the momentum is set between $[85\%, 95\%]$ . During the training, we sample region proposals with IoU>0.5 and conduct proposal augmentation following PointRCNN [23]. A number of 128 raw Li-DAR points are randomly sampled for each proposal. The voxel size $v$ and the points-per-voxel $k$ used in voxel-based sampling are set to 0.4 and 32, respectively. The feature dimension of each point in the learning block is set to 256. At the training stage, we use the intermediate supervision by adding loss to the output of each learning block and sum all the intermediate losses to train the model. At the test stage, we only use the bounding boxes and the confidence scores predicted from the last learning block. + +# 4.2. Results on WOD + +Waymo Validation Set. Tab. 3 compares the performance of MSF and the current state-of-the-art methods. As can be seen, multi-frame detectors generally outperform single-frame methods. The best two multi-frame methods by far are MPPNet [2] based on proposal trajectory and + +CenterFormer [37] based on transformer model. Using the same number of frames, our MSF method significantly outperforms CenterFormer by $3.4\%$ APH and $1.2\%$ APH on LEVEL1 and LEVEL2, respectively. This demonstrates that fusing multiple frame feature at the region level is more effective than the fusion with convolutional features. Compared with MPPNet [2], which is also based on region-level fusion, our method outperforms it on almost all cases, except for the APH of Vehicle LEVEL1. It should be noted that our MSF method only use 8 frames, while MPPNet uses 16 frames. MSF uses different pooling sizes in different frames, therefore generating multi-level contextual features for each object. Specifically, MSF models with 4- and 8-frames achieve the mAPH of $74.62\%$ and $75.13\%$ , respectively, recording new state-of-the-arts. In addition, both CenterFormer and MPPNet extract features on each frame of the sequence with the need of a memory bank to store the intermediate results of preceding frames, while our method performs proposal generation only on the current frame, which is more practical to online inference. + +Waymo Test Set. In Tab. 4, we show the results of our 8-frame model by submitting the prediction result to the online server for evaluation. Our method outperforms all the previously published methods. In particular, the improvements on Vehicle and Pedestrian classes are significant, outperforming the best competitor, i.e., CenterFormer, by $1.91\%$ APH and $1.89\%$ APH on LEVEL2, respectively. + +Table 4. Performance comparison on the test set of Waymo Open Dataset. + +
MethodALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
PointPillar [9]--68.1060.1068.00/55.5061.40/50.10--
StarNet [15]--61.0054.5067.80/59.9061.10/54.00--
M3DETR [5]67.161.977.7/77.170.5/70.068.2/58.560.6/52.067.3/65.765.3/63.8
3D-MAN [31]--78.2869.9869.97/65.9863.98/60.26--
PV-RCNN++ [22]75.770.281.6/81.273.9/73.580.4/75.074.1/69.071.9/70.869.3/68.2
CenterPoint [33]77.271.981.1/80.673.4/73.080.5/77.374.6/71.574.6/73.772.2/71.3
RSN [26]--80.3071.6078.90/75.6070.70/67.80--
SST-3f [3]78.372.881.0/80.673.1/72.783.3/79.776.9/73.575.7/74.673.2/72.2
MPPNet [2]80.5975.6784.27/83.8877.29/76.9184.12/81.5278.44/75.9377.11/76.3674.91/74.18
CenterFormer [37]80.9176.2985.36/84.9478.68/78.2885.22/ 82.4880.09/77.4276.21/75.3274.04/73.17
MSF (ours)81.7476.9686.07/85.6779.20/78.8285.99/83.1080.61/77.8277.29/76.4475.09/74.25
+ +Table 5. Ablation experiments on the WOD validation set. "ME" refers to using Motion Embedding and "SA" refers to using Self-Attention modules. "Q" means using query-based decoder layer to generate global representation of proposal features and "M" means using single max-pooling layer for global feature generation. APH scores on LEVEL1 and LEVEL2 are reported. + +
MESABiFADecoderL1L2
✓✓✓Q80.2074.62
X✓✓Q80.03 (-0.17)74.50 (-0.12)
✓X✓Q76.51 (-3.69)71.91 (-2.71)
✓✓XQ78.25 (-1.95)73.11 (-1.51)
✓✓✓M79.54 (-0.66)74.08 (-0.54)
+ +Table 6. Bidirectional feature aggregation vs. unidirectional feature aggregation. + +
ForwardBackwardL1L2
✓✓80.2074.62
✓×79.64 (-0.56)74.35 (-0.27)
×✓78.97 (-1.23)73.77 (-0.85)
+ +Table 7. The performance by integrating the proxy points and MLP mixer from MPPNet. "SA" and "Mixer" denote the self-attention and MLP Mixer [2], respectively. The APH scores of LEVEL2 on three object classes are reported. + +
ConfigVehicle.Pedestrian.Cyclist
Raw + SA73.3575.1175.40
Proxy + SA73.12 (-0.23)74.20 (-0.91)74.32 (-1.08)
Proxy + Mixer73.45(+0.10)74.13 (-0.98)74.39 (-1.01)
+ +# 4.3. Ablation Studies + +In this section, we conduct in-depth analysis on MSF by evaluating the effectiveness of each individual component of it. We report the APH metric of our 4-frame model on the WOD validation set. The ablation study results are shown + +in Tab. 5 to Tab. 8, respectively. + +Motion embedding. As can be seen from the $1^{\mathrm{st}}$ and $2^{\mathrm{nd}}$ rows of Tab. 5, without motion encoding, the performance of our proposed MSF will drop by $0.17\%$ on LEVEL1 and $0.12\%$ in LEVEL2. This indicates that the motion information is beneficial to infer the object's geometry information in point cloud sequences. + +Self-attention. As can be seen from the $1^{\text{st}}$ and $3^{\text{rd}}$ rows of Tab. 5, without using self-attention for spatial interactions, MSF will suffer from significant performance drop, $3.69\%$ on LEVEL1 and $2.71\%$ on LEVEL2. This indicates that the intra-frame interaction is vital for learning internal geometry information of the proposals. + +Bidirectional feature aggregation. From the $1^{\text{st}}$ and $4^{\text{th}}$ rows of Tab. 5, we can find that enforcing temporal interactions among hidden features of proposals by using the proposed BiFA module can bring an improvement of $1.95\%$ on LEVEL1 and $1.51\%$ on LEVEL2, which demonstrates the benefits of exploring long-term dependencies in the point cloud sequence. We also conduct experiments by performing unidirectional feature aggregation with either forward or backward path. The results are shown in Tab. 6, from which we see that the unidirectional model obtains lower APH scores than that of the bidirectional model. + +Query-based decoder layer. As shown in the last row of Tab. 5, by replacing the final query-based decoder layer with a single max-pooling layer, the performance will drop slightly by $0.66\%$ and $0.54\%$ . This is because the final proposal representation after decoder can be regarded as a weighted sum of all point features, and the decoding weights of different points can intrinsically introduce more dynamics for feature representation. + +Spatial modeling with proxy points and MLP mixer. We perform an analysis by integrating the proxy points and MLP mixer developed in our best competitor MPPNet [2] into MSP. We first follow MPPNet to generate proxy point + +Table 8. The APH (L2) scores on objects with different velocities using the 8-frame model. The objects are categorized into stationary $(<0.2\mathrm{m/s})$ , slow $(0.2\sim1\mathrm{m/s})$ , medium $(1\sim6\mathrm{m/s})$ and fast $(>6\mathrm{m/s})$ groups. + +
StationarySlowMediumFast
γ=1.070.3270.3075.6080.32
γ=1.170.4371.1777.5682.48
γ=1.270.3571.2377.4482.31
+ +of each proposal and apply a set abstraction [17] to aggregate the point-wise features to every proxy point. As shown in the $1^{\mathrm{st}}$ and $2^{\mathrm{nd}}$ rows of Tab. 7, using features of proxy points will degrade the performance on Vehicle and more significantly on Pedestrian and Cyclist classes. After replacing the self-attention module with the MLP mixer, the performance degradation on Pedestrian and Cyclist classes still exists, as shown in the $3^{\mathrm{rd}}$ row of Tab. 7. This phenomenon is contradictory to the conclusion in MPPNet that using proxy points to formulate proposal features is more effective than using raw points. We believe this is because MPPNet applies per-frame detection, and thus the proposals across frames may have different sizes, while proxy points can provide consistent representations for different frames. In contrast, our method uses propagated proposals with the same size over the sequence, and hence our proposal features are naturally on the same scale. For small objects such as pedestrian and cyclist, the features from raw points can provide more fine-grained details than proxy points. + +Effects of $\gamma$ . We evaluate how $\gamma$ affects the performance on the objects with different speeds. As shown in Tab. 8, gradually expanding the proposal region with $\gamma = 1.1$ will bring substantial improvements on the moving objects. This demonstrates that our model is able to capture the information of fast moving objects, even in the distant frames. Interestingly, we also see improvements on slow and stationary objects. This is because the different pooling sizes could provide multi-level contextual information, which is beneficial for detecting the objects. As shown in the $3^{\mathrm{rd}}$ row of Tab. 8, no further improvements can be made by increasing the value of $\gamma$ to 1.2. + +# 4.4. Runtime Analysis + +Fig. 4 illustrates the latency decomposition of different methods. Here the latency is computed as the average runtime over 100 samples that are randomly drawn from WOD validation set, which is measured by a single Nvidia GeForce RTX A6000 GPU. Since all the methods employ the same backbone network, we exclude the runtime of the backbone from the latency for better comparison. For MPPNet and CenterFormer, we assume that the features from the preceding frames are already available in the memory bank. As can be seen, MPPNet costs much time in the feature encoding stage because it performs set-abstraction [17] to + +![](images/645401f9ba61a5b294e407b6e118c4ac15d7e17b14799607950fd91c1aa03ee6.jpg) +Figure 4. Comparison of the runtime of different methods. + +Table 9. The runtime decomposition of MPPNet and MSF. + +
MPPNetMSF
MLP-Mixer 75 msCross-Attention 24 msSelf-Attention 36 msBiFA 12 ms
+ +group raw points on the proxy points, which requires manipulating point clouds in continuous space. In comparison, our method takes only 4 milliseconds to encode raw points as proposal features. The overall latency of MPP-Net, CenterFormer and MSF are 99, 80 and 50 milliseconds, respectively. Tab. 9 further decomposes the network latency regarding the spatial and temporal modules in MPP-Net and MSF. As can be seen, the self-attention module and BiFA module are much more efficient than MLP-Mixer and cross-attention module, respectively. Thanks to our optimized point cloud pooling, MSF achieves higher efficiency even compared with CenterFormer, which performs feature fusion directly on the convolutional features. + +# 5. Conclusion + +We presented a novel motion-guided sequential fusion method, namely MSF, for 3D object detection from point cloud sequence. Unlike previous multi-frame detectors that performed feature extraction on each frame of the sequence, MSF adopted a proposal propagation strategy to mine points-of-interest based on the proposals generated on the current frame, therefore reducing the redundant computations and relieving reliance on the preceding results. A bidirectional feature aggregation module was proposed to enable cross-frame interaction between proposal features, facilitating MSF to capture long-term dependencies over the sequence. Besides, we optimized the point cloud pooling process, allowing MSF to process large-scale point cloud sequences in milliseconds. The proposed MSF achieved state-of-the-art performance on the Waymo Open Datasets and it was more efficient than other multi-frame detectors. In future research, we plan to extend MSF to generate detection priors on the future frames to further reduce the overall computation of multi-frame detection. + +# References + +[1] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1 +[2] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2, 3, 5, 6, 7 +[3] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8458-8468, June 2022. 2, 6, 7 +[4] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and ZhaoXiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2918-2927, October 2021. 2 +[5] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (CVPR), pages 772-782, 2022. 7 +[6] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8417-8427, June 2022. 1, 6 +[7] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11873-11882, 2020. 1, 2 +[8] Rui Huang, Wanyue Zhang, Abhijit Kundu, Caroline Pantofoaru, David A Ross, Thomas Funkhouser, and Alireza Fathi. An lstm approach to temporal 3d object detection in lidar point clouds. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2020. 1, 2 +[9] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 1, 2, 6, 7 +[10] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6 +[11] Zhidong Liang, Zehan Zhang, Ming Zhang, Xian Zhao, and Shiliang Pu. Rangeoudet: Range image based real-time 3d object detector optimized by intersection over union. In Proceedings of the IEEE/CVF Conference on Computer Vision + +and Pattern Recognition (CVPR), pages 7140-7149, June 2021. 2 +[12] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019. 2 +[13] Wenjie Luo, Bin Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3569-3577, 2018. 2 +[14] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3164-3173, October 2021. 1, 5, 6 +[15] Jiquan Ngiam, Benjamin Caine, Wei Han, Brandon Yang, Yuning Chai, Pei Sun, Yin Zhou, Xi Yi, Ouais Alsharif, Patrick Nguyen, et al. Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069, 2019. 7 +[16] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3d object detection in point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2 +[17] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 2, 8 +[18] Charles R. Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6134-6144, June 2021. 1, 2, 3 +[19] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2743-2752, October 2021. 2, 4, 5 +[20] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: High-performance pillar-based 3d object detection. 2022. 6 +[21] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2, 6 +[22] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. arXiv preprint arXiv:2102.00463, 2021. 6, 7 +[23] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincarnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-779, 2019. 2, 6 +[24] Shaoshuai Shi, Zhe Wang, Xiaogang Wang, and Hongsheng Li. Part-a $^2$ net: 3d part-aware and aggregation neural net + +work for object detection from point cloud. arXiv preprint arXiv:1907.03670, 2019. 6 +[25] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. 2022. 6 +[26] Pei Sun, Weiyue Wang, Yuning Chai, Gamaleldin El-sayed, Alex Bewley, Xiao Zhang, Cristian Sminchisescu, and Dragomir Anguelov. Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR, pages 5725-5734, 2021. 6, 7 +[27] Dominic Zeng Wang and Ingmar Posner. Voting for voting in online point cloud object detection. Robotics: Science and Systems, 1(3):10-15607, 2015. 1 +[28] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 2, 5, 6 +[29] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2 +[30] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE international conference on computer vision (ICCV), 2019. 2, 5 +[31] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1863-1872, June 2021. 1, 2, 6, 7 +[32] Junbo Yin, Jianbing Shen, Chenye Guan, Dingfu Zhou, and Ruigang Yang. Lidar-based online 3d video object detection with graph-based message passing and spatiotemporal transformer attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2 +[33] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11784-11793, June 2021. 2, 3, 5, 6, 7 +[34] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18953-18962, June 2022. 2 +[35] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18953-18962, 2022. 6 +[36] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2018. 1, 2, 4, 5 + +[37] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2022. 1, 6, 7 \ No newline at end of file diff --git a/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/images.zip b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18839170e0ed682821dc769159355fd17382b277 --- /dev/null +++ b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41e0efee246ee48e3d994e435e1cdcc783130944b9568334e600b1bf3ef08c36 +size 740179 diff --git a/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/layout.json b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc65aaae6f9a49e7bfdcb8b14fb14812f1a9e6ac --- /dev/null +++ b/2023/MSF_ Motion-Guided Sequential Fusion for Efficient 3D Object Detection From Point Cloud Sequences/layout.json @@ -0,0 +1,7809 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 102, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 102, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 125, + 102, + 470, + 140 + ], + "type": "text", + "content": "MSF: Motion-guided Sequential Fusion for Efficient 3D Object Detection from Point Cloud Sequences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 161, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 161, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 131, + 161, + 465, + 190 + ], + "type": "text", + "content": "Chenhang He* Ruihuang Li* Yabin Zhang Shuai Li Lei Zhang† The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 191, + 462, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 191, + 462, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 462, + 205 + ], + "type": "text", + "content": "{csche, csrhli, csybhang, cssli, cslzhang}@comp.polyu.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "text", + "content": "Point cloud sequences are commonly used to accurately detect 3D objects in applications such as autonomous driving. Current top-performing multi-frame detectors mostly follow a Detect-and-Fuse framework, which extracts features from each frame of the sequence and fuses them to detect the objects in the current frame. However, this inevitably leads to redundant computation since adjacent frames are highly correlated. In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, which exploits the continuity of object motion to mine useful sequential contexts for object detection in the current frame. We first generate 3D proposals on the current frame and propagate them to preceding frames based on the estimated velocities. The points-of-interest are then pooled from the sequence and encoded as proposal features. A novel Bidirectional Feature Aggregation (BiFA) module is further proposed to facilitate the interactions of proposal features across frames. Besides, we optimize the point cloud pooling by a voxel-based sampling technique so that millions of points can be processed in several milliseconds. The proposed MSF method achieves not only better efficiency than other multi-frame detectors but also leading accuracy, with " + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "inline_equation", + "content": "83.12\\%" + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "inline_equation", + "content": "78.30\\%" + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "text", + "content": " mAP on the LEVEL1 and LEVEL2 test sets of Waymo Open Dataset, respectively. Codes can be found at https://github.com/skyhehel23/MSF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "type": "text", + "content": "3D object detection [1,2,6,7,9,14,21,27-29,36] is one of the key technologies in autonomous driving, which helps the vehicle to better understand the surrounding environment and make critical decisions in the downstream tasks. As an indispensable sensing device in autonomous driving systems, LiDAR collects 3D measurements of the scene in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 229, + 457, + 384 + ], + "blocks": [ + { + "bbox": [ + 312, + 229, + 457, + 384 + ], + "lines": [ + { + "bbox": [ + 312, + 229, + 457, + 384 + ], + "spans": [ + { + "bbox": [ + 312, + 229, + 457, + 384 + ], + "type": "image", + "image_path": "16385f200141b776c4a846848ac3ffa1b9ab7eeb7285b35ce2bb067b73d9d325.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 388, + 547, + 444 + ], + "lines": [ + { + "bbox": [ + 305, + 388, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 547, + 444 + ], + "type": "text", + "content": "Figure 1. (a) The \"Detect-and-Fuse\" framework extracts features from each frame of the sequence and then fuses them, while (b) our proposed \"Motion-guided Sequential Fusion\" (MSF) method generates proposals on the current frame and propagates them to preceding frames to explore useful contexts in the sequence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 460, + 228, + 544, + 385 + ], + "blocks": [ + { + "bbox": [ + 460, + 228, + 544, + 385 + ], + "lines": [ + { + "bbox": [ + 460, + 228, + 544, + 385 + ], + "spans": [ + { + "bbox": [ + 460, + 228, + 544, + 385 + ], + "type": "image", + "image_path": "8ae0c5cec61df169bef6a172848d8978ddc21cf119d375f5cb6a2f001b3a1c5b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "text", + "content": "the form of point clouds. However, LiDAR can only produce partial view of the scene at a time, and the sparse and incomplete representation of point clouds brings considerable challenges to the 3D object detection task. In practice, the LiDAR sensor will continuously sense the environment and produce a sequence of point cloud frames over time. The multi-frame data can provide a denser representation of the scene as the vehicle moves. Therefore, how to fuse these multi-frame point cloud data for more accurate object detection is worth deep investigation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "Recent works mainly focus on deep feature fusion with multi-frame point clouds, for example, aggregating dense birds-eye-view features via Transformer models [31, 37], passing the voxel features to LSTM [8] or GRU [32] modules for temporal modeling. Some top-performing detectors [2, 18] focus on fusing proposal features, where a tracker is employed to associate the 3D proposals across frames, and a region-based network is applied to refine the current proposals by incorporating contextual features from the proposal trajectories. These approaches generally follow a \"Detect-and-Fuse\" framework, as shown in Fig. 1(a), where the model requires to process each frame of the sequence," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5196" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "and the predictions on the current frame rely on the results of preceding frames. Since online detection is a causal system, such a detection framework might cause significant delay if the network is still processing a preceding frame when the current frame is loaded." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 395 + ], + "type": "text", + "content": "In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, as shown in Fig. 1(b), which leverages the continuity of object motion to extract useful contexts from point cloud sequences and improve the detection of current frame. Specifically, considering that the motions of objects are relatively smooth in a short sequence, we propagate the proposals generated on current frame to preceding frames based on the velocities of objects, and sample reliable points-of-interest from the sequence. In this way, we bypass extracting features on each frame of the sequence, which reduces the redundant computation and reliance on the results of preceding frames. The sampled points are then transformed to proposal features via two encoding schemes and passed to a region-based network for further refinement. Specifically, a self-attention module is employed to enhance the interaction of point features within proposals, while a novel Bidirectional Feature Aggregation (BiFA) module is proposed to enforce the information exchange between proposals across frames. The refined proposal features consequently capture both spatial details and long-term dependencies over the sequence, leading to more accurate bounding-box prediction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "type": "text", + "content": "It is found that the existing point cloud pooling methods [2, 19, 23, 30] are inefficient, taking more than 40 milliseconds when processing millions of points from sequential point clouds. We find that the major bottleneck lies in the heavy computation of pair-wise distances between " + }, + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "type": "text", + "content": " points and " + }, + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "type": "text", + "content": " proposals, which costs " + }, + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(nm)" + }, + { + "bbox": [ + 46, + 396, + 287, + 525 + ], + "type": "text", + "content": " complexity. To further improve the efficiency, we optimize the point cloud pooling with a voxel sampling technique. The improved pooling operation is of linear complexity and can process millions of points in several milliseconds, more than eight times faster than the original method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 527, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 527, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 59, + 527, + 286, + 538 + ], + "type": "text", + "content": "Overall, our contributions can be summarized as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 544, + 286, + 672 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 57, + 544, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 286, + 590 + ], + "type": "text", + "content": "- An efficient Motion-guided Sequential Fusion (MSF) method is proposed to fuse multi-frame point clouds at region level by propagating the proposals of current frame to preceding frames based on the object motions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 595, + 286, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 595, + 286, + 629 + ], + "spans": [ + { + "bbox": [ + 57, + 595, + 286, + 629 + ], + "type": "text", + "content": "- A novel Bidirectional Feature Aggregation (BiFA) module is introduced to facilitate the interactions of proposal features across frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 637, + 286, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 286, + 672 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 286, + 672 + ], + "type": "text", + "content": "- The point cloud pooling method is optimized with a voxel-based sampling technique, significantly reducing the runtime on large-scale point cloud sequence." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "The proposed MSF method is validated on the challenging Waymo Open Dataset, and it achieves leading accuracy on the LEVEL1 and LEVEL2 test sets with fast speed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 95, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 346 + ], + "type": "text", + "content": "Single-frame 3D object detection. Recent research on single-frame 3D object detection is mainly focused on representation learning on point clouds. Voxel-based detectors [28, 33, 36] rasterize the point cloud into volumetric representation, followed by 3D CNN to extract dense features. Some works convert point clouds into 2D birds-eye-view [9] or range view [4, 11] representations, and process them with more efficient 2D CNN. Following PointNet++ [17], point-based methods [16, 23, 29, 30, 34] directly process point clouds in continuous space, and extract highly-semantic features through a series of downsampling and set abstraction layers. Voxel-point approaches [7, 12, 21] employ a hybrid representation, where the flexible conversion between voxel-based and point-based representations are explored, leading to better balance between efficiency and performance. Our method employs a high-quality voxel-based detector CenterPoint [33] as the proposal generation network to predict 3D proposals of current frame and their motions. We then employ an efficient region-based network to further refine these proposals by mining sequential points from point cloud sequence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 350, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 638 + ], + "type": "text", + "content": "3D object detection from point cloud sequence. Multi-frame point clouds provide richer 3D information of the environment. While some single-frame detectors [3, 33] can be adapted to point cloud sequence by simply concatenating multi-frame point cloud as the input, the improvements are typically marginal and the performance can be even worse when encountering moving objects. Fast-and-furious [13] explores an intermediate fusion to align multi-frame point cloud by concatenating the hidden feature maps of the backbone network. However, it still suffers from the misalignment brought by the fast-moving objects in long sequence. Recent approaches [8, 32] demonstrate that an in-depth fusion can be achieved with recurrent networks. Unfortunately, the use of a single memory to store and update features across frames builds a potential bottleneck. To resolve such limitations, 3D-MAN [31] first attempts to employ the attention mechanism to align different views of 3D objects and then exploits a memory bank to store and aggregate multi-frame features for long sequence. Recently, Offboard3D [18] and MPPNet [2] improve much the detection performance, where they associate the detected boxes from each frame of the sequence as proposal trajectories, and extract high-quality proposal features by sampling sequential point cloud on the trajectories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Our MSF method also samples points from the sequence, but it differs from those methods with proposal trajectories [2, 18] in that we only generate proposals on the current frame and propagate them to explore features in preceding frames. This makes our method much more efficient and favorable to online detection systems." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5197" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 68, + 293, + 284 + ], + "blocks": [ + { + "bbox": [ + 49, + 68, + 293, + 284 + ], + "lines": [ + { + "bbox": [ + 49, + 68, + 293, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 68, + 293, + 284 + ], + "type": "image", + "image_path": "4f2c2d8d78724a94d855046ec0cf9856ee549fcf4929dfbc796aa31204874c53.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 546, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 546, + 360 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our proposed Motion-guided Sequential Fusion (MSF) approach. By taking a point cloud sequence as input, MSF employs a region proposal network to generate proposals on the current frame and sample points-of-interest from the sequence by using motion-guided sequential pooling. The sampled points are encoded as high-dimensional proposal features and passed to a region-based network, where three learning blocks are consequently applied to refine the proposal features. A Bidirectional Feature Aggregation (BiFA) module is introduced in the region-based network to facilitate the interactions of proposal features across frames. The red and blue cubes represent single-point features from the current frame and preceding frame, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 68, + 545, + 283 + ], + "blocks": [ + { + "bbox": [ + 294, + 68, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 294, + 68, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 294, + 68, + 545, + 283 + ], + "type": "image", + "image_path": "46252ebc4e81f5e1dc003f9b119be1c35fd45f78eefa90de399916842dc056cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 430, + 274, + 487 + ], + "blocks": [ + { + "bbox": [ + 46, + 378, + 288, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 288, + 422 + ], + "type": "text", + "content": "Table 1. Recall rates of foreground points by using per-frame detection based proposal trajectory method [2] and our motion guided proposal generation method. We employ the CenterPoint [33] as proposal generator and evaluate on Waymo validation split." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 430, + 274, + 487 + ], + "lines": [ + { + "bbox": [ + 58, + 430, + 274, + 487 + ], + "spans": [ + { + "bbox": [ + 58, + 430, + 274, + 487 + ], + "type": "table", + "html": "
4-frame8-frame16-frame
Trajectory [2]93.2%92.8%90.5%
Ours (γ = 1.0)92.3%87.5%78.3%
Ours (γ = 1.1)93.5%91.7%87.3%
", + "image_path": "7b7974c2362f8ff564fcbb2c6a4fc8e678cea421762f11f594d66445f3c25f50.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 493, + 233, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 233, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 233, + 507 + ], + "type": "text", + "content": "3. Motion-guided Sequential Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 514, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 647 + ], + "type": "text", + "content": "This section presents our Motion-guided Sequential Fusion (MSF) approach for efficient 3D object detection on point cloud sequences. The overall architecture of MSF is illustrated in Fig. 2. In Sec. 3.1, we describe the details of motion-guided sequential pooling, which effectively mines reliable sequential points-of-interest based on the proposals of current frame. In Sec. 3.2, we present the region-based network, including the formulation of proposal features and a novel bidirectional feature aggregation module. In Sec. 3.3, we demonstrate a voxel-based sampling technique to accelerate the current point cloud pooling method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 230, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 230, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 230, + 671 + ], + "type": "text", + "content": "3.1. Motion-guided Sequential Pooling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Current multi-frame detection methods [2,18] mostly explore proposal trajectories to generate high-quality point cloud representations. However, such a scheme relies on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": "frame-by-frame proposal generation, which is not suitable for online detection systems. We observe that in a point cloud sequence, although objects move at different speeds, their motions are relatively smooth. That is to say, we can estimate their motion displacements and roughly localize their positions in preceding frames. To this end, given a point cloud sequence " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\{I^t\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": ", we propose to propagate the proposals generated on the current frame " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "I^T" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": " to preceding frames " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\{I^t\\}_{t=1}^{T-1}" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": " based on their estimated velocities. Since moving objects may slightly deviate from the estimated positions in the preceding frames, we sample the points-of-interest in a cylindrical region of each proposal and gradually increase the diameter of the region by a factor " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": " as the proposal propagates. Let's denote a proposal of current frame as " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "(p_x, p_y, p_z, w, l, h, \\theta)" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "(p_x, p_y, p_z)" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": " denotes its center location, " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "w, l, h" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": " denote its width, length, height and yaw angle, respectively. Suppose that the object has a unit-time velocity " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\vec{v} = (v_x, v_y)" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": ". The corresponding points-of-interest " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "(x^t, y^t, z^t)" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": " sampled from frame " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": " will satisfy the following condition:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 624, + 545, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 624, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 310, + 624, + 545, + 648 + ], + "type": "interline_equation", + "content": "\\left(x ^ {t} - p _ {x} + v _ {x} \\cdot \\Delta t\\right) ^ {2} + \\left(y ^ {t} - p _ {y} + v _ {y} \\cdot \\Delta t\\right) ^ {2} < \\left(\\frac {d ^ {t}}{2}\\right) ^ {2}, \\tag {1}", + "image_path": "2801601dafe0d26d7183f63e1f9a7334a26257b06d34e2f9cf48aa0496a054eb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\Delta t = T - t" + }, + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "type": "text", + "content": " is the time offset of frame " + }, + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "type": "inline_equation", + "content": "d^t = \\sqrt{(w^2 + l^2)}\\cdot \\gamma^{\\Delta t + 1}" + }, + { + "bbox": [ + 306, + 653, + 545, + 677 + ], + "type": "text", + "content": " is the diameter of cylindrical region." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In our preliminary study, we compare the overall recall rates of foreground points between our method and the trajectory-based method [2]. As can be seen in Tab. 1," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5198" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "our method can achieve very close result to the proposal trajectory method on 4-frame sequences. For 8-frame sequences, since it is difficult to estimate the positions of fast-moving objects on distant frames, we use " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\gamma = 1.1" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " to recall more points and obtain comparable results to the proposal trajectory method. The recall rates only drop slightly even when 16-frame sequences are used. Interestingly, we find that setting " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\gamma = 1.1" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " is not only beneficial for detecting fast-moving objects but also beneficial for improving the performance on slow and stationary objects. We believe this is because the points sampled from regions of different sizes contain multi-level contextual information about the objects, which will be further discussed in Sec. 4.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 236, + 177, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 177, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 177, + 248 + ], + "type": "text", + "content": "3.2. Region-based Network" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "text", + "content": "Proposal feature encoding. After sampling " + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "text", + "content": " points in each proposal, we adopt two encoding schemes to generate proposal features. We first follow [19] to calculate the relative offsets between each point-of-interest " + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "inline_equation", + "content": "l_i^t \\in I^t" + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "text", + "content": " and the nine key-points (eight corner points plus one center point) of the proposal box " + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "inline_equation", + "content": "\\{b_j^t \\in P^t : j = 0, \\dots, 8\\}" + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "text", + "content": ". The resulted offsets are then converted to spherical coordinates and transformed, via a Multi-layer Perceptron (MLP), to a geometric embedding " + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "inline_equation", + "content": "g^t \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 46, + 254, + 290, + 386 + ], + "type": "text", + "content": " that encodes the spatial correlation between the sampled point and the proposal box. The encoding scheme can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 395, + 287, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 395, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 72, + 395, + 287, + 410 + ], + "type": "interline_equation", + "content": "g _ {i} ^ {t} = \\operatorname {M L P} \\left(S \\left(\\left\\{l _ {i} ^ {t} - b _ {j} ^ {t} \\right\\} _ {j = 0} ^ {8}\\right)\\right), \\text {f o r} i = 1, \\dots , K, \\tag {2}", + "image_path": "a2185fe9eb708d69438008a5c372852b14740b9d1a2688eac5788d43cba0f94e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "S:(x,y,z)\\to (r,\\theta ,\\phi)" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": " denotes the spherical transform where " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "r = \\sqrt{x^2 + y^2 + z^2}" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "\\theta = \\arcsin (z / r)" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "\\phi = \\arctan (y / x)" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": " respectively. The second scheme produces a motion embedding " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "m^t\\in \\mathbb{R}^{K\\times D}" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": ", which encodes the displacements of the points-of-interest at each frame " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "I^t" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": " relative to the key-points of the proposal boxes " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "b^{0}" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": " in the first frame. The time offsets " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": " are also concatenated, resulting in the motion embedding at frame " + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 418, + 289, + 515 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 523, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 523, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 51, + 523, + 287, + 548 + ], + "type": "interline_equation", + "content": "m _ {i} ^ {t} = \\operatorname {M L P} \\left(\\operatorname {C o n c a t} \\left(\\left\\{l _ {i} ^ {t} - b _ {j} ^ {0} \\right\\} _ {j = 0} ^ {8}, \\Delta t\\right)\\right), \\text {f o r} i = 1, \\dots , K. \\tag {3}", + "image_path": "3b515c95342e1e65c7eb91d1619ce8b99247b77cef53472730fb775b6092127c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 548, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 573 + ], + "type": "text", + "content": "The proposal feature " + }, + { + "bbox": [ + 46, + 548, + 287, + 573 + ], + "type": "inline_equation", + "content": "f^t \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 46, + 548, + 287, + 573 + ], + "type": "text", + "content": " can be formulated as the summation of geometric and motion embeddings:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 582, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 582, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 135, + 582, + 287, + 597 + ], + "type": "interline_equation", + "content": "f ^ {t} = g ^ {t} + m ^ {t}. \\tag {4}", + "image_path": "fb1511941b5fa887a7ebbbb0755215a8bf8f53307db9706e0eaeafe1a82caa93.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Bidirectional feature aggregation. MSF employs a region-based network to explore spatial-temporal dependencies among proposal features and fuse them into global representations for final bounding-box refinement. As shown in Fig. 2, the region-based network is composed of three learning blocks. Each block consists of a traditional Multi-Head Self-Attention (MHSA) layer, followed by a Feed-Forward Network (FFN) with residual connection, and our proposed Bidirectional Feature Aggregation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "(BiFA) module. The MHSA layer aims to encode rich spatial relationships and point dependencies in the proposal for refining point features, while the proposed BiFA module aims to encode temporal information by facilitating information exchange between proposals across frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "text", + "content": "Specifically, BiFA involves a forward path and a backward path, representing two ways of information flow along the sequence. Since proposal features have an unordered representation, we leverage the Max-pool&Repeat [36] tensor manipulation to obtain a summarized contextual features. In the forward path, aside from the first frame in the sequence, which is concatenated with its own contextual features, each of the other frames is concatenated with the contextual features of its preceding frame along channel dimension. A point-wise convolutional layer is thereby employed to refine the concatenated feature and halve their channels. Given proposal features " + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "inline_equation", + "content": "f^t" + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "inline_equation", + "content": "f^{t-1}" + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "text", + "content": " from the current frame and its preceding frame, the forward output of frame " + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "text", + "content": ", denote by " + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "inline_equation", + "content": "h_F^t" + }, + { + "bbox": [ + 304, + 133, + 546, + 300 + ], + "type": "text", + "content": ", can be obtained by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 304, + 546, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 304, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 546, + 319 + ], + "type": "interline_equation", + "content": "h _ {F} ^ {t} = \\operatorname {C o n v} \\left(\\operatorname {C o n c a t} \\left(f ^ {t}, \\text {R e p e a t} \\circ \\operatorname {M a x - p o o l} \\left(f ^ {t - 1}\\right)\\right)\\right). \\tag {5}", + "image_path": "563e8caa097d4eb3b6a8d4fcd0ec808e45056dad888203496ac3429a5863c9fe.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "type": "text", + "content": "Unfortunately, introducing the forward path only will lead to information imbalance among different frames, i.e., the current frame receives information from other frames, whereas the last preceding frame receives no information from the sequence. To overcome this limitation, we augment the backward path, where the features of frame " + }, + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "type": "text", + "content": " are aggregated with the contextual features of frame " + }, + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "type": "text", + "content": ", resulting in the backward output " + }, + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "type": "inline_equation", + "content": "h_B^t" + }, + { + "bbox": [ + 304, + 323, + 545, + 421 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 425, + 545, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 425, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 310, + 425, + 545, + 440 + ], + "type": "interline_equation", + "content": "h _ {B} ^ {t} = \\operatorname {C o n v} \\left(\\operatorname {C o n c a t} \\left(h _ {F} ^ {t}, \\text {R e p e a t} \\circ \\operatorname {M a x - p o o l} \\left(h _ {F} ^ {t + 1}\\right)\\right)\\right). \\tag {6}", + "image_path": "6dc3cd0b34957b4df6335b718cf23aa652ec7085bfd1aa132f100e8b560f9d0e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 445, + 546, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 563 + ], + "type": "text", + "content": "In this way, each frame can simultaneously exchange information with its two adjacent frames from both directions, and the information can be propagated to more distant frames after three learning blocks. The resulted proposal features can capture long-term dependencies over the point cloud sequence. It is worthy noting that, the parameters of the convolutional layer in the same path can be shared and all the tensor operations can be performed in parallel. This makes our BiFA module lightweight and much more efficient than other multi-frame fusion modules." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "content": "Outputs and loss function. Finally, we aggregate the point-wise features of each proposal into a global representation through a query-based transformer decoder layer, where a learnable query feature " + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "content": " attends to each point vector of proposal features " + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "inline_equation", + "content": "h^t \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "content": " through a cross-attention module. The decoder layer can be depicted as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 652, + 545, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 652, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 361, + 652, + 545, + 667 + ], + "type": "interline_equation", + "content": "\\hat {e} ^ {t} = \\operatorname {A t t e n t i o n} \\left(q, h ^ {t}, h ^ {t}\\right) + h ^ {t}, \\tag {7}", + "image_path": "8442ba5cff35b682fc37d5d978a47734b5c0c2f99cd7fd64bbec258d556bb512.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 362, + 669, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 669, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 362, + 669, + 545, + 683 + ], + "type": "interline_equation", + "content": "e ^ {t} = \\operatorname {F F N} \\left(\\hat {e} ^ {t}\\right) + \\hat {e} ^ {t}. \\tag {8}", + "image_path": "3932ecde9127bd524cbc2f374bf4befcd61cfc530c2e5c5eb4369733d91c03a2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "The decoded outputs across frames " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{e^t\\}_{t=0}^T" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " are concatenated and passed to a group of detection heads for" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5199" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 542, + 186 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 542, + 186 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 542, + 186 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 542, + 186 + ], + "type": "image", + "image_path": "7d5bfb18b35dc3d2f27ac221e15715f6fbe352e33ef0c869c67ca19a376ca798.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 195, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 547, + 220 + ], + "type": "text", + "content": "Figure 3. Illustration of our optimized point cloud pooling method. We first perform intra-voxel sampling to keep a fixed number of points in each voxel. Then we query " + }, + { + "bbox": [ + 46, + 195, + 547, + 220 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 46, + 195, + 547, + 220 + ], + "type": "text", + "content": " voxels fields for each proposal and uniformly draw points from the non-empty voxels within." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 262, + 272, + 300 + ], + "blocks": [ + { + "bbox": [ + 46, + 237, + 288, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 260 + ], + "type": "text", + "content": "Table 2. The latency of point cloud pooling on 1-frame, 4-frames and 8-frames sequences." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 262, + 272, + 300 + ], + "lines": [ + { + "bbox": [ + 57, + 262, + 272, + 300 + ], + "spans": [ + { + "bbox": [ + 57, + 262, + 272, + 300 + ], + "type": "table", + "html": "
N=168kN=674kN=1382k
Cylindrical Pooling8.2ms25.2ms40.1ms
Our Optimized2.3ms3.4ms5.0ms
", + "image_path": "2038d39f89e82fe8918bdc191e9d0f409aed93499821c519243efa5781ae5ecd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "type": "text", + "content": "boundong-box refinement. The overall loss function " + }, + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "type": "text", + "content": " is the summation of the confidence prediction loss " + }, + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{conf}}" + }, + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "type": "text", + "content": " and the box regression loss " + }, + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 46, + 310, + 287, + 348 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 356, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 356, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 116, + 356, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n f}} + \\alpha \\mathcal {L} _ {\\text {r e g}}, \\tag {9}", + "image_path": "4dafaa34ba31b7ba27cecf0d4a6bd47a684dfa62e49406ff40dcc71dab813323.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "text", + "content": " is a balancing hyper-parameter. We adopt the same binary cross entropy loss and box regression loss employed in CT3D [19] as our " + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{conf}}" + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 420, + 208, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 208, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 208, + 434 + ], + "type": "text", + "content": "3.3. Efficient Point Cloud Pooling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": "In this subsection, we optimize the point cloud pooling method to sample a fixed number of points more efficiently from the cylindrical region of each proposal. As shown in Fig. 3, we perform proposal-based point sampling in two steps, i.e., the intra-voxel sampling and the voxel-field sampling. In the first step, the input space is discretized into a voxel grid with voxel size " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": ", and each point corresponds to a voxel coordinates " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "(\\lfloor \\frac{x}{v} \\rfloor, \\lfloor \\frac{y}{v} \\rfloor)" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": ". Here the voxel partition in " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": "-axis is omitted considering that the cylindrical regions have unlimited height. Then, up to " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": " points are kept in each voxel and padding points are inserted if the voxel has less than " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": " points. Due to the high memory consumption required to store all voxels in a dense grid, we follow [14] to store only non-empty voxels in a contiguous memory and use a hash table to store the mappings between the coordinates of non-empty voxels and their indices in memory space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "In the second step, we first query " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "-by-" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " voxel-field for each proposal and calculate the coordinates of the voxels within. Next, we convert the coordinates of voxel into hashed keys and look up the table for querying the sampled points generated from the first hierarchy. The hash-table will return \"-1\" if no keys are found, which means it is an empty voxel. The queried points are then drawn from these" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "type": "text", + "content": "voxels and stored in an output buffer if the following conditions are met: 1) it is a valid point and 2) it falls into the cylindrical region, i.e. satisfying Eq. 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "text", + "content": "Complexity and efficiency We perform an in-depth analysis of our optimized pooling method and the previous cylindrical pooling methods [2, 19, 30] in continuous space. Given " + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "text", + "content": " points, " + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "text", + "content": " proposals and the requirement of sampling " + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "text", + "content": " points in each proposal, the original method costs " + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(NM) + \\mathcal{O}(MK)" + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "text", + "content": " complexity due to the calculation of pair-wise distances between points and proposals. In contrast, our optimized version costs " + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N) + \\mathcal{O}(M) + \\mathcal{O}(MK)" + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "text", + "content": " complexity, where the first term is the cost of intra-voxel sampling, the second term is the cost of querying voxel-field for each proposal and the third term is the cost of drawing points from the queried voxels. We evaluate the latency of point cloud pooling on sequences with different lengths. As shown in Tab. 2, our optimized pooling method can achieve an " + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "inline_equation", + "content": "8\\times" + }, + { + "bbox": [ + 304, + 276, + 546, + 456 + ], + "type": "text", + "content": " speedup over the original pooling method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 469, + 382, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 382, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 382, + 483 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 489, + 497, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 497, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 497, + 502 + ], + "type": "text", + "content": "4.1. Dataset and Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 509, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 664 + ], + "type": "text", + "content": "Dataset. We evaluate our MSF on Waymo Open Dataset (WOD), which is a large-scale multi-modality dataset for autonomous driving. WOD contains 1,150 sequences, which are divided into 798 training, 202 validation, and 150 testing sequences, respectively. Each sequence is 20s long, captured by a 64-line LiDAR sensor at " + }, + { + "bbox": [ + 304, + 509, + 545, + 664 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 509, + 545, + 664 + ], + "type": "text", + "content": " frequency. The evaluation metrics used in WOD are mean average precision (mAP) and mAP weighted by heading accuracy (mAPH). Three object classes, \"Vehicle\", \"Pedestrian\" and \"Cyclist\", are evaluated. Each object class is further categorized into two levels of difficulties, LEVEL1 and LEVEL2. The former refers to objects with more than 5 points and the latter refers to objects with less than 5 but at least 1 point." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Implementation. We employ the traditional Center-Point [28, 33, 36] as our region proposal network (RPN). To include motion information, we concatenate four adjacent frames at input and add one additional head for pre" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5200" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 91, + 547, + 364 + ], + "blocks": [ + { + "bbox": [ + 150, + 71, + 442, + 82 + ], + "lines": [ + { + "bbox": [ + 150, + 71, + 442, + 82 + ], + "spans": [ + { + "bbox": [ + 150, + 71, + 442, + 82 + ], + "type": "text", + "content": "Table 3. Performance comparison on the validation set of Waymo Open Dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 91, + 547, + 364 + ], + "lines": [ + { + "bbox": [ + 47, + 91, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 547, + 364 + ], + "type": "table", + "html": "
MethodFramesALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
SECOND [28]163.0557.2372.27/71.6963.85/63.3368.70/58.1860.72/51.3160.62/59.2858.34/57.05
PointPillar [9]163.3357.5371.60/71.0063.10/62.5070.60/56.7062.90/50.2064.40/62.3061.90/59.90
IA-SSD [35]164.4858.0870.53/69.6761.55/60.8069.38/58.4760.30/50.7367.67/65.3064.98/62.71
LiDAR R-CNN [10]166.2060.1073.50/73.0064.70/64.2071.20/58.7063.10/51.7068.60/66.9066.10/64.40
RSN [26]1--75.10/74.6066.00/65.5077.80/72.7068.30/63.70--
PV-RCNN [21]169.6363.3377.51/76.8968.98/68.4175.01/65.6566.04/57.6167.81/66.3565.39/63.98
Part-A2 [24]170.2563.8477.05/76.5168.47/67.9775.24/66.8766.18/58.6268.60/67.3666.13/64.93
Centerpoint [33]1-65.50--/66.20--/62.60--/67.60
VoTR [14]1--74.95/74.2565.91/65.29----
VoxSeT [6]172.2466.2274.50/74.0365.99/65.5680.03/72.4272.45/65.3971.56/70.2968.95/67.73
SST-1f [3]1--76.22/75.7968.04/67.6481.39/74.0572.82/65.93--
SWFormer-1f [25]1--77.8/77.369.2/68.880.9/72.772.5/64.9--
PillarNet [20]174.6068.4379.09/78.5970.92/70.4680.59/74.0172.28/66.1772.29/71.2169.72/68.67
PV-RCNN++ [22]175.2168.6179.10/78.6370.34/69.9180.62/74.6271.86/66.3073.49/72.3870.70/69.62
3D-MAN [31]16--74.53/74.0367.61/67.1471.7/67.762.6/59.0--
SST-3f [3]3--78.66/78.2169.98/69.5783.81/80.1475.94/72.37--
SWFormer-3f [25]3--79.4/78.971.1/70.682.9/79.074.8/71.1--
CenterFormer [37]477.073.278.1/77.673.4/72.981.7/78.677.2/74.275.6/74.873.4/72.6
CenterFormer [37]877.373.778.8/78.374.3/73.882.1/79.377.8/75.075.2/74.473.2/72.3
MPPNet [2]479.8374.2281.54/81.0674.07/73.6184.56/81.9477.20/74.6777.15/76.5075.01/74.38
MPPNet [2]1680.4074.8582.74/82.2875.41/74.9684.69/82.2577.43/75.0677.28/76.6675.13/74.52
MSF (ours)480.2074.6281.36/80.8773.81/73.3585.05/82.1077.92/75.1178.40/77.6176.17/75.40
MSF (ours)880.6575.4682.83/82.0175.76/75.3185.24/82.2178.32/75.6178.52/77.7476.32/75.47
", + "image_path": "890ed456711b4a397346682fb3913f541f92eec4b3fd0faf19440cfebbc9e9bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "text", + "content": "dicting the velocity of objects. In our experiments, we first train RPN using the official settings of OpenPCDet" + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "text", + "content": ", and use it to generate proposals of WOD. Based on these proposals, we then train our region-based network for 6 epochs by using the ADAM optimizer with an initial learning rate of 0.003 and a batch size of 16. The learning rate is decayed with One-Cycle policy and the momentum is set between " + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "inline_equation", + "content": "[85\\%, 95\\%]" + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "text", + "content": ". During the training, we sample region proposals with IoU>0.5 and conduct proposal augmentation following PointRCNN [23]. A number of 128 raw Li-DAR points are randomly sampled for each proposal. The voxel size " + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "text", + "content": " and the points-per-voxel " + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 377, + 289, + 605 + ], + "type": "text", + "content": " used in voxel-based sampling are set to 0.4 and 32, respectively. The feature dimension of each point in the learning block is set to 256. At the training stage, we use the intermediate supervision by adding loss to the output of each learning block and sum all the intermediate losses to train the model. At the test stage, we only use the bounding boxes and the confidence scores predicted from the last learning block." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 613, + 149, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 149, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 149, + 625 + ], + "type": "text", + "content": "4.2. Results on WOD" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 633, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 633, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 633, + 287, + 694 + ], + "type": "text", + "content": "Waymo Validation Set. Tab. 3 compares the performance of MSF and the current state-of-the-art methods. As can be seen, multi-frame detectors generally outperform single-frame methods. The best two multi-frame methods by far are MPPNet [2] based on proposal trajectory and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "text", + "content": "CenterFormer [37] based on transformer model. Using the same number of frames, our MSF method significantly outperforms CenterFormer by " + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "text", + "content": " APH and " + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "text", + "content": " APH on LEVEL1 and LEVEL2, respectively. This demonstrates that fusing multiple frame feature at the region level is more effective than the fusion with convolutional features. Compared with MPPNet [2], which is also based on region-level fusion, our method outperforms it on almost all cases, except for the APH of Vehicle LEVEL1. It should be noted that our MSF method only use 8 frames, while MPPNet uses 16 frames. MSF uses different pooling sizes in different frames, therefore generating multi-level contextual features for each object. Specifically, MSF models with 4- and 8-frames achieve the mAPH of " + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "inline_equation", + "content": "74.62\\%" + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "inline_equation", + "content": "75.13\\%" + }, + { + "bbox": [ + 304, + 377, + 547, + 617 + ], + "type": "text", + "content": ", respectively, recording new state-of-the-arts. In addition, both CenterFormer and MPPNet extract features on each frame of the sequence with the need of a memory bank to store the intermediate results of preceding frames, while our method performs proposal generation only on the current frame, which is more practical to online inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Waymo Test Set. In Tab. 4, we show the results of our 8-frame model by submitting the prediction result to the online server for evaluation. Our method outperforms all the previously published methods. In particular, the improvements on Vehicle and Pedestrian classes are significant, outperforming the best competitor, i.e., CenterFormer, by " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.91\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": " APH and " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.89\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": " APH on LEVEL2, respectively." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "text", + "content": "1 https://github.com/open-mmlab/OpenPCDet/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "5201" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 91, + 548, + 255 + ], + "blocks": [ + { + "bbox": [ + 162, + 71, + 430, + 82 + ], + "lines": [ + { + "bbox": [ + 162, + 71, + 430, + 82 + ], + "spans": [ + { + "bbox": [ + 162, + 71, + 430, + 82 + ], + "type": "text", + "content": "Table 4. Performance comparison on the test set of Waymo Open Dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 91, + 548, + 255 + ], + "lines": [ + { + "bbox": [ + 48, + 91, + 548, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 91, + 548, + 255 + ], + "type": "table", + "html": "
MethodALL (3D mAPH)Vehicle (AP/APH)Pedestrian (AP/APH)Cyclist (AP/APH)
L1L2L1L2L1L2L1L2
PointPillar [9]--68.1060.1068.00/55.5061.40/50.10--
StarNet [15]--61.0054.5067.80/59.9061.10/54.00--
M3DETR [5]67.161.977.7/77.170.5/70.068.2/58.560.6/52.067.3/65.765.3/63.8
3D-MAN [31]--78.2869.9869.97/65.9863.98/60.26--
PV-RCNN++ [22]75.770.281.6/81.273.9/73.580.4/75.074.1/69.071.9/70.869.3/68.2
CenterPoint [33]77.271.981.1/80.673.4/73.080.5/77.374.6/71.574.6/73.772.2/71.3
RSN [26]--80.3071.6078.90/75.6070.70/67.80--
SST-3f [3]78.372.881.0/80.673.1/72.783.3/79.776.9/73.575.7/74.673.2/72.2
MPPNet [2]80.5975.6784.27/83.8877.29/76.9184.12/81.5278.44/75.9377.11/76.3674.91/74.18
CenterFormer [37]80.9176.2985.36/84.9478.68/78.2885.22/ 82.4880.09/77.4276.21/75.3274.04/73.17
MSF (ours)81.7476.9686.07/85.6779.20/78.8285.99/83.1080.61/77.8277.29/76.4475.09/74.25
", + "image_path": "7e9729a2a56844773312ceddd079113b8d2d33366a1218bf200d7cdecbcf44b2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 350, + 287, + 422 + ], + "blocks": [ + { + "bbox": [ + 46, + 275, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 287, + 342 + ], + "type": "text", + "content": "Table 5. Ablation experiments on the WOD validation set. \"ME\" refers to using Motion Embedding and \"SA\" refers to using Self-Attention modules. \"Q\" means using query-based decoder layer to generate global representation of proposal features and \"M\" means using single max-pooling layer for global feature generation. APH scores on LEVEL1 and LEVEL2 are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 350, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 422 + ], + "type": "table", + "html": "
MESABiFADecoderL1L2
✓✓✓Q80.2074.62
X✓✓Q80.03 (-0.17)74.50 (-0.12)
✓X✓Q76.51 (-3.69)71.91 (-2.71)
✓✓XQ78.25 (-1.95)73.11 (-1.51)
✓✓✓M79.54 (-0.66)74.08 (-0.54)
", + "image_path": "c2fd076674a619af93cd622c1e62a2edaf9c9087df8253340d856a7fd6c13b41.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 464, + 274, + 515 + ], + "blocks": [ + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "type": "text", + "content": "Table 6. Bidirectional feature aggregation vs. unidirectional feature aggregation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 464, + 274, + 515 + ], + "lines": [ + { + "bbox": [ + 58, + 464, + 274, + 515 + ], + "spans": [ + { + "bbox": [ + 58, + 464, + 274, + 515 + ], + "type": "table", + "html": "
ForwardBackwardL1L2
✓✓80.2074.62
✓×79.64 (-0.56)74.35 (-0.27)
×✓78.97 (-1.23)73.77 (-0.85)
", + "image_path": "b47319960727f38a1a4f1e94b630524d397c8737f75875a692e1e14e65162fd6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 578, + 287, + 627 + ], + "blocks": [ + { + "bbox": [ + 46, + 525, + 287, + 570 + ], + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 570 + ], + "type": "text", + "content": "Table 7. The performance by integrating the proxy points and MLP mixer from MPPNet. \"SA\" and \"Mixer\" denote the self-attention and MLP Mixer [2], respectively. The APH scores of LEVEL2 on three object classes are reported." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 578, + 287, + 627 + ], + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 627 + ], + "type": "table", + "html": "
ConfigVehicle.Pedestrian.Cyclist
Raw + SA73.3575.1175.40
Proxy + SA73.12 (-0.23)74.20 (-0.91)74.32 (-1.08)
Proxy + Mixer73.45(+0.10)74.13 (-0.98)74.39 (-1.01)
", + "image_path": "5490bcc10e3be1fc93184d349ed145b276f094490e0287315e10885e142f3e8e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "In this section, we conduct in-depth analysis on MSF by evaluating the effectiveness of each individual component of it. We report the APH metric of our 4-frame model on the WOD validation set. The ablation study results are shown" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 277, + 435, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 435, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 435, + 289 + ], + "type": "text", + "content": "in Tab. 5 to Tab. 8, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": "Motion embedding. As can be seen from the " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " rows of Tab. 5, without motion encoding, the performance of our proposed MSF will drop by " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "0.17\\%" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " on LEVEL1 and " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "0.12\\%" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " in LEVEL2. This indicates that the motion information is beneficial to infer the object's geometry information in point cloud sequences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": "Self-attention. As can be seen from the " + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "inline_equation", + "content": "3^{\\text{rd}}" + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": " rows of Tab. 5, without using self-attention for spatial interactions, MSF will suffer from significant performance drop, " + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "inline_equation", + "content": "3.69\\%" + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": " on LEVEL1 and " + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "inline_equation", + "content": "2.71\\%" + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": " on LEVEL2. This indicates that the intra-frame interaction is vital for learning internal geometry information of the proposals." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "text", + "content": "Bidirectional feature aggregation. From the " + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "inline_equation", + "content": "4^{\\text{th}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "text", + "content": " rows of Tab. 5, we can find that enforcing temporal interactions among hidden features of proposals by using the proposed BiFA module can bring an improvement of " + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "inline_equation", + "content": "1.95\\%" + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "text", + "content": " on LEVEL1 and " + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "inline_equation", + "content": "1.51\\%" + }, + { + "bbox": [ + 304, + 436, + 545, + 567 + ], + "type": "text", + "content": " on LEVEL2, which demonstrates the benefits of exploring long-term dependencies in the point cloud sequence. We also conduct experiments by performing unidirectional feature aggregation with either forward or backward path. The results are shown in Tab. 6, from which we see that the unidirectional model obtains lower APH scores than that of the bidirectional model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "content": "Query-based decoder layer. As shown in the last row of Tab. 5, by replacing the final query-based decoder layer with a single max-pooling layer, the performance will drop slightly by " + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "inline_equation", + "content": "0.66\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "inline_equation", + "content": "0.54\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "content": ". This is because the final proposal representation after decoder can be regarded as a weighted sum of all point features, and the decoding weights of different points can intrinsically introduce more dynamics for feature representation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Spatial modeling with proxy points and MLP mixer. We perform an analysis by integrating the proxy points and MLP mixer developed in our best competitor MPPNet [2] into MSP. We first follow MPPNet to generate proxy point" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5202" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 123, + 272, + 178 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 8. The APH (L2) scores on objects with different velocities using the 8-frame model. The objects are categorized into stationary " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "(<0.2\\mathrm{m/s})" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": ", slow " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "(0.2\\sim1\\mathrm{m/s})" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": ", medium " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "(1\\sim6\\mathrm{m/s})" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": " and fast " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "(>6\\mathrm{m/s})" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": " groups." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 123, + 272, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 123, + 272, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 123, + 272, + 178 + ], + "type": "table", + "html": "
StationarySlowMediumFast
γ=1.070.3270.3075.6080.32
γ=1.170.4371.1777.5682.48
γ=1.270.3571.2377.4482.31
", + "image_path": "fad166e64ec156533aec0dfff0dacede22b04e399a4556d659cb2ca74aca14b6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "type": "text", + "content": "of each proposal and apply a set abstraction [17] to aggregate the point-wise features to every proxy point. As shown in the " + }, + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "type": "text", + "content": " rows of Tab. 7, using features of proxy points will degrade the performance on Vehicle and more significantly on Pedestrian and Cyclist classes. After replacing the self-attention module with the MLP mixer, the performance degradation on Pedestrian and Cyclist classes still exists, as shown in the " + }, + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 46, + 185, + 287, + 411 + ], + "type": "text", + "content": " row of Tab. 7. This phenomenon is contradictory to the conclusion in MPPNet that using proxy points to formulate proposal features is more effective than using raw points. We believe this is because MPPNet applies per-frame detection, and thus the proposals across frames may have different sizes, while proxy points can provide consistent representations for different frames. In contrast, our method uses propagated proposals with the same size over the sequence, and hence our proposal features are naturally on the same scale. For small objects such as pedestrian and cyclist, the features from raw points can provide more fine-grained details than proxy points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "content": "Effects of " + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "content": ". We evaluate how " + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "content": " affects the performance on the objects with different speeds. As shown in Tab. 8, gradually expanding the proposal region with " + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\gamma = 1.1" + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "content": " will bring substantial improvements on the moving objects. This demonstrates that our model is able to capture the information of fast moving objects, even in the distant frames. Interestingly, we also see improvements on slow and stationary objects. This is because the different pooling sizes could provide multi-level contextual information, which is beneficial for detecting the objects. As shown in the " + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "content": " row of Tab. 8, no further improvements can be made by increasing the value of " + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "content": " to 1.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 563, + 153, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 153, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 153, + 575 + ], + "type": "text", + "content": "4.4. Runtime Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Fig. 4 illustrates the latency decomposition of different methods. Here the latency is computed as the average runtime over 100 samples that are randomly drawn from WOD validation set, which is measured by a single Nvidia GeForce RTX A6000 GPU. Since all the methods employ the same backbone network, we exclude the runtime of the backbone from the latency for better comparison. For MPPNet and CenterFormer, we assume that the features from the preceding frames are already available in the memory bank. As can be seen, MPPNet costs much time in the feature encoding stage because it performs set-abstraction [17] to" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 305, + 71, + 545, + 196 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 196 + ], + "type": "image", + "image_path": "645401f9ba61a5b294e407b6e118c4ac15d7e17b14799607950fd91c1aa03ee6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 205, + 531, + 216 + ], + "lines": [ + { + "bbox": [ + 318, + 205, + 531, + 216 + ], + "spans": [ + { + "bbox": [ + 318, + 205, + 531, + 216 + ], + "type": "text", + "content": "Figure 4. Comparison of the runtime of different methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 243, + 545, + 282 + ], + "blocks": [ + { + "bbox": [ + 318, + 229, + 532, + 240 + ], + "lines": [ + { + "bbox": [ + 318, + 229, + 532, + 240 + ], + "spans": [ + { + "bbox": [ + 318, + 229, + 532, + 240 + ], + "type": "text", + "content": "Table 9. The runtime decomposition of MPPNet and MSF." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 243, + 545, + 282 + ], + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 282 + ], + "type": "table", + "html": "
MPPNetMSF
MLP-Mixer 75 msCross-Attention 24 msSelf-Attention 36 msBiFA 12 ms
", + "image_path": "4e596edc04c94dea7b0d5743ab0512942c363708efef1d4a1718add05ddd9e15.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 297, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 452 + ], + "type": "text", + "content": "group raw points on the proxy points, which requires manipulating point clouds in continuous space. In comparison, our method takes only 4 milliseconds to encode raw points as proposal features. The overall latency of MPP-Net, CenterFormer and MSF are 99, 80 and 50 milliseconds, respectively. Tab. 9 further decomposes the network latency regarding the spatial and temporal modules in MPP-Net and MSF. As can be seen, the self-attention module and BiFA module are much more efficient than MLP-Mixer and cross-attention module, respectively. Thanks to our optimized point cloud pooling, MSF achieves higher efficiency even compared with CenterFormer, which performs feature fusion directly on the convolutional features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 465, + 378, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 378, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 378, + 477 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "We presented a novel motion-guided sequential fusion method, namely MSF, for 3D object detection from point cloud sequence. Unlike previous multi-frame detectors that performed feature extraction on each frame of the sequence, MSF adopted a proposal propagation strategy to mine points-of-interest based on the proposals generated on the current frame, therefore reducing the redundant computations and relieving reliance on the preceding results. A bidirectional feature aggregation module was proposed to enable cross-frame interaction between proposal features, facilitating MSF to capture long-term dependencies over the sequence. Besides, we optimized the point cloud pooling process, allowing MSF to process large-scale point cloud sequences in milliseconds. The proposed MSF achieved state-of-the-art performance on the Waymo Open Datasets and it was more efficient than other multi-frame detectors. In future research, we plan to extend MSF to generate detection priors on the future frames to further reduce the overall computation of multi-frame detection." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5203" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 204 + ], + "type": "text", + "content": "[2] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 270 + ], + "type": "text", + "content": "[3] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8458-8468, June 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 273, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 328 + ], + "type": "text", + "content": "[4] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and ZhaoXiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2918-2927, October 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 330, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 396 + ], + "type": "text", + "content": "[5] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (CVPR), pages 772-782, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 397, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 453 + ], + "type": "text", + "content": "[6] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8417-8427, June 2022. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "text", + "content": "[7] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11873-11882, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 510, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 566 + ], + "type": "text", + "content": "[8] Rui Huang, Wanyue Zhang, Abhijit Kundu, Caroline Pantofoaru, David A Ross, Thomas Funkhouser, and Alireza Fathi. An lstm approach to temporal 3d object detection in lidar point clouds. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 567, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 287, + 621 + ], + "type": "text", + "content": "[9] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "type": "text", + "content": "[10] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "type": "text", + "content": "[11] Zhidong Liang, Zehan Zhang, Ming Zhang, Xian Zhao, and Shiliang Pu. Rangeoudet: Range image based real-time 3d object detector optimized by intersection over union. In Proceedings of the IEEE/CVF Conference on Computer Vision" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "and Pattern Recognition (CVPR), pages 7140-7149, June 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "type": "text", + "content": "[12] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 130, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 547, + 185 + ], + "type": "text", + "content": "[13] Wenjie Luo, Bin Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3569-3577, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "type": "text", + "content": "[14] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3164-3173, October 2021. 1, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 243, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 547, + 297 + ], + "type": "text", + "content": "[15] Jiquan Ngiam, Benjamin Caine, Wei Han, Brandon Yang, Yuning Chai, Pei Sun, Yin Zhou, Xi Yi, Ouais Alsharif, Patrick Nguyen, et al. Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 298, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 547, + 342 + ], + "type": "text", + "content": "[16] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3d object detection in point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 344, + 547, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 547, + 388 + ], + "type": "text", + "content": "[17] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 443 + ], + "type": "text", + "content": "[18] Charles R. Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6134-6144, June 2021. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 444, + 547, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 547, + 509 + ], + "type": "text", + "content": "[19] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2743-2752, October 2021. 2, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 510, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 547, + 533 + ], + "type": "text", + "content": "[20] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: High-performance pillar-based 3d object detection. 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 535, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 547, + 589 + ], + "type": "text", + "content": "[21] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "type": "text", + "content": "[22] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. arXiv preprint arXiv:2102.00463, 2021. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 547, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 691 + ], + "type": "text", + "content": "[23] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincarnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-779, 2019. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "type": "text", + "content": "[24] Shaoshuai Shi, Zhe Wang, Xiaogang Wang, and Hongsheng Li. Part-a" + }, + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "type": "text", + "content": " net: 3d part-aware and aggregation neural net" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5204" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "work for object detection from point cloud. arXiv preprint arXiv:1907.03670, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 137 + ], + "type": "text", + "content": "[25] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 205 + ], + "type": "text", + "content": "[26] Pei Sun, Weiyue Wang, Yuning Chai, Gamaleldin El-sayed, Alex Bewley, Xiao Zhang, Cristian Sminchisescu, and Dragomir Anguelov. Rsn: Range sparse net for efficient, accurate lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR, pages 5725-5734, 2021. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 288, + 239 + ], + "type": "text", + "content": "[27] Dominic Zeng Wang and Ingmar Posner. Voting for voting in online point cloud object detection. Robotics: Science and Systems, 1(3):10-15607, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 240, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 288, + 271 + ], + "type": "text", + "content": "[28] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 273, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 273, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 273, + 288, + 316 + ], + "type": "text", + "content": "[29] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 317, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 288, + 360 + ], + "type": "text", + "content": "[30] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. STD: sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE international conference on computer vision (ICCV), 2019. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "text", + "content": "[31] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1863-1872, June 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 288, + 480 + ], + "type": "text", + "content": "[32] Junbo Yin, Jianbing Shen, Chenye Guan, Dingfu Zhou, and Ruigang Yang. Lidar-based online 3d video object detection with graph-based message passing and spatiotemporal transformer attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 483, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 288, + 536 + ], + "type": "text", + "content": "[33] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11784-11793, June 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 602 + ], + "type": "text", + "content": "[34] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18953-18962, June 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 604, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 288, + 668 + ], + "type": "text", + "content": "[35] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18953-18962, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[36] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2018. 1, 2, 4, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 126 + ], + "type": "text", + "content": "[37] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2022. 1, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5205" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_content_list.json b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..15a5278e840a874fd1e669e655fc9e17458e2a40 --- /dev/null +++ b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_content_list.json @@ -0,0 +1,1821 @@ +[ + { + "type": "text", + "text": "MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID", + "text_level": 1, + "bbox": [ + 89, + 161, + 878, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianyang Gu", + "bbox": [ + 205, + 207, + 312, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Wang", + "bbox": [ + 337, + 207, + 425, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Luo3", + "bbox": [ + 447, + 207, + 526, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Chen", + "bbox": [ + 547, + 207, + 645, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Jiang $^{1*}$", + "bbox": [ + 667, + 207, + 761, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuqiang Fang", + "bbox": [ + 228, + 224, + 347, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanghang Zhang", + "bbox": [ + 369, + 224, + 522, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang You", + "bbox": [ + 544, + 224, + 631, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jian Zhao $^{7*}$", + "bbox": [ + 652, + 224, + 744, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University", + "bbox": [ + 181, + 239, + 347, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National University of Singapore", + "bbox": [ + 369, + 239, + 638, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Alibaba Group", + "bbox": [ + 655, + 239, + 786, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ OPPO Research Institute", + "bbox": [ + 164, + 256, + 366, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5Space Engineering University", + "bbox": [ + 369, + 256, + 633, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6Peking University", + "bbox": [ + 655, + 256, + 805, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{7}$ Institute of North Electronic Equipment", + "bbox": [ + 324, + 273, + 645, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gu_jianyang, jiangwei_zju}@zju.edu.cn zhaiqian90@u.nus.edu", + "bbox": [ + 225, + 290, + 737, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 336, + 310, + 350 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Architecture Search (NAS) has been increasingly appealing to the society of object Re-Identification (ReID), for that task-specific architectures significantly improve the retrieval performance. Previous works explore new optimizing targets and search spaces for NAS ReID, yet they neglect the difference of training schemes between image classification and ReID. In this work, we propose a novel Twins Contrastive Mechanism (TCM) to provide more appropriate supervision for ReID architecture search. TCM reduces the category overlaps between the training and validation data, and assists NAS in simulating real-world ReID training schemes. We then design a Multi-Scale Interaction (MSI) search space to search for rational interaction operations between multi-scale features. In addition, we introduce a Spatial Alignment Module (SAM) to further enhance the attention consistency confronted with images from different sources. Under the proposed NAS scheme, a specific architecture is automatically searched, named as MSINet. Extensive experiments demonstrate that our method surpasses state-of-the-art ReID methods on both indomain and cross-domain scenarios. Source code available in https://github.com/vimar-gu/MSINet.", + "bbox": [ + 75, + 365, + 472, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 695, + 206, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object re-identification (Re-ID) aims at retrieving specific object instances across different views [39, 40, 57, 65, 70], which attracts much attention in computer vision community due to its wide-range applications. Previous works have achieved great progresses on both supervised [42, 49, 58] and unsupervised ReID tasks [17, 50, 78], most of which adopts backbone models originally designed for general image classification tasks [20, 52].", + "bbox": [ + 75, + 719, + 468, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent literature [64, 75] has shown that applying differ", + "bbox": [ + 95, + 831, + 468, + 845 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f45c96a1a85b6862cf03b220f4f50423d45eb5257094decaf6cc688540b11d7d.jpg", + "image_caption": [ + "hood" + ], + "image_footnote": [], + "bbox": [ + 504, + 336, + 608, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d56fa0f9226451dc5a875aef99222e193cb7719a117b1782c25cb1a45a3ea4ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 337, + 769, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79556eb58cd1edc26469fa6bae015918a798943de8b82f5f7c567461ad4b04bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 338, + 880, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/054c15f3b8ca7fd16fe1809275a1d8dcdfccbebd0c8bc0165c7ae1f9f5fbbf3e.jpg", + "image_caption": [ + "luggage carrier", + "Figure 1. The left panel shows the example activation maps of ResNet50 (1st row) and MSINet (2nd row). The right panel shows the average distances between the most similar 10 negative samples and each query image at the inference. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 502, + 413, + 606, + 490 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87ea7dbb407db1dfdbec3588fa670b59bf88a178a381deb80799c4a7e2522fad.jpg", + "image_caption": [ + "Distance between top-10 neg. samples and each query image" + ], + "image_footnote": [], + "bbox": [ + 667, + 415, + 769, + 490 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ment architectures on ReID leads to large performance variations. Some works employ Neural Architecture Search (NAS) for ReID [28, 45]. The proposed optimizing targets and search spaces stably improve the model performance, yet the main search scheme still follows traditional NAS methods designed for general classification tasks [12, 36]. As an open-set task, ReID contains different categories in the training and validation sets [64, 71], while the two sets share exactly the same categories in standard classification tasks [10], which is also followed by traditional NAS methods. The incompatibility between search schemes and real-world training schemes makes the searched architecture sub-optimal for ReID. Moreover, ReID is required to distinguish more subtle distinctions among fine-grained instances compared with image-level classification [48, 63]. Some previous works [4, 44, 68, 75] have manifested that local perspectives and multi-scale features are discriminative for ReID. However, current utilizations of these features are mostly empirically designed, which can be more flexible according to the characteristics of different network layers.", + "bbox": [ + 497, + 589, + 890, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 84 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Co-corresponding authors", + "bbox": [ + 95, + 853, + 240, + 865 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19243", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a novel NAS scheme aiming at addressing the aforementioned challenges. In order to simulate the real-world ReID training schemes, a Twins Contrastive Mechanism (TCM) is proposed to unbind the categories of the training and validation sets. An adjustable overlap ratio of categories builds up the compatibility between NAS and ReID, which provides more appropriate supervision for ReID architecture search. Moreover, to search for more rational utilizations of multi-scale features, we design a Multi-Scale Interaction (MSI) search space. The MSI space focuses on interaction operations between multi-scale features along the shallow and deep layers of the network, which guides the features to promote each other. Additionally, to further improve the generalization capability, we propose a Spatial Alignment Module (SAM) to enhance the attention consistency of the model confronted with images from different sources. With the above NAS scheme, we obtain a light-weight yet effective model architecture, denoted as Multi-Scale Interaction Net (MSINet).", + "bbox": [ + 75, + 125, + 472, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We visualize the example activation maps of our proposed MSINet and ResNet50 [20] trained on VeRi-776 [38, 39] in Fig. 1. Compared to ResNet50, MSINet focuses on more unique distinctions with specific semantic information to recognize instances. Besides, MSINet largely increases the distance margin between query image and corresponding negative samples, reflecting extraordinary discriminative capability. Extensive experiments demonstrate that MSINet surpasses state-of-the-art (SOTA) ReID methods on both in-domain and cross-domain scenarios. Our source codes are available in the supplementary material.", + "bbox": [ + 75, + 388, + 472, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 95, + 541, + 400, + 553 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to build the NAS search scheme according to the real-world ReID training schemes, which provides more appropriate supervision for the ReID architecture search.", + "- We propose a novel search space based on the Multi-Scale Interaction (MSI) operations and a Spatial Alignment Module (SAM) to improve the model performance on in-domain and cross-domain scenarios.", + "- We construct a light-weight yet effective architecture for ReID tasks, denoted as MSINet. With only 2.3M parameters, MSINet surpasses ResNet50 [20] by $9\\%$ mAP on MSMT17 [60] and $16\\%$ mAP on MSMT17 $\\rightarrow$ Market-1501 [69]." + ], + "bbox": [ + 95, + 562, + 468, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 77, + 773, + 226, + 788 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Architecture Search. NAS has been increasingly appealing to the computer vision society, due to its automatic architecture designing characteristics. NAS methods can be roughly separated into four categories: reinforcement learning [1, 77], evolutionary algorithms [35, 46], gra", + "bbox": [ + 75, + 796, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dient descent [36, 43] and performance prediction [11, 34]. Liu et al. establish a differentiable architecture search (DARTS) method [36], which improves the practicability of NAS by a large extent. Some later works further improve the structure through sampling strategy [62], network pruning [2, 8], progressive learning [5], collaborative competition [7], etc. Most of NAS works focus on general image classification tasks, where the training and validation sets share the exact same categories. Following the setting, however, leads to incompatibility with the real-world training schemes of object ReID. In this work, we unbind the category bond between the two sets and propose a novel search scheme suitable for ReID.", + "bbox": [ + 497, + 125, + 890, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ReID Network Design. Current ReID works mostly adopt backbones designed for image classification [20, 23, 47, 52]. Some works [14, 31, 67] design attention modules based on the common backbones to unearth their potential on distinguishing local distinctions. However, these methods usually lead to large calculation consumption.", + "bbox": [ + 497, + 305, + 890, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are also several works focusing on designing ReID-specific architectures. Li et al. present a Filter Pairing Neural Network to dynamically match patches in the feature maps [30]. Wang et al. separate and regroup the features of two samples with a WConv layer [56]. Guo et al. extract multi-scale features to directly evaluate the similarity between samples [18]. However, the siamese structure is inconvenient when conducting retrieval on large galleries. Zhou et al. aggregate multi-scale information to achieve high accuracy with small computing consumption [75]. Quan et al. introduce a part-aware module into the DARTS search space [36, 45]. Li et al. propose a new search space in regard to receptive field scales [28]. These methods have excellent performance on limited parameter scales, but fail to surpass those networks with complex structures. Different from previous works, we design a light-weight searching structure focusing on rational interaction operations between multi-scale features. The searched MSINet surpasses SOTA methods on both in-domain and cross-domain tasks.", + "bbox": [ + 497, + 388, + 892, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 499, + 661, + 598, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to construct an effective NAS scheme to search for a light-weight backbone architecture suitable for ReID tasks. Based on the training schemes of ReID, we propose a novel Twins Contrastive Mechanism to provide more appropriate supervision for the search process. Aiming at rational interaction between multi-scale features, we design a Multi-Scale Interaction search space. We further introduce a Spatial Alignment Module to improve the generalization capability with limited parameter growth.", + "bbox": [ + 497, + 685, + 890, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Twins Contrastive Mechanism", + "text_level": 1, + "bbox": [ + 499, + 817, + 769, + 831 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NAS aims at automatically searching for the optimal network architecture for certain data. Inspired by [36], a basic", + "bbox": [ + 499, + 838, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19244", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93629ce15e822e4f454f33347d05317c4ccacf32239c4a665cee6a5fbe229fd9.jpg", + "image_caption": [ + "Figure 2. The model structure of the proposed MSINet. The input can be either person or vehicle. Inside a cell, the input is separated to two branches, with different receptive field scales. The interaction module exchanges information between two branches. Architecture search automatically selects the most appropriate interaction for each cell." + ], + "image_footnote": [], + "bbox": [ + 102, + 123, + 868, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "differentiable architecture search scheme is established. We define the ordinary model parameters as $\\omega$ , and architecture parameters as $\\alpha$ . For network layer $i$ with a search space of $\\mathcal{O}$ , $\\alpha_{i}$ controls the weight of each operation $o$ in the space. The features are parallelly passed through all the operations, and the final output is formulated by the softmax-weighted sum of operation outputs:", + "bbox": [ + 75, + 359, + 472, + 458 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\mathbf {x} _ {i}\\right) = \\sum_ {o \\in \\mathcal {O}} \\frac {\\exp \\left\\{\\alpha_ {i} ^ {o} \\right\\}}{\\sum_ {o ^ {\\prime} \\in \\mathcal {O}} \\exp \\left\\{\\alpha_ {i} ^ {o ^ {\\prime}} \\right\\}} \\cdot o \\left(\\mathbf {x} _ {i}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 134, + 469, + 468, + 509 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The search process is conducted in an alternative manner. Training data is utilized to update the model parameters, and validation data is then employed to update the architecture parameters. For most NAS methods designed for image classification tasks, the training and validation data share exactly the same categories and a linear classification layer for loss calculation.", + "bbox": [ + 75, + 519, + 468, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from standard image classification, as an open-set retrieval task, ReID has different categories in the training and validation sets. The incompatibility between search schemes and real-world training schemes might lead to suboptimal searching results. Accordingly, we propose a novel Twins Contrastive Mechanism (TCM) for NAS ReID training. Specifically, we employ two independent auxiliary memories $\\mathcal{C}_{tr}$ and $\\mathcal{C}_{val}$ to store the embedded features of the training and validation data, respectively. The memories are initialized with the centroid features, which are calculated by averaging the features of each category. At each iteration, the training loss is first calculated with $\\mathcal{C}_{tr}$ for model parameter updating. Given an embedded feature $f$ with category label $j$ , the contrastive classification loss is calculated with:", + "bbox": [ + 75, + 617, + 470, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t r} ^ {c l s} = - \\log \\frac {\\exp (\\mathbf {f} \\cdot \\mathbf {c} _ {t r} ^ {j} / \\tau)}{\\sum_ {n = 0} ^ {N _ {t r} ^ {c}} \\exp (\\mathbf {f} \\cdot \\mathbf {c} _ {t r} ^ {n} / \\tau)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 149, + 834, + 468, + 871 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_{tr}^n$ represents the memorized feature of category $n$ , $N_{tr}^c$ stands for the total number of categories in the training set, and $\\tau$ is the temperature parameter, which is set as 0.05 empirically [17]. After updating the model parameters, the embedded feature $\\mathbf{f}$ with category label $j$ is integrated into the corresponding memorized feature $\\mathbf{c}_{tr}^j$ by:", + "bbox": [ + 497, + 359, + 890, + 445 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {t r} ^ {j} \\leftarrow \\beta \\mathbf {c} _ {t r} ^ {j} + (1 - \\beta) \\mathbf {f}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 452, + 890, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta$ is set as 0.2 empirically [17]. Then the updated model is evaluated on the validation data to generate to validation loss with $C_{val}$ replacing $C_{tr}$ in Eq. 2. The architecture parameter is then updated with the validation loss to finish an iteration.", + "bbox": [ + 497, + 477, + 890, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the loss calculation does not rely on the linear classification layer, the categories of the training and validation sets are unbound. We are able to dynamically adjust the category overlap ratio in these two sets. The advantages of a proper overlap ratio are summarized as two folds. Firstly, TCM better simulates the real-world training of ReID and helps the model focus on truly discriminative distinctions. The differences between the training and validation data improves the generalization capability of the model. Secondly, a relatively small proportion of overlapped categories stabilizes the architecture parameter update through a consistent optimizing target with the model parameter update.", + "bbox": [ + 497, + 547, + 892, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Scale Interaction Space", + "text_level": 1, + "bbox": [ + 499, + 720, + 764, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although the local perspective and multi-scale features have already been investigated in previous ReID works [4, 28, 44, 51, 68, 72, 75, 76], the utilization of these information is mainly empirically designed aggregation, which is monotonous and restrained. We argue that on the one hand, the rational utilization of multi-scale features should be dynamically adjusted along the shallow and deep layers of the network. On the other hand, introducing interaction other than aggregation creates direct information exchange, and", + "bbox": [ + 497, + 741, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19245", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5523ff0813347b36722b461219b27c29c12088ba4bdffab88dbf652e82b6f9aa.jpg", + "image_caption": [ + "Figure 3. The aligning pipeline of the proposed Spatial Alignment Module. The correlation activation vectors are calculated between the anchor feature and all features in a mini-batch. The positive vectors are aligned with the learnable self-activation, and the negative vectors are aligned with each other. The structure of PAM is shown on the left." + ], + "image_footnote": [], + "bbox": [ + 100, + 123, + 868, + 280 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/09534fbfebbdde93fedc1e8704b85ddf408432c67792b2bf91da8410fb178579.jpg", + "table_caption": [ + "Table 1. The detailed interaction operation in the proposed MSINet architecture. N: None; E: Exchange; G: Channel Gate; A: Cross Attention." + ], + "table_footnote": [], + "table_body": "
Cell #1Cell #2Cell #3Cell #4Cell #5Cell #6
123456789101112
GGEGAGGNGAEA
", + "bbox": [ + 78, + 400, + 467, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "makes fuller use of multi-scale features. Therefore, we propose a novel Multi-Scale Interaction (MSI) search space to establish a light-weight architecture suitable for ReID.", + "bbox": [ + 75, + 473, + 468, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the network is mainly grouped with MSI cells and down-sample blocks, which is generally consistent with OSNet [75]. In each cell, the input features are passed through two branches with different receptive field scales. To reduce the calculation burden of the network, for the layers inside each branch, we adopt the stack of $1 \\times 1$ convolution and multiple depth-wise $3 \\times 3$ convolution to implement specific scales. A scale ratio $\\rho$ of 3:1 is selected for the two branches. These two branches do not share model parameters, except for the Interaction Modules (IM). IM introduces information exchange for the two branches. There are 4 operation options for the IM. With the two-branch input features defined as $(\\mathbf{x}_1, \\mathbf{x}_2)$ , the operations can be formulated as:", + "bbox": [ + 75, + 516, + 468, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "None. None operation involves no parameters, and outputs exactly the input features $(\\mathbf{x}_1,\\mathbf{x}_2)$ .", + "bbox": [ + 75, + 711, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exchange. Exchange acts as the strongest interaction among all options. It directly exchanges the features for the two branches and outputs $(\\mathbf{x}_2, \\mathbf{x}_1)$ . Exchange contains no extra parameters, as well.", + "bbox": [ + 75, + 739, + 468, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Channel Gate. Channel gate introduces a Multi-Layer Perceptron (MLP) to generate a channel-wise attention gate [61, 75] as:", + "bbox": [ + 75, + 796, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (\\mathbf {x}) = \\sigma (M L P (\\mathbf {x})), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 852, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d19bcb32e290dde3c7cf4dc867064efe3f68de7668d2d5791e3709501adffe4.jpg", + "image_caption": [ + "Figure 4. Output feature maps of MSI cells at different layers." + ], + "image_footnote": [], + "bbox": [ + 505, + 350, + 905, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and returns $(G(\\mathbf{x}_1) \\cdot \\mathbf{x}_1, G(\\mathbf{x}_2) \\cdot \\mathbf{x}_2)$ . The MLP is composed of 2 fully connected layers and its parameters are shared for both branches. Thereby it achieves interaction by jointly screening discriminative feature channels.", + "bbox": [ + 497, + 491, + 890, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross Attention. Traditional channel attention module calculates the channel correlation inside a single feature map [15]. The original feature map $\\mathbf{x} \\in \\mathcal{R}^{C \\times H \\times W}$ is firstly reshaped into the query feature $\\tilde{\\mathbf{x}} \\in \\mathcal{R}^{C \\times N}$ , where $N = H \\times W$ . Then the correlation activation is calculated by performing a matrix multiplication between the query feature $\\tilde{\\mathbf{x}}$ and the key feature $\\tilde{\\mathbf{x}}^{\\top}$ . We propose to exchange the keys of the two branches to explicitly calculate the correlation between each other. The correlation activation is then transformed to a mask, and is added up to the original features with a learnable proportion.", + "bbox": [ + 497, + 547, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After interaction, the multi-scale branches are fused through a sum operation. It is worth noting that the extra parameters brought by multiple interaction options are limited, which enables searching for each cell along the whole network independently. At the beginning of the network, we employ the same stem module as that in OSNet [75], containing a $7 \\times 7$ convolutional layer and a $3 \\times 3$ max pooling with a stride of 2. After the searching process, the interaction operation $o$ with the largest weight $\\alpha_{i}^{o}$ at each layer is reserved to form the searched architecture.", + "bbox": [ + 497, + 699, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After searching the architecture, the model is validated on various Re-ID tasks. The training is constrained by the", + "bbox": [ + 497, + 838, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19246", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1783de226bd1396dd2f95ee7a8c219fa06e0b94a2d000fac1d24cc9b112ec0fe.jpg", + "table_caption": [ + "Table 2. Supervised performance on object ReID datasets. The results in the top part are trained from scratch, and those in the bottom part are pre-trained on ImageNet in advance. As the compared methods are originally proposed for person ReID, we reproduce the results in vehicle datasets. * indicates that the results of person ReID are reproduced by us. The evaluation results of architecture searched on VR can be found in the supplementary material." + ], + "table_footnote": [], + "table_body": "
MethodParamsInference TimeMMSVRVIDMS→MVR→VID
R-1↑mAP↑R-1↑mAP↑R-1↑mAP↑R-1↑R-5↑R-1↑mAP↑R-1↑R-5↑
ResNet50* [42]~24M1×85.768.348.025.792.869.970.676.6--
OSNet [75]2.2M0.79×93.681.071.043.395.472.876.088.7----
CDNet [28]1.8M0.67×93.783.773.748.594.373.074.588.8----
MSINet2.3M0.71×94.687.076.052.595.975.076.589.8----
ResNet50* [42]~24M1×94.585.975.550.494.573.676.589.958.831.842.861.9
OSNet [75]2.2M0.79×94.884.978.752.995.576.476.088.666.637.546.563.1
CDNet [28]1.8M0.67×95.186.078.954.7--------
MSINet2.3M0.71×95.389.681.059.696.878.877.991.774.946.248.065.6
MSINet-SAM2.4M0.71×95.589.980.759.596.779.078.091.976.348.449.066.8
", + "bbox": [ + 95, + 185, + 875, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "classification id loss and the triplet loss, formulated by:", + "bbox": [ + 75, + 380, + 442, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i d} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - \\log \\left(\\frac {\\exp \\mathbf {W} _ {i} ^ {\\top} \\mathbf {f} _ {i}}{\\sum_ {j} \\exp \\mathbf {W} _ {j} ^ {\\top} \\mathbf {f} _ {i}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 404, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_i$ is a feature vector, the corresponding classifier weight of which is $\\mathbf{W}_i$ , and", + "bbox": [ + 75, + 451, + 468, + 480 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t r i} = \\left[ \\mathcal {D} \\left(\\mathbf {f} _ {a}, \\mathbf {f} _ {p}\\right) - \\mathcal {D} \\left(\\mathbf {f} _ {a}, \\mathbf {f} _ {n}\\right) + m \\right] _ {+}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 144, + 490, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_a, \\mathbf{f}_p, \\mathbf{f}_n$ are the embedded features for the anchor, the hardest positive and negative samples in a mini-batch, $\\mathcal{D}(\\cdot, \\cdot)$ is the Euclidean distance, $m$ is the margin parameter, and $[\\cdot]_{+}$ is the $\\max(\\cdot, 0)$ function.", + "bbox": [ + 75, + 516, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Spatial Alignment Module", + "text_level": 1, + "bbox": [ + 77, + 580, + 315, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The retrieval precision of object ReID tasks are largely affected by the variation of appearances such as poses, illumination and occlusion when the camera conditions change. In order that the model correctly and consistently focuses on the discriminative spatial positions, we design a Spatial Alignment Module (SAM) to explicitly align the spatial attention between images, as shown in Fig. 3.", + "bbox": [ + 75, + 602, + 468, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first calculate the position-wise correlation activation map $\\mathbf{A}$ between the feature maps in a minibatch. The activation between sample $i$ and $j$ can be formulated as: $\\mathbf{A}(i,j) = \\tilde{\\mathbf{x}}_j^\\top \\times \\tilde{\\mathbf{x}}_i$ , where $\\tilde{\\mathbf{x}} \\in \\mathcal{R}^{C \\times N}$ is reshaped from the original feature $\\mathbf{x} \\in \\mathcal{R}^{C \\times H \\times W}$ . Then we take the maximum activation for each position of sample $i$ as:", + "bbox": [ + 75, + 699, + 468, + 783 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} (i, j) = \\max _ {d i m = 1} \\mathbf {A} (i, j). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 794, + 468, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above process is denoted as \"Mutual Conv\" in Fig. 3. We evaluate the consistency between activation vectors with cosine similarity. For negative samples specifically, there", + "bbox": [ + 75, + 824, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can be many different hints for recognition, some of which might be inappropriate, such as the backgrounds. By aligning all the correlations for sample $i$ , we hope that the network can correct some attention bias and consistently focus on discriminative positions.", + "bbox": [ + 497, + 380, + 890, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, through aligning positive sample pairs, the ID-related features are expected to be emphasized, which cannot be achieved by aligning negative pairs. Therefore, we introduce an extra position activation module (PAM) to generate supervision for the alignment between positive pairs. The spatial alignment loss is formulated as:", + "bbox": [ + 497, + 451, + 892, + 534 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s a} (i) = \\frac {1}{N _ {+}} \\sum_ {p \\in \\mathcal {I} _ {+}} \\left(1 - S (\\hat {\\mathbf {a}} (i), \\mathbf {a} (i, p))\\right) + \\\\ \\frac {1}{N _ {-}} \\sum_ {n _ {1}, n _ {2} \\in \\mathcal {I} _ {-}} (1 - S (\\mathbf {a} (i, n _ {1}), \\mathbf {a} (i, n _ {2}))), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 547, + 842, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(8)", + "bbox": [ + 867, + 578, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_{+}$ contains positive indices for sample $i$ , the total number of which is $N_{+}$ , and vice versa. $\\hat{\\mathbf{a}}(i)$ stands for the generated activation vector for positive sample alignment, and $S(\\cdot, \\cdot)$ is the cosine similarity.", + "bbox": [ + 497, + 633, + 892, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 499, + 707, + 633, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 499, + 732, + 786, + 745 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed method is evaluated on two person ReID datasets Market-1501 [69], MSMT17 [60], and two vehicle ReID datasets VeRi-776 [38, 39] and VehicleID [37]. For simplicity, the four datasets are denoted as M, MS, VR and VID in the following sections, respectively. Evaluation metrics include Cumulative Matching Characteristic (CMC) and mean average precision (mAP), which are commonly utilized on ReID tasks.", + "bbox": [ + 497, + 755, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19247", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e3eeeaf9c314078637d834686c26ffdf1a3fa8e69ce511c4fe7d50a6d4996ce5.jpg", + "table_caption": [ + "Table 3. Supervised performance comparison between MSINet and SOTA methods on M and MS datasets." + ], + "table_footnote": [], + "table_body": "
MethodMMS
R-1↑mAP↑R-1↑mAP↑
PCB [51]93.881.668.240.4
MGN [55]95.786.976.952.1
OSNet [75]93.681.071.043.3
IANet [22]94.483.175.546.8
DGNet [73]94.886.077.252.3
Auto-ReID [45]94.585.1--
SAN [26]96.188.079.255.7
CDNet [28]95.186.078.954.7
BAT-Net [14]95.187.479.556.8
SFT [41]94.187.579.058.3
CTF [66]94.887.7--
RGA-SC [67]96.188.480.357.5
MSINet95.389.681.059.6
", + "bbox": [ + 122, + 159, + 425, + 365 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Architecture Search", + "text_level": 1, + "bbox": [ + 77, + 387, + 265, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct the searching process on MSMT17. SGD is adopted for model parameter update with an initial learning rate of 0.025. The model is trained for 350 epochs in total. We adopt a warm-up strategy for the first 10 epochs. Then the learning rate is decayed by 0.1 at 150, 225 and 300 epochs, respectively. Adam [27] is adopted for the architecture parameter update with an initial learning rate of 0.002. The learning rate is decayed at the same pace. The images are reshaped to $256 \\times 128$ for person and $256 \\times 256$ for vehicles. Data augmentation includes random flip, random crop and random erasing [74]. The searched architecture is presented in Tab. 1. The \"MSINet\" in the following experiment sections refers to this architecture.", + "bbox": [ + 75, + 408, + 467, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We visualize the feature maps extracted by each MSI cell in Fig. 4. At the shallow layers of the network, the kernels mainly focus on overall contour information. Channel gate helps to filter out inferior information, such as the background. As we approach deeper layers, the extracted features each have specific semantic information, where cross attention is more likely to be selected for the interaction. It indicates that cross attention is more rational for exchanging high-level semantic information.", + "bbox": [ + 75, + 587, + 467, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with Other Backbones", + "text_level": 1, + "bbox": [ + 77, + 720, + 381, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare our proposed MSINet with ResNet50 and recent proposed light-weight backbones in both in-domain and cross-domain ReID scenarios.", + "bbox": [ + 77, + 742, + 467, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-Domain Tasks. We adopt a two-group supervised evaluation scheme similar to that in [28, 75]: training from scratch and fine-tuning ImageNet [10] pre-trained models. The training parameters for both schemes are kept the same as that in architecture search, except for an initial learning rate of 0.065. Triplet loss and cross entropy loss are adopted", + "bbox": [ + 77, + 783, + 467, + 865 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4ab284cb5208d20a3833265493029730dec2eb4a10b14c5bfc0638a0e9e789e0.jpg", + "table_caption": [ + "Table 4. Unsupervised performance applying MSINet to SOTA methods for USL on M and UDA on $\\mathrm{M} \\rightarrow \\mathrm{{MS}}$ ." + ], + "table_footnote": [], + "table_body": "
MethodMM→MS
R-1↑mAP↑R-1↑mAP↑
MMCL [54]80.345.540.815.1
MMT [16]--50.124.0
JVTC+ [29]79.547.548.625.1
CycAs [59]84.864.8--
GCL [3]87.366.851.127.0
MPRD [24]83.051.0--
SpCL [17]88.173.153.726.8
GS [19]92.379.2--
GS+MSINet91.781.5--
HDCRL [6]92.481.7--
HDCRL+MSINet92.982.7--
IDM [9]--61.333.5
IDM+MSINet--66.037.8
", + "bbox": [ + 542, + 160, + 848, + 365 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the parameter update. The margin $m$ in Eq. 6 is set as 0.3. [28] adopts an FBLNeck. We also employ the same structure. The results are shown in Tab. 2.", + "bbox": [ + 499, + 391, + 890, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ResNet50 is the most commonly utilized backbone network in ReID tasks, yet holds the worst performance. Moreover, ResNet50 largely depends on ImageNet pre-training, while MSINet without pre-training has already surpassed pre-trained ResNet50 on all metrics. Compared with the other datasets, MS contains more variations on illumination, background and camera pose, and brings a large performance gap between ResNet50 and other methods. It also validates the inadequacy of image classification networks on ReID tasks. OSNet [75] and CDNet [28] are recently proposed architectures designed specifically for ReID tasks. Both architectures focus on fusing multi-scale features to better suit ReID. CDNet employs a traditional NAS scheme to search for the proper receptive field scales for each cell. MSINet fixes the receptive field scale and instead selects optimal interaction operations inside each cell. With only a bit more parameters, MSINet surpasses all the other backbones by a large margin.", + "bbox": [ + 499, + 434, + 890, + 682 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-Domain Tasks. Cross-domain experiments verify the generalization capability of the model. Following previous domain generalizable ReID works [25, 33], data augmentation is adjusted to random flip, random crop and color jittering. The model is pre-trained and fine-tuned for 250 epochs to avoid over-fitting. The other settings are kept the same as supervision scenes. With no present pre-trained models for CDNet [28], it is excluded from this section.", + "bbox": [ + 499, + 685, + 890, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 shows that ResNet50 can be easily interfered by different image styles confronted with new image domains. OSNet learns multi-scale features with specific semantic information for ReID, which is domain invariant to some extent. Our proposed search scheme also takes into ac", + "bbox": [ + 499, + 796, + 890, + 865 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19248", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51c2fbd7e148acffa94d21e9d5888f89d2f1d69179c45870fdab237151ac504b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 90, + 125, + 243, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee5a3bfa957f5cbf48be410b91cbae67f51a37e44d21cae0ad9d9e3e08e97fcd.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 248, + 127, + 408, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fb17dd76fc40c0ad73e6f3fdeb75235051a3d3d4eaf379aedcc200ae10e1eb1.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 413, + 125, + 561, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84d3e0d04855a8721be3990a139f9aa43aa4625c7d735ac08847e1c2b9bc49d8.jpg", + "image_caption": [ + "(d)", + "Figure 5. Ablation studies on (a) search scheme; (b) architecture; (c) multi-scale aggregation; (d) scale ratio $\\rho$ ; (e) spatial alignment weight." + ], + "image_footnote": [], + "bbox": [ + 566, + 125, + 712, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/480a979402b8bf2edbe704b3d93a387f6878636a7f76d1d2ab31c89b772c7e41.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 715, + 125, + 877, + 257 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5def09c6db2f2dd8ba720a3916a0266860d01e6ac321479a87edc347bff5814.jpg", + "table_caption": [ + "Table 5. The effectiveness of each components in SAM. Both in-domain performance on VR and cross-domain performance on $\\mathrm{VR}\\rightarrow \\mathrm{VID}$ is evaluated." + ], + "table_footnote": [], + "table_body": "
ModelPos.Neg.AlignVRVR→VID
R-1↑mAP↑R-1↑R-5↑
MSINet✓-96.778.548.065.6
Self96.778.148.466.0
✓Self96.778.448.566.2
✓✓Unified96.678.348.365.8
✓✓Separated96.878.648.766.4
✓✓PAM-Self96.779.049.066.8
OSNet✓✓None95.576.446.563.1
PAM-Self95.976.347.563.3
", + "bbox": [ + 80, + 363, + 465, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "count the generalization capability of the model. By partly separating the categories for training and validation sets, the searched interaction operations generalize well confronted with new image domains. Except for discrimination, MSINet also surpasses the other backbones on cross-domain tasks by a large margin with faster inference speed.", + "bbox": [ + 75, + 536, + 468, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we introduce SAM into the model, which aligns the spatial correlations between images. A weighted sum of ReID loss and spatial alignment loss is utilized when training the network with SAM. The weight of spatial alignment loss is set as $\\lambda_{sa} = 2.0$ . Without extra inference consumption or damages on the supervised performance, SAM further boosts the generalization capability of MSINet.", + "bbox": [ + 75, + 622, + 468, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 75, + 732, + 440, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 further illustrates the supervision performance comparison of our proposed MSINet with the SOTA methods on M and MS datasets. With much less parameters than most of the compared methods, MSINet achieves a retrieval accuracy comparable to that of more complicated ones. Auto-ReID [45] first designs a NAS scheme for ReID, yet the DARTS-style architecture contains 13M parameters. RGA-SC [67] carefully designs a relation-aware global at", + "bbox": [ + 75, + 755, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tention module. MSINet achieves even higher performance with less training consumption, which validates the superiority of selecting rational interaction.", + "bbox": [ + 497, + 317, + 890, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate the model performance replacing the backbone network from ResNet50 to MSINet for SOTA unsupervised ReID methods in Tab. 4. For purely unsupervised learning (USL) method GS [19] on M dataset, MSINet performs slightly lower on rank-1, yet has a large superiority on mAP. For HDCRL [6], MSINet shows obvious superiority over ResNet50. For unsupervised domain adaptation (UDA) method IDM [9] on $\\mathrm{M}\\rightarrow \\mathrm{MS}$ task, MSINet surpasses ResNet50 by a large margin, which further proves that the TCM brings outstanding generalization capability to the searched architecture.", + "bbox": [ + 497, + 359, + 892, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 499, + 524, + 663, + 538 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Architecture Search. To verify the effectiveness, we conduct supervised training on MS with different search schemes in Fig. 5 (a). Under the standard classification scheme (\"CE Overlap\"), the searched model performs poor. Replacing the cross entropy loss to contrastive loss (\"TCM Overlap\") only brings a slight improvement. The complete TCM framework unbounds the categories between the training and validation set, and thereby improves the performance by a large margin.", + "bbox": [ + 497, + 547, + 890, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 5 (b) we compare the performance of different architectures. Firstly, we validate 4 models each with a unique interaction operation from the 4 options in the search space. None and Exchange, with no trainable parameters, achieve poor performance. Channel gate introduces channel-wise attention, whose model performs the best among 4 options. Cross attention exchanges the key features for the two branches. Over-frequent exchange interferes the ordinary feature extraction and degrades the network performance. Through appropriately arranging interaction operations along the architecture, MSINet surpasses all the above 4 models. Random architecture, on the other hand, shows no rational appliance of interaction operations, which validates that the proposed search scheme helps find", + "bbox": [ + 497, + 672, + 892, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19249", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d4a46fc93d9d7ed3c58dc330b48071522826c61fd30a85ba9439a616033a4c84.jpg", + "table_caption": [ + "Table 6. Supervised performance comparison between MSINet and Transformers on VR and MS datasets." + ], + "table_footnote": [], + "table_body": "
MethodParamsInference TimeMSVR
R-1↑mAP↑R-1↑mAP↑
DeiT-S [21]~22M0.97x76.355.295.576.3
DeiT-B [21]~86M1.79x81.961.495.978.4
ViT-B [21]~86M1.79x81.861.096.578.2
MSINet2.3M0.71x81.059.696.878.8
", + "bbox": [ + 78, + 159, + 468, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8316f43b4fa82cca870622977851a032b0a78b4336f37cc468b944698b660efd.jpg", + "image_caption": [ + "Figure 6. Example top-15 retrieved sequences comparison on VR. Appearance differences caused by variant camera conditions are well addressed by the proposed MSINet. Visualization of person ReID can be found in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 80, + 263, + 463, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "suitable architectures for ReID.", + "bbox": [ + 75, + 491, + 285, + 504 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Spatial Alignment Module. We validate the effectiveness of each components of SAM on the VR $\\rightarrow$ VID cross-domain experiment in Tab. 5. Firstly, we introduce the spatial alignment for positive and negative sample pairs, respectively. Each of them brings certain performance improvements. However, a unified alignment for all sample pairs damages ID-related features and degrades the performance instead. Therefore, we separate the alignment of positive and negative samples, which retains some discriminative features and integrates the effect of both aspects. The extra PAM for positive sample alignment further guarantees the focus on ID-related positions and achieves the best performance. We also conduct in-domain experiment on VR to prove that SAM improves the generalization capability without sacrificing the supervision performance. Adding SAM to OSNet receives similar results, which validates the universality of SAM.", + "bbox": [ + 75, + 505, + 468, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fusing Operation. After interaction, the multi-scale features are fused by sum operation. We investigate several fusing options on MS training from scratch in Fig. 5 (c). Subtracting (\"Minus\") a branch from the other leads to about the same results as \"Sum\" while multiplication (\"Mul\") performs poorly.", + "bbox": [ + 75, + 741, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with Transformer. Transformer, as a new architecture, has recently been continuously making progresses in many computer vision domains [13, 53], includ", + "bbox": [ + 75, + 825, + 468, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing ReID [21, 32]. We compare the model performance with some baseline Transformer models in Tab. 6. DeiT-B and ViT-B [21] achieves higher performance on MSMT17, with much larger calculation burden compared with our proposed MSINet. On VeRi-776, MSINet surpasses all the baseline Transformer methods. It proves that rational interaction operations between multi-scale features are capable to assist light-weighted pure-CNN models to obtain comparable performance with complex Transformers.", + "bbox": [ + 497, + 125, + 890, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter Analysis. Firstly, we study the influence of different receptive field scale ratios $\\rho$ inside an MSI cell on MS training from scratch in Fig. 5 (d). Introducing scale differences between branches improves the model performance significantly, and subsequent increases brings more modest impacts. Considering both parameter scales and model performance, the ratio of 3:1 is selected for MSINet.", + "bbox": [ + 497, + 250, + 892, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Secondly, the model performance fluctuation influenced by spatial alignment weight is visualized in Fig. 5 (e). The experiment is conducted on the VR $\\rightarrow$ VID cross-domain scenario. Employing the alignment generally makes a positive impact on the generalization capability of the model. The optimal loss weight $\\lambda_{sa}$ locates at 2.0.", + "bbox": [ + 497, + 347, + 890, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization Results. We visualize the top-15 retrieved sequences and the corresponding distances from an query image on VR in Fig. 6. By comparison, ResNet50 mainly focuses on general appearance features, where the top-rated negative samples share similar car bodies. MSINet, oppositely, concentrates on discriminative distinctions, an empty car hopper in this case, and creates an evident distance gap between positive and negative samples. More details can be seen in the supplementation material.", + "bbox": [ + 497, + 431, + 890, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 568, + 616, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we design a Twins Contrastive Mechanism for NAS to build the compatibility with ReID. The task-specific search scheme provides the searching process with more appropriate supervision. A Multi-Scale Interaction search space is proposed to establish rational and flexible utilization of multi-scale features. With a Spatial Alignment Module, our proposed MSINet achieves SOTA performance on both supervision and cross-domain scenarios with limited parameter amount. We hope the proposed approach could inspire more works focusing on designing network architectures suitable for ReID tasks.", + "bbox": [ + 497, + 593, + 890, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 499, + 758, + 678, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by National Natural Science Foundation of China (62173302, 62006244), China Scholarship Council (202206320302), Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and AI Singapore Programme (AISG2-PhD-2021-08-008).", + "bbox": [ + 497, + 783, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19250", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 171, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Irwan Bello, Barret Zoph, Vijay Vasudevan, and Quoc V Le. Neural optimizer search with reinforcement learning. In ICML, pages 459-468, 2017. 2", + "[2] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. In ICLR, 2018. 2", + "[3] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Joint generative and contrastive learning for unsupervised person re-identification. In CVPR, pages 2004–2013, 2021. 6", + "[4] Tsai-Shien Chen, Chih-Ting Liu, Chih-Wei Wu, and Shao-Yi Chien. Orientation-aware vehicle re-identification with semantics-guided part attention network. In ECCV, pages 330–346, 2020. 1, 3", + "[5] Xin Chen, Lingxi Xie, Jun Wu, and Qi Tian. Progressive differentiable architecture search: Bridging the depth gap between search and evaluation. In ICCV, pages 1294-1303, 2019. 2", + "[6] De Cheng, Jingyu Zhou, Nannan Wang, and Xinbo Gao. Hybrid dynamic contrast and probability distillation for unsupervised person re-id. IEEE Transactions on Image Processing, 31:3334-3346, 2022. 6, 7", + "[7] Xiangxiang Chu, Tianbao Zhou, Bo Zhang, and Jixiang Li. Fair darts: Eliminating unfair advantages in differentiable architecture search. In ECCV, pages 465-480, 2020. 2", + "[8] Xiyang Dai, Dongdong Chen, Mengchen Liu, Yinpeng Chen, and Lu Yuan. Da-nas: Data adapted pruning for efficient neural architecture search. In ECCV, pages 584-600, 2020. 2", + "[9] Yongxing Dai, Jun Liu, Yifan Sun, Zekun Tong, Chi Zhang, and Ling-Yu Duan. Idm: An intermediate domain module for domain adaptive person re-id. In ICCV, pages 11864-11874, 2021. 6, 7", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 1, 6", + "[11] Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. Speeding up automatic hyperparameter optimization of deep neural networks by extrapolation of learning curves. In *IJCAI*, 2015. 2", + "[12] Xuanyi Dong and Yi Yang. Searching for a robust neural architecture in fourgpu hours. In CVPR, pages 1761-1770, 2019. 1", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 8", + "[14] Pengfei Fang, Jieming Zhou, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Bilinear attention networks for person retrieval. In ICCV, pages 8030-8039, 2019. 2, 6", + "[15] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 4" + ], + "bbox": [ + 78, + 148, + 468, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In ICLR, 2019. 6", + "[17] Yixiao Ge, Dapeng Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. In NeurIPS, 2020. 1, 3, 6", + "[18] Yiluan Guo and Ngai-Man Cheung. Efficient and deep person re-identification using multi-level similarity. In CVPR, pages 2335–2344, 2018. 2", + "[19] Xumeng Han, Xuehui Yu, Nan Jiang, Guorong Li, Jian Zhao, Qixiang Ye, and Zhenjun Han. Group sampling for unsupervised person re-identification. arXiv preprint arXiv:2107.03024, 2021. 6, 7", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2", + "[21] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, pages 15013–15022, 2021. 8", + "[22] Ruibing Hou, Bingpeng Ma, Hong Chang, Xinqian Gu, Shiguang Shan, and Xilin Chen. Interaction-and-aggregation network for person re-identification. In CVPR, pages 9317–9326, 2019. 6", + "[23] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 2", + "[24] Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Meta pairwise relationship distillation for unsupervised person re-identification. In ICCV, pages 3661–3670, 2021. 6", + "[25] Xin Jin, Cuiling Lan, Wenjun Zeng, Zhibo Chen, and Li Zhang. Style normalization and restitution for generalizable person re-identification. In CVPR, pages 3143–3152, 2020. 6", + "[26] Xin Jin, Cuiling Lan, Wenjun Zeng, Guoqiang Wei, and Zhibo Chen. Semantics-aligned representation learning for person re-identification. In AAAI, pages 11173–11180, 2020. 6", + "[27] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6", + "[28] Hanjun Li, Gaojie Wu, and Wei-Shi Zheng. Combined depth space based architecture search for person re-identification. In CVPR, pages 6729–6738, 2021. 1, 2, 3, 5, 6", + "[29] Jianing Li and Shiliang Zhang. Joint visual and temporal consistency for unsupervised domain adaptive person re-identification. In ECCV, pages 483–499, 2020. 6", + "[30] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deepreid: Deep filter pairing neural network for person re-identification. In CVPR, pages 152–159, 2014. 2", + "[31] Wei Li, Xiatian Zhu, and Shaogang Gong. Harmonious attention network for person re-identification. In CVPR, pages 2285–2294, 2018. 2", + "[32] Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded person re-identification with part-aware transformer. In CVPR, pages 2898–2907, 2021. 8" + ], + "bbox": [ + 502, + 125, + 890, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19251", + "bbox": [ + 465, + 887, + 502, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Shengcai Liao and Ling Shao. Interpretable and generalizable person re-identification with query-adaptive convolution and temporal lifting. In ECCV, pages 456–474, 2020. 6", + "[34] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shiens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In ECCV, pages 19–34, 2018. 2", + "[35] Hanxiao Liu, Karen Simonyan, Oriol Vinyals, Chrisantha Fernando, and Koray Kavukcuoglu. Hierarchical representations for efficient architecture search. In ICLR, 2018. 2", + "[36] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In ICLR, 2018. 1, 2", + "[37] Hongye Liu, Yonghong Tian, Yaowei Yang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In CVPR, pages 2167-2175, 2016. 5", + "[38] Xinchen Liu, Wu Liu, Huadong Ma, and Huiyuan Fu. Large-scale vehicle re-identification in urban surveillance videos. In ICME, pages 1–6, 2016. 2, 5", + "[39] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In ECCV, pages 869–884, 2016. 1, 2, 5", + "[40] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. Provid: Progressive and multimodal vehicle reidentification for large-scale urban surveillance. T-MM, 20(3):645-658, 2017. 1", + "[41] Chuanchen Luo, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Spectral feature transformation for person re-identification. In ICCV, pages 4976–4985, 2019. 6", + "[42] Hao Luo, Wei Jiang, Youzhi Gu, Fuxu Liu, Xingyu Liao, Shenqi Lai, and Jianyang Gu. A strong baseline and batch normalization neck for deep person re-identification. T-MM, 22(10):2597-2609, 2019. 1, 5", + "[43] Renqian Luo, Fei Tian, Tao Qin, Enhong Chen, and Tie-Yan Liu. Neural architecture optimization. In NeurIPS, 2018. 2", + "[44] Xuelin Qian, Yanwei Fu, Yu-Gang Jiang, Tao Xiang, and Xiangyang Xue. Multi-scale deep learning architectures for person re-identification. In ICCV, pages 5399–5408, 2017. 1, 3", + "[45] Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, and Yi Yang. Auto-reid: Searching for a part-aware convnet for person re-identification. In ICCV, pages 3750–3759, 2019. 1, 2, 6, 7", + "[46] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In AAAI, pages 4780-4789, 2019. 2", + "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 2", + "[48] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 1", + "[49] Yantao Shen, Tong Xiao, Hongsheng Li, Shuai Yi, and Xiaogang Wang. Learning deep neural networks for vehicle re-id with visual-spatio-temporal path proposals. In ICCV, pages 1900-1909, 2017. 1" + ], + "bbox": [ + 78, + 125, + 468, + 864 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Liangchen Song, Cheng Wang, Lefei Zhang, Bo Du, Qian Zhang, Chang Huang, and Xinggang Wang. Unsupervised domain adaptive re-identification: Theory and practice. PR, 102:107173, 2020. 1", + "[51] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV, pages 480-496, 2018. 3, 6", + "[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 1, 2", + "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 8", + "[54] Dongkai Wang and Shiliang Zhang. Unsupervised person re-identification via multi-label classification. In CVPR, pages 10981-10990, 2020. 6", + "[55] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACMMM, pages 274–282, 2018. 6", + "[56] Yicheng Wang, Zhenzhong Chen, Feng Wu, and Gang Wang. Person re-identification with cascaded pairwise convolutions. In CVPR, pages 1470–1478, 2018. 2", + "[57] Zhikang Wang, Lihuo He, Xiaoguang Tu, Jian Zhao, Xinbo Gao, Shengmei Shen, and Jiashi Feng. Robust video-based person re-identification by hierarchical mining. T-CSVT, 2021. 1", + "[58] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In ICCV, pages 379–387, 2017. 1", + "[59] Zhongdao Wang, Jingwei Zhang, Liang Zheng, Yixuan Liu, Yifan Sun, Yali Li, and Shengjin Wang. Cycles: Self-supervised cycle association for learning re-identifiable descriptions. In ECCV, pages 72–88, 2020. 6", + "[60] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In CVPR, pages 79–88, 2018. 2, 5", + "[61] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 4", + "[62] Yuhui Xu, Lingxi Xie, Xiaopeng Zhang, Xin Chen, Guo-Jun Qi, Qi Tian, and Hongkai Xiong. Pc-darts: Partial channel connections for memory-efficient architecture search. In ICLR, 2019. 2", + "[63] Cheng Yan, Guansong Pang, Xiao Bai, Changhong Liu, Ning Xin, Lin Gu, and Jun Zhou. Beyond triplet loss: person re-identification with fine-grained difference-aware pairwise loss. T-MM, 2021. 1", + "[64] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C.H. Hoi. Deep learning for person re-identification: A survey and outlook. T-PAMI, 2021. 1" + ], + "bbox": [ + 502, + 127, + 890, + 865 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19252", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Deep metric learning for person re-identification. In ICPR, pages 34–39, 2014. 1", + "[66] Anguo Zhang, Yueming Gao, Yuzhen Niu, Wenxi Liu, and Yongcheng Zhou. Coarse-to-fine person re-identification with auxiliary-domain classification and second-order information bottleneck. In CVPR, pages 598–607, 2021. 6", + "[67] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Xin Jin, and Zhibo Chen. Relation-aware global attention for person re-identification. In CVPR, pages 3186–3195, 2020. 2, 6, 7", + "[68] Aihua Zheng, Xianmin Lin, Jiacheng Dong, Wenzhong Wang, Jin Tang, and Bin Luo. Multi-scale attention vehicle re-identification. Neural Computing and Applications, 32(23):17489-17503, 2020. 1, 3", + "[69] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, pages 1116–1124, 2015. 2, 5", + "[70] Liang Zheng, Yi Yang, and Alexander G Hauptmann. Person re-identification: Past, present and future. arXiv preprint arXiv:1610.02984, 2016. 1", + "[71] Liang Zheng, Yi Yang, and Qi Tian. Sift meets cnn: A decade survey of instance retrieval. T-PAMI, 40(5):1224-1244, 2017. 1", + "[72] Meng Zheng, Srikrishna Karanam, Ziyan Wu, and Richard J Radke. Re-identification with consistent attentive siamese networks. In CVPR, pages 5735–5744, 2019. 3", + "[73] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In CVPR, pages 2138–2147, 2019. 6", + "[74] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, pages 13001-13008, 2020. 6", + "[75] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re-identification. In ICCV, pages 3702–3712, 2019. 1, 2, 3, 4, 5, 6", + "[76] Sanping Zhou, Fei Wang, Zeyi Huang, and Jinjun Wang. Discriminative feature learning with consistent attention regularization for person re-identification. In ICCV, pages 8040–8049, 2019. 3", + "[77] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697-8710, 2018. 2", + "[78] Yang Zou, Xiaodong Yang, Zhiding Yu, BVK Vijaya Kumar, and Jan Kautz. Joint disentangling and adaptation for cross-domain person re-identification. In ECCV, pages 87–104, 2020. 1" + ], + "bbox": [ + 78, + 125, + 468, + 747 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19253", + "bbox": [ + 467, + 887, + 504, + 897 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_model.json b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0e7aeaf50e23a0d955cf4085839f67f7d897e91f --- /dev/null +++ b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_model.json @@ -0,0 +1,2752 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.085 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.792, + 0.088 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.162, + 0.88, + 0.182 + ], + "angle": 0, + "content": "MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.208, + 0.314, + 0.225 + ], + "angle": 0, + "content": "Jianyang Gu" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.208, + 0.426, + 0.225 + ], + "angle": 0, + "content": "Kai Wang" + }, + { + "type": "text", + "bbox": [ + 0.448, + 0.208, + 0.527, + 0.225 + ], + "angle": 0, + "content": "Hao Luo3" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.208, + 0.647, + 0.225 + ], + "angle": 0, + "content": "Chen Chen" + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.208, + 0.762, + 0.225 + ], + "angle": 0, + "content": "Wei Jiang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.225, + 0.348, + 0.241 + ], + "angle": 0, + "content": "Yuqiang Fang" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.225, + 0.523, + 0.242 + ], + "angle": 0, + "content": "Shanghang Zhang" + }, + { + "type": "text", + "bbox": [ + 0.546, + 0.225, + 0.633, + 0.241 + ], + "angle": 0, + "content": "Yang You" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.225, + 0.746, + 0.24 + ], + "angle": 0, + "content": "Jian Zhao\\(^{7*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.241, + 0.348, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.241, + 0.639, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.241, + 0.788, + 0.257 + ], + "angle": 0, + "content": "3Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.257, + 0.368, + 0.273 + ], + "angle": 0, + "content": "\\(^{4}\\)OPPO Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.257, + 0.635, + 0.274 + ], + "angle": 0, + "content": "5Space Engineering University" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.257, + 0.806, + 0.274 + ], + "angle": 0, + "content": "6Peking University" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.274, + 0.646, + 0.29 + ], + "angle": 0, + "content": "\\(^{7}\\)Institute of North Electronic Equipment" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.292, + 0.739, + 0.305 + ], + "angle": 0, + "content": "{gu_jianyang, jiangwei_zju}@zju.edu.cn zhaiqian90@u.nus.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.337, + 0.312, + 0.351 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.671 + ], + "angle": 0, + "content": "Neural Architecture Search (NAS) has been increasingly appealing to the society of object Re-Identification (ReID), for that task-specific architectures significantly improve the retrieval performance. Previous works explore new optimizing targets and search spaces for NAS ReID, yet they neglect the difference of training schemes between image classification and ReID. In this work, we propose a novel Twins Contrastive Mechanism (TCM) to provide more appropriate supervision for ReID architecture search. TCM reduces the category overlaps between the training and validation data, and assists NAS in simulating real-world ReID training schemes. We then design a Multi-Scale Interaction (MSI) search space to search for rational interaction operations between multi-scale features. In addition, we introduce a Spatial Alignment Module (SAM) to further enhance the attention consistency confronted with images from different sources. Under the proposed NAS scheme, a specific architecture is automatically searched, named as MSINet. Extensive experiments demonstrate that our method surpasses state-of-the-art ReID methods on both indomain and cross-domain scenarios. Source code available in https://github.com/vimar-gu/MSINet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.208, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.831 + ], + "angle": 0, + "content": "Object re-identification (Re-ID) aims at retrieving specific object instances across different views [39, 40, 57, 65, 70], which attracts much attention in computer vision community due to its wide-range applications. Previous works have achieved great progresses on both supervised [42, 49, 58] and unsupervised ReID tasks [17, 50, 78], most of which adopts backbone models originally designed for general image classification tasks [20, 52]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.832, + 0.469, + 0.846 + ], + "angle": 0, + "content": "Recent literature [64, 75] has shown that applying differ" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.337, + 0.609, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.353, + 0.655, + 0.366 + ], + "angle": 0, + "content": "hood" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.338, + 0.77, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.339, + 0.882, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.473, + 0.891, + 0.489 + ], + "angle": 0, + "content": "Distance between top-10 neg. samples and each query image" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.414, + 0.607, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.416, + 0.664, + 0.439 + ], + "angle": 0, + "content": "luggage carrier" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.416, + 0.771, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.502, + 0.893, + 0.554 + ], + "angle": 0, + "content": "Figure 1. The left panel shows the example activation maps of ResNet50 (1st row) and MSINet (2nd row). The right panel shows the average distances between the most similar 10 negative samples and each query image at the inference. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.868 + ], + "angle": 0, + "content": "ment architectures on ReID leads to large performance variations. Some works employ Neural Architecture Search (NAS) for ReID [28, 45]. The proposed optimizing targets and search spaces stably improve the model performance, yet the main search scheme still follows traditional NAS methods designed for general classification tasks [12, 36]. As an open-set task, ReID contains different categories in the training and validation sets [64, 71], while the two sets share exactly the same categories in standard classification tasks [10], which is also followed by traditional NAS methods. The incompatibility between search schemes and real-world training schemes makes the searched architecture sub-optimal for ReID. Moreover, ReID is required to distinguish more subtle distinctions among fine-grained instances compared with image-level classification [48, 63]. Some previous works [4, 44, 68, 75] have manifested that local perspectives and multi-scale features are discriminative for ReID. However, current utilizations of these features are mostly empirically designed, which can be more flexible according to the characteristics of different network layers." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.854, + 0.242, + 0.866 + ], + "angle": 0, + "content": "*Co-corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "19243" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.473, + 0.388 + ], + "angle": 0, + "content": "In this work, we propose a novel NAS scheme aiming at addressing the aforementioned challenges. In order to simulate the real-world ReID training schemes, a Twins Contrastive Mechanism (TCM) is proposed to unbind the categories of the training and validation sets. An adjustable overlap ratio of categories builds up the compatibility between NAS and ReID, which provides more appropriate supervision for ReID architecture search. Moreover, to search for more rational utilizations of multi-scale features, we design a Multi-Scale Interaction (MSI) search space. The MSI space focuses on interaction operations between multi-scale features along the shallow and deep layers of the network, which guides the features to promote each other. Additionally, to further improve the generalization capability, we propose a Spatial Alignment Module (SAM) to enhance the attention consistency of the model confronted with images from different sources. With the above NAS scheme, we obtain a light-weight yet effective model architecture, denoted as Multi-Scale Interaction Net (MSINet)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.389, + 0.473, + 0.542 + ], + "angle": 0, + "content": "We visualize the example activation maps of our proposed MSINet and ResNet50 [20] trained on VeRi-776 [38, 39] in Fig. 1. Compared to ResNet50, MSINet focuses on more unique distinctions with specific semantic information to recognize instances. Besides, MSINet largely increases the distance margin between query image and corresponding negative samples, reflecting extraordinary discriminative capability. Extensive experiments demonstrate that MSINet surpasses state-of-the-art (SOTA) ReID methods on both in-domain and cross-domain scenarios. Our source codes are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.542, + 0.4, + 0.554 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.564, + 0.47, + 0.62 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to build the NAS search scheme according to the real-world ReID training schemes, which provides more appropriate supervision for the ReID architecture search." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.629, + 0.47, + 0.683 + ], + "angle": 0, + "content": "- We propose a novel search space based on the Multi-Scale Interaction (MSI) operations and a Spatial Alignment Module (SAM) to improve the model performance on in-domain and cross-domain scenarios." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.693, + 0.47, + 0.762 + ], + "angle": 0, + "content": "- We construct a light-weight yet effective architecture for ReID tasks, denoted as MSINet. With only 2.3M parameters, MSINet surpasses ResNet50 [20] by \\(9\\%\\) mAP on MSMT17 [60] and \\(16\\%\\) mAP on MSMT17 \\(\\rightarrow\\) Market-1501 [69]." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.564, + 0.47, + 0.762 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.774, + 0.227, + 0.789 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.797, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Neural Architecture Search. NAS has been increasingly appealing to the computer vision society, due to its automatic architecture designing characteristics. NAS methods can be roughly separated into four categories: reinforcement learning [1, 77], evolutionary algorithms [35, 46], gra" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.305 + ], + "angle": 0, + "content": "dient descent [36, 43] and performance prediction [11, 34]. Liu et al. establish a differentiable architecture search (DARTS) method [36], which improves the practicability of NAS by a large extent. Some later works further improve the structure through sampling strategy [62], network pruning [2, 8], progressive learning [5], collaborative competition [7], etc. Most of NAS works focus on general image classification tasks, where the training and validation sets share the exact same categories. Following the setting, however, leads to incompatibility with the real-world training schemes of object ReID. In this work, we unbind the category bond between the two sets and propose a novel search scheme suitable for ReID." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.892, + 0.389 + ], + "angle": 0, + "content": "ReID Network Design. Current ReID works mostly adopt backbones designed for image classification [20, 23, 47, 52]. Some works [14, 31, 67] design attention modules based on the common backbones to unearth their potential on distinguishing local distinctions. However, these methods usually lead to large calculation consumption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.893, + 0.651 + ], + "angle": 0, + "content": "There are also several works focusing on designing ReID-specific architectures. Li et al. present a Filter Pairing Neural Network to dynamically match patches in the feature maps [30]. Wang et al. separate and regroup the features of two samples with a WConv layer [56]. Guo et al. extract multi-scale features to directly evaluate the similarity between samples [18]. However, the siamese structure is inconvenient when conducting retrieval on large galleries. Zhou et al. aggregate multi-scale information to achieve high accuracy with small computing consumption [75]. Quan et al. introduce a part-aware module into the DARTS search space [36, 45]. Li et al. propose a new search space in regard to receptive field scales [28]. These methods have excellent performance on limited parameter scales, but fail to surpass those networks with complex structures. Different from previous works, we design a light-weight searching structure focusing on rational interaction operations between multi-scale features. The searched MSINet surpasses SOTA methods on both in-domain and cross-domain tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.662, + 0.599, + 0.677 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.811 + ], + "angle": 0, + "content": "Our goal is to construct an effective NAS scheme to search for a light-weight backbone architecture suitable for ReID tasks. Based on the training schemes of ReID, we propose a novel Twins Contrastive Mechanism to provide more appropriate supervision for the search process. Aiming at rational interaction between multi-scale features, we design a Multi-Scale Interaction search space. We further introduce a Spatial Alignment Module to improve the generalization capability with limited parameter growth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.77, + 0.832 + ], + "angle": 0, + "content": "3.1. Twins Contrastive Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.868 + ], + "angle": 0, + "content": "NAS aims at automatically searching for the optimal network architecture for certain data. Inspired by [36], a basic" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19244" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.124, + 0.87, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.893, + 0.339 + ], + "angle": 0, + "content": "Figure 2. The model structure of the proposed MSINet. The input can be either person or vehicle. Inside a cell, the input is separated to two branches, with different receptive field scales. The interaction module exchanges information between two branches. Architecture search automatically selects the most appropriate interaction for each cell." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.473, + 0.459 + ], + "angle": 0, + "content": "differentiable architecture search scheme is established. We define the ordinary model parameters as \\(\\omega\\), and architecture parameters as \\(\\alpha\\). For network layer \\(i\\) with a search space of \\(\\mathcal{O}\\), \\(\\alpha_{i}\\) controls the weight of each operation \\(o\\) in the space. The features are parallelly passed through all the operations, and the final output is formulated by the softmax-weighted sum of operation outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.47, + 0.47, + 0.51 + ], + "angle": 0, + "content": "\\[\nf \\left(\\mathbf {x} _ {i}\\right) = \\sum_ {o \\in \\mathcal {O}} \\frac {\\exp \\left\\{\\alpha_ {i} ^ {o} \\right\\}}{\\sum_ {o ^ {\\prime} \\in \\mathcal {O}} \\exp \\left\\{\\alpha_ {i} ^ {o ^ {\\prime}} \\right\\}} \\cdot o \\left(\\mathbf {x} _ {i}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.617 + ], + "angle": 0, + "content": "The search process is conducted in an alternative manner. Training data is utilized to update the model parameters, and validation data is then employed to update the architecture parameters. For most NAS methods designed for image classification tasks, the training and validation data share exactly the same categories and a linear classification layer for loss calculation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.471, + 0.825 + ], + "angle": 0, + "content": "Different from standard image classification, as an open-set retrieval task, ReID has different categories in the training and validation sets. The incompatibility between search schemes and real-world training schemes might lead to suboptimal searching results. Accordingly, we propose a novel Twins Contrastive Mechanism (TCM) for NAS ReID training. Specifically, we employ two independent auxiliary memories \\( \\mathcal{C}_{tr} \\) and \\( \\mathcal{C}_{val} \\) to store the embedded features of the training and validation data, respectively. The memories are initialized with the centroid features, which are calculated by averaging the features of each category. At each iteration, the training loss is first calculated with \\( \\mathcal{C}_{tr} \\) for model parameter updating. Given an embedded feature \\( f \\) with category label \\( j \\), the contrastive classification loss is calculated with:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.835, + 0.47, + 0.872 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t r} ^ {c l s} = - \\log \\frac {\\exp (\\mathbf {f} \\cdot \\mathbf {c} _ {t r} ^ {j} / \\tau)}{\\sum_ {n = 0} ^ {N _ {t r} ^ {c}} \\exp (\\mathbf {f} \\cdot \\mathbf {c} _ {t r} ^ {n} / \\tau)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.446 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_{tr}^n\\) represents the memorized feature of category \\(n\\), \\(N_{tr}^c\\) stands for the total number of categories in the training set, and \\(\\tau\\) is the temperature parameter, which is set as 0.05 empirically [17]. After updating the model parameters, the embedded feature \\(\\mathbf{f}\\) with category label \\(j\\) is integrated into the corresponding memorized feature \\(\\mathbf{c}_{tr}^j\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.453, + 0.892, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {t r} ^ {j} \\leftarrow \\beta \\mathbf {c} _ {t r} ^ {j} + (1 - \\beta) \\mathbf {f}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.546 + ], + "angle": 0, + "content": "where \\(\\beta\\) is set as 0.2 empirically [17]. Then the updated model is evaluated on the validation data to generate to validation loss with \\(C_{val}\\) replacing \\(C_{tr}\\) in Eq. 2. The architecture parameter is then updated with the validation loss to finish an iteration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.893, + 0.715 + ], + "angle": 0, + "content": "As the loss calculation does not rely on the linear classification layer, the categories of the training and validation sets are unbound. We are able to dynamically adjust the category overlap ratio in these two sets. The advantages of a proper overlap ratio are summarized as two folds. Firstly, TCM better simulates the real-world training of ReID and helps the model focus on truly discriminative distinctions. The differences between the training and validation data improves the generalization capability of the model. Secondly, a relatively small proportion of overlapped categories stabilizes the architecture parameter update through a consistent optimizing target with the model parameter update." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.766, + 0.737 + ], + "angle": 0, + "content": "3.2. Multi-Scale Interaction Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Although the local perspective and multi-scale features have already been investigated in previous ReID works [4, 28, 44, 51, 68, 72, 75, 76], the utilization of these information is mainly empirically designed aggregation, which is monotonous and restrained. We argue that on the one hand, the rational utilization of multi-scale features should be dynamically adjusted along the shallow and deep layers of the network. On the other hand, introducing interaction other than aggregation creates direct information exchange, and" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "19245" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.124, + 0.87, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.291, + 0.893, + 0.331 + ], + "angle": 0, + "content": "Figure 3. The aligning pipeline of the proposed Spatial Alignment Module. The correlation activation vectors are calculated between the anchor feature and all features in a mini-batch. The positive vectors are aligned with the learnable self-activation, and the negative vectors are aligned with each other. The structure of PAM is shown on the left." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.352, + 0.47, + 0.39 + ], + "angle": 0, + "content": "Table 1. The detailed interaction operation in the proposed MSINet architecture. N: None; E: Exchange; G: Channel Gate; A: Cross Attention." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.401, + 0.468, + 0.449 + ], + "angle": 0, + "content": "
Cell #1Cell #2Cell #3Cell #4Cell #5Cell #6
123456789101112
GGEGAGGNGAEA
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.469, + 0.516 + ], + "angle": 0, + "content": "makes fuller use of multi-scale features. Therefore, we propose a novel Multi-Scale Interaction (MSI) search space to establish a light-weight architecture suitable for ReID." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.469, + 0.71 + ], + "angle": 0, + "content": "As shown in Fig. 2, the network is mainly grouped with MSI cells and down-sample blocks, which is generally consistent with OSNet [75]. In each cell, the input features are passed through two branches with different receptive field scales. To reduce the calculation burden of the network, for the layers inside each branch, we adopt the stack of \\(1 \\times 1\\) convolution and multiple depth-wise \\(3 \\times 3\\) convolution to implement specific scales. A scale ratio \\(\\rho\\) of 3:1 is selected for the two branches. These two branches do not share model parameters, except for the Interaction Modules (IM). IM introduces information exchange for the two branches. There are 4 operation options for the IM. With the two-branch input features defined as \\((\\mathbf{x}_1, \\mathbf{x}_2)\\), the operations can be formulated as:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.712, + 0.469, + 0.74 + ], + "angle": 0, + "content": "None. None operation involves no parameters, and outputs exactly the input features \\((\\mathbf{x}_1,\\mathbf{x}_2)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Exchange. Exchange acts as the strongest interaction among all options. It directly exchanges the features for the two branches and outputs \\((\\mathbf{x}_2, \\mathbf{x}_1)\\). Exchange contains no extra parameters, as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.469, + 0.839 + ], + "angle": 0, + "content": "Channel Gate. Channel gate introduces a Multi-Layer Perceptron (MLP) to generate a channel-wise attention gate [61, 75] as:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.853, + 0.469, + 0.868 + ], + "angle": 0, + "content": "\\[\nG (\\mathbf {x}) = \\sigma (M L P (\\mathbf {x})), \\tag {4}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.351, + 0.906, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.455, + 0.879, + 0.468 + ], + "angle": 0, + "content": "Figure 4. Output feature maps of MSI cells at different layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.547 + ], + "angle": 0, + "content": "and returns \\((G(\\mathbf{x}_1) \\cdot \\mathbf{x}_1, G(\\mathbf{x}_2) \\cdot \\mathbf{x}_2)\\). The MLP is composed of 2 fully connected layers and its parameters are shared for both branches. Thereby it achieves interaction by jointly screening discriminative feature channels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Cross Attention. Traditional channel attention module calculates the channel correlation inside a single feature map [15]. The original feature map \\(\\mathbf{x} \\in \\mathcal{R}^{C \\times H \\times W}\\) is firstly reshaped into the query feature \\(\\tilde{\\mathbf{x}} \\in \\mathcal{R}^{C \\times N}\\), where \\(N = H \\times W\\). Then the correlation activation is calculated by performing a matrix multiplication between the query feature \\(\\tilde{\\mathbf{x}}\\) and the key feature \\(\\tilde{\\mathbf{x}}^{\\top}\\). We propose to exchange the keys of the two branches to explicitly calculate the correlation between each other. The correlation activation is then transformed to a mask, and is added up to the original features with a learnable proportion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.838 + ], + "angle": 0, + "content": "After interaction, the multi-scale branches are fused through a sum operation. It is worth noting that the extra parameters brought by multiple interaction options are limited, which enables searching for each cell along the whole network independently. At the beginning of the network, we employ the same stem module as that in OSNet [75], containing a \\( 7 \\times 7 \\) convolutional layer and a \\( 3 \\times 3 \\) max pooling with a stride of 2. After the searching process, the interaction operation \\( o \\) with the largest weight \\( \\alpha_{i}^{o} \\) at each layer is reserved to form the searched architecture." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.892, + 0.867 + ], + "angle": 0, + "content": "After searching the architecture, the model is validated on various Re-ID tasks. The training is constrained by the" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "19246" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.124, + 0.895, + 0.178 + ], + "angle": 0, + "content": "Table 2. Supervised performance on object ReID datasets. The results in the top part are trained from scratch, and those in the bottom part are pre-trained on ImageNet in advance. As the compared methods are originally proposed for person ReID, we reproduce the results in vehicle datasets. * indicates that the results of person ReID are reproduced by us. The evaluation results of architecture searched on VR can be found in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.186, + 0.877, + 0.348 + ], + "angle": 0, + "content": "
MethodParamsInference TimeMMSVRVIDMS→MVR→VID
R-1↑mAP↑R-1↑mAP↑R-1↑mAP↑R-1↑R-5↑R-1↑mAP↑R-1↑R-5↑
ResNet50* [42]~24M1×85.768.348.025.792.869.970.676.6--
OSNet [75]2.2M0.79×93.681.071.043.395.472.876.088.7----
CDNet [28]1.8M0.67×93.783.773.748.594.373.074.588.8----
MSINet2.3M0.71×94.687.076.052.595.975.076.589.8----
ResNet50* [42]~24M1×94.585.975.550.494.573.676.589.958.831.842.861.9
OSNet [75]2.2M0.79×94.884.978.752.995.576.476.088.666.637.546.563.1
CDNet [28]1.8M0.67×95.186.078.954.7--------
MSINet2.3M0.71×95.389.681.059.696.878.877.991.774.946.248.065.6
MSINet-SAM2.4M0.71×95.589.980.759.596.779.078.091.976.348.449.066.8
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.381, + 0.443, + 0.396 + ], + "angle": 0, + "content": "classification id loss and the triplet loss, formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.405, + 0.47, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i d} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - \\log \\left(\\frac {\\exp \\mathbf {W} _ {i} ^ {\\top} \\mathbf {f} _ {i}}{\\sum_ {j} \\exp \\mathbf {W} _ {j} ^ {\\top} \\mathbf {f} _ {i}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.452, + 0.47, + 0.482 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_i\\) is a feature vector, the corresponding classifier weight of which is \\(\\mathbf{W}_i\\), and" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.491, + 0.47, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t r i} = \\left[ \\mathcal {D} \\left(\\mathbf {f} _ {a}, \\mathbf {f} _ {p}\\right) - \\mathcal {D} \\left(\\mathbf {f} _ {a}, \\mathbf {f} _ {n}\\right) + m \\right] _ {+}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.517, + 0.47, + 0.575 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_a, \\mathbf{f}_p, \\mathbf{f}_n\\) are the embedded features for the anchor, the hardest positive and negative samples in a mini-batch, \\(\\mathcal{D}(\\cdot, \\cdot)\\) is the Euclidean distance, \\(m\\) is the margin parameter, and \\([\\cdot]_{+}\\) is the \\(\\max(\\cdot, 0)\\) function." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.581, + 0.317, + 0.596 + ], + "angle": 0, + "content": "3.3. Spatial Alignment Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.469, + 0.699 + ], + "angle": 0, + "content": "The retrieval precision of object ReID tasks are largely affected by the variation of appearances such as poses, illumination and occlusion when the camera conditions change. In order that the model correctly and consistently focuses on the discriminative spatial positions, we design a Spatial Alignment Module (SAM) to explicitly align the spatial attention between images, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.469, + 0.785 + ], + "angle": 0, + "content": "Specifically, we first calculate the position-wise correlation activation map \\(\\mathbf{A}\\) between the feature maps in a minibatch. The activation between sample \\(i\\) and \\(j\\) can be formulated as: \\(\\mathbf{A}(i,j) = \\tilde{\\mathbf{x}}_j^\\top \\times \\tilde{\\mathbf{x}}_i\\), where \\(\\tilde{\\mathbf{x}} \\in \\mathcal{R}^{C \\times N}\\) is reshaped from the original feature \\(\\mathbf{x} \\in \\mathcal{R}^{C \\times H \\times W}\\). Then we take the maximum activation for each position of sample \\(i\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.795, + 0.47, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} (i, j) = \\max _ {d i m = 1} \\mathbf {A} (i, j). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.868 + ], + "angle": 0, + "content": "The above process is denoted as \"Mutual Conv\" in Fig. 3. We evaluate the consistency between activation vectors with cosine similarity. For negative samples specifically, there" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.45 + ], + "angle": 0, + "content": "can be many different hints for recognition, some of which might be inappropriate, such as the backgrounds. By aligning all the correlations for sample \\( i \\), we hope that the network can correct some attention bias and consistently focus on discriminative positions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.893, + 0.535 + ], + "angle": 0, + "content": "However, through aligning positive sample pairs, the ID-related features are expected to be emphasized, which cannot be achieved by aligning negative pairs. Therefore, we introduce an extra position activation module (PAM) to generate supervision for the alignment between positive pairs. The spatial alignment loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.548, + 0.843, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s a} (i) = \\frac {1}{N _ {+}} \\sum_ {p \\in \\mathcal {I} _ {+}} \\left(1 - S (\\hat {\\mathbf {a}} (i), \\mathbf {a} (i, p))\\right) + \\\\ \\frac {1}{N _ {-}} \\sum_ {n _ {1}, n _ {2} \\in \\mathcal {I} _ {-}} (1 - S (\\mathbf {a} (i, n _ {1}), \\mathbf {a} (i, n _ {2}))), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.868, + 0.579, + 0.892, + 0.591 + ], + "angle": 0, + "content": "(8)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.893, + 0.69 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_{+}\\) contains positive indices for sample \\(i\\), the total number of which is \\(N_{+}\\), and vice versa. \\(\\hat{\\mathbf{a}}(i)\\) stands for the generated activation vector for positive sample alignment, and \\(S(\\cdot, \\cdot)\\) is the cosine similarity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.634, + 0.724 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.733, + 0.787, + 0.747 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Our proposed method is evaluated on two person ReID datasets Market-1501 [69], MSMT17 [60], and two vehicle ReID datasets VeRi-776 [38, 39] and VehicleID [37]. For simplicity, the four datasets are denoted as M, MS, VR and VID in the following sections, respectively. Evaluation metrics include Cumulative Matching Characteristic (CMC) and mean average precision (mAP), which are commonly utilized on ReID tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "19247" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.124, + 0.47, + 0.15 + ], + "angle": 0, + "content": "Table 3. Supervised performance comparison between MSINet and SOTA methods on M and MS datasets." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.16, + 0.426, + 0.366 + ], + "angle": 0, + "content": "
MethodMMS
R-1↑mAP↑R-1↑mAP↑
PCB [51]93.881.668.240.4
MGN [55]95.786.976.952.1
OSNet [75]93.681.071.043.3
IANet [22]94.483.175.546.8
DGNet [73]94.886.077.252.3
Auto-ReID [45]94.585.1--
SAN [26]96.188.079.255.7
CDNet [28]95.186.078.954.7
BAT-Net [14]95.187.479.556.8
SFT [41]94.187.579.058.3
CTF [66]94.887.7--
RGA-SC [67]96.188.480.357.5
MSINet95.389.681.059.6
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.388, + 0.267, + 0.401 + ], + "angle": 0, + "content": "4.2. Architecture Search" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.409, + 0.468, + 0.588 + ], + "angle": 0, + "content": "We conduct the searching process on MSMT17. SGD is adopted for model parameter update with an initial learning rate of 0.025. The model is trained for 350 epochs in total. We adopt a warm-up strategy for the first 10 epochs. Then the learning rate is decayed by 0.1 at 150, 225 and 300 epochs, respectively. Adam [27] is adopted for the architecture parameter update with an initial learning rate of 0.002. The learning rate is decayed at the same pace. The images are reshaped to \\(256 \\times 128\\) for person and \\(256 \\times 256\\) for vehicles. Data augmentation includes random flip, random crop and random erasing [74]. The searched architecture is presented in Tab. 1. The \"MSINet\" in the following experiment sections refers to this architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.589, + 0.468, + 0.713 + ], + "angle": 0, + "content": "We visualize the feature maps extracted by each MSI cell in Fig. 4. At the shallow layers of the network, the kernels mainly focus on overall contour information. Channel gate helps to filter out inferior information, such as the background. As we approach deeper layers, the extracted features each have specific semantic information, where cross attention is more likely to be selected for the interaction. It indicates that cross attention is more rational for exchanging high-level semantic information." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.721, + 0.383, + 0.735 + ], + "angle": 0, + "content": "4.3. Comparison with Other Backbones" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.743, + 0.468, + 0.782 + ], + "angle": 0, + "content": "We first compare our proposed MSINet with ResNet50 and recent proposed light-weight backbones in both in-domain and cross-domain ReID scenarios." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.784, + 0.468, + 0.866 + ], + "angle": 0, + "content": "In-Domain Tasks. We adopt a two-group supervised evaluation scheme similar to that in [28, 75]: training from scratch and fine-tuning ImageNet [10] pre-trained models. The training parameters for both schemes are kept the same as that in architecture search, except for an initial learning rate of 0.065. Triplet loss and cross entropy loss are adopted" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.124, + 0.892, + 0.15 + ], + "angle": 0, + "content": "Table 4. Unsupervised performance applying MSINet to SOTA methods for USL on M and UDA on \\( \\mathrm{M} \\rightarrow \\mathrm{{MS}} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.161, + 0.85, + 0.366 + ], + "angle": 0, + "content": "
MethodMM→MS
R-1↑mAP↑R-1↑mAP↑
MMCL [54]80.345.540.815.1
MMT [16]--50.124.0
JVTC+ [29]79.547.548.625.1
CycAs [59]84.864.8--
GCL [3]87.366.851.127.0
MPRD [24]83.051.0--
SpCL [17]88.173.153.726.8
GS [19]92.379.2--
GS+MSINet91.781.5--
HDCRL [6]92.481.7--
HDCRL+MSINet92.982.7--
IDM [9]--61.333.5
IDM+MSINet--66.037.8
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.393, + 0.891, + 0.432 + ], + "angle": 0, + "content": "for the parameter update. The margin \\( m \\) in Eq. 6 is set as 0.3. [28] adopts an FBLNeck. We also employ the same structure. The results are shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.435, + 0.892, + 0.684 + ], + "angle": 0, + "content": "ResNet50 is the most commonly utilized backbone network in ReID tasks, yet holds the worst performance. Moreover, ResNet50 largely depends on ImageNet pre-training, while MSINet without pre-training has already surpassed pre-trained ResNet50 on all metrics. Compared with the other datasets, MS contains more variations on illumination, background and camera pose, and brings a large performance gap between ResNet50 and other methods. It also validates the inadequacy of image classification networks on ReID tasks. OSNet [75] and CDNet [28] are recently proposed architectures designed specifically for ReID tasks. Both architectures focus on fusing multi-scale features to better suit ReID. CDNet employs a traditional NAS scheme to search for the proper receptive field scales for each cell. MSINet fixes the receptive field scale and instead selects optimal interaction operations inside each cell. With only a bit more parameters, MSINet surpasses all the other backbones by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.686, + 0.892, + 0.796 + ], + "angle": 0, + "content": "Cross-Domain Tasks. Cross-domain experiments verify the generalization capability of the model. Following previous domain generalizable ReID works [25, 33], data augmentation is adjusted to random flip, random crop and color jittering. The model is pre-trained and fine-tuned for 250 epochs to avoid over-fitting. The other settings are kept the same as supervision scenes. With no present pre-trained models for CDNet [28], it is excluded from this section." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.798, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Tab. 2 shows that ResNet50 can be easily interfered by different image styles confronted with new image domains. OSNet learns multi-scale features with specific semantic information for ReID, which is domain invariant to some extent. Our proposed search scheme also takes into ac" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "19248" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.126, + 0.245, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.258, + 0.19, + 0.27 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.128, + 0.41, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.258, + 0.351, + 0.27 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.127, + 0.563, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.258, + 0.51, + 0.27 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.127, + 0.714, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.258, + 0.655, + 0.27 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.127, + 0.878, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.258, + 0.812, + 0.27 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.28, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Figure 5. Ablation studies on (a) search scheme; (b) architecture; (c) multi-scale aggregation; (d) scale ratio \\(\\rho\\); (e) spatial alignment weight." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.316, + 0.471, + 0.355 + ], + "angle": 0, + "content": "Table 5. The effectiveness of each components in SAM. Both in-domain performance on VR and cross-domain performance on \\(\\mathrm{VR}\\rightarrow \\mathrm{VID}\\) is evaluated." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.364, + 0.467, + 0.513 + ], + "angle": 0, + "content": "
ModelPos.Neg.AlignVRVR→VID
R-1↑mAP↑R-1↑R-5↑
MSINet✓-96.778.548.065.6
Self96.778.148.466.0
✓Self96.778.448.566.2
✓✓Unified96.678.348.365.8
✓✓Separated96.878.648.766.4
✓✓PAM-Self96.779.049.066.8
OSNet✓✓None95.576.446.563.1
PAM-Self95.976.347.563.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.622 + ], + "angle": 0, + "content": "count the generalization capability of the model. By partly separating the categories for training and validation sets, the searched interaction operations generalize well confronted with new image domains. Except for discrimination, MSINet also surpasses the other backbones on cross-domain tasks by a large margin with faster inference speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.469, + 0.721 + ], + "angle": 0, + "content": "Additionally, we introduce SAM into the model, which aligns the spatial correlations between images. A weighted sum of ReID loss and spatial alignment loss is utilized when training the network with SAM. The weight of spatial alignment loss is set as \\(\\lambda_{sa} = 2.0\\). Without extra inference consumption or damages on the supervised performance, SAM further boosts the generalization capability of MSINet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.733, + 0.442, + 0.749 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Tab. 3 further illustrates the supervision performance comparison of our proposed MSINet with the SOTA methods on M and MS datasets. With much less parameters than most of the compared methods, MSINet achieves a retrieval accuracy comparable to that of more complicated ones. Auto-ReID [45] first designs a NAS scheme for ReID, yet the DARTS-style architecture contains 13M parameters. RGA-SC [67] carefully designs a relation-aware global at" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.359 + ], + "angle": 0, + "content": "tention module. MSINet achieves even higher performance with less training consumption, which validates the superiority of selecting rational interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.893, + 0.513 + ], + "angle": 0, + "content": "We also evaluate the model performance replacing the backbone network from ResNet50 to MSINet for SOTA unsupervised ReID methods in Tab. 4. For purely unsupervised learning (USL) method GS [19] on M dataset, MSINet performs slightly lower on rank-1, yet has a large superiority on mAP. For HDCRL [6], MSINet shows obvious superiority over ResNet50. For unsupervised domain adaptation (UDA) method IDM [9] on \\(\\mathrm{M}\\rightarrow \\mathrm{MS}\\) task, MSINet surpasses ResNet50 by a large margin, which further proves that the TCM brings outstanding generalization capability to the searched architecture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.665, + 0.539 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.672 + ], + "angle": 0, + "content": "Effectiveness of Architecture Search. To verify the effectiveness, we conduct supervised training on MS with different search schemes in Fig. 5 (a). Under the standard classification scheme (\"CE Overlap\"), the searched model performs poor. Replacing the cross entropy loss to contrastive loss (\"TCM Overlap\") only brings a slight improvement. The complete TCM framework unbounds the categories between the training and validation set, and thereby improves the performance by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.893, + 0.868 + ], + "angle": 0, + "content": "In Fig. 5 (b) we compare the performance of different architectures. Firstly, we validate 4 models each with a unique interaction operation from the 4 options in the search space. None and Exchange, with no trainable parameters, achieve poor performance. Channel gate introduces channel-wise attention, whose model performs the best among 4 options. Cross attention exchanges the key features for the two branches. Over-frequent exchange interferes the ordinary feature extraction and degrades the network performance. Through appropriately arranging interaction operations along the architecture, MSINet surpasses all the above 4 models. Random architecture, on the other hand, shows no rational appliance of interaction operations, which validates that the proposed search scheme helps find" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19249" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.124, + 0.47, + 0.15 + ], + "angle": 0, + "content": "Table 6. Supervised performance comparison between MSINet and Transformers on VR and MS datasets." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.16, + 0.47, + 0.253 + ], + "angle": 0, + "content": "
MethodParamsInference TimeMSVR
R-1↑mAP↑R-1↑mAP↑
DeiT-S [21]~22M0.97x76.355.295.576.3
DeiT-B [21]~86M1.79x81.961.495.978.4
ViT-B [21]~86M1.79x81.861.096.578.2
MSINet2.3M0.71x81.059.696.878.8
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.264, + 0.465, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.468 + ], + "angle": 0, + "content": "Figure 6. Example top-15 retrieved sequences comparison on VR. Appearance differences caused by variant camera conditions are well addressed by the proposed MSINet. Visualization of person ReID can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.492, + 0.286, + 0.505 + ], + "angle": 0, + "content": "suitable architectures for ReID." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.469, + 0.741 + ], + "angle": 0, + "content": "Effectiveness of Spatial Alignment Module. We validate the effectiveness of each components of SAM on the VR \\(\\rightarrow\\) VID cross-domain experiment in Tab. 5. Firstly, we introduce the spatial alignment for positive and negative sample pairs, respectively. Each of them brings certain performance improvements. However, a unified alignment for all sample pairs damages ID-related features and degrades the performance instead. Therefore, we separate the alignment of positive and negative samples, which retains some discriminative features and integrates the effect of both aspects. The extra PAM for positive sample alignment further guarantees the focus on ID-related positions and achieves the best performance. We also conduct in-domain experiment on VR to prove that SAM improves the generalization capability without sacrificing the supervision performance. Adding SAM to OSNet receives similar results, which validates the universality of SAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Fusing Operation. After interaction, the multi-scale features are fused by sum operation. We investigate several fusing options on MS training from scratch in Fig. 5 (c). Subtracting (\"Minus\") a branch from the other leads to about the same results as \"Sum\" while multiplication (\"Mul\") performs poorly." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Comparison with Transformer. Transformer, as a new architecture, has recently been continuously making progresses in many computer vision domains [13, 53], includ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.251 + ], + "angle": 0, + "content": "ing ReID [21, 32]. We compare the model performance with some baseline Transformer models in Tab. 6. DeiT-B and ViT-B [21] achieves higher performance on MSMT17, with much larger calculation burden compared with our proposed MSINet. On VeRi-776, MSINet surpasses all the baseline Transformer methods. It proves that rational interaction operations between multi-scale features are capable to assist light-weighted pure-CNN models to obtain comparable performance with complex Transformers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.893, + 0.348 + ], + "angle": 0, + "content": "Parameter Analysis. Firstly, we study the influence of different receptive field scale ratios \\(\\rho\\) inside an MSI cell on MS training from scratch in Fig. 5 (d). Introducing scale differences between branches improves the model performance significantly, and subsequent increases brings more modest impacts. Considering both parameter scales and model performance, the ratio of 3:1 is selected for MSINet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Secondly, the model performance fluctuation influenced by spatial alignment weight is visualized in Fig. 5 (e). The experiment is conducted on the VR \\(\\rightarrow\\) VID cross-domain scenario. Employing the alignment generally makes a positive impact on the generalization capability of the model. The optimal loss weight \\(\\lambda_{sa}\\) locates at 2.0." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.556 + ], + "angle": 0, + "content": "Visualization Results. We visualize the top-15 retrieved sequences and the corresponding distances from an query image on VR in Fig. 6. By comparison, ResNet50 mainly focuses on general appearance features, where the top-rated negative samples share similar car bodies. MSINet, oppositely, concentrates on discriminative distinctions, an empty car hopper in this case, and creates an evident distance gap between positive and negative samples. More details can be seen in the supplementation material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.57, + 0.618, + 0.584 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.745 + ], + "angle": 0, + "content": "In this paper, we design a Twins Contrastive Mechanism for NAS to build the compatibility with ReID. The task-specific search scheme provides the searching process with more appropriate supervision. A Multi-Scale Interaction search space is proposed to establish rational and flexible utilization of multi-scale features. With a Spatial Alignment Module, our proposed MSINet achieves SOTA performance on both supervision and cross-domain scenarios with limited parameter amount. We hope the proposed approach could inspire more works focusing on designing network architectures suitable for ReID tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.76, + 0.68, + 0.776 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.867 + ], + "angle": 0, + "content": "This work is partially supported by National Natural Science Foundation of China (62173302, 62006244), China Scholarship Council (202206320302), Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and AI Singapore Programme (AISG2-PhD-2021-08-008)." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19250" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.173, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.469, + 0.188 + ], + "angle": 0, + "content": "[1] Irwan Bello, Barret Zoph, Vijay Vasudevan, and Quoc V Le. Neural optimizer search with reinforcement learning. In ICML, pages 459-468, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[2] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.469, + 0.281 + ], + "angle": 0, + "content": "[3] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Joint generative and contrastive learning for unsupervised person re-identification. In CVPR, pages 2004–2013, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[4] Tsai-Shien Chen, Chih-Ting Liu, Chih-Wei Wu, and Shao-Yi Chien. Orientation-aware vehicle re-identification with semantics-guided part attention network. In ECCV, pages 330–346, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.336, + 0.469, + 0.386 + ], + "angle": 0, + "content": "[5] Xin Chen, Lingxi Xie, Jun Wu, and Qi Tian. Progressive differentiable architecture search: Bridging the depth gap between search and evaluation. In ICCV, pages 1294-1303, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[6] De Cheng, Jingyu Zhou, Nannan Wang, and Xinbo Gao. Hybrid dynamic contrast and probability distillation for unsupervised person re-id. IEEE Transactions on Image Processing, 31:3334-3346, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.442, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[7] Xiangxiang Chu, Tianbao Zhou, Bo Zhang, and Jixiang Li. Fair darts: Eliminating unfair advantages in differentiable architecture search. In ECCV, pages 465-480, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.483, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[8] Xiyang Dai, Dongdong Chen, Mengchen Liu, Yinpeng Chen, and Lu Yuan. Da-nas: Data adapted pruning for efficient neural architecture search. In ECCV, pages 584-600, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[9] Yongxing Dai, Jun Liu, Yifan Sun, Zekun Tong, Chi Zhang, and Ling-Yu Duan. Idm: An intermediate domain module for domain adaptive person re-id. In ICCV, pages 11864-11874, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.679 + ], + "angle": 0, + "content": "[11] Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. Speeding up automatic hyperparameter optimization of deep neural networks by extrapolation of learning curves. In *IJCAI*, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.682, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[12] Xuanyi Dong and Yi Yang. Searching for a robust neural architecture in fourgpu hours. In CVPR, pages 1761-1770, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.723, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[14] Pengfei Fang, Jieming Zhou, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Bilinear attention networks for person retrieval. In ICCV, pages 8030-8039, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.828, + 0.469, + 0.867 + ], + "angle": 0, + "content": "[15] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.166 + ], + "angle": 0, + "content": "[16] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[17] Yixiao Ge, Dapeng Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. In NeurIPS, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.257 + ], + "angle": 0, + "content": "[18] Yiluan Guo and Ngai-Man Cheung. Efficient and deep person re-identification using multi-level similarity. In CVPR, pages 2335–2344, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.258, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[19] Xumeng Han, Xuehui Yu, Nan Jiang, Guorong Li, Jian Zhao, Qixiang Ye, and Zhenjun Han. Group sampling for unsupervised person re-identification. arXiv preprint arXiv:2107.03024, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.386 + ], + "angle": 0, + "content": "[21] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, pages 15013–15022, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.388, + 0.892, + 0.436 + ], + "angle": 0, + "content": "[22] Ruibing Hou, Bingpeng Ma, Hong Chang, Xinqian Gu, Shiguang Shan, and Xilin Chen. Interaction-and-aggregation network for person re-identification. In CVPR, pages 9317–9326, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.892, + 0.477 + ], + "angle": 0, + "content": "[23] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.479, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[24] Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Meta pairwise relationship distillation for unsupervised person re-identification. In ICCV, pages 3661–3670, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[25] Xin Jin, Cuiling Lan, Wenjun Zeng, Zhibo Chen, and Li Zhang. Style normalization and restitution for generalizable person re-identification. In CVPR, pages 3143–3152, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[26] Xin Jin, Cuiling Lan, Wenjun Zeng, Guoqiang Wei, and Zhibo Chen. Semantics-aligned representation learning for person re-identification. In AAAI, pages 11173–11180, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[27] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[28] Hanjun Li, Gaojie Wu, and Wei-Shi Zheng. Combined depth space based architecture search for person re-identification. In CVPR, pages 6729–6738, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.699, + 0.892, + 0.736 + ], + "angle": 0, + "content": "[29] Jianing Li and Shiliang Zhang. Joint visual and temporal consistency for unsupervised domain adaptive person re-identification. In ECCV, pages 483–499, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.738, + 0.892, + 0.776 + ], + "angle": 0, + "content": "[30] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deepreid: Deep filter pairing neural network for person re-identification. In CVPR, pages 152–159, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.778, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[31] Wei Li, Xiatian Zhu, and Shaogang Gong. Harmonious attention network for person re-identification. In CVPR, pages 2285–2294, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[32] Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded person re-identification with part-aware transformer. In CVPR, pages 2898–2907, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.503, + 0.898 + ], + "angle": 0, + "content": "19251" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.469, + 0.165 + ], + "angle": 0, + "content": "[33] Shengcai Liao and Ling Shao. Interpretable and generalizable person re-identification with query-adaptive convolution and temporal lifting. In ECCV, pages 456–474, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.166, + 0.469, + 0.215 + ], + "angle": 0, + "content": "[34] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shiens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In ECCV, pages 19–34, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.217, + 0.469, + 0.254 + ], + "angle": 0, + "content": "[35] Hanxiao Liu, Karen Simonyan, Oriol Vinyals, Chrisantha Fernando, and Koray Kavukcuoglu. Hierarchical representations for efficient architecture search. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.468, + 0.279 + ], + "angle": 0, + "content": "[36] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In ICLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[37] Hongye Liu, Yonghong Tian, Yaowei Yang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In CVPR, pages 2167-2175, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.331, + 0.468, + 0.369 + ], + "angle": 0, + "content": "[38] Xinchen Liu, Wu Liu, Huadong Ma, and Huiyuan Fu. Large-scale vehicle re-identification in urban surveillance videos. In ICME, pages 1–6, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.37, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[39] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In ECCV, pages 869–884, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[40] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. Provid: Progressive and multimodal vehicle reidentification for large-scale urban surveillance. T-MM, 20(3):645-658, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[41] Chuanchen Luo, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Spectral feature transformation for person re-identification. In ICCV, pages 4976–4985, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[42] Hao Luo, Wei Jiang, Youzhi Gu, Fuxu Liu, Xingyu Liao, Shenqi Lai, and Jianyang Gu. A strong baseline and batch normalization neck for deep person re-identification. T-MM, 22(10):2597-2609, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[43] Renqian Luo, Fei Tian, Tao Qin, Enhong Chen, and Tie-Yan Liu. Neural architecture optimization. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[44] Xuelin Qian, Yanwei Fu, Yu-Gang Jiang, Tao Xiang, and Xiangyang Xue. Multi-scale deep learning architectures for person re-identification. In ICCV, pages 5399–5408, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.638, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[45] Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, and Yi Yang. Auto-reid: Searching for a part-aware convnet for person re-identification. In ICCV, pages 3750–3759, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[46] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In AAAI, pages 4780-4789, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.778, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[48] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.865 + ], + "angle": 0, + "content": "[49] Yantao Shen, Tong Xiao, Hongsheng Li, Shuai Yi, and Xiaogang Wang. Learning deep neural networks for vehicle re-id with visual-spatio-temporal path proposals. In ICCV, pages 1900-1909, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.469, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[50] Liangchen Song, Cheng Wang, Lefei Zhang, Bo Du, Qian Zhang, Chang Huang, and Xinggang Wang. Unsupervised domain adaptive re-identification: Theory and practice. PR, 102:107173, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[51] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV, pages 480-496, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.282 + ], + "angle": 0, + "content": "[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.285, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.338, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[54] Dongkai Wang and Shiliang Zhang. Unsupervised person re-identification via multi-label classification. In CVPR, pages 10981-10990, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[55] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACMMM, pages 274–282, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[56] Yicheng Wang, Zhenzhong Chen, Feng Wu, and Gang Wang. Person re-identification with cascaded pairwise convolutions. In CVPR, pages 1470–1478, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.471, + 0.892, + 0.521 + ], + "angle": 0, + "content": "[57] Zhikang Wang, Lihuo He, Xiaoguang Tu, Jian Zhao, Xinbo Gao, Shengmei Shen, and Jiashi Feng. Robust video-based person re-identification by hierarchical mining. T-CSVT, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.524, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[58] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In ICCV, pages 379–387, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[59] Zhongdao Wang, Jingwei Zhang, Liang Zheng, Yixuan Liu, Yifan Sun, Yali Li, and Shengjin Wang. Cycles: Self-supervised cycle association for learning re-identifiable descriptions. In ECCV, pages 72–88, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.68 + ], + "angle": 0, + "content": "[60] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In CVPR, pages 79–88, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.682, + 0.892, + 0.721 + ], + "angle": 0, + "content": "[61] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.722, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[62] Yuhui Xu, Lingxi Xie, Xiaopeng Zhang, Xin Chen, Guo-Jun Qi, Qi Tian, and Hongkai Xiong. Pc-darts: Partial channel connections for memory-efficient architecture search. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.825 + ], + "angle": 0, + "content": "[63] Cheng Yan, Guansong Pang, Xiao Bai, Changhong Liu, Ning Xin, Lin Gu, and Jun Zhou. Beyond triplet loss: person re-identification with fine-grained difference-aware pairwise loss. T-MM, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.828, + 0.892, + 0.866 + ], + "angle": 0, + "content": "[64] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C.H. Hoi. Deep learning for person re-identification: A survey and outlook. T-PAMI, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "19252" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[65] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Deep metric learning for person re-identification. In ICPR, pages 34–39, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[66] Anguo Zhang, Yueming Gao, Yuzhen Niu, Wenxi Liu, and Yongcheng Zhou. Coarse-to-fine person re-identification with auxiliary-domain classification and second-order information bottleneck. In CVPR, pages 598–607, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.257 + ], + "angle": 0, + "content": "[67] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Xin Jin, and Zhibo Chen. Relation-aware global attention for person re-identification. In CVPR, pages 3186–3195, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.258, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[68] Aihua Zheng, Xianmin Lin, Jiacheng Dong, Wenzhong Wang, Jin Tang, and Bin Luo. Multi-scale attention vehicle re-identification. Neural Computing and Applications, 32(23):17489-17503, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[69] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, pages 1116–1124, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.386 + ], + "angle": 0, + "content": "[70] Liang Zheng, Yi Yang, and Alexander G Hauptmann. Person re-identification: Past, present and future. arXiv preprint arXiv:1610.02984, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.388, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[71] Liang Zheng, Yi Yang, and Qi Tian. Sift meets cnn: A decade survey of instance retrieval. T-PAMI, 40(5):1224-1244, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[72] Meng Zheng, Srikrishna Karanam, Ziyan Wu, and Richard J Radke. Re-identification with consistent attentive siamese networks. In CVPR, pages 5735–5744, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.466, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[73] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In CVPR, pages 2138–2147, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.554 + ], + "angle": 0, + "content": "[74] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, pages 13001-13008, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.557, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[75] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re-identification. In ICCV, pages 3702–3712, 2019. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[76] Sanping Zhou, Fei Wang, Zeyi Huang, and Jinjun Wang. Discriminative feature learning with consistent attention regularization for person re-identification. In ICCV, pages 8040–8049, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.698 + ], + "angle": 0, + "content": "[77] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697-8710, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[78] Yang Zou, Xiaodong Yang, Zhiding Yu, BVK Vijaya Kumar, and Jan Kautz. Joint disentangling and adaptation for cross-domain person re-identification. In ECCV, pages 87–104, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.748 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.468, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "19253" + } + ] +] \ No newline at end of file diff --git a/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_origin.pdf b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a38bd4722d83e6fc1e2a43bc442686c9290acc5e --- /dev/null +++ b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/5b190a6d-e663-4703-8784-f68b7d4a2470_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39c718f477d2d735f2e4cbbef6afdaff1fcd5f5204f74b43d2c1f3045244e8f3 +size 862087 diff --git a/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/full.md b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e3d398a71397abc04e0ad72d7e729a2fd606b1ab --- /dev/null +++ b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/full.md @@ -0,0 +1,382 @@ +# MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID + +Jianyang Gu + +Kai Wang + +Hao Luo3 + +Chen Chen + +Wei Jiang $^{1*}$ + +Yuqiang Fang + +Shanghang Zhang + +Yang You + +Jian Zhao $^{7*}$ + +$^{1}$ Zhejiang University + +$^{2}$ National University of Singapore + +3Alibaba Group + +$^{4}$ OPPO Research Institute + +5Space Engineering University + +6Peking University + +$^{7}$ Institute of North Electronic Equipment + +{gu_jianyang, jiangwei_zju}@zju.edu.cn zhaiqian90@u.nus.edu + +# Abstract + +Neural Architecture Search (NAS) has been increasingly appealing to the society of object Re-Identification (ReID), for that task-specific architectures significantly improve the retrieval performance. Previous works explore new optimizing targets and search spaces for NAS ReID, yet they neglect the difference of training schemes between image classification and ReID. In this work, we propose a novel Twins Contrastive Mechanism (TCM) to provide more appropriate supervision for ReID architecture search. TCM reduces the category overlaps between the training and validation data, and assists NAS in simulating real-world ReID training schemes. We then design a Multi-Scale Interaction (MSI) search space to search for rational interaction operations between multi-scale features. In addition, we introduce a Spatial Alignment Module (SAM) to further enhance the attention consistency confronted with images from different sources. Under the proposed NAS scheme, a specific architecture is automatically searched, named as MSINet. Extensive experiments demonstrate that our method surpasses state-of-the-art ReID methods on both indomain and cross-domain scenarios. Source code available in https://github.com/vimar-gu/MSINet. + +# 1. Introduction + +Object re-identification (Re-ID) aims at retrieving specific object instances across different views [39, 40, 57, 65, 70], which attracts much attention in computer vision community due to its wide-range applications. Previous works have achieved great progresses on both supervised [42, 49, 58] and unsupervised ReID tasks [17, 50, 78], most of which adopts backbone models originally designed for general image classification tasks [20, 52]. + +Recent literature [64, 75] has shown that applying differ + +![](images/f45c96a1a85b6862cf03b220f4f50423d45eb5257094decaf6cc688540b11d7d.jpg) +hood + +![](images/d56fa0f9226451dc5a875aef99222e193cb7719a117b1782c25cb1a45a3ea4ae.jpg) + +![](images/79556eb58cd1edc26469fa6bae015918a798943de8b82f5f7c567461ad4b04bd.jpg) + +![](images/054c15f3b8ca7fd16fe1809275a1d8dcdfccbebd0c8bc0165c7ae1f9f5fbbf3e.jpg) +luggage carrier +Figure 1. The left panel shows the example activation maps of ResNet50 (1st row) and MSINet (2nd row). The right panel shows the average distances between the most similar 10 negative samples and each query image at the inference. Best viewed in color. + +![](images/87ea7dbb407db1dfdbec3588fa670b59bf88a178a381deb80799c4a7e2522fad.jpg) +Distance between top-10 neg. samples and each query image + +ment architectures on ReID leads to large performance variations. Some works employ Neural Architecture Search (NAS) for ReID [28, 45]. The proposed optimizing targets and search spaces stably improve the model performance, yet the main search scheme still follows traditional NAS methods designed for general classification tasks [12, 36]. As an open-set task, ReID contains different categories in the training and validation sets [64, 71], while the two sets share exactly the same categories in standard classification tasks [10], which is also followed by traditional NAS methods. The incompatibility between search schemes and real-world training schemes makes the searched architecture sub-optimal for ReID. Moreover, ReID is required to distinguish more subtle distinctions among fine-grained instances compared with image-level classification [48, 63]. Some previous works [4, 44, 68, 75] have manifested that local perspectives and multi-scale features are discriminative for ReID. However, current utilizations of these features are mostly empirically designed, which can be more flexible according to the characteristics of different network layers. + +In this work, we propose a novel NAS scheme aiming at addressing the aforementioned challenges. In order to simulate the real-world ReID training schemes, a Twins Contrastive Mechanism (TCM) is proposed to unbind the categories of the training and validation sets. An adjustable overlap ratio of categories builds up the compatibility between NAS and ReID, which provides more appropriate supervision for ReID architecture search. Moreover, to search for more rational utilizations of multi-scale features, we design a Multi-Scale Interaction (MSI) search space. The MSI space focuses on interaction operations between multi-scale features along the shallow and deep layers of the network, which guides the features to promote each other. Additionally, to further improve the generalization capability, we propose a Spatial Alignment Module (SAM) to enhance the attention consistency of the model confronted with images from different sources. With the above NAS scheme, we obtain a light-weight yet effective model architecture, denoted as Multi-Scale Interaction Net (MSINet). + +We visualize the example activation maps of our proposed MSINet and ResNet50 [20] trained on VeRi-776 [38, 39] in Fig. 1. Compared to ResNet50, MSINet focuses on more unique distinctions with specific semantic information to recognize instances. Besides, MSINet largely increases the distance margin between query image and corresponding negative samples, reflecting extraordinary discriminative capability. Extensive experiments demonstrate that MSINet surpasses state-of-the-art (SOTA) ReID methods on both in-domain and cross-domain scenarios. Our source codes are available in the supplementary material. + +Our contributions are summarized as follows: + +- To the best of our knowledge, we are the first to build the NAS search scheme according to the real-world ReID training schemes, which provides more appropriate supervision for the ReID architecture search. +- We propose a novel search space based on the Multi-Scale Interaction (MSI) operations and a Spatial Alignment Module (SAM) to improve the model performance on in-domain and cross-domain scenarios. +- We construct a light-weight yet effective architecture for ReID tasks, denoted as MSINet. With only 2.3M parameters, MSINet surpasses ResNet50 [20] by $9\%$ mAP on MSMT17 [60] and $16\%$ mAP on MSMT17 $\rightarrow$ Market-1501 [69]. + +# 2. Related Works + +Neural Architecture Search. NAS has been increasingly appealing to the computer vision society, due to its automatic architecture designing characteristics. NAS methods can be roughly separated into four categories: reinforcement learning [1, 77], evolutionary algorithms [35, 46], gra + +dient descent [36, 43] and performance prediction [11, 34]. Liu et al. establish a differentiable architecture search (DARTS) method [36], which improves the practicability of NAS by a large extent. Some later works further improve the structure through sampling strategy [62], network pruning [2, 8], progressive learning [5], collaborative competition [7], etc. Most of NAS works focus on general image classification tasks, where the training and validation sets share the exact same categories. Following the setting, however, leads to incompatibility with the real-world training schemes of object ReID. In this work, we unbind the category bond between the two sets and propose a novel search scheme suitable for ReID. + +ReID Network Design. Current ReID works mostly adopt backbones designed for image classification [20, 23, 47, 52]. Some works [14, 31, 67] design attention modules based on the common backbones to unearth their potential on distinguishing local distinctions. However, these methods usually lead to large calculation consumption. + +There are also several works focusing on designing ReID-specific architectures. Li et al. present a Filter Pairing Neural Network to dynamically match patches in the feature maps [30]. Wang et al. separate and regroup the features of two samples with a WConv layer [56]. Guo et al. extract multi-scale features to directly evaluate the similarity between samples [18]. However, the siamese structure is inconvenient when conducting retrieval on large galleries. Zhou et al. aggregate multi-scale information to achieve high accuracy with small computing consumption [75]. Quan et al. introduce a part-aware module into the DARTS search space [36, 45]. Li et al. propose a new search space in regard to receptive field scales [28]. These methods have excellent performance on limited parameter scales, but fail to surpass those networks with complex structures. Different from previous works, we design a light-weight searching structure focusing on rational interaction operations between multi-scale features. The searched MSINet surpasses SOTA methods on both in-domain and cross-domain tasks. + +# 3. Methods + +Our goal is to construct an effective NAS scheme to search for a light-weight backbone architecture suitable for ReID tasks. Based on the training schemes of ReID, we propose a novel Twins Contrastive Mechanism to provide more appropriate supervision for the search process. Aiming at rational interaction between multi-scale features, we design a Multi-Scale Interaction search space. We further introduce a Spatial Alignment Module to improve the generalization capability with limited parameter growth. + +# 3.1. Twins Contrastive Mechanism + +NAS aims at automatically searching for the optimal network architecture for certain data. Inspired by [36], a basic + +![](images/93629ce15e822e4f454f33347d05317c4ccacf32239c4a665cee6a5fbe229fd9.jpg) +Figure 2. The model structure of the proposed MSINet. The input can be either person or vehicle. Inside a cell, the input is separated to two branches, with different receptive field scales. The interaction module exchanges information between two branches. Architecture search automatically selects the most appropriate interaction for each cell. + +differentiable architecture search scheme is established. We define the ordinary model parameters as $\omega$ , and architecture parameters as $\alpha$ . For network layer $i$ with a search space of $\mathcal{O}$ , $\alpha_{i}$ controls the weight of each operation $o$ in the space. The features are parallelly passed through all the operations, and the final output is formulated by the softmax-weighted sum of operation outputs: + +$$ +f \left(\mathbf {x} _ {i}\right) = \sum_ {o \in \mathcal {O}} \frac {\exp \left\{\alpha_ {i} ^ {o} \right\}}{\sum_ {o ^ {\prime} \in \mathcal {O}} \exp \left\{\alpha_ {i} ^ {o ^ {\prime}} \right\}} \cdot o \left(\mathbf {x} _ {i}\right). \tag {1} +$$ + +The search process is conducted in an alternative manner. Training data is utilized to update the model parameters, and validation data is then employed to update the architecture parameters. For most NAS methods designed for image classification tasks, the training and validation data share exactly the same categories and a linear classification layer for loss calculation. + +Different from standard image classification, as an open-set retrieval task, ReID has different categories in the training and validation sets. The incompatibility between search schemes and real-world training schemes might lead to suboptimal searching results. Accordingly, we propose a novel Twins Contrastive Mechanism (TCM) for NAS ReID training. Specifically, we employ two independent auxiliary memories $\mathcal{C}_{tr}$ and $\mathcal{C}_{val}$ to store the embedded features of the training and validation data, respectively. The memories are initialized with the centroid features, which are calculated by averaging the features of each category. At each iteration, the training loss is first calculated with $\mathcal{C}_{tr}$ for model parameter updating. Given an embedded feature $f$ with category label $j$ , the contrastive classification loss is calculated with: + +$$ +\mathcal {L} _ {t r} ^ {c l s} = - \log \frac {\exp (\mathbf {f} \cdot \mathbf {c} _ {t r} ^ {j} / \tau)}{\sum_ {n = 0} ^ {N _ {t r} ^ {c}} \exp (\mathbf {f} \cdot \mathbf {c} _ {t r} ^ {n} / \tau)}, \tag {2} +$$ + +where $\mathbf{c}_{tr}^n$ represents the memorized feature of category $n$ , $N_{tr}^c$ stands for the total number of categories in the training set, and $\tau$ is the temperature parameter, which is set as 0.05 empirically [17]. After updating the model parameters, the embedded feature $\mathbf{f}$ with category label $j$ is integrated into the corresponding memorized feature $\mathbf{c}_{tr}^j$ by: + +$$ +\mathbf {c} _ {t r} ^ {j} \leftarrow \beta \mathbf {c} _ {t r} ^ {j} + (1 - \beta) \mathbf {f}, \tag {3} +$$ + +where $\beta$ is set as 0.2 empirically [17]. Then the updated model is evaluated on the validation data to generate to validation loss with $C_{val}$ replacing $C_{tr}$ in Eq. 2. The architecture parameter is then updated with the validation loss to finish an iteration. + +As the loss calculation does not rely on the linear classification layer, the categories of the training and validation sets are unbound. We are able to dynamically adjust the category overlap ratio in these two sets. The advantages of a proper overlap ratio are summarized as two folds. Firstly, TCM better simulates the real-world training of ReID and helps the model focus on truly discriminative distinctions. The differences between the training and validation data improves the generalization capability of the model. Secondly, a relatively small proportion of overlapped categories stabilizes the architecture parameter update through a consistent optimizing target with the model parameter update. + +# 3.2. Multi-Scale Interaction Space + +Although the local perspective and multi-scale features have already been investigated in previous ReID works [4, 28, 44, 51, 68, 72, 75, 76], the utilization of these information is mainly empirically designed aggregation, which is monotonous and restrained. We argue that on the one hand, the rational utilization of multi-scale features should be dynamically adjusted along the shallow and deep layers of the network. On the other hand, introducing interaction other than aggregation creates direct information exchange, and + +![](images/5523ff0813347b36722b461219b27c29c12088ba4bdffab88dbf652e82b6f9aa.jpg) +Figure 3. The aligning pipeline of the proposed Spatial Alignment Module. The correlation activation vectors are calculated between the anchor feature and all features in a mini-batch. The positive vectors are aligned with the learnable self-activation, and the negative vectors are aligned with each other. The structure of PAM is shown on the left. + +Table 1. The detailed interaction operation in the proposed MSINet architecture. N: None; E: Exchange; G: Channel Gate; A: Cross Attention. + +
Cell #1Cell #2Cell #3Cell #4Cell #5Cell #6
123456789101112
GGEGAGGNGAEA
+ +makes fuller use of multi-scale features. Therefore, we propose a novel Multi-Scale Interaction (MSI) search space to establish a light-weight architecture suitable for ReID. + +As shown in Fig. 2, the network is mainly grouped with MSI cells and down-sample blocks, which is generally consistent with OSNet [75]. In each cell, the input features are passed through two branches with different receptive field scales. To reduce the calculation burden of the network, for the layers inside each branch, we adopt the stack of $1 \times 1$ convolution and multiple depth-wise $3 \times 3$ convolution to implement specific scales. A scale ratio $\rho$ of 3:1 is selected for the two branches. These two branches do not share model parameters, except for the Interaction Modules (IM). IM introduces information exchange for the two branches. There are 4 operation options for the IM. With the two-branch input features defined as $(\mathbf{x}_1, \mathbf{x}_2)$ , the operations can be formulated as: + +None. None operation involves no parameters, and outputs exactly the input features $(\mathbf{x}_1,\mathbf{x}_2)$ . + +Exchange. Exchange acts as the strongest interaction among all options. It directly exchanges the features for the two branches and outputs $(\mathbf{x}_2, \mathbf{x}_1)$ . Exchange contains no extra parameters, as well. + +Channel Gate. Channel gate introduces a Multi-Layer Perceptron (MLP) to generate a channel-wise attention gate [61, 75] as: + +$$ +G (\mathbf {x}) = \sigma (M L P (\mathbf {x})), \tag {4} +$$ + +![](images/0d19bcb32e290dde3c7cf4dc867064efe3f68de7668d2d5791e3709501adffe4.jpg) +Figure 4. Output feature maps of MSI cells at different layers. + +and returns $(G(\mathbf{x}_1) \cdot \mathbf{x}_1, G(\mathbf{x}_2) \cdot \mathbf{x}_2)$ . The MLP is composed of 2 fully connected layers and its parameters are shared for both branches. Thereby it achieves interaction by jointly screening discriminative feature channels. + +Cross Attention. Traditional channel attention module calculates the channel correlation inside a single feature map [15]. The original feature map $\mathbf{x} \in \mathcal{R}^{C \times H \times W}$ is firstly reshaped into the query feature $\tilde{\mathbf{x}} \in \mathcal{R}^{C \times N}$ , where $N = H \times W$ . Then the correlation activation is calculated by performing a matrix multiplication between the query feature $\tilde{\mathbf{x}}$ and the key feature $\tilde{\mathbf{x}}^{\top}$ . We propose to exchange the keys of the two branches to explicitly calculate the correlation between each other. The correlation activation is then transformed to a mask, and is added up to the original features with a learnable proportion. + +After interaction, the multi-scale branches are fused through a sum operation. It is worth noting that the extra parameters brought by multiple interaction options are limited, which enables searching for each cell along the whole network independently. At the beginning of the network, we employ the same stem module as that in OSNet [75], containing a $7 \times 7$ convolutional layer and a $3 \times 3$ max pooling with a stride of 2. After the searching process, the interaction operation $o$ with the largest weight $\alpha_{i}^{o}$ at each layer is reserved to form the searched architecture. + +After searching the architecture, the model is validated on various Re-ID tasks. The training is constrained by the + +Table 2. Supervised performance on object ReID datasets. The results in the top part are trained from scratch, and those in the bottom part are pre-trained on ImageNet in advance. As the compared methods are originally proposed for person ReID, we reproduce the results in vehicle datasets. * indicates that the results of person ReID are reproduced by us. The evaluation results of architecture searched on VR can be found in the supplementary material. + +
MethodParamsInference TimeMMSVRVIDMS→MVR→VID
R-1↑mAP↑R-1↑mAP↑R-1↑mAP↑R-1↑R-5↑R-1↑mAP↑R-1↑R-5↑
ResNet50* [42]~24M1×85.768.348.025.792.869.970.676.6--
OSNet [75]2.2M0.79×93.681.071.043.395.472.876.088.7----
CDNet [28]1.8M0.67×93.783.773.748.594.373.074.588.8----
MSINet2.3M0.71×94.687.076.052.595.975.076.589.8----
ResNet50* [42]~24M1×94.585.975.550.494.573.676.589.958.831.842.861.9
OSNet [75]2.2M0.79×94.884.978.752.995.576.476.088.666.637.546.563.1
CDNet [28]1.8M0.67×95.186.078.954.7--------
MSINet2.3M0.71×95.389.681.059.696.878.877.991.774.946.248.065.6
MSINet-SAM2.4M0.71×95.589.980.759.596.779.078.091.976.348.449.066.8
+ +classification id loss and the triplet loss, formulated by: + +$$ +\mathcal {L} _ {i d} = \frac {1}{N} \sum_ {i = 1} ^ {N} - \log \left(\frac {\exp \mathbf {W} _ {i} ^ {\top} \mathbf {f} _ {i}}{\sum_ {j} \exp \mathbf {W} _ {j} ^ {\top} \mathbf {f} _ {i}}\right), \tag {5} +$$ + +where $\mathbf{f}_i$ is a feature vector, the corresponding classifier weight of which is $\mathbf{W}_i$ , and + +$$ +\mathcal {L} _ {t r i} = \left[ \mathcal {D} \left(\mathbf {f} _ {a}, \mathbf {f} _ {p}\right) - \mathcal {D} \left(\mathbf {f} _ {a}, \mathbf {f} _ {n}\right) + m \right] _ {+}, \tag {6} +$$ + +where $\mathbf{f}_a, \mathbf{f}_p, \mathbf{f}_n$ are the embedded features for the anchor, the hardest positive and negative samples in a mini-batch, $\mathcal{D}(\cdot, \cdot)$ is the Euclidean distance, $m$ is the margin parameter, and $[\cdot]_{+}$ is the $\max(\cdot, 0)$ function. + +# 3.3. Spatial Alignment Module + +The retrieval precision of object ReID tasks are largely affected by the variation of appearances such as poses, illumination and occlusion when the camera conditions change. In order that the model correctly and consistently focuses on the discriminative spatial positions, we design a Spatial Alignment Module (SAM) to explicitly align the spatial attention between images, as shown in Fig. 3. + +Specifically, we first calculate the position-wise correlation activation map $\mathbf{A}$ between the feature maps in a minibatch. The activation between sample $i$ and $j$ can be formulated as: $\mathbf{A}(i,j) = \tilde{\mathbf{x}}_j^\top \times \tilde{\mathbf{x}}_i$ , where $\tilde{\mathbf{x}} \in \mathcal{R}^{C \times N}$ is reshaped from the original feature $\mathbf{x} \in \mathcal{R}^{C \times H \times W}$ . Then we take the maximum activation for each position of sample $i$ as: + +$$ +\mathbf {a} (i, j) = \max _ {d i m = 1} \mathbf {A} (i, j). \tag {7} +$$ + +The above process is denoted as "Mutual Conv" in Fig. 3. We evaluate the consistency between activation vectors with cosine similarity. For negative samples specifically, there + +can be many different hints for recognition, some of which might be inappropriate, such as the backgrounds. By aligning all the correlations for sample $i$ , we hope that the network can correct some attention bias and consistently focus on discriminative positions. + +However, through aligning positive sample pairs, the ID-related features are expected to be emphasized, which cannot be achieved by aligning negative pairs. Therefore, we introduce an extra position activation module (PAM) to generate supervision for the alignment between positive pairs. The spatial alignment loss is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {s a} (i) = \frac {1}{N _ {+}} \sum_ {p \in \mathcal {I} _ {+}} \left(1 - S (\hat {\mathbf {a}} (i), \mathbf {a} (i, p))\right) + \\ \frac {1}{N _ {-}} \sum_ {n _ {1}, n _ {2} \in \mathcal {I} _ {-}} (1 - S (\mathbf {a} (i, n _ {1}), \mathbf {a} (i, n _ {2}))), \\ \end{array} +$$ + +(8) + +where $\mathcal{I}_{+}$ contains positive indices for sample $i$ , the total number of which is $N_{+}$ , and vice versa. $\hat{\mathbf{a}}(i)$ stands for the generated activation vector for positive sample alignment, and $S(\cdot, \cdot)$ is the cosine similarity. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +Our proposed method is evaluated on two person ReID datasets Market-1501 [69], MSMT17 [60], and two vehicle ReID datasets VeRi-776 [38, 39] and VehicleID [37]. For simplicity, the four datasets are denoted as M, MS, VR and VID in the following sections, respectively. Evaluation metrics include Cumulative Matching Characteristic (CMC) and mean average precision (mAP), which are commonly utilized on ReID tasks. + +Table 3. Supervised performance comparison between MSINet and SOTA methods on M and MS datasets. + +
MethodMMS
R-1↑mAP↑R-1↑mAP↑
PCB [51]93.881.668.240.4
MGN [55]95.786.976.952.1
OSNet [75]93.681.071.043.3
IANet [22]94.483.175.546.8
DGNet [73]94.886.077.252.3
Auto-ReID [45]94.585.1--
SAN [26]96.188.079.255.7
CDNet [28]95.186.078.954.7
BAT-Net [14]95.187.479.556.8
SFT [41]94.187.579.058.3
CTF [66]94.887.7--
RGA-SC [67]96.188.480.357.5
MSINet95.389.681.059.6
+ +# 4.2. Architecture Search + +We conduct the searching process on MSMT17. SGD is adopted for model parameter update with an initial learning rate of 0.025. The model is trained for 350 epochs in total. We adopt a warm-up strategy for the first 10 epochs. Then the learning rate is decayed by 0.1 at 150, 225 and 300 epochs, respectively. Adam [27] is adopted for the architecture parameter update with an initial learning rate of 0.002. The learning rate is decayed at the same pace. The images are reshaped to $256 \times 128$ for person and $256 \times 256$ for vehicles. Data augmentation includes random flip, random crop and random erasing [74]. The searched architecture is presented in Tab. 1. The "MSINet" in the following experiment sections refers to this architecture. + +We visualize the feature maps extracted by each MSI cell in Fig. 4. At the shallow layers of the network, the kernels mainly focus on overall contour information. Channel gate helps to filter out inferior information, such as the background. As we approach deeper layers, the extracted features each have specific semantic information, where cross attention is more likely to be selected for the interaction. It indicates that cross attention is more rational for exchanging high-level semantic information. + +# 4.3. Comparison with Other Backbones + +We first compare our proposed MSINet with ResNet50 and recent proposed light-weight backbones in both in-domain and cross-domain ReID scenarios. + +In-Domain Tasks. We adopt a two-group supervised evaluation scheme similar to that in [28, 75]: training from scratch and fine-tuning ImageNet [10] pre-trained models. The training parameters for both schemes are kept the same as that in architecture search, except for an initial learning rate of 0.065. Triplet loss and cross entropy loss are adopted + +Table 4. Unsupervised performance applying MSINet to SOTA methods for USL on M and UDA on $\mathrm{M} \rightarrow \mathrm{{MS}}$ . + +
MethodMM→MS
R-1↑mAP↑R-1↑mAP↑
MMCL [54]80.345.540.815.1
MMT [16]--50.124.0
JVTC+ [29]79.547.548.625.1
CycAs [59]84.864.8--
GCL [3]87.366.851.127.0
MPRD [24]83.051.0--
SpCL [17]88.173.153.726.8
GS [19]92.379.2--
GS+MSINet91.781.5--
HDCRL [6]92.481.7--
HDCRL+MSINet92.982.7--
IDM [9]--61.333.5
IDM+MSINet--66.037.8
+ +for the parameter update. The margin $m$ in Eq. 6 is set as 0.3. [28] adopts an FBLNeck. We also employ the same structure. The results are shown in Tab. 2. + +ResNet50 is the most commonly utilized backbone network in ReID tasks, yet holds the worst performance. Moreover, ResNet50 largely depends on ImageNet pre-training, while MSINet without pre-training has already surpassed pre-trained ResNet50 on all metrics. Compared with the other datasets, MS contains more variations on illumination, background and camera pose, and brings a large performance gap between ResNet50 and other methods. It also validates the inadequacy of image classification networks on ReID tasks. OSNet [75] and CDNet [28] are recently proposed architectures designed specifically for ReID tasks. Both architectures focus on fusing multi-scale features to better suit ReID. CDNet employs a traditional NAS scheme to search for the proper receptive field scales for each cell. MSINet fixes the receptive field scale and instead selects optimal interaction operations inside each cell. With only a bit more parameters, MSINet surpasses all the other backbones by a large margin. + +Cross-Domain Tasks. Cross-domain experiments verify the generalization capability of the model. Following previous domain generalizable ReID works [25, 33], data augmentation is adjusted to random flip, random crop and color jittering. The model is pre-trained and fine-tuned for 250 epochs to avoid over-fitting. The other settings are kept the same as supervision scenes. With no present pre-trained models for CDNet [28], it is excluded from this section. + +Tab. 2 shows that ResNet50 can be easily interfered by different image styles confronted with new image domains. OSNet learns multi-scale features with specific semantic information for ReID, which is domain invariant to some extent. Our proposed search scheme also takes into ac + +![](images/51c2fbd7e148acffa94d21e9d5888f89d2f1d69179c45870fdab237151ac504b.jpg) +(a) + +![](images/ee5a3bfa957f5cbf48be410b91cbae67f51a37e44d21cae0ad9d9e3e08e97fcd.jpg) +(b) + +![](images/7fb17dd76fc40c0ad73e6f3fdeb75235051a3d3d4eaf379aedcc200ae10e1eb1.jpg) +(c) + +![](images/84d3e0d04855a8721be3990a139f9aa43aa4625c7d735ac08847e1c2b9bc49d8.jpg) +(d) +Figure 5. Ablation studies on (a) search scheme; (b) architecture; (c) multi-scale aggregation; (d) scale ratio $\rho$ ; (e) spatial alignment weight. + +![](images/480a979402b8bf2edbe704b3d93a387f6878636a7f76d1d2ab31c89b772c7e41.jpg) +(e) + +Table 5. The effectiveness of each components in SAM. Both in-domain performance on VR and cross-domain performance on $\mathrm{VR}\rightarrow \mathrm{VID}$ is evaluated. + +
ModelPos.Neg.AlignVRVR→VID
R-1↑mAP↑R-1↑R-5↑
MSINet✓-96.778.548.065.6
Self96.778.148.466.0
✓Self96.778.448.566.2
✓✓Unified96.678.348.365.8
✓✓Separated96.878.648.766.4
✓✓PAM-Self96.779.049.066.8
OSNet✓✓None95.576.446.563.1
PAM-Self95.976.347.563.3
+ +count the generalization capability of the model. By partly separating the categories for training and validation sets, the searched interaction operations generalize well confronted with new image domains. Except for discrimination, MSINet also surpasses the other backbones on cross-domain tasks by a large margin with faster inference speed. + +Additionally, we introduce SAM into the model, which aligns the spatial correlations between images. A weighted sum of ReID loss and spatial alignment loss is utilized when training the network with SAM. The weight of spatial alignment loss is set as $\lambda_{sa} = 2.0$ . Without extra inference consumption or damages on the supervised performance, SAM further boosts the generalization capability of MSINet. + +# 4.4. Comparison with State-of-the-art Methods + +Tab. 3 further illustrates the supervision performance comparison of our proposed MSINet with the SOTA methods on M and MS datasets. With much less parameters than most of the compared methods, MSINet achieves a retrieval accuracy comparable to that of more complicated ones. Auto-ReID [45] first designs a NAS scheme for ReID, yet the DARTS-style architecture contains 13M parameters. RGA-SC [67] carefully designs a relation-aware global at + +tention module. MSINet achieves even higher performance with less training consumption, which validates the superiority of selecting rational interaction. + +We also evaluate the model performance replacing the backbone network from ResNet50 to MSINet for SOTA unsupervised ReID methods in Tab. 4. For purely unsupervised learning (USL) method GS [19] on M dataset, MSINet performs slightly lower on rank-1, yet has a large superiority on mAP. For HDCRL [6], MSINet shows obvious superiority over ResNet50. For unsupervised domain adaptation (UDA) method IDM [9] on $\mathrm{M}\rightarrow \mathrm{MS}$ task, MSINet surpasses ResNet50 by a large margin, which further proves that the TCM brings outstanding generalization capability to the searched architecture. + +# 4.5. Ablation Studies + +Effectiveness of Architecture Search. To verify the effectiveness, we conduct supervised training on MS with different search schemes in Fig. 5 (a). Under the standard classification scheme ("CE Overlap"), the searched model performs poor. Replacing the cross entropy loss to contrastive loss ("TCM Overlap") only brings a slight improvement. The complete TCM framework unbounds the categories between the training and validation set, and thereby improves the performance by a large margin. + +In Fig. 5 (b) we compare the performance of different architectures. Firstly, we validate 4 models each with a unique interaction operation from the 4 options in the search space. None and Exchange, with no trainable parameters, achieve poor performance. Channel gate introduces channel-wise attention, whose model performs the best among 4 options. Cross attention exchanges the key features for the two branches. Over-frequent exchange interferes the ordinary feature extraction and degrades the network performance. Through appropriately arranging interaction operations along the architecture, MSINet surpasses all the above 4 models. Random architecture, on the other hand, shows no rational appliance of interaction operations, which validates that the proposed search scheme helps find + +Table 6. Supervised performance comparison between MSINet and Transformers on VR and MS datasets. + +
MethodParamsInference TimeMSVR
R-1↑mAP↑R-1↑mAP↑
DeiT-S [21]~22M0.97x76.355.295.576.3
DeiT-B [21]~86M1.79x81.961.495.978.4
ViT-B [21]~86M1.79x81.861.096.578.2
MSINet2.3M0.71x81.059.696.878.8
+ +![](images/8316f43b4fa82cca870622977851a032b0a78b4336f37cc468b944698b660efd.jpg) +Figure 6. Example top-15 retrieved sequences comparison on VR. Appearance differences caused by variant camera conditions are well addressed by the proposed MSINet. Visualization of person ReID can be found in the supplementary material. + +suitable architectures for ReID. + +Effectiveness of Spatial Alignment Module. We validate the effectiveness of each components of SAM on the VR $\rightarrow$ VID cross-domain experiment in Tab. 5. Firstly, we introduce the spatial alignment for positive and negative sample pairs, respectively. Each of them brings certain performance improvements. However, a unified alignment for all sample pairs damages ID-related features and degrades the performance instead. Therefore, we separate the alignment of positive and negative samples, which retains some discriminative features and integrates the effect of both aspects. The extra PAM for positive sample alignment further guarantees the focus on ID-related positions and achieves the best performance. We also conduct in-domain experiment on VR to prove that SAM improves the generalization capability without sacrificing the supervision performance. Adding SAM to OSNet receives similar results, which validates the universality of SAM. + +Fusing Operation. After interaction, the multi-scale features are fused by sum operation. We investigate several fusing options on MS training from scratch in Fig. 5 (c). Subtracting ("Minus") a branch from the other leads to about the same results as "Sum" while multiplication ("Mul") performs poorly. + +Comparison with Transformer. Transformer, as a new architecture, has recently been continuously making progresses in many computer vision domains [13, 53], includ + +ing ReID [21, 32]. We compare the model performance with some baseline Transformer models in Tab. 6. DeiT-B and ViT-B [21] achieves higher performance on MSMT17, with much larger calculation burden compared with our proposed MSINet. On VeRi-776, MSINet surpasses all the baseline Transformer methods. It proves that rational interaction operations between multi-scale features are capable to assist light-weighted pure-CNN models to obtain comparable performance with complex Transformers. + +Parameter Analysis. Firstly, we study the influence of different receptive field scale ratios $\rho$ inside an MSI cell on MS training from scratch in Fig. 5 (d). Introducing scale differences between branches improves the model performance significantly, and subsequent increases brings more modest impacts. Considering both parameter scales and model performance, the ratio of 3:1 is selected for MSINet. + +Secondly, the model performance fluctuation influenced by spatial alignment weight is visualized in Fig. 5 (e). The experiment is conducted on the VR $\rightarrow$ VID cross-domain scenario. Employing the alignment generally makes a positive impact on the generalization capability of the model. The optimal loss weight $\lambda_{sa}$ locates at 2.0. + +Visualization Results. We visualize the top-15 retrieved sequences and the corresponding distances from an query image on VR in Fig. 6. By comparison, ResNet50 mainly focuses on general appearance features, where the top-rated negative samples share similar car bodies. MSINet, oppositely, concentrates on discriminative distinctions, an empty car hopper in this case, and creates an evident distance gap between positive and negative samples. More details can be seen in the supplementation material. + +# 5. Conclusion + +In this paper, we design a Twins Contrastive Mechanism for NAS to build the compatibility with ReID. The task-specific search scheme provides the searching process with more appropriate supervision. A Multi-Scale Interaction search space is proposed to establish rational and flexible utilization of multi-scale features. With a Spatial Alignment Module, our proposed MSINet achieves SOTA performance on both supervision and cross-domain scenarios with limited parameter amount. We hope the proposed approach could inspire more works focusing on designing network architectures suitable for ReID tasks. + +# 6. Acknowledgement + +This work is partially supported by National Natural Science Foundation of China (62173302, 62006244), China Scholarship Council (202206320302), Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and AI Singapore Programme (AISG2-PhD-2021-08-008). + +# References + +[1] Irwan Bello, Barret Zoph, Vijay Vasudevan, and Quoc V Le. Neural optimizer search with reinforcement learning. In ICML, pages 459-468, 2017. 2 +[2] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. In ICLR, 2018. 2 +[3] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Joint generative and contrastive learning for unsupervised person re-identification. In CVPR, pages 2004–2013, 2021. 6 +[4] Tsai-Shien Chen, Chih-Ting Liu, Chih-Wei Wu, and Shao-Yi Chien. Orientation-aware vehicle re-identification with semantics-guided part attention network. In ECCV, pages 330–346, 2020. 1, 3 +[5] Xin Chen, Lingxi Xie, Jun Wu, and Qi Tian. Progressive differentiable architecture search: Bridging the depth gap between search and evaluation. In ICCV, pages 1294-1303, 2019. 2 +[6] De Cheng, Jingyu Zhou, Nannan Wang, and Xinbo Gao. Hybrid dynamic contrast and probability distillation for unsupervised person re-id. IEEE Transactions on Image Processing, 31:3334-3346, 2022. 6, 7 +[7] Xiangxiang Chu, Tianbao Zhou, Bo Zhang, and Jixiang Li. Fair darts: Eliminating unfair advantages in differentiable architecture search. In ECCV, pages 465-480, 2020. 2 +[8] Xiyang Dai, Dongdong Chen, Mengchen Liu, Yinpeng Chen, and Lu Yuan. Da-nas: Data adapted pruning for efficient neural architecture search. In ECCV, pages 584-600, 2020. 2 +[9] Yongxing Dai, Jun Liu, Yifan Sun, Zekun Tong, Chi Zhang, and Ling-Yu Duan. Idm: An intermediate domain module for domain adaptive person re-id. In ICCV, pages 11864-11874, 2021. 6, 7 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 1, 6 +[11] Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. Speeding up automatic hyperparameter optimization of deep neural networks by extrapolation of learning curves. In *IJCAI*, 2015. 2 +[12] Xuanyi Dong and Yi Yang. Searching for a robust neural architecture in fourgpu hours. In CVPR, pages 1761-1770, 2019. 1 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 8 +[14] Pengfei Fang, Jieming Zhou, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Bilinear attention networks for person retrieval. In ICCV, pages 8030-8039, 2019. 2, 6 +[15] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 4 + +[16] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In ICLR, 2019. 6 +[17] Yixiao Ge, Dapeng Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. In NeurIPS, 2020. 1, 3, 6 +[18] Yiluan Guo and Ngai-Man Cheung. Efficient and deep person re-identification using multi-level similarity. In CVPR, pages 2335–2344, 2018. 2 +[19] Xumeng Han, Xuehui Yu, Nan Jiang, Guorong Li, Jian Zhao, Qixiang Ye, and Zhenjun Han. Group sampling for unsupervised person re-identification. arXiv preprint arXiv:2107.03024, 2021. 6, 7 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2 +[21] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, pages 15013–15022, 2021. 8 +[22] Ruibing Hou, Bingpeng Ma, Hong Chang, Xinqian Gu, Shiguang Shan, and Xilin Chen. Interaction-and-aggregation network for person re-identification. In CVPR, pages 9317–9326, 2019. 6 +[23] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 2 +[24] Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Meta pairwise relationship distillation for unsupervised person re-identification. In ICCV, pages 3661–3670, 2021. 6 +[25] Xin Jin, Cuiling Lan, Wenjun Zeng, Zhibo Chen, and Li Zhang. Style normalization and restitution for generalizable person re-identification. In CVPR, pages 3143–3152, 2020. 6 +[26] Xin Jin, Cuiling Lan, Wenjun Zeng, Guoqiang Wei, and Zhibo Chen. Semantics-aligned representation learning for person re-identification. In AAAI, pages 11173–11180, 2020. 6 +[27] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6 +[28] Hanjun Li, Gaojie Wu, and Wei-Shi Zheng. Combined depth space based architecture search for person re-identification. In CVPR, pages 6729–6738, 2021. 1, 2, 3, 5, 6 +[29] Jianing Li and Shiliang Zhang. Joint visual and temporal consistency for unsupervised domain adaptive person re-identification. In ECCV, pages 483–499, 2020. 6 +[30] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deepreid: Deep filter pairing neural network for person re-identification. In CVPR, pages 152–159, 2014. 2 +[31] Wei Li, Xiatian Zhu, and Shaogang Gong. Harmonious attention network for person re-identification. In CVPR, pages 2285–2294, 2018. 2 +[32] Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded person re-identification with part-aware transformer. In CVPR, pages 2898–2907, 2021. 8 + +[33] Shengcai Liao and Ling Shao. Interpretable and generalizable person re-identification with query-adaptive convolution and temporal lifting. In ECCV, pages 456–474, 2020. 6 +[34] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shiens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In ECCV, pages 19–34, 2018. 2 +[35] Hanxiao Liu, Karen Simonyan, Oriol Vinyals, Chrisantha Fernando, and Koray Kavukcuoglu. Hierarchical representations for efficient architecture search. In ICLR, 2018. 2 +[36] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In ICLR, 2018. 1, 2 +[37] Hongye Liu, Yonghong Tian, Yaowei Yang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In CVPR, pages 2167-2175, 2016. 5 +[38] Xinchen Liu, Wu Liu, Huadong Ma, and Huiyuan Fu. Large-scale vehicle re-identification in urban surveillance videos. In ICME, pages 1–6, 2016. 2, 5 +[39] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In ECCV, pages 869–884, 2016. 1, 2, 5 +[40] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. Provid: Progressive and multimodal vehicle reidentification for large-scale urban surveillance. T-MM, 20(3):645-658, 2017. 1 +[41] Chuanchen Luo, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Spectral feature transformation for person re-identification. In ICCV, pages 4976–4985, 2019. 6 +[42] Hao Luo, Wei Jiang, Youzhi Gu, Fuxu Liu, Xingyu Liao, Shenqi Lai, and Jianyang Gu. A strong baseline and batch normalization neck for deep person re-identification. T-MM, 22(10):2597-2609, 2019. 1, 5 +[43] Renqian Luo, Fei Tian, Tao Qin, Enhong Chen, and Tie-Yan Liu. Neural architecture optimization. In NeurIPS, 2018. 2 +[44] Xuelin Qian, Yanwei Fu, Yu-Gang Jiang, Tao Xiang, and Xiangyang Xue. Multi-scale deep learning architectures for person re-identification. In ICCV, pages 5399–5408, 2017. 1, 3 +[45] Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, and Yi Yang. Auto-reid: Searching for a part-aware convnet for person re-identification. In ICCV, pages 3750–3759, 2019. 1, 2, 6, 7 +[46] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In AAAI, pages 4780-4789, 2019. 2 +[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 2 +[48] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 1 +[49] Yantao Shen, Tong Xiao, Hongsheng Li, Shuai Yi, and Xiaogang Wang. Learning deep neural networks for vehicle re-id with visual-spatio-temporal path proposals. In ICCV, pages 1900-1909, 2017. 1 + +[50] Liangchen Song, Cheng Wang, Lefei Zhang, Bo Du, Qian Zhang, Chang Huang, and Xinggang Wang. Unsupervised domain adaptive re-identification: Theory and practice. PR, 102:107173, 2020. 1 +[51] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV, pages 480-496, 2018. 3, 6 +[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 1, 2 +[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 8 +[54] Dongkai Wang and Shiliang Zhang. Unsupervised person re-identification via multi-label classification. In CVPR, pages 10981-10990, 2020. 6 +[55] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACMMM, pages 274–282, 2018. 6 +[56] Yicheng Wang, Zhenzhong Chen, Feng Wu, and Gang Wang. Person re-identification with cascaded pairwise convolutions. In CVPR, pages 1470–1478, 2018. 2 +[57] Zhikang Wang, Lihuo He, Xiaoguang Tu, Jian Zhao, Xinbo Gao, Shengmei Shen, and Jiashi Feng. Robust video-based person re-identification by hierarchical mining. T-CSVT, 2021. 1 +[58] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In ICCV, pages 379–387, 2017. 1 +[59] Zhongdao Wang, Jingwei Zhang, Liang Zheng, Yixuan Liu, Yifan Sun, Yali Li, and Shengjin Wang. Cycles: Self-supervised cycle association for learning re-identifiable descriptions. In ECCV, pages 72–88, 2020. 6 +[60] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In CVPR, pages 79–88, 2018. 2, 5 +[61] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 4 +[62] Yuhui Xu, Lingxi Xie, Xiaopeng Zhang, Xin Chen, Guo-Jun Qi, Qi Tian, and Hongkai Xiong. Pc-darts: Partial channel connections for memory-efficient architecture search. In ICLR, 2019. 2 +[63] Cheng Yan, Guansong Pang, Xiao Bai, Changhong Liu, Ning Xin, Lin Gu, and Jun Zhou. Beyond triplet loss: person re-identification with fine-grained difference-aware pairwise loss. T-MM, 2021. 1 +[64] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C.H. Hoi. Deep learning for person re-identification: A survey and outlook. T-PAMI, 2021. 1 + +[65] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Deep metric learning for person re-identification. In ICPR, pages 34–39, 2014. 1 +[66] Anguo Zhang, Yueming Gao, Yuzhen Niu, Wenxi Liu, and Yongcheng Zhou. Coarse-to-fine person re-identification with auxiliary-domain classification and second-order information bottleneck. In CVPR, pages 598–607, 2021. 6 +[67] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Xin Jin, and Zhibo Chen. Relation-aware global attention for person re-identification. In CVPR, pages 3186–3195, 2020. 2, 6, 7 +[68] Aihua Zheng, Xianmin Lin, Jiacheng Dong, Wenzhong Wang, Jin Tang, and Bin Luo. Multi-scale attention vehicle re-identification. Neural Computing and Applications, 32(23):17489-17503, 2020. 1, 3 +[69] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, pages 1116–1124, 2015. 2, 5 +[70] Liang Zheng, Yi Yang, and Alexander G Hauptmann. Person re-identification: Past, present and future. arXiv preprint arXiv:1610.02984, 2016. 1 +[71] Liang Zheng, Yi Yang, and Qi Tian. Sift meets cnn: A decade survey of instance retrieval. T-PAMI, 40(5):1224-1244, 2017. 1 +[72] Meng Zheng, Srikrishna Karanam, Ziyan Wu, and Richard J Radke. Re-identification with consistent attentive siamese networks. In CVPR, pages 5735–5744, 2019. 3 +[73] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In CVPR, pages 2138–2147, 2019. 6 +[74] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, pages 13001-13008, 2020. 6 +[75] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re-identification. In ICCV, pages 3702–3712, 2019. 1, 2, 3, 4, 5, 6 +[76] Sanping Zhou, Fei Wang, Zeyi Huang, and Jinjun Wang. Discriminative feature learning with consistent attention regularization for person re-identification. In ICCV, pages 8040–8049, 2019. 3 +[77] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697-8710, 2018. 2 +[78] Yang Zou, Xiaodong Yang, Zhiding Yu, BVK Vijaya Kumar, and Jan Kautz. Joint disentangling and adaptation for cross-domain person re-identification. In ECCV, pages 87–104, 2020. 1 \ No newline at end of file diff --git a/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/images.zip b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..37230b8c77bab0f46a9732b1b3757d7eecd7e0c0 --- /dev/null +++ b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01d3180949f87ee0cf574eabc3ab09fa7368a01cfbff1c26c3254c66211f742c +size 558804 diff --git a/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/layout.json b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a4e02dbb5c75adb95c97f0fbad925029ebfef557 --- /dev/null +++ b/2023/MSINet_ Twins Contrastive Search of Multi-Scale Interaction for Object ReID/layout.json @@ -0,0 +1,9951 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 136, + 523, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 523, + 153 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 523, + 153 + ], + "type": "text", + "content": "MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 175, + 186, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 175, + 186, + 189 + ], + "spans": [ + { + "bbox": [ + 122, + 175, + 186, + 189 + ], + "type": "text", + "content": "Jianyang Gu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 175, + 253, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 175, + 253, + 189 + ], + "spans": [ + { + "bbox": [ + 201, + 175, + 253, + 189 + ], + "type": "text", + "content": "Kai Wang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 266, + 175, + 313, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 175, + 313, + 189 + ], + "spans": [ + { + "bbox": [ + 266, + 175, + 313, + 189 + ], + "type": "text", + "content": "Hao Luo3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 175, + 384, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 175, + 384, + 189 + ], + "spans": [ + { + "bbox": [ + 326, + 175, + 384, + 189 + ], + "type": "text", + "content": "Chen Chen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 397, + 175, + 453, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 175, + 453, + 189 + ], + "spans": [ + { + "bbox": [ + 397, + 175, + 453, + 189 + ], + "type": "text", + "content": "Wei Jiang" + }, + { + "bbox": [ + 397, + 175, + 453, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 189, + 207, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 189, + 207, + 202 + ], + "spans": [ + { + "bbox": [ + 136, + 189, + 207, + 202 + ], + "type": "text", + "content": "Yuqiang Fang" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 220, + 189, + 311, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 189, + 311, + 203 + ], + "spans": [ + { + "bbox": [ + 220, + 189, + 311, + 203 + ], + "type": "text", + "content": "Shanghang Zhang" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 189, + 376, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 189, + 376, + 202 + ], + "spans": [ + { + "bbox": [ + 324, + 189, + 376, + 202 + ], + "type": "text", + "content": "Yang You" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 189, + 443, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 189, + 443, + 202 + ], + "spans": [ + { + "bbox": [ + 388, + 189, + 443, + 202 + ], + "type": "text", + "content": "Jian Zhao" + }, + { + "bbox": [ + 388, + 189, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{7*}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 108, + 202, + 207, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 202, + 207, + 216 + ], + "spans": [ + { + "bbox": [ + 108, + 202, + 207, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 202, + 207, + 216 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 220, + 202, + 380, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 202, + 380, + 216 + ], + "spans": [ + { + "bbox": [ + 220, + 202, + 380, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 220, + 202, + 380, + 216 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 390, + 202, + 468, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 202, + 468, + 216 + ], + "spans": [ + { + "bbox": [ + 390, + 202, + 468, + 216 + ], + "type": "text", + "content": "3Alibaba Group" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 98, + 216, + 218, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 216, + 218, + 229 + ], + "spans": [ + { + "bbox": [ + 98, + 216, + 218, + 229 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 98, + 216, + 218, + 229 + ], + "type": "text", + "content": "OPPO Research Institute" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 220, + 216, + 377, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 216, + 377, + 230 + ], + "spans": [ + { + "bbox": [ + 220, + 216, + 377, + 230 + ], + "type": "text", + "content": "5Space Engineering University" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 390, + 216, + 479, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 216, + 479, + 230 + ], + "spans": [ + { + "bbox": [ + 390, + 216, + 479, + 230 + ], + "type": "text", + "content": "6Peking University" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 193, + 230, + 384, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 230, + 384, + 244 + ], + "spans": [ + { + "bbox": [ + 193, + 230, + 384, + 244 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 193, + 230, + 384, + 244 + ], + "type": "text", + "content": "Institute of North Electronic Equipment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 245, + 439, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 245, + 439, + 256 + ], + "spans": [ + { + "bbox": [ + 134, + 245, + 439, + 256 + ], + "type": "text", + "content": "{gu_jianyang, jiangwei_zju}@zju.edu.cn zhaiqian90@u.nus.edu" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 139, + 283, + 185, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 283, + 185, + 295 + ], + "spans": [ + { + "bbox": [ + 139, + 283, + 185, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 45, + 308, + 281, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 308, + 281, + 564 + ], + "spans": [ + { + "bbox": [ + 45, + 308, + 281, + 564 + ], + "type": "text", + "content": "Neural Architecture Search (NAS) has been increasingly appealing to the society of object Re-Identification (ReID), for that task-specific architectures significantly improve the retrieval performance. Previous works explore new optimizing targets and search spaces for NAS ReID, yet they neglect the difference of training schemes between image classification and ReID. In this work, we propose a novel Twins Contrastive Mechanism (TCM) to provide more appropriate supervision for ReID architecture search. TCM reduces the category overlaps between the training and validation data, and assists NAS in simulating real-world ReID training schemes. We then design a Multi-Scale Interaction (MSI) search space to search for rational interaction operations between multi-scale features. In addition, we introduce a Spatial Alignment Module (SAM) to further enhance the attention consistency confronted with images from different sources. Under the proposed NAS scheme, a specific architecture is automatically searched, named as MSINet. Extensive experiments demonstrate that our method surpasses state-of-the-art ReID methods on both indomain and cross-domain scenarios. Source code available in https://github.com/vimar-gu/MSINet." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 586, + 123, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 123, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 123, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 45, + 606, + 279, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 606, + 279, + 699 + ], + "spans": [ + { + "bbox": [ + 45, + 606, + 279, + 699 + ], + "type": "text", + "content": "Object re-identification (Re-ID) aims at retrieving specific object instances across different views [39, 40, 57, 65, 70], which attracts much attention in computer vision community due to its wide-range applications. Previous works have achieved great progresses on both supervised [42, 49, 58] and unsupervised ReID tasks [17, 50, 78], most of which adopts backbone models originally designed for general image classification tasks [20, 52]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 57, + 700, + 279, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 700, + 279, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 700, + 279, + 712 + ], + "type": "text", + "content": "Recent literature [64, 75] has shown that applying differ" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 300, + 283, + 362, + 346 + ], + "blocks": [ + { + "bbox": [ + 300, + 283, + 362, + 346 + ], + "lines": [ + { + "bbox": [ + 300, + 283, + 362, + 346 + ], + "spans": [ + { + "bbox": [ + 300, + 283, + 362, + 346 + ], + "type": "image", + "image_path": "f45c96a1a85b6862cf03b220f4f50423d45eb5257094decaf6cc688540b11d7d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 297, + 389, + 308 + ], + "lines": [ + { + "bbox": [ + 364, + 297, + 389, + 308 + ], + "spans": [ + { + "bbox": [ + 364, + 297, + 389, + 308 + ], + "type": "text", + "content": "hood" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 396, + 284, + 458, + 346 + ], + "blocks": [ + { + "bbox": [ + 396, + 284, + 458, + 346 + ], + "lines": [ + { + "bbox": [ + 396, + 284, + 458, + 346 + ], + "spans": [ + { + "bbox": [ + 396, + 284, + 458, + 346 + ], + "type": "image", + "image_path": "d56fa0f9226451dc5a875aef99222e193cb7719a117b1782c25cb1a45a3ea4ae.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 461, + 285, + 524, + 394 + ], + "blocks": [ + { + "bbox": [ + 461, + 285, + 524, + 394 + ], + "lines": [ + { + "bbox": [ + 461, + 285, + 524, + 394 + ], + "spans": [ + { + "bbox": [ + 461, + 285, + 524, + 394 + ], + "type": "image", + "image_path": "79556eb58cd1edc26469fa6bae015918a798943de8b82f5f7c567461ad4b04bd.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 299, + 348, + 361, + 413 + ], + "blocks": [ + { + "bbox": [ + 299, + 348, + 361, + 413 + ], + "lines": [ + { + "bbox": [ + 299, + 348, + 361, + 413 + ], + "spans": [ + { + "bbox": [ + 299, + 348, + 361, + 413 + ], + "type": "image", + "image_path": "054c15f3b8ca7fd16fe1809275a1d8dcdfccbebd0c8bc0165c7ae1f9f5fbbf3e.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 350, + 395, + 369 + ], + "lines": [ + { + "bbox": [ + 362, + 350, + 395, + 369 + ], + "spans": [ + { + "bbox": [ + 362, + 350, + 395, + 369 + ], + "type": "text", + "content": "luggage carrier" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 422, + 531, + 466 + ], + "lines": [ + { + "bbox": [ + 296, + 422, + 531, + 466 + ], + "spans": [ + { + "bbox": [ + 296, + 422, + 531, + 466 + ], + "type": "text", + "content": "Figure 1. The left panel shows the example activation maps of ResNet50 (1st row) and MSINet (2nd row). The right panel shows the average distances between the most similar 10 negative samples and each query image at the inference. Best viewed in color." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 397, + 350, + 458, + 413 + ], + "blocks": [ + { + "bbox": [ + 461, + 398, + 530, + 411 + ], + "lines": [ + { + "bbox": [ + 461, + 398, + 530, + 411 + ], + "spans": [ + { + "bbox": [ + 461, + 398, + 530, + 411 + ], + "type": "text", + "content": "Distance between top-10 neg. samples and each query image" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 350, + 458, + 413 + ], + "lines": [ + { + "bbox": [ + 397, + 350, + 458, + 413 + ], + "spans": [ + { + "bbox": [ + 397, + 350, + 458, + 413 + ], + "type": "image", + "image_path": "87ea7dbb407db1dfdbec3588fa670b59bf88a178a381deb80799c4a7e2522fad.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 296, + 496, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 496, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 496, + 530, + 730 + ], + "type": "text", + "content": "ment architectures on ReID leads to large performance variations. Some works employ Neural Architecture Search (NAS) for ReID [28, 45]. The proposed optimizing targets and search spaces stably improve the model performance, yet the main search scheme still follows traditional NAS methods designed for general classification tasks [12, 36]. As an open-set task, ReID contains different categories in the training and validation sets [64, 71], while the two sets share exactly the same categories in standard classification tasks [10], which is also followed by traditional NAS methods. The incompatibility between search schemes and real-world training schemes makes the searched architecture sub-optimal for ReID. Moreover, ReID is required to distinguish more subtle distinctions among fine-grained instances compared with image-level classification [48, 63]. Some previous works [4, 44, 68, 75] have manifested that local perspectives and multi-scale features are discriminative for ReID. However, current utilizations of these features are mostly empirically designed, which can be more flexible according to the characteristics of different network layers." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 719, + 143, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 719, + 143, + 729 + ], + "spans": [ + { + "bbox": [ + 57, + 719, + 143, + 729 + ], + "type": "text", + "content": "*Co-corresponding authors" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "19243" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 326 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 326 + ], + "type": "text", + "content": "In this work, we propose a novel NAS scheme aiming at addressing the aforementioned challenges. In order to simulate the real-world ReID training schemes, a Twins Contrastive Mechanism (TCM) is proposed to unbind the categories of the training and validation sets. An adjustable overlap ratio of categories builds up the compatibility between NAS and ReID, which provides more appropriate supervision for ReID architecture search. Moreover, to search for more rational utilizations of multi-scale features, we design a Multi-Scale Interaction (MSI) search space. The MSI space focuses on interaction operations between multi-scale features along the shallow and deep layers of the network, which guides the features to promote each other. Additionally, to further improve the generalization capability, we propose a Spatial Alignment Module (SAM) to enhance the attention consistency of the model confronted with images from different sources. With the above NAS scheme, we obtain a light-weight yet effective model architecture, denoted as Multi-Scale Interaction Net (MSINet)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 327, + 281, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 327, + 281, + 456 + ], + "spans": [ + { + "bbox": [ + 45, + 327, + 281, + 456 + ], + "type": "text", + "content": "We visualize the example activation maps of our proposed MSINet and ResNet50 [20] trained on VeRi-776 [38, 39] in Fig. 1. Compared to ResNet50, MSINet focuses on more unique distinctions with specific semantic information to recognize instances. Besides, MSINet largely increases the distance margin between query image and corresponding negative samples, reflecting extraordinary discriminative capability. Extensive experiments demonstrate that MSINet surpasses state-of-the-art (SOTA) ReID methods on both in-domain and cross-domain scenarios. Our source codes are available in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 456, + 238, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 238, + 466 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 238, + 466 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 474, + 279, + 641 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 57, + 474, + 279, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 474, + 279, + 522 + ], + "spans": [ + { + "bbox": [ + 57, + 474, + 279, + 522 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to build the NAS search scheme according to the real-world ReID training schemes, which provides more appropriate supervision for the ReID architecture search." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 529, + 279, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 529, + 279, + 575 + ], + "spans": [ + { + "bbox": [ + 57, + 529, + 279, + 575 + ], + "type": "text", + "content": "- We propose a novel search space based on the Multi-Scale Interaction (MSI) operations and a Spatial Alignment Module (SAM) to improve the model performance on in-domain and cross-domain scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "spans": [ + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "type": "text", + "content": "- We construct a light-weight yet effective architecture for ReID tasks, denoted as MSINet. With only 2.3M parameters, MSINet surpasses ResNet50 [20] by " + }, + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "type": "text", + "content": " mAP on MSMT17 [60] and " + }, + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "type": "text", + "content": " mAP on MSMT17 " + }, + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 57, + 583, + 279, + 641 + ], + "type": "text", + "content": " Market-1501 [69]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 651, + 135, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 135, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 135, + 664 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": "Neural Architecture Search. NAS has been increasingly appealing to the computer vision society, due to its automatic architecture designing characteristics. NAS methods can be roughly separated into four categories: reinforcement learning [1, 77], evolutionary algorithms [35, 46], gra" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 106, + 530, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 256 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 256 + ], + "type": "text", + "content": "dient descent [36, 43] and performance prediction [11, 34]. Liu et al. establish a differentiable architecture search (DARTS) method [36], which improves the practicability of NAS by a large extent. Some later works further improve the structure through sampling strategy [62], network pruning [2, 8], progressive learning [5], collaborative competition [7], etc. Most of NAS works focus on general image classification tasks, where the training and validation sets share the exact same categories. Following the setting, however, leads to incompatibility with the real-world training schemes of object ReID. In this work, we unbind the category bond between the two sets and propose a novel search scheme suitable for ReID." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 257, + 530, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 257, + 530, + 327 + ], + "spans": [ + { + "bbox": [ + 296, + 257, + 530, + 327 + ], + "type": "text", + "content": "ReID Network Design. Current ReID works mostly adopt backbones designed for image classification [20, 23, 47, 52]. Some works [14, 31, 67] design attention modules based on the common backbones to unearth their potential on distinguishing local distinctions. However, these methods usually lead to large calculation consumption." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 327, + 531, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 327, + 531, + 548 + ], + "spans": [ + { + "bbox": [ + 296, + 327, + 531, + 548 + ], + "type": "text", + "content": "There are also several works focusing on designing ReID-specific architectures. Li et al. present a Filter Pairing Neural Network to dynamically match patches in the feature maps [30]. Wang et al. separate and regroup the features of two samples with a WConv layer [56]. Guo et al. extract multi-scale features to directly evaluate the similarity between samples [18]. However, the siamese structure is inconvenient when conducting retrieval on large galleries. Zhou et al. aggregate multi-scale information to achieve high accuracy with small computing consumption [75]. Quan et al. introduce a part-aware module into the DARTS search space [36, 45]. Li et al. propose a new search space in regard to receptive field scales [28]. These methods have excellent performance on limited parameter scales, but fail to surpass those networks with complex structures. Different from previous works, we design a light-weight searching structure focusing on rational interaction operations between multi-scale features. The searched MSINet surpasses SOTA methods on both in-domain and cross-domain tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 557, + 356, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 557, + 356, + 570 + ], + "spans": [ + { + "bbox": [ + 297, + 557, + 356, + 570 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "spans": [ + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "text", + "content": "Our goal is to construct an effective NAS scheme to search for a light-weight backbone architecture suitable for ReID tasks. Based on the training schemes of ReID, we propose a novel Twins Contrastive Mechanism to provide more appropriate supervision for the search process. Aiming at rational interaction between multi-scale features, we design a Multi-Scale Interaction search space. We further introduce a Spatial Alignment Module to improve the generalization capability with limited parameter growth." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 688, + 458, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 688, + 458, + 700 + ], + "spans": [ + { + "bbox": [ + 297, + 688, + 458, + 700 + ], + "type": "text", + "content": "3.1. Twins Contrastive Mechanism" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": "NAS aims at automatically searching for the optimal network architecture for certain data. Inspired by [36], a basic" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19244" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 104, + 517, + 241 + ], + "blocks": [ + { + "bbox": [ + 61, + 104, + 517, + 241 + ], + "lines": [ + { + "bbox": [ + 61, + 104, + 517, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 104, + 517, + 241 + ], + "type": "image", + "image_path": "93629ce15e822e4f454f33347d05317c4ccacf32239c4a665cee6a5fbe229fd9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 250, + 531, + 285 + ], + "lines": [ + { + "bbox": [ + 45, + 250, + 531, + 285 + ], + "spans": [ + { + "bbox": [ + 45, + 250, + 531, + 285 + ], + "type": "text", + "content": "Figure 2. The model structure of the proposed MSINet. The input can be either person or vehicle. Inside a cell, the input is separated to two branches, with different receptive field scales. The interaction module exchanges information between two branches. Architecture search automatically selects the most appropriate interaction for each cell." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "spans": [ + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "text", + "content": "differentiable architecture search scheme is established. We define the ordinary model parameters as " + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "text", + "content": ", and architecture parameters as " + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "text", + "content": ". For network layer " + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "text", + "content": " with a search space of " + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "text", + "content": " controls the weight of each operation " + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 45, + 303, + 281, + 386 + ], + "type": "text", + "content": " in the space. The features are parallelly passed through all the operations, and the final output is formulated by the softmax-weighted sum of operation outputs:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 395, + 279, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 395, + 279, + 429 + ], + "spans": [ + { + "bbox": [ + 80, + 395, + 279, + 429 + ], + "type": "interline_equation", + "content": "f \\left(\\mathbf {x} _ {i}\\right) = \\sum_ {o \\in \\mathcal {O}} \\frac {\\exp \\left\\{\\alpha_ {i} ^ {o} \\right\\}}{\\sum_ {o ^ {\\prime} \\in \\mathcal {O}} \\exp \\left\\{\\alpha_ {i} ^ {o ^ {\\prime}} \\right\\}} \\cdot o \\left(\\mathbf {x} _ {i}\\right). \\tag {1}", + "image_path": "8fe7eb50fe3543b82e4a7aa5b070e1c368cace7f1ca1a8f1ab5345c57cc41d9c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 437, + 279, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 437, + 279, + 519 + ], + "spans": [ + { + "bbox": [ + 45, + 437, + 279, + 519 + ], + "type": "text", + "content": "The search process is conducted in an alternative manner. Training data is utilized to update the model parameters, and validation data is then employed to update the architecture parameters. For most NAS methods designed for image classification tasks, the training and validation data share exactly the same categories and a linear classification layer for loss calculation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "spans": [ + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "text", + "content": "Different from standard image classification, as an open-set retrieval task, ReID has different categories in the training and validation sets. The incompatibility between search schemes and real-world training schemes might lead to suboptimal searching results. Accordingly, we propose a novel Twins Contrastive Mechanism (TCM) for NAS ReID training. Specifically, we employ two independent auxiliary memories " + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{tr}" + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{val}" + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "text", + "content": " to store the embedded features of the training and validation data, respectively. The memories are initialized with the centroid features, which are calculated by averaging the features of each category. At each iteration, the training loss is first calculated with " + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{tr}" + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "text", + "content": " for model parameter updating. Given an embedded feature " + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "text", + "content": " with category label " + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 45, + 520, + 280, + 694 + ], + "type": "text", + "content": ", the contrastive classification loss is calculated with:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 703, + 279, + 734 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 703, + 279, + 734 + ], + "spans": [ + { + "bbox": [ + 89, + 703, + 279, + 734 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t r} ^ {c l s} = - \\log \\frac {\\exp (\\mathbf {f} \\cdot \\mathbf {c} _ {t r} ^ {j} / \\tau)}{\\sum_ {n = 0} ^ {N _ {t r} ^ {c}} \\exp (\\mathbf {f} \\cdot \\mathbf {c} _ {t r} ^ {n} / \\tau)}, \\tag {2}", + "image_path": "6818a1d7d7e633edda0eaca6efdc7dc38344642108fa2ab51b088cdbffaaf3d6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "spans": [ + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{tr}^n" + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "content": " represents the memorized feature of category " + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "inline_equation", + "content": "N_{tr}^c" + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "content": " stands for the total number of categories in the training set, and " + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "content": " is the temperature parameter, which is set as 0.05 empirically [17]. After updating the model parameters, the embedded feature " + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "content": " with category label " + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "content": " is integrated into the corresponding memorized feature " + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{tr}^j" + }, + { + "bbox": [ + 296, + 303, + 530, + 375 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 364, + 381, + 530, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 381, + 530, + 395 + ], + "spans": [ + { + "bbox": [ + 364, + 381, + 530, + 395 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {t r} ^ {j} \\leftarrow \\beta \\mathbf {c} _ {t r} ^ {j} + (1 - \\beta) \\mathbf {f}, \\tag {3}", + "image_path": "c309c748e799e5778d8bc76f2a9c4d946103b4d1cb949a1cd3b7d3bd4a857857.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "spans": [ + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "type": "text", + "content": " is set as 0.2 empirically [17]. Then the updated model is evaluated on the validation data to generate to validation loss with " + }, + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "type": "inline_equation", + "content": "C_{val}" + }, + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "type": "text", + "content": " replacing " + }, + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "type": "inline_equation", + "content": "C_{tr}" + }, + { + "bbox": [ + 296, + 402, + 530, + 459 + ], + "type": "text", + "content": " in Eq. 2. The architecture parameter is then updated with the validation loss to finish an iteration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 461, + 531, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 461, + 531, + 602 + ], + "spans": [ + { + "bbox": [ + 296, + 461, + 531, + 602 + ], + "type": "text", + "content": "As the loss calculation does not rely on the linear classification layer, the categories of the training and validation sets are unbound. We are able to dynamically adjust the category overlap ratio in these two sets. The advantages of a proper overlap ratio are summarized as two folds. Firstly, TCM better simulates the real-world training of ReID and helps the model focus on truly discriminative distinctions. The differences between the training and validation data improves the generalization capability of the model. Secondly, a relatively small proportion of overlapped categories stabilizes the architecture parameter update through a consistent optimizing target with the model parameter update." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 607, + 455, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 607, + 455, + 620 + ], + "spans": [ + { + "bbox": [ + 297, + 607, + 455, + 620 + ], + "type": "text", + "content": "3.2. Multi-Scale Interaction Space" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 624, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 624, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 624, + 530, + 730 + ], + "type": "text", + "content": "Although the local perspective and multi-scale features have already been investigated in previous ReID works [4, 28, 44, 51, 68, 72, 75, 76], the utilization of these information is mainly empirically designed aggregation, which is monotonous and restrained. We argue that on the one hand, the rational utilization of multi-scale features should be dynamically adjusted along the shallow and deep layers of the network. On the other hand, introducing interaction other than aggregation creates direct information exchange, and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "19245" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 104, + 517, + 236 + ], + "blocks": [ + { + "bbox": [ + 60, + 104, + 517, + 236 + ], + "lines": [ + { + "bbox": [ + 60, + 104, + 517, + 236 + ], + "spans": [ + { + "bbox": [ + 60, + 104, + 517, + 236 + ], + "type": "image", + "image_path": "5523ff0813347b36722b461219b27c29c12088ba4bdffab88dbf652e82b6f9aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 245, + 531, + 278 + ], + "lines": [ + { + "bbox": [ + 45, + 245, + 531, + 278 + ], + "spans": [ + { + "bbox": [ + 45, + 245, + 531, + 278 + ], + "type": "text", + "content": "Figure 3. The aligning pipeline of the proposed Spatial Alignment Module. The correlation activation vectors are calculated between the anchor feature and all features in a mini-batch. The positive vectors are aligned with the learnable self-activation, and the negative vectors are aligned with each other. The structure of PAM is shown on the left." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 337, + 278, + 378 + ], + "blocks": [ + { + "bbox": [ + 45, + 296, + 279, + 328 + ], + "lines": [ + { + "bbox": [ + 45, + 296, + 279, + 328 + ], + "spans": [ + { + "bbox": [ + 45, + 296, + 279, + 328 + ], + "type": "text", + "content": "Table 1. The detailed interaction operation in the proposed MSINet architecture. N: None; E: Exchange; G: Channel Gate; A: Cross Attention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 337, + 278, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 337, + 278, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 278, + 378 + ], + "type": "table", + "html": "
Cell #1Cell #2Cell #3Cell #4Cell #5Cell #6
123456789101112
GGEGAGGNGAEA
", + "image_path": "09534fbfebbdde93fedc1e8704b85ddf408432c67792b2bf91da8410fb178579.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 399, + 279, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 399, + 279, + 434 + ], + "spans": [ + { + "bbox": [ + 45, + 399, + 279, + 434 + ], + "type": "text", + "content": "makes fuller use of multi-scale features. Therefore, we propose a novel Multi-Scale Interaction (MSI) search space to establish a light-weight architecture suitable for ReID." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "spans": [ + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "text", + "content": "As shown in Fig. 2, the network is mainly grouped with MSI cells and down-sample blocks, which is generally consistent with OSNet [75]. In each cell, the input features are passed through two branches with different receptive field scales. To reduce the calculation burden of the network, for the layers inside each branch, we adopt the stack of " + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "text", + "content": " convolution and multiple depth-wise " + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "text", + "content": " convolution to implement specific scales. A scale ratio " + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "text", + "content": " of 3:1 is selected for the two branches. These two branches do not share model parameters, except for the Interaction Modules (IM). IM introduces information exchange for the two branches. There are 4 operation options for the IM. With the two-branch input features defined as " + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_1, \\mathbf{x}_2)" + }, + { + "bbox": [ + 45, + 435, + 279, + 597 + ], + "type": "text", + "content": ", the operations can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 599, + 279, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 599, + 279, + 623 + ], + "spans": [ + { + "bbox": [ + 45, + 599, + 279, + 623 + ], + "type": "text", + "content": "None. None operation involves no parameters, and outputs exactly the input features " + }, + { + "bbox": [ + 45, + 599, + 279, + 623 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_1,\\mathbf{x}_2)" + }, + { + "bbox": [ + 45, + 599, + 279, + 623 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 623, + 279, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 623, + 279, + 670 + ], + "spans": [ + { + "bbox": [ + 45, + 623, + 279, + 670 + ], + "type": "text", + "content": "Exchange. Exchange acts as the strongest interaction among all options. It directly exchanges the features for the two branches and outputs " + }, + { + "bbox": [ + 45, + 623, + 279, + 670 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_2, \\mathbf{x}_1)" + }, + { + "bbox": [ + 45, + 623, + 279, + 670 + ], + "type": "text", + "content": ". Exchange contains no extra parameters, as well." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 671, + 279, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 706 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 706 + ], + "type": "text", + "content": "Channel Gate. Channel gate introduces a Multi-Layer Perceptron (MLP) to generate a channel-wise attention gate [61, 75] as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 718, + 279, + 730 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 718, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 114, + 718, + 279, + 730 + ], + "type": "interline_equation", + "content": "G (\\mathbf {x}) = \\sigma (M L P (\\mathbf {x})), \\tag {4}", + "image_path": "d86f35816f3c42c874ea2d5f0feb41347f00700850896fdac81fff059f9c2d10.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 301, + 295, + 539, + 374 + ], + "blocks": [ + { + "bbox": [ + 301, + 295, + 539, + 374 + ], + "lines": [ + { + "bbox": [ + 301, + 295, + 539, + 374 + ], + "spans": [ + { + "bbox": [ + 301, + 295, + 539, + 374 + ], + "type": "image", + "image_path": "0d19bcb32e290dde3c7cf4dc867064efe3f68de7668d2d5791e3709501adffe4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 383, + 523, + 394 + ], + "lines": [ + { + "bbox": [ + 303, + 383, + 523, + 394 + ], + "spans": [ + { + "bbox": [ + 303, + 383, + 523, + 394 + ], + "type": "text", + "content": "Figure 4. Output feature maps of MSI cells at different layers." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 414, + 530, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 414, + 530, + 460 + ], + "spans": [ + { + "bbox": [ + 296, + 414, + 530, + 460 + ], + "type": "text", + "content": "and returns " + }, + { + "bbox": [ + 296, + 414, + 530, + 460 + ], + "type": "inline_equation", + "content": "(G(\\mathbf{x}_1) \\cdot \\mathbf{x}_1, G(\\mathbf{x}_2) \\cdot \\mathbf{x}_2)" + }, + { + "bbox": [ + 296, + 414, + 530, + 460 + ], + "type": "text", + "content": ". The MLP is composed of 2 fully connected layers and its parameters are shared for both branches. Thereby it achieves interaction by jointly screening discriminative feature channels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "spans": [ + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "text", + "content": "Cross Attention. Traditional channel attention module calculates the channel correlation inside a single feature map [15]. The original feature map " + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "text", + "content": " is firstly reshaped into the query feature " + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} \\in \\mathcal{R}^{C \\times N}" + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "inline_equation", + "content": "N = H \\times W" + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "text", + "content": ". Then the correlation activation is calculated by performing a matrix multiplication between the query feature " + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "text", + "content": " and the key feature " + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}^{\\top}" + }, + { + "bbox": [ + 296, + 461, + 530, + 589 + ], + "type": "text", + "content": ". We propose to exchange the keys of the two branches to explicitly calculate the correlation between each other. The correlation activation is then transformed to a mask, and is added up to the original features with a learnable proportion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "spans": [ + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "text", + "content": "After interaction, the multi-scale branches are fused through a sum operation. It is worth noting that the extra parameters brought by multiple interaction options are limited, which enables searching for each cell along the whole network independently. At the beginning of the network, we employ the same stem module as that in OSNet [75], containing a " + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "text", + "content": " convolutional layer and a " + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "text", + "content": " max pooling with a stride of 2. After the searching process, the interaction operation " + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "text", + "content": " with the largest weight " + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{o}" + }, + { + "bbox": [ + 296, + 589, + 530, + 705 + ], + "type": "text", + "content": " at each layer is reserved to form the searched architecture." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "type": "text", + "content": "After searching the architecture, the model is validated on various Re-ID tasks. The training is constrained by the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "19246" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 156, + 521, + 293 + ], + "blocks": [ + { + "bbox": [ + 45, + 104, + 532, + 149 + ], + "lines": [ + { + "bbox": [ + 45, + 104, + 532, + 149 + ], + "spans": [ + { + "bbox": [ + 45, + 104, + 532, + 149 + ], + "type": "text", + "content": "Table 2. Supervised performance on object ReID datasets. The results in the top part are trained from scratch, and those in the bottom part are pre-trained on ImageNet in advance. As the compared methods are originally proposed for person ReID, we reproduce the results in vehicle datasets. * indicates that the results of person ReID are reproduced by us. The evaluation results of architecture searched on VR can be found in the supplementary material." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 156, + 521, + 293 + ], + "lines": [ + { + "bbox": [ + 57, + 156, + 521, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 156, + 521, + 293 + ], + "type": "table", + "html": "
MethodParamsInference TimeMMSVRVIDMS→MVR→VID
R-1↑mAP↑R-1↑mAP↑R-1↑mAP↑R-1↑R-5↑R-1↑mAP↑R-1↑R-5↑
ResNet50* [42]~24M1×85.768.348.025.792.869.970.676.6--
OSNet [75]2.2M0.79×93.681.071.043.395.472.876.088.7----
CDNet [28]1.8M0.67×93.783.773.748.594.373.074.588.8----
MSINet2.3M0.71×94.687.076.052.595.975.076.589.8----
ResNet50* [42]~24M1×94.585.975.550.494.573.676.589.958.831.842.861.9
OSNet [75]2.2M0.79×94.884.978.752.995.576.476.088.666.637.546.563.1
CDNet [28]1.8M0.67×95.186.078.954.7--------
MSINet2.3M0.71×95.389.681.059.696.878.877.991.774.946.248.065.6
MSINet-SAM2.4M0.71×95.589.980.759.596.779.078.091.976.348.449.066.8
", + "image_path": "1783de226bd1396dd2f95ee7a8c219fa06e0b94a2d000fac1d24cc9b112ec0fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 320, + 263, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 320, + 263, + 333 + ], + "spans": [ + { + "bbox": [ + 45, + 320, + 263, + 333 + ], + "type": "text", + "content": "classification id loss and the triplet loss, formulated by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 341, + 279, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 341, + 279, + 374 + ], + "spans": [ + { + "bbox": [ + 82, + 341, + 279, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i d} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - \\log \\left(\\frac {\\exp \\mathbf {W} _ {i} ^ {\\top} \\mathbf {f} _ {i}}{\\sum_ {j} \\exp \\mathbf {W} _ {j} ^ {\\top} \\mathbf {f} _ {i}}\\right), \\tag {5}", + "image_path": "00debd0e70861856953ae4658f308390a1dfac4f69535ac4319563a1caea162d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 380, + 279, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 380, + 279, + 405 + ], + "spans": [ + { + "bbox": [ + 45, + 380, + 279, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 380, + 279, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 45, + 380, + 279, + 405 + ], + "type": "text", + "content": " is a feature vector, the corresponding classifier weight of which is " + }, + { + "bbox": [ + 45, + 380, + 279, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i" + }, + { + "bbox": [ + 45, + 380, + 279, + 405 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 413, + 279, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 413, + 279, + 428 + ], + "spans": [ + { + "bbox": [ + 86, + 413, + 279, + 428 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t r i} = \\left[ \\mathcal {D} \\left(\\mathbf {f} _ {a}, \\mathbf {f} _ {p}\\right) - \\mathcal {D} \\left(\\mathbf {f} _ {a}, \\mathbf {f} _ {n}\\right) + m \\right] _ {+}, \\tag {6}", + "image_path": "baeee56c7f808ab88c12968d99111d079d7aa1919f39c420200df9f25b056fb1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "spans": [ + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_a, \\mathbf{f}_p, \\mathbf{f}_n" + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "text", + "content": " are the embedded features for the anchor, the hardest positive and negative samples in a mini-batch, " + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "text", + "content": " is the Euclidean distance, " + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "text", + "content": " is the margin parameter, and " + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "inline_equation", + "content": "[\\cdot]_{+}" + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "inline_equation", + "content": "\\max(\\cdot, 0)" + }, + { + "bbox": [ + 45, + 435, + 279, + 484 + ], + "type": "text", + "content": " function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 489, + 188, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 188, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 188, + 501 + ], + "type": "text", + "content": "3.3. Spatial Alignment Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 507, + 279, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 507, + 279, + 588 + ], + "spans": [ + { + "bbox": [ + 45, + 507, + 279, + 588 + ], + "type": "text", + "content": "The retrieval precision of object ReID tasks are largely affected by the variation of appearances such as poses, illumination and occlusion when the camera conditions change. In order that the model correctly and consistently focuses on the discriminative spatial positions, we design a Spatial Alignment Module (SAM) to explicitly align the spatial attention between images, as shown in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "spans": [ + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "content": "Specifically, we first calculate the position-wise correlation activation map " + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "content": " between the feature maps in a minibatch. The activation between sample " + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "content": " can be formulated as: " + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(i,j) = \\tilde{\\mathbf{x}}_j^\\top \\times \\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} \\in \\mathcal{R}^{C \\times N}" + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "content": " is reshaped from the original feature " + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "content": ". Then we take the maximum activation for each position of sample " + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 589, + 279, + 660 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 669, + 279, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 669, + 279, + 687 + ], + "spans": [ + { + "bbox": [ + 113, + 669, + 279, + 687 + ], + "type": "interline_equation", + "content": "\\mathbf {a} (i, j) = \\max _ {d i m = 1} \\mathbf {A} (i, j). \\tag {7}", + "image_path": "45334e92f57a9ed0fca442122223bb8e53bdecd592d4e197f265d7255a6a039e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": "The above process is denoted as \"Mutual Conv\" in Fig. 3. We evaluate the consistency between activation vectors with cosine similarity. For negative samples specifically, there" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 320, + 530, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 320, + 530, + 378 + ], + "spans": [ + { + "bbox": [ + 296, + 320, + 530, + 378 + ], + "type": "text", + "content": "can be many different hints for recognition, some of which might be inappropriate, such as the backgrounds. By aligning all the correlations for sample " + }, + { + "bbox": [ + 296, + 320, + 530, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 296, + 320, + 530, + 378 + ], + "type": "text", + "content": ", we hope that the network can correct some attention bias and consistently focus on discriminative positions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 380, + 531, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 380, + 531, + 450 + ], + "spans": [ + { + "bbox": [ + 296, + 380, + 531, + 450 + ], + "type": "text", + "content": "However, through aligning positive sample pairs, the ID-related features are expected to be emphasized, which cannot be achieved by aligning negative pairs. Therefore, we introduce an extra position activation module (PAM) to generate supervision for the alignment between positive pairs. The spatial alignment loss is formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 461, + 501, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 461, + 501, + 522 + ], + "spans": [ + { + "bbox": [ + 326, + 461, + 501, + 522 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {s a} (i) = \\frac {1}{N _ {+}} \\sum_ {p \\in \\mathcal {I} _ {+}} \\left(1 - S (\\hat {\\mathbf {a}} (i), \\mathbf {a} (i, p))\\right) + \\\\ \\frac {1}{N _ {-}} \\sum_ {n _ {1}, n _ {2} \\in \\mathcal {I} _ {-}} (1 - S (\\mathbf {a} (i, n _ {1}), \\mathbf {a} (i, n _ {2}))), \\\\ \\end{array}", + "image_path": "e6fc436fa0607c14687fb9a2682161ad933b252887e774b74b95597046629c61.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 516, + 487, + 530, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 516, + 487, + 530, + 497 + ], + "spans": [ + { + "bbox": [ + 516, + 487, + 530, + 497 + ], + "type": "text", + "content": "(8)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "spans": [ + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{+}" + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "text", + "content": " contains positive indices for sample " + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "text", + "content": ", the total number of which is " + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "inline_equation", + "content": "N_{+}" + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "text", + "content": ", and vice versa. " + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}(i)" + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "text", + "content": " stands for the generated activation vector for positive sample alignment, and " + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "inline_equation", + "content": "S(\\cdot, \\cdot)" + }, + { + "bbox": [ + 296, + 533, + 531, + 580 + ], + "type": "text", + "content": " is the cosine similarity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 596, + 377, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 596, + 377, + 609 + ], + "spans": [ + { + "bbox": [ + 297, + 596, + 377, + 609 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 617, + 468, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 617, + 468, + 628 + ], + "spans": [ + { + "bbox": [ + 297, + 617, + 468, + 628 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "content": "Our proposed method is evaluated on two person ReID datasets Market-1501 [69], MSMT17 [60], and two vehicle ReID datasets VeRi-776 [38, 39] and VehicleID [37]. For simplicity, the four datasets are denoted as M, MS, VR and VID in the following sections, respectively. Evaluation metrics include Cumulative Matching Characteristic (CMC) and mean average precision (mAP), which are commonly utilized on ReID tasks." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "19247" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 134, + 253, + 308 + ], + "blocks": [ + { + "bbox": [ + 46, + 104, + 279, + 126 + ], + "lines": [ + { + "bbox": [ + 46, + 104, + 279, + 126 + ], + "spans": [ + { + "bbox": [ + 46, + 104, + 279, + 126 + ], + "type": "text", + "content": "Table 3. Supervised performance comparison between MSINet and SOTA methods on M and MS datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 134, + 253, + 308 + ], + "lines": [ + { + "bbox": [ + 73, + 134, + 253, + 308 + ], + "spans": [ + { + "bbox": [ + 73, + 134, + 253, + 308 + ], + "type": "table", + "html": "
MethodMMS
R-1↑mAP↑R-1↑mAP↑
PCB [51]93.881.668.240.4
MGN [55]95.786.976.952.1
OSNet [75]93.681.071.043.3
IANet [22]94.483.175.546.8
DGNet [73]94.886.077.252.3
Auto-ReID [45]94.585.1--
SAN [26]96.188.079.255.7
CDNet [28]95.186.078.954.7
BAT-Net [14]95.187.479.556.8
SFT [41]94.187.579.058.3
CTF [66]94.887.7--
RGA-SC [67]96.188.480.357.5
MSINet95.389.681.059.6
", + "image_path": "e3eeeaf9c314078637d834686c26ffdf1a3fa8e69ce511c4fe7d50a6d4996ce5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 326, + 158, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 158, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 158, + 337 + ], + "type": "text", + "content": "4.2. Architecture Search" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 344, + 278, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 344, + 278, + 495 + ], + "spans": [ + { + "bbox": [ + 45, + 344, + 278, + 495 + ], + "type": "text", + "content": "We conduct the searching process on MSMT17. SGD is adopted for model parameter update with an initial learning rate of 0.025. The model is trained for 350 epochs in total. We adopt a warm-up strategy for the first 10 epochs. Then the learning rate is decayed by 0.1 at 150, 225 and 300 epochs, respectively. Adam [27] is adopted for the architecture parameter update with an initial learning rate of 0.002. The learning rate is decayed at the same pace. The images are reshaped to " + }, + { + "bbox": [ + 45, + 344, + 278, + 495 + ], + "type": "inline_equation", + "content": "256 \\times 128" + }, + { + "bbox": [ + 45, + 344, + 278, + 495 + ], + "type": "text", + "content": " for person and " + }, + { + "bbox": [ + 45, + 344, + 278, + 495 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 45, + 344, + 278, + 495 + ], + "type": "text", + "content": " for vehicles. Data augmentation includes random flip, random crop and random erasing [74]. The searched architecture is presented in Tab. 1. The \"MSINet\" in the following experiment sections refers to this architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 495, + 278, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 495, + 278, + 600 + ], + "spans": [ + { + "bbox": [ + 45, + 495, + 278, + 600 + ], + "type": "text", + "content": "We visualize the feature maps extracted by each MSI cell in Fig. 4. At the shallow layers of the network, the kernels mainly focus on overall contour information. Channel gate helps to filter out inferior information, such as the background. As we approach deeper layers, the extracted features each have specific semantic information, where cross attention is more likely to be selected for the interaction. It indicates that cross attention is more rational for exchanging high-level semantic information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 607, + 227, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 227, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 227, + 618 + ], + "type": "text", + "content": "4.3. Comparison with Other Backbones" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 625, + 278, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 278, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 278, + 658 + ], + "type": "text", + "content": "We first compare our proposed MSINet with ResNet50 and recent proposed light-weight backbones in both in-domain and cross-domain ReID scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 660, + 278, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 660, + 278, + 729 + ], + "spans": [ + { + "bbox": [ + 46, + 660, + 278, + 729 + ], + "type": "text", + "content": "In-Domain Tasks. We adopt a two-group supervised evaluation scheme similar to that in [28, 75]: training from scratch and fine-tuning ImageNet [10] pre-trained models. The training parameters for both schemes are kept the same as that in architecture search, except for an initial learning rate of 0.065. Triplet loss and cross entropy loss are adopted" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 323, + 135, + 505, + 308 + ], + "blocks": [ + { + "bbox": [ + 297, + 104, + 530, + 126 + ], + "lines": [ + { + "bbox": [ + 297, + 104, + 530, + 126 + ], + "spans": [ + { + "bbox": [ + 297, + 104, + 530, + 126 + ], + "type": "text", + "content": "Table 4. Unsupervised performance applying MSINet to SOTA methods for USL on M and UDA on " + }, + { + "bbox": [ + 297, + 104, + 530, + 126 + ], + "type": "inline_equation", + "content": "\\mathrm{M} \\rightarrow \\mathrm{{MS}}" + }, + { + "bbox": [ + 297, + 104, + 530, + 126 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 135, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 323, + 135, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 323, + 135, + 505, + 308 + ], + "type": "table", + "html": "
MethodMM→MS
R-1↑mAP↑R-1↑mAP↑
MMCL [54]80.345.540.815.1
MMT [16]--50.124.0
JVTC+ [29]79.547.548.625.1
CycAs [59]84.864.8--
GCL [3]87.366.851.127.0
MPRD [24]83.051.0--
SpCL [17]88.173.153.726.8
GS [19]92.379.2--
GS+MSINet91.781.5--
HDCRL [6]92.481.7--
HDCRL+MSINet92.982.7--
IDM [9]--61.333.5
IDM+MSINet--66.037.8
", + "image_path": "4ab284cb5208d20a3833265493029730dec2eb4a10b14c5bfc0638a0e9e789e0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 330, + 530, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 330, + 530, + 363 + ], + "spans": [ + { + "bbox": [ + 297, + 330, + 530, + 363 + ], + "type": "text", + "content": "for the parameter update. The margin " + }, + { + "bbox": [ + 297, + 330, + 530, + 363 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 297, + 330, + 530, + 363 + ], + "type": "text", + "content": " in Eq. 6 is set as 0.3. [28] adopts an FBLNeck. We also employ the same structure. The results are shown in Tab. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 366, + 530, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 366, + 530, + 575 + ], + "spans": [ + { + "bbox": [ + 297, + 366, + 530, + 575 + ], + "type": "text", + "content": "ResNet50 is the most commonly utilized backbone network in ReID tasks, yet holds the worst performance. Moreover, ResNet50 largely depends on ImageNet pre-training, while MSINet without pre-training has already surpassed pre-trained ResNet50 on all metrics. Compared with the other datasets, MS contains more variations on illumination, background and camera pose, and brings a large performance gap between ResNet50 and other methods. It also validates the inadequacy of image classification networks on ReID tasks. OSNet [75] and CDNet [28] are recently proposed architectures designed specifically for ReID tasks. Both architectures focus on fusing multi-scale features to better suit ReID. CDNet employs a traditional NAS scheme to search for the proper receptive field scales for each cell. MSINet fixes the receptive field scale and instead selects optimal interaction operations inside each cell. With only a bit more parameters, MSINet surpasses all the other backbones by a large margin." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 577, + 530, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 577, + 530, + 670 + ], + "spans": [ + { + "bbox": [ + 297, + 577, + 530, + 670 + ], + "type": "text", + "content": "Cross-Domain Tasks. Cross-domain experiments verify the generalization capability of the model. Following previous domain generalizable ReID works [25, 33], data augmentation is adjusted to random flip, random crop and color jittering. The model is pre-trained and fine-tuned for 250 epochs to avoid over-fitting. The other settings are kept the same as supervision scenes. With no present pre-trained models for CDNet [28], it is excluded from this section." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 671, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 671, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 297, + 671, + 530, + 729 + ], + "type": "text", + "content": "Tab. 2 shows that ResNet50 can be easily interfered by different image styles confronted with new image domains. OSNet learns multi-scale features with specific semantic information for ReID, which is domain invariant to some extent. Our proposed search scheme also takes into ac" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "19248" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 106, + 145, + 215 + ], + "blocks": [ + { + "bbox": [ + 54, + 106, + 145, + 215 + ], + "lines": [ + { + "bbox": [ + 54, + 106, + 145, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 106, + 145, + 215 + ], + "type": "image", + "image_path": "51c2fbd7e148acffa94d21e9d5888f89d2f1d69179c45870fdab237151ac504b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 217, + 113, + 227 + ], + "lines": [ + { + "bbox": [ + 101, + 217, + 113, + 227 + ], + "spans": [ + { + "bbox": [ + 101, + 217, + 113, + 227 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 148, + 107, + 243, + 215 + ], + "blocks": [ + { + "bbox": [ + 148, + 107, + 243, + 215 + ], + "lines": [ + { + "bbox": [ + 148, + 107, + 243, + 215 + ], + "spans": [ + { + "bbox": [ + 148, + 107, + 243, + 215 + ], + "type": "image", + "image_path": "ee5a3bfa957f5cbf48be410b91cbae67f51a37e44d21cae0ad9d9e3e08e97fcd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 217, + 208, + 227 + ], + "lines": [ + { + "bbox": [ + 196, + 217, + 208, + 227 + ], + "spans": [ + { + "bbox": [ + 196, + 217, + 208, + 227 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 106, + 334, + 216 + ], + "blocks": [ + { + "bbox": [ + 246, + 106, + 334, + 216 + ], + "lines": [ + { + "bbox": [ + 246, + 106, + 334, + 216 + ], + "spans": [ + { + "bbox": [ + 246, + 106, + 334, + 216 + ], + "type": "image", + "image_path": "7fb17dd76fc40c0ad73e6f3fdeb75235051a3d3d4eaf379aedcc200ae10e1eb1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 217, + 303, + 227 + ], + "lines": [ + { + "bbox": [ + 291, + 217, + 303, + 227 + ], + "spans": [ + { + "bbox": [ + 291, + 217, + 303, + 227 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 337, + 106, + 424, + 217 + ], + "blocks": [ + { + "bbox": [ + 337, + 106, + 424, + 217 + ], + "lines": [ + { + "bbox": [ + 337, + 106, + 424, + 217 + ], + "spans": [ + { + "bbox": [ + 337, + 106, + 424, + 217 + ], + "type": "image", + "image_path": "84d3e0d04855a8721be3990a139f9aa43aa4625c7d735ac08847e1c2b9bc49d8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 217, + 389, + 227 + ], + "lines": [ + { + "bbox": [ + 377, + 217, + 389, + 227 + ], + "spans": [ + { + "bbox": [ + 377, + 217, + 389, + 227 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 235, + 530, + 247 + ], + "lines": [ + { + "bbox": [ + 45, + 235, + 530, + 247 + ], + "spans": [ + { + "bbox": [ + 45, + 235, + 530, + 247 + ], + "type": "text", + "content": "Figure 5. Ablation studies on (a) search scheme; (b) architecture; (c) multi-scale aggregation; (d) scale ratio " + }, + { + "bbox": [ + 45, + 235, + 530, + 247 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 45, + 235, + 530, + 247 + ], + "type": "text", + "content": "; (e) spatial alignment weight." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 426, + 106, + 522, + 217 + ], + "blocks": [ + { + "bbox": [ + 426, + 106, + 522, + 217 + ], + "lines": [ + { + "bbox": [ + 426, + 106, + 522, + 217 + ], + "spans": [ + { + "bbox": [ + 426, + 106, + 522, + 217 + ], + "type": "image", + "image_path": "480a979402b8bf2edbe704b3d93a387f6878636a7f76d1d2ab31c89b772c7e41.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 217, + 483, + 227 + ], + "lines": [ + { + "bbox": [ + 471, + 217, + 483, + 227 + ], + "spans": [ + { + "bbox": [ + 471, + 217, + 483, + 227 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 48, + 306, + 277, + 431 + ], + "blocks": [ + { + "bbox": [ + 45, + 266, + 280, + 298 + ], + "lines": [ + { + "bbox": [ + 45, + 266, + 280, + 298 + ], + "spans": [ + { + "bbox": [ + 45, + 266, + 280, + 298 + ], + "type": "text", + "content": "Table 5. The effectiveness of each components in SAM. Both in-domain performance on VR and cross-domain performance on " + }, + { + "bbox": [ + 45, + 266, + 280, + 298 + ], + "type": "inline_equation", + "content": "\\mathrm{VR}\\rightarrow \\mathrm{VID}" + }, + { + "bbox": [ + 45, + 266, + 280, + 298 + ], + "type": "text", + "content": " is evaluated." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 306, + 277, + 431 + ], + "lines": [ + { + "bbox": [ + 48, + 306, + 277, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 306, + 277, + 431 + ], + "type": "table", + "html": "
ModelPos.Neg.AlignVRVR→VID
R-1↑mAP↑R-1↑R-5↑
MSINet✓-96.778.548.065.6
Self96.778.148.466.0
✓Self96.778.448.566.2
✓✓Unified96.678.348.365.8
✓✓Separated96.878.648.766.4
✓✓PAM-Self96.779.049.066.8
OSNet✓✓None95.576.446.563.1
PAM-Self95.976.347.563.3
", + "image_path": "c5def09c6db2f2dd8ba720a3916a0266860d01e6ac321479a87edc347bff5814.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 452, + 279, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 452, + 279, + 523 + ], + "spans": [ + { + "bbox": [ + 45, + 452, + 279, + 523 + ], + "type": "text", + "content": "count the generalization capability of the model. By partly separating the categories for training and validation sets, the searched interaction operations generalize well confronted with new image domains. Except for discrimination, MSINet also surpasses the other backbones on cross-domain tasks by a large margin with faster inference speed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 524, + 279, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 524, + 279, + 607 + ], + "spans": [ + { + "bbox": [ + 45, + 524, + 279, + 607 + ], + "type": "text", + "content": "Additionally, we introduce SAM into the model, which aligns the spatial correlations between images. A weighted sum of ReID loss and spatial alignment loss is utilized when training the network with SAM. The weight of spatial alignment loss is set as " + }, + { + "bbox": [ + 45, + 524, + 279, + 607 + ], + "type": "inline_equation", + "content": "\\lambda_{sa} = 2.0" + }, + { + "bbox": [ + 45, + 524, + 279, + 607 + ], + "type": "text", + "content": ". Without extra inference consumption or damages on the supervised performance, SAM further boosts the generalization capability of MSINet." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 617, + 262, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 617, + 262, + 630 + ], + "spans": [ + { + "bbox": [ + 45, + 617, + 262, + 630 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": "Tab. 3 further illustrates the supervision performance comparison of our proposed MSINet with the SOTA methods on M and MS datasets. With much less parameters than most of the compared methods, MSINet achieves a retrieval accuracy comparable to that of more complicated ones. Auto-ReID [45] first designs a NAS scheme for ReID, yet the DARTS-style architecture contains 13M parameters. RGA-SC [67] carefully designs a relation-aware global at" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 267, + 530, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 267, + 530, + 302 + ], + "spans": [ + { + "bbox": [ + 296, + 267, + 530, + 302 + ], + "type": "text", + "content": "tention module. MSINet achieves even higher performance with less training consumption, which validates the superiority of selecting rational interaction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 303, + 531, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 303, + 531, + 431 + ], + "spans": [ + { + "bbox": [ + 296, + 303, + 531, + 431 + ], + "type": "text", + "content": "We also evaluate the model performance replacing the backbone network from ResNet50 to MSINet for SOTA unsupervised ReID methods in Tab. 4. For purely unsupervised learning (USL) method GS [19] on M dataset, MSINet performs slightly lower on rank-1, yet has a large superiority on mAP. For HDCRL [6], MSINet shows obvious superiority over ResNet50. For unsupervised domain adaptation (UDA) method IDM [9] on " + }, + { + "bbox": [ + 296, + 303, + 531, + 431 + ], + "type": "inline_equation", + "content": "\\mathrm{M}\\rightarrow \\mathrm{MS}" + }, + { + "bbox": [ + 296, + 303, + 531, + 431 + ], + "type": "text", + "content": " task, MSINet surpasses ResNet50 by a large margin, which further proves that the TCM brings outstanding generalization capability to the searched architecture." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 297, + 442, + 395, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 442, + 395, + 453 + ], + "spans": [ + { + "bbox": [ + 297, + 442, + 395, + 453 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 461, + 530, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 461, + 530, + 565 + ], + "spans": [ + { + "bbox": [ + 296, + 461, + 530, + 565 + ], + "type": "text", + "content": "Effectiveness of Architecture Search. To verify the effectiveness, we conduct supervised training on MS with different search schemes in Fig. 5 (a). Under the standard classification scheme (\"CE Overlap\"), the searched model performs poor. Replacing the cross entropy loss to contrastive loss (\"TCM Overlap\") only brings a slight improvement. The complete TCM framework unbounds the categories between the training and validation set, and thereby improves the performance by a large margin." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 566, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 566, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 566, + 531, + 730 + ], + "type": "text", + "content": "In Fig. 5 (b) we compare the performance of different architectures. Firstly, we validate 4 models each with a unique interaction operation from the 4 options in the search space. None and Exchange, with no trainable parameters, achieve poor performance. Channel gate introduces channel-wise attention, whose model performs the best among 4 options. Cross attention exchanges the key features for the two branches. Over-frequent exchange interferes the ordinary feature extraction and degrades the network performance. Through appropriately arranging interaction operations along the architecture, MSINet surpasses all the above 4 models. Random architecture, on the other hand, shows no rational appliance of interaction operations, which validates that the proposed search scheme helps find" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19249" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 134, + 279, + 213 + ], + "blocks": [ + { + "bbox": [ + 46, + 104, + 279, + 126 + ], + "lines": [ + { + "bbox": [ + 46, + 104, + 279, + 126 + ], + "spans": [ + { + "bbox": [ + 46, + 104, + 279, + 126 + ], + "type": "text", + "content": "Table 6. Supervised performance comparison between MSINet and Transformers on VR and MS datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 134, + 279, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 134, + 279, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 279, + 213 + ], + "type": "table", + "html": "
MethodParamsInference TimeMSVR
R-1↑mAP↑R-1↑mAP↑
DeiT-S [21]~22M0.97x76.355.295.576.3
DeiT-B [21]~86M1.79x81.961.495.978.4
ViT-B [21]~86M1.79x81.861.096.578.2
MSINet2.3M0.71x81.059.696.878.8
", + "image_path": "d4a46fc93d9d7ed3c58dc330b48071522826c61fd30a85ba9439a616033a4c84.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 222, + 276, + 341 + ], + "blocks": [ + { + "bbox": [ + 48, + 222, + 276, + 341 + ], + "lines": [ + { + "bbox": [ + 48, + 222, + 276, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 276, + 341 + ], + "type": "image", + "image_path": "8316f43b4fa82cca870622977851a032b0a78b4336f37cc468b944698b660efd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 350, + 279, + 394 + ], + "lines": [ + { + "bbox": [ + 45, + 350, + 279, + 394 + ], + "spans": [ + { + "bbox": [ + 45, + 350, + 279, + 394 + ], + "type": "text", + "content": "Figure 6. Example top-15 retrieved sequences comparison on VR. Appearance differences caused by variant camera conditions are well addressed by the proposed MSINet. Visualization of person ReID can be found in the supplementary material." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 414, + 170, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 414, + 170, + 425 + ], + "spans": [ + { + "bbox": [ + 45, + 414, + 170, + 425 + ], + "type": "text", + "content": "suitable architectures for ReID." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 426, + 279, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 426, + 279, + 623 + ], + "spans": [ + { + "bbox": [ + 45, + 426, + 279, + 623 + ], + "type": "text", + "content": "Effectiveness of Spatial Alignment Module. We validate the effectiveness of each components of SAM on the VR " + }, + { + "bbox": [ + 45, + 426, + 279, + 623 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 45, + 426, + 279, + 623 + ], + "type": "text", + "content": " VID cross-domain experiment in Tab. 5. Firstly, we introduce the spatial alignment for positive and negative sample pairs, respectively. Each of them brings certain performance improvements. However, a unified alignment for all sample pairs damages ID-related features and degrades the performance instead. Therefore, we separate the alignment of positive and negative samples, which retains some discriminative features and integrates the effect of both aspects. The extra PAM for positive sample alignment further guarantees the focus on ID-related positions and achieves the best performance. We also conduct in-domain experiment on VR to prove that SAM improves the generalization capability without sacrificing the supervision performance. Adding SAM to OSNet receives similar results, which validates the universality of SAM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 624, + 279, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 624, + 279, + 694 + ], + "spans": [ + { + "bbox": [ + 45, + 624, + 279, + 694 + ], + "type": "text", + "content": "Fusing Operation. After interaction, the multi-scale features are fused by sum operation. We investigate several fusing options on MS training from scratch in Fig. 5 (c). Subtracting (\"Minus\") a branch from the other leads to about the same results as \"Sum\" while multiplication (\"Mul\") performs poorly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 695, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 695, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 695, + 279, + 730 + ], + "type": "text", + "content": "Comparison with Transformer. Transformer, as a new architecture, has recently been continuously making progresses in many computer vision domains [13, 53], includ" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": "ing ReID [21, 32]. We compare the model performance with some baseline Transformer models in Tab. 6. DeiT-B and ViT-B [21] achieves higher performance on MSMT17, with much larger calculation burden compared with our proposed MSINet. On VeRi-776, MSINet surpasses all the baseline Transformer methods. It proves that rational interaction operations between multi-scale features are capable to assist light-weighted pure-CNN models to obtain comparable performance with complex Transformers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 211, + 531, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 211, + 531, + 293 + ], + "spans": [ + { + "bbox": [ + 296, + 211, + 531, + 293 + ], + "type": "text", + "content": "Parameter Analysis. Firstly, we study the influence of different receptive field scale ratios " + }, + { + "bbox": [ + 296, + 211, + 531, + 293 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 296, + 211, + 531, + 293 + ], + "type": "text", + "content": " inside an MSI cell on MS training from scratch in Fig. 5 (d). Introducing scale differences between branches improves the model performance significantly, and subsequent increases brings more modest impacts. Considering both parameter scales and model performance, the ratio of 3:1 is selected for MSINet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 293, + 530, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 293, + 530, + 362 + ], + "spans": [ + { + "bbox": [ + 296, + 293, + 530, + 362 + ], + "type": "text", + "content": "Secondly, the model performance fluctuation influenced by spatial alignment weight is visualized in Fig. 5 (e). The experiment is conducted on the VR " + }, + { + "bbox": [ + 296, + 293, + 530, + 362 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 296, + 293, + 530, + 362 + ], + "type": "text", + "content": " VID cross-domain scenario. Employing the alignment generally makes a positive impact on the generalization capability of the model. The optimal loss weight " + }, + { + "bbox": [ + 296, + 293, + 530, + 362 + ], + "type": "inline_equation", + "content": "\\lambda_{sa}" + }, + { + "bbox": [ + 296, + 293, + 530, + 362 + ], + "type": "text", + "content": " locates at 2.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 363, + 530, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 363, + 530, + 468 + ], + "spans": [ + { + "bbox": [ + 296, + 363, + 530, + 468 + ], + "type": "text", + "content": "Visualization Results. We visualize the top-15 retrieved sequences and the corresponding distances from an query image on VR in Fig. 6. By comparison, ResNet50 mainly focuses on general appearance features, where the top-rated negative samples share similar car bodies. MSINet, oppositely, concentrates on discriminative distinctions, an empty car hopper in this case, and creates an evident distance gap between positive and negative samples. More details can be seen in the supplementation material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 479, + 367, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 479, + 367, + 491 + ], + "spans": [ + { + "bbox": [ + 297, + 479, + 367, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 500, + 530, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 500, + 530, + 627 + ], + "spans": [ + { + "bbox": [ + 296, + 500, + 530, + 627 + ], + "type": "text", + "content": "In this paper, we design a Twins Contrastive Mechanism for NAS to build the compatibility with ReID. The task-specific search scheme provides the searching process with more appropriate supervision. A Multi-Scale Interaction search space is proposed to establish rational and flexible utilization of multi-scale features. With a Spatial Alignment Module, our proposed MSINet achieves SOTA performance on both supervision and cross-domain scenarios with limited parameter amount. We hope the proposed approach could inspire more works focusing on designing network architectures suitable for ReID tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 639, + 404, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 639, + 404, + 653 + ], + "spans": [ + { + "bbox": [ + 297, + 639, + 404, + 653 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "content": "This work is partially supported by National Natural Science Foundation of China (62173302, 62006244), China Scholarship Council (202206320302), Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and AI Singapore Programme (AISG2-PhD-2021-08-008)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19250" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 125, + 279, + 730 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 51, + 125, + 279, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 125, + 279, + 158 + ], + "spans": [ + { + "bbox": [ + 51, + 125, + 279, + 158 + ], + "type": "text", + "content": "[1] Irwan Bello, Barret Zoph, Vijay Vasudevan, and Quoc V Le. Neural optimizer search with reinforcement learning. In ICML, pages 459-468, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 159, + 279, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 159, + 279, + 191 + ], + "spans": [ + { + "bbox": [ + 51, + 159, + 279, + 191 + ], + "type": "text", + "content": "[2] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. In ICLR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 193, + 279, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 193, + 279, + 236 + ], + "spans": [ + { + "bbox": [ + 51, + 193, + 279, + 236 + ], + "type": "text", + "content": "[3] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Joint generative and contrastive learning for unsupervised person re-identification. In CVPR, pages 2004–2013, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 238, + 279, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 238, + 279, + 280 + ], + "spans": [ + { + "bbox": [ + 51, + 238, + 279, + 280 + ], + "type": "text", + "content": "[4] Tsai-Shien Chen, Chih-Ting Liu, Chih-Wei Wu, and Shao-Yi Chien. Orientation-aware vehicle re-identification with semantics-guided part attention network. In ECCV, pages 330–346, 2020. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 282, + 279, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 282, + 279, + 325 + ], + "spans": [ + { + "bbox": [ + 51, + 282, + 279, + 325 + ], + "type": "text", + "content": "[5] Xin Chen, Lingxi Xie, Jun Wu, and Qi Tian. Progressive differentiable architecture search: Bridging the depth gap between search and evaluation. In ICCV, pages 1294-1303, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 327, + 279, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 327, + 279, + 370 + ], + "spans": [ + { + "bbox": [ + 51, + 327, + 279, + 370 + ], + "type": "text", + "content": "[6] De Cheng, Jingyu Zhou, Nannan Wang, and Xinbo Gao. Hybrid dynamic contrast and probability distillation for unsupervised person re-id. IEEE Transactions on Image Processing, 31:3334-3346, 2022. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 372, + 279, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 372, + 279, + 405 + ], + "spans": [ + { + "bbox": [ + 51, + 372, + 279, + 405 + ], + "type": "text", + "content": "[7] Xiangxiang Chu, Tianbao Zhou, Bo Zhang, and Jixiang Li. Fair darts: Eliminating unfair advantages in differentiable architecture search. In ECCV, pages 465-480, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 406, + 279, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 406, + 279, + 448 + ], + "spans": [ + { + "bbox": [ + 51, + 406, + 279, + 448 + ], + "type": "text", + "content": "[8] Xiyang Dai, Dongdong Chen, Mengchen Liu, Yinpeng Chen, and Lu Yuan. Da-nas: Data adapted pruning for efficient neural architecture search. In ECCV, pages 584-600, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 451, + 279, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 451, + 279, + 492 + ], + "spans": [ + { + "bbox": [ + 51, + 451, + 279, + 492 + ], + "type": "text", + "content": "[9] Yongxing Dai, Jun Liu, Yifan Sun, Zekun Tong, Chi Zhang, and Ling-Yu Duan. Idm: An intermediate domain module for domain adaptive person re-id. In ICCV, pages 11864-11874, 2021. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 495, + 279, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 279, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 279, + 527 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 529, + 279, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 279, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 279, + 571 + ], + "type": "text", + "content": "[11] Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. Speeding up automatic hyperparameter optimization of deep neural networks by extrapolation of learning curves. In *IJCAI*, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 574, + 279, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 279, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 279, + 605 + ], + "type": "text", + "content": "[12] Xuanyi Dong and Yi Yang. Searching for a robust neural architecture in fourgpu hours. In CVPR, pages 1761-1770, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 608, + 279, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 279, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 279, + 660 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 663, + 279, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 279, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 279, + 696 + ], + "type": "text", + "content": "[14] Pengfei Fang, Jieming Zhou, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Bilinear attention networks for person retrieval. In ICCV, pages 8030-8039, 2019. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "type": "text", + "content": "[15] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 730 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 299, + 106, + 530, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 106, + 530, + 139 + ], + "spans": [ + { + "bbox": [ + 299, + 106, + 530, + 139 + ], + "type": "text", + "content": "[16] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In ICLR, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "type": "text", + "content": "[17] Yixiao Ge, Dapeng Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. In NeurIPS, 2020. 1, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 184, + 530, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 184, + 530, + 216 + ], + "spans": [ + { + "bbox": [ + 299, + 184, + 530, + 216 + ], + "type": "text", + "content": "[18] Yiluan Guo and Ngai-Man Cheung. Efficient and deep person re-identification using multi-level similarity. In CVPR, pages 2335–2344, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 217, + 530, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 217, + 530, + 258 + ], + "spans": [ + { + "bbox": [ + 299, + 217, + 530, + 258 + ], + "type": "text", + "content": "[19] Xumeng Han, Xuehui Yu, Nan Jiang, Guorong Li, Jian Zhao, Qixiang Ye, and Zhenjun Han. Group sampling for unsupervised person re-identification. arXiv preprint arXiv:2107.03024, 2021. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 260, + 530, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 260, + 530, + 292 + ], + "spans": [ + { + "bbox": [ + 299, + 260, + 530, + 292 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 293, + 530, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 293, + 530, + 325 + ], + "spans": [ + { + "bbox": [ + 299, + 293, + 530, + 325 + ], + "type": "text", + "content": "[21] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, pages 15013–15022, 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 326, + 530, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 326, + 530, + 367 + ], + "spans": [ + { + "bbox": [ + 299, + 326, + 530, + 367 + ], + "type": "text", + "content": "[22] Ruibing Hou, Bingpeng Ma, Hong Chang, Xinqian Gu, Shiguang Shan, and Xilin Chen. Interaction-and-aggregation network for person re-identification. In CVPR, pages 9317–9326, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 369, + 530, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 369, + 530, + 401 + ], + "spans": [ + { + "bbox": [ + 299, + 369, + 530, + 401 + ], + "type": "text", + "content": "[23] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 403, + 530, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 403, + 530, + 445 + ], + "spans": [ + { + "bbox": [ + 299, + 403, + 530, + 445 + ], + "type": "text", + "content": "[24] Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Meta pairwise relationship distillation for unsupervised person re-identification. In ICCV, pages 3661–3670, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 447, + 530, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 447, + 530, + 487 + ], + "spans": [ + { + "bbox": [ + 299, + 447, + 530, + 487 + ], + "type": "text", + "content": "[25] Xin Jin, Cuiling Lan, Wenjun Zeng, Zhibo Chen, and Li Zhang. Style normalization and restitution for generalizable person re-identification. In CVPR, pages 3143–3152, 2020. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 489, + 530, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 489, + 530, + 530 + ], + "spans": [ + { + "bbox": [ + 299, + 489, + 530, + 530 + ], + "type": "text", + "content": "[26] Xin Jin, Cuiling Lan, Wenjun Zeng, Guoqiang Wei, and Zhibo Chen. Semantics-aligned representation learning for person re-identification. In AAAI, pages 11173–11180, 2020. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 532, + 530, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 532, + 530, + 554 + ], + "spans": [ + { + "bbox": [ + 299, + 532, + 530, + 554 + ], + "type": "text", + "content": "[27] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 555, + 530, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 555, + 530, + 586 + ], + "spans": [ + { + "bbox": [ + 299, + 555, + 530, + 586 + ], + "type": "text", + "content": "[28] Hanjun Li, Gaojie Wu, and Wei-Shi Zheng. Combined depth space based architecture search for person re-identification. In CVPR, pages 6729–6738, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 588, + 530, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 588, + 530, + 619 + ], + "spans": [ + { + "bbox": [ + 299, + 588, + 530, + 619 + ], + "type": "text", + "content": "[29] Jianing Li and Shiliang Zhang. Joint visual and temporal consistency for unsupervised domain adaptive person re-identification. In ECCV, pages 483–499, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 621, + 530, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 621, + 530, + 653 + ], + "spans": [ + { + "bbox": [ + 299, + 621, + 530, + 653 + ], + "type": "text", + "content": "[30] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deepreid: Deep filter pairing neural network for person re-identification. In CVPR, pages 152–159, 2014. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 655, + 530, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 655, + 530, + 685 + ], + "spans": [ + { + "bbox": [ + 299, + 655, + 530, + 685 + ], + "type": "text", + "content": "[31] Wei Li, Xiatian Zhu, and Shaogang Gong. Harmonious attention network for person re-identification. In CVPR, pages 2285–2294, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "type": "text", + "content": "[32] Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded person re-identification with part-aware transformer. In CVPR, pages 2898–2907, 2021. 8" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "type": "text", + "content": "19251" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 728 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[33] Shengcai Liao and Ling Shao. Interpretable and generalizable person re-identification with query-adaptive convolution and temporal lifting. In ECCV, pages 456–474, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 139, + 279, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 279, + 181 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 279, + 181 + ], + "type": "text", + "content": "[34] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shiens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In ECCV, pages 19–34, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 182, + 279, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 279, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 279, + 213 + ], + "type": "text", + "content": "[35] Hanxiao Liu, Karen Simonyan, Oriol Vinyals, Chrisantha Fernando, and Koray Kavukcuoglu. Hierarchical representations for efficient architecture search. In ICLR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 214, + 278, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 278, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 278, + 234 + ], + "type": "text", + "content": "[36] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In ICLR, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 236, + 279, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 279, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 279, + 277 + ], + "type": "text", + "content": "[37] Hongye Liu, Yonghong Tian, Yaowei Yang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In CVPR, pages 2167-2175, 2016. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 278, + 278, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 278, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 278, + 310 + ], + "type": "text", + "content": "[38] Xinchen Liu, Wu Liu, Huadong Ma, and Huiyuan Fu. Large-scale vehicle re-identification in urban surveillance videos. In ICME, pages 1–6, 2016. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 311, + 279, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 279, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 279, + 352 + ], + "type": "text", + "content": "[39] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In ECCV, pages 869–884, 2016. 1, 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 354, + 279, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 279, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 279, + 394 + ], + "type": "text", + "content": "[40] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. Provid: Progressive and multimodal vehicle reidentification for large-scale urban surveillance. T-MM, 20(3):645-658, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 396, + 279, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 279, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 279, + 428 + ], + "type": "text", + "content": "[41] Chuanchen Luo, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Spectral feature transformation for person re-identification. In ICCV, pages 4976–4985, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 429, + 279, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 279, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 279, + 471 + ], + "type": "text", + "content": "[42] Hao Luo, Wei Jiang, Youzhi Gu, Fuxu Liu, Xingyu Liao, Shenqi Lai, and Jianyang Gu. A strong baseline and batch normalization neck for deep person re-identification. T-MM, 22(10):2597-2609, 2019. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 472, + 279, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 279, + 492 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 279, + 492 + ], + "type": "text", + "content": "[43] Renqian Luo, Fei Tian, Tao Qin, Enhong Chen, and Tie-Yan Liu. Neural architecture optimization. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 494, + 279, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 279, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 279, + 535 + ], + "type": "text", + "content": "[44] Xuelin Qian, Yanwei Fu, Yu-Gang Jiang, Tao Xiang, and Xiangyang Xue. Multi-scale deep learning architectures for person re-identification. In ICCV, pages 5399–5408, 2017. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 537, + 279, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 279, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 279, + 577 + ], + "type": "text", + "content": "[45] Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, and Yi Yang. Auto-reid: Searching for a part-aware convnet for person re-identification. In ICCV, pages 3750–3759, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 579, + 279, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 279, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 279, + 611 + ], + "type": "text", + "content": "[46] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In AAAI, pages 4780-4789, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 612, + 279, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 279, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 279, + 652 + ], + "type": "text", + "content": "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 655, + 279, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 279, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 279, + 686 + ], + "type": "text", + "content": "[48] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 687, + 279, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 279, + 728 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 279, + 728 + ], + "type": "text", + "content": "[49] Yantao Shen, Tong Xiao, Hongsheng Li, Shuai Yi, and Xiaogang Wang. Learning deep neural networks for vehicle re-id with visual-spatio-temporal path proposals. In ICCV, pages 1900-1909, 2017. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 729 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "text", + "content": "[50] Liangchen Song, Cheng Wang, Lefei Zhang, Bo Du, Qian Zhang, Chang Huang, and Xinggang Wang. Unsupervised domain adaptive re-identification: Theory and practice. PR, 102:107173, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 151, + 530, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 151, + 530, + 193 + ], + "spans": [ + { + "bbox": [ + 299, + 151, + 530, + 193 + ], + "type": "text", + "content": "[51] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV, pages 480-496, 2018. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 195, + 530, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 195, + 530, + 237 + ], + "spans": [ + { + "bbox": [ + 299, + 195, + 530, + 237 + ], + "type": "text", + "content": "[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 239, + 530, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 239, + 530, + 282 + ], + "spans": [ + { + "bbox": [ + 299, + 239, + 530, + 282 + ], + "type": "text", + "content": "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 284, + 530, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 284, + 530, + 315 + ], + "spans": [ + { + "bbox": [ + 299, + 284, + 530, + 315 + ], + "type": "text", + "content": "[54] Dongkai Wang and Shiliang Zhang. Unsupervised person re-identification via multi-label classification. In CVPR, pages 10981-10990, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 318, + 530, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 318, + 530, + 360 + ], + "spans": [ + { + "bbox": [ + 299, + 318, + 530, + 360 + ], + "type": "text", + "content": "[55] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACMMM, pages 274–282, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 362, + 530, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 362, + 530, + 394 + ], + "spans": [ + { + "bbox": [ + 299, + 362, + 530, + 394 + ], + "type": "text", + "content": "[56] Yicheng Wang, Zhenzhong Chen, Feng Wu, and Gang Wang. Person re-identification with cascaded pairwise convolutions. In CVPR, pages 1470–1478, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 396, + 530, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 396, + 530, + 438 + ], + "spans": [ + { + "bbox": [ + 299, + 396, + 530, + 438 + ], + "type": "text", + "content": "[57] Zhikang Wang, Lihuo He, Xiaoguang Tu, Jian Zhao, Xinbo Gao, Shengmei Shen, and Jiashi Feng. Robust video-based person re-identification by hierarchical mining. T-CSVT, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 441, + 530, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 441, + 530, + 494 + ], + "spans": [ + { + "bbox": [ + 299, + 441, + 530, + 494 + ], + "type": "text", + "content": "[58] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In ICCV, pages 379–387, 2017. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 495, + 530, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 495, + 530, + 538 + ], + "spans": [ + { + "bbox": [ + 299, + 495, + 530, + 538 + ], + "type": "text", + "content": "[59] Zhongdao Wang, Jingwei Zhang, Liang Zheng, Yixuan Liu, Yifan Sun, Yali Li, and Shengjin Wang. Cycles: Self-supervised cycle association for learning re-identifiable descriptions. In ECCV, pages 72–88, 2020. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 540, + 530, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 540, + 530, + 572 + ], + "spans": [ + { + "bbox": [ + 299, + 540, + 530, + 572 + ], + "type": "text", + "content": "[60] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In CVPR, pages 79–88, 2018. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 574, + 530, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 574, + 530, + 607 + ], + "spans": [ + { + "bbox": [ + 299, + 574, + 530, + 607 + ], + "type": "text", + "content": "[61] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 607, + 530, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 607, + 530, + 650 + ], + "spans": [ + { + "bbox": [ + 299, + 607, + 530, + 650 + ], + "type": "text", + "content": "[62] Yuhui Xu, Lingxi Xie, Xiaopeng Zhang, Xin Chen, Guo-Jun Qi, Qi Tian, and Hongkai Xiong. Pc-darts: Partial channel connections for memory-efficient architecture search. In ICLR, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 652, + 530, + 694 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 652, + 530, + 694 + ], + "spans": [ + { + "bbox": [ + 299, + 652, + 530, + 694 + ], + "type": "text", + "content": "[63] Cheng Yan, Guansong Pang, Xiao Bai, Changhong Liu, Ning Xin, Lin Gu, and Jun Zhou. Beyond triplet loss: person re-identification with fine-grained difference-aware pairwise loss. T-MM, 2021. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 697, + 530, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 697, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 697, + 530, + 729 + ], + "type": "text", + "content": "[64] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C.H. Hoi. Deep learning for person re-identification: A survey and outlook. T-PAMI, 2021. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "19252" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 629 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[65] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Deep metric learning for person re-identification. In ICPR, pages 34–39, 2014. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "type": "text", + "content": "[66] Anguo Zhang, Yueming Gao, Yuzhen Niu, Wenxi Liu, and Yongcheng Zhou. Coarse-to-fine person re-identification with auxiliary-domain classification and second-order information bottleneck. In CVPR, pages 598–607, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 183, + 279, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 183, + 279, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 279, + 216 + ], + "type": "text", + "content": "[67] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Xin Jin, and Zhibo Chen. Relation-aware global attention for person re-identification. In CVPR, pages 3186–3195, 2020. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 217, + 279, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 279, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 279, + 258 + ], + "type": "text", + "content": "[68] Aihua Zheng, Xianmin Lin, Jiacheng Dong, Wenzhong Wang, Jin Tang, and Bin Luo. Multi-scale attention vehicle re-identification. Neural Computing and Applications, 32(23):17489-17503, 2020. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 260, + 279, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 279, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 279, + 292 + ], + "type": "text", + "content": "[69] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, pages 1116–1124, 2015. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 293, + 279, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 279, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 279, + 325 + ], + "type": "text", + "content": "[70] Liang Zheng, Yi Yang, and Alexander G Hauptmann. Person re-identification: Past, present and future. arXiv preprint arXiv:1610.02984, 2016. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 326, + 279, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 279, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 279, + 357 + ], + "type": "text", + "content": "[71] Liang Zheng, Yi Yang, and Qi Tian. Sift meets cnn: A decade survey of instance retrieval. T-PAMI, 40(5):1224-1244, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 359, + 279, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 279, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 279, + 390 + ], + "type": "text", + "content": "[72] Meng Zheng, Srikrishna Karanam, Ziyan Wu, and Richard J Radke. Re-identification with consistent attentive siamese networks. In CVPR, pages 5735–5744, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 392, + 279, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 279, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 279, + 433 + ], + "type": "text", + "content": "[73] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In CVPR, pages 2138–2147, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 436, + 279, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 279, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 279, + 466 + ], + "type": "text", + "content": "[74] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, pages 13001-13008, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 468, + 279, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 279, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 279, + 510 + ], + "type": "text", + "content": "[75] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re-identification. In ICCV, pages 3702–3712, 2019. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 512, + 279, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 279, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 279, + 554 + ], + "type": "text", + "content": "[76] Sanping Zhou, Fei Wang, Zeyi Huang, and Jinjun Wang. Discriminative feature learning with consistent attention regularization for person re-identification. In ICCV, pages 8040–8049, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 555, + 279, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 279, + 587 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 279, + 587 + ], + "type": "text", + "content": "[77] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697-8710, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 589, + 279, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 279, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 279, + 629 + ], + "type": "text", + "content": "[78] Yang Zou, Xiaodong Yang, Zhiding Yu, BVK Vijaya Kumar, and Jan Kautz. Joint disentangling and adaptation for cross-domain person re-identification. In ECCV, pages 87–104, 2020. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "type": "text", + "content": "19253" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_content_list.json b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fed071eae0c736d6005b47f4c55340b7b9386bd0 --- /dev/null +++ b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_content_list.json @@ -0,0 +1,1368 @@ +[ + { + "type": "text", + "text": "MSMDFusion: Fusing LiDAR and Camera at Multiple Scales with Multi-Depth Seeds for 3D Object Detection", + "text_level": 1, + "bbox": [ + 148, + 128, + 821, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Jiao $^{1,2*}$ Zequn Jie $^{3*}$ Shaoxiang Chen $^{3}$ Jingjing Chen $^{1,2\\dagger}$ Lin Ma $^{3\\dagger}$ Yu-Gang Jiang $^{1,2}$", + "bbox": [ + 91, + 200, + 877, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University", + "bbox": [ + 173, + 220, + 797, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai Collaborative Innovation Center of Intelligent Visual Computing \n $^{3}$ Meituan", + "bbox": [ + 184, + 238, + 784, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fusing LiDAR and camera information is essential for accurate and reliable 3D object detection in autonomous driving systems. This is challenging due to the difficulty of combining multi-granularity geometric and semantic features from two drastically different modalities. Recent approaches aim at exploring the semantic densities of camera features through lifting points in 2D camera images (referred to as \"seeds\") into 3D space, and then incorporate 2D semantics via cross-modal interaction or fusion techniques. However, depth information is under-investigated in these approaches when lifting points into 3D space, thus 2D semantics can not be reliably fused with 3D points. Moreover, their multi-modal fusion strategy, which is implemented as concatenation or attention, either can not effectively fuse 2D and 3D information or is unable to perform fine-grained interactions in the voxel space. To this end, we propose a novel framework with better utilization of the depth information and fine-grained cross-modal interaction between LiDAR and camera, which consists of two important components. First, a Multi-Depth Unprojection (MDU) method is used to enhance the depth quality of the lifted points at each interaction level. Second, a Gated Modality-Aware Convolution (GMA-Conv) block is applied to modulate voxels involved with the camera modality in a fine-grained manner and then aggregate multi-modal features into a unified space. Together they provide the detection head with more comprehensive features from LiDAR and camera. On the nuScenes test benchmark, our proposed method, abbreviated as MSMD-Fusion, achieves state-of-the-art results on both 3D object detection and tracking tasks without using test-time-augmentation and ensemble techniques. The code is available at https://github.com/SxJyJay/MSMDFusion.", + "bbox": [ + 75, + 342, + 473, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7581a5d6ba857f2a80165958d2500486eb0afe2ee11843b90dffca75505b1034.jpg", + "image_caption": [ + "Figure 1. Illustration of our MSMDFusion pipeline. The yellow arrows indicate information passing or interaction between the Li-DAR and camera modalities." + ], + "image_footnote": [], + "bbox": [ + 503, + 309, + 893, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 525, + 632, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detecting 3D objects [1, 22, 29] is regarded as a fundamental task for autonomous driving. Aiming at the robust environmental perception, LiDARs and cameras are widely equipped on autonomous driving vehicles since they can provide complementary information. Characterized by point clouds, LiDARs can capture accurate spatial information, while cameras contain rich semantics and context with images. Therefore, developing multi-modal detectors that enjoy the benefits of the two worlds is promising. Such an idea has catalyzed the emergence of a set of recent researches [1, 3, 12, 13, 15, 17, 19, 20, 25, 26]. Early works [1, 3, 11, 12, 18-20, 26] perform LiDAR-camera fusion by projecting 3D LiDAR points (or region proposals generated from them) onto 2D image planes to collect useful 2D semantics. However, such a paradigm suffers from the signal density mismatching of multi-modality sensors. Since LiDAR points are much sparser than camera pixels, this projection manner will inevitably waste semantically rich 2D features [1, 13].", + "bbox": [ + 496, + 551, + 890, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, another paradigm [9,10,13,15,25] for LiDAR-camera fusion has emerged. Instead of collecting 2D semantic features with 3D queries, these methods first estimate depths of pixels, and then directly lift 2D pixels with", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 96, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21643", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "their semantics to the 3D world (we refer to these pixels and corresponding lifted points as \"seeds\" and \"virtual points\" in this paper) to fuse with the real 3D point cloud. Two methods with the same name of BEVFusion [13, 15] treat every image feature pixel as a seed and generate virtual points in the BEV space. MVP [25] and VFF [10] sample pixels from foreground regions and lift them to the voxel space. Benefited from the dense virtual points, such a paradigm not only maintains the semantic consistency in the image [15], but also complements the geometric cues for the sparse LiDAR point cloud [25].", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant improvements have been made, existing methods in this line suffer from two major problems, which hampers benefiting from virtual points. First, depth, as the key to the quality of virtual points, is underinvestigated in generating virtual points. On the one hand, depth directly determines the spatial location in 3D space of a seed via perspective projection which can thereby significantly affect 3D object detection results. On the other hand, depth can also enhance the semantics carried by virtual points by providing color-insensitive cues in describing objects [27], since combining RGB information with depth guidance correlates camera pixels of similar depths and enables them to jointly contribute to capturing instance-related semantics when lifted as virtual points. Existing multi-modal detectors [9, 13, 15] mainly pay attention on interacting LiDAR points with camera virtual points, while ignoring the importance of seed depths in generating the virtual points.", + "bbox": [ + 75, + 260, + 472, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, the fine-grained cross-modal interaction between virtual points and 3D points in the uncompressed space (e.g., the voxel space) is crucial but non-trivial. Generated virtual points are geometrically and semantically inconsistent with real LiDAR points due to imperfect depths and inherent modality gap. Hence, in order to benefit from the semantically rich virtual points, it is necessary to adaptively select helpful information from virtual points under the guidance of real LiDAR points in a fine-grained and controllable manner. However, such a cross-modal interaction is constrained by the intensive memory and computation cost brought by the massive amounts and unstructured nature of point cloud data. Alternatively, existing approaches combine the multi-modal information with simple concatenate [25] or add operations [9] in the voxel space, or perform cross-attention in a compressed BEV space [23].", + "bbox": [ + 75, + 534, + 472, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aiming at unlocking the potential of virtual points and addressing the drawbacks of existing methods, we propose a multi-scale fusion framework, called MSMDFusion, and within each scale, there are two key novel designs, namely the Multi-Depth Unprojection (MDU) and Gated Modality-Aware Convolution (GMA-Conv). As shown in Fig.1, MDU is mainly for enhancing the quality of generated virtual points in terms of geometric accuracy and", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "semantic richness. To be specific, when lifting 2D seeds from image into the 3D space, multiple depths are explored within a reference neighborhood to generate virtual points with more reliable depth. Next, the camera feature and depth are combined to produce depth-aware features as stronger 2D semantics to decorate these virtual points. GMA-Conv takes real LiDAR points and generated virtual points as inputs, and performs fine-grained interaction in a select-then-aggregate manner. Concretely, we first adaptively select useful information from camera voxel features under the guidance of reference LiDAR voxels, then aggregate them grouped sparse convolutions for sufficient multimodal interaction. We also specifically adopt a voxel subsampling strategy to efficiently obtain reliable LiDAR references when implementing our GMA-Conv.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, with the resulting multi-modal voxel features from multiple scales, we further associate them with cascade connections across scales to aggregate multi-granularity information. With the above designs, the camera semantics encapsulated in the virtual points are consistently combined with LiDAR points, and thereby providing a stronger multi-modal feature representation for boosting the 3D object detection. As shown in the Table 3, with 100 times fewer generated virtual points than the two BEVFusion methods [13, 15], our MSMDFusion can still achieve state-of-the-art performances.", + "bbox": [ + 496, + 318, + 893, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions lie in threefold: (1) We propose a novel MSMDFusion approach, which encourages sufficient LiDAR-Camera feature fusion in the multiscale voxel space. (2) Within each scale, we propose a Multi-Depth Unprojection strategy (MDU) to promote virtual points generation with better locations and semantics by fully leveraging depth of pixels, as well as a Gated Modality-Aware Convolution (GMA-Conv) to achieve fine-grained controllable multi-modal interaction. (3) Extensive experimental results on the large-scale nuScenes dataset prove the effectiveness of our MSMDFusion and its components. We achieve state-of-the-art performances with $71.5\\%$ mAP and $74.0\\%$ NDS on the challenging nuScenes detection track using single model1. When combining the simple greedy tracking strategy [24], our method also achieves strong tracking results with $74.0\\%$ AMOTA.", + "bbox": [ + 496, + 486, + 895, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 744, + 640, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, fusing LiDAR and camera signals in 3D detection has attracted increasing interest. Early works [3, 6, 12, 19, 20, 26] project 3D queries to 2D camera images for collecting useful semantics. Under such a paradigm, MV3D [3] and AVOD [7] associate 3D proposals with 2D RoI features. While PointPainting [19] and PointAugment-", + "bbox": [ + 496, + 771, + 893, + 863 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We do not use Test-Time Augmentation (TTA) and model ensemble technique.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21644", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1439e75b70d6e192392497352d904d9fcd101bf660b9b3d9cacbabe687410320.jpg", + "image_caption": [ + "Figure 2. The framework of our proposed MSMDFusion. Given a LiDAR point cloud and corresponding multi-view camera images, MSMDFusion first separately transforms them into the voxel space and obtains multi-scale LiDAR and camera features. Then, the LiDAR and camera features at each scale interact through our proposed Multi-Depth Unprojection strategy (MDU) and Gated Modality-Aware Convolution block (GMA-Conv, details in Fig.4). The outputs of GMA-Conv at each scale are further aggregated for combining multi-granularity information. Finally, the resulting multi-modal and LiDAR voxel features are flattened into the BEV space for the final prediction. $\\oplus$ represents voxel addition and \"spconv\" indicates 3D sparse convolution." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing [20] directly decorate raw 3D points with 2D semantics. EP-Net [6] and 3D-CVF [26] perform multi-modal fusion at both point level and proposal level. However, since 3D points are inherently sparse, such a hard association approach wastes the dense semantic information in 2D features. Recently, multi-modal 3D detectors [13, 15, 23, 25] lift dense 2D seeds to 3D space for learning the 2D-3D joint representation in a shared space. Two BEVFusion methods [13, 15] densely lift every image feature pixel into 3D space and then encode these lifted points as another BEV (bird's-eye view) map to fuse with the BEV map obtained using the original 3D point cloud. LiDAR and camera signals will finally share a common representation space, but they do not interact at any stage before BEV fusion. MVP [25] selects seeds from foreground objects and unprojected them into the LiDAR frame for points cloud augmentation. With the seeds sampling strategy similar to MVP, UVTR [9] constructs both LiDAR and camera branches in the voxel space to maintain the modality-specific information, and then interact them with voxel add operation. However, the LiDAR and camera branches are constructed in a single scale of voxel space, losing the benefits of different levels of feature abstraction in multi-scale voxel space. Moreover, merging LiDAR and camera voxels with simple addition limits their fine-grained interaction.", + "bbox": [ + 75, + 491, + 472, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 489, + 591, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework Overview", + "text_level": 1, + "bbox": [ + 500, + 513, + 705, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overall view of MSMDFusion is shown in Fig.2. Given a LiDAR point cloud and corresponding multiview camera images as inputs, MSMDFusion first extracts multi-scale features from both modalities in the voxel space. Then, LiDAR-camera interaction is performed within multi-scale voxel space to properly combine multi-granularity information from both modalities. At each scale, we specifically design a Multi-Depth Unprojection (MDU) strategy to obtain high-quality virtual points in the voxel space, and a Gated Modality-Aware Convolution block (GMA-Conv) for effective LiDAR-camera interaction and fusion. We also introduce cross-scale connections to progressively combine features of different granularities. Afterward, the deeply interacted multi-modal features, together with the LiDAR features, are transformed into the BEV space and fed to the BEV encoder and detection head for the final prediction.", + "bbox": [ + 496, + 537, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. LiDAR and Camera Feature Extraction", + "text_level": 1, + "bbox": [ + 500, + 801, + 841, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first extract high-level features from the raw inputs of LiDAR and camera. For a given LiDAR point cloud, we voxelize the points and extract their features in the voxel space using a set of 3D sparse convolutional blocks [4] following prior arts [1, 13, 15]. Each convolutional block out", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21645", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7eca596c2913e40ab3a1763b89398cbca7e2c909445143e62e77b2f288f402f7.jpg", + "image_caption": [ + "large depth bias moderate depth bias", + "(a) Nearest neighbor retrieval", + "Figure 3. Comparison of the effects of retrieving reference points from nearest neighbor (MVP) and K-nearest neighbor (Ours). The blue circle and yellow star represent seeds and reference points, respectively. To keep the figure concise, we only depict the case that $\\mathrm{K} = 2$ . Our strategy can obtain more reliable depths by exploring more neighbors." + ], + "image_footnote": [], + "bbox": [ + 96, + 104, + 253, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dbd763435d95891ae079e5fda7da272a8252a336cd42646e1d1028298d998970.jpg", + "image_caption": [ + "(b) K Nearest neighbor retrieval (Ours)" + ], + "image_footnote": [], + "bbox": [ + 285, + 94, + 444, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "puts voxel features of different scales, representing different levels of abstraction for point cloud, and these multiscale features will further interact with their 2D counterparts through our proposed GMA-Conv blocks.", + "bbox": [ + 75, + 335, + 467, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the multi-view camera images, we adopt ResNet50 with FPN as the image backbone to extract multi-scale image features, which contains rich object semantics. However, it is non-trivial to transfer these useful semantics to the 3D voxel field, since images are inherently 2D data. Motivated by recent works [13, 15, 25] which select seeds from the image plane and estimate their depths to lift them as 3D virtual points, we further propose a Multi-Depth Unprojection strategy to lift 2D seeds, which could mitigate the drawbacks of prior works. The details will be presented in the following subsections.", + "bbox": [ + 75, + 397, + 467, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Modal Interaction in Voxel Space", + "text_level": 1, + "bbox": [ + 76, + 573, + 416, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The extracted LiDAR and camera features from the previous stage respectively focus on geometric and semantic information, hence the goal of multi-modal interaction is to properly fuse these features into a unified rich representation of the scene. As shown in the Fig.2, the components of LiDAR-camera interaction within each scale in the voxel space are conceptually simple: 1) Multi-Depth Unprojection (MDU) strategy is introduced to improve the quality of the generated virtual points and 2) Gated Modality-Aware Convolution (GMA-Conv) block is designed for controllable LiDAR-camera fusion. In the following parts, we will first elaborate on MDU and GMA-Conv within a specific scale, then on top of them, we describe the complete multiscale fusion process.", + "bbox": [ + 75, + 597, + 467, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Multi-Depth Unprojection", + "text_level": 1, + "bbox": [ + 76, + 830, + 313, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To lift a 2D seed in the pixel space to the 3D space (i.e., the unprojection operation), the depth associated with each seed should be estimated [21]. To obtain an initial reli", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "able depth estimation, we use the strategy from MVP [25] as the basis as shown in Fig.3(a). Formally, given a point cloud $\\mathcal{P} = \\{(x_i,y_i,z_i)\\}_{i = 1}^{N_P}$ and multi-view images as inputs, MVP first projects 3D points onto 2D images and preserves points falling within the 2D foreground instance masks. These 3D points are kept as reference points to provide depth, and they are denoted as $\\mathcal{R} = \\{(u_i,v_i,d_i)\\}_{i = 1}^{N_R}$ , where $u_{i}$ and $v_{i}$ are pixel coordinates and $d_{i}$ is the real depth. Then, a set of seeds $S = \\{(u_i^s,v_i^s)\\}_{i = 1}^{N_S}$ are uniformly sampled from each instance mask on the image, and each of them will retrieve a real depth from its nearest reference point as its estimated depth. Finally the sampled seeds with estimated depths are unprojected to 3D as virtual points.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Though effective, such a strategy ignores the fact that spatial proximity in 2D images can not be guaranteed in 3D, and this may lead to inaccurate depth estimation as shown in the red circle of Fig.3(a). Toward this end, we propose to equip every seed with multiple depths by retrieving K-nearest reference points, which can be regarded as a soft strategy to achieve more reliable depth estimation. As shown in the green circle of the Fig.3(b), each seed can be unprojected using multiple (K) depths from neighboring reference points, which generates K virtual points to improve its recall of the actual 3D points. We provide more detailed discussion on the effect of multi-depth seeds in Sec.4.3.3.", + "bbox": [ + 496, + 287, + 890, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These generated virtual points will be further decorated to make the best use of semantics encapsulated in images. But different from MVP [25] that simply decorates virtual points with class scores predicted by a pretrained 2D detector, we make an early interaction between image features and sparse depth information from reference points to generate depth-aware semantic features. These depth-aware semantic features are then used to decorate virtual points in an adaptive manner controlled by their corresponding seed depths. The whole process can be trained end-to-end. Concretely, we concatenate the camera image feature $C$ with a sparse depth map containing all depths of reference points $\\mathcal{R}$ in that image, and then fuse them with a convolutional layer to obtain the depth-aware semantic feature $C^d$ . For the K virtual points generated by the same seed, the semantic feature of the corresponding pixel should also contribute differently along the projection ray in the 3D space [16]. Thus we compute a dynamic weight factor for each individual depth to modulate the semantic feature. The process for a specific seed $(u_i^s, v_i^s)$ can be formally described as:", + "bbox": [ + 496, + 483, + 890, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s _ {i} ^ {k} = \\operatorname {S i g m o i d} \\left(\\operatorname {L i n e a r} \\left([ C ^ {d} [ u _ {i} ^ {s}, v _ {i} ^ {s} ]; d _ {i} ^ {s, k} ]\\right)\\right) \\tag {1} \\\\ c _ {i} ^ {k} = C ^ {d} [ u _ {i} ^ {s}, v _ {i} ^ {s} ] * s _ {i} ^ {k} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 791, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{i}^{s,k}$ is the $k$ -th estimated depth of seed $(u_{i}^{s}, v_{i}^{s})$ , $s_{i}^{k} \\in [0,1]$ represents its weight factor, and $c^{k}$ is the modulated semantic feature, which will be used to decorate the corresponding unprojected virtual point $(u_{i}^{s}, v_{i}^{s}, d_{i}^{s,k})$ .", + "bbox": [ + 496, + 838, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21646", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eefbd71802db6811b3414bcb8786caddd5cfa9cc2ec90d279abfdd39c418ec4a.jpg", + "image_caption": [ + "Figure 4. Illustration of the select-then-aggregate process in the Gated Modality-Aware Convolution. Colors indicate different modalities: yellow for camera, blue for LiDAR, and red for LiDAR-camera combined. Cubes with shadows represent voxels selected with gates." + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 810, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The resulting virtual points with decorated features will be voxelized to match the output resolution of the corresponding scale of LiDAR branch, in order to perform interaction in the voxel space.", + "bbox": [ + 75, + 275, + 470, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Gated Modality-Aware Convolution", + "text_level": 1, + "bbox": [ + 76, + 358, + 380, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the resulting LiDAR and camera voxel features that share the same spatial resolution (i.e., the vertically aligned blue and yellow voxels in Fig.2), we aim to interact them in a fine-grained and controllable manner with our proposed Gated Modality-Aware Convolution (GMA-Conv) block in each scale. As shown in Fig.4, we first group the voxels according to their modalities: LiDAR-only (blue voxels), camera-only (yellow voxels), as well as LiDAR and camera combined (red voxels), and they are denoted as $f^L$ , $f^C$ , and $f^{LC}$ , respectively. Then, more fine-grained multi-modal interaction is conducted in a select-then-aggregate manner.", + "bbox": [ + 75, + 382, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Select. Motivated by the fact that LiDAR-based detectors usually surpass their camera-based counterparts with a large margin, we take LiDAR as the guiding modality to select useful information from camera features. Specifically, we design a gate conditioned on the LiDAR feature to control and update its camera counterpart, which can be formally described as:", + "bbox": [ + 75, + 549, + 468, + 655 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f _ {i}} ^ {C} = \\operatorname {R e L U} \\left(\\operatorname {L i n e a r} \\left(f _ {j} ^ {L}\\right)\\right) * f _ {i} ^ {C} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 667, + 468, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{f_i^C}$ is the updated camera voxel feature, and $i$ and $j$ are voxel indices of paired camera and LiDAR voxel features ( $f_j^L$ can be regarded as the reference voxel for $f_i^C$ ). The key to make this selection process effective is how to efficiently find reliable reference voxels. A straightforward way is to retrieve the nearest LiDAR voxels for each camera-only voxel $f_i^C$ from $f^L$ . And similarly for LiDAR-camera voxels $f_k^{LC}$ , the reference can be retrieved from both the LiDAR and camera modalities (i.e., $f^L$ and $f^C$ ). Although nearby neighbors from more reliable modalities can serve as a good reference, such a retrieval process is infeasible to be implemented with exhaustive pairwise distance calculation, because the computation and memory", + "bbox": [ + 75, + 702, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cost brought by massive amounts of LiDAR and camera voxels is intractable, and we will give more analysis and our solution in the supplementary material.", + "bbox": [ + 496, + 276, + 890, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aggregate. With the original LiDAR-only voxels $f^L$ , as well as the updated camera-only and LiDAR-camera combined voxels $\\tilde{f}^{C}$ and $\\tilde{f}^{LC}$ , we first transform their modality-specific representation into an intermediate joint space with 3D sparse convolutions. Then, within the joint space, we further combine all voxels and promote their interaction through another 3D sparse convolution as shown in Fig.4. The resulting multi-modal voxel features are denoted as $F^M$ .", + "bbox": [ + 496, + 321, + 892, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Multi-Scale Progressive Interaction", + "text_level": 1, + "bbox": [ + 498, + 479, + 802, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After performing MDU and GMA-Conv within different scales in the voxel space, multi-modal voxel features $\\{F_i^M\\}$ are generated for each scale. To further aggregate these multi-granularity information from $\\{F_i^M\\}$ , we introduce cascade connections across different scales as shown in Fig.2, which can be formulated as:", + "bbox": [ + 496, + 503, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {i + 1} ^ {M} = F _ {i + 1} ^ {M} + \\text {D o w n S a m p l e} \\left(\\hat {F} _ {i} ^ {M}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 606, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the voxel downsample operation DownSample() is applied to align their spatial resolutions, and $\\hat{F}_{i+1}^{M}$ is the resulting voxel features that combine multi-scale information from the current and previous scales.", + "bbox": [ + 496, + 638, + 890, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the above designs, multi-granularity LiDAR-camera features can be thoroughly interacted at multiple scales in the voxel space. The final output multi-modal voxel features can serve as a powerful representation for the subsequent modules.", + "bbox": [ + 496, + 699, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Prediction in BEV Space", + "text_level": 1, + "bbox": [ + 500, + 786, + 725, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the workflow of prevalent voxel-based 3D object detectors [22, 24, 29], we compress the height of voxel features to transform them into the BEV (bird's-eye view) space. Specifically, we first fuse the LiDAR and multimodal BEV features with a lightweight 2D convolution block. Then, the BEV feature containing enhanced multi", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21647", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e60cd27e88908c01c07fe4893a17bc118d9adb9aea046c7f0fde9cc169b2c7b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModalitymAPNDSCarTruckC.V.BusTrailerBarrierMotor.BikePed.T.C.
PointPillar [8]L40.155.076.031.011.332.136.656.434.214.064.045.6
CenterPoint [24]L60.367.385.253.520.063.656.071.159.530.784.678.4
TransFusion-L [1]L65.570.286.256.728.266.358.878.268.344.286.182.0
PointPainting [19]LC46.458.177.935.815.836.237.360.241.524.173.362.4
3D-CVF [26]LC52.762.383.045.015.948.849.665.951.230.474.262.9
PointAugmenting [20]LC66.871.087.557.328.065.260.772.674.350.987.983.6
MVP [25]LC66.470.586.858.526.167.457.374.870.049.389.185.0
TransFusion [1]LC68.971.787.160.033.168.360.878.173.652.988.486.7
VFF [10]LC68.472.486.858.132.170.261.073.978.552.987.183.8
BEVFusion [13]LC69.271.888.160.934.469.362.178.272.252.289.285.2
BEVFusion [15]LC70.272.988.660.139.369.863.880.074.151.089.286.5
UVTR [9]LC67.171.187.556.033.867.559.573.073.454.886.379.6
DeepInteraction [23]LC70.873.487.960.237.570.863.880.475.454.590.387.0
MSMDFusion (Ours)LC71.574.088.461.035.271.464.280.776.958.390.688.1
", + "bbox": [ + 81, + 88, + 893, + 314 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6fd593ed2264cf4f64064caf160348ecf92822b74438fbc415edf39aee5af12e.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art methods on nuScenes test set. Note that these are all single-model results without ensemble or test-time augmentation. We highlight the best performances across all methods with bold." + ], + "table_footnote": [], + "table_body": "
MethodModalityAMOTA ↑AMOTP ↓TP ↑FP ↓FN ↓Recall ↑
CenterPoint [24]L63.80.60695877186122292867.5
TransFusion [1]LC71.80.55196775162322184675.8
UVTR [9]LC70.10.68698434156152019075.0
BEVFusion [15]*LC74.10.40399664199971939577.9
MSMDFusion (Ours)LC74.00.54998624147891985376.3
", + "bbox": [ + 181, + 364, + 794, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with leading detectors that use a greedy tracker on nuScenes tracking task. * represents that BEVFusion uses an ensemble model, while we only use a single model. We use **bold** to denote the best performances among all approaches with a single model.", + "bbox": [ + 75, + 484, + 892, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "modality information is fed into a conventional BEV encoder and a detection head for the final prediction.", + "bbox": [ + 75, + 542, + 470, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 588, + 209, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 613, + 267, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Dataset and Metrics", + "text_level": 1, + "bbox": [ + 76, + 637, + 272, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The nuScenes [2] dataset is a large-scale autonomous driving benchmark including 10,000 driving scenarios in total, which are split into 700, 150, and 150 scenes for training, validation, and testing, respectively. For detection, nuScenes defines a set of evaluation protocols, including the nuScenes Detection Score (NDS), mean Average Precision (mAP), as well as five True Positive (TP) metrics, namely mean Average Translation Error (mATE), mean Average Scale Error (mASE), mean Average Orientation Error (mAOE), mean Average Velocity Error (mAVE) and mean Average Attribute Error (mAAE). We report mAP, which is the mean of the average precision across ten classes under distance thresholds of $0.5\\mathrm{m}$ , $1\\mathrm{m}$ , $2\\mathrm{m}$ , $4\\mathrm{m}$ . NDS is the weighted combination of mAP, mATE, mASE, mAOE, mAVE and mAAE.", + "bbox": [ + 75, + 657, + 468, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details We use ResNet-50 [5] with", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FPN [14] as the image backbone and VoxelNet [22] as the LiDAR backbone. We set the image size to $448 \\times 800$ , and voxel size as $(0.075m, 0.075m, 0.2m)$ following [1]. To make the best use of multi-scale semantics, we extract features from four levels of the FPN (C2 to C5). Following MVP [25], we use CenterNet2 [28] as 2D detector when generating virtual points, and 50 seeds are selected on each instance unless otherwise specified. Our model training has two stages: (1) We first train a LiDAR-only detector for 20 epochs as our 3D backbone. (2) We then connect the proposed LiDAR-camera fusion modules with the 3D backbone for a joint training of another 6 epochs. The data augmentation strategies and training schedules are the same as prior works [24, 30]. We do not use Test-Time Augmentation (TTA) or multi-model ensemble during inference.", + "bbox": [ + 496, + 542, + 890, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 781, + 794, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our MSMDFusion with state-of-the-art approaches on the nuScenes test set. Overall, Table 1 shows that our method surpasses all existing methods, and", + "bbox": [ + 496, + 804, + 892, + 851 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}$ NVPF of BEVFusion [13] is calculated with its official code. BEVFusion [15] reports its NVPF in the paper. The former has a higher NVPF because it generates virtual points from image features of larger resolution.", + "bbox": [ + 498, + 862, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21648", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e80981cabae98f100d1df3183d9be0c9bb008da38982ae659c894bdb32555a20.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNVPF2mAPNDSFPS
BEVFusion [13]5M69.271.80.7
BEVFusion [15]2M70.272.98.4
MSMDFusion(Ours)16k71.574.02.1
", + "bbox": [ + 102, + 88, + 446, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves the new state-of-the-art performances of $71.5\\mathrm{mAP}$ and 74.0 NDS. Meanwhile, our method maintains consistent performance advantages on most object categories, especially on the challenging category: Bike, where a $3.5\\%$ absolute AP gain over the strongest competitor is achieved.", + "bbox": [ + 75, + 243, + 467, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our MSMDFusion and two strong BEVFusion methods all fuse LiDAR and camera signals via generating 3D virtual points from 2D seeds, we specifically compare our method with them in terms of the number of generated virtual points per LiDAR frame as shown in Table 3. The results suggest that although with 100 times fewer generated virtual points (16k vs 2M/5M) than them, our MSMDFusion still outperforms them in terms of both mAP and NDS, which proves that our method can better utilize 2D semantics. Meanwhile, we also compare the FPS of our method and two BEVFusion approaches. BEVFusion [15] can achieve the fastest inference speed due to its efficient camera-to-BEV transformation implemented with CUDA. Without such a hardware-friendly implementation, BEVFusion [13] suffers from processing a huge amount of virtual points, while our MSMDFusion runs much faster although performing multi-scale fine-grained fusion in the voxel space.", + "bbox": [ + 75, + 319, + 467, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, to demonstrate that our MSMDFusion can also generalize to other downstream tasks, we further perform the tracking task on nuScenes. For fair comparison, we only compare with single-model methods (i.e., without TTA or ensemble) using the greedy tracker introduced in [24]. As shown in Table 2, our method consistent outperforms other single model competitors across all evaluation metrics. Meanwhile, although without using ensemble techniques, our method also achieve competitive results with the ensemble BEVFusion [15] method.", + "bbox": [ + 75, + 592, + 467, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comprehensive Analysis", + "text_level": 1, + "bbox": [ + 76, + 753, + 302, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following UVTR [9], we conduct the following ablation studies on a randomly sampled 1/4 split of nuScenes training set for efficiency.", + "bbox": [ + 76, + 777, + 467, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Ablation of proposed components", + "text_level": 1, + "bbox": [ + 76, + 845, + 364, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies for each of our proposed components as shown in Table 4. We use", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d731b2edd95c5de3b7cc551489b8180c81c200ad684eef1813dadbe67f8fb307.jpg", + "table_caption": [ + "Table 3. Number of Virtual points Per Frame (NVPF) and performance comparison with two strongest methods on nuScenes test set. The FPS is tested on a single NVIDIA RTX 3090 GPU." + ], + "table_footnote": [], + "table_body": "
#MDU*MDUGMA-ConvmAPNDS
160.2665.62
2✓64.3367.03
3✓65.5968.04
4✓✓65.9768.24
5✓✓66.9368.93
", + "bbox": [ + 532, + 88, + 862, + 184 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c46d4e4c858e7765dbba96732f6158e03cb6110d77790a1d78cdff749c4f7697.jpg", + "table_caption": [ + "Table 4. Ablation studies for our proposed components on the nuScenes validation set. MDU* is a degraded version of MDU by discarding the depth-aware feature generating process." + ], + "table_footnote": [], + "table_body": "
#C2C3C4C5mAPNDS
160.2665.62
2✓65.5167.92
3✓64.9967.97
4✓64.9767.66
5✓64.5267.61
6✓✓66.3468.63
7✓✓✓66.8068.86
8✓✓✓✓66.9368.93
", + "bbox": [ + 563, + 252, + 831, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Effects of LiDAR-camera interaction at different scales, C2-C5 represent the scales where the spatial resolutions of the features decrease from C2 to C5.", + "bbox": [ + 498, + 401, + 890, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TransFusion-L [1] as our baseline (#1), and perform multiscale fusion across all experiments.", + "bbox": [ + 498, + 460, + 888, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the results of the Table 4, we have the following observations. (i) Introducing virtual points for multi-modal interaction (#2 and #3) brings evident improvements over the baseline, which proves that fusing dense 2D semantics from virtual points with LiDAR features in the voxel space can significantly enhance the LiDAR-only detector. (ii) An early depth-color interaction (#2 and #3) can provide stronger semantics, which further boosts the benefits of interaction between LiDAR and virtual points. (iii) Using GMA-Conv to select helpful information from virtual points (#4 and #5) can bring further improvements over absorbing all information from virtual points without distinction (#2 and #3). (iv) With all these components combined (#5), the final mAP and NDS are significantly improved from $60.26\\%$ to $66.93\\%$ and $65.62\\%$ to $68.93\\%$ , respectively, which proves the effectiveness of our designs.", + "bbox": [ + 496, + 491, + 890, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Effects of multi-scale interaction", + "text_level": 1, + "bbox": [ + 500, + 755, + 779, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also investigate the effects of performing LiDAR-camera interaction at different scales. As shown in Table 5, for models with single-scale interaction (#2-#4), interacting at scales of larger spatial resolution generally leads to larger performance gains over the baseline (#1), which mainly benefits from more fine-grained image and LiDAR features. If the multi-modal interaction is performed at more scales (#6-#8), multi-granularity information can be", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21649", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ca241856c47795e9bbfad59dbfd78f6131980b457dfde1cdf93af3b9213e3db.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KNNVPFmAPNDS
1502.6K65.8668.24
3508.0K66.4968.88
65016.0K66.9368.93
105026.6K66.6168.85
120010.6K65.5968.00
", + "bbox": [ + 158, + 88, + 393, + 185 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Effects of the number of depths per seed (K) and number of seeds per instance (N) in multi-depth unprojection. NVPF represents the number of virtual points per frame.", + "bbox": [ + 75, + 195, + 470, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "integrated into our progressive interaction process to consistently achieve better performances than single-scale models. The complete multi-scale interaction model achieves performance gains of $\\sim 2\\%$ mAP and $\\sim 1\\%$ NDS over the best single-scale model, which indicates that multi-scale interaction is beneficial for comprehensive multi-modal fusion.", + "bbox": [ + 75, + 262, + 468, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9de000808512cc8e635f2bab6019cb6912ceaa23482df945daaa1c9729196f9b.jpg", + "image_caption": [ + "Figure 5. Mean error and recall of the virtual points computed with different numbers (K) of retrieved depths per seed." + ], + "image_footnote": [], + "bbox": [ + 81, + 380, + 464, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3 Discussion on MDU", + "text_level": 1, + "bbox": [ + 76, + 588, + 269, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed MDU is a simple yet effective strategy to lift 2D seeds from image to 3D space, and we would like to further discuss one critical problem: will the multiple depths for a seed introduce geometric noise and diminish the benefits? We conducted a quantitative experiment on the nuScenes dataset to investigate this problem. Recall that original point cloud, their projection on image plane (i.e., reference points) and seeds are denoted as $\\mathcal{P}$ , $\\mathcal{R}$ and $\\mathcal{S}$ , respectively. We fist preserve a subset $\\mathcal{R}^e = \\{(u_i^e,v_i^e,d_i^e)\\}_{i=1}^{N^e}$ of the reference points $\\mathcal{R}$ , and use all $\\{(u_i^e,v_i^e)\\}$ as seeds to generate virtual points $\\hat{\\mathcal{P}}^e$ by retrieving real depths from the remaining reference points $\\mathcal{R} - \\mathcal{R}^e$ as done in MDU. Since the real 3D coordinates $\\mathcal{P}^e$ of the preserved reference points $\\mathcal{R}^e$ are available, we then measure two important metrics for the generated virtual points $\\hat{\\mathcal{P}}^e$ : (1) The mean distance error in 3D space between $\\hat{\\mathcal{P}}^e$ and $\\mathcal{P}^e$ , which reflects the overall accuracy of the virtual points' coordinates. (2) The recall rate that is the ratio of points in $\\mathcal{P}^e$ which fall within a small neighborhood of at", + "bbox": [ + 75, + 611, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "least one point in $\\hat{\\mathcal{P}}^e$ , which represents MDU's ability to capture unseen real 3D points. We also vary the number of depths per seed (K) to show its effect. As shown in Fig.5, when the number of K increases from 1 to 6, the recall rate of real points improves dramatically, meanwhile, the mean distance error does not increase significantly. But a too large value of K (e.g., 10) can inevitably introduce noise. This proves that with a properly chosen number of depths, MDU does not lead to increased noise in depth estimation and can generate virtual points with reliable depths while also covering more real points than a single-depth strategy.", + "bbox": [ + 496, + 90, + 890, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further inspect the effects of K on the final detection performance, we keep other settings unchanged and vary K to compare performances. As Table 6 shows, increasing K from 1 to 6 can gradually improve these models' performances. As we previously analyzed in Fig.5, such improvements can be mainly attributed to the increased recall and tolerable distance error of the virtual points. However, when K increases to 10, the model's performance drops because the extra noise diminishes the benefits of correctly generated virtual points. Moreover, we also study whether increasing the number of seeds per instance N can achieve similar improvements, and the results when K is set to 1 and N is set to 200 are shown in the last row of Table 6. It can be seen that simply increasing seeds cannot bring improvements as our multi-depth strategy, but can even hurt performance due to more noise from single-depth unprojection.", + "bbox": [ + 496, + 262, + 892, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 550, + 617, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce MSMDFusion, a novel LiDAR-camera fusion framework for 3D object detection, which facilitates multi-modal interaction across multiple scales. Our approach employs a Multi-Depth Unprojection (MDU) strategy to obtain reliable virtual points from images, and applies Gated Modality-Aware Convolution (GMA-Conv) at each scale to enable fine-grained, controllable multi-modal LiDAR-camera fusion. We further combine multi-granularity information across scales to generate comprehensive features for the final prediction. Extensive experiments demonstrate the effectiveness of these components, with our method ultimately achieving state-of-the-art performance on the nuScenes dataset.", + "bbox": [ + 496, + 580, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 809, + 687, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project was supported by National Key R&D Program of China (No. 2020AAA0140001), NSFC (No. 62072116), and Science and Technology Commission of Shanghai Municipality (No. 21JC1400600).", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21650", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 1, 3, 6, 7", + "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6", + "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1, 2", + "[4] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 3", + "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[6] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In European Conference on Computer Vision, pages 35-52. Springer, 2020. 2, 3", + "[7] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. 2", + "[8] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 6", + "[9] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. 1, 2, 3, 6, 7", + "[10] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1120-1129, 2022. 1, 2, 6", + "[11] Ming Liang, Bin Yang, Yun Chen, Rui Hu, and Raquel Urtasun. Multi-task multi-sensor fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7345-7353, 2019. 1", + "[12] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 1, 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 1, 2, 3, 4, 6, 7", + "[14] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 6", + "[15] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 1, 2, 3, 4, 6, 7", + "[16] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 4", + "[17] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 1", + "[18] Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA), pages 7276-7282. IEEE, 2019. 1", + "[19] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 1, 2, 6", + "[20] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 1, 2, 3, 6", + "[21] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 4", + "[22] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 5, 6", + "[23] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. arXiv preprint arXiv:2208.11112, 2022. 2, 3, 6", + "[24] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2, 5, 6, 7", + "[25] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. 1, 2, 3, 4, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21651", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In European Conference on Computer Vision, pages 720-736. Springer, 2020. 1, 2, 3, 6", + "[27] Zhijie Zhang, Yan Liu, Junjie Chen, Li Niu, and Liqing Zhang. Depth privileged object detection in indoor scenes via deformation hallucination. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3456-3464, 2021. 2", + "[28] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 6", + "[29] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1, 5", + "[30] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.6" + ], + "bbox": [ + 78, + 90, + 470, + 386 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21652", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_model.json b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be9bf615cc90501d2ae4ad7af77edede0efc64c2 --- /dev/null +++ b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_model.json @@ -0,0 +1,1738 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.13, + 0.823, + 0.177 + ], + "angle": 0, + "content": "MSMDFusion: Fusing LiDAR and Camera at Multiple Scales with Multi-Depth Seeds for 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.202, + 0.878, + 0.221 + ], + "angle": 0, + "content": "Yang Jiao\\(^{1,2*}\\) Zequn Jie\\(^{3*}\\) Shaoxiang Chen\\(^{3}\\) Jingjing Chen\\(^{1,2\\dagger}\\) Lin Ma\\(^{3\\dagger}\\) Yu-Gang Jiang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.221, + 0.798, + 0.239 + ], + "angle": 0, + "content": "1 Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.239, + 0.785, + 0.273 + ], + "angle": 0, + "content": "\\(^{2}\\) Shanghai Collaborative Innovation Center of Intelligent Visual Computing \n\\(^{3}\\)Meituan" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.474, + 0.844 + ], + "angle": 0, + "content": "Fusing LiDAR and camera information is essential for accurate and reliable 3D object detection in autonomous driving systems. This is challenging due to the difficulty of combining multi-granularity geometric and semantic features from two drastically different modalities. Recent approaches aim at exploring the semantic densities of camera features through lifting points in 2D camera images (referred to as \"seeds\") into 3D space, and then incorporate 2D semantics via cross-modal interaction or fusion techniques. However, depth information is under-investigated in these approaches when lifting points into 3D space, thus 2D semantics can not be reliably fused with 3D points. Moreover, their multi-modal fusion strategy, which is implemented as concatenation or attention, either can not effectively fuse 2D and 3D information or is unable to perform fine-grained interactions in the voxel space. To this end, we propose a novel framework with better utilization of the depth information and fine-grained cross-modal interaction between LiDAR and camera, which consists of two important components. First, a Multi-Depth Unprojection (MDU) method is used to enhance the depth quality of the lifted points at each interaction level. Second, a Gated Modality-Aware Convolution (GMA-Conv) block is applied to modulate voxels involved with the camera modality in a fine-grained manner and then aggregate multi-modal features into a unified space. Together they provide the detection head with more comprehensive features from LiDAR and camera. On the nuScenes test benchmark, our proposed method, abbreviated as MSMD-Fusion, achieves state-of-the-art results on both 3D object detection and tracking tasks without using test-time-augmentation and ensemble techniques. The code is available at https://github.com/SxJyJay/MSMDFusion." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.31, + 0.895, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.456, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Figure 1. Illustration of our MSMDFusion pipeline. The yellow arrows indicate information passing or interaction between the Li-DAR and camera modalities." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.526, + 0.633, + 0.542 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Detecting 3D objects [1, 22, 29] is regarded as a fundamental task for autonomous driving. Aiming at the robust environmental perception, LiDARs and cameras are widely equipped on autonomous driving vehicles since they can provide complementary information. Characterized by point clouds, LiDARs can capture accurate spatial information, while cameras contain rich semantics and context with images. Therefore, developing multi-modal detectors that enjoy the benefits of the two worlds is promising. Such an idea has catalyzed the emergence of a set of recent researches [1, 3, 12, 13, 15, 17, 19, 20, 25, 26]. Early works [1, 3, 11, 12, 18-20, 26] perform LiDAR-camera fusion by projecting 3D LiDAR points (or region proposals generated from them) onto 2D image planes to collect useful 2D semantics. However, such a paradigm suffers from the signal density mismatching of multi-modality sensors. Since LiDAR points are much sparser than camera pixels, this projection manner will inevitably waste semantically rich 2D features [1, 13]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently, another paradigm [9,10,13,15,25] for LiDAR-camera fusion has emerged. Instead of collecting 2D semantic features with 3D queries, these methods first estimate depths of pixels, and then directly lift 2D pixels with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.227, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21643" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "their semantics to the 3D world (we refer to these pixels and corresponding lifted points as \"seeds\" and \"virtual points\" in this paper) to fuse with the real 3D point cloud. Two methods with the same name of BEVFusion [13, 15] treat every image feature pixel as a seed and generate virtual points in the BEV space. MVP [25] and VFF [10] sample pixels from foreground regions and lift them to the voxel space. Benefited from the dense virtual points, such a paradigm not only maintains the semantic consistency in the image [15], but also complements the geometric cues for the sparse LiDAR point cloud [25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.534 + ], + "angle": 0, + "content": "Despite significant improvements have been made, existing methods in this line suffer from two major problems, which hampers benefiting from virtual points. First, depth, as the key to the quality of virtual points, is underinvestigated in generating virtual points. On the one hand, depth directly determines the spatial location in 3D space of a seed via perspective projection which can thereby significantly affect 3D object detection results. On the other hand, depth can also enhance the semantics carried by virtual points by providing color-insensitive cues in describing objects [27], since combining RGB information with depth guidance correlates camera pixels of similar depths and enables them to jointly contribute to capturing instance-related semantics when lifted as virtual points. Existing multi-modal detectors [9, 13, 15] mainly pay attention on interacting LiDAR points with camera virtual points, while ignoring the importance of seed depths in generating the virtual points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.779 + ], + "angle": 0, + "content": "Second, the fine-grained cross-modal interaction between virtual points and 3D points in the uncompressed space (e.g., the voxel space) is crucial but non-trivial. Generated virtual points are geometrically and semantically inconsistent with real LiDAR points due to imperfect depths and inherent modality gap. Hence, in order to benefit from the semantically rich virtual points, it is necessary to adaptively select helpful information from virtual points under the guidance of real LiDAR points in a fine-grained and controllable manner. However, such a cross-modal interaction is constrained by the intensive memory and computation cost brought by the massive amounts and unstructured nature of point cloud data. Alternatively, existing approaches combine the multi-modal information with simple concatenate [25] or add operations [9] in the voxel space, or perform cross-attention in a compressed BEV space [23]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Aiming at unlocking the potential of virtual points and addressing the drawbacks of existing methods, we propose a multi-scale fusion framework, called MSMDFusion, and within each scale, there are two key novel designs, namely the Multi-Depth Unprojection (MDU) and Gated Modality-Aware Convolution (GMA-Conv). As shown in Fig.1, MDU is mainly for enhancing the quality of generated virtual points in terms of geometric accuracy and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "semantic richness. To be specific, when lifting 2D seeds from image into the 3D space, multiple depths are explored within a reference neighborhood to generate virtual points with more reliable depth. Next, the camera feature and depth are combined to produce depth-aware features as stronger 2D semantics to decorate these virtual points. GMA-Conv takes real LiDAR points and generated virtual points as inputs, and performs fine-grained interaction in a select-then-aggregate manner. Concretely, we first adaptively select useful information from camera voxel features under the guidance of reference LiDAR voxels, then aggregate them grouped sparse convolutions for sufficient multimodal interaction. We also specifically adopt a voxel subsampling strategy to efficiently obtain reliable LiDAR references when implementing our GMA-Conv." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.485 + ], + "angle": 0, + "content": "Finally, with the resulting multi-modal voxel features from multiple scales, we further associate them with cascade connections across scales to aggregate multi-granularity information. With the above designs, the camera semantics encapsulated in the virtual points are consistently combined with LiDAR points, and thereby providing a stronger multi-modal feature representation for boosting the 3D object detection. As shown in the Table 3, with 100 times fewer generated virtual points than the two BEVFusion methods [13, 15], our MSMDFusion can still achieve state-of-the-art performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.896, + 0.729 + ], + "angle": 0, + "content": "In summary, our contributions lie in threefold: (1) We propose a novel MSMDFusion approach, which encourages sufficient LiDAR-Camera feature fusion in the multiscale voxel space. (2) Within each scale, we propose a Multi-Depth Unprojection strategy (MDU) to promote virtual points generation with better locations and semantics by fully leveraging depth of pixels, as well as a Gated Modality-Aware Convolution (GMA-Conv) to achieve fine-grained controllable multi-modal interaction. (3) Extensive experimental results on the large-scale nuScenes dataset prove the effectiveness of our MSMDFusion and its components. We achieve state-of-the-art performances with \\(71.5\\%\\) mAP and \\(74.0\\%\\) NDS on the challenging nuScenes detection track using single model1. When combining the simple greedy tracking strategy [24], our method also achieves strong tracking results with \\(74.0\\%\\) AMOTA." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.745, + 0.642, + 0.761 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.895, + 0.864 + ], + "angle": 0, + "content": "Recently, fusing LiDAR and camera signals in 3D detection has attracted increasing interest. Early works [3, 6, 12, 19, 20, 26] project 3D queries to 2D camera images for collecting useful semantics. Under such a paradigm, MV3D [3] and AVOD [7] associate 3D proposals with 2D RoI features. While PointPainting [19] and PointAugment-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1We do not use Test-Time Augmentation (TTA) and model ensemble technique." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21644" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.381, + 0.897, + 0.466 + ], + "angle": 0, + "content": "Figure 2. The framework of our proposed MSMDFusion. Given a LiDAR point cloud and corresponding multi-view camera images, MSMDFusion first separately transforms them into the voxel space and obtains multi-scale LiDAR and camera features. Then, the LiDAR and camera features at each scale interact through our proposed Multi-Depth Unprojection strategy (MDU) and Gated Modality-Aware Convolution block (GMA-Conv, details in Fig.4). The outputs of GMA-Conv at each scale are further aggregated for combining multi-granularity information. Finally, the resulting multi-modal and LiDAR voxel features are flattened into the BEV space for the final prediction. \\(\\oplus\\) represents voxel addition and \"spconv\" indicates 3D sparse convolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.473, + 0.87 + ], + "angle": 0, + "content": "ing [20] directly decorate raw 3D points with 2D semantics. EP-Net [6] and 3D-CVF [26] perform multi-modal fusion at both point level and proposal level. However, since 3D points are inherently sparse, such a hard association approach wastes the dense semantic information in 2D features. Recently, multi-modal 3D detectors [13, 15, 23, 25] lift dense 2D seeds to 3D space for learning the 2D-3D joint representation in a shared space. Two BEVFusion methods [13, 15] densely lift every image feature pixel into 3D space and then encode these lifted points as another BEV (bird's-eye view) map to fuse with the BEV map obtained using the original 3D point cloud. LiDAR and camera signals will finally share a common representation space, but they do not interact at any stage before BEV fusion. MVP [25] selects seeds from foreground objects and unprojected them into the LiDAR frame for points cloud augmentation. With the seeds sampling strategy similar to MVP, UVTR [9] constructs both LiDAR and camera branches in the voxel space to maintain the modality-specific information, and then interact them with voxel add operation. However, the LiDAR and camera branches are constructed in a single scale of voxel space, losing the benefits of different levels of feature abstraction in multi-scale voxel space. Moreover, merging LiDAR and camera voxels with simple addition limits their fine-grained interaction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.49, + 0.593, + 0.506 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.707, + 0.53 + ], + "angle": 0, + "content": "3.1. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.795 + ], + "angle": 0, + "content": "An overall view of MSMDFusion is shown in Fig.2. Given a LiDAR point cloud and corresponding multiview camera images as inputs, MSMDFusion first extracts multi-scale features from both modalities in the voxel space. Then, LiDAR-camera interaction is performed within multi-scale voxel space to properly combine multi-granularity information from both modalities. At each scale, we specifically design a Multi-Depth Unprojection (MDU) strategy to obtain high-quality virtual points in the voxel space, and a Gated Modality-Aware Convolution block (GMA-Conv) for effective LiDAR-camera interaction and fusion. We also introduce cross-scale connections to progressively combine features of different granularities. Afterward, the deeply interacted multi-modal features, together with the LiDAR features, are transformed into the BEV space and fed to the BEV encoder and detection head for the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.842, + 0.817 + ], + "angle": 0, + "content": "3.2. LiDAR and Camera Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We first extract high-level features from the raw inputs of LiDAR and camera. For a given LiDAR point cloud, we voxelize the points and extract their features in the voxel space using a set of 3D sparse convolutional blocks [4] following prior arts [1, 13, 15]. Each convolutional block out" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21645" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.094, + 0.273, + 0.105 + ], + "angle": 0, + "content": "large depth bias moderate depth bias" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.105, + 0.254, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.198, + 0.239, + 0.208 + ], + "angle": 0, + "content": "(a) Nearest neighbor retrieval" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.095, + 0.445, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.199, + 0.462, + 0.21 + ], + "angle": 0, + "content": "(b) K Nearest neighbor retrieval (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.47, + 0.308 + ], + "angle": 0, + "content": "Figure 3. Comparison of the effects of retrieving reference points from nearest neighbor (MVP) and K-nearest neighbor (Ours). The blue circle and yellow star represent seeds and reference points, respectively. To keep the figure concise, we only depict the case that \\( \\mathrm{K} = 2 \\). Our strategy can obtain more reliable depths by exploring more neighbors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.468, + 0.396 + ], + "angle": 0, + "content": "puts voxel features of different scales, representing different levels of abstraction for point cloud, and these multiscale features will further interact with their 2D counterparts through our proposed GMA-Conv blocks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.468, + 0.563 + ], + "angle": 0, + "content": "For the multi-view camera images, we adopt ResNet50 with FPN as the image backbone to extract multi-scale image features, which contains rich object semantics. However, it is non-trivial to transfer these useful semantics to the 3D voxel field, since images are inherently 2D data. Motivated by recent works [13, 15, 25] which select seeds from the image plane and estimate their depths to lift them as 3D virtual points, we further propose a Multi-Depth Unprojection strategy to lift 2D seeds, which could mitigate the drawbacks of prior works. The details will be presented in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.418, + 0.59 + ], + "angle": 0, + "content": "3.3. Multi-Modal Interaction in Voxel Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.468, + 0.81 + ], + "angle": 0, + "content": "The extracted LiDAR and camera features from the previous stage respectively focus on geometric and semantic information, hence the goal of multi-modal interaction is to properly fuse these features into a unified rich representation of the scene. As shown in the Fig.2, the components of LiDAR-camera interaction within each scale in the voxel space are conceptually simple: 1) Multi-Depth Unprojection (MDU) strategy is introduced to improve the quality of the generated virtual points and 2) Gated Modality-Aware Convolution (GMA-Conv) block is designed for controllable LiDAR-camera fusion. In the following parts, we will first elaborate on MDU and GMA-Conv within a specific scale, then on top of them, we describe the complete multiscale fusion process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.315, + 0.846 + ], + "angle": 0, + "content": "3.3.1 Multi-Depth Unprojection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To lift a 2D seed in the pixel space to the 3D space (i.e., the unprojection operation), the depth associated with each seed should be estimated [21]. To obtain an initial reli" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "able depth estimation, we use the strategy from MVP [25] as the basis as shown in Fig.3(a). Formally, given a point cloud \\(\\mathcal{P} = \\{(x_i,y_i,z_i)\\}_{i = 1}^{N_P}\\) and multi-view images as inputs, MVP first projects 3D points onto 2D images and preserves points falling within the 2D foreground instance masks. These 3D points are kept as reference points to provide depth, and they are denoted as \\(\\mathcal{R} = \\{(u_i,v_i,d_i)\\}_{i = 1}^{N_R}\\), where \\(u_{i}\\) and \\(v_{i}\\) are pixel coordinates and \\(d_{i}\\) is the real depth. Then, a set of seeds \\(S = \\{(u_i^s,v_i^s)\\}_{i = 1}^{N_S}\\) are uniformly sampled from each instance mask on the image, and each of them will retrieve a real depth from its nearest reference point as its estimated depth. Finally the sampled seeds with estimated depths are unprojected to 3D as virtual points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Though effective, such a strategy ignores the fact that spatial proximity in 2D images can not be guaranteed in 3D, and this may lead to inaccurate depth estimation as shown in the red circle of Fig.3(a). Toward this end, we propose to equip every seed with multiple depths by retrieving K-nearest reference points, which can be regarded as a soft strategy to achieve more reliable depth estimation. As shown in the green circle of the Fig.3(b), each seed can be unprojected using multiple (K) depths from neighboring reference points, which generates K virtual points to improve its recall of the actual 3D points. We provide more detailed discussion on the effect of multi-depth seeds in Sec.4.3.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.787 + ], + "angle": 0, + "content": "These generated virtual points will be further decorated to make the best use of semantics encapsulated in images. But different from MVP [25] that simply decorates virtual points with class scores predicted by a pretrained 2D detector, we make an early interaction between image features and sparse depth information from reference points to generate depth-aware semantic features. These depth-aware semantic features are then used to decorate virtual points in an adaptive manner controlled by their corresponding seed depths. The whole process can be trained end-to-end. Concretely, we concatenate the camera image feature \\( C \\) with a sparse depth map containing all depths of reference points \\( \\mathcal{R} \\) in that image, and then fuse them with a convolutional layer to obtain the depth-aware semantic feature \\( C^d \\). For the K virtual points generated by the same seed, the semantic feature of the corresponding pixel should also contribute differently along the projection ray in the 3D space [16]. Thus we compute a dynamic weight factor for each individual depth to modulate the semantic feature. The process for a specific seed \\( (u_i^s, v_i^s) \\) can be formally described as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.792, + 0.892, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s _ {i} ^ {k} = \\operatorname {S i g m o i d} \\left(\\operatorname {L i n e a r} \\left([ C ^ {d} [ u _ {i} ^ {s}, v _ {i} ^ {s} ]; d _ {i} ^ {s, k} ]\\right)\\right) \\tag {1} \\\\ c _ {i} ^ {k} = C ^ {d} [ u _ {i} ^ {s}, v _ {i} ^ {s} ] * s _ {i} ^ {k} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where \\( d_{i}^{s,k} \\) is the \\( k \\)-th estimated depth of seed \\( (u_{i}^{s}, v_{i}^{s}) \\), \\( s_{i}^{k} \\in [0,1] \\) represents its weight factor, and \\( c^{k} \\) is the modulated semantic feature, which will be used to decorate the corresponding unprojected virtual point \\( (u_{i}^{s}, v_{i}^{s}, d_{i}^{s,k}) \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21646" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.089, + 0.812, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.895, + 0.252 + ], + "angle": 0, + "content": "Figure 4. Illustration of the select-then-aggregate process in the Gated Modality-Aware Convolution. Colors indicate different modalities: yellow for camera, blue for LiDAR, and red for LiDAR-camera combined. Cubes with shadows represent voxels selected with gates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.471, + 0.338 + ], + "angle": 0, + "content": "The resulting virtual points with decorated features will be voxelized to match the output resolution of the corresponding scale of LiDAR branch, in order to perform interaction in the voxel space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.382, + 0.374 + ], + "angle": 0, + "content": "3.3.2 Gated Modality-Aware Convolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.549 + ], + "angle": 0, + "content": "With the resulting LiDAR and camera voxel features that share the same spatial resolution (i.e., the vertically aligned blue and yellow voxels in Fig.2), we aim to interact them in a fine-grained and controllable manner with our proposed Gated Modality-Aware Convolution (GMA-Conv) block in each scale. As shown in Fig.4, we first group the voxels according to their modalities: LiDAR-only (blue voxels), camera-only (yellow voxels), as well as LiDAR and camera combined (red voxels), and they are denoted as \\( f^L \\), \\( f^C \\), and \\( f^{LC} \\), respectively. Then, more fine-grained multi-modal interaction is conducted in a select-then-aggregate manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Select. Motivated by the fact that LiDAR-based detectors usually surpass their camera-based counterparts with a large margin, we take LiDAR as the guiding modality to select useful information from camera features. Specifically, we design a gate conditioned on the LiDAR feature to control and update its camera counterpart, which can be formally described as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.668, + 0.47, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\tilde {f _ {i}} ^ {C} = \\operatorname {R e L U} \\left(\\operatorname {L i n e a r} \\left(f _ {j} ^ {L}\\right)\\right) * f _ {i} ^ {C} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\( \\tilde{f_i^C} \\) is the updated camera voxel feature, and \\( i \\) and \\( j \\) are voxel indices of paired camera and LiDAR voxel features (\\( f_j^L \\) can be regarded as the reference voxel for \\( f_i^C \\)). The key to make this selection process effective is how to efficiently find reliable reference voxels. A straightforward way is to retrieve the nearest LiDAR voxels for each camera-only voxel \\( f_i^C \\) from \\( f^L \\). And similarly for LiDAR-camera voxels \\( f_k^{LC} \\), the reference can be retrieved from both the LiDAR and camera modalities (i.e., \\( f^L \\) and \\( f^C \\)). Although nearby neighbors from more reliable modalities can serve as a good reference, such a retrieval process is infeasible to be implemented with exhaustive pairwise distance calculation, because the computation and memory" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.321 + ], + "angle": 0, + "content": "cost brought by massive amounts of LiDAR and camera voxels is intractable, and we will give more analysis and our solution in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.893, + 0.458 + ], + "angle": 0, + "content": "Aggregate. With the original LiDAR-only voxels \\( f^L \\), as well as the updated camera-only and LiDAR-camera combined voxels \\( \\tilde{f}^{C} \\) and \\( \\tilde{f}^{LC} \\), we first transform their modality-specific representation into an intermediate joint space with 3D sparse convolutions. Then, within the joint space, we further combine all voxels and promote their interaction through another 3D sparse convolution as shown in Fig.4. The resulting multi-modal voxel features are denoted as \\( F^M \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.48, + 0.803, + 0.495 + ], + "angle": 0, + "content": "3.3.3 Multi-Scale Progressive Interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.596 + ], + "angle": 0, + "content": "After performing MDU and GMA-Conv within different scales in the voxel space, multi-modal voxel features \\(\\{F_i^M\\}\\) are generated for each scale. To further aggregate these multi-granularity information from \\(\\{F_i^M\\}\\), we introduce cascade connections across different scales as shown in Fig.2, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.607, + 0.892, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {i + 1} ^ {M} = F _ {i + 1} ^ {M} + \\text {D o w n S a m p l e} \\left(\\hat {F} _ {i} ^ {M}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.699 + ], + "angle": 0, + "content": "where the voxel downsample operation DownSample() is applied to align their spatial resolutions, and \\(\\hat{F}_{i+1}^{M}\\) is the resulting voxel features that combine multi-scale information from the current and previous scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.775 + ], + "angle": 0, + "content": "With the above designs, multi-granularity LiDAR-camera features can be thoroughly interacted at multiple scales in the voxel space. The final output multi-modal voxel features can serve as a powerful representation for the subsequent modules." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.727, + 0.803 + ], + "angle": 0, + "content": "3.4. Prediction in BEV Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Following the workflow of prevalent voxel-based 3D object detectors [22, 24, 29], we compress the height of voxel features to transform them into the BEV (bird's-eye view) space. Specifically, we first fuse the LiDAR and multimodal BEV features with a lightweight 2D convolution block. Then, the BEV feature containing enhanced multi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21647" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.315 + ], + "angle": 0, + "content": "
MethodModalitymAPNDSCarTruckC.V.BusTrailerBarrierMotor.BikePed.T.C.
PointPillar [8]L40.155.076.031.011.332.136.656.434.214.064.045.6
CenterPoint [24]L60.367.385.253.520.063.656.071.159.530.784.678.4
TransFusion-L [1]L65.570.286.256.728.266.358.878.268.344.286.182.0
PointPainting [19]LC46.458.177.935.815.836.237.360.241.524.173.362.4
3D-CVF [26]LC52.762.383.045.015.948.849.665.951.230.474.262.9
PointAugmenting [20]LC66.871.087.557.328.065.260.772.674.350.987.983.6
MVP [25]LC66.470.586.858.526.167.457.374.870.049.389.185.0
TransFusion [1]LC68.971.787.160.033.168.360.878.173.652.988.486.7
VFF [10]LC68.472.486.858.132.170.261.073.978.552.987.183.8
BEVFusion [13]LC69.271.888.160.934.469.362.178.272.252.289.285.2
BEVFusion [15]LC70.272.988.660.139.369.863.880.074.151.089.286.5
UVTR [9]LC67.171.187.556.033.867.559.573.073.454.886.379.6
DeepInteraction [23]LC70.873.487.960.237.570.863.880.475.454.590.387.0
MSMDFusion (Ours)LC71.574.088.461.035.271.464.280.776.958.390.688.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.324, + 0.893, + 0.355 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on nuScenes test set. Note that these are all single-model results without ensemble or test-time augmentation. We highlight the best performances across all methods with bold." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.366, + 0.795, + 0.476 + ], + "angle": 0, + "content": "
MethodModalityAMOTA ↑AMOTP ↓TP ↑FP ↓FN ↓Recall ↑
CenterPoint [24]L63.80.60695877186122292867.5
TransFusion [1]LC71.80.55196775162322184675.8
UVTR [9]LC70.10.68698434156152019075.0
BEVFusion [15]*LC74.10.40399664199971939577.9
MSMDFusion (Ours)LC74.00.54998624147891985376.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.485, + 0.893, + 0.528 + ], + "angle": 0, + "content": "Table 2. Comparison with leading detectors that use a greedy tracker on nuScenes tracking task. * represents that BEVFusion uses an ensemble model, while we only use a single model. We use **bold** to denote the best performances among all approaches with a single model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.471, + 0.575 + ], + "angle": 0, + "content": "modality information is fed into a conventional BEV encoder and a detection head for the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.21, + 0.606 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.614, + 0.268, + 0.631 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.638, + 0.273, + 0.652 + ], + "angle": 0, + "content": "4.1.1 Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.884 + ], + "angle": 0, + "content": "The nuScenes [2] dataset is a large-scale autonomous driving benchmark including 10,000 driving scenarios in total, which are split into 700, 150, and 150 scenes for training, validation, and testing, respectively. For detection, nuScenes defines a set of evaluation protocols, including the nuScenes Detection Score (NDS), mean Average Precision (mAP), as well as five True Positive (TP) metrics, namely mean Average Translation Error (mATE), mean Average Scale Error (mASE), mean Average Orientation Error (mAOE), mean Average Velocity Error (mAVE) and mean Average Attribute Error (mAAE). We report mAP, which is the mean of the average precision across ten classes under distance thresholds of \\(0.5\\mathrm{m}\\), \\(1\\mathrm{m}\\), \\(2\\mathrm{m}\\), \\(4\\mathrm{m}\\). NDS is the weighted combination of mAP, mATE, mASE, mAOE, mAVE and mAAE." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Implementation Details We use ResNet-50 [5] with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.771 + ], + "angle": 0, + "content": "FPN [14] as the image backbone and VoxelNet [22] as the LiDAR backbone. We set the image size to \\(448 \\times 800\\), and voxel size as \\((0.075m, 0.075m, 0.2m)\\) following [1]. To make the best use of multi-scale semantics, we extract features from four levels of the FPN (C2 to C5). Following MVP [25], we use CenterNet2 [28] as 2D detector when generating virtual points, and 50 seeds are selected on each instance unless otherwise specified. Our model training has two stages: (1) We first train a LiDAR-only detector for 20 epochs as our 3D backbone. (2) We then connect the proposed LiDAR-camera fusion modules with the 3D backbone for a joint training of another 6 epochs. The data augmentation strategies and training schedules are the same as prior works [24, 30]. We do not use Test-Time Augmentation (TTA) or multi-model ensemble during inference." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.782, + 0.795, + 0.798 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.805, + 0.893, + 0.852 + ], + "angle": 0, + "content": "We compare our MSMDFusion with state-of-the-art approaches on the nuScenes test set. Overall, Table 1 shows that our method surpasses all existing methods, and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.894, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)NVPF of BEVFusion [13] is calculated with its official code. BEVFusion [15] reports its NVPF in the paper. The former has a higher NVPF because it generates virtual points from image features of larger resolution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21648" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.447, + 0.164 + ], + "angle": 0, + "content": "
MethodNVPF2mAPNDSFPS
BEVFusion [13]5M69.271.80.7
BEVFusion [15]2M70.272.98.4
MSMDFusion(Ours)16k71.574.02.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.175, + 0.47, + 0.217 + ], + "angle": 0, + "content": "Table 3. Number of Virtual points Per Frame (NVPF) and performance comparison with two strongest methods on nuScenes test set. The FPS is tested on a single NVIDIA RTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.468, + 0.32 + ], + "angle": 0, + "content": "achieves the new state-of-the-art performances of \\(71.5\\mathrm{mAP}\\) and 74.0 NDS. Meanwhile, our method maintains consistent performance advantages on most object categories, especially on the challenging category: Bike, where a \\(3.5\\%\\) absolute AP gain over the strongest competitor is achieved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.591 + ], + "angle": 0, + "content": "Since our MSMDFusion and two strong BEVFusion methods all fuse LiDAR and camera signals via generating 3D virtual points from 2D seeds, we specifically compare our method with them in terms of the number of generated virtual points per LiDAR frame as shown in Table 3. The results suggest that although with 100 times fewer generated virtual points (16k vs 2M/5M) than them, our MSMDFusion still outperforms them in terms of both mAP and NDS, which proves that our method can better utilize 2D semantics. Meanwhile, we also compare the FPS of our method and two BEVFusion approaches. BEVFusion [15] can achieve the fastest inference speed due to its efficient camera-to-BEV transformation implemented with CUDA. Without such a hardware-friendly implementation, BEVFusion [13] suffers from processing a huge amount of virtual points, while our MSMDFusion runs much faster although performing multi-scale fine-grained fusion in the voxel space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.468, + 0.744 + ], + "angle": 0, + "content": "Besides, to demonstrate that our MSMDFusion can also generalize to other downstream tasks, we further perform the tracking task on nuScenes. For fair comparison, we only compare with single-model methods (i.e., without TTA or ensemble) using the greedy tracker introduced in [24]. As shown in Table 2, our method consistent outperforms other single model competitors across all evaluation metrics. Meanwhile, although without using ensemble techniques, our method also achieve competitive results with the ensemble BEVFusion [15] method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.303, + 0.772 + ], + "angle": 0, + "content": "4.3. Comprehensive Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.779, + 0.468, + 0.825 + ], + "angle": 0, + "content": "Following UVTR [9], we conduct the following ablation studies on a randomly sampled 1/4 split of nuScenes training set for efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.365, + 0.862 + ], + "angle": 0, + "content": "4.3.1 Ablation of proposed components" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies for each of our proposed components as shown in Table 4. We use" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.089, + 0.864, + 0.185 + ], + "angle": 0, + "content": "
#MDU*MDUGMA-ConvmAPNDS
160.2665.62
2✓64.3367.03
3✓65.5968.04
4✓✓65.9768.24
5✓✓66.9368.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.197, + 0.892, + 0.239 + ], + "angle": 0, + "content": "Table 4. Ablation studies for our proposed components on the nuScenes validation set. MDU* is a degraded version of MDU by discarding the depth-aware feature generating process." + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.253, + 0.832, + 0.391 + ], + "angle": 0, + "content": "
#C2C3C4C5mAPNDS
160.2665.62
2✓65.5167.92
3✓64.9967.97
4✓64.9767.66
5✓64.5267.61
6✓✓66.3468.63
7✓✓✓66.8068.86
8✓✓✓✓66.9368.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.402, + 0.892, + 0.443 + ], + "angle": 0, + "content": "Table 5. Effects of LiDAR-camera interaction at different scales, C2-C5 represent the scales where the spatial resolutions of the features decrease from C2 to C5." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.89, + 0.49 + ], + "angle": 0, + "content": "TransFusion-L [1] as our baseline (#1), and perform multiscale fusion across all experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.733 + ], + "angle": 0, + "content": "From the results of the Table 4, we have the following observations. (i) Introducing virtual points for multi-modal interaction (#2 and #3) brings evident improvements over the baseline, which proves that fusing dense 2D semantics from virtual points with LiDAR features in the voxel space can significantly enhance the LiDAR-only detector. (ii) An early depth-color interaction (#2 and #3) can provide stronger semantics, which further boosts the benefits of interaction between LiDAR and virtual points. (iii) Using GMA-Conv to select helpful information from virtual points (#4 and #5) can bring further improvements over absorbing all information from virtual points without distinction (#2 and #3). (iv) With all these components combined (#5), the final mAP and NDS are significantly improved from \\(60.26\\%\\) to \\(66.93\\%\\) and \\(65.62\\%\\) to \\(68.93\\%\\), respectively, which proves the effectiveness of our designs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.78, + 0.769 + ], + "angle": 0, + "content": "4.3.2 Effects of multi-scale interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also investigate the effects of performing LiDAR-camera interaction at different scales. As shown in Table 5, for models with single-scale interaction (#2-#4), interacting at scales of larger spatial resolution generally leads to larger performance gains over the baseline (#1), which mainly benefits from more fine-grained image and LiDAR features. If the multi-modal interaction is performed at more scales (#6-#8), multi-granularity information can be" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21649" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.394, + 0.186 + ], + "angle": 0, + "content": "
KNNVPFmAPNDS
1502.6K65.8668.24
3508.0K66.4968.88
65016.0K66.9368.93
105026.6K66.6168.85
120010.6K65.5968.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.196, + 0.471, + 0.239 + ], + "angle": 0, + "content": "Table 6. Effects of the number of depths per seed (K) and number of seeds per instance (N) in multi-depth unprojection. NVPF represents the number of virtual points per frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.368 + ], + "angle": 0, + "content": "integrated into our progressive interaction process to consistently achieve better performances than single-scale models. The complete multi-scale interaction model achieves performance gains of \\(\\sim 2\\%\\) mAP and \\(\\sim 1\\%\\) NDS over the best single-scale model, which indicates that multi-scale interaction is beneficial for comprehensive multi-modal fusion." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.381, + 0.465, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Figure 5. Mean error and recall of the virtual points computed with different numbers (K) of retrieved depths per seed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.27, + 0.604 + ], + "angle": 0, + "content": "4.3.3 Discussion on MDU" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our proposed MDU is a simple yet effective strategy to lift 2D seeds from image to 3D space, and we would like to further discuss one critical problem: will the multiple depths for a seed introduce geometric noise and diminish the benefits? We conducted a quantitative experiment on the nuScenes dataset to investigate this problem. Recall that original point cloud, their projection on image plane (i.e., reference points) and seeds are denoted as \\(\\mathcal{P}\\), \\(\\mathcal{R}\\) and \\(\\mathcal{S}\\), respectively. We fist preserve a subset \\(\\mathcal{R}^e = \\{(u_i^e,v_i^e,d_i^e)\\}_{i=1}^{N^e}\\) of the reference points \\(\\mathcal{R}\\), and use all \\(\\{(u_i^e,v_i^e)\\}\\) as seeds to generate virtual points \\(\\hat{\\mathcal{P}}^e\\) by retrieving real depths from the remaining reference points \\(\\mathcal{R} - \\mathcal{R}^e\\) as done in MDU. Since the real 3D coordinates \\(\\mathcal{P}^e\\) of the preserved reference points \\(\\mathcal{R}^e\\) are available, we then measure two important metrics for the generated virtual points \\(\\hat{\\mathcal{P}}^e\\): (1) The mean distance error in 3D space between \\(\\hat{\\mathcal{P}}^e\\) and \\(\\mathcal{P}^e\\), which reflects the overall accuracy of the virtual points' coordinates. (2) The recall rate that is the ratio of points in \\(\\mathcal{P}^e\\) which fall within a small neighborhood of at" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.259 + ], + "angle": 0, + "content": "least one point in \\(\\hat{\\mathcal{P}}^e\\), which represents MDU's ability to capture unseen real 3D points. We also vary the number of depths per seed (K) to show its effect. As shown in Fig.5, when the number of K increases from 1 to 6, the recall rate of real points improves dramatically, meanwhile, the mean distance error does not increase significantly. But a too large value of K (e.g., 10) can inevitably introduce noise. This proves that with a properly chosen number of depths, MDU does not lead to increased noise in depth estimation and can generate virtual points with reliable depths while also covering more real points than a single-depth strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.893, + 0.52 + ], + "angle": 0, + "content": "To further inspect the effects of K on the final detection performance, we keep other settings unchanged and vary K to compare performances. As Table 6 shows, increasing K from 1 to 6 can gradually improve these models' performances. As we previously analyzed in Fig.5, such improvements can be mainly attributed to the increased recall and tolerable distance error of the virtual points. However, when K increases to 10, the model's performance drops because the extra noise diminishes the benefits of correctly generated virtual points. Moreover, we also study whether increasing the number of seeds per instance N can achieve similar improvements, and the results when K is set to 1 and N is set to 200 are shown in the last row of Table 6. It can be seen that simply increasing seeds cannot bring improvements as our multi-depth strategy, but can even hurt performance due to more noise from single-depth unprojection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.619, + 0.567 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In this paper, we introduce MSMDFusion, a novel LiDAR-camera fusion framework for 3D object detection, which facilitates multi-modal interaction across multiple scales. Our approach employs a Multi-Depth Unprojection (MDU) strategy to obtain reliable virtual points from images, and applies Gated Modality-Aware Convolution (GMA-Conv) at each scale to enable fine-grained, controllable multi-modal LiDAR-camera fusion. We further combine multi-granularity information across scales to generate comprehensive features for the final prediction. Extensive experiments demonstrate the effectiveness of these components, with our method ultimately achieving state-of-the-art performance on the nuScenes dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.81, + 0.688, + 0.827 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This project was supported by National Key R&D Program of China (No. 2020AAA0140001), NSFC (No. 62072116), and Science and Technology Commission of Shanghai Municipality (No. 21JC1400600)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21650" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[4] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.515 + ], + "angle": 0, + "content": "[6] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In European Conference on Computer Vision, pages 35-52. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.586 + ], + "angle": 0, + "content": "[7] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[8] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.66, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[9] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[10] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1120-1129, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Ming Liang, Bin Yang, Yun Chen, Rui Hu, and Raquel Urtasun. Multi-task multi-sensor fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7345-7353, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[15] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[16] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[18] Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA), pages 7276-7282. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[19] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[20] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[21] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[22] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[23] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. arXiv preprint arXiv:2208.11112, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21651" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In European Conference on Computer Vision, pages 720-736. Springer, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[27] Zhijie Zhang, Yan Liu, Junjie Chen, Li Niu, and Liqing Zhang. Depth privileged object detection in indoor scenes via deformation hallucination. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3456-3464, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[28] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.472, + 0.331 + ], + "angle": 0, + "content": "[29] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[30] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21652" + } + ] +] \ No newline at end of file diff --git a/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_origin.pdf b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..919f119c9ac77efb168347d881c94089d6aed9da --- /dev/null +++ b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/2069ca18-1048-442b-b642-22315aec0f29_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f0940a2edccccdb78caded266327ec37c8f67cb9843ee03241d8686ec89d196 +size 2086768 diff --git a/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/full.md b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..407a594b1fe24b44b57b96c75cde81d59cad19db --- /dev/null +++ b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/full.md @@ -0,0 +1,243 @@ +# MSMDFusion: Fusing LiDAR and Camera at Multiple Scales with Multi-Depth Seeds for 3D Object Detection + +Yang Jiao $^{1,2*}$ Zequn Jie $^{3*}$ Shaoxiang Chen $^{3}$ Jingjing Chen $^{1,2\dagger}$ Lin Ma $^{3\dagger}$ Yu-Gang Jiang $^{1,2}$ + +1 Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University + +$^{2}$ Shanghai Collaborative Innovation Center of Intelligent Visual Computing + $^{3}$ Meituan + +# Abstract + +Fusing LiDAR and camera information is essential for accurate and reliable 3D object detection in autonomous driving systems. This is challenging due to the difficulty of combining multi-granularity geometric and semantic features from two drastically different modalities. Recent approaches aim at exploring the semantic densities of camera features through lifting points in 2D camera images (referred to as "seeds") into 3D space, and then incorporate 2D semantics via cross-modal interaction or fusion techniques. However, depth information is under-investigated in these approaches when lifting points into 3D space, thus 2D semantics can not be reliably fused with 3D points. Moreover, their multi-modal fusion strategy, which is implemented as concatenation or attention, either can not effectively fuse 2D and 3D information or is unable to perform fine-grained interactions in the voxel space. To this end, we propose a novel framework with better utilization of the depth information and fine-grained cross-modal interaction between LiDAR and camera, which consists of two important components. First, a Multi-Depth Unprojection (MDU) method is used to enhance the depth quality of the lifted points at each interaction level. Second, a Gated Modality-Aware Convolution (GMA-Conv) block is applied to modulate voxels involved with the camera modality in a fine-grained manner and then aggregate multi-modal features into a unified space. Together they provide the detection head with more comprehensive features from LiDAR and camera. On the nuScenes test benchmark, our proposed method, abbreviated as MSMD-Fusion, achieves state-of-the-art results on both 3D object detection and tracking tasks without using test-time-augmentation and ensemble techniques. The code is available at https://github.com/SxJyJay/MSMDFusion. + +![](images/7581a5d6ba857f2a80165958d2500486eb0afe2ee11843b90dffca75505b1034.jpg) +Figure 1. Illustration of our MSMDFusion pipeline. The yellow arrows indicate information passing or interaction between the Li-DAR and camera modalities. + +# 1. Introduction + +Detecting 3D objects [1, 22, 29] is regarded as a fundamental task for autonomous driving. Aiming at the robust environmental perception, LiDARs and cameras are widely equipped on autonomous driving vehicles since they can provide complementary information. Characterized by point clouds, LiDARs can capture accurate spatial information, while cameras contain rich semantics and context with images. Therefore, developing multi-modal detectors that enjoy the benefits of the two worlds is promising. Such an idea has catalyzed the emergence of a set of recent researches [1, 3, 12, 13, 15, 17, 19, 20, 25, 26]. Early works [1, 3, 11, 12, 18-20, 26] perform LiDAR-camera fusion by projecting 3D LiDAR points (or region proposals generated from them) onto 2D image planes to collect useful 2D semantics. However, such a paradigm suffers from the signal density mismatching of multi-modality sensors. Since LiDAR points are much sparser than camera pixels, this projection manner will inevitably waste semantically rich 2D features [1, 13]. + +Recently, another paradigm [9,10,13,15,25] for LiDAR-camera fusion has emerged. Instead of collecting 2D semantic features with 3D queries, these methods first estimate depths of pixels, and then directly lift 2D pixels with + +their semantics to the 3D world (we refer to these pixels and corresponding lifted points as "seeds" and "virtual points" in this paper) to fuse with the real 3D point cloud. Two methods with the same name of BEVFusion [13, 15] treat every image feature pixel as a seed and generate virtual points in the BEV space. MVP [25] and VFF [10] sample pixels from foreground regions and lift them to the voxel space. Benefited from the dense virtual points, such a paradigm not only maintains the semantic consistency in the image [15], but also complements the geometric cues for the sparse LiDAR point cloud [25]. + +Despite significant improvements have been made, existing methods in this line suffer from two major problems, which hampers benefiting from virtual points. First, depth, as the key to the quality of virtual points, is underinvestigated in generating virtual points. On the one hand, depth directly determines the spatial location in 3D space of a seed via perspective projection which can thereby significantly affect 3D object detection results. On the other hand, depth can also enhance the semantics carried by virtual points by providing color-insensitive cues in describing objects [27], since combining RGB information with depth guidance correlates camera pixels of similar depths and enables them to jointly contribute to capturing instance-related semantics when lifted as virtual points. Existing multi-modal detectors [9, 13, 15] mainly pay attention on interacting LiDAR points with camera virtual points, while ignoring the importance of seed depths in generating the virtual points. + +Second, the fine-grained cross-modal interaction between virtual points and 3D points in the uncompressed space (e.g., the voxel space) is crucial but non-trivial. Generated virtual points are geometrically and semantically inconsistent with real LiDAR points due to imperfect depths and inherent modality gap. Hence, in order to benefit from the semantically rich virtual points, it is necessary to adaptively select helpful information from virtual points under the guidance of real LiDAR points in a fine-grained and controllable manner. However, such a cross-modal interaction is constrained by the intensive memory and computation cost brought by the massive amounts and unstructured nature of point cloud data. Alternatively, existing approaches combine the multi-modal information with simple concatenate [25] or add operations [9] in the voxel space, or perform cross-attention in a compressed BEV space [23]. + +Aiming at unlocking the potential of virtual points and addressing the drawbacks of existing methods, we propose a multi-scale fusion framework, called MSMDFusion, and within each scale, there are two key novel designs, namely the Multi-Depth Unprojection (MDU) and Gated Modality-Aware Convolution (GMA-Conv). As shown in Fig.1, MDU is mainly for enhancing the quality of generated virtual points in terms of geometric accuracy and + +semantic richness. To be specific, when lifting 2D seeds from image into the 3D space, multiple depths are explored within a reference neighborhood to generate virtual points with more reliable depth. Next, the camera feature and depth are combined to produce depth-aware features as stronger 2D semantics to decorate these virtual points. GMA-Conv takes real LiDAR points and generated virtual points as inputs, and performs fine-grained interaction in a select-then-aggregate manner. Concretely, we first adaptively select useful information from camera voxel features under the guidance of reference LiDAR voxels, then aggregate them grouped sparse convolutions for sufficient multimodal interaction. We also specifically adopt a voxel subsampling strategy to efficiently obtain reliable LiDAR references when implementing our GMA-Conv. + +Finally, with the resulting multi-modal voxel features from multiple scales, we further associate them with cascade connections across scales to aggregate multi-granularity information. With the above designs, the camera semantics encapsulated in the virtual points are consistently combined with LiDAR points, and thereby providing a stronger multi-modal feature representation for boosting the 3D object detection. As shown in the Table 3, with 100 times fewer generated virtual points than the two BEVFusion methods [13, 15], our MSMDFusion can still achieve state-of-the-art performances. + +In summary, our contributions lie in threefold: (1) We propose a novel MSMDFusion approach, which encourages sufficient LiDAR-Camera feature fusion in the multiscale voxel space. (2) Within each scale, we propose a Multi-Depth Unprojection strategy (MDU) to promote virtual points generation with better locations and semantics by fully leveraging depth of pixels, as well as a Gated Modality-Aware Convolution (GMA-Conv) to achieve fine-grained controllable multi-modal interaction. (3) Extensive experimental results on the large-scale nuScenes dataset prove the effectiveness of our MSMDFusion and its components. We achieve state-of-the-art performances with $71.5\%$ mAP and $74.0\%$ NDS on the challenging nuScenes detection track using single model1. When combining the simple greedy tracking strategy [24], our method also achieves strong tracking results with $74.0\%$ AMOTA. + +# 2. Related Work + +Recently, fusing LiDAR and camera signals in 3D detection has attracted increasing interest. Early works [3, 6, 12, 19, 20, 26] project 3D queries to 2D camera images for collecting useful semantics. Under such a paradigm, MV3D [3] and AVOD [7] associate 3D proposals with 2D RoI features. While PointPainting [19] and PointAugment- + +![](images/1439e75b70d6e192392497352d904d9fcd101bf660b9b3d9cacbabe687410320.jpg) +Figure 2. The framework of our proposed MSMDFusion. Given a LiDAR point cloud and corresponding multi-view camera images, MSMDFusion first separately transforms them into the voxel space and obtains multi-scale LiDAR and camera features. Then, the LiDAR and camera features at each scale interact through our proposed Multi-Depth Unprojection strategy (MDU) and Gated Modality-Aware Convolution block (GMA-Conv, details in Fig.4). The outputs of GMA-Conv at each scale are further aggregated for combining multi-granularity information. Finally, the resulting multi-modal and LiDAR voxel features are flattened into the BEV space for the final prediction. $\oplus$ represents voxel addition and "spconv" indicates 3D sparse convolution. + +ing [20] directly decorate raw 3D points with 2D semantics. EP-Net [6] and 3D-CVF [26] perform multi-modal fusion at both point level and proposal level. However, since 3D points are inherently sparse, such a hard association approach wastes the dense semantic information in 2D features. Recently, multi-modal 3D detectors [13, 15, 23, 25] lift dense 2D seeds to 3D space for learning the 2D-3D joint representation in a shared space. Two BEVFusion methods [13, 15] densely lift every image feature pixel into 3D space and then encode these lifted points as another BEV (bird's-eye view) map to fuse with the BEV map obtained using the original 3D point cloud. LiDAR and camera signals will finally share a common representation space, but they do not interact at any stage before BEV fusion. MVP [25] selects seeds from foreground objects and unprojected them into the LiDAR frame for points cloud augmentation. With the seeds sampling strategy similar to MVP, UVTR [9] constructs both LiDAR and camera branches in the voxel space to maintain the modality-specific information, and then interact them with voxel add operation. However, the LiDAR and camera branches are constructed in a single scale of voxel space, losing the benefits of different levels of feature abstraction in multi-scale voxel space. Moreover, merging LiDAR and camera voxels with simple addition limits their fine-grained interaction. + +# 3. Method + +# 3.1. Framework Overview + +An overall view of MSMDFusion is shown in Fig.2. Given a LiDAR point cloud and corresponding multiview camera images as inputs, MSMDFusion first extracts multi-scale features from both modalities in the voxel space. Then, LiDAR-camera interaction is performed within multi-scale voxel space to properly combine multi-granularity information from both modalities. At each scale, we specifically design a Multi-Depth Unprojection (MDU) strategy to obtain high-quality virtual points in the voxel space, and a Gated Modality-Aware Convolution block (GMA-Conv) for effective LiDAR-camera interaction and fusion. We also introduce cross-scale connections to progressively combine features of different granularities. Afterward, the deeply interacted multi-modal features, together with the LiDAR features, are transformed into the BEV space and fed to the BEV encoder and detection head for the final prediction. + +# 3.2. LiDAR and Camera Feature Extraction + +We first extract high-level features from the raw inputs of LiDAR and camera. For a given LiDAR point cloud, we voxelize the points and extract their features in the voxel space using a set of 3D sparse convolutional blocks [4] following prior arts [1, 13, 15]. Each convolutional block out + +![](images/7eca596c2913e40ab3a1763b89398cbca7e2c909445143e62e77b2f288f402f7.jpg) +large depth bias moderate depth bias +(a) Nearest neighbor retrieval +Figure 3. Comparison of the effects of retrieving reference points from nearest neighbor (MVP) and K-nearest neighbor (Ours). The blue circle and yellow star represent seeds and reference points, respectively. To keep the figure concise, we only depict the case that $\mathrm{K} = 2$ . Our strategy can obtain more reliable depths by exploring more neighbors. + +![](images/dbd763435d95891ae079e5fda7da272a8252a336cd42646e1d1028298d998970.jpg) +(b) K Nearest neighbor retrieval (Ours) + +puts voxel features of different scales, representing different levels of abstraction for point cloud, and these multiscale features will further interact with their 2D counterparts through our proposed GMA-Conv blocks. + +For the multi-view camera images, we adopt ResNet50 with FPN as the image backbone to extract multi-scale image features, which contains rich object semantics. However, it is non-trivial to transfer these useful semantics to the 3D voxel field, since images are inherently 2D data. Motivated by recent works [13, 15, 25] which select seeds from the image plane and estimate their depths to lift them as 3D virtual points, we further propose a Multi-Depth Unprojection strategy to lift 2D seeds, which could mitigate the drawbacks of prior works. The details will be presented in the following subsections. + +# 3.3. Multi-Modal Interaction in Voxel Space + +The extracted LiDAR and camera features from the previous stage respectively focus on geometric and semantic information, hence the goal of multi-modal interaction is to properly fuse these features into a unified rich representation of the scene. As shown in the Fig.2, the components of LiDAR-camera interaction within each scale in the voxel space are conceptually simple: 1) Multi-Depth Unprojection (MDU) strategy is introduced to improve the quality of the generated virtual points and 2) Gated Modality-Aware Convolution (GMA-Conv) block is designed for controllable LiDAR-camera fusion. In the following parts, we will first elaborate on MDU and GMA-Conv within a specific scale, then on top of them, we describe the complete multiscale fusion process. + +# 3.3.1 Multi-Depth Unprojection + +To lift a 2D seed in the pixel space to the 3D space (i.e., the unprojection operation), the depth associated with each seed should be estimated [21]. To obtain an initial reli + +able depth estimation, we use the strategy from MVP [25] as the basis as shown in Fig.3(a). Formally, given a point cloud $\mathcal{P} = \{(x_i,y_i,z_i)\}_{i = 1}^{N_P}$ and multi-view images as inputs, MVP first projects 3D points onto 2D images and preserves points falling within the 2D foreground instance masks. These 3D points are kept as reference points to provide depth, and they are denoted as $\mathcal{R} = \{(u_i,v_i,d_i)\}_{i = 1}^{N_R}$ , where $u_{i}$ and $v_{i}$ are pixel coordinates and $d_{i}$ is the real depth. Then, a set of seeds $S = \{(u_i^s,v_i^s)\}_{i = 1}^{N_S}$ are uniformly sampled from each instance mask on the image, and each of them will retrieve a real depth from its nearest reference point as its estimated depth. Finally the sampled seeds with estimated depths are unprojected to 3D as virtual points. + +Though effective, such a strategy ignores the fact that spatial proximity in 2D images can not be guaranteed in 3D, and this may lead to inaccurate depth estimation as shown in the red circle of Fig.3(a). Toward this end, we propose to equip every seed with multiple depths by retrieving K-nearest reference points, which can be regarded as a soft strategy to achieve more reliable depth estimation. As shown in the green circle of the Fig.3(b), each seed can be unprojected using multiple (K) depths from neighboring reference points, which generates K virtual points to improve its recall of the actual 3D points. We provide more detailed discussion on the effect of multi-depth seeds in Sec.4.3.3. + +These generated virtual points will be further decorated to make the best use of semantics encapsulated in images. But different from MVP [25] that simply decorates virtual points with class scores predicted by a pretrained 2D detector, we make an early interaction between image features and sparse depth information from reference points to generate depth-aware semantic features. These depth-aware semantic features are then used to decorate virtual points in an adaptive manner controlled by their corresponding seed depths. The whole process can be trained end-to-end. Concretely, we concatenate the camera image feature $C$ with a sparse depth map containing all depths of reference points $\mathcal{R}$ in that image, and then fuse them with a convolutional layer to obtain the depth-aware semantic feature $C^d$ . For the K virtual points generated by the same seed, the semantic feature of the corresponding pixel should also contribute differently along the projection ray in the 3D space [16]. Thus we compute a dynamic weight factor for each individual depth to modulate the semantic feature. The process for a specific seed $(u_i^s, v_i^s)$ can be formally described as: + +$$ +\begin{array}{l} s _ {i} ^ {k} = \operatorname {S i g m o i d} \left(\operatorname {L i n e a r} \left([ C ^ {d} [ u _ {i} ^ {s}, v _ {i} ^ {s} ]; d _ {i} ^ {s, k} ]\right)\right) \tag {1} \\ c _ {i} ^ {k} = C ^ {d} [ u _ {i} ^ {s}, v _ {i} ^ {s} ] * s _ {i} ^ {k} \\ \end{array} +$$ + +where $d_{i}^{s,k}$ is the $k$ -th estimated depth of seed $(u_{i}^{s}, v_{i}^{s})$ , $s_{i}^{k} \in [0,1]$ represents its weight factor, and $c^{k}$ is the modulated semantic feature, which will be used to decorate the corresponding unprojected virtual point $(u_{i}^{s}, v_{i}^{s}, d_{i}^{s,k})$ . + +![](images/eefbd71802db6811b3414bcb8786caddd5cfa9cc2ec90d279abfdd39c418ec4a.jpg) +Figure 4. Illustration of the select-then-aggregate process in the Gated Modality-Aware Convolution. Colors indicate different modalities: yellow for camera, blue for LiDAR, and red for LiDAR-camera combined. Cubes with shadows represent voxels selected with gates. + +The resulting virtual points with decorated features will be voxelized to match the output resolution of the corresponding scale of LiDAR branch, in order to perform interaction in the voxel space. + +# 3.3.2 Gated Modality-Aware Convolution + +With the resulting LiDAR and camera voxel features that share the same spatial resolution (i.e., the vertically aligned blue and yellow voxels in Fig.2), we aim to interact them in a fine-grained and controllable manner with our proposed Gated Modality-Aware Convolution (GMA-Conv) block in each scale. As shown in Fig.4, we first group the voxels according to their modalities: LiDAR-only (blue voxels), camera-only (yellow voxels), as well as LiDAR and camera combined (red voxels), and they are denoted as $f^L$ , $f^C$ , and $f^{LC}$ , respectively. Then, more fine-grained multi-modal interaction is conducted in a select-then-aggregate manner. + +Select. Motivated by the fact that LiDAR-based detectors usually surpass their camera-based counterparts with a large margin, we take LiDAR as the guiding modality to select useful information from camera features. Specifically, we design a gate conditioned on the LiDAR feature to control and update its camera counterpart, which can be formally described as: + +$$ +\tilde {f _ {i}} ^ {C} = \operatorname {R e L U} \left(\operatorname {L i n e a r} \left(f _ {j} ^ {L}\right)\right) * f _ {i} ^ {C} \tag {2} +$$ + +where $\tilde{f_i^C}$ is the updated camera voxel feature, and $i$ and $j$ are voxel indices of paired camera and LiDAR voxel features ( $f_j^L$ can be regarded as the reference voxel for $f_i^C$ ). The key to make this selection process effective is how to efficiently find reliable reference voxels. A straightforward way is to retrieve the nearest LiDAR voxels for each camera-only voxel $f_i^C$ from $f^L$ . And similarly for LiDAR-camera voxels $f_k^{LC}$ , the reference can be retrieved from both the LiDAR and camera modalities (i.e., $f^L$ and $f^C$ ). Although nearby neighbors from more reliable modalities can serve as a good reference, such a retrieval process is infeasible to be implemented with exhaustive pairwise distance calculation, because the computation and memory + +cost brought by massive amounts of LiDAR and camera voxels is intractable, and we will give more analysis and our solution in the supplementary material. + +Aggregate. With the original LiDAR-only voxels $f^L$ , as well as the updated camera-only and LiDAR-camera combined voxels $\tilde{f}^{C}$ and $\tilde{f}^{LC}$ , we first transform their modality-specific representation into an intermediate joint space with 3D sparse convolutions. Then, within the joint space, we further combine all voxels and promote their interaction through another 3D sparse convolution as shown in Fig.4. The resulting multi-modal voxel features are denoted as $F^M$ . + +# 3.3.3 Multi-Scale Progressive Interaction + +After performing MDU and GMA-Conv within different scales in the voxel space, multi-modal voxel features $\{F_i^M\}$ are generated for each scale. To further aggregate these multi-granularity information from $\{F_i^M\}$ , we introduce cascade connections across different scales as shown in Fig.2, which can be formulated as: + +$$ +\hat {F} _ {i + 1} ^ {M} = F _ {i + 1} ^ {M} + \text {D o w n S a m p l e} \left(\hat {F} _ {i} ^ {M}\right), \tag {3} +$$ + +where the voxel downsample operation DownSample() is applied to align their spatial resolutions, and $\hat{F}_{i+1}^{M}$ is the resulting voxel features that combine multi-scale information from the current and previous scales. + +With the above designs, multi-granularity LiDAR-camera features can be thoroughly interacted at multiple scales in the voxel space. The final output multi-modal voxel features can serve as a powerful representation for the subsequent modules. + +# 3.4. Prediction in BEV Space + +Following the workflow of prevalent voxel-based 3D object detectors [22, 24, 29], we compress the height of voxel features to transform them into the BEV (bird's-eye view) space. Specifically, we first fuse the LiDAR and multimodal BEV features with a lightweight 2D convolution block. Then, the BEV feature containing enhanced multi + +
MethodModalitymAPNDSCarTruckC.V.BusTrailerBarrierMotor.BikePed.T.C.
PointPillar [8]L40.155.076.031.011.332.136.656.434.214.064.045.6
CenterPoint [24]L60.367.385.253.520.063.656.071.159.530.784.678.4
TransFusion-L [1]L65.570.286.256.728.266.358.878.268.344.286.182.0
PointPainting [19]LC46.458.177.935.815.836.237.360.241.524.173.362.4
3D-CVF [26]LC52.762.383.045.015.948.849.665.951.230.474.262.9
PointAugmenting [20]LC66.871.087.557.328.065.260.772.674.350.987.983.6
MVP [25]LC66.470.586.858.526.167.457.374.870.049.389.185.0
TransFusion [1]LC68.971.787.160.033.168.360.878.173.652.988.486.7
VFF [10]LC68.472.486.858.132.170.261.073.978.552.987.183.8
BEVFusion [13]LC69.271.888.160.934.469.362.178.272.252.289.285.2
BEVFusion [15]LC70.272.988.660.139.369.863.880.074.151.089.286.5
UVTR [9]LC67.171.187.556.033.867.559.573.073.454.886.379.6
DeepInteraction [23]LC70.873.487.960.237.570.863.880.475.454.590.387.0
MSMDFusion (Ours)LC71.574.088.461.035.271.464.280.776.958.390.688.1
+ +Table 1. Comparison with state-of-the-art methods on nuScenes test set. Note that these are all single-model results without ensemble or test-time augmentation. We highlight the best performances across all methods with bold. + +
MethodModalityAMOTA ↑AMOTP ↓TP ↑FP ↓FN ↓Recall ↑
CenterPoint [24]L63.80.60695877186122292867.5
TransFusion [1]LC71.80.55196775162322184675.8
UVTR [9]LC70.10.68698434156152019075.0
BEVFusion [15]*LC74.10.40399664199971939577.9
MSMDFusion (Ours)LC74.00.54998624147891985376.3
+ +Table 2. Comparison with leading detectors that use a greedy tracker on nuScenes tracking task. * represents that BEVFusion uses an ensemble model, while we only use a single model. We use **bold** to denote the best performances among all approaches with a single model. + +modality information is fed into a conventional BEV encoder and a detection head for the final prediction. + +# 4. Experiments + +# 4.1. Experimental Setup + +# 4.1.1 Dataset and Metrics + +The nuScenes [2] dataset is a large-scale autonomous driving benchmark including 10,000 driving scenarios in total, which are split into 700, 150, and 150 scenes for training, validation, and testing, respectively. For detection, nuScenes defines a set of evaluation protocols, including the nuScenes Detection Score (NDS), mean Average Precision (mAP), as well as five True Positive (TP) metrics, namely mean Average Translation Error (mATE), mean Average Scale Error (mASE), mean Average Orientation Error (mAOE), mean Average Velocity Error (mAVE) and mean Average Attribute Error (mAAE). We report mAP, which is the mean of the average precision across ten classes under distance thresholds of $0.5\mathrm{m}$ , $1\mathrm{m}$ , $2\mathrm{m}$ , $4\mathrm{m}$ . NDS is the weighted combination of mAP, mATE, mASE, mAOE, mAVE and mAAE. + +Implementation Details We use ResNet-50 [5] with + +FPN [14] as the image backbone and VoxelNet [22] as the LiDAR backbone. We set the image size to $448 \times 800$ , and voxel size as $(0.075m, 0.075m, 0.2m)$ following [1]. To make the best use of multi-scale semantics, we extract features from four levels of the FPN (C2 to C5). Following MVP [25], we use CenterNet2 [28] as 2D detector when generating virtual points, and 50 seeds are selected on each instance unless otherwise specified. Our model training has two stages: (1) We first train a LiDAR-only detector for 20 epochs as our 3D backbone. (2) We then connect the proposed LiDAR-camera fusion modules with the 3D backbone for a joint training of another 6 epochs. The data augmentation strategies and training schedules are the same as prior works [24, 30]. We do not use Test-Time Augmentation (TTA) or multi-model ensemble during inference. + +# 4.2. Comparison with State-of-the-art + +We compare our MSMDFusion with state-of-the-art approaches on the nuScenes test set. Overall, Table 1 shows that our method surpasses all existing methods, and + +
MethodNVPF2mAPNDSFPS
BEVFusion [13]5M69.271.80.7
BEVFusion [15]2M70.272.98.4
MSMDFusion(Ours)16k71.574.02.1
+ +achieves the new state-of-the-art performances of $71.5\mathrm{mAP}$ and 74.0 NDS. Meanwhile, our method maintains consistent performance advantages on most object categories, especially on the challenging category: Bike, where a $3.5\%$ absolute AP gain over the strongest competitor is achieved. + +Since our MSMDFusion and two strong BEVFusion methods all fuse LiDAR and camera signals via generating 3D virtual points from 2D seeds, we specifically compare our method with them in terms of the number of generated virtual points per LiDAR frame as shown in Table 3. The results suggest that although with 100 times fewer generated virtual points (16k vs 2M/5M) than them, our MSMDFusion still outperforms them in terms of both mAP and NDS, which proves that our method can better utilize 2D semantics. Meanwhile, we also compare the FPS of our method and two BEVFusion approaches. BEVFusion [15] can achieve the fastest inference speed due to its efficient camera-to-BEV transformation implemented with CUDA. Without such a hardware-friendly implementation, BEVFusion [13] suffers from processing a huge amount of virtual points, while our MSMDFusion runs much faster although performing multi-scale fine-grained fusion in the voxel space. + +Besides, to demonstrate that our MSMDFusion can also generalize to other downstream tasks, we further perform the tracking task on nuScenes. For fair comparison, we only compare with single-model methods (i.e., without TTA or ensemble) using the greedy tracker introduced in [24]. As shown in Table 2, our method consistent outperforms other single model competitors across all evaluation metrics. Meanwhile, although without using ensemble techniques, our method also achieve competitive results with the ensemble BEVFusion [15] method. + +# 4.3. Comprehensive Analysis + +Following UVTR [9], we conduct the following ablation studies on a randomly sampled 1/4 split of nuScenes training set for efficiency. + +# 4.3.1 Ablation of proposed components + +We conduct comprehensive ablation studies for each of our proposed components as shown in Table 4. We use + +Table 3. Number of Virtual points Per Frame (NVPF) and performance comparison with two strongest methods on nuScenes test set. The FPS is tested on a single NVIDIA RTX 3090 GPU. + +
#MDU*MDUGMA-ConvmAPNDS
160.2665.62
2✓64.3367.03
3✓65.5968.04
4✓✓65.9768.24
5✓✓66.9368.93
+ +Table 4. Ablation studies for our proposed components on the nuScenes validation set. MDU* is a degraded version of MDU by discarding the depth-aware feature generating process. + +
#C2C3C4C5mAPNDS
160.2665.62
2✓65.5167.92
3✓64.9967.97
4✓64.9767.66
5✓64.5267.61
6✓✓66.3468.63
7✓✓✓66.8068.86
8✓✓✓✓66.9368.93
+ +Table 5. Effects of LiDAR-camera interaction at different scales, C2-C5 represent the scales where the spatial resolutions of the features decrease from C2 to C5. + +TransFusion-L [1] as our baseline (#1), and perform multiscale fusion across all experiments. + +From the results of the Table 4, we have the following observations. (i) Introducing virtual points for multi-modal interaction (#2 and #3) brings evident improvements over the baseline, which proves that fusing dense 2D semantics from virtual points with LiDAR features in the voxel space can significantly enhance the LiDAR-only detector. (ii) An early depth-color interaction (#2 and #3) can provide stronger semantics, which further boosts the benefits of interaction between LiDAR and virtual points. (iii) Using GMA-Conv to select helpful information from virtual points (#4 and #5) can bring further improvements over absorbing all information from virtual points without distinction (#2 and #3). (iv) With all these components combined (#5), the final mAP and NDS are significantly improved from $60.26\%$ to $66.93\%$ and $65.62\%$ to $68.93\%$ , respectively, which proves the effectiveness of our designs. + +# 4.3.2 Effects of multi-scale interaction + +We also investigate the effects of performing LiDAR-camera interaction at different scales. As shown in Table 5, for models with single-scale interaction (#2-#4), interacting at scales of larger spatial resolution generally leads to larger performance gains over the baseline (#1), which mainly benefits from more fine-grained image and LiDAR features. If the multi-modal interaction is performed at more scales (#6-#8), multi-granularity information can be + +
KNNVPFmAPNDS
1502.6K65.8668.24
3508.0K66.4968.88
65016.0K66.9368.93
105026.6K66.6168.85
120010.6K65.5968.00
+ +Table 6. Effects of the number of depths per seed (K) and number of seeds per instance (N) in multi-depth unprojection. NVPF represents the number of virtual points per frame. + +integrated into our progressive interaction process to consistently achieve better performances than single-scale models. The complete multi-scale interaction model achieves performance gains of $\sim 2\%$ mAP and $\sim 1\%$ NDS over the best single-scale model, which indicates that multi-scale interaction is beneficial for comprehensive multi-modal fusion. + +![](images/9de000808512cc8e635f2bab6019cb6912ceaa23482df945daaa1c9729196f9b.jpg) +Figure 5. Mean error and recall of the virtual points computed with different numbers (K) of retrieved depths per seed. + +# 4.3.3 Discussion on MDU + +Our proposed MDU is a simple yet effective strategy to lift 2D seeds from image to 3D space, and we would like to further discuss one critical problem: will the multiple depths for a seed introduce geometric noise and diminish the benefits? We conducted a quantitative experiment on the nuScenes dataset to investigate this problem. Recall that original point cloud, their projection on image plane (i.e., reference points) and seeds are denoted as $\mathcal{P}$ , $\mathcal{R}$ and $\mathcal{S}$ , respectively. We fist preserve a subset $\mathcal{R}^e = \{(u_i^e,v_i^e,d_i^e)\}_{i=1}^{N^e}$ of the reference points $\mathcal{R}$ , and use all $\{(u_i^e,v_i^e)\}$ as seeds to generate virtual points $\hat{\mathcal{P}}^e$ by retrieving real depths from the remaining reference points $\mathcal{R} - \mathcal{R}^e$ as done in MDU. Since the real 3D coordinates $\mathcal{P}^e$ of the preserved reference points $\mathcal{R}^e$ are available, we then measure two important metrics for the generated virtual points $\hat{\mathcal{P}}^e$ : (1) The mean distance error in 3D space between $\hat{\mathcal{P}}^e$ and $\mathcal{P}^e$ , which reflects the overall accuracy of the virtual points' coordinates. (2) The recall rate that is the ratio of points in $\mathcal{P}^e$ which fall within a small neighborhood of at + +least one point in $\hat{\mathcal{P}}^e$ , which represents MDU's ability to capture unseen real 3D points. We also vary the number of depths per seed (K) to show its effect. As shown in Fig.5, when the number of K increases from 1 to 6, the recall rate of real points improves dramatically, meanwhile, the mean distance error does not increase significantly. But a too large value of K (e.g., 10) can inevitably introduce noise. This proves that with a properly chosen number of depths, MDU does not lead to increased noise in depth estimation and can generate virtual points with reliable depths while also covering more real points than a single-depth strategy. + +To further inspect the effects of K on the final detection performance, we keep other settings unchanged and vary K to compare performances. As Table 6 shows, increasing K from 1 to 6 can gradually improve these models' performances. As we previously analyzed in Fig.5, such improvements can be mainly attributed to the increased recall and tolerable distance error of the virtual points. However, when K increases to 10, the model's performance drops because the extra noise diminishes the benefits of correctly generated virtual points. Moreover, we also study whether increasing the number of seeds per instance N can achieve similar improvements, and the results when K is set to 1 and N is set to 200 are shown in the last row of Table 6. It can be seen that simply increasing seeds cannot bring improvements as our multi-depth strategy, but can even hurt performance due to more noise from single-depth unprojection. + +# 5. Conclusion + +In this paper, we introduce MSMDFusion, a novel LiDAR-camera fusion framework for 3D object detection, which facilitates multi-modal interaction across multiple scales. Our approach employs a Multi-Depth Unprojection (MDU) strategy to obtain reliable virtual points from images, and applies Gated Modality-Aware Convolution (GMA-Conv) at each scale to enable fine-grained, controllable multi-modal LiDAR-camera fusion. We further combine multi-granularity information across scales to generate comprehensive features for the final prediction. Extensive experiments demonstrate the effectiveness of these components, with our method ultimately achieving state-of-the-art performance on the nuScenes dataset. + +# 6. Acknowledgements + +This project was supported by National Key R&D Program of China (No. 2020AAA0140001), NSFC (No. 62072116), and Science and Technology Commission of Shanghai Municipality (No. 21JC1400600). + +# References + +[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 1, 3, 6, 7 +[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6 +[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1, 2 +[4] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 3 +[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[6] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In European Conference on Computer Vision, pages 35-52. Springer, 2020. 2, 3 +[7] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. 2 +[8] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 6 +[9] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. 1, 2, 3, 6, 7 +[10] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1120-1129, 2022. 1, 2, 6 +[11] Ming Liang, Bin Yang, Yun Chen, Rui Hu, and Raquel Urtasun. Multi-task multi-sensor fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7345-7353, 2019. 1 +[12] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 1, 2 + +[13] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 1, 2, 3, 4, 6, 7 +[14] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 6 +[15] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 1, 2, 3, 4, 6, 7 +[16] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 4 +[17] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 1 +[18] Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA), pages 7276-7282. IEEE, 2019. 1 +[19] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 1, 2, 6 +[20] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 1, 2, 3, 6 +[21] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 4 +[22] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 5, 6 +[23] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. arXiv preprint arXiv:2208.11112, 2022. 2, 3, 6 +[24] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2, 5, 6, 7 +[25] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. 1, 2, 3, 4, 6 + +[26] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In European Conference on Computer Vision, pages 720-736. Springer, 2020. 1, 2, 3, 6 +[27] Zhijie Zhang, Yan Liu, Junjie Chen, Li Niu, and Liqing Zhang. Depth privileged object detection in indoor scenes via deformation hallucination. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3456-3464, 2021. 2 +[28] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 6 +[29] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1, 5 +[30] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.6 \ No newline at end of file diff --git a/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/images.zip b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a3f4cf0a4a9e7e017d0ce4671ba287725e4cf5f3 --- /dev/null +++ b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9f450af17588532b7fa9b0ff4d25a9009c0e3d41e3381f260ede953228bc845 +size 471639 diff --git a/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/layout.json b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..61300368b1f2e170815ba287af08d09750ef0390 --- /dev/null +++ b/2023/MSMDFusion_ Fusing LiDAR and Camera at Multiple Scales With Multi-Depth Seeds for 3D Object Detection/layout.json @@ -0,0 +1,6877 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 503, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 503, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 503, + 140 + ], + "type": "text", + "content": "MSMDFusion: Fusing LiDAR and Camera at Multiple Scales with Multi-Depth Seeds for 3D Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "text", + "content": "Yang Jiao" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "text", + "content": " Zequn Jie" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "text", + "content": " Shaoxiang Chen" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "text", + "content": " Jingjing Chen" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "text", + "content": " Lin Ma" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "text", + "content": " Yu-Gang Jiang" + }, + { + "bbox": [ + 56, + 159, + 537, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 175, + 488, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 175, + 488, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 488, + 189 + ], + "type": "text", + "content": "1 Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 189, + 480, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 189, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 113, + 189, + 480, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 189, + 480, + 216 + ], + "type": "text", + "content": " Shanghai Collaborative Innovation Center of Intelligent Visual Computing \n" + }, + { + "bbox": [ + 113, + 189, + 480, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 189, + 480, + 216 + ], + "type": "text", + "content": "Meituan" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 271, + 290, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 290, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 290, + 668 + ], + "type": "text", + "content": "Fusing LiDAR and camera information is essential for accurate and reliable 3D object detection in autonomous driving systems. This is challenging due to the difficulty of combining multi-granularity geometric and semantic features from two drastically different modalities. Recent approaches aim at exploring the semantic densities of camera features through lifting points in 2D camera images (referred to as \"seeds\") into 3D space, and then incorporate 2D semantics via cross-modal interaction or fusion techniques. However, depth information is under-investigated in these approaches when lifting points into 3D space, thus 2D semantics can not be reliably fused with 3D points. Moreover, their multi-modal fusion strategy, which is implemented as concatenation or attention, either can not effectively fuse 2D and 3D information or is unable to perform fine-grained interactions in the voxel space. To this end, we propose a novel framework with better utilization of the depth information and fine-grained cross-modal interaction between LiDAR and camera, which consists of two important components. First, a Multi-Depth Unprojection (MDU) method is used to enhance the depth quality of the lifted points at each interaction level. Second, a Gated Modality-Aware Convolution (GMA-Conv) block is applied to modulate voxels involved with the camera modality in a fine-grained manner and then aggregate multi-modal features into a unified space. Together they provide the detection head with more comprehensive features from LiDAR and camera. On the nuScenes test benchmark, our proposed method, abbreviated as MSMD-Fusion, achieves state-of-the-art results on both 3D object detection and tracking tasks without using test-time-augmentation and ensemble techniques. The code is available at https://github.com/SxJyJay/MSMDFusion." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 245, + 547, + 353 + ], + "blocks": [ + { + "bbox": [ + 308, + 245, + 547, + 353 + ], + "lines": [ + { + "bbox": [ + 308, + 245, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 547, + 353 + ], + "type": "image", + "image_path": "7581a5d6ba857f2a80165958d2500486eb0afe2ee11843b90dffca75505b1034.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 361, + 545, + 394 + ], + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 394 + ], + "type": "text", + "content": "Figure 1. Illustration of our MSMDFusion pipeline. The yellow arrows indicate information passing or interaction between the Li-DAR and camera modalities." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 416, + 387, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 416, + 387, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 416, + 387, + 429 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": "Detecting 3D objects [1, 22, 29] is regarded as a fundamental task for autonomous driving. Aiming at the robust environmental perception, LiDARs and cameras are widely equipped on autonomous driving vehicles since they can provide complementary information. Characterized by point clouds, LiDARs can capture accurate spatial information, while cameras contain rich semantics and context with images. Therefore, developing multi-modal detectors that enjoy the benefits of the two worlds is promising. Such an idea has catalyzed the emergence of a set of recent researches [1, 3, 12, 13, 15, 17, 19, 20, 25, 26]. Early works [1, 3, 11, 12, 18-20, 26] perform LiDAR-camera fusion by projecting 3D LiDAR points (or region proposals generated from them) onto 2D image planes to collect useful 2D semantics. However, such a paradigm suffers from the signal density mismatching of multi-modality sensors. Since LiDAR points are much sparser than camera pixels, this projection manner will inevitably waste semantically rich 2D features [1, 13]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Recently, another paradigm [9,10,13,15,25] for LiDAR-camera fusion has emerged. Instead of collecting 2D semantic features with 3D queries, these methods first estimate depths of pixels, and then directly lift 2D pixels with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21643" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "their semantics to the 3D world (we refer to these pixels and corresponding lifted points as \"seeds\" and \"virtual points\" in this paper) to fuse with the real 3D point cloud. Two methods with the same name of BEVFusion [13, 15] treat every image feature pixel as a seed and generate virtual points in the BEV space. MVP [25] and VFF [10] sample pixels from foreground regions and lift them to the voxel space. Benefited from the dense virtual points, such a paradigm not only maintains the semantic consistency in the image [15], but also complements the geometric cues for the sparse LiDAR point cloud [25]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 422 + ], + "type": "text", + "content": "Despite significant improvements have been made, existing methods in this line suffer from two major problems, which hampers benefiting from virtual points. First, depth, as the key to the quality of virtual points, is underinvestigated in generating virtual points. On the one hand, depth directly determines the spatial location in 3D space of a seed via perspective projection which can thereby significantly affect 3D object detection results. On the other hand, depth can also enhance the semantics carried by virtual points by providing color-insensitive cues in describing objects [27], since combining RGB information with depth guidance correlates camera pixels of similar depths and enables them to jointly contribute to capturing instance-related semantics when lifted as virtual points. Existing multi-modal detectors [9, 13, 15] mainly pay attention on interacting LiDAR points with camera virtual points, while ignoring the importance of seed depths in generating the virtual points." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 423, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 616 + ], + "type": "text", + "content": "Second, the fine-grained cross-modal interaction between virtual points and 3D points in the uncompressed space (e.g., the voxel space) is crucial but non-trivial. Generated virtual points are geometrically and semantically inconsistent with real LiDAR points due to imperfect depths and inherent modality gap. Hence, in order to benefit from the semantically rich virtual points, it is necessary to adaptively select helpful information from virtual points under the guidance of real LiDAR points in a fine-grained and controllable manner. However, such a cross-modal interaction is constrained by the intensive memory and computation cost brought by the massive amounts and unstructured nature of point cloud data. Alternatively, existing approaches combine the multi-modal information with simple concatenate [25] or add operations [9] in the voxel space, or perform cross-attention in a compressed BEV space [23]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Aiming at unlocking the potential of virtual points and addressing the drawbacks of existing methods, we propose a multi-scale fusion framework, called MSMDFusion, and within each scale, there are two key novel designs, namely the Multi-Depth Unprojection (MDU) and Gated Modality-Aware Convolution (GMA-Conv). As shown in Fig.1, MDU is mainly for enhancing the quality of generated virtual points in terms of geometric accuracy and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "semantic richness. To be specific, when lifting 2D seeds from image into the 3D space, multiple depths are explored within a reference neighborhood to generate virtual points with more reliable depth. Next, the camera feature and depth are combined to produce depth-aware features as stronger 2D semantics to decorate these virtual points. GMA-Conv takes real LiDAR points and generated virtual points as inputs, and performs fine-grained interaction in a select-then-aggregate manner. Concretely, we first adaptively select useful information from camera voxel features under the guidance of reference LiDAR voxels, then aggregate them grouped sparse convolutions for sufficient multimodal interaction. We also specifically adopt a voxel subsampling strategy to efficiently obtain reliable LiDAR references when implementing our GMA-Conv." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 252, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 384 + ], + "type": "text", + "content": "Finally, with the resulting multi-modal voxel features from multiple scales, we further associate them with cascade connections across scales to aggregate multi-granularity information. With the above designs, the camera semantics encapsulated in the virtual points are consistently combined with LiDAR points, and thereby providing a stronger multi-modal feature representation for boosting the 3D object detection. As shown in the Table 3, with 100 times fewer generated virtual points than the two BEVFusion methods [13, 15], our MSMDFusion can still achieve state-of-the-art performances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "type": "text", + "content": "In summary, our contributions lie in threefold: (1) We propose a novel MSMDFusion approach, which encourages sufficient LiDAR-Camera feature fusion in the multiscale voxel space. (2) Within each scale, we propose a Multi-Depth Unprojection strategy (MDU) to promote virtual points generation with better locations and semantics by fully leveraging depth of pixels, as well as a Gated Modality-Aware Convolution (GMA-Conv) to achieve fine-grained controllable multi-modal interaction. (3) Extensive experimental results on the large-scale nuScenes dataset prove the effectiveness of our MSMDFusion and its components. We achieve state-of-the-art performances with " + }, + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "type": "inline_equation", + "content": "71.5\\%" + }, + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "type": "inline_equation", + "content": "74.0\\%" + }, + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "type": "text", + "content": " NDS on the challenging nuScenes detection track using single model1. When combining the simple greedy tracking strategy [24], our method also achieves strong tracking results with " + }, + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "type": "inline_equation", + "content": "74.0\\%" + }, + { + "bbox": [ + 304, + 385, + 548, + 577 + ], + "type": "text", + "content": " AMOTA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 590, + 392, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 392, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 392, + 602 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 611, + 547, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 547, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 547, + 684 + ], + "type": "text", + "content": "Recently, fusing LiDAR and camera signals in 3D detection has attracted increasing interest. Early works [3, 6, 12, 19, 20, 26] project 3D queries to 2D camera images for collecting useful semantics. Under such a paradigm, MV3D [3] and AVOD [7] associate 3D proposals with 2D RoI features. While PointPainting [19] and PointAugment-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1We do not use Test-Time Augmentation (TTA) and model ensemble technique." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21644" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 293 + ], + "type": "image", + "image_path": "1439e75b70d6e192392497352d904d9fcd101bf660b9b3d9cacbabe687410320.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 301, + 548, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 548, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 548, + 369 + ], + "type": "text", + "content": "Figure 2. The framework of our proposed MSMDFusion. Given a LiDAR point cloud and corresponding multi-view camera images, MSMDFusion first separately transforms them into the voxel space and obtains multi-scale LiDAR and camera features. Then, the LiDAR and camera features at each scale interact through our proposed Multi-Depth Unprojection strategy (MDU) and Gated Modality-Aware Convolution block (GMA-Conv, details in Fig.4). The outputs of GMA-Conv at each scale are further aggregated for combining multi-granularity information. Finally, the resulting multi-modal and LiDAR voxel features are flattened into the BEV space for the final prediction. " + }, + { + "bbox": [ + 46, + 301, + 548, + 369 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 301, + 548, + 369 + ], + "type": "text", + "content": " represents voxel addition and \"spconv\" indicates 3D sparse convolution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 389, + 289, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 289, + 689 + ], + "type": "text", + "content": "ing [20] directly decorate raw 3D points with 2D semantics. EP-Net [6] and 3D-CVF [26] perform multi-modal fusion at both point level and proposal level. However, since 3D points are inherently sparse, such a hard association approach wastes the dense semantic information in 2D features. Recently, multi-modal 3D detectors [13, 15, 23, 25] lift dense 2D seeds to 3D space for learning the 2D-3D joint representation in a shared space. Two BEVFusion methods [13, 15] densely lift every image feature pixel into 3D space and then encode these lifted points as another BEV (bird's-eye view) map to fuse with the BEV map obtained using the original 3D point cloud. LiDAR and camera signals will finally share a common representation space, but they do not interact at any stage before BEV fusion. MVP [25] selects seeds from foreground objects and unprojected them into the LiDAR frame for points cloud augmentation. With the seeds sampling strategy similar to MVP, UVTR [9] constructs both LiDAR and camera branches in the voxel space to maintain the modality-specific information, and then interact them with voxel add operation. However, the LiDAR and camera branches are constructed in a single scale of voxel space, losing the benefits of different levels of feature abstraction in multi-scale voxel space. Moreover, merging LiDAR and camera voxels with simple addition limits their fine-grained interaction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 388, + 362, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 362, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 362, + 400 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 407, + 432, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 432, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 432, + 419 + ], + "type": "text", + "content": "3.1. Framework Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 426, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 629 + ], + "type": "text", + "content": "An overall view of MSMDFusion is shown in Fig.2. Given a LiDAR point cloud and corresponding multiview camera images as inputs, MSMDFusion first extracts multi-scale features from both modalities in the voxel space. Then, LiDAR-camera interaction is performed within multi-scale voxel space to properly combine multi-granularity information from both modalities. At each scale, we specifically design a Multi-Depth Unprojection (MDU) strategy to obtain high-quality virtual points in the voxel space, and a Gated Modality-Aware Convolution block (GMA-Conv) for effective LiDAR-camera interaction and fusion. We also introduce cross-scale connections to progressively combine features of different granularities. Afterward, the deeply interacted multi-modal features, together with the LiDAR features, are transformed into the BEV space and fed to the BEV encoder and detection head for the final prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 635, + 515, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 515, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 515, + 647 + ], + "type": "text", + "content": "3.2. LiDAR and Camera Feature Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "We first extract high-level features from the raw inputs of LiDAR and camera. For a given LiDAR point cloud, we voxelize the points and extract their features in the voxel space using a set of 3D sparse convolutional blocks [4] following prior arts [1, 13, 15]. Each convolutional block out" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21645" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 83, + 155, + 156 + ], + "blocks": [ + { + "bbox": [ + 57, + 74, + 167, + 83 + ], + "lines": [ + { + "bbox": [ + 57, + 74, + 167, + 83 + ], + "spans": [ + { + "bbox": [ + 57, + 74, + 167, + 83 + ], + "type": "text", + "content": "large depth bias moderate depth bias" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 83, + 155, + 156 + ], + "lines": [ + { + "bbox": [ + 59, + 83, + 155, + 156 + ], + "spans": [ + { + "bbox": [ + 59, + 83, + 155, + 156 + ], + "type": "image", + "image_path": "7eca596c2913e40ab3a1763b89398cbca7e2c909445143e62e77b2f288f402f7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 156, + 146, + 164 + ], + "lines": [ + { + "bbox": [ + 53, + 156, + 146, + 164 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 146, + 164 + ], + "type": "text", + "content": "(a) Nearest neighbor retrieval" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 178, + 287, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 287, + 243 + ], + "type": "text", + "content": "Figure 3. Comparison of the effects of retrieving reference points from nearest neighbor (MVP) and K-nearest neighbor (Ours). The blue circle and yellow star represent seeds and reference points, respectively. To keep the figure concise, we only depict the case that " + }, + { + "bbox": [ + 46, + 178, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 2" + }, + { + "bbox": [ + 46, + 178, + 287, + 243 + ], + "type": "text", + "content": ". Our strategy can obtain more reliable depths by exploring more neighbors." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 175, + 75, + 272, + 156 + ], + "blocks": [ + { + "bbox": [ + 175, + 75, + 272, + 156 + ], + "lines": [ + { + "bbox": [ + 175, + 75, + 272, + 156 + ], + "spans": [ + { + "bbox": [ + 175, + 75, + 272, + 156 + ], + "type": "image", + "image_path": "dbd763435d95891ae079e5fda7da272a8252a336cd42646e1d1028298d998970.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 157, + 282, + 166 + ], + "lines": [ + { + "bbox": [ + 160, + 157, + 282, + 166 + ], + "spans": [ + { + "bbox": [ + 160, + 157, + 282, + 166 + ], + "type": "text", + "content": "(b) K Nearest neighbor retrieval (Ours)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 266, + 286, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 286, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 286, + 313 + ], + "type": "text", + "content": "puts voxel features of different scales, representing different levels of abstraction for point cloud, and these multiscale features will further interact with their 2D counterparts through our proposed GMA-Conv blocks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 315, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 286, + 445 + ], + "type": "text", + "content": "For the multi-view camera images, we adopt ResNet50 with FPN as the image backbone to extract multi-scale image features, which contains rich object semantics. However, it is non-trivial to transfer these useful semantics to the 3D voxel field, since images are inherently 2D data. Motivated by recent works [13, 15, 25] which select seeds from the image plane and estimate their depths to lift them as 3D virtual points, we further propose a Multi-Depth Unprojection strategy to lift 2D seeds, which could mitigate the drawbacks of prior works. The details will be presented in the following subsections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 454, + 255, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 255, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 255, + 467 + ], + "type": "text", + "content": "3.3. Multi-Modal Interaction in Voxel Space" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 473, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 286, + 641 + ], + "type": "text", + "content": "The extracted LiDAR and camera features from the previous stage respectively focus on geometric and semantic information, hence the goal of multi-modal interaction is to properly fuse these features into a unified rich representation of the scene. As shown in the Fig.2, the components of LiDAR-camera interaction within each scale in the voxel space are conceptually simple: 1) Multi-Depth Unprojection (MDU) strategy is introduced to improve the quality of the generated virtual points and 2) Gated Modality-Aware Convolution (GMA-Conv) block is designed for controllable LiDAR-camera fusion. In the following parts, we will first elaborate on MDU and GMA-Conv within a specific scale, then on top of them, we describe the complete multiscale fusion process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 658, + 192, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 192, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 192, + 670 + ], + "type": "text", + "content": "3.3.1 Multi-Depth Unprojection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "To lift a 2D seed in the pixel space to the 3D space (i.e., the unprojection operation), the depth associated with each seed should be estimated [21]. To obtain an initial reli" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "able depth estimation, we use the strategy from MVP [25] as the basis as shown in Fig.3(a). Formally, given a point cloud " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{(x_i,y_i,z_i)\\}_{i = 1}^{N_P}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " and multi-view images as inputs, MVP first projects 3D points onto 2D images and preserves points falling within the 2D foreground instance masks. These 3D points are kept as reference points to provide depth, and they are denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{(u_i,v_i,d_i)\\}_{i = 1}^{N_R}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " are pixel coordinates and " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " is the real depth. Then, a set of seeds " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "S = \\{(u_i^s,v_i^s)\\}_{i = 1}^{N_S}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " are uniformly sampled from each instance mask on the image, and each of them will retrieve a real depth from its nearest reference point as its estimated depth. Finally the sampled seeds with estimated depths are unprojected to 3D as virtual points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 228, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 382 + ], + "type": "text", + "content": "Though effective, such a strategy ignores the fact that spatial proximity in 2D images can not be guaranteed in 3D, and this may lead to inaccurate depth estimation as shown in the red circle of Fig.3(a). Toward this end, we propose to equip every seed with multiple depths by retrieving K-nearest reference points, which can be regarded as a soft strategy to achieve more reliable depth estimation. As shown in the green circle of the Fig.3(b), each seed can be unprojected using multiple (K) depths from neighboring reference points, which generates K virtual points to improve its recall of the actual 3D points. We provide more detailed discussion on the effect of multi-depth seeds in Sec.4.3.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "text", + "content": "These generated virtual points will be further decorated to make the best use of semantics encapsulated in images. But different from MVP [25] that simply decorates virtual points with class scores predicted by a pretrained 2D detector, we make an early interaction between image features and sparse depth information from reference points to generate depth-aware semantic features. These depth-aware semantic features are then used to decorate virtual points in an adaptive manner controlled by their corresponding seed depths. The whole process can be trained end-to-end. Concretely, we concatenate the camera image feature " + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "text", + "content": " with a sparse depth map containing all depths of reference points " + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "text", + "content": " in that image, and then fuse them with a convolutional layer to obtain the depth-aware semantic feature " + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "inline_equation", + "content": "C^d" + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "text", + "content": ". For the K virtual points generated by the same seed, the semantic feature of the corresponding pixel should also contribute differently along the projection ray in the 3D space [16]. Thus we compute a dynamic weight factor for each individual depth to modulate the semantic feature. The process for a specific seed " + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "inline_equation", + "content": "(u_i^s, v_i^s)" + }, + { + "bbox": [ + 304, + 383, + 545, + 623 + ], + "type": "text", + "content": " can be formally described as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 336, + 627, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 627, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 336, + 627, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s _ {i} ^ {k} = \\operatorname {S i g m o i d} \\left(\\operatorname {L i n e a r} \\left([ C ^ {d} [ u _ {i} ^ {s}, v _ {i} ^ {s} ]; d _ {i} ^ {s, k} ]\\right)\\right) \\tag {1} \\\\ c _ {i} ^ {k} = C ^ {d} [ u _ {i} ^ {s}, v _ {i} ^ {s} ] * s _ {i} ^ {k} \\\\ \\end{array}", + "image_path": "cddeb11fcae9872f73664b6f40637af915a33a3c4ff93707c111de8611ff90ac.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "inline_equation", + "content": "d_{i}^{s,k}" + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "text", + "content": "-th estimated depth of seed " + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "inline_equation", + "content": "(u_{i}^{s}, v_{i}^{s})" + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "inline_equation", + "content": "s_{i}^{k} \\in [0,1]" + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "text", + "content": " represents its weight factor, and " + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "inline_equation", + "content": "c^{k}" + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "text", + "content": " is the modulated semantic feature, which will be used to decorate the corresponding unprojected virtual point " + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "inline_equation", + "content": "(u_{i}^{s}, v_{i}^{s}, d_{i}^{s,k})" + }, + { + "bbox": [ + 304, + 664, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21646" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 496, + 167 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 496, + 167 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 496, + 167 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 496, + 167 + ], + "type": "image", + "image_path": "eefbd71802db6811b3414bcb8786caddd5cfa9cc2ec90d279abfdd39c418ec4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 547, + 199 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 547, + 199 + ], + "type": "text", + "content": "Figure 4. Illustration of the select-then-aggregate process in the Gated Modality-Aware Convolution. Colors indicate different modalities: yellow for camera, blue for LiDAR, and red for LiDAR-camera combined. Cubes with shadows represent voxels selected with gates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 288, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 288, + 267 + ], + "type": "text", + "content": "The resulting virtual points with decorated features will be voxelized to match the output resolution of the corresponding scale of LiDAR branch, in order to perform interaction in the voxel space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 284, + 233, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 233, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 233, + 296 + ], + "type": "text", + "content": "3.3.2 Gated Modality-Aware Convolution" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "type": "text", + "content": "With the resulting LiDAR and camera voxel features that share the same spatial resolution (i.e., the vertically aligned blue and yellow voxels in Fig.2), we aim to interact them in a fine-grained and controllable manner with our proposed Gated Modality-Aware Convolution (GMA-Conv) block in each scale. As shown in Fig.4, we first group the voxels according to their modalities: LiDAR-only (blue voxels), camera-only (yellow voxels), as well as LiDAR and camera combined (red voxels), and they are denoted as " + }, + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "type": "inline_equation", + "content": "f^L" + }, + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "type": "inline_equation", + "content": "f^C" + }, + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "type": "inline_equation", + "content": "f^{LC}" + }, + { + "bbox": [ + 46, + 303, + 287, + 434 + ], + "type": "text", + "content": ", respectively. Then, more fine-grained multi-modal interaction is conducted in a select-then-aggregate manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "type": "text", + "content": "Select. Motivated by the fact that LiDAR-based detectors usually surpass their camera-based counterparts with a large margin, we take LiDAR as the guiding modality to select useful information from camera features. Specifically, we design a gate conditioned on the LiDAR feature to control and update its camera counterpart, which can be formally described as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 529, + 287, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 529, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 101, + 529, + 287, + 546 + ], + "type": "interline_equation", + "content": "\\tilde {f _ {i}} ^ {C} = \\operatorname {R e L U} \\left(\\operatorname {L i n e a r} \\left(f _ {j} ^ {L}\\right)\\right) * f _ {i} ^ {C} \\tag {2}", + "image_path": "2906526bc3d7825127123c22c7cac38bd894ea4ea452dae879c4cc38456a73ea.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{f_i^C}" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": " is the updated camera voxel feature, and " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": " are voxel indices of paired camera and LiDAR voxel features (" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_j^L" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": " can be regarded as the reference voxel for " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_i^C" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": "). The key to make this selection process effective is how to efficiently find reliable reference voxels. A straightforward way is to retrieve the nearest LiDAR voxels for each camera-only voxel " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_i^C" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "f^L" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": ". And similarly for LiDAR-camera voxels " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_k^{LC}" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": ", the reference can be retrieved from both the LiDAR and camera modalities (i.e., " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "f^L" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "f^C" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": "). Although nearby neighbors from more reliable modalities can serve as a good reference, such a retrieval process is infeasible to be implemented with exhaustive pairwise distance calculation, because the computation and memory" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 219, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 254 + ], + "type": "text", + "content": "cost brought by massive amounts of LiDAR and camera voxels is intractable, and we will give more analysis and our solution in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "text", + "content": "Aggregate. With the original LiDAR-only voxels " + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "inline_equation", + "content": "f^L" + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "text", + "content": ", as well as the updated camera-only and LiDAR-camera combined voxels " + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "inline_equation", + "content": "\\tilde{f}^{C}" + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "inline_equation", + "content": "\\tilde{f}^{LC}" + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "text", + "content": ", we first transform their modality-specific representation into an intermediate joint space with 3D sparse convolutions. Then, within the joint space, we further combine all voxels and promote their interaction through another 3D sparse convolution as shown in Fig.4. The resulting multi-modal voxel features are denoted as " + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "inline_equation", + "content": "F^M" + }, + { + "bbox": [ + 304, + 255, + 546, + 362 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 380, + 491, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 491, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 491, + 392 + ], + "type": "text", + "content": "3.3.3 Multi-Scale Progressive Interaction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": "After performing MDU and GMA-Conv within different scales in the voxel space, multi-modal voxel features " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\{F_i^M\\}" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": " are generated for each scale. To further aggregate these multi-granularity information from " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\{F_i^M\\}" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": ", we introduce cascade connections across different scales as shown in Fig.2, which can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 480, + 545, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 480, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 351, + 480, + 545, + 497 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {i + 1} ^ {M} = F _ {i + 1} ^ {M} + \\text {D o w n S a m p l e} \\left(\\hat {F} _ {i} ^ {M}\\right), \\tag {3}", + "image_path": "469fa91c2ae6d6ca3bc99d49ef0e950068c331d7cd7d162fe0fa5d3ed65c1161.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "text", + "content": "where the voxel downsample operation DownSample() is applied to align their spatial resolutions, and " + }, + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\hat{F}_{i+1}^{M}" + }, + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "text", + "content": " is the resulting voxel features that combine multi-scale information from the current and previous scales." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 554, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 613 + ], + "type": "text", + "content": "With the above designs, multi-granularity LiDAR-camera features can be thoroughly interacted at multiple scales in the voxel space. The final output multi-modal voxel features can serve as a powerful representation for the subsequent modules." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 623, + 444, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 444, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 444, + 635 + ], + "type": "text", + "content": "3.4. Prediction in BEV Space" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Following the workflow of prevalent voxel-based 3D object detectors [22, 24, 29], we compress the height of voxel features to transform them into the BEV (bird's-eye view) space. Specifically, we first fuse the LiDAR and multimodal BEV features with a lightweight 2D convolution block. Then, the BEV feature containing enhanced multi" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21647" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 249 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 249 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 249 + ], + "type": "table", + "html": "
MethodModalitymAPNDSCarTruckC.V.BusTrailerBarrierMotor.BikePed.T.C.
PointPillar [8]L40.155.076.031.011.332.136.656.434.214.064.045.6
CenterPoint [24]L60.367.385.253.520.063.656.071.159.530.784.678.4
TransFusion-L [1]L65.570.286.256.728.266.358.878.268.344.286.182.0
PointPainting [19]LC46.458.177.935.815.836.237.360.241.524.173.362.4
3D-CVF [26]LC52.762.383.045.015.948.849.665.951.230.474.262.9
PointAugmenting [20]LC66.871.087.557.328.065.260.772.674.350.987.983.6
MVP [25]LC66.470.586.858.526.167.457.374.870.049.389.185.0
TransFusion [1]LC68.971.787.160.033.168.360.878.173.652.988.486.7
VFF [10]LC68.472.486.858.132.170.261.073.978.552.987.183.8
BEVFusion [13]LC69.271.888.160.934.469.362.178.272.252.289.285.2
BEVFusion [15]LC70.272.988.660.139.369.863.880.074.151.089.286.5
UVTR [9]LC67.171.187.556.033.867.559.573.073.454.886.379.6
DeepInteraction [23]LC70.873.487.960.237.570.863.880.475.454.590.387.0
MSMDFusion (Ours)LC71.574.088.461.035.271.464.280.776.958.390.688.1
", + "image_path": "e60cd27e88908c01c07fe4893a17bc118d9adb9aea046c7f0fde9cc169b2c7b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 111, + 289, + 486, + 376 + ], + "blocks": [ + { + "bbox": [ + 46, + 256, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 546, + 281 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on nuScenes test set. Note that these are all single-model results without ensemble or test-time augmentation. We highlight the best performances across all methods with bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 289, + 486, + 376 + ], + "lines": [ + { + "bbox": [ + 111, + 289, + 486, + 376 + ], + "spans": [ + { + "bbox": [ + 111, + 289, + 486, + 376 + ], + "type": "table", + "html": "
MethodModalityAMOTA ↑AMOTP ↓TP ↑FP ↓FN ↓Recall ↑
CenterPoint [24]L63.80.60695877186122292867.5
TransFusion [1]LC71.80.55196775162322184675.8
UVTR [9]LC70.10.68698434156152019075.0
BEVFusion [15]*LC74.10.40399664199971939577.9
MSMDFusion (Ours)LC74.00.54998624147891985376.3
", + "image_path": "6fd593ed2264cf4f64064caf160348ecf92822b74438fbc415edf39aee5af12e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 546, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 384, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 546, + 418 + ], + "type": "text", + "content": "Table 2. Comparison with leading detectors that use a greedy tracker on nuScenes tracking task. * represents that BEVFusion uses an ensemble model, while we only use a single model. We use **bold** to denote the best performances among all approaches with a single model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 430, + 288, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 455 + ], + "type": "text", + "content": "modality information is fed into a conventional BEV encoder and a detection head for the final prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 466, + 128, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 128, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 128, + 479 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 486, + 164, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 164, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 164, + 499 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 505, + 167, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 167, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 167, + 516 + ], + "type": "text", + "content": "4.1.1 Dataset and Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": "The nuScenes [2] dataset is a large-scale autonomous driving benchmark including 10,000 driving scenarios in total, which are split into 700, 150, and 150 scenes for training, validation, and testing, respectively. For detection, nuScenes defines a set of evaluation protocols, including the nuScenes Detection Score (NDS), mean Average Precision (mAP), as well as five True Positive (TP) metrics, namely mean Average Translation Error (mATE), mean Average Scale Error (mASE), mean Average Orientation Error (mAOE), mean Average Velocity Error (mAVE) and mean Average Attribute Error (mAAE). We report mAP, which is the mean of the average precision across ten classes under distance thresholds of " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "2\\mathrm{m}" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "inline_equation", + "content": "4\\mathrm{m}" + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": ". NDS is the weighted combination of mAP, mATE, mASE, mAOE, mAVE and mAAE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "Implementation Details We use ResNet-50 [5] with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 430, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 610 + ], + "type": "text", + "content": "FPN [14] as the image backbone and VoxelNet [22] as the LiDAR backbone. We set the image size to " + }, + { + "bbox": [ + 304, + 430, + 545, + 610 + ], + "type": "inline_equation", + "content": "448 \\times 800" + }, + { + "bbox": [ + 304, + 430, + 545, + 610 + ], + "type": "text", + "content": ", and voxel size as " + }, + { + "bbox": [ + 304, + 430, + 545, + 610 + ], + "type": "inline_equation", + "content": "(0.075m, 0.075m, 0.2m)" + }, + { + "bbox": [ + 304, + 430, + 545, + 610 + ], + "type": "text", + "content": " following [1]. To make the best use of multi-scale semantics, we extract features from four levels of the FPN (C2 to C5). Following MVP [25], we use CenterNet2 [28] as 2D detector when generating virtual points, and 50 seeds are selected on each instance unless otherwise specified. Our model training has two stages: (1) We first train a LiDAR-only detector for 20 epochs as our 3D backbone. (2) We then connect the proposed LiDAR-camera fusion modules with the 3D backbone for a joint training of another 6 epochs. The data augmentation strategies and training schedules are the same as prior works [24, 30]. We do not use Test-Time Augmentation (TTA) or multi-model ensemble during inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 619, + 486, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 619, + 486, + 632 + ], + "spans": [ + { + "bbox": [ + 305, + 619, + 486, + 632 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 637, + 546, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 637, + 546, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 637, + 546, + 674 + ], + "type": "text", + "content": "We compare our MSMDFusion with state-of-the-art approaches on the nuScenes test set. Overall, Table 1 shows that our method surpasses all existing methods, and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "text", + "content": "NVPF of BEVFusion [13] is calculated with its official code. BEVFusion [15] reports its NVPF in the paper. The former has a higher NVPF because it generates virtual points from image features of larger resolution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21648" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 273, + 129 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 273, + 129 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 273, + 129 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 273, + 129 + ], + "type": "table", + "html": "
MethodNVPF2mAPNDSFPS
BEVFusion [13]5M69.271.80.7
BEVFusion [15]2M70.272.98.4
MSMDFusion(Ours)16k71.574.02.1
", + "image_path": "e80981cabae98f100d1df3183d9be0c9bb008da38982ae659c894bdb32555a20.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 286, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 286, + 253 + ], + "type": "text", + "content": "achieves the new state-of-the-art performances of " + }, + { + "bbox": [ + 46, + 193, + 286, + 253 + ], + "type": "inline_equation", + "content": "71.5\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 193, + 286, + 253 + ], + "type": "text", + "content": " and 74.0 NDS. Meanwhile, our method maintains consistent performance advantages on most object categories, especially on the challenging category: Bike, where a " + }, + { + "bbox": [ + 46, + 193, + 286, + 253 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 46, + 193, + 286, + 253 + ], + "type": "text", + "content": " absolute AP gain over the strongest competitor is achieved." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 253, + 286, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 468 + ], + "type": "text", + "content": "Since our MSMDFusion and two strong BEVFusion methods all fuse LiDAR and camera signals via generating 3D virtual points from 2D seeds, we specifically compare our method with them in terms of the number of generated virtual points per LiDAR frame as shown in Table 3. The results suggest that although with 100 times fewer generated virtual points (16k vs 2M/5M) than them, our MSMDFusion still outperforms them in terms of both mAP and NDS, which proves that our method can better utilize 2D semantics. Meanwhile, we also compare the FPS of our method and two BEVFusion approaches. BEVFusion [15] can achieve the fastest inference speed due to its efficient camera-to-BEV transformation implemented with CUDA. Without such a hardware-friendly implementation, BEVFusion [13] suffers from processing a huge amount of virtual points, while our MSMDFusion runs much faster although performing multi-scale fine-grained fusion in the voxel space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 286, + 589 + ], + "type": "text", + "content": "Besides, to demonstrate that our MSMDFusion can also generalize to other downstream tasks, we further perform the tracking task on nuScenes. For fair comparison, we only compare with single-model methods (i.e., without TTA or ensemble) using the greedy tracker introduced in [24]. As shown in Table 2, our method consistent outperforms other single model competitors across all evaluation metrics. Meanwhile, although without using ensemble techniques, our method also achieve competitive results with the ensemble BEVFusion [15] method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 185, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 185, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 185, + 611 + ], + "type": "text", + "content": "4.3. Comprehensive Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 616, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 286, + 653 + ], + "type": "text", + "content": "Following UVTR [9], we conduct the following ablation studies on a randomly sampled 1/4 split of nuScenes training set for efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 223, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 223, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 223, + 682 + ], + "type": "text", + "content": "4.3.1 Ablation of proposed components" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies for each of our proposed components as shown in Table 4. We use" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 70, + 528, + 146 + ], + "blocks": [ + { + "bbox": [ + 47, + 138, + 287, + 171 + ], + "lines": [ + { + "bbox": [ + 47, + 138, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 287, + 171 + ], + "type": "text", + "content": "Table 3. Number of Virtual points Per Frame (NVPF) and performance comparison with two strongest methods on nuScenes test set. The FPS is tested on a single NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 70, + 528, + 146 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 528, + 146 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 528, + 146 + ], + "type": "table", + "html": "
#MDU*MDUGMA-ConvmAPNDS
160.2665.62
2✓64.3367.03
3✓65.5968.04
4✓✓65.9768.24
5✓✓66.9368.93
", + "image_path": "d731b2edd95c5de3b7cc551489b8180c81c200ad684eef1813dadbe67f8fb307.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 345, + 200, + 509, + 309 + ], + "blocks": [ + { + "bbox": [ + 305, + 156, + 545, + 189 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 545, + 189 + ], + "type": "text", + "content": "Table 4. Ablation studies for our proposed components on the nuScenes validation set. MDU* is a degraded version of MDU by discarding the depth-aware feature generating process." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 200, + 509, + 309 + ], + "lines": [ + { + "bbox": [ + 345, + 200, + 509, + 309 + ], + "spans": [ + { + "bbox": [ + 345, + 200, + 509, + 309 + ], + "type": "table", + "html": "
#C2C3C4C5mAPNDS
160.2665.62
2✓65.5167.92
3✓64.9967.97
4✓64.9767.66
5✓64.5267.61
6✓✓66.3468.63
7✓✓✓66.8068.86
8✓✓✓✓66.9368.93
", + "image_path": "c46d4e4c858e7765dbba96732f6158e03cb6110d77790a1d78cdff749c4f7697.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 318, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 305, + 318, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 545, + 350 + ], + "type": "text", + "content": "Table 5. Effects of LiDAR-camera interaction at different scales, C2-C5 represent the scales where the spatial resolutions of the features decrease from C2 to C5." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 365, + 544, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 544, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 544, + 388 + ], + "type": "text", + "content": "TransFusion-L [1] as our baseline (#1), and perform multiscale fusion across all experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "text", + "content": "From the results of the Table 4, we have the following observations. (i) Introducing virtual points for multi-modal interaction (#2 and #3) brings evident improvements over the baseline, which proves that fusing dense 2D semantics from virtual points with LiDAR features in the voxel space can significantly enhance the LiDAR-only detector. (ii) An early depth-color interaction (#2 and #3) can provide stronger semantics, which further boosts the benefits of interaction between LiDAR and virtual points. (iii) Using GMA-Conv to select helpful information from virtual points (#4 and #5) can bring further improvements over absorbing all information from virtual points without distinction (#2 and #3). (iv) With all these components combined (#5), the final mAP and NDS are significantly improved from " + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "inline_equation", + "content": "60.26\\%" + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "inline_equation", + "content": "66.93\\%" + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "inline_equation", + "content": "65.62\\%" + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "inline_equation", + "content": "68.93\\%" + }, + { + "bbox": [ + 304, + 389, + 545, + 580 + ], + "type": "text", + "content": ", respectively, which proves the effectiveness of our designs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 598, + 477, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 477, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 477, + 609 + ], + "type": "text", + "content": "4.3.2 Effects of multi-scale interaction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We also investigate the effects of performing LiDAR-camera interaction at different scales. As shown in Table 5, for models with single-scale interaction (#2-#4), interacting at scales of larger spatial resolution generally leads to larger performance gains over the baseline (#1), which mainly benefits from more fine-grained image and LiDAR features. If the multi-modal interaction is performed at more scales (#6-#8), multi-granularity information can be" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21649" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 241, + 147 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 241, + 147 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 241, + 147 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 241, + 147 + ], + "type": "table", + "html": "
KNNVPFmAPNDS
1502.6K65.8668.24
3508.0K66.4968.88
65016.0K66.9368.93
105026.6K66.6168.85
120010.6K65.5968.00
", + "image_path": "7ca241856c47795e9bbfad59dbfd78f6131980b457dfde1cdf93af3b9213e3db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 155, + 288, + 189 + ], + "lines": [ + { + "bbox": [ + 46, + 155, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 155, + 288, + 189 + ], + "type": "text", + "content": "Table 6. Effects of the number of depths per seed (K) and number of seeds per instance (N) in multi-depth unprojection. NVPF represents the number of virtual points per frame." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 208, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 291 + ], + "type": "text", + "content": "integrated into our progressive interaction process to consistently achieve better performances than single-scale models. The complete multi-scale interaction model achieves performance gains of " + }, + { + "bbox": [ + 46, + 208, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\sim 2\\%" + }, + { + "bbox": [ + 46, + 208, + 287, + 291 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 46, + 208, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\sim 1\\%" + }, + { + "bbox": [ + 46, + 208, + 287, + 291 + ], + "type": "text", + "content": " NDS over the best single-scale model, which indicates that multi-scale interaction is beneficial for comprehensive multi-modal fusion." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 301, + 284, + 405 + ], + "blocks": [ + { + "bbox": [ + 50, + 301, + 284, + 405 + ], + "lines": [ + { + "bbox": [ + 50, + 301, + 284, + 405 + ], + "spans": [ + { + "bbox": [ + 50, + 301, + 284, + 405 + ], + "type": "image", + "image_path": "9de000808512cc8e635f2bab6019cb6912ceaa23482df945daaa1c9729196f9b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "type": "text", + "content": "Figure 5. Mean error and recall of the virtual points computed with different numbers (K) of retrieved depths per seed." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 466, + 165, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 165, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 165, + 478 + ], + "type": "text", + "content": "4.3.3 Discussion on MDU" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": "Our proposed MDU is a simple yet effective strategy to lift 2D seeds from image to 3D space, and we would like to further discuss one critical problem: will the multiple depths for a seed introduce geometric noise and diminish the benefits? We conducted a quantitative experiment on the nuScenes dataset to investigate this problem. Recall that original point cloud, their projection on image plane (i.e., reference points) and seeds are denoted as " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": ", respectively. We fist preserve a subset " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^e = \\{(u_i^e,v_i^e,d_i^e)\\}_{i=1}^{N^e}" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " of the reference points " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": ", and use all " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{(u_i^e,v_i^e)\\}" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " as seeds to generate virtual points " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^e" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " by retrieving real depths from the remaining reference points " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R} - \\mathcal{R}^e" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " as done in MDU. Since the real 3D coordinates " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^e" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " of the preserved reference points " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^e" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " are available, we then measure two important metrics for the generated virtual points " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^e" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": ": (1) The mean distance error in 3D space between " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^e" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^e" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": ", which reflects the overall accuracy of the virtual points' coordinates. (2) The recall rate that is the ratio of points in " + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^e" + }, + { + "bbox": [ + 46, + 484, + 288, + 715 + ], + "type": "text", + "content": " which fall within a small neighborhood of at" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": "least one point in " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^e" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": ", which represents MDU's ability to capture unseen real 3D points. We also vary the number of depths per seed (K) to show its effect. As shown in Fig.5, when the number of K increases from 1 to 6, the recall rate of real points improves dramatically, meanwhile, the mean distance error does not increase significantly. But a too large value of K (e.g., 10) can inevitably introduce noise. This proves that with a properly chosen number of depths, MDU does not lead to increased noise in depth estimation and can generate virtual points with reliable depths while also covering more real points than a single-depth strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 208, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 546, + 411 + ], + "type": "text", + "content": "To further inspect the effects of K on the final detection performance, we keep other settings unchanged and vary K to compare performances. As Table 6 shows, increasing K from 1 to 6 can gradually improve these models' performances. As we previously analyzed in Fig.5, such improvements can be mainly attributed to the increased recall and tolerable distance error of the virtual points. However, when K increases to 10, the model's performance drops because the extra noise diminishes the benefits of correctly generated virtual points. Moreover, we also study whether increasing the number of seeds per instance N can achieve similar improvements, and the results when K is set to 1 and N is set to 200 are shown in the last row of Table 6. It can be seen that simply increasing seeds cannot bring improvements as our multi-depth strategy, but can even hurt performance due to more noise from single-depth unprojection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 436, + 378, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 378, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 378, + 449 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 460, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 617 + ], + "type": "text", + "content": "In this paper, we introduce MSMDFusion, a novel LiDAR-camera fusion framework for 3D object detection, which facilitates multi-modal interaction across multiple scales. Our approach employs a Multi-Depth Unprojection (MDU) strategy to obtain reliable virtual points from images, and applies Gated Modality-Aware Convolution (GMA-Conv) at each scale to enable fine-grained, controllable multi-modal LiDAR-camera fusion. We further combine multi-granularity information across scales to generate comprehensive features for the final prediction. Extensive experiments demonstrate the effectiveness of these components, with our method ultimately achieving state-of-the-art performance on the nuScenes dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 641, + 421, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 641, + 421, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 641, + 421, + 654 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "This project was supported by National Key R&D Program of China (No. 2020AAA0140001), NSFC (No. 62072116), and Science and Technology Commission of Shanghai Municipality (No. 21JC1400600)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21650" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 157 + ], + "type": "text", + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "text", + "content": "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "type": "text", + "content": "[4] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "text", + "content": "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "text", + "content": "[6] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In European Conference on Computer Vision, pages 35-52. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 464 + ], + "type": "text", + "content": "[7] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "type": "text", + "content": "[8] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 522, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 287, + 565 + ], + "type": "text", + "content": "[9] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "text", + "content": "[10] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1120-1129, 2022. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[11] Ming Liang, Bin Yang, Yun Chen, Rui Hu, and Raquel Urtasun. Multi-task multi-sensor fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7345-7353, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[12] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[14] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "text", + "content": "[15] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[16] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "text", + "content": "[17] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "type": "text", + "content": "[18] Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA), pages 7276-7282. IEEE, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[19] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[20] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "text", + "content": "[21] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "type": "text", + "content": "[22] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[23] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. arXiv preprint arXiv:2208.11112, 2022. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[24] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[25] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21651" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 306 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In European Conference on Computer Vision, pages 720-736. Springer, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[27] Zhijie Zhang, Yan Liu, Junjie Chen, Li Niu, and Liqing Zhang. Depth privileged object detection in indoor scenes via deformation hallucination. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3456-3464, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "text", + "content": "[28] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "text", + "content": "[29] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 306 + ], + "type": "text", + "content": "[30] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.6" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21652" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_content_list.json b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c88a4db8fba30f0cf1633d788addc5ffbfab719b --- /dev/null +++ b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_content_list.json @@ -0,0 +1,1673 @@ +[ + { + "type": "text", + "text": "MSeg3D: Multi-modal 3D Semantic Segmentation for Autonomous Driving", + "text_level": 1, + "bbox": [ + 106, + 128, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiale Li $^{1}$ Hang Dai $^{2*}$ Hao Han $^{3}$ Yong Ding $^{3*}$", + "bbox": [ + 279, + 179, + 692, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1College of Information Science and Electronic Engineering, Zhejiang University", + "bbox": [ + 163, + 198, + 805, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Computing Science, University of Glasgow", + "bbox": [ + 267, + 215, + 699, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Micro-Nano Electronics, Zhejiang University", + "bbox": [ + 261, + 233, + 707, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Corresponding authors{Hang.Dai@glasgow.ac.uk, dingyong09@zju.edu.cn}.", + "bbox": [ + 171, + 253, + 785, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR and camera are two modalities available for 3D semantic segmentation in autonomous driving. The popular LiDAR-only methods severely suffer from inferior segmentation on small and distant objects due to insufficient laser points, while the robust multi-modal solution is under-explored, where we investigate three crucial inherent difficulties: modality heterogeneity, limited sensor field of view intersection, and multi-modal data augmentation. We propose a multi-modal 3D semantic segmentation model (MSeg3D) with joint intra-modal feature extraction and inter-modal feature fusion to mitigate the modality heterogeneity. The multi-modal fusion in MSeg3D consists of geometry-based feature fusion GF-Phase, cross-modal feature completion, and semantic-based feature fusion SF-Phase on all visible points. The multi-modal data augmentation is reinvigorated by applying asymmetric transformations on LiDAR point cloud and multi-camera images individually, which benefits the model training with diversified augmentation transformations. MSeg3D achieves state-of-the-art results on nuScenes, Waymo, and SemanticKITTI datasets. Under the malfunctioning multicamera input and the multi-frame point clouds input, MSeg3D still shows robustness and improves the LiDAR-only baseline. Our code is publicly available at https://github.com/jialeli1/lidarseg3d.", + "bbox": [ + 75, + 316, + 472, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 712, + 209, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene understanding for safe autonomous driving can be achieved through semantic segmentation using camera 2D images and LiDAR 3D point clouds, which densely classifies each smallest sensing unit of the modality. The image-based 2D semantic segmentation has been developed with massive solid studies [12, 34, 61, 63]. The camera image has rich appearance information about the object but severely suffers from illumination, varying object scales, and indirect applications in the 3D world. Another modality, LiDAR point cloud, drives 3D semantic segmentation with laser points [1, 3, 11, 37]. Unfortunately, irregular", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "laser points are too sparse to capture the details of objects. The inaccurate segmentation appears especially on small and distant objects. The other under-explored direction is using multi-modal data to increase both the robustness and accuracy in 3D semantic segmentation [67].", + "bbox": [ + 496, + 287, + 893, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the conceptual superiority, the development of multi-modal segmentation model is still nontrivial, lagging behind the single-modal methods [27]. We rationally attribute the difficulties to the three following aspects. i) Heterogeneity between modalities. Due to sparse points and dense pixels, point cloud feature extractors [14, 31] and image feature extractors [15, 44, 47] are developed distinctly. Separate intra-modal feature extractors are used to address the heterogeneity [13, 20, 25, 42, 51], but the lack of joint optimization leads to suboptimal features from irrelevant network parameters. ii) Limited intersection on the field of view (FOV) between sensors. Only the points falling into the intersected FOV are geometrically associated with multi-modal data, while simply considering the intersected multi-modal data is not sufficient to be practically applicable. Performing fusion solely in the limited FOV intersection like [20,67] results in unsatisfactory overall segmentation performance as shown in Fig. 1. iii) Multi-modal data augmentation. For example, PMF [67] uses only several 2D augmentations for spatially aligned point cloud projection image and camera RGB image. Under the constraint of modal alignment or 2D representation of point cloud, the multimodal segmentation works [20, 25, 67] discard many useful and critical point cloud augmentation transformations with sacrificed perception performance [36, 48].", + "bbox": [ + 496, + 367, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accordingly, we propose a top-performing multi-modal 3D semantic segmentation method termed MSeg3D, inherently motivated by addressing the aforementioned three technical difficulties. i) Unlike separately extracting modal features in existing methods [13, 25, 42, 51], we jointly optimize intra-modal feature extraction and inter-modal feature fusion to drive maximum correlation and complementarity between heterogeneous modalities. ii) To overcome the disregarded multi-modal fusion outside FOV", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21694", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "intersection [25, 67], we propose a cross-modal feature completion and a semantic-based feature fusion phase SF-Phase to collaborate with the geometry-based feature fusion phase GF-Phase. For points outside the FOV intersection, the former completes the missing camera features using predicted pseudo-camera features, under the explicit guidance of cross-modal supervision. For all the points outside and inside the FOV intersection, the later SF-Phase leverages the multi-head attention [41] to model the semantic relation between point and interested categories so that we can attentively fuse the semantic embeddings aggregated from all the visible fields to each point. iii) The challenging multi-modal data augmentation is reinvigorated by being decomposed as the asymmetric transformations in the LiDAR, camera worlds, and local cameras, which enables flexible permutation to enrich training samples.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the proposed components accumulated in Fig. 1, mIoU and $\\mathrm{mIoU}^1$ are significantly increased while the gaps between mIoU and $\\mathrm{mIoU}^1$ are gradually decreased. Our contributions are four-fold: i) We propose a multi-modal segmentation model MSeg3D with joint intra-modal feature extraction and inter-modal feature fusion, achieving state-of-the-art 3D segmentation performance on the competitive nuScenes [3], Waymo [37], and SemanticKITTI [1] datasets for autonomous driving. The proposed framework won $2^{nd}$ place in the Waymo 3D semantic segmentation challenge at CVPR 2022. ii) We propose a cross-modal feature completion and a semantic-based feature fusion phase. To our best knowledge, it is the first time to address the overlooked and inapplicable multi-modal fusion outside the sensor FOV intersection. iii) By applying augmentation transformations asymmetrically on point cloud and images, the proposed asymmetrical multi-modal data augmentation significantly increases the diversity of multi-modal samples for training model with robust improvements. iv) Extensive experimental analyses on the improvement and robustness of our method clearly investigate our designs.", + "bbox": [ + 75, + 333, + 472, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 662, + 218, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR-only 3D Semantic Segmentation is promoted by SemanticKITTI [1], nuScenes [3], and Waymo [37] datasets. The methods follow the U-Net [33] architecture, but progress from three point cloud representations. i) Point. The point methods derived from PointNet++ [31] cost heavy computation on sampling and gathering disordered neighbors, especially on a large-scale LiDAR point cloud. The major point methods [17, 32, 39, 53] perform well on small synthetic point cloud [56] rather than sparse LiDAR point cloud. ii) 2D images. PolarNetseries [62, 65] and others [26, 40, 46, 49] project 3D point", + "bbox": [ + 75, + 686, + 470, + 854 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa4344f810a7997690cdd94c737b4fce6b7a142acc6ee40926ee7928c15aad3e.jpg", + "image_caption": [ + "Figure 1. Performance (mIoU and mIoU $^1$ ) evaluation on nuScenes [3] (a) and Waymo [37] (b) validation sets. All points are distinguished as points inside the sensor FOV intersection (points inside) and points outside the sensor FOV intersection (points outside). We cumulatively add (“+”) the proposed components." + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 870, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cloud as 2D images in bird's-eye-view and range-view, achieving efficient LiDAR segmentation with 2D CNNs. But the 3D-to-2D projection damages 3D information and performance. iii) 3D Voxel. The state-of-the-art SPVNAS [38], Cylinder3D [66] and SDSeg3D [21] explicitly explore 3D structure information in different 3D coordinate systems, which efficiently perform 3D sparse convolutions on non-empty voxels [8, 54]. The point and image representations also potentially facilitate voxel features in RPVNet [50] and AF2-S3Net [7]. We also perform effective and efficient 3D voxel feature learning.", + "bbox": [ + 496, + 400, + 890, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal 3D Semantic Segmentation is investigated in [13, 20, 25, 67] using LiDAR and camera with unsatisfactory performance. The suboptimal feature sources of RGB value or separately learned CNN feature are warped into point cloud range image as additional input features [20, 25]. Recently, PMF [67] projects point cloud onto the image plane for fusion with image data, which is obviously limited by camera FOV. They discard the fusion and segmentation over points outside. Despite the improved segmentation performance within the FOV, another LiDAR model is required to take responsibility for the remaining area [43, 67]. Such inequity and inconvenience have prevented researchers from focusing on multi-modal 3D semantic segmentation. The image format of point cloud in the aforementioned works limits them to do only image augmentation such as 2D flipping rather than 3D point cloud augmentation. Instead, we point out the necessity of point cloud augmentation for our 3D segmentation network. Unlike another multi-modal 3D object detector PointAugmenting [43] that preserves the point cloud augmentation while keeping the image unchanged, our method starts from both modalities with", + "bbox": [ + 496, + 569, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1In the following text, mIoU1 denotes the segmentation performance evaluated on only the points inside FOV intersection by excluding the points outside like PMF [67] and other methods [20, 25].", + "bbox": [ + 75, + 862, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21695", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/47cad5a6f8e5c2f3cc200593163ae883e4f473d4ed37a6d0a0725bc124295b7d.jpg", + "image_caption": [ + "Figure 2. Overview of our multi-modal 3D semantic segmentation model (MSeg3D). For multi-modal feature fusion, GF-Phase mainly includes the Geometry-based Feature Fusion Module (GFFM), while SF-Phase consists of LiDAR Semantic Feature Aggregation Module (SFAM), camera SFAM, and Semantic-based Feature Fusion Module (SFFM)." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "asymmetric transformations. Besides, 2D3DNet [13] aims to train a 3D model using reduced 3D annotations but more unlabeled data and additional manual 2D annotations to train a 2D model for obtaining 2D segmentation in advance. The 2D segmentation is also painted onto the points as the additional input of 3D model [42]. Such a phased training approach lacks joint 2D-3D optimization. Our motivation, technique, and results are distinguished from them.", + "bbox": [ + 75, + 330, + 472, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 459, + 171, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 484, + 205, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Setup. Let $\\{\\mathbf{P}_{\\mathrm{in}},\\mathbf{X}_{\\mathrm{in}}\\}$ be a multi-modal sample, where $\\mathbf{P}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{point}}\\times C_{\\mathrm{in}}}$ denotes LiDAR point cloud of $N_{\\mathrm{point}}$ points with $C_\\mathrm{in}$ -dimensional input features (e.g., 3D coordinates and reflectance) and $\\mathbf{X}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{cam}}\\times 3\\times H_{\\mathrm{in}}\\times W_{\\mathrm{in}}}$ denotes multi-camera RGB images of $N_{\\mathrm{cam}}$ cameras. Based on sensor calibration, the point with 3D coordinates $(x,y,z)$ can be projected onto the $c$ -th local camera image with the pixel coordinates $(c,u,v)$ . $N_{\\mathrm{cls}}$ semantic categories for each point in the 3D point cloud.", + "bbox": [ + 75, + 507, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR Feature Extraction. Given the input $\\{\\mathbf{P}_{\\mathrm{in}},\\mathbf{X}_{\\mathrm{in}}\\}$ in Fig. 2, we first conduct intra-modal feature extraction with parallel backbones for addressing the heterogeneity between point cloud and images. Due to the superiority of 3D voxel, we perform voxel-based LiDAR feature extraction by a general sparse 3D U-Net [9, 35] in Supplementary Material (SM). The input point cloud is first divided by a quantization step $d$ then grouped into non-empty voxels by averaging the point-wise initial features of the local points inside a voxel. These non-empty voxels $\\mathbf{V}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{voxel}}\\times C_{\\mathrm{in}}}$ constitute the sparse tensor input to point cloud backbone for learning the expressive voxel features $\\mathbf{V}\\in \\mathbb{R}^{N_{\\mathrm{voxel}}\\times C_{\\mathrm{voxel}}}$ .", + "bbox": [ + 75, + 643, + 470, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera Feature Extraction. To enhance the LiDAR features with jointly-optimized camera features, we use a trainable image backbone (HRNet-w48 [44] by default) to non-linearly project $\\mathbf{X}_{\\mathrm{in}}$ as $\\mathbf{X} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times C_{\\mathrm{img}} \\times H \\times W}$ , with down-sampled shape $H \\times W$ and increased channels $C_{\\mathrm{img}}$ .", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The point cloud and image backbones can be flexibly selected from various mature networks. Moreover, we jointly optimize both backbones in the overall 3D segmentation model to learn relevant semantic feature representations from different modalities. The expressive voxel features $\\mathbf{V}$ and image features $\\mathbf{X}$ guarantee the inter-modal feature fusion, which consists of GF-Phase, cross-modal feature completion, and SF-Phase in the right half of Fig. 2.", + "bbox": [ + 496, + 330, + 893, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. GF-Phase: Geometry-based Feature Fusion", + "text_level": 1, + "bbox": [ + 498, + 459, + 870, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The point-centric segmentation motivates us to warp voxel features $\\mathbf{V}$ and image feature maps $\\mathbf{X}$ to points in $\\mathbf{P}_{\\mathrm{in}}$ as point-wise LiDAR features $\\mathbf{F}_{\\mathrm{lidar}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{voxel}}}$ and point-wise camera features $\\mathbf{F}_{\\mathrm{cam}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{img}}}$ by using the geometric association of LiDAR and multi-camera.", + "bbox": [ + 496, + 482, + 890, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning Point Feature From Voxels and Pixels. We devoxelize voxel features $\\mathbf{V}$ as point-wise LiDAR features $\\mathbf{F}_{\\mathrm{lidar}} = [f_{\\mathrm{lidar},i}]_{i=1}^{i=N_{\\mathrm{point}}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{voxel}}}$ for augmenting the point-wise fusion and segmentation. Given a point, we interpolate its point feature from the three nearest neighboring voxels [22]. For the $i$ -th point $(x_i, y_i, z_i)$ with LiDAR feature $f_{\\mathrm{lidar},i}$ in $\\mathbf{F}_{\\mathrm{lidar}}$ , we use the known point-wise image coordinates $(c_i, u_i, v_i)$ to perform bilinear interpolation $\\mathcal{I}$ on the image feature maps $\\mathbf{X}[c_i] \\in \\mathbb{R}^{C_{\\mathrm{img}} \\times H \\times W}$ within the $c_i$ -th local camera as Eq. 1. Zeros are temporarily padded to those points outside the camera FOV, which allows each point to be decorated with the $C_{\\mathrm{img}}$ -dimensional feature $f_{\\mathrm{cam},i}$ . To distinguish points outside, we set a binary mask $\\mathbf{B}$ of $N_{\\mathrm{point}}$ elements with 0 for them and 1 otherwise.", + "bbox": [ + 496, + 558, + 892, + 772 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {c a m}, i} = \\mathcal {I} \\left(\\mathbf {X} \\left[ c _ {i} \\right], \\frac {H}{H _ {\\mathrm {i n}}} u _ {i}, \\frac {W}{W _ {\\mathrm {i n}}} v _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 779, + 890, + 810 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {c a m}} = \\left[ f _ {\\mathrm {c a m}, i} \\right] _ {i = 1} ^ {i = N _ {\\text {p o i n t}}} \\in \\mathbb {R} ^ {N _ {\\text {p o i n t}} \\times C _ {\\mathrm {i m g}}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 811, + 890, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometry-based Feature Fusion Module (GFFM) first projects $f_{\\mathrm{lidar},i}$ and $f_{\\mathrm{cam},i}$ as $C_{\\mathrm{int}}$ -dimensional by using fully connected layers $\\mathcal{F}_{\\mathrm{lidar}}$ and $\\mathcal{F}_{\\mathrm{cam}}$ , and concatenates (\"©\") them together for subsequent learnable", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21696", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fusion using another MLP with $C_{\\mathrm{gfused}}$ output channels as $f_{\\mathrm{gfused},i} = MLP(\\mathcal{F}_{\\mathrm{lidar}}(f_{\\mathrm{lidar},i}) \\odot \\mathcal{F}_{\\mathrm{cam}}(f_{\\mathrm{cam},i}))$ and $\\mathbf{F}_{\\mathrm{gfused}} = [f_{\\mathrm{gfused},i}]_{i=1}^{i=N_{\\mathrm{point}}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{gfused}}}$ . The GF-Phase combines the multi-modal features as $\\mathbf{F}_{\\mathrm{gfused}}$ mainly with geometry clues, so-called the geometry-based feature fusion phase.", + "bbox": [ + 76, + 90, + 468, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. SF-Phase: Semantic-based Feature Fusion", + "text_level": 1, + "bbox": [ + 76, + 176, + 437, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Though the simple yet effective GF-Phase yields appreciable performance gains inside FOV intersection through unbiased considerations of both modalities, three drawbacks remain to be addressed. i) The points outside still cannot retrieve their realistic camera features. ii) The relative importance between modalities varies in different areas [5, 58], which is overlooked in GF-Phase. iii) Explicit relational modeling for category-wise semantic representations [18, 19] is not available. As shown in Fig. 2, we further propose the SF-Phase after GF-Phase. Beyond the geometric association in euclidean space, we aggregate LiDAR features and camera features into category-wise semantic embeddings as $\\mathbf{E}_{\\mathrm{lidar}}$ and $\\mathbf{E}_{\\mathrm{cam}}$ for performing multi-modal fusion in a hidden semantic space.", + "bbox": [ + 75, + 199, + 468, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LiDAR Semantic Feature Aggregation Module (LiDAR SFAM) starts from voxel features $\\mathbf{V}$ and ends with aggregated $N_{\\mathrm{cls}}$ LiDAR semantic embeddings in $\\mathbf{E}_{\\mathrm{lidar}}$ . For $N_{\\mathrm{cls}}$ categories, we assume $\\mathbf{E}_{\\mathrm{lidar}} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times C_{\\mathrm{voxel}}}$ and a distribution matrix $\\mathbf{D}_{\\mathrm{lidar}} \\in (0,1)^{N_{\\mathrm{cls}} \\times N_{\\mathrm{voxel}}}$ . The $\\mathbf{D}[j,i]$ closer to 1 indicates that the $i$ -th voxel is more likely to belong to the $j$ -th category than other voxels, so that it contributes more to describing this category in $\\mathbf{E}_{\\mathrm{lidar}}[j]$ . Otherwise, the $i$ -th voxel should contribute less to $\\mathbf{E}_{\\mathrm{lidar}}[j]$ . Now we can compute $\\mathbf{E}_{\\mathrm{lidar}}$ by matrix product as $\\mathbf{E}_{\\mathrm{lidar}} = \\mathbf{D}_{\\mathrm{lidar}}\\mathbf{V}$ . Two more steps (Eqs. $3 \\sim 4$ ) are required to compute $\\mathbf{D}_{\\mathrm{lidar}}$ . We first predict an intermediate segmentation $\\mathbf{D}_{\\mathrm{lidar}}' \\in \\mathbb{R}^{N_{\\mathrm{voxel}} \\times N_{\\mathrm{cls}}}$ on the voxel features $\\mathbf{V}$ using an MLP-based auxiliary voxel segmentation head $\\mathcal{H}_{\\mathrm{voxel}}$ . We then perform a spatial softmax among voxels to normalize the values into $(0,1)$ for each category.", + "bbox": [ + 75, + 411, + 470, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime} = \\mathcal {H} _ {\\text {v o x e l}} (\\mathbf {V}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 659, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\text {l i d a r}} ^ {T} = \\left[ \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime} [:, j ]\\right) \\right] _ {j = 1} ^ {j = N _ {\\mathrm {c l s}}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 678, + 468, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For more accurate $\\mathbf{D}_{\\mathrm{lidar}}$ , we explicitly guide $\\mathbf{D}_{\\mathrm{lidar}}^{\\prime}$ by using the voxel segmentation supervision in Eq. 15.", + "bbox": [ + 75, + 704, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera Semantic Feature Aggregation Module (Camera SFAM) similarly aggregates $\\mathbf{E}_{\\mathrm{cam}} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times C_{\\mathrm{img}}}$ from image feature maps $\\mathbf{X} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times C_{\\mathrm{img}} \\times H \\times W}$ using another distribution matrix $\\mathbf{D}_{\\mathrm{cam}} \\in (0,1)^{N_{\\mathrm{cls}} \\times N_{\\mathrm{pixel}}}$ and another auxiliary image segmentation head $\\mathcal{H}_{\\mathrm{img}}$ implemented with the simple FCN head [34]. The difference lies in that the camera SFAM is designed with all the $N_{\\mathrm{pixel}}$ pixels of image feature maps across $N_{\\mathrm{cam}}$ local cameras, where $N_{\\mathrm{pixel}}$ is $N_{\\mathrm{cam}} \\times H \\times W$ .", + "bbox": [ + 75, + 734, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $\\mathcal{H}_{\\mathrm{img}}$ first predicts the intermediate segmentation $\\mathbf{D}_{\\mathrm{img}}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times N_{\\mathrm{cls}} \\times H \\times W}$ by $\\mathcal{H}_{\\mathrm{img}}(\\mathbf{X})$ . Note that $\\mathbf{D}_{\\mathrm{img}}^{\\prime}$", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\mathbf{X}$ are then rearranged into $\\mathbf{D}_{\\mathrm{cam}}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times N_{\\mathrm{pixel}}}$ and $\\mathbf{X}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{pixel}} \\times C_{\\mathrm{img}}}$ by transposing and reshaping operations. The final camera semantic embeddings $\\mathbf{E}_{\\mathrm{cam}}$ in Eq. 7 is the matrix product of $\\mathbf{D}_{\\mathrm{cam}}$ and $\\mathbf{X}^{\\prime}$ , where the normalized $\\mathbf{D}_{\\mathrm{cam}}$ is the result of applying spatial softmax on $\\mathbf{D}_{\\mathrm{cam}}^{\\prime}$ as Eq. 6.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\mathrm {i m g}} ^ {\\prime} = \\mathcal {H} _ {\\mathrm {i m g}} (\\mathbf {X}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 186, + 890, + 205 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\mathrm {c a m}} = \\left[ \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {\\mathrm {c a m}} ^ {\\prime} [ j,: ]\\right) \\right] _ {j = 1} ^ {j = N _ {\\mathrm {c l s}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 208, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {\\mathrm {c a m}} = \\mathbf {D} _ {\\mathrm {c a m}} \\mathbf {X} ^ {\\prime}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 229, + 890, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the image segmentation annotations are unavailable for guiding the $\\mathbf{D}_{\\mathrm{img}}^{\\prime}$ in such 3D semantic segmentation datasets [3, 37]. Nevertheless, we also address this by the cross-modal semantic supervision scheme, which is detailed in Eq. 16 and Eq. 17.", + "bbox": [ + 496, + 253, + 890, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic-based Feature Fusion Module (SFFM). As shown in the right dashed box in Fig. 2, given the input of point-wise geometry-based fused features $\\mathbf{F}_{\\mathrm{gfused}}$ , category-wise semantic embeddings $\\mathbf{E}_{\\mathrm{lidar}}$ and $\\mathbf{E}_{\\mathrm{cam}}$ , the SFFM first projects them into a $C_{\\mathrm{sfused}}$ -dimensional space as $\\mathbf{F}_{\\mathrm{proj}} = Proj_1(\\mathbf{F}_{\\mathrm{gfused}})$ , and $\\mathbf{E} = Proj_2(\\mathbf{E}_{\\mathrm{lidar}}) \\odot Proj_3(\\mathbf{E}_{\\mathrm{cam}})$ , respectively. Before stacking $K$ gray blocks in SFFM, we concatenate the projected multi-modal semantic embeddings together as $\\mathbf{E} \\in \\mathbb{R}^{2N_{\\mathrm{cls}} \\times C_{\\mathrm{sfused}}}$ , which can be regarded as a dictionary that describes the typical characteristics for each category from the perspectives of LiDAR and cameras.", + "bbox": [ + 496, + 328, + 892, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each gray block, we model the semantic relations among all the $2N_{\\mathrm{cls}}$ category-wise semantic embeddings using the Multi-Head Self-Attention (MHSA) [41] as Eq. 8, where Norm indicates the LayerNorm operation. Let $C_{\\mathrm{shsa}}$ be $C_{\\mathrm{sfused}} / N_H$ , the MHSA can be decomposed into $N_{\\mathrm{H}}$ Single-Head Self-Attention (SHSA) operations with $\\mathbf{E}_h \\in \\mathbb{R}^{2N_{\\mathrm{cls}} \\times C_{\\mathrm{shsa}}}$ as", + "bbox": [ + 496, + 494, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {\\mathrm {m h s a}} = \\operatorname {N o r m} (\\mathbf {E} + M H S A (\\mathbf {E})), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 606, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM H S A (\\mathbf {E}) = \\left[ S H S A \\left(\\mathbf {E} _ {h}\\right) \\right] _ {h = 1} ^ {h = N _ {\\mathrm {H}}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 625, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS H S A \\left(\\mathbf {E} _ {h}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {h} \\mathbf {K} _ {h} ^ {T}}{\\sqrt {C _ {\\mathrm {s h s a}}}}\\right) \\mathbf {V} _ {h}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 647, + 890, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $\\mathbf{E}$ includes $\\mathbf{E}_{\\mathrm{lidar}}$ and $\\mathbf{E}_{\\mathrm{cam}}$ , the attention matrix $Softmax\\left(\\frac{\\mathbf{Q}_h\\mathbf{K}_h^T}{\\sqrt{C_{\\mathrm{shsa}}}}\\right)\\in \\mathbb{R}^{2N_{\\mathrm{cls}}\\times 2N_{\\mathrm{cls}}}$ intuitively includes four parts of relation modeling on $\\mathbf{E}_{\\mathrm{lidar}}\\rightarrow \\mathbf{E}_{\\mathrm{lidar}},\\mathbf{E}_{\\mathrm{lidar}}\\rightarrow \\mathbf{E}_{\\mathrm{cam}},$ $\\mathbf{E}_{\\mathrm{cam}}\\rightarrow \\mathbf{E}_{\\mathrm{lidar}}$ , and $\\mathbf{E}_{\\mathrm{cam}}\\rightarrow \\mathbf{E}_{\\mathrm{cam}}$", + "bbox": [ + 496, + 685, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After updating $\\mathbf{E}$ as $\\mathbf{E}_{\\mathrm{mhsa}}$ , we perform the further fusion between the point-wise projected features $\\mathbf{F}_{\\mathrm{proj}}$ and the semantic embeddings $\\mathbf{E}_{\\mathrm{mhsa}}$ , using the Multi-Head Cross-Attention (MHCA) for semantic relation modeling in Eq 12 and the Feed-Forward Network (FFN) for feature embedding in Eq. 11, respectively.", + "bbox": [ + 496, + 750, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {s f u s e d}} = \\operatorname {N o r m} \\left(\\mathbf {F} _ {\\text {m h c a}} + F F N \\left(\\mathbf {F} _ {\\text {m h c a}}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 845, + 890, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {m h c a}} = \\operatorname {N o r m} \\left(\\mathbf {F} _ {\\mathrm {p r o j}} + M H C A \\left(\\mathbf {F} _ {\\mathrm {p r o j}}, \\mathbf {E} _ {\\mathrm {m h s a}}, \\mathbf {E} _ {\\mathrm {m h s a}}\\right)\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 864, + 890, + 895 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21697", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike Eq. 8, the $MHCA$ in Eq. 12 sets the query $\\mathbf{Q}$ from point-wise feature $\\mathbf{F}_{\\mathrm{proj}}$ , and sets the key $\\mathbf{K}$ and value $\\mathbf{V}$ from semantic embeddings $\\mathbf{E}_{\\mathrm{mhsa}}$ . Thus, an attention matrix with shape $N_{\\mathrm{point}} \\times 2N_{\\mathrm{cls}}$ is computed to attentively aggregate the more important semantic embeddings respective to individual points, beyond the paired geometric association in GF-Phase. Following the common usages of multi-head attention [4], $K$ (i.e., 6) gray blocks are stacked.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion on SFFM. As a useful and general technique [4,28,47], multi-head attention [41], although not proposed in this paper, is effectively tailored to multi-modal feature fusion with different motivations and designs by us. i) In Eq. 12, the $MHCA$ enables the point-wise feature to attend to multi-modal semantic embeddings, so that both points inside and outside can be consistently supported by expressive multi-modal semantic embeddings. ii) The $MHCA$ attention matrix computes the relative importance of two modalities to each point, improving the unbiased considerations of modalities in GF-Phase. iii) The $MHSA$ in Eq. 8 explicitly models the category-wise intra-modal and inter-modal semantic relations, deriving the commonality learning. iv) The LiDAR SFAM and Camera SFAM aggregate the long sequences of voxel features $V$ and image features $X'$ into the short sequences of $E_{\\text{lidar}}$ and $E_{\\text{cam}}$ with $N_{\\text{voxel}} / N_{\\text{cls}}$ and $N_{\\text{pixel}} / N_{\\text{cls}}$ times, which enables efficient computation of the multi-head attention.", + "bbox": [ + 75, + 212, + 470, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Cross-modal Feature Completion", + "text_level": 1, + "bbox": [ + 76, + 496, + 369, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, a cross-modal feature completion module with pixel-to-point loss $\\mathbf{L}_{\\mathrm{pixel2point}}$ (Eq. 13) is set, where the point-wise pseudo-camera feature $\\mathbf{F}_{\\mathrm{pcam}}$ is mapped from the point-wise LiDAR features $\\mathbf{F}_{\\mathrm{lidar}}$ by another MLP-based $\\mathcal{H}_{\\mathrm{pcam}}$ as $\\mathcal{H}_{\\mathrm{pcam}}(\\mathbf{F}_{\\mathrm{lidar}})$ . In practice, we compute the mean square error loss $\\mathcal{L}_{\\mathrm{mse}}$ between the $\\mathbf{F}_{\\mathrm{pcam}}$ and $\\mathbf{F}_{\\mathrm{cam}}$ of the points inside for learning from the correctly paired cross-modal features relationship.", + "bbox": [ + 75, + 520, + 468, + 641 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {\\text {p i x e l 2 p o i n t}} = \\mathcal {L} _ {\\mathrm {m s e}} \\left(\\mathbf {B F} _ {\\mathrm {p c a m}}, \\mathbf {B D e t a c h} \\left(\\mathbf {F} _ {\\mathrm {c a m}}\\right)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 655, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {c a m}} [ i,: ] = \\mathbf {F} _ {\\mathrm {p c a m}} [ i,: ] | _ {i \\in \\{j \\mid \\mathbf {B} [ j ] = 0 \\}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 674, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the points outside are ignored by the binary mask $\\mathbf{B}$ , which is mentioned in GF-Phase. The gradients of $\\mathbf{F}_{\\mathrm{cam}}$ are also detached for optimizing the learning of $\\mathbf{F}_{\\mathrm{pcam}}$ .", + "bbox": [ + 75, + 703, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ such a cross-modal feature completion module due to two motivations: i) Optimizing $\\mathbf{L}_{\\mathrm{pixel2point}}$ forces LiDAR features $\\mathbf{F}_{\\mathrm{lidar}}$ to imitate camera features $\\mathbf{F}_{\\mathrm{cam}}$ with $\\mathcal{H}_{\\mathrm{pcam}}$ , then the learned pseudo-image features $\\mathbf{F}_{\\mathrm{pcam}}$ can be switched to replace the padded zeros in $\\mathbf{F}_{\\mathrm{cam}}$ as Eq. 14 in the inference stage, which serves as the cross-modal feature completion to enhance the feature learning. Thus, we further reduce feature gaps between points outside and inside. ii) $\\mathcal{H}_{\\mathrm{pcam}}$ transfers rich appearance priors from the camera branch to the LiDAR branch with an effective", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "consistency constraint for enhancing the intra-modal feature learning in the training stage.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Cross-modal Semantic Supervision", + "text_level": 1, + "bbox": [ + 500, + 130, + 803, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Supervision. Let $\\mathbf{Y}$ of $N_{\\mathrm{point}}$ elements be the point-wise 3D semantic segmentation labels. The value of $\\mathbf{Y}[i]$ is in $[0, N_{\\mathrm{cls}} - 1]$ , where 0 denotes the ignored category. An MLP based point segmentation head $\\mathcal{H}_{\\mathrm{point}}$ is built on $\\mathbf{F}_{\\mathrm{sfused}}$ for the 3D segmentation prediction $\\hat{\\mathbf{Y}} = \\mathcal{H}_{\\mathrm{point}}(\\mathbf{F}_{\\mathrm{sfused}})$ . Following [6, 21, 66], we adopt point loss $\\mathbf{L}_{\\mathrm{point}}$ as a combination of cross-entropy loss $\\mathcal{L}_{\\mathrm{ce}}$ and lovasz-softmax loss $\\mathcal{L}_{\\mathrm{lovasz}}$ [2] as $\\mathcal{L}_{\\mathrm{ce}}(\\hat{\\mathbf{Y}}, \\mathbf{Y}) + \\mathcal{L}_{\\mathrm{lovasz}}(\\hat{\\mathbf{Y}}, \\mathbf{Y})$ .", + "bbox": [ + 496, + 152, + 890, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point-to-voxel Supervision. For guiding the $\\mathbf{D}_{\\mathrm{lidar}}^{\\prime}$ in Eq. 3, the voxel loss $\\mathbf{L}_{\\mathrm{p2v}}$ is defined as:", + "bbox": [ + 498, + 277, + 890, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {\\mathrm {p} 2 \\mathrm {v}} = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime}, \\mathbf {Y} _ {\\mathrm {p} 2 \\mathrm {v}}\\right) + \\mathcal {L} _ {\\text {l o v a s z}} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime}, \\mathbf {Y} _ {\\mathrm {p} 2 \\mathrm {v}}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 318, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the voxel labels $\\mathbf{Y}_{\\mathrm{p2v}}$ of $N_{\\mathrm{voxel}}$ elements can be determined from the labels of points in the voxel. To avoid ambiguity in $\\mathbf{L}_{\\mathrm{p2v}}$ , the label of the voxel containing points of multiple categories is set to zero as the ignored category.", + "bbox": [ + 496, + 345, + 890, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point-to-pixel Supervision. The problem of how to use only point labels to correctly guide $\\mathbf{D}_{\\mathrm{img}}^{\\prime}$ (Eq. 5) is still unsolved. For $\\mathbf{D}_{\\mathrm{img}}^{\\prime}$ of each training sample, we initialize a $N_{\\mathrm{cam}} \\times H \\times W$ map $\\mathbf{Y}_{\\mathrm{point2pixel}}$ of all zeros as the image segmentation label. We then project each point onto the down-sampled image planes and retrieve the point label $\\mathbf{Y}[i]$ for the corresponding nearest $(\\langle \\cdot \\rangle)$ pixel as Eq. 16. Although the generated image label $\\mathbf{Y}_{\\mathrm{point2pixel}}$ is sparse, it provides sufficient supervision. In SM, Fig. S2 and Fig. S3 provide visualizations of $\\mathbf{D}_{\\mathrm{img}}^{\\prime}$ to show that these sparse point-to-pixel supervisions can propagate to other regions with similar appearance, which guides $\\mathbf{D}_{\\mathrm{img}}^{\\prime}$ correctly. Since $\\mathbf{Y}_{\\mathrm{point2pixel}}$ is sparse, we only compute the basic cross-entropy loss $\\mathcal{L}_{\\mathrm{ce}}$ in Eq. 17.", + "bbox": [ + 496, + 406, + 892, + 618 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {\\text {p o i n t 2 p i x e l}} \\left[ c _ {i}, \\left\\langle \\frac {H}{H _ {\\mathrm {i n}}} u _ {i} \\right\\rangle , \\left\\langle \\frac {W}{W _ {\\mathrm {i n}}} u _ {i} \\right\\rangle \\right] = \\mathbf {Y} [ i ], \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 626, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {\\text {p o i n t 2 p i x e l}} = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {D} _ {\\mathrm {i m g}} ^ {\\prime}, \\mathbf {Y} _ {\\text {p o i n t 2 p i x e l}}\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 660, + 890, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total Loss Function. To jointly optimize all the modules in the proposed segmentation model, the total loss combines loss terms across different modalities into $\\mathbf{L}$ as $\\alpha_{1}\\mathbf{L}_{\\mathrm{point}} + \\alpha_{2}\\mathbf{L}_{\\mathrm{p2v}} + \\alpha_{3}\\mathbf{L}_{\\mathrm{point2pixel}} + \\alpha_{4}\\mathbf{L}_{\\mathrm{pixel2point}}$ , where $\\alpha_{1} \\sim \\alpha_{4}$ are set as 1.0, 1.0, 0.5, 1.0 to balance loss terms.", + "bbox": [ + 496, + 686, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Asymmetric Multi-modal Data Augmentation", + "text_level": 1, + "bbox": [ + 500, + 771, + 885, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The segmentation output and our multi-modal fusion mechanism are point-centered. In Eq. 1 and Eq. 16, the point and the pixel always can be bridged by the coordinate pair of $(x_{i},y_{i},z_{i})$ and $(c_i,u_i,v_i)$ , as long as we pre-compute the coordinate pairs and keep the order of coordinate pairs among all the points in synchronization. With this ordering constraint, we have the flexibility", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21698", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/340afb42fa332033cc19eba41c2a05b36f1401801c965b99d52e39233a930221.jpg", + "table_caption": [ + "Table 1. Data augmentation transformations on LiDAR (L) point cloud and camera (C) images. We only treat random flipping as a naive symmetric transformation that can be simply applied to both modalities simultaneously." + ], + "table_footnote": [], + "table_body": "
L-only·Global rotation around the Z axis with a random angle in [−π/4, +π/4]. \n·Global translation with a random vector (Δx, Δy, Δz) sampled from a Gaussian distribution with mean zero and the standard deviation 0.5. \n·Global scaling with a random scaling factor in [0.95, 1.05]. \n…
Symmetric·Random flipping along the X, Y axis with probability 0.5.
C-only·Scaling with a random ratio in [1.0, 1.5]. \n·Horizontal rotation with a random angle in [−1°, 1°]. \n·Random cropping with a size (Hin, Win). \n·Color jitter from torchvision [30] with random parameters of 0.3 for brightness, 0.3 for contrast, 0.3 for saturation, 0.1 for hue. \n·JPEG with a random compression ratio in [30, 70] and probability 0.5 [64]. \n…
", + "bbox": [ + 76, + 146, + 470, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to decouple multi-modal data augmentation: i) We can preserve all the LiDAR transformations in Tab. 1 with the images and the coordinates $(c, u, v)$ unchanged. ii) Once coordinates $(c, u, v)$ are synchronously transformed with images, we can further apply the camera-only transformations in Tab. 1 to images, which are asymmetric to the LiDAR transformations. iii) We can apply independent transformations between the images of local cameras since the image view is naturally sliced by local cameras.", + "bbox": [ + 75, + 306, + 468, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The asymmetry lies in not only the LiDAR and camera world but also the local cameras. Thus, we significantly diversify the multi-modal data augmentation for 3D semantic segmentation using the introduced camera transformations with the preserved LiDAR transformations. More potential transformations can be exploited beyond Tab. 1.", + "bbox": [ + 75, + 443, + 468, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 549, + 202, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 574, + 267, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nuScenes Dataset splits 28,130 training samples, 6,019 validation samples, and 6,008 testing samples [3, 11]. Each nuScenes sample contains relatively sparser point cloud of 32 beams and RGB images captured by 6 cameras: front, front-left, front-right, back, back-left, and back-right. Points outside are projected below the image bottom due to different vertical FOVs of LiDAR and cameras. Following official protocol, the $N_{\\mathrm{cls}}$ is 17. Semantic segmentation annotations are only on point clouds and not on images.", + "bbox": [ + 75, + 597, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Waymo Dataset for 3D semantic segmentation includes 23,691 training samples, 5,976 validation samples, and 2,982 testing samples [37]. Each sample contains the point cloud of 64 beams and the RGB images captured by 5 cameras: front, front-left, front-right, side-left, and side-right. There is no rear camera on Waymo ego-vehicle, hence more points outside unfairly increase the difficulty of multimodal segmentation. The $N_{\\mathrm{cls}}$ is 23. Similarly, semantic segmentation annotations are only on point clouds.", + "bbox": [ + 75, + 734, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SemanticKITTI Dataset is collected by a LiDAR with 64 beams [1]. Following [10,50,66,67], we use sequences", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "00 to 10 (excluding 08) with 19,130 training samples and sequence 08 with 4,071 validation samples. It provides only the images of the front-view camera. The $N_{\\mathrm{cls}}$ is 20.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric is the mean Intersection-over-Union (mIoU) defined as $\\frac{1}{C}\\sum_{c = 1}^{N_{\\mathrm{cls}} - 1}\\frac{TP_c}{TP_c + FP_c + FN_c}$ , where $TP_{c}, FP_{c}, FN_{c}$ denote the number of true positive, false positive, and false negative predictions for the $c$ -th category out of $N_{\\mathrm{cls}} - 1$ valid categories. Besides, nuScenes computes another metric fwIoU by weighting category-wise IoU with the point frequency of its category.", + "bbox": [ + 496, + 137, + 890, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. Although lightweight LiDRA-only baseline models can be trained with larger batchsize and more epochs to obtain relatively better results, we still train all the models under the same schedule: a batch of 32 random samples distributed on 16 Tesla V100 GPUs with 24 epochs. More implementation details on point cloud voxelization, network architecture, model training, and inference are all included in SM.", + "bbox": [ + 496, + 243, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 498, + 376, + 864, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on nuScenes. From Tab. 2, multi-modal methods counterintuitively lag far behind LiDAR-only methods. In the top-20 submissions on nuScenes benchmark [27], there are only four multi-modal methods. Details of Cylinder3D++, SPVCNN++, LIFusion, and CPFusion are not available till submission. Our framework achieves the best mIoU and fwIoU with multi-modal fusion mechanism and asymmetric multi-modal data augmentation. Especially, our method achieves superior performance on challenging small objects such as pedestrians and traffic cones, where the laser points are typically insufficient for LiDAR-only methods. Our method also significantly outperforms all public and private multi-modal submissions in Tab. 2.", + "bbox": [ + 496, + 400, + 890, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Waymo. We perform evaluation on Waymo benchmark [45] in Tab. 3, where the LiDAR sensor provides denser laser points and the multi-camera excludes the rearview. Such denser point cloud and incomplete camera FOV unfairly weaken the advantages of multiple modalities. Although, our method still achieves state-of-the-art performance using a general point cloud backbone network [9, 35]. We believe that the performance can be improved by incorporating stronger point cloud backbone networks from other state-of-the-art LiDAR-only methods.", + "bbox": [ + 496, + 598, + 890, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on SemanticKITTI. While comparisons on the highly competitive nuScenes and Waymo datasets validate the effectiveness of our method, Tab. 4 shows the experimental comparison on SemanticKITTI [1]. Since SemanticKITTI provides only front-view images, PMF [67] evaluates the multi-modal segmentation performance on points inside the FOV intersection. We follow PMF to report $\\mathrm{mIoU}^1$ on points inside in Tab. 4. Our method not only shows improvements to the LiDAR-only methods but also outperforms other multi-modal methods.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21699", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dff86aca43bddd43c7e6409d3043117a66ce475a905bd63cd00d2290ab1d865e.jpg", + "table_caption": [ + "Table 2. Performance comparison on nuScenes testing set [27]. The modalities available on nuScenes include LiDAR (L), Camera (C), and Radar (R). Top-1 results are in bold. *Submission entries without a published paper by the CVPR 2023 deadline of Nov 11, 2022." + ], + "table_footnote": [], + "table_body": "
MethodModalitymIoUfIoUbarrierbicyclebuscarcons. vehiclemotorcyclepedestriantraffic-conetrailertruckdriveableothersidewalkterrainmammadevegetation
PolarNet [62]L69.4287.3872.1616.8177.0186.5351.1469.6564.8054.1169.7063.5396.6467.1477.7072.1387.1384.47
JS3C-Net [52]L73.6088.0680.1426.1587.7984.5455.1772.5671.2866.2676.7971.1196.8064.4776.8674.0987.4886.10
Cylinder3D [66]L77.1689.9282.7629.7584.3489.4163.0379.2977.2173.4084.5569.1797.6670.2480.2975.5190.4187.55
AMVNet [24]L77.2790.0880.6431.9681.7388.9367.0784.3376.1173.4884.8767.3097.3767.3779.4175.4591.4588.69
SPVNAS [38]L77.3589.6880.0029.9891.9290.8164.6878.9975.6270.9481.0174.6497.4469.2379.9576.1089.2887.06
Cylinder3D++[66]L77.8689.9382.7633.8984.3489.4169.6379.4277.2673.4084.5569.4197.6670.2480.2975.5190.4287.55
AF2S3Net [7]L78.3488.5178.8752.2189.9384.1777.4274,3077.3271.9583.8873.7897.1366.4777.5174.0187.6986.80
SPVCNN++[38]L81.1290.9786.3543.1391.9092.1875.9075.7283.4477.3186.8277.3697.6971.2281.0877.1991.6788.98
LIFusion* [27]LC75.7489.3258.1336.3086.6784.2859.9679.6980.3077.7783.2368.7497.1868.1977.0474.4591.0388.95
PMF [67]LC77.0389.3482.1140.3380.9486.4263.7279.2279.7575.8681.1767.0597.2867.6978.0574.4889.9488.46
CPFusion* [27]LCR77.7289.1983.6737.0389.0286.2470.0877.4778.0774.5382.7867.9496.6468.2479.5374.9190.4786.95
2D3DNet [13]LC79.9690.0883.0159.3587.9985.0963.7084.3981.9575.9684.7971.9396.8867.3579.8175.9692.0589.18
MSeg3DLC81.1491.3583.1142.4694.9292.0167.1078.5885.6680.4787.5377.3297.7469.8281.2277.8392.3590.07
", + "bbox": [ + 176, + 119, + 789, + 297 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c9263821200625948778f16311ddf1a46460f48f6deab9290e89cc154147bba.jpg", + "table_caption": [ + "Table 3. Performance comparison on Waymo testing set [45]. *Submission entries without a published paper by the CVPR 2023 deadline of Nov 11, 2022." + ], + "table_footnote": [], + "table_body": "
MethodmIoU
LiDARMultiNet* [55]71.13
HorizonSegExpert* [45]69.44
SPVCNN++ [38]67.70
PolarFuse* [45]67.28
SalsaNext [40]55.85
MSeg3D70.51
", + "bbox": [ + 191, + 349, + 352, + 439 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da7abc9bac544f86885da766b41c964825017da2c32e35855c734f13c1d6ff9a.jpg", + "table_caption": [ + "Table 4. Performance comparison on SemanticKITTI [1] validation set following PMF [67]. The results of the other methods are from PMF paper." + ], + "table_footnote": [], + "table_body": "
MethodModalitymIoU1
SalsaNext [10]L59.4
SPVNAS [38]L62.3
Cylinder3D [66]L64.9
PointPainting [42]LC54.5
RGBAL [25]LC56.2
PMF [67]LC63.9
L-BaselineL64.8
MSeg3DLC66.7
", + "bbox": [ + 181, + 488, + 362, + 592 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb8ba261cd3e087baa4aa82d7506cbacf5e02ab0164bf609537ab8926f18dee6.jpg", + "table_caption": [ + "Table 5. Analysis on the performance gap on all points and points inside, and the data augmentation (DA). Only GF-Phase is simply used for multi-modal fusion (M-Fusion) here. DA includes LiDAR DA (L-DA) and Multi-modal DA (M-DA)." + ], + "table_footnote": [], + "table_body": "
LiDARM-FusionDAnuScenesWaymo
mIoUmIoU1mIoUmIoU1
✓×L-DA72.0070.7667.4867.41
✓GF-Phase×68.1076.6859.7764.79
✓GF-PhaseL-DA71.3577.4860.3465.70
✓GF-PhaseM-DA72.3978.6563.9767.94
", + "bbox": [ + 129, + 655, + 411, + 720 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/211504e3b86939ecae8625bbc207ebf798029391961821c543764b84557fcab9.jpg", + "table_caption": [ + "Table 6. Analysis on multi-modal feature fusion of GF-Phase, cross-modal (CM) feature completion, and SF-Phase. The supervised CM feature completion is decomposed as $\\mathbf{L}_{\\mathrm{pixel2point}}$ (Eq. 13) and \"Comp.\" (Eq. 14). The gap can be formulated as mIoU - mIoU1." + ], + "table_footnote": [], + "table_body": "
LiDARM-FusionCM Feature CompletionnuScenesWaymo
GF-PhaseSF-PhaseLpixel2pointComp.mIoUmIoU1GapmIoUmIoU1Gap
✓××××72.0070.76+1.2467.4867.41+0.07
✓✓×××72.3978.65-6.2663.9767.94-3.97
✓✓×✓×76.4479.10-2.6667.1369.06-1.93
✓✓×✓✓78.2879.10-0.8267.8969.06-1.17
✓✓✓✓✓80.0080.10-0.1069.6370.19-0.56
", + "bbox": [ + 76, + 799, + 468, + 875 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 306, + 651, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Difficulties Arising from Multi-modality are investigated in Tab. 5 from two perspectives of the performance gap between all points and points inside, as well as the multi-modal data augmentation. The experiments in Tab. 5 start with a vanilla variant of the multi-modal model using the basic GF-Phase fusion without the multi-modal data augmentation. The mIoU is relatively worse than the mIoU1, since the geometry-based feature fusion ideally ignores the inevitable points outside with the missing camera features [20, 25, 67], which also motivates us to close the performance gaps between mIoU and mIoU1 by using the cross-modal feature completion and semantic-based SF-Phase. Besides, in rows 2 and 1, the multi-modal segmentation performance will be worsened if we deprecated the augmentation transformations that cannot be directly applied to both modalities. However, the proposed asymmetric multi-modal data augmentation allows the accumulation of the transformations on point cloud and image, achieving the best performance on both datasets.", + "bbox": [ + 496, + 325, + 890, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-modal Feature Fusion Modules. Tab. 6 shows that our supervised cross-modal feature completion and SF-Phase benefit both mIoU and mIoU1 with gradually narrowed gaps. In row 3, Lpixel2point can also bring improvements, which indicates that transferring appearance information from dense image to sparse point cloud is applicable and beneficial for joint feature learning. In row 4, completing camera features with pseudo-camera features facilitates the feature gap reduction between points inside and outside. Eventually, the SF-Phase further narrows the gaps between mIoU and mIoU1 to the smallest values of 0.10 and 0.56, addressing the limitation of geometric associations by semantic-based feature fusion. Thus, we set the model in the last row as our final framework.", + "bbox": [ + 496, + 612, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mIoU Breakdown Over Distance. Fig. 3 presents the distance-based evaluation corresponding to the models in Tab. 6. The LiDAR-only model degrades at long distances due to more sparse points. Instead, the multimodal model in row 5 effectively alleviates the performance", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21700", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/681d2b87da08b5ef66710f1cc90178665af009dd34df04811f476cfdf7247b34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 88, + 274, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e45d608929ad29f60ff5ed2080c7d5cee422695e4c87498a8db33786491219ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 88, + 454, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eab315ed50cfff5c2985d01f4bbdedf11e962b7c799ca9d0df502695d56cd55e.jpg", + "image_caption": [ + "Figure 3. Distance-based mIoU evaluation on nuScenes and Waymo validation (Val.) sets using the models in Tab. 6." + ], + "image_footnote": [], + "bbox": [ + 86, + 237, + 274, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4bf6a4846b6c0732af2849c732a60edeec6439c2510c4702c2bc2f20b39bdc77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 237, + 454, + 402 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9eab1f212612c2214d63de2e32e49fa4c733deb5bcdde9c557984d222aaa5eba.jpg", + "table_caption": [ + "Table 7. Robustness analysis on MSeg3D by removing some cameras as malfunction. \"#Camera\" denotes the number of available cameras. \"×\" denotes LiDAR-only baseline." + ], + "table_footnote": [], + "table_body": "
#Camera6543210×
nuScenes80.0078.6977.6276.6976.0175.4474.4772.00
Waymo-69.6369.1868.8668.1368.0968.0467.48
", + "bbox": [ + 114, + 478, + 429, + 523 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df11a89135ee8eab8b318d0fb41ba25a937280986fb686c1c15e4ab7528ee93e.jpg", + "table_caption": [ + "Table 8. Robustness analysis on MSeg3D with multi-frame point clouds input. \"#L-Frame\" is the frame number for LiDAR." + ], + "table_footnote": [], + "table_body": "
nuScenes#L-Frame1102025304025
Camera××××××√
mIoU72.0074.6675.3775.7775.2875.1581.12
Waymo#L-Frame151015203010
Camera××××××√
mIoU67.4869.1869.4568.8868.7868.6470.20
", + "bbox": [ + 114, + 558, + 429, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "degradation. In Fig. 3 (a) and (b), improvements between the multi-modal model in row 5 and the LiDAR-only model in row 1 are also increased along the distance due to increased sparsity. The reasons why the improvement of the multi-modal model on Waymo is not as similar as that on nuScenes are further analyzed in Fig. 3 (c) and (d). The Waymo point cloud has denser points, which reduces the dependence on the image. More points outside are distributed along different distances due to no rear cameras on Waymo. Although more points outside hinder the applicability of multi-modal fusion, our final multi-modal model still yields a low-performance gap of $0.56\\mathrm{mIoU}$ on Waymo in row 5.", + "bbox": [ + 75, + 643, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness Against Camera Malfunction. In Tab. 7, our MSeg3D performs properly under the unfavorable condition of camera malfunction, and it can easily deal with the practical situation without switching the model.", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3c365163f4d9d3a7511759012ca2cbaec224cadf72f381c3ebc9d839e29eca46.jpg", + "table_caption": [ + "Table 9. Scalability analysis by barely varying the backbone settings on nuScenes validation set." + ], + "table_footnote": [], + "table_body": "
Image BackbonemIoU#Params(M)Latency(s)
×72.0021.280.083
SegFormer-B0 [47]78.1425.330.204
SegFormer-B5 [47]78.89103.480.479
HRNet-w18 [44]79.2131.560.265
ResNet101 [15]79.3664.600.567
HRNet-w48 [44]80.0087.340.445
", + "bbox": [ + 575, + 117, + 815, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that even with all cameras removed, MSeg3D still outperforms the LiDAR-only baseline, where our cross-modal feature completion supervision provides effective cross-modal information transfer in training.", + "bbox": [ + 496, + 213, + 890, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness Against Multi-frame Point Clouds Input. Since the densified points alleviate the sparsity, collapsing laser points from neighboring frames usually boosts LiDAR-only 3D perception [16, 52]. In Tab. 8, we follow [57] collapsing multiple previous frames to the current frame from the provided ego-vehicle motion information. For LiDAR-only model, the improvements are saturated with 25 and 10 frames on nuScenes and Waymo. Under such conditions, our MSeg3D still achieves further improvements in the last column, which shows that multi-frame point clouds input can be an optional extension.", + "bbox": [ + 496, + 275, + 890, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complexity Scalability. From Tab. 9, a lightweight image backbone also delivers significant performance gains due to multi-modal input. The best-performing HRNet-w48 is our default image backbone. However, the multi-camera image input makes the efficiency bottleneck lie in the image backbone, which deserves in-depth study for autonomous driving [23]. We believe that real-time image segmentation networks such as BiSeNet [59,60] and Fast-SCNN [29] can accelerate our approach. Overall, our MSeg3D can adapt to complexity scalability among various backbones, which flexibly trades off performance and efficiency.", + "bbox": [ + 496, + 443, + 890, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 619, + 617, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel multi-modal 3D semantic segmentation method termed MSeg3D for autonomous driving, based on LiDAR and multi-camera sensors. Our cross-modal feature completion and semantic-based feature fusion solve the overlooked problem, where multi-modal fusion can occur only in the sensor FOV intersection. The proposed asymmetric multi-modal data augmentation enables the multi-modal segmentation model to be effectively trained with reliable performance. Our method achieves state-of-the-art performance on nuScenes, Waymo, and SemanticKITTI. The comprehensive experiments validate improvements and robustness. We hope that our work can inspire further investigation into multi-modal fusion for 3D semantic segmentation in autonomous driving.", + "bbox": [ + 496, + 642, + 890, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: This work was supported by the National Key Research and Development Program of China (2018YFE0183900) and YUNJI Technology Co. Ltd.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jürgen Gall. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In ICCV, pages 9296-9306, 2019. 1, 2, 6, 7", + "[2] Maxim Berman, Amal Rannen Triki, and Matthew B. Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, pages 4413-4421, 2018. 5", + "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, pages 11618-11628, 2020. 1, 2, 4, 6", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 5", + "[5] Bowen Cheng, Ross B. Girshick, Piotr Dólar, Alexander C. Berg, and Alexander Kirillov. Boundary IoU: Improving object-centric image segmentation evaluation. In CVPR, pages 15334-15342, 2021. 4", + "[6] Hui-Xian Cheng, Xian-Feng Han, and Guo-Qiang Xiao. CENet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In ICME, pages 1-6. IEEE, 2022. 5", + "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (AF)2-S3Net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. 2, 7", + "[8] Christopher B. Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2", + "[9] OpenPCDet Contributors. OpenPCDet: An open-source toolbox for 3D object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2023. 3, 6", + "[10] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. SalsaNext: fast, uncertainty-aware semantic segmentation of LiDAR point clouds for autonomous driving. arXiv preprint arXiv:2003.03653, 2020. 6, 7", + "[11] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic NuScenes: A large-scale benchmark for LiDAR panoptic segmentation and tracking. IEEE Robotics Autom. Lett., 7(2):3795-3802, 2022. 1, 6", + "[12] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 1", + "[13] Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Pantofaru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas A. Funkhouser. Learning 3D semantic segmentation with only 2D image supervision. In 3DV, pages 361-372. IEEE, 2021. 1, 2, 3, 7", + "[14] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D semantic segmentation with submanifold sparse" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "convolutional networks. In CVPR, pages 9224-9232, 2018. 1", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 8", + "[16] Peiyun Hu, Jason Ziglar, David Held, and Deva Ramanan. What you see is what you get: Exploiting visibility for 3D object detection. In CVPR, pages 10998-11006, 2020. 8", + "[17] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. RandLA-Net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11105–11114, 2020. 2", + "[18] Zhenchao Jin, Tao Gong, Dongdong Yu, Q. Chu, Jian Wang, Changhu Wang, and Jie Shao. Mining contextual information beyond image for semantic segmentation. In ICCV, pages 7211-7221, 2021. 4", + "[19] Zhenchao Jin, Bin Liu, Qi Chu, and Nenghai Yu. ISNet: Integrate image-level and semantic-level context for semantic segmentation. In ICCV, pages 7169-7178, 2021. 4", + "[20] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. FuseSeg: LiDAR point cloud segmentation fusing multi-modal data. In WACV, pages 1863-1872, 2020. 1, 2, 7", + "[21] Jiale Li, Hang Dai, and Yong Ding. Self-distillation for robust LiDAR semantic segmentation in autonomous driving. In ECCV, volume 13688, pages 659-676, 2022. 2, 5", + "[22] Jiale Li, Hang Dai, Ling Shao, and Yong Ding. From voxel to point: IoU-guided 3D object detection for point cloud with voxel-to-point decoder. In ACM MM, pages 4622-4631, 2021. 3", + "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 8", + "[24] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. AMVNet: Assertion-based multi-view fusion network for LiDAR semantic segmentation. CoRR, abs/2012.04934, 2020. 7", + "[25] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Kumar Yogamani. RGB and LiDAR fusion based 3D semantic segmentation for autonomous driving. In ITSC, pages 7-12, 2019. 1, 2, 7", + "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. RangeNet++: Fast and accurate LiDAR semantic segmentation. In IROS, pages 4213-4220, 2019. 2", + "[27] nuScenes. nuScenes leaderboard of lidar segmentation task. https://www.nuscenes.org/lidar-segmentation?externalData $\\equiv$ all&mapData $=$ all&modalities $\\equiv$ Any, 2023.1,6,7", + "[28] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3D object detection with pointformer. In CVPR, pages 7463-7472, 2021. 5", + "[29] Rudra P. K. Poudel, Stephan Liwicki, and Roberto Cipolla. Fast-SCNN: Fast semantic segmentation network. In BMVC, page 289, 2019. 8" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21702", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Pytorch. Pytorch document of color jitter. https://torch.org/vision/stable/generated/torchvision.transforms.ColorJitter.html?highlight=colorjitter,2023.6", + "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1, 2", + "[32] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 2", + "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells III, and Alejandro F. Frangi, editors, MICCAI, volume 9351, pages 234–241, 2015. 2", + "[34] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. IEEE TPAMI, 39(4):640-651, 2017. 1, 4", + "[35] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3D object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2021. 3, 6", + "[36] Vishwanath A. Sindagi, Yin Zhou, and Oncel Tuzel. MVX-Net: Multimodal VoxelNet for 3D object detection. In ICRA, pages 7276–7282, 2019. 1", + "[37] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2443-2451, 2020. 1, 2, 4, 6", + "[38] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3D architectures with sparse point-voxel convolution. In ECCV, volume 12373, pages 685–702, 2020. 2, 7", + "[39] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPConv: Flexible and deformable convolution for point clouds. In ICCV, pages 6410–6419, 2019. 2", + "[40] Cortinhal Tiago, Tzelepis George, and Erdal Aksoy Eren. SalsaNext: Fast, uncertainty-aware semantic segmentation of LiDAR point clouds. In Advances in Visual Computing, pages 207-222, 2020. 2, 7", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2, 4, 5", + "[42] Sourabh Vora, Alex H. Lang, Bassam Helou, and Oscar Beijbom. PointPainting: Sequential fusion for 3D object detection. In CVPR, pages 4603-4611, 2020. 1, 3, 7", + "[43] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. PointAugmenting: Cross-modal augmentation for 3D object detection. In CVPR, pages 11794–11803, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349-3364, 2021. 1, 3, 8", + "[45] Waymo. Waymo Open Dataset leaderboard of 3D semantic segmentation challenge. https://waymo.com/open/challenges/2022/3d-semantic-segmentation/, 2023.6,7", + "[46] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. SqueezeSegV2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a LiDAR point cloud. In ICRA, pages 4376-4382, 2019. 2", + "[47] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 1, 5, 8", + "[48] Liang Xie, Chao Xiang, Zhengxu Yu, Guodong Xu, Zheng Yang, Deng Cai, and Xiaofei He. PI-RCNN: an efficient multi-sensor 3D object detector with point-based attentive cont-conv fusion module. In AAAI, pages 12460–12467, 2020. 1", + "[49] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In ECCV, pages 1-19, 2020. 2", + "[50] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVNet: A deep and efficient range-point-voxel fusion network for LiDAR point cloud segmentation. In ICCV, pages 16004-16013, 2021. 2, 6", + "[51] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Bin Zhou, and Liangjun Zhang. FusionPainting: Multimodal fusion with adaptive attention for 3D object detection. In ITSC, pages 3047-3054, 2021. 1", + "[52] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 7, 8", + "[53] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. PointASNL: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, pages 5588-5597, 2020. 2", + "[54] Yan Yan, Yuxing Mao, and Bo Li. SECOND: sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2", + "[55] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. LidarMultiNet: Towards a unified multi-task network for LiDAR perception. CoRR, abs/2209.09385, 2022. 7", + "[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6):210:1-210:12, 2016. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21703", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3D object detection and tracking. In CVPR, pages 11784-11793, 2021. 8", + "[58] Tianwei Yin, Xingyi Zhou, and Philipp Krähenbuhl. Multimodal virtual point 3D detection. In NeurIPS, pages 16494-16507, 2021. 4", + "[59] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. BiSeNet V2: bilateral network with guided aggregation for real-time semantic segmentation. IJCV, 129(11):3051-3068, 2021. 8", + "[60] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. BiSeNet: Bilateral segmentation network for real-time semantic segmentation. In ECCV, volume 11217, pages 334-349, 2018. 8", + "[61] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In ECCV, volume 12351, pages 173-190, 2020. 1", + "[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. PolarNet: An improved grid representation for online LiDAR point clouds semantic segmentation. In CVPR, pages 9598-9607, 2020. 2, 7", + "[63] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 6230-6239, 2017. 1", + "[64] Stephan Zheng, Yang Song, Thomas Leung, and Ian J. Goodfellow. Improving the robustness of deep neural networks via stability training. In CVPR, pages 4480-4488, 2016. 6", + "[65] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-PolarNet: Proposal-free LiDAR point cloud panoptic segmentation. In CVPR, pages 13194-13203, 2021. 2", + "[66] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation. In CVPR, pages 9939-9948, 2021. 2, 5, 6, 7", + "[67] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multisensor fusion for 3D LiDAR semantic segmentation. In ICCV, pages 16260-16270, 2021. 1, 2, 6, 7" + ], + "bbox": [ + 78, + 90, + 467, + 671 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21704", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_model.json b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_model.json new file mode 100644 index 0000000000000000000000000000000000000000..50f4dc70182a19ae49391a8dcfdacac415a47d4e --- /dev/null +++ b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.865, + 0.154 + ], + "angle": 0, + "content": "MSeg3D: Multi-modal 3D Semantic Segmentation for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.18, + 0.693, + 0.199 + ], + "angle": 0, + "content": "Jiale Li\\(^{1}\\) Hang Dai\\(^{2*}\\) Hao Han\\(^{3}\\) Yong Ding\\(^{3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.199, + 0.806, + 0.217 + ], + "angle": 0, + "content": "1College of Information Science and Electronic Engineering, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.217, + 0.7, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Computing Science, University of Glasgow" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.234, + 0.709, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Micro-Nano Electronics, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.254, + 0.787, + 0.27 + ], + "angle": 0, + "content": "*Corresponding authors{Hang.Dai@glasgow.ac.uk, dingyong09@zju.edu.cn}." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.697 + ], + "angle": 0, + "content": "LiDAR and camera are two modalities available for 3D semantic segmentation in autonomous driving. The popular LiDAR-only methods severely suffer from inferior segmentation on small and distant objects due to insufficient laser points, while the robust multi-modal solution is under-explored, where we investigate three crucial inherent difficulties: modality heterogeneity, limited sensor field of view intersection, and multi-modal data augmentation. We propose a multi-modal 3D semantic segmentation model (MSeg3D) with joint intra-modal feature extraction and inter-modal feature fusion to mitigate the modality heterogeneity. The multi-modal fusion in MSeg3D consists of geometry-based feature fusion GF-Phase, cross-modal feature completion, and semantic-based feature fusion SF-Phase on all visible points. The multi-modal data augmentation is reinvigorated by applying asymmetric transformations on LiDAR point cloud and multi-camera images individually, which benefits the model training with diversified augmentation transformations. MSeg3D achieves state-of-the-art results on nuScenes, Waymo, and SemanticKITTI datasets. Under the malfunctioning multicamera input and the multi-frame point clouds input, MSeg3D still shows robustness and improves the LiDAR-only baseline. Our code is publicly available at https://github.com/jialeli1/lidarseg3d." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.713, + 0.21, + 0.728 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Scene understanding for safe autonomous driving can be achieved through semantic segmentation using camera 2D images and LiDAR 3D point clouds, which densely classifies each smallest sensing unit of the modality. The image-based 2D semantic segmentation has been developed with massive solid studies [12, 34, 61, 63]. The camera image has rich appearance information about the object but severely suffers from illumination, varying object scales, and indirect applications in the 3D world. Another modality, LiDAR point cloud, drives 3D semantic segmentation with laser points [1, 3, 11, 37]. Unfortunately, irregular" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.364 + ], + "angle": 0, + "content": "laser points are too sparse to capture the details of objects. The inaccurate segmentation appears especially on small and distant objects. The other under-explored direction is using multi-modal data to increase both the robustness and accuracy in 3D semantic segmentation [67]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Despite the conceptual superiority, the development of multi-modal segmentation model is still nontrivial, lagging behind the single-modal methods [27]. We rationally attribute the difficulties to the three following aspects. i) Heterogeneity between modalities. Due to sparse points and dense pixels, point cloud feature extractors [14, 31] and image feature extractors [15, 44, 47] are developed distinctly. Separate intra-modal feature extractors are used to address the heterogeneity [13, 20, 25, 42, 51], but the lack of joint optimization leads to suboptimal features from irrelevant network parameters. ii) Limited intersection on the field of view (FOV) between sensors. Only the points falling into the intersected FOV are geometrically associated with multi-modal data, while simply considering the intersected multi-modal data is not sufficient to be practically applicable. Performing fusion solely in the limited FOV intersection like [20,67] results in unsatisfactory overall segmentation performance as shown in Fig. 1. iii) Multi-modal data augmentation. For example, PMF [67] uses only several 2D augmentations for spatially aligned point cloud projection image and camera RGB image. Under the constraint of modal alignment or 2D representation of point cloud, the multimodal segmentation works [20, 25, 67] discard many useful and critical point cloud augmentation transformations with sacrificed perception performance [36, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Accordingly, we propose a top-performing multi-modal 3D semantic segmentation method termed MSeg3D, inherently motivated by addressing the aforementioned three technical difficulties. i) Unlike separately extracting modal features in existing methods [13, 25, 42, 51], we jointly optimize intra-modal feature extraction and inter-modal feature fusion to drive maximum correlation and complementarity between heterogeneous modalities. ii) To overcome the disregarded multi-modal fusion outside FOV" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21694" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "intersection [25, 67], we propose a cross-modal feature completion and a semantic-based feature fusion phase SF-Phase to collaborate with the geometry-based feature fusion phase GF-Phase. For points outside the FOV intersection, the former completes the missing camera features using predicted pseudo-camera features, under the explicit guidance of cross-modal supervision. For all the points outside and inside the FOV intersection, the later SF-Phase leverages the multi-head attention [41] to model the semantic relation between point and interested categories so that we can attentively fuse the semantic embeddings aggregated from all the visible fields to each point. iii) The challenging multi-modal data augmentation is reinvigorated by being decomposed as the asymmetric transformations in the LiDAR, camera worlds, and local cameras, which enables flexible permutation to enrich training samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.651 + ], + "angle": 0, + "content": "As the proposed components accumulated in Fig. 1, mIoU and \\(\\mathrm{mIoU}^1\\) are significantly increased while the gaps between mIoU and \\(\\mathrm{mIoU}^1\\) are gradually decreased. Our contributions are four-fold: i) We propose a multi-modal segmentation model MSeg3D with joint intra-modal feature extraction and inter-modal feature fusion, achieving state-of-the-art 3D segmentation performance on the competitive nuScenes [3], Waymo [37], and SemanticKITTI [1] datasets for autonomous driving. The proposed framework won \\(2^{nd}\\) place in the Waymo 3D semantic segmentation challenge at CVPR 2022. ii) We propose a cross-modal feature completion and a semantic-based feature fusion phase. To our best knowledge, it is the first time to address the overlooked and inapplicable multi-modal fusion outside the sensor FOV intersection. iii) By applying augmentation transformations asymmetrically on point cloud and images, the proposed asymmetrical multi-modal data augmentation significantly increases the diversity of multi-modal samples for training model with robust improvements. iv) Extensive experimental analyses on the improvement and robustness of our method clearly investigate our designs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.22, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.472, + 0.856 + ], + "angle": 0, + "content": "LiDAR-only 3D Semantic Segmentation is promoted by SemanticKITTI [1], nuScenes [3], and Waymo [37] datasets. The methods follow the U-Net [33] architecture, but progress from three point cloud representations. i) Point. The point methods derived from PointNet++ [31] cost heavy computation on sampling and gathering disordered neighbors, especially on a large-scale LiDAR point cloud. The major point methods [17, 32, 39, 53] perform well on small synthetic point cloud [56] rather than sparse LiDAR point cloud. ii) 2D images. PolarNetseries [62, 65] and others [26, 40, 46, 49] project 3D point" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.903 + ], + "angle": 0, + "content": "1In the following text, mIoU1 denotes the segmentation performance evaluated on only the points inside FOV intersection by excluding the points outside like PMF [67] and other methods [20, 25]." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.089, + 0.872, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Figure 1. Performance (mIoU and mIoU\\(^1\\)) evaluation on nuScenes [3] (a) and Waymo [37] (b) validation sets. All points are distinguished as points inside the sensor FOV intersection (points inside) and points outside the sensor FOV intersection (points outside). We cumulatively add (“+”) the proposed components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.568 + ], + "angle": 0, + "content": "cloud as 2D images in bird's-eye-view and range-view, achieving efficient LiDAR segmentation with 2D CNNs. But the 3D-to-2D projection damages 3D information and performance. iii) 3D Voxel. The state-of-the-art SPVNAS [38], Cylinder3D [66] and SDSeg3D [21] explicitly explore 3D structure information in different 3D coordinate systems, which efficiently perform 3D sparse convolutions on non-empty voxels [8, 54]. The point and image representations also potentially facilitate voxel features in RPVNet [50] and AF2-S3Net [7]. We also perform effective and efficient 3D voxel feature learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Multi-modal 3D Semantic Segmentation is investigated in [13, 20, 25, 67] using LiDAR and camera with unsatisfactory performance. The suboptimal feature sources of RGB value or separately learned CNN feature are warped into point cloud range image as additional input features [20, 25]. Recently, PMF [67] projects point cloud onto the image plane for fusion with image data, which is obviously limited by camera FOV. They discard the fusion and segmentation over points outside. Despite the improved segmentation performance within the FOV, another LiDAR model is required to take responsibility for the remaining area [43, 67]. Such inequity and inconvenience have prevented researchers from focusing on multi-modal 3D semantic segmentation. The image format of point cloud in the aforementioned works limits them to do only image augmentation such as 2D flipping rather than 3D point cloud augmentation. Instead, we point out the necessity of point cloud augmentation for our 3D segmentation network. Unlike another multi-modal 3D object detector PointAugmenting [43] that preserves the point cloud augmentation while keeping the image unchanged, our method starts from both modalities with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21695" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 2. Overview of our multi-modal 3D semantic segmentation model (MSeg3D). For multi-modal feature fusion, GF-Phase mainly includes the Geometry-based Feature Fusion Module (GFFM), while SF-Phase consists of LiDAR Semantic Feature Aggregation Module (SFAM), camera SFAM, and Semantic-based Feature Fusion Module (SFFM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.473, + 0.452 + ], + "angle": 0, + "content": "asymmetric transformations. Besides, 2D3DNet [13] aims to train a 3D model using reduced 3D annotations but more unlabeled data and additional manual 2D annotations to train a 2D model for obtaining 2D segmentation in advance. The 2D segmentation is also painted onto the points as the additional input of 3D model [42]. Such a phased training approach lacks joint 2D-3D optimization. Our motivation, technique, and results are distinguished from them." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.172, + 0.476 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.206, + 0.501 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.645 + ], + "angle": 0, + "content": "Problem Setup. Let \\(\\{\\mathbf{P}_{\\mathrm{in}},\\mathbf{X}_{\\mathrm{in}}\\}\\) be a multi-modal sample, where \\(\\mathbf{P}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{point}}\\times C_{\\mathrm{in}}}\\) denotes LiDAR point cloud of \\(N_{\\mathrm{point}}\\) points with \\(C_\\mathrm{in}\\)-dimensional input features (e.g., 3D coordinates and reflectance) and \\(\\mathbf{X}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{cam}}\\times 3\\times H_{\\mathrm{in}}\\times W_{\\mathrm{in}}}\\) denotes multi-camera RGB images of \\(N_{\\mathrm{cam}}\\) cameras. Based on sensor calibration, the point with 3D coordinates \\((x,y,z)\\) can be projected onto the \\(c\\)-th local camera image with the pixel coordinates \\((c,u,v)\\). \\(N_{\\mathrm{cls}}\\) semantic categories for each point in the 3D point cloud." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.825 + ], + "angle": 0, + "content": "LiDAR Feature Extraction. Given the input \\(\\{\\mathbf{P}_{\\mathrm{in}},\\mathbf{X}_{\\mathrm{in}}\\}\\) in Fig. 2, we first conduct intra-modal feature extraction with parallel backbones for addressing the heterogeneity between point cloud and images. Due to the superiority of 3D voxel, we perform voxel-based LiDAR feature extraction by a general sparse 3D U-Net [9, 35] in Supplementary Material (SM). The input point cloud is first divided by a quantization step \\(d\\) then grouped into non-empty voxels by averaging the point-wise initial features of the local points inside a voxel. These non-empty voxels \\(\\mathbf{V}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{voxel}}\\times C_{\\mathrm{in}}}\\) constitute the sparse tensor input to point cloud backbone for learning the expressive voxel features \\(\\mathbf{V}\\in \\mathbb{R}^{N_{\\mathrm{voxel}}\\times C_{\\mathrm{voxel}}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Camera Feature Extraction. To enhance the LiDAR features with jointly-optimized camera features, we use a trainable image backbone (HRNet-w48 [44] by default) to non-linearly project \\(\\mathbf{X}_{\\mathrm{in}}\\) as \\(\\mathbf{X} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times C_{\\mathrm{img}} \\times H \\times W}\\), with down-sampled shape \\(H \\times W\\) and increased channels \\(C_{\\mathrm{img}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.895, + 0.452 + ], + "angle": 0, + "content": "The point cloud and image backbones can be flexibly selected from various mature networks. Moreover, we jointly optimize both backbones in the overall 3D segmentation model to learn relevant semantic feature representations from different modalities. The expressive voxel features \\(\\mathbf{V}\\) and image features \\(\\mathbf{X}\\) guarantee the inter-modal feature fusion, which consists of GF-Phase, cross-modal feature completion, and SF-Phase in the right half of Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.46, + 0.871, + 0.476 + ], + "angle": 0, + "content": "3.2. GF-Phase: Geometry-based Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.558 + ], + "angle": 0, + "content": "The point-centric segmentation motivates us to warp voxel features \\(\\mathbf{V}\\) and image feature maps \\(\\mathbf{X}\\) to points in \\(\\mathbf{P}_{\\mathrm{in}}\\) as point-wise LiDAR features \\(\\mathbf{F}_{\\mathrm{lidar}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{voxel}}}\\) and point-wise camera features \\(\\mathbf{F}_{\\mathrm{cam}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{img}}}\\) by using the geometric association of LiDAR and multi-camera." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.893, + 0.773 + ], + "angle": 0, + "content": "Learning Point Feature From Voxels and Pixels. We devoxelize voxel features \\(\\mathbf{V}\\) as point-wise LiDAR features \\(\\mathbf{F}_{\\mathrm{lidar}} = [f_{\\mathrm{lidar},i}]_{i=1}^{i=N_{\\mathrm{point}}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{voxel}}}\\) for augmenting the point-wise fusion and segmentation. Given a point, we interpolate its point feature from the three nearest neighboring voxels [22]. For the \\(i\\)-th point \\((x_i, y_i, z_i)\\) with LiDAR feature \\(f_{\\mathrm{lidar},i}\\) in \\(\\mathbf{F}_{\\mathrm{lidar}}\\), we use the known point-wise image coordinates \\((c_i, u_i, v_i)\\) to perform bilinear interpolation \\(\\mathcal{I}\\) on the image feature maps \\(\\mathbf{X}[c_i] \\in \\mathbb{R}^{C_{\\mathrm{img}} \\times H \\times W}\\) within the \\(c_i\\)-th local camera as Eq. 1. Zeros are temporarily padded to those points outside the camera FOV, which allows each point to be decorated with the \\(C_{\\mathrm{img}}\\)-dimensional feature \\(f_{\\mathrm{cam},i}\\). To distinguish points outside, we set a binary mask \\(\\mathbf{B}\\) of \\(N_{\\mathrm{point}}\\) elements with 0 for them and 1 otherwise." + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.78, + 0.891, + 0.811 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {c a m}, i} = \\mathcal {I} \\left(\\mathbf {X} \\left[ c _ {i} \\right], \\frac {H}{H _ {\\mathrm {i n}}} u _ {i}, \\frac {W}{W _ {\\mathrm {i n}}} v _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.813, + 0.891, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {c a m}} = \\left[ f _ {\\mathrm {c a m}, i} \\right] _ {i = 1} ^ {i = N _ {\\text {p o i n t}}} \\in \\mathbb {R} ^ {N _ {\\text {p o i n t}} \\times C _ {\\mathrm {i m g}}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Geometry-based Feature Fusion Module (GFFM) first projects \\( f_{\\mathrm{lidar},i} \\) and \\( f_{\\mathrm{cam},i} \\) as \\( C_{\\mathrm{int}} \\)-dimensional by using fully connected layers \\( \\mathcal{F}_{\\mathrm{lidar}} \\) and \\( \\mathcal{F}_{\\mathrm{cam}} \\), and concatenates (\"©\") them together for subsequent learnable" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21696" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.171 + ], + "angle": 0, + "content": "fusion using another MLP with \\( C_{\\mathrm{gfused}} \\) output channels as \\( f_{\\mathrm{gfused},i} = MLP(\\mathcal{F}_{\\mathrm{lidar}}(f_{\\mathrm{lidar},i}) \\odot \\mathcal{F}_{\\mathrm{cam}}(f_{\\mathrm{cam},i})) \\) and \\( \\mathbf{F}_{\\mathrm{gfused}} = [f_{\\mathrm{gfused},i}]_{i=1}^{i=N_{\\mathrm{point}}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{gfused}}} \\). The GF-Phase combines the multi-modal features as \\( \\mathbf{F}_{\\mathrm{gfused}} \\) mainly with geometry clues, so-called the geometry-based feature fusion phase." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.178, + 0.439, + 0.193 + ], + "angle": 0, + "content": "3.3. SF-Phase: Semantic-based Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.412 + ], + "angle": 0, + "content": "Though the simple yet effective GF-Phase yields appreciable performance gains inside FOV intersection through unbiased considerations of both modalities, three drawbacks remain to be addressed. i) The points outside still cannot retrieve their realistic camera features. ii) The relative importance between modalities varies in different areas [5, 58], which is overlooked in GF-Phase. iii) Explicit relational modeling for category-wise semantic representations [18, 19] is not available. As shown in Fig. 2, we further propose the SF-Phase after GF-Phase. Beyond the geometric association in euclidean space, we aggregate LiDAR features and camera features into category-wise semantic embeddings as \\(\\mathbf{E}_{\\mathrm{lidar}}\\) and \\(\\mathbf{E}_{\\mathrm{cam}}\\) for performing multi-modal fusion in a hidden semantic space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.471, + 0.654 + ], + "angle": 0, + "content": "LiDAR Semantic Feature Aggregation Module (LiDAR SFAM) starts from voxel features \\(\\mathbf{V}\\) and ends with aggregated \\(N_{\\mathrm{cls}}\\) LiDAR semantic embeddings in \\(\\mathbf{E}_{\\mathrm{lidar}}\\). For \\(N_{\\mathrm{cls}}\\) categories, we assume \\(\\mathbf{E}_{\\mathrm{lidar}} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times C_{\\mathrm{voxel}}}\\) and a distribution matrix \\(\\mathbf{D}_{\\mathrm{lidar}} \\in (0,1)^{N_{\\mathrm{cls}} \\times N_{\\mathrm{voxel}}}\\). The \\(\\mathbf{D}[j,i]\\) closer to 1 indicates that the \\(i\\)-th voxel is more likely to belong to the \\(j\\)-th category than other voxels, so that it contributes more to describing this category in \\(\\mathbf{E}_{\\mathrm{lidar}}[j]\\). Otherwise, the \\(i\\)-th voxel should contribute less to \\(\\mathbf{E}_{\\mathrm{lidar}}[j]\\). Now we can compute \\(\\mathbf{E}_{\\mathrm{lidar}}\\) by matrix product as \\(\\mathbf{E}_{\\mathrm{lidar}} = \\mathbf{D}_{\\mathrm{lidar}}\\mathbf{V}\\). Two more steps (Eqs. \\(3 \\sim 4\\)) are required to compute \\(\\mathbf{D}_{\\mathrm{lidar}}\\). We first predict an intermediate segmentation \\(\\mathbf{D}_{\\mathrm{lidar}}' \\in \\mathbb{R}^{N_{\\mathrm{voxel}} \\times N_{\\mathrm{cls}}}\\) on the voxel features \\(\\mathbf{V}\\) using an MLP-based auxiliary voxel segmentation head \\(\\mathcal{H}_{\\mathrm{voxel}}\\). We then perform a spatial softmax among voxels to normalize the values into \\((0,1)\\) for each category." + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.66, + 0.469, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime} = \\mathcal {H} _ {\\text {v o x e l}} (\\mathbf {V}), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.679, + 0.469, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\text {l i d a r}} ^ {T} = \\left[ \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime} [:, j ]\\right) \\right] _ {j = 1} ^ {j = N _ {\\mathrm {c l s}}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.734 + ], + "angle": 0, + "content": "For more accurate \\(\\mathbf{D}_{\\mathrm{lidar}}\\), we explicitly guide \\(\\mathbf{D}_{\\mathrm{lidar}}^{\\prime}\\) by using the voxel segmentation supervision in Eq. 15." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Camera Semantic Feature Aggregation Module (Camera SFAM) similarly aggregates \\(\\mathbf{E}_{\\mathrm{cam}} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times C_{\\mathrm{img}}}\\) from image feature maps \\(\\mathbf{X} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times C_{\\mathrm{img}} \\times H \\times W}\\) using another distribution matrix \\(\\mathbf{D}_{\\mathrm{cam}} \\in (0,1)^{N_{\\mathrm{cls}} \\times N_{\\mathrm{pixel}}}\\) and another auxiliary image segmentation head \\(\\mathcal{H}_{\\mathrm{img}}\\) implemented with the simple FCN head [34]. The difference lies in that the camera SFAM is designed with all the \\(N_{\\mathrm{pixel}}\\) pixels of image feature maps across \\(N_{\\mathrm{cam}}\\) local cameras, where \\(N_{\\mathrm{pixel}}\\) is \\(N_{\\mathrm{cam}} \\times H \\times W\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.904 + ], + "angle": 0, + "content": "The \\(\\mathcal{H}_{\\mathrm{img}}\\) first predicts the intermediate segmentation \\(\\mathbf{D}_{\\mathrm{img}}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times N_{\\mathrm{cls}} \\times H \\times W}\\) by \\(\\mathcal{H}_{\\mathrm{img}}(\\mathbf{X})\\). Note that \\(\\mathbf{D}_{\\mathrm{img}}^{\\prime}\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.168 + ], + "angle": 0, + "content": "and \\(\\mathbf{X}\\) are then rearranged into \\(\\mathbf{D}_{\\mathrm{cam}}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times N_{\\mathrm{pixel}}}\\) and \\(\\mathbf{X}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{pixel}} \\times C_{\\mathrm{img}}}\\) by transposing and reshaping operations. The final camera semantic embeddings \\(\\mathbf{E}_{\\mathrm{cam}}\\) in Eq. 7 is the matrix product of \\(\\mathbf{D}_{\\mathrm{cam}}\\) and \\(\\mathbf{X}^{\\prime}\\), where the normalized \\(\\mathbf{D}_{\\mathrm{cam}}\\) is the result of applying spatial softmax on \\(\\mathbf{D}_{\\mathrm{cam}}^{\\prime}\\) as Eq. 6." + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.188, + 0.891, + 0.206 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\mathrm {i m g}} ^ {\\prime} = \\mathcal {H} _ {\\mathrm {i m g}} (\\mathbf {X}), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.209, + 0.891, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\mathrm {c a m}} = \\left[ \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {\\mathrm {c a m}} ^ {\\prime} [ j,: ]\\right) \\right] _ {j = 1} ^ {j = N _ {\\mathrm {c l s}}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.231, + 0.891, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {\\mathrm {c a m}} = \\mathbf {D} _ {\\mathrm {c a m}} \\mathbf {X} ^ {\\prime}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.892, + 0.328 + ], + "angle": 0, + "content": "Note that the image segmentation annotations are unavailable for guiding the \\(\\mathbf{D}_{\\mathrm{img}}^{\\prime}\\) in such 3D semantic segmentation datasets [3, 37]. Nevertheless, we also address this by the cross-modal semantic supervision scheme, which is detailed in Eq. 16 and Eq. 17." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.893, + 0.495 + ], + "angle": 0, + "content": "Semantic-based Feature Fusion Module (SFFM). As shown in the right dashed box in Fig. 2, given the input of point-wise geometry-based fused features \\(\\mathbf{F}_{\\mathrm{gfused}}\\), category-wise semantic embeddings \\(\\mathbf{E}_{\\mathrm{lidar}}\\) and \\(\\mathbf{E}_{\\mathrm{cam}}\\), the SFFM first projects them into a \\(C_{\\mathrm{sfused}}\\)-dimensional space as \\(\\mathbf{F}_{\\mathrm{proj}} = Proj_1(\\mathbf{F}_{\\mathrm{gfused}})\\), and \\(\\mathbf{E} = Proj_2(\\mathbf{E}_{\\mathrm{lidar}}) \\odot Proj_3(\\mathbf{E}_{\\mathrm{cam}})\\), respectively. Before stacking \\(K\\) gray blocks in SFFM, we concatenate the projected multi-modal semantic embeddings together as \\(\\mathbf{E} \\in \\mathbb{R}^{2N_{\\mathrm{cls}} \\times C_{\\mathrm{sfused}}}\\), which can be regarded as a dictionary that describes the typical characteristics for each category from the perspectives of LiDAR and cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.6 + ], + "angle": 0, + "content": "In each gray block, we model the semantic relations among all the \\(2N_{\\mathrm{cls}}\\) category-wise semantic embeddings using the Multi-Head Self-Attention (MHSA) [41] as Eq. 8, where Norm indicates the LayerNorm operation. Let \\(C_{\\mathrm{shsa}}\\) be \\(C_{\\mathrm{sfused}} / N_H\\), the MHSA can be decomposed into \\(N_{\\mathrm{H}}\\) Single-Head Self-Attention (SHSA) operations with \\(\\mathbf{E}_h \\in \\mathbb{R}^{2N_{\\mathrm{cls}} \\times C_{\\mathrm{shsa}}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.607, + 0.891, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {\\mathrm {m h s a}} = \\operatorname {N o r m} (\\mathbf {E} + M H S A (\\mathbf {E})), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.626, + 0.891, + 0.645 + ], + "angle": 0, + "content": "\\[\nM H S A (\\mathbf {E}) = \\left[ S H S A \\left(\\mathbf {E} _ {h}\\right) \\right] _ {h = 1} ^ {h = N _ {\\mathrm {H}}}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.648, + 0.891, + 0.682 + ], + "angle": 0, + "content": "\\[\nS H S A \\left(\\mathbf {E} _ {h}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {h} \\mathbf {K} _ {h} ^ {T}}{\\sqrt {C _ {\\mathrm {s h s a}}}}\\right) \\mathbf {V} _ {h}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Since \\(\\mathbf{E}\\) includes \\(\\mathbf{E}_{\\mathrm{lidar}}\\) and \\(\\mathbf{E}_{\\mathrm{cam}}\\), the attention matrix \\(Softmax\\left(\\frac{\\mathbf{Q}_h\\mathbf{K}_h^T}{\\sqrt{C_{\\mathrm{shsa}}}}\\right)\\in \\mathbb{R}^{2N_{\\mathrm{cls}}\\times 2N_{\\mathrm{cls}}}\\) intuitively includes four parts of relation modeling on \\(\\mathbf{E}_{\\mathrm{lidar}}\\rightarrow \\mathbf{E}_{\\mathrm{lidar}},\\mathbf{E}_{\\mathrm{lidar}}\\rightarrow \\mathbf{E}_{\\mathrm{cam}},\\) \\(\\mathbf{E}_{\\mathrm{cam}}\\rightarrow \\mathbf{E}_{\\mathrm{lidar}}\\), and \\(\\mathbf{E}_{\\mathrm{cam}}\\rightarrow \\mathbf{E}_{\\mathrm{cam}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.892, + 0.841 + ], + "angle": 0, + "content": "After updating \\(\\mathbf{E}\\) as \\(\\mathbf{E}_{\\mathrm{mhsa}}\\), we perform the further fusion between the point-wise projected features \\(\\mathbf{F}_{\\mathrm{proj}}\\) and the semantic embeddings \\(\\mathbf{E}_{\\mathrm{mhsa}}\\), using the Multi-Head Cross-Attention (MHCA) for semantic relation modeling in Eq 12 and the Feed-Forward Network (FFN) for feature embedding in Eq. 11, respectively." + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.847, + 0.891, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {s f u s e d}} = \\operatorname {N o r m} \\left(\\mathbf {F} _ {\\text {m h c a}} + F F N \\left(\\mathbf {F} _ {\\text {m h c a}}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.866, + 0.891, + 0.896 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {m h c a}} = \\operatorname {N o r m} \\left(\\mathbf {F} _ {\\mathrm {p r o j}} + M H C A \\left(\\mathbf {F} _ {\\mathrm {p r o j}}, \\mathbf {E} _ {\\mathrm {m h s a}}, \\mathbf {E} _ {\\mathrm {m h s a}}\\right)\\right). \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21697" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.212 + ], + "angle": 0, + "content": "Unlike Eq. 8, the \\( MHCA \\) in Eq. 12 sets the query \\( \\mathbf{Q} \\) from point-wise feature \\( \\mathbf{F}_{\\mathrm{proj}} \\), and sets the key \\( \\mathbf{K} \\) and value \\( \\mathbf{V} \\) from semantic embeddings \\( \\mathbf{E}_{\\mathrm{mhsa}} \\). Thus, an attention matrix with shape \\( N_{\\mathrm{point}} \\times 2N_{\\mathrm{cls}} \\) is computed to attentively aggregate the more important semantic embeddings respective to individual points, beyond the paired geometric association in GF-Phase. Following the common usages of multi-head attention [4], \\( K \\) (i.e., 6) gray blocks are stacked." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.485 + ], + "angle": 0, + "content": "Discussion on SFFM. As a useful and general technique [4,28,47], multi-head attention [41], although not proposed in this paper, is effectively tailored to multi-modal feature fusion with different motivations and designs by us. i) In Eq. 12, the \\( MHCA \\) enables the point-wise feature to attend to multi-modal semantic embeddings, so that both points inside and outside can be consistently supported by expressive multi-modal semantic embeddings. ii) The \\( MHCA \\) attention matrix computes the relative importance of two modalities to each point, improving the unbiased considerations of modalities in GF-Phase. iii) The \\( MHSA \\) in Eq. 8 explicitly models the category-wise intra-modal and inter-modal semantic relations, deriving the commonality learning. iv) The LiDAR SFAM and Camera SFAM aggregate the long sequences of voxel features \\( V \\) and image features \\( X' \\) into the short sequences of \\( E_{\\text{lidar}} \\) and \\( E_{\\text{cam}} \\) with \\( N_{\\text{voxel}} / N_{\\text{cls}} \\) and \\( N_{\\text{pixel}} / N_{\\text{cls}} \\) times, which enables efficient computation of the multi-head attention." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.37, + 0.513 + ], + "angle": 0, + "content": "3.4. Cross-modal Feature Completion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.469, + 0.642 + ], + "angle": 0, + "content": "As shown in Fig. 2, a cross-modal feature completion module with pixel-to-point loss \\(\\mathbf{L}_{\\mathrm{pixel2point}}\\) (Eq. 13) is set, where the point-wise pseudo-camera feature \\(\\mathbf{F}_{\\mathrm{pcam}}\\) is mapped from the point-wise LiDAR features \\(\\mathbf{F}_{\\mathrm{lidar}}\\) by another MLP-based \\(\\mathcal{H}_{\\mathrm{pcam}}\\) as \\(\\mathcal{H}_{\\mathrm{pcam}}(\\mathbf{F}_{\\mathrm{lidar}})\\). In practice, we compute the mean square error loss \\(\\mathcal{L}_{\\mathrm{mse}}\\) between the \\(\\mathbf{F}_{\\mathrm{pcam}}\\) and \\(\\mathbf{F}_{\\mathrm{cam}}\\) of the points inside for learning from the correctly paired cross-modal features relationship." + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.656, + 0.469, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {\\text {p i x e l 2 p o i n t}} = \\mathcal {L} _ {\\mathrm {m s e}} \\left(\\mathbf {B F} _ {\\mathrm {p c a m}}, \\mathbf {B D e t a c h} \\left(\\mathbf {F} _ {\\mathrm {c a m}}\\right)\\right). \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.675, + 0.469, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {c a m}} [ i,: ] = \\mathbf {F} _ {\\mathrm {p c a m}} [ i,: ] | _ {i \\in \\{j \\mid \\mathbf {B} [ j ] = 0 \\}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Note that the points outside are ignored by the binary mask \\(\\mathbf{B}\\), which is mentioned in GF-Phase. The gradients of \\(\\mathbf{F}_{\\mathrm{cam}}\\) are also detached for optimizing the learning of \\(\\mathbf{F}_{\\mathrm{pcam}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We employ such a cross-modal feature completion module due to two motivations: i) Optimizing \\(\\mathbf{L}_{\\mathrm{pixel2point}}\\) forces LiDAR features \\(\\mathbf{F}_{\\mathrm{lidar}}\\) to imitate camera features \\(\\mathbf{F}_{\\mathrm{cam}}\\) with \\(\\mathcal{H}_{\\mathrm{pcam}}\\), then the learned pseudo-image features \\(\\mathbf{F}_{\\mathrm{pcam}}\\) can be switched to replace the padded zeros in \\(\\mathbf{F}_{\\mathrm{cam}}\\) as Eq. 14 in the inference stage, which serves as the cross-modal feature completion to enhance the feature learning. Thus, we further reduce feature gaps between points outside and inside. ii) \\(\\mathcal{H}_{\\mathrm{pcam}}\\) transfers rich appearance priors from the camera branch to the LiDAR branch with an effective" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "consistency constraint for enhancing the intra-modal feature learning in the training stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.131, + 0.805, + 0.147 + ], + "angle": 0, + "content": "3.5. Cross-modal Semantic Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.892, + 0.279 + ], + "angle": 0, + "content": "Point Supervision. Let \\(\\mathbf{Y}\\) of \\(N_{\\mathrm{point}}\\) elements be the point-wise 3D semantic segmentation labels. The value of \\(\\mathbf{Y}[i]\\) is in \\([0, N_{\\mathrm{cls}} - 1]\\), where 0 denotes the ignored category. An MLP based point segmentation head \\(\\mathcal{H}_{\\mathrm{point}}\\) is built on \\(\\mathbf{F}_{\\mathrm{sfused}}\\) for the 3D segmentation prediction \\(\\hat{\\mathbf{Y}} = \\mathcal{H}_{\\mathrm{point}}(\\mathbf{F}_{\\mathrm{sfused}})\\). Following [6, 21, 66], we adopt point loss \\(\\mathbf{L}_{\\mathrm{point}}\\) as a combination of cross-entropy loss \\(\\mathcal{L}_{\\mathrm{ce}}\\) and lovasz-softmax loss \\(\\mathcal{L}_{\\mathrm{lovasz}}\\) [2] as \\(\\mathcal{L}_{\\mathrm{ce}}(\\hat{\\mathbf{Y}}, \\mathbf{Y}) + \\mathcal{L}_{\\mathrm{lovasz}}(\\hat{\\mathbf{Y}}, \\mathbf{Y})\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.279, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Point-to-voxel Supervision. For guiding the \\(\\mathbf{D}_{\\mathrm{lidar}}^{\\prime}\\) in Eq. 3, the voxel loss \\(\\mathbf{L}_{\\mathrm{p2v}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.319, + 0.892, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {\\mathrm {p} 2 \\mathrm {v}} = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime}, \\mathbf {Y} _ {\\mathrm {p} 2 \\mathrm {v}}\\right) + \\mathcal {L} _ {\\text {l o v a s z}} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime}, \\mathbf {Y} _ {\\mathrm {p} 2 \\mathrm {v}}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.407 + ], + "angle": 0, + "content": "where the voxel labels \\(\\mathbf{Y}_{\\mathrm{p2v}}\\) of \\(N_{\\mathrm{voxel}}\\) elements can be determined from the labels of points in the voxel. To avoid ambiguity in \\(\\mathbf{L}_{\\mathrm{p2v}}\\), the label of the voxel containing points of multiple categories is set to zero as the ignored category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.893, + 0.619 + ], + "angle": 0, + "content": "Point-to-pixel Supervision. The problem of how to use only point labels to correctly guide \\(\\mathbf{D}_{\\mathrm{img}}^{\\prime}\\) (Eq. 5) is still unsolved. For \\(\\mathbf{D}_{\\mathrm{img}}^{\\prime}\\) of each training sample, we initialize a \\(N_{\\mathrm{cam}} \\times H \\times W\\) map \\(\\mathbf{Y}_{\\mathrm{point2pixel}}\\) of all zeros as the image segmentation label. We then project each point onto the down-sampled image planes and retrieve the point label \\(\\mathbf{Y}[i]\\) for the corresponding nearest \\((\\langle \\cdot \\rangle)\\) pixel as Eq. 16. Although the generated image label \\(\\mathbf{Y}_{\\mathrm{point2pixel}}\\) is sparse, it provides sufficient supervision. In SM, Fig. S2 and Fig. S3 provide visualizations of \\(\\mathbf{D}_{\\mathrm{img}}^{\\prime}\\) to show that these sparse point-to-pixel supervisions can propagate to other regions with similar appearance, which guides \\(\\mathbf{D}_{\\mathrm{img}}^{\\prime}\\) correctly. Since \\(\\mathbf{Y}_{\\mathrm{point2pixel}}\\) is sparse, we only compute the basic cross-entropy loss \\(\\mathcal{L}_{\\mathrm{ce}}\\) in Eq. 17." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.627, + 0.892, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {\\text {p o i n t 2 p i x e l}} \\left[ c _ {i}, \\left\\langle \\frac {H}{H _ {\\mathrm {i n}}} u _ {i} \\right\\rangle , \\left\\langle \\frac {W}{W _ {\\mathrm {i n}}} u _ {i} \\right\\rangle \\right] = \\mathbf {Y} [ i ], \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.661, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {\\text {p o i n t 2 p i x e l}} = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {D} _ {\\mathrm {i m g}} ^ {\\prime}, \\mathbf {Y} _ {\\text {p o i n t 2 p i x e l}}\\right). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Total Loss Function. To jointly optimize all the modules in the proposed segmentation model, the total loss combines loss terms across different modalities into \\(\\mathbf{L}\\) as \\(\\alpha_{1}\\mathbf{L}_{\\mathrm{point}} + \\alpha_{2}\\mathbf{L}_{\\mathrm{p2v}} + \\alpha_{3}\\mathbf{L}_{\\mathrm{point2pixel}} + \\alpha_{4}\\mathbf{L}_{\\mathrm{pixel2point}}\\), where \\(\\alpha_{1} \\sim \\alpha_{4}\\) are set as 1.0, 1.0, 0.5, 1.0 to balance loss terms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.886, + 0.789 + ], + "angle": 0, + "content": "3.6. Asymmetric Multi-modal Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The segmentation output and our multi-modal fusion mechanism are point-centered. In Eq. 1 and Eq. 16, the point and the pixel always can be bridged by the coordinate pair of \\((x_{i},y_{i},z_{i})\\) and \\((c_i,u_i,v_i)\\), as long as we pre-compute the coordinate pairs and keep the order of coordinate pairs among all the points in synchronization. With this ordering constraint, we have the flexibility" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21698" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.146 + ], + "angle": 0, + "content": "Table 1. Data augmentation transformations on LiDAR (L) point cloud and camera (C) images. We only treat random flipping as a naive symmetric transformation that can be simply applied to both modalities simultaneously." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.147, + 0.472, + 0.297 + ], + "angle": 0, + "content": "
L-only·Global rotation around the Z axis with a random angle in [−π/4, +π/4]. \n·Global translation with a random vector (Δx, Δy, Δz) sampled from a Gaussian distribution with mean zero and the standard deviation 0.5. \n·Global scaling with a random scaling factor in [0.95, 1.05]. \n…
Symmetric·Random flipping along the X, Y axis with probability 0.5.
C-only·Scaling with a random ratio in [1.0, 1.5]. \n·Horizontal rotation with a random angle in [−1°, 1°]. \n·Random cropping with a size (Hin, Win). \n·Color jitter from torchvision [30] with random parameters of 0.3 for brightness, 0.3 for contrast, 0.3 for saturation, 0.1 for hue. \n·JPEG with a random compression ratio in [30, 70] and probability 0.5 [64]. \n…
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.444 + ], + "angle": 0, + "content": "to decouple multi-modal data augmentation: i) We can preserve all the LiDAR transformations in Tab. 1 with the images and the coordinates \\((c, u, v)\\) unchanged. ii) Once coordinates \\((c, u, v)\\) are synchronously transformed with images, we can further apply the camera-only transformations in Tab. 1 to images, which are asymmetric to the LiDAR transformations. iii) We can apply independent transformations between the images of local cameras since the image view is naturally sliced by local cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.536 + ], + "angle": 0, + "content": "The asymmetry lies in not only the LiDAR and camera world but also the local cameras. Thus, we significantly diversify the multi-modal data augmentation for 3D semantic segmentation using the introduced camera transformations with the preserved LiDAR transformations. More potential transformations can be exploited beyond Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.55, + 0.203, + 0.567 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.268, + 0.592 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.734 + ], + "angle": 0, + "content": "nuScenes Dataset splits 28,130 training samples, 6,019 validation samples, and 6,008 testing samples [3, 11]. Each nuScenes sample contains relatively sparser point cloud of 32 beams and RGB images captured by 6 cameras: front, front-left, front-right, back, back-left, and back-right. Points outside are projected below the image bottom due to different vertical FOVs of LiDAR and cameras. Following official protocol, the \\( N_{\\mathrm{cls}} \\) is 17. Semantic segmentation annotations are only on point clouds and not on images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Waymo Dataset for 3D semantic segmentation includes 23,691 training samples, 5,976 validation samples, and 2,982 testing samples [37]. Each sample contains the point cloud of 64 beams and the RGB images captured by 5 cameras: front, front-left, front-right, side-left, and side-right. There is no rear camera on Waymo ego-vehicle, hence more points outside unfairly increase the difficulty of multimodal segmentation. The \\( N_{\\mathrm{cls}} \\) is 23. Similarly, semantic segmentation annotations are only on point clouds." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "SemanticKITTI Dataset is collected by a LiDAR with 64 beams [1]. Following [10,50,66,67], we use sequences" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "00 to 10 (excluding 08) with 19,130 training samples and sequence 08 with 4,071 validation samples. It provides only the images of the front-view camera. The \\(N_{\\mathrm{cls}}\\) is 20." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Evaluation Metric is the mean Intersection-over-Union (mIoU) defined as \\(\\frac{1}{C}\\sum_{c = 1}^{N_{\\mathrm{cls}} - 1}\\frac{TP_c}{TP_c + FP_c + FN_c}\\), where \\(TP_{c}, FP_{c}, FN_{c}\\) denote the number of true positive, false positive, and false negative predictions for the \\(c\\)-th category out of \\(N_{\\mathrm{cls}} - 1\\) valid categories. Besides, nuScenes computes another metric fwIoU by weighting category-wise IoU with the point frequency of its category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Settings. Although lightweight LiDRA-only baseline models can be trained with larger batchsize and more epochs to obtain relatively better results, we still train all the models under the same schedule: a batch of 32 random samples distributed on 16 Tesla V100 GPUs with 24 epochs. More implementation details on point cloud voxelization, network architecture, model training, and inference are all included in SM." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.377, + 0.866, + 0.394 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Results on nuScenes. From Tab. 2, multi-modal methods counterintuitively lag far behind LiDAR-only methods. In the top-20 submissions on nuScenes benchmark [27], there are only four multi-modal methods. Details of Cylinder3D++, SPVCNN++, LIFusion, and CPFusion are not available till submission. Our framework achieves the best mIoU and fwIoU with multi-modal fusion mechanism and asymmetric multi-modal data augmentation. Especially, our method achieves superior performance on challenging small objects such as pedestrians and traffic cones, where the laser points are typically insufficient for LiDAR-only methods. Our method also significantly outperforms all public and private multi-modal submissions in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Results on Waymo. We perform evaluation on Waymo benchmark [45] in Tab. 3, where the LiDAR sensor provides denser laser points and the multi-camera excludes the rearview. Such denser point cloud and incomplete camera FOV unfairly weaken the advantages of multiple modalities. Although, our method still achieves state-of-the-art performance using a general point cloud backbone network [9, 35]. We believe that the performance can be improved by incorporating stronger point cloud backbone networks from other state-of-the-art LiDAR-only methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results on SemanticKITTI. While comparisons on the highly competitive nuScenes and Waymo datasets validate the effectiveness of our method, Tab. 4 shows the experimental comparison on SemanticKITTI [1]. Since SemanticKITTI provides only front-view images, PMF [67] evaluates the multi-modal segmentation performance on points inside the FOV intersection. We follow PMF to report \\(\\mathrm{mIoU}^1\\) on points inside in Tab. 4. Our method not only shows improvements to the LiDAR-only methods but also outperforms other multi-modal methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21699" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2. Performance comparison on nuScenes testing set [27]. The modalities available on nuScenes include LiDAR (L), Camera (C), and Radar (R). Top-1 results are in bold. *Submission entries without a published paper by the CVPR 2023 deadline of Nov 11, 2022." + }, + { + "type": "table", + "bbox": [ + 0.178, + 0.12, + 0.79, + 0.299 + ], + "angle": 0, + "content": "
MethodModalitymIoUfIoUbarrierbicyclebuscarcons. vehiclemotorcyclepedestriantraffic-conetrailertruckdriveableothersidewalkterrainmammadevegetation
PolarNet [62]L69.4287.3872.1616.8177.0186.5351.1469.6564.8054.1169.7063.5396.6467.1477.7072.1387.1384.47
JS3C-Net [52]L73.6088.0680.1426.1587.7984.5455.1772.5671.2866.2676.7971.1196.8064.4776.8674.0987.4886.10
Cylinder3D [66]L77.1689.9282.7629.7584.3489.4163.0379.2977.2173.4084.5569.1797.6670.2480.2975.5190.4187.55
AMVNet [24]L77.2790.0880.6431.9681.7388.9367.0784.3376.1173.4884.8767.3097.3767.3779.4175.4591.4588.69
SPVNAS [38]L77.3589.6880.0029.9891.9290.8164.6878.9975.6270.9481.0174.6497.4469.2379.9576.1089.2887.06
Cylinder3D++[66]L77.8689.9382.7633.8984.3489.4169.6379.4277.2673.4084.5569.4197.6670.2480.2975.5190.4287.55
AF2S3Net [7]L78.3488.5178.8752.2189.9384.1777.4274,3077.3271.9583.8873.7897.1366.4777.5174.0187.6986.80
SPVCNN++[38]L81.1290.9786.3543.1391.9092.1875.9075.7283.4477.3186.8277.3697.6971.2281.0877.1991.6788.98
LIFusion* [27]LC75.7489.3258.1336.3086.6784.2859.9679.6980.3077.7783.2368.7497.1868.1977.0474.4591.0388.95
PMF [67]LC77.0389.3482.1140.3380.9486.4263.7279.2279.7575.8681.1767.0597.2867.6978.0574.4889.9488.46
CPFusion* [27]LCR77.7289.1983.6737.0389.0286.2470.0877.4778.0774.5382.7867.9496.6468.2479.5374.9190.4786.95
2D3DNet [13]LC79.9690.0883.0159.3587.9985.0963.7084.3981.9575.9684.7971.9396.8867.3579.8175.9692.0589.18
MSeg3DLC81.1491.3583.1142.4694.9292.0167.1078.5885.6680.4787.5377.3297.7469.8281.2277.8392.3590.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.307, + 0.468, + 0.346 + ], + "angle": 0, + "content": "Table 3. Performance comparison on Waymo testing set [45]. *Submission entries without a published paper by the CVPR 2023 deadline of Nov 11, 2022." + }, + { + "type": "table", + "bbox": [ + 0.192, + 0.35, + 0.353, + 0.44 + ], + "angle": 0, + "content": "
MethodmIoU
LiDARMultiNet* [55]71.13
HorizonSegExpert* [45]69.44
SPVCNN++ [38]67.70
PolarFuse* [45]67.28
SalsaNext [40]55.85
MSeg3D70.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.488 + ], + "angle": 0, + "content": "Table 4. Performance comparison on SemanticKITTI [1] validation set following PMF [67]. The results of the other methods are from PMF paper." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.489, + 0.363, + 0.593 + ], + "angle": 0, + "content": "
MethodModalitymIoU1
SalsaNext [10]L59.4
SPVNAS [38]L62.3
Cylinder3D [66]L64.9
PointPainting [42]LC54.5
RGBAL [25]LC56.2
PMF [67]LC63.9
L-BaselineL64.8
MSeg3DLC66.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.654 + ], + "angle": 0, + "content": "Table 5. Analysis on the performance gap on all points and points inside, and the data augmentation (DA). Only GF-Phase is simply used for multi-modal fusion (M-Fusion) here. DA includes LiDAR DA (L-DA) and Multi-modal DA (M-DA)." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.656, + 0.413, + 0.722 + ], + "angle": 0, + "content": "
LiDARM-FusionDAnuScenesWaymo
mIoUmIoU1mIoUmIoU1
✓×L-DA72.0070.7667.4867.41
✓GF-Phase×68.1076.6859.7764.79
✓GF-PhaseL-DA71.3577.4860.3465.70
✓GF-PhaseM-DA72.3978.6563.9767.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.728, + 0.468, + 0.796 + ], + "angle": 0, + "content": "Table 6. Analysis on multi-modal feature fusion of GF-Phase, cross-modal (CM) feature completion, and SF-Phase. The supervised CM feature completion is decomposed as \\(\\mathbf{L}_{\\mathrm{pixel2point}}\\) (Eq. 13) and \"Comp.\" (Eq. 14). The gap can be formulated as mIoU - mIoU1." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.8, + 0.47, + 0.876 + ], + "angle": 0, + "content": "
LiDARM-FusionCM Feature CompletionnuScenesWaymo
GF-PhaseSF-PhaseLpixel2pointComp.mIoUmIoU1GapmIoUmIoU1Gap
✓××××72.0070.76+1.2467.4867.41+0.07
✓✓×××72.3978.65-6.2663.9767.94-3.97
✓✓×✓×76.4479.10-2.6667.1369.06-1.93
✓✓×✓✓78.2879.10-0.8267.8969.06-1.17
✓✓✓✓✓80.0080.10-0.1069.6370.19-0.56
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.308, + 0.653, + 0.324 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Difficulties Arising from Multi-modality are investigated in Tab. 5 from two perspectives of the performance gap between all points and points inside, as well as the multi-modal data augmentation. The experiments in Tab. 5 start with a vanilla variant of the multi-modal model using the basic GF-Phase fusion without the multi-modal data augmentation. The mIoU is relatively worse than the mIoU1, since the geometry-based feature fusion ideally ignores the inevitable points outside with the missing camera features [20, 25, 67], which also motivates us to close the performance gaps between mIoU and mIoU1 by using the cross-modal feature completion and semantic-based SF-Phase. Besides, in rows 2 and 1, the multi-modal segmentation performance will be worsened if we deprecated the augmentation transformations that cannot be directly applied to both modalities. However, the proposed asymmetric multi-modal data augmentation allows the accumulation of the transformations on point cloud and image, achieving the best performance on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Multi-modal Feature Fusion Modules. Tab. 6 shows that our supervised cross-modal feature completion and SF-Phase benefit both mIoU and mIoU1 with gradually narrowed gaps. In row 3, Lpixel2point can also bring improvements, which indicates that transferring appearance information from dense image to sparse point cloud is applicable and beneficial for joint feature learning. In row 4, completing camera features with pseudo-camera features facilitates the feature gap reduction between points inside and outside. Eventually, the SF-Phase further narrows the gaps between mIoU and mIoU1 to the smallest values of 0.10 and 0.56, addressing the limitation of geometric associations by semantic-based feature fusion. Thus, we set the model in the last row as our final framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "mIoU Breakdown Over Distance. Fig. 3 presents the distance-based evaluation corresponding to the models in Tab. 6. The LiDAR-only model degrades at long distances due to more sparse points. Instead, the multimodal model in row 5 effectively alleviates the performance" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21700" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.275, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.089, + 0.455, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.238, + 0.276, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.238, + 0.455, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.405, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Figure 3. Distance-based mIoU evaluation on nuScenes and Waymo validation (Val.) sets using the models in Tab. 6." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.439, + 0.47, + 0.478 + ], + "angle": 0, + "content": "Table 7. Robustness analysis on MSeg3D by removing some cameras as malfunction. \"#Camera\" denotes the number of available cameras. \"×\" denotes LiDAR-only baseline." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.479, + 0.43, + 0.524 + ], + "angle": 0, + "content": "
#Camera6543210×
nuScenes80.0078.6977.6276.6976.0175.4474.4772.00
Waymo-69.6369.1868.8668.1368.0968.0467.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.532, + 0.469, + 0.559 + ], + "angle": 0, + "content": "Table 8. Robustness analysis on MSeg3D with multi-frame point clouds input. \"#L-Frame\" is the frame number for LiDAR." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.559, + 0.43, + 0.634 + ], + "angle": 0, + "content": "
nuScenes#L-Frame1102025304025
Camera××××××√
mIoU72.0074.6675.3775.7775.2875.1581.12
Waymo#L-Frame151015203010
Camera××××××√
mIoU67.4869.1869.4568.8868.7868.6470.20
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.839 + ], + "angle": 0, + "content": "degradation. In Fig. 3 (a) and (b), improvements between the multi-modal model in row 5 and the LiDAR-only model in row 1 are also increased along the distance due to increased sparsity. The reasons why the improvement of the multi-modal model on Waymo is not as similar as that on nuScenes are further analyzed in Fig. 3 (c) and (d). The Waymo point cloud has denser points, which reduces the dependence on the image. More points outside are distributed along different distances due to no rear cameras on Waymo. Although more points outside hinder the applicability of multi-modal fusion, our final multi-modal model still yields a low-performance gap of \\(0.56\\mathrm{mIoU}\\) on Waymo in row 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Robustness Against Camera Malfunction. In Tab. 7, our MSeg3D performs properly under the unfavorable condition of camera malfunction, and it can easily deal with the practical situation without switching the model." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 9. Scalability analysis by barely varying the backbone settings on nuScenes validation set." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.118, + 0.816, + 0.203 + ], + "angle": 0, + "content": "
Image BackbonemIoU#Params(M)Latency(s)
×72.0021.280.083
SegFormer-B0 [47]78.1425.330.204
SegFormer-B5 [47]78.89103.480.479
HRNet-w18 [44]79.2131.560.265
ResNet101 [15]79.3664.600.567
HRNet-w48 [44]80.0087.340.445
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.274 + ], + "angle": 0, + "content": "Note that even with all cameras removed, MSeg3D still outperforms the LiDAR-only baseline, where our cross-modal feature completion supervision provides effective cross-modal information transfer in training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Robustness Against Multi-frame Point Clouds Input. Since the densified points alleviate the sparsity, collapsing laser points from neighboring frames usually boosts LiDAR-only 3D perception [16, 52]. In Tab. 8, we follow [57] collapsing multiple previous frames to the current frame from the provided ego-vehicle motion information. For LiDAR-only model, the improvements are saturated with 25 and 10 frames on nuScenes and Waymo. Under such conditions, our MSeg3D still achieves further improvements in the last column, which shows that multi-frame point clouds input can be an optional extension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Complexity Scalability. From Tab. 9, a lightweight image backbone also delivers significant performance gains due to multi-modal input. The best-performing HRNet-w48 is our default image backbone. However, the multi-camera image input makes the efficiency bottleneck lie in the image backbone, which deserves in-depth study for autonomous driving [23]. We believe that real-time image segmentation networks such as BiSeNet [59,60] and Fast-SCNN [29] can accelerate our approach. Overall, our MSeg3D can adapt to complexity scalability among various backbones, which flexibly trades off performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.619, + 0.637 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.853 + ], + "angle": 0, + "content": "We propose a novel multi-modal 3D semantic segmentation method termed MSeg3D for autonomous driving, based on LiDAR and multi-camera sensors. Our cross-modal feature completion and semantic-based feature fusion solve the overlooked problem, where multi-modal fusion can occur only in the sensor FOV intersection. The proposed asymmetric multi-modal data augmentation enables the multi-modal segmentation model to be effectively trained with reliable performance. Our method achieves state-of-the-art performance on nuScenes, Waymo, and SemanticKITTI. The comprehensive experiments validate improvements and robustness. We hope that our work can inspire further investigation into multi-modal fusion for 3D semantic segmentation in autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment: This work was supported by the National Key Research and Development Program of China (2018YFE0183900) and YUNJI Technology Co. Ltd." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21701" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jürgen Gall. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In ICCV, pages 9296-9306, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.47, + 0.24 + ], + "angle": 0, + "content": "[2] Maxim Berman, Amal Rannen Triki, and Matthew B. Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, pages 4413-4421, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, pages 11618-11628, 2020. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ross B. Girshick, Piotr Dólar, Alexander C. Berg, and Alexander Kirillov. Boundary IoU: Improving object-centric image segmentation evaluation. In CVPR, pages 15334-15342, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.469, + 0.478 + ], + "angle": 0, + "content": "[6] Hui-Xian Cheng, Xian-Feng Han, and Guo-Qiang Xiao. CENet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In ICME, pages 1-6. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (AF)2-S3Net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[8] Christopher B. Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[9] OpenPCDet Contributors. OpenPCDet: An open-source toolbox for 3D object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[10] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. SalsaNext: fast, uncertainty-aware semantic segmentation of LiDAR point clouds for autonomous driving. arXiv preprint arXiv:2003.03653, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[11] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic NuScenes: A large-scale benchmark for LiDAR panoptic segmentation and tracking. IEEE Robotics Autom. Lett., 7(2):3795-3802, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[12] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Pantofaru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas A. Funkhouser. Learning 3D semantic segmentation with only 2D image supervision. In 3DV, pages 361-372. IEEE, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D semantic segmentation with submanifold sparse" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "convolutional networks. In CVPR, pages 9224-9232, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.167, + 0.892, + 0.208 + ], + "angle": 0, + "content": "[16] Peiyun Hu, Jason Ziglar, David Held, and Deva Ramanan. What you see is what you get: Exploiting visibility for 3D object detection. In CVPR, pages 10998-11006, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.211, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[17] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. RandLA-Net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11105–11114, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.268, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[18] Zhenchao Jin, Tao Gong, Dongdong Yu, Q. Chu, Jian Wang, Changhu Wang, and Jie Shao. Mining contextual information beyond image for semantic segmentation. In ICCV, pages 7211-7221, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.892, + 0.367 + ], + "angle": 0, + "content": "[19] Zhenchao Jin, Bin Liu, Qi Chu, and Nenghai Yu. ISNet: Integrate image-level and semantic-level context for semantic segmentation. In ICCV, pages 7169-7178, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.892, + 0.423 + ], + "angle": 0, + "content": "[20] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. FuseSeg: LiDAR point cloud segmentation fusing multi-modal data. In WACV, pages 1863-1872, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.426, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[21] Jiale Li, Hang Dai, and Yong Ding. Self-distillation for robust LiDAR semantic segmentation in autonomous driving. In ECCV, volume 13688, pages 659-676, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[22] Jiale Li, Hang Dai, Ling Shao, and Yong Ding. From voxel to point: IoU-guided 3D object detection for point cloud with voxel-to-point decoder. In ACM MM, pages 4622-4631, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.542, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.6, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[24] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. AMVNet: Assertion-based multi-view fusion network for LiDAR semantic segmentation. CoRR, abs/2012.04934, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[25] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Kumar Yogamani. RGB and LiDAR fusion based 3D semantic segmentation for autonomous driving. In ITSC, pages 7-12, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.715, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. RangeNet++: Fast and accurate LiDAR semantic segmentation. In IROS, pages 4213-4220, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[27] nuScenes. nuScenes leaderboard of lidar segmentation task. https://www.nuscenes.org/lidar-segmentation?externalData \\(\\equiv\\) all&mapData \\(=\\) all&modalities \\(\\equiv\\) Any, 2023.1,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[28] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3D object detection with pointformer. In CVPR, pages 7463-7472, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Rudra P. K. Poudel, Stephan Liwicki, and Roberto Cipolla. Fast-SCNN: Fast semantic segmentation network. In BMVC, page 289, 2019. 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21702" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[30] Pytorch. Pytorch document of color jitter. https://torch.org/vision/stable/generated/torchvision.transforms.ColorJitter.html?highlight=colorjitter,2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[32] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells III, and Alejandro F. Frangi, editors, MICCAI, volume 9351, pages 234–241, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[34] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. IEEE TPAMI, 39(4):640-651, 2017. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[35] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3D object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[36] Vishwanath A. Sindagi, Yin Zhou, and Oncel Tuzel. MVX-Net: Multimodal VoxelNet for 3D object detection. In ICRA, pages 7276–7282, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[37] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2443-2451, 2020. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[38] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3D architectures with sparse point-voxel convolution. In ECCV, volume 12373, pages 685–702, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[39] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPConv: Flexible and deformable convolution for point clouds. In ICCV, pages 6410–6419, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[40] Cortinhal Tiago, Tzelepis George, and Erdal Aksoy Eren. SalsaNext: Fast, uncertainty-aware semantic segmentation of LiDAR point clouds. In Advances in Visual Computing, pages 207-222, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[42] Sourabh Vora, Alex H. Lang, Bassam Helou, and Oscar Beijbom. PointPainting: Sequential fusion for 3D object detection. In CVPR, pages 4603-4611, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. PointAugmenting: Cross-modal augmentation for 3D object detection. In CVPR, pages 11794–11803, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[44] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349-3364, 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[45] Waymo. Waymo Open Dataset leaderboard of 3D semantic segmentation challenge. https://waymo.com/open/challenges/2022/3d-semantic-segmentation/, 2023.6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[46] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. SqueezeSegV2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a LiDAR point cloud. In ICRA, pages 4376-4382, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.295, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[47] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.353, + 0.892, + 0.421 + ], + "angle": 0, + "content": "[48] Liang Xie, Chao Xiang, Zhengxu Yu, Guodong Xu, Zheng Yang, Deng Cai, and Xiaofei He. PI-RCNN: an efficient multi-sensor 3D object detector with point-based attentive cont-conv fusion module. In AAAI, pages 12460–12467, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.424, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[49] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In ECCV, pages 1-19, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[50] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVNet: A deep and efficient range-point-voxel fusion network for LiDAR point cloud segmentation. In ICCV, pages 16004-16013, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.892, + 0.596 + ], + "angle": 0, + "content": "[51] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Bin Zhou, and Liangjun Zhang. FusionPainting: Multimodal fusion with adaptive attention for 3D object detection. In ITSC, pages 3047-3054, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.599, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[52] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.671, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[53] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. PointASNL: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, pages 5588-5597, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[54] Yan Yan, Yuxing Mao, and Bo Li. SECOND: sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[55] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. LidarMultiNet: Towards a unified multi-task network for LiDAR perception. CoRR, abs/2209.09385, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6):210:1-210:12, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21703" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[57] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3D object detection and tracking. In CVPR, pages 11784-11793, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[58] Tianwei Yin, Xingyi Zhou, and Philipp Krähenbuhl. Multimodal virtual point 3D detection. In NeurIPS, pages 16494-16507, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[59] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. BiSeNet V2: bilateral network with guided aggregation for real-time semantic segmentation. IJCV, 129(11):3051-3068, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[60] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. BiSeNet: Bilateral segmentation network for real-time semantic segmentation. In ECCV, volume 11217, pages 334-349, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[61] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In ECCV, volume 12351, pages 173-190, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. PolarNet: An improved grid representation for online LiDAR point clouds semantic segmentation. In CVPR, pages 9598-9607, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[63] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 6230-6239, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.448, + 0.468, + 0.5 + ], + "angle": 0, + "content": "[64] Stephan Zheng, Yang Song, Thomas Leung, and Ian J. Goodfellow. Improving the robustness of deep neural networks via stability training. In CVPR, pages 4480-4488, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.468, + 0.544 + ], + "angle": 0, + "content": "[65] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-PolarNet: Proposal-free LiDAR point cloud panoptic segmentation. In CVPR, pages 13194-13203, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.468, + 0.613 + ], + "angle": 0, + "content": "[66] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation. In CVPR, pages 9939-9948, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.468, + 0.672 + ], + "angle": 0, + "content": "[67] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multisensor fusion for 3D LiDAR semantic segmentation. In ICCV, pages 16260-16270, 2021. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21704" + } + ] +] \ No newline at end of file diff --git a/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_origin.pdf b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c0e275bb8cbaff418296853959acdcd600afb4f9 --- /dev/null +++ b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/260fb1ca-a74f-4680-81e2-e4e9fb748eda_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a51298f1d239d434259f51da1d52ab04f663658783cf502ebbb2fbd54018e0c +size 1199812 diff --git a/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/full.md b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ef5942fda1adbc7ab9edf0a3dffb975d0491a23 --- /dev/null +++ b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/full.md @@ -0,0 +1,369 @@ +# MSeg3D: Multi-modal 3D Semantic Segmentation for Autonomous Driving + +Jiale Li $^{1}$ Hang Dai $^{2*}$ Hao Han $^{3}$ Yong Ding $^{3*}$ + +1College of Information Science and Electronic Engineering, Zhejiang University + +$^{2}$ School of Computing Science, University of Glasgow + +$^{3}$ School of Micro-Nano Electronics, Zhejiang University + +*Corresponding authors{Hang.Dai@glasgow.ac.uk, dingyong09@zju.edu.cn}. + +# Abstract + +LiDAR and camera are two modalities available for 3D semantic segmentation in autonomous driving. The popular LiDAR-only methods severely suffer from inferior segmentation on small and distant objects due to insufficient laser points, while the robust multi-modal solution is under-explored, where we investigate three crucial inherent difficulties: modality heterogeneity, limited sensor field of view intersection, and multi-modal data augmentation. We propose a multi-modal 3D semantic segmentation model (MSeg3D) with joint intra-modal feature extraction and inter-modal feature fusion to mitigate the modality heterogeneity. The multi-modal fusion in MSeg3D consists of geometry-based feature fusion GF-Phase, cross-modal feature completion, and semantic-based feature fusion SF-Phase on all visible points. The multi-modal data augmentation is reinvigorated by applying asymmetric transformations on LiDAR point cloud and multi-camera images individually, which benefits the model training with diversified augmentation transformations. MSeg3D achieves state-of-the-art results on nuScenes, Waymo, and SemanticKITTI datasets. Under the malfunctioning multicamera input and the multi-frame point clouds input, MSeg3D still shows robustness and improves the LiDAR-only baseline. Our code is publicly available at https://github.com/jialeli1/lidarseg3d. + +# 1. Introduction + +Scene understanding for safe autonomous driving can be achieved through semantic segmentation using camera 2D images and LiDAR 3D point clouds, which densely classifies each smallest sensing unit of the modality. The image-based 2D semantic segmentation has been developed with massive solid studies [12, 34, 61, 63]. The camera image has rich appearance information about the object but severely suffers from illumination, varying object scales, and indirect applications in the 3D world. Another modality, LiDAR point cloud, drives 3D semantic segmentation with laser points [1, 3, 11, 37]. Unfortunately, irregular + +laser points are too sparse to capture the details of objects. The inaccurate segmentation appears especially on small and distant objects. The other under-explored direction is using multi-modal data to increase both the robustness and accuracy in 3D semantic segmentation [67]. + +Despite the conceptual superiority, the development of multi-modal segmentation model is still nontrivial, lagging behind the single-modal methods [27]. We rationally attribute the difficulties to the three following aspects. i) Heterogeneity between modalities. Due to sparse points and dense pixels, point cloud feature extractors [14, 31] and image feature extractors [15, 44, 47] are developed distinctly. Separate intra-modal feature extractors are used to address the heterogeneity [13, 20, 25, 42, 51], but the lack of joint optimization leads to suboptimal features from irrelevant network parameters. ii) Limited intersection on the field of view (FOV) between sensors. Only the points falling into the intersected FOV are geometrically associated with multi-modal data, while simply considering the intersected multi-modal data is not sufficient to be practically applicable. Performing fusion solely in the limited FOV intersection like [20,67] results in unsatisfactory overall segmentation performance as shown in Fig. 1. iii) Multi-modal data augmentation. For example, PMF [67] uses only several 2D augmentations for spatially aligned point cloud projection image and camera RGB image. Under the constraint of modal alignment or 2D representation of point cloud, the multimodal segmentation works [20, 25, 67] discard many useful and critical point cloud augmentation transformations with sacrificed perception performance [36, 48]. + +Accordingly, we propose a top-performing multi-modal 3D semantic segmentation method termed MSeg3D, inherently motivated by addressing the aforementioned three technical difficulties. i) Unlike separately extracting modal features in existing methods [13, 25, 42, 51], we jointly optimize intra-modal feature extraction and inter-modal feature fusion to drive maximum correlation and complementarity between heterogeneous modalities. ii) To overcome the disregarded multi-modal fusion outside FOV + +intersection [25, 67], we propose a cross-modal feature completion and a semantic-based feature fusion phase SF-Phase to collaborate with the geometry-based feature fusion phase GF-Phase. For points outside the FOV intersection, the former completes the missing camera features using predicted pseudo-camera features, under the explicit guidance of cross-modal supervision. For all the points outside and inside the FOV intersection, the later SF-Phase leverages the multi-head attention [41] to model the semantic relation between point and interested categories so that we can attentively fuse the semantic embeddings aggregated from all the visible fields to each point. iii) The challenging multi-modal data augmentation is reinvigorated by being decomposed as the asymmetric transformations in the LiDAR, camera worlds, and local cameras, which enables flexible permutation to enrich training samples. + +As the proposed components accumulated in Fig. 1, mIoU and $\mathrm{mIoU}^1$ are significantly increased while the gaps between mIoU and $\mathrm{mIoU}^1$ are gradually decreased. Our contributions are four-fold: i) We propose a multi-modal segmentation model MSeg3D with joint intra-modal feature extraction and inter-modal feature fusion, achieving state-of-the-art 3D segmentation performance on the competitive nuScenes [3], Waymo [37], and SemanticKITTI [1] datasets for autonomous driving. The proposed framework won $2^{nd}$ place in the Waymo 3D semantic segmentation challenge at CVPR 2022. ii) We propose a cross-modal feature completion and a semantic-based feature fusion phase. To our best knowledge, it is the first time to address the overlooked and inapplicable multi-modal fusion outside the sensor FOV intersection. iii) By applying augmentation transformations asymmetrically on point cloud and images, the proposed asymmetrical multi-modal data augmentation significantly increases the diversity of multi-modal samples for training model with robust improvements. iv) Extensive experimental analyses on the improvement and robustness of our method clearly investigate our designs. + +# 2. Related Work + +LiDAR-only 3D Semantic Segmentation is promoted by SemanticKITTI [1], nuScenes [3], and Waymo [37] datasets. The methods follow the U-Net [33] architecture, but progress from three point cloud representations. i) Point. The point methods derived from PointNet++ [31] cost heavy computation on sampling and gathering disordered neighbors, especially on a large-scale LiDAR point cloud. The major point methods [17, 32, 39, 53] perform well on small synthetic point cloud [56] rather than sparse LiDAR point cloud. ii) 2D images. PolarNetseries [62, 65] and others [26, 40, 46, 49] project 3D point + +![](images/aa4344f810a7997690cdd94c737b4fce6b7a142acc6ee40926ee7928c15aad3e.jpg) +Figure 1. Performance (mIoU and mIoU $^1$ ) evaluation on nuScenes [3] (a) and Waymo [37] (b) validation sets. All points are distinguished as points inside the sensor FOV intersection (points inside) and points outside the sensor FOV intersection (points outside). We cumulatively add (“+”) the proposed components. + +cloud as 2D images in bird's-eye-view and range-view, achieving efficient LiDAR segmentation with 2D CNNs. But the 3D-to-2D projection damages 3D information and performance. iii) 3D Voxel. The state-of-the-art SPVNAS [38], Cylinder3D [66] and SDSeg3D [21] explicitly explore 3D structure information in different 3D coordinate systems, which efficiently perform 3D sparse convolutions on non-empty voxels [8, 54]. The point and image representations also potentially facilitate voxel features in RPVNet [50] and AF2-S3Net [7]. We also perform effective and efficient 3D voxel feature learning. + +Multi-modal 3D Semantic Segmentation is investigated in [13, 20, 25, 67] using LiDAR and camera with unsatisfactory performance. The suboptimal feature sources of RGB value or separately learned CNN feature are warped into point cloud range image as additional input features [20, 25]. Recently, PMF [67] projects point cloud onto the image plane for fusion with image data, which is obviously limited by camera FOV. They discard the fusion and segmentation over points outside. Despite the improved segmentation performance within the FOV, another LiDAR model is required to take responsibility for the remaining area [43, 67]. Such inequity and inconvenience have prevented researchers from focusing on multi-modal 3D semantic segmentation. The image format of point cloud in the aforementioned works limits them to do only image augmentation such as 2D flipping rather than 3D point cloud augmentation. Instead, we point out the necessity of point cloud augmentation for our 3D segmentation network. Unlike another multi-modal 3D object detector PointAugmenting [43] that preserves the point cloud augmentation while keeping the image unchanged, our method starts from both modalities with + +![](images/47cad5a6f8e5c2f3cc200593163ae883e4f473d4ed37a6d0a0725bc124295b7d.jpg) +Figure 2. Overview of our multi-modal 3D semantic segmentation model (MSeg3D). For multi-modal feature fusion, GF-Phase mainly includes the Geometry-based Feature Fusion Module (GFFM), while SF-Phase consists of LiDAR Semantic Feature Aggregation Module (SFAM), camera SFAM, and Semantic-based Feature Fusion Module (SFFM). + +asymmetric transformations. Besides, 2D3DNet [13] aims to train a 3D model using reduced 3D annotations but more unlabeled data and additional manual 2D annotations to train a 2D model for obtaining 2D segmentation in advance. The 2D segmentation is also painted onto the points as the additional input of 3D model [42]. Such a phased training approach lacks joint 2D-3D optimization. Our motivation, technique, and results are distinguished from them. + +# 3. Method + +# 3.1. Preliminary + +Problem Setup. Let $\{\mathbf{P}_{\mathrm{in}},\mathbf{X}_{\mathrm{in}}\}$ be a multi-modal sample, where $\mathbf{P}_{\mathrm{in}}\in \mathbb{R}^{N_{\mathrm{point}}\times C_{\mathrm{in}}}$ denotes LiDAR point cloud of $N_{\mathrm{point}}$ points with $C_\mathrm{in}$ -dimensional input features (e.g., 3D coordinates and reflectance) and $\mathbf{X}_{\mathrm{in}}\in \mathbb{R}^{N_{\mathrm{cam}}\times 3\times H_{\mathrm{in}}\times W_{\mathrm{in}}}$ denotes multi-camera RGB images of $N_{\mathrm{cam}}$ cameras. Based on sensor calibration, the point with 3D coordinates $(x,y,z)$ can be projected onto the $c$ -th local camera image with the pixel coordinates $(c,u,v)$ . $N_{\mathrm{cls}}$ semantic categories for each point in the 3D point cloud. + +LiDAR Feature Extraction. Given the input $\{\mathbf{P}_{\mathrm{in}},\mathbf{X}_{\mathrm{in}}\}$ in Fig. 2, we first conduct intra-modal feature extraction with parallel backbones for addressing the heterogeneity between point cloud and images. Due to the superiority of 3D voxel, we perform voxel-based LiDAR feature extraction by a general sparse 3D U-Net [9, 35] in Supplementary Material (SM). The input point cloud is first divided by a quantization step $d$ then grouped into non-empty voxels by averaging the point-wise initial features of the local points inside a voxel. These non-empty voxels $\mathbf{V}_{\mathrm{in}}\in \mathbb{R}^{N_{\mathrm{voxel}}\times C_{\mathrm{in}}}$ constitute the sparse tensor input to point cloud backbone for learning the expressive voxel features $\mathbf{V}\in \mathbb{R}^{N_{\mathrm{voxel}}\times C_{\mathrm{voxel}}}$ . + +Camera Feature Extraction. To enhance the LiDAR features with jointly-optimized camera features, we use a trainable image backbone (HRNet-w48 [44] by default) to non-linearly project $\mathbf{X}_{\mathrm{in}}$ as $\mathbf{X} \in \mathbb{R}^{N_{\mathrm{cam}} \times C_{\mathrm{img}} \times H \times W}$ , with down-sampled shape $H \times W$ and increased channels $C_{\mathrm{img}}$ . + +The point cloud and image backbones can be flexibly selected from various mature networks. Moreover, we jointly optimize both backbones in the overall 3D segmentation model to learn relevant semantic feature representations from different modalities. The expressive voxel features $\mathbf{V}$ and image features $\mathbf{X}$ guarantee the inter-modal feature fusion, which consists of GF-Phase, cross-modal feature completion, and SF-Phase in the right half of Fig. 2. + +# 3.2. GF-Phase: Geometry-based Feature Fusion + +The point-centric segmentation motivates us to warp voxel features $\mathbf{V}$ and image feature maps $\mathbf{X}$ to points in $\mathbf{P}_{\mathrm{in}}$ as point-wise LiDAR features $\mathbf{F}_{\mathrm{lidar}} \in \mathbb{R}^{N_{\mathrm{point}} \times C_{\mathrm{voxel}}}$ and point-wise camera features $\mathbf{F}_{\mathrm{cam}} \in \mathbb{R}^{N_{\mathrm{point}} \times C_{\mathrm{img}}}$ by using the geometric association of LiDAR and multi-camera. + +Learning Point Feature From Voxels and Pixels. We devoxelize voxel features $\mathbf{V}$ as point-wise LiDAR features $\mathbf{F}_{\mathrm{lidar}} = [f_{\mathrm{lidar},i}]_{i=1}^{i=N_{\mathrm{point}}} \in \mathbb{R}^{N_{\mathrm{point}} \times C_{\mathrm{voxel}}}$ for augmenting the point-wise fusion and segmentation. Given a point, we interpolate its point feature from the three nearest neighboring voxels [22]. For the $i$ -th point $(x_i, y_i, z_i)$ with LiDAR feature $f_{\mathrm{lidar},i}$ in $\mathbf{F}_{\mathrm{lidar}}$ , we use the known point-wise image coordinates $(c_i, u_i, v_i)$ to perform bilinear interpolation $\mathcal{I}$ on the image feature maps $\mathbf{X}[c_i] \in \mathbb{R}^{C_{\mathrm{img}} \times H \times W}$ within the $c_i$ -th local camera as Eq. 1. Zeros are temporarily padded to those points outside the camera FOV, which allows each point to be decorated with the $C_{\mathrm{img}}$ -dimensional feature $f_{\mathrm{cam},i}$ . To distinguish points outside, we set a binary mask $\mathbf{B}$ of $N_{\mathrm{point}}$ elements with 0 for them and 1 otherwise. + +$$ +f _ {\mathrm {c a m}, i} = \mathcal {I} \left(\mathbf {X} \left[ c _ {i} \right], \frac {H}{H _ {\mathrm {i n}}} u _ {i}, \frac {W}{W _ {\mathrm {i n}}} v _ {i}\right), \tag {1} +$$ + +$$ +\mathbf {F} _ {\mathrm {c a m}} = \left[ f _ {\mathrm {c a m}, i} \right] _ {i = 1} ^ {i = N _ {\text {p o i n t}}} \in \mathbb {R} ^ {N _ {\text {p o i n t}} \times C _ {\mathrm {i m g}}}. \tag {2} +$$ + +Geometry-based Feature Fusion Module (GFFM) first projects $f_{\mathrm{lidar},i}$ and $f_{\mathrm{cam},i}$ as $C_{\mathrm{int}}$ -dimensional by using fully connected layers $\mathcal{F}_{\mathrm{lidar}}$ and $\mathcal{F}_{\mathrm{cam}}$ , and concatenates ("©") them together for subsequent learnable + +fusion using another MLP with $C_{\mathrm{gfused}}$ output channels as $f_{\mathrm{gfused},i} = MLP(\mathcal{F}_{\mathrm{lidar}}(f_{\mathrm{lidar},i}) \odot \mathcal{F}_{\mathrm{cam}}(f_{\mathrm{cam},i}))$ and $\mathbf{F}_{\mathrm{gfused}} = [f_{\mathrm{gfused},i}]_{i=1}^{i=N_{\mathrm{point}}} \in \mathbb{R}^{N_{\mathrm{point}} \times C_{\mathrm{gfused}}}$ . The GF-Phase combines the multi-modal features as $\mathbf{F}_{\mathrm{gfused}}$ mainly with geometry clues, so-called the geometry-based feature fusion phase. + +# 3.3. SF-Phase: Semantic-based Feature Fusion + +Though the simple yet effective GF-Phase yields appreciable performance gains inside FOV intersection through unbiased considerations of both modalities, three drawbacks remain to be addressed. i) The points outside still cannot retrieve their realistic camera features. ii) The relative importance between modalities varies in different areas [5, 58], which is overlooked in GF-Phase. iii) Explicit relational modeling for category-wise semantic representations [18, 19] is not available. As shown in Fig. 2, we further propose the SF-Phase after GF-Phase. Beyond the geometric association in euclidean space, we aggregate LiDAR features and camera features into category-wise semantic embeddings as $\mathbf{E}_{\mathrm{lidar}}$ and $\mathbf{E}_{\mathrm{cam}}$ for performing multi-modal fusion in a hidden semantic space. + +LiDAR Semantic Feature Aggregation Module (LiDAR SFAM) starts from voxel features $\mathbf{V}$ and ends with aggregated $N_{\mathrm{cls}}$ LiDAR semantic embeddings in $\mathbf{E}_{\mathrm{lidar}}$ . For $N_{\mathrm{cls}}$ categories, we assume $\mathbf{E}_{\mathrm{lidar}} \in \mathbb{R}^{N_{\mathrm{cls}} \times C_{\mathrm{voxel}}}$ and a distribution matrix $\mathbf{D}_{\mathrm{lidar}} \in (0,1)^{N_{\mathrm{cls}} \times N_{\mathrm{voxel}}}$ . The $\mathbf{D}[j,i]$ closer to 1 indicates that the $i$ -th voxel is more likely to belong to the $j$ -th category than other voxels, so that it contributes more to describing this category in $\mathbf{E}_{\mathrm{lidar}}[j]$ . Otherwise, the $i$ -th voxel should contribute less to $\mathbf{E}_{\mathrm{lidar}}[j]$ . Now we can compute $\mathbf{E}_{\mathrm{lidar}}$ by matrix product as $\mathbf{E}_{\mathrm{lidar}} = \mathbf{D}_{\mathrm{lidar}}\mathbf{V}$ . Two more steps (Eqs. $3 \sim 4$ ) are required to compute $\mathbf{D}_{\mathrm{lidar}}$ . We first predict an intermediate segmentation $\mathbf{D}_{\mathrm{lidar}}' \in \mathbb{R}^{N_{\mathrm{voxel}} \times N_{\mathrm{cls}}}$ on the voxel features $\mathbf{V}$ using an MLP-based auxiliary voxel segmentation head $\mathcal{H}_{\mathrm{voxel}}$ . We then perform a spatial softmax among voxels to normalize the values into $(0,1)$ for each category. + +$$ +\mathbf {D} _ {\text {l i d a r}} ^ {\prime} = \mathcal {H} _ {\text {v o x e l}} (\mathbf {V}), \tag {3} +$$ + +$$ +\mathbf {D} _ {\text {l i d a r}} ^ {T} = \left[ \operatorname {S o f t m a x} \left(\mathbf {D} _ {\text {l i d a r}} ^ {\prime} [:, j ]\right) \right] _ {j = 1} ^ {j = N _ {\mathrm {c l s}}}. \tag {4} +$$ + +For more accurate $\mathbf{D}_{\mathrm{lidar}}$ , we explicitly guide $\mathbf{D}_{\mathrm{lidar}}^{\prime}$ by using the voxel segmentation supervision in Eq. 15. + +Camera Semantic Feature Aggregation Module (Camera SFAM) similarly aggregates $\mathbf{E}_{\mathrm{cam}} \in \mathbb{R}^{N_{\mathrm{cls}} \times C_{\mathrm{img}}}$ from image feature maps $\mathbf{X} \in \mathbb{R}^{N_{\mathrm{cam}} \times C_{\mathrm{img}} \times H \times W}$ using another distribution matrix $\mathbf{D}_{\mathrm{cam}} \in (0,1)^{N_{\mathrm{cls}} \times N_{\mathrm{pixel}}}$ and another auxiliary image segmentation head $\mathcal{H}_{\mathrm{img}}$ implemented with the simple FCN head [34]. The difference lies in that the camera SFAM is designed with all the $N_{\mathrm{pixel}}$ pixels of image feature maps across $N_{\mathrm{cam}}$ local cameras, where $N_{\mathrm{pixel}}$ is $N_{\mathrm{cam}} \times H \times W$ . + +The $\mathcal{H}_{\mathrm{img}}$ first predicts the intermediate segmentation $\mathbf{D}_{\mathrm{img}}^{\prime} \in \mathbb{R}^{N_{\mathrm{cam}} \times N_{\mathrm{cls}} \times H \times W}$ by $\mathcal{H}_{\mathrm{img}}(\mathbf{X})$ . Note that $\mathbf{D}_{\mathrm{img}}^{\prime}$ + +and $\mathbf{X}$ are then rearranged into $\mathbf{D}_{\mathrm{cam}}^{\prime} \in \mathbb{R}^{N_{\mathrm{cls}} \times N_{\mathrm{pixel}}}$ and $\mathbf{X}^{\prime} \in \mathbb{R}^{N_{\mathrm{pixel}} \times C_{\mathrm{img}}}$ by transposing and reshaping operations. The final camera semantic embeddings $\mathbf{E}_{\mathrm{cam}}$ in Eq. 7 is the matrix product of $\mathbf{D}_{\mathrm{cam}}$ and $\mathbf{X}^{\prime}$ , where the normalized $\mathbf{D}_{\mathrm{cam}}$ is the result of applying spatial softmax on $\mathbf{D}_{\mathrm{cam}}^{\prime}$ as Eq. 6. + +$$ +\mathbf {D} _ {\mathrm {i m g}} ^ {\prime} = \mathcal {H} _ {\mathrm {i m g}} (\mathbf {X}), \tag {5} +$$ + +$$ +\mathbf {D} _ {\mathrm {c a m}} = \left[ \operatorname {S o f t m a x} \left(\mathbf {D} _ {\mathrm {c a m}} ^ {\prime} [ j,: ]\right) \right] _ {j = 1} ^ {j = N _ {\mathrm {c l s}}}, \tag {6} +$$ + +$$ +\mathbf {E} _ {\mathrm {c a m}} = \mathbf {D} _ {\mathrm {c a m}} \mathbf {X} ^ {\prime}. \tag {7} +$$ + +Note that the image segmentation annotations are unavailable for guiding the $\mathbf{D}_{\mathrm{img}}^{\prime}$ in such 3D semantic segmentation datasets [3, 37]. Nevertheless, we also address this by the cross-modal semantic supervision scheme, which is detailed in Eq. 16 and Eq. 17. + +Semantic-based Feature Fusion Module (SFFM). As shown in the right dashed box in Fig. 2, given the input of point-wise geometry-based fused features $\mathbf{F}_{\mathrm{gfused}}$ , category-wise semantic embeddings $\mathbf{E}_{\mathrm{lidar}}$ and $\mathbf{E}_{\mathrm{cam}}$ , the SFFM first projects them into a $C_{\mathrm{sfused}}$ -dimensional space as $\mathbf{F}_{\mathrm{proj}} = Proj_1(\mathbf{F}_{\mathrm{gfused}})$ , and $\mathbf{E} = Proj_2(\mathbf{E}_{\mathrm{lidar}}) \odot Proj_3(\mathbf{E}_{\mathrm{cam}})$ , respectively. Before stacking $K$ gray blocks in SFFM, we concatenate the projected multi-modal semantic embeddings together as $\mathbf{E} \in \mathbb{R}^{2N_{\mathrm{cls}} \times C_{\mathrm{sfused}}}$ , which can be regarded as a dictionary that describes the typical characteristics for each category from the perspectives of LiDAR and cameras. + +In each gray block, we model the semantic relations among all the $2N_{\mathrm{cls}}$ category-wise semantic embeddings using the Multi-Head Self-Attention (MHSA) [41] as Eq. 8, where Norm indicates the LayerNorm operation. Let $C_{\mathrm{shsa}}$ be $C_{\mathrm{sfused}} / N_H$ , the MHSA can be decomposed into $N_{\mathrm{H}}$ Single-Head Self-Attention (SHSA) operations with $\mathbf{E}_h \in \mathbb{R}^{2N_{\mathrm{cls}} \times C_{\mathrm{shsa}}}$ as + +$$ +\mathbf {E} _ {\mathrm {m h s a}} = \operatorname {N o r m} (\mathbf {E} + M H S A (\mathbf {E})), \tag {8} +$$ + +$$ +M H S A (\mathbf {E}) = \left[ S H S A \left(\mathbf {E} _ {h}\right) \right] _ {h = 1} ^ {h = N _ {\mathrm {H}}}, \tag {9} +$$ + +$$ +S H S A \left(\mathbf {E} _ {h}\right) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {h} \mathbf {K} _ {h} ^ {T}}{\sqrt {C _ {\mathrm {s h s a}}}}\right) \mathbf {V} _ {h}. \tag {10} +$$ + +Since $\mathbf{E}$ includes $\mathbf{E}_{\mathrm{lidar}}$ and $\mathbf{E}_{\mathrm{cam}}$ , the attention matrix $Softmax\left(\frac{\mathbf{Q}_h\mathbf{K}_h^T}{\sqrt{C_{\mathrm{shsa}}}}\right)\in \mathbb{R}^{2N_{\mathrm{cls}}\times 2N_{\mathrm{cls}}}$ intuitively includes four parts of relation modeling on $\mathbf{E}_{\mathrm{lidar}}\rightarrow \mathbf{E}_{\mathrm{lidar}},\mathbf{E}_{\mathrm{lidar}}\rightarrow \mathbf{E}_{\mathrm{cam}},$ $\mathbf{E}_{\mathrm{cam}}\rightarrow \mathbf{E}_{\mathrm{lidar}}$ , and $\mathbf{E}_{\mathrm{cam}}\rightarrow \mathbf{E}_{\mathrm{cam}}$ + +After updating $\mathbf{E}$ as $\mathbf{E}_{\mathrm{mhsa}}$ , we perform the further fusion between the point-wise projected features $\mathbf{F}_{\mathrm{proj}}$ and the semantic embeddings $\mathbf{E}_{\mathrm{mhsa}}$ , using the Multi-Head Cross-Attention (MHCA) for semantic relation modeling in Eq 12 and the Feed-Forward Network (FFN) for feature embedding in Eq. 11, respectively. + +$$ +\mathbf {F} _ {\text {s f u s e d}} = \operatorname {N o r m} \left(\mathbf {F} _ {\text {m h c a}} + F F N \left(\mathbf {F} _ {\text {m h c a}}\right)\right), \tag {11} +$$ + +$$ +\mathbf {F} _ {\mathrm {m h c a}} = \operatorname {N o r m} \left(\mathbf {F} _ {\mathrm {p r o j}} + M H C A \left(\mathbf {F} _ {\mathrm {p r o j}}, \mathbf {E} _ {\mathrm {m h s a}}, \mathbf {E} _ {\mathrm {m h s a}}\right)\right). \tag {12} +$$ + +Unlike Eq. 8, the $MHCA$ in Eq. 12 sets the query $\mathbf{Q}$ from point-wise feature $\mathbf{F}_{\mathrm{proj}}$ , and sets the key $\mathbf{K}$ and value $\mathbf{V}$ from semantic embeddings $\mathbf{E}_{\mathrm{mhsa}}$ . Thus, an attention matrix with shape $N_{\mathrm{point}} \times 2N_{\mathrm{cls}}$ is computed to attentively aggregate the more important semantic embeddings respective to individual points, beyond the paired geometric association in GF-Phase. Following the common usages of multi-head attention [4], $K$ (i.e., 6) gray blocks are stacked. + +Discussion on SFFM. As a useful and general technique [4,28,47], multi-head attention [41], although not proposed in this paper, is effectively tailored to multi-modal feature fusion with different motivations and designs by us. i) In Eq. 12, the $MHCA$ enables the point-wise feature to attend to multi-modal semantic embeddings, so that both points inside and outside can be consistently supported by expressive multi-modal semantic embeddings. ii) The $MHCA$ attention matrix computes the relative importance of two modalities to each point, improving the unbiased considerations of modalities in GF-Phase. iii) The $MHSA$ in Eq. 8 explicitly models the category-wise intra-modal and inter-modal semantic relations, deriving the commonality learning. iv) The LiDAR SFAM and Camera SFAM aggregate the long sequences of voxel features $V$ and image features $X'$ into the short sequences of $E_{\text{lidar}}$ and $E_{\text{cam}}$ with $N_{\text{voxel}} / N_{\text{cls}}$ and $N_{\text{pixel}} / N_{\text{cls}}$ times, which enables efficient computation of the multi-head attention. + +# 3.4. Cross-modal Feature Completion + +As shown in Fig. 2, a cross-modal feature completion module with pixel-to-point loss $\mathbf{L}_{\mathrm{pixel2point}}$ (Eq. 13) is set, where the point-wise pseudo-camera feature $\mathbf{F}_{\mathrm{pcam}}$ is mapped from the point-wise LiDAR features $\mathbf{F}_{\mathrm{lidar}}$ by another MLP-based $\mathcal{H}_{\mathrm{pcam}}$ as $\mathcal{H}_{\mathrm{pcam}}(\mathbf{F}_{\mathrm{lidar}})$ . In practice, we compute the mean square error loss $\mathcal{L}_{\mathrm{mse}}$ between the $\mathbf{F}_{\mathrm{pcam}}$ and $\mathbf{F}_{\mathrm{cam}}$ of the points inside for learning from the correctly paired cross-modal features relationship. + +$$ +\mathbf {L} _ {\text {p i x e l 2 p o i n t}} = \mathcal {L} _ {\mathrm {m s e}} \left(\mathbf {B F} _ {\mathrm {p c a m}}, \mathbf {B D e t a c h} \left(\mathbf {F} _ {\mathrm {c a m}}\right)\right). \tag {13} +$$ + +$$ +\mathbf {F} _ {\mathrm {c a m}} [ i,: ] = \mathbf {F} _ {\mathrm {p c a m}} [ i,: ] | _ {i \in \{j \mid \mathbf {B} [ j ] = 0 \}}. \tag {14} +$$ + +Note that the points outside are ignored by the binary mask $\mathbf{B}$ , which is mentioned in GF-Phase. The gradients of $\mathbf{F}_{\mathrm{cam}}$ are also detached for optimizing the learning of $\mathbf{F}_{\mathrm{pcam}}$ . + +We employ such a cross-modal feature completion module due to two motivations: i) Optimizing $\mathbf{L}_{\mathrm{pixel2point}}$ forces LiDAR features $\mathbf{F}_{\mathrm{lidar}}$ to imitate camera features $\mathbf{F}_{\mathrm{cam}}$ with $\mathcal{H}_{\mathrm{pcam}}$ , then the learned pseudo-image features $\mathbf{F}_{\mathrm{pcam}}$ can be switched to replace the padded zeros in $\mathbf{F}_{\mathrm{cam}}$ as Eq. 14 in the inference stage, which serves as the cross-modal feature completion to enhance the feature learning. Thus, we further reduce feature gaps between points outside and inside. ii) $\mathcal{H}_{\mathrm{pcam}}$ transfers rich appearance priors from the camera branch to the LiDAR branch with an effective + +consistency constraint for enhancing the intra-modal feature learning in the training stage. + +# 3.5. Cross-modal Semantic Supervision + +Point Supervision. Let $\mathbf{Y}$ of $N_{\mathrm{point}}$ elements be the point-wise 3D semantic segmentation labels. The value of $\mathbf{Y}[i]$ is in $[0, N_{\mathrm{cls}} - 1]$ , where 0 denotes the ignored category. An MLP based point segmentation head $\mathcal{H}_{\mathrm{point}}$ is built on $\mathbf{F}_{\mathrm{sfused}}$ for the 3D segmentation prediction $\hat{\mathbf{Y}} = \mathcal{H}_{\mathrm{point}}(\mathbf{F}_{\mathrm{sfused}})$ . Following [6, 21, 66], we adopt point loss $\mathbf{L}_{\mathrm{point}}$ as a combination of cross-entropy loss $\mathcal{L}_{\mathrm{ce}}$ and lovasz-softmax loss $\mathcal{L}_{\mathrm{lovasz}}$ [2] as $\mathcal{L}_{\mathrm{ce}}(\hat{\mathbf{Y}}, \mathbf{Y}) + \mathcal{L}_{\mathrm{lovasz}}(\hat{\mathbf{Y}}, \mathbf{Y})$ . + +Point-to-voxel Supervision. For guiding the $\mathbf{D}_{\mathrm{lidar}}^{\prime}$ in Eq. 3, the voxel loss $\mathbf{L}_{\mathrm{p2v}}$ is defined as: + +$$ +\mathbf {L} _ {\mathrm {p} 2 \mathrm {v}} = \mathcal {L} _ {\mathrm {c e}} \left(\mathbf {D} _ {\text {l i d a r}} ^ {\prime}, \mathbf {Y} _ {\mathrm {p} 2 \mathrm {v}}\right) + \mathcal {L} _ {\text {l o v a s z}} \left(\mathbf {D} _ {\text {l i d a r}} ^ {\prime}, \mathbf {Y} _ {\mathrm {p} 2 \mathrm {v}}\right), \tag {15} +$$ + +where the voxel labels $\mathbf{Y}_{\mathrm{p2v}}$ of $N_{\mathrm{voxel}}$ elements can be determined from the labels of points in the voxel. To avoid ambiguity in $\mathbf{L}_{\mathrm{p2v}}$ , the label of the voxel containing points of multiple categories is set to zero as the ignored category. + +Point-to-pixel Supervision. The problem of how to use only point labels to correctly guide $\mathbf{D}_{\mathrm{img}}^{\prime}$ (Eq. 5) is still unsolved. For $\mathbf{D}_{\mathrm{img}}^{\prime}$ of each training sample, we initialize a $N_{\mathrm{cam}} \times H \times W$ map $\mathbf{Y}_{\mathrm{point2pixel}}$ of all zeros as the image segmentation label. We then project each point onto the down-sampled image planes and retrieve the point label $\mathbf{Y}[i]$ for the corresponding nearest $(\langle \cdot \rangle)$ pixel as Eq. 16. Although the generated image label $\mathbf{Y}_{\mathrm{point2pixel}}$ is sparse, it provides sufficient supervision. In SM, Fig. S2 and Fig. S3 provide visualizations of $\mathbf{D}_{\mathrm{img}}^{\prime}$ to show that these sparse point-to-pixel supervisions can propagate to other regions with similar appearance, which guides $\mathbf{D}_{\mathrm{img}}^{\prime}$ correctly. Since $\mathbf{Y}_{\mathrm{point2pixel}}$ is sparse, we only compute the basic cross-entropy loss $\mathcal{L}_{\mathrm{ce}}$ in Eq. 17. + +$$ +\mathbf {Y} _ {\text {p o i n t 2 p i x e l}} \left[ c _ {i}, \left\langle \frac {H}{H _ {\mathrm {i n}}} u _ {i} \right\rangle , \left\langle \frac {W}{W _ {\mathrm {i n}}} u _ {i} \right\rangle \right] = \mathbf {Y} [ i ], \tag {16} +$$ + +$$ +\mathbf {L} _ {\text {p o i n t 2 p i x e l}} = \mathcal {L} _ {\mathrm {c e}} \left(\mathbf {D} _ {\mathrm {i m g}} ^ {\prime}, \mathbf {Y} _ {\text {p o i n t 2 p i x e l}}\right). \tag {17} +$$ + +Total Loss Function. To jointly optimize all the modules in the proposed segmentation model, the total loss combines loss terms across different modalities into $\mathbf{L}$ as $\alpha_{1}\mathbf{L}_{\mathrm{point}} + \alpha_{2}\mathbf{L}_{\mathrm{p2v}} + \alpha_{3}\mathbf{L}_{\mathrm{point2pixel}} + \alpha_{4}\mathbf{L}_{\mathrm{pixel2point}}$ , where $\alpha_{1} \sim \alpha_{4}$ are set as 1.0, 1.0, 0.5, 1.0 to balance loss terms. + +# 3.6. Asymmetric Multi-modal Data Augmentation + +The segmentation output and our multi-modal fusion mechanism are point-centered. In Eq. 1 and Eq. 16, the point and the pixel always can be bridged by the coordinate pair of $(x_{i},y_{i},z_{i})$ and $(c_i,u_i,v_i)$ , as long as we pre-compute the coordinate pairs and keep the order of coordinate pairs among all the points in synchronization. With this ordering constraint, we have the flexibility + +Table 1. Data augmentation transformations on LiDAR (L) point cloud and camera (C) images. We only treat random flipping as a naive symmetric transformation that can be simply applied to both modalities simultaneously. + +
L-only·Global rotation around the Z axis with a random angle in [−π/4, +π/4]. +·Global translation with a random vector (Δx, Δy, Δz) sampled from a Gaussian distribution with mean zero and the standard deviation 0.5. +·Global scaling with a random scaling factor in [0.95, 1.05]. +…
Symmetric·Random flipping along the X, Y axis with probability 0.5.
C-only·Scaling with a random ratio in [1.0, 1.5]. +·Horizontal rotation with a random angle in [−1°, 1°]. +·Random cropping with a size (Hin, Win). +·Color jitter from torchvision [30] with random parameters of 0.3 for brightness, 0.3 for contrast, 0.3 for saturation, 0.1 for hue. +·JPEG with a random compression ratio in [30, 70] and probability 0.5 [64]. +…
+ +to decouple multi-modal data augmentation: i) We can preserve all the LiDAR transformations in Tab. 1 with the images and the coordinates $(c, u, v)$ unchanged. ii) Once coordinates $(c, u, v)$ are synchronously transformed with images, we can further apply the camera-only transformations in Tab. 1 to images, which are asymmetric to the LiDAR transformations. iii) We can apply independent transformations between the images of local cameras since the image view is naturally sliced by local cameras. + +The asymmetry lies in not only the LiDAR and camera world but also the local cameras. Thus, we significantly diversify the multi-modal data augmentation for 3D semantic segmentation using the introduced camera transformations with the preserved LiDAR transformations. More potential transformations can be exploited beyond Tab. 1. + +# 4. Experiment + +# 4.1. Experimental Setup + +nuScenes Dataset splits 28,130 training samples, 6,019 validation samples, and 6,008 testing samples [3, 11]. Each nuScenes sample contains relatively sparser point cloud of 32 beams and RGB images captured by 6 cameras: front, front-left, front-right, back, back-left, and back-right. Points outside are projected below the image bottom due to different vertical FOVs of LiDAR and cameras. Following official protocol, the $N_{\mathrm{cls}}$ is 17. Semantic segmentation annotations are only on point clouds and not on images. + +Waymo Dataset for 3D semantic segmentation includes 23,691 training samples, 5,976 validation samples, and 2,982 testing samples [37]. Each sample contains the point cloud of 64 beams and the RGB images captured by 5 cameras: front, front-left, front-right, side-left, and side-right. There is no rear camera on Waymo ego-vehicle, hence more points outside unfairly increase the difficulty of multimodal segmentation. The $N_{\mathrm{cls}}$ is 23. Similarly, semantic segmentation annotations are only on point clouds. + +SemanticKITTI Dataset is collected by a LiDAR with 64 beams [1]. Following [10,50,66,67], we use sequences + +00 to 10 (excluding 08) with 19,130 training samples and sequence 08 with 4,071 validation samples. It provides only the images of the front-view camera. The $N_{\mathrm{cls}}$ is 20. + +Evaluation Metric is the mean Intersection-over-Union (mIoU) defined as $\frac{1}{C}\sum_{c = 1}^{N_{\mathrm{cls}} - 1}\frac{TP_c}{TP_c + FP_c + FN_c}$ , where $TP_{c}, FP_{c}, FN_{c}$ denote the number of true positive, false positive, and false negative predictions for the $c$ -th category out of $N_{\mathrm{cls}} - 1$ valid categories. Besides, nuScenes computes another metric fwIoU by weighting category-wise IoU with the point frequency of its category. + +Settings. Although lightweight LiDRA-only baseline models can be trained with larger batchsize and more epochs to obtain relatively better results, we still train all the models under the same schedule: a batch of 32 random samples distributed on 16 Tesla V100 GPUs with 24 epochs. More implementation details on point cloud voxelization, network architecture, model training, and inference are all included in SM. + +# 4.2. Comparison with State-of-the-art Methods + +Results on nuScenes. From Tab. 2, multi-modal methods counterintuitively lag far behind LiDAR-only methods. In the top-20 submissions on nuScenes benchmark [27], there are only four multi-modal methods. Details of Cylinder3D++, SPVCNN++, LIFusion, and CPFusion are not available till submission. Our framework achieves the best mIoU and fwIoU with multi-modal fusion mechanism and asymmetric multi-modal data augmentation. Especially, our method achieves superior performance on challenging small objects such as pedestrians and traffic cones, where the laser points are typically insufficient for LiDAR-only methods. Our method also significantly outperforms all public and private multi-modal submissions in Tab. 2. + +Results on Waymo. We perform evaluation on Waymo benchmark [45] in Tab. 3, where the LiDAR sensor provides denser laser points and the multi-camera excludes the rearview. Such denser point cloud and incomplete camera FOV unfairly weaken the advantages of multiple modalities. Although, our method still achieves state-of-the-art performance using a general point cloud backbone network [9, 35]. We believe that the performance can be improved by incorporating stronger point cloud backbone networks from other state-of-the-art LiDAR-only methods. + +Results on SemanticKITTI. While comparisons on the highly competitive nuScenes and Waymo datasets validate the effectiveness of our method, Tab. 4 shows the experimental comparison on SemanticKITTI [1]. Since SemanticKITTI provides only front-view images, PMF [67] evaluates the multi-modal segmentation performance on points inside the FOV intersection. We follow PMF to report $\mathrm{mIoU}^1$ on points inside in Tab. 4. Our method not only shows improvements to the LiDAR-only methods but also outperforms other multi-modal methods. + +Table 2. Performance comparison on nuScenes testing set [27]. The modalities available on nuScenes include LiDAR (L), Camera (C), and Radar (R). Top-1 results are in bold. *Submission entries without a published paper by the CVPR 2023 deadline of Nov 11, 2022. + +
MethodModalitymIoUfIoUbarrierbicyclebuscarcons. vehiclemotorcyclepedestriantraffic-conetrailertruckdriveableothersidewalkterrainmammadevegetation
PolarNet [62]L69.4287.3872.1616.8177.0186.5351.1469.6564.8054.1169.7063.5396.6467.1477.7072.1387.1384.47
JS3C-Net [52]L73.6088.0680.1426.1587.7984.5455.1772.5671.2866.2676.7971.1196.8064.4776.8674.0987.4886.10
Cylinder3D [66]L77.1689.9282.7629.7584.3489.4163.0379.2977.2173.4084.5569.1797.6670.2480.2975.5190.4187.55
AMVNet [24]L77.2790.0880.6431.9681.7388.9367.0784.3376.1173.4884.8767.3097.3767.3779.4175.4591.4588.69
SPVNAS [38]L77.3589.6880.0029.9891.9290.8164.6878.9975.6270.9481.0174.6497.4469.2379.9576.1089.2887.06
Cylinder3D++[66]L77.8689.9382.7633.8984.3489.4169.6379.4277.2673.4084.5569.4197.6670.2480.2975.5190.4287.55
AF2S3Net [7]L78.3488.5178.8752.2189.9384.1777.4274,3077.3271.9583.8873.7897.1366.4777.5174.0187.6986.80
SPVCNN++[38]L81.1290.9786.3543.1391.9092.1875.9075.7283.4477.3186.8277.3697.6971.2281.0877.1991.6788.98
LIFusion* [27]LC75.7489.3258.1336.3086.6784.2859.9679.6980.3077.7783.2368.7497.1868.1977.0474.4591.0388.95
PMF [67]LC77.0389.3482.1140.3380.9486.4263.7279.2279.7575.8681.1767.0597.2867.6978.0574.4889.9488.46
CPFusion* [27]LCR77.7289.1983.6737.0389.0286.2470.0877.4778.0774.5382.7867.9496.6468.2479.5374.9190.4786.95
2D3DNet [13]LC79.9690.0883.0159.3587.9985.0963.7084.3981.9575.9684.7971.9396.8867.3579.8175.9692.0589.18
MSeg3DLC81.1491.3583.1142.4694.9292.0167.1078.5885.6680.4787.5377.3297.7469.8281.2277.8392.3590.07
+ +Table 3. Performance comparison on Waymo testing set [45]. *Submission entries without a published paper by the CVPR 2023 deadline of Nov 11, 2022. + +
MethodmIoU
LiDARMultiNet* [55]71.13
HorizonSegExpert* [45]69.44
SPVCNN++ [38]67.70
PolarFuse* [45]67.28
SalsaNext [40]55.85
MSeg3D70.51
+ +Table 4. Performance comparison on SemanticKITTI [1] validation set following PMF [67]. The results of the other methods are from PMF paper. + +
MethodModalitymIoU1
SalsaNext [10]L59.4
SPVNAS [38]L62.3
Cylinder3D [66]L64.9
PointPainting [42]LC54.5
RGBAL [25]LC56.2
PMF [67]LC63.9
L-BaselineL64.8
MSeg3DLC66.7
+ +Table 5. Analysis on the performance gap on all points and points inside, and the data augmentation (DA). Only GF-Phase is simply used for multi-modal fusion (M-Fusion) here. DA includes LiDAR DA (L-DA) and Multi-modal DA (M-DA). + +
LiDARM-FusionDAnuScenesWaymo
mIoUmIoU1mIoUmIoU1
✓×L-DA72.0070.7667.4867.41
✓GF-Phase×68.1076.6859.7764.79
✓GF-PhaseL-DA71.3577.4860.3465.70
✓GF-PhaseM-DA72.3978.6563.9767.94
+ +Table 6. Analysis on multi-modal feature fusion of GF-Phase, cross-modal (CM) feature completion, and SF-Phase. The supervised CM feature completion is decomposed as $\mathbf{L}_{\mathrm{pixel2point}}$ (Eq. 13) and "Comp." (Eq. 14). The gap can be formulated as mIoU - mIoU1. + +
LiDARM-FusionCM Feature CompletionnuScenesWaymo
GF-PhaseSF-PhaseLpixel2pointComp.mIoUmIoU1GapmIoUmIoU1Gap
✓××××72.0070.76+1.2467.4867.41+0.07
✓✓×××72.3978.65-6.2663.9767.94-3.97
✓✓×✓×76.4479.10-2.6667.1369.06-1.93
✓✓×✓✓78.2879.10-0.8267.8969.06-1.17
✓✓✓✓✓80.0080.10-0.1069.6370.19-0.56
+ +# 4.3. Ablation Study + +Difficulties Arising from Multi-modality are investigated in Tab. 5 from two perspectives of the performance gap between all points and points inside, as well as the multi-modal data augmentation. The experiments in Tab. 5 start with a vanilla variant of the multi-modal model using the basic GF-Phase fusion without the multi-modal data augmentation. The mIoU is relatively worse than the mIoU1, since the geometry-based feature fusion ideally ignores the inevitable points outside with the missing camera features [20, 25, 67], which also motivates us to close the performance gaps between mIoU and mIoU1 by using the cross-modal feature completion and semantic-based SF-Phase. Besides, in rows 2 and 1, the multi-modal segmentation performance will be worsened if we deprecated the augmentation transformations that cannot be directly applied to both modalities. However, the proposed asymmetric multi-modal data augmentation allows the accumulation of the transformations on point cloud and image, achieving the best performance on both datasets. + +Multi-modal Feature Fusion Modules. Tab. 6 shows that our supervised cross-modal feature completion and SF-Phase benefit both mIoU and mIoU1 with gradually narrowed gaps. In row 3, Lpixel2point can also bring improvements, which indicates that transferring appearance information from dense image to sparse point cloud is applicable and beneficial for joint feature learning. In row 4, completing camera features with pseudo-camera features facilitates the feature gap reduction between points inside and outside. Eventually, the SF-Phase further narrows the gaps between mIoU and mIoU1 to the smallest values of 0.10 and 0.56, addressing the limitation of geometric associations by semantic-based feature fusion. Thus, we set the model in the last row as our final framework. + +mIoU Breakdown Over Distance. Fig. 3 presents the distance-based evaluation corresponding to the models in Tab. 6. The LiDAR-only model degrades at long distances due to more sparse points. Instead, the multimodal model in row 5 effectively alleviates the performance + +![](images/681d2b87da08b5ef66710f1cc90178665af009dd34df04811f476cfdf7247b34.jpg) + +![](images/e45d608929ad29f60ff5ed2080c7d5cee422695e4c87498a8db33786491219ce.jpg) + +![](images/eab315ed50cfff5c2985d01f4bbdedf11e962b7c799ca9d0df502695d56cd55e.jpg) +Figure 3. Distance-based mIoU evaluation on nuScenes and Waymo validation (Val.) sets using the models in Tab. 6. + +![](images/4bf6a4846b6c0732af2849c732a60edeec6439c2510c4702c2bc2f20b39bdc77.jpg) + +Table 7. Robustness analysis on MSeg3D by removing some cameras as malfunction. "#Camera" denotes the number of available cameras. "×" denotes LiDAR-only baseline. + +
#Camera6543210×
nuScenes80.0078.6977.6276.6976.0175.4474.4772.00
Waymo-69.6369.1868.8668.1368.0968.0467.48
+ +Table 8. Robustness analysis on MSeg3D with multi-frame point clouds input. "#L-Frame" is the frame number for LiDAR. + +
nuScenes#L-Frame1102025304025
Camera××××××√
mIoU72.0074.6675.3775.7775.2875.1581.12
Waymo#L-Frame151015203010
Camera××××××√
mIoU67.4869.1869.4568.8868.7868.6470.20
+ +degradation. In Fig. 3 (a) and (b), improvements between the multi-modal model in row 5 and the LiDAR-only model in row 1 are also increased along the distance due to increased sparsity. The reasons why the improvement of the multi-modal model on Waymo is not as similar as that on nuScenes are further analyzed in Fig. 3 (c) and (d). The Waymo point cloud has denser points, which reduces the dependence on the image. More points outside are distributed along different distances due to no rear cameras on Waymo. Although more points outside hinder the applicability of multi-modal fusion, our final multi-modal model still yields a low-performance gap of $0.56\mathrm{mIoU}$ on Waymo in row 5. + +Robustness Against Camera Malfunction. In Tab. 7, our MSeg3D performs properly under the unfavorable condition of camera malfunction, and it can easily deal with the practical situation without switching the model. + +Table 9. Scalability analysis by barely varying the backbone settings on nuScenes validation set. + +
Image BackbonemIoU#Params(M)Latency(s)
×72.0021.280.083
SegFormer-B0 [47]78.1425.330.204
SegFormer-B5 [47]78.89103.480.479
HRNet-w18 [44]79.2131.560.265
ResNet101 [15]79.3664.600.567
HRNet-w48 [44]80.0087.340.445
+ +Note that even with all cameras removed, MSeg3D still outperforms the LiDAR-only baseline, where our cross-modal feature completion supervision provides effective cross-modal information transfer in training. + +Robustness Against Multi-frame Point Clouds Input. Since the densified points alleviate the sparsity, collapsing laser points from neighboring frames usually boosts LiDAR-only 3D perception [16, 52]. In Tab. 8, we follow [57] collapsing multiple previous frames to the current frame from the provided ego-vehicle motion information. For LiDAR-only model, the improvements are saturated with 25 and 10 frames on nuScenes and Waymo. Under such conditions, our MSeg3D still achieves further improvements in the last column, which shows that multi-frame point clouds input can be an optional extension. + +Complexity Scalability. From Tab. 9, a lightweight image backbone also delivers significant performance gains due to multi-modal input. The best-performing HRNet-w48 is our default image backbone. However, the multi-camera image input makes the efficiency bottleneck lie in the image backbone, which deserves in-depth study for autonomous driving [23]. We believe that real-time image segmentation networks such as BiSeNet [59,60] and Fast-SCNN [29] can accelerate our approach. Overall, our MSeg3D can adapt to complexity scalability among various backbones, which flexibly trades off performance and efficiency. + +# 5. Conclusion + +We propose a novel multi-modal 3D semantic segmentation method termed MSeg3D for autonomous driving, based on LiDAR and multi-camera sensors. Our cross-modal feature completion and semantic-based feature fusion solve the overlooked problem, where multi-modal fusion can occur only in the sensor FOV intersection. The proposed asymmetric multi-modal data augmentation enables the multi-modal segmentation model to be effectively trained with reliable performance. Our method achieves state-of-the-art performance on nuScenes, Waymo, and SemanticKITTI. The comprehensive experiments validate improvements and robustness. We hope that our work can inspire further investigation into multi-modal fusion for 3D semantic segmentation in autonomous driving. + +Acknowledgment: This work was supported by the National Key Research and Development Program of China (2018YFE0183900) and YUNJI Technology Co. Ltd. + +# References + +[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jürgen Gall. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In ICCV, pages 9296-9306, 2019. 1, 2, 6, 7 +[2] Maxim Berman, Amal Rannen Triki, and Matthew B. Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, pages 4413-4421, 2018. 5 +[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, pages 11618-11628, 2020. 1, 2, 4, 6 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 5 +[5] Bowen Cheng, Ross B. Girshick, Piotr Dólar, Alexander C. Berg, and Alexander Kirillov. Boundary IoU: Improving object-centric image segmentation evaluation. In CVPR, pages 15334-15342, 2021. 4 +[6] Hui-Xian Cheng, Xian-Feng Han, and Guo-Qiang Xiao. CENet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In ICME, pages 1-6. IEEE, 2022. 5 +[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (AF)2-S3Net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. 2, 7 +[8] Christopher B. Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2 +[9] OpenPCDet Contributors. OpenPCDet: An open-source toolbox for 3D object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2023. 3, 6 +[10] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. SalsaNext: fast, uncertainty-aware semantic segmentation of LiDAR point clouds for autonomous driving. arXiv preprint arXiv:2003.03653, 2020. 6, 7 +[11] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic NuScenes: A large-scale benchmark for LiDAR panoptic segmentation and tracking. IEEE Robotics Autom. Lett., 7(2):3795-3802, 2022. 1, 6 +[12] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 1 +[13] Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Pantofaru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas A. Funkhouser. Learning 3D semantic segmentation with only 2D image supervision. In 3DV, pages 361-372. IEEE, 2021. 1, 2, 3, 7 +[14] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D semantic segmentation with submanifold sparse + +convolutional networks. In CVPR, pages 9224-9232, 2018. 1 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 8 +[16] Peiyun Hu, Jason Ziglar, David Held, and Deva Ramanan. What you see is what you get: Exploiting visibility for 3D object detection. In CVPR, pages 10998-11006, 2020. 8 +[17] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. RandLA-Net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11105–11114, 2020. 2 +[18] Zhenchao Jin, Tao Gong, Dongdong Yu, Q. Chu, Jian Wang, Changhu Wang, and Jie Shao. Mining contextual information beyond image for semantic segmentation. In ICCV, pages 7211-7221, 2021. 4 +[19] Zhenchao Jin, Bin Liu, Qi Chu, and Nenghai Yu. ISNet: Integrate image-level and semantic-level context for semantic segmentation. In ICCV, pages 7169-7178, 2021. 4 +[20] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. FuseSeg: LiDAR point cloud segmentation fusing multi-modal data. In WACV, pages 1863-1872, 2020. 1, 2, 7 +[21] Jiale Li, Hang Dai, and Yong Ding. Self-distillation for robust LiDAR semantic segmentation in autonomous driving. In ECCV, volume 13688, pages 659-676, 2022. 2, 5 +[22] Jiale Li, Hang Dai, Ling Shao, and Yong Ding. From voxel to point: IoU-guided 3D object detection for point cloud with voxel-to-point decoder. In ACM MM, pages 4622-4631, 2021. 3 +[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 8 +[24] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. AMVNet: Assertion-based multi-view fusion network for LiDAR semantic segmentation. CoRR, abs/2012.04934, 2020. 7 +[25] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Kumar Yogamani. RGB and LiDAR fusion based 3D semantic segmentation for autonomous driving. In ITSC, pages 7-12, 2019. 1, 2, 7 +[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. RangeNet++: Fast and accurate LiDAR semantic segmentation. In IROS, pages 4213-4220, 2019. 2 +[27] nuScenes. nuScenes leaderboard of lidar segmentation task. https://www.nuscenes.org/lidar-segmentation?externalData $\equiv$ all&mapData $=$ all&modalities $\equiv$ Any, 2023.1,6,7 +[28] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3D object detection with pointformer. In CVPR, pages 7463-7472, 2021. 5 +[29] Rudra P. K. Poudel, Stephan Liwicki, and Roberto Cipolla. Fast-SCNN: Fast semantic segmentation network. In BMVC, page 289, 2019. 8 + +[30] Pytorch. Pytorch document of color jitter. https://torch.org/vision/stable/generated/torchvision.transforms.ColorJitter.html?highlight=colorjitter,2023.6 +[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1, 2 +[32] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 2 +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells III, and Alejandro F. Frangi, editors, MICCAI, volume 9351, pages 234–241, 2015. 2 +[34] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. IEEE TPAMI, 39(4):640-651, 2017. 1, 4 +[35] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3D object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2021. 3, 6 +[36] Vishwanath A. Sindagi, Yin Zhou, and Oncel Tuzel. MVX-Net: Multimodal VoxelNet for 3D object detection. In ICRA, pages 7276–7282, 2019. 1 +[37] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2443-2451, 2020. 1, 2, 4, 6 +[38] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3D architectures with sparse point-voxel convolution. In ECCV, volume 12373, pages 685–702, 2020. 2, 7 +[39] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPConv: Flexible and deformable convolution for point clouds. In ICCV, pages 6410–6419, 2019. 2 +[40] Cortinhal Tiago, Tzelepis George, and Erdal Aksoy Eren. SalsaNext: Fast, uncertainty-aware semantic segmentation of LiDAR point clouds. In Advances in Visual Computing, pages 207-222, 2020. 2, 7 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2, 4, 5 +[42] Sourabh Vora, Alex H. Lang, Bassam Helou, and Oscar Beijbom. PointPainting: Sequential fusion for 3D object detection. In CVPR, pages 4603-4611, 2020. 1, 3, 7 +[43] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. PointAugmenting: Cross-modal augmentation for 3D object detection. In CVPR, pages 11794–11803, 2021. 2 + +[44] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349-3364, 2021. 1, 3, 8 +[45] Waymo. Waymo Open Dataset leaderboard of 3D semantic segmentation challenge. https://waymo.com/open/challenges/2022/3d-semantic-segmentation/, 2023.6,7 +[46] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. SqueezeSegV2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a LiDAR point cloud. In ICRA, pages 4376-4382, 2019. 2 +[47] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 1, 5, 8 +[48] Liang Xie, Chao Xiang, Zhengxu Yu, Guodong Xu, Zheng Yang, Deng Cai, and Xiaofei He. PI-RCNN: an efficient multi-sensor 3D object detector with point-based attentive cont-conv fusion module. In AAAI, pages 12460–12467, 2020. 1 +[49] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In ECCV, pages 1-19, 2020. 2 +[50] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVNet: A deep and efficient range-point-voxel fusion network for LiDAR point cloud segmentation. In ICCV, pages 16004-16013, 2021. 2, 6 +[51] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Bin Zhou, and Liangjun Zhang. FusionPainting: Multimodal fusion with adaptive attention for 3D object detection. In ITSC, pages 3047-3054, 2021. 1 +[52] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 7, 8 +[53] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. PointASNL: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, pages 5588-5597, 2020. 2 +[54] Yan Yan, Yuxing Mao, and Bo Li. SECOND: sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2 +[55] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. LidarMultiNet: Towards a unified multi-task network for LiDAR perception. CoRR, abs/2209.09385, 2022. 7 +[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6):210:1-210:12, 2016. 2 + +[57] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3D object detection and tracking. In CVPR, pages 11784-11793, 2021. 8 +[58] Tianwei Yin, Xingyi Zhou, and Philipp Krähenbuhl. Multimodal virtual point 3D detection. In NeurIPS, pages 16494-16507, 2021. 4 +[59] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. BiSeNet V2: bilateral network with guided aggregation for real-time semantic segmentation. IJCV, 129(11):3051-3068, 2021. 8 +[60] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. BiSeNet: Bilateral segmentation network for real-time semantic segmentation. In ECCV, volume 11217, pages 334-349, 2018. 8 +[61] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In ECCV, volume 12351, pages 173-190, 2020. 1 +[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. PolarNet: An improved grid representation for online LiDAR point clouds semantic segmentation. In CVPR, pages 9598-9607, 2020. 2, 7 +[63] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 6230-6239, 2017. 1 +[64] Stephan Zheng, Yang Song, Thomas Leung, and Ian J. Goodfellow. Improving the robustness of deep neural networks via stability training. In CVPR, pages 4480-4488, 2016. 6 +[65] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-PolarNet: Proposal-free LiDAR point cloud panoptic segmentation. In CVPR, pages 13194-13203, 2021. 2 +[66] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation. In CVPR, pages 9939-9948, 2021. 2, 5, 6, 7 +[67] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multisensor fusion for 3D LiDAR semantic segmentation. In ICCV, pages 16260-16270, 2021. 1, 2, 6, 7 \ No newline at end of file diff --git a/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/images.zip b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9666ccacfa3aa26a4c648669674cfdd8a2a9f64a --- /dev/null +++ b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a70eff2087d9a8b76728d88f23b48df7fb82bd8fa9e912b75ee1dcf0adc83de +size 565627 diff --git a/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/layout.json b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..689cbfed766ca038134cc768f6a60327c509f53e --- /dev/null +++ b/2023/MSeg3D_ Multi-Modal 3D Semantic Segmentation for Autonomous Driving/layout.json @@ -0,0 +1,11486 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 529, + 121 + ], + "type": "text", + "content": "MSeg3D: Multi-modal 3D Semantic Segmentation for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "spans": [ + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "text", + "content": "Jiale Li" + }, + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "text", + "content": " Hang Dai" + }, + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "text", + "content": " Hao Han" + }, + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "text", + "content": " Yong Ding" + }, + { + "bbox": [ + 171, + 142, + 424, + 157 + ], + "type": "inline_equation", + "content": "^{3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 157, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 157, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 100, + 157, + 493, + 171 + ], + "type": "text", + "content": "1College of Information Science and Electronic Engineering, Zhejiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 171, + 428, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 171, + 428, + 185 + ], + "spans": [ + { + "bbox": [ + 164, + 171, + 428, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 171, + 428, + 185 + ], + "type": "text", + "content": "School of Computing Science, University of Glasgow" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 185, + 433, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 185, + 433, + 200 + ], + "spans": [ + { + "bbox": [ + 160, + 185, + 433, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 160, + 185, + 433, + 200 + ], + "type": "text", + "content": "School of Micro-Nano Electronics, Zhejiang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 201, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 201, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 481, + 213 + ], + "type": "text", + "content": "*Corresponding authors{Hang.Dai@glasgow.ac.uk, dingyong09@zju.edu.cn}." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 251, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 552 + ], + "type": "text", + "content": "LiDAR and camera are two modalities available for 3D semantic segmentation in autonomous driving. The popular LiDAR-only methods severely suffer from inferior segmentation on small and distant objects due to insufficient laser points, while the robust multi-modal solution is under-explored, where we investigate three crucial inherent difficulties: modality heterogeneity, limited sensor field of view intersection, and multi-modal data augmentation. We propose a multi-modal 3D semantic segmentation model (MSeg3D) with joint intra-modal feature extraction and inter-modal feature fusion to mitigate the modality heterogeneity. The multi-modal fusion in MSeg3D consists of geometry-based feature fusion GF-Phase, cross-modal feature completion, and semantic-based feature fusion SF-Phase on all visible points. The multi-modal data augmentation is reinvigorated by applying asymmetric transformations on LiDAR point cloud and multi-camera images individually, which benefits the model training with diversified augmentation transformations. MSeg3D achieves state-of-the-art results on nuScenes, Waymo, and SemanticKITTI datasets. Under the malfunctioning multicamera input and the multi-frame point clouds input, MSeg3D still shows robustness and improves the LiDAR-only baseline. Our code is publicly available at https://github.com/jialeli1/lidarseg3d." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 564, + 128, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 128, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 128, + 576 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Scene understanding for safe autonomous driving can be achieved through semantic segmentation using camera 2D images and LiDAR 3D point clouds, which densely classifies each smallest sensing unit of the modality. The image-based 2D semantic segmentation has been developed with massive solid studies [12, 34, 61, 63]. The camera image has rich appearance information about the object but severely suffers from illumination, varying object scales, and indirect applications in the 3D world. Another modality, LiDAR point cloud, drives 3D semantic segmentation with laser points [1, 3, 11, 37]. Unfortunately, irregular" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 547, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 288 + ], + "type": "text", + "content": "laser points are too sparse to capture the details of objects. The inaccurate segmentation appears especially on small and distant objects. The other under-explored direction is using multi-modal data to increase both the robustness and accuracy in 3D semantic segmentation [67]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 291, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 547, + 602 + ], + "type": "text", + "content": "Despite the conceptual superiority, the development of multi-modal segmentation model is still nontrivial, lagging behind the single-modal methods [27]. We rationally attribute the difficulties to the three following aspects. i) Heterogeneity between modalities. Due to sparse points and dense pixels, point cloud feature extractors [14, 31] and image feature extractors [15, 44, 47] are developed distinctly. Separate intra-modal feature extractors are used to address the heterogeneity [13, 20, 25, 42, 51], but the lack of joint optimization leads to suboptimal features from irrelevant network parameters. ii) Limited intersection on the field of view (FOV) between sensors. Only the points falling into the intersected FOV are geometrically associated with multi-modal data, while simply considering the intersected multi-modal data is not sufficient to be practically applicable. Performing fusion solely in the limited FOV intersection like [20,67] results in unsatisfactory overall segmentation performance as shown in Fig. 1. iii) Multi-modal data augmentation. For example, PMF [67] uses only several 2D augmentations for spatially aligned point cloud projection image and camera RGB image. Under the constraint of modal alignment or 2D representation of point cloud, the multimodal segmentation works [20, 25, 67] discard many useful and critical point cloud augmentation transformations with sacrificed perception performance [36, 48]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Accordingly, we propose a top-performing multi-modal 3D semantic segmentation method termed MSeg3D, inherently motivated by addressing the aforementioned three technical difficulties. i) Unlike separately extracting modal features in existing methods [13, 25, 42, 51], we jointly optimize intra-modal feature extraction and inter-modal feature fusion to drive maximum correlation and complementarity between heterogeneous modalities. ii) To overcome the disregarded multi-modal fusion outside FOV" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21694" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "intersection [25, 67], we propose a cross-modal feature completion and a semantic-based feature fusion phase SF-Phase to collaborate with the geometry-based feature fusion phase GF-Phase. For points outside the FOV intersection, the former completes the missing camera features using predicted pseudo-camera features, under the explicit guidance of cross-modal supervision. For all the points outside and inside the FOV intersection, the later SF-Phase leverages the multi-head attention [41] to model the semantic relation between point and interested categories so that we can attentively fuse the semantic embeddings aggregated from all the visible fields to each point. iii) The challenging multi-modal data augmentation is reinvigorated by being decomposed as the asymmetric transformations in the LiDAR, camera worlds, and local cameras, which enables flexible permutation to enrich training samples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "text", + "content": "As the proposed components accumulated in Fig. 1, mIoU and " + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}^1" + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "text", + "content": " are significantly increased while the gaps between mIoU and " + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}^1" + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "text", + "content": " are gradually decreased. Our contributions are four-fold: i) We propose a multi-modal segmentation model MSeg3D with joint intra-modal feature extraction and inter-modal feature fusion, achieving state-of-the-art 3D segmentation performance on the competitive nuScenes [3], Waymo [37], and SemanticKITTI [1] datasets for autonomous driving. The proposed framework won " + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "text", + "content": " place in the Waymo 3D semantic segmentation challenge at CVPR 2022. ii) We propose a cross-modal feature completion and a semantic-based feature fusion phase. To our best knowledge, it is the first time to address the overlooked and inapplicable multi-modal fusion outside the sensor FOV intersection. iii) By applying augmentation transformations asymmetrically on point cloud and images, the proposed asymmetrical multi-modal data augmentation significantly increases the diversity of multi-modal samples for training model with robust improvements. iv) Extensive experimental analyses on the improvement and robustness of our method clearly investigate our designs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": "LiDAR-only 3D Semantic Segmentation is promoted by SemanticKITTI [1], nuScenes [3], and Waymo [37] datasets. The methods follow the U-Net [33] architecture, but progress from three point cloud representations. i) Point. The point methods derived from PointNet++ [31] cost heavy computation on sampling and gathering disordered neighbors, especially on a large-scale LiDAR point cloud. The major point methods [17, 32, 39, 53] perform well on small synthetic point cloud [56] rather than sparse LiDAR point cloud. ii) 2D images. PolarNetseries [62, 65] and others [26, 40, 46, 49] project 3D point" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 533, + 257 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 533, + 257 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 533, + 257 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 533, + 257 + ], + "type": "image", + "image_path": "aa4344f810a7997690cdd94c737b4fce6b7a142acc6ee40926ee7928c15aad3e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 257, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 312 + ], + "type": "text", + "content": "Figure 1. Performance (mIoU and mIoU" + }, + { + "bbox": [ + 304, + 257, + 547, + 312 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 257, + 547, + 312 + ], + "type": "text", + "content": ") evaluation on nuScenes [3] (a) and Waymo [37] (b) validation sets. All points are distinguished as points inside the sensor FOV intersection (points inside) and points outside the sensor FOV intersection (points outside). We cumulatively add (“+”) the proposed components." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 317, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 449 + ], + "type": "text", + "content": "cloud as 2D images in bird's-eye-view and range-view, achieving efficient LiDAR segmentation with 2D CNNs. But the 3D-to-2D projection damages 3D information and performance. iii) 3D Voxel. The state-of-the-art SPVNAS [38], Cylinder3D [66] and SDSeg3D [21] explicitly explore 3D structure information in different 3D coordinate systems, which efficiently perform 3D sparse convolutions on non-empty voxels [8, 54]. The point and image representations also potentially facilitate voxel features in RPVNet [50] and AF2-S3Net [7]. We also perform effective and efficient 3D voxel feature learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "type": "text", + "content": "Multi-modal 3D Semantic Segmentation is investigated in [13, 20, 25, 67] using LiDAR and camera with unsatisfactory performance. The suboptimal feature sources of RGB value or separately learned CNN feature are warped into point cloud range image as additional input features [20, 25]. Recently, PMF [67] projects point cloud onto the image plane for fusion with image data, which is obviously limited by camera FOV. They discard the fusion and segmentation over points outside. Despite the improved segmentation performance within the FOV, another LiDAR model is required to take responsibility for the remaining area [43, 67]. Such inequity and inconvenience have prevented researchers from focusing on multi-modal 3D semantic segmentation. The image format of point cloud in the aforementioned works limits them to do only image augmentation such as 2D flipping rather than 3D point cloud augmentation. Instead, we point out the necessity of point cloud augmentation for our 3D segmentation network. Unlike another multi-modal 3D object detector PointAugmenting [43] that preserves the point cloud augmentation while keeping the image unchanged, our method starts from both modalities with" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "type": "text", + "content": "1In the following text, mIoU1 denotes the segmentation performance evaluated on only the points inside FOV intersection by excluding the points outside like PMF [67] and other methods [20, 25]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21695" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 217 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 217 + ], + "type": "image", + "image_path": "47cad5a6f8e5c2f3cc200593163ae883e4f473d4ed37a6d0a0725bc124295b7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": "Figure 2. Overview of our multi-modal 3D semantic segmentation model (MSeg3D). For multi-modal feature fusion, GF-Phase mainly includes the Geometry-based Feature Fusion Module (GFFM), while SF-Phase consists of LiDAR Semantic Feature Aggregation Module (SFAM), camera SFAM, and Semantic-based Feature Fusion Module (SFFM)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 289, + 357 + ], + "type": "text", + "content": "asymmetric transformations. Besides, 2D3DNet [13] aims to train a 3D model using reduced 3D annotations but more unlabeled data and additional manual 2D annotations to train a 2D model for obtaining 2D segmentation in advance. The 2D segmentation is also painted onto the points as the additional input of 3D model [42]. Such a phased training approach lacks joint 2D-3D optimization. Our motivation, technique, and results are distinguished from them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 364, + 105, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 105, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 105, + 376 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 384, + 126, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 126, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 126, + 396 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": "Problem Setup. Let " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_{\\mathrm{in}},\\mathbf{X}_{\\mathrm{in}}\\}" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": " be a multi-modal sample, where " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{point}}\\times C_{\\mathrm{in}}}" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": " denotes LiDAR point cloud of " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{point}}" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": " points with " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "C_\\mathrm{in}" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": "-dimensional input features (e.g., 3D coordinates and reflectance) and " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{cam}}\\times 3\\times H_{\\mathrm{in}}\\times W_{\\mathrm{in}}}" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": " denotes multi-camera RGB images of " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": " cameras. Based on sensor calibration, the point with 3D coordinates " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": " can be projected onto the " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": "-th local camera image with the pixel coordinates " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "(c,u,v)" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": " semantic categories for each point in the 3D point cloud." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "text", + "content": "LiDAR Feature Extraction. Given the input " + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_{\\mathrm{in}},\\mathbf{X}_{\\mathrm{in}}\\}" + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "text", + "content": " in Fig. 2, we first conduct intra-modal feature extraction with parallel backbones for addressing the heterogeneity between point cloud and images. Due to the superiority of 3D voxel, we perform voxel-based LiDAR feature extraction by a general sparse 3D U-Net [9, 35] in Supplementary Material (SM). The input point cloud is first divided by a quantization step " + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "text", + "content": " then grouped into non-empty voxels by averaging the point-wise initial features of the local points inside a voxel. These non-empty voxels " + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathrm{in}}\\in \\mathbb{R}^{N_{\\mathrm{voxel}}\\times C_{\\mathrm{in}}}" + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "text", + "content": " constitute the sparse tensor input to point cloud backbone for learning the expressive voxel features " + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{V}\\in \\mathbb{R}^{N_{\\mathrm{voxel}}\\times C_{\\mathrm{voxel}}}" + }, + { + "bbox": [ + 46, + 510, + 288, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Camera Feature Extraction. To enhance the LiDAR features with jointly-optimized camera features, we use a trainable image backbone (HRNet-w48 [44] by default) to non-linearly project " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times C_{\\mathrm{img}} \\times H \\times W}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", with down-sampled shape " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " and increased channels " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 262, + 547, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 357 + ], + "type": "text", + "content": "The point cloud and image backbones can be flexibly selected from various mature networks. Moreover, we jointly optimize both backbones in the overall 3D segmentation model to learn relevant semantic feature representations from different modalities. The expressive voxel features " + }, + { + "bbox": [ + 304, + 262, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 262, + 547, + 357 + ], + "type": "text", + "content": " and image features " + }, + { + "bbox": [ + 304, + 262, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 262, + 547, + 357 + ], + "type": "text", + "content": " guarantee the inter-modal feature fusion, which consists of GF-Phase, cross-modal feature completion, and SF-Phase in the right half of Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 364, + 533, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 533, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 533, + 376 + ], + "type": "text", + "content": "3.2. GF-Phase: Geometry-based Feature Fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "text", + "content": "The point-centric segmentation motivates us to warp voxel features " + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "text", + "content": " and image feature maps " + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "text", + "content": " to points in " + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "text", + "content": " as point-wise LiDAR features " + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{lidar}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{voxel}}}" + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "text", + "content": " and point-wise camera features " + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{cam}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{img}}}" + }, + { + "bbox": [ + 304, + 382, + 545, + 441 + ], + "type": "text", + "content": " by using the geometric association of LiDAR and multi-camera." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": "Learning Point Feature From Voxels and Pixels. We devoxelize voxel features " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " as point-wise LiDAR features " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{lidar}} = [f_{\\mathrm{lidar},i}]_{i=1}^{i=N_{\\mathrm{point}}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{voxel}}}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " for augmenting the point-wise fusion and segmentation. Given a point, we interpolate its point feature from the three nearest neighboring voxels [22]. For the " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": "-th point " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "(x_i, y_i, z_i)" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " with LiDAR feature " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{lidar},i}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": ", we use the known point-wise image coordinates " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "(c_i, u_i, v_i)" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " to perform bilinear interpolation " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " on the image feature maps " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{X}[c_i] \\in \\mathbb{R}^{C_{\\mathrm{img}} \\times H \\times W}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " within the " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": "-th local camera as Eq. 1. Zeros are temporarily padded to those points outside the camera FOV, which allows each point to be decorated with the " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{img}}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": "-dimensional feature " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cam},i}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": ". To distinguish points outside, we set a binary mask " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{point}}" + }, + { + "bbox": [ + 304, + 442, + 546, + 612 + ], + "type": "text", + "content": " elements with 0 for them and 1 otherwise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 617, + 545, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 617, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 348, + 617, + 545, + 642 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {c a m}, i} = \\mathcal {I} \\left(\\mathbf {X} \\left[ c _ {i} \\right], \\frac {H}{H _ {\\mathrm {i n}}} u _ {i}, \\frac {W}{W _ {\\mathrm {i n}}} v _ {i}\\right), \\tag {1}", + "image_path": "152487717010c2e49298661a5bff4b3f33da23569bf14bac722a0083472e13d2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 643, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 643, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 353, + 643, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {c a m}} = \\left[ f _ {\\mathrm {c a m}, i} \\right] _ {i = 1} ^ {i = N _ {\\text {p o i n t}}} \\in \\mathbb {R} ^ {N _ {\\text {p o i n t}} \\times C _ {\\mathrm {i m g}}}. \\tag {2}", + "image_path": "957f978fcc4b403e6da7435739ffeb6a41a6ffbc6773ab17676a694294808eb3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Geometry-based Feature Fusion Module (GFFM) first projects " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{lidar},i}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cam},i}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{int}}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-dimensional by using fully connected layers " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", and concatenates (\"©\") them together for subsequent learnable" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21696" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": "fusion using another MLP with " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{gfused}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " output channels as " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{gfused},i} = MLP(\\mathcal{F}_{\\mathrm{lidar}}(f_{\\mathrm{lidar},i}) \\odot \\mathcal{F}_{\\mathrm{cam}}(f_{\\mathrm{cam},i}))" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{gfused}} = [f_{\\mathrm{gfused},i}]_{i=1}^{i=N_{\\mathrm{point}}} \\in \\mathbb{R}^{N_{\\mathrm{point}} \\times C_{\\mathrm{gfused}}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": ". The GF-Phase combines the multi-modal features as " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{gfused}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " mainly with geometry clues, so-called the geometry-based feature fusion phase." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 268, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 268, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 268, + 152 + ], + "type": "text", + "content": "3.3. SF-Phase: Semantic-based Feature Fusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 158, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 326 + ], + "type": "text", + "content": "Though the simple yet effective GF-Phase yields appreciable performance gains inside FOV intersection through unbiased considerations of both modalities, three drawbacks remain to be addressed. i) The points outside still cannot retrieve their realistic camera features. ii) The relative importance between modalities varies in different areas [5, 58], which is overlooked in GF-Phase. iii) Explicit relational modeling for category-wise semantic representations [18, 19] is not available. As shown in Fig. 2, we further propose the SF-Phase after GF-Phase. Beyond the geometric association in euclidean space, we aggregate LiDAR features and camera features into category-wise semantic embeddings as " + }, + { + "bbox": [ + 46, + 158, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 46, + 158, + 287, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 158, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 158, + 287, + 326 + ], + "type": "text", + "content": " for performing multi-modal fusion in a hidden semantic space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": "LiDAR Semantic Feature Aggregation Module (LiDAR SFAM) starts from voxel features " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " and ends with aggregated " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " LiDAR semantic embeddings in " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " categories, we assume " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times C_{\\mathrm{voxel}}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " and a distribution matrix " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{lidar}} \\in (0,1)^{N_{\\mathrm{cls}} \\times N_{\\mathrm{voxel}}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{D}[j,i]" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " closer to 1 indicates that the " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": "-th voxel is more likely to belong to the " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": "-th category than other voxels, so that it contributes more to describing this category in " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}}[j]" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": ". Otherwise, the " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": "-th voxel should contribute less to " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}}[j]" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": ". Now we can compute " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " by matrix product as " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}} = \\mathbf{D}_{\\mathrm{lidar}}\\mathbf{V}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": ". Two more steps (Eqs. " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "3 \\sim 4" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": ") are required to compute " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": ". We first predict an intermediate segmentation " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{lidar}}' \\in \\mathbb{R}^{N_{\\mathrm{voxel}} \\times N_{\\mathrm{cls}}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " on the voxel features " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " using an MLP-based auxiliary voxel segmentation head " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{voxel}}" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": ". We then perform a spatial softmax among voxels to normalize the values into " + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 46, + 326, + 288, + 517 + ], + "type": "text", + "content": " for each category." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 522, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 522, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 88, + 522, + 287, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime} = \\mathcal {H} _ {\\text {v o x e l}} (\\mathbf {V}), \\tag {3}", + "image_path": "f6a04d27ffcd1eaa05ba051c2e15eb3415f616bdbc697610f865a9cc47e87c45.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 537, + 287, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 537, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 88, + 537, + 287, + 554 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\text {l i d a r}} ^ {T} = \\left[ \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime} [:, j ]\\right) \\right] _ {j = 1} ^ {j = N _ {\\mathrm {c l s}}}. \\tag {4}", + "image_path": "2ee15a2181fd262eeb6f1c5b53e7dd787c179358dc75644173026ab1a01dd33c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 581 + ], + "type": "text", + "content": "For more accurate " + }, + { + "bbox": [ + 46, + 558, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 46, + 558, + 287, + 581 + ], + "type": "text", + "content": ", we explicitly guide " + }, + { + "bbox": [ + 46, + 558, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{lidar}}^{\\prime}" + }, + { + "bbox": [ + 46, + 558, + 287, + 581 + ], + "type": "text", + "content": " by using the voxel segmentation supervision in Eq. 15." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": "Camera Semantic Feature Aggregation Module (Camera SFAM) similarly aggregates " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{cam}} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times C_{\\mathrm{img}}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " from image feature maps " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times C_{\\mathrm{img}} \\times H \\times W}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " using another distribution matrix " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{cam}} \\in (0,1)^{N_{\\mathrm{cls}} \\times N_{\\mathrm{pixel}}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " and another auxiliary image segmentation head " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " implemented with the simple FCN head [34]. The difference lies in that the camera SFAM is designed with all the " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{pixel}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " pixels of image feature maps across " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " local cameras, where " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{pixel}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cam}} \\times H \\times W" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{img}}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " first predicts the intermediate segmentation " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{img}}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{cam}} \\times N_{\\mathrm{cls}} \\times H \\times W}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{img}}(\\mathbf{X})" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{img}}^{\\prime}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " are then rearranged into " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{cam}}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{cls}} \\times N_{\\mathrm{pixel}}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\prime} \\in \\mathbb{R}^{N_{\\mathrm{pixel}} \\times C_{\\mathrm{img}}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " by transposing and reshaping operations. The final camera semantic embeddings " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " in Eq. 7 is the matrix product of " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\prime}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ", where the normalized " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the result of applying spatial softmax on " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{cam}}^{\\prime}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " as Eq. 6." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 148, + 545, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 148, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 348, + 148, + 545, + 163 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\mathrm {i m g}} ^ {\\prime} = \\mathcal {H} _ {\\mathrm {i m g}} (\\mathbf {X}), \\tag {5}", + "image_path": "c2a103d2804696b77c9cf9660589939b9ba583a80763dca0d549d49e6dfb0048.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 165, + 545, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 165, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 348, + 165, + 545, + 180 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\mathrm {c a m}} = \\left[ \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {\\mathrm {c a m}} ^ {\\prime} [ j,: ]\\right) \\right] _ {j = 1} ^ {j = N _ {\\mathrm {c l s}}}, \\tag {6}", + "image_path": "080ea0df7be288f1f77f6adf951eca15214978a8bb0e96a3f739c4098a1026c1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 350, + 182, + 545, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 182, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 350, + 182, + 545, + 196 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {\\mathrm {c a m}} = \\mathbf {D} _ {\\mathrm {c a m}} \\mathbf {X} ^ {\\prime}. \\tag {7}", + "image_path": "cfe15506f9b8dab3c1f715695f56d1bea9d8c72f6a2ecd752cf2b53033d23fce.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 201, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 259 + ], + "type": "text", + "content": "Note that the image segmentation annotations are unavailable for guiding the " + }, + { + "bbox": [ + 304, + 201, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{img}}^{\\prime}" + }, + { + "bbox": [ + 304, + 201, + 545, + 259 + ], + "type": "text", + "content": " in such 3D semantic segmentation datasets [3, 37]. Nevertheless, we also address this by the cross-modal semantic supervision scheme, which is detailed in Eq. 16 and Eq. 17." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": "Semantic-based Feature Fusion Module (SFFM). As shown in the right dashed box in Fig. 2, given the input of point-wise geometry-based fused features " + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{gfused}}" + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": ", category-wise semantic embeddings " + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": ", the SFFM first projects them into a " + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{sfused}}" + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": "-dimensional space as " + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{proj}} = Proj_1(\\mathbf{F}_{\\mathrm{gfused}})" + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = Proj_2(\\mathbf{E}_{\\mathrm{lidar}}) \\odot Proj_3(\\mathbf{E}_{\\mathrm{cam}})" + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": ", respectively. Before stacking " + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": " gray blocks in SFFM, we concatenate the projected multi-modal semantic embeddings together as " + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{2N_{\\mathrm{cls}} \\times C_{\\mathrm{sfused}}}" + }, + { + "bbox": [ + 304, + 260, + 546, + 392 + ], + "type": "text", + "content": ", which can be regarded as a dictionary that describes the typical characteristics for each category from the perspectives of LiDAR and cameras." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "text", + "content": "In each gray block, we model the semantic relations among all the " + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "inline_equation", + "content": "2N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "text", + "content": " category-wise semantic embeddings using the Multi-Head Self-Attention (MHSA) [41] as Eq. 8, where Norm indicates the LayerNorm operation. Let " + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{shsa}}" + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{sfused}} / N_H" + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "text", + "content": ", the MHSA can be decomposed into " + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{H}}" + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "text", + "content": " Single-Head Self-Attention (SHSA) operations with " + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_h \\in \\mathbb{R}^{2N_{\\mathrm{cls}} \\times C_{\\mathrm{shsa}}}" + }, + { + "bbox": [ + 304, + 392, + 545, + 475 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 480, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 480, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 366, + 480, + 545, + 493 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {\\mathrm {m h s a}} = \\operatorname {N o r m} (\\mathbf {E} + M H S A (\\mathbf {E})), \\tag {8}", + "image_path": "71acad9d5d9d9fa5702849fc8b6604e0a30d79c6e1432b195835db770a798ffa.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 340, + 495, + 545, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 495, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 340, + 495, + 545, + 510 + ], + "type": "interline_equation", + "content": "M H S A (\\mathbf {E}) = \\left[ S H S A \\left(\\mathbf {E} _ {h}\\right) \\right] _ {h = 1} ^ {h = N _ {\\mathrm {H}}}, \\tag {9}", + "image_path": "858d599ccc1e2985d1677d85c08b0a0032fc1d6f20f5241314bcb91d77a62f9d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 339, + 513, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 513, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 339, + 513, + 545, + 540 + ], + "type": "interline_equation", + "content": "S H S A \\left(\\mathbf {E} _ {h}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {h} \\mathbf {K} _ {h} ^ {T}}{\\sqrt {C _ {\\mathrm {s h s a}}}}\\right) \\mathbf {V} _ {h}. \\tag {10}", + "image_path": "b8faba12a2ed410cee53a34a4201db10df63918fe2e81619b24fd2c356d9c65e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "text", + "content": " includes " + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "text", + "content": ", the attention matrix " + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "inline_equation", + "content": "Softmax\\left(\\frac{\\mathbf{Q}_h\\mathbf{K}_h^T}{\\sqrt{C_{\\mathrm{shsa}}}}\\right)\\in \\mathbb{R}^{2N_{\\mathrm{cls}}\\times 2N_{\\mathrm{cls}}}" + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "text", + "content": " intuitively includes four parts of relation modeling on " + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{lidar}}\\rightarrow \\mathbf{E}_{\\mathrm{lidar}},\\mathbf{E}_{\\mathrm{lidar}}\\rightarrow \\mathbf{E}_{\\mathrm{cam}}," + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{cam}}\\rightarrow \\mathbf{E}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 543, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{cam}}\\rightarrow \\mathbf{E}_{\\mathrm{cam}}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "text", + "content": "After updating " + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{mhsa}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "text", + "content": ", we perform the further fusion between the point-wise projected features " + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "text", + "content": " and the semantic embeddings " + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{mhsa}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "text", + "content": ", using the Multi-Head Cross-Attention (MHCA) for semantic relation modeling in Eq 12 and the Feed-Forward Network (FFN) for feature embedding in Eq. 11, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 670, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {s f u s e d}} = \\operatorname {N o r m} \\left(\\mathbf {F} _ {\\text {m h c a}} + F F N \\left(\\mathbf {F} _ {\\text {m h c a}}\\right)\\right), \\tag {11}", + "image_path": "681a05202b8b2433abe091af11636b89aa12f6696a9f881efe12b9b2f7f1c3ea.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 685, + 545, + 709 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 685, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 315, + 685, + 545, + 709 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {m h c a}} = \\operatorname {N o r m} \\left(\\mathbf {F} _ {\\mathrm {p r o j}} + M H C A \\left(\\mathbf {F} _ {\\mathrm {p r o j}}, \\mathbf {E} _ {\\mathrm {m h s a}}, \\mathbf {E} _ {\\mathrm {m h s a}}\\right)\\right). \\tag {12}", + "image_path": "002cde9a7512426c09b147c136e9c88a7cd80be0f0a0a87c618a1ef7059bacf9.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21697" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "Unlike Eq. 8, the " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "MHCA" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": " in Eq. 12 sets the query " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": " from point-wise feature " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": ", and sets the key " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": " from semantic embeddings " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{mhsa}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": ". Thus, an attention matrix with shape " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{point}} \\times 2N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": " is computed to attentively aggregate the more important semantic embeddings respective to individual points, beyond the paired geometric association in GF-Phase. Following the common usages of multi-head attention [4], " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": " (i.e., 6) gray blocks are stacked." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": "Discussion on SFFM. As a useful and general technique [4,28,47], multi-head attention [41], although not proposed in this paper, is effectively tailored to multi-modal feature fusion with different motivations and designs by us. i) In Eq. 12, the " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "MHCA" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " enables the point-wise feature to attend to multi-modal semantic embeddings, so that both points inside and outside can be consistently supported by expressive multi-modal semantic embeddings. ii) The " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "MHCA" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " attention matrix computes the relative importance of two modalities to each point, improving the unbiased considerations of modalities in GF-Phase. iii) The " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "MHSA" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " in Eq. 8 explicitly models the category-wise intra-modal and inter-modal semantic relations, deriving the commonality learning. iv) The LiDAR SFAM and Camera SFAM aggregate the long sequences of voxel features " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " and image features " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " into the short sequences of " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "E_{\\text{lidar}}" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "E_{\\text{cam}}" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "N_{\\text{voxel}} / N_{\\text{cls}}" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "inline_equation", + "content": "N_{\\text{pixel}} / N_{\\text{cls}}" + }, + { + "bbox": [ + 46, + 168, + 288, + 384 + ], + "type": "text", + "content": " times, which enables efficient computation of the multi-head attention." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 393, + 226, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 226, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 226, + 406 + ], + "type": "text", + "content": "3.4. Cross-modal Feature Completion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": "As shown in Fig. 2, a cross-modal feature completion module with pixel-to-point loss " + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{pixel2point}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": " (Eq. 13) is set, where the point-wise pseudo-camera feature " + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{pcam}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": " is mapped from the point-wise LiDAR features " + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": " by another MLP-based " + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{pcam}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{pcam}}(\\mathbf{F}_{\\mathrm{lidar}})" + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": ". In practice, we compute the mean square error loss " + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mse}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": " between the " + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{pcam}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 508 + ], + "type": "text", + "content": " of the points inside for learning from the correctly paired cross-modal features relationship." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 519, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 519, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 64, + 519, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {\\text {p i x e l 2 p o i n t}} = \\mathcal {L} _ {\\mathrm {m s e}} \\left(\\mathbf {B F} _ {\\mathrm {p c a m}}, \\mathbf {B D e t a c h} \\left(\\mathbf {F} _ {\\mathrm {c a m}}\\right)\\right). \\tag {13}", + "image_path": "acf5b348f9f20276fd36c066a256e1c73703be6e99d7842d5c0ddf8c72951a5a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 534, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 534, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 69, + 534, + 287, + 548 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {c a m}} [ i,: ] = \\mathbf {F} _ {\\mathrm {p c a m}} [ i,: ] | _ {i \\in \\{j \\mid \\mathbf {B} [ j ] = 0 \\}}. \\tag {14}", + "image_path": "8667dffa83d56738592af1870053e5d4f122cd5c20fbdfd1633a1314be45c7d7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": "Note that the points outside are ignored by the binary mask " + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": ", which is mentioned in GF-Phase. The gradients of " + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": " are also detached for optimizing the learning of " + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{pcam}}" + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "We employ such a cross-modal feature completion module due to two motivations: i) Optimizing " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{pixel2point}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " forces LiDAR features " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{lidar}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " to imitate camera features " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{pcam}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", then the learned pseudo-image features " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{pcam}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " can be switched to replace the padded zeros in " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " as Eq. 14 in the inference stage, which serves as the cross-modal feature completion to enhance the feature learning. Thus, we further reduce feature gaps between points outside and inside. ii) " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{pcam}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " transfers rich appearance priors from the camera branch to the LiDAR branch with an effective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "consistency constraint for enhancing the intra-modal feature learning in the training stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 103, + 492, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 103, + 492, + 116 + ], + "spans": [ + { + "bbox": [ + 306, + 103, + 492, + 116 + ], + "type": "text", + "content": "3.5. Cross-modal Semantic Supervision" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": "Point Supervision. Let " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{point}}" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": " elements be the point-wise 3D semantic segmentation labels. The value of " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}[i]" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": " is in " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "[0, N_{\\mathrm{cls}} - 1]" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": ", where 0 denotes the ignored category. An MLP based point segmentation head " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{point}}" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": " is built on " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{sfused}}" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": " for the 3D segmentation prediction " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}} = \\mathcal{H}_{\\mathrm{point}}(\\mathbf{F}_{\\mathrm{sfused}})" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": ". Following [6, 21, 66], we adopt point loss " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{point}}" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": " as a combination of cross-entropy loss " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": " and lovasz-softmax loss " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lovasz}}" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": " [2] as " + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}(\\hat{\\mathbf{Y}}, \\mathbf{Y}) + \\mathcal{L}_{\\mathrm{lovasz}}(\\hat{\\mathbf{Y}}, \\mathbf{Y})" + }, + { + "bbox": [ + 304, + 121, + 545, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 220, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 245 + ], + "type": "text", + "content": "Point-to-voxel Supervision. For guiding the " + }, + { + "bbox": [ + 305, + 220, + 545, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{lidar}}^{\\prime}" + }, + { + "bbox": [ + 305, + 220, + 545, + 245 + ], + "type": "text", + "content": " in Eq. 3, the voxel loss " + }, + { + "bbox": [ + 305, + 220, + 545, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{p2v}}" + }, + { + "bbox": [ + 305, + 220, + 545, + 245 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 252, + 545, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 252, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 319, + 252, + 545, + 266 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {\\mathrm {p} 2 \\mathrm {v}} = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime}, \\mathbf {Y} _ {\\mathrm {p} 2 \\mathrm {v}}\\right) + \\mathcal {L} _ {\\text {l o v a s z}} \\left(\\mathbf {D} _ {\\text {l i d a r}} ^ {\\prime}, \\mathbf {Y} _ {\\mathrm {p} 2 \\mathrm {v}}\\right), \\tag {15}", + "image_path": "ebbf8abbb54d3b07439d02712239e7d8fa3fc7ef1a0e53765f7237651812d95b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "type": "text", + "content": "where the voxel labels " + }, + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathrm{p2v}}" + }, + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{voxel}}" + }, + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "type": "text", + "content": " elements can be determined from the labels of points in the voxel. To avoid ambiguity in " + }, + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{p2v}}" + }, + { + "bbox": [ + 304, + 274, + 545, + 322 + ], + "type": "text", + "content": ", the label of the voxel containing points of multiple categories is set to zero as the ignored category." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": "Point-to-pixel Supervision. The problem of how to use only point labels to correctly guide " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{img}}^{\\prime}" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " (Eq. 5) is still unsolved. For " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{img}}^{\\prime}" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " of each training sample, we initialize a " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cam}} \\times H \\times W" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " map " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathrm{point2pixel}}" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " of all zeros as the image segmentation label. We then project each point onto the down-sampled image planes and retrieve the point label " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}[i]" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " for the corresponding nearest " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "(\\langle \\cdot \\rangle)" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " pixel as Eq. 16. Although the generated image label " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathrm{point2pixel}}" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " is sparse, it provides sufficient supervision. In SM, Fig. S2 and Fig. S3 provide visualizations of " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{img}}^{\\prime}" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " to show that these sparse point-to-pixel supervisions can propagate to other regions with similar appearance, which guides " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{img}}^{\\prime}" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " correctly. Since " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathrm{point2pixel}}" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " is sparse, we only compute the basic cross-entropy loss " + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 304, + 322, + 546, + 490 + ], + "type": "text", + "content": " in Eq. 17." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 496, + 545, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 496, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 341, + 496, + 545, + 521 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {\\text {p o i n t 2 p i x e l}} \\left[ c _ {i}, \\left\\langle \\frac {H}{H _ {\\mathrm {i n}}} u _ {i} \\right\\rangle , \\left\\langle \\frac {W}{W _ {\\mathrm {i n}}} u _ {i} \\right\\rangle \\right] = \\mathbf {Y} [ i ], \\tag {16}", + "image_path": "6667a3a4d43db68571468ddc62db3187bc89f60384966a70266075b46cec496a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 523, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 523, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 342, + 523, + 545, + 537 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {\\text {p o i n t 2 p i x e l}} = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {D} _ {\\mathrm {i m g}} ^ {\\prime}, \\mathbf {Y} _ {\\text {p o i n t 2 p i x e l}}\\right). \\tag {17}", + "image_path": "0f1af408f0d0c7391dffdd000157800882ae7f3cb35831cc2c06ebbee94c189c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": "Total Loss Function. To jointly optimize all the modules in the proposed segmentation model, the total loss combines loss terms across different modalities into " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\alpha_{1}\\mathbf{L}_{\\mathrm{point}} + \\alpha_{2}\\mathbf{L}_{\\mathrm{p2v}} + \\alpha_{3}\\mathbf{L}_{\\mathrm{point2pixel}} + \\alpha_{4}\\mathbf{L}_{\\mathrm{pixel2point}}" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\alpha_{1} \\sim \\alpha_{4}" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " are set as 1.0, 1.0, 0.5, 1.0 to balance loss terms." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 611, + 542, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 542, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 542, + 624 + ], + "type": "text", + "content": "3.6. Asymmetric Multi-modal Data Augmentation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "The segmentation output and our multi-modal fusion mechanism are point-centered. In Eq. 1 and Eq. 16, the point and the pixel always can be bridged by the coordinate pair of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i},z_{i})" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "(c_i,u_i,v_i)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", as long as we pre-compute the coordinate pairs and keep the order of coordinate pairs among all the points in synchronization. With this ordering constraint, we have the flexibility" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21698" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 116, + 288, + 235 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "type": "text", + "content": "Table 1. Data augmentation transformations on LiDAR (L) point cloud and camera (C) images. We only treat random flipping as a naive symmetric transformation that can be simply applied to both modalities simultaneously." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 116, + 288, + 235 + ], + "lines": [ + { + "bbox": [ + 47, + 116, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 288, + 235 + ], + "type": "table", + "html": "
L-only·Global rotation around the Z axis with a random angle in [−π/4, +π/4]. \n·Global translation with a random vector (Δx, Δy, Δz) sampled from a Gaussian distribution with mean zero and the standard deviation 0.5. \n·Global scaling with a random scaling factor in [0.95, 1.05]. \n…
Symmetric·Random flipping along the X, Y axis with probability 0.5.
C-only·Scaling with a random ratio in [1.0, 1.5]. \n·Horizontal rotation with a random angle in [−1°, 1°]. \n·Random cropping with a size (Hin, Win). \n·Color jitter from torchvision [30] with random parameters of 0.3 for brightness, 0.3 for contrast, 0.3 for saturation, 0.1 for hue. \n·JPEG with a random compression ratio in [30, 70] and probability 0.5 [64]. \n…
", + "image_path": "340afb42fa332033cc19eba41c2a05b36f1401801c965b99d52e39233a930221.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 351 + ], + "type": "text", + "content": "to decouple multi-modal data augmentation: i) We can preserve all the LiDAR transformations in Tab. 1 with the images and the coordinates " + }, + { + "bbox": [ + 46, + 243, + 287, + 351 + ], + "type": "inline_equation", + "content": "(c, u, v)" + }, + { + "bbox": [ + 46, + 243, + 287, + 351 + ], + "type": "text", + "content": " unchanged. ii) Once coordinates " + }, + { + "bbox": [ + 46, + 243, + 287, + 351 + ], + "type": "inline_equation", + "content": "(c, u, v)" + }, + { + "bbox": [ + 46, + 243, + 287, + 351 + ], + "type": "text", + "content": " are synchronously transformed with images, we can further apply the camera-only transformations in Tab. 1 to images, which are asymmetric to the LiDAR transformations. iii) We can apply independent transformations between the images of local cameras since the image view is naturally sliced by local cameras." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 351, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 424 + ], + "type": "text", + "content": "The asymmetry lies in not only the LiDAR and camera world but also the local cameras. Thus, we significantly diversify the multi-modal data augmentation for 3D semantic segmentation using the introduced camera transformations with the preserved LiDAR transformations. More potential transformations can be exploited beyond Tab. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 435, + 124, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 124, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 124, + 449 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 455, + 164, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 164, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 164, + 468 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": "nuScenes Dataset splits 28,130 training samples, 6,019 validation samples, and 6,008 testing samples [3, 11]. Each nuScenes sample contains relatively sparser point cloud of 32 beams and RGB images captured by 6 cameras: front, front-left, front-right, back, back-left, and back-right. Points outside are projected below the image bottom due to different vertical FOVs of LiDAR and cameras. Following official protocol, the " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": " is 17. Semantic segmentation annotations are only on point clouds and not on images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": "Waymo Dataset for 3D semantic segmentation includes 23,691 training samples, 5,976 validation samples, and 2,982 testing samples [37]. Each sample contains the point cloud of 64 beams and the RGB images captured by 5 cameras: front, front-left, front-right, side-left, and side-right. There is no rear camera on Waymo ego-vehicle, hence more points outside unfairly increase the difficulty of multimodal segmentation. The " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " is 23. Similarly, semantic segmentation annotations are only on point clouds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "SemanticKITTI Dataset is collected by a LiDAR with 64 beams [1]. Following [10,50,66,67], we use sequences" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "00 to 10 (excluding 08) with 19,130 training samples and sequence 08 with 4,071 validation samples. It provides only the images of the front-view camera. The " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " is 20." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "text", + "content": "Evaluation Metric is the mean Intersection-over-Union (mIoU) defined as " + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\frac{1}{C}\\sum_{c = 1}^{N_{\\mathrm{cls}} - 1}\\frac{TP_c}{TP_c + FP_c + FN_c}" + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "inline_equation", + "content": "TP_{c}, FP_{c}, FN_{c}" + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "text", + "content": " denote the number of true positive, false positive, and false negative predictions for the " + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "text", + "content": "-th category out of " + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cls}} - 1" + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "text", + "content": " valid categories. Besides, nuScenes computes another metric fwIoU by weighting category-wise IoU with the point frequency of its category." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 193, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 289 + ], + "type": "text", + "content": "Settings. Although lightweight LiDRA-only baseline models can be trained with larger batchsize and more epochs to obtain relatively better results, we still train all the models under the same schedule: a batch of 32 random samples distributed on 16 Tesla V100 GPUs with 24 epochs. More implementation details on point cloud voxelization, network architecture, model training, and inference are all included in SM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 298, + 529, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 529, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 529, + 312 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 317, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 472 + ], + "type": "text", + "content": "Results on nuScenes. From Tab. 2, multi-modal methods counterintuitively lag far behind LiDAR-only methods. In the top-20 submissions on nuScenes benchmark [27], there are only four multi-modal methods. Details of Cylinder3D++, SPVCNN++, LIFusion, and CPFusion are not available till submission. Our framework achieves the best mIoU and fwIoU with multi-modal fusion mechanism and asymmetric multi-modal data augmentation. Especially, our method achieves superior performance on challenging small objects such as pedestrians and traffic cones, where the laser points are typically insufficient for LiDAR-only methods. Our method also significantly outperforms all public and private multi-modal submissions in Tab. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "type": "text", + "content": "Results on Waymo. We perform evaluation on Waymo benchmark [45] in Tab. 3, where the LiDAR sensor provides denser laser points and the multi-camera excludes the rearview. Such denser point cloud and incomplete camera FOV unfairly weaken the advantages of multiple modalities. Although, our method still achieves state-of-the-art performance using a general point cloud backbone network [9, 35]. We believe that the performance can be improved by incorporating stronger point cloud backbone networks from other state-of-the-art LiDAR-only methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Results on SemanticKITTI. While comparisons on the highly competitive nuScenes and Waymo datasets validate the effectiveness of our method, Tab. 4 shows the experimental comparison on SemanticKITTI [1]. Since SemanticKITTI provides only front-view images, PMF [67] evaluates the multi-modal segmentation performance on points inside the FOV intersection. We follow PMF to report " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}^1" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " on points inside in Tab. 4. Our method not only shows improvements to the LiDAR-only methods but also outperforms other multi-modal methods." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21699" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 95, + 483, + 236 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 2. Performance comparison on nuScenes testing set [27]. The modalities available on nuScenes include LiDAR (L), Camera (C), and Radar (R). Top-1 results are in bold. *Submission entries without a published paper by the CVPR 2023 deadline of Nov 11, 2022." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 108, + 95, + 483, + 236 + ], + "lines": [ + { + "bbox": [ + 108, + 95, + 483, + 236 + ], + "spans": [ + { + "bbox": [ + 108, + 95, + 483, + 236 + ], + "type": "table", + "html": "
MethodModalitymIoUfIoUbarrierbicyclebuscarcons. vehiclemotorcyclepedestriantraffic-conetrailertruckdriveableothersidewalkterrainmammadevegetation
PolarNet [62]L69.4287.3872.1616.8177.0186.5351.1469.6564.8054.1169.7063.5396.6467.1477.7072.1387.1384.47
JS3C-Net [52]L73.6088.0680.1426.1587.7984.5455.1772.5671.2866.2676.7971.1196.8064.4776.8674.0987.4886.10
Cylinder3D [66]L77.1689.9282.7629.7584.3489.4163.0379.2977.2173.4084.5569.1797.6670.2480.2975.5190.4187.55
AMVNet [24]L77.2790.0880.6431.9681.7388.9367.0784.3376.1173.4884.8767.3097.3767.3779.4175.4591.4588.69
SPVNAS [38]L77.3589.6880.0029.9891.9290.8164.6878.9975.6270.9481.0174.6497.4469.2379.9576.1089.2887.06
Cylinder3D++[66]L77.8689.9382.7633.8984.3489.4169.6379.4277.2673.4084.5569.4197.6670.2480.2975.5190.4287.55
AF2S3Net [7]L78.3488.5178.8752.2189.9384.1777.4274,3077.3271.9583.8873.7897.1366.4777.5174.0187.6986.80
SPVCNN++[38]L81.1290.9786.3543.1391.9092.1875.9075.7283.4477.3186.8277.3697.6971.2281.0877.1991.6788.98
LIFusion* [27]LC75.7489.3258.1336.3086.6784.2859.9679.6980.3077.7783.2368.7497.1868.1977.0474.4591.0388.95
PMF [67]LC77.0389.3482.1140.3380.9486.4263.7279.2279.7575.8681.1767.0597.2867.6978.0574.4889.9488.46
CPFusion* [27]LCR77.7289.1983.6737.0389.0286.2470.0877.4778.0774.5382.7867.9496.6468.2479.5374.9190.4786.95
2D3DNet [13]LC79.9690.0883.0159.3587.9985.0963.7084.3981.9575.9684.7971.9396.8867.3579.8175.9692.0589.18
MSeg3DLC81.1491.3583.1142.4694.9292.0167.1078.5885.6680.4787.5377.3297.7469.8281.2277.8392.3590.07
", + "image_path": "dff86aca43bddd43c7e6409d3043117a66ce475a905bd63cd00d2290ab1d865e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 117, + 277, + 216, + 348 + ], + "blocks": [ + { + "bbox": [ + 46, + 243, + 286, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 286, + 274 + ], + "type": "text", + "content": "Table 3. Performance comparison on Waymo testing set [45]. *Submission entries without a published paper by the CVPR 2023 deadline of Nov 11, 2022." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 117, + 277, + 216, + 348 + ], + "lines": [ + { + "bbox": [ + 117, + 277, + 216, + 348 + ], + "spans": [ + { + "bbox": [ + 117, + 277, + 216, + 348 + ], + "type": "table", + "html": "
MethodmIoU
LiDARMultiNet* [55]71.13
HorizonSegExpert* [45]69.44
SPVCNN++ [38]67.70
PolarFuse* [45]67.28
SalsaNext [40]55.85
MSeg3D70.51
", + "image_path": "3c9263821200625948778f16311ddf1a46460f48f6deab9290e89cc154147bba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 111, + 387, + 222, + 469 + ], + "blocks": [ + { + "bbox": [ + 46, + 354, + 287, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 386 + ], + "type": "text", + "content": "Table 4. Performance comparison on SemanticKITTI [1] validation set following PMF [67]. The results of the other methods are from PMF paper." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 387, + 222, + 469 + ], + "lines": [ + { + "bbox": [ + 111, + 387, + 222, + 469 + ], + "spans": [ + { + "bbox": [ + 111, + 387, + 222, + 469 + ], + "type": "table", + "html": "
MethodModalitymIoU1
SalsaNext [10]L59.4
SPVNAS [38]L62.3
Cylinder3D [66]L64.9
PointPainting [42]LC54.5
RGBAL [25]LC56.2
PMF [67]LC63.9
L-BaselineL64.8
MSeg3DLC66.7
", + "image_path": "da7abc9bac544f86885da766b41c964825017da2c32e35855c734f13c1d6ff9a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 79, + 519, + 252, + 571 + ], + "blocks": [ + { + "bbox": [ + 46, + 474, + 286, + 517 + ], + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 517 + ], + "type": "text", + "content": "Table 5. Analysis on the performance gap on all points and points inside, and the data augmentation (DA). Only GF-Phase is simply used for multi-modal fusion (M-Fusion) here. DA includes LiDAR DA (L-DA) and Multi-modal DA (M-DA)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 519, + 252, + 571 + ], + "lines": [ + { + "bbox": [ + 79, + 519, + 252, + 571 + ], + "spans": [ + { + "bbox": [ + 79, + 519, + 252, + 571 + ], + "type": "table", + "html": "
LiDARM-FusionDAnuScenesWaymo
mIoUmIoU1mIoUmIoU1
✓×L-DA72.0070.7667.4867.41
✓GF-Phase×68.1076.6859.7764.79
✓GF-PhaseL-DA71.3577.4860.3465.70
✓GF-PhaseM-DA72.3978.6563.9767.94
", + "image_path": "fb8ba261cd3e087baa4aa82d7506cbacf5e02ab0164bf609537ab8926f18dee6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 47, + 633, + 287, + 693 + ], + "blocks": [ + { + "bbox": [ + 46, + 576, + 286, + 630 + ], + "lines": [ + { + "bbox": [ + 46, + 576, + 286, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 286, + 630 + ], + "type": "text", + "content": "Table 6. Analysis on multi-modal feature fusion of GF-Phase, cross-modal (CM) feature completion, and SF-Phase. The supervised CM feature completion is decomposed as " + }, + { + "bbox": [ + 46, + 576, + 286, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{pixel2point}}" + }, + { + "bbox": [ + 46, + 576, + 286, + 630 + ], + "type": "text", + "content": " (Eq. 13) and \"Comp.\" (Eq. 14). The gap can be formulated as mIoU - mIoU1." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 633, + 287, + 693 + ], + "lines": [ + { + "bbox": [ + 47, + 633, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 287, + 693 + ], + "type": "table", + "html": "
LiDARM-FusionCM Feature CompletionnuScenesWaymo
GF-PhaseSF-PhaseLpixel2pointComp.mIoUmIoU1GapmIoUmIoU1Gap
✓××××72.0070.76+1.2467.4867.41+0.07
✓✓×××72.3978.65-6.2663.9767.94-3.97
✓✓×✓×76.4479.10-2.6667.1369.06-1.93
✓✓×✓✓78.2879.10-0.8267.8969.06-1.17
✓✓✓✓✓80.0080.10-0.1069.6370.19-0.56
", + "image_path": "211504e3b86939ecae8625bbc207ebf798029391961821c543764b84557fcab9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 243, + 399, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 399, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 399, + 256 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 484 + ], + "type": "text", + "content": "Difficulties Arising from Multi-modality are investigated in Tab. 5 from two perspectives of the performance gap between all points and points inside, as well as the multi-modal data augmentation. The experiments in Tab. 5 start with a vanilla variant of the multi-modal model using the basic GF-Phase fusion without the multi-modal data augmentation. The mIoU is relatively worse than the mIoU1, since the geometry-based feature fusion ideally ignores the inevitable points outside with the missing camera features [20, 25, 67], which also motivates us to close the performance gaps between mIoU and mIoU1 by using the cross-modal feature completion and semantic-based SF-Phase. Besides, in rows 2 and 1, the multi-modal segmentation performance will be worsened if we deprecated the augmentation transformations that cannot be directly applied to both modalities. However, the proposed asymmetric multi-modal data augmentation allows the accumulation of the transformations on point cloud and image, achieving the best performance on both datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 652 + ], + "type": "text", + "content": "Multi-modal Feature Fusion Modules. Tab. 6 shows that our supervised cross-modal feature completion and SF-Phase benefit both mIoU and mIoU1 with gradually narrowed gaps. In row 3, Lpixel2point can also bring improvements, which indicates that transferring appearance information from dense image to sparse point cloud is applicable and beneficial for joint feature learning. In row 4, completing camera features with pseudo-camera features facilitates the feature gap reduction between points inside and outside. Eventually, the SF-Phase further narrows the gaps between mIoU and mIoU1 to the smallest values of 0.10 and 0.56, addressing the limitation of geometric associations by semantic-based feature fusion. Thus, we set the model in the last row as our final framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "mIoU Breakdown Over Distance. Fig. 3 presents the distance-based evaluation corresponding to the models in Tab. 6. The LiDAR-only model degrades at long distances due to more sparse points. Instead, the multimodal model in row 5 effectively alleviates the performance" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21700" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 168, + 187 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 168, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 168, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 168, + 187 + ], + "type": "image", + "image_path": "681d2b87da08b5ef66710f1cc90178665af009dd34df04811f476cfdf7247b34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 169, + 70, + 278, + 187 + ], + "blocks": [ + { + "bbox": [ + 169, + 70, + 278, + 187 + ], + "lines": [ + { + "bbox": [ + 169, + 70, + 278, + 187 + ], + "spans": [ + { + "bbox": [ + 169, + 70, + 278, + 187 + ], + "type": "image", + "image_path": "e45d608929ad29f60ff5ed2080c7d5cee422695e4c87498a8db33786491219ce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 188, + 168, + 319 + ], + "blocks": [ + { + "bbox": [ + 53, + 188, + 168, + 319 + ], + "lines": [ + { + "bbox": [ + 53, + 188, + 168, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 188, + 168, + 319 + ], + "type": "image", + "image_path": "eab315ed50cfff5c2985d01f4bbdedf11e962b7c799ca9d0df502695d56cd55e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 320, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 47, + 320, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 287, + 342 + ], + "type": "text", + "content": "Figure 3. Distance-based mIoU evaluation on nuScenes and Waymo validation (Val.) sets using the models in Tab. 6." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 188, + 278, + 319 + ], + "blocks": [ + { + "bbox": [ + 169, + 188, + 278, + 319 + ], + "lines": [ + { + "bbox": [ + 169, + 188, + 278, + 319 + ], + "spans": [ + { + "bbox": [ + 169, + 188, + 278, + 319 + ], + "type": "image", + "image_path": "4bf6a4846b6c0732af2849c732a60edeec6439c2510c4702c2bc2f20b39bdc77.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 70, + 379, + 263, + 415 + ], + "blocks": [ + { + "bbox": [ + 47, + 347, + 287, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 347, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 287, + 378 + ], + "type": "text", + "content": "Table 7. Robustness analysis on MSeg3D by removing some cameras as malfunction. \"#Camera\" denotes the number of available cameras. \"×\" denotes LiDAR-only baseline." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 379, + 263, + 415 + ], + "lines": [ + { + "bbox": [ + 70, + 379, + 263, + 415 + ], + "spans": [ + { + "bbox": [ + 70, + 379, + 263, + 415 + ], + "type": "table", + "html": "
#Camera6543210×
nuScenes80.0078.6977.6276.6976.0175.4474.4772.00
Waymo-69.6369.1868.8668.1368.0968.0467.48
", + "image_path": "9eab1f212612c2214d63de2e32e49fa4c733deb5bcdde9c557984d222aaa5eba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 70, + 442, + 263, + 502 + ], + "blocks": [ + { + "bbox": [ + 47, + 421, + 287, + 442 + ], + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 442 + ], + "type": "text", + "content": "Table 8. Robustness analysis on MSeg3D with multi-frame point clouds input. \"#L-Frame\" is the frame number for LiDAR." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 442, + 263, + 502 + ], + "lines": [ + { + "bbox": [ + 70, + 442, + 263, + 502 + ], + "spans": [ + { + "bbox": [ + 70, + 442, + 263, + 502 + ], + "type": "table", + "html": "
nuScenes#L-Frame1102025304025
Camera××××××√
mIoU72.0074.6675.3775.7775.2875.1581.12
Waymo#L-Frame151015203010
Camera××××××√
mIoU67.4869.1869.4568.8868.7868.6470.20
", + "image_path": "df11a89135ee8eab8b318d0fb41ba25a937280986fb686c1c15e4ab7528ee93e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 510, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 664 + ], + "type": "text", + "content": "degradation. In Fig. 3 (a) and (b), improvements between the multi-modal model in row 5 and the LiDAR-only model in row 1 are also increased along the distance due to increased sparsity. The reasons why the improvement of the multi-modal model on Waymo is not as similar as that on nuScenes are further analyzed in Fig. 3 (c) and (d). The Waymo point cloud has denser points, which reduces the dependence on the image. More points outside are distributed along different distances due to no rear cameras on Waymo. Although more points outside hinder the applicability of multi-modal fusion, our final multi-modal model still yields a low-performance gap of " + }, + { + "bbox": [ + 46, + 510, + 287, + 664 + ], + "type": "inline_equation", + "content": "0.56\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 510, + 287, + 664 + ], + "type": "text", + "content": " on Waymo in row 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Robustness Against Camera Malfunction. In Tab. 7, our MSeg3D performs properly under the unfavorable condition of camera malfunction, and it can easily deal with the practical situation without switching the model." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 352, + 93, + 499, + 160 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "type": "text", + "content": "Table 9. Scalability analysis by barely varying the backbone settings on nuScenes validation set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 93, + 499, + 160 + ], + "lines": [ + { + "bbox": [ + 352, + 93, + 499, + 160 + ], + "spans": [ + { + "bbox": [ + 352, + 93, + 499, + 160 + ], + "type": "table", + "html": "
Image BackbonemIoU#Params(M)Latency(s)
×72.0021.280.083
SegFormer-B0 [47]78.1425.330.204
SegFormer-B5 [47]78.89103.480.479
HRNet-w18 [44]79.2131.560.265
ResNet101 [15]79.3664.600.567
HRNet-w48 [44]80.0087.340.445
", + "image_path": "3c365163f4d9d3a7511759012ca2cbaec224cadf72f381c3ebc9d839e29eca46.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 169, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 217 + ], + "type": "text", + "content": "Note that even with all cameras removed, MSeg3D still outperforms the LiDAR-only baseline, where our cross-modal feature completion supervision provides effective cross-modal information transfer in training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 218, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 350 + ], + "type": "text", + "content": "Robustness Against Multi-frame Point Clouds Input. Since the densified points alleviate the sparsity, collapsing laser points from neighboring frames usually boosts LiDAR-only 3D perception [16, 52]. In Tab. 8, we follow [57] collapsing multiple previous frames to the current frame from the provided ego-vehicle motion information. For LiDAR-only model, the improvements are saturated with 25 and 10 frames on nuScenes and Waymo. Under such conditions, our MSeg3D still achieves further improvements in the last column, which shows that multi-frame point clouds input can be an optional extension." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "content": "Complexity Scalability. From Tab. 9, a lightweight image backbone also delivers significant performance gains due to multi-modal input. The best-performing HRNet-w48 is our default image backbone. However, the multi-camera image input makes the efficiency bottleneck lie in the image backbone, which deserves in-depth study for autonomous driving [23]. We believe that real-time image segmentation networks such as BiSeNet [59,60] and Fast-SCNN [29] can accelerate our approach. Overall, our MSeg3D can adapt to complexity scalability among various backbones, which flexibly trades off performance and efficiency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 491, + 378, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 378, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 378, + 504 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 509, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 675 + ], + "type": "text", + "content": "We propose a novel multi-modal 3D semantic segmentation method termed MSeg3D for autonomous driving, based on LiDAR and multi-camera sensors. Our cross-modal feature completion and semantic-based feature fusion solve the overlooked problem, where multi-modal fusion can occur only in the sensor FOV intersection. The proposed asymmetric multi-modal data augmentation enables the multi-modal segmentation model to be effectively trained with reliable performance. Our method achieves state-of-the-art performance on nuScenes, Waymo, and SemanticKITTI. The comprehensive experiments validate improvements and robustness. We hope that our work can inspire further investigation into multi-modal fusion for 3D semantic segmentation in autonomous driving." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment: This work was supported by the National Key Research and Development Program of China (2018YFE0183900) and YUNJI Technology Co. Ltd." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21701" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jürgen Gall. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In ICCV, pages 9296-9306, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 287, + 190 + ], + "type": "text", + "content": "[2] Maxim Berman, Amal Rannen Triki, and Matthew B. Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, pages 4413-4421, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, pages 11618-11628, 2020. 1, 2, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 289 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ross B. Girshick, Piotr Dólar, Alexander C. Berg, and Alexander Kirillov. Boundary IoU: Improving object-centric image segmentation evaluation. In CVPR, pages 15334-15342, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "type": "text", + "content": "[6] Hui-Xian Cheng, Xian-Feng Han, and Guo-Qiang Xiao. CENet: Toward concise and efficient LiDAR semantic segmentation for autonomous driving. In ICME, pages 1-6. IEEE, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "text", + "content": "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (AF)2-S3Net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "type": "text", + "content": "[8] Christopher B. Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 500 + ], + "type": "text", + "content": "[9] OpenPCDet Contributors. OpenPCDet: An open-source toolbox for 3D object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2023. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "text", + "content": "[10] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. SalsaNext: fast, uncertainty-aware semantic segmentation of LiDAR point clouds for autonomous driving. arXiv preprint arXiv:2003.03653, 2020. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "text", + "content": "[11] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic NuScenes: A large-scale benchmark for LiDAR panoptic segmentation and tracking. IEEE Robotics Autom. Lett., 7(2):3795-3802, 2022. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[12] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, pages 3146-3154, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[13] Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Pantofaru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas A. Funkhouser. Learning 3D semantic segmentation with only 2D image supervision. In 3DV, pages 361-372. IEEE, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D semantic segmentation with submanifold sparse" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "convolutional networks. In CVPR, pages 9224-9232, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 132, + 545, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 132, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 308, + 132, + 545, + 164 + ], + "type": "text", + "content": "[16] Peiyun Hu, Jason Ziglar, David Held, and Deva Ramanan. What you see is what you get: Exploiting visibility for 3D object detection. In CVPR, pages 10998-11006, 2020. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 167, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 167, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 308, + 167, + 545, + 209 + ], + "type": "text", + "content": "[17] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. RandLA-Net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11105–11114, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "type": "text", + "content": "[18] Zhenchao Jin, Tao Gong, Dongdong Yu, Q. Chu, Jian Wang, Changhu Wang, and Jie Shao. Mining contextual information beyond image for semantic segmentation. In ICCV, pages 7211-7221, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 258, + 545, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 290 + ], + "type": "text", + "content": "[19] Zhenchao Jin, Bin Liu, Qi Chu, and Nenghai Yu. ISNet: Integrate image-level and semantic-level context for semantic segmentation. In ICCV, pages 7169-7178, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "type": "text", + "content": "[20] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. FuseSeg: LiDAR point cloud segmentation fusing multi-modal data. In WACV, pages 1863-1872, 2020. 1, 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 337, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 545, + 380 + ], + "type": "text", + "content": "[21] Jiale Li, Hang Dai, and Yong Ding. Self-distillation for robust LiDAR semantic segmentation in autonomous driving. In ECCV, volume 13688, pages 659-676, 2022. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[22] Jiale Li, Hang Dai, Ling Shao, and Yong Ding. From voxel to point: IoU-guided 3D object detection for point cloud with voxel-to-point decoder. In ACM MM, pages 4622-4631, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 429, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 429, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 308, + 429, + 545, + 472 + ], + "type": "text", + "content": "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 475, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 517 + ], + "type": "text", + "content": "[24] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. AMVNet: Assertion-based multi-view fusion network for LiDAR semantic segmentation. CoRR, abs/2012.04934, 2020. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 520, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 545, + 563 + ], + "type": "text", + "content": "[25] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Kumar Yogamani. RGB and LiDAR fusion based 3D semantic segmentation for autonomous driving. In ITSC, pages 7-12, 2019. 1, 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 566, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 566, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 545, + 598 + ], + "type": "text", + "content": "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. RangeNet++: Fast and accurate LiDAR semantic segmentation. In IROS, pages 4213-4220, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "text", + "content": "[27] nuScenes. nuScenes leaderboard of lidar segmentation task. https://www.nuscenes.org/lidar-segmentation?externalData " + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "text", + "content": " all&mapData " + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "text", + "content": " all&modalities " + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "text", + "content": " Any, 2023.1,6,7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "text", + "content": "[28] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3D object detection with pointformer. In CVPR, pages 7463-7472, 2021. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[29] Rudra P. K. Poudel, Stephan Liwicki, and Roberto Cipolla. Fast-SCNN: Fast semantic segmentation network. In BMVC, page 289, 2019. 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21702" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[30] Pytorch. Pytorch document of color jitter. https://torch.org/vision/stable/generated/torchvision.transforms.ColorJitter.html?highlight=colorjitter,2023.6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "text", + "content": "[32] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells III, and Alejandro F. Frangi, editors, MICCAI, volume 9351, pages 234–241, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[34] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. IEEE TPAMI, 39(4):640-651, 2017. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "text", + "content": "[35] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3D object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2021. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "text", + "content": "[36] Vishwanath A. Sindagi, Yin Zhou, and Oncel Tuzel. MVX-Net: Multimodal VoxelNet for 3D object detection. In ICRA, pages 7276–7282, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 465 + ], + "type": "text", + "content": "[37] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2443-2451, 2020. 1, 2, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "text", + "content": "[38] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3D architectures with sparse point-voxel convolution. In ECCV, volume 12373, pages 685–702, 2020. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "text", + "content": "[39] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPConv: Flexible and deformable convolution for point clouds. In ICCV, pages 6410–6419, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "text", + "content": "[40] Cortinhal Tiago, Tzelepis George, and Erdal Aksoy Eren. SalsaNext: Fast, uncertainty-aware semantic segmentation of LiDAR point clouds. In Advances in Visual Computing, pages 207-222, 2020. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[42] Sourabh Vora, Alex H. Lang, Bassam Helou, and Oscar Beijbom. PointPainting: Sequential fusion for 3D object detection. In CVPR, pages 4603-4611, 2020. 1, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[43] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. PointAugmenting: Cross-modal augmentation for 3D object detection. In CVPR, pages 11794–11803, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[44] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. IEEE TPAMI, 43(10):3349-3364, 2021. 1, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "text", + "content": "[45] Waymo. Waymo Open Dataset leaderboard of 3D semantic segmentation challenge. https://waymo.com/open/challenges/2022/3d-semantic-segmentation/, 2023.6,7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 176, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 231 + ], + "type": "text", + "content": "[46] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. SqueezeSegV2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a LiDAR point cloud. In ICRA, pages 4376-4382, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 233, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 233, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 233, + 545, + 277 + ], + "type": "text", + "content": "[47] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 1, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 279, + 545, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 279, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 308, + 279, + 545, + 333 + ], + "type": "text", + "content": "[48] Liang Xie, Chao Xiang, Zhengxu Yu, Guodong Xu, Zheng Yang, Deng Cai, and Xiaofei He. PI-RCNN: an efficient multi-sensor 3D object detector with point-based attentive cont-conv fusion module. In AAAI, pages 12460–12467, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 335, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 335, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 545, + 380 + ], + "type": "text", + "content": "[49] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In ECCV, pages 1-19, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "text", + "content": "[50] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVNet: A deep and efficient range-point-voxel fusion network for LiDAR point cloud segmentation. In ICCV, pages 16004-16013, 2021. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 428, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 545, + 472 + ], + "type": "text", + "content": "[51] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Bin Zhou, and Liangjun Zhang. FusionPainting: Multimodal fusion with adaptive attention for 3D object detection. In ITSC, pages 3047-3054, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 474, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 474, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 308, + 474, + 545, + 528 + ], + "type": "text", + "content": "[52] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 531, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 531, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 531, + 545, + 575 + ], + "type": "text", + "content": "[53] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. PointASNL: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, pages 5588-5597, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 577, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 609 + ], + "type": "text", + "content": "[54] Yan Yan, Yuxing Mao, and Bo Li. SECOND: sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 612, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 655 + ], + "type": "text", + "content": "[55] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. LidarMultiNet: Towards a unified multi-task network for LiDAR perception. CoRR, abs/2209.09385, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6):210:1-210:12, 2016. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21703" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 532 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[57] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3D object detection and tracking. In CVPR, pages 11784-11793, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 138 + ], + "type": "text", + "content": "[58] Tianwei Yin, Xingyi Zhou, and Philipp Krähenbuhl. Multimodal virtual point 3D detection. In NeurIPS, pages 16494-16507, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "text", + "content": "[59] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. BiSeNet V2: bilateral network with guided aggregation for real-time semantic segmentation. IJCV, 129(11):3051-3068, 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "text", + "content": "[60] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. BiSeNet: Bilateral segmentation network for real-time semantic segmentation. In ECCV, volume 11217, pages 334-349, 2018. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 286, + 262 + ], + "type": "text", + "content": "[61] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In ECCV, volume 12351, pages 173-190, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 286, + 318 + ], + "type": "text", + "content": "[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. PolarNet: An improved grid representation for online LiDAR point clouds semantic segmentation. In CVPR, pages 9598-9607, 2020. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 321, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 286, + 352 + ], + "type": "text", + "content": "[63] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 6230-6239, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 396 + ], + "type": "text", + "content": "[64] Stephan Zheng, Yang Song, Thomas Leung, and Ian J. Goodfellow. Improving the robustness of deep neural networks via stability training. In CVPR, pages 4480-4488, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "type": "text", + "content": "[65] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-PolarNet: Proposal-free LiDAR point cloud panoptic segmentation. In CVPR, pages 13194-13203, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 434, + 286, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 286, + 485 + ], + "type": "text", + "content": "[66] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation. In CVPR, pages 9939-9948, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 489, + 286, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 286, + 532 + ], + "type": "text", + "content": "[67] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multisensor fusion for 3D LiDAR semantic segmentation. In ICCV, pages 16260-16270, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21704" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_content_list.json b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dbb501e5578f6351b601489586e891305fc5dddc --- /dev/null +++ b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_content_list.json @@ -0,0 +1,1399 @@ +[ + { + "type": "text", + "text": "MV-JAR: Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training", + "text_level": 1, + "bbox": [ + 133, + 128, + 836, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Runsen Xu $^{1,2}$ Tai Wang $^{1,2}$ Wenwei Zhang $^{3,2}$ Runjian Chen $^{4}$ Jinkun Cao $^{5}$ Jiangmiao Pang $^{2\\boxtimes}$ Dahua Lin $^{1,2}$", + "bbox": [ + 168, + 200, + 799, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Chinese University of Hong Kong 2Shanghai AI Laboratory 3S-Lab, NTU 4The University of Hong Kong 5Carnegie Mellon University", + "bbox": [ + 156, + 239, + 812, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{runsenxu,wt019,dhlin}@ie.cuhk.edu.hk, wenwei001@ntu.edu.sg, rjchen@connect.hku.hk, jinkunc@andrew.cmu.edu, pangjiangmiao@gmail.com", + "bbox": [ + 119, + 276, + 844, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 359, + 312, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces the Masked Voxel Jigsaw and Reconstruction (MV-JAR) method for LiDAR-based self-supervised pre-training and a carefully designed data-efficient 3D object detection benchmark on the Waymo dataset. Inspired by the scene-voxel-point hierarchy in downstream 3D object detectors, we design masking and reconstruction strategies accounting for voxel distributions in the scene and local point distributions within the voxel. We employ a Reversed-Furthest-Voxel-Sampling strategy to address the uneven distribution of LiDAR points and propose MV-JAR, which combines two techniques for modeling the aforementioned distributions, resulting in superior performance. Our experiments reveal limitations in previous data-efficient experiments, which uniformly sample fine-tuning splits with varying data proportions from each LiDAR sequence, leading to similar data diversity across splits. To address this, we propose a new benchmark that samples scene sequences for diverse fine-tuning splits, ensuring adequate model convergence and providing a more accurate evaluation of pre-training methods. Experiments on our Waymo benchmark and the KITTI dataset demonstrate that MV-JAR consistently and significantly improves 3D detection performance across various data scales, achieving up to a $6.3\\%$ increase in mAPH compared to training from scratch. Codes and the benchmark are available at https://github.com/SmartBot-PJLab/MV-JAR.", + "bbox": [ + 75, + 391, + 472, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised pre-training has gained considerable attention, owing to its exceptional performance in visual representation learning. Recent advancements in contrastive", + "bbox": [ + 75, + 832, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc99e3383e48fa7a358cd24437bab93ddf8278ebaef1bf496a601e560e915666.jpg", + "image_caption": [ + "Figure 1. 3D object detection results on the Waymo dataset. Our MV-JAR pre-training accelerates model convergence and greatly improves the performance with limited fine-tuning data." + ], + "image_footnote": [], + "bbox": [ + 501, + 343, + 702, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27eed56131e7787e728f4124f47081fe7fb9186f4e67cde9a5c4e39ba0cb321c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 343, + 890, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learning [4, 6, 8, 16, 45] and masked autoencoders [1, 7, 15, 36, 47] for images have sparked interest among researchers and facilitated progress in modalities such as point clouds.", + "bbox": [ + 496, + 612, + 890, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, LiDAR point clouds differ from images and dense point clouds obtained by reconstruction as they are naturally sparse, unorganized, and irregularly distributed. Developing effective self-supervised proxy tasks for these unique properties remains an open challenge. Constructing matching pairs for contrastive learning in geometry-dominant scenes is more difficult [20, 40], as points or regions with similar geometry may be assigned as negative samples, leading to ambiguity during training. To address this, our study explores masked voxel modeling paradigms for effective LiDAR-based self-supervised pre-training.", + "bbox": [ + 496, + 657, + 892, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Downstream LiDAR-based 3D object detectors [12, 19, 33, 38, 41, 50] typically quantize the 3D space into voxels and encode point features within them. Unlike pixels, which are represented by RGB values, the 3D space presents a scene-voxel-point hierarchy, introducing new", + "bbox": [ + 496, + 825, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 99, + 886, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13445", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenges for masked modeling. Inspired by this, we design masking and reconstruction strategies that consider voxel distributions in the scene and local point distributions in the voxel. Our proposed method, Masked Voxel Jigsaw And Reconstruction (MV-JAR), harnesses the strengths of both voxel and point distributions to improve performance.", + "bbox": [ + 78, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To account for the uneven distribution of LiDAR points, we first employ a Reversed-Furthest-Voxel-Sampling (R-FVS) strategy that samples voxels to mask based on their sparseness. This approach prevents masking the furthest distributed voxels, thereby avoiding information loss in regions with sparse points. To model voxel distributions, we propose Masked Voxel Jigsaw (MVJ), which masks the voxel coordinates while preserving the local shape of each voxel, enabling scene reconstruction akin to solving a jigsaw puzzle. For modeling local point distributions, we introduce Masked Voxel Reconstruction (MVR), which masks all coordinates of points within the voxel but retains one point as a hint for reconstruction. Combining these two methods enhances masked voxel modeling.", + "bbox": [ + 78, + 181, + 467, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments indicate that existing data-efficient experiments [20, 40] inadequately evaluate the effectiveness of various pre-training methods. The current benchmarks, which uniformly sample frames from each data sequence to create diverse fine-tuning splits, exhibit similar data diversity due to the proximity of neighboring frames in a sequence [3, 14, 30]. Moreover, these experiments train models for the same number of epochs across different fine-tuning splits, potentially leading to incomplete convergence. As a result, the benefits of pre-trained representations become indistinguishable across splits once the object detector is sufficiently trained on the fine-tuning data. To address these shortcomings, we propose sampling scene sequences to form diverse fine-tuning splits and establish a new data-efficient 3D object detection benchmark on the Waymo [30] dataset, ensuring sufficient model convergence for a more accurate evaluation.", + "bbox": [ + 78, + 393, + 467, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We employ the Transformer-based SST [12] as our detector and pre-train its backbone for downstream detection tasks. Comprehensive experiments on the Waymo and KITTI [14] datasets demonstrate that our pre-training method significantly enhances the model's performance and convergence speed in downstream tasks. Notably, it improves detection performance by $6.3\\%$ mAPH when using only $5\\%$ of scenes for fine-tuning and reduces training time by half when utilizing the entire dataset (Fig. 1). With the representation pre-trained by MV-JAR, the 3D object detectors pre-trained on Waymo also exhibit generalizability when transferred to KITTI.", + "bbox": [ + 78, + 648, + 467, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 78, + 844, + 215, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning for point clouds. Annotating point clouds demands significant effort, necessitating self", + "bbox": [ + 78, + 869, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "supervised pre-training methods. Prior approaches primarily focus on object CAD models [21, 26, 29, 39, 42, 44] and indoor scenes [17, 35, 46]. Point-BERT [42] applies BERT-like paradigms for point cloud recognition, while Point-MAE [26] reconstructs point patches without the tokenizer. To acquire scene-level representation, PointContrast [35] introduces a contrastive learning approach that compares points in two static partial views of a reconstructed indoor scene. Its successor, SceneContrast [17], incorporates spatial information into the contrastive learning framework. However, LiDAR point clouds are irregular and dynamic, with mainstream LiDAR perception models often exhibiting distinct architectures, which obstructs the direct adaptation of object-level and indoor scene pre-training methods. STRL [18] employs contrastive learning with two temporally-correlated frames for spatiotemporal representation. GCC-3D [20] presents a framework that combines geometry-aware contrast and pseudo-instance clustering harmonization. ProposalContrast [40] targets region-level contrastive learning to improve 3D detection. CO3 [5] leverages infrastructure-vehicle-cooperation point clouds to construct effective contrastive views. In contrast, our work explores masked voxel modeling, diverging from the prevailing contrastive learning paradigm.", + "bbox": [ + 501, + 90, + 888, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked autoencoders for self-supervised pre-training. Masked language modeling plays a pivotal role in self-supervised pre-training for Transformer-based networks in natural language processing [2, 10, 28]. This approach typically masks portions of the input and pre-trains networks to predict the original information. With the successful integration of Transformers into computer vision [11], researchers have increasingly focused on masked image modeling [1, 7, 15, 36, 43, 47] to mitigate the data-intensive issue of ViT [11]. BEiT [1] employs a tokenizer to generate discrete tokens for image patches and utilizes a BERT-like framework to pre-train ViT. MAE [15] introduces an asymmetric autoencoder for reconstructing RGB pixels of original images, eliminating the need for an extra tokenizer. SimMiM [36] encodes masked raw images with Transformers and employs a lightweight prediction head for recovery. Demonstrating considerable potential in image recognition, masked autoencoders have been applied in video understanding [13, 32], medical image analysis [48], and audio processing [37]. In this work, we investigate the application of this powerful pre-training technique to LiDAR point clouds using voxel representation.", + "bbox": [ + 503, + 458, + 888, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer-based 3D object detection. The recent success of Vision Transformers [11] has inspired extensive research into the application of Transformer-based architectures for 3D object detection [12, 23-25, 31]. 3DETR [24] presents an end-to-end object detection framework with modified Transformer blocks for 3D point clouds. Pointformer [25] employs hierarchical Transformer blocks to", + "bbox": [ + 503, + 794, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13446", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2f89105dc3364d6ded51795534fdab51ff4bed01f320412b3f85aec31ece80bc.jpg", + "image_caption": [ + "Figure 2. Overview of Masked Voxel Jigsaw and Reconstruction (MV-JAR). Our R-FVS sampling method selects a specific ratio of nonempty voxels to mask. For position-masked voxels, we mask the absolute coordinates of all points within the voxel while maintaining the relative coordinates. For shape-masked voxels, we mask both absolute and relative coordinates of all points, preserving only one point. Features are extracted from all non-empty voxels, and masked features are decoded by a dedicated head to recover corresponding targets." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "extract point features for 3D detection. Voxel Transformer [23] develops a Transformer-based 3D backbone to establish long-range relationships between voxels. Recently, SST [12] introduced a single-stride sparse Transformer as a replacement for the PointPillars [19] backbone, abandoning the multi-resolution strategy to enhance the detection of small objects. Emulating the SwinTransformer [22], SST partitions the space into windows and calculates attention only within each window for efficiency. This approach demonstrates impressive detection results, and we adopt it for our experiments.", + "bbox": [ + 75, + 369, + 472, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 550, + 212, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present an overview of a general masked voxel modeling framework for LiDAR point clouds and our Reversed-Furthest-Voxel-Sampling strategy to tackle the uneven distribution of LiDAR points (Sec. 3.1). Masked Voxel Jigsaw (MVJ) (Sec. 3.2) and Masked Voxel Reconstruction (MVR) (Sec. 3.3) are both instantiations of the general framework that learns the voxel distributions and local point distributions, respectively. Masked Voxel Jigsaw and Reconstruction (MV-JAR) consist of these two tasks to synergically learn the data distributions at two levels (Sec. 3.4).", + "bbox": [ + 75, + 578, + 468, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Masked Voxel Modeling and Sampling", + "text_level": 1, + "bbox": [ + 76, + 755, + 410, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, a general masked voxel modeling framework converts LiDAR point clouds into voxels and samples a proportion of non-empty voxels using a specific sampling strategy and masking ratio. Various features within the voxel can be masked for a particular learning target in this general framework to instantiate different self-supervised learning tasks. Specifically, task-specific features of the points in the sampled voxels are replaced with", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a learnable vector. Subsequently, all non-empty voxels are encoded by a voxel encoder, processed by a backbone network to extract features, and followed by a lightweight decoder that reconstructs task-specific targets from each voxel feature. Each stage is described in the following sections. In this paper, we employ SST [12], which demonstrates superior performance in 3D detection. The pre-trained weights of the voxel encoder and the SST backbone are utilized as initialization for the downstream 3D detection task.", + "bbox": [ + 496, + 369, + 890, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Voxelization. Let $p$ denote a point with coordinates $x, y, z$ . Given a point cloud $P = \\{p_i = [x_i, y_i, z_i]^T \\in \\mathbb{R}^3\\}_{i=1\\dots n}$ containing $n$ points and range $W, H, D$ , we partition the space into a 3D grid using voxel sizes $v_W, v_H, v_D$ along the three axes, grouping points in the same voxel together. As a common technique [12, 19, 49], we decorate each point with $x^c, y^c, z^c, x^v, y^v, z^v$ , where the superscript $c$ denotes distance to the clustering center of all points in each voxel and the superscript $v$ denotes distance to the voxel center. Considering the sparse and irregular nature of LiDAR point clouds, we focus only on non-empty voxels. Assuming non-empty voxels contain at most $T$ points and a total of $N$ non-empty voxels, we obtain the voxel-based point cloud representation as $M = \\{V_i\\}_{i=1\\dots N}$ , where $V_i = \\{p_j = [x_j, y_j, z_j, x_j^c, y_j^c, z_j^c, x_j^v, y_j^v, z_j^v]^T \\in \\mathbb{R}^9\\}_{j=1\\dots t_i}$ and $t_i \\leq T$ . We assume each non-empty voxel contains $T$ points for simplicity. We also maintain each voxel's coordinate in the 3D grid with $C_i \\in \\mathbb{R}^3$ .", + "bbox": [ + 496, + 506, + 892, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked voxel sampling. A key component of masked autoencoders is the mask sampling strategy. Unlike 2D pixels that uniformly distribute on the image plane, voxels are unevenly distributed due to the sparse and irregular nature of LiDAR point clouds. Random sampling may result in information loss in regions with sparse points. Inspired by furthest point sampling [27] (FPS), which samples an evenly distributed point set while maintaining the key struc", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13447", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture of the data, we propose a Reversed-Furthest-Voxel-Sampling (R-FVS) scheme. Given a masking ratio $r$ , we apply FPS to select $\\lfloor N(1 - r)\\rfloor$ voxels from all non-empty voxels according to each voxel's coordinate $C_i$ . These sampled voxels are kept to preserve points in sparse regions and maintain key geometric structures in the data, while the remaining unsampled voxels are masked during training.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masking voxels. We directly mask the raw inputs of voxels, providing a unified formulation for the masked voxel modeling framework. For each of the remaining unsampled $\\lceil Nr \\rceil$ voxels to be masked, we replace some of the original features of each point in the voxel with a shared learnable vector, denoted as a mask token $m$ . The dimensions of $m$ and the replaced point features vary for different tasks.", + "bbox": [ + 75, + 198, + 467, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Voxel encoding and decoding. All the non-empty voxels are then encoded by a voxel encoder (VE) and the SST backbone. As the features of masked voxels are also extracted by the backbone, we only need a lightweight task-specific MLP head to decode and predict their original information [36]. The loss is computed only on masked voxels, following previous practices [15, 36].", + "bbox": [ + 75, + 304, + 467, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Masked Voxel Jigsaw (MVJ)", + "text_level": 1, + "bbox": [ + 76, + 421, + 331, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object detection in LiDAR point clouds necessitates a representation capable of capturing contextual information, such as the distribution and relationships among voxels. We propose a jigsaw puzzle that obscures position information, compelling the model to learn these relationships.", + "bbox": [ + 75, + 446, + 467, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masking position information. Given a position-masked voxel, we replace the absolute coordinates $x_{j}, y_{j}, z_{j}$ of each point with a mask token $m_{v} \\in \\mathbb{R}^{3}$ , while preserving the local shape by retaining the decorated coordinates of each point within the voxel (Fig. 2). The masked voxel is then represented as $V_{\\mathrm{masked}} = \\{p_{j} = [m_{v}, x_{j}^{c}, y_{j}^{c}, z_{j}^{c}, x_{j}^{v}, y_{j}^{v}, z_{j}^{v}]^{T} \\in \\mathbb{R}^{9}\\}_{j=1\\dots T}$ . All voxels are subsequently input to the voxel encoder and the SST backbone. Notably, positional embeddings are not added to the voxels before being fed to the backbone, as the model is tasked with predicting position information.", + "bbox": [ + 75, + 522, + 467, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prediction target and loss function. Instead of directly predicting the absolute position of the masked voxels, we formulate a classification task to facilitate optimization. Owing to the extensive range of LiDAR point clouds, the SST backbone partitions the 3D voxelized grid into windows, akin to the Swin-transformer [22], to enable efficient attention calculation within windows. We employ the same window partitioning method, requiring the model to predict only the relative index of the masked voxel within its corresponding window. Assuming a 3D window contains $N_x, N_y, N_z$ voxels along the $x, y, z$ axes and the voxel coordinates of a masked voxel are $X, Y, Z$ , the relative coordinates $I_x, I_y, I_z$ of the masked voxel with respect to its window are calculated as $I_x = X \\mod N_x, I_y = Y \\mod N_y$ .", + "bbox": [ + 75, + 688, + 467, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mod $N_{y}, I_{z} = Z \\mod N_{z}$ . The relative index is then given by $I = I_{x} + I_{y}N_{x} + I_{z}N_{x}N_{y}$ . The prediction head outputs a classification vector $\\hat{v} \\in \\mathbb{R}^{N_xN_yN_z}$ representing the probabilities of the masked voxel occupying each position. The prediction loss for all masked voxel is calculated using the Cross-Entropy loss as follows:", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M V J} = \\frac {1}{R _ {p}} \\sum_ {i = 1} ^ {R _ {p}} \\text {C r o s s E n t r o p y} (\\hat {v} _ {i}, I _ {i}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 188, + 890, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R_{p}$ denotes the number of position-masked voxels.", + "bbox": [ + 500, + 236, + 877, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Masked Voxel Reconstruction (MVR)", + "text_level": 1, + "bbox": [ + 500, + 258, + 823, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To incorporate local shape information within each voxel, we introduce masked voxel reconstruction (MVR). MVR masks the absolute and relative coordinates of all points in shape-masked voxels to learn point distributions, preserving only one point to provide the voxel position.", + "bbox": [ + 496, + 282, + 890, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masking voxel shapes. In a given voxel, both the original and decorated coordinates of points represent the local shape of the voxel (Fig. 2). We employ a shared learnable token $m_p \\in \\mathbb{R}^9$ to replace all point features in every shape-masked voxel, except for one point. This particular point conveys the voxel's positional information without revealing the voxel shape, which would render the reconstruction task trivial. Consequently, the masked voxel becomes $V_{\\mathrm{masked}} = \\{[x_1, y_1, z_1, x_1^c, y_1^c, z_1^c, x_1^v, y_1^v, z_1^v]\\} \\cup \\{p_j = m_p\\}_{j=2\\dots T}$ .", + "bbox": [ + 496, + 358, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction target and loss function. We reconstruct each masked voxel and use the $L2$ Chamfer Distance loss to measure the discrepancy between the reconstructed and target point distributions. This loss function is insensitive to point density [34], which is crucial since each voxel may contain varying numbers of points. Assuming each reconstructed voxel contains $n$ points, the reconstruction head outputs a vector containing the coordinates of each point in the masked voxel, $\\hat{v} \\in \\mathbb{R}^{3n}$ . To pre-train the model more stably, we normalize the target point distribution using each point's distance to the voxel center as ground truth and scale the distance between 0 and 1. Supposing each reconstructed masked voxel is $\\hat{V} = \\{\\hat{p}_j = [\\hat{x}_j, \\hat{y}_j, \\hat{z}_j]\\}_{j=1\\dots n}$ and the ground truth is $V = \\{p_j = [x_j^v, y_j^v, z_j^v]\\}_{j=1\\dots t_i}$ , the total loss across all masked voxels is given by:", + "bbox": [ + 496, + 508, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M V R} = \\frac {1}{R _ {s}} \\sum_ {i = 1} ^ {R _ {s}} L _ {C D} \\left(\\hat {V} _ {i}, V _ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 741, + 890, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R_{s}$ represents the number of shape-masked voxels and $L_{CD}$ is the $L2$ Chamfer Distance loss calculated as:", + "bbox": [ + 496, + 787, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {C D} = \\frac {1}{\\left| \\hat {V} _ {i} \\right|} \\sum_ {\\hat {p} \\in \\hat {V} _ {i}} \\min _ {p \\in V _ {i}} \\| \\hat {p} - p \\| _ {2} ^ {2} \\tag {3} \\\\ + \\frac {1}{| V _ {i} |} \\sum_ {p \\in V _ {i}} \\min _ {\\hat {p} \\in \\hat {V} _ {i}} \\| p - \\hat {p} \\| _ {2} ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 823, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13448", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Joint Pre-Training", + "text_level": 1, + "bbox": [ + 76, + 90, + 256, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We jointly pre-train the model using both MVJ and MVR tasks to enable the model to learn both point and voxel distributions. We use R-FVS to sample $R_{s}$ voxels for MVR and $R_{p}$ voxels for MVJ. As illustrated in Fig. 2, the voxel encoder and the backbone extract features from both types of masked voxels and unmasked voxels. Masked features are decoded to recover their corresponding targets, and Cross-Entropy loss and Chamfer Distance loss are utilized for supervision. The total loss for joint pre-training is expressed as follows:", + "bbox": [ + 75, + 112, + 470, + 263 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\alpha L _ {M V J} + \\beta L _ {M V R} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 276, + 468, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are balancing coefficients.", + "bbox": [ + 76, + 303, + 352, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Data-Efficient Benchmark on Waymo", + "text_level": 1, + "bbox": [ + 76, + 330, + 416, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the pre-trained representation, we transfer the weights of the voxel encoder and the backbone for object detection. Previous benchmarks fail to effectively reveal the effectiveness of pre-training strategies due to data diversity issues. We propose a new data-efficient benchmark to address these limitations.", + "bbox": [ + 75, + 357, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observation: Incomplete model convergence. Prior works [20, 40] fine-tune the model on different splits with the same epochs, causing reduced iteration numbers when dataset sizes decrease. As illustrated in Fig. 3a, by training for more epochs (but the same iterations), the performance of SST significantly improves, suggesting that the models may not fully converge on smaller splits when training epochs remain constant. This entangled factor prevents accurate evaluation of pre-training strategies.", + "bbox": [ + 75, + 446, + 468, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Issue: Fine-Tuning splits with similar data diversities. With the same training iterations, models trained on different splits display comparable performance, despite significant differences in data amounts (Fig. 3a). We find that scene diversity is a crucial factor. Previous works [20, 40] uniformly sample the entire Waymo training set to create different fine-tuning splits, for instance, sampling one frame from every two consecutive frames for a $50\\%$ split. However, neighboring frames in self-driving datasets are similar due to the short time difference between consecutive frames (e.g., only $0.1s$ in Waymo [30]). This results in similar scene diversity across splits, leading to comparable detection performance using the same training iterations.", + "bbox": [ + 75, + 584, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We argue that such an experimental setting does not accurately represent a typical application scenario of self-supervised pre-training, where annotated fine-tuning data is less abundant and diverse compared to pre-training data. Furthermore, it fails to evaluate how pre-training benefits downstream tasks when varying amounts and diversities of labeled data are available. We provide an additional example in the supplementary material to further illustrate", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2608633e5dbcc7499a8da46049e8c7601e74992ba225c2fc0fd9c527a902683e.jpg", + "image_caption": [ + "(a) Existing" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 692, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59d89ebeb26e2e21bb5ea0b6aae706219fd9f3d8b6e6ee6784cb051e7ec6ca52.jpg", + "image_caption": [ + "(b) Ours", + "Figure 3. Comparison of data-efficient benchmarks. Previous uniformly sampled splits exhibit similar diversity, as evidenced by similar performance with the same training iterations. Our benchmark features varying data amounts and diversities, ensuring model convergence. Dashed lines show baseline performances." + ], + "image_footnote": [], + "bbox": [ + 697, + 88, + 890, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "these issues with previous benchmarks. Our proposed data-efficient benchmark addresses these issues, offering a more precise evaluation of pre-training strategies.", + "bbox": [ + 496, + 311, + 890, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solution: Sequence-Based data sampling. We sample fine-tuning data using different scene sequences rather than uniformly sampling varying numbers of frames from identical sequences. The Waymo training split comprises 798 distinct scene sequences, each lasting $20s$ with approximately 200 frames [30]. We randomly sample $5\\%$ , $10\\%$ , $20\\%$ , $50\\%$ of the scene sequences, respectively, using all frames within the sampled sequences for fine-tuning. Consequently, data diversities and amounts vary across each split. Each larger split also encompasses the smaller ones, enabling measurement of performance changes with additional fine-tuning data. Furthermore, we randomly sample three $5\\%$ and $10\\%$ splits and report average results to reduce variance on these smaller sets (see our supplementary material).", + "bbox": [ + 496, + 356, + 892, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For different fine-tuning splits, we train the model until performance saturates (i.e., overfitting) and determine the number of epochs used for experiments accordingly. For example, for $5\\%$ data, we try $\\{48, 60, 66, 72, 78, 84\\}$ epochs and ultimately select 72. The baseline performances across various training iterations are shown in Fig. 3b. With our new benchmark, baseline performance and convergence speed vary as anticipated for different splits.", + "bbox": [ + 496, + 566, + 890, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 700, + 633, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 726, + 705, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. The Waymo dataset [30] consists of 1150 self-driving scenes with 798 scenes as the training split, 202 scenes as the validation split, and 150 scenes as the testing split. We pre-train on the training split, which contains 158,240 annotated frames. For the downstream data-efficient 3D object detection task, we fine-tune the model using our sampled splits (as described in Sec. 4) and evaluate performance on the standard Waymo validation split with 3 classes (cars, pedestrians, and cyclists) and two difficulty levels (L1 and L2). We adopt L2 mean average preci", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13449", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/818118d44b028a1012bdd330c3327d5dd8a76846789cd53ee58b165283d50811.jpg", + "table_caption": [ + "Table 1. Data-efficient 3D object detection results of SST on the Waymo validation set. SST is pre-trained using various self-supervised methods on Waymo training split, with weights applied for the downstream detection task. \"Random initialization\" denotes training from scratch. The model performances are shown using different amounts and diversities of fine-tuning data." + ], + "table_footnote": [], + "table_body": "
Data amountInitializationOverallCarPedestrianCyclist
L2 mAPL2 mAPHL2 mAPL2 mAPHL2 mAPL2 mAPHL2 mAPL2 mAPH
5%Random44.4140.3451.0150.4952.7442.2629.4928.27
PointContrast [35]45.3241.3052.1251.6153.6843.2230.1629.09
ProposalContrast [40]46.6242.5852.6752.1954.3143.8232.8731.72
MV-JAR (Ours)50.52+6.1146.68+6.3456.4756.0157.6547.6937.4436.33
10%Random54.3150.4654.8454.3760.5550.7147.5546.29
PointContrast [35]53.6949.9454.7654.3059.7550.1246.5745.39
ProposalContrast [40]53.8950.1355.1854.7160.0150.3946.4845.28
MV-JAR (Ours)57.44+3.1354.06+3.6058.4358.0063.2854.6650.6349.52
20%Random60.1656.7858.7958.3565.6357.0456.0754.94
PointContrast [35]59.3555.7858.6458.1864.3955.4355.0253.73
ProposalContrast [40]59.5255.9158.6958.2264.5355.4555.3654.07
MV-JAR (Ours)62.28+2.1159.15+2.3761.8861.4566.9859.0257.9857.00
50%Random66.4363.3663.8163.3870.7863.0564.7163.66
PointContrast [35]65.5162.2162.6662.2369.8261.5364.0462.86
ProposalContrast [40]65.7662.4962.9362.5070.0961.8664.2663.11
MV-JAR (Ours)66.70+0.2763.69+0.3364.3063.8971.1463.5764.6563.63
100%Random68.5065.5464.9664.5672.3864.8968.1767.17
PointContrast [35]68.0664.8464.2463.8271.9263.8168.0366.89
ProposalContrast [40]68.1765.0164.4264.0071.9463.9468.1667.10
MV-JAR (Ours)69.16+0.6666.20+0.6665.5265.1272.7765.2869.1968.20
", + "bbox": [ + 99, + 142, + 874, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sion (L2 mAP) and L2 mean average precision with heading (L2 mAPH) as main evaluation metrics. We also evaluate the model on the KITTI dataset [14] for the detection task. It contains 7481 training samples and 7518 test samples.", + "bbox": [ + 75, + 467, + 468, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We employ the official implementation and training settings of SST [12] for pre-training and fine-tuning unless specified otherwise. The Waymo dataset's point cloud range is $W \\times H \\times D = 149.76m \\times 149.76m \\times 6m$ , and the voxel size is $v_{W} \\times v_{H} \\times v_{D} = 0.32m \\times 0.32m \\times 6m$ . For MVR, we predict 15 points per voxel with a masking ratio of 0.05. For MVJ, the window size is $N_{x} \\times N_{y} \\times N_{z} = 12 \\times 12 \\times 1$ and the masking ratio is 0.1. We set the loss weight $\\alpha = \\beta = 1$ , pre-train for 6 epochs with an initial learning rate of $5e - 6$ . When fine-tuning, we initialize the voxel encoder and SST backbone with pre-trained weights, and other training hyperparameters remain unchanged. We set KITTI's point cloud range to $69.12m \\times 79.36m \\times 4m$ , voxel size to $0.32m \\times 0.32m \\times 4m$ , and train for 160 epochs. We use the AdamW optimizer with a batch size of 8 and a cyclic learning rate scheduler with cosine annealing for all training.", + "bbox": [ + 75, + 529, + 472, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results on Waymo", + "text_level": 1, + "bbox": [ + 76, + 800, + 302, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We fine-tune the model using data splits described in Sec. 4. For $\\{5\\%, 10\\%, 20\\%, 50\\%, 100\\}$ data, the model is trained for $\\{72, 60, 42, 36, 24\\}$ epochs. We use the randomly initialized model's performances as our baselines. We also include experiments with models using", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "convolutional backbones in the supplementary material.", + "bbox": [ + 498, + 467, + 867, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results with different initialization. Tab. 1 presents the SST model's performances with various initializations on our proposed data-efficient benchmark. Our pre-training method consistently improves SST baselines across all finetuning data amounts, especially when data is scarce. Our method achieves up to $15.7\\%$ relative improvement, from $40.34\\%$ L2 mAPH to $46.68\\%$ L2 mAPH, when using only $5\\%$ fine-tuning data. The performance gain exists across object categories, indicating effective general representation learning. As more fine-tuning data is introduced, the performance gain persists, and when fine-tuning with the entire Waymo training set, our pre-training enhances detection performance from $65.54\\%$ mAPH to $66.20\\%$ mAPH.", + "bbox": [ + 496, + 486, + 890, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that with $50\\%$ and $100\\%$ fine-tuning data, the improvements from our pre-training are not as significant as with less data. We observe that the baseline result on $50\\%$ data (63.36% mAPH) is comparable to that on $100\\%$ data (65.54% mAPH), suggesting that training data is no longer the bottleneck. The original SST, with only 2.1M parameters, is a lightweight network. We scale up SST to 8.3M parameters by using larger hidden layer channels and observe larger benefits from pre-training on the $50\\%$ fine-tuning split, from $63.13\\%$ L2 mAPH to $64.24\\%$ L2 mAPH. Therefore, we argue that the SST model's capacity, rather than the fine-tuning data amount, becomes the bottleneck for detection performance.", + "bbox": [ + 496, + 685, + 892, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with contrastive learning methods. We", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13450", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c7e29b766c5c852d729b1cb467b4cea66f5da5c6d70132b8ecdabceedb8a3857.jpg", + "table_caption": [ + "Table 2. Transferring results on the KITTI validation split. We pre-train SST with the Waymo training split and fine-tune with the KITTI training split." + ], + "table_footnote": [], + "table_body": "
InitializationOverall
EasyMod.Hard
Random74.7163.4360.00
PointContrast [35]73.3562.5359.01
ProposalContrast [40]73.6363.3459.40
MV-JAR (Ours)75.2263.8060.35
", + "bbox": [ + 142, + 141, + 408, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "utilize the official implementation of the recently proposed ProposalContrast [40] and reimplement PointContrast [35] for comparison. As demonstrated in Tab. 1, both contrastive learning methods offer benefits only with $5\\%$ fine-tuning data, and their improvements diminish with higher baseline performances. This may be due to the difficulty in smoothly learning representations. A key challenge in LiDAR-based contrastive learning is constructing matched pairs. The point pairs used by PointContrast face the issue of spatially adjacent points with similar features being assigned as negative samples. We observe difficulty in convergence when sampling excessive pairs. ProposalContrast alleviates the issue by contrasting regions but does not entirely resolve it. Alternatively, our method offers a new avenue for developing self-supervised pre-training through masked voxel modeling on LiDAR point clouds.", + "bbox": [ + 75, + 258, + 468, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Pre-training Accelerates Convergence", + "text_level": 1, + "bbox": [ + 76, + 510, + 403, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate how MV-JAR pre-training accelerates convergence, we fine-tune the SST model on the entire Waymo training split. We train the SST model for different epochs and report the performance of both fine-tuning and training from scratch. As illustrated in Fig. 1, MV-JAR pre-training significantly enhances the convergence speed of SST. By fine-tuning for only 3 epochs, the SST model achieves $62.48\\%$ L2 mAPH. Remarkably, fine-tuning for just 12 epochs results in a performance comparable to training from scratch for 24 epochs, effectively halving the convergence time. As the number of fine-tuning epochs increases, the performance gain provided by MV-JAR does not wane. Pre-training can still improve the model's converged performance by $0.74\\%$ L2 mAPH, highlighting the superiority of the pre-trained representation.", + "bbox": [ + 75, + 534, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Transferring Results on KITTI", + "text_level": 1, + "bbox": [ + 76, + 770, + 354, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the transferability of the learned representation, we fine-tune various pre-trained SST models on the KITTI dataset. The original SST paper [12] did not conduct experiments on the KITTI dataset; thus, we follow the training schedule and setup of SST and PointPillars [19] within the MMDetection3D framework [9]. As demonstrated in Tab. 2, the performance gain from our MV-JAR pre-training", + "bbox": [ + 75, + 794, + 467, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "persists when the representation is transferred to a different domain, indicating that the model learns a generic representation through pre-training. While the KITTI training samples account for approximately $5\\%$ of the entire Waymo training split, the relative improvement is much smaller than transferring to the $5\\%$ Waymo split. We hypothesize that this may be attributed to the domain gap between the Waymo and KITTI datasets.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 224, + 663, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to investigate the effectiveness of our proposed Reversed-Furthest-Voxel-Sampling (R-FVS) strategy. Additionally, we explore the optimal masking ratio and the impact of MVR and MVJ pre-training. All pre-training experiments utilize the entire Waymo training split, and we fine-tune the model on one of our $5\\%$ data splits.", + "bbox": [ + 496, + 248, + 890, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask sampling strategy. Our R-FVS mask sampling strategy masks voxels that are not sampled by FPS, aiming to avoid masking voxels located in sparse regions. We also examine random sampling and an opposing sampling strategy called FVS, which masks voxels sampled by FPS. With FVS sampling, regions with sparse points are more likely to be masked, leading to greater information loss compared to random sampling and R-FVS sampling. Tab. 3 presents the fine-tuning results of pre-training with the three sampling strategies. R-FVS sampling performs the best, while FVS performs the worst. These findings confirm that minimizing information loss benefits pre-training and validate the effectiveness of our R-FVS strategy.", + "bbox": [ + 496, + 354, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Masking ratio. Tab. 4 illustrates the impact of different masking ratio combinations. Specifically, MVR with a 0.1 masking ratio and MVJ with a 0.05 masking ratio yield the best results. With this combination, the overall masking ratio is 0.15, which is significantly lower than the masking ratios in the image [15, 36] or video domain [13, 32]. A high masking ratio in masked autoencoders typically indicates information redundancy [13, 15, 32, 36]. We postulate that two factors contribute to our low masking ratio: 1) LiDAR point clouds represent a vast space where each meaningful object occupies only a small fraction. The information on each object is not highly redundant, especially when the objects are distant from the sensor. This contrasts with images in datasets like ImageNet, where objects often encompass a significant portion of the image. 2) The SST model calculates attention only within a partitioned window, which restricts the information used. We leave this question for future exploration.", + "bbox": [ + 496, + 551, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pre-training effects of MVR and MVJ. We investigate the individual pre-training effects of MVR and MVJ by optimizing the masking ratio for each task and reporting the fine-tuning performances in Tab. 5. Both MVR and MVJ pre-training enhance performance compared to random ini", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13451", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/122ade527333146bfb78f1e4479a3233e9ad229d33c00b1ebdc20c6cbff47fa8.jpg", + "image_caption": [ + "Figure 4. Visualization of the reconstruction results of MV-JAR on the Waymo validation set, compared with ground truths." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 480, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8054ab95e8e42dbecbfa41a75bd0d776b159a5be7621c11e4b0fe948951c8d30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 93, + 888, + 337 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b12d36ebc59d7493558f4cb53fc0826b4477b3a00a7458efd400af1cbf787ff2.jpg", + "table_caption": [ + "Table 3. Ablation study on mask sampling strategies." + ], + "table_footnote": [], + "table_body": "
Mask samplingOverall
L2 mAPL2 mAPH
FVS49.7345.88
Random49.8646.06
R-FVS50.5246.68
", + "bbox": [ + 93, + 421, + 313, + 501 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6db408c86e77732758377b00ce744551e2d216f84e25032abf947a6819a7304d.jpg", + "table_caption": [ + "Table 4. Ablation study on masking ratios." + ], + "table_footnote": [], + "table_body": "
Masking ratioOverall
MVJMVRL2 mAPL2 mAPH
0.30.149.7446.05
0.050.149.6745.92
0.10.149.8545.97
0.10.0550.5246.68
0.10.350.2546.35
", + "bbox": [ + 375, + 402, + 602, + 506 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/13c1cd4f80ea2da424b1e6f6abcc878beda39ece09b43ad5a8b2d3dcd4f31711.jpg", + "table_caption": [ + "Table 5. Ablation study on pre-training effects of MVR and MVJ." + ], + "table_footnote": [], + "table_body": "
TaskOverall
L2 mAPL2 mAPH
MVR46.5742.73
MVJ49.9646.10
MV-JAR50.5246.68
", + "bbox": [ + 679, + 421, + 870, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tialization, with MVJ outperforming MVR. This finding suggests that capturing voxel distributions plays a more significant role in representation learning, which is consistent with the fact that LiDAR detectors process points by down-sampling them into voxels. Our integrated MV-JAR method demonstrates further improvement over MVR and MVJ, validating the effectiveness of jointly capturing point and voxel distributions. Additionally, we offer an analysis of improvements across various distances in our supplementary material.", + "bbox": [ + 75, + 530, + 468, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Visualization", + "text_level": 1, + "bbox": [ + 76, + 694, + 212, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 displays the MV-JAR reconstruction results on the Waymo validation set, including MVJ and MVR outcomes. In our pre-training, MVJ achieves approximately $89\\%$ classification accuracy of masked voxels on the validation set, accurately reconstructing masked voxels in the correct positions most of the time. MVR can capture the point distributions within voxels, evidenced by reconstructed ground points aligning with ground circles. However, MVR struggles to capture detailed distributions, likely due to the discrete sampling of LiDAR points from continuous surfaces, resulting in considerable variability in the data and increased difficulty in capturing finer details.", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 529, + 625, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce the Masked Voxel Jigsaw and Reconstruction (MV-JAR) pre-training method for LiDAR detectors. MV-JAR captures both point and voxel distributions of LiDAR point clouds, enabling models to learn effective and generic representations. We also develop a Reversed-Furthest-Voxel-Sampling strategy to address the uneven distribution of LiDAR points. Comprehensive experiments on the Waymo and KITTI datasets show that our method consistently and significantly enhances the detector's performance across different data scale regimes. MV-JAR offers a promising alternative for LiDAR-based self-supervised pre-training through Masked Voxel Modeling. Additionally, we establish a new data-efficient benchmark on the Waymo dataset, incorporating fine-tuning splits with diverse data variations. This benchmark effectively assesses the impact of pre-training on downstream tasks with varying amounts and diversities of labeled data.", + "bbox": [ + 496, + 561, + 890, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project is funded in part by the Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd. under the Innovation and Technology Commission (ITC)'s InnoHK.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13452", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv:2106.08254, 2021. 1, 2", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2", + "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. arXiv:1903.11027, 2019. 2", + "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 1", + "[5] Runjian Chen, Yao Mu, Runsen Xu, Wenqi Shao, Chenhan Jiang, Hang Xu, Yu Qiao, Zhenguo Li, and Ping Luo. CO3: Cooperative unsupervised 3d representation learning for autonomous driving. In ICLR, 2023. 2", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 1", + "[7] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv:2202.03026, 2022.1, 2", + "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297, 2020. 1", + "[9] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.7", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.2", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[12] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, 2022. 1, 2, 3, 6, 7" + ], + "bbox": [ + 78, + 116, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv:2205.09113, 2022. 2, 7", + "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 6", + "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2, 4, 7", + "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 1", + "[17] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2", + "[18] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In CVPR, 2021. 2", + "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 1, 3, 7", + "[20] Hanxue Liang, Chenhan Jiang, Dapeng Feng, Xin Chen, Hang Xu, Xiaodan Liang, Wei Zhang, Zhenguo Li, and Luc Van Gool. Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In CVPR, 2021. 1, 2, 5", + "[21] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 2", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In CVPR, 2021. 3, 4", + "[23] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In CVPR, 2021. 2, 3", + "[24] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In CVPR, 2021. 2", + "[25] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In CVPR, 2021. 2", + "[26] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2", + "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3", + "[28] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2", + "[29] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. In NeurIPS, 2019. 2", + "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou," + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13453", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In CVPR, 2020. 2, 5", + "[31] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. In ECCV, 2022. 2", + "[32] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 2, 7", + "[33] Tai Wang, Xinge Zhu, and Dahua Lin. Reconfigurable voxels: A new representation for lidar-based point clouds. In CoRL, 2020. 1", + "[34] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. In NeurIPS, 2021. 4", + "[35] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 2, 6, 7", + "[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 1, 2, 4, 7", + "[37] Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, Christoph Feichtenhofer, et al. Masked autoencoders that listen. arXiv:2207.06405, 2022. 2", + "[38] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 2018. 1", + "[39] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In CVPR, 2018. 2", + "[40] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV,2022.1,2,5,6,7", + "[41] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 1", + "[42] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 2", + "[43] Shuangfei Zhai, Navdeep Jaitly, Jason Ramapuram, Dan Busbridge, Tatiana Likhomanenko, Joseph Yitan Cheng, Walter Talbott, Chen Huang, Hanlin Goh, and Joshua Susskind. Position prediction as an effective pretraining strategy. In ICML, 2022. 2", + "[44] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv:2205.14401, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. Dense siamese network for dense unsupervised learning. In ECCV, 2022. 1", + "[46] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In CVPR, 2021. 2", + "[47] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv:2111.07832, 2021. 1, 2", + "[48] Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image analysis. arXiv:2203.05573, 2022. 2", + "[49] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 3", + "[50] Xinge Zhu, Yuexin Ma, Tai Wang, Yan Xu, Jianping Shi, and Dahua Lin. Ssn: Shape signature networks for multi-class object detection from point clouds. In ECCV, 2020. 1" + ], + "bbox": [ + 501, + 92, + 890, + 347 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13454", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_model.json b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..617f2da137366b4e789cf9522786810a815fd39a --- /dev/null +++ b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.13, + 0.838, + 0.177 + ], + "angle": 0, + "content": "MV-JAR: Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.202, + 0.8, + 0.239 + ], + "angle": 0, + "content": "Runsen Xu\\(^{1,2}\\) Tai Wang\\(^{1,2}\\) Wenwei Zhang\\(^{3,2}\\) Runjian Chen\\(^{4}\\) Jinkun Cao\\(^{5}\\) Jiangmiao Pang\\(^{2\\boxtimes}\\) Dahua Lin\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.24, + 0.813, + 0.275 + ], + "angle": 0, + "content": "1The Chinese University of Hong Kong 2Shanghai AI Laboratory 3S-Lab, NTU 4The University of Hong Kong 5Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.277, + 0.846, + 0.31 + ], + "angle": 0, + "content": "{runsenxu,wt019,dhlin}@ie.cuhk.edu.hk, wenwei001@ntu.edu.sg, rjchen@connect.hku.hk, jinkunc@andrew.cmu.edu, pangjiangmiao@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.36, + 0.313, + 0.375 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.473, + 0.784 + ], + "angle": 0, + "content": "This paper introduces the Masked Voxel Jigsaw and Reconstruction (MV-JAR) method for LiDAR-based self-supervised pre-training and a carefully designed data-efficient 3D object detection benchmark on the Waymo dataset. Inspired by the scene-voxel-point hierarchy in downstream 3D object detectors, we design masking and reconstruction strategies accounting for voxel distributions in the scene and local point distributions within the voxel. We employ a Reversed-Furthest-Voxel-Sampling strategy to address the uneven distribution of LiDAR points and propose MV-JAR, which combines two techniques for modeling the aforementioned distributions, resulting in superior performance. Our experiments reveal limitations in previous data-efficient experiments, which uniformly sample fine-tuning splits with varying data proportions from each LiDAR sequence, leading to similar data diversity across splits. To address this, we propose a new benchmark that samples scene sequences for diverse fine-tuning splits, ensuring adequate model convergence and providing a more accurate evaluation of pre-training methods. Experiments on our Waymo benchmark and the KITTI dataset demonstrate that MV-JAR consistently and significantly improves 3D detection performance across various data scales, achieving up to a \\(6.3\\%\\) increase in mAPH compared to training from scratch. Codes and the benchmark are available at https://github.com/SmartBot-PJLab/MV-JAR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.808, + 0.21, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.833, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Self-supervised pre-training has gained considerable attention, owing to its exceptional performance in visual representation learning. Recent advancements in contrastive" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.344, + 0.703, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.344, + 0.892, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.556, + 0.894, + 0.599 + ], + "angle": 0, + "content": "Figure 1. 3D object detection results on the Waymo dataset. Our MV-JAR pre-training accelerates model convergence and greatly improves the performance with limited fine-tuning data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.658 + ], + "angle": 0, + "content": "learning [4, 6, 8, 16, 45] and masked autoencoders [1, 7, 15, 36, 47] for images have sparked interest among researchers and facilitated progress in modalities such as point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.825 + ], + "angle": 0, + "content": "However, LiDAR point clouds differ from images and dense point clouds obtained by reconstruction as they are naturally sparse, unorganized, and irregularly distributed. Developing effective self-supervised proxy tasks for these unique properties remains an open challenge. Constructing matching pairs for contrastive learning in geometry-dominant scenes is more difficult [20, 40], as points or regions with similar geometry may be assigned as negative samples, leading to ambiguity during training. To address this, our study explores masked voxel modeling paradigms for effective LiDAR-based self-supervised pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Downstream LiDAR-based 3D object detectors [12, 19, 33, 38, 41, 50] typically quantize the 3D space into voxels and encode point features within them. Unlike pixels, which are represented by RGB values, the 3D space presents a scene-voxel-point hierarchy, introducing new" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.24, + 0.9 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13445" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "challenges for masked modeling. Inspired by this, we design masking and reconstruction strategies that consider voxel distributions in the scene and local point distributions in the voxel. Our proposed method, Masked Voxel Jigsaw And Reconstruction (MV-JAR), harnesses the strengths of both voxel and point distributions to improve performance." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.183, + 0.468, + 0.393 + ], + "angle": 0, + "content": "To account for the uneven distribution of LiDAR points, we first employ a Reversed-Furthest-Voxel-Sampling (R-FVS) strategy that samples voxels to mask based on their sparseness. This approach prevents masking the furthest distributed voxels, thereby avoiding information loss in regions with sparse points. To model voxel distributions, we propose Masked Voxel Jigsaw (MVJ), which masks the voxel coordinates while preserving the local shape of each voxel, enabling scene reconstruction akin to solving a jigsaw puzzle. For modeling local point distributions, we introduce Masked Voxel Reconstruction (MVR), which masks all coordinates of points within the voxel but retains one point as a hint for reconstruction. Combining these two methods enhances masked voxel modeling." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.394, + 0.468, + 0.649 + ], + "angle": 0, + "content": "Our experiments indicate that existing data-efficient experiments [20, 40] inadequately evaluate the effectiveness of various pre-training methods. The current benchmarks, which uniformly sample frames from each data sequence to create diverse fine-tuning splits, exhibit similar data diversity due to the proximity of neighboring frames in a sequence [3, 14, 30]. Moreover, these experiments train models for the same number of epochs across different fine-tuning splits, potentially leading to incomplete convergence. As a result, the benefits of pre-trained representations become indistinguishable across splits once the object detector is sufficiently trained on the fine-tuning data. To address these shortcomings, we propose sampling scene sequences to form diverse fine-tuning splits and establish a new data-efficient 3D object detection benchmark on the Waymo [30] dataset, ensuring sufficient model convergence for a more accurate evaluation." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.65, + 0.468, + 0.83 + ], + "angle": 0, + "content": "We employ the Transformer-based SST [12] as our detector and pre-train its backbone for downstream detection tasks. Comprehensive experiments on the Waymo and KITTI [14] datasets demonstrate that our pre-training method significantly enhances the model's performance and convergence speed in downstream tasks. Notably, it improves detection performance by \\(6.3\\%\\) mAPH when using only \\(5\\%\\) of scenes for fine-tuning and reduces training time by half when utilizing the entire dataset (Fig. 1). With the representation pre-trained by MV-JAR, the 3D object detectors pre-trained on Waymo also exhibit generalizability when transferred to KITTI." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.845, + 0.217, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Self-supervised learning for point clouds. Annotating point clouds demands significant effort, necessitating self" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.89, + 0.453 + ], + "angle": 0, + "content": "supervised pre-training methods. Prior approaches primarily focus on object CAD models [21, 26, 29, 39, 42, 44] and indoor scenes [17, 35, 46]. Point-BERT [42] applies BERT-like paradigms for point cloud recognition, while Point-MAE [26] reconstructs point patches without the tokenizer. To acquire scene-level representation, PointContrast [35] introduces a contrastive learning approach that compares points in two static partial views of a reconstructed indoor scene. Its successor, SceneContrast [17], incorporates spatial information into the contrastive learning framework. However, LiDAR point clouds are irregular and dynamic, with mainstream LiDAR perception models often exhibiting distinct architectures, which obstructs the direct adaptation of object-level and indoor scene pre-training methods. STRL [18] employs contrastive learning with two temporally-correlated frames for spatiotemporal representation. GCC-3D [20] presents a framework that combines geometry-aware contrast and pseudo-instance clustering harmonization. ProposalContrast [40] targets region-level contrastive learning to improve 3D detection. CO3 [5] leverages infrastructure-vehicle-cooperation point clouds to construct effective contrastive views. In contrast, our work explores masked voxel modeling, diverging from the prevailing contrastive learning paradigm." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.459, + 0.89, + 0.789 + ], + "angle": 0, + "content": "Masked autoencoders for self-supervised pre-training. Masked language modeling plays a pivotal role in self-supervised pre-training for Transformer-based networks in natural language processing [2, 10, 28]. This approach typically masks portions of the input and pre-trains networks to predict the original information. With the successful integration of Transformers into computer vision [11], researchers have increasingly focused on masked image modeling [1, 7, 15, 36, 43, 47] to mitigate the data-intensive issue of ViT [11]. BEiT [1] employs a tokenizer to generate discrete tokens for image patches and utilizes a BERT-like framework to pre-train ViT. MAE [15] introduces an asymmetric autoencoder for reconstructing RGB pixels of original images, eliminating the need for an extra tokenizer. SimMiM [36] encodes masked raw images with Transformers and employs a lightweight prediction head for recovery. Demonstrating considerable potential in image recognition, masked autoencoders have been applied in video understanding [13, 32], medical image analysis [48], and audio processing [37]. In this work, we investigate the application of this powerful pre-training technique to LiDAR point clouds using voxel representation." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.795, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Transformer-based 3D object detection. The recent success of Vision Transformers [11] has inspired extensive research into the application of Transformer-based architectures for 3D object detection [12, 23-25, 31]. 3DETR [24] presents an end-to-end object detection framework with modified Transformer blocks for 3D point clouds. Pointformer [25] employs hierarchical Transformer blocks to" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13446" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.893, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Overview of Masked Voxel Jigsaw and Reconstruction (MV-JAR). Our R-FVS sampling method selects a specific ratio of nonempty voxels to mask. For position-masked voxels, we mask the absolute coordinates of all points within the voxel while maintaining the relative coordinates. For shape-masked voxels, we mask both absolute and relative coordinates of all points, preserving only one point. Features are extracted from all non-empty voxels, and masked features are decoded by a dedicated head to recover corresponding targets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.536 + ], + "angle": 0, + "content": "extract point features for 3D detection. Voxel Transformer [23] develops a Transformer-based 3D backbone to establish long-range relationships between voxels. Recently, SST [12] introduced a single-stride sparse Transformer as a replacement for the PointPillars [19] backbone, abandoning the multi-resolution strategy to enhance the detection of small objects. Emulating the SwinTransformer [22], SST partitions the space into windows and calculates attention only within each window for efficiency. This approach demonstrates impressive detection results, and we adopt it for our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.214, + 0.57 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.744 + ], + "angle": 0, + "content": "In this section, we first present an overview of a general masked voxel modeling framework for LiDAR point clouds and our Reversed-Furthest-Voxel-Sampling strategy to tackle the uneven distribution of LiDAR points (Sec. 3.1). Masked Voxel Jigsaw (MVJ) (Sec. 3.2) and Masked Voxel Reconstruction (MVR) (Sec. 3.3) are both instantiations of the general framework that learns the voxel distributions and local point distributions, respectively. Masked Voxel Jigsaw and Reconstruction (MV-JAR) consist of these two tasks to synergically learn the data distributions at two levels (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.411, + 0.773 + ], + "angle": 0, + "content": "3.1. Masked Voxel Modeling and Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, a general masked voxel modeling framework converts LiDAR point clouds into voxels and samples a proportion of non-empty voxels using a specific sampling strategy and masking ratio. Various features within the voxel can be masked for a particular learning target in this general framework to instantiate different self-supervised learning tasks. Specifically, task-specific features of the points in the sampled voxels are replaced with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.504 + ], + "angle": 0, + "content": "a learnable vector. Subsequently, all non-empty voxels are encoded by a voxel encoder, processed by a backbone network to extract features, and followed by a lightweight decoder that reconstructs task-specific targets from each voxel feature. Each stage is described in the following sections. In this paper, we employ SST [12], which demonstrates superior performance in 3D detection. The pre-trained weights of the voxel encoder and the SST backbone are utilized as initialization for the downstream 3D detection task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Voxelization. Let \\( p \\) denote a point with coordinates \\( x, y, z \\). Given a point cloud \\( P = \\{p_i = [x_i, y_i, z_i]^T \\in \\mathbb{R}^3\\}_{i=1\\dots n} \\) containing \\( n \\) points and range \\( W, H, D \\), we partition the space into a 3D grid using voxel sizes \\( v_W, v_H, v_D \\) along the three axes, grouping points in the same voxel together. As a common technique [12, 19, 49], we decorate each point with \\( x^c, y^c, z^c, x^v, y^v, z^v \\), where the superscript \\( c \\) denotes distance to the clustering center of all points in each voxel and the superscript \\( v \\) denotes distance to the voxel center. Considering the sparse and irregular nature of LiDAR point clouds, we focus only on non-empty voxels. Assuming non-empty voxels contain at most \\( T \\) points and a total of \\( N \\) non-empty voxels, we obtain the voxel-based point cloud representation as \\( M = \\{V_i\\}_{i=1\\dots N} \\), where \\( V_i = \\{p_j = [x_j, y_j, z_j, x_j^c, y_j^c, z_j^c, x_j^v, y_j^v, z_j^v]^T \\in \\mathbb{R}^9\\}_{j=1\\dots t_i} \\) and \\( t_i \\leq T \\). We assume each non-empty voxel contains \\( T \\) points for simplicity. We also maintain each voxel's coordinate in the 3D grid with \\( C_i \\in \\mathbb{R}^3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Masked voxel sampling. A key component of masked autoencoders is the mask sampling strategy. Unlike 2D pixels that uniformly distribute on the image plane, voxels are unevenly distributed due to the sparse and irregular nature of LiDAR point clouds. Random sampling may result in information loss in regions with sparse points. Inspired by furthest point sampling [27] (FPS), which samples an evenly distributed point set while maintaining the key struc" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13447" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "ture of the data, we propose a Reversed-Furthest-Voxel-Sampling (R-FVS) scheme. Given a masking ratio \\( r \\), we apply FPS to select \\( \\lfloor N(1 - r)\\rfloor \\) voxels from all non-empty voxels according to each voxel's coordinate \\( C_i \\). These sampled voxels are kept to preserve points in sparse regions and maintain key geometric structures in the data, while the remaining unsampled voxels are masked during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.468, + 0.304 + ], + "angle": 0, + "content": "Masking voxels. We directly mask the raw inputs of voxels, providing a unified formulation for the masked voxel modeling framework. For each of the remaining unsampled \\(\\lceil Nr \\rceil\\) voxels to be masked, we replace some of the original features of each point in the voxel with a shared learnable vector, denoted as a mask token \\(m\\). The dimensions of \\(m\\) and the replaced point features vary for different tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.468, + 0.411 + ], + "angle": 0, + "content": "Voxel encoding and decoding. All the non-empty voxels are then encoded by a voxel encoder (VE) and the SST backbone. As the features of masked voxels are also extracted by the backbone, we only need a lightweight task-specific MLP head to decode and predict their original information [36]. The loss is computed only on masked voxels, following previous practices [15, 36]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.422, + 0.332, + 0.438 + ], + "angle": 0, + "content": "3.2. Masked Voxel Jigsaw (MVJ)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.468, + 0.522 + ], + "angle": 0, + "content": "Object detection in LiDAR point clouds necessitates a representation capable of capturing contextual information, such as the distribution and relationships among voxels. We propose a jigsaw puzzle that obscures position information, compelling the model to learn these relationships." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.468, + 0.688 + ], + "angle": 0, + "content": "Masking position information. Given a position-masked voxel, we replace the absolute coordinates \\( x_{j}, y_{j}, z_{j} \\) of each point with a mask token \\( m_{v} \\in \\mathbb{R}^{3} \\), while preserving the local shape by retaining the decorated coordinates of each point within the voxel (Fig. 2). The masked voxel is then represented as \\( V_{\\mathrm{masked}} = \\{p_{j} = [m_{v}, x_{j}^{c}, y_{j}^{c}, z_{j}^{c}, x_{j}^{v}, y_{j}^{v}, z_{j}^{v}]^{T} \\in \\mathbb{R}^{9}\\}_{j=1\\dots T} \\). All voxels are subsequently input to the voxel encoder and the SST backbone. Notably, positional embeddings are not added to the voxels before being fed to the backbone, as the model is tasked with predicting position information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Prediction target and loss function. Instead of directly predicting the absolute position of the masked voxels, we formulate a classification task to facilitate optimization. Owing to the extensive range of LiDAR point clouds, the SST backbone partitions the 3D voxelized grid into windows, akin to the Swin-transformer [22], to enable efficient attention calculation within windows. We employ the same window partitioning method, requiring the model to predict only the relative index of the masked voxel within its corresponding window. Assuming a 3D window contains \\( N_x, N_y, N_z \\) voxels along the \\( x, y, z \\) axes and the voxel coordinates of a masked voxel are \\( X, Y, Z \\), the relative coordinates \\( I_x, I_y, I_z \\) of the masked voxel with respect to its window are calculated as \\( I_x = X \\mod N_x, I_y = Y \\mod N_y \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "mod \\(N_{y}, I_{z} = Z \\mod N_{z}\\). The relative index is then given by \\(I = I_{x} + I_{y}N_{x} + I_{z}N_{x}N_{y}\\). The prediction head outputs a classification vector \\(\\hat{v} \\in \\mathbb{R}^{N_xN_yN_z}\\) representing the probabilities of the masked voxel occupying each position. The prediction loss for all masked voxel is calculated using the Cross-Entropy loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.189, + 0.892, + 0.232 + ], + "angle": 0, + "content": "\\[\nL _ {M V J} = \\frac {1}{R _ {p}} \\sum_ {i = 1} ^ {R _ {p}} \\text {C r o s s E n t r o p y} (\\hat {v} _ {i}, I _ {i}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.237, + 0.879, + 0.253 + ], + "angle": 0, + "content": "where \\(R_{p}\\) denotes the number of position-masked voxels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.26, + 0.824, + 0.275 + ], + "angle": 0, + "content": "3.3. Masked Voxel Reconstruction (MVR)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.892, + 0.358 + ], + "angle": 0, + "content": "To incorporate local shape information within each voxel, we introduce masked voxel reconstruction (MVR). MVR masks the absolute and relative coordinates of all points in shape-masked voxels to learn point distributions, preserving only one point to provide the voxel position." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.51 + ], + "angle": 0, + "content": "Masking voxel shapes. In a given voxel, both the original and decorated coordinates of points represent the local shape of the voxel (Fig. 2). We employ a shared learnable token \\( m_p \\in \\mathbb{R}^9 \\) to replace all point features in every shape-masked voxel, except for one point. This particular point conveys the voxel's positional information without revealing the voxel shape, which would render the reconstruction task trivial. Consequently, the masked voxel becomes \\( V_{\\mathrm{masked}} = \\{[x_1, y_1, z_1, x_1^c, y_1^c, z_1^c, x_1^v, y_1^v, z_1^v]\\} \\cup \\{p_j = m_p\\}_{j=2\\dots T} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.736 + ], + "angle": 0, + "content": "Reconstruction target and loss function. We reconstruct each masked voxel and use the \\(L2\\) Chamfer Distance loss to measure the discrepancy between the reconstructed and target point distributions. This loss function is insensitive to point density [34], which is crucial since each voxel may contain varying numbers of points. Assuming each reconstructed voxel contains \\(n\\) points, the reconstruction head outputs a vector containing the coordinates of each point in the masked voxel, \\(\\hat{v} \\in \\mathbb{R}^{3n}\\). To pre-train the model more stably, we normalize the target point distribution using each point's distance to the voxel center as ground truth and scale the distance between 0 and 1. Supposing each reconstructed masked voxel is \\(\\hat{V} = \\{\\hat{p}_j = [\\hat{x}_j, \\hat{y}_j, \\hat{z}_j]\\}_{j=1\\dots n}\\) and the ground truth is \\(V = \\{p_j = [x_j^v, y_j^v, z_j^v]\\}_{j=1\\dots t_i}\\), the total loss across all masked voxels is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.742, + 0.892, + 0.782 + ], + "angle": 0, + "content": "\\[\nL _ {M V R} = \\frac {1}{R _ {s}} \\sum_ {i = 1} ^ {R _ {s}} L _ {C D} \\left(\\hat {V} _ {i}, V _ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.891, + 0.818 + ], + "angle": 0, + "content": "where \\(R_{s}\\) represents the number of shape-masked voxels and \\(L_{CD}\\) is the \\(L2\\) Chamfer Distance loss calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.824, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {C D} = \\frac {1}{\\left| \\hat {V} _ {i} \\right|} \\sum_ {\\hat {p} \\in \\hat {V} _ {i}} \\min _ {p \\in V _ {i}} \\| \\hat {p} - p \\| _ {2} ^ {2} \\tag {3} \\\\ + \\frac {1}{| V _ {i} |} \\sum_ {p \\in V _ {i}} \\min _ {\\hat {p} \\in \\hat {V} _ {i}} \\| p - \\hat {p} \\| _ {2} ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13448" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "3.4. Joint Pre-Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.471, + 0.265 + ], + "angle": 0, + "content": "We jointly pre-train the model using both MVJ and MVR tasks to enable the model to learn both point and voxel distributions. We use R-FVS to sample \\(R_{s}\\) voxels for MVR and \\(R_{p}\\) voxels for MVJ. As illustrated in Fig. 2, the voxel encoder and the backbone extract features from both types of masked voxels and unmasked voxels. Masked features are decoded to recover their corresponding targets, and Cross-Entropy loss and Chamfer Distance loss are utilized for supervision. The total loss for joint pre-training is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.277, + 0.469, + 0.293 + ], + "angle": 0, + "content": "\\[\nL = \\alpha L _ {M V J} + \\beta L _ {M V R} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.304, + 0.354, + 0.319 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are balancing coefficients." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.332, + 0.418, + 0.35 + ], + "angle": 0, + "content": "4. Data-Efficient Benchmark on Waymo" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.469, + 0.447 + ], + "angle": 0, + "content": "To evaluate the pre-trained representation, we transfer the weights of the voxel encoder and the backbone for object detection. Previous benchmarks fail to effectively reveal the effectiveness of pre-training strategies due to data diversity issues. We propose a new data-efficient benchmark to address these limitations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.469, + 0.584 + ], + "angle": 0, + "content": "Observation: Incomplete model convergence. Prior works [20, 40] fine-tune the model on different splits with the same epochs, causing reduced iteration numbers when dataset sizes decrease. As illustrated in Fig. 3a, by training for more epochs (but the same iterations), the performance of SST significantly improves, suggesting that the models may not fully converge on smaller splits when training epochs remain constant. This entangled factor prevents accurate evaluation of pre-training strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Issue: Fine-Tuning splits with similar data diversities. With the same training iterations, models trained on different splits display comparable performance, despite significant differences in data amounts (Fig. 3a). We find that scene diversity is a crucial factor. Previous works [20, 40] uniformly sample the entire Waymo training set to create different fine-tuning splits, for instance, sampling one frame from every two consecutive frames for a \\(50\\%\\) split. However, neighboring frames in self-driving datasets are similar due to the short time difference between consecutive frames (e.g., only \\(0.1s\\) in Waymo [30]). This results in similar scene diversity across splits, leading to comparable detection performance using the same training iterations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We argue that such an experimental setting does not accurately represent a typical application scenario of self-supervised pre-training, where annotated fine-tuning data is less abundant and diverse compared to pre-training data. Furthermore, it fails to evaluate how pre-training benefits downstream tasks when varying amounts and diversities of labeled data are available. We provide an additional example in the supplementary material to further illustrate" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.693, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.192, + 0.631, + 0.204 + ], + "angle": 0, + "content": "(a) Existing" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.891, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.192, + 0.819, + 0.204 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.216, + 0.894, + 0.287 + ], + "angle": 0, + "content": "Figure 3. Comparison of data-efficient benchmarks. Previous uniformly sampled splits exhibit similar diversity, as evidenced by similar performance with the same training iterations. Our benchmark features varying data amounts and diversities, ensuring model convergence. Dashed lines show baseline performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.357 + ], + "angle": 0, + "content": "these issues with previous benchmarks. Our proposed data-efficient benchmark addresses these issues, offering a more precise evaluation of pre-training strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.893, + 0.568 + ], + "angle": 0, + "content": "Solution: Sequence-Based data sampling. We sample fine-tuning data using different scene sequences rather than uniformly sampling varying numbers of frames from identical sequences. The Waymo training split comprises 798 distinct scene sequences, each lasting \\(20s\\) with approximately 200 frames [30]. We randomly sample \\(5\\%\\), \\(10\\%\\), \\(20\\%\\), \\(50\\%\\) of the scene sequences, respectively, using all frames within the sampled sequences for fine-tuning. Consequently, data diversities and amounts vary across each split. Each larger split also encompasses the smaller ones, enabling measurement of performance changes with additional fine-tuning data. Furthermore, we randomly sample three \\(5\\%\\) and \\(10\\%\\) splits and report average results to reduce variance on these smaller sets (see our supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.69 + ], + "angle": 0, + "content": "For different fine-tuning splits, we train the model until performance saturates (i.e., overfitting) and determine the number of epochs used for experiments accordingly. For example, for \\(5\\%\\) data, we try \\(\\{48, 60, 66, 72, 78, 84\\}\\) epochs and ultimately select 72. The baseline performances across various training iterations are shown in Fig. 3b. With our new benchmark, baseline performance and convergence speed vary as anticipated for different splits." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.702, + 0.634, + 0.719 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.707, + 0.743 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. The Waymo dataset [30] consists of 1150 self-driving scenes with 798 scenes as the training split, 202 scenes as the validation split, and 150 scenes as the testing split. We pre-train on the training split, which contains 158,240 annotated frames. For the downstream data-efficient 3D object detection task, we fine-tune the model using our sampled splits (as described in Sec. 4) and evaluate performance on the standard Waymo validation split with 3 classes (cars, pedestrians, and cyclists) and two difficulty levels (L1 and L2). We adopt L2 mean average preci" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13449" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Data-efficient 3D object detection results of SST on the Waymo validation set. SST is pre-trained using various self-supervised methods on Waymo training split, with weights applied for the downstream detection task. \"Random initialization\" denotes training from scratch. The model performances are shown using different amounts and diversities of fine-tuning data." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.143, + 0.875, + 0.445 + ], + "angle": 0, + "content": "
Data amountInitializationOverallCarPedestrianCyclist
L2 mAPL2 mAPHL2 mAPL2 mAPHL2 mAPL2 mAPHL2 mAPL2 mAPH
5%Random44.4140.3451.0150.4952.7442.2629.4928.27
PointContrast [35]45.3241.3052.1251.6153.6843.2230.1629.09
ProposalContrast [40]46.6242.5852.6752.1954.3143.8232.8731.72
MV-JAR (Ours)50.52+6.1146.68+6.3456.4756.0157.6547.6937.4436.33
10%Random54.3150.4654.8454.3760.5550.7147.5546.29
PointContrast [35]53.6949.9454.7654.3059.7550.1246.5745.39
ProposalContrast [40]53.8950.1355.1854.7160.0150.3946.4845.28
MV-JAR (Ours)57.44+3.1354.06+3.6058.4358.0063.2854.6650.6349.52
20%Random60.1656.7858.7958.3565.6357.0456.0754.94
PointContrast [35]59.3555.7858.6458.1864.3955.4355.0253.73
ProposalContrast [40]59.5255.9158.6958.2264.5355.4555.3654.07
MV-JAR (Ours)62.28+2.1159.15+2.3761.8861.4566.9859.0257.9857.00
50%Random66.4363.3663.8163.3870.7863.0564.7163.66
PointContrast [35]65.5162.2162.6662.2369.8261.5364.0462.86
ProposalContrast [40]65.7662.4962.9362.5070.0961.8664.2663.11
MV-JAR (Ours)66.70+0.2763.69+0.3364.3063.8971.1463.5764.6563.63
100%Random68.5065.5464.9664.5672.3864.8968.1767.17
PointContrast [35]68.0664.8464.2463.8271.9263.8168.0366.89
ProposalContrast [40]68.1765.0164.4264.0071.9463.9468.1667.10
MV-JAR (Ours)69.16+0.6666.20+0.6665.5265.1272.7765.2869.1968.20
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.47, + 0.529 + ], + "angle": 0, + "content": "sion (L2 mAP) and L2 mean average precision with heading (L2 mAPH) as main evaluation metrics. We also evaluate the model on the KITTI dataset [14] for the detection task. It contains 7481 training samples and 7518 test samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.789 + ], + "angle": 0, + "content": "Implementation details. We employ the official implementation and training settings of SST [12] for pre-training and fine-tuning unless specified otherwise. The Waymo dataset's point cloud range is \\( W \\times H \\times D = 149.76m \\times 149.76m \\times 6m \\), and the voxel size is \\( v_{W} \\times v_{H} \\times v_{D} = 0.32m \\times 0.32m \\times 6m \\). For MVR, we predict 15 points per voxel with a masking ratio of 0.05. For MVJ, the window size is \\( N_{x} \\times N_{y} \\times N_{z} = 12 \\times 12 \\times 1 \\) and the masking ratio is 0.1. We set the loss weight \\( \\alpha = \\beta = 1 \\), pre-train for 6 epochs with an initial learning rate of \\( 5e - 6 \\). When fine-tuning, we initialize the voxel encoder and SST backbone with pre-trained weights, and other training hyperparameters remain unchanged. We set KITTI's point cloud range to \\( 69.12m \\times 79.36m \\times 4m \\), voxel size to \\( 0.32m \\times 0.32m \\times 4m \\), and train for 160 epochs. We use the AdamW optimizer with a batch size of 8 and a cyclic learning rate scheduler with cosine annealing for all training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.303, + 0.817 + ], + "angle": 0, + "content": "5.2. Main Results on Waymo" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Baselines. We fine-tune the model using data splits described in Sec. 4. For \\(\\{5\\%, 10\\%, 20\\%, 50\\%, 100\\}\\) data, the model is trained for \\(\\{72, 60, 42, 36, 24\\}\\) epochs. We use the randomly initialized model's performances as our baselines. We also include experiments with models using" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.869, + 0.484 + ], + "angle": 0, + "content": "convolutional backbones in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.683 + ], + "angle": 0, + "content": "Results with different initialization. Tab. 1 presents the SST model's performances with various initializations on our proposed data-efficient benchmark. Our pre-training method consistently improves SST baselines across all finetuning data amounts, especially when data is scarce. Our method achieves up to \\(15.7\\%\\) relative improvement, from \\(40.34\\%\\) L2 mAPH to \\(46.68\\%\\) L2 mAPH, when using only \\(5\\%\\) fine-tuning data. The performance gain exists across object categories, indicating effective general representation learning. As more fine-tuning data is introduced, the performance gain persists, and when fine-tuning with the entire Waymo training set, our pre-training enhances detection performance from \\(65.54\\%\\) mAPH to \\(66.20\\%\\) mAPH." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.893, + 0.883 + ], + "angle": 0, + "content": "We note that with \\(50\\%\\) and \\(100\\%\\) fine-tuning data, the improvements from our pre-training are not as significant as with less data. We observe that the baseline result on \\(50\\%\\) data (63.36% mAPH) is comparable to that on \\(100\\%\\) data (65.54% mAPH), suggesting that training data is no longer the bottleneck. The original SST, with only 2.1M parameters, is a lightweight network. We scale up SST to 8.3M parameters by using larger hidden layer channels and observe larger benefits from pre-training on the \\(50\\%\\) fine-tuning split, from \\(63.13\\%\\) L2 mAPH to \\(64.24\\%\\) L2 mAPH. Therefore, we argue that the SST model's capacity, rather than the fine-tuning data amount, becomes the bottleneck for detection performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Comparisons with contrastive learning methods. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13450" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.132 + ], + "angle": 0, + "content": "Table 2. Transferring results on the KITTI validation split. We pre-train SST with the Waymo training split and fine-tune with the KITTI training split." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.142, + 0.409, + 0.234 + ], + "angle": 0, + "content": "
InitializationOverall
EasyMod.Hard
Random74.7163.4360.00
PointContrast [35]73.3562.5359.01
ProposalContrast [40]73.6363.3459.40
MV-JAR (Ours)75.2263.8060.35
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.501 + ], + "angle": 0, + "content": "utilize the official implementation of the recently proposed ProposalContrast [40] and reimplement PointContrast [35] for comparison. As demonstrated in Tab. 1, both contrastive learning methods offer benefits only with \\(5\\%\\) fine-tuning data, and their improvements diminish with higher baseline performances. This may be due to the difficulty in smoothly learning representations. A key challenge in LiDAR-based contrastive learning is constructing matched pairs. The point pairs used by PointContrast face the issue of spatially adjacent points with similar features being assigned as negative samples. We observe difficulty in convergence when sampling excessive pairs. ProposalContrast alleviates the issue by contrasting regions but does not entirely resolve it. Alternatively, our method offers a new avenue for developing self-supervised pre-training through masked voxel modeling on LiDAR point clouds." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.511, + 0.405, + 0.527 + ], + "angle": 0, + "content": "5.3. Pre-training Accelerates Convergence" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.763 + ], + "angle": 0, + "content": "To investigate how MV-JAR pre-training accelerates convergence, we fine-tune the SST model on the entire Waymo training split. We train the SST model for different epochs and report the performance of both fine-tuning and training from scratch. As illustrated in Fig. 1, MV-JAR pre-training significantly enhances the convergence speed of SST. By fine-tuning for only 3 epochs, the SST model achieves \\(62.48\\%\\) L2 mAPH. Remarkably, fine-tuning for just 12 epochs results in a performance comparable to training from scratch for 24 epochs, effectively halving the convergence time. As the number of fine-tuning epochs increases, the performance gain provided by MV-JAR does not wane. Pre-training can still improve the model's converged performance by \\(0.74\\%\\) L2 mAPH, highlighting the superiority of the pre-trained representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.355, + 0.788 + ], + "angle": 0, + "content": "5.4. Transferring Results on KITTI" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.903 + ], + "angle": 0, + "content": "To evaluate the transferability of the learned representation, we fine-tune various pre-trained SST models on the KITTI dataset. The original SST paper [12] did not conduct experiments on the KITTI dataset; thus, we follow the training schedule and setup of SST and PointPillars [19] within the MMDetection3D framework [9]. As demonstrated in Tab. 2, the performance gain from our MV-JAR pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "persists when the representation is transferred to a different domain, indicating that the model learns a generic representation through pre-training. While the KITTI training samples account for approximately \\(5\\%\\) of the entire Waymo training split, the relative improvement is much smaller than transferring to the \\(5\\%\\) Waymo split. We hypothesize that this may be attributed to the domain gap between the Waymo and KITTI datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.225, + 0.665, + 0.24 + ], + "angle": 0, + "content": "5.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.892, + 0.354 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to investigate the effectiveness of our proposed Reversed-Furthest-Voxel-Sampling (R-FVS) strategy. Additionally, we explore the optimal masking ratio and the impact of MVR and MVJ pre-training. All pre-training experiments utilize the entire Waymo training split, and we fine-tune the model on one of our \\(5\\%\\) data splits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Mask sampling strategy. Our R-FVS mask sampling strategy masks voxels that are not sampled by FPS, aiming to avoid masking voxels located in sparse regions. We also examine random sampling and an opposing sampling strategy called FVS, which masks voxels sampled by FPS. With FVS sampling, regions with sparse points are more likely to be masked, leading to greater information loss compared to random sampling and R-FVS sampling. Tab. 3 presents the fine-tuning results of pre-training with the three sampling strategies. R-FVS sampling performs the best, while FVS performs the worst. These findings confirm that minimizing information loss benefits pre-training and validate the effectiveness of our R-FVS strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Masking ratio. Tab. 4 illustrates the impact of different masking ratio combinations. Specifically, MVR with a 0.1 masking ratio and MVJ with a 0.05 masking ratio yield the best results. With this combination, the overall masking ratio is 0.15, which is significantly lower than the masking ratios in the image [15, 36] or video domain [13, 32]. A high masking ratio in masked autoencoders typically indicates information redundancy [13, 15, 32, 36]. We postulate that two factors contribute to our low masking ratio: 1) LiDAR point clouds represent a vast space where each meaningful object occupies only a small fraction. The information on each object is not highly redundant, especially when the objects are distant from the sensor. This contrasts with images in datasets like ImageNet, where objects often encompass a significant portion of the image. 2) The SST model calculates attention only within a partitioned window, which restricts the information used. We leave this question for future exploration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Pre-training effects of MVR and MVJ. We investigate the individual pre-training effects of MVR and MVJ by optimizing the masking ratio for each task and reporting the fine-tuning performances in Tab. 5. Both MVR and MVJ pre-training enhance performance compared to random ini" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13451" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.481, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.094, + 0.89, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.349, + 0.848, + 0.364 + ], + "angle": 0, + "content": "Figure 4. Visualization of the reconstruction results of MV-JAR on the Waymo validation set, compared with ground truths." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.383, + 0.325, + 0.411 + ], + "angle": 0, + "content": "Table 3. Ablation study on mask sampling strategies." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.422, + 0.314, + 0.502 + ], + "angle": 0, + "content": "
Mask samplingOverall
L2 mAPL2 mAPH
FVS49.7345.88
Random49.8646.06
R-FVS50.5246.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.36, + 0.379, + 0.614, + 0.393 + ], + "angle": 0, + "content": "Table 4. Ablation study on masking ratios." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.403, + 0.603, + 0.507 + ], + "angle": 0, + "content": "
Masking ratioOverall
MVJMVRL2 mAPL2 mAPH
0.30.149.7446.05
0.050.149.6745.92
0.10.149.8545.97
0.10.0550.5246.68
0.10.350.2546.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.649, + 0.383, + 0.897, + 0.41 + ], + "angle": 0, + "content": "Table 5. Ablation study on pre-training effects of MVR and MVJ." + }, + { + "type": "table", + "bbox": [ + 0.68, + 0.422, + 0.871, + 0.501 + ], + "angle": 0, + "content": "
TaskOverall
L2 mAPL2 mAPH
MVR46.5742.73
MVJ49.9646.10
MV-JAR50.5246.68
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.681 + ], + "angle": 0, + "content": "tialization, with MVJ outperforming MVR. This finding suggests that capturing voxel distributions plays a more significant role in representation learning, which is consistent with the fact that LiDAR detectors process points by down-sampling them into voxels. Our integrated MV-JAR method demonstrates further improvement over MVR and MVJ, validating the effectiveness of jointly capturing point and voxel distributions. Additionally, we offer an analysis of improvements across various distances in our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.214, + 0.71 + ], + "angle": 0, + "content": "5.6. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Figure 4 displays the MV-JAR reconstruction results on the Waymo validation set, including MVJ and MVR outcomes. In our pre-training, MVJ achieves approximately \\(89\\%\\) classification accuracy of masked voxels on the validation set, accurately reconstructing masked voxels in the correct positions most of the time. MVR can capture the point distributions within voxels, evidenced by reconstructed ground points aligning with ground circles. However, MVR struggles to capture detailed distributions, likely due to the discrete sampling of LiDAR points from continuous surfaces, resulting in considerable variability in the data and increased difficulty in capturing finer details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.627, + 0.545 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.819 + ], + "angle": 0, + "content": "In this paper, we introduce the Masked Voxel Jigsaw and Reconstruction (MV-JAR) pre-training method for LiDAR detectors. MV-JAR captures both point and voxel distributions of LiDAR point clouds, enabling models to learn effective and generic representations. We also develop a Reversed-Furthest-Voxel-Sampling strategy to address the uneven distribution of LiDAR points. Comprehensive experiments on the Waymo and KITTI datasets show that our method consistently and significantly enhances the detector's performance across different data scale regimes. MV-JAR offers a promising alternative for LiDAR-based self-supervised pre-training through Masked Voxel Modeling. Additionally, we establish a new data-efficient benchmark on the Waymo dataset, incorporating fine-tuning splits with diverse data variations. This benchmark effectively assesses the impact of pre-training on downstream tasks with varying amounts and diversities of labeled data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This project is funded in part by the Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd. under the Innovation and Technology Commission (ITC)'s InnoHK." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13452" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.468, + 0.144 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv:2106.08254, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.147, + 0.468, + 0.286 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.29, + 0.468, + 0.358 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. arXiv:1903.11027, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.362, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.421, + 0.468, + 0.476 + ], + "angle": 0, + "content": "[5] Runjian Chen, Yao Mu, Runsen Xu, Wenqi Shao, Chenhan Jiang, Hang Xu, Yu Qiao, Zhenguo Li, and Ping Luo. CO3: Cooperative unsupervised 3d representation learning for autonomous driving. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.479, + 0.468, + 0.521 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.468, + 0.593 + ], + "angle": 0, + "content": "[7] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv:2202.03026, 2022.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.597, + 0.468, + 0.638 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.642, + 0.468, + 0.696 + ], + "angle": 0, + "content": "[9] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.468, + 0.754 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.468, + 0.841 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[12] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[13] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv:2205.09113, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[17] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[18] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[20] Hanxue Liang, Chenhan Jiang, Dapeng Feng, Xin Chen, Hang Xu, Xiaodan Liang, Wei Zhang, Zhenguo Li, and Luc Van Gool. Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In CVPR, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[21] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[23] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[24] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[25] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[26] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[28] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[29] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13453" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In CVPR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[31] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[32] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.294, + 0.469, + 0.334 + ], + "angle": 0, + "content": "[33] Tai Wang, Xinge Zhu, and Dahua Lin. Reconfigurable voxels: A new representation for lidar-based point clouds. In CoRL, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.337, + 0.469, + 0.391 + ], + "angle": 0, + "content": "[34] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.395, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[35] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[37] Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, Christoph Feichtenhofer, et al. Masked autoencoders that listen. arXiv:2207.06405, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.469, + 0.582 + ], + "angle": 0, + "content": "[38] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[39] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[40] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV,2022.1,2,5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[41] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[42] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[43] Shuangfei Zhai, Navdeep Jaitly, Jason Ramapuram, Dan Busbridge, Tatiana Likhomanenko, Joseph Yitan Cheng, Walter Talbott, Chen Huang, Hanlin Goh, and Joshua Susskind. Position prediction as an effective pretraining strategy. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[44] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv:2205.14401, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[45] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. Dense siamese network for dense unsupervised learning. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[46] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[47] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv:2111.07832, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[48] Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image analysis. arXiv:2203.05573, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[49] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[50] Xinge Zhu, Yuexin Ma, Tai Wang, Yan Xu, Jianping Shi, and Dahua Lin. Ssn: Shape signature networks for multi-class object detection from point clouds. In ECCV, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13454" + } + ] +] \ No newline at end of file diff --git a/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_origin.pdf b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..37d978de7489ad460621220e987a35bd2e4c4d2c --- /dev/null +++ b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/12c44818-ff06-4c0a-af41-417eda66e48c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22f98e40f3b9bf8cc848958bfd06b56e3d6e6d1781d1adfed8d46522c3e484a4 +size 392710 diff --git a/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/full.md b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..438b49a23c10192142add55045fff348f6e82dff --- /dev/null +++ b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/full.md @@ -0,0 +1,274 @@ +# MV-JAR: Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training + +Runsen Xu $^{1,2}$ Tai Wang $^{1,2}$ Wenwei Zhang $^{3,2}$ Runjian Chen $^{4}$ Jinkun Cao $^{5}$ Jiangmiao Pang $^{2\boxtimes}$ Dahua Lin $^{1,2}$ + +1The Chinese University of Hong Kong 2Shanghai AI Laboratory 3S-Lab, NTU 4The University of Hong Kong 5Carnegie Mellon University + +{runsenxu,wt019,dhlin}@ie.cuhk.edu.hk, wenwei001@ntu.edu.sg, rjchen@connect.hku.hk, jinkunc@andrew.cmu.edu, pangjiangmiao@gmail.com + +# Abstract + +This paper introduces the Masked Voxel Jigsaw and Reconstruction (MV-JAR) method for LiDAR-based self-supervised pre-training and a carefully designed data-efficient 3D object detection benchmark on the Waymo dataset. Inspired by the scene-voxel-point hierarchy in downstream 3D object detectors, we design masking and reconstruction strategies accounting for voxel distributions in the scene and local point distributions within the voxel. We employ a Reversed-Furthest-Voxel-Sampling strategy to address the uneven distribution of LiDAR points and propose MV-JAR, which combines two techniques for modeling the aforementioned distributions, resulting in superior performance. Our experiments reveal limitations in previous data-efficient experiments, which uniformly sample fine-tuning splits with varying data proportions from each LiDAR sequence, leading to similar data diversity across splits. To address this, we propose a new benchmark that samples scene sequences for diverse fine-tuning splits, ensuring adequate model convergence and providing a more accurate evaluation of pre-training methods. Experiments on our Waymo benchmark and the KITTI dataset demonstrate that MV-JAR consistently and significantly improves 3D detection performance across various data scales, achieving up to a $6.3\%$ increase in mAPH compared to training from scratch. Codes and the benchmark are available at https://github.com/SmartBot-PJLab/MV-JAR. + +# 1. Introduction + +Self-supervised pre-training has gained considerable attention, owing to its exceptional performance in visual representation learning. Recent advancements in contrastive + +![](images/bc99e3383e48fa7a358cd24437bab93ddf8278ebaef1bf496a601e560e915666.jpg) +Figure 1. 3D object detection results on the Waymo dataset. Our MV-JAR pre-training accelerates model convergence and greatly improves the performance with limited fine-tuning data. + +![](images/27eed56131e7787e728f4124f47081fe7fb9186f4e67cde9a5c4e39ba0cb321c.jpg) + +learning [4, 6, 8, 16, 45] and masked autoencoders [1, 7, 15, 36, 47] for images have sparked interest among researchers and facilitated progress in modalities such as point clouds. + +However, LiDAR point clouds differ from images and dense point clouds obtained by reconstruction as they are naturally sparse, unorganized, and irregularly distributed. Developing effective self-supervised proxy tasks for these unique properties remains an open challenge. Constructing matching pairs for contrastive learning in geometry-dominant scenes is more difficult [20, 40], as points or regions with similar geometry may be assigned as negative samples, leading to ambiguity during training. To address this, our study explores masked voxel modeling paradigms for effective LiDAR-based self-supervised pre-training. + +Downstream LiDAR-based 3D object detectors [12, 19, 33, 38, 41, 50] typically quantize the 3D space into voxels and encode point features within them. Unlike pixels, which are represented by RGB values, the 3D space presents a scene-voxel-point hierarchy, introducing new + +challenges for masked modeling. Inspired by this, we design masking and reconstruction strategies that consider voxel distributions in the scene and local point distributions in the voxel. Our proposed method, Masked Voxel Jigsaw And Reconstruction (MV-JAR), harnesses the strengths of both voxel and point distributions to improve performance. + +To account for the uneven distribution of LiDAR points, we first employ a Reversed-Furthest-Voxel-Sampling (R-FVS) strategy that samples voxels to mask based on their sparseness. This approach prevents masking the furthest distributed voxels, thereby avoiding information loss in regions with sparse points. To model voxel distributions, we propose Masked Voxel Jigsaw (MVJ), which masks the voxel coordinates while preserving the local shape of each voxel, enabling scene reconstruction akin to solving a jigsaw puzzle. For modeling local point distributions, we introduce Masked Voxel Reconstruction (MVR), which masks all coordinates of points within the voxel but retains one point as a hint for reconstruction. Combining these two methods enhances masked voxel modeling. + +Our experiments indicate that existing data-efficient experiments [20, 40] inadequately evaluate the effectiveness of various pre-training methods. The current benchmarks, which uniformly sample frames from each data sequence to create diverse fine-tuning splits, exhibit similar data diversity due to the proximity of neighboring frames in a sequence [3, 14, 30]. Moreover, these experiments train models for the same number of epochs across different fine-tuning splits, potentially leading to incomplete convergence. As a result, the benefits of pre-trained representations become indistinguishable across splits once the object detector is sufficiently trained on the fine-tuning data. To address these shortcomings, we propose sampling scene sequences to form diverse fine-tuning splits and establish a new data-efficient 3D object detection benchmark on the Waymo [30] dataset, ensuring sufficient model convergence for a more accurate evaluation. + +We employ the Transformer-based SST [12] as our detector and pre-train its backbone for downstream detection tasks. Comprehensive experiments on the Waymo and KITTI [14] datasets demonstrate that our pre-training method significantly enhances the model's performance and convergence speed in downstream tasks. Notably, it improves detection performance by $6.3\%$ mAPH when using only $5\%$ of scenes for fine-tuning and reduces training time by half when utilizing the entire dataset (Fig. 1). With the representation pre-trained by MV-JAR, the 3D object detectors pre-trained on Waymo also exhibit generalizability when transferred to KITTI. + +# 2. Related Work + +Self-supervised learning for point clouds. Annotating point clouds demands significant effort, necessitating self + +supervised pre-training methods. Prior approaches primarily focus on object CAD models [21, 26, 29, 39, 42, 44] and indoor scenes [17, 35, 46]. Point-BERT [42] applies BERT-like paradigms for point cloud recognition, while Point-MAE [26] reconstructs point patches without the tokenizer. To acquire scene-level representation, PointContrast [35] introduces a contrastive learning approach that compares points in two static partial views of a reconstructed indoor scene. Its successor, SceneContrast [17], incorporates spatial information into the contrastive learning framework. However, LiDAR point clouds are irregular and dynamic, with mainstream LiDAR perception models often exhibiting distinct architectures, which obstructs the direct adaptation of object-level and indoor scene pre-training methods. STRL [18] employs contrastive learning with two temporally-correlated frames for spatiotemporal representation. GCC-3D [20] presents a framework that combines geometry-aware contrast and pseudo-instance clustering harmonization. ProposalContrast [40] targets region-level contrastive learning to improve 3D detection. CO3 [5] leverages infrastructure-vehicle-cooperation point clouds to construct effective contrastive views. In contrast, our work explores masked voxel modeling, diverging from the prevailing contrastive learning paradigm. + +Masked autoencoders for self-supervised pre-training. Masked language modeling plays a pivotal role in self-supervised pre-training for Transformer-based networks in natural language processing [2, 10, 28]. This approach typically masks portions of the input and pre-trains networks to predict the original information. With the successful integration of Transformers into computer vision [11], researchers have increasingly focused on masked image modeling [1, 7, 15, 36, 43, 47] to mitigate the data-intensive issue of ViT [11]. BEiT [1] employs a tokenizer to generate discrete tokens for image patches and utilizes a BERT-like framework to pre-train ViT. MAE [15] introduces an asymmetric autoencoder for reconstructing RGB pixels of original images, eliminating the need for an extra tokenizer. SimMiM [36] encodes masked raw images with Transformers and employs a lightweight prediction head for recovery. Demonstrating considerable potential in image recognition, masked autoencoders have been applied in video understanding [13, 32], medical image analysis [48], and audio processing [37]. In this work, we investigate the application of this powerful pre-training technique to LiDAR point clouds using voxel representation. + +Transformer-based 3D object detection. The recent success of Vision Transformers [11] has inspired extensive research into the application of Transformer-based architectures for 3D object detection [12, 23-25, 31]. 3DETR [24] presents an end-to-end object detection framework with modified Transformer blocks for 3D point clouds. Pointformer [25] employs hierarchical Transformer blocks to + +![](images/2f89105dc3364d6ded51795534fdab51ff4bed01f320412b3f85aec31ece80bc.jpg) +Figure 2. Overview of Masked Voxel Jigsaw and Reconstruction (MV-JAR). Our R-FVS sampling method selects a specific ratio of nonempty voxels to mask. For position-masked voxels, we mask the absolute coordinates of all points within the voxel while maintaining the relative coordinates. For shape-masked voxels, we mask both absolute and relative coordinates of all points, preserving only one point. Features are extracted from all non-empty voxels, and masked features are decoded by a dedicated head to recover corresponding targets. + +extract point features for 3D detection. Voxel Transformer [23] develops a Transformer-based 3D backbone to establish long-range relationships between voxels. Recently, SST [12] introduced a single-stride sparse Transformer as a replacement for the PointPillars [19] backbone, abandoning the multi-resolution strategy to enhance the detection of small objects. Emulating the SwinTransformer [22], SST partitions the space into windows and calculates attention only within each window for efficiency. This approach demonstrates impressive detection results, and we adopt it for our experiments. + +# 3. Methodology + +In this section, we first present an overview of a general masked voxel modeling framework for LiDAR point clouds and our Reversed-Furthest-Voxel-Sampling strategy to tackle the uneven distribution of LiDAR points (Sec. 3.1). Masked Voxel Jigsaw (MVJ) (Sec. 3.2) and Masked Voxel Reconstruction (MVR) (Sec. 3.3) are both instantiations of the general framework that learns the voxel distributions and local point distributions, respectively. Masked Voxel Jigsaw and Reconstruction (MV-JAR) consist of these two tasks to synergically learn the data distributions at two levels (Sec. 3.4). + +# 3.1. Masked Voxel Modeling and Sampling + +As illustrated in Fig. 2, a general masked voxel modeling framework converts LiDAR point clouds into voxels and samples a proportion of non-empty voxels using a specific sampling strategy and masking ratio. Various features within the voxel can be masked for a particular learning target in this general framework to instantiate different self-supervised learning tasks. Specifically, task-specific features of the points in the sampled voxels are replaced with + +a learnable vector. Subsequently, all non-empty voxels are encoded by a voxel encoder, processed by a backbone network to extract features, and followed by a lightweight decoder that reconstructs task-specific targets from each voxel feature. Each stage is described in the following sections. In this paper, we employ SST [12], which demonstrates superior performance in 3D detection. The pre-trained weights of the voxel encoder and the SST backbone are utilized as initialization for the downstream 3D detection task. + +Voxelization. Let $p$ denote a point with coordinates $x, y, z$ . Given a point cloud $P = \{p_i = [x_i, y_i, z_i]^T \in \mathbb{R}^3\}_{i=1\dots n}$ containing $n$ points and range $W, H, D$ , we partition the space into a 3D grid using voxel sizes $v_W, v_H, v_D$ along the three axes, grouping points in the same voxel together. As a common technique [12, 19, 49], we decorate each point with $x^c, y^c, z^c, x^v, y^v, z^v$ , where the superscript $c$ denotes distance to the clustering center of all points in each voxel and the superscript $v$ denotes distance to the voxel center. Considering the sparse and irregular nature of LiDAR point clouds, we focus only on non-empty voxels. Assuming non-empty voxels contain at most $T$ points and a total of $N$ non-empty voxels, we obtain the voxel-based point cloud representation as $M = \{V_i\}_{i=1\dots N}$ , where $V_i = \{p_j = [x_j, y_j, z_j, x_j^c, y_j^c, z_j^c, x_j^v, y_j^v, z_j^v]^T \in \mathbb{R}^9\}_{j=1\dots t_i}$ and $t_i \leq T$ . We assume each non-empty voxel contains $T$ points for simplicity. We also maintain each voxel's coordinate in the 3D grid with $C_i \in \mathbb{R}^3$ . + +Masked voxel sampling. A key component of masked autoencoders is the mask sampling strategy. Unlike 2D pixels that uniformly distribute on the image plane, voxels are unevenly distributed due to the sparse and irregular nature of LiDAR point clouds. Random sampling may result in information loss in regions with sparse points. Inspired by furthest point sampling [27] (FPS), which samples an evenly distributed point set while maintaining the key struc + +ture of the data, we propose a Reversed-Furthest-Voxel-Sampling (R-FVS) scheme. Given a masking ratio $r$ , we apply FPS to select $\lfloor N(1 - r)\rfloor$ voxels from all non-empty voxels according to each voxel's coordinate $C_i$ . These sampled voxels are kept to preserve points in sparse regions and maintain key geometric structures in the data, while the remaining unsampled voxels are masked during training. + +Masking voxels. We directly mask the raw inputs of voxels, providing a unified formulation for the masked voxel modeling framework. For each of the remaining unsampled $\lceil Nr \rceil$ voxels to be masked, we replace some of the original features of each point in the voxel with a shared learnable vector, denoted as a mask token $m$ . The dimensions of $m$ and the replaced point features vary for different tasks. + +Voxel encoding and decoding. All the non-empty voxels are then encoded by a voxel encoder (VE) and the SST backbone. As the features of masked voxels are also extracted by the backbone, we only need a lightweight task-specific MLP head to decode and predict their original information [36]. The loss is computed only on masked voxels, following previous practices [15, 36]. + +# 3.2. Masked Voxel Jigsaw (MVJ) + +Object detection in LiDAR point clouds necessitates a representation capable of capturing contextual information, such as the distribution and relationships among voxels. We propose a jigsaw puzzle that obscures position information, compelling the model to learn these relationships. + +Masking position information. Given a position-masked voxel, we replace the absolute coordinates $x_{j}, y_{j}, z_{j}$ of each point with a mask token $m_{v} \in \mathbb{R}^{3}$ , while preserving the local shape by retaining the decorated coordinates of each point within the voxel (Fig. 2). The masked voxel is then represented as $V_{\mathrm{masked}} = \{p_{j} = [m_{v}, x_{j}^{c}, y_{j}^{c}, z_{j}^{c}, x_{j}^{v}, y_{j}^{v}, z_{j}^{v}]^{T} \in \mathbb{R}^{9}\}_{j=1\dots T}$ . All voxels are subsequently input to the voxel encoder and the SST backbone. Notably, positional embeddings are not added to the voxels before being fed to the backbone, as the model is tasked with predicting position information. + +Prediction target and loss function. Instead of directly predicting the absolute position of the masked voxels, we formulate a classification task to facilitate optimization. Owing to the extensive range of LiDAR point clouds, the SST backbone partitions the 3D voxelized grid into windows, akin to the Swin-transformer [22], to enable efficient attention calculation within windows. We employ the same window partitioning method, requiring the model to predict only the relative index of the masked voxel within its corresponding window. Assuming a 3D window contains $N_x, N_y, N_z$ voxels along the $x, y, z$ axes and the voxel coordinates of a masked voxel are $X, Y, Z$ , the relative coordinates $I_x, I_y, I_z$ of the masked voxel with respect to its window are calculated as $I_x = X \mod N_x, I_y = Y \mod N_y$ . + +mod $N_{y}, I_{z} = Z \mod N_{z}$ . The relative index is then given by $I = I_{x} + I_{y}N_{x} + I_{z}N_{x}N_{y}$ . The prediction head outputs a classification vector $\hat{v} \in \mathbb{R}^{N_xN_yN_z}$ representing the probabilities of the masked voxel occupying each position. The prediction loss for all masked voxel is calculated using the Cross-Entropy loss as follows: + +$$ +L _ {M V J} = \frac {1}{R _ {p}} \sum_ {i = 1} ^ {R _ {p}} \text {C r o s s E n t r o p y} (\hat {v} _ {i}, I _ {i}), \tag {1} +$$ + +where $R_{p}$ denotes the number of position-masked voxels. + +# 3.3. Masked Voxel Reconstruction (MVR) + +To incorporate local shape information within each voxel, we introduce masked voxel reconstruction (MVR). MVR masks the absolute and relative coordinates of all points in shape-masked voxels to learn point distributions, preserving only one point to provide the voxel position. + +Masking voxel shapes. In a given voxel, both the original and decorated coordinates of points represent the local shape of the voxel (Fig. 2). We employ a shared learnable token $m_p \in \mathbb{R}^9$ to replace all point features in every shape-masked voxel, except for one point. This particular point conveys the voxel's positional information without revealing the voxel shape, which would render the reconstruction task trivial. Consequently, the masked voxel becomes $V_{\mathrm{masked}} = \{[x_1, y_1, z_1, x_1^c, y_1^c, z_1^c, x_1^v, y_1^v, z_1^v]\} \cup \{p_j = m_p\}_{j=2\dots T}$ . + +Reconstruction target and loss function. We reconstruct each masked voxel and use the $L2$ Chamfer Distance loss to measure the discrepancy between the reconstructed and target point distributions. This loss function is insensitive to point density [34], which is crucial since each voxel may contain varying numbers of points. Assuming each reconstructed voxel contains $n$ points, the reconstruction head outputs a vector containing the coordinates of each point in the masked voxel, $\hat{v} \in \mathbb{R}^{3n}$ . To pre-train the model more stably, we normalize the target point distribution using each point's distance to the voxel center as ground truth and scale the distance between 0 and 1. Supposing each reconstructed masked voxel is $\hat{V} = \{\hat{p}_j = [\hat{x}_j, \hat{y}_j, \hat{z}_j]\}_{j=1\dots n}$ and the ground truth is $V = \{p_j = [x_j^v, y_j^v, z_j^v]\}_{j=1\dots t_i}$ , the total loss across all masked voxels is given by: + +$$ +L _ {M V R} = \frac {1}{R _ {s}} \sum_ {i = 1} ^ {R _ {s}} L _ {C D} \left(\hat {V} _ {i}, V _ {i}\right), \tag {2} +$$ + +where $R_{s}$ represents the number of shape-masked voxels and $L_{CD}$ is the $L2$ Chamfer Distance loss calculated as: + +$$ +\begin{array}{l} L _ {C D} = \frac {1}{\left| \hat {V} _ {i} \right|} \sum_ {\hat {p} \in \hat {V} _ {i}} \min _ {p \in V _ {i}} \| \hat {p} - p \| _ {2} ^ {2} \tag {3} \\ + \frac {1}{| V _ {i} |} \sum_ {p \in V _ {i}} \min _ {\hat {p} \in \hat {V} _ {i}} \| p - \hat {p} \| _ {2} ^ {2}. \\ \end{array} +$$ + +# 3.4. Joint Pre-Training + +We jointly pre-train the model using both MVJ and MVR tasks to enable the model to learn both point and voxel distributions. We use R-FVS to sample $R_{s}$ voxels for MVR and $R_{p}$ voxels for MVJ. As illustrated in Fig. 2, the voxel encoder and the backbone extract features from both types of masked voxels and unmasked voxels. Masked features are decoded to recover their corresponding targets, and Cross-Entropy loss and Chamfer Distance loss are utilized for supervision. The total loss for joint pre-training is expressed as follows: + +$$ +L = \alpha L _ {M V J} + \beta L _ {M V R} \tag {4} +$$ + +where $\alpha$ and $\beta$ are balancing coefficients. + +# 4. Data-Efficient Benchmark on Waymo + +To evaluate the pre-trained representation, we transfer the weights of the voxel encoder and the backbone for object detection. Previous benchmarks fail to effectively reveal the effectiveness of pre-training strategies due to data diversity issues. We propose a new data-efficient benchmark to address these limitations. + +Observation: Incomplete model convergence. Prior works [20, 40] fine-tune the model on different splits with the same epochs, causing reduced iteration numbers when dataset sizes decrease. As illustrated in Fig. 3a, by training for more epochs (but the same iterations), the performance of SST significantly improves, suggesting that the models may not fully converge on smaller splits when training epochs remain constant. This entangled factor prevents accurate evaluation of pre-training strategies. + +Issue: Fine-Tuning splits with similar data diversities. With the same training iterations, models trained on different splits display comparable performance, despite significant differences in data amounts (Fig. 3a). We find that scene diversity is a crucial factor. Previous works [20, 40] uniformly sample the entire Waymo training set to create different fine-tuning splits, for instance, sampling one frame from every two consecutive frames for a $50\%$ split. However, neighboring frames in self-driving datasets are similar due to the short time difference between consecutive frames (e.g., only $0.1s$ in Waymo [30]). This results in similar scene diversity across splits, leading to comparable detection performance using the same training iterations. + +We argue that such an experimental setting does not accurately represent a typical application scenario of self-supervised pre-training, where annotated fine-tuning data is less abundant and diverse compared to pre-training data. Furthermore, it fails to evaluate how pre-training benefits downstream tasks when varying amounts and diversities of labeled data are available. We provide an additional example in the supplementary material to further illustrate + +![](images/2608633e5dbcc7499a8da46049e8c7601e74992ba225c2fc0fd9c527a902683e.jpg) +(a) Existing + +![](images/59d89ebeb26e2e21bb5ea0b6aae706219fd9f3d8b6e6ee6784cb051e7ec6ca52.jpg) +(b) Ours +Figure 3. Comparison of data-efficient benchmarks. Previous uniformly sampled splits exhibit similar diversity, as evidenced by similar performance with the same training iterations. Our benchmark features varying data amounts and diversities, ensuring model convergence. Dashed lines show baseline performances. + +these issues with previous benchmarks. Our proposed data-efficient benchmark addresses these issues, offering a more precise evaluation of pre-training strategies. + +Solution: Sequence-Based data sampling. We sample fine-tuning data using different scene sequences rather than uniformly sampling varying numbers of frames from identical sequences. The Waymo training split comprises 798 distinct scene sequences, each lasting $20s$ with approximately 200 frames [30]. We randomly sample $5\%$ , $10\%$ , $20\%$ , $50\%$ of the scene sequences, respectively, using all frames within the sampled sequences for fine-tuning. Consequently, data diversities and amounts vary across each split. Each larger split also encompasses the smaller ones, enabling measurement of performance changes with additional fine-tuning data. Furthermore, we randomly sample three $5\%$ and $10\%$ splits and report average results to reduce variance on these smaller sets (see our supplementary material). + +For different fine-tuning splits, we train the model until performance saturates (i.e., overfitting) and determine the number of epochs used for experiments accordingly. For example, for $5\%$ data, we try $\{48, 60, 66, 72, 78, 84\}$ epochs and ultimately select 72. The baseline performances across various training iterations are shown in Fig. 3b. With our new benchmark, baseline performance and convergence speed vary as anticipated for different splits. + +# 5. Experiments + +# 5.1. Experimental Settings + +Datasets. The Waymo dataset [30] consists of 1150 self-driving scenes with 798 scenes as the training split, 202 scenes as the validation split, and 150 scenes as the testing split. We pre-train on the training split, which contains 158,240 annotated frames. For the downstream data-efficient 3D object detection task, we fine-tune the model using our sampled splits (as described in Sec. 4) and evaluate performance on the standard Waymo validation split with 3 classes (cars, pedestrians, and cyclists) and two difficulty levels (L1 and L2). We adopt L2 mean average preci + +Table 1. Data-efficient 3D object detection results of SST on the Waymo validation set. SST is pre-trained using various self-supervised methods on Waymo training split, with weights applied for the downstream detection task. "Random initialization" denotes training from scratch. The model performances are shown using different amounts and diversities of fine-tuning data. + +
Data amountInitializationOverallCarPedestrianCyclist
L2 mAPL2 mAPHL2 mAPL2 mAPHL2 mAPL2 mAPHL2 mAPL2 mAPH
5%Random44.4140.3451.0150.4952.7442.2629.4928.27
PointContrast [35]45.3241.3052.1251.6153.6843.2230.1629.09
ProposalContrast [40]46.6242.5852.6752.1954.3143.8232.8731.72
MV-JAR (Ours)50.52+6.1146.68+6.3456.4756.0157.6547.6937.4436.33
10%Random54.3150.4654.8454.3760.5550.7147.5546.29
PointContrast [35]53.6949.9454.7654.3059.7550.1246.5745.39
ProposalContrast [40]53.8950.1355.1854.7160.0150.3946.4845.28
MV-JAR (Ours)57.44+3.1354.06+3.6058.4358.0063.2854.6650.6349.52
20%Random60.1656.7858.7958.3565.6357.0456.0754.94
PointContrast [35]59.3555.7858.6458.1864.3955.4355.0253.73
ProposalContrast [40]59.5255.9158.6958.2264.5355.4555.3654.07
MV-JAR (Ours)62.28+2.1159.15+2.3761.8861.4566.9859.0257.9857.00
50%Random66.4363.3663.8163.3870.7863.0564.7163.66
PointContrast [35]65.5162.2162.6662.2369.8261.5364.0462.86
ProposalContrast [40]65.7662.4962.9362.5070.0961.8664.2663.11
MV-JAR (Ours)66.70+0.2763.69+0.3364.3063.8971.1463.5764.6563.63
100%Random68.5065.5464.9664.5672.3864.8968.1767.17
PointContrast [35]68.0664.8464.2463.8271.9263.8168.0366.89
ProposalContrast [40]68.1765.0164.4264.0071.9463.9468.1667.10
MV-JAR (Ours)69.16+0.6666.20+0.6665.5265.1272.7765.2869.1968.20
+ +sion (L2 mAP) and L2 mean average precision with heading (L2 mAPH) as main evaluation metrics. We also evaluate the model on the KITTI dataset [14] for the detection task. It contains 7481 training samples and 7518 test samples. + +Implementation details. We employ the official implementation and training settings of SST [12] for pre-training and fine-tuning unless specified otherwise. The Waymo dataset's point cloud range is $W \times H \times D = 149.76m \times 149.76m \times 6m$ , and the voxel size is $v_{W} \times v_{H} \times v_{D} = 0.32m \times 0.32m \times 6m$ . For MVR, we predict 15 points per voxel with a masking ratio of 0.05. For MVJ, the window size is $N_{x} \times N_{y} \times N_{z} = 12 \times 12 \times 1$ and the masking ratio is 0.1. We set the loss weight $\alpha = \beta = 1$ , pre-train for 6 epochs with an initial learning rate of $5e - 6$ . When fine-tuning, we initialize the voxel encoder and SST backbone with pre-trained weights, and other training hyperparameters remain unchanged. We set KITTI's point cloud range to $69.12m \times 79.36m \times 4m$ , voxel size to $0.32m \times 0.32m \times 4m$ , and train for 160 epochs. We use the AdamW optimizer with a batch size of 8 and a cyclic learning rate scheduler with cosine annealing for all training. + +# 5.2. Main Results on Waymo + +Baselines. We fine-tune the model using data splits described in Sec. 4. For $\{5\%, 10\%, 20\%, 50\%, 100\}$ data, the model is trained for $\{72, 60, 42, 36, 24\}$ epochs. We use the randomly initialized model's performances as our baselines. We also include experiments with models using + +convolutional backbones in the supplementary material. + +Results with different initialization. Tab. 1 presents the SST model's performances with various initializations on our proposed data-efficient benchmark. Our pre-training method consistently improves SST baselines across all finetuning data amounts, especially when data is scarce. Our method achieves up to $15.7\%$ relative improvement, from $40.34\%$ L2 mAPH to $46.68\%$ L2 mAPH, when using only $5\%$ fine-tuning data. The performance gain exists across object categories, indicating effective general representation learning. As more fine-tuning data is introduced, the performance gain persists, and when fine-tuning with the entire Waymo training set, our pre-training enhances detection performance from $65.54\%$ mAPH to $66.20\%$ mAPH. + +We note that with $50\%$ and $100\%$ fine-tuning data, the improvements from our pre-training are not as significant as with less data. We observe that the baseline result on $50\%$ data (63.36% mAPH) is comparable to that on $100\%$ data (65.54% mAPH), suggesting that training data is no longer the bottleneck. The original SST, with only 2.1M parameters, is a lightweight network. We scale up SST to 8.3M parameters by using larger hidden layer channels and observe larger benefits from pre-training on the $50\%$ fine-tuning split, from $63.13\%$ L2 mAPH to $64.24\%$ L2 mAPH. Therefore, we argue that the SST model's capacity, rather than the fine-tuning data amount, becomes the bottleneck for detection performance. + +Comparisons with contrastive learning methods. We + +Table 2. Transferring results on the KITTI validation split. We pre-train SST with the Waymo training split and fine-tune with the KITTI training split. + +
InitializationOverall
EasyMod.Hard
Random74.7163.4360.00
PointContrast [35]73.3562.5359.01
ProposalContrast [40]73.6363.3459.40
MV-JAR (Ours)75.2263.8060.35
+ +utilize the official implementation of the recently proposed ProposalContrast [40] and reimplement PointContrast [35] for comparison. As demonstrated in Tab. 1, both contrastive learning methods offer benefits only with $5\%$ fine-tuning data, and their improvements diminish with higher baseline performances. This may be due to the difficulty in smoothly learning representations. A key challenge in LiDAR-based contrastive learning is constructing matched pairs. The point pairs used by PointContrast face the issue of spatially adjacent points with similar features being assigned as negative samples. We observe difficulty in convergence when sampling excessive pairs. ProposalContrast alleviates the issue by contrasting regions but does not entirely resolve it. Alternatively, our method offers a new avenue for developing self-supervised pre-training through masked voxel modeling on LiDAR point clouds. + +# 5.3. Pre-training Accelerates Convergence + +To investigate how MV-JAR pre-training accelerates convergence, we fine-tune the SST model on the entire Waymo training split. We train the SST model for different epochs and report the performance of both fine-tuning and training from scratch. As illustrated in Fig. 1, MV-JAR pre-training significantly enhances the convergence speed of SST. By fine-tuning for only 3 epochs, the SST model achieves $62.48\%$ L2 mAPH. Remarkably, fine-tuning for just 12 epochs results in a performance comparable to training from scratch for 24 epochs, effectively halving the convergence time. As the number of fine-tuning epochs increases, the performance gain provided by MV-JAR does not wane. Pre-training can still improve the model's converged performance by $0.74\%$ L2 mAPH, highlighting the superiority of the pre-trained representation. + +# 5.4. Transferring Results on KITTI + +To evaluate the transferability of the learned representation, we fine-tune various pre-trained SST models on the KITTI dataset. The original SST paper [12] did not conduct experiments on the KITTI dataset; thus, we follow the training schedule and setup of SST and PointPillars [19] within the MMDetection3D framework [9]. As demonstrated in Tab. 2, the performance gain from our MV-JAR pre-training + +persists when the representation is transferred to a different domain, indicating that the model learns a generic representation through pre-training. While the KITTI training samples account for approximately $5\%$ of the entire Waymo training split, the relative improvement is much smaller than transferring to the $5\%$ Waymo split. We hypothesize that this may be attributed to the domain gap between the Waymo and KITTI datasets. + +# 5.5. Ablation Studies + +In this section, we conduct ablation studies to investigate the effectiveness of our proposed Reversed-Furthest-Voxel-Sampling (R-FVS) strategy. Additionally, we explore the optimal masking ratio and the impact of MVR and MVJ pre-training. All pre-training experiments utilize the entire Waymo training split, and we fine-tune the model on one of our $5\%$ data splits. + +Mask sampling strategy. Our R-FVS mask sampling strategy masks voxels that are not sampled by FPS, aiming to avoid masking voxels located in sparse regions. We also examine random sampling and an opposing sampling strategy called FVS, which masks voxels sampled by FPS. With FVS sampling, regions with sparse points are more likely to be masked, leading to greater information loss compared to random sampling and R-FVS sampling. Tab. 3 presents the fine-tuning results of pre-training with the three sampling strategies. R-FVS sampling performs the best, while FVS performs the worst. These findings confirm that minimizing information loss benefits pre-training and validate the effectiveness of our R-FVS strategy. + +Masking ratio. Tab. 4 illustrates the impact of different masking ratio combinations. Specifically, MVR with a 0.1 masking ratio and MVJ with a 0.05 masking ratio yield the best results. With this combination, the overall masking ratio is 0.15, which is significantly lower than the masking ratios in the image [15, 36] or video domain [13, 32]. A high masking ratio in masked autoencoders typically indicates information redundancy [13, 15, 32, 36]. We postulate that two factors contribute to our low masking ratio: 1) LiDAR point clouds represent a vast space where each meaningful object occupies only a small fraction. The information on each object is not highly redundant, especially when the objects are distant from the sensor. This contrasts with images in datasets like ImageNet, where objects often encompass a significant portion of the image. 2) The SST model calculates attention only within a partitioned window, which restricts the information used. We leave this question for future exploration. + +Pre-training effects of MVR and MVJ. We investigate the individual pre-training effects of MVR and MVJ by optimizing the masking ratio for each task and reporting the fine-tuning performances in Tab. 5. Both MVR and MVJ pre-training enhance performance compared to random ini + +![](images/122ade527333146bfb78f1e4479a3233e9ad229d33c00b1ebdc20c6cbff47fa8.jpg) +Figure 4. Visualization of the reconstruction results of MV-JAR on the Waymo validation set, compared with ground truths. + +![](images/8054ab95e8e42dbecbfa41a75bd0d776b159a5be7621c11e4b0fe948951c8d30.jpg) + +Table 3. Ablation study on mask sampling strategies. + +
Mask samplingOverall
L2 mAPL2 mAPH
FVS49.7345.88
Random49.8646.06
R-FVS50.5246.68
+ +Table 4. Ablation study on masking ratios. + +
Masking ratioOverall
MVJMVRL2 mAPL2 mAPH
0.30.149.7446.05
0.050.149.6745.92
0.10.149.8545.97
0.10.0550.5246.68
0.10.350.2546.35
+ +Table 5. Ablation study on pre-training effects of MVR and MVJ. + +
TaskOverall
L2 mAPL2 mAPH
MVR46.5742.73
MVJ49.9646.10
MV-JAR50.5246.68
+ +tialization, with MVJ outperforming MVR. This finding suggests that capturing voxel distributions plays a more significant role in representation learning, which is consistent with the fact that LiDAR detectors process points by down-sampling them into voxels. Our integrated MV-JAR method demonstrates further improvement over MVR and MVJ, validating the effectiveness of jointly capturing point and voxel distributions. Additionally, we offer an analysis of improvements across various distances in our supplementary material. + +# 5.6. Visualization + +Figure 4 displays the MV-JAR reconstruction results on the Waymo validation set, including MVJ and MVR outcomes. In our pre-training, MVJ achieves approximately $89\%$ classification accuracy of masked voxels on the validation set, accurately reconstructing masked voxels in the correct positions most of the time. MVR can capture the point distributions within voxels, evidenced by reconstructed ground points aligning with ground circles. However, MVR struggles to capture detailed distributions, likely due to the discrete sampling of LiDAR points from continuous surfaces, resulting in considerable variability in the data and increased difficulty in capturing finer details. + +# 6. Conclusions + +In this paper, we introduce the Masked Voxel Jigsaw and Reconstruction (MV-JAR) pre-training method for LiDAR detectors. MV-JAR captures both point and voxel distributions of LiDAR point clouds, enabling models to learn effective and generic representations. We also develop a Reversed-Furthest-Voxel-Sampling strategy to address the uneven distribution of LiDAR points. Comprehensive experiments on the Waymo and KITTI datasets show that our method consistently and significantly enhances the detector's performance across different data scale regimes. MV-JAR offers a promising alternative for LiDAR-based self-supervised pre-training through Masked Voxel Modeling. Additionally, we establish a new data-efficient benchmark on the Waymo dataset, incorporating fine-tuning splits with diverse data variations. This benchmark effectively assesses the impact of pre-training on downstream tasks with varying amounts and diversities of labeled data. + +Acknowledgements. This project is funded in part by the Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd. under the Innovation and Technology Commission (ITC)'s InnoHK. + +# References + +[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv:2106.08254, 2021. 1, 2 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2 +[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. arXiv:1903.11027, 2019. 2 +[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 1 +[5] Runjian Chen, Yao Mu, Runsen Xu, Wenqi Shao, Chenhan Jiang, Hang Xu, Yu Qiao, Zhenguo Li, and Ping Luo. CO3: Cooperative unsupervised 3d representation learning for autonomous driving. In ICLR, 2023. 2 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 1 +[7] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv:2202.03026, 2022.1, 2 +[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297, 2020. 1 +[9] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.7 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.2 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[12] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, 2022. 1, 2, 3, 6, 7 + +[13] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv:2205.09113, 2022. 2, 7 +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 6 +[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2, 4, 7 +[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 1 +[17] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2 +[18] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In CVPR, 2021. 2 +[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 1, 3, 7 +[20] Hanxue Liang, Chenhan Jiang, Dapeng Feng, Xin Chen, Hang Xu, Xiaodan Liang, Wei Zhang, Zhenguo Li, and Luc Van Gool. Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In CVPR, 2021. 1, 2, 5 +[21] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 2 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In CVPR, 2021. 3, 4 +[23] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In CVPR, 2021. 2, 3 +[24] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In CVPR, 2021. 2 +[25] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In CVPR, 2021. 2 +[26] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2 +[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3 +[28] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2 +[29] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. In NeurIPS, 2019. 2 +[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, + +Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In CVPR, 2020. 2, 5 +[31] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. In ECCV, 2022. 2 +[32] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 2, 7 +[33] Tai Wang, Xinge Zhu, and Dahua Lin. Reconfigurable voxels: A new representation for lidar-based point clouds. In CoRL, 2020. 1 +[34] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. In NeurIPS, 2021. 4 +[35] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 2, 6, 7 +[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 1, 2, 4, 7 +[37] Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, Christoph Feichtenhofer, et al. Masked autoencoders that listen. arXiv:2207.06405, 2022. 2 +[38] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 2018. 1 +[39] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In CVPR, 2018. 2 +[40] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV,2022.1,2,5,6,7 +[41] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 1 +[42] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 2 +[43] Shuangfei Zhai, Navdeep Jaitly, Jason Ramapuram, Dan Busbridge, Tatiana Likhomanenko, Joseph Yitan Cheng, Walter Talbott, Chen Huang, Hanlin Goh, and Joshua Susskind. Position prediction as an effective pretraining strategy. In ICML, 2022. 2 +[44] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv:2205.14401, 2022. 2 + +[45] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. Dense siamese network for dense unsupervised learning. In ECCV, 2022. 1 +[46] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In CVPR, 2021. 2 +[47] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv:2111.07832, 2021. 1, 2 +[48] Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image analysis. arXiv:2203.05573, 2022. 2 +[49] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 3 +[50] Xinge Zhu, Yuexin Ma, Tai Wang, Yan Xu, Jianping Shi, and Dahua Lin. Ssn: Shape signature networks for multi-class object detection from point clouds. In ECCV, 2020. 1 \ No newline at end of file diff --git a/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/images.zip b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b0f5f5b8d80c11287967de40c7572be1c381b798 --- /dev/null +++ b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b189f5cb58988b165e5de5a59e853be2a71fe737a54cdd695aeee9422a37d0d +size 464578 diff --git a/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/layout.json b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c367d983a4de79d5f5aeaec36247f6e1937c052 --- /dev/null +++ b/2023/MV-JAR_ Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training/layout.json @@ -0,0 +1,7984 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "type": "text", + "content": "MV-JAR: Masked Voxel Jigsaw and Reconstruction for LiDAR-Based Self-Supervised Pre-Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "spans": [ + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "text", + "content": "Runsen Xu" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "text", + "content": " Tai Wang" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "text", + "content": " Wenwei Zhang" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "inline_equation", + "content": "^{3,2}" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "text", + "content": " Runjian Chen" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "text", + "content": " Jinkun Cao" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "text", + "content": " Jiangmiao Pang" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "text", + "content": " Dahua Lin" + }, + { + "bbox": [ + 103, + 159, + 489, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 190, + 497, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 190, + 497, + 217 + ], + "spans": [ + { + "bbox": [ + 96, + 190, + 497, + 217 + ], + "type": "text", + "content": "1The Chinese University of Hong Kong 2Shanghai AI Laboratory 3S-Lab, NTU 4The University of Hong Kong 5Carnegie Mellon University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 219, + 517, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 219, + 517, + 245 + ], + "spans": [ + { + "bbox": [ + 73, + 219, + 517, + 245 + ], + "type": "text", + "content": "{runsenxu,wt019,dhlin}@ie.cuhk.edu.hk, wenwei001@ntu.edu.sg, rjchen@connect.hku.hk, jinkunc@andrew.cmu.edu, pangjiangmiao@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 285, + 191, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 285, + 191, + 297 + ], + "spans": [ + { + "bbox": [ + 143, + 285, + 191, + 297 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 310, + 289, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 289, + 620 + ], + "type": "text", + "content": "This paper introduces the Masked Voxel Jigsaw and Reconstruction (MV-JAR) method for LiDAR-based self-supervised pre-training and a carefully designed data-efficient 3D object detection benchmark on the Waymo dataset. Inspired by the scene-voxel-point hierarchy in downstream 3D object detectors, we design masking and reconstruction strategies accounting for voxel distributions in the scene and local point distributions within the voxel. We employ a Reversed-Furthest-Voxel-Sampling strategy to address the uneven distribution of LiDAR points and propose MV-JAR, which combines two techniques for modeling the aforementioned distributions, resulting in superior performance. Our experiments reveal limitations in previous data-efficient experiments, which uniformly sample fine-tuning splits with varying data proportions from each LiDAR sequence, leading to similar data diversity across splits. To address this, we propose a new benchmark that samples scene sequences for diverse fine-tuning splits, ensuring adequate model convergence and providing a more accurate evaluation of pre-training methods. Experiments on our Waymo benchmark and the KITTI dataset demonstrate that MV-JAR consistently and significantly improves 3D detection performance across various data scales, achieving up to a " + }, + { + "bbox": [ + 46, + 310, + 289, + 620 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 46, + 310, + 289, + 620 + ], + "type": "text", + "content": " increase in mAPH compared to training from scratch. Codes and the benchmark are available at https://github.com/SmartBot-PJLab/MV-JAR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "content": "Self-supervised pre-training has gained considerable attention, owing to its exceptional performance in visual representation learning. Recent advancements in contrastive" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 272, + 430, + 430 + ], + "blocks": [ + { + "bbox": [ + 307, + 272, + 430, + 430 + ], + "lines": [ + { + "bbox": [ + 307, + 272, + 430, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 430, + 430 + ], + "type": "image", + "image_path": "bc99e3383e48fa7a358cd24437bab93ddf8278ebaef1bf496a601e560e915666.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 440, + 547, + 474 + ], + "lines": [ + { + "bbox": [ + 305, + 440, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 440, + 547, + 474 + ], + "type": "text", + "content": "Figure 1. 3D object detection results on the Waymo dataset. Our MV-JAR pre-training accelerates model convergence and greatly improves the performance with limited fine-tuning data." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 426, + 272, + 545, + 430 + ], + "blocks": [ + { + "bbox": [ + 426, + 272, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 426, + 272, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 426, + 272, + 545, + 430 + ], + "type": "image", + "image_path": "27eed56131e7787e728f4124f47081fe7fb9186f4e67cde9a5c4e39ba0cb321c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 521 + ], + "type": "text", + "content": "learning [4, 6, 8, 16, 45] and masked autoencoders [1, 7, 15, 36, 47] for images have sparked interest among researchers and facilitated progress in modalities such as point clouds." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 653 + ], + "type": "text", + "content": "However, LiDAR point clouds differ from images and dense point clouds obtained by reconstruction as they are naturally sparse, unorganized, and irregularly distributed. Developing effective self-supervised proxy tasks for these unique properties remains an open challenge. Constructing matching pairs for contrastive learning in geometry-dominant scenes is more difficult [20, 40], as points or regions with similar geometry may be assigned as negative samples, leading to ambiguity during training. To address this, our study explores masked voxel modeling paradigms for effective LiDAR-based self-supervised pre-training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": "Downstream LiDAR-based 3D object detectors [12, 19, 33, 38, 41, 50] typically quantize the 3D space into voxels and encode point features within them. Unlike pixels, which are represented by RGB values, the 3D space presents a scene-voxel-point hierarchy, introducing new" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 146, + 712 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13445" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 144 + ], + "type": "text", + "content": "challenges for masked modeling. Inspired by this, we design masking and reconstruction strategies that consider voxel distributions in the scene and local point distributions in the voxel. Our proposed method, Masked Voxel Jigsaw And Reconstruction (MV-JAR), harnesses the strengths of both voxel and point distributions to improve performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 144, + 286, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 144, + 286, + 311 + ], + "spans": [ + { + "bbox": [ + 48, + 144, + 286, + 311 + ], + "type": "text", + "content": "To account for the uneven distribution of LiDAR points, we first employ a Reversed-Furthest-Voxel-Sampling (R-FVS) strategy that samples voxels to mask based on their sparseness. This approach prevents masking the furthest distributed voxels, thereby avoiding information loss in regions with sparse points. To model voxel distributions, we propose Masked Voxel Jigsaw (MVJ), which masks the voxel coordinates while preserving the local shape of each voxel, enabling scene reconstruction akin to solving a jigsaw puzzle. For modeling local point distributions, we introduce Masked Voxel Reconstruction (MVR), which masks all coordinates of points within the voxel but retains one point as a hint for reconstruction. Combining these two methods enhances masked voxel modeling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 312, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 312, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 286, + 514 + ], + "type": "text", + "content": "Our experiments indicate that existing data-efficient experiments [20, 40] inadequately evaluate the effectiveness of various pre-training methods. The current benchmarks, which uniformly sample frames from each data sequence to create diverse fine-tuning splits, exhibit similar data diversity due to the proximity of neighboring frames in a sequence [3, 14, 30]. Moreover, these experiments train models for the same number of epochs across different fine-tuning splits, potentially leading to incomplete convergence. As a result, the benefits of pre-trained representations become indistinguishable across splits once the object detector is sufficiently trained on the fine-tuning data. To address these shortcomings, we propose sampling scene sequences to form diverse fine-tuning splits and establish a new data-efficient 3D object detection benchmark on the Waymo [30] dataset, ensuring sufficient model convergence for a more accurate evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 514, + 286, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 657 + ], + "type": "text", + "content": "We employ the Transformer-based SST [12] as our detector and pre-train its backbone for downstream detection tasks. Comprehensive experiments on the Waymo and KITTI [14] datasets demonstrate that our pre-training method significantly enhances the model's performance and convergence speed in downstream tasks. Notably, it improves detection performance by " + }, + { + "bbox": [ + 48, + 514, + 286, + 657 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 48, + 514, + 286, + 657 + ], + "type": "text", + "content": " mAPH when using only " + }, + { + "bbox": [ + 48, + 514, + 286, + 657 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 48, + 514, + 286, + 657 + ], + "type": "text", + "content": " of scenes for fine-tuning and reduces training time by half when utilizing the entire dataset (Fig. 1). With the representation pre-trained by MV-JAR, the 3D object detectors pre-trained on Waymo also exhibit generalizability when transferred to KITTI." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 669, + 132, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 132, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 132, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 286, + 712 + ], + "type": "text", + "content": "Self-supervised learning for point clouds. Annotating point clouds demands significant effort, necessitating self" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 72, + 544, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 544, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 544, + 358 + ], + "type": "text", + "content": "supervised pre-training methods. Prior approaches primarily focus on object CAD models [21, 26, 29, 39, 42, 44] and indoor scenes [17, 35, 46]. Point-BERT [42] applies BERT-like paradigms for point cloud recognition, while Point-MAE [26] reconstructs point patches without the tokenizer. To acquire scene-level representation, PointContrast [35] introduces a contrastive learning approach that compares points in two static partial views of a reconstructed indoor scene. Its successor, SceneContrast [17], incorporates spatial information into the contrastive learning framework. However, LiDAR point clouds are irregular and dynamic, with mainstream LiDAR perception models often exhibiting distinct architectures, which obstructs the direct adaptation of object-level and indoor scene pre-training methods. STRL [18] employs contrastive learning with two temporally-correlated frames for spatiotemporal representation. GCC-3D [20] presents a framework that combines geometry-aware contrast and pseudo-instance clustering harmonization. ProposalContrast [40] targets region-level contrastive learning to improve 3D detection. CO3 [5] leverages infrastructure-vehicle-cooperation point clouds to construct effective contrastive views. In contrast, our work explores masked voxel modeling, diverging from the prevailing contrastive learning paradigm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 363, + 544, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 363, + 544, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 363, + 544, + 624 + ], + "type": "text", + "content": "Masked autoencoders for self-supervised pre-training. Masked language modeling plays a pivotal role in self-supervised pre-training for Transformer-based networks in natural language processing [2, 10, 28]. This approach typically masks portions of the input and pre-trains networks to predict the original information. With the successful integration of Transformers into computer vision [11], researchers have increasingly focused on masked image modeling [1, 7, 15, 36, 43, 47] to mitigate the data-intensive issue of ViT [11]. BEiT [1] employs a tokenizer to generate discrete tokens for image patches and utilizes a BERT-like framework to pre-train ViT. MAE [15] introduces an asymmetric autoencoder for reconstructing RGB pixels of original images, eliminating the need for an extra tokenizer. SimMiM [36] encodes masked raw images with Transformers and employs a lightweight prediction head for recovery. Demonstrating considerable potential in image recognition, masked autoencoders have been applied in video understanding [13, 32], medical image analysis [48], and audio processing [37]. In this work, we investigate the application of this powerful pre-training technique to LiDAR point clouds using voxel representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 629, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 629, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 629, + 544, + 712 + ], + "type": "text", + "content": "Transformer-based 3D object detection. The recent success of Vision Transformers [11] has inspired extensive research into the application of Transformer-based architectures for 3D object detection [12, 23-25, 31]. 3DETR [24] presents an end-to-end object detection framework with modified Transformer blocks for 3D point clouds. Pointformer [25] employs hierarchical Transformer blocks to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "13446" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 220 + ], + "type": "image", + "image_path": "2f89105dc3364d6ded51795534fdab51ff4bed01f320412b3f85aec31ece80bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 546, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 546, + 273 + ], + "type": "text", + "content": "Figure 2. Overview of Masked Voxel Jigsaw and Reconstruction (MV-JAR). Our R-FVS sampling method selects a specific ratio of nonempty voxels to mask. For position-masked voxels, we mask the absolute coordinates of all points within the voxel while maintaining the relative coordinates. For shape-masked voxels, we mask both absolute and relative coordinates of all points, preserving only one point. Features are extracted from all non-empty voxels, and masked features are decoded by a dedicated head to recover corresponding targets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 289, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 424 + ], + "type": "text", + "content": "extract point features for 3D detection. Voxel Transformer [23] develops a Transformer-based 3D backbone to establish long-range relationships between voxels. Recently, SST [12] introduced a single-stride sparse Transformer as a replacement for the PointPillars [19] backbone, abandoning the multi-resolution strategy to enhance the detection of small objects. Emulating the SwinTransformer [22], SST partitions the space into windows and calculates attention only within each window for efficiency. This approach demonstrates impressive detection results, and we adopt it for our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 436, + 130, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 130, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 130, + 451 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 458, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 589 + ], + "type": "text", + "content": "In this section, we first present an overview of a general masked voxel modeling framework for LiDAR point clouds and our Reversed-Furthest-Voxel-Sampling strategy to tackle the uneven distribution of LiDAR points (Sec. 3.1). Masked Voxel Jigsaw (MVJ) (Sec. 3.2) and Masked Voxel Reconstruction (MVR) (Sec. 3.3) are both instantiations of the general framework that learns the voxel distributions and local point distributions, respectively. Masked Voxel Jigsaw and Reconstruction (MV-JAR) consist of these two tasks to synergically learn the data distributions at two levels (Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 598, + 251, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 251, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 251, + 612 + ], + "type": "text", + "content": "3.1. Masked Voxel Modeling and Sampling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "As illustrated in Fig. 2, a general masked voxel modeling framework converts LiDAR point clouds into voxels and samples a proportion of non-empty voxels using a specific sampling strategy and masking ratio. Various features within the voxel can be masked for a particular learning target in this general framework to instantiate different self-supervised learning tasks. Specifically, task-specific features of the points in the sampled voxels are replaced with" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 293, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 399 + ], + "type": "text", + "content": "a learnable vector. Subsequently, all non-empty voxels are encoded by a voxel encoder, processed by a backbone network to extract features, and followed by a lightweight decoder that reconstructs task-specific targets from each voxel feature. Each stage is described in the following sections. In this paper, we employ SST [12], which demonstrates superior performance in 3D detection. The pre-trained weights of the voxel encoder and the SST backbone are utilized as initialization for the downstream 3D detection task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": "Voxelization. Let " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " denote a point with coordinates " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": ". Given a point cloud " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "P = \\{p_i = [x_i, y_i, z_i]^T \\in \\mathbb{R}^3\\}_{i=1\\dots n}" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " points and range " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "W, H, D" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": ", we partition the space into a 3D grid using voxel sizes " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "v_W, v_H, v_D" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " along the three axes, grouping points in the same voxel together. As a common technique [12, 19, 49], we decorate each point with " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "x^c, y^c, z^c, x^v, y^v, z^v" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": ", where the superscript " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " denotes distance to the clustering center of all points in each voxel and the superscript " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " denotes distance to the voxel center. Considering the sparse and irregular nature of LiDAR point clouds, we focus only on non-empty voxels. Assuming non-empty voxels contain at most " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " points and a total of " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " non-empty voxels, we obtain the voxel-based point cloud representation as " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "M = \\{V_i\\}_{i=1\\dots N}" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "V_i = \\{p_j = [x_j, y_j, z_j, x_j^c, y_j^c, z_j^c, x_j^v, y_j^v, z_j^v]^T \\in \\mathbb{R}^9\\}_{j=1\\dots t_i}" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "t_i \\leq T" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": ". We assume each non-empty voxel contains " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": " points for simplicity. We also maintain each voxel's coordinate in the 3D grid with " + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "inline_equation", + "content": "C_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 401, + 546, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Masked voxel sampling. A key component of masked autoencoders is the mask sampling strategy. Unlike 2D pixels that uniformly distribute on the image plane, voxels are unevenly distributed due to the sparse and irregular nature of LiDAR point clouds. Random sampling may result in information loss in regions with sparse points. Inspired by furthest point sampling [27] (FPS), which samples an evenly distributed point set while maintaining the key struc" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "13447" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "ture of the data, we propose a Reversed-Furthest-Voxel-Sampling (R-FVS) scheme. Given a masking ratio " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": ", we apply FPS to select " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\lfloor N(1 - r)\\rfloor" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " voxels from all non-empty voxels according to each voxel's coordinate " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": ". These sampled voxels are kept to preserve points in sparse regions and maintain key geometric structures in the data, while the remaining unsampled voxels are masked during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "text", + "content": "Masking voxels. We directly mask the raw inputs of voxels, providing a unified formulation for the masked voxel modeling framework. For each of the remaining unsampled " + }, + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\lceil Nr \\rceil" + }, + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "text", + "content": " voxels to be masked, we replace some of the original features of each point in the voxel with a shared learnable vector, denoted as a mask token " + }, + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "text", + "content": ". The dimensions of " + }, + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "text", + "content": " and the replaced point features vary for different tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 286, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 286, + 325 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 286, + 325 + ], + "type": "text", + "content": "Voxel encoding and decoding. All the non-empty voxels are then encoded by a voxel encoder (VE) and the SST backbone. As the features of masked voxels are also extracted by the backbone, we only need a lightweight task-specific MLP head to decode and predict their original information [36]. The loss is computed only on masked voxels, following previous practices [15, 36]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 334, + 203, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 203, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 203, + 346 + ], + "type": "text", + "content": "3.2. Masked Voxel Jigsaw (MVJ)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 286, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 286, + 413 + ], + "type": "text", + "content": "Object detection in LiDAR point clouds necessitates a representation capable of capturing contextual information, such as the distribution and relationships among voxels. We propose a jigsaw puzzle that obscures position information, compelling the model to learn these relationships." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "type": "text", + "content": "Masking position information. Given a position-masked voxel, we replace the absolute coordinates " + }, + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "type": "inline_equation", + "content": "x_{j}, y_{j}, z_{j}" + }, + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "type": "text", + "content": " of each point with a mask token " + }, + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "type": "inline_equation", + "content": "m_{v} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "type": "text", + "content": ", while preserving the local shape by retaining the decorated coordinates of each point within the voxel (Fig. 2). The masked voxel is then represented as " + }, + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{masked}} = \\{p_{j} = [m_{v}, x_{j}^{c}, y_{j}^{c}, z_{j}^{c}, x_{j}^{v}, y_{j}^{v}, z_{j}^{v}]^{T} \\in \\mathbb{R}^{9}\\}_{j=1\\dots T}" + }, + { + "bbox": [ + 46, + 414, + 286, + 544 + ], + "type": "text", + "content": ". All voxels are subsequently input to the voxel encoder and the SST backbone. Notably, positional embeddings are not added to the voxels before being fed to the backbone, as the model is tasked with predicting position information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "text", + "content": "Prediction target and loss function. Instead of directly predicting the absolute position of the masked voxels, we formulate a classification task to facilitate optimization. Owing to the extensive range of LiDAR point clouds, the SST backbone partitions the 3D voxelized grid into windows, akin to the Swin-transformer [22], to enable efficient attention calculation within windows. We employ the same window partitioning method, requiring the model to predict only the relative index of the masked voxel within its corresponding window. Assuming a 3D window contains " + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "inline_equation", + "content": "N_x, N_y, N_z" + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "text", + "content": " voxels along the " + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "text", + "content": " axes and the voxel coordinates of a masked voxel are " + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "inline_equation", + "content": "X, Y, Z" + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "text", + "content": ", the relative coordinates " + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "inline_equation", + "content": "I_x, I_y, I_z" + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "text", + "content": " of the masked voxel with respect to its window are calculated as " + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "inline_equation", + "content": "I_x = X \\mod N_x, I_y = Y \\mod N_y" + }, + { + "bbox": [ + 46, + 545, + 286, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "mod " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "N_{y}, I_{z} = Z \\mod N_{z}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". The relative index is then given by " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "I = I_{x} + I_{y}N_{x} + I_{z}N_{x}N_{y}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". The prediction head outputs a classification vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\hat{v} \\in \\mathbb{R}^{N_xN_yN_z}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " representing the probabilities of the masked voxel occupying each position. The prediction loss for all masked voxel is calculated using the Cross-Entropy loss as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 345, + 149, + 545, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 149, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 345, + 149, + 545, + 183 + ], + "type": "interline_equation", + "content": "L _ {M V J} = \\frac {1}{R _ {p}} \\sum_ {i = 1} ^ {R _ {p}} \\text {C r o s s E n t r o p y} (\\hat {v} _ {i}, I _ {i}), \\tag {1}", + "image_path": "e51650335426350f7e2c1ecd8c5b31d3a54d6201100155c5d84c7db9c48a4426.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 187, + 537, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 187, + 537, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 187, + 537, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 187, + 537, + 200 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 306, + 187, + 537, + 200 + ], + "type": "text", + "content": " denotes the number of position-masked voxels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 205, + 504, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 504, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 504, + 217 + ], + "type": "text", + "content": "3.3. Masked Voxel Reconstruction (MVR)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 224, + 545, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 283 + ], + "type": "text", + "content": "To incorporate local shape information within each voxel, we introduce masked voxel reconstruction (MVR). MVR masks the absolute and relative coordinates of all points in shape-masked voxels to learn point distributions, preserving only one point to provide the voxel position." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 284, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 403 + ], + "type": "text", + "content": "Masking voxel shapes. In a given voxel, both the original and decorated coordinates of points represent the local shape of the voxel (Fig. 2). We employ a shared learnable token " + }, + { + "bbox": [ + 304, + 284, + 545, + 403 + ], + "type": "inline_equation", + "content": "m_p \\in \\mathbb{R}^9" + }, + { + "bbox": [ + 304, + 284, + 545, + 403 + ], + "type": "text", + "content": " to replace all point features in every shape-masked voxel, except for one point. This particular point conveys the voxel's positional information without revealing the voxel shape, which would render the reconstruction task trivial. Consequently, the masked voxel becomes " + }, + { + "bbox": [ + 304, + 284, + 545, + 403 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{masked}} = \\{[x_1, y_1, z_1, x_1^c, y_1^c, z_1^c, x_1^v, y_1^v, z_1^v]\\} \\cup \\{p_j = m_p\\}_{j=2\\dots T}" + }, + { + "bbox": [ + 304, + 284, + 545, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "text", + "content": "Reconstruction target and loss function. We reconstruct each masked voxel and use the " + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "text", + "content": " Chamfer Distance loss to measure the discrepancy between the reconstructed and target point distributions. This loss function is insensitive to point density [34], which is crucial since each voxel may contain varying numbers of points. Assuming each reconstructed voxel contains " + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "text", + "content": " points, the reconstruction head outputs a vector containing the coordinates of each point in the masked voxel, " + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\hat{v} \\in \\mathbb{R}^{3n}" + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "text", + "content": ". To pre-train the model more stably, we normalize the target point distribution using each point's distance to the voxel center as ground truth and scale the distance between 0 and 1. Supposing each reconstructed masked voxel is " + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\hat{V} = \\{\\hat{p}_j = [\\hat{x}_j, \\hat{y}_j, \\hat{z}_j]\\}_{j=1\\dots n}" + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "text", + "content": " and the ground truth is " + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "inline_equation", + "content": "V = \\{p_j = [x_j^v, y_j^v, z_j^v]\\}_{j=1\\dots t_i}" + }, + { + "bbox": [ + 304, + 403, + 545, + 582 + ], + "type": "text", + "content": ", the total loss across all masked voxels is given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 587, + 545, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 587, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 361, + 587, + 545, + 619 + ], + "type": "interline_equation", + "content": "L _ {M V R} = \\frac {1}{R _ {s}} \\sum_ {i = 1} ^ {R _ {s}} L _ {C D} \\left(\\hat {V} _ {i}, V _ {i}\\right), \\tag {2}", + "image_path": "24ef24f97027ca932203ffcef55cd0650f8595b3e8003eb4f2e38a1b17ec748f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "type": "text", + "content": " represents the number of shape-masked voxels and " + }, + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "type": "inline_equation", + "content": "L_{CD}" + }, + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 304, + 624, + 545, + 647 + ], + "type": "text", + "content": " Chamfer Distance loss calculated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 359, + 652, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 652, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 359, + 652, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {C D} = \\frac {1}{\\left| \\hat {V} _ {i} \\right|} \\sum_ {\\hat {p} \\in \\hat {V} _ {i}} \\min _ {p \\in V _ {i}} \\| \\hat {p} - p \\| _ {2} ^ {2} \\tag {3} \\\\ + \\frac {1}{| V _ {i} |} \\sum_ {p \\in V _ {i}} \\min _ {\\hat {p} \\in \\hat {V} _ {i}} \\| p - \\hat {p} \\| _ {2} ^ {2}. \\\\ \\end{array}", + "image_path": "996a59103109e850b12ef813223865bb58ca9e8819c1386b84da94fe854aad22.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13448" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "type": "text", + "content": "3.4. Joint Pre-Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 288, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 288, + 209 + ], + "type": "text", + "content": "We jointly pre-train the model using both MVJ and MVR tasks to enable the model to learn both point and voxel distributions. We use R-FVS to sample " + }, + { + "bbox": [ + 46, + 89, + 288, + 209 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 46, + 89, + 288, + 209 + ], + "type": "text", + "content": " voxels for MVR and " + }, + { + "bbox": [ + 46, + 89, + 288, + 209 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 46, + 89, + 288, + 209 + ], + "type": "text", + "content": " voxels for MVJ. As illustrated in Fig. 2, the voxel encoder and the backbone extract features from both types of masked voxels and unmasked voxels. Masked features are decoded to recover their corresponding targets, and Cross-Entropy loss and Chamfer Distance loss are utilized for supervision. The total loss for joint pre-training is expressed as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 219, + 287, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 219, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 116, + 219, + 287, + 232 + ], + "type": "interline_equation", + "content": "L = \\alpha L _ {M V J} + \\beta L _ {M V R} \\tag {4}", + "image_path": "d23784239ff8696006ad5af5908396645e8cd39cb9aa8cad6329dedd441371f8.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 240, + 216, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 216, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 216, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 240, + 216, + 252 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 240, + 216, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 240, + 216, + 252 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 240, + 216, + 252 + ], + "type": "text", + "content": " are balancing coefficients." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 262, + 255, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 255, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 255, + 277 + ], + "type": "text", + "content": "4. Data-Efficient Benchmark on Waymo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 283, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 354 + ], + "type": "text", + "content": "To evaluate the pre-trained representation, we transfer the weights of the voxel encoder and the backbone for object detection. Previous benchmarks fail to effectively reveal the effectiveness of pre-training strategies due to data diversity issues. We propose a new data-efficient benchmark to address these limitations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 462 + ], + "type": "text", + "content": "Observation: Incomplete model convergence. Prior works [20, 40] fine-tune the model on different splits with the same epochs, causing reduced iteration numbers when dataset sizes decrease. As illustrated in Fig. 3a, by training for more epochs (but the same iterations), the performance of SST significantly improves, suggesting that the models may not fully converge on smaller splits when training epochs remain constant. This entangled factor prevents accurate evaluation of pre-training strategies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "text", + "content": "Issue: Fine-Tuning splits with similar data diversities. With the same training iterations, models trained on different splits display comparable performance, despite significant differences in data amounts (Fig. 3a). We find that scene diversity is a crucial factor. Previous works [20, 40] uniformly sample the entire Waymo training set to create different fine-tuning splits, for instance, sampling one frame from every two consecutive frames for a " + }, + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "text", + "content": " split. However, neighboring frames in self-driving datasets are similar due to the short time difference between consecutive frames (e.g., only " + }, + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "inline_equation", + "content": "0.1s" + }, + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "text", + "content": " in Waymo [30]). This results in similar scene diversity across splits, leading to comparable detection performance using the same training iterations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "We argue that such an experimental setting does not accurately represent a typical application scenario of self-supervised pre-training, where annotated fine-tuning data is less abundant and diverse compared to pre-training data. Furthermore, it fails to evaluate how pre-training benefits downstream tasks when varying amounts and diversities of labeled data are available. We provide an additional example in the supplementary material to further illustrate" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 424, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 424, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 424, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 424, + 148 + ], + "type": "image", + "image_path": "2608633e5dbcc7499a8da46049e8c7601e74992ba225c2fc0fd9c527a902683e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 152, + 386, + 161 + ], + "lines": [ + { + "bbox": [ + 347, + 152, + 386, + 161 + ], + "spans": [ + { + "bbox": [ + 347, + 152, + 386, + 161 + ], + "type": "text", + "content": "(a) Existing" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 545, + 148 + ], + "type": "image", + "image_path": "59d89ebeb26e2e21bb5ea0b6aae706219fd9f3d8b6e6ee6784cb051e7ec6ca52.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 152, + 501, + 161 + ], + "lines": [ + { + "bbox": [ + 473, + 152, + 501, + 161 + ], + "spans": [ + { + "bbox": [ + 473, + 152, + 501, + 161 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 171, + 547, + 227 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 227 + ], + "type": "text", + "content": "Figure 3. Comparison of data-efficient benchmarks. Previous uniformly sampled splits exhibit similar diversity, as evidenced by similar performance with the same training iterations. Our benchmark features varying data amounts and diversities, ensuring model convergence. Dashed lines show baseline performances." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 247, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 282 + ], + "type": "text", + "content": "these issues with previous benchmarks. Our proposed data-efficient benchmark addresses these issues, offering a more precise evaluation of pre-training strategies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "content": "Solution: Sequence-Based data sampling. We sample fine-tuning data using different scene sequences rather than uniformly sampling varying numbers of frames from identical sequences. The Waymo training split comprises 798 distinct scene sequences, each lasting " + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "inline_equation", + "content": "20s" + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "content": " with approximately 200 frames [30]. We randomly sample " + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "content": " of the scene sequences, respectively, using all frames within the sampled sequences for fine-tuning. Consequently, data diversities and amounts vary across each split. Each larger split also encompasses the smaller ones, enabling measurement of performance changes with additional fine-tuning data. Furthermore, we randomly sample three " + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 282, + 546, + 449 + ], + "type": "text", + "content": " splits and report average results to reduce variance on these smaller sets (see our supplementary material)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 449, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 546 + ], + "type": "text", + "content": "For different fine-tuning splits, we train the model until performance saturates (i.e., overfitting) and determine the number of epochs used for experiments accordingly. For example, for " + }, + { + "bbox": [ + 304, + 449, + 545, + 546 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 449, + 545, + 546 + ], + "type": "text", + "content": " data, we try " + }, + { + "bbox": [ + 304, + 449, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\{48, 60, 66, 72, 78, 84\\}" + }, + { + "bbox": [ + 304, + 449, + 545, + 546 + ], + "type": "text", + "content": " epochs and ultimately select 72. The baseline performances across various training iterations are shown in Fig. 3b. With our new benchmark, baseline performance and convergence speed vary as anticipated for different splits." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 388, + 569 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 575, + 432, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 432, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 432, + 588 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Datasets. The Waymo dataset [30] consists of 1150 self-driving scenes with 798 scenes as the training split, 202 scenes as the validation split, and 150 scenes as the testing split. We pre-train on the training split, which contains 158,240 annotated frames. For the downstream data-efficient 3D object detection task, we fine-tune the model using our sampled splits (as described in Sec. 4) and evaluate performance on the standard Waymo validation split with 3 classes (cars, pedestrians, and cyclists) and two difficulty levels (L1 and L2). We adopt L2 mean average preci" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13449" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 113, + 535, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Data-efficient 3D object detection results of SST on the Waymo validation set. SST is pre-trained using various self-supervised methods on Waymo training split, with weights applied for the downstream detection task. \"Random initialization\" denotes training from scratch. The model performances are shown using different amounts and diversities of fine-tuning data." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 113, + 535, + 352 + ], + "lines": [ + { + "bbox": [ + 61, + 113, + 535, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 113, + 535, + 352 + ], + "type": "table", + "html": "
Data amountInitializationOverallCarPedestrianCyclist
L2 mAPL2 mAPHL2 mAPL2 mAPHL2 mAPL2 mAPHL2 mAPL2 mAPH
5%Random44.4140.3451.0150.4952.7442.2629.4928.27
PointContrast [35]45.3241.3052.1251.6153.6843.2230.1629.09
ProposalContrast [40]46.6242.5852.6752.1954.3143.8232.8731.72
MV-JAR (Ours)50.52+6.1146.68+6.3456.4756.0157.6547.6937.4436.33
10%Random54.3150.4654.8454.3760.5550.7147.5546.29
PointContrast [35]53.6949.9454.7654.3059.7550.1246.5745.39
ProposalContrast [40]53.8950.1355.1854.7160.0150.3946.4845.28
MV-JAR (Ours)57.44+3.1354.06+3.6058.4358.0063.2854.6650.6349.52
20%Random60.1656.7858.7958.3565.6357.0456.0754.94
PointContrast [35]59.3555.7858.6458.1864.3955.4355.0253.73
ProposalContrast [40]59.5255.9158.6958.2264.5355.4555.3654.07
MV-JAR (Ours)62.28+2.1159.15+2.3761.8861.4566.9859.0257.9857.00
50%Random66.4363.3663.8163.3870.7863.0564.7163.66
PointContrast [35]65.5162.2162.6662.2369.8261.5364.0462.86
ProposalContrast [40]65.7662.4962.9362.5070.0961.8664.2663.11
MV-JAR (Ours)66.70+0.2763.69+0.3364.3063.8971.1463.5764.6563.63
100%Random68.5065.5464.9664.5672.3864.8968.1767.17
PointContrast [35]68.0664.8464.2463.8271.9263.8168.0366.89
ProposalContrast [40]68.1765.0164.4264.0071.9463.9468.1667.10
MV-JAR (Ours)69.16+0.6666.20+0.6665.5265.1272.7765.2869.1968.20
", + "image_path": "818118d44b028a1012bdd330c3327d5dd8a76846789cd53ee58b165283d50811.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 370, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 418 + ], + "type": "text", + "content": "sion (L2 mAP) and L2 mean average precision with heading (L2 mAPH) as main evaluation metrics. We also evaluate the model on the KITTI dataset [14] for the detection task. It contains 7481 training samples and 7518 test samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "content": "Implementation details. We employ the official implementation and training settings of SST [12] for pre-training and fine-tuning unless specified otherwise. The Waymo dataset's point cloud range is " + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "inline_equation", + "content": "W \\times H \\times D = 149.76m \\times 149.76m \\times 6m" + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "content": ", and the voxel size is " + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "inline_equation", + "content": "v_{W} \\times v_{H} \\times v_{D} = 0.32m \\times 0.32m \\times 6m" + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "content": ". For MVR, we predict 15 points per voxel with a masking ratio of 0.05. For MVJ, the window size is " + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "inline_equation", + "content": "N_{x} \\times N_{y} \\times N_{z} = 12 \\times 12 \\times 1" + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "content": " and the masking ratio is 0.1. We set the loss weight " + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta = 1" + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "content": ", pre-train for 6 epochs with an initial learning rate of " + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "inline_equation", + "content": "5e - 6" + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "content": ". When fine-tuning, we initialize the voxel encoder and SST backbone with pre-trained weights, and other training hyperparameters remain unchanged. We set KITTI's point cloud range to " + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "inline_equation", + "content": "69.12m \\times 79.36m \\times 4m" + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "content": ", voxel size to " + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "inline_equation", + "content": "0.32m \\times 0.32m \\times 4m" + }, + { + "bbox": [ + 46, + 419, + 289, + 624 + ], + "type": "text", + "content": ", and train for 160 epochs. We use the AdamW optimizer with a batch size of 8 and a cyclic learning rate scheduler with cosine annealing for all training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 634, + 185, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 185, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 185, + 647 + ], + "type": "text", + "content": "5.2. Main Results on Waymo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Baselines. We fine-tune the model using data splits described in Sec. 4. For " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{5\\%, 10\\%, 20\\%, 50\\%, 100\\}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " data, the model is trained for " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{72, 60, 42, 36, 24\\}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " epochs. We use the randomly initialized model's performances as our baselines. We also include experiments with models using" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 370, + 531, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 531, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 531, + 383 + ], + "type": "text", + "content": "convolutional backbones in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "text", + "content": "Results with different initialization. Tab. 1 presents the SST model's performances with various initializations on our proposed data-efficient benchmark. Our pre-training method consistently improves SST baselines across all finetuning data amounts, especially when data is scarce. Our method achieves up to " + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "text", + "content": " relative improvement, from " + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "inline_equation", + "content": "40.34\\%" + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "text", + "content": " L2 mAPH to " + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "inline_equation", + "content": "46.68\\%" + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "text", + "content": " L2 mAPH, when using only " + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "text", + "content": " fine-tuning data. The performance gain exists across object categories, indicating effective general representation learning. As more fine-tuning data is introduced, the performance gain persists, and when fine-tuning with the entire Waymo training set, our pre-training enhances detection performance from " + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "inline_equation", + "content": "65.54\\%" + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "text", + "content": " mAPH to " + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "inline_equation", + "content": "66.20\\%" + }, + { + "bbox": [ + 304, + 385, + 545, + 540 + ], + "type": "text", + "content": " mAPH." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "content": "We note that with " + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "content": " fine-tuning data, the improvements from our pre-training are not as significant as with less data. We observe that the baseline result on " + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "content": " data (63.36% mAPH) is comparable to that on " + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "content": " data (65.54% mAPH), suggesting that training data is no longer the bottleneck. The original SST, with only 2.1M parameters, is a lightweight network. We scale up SST to 8.3M parameters by using larger hidden layer channels and observe larger benefits from pre-training on the " + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "content": " fine-tuning split, from " + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "inline_equation", + "content": "63.13\\%" + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "content": " L2 mAPH to " + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "inline_equation", + "content": "64.24\\%" + }, + { + "bbox": [ + 304, + 543, + 546, + 699 + ], + "type": "text", + "content": " L2 mAPH. Therefore, we argue that the SST model's capacity, rather than the fine-tuning data amount, becomes the bottleneck for detection performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Comparisons with contrastive learning methods. We" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13450" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 112, + 250, + 185 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "text", + "content": "Table 2. Transferring results on the KITTI validation split. We pre-train SST with the Waymo training split and fine-tune with the KITTI training split." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 112, + 250, + 185 + ], + "lines": [ + { + "bbox": [ + 87, + 112, + 250, + 185 + ], + "spans": [ + { + "bbox": [ + 87, + 112, + 250, + 185 + ], + "type": "table", + "html": "
InitializationOverall
EasyMod.Hard
Random74.7163.4360.00
PointContrast [35]73.3562.5359.01
ProposalContrast [40]73.6363.3459.40
MV-JAR (Ours)75.2263.8060.35
", + "image_path": "c7e29b766c5c852d729b1cb467b4cea66f5da5c6d70132b8ecdabceedb8a3857.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 396 + ], + "type": "text", + "content": "utilize the official implementation of the recently proposed ProposalContrast [40] and reimplement PointContrast [35] for comparison. As demonstrated in Tab. 1, both contrastive learning methods offer benefits only with " + }, + { + "bbox": [ + 46, + 205, + 287, + 396 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 205, + 287, + 396 + ], + "type": "text", + "content": " fine-tuning data, and their improvements diminish with higher baseline performances. This may be due to the difficulty in smoothly learning representations. A key challenge in LiDAR-based contrastive learning is constructing matched pairs. The point pairs used by PointContrast face the issue of spatially adjacent points with similar features being assigned as negative samples. We observe difficulty in convergence when sampling excessive pairs. ProposalContrast alleviates the issue by contrasting regions but does not entirely resolve it. Alternatively, our method offers a new avenue for developing self-supervised pre-training through masked voxel modeling on LiDAR point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 404, + 247, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 247, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 247, + 417 + ], + "type": "text", + "content": "5.3. Pre-training Accelerates Convergence" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 423, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 604 + ], + "type": "text", + "content": "To investigate how MV-JAR pre-training accelerates convergence, we fine-tune the SST model on the entire Waymo training split. We train the SST model for different epochs and report the performance of both fine-tuning and training from scratch. As illustrated in Fig. 1, MV-JAR pre-training significantly enhances the convergence speed of SST. By fine-tuning for only 3 epochs, the SST model achieves " + }, + { + "bbox": [ + 46, + 423, + 287, + 604 + ], + "type": "inline_equation", + "content": "62.48\\%" + }, + { + "bbox": [ + 46, + 423, + 287, + 604 + ], + "type": "text", + "content": " L2 mAPH. Remarkably, fine-tuning for just 12 epochs results in a performance comparable to training from scratch for 24 epochs, effectively halving the convergence time. As the number of fine-tuning epochs increases, the performance gain provided by MV-JAR does not wane. Pre-training can still improve the model's converged performance by " + }, + { + "bbox": [ + 46, + 423, + 287, + 604 + ], + "type": "inline_equation", + "content": "0.74\\%" + }, + { + "bbox": [ + 46, + 423, + 287, + 604 + ], + "type": "text", + "content": " L2 mAPH, highlighting the superiority of the pre-trained representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 610, + 217, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 217, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 217, + 624 + ], + "type": "text", + "content": "5.4. Transferring Results on KITTI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 286, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 715 + ], + "type": "text", + "content": "To evaluate the transferability of the learned representation, we fine-tune various pre-trained SST models on the KITTI dataset. The original SST paper [12] did not conduct experiments on the KITTI dataset; thus, we follow the training schedule and setup of SST and PointPillars [19] within the MMDetection3D framework [9]. As demonstrated in Tab. 2, the performance gain from our MV-JAR pre-training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "persists when the representation is transferred to a different domain, indicating that the model learns a generic representation through pre-training. While the KITTI training samples account for approximately " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " of the entire Waymo training split, the relative improvement is much smaller than transferring to the " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " Waymo split. We hypothesize that this may be attributed to the domain gap between the Waymo and KITTI datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 178, + 406, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 406, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 406, + 190 + ], + "type": "text", + "content": "5.5. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 197, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 545, + 280 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to investigate the effectiveness of our proposed Reversed-Furthest-Voxel-Sampling (R-FVS) strategy. Additionally, we explore the optimal masking ratio and the impact of MVR and MVJ pre-training. All pre-training experiments utilize the entire Waymo training split, and we fine-tune the model on one of our " + }, + { + "bbox": [ + 304, + 197, + 545, + 280 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 197, + 545, + 280 + ], + "type": "text", + "content": " data splits." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": "Mask sampling strategy. Our R-FVS mask sampling strategy masks voxels that are not sampled by FPS, aiming to avoid masking voxels located in sparse regions. We also examine random sampling and an opposing sampling strategy called FVS, which masks voxels sampled by FPS. With FVS sampling, regions with sparse points are more likely to be masked, leading to greater information loss compared to random sampling and R-FVS sampling. Tab. 3 presents the fine-tuning results of pre-training with the three sampling strategies. R-FVS sampling performs the best, while FVS performs the worst. These findings confirm that minimizing information loss benefits pre-training and validate the effectiveness of our R-FVS strategy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 652 + ], + "type": "text", + "content": "Masking ratio. Tab. 4 illustrates the impact of different masking ratio combinations. Specifically, MVR with a 0.1 masking ratio and MVJ with a 0.05 masking ratio yield the best results. With this combination, the overall masking ratio is 0.15, which is significantly lower than the masking ratios in the image [15, 36] or video domain [13, 32]. A high masking ratio in masked autoencoders typically indicates information redundancy [13, 15, 32, 36]. We postulate that two factors contribute to our low masking ratio: 1) LiDAR point clouds represent a vast space where each meaningful object occupies only a small fraction. The information on each object is not highly redundant, especially when the objects are distant from the sensor. This contrasts with images in datasets like ImageNet, where objects often encompass a significant portion of the image. 2) The SST model calculates attention only within a partitioned window, which restricts the information used. We leave this question for future exploration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "Pre-training effects of MVR and MVJ. We investigate the individual pre-training effects of MVR and MVJ by optimizing the masking ratio for each task and reporting the fine-tuning performances in Tab. 5. Both MVR and MVJ pre-training enhance performance compared to random ini" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13451" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 294, + 267 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 294, + 267 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 294, + 267 + ], + "type": "image", + "image_path": "122ade527333146bfb78f1e4479a3233e9ad229d33c00b1ebdc20c6cbff47fa8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 276, + 518, + 288 + ], + "lines": [ + { + "bbox": [ + 72, + 276, + 518, + 288 + ], + "spans": [ + { + "bbox": [ + 72, + 276, + 518, + 288 + ], + "type": "text", + "content": "Figure 4. Visualization of the reconstruction results of MV-JAR on the Waymo validation set, compared with ground truths." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 74, + 544, + 267 + ], + "blocks": [ + { + "bbox": [ + 296, + 74, + 544, + 267 + ], + "lines": [ + { + "bbox": [ + 296, + 74, + 544, + 267 + ], + "spans": [ + { + "bbox": [ + 296, + 74, + 544, + 267 + ], + "type": "image", + "image_path": "8054ab95e8e42dbecbfa41a75bd0d776b159a5be7621c11e4b0fe948951c8d30.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 334, + 192, + 397 + ], + "blocks": [ + { + "bbox": [ + 47, + 303, + 198, + 325 + ], + "lines": [ + { + "bbox": [ + 47, + 303, + 198, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 198, + 325 + ], + "type": "text", + "content": "Table 3. Ablation study on mask sampling strategies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 334, + 192, + 397 + ], + "lines": [ + { + "bbox": [ + 57, + 334, + 192, + 397 + ], + "spans": [ + { + "bbox": [ + 57, + 334, + 192, + 397 + ], + "type": "table", + "html": "
Mask samplingOverall
L2 mAPL2 mAPH
FVS49.7345.88
Random49.8646.06
R-FVS50.5246.68
", + "image_path": "b12d36ebc59d7493558f4cb53fc0826b4477b3a00a7458efd400af1cbf787ff2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 230, + 319, + 369, + 401 + ], + "blocks": [ + { + "bbox": [ + 220, + 300, + 375, + 311 + ], + "lines": [ + { + "bbox": [ + 220, + 300, + 375, + 311 + ], + "spans": [ + { + "bbox": [ + 220, + 300, + 375, + 311 + ], + "type": "text", + "content": "Table 4. Ablation study on masking ratios." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 230, + 319, + 369, + 401 + ], + "lines": [ + { + "bbox": [ + 230, + 319, + 369, + 401 + ], + "spans": [ + { + "bbox": [ + 230, + 319, + 369, + 401 + ], + "type": "table", + "html": "
Masking ratioOverall
MVJMVRL2 mAPL2 mAPH
0.30.149.7446.05
0.050.149.6745.92
0.10.149.8545.97
0.10.0550.5246.68
0.10.350.2546.35
", + "image_path": "6db408c86e77732758377b00ce744551e2d216f84e25032abf947a6819a7304d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 416, + 334, + 533, + 396 + ], + "blocks": [ + { + "bbox": [ + 397, + 303, + 548, + 324 + ], + "lines": [ + { + "bbox": [ + 397, + 303, + 548, + 324 + ], + "spans": [ + { + "bbox": [ + 397, + 303, + 548, + 324 + ], + "type": "text", + "content": "Table 5. Ablation study on pre-training effects of MVR and MVJ." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 416, + 334, + 533, + 396 + ], + "lines": [ + { + "bbox": [ + 416, + 334, + 533, + 396 + ], + "spans": [ + { + "bbox": [ + 416, + 334, + 533, + 396 + ], + "type": "table", + "html": "
TaskOverall
L2 mAPL2 mAPH
MVR46.5742.73
MVJ49.9646.10
MV-JAR50.5246.68
", + "image_path": "13c1cd4f80ea2da424b1e6f6abcc878beda39ece09b43ad5a8b2d3dcd4f31711.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 420, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 539 + ], + "type": "text", + "content": "tialization, with MVJ outperforming MVR. This finding suggests that capturing voxel distributions plays a more significant role in representation learning, which is consistent with the fact that LiDAR detectors process points by down-sampling them into voxels. Our integrated MV-JAR method demonstrates further improvement over MVR and MVJ, validating the effectiveness of jointly capturing point and voxel distributions. Additionally, we offer an analysis of improvements across various distances in our supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 550, + 130, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 130, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 130, + 562 + ], + "type": "text", + "content": "5.6. Visualization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Figure 4 displays the MV-JAR reconstruction results on the Waymo validation set, including MVJ and MVR outcomes. In our pre-training, MVJ achieves approximately " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " classification accuracy of masked voxels on the validation set, accurately reconstructing masked voxels in the correct positions most of the time. MVR can capture the point distributions within voxels, evidenced by reconstructed ground points aligning with ground circles. However, MVR struggles to capture detailed distributions, likely due to the discrete sampling of LiDAR points from continuous surfaces, resulting in considerable variability in the data and increased difficulty in capturing finer details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 419, + 383, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 383, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 383, + 431 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 445, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 648 + ], + "type": "text", + "content": "In this paper, we introduce the Masked Voxel Jigsaw and Reconstruction (MV-JAR) pre-training method for LiDAR detectors. MV-JAR captures both point and voxel distributions of LiDAR point clouds, enabling models to learn effective and generic representations. We also develop a Reversed-Furthest-Voxel-Sampling strategy to address the uneven distribution of LiDAR points. Comprehensive experiments on the Waymo and KITTI datasets show that our method consistently and significantly enhances the detector's performance across different data scale regimes. MV-JAR offers a promising alternative for LiDAR-based self-supervised pre-training through Masked Voxel Modeling. Additionally, we establish a new data-efficient benchmark on the Waymo dataset, incorporating fine-tuning splits with diverse data variations. This benchmark effectively assesses the impact of pre-training on downstream tasks with varying amounts and diversities of labeled data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This project is funded in part by the Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd. under the Innovation and Technology Commission (ITC)'s InnoHK." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13452" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 92, + 286, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 286, + 114 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 286, + 114 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv:2106.08254, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 116, + 286, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 116, + 286, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 116, + 286, + 226 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 229, + 286, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 229, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 229, + 286, + 283 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. arXiv:1903.11027, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 286, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 286, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 286, + 286, + 330 + ], + "type": "text", + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 333, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 333, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 333, + 286, + 376 + ], + "type": "text", + "content": "[5] Runjian Chen, Yao Mu, Runsen Xu, Wenqi Shao, Chenhan Jiang, Hang Xu, Yu Qiao, Zhenguo Li, and Ping Luo. CO3: Cooperative unsupervised 3d representation learning for autonomous driving. In ICLR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 379, + 286, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 379, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 379, + 286, + 412 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 415, + 286, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 286, + 469 + ], + "type": "text", + "content": "[7] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv:2202.03026, 2022.1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 472, + 286, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 472, + 286, + 505 + ], + "spans": [ + { + "bbox": [ + 53, + 472, + 286, + 505 + ], + "type": "text", + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 508, + 286, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 508, + 286, + 551 + ], + "spans": [ + { + "bbox": [ + 53, + 508, + 286, + 551 + ], + "type": "text", + "content": "[9] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 554, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 286, + 597 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 286, + 666 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "text", + "content": "[12] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[13] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv:2205.09113, 2022. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "type": "text", + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2, 4, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "type": "text", + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "text", + "content": "[17] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "text", + "content": "[18] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In CVPR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 276, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 317 + ], + "type": "text", + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 1, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "text", + "content": "[20] Hanxue Liang, Chenhan Jiang, Dapeng Feng, Xin Chen, Hang Xu, Xiaodan Liang, Wei Zhang, Zhenguo Li, and Luc Van Gool. Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In CVPR, 2021. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 408 + ], + "type": "text", + "content": "[21] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "text", + "content": "[23] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "text", + "content": "[24] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 522, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 553 + ], + "type": "text", + "content": "[25] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In CVPR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 556, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 588 + ], + "type": "text", + "content": "[26] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 622 + ], + "type": "text", + "content": "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "text", + "content": "[28] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[29] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou," + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13453" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 139 + ], + "type": "text", + "content": "Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in Perception for Autonomous Driving: Waymo Open Dataset. In CVPR, 2020. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[31] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. In ECCV, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "text", + "content": "[32] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 232, + 287, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 264 + ], + "type": "text", + "content": "[33] Tai Wang, Xinge Zhu, and Dahua Lin. Reconfigurable voxels: A new representation for lidar-based point clouds. In CoRL, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 266, + 287, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 287, + 309 + ], + "type": "text", + "content": "[34] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 312, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 312, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 287, + 356 + ], + "type": "text", + "content": "[35] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 358, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 287, + 401 + ], + "type": "text", + "content": "[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 1, 2, 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "text", + "content": "[37] Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, Christoph Feichtenhofer, et al. Masked autoencoders that listen. arXiv:2207.06405, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 439, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 287, + 460 + ], + "type": "text", + "content": "[38] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 463, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 495 + ], + "type": "text", + "content": "[39] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "type": "text", + "content": "[40] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV,2022.1,2,5,6,7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 565 + ], + "type": "text", + "content": "[41] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[42] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[43] Shuangfei Zhai, Navdeep Jaitly, Jason Ramapuram, Dan Busbridge, Tatiana Likhomanenko, Joseph Yitan Cheng, Walter Talbott, Chen Huang, Hanlin Goh, and Joshua Susskind. Position prediction as an effective pretraining strategy. In ICML, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[44] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv:2205.14401, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 275 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[45] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. Dense siamese network for dense unsupervised learning. In ECCV, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[46] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[47] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv:2111.07832, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[48] Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image analysis. arXiv:2203.05573, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "type": "text", + "content": "[49] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "type": "text", + "content": "[50] Xinge Zhu, Yuexin Ma, Tai Wang, Yan Xu, Jianping Shi, and Dahua Lin. Ssn: Shape signature networks for multi-class object detection from point clouds. In ECCV, 2020. 1" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13454" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_content_list.json b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1625bc36921bdc8ec4065bdc492d07c274364ce3 --- /dev/null +++ b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_content_list.json @@ -0,0 +1,1847 @@ +[ + { + "type": "text", + "text": "MVImgNet: A Large-scale Dataset of Multi-view Images", + "text_level": 1, + "bbox": [ + 197, + 130, + 772, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/072d33ab3934ec8009ffb05b7f796dd1a196075de0e457c63ed880ec12548083.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 178, + 869, + 308 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0cfd9e332b747d0067e24e4e45f857297361add9a2b3aff92d4e8a3fa67e2b7.jpg", + "image_caption": [ + "Figure 1. It is time to embrace MVImgNet! We introduce MVImgnet, a large-scale dataset of multi-view images, which is efficiently collected by shooting videos of real-world objects. It enjoys 3D-aware signals from multi-view consistency, being a soft bridge between 2D and 3D vision. Through dense reconstruction on MVImgNet, we also present a large-scale real-world 3D object point cloud dataset – MVPNet. Exterior: Examples of various multi-view images in MVImgNet (see Fig. 3 more intuitively). Interior: Instances of colorful point clouds from MVPNet (see Fig. 7 more clearly) are assembled into a stereo sign of 'MVImgNet'." + ], + "image_footnote": [], + "bbox": [ + 78, + 340, + 890, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 616, + 313, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Being data-driven is one of the most iconic properties of deep learning algorithms. The birth of ImageNet [24] drives a remarkable trend of 'learning from large-scale data' in computer vision. Pretraining on ImageNet to obtain rich universal representations has been manifested to benefit various 2D visual tasks, and becomes a standard in 2D vision. However, due to the laborious collection of real-world 3D data, there is yet no generic dataset serving as a counterpart of ImageNet in 3D vision, thus how such a dataset can impact the 3D community is unraveled. To remedy this defect, we introduce MVImgNet, a large-scale dataset of multi-view images, which is highly convenient to gain by shooting videos of real-world objects in human daily life. It contains 6.5 million frames from 219,188 videos crossing objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds.", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The multi-view attribute endows our dataset with 3D-aware signals, making it a soft bridge between 2D and 3D vision.", + "bbox": [ + 500, + 617, + 890, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We conduct pilot studies for probing the potential of MVImgNet on a variety of 3D and 2D visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising performance, remaining lots of possibilities for future explorations.", + "bbox": [ + 496, + 659, + 893, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Besides, via dense reconstruction on MVImgNet, a 3D object point cloud dataset is derived, called MVPNet, covering 87,200 samples from 150 categories, with the class label on each point cloud. Experiments show that MVPNet can benefit the real-world 3D object classification while posing new challenges to point cloud understanding.", + "bbox": [ + 496, + 761, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MVImgNet and MVPNet will be public, hoping to inspire the broader vision community.", + "bbox": [ + 498, + 863, + 890, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9150", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 90, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Being data-driven, also known as data-hungry, is one of the most important attributes of deep learning algorithms. By training on large-scale datasets, deep neural networks are able to extract rich representations. In the past few years, the computer vision community has witnessed the bloom of such 'learning from data' regime [43, 44, 55], after the birth of ImageNet [24] – the pioneer of large-scale real-world image datasets. Notably, pretraining on ImageNet is well-proven to boost the model performance when transferring the pretrained weights into not only high-level [34, 41, 42, 60, 61] but also low-level visual tasks [15, 57], and becomes a de-facto standard in 2D. Recently, various 3D datasets [5, 9, 11, 23, 33, 96, 104] are produced to facilitate 3D visual applications.", + "bbox": [ + 76, + 114, + 467, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, due to the non-trivial scanning and laborious labeling of real-world 3D data (commonly organized in point clouds or meshes), existing 3D datasets are either synthetic or their scales are not comparable with ImageNet [24]. Consequently, unlike in 2D vision where models are usually pretrained on ImageNet to gain universal representation or commonsense knowledge, most of the current methods in 3D area are directly trained and evaluated on particular datasets for solving specific 3D visual tasks (e.g., NeRF dataset [64] and ShapeNet [11] for novel view synthesis, ModelNet [96] and ScanObjectNN [88] for object classification, KITTI [33] and ScanNet [23] for scene understanding). Here, two crucial and successive issues can be induced: (1) There is still no generic dataset in 3D vision, as a counterpart of ImageNet in 2D. (2) What benefit such a dataset can endow to 3D community is yet unknown. In this paper, we focus on investigating these two problems and set two corresponding targets: Build the primary dataset, then explore its effect through experiments.", + "bbox": [ + 76, + 327, + 467, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Milestone 1 - Dataset:", + "text_level": 1, + "bbox": [ + 96, + 614, + 253, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For a clearer picture of the first goal, we start by carefully revisiting existing 3D datasets as well as ImageNet [24]. i) 3D synthetic datasets [11, 96] provide rich 3D CAD models. However, they lack real-world cues (e.g., context, occlusions, noises), which are indispensable for model robustness in practical applications. ScanObjectNN [88] extracts real-world 3D objects from indoor scene data, but is limited in scale. For 3D scene-level dataset [5, 10, 23, 33, 37, 81], their scales are still constrained by the laborious scanning and labeling (e.g., millions of points per scene). Additionally, they contain specific inner-domain knowledge such as a particularly intricate indoor room or outdoor driving configurations, making it hard for general transfer learning. ii) Although ImageNet [24] contains the most comprehensive real-world objects, it only describes a 2D world that misses 3D-aware signals. Since humans live in a 3D world, 3D consciousness is vitally important for realizing humanlike intelligence and solving real-life visual problems.", + "bbox": [ + 76, + 628, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above review, our dataset is created from a new insight - multi-view images, as a soft bridge between 2D and 3D. It lies several benefits to remedying the aforementioned defects. Such data can be easily gained in considerable sizes via shooting an object around different views on common mobile devices with cameras (e.g., smartphones), which can be collected by crowd-sourcing in real world. Moreover, the multi-view constraint can bring natural 3D visual signals (later experiments show that this not only benefits 3D tasks but also 2D image understanding). To this end, we build MVImgNet, containing 6.5 million frames from 219,188 videos crossing real-life objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds. You may take a glance at our MVImgNet from Fig. 1.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Milestone 2 - Experimental Exploration:", + "text_level": 1, + "bbox": [ + 517, + 316, + 805, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Now facing the second goal of this paper, we attempt to probe the power of our dataset by conducting some pilot experiments. Leveraging the multi-view nature of the data, we start by focusing on the view-based 3D reconstruction task and demonstrate that pretraining on MVImgNet can not only benefit the generalization ability of NeRF (Sec. 4.1), but also data-efficient multi-view stereo (Sec. 4.2). Moreover, for image understanding, although humans can easily recognize one object from different viewpoints, deep learning models can hardly do that robustly [26,38]. Considering MVImgNet provides numerous images of a particular object from different viewpoints, we verify that MVImgNet-pretrained models are endowed with decent view consistency in general image classification (supervised learning in Sec. 5.1, self-supervised contrastive learning in Sec. 5.2) and salient object detection (Sec. 5.3).", + "bbox": [ + 496, + 333, + 890, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bonus - A New 3D Point Cloud Dataset - MVPNet:", + "text_level": 1, + "bbox": [ + 517, + 574, + 880, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through dense reconstruction on MVImgNet, a new 3D object point cloud dataset is derived, called MVPNet, which contains 87,200 point clouds with 150 categories, with the class label on each point cloud (see Fig. 7). Experiments show that MVPNet not only benefits the real-world 3D object classification task but also poses new challenges and prospects to point cloud understanding (Sec. 6).", + "bbox": [ + 496, + 589, + 890, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MVImgNet and MVPNet will be public, hoping to inspire the broader vision community.", + "bbox": [ + 496, + 695, + 890, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 736, + 640, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-view image datasets. The MNIST database [1] is one of the most pioneering datasets, composed of 70k monochrome images of handwritten digits. The CIFAR10 and CIFAR100 datasets proposed by [54], respectively collect 60k tiny color images $(32\\times 32)$ of various common objects or animals in 10 and 100 classes. ImageNet [24] is presented with a large scale, high accuracy, large diversity, and hierarchical structure, which provides opportunities for training deep neural networks. In detection", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9151", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and segmentation tasks, MSCOCO [58] is one of the most popular datasets containing 328k images with rich annotations. Some other datasets include PASCAL VOC [28], Visual Genome [53], Cityscapes [22], MPII [4], etc. Although these datasets facilitate the development of deep visual learning, they mainly serve for 2D single-view image understanding, which limits their applications in 3D vision.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video datasets. Another line is video datasets. Pioneering works construct the HMDB-51 [56] and UCF-101 [82] dataset. Afterward, the ActivityNet [8] and Kinetics [49] datasets are constructed of a larger scale and variation, of which the latter has collected 650k video clips that cover 700 classes. Besides, some datasets are built for human pose estimation [48, 90] and object detection/segmentation/tracking [69, 95, 112]. IEMOCAP [7] provides video data for the task of multimodal emotion recognition. MSVD [14] and MSR-VTT [103] annotate videos with extra captions. Further, HowTo100M [63] proposes a larger-scale one of 136 million samples from narrated instructional videos. ActivityNet Captions [52] introduces the task of dense-captioning events and constructs a dataset with 20k videos. These datasets are primarily for video understanding which is different from our objective.", + "bbox": [ + 75, + 199, + 472, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D datasets. As the applications in 3D vision attract increasing attention, various 3D datasets are proposed. One line of works focuses on indoor scenes [5, 10, 23, 47, 78, 81, 100], where S3DIS [5] and ScanNet [23] are two of most popular datasets. In addition, some works provide 3D object point clouds for contextual object surface reconstruction [6, 29]. 3D outdoor scenes are scanned via LiDAR sensors [9, 33, 37, 83] for autonomous driving. ShapeNet [11] and ModelNet [96] are two object-centric datasets that provide rich 3D Computer-Aided Design (CAD) models for shape analysis, followed by similar datasets [51, 98], which are usually low-quality, untextured, and have a domain gap with real-world objects. Another line of works [20, 21, 68, 79, 89, 104] advocate real 3D objects but are still limited in scale. We close this gap by shooting multi-view images of real-world objects, which capture the 3D awareness while allowing a scalable collection.", + "bbox": [ + 75, + 444, + 472, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-view image datasets. Multi-view image data is recently regarded as the source of 3D reconstruction or novel view synthesis. Early works collect multi-view images of real objects but only provide 3D models that are approximated [99] or for only a few instances [20]. Henzler et al. [45] contribute a larger video dataset to benchmark the task of 3D reconstruction. Another concurrent dataset, Objectron [3], annotates 3D bounding boxes and ground planes for all objects but lacks camera poses or point clouds for all 15k videos. GSO [27] gets clean 3D models with textures via scanning common household items but includes only a limited number of samples. Some works also construct synthetic multi-view datasets [11, 32, 86, 101, 108]. Neverthe", + "bbox": [ + 75, + 704, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0cc67aefea19b21664fb036c08620ef3f224fa3308c3818077fba07a467c05bd.jpg", + "image_caption": [ + "Figure 2. The efficient data acquisition pipeline of MVImgNet." + ], + "image_footnote": [], + "bbox": [ + 513, + 89, + 883, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "less, these datasets have a small scale and category range for special tasks, which limits the robust and generic learning of 3D deep models. We construct a large-scale multi-view image dataset - MVImgNet, which contains 219k videos for real-world objects in a wide range of 238 categories.", + "bbox": [ + 496, + 233, + 890, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Special discussions on CO3D [74]. Very recently, CO3D extends [45] to 19k videos covering 50 object categories. Although we share a similar idea in the light of object-centric multi-view data collections, our MVImgNet enjoys conspicuously larger scales.", + "bbox": [ + 496, + 309, + 890, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More importantly, we have fundamentally different motivations and insights. The primary motivation of CO3D is exactly pre-set, which is to transfer the 3D reconstruction training/evaluation from synthetic datasets into the real-world setup. In contrast, MVImgNet examines existing 3D datasets by re-walking the past development journey in the 2D domain. Concretely, through reflecting on how ImageNet [24] demonstrates its generic impact on data-hungry algorithms, we attempt to build a 3D counterpart of ImageNet and choose multi-view images as a soft bridge between 2D and 3D. Instead of nailing down a special target, we conduct a variety of pilot studies on how MVImgNet can benefit miscellaneous visual tasks, aiming to inspire and retain lots of possibilities for the broader vision community.", + "bbox": [ + 496, + 385, + 892, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the later experiments, our datasets indeed show greater power than CO3D on different visual challenges.", + "bbox": [ + 498, + 597, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Basis - MVImgNet Dataset", + "text_level": 1, + "bbox": [ + 500, + 641, + 787, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the whole data acquisition pipeline of MVImgNet is highly efficient, which is illustrated below.", + "bbox": [ + 498, + 667, + 890, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Raw Data Preparation", + "text_level": 1, + "bbox": [ + 500, + 708, + 709, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building a large-scale dataset is always challenging, due to not only the laborious data collection but also the nontrivial annotation, which is especially critical for real-world 3D data [5, 10, 23]. Thanks to the rapid development of mobile devices, shooting a video around an object in the wild becomes highly convenient and accessible in our daily life, which makes multi-view images be easily gained by crowd-sourcing.", + "bbox": [ + 496, + 731, + 890, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Composition setup. We set up some constraints on the ratio. Depending on the category, each class is initially set with the different expected number of video cap", + "bbox": [ + 496, + 856, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9152", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/de14d30fc652844d3d8af2dca54c443ff8fd882a3757fbf0fae7a90b4551b041.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AnnotationCollectedValidGPU hours
Sparse219,188215,7553,806.8
Segmentation104,261\\2,316.9
Dense98,89987,20025,122.4
", + "bbox": [ + 166, + 88, + 379, + 133 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/562369e864751c37a3d1bcf1022e8eafa7d503365c53f052752077b12873a1c7.jpg", + "table_caption": [ + "Table 1. Data statistics, including collected amount, valid amount after cleaning, and the GPU hours for processing." + ], + "table_footnote": [ + "Table 2. Comparison between our datasets and related ones. \"pcl\" denotes point clouds, \"bbox\" means bounding boxes." + ], + "table_body": "
DatasetReal# of objects# of categoriesMulti-view3D-GT
ShapeNet [11]×51k55renderCAD model
ModelNet [96]×12k40renderCAD model
Choi et al. [20]✓2k9360° capturedRGB-D scan
Objectron [3]✓15k9limited3D bbox, pcl
GoogleScan [27]✓2kNA360° capturedRGB-D scan
Henzler et al. [45]✓2k7360° capturedpcl
ScanObjectNN [88]✓14k15limitedpcl
CO3D [74]✓19k50360° captured-
CO3D-pcl [74]✓5k50360° capturedpcl
MVIImgNet (ours)✓220k238180° captured-
MVPNet (ours)✓87k150180° capturedpcl
", + "bbox": [ + 78, + 176, + 467, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tures regarding their generality, e.g., the number of captures for “bottles”, “bags” and “snacks” is planned to be around 2000, while the number of “chips”, “apples” and “guitars” is set to about 1000. This setting is closer to real life.", + "bbox": [ + 75, + 354, + 468, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video capture. How to capture videos directly affect the quality of our data, so we draw up the following requirements as guidance for the captured videos: 1) The length of each video should be about 10 seconds. 2) The frames in the video should not be blurred. 3) Each video should capture $180^{\\circ}$ or $360^{\\circ}$ view of the object as much as possible. 4) The proportion of the object in the video should be above $15\\%$ . 5) Each video can only contain one class of principal object. 6) The captured object should be stereoscopic.", + "bbox": [ + 75, + 417, + 468, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Crowdsourcing. We employ around 1000 normal collectors from different professions (e.g., teacher, doctor, student, cook, babysitter) and ages $(20\\sim 50)$ . Each of them is asked to take several videos in their daily life, (i.e., denoting diverse real-world environments) and upload them to the backend. Meanwhile, about 200 new well-trained expert data cleaners are responsible to review each submission and assure it fulfills the aforementioned capture requirements, when they may report some feedback or directly delete the unqualified submissions. The whole procedure ensures both the diversity and quality of the raw videos.", + "bbox": [ + 75, + 556, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Processing", + "text_level": 1, + "bbox": [ + 76, + 736, + 236, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each qualified video submission, we exploit an automatic process to obtain the common 2D and 3D annotations, including object masks, camera intrinsic and extrinsic, depth maps, and point clouds.", + "bbox": [ + 75, + 760, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse reconstruction. Following the procedure of [64, 65], the sparse reconstruction aims to reconstruct the camera intrinsic and extrinsic for each video, by applying the COLMAP Structure-from-Motion (SfM) algorithm [75] on a series of equal-time-interval chosen frames.", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b153f4934935b54f497c7f3de180a47c3741d5d3a51ea8e2c70c2ba963d5eb4f.jpg", + "image_caption": [ + "Figure 3. Some frames sampled from MVImgNet." + ], + "image_footnote": [], + "bbox": [ + 549, + 88, + 844, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Foreground object segmentation. Each frame extracted from the original video is fed to the CarveKit [70] for generating the binary foreground object mask, not only benefiting the dense reconstruction but also contributing to the further step of salient object detection (Sec. 5.3).", + "bbox": [ + 496, + 250, + 890, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dense reconstruction. With the sparse model output from COLMAP SfM, we employ multi-view stereo [76] of COLMAP to generate the dense depth and normal maps for each frame. We extract the depths of an object using the binary foreground masks, which are then back-projected and fused according to the normal information, yielding a densely reconstructed point cloud for each video. Finally, the point cloud is manually cleaned by: 1) Delete the object with obvious noisy or extremely sparse reconstruction. 2) Remove all backgrounds. The final derived 3D point cloud dataset - MVPNet, is illustrated in Sec. 6.1.", + "bbox": [ + 496, + 327, + 892, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Summary", + "text_level": 1, + "bbox": [ + 500, + 503, + 671, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Statistics. The statistics of MVImgNet are shown in Tab. 1, and Tab. 2 compares our datasets with other alternatives. MVImgNet includes 238 object classes, from 6.5 million frames of 219,188 videos. Fig. 3 shows some samples of MVImgNet. The annotation comprehensively covers object masks, camera parameters, and point clouds.", + "bbox": [ + 496, + 529, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Category. We leverage WordNet [66] taxonomy that is used by ImageNet [24] to describe multi-hierarchy categories of objects and define 238 common classes. Unlike ImageNet which contains various plants and animals (nature-centric), the objects in our MVImgNet are found or used in human daily life (human-centric), where 65 classes overlap with ImageNet. The detailed category taxonomy, per-category data distributions, and more sample visualizations of MVImgNet are illustrated in the supplementary material.", + "bbox": [ + 496, + 623, + 890, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3D Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 789, + 679, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Radiance Field Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 814, + 772, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-review. Recently, a series of generalizable Neural Radiance Fields (NeRF) variants [13, 87, 93, 111] have been proposed to reconstruct radiance field on-the-fly from one/few-shot source views for novel view synthesis.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9153", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e719751f302dc60027b387e478ddd7ff891b0970fdce2fb7945f4bbee86b9031.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDiffuse PSNR↑Synthetic SSIM↑360° LPIPS↓Realistic PSNR↑Synthetic SSIM↑360° LPIPS↓Real-world PSNR↑360° SSIM↑Objects LPIPS↓
Train from scratch37.170.9900.01725.490.9160.10022.180.7140.365
MVIImgNet-pretrained37.660.9900.01427.260.9300.07124.670.7400.310
", + "bbox": [ + 83, + 92, + 580, + 141 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9b3f7d3a9720c9780347710804d13415406741139420bef2bed17954f2d0ac55.jpg", + "table_caption": [ + "Table 3. NeRF quantitative results on three different levels of test sets." + ], + "table_footnote": [], + "table_body": "
Where to pretrainReal-world 360°Objects [74]
PSNR↑SSIM↑LPIPS↓
CO3D [74]24.010.7320.339
MVIImgNet-small24.080.7360.316
MVIImgNet24.670.7400.310
", + "bbox": [ + 640, + 89, + 859, + 145 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4. NeRF quantitative comparison with CO3D [74].", + "bbox": [ + 581, + 148, + 918, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8cf5e9928a1762f9647ef48dedd1dd33aa4f416e2e8dc4f58b16937c3286e65e.jpg", + "image_caption": [ + "Figure 4. NeRF qualitative comparison of train-from-scratch IBRNet, CO3D [74]-pretrained IBRNet and MVImgNet-pretrained IBRNet [93]." + ], + "image_footnote": [], + "bbox": [ + 89, + 167, + 450, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "What can MVImgNet do? Training NeRFs that can generalize to unseen objects requires learning 3D priors from a huge amount of multi-view images. Existing state-of-the-arts either resort to learning on large-scale synthetic data [13, 87, 111], or adopt a mixed use of synthetic data and a small self-collected real dataset [93]. However, the synthetic data introduces a big domain gap with real-world objects, only a few real scenes cannot remedy this defect. We argue that: our MVImgNet perfectly fits the huge data demand of learning-based generalizable NeRF methods.", + "bbox": [ + 75, + 354, + 468, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify this, we choose IBRNet [93] as the baseline and conduct an empirical study. We pretrain IBRNet on the full MVImgNet dataset then finetune on the training datasets used in IBRNet [93] for a few iterations, and compare with the original train-from-scratch IBRNet model in terms of generalization capability (i.e., generalizing to unseen scenes with only few-shot inputs). For fairness, we evaluate all methods under the same protocol of IBRNet.", + "bbox": [ + 75, + 506, + 468, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here comes a challenge about how to evaluate the generalization ability of NeRFs, since there is no official benchmark for this. To this end, we employ three different third-party object-centric datasets to form the test set. 1) The diffuse synthetic $360^{\\circ}$ object dataset [80] which contains 4 Lambertian objects. 2) The realistic synthetic $360^{\\circ}$ object dataset [64] which consists of 8 realistically fabricated objects. 3) The real-world $360^{\\circ}$ object dataset [74] which includes 88 real-world scenes from different lighting conditions. To summarize, the whole test set comprises 100 objects from 56 distinct categories, ranging from the synthetic domain to the real-world domain, which is considered to be an impartial evaluation set for generalization ability.", + "bbox": [ + 75, + 628, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The quantitative and qualitative results are respectively shown in Tab. 3 and Fig. 4. Evidently, pretraining on MVImgNet improves the generalization ability of the model by a large margin. Moreover, we perform the same pretraining on CO3D [74] and MVImgNet-small (a random subset", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1089e7076a63296b6f70a630ede941cf05604a67a5ae71d7596330049dce03e4.jpg", + "image_caption": [ + "Figure 5. PSNR of IBRNet pretrained by a different number of MVImgNet data (higher is better)." + ], + "image_footnote": [], + "bbox": [ + 562, + 170, + 830, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of MVImgNet, which owns the same scale as CO3D). As shown in Tab. 4 and Fig. 4, MVImgNet shows greater power than CO3D [74]. The implementation details can be found in the supplementary material.", + "bbox": [ + 496, + 316, + 892, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More data, more power. Fig. 5 shows that an apparent rising trend of generalization metrics can be observed with the increase of training data.", + "bbox": [ + 496, + 378, + 893, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Multi-view Stereo", + "text_level": 1, + "bbox": [ + 500, + 434, + 671, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-review. Multi-view Stereo (MVS) [30] is a classical task in 3D computer vision, with the goal of reconstructing 3D geometry from multi-view images. Conventional methods [31,46,77,114] reconstruct 3D geometry by finding the patches matched with different images and estimating the depth according to the extrinsic camera. In recent years, deep learning methods are introduced into MVS to solve the issues such as weak textures, and nonlaplacian spheres. Two popular end-to-end methods MVSNet [109] and R-MVSNet [110] propose to encode the multi-view images and build a cost volume for predicting depth maps. Under the supervision of large-scale RGB-D datasets [2], they outperform conventional patch-matched-based approaches. However, these methods require massive amounts of RGB-D data, which is always difficult to acquire. This drives the emergence of self-supervised MVS methods [12,25,102,107].", + "bbox": [ + 496, + 460, + 892, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "What can MVImgNet do? We demonstrate that our MVImgNet is capable of benefiting the data-efficient MVS with limited training examples, which is practically meaningful considering MVS always requires the burdensome collection of RGB-D data.", + "bbox": [ + 496, + 719, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We pretrain a self-supervised MVS method, JDACS [102], on MVImgNet. Our implementation follows the original settings of JDACS. Then, we select DTU [2] dataset with 79 training samples and 22 test samples. We perform the data-efficient evaluation with limited training data, where the MVImgNet-pretrained MVsNet is finetuned on $5\\%$ , $15\\%$ , and $25\\%$ of DTU training samples, and eval", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9154", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/191585e7ae228e41128c6d5014d434a52ccdc60c6843314498da9eb06330311c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ratio2mm↑4mm↑8mm↑
5%48.61 / 50.1065.18 / 67.9775.69 / 78.58
15%54.74 / 54.6870.96 / 72.0480.32 / 81.73
25%57.29 / 58.6374.41 / 75.2083.68 / 84.28
", + "bbox": [ + 148, + 88, + 400, + 140 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2c83d1aaa41c5e0e03c2fb0de4f33113920e45f39017f824a5d37e4212bf4be9.jpg", + "table_caption": [ + "Table 5. MVS depth map accuracy on DTU evaluation set under different ratios of DTU training samples, in terms of training from scratch / MVImgNet-pretrained (higher is better)." + ], + "table_footnote": [], + "table_body": "
Where to pretrain2mm↑4mm↑8mm↑
CO3D46.1866.7278.39
MVIImgNet-small47.7366.5378.15
MVIImgNet50.1067.9778.58
", + "bbox": [ + 153, + 193, + 393, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "uated on the DTU test set. Tab. 5 reports the accuracy given different thresholds of the error between the predicted and ground truth depth map, which indicates that the MVImgNet-pretrained model is capable of improving the model trained from scratch by a large margin under the data-efficient setup. Furthermore, we perform pretraining using CO3D [74] and MVImgNet-small. Tab. 6 indicates that our MVImgNet is stronger than CO3D, and MVImgNet-small shows comparable power as CO3D. For more implementation details, please see the supplementary material.", + "bbox": [ + 75, + 306, + 470, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We advocate benchmarking NeRF and MVS methods with the help of pretraining on MVImgNet.", + "bbox": [ + 76, + 458, + 468, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. View-consistent Image Understanding", + "text_level": 1, + "bbox": [ + 76, + 503, + 419, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. View-consistent Image Classification", + "text_level": 1, + "bbox": [ + 76, + 527, + 393, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-review. As indicated by Dong et al. [26], although humans can easily recognize one object from different views, deep learning models can hardly do that robustly. MVImgNet provides numerous images from different viewpoints, so we hope to enhance the model's view consistency with the help of MVImgNet, which is significantly important for realizing human-like intelligence.", + "bbox": [ + 75, + 551, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What can MVImgNet do? One naive approach is to finetune the ImageNet-pretrained ResNet-50 [44] on our MVImgNet. However, such an approach will be problematic due to the categories of two datasets are very distinct, which may cause catastrophic forgetting issues [50].", + "bbox": [ + 75, + 657, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For this reason, we create a new training set, namely MVI-Mix, by mixing the original ImageNet data with MVIImgNet. Specifically, we randomly sample 5 consecutive frames of each video in MVIImgNet, and mix the multi-view images with original ImageNet data. We also build another two artificial training sets, namely MVI-Gap and MVI-Aug, that only differ MVI-Mix with image views. MVI-Gap samples the 5 non-consecutive frames that have much larger view differences. MVI-Aug samples 1 frame from each video of MVI-Mix and applies 4 different data augmentations (i.e., random color jittering, grid mask, ro", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bf877891f99ee5498dcf221a637f78cd142cfe01dd6ecfee3dfc8b3712cbc891.jpg", + "table_caption": [ + "Table 6. MVS depth map accuracy comparison with CO3D [74] on DTU evaluation set. Pretrain on MVImgNet, MVImgNet-small or CO3D, and finetune using $5\\%$ of DTU training samples." + ], + "table_footnote": [], + "table_body": "
DatasetConfidence VarAccuracy
ImageNet-only0.20753.09%
MVI-Aug0.10571.48%
MVI-Gap0.10377.23%
MVI-Mix0.10277.31%
", + "bbox": [ + 588, + 88, + 803, + 148 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3e8ea823817d7ade858f0d548af5bc015147a2c803caed94c16896bdb83c6e5.jpg", + "table_caption": [ + "Table 7. View-consistent image classification results on MVImgNet test set using fully-supervised ResNet-50 [44]. Adding MVImgNet for training improves the view consistency (smaller Var and higher Acc indicate better view consistency)." + ], + "table_footnote": [], + "table_body": "
Train: CO3DI-MixTrain: MVI-Mix
Test: CO3DVar: 0.104Var: 0.155 (+0.051)
Acc: 91.25%Acc: 84.83% (-6.42%)
Test:MVIImgNetVar: 0.188 (+0.086)Var: 0.102
Acc: 57.93% (-19.38%)Test: 77.31%
", + "bbox": [ + 560, + 218, + 834, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 8. View-consistent image classification comparison with CO3D [74]. The performance drops when train on CO3DI-Mix and test on MVIImgNet (blue number) are much larger than the opposite (red number), which means that MVI-Mix pretrained model is more robust to multiview consistency than CO3DI-Mix pretrained model.", + "bbox": [ + 496, + 277, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tation and erase, respectively) to it, which aims to differentiate the multi-view augmentation from the normal data augmentations. We compare the variance of the softmax confidence and the accuracy on MVImgNet test set for examining the view consistency. As illustrated in Tab. 7, adding our MVImgNet data for training can effectively improve the model's view consistency and accuracy. The test dataset used for all experiments is MVImgNet test dataset. In addition to using the convolution-based network (i.e., ResNet-50), we also apply the Transformer-based architecture, DeiT-Tiny [85]. The results are $48.76\\%$ accuracy and 0.225 Var on ImageNet-only, VS 0.122 Var $(\\mathbf{0.103} \\downarrow)$ and $73.88\\%$ $(25.12 \\uparrow)$ accuracy on MVI-Mix. This leads to a unanimous conclusion.", + "bbox": [ + 496, + 378, + 890, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further construct CO3DI-Mix (corresponding to MVI-Mix) as the mixer of CO3D [74] & ImageNet under the same ratio of MVImgNet & ImageNet in MVI-Mix. As illustrated in Tab. 8, MVImgNet brings more benefits than CO3D for view-consistent image recognition.", + "bbox": [ + 496, + 590, + 890, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. View-consistent Contrastive Learning", + "text_level": 1, + "bbox": [ + 500, + 679, + 823, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-review. Contrastive Learning (CL) is one mainstream of self-supervised training techniques [16-19, 35, 40, 84]. One key factor of CL is the construction of positive/negative pairs, e.g., MoCo-v2 [17] treats an image with different augmentations, a.k.a. views, as the positive pair.", + "bbox": [ + 496, + 703, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What can MVImgNet do? One natural question is: can the viewpoints of MVImgNet serve as the positive pairs for $CL$ ? To answer this question, we finetune the off-the-shelf ImageNet-pretrained MoCo-v2 on MVImgNet. For each iteration, we randomly sample two frames from the same video as the positive pair and apply the original data augmentations used in MoCo-v2 [17]. Meanwhile, the frames from other videos will be treated as their negatives.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9155", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a5a8621e069763890cc7258e080873e3f2436adb703920fb3047901783ee983c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training schemeConfidence VarAccuracy
w/o finetune0.09870.26%
w/ finetune0.08671.22%
", + "bbox": [ + 147, + 88, + 403, + 133 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 9. View-consistency classification results under the self-supervised contrastive learning regime on MVImgNet test set. Finetuning MoCo-v2 [17] on MVImgNet improves the view consistency.", + "bbox": [ + 76, + 137, + 468, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8fca2ba8a21e7fee99ee8035da0a5ba775212954ee03eea21d8d4e62f0d9c79b.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison between MVImgNet-finetuned U2Net [73] and original U2Net for salient object detection (SOD). Left: finetuning on MVImgNet improves the performance on a hard view. Right: finetuning on MVImgNet improves the performance on two consecutive hard views." + ], + "image_footnote": [], + "bbox": [ + 98, + 202, + 450, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The MVImgNet-finetuned model is finally evaluated on the MVImgNet test dataset for examining the view consistency, where we compare the variance and mean of the softmax confidence and the accuracy. As Tab. 9 shows, finetuning on MVImgNet with CL can also improve both the model's view consistency and accuracy. More implementation details of view-consistent image classification is described in the supplementary material.", + "bbox": [ + 75, + 383, + 468, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the future, it is highly recommended to regard view-consistency as a criterion for evaluating the image recognition task, and utilize MVImgNet to pretrain the models.", + "bbox": [ + 76, + 506, + 468, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. View-consistent SOD", + "text_level": 1, + "bbox": [ + 76, + 560, + 274, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pre-review. Salient Object Detection (SOD) aims to segment the most visually prominent objects in an image. Although remarkable progress has been made recently, it remains lots of challenges.", + "bbox": [ + 76, + 583, + 468, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What can MVImgNet do? We test a state-of-the-art SOD model U2Net [73] on our MVImgNet. As Fig. 6 shows, U2Net failed to segment \"hard\" views, even though some views can be segmented with few flaws. Despite such a disappointment, the inconsistent predictions of different views caught our attention: can we improve the SOD models with multi-view consistency?", + "bbox": [ + 75, + 643, + 468, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose to leverage the multi-view consistency to improve SOD with the help of optical flows. Specifically, given two consecutive frames, we first calculate their optical flow and warp the flow to the segmentation mask of one of the frames. Then, a consistency loss can be calculated between the warped mask and the mask of another frame. For ease of implementation, we directly finetune U2Net on our MVImgNet. To prevent the catastrophic forgetting issue [50], in addition to the MVImgNet, we also use the original training data DUTS-TR [92]. For the \"hard\" views", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40b0bba5864a6bb332e007b8ce09103b4297013eedba4fcb071aabe23123c508.jpg", + "image_caption": [ + "Figure 7. Some 3D point clouds sampled from MVPNet." + ], + "image_footnote": [], + "bbox": [ + 540, + 89, + 851, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$(\\mathrm{IoU} \\leq 0.7)$ on our MVImgNet test set, the model finetuned on our MVImgNet can bring a $4.1\\%$ IoU improvement (see Fig. 6 for qualitative comparison). The implementation details of view-consistent SOD can be found in the supplementary material.", + "bbox": [ + 498, + 226, + 890, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. MVPNet for 3D Understanding", + "text_level": 1, + "bbox": [ + 500, + 316, + 785, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. MVPNet Dataset", + "text_level": 1, + "bbox": [ + 500, + 340, + 666, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Derived from the dense reconstruction on MVImgNet (as mentioned in Sec. 3.2), a new large-scale real-world 3D object point cloud dataset – MVPNet, is born, which contains 87,200 point clouds with 150 categories. Fig. 7 shows some examples of MVPNet. As listed in Tab. 2, compared with existing 3D object datasets, our MVPNet contains a conspicuously richer amount of real-world object point clouds, with abundant categories covering many common objects in the real life. The detailed category taxonomy per-category data distributions of MVPNet are illustrated in the supplementary material.", + "bbox": [ + 496, + 364, + 890, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.3D Point Cloud Classification", + "text_level": 1, + "bbox": [ + 500, + 540, + 761, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we focus on point cloud classification. We believe that future works may also utilize our dataset to help much more 3D understanding tasks such as indoor-scene parsing, outdoor-environment perception, pose estimation, and robotics manipulation.", + "bbox": [ + 496, + 563, + 890, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pre-review. We advocate paying more attention to real-world setup, which is more feasible for real applications. ScanObjectNN [89] has been manifested as the most challenging point cloud classification benchmark so far, so we choose it for the major comparison with our MVPNet.", + "bbox": [ + 496, + 641, + 890, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What can MVPNet do? We show that pretraining on MVPNet is able to aid the performance of real-world point cloud classification [36, 39, 62, 71, 72, 94, 97, 105, 106]. We pretrain several models on MVPNet, and finetune them on ScanObjectNN.", + "bbox": [ + 496, + 719, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two settings are considered for evaluation. First is PB_T50_RS in ScanObjectNN with small perturbation, translation and rotation on point cloud. Another is adding heavy rotation on PB_T50_RS, to create a more challenging setting. The results are shown in Tab. 10, where pretraining on MVPNet is able to increase the classification accuracy under most circumstances.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9156", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ffff2dac759f99f6085ff3404e5c238e68a6105f5e250a454b2f368236476c70.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAdd Random RotationPB_T50_RS
from scratchpretrainedfrom scratchpretrained
PointNet [71]60.57 / 55.2064.25 / 59.2970.63 / 67.2867.73 / 64.12
PointNet++ [72]76.50 / 73.4278.76 / 76.5478.80 / 75.7080.22 / 76.91
DGCNN [94]80.50 / 78.4580.42 / 78.2079.44 / 76.2482.36 / 80.08
PointMLP [62]83.69 / 82.5484.87 / 83.7185.64 / 84.1485.98 / 84.38
CurveNet [97]73.96 / 69.9678.99 / 76.5974.27 / 69.4383.68 / 81.17
GDANet [106]80.33 / 79.1483.59 / 82.2979.01 / 75.9183.90 / 82.51
PAConv [105]70.91 / 65.7076.21 / 72.4772.88 / 68.6076.91 / 73.45
PCT [36]81.32 / 79.1782.08 / 80.4177.46 / 73.6484.20 / 81.94
PointMAE [67]83.17 / 80.7586.19 / 84.6077.34 / 73.5284.13 / 81.92
", + "bbox": [ + 101, + 88, + 446, + 204 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b3fd1156db226393524886775b225b1000d1e56f4532d742a3c657dba2987a6a.jpg", + "table_caption": [ + "Table 10. ScanObjectNN real-world point cloud classification results. The comparison is between the train-from-scratch model and MVPNet pretrained model. The metric is overall / average accuracy." + ], + "table_footnote": [], + "table_body": "
MethodPB_T50_RS
PreTr. on MVPNetPreTr. on MVPNet-smallPreTr. on CO3D
PointNet++78.90 / 77.1179.00 / 76.7878.79 / 77.20
CurveNet83.68 / 81.1773.92 / 68.9273.78 / 69.62
PointMAE84.13 / 81.9281.85 / 79.2581.40 / 78.87
", + "bbox": [ + 119, + 268, + 426, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Self-supervised Point Cloud Pretraining", + "text_level": 1, + "bbox": [ + 76, + 391, + 419, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pre-review. Self-supervised learning has been exploited for 3D object point cloud understanding [59,67,91,113].", + "bbox": [ + 76, + 417, + 467, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What can MVPNet do? In Tab. 10, PointMAE [67] pretrained on our MVPNet outperforms the state-of-the-art methods when finetuned on ScanObjectNN, proving the benefit of MVPNet on the self-supervised learning regime for real-world point cloud classification. Tab. 11 also shows that pretraining on MVPNet is more powerful than CO3D [74] for real-world point cloud classification.", + "bbox": [ + 76, + 449, + 468, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. MVPNet Benchmark Challenge", + "text_level": 1, + "bbox": [ + 76, + 564, + 356, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the MVPNet benchmark challenge for real-world point cloud classification, which contains 69,760 training and 17,440 testing samples. The results of various methods are shown in Tab. 12.", + "bbox": [ + 76, + 587, + 468, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MVPNet is more challenging than ScanObjectNN. We firstly train models on ScanObjectNN, then conduct the test on MVPNet, which is concluded in Tab. 13. On the opposite, we also train on MVPNet and test on ScanObjectNN, which is listed in Tab. 14. Comparing Tab. 13 and Tab. 14, the accuracy drops are significantly larger when training on ScanObjectNN and testing on MVPNet, which verifies the greater challenge of our MVPNet. All the experiments in 3D understanding strictly follow the original settings of the selected backbone networks.", + "bbox": [ + 76, + 650, + 468, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MVPNet is suggested for investigating 3D point cloud understanding in the future.", + "bbox": [ + 76, + 801, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced MVImgNet, a large-scale dataset of multi-view images, which is efficiently collected by shoot", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4b4ffc7c0eb70ff3f42327610f69272f70b21a148933740a1b90d78a2d180b78.jpg", + "table_caption": [ + "Table 11. ScanObjectNN real-world point cloud classification comparison with CO3D [74]. Pretrain on MVImgNet, MVImgNet-small or CO3D, and finetune on ScanObjectNN. The metric is overall / average accuracy." + ], + "table_footnote": [], + "table_body": "
MethodOverall / Average Accuracy
PointNet [71]70.72 / 54.46
PointNet++ [72]79.15 / 58.24
DGCNN [94]86.49 / 63.98
PointMLP [62]88.89 / 73.64
CurveNet [97]88.88 / 75.37
GDANet [106]89.54 / 68.41
PACConv [105]83.35 / 59.13
PCT [36]91.49 / 75.41
", + "bbox": [ + 589, + 88, + 802, + 184 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a4afc91ba65f48cc14cc323ea8127e3bfc4eb1bcbed8fd5f93b90084bd372c67.jpg", + "table_caption": [ + "Table 12. Quantitative results on our new MVPNet benchmark for real-world point cloud classification." + ], + "table_footnote": [], + "table_body": "
Methodtrain: MVPNet test: MVPNettrain: ScanObjectNN test: MVPNet
PointNet++ [72]79.15 / 58.2415.13 (-64.02) / 7.96 (-50.28)
CurveNet [97]88.88 / 75.3746.95 (-41.93) / 28.24 (-47.13)
", + "bbox": [ + 549, + 224, + 843, + 272 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/275ca20a8c4cf7fffb7e3d924a2ef877ab41570ac28ca717efad266bfa11ab37.jpg", + "table_caption": [ + "Table 13. Quantitative comparison while training on ScanObjectNN, testing on MVPNet. The evaluation metric is overall / average accuracy." + ], + "table_footnote": [], + "table_body": "
Methodtrain: ScanObjectNN\ntest: ScanObjectNNtrain: MVPNet\ntest: ScanObjectNN
PointNet++ [72]76.50 / 73.4240.35 (-36.15) / 33.59 (-39.83)
CurveNet [97]73.96 / 69.9651.84 (-22.12) / 46.27 (-23.69)
", + "bbox": [ + 542, + 319, + 849, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 14. Quantitative comparison while training on MVPNet, testing on ScanObjectNN. The evaluation metric is overall / average accuracy.", + "bbox": [ + 500, + 369, + 890, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing videos of real-world objects. The multi-view nature endows our dataset with 3D-aware visual signals, making MVImgNet a soft bridge to link 2D and 3D vision. To probe the power of MVImgNet, we conduct a host of pilot experiments on various visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising effectiveness, expecting more future explorations. As a bonus of MVImgNet, a point cloud dataset - MVPNet is derived. Experiments show MVPNet can benefit real-world 3D object classification. As a broader impact on society, our datasets delineates a world - that is closer to a colorful and vivid real $3D$ world - where we human lives.", + "bbox": [ + 496, + 421, + 890, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 632, + 656, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work was supported in part by NSFC with Grant No. 62293482, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone. It was also partially supported by NSFC-62172348, Outstanding Yound Fund of Guangdong Province with No. 2023B1515020055 and Shenzhen General Project with No. JCYJ20220530143604010, the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). It was also sponsored by CCF-Tencent Open Research Fund.", + "bbox": [ + 496, + 657, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9157", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gradient-based learning applied to document recognition, 1998. 2", + "[2] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 2016. 5", + "[3] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In CVPR, 2021. 3, 4", + "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, 2014. 3", + "[5] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 3", + "[6] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In CVPR, 2019. 3", + "[7] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 2008. 3", + "[8] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3", + "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3", + "[10] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 2, 3", + "[11] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3, 4", + "[12] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rcomvsnet: Unsupervised multi-view stereo with neural rendering. arXiv preprint arXiv:2203.03949, 2022. 5", + "[13] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 4, 5", + "[14] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In The 49th annual meeting of the association for computational linguistics: human language technologies, 2011. 3" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2", + "[16] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 6", + "[17] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 6, 7", + "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 6", + "[19] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 6", + "[20] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 3, 4", + "[21] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 3", + "[22] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3", + "[23] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3", + "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 3, 4", + "[25] Yikang Ding, Qingtian Zhu, Xiangyue Liu, Wentao Yuan, Haotian Zhang, and Chi Zhang. Kd-mvs: Knowledge distillation based self-supervised learning for multi-view stereo. In ECCV, 2022. 5", + "[26] Yinpeng Dong, Shouwei Ruan, Hang Su, Caixin Kang, Xingxing Wei, and Jun Zhu. Viewfool: Evaluating the robustness of visual recognition to adversarial viewpoints. arXiv preprint arXiv:2210.03895, 2022. 2, 6", + "[27] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv preprint arXiv:2204.11918, 2022. 3, 4", + "[28] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2015. 3", + "[29] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In ICCV, 2021. 3", + "[30] Yasutaka Furukawa and Carlos Hernández. Multi-view stereo: A tutorial. Foundations and Trends in Computer Graphics and Vision, 2015. 5" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9158", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] David Gallup, Jan-Michael Frahm, Philippos Mordohai, and Marc Pollefeys. Variable baseline/resolution stereo. In CVPR, 2008. 5", + "[32] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In CVPR, 2022. 3", + "[33] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 3", + "[34] Ross Girshick. Fast r-cnn. In ICCV, 2015. 2", + "[35] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In NeurIPS, 2020. 6", + "[36] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 2021. 7, 8", + "[37] Timo Hackel, Nikolay Savinov, Luber Ladicky, Jan Wegner, Konrad Schindler, and Marc Pollefeys. Semantic3d.net: A new large-scale point cloud classification benchmark. In ISPRS, 2017. 2, 3", + "[38] Abdullah Hamdi and Bernard Ghanem. Towards analyzing semantic robustness of deep neural networks. arXiv preprint arXiv:1904.04621, 2019. 2", + "[39] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In ICCV, 2021. 7", + "[40] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, June 2020. 6", + "[41] Kaiming He, Ross Girshick, and Piotr Dólar. Rethinking imagenet pre-training. arXiv preprint arXiv:1811.08883, 2018.2", + "[42] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2", + "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification. In ICCV, 2015. 2", + "[44] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 6", + "[45] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In CVPR, 2021. 3, 4", + "[46] Xiaoyan Hu and Philippos Mordohai. Least commitment, viewpoint-based, multi-view stereo. In International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission, 2012. 5", + "[47] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 3DV, 2016. 3" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, 2013. 3", + "[49] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 3", + "[50] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 6, 7", + "[51] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In CVPR, 2019. 3", + "[52] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 3", + "[53] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 3", + "[54] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2", + "[55] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 2", + "[56] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 3", + "[57] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. arXiv preprint arXiv:2108.10257, 2021. 2", + "[58] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3", + "[59] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 8", + "[60] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2", + "[61] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 2", + "[62] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 7, 8", + "[63] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic." + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9159", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3", + "[64] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 5", + "[65] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Trans. Graph., 2019. 4", + "[66] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 4", + "[67] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 8", + "[68] Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M Seitz. Photoshop: photorealistic materials for large-scale shape collections. ACM Trans. Graph., 2018. 3", + "[69] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 3", + "[70] Open Source Project. CarveKit - image-background-remove-tool. 4", + "[71] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7, 8", + "[72] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 7, 8", + "[73] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 2020. 7", + "[74] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 3, 4, 5, 6, 8", + "[75] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 4", + "[76] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 4", + "[77] Shuhan Shen. Accurate multiple view 3d reconstruction using patch-based stereo for large-scale scenes. TIP, 2013. 5", + "[78] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3", + "[79] Arjun Singh, James Sha, Karthik S Narayan, Tudor Achim, and Pieter Abbeel. Bigbird: A large-scale 3d database of object instances. In ICRA, 2014. 3" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, 2019. 5", + "[81] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 2, 3", + "[82] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3", + "[83] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 3", + "[84] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. 6", + "[85] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 6", + "[86] Jonathan Tremblay, Moustafa Meshry, Alex Evans, Jan Kautz, Alexander Keller, Sameh Khamis, Charles Loop, Nathan Morrical, Koki Nagano, Towaki Takikawa, et al. Rtmv: A ray-traced multi-view synthetic dataset for novel view synthesis. arXiv preprint arXiv:2205.07058, 2022. 3", + "[87] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, 2021. 4, 5", + "[88] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 4", + "[89] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 3, 7", + "[90] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3", + "[91] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 8", + "[92] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 7", + "[93] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 4, 5", + "[94] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + ], + "bbox": [ + 509, + 90, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9160", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graph cnn for learning on point clouds. ACM Trans. Graph., 2019. 7, 8", + "[95] Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Online object tracking: A benchmark. In CVPR, 2013. 3", + "[96] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 3, 4", + "[97] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In ICCV, 2021. 7, 8", + "[98] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In ECCV, 2016. 3", + "[99] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond pascal: A benchmark for 3d object detection in the wild. In WACV, 2014. 3", + "[100] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In ICCV, 2013. 3", + "[101] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. IJCV, 2020. 3", + "[102] Hongbin Xu, Zhipeng Zhou, Yu Qiao, Wenxiong Kang, and Qiuxia Wu. Self-supervised multi-view stereo via effective co-segmentation and data-augmentation. In AAAI, 2021. 5", + "[103] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 3", + "[104] Mutian Xu, Pei Chen, Haolin Liu, and Xiaoguang Han. Toscane: A large-scale dataset for understanding 3d tabletop scenes. In ECCV, 2022. 2, 3", + "[105] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 7, 8", + "[106] Mutian Xu, Junhao Zhang, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi, and Yu Qiao. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In AAAI, 2021. 7, 8", + "[107] Jiayu Yang, Jose M Alvarez, and Miaomiao Liu. Self-supervised learning of depth inference for multi-view stereo. In CVPR, 2021. 5", + "[108] Zhenpei Yang, Zaiwei Zhang, and Qixing Huang. Hm3d-abo: A photo-realistic dataset for object-centric multi-view 3d reconstruction. arXiv preprint arXiv:2206.12356, 2022. 3", + "[109] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 5", + "[110] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In CVPR, 2019. 5", + "[111] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 4, 5" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[112] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 3", + "[113] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 8", + "[114] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In CVPR, 2014. 5" + ], + "bbox": [ + 501, + 92, + 890, + 246 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9161", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_model.json b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..33d08ef6e2054e9beafd1bd076133a8b864bdafb --- /dev/null +++ b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_model.json @@ -0,0 +1,3084 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.198, + 0.131, + 0.773, + 0.154 + ], + "angle": 0, + "content": "MVImgNet: A Large-scale Dataset of Multi-view Images" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.179, + 0.87, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.341, + 0.891, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.534, + 0.893, + 0.605 + ], + "angle": 0, + "content": "Figure 1. It is time to embrace MVImgNet! We introduce MVImgnet, a large-scale dataset of multi-view images, which is efficiently collected by shooting videos of real-world objects. It enjoys 3D-aware signals from multi-view consistency, being a soft bridge between 2D and 3D vision. Through dense reconstruction on MVImgNet, we also present a large-scale real-world 3D object point cloud dataset – MVPNet. Exterior: Examples of various multi-view images in MVImgNet (see Fig. 3 more intuitively). Interior: Instances of colorful point clouds from MVPNet (see Fig. 7 more clearly) are assembled into a stereo sign of 'MVImgNet'." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.617, + 0.314, + 0.634 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Being data-driven is one of the most iconic properties of deep learning algorithms. The birth of ImageNet [24] drives a remarkable trend of 'learning from large-scale data' in computer vision. Pretraining on ImageNet to obtain rich universal representations has been manifested to benefit various 2D visual tasks, and becomes a standard in 2D vision. However, due to the laborious collection of real-world 3D data, there is yet no generic dataset serving as a counterpart of ImageNet in 3D vision, thus how such a dataset can impact the 3D community is unraveled. To remedy this defect, we introduce MVImgNet, a large-scale dataset of multi-view images, which is highly convenient to gain by shooting videos of real-world objects in human daily life. It contains 6.5 million frames from 219,188 videos crossing objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.618, + 0.892, + 0.648 + ], + "angle": 0, + "content": "The multi-view attribute endows our dataset with 3D-aware signals, making it a soft bridge between 2D and 3D vision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.752 + ], + "angle": 0, + "content": "We conduct pilot studies for probing the potential of MVImgNet on a variety of 3D and 2D visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising performance, remaining lots of possibilities for future explorations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.894, + 0.854 + ], + "angle": 0, + "content": "Besides, via dense reconstruction on MVImgNet, a 3D object point cloud dataset is derived, called MVPNet, covering 87,200 samples from 150 categories, with the class label on each point cloud. Experiments show that MVPNet can benefit the real-world 3D object classification while posing new challenges to point cloud understanding." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.864, + 0.892, + 0.894 + ], + "angle": 0, + "content": "MVImgNet and MVPNet will be public, hoping to inspire the broader vision community." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9150" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.208, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.468, + 0.326 + ], + "angle": 0, + "content": "Being data-driven, also known as data-hungry, is one of the most important attributes of deep learning algorithms. By training on large-scale datasets, deep neural networks are able to extract rich representations. In the past few years, the computer vision community has witnessed the bloom of such 'learning from data' regime [43, 44, 55], after the birth of ImageNet [24] – the pioneer of large-scale real-world image datasets. Notably, pretraining on ImageNet is well-proven to boost the model performance when transferring the pretrained weights into not only high-level [34, 41, 42, 60, 61] but also low-level visual tasks [15, 57], and becomes a de-facto standard in 2D. Recently, various 3D datasets [5, 9, 11, 23, 33, 96, 104] are produced to facilitate 3D visual applications." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.328, + 0.468, + 0.614 + ], + "angle": 0, + "content": "However, due to the non-trivial scanning and laborious labeling of real-world 3D data (commonly organized in point clouds or meshes), existing 3D datasets are either synthetic or their scales are not comparable with ImageNet [24]. Consequently, unlike in 2D vision where models are usually pretrained on ImageNet to gain universal representation or commonsense knowledge, most of the current methods in 3D area are directly trained and evaluated on particular datasets for solving specific 3D visual tasks (e.g., NeRF dataset [64] and ShapeNet [11] for novel view synthesis, ModelNet [96] and ScanObjectNN [88] for object classification, KITTI [33] and ScanNet [23] for scene understanding). Here, two crucial and successive issues can be induced: (1) There is still no generic dataset in 3D vision, as a counterpart of ImageNet in 2D. (2) What benefit such a dataset can endow to 3D community is yet unknown. In this paper, we focus on investigating these two problems and set two corresponding targets: Build the primary dataset, then explore its effect through experiments." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.615, + 0.254, + 0.628 + ], + "angle": 0, + "content": "Milestone 1 - Dataset:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.63, + 0.468, + 0.901 + ], + "angle": 0, + "content": "For a clearer picture of the first goal, we start by carefully revisiting existing 3D datasets as well as ImageNet [24]. i) 3D synthetic datasets [11, 96] provide rich 3D CAD models. However, they lack real-world cues (e.g., context, occlusions, noises), which are indispensable for model robustness in practical applications. ScanObjectNN [88] extracts real-world 3D objects from indoor scene data, but is limited in scale. For 3D scene-level dataset [5, 10, 23, 33, 37, 81], their scales are still constrained by the laborious scanning and labeling (e.g., millions of points per scene). Additionally, they contain specific inner-domain knowledge such as a particularly intricate indoor room or outdoor driving configurations, making it hard for general transfer learning. ii) Although ImageNet [24] contains the most comprehensive real-world objects, it only describes a 2D world that misses 3D-aware signals. Since humans live in a 3D world, 3D consciousness is vitally important for realizing humanlike intelligence and solving real-life visual problems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Based on the above review, our dataset is created from a new insight - multi-view images, as a soft bridge between 2D and 3D. It lies several benefits to remedying the aforementioned defects. Such data can be easily gained in considerable sizes via shooting an object around different views on common mobile devices with cameras (e.g., smartphones), which can be collected by crowd-sourcing in real world. Moreover, the multi-view constraint can bring natural 3D visual signals (later experiments show that this not only benefits 3D tasks but also 2D image understanding). To this end, we build MVImgNet, containing 6.5 million frames from 219,188 videos crossing real-life objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds. You may take a glance at our MVImgNet from Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.318, + 0.807, + 0.333 + ], + "angle": 0, + "content": "Milestone 2 - Experimental Exploration:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.574 + ], + "angle": 0, + "content": "Now facing the second goal of this paper, we attempt to probe the power of our dataset by conducting some pilot experiments. Leveraging the multi-view nature of the data, we start by focusing on the view-based 3D reconstruction task and demonstrate that pretraining on MVImgNet can not only benefit the generalization ability of NeRF (Sec. 4.1), but also data-efficient multi-view stereo (Sec. 4.2). Moreover, for image understanding, although humans can easily recognize one object from different viewpoints, deep learning models can hardly do that robustly [26,38]. Considering MVImgNet provides numerous images of a particular object from different viewpoints, we verify that MVImgNet-pretrained models are endowed with decent view consistency in general image classification (supervised learning in Sec. 5.1, self-supervised contrastive learning in Sec. 5.2) and salient object detection (Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.575, + 0.882, + 0.589 + ], + "angle": 0, + "content": "Bonus - A New 3D Point Cloud Dataset - MVPNet:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.891, + 0.695 + ], + "angle": 0, + "content": "Through dense reconstruction on MVImgNet, a new 3D object point cloud dataset is derived, called MVPNet, which contains 87,200 point clouds with 150 categories, with the class label on each point cloud (see Fig. 7). Experiments show that MVPNet not only benefits the real-world 3D object classification task but also poses new challenges and prospects to point cloud understanding (Sec. 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.891, + 0.726 + ], + "angle": 0, + "content": "MVImgNet and MVPNet will be public, hoping to inspire the broader vision community." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.737, + 0.641, + 0.753 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Single-view image datasets. The MNIST database [1] is one of the most pioneering datasets, composed of 70k monochrome images of handwritten digits. The CIFAR10 and CIFAR100 datasets proposed by [54], respectively collect 60k tiny color images \\((32\\times 32)\\) of various common objects or animals in 10 and 100 classes. ImageNet [24] is presented with a large scale, high accuracy, large diversity, and hierarchical structure, which provides opportunities for training deep neural networks. In detection" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "and segmentation tasks, MSCOCO [58] is one of the most popular datasets containing 328k images with rich annotations. Some other datasets include PASCAL VOC [28], Visual Genome [53], Cityscapes [22], MPII [4], etc. Although these datasets facilitate the development of deep visual learning, they mainly serve for 2D single-view image understanding, which limits their applications in 3D vision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.473, + 0.442 + ], + "angle": 0, + "content": "Video datasets. Another line is video datasets. Pioneering works construct the HMDB-51 [56] and UCF-101 [82] dataset. Afterward, the ActivityNet [8] and Kinetics [49] datasets are constructed of a larger scale and variation, of which the latter has collected 650k video clips that cover 700 classes. Besides, some datasets are built for human pose estimation [48, 90] and object detection/segmentation/tracking [69, 95, 112]. IEMOCAP [7] provides video data for the task of multimodal emotion recognition. MSVD [14] and MSR-VTT [103] annotate videos with extra captions. Further, HowTo100M [63] proposes a larger-scale one of 136 million samples from narrated instructional videos. ActivityNet Captions [52] introduces the task of dense-captioning events and constructs a dataset with 20k videos. These datasets are primarily for video understanding which is different from our objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.702 + ], + "angle": 0, + "content": "3D datasets. As the applications in 3D vision attract increasing attention, various 3D datasets are proposed. One line of works focuses on indoor scenes [5, 10, 23, 47, 78, 81, 100], where S3DIS [5] and ScanNet [23] are two of most popular datasets. In addition, some works provide 3D object point clouds for contextual object surface reconstruction [6, 29]. 3D outdoor scenes are scanned via LiDAR sensors [9, 33, 37, 83] for autonomous driving. ShapeNet [11] and ModelNet [96] are two object-centric datasets that provide rich 3D Computer-Aided Design (CAD) models for shape analysis, followed by similar datasets [51, 98], which are usually low-quality, untextured, and have a domain gap with real-world objects. Another line of works [20, 21, 68, 79, 89, 104] advocate real 3D objects but are still limited in scale. We close this gap by shooting multi-view images of real-world objects, which capture the 3D awareness while allowing a scalable collection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Multi-view image datasets. Multi-view image data is recently regarded as the source of 3D reconstruction or novel view synthesis. Early works collect multi-view images of real objects but only provide 3D models that are approximated [99] or for only a few instances [20]. Henzler et al. [45] contribute a larger video dataset to benchmark the task of 3D reconstruction. Another concurrent dataset, Objectron [3], annotates 3D bounding boxes and ground planes for all objects but lacks camera poses or point clouds for all 15k videos. GSO [27] gets clean 3D models with textures via scanning common household items but includes only a limited number of samples. Some works also construct synthetic multi-view datasets [11, 32, 86, 101, 108]. Neverthe" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.09, + 0.885, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.207, + 0.89, + 0.221 + ], + "angle": 0, + "content": "Figure 2. The efficient data acquisition pipeline of MVImgNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.31 + ], + "angle": 0, + "content": "less, these datasets have a small scale and category range for special tasks, which limits the robust and generic learning of 3D deep models. We construct a large-scale multi-view image dataset - MVImgNet, which contains 219k videos for real-world objects in a wide range of 238 categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Special discussions on CO3D [74]. Very recently, CO3D extends [45] to 19k videos covering 50 object categories. Although we share a similar idea in the light of object-centric multi-view data collections, our MVImgNet enjoys conspicuously larger scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.893, + 0.597 + ], + "angle": 0, + "content": "More importantly, we have fundamentally different motivations and insights. The primary motivation of CO3D is exactly pre-set, which is to transfer the 3D reconstruction training/evaluation from synthetic datasets into the real-world setup. In contrast, MVImgNet examines existing 3D datasets by re-walking the past development journey in the 2D domain. Concretely, through reflecting on how ImageNet [24] demonstrates its generic impact on data-hungry algorithms, we attempt to build a 3D counterpart of ImageNet and choose multi-view images as a soft bridge between 2D and 3D. Instead of nailing down a special target, we conduct a variety of pilot studies on how MVImgNet can benefit miscellaneous visual tasks, aiming to inspire and retain lots of possibilities for the broader vision community." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.629 + ], + "angle": 0, + "content": "In the later experiments, our datasets indeed show greater power than CO3D on different visual challenges." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.642, + 0.788, + 0.659 + ], + "angle": 0, + "content": "3. The Basis - MVImgNet Dataset" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.668, + 0.892, + 0.699 + ], + "angle": 0, + "content": "As shown in Fig. 2, the whole data acquisition pipeline of MVImgNet is highly efficient, which is illustrated below." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.71, + 0.725 + ], + "angle": 0, + "content": "3.1. Raw Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Building a large-scale dataset is always challenging, due to not only the laborious data collection but also the nontrivial annotation, which is especially critical for real-world 3D data [5, 10, 23]. Thanks to the rapid development of mobile devices, shooting a video around an object in the wild becomes highly convenient and accessible in our daily life, which makes multi-view images be easily gained by crowd-sourcing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.857, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Composition setup. We set up some constraints on the ratio. Depending on the category, each class is initially set with the different expected number of video cap" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9152" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.168, + 0.089, + 0.38, + 0.134 + ], + "angle": 0, + "content": "
AnnotationCollectedValidGPU hours
Sparse219,188215,7553,806.8
Segmentation104,261\\2,316.9
Dense98,89987,20025,122.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.137, + 0.47, + 0.166 + ], + "angle": 0, + "content": "Table 1. Data statistics, including collected amount, valid amount after cleaning, and the GPU hours for processing." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.177, + 0.468, + 0.311 + ], + "angle": 0, + "content": "
DatasetReal# of objects# of categoriesMulti-view3D-GT
ShapeNet [11]×51k55renderCAD model
ModelNet [96]×12k40renderCAD model
Choi et al. [20]✓2k9360° capturedRGB-D scan
Objectron [3]✓15k9limited3D bbox, pcl
GoogleScan [27]✓2kNA360° capturedRGB-D scan
Henzler et al. [45]✓2k7360° capturedpcl
ScanObjectNN [88]✓14k15limitedpcl
CO3D [74]✓19k50360° captured-
CO3D-pcl [74]✓5k50360° capturedpcl
MVIImgNet (ours)✓220k238180° captured-
MVPNet (ours)✓87k150180° capturedpcl
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.313, + 0.468, + 0.34 + ], + "angle": 0, + "content": "Table 2. Comparison between our datasets and related ones. \"pcl\" denotes point clouds, \"bbox\" means bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.469, + 0.415 + ], + "angle": 0, + "content": "tures regarding their generality, e.g., the number of captures for “bottles”, “bags” and “snacks” is planned to be around 2000, while the number of “chips”, “apples” and “guitars” is set to about 1000. This setting is closer to real life." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.469, + 0.555 + ], + "angle": 0, + "content": "Video capture. How to capture videos directly affect the quality of our data, so we draw up the following requirements as guidance for the captured videos: 1) The length of each video should be about 10 seconds. 2) The frames in the video should not be blurred. 3) Each video should capture \\(180^{\\circ}\\) or \\(360^{\\circ}\\) view of the object as much as possible. 4) The proportion of the object in the video should be above \\(15\\%\\). 5) Each video can only contain one class of principal object. 6) The captured object should be stereoscopic." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.724 + ], + "angle": 0, + "content": "Crowdsourcing. We employ around 1000 normal collectors from different professions (e.g., teacher, doctor, student, cook, babysitter) and ages \\((20\\sim 50)\\). Each of them is asked to take several videos in their daily life, (i.e., denoting diverse real-world environments) and upload them to the backend. Meanwhile, about 200 new well-trained expert data cleaners are responsible to review each submission and assure it fulfills the aforementioned capture requirements, when they may report some feedback or directly delete the unqualified submissions. The whole procedure ensures both the diversity and quality of the raw videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.737, + 0.238, + 0.754 + ], + "angle": 0, + "content": "3.2. Data Processing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.469, + 0.822 + ], + "angle": 0, + "content": "For each qualified video submission, we exploit an automatic process to obtain the common 2D and 3D annotations, including object masks, camera intrinsic and extrinsic, depth maps, and point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Sparse reconstruction. Following the procedure of [64, 65], the sparse reconstruction aims to reconstruct the camera intrinsic and extrinsic for each video, by applying the COLMAP Structure-from-Motion (SfM) algorithm [75] on a series of equal-time-interval chosen frames." + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.089, + 0.845, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.226, + 0.848, + 0.24 + ], + "angle": 0, + "content": "Figure 3. Some frames sampled from MVImgNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.325 + ], + "angle": 0, + "content": "Foreground object segmentation. Each frame extracted from the original video is fed to the CarveKit [70] for generating the binary foreground object mask, not only benefiting the dense reconstruction but also contributing to the further step of salient object detection (Sec. 5.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Dense reconstruction. With the sparse model output from COLMAP SfM, we employ multi-view stereo [76] of COLMAP to generate the dense depth and normal maps for each frame. We extract the depths of an object using the binary foreground masks, which are then back-projected and fused according to the normal information, yielding a densely reconstructed point cloud for each video. Finally, the point cloud is manually cleaned by: 1) Delete the object with obvious noisy or extremely sparse reconstruction. 2) Remove all backgrounds. The final derived 3D point cloud dataset - MVPNet, is illustrated in Sec. 6.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.673, + 0.52 + ], + "angle": 0, + "content": "3.3. Dataset Summary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Statistics. The statistics of MVImgNet are shown in Tab. 1, and Tab. 2 compares our datasets with other alternatives. MVImgNet includes 238 object classes, from 6.5 million frames of 219,188 videos. Fig. 3 shows some samples of MVImgNet. The annotation comprehensively covers object masks, camera parameters, and point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Category. We leverage WordNet [66] taxonomy that is used by ImageNet [24] to describe multi-hierarchy categories of objects and define 238 common classes. Unlike ImageNet which contains various plants and animals (nature-centric), the objects in our MVImgNet are found or used in human daily life (human-centric), where 65 classes overlap with ImageNet. The detailed category taxonomy, per-category data distributions, and more sample visualizations of MVImgNet are illustrated in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.79, + 0.681, + 0.805 + ], + "angle": 0, + "content": "4.3D Reconstruction" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.773, + 0.83 + ], + "angle": 0, + "content": "4.1. Radiance Field Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pre-review. Recently, a series of generalizable Neural Radiance Fields (NeRF) variants [13, 87, 93, 111] have been proposed to reconstruct radiance field on-the-fly from one/few-shot source views for novel view synthesis." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9153" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.093, + 0.581, + 0.142 + ], + "angle": 0, + "content": "
MethodDiffuse PSNR↑Synthetic SSIM↑360° LPIPS↓Realistic PSNR↑Synthetic SSIM↑360° LPIPS↓Real-world PSNR↑360° SSIM↑Objects LPIPS↓
Train from scratch37.170.9900.01725.490.9160.10022.180.7140.365
MVIImgNet-pretrained37.660.9900.01427.260.9300.07124.670.7400.310
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.144, + 0.541, + 0.157 + ], + "angle": 0, + "content": "Table 3. NeRF quantitative results on three different levels of test sets." + }, + { + "type": "table", + "bbox": [ + 0.642, + 0.09, + 0.861, + 0.146 + ], + "angle": 0, + "content": "
Where to pretrainReal-world 360°Objects [74]
PSNR↑SSIM↑LPIPS↓
CO3D [74]24.010.7320.339
MVIImgNet-small24.080.7360.316
MVIImgNet24.670.7400.310
" + }, + { + "type": "table_caption", + "bbox": [ + 0.582, + 0.149, + 0.919, + 0.162 + ], + "angle": 0, + "content": "Table 4. NeRF quantitative comparison with CO3D [74]." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.168, + 0.451, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.471, + 0.342 + ], + "angle": 0, + "content": "Figure 4. NeRF qualitative comparison of train-from-scratch IBRNet, CO3D [74]-pretrained IBRNet and MVImgNet-pretrained IBRNet [93]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.506 + ], + "angle": 0, + "content": "What can MVImgNet do? Training NeRFs that can generalize to unseen objects requires learning 3D priors from a huge amount of multi-view images. Existing state-of-the-arts either resort to learning on large-scale synthetic data [13, 87, 111], or adopt a mixed use of synthetic data and a small self-collected real dataset [93]. However, the synthetic data introduces a big domain gap with real-world objects, only a few real scenes cannot remedy this defect. We argue that: our MVImgNet perfectly fits the huge data demand of learning-based generalizable NeRF methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.628 + ], + "angle": 0, + "content": "To verify this, we choose IBRNet [93] as the baseline and conduct an empirical study. We pretrain IBRNet on the full MVImgNet dataset then finetune on the training datasets used in IBRNet [93] for a few iterations, and compare with the original train-from-scratch IBRNet model in terms of generalization capability (i.e., generalizing to unseen scenes with only few-shot inputs). For fairness, we evaluate all methods under the same protocol of IBRNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Here comes a challenge about how to evaluate the generalization ability of NeRFs, since there is no official benchmark for this. To this end, we employ three different third-party object-centric datasets to form the test set. 1) The diffuse synthetic \\(360^{\\circ}\\) object dataset [80] which contains 4 Lambertian objects. 2) The realistic synthetic \\(360^{\\circ}\\) object dataset [64] which consists of 8 realistically fabricated objects. 3) The real-world \\(360^{\\circ}\\) object dataset [74] which includes 88 real-world scenes from different lighting conditions. To summarize, the whole test set comprises 100 objects from 56 distinct categories, ranging from the synthetic domain to the real-world domain, which is considered to be an impartial evaluation set for generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The quantitative and qualitative results are respectively shown in Tab. 3 and Fig. 4. Evidently, pretraining on MVImgNet improves the generalization ability of the model by a large margin. Moreover, we perform the same pretraining on CO3D [74] and MVImgNet-small (a random subset" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.171, + 0.831, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.277, + 0.894, + 0.307 + ], + "angle": 0, + "content": "Figure 5. PSNR of IBRNet pretrained by a different number of MVImgNet data (higher is better)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.893, + 0.378 + ], + "angle": 0, + "content": "of MVImgNet, which owns the same scale as CO3D). As shown in Tab. 4 and Fig. 4, MVImgNet shows greater power than CO3D [74]. The implementation details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.894, + 0.426 + ], + "angle": 0, + "content": "More data, more power. Fig. 5 shows that an apparent rising trend of generalization metrics can be observed with the increase of training data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.673, + 0.45 + ], + "angle": 0, + "content": "4.2. Multi-view Stereo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.718 + ], + "angle": 0, + "content": "Pre-review. Multi-view Stereo (MVS) [30] is a classical task in 3D computer vision, with the goal of reconstructing 3D geometry from multi-view images. Conventional methods [31,46,77,114] reconstruct 3D geometry by finding the patches matched with different images and estimating the depth according to the extrinsic camera. In recent years, deep learning methods are introduced into MVS to solve the issues such as weak textures, and nonlaplacian spheres. Two popular end-to-end methods MVSNet [109] and R-MVSNet [110] propose to encode the multi-view images and build a cost volume for predicting depth maps. Under the supervision of large-scale RGB-D datasets [2], they outperform conventional patch-matched-based approaches. However, these methods require massive amounts of RGB-D data, which is always difficult to acquire. This drives the emergence of self-supervised MVS methods [12,25,102,107]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.795 + ], + "angle": 0, + "content": "What can MVImgNet do? We demonstrate that our MVImgNet is capable of benefiting the data-efficient MVS with limited training examples, which is practically meaningful considering MVS always requires the burdensome collection of RGB-D data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We pretrain a self-supervised MVS method, JDACS [102], on MVImgNet. Our implementation follows the original settings of JDACS. Then, we select DTU [2] dataset with 79 training samples and 22 test samples. We perform the data-efficient evaluation with limited training data, where the MVImgNet-pretrained MVsNet is finetuned on \\(5\\%\\), \\(15\\%\\), and \\(25\\%\\) of DTU training samples, and eval" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9154" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.089, + 0.401, + 0.141 + ], + "angle": 0, + "content": "
Ratio2mm↑4mm↑8mm↑
5%48.61 / 50.1065.18 / 67.9775.69 / 78.58
15%54.74 / 54.6870.96 / 72.0480.32 / 81.73
25%57.29 / 58.6374.41 / 75.2083.68 / 84.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.142, + 0.47, + 0.185 + ], + "angle": 0, + "content": "Table 5. MVS depth map accuracy on DTU evaluation set under different ratios of DTU training samples, in terms of training from scratch / MVImgNet-pretrained (higher is better)." + }, + { + "type": "table", + "bbox": [ + 0.155, + 0.194, + 0.395, + 0.247 + ], + "angle": 0, + "content": "
Where to pretrain2mm↑4mm↑8mm↑
CO3D46.1866.7278.39
MVIImgNet-small47.7366.5378.15
MVIImgNet50.1067.9778.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.292 + ], + "angle": 0, + "content": "Table 6. MVS depth map accuracy comparison with CO3D [74] on DTU evaluation set. Pretrain on MVImgNet, MVImgNet-small or CO3D, and finetune using \\(5\\%\\) of DTU training samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.471, + 0.458 + ], + "angle": 0, + "content": "uated on the DTU test set. Tab. 5 reports the accuracy given different thresholds of the error between the predicted and ground truth depth map, which indicates that the MVImgNet-pretrained model is capable of improving the model trained from scratch by a large margin under the data-efficient setup. Furthermore, we perform pretraining using CO3D [74] and MVImgNet-small. Tab. 6 indicates that our MVImgNet is stronger than CO3D, and MVImgNet-small shows comparable power as CO3D. For more implementation details, please see the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.469, + 0.489 + ], + "angle": 0, + "content": "We advocate benchmarking NeRF and MVS methods with the help of pretraining on MVImgNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.504, + 0.421, + 0.522 + ], + "angle": 0, + "content": "5. View-consistent Image Understanding" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.529, + 0.394, + 0.545 + ], + "angle": 0, + "content": "5.1. View-consistent Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Pre-review. As indicated by Dong et al. [26], although humans can easily recognize one object from different views, deep learning models can hardly do that robustly. MVImgNet provides numerous images from different viewpoints, so we hope to enhance the model's view consistency with the help of MVImgNet, which is significantly important for realizing human-like intelligence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.734 + ], + "angle": 0, + "content": "What can MVImgNet do? One naive approach is to finetune the ImageNet-pretrained ResNet-50 [44] on our MVImgNet. However, such an approach will be problematic due to the categories of two datasets are very distinct, which may cause catastrophic forgetting issues [50]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.47, + 0.902 + ], + "angle": 0, + "content": "For this reason, we create a new training set, namely MVI-Mix, by mixing the original ImageNet data with MVIImgNet. Specifically, we randomly sample 5 consecutive frames of each video in MVIImgNet, and mix the multi-view images with original ImageNet data. We also build another two artificial training sets, namely MVI-Gap and MVI-Aug, that only differ MVI-Mix with image views. MVI-Gap samples the 5 non-consecutive frames that have much larger view differences. MVI-Aug samples 1 frame from each video of MVI-Mix and applies 4 different data augmentations (i.e., random color jittering, grid mask, ro" + }, + { + "type": "table", + "bbox": [ + 0.589, + 0.089, + 0.805, + 0.15 + ], + "angle": 0, + "content": "
DatasetConfidence VarAccuracy
ImageNet-only0.20753.09%
MVI-Aug0.10571.48%
MVI-Gap0.10377.23%
MVI-Mix0.10277.31%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.153, + 0.892, + 0.208 + ], + "angle": 0, + "content": "Table 7. View-consistent image classification results on MVImgNet test set using fully-supervised ResNet-50 [44]. Adding MVImgNet for training improves the view consistency (smaller Var and higher Acc indicate better view consistency)." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.219, + 0.835, + 0.277 + ], + "angle": 0, + "content": "
Train: CO3DI-MixTrain: MVI-Mix
Test: CO3DVar: 0.104Var: 0.155 (+0.051)
Acc: 91.25%Acc: 84.83% (-6.42%)
Test:MVIImgNetVar: 0.188 (+0.086)Var: 0.102
Acc: 57.93% (-19.38%)Test: 77.31%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.361 + ], + "angle": 0, + "content": "Table 8. View-consistent image classification comparison with CO3D [74]. The performance drops when train on CO3DI-Mix and test on MVIImgNet (blue number) are much larger than the opposite (red number), which means that MVI-Mix pretrained model is more robust to multiview consistency than CO3DI-Mix pretrained model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.59 + ], + "angle": 0, + "content": "tation and erase, respectively) to it, which aims to differentiate the multi-view augmentation from the normal data augmentations. We compare the variance of the softmax confidence and the accuracy on MVImgNet test set for examining the view consistency. As illustrated in Tab. 7, adding our MVImgNet data for training can effectively improve the model's view consistency and accuracy. The test dataset used for all experiments is MVImgNet test dataset. In addition to using the convolution-based network (i.e., ResNet-50), we also apply the Transformer-based architecture, DeiT-Tiny [85]. The results are \\(48.76\\%\\) accuracy and 0.225 Var on ImageNet-only, VS 0.122 Var \\((\\mathbf{0.103} \\downarrow)\\) and \\(73.88\\%\\) \\((25.12 \\uparrow)\\) accuracy on MVI-Mix. This leads to a unanimous conclusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.668 + ], + "angle": 0, + "content": "We further construct CO3DI-Mix (corresponding to MVI-Mix) as the mixer of CO3D [74] & ImageNet under the same ratio of MVImgNet & ImageNet in MVI-Mix. As illustrated in Tab. 8, MVImgNet brings more benefits than CO3D for view-consistent image recognition." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.825, + 0.696 + ], + "angle": 0, + "content": "5.2. View-consistent Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Pre-review. Contrastive Learning (CL) is one mainstream of self-supervised training techniques [16-19, 35, 40, 84]. One key factor of CL is the construction of positive/negative pairs, e.g., MoCo-v2 [17] treats an image with different augmentations, a.k.a. views, as the positive pair." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "What can MVImgNet do? One natural question is: can the viewpoints of MVImgNet serve as the positive pairs for \\(CL\\)? To answer this question, we finetune the off-the-shelf ImageNet-pretrained MoCo-v2 on MVImgNet. For each iteration, we randomly sample two frames from the same video as the positive pair and apply the original data augmentations used in MoCo-v2 [17]. Meanwhile, the frames from other videos will be treated as their negatives." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9155" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.089, + 0.404, + 0.135 + ], + "angle": 0, + "content": "
Training schemeConfidence VarAccuracy
w/o finetune0.09870.26%
w/ finetune0.08671.22%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.138, + 0.47, + 0.193 + ], + "angle": 0, + "content": "Table 9. View-consistency classification results under the self-supervised contrastive learning regime on MVImgNet test set. Finetuning MoCo-v2 [17] on MVImgNet improves the view consistency." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.203, + 0.451, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.305, + 0.47, + 0.374 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison between MVImgNet-finetuned U2Net [73] and original U2Net for salient object detection (SOD). Left: finetuning on MVImgNet improves the performance on a hard view. Right: finetuning on MVImgNet improves the performance on two consecutive hard views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.469, + 0.506 + ], + "angle": 0, + "content": "The MVImgNet-finetuned model is finally evaluated on the MVImgNet test dataset for examining the view consistency, where we compare the variance and mean of the softmax confidence and the accuracy. As Tab. 9 shows, finetuning on MVImgNet with CL can also improve both the model's view consistency and accuracy. More implementation details of view-consistent image classification is described in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.507, + 0.469, + 0.552 + ], + "angle": 0, + "content": "In the future, it is highly recommended to regard view-consistency as a criterion for evaluating the image recognition task, and utilize MVImgNet to pretrain the models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.275, + 0.575 + ], + "angle": 0, + "content": "5.3. View-consistent SOD" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.584, + 0.469, + 0.644 + ], + "angle": 0, + "content": "Pre-review. Salient Object Detection (SOD) aims to segment the most visually prominent objects in an image. Although remarkable progress has been made recently, it remains lots of challenges." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.75 + ], + "angle": 0, + "content": "What can MVImgNet do? We test a state-of-the-art SOD model U2Net [73] on our MVImgNet. As Fig. 6 shows, U2Net failed to segment \"hard\" views, even though some views can be segmented with few flaws. Despite such a disappointment, the inconsistent predictions of different views caught our attention: can we improve the SOD models with multi-view consistency?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.751, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We propose to leverage the multi-view consistency to improve SOD with the help of optical flows. Specifically, given two consecutive frames, we first calculate their optical flow and warp the flow to the segmentation mask of one of the frames. Then, a consistency loss can be calculated between the warped mask and the mask of another frame. For ease of implementation, we directly finetune U2Net on our MVImgNet. To prevent the catastrophic forgetting issue [50], in addition to the MVImgNet, we also use the original training data DUTS-TR [92]. For the \"hard\" views" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.09, + 0.852, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.201, + 0.868, + 0.215 + ], + "angle": 0, + "content": "Figure 7. Some 3D point clouds sampled from MVPNet." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.227, + 0.892, + 0.304 + ], + "angle": 0, + "content": "\\((\\mathrm{IoU} \\leq 0.7)\\) on our MVImgNet test set, the model finetuned on our MVImgNet can bring a \\(4.1\\%\\) IoU improvement (see Fig. 6 for qualitative comparison). The implementation details of view-consistent SOD can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.787, + 0.334 + ], + "angle": 0, + "content": "6. MVPNet for 3D Understanding" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.342, + 0.667, + 0.356 + ], + "angle": 0, + "content": "6.1. MVPNet Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Derived from the dense reconstruction on MVImgNet (as mentioned in Sec. 3.2), a new large-scale real-world 3D object point cloud dataset – MVPNet, is born, which contains 87,200 point clouds with 150 categories. Fig. 7 shows some examples of MVPNet. As listed in Tab. 2, compared with existing 3D object datasets, our MVPNet contains a conspicuously richer amount of real-world object point clouds, with abundant categories covering many common objects in the real life. The detailed category taxonomy per-category data distributions of MVPNet are illustrated in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.762, + 0.555 + ], + "angle": 0, + "content": "6.2.3D Point Cloud Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.639 + ], + "angle": 0, + "content": "In this work, we focus on point cloud classification. We believe that future works may also utilize our dataset to help much more 3D understanding tasks such as indoor-scene parsing, outdoor-environment perception, pose estimation, and robotics manipulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Pre-review. We advocate paying more attention to real-world setup, which is more feasible for real applications. ScanObjectNN [89] has been manifested as the most challenging point cloud classification benchmark so far, so we choose it for the major comparison with our MVPNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.795 + ], + "angle": 0, + "content": "What can MVPNet do? We show that pretraining on MVPNet is able to aid the performance of real-world point cloud classification [36, 39, 62, 71, 72, 94, 97, 105, 106]. We pretrain several models on MVPNet, and finetune them on ScanObjectNN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Two settings are considered for evaluation. First is PB_T50_RS in ScanObjectNN with small perturbation, translation and rotation on point cloud. Another is adding heavy rotation on PB_T50_RS, to create a more challenging setting. The results are shown in Tab. 10, where pretraining on MVPNet is able to increase the classification accuracy under most circumstances." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9156" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.447, + 0.205 + ], + "angle": 0, + "content": "
MethodAdd Random RotationPB_T50_RS
from scratchpretrainedfrom scratchpretrained
PointNet [71]60.57 / 55.2064.25 / 59.2970.63 / 67.2867.73 / 64.12
PointNet++ [72]76.50 / 73.4278.76 / 76.5478.80 / 75.7080.22 / 76.91
DGCNN [94]80.50 / 78.4580.42 / 78.2079.44 / 76.2482.36 / 80.08
PointMLP [62]83.69 / 82.5484.87 / 83.7185.64 / 84.1485.98 / 84.38
CurveNet [97]73.96 / 69.9678.99 / 76.5974.27 / 69.4383.68 / 81.17
GDANet [106]80.33 / 79.1483.59 / 82.2979.01 / 75.9183.90 / 82.51
PAConv [105]70.91 / 65.7076.21 / 72.4772.88 / 68.6076.91 / 73.45
PCT [36]81.32 / 79.1782.08 / 80.4177.46 / 73.6484.20 / 81.94
PointMAE [67]83.17 / 80.7586.19 / 84.6077.34 / 73.5284.13 / 81.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.208, + 0.47, + 0.264 + ], + "angle": 0, + "content": "Table 10. ScanObjectNN real-world point cloud classification results. The comparison is between the train-from-scratch model and MVPNet pretrained model. The metric is overall / average accuracy." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.27, + 0.428, + 0.324 + ], + "angle": 0, + "content": "
MethodPB_T50_RS
PreTr. on MVPNetPreTr. on MVPNet-smallPreTr. on CO3D
PointNet++78.90 / 77.1179.00 / 76.7878.79 / 77.20
CurveNet83.68 / 81.1773.92 / 68.9273.78 / 69.62
PointMAE84.13 / 81.9281.85 / 79.2581.40 / 78.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.327, + 0.469, + 0.381 + ], + "angle": 0, + "content": "Table 11. ScanObjectNN real-world point cloud classification comparison with CO3D [74]. Pretrain on MVImgNet, MVImgNet-small or CO3D, and finetune on ScanObjectNN. The metric is overall / average accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.392, + 0.42, + 0.409 + ], + "angle": 0, + "content": "6.3. Self-supervised Point Cloud Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.418, + 0.468, + 0.447 + ], + "angle": 0, + "content": "Pre-review. Self-supervised learning has been exploited for 3D object point cloud understanding [59,67,91,113]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.469, + 0.555 + ], + "angle": 0, + "content": "What can MVPNet do? In Tab. 10, PointMAE [67] pretrained on our MVPNet outperforms the state-of-the-art methods when finetuned on ScanObjectNN, proving the benefit of MVPNet on the self-supervised learning regime for real-world point cloud classification. Tab. 11 also shows that pretraining on MVPNet is more powerful than CO3D [74] for real-world point cloud classification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.565, + 0.357, + 0.581 + ], + "angle": 0, + "content": "6.4. MVPNet Benchmark Challenge" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.588, + 0.469, + 0.648 + ], + "angle": 0, + "content": "We present the MVPNet benchmark challenge for real-world point cloud classification, which contains 69,760 training and 17,440 testing samples. The results of various methods are shown in Tab. 12." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.651, + 0.469, + 0.8 + ], + "angle": 0, + "content": "MVPNet is more challenging than ScanObjectNN. We firstly train models on ScanObjectNN, then conduct the test on MVPNet, which is concluded in Tab. 13. On the opposite, we also train on MVPNet and test on ScanObjectNN, which is listed in Tab. 14. Comparing Tab. 13 and Tab. 14, the accuracy drops are significantly larger when training on ScanObjectNN and testing on MVPNet, which verifies the greater challenge of our MVPNet. All the experiments in 3D understanding strictly follow the original settings of the selected backbone networks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.802, + 0.469, + 0.832 + ], + "angle": 0, + "content": "MVPNet is suggested for investigating 3D point cloud understanding in the future." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.195, + 0.86 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "We have introduced MVImgNet, a large-scale dataset of multi-view images, which is efficiently collected by shoot" + }, + { + "type": "table", + "bbox": [ + 0.591, + 0.089, + 0.803, + 0.185 + ], + "angle": 0, + "content": "
MethodOverall / Average Accuracy
PointNet [71]70.72 / 54.46
PointNet++ [72]79.15 / 58.24
DGCNN [94]86.49 / 63.98
PointMLP [62]88.89 / 73.64
CurveNet [97]88.88 / 75.37
GDANet [106]89.54 / 68.41
PACConv [105]83.35 / 59.13
PCT [36]91.49 / 75.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.188, + 0.892, + 0.215 + ], + "angle": 0, + "content": "Table 12. Quantitative results on our new MVPNet benchmark for real-world point cloud classification." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.225, + 0.844, + 0.273 + ], + "angle": 0, + "content": "
Methodtrain: MVPNet test: MVPNettrain: ScanObjectNN test: MVPNet
PointNet++ [72]79.15 / 58.2415.13 (-64.02) / 7.96 (-50.28)
CurveNet [97]88.88 / 75.3746.95 (-41.93) / 28.24 (-47.13)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.276, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Table 13. Quantitative comparison while training on ScanObjectNN, testing on MVPNet. The evaluation metric is overall / average accuracy." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.32, + 0.851, + 0.366 + ], + "angle": 0, + "content": "
Methodtrain: ScanObjectNN\ntest: ScanObjectNNtrain: MVPNet\ntest: ScanObjectNN
PointNet++ [72]76.50 / 73.4240.35 (-36.15) / 33.59 (-39.83)
CurveNet [97]73.96 / 69.9651.84 (-22.12) / 46.27 (-23.69)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.37, + 0.892, + 0.411 + ], + "angle": 0, + "content": "Table 14. Quantitative comparison while training on MVPNet, testing on ScanObjectNN. The evaluation metric is overall / average accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.618 + ], + "angle": 0, + "content": "ing videos of real-world objects. The multi-view nature endows our dataset with 3D-aware visual signals, making MVImgNet a soft bridge to link 2D and 3D vision. To probe the power of MVImgNet, we conduct a host of pilot experiments on various visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising effectiveness, expecting more future explorations. As a bonus of MVImgNet, a point cloud dataset - MVPNet is derived. Experiments show MVPNet can benefit real-world 3D object classification. As a broader impact on society, our datasets delineates a world - that is closer to a colorful and vivid real \\(3D\\) world - where we human lives." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.633, + 0.658, + 0.649 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.899 + ], + "angle": 0, + "content": "The work was supported in part by NSFC with Grant No. 62293482, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone. It was also partially supported by NSFC-62172348, Outstanding Yound Fund of Guangdong Province with No. 2023B1515020055 and Shenzhen General Project with No. JCYJ20220530143604010, the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). It was also sponsored by CCF-Tencent Open Research Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9157" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.468, + 0.141 + ], + "angle": 0, + "content": "[1] Gradient-based learning applied to document recognition, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.144, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[2] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.187, + 0.469, + 0.241 + ], + "angle": 0, + "content": "[3] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.243, + 0.469, + 0.284 + ], + "angle": 0, + "content": "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.285, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[5] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.341, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[6] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.398, + 0.469, + 0.466 + ], + "angle": 0, + "content": "[7] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.468, + 0.469, + 0.521 + ], + "angle": 0, + "content": "[8] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.524, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.594, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[10] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.664, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[11] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.734, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[12] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rcomvsnet: Unsupervised multi-view stereo with neural rendering. arXiv preprint arXiv:2203.03949, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[13] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In The 49th annual meeting of the association for computational linguistics: human language technologies, 2011. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[16] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.191, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[17] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.234, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[19] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[20] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.345, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[21] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.415, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[22] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[23] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.582, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[25] Yikang Ding, Qingtian Zhu, Xiangyue Liu, Wentao Yuan, Haotian Zhang, and Chi Zhang. Kd-mvs: Knowledge distillation based self-supervised learning for multi-view stereo. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[26] Yinpeng Dong, Shouwei Ruan, Hang Su, Caixin Kang, Xingxing Wei, and Jun Zhu. Viewfool: Evaluating the robustness of visual recognition to adversarial viewpoints. arXiv preprint arXiv:2210.03895, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.693, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[27] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv preprint arXiv:2204.11918, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.763, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[28] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.805, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[29] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Yasutaka Furukawa and Carlos Hernández. Multi-view stereo: A tutorial. Foundations and Trends in Computer Graphics and Vision, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9158" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[31] David Gallup, Jan-Michael Frahm, Philippos Mordohai, and Marc Pollefeys. Variable baseline/resolution stereo. In CVPR, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.472, + 0.192 + ], + "angle": 0, + "content": "[32] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.194, + 0.472, + 0.235 + ], + "angle": 0, + "content": "[33] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.237, + 0.385, + 0.251 + ], + "angle": 0, + "content": "[34] Ross Girshick. Fast r-cnn. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.253, + 0.472, + 0.349 + ], + "angle": 0, + "content": "[35] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In NeurIPS, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.352, + 0.472, + 0.394 + ], + "angle": 0, + "content": "[36] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.472, + 0.451 + ], + "angle": 0, + "content": "[37] Timo Hackel, Nikolay Savinov, Luber Ladicky, Jan Wegner, Konrad Schindler, and Marc Pollefeys. Semantic3d.net: A new large-scale point cloud classification benchmark. In ISPRS, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.472, + 0.495 + ], + "angle": 0, + "content": "[38] Abdullah Hamdi and Bernard Ghanem. Towards analyzing semantic robustness of deep neural networks. arXiv preprint arXiv:1904.04621, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.472, + 0.538 + ], + "angle": 0, + "content": "[39] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.54, + 0.472, + 0.582 + ], + "angle": 0, + "content": "[40] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, June 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.584, + 0.472, + 0.625 + ], + "angle": 0, + "content": "[41] Kaiming He, Ross Girshick, and Piotr Dólar. Rethinking imagenet pre-training. arXiv preprint arXiv:1811.08883, 2018.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.627, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[42] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.657, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.701, + 0.472, + 0.742 + ], + "angle": 0, + "content": "[44] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.744, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[45] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.802, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[46] Xiaoyan Hu and Philippos Mordohai. Least commitment, viewpoint-based, multi-view stereo. In International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[47] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 3DV, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[48] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[49] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.206, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[50] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[51] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[52] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.376, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[53] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.447, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[54] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.476, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[55] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.518, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[56] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.561, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[57] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. arXiv preprint arXiv:2108.10257, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[58] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.674, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[59] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.717, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[60] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[61] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.816, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[62] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[63] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[64] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.18, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[65] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Trans. Graph., 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.251, + 0.47, + 0.306 + ], + "angle": 0, + "content": "[66] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.309, + 0.471, + 0.351 + ], + "angle": 0, + "content": "[67] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.352, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[68] Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M Seitz. Photoshop: photorealistic materials for large-scale shape collections. ACM Trans. Graph., 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[69] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[70] Open Source Project. CarveKit - image-background-remove-tool. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.482, + 0.471, + 0.525 + ], + "angle": 0, + "content": "[71] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.526, + 0.471, + 0.569 + ], + "angle": 0, + "content": "[72] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.57, + 0.471, + 0.626 + ], + "angle": 0, + "content": "[73] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.628, + 0.471, + 0.696 + ], + "angle": 0, + "content": "[74] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.698, + 0.471, + 0.726 + ], + "angle": 0, + "content": "[75] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.728, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[76] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.772, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[77] Shuhan Shen. Accurate multiple view 3d reconstruction using patch-based stereo for large-scale scenes. TIP, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.815, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[78] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[79] Arjun Singh, James Sha, Karthik S Narayan, Tudor Achim, and Pieter Abbeel. Bigbird: A large-scale 3d database of object instances. In ICRA, 2014. 3" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[80] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[81] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[82] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.247, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[83] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.316, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[84] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[85] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[86] Jonathan Tremblay, Moustafa Meshry, Alex Evans, Jan Kautz, Alexander Keller, Sameh Khamis, Charles Loop, Nathan Morrical, Koki Nagano, Towaki Takikawa, et al. Rtmv: A ray-traced multi-view synthetic dataset for novel view synthesis. arXiv preprint arXiv:2205.07058, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.497, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[87] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[88] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[89] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.651, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[90] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[91] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.749, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[92] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[93] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[94] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9160" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "graph cnn for learning on point clouds. ACM Trans. Graph., 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.468, + 0.149 + ], + "angle": 0, + "content": "[95] Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Online object tracking: A benchmark. In CVPR, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.151, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[96] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.208, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[97] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In ICCV, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.25, + 0.468, + 0.305 + ], + "angle": 0, + "content": "[98] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.307, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[99] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond pascal: A benchmark for 3d object detection in the wild. In WACV, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.468, + 0.39 + ], + "angle": 0, + "content": "[100] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In ICCV, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[101] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. IJCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[102] Hongbin Xu, Zhipeng Zhou, Yu Qiao, Wenxiong Kang, and Qiuxia Wu. Self-supervised multi-view stereo via effective co-segmentation and data-augmentation. In AAAI, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.468, + 0.531 + ], + "angle": 0, + "content": "[103] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[104] Mutian Xu, Pei Chen, Haolin Liu, and Xiaoguang Han. Toscane: A large-scale dataset for understanding 3d tabletop scenes. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.468, + 0.617 + ], + "angle": 0, + "content": "[105] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.468, + 0.673 + ], + "angle": 0, + "content": "[106] Mutian Xu, Junhao Zhang, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi, and Yu Qiao. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In AAAI, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[107] Jiayu Yang, Jose M Alvarez, and Miaomiao Liu. Self-supervised learning of depth inference for multi-view stereo. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[108] Zhenpei Yang, Zaiwei Zhang, and Qixing Huang. Hm3d-abo: A photo-realistic dataset for object-centric multi-view 3d reconstruction. arXiv preprint arXiv:2206.12356, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[109] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[110] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[111] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[112] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[113] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[114] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In CVPR, 2014. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9161" + } + ] +] \ No newline at end of file diff --git a/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_origin.pdf b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..17fd64fe53f4524dcf0e84f588ff40b8aa626f0d --- /dev/null +++ b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/3e031e8b-eec6-45a0-a28e-e6489370a07d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ed6f54b5c40ccd9d4e11e4a43905c3d4603bffd7b51defbe5da09741ff2b437 +size 8249933 diff --git a/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/full.md b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3c3f5778730c3e90d5bb58e69b2eff0482ac39db --- /dev/null +++ b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/full.md @@ -0,0 +1,398 @@ +# MVImgNet: A Large-scale Dataset of Multi-view Images + +![](images/072d33ab3934ec8009ffb05b7f796dd1a196075de0e457c63ed880ec12548083.jpg) + +![](images/c0cfd9e332b747d0067e24e4e45f857297361add9a2b3aff92d4e8a3fa67e2b7.jpg) +Figure 1. It is time to embrace MVImgNet! We introduce MVImgnet, a large-scale dataset of multi-view images, which is efficiently collected by shooting videos of real-world objects. It enjoys 3D-aware signals from multi-view consistency, being a soft bridge between 2D and 3D vision. Through dense reconstruction on MVImgNet, we also present a large-scale real-world 3D object point cloud dataset – MVPNet. Exterior: Examples of various multi-view images in MVImgNet (see Fig. 3 more intuitively). Interior: Instances of colorful point clouds from MVPNet (see Fig. 7 more clearly) are assembled into a stereo sign of 'MVImgNet'. + +# Abstract + +Being data-driven is one of the most iconic properties of deep learning algorithms. The birth of ImageNet [24] drives a remarkable trend of 'learning from large-scale data' in computer vision. Pretraining on ImageNet to obtain rich universal representations has been manifested to benefit various 2D visual tasks, and becomes a standard in 2D vision. However, due to the laborious collection of real-world 3D data, there is yet no generic dataset serving as a counterpart of ImageNet in 3D vision, thus how such a dataset can impact the 3D community is unraveled. To remedy this defect, we introduce MVImgNet, a large-scale dataset of multi-view images, which is highly convenient to gain by shooting videos of real-world objects in human daily life. It contains 6.5 million frames from 219,188 videos crossing objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds. + +The multi-view attribute endows our dataset with 3D-aware signals, making it a soft bridge between 2D and 3D vision. + +We conduct pilot studies for probing the potential of MVImgNet on a variety of 3D and 2D visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising performance, remaining lots of possibilities for future explorations. + +Besides, via dense reconstruction on MVImgNet, a 3D object point cloud dataset is derived, called MVPNet, covering 87,200 samples from 150 categories, with the class label on each point cloud. Experiments show that MVPNet can benefit the real-world 3D object classification while posing new challenges to point cloud understanding. + +MVImgNet and MVPNet will be public, hoping to inspire the broader vision community. + +# 1. Introduction + +Being data-driven, also known as data-hungry, is one of the most important attributes of deep learning algorithms. By training on large-scale datasets, deep neural networks are able to extract rich representations. In the past few years, the computer vision community has witnessed the bloom of such 'learning from data' regime [43, 44, 55], after the birth of ImageNet [24] – the pioneer of large-scale real-world image datasets. Notably, pretraining on ImageNet is well-proven to boost the model performance when transferring the pretrained weights into not only high-level [34, 41, 42, 60, 61] but also low-level visual tasks [15, 57], and becomes a de-facto standard in 2D. Recently, various 3D datasets [5, 9, 11, 23, 33, 96, 104] are produced to facilitate 3D visual applications. + +However, due to the non-trivial scanning and laborious labeling of real-world 3D data (commonly organized in point clouds or meshes), existing 3D datasets are either synthetic or their scales are not comparable with ImageNet [24]. Consequently, unlike in 2D vision where models are usually pretrained on ImageNet to gain universal representation or commonsense knowledge, most of the current methods in 3D area are directly trained and evaluated on particular datasets for solving specific 3D visual tasks (e.g., NeRF dataset [64] and ShapeNet [11] for novel view synthesis, ModelNet [96] and ScanObjectNN [88] for object classification, KITTI [33] and ScanNet [23] for scene understanding). Here, two crucial and successive issues can be induced: (1) There is still no generic dataset in 3D vision, as a counterpart of ImageNet in 2D. (2) What benefit such a dataset can endow to 3D community is yet unknown. In this paper, we focus on investigating these two problems and set two corresponding targets: Build the primary dataset, then explore its effect through experiments. + +# Milestone 1 - Dataset: + +For a clearer picture of the first goal, we start by carefully revisiting existing 3D datasets as well as ImageNet [24]. i) 3D synthetic datasets [11, 96] provide rich 3D CAD models. However, they lack real-world cues (e.g., context, occlusions, noises), which are indispensable for model robustness in practical applications. ScanObjectNN [88] extracts real-world 3D objects from indoor scene data, but is limited in scale. For 3D scene-level dataset [5, 10, 23, 33, 37, 81], their scales are still constrained by the laborious scanning and labeling (e.g., millions of points per scene). Additionally, they contain specific inner-domain knowledge such as a particularly intricate indoor room or outdoor driving configurations, making it hard for general transfer learning. ii) Although ImageNet [24] contains the most comprehensive real-world objects, it only describes a 2D world that misses 3D-aware signals. Since humans live in a 3D world, 3D consciousness is vitally important for realizing humanlike intelligence and solving real-life visual problems. + +Based on the above review, our dataset is created from a new insight - multi-view images, as a soft bridge between 2D and 3D. It lies several benefits to remedying the aforementioned defects. Such data can be easily gained in considerable sizes via shooting an object around different views on common mobile devices with cameras (e.g., smartphones), which can be collected by crowd-sourcing in real world. Moreover, the multi-view constraint can bring natural 3D visual signals (later experiments show that this not only benefits 3D tasks but also 2D image understanding). To this end, we build MVImgNet, containing 6.5 million frames from 219,188 videos crossing real-life objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds. You may take a glance at our MVImgNet from Fig. 1. + +# Milestone 2 - Experimental Exploration: + +Now facing the second goal of this paper, we attempt to probe the power of our dataset by conducting some pilot experiments. Leveraging the multi-view nature of the data, we start by focusing on the view-based 3D reconstruction task and demonstrate that pretraining on MVImgNet can not only benefit the generalization ability of NeRF (Sec. 4.1), but also data-efficient multi-view stereo (Sec. 4.2). Moreover, for image understanding, although humans can easily recognize one object from different viewpoints, deep learning models can hardly do that robustly [26,38]. Considering MVImgNet provides numerous images of a particular object from different viewpoints, we verify that MVImgNet-pretrained models are endowed with decent view consistency in general image classification (supervised learning in Sec. 5.1, self-supervised contrastive learning in Sec. 5.2) and salient object detection (Sec. 5.3). + +# Bonus - A New 3D Point Cloud Dataset - MVPNet: + +Through dense reconstruction on MVImgNet, a new 3D object point cloud dataset is derived, called MVPNet, which contains 87,200 point clouds with 150 categories, with the class label on each point cloud (see Fig. 7). Experiments show that MVPNet not only benefits the real-world 3D object classification task but also poses new challenges and prospects to point cloud understanding (Sec. 6). + +MVImgNet and MVPNet will be public, hoping to inspire the broader vision community. + +# 2. Related Work + +Single-view image datasets. The MNIST database [1] is one of the most pioneering datasets, composed of 70k monochrome images of handwritten digits. The CIFAR10 and CIFAR100 datasets proposed by [54], respectively collect 60k tiny color images $(32\times 32)$ of various common objects or animals in 10 and 100 classes. ImageNet [24] is presented with a large scale, high accuracy, large diversity, and hierarchical structure, which provides opportunities for training deep neural networks. In detection + +and segmentation tasks, MSCOCO [58] is one of the most popular datasets containing 328k images with rich annotations. Some other datasets include PASCAL VOC [28], Visual Genome [53], Cityscapes [22], MPII [4], etc. Although these datasets facilitate the development of deep visual learning, they mainly serve for 2D single-view image understanding, which limits their applications in 3D vision. + +Video datasets. Another line is video datasets. Pioneering works construct the HMDB-51 [56] and UCF-101 [82] dataset. Afterward, the ActivityNet [8] and Kinetics [49] datasets are constructed of a larger scale and variation, of which the latter has collected 650k video clips that cover 700 classes. Besides, some datasets are built for human pose estimation [48, 90] and object detection/segmentation/tracking [69, 95, 112]. IEMOCAP [7] provides video data for the task of multimodal emotion recognition. MSVD [14] and MSR-VTT [103] annotate videos with extra captions. Further, HowTo100M [63] proposes a larger-scale one of 136 million samples from narrated instructional videos. ActivityNet Captions [52] introduces the task of dense-captioning events and constructs a dataset with 20k videos. These datasets are primarily for video understanding which is different from our objective. + +3D datasets. As the applications in 3D vision attract increasing attention, various 3D datasets are proposed. One line of works focuses on indoor scenes [5, 10, 23, 47, 78, 81, 100], where S3DIS [5] and ScanNet [23] are two of most popular datasets. In addition, some works provide 3D object point clouds for contextual object surface reconstruction [6, 29]. 3D outdoor scenes are scanned via LiDAR sensors [9, 33, 37, 83] for autonomous driving. ShapeNet [11] and ModelNet [96] are two object-centric datasets that provide rich 3D Computer-Aided Design (CAD) models for shape analysis, followed by similar datasets [51, 98], which are usually low-quality, untextured, and have a domain gap with real-world objects. Another line of works [20, 21, 68, 79, 89, 104] advocate real 3D objects but are still limited in scale. We close this gap by shooting multi-view images of real-world objects, which capture the 3D awareness while allowing a scalable collection. + +Multi-view image datasets. Multi-view image data is recently regarded as the source of 3D reconstruction or novel view synthesis. Early works collect multi-view images of real objects but only provide 3D models that are approximated [99] or for only a few instances [20]. Henzler et al. [45] contribute a larger video dataset to benchmark the task of 3D reconstruction. Another concurrent dataset, Objectron [3], annotates 3D bounding boxes and ground planes for all objects but lacks camera poses or point clouds for all 15k videos. GSO [27] gets clean 3D models with textures via scanning common household items but includes only a limited number of samples. Some works also construct synthetic multi-view datasets [11, 32, 86, 101, 108]. Neverthe + +![](images/0cc67aefea19b21664fb036c08620ef3f224fa3308c3818077fba07a467c05bd.jpg) +Figure 2. The efficient data acquisition pipeline of MVImgNet. + +less, these datasets have a small scale and category range for special tasks, which limits the robust and generic learning of 3D deep models. We construct a large-scale multi-view image dataset - MVImgNet, which contains 219k videos for real-world objects in a wide range of 238 categories. + +Special discussions on CO3D [74]. Very recently, CO3D extends [45] to 19k videos covering 50 object categories. Although we share a similar idea in the light of object-centric multi-view data collections, our MVImgNet enjoys conspicuously larger scales. + +More importantly, we have fundamentally different motivations and insights. The primary motivation of CO3D is exactly pre-set, which is to transfer the 3D reconstruction training/evaluation from synthetic datasets into the real-world setup. In contrast, MVImgNet examines existing 3D datasets by re-walking the past development journey in the 2D domain. Concretely, through reflecting on how ImageNet [24] demonstrates its generic impact on data-hungry algorithms, we attempt to build a 3D counterpart of ImageNet and choose multi-view images as a soft bridge between 2D and 3D. Instead of nailing down a special target, we conduct a variety of pilot studies on how MVImgNet can benefit miscellaneous visual tasks, aiming to inspire and retain lots of possibilities for the broader vision community. + +In the later experiments, our datasets indeed show greater power than CO3D on different visual challenges. + +# 3. The Basis - MVImgNet Dataset + +As shown in Fig. 2, the whole data acquisition pipeline of MVImgNet is highly efficient, which is illustrated below. + +# 3.1. Raw Data Preparation + +Building a large-scale dataset is always challenging, due to not only the laborious data collection but also the nontrivial annotation, which is especially critical for real-world 3D data [5, 10, 23]. Thanks to the rapid development of mobile devices, shooting a video around an object in the wild becomes highly convenient and accessible in our daily life, which makes multi-view images be easily gained by crowd-sourcing. + +Composition setup. We set up some constraints on the ratio. Depending on the category, each class is initially set with the different expected number of video cap + +
AnnotationCollectedValidGPU hours
Sparse219,188215,7553,806.8
Segmentation104,261\2,316.9
Dense98,89987,20025,122.4
+ +Table 1. Data statistics, including collected amount, valid amount after cleaning, and the GPU hours for processing. + +
DatasetReal# of objects# of categoriesMulti-view3D-GT
ShapeNet [11]×51k55renderCAD model
ModelNet [96]×12k40renderCAD model
Choi et al. [20]✓2k9360° capturedRGB-D scan
Objectron [3]✓15k9limited3D bbox, pcl
GoogleScan [27]✓2kNA360° capturedRGB-D scan
Henzler et al. [45]✓2k7360° capturedpcl
ScanObjectNN [88]✓14k15limitedpcl
CO3D [74]✓19k50360° captured-
CO3D-pcl [74]✓5k50360° capturedpcl
MVIImgNet (ours)✓220k238180° captured-
MVPNet (ours)✓87k150180° capturedpcl
+ +Table 2. Comparison between our datasets and related ones. "pcl" denotes point clouds, "bbox" means bounding boxes. + +tures regarding their generality, e.g., the number of captures for “bottles”, “bags” and “snacks” is planned to be around 2000, while the number of “chips”, “apples” and “guitars” is set to about 1000. This setting is closer to real life. + +Video capture. How to capture videos directly affect the quality of our data, so we draw up the following requirements as guidance for the captured videos: 1) The length of each video should be about 10 seconds. 2) The frames in the video should not be blurred. 3) Each video should capture $180^{\circ}$ or $360^{\circ}$ view of the object as much as possible. 4) The proportion of the object in the video should be above $15\%$ . 5) Each video can only contain one class of principal object. 6) The captured object should be stereoscopic. + +Crowdsourcing. We employ around 1000 normal collectors from different professions (e.g., teacher, doctor, student, cook, babysitter) and ages $(20\sim 50)$ . Each of them is asked to take several videos in their daily life, (i.e., denoting diverse real-world environments) and upload them to the backend. Meanwhile, about 200 new well-trained expert data cleaners are responsible to review each submission and assure it fulfills the aforementioned capture requirements, when they may report some feedback or directly delete the unqualified submissions. The whole procedure ensures both the diversity and quality of the raw videos. + +# 3.2. Data Processing + +For each qualified video submission, we exploit an automatic process to obtain the common 2D and 3D annotations, including object masks, camera intrinsic and extrinsic, depth maps, and point clouds. + +Sparse reconstruction. Following the procedure of [64, 65], the sparse reconstruction aims to reconstruct the camera intrinsic and extrinsic for each video, by applying the COLMAP Structure-from-Motion (SfM) algorithm [75] on a series of equal-time-interval chosen frames. + +![](images/b153f4934935b54f497c7f3de180a47c3741d5d3a51ea8e2c70c2ba963d5eb4f.jpg) +Figure 3. Some frames sampled from MVImgNet. + +Foreground object segmentation. Each frame extracted from the original video is fed to the CarveKit [70] for generating the binary foreground object mask, not only benefiting the dense reconstruction but also contributing to the further step of salient object detection (Sec. 5.3). + +Dense reconstruction. With the sparse model output from COLMAP SfM, we employ multi-view stereo [76] of COLMAP to generate the dense depth and normal maps for each frame. We extract the depths of an object using the binary foreground masks, which are then back-projected and fused according to the normal information, yielding a densely reconstructed point cloud for each video. Finally, the point cloud is manually cleaned by: 1) Delete the object with obvious noisy or extremely sparse reconstruction. 2) Remove all backgrounds. The final derived 3D point cloud dataset - MVPNet, is illustrated in Sec. 6.1. + +# 3.3. Dataset Summary + +Statistics. The statistics of MVImgNet are shown in Tab. 1, and Tab. 2 compares our datasets with other alternatives. MVImgNet includes 238 object classes, from 6.5 million frames of 219,188 videos. Fig. 3 shows some samples of MVImgNet. The annotation comprehensively covers object masks, camera parameters, and point clouds. + +Category. We leverage WordNet [66] taxonomy that is used by ImageNet [24] to describe multi-hierarchy categories of objects and define 238 common classes. Unlike ImageNet which contains various plants and animals (nature-centric), the objects in our MVImgNet are found or used in human daily life (human-centric), where 65 classes overlap with ImageNet. The detailed category taxonomy, per-category data distributions, and more sample visualizations of MVImgNet are illustrated in the supplementary material. + +# 4.3D Reconstruction + +# 4.1. Radiance Field Reconstruction + +Pre-review. Recently, a series of generalizable Neural Radiance Fields (NeRF) variants [13, 87, 93, 111] have been proposed to reconstruct radiance field on-the-fly from one/few-shot source views for novel view synthesis. + +
MethodDiffuse PSNR↑Synthetic SSIM↑360° LPIPS↓Realistic PSNR↑Synthetic SSIM↑360° LPIPS↓Real-world PSNR↑360° SSIM↑Objects LPIPS↓
Train from scratch37.170.9900.01725.490.9160.10022.180.7140.365
MVIImgNet-pretrained37.660.9900.01427.260.9300.07124.670.7400.310
+ +Table 3. NeRF quantitative results on three different levels of test sets. + +
Where to pretrainReal-world 360°Objects [74]
PSNR↑SSIM↑LPIPS↓
CO3D [74]24.010.7320.339
MVIImgNet-small24.080.7360.316
MVIImgNet24.670.7400.310
+ +Table 4. NeRF quantitative comparison with CO3D [74]. + +![](images/8cf5e9928a1762f9647ef48dedd1dd33aa4f416e2e8dc4f58b16937c3286e65e.jpg) +Figure 4. NeRF qualitative comparison of train-from-scratch IBRNet, CO3D [74]-pretrained IBRNet and MVImgNet-pretrained IBRNet [93]. + +What can MVImgNet do? Training NeRFs that can generalize to unseen objects requires learning 3D priors from a huge amount of multi-view images. Existing state-of-the-arts either resort to learning on large-scale synthetic data [13, 87, 111], or adopt a mixed use of synthetic data and a small self-collected real dataset [93]. However, the synthetic data introduces a big domain gap with real-world objects, only a few real scenes cannot remedy this defect. We argue that: our MVImgNet perfectly fits the huge data demand of learning-based generalizable NeRF methods. + +To verify this, we choose IBRNet [93] as the baseline and conduct an empirical study. We pretrain IBRNet on the full MVImgNet dataset then finetune on the training datasets used in IBRNet [93] for a few iterations, and compare with the original train-from-scratch IBRNet model in terms of generalization capability (i.e., generalizing to unseen scenes with only few-shot inputs). For fairness, we evaluate all methods under the same protocol of IBRNet. + +Here comes a challenge about how to evaluate the generalization ability of NeRFs, since there is no official benchmark for this. To this end, we employ three different third-party object-centric datasets to form the test set. 1) The diffuse synthetic $360^{\circ}$ object dataset [80] which contains 4 Lambertian objects. 2) The realistic synthetic $360^{\circ}$ object dataset [64] which consists of 8 realistically fabricated objects. 3) The real-world $360^{\circ}$ object dataset [74] which includes 88 real-world scenes from different lighting conditions. To summarize, the whole test set comprises 100 objects from 56 distinct categories, ranging from the synthetic domain to the real-world domain, which is considered to be an impartial evaluation set for generalization ability. + +The quantitative and qualitative results are respectively shown in Tab. 3 and Fig. 4. Evidently, pretraining on MVImgNet improves the generalization ability of the model by a large margin. Moreover, we perform the same pretraining on CO3D [74] and MVImgNet-small (a random subset + +![](images/1089e7076a63296b6f70a630ede941cf05604a67a5ae71d7596330049dce03e4.jpg) +Figure 5. PSNR of IBRNet pretrained by a different number of MVImgNet data (higher is better). + +of MVImgNet, which owns the same scale as CO3D). As shown in Tab. 4 and Fig. 4, MVImgNet shows greater power than CO3D [74]. The implementation details can be found in the supplementary material. + +More data, more power. Fig. 5 shows that an apparent rising trend of generalization metrics can be observed with the increase of training data. + +# 4.2. Multi-view Stereo + +Pre-review. Multi-view Stereo (MVS) [30] is a classical task in 3D computer vision, with the goal of reconstructing 3D geometry from multi-view images. Conventional methods [31,46,77,114] reconstruct 3D geometry by finding the patches matched with different images and estimating the depth according to the extrinsic camera. In recent years, deep learning methods are introduced into MVS to solve the issues such as weak textures, and nonlaplacian spheres. Two popular end-to-end methods MVSNet [109] and R-MVSNet [110] propose to encode the multi-view images and build a cost volume for predicting depth maps. Under the supervision of large-scale RGB-D datasets [2], they outperform conventional patch-matched-based approaches. However, these methods require massive amounts of RGB-D data, which is always difficult to acquire. This drives the emergence of self-supervised MVS methods [12,25,102,107]. + +What can MVImgNet do? We demonstrate that our MVImgNet is capable of benefiting the data-efficient MVS with limited training examples, which is practically meaningful considering MVS always requires the burdensome collection of RGB-D data. + +We pretrain a self-supervised MVS method, JDACS [102], on MVImgNet. Our implementation follows the original settings of JDACS. Then, we select DTU [2] dataset with 79 training samples and 22 test samples. We perform the data-efficient evaluation with limited training data, where the MVImgNet-pretrained MVsNet is finetuned on $5\%$ , $15\%$ , and $25\%$ of DTU training samples, and eval + +
Ratio2mm↑4mm↑8mm↑
5%48.61 / 50.1065.18 / 67.9775.69 / 78.58
15%54.74 / 54.6870.96 / 72.0480.32 / 81.73
25%57.29 / 58.6374.41 / 75.2083.68 / 84.28
+ +Table 5. MVS depth map accuracy on DTU evaluation set under different ratios of DTU training samples, in terms of training from scratch / MVImgNet-pretrained (higher is better). + +
Where to pretrain2mm↑4mm↑8mm↑
CO3D46.1866.7278.39
MVIImgNet-small47.7366.5378.15
MVIImgNet50.1067.9778.58
+ +uated on the DTU test set. Tab. 5 reports the accuracy given different thresholds of the error between the predicted and ground truth depth map, which indicates that the MVImgNet-pretrained model is capable of improving the model trained from scratch by a large margin under the data-efficient setup. Furthermore, we perform pretraining using CO3D [74] and MVImgNet-small. Tab. 6 indicates that our MVImgNet is stronger than CO3D, and MVImgNet-small shows comparable power as CO3D. For more implementation details, please see the supplementary material. + +We advocate benchmarking NeRF and MVS methods with the help of pretraining on MVImgNet. + +# 5. View-consistent Image Understanding + +# 5.1. View-consistent Image Classification + +Pre-review. As indicated by Dong et al. [26], although humans can easily recognize one object from different views, deep learning models can hardly do that robustly. MVImgNet provides numerous images from different viewpoints, so we hope to enhance the model's view consistency with the help of MVImgNet, which is significantly important for realizing human-like intelligence. + +What can MVImgNet do? One naive approach is to finetune the ImageNet-pretrained ResNet-50 [44] on our MVImgNet. However, such an approach will be problematic due to the categories of two datasets are very distinct, which may cause catastrophic forgetting issues [50]. + +For this reason, we create a new training set, namely MVI-Mix, by mixing the original ImageNet data with MVIImgNet. Specifically, we randomly sample 5 consecutive frames of each video in MVIImgNet, and mix the multi-view images with original ImageNet data. We also build another two artificial training sets, namely MVI-Gap and MVI-Aug, that only differ MVI-Mix with image views. MVI-Gap samples the 5 non-consecutive frames that have much larger view differences. MVI-Aug samples 1 frame from each video of MVI-Mix and applies 4 different data augmentations (i.e., random color jittering, grid mask, ro + +Table 6. MVS depth map accuracy comparison with CO3D [74] on DTU evaluation set. Pretrain on MVImgNet, MVImgNet-small or CO3D, and finetune using $5\%$ of DTU training samples. + +
DatasetConfidence VarAccuracy
ImageNet-only0.20753.09%
MVI-Aug0.10571.48%
MVI-Gap0.10377.23%
MVI-Mix0.10277.31%
+ +Table 7. View-consistent image classification results on MVImgNet test set using fully-supervised ResNet-50 [44]. Adding MVImgNet for training improves the view consistency (smaller Var and higher Acc indicate better view consistency). + +
Train: CO3DI-MixTrain: MVI-Mix
Test: CO3DVar: 0.104Var: 0.155 (+0.051)
Acc: 91.25%Acc: 84.83% (-6.42%)
Test:MVIImgNetVar: 0.188 (+0.086)Var: 0.102
Acc: 57.93% (-19.38%)Test: 77.31%
+ +Table 8. View-consistent image classification comparison with CO3D [74]. The performance drops when train on CO3DI-Mix and test on MVIImgNet (blue number) are much larger than the opposite (red number), which means that MVI-Mix pretrained model is more robust to multiview consistency than CO3DI-Mix pretrained model. + +tation and erase, respectively) to it, which aims to differentiate the multi-view augmentation from the normal data augmentations. We compare the variance of the softmax confidence and the accuracy on MVImgNet test set for examining the view consistency. As illustrated in Tab. 7, adding our MVImgNet data for training can effectively improve the model's view consistency and accuracy. The test dataset used for all experiments is MVImgNet test dataset. In addition to using the convolution-based network (i.e., ResNet-50), we also apply the Transformer-based architecture, DeiT-Tiny [85]. The results are $48.76\%$ accuracy and 0.225 Var on ImageNet-only, VS 0.122 Var $(\mathbf{0.103} \downarrow)$ and $73.88\%$ $(25.12 \uparrow)$ accuracy on MVI-Mix. This leads to a unanimous conclusion. + +We further construct CO3DI-Mix (corresponding to MVI-Mix) as the mixer of CO3D [74] & ImageNet under the same ratio of MVImgNet & ImageNet in MVI-Mix. As illustrated in Tab. 8, MVImgNet brings more benefits than CO3D for view-consistent image recognition. + +# 5.2. View-consistent Contrastive Learning + +Pre-review. Contrastive Learning (CL) is one mainstream of self-supervised training techniques [16-19, 35, 40, 84]. One key factor of CL is the construction of positive/negative pairs, e.g., MoCo-v2 [17] treats an image with different augmentations, a.k.a. views, as the positive pair. + +What can MVImgNet do? One natural question is: can the viewpoints of MVImgNet serve as the positive pairs for $CL$ ? To answer this question, we finetune the off-the-shelf ImageNet-pretrained MoCo-v2 on MVImgNet. For each iteration, we randomly sample two frames from the same video as the positive pair and apply the original data augmentations used in MoCo-v2 [17]. Meanwhile, the frames from other videos will be treated as their negatives. + +
Training schemeConfidence VarAccuracy
w/o finetune0.09870.26%
w/ finetune0.08671.22%
+ +Table 9. View-consistency classification results under the self-supervised contrastive learning regime on MVImgNet test set. Finetuning MoCo-v2 [17] on MVImgNet improves the view consistency. + +![](images/8fca2ba8a21e7fee99ee8035da0a5ba775212954ee03eea21d8d4e62f0d9c79b.jpg) +Figure 6. Qualitative comparison between MVImgNet-finetuned U2Net [73] and original U2Net for salient object detection (SOD). Left: finetuning on MVImgNet improves the performance on a hard view. Right: finetuning on MVImgNet improves the performance on two consecutive hard views. + +The MVImgNet-finetuned model is finally evaluated on the MVImgNet test dataset for examining the view consistency, where we compare the variance and mean of the softmax confidence and the accuracy. As Tab. 9 shows, finetuning on MVImgNet with CL can also improve both the model's view consistency and accuracy. More implementation details of view-consistent image classification is described in the supplementary material. + +In the future, it is highly recommended to regard view-consistency as a criterion for evaluating the image recognition task, and utilize MVImgNet to pretrain the models. + +# 5.3. View-consistent SOD + +Pre-review. Salient Object Detection (SOD) aims to segment the most visually prominent objects in an image. Although remarkable progress has been made recently, it remains lots of challenges. + +What can MVImgNet do? We test a state-of-the-art SOD model U2Net [73] on our MVImgNet. As Fig. 6 shows, U2Net failed to segment "hard" views, even though some views can be segmented with few flaws. Despite such a disappointment, the inconsistent predictions of different views caught our attention: can we improve the SOD models with multi-view consistency? + +We propose to leverage the multi-view consistency to improve SOD with the help of optical flows. Specifically, given two consecutive frames, we first calculate their optical flow and warp the flow to the segmentation mask of one of the frames. Then, a consistency loss can be calculated between the warped mask and the mask of another frame. For ease of implementation, we directly finetune U2Net on our MVImgNet. To prevent the catastrophic forgetting issue [50], in addition to the MVImgNet, we also use the original training data DUTS-TR [92]. For the "hard" views + +![](images/40b0bba5864a6bb332e007b8ce09103b4297013eedba4fcb071aabe23123c508.jpg) +Figure 7. Some 3D point clouds sampled from MVPNet. + +$(\mathrm{IoU} \leq 0.7)$ on our MVImgNet test set, the model finetuned on our MVImgNet can bring a $4.1\%$ IoU improvement (see Fig. 6 for qualitative comparison). The implementation details of view-consistent SOD can be found in the supplementary material. + +# 6. MVPNet for 3D Understanding + +# 6.1. MVPNet Dataset + +Derived from the dense reconstruction on MVImgNet (as mentioned in Sec. 3.2), a new large-scale real-world 3D object point cloud dataset – MVPNet, is born, which contains 87,200 point clouds with 150 categories. Fig. 7 shows some examples of MVPNet. As listed in Tab. 2, compared with existing 3D object datasets, our MVPNet contains a conspicuously richer amount of real-world object point clouds, with abundant categories covering many common objects in the real life. The detailed category taxonomy per-category data distributions of MVPNet are illustrated in the supplementary material. + +# 6.2.3D Point Cloud Classification + +In this work, we focus on point cloud classification. We believe that future works may also utilize our dataset to help much more 3D understanding tasks such as indoor-scene parsing, outdoor-environment perception, pose estimation, and robotics manipulation. + +Pre-review. We advocate paying more attention to real-world setup, which is more feasible for real applications. ScanObjectNN [89] has been manifested as the most challenging point cloud classification benchmark so far, so we choose it for the major comparison with our MVPNet. + +What can MVPNet do? We show that pretraining on MVPNet is able to aid the performance of real-world point cloud classification [36, 39, 62, 71, 72, 94, 97, 105, 106]. We pretrain several models on MVPNet, and finetune them on ScanObjectNN. + +Two settings are considered for evaluation. First is PB_T50_RS in ScanObjectNN with small perturbation, translation and rotation on point cloud. Another is adding heavy rotation on PB_T50_RS, to create a more challenging setting. The results are shown in Tab. 10, where pretraining on MVPNet is able to increase the classification accuracy under most circumstances. + +
MethodAdd Random RotationPB_T50_RS
from scratchpretrainedfrom scratchpretrained
PointNet [71]60.57 / 55.2064.25 / 59.2970.63 / 67.2867.73 / 64.12
PointNet++ [72]76.50 / 73.4278.76 / 76.5478.80 / 75.7080.22 / 76.91
DGCNN [94]80.50 / 78.4580.42 / 78.2079.44 / 76.2482.36 / 80.08
PointMLP [62]83.69 / 82.5484.87 / 83.7185.64 / 84.1485.98 / 84.38
CurveNet [97]73.96 / 69.9678.99 / 76.5974.27 / 69.4383.68 / 81.17
GDANet [106]80.33 / 79.1483.59 / 82.2979.01 / 75.9183.90 / 82.51
PAConv [105]70.91 / 65.7076.21 / 72.4772.88 / 68.6076.91 / 73.45
PCT [36]81.32 / 79.1782.08 / 80.4177.46 / 73.6484.20 / 81.94
PointMAE [67]83.17 / 80.7586.19 / 84.6077.34 / 73.5284.13 / 81.92
+ +Table 10. ScanObjectNN real-world point cloud classification results. The comparison is between the train-from-scratch model and MVPNet pretrained model. The metric is overall / average accuracy. + +
MethodPB_T50_RS
PreTr. on MVPNetPreTr. on MVPNet-smallPreTr. on CO3D
PointNet++78.90 / 77.1179.00 / 76.7878.79 / 77.20
CurveNet83.68 / 81.1773.92 / 68.9273.78 / 69.62
PointMAE84.13 / 81.9281.85 / 79.2581.40 / 78.87
+ +# 6.3. Self-supervised Point Cloud Pretraining + +Pre-review. Self-supervised learning has been exploited for 3D object point cloud understanding [59,67,91,113]. + +What can MVPNet do? In Tab. 10, PointMAE [67] pretrained on our MVPNet outperforms the state-of-the-art methods when finetuned on ScanObjectNN, proving the benefit of MVPNet on the self-supervised learning regime for real-world point cloud classification. Tab. 11 also shows that pretraining on MVPNet is more powerful than CO3D [74] for real-world point cloud classification. + +# 6.4. MVPNet Benchmark Challenge + +We present the MVPNet benchmark challenge for real-world point cloud classification, which contains 69,760 training and 17,440 testing samples. The results of various methods are shown in Tab. 12. + +MVPNet is more challenging than ScanObjectNN. We firstly train models on ScanObjectNN, then conduct the test on MVPNet, which is concluded in Tab. 13. On the opposite, we also train on MVPNet and test on ScanObjectNN, which is listed in Tab. 14. Comparing Tab. 13 and Tab. 14, the accuracy drops are significantly larger when training on ScanObjectNN and testing on MVPNet, which verifies the greater challenge of our MVPNet. All the experiments in 3D understanding strictly follow the original settings of the selected backbone networks. + +MVPNet is suggested for investigating 3D point cloud understanding in the future. + +# 7. Conclusion + +We have introduced MVImgNet, a large-scale dataset of multi-view images, which is efficiently collected by shoot + +Table 11. ScanObjectNN real-world point cloud classification comparison with CO3D [74]. Pretrain on MVImgNet, MVImgNet-small or CO3D, and finetune on ScanObjectNN. The metric is overall / average accuracy. + +
MethodOverall / Average Accuracy
PointNet [71]70.72 / 54.46
PointNet++ [72]79.15 / 58.24
DGCNN [94]86.49 / 63.98
PointMLP [62]88.89 / 73.64
CurveNet [97]88.88 / 75.37
GDANet [106]89.54 / 68.41
PACConv [105]83.35 / 59.13
PCT [36]91.49 / 75.41
+ +Table 12. Quantitative results on our new MVPNet benchmark for real-world point cloud classification. + +
Methodtrain: MVPNet test: MVPNettrain: ScanObjectNN test: MVPNet
PointNet++ [72]79.15 / 58.2415.13 (-64.02) / 7.96 (-50.28)
CurveNet [97]88.88 / 75.3746.95 (-41.93) / 28.24 (-47.13)
+ +Table 13. Quantitative comparison while training on ScanObjectNN, testing on MVPNet. The evaluation metric is overall / average accuracy. + +
Methodtrain: ScanObjectNN +test: ScanObjectNNtrain: MVPNet +test: ScanObjectNN
PointNet++ [72]76.50 / 73.4240.35 (-36.15) / 33.59 (-39.83)
CurveNet [97]73.96 / 69.9651.84 (-22.12) / 46.27 (-23.69)
+ +Table 14. Quantitative comparison while training on MVPNet, testing on ScanObjectNN. The evaluation metric is overall / average accuracy. + +ing videos of real-world objects. The multi-view nature endows our dataset with 3D-aware visual signals, making MVImgNet a soft bridge to link 2D and 3D vision. To probe the power of MVImgNet, we conduct a host of pilot experiments on various visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising effectiveness, expecting more future explorations. As a bonus of MVImgNet, a point cloud dataset - MVPNet is derived. Experiments show MVPNet can benefit real-world 3D object classification. As a broader impact on society, our datasets delineates a world - that is closer to a colorful and vivid real $3D$ world - where we human lives. + +# Acknowledgments + +The work was supported in part by NSFC with Grant No. 62293482, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone. It was also partially supported by NSFC-62172348, Outstanding Yound Fund of Guangdong Province with No. 2023B1515020055 and Shenzhen General Project with No. JCYJ20220530143604010, the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). It was also sponsored by CCF-Tencent Open Research Fund. + +# References + +[1] Gradient-based learning applied to document recognition, 1998. 2 +[2] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 2016. 5 +[3] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In CVPR, 2021. 3, 4 +[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, 2014. 3 +[5] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 3 +[6] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In CVPR, 2019. 3 +[7] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 2008. 3 +[8] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3 +[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3 +[10] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 2, 3 +[11] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3, 4 +[12] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rcomvsnet: Unsupervised multi-view stereo with neural rendering. arXiv preprint arXiv:2203.03949, 2022. 5 +[13] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 4, 5 +[14] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In The 49th annual meeting of the association for computational linguistics: human language technologies, 2011. 3 + +[15] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2 +[16] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 6 +[17] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 6, 7 +[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 6 +[19] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 6 +[20] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 3, 4 +[21] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 3 +[22] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3 +[23] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3 +[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 3, 4 +[25] Yikang Ding, Qingtian Zhu, Xiangyue Liu, Wentao Yuan, Haotian Zhang, and Chi Zhang. Kd-mvs: Knowledge distillation based self-supervised learning for multi-view stereo. In ECCV, 2022. 5 +[26] Yinpeng Dong, Shouwei Ruan, Hang Su, Caixin Kang, Xingxing Wei, and Jun Zhu. Viewfool: Evaluating the robustness of visual recognition to adversarial viewpoints. arXiv preprint arXiv:2210.03895, 2022. 2, 6 +[27] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv preprint arXiv:2204.11918, 2022. 3, 4 +[28] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2015. 3 +[29] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In ICCV, 2021. 3 +[30] Yasutaka Furukawa and Carlos Hernández. Multi-view stereo: A tutorial. Foundations and Trends in Computer Graphics and Vision, 2015. 5 + +[31] David Gallup, Jan-Michael Frahm, Philippos Mordohai, and Marc Pollefeys. Variable baseline/resolution stereo. In CVPR, 2008. 5 +[32] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In CVPR, 2022. 3 +[33] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 3 +[34] Ross Girshick. Fast r-cnn. In ICCV, 2015. 2 +[35] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In NeurIPS, 2020. 6 +[36] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 2021. 7, 8 +[37] Timo Hackel, Nikolay Savinov, Luber Ladicky, Jan Wegner, Konrad Schindler, and Marc Pollefeys. Semantic3d.net: A new large-scale point cloud classification benchmark. In ISPRS, 2017. 2, 3 +[38] Abdullah Hamdi and Bernard Ghanem. Towards analyzing semantic robustness of deep neural networks. arXiv preprint arXiv:1904.04621, 2019. 2 +[39] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In ICCV, 2021. 7 +[40] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, June 2020. 6 +[41] Kaiming He, Ross Girshick, and Piotr Dólar. Rethinking imagenet pre-training. arXiv preprint arXiv:1811.08883, 2018.2 +[42] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2 +[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification. In ICCV, 2015. 2 +[44] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 6 +[45] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In CVPR, 2021. 3, 4 +[46] Xiaoyan Hu and Philippos Mordohai. Least commitment, viewpoint-based, multi-view stereo. In International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission, 2012. 5 +[47] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 3DV, 2016. 3 + +[48] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, 2013. 3 +[49] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 3 +[50] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 6, 7 +[51] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In CVPR, 2019. 3 +[52] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 3 +[53] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 3 +[54] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2 +[55] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 2 +[56] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 3 +[57] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. arXiv preprint arXiv:2108.10257, 2021. 2 +[58] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3 +[59] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 8 +[60] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2 +[61] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 2 +[62] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 7, 8 +[63] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. + +Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3 +[64] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 5 +[65] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Trans. Graph., 2019. 4 +[66] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 4 +[67] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 8 +[68] Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M Seitz. Photoshop: photorealistic materials for large-scale shape collections. ACM Trans. Graph., 2018. 3 +[69] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 3 +[70] Open Source Project. CarveKit - image-background-remove-tool. 4 +[71] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7, 8 +[72] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 7, 8 +[73] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 2020. 7 +[74] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 3, 4, 5, 6, 8 +[75] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 4 +[76] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 4 +[77] Shuhan Shen. Accurate multiple view 3d reconstruction using patch-based stereo for large-scale scenes. TIP, 2013. 5 +[78] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3 +[79] Arjun Singh, James Sha, Karthik S Narayan, Tudor Achim, and Pieter Abbeel. Bigbird: A large-scale 3d database of object instances. In ICRA, 2014. 3 + +[80] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, 2019. 5 +[81] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 2, 3 +[82] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3 +[83] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 3 +[84] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. 6 +[85] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 6 +[86] Jonathan Tremblay, Moustafa Meshry, Alex Evans, Jan Kautz, Alexander Keller, Sameh Khamis, Charles Loop, Nathan Morrical, Koki Nagano, Towaki Takikawa, et al. Rtmv: A ray-traced multi-view synthetic dataset for novel view synthesis. arXiv preprint arXiv:2205.07058, 2022. 3 +[87] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, 2021. 4, 5 +[88] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 4 +[89] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 3, 7 +[90] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3 +[91] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 8 +[92] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 7 +[93] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 4, 5 +[94] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic + +graph cnn for learning on point clouds. ACM Trans. Graph., 2019. 7, 8 +[95] Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Online object tracking: A benchmark. In CVPR, 2013. 3 +[96] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 3, 4 +[97] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In ICCV, 2021. 7, 8 +[98] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In ECCV, 2016. 3 +[99] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond pascal: A benchmark for 3d object detection in the wild. In WACV, 2014. 3 +[100] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In ICCV, 2013. 3 +[101] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. IJCV, 2020. 3 +[102] Hongbin Xu, Zhipeng Zhou, Yu Qiao, Wenxiong Kang, and Qiuxia Wu. Self-supervised multi-view stereo via effective co-segmentation and data-augmentation. In AAAI, 2021. 5 +[103] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 3 +[104] Mutian Xu, Pei Chen, Haolin Liu, and Xiaoguang Han. Toscane: A large-scale dataset for understanding 3d tabletop scenes. In ECCV, 2022. 2, 3 +[105] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 7, 8 +[106] Mutian Xu, Junhao Zhang, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi, and Yu Qiao. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In AAAI, 2021. 7, 8 +[107] Jiayu Yang, Jose M Alvarez, and Miaomiao Liu. Self-supervised learning of depth inference for multi-view stereo. In CVPR, 2021. 5 +[108] Zhenpei Yang, Zaiwei Zhang, and Qixing Huang. Hm3d-abo: A photo-realistic dataset for object-centric multi-view 3d reconstruction. arXiv preprint arXiv:2206.12356, 2022. 3 +[109] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 5 +[110] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In CVPR, 2019. 5 +[111] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 4, 5 + +[112] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 3 +[113] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 8 +[114] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In CVPR, 2014. 5 \ No newline at end of file diff --git a/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/images.zip b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..be54a0b83e32a5994021fc987376505c75e95832 --- /dev/null +++ b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fdd38b7c51c25518ea2ef86435f0e6e7084a303d9cb5b39b2caf449884d432c +size 633115 diff --git a/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/layout.json b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5befb4011e9914b660aa2c638892c7a05dc58020 --- /dev/null +++ b/2023/MVImgNet_ A Large-Scale Dataset of Multi-View Images/layout.json @@ -0,0 +1,9818 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "type": "text", + "content": "MVImgNet: A Large-scale Dataset of Multi-view Images" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 141, + 532, + 244 + ], + "blocks": [ + { + "bbox": [ + 61, + 141, + 532, + 244 + ], + "lines": [ + { + "bbox": [ + 61, + 141, + 532, + 244 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 532, + 244 + ], + "type": "image", + "image_path": "072d33ab3934ec8009ffb05b7f796dd1a196075de0e457c63ed880ec12548083.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 270, + 545, + 419 + ], + "blocks": [ + { + "bbox": [ + 48, + 270, + 545, + 419 + ], + "lines": [ + { + "bbox": [ + 48, + 270, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 545, + 419 + ], + "type": "image", + "image_path": "c0cfd9e332b747d0067e24e4e45f857297361add9a2b3aff92d4e8a3fa67e2b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 422, + 546, + 479 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 546, + 479 + ], + "type": "text", + "content": "Figure 1. It is time to embrace MVImgNet! We introduce MVImgnet, a large-scale dataset of multi-view images, which is efficiently collected by shooting videos of real-world objects. It enjoys 3D-aware signals from multi-view consistency, being a soft bridge between 2D and 3D vision. Through dense reconstruction on MVImgNet, we also present a large-scale real-world 3D object point cloud dataset – MVPNet. Exterior: Examples of various multi-view images in MVImgNet (see Fig. 3 more intuitively). Interior: Instances of colorful point clouds from MVPNet (see Fig. 7 more clearly) are assembled into a stereo sign of 'MVImgNet'." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 488, + 192, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 488, + 192, + 502 + ], + "spans": [ + { + "bbox": [ + 143, + 488, + 192, + 502 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "Being data-driven is one of the most iconic properties of deep learning algorithms. The birth of ImageNet [24] drives a remarkable trend of 'learning from large-scale data' in computer vision. Pretraining on ImageNet to obtain rich universal representations has been manifested to benefit various 2D visual tasks, and becomes a standard in 2D vision. However, due to the laborious collection of real-world 3D data, there is yet no generic dataset serving as a counterpart of ImageNet in 3D vision, thus how such a dataset can impact the 3D community is unraveled. To remedy this defect, we introduce MVImgNet, a large-scale dataset of multi-view images, which is highly convenient to gain by shooting videos of real-world objects in human daily life. It contains 6.5 million frames from 219,188 videos crossing objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 489, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 545, + 513 + ], + "type": "text", + "content": "The multi-view attribute endows our dataset with 3D-aware signals, making it a soft bridge between 2D and 3D vision." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 547, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 595 + ], + "type": "text", + "content": "We conduct pilot studies for probing the potential of MVImgNet on a variety of 3D and 2D visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising performance, remaining lots of possibilities for future explorations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 603, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 547, + 676 + ], + "type": "text", + "content": "Besides, via dense reconstruction on MVImgNet, a 3D object point cloud dataset is derived, called MVPNet, covering 87,200 samples from 150 categories, with the class label on each point cloud. Experiments show that MVPNet can benefit the real-world 3D object classification while posing new challenges to point cloud understanding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 684, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 684, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 305, + 684, + 545, + 708 + ], + "type": "text", + "content": "MVImgNet and MVPNet will be public, hoping to inspire the broader vision community." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9150" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 286, + 258 + ], + "type": "text", + "content": "Being data-driven, also known as data-hungry, is one of the most important attributes of deep learning algorithms. By training on large-scale datasets, deep neural networks are able to extract rich representations. In the past few years, the computer vision community has witnessed the bloom of such 'learning from data' regime [43, 44, 55], after the birth of ImageNet [24] – the pioneer of large-scale real-world image datasets. Notably, pretraining on ImageNet is well-proven to boost the model performance when transferring the pretrained weights into not only high-level [34, 41, 42, 60, 61] but also low-level visual tasks [15, 57], and becomes a de-facto standard in 2D. Recently, various 3D datasets [5, 9, 11, 23, 33, 96, 104] are produced to facilitate 3D visual applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 259, + 286, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 286, + 486 + ], + "type": "text", + "content": "However, due to the non-trivial scanning and laborious labeling of real-world 3D data (commonly organized in point clouds or meshes), existing 3D datasets are either synthetic or their scales are not comparable with ImageNet [24]. Consequently, unlike in 2D vision where models are usually pretrained on ImageNet to gain universal representation or commonsense knowledge, most of the current methods in 3D area are directly trained and evaluated on particular datasets for solving specific 3D visual tasks (e.g., NeRF dataset [64] and ShapeNet [11] for novel view synthesis, ModelNet [96] and ScanObjectNN [88] for object classification, KITTI [33] and ScanNet [23] for scene understanding). Here, two crucial and successive issues can be induced: (1) There is still no generic dataset in 3D vision, as a counterpart of ImageNet in 2D. (2) What benefit such a dataset can endow to 3D community is yet unknown. In this paper, we focus on investigating these two problems and set two corresponding targets: Build the primary dataset, then explore its effect through experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 487, + 155, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 487, + 155, + 497 + ], + "spans": [ + { + "bbox": [ + 59, + 487, + 155, + 497 + ], + "type": "text", + "content": "Milestone 1 - Dataset:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 286, + 713 + ], + "type": "text", + "content": "For a clearer picture of the first goal, we start by carefully revisiting existing 3D datasets as well as ImageNet [24]. i) 3D synthetic datasets [11, 96] provide rich 3D CAD models. However, they lack real-world cues (e.g., context, occlusions, noises), which are indispensable for model robustness in practical applications. ScanObjectNN [88] extracts real-world 3D objects from indoor scene data, but is limited in scale. For 3D scene-level dataset [5, 10, 23, 33, 37, 81], their scales are still constrained by the laborious scanning and labeling (e.g., millions of points per scene). Additionally, they contain specific inner-domain knowledge such as a particularly intricate indoor room or outdoor driving configurations, making it hard for general transfer learning. ii) Although ImageNet [24] contains the most comprehensive real-world objects, it only describes a 2D world that misses 3D-aware signals. Since humans live in a 3D world, 3D consciousness is vitally important for realizing humanlike intelligence and solving real-life visual problems." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "Based on the above review, our dataset is created from a new insight - multi-view images, as a soft bridge between 2D and 3D. It lies several benefits to remedying the aforementioned defects. Such data can be easily gained in considerable sizes via shooting an object around different views on common mobile devices with cameras (e.g., smartphones), which can be collected by crowd-sourcing in real world. Moreover, the multi-view constraint can bring natural 3D visual signals (later experiments show that this not only benefits 3D tasks but also 2D image understanding). To this end, we build MVImgNet, containing 6.5 million frames from 219,188 videos crossing real-life objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds. You may take a glance at our MVImgNet from Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 251, + 493, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 493, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 493, + 263 + ], + "type": "text", + "content": "Milestone 2 - Experimental Exploration:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 264, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 454 + ], + "type": "text", + "content": "Now facing the second goal of this paper, we attempt to probe the power of our dataset by conducting some pilot experiments. Leveraging the multi-view nature of the data, we start by focusing on the view-based 3D reconstruction task and demonstrate that pretraining on MVImgNet can not only benefit the generalization ability of NeRF (Sec. 4.1), but also data-efficient multi-view stereo (Sec. 4.2). Moreover, for image understanding, although humans can easily recognize one object from different viewpoints, deep learning models can hardly do that robustly [26,38]. Considering MVImgNet provides numerous images of a particular object from different viewpoints, we verify that MVImgNet-pretrained models are endowed with decent view consistency in general image classification (supervised learning in Sec. 5.1, self-supervised contrastive learning in Sec. 5.2) and salient object detection (Sec. 5.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 455, + 539, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 539, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 539, + 466 + ], + "type": "text", + "content": "Bonus - A New 3D Point Cloud Dataset - MVPNet:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 467, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 550 + ], + "type": "text", + "content": "Through dense reconstruction on MVImgNet, a new 3D object point cloud dataset is derived, called MVPNet, which contains 87,200 point clouds with 150 categories, with the class label on each point cloud (see Fig. 7). Experiments show that MVPNet not only benefits the real-world 3D object classification task but also poses new challenges and prospects to point cloud understanding (Sec. 6)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 551, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 545, + 574 + ], + "type": "text", + "content": "MVImgNet and MVPNet will be public, hoping to inspire the broader vision community." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 583, + 392, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 583, + 392, + 596 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 392, + 596 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Single-view image datasets. The MNIST database [1] is one of the most pioneering datasets, composed of 70k monochrome images of handwritten digits. The CIFAR10 and CIFAR100 datasets proposed by [54], respectively collect 60k tiny color images " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "(32\\times 32)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " of various common objects or animals in 10 and 100 classes. ImageNet [24] is presented with a large scale, high accuracy, large diversity, and hierarchical structure, which provides opportunities for training deep neural networks. In detection" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9151" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "and segmentation tasks, MSCOCO [58] is one of the most popular datasets containing 328k images with rich annotations. Some other datasets include PASCAL VOC [28], Visual Genome [53], Cityscapes [22], MPII [4], etc. Although these datasets facilitate the development of deep visual learning, they mainly serve for 2D single-view image understanding, which limits their applications in 3D vision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 289, + 350 + ], + "type": "text", + "content": "Video datasets. Another line is video datasets. Pioneering works construct the HMDB-51 [56] and UCF-101 [82] dataset. Afterward, the ActivityNet [8] and Kinetics [49] datasets are constructed of a larger scale and variation, of which the latter has collected 650k video clips that cover 700 classes. Besides, some datasets are built for human pose estimation [48, 90] and object detection/segmentation/tracking [69, 95, 112]. IEMOCAP [7] provides video data for the task of multimodal emotion recognition. MSVD [14] and MSR-VTT [103] annotate videos with extra captions. Further, HowTo100M [63] proposes a larger-scale one of 136 million samples from narrated instructional videos. ActivityNet Captions [52] introduces the task of dense-captioning events and constructs a dataset with 20k videos. These datasets are primarily for video understanding which is different from our objective." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 555 + ], + "type": "text", + "content": "3D datasets. As the applications in 3D vision attract increasing attention, various 3D datasets are proposed. One line of works focuses on indoor scenes [5, 10, 23, 47, 78, 81, 100], where S3DIS [5] and ScanNet [23] are two of most popular datasets. In addition, some works provide 3D object point clouds for contextual object surface reconstruction [6, 29]. 3D outdoor scenes are scanned via LiDAR sensors [9, 33, 37, 83] for autonomous driving. ShapeNet [11] and ModelNet [96] are two object-centric datasets that provide rich 3D Computer-Aided Design (CAD) models for shape analysis, followed by similar datasets [51, 98], which are usually low-quality, untextured, and have a domain gap with real-world objects. Another line of works [20, 21, 68, 79, 89, 104] advocate real 3D objects but are still limited in scale. We close this gap by shooting multi-view images of real-world objects, which capture the 3D awareness while allowing a scalable collection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": "Multi-view image datasets. Multi-view image data is recently regarded as the source of 3D reconstruction or novel view synthesis. Early works collect multi-view images of real objects but only provide 3D models that are approximated [99] or for only a few instances [20]. Henzler et al. [45] contribute a larger video dataset to benchmark the task of 3D reconstruction. Another concurrent dataset, Objectron [3], annotates 3D bounding boxes and ground planes for all objects but lacks camera poses or point clouds for all 15k videos. GSO [27] gets clean 3D models with textures via scanning common household items but includes only a limited number of samples. Some works also construct synthetic multi-view datasets [11, 32, 86, 101, 108]. Neverthe" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 314, + 71, + 541, + 160 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 541, + 160 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 541, + 160 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 541, + 160 + ], + "type": "image", + "image_path": "0cc67aefea19b21664fb036c08620ef3f224fa3308c3818077fba07a467c05bd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 163, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 306, + 163, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 544, + 175 + ], + "type": "text", + "content": "Figure 2. The efficient data acquisition pipeline of MVImgNet." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 185, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 245 + ], + "type": "text", + "content": "less, these datasets have a small scale and category range for special tasks, which limits the robust and generic learning of 3D deep models. We construct a large-scale multi-view image dataset - MVImgNet, which contains 219k videos for real-world objects in a wide range of 238 categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 245, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 305 + ], + "type": "text", + "content": "Special discussions on CO3D [74]. Very recently, CO3D extends [45] to 19k videos covering 50 object categories. Although we share a similar idea in the light of object-centric multi-view data collections, our MVImgNet enjoys conspicuously larger scales." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 305, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 546, + 472 + ], + "type": "text", + "content": "More importantly, we have fundamentally different motivations and insights. The primary motivation of CO3D is exactly pre-set, which is to transfer the 3D reconstruction training/evaluation from synthetic datasets into the real-world setup. In contrast, MVImgNet examines existing 3D datasets by re-walking the past development journey in the 2D domain. Concretely, through reflecting on how ImageNet [24] demonstrates its generic impact on data-hungry algorithms, we attempt to build a 3D counterpart of ImageNet and choose multi-view images as a soft bridge between 2D and 3D. Instead of nailing down a special target, we conduct a variety of pilot studies on how MVImgNet can benefit miscellaneous visual tasks, aiming to inspire and retain lots of possibilities for the broader vision community." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 473, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 498 + ], + "type": "text", + "content": "In the later experiments, our datasets indeed show greater power than CO3D on different visual challenges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 508, + 482, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 482, + 521 + ], + "type": "text", + "content": "3. The Basis - MVImgNet Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 529, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 529, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 545, + 553 + ], + "type": "text", + "content": "As shown in Fig. 2, the whole data acquisition pipeline of MVImgNet is highly efficient, which is illustrated below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 561, + 434, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 434, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 434, + 574 + ], + "type": "text", + "content": "3.1. Raw Data Preparation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "type": "text", + "content": "Building a large-scale dataset is always challenging, due to not only the laborious data collection but also the nontrivial annotation, which is especially critical for real-world 3D data [5, 10, 23]. Thanks to the rapid development of mobile devices, shooting a video around an object in the wild becomes highly convenient and accessible in our daily life, which makes multi-view images be easily gained by crowd-sourcing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 678, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 678, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 678, + 545, + 713 + ], + "type": "text", + "content": "Composition setup. We set up some constraints on the ratio. Depending on the category, each class is initially set with the different expected number of video cap" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9152" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 70, + 232, + 106 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 232, + 106 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 232, + 106 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 232, + 106 + ], + "type": "table", + "html": "
AnnotationCollectedValidGPU hours
Sparse219,188215,7553,806.8
Segmentation104,261\\2,316.9
Dense98,89987,20025,122.4
", + "image_path": "de14d30fc652844d3d8af2dca54c443ff8fd882a3757fbf0fae7a90b4551b041.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 140, + 286, + 246 + ], + "blocks": [ + { + "bbox": [ + 47, + 108, + 287, + 131 + ], + "lines": [ + { + "bbox": [ + 47, + 108, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 287, + 131 + ], + "type": "text", + "content": "Table 1. Data statistics, including collected amount, valid amount after cleaning, and the GPU hours for processing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 140, + 286, + 246 + ], + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 246 + ], + "type": "table", + "html": "
DatasetReal# of objects# of categoriesMulti-view3D-GT
ShapeNet [11]×51k55renderCAD model
ModelNet [96]×12k40renderCAD model
Choi et al. [20]✓2k9360° capturedRGB-D scan
Objectron [3]✓15k9limited3D bbox, pcl
GoogleScan [27]✓2kNA360° capturedRGB-D scan
Henzler et al. [45]✓2k7360° capturedpcl
ScanObjectNN [88]✓14k15limitedpcl
CO3D [74]✓19k50360° captured-
CO3D-pcl [74]✓5k50360° capturedpcl
MVIImgNet (ours)✓220k238180° captured-
MVPNet (ours)✓87k150180° capturedpcl
", + "image_path": "562369e864751c37a3d1bcf1022e8eafa7d503365c53f052752077b12873a1c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 247, + 286, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 286, + 269 + ], + "type": "text", + "content": "Table 2. Comparison between our datasets and related ones. \"pcl\" denotes point clouds, \"bbox\" means bounding boxes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 328 + ], + "type": "text", + "content": "tures regarding their generality, e.g., the number of captures for “bottles”, “bags” and “snacks” is planned to be around 2000, while the number of “chips”, “apples” and “guitars” is set to about 1000. This setting is closer to real life." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "type": "text", + "content": "Video capture. How to capture videos directly affect the quality of our data, so we draw up the following requirements as guidance for the captured videos: 1) The length of each video should be about 10 seconds. 2) The frames in the video should not be blurred. 3) Each video should capture " + }, + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "type": "text", + "content": " view of the object as much as possible. 4) The proportion of the object in the video should be above " + }, + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 331, + 287, + 439 + ], + "type": "text", + "content": ". 5) Each video can only contain one class of principal object. 6) The captured object should be stereoscopic." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 441, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 573 + ], + "type": "text", + "content": "Crowdsourcing. We employ around 1000 normal collectors from different professions (e.g., teacher, doctor, student, cook, babysitter) and ages " + }, + { + "bbox": [ + 46, + 441, + 287, + 573 + ], + "type": "inline_equation", + "content": "(20\\sim 50)" + }, + { + "bbox": [ + 46, + 441, + 287, + 573 + ], + "type": "text", + "content": ". Each of them is asked to take several videos in their daily life, (i.e., denoting diverse real-world environments) and upload them to the backend. Meanwhile, about 200 new well-trained expert data cleaners are responsible to review each submission and assure it fulfills the aforementioned capture requirements, when they may report some feedback or directly delete the unqualified submissions. The whole procedure ensures both the diversity and quality of the raw videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 583, + 145, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 145, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 145, + 597 + ], + "type": "text", + "content": "3.2. Data Processing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 602, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 651 + ], + "type": "text", + "content": "For each qualified video submission, we exploit an automatic process to obtain the common 2D and 3D annotations, including object masks, camera intrinsic and extrinsic, depth maps, and point clouds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Sparse reconstruction. Following the procedure of [64, 65], the sparse reconstruction aims to reconstruct the camera intrinsic and extrinsic for each video, by applying the COLMAP Structure-from-Motion (SfM) algorithm [75] on a series of equal-time-interval chosen frames." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 336, + 70, + 517, + 175 + ], + "blocks": [ + { + "bbox": [ + 336, + 70, + 517, + 175 + ], + "lines": [ + { + "bbox": [ + 336, + 70, + 517, + 175 + ], + "spans": [ + { + "bbox": [ + 336, + 70, + 517, + 175 + ], + "type": "image", + "image_path": "b153f4934935b54f497c7f3de180a47c3741d5d3a51ea8e2c70c2ba963d5eb4f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 178, + 518, + 190 + ], + "lines": [ + { + "bbox": [ + 331, + 178, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 331, + 178, + 518, + 190 + ], + "type": "text", + "content": "Figure 3. Some frames sampled from MVImgNet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 198, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 257 + ], + "type": "text", + "content": "Foreground object segmentation. Each frame extracted from the original video is fed to the CarveKit [70] for generating the binary foreground object mask, not only benefiting the dense reconstruction but also contributing to the further step of salient object detection (Sec. 5.3)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 259, + 546, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 546, + 390 + ], + "type": "text", + "content": "Dense reconstruction. With the sparse model output from COLMAP SfM, we employ multi-view stereo [76] of COLMAP to generate the dense depth and normal maps for each frame. We extract the depths of an object using the binary foreground masks, which are then back-projected and fused according to the normal information, yielding a densely reconstructed point cloud for each video. Finally, the point cloud is manually cleaned by: 1) Delete the object with obvious noisy or extremely sparse reconstruction. 2) Remove all backgrounds. The final derived 3D point cloud dataset - MVPNet, is illustrated in Sec. 6.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 399, + 411, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 411, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 411, + 411 + ], + "type": "text", + "content": "3.3. Dataset Summary" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 419, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 491 + ], + "type": "text", + "content": "Statistics. The statistics of MVImgNet are shown in Tab. 1, and Tab. 2 compares our datasets with other alternatives. MVImgNet includes 238 object classes, from 6.5 million frames of 219,188 videos. Fig. 3 shows some samples of MVImgNet. The annotation comprehensively covers object masks, camera parameters, and point clouds." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 494, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 612 + ], + "type": "text", + "content": "Category. We leverage WordNet [66] taxonomy that is used by ImageNet [24] to describe multi-hierarchy categories of objects and define 238 common classes. Unlike ImageNet which contains various plants and animals (nature-centric), the objects in our MVImgNet are found or used in human daily life (human-centric), where 65 classes overlap with ImageNet. The detailed category taxonomy, per-category data distributions, and more sample visualizations of MVImgNet are illustrated in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 625, + 416, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 625, + 416, + 637 + ], + "spans": [ + { + "bbox": [ + 306, + 625, + 416, + 637 + ], + "type": "text", + "content": "4.3D Reconstruction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 645, + 473, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 473, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 473, + 657 + ], + "type": "text", + "content": "4.1. Radiance Field Reconstruction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Pre-review. Recently, a series of generalizable Neural Radiance Fields (NeRF) variants [13, 87, 93, 111] have been proposed to reconstruct radiance field on-the-fly from one/few-shot source views for novel view synthesis." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9153" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 73, + 355, + 112 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 355, + 112 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 355, + 112 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 355, + 112 + ], + "type": "table", + "html": "
MethodDiffuse PSNR↑Synthetic SSIM↑360° LPIPS↓Realistic PSNR↑Synthetic SSIM↑360° LPIPS↓Real-world PSNR↑360° SSIM↑Objects LPIPS↓
Train from scratch37.170.9900.01725.490.9160.10022.180.7140.365
MVIImgNet-pretrained37.660.9900.01427.260.9300.07124.670.7400.310
", + "image_path": "e719751f302dc60027b387e478ddd7ff891b0970fdce2fb7945f4bbee86b9031.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 392, + 71, + 526, + 115 + ], + "blocks": [ + { + "bbox": [ + 74, + 114, + 331, + 124 + ], + "lines": [ + { + "bbox": [ + 74, + 114, + 331, + 124 + ], + "spans": [ + { + "bbox": [ + 74, + 114, + 331, + 124 + ], + "type": "text", + "content": "Table 3. NeRF quantitative results on three different levels of test sets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 392, + 71, + 526, + 115 + ], + "lines": [ + { + "bbox": [ + 392, + 71, + 526, + 115 + ], + "spans": [ + { + "bbox": [ + 392, + 71, + 526, + 115 + ], + "type": "table", + "html": "
Where to pretrainReal-world 360°Objects [74]
PSNR↑SSIM↑LPIPS↓
CO3D [74]24.010.7320.339
MVIImgNet-small24.080.7360.316
MVIImgNet24.670.7400.310
", + "image_path": "9b3f7d3a9720c9780347710804d13415406741139420bef2bed17954f2d0ac55.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 356, + 118, + 562, + 128 + ], + "lines": [ + { + "bbox": [ + 356, + 118, + 562, + 128 + ], + "spans": [ + { + "bbox": [ + 356, + 118, + 562, + 128 + ], + "type": "text", + "content": "Table 4. NeRF quantitative comparison with CO3D [74]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 55, + 133, + 276, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 133, + 276, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 133, + 276, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 276, + 232 + ], + "type": "image", + "image_path": "8cf5e9928a1762f9647ef48dedd1dd33aa4f416e2e8dc4f58b16937c3286e65e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 288, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 270 + ], + "type": "text", + "content": "Figure 4. NeRF qualitative comparison of train-from-scratch IBRNet, CO3D [74]-pretrained IBRNet and MVImgNet-pretrained IBRNet [93]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "content": "What can MVImgNet do? Training NeRFs that can generalize to unseen objects requires learning 3D priors from a huge amount of multi-view images. Existing state-of-the-arts either resort to learning on large-scale synthetic data [13, 87, 111], or adopt a mixed use of synthetic data and a small self-collected real dataset [93]. However, the synthetic data introduces a big domain gap with real-world objects, only a few real scenes cannot remedy this defect. We argue that: our MVImgNet perfectly fits the huge data demand of learning-based generalizable NeRF methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 401, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 497 + ], + "type": "text", + "content": "To verify this, we choose IBRNet [93] as the baseline and conduct an empirical study. We pretrain IBRNet on the full MVImgNet dataset then finetune on the training datasets used in IBRNet [93] for a few iterations, and compare with the original train-from-scratch IBRNet model in terms of generalization capability (i.e., generalizing to unseen scenes with only few-shot inputs). For fairness, we evaluate all methods under the same protocol of IBRNet." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": "Here comes a challenge about how to evaluate the generalization ability of NeRFs, since there is no official benchmark for this. To this end, we employ three different third-party object-centric datasets to form the test set. 1) The diffuse synthetic " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " object dataset [80] which contains 4 Lambertian objects. 2) The realistic synthetic " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " object dataset [64] which consists of 8 realistically fabricated objects. 3) The real-world " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " object dataset [74] which includes 88 real-world scenes from different lighting conditions. To summarize, the whole test set comprises 100 objects from 56 distinct categories, ranging from the synthetic domain to the real-world domain, which is considered to be an impartial evaluation set for generalization ability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "The quantitative and qualitative results are respectively shown in Tab. 3 and Fig. 4. Evidently, pretraining on MVImgNet improves the generalization ability of the model by a large margin. Moreover, we perform the same pretraining on CO3D [74] and MVImgNet-small (a random subset" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 344, + 135, + 508, + 213 + ], + "blocks": [ + { + "bbox": [ + 344, + 135, + 508, + 213 + ], + "lines": [ + { + "bbox": [ + 344, + 135, + 508, + 213 + ], + "spans": [ + { + "bbox": [ + 344, + 135, + 508, + 213 + ], + "type": "image", + "image_path": "1089e7076a63296b6f70a630ede941cf05604a67a5ae71d7596330049dce03e4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 219, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 306, + 219, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 219, + 547, + 243 + ], + "type": "text", + "content": "Figure 5. PSNR of IBRNet pretrained by a different number of MVImgNet data (higher is better)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 251, + 546, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 546, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 546, + 299 + ], + "type": "text", + "content": "of MVImgNet, which owns the same scale as CO3D). As shown in Tab. 4 and Fig. 4, MVImgNet shows greater power than CO3D [74]. The implementation details can be found in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 300, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 337 + ], + "type": "text", + "content": "More data, more power. Fig. 5 shows that an apparent rising trend of generalization metrics can be observed with the increase of training data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 344, + 411, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 411, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 411, + 356 + ], + "type": "text", + "content": "4.2. Multi-view Stereo" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 365, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 568 + ], + "type": "text", + "content": "Pre-review. Multi-view Stereo (MVS) [30] is a classical task in 3D computer vision, with the goal of reconstructing 3D geometry from multi-view images. Conventional methods [31,46,77,114] reconstruct 3D geometry by finding the patches matched with different images and estimating the depth according to the extrinsic camera. In recent years, deep learning methods are introduced into MVS to solve the issues such as weak textures, and nonlaplacian spheres. Two popular end-to-end methods MVSNet [109] and R-MVSNet [110] propose to encode the multi-view images and build a cost volume for predicting depth maps. Under the supervision of large-scale RGB-D datasets [2], they outperform conventional patch-matched-based approaches. However, these methods require massive amounts of RGB-D data, which is always difficult to acquire. This drives the emergence of self-supervised MVS methods [12,25,102,107]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 629 + ], + "type": "text", + "content": "What can MVImgNet do? We demonstrate that our MVImgNet is capable of benefiting the data-efficient MVS with limited training examples, which is practically meaningful considering MVS always requires the burdensome collection of RGB-D data." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "We pretrain a self-supervised MVS method, JDACS [102], on MVImgNet. Our implementation follows the original settings of JDACS. Then, we select DTU [2] dataset with 79 training samples and 22 test samples. We perform the data-efficient evaluation with limited training data, where the MVImgNet-pretrained MVsNet is finetuned on " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " of DTU training samples, and eval" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9154" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 70, + 245, + 111 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 245, + 111 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 245, + 111 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 245, + 111 + ], + "type": "table", + "html": "
Ratio2mm↑4mm↑8mm↑
5%48.61 / 50.1065.18 / 67.9775.69 / 78.58
15%54.74 / 54.6870.96 / 72.0480.32 / 81.73
25%57.29 / 58.6374.41 / 75.2083.68 / 84.28
", + "image_path": "191585e7ae228e41128c6d5014d434a52ccdc60c6843314498da9eb06330311c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 94, + 153, + 241, + 195 + ], + "blocks": [ + { + "bbox": [ + 46, + 112, + 287, + 146 + ], + "lines": [ + { + "bbox": [ + 46, + 112, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 46, + 112, + 287, + 146 + ], + "type": "text", + "content": "Table 5. MVS depth map accuracy on DTU evaluation set under different ratios of DTU training samples, in terms of training from scratch / MVImgNet-pretrained (higher is better)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 153, + 241, + 195 + ], + "lines": [ + { + "bbox": [ + 94, + 153, + 241, + 195 + ], + "spans": [ + { + "bbox": [ + 94, + 153, + 241, + 195 + ], + "type": "table", + "html": "
Where to pretrain2mm↑4mm↑8mm↑
CO3D46.1866.7278.39
MVIImgNet-small47.7366.5378.15
MVIImgNet50.1067.9778.58
", + "image_path": "2c83d1aaa41c5e0e03c2fb0de4f33113920e45f39017f824a5d37e4212bf4be9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 288, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 362 + ], + "type": "text", + "content": "uated on the DTU test set. Tab. 5 reports the accuracy given different thresholds of the error between the predicted and ground truth depth map, which indicates that the MVImgNet-pretrained model is capable of improving the model trained from scratch by a large margin under the data-efficient setup. Furthermore, we perform pretraining using CO3D [74] and MVImgNet-small. Tab. 6 indicates that our MVImgNet is stronger than CO3D, and MVImgNet-small shows comparable power as CO3D. For more implementation details, please see the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 363, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 287, + 387 + ], + "type": "text", + "content": "We advocate benchmarking NeRF and MVS methods with the help of pretraining on MVImgNet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 399, + 257, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 257, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 257, + 413 + ], + "type": "text", + "content": "5. View-consistent Image Understanding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 418, + 241, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 241, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 241, + 431 + ], + "type": "text", + "content": "5.1. View-consistent Image Classification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": "Pre-review. As indicated by Dong et al. [26], although humans can easily recognize one object from different views, deep learning models can hardly do that robustly. MVImgNet provides numerous images from different viewpoints, so we hope to enhance the model's view consistency with the help of MVImgNet, which is significantly important for realizing human-like intelligence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 581 + ], + "type": "text", + "content": "What can MVImgNet do? One naive approach is to finetune the ImageNet-pretrained ResNet-50 [44] on our MVImgNet. However, such an approach will be problematic due to the categories of two datasets are very distinct, which may cause catastrophic forgetting issues [50]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "For this reason, we create a new training set, namely MVI-Mix, by mixing the original ImageNet data with MVIImgNet. Specifically, we randomly sample 5 consecutive frames of each video in MVIImgNet, and mix the multi-view images with original ImageNet data. We also build another two artificial training sets, namely MVI-Gap and MVI-Aug, that only differ MVI-Mix with image views. MVI-Gap samples the 5 non-consecutive frames that have much larger view differences. MVI-Aug samples 1 frame from each video of MVI-Mix and applies 4 different data augmentations (i.e., random color jittering, grid mask, ro" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 360, + 70, + 492, + 118 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 287, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 231 + ], + "type": "text", + "content": "Table 6. MVS depth map accuracy comparison with CO3D [74] on DTU evaluation set. Pretrain on MVImgNet, MVImgNet-small or CO3D, and finetune using " + }, + { + "bbox": [ + 46, + 198, + 287, + 231 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 198, + 287, + 231 + ], + "type": "text", + "content": " of DTU training samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 360, + 70, + 492, + 118 + ], + "lines": [ + { + "bbox": [ + 360, + 70, + 492, + 118 + ], + "spans": [ + { + "bbox": [ + 360, + 70, + 492, + 118 + ], + "type": "table", + "html": "
DatasetConfidence VarAccuracy
ImageNet-only0.20753.09%
MVI-Aug0.10571.48%
MVI-Gap0.10377.23%
MVI-Mix0.10277.31%
", + "image_path": "bf877891f99ee5498dcf221a637f78cd142cfe01dd6ecfee3dfc8b3712cbc891.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 343, + 173, + 511, + 219 + ], + "blocks": [ + { + "bbox": [ + 305, + 121, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 305, + 121, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 545, + 164 + ], + "type": "text", + "content": "Table 7. View-consistent image classification results on MVImgNet test set using fully-supervised ResNet-50 [44]. Adding MVImgNet for training improves the view consistency (smaller Var and higher Acc indicate better view consistency)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 173, + 511, + 219 + ], + "lines": [ + { + "bbox": [ + 343, + 173, + 511, + 219 + ], + "spans": [ + { + "bbox": [ + 343, + 173, + 511, + 219 + ], + "type": "table", + "html": "
Train: CO3DI-MixTrain: MVI-Mix
Test: CO3DVar: 0.104Var: 0.155 (+0.051)
Acc: 91.25%Acc: 84.83% (-6.42%)
Test:MVIImgNetVar: 0.188 (+0.086)Var: 0.102
Acc: 57.93% (-19.38%)Test: 77.31%
", + "image_path": "c3e8ea823817d7ade858f0d548af5bc015147a2c803caed94c16896bdb83c6e5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 220, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 285 + ], + "type": "text", + "content": "Table 8. View-consistent image classification comparison with CO3D [74]. The performance drops when train on CO3DI-Mix and test on MVIImgNet (blue number) are much larger than the opposite (red number), which means that MVI-Mix pretrained model is more robust to multiview consistency than CO3DI-Mix pretrained model." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "text", + "content": "tation and erase, respectively) to it, which aims to differentiate the multi-view augmentation from the normal data augmentations. We compare the variance of the softmax confidence and the accuracy on MVImgNet test set for examining the view consistency. As illustrated in Tab. 7, adding our MVImgNet data for training can effectively improve the model's view consistency and accuracy. The test dataset used for all experiments is MVImgNet test dataset. In addition to using the convolution-based network (i.e., ResNet-50), we also apply the Transformer-based architecture, DeiT-Tiny [85]. The results are " + }, + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "inline_equation", + "content": "48.76\\%" + }, + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "text", + "content": " accuracy and 0.225 Var on ImageNet-only, VS 0.122 Var " + }, + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "inline_equation", + "content": "(\\mathbf{0.103} \\downarrow)" + }, + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "inline_equation", + "content": "73.88\\%" + }, + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "inline_equation", + "content": "(25.12 \\uparrow)" + }, + { + "bbox": [ + 304, + 300, + 545, + 467 + ], + "type": "text", + "content": " accuracy on MVI-Mix. This leads to a unanimous conclusion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 468, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 529 + ], + "type": "text", + "content": "We further construct CO3DI-Mix (corresponding to MVI-Mix) as the mixer of CO3D [74] & ImageNet under the same ratio of MVImgNet & ImageNet in MVI-Mix. As illustrated in Tab. 8, MVImgNet brings more benefits than CO3D for view-consistent image recognition." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 538, + 504, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 504, + 551 + ], + "type": "text", + "content": "5.2. View-consistent Contrastive Learning" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 616 + ], + "type": "text", + "content": "Pre-review. Contrastive Learning (CL) is one mainstream of self-supervised training techniques [16-19, 35, 40, 84]. One key factor of CL is the construction of positive/negative pairs, e.g., MoCo-v2 [17] treats an image with different augmentations, a.k.a. views, as the positive pair." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "What can MVImgNet do? One natural question is: can the viewpoints of MVImgNet serve as the positive pairs for " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "CL" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "? To answer this question, we finetune the off-the-shelf ImageNet-pretrained MoCo-v2 on MVImgNet. For each iteration, we randomly sample two frames from the same video as the positive pair and apply the original data augmentations used in MoCo-v2 [17]. Meanwhile, the frames from other videos will be treated as their negatives." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9155" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 70, + 247, + 106 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 247, + 106 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 247, + 106 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 247, + 106 + ], + "type": "table", + "html": "
Training schemeConfidence VarAccuracy
w/o finetune0.09870.26%
w/ finetune0.08671.22%
", + "image_path": "a5a8621e069763890cc7258e080873e3f2436adb703920fb3047901783ee983c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 109, + 287, + 152 + ], + "lines": [ + { + "bbox": [ + 47, + 109, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 287, + 152 + ], + "type": "text", + "content": "Table 9. View-consistency classification results under the self-supervised contrastive learning regime on MVImgNet test set. Finetuning MoCo-v2 [17] on MVImgNet improves the view consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 60, + 160, + 276, + 237 + ], + "blocks": [ + { + "bbox": [ + 60, + 160, + 276, + 237 + ], + "lines": [ + { + "bbox": [ + 60, + 160, + 276, + 237 + ], + "spans": [ + { + "bbox": [ + 60, + 160, + 276, + 237 + ], + "type": "image", + "image_path": "8fca2ba8a21e7fee99ee8035da0a5ba775212954ee03eea21d8d4e62f0d9c79b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 241, + 287, + 296 + ], + "lines": [ + { + "bbox": [ + 47, + 241, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 287, + 296 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison between MVImgNet-finetuned U2Net [73] and original U2Net for salient object detection (SOD). Left: finetuning on MVImgNet improves the performance on a hard view. Right: finetuning on MVImgNet improves the performance on two consecutive hard views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 400 + ], + "type": "text", + "content": "The MVImgNet-finetuned model is finally evaluated on the MVImgNet test dataset for examining the view consistency, where we compare the variance and mean of the softmax confidence and the accuracy. As Tab. 9 shows, finetuning on MVImgNet with CL can also improve both the model's view consistency and accuracy. More implementation details of view-consistent image classification is described in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": "In the future, it is highly recommended to regard view-consistency as a criterion for evaluating the image recognition task, and utilize MVImgNet to pretrain the models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 444, + 168, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 168, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 168, + 455 + ], + "type": "text", + "content": "5.3. View-consistent SOD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": "Pre-review. Salient Object Detection (SOD) aims to segment the most visually prominent objects in an image. Although remarkable progress has been made recently, it remains lots of challenges." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": "What can MVImgNet do? We test a state-of-the-art SOD model U2Net [73] on our MVImgNet. As Fig. 6 shows, U2Net failed to segment \"hard\" views, even though some views can be segmented with few flaws. Despite such a disappointment, the inconsistent predictions of different views caught our attention: can we improve the SOD models with multi-view consistency?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "We propose to leverage the multi-view consistency to improve SOD with the help of optical flows. Specifically, given two consecutive frames, we first calculate their optical flow and warp the flow to the segmentation mask of one of the frames. Then, a consistency loss can be calculated between the warped mask and the mask of another frame. For ease of implementation, we directly finetune U2Net on our MVImgNet. To prevent the catastrophic forgetting issue [50], in addition to the MVImgNet, we also use the original training data DUTS-TR [92]. For the \"hard\" views" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 71, + 521, + 155 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 521, + 155 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 521, + 155 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 521, + 155 + ], + "type": "image", + "image_path": "40b0bba5864a6bb332e007b8ce09103b4297013eedba4fcb071aabe23123c508.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 159, + 531, + 170 + ], + "lines": [ + { + "bbox": [ + 320, + 159, + 531, + 170 + ], + "spans": [ + { + "bbox": [ + 320, + 159, + 531, + 170 + ], + "type": "text", + "content": "Figure 7. Some 3D point clouds sampled from MVPNet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 179, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 179, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 545, + 240 + ], + "type": "inline_equation", + "content": "(\\mathrm{IoU} \\leq 0.7)" + }, + { + "bbox": [ + 305, + 179, + 545, + 240 + ], + "type": "text", + "content": " on our MVImgNet test set, the model finetuned on our MVImgNet can bring a " + }, + { + "bbox": [ + 305, + 179, + 545, + 240 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 305, + 179, + 545, + 240 + ], + "type": "text", + "content": " IoU improvement (see Fig. 6 for qualitative comparison). The implementation details of view-consistent SOD can be found in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 251, + 481, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 481, + 264 + ], + "type": "text", + "content": "6. MVPNet for 3D Understanding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 270, + 408, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 408, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 408, + 281 + ], + "type": "text", + "content": "6.1. MVPNet Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 289, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 420 + ], + "type": "text", + "content": "Derived from the dense reconstruction on MVImgNet (as mentioned in Sec. 3.2), a new large-scale real-world 3D object point cloud dataset – MVPNet, is born, which contains 87,200 point clouds with 150 categories. Fig. 7 shows some examples of MVPNet. As listed in Tab. 2, compared with existing 3D object datasets, our MVPNet contains a conspicuously richer amount of real-world object point clouds, with abundant categories covering many common objects in the real life. The detailed category taxonomy per-category data distributions of MVPNet are illustrated in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 428, + 466, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 466, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 466, + 439 + ], + "type": "text", + "content": "6.2.3D Point Cloud Classification" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "type": "text", + "content": "In this work, we focus on point cloud classification. We believe that future works may also utilize our dataset to help much more 3D understanding tasks such as indoor-scene parsing, outdoor-environment perception, pose estimation, and robotics manipulation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 508, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 567 + ], + "type": "text", + "content": "Pre-review. We advocate paying more attention to real-world setup, which is more feasible for real applications. ScanObjectNN [89] has been manifested as the most challenging point cloud classification benchmark so far, so we choose it for the major comparison with our MVPNet." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 570, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 629 + ], + "type": "text", + "content": "What can MVPNet do? We show that pretraining on MVPNet is able to aid the performance of real-world point cloud classification [36, 39, 62, 71, 72, 94, 97, 105, 106]. We pretrain several models on MVPNet, and finetune them on ScanObjectNN." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Two settings are considered for evaluation. First is PB_T50_RS in ScanObjectNN with small perturbation, translation and rotation on point cloud. Another is adding heavy rotation on PB_T50_RS, to create a more challenging setting. The results are shown in Tab. 10, where pretraining on MVPNet is able to increase the classification accuracy under most circumstances." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9156" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 273, + 162 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 273, + 162 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 273, + 162 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 273, + 162 + ], + "type": "table", + "html": "
MethodAdd Random RotationPB_T50_RS
from scratchpretrainedfrom scratchpretrained
PointNet [71]60.57 / 55.2064.25 / 59.2970.63 / 67.2867.73 / 64.12
PointNet++ [72]76.50 / 73.4278.76 / 76.5478.80 / 75.7080.22 / 76.91
DGCNN [94]80.50 / 78.4580.42 / 78.2079.44 / 76.2482.36 / 80.08
PointMLP [62]83.69 / 82.5484.87 / 83.7185.64 / 84.1485.98 / 84.38
CurveNet [97]73.96 / 69.9678.99 / 76.5974.27 / 69.4383.68 / 81.17
GDANet [106]80.33 / 79.1483.59 / 82.2979.01 / 75.9183.90 / 82.51
PAConv [105]70.91 / 65.7076.21 / 72.4772.88 / 68.6076.91 / 73.45
PCT [36]81.32 / 79.1782.08 / 80.4177.46 / 73.6484.20 / 81.94
PointMAE [67]83.17 / 80.7586.19 / 84.6077.34 / 73.5284.13 / 81.92
", + "image_path": "ffff2dac759f99f6085ff3404e5c238e68a6105f5e250a454b2f368236476c70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 213, + 261, + 256 + ], + "blocks": [ + { + "bbox": [ + 47, + 164, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 47, + 164, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 287, + 209 + ], + "type": "text", + "content": "Table 10. ScanObjectNN real-world point cloud classification results. The comparison is between the train-from-scratch model and MVPNet pretrained model. The metric is overall / average accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 213, + 261, + 256 + ], + "lines": [ + { + "bbox": [ + 73, + 213, + 261, + 256 + ], + "spans": [ + { + "bbox": [ + 73, + 213, + 261, + 256 + ], + "type": "table", + "html": "
MethodPB_T50_RS
PreTr. on MVPNetPreTr. on MVPNet-smallPreTr. on CO3D
PointNet++78.90 / 77.1179.00 / 76.7878.79 / 77.20
CurveNet83.68 / 81.1773.92 / 68.9273.78 / 69.62
PointMAE84.13 / 81.9281.85 / 79.2581.40 / 78.87
", + "image_path": "b3fd1156db226393524886775b225b1000d1e56f4532d742a3c657dba2987a6a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 310, + 257, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 257, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 257, + 323 + ], + "type": "text", + "content": "6.3. Self-supervised Point Cloud Pretraining" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 286, + 354 + ], + "type": "text", + "content": "Pre-review. Self-supervised learning has been exploited for 3D object point cloud understanding [59,67,91,113]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 356, + 287, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 287, + 439 + ], + "type": "text", + "content": "What can MVPNet do? In Tab. 10, PointMAE [67] pretrained on our MVPNet outperforms the state-of-the-art methods when finetuned on ScanObjectNN, proving the benefit of MVPNet on the self-supervised learning regime for real-world point cloud classification. Tab. 11 also shows that pretraining on MVPNet is more powerful than CO3D [74] for real-world point cloud classification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 447, + 218, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 218, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 218, + 460 + ], + "type": "text", + "content": "6.4. MVPNet Benchmark Challenge" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 465, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 287, + 513 + ], + "type": "text", + "content": "We present the MVPNet benchmark challenge for real-world point cloud classification, which contains 69,760 training and 17,440 testing samples. The results of various methods are shown in Tab. 12." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 515, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 287, + 633 + ], + "type": "text", + "content": "MVPNet is more challenging than ScanObjectNN. We firstly train models on ScanObjectNN, then conduct the test on MVPNet, which is concluded in Tab. 13. On the opposite, we also train on MVPNet and test on ScanObjectNN, which is listed in Tab. 14. Comparing Tab. 13 and Tab. 14, the accuracy drops are significantly larger when training on ScanObjectNN and testing on MVPNet, which verifies the greater challenge of our MVPNet. All the experiments in 3D understanding strictly follow the original settings of the selected backbone networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 658 + ], + "type": "text", + "content": "MVPNet is suggested for investigating 3D point cloud understanding in the future." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "We have introduced MVImgNet, a large-scale dataset of multi-view images, which is efficiently collected by shoot" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 361, + 70, + 491, + 146 + ], + "blocks": [ + { + "bbox": [ + 47, + 258, + 287, + 301 + ], + "lines": [ + { + "bbox": [ + 47, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 287, + 301 + ], + "type": "text", + "content": "Table 11. ScanObjectNN real-world point cloud classification comparison with CO3D [74]. Pretrain on MVImgNet, MVImgNet-small or CO3D, and finetune on ScanObjectNN. The metric is overall / average accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 361, + 70, + 491, + 146 + ], + "lines": [ + { + "bbox": [ + 361, + 70, + 491, + 146 + ], + "spans": [ + { + "bbox": [ + 361, + 70, + 491, + 146 + ], + "type": "table", + "html": "
MethodOverall / Average Accuracy
PointNet [71]70.72 / 54.46
PointNet++ [72]79.15 / 58.24
DGCNN [94]86.49 / 63.98
PointMLP [62]88.89 / 73.64
CurveNet [97]88.88 / 75.37
GDANet [106]89.54 / 68.41
PACConv [105]83.35 / 59.13
PCT [36]91.49 / 75.41
", + "image_path": "4b4ffc7c0eb70ff3f42327610f69272f70b21a148933740a1b90d78a2d180b78.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 336, + 178, + 516, + 216 + ], + "blocks": [ + { + "bbox": [ + 306, + 148, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 306, + 148, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 545, + 170 + ], + "type": "text", + "content": "Table 12. Quantitative results on our new MVPNet benchmark for real-world point cloud classification." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 336, + 178, + 516, + 216 + ], + "lines": [ + { + "bbox": [ + 336, + 178, + 516, + 216 + ], + "spans": [ + { + "bbox": [ + 336, + 178, + 516, + 216 + ], + "type": "table", + "html": "
Methodtrain: MVPNet test: MVPNettrain: ScanObjectNN test: MVPNet
PointNet++ [72]79.15 / 58.2415.13 (-64.02) / 7.96 (-50.28)
CurveNet [97]88.88 / 75.3746.95 (-41.93) / 28.24 (-47.13)
", + "image_path": "a4afc91ba65f48cc14cc323ea8127e3bfc4eb1bcbed8fd5f93b90084bd372c67.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 332, + 253, + 520, + 289 + ], + "blocks": [ + { + "bbox": [ + 306, + 218, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 306, + 218, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 218, + 545, + 251 + ], + "type": "text", + "content": "Table 13. Quantitative comparison while training on ScanObjectNN, testing on MVPNet. The evaluation metric is overall / average accuracy." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 253, + 520, + 289 + ], + "lines": [ + { + "bbox": [ + 332, + 253, + 520, + 289 + ], + "spans": [ + { + "bbox": [ + 332, + 253, + 520, + 289 + ], + "type": "table", + "html": "
Methodtrain: ScanObjectNN\ntest: ScanObjectNNtrain: MVPNet\ntest: ScanObjectNN
PointNet++ [72]76.50 / 73.4240.35 (-36.15) / 33.59 (-39.83)
CurveNet [97]73.96 / 69.9651.84 (-22.12) / 46.27 (-23.69)
", + "image_path": "275ca20a8c4cf7fffb7e3d924a2ef877ab41570ac28ca717efad266bfa11ab37.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 293, + 545, + 325 + ], + "lines": [ + { + "bbox": [ + 306, + 293, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 545, + 325 + ], + "type": "text", + "content": "Table 14. Quantitative comparison while training on MVPNet, testing on ScanObjectNN. The evaluation metric is overall / average accuracy." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 334, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 489 + ], + "type": "text", + "content": "ing videos of real-world objects. The multi-view nature endows our dataset with 3D-aware visual signals, making MVImgNet a soft bridge to link 2D and 3D vision. To probe the power of MVImgNet, we conduct a host of pilot experiments on various visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising effectiveness, expecting more future explorations. As a bonus of MVImgNet, a point cloud dataset - MVPNet is derived. Experiments show MVPNet can benefit real-world 3D object classification. As a broader impact on society, our datasets delineates a world - that is closer to a colorful and vivid real " + }, + { + "bbox": [ + 304, + 334, + 545, + 489 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 304, + 334, + 545, + 489 + ], + "type": "text", + "content": " world - where we human lives." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 501, + 402, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 402, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 402, + 514 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 521, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 712 + ], + "type": "text", + "content": "The work was supported in part by NSFC with Grant No. 62293482, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone. It was also partially supported by NSFC-62172348, Outstanding Yound Fund of Guangdong Province with No. 2023B1515020055 and Shenzhen General Project with No. JCYJ20220530143604010, the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). It was also sponsored by CCF-Tencent Open Research Fund." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9157" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 286, + 111 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 286, + 111 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 286, + 111 + ], + "type": "text", + "content": "[1] Gradient-based learning applied to document recognition, 1998. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 114, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 114, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 58, + 114, + 287, + 146 + ], + "type": "text", + "content": "[2] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 148, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 148, + 287, + 190 + ], + "type": "text", + "content": "[3] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 192, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 192, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 57, + 192, + 287, + 224 + ], + "type": "text", + "content": "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 225, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 225, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 57, + 225, + 287, + 267 + ], + "type": "text", + "content": "[5] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 270, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 270, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 57, + 270, + 287, + 312 + ], + "type": "text", + "content": "[6] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In CVPR, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 315, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 315, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 315, + 287, + 369 + ], + "type": "text", + "content": "[7] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 2008. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 370, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 370, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 57, + 370, + 287, + 412 + ], + "type": "text", + "content": "[8] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 415, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 415, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 287, + 467 + ], + "type": "text", + "content": "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 470, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 470, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 470, + 287, + 523 + ], + "type": "text", + "content": "[10] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "type": "text", + "content": "[11] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 581, + 287, + 624 + ], + "type": "text", + "content": "[12] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rcomvsnet: Unsupervised multi-view stereo with neural rendering. arXiv preprint arXiv:2203.03949, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 287, + 668 + ], + "type": "text", + "content": "[13] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[14] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In The 49th annual meeting of the association for computational linguistics: human language technologies, 2011. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 149 + ], + "type": "text", + "content": "[16] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 151, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 151, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 151, + 545, + 183 + ], + "type": "text", + "content": "[17] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 185, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 185, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 185, + 545, + 205 + ], + "type": "text", + "content": "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "type": "text", + "content": "[19] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 545, + 271 + ], + "type": "text", + "content": "[20] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 273, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 273, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 273, + 545, + 327 + ], + "type": "text", + "content": "[21] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 328, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 328, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 312, + 328, + 545, + 380 + ], + "type": "text", + "content": "[22] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 545, + 425 + ], + "type": "text", + "content": "[23] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 459 + ], + "type": "text", + "content": "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "type": "text", + "content": "[25] Yikang Ding, Qingtian Zhu, Xiangyue Liu, Wentao Yuan, Haotian Zhang, and Chi Zhang. Kd-mvs: Knowledge distillation based self-supervised learning for multi-view stereo. In ECCV, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 545, + 547 + ], + "type": "text", + "content": "[26] Yinpeng Dong, Shouwei Ruan, Hang Su, Caixin Kang, Xingxing Wei, and Jun Zhu. Viewfool: Evaluating the robustness of visual recognition to adversarial viewpoints. arXiv preprint arXiv:2210.03895, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 545, + 602 + ], + "type": "text", + "content": "[27] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv preprint arXiv:2204.11918, 2022. 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 604, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 604, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 604, + 545, + 635 + ], + "type": "text", + "content": "[28] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2015. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 637, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 637, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 637, + 545, + 679 + ], + "type": "text", + "content": "[29] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In ICCV, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[30] Yasutaka Furukawa and Carlos Hernández. Multi-view stereo: A tutorial. Foundations and Trends in Computer Graphics and Vision, 2015. 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9158" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[31] David Gallup, Jan-Michael Frahm, Philippos Mordohai, and Marc Pollefeys. Variable baseline/resolution stereo. In CVPR, 2008. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 152 + ], + "type": "text", + "content": "[32] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In CVPR, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 153, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 153, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 153, + 288, + 186 + ], + "type": "text", + "content": "[33] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 187, + 235, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 187, + 235, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 187, + 235, + 198 + ], + "type": "text", + "content": "[34] Ross Girshick. Fast r-cnn. In ICCV, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 200, + 288, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 200, + 288, + 276 + ], + "spans": [ + { + "bbox": [ + 53, + 200, + 288, + 276 + ], + "type": "text", + "content": "[35] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In NeurIPS, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 278, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 278, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 278, + 288, + 312 + ], + "type": "text", + "content": "[36] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 2021. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "type": "text", + "content": "[37] Timo Hackel, Nikolay Savinov, Luber Ladicky, Jan Wegner, Konrad Schindler, and Marc Pollefeys. Semantic3d.net: A new large-scale point cloud classification benchmark. In ISPRS, 2017. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 392 + ], + "type": "text", + "content": "[38] Abdullah Hamdi and Bernard Ghanem. Towards analyzing semantic robustness of deep neural networks. arXiv preprint arXiv:1904.04621, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 393, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 288, + 426 + ], + "type": "text", + "content": "[39] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In ICCV, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 427, + 288, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 427, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 53, + 427, + 288, + 460 + ], + "type": "text", + "content": "[40] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, June 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 462, + 288, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 462, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 462, + 288, + 495 + ], + "type": "text", + "content": "[41] Kaiming He, Ross Girshick, and Piotr Dólar. Rethinking imagenet pre-training. arXiv preprint arXiv:1811.08883, 2018.2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 496, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 496, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 496, + 288, + 518 + ], + "type": "text", + "content": "[42] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 520, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 520, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 520, + 288, + 553 + ], + "type": "text", + "content": "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification. In ICCV, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 555, + 288, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 288, + 587 + ], + "type": "text", + "content": "[44] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 589, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 589, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 589, + 288, + 633 + ], + "type": "text", + "content": "[45] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "type": "text", + "content": "[46] Xiaoyan Hu and Philippos Mordohai. Least commitment, viewpoint-based, multi-view stereo. In International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission, 2012. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "type": "text", + "content": "[47] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 3DV, 2016. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[48] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, 2013. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "type": "text", + "content": "[49] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "type": "text", + "content": "[50] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 219, + 545, + 262 + ], + "type": "text", + "content": "[51] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In CVPR, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 296 + ], + "type": "text", + "content": "[52] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 545, + 352 + ], + "type": "text", + "content": "[53] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 354, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 375 + ], + "type": "text", + "content": "[54] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 376, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 545, + 408 + ], + "type": "text", + "content": "[55] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 410, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 410, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 312, + 410, + 545, + 443 + ], + "type": "text", + "content": "[56] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 312, + 444, + 545, + 487 + ], + "type": "text", + "content": "[57] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. arXiv preprint arXiv:2108.10257, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 312, + 488, + 545, + 532 + ], + "type": "text", + "content": "[58] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 533, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 545, + 566 + ], + "type": "text", + "content": "[59] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 567, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 567, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 312, + 567, + 545, + 611 + ], + "type": "text", + "content": "[60] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 545, + 644 + ], + "type": "text", + "content": "[61] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 646, + 545, + 689 + ], + "type": "text", + "content": "[62] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 7, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[63] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9159" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "text", + "content": "Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "type": "text", + "content": "[64] B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ramamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 142, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 288, + 196 + ], + "type": "text", + "content": "[65] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Trans. Graph., 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 287, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 287, + 242 + ], + "type": "text", + "content": "[66] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 244, + 288, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 244, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 244, + 288, + 277 + ], + "type": "text", + "content": "[67] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 278, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 278, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 278, + 288, + 312 + ], + "type": "text", + "content": "[68] Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M Seitz. Photoshop: photorealistic materials for large-scale shape collections. ACM Trans. Graph., 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 357 + ], + "type": "text", + "content": "[69] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 379 + ], + "type": "text", + "content": "[70] Open Source Project. CarveKit - image-background-remove-tool. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 381, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 288, + 415 + ], + "type": "text", + "content": "[71] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 416, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 416, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 53, + 416, + 288, + 450 + ], + "type": "text", + "content": "[72] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 451, + 288, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 451, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 451, + 288, + 495 + ], + "type": "text", + "content": "[73] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 2020. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 497, + 288, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 497, + 288, + 551 + ], + "spans": [ + { + "bbox": [ + 53, + 497, + 288, + 551 + ], + "type": "text", + "content": "[74] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 552, + 288, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 552, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 53, + 552, + 288, + 574 + ], + "type": "text", + "content": "[75] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 576, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 576, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 576, + 288, + 609 + ], + "type": "text", + "content": "[76] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 611, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 611, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 611, + 288, + 643 + ], + "type": "text", + "content": "[77] Shuhan Shen. Accurate multiple view 3d reconstruction using patch-based stereo for large-scale scenes. TIP, 2013. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 645, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 645, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 645, + 288, + 678 + ], + "type": "text", + "content": "[78] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "type": "text", + "content": "[79] Arjun Singh, James Sha, Karthik S Narayan, Tudor Achim, and Pieter Abbeel. Bigbird: A large-scale 3d database of object instances. In ICRA, 2014. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 116 + ], + "type": "text", + "content": "[80] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, 2019. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 149 + ], + "type": "text", + "content": "[81] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 150, + 545, + 194 + ], + "type": "text", + "content": "[82] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 195, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 545, + 248 + ], + "type": "text", + "content": "[83] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 545, + 293 + ], + "type": "text", + "content": "[84] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 312, + 294, + 545, + 337 + ], + "type": "text", + "content": "[85] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "type": "text", + "content": "[86] Jonathan Tremblay, Moustafa Meshry, Alex Evans, Jan Kautz, Alexander Keller, Sameh Khamis, Charles Loop, Nathan Morrical, Koki Nagano, Towaki Takikawa, et al. Rtmv: A ray-traced multi-view synthetic dataset for novel view synthesis. arXiv preprint arXiv:2205.07058, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 393, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 393, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 312, + 393, + 545, + 425 + ], + "type": "text", + "content": "[87] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, 2021. 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "type": "text", + "content": "[88] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 471, + 545, + 514 + ], + "type": "text", + "content": "[89] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 3, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 515, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 557 + ], + "type": "text", + "content": "[90] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 545, + 592 + ], + "type": "text", + "content": "[91] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "type": "text", + "content": "[92] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "type": "text", + "content": "[93] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 4, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[94] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9160" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "type": "text", + "content": "graph cnn for learning on point clouds. ACM Trans. Graph., 2019. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 286, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 286, + 118 + ], + "type": "text", + "content": "[95] Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Online object tracking: A benchmark. In CVPR, 2013. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 119, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 119, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 54, + 119, + 286, + 162 + ], + "type": "text", + "content": "[96] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 164, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 164, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 54, + 164, + 286, + 196 + ], + "type": "text", + "content": "[97] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In ICCV, 2021. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 198, + 286, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 198, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 286, + 241 + ], + "type": "text", + "content": "[98] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In ECCV, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 243, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 286, + 274 + ], + "type": "text", + "content": "[99] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond pascal: A benchmark for 3d object detection in the wild. In WACV, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 276, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 286, + 308 + ], + "type": "text", + "content": "[100] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In ICCV, 2013. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 310, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 286, + 353 + ], + "type": "text", + "content": "[101] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. IJCV, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 286, + 387 + ], + "type": "text", + "content": "[102] Hongbin Xu, Zhipeng Zhou, Yu Qiao, Wenxiong Kang, and Qiuxia Wu. Self-supervised multi-view stereo via effective co-segmentation and data-augmentation. In AAAI, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 388, + 286, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 286, + 420 + ], + "type": "text", + "content": "[103] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 422, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 286, + 453 + ], + "type": "text", + "content": "[104] Mutian Xu, Pei Chen, Haolin Liu, and Xiaoguang Han. Toscane: A large-scale dataset for understanding 3d tabletop scenes. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 456, + 286, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 488 + ], + "type": "text", + "content": "[105] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 490, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 286, + 533 + ], + "type": "text", + "content": "[106] Mutian Xu, Junhao Zhang, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi, and Yu Qiao. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In AAAI, 2021. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 535, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 286, + 566 + ], + "type": "text", + "content": "[107] Jiayu Yang, Jose M Alvarez, and Miaomiao Liu. Self-supervised learning of depth inference for multi-view stereo. In CVPR, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 568, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 610 + ], + "type": "text", + "content": "[108] Zhenpei Yang, Zaiwei Zhang, and Qixing Huang. Hm3d-abo: A photo-realistic dataset for object-centric multi-view 3d reconstruction. arXiv preprint arXiv:2206.12356, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "type": "text", + "content": "[109] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "type": "text", + "content": "[110] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In CVPR, 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[111] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 4, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 195 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[112] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[113] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[114] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In CVPR, 2014. 5" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9161" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_content_list.json b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..83164e28a2bd871029e31b6143990a6a2ec11588 --- /dev/null +++ b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_content_list.json @@ -0,0 +1,1733 @@ +[ + { + "type": "text", + "text": "MaLP: Manipulation Localization Using a Proactive Scheme", + "text_level": 1, + "bbox": [ + 178, + 130, + 792, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vishal Asnani1, Xi Yin2, Tal Hassner2, Xiaoming Liu1 \n1*Michigan State University, 2Meta AI", + "bbox": [ + 250, + 179, + 715, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1{asnanivi, liuxm}@msu.edu, 2{yinxi, thassner}@meta.com", + "bbox": [ + 246, + 217, + 722, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advancements in the generation quality of various Generative Models (GMs) has made it necessary to not only perform binary manipulation detection but also localize the modified pixels in an image. However, prior works termed as passive for manipulation localization exhibit poor generalization performance over unseen GMs and attribute modifications. To combat this issue, we propose a proactive scheme for manipulation localization, termed MaLP. We encrypt the real images by adding a learned template. If the image is manipulated by any GM, this added protection from the template not only aids binary detection but also helps in identifying the pixels modified by the GM. The template is learned by leveraging local and global-level features estimated by a two-branch architecture. We show that MaLP performs better than prior passive works. We also show the generalizability of MaLP by testing on 22 different GMs, providing a benchmark for future research on manipulation localization. Finally, we show that MaLP can be used as a discriminator for improving the generation quality of GMs. Our models/codes are available at www.github.com/vishal3477/pro_LOC.", + "bbox": [ + 75, + 300, + 473, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 646, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We witness numerous Generative Models (GMs) [8, 9, 15, 17, 23-25, 28, 34, 39, 44, 50, 52, 60] being proposed to generate realistic-looking images. These GMs can not only generate an entirely new image [23, 24], but also perform partial manipulation of an input image [9, 9, 28, 60]. The proliferation of these GMs has made it easier to manipulate personal media for malicious use. Prior methods to combat manipulated media focus on binary detection [1, 2, 5, 11, 14, 30, 45, 48, 55, 56], using mouth movement, model parsing, hand-crafted features, etc.", + "bbox": [ + 75, + 671, + 468, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works go one step further than detection, i.e. manipulation localization, which is defined as follows: given", + "bbox": [ + 75, + 823, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/255dcfe857989d1c6bbc0bd011e3dbfca1999f291260392b4df061bf352637da.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 718, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1fdf50dd8a64232386fce9fd8f9d76bd337ad02ae971637e30a3ddf39b11d8dd.jpg", + "image_caption": [ + "(b)", + "Figure 1. (a) High-level idea of MaLP. We encrypt the image by adding a learnable template, which helps to estimate the fakeness map. (b) The cosine similarity (CS) between ground-truth and predicted fakeness maps for 22 unseen GMs. The performance is better for almost all GMs when using our proactive approach." + ], + "image_footnote": [], + "bbox": [ + 720, + 268, + 888, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a partially manipulated image by a GM (e.g. STGAN [28] modifying hair colors of a face image), the goal is to identify which pixels are modified by estimating a fakeness map [21]. Identifying modified pixels helps to determine the severity of the fakeness in the image, and aid mediaforensics [11, 21]. Also, manipulation localization provides an understanding of the attacker's intent for modification which may further benefit identifying attack toolchains used [13].", + "bbox": [ + 496, + 501, + 892, + 637 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent methods for manipulation localization [27,37,49] focus on estimating the manipulation mask of face-swapped images. They localize modified facial attributes by leveraging attention mechanisms [11], patch-based classifier [4], and face-parsing [21]. The main drawback of these methods is that they do not generalize well to GMs unseen in training. That is when the test images and training images are modified by different GMs, which will likely happen given the vast number of existing GMs. Thus, our work aims for a localization method generalizable to unseen GMs.", + "bbox": [ + 496, + 640, + 893, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "All aforementioned methods are based on a passive scheme as the method receives an image as is for estimation. Recently, proactive methods are gaining success for deepfake tasks such as detection [1], disruption [46, 57], and tagging [51]. These methods are considered proactive as they add different types of signals known as templates for encrypting the image before it is manipulated by a GM.", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*All data sourcing, modeling codes, and experiments were developed at Michigan State University. Meta did not obtain the data/codes or conduct any experiments in this work.", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12343", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This template can be one-hot encoding [51], adversarial perturbation [46], or a learnable noise [1], and is optimized to improve the performance of the defined tasks.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by [1], we propose a Proactive scheme for MAnipulation Localization, termed as MaLP, in order to improve generalization. Specifically, MaLP learns an optimized template which, when added to real images, would improve manipulation localization, should they get manipulated. This manipulation can be done by an unseen GM trained on either in-domain or out-of-domain datasets. Furthermore, face manipulation may involve modifying facial attributes unseen in training (e.g. train on hair color modification yet test on gender modification). MaLP incorporates three modules that focus on encryption, detection, and localization. The encryption module selects and adds the template from the template set to the real images. These encrypted images are further processed by localization and detection modules to perform the respective tasks.", + "bbox": [ + 75, + 138, + 468, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Designing a proactive manipulation localization approach comes with several challenges. First, it is not straightforward to formulate constraints for learning the template unsupervisedly. Second, calculating a fakeness map at the same resolution as the input image is computationally expensive if the decision for each pixel has to be made. Prior works [4, 11] either down-sample the images or use a patch-wise approach, both of which result in inaccurate low-resolution fakeness maps. Lastly, the templates should be generalizable to localize modified regions from unseen GMs.", + "bbox": [ + 75, + 367, + 468, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design a two-branch architecture consisting of a shallow CNN network and a transformer to optimize the template during training. While the former leverages local-level features due to its shallow depth, the latter focuses on global-level features to better capture the affinity of the far apart regions. The joint training of both networks enables the MaLP to learn a better template, having embedded the information of both levels. During inference, the CNN network alone is sufficient to estimate the fakeness map with a higher inference efficiency. Compared to prior passive works [11, 21], MaLP improves the generalization performance on unseen GMs. We also demonstrate that MaLP can be used as a discriminator for fine-tuning conventional GMs to improve the quality of GM-generated images.", + "bbox": [ + 75, + 534, + 468, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make the following contributions.", + "bbox": [ + 94, + 748, + 429, + 762 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to propose a proactive scheme for image manipulation localization, applicable to both face and generic images.", + "- Our novel two-branch architecture uses both local and global level features to learn a set of templates in an unsupervised manner. The framework is guided by constraints based on template recovery, fakeness maps classification, and high cosine similarity between predicted and ground-truth fakeness maps." + ], + "bbox": [ + 76, + 763, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/90ef4124fc4762a168c21ea4da4878d0e54582bcbe98a6dff7f50b86cf55d690.jpg", + "table_caption": [ + "Table 1. Comparison of our approach with prior works on manipulation localization and proactive schemes. We show the generalization ability of all works across different facial attribute modifications, unseen GMs trained on datasets with the same domain (in-domain) and different domains (out-domain). [Keys: Attr.: Attributes, Imp.: Improving, L.: Localization, D.: Detection]" + ], + "table_footnote": [], + "table_body": "
WorkSchemeTaskTemplateGeneralizationImp. GM
Attr.In-domainOut-domain
[51]ProactiveTagFix✓✓✘✘
[47]ProactiveDisruptLearn✓✘✘✘
[46]ProactiveDisruptLearn✓✓✘✘
[57]ProactiveDisruptLearn✓✘✘✘
[1]ProactiveD.Learn✘✓✓✘
[37]PassiveL. + D.-✘✘✘✘
[49]PassiveL. + D.-✘✘✘✘
[27]PassiveL. + D.-✘✓✘✘
[11]PassiveL. + D.-✓✓✘✘
[4]PassiveL. + D.-✘✓✘✘
[21]PassiveL. + D.-✓✓✘✘
MaLPProactiveL. + D.Learn✓✓✓✓
", + "bbox": [ + 501, + 172, + 888, + 319 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MaLP can be used as a plug-and-play discriminator module to fine-tune the generative model to improve the quality of the generated images.", + "- Our method outperforms State-of-The-Art (SoTA) methods in manipulation localization and detection. Furthermore, our method generalizes well to GMs and modified attributes unseen in training. To facilitate the research of localization, we develop a benchmark for evaluating the generalization of manipulation localization, on images where the train and test GMs are different." + ], + "bbox": [ + 500, + 335, + 890, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 503, + 640, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Manipulation Localization. Prior works tackle manipulation localization by adopting a passive scheme. Some of them focus on forgery attacks like removal, copymove, and splicing using multi-task learning [37]. Songsriin et al. [49] leverage facial landmarks [10] for manipulation localization. Li et al. [27] estimate the blended boundary for forged face-swap images. [11] uses an attention mechanism to leverage the relationship between pixels and [4] uses a patch-based classifier to estimate modified regions. Recently, Huang et al. [21] utilize gray-scale maps as ground truth for manipulation localization and leverage face parsing with an attention mechanism for prediction. The passive methods discussed above suffer from the generalization issue [4, 10, 11, 21, 37, 49] and estimate a low-resolution fakeness map [11] which is less accurate for the localization purpose. MaLP generalizes better to modified attributes and GMs unseen in training.", + "bbox": [ + 496, + 532, + 890, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Proactive Scheme. Recently, proactive schemes are developed for various tasks. Wang et al. [51] leverage the recovery of embedded one-hot encoding messages to perform deepfake tagging. A small perturbation is added onto the images by Segalis et al. [47] to disrupt the output of a GM. The same task is performed by Ruiz et al. [46] and Yeh et al. [57], both adding adversarial noise onto the in", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12344", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "put images. Asnani et al. [1] propose a framework based on adding a learnable template to input images for generalized manipulation detection. Unlike prior works, which focus on binary detection, deepfake disruption, or tagging, our work emphasizes on manipulation localization. We show the comparison of our approach with prior works in Tab. 1.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Manipulation Detection. The advancement in manipulation detection keeps reaching new heights. Prior works propose to combat deepfakes by exploiting frequency domain patterns [53], up-sampling artifacts [59], model parsing [2], hand-crafted features [35], lip motions [45], and self-attention [11]. Recent methods use self-blended images [48], real-time deviations [14], and self-supervised learning with adversarial training [5]. Finally, methods based on contrastive learning [56] and proactive scheme [1] have explicitly focused on generalized manipulation detection across unknown GMs.", + "bbox": [ + 75, + 185, + 470, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 76, + 369, + 267, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 397, + 279, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Passive Manipulation Localization Let $\\pmb{I}^{R}$ be a set of real images that are manipulated by a GM $G$ to output the set of manipulated images $G(\\pmb{I}^{R})$ . Prior passive works perform manipulation localization by estimating the fakeness map $M_{pred}$ with the following objective:", + "bbox": [ + 76, + 425, + 468, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\min _ {\\theta_ {\\mathcal {E}}} \\left\\{\\sum_ {j} \\left(\\left| \\left| \\mathcal {E} \\left(G \\left(I _ {j} ^ {R}\\right); \\theta_ {\\mathcal {E}}\\right) - M _ {G T} \\right| \\right| _ {2}\\right) \\right\\}, \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 508, + 468, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ denotes the passive framework with parameters $\\theta_{\\mathcal{E}}$ and $M_{GT}$ is the ground-truth fakeness map.", + "bbox": [ + 76, + 555, + 467, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To represent the fakeness map, some prior methods [11, 27, 49] choose a binary map by applying a threshold on the difference between the real and manipulated images. This is undesirable as the threshold selection is highly subjective and sensitive, leading to inaccurate fakeness maps. Therefore, we adopt the continuous gray-scale map for calculating the ground-truth fakeness maps [21], formulated as:", + "bbox": [ + 75, + 585, + 468, + 683 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {G T} = \\operatorname {G r a y} \\left(\\left| \\boldsymbol {I} ^ {R} - G \\left(\\boldsymbol {I} ^ {R}\\right) \\right|\\right) / 2 5 5, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 693, + 468, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Gray(.)$ converts the image to gray-scale.", + "bbox": [ + 78, + 720, + 403, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proactive Scheme Asnani et al. [1] define adding the template as a transformation $\\mathcal{T}$ applied to images $I^R$ , resulting in the encrypted images $\\mathcal{T}(I^R)$ . The added template acts as a signature of the defender and is learned during the training, aiming to improve the performance of the task at hand, e.g. detection, disruption, and tagging. Motivated by [1] that uses multiple templates, we have a set of $n$ orthogonal templates $\\mathcal{S} = \\{S_1, S_2, \\ldots, S_n\\}$ where $S_i \\in \\mathbb{R}^{128 \\times 128}$ , for a real image $I_j^R \\in I^R$ , transformation $\\mathcal{T}$ is defined as:", + "bbox": [ + 75, + 739, + 468, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right) = \\boldsymbol {I} _ {j} ^ {R} + \\boldsymbol {S} _ {i}, \\text {w h e r e} i \\in \\{1, 2, \\dots , n \\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 883, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The templates are optimized such that adding them to the real images wouldn't result in a noticeable visual difference, yet helps manipulation localization.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proactive Manipulation Localization. Unlike the passive schemes [11, 21, 27, 37], we learn an optimal template set to help manipulation localization. For the encrypted images $\\mathcal{T}(\\boldsymbol{I}^R)$ , we formulate the estimation of the fakeness map as:", + "bbox": [ + 498, + 138, + 890, + 198 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\min _ {\\theta_ {\\mathcal {E} _ {P}}, \\boldsymbol {S} _ {i}} \\left\\{\\sum_ {j} \\left(\\left| \\left| \\mathcal {E} _ {P} \\left(G \\left(\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right)\\right); \\theta_ {\\mathcal {E} _ {P}}\\right) - \\boldsymbol {M} _ {G T} \\right| \\right| _ {2}\\right) \\right\\}. \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 212, + 890, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_P$ is the proactive framework with parameters $\\theta_{\\mathcal{E}_P}$ .", + "bbox": [ + 501, + 258, + 888, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, as the output of the GM has changed from images in set $G(\\mathbf{I}^R)$ to images in set $G(\\mathcal{T}(\\mathbf{I}^R))$ , in our proactive approach, the calculation of the ground-truth fakeness map shall be changed from Eq. 2 to the follows:", + "bbox": [ + 498, + 275, + 890, + 333 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {G T} = \\operatorname {G r a y} \\left(\\left| \\boldsymbol {I} ^ {R} - G \\left(\\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right)\\right) \\right|\\right) / 2 5 5. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 345, + 890, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Manipulation Localization", + "text_level": 1, + "bbox": [ + 500, + 391, + 740, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MaLP consists of three modules: encryption, localization, and detection. The encryption module is used to encrypt the real images. The localization module estimates the fakeness map using a two-branch architecture. The detection module performs binary detection for the encrypted and manipulated images by recovering the template and using the classifier in the localization module. All three modules, as detailed next, are trained in an end-to-end manner.", + "bbox": [ + 498, + 417, + 890, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Encryption Module", + "text_level": 1, + "bbox": [ + 500, + 566, + 689, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the procedure in [1], we add a randomly selected learnable template from the template set to a real image. We control the strength of the added template using a hyperparameter $m$ , which prevents the degradation of the image quality. The encryption process is summarised below:", + "bbox": [ + 498, + 593, + 890, + 664 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}\\right) = \\boldsymbol {I} _ {j} ^ {R} + m \\times \\boldsymbol {S} _ {i} \\text {w h e r e} i = \\operatorname {R a n d} (1, 2, \\dots , n). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 676, + 890, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We select the value of $m$ as $30\\%$ for our framework.", + "bbox": [ + 500, + 707, + 844, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We optimize the template set by focusing on properties like low magnitude, orthogonality, and high-frequency content [1]. The properties are applied as constraints as follows.", + "bbox": [ + 498, + 722, + 890, + 765 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {T} = \\lambda_ {1} \\times \\sum_ {i = 1} ^ {n} \\left\\| \\boldsymbol {S} _ {i} \\right\\| _ {2} + \\lambda_ {2} \\times \\sum_ {\\substack {i, j = 1 \\\\ i \\neq j}} ^ {n} \\operatorname {CS} \\left(\\boldsymbol {S} _ {i}, \\boldsymbol {S} _ {j}\\right) + \\lambda_ {3} \\times \\left\\| \\mathcal {L} (\\mathfrak {F} (\\boldsymbol {S})) \\right\\| _ {2}, \\tag{7}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 779, + 893, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where CS is the cosine similarity, $\\mathcal{L}$ is the low-pass filter, $\\mathfrak{F}$ is the fourier transform, $\\lambda_1,\\lambda_2,\\lambda_3$ are weights for losses of low magnitude, orthogonality and high-frequency content, respectively.", + "bbox": [ + 498, + 835, + 890, + 896 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12345", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/601d127cebb1148b646b0d3f2a73692e77c71504fbf39a5a4281021d4de09cca.jpg", + "image_caption": [ + "Figure 2. The overview of MaLP. It includes three modules: encryption, localization, and detection. We randomly select a template from the template set and add it to the real image as encryption. The GM is used in inference mode to manipulate the encrypted image. The detection module recovers the added template for binary detection. The localization module uses a two-branch architecture to estimate the fakeness map. Lastly, we apply the classifier to the fakeness map to better distinguish them from each other. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 880, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Localization Module", + "text_level": 1, + "bbox": [ + 76, + 417, + 274, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To design the localization module, we consider two desired properties: a larger receptive field for fakeness map estimation and high inference efficiency. A network with a large receptive field will consider far-apart regions in learning features for localization. Yet, large receptive fields normally come from deeper networks, implying slower inference.", + "bbox": [ + 75, + 446, + 468, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In light of these properties, we design a two-branch architecture consisting of a shallow CNN network $\\mathcal{E}_C$ and a ViT transformer [12] $\\mathcal{E}_T$ (see Fig. 2). The intuition is to have one shallow branch to capture local features, and one deeper branch to capture global features. While training with both branches helps to learn better templates, in inference we only use the shallow branch for a higher efficiency. Specifically, the shallow CNN network has 10 layers which is efficient in inference but can only capture the local features due to small receptive fields. To capture global information, we adopt the ViT transformer. With the self-attention between the image patches, the transformer can estimate the fakeness map considering the far-apart regions.", + "bbox": [ + 75, + 539, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Both the CNN and transformer are trained jointly to estimate a better template set, resembling the concept of the ensemble of networks. We empirically show that training both networks simultaneously results in higher performance than training either network separately. As the shallow CNN network is much faster in inference than the transformer, we use the transformer only in training to optimize the templates and switch off the transformer branch in inference.", + "bbox": [ + 75, + 736, + 468, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate the fakeness map, we leverage the supervision of the ground-truth fakeness map in Eq. 5. For fake images, we maximize the cosine similarity $(CS)$ and struc", + "bbox": [ + 75, + 857, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tural similarity index measure $(SS)$ between the predicted and ground-truth fakeness map. However, the fakeness map should be a zero image for encrypted images. Therefore, we apply an $L_{2}$ loss [21] to minimize the predicted map to zero for encrypted images. To maximize the difference between the two fakeness maps, we further minimize the cosine similarity between the predicted map from encrypted images and $M_{GT}$ . The localization loss is defined as:", + "bbox": [ + 496, + 417, + 893, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {L} = \\left\\{ \\begin{array}{l l} \\left\\{\\lambda_ {4} \\times | | \\mathcal {E} _ {C / T} (\\boldsymbol {I}) | | _ {2} ^ {2} + \\right. & \\text {i f} \\boldsymbol {I} \\in \\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right) \\\\ \\left. \\lambda_ {5} \\times \\operatorname {C S} \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right) \\right\\} & \\\\ \\left\\{\\lambda_ {6} \\times (1 - \\operatorname {C S} \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right)) + \\right. & \\text {i f} \\boldsymbol {I} \\in G \\left(\\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right)\\right) \\\\ \\left. \\lambda_ {7} \\times (1 - S S \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right)) \\right\\}. & \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 542, + 893, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we have a classifier to make a binary decision of real vs. fake using the fakeness maps. This classifier is included in the framework to aid the detection module for binary detection of the input images, which will be discussed in Sec. 3.2.3. Another reason to have the classifier is to make the fakeness maps from encrypted and fake images to be distinguishable. We find that this design allows our training to converge much faster.", + "bbox": [ + 496, + 638, + 893, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Detection Module", + "text_level": 1, + "bbox": [ + 500, + 787, + 679, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To leverage the added template for manipulation detection, we perform template recovery using encoder $\\mathcal{E}_E$ . We follow the procedure in [1] to recover the added template from the encrypted images by maximizing the cosine similarity between $S$ and $S_R$ . However, for manipulated images, we minimize the cosine similarity between the recovered tem", + "bbox": [ + 496, + 816, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12346", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/567c396b3f3cbecd22bc5b75a218786205e9d437871399760f9c8abde3756f81.jpg", + "image_caption": [ + "(a)", + "(b)", + "(c)", + "(d)", + "(e)", + "(f)", + "Figure 3. Visualization of fakeness maps for faces and generic images showing generalization across unseen attribute modifications and GMs: (a) real image, (b) encrypted image, (c) manipulated image, (d) $M_{GT}$ , (e) predicted fakeness map for encrypted images, and (f) predicted fakeness map for manipulated images. The first column shows the manipulation of (seen GM, seen attribute modification) i.e. (STGAN, bald). Following two columns show the manipulation of (seen GM, unseen attribute modification) i.e. (STGAN, [bangs, pale skin]. The fourth and fifth columns show manipulation of unseen GM, GauGAN for non-face images. The last column shows manipulation by unseen GM, DRIT. We see that the fakeness map of manipulated images is more bright and similar to $M_{GT}$ , while the real fakeness map is more close to zero. We use the cmap as \"pink\" to better visualize the fakeness map. All face images come from SiWM-v2 data [18]." + ], + "image_footnote": [], + "bbox": [ + 107, + 92, + 467, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "plate $(S_R)$ and all the templates in the template set $S$ .", + "bbox": [ + 76, + 614, + 433, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {R} = \\left\\{ \\begin{array}{l l} \\lambda_ {8} \\times (1 - \\operatorname {C S} (\\boldsymbol {S}, \\boldsymbol {S} _ {R})) & \\text {i f} x \\in \\mathcal {T} (\\boldsymbol {I} ^ {R}) \\\\ \\lambda_ {9} \\times \\left(\\sum_ {i = 1} ^ {n} (\\operatorname {C S} (\\boldsymbol {S} _ {i}, \\boldsymbol {S} _ {R}))\\right) & \\text {i f} x \\in G (\\mathcal {T} (\\boldsymbol {I} ^ {R})). \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 636, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, we leverage our estimated fakeness map to help manipulation detection. As discussed in the previous section, we apply a classifier $\\mathcal{C}$ to perform binary classification of the predicted fakeness map for the encrypted and fake images. The logits of the classifier are further combined with the cosine similarity of the recovered template. The averaged logits are back-propagated using the binary cross-entropy constraint. This not only improves the performance of manipulation detection but also helps manipulation localization. Therefore, we apply the binary cross entropy loss on the averaged logits as follows:", + "bbox": [ + 75, + 676, + 472, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} J _ {C} = \\lambda_ {1 0} \\times - \\sum_ {j} \\left\\{y _ {j}. \\log \\left[ \\frac {\\mathcal {C} \\left(\\boldsymbol {X} _ {j}\\right) + C S \\left(\\boldsymbol {S} _ {R} , \\boldsymbol {S}\\right)}{2} \\right] - \\right. \\tag {10} \\\\ (1 - y _ {j}). \\log \\left[ 1 - \\frac {\\mathcal {C} (\\boldsymbol {X} _ {j}) + C S (\\boldsymbol {S} _ {R} , \\boldsymbol {S})}{2} \\right] \\Bigg \\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 835, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/814a3b097cbeaa86e41d07d701e1874fb1b699a4dc141fd5233a28f6ffb6c5a8.jpg", + "table_caption": [ + "Table 2. Manipulation localization comparison with prior works." + ], + "table_footnote": [], + "table_body": "
MethodLocalizationDetection
CS ↑PSNR ↑SSIM ↑Accuracy ↑EER ↓AUC ↑
[11]0.62306.2140.21780.99750.00500.9975
[21]0.883122.8900.78760.99450.00770.9998
MaLP0.939423.0200.73120.99910.00721.0
", + "bbox": [ + 504, + 104, + 890, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{j}$ is the class label, $\\pmb{S}$ and $S_{R}$ are the added and recovered template respectively.", + "bbox": [ + 498, + 195, + 890, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our framework is trained in an end-to-end manner with the overall loss function as follows:", + "bbox": [ + 498, + 224, + 890, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ = J _ {T} + J _ {R} + J _ {C} + J _ {L}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 261, + 890, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. MaLP as A Discriminator", + "text_level": 1, + "bbox": [ + 500, + 284, + 735, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One application of MaLP is to leverage our proposed localization module as a discriminator for improving the quality of the manipulated images. MaLP performs binary classification by estimating a fakeness map, which can be used as an objective. This results in output images being resilient to manipulation localization, thereby lowering the performance of our framework.", + "bbox": [ + 496, + 306, + 890, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use MaLP as a plug-and-play discriminator to improve image generation quality through fine-tuning pretrained GMs. The generation quality and manipulation localization will compete head-to-head, resulting in a better quality of the manipulated images. We define the fine-tuning objective for the GM as follows:", + "bbox": [ + 496, + 411, + 890, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta_ {G}} \\max _ {\\theta_ {M a L P}, \\boldsymbol {S} _ {i}} \\left\\{\\sum_ {j} \\left(\\mathbb {E} \\left[ l o g \\left(\\mathcal {E} _ {M a L P} \\left(\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}\\right)\\right); \\theta_ {M a L P}\\right) \\right] + \\right. \\right. \\\\ \\left. \\mathbb {E} \\left[ 1 - \\log \\left(\\mathcal {E} _ {M a L P} \\left(G \\left(\\mathcal {T} \\left(I _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right); \\theta_ {G}\\right); \\theta_ {M a L P}\\right)\\right) \\right]\\right) \\Bigg \\}. \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 503, + 890, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_{MaLP}$ is our framework with $\\theta_{MaLP}$ parameters.", + "bbox": [ + 500, + 590, + 877, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 619, + 633, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 643, + 689, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings Following the settings in [21], we use STGAN [28] to manipulate images from CelebA [31] dataset and train on bald facial attribute modification. In order to evaluate the generalization of image manipulation localization, we construct a new benchmark that consists of 200 real images of 22 different GMs on various data domains. The real images are chosen from the dataset on which the GM is trained on. The list of GMs, datasets and implementation details are provided in the supplementary.", + "bbox": [ + 496, + 670, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics We use cosine similarity (CS), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) as adopted by [21] to evaluate manipulation localization since the GT is a continuous map. For binary detection, we use the area under the curve (AUC), equal error rate (EER), and accuracy score [21].", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12347", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/eeb2ff54ce3edacd16014f59d9e93c31efb4cbab2cbdf95cae18677a4d3282d4.jpg", + "table_caption": [ + "Table 3. Comparison of localization performance across unseen GMs and attribute modifications. We train on STGAN bald/smile attribute modification and test on AttGAN/StyleGAN." + ], + "table_footnote": [], + "table_body": "
MethodCosine similarity ↑(AttGAN)Cosine similarity ↑ (StyleGAN)
BaldBlack HairEyeglassesSmileAgeGender
[21]0.81410.69320.69500.61760.31410.6470
MaLP0.82010.79400.85570.81590.82550.8016
", + "bbox": [ + 81, + 130, + 468, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 76, + 199, + 321, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our results with [21] and [11] for manipulation localization. The results are shown in Tab. 2. MaLP has higher cosine similarity and similar PSNR for localization compared to [21]. However, we observe a dip in SSIM. This might be because of the degradation caused by adding our template to the real images and then performing the manipulation. The learned template helps localize the manipulated regions better, as demonstrated by cosine similarity, but the degradation affects SSIM and PSNR. We also compare the performance of real vs. fake binary detection. As expected, our proposed proactive approach outperforms the passive methods with a perfect AUC and near-perfect accuracy. We also show visual examples of fakeness maps for images modified by unseen GMs in Fig. 3. MaLP is able to estimate the fakeness map for unseen modifications and GMs across face/generic image datasets.", + "bbox": [ + 75, + 223, + 470, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Generalization", + "text_level": 1, + "bbox": [ + 76, + 476, + 227, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Across Attribute Modifications Following the settings in [21], we evaluate the performance of MaLP across unseen attribute modifications. Specifically, we train MaLP using STGAN with the bald/smile attribute modification and test it on unseen attribute modifications with unseen GMs: AttGAN/StyleGAN. As shown in Tab. 3, MaLP is more generalizable to all unseen attribute modifications. Furthermore, AttGAN shares the high-level architecture with STGAN but not with StyleGAN. We observe a significant increase in localization performance for StyleGAN compared to AttGAN. This shows that, unlike our MaLP, passive works perform much worse if the test GM doesn't share any similarity with the training GM.", + "bbox": [ + 75, + 503, + 468, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Across GMs Although [21] tries to show generalization across unseen GMs; it is limited by the GMs within the same domain of the dataset used in training. We propose a benchmark to evaluate the generalization performance for future manipulation localization works that consists of 22 different GMs in various domains. We select GMs that are publicly released and can perform partial manipulation.", + "bbox": [ + 75, + 703, + 468, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As no open-source code base is available for [21], we train a passive approach using a ResNet50 [19] network to estimate the fakeness map as the baseline for comparison. Further, we compare our approach with [4, 11]. Although [4, 11] estimate a fakeness map, it has at least $5 \\times$ lower resolution compared to input images due to", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae2e28d09a67750ebd9891c7e909bd33228f3a8ceda7eef3d3d3711fb19c0658.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 887, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c13de7d103ce68c48e9ac785738e75663e8b59909114dbcc440105ee2af30208.jpg", + "image_caption": [ + "0.58", + "0.62", + "0.61", + "0.59", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 534, + 178, + 887, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d462d6840d020c7f4ea6ebb61189e6f10db23bffb6cffceec47b00e6715b3af.jpg", + "image_caption": [ + "0.42", + "0.45", + "0.40", + "0.44", + "Figure 4. Visualization of (a) encrypted images, (b) manipulated images before fine-tuning, and (c) manipulated images after fine-tuning. The generation quality has improved after we fine-tune the GM using our framework as a discriminator. The artifacts in the images have been reduced, and the face skin color is less pale and more realistic. We also specify the cosine similarity of the predicted fakeness map and $M_{GT}$ . The GM is able to decrease the performance of our framework after fine-tuning. All face images come from SiWM-v2 data [18]." + ], + "image_footnote": [], + "bbox": [ + 534, + 267, + 887, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "their patch-based methodology. For a fair comparison, we rescale their predicted fakeness maps to the resolution of $M_{GT}$ . We compare the cosine similarity in Tab. 4. MaLP is able to outperform all the baselines for almost all GMs, which proves the effectiveness of the proactive scheme.", + "bbox": [ + 498, + 479, + 890, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate the performance of $\\mathcal{E}_C$ for high-resolution images. For encryption, we upsample the $128\\times 128$ template to the original resolution of images and evaluate $\\mathcal{E}_C$ on these higher resolution encrypted images. We observe similar performance of $\\mathcal{E}_C$ for higher resolution images in Tab. 4, proving the versatility of $\\mathcal{E}_C$ to image sizes.", + "bbox": [ + 498, + 556, + 890, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Improving Quality of GMs", + "text_level": 1, + "bbox": [ + 500, + 664, + 741, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We fine-tune the GM into fooling our framework to generate a fakeness map as a zero image. This process results in better-quality images. Initially, we train MaLP with the pretrained GM so that it can perform manipulation localization. Next, to fine-tune the GM, we adopt two strategies. First, we freeze MaLP and fine-tune the GM only. Second, we fine-tune both the GM and the MaLP but update the MaLP with a lower learning rate. The result for fine-tuning StarGAN is shown in Tab. 5. We observe that for both strategies, MaLP reduces the FID score of StarGAN. We also show some visual examples in Fig. 4. We see that the images are of better quality after fine-tuning, and many artifacts in the images manipulated by the pretrained model are removed.", + "bbox": [ + 496, + 689, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12348", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/25112bda1f9ecf6a0e0b876ce1b34db1e6bdfe20204d52a6a9ac81997e3e927e.jpg", + "table_caption": [ + "Table 4. Benchmark for manipulation localization across 22 different unseen GMs, showing cosine similarity between ground-truth and predicted fakeness maps. We compare our proactive vs. passive baselines [4, 11, 19] approach to highlight the generalization ability of our MaLP. We scale the images to $128^{2}$ for \"sc.\" and keep the resolution as is for \"no sc\"." + ], + "table_footnote": [], + "table_body": "
GMSEAN [62]StarGAN [8]CycleGAN [60]GauGAN [39]Con.Enc. [41]StarGAN2 [9]ALAE [42]BiGAN [61]AuGAN [60]GANim [43]DRGAN [50]ILVR [7]
Resolution256212822562256212822562256225623402128212822562
ResNet50 [19]0.86140.75130.67150.76150.86390.81960.67660.65140.66390.68710.80290.7018
[4]0.75140.71110.79810.80160.78940.70260.71560.72170.75160.76120.71150.7851
[11]0.79610.78870.80140.82560.85410.70340.75490.78050.72320.84570.72390.7854
MaLP (sc.)0.93760.87180.91280.92510.85460.88360.91920.91810.88940.96250.75120.8003
MaLP (no sc.)0.92580.87180.92450.91250.85460.87850.91410.92290.91490.96250.75120.8359
GMDRIT [26]Pix2Pix [22]CounGAN [38]DualGAN [58]ESRGAN [54]UNIT [29]MUNIT [20]ColGAN [36]GDWCT [6]RePaint [32]Average
Resolution256225621282256210242512 × 931256 × 512128212822562-
ResNet50 [19]0.74860.67190.72930.73650.87030.70830.66010.75960.83500.65120.7401
[4]0.78710.77690.81460.75690.81680.80640.67880.76100.86910.75160.7645
[11]0.81200.77810.85590.77210.82410.80860.70970.78740.88790.76960.7903
MaLP (sc.)0.88670.89150.93260.88720.83480.82140.75650.80960.93840.81020.8725
MaLP (no sc.)0.90840.87140.93260.84320.87430.83910.78600.80960.93840.82900.8773
", + "bbox": [ + 78, + 133, + 890, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0ce6b7fc05bfeeeb3af8f4462d4f1758712437f8b48e344a3a81c23e3f82d880.jpg", + "table_caption": [ + "Table 5. FID score comparison for the application of our approach as a discriminator for improving the generation quality of the GM" + ], + "table_footnote": [], + "table_body": "
StateFine-tuneStarGAN FID ↓
Before-60.49
AfterG51.91
G + MaLP52.07
", + "bbox": [ + 163, + 316, + 377, + 367 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f00ce0705f6657ef1d4d991066087b2a9c8008da761c349f384e38155af78e46.jpg", + "table_caption": [ + "Table 6. Comparison with prior binary detection works. [Keys: D.M.: Detection module, L.M.: Localization module]" + ], + "table_footnote": [], + "table_body": "
MethodTrain GMSet sizeTest GM Average precision (%)↑
CycleGANStarGANGauGAN
Nataraj et al. [35]CycleGAN-10088.2056.20
Zhang et al. [59]AutoGAN-10010061.00
Wang et al. [53]ProGAN-84.0010067.00
Asnani et al. [1]STGAN194.0010069.50
MaLP (D.M.)STGAN194.1010069.61
MaLP (D.M. + L.M.)STGAN194.3010072.16
", + "bbox": [ + 81, + 398, + 467, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Other Comparisons", + "text_level": 1, + "bbox": [ + 76, + 503, + 266, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Binary Detection We compare with prior proactive and passive approaches for binary manipulation detection [1, 35, 53, 59]. We adopt the evaluation protocol in [1] to test on images manipulated by CycleGAN, StarGAN, and GauGAN. We are able to perform similar to [1] as shown in Tab. 6. We have better average precision than passive schemes and generalize well to GMs unseen in training. We also conduct experiments to see whether localization can help binary detection to improve the performance, as mentioned in Sec. 3.2.3. The combined predictions' results are better than just using the detection module as shown in Tab. 6. This is intuitive as the localization module provides extra information, thereby increasing the performance.", + "bbox": [ + 75, + 530, + 468, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference Speed We compare the inference speed of our MaLP against prior work. [21] uses Deeplabv3-ResNet101 model from PyTorch [40]. In our generalization benchmark shown in Sec. 4.3, we use the ResNet50 model for training the passive baseline. The inference speed per image on an NVIDIA K80 GPU for Deeplabv3-ResNet101, ResNet50, and MaLP are 75.61, 52.66, and $29.26~\\mathrm{ms}$ , respectively. MaLP takes less than half the inference time compared to [21] due to our shallow CNN network.", + "bbox": [ + 75, + 729, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adversarial Attack Our framework can be considered as an adversarial attack on real images to aid manipula-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/469401c11a970ce163e9c6c134b7a1df1705851736fa05095285dd8758e607a8.jpg", + "table_caption": [ + "Table 7. Comparison with adversarial attack methods." + ], + "table_footnote": [], + "table_body": "
MethodSchemeCosine similarity↑
BaldBlack HairEyeglasses
Huang et al. [21]Passive0.81410.69320.6950
PGD [33]Proactive0.80510.75140.8358
FGSM [16]Proactive0.81110.78820.8512
CW [3]Proactive0.80140.83440.8405
MaLPProactive0.82010.79400.8557
", + "bbox": [ + 504, + 303, + 890, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion localization. Therefore, it is vital to contrast the performance between our approach and classic adversarial attacks. For this purpose, we perform experiments that make use of adversarial attacks, namely PGD [33], CW [3], and FGSM [16] to guide the learning of the added template. We evaluate on unseen GM AttGAN for unseen attribute modifications. We show the performance comparison in Tab. 7. MaLP has higher cosine similarity across some unseen facial attribute modifications compared to adversarial attacks. This can be explained as the adversarial attack methods being over-fitted to training parameters (data, target network etc.). Therefore, if the testing data is changed with unseen attribute modifications by GMs, the performance of adversarial attacks degrades. Further, these attacks are analogous to our MaLP as a proactive scheme which, in general, have better performance than passive works.", + "bbox": [ + 496, + 416, + 892, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Robustness Against Degradations It is necessary to test the robustness of our proposed approach against various types of real-world image editing degradations. We evaluate our method on degradations applied during testing as adopted by [21], which include JPEG compression, blurring, adding noise, and low resolution. The results are shown in Fig. 5. Our proposed MaLP is more robust to real-world degradations than passive schemes.", + "bbox": [ + 496, + 662, + 890, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablations", + "text_level": 1, + "bbox": [ + 500, + 797, + 609, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two-branch Architecture As described in Sec. 3.2.2, MaLP adopts a two-branch architecture to predict the fairness map using the local-level and global-level features, which are estimated by a shallow CNN and a transformer. We ablate by training each branch separately to show the", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12349", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/996afe55f94d70e619084c2b3567fe3017ddd9a25bf2498576d48dfa3e1218b9.jpg", + "image_caption": [ + "Figure 5. Comparison of our approach's robustness against common image editing degradations." + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 467, + 188 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/30e144141c2d7e555280f6154fa09099ba063c1b2ac337cb27859ec58307902f.jpg", + "table_caption": [ + "Table 8. Ablation of two-branch architecture. CNN is a shallow network with 10 layers. Training each branch separately has worse localization results than combining them." + ], + "table_footnote": [], + "table_body": "
Network trainedCosine similarity ↑Accuracy ↑
CNN only0.89610.9801
Transformer only0.88480.9856
CNN + ResNet500.86470.9512
CNN + Transformer0.93940.9981
", + "bbox": [ + 86, + 266, + 450, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "effectiveness of combining them. As shown in Tab. 8, if the individual network is trained separately, the performance is lower than the two-branch architecture. Next, to show the efficacy of the transformer, we use a ResNet50 network in place of the transformer to predict the fakeness map. We observe that the performance is even worse than using only the transformer. ResNet50 lacks the added advantage of self-attention in the transformer, which estimates the global-level features much better than a CNN network.", + "bbox": [ + 75, + 361, + 467, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Constraints MaLP leverages different constraints to estimate the fakeness map using an optimized template. We perform an ablation by removing each constraint separately, showing the importance of every constraint. Tab. 9 shows the cosine similarity for localization and accuracy for detection. Removing either the classifier or recovery constraint results in lower detection performance. This is expected as we leverage logits from both $\\mathcal{C}$ and $\\varepsilon_{E}$ , and removing the constraint for one network will hurt the logits of the other network. Furthermore, removing the template constraint results in a decrease in performance. Although the gap is small, the template is not properly optimized to have lower magnitude and high-frequency content.", + "bbox": [ + 75, + 501, + 467, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Removing the localization constraint and just applying a $L_{2}$ loss for supervising fakeness maps result in a significant performance drop for both localization and detection, showing the necessity of this constraint. Finally, we show the importance of a learnable template by not optimizing it during the training of MaLP. This hurts the performance a lot, similar to removing the localization constraint. Both these observations prove that our localization constraint and learnable template are important components of MaLP.", + "bbox": [ + 75, + 699, + 467, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Template Set Size We perform an ablation to vary the size of the template set $S$ . Having multiple templates will improve security if an attacker tries to reverse engineer the template from encrypted images. The results are shown in", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dce03e1c133c93006fe6f2fc36a4b1584f906d897f9eb84bb5cea6cdceb376ca.jpg", + "table_caption": [ + "Table 9. Ablation of constraints used in training our framework." + ], + "table_footnote": [], + "table_body": "
Constraint removedCosine similarity ↑Accuracy ↑
Classifier constraint JC0.93190.9814
Template constraint JT0.91430.9803
Localization constraint JL0.88140.9539
Recovery constraint JR0.92060.9780
Fixed template0.88870.9514
Nothing (MaLP)0.93940.9991
", + "bbox": [ + 503, + 103, + 885, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a883284c91c6560a0e7db0000cfbb0b571e604bc747af57c583e964b60f02ef4.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 215, + 694, + 296 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6af8e81bd60cd8193c3050d37bb2df239eb871a1cf7b6760ab9b67260b2d09fe.jpg", + "image_caption": [ + "(b)", + "Figure 6. Ablation study on hyperparameters used in our framework: set size and signal strength." + ], + "image_footnote": [], + "bbox": [ + 694, + 215, + 887, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 6 (a). The cosine similarity takes a dip when the set size is increased. We also observe the inter-template cosine similarity, which remains constant at a high value of around 0.74 for all templates. This is against the findings of [1]. Localization is a more challenging task than binary detection. Therefore, it is less likely to find different templates for our MaLP in the given feature space compared to [1].", + "bbox": [ + 496, + 349, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Signal Strength We vary the template strength hyperparameter $m$ to find its impact on the performance. As shown in Fig. 6 (b), the cosine similarity increases as we increase the strength of the added template. However, this comes with the lower visual quality of the encrypted images if the template strength is increased. The performance doesn't vary much after $m = 30\\%$ , which we use for MaLP.", + "bbox": [ + 496, + 459, + 890, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 582, + 617, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper focuses on manipulation localization using a proactive scheme (MaLP). We propose to improve the generalization of manipulation localization across unseen GM and facial attribute modifications. We add an optimal template onto the real images and estimate the fakeness map via a two-branch architecture using local and global-level features. MaLP outperforms prior works with much stronger generalization capabilities, as demonstrated by our proposed evaluation benchmark with 22 different GMs in various domains. We show an application of MaLP in finetuning GMs to improve generation quality.", + "bbox": [ + 496, + 609, + 890, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations First, the number of publicly available GMs is limited. More thorough testing on many different GMs might give more insights into the problem of generalizable manipulation localization. Second, we show that our MaLP can be used to fine-tune the GMs to improve image generation quality. However, it is based on the pretrained GM. Using our method to train a GM from scratch can be an interesting direction to explore in the future.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12350", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vishal Asnani, Xi Yin, Tal Hassner, Sijia Liu, and Xiaoming Liu. Proactive image manipulation detection. In CVPR, 2022. 1, 2, 3, 4, 7, 8", + "[2] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Reverse engineering of generative models: Inferring model hyperparameters from generated images. arXiv preprint arXiv:2106.07873, 2021. 1, 3", + "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In SSP, 2017. 7", + "[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020. 1, 2, 6, 7", + "[5] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, 2022. 1, 3", + "[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In CVPR, 2019. 7", + "[7] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 7", + "[8] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, 2018. 1, 7", + "[9] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 1, 7", + "[10] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, 2017. 2", + "[11] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In CVPR, 2020. 1, 2, 3, 5, 6, 7", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 4", + "[13] Bruce Draper. Reverse engineering of deceptions (red). https://www.darpa.mil/program/reverse-engineering-of-deceptions.1", + "[14] Candice R Gerstner and Hany Farid. Detecting real-time deep-fake videos using active illumination. In CVPR, 2022. 1, 3", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1", + "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 7", + "[17] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vec" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tor quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 1", + "[18] Xiao Guo, Yaojie Liu, Anil Jain, and Xiaoming Liu. Multidomain learning for updating face anti-spoofing models. In ECCV, 2022. 5, 6", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 7", + "[20] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In ECCV, 2018. 7", + "[21] Yihao Huang, Felix Juefei-Xu, Qing Guo, Yang Liu, and Geguang Pu. FakeLocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information Forensics and Security, 17:2657-2672, 2022. 1, 2, 3, 4, 5, 6, 7", + "[22] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 7", + "[23] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 1", + "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1", + "[25] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 1", + "[26] Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Kumar Singh, and Ming-Hsuan Yang. Diverse imaged-to-image translation via disentangled representations. In ECCV, 2018. 7", + "[27] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, 2020. 1, 2, 3", + "[28] Ming Liu, Yukang Ding, Min Xia, Xiao Liu, Errui Ding, Wangmeng Zuo, and Shilei Wen. STGAN: A unified selective transfer network for arbitrary image attribute editing. In CVPR, 2019. 1, 5", + "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In NeurIPS, 2017. 7", + "[30] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. PSCC-Net: Progressive spatio-channel correlation network for image manipulation detection and localization. In arXiv preprint arXiv:2103.10596, 2021. 1", + "[31] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5", + "[32] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 7", + "[33] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 7", + "[34] Safa C. Medin, Bernhard Egger, Anoop Cherian, Ye Wang, Joshua B. Tenenbaum, Xiaoming Liu, and Tim K. Marks." + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12351", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "MOST-GAN: 3D morphable StyleGAN for disentangled face image manipulation. In AAAI, 2022. 1", + "[35] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, BS Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, and Amit K Roy-Chowdhury. Detecting GAN generated fake images using co-occurrence matrices. Electronic Imaging, 2019:532-1, 2019. 3, 7", + "[36] Kamyar Nazeri, Eric Ng, and Mehran Ebrahimi. Image colorization using generative adversarial networks. In AMDO, 2018. 7", + "[37] Huy H Nguyen, Fuming Fang, Junichi Yamagishi, and Isao Echizen. Multi-task learning for detecting and segmenting manipulated facial images and videos. In BTAS, 2019. 1, 2, 3", + "[38] Ori Nizan and Ayellet Tal. Breaking the cycle - colleagues are all you need. In CVPR, 2020. 7", + "[39] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. GauGAN: semantic image synthesis with spatially adaptive normalization. In ACM, 2019. 1, 7", + "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 7", + "[41] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 7", + "[42] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In CVPR, 2020. 7", + "[43] Albert Pumarola, Antonio Agudo, Aleix M Martinez, Alberto Sanfeliu, and Francesc Moreno-Noguer. Ganimation: Anatomically-aware facial animation from a single image. In ECCV, 2018. 7", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1", + "[45] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: Learning to detect manipulated facial images. In CVPR, 2019. 1, 3", + "[46] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In ECCV, 2020. 1, 2", + "[47] Eran Segalis and Eran Galili. OGAN: Disrupting deep-fakes with an adversarial attack that survives training. arXiv preprint arXiv:2006.12247, 2020. 2", + "[48] Kaede Shiohara and Toshihiko Yamasaki. Detecting deep-fakes with self-blended images. In CVPR, 2022. 1, 3", + "[49] Kritaphat Songsri-in and Stefanos Zafeiriou. Complement face forensic detection and localization with facial landmarks. arXiv preprint arXiv:1910.05455, 2019. 1, 2, 3", + "[50] Luan Tran, Xi Yin, and Xiaoming Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, 2017. 1, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Run Wang, Felix Juefei-Xu, Meng Luo, Yang Liu, and Lina Wang. FakeTagger: Robust safeguards against deepfake dissemination via provenance tracking. In ACMM, 2021. 1, 2", + "[52] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Sketch your own gan. In ICCV, 2021. 1", + "[53] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. CNN-generated images are surprisingly easy to spot... for now. In CVPR, 2020. 3, 7", + "[54] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In CVPR, 2021. 7", + "[55] Xi Wu, Zhen Xie, YuTao Gao, and Yu Xiao. SSTNET: Detecting manipulated faces through spatial, steganalysis and temporal features. In ICASSP, 2020. 1", + "[56] Ying Xu, Kiran Raja, and Marius Pedersen. Supervised contrastive learning for generalizable and explainable deepfakes detection. In WACV, 2022. 1, 3", + "[57] Chin-Yuan Yeh, Hsi-Wen Chen, Shang-Lun Tsai, and Sheng-De Wang. Disrupting image-translation-based deepfake algorithms with adversarial attacks. In WACVW, 2020. 1, 2", + "[58] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. DualGAN: Unsupervised dual learning for image-to-image translation. In CVPR, 2017. 7", + "[59] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in GAN fake images. In WIFS, 2019. 3, 7", + "[60] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 1, 7", + "[61] Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darryll, Alexei A Efros, Oliver Wang, and Eli Shechtman. Toward multimodal image-to-image translation. In NeurIPS, 2017. 7", + "[62] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. SEAN: Image synthesis with semantic region-adaptive normalization. In CVPR, 2020. 7" + ], + "bbox": [ + 501, + 92, + 890, + 616 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12352", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_model.json b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ab340e529a889c415d53663e8e58af055d2398dc --- /dev/null +++ b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_model.json @@ -0,0 +1,2673 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.179, + 0.131, + 0.793, + 0.152 + ], + "angle": 0, + "content": "MaLP: Manipulation Localization Using a Proactive Scheme" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.18, + 0.717, + 0.217 + ], + "angle": 0, + "content": "Vishal Asnani1, Xi Yin2, Tal Hassner2, Xiaoming Liu1 \n1*Michigan State University, 2Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.218, + 0.723, + 0.234 + ], + "angle": 0, + "content": "1{asnanivi, liuxm}@msu.edu, 2{yinxi, thassner}@meta.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.619 + ], + "angle": 0, + "content": "Advancements in the generation quality of various Generative Models (GMs) has made it necessary to not only perform binary manipulation detection but also localize the modified pixels in an image. However, prior works termed as passive for manipulation localization exhibit poor generalization performance over unseen GMs and attribute modifications. To combat this issue, we propose a proactive scheme for manipulation localization, termed MaLP. We encrypt the real images by adding a learned template. If the image is manipulated by any GM, this added protection from the template not only aids binary detection but also helps in identifying the pixels modified by the GM. The template is learned by leveraging local and global-level features estimated by a two-branch architecture. We show that MaLP performs better than prior passive works. We also show the generalizability of MaLP by testing on 22 different GMs, providing a benchmark for future research on manipulation localization. Finally, we show that MaLP can be used as a discriminator for improving the generation quality of GMs. Our models/codes are available at www.github.com/vishal3477/pro_LOC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.21, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.822 + ], + "angle": 0, + "content": "We witness numerous Generative Models (GMs) [8, 9, 15, 17, 23-25, 28, 34, 39, 44, 50, 52, 60] being proposed to generate realistic-looking images. These GMs can not only generate an entirely new image [23, 24], but also perform partial manipulation of an input image [9, 9, 28, 60]. The proliferation of these GMs has made it easier to manipulate personal media for malicious use. Prior methods to combat manipulated media focus on binary detection [1, 2, 5, 11, 14, 30, 45, 48, 55, 56], using mouth movement, model parsing, hand-crafted features, etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.47, + 0.855 + ], + "angle": 0, + "content": "Recent works go one step further than detection, i.e. manipulation localization, which is defined as follows: given" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.27, + 0.72, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.398, + 0.614, + 0.41 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.27, + 0.889, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.397, + 0.823, + 0.41 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.411, + 0.895, + 0.482 + ], + "angle": 0, + "content": "Figure 1. (a) High-level idea of MaLP. We encrypt the image by adding a learnable template, which helps to estimate the fakeness map. (b) The cosine similarity (CS) between ground-truth and predicted fakeness maps for 22 unseen GMs. The performance is better for almost all GMs when using our proactive approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.893, + 0.638 + ], + "angle": 0, + "content": "a partially manipulated image by a GM (e.g. STGAN [28] modifying hair colors of a face image), the goal is to identify which pixels are modified by estimating a fakeness map [21]. Identifying modified pixels helps to determine the severity of the fakeness in the image, and aid mediaforensics [11, 21]. Also, manipulation localization provides an understanding of the attacker's intent for modification which may further benefit identifying attack toolchains used [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.895, + 0.792 + ], + "angle": 0, + "content": "Recent methods for manipulation localization [27,37,49] focus on estimating the manipulation mask of face-swapped images. They localize modified facial attributes by leveraging attention mechanisms [11], patch-based classifier [4], and face-parsing [21]. The main drawback of these methods is that they do not generalize well to GMs unseen in training. That is when the test images and training images are modified by different GMs, which will likely happen given the vast number of existing GMs. Thus, our work aims for a localization method generalizable to unseen GMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "All aforementioned methods are based on a passive scheme as the method receives an image as is for estimation. Recently, proactive methods are gaining success for deepfake tasks such as detection [1], disruption [46, 57], and tagging [51]. These methods are considered proactive as they add different types of signals known as templates for encrypting the image before it is manipulated by a GM." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*All data sourcing, modeling codes, and experiments were developed at Michigan State University. Meta did not obtain the data/codes or conduct any experiments in this work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12343" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "This template can be one-hot encoding [51], adversarial perturbation [46], or a learnable noise [1], and is optimized to improve the performance of the defined tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.365 + ], + "angle": 0, + "content": "Motivated by [1], we propose a Proactive scheme for MAnipulation Localization, termed as MaLP, in order to improve generalization. Specifically, MaLP learns an optimized template which, when added to real images, would improve manipulation localization, should they get manipulated. This manipulation can be done by an unseen GM trained on either in-domain or out-of-domain datasets. Furthermore, face manipulation may involve modifying facial attributes unseen in training (e.g. train on hair color modification yet test on gender modification). MaLP incorporates three modules that focus on encryption, detection, and localization. The encryption module selects and adds the template from the template set to the real images. These encrypted images are further processed by localization and detection modules to perform the respective tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.533 + ], + "angle": 0, + "content": "Designing a proactive manipulation localization approach comes with several challenges. First, it is not straightforward to formulate constraints for learning the template unsupervisedly. Second, calculating a fakeness map at the same resolution as the input image is computationally expensive if the decision for each pixel has to be made. Prior works [4, 11] either down-sample the images or use a patch-wise approach, both of which result in inaccurate low-resolution fakeness maps. Lastly, the templates should be generalizable to localize modified regions from unseen GMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.747 + ], + "angle": 0, + "content": "We design a two-branch architecture consisting of a shallow CNN network and a transformer to optimize the template during training. While the former leverages local-level features due to its shallow depth, the latter focuses on global-level features to better capture the affinity of the far apart regions. The joint training of both networks enables the MaLP to learn a better template, having embedded the information of both levels. During inference, the CNN network alone is sufficient to estimate the fakeness map with a higher inference efficiency. Compared to prior passive works [11, 21], MaLP improves the generalization performance on unseen GMs. We also demonstrate that MaLP can be used as a discriminator for fine-tuning conventional GMs to improve the quality of GM-generated images." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.749, + 0.43, + 0.763 + ], + "angle": 0, + "content": "In summary, we make the following contributions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.765, + 0.469, + 0.81 + ], + "angle": 0, + "content": "- We are the first to propose a proactive scheme for image manipulation localization, applicable to both face and generic images." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Our novel two-branch architecture uses both local and global level features to learn a set of templates in an unsupervised manner. The framework is guided by constraints based on template recovery, fakeness maps classification, and high cosine similarity between predicted and ground-truth fakeness maps." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.173 + ], + "angle": 0, + "content": "Table 1. Comparison of our approach with prior works on manipulation localization and proactive schemes. We show the generalization ability of all works across different facial attribute modifications, unseen GMs trained on datasets with the same domain (in-domain) and different domains (out-domain). [Keys: Attr.: Attributes, Imp.: Improving, L.: Localization, D.: Detection]" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.173, + 0.89, + 0.32 + ], + "angle": 0, + "content": "
WorkSchemeTaskTemplateGeneralizationImp. GM
Attr.In-domainOut-domain
[51]ProactiveTagFix✓✓✘✘
[47]ProactiveDisruptLearn✓✘✘✘
[46]ProactiveDisruptLearn✓✓✘✘
[57]ProactiveDisruptLearn✓✘✘✘
[1]ProactiveD.Learn✘✓✓✘
[37]PassiveL. + D.-✘✘✘✘
[49]PassiveL. + D.-✘✘✘✘
[27]PassiveL. + D.-✘✓✘✘
[11]PassiveL. + D.-✓✓✘✘
[4]PassiveL. + D.-✘✓✘✘
[21]PassiveL. + D.-✓✓✘✘
MaLPProactiveL. + D.Learn✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.337, + 0.891, + 0.381 + ], + "angle": 0, + "content": "- MaLP can be used as a plug-and-play discriminator module to fine-tune the generative model to improve the quality of the generated images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.383, + 0.892, + 0.487 + ], + "angle": 0, + "content": "- Our method outperforms State-of-The-Art (SoTA) methods in manipulation localization and detection. Furthermore, our method generalizes well to GMs and modified attributes unseen in training. To facilitate the research of localization, we develop a benchmark for evaluating the generalization of manipulation localization, on images where the train and test GMs are different." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.337, + 0.892, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.504, + 0.642, + 0.52 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Manipulation Localization. Prior works tackle manipulation localization by adopting a passive scheme. Some of them focus on forgery attacks like removal, copymove, and splicing using multi-task learning [37]. Songsriin et al. [49] leverage facial landmarks [10] for manipulation localization. Li et al. [27] estimate the blended boundary for forged face-swap images. [11] uses an attention mechanism to leverage the relationship between pixels and [4] uses a patch-based classifier to estimate modified regions. Recently, Huang et al. [21] utilize gray-scale maps as ground truth for manipulation localization and leverage face parsing with an attention mechanism for prediction. The passive methods discussed above suffer from the generalization issue [4, 10, 11, 21, 37, 49] and estimate a low-resolution fakeness map [11] which is less accurate for the localization purpose. MaLP generalizes better to modified attributes and GMs unseen in training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Proactive Scheme. Recently, proactive schemes are developed for various tasks. Wang et al. [51] leverage the recovery of embedded one-hot encoding messages to perform deepfake tagging. A small perturbation is added onto the images by Segalis et al. [47] to disrupt the output of a GM. The same task is performed by Ruiz et al. [46] and Yeh et al. [57], both adding adversarial noise onto the in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12344" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.183 + ], + "angle": 0, + "content": "put images. Asnani et al. [1] propose a framework based on adding a learnable template to input images for generalized manipulation detection. Unlike prior works, which focus on binary detection, deepfake disruption, or tagging, our work emphasizes on manipulation localization. We show the comparison of our approach with prior works in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.471, + 0.352 + ], + "angle": 0, + "content": "Manipulation Detection. The advancement in manipulation detection keeps reaching new heights. Prior works propose to combat deepfakes by exploiting frequency domain patterns [53], up-sampling artifacts [59], model parsing [2], hand-crafted features [35], lip motions [45], and self-attention [11]. Recent methods use self-blended images [48], real-time deviations [14], and self-supervised learning with adversarial training [5]. Finally, methods based on contrastive learning [56] and proactive scheme [1] have explicitly focused on generalized manipulation detection across unknown GMs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.37, + 0.269, + 0.387 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.398, + 0.281, + 0.413 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.426, + 0.47, + 0.5 + ], + "angle": 0, + "content": "Passive Manipulation Localization Let \\( \\pmb{I}^{R} \\) be a set of real images that are manipulated by a GM \\( G \\) to output the set of manipulated images \\( G(\\pmb{I}^{R}) \\). Prior passive works perform manipulation localization by estimating the fakeness map \\( M_{pred} \\) with the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.51, + 0.469, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\left. \\min _ {\\theta_ {\\mathcal {E}}} \\left\\{\\sum_ {j} \\left(\\left| \\left| \\mathcal {E} \\left(G \\left(I _ {j} ^ {R}\\right); \\theta_ {\\mathcal {E}}\\right) - M _ {G T} \\right| \\right| _ {2}\\right) \\right\\}, \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.556, + 0.468, + 0.586 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) denotes the passive framework with parameters \\(\\theta_{\\mathcal{E}}\\) and \\(M_{GT}\\) is the ground-truth fakeness map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.469, + 0.684 + ], + "angle": 0, + "content": "To represent the fakeness map, some prior methods [11, 27, 49] choose a binary map by applying a threshold on the difference between the real and manipulated images. This is undesirable as the threshold selection is highly subjective and sensitive, leading to inaccurate fakeness maps. Therefore, we adopt the continuous gray-scale map for calculating the ground-truth fakeness maps [21], formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.694, + 0.469, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {G T} = \\operatorname {G r a y} \\left(\\left| \\boldsymbol {I} ^ {R} - G \\left(\\boldsymbol {I} ^ {R}\\right) \\right|\\right) / 2 5 5, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.722, + 0.404, + 0.737 + ], + "angle": 0, + "content": "where \\(Gray(.)\\) converts the image to gray-scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.469, + 0.873 + ], + "angle": 0, + "content": "Proactive Scheme Asnani et al. [1] define adding the template as a transformation \\(\\mathcal{T}\\) applied to images \\(I^R\\), resulting in the encrypted images \\(\\mathcal{T}(I^R)\\). The added template acts as a signature of the defender and is learned during the training, aiming to improve the performance of the task at hand, e.g. detection, disruption, and tagging. Motivated by [1] that uses multiple templates, we have a set of \\(n\\) orthogonal templates \\(\\mathcal{S} = \\{S_1, S_2, \\ldots, S_n\\}\\) where \\(S_i \\in \\mathbb{R}^{128 \\times 128}\\), for a real image \\(I_j^R \\in I^R\\), transformation \\(\\mathcal{T}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.884, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right) = \\boldsymbol {I} _ {j} ^ {R} + \\boldsymbol {S} _ {i}, \\text {w h e r e} i \\in \\{1, 2, \\dots , n \\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "The templates are optimized such that adding them to the real images wouldn't result in a noticeable visual difference, yet helps manipulation localization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.14, + 0.892, + 0.199 + ], + "angle": 0, + "content": "Proactive Manipulation Localization. Unlike the passive schemes [11, 21, 27, 37], we learn an optimal template set to help manipulation localization. For the encrypted images \\(\\mathcal{T}(\\boldsymbol{I}^R)\\), we formulate the estimation of the fakeness map as:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.213, + 0.892, + 0.247 + ], + "angle": 0, + "content": "\\[\n\\left. \\min _ {\\theta_ {\\mathcal {E} _ {P}}, \\boldsymbol {S} _ {i}} \\left\\{\\sum_ {j} \\left(\\left| \\left| \\mathcal {E} _ {P} \\left(G \\left(\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right)\\right); \\theta_ {\\mathcal {E} _ {P}}\\right) - \\boldsymbol {M} _ {G T} \\right| \\right| _ {2}\\right) \\right\\}. \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.26, + 0.89, + 0.275 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_P\\) is the proactive framework with parameters \\(\\theta_{\\mathcal{E}_P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.276, + 0.892, + 0.334 + ], + "angle": 0, + "content": "However, as the output of the GM has changed from images in set \\( G(\\mathbf{I}^R) \\) to images in set \\( G(\\mathcal{T}(\\mathbf{I}^R)) \\), in our proactive approach, the calculation of the ground-truth fakeness map shall be changed from Eq. 2 to the follows:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.346, + 0.892, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {G T} = \\operatorname {G r a y} \\left(\\left| \\boldsymbol {I} ^ {R} - G \\left(\\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right)\\right) \\right|\\right) / 2 5 5. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.392, + 0.741, + 0.408 + ], + "angle": 0, + "content": "3.2. Manipulation Localization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.419, + 0.892, + 0.54 + ], + "angle": 0, + "content": "MaLP consists of three modules: encryption, localization, and detection. The encryption module is used to encrypt the real images. The localization module estimates the fakeness map using a two-branch architecture. The detection module performs binary detection for the encrypted and manipulated images by recovering the template and using the classifier in the localization module. All three modules, as detailed next, are trained in an end-to-end manner." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.691, + 0.582 + ], + "angle": 0, + "content": "3.2.1 Encryption Module" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.594, + 0.892, + 0.665 + ], + "angle": 0, + "content": "Following the procedure in [1], we add a randomly selected learnable template from the template set to a real image. We control the strength of the added template using a hyperparameter \\( m \\), which prevents the degradation of the image quality. The encryption process is summarised below:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.677, + 0.891, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}\\right) = \\boldsymbol {I} _ {j} ^ {R} + m \\times \\boldsymbol {S} _ {i} \\text {w h e r e} i = \\operatorname {R a n d} (1, 2, \\dots , n). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.708, + 0.845, + 0.722 + ], + "angle": 0, + "content": "We select the value of \\( m \\) as \\( 30\\% \\) for our framework." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.723, + 0.892, + 0.766 + ], + "angle": 0, + "content": "We optimize the template set by focusing on properties like low magnitude, orthogonality, and high-frequency content [1]. The properties are applied as constraints as follows." + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.78, + 0.894, + 0.835 + ], + "angle": 0, + "content": "\\[\nJ _ {T} = \\lambda_ {1} \\times \\sum_ {i = 1} ^ {n} \\left\\| \\boldsymbol {S} _ {i} \\right\\| _ {2} + \\lambda_ {2} \\times \\sum_ {\\substack {i, j = 1 \\\\ i \\neq j}} ^ {n} \\operatorname {CS} \\left(\\boldsymbol {S} _ {i}, \\boldsymbol {S} _ {j}\\right) + \\lambda_ {3} \\times \\left\\| \\mathcal {L} (\\mathfrak {F} (\\boldsymbol {S})) \\right\\| _ {2}, \\tag{7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.836, + 0.892, + 0.897 + ], + "angle": 0, + "content": "where CS is the cosine similarity, \\(\\mathcal{L}\\) is the low-pass filter, \\(\\mathfrak{F}\\) is the fourier transform, \\(\\lambda_1,\\lambda_2,\\lambda_3\\) are weights for losses of low magnitude, orthogonality and high-frequency content, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12345" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.882, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.347, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 2. The overview of MaLP. It includes three modules: encryption, localization, and detection. We randomly select a template from the template set and add it to the real image as encryption. The GM is used in inference mode to manipulate the encrypted image. The detection module recovers the added template for binary detection. The localization module uses a two-branch architecture to estimate the fakeness map. Lastly, we apply the classifier to the fakeness map to better distinguish them from each other. Best viewed in color." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.419, + 0.276, + 0.433 + ], + "angle": 0, + "content": "3.2.2 Localization Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.538 + ], + "angle": 0, + "content": "To design the localization module, we consider two desired properties: a larger receptive field for fakeness map estimation and high inference efficiency. A network with a large receptive field will consider far-apart regions in learning features for localization. Yet, large receptive fields normally come from deeper networks, implying slower inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.736 + ], + "angle": 0, + "content": "In light of these properties, we design a two-branch architecture consisting of a shallow CNN network \\(\\mathcal{E}_C\\) and a ViT transformer [12] \\(\\mathcal{E}_T\\) (see Fig. 2). The intuition is to have one shallow branch to capture local features, and one deeper branch to capture global features. While training with both branches helps to learn better templates, in inference we only use the shallow branch for a higher efficiency. Specifically, the shallow CNN network has 10 layers which is efficient in inference but can only capture the local features due to small receptive fields. To capture global information, we adopt the ViT transformer. With the self-attention between the image patches, the transformer can estimate the fakeness map considering the far-apart regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.469, + 0.857 + ], + "angle": 0, + "content": "Both the CNN and transformer are trained jointly to estimate a better template set, resembling the concept of the ensemble of networks. We empirically show that training both networks simultaneously results in higher performance than training either network separately. As the shallow CNN network is much faster in inference than the transformer, we use the transformer only in training to optimize the templates and switch off the transformer branch in inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.858, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To estimate the fakeness map, we leverage the supervision of the ground-truth fakeness map in Eq. 5. For fake images, we maximize the cosine similarity \\((CS)\\) and struc" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.895, + 0.532 + ], + "angle": 0, + "content": "tural similarity index measure \\((SS)\\) between the predicted and ground-truth fakeness map. However, the fakeness map should be a zero image for encrypted images. Therefore, we apply an \\(L_{2}\\) loss [21] to minimize the predicted map to zero for encrypted images. To maximize the difference between the two fakeness maps, we further minimize the cosine similarity between the predicted map from encrypted images and \\(M_{GT}\\). The localization loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.543, + 0.895, + 0.639 + ], + "angle": 0, + "content": "\\[\nJ _ {L} = \\left\\{ \\begin{array}{l l} \\left\\{\\lambda_ {4} \\times | | \\mathcal {E} _ {C / T} (\\boldsymbol {I}) | | _ {2} ^ {2} + \\right. & \\text {i f} \\boldsymbol {I} \\in \\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right) \\\\ \\left. \\lambda_ {5} \\times \\operatorname {C S} \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right) \\right\\} & \\\\ \\left\\{\\lambda_ {6} \\times (1 - \\operatorname {C S} \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right)) + \\right. & \\text {i f} \\boldsymbol {I} \\in G \\left(\\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right)\\right) \\\\ \\left. \\lambda_ {7} \\times (1 - S S \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right)) \\right\\}. & \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Finally, we have a classifier to make a binary decision of real vs. fake using the fakeness maps. This classifier is included in the framework to aid the detection module for binary detection of the input images, which will be discussed in Sec. 3.2.3. Another reason to have the classifier is to make the fakeness maps from encrypted and fake images to be distinguishable. We find that this design allows our training to converge much faster." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.789, + 0.681, + 0.804 + ], + "angle": 0, + "content": "3.2.3 Detection Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To leverage the added template for manipulation detection, we perform template recovery using encoder \\(\\mathcal{E}_E\\). We follow the procedure in [1] to recover the added template from the encrypted images by maximizing the cosine similarity between \\(S\\) and \\(S_R\\). However, for manipulated images, we minimize the cosine similarity between the recovered tem" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12346" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.11, + 0.102, + 0.121 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.156, + 0.102, + 0.167 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.201, + 0.102, + 0.213 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.247, + 0.102, + 0.259 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.294, + 0.102, + 0.305 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.343, + 0.102, + 0.355 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.473, + 0.6 + ], + "angle": 0, + "content": "Figure 3. Visualization of fakeness maps for faces and generic images showing generalization across unseen attribute modifications and GMs: (a) real image, (b) encrypted image, (c) manipulated image, (d) \\( M_{GT} \\), (e) predicted fakeness map for encrypted images, and (f) predicted fakeness map for manipulated images. The first column shows the manipulation of (seen GM, seen attribute modification) i.e. (STGAN, bald). Following two columns show the manipulation of (seen GM, unseen attribute modification) i.e. (STGAN, [bangs, pale skin]. The fourth and fifth columns show manipulation of unseen GM, GauGAN for non-face images. The last column shows manipulation by unseen GM, DRIT. We see that the fakeness map of manipulated images is more bright and similar to \\( M_{GT} \\), while the real fakeness map is more close to zero. We use the cmap as \"pink\" to better visualize the fakeness map. All face images come from SiWM-v2 data [18]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.615, + 0.434, + 0.631 + ], + "angle": 0, + "content": "plate \\((S_R)\\) and all the templates in the template set \\(S\\)." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.637, + 0.47, + 0.673 + ], + "angle": 0, + "content": "\\[\nJ _ {R} = \\left\\{ \\begin{array}{l l} \\lambda_ {8} \\times (1 - \\operatorname {C S} (\\boldsymbol {S}, \\boldsymbol {S} _ {R})) & \\text {i f} x \\in \\mathcal {T} (\\boldsymbol {I} ^ {R}) \\\\ \\lambda_ {9} \\times \\left(\\sum_ {i = 1} ^ {n} (\\operatorname {C S} (\\boldsymbol {S} _ {i}, \\boldsymbol {S} _ {R}))\\right) & \\text {i f} x \\in G (\\mathcal {T} (\\boldsymbol {I} ^ {R})). \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.473, + 0.832 + ], + "angle": 0, + "content": "Further, we leverage our estimated fakeness map to help manipulation detection. As discussed in the previous section, we apply a classifier \\(\\mathcal{C}\\) to perform binary classification of the predicted fakeness map for the encrypted and fake images. The logits of the classifier are further combined with the cosine similarity of the recovered template. The averaged logits are back-propagated using the binary cross-entropy constraint. This not only improves the performance of manipulation detection but also helps manipulation localization. Therefore, we apply the binary cross entropy loss on the averaged logits as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.837, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} J _ {C} = \\lambda_ {1 0} \\times - \\sum_ {j} \\left\\{y _ {j}. \\log \\left[ \\frac {\\mathcal {C} \\left(\\boldsymbol {X} _ {j}\\right) + C S \\left(\\boldsymbol {S} _ {R} , \\boldsymbol {S}\\right)}{2} \\right] - \\right. \\tag {10} \\\\ (1 - y _ {j}). \\log \\left[ 1 - \\frac {\\mathcal {C} (\\boldsymbol {X} _ {j}) + C S (\\boldsymbol {S} _ {R} , \\boldsymbol {S})}{2} \\right] \\Bigg \\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.09, + 0.888, + 0.104 + ], + "angle": 0, + "content": "Table 2. Manipulation localization comparison with prior works." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.105, + 0.892, + 0.167 + ], + "angle": 0, + "content": "
MethodLocalizationDetection
CS ↑PSNR ↑SSIM ↑Accuracy ↑EER ↓AUC ↑
[11]0.62306.2140.21780.99750.00500.9975
[21]0.883122.8900.78760.99450.00770.9998
MaLP0.939423.0200.73120.99910.00721.0
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.196, + 0.892, + 0.226 + ], + "angle": 0, + "content": "where \\(y_{j}\\) is the class label, \\(\\pmb{S}\\) and \\(S_{R}\\) are the added and recovered template respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.226, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Our framework is trained in an end-to-end manner with the overall loss function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.262, + 0.892, + 0.277 + ], + "angle": 0, + "content": "\\[\nJ = J _ {T} + J _ {R} + J _ {C} + J _ {L}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.285, + 0.736, + 0.299 + ], + "angle": 0, + "content": "3.3. MaLP as A Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.412 + ], + "angle": 0, + "content": "One application of MaLP is to leverage our proposed localization module as a discriminator for improving the quality of the manipulated images. MaLP performs binary classification by estimating a fakeness map, which can be used as an objective. This results in output images being resilient to manipulation localization, thereby lowering the performance of our framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.497 + ], + "angle": 0, + "content": "We use MaLP as a plug-and-play discriminator to improve image generation quality through fine-tuning pretrained GMs. The generation quality and manipulation localization will compete head-to-head, resulting in a better quality of the manipulated images. We define the fine-tuning objective for the GM as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.504, + 0.892, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta_ {G}} \\max _ {\\theta_ {M a L P}, \\boldsymbol {S} _ {i}} \\left\\{\\sum_ {j} \\left(\\mathbb {E} \\left[ l o g \\left(\\mathcal {E} _ {M a L P} \\left(\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}\\right)\\right); \\theta_ {M a L P}\\right) \\right] + \\right. \\right. \\\\ \\left. \\mathbb {E} \\left[ 1 - \\log \\left(\\mathcal {E} _ {M a L P} \\left(G \\left(\\mathcal {T} \\left(I _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right); \\theta_ {G}\\right); \\theta_ {M a L P}\\right)\\right) \\right]\\right) \\Bigg \\}. \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.592, + 0.879, + 0.608 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_{MaLP}\\) is our framework with \\(\\theta_{MaLP}\\) parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.634, + 0.636 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.644, + 0.691, + 0.661 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Settings Following the settings in [21], we use STGAN [28] to manipulate images from CelebA [31] dataset and train on bald facial attribute modification. In order to evaluate the generalization of image manipulation localization, we construct a new benchmark that consists of 200 real images of 22 different GMs on various data domains. The real images are chosen from the dataset on which the GM is trained on. The list of GMs, datasets and implementation details are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics We use cosine similarity (CS), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) as adopted by [21] to evaluate manipulation localization since the GT is a continuous map. For binary detection, we use the area under the curve (AUC), equal error rate (EER), and accuracy score [21]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12347" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.131 + ], + "angle": 0, + "content": "Table 3. Comparison of localization performance across unseen GMs and attribute modifications. We train on STGAN bald/smile attribute modification and test on AttGAN/StyleGAN." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.131, + 0.47, + 0.18 + ], + "angle": 0, + "content": "
MethodCosine similarity ↑(AttGAN)Cosine similarity ↑ (StyleGAN)
BaldBlack HairEyeglassesSmileAgeGender
[21]0.81410.69320.69500.61760.31410.6470
MaLP0.82010.79400.85570.81590.82550.8016
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.2, + 0.323, + 0.216 + ], + "angle": 0, + "content": "4.2. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.471, + 0.466 + ], + "angle": 0, + "content": "We compare our results with [21] and [11] for manipulation localization. The results are shown in Tab. 2. MaLP has higher cosine similarity and similar PSNR for localization compared to [21]. However, we observe a dip in SSIM. This might be because of the degradation caused by adding our template to the real images and then performing the manipulation. The learned template helps localize the manipulated regions better, as demonstrated by cosine similarity, but the degradation affects SSIM and PSNR. We also compare the performance of real vs. fake binary detection. As expected, our proposed proactive approach outperforms the passive methods with a perfect AUC and near-perfect accuracy. We also show visual examples of fakeness maps for images modified by unseen GMs in Fig. 3. MaLP is able to estimate the fakeness map for unseen modifications and GMs across face/generic image datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.477, + 0.228, + 0.492 + ], + "angle": 0, + "content": "4.3. Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.7 + ], + "angle": 0, + "content": "Across Attribute Modifications Following the settings in [21], we evaluate the performance of MaLP across unseen attribute modifications. Specifically, we train MaLP using STGAN with the bald/smile attribute modification and test it on unseen attribute modifications with unseen GMs: AttGAN/StyleGAN. As shown in Tab. 3, MaLP is more generalizable to all unseen attribute modifications. Furthermore, AttGAN shares the high-level architecture with STGAN but not with StyleGAN. We observe a significant increase in localization performance for StyleGAN compared to AttGAN. This shows that, unlike our MaLP, passive works perform much worse if the test GM doesn't share any similarity with the training GM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Across GMs Although [21] tries to show generalization across unseen GMs; it is limited by the GMs within the same domain of the dataset used in training. We propose a benchmark to evaluate the generalization performance for future manipulation localization works that consists of 22 different GMs in various domains. We select GMs that are publicly released and can perform partial manipulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As no open-source code base is available for [21], we train a passive approach using a ResNet50 [19] network to estimate the fakeness map as the baseline for comparison. Further, we compare our approach with [4, 11]. Although [4, 11] estimate a fakeness map, it has at least \\(5 \\times\\) lower resolution compared to input images due to" + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.117, + 0.526, + 0.131 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.091, + 0.888, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.168, + 0.594, + 0.178 + ], + "angle": 0, + "content": "0.58" + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.168, + 0.681, + 0.178 + ], + "angle": 0, + "content": "0.62" + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.168, + 0.769, + 0.178 + ], + "angle": 0, + "content": "0.61" + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.168, + 0.857, + 0.178 + ], + "angle": 0, + "content": "0.59" + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.206, + 0.526, + 0.22 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.179, + 0.888, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.256, + 0.593, + 0.267 + ], + "angle": 0, + "content": "0.42" + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.257, + 0.681, + 0.267 + ], + "angle": 0, + "content": "0.45" + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.256, + 0.77, + 0.267 + ], + "angle": 0, + "content": "0.40" + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.257, + 0.857, + 0.267 + ], + "angle": 0, + "content": "0.44" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.268, + 0.888, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.337, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Figure 4. Visualization of (a) encrypted images, (b) manipulated images before fine-tuning, and (c) manipulated images after fine-tuning. The generation quality has improved after we fine-tune the GM using our framework as a discriminator. The artifacts in the images have been reduced, and the face skin color is less pale and more realistic. We also specify the cosine similarity of the predicted fakeness map and \\( M_{GT} \\). The GM is able to decrease the performance of our framework after fine-tuning. All face images come from SiWM-v2 data [18]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.481, + 0.892, + 0.556 + ], + "angle": 0, + "content": "their patch-based methodology. For a fair comparison, we rescale their predicted fakeness maps to the resolution of \\( M_{GT} \\). We compare the cosine similarity in Tab. 4. MaLP is able to outperform all the baselines for almost all GMs, which proves the effectiveness of the proactive scheme." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.558, + 0.892, + 0.649 + ], + "angle": 0, + "content": "We also evaluate the performance of \\(\\mathcal{E}_C\\) for high-resolution images. For encryption, we upsample the \\(128\\times 128\\) template to the original resolution of images and evaluate \\(\\mathcal{E}_C\\) on these higher resolution encrypted images. We observe similar performance of \\(\\mathcal{E}_C\\) for higher resolution images in Tab. 4, proving the versatility of \\(\\mathcal{E}_C\\) to image sizes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.742, + 0.681 + ], + "angle": 0, + "content": "4.4. Improving Quality of GMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We fine-tune the GM into fooling our framework to generate a fakeness map as a zero image. This process results in better-quality images. Initially, we train MaLP with the pretrained GM so that it can perform manipulation localization. Next, to fine-tune the GM, we adopt two strategies. First, we freeze MaLP and fine-tune the GM only. Second, we fine-tune both the GM and the MaLP but update the MaLP with a lower learning rate. The result for fine-tuning StarGAN is shown in Tab. 5. We observe that for both strategies, MaLP reduces the FID score of StarGAN. We also show some visual examples in Fig. 4. We see that the images are of better quality after fine-tuning, and many artifacts in the images manipulated by the pretrained model are removed." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12348" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 4. Benchmark for manipulation localization across 22 different unseen GMs, showing cosine similarity between ground-truth and predicted fakeness maps. We compare our proactive vs. passive baselines [4, 11, 19] approach to highlight the generalization ability of our MaLP. We scale the images to \\(128^{2}\\) for \"sc.\" and keep the resolution as is for \"no sc\"." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.135, + 0.891, + 0.273 + ], + "angle": 0, + "content": "
GMSEAN [62]StarGAN [8]CycleGAN [60]GauGAN [39]Con.Enc. [41]StarGAN2 [9]ALAE [42]BiGAN [61]AuGAN [60]GANim [43]DRGAN [50]ILVR [7]
Resolution256212822562256212822562256225623402128212822562
ResNet50 [19]0.86140.75130.67150.76150.86390.81960.67660.65140.66390.68710.80290.7018
[4]0.75140.71110.79810.80160.78940.70260.71560.72170.75160.76120.71150.7851
[11]0.79610.78870.80140.82560.85410.70340.75490.78050.72320.84570.72390.7854
MaLP (sc.)0.93760.87180.91280.92510.85460.88360.91920.91810.88940.96250.75120.8003
MaLP (no sc.)0.92580.87180.92450.91250.85460.87850.91410.92290.91490.96250.75120.8359
GMDRIT [26]Pix2Pix [22]CounGAN [38]DualGAN [58]ESRGAN [54]UNIT [29]MUNIT [20]ColGAN [36]GDWCT [6]RePaint [32]Average
Resolution256225621282256210242512 × 931256 × 512128212822562-
ResNet50 [19]0.74860.67190.72930.73650.87030.70830.66010.75960.83500.65120.7401
[4]0.78710.77690.81460.75690.81680.80640.67880.76100.86910.75160.7645
[11]0.81200.77810.85590.77210.82410.80860.70970.78740.88790.76960.7903
MaLP (sc.)0.88670.89150.93260.88720.83480.82140.75650.80960.93840.81020.8725
MaLP (no sc.)0.90840.87140.93260.84320.87430.83910.78600.80960.93840.82900.8773
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.29, + 0.47, + 0.317 + ], + "angle": 0, + "content": "Table 5. FID score comparison for the application of our approach as a discriminator for improving the generation quality of the GM" + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.318, + 0.378, + 0.368 + ], + "angle": 0, + "content": "
StateFine-tuneStarGAN FID ↓
Before-60.49
AfterG51.91
G + MaLP52.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.371, + 0.468, + 0.398 + ], + "angle": 0, + "content": "Table 6. Comparison with prior binary detection works. [Keys: D.M.: Detection module, L.M.: Localization module]" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.399, + 0.468, + 0.489 + ], + "angle": 0, + "content": "
MethodTrain GMSet sizeTest GM Average precision (%)↑
CycleGANStarGANGauGAN
Nataraj et al. [35]CycleGAN-10088.2056.20
Zhang et al. [59]AutoGAN-10010061.00
Wang et al. [53]ProGAN-84.0010067.00
Asnani et al. [1]STGAN194.0010069.50
MaLP (D.M.)STGAN194.1010069.61
MaLP (D.M. + L.M.)STGAN194.3010072.16
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.504, + 0.267, + 0.521 + ], + "angle": 0, + "content": "4.5. Other Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.728 + ], + "angle": 0, + "content": "Binary Detection We compare with prior proactive and passive approaches for binary manipulation detection [1, 35, 53, 59]. We adopt the evaluation protocol in [1] to test on images manipulated by CycleGAN, StarGAN, and GauGAN. We are able to perform similar to [1] as shown in Tab. 6. We have better average precision than passive schemes and generalize well to GMs unseen in training. We also conduct experiments to see whether localization can help binary detection to improve the performance, as mentioned in Sec. 3.2.3. The combined predictions' results are better than just using the detection module as shown in Tab. 6. This is intuitive as the localization module provides extra information, thereby increasing the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Inference Speed We compare the inference speed of our MaLP against prior work. [21] uses Deeplabv3-ResNet101 model from PyTorch [40]. In our generalization benchmark shown in Sec. 4.3, we use the ResNet50 model for training the passive baseline. The inference speed per image on an NVIDIA K80 GPU for Deeplabv3-ResNet101, ResNet50, and MaLP are 75.61, 52.66, and \\(29.26~\\mathrm{ms}\\), respectively. MaLP takes less than half the inference time compared to [21] due to our shallow CNN network." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Adversarial Attack Our framework can be considered as an adversarial attack on real images to aid manipula-" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.29, + 0.857, + 0.303 + ], + "angle": 0, + "content": "Table 7. Comparison with adversarial attack methods." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.304, + 0.892, + 0.398 + ], + "angle": 0, + "content": "
MethodSchemeCosine similarity↑
BaldBlack HairEyeglasses
Huang et al. [21]Passive0.81410.69320.6950
PGD [33]Proactive0.80510.75140.8358
FGSM [16]Proactive0.81110.78820.8512
CW [3]Proactive0.80140.83440.8405
MaLPProactive0.82010.79400.8557
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.893, + 0.658 + ], + "angle": 0, + "content": "tion localization. Therefore, it is vital to contrast the performance between our approach and classic adversarial attacks. For this purpose, we perform experiments that make use of adversarial attacks, namely PGD [33], CW [3], and FGSM [16] to guide the learning of the added template. We evaluate on unseen GM AttGAN for unseen attribute modifications. We show the performance comparison in Tab. 7. MaLP has higher cosine similarity across some unseen facial attribute modifications compared to adversarial attacks. This can be explained as the adversarial attack methods being over-fitted to training parameters (data, target network etc.). Therefore, if the testing data is changed with unseen attribute modifications by GMs, the performance of adversarial attacks degrades. Further, these attacks are analogous to our MaLP as a proactive scheme which, in general, have better performance than passive works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.785 + ], + "angle": 0, + "content": "Model Robustness Against Degradations It is necessary to test the robustness of our proposed approach against various types of real-world image editing degradations. We evaluate our method on degradations applied during testing as adopted by [21], which include JPEG compression, blurring, adding noise, and low resolution. The results are shown in Fig. 5. Our proposed MaLP is more robust to real-world degradations than passive schemes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.798, + 0.611, + 0.812 + ], + "angle": 0, + "content": "4.6. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Two-branch Architecture As described in Sec. 3.2.2, MaLP adopts a two-branch architecture to predict the fairness map using the local-level and global-level features, which are estimated by a shallow CNN and a transformer. We ablate by training each branch separately to show the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12349" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.468, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.191, + 0.468, + 0.218 + ], + "angle": 0, + "content": "Figure 5. Comparison of our approach's robustness against common image editing degradations." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.224, + 0.468, + 0.265 + ], + "angle": 0, + "content": "Table 8. Ablation of two-branch architecture. CNN is a shallow network with 10 layers. Training each branch separately has worse localization results than combining them." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.267, + 0.452, + 0.343 + ], + "angle": 0, + "content": "
Network trainedCosine similarity ↑Accuracy ↑
CNN only0.89610.9801
Transformer only0.88480.9856
CNN + ResNet500.86470.9512
CNN + Transformer0.93940.9981
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.468, + 0.496 + ], + "angle": 0, + "content": "effectiveness of combining them. As shown in Tab. 8, if the individual network is trained separately, the performance is lower than the two-branch architecture. Next, to show the efficacy of the transformer, we use a ResNet50 network in place of the transformer to predict the fakeness map. We observe that the performance is even worse than using only the transformer. ResNet50 lacks the added advantage of self-attention in the transformer, which estimates the global-level features much better than a CNN network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.468, + 0.698 + ], + "angle": 0, + "content": "Constraints MaLP leverages different constraints to estimate the fakeness map using an optimized template. We perform an ablation by removing each constraint separately, showing the importance of every constraint. Tab. 9 shows the cosine similarity for localization and accuracy for detection. Removing either the classifier or recovery constraint results in lower detection performance. This is expected as we leverage logits from both \\(\\mathcal{C}\\) and \\(\\varepsilon_{E}\\), and removing the constraint for one network will hurt the logits of the other network. Furthermore, removing the template constraint results in a decrease in performance. Although the gap is small, the template is not properly optimized to have lower magnitude and high-frequency content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.468, + 0.835 + ], + "angle": 0, + "content": "Removing the localization constraint and just applying a \\( L_{2} \\) loss for supervising fakeness maps result in a significant performance drop for both localization and detection, showing the necessity of this constraint. Finally, we show the importance of a learnable template by not optimizing it during the training of MaLP. This hurts the performance a lot, similar to removing the localization constraint. Both these observations prove that our localization constraint and learnable template are important components of MaLP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Template Set Size We perform an ablation to vary the size of the template set \\( S \\). Having multiple templates will improve security if an attacker tries to reverse engineer the template from encrypted images. The results are shown in" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.09, + 0.885, + 0.103 + ], + "angle": 0, + "content": "Table 9. Ablation of constraints used in training our framework." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.104, + 0.887, + 0.205 + ], + "angle": 0, + "content": "
Constraint removedCosine similarity ↑Accuracy ↑
Classifier constraint JC0.93190.9814
Template constraint JT0.91430.9803
Localization constraint JL0.88140.9539
Recovery constraint JR0.92060.9780
Fixed template0.88870.9514
Nothing (MaLP)0.93940.9991
" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.216, + 0.696, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.298, + 0.618, + 0.308 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.216, + 0.888, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.298, + 0.808, + 0.308 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.308, + 0.89, + 0.335 + ], + "angle": 0, + "content": "Figure 6. Ablation study on hyperparameters used in our framework: set size and signal strength." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Fig. 6 (a). The cosine similarity takes a dip when the set size is increased. We also observe the inter-template cosine similarity, which remains constant at a high value of around 0.74 for all templates. This is against the findings of [1]. Localization is a more challenging task than binary detection. Therefore, it is less likely to find different templates for our MaLP in the given feature space compared to [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Signal Strength We vary the template strength hyperparameter \\( m \\) to find its impact on the performance. As shown in Fig. 6 (b), the cosine similarity increases as we increase the strength of the added template. However, this comes with the lower visual quality of the encrypted images if the template strength is increased. The performance doesn't vary much after \\( m = 30\\% \\), which we use for MaLP." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.583, + 0.618, + 0.599 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.776 + ], + "angle": 0, + "content": "This paper focuses on manipulation localization using a proactive scheme (MaLP). We propose to improve the generalization of manipulation localization across unseen GM and facial attribute modifications. We add an optimal template onto the real images and estimate the fakeness map via a two-branch architecture using local and global-level features. MaLP outperforms prior works with much stronger generalization capabilities, as demonstrated by our proposed evaluation benchmark with 22 different GMs in various domains. We show an application of MaLP in finetuning GMs to improve generation quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitations First, the number of publicly available GMs is limited. More thorough testing on many different GMs might give more insights into the problem of generalizable manipulation localization. Second, we show that our MaLP can be used to fine-tune the GMs to improve image generation quality. However, it is based on the pretrained GM. Using our method to train a GM from scratch can be an interesting direction to explore in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12350" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Vishal Asnani, Xi Yin, Tal Hassner, Sijia Liu, and Xiaoming Liu. Proactive image manipulation detection. In CVPR, 2022. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Reverse engineering of generative models: Inferring model hyperparameters from generated images. arXiv preprint arXiv:2106.07873, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.242 + ], + "angle": 0, + "content": "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In SSP, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[5] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[7] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[8] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[9] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[10] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[11] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In CVPR, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[13] Bruce Draper. Reverse engineering of deceptions (red). https://www.darpa.mil/program/reverse-engineering-of-deceptions.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[14] Candice R Gerstner and Hany Farid. Detecting real-time deep-fake videos using active illumination. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[17] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vec" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tor quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[18] Xiao Guo, Yaojie Liu, Anil Jain, and Xiaoming Liu. Multidomain learning for updating face anti-spoofing models. In ECCV, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[20] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[21] Yihao Huang, Felix Juefei-Xu, Qing Guo, Yang Liu, and Geguang Pu. FakeLocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information Forensics and Security, 17:2657-2672, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[22] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[23] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[25] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[26] Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Kumar Singh, and Ming-Hsuan Yang. Diverse imaged-to-image translation via disentangled representations. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[27] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[28] Ming Liu, Yukang Ding, Min Xia, Xiao Liu, Errui Ding, Wangmeng Zuo, and Shilei Wen. STGAN: A unified selective transfer network for arbitrary image attribute editing. In CVPR, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In NeurIPS, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[30] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. PSCC-Net: Progressive spatio-channel correlation network for image manipulation detection and localization. In arXiv preprint arXiv:2103.10596, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[31] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[32] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[33] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[34] Safa C. Medin, Bernhard Egger, Anoop Cherian, Ye Wang, Joshua B. Tenenbaum, Xiaoming Liu, and Tim K. Marks." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12351" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "MOST-GAN: 3D morphable StyleGAN for disentangled face image manipulation. In AAAI, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[35] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, BS Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, and Amit K Roy-Chowdhury. Detecting GAN generated fake images using co-occurrence matrices. Electronic Imaging, 2019:532-1, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[36] Kamyar Nazeri, Eric Ng, and Mehran Ebrahimi. Image colorization using generative adversarial networks. In AMDO, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.237, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[37] Huy H Nguyen, Fuming Fang, Junichi Yamagishi, and Isao Echizen. Multi-task learning for detecting and segmenting manipulated facial images and videos. In BTAS, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[38] Ori Nizan and Ayellet Tal. Breaking the cycle - colleagues are all you need. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[39] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. GauGAN: semantic image synthesis with spatially adaptive normalization. In ACM, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[41] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[42] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[43] Albert Pumarola, Antonio Agudo, Aleix M Martinez, Alberto Sanfeliu, and Francesc Moreno-Noguer. Ganimation: Anatomically-aware facial animation from a single image. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.469, + 0.682 + ], + "angle": 0, + "content": "[45] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: Learning to detect manipulated facial images. In CVPR, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.739 + ], + "angle": 0, + "content": "[46] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[47] Eran Segalis and Eran Galili. OGAN: Disrupting deep-fakes with an adversarial attack that survives training. arXiv preprint arXiv:2006.12247, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[48] Kaede Shiohara and Toshihiko Yamasaki. Detecting deep-fakes with self-blended images. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[49] Kritaphat Songsri-in and Stefanos Zafeiriou. Complement face forensic detection and localization with facial landmarks. arXiv preprint arXiv:1910.05455, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[50] Luan Tran, Xi Yin, and Xiaoming Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, 2017. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[51] Run Wang, Felix Juefei-Xu, Meng Luo, Yang Liu, and Lina Wang. FakeTagger: Robust safeguards against deepfake dissemination via provenance tracking. In ACMM, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[52] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Sketch your own gan. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[53] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. CNN-generated images are surprisingly easy to spot... for now. In CVPR, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[54] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[55] Xi Wu, Zhen Xie, YuTao Gao, and Yu Xiao. SSTNET: Detecting manipulated faces through spatial, steganalysis and temporal features. In ICASSP, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[56] Ying Xu, Kiran Raja, and Marius Pedersen. Supervised contrastive learning for generalizable and explainable deepfakes detection. In WACV, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[57] Chin-Yuan Yeh, Hsi-Wen Chen, Shang-Lun Tsai, and Sheng-De Wang. Disrupting image-translation-based deepfake algorithms with adversarial attacks. In WACVW, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[58] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. DualGAN: Unsupervised dual learning for image-to-image translation. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[59] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in GAN fake images. In WIFS, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[60] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.521, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[61] Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darryll, Alexei A Efros, Oliver Wang, and Eli Shechtman. Toward multimodal image-to-image translation. In NeurIPS, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[62] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. SEAN: Image synthesis with semantic region-adaptive normalization. In CVPR, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12352" + } + ] +] \ No newline at end of file diff --git a/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_origin.pdf b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a4ed6e939a3c2eef4e53f2852e52986594a489c8 --- /dev/null +++ b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/f5c5562b-3811-4dcc-b7bd-6a7a14d3f217_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:238a988ce28b8d069dc3d36fb629e945ff40ddf8534d067e77249182445da339 +size 7874439 diff --git a/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/full.md b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/full.md new file mode 100644 index 0000000000000000000000000000000000000000..223ff0a91ed4e1584f337d5fa7f3789cd1c54dc9 --- /dev/null +++ b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/full.md @@ -0,0 +1,380 @@ +# MaLP: Manipulation Localization Using a Proactive Scheme + +Vishal Asnani1, Xi Yin2, Tal Hassner2, Xiaoming Liu1 +1*Michigan State University, 2Meta AI + +1{asnanivi, liuxm}@msu.edu, 2{yinxi, thassner}@meta.com + +# Abstract + +Advancements in the generation quality of various Generative Models (GMs) has made it necessary to not only perform binary manipulation detection but also localize the modified pixels in an image. However, prior works termed as passive for manipulation localization exhibit poor generalization performance over unseen GMs and attribute modifications. To combat this issue, we propose a proactive scheme for manipulation localization, termed MaLP. We encrypt the real images by adding a learned template. If the image is manipulated by any GM, this added protection from the template not only aids binary detection but also helps in identifying the pixels modified by the GM. The template is learned by leveraging local and global-level features estimated by a two-branch architecture. We show that MaLP performs better than prior passive works. We also show the generalizability of MaLP by testing on 22 different GMs, providing a benchmark for future research on manipulation localization. Finally, we show that MaLP can be used as a discriminator for improving the generation quality of GMs. Our models/codes are available at www.github.com/vishal3477/pro_LOC. + +# 1. Introduction + +We witness numerous Generative Models (GMs) [8, 9, 15, 17, 23-25, 28, 34, 39, 44, 50, 52, 60] being proposed to generate realistic-looking images. These GMs can not only generate an entirely new image [23, 24], but also perform partial manipulation of an input image [9, 9, 28, 60]. The proliferation of these GMs has made it easier to manipulate personal media for malicious use. Prior methods to combat manipulated media focus on binary detection [1, 2, 5, 11, 14, 30, 45, 48, 55, 56], using mouth movement, model parsing, hand-crafted features, etc. + +Recent works go one step further than detection, i.e. manipulation localization, which is defined as follows: given + +![](images/255dcfe857989d1c6bbc0bd011e3dbfca1999f291260392b4df061bf352637da.jpg) +(a) + +![](images/1fdf50dd8a64232386fce9fd8f9d76bd337ad02ae971637e30a3ddf39b11d8dd.jpg) +(b) +Figure 1. (a) High-level idea of MaLP. We encrypt the image by adding a learnable template, which helps to estimate the fakeness map. (b) The cosine similarity (CS) between ground-truth and predicted fakeness maps for 22 unseen GMs. The performance is better for almost all GMs when using our proactive approach. + +a partially manipulated image by a GM (e.g. STGAN [28] modifying hair colors of a face image), the goal is to identify which pixels are modified by estimating a fakeness map [21]. Identifying modified pixels helps to determine the severity of the fakeness in the image, and aid mediaforensics [11, 21]. Also, manipulation localization provides an understanding of the attacker's intent for modification which may further benefit identifying attack toolchains used [13]. + +Recent methods for manipulation localization [27,37,49] focus on estimating the manipulation mask of face-swapped images. They localize modified facial attributes by leveraging attention mechanisms [11], patch-based classifier [4], and face-parsing [21]. The main drawback of these methods is that they do not generalize well to GMs unseen in training. That is when the test images and training images are modified by different GMs, which will likely happen given the vast number of existing GMs. Thus, our work aims for a localization method generalizable to unseen GMs. + +All aforementioned methods are based on a passive scheme as the method receives an image as is for estimation. Recently, proactive methods are gaining success for deepfake tasks such as detection [1], disruption [46, 57], and tagging [51]. These methods are considered proactive as they add different types of signals known as templates for encrypting the image before it is manipulated by a GM. + +This template can be one-hot encoding [51], adversarial perturbation [46], or a learnable noise [1], and is optimized to improve the performance of the defined tasks. + +Motivated by [1], we propose a Proactive scheme for MAnipulation Localization, termed as MaLP, in order to improve generalization. Specifically, MaLP learns an optimized template which, when added to real images, would improve manipulation localization, should they get manipulated. This manipulation can be done by an unseen GM trained on either in-domain or out-of-domain datasets. Furthermore, face manipulation may involve modifying facial attributes unseen in training (e.g. train on hair color modification yet test on gender modification). MaLP incorporates three modules that focus on encryption, detection, and localization. The encryption module selects and adds the template from the template set to the real images. These encrypted images are further processed by localization and detection modules to perform the respective tasks. + +Designing a proactive manipulation localization approach comes with several challenges. First, it is not straightforward to formulate constraints for learning the template unsupervisedly. Second, calculating a fakeness map at the same resolution as the input image is computationally expensive if the decision for each pixel has to be made. Prior works [4, 11] either down-sample the images or use a patch-wise approach, both of which result in inaccurate low-resolution fakeness maps. Lastly, the templates should be generalizable to localize modified regions from unseen GMs. + +We design a two-branch architecture consisting of a shallow CNN network and a transformer to optimize the template during training. While the former leverages local-level features due to its shallow depth, the latter focuses on global-level features to better capture the affinity of the far apart regions. The joint training of both networks enables the MaLP to learn a better template, having embedded the information of both levels. During inference, the CNN network alone is sufficient to estimate the fakeness map with a higher inference efficiency. Compared to prior passive works [11, 21], MaLP improves the generalization performance on unseen GMs. We also demonstrate that MaLP can be used as a discriminator for fine-tuning conventional GMs to improve the quality of GM-generated images. + +In summary, we make the following contributions. + +- We are the first to propose a proactive scheme for image manipulation localization, applicable to both face and generic images. +- Our novel two-branch architecture uses both local and global level features to learn a set of templates in an unsupervised manner. The framework is guided by constraints based on template recovery, fakeness maps classification, and high cosine similarity between predicted and ground-truth fakeness maps. + +Table 1. Comparison of our approach with prior works on manipulation localization and proactive schemes. We show the generalization ability of all works across different facial attribute modifications, unseen GMs trained on datasets with the same domain (in-domain) and different domains (out-domain). [Keys: Attr.: Attributes, Imp.: Improving, L.: Localization, D.: Detection] + +
WorkSchemeTaskTemplateGeneralizationImp. GM
Attr.In-domainOut-domain
[51]ProactiveTagFix✓✓✘✘
[47]ProactiveDisruptLearn✓✘✘✘
[46]ProactiveDisruptLearn✓✓✘✘
[57]ProactiveDisruptLearn✓✘✘✘
[1]ProactiveD.Learn✘✓✓✘
[37]PassiveL. + D.-✘✘✘✘
[49]PassiveL. + D.-✘✘✘✘
[27]PassiveL. + D.-✘✓✘✘
[11]PassiveL. + D.-✓✓✘✘
[4]PassiveL. + D.-✘✓✘✘
[21]PassiveL. + D.-✓✓✘✘
MaLPProactiveL. + D.Learn✓✓✓✓
+ +- MaLP can be used as a plug-and-play discriminator module to fine-tune the generative model to improve the quality of the generated images. +- Our method outperforms State-of-The-Art (SoTA) methods in manipulation localization and detection. Furthermore, our method generalizes well to GMs and modified attributes unseen in training. To facilitate the research of localization, we develop a benchmark for evaluating the generalization of manipulation localization, on images where the train and test GMs are different. + +# 2. Related Work + +Manipulation Localization. Prior works tackle manipulation localization by adopting a passive scheme. Some of them focus on forgery attacks like removal, copymove, and splicing using multi-task learning [37]. Songsriin et al. [49] leverage facial landmarks [10] for manipulation localization. Li et al. [27] estimate the blended boundary for forged face-swap images. [11] uses an attention mechanism to leverage the relationship between pixels and [4] uses a patch-based classifier to estimate modified regions. Recently, Huang et al. [21] utilize gray-scale maps as ground truth for manipulation localization and leverage face parsing with an attention mechanism for prediction. The passive methods discussed above suffer from the generalization issue [4, 10, 11, 21, 37, 49] and estimate a low-resolution fakeness map [11] which is less accurate for the localization purpose. MaLP generalizes better to modified attributes and GMs unseen in training. + +Proactive Scheme. Recently, proactive schemes are developed for various tasks. Wang et al. [51] leverage the recovery of embedded one-hot encoding messages to perform deepfake tagging. A small perturbation is added onto the images by Segalis et al. [47] to disrupt the output of a GM. The same task is performed by Ruiz et al. [46] and Yeh et al. [57], both adding adversarial noise onto the in + +put images. Asnani et al. [1] propose a framework based on adding a learnable template to input images for generalized manipulation detection. Unlike prior works, which focus on binary detection, deepfake disruption, or tagging, our work emphasizes on manipulation localization. We show the comparison of our approach with prior works in Tab. 1. + +Manipulation Detection. The advancement in manipulation detection keeps reaching new heights. Prior works propose to combat deepfakes by exploiting frequency domain patterns [53], up-sampling artifacts [59], model parsing [2], hand-crafted features [35], lip motions [45], and self-attention [11]. Recent methods use self-blended images [48], real-time deviations [14], and self-supervised learning with adversarial training [5]. Finally, methods based on contrastive learning [56] and proactive scheme [1] have explicitly focused on generalized manipulation detection across unknown GMs. + +# 3. Proposed Approach + +# 3.1. Problem Formulation + +Passive Manipulation Localization Let $\pmb{I}^{R}$ be a set of real images that are manipulated by a GM $G$ to output the set of manipulated images $G(\pmb{I}^{R})$ . Prior passive works perform manipulation localization by estimating the fakeness map $M_{pred}$ with the following objective: + +$$ +\left. \min _ {\theta_ {\mathcal {E}}} \left\{\sum_ {j} \left(\left| \left| \mathcal {E} \left(G \left(I _ {j} ^ {R}\right); \theta_ {\mathcal {E}}\right) - M _ {G T} \right| \right| _ {2}\right) \right\}, \right. \tag {1} +$$ + +where $\mathcal{E}$ denotes the passive framework with parameters $\theta_{\mathcal{E}}$ and $M_{GT}$ is the ground-truth fakeness map. + +To represent the fakeness map, some prior methods [11, 27, 49] choose a binary map by applying a threshold on the difference between the real and manipulated images. This is undesirable as the threshold selection is highly subjective and sensitive, leading to inaccurate fakeness maps. Therefore, we adopt the continuous gray-scale map for calculating the ground-truth fakeness maps [21], formulated as: + +$$ +\boldsymbol {M} _ {G T} = \operatorname {G r a y} \left(\left| \boldsymbol {I} ^ {R} - G \left(\boldsymbol {I} ^ {R}\right) \right|\right) / 2 5 5, \tag {2} +$$ + +where $Gray(.)$ converts the image to gray-scale. + +Proactive Scheme Asnani et al. [1] define adding the template as a transformation $\mathcal{T}$ applied to images $I^R$ , resulting in the encrypted images $\mathcal{T}(I^R)$ . The added template acts as a signature of the defender and is learned during the training, aiming to improve the performance of the task at hand, e.g. detection, disruption, and tagging. Motivated by [1] that uses multiple templates, we have a set of $n$ orthogonal templates $\mathcal{S} = \{S_1, S_2, \ldots, S_n\}$ where $S_i \in \mathbb{R}^{128 \times 128}$ , for a real image $I_j^R \in I^R$ , transformation $\mathcal{T}$ is defined as: + +$$ +\mathcal {T} \left(\boldsymbol {I} _ {j} ^ {R}; \boldsymbol {S} _ {i}\right) = \boldsymbol {I} _ {j} ^ {R} + \boldsymbol {S} _ {i}, \text {w h e r e} i \in \{1, 2, \dots , n \}. \tag {3} +$$ + +The templates are optimized such that adding them to the real images wouldn't result in a noticeable visual difference, yet helps manipulation localization. + +Proactive Manipulation Localization. Unlike the passive schemes [11, 21, 27, 37], we learn an optimal template set to help manipulation localization. For the encrypted images $\mathcal{T}(\boldsymbol{I}^R)$ , we formulate the estimation of the fakeness map as: + +$$ +\left. \min _ {\theta_ {\mathcal {E} _ {P}}, \boldsymbol {S} _ {i}} \left\{\sum_ {j} \left(\left| \left| \mathcal {E} _ {P} \left(G \left(\mathcal {T} \left(\boldsymbol {I} _ {j} ^ {R}; \boldsymbol {S} _ {i}\right)\right); \theta_ {\mathcal {E} _ {P}}\right) - \boldsymbol {M} _ {G T} \right| \right| _ {2}\right) \right\}. \right. \tag {4} +$$ + +where $\mathcal{E}_P$ is the proactive framework with parameters $\theta_{\mathcal{E}_P}$ . + +However, as the output of the GM has changed from images in set $G(\mathbf{I}^R)$ to images in set $G(\mathcal{T}(\mathbf{I}^R))$ , in our proactive approach, the calculation of the ground-truth fakeness map shall be changed from Eq. 2 to the follows: + +$$ +\boldsymbol {M} _ {G T} = \operatorname {G r a y} \left(\left| \boldsymbol {I} ^ {R} - G \left(\mathcal {T} \left(\boldsymbol {I} ^ {R}\right)\right) \right|\right) / 2 5 5. \tag {5} +$$ + +# 3.2. Manipulation Localization + +MaLP consists of three modules: encryption, localization, and detection. The encryption module is used to encrypt the real images. The localization module estimates the fakeness map using a two-branch architecture. The detection module performs binary detection for the encrypted and manipulated images by recovering the template and using the classifier in the localization module. All three modules, as detailed next, are trained in an end-to-end manner. + +# 3.2.1 Encryption Module + +Following the procedure in [1], we add a randomly selected learnable template from the template set to a real image. We control the strength of the added template using a hyperparameter $m$ , which prevents the degradation of the image quality. The encryption process is summarised below: + +$$ +\mathcal {T} \left(\boldsymbol {I} _ {j} ^ {R}\right) = \boldsymbol {I} _ {j} ^ {R} + m \times \boldsymbol {S} _ {i} \text {w h e r e} i = \operatorname {R a n d} (1, 2, \dots , n). \tag {6} +$$ + +We select the value of $m$ as $30\%$ for our framework. + +We optimize the template set by focusing on properties like low magnitude, orthogonality, and high-frequency content [1]. The properties are applied as constraints as follows. + +$$ +J _ {T} = \lambda_ {1} \times \sum_ {i = 1} ^ {n} \left\| \boldsymbol {S} _ {i} \right\| _ {2} + \lambda_ {2} \times \sum_ {\substack {i, j = 1 \\ i \neq j}} ^ {n} \operatorname {CS} \left(\boldsymbol {S} _ {i}, \boldsymbol {S} _ {j}\right) + \lambda_ {3} \times \left\| \mathcal {L} (\mathfrak {F} (\boldsymbol {S})) \right\| _ {2}, \tag{7} +$$ + +where CS is the cosine similarity, $\mathcal{L}$ is the low-pass filter, $\mathfrak{F}$ is the fourier transform, $\lambda_1,\lambda_2,\lambda_3$ are weights for losses of low magnitude, orthogonality and high-frequency content, respectively. + +![](images/601d127cebb1148b646b0d3f2a73692e77c71504fbf39a5a4281021d4de09cca.jpg) +Figure 2. The overview of MaLP. It includes three modules: encryption, localization, and detection. We randomly select a template from the template set and add it to the real image as encryption. The GM is used in inference mode to manipulate the encrypted image. The detection module recovers the added template for binary detection. The localization module uses a two-branch architecture to estimate the fakeness map. Lastly, we apply the classifier to the fakeness map to better distinguish them from each other. Best viewed in color. + +# 3.2.2 Localization Module + +To design the localization module, we consider two desired properties: a larger receptive field for fakeness map estimation and high inference efficiency. A network with a large receptive field will consider far-apart regions in learning features for localization. Yet, large receptive fields normally come from deeper networks, implying slower inference. + +In light of these properties, we design a two-branch architecture consisting of a shallow CNN network $\mathcal{E}_C$ and a ViT transformer [12] $\mathcal{E}_T$ (see Fig. 2). The intuition is to have one shallow branch to capture local features, and one deeper branch to capture global features. While training with both branches helps to learn better templates, in inference we only use the shallow branch for a higher efficiency. Specifically, the shallow CNN network has 10 layers which is efficient in inference but can only capture the local features due to small receptive fields. To capture global information, we adopt the ViT transformer. With the self-attention between the image patches, the transformer can estimate the fakeness map considering the far-apart regions. + +Both the CNN and transformer are trained jointly to estimate a better template set, resembling the concept of the ensemble of networks. We empirically show that training both networks simultaneously results in higher performance than training either network separately. As the shallow CNN network is much faster in inference than the transformer, we use the transformer only in training to optimize the templates and switch off the transformer branch in inference. + +To estimate the fakeness map, we leverage the supervision of the ground-truth fakeness map in Eq. 5. For fake images, we maximize the cosine similarity $(CS)$ and struc + +tural similarity index measure $(SS)$ between the predicted and ground-truth fakeness map. However, the fakeness map should be a zero image for encrypted images. Therefore, we apply an $L_{2}$ loss [21] to minimize the predicted map to zero for encrypted images. To maximize the difference between the two fakeness maps, we further minimize the cosine similarity between the predicted map from encrypted images and $M_{GT}$ . The localization loss is defined as: + +$$ +J _ {L} = \left\{ \begin{array}{l l} \left\{\lambda_ {4} \times | | \mathcal {E} _ {C / T} (\boldsymbol {I}) | | _ {2} ^ {2} + \right. & \text {i f} \boldsymbol {I} \in \mathcal {T} \left(\boldsymbol {I} ^ {R}\right) \\ \left. \lambda_ {5} \times \operatorname {C S} \left(\mathcal {E} _ {C / T} (\boldsymbol {I}), M _ {G T}\right) \right\} & \\ \left\{\lambda_ {6} \times (1 - \operatorname {C S} \left(\mathcal {E} _ {C / T} (\boldsymbol {I}), M _ {G T}\right)) + \right. & \text {i f} \boldsymbol {I} \in G \left(\mathcal {T} \left(\boldsymbol {I} ^ {R}\right)\right) \\ \left. \lambda_ {7} \times (1 - S S \left(\mathcal {E} _ {C / T} (\boldsymbol {I}), M _ {G T}\right)) \right\}. & \end{array} \right. \tag {8} +$$ + +Finally, we have a classifier to make a binary decision of real vs. fake using the fakeness maps. This classifier is included in the framework to aid the detection module for binary detection of the input images, which will be discussed in Sec. 3.2.3. Another reason to have the classifier is to make the fakeness maps from encrypted and fake images to be distinguishable. We find that this design allows our training to converge much faster. + +# 3.2.3 Detection Module + +To leverage the added template for manipulation detection, we perform template recovery using encoder $\mathcal{E}_E$ . We follow the procedure in [1] to recover the added template from the encrypted images by maximizing the cosine similarity between $S$ and $S_R$ . However, for manipulated images, we minimize the cosine similarity between the recovered tem + +![](images/567c396b3f3cbecd22bc5b75a218786205e9d437871399760f9c8abde3756f81.jpg) +(a) +(b) +(c) +(d) +(e) +(f) +Figure 3. Visualization of fakeness maps for faces and generic images showing generalization across unseen attribute modifications and GMs: (a) real image, (b) encrypted image, (c) manipulated image, (d) $M_{GT}$ , (e) predicted fakeness map for encrypted images, and (f) predicted fakeness map for manipulated images. The first column shows the manipulation of (seen GM, seen attribute modification) i.e. (STGAN, bald). Following two columns show the manipulation of (seen GM, unseen attribute modification) i.e. (STGAN, [bangs, pale skin]. The fourth and fifth columns show manipulation of unseen GM, GauGAN for non-face images. The last column shows manipulation by unseen GM, DRIT. We see that the fakeness map of manipulated images is more bright and similar to $M_{GT}$ , while the real fakeness map is more close to zero. We use the cmap as "pink" to better visualize the fakeness map. All face images come from SiWM-v2 data [18]. + +plate $(S_R)$ and all the templates in the template set $S$ . + +$$ +J _ {R} = \left\{ \begin{array}{l l} \lambda_ {8} \times (1 - \operatorname {C S} (\boldsymbol {S}, \boldsymbol {S} _ {R})) & \text {i f} x \in \mathcal {T} (\boldsymbol {I} ^ {R}) \\ \lambda_ {9} \times \left(\sum_ {i = 1} ^ {n} (\operatorname {C S} (\boldsymbol {S} _ {i}, \boldsymbol {S} _ {R}))\right) & \text {i f} x \in G (\mathcal {T} (\boldsymbol {I} ^ {R})). \end{array} \right. \tag {9} +$$ + +Further, we leverage our estimated fakeness map to help manipulation detection. As discussed in the previous section, we apply a classifier $\mathcal{C}$ to perform binary classification of the predicted fakeness map for the encrypted and fake images. The logits of the classifier are further combined with the cosine similarity of the recovered template. The averaged logits are back-propagated using the binary cross-entropy constraint. This not only improves the performance of manipulation detection but also helps manipulation localization. Therefore, we apply the binary cross entropy loss on the averaged logits as follows: + +$$ +\begin{array}{l} J _ {C} = \lambda_ {1 0} \times - \sum_ {j} \left\{y _ {j}. \log \left[ \frac {\mathcal {C} \left(\boldsymbol {X} _ {j}\right) + C S \left(\boldsymbol {S} _ {R} , \boldsymbol {S}\right)}{2} \right] - \right. \tag {10} \\ (1 - y _ {j}). \log \left[ 1 - \frac {\mathcal {C} (\boldsymbol {X} _ {j}) + C S (\boldsymbol {S} _ {R} , \boldsymbol {S})}{2} \right] \Bigg \}, \\ \end{array} +$$ + +Table 2. Manipulation localization comparison with prior works. + +
MethodLocalizationDetection
CS ↑PSNR ↑SSIM ↑Accuracy ↑EER ↓AUC ↑
[11]0.62306.2140.21780.99750.00500.9975
[21]0.883122.8900.78760.99450.00770.9998
MaLP0.939423.0200.73120.99910.00721.0
+ +where $y_{j}$ is the class label, $\pmb{S}$ and $S_{R}$ are the added and recovered template respectively. + +Our framework is trained in an end-to-end manner with the overall loss function as follows: + +$$ +J = J _ {T} + J _ {R} + J _ {C} + J _ {L}. \tag {11} +$$ + +# 3.3. MaLP as A Discriminator + +One application of MaLP is to leverage our proposed localization module as a discriminator for improving the quality of the manipulated images. MaLP performs binary classification by estimating a fakeness map, which can be used as an objective. This results in output images being resilient to manipulation localization, thereby lowering the performance of our framework. + +We use MaLP as a plug-and-play discriminator to improve image generation quality through fine-tuning pretrained GMs. The generation quality and manipulation localization will compete head-to-head, resulting in a better quality of the manipulated images. We define the fine-tuning objective for the GM as follows: + +$$ +\begin{array}{l} \min _ {\theta_ {G}} \max _ {\theta_ {M a L P}, \boldsymbol {S} _ {i}} \left\{\sum_ {j} \left(\mathbb {E} \left[ l o g \left(\mathcal {E} _ {M a L P} \left(\mathcal {T} \left(\boldsymbol {I} _ {j} ^ {R}\right)\right); \theta_ {M a L P}\right) \right] + \right. \right. \\ \left. \mathbb {E} \left[ 1 - \log \left(\mathcal {E} _ {M a L P} \left(G \left(\mathcal {T} \left(I _ {j} ^ {R}; \boldsymbol {S} _ {i}\right); \theta_ {G}\right); \theta_ {M a L P}\right)\right) \right]\right) \Bigg \}. \tag {12} \\ \end{array} +$$ + +where $\mathcal{E}_{MaLP}$ is our framework with $\theta_{MaLP}$ parameters. + +# 4. Experiments + +# 4.1. Experimental Setup + +Settings Following the settings in [21], we use STGAN [28] to manipulate images from CelebA [31] dataset and train on bald facial attribute modification. In order to evaluate the generalization of image manipulation localization, we construct a new benchmark that consists of 200 real images of 22 different GMs on various data domains. The real images are chosen from the dataset on which the GM is trained on. The list of GMs, datasets and implementation details are provided in the supplementary. + +Evaluation Metrics We use cosine similarity (CS), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) as adopted by [21] to evaluate manipulation localization since the GT is a continuous map. For binary detection, we use the area under the curve (AUC), equal error rate (EER), and accuracy score [21]. + +Table 3. Comparison of localization performance across unseen GMs and attribute modifications. We train on STGAN bald/smile attribute modification and test on AttGAN/StyleGAN. + +
MethodCosine similarity ↑(AttGAN)Cosine similarity ↑ (StyleGAN)
BaldBlack HairEyeglassesSmileAgeGender
[21]0.81410.69320.69500.61760.31410.6470
MaLP0.82010.79400.85570.81590.82550.8016
+ +# 4.2. Comparison with Baselines + +We compare our results with [21] and [11] for manipulation localization. The results are shown in Tab. 2. MaLP has higher cosine similarity and similar PSNR for localization compared to [21]. However, we observe a dip in SSIM. This might be because of the degradation caused by adding our template to the real images and then performing the manipulation. The learned template helps localize the manipulated regions better, as demonstrated by cosine similarity, but the degradation affects SSIM and PSNR. We also compare the performance of real vs. fake binary detection. As expected, our proposed proactive approach outperforms the passive methods with a perfect AUC and near-perfect accuracy. We also show visual examples of fakeness maps for images modified by unseen GMs in Fig. 3. MaLP is able to estimate the fakeness map for unseen modifications and GMs across face/generic image datasets. + +# 4.3. Generalization + +Across Attribute Modifications Following the settings in [21], we evaluate the performance of MaLP across unseen attribute modifications. Specifically, we train MaLP using STGAN with the bald/smile attribute modification and test it on unseen attribute modifications with unseen GMs: AttGAN/StyleGAN. As shown in Tab. 3, MaLP is more generalizable to all unseen attribute modifications. Furthermore, AttGAN shares the high-level architecture with STGAN but not with StyleGAN. We observe a significant increase in localization performance for StyleGAN compared to AttGAN. This shows that, unlike our MaLP, passive works perform much worse if the test GM doesn't share any similarity with the training GM. + +Across GMs Although [21] tries to show generalization across unseen GMs; it is limited by the GMs within the same domain of the dataset used in training. We propose a benchmark to evaluate the generalization performance for future manipulation localization works that consists of 22 different GMs in various domains. We select GMs that are publicly released and can perform partial manipulation. + +As no open-source code base is available for [21], we train a passive approach using a ResNet50 [19] network to estimate the fakeness map as the baseline for comparison. Further, we compare our approach with [4, 11]. Although [4, 11] estimate a fakeness map, it has at least $5 \times$ lower resolution compared to input images due to + +![](images/ae2e28d09a67750ebd9891c7e909bd33228f3a8ceda7eef3d3d3711fb19c0658.jpg) +(a) + +![](images/c13de7d103ce68c48e9ac785738e75663e8b59909114dbcc440105ee2af30208.jpg) +0.58 +0.62 +0.61 +0.59 +(b) + +![](images/9d462d6840d020c7f4ea6ebb61189e6f10db23bffb6cffceec47b00e6715b3af.jpg) +0.42 +0.45 +0.40 +0.44 +Figure 4. Visualization of (a) encrypted images, (b) manipulated images before fine-tuning, and (c) manipulated images after fine-tuning. The generation quality has improved after we fine-tune the GM using our framework as a discriminator. The artifacts in the images have been reduced, and the face skin color is less pale and more realistic. We also specify the cosine similarity of the predicted fakeness map and $M_{GT}$ . The GM is able to decrease the performance of our framework after fine-tuning. All face images come from SiWM-v2 data [18]. + +their patch-based methodology. For a fair comparison, we rescale their predicted fakeness maps to the resolution of $M_{GT}$ . We compare the cosine similarity in Tab. 4. MaLP is able to outperform all the baselines for almost all GMs, which proves the effectiveness of the proactive scheme. + +We also evaluate the performance of $\mathcal{E}_C$ for high-resolution images. For encryption, we upsample the $128\times 128$ template to the original resolution of images and evaluate $\mathcal{E}_C$ on these higher resolution encrypted images. We observe similar performance of $\mathcal{E}_C$ for higher resolution images in Tab. 4, proving the versatility of $\mathcal{E}_C$ to image sizes. + +# 4.4. Improving Quality of GMs + +We fine-tune the GM into fooling our framework to generate a fakeness map as a zero image. This process results in better-quality images. Initially, we train MaLP with the pretrained GM so that it can perform manipulation localization. Next, to fine-tune the GM, we adopt two strategies. First, we freeze MaLP and fine-tune the GM only. Second, we fine-tune both the GM and the MaLP but update the MaLP with a lower learning rate. The result for fine-tuning StarGAN is shown in Tab. 5. We observe that for both strategies, MaLP reduces the FID score of StarGAN. We also show some visual examples in Fig. 4. We see that the images are of better quality after fine-tuning, and many artifacts in the images manipulated by the pretrained model are removed. + +Table 4. Benchmark for manipulation localization across 22 different unseen GMs, showing cosine similarity between ground-truth and predicted fakeness maps. We compare our proactive vs. passive baselines [4, 11, 19] approach to highlight the generalization ability of our MaLP. We scale the images to $128^{2}$ for "sc." and keep the resolution as is for "no sc". + +
GMSEAN [62]StarGAN [8]CycleGAN [60]GauGAN [39]Con.Enc. [41]StarGAN2 [9]ALAE [42]BiGAN [61]AuGAN [60]GANim [43]DRGAN [50]ILVR [7]
Resolution256212822562256212822562256225623402128212822562
ResNet50 [19]0.86140.75130.67150.76150.86390.81960.67660.65140.66390.68710.80290.7018
[4]0.75140.71110.79810.80160.78940.70260.71560.72170.75160.76120.71150.7851
[11]0.79610.78870.80140.82560.85410.70340.75490.78050.72320.84570.72390.7854
MaLP (sc.)0.93760.87180.91280.92510.85460.88360.91920.91810.88940.96250.75120.8003
MaLP (no sc.)0.92580.87180.92450.91250.85460.87850.91410.92290.91490.96250.75120.8359
GMDRIT [26]Pix2Pix [22]CounGAN [38]DualGAN [58]ESRGAN [54]UNIT [29]MUNIT [20]ColGAN [36]GDWCT [6]RePaint [32]Average
Resolution256225621282256210242512 × 931256 × 512128212822562-
ResNet50 [19]0.74860.67190.72930.73650.87030.70830.66010.75960.83500.65120.7401
[4]0.78710.77690.81460.75690.81680.80640.67880.76100.86910.75160.7645
[11]0.81200.77810.85590.77210.82410.80860.70970.78740.88790.76960.7903
MaLP (sc.)0.88670.89150.93260.88720.83480.82140.75650.80960.93840.81020.8725
MaLP (no sc.)0.90840.87140.93260.84320.87430.83910.78600.80960.93840.82900.8773
+ +Table 5. FID score comparison for the application of our approach as a discriminator for improving the generation quality of the GM + +
StateFine-tuneStarGAN FID ↓
Before-60.49
AfterG51.91
G + MaLP52.07
+ +Table 6. Comparison with prior binary detection works. [Keys: D.M.: Detection module, L.M.: Localization module] + +
MethodTrain GMSet sizeTest GM Average precision (%)↑
CycleGANStarGANGauGAN
Nataraj et al. [35]CycleGAN-10088.2056.20
Zhang et al. [59]AutoGAN-10010061.00
Wang et al. [53]ProGAN-84.0010067.00
Asnani et al. [1]STGAN194.0010069.50
MaLP (D.M.)STGAN194.1010069.61
MaLP (D.M. + L.M.)STGAN194.3010072.16
+ +# 4.5. Other Comparisons + +Binary Detection We compare with prior proactive and passive approaches for binary manipulation detection [1, 35, 53, 59]. We adopt the evaluation protocol in [1] to test on images manipulated by CycleGAN, StarGAN, and GauGAN. We are able to perform similar to [1] as shown in Tab. 6. We have better average precision than passive schemes and generalize well to GMs unseen in training. We also conduct experiments to see whether localization can help binary detection to improve the performance, as mentioned in Sec. 3.2.3. The combined predictions' results are better than just using the detection module as shown in Tab. 6. This is intuitive as the localization module provides extra information, thereby increasing the performance. + +Inference Speed We compare the inference speed of our MaLP against prior work. [21] uses Deeplabv3-ResNet101 model from PyTorch [40]. In our generalization benchmark shown in Sec. 4.3, we use the ResNet50 model for training the passive baseline. The inference speed per image on an NVIDIA K80 GPU for Deeplabv3-ResNet101, ResNet50, and MaLP are 75.61, 52.66, and $29.26~\mathrm{ms}$ , respectively. MaLP takes less than half the inference time compared to [21] due to our shallow CNN network. + +Adversarial Attack Our framework can be considered as an adversarial attack on real images to aid manipula- + +Table 7. Comparison with adversarial attack methods. + +
MethodSchemeCosine similarity↑
BaldBlack HairEyeglasses
Huang et al. [21]Passive0.81410.69320.6950
PGD [33]Proactive0.80510.75140.8358
FGSM [16]Proactive0.81110.78820.8512
CW [3]Proactive0.80140.83440.8405
MaLPProactive0.82010.79400.8557
+ +tion localization. Therefore, it is vital to contrast the performance between our approach and classic adversarial attacks. For this purpose, we perform experiments that make use of adversarial attacks, namely PGD [33], CW [3], and FGSM [16] to guide the learning of the added template. We evaluate on unseen GM AttGAN for unseen attribute modifications. We show the performance comparison in Tab. 7. MaLP has higher cosine similarity across some unseen facial attribute modifications compared to adversarial attacks. This can be explained as the adversarial attack methods being over-fitted to training parameters (data, target network etc.). Therefore, if the testing data is changed with unseen attribute modifications by GMs, the performance of adversarial attacks degrades. Further, these attacks are analogous to our MaLP as a proactive scheme which, in general, have better performance than passive works. + +Model Robustness Against Degradations It is necessary to test the robustness of our proposed approach against various types of real-world image editing degradations. We evaluate our method on degradations applied during testing as adopted by [21], which include JPEG compression, blurring, adding noise, and low resolution. The results are shown in Fig. 5. Our proposed MaLP is more robust to real-world degradations than passive schemes. + +# 4.6. Ablations + +Two-branch Architecture As described in Sec. 3.2.2, MaLP adopts a two-branch architecture to predict the fairness map using the local-level and global-level features, which are estimated by a shallow CNN and a transformer. We ablate by training each branch separately to show the + +![](images/996afe55f94d70e619084c2b3567fe3017ddd9a25bf2498576d48dfa3e1218b9.jpg) +Figure 5. Comparison of our approach's robustness against common image editing degradations. + +Table 8. Ablation of two-branch architecture. CNN is a shallow network with 10 layers. Training each branch separately has worse localization results than combining them. + +
Network trainedCosine similarity ↑Accuracy ↑
CNN only0.89610.9801
Transformer only0.88480.9856
CNN + ResNet500.86470.9512
CNN + Transformer0.93940.9981
+ +effectiveness of combining them. As shown in Tab. 8, if the individual network is trained separately, the performance is lower than the two-branch architecture. Next, to show the efficacy of the transformer, we use a ResNet50 network in place of the transformer to predict the fakeness map. We observe that the performance is even worse than using only the transformer. ResNet50 lacks the added advantage of self-attention in the transformer, which estimates the global-level features much better than a CNN network. + +Constraints MaLP leverages different constraints to estimate the fakeness map using an optimized template. We perform an ablation by removing each constraint separately, showing the importance of every constraint. Tab. 9 shows the cosine similarity for localization and accuracy for detection. Removing either the classifier or recovery constraint results in lower detection performance. This is expected as we leverage logits from both $\mathcal{C}$ and $\varepsilon_{E}$ , and removing the constraint for one network will hurt the logits of the other network. Furthermore, removing the template constraint results in a decrease in performance. Although the gap is small, the template is not properly optimized to have lower magnitude and high-frequency content. + +Removing the localization constraint and just applying a $L_{2}$ loss for supervising fakeness maps result in a significant performance drop for both localization and detection, showing the necessity of this constraint. Finally, we show the importance of a learnable template by not optimizing it during the training of MaLP. This hurts the performance a lot, similar to removing the localization constraint. Both these observations prove that our localization constraint and learnable template are important components of MaLP. + +Template Set Size We perform an ablation to vary the size of the template set $S$ . Having multiple templates will improve security if an attacker tries to reverse engineer the template from encrypted images. The results are shown in + +Table 9. Ablation of constraints used in training our framework. + +
Constraint removedCosine similarity ↑Accuracy ↑
Classifier constraint JC0.93190.9814
Template constraint JT0.91430.9803
Localization constraint JL0.88140.9539
Recovery constraint JR0.92060.9780
Fixed template0.88870.9514
Nothing (MaLP)0.93940.9991
+ +![](images/a883284c91c6560a0e7db0000cfbb0b571e604bc747af57c583e964b60f02ef4.jpg) +(a) + +![](images/6af8e81bd60cd8193c3050d37bb2df239eb871a1cf7b6760ab9b67260b2d09fe.jpg) +(b) +Figure 6. Ablation study on hyperparameters used in our framework: set size and signal strength. + +Fig. 6 (a). The cosine similarity takes a dip when the set size is increased. We also observe the inter-template cosine similarity, which remains constant at a high value of around 0.74 for all templates. This is against the findings of [1]. Localization is a more challenging task than binary detection. Therefore, it is less likely to find different templates for our MaLP in the given feature space compared to [1]. + +Signal Strength We vary the template strength hyperparameter $m$ to find its impact on the performance. As shown in Fig. 6 (b), the cosine similarity increases as we increase the strength of the added template. However, this comes with the lower visual quality of the encrypted images if the template strength is increased. The performance doesn't vary much after $m = 30\%$ , which we use for MaLP. + +# 5. Conclusion + +This paper focuses on manipulation localization using a proactive scheme (MaLP). We propose to improve the generalization of manipulation localization across unseen GM and facial attribute modifications. We add an optimal template onto the real images and estimate the fakeness map via a two-branch architecture using local and global-level features. MaLP outperforms prior works with much stronger generalization capabilities, as demonstrated by our proposed evaluation benchmark with 22 different GMs in various domains. We show an application of MaLP in finetuning GMs to improve generation quality. + +Limitations First, the number of publicly available GMs is limited. More thorough testing on many different GMs might give more insights into the problem of generalizable manipulation localization. Second, we show that our MaLP can be used to fine-tune the GMs to improve image generation quality. However, it is based on the pretrained GM. Using our method to train a GM from scratch can be an interesting direction to explore in the future. + +# References + +[1] Vishal Asnani, Xi Yin, Tal Hassner, Sijia Liu, and Xiaoming Liu. Proactive image manipulation detection. In CVPR, 2022. 1, 2, 3, 4, 7, 8 +[2] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Reverse engineering of generative models: Inferring model hyperparameters from generated images. arXiv preprint arXiv:2106.07873, 2021. 1, 3 +[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In SSP, 2017. 7 +[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020. 1, 2, 6, 7 +[5] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, 2022. 1, 3 +[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In CVPR, 2019. 7 +[7] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 7 +[8] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, 2018. 1, 7 +[9] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 1, 7 +[10] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, 2017. 2 +[11] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In CVPR, 2020. 1, 2, 3, 5, 6, 7 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 4 +[13] Bruce Draper. Reverse engineering of deceptions (red). https://www.darpa.mil/program/reverse-engineering-of-deceptions.1 +[14] Candice R Gerstner and Hany Farid. Detecting real-time deep-fake videos using active illumination. In CVPR, 2022. 1, 3 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1 +[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 7 +[17] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vec + +tor quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 1 +[18] Xiao Guo, Yaojie Liu, Anil Jain, and Xiaoming Liu. Multidomain learning for updating face anti-spoofing models. In ECCV, 2022. 5, 6 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 7 +[20] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In ECCV, 2018. 7 +[21] Yihao Huang, Felix Juefei-Xu, Qing Guo, Yang Liu, and Geguang Pu. FakeLocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information Forensics and Security, 17:2657-2672, 2022. 1, 2, 3, 4, 5, 6, 7 +[22] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 7 +[23] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 1 +[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1 +[25] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 1 +[26] Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Kumar Singh, and Ming-Hsuan Yang. Diverse imaged-to-image translation via disentangled representations. In ECCV, 2018. 7 +[27] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, 2020. 1, 2, 3 +[28] Ming Liu, Yukang Ding, Min Xia, Xiao Liu, Errui Ding, Wangmeng Zuo, and Shilei Wen. STGAN: A unified selective transfer network for arbitrary image attribute editing. In CVPR, 2019. 1, 5 +[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In NeurIPS, 2017. 7 +[30] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. PSCC-Net: Progressive spatio-channel correlation network for image manipulation detection and localization. In arXiv preprint arXiv:2103.10596, 2021. 1 +[31] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5 +[32] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 7 +[33] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 7 +[34] Safa C. Medin, Bernhard Egger, Anoop Cherian, Ye Wang, Joshua B. Tenenbaum, Xiaoming Liu, and Tim K. Marks. + +MOST-GAN: 3D morphable StyleGAN for disentangled face image manipulation. In AAAI, 2022. 1 +[35] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, BS Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, and Amit K Roy-Chowdhury. Detecting GAN generated fake images using co-occurrence matrices. Electronic Imaging, 2019:532-1, 2019. 3, 7 +[36] Kamyar Nazeri, Eric Ng, and Mehran Ebrahimi. Image colorization using generative adversarial networks. In AMDO, 2018. 7 +[37] Huy H Nguyen, Fuming Fang, Junichi Yamagishi, and Isao Echizen. Multi-task learning for detecting and segmenting manipulated facial images and videos. In BTAS, 2019. 1, 2, 3 +[38] Ori Nizan and Ayellet Tal. Breaking the cycle - colleagues are all you need. In CVPR, 2020. 7 +[39] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. GauGAN: semantic image synthesis with spatially adaptive normalization. In ACM, 2019. 1, 7 +[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 7 +[41] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 7 +[42] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In CVPR, 2020. 7 +[43] Albert Pumarola, Antonio Agudo, Aleix M Martinez, Alberto Sanfeliu, and Francesc Moreno-Noguer. Ganimation: Anatomically-aware facial animation from a single image. In ECCV, 2018. 7 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1 +[45] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: Learning to detect manipulated facial images. In CVPR, 2019. 1, 3 +[46] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In ECCV, 2020. 1, 2 +[47] Eran Segalis and Eran Galili. OGAN: Disrupting deep-fakes with an adversarial attack that survives training. arXiv preprint arXiv:2006.12247, 2020. 2 +[48] Kaede Shiohara and Toshihiko Yamasaki. Detecting deep-fakes with self-blended images. In CVPR, 2022. 1, 3 +[49] Kritaphat Songsri-in and Stefanos Zafeiriou. Complement face forensic detection and localization with facial landmarks. arXiv preprint arXiv:1910.05455, 2019. 1, 2, 3 +[50] Luan Tran, Xi Yin, and Xiaoming Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, 2017. 1, 7 + +[51] Run Wang, Felix Juefei-Xu, Meng Luo, Yang Liu, and Lina Wang. FakeTagger: Robust safeguards against deepfake dissemination via provenance tracking. In ACMM, 2021. 1, 2 +[52] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Sketch your own gan. In ICCV, 2021. 1 +[53] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. CNN-generated images are surprisingly easy to spot... for now. In CVPR, 2020. 3, 7 +[54] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In CVPR, 2021. 7 +[55] Xi Wu, Zhen Xie, YuTao Gao, and Yu Xiao. SSTNET: Detecting manipulated faces through spatial, steganalysis and temporal features. In ICASSP, 2020. 1 +[56] Ying Xu, Kiran Raja, and Marius Pedersen. Supervised contrastive learning for generalizable and explainable deepfakes detection. In WACV, 2022. 1, 3 +[57] Chin-Yuan Yeh, Hsi-Wen Chen, Shang-Lun Tsai, and Sheng-De Wang. Disrupting image-translation-based deepfake algorithms with adversarial attacks. In WACVW, 2020. 1, 2 +[58] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. DualGAN: Unsupervised dual learning for image-to-image translation. In CVPR, 2017. 7 +[59] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in GAN fake images. In WIFS, 2019. 3, 7 +[60] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 1, 7 +[61] Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darryll, Alexei A Efros, Oliver Wang, and Eli Shechtman. Toward multimodal image-to-image translation. In NeurIPS, 2017. 7 +[62] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. SEAN: Image synthesis with semantic region-adaptive normalization. In CVPR, 2020. 7 \ No newline at end of file diff --git a/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/images.zip b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c5e33d4a87f9d1ad0a647cc0f43dd6b498898787 --- /dev/null +++ b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ea8fd73712aac017df3b53162659f63605c4fae214fae9dcf3c7da26e395415 +size 656748 diff --git a/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/layout.json b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac218160924180b450378fb5639daf69058f8dd0 --- /dev/null +++ b/2023/MaLP_ Manipulation Localization Using a Proactive Scheme/layout.json @@ -0,0 +1,9444 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 109, + 103, + 485, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 103, + 485, + 120 + ], + "spans": [ + { + "bbox": [ + 109, + 103, + 485, + 120 + ], + "type": "text", + "content": "MaLP: Manipulation Localization Using a Proactive Scheme" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "spans": [ + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "content": "Vishal Asnani1, Xi Yin2, Tal Hassner2, Xiaoming Liu1 \n1*Michigan State University, 2Meta AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 172, + 442, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 172, + 442, + 185 + ], + "spans": [ + { + "bbox": [ + 151, + 172, + 442, + 185 + ], + "type": "text", + "content": "1{asnanivi, liuxm}@msu.edu, 2{yinxi, thassner}@meta.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "content": "Advancements in the generation quality of various Generative Models (GMs) has made it necessary to not only perform binary manipulation detection but also localize the modified pixels in an image. However, prior works termed as passive for manipulation localization exhibit poor generalization performance over unseen GMs and attribute modifications. To combat this issue, we propose a proactive scheme for manipulation localization, termed MaLP. We encrypt the real images by adding a learned template. If the image is manipulated by any GM, this added protection from the template not only aids binary detection but also helps in identifying the pixels modified by the GM. The template is learned by leveraging local and global-level features estimated by a two-branch architecture. We show that MaLP performs better than prior passive works. We also show the generalizability of MaLP by testing on 22 different GMs, providing a benchmark for future research on manipulation localization. Finally, we show that MaLP can be used as a discriminator for improving the generation quality of GMs. Our models/codes are available at www.github.com/vishal3477/pro_LOC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "type": "text", + "content": "We witness numerous Generative Models (GMs) [8, 9, 15, 17, 23-25, 28, 34, 39, 44, 50, 52, 60] being proposed to generate realistic-looking images. These GMs can not only generate an entirely new image [23, 24], but also perform partial manipulation of an input image [9, 9, 28, 60]. The proliferation of these GMs has made it easier to manipulate personal media for malicious use. Prior methods to combat manipulated media focus on binary detection [1, 2, 5, 11, 14, 30, 45, 48, 55, 56], using mouth movement, model parsing, hand-crafted features, etc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 287, + 677 + ], + "type": "text", + "content": "Recent works go one step further than detection, i.e. manipulation localization, which is defined as follows: given" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 440, + 308 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 440, + 308 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 440, + 308 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 440, + 308 + ], + "type": "image", + "image_path": "255dcfe857989d1c6bbc0bd011e3dbfca1999f291260392b4df061bf352637da.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 315, + 375, + 324 + ], + "lines": [ + { + "bbox": [ + 364, + 315, + 375, + 324 + ], + "spans": [ + { + "bbox": [ + 364, + 315, + 375, + 324 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 441, + 213, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 441, + 213, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 441, + 213, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 441, + 213, + 544, + 312 + ], + "type": "image", + "image_path": "1fdf50dd8a64232386fce9fd8f9d76bd337ad02ae971637e30a3ddf39b11d8dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 314, + 503, + 324 + ], + "lines": [ + { + "bbox": [ + 490, + 314, + 503, + 324 + ], + "spans": [ + { + "bbox": [ + 490, + 314, + 503, + 324 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 325, + 547, + 381 + ], + "lines": [ + { + "bbox": [ + 305, + 325, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 547, + 381 + ], + "type": "text", + "content": "Figure 1. (a) High-level idea of MaLP. We encrypt the image by adding a learnable template, which helps to estimate the fakeness map. (b) The cosine similarity (CS) between ground-truth and predicted fakeness maps for 22 unseen GMs. The performance is better for almost all GMs when using our proactive approach." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 397, + 546, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 546, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 546, + 505 + ], + "type": "text", + "content": "a partially manipulated image by a GM (e.g. STGAN [28] modifying hair colors of a face image), the goal is to identify which pixels are modified by estimating a fakeness map [21]. Identifying modified pixels helps to determine the severity of the fakeness in the image, and aid mediaforensics [11, 21]. Also, manipulation localization provides an understanding of the attacker's intent for modification which may further benefit identifying attack toolchains used [13]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "text", + "content": "Recent methods for manipulation localization [27,37,49] focus on estimating the manipulation mask of face-swapped images. They localize modified facial attributes by leveraging attention mechanisms [11], patch-based classifier [4], and face-parsing [21]. The main drawback of these methods is that they do not generalize well to GMs unseen in training. That is when the test images and training images are modified by different GMs, which will likely happen given the vast number of existing GMs. Thus, our work aims for a localization method generalizable to unseen GMs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "All aforementioned methods are based on a passive scheme as the method receives an image as is for estimation. Recently, proactive methods are gaining success for deepfake tasks such as detection [1], disruption [46, 57], and tagging [51]. These methods are considered proactive as they add different types of signals known as templates for encrypting the image before it is manipulated by a GM." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "*All data sourcing, modeling codes, and experiments were developed at Michigan State University. Meta did not obtain the data/codes or conduct any experiments in this work." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12343" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "This template can be one-hot encoding [51], adversarial perturbation [46], or a learnable noise [1], and is optimized to improve the performance of the defined tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 289 + ], + "type": "text", + "content": "Motivated by [1], we propose a Proactive scheme for MAnipulation Localization, termed as MaLP, in order to improve generalization. Specifically, MaLP learns an optimized template which, when added to real images, would improve manipulation localization, should they get manipulated. This manipulation can be done by an unseen GM trained on either in-domain or out-of-domain datasets. Furthermore, face manipulation may involve modifying facial attributes unseen in training (e.g. train on hair color modification yet test on gender modification). MaLP incorporates three modules that focus on encryption, detection, and localization. The encryption module selects and adds the template from the template set to the real images. These encrypted images are further processed by localization and detection modules to perform the respective tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 422 + ], + "type": "text", + "content": "Designing a proactive manipulation localization approach comes with several challenges. First, it is not straightforward to formulate constraints for learning the template unsupervisedly. Second, calculating a fakeness map at the same resolution as the input image is computationally expensive if the decision for each pixel has to be made. Prior works [4, 11] either down-sample the images or use a patch-wise approach, both of which result in inaccurate low-resolution fakeness maps. Lastly, the templates should be generalizable to localize modified regions from unseen GMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 591 + ], + "type": "text", + "content": "We design a two-branch architecture consisting of a shallow CNN network and a transformer to optimize the template during training. While the former leverages local-level features due to its shallow depth, the latter focuses on global-level features to better capture the affinity of the far apart regions. The joint training of both networks enables the MaLP to learn a better template, having embedded the information of both levels. During inference, the CNN network alone is sufficient to estimate the fakeness map with a higher inference efficiency. Compared to prior passive works [11, 21], MaLP improves the generalization performance on unseen GMs. We also demonstrate that MaLP can be used as a discriminator for fine-tuning conventional GMs to improve the quality of GM-generated images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 593, + 263, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 593, + 263, + 604 + ], + "spans": [ + { + "bbox": [ + 58, + 593, + 263, + 604 + ], + "type": "text", + "content": "In summary, we make the following contributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 605, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 641 + ], + "type": "text", + "content": "- We are the first to propose a proactive scheme for image manipulation localization, applicable to both face and generic images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "content": "- Our novel two-branch architecture uses both local and global level features to learn a set of templates in an unsupervised manner. The framework is guided by constraints based on template recovery, fakeness maps classification, and high cosine similarity between predicted and ground-truth fakeness maps." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 307, + 137, + 544, + 253 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 137 + ], + "type": "text", + "content": "Table 1. Comparison of our approach with prior works on manipulation localization and proactive schemes. We show the generalization ability of all works across different facial attribute modifications, unseen GMs trained on datasets with the same domain (in-domain) and different domains (out-domain). [Keys: Attr.: Attributes, Imp.: Improving, L.: Localization, D.: Detection]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 137, + 544, + 253 + ], + "lines": [ + { + "bbox": [ + 307, + 137, + 544, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 544, + 253 + ], + "type": "table", + "html": "
WorkSchemeTaskTemplateGeneralizationImp. GM
Attr.In-domainOut-domain
[51]ProactiveTagFix✓✓✘✘
[47]ProactiveDisruptLearn✓✘✘✘
[46]ProactiveDisruptLearn✓✓✘✘
[57]ProactiveDisruptLearn✓✘✘✘
[1]ProactiveD.Learn✘✓✓✘
[37]PassiveL. + D.-✘✘✘✘
[49]PassiveL. + D.-✘✘✘✘
[27]PassiveL. + D.-✘✓✘✘
[11]PassiveL. + D.-✓✓✘✘
[4]PassiveL. + D.-✘✓✘✘
[21]PassiveL. + D.-✓✓✘✘
MaLPProactiveL. + D.Learn✓✓✓✓
", + "image_path": "90ef4124fc4762a168c21ea4da4878d0e54582bcbe98a6dff7f50b86cf55d690.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 266, + 545, + 385 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 266, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 266, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 545, + 301 + ], + "type": "text", + "content": "- MaLP can be used as a plug-and-play discriminator module to fine-tune the generative model to improve the quality of the generated images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 303, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 545, + 385 + ], + "type": "text", + "content": "- Our method outperforms State-of-The-Art (SoTA) methods in manipulation localization and detection. Furthermore, our method generalizes well to GMs and modified attributes unseen in training. To facilitate the research of localization, we develop a benchmark for evaluating the generalization of manipulation localization, on images where the train and test GMs are different." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 399, + 392, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 392, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 392, + 411 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 626 + ], + "type": "text", + "content": "Manipulation Localization. Prior works tackle manipulation localization by adopting a passive scheme. Some of them focus on forgery attacks like removal, copymove, and splicing using multi-task learning [37]. Songsriin et al. [49] leverage facial landmarks [10] for manipulation localization. Li et al. [27] estimate the blended boundary for forged face-swap images. [11] uses an attention mechanism to leverage the relationship between pixels and [4] uses a patch-based classifier to estimate modified regions. Recently, Huang et al. [21] utilize gray-scale maps as ground truth for manipulation localization and leverage face parsing with an attention mechanism for prediction. The passive methods discussed above suffer from the generalization issue [4, 10, 11, 21, 37, 49] and estimate a low-resolution fakeness map [11] which is less accurate for the localization purpose. MaLP generalizes better to modified attributes and GMs unseen in training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Proactive Scheme. Recently, proactive schemes are developed for various tasks. Wang et al. [51] leverage the recovery of embedded one-hot encoding messages to perform deepfake tagging. A small perturbation is added onto the images by Segalis et al. [47] to disrupt the output of a GM. The same task is performed by Ruiz et al. [46] and Yeh et al. [57], both adding adversarial noise onto the in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12344" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "put images. Asnani et al. [1] propose a framework based on adding a learnable template to input images for generalized manipulation detection. Unlike prior works, which focus on binary detection, deepfake disruption, or tagging, our work emphasizes on manipulation localization. We show the comparison of our approach with prior works in Tab. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 288, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 288, + 278 + ], + "type": "text", + "content": "Manipulation Detection. The advancement in manipulation detection keeps reaching new heights. Prior works propose to combat deepfakes by exploiting frequency domain patterns [53], up-sampling artifacts [59], model parsing [2], hand-crafted features [35], lip motions [45], and self-attention [11]. Recent methods use self-blended images [48], real-time deviations [14], and self-supervised learning with adversarial training [5]. Finally, methods based on contrastive learning [56] and proactive scheme [1] have explicitly focused on generalized manipulation detection across unknown GMs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 293, + 164, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 164, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 164, + 306 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 315, + 171, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 171, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 171, + 327 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": "Passive Manipulation Localization Let " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\pmb{I}^{R}" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " be a set of real images that are manipulated by a GM " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " to output the set of manipulated images " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "G(\\pmb{I}^{R})" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": ". Prior passive works perform manipulation localization by estimating the fakeness map " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "M_{pred}" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " with the following objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 403, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 403, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 83, + 403, + 287, + 430 + ], + "type": "interline_equation", + "content": "\\left. \\min _ {\\theta_ {\\mathcal {E}}} \\left\\{\\sum_ {j} \\left(\\left| \\left| \\mathcal {E} \\left(G \\left(I _ {j} ^ {R}\\right); \\theta_ {\\mathcal {E}}\\right) - M _ {G T} \\right| \\right| _ {2}\\right) \\right\\}, \\right. \\tag {1}", + "image_path": "b3ebe42c54e20db6526531cac6e485f50320ef455e97c3d0fa789e550fe5172a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "type": "text", + "content": " denotes the passive framework with parameters " + }, + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{E}}" + }, + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "type": "inline_equation", + "content": "M_{GT}" + }, + { + "bbox": [ + 47, + 440, + 286, + 464 + ], + "type": "text", + "content": " is the ground-truth fakeness map." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 464, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 541 + ], + "type": "text", + "content": "To represent the fakeness map, some prior methods [11, 27, 49] choose a binary map by applying a threshold on the difference between the real and manipulated images. This is undesirable as the threshold selection is highly subjective and sensitive, leading to inaccurate fakeness maps. Therefore, we adopt the continuous gray-scale map for calculating the ground-truth fakeness maps [21], formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 549, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 549, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 96, + 549, + 287, + 562 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {G T} = \\operatorname {G r a y} \\left(\\left| \\boldsymbol {I} ^ {R} - G \\left(\\boldsymbol {I} ^ {R}\\right) \\right|\\right) / 2 5 5, \\tag {2}", + "image_path": "d26cb5711f8f9238ec8d3886c47df9c6fe4dff75d2d0d90ccb9b88ae510036fb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 571, + 247, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 247, + 583 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 247, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 48, + 571, + 247, + 583 + ], + "type": "inline_equation", + "content": "Gray(.)" + }, + { + "bbox": [ + 48, + 571, + 247, + 583 + ], + "type": "text", + "content": " converts the image to gray-scale." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": "Proactive Scheme Asnani et al. [1] define adding the template as a transformation " + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": " applied to images " + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "inline_equation", + "content": "I^R" + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": ", resulting in the encrypted images " + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(I^R)" + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": ". The added template acts as a signature of the defender and is learned during the training, aiming to improve the performance of the task at hand, e.g. detection, disruption, and tagging. Motivated by [1] that uses multiple templates, we have a set of " + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": " orthogonal templates " + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{S_1, S_2, \\ldots, S_n\\}" + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "inline_equation", + "content": "S_i \\in \\mathbb{R}^{128 \\times 128}" + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": ", for a real image " + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "inline_equation", + "content": "I_j^R \\in I^R" + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": ", transformation " + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 586, + 287, + 691 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 700, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 700, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 77, + 700, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right) = \\boldsymbol {I} _ {j} ^ {R} + \\boldsymbol {S} _ {i}, \\text {w h e r e} i \\in \\{1, 2, \\dots , n \\}. \\tag {3}", + "image_path": "24bccc3dac162ff8ce0f94ea2734f7d33a90300330319478f9326637d1a4b882.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "The templates are optimized such that adding them to the real images wouldn't result in a noticeable visual difference, yet helps manipulation localization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 110, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 110, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 305, + 110, + 545, + 157 + ], + "type": "text", + "content": "Proactive Manipulation Localization. Unlike the passive schemes [11, 21, 27, 37], we learn an optimal template set to help manipulation localization. For the encrypted images " + }, + { + "bbox": [ + 305, + 110, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\boldsymbol{I}^R)" + }, + { + "bbox": [ + 305, + 110, + 545, + 157 + ], + "type": "text", + "content": ", we formulate the estimation of the fakeness map as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 168, + 545, + 195 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 168, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 168, + 545, + 195 + ], + "type": "interline_equation", + "content": "\\left. \\min _ {\\theta_ {\\mathcal {E} _ {P}}, \\boldsymbol {S} _ {i}} \\left\\{\\sum_ {j} \\left(\\left| \\left| \\mathcal {E} _ {P} \\left(G \\left(\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right)\\right); \\theta_ {\\mathcal {E} _ {P}}\\right) - \\boldsymbol {M} _ {G T} \\right| \\right| _ {2}\\right) \\right\\}. \\right. \\tag {4}", + "image_path": "c7073194ef6b722cd21ca1681f4da18bdd3491ba66a8d9a9922437c95ed047a0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 205, + 544, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 544, + 217 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 307, + 205, + 544, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_P" + }, + { + "bbox": [ + 307, + 205, + 544, + 217 + ], + "type": "text", + "content": " is the proactive framework with parameters " + }, + { + "bbox": [ + 307, + 205, + 544, + 217 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{E}_P}" + }, + { + "bbox": [ + 307, + 205, + 544, + 217 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 218, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 218, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 545, + 264 + ], + "type": "text", + "content": "However, as the output of the GM has changed from images in set " + }, + { + "bbox": [ + 305, + 218, + 545, + 264 + ], + "type": "inline_equation", + "content": "G(\\mathbf{I}^R)" + }, + { + "bbox": [ + 305, + 218, + 545, + 264 + ], + "type": "text", + "content": " to images in set " + }, + { + "bbox": [ + 305, + 218, + 545, + 264 + ], + "type": "inline_equation", + "content": "G(\\mathcal{T}(\\mathbf{I}^R))" + }, + { + "bbox": [ + 305, + 218, + 545, + 264 + ], + "type": "text", + "content": ", in our proactive approach, the calculation of the ground-truth fakeness map shall be changed from Eq. 2 to the follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 347, + 274, + 545, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 274, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 347, + 274, + 545, + 287 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {G T} = \\operatorname {G r a y} \\left(\\left| \\boldsymbol {I} ^ {R} - G \\left(\\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right)\\right) \\right|\\right) / 2 5 5. \\tag {5}", + "image_path": "53e9caf7b20357c6f6be2340f2ca0669fc179eba629f79e27a9be24eece40c91.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 310, + 453, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 453, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 453, + 323 + ], + "type": "text", + "content": "3.2. Manipulation Localization" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 331, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 427 + ], + "type": "text", + "content": "MaLP consists of three modules: encryption, localization, and detection. The encryption module is used to encrypt the real images. The localization module estimates the fakeness map using a two-branch architecture. The detection module performs binary detection for the encrypted and manipulated images by recovering the template and using the classifier in the localization module. All three modules, as detailed next, are trained in an end-to-end manner." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 449, + 422, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 422, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 422, + 460 + ], + "type": "text", + "content": "3.2.1 Encryption Module" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 470, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 470, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 545, + 526 + ], + "type": "text", + "content": "Following the procedure in [1], we add a randomly selected learnable template from the template set to a real image. We control the strength of the added template using a hyperparameter " + }, + { + "bbox": [ + 305, + 470, + 545, + 526 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 470, + 545, + 526 + ], + "type": "text", + "content": ", which prevents the degradation of the image quality. The encryption process is summarised below:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 536, + 545, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 536, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 318, + 536, + 545, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}\\right) = \\boldsymbol {I} _ {j} ^ {R} + m \\times \\boldsymbol {S} _ {i} \\text {w h e r e} i = \\operatorname {R a n d} (1, 2, \\dots , n). \\tag {6}", + "image_path": "c5c7c74ade92c2c7bf513023643df174a2fbc31ab4fda1eeefcfe933769f98f3.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 560, + 517, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 517, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 517, + 571 + ], + "type": "text", + "content": "We select the value of " + }, + { + "bbox": [ + 306, + 560, + 517, + 571 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 306, + 560, + 517, + 571 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 306, + 560, + 517, + 571 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 306, + 560, + 517, + 571 + ], + "type": "text", + "content": " for our framework." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 572, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 572, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 305, + 572, + 545, + 606 + ], + "type": "text", + "content": "We optimize the template set by focusing on properties like low magnitude, orthogonality, and high-frequency content [1]. The properties are applied as constraints as follows." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 617, + 547, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 547, + 661 + ], + "type": "interline_equation", + "content": "J _ {T} = \\lambda_ {1} \\times \\sum_ {i = 1} ^ {n} \\left\\| \\boldsymbol {S} _ {i} \\right\\| _ {2} + \\lambda_ {2} \\times \\sum_ {\\substack {i, j = 1 \\\\ i \\neq j}} ^ {n} \\operatorname {CS} \\left(\\boldsymbol {S} _ {i}, \\boldsymbol {S} _ {j}\\right) + \\lambda_ {3} \\times \\left\\| \\mathcal {L} (\\mathfrak {F} (\\boldsymbol {S})) \\right\\| _ {2}, \\tag{7}", + "image_path": "f07fd3719f177f37c83dff723ae51e97a15c27182a28d729481ab0ab07085294.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "type": "text", + "content": "where CS is the cosine similarity, " + }, + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "type": "text", + "content": " is the low-pass filter, " + }, + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "type": "inline_equation", + "content": "\\mathfrak{F}" + }, + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "type": "text", + "content": " is the fourier transform, " + }, + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "type": "inline_equation", + "content": "\\lambda_1,\\lambda_2,\\lambda_3" + }, + { + "bbox": [ + 305, + 662, + 545, + 710 + ], + "type": "text", + "content": " are weights for losses of low magnitude, orthogonality and high-frequency content, respectively." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12345" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 539, + 273 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 539, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 539, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 539, + 273 + ], + "type": "image", + "image_path": "601d127cebb1148b646b0d3f2a73692e77c71504fbf39a5a4281021d4de09cca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 319 + ], + "type": "text", + "content": "Figure 2. The overview of MaLP. It includes three modules: encryption, localization, and detection. We randomly select a template from the template set and add it to the real image as encryption. The GM is used in inference mode to manipulate the encrypted image. The detection module recovers the added template for binary detection. The localization module uses a two-branch architecture to estimate the fakeness map. Lastly, we apply the classifier to the fakeness map to better distinguish them from each other. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 331, + 168, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 168, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 168, + 342 + ], + "type": "text", + "content": "3.2.2 Localization Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "text", + "content": "To design the localization module, we consider two desired properties: a larger receptive field for fakeness map estimation and high inference efficiency. A network with a large receptive field will consider far-apart regions in learning features for localization. Yet, large receptive fields normally come from deeper networks, implying slower inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 427, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 582 + ], + "type": "text", + "content": "In light of these properties, we design a two-branch architecture consisting of a shallow CNN network " + }, + { + "bbox": [ + 46, + 427, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_C" + }, + { + "bbox": [ + 46, + 427, + 287, + 582 + ], + "type": "text", + "content": " and a ViT transformer [12] " + }, + { + "bbox": [ + 46, + 427, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_T" + }, + { + "bbox": [ + 46, + 427, + 287, + 582 + ], + "type": "text", + "content": " (see Fig. 2). The intuition is to have one shallow branch to capture local features, and one deeper branch to capture global features. While training with both branches helps to learn better templates, in inference we only use the shallow branch for a higher efficiency. Specifically, the shallow CNN network has 10 layers which is efficient in inference but can only capture the local features due to small receptive fields. To capture global information, we adopt the ViT transformer. With the self-attention between the image patches, the transformer can estimate the fakeness map considering the far-apart regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 583, + 287, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 678 + ], + "type": "text", + "content": "Both the CNN and transformer are trained jointly to estimate a better template set, resembling the concept of the ensemble of networks. We empirically show that training both networks simultaneously results in higher performance than training either network separately. As the shallow CNN network is much faster in inference than the transformer, we use the transformer only in training to optimize the templates and switch off the transformer branch in inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 679, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 679, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 679, + 287, + 713 + ], + "type": "text", + "content": "To estimate the fakeness map, we leverage the supervision of the ground-truth fakeness map in Eq. 5. For fake images, we maximize the cosine similarity " + }, + { + "bbox": [ + 46, + 679, + 287, + 713 + ], + "type": "inline_equation", + "content": "(CS)" + }, + { + "bbox": [ + 46, + 679, + 287, + 713 + ], + "type": "text", + "content": " and struc" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "type": "text", + "content": "tural similarity index measure " + }, + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "type": "inline_equation", + "content": "(SS)" + }, + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "type": "text", + "content": " between the predicted and ground-truth fakeness map. However, the fakeness map should be a zero image for encrypted images. Therefore, we apply an " + }, + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "type": "text", + "content": " loss [21] to minimize the predicted map to zero for encrypted images. To maximize the difference between the two fakeness maps, we further minimize the cosine similarity between the predicted map from encrypted images and " + }, + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "type": "inline_equation", + "content": "M_{GT}" + }, + { + "bbox": [ + 304, + 331, + 547, + 421 + ], + "type": "text", + "content": ". The localization loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 430, + 547, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 547, + 506 + ], + "type": "interline_equation", + "content": "J _ {L} = \\left\\{ \\begin{array}{l l} \\left\\{\\lambda_ {4} \\times | | \\mathcal {E} _ {C / T} (\\boldsymbol {I}) | | _ {2} ^ {2} + \\right. & \\text {i f} \\boldsymbol {I} \\in \\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right) \\\\ \\left. \\lambda_ {5} \\times \\operatorname {C S} \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right) \\right\\} & \\\\ \\left\\{\\lambda_ {6} \\times (1 - \\operatorname {C S} \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right)) + \\right. & \\text {i f} \\boldsymbol {I} \\in G \\left(\\mathcal {T} \\left(\\boldsymbol {I} ^ {R}\\right)\\right) \\\\ \\left. \\lambda_ {7} \\times (1 - S S \\left(\\mathcal {E} _ {C / T} (\\boldsymbol {I}), M _ {G T}\\right)) \\right\\}. & \\end{array} \\right. \\tag {8}", + "image_path": "c8ed6f5ffd69938fbc804e54aea4d89503b77c3072c44bd32b45bcfa0ffd9566.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 506, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 602 + ], + "type": "text", + "content": "Finally, we have a classifier to make a binary decision of real vs. fake using the fakeness maps. This classifier is included in the framework to aid the detection module for binary detection of the input images, which will be discussed in Sec. 3.2.3. Another reason to have the classifier is to make the fakeness maps from encrypted and fake images to be distinguishable. We find that this design allows our training to converge much faster." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 624, + 416, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 416, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 416, + 636 + ], + "type": "text", + "content": "3.2.3 Detection Module" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "type": "text", + "content": "To leverage the added template for manipulation detection, we perform template recovery using encoder " + }, + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_E" + }, + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "type": "text", + "content": ". We follow the procedure in [1] to recover the added template from the encrypted images by maximizing the cosine similarity between " + }, + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "type": "inline_equation", + "content": "S_R" + }, + { + "bbox": [ + 304, + 647, + 547, + 713 + ], + "type": "text", + "content": ". However, for manipulated images, we minimize the cosine similarity between the recovered tem" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12346" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 73, + 286, + 310 + ], + "blocks": [ + { + "bbox": [ + 52, + 87, + 62, + 95 + ], + "lines": [ + { + "bbox": [ + 52, + 87, + 62, + 95 + ], + "spans": [ + { + "bbox": [ + 52, + 87, + 62, + 95 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 123, + 62, + 132 + ], + "lines": [ + { + "bbox": [ + 52, + 123, + 62, + 132 + ], + "spans": [ + { + "bbox": [ + 52, + 123, + 62, + 132 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 159, + 62, + 168 + ], + "lines": [ + { + "bbox": [ + 52, + 159, + 62, + 168 + ], + "spans": [ + { + "bbox": [ + 52, + 159, + 62, + 168 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 195, + 62, + 205 + ], + "lines": [ + { + "bbox": [ + 52, + 195, + 62, + 205 + ], + "spans": [ + { + "bbox": [ + 52, + 195, + 62, + 205 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 232, + 62, + 241 + ], + "lines": [ + { + "bbox": [ + 52, + 232, + 62, + 241 + ], + "spans": [ + { + "bbox": [ + 52, + 232, + 62, + 241 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 271, + 62, + 281 + ], + "lines": [ + { + "bbox": [ + 52, + 271, + 62, + 281 + ], + "spans": [ + { + "bbox": [ + 52, + 271, + 62, + 281 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 73, + 286, + 310 + ], + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 310 + ], + "type": "image", + "image_path": "567c396b3f3cbecd22bc5b75a218786205e9d437871399760f9c8abde3756f81.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 310, + 289, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 289, + 475 + ], + "type": "text", + "content": "Figure 3. Visualization of fakeness maps for faces and generic images showing generalization across unseen attribute modifications and GMs: (a) real image, (b) encrypted image, (c) manipulated image, (d) " + }, + { + "bbox": [ + 46, + 310, + 289, + 475 + ], + "type": "inline_equation", + "content": "M_{GT}" + }, + { + "bbox": [ + 46, + 310, + 289, + 475 + ], + "type": "text", + "content": ", (e) predicted fakeness map for encrypted images, and (f) predicted fakeness map for manipulated images. The first column shows the manipulation of (seen GM, seen attribute modification) i.e. (STGAN, bald). Following two columns show the manipulation of (seen GM, unseen attribute modification) i.e. (STGAN, [bangs, pale skin]. The fourth and fifth columns show manipulation of unseen GM, GauGAN for non-face images. The last column shows manipulation by unseen GM, DRIT. We see that the fakeness map of manipulated images is more bright and similar to " + }, + { + "bbox": [ + 46, + 310, + 289, + 475 + ], + "type": "inline_equation", + "content": "M_{GT}" + }, + { + "bbox": [ + 46, + 310, + 289, + 475 + ], + "type": "text", + "content": ", while the real fakeness map is more close to zero. We use the cmap as \"pink\" to better visualize the fakeness map. All face images come from SiWM-v2 data [18]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 487, + 265, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 265, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 265, + 499 + ], + "type": "text", + "content": "plate " + }, + { + "bbox": [ + 47, + 487, + 265, + 499 + ], + "type": "inline_equation", + "content": "(S_R)" + }, + { + "bbox": [ + 47, + 487, + 265, + 499 + ], + "type": "text", + "content": " and all the templates in the template set " + }, + { + "bbox": [ + 47, + 487, + 265, + 499 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 487, + 265, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 504, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 287, + 533 + ], + "type": "interline_equation", + "content": "J _ {R} = \\left\\{ \\begin{array}{l l} \\lambda_ {8} \\times (1 - \\operatorname {C S} (\\boldsymbol {S}, \\boldsymbol {S} _ {R})) & \\text {i f} x \\in \\mathcal {T} (\\boldsymbol {I} ^ {R}) \\\\ \\lambda_ {9} \\times \\left(\\sum_ {i = 1} ^ {n} (\\operatorname {C S} (\\boldsymbol {S} _ {i}, \\boldsymbol {S} _ {R}))\\right) & \\text {i f} x \\in G (\\mathcal {T} (\\boldsymbol {I} ^ {R})). \\end{array} \\right. \\tag {9}", + "image_path": "6262a9b897786623edda969f8ff2abba297497b3b68ffc678463df77e4d9648a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 536, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 289, + 658 + ], + "type": "text", + "content": "Further, we leverage our estimated fakeness map to help manipulation detection. As discussed in the previous section, we apply a classifier " + }, + { + "bbox": [ + 46, + 536, + 289, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 536, + 289, + 658 + ], + "type": "text", + "content": " to perform binary classification of the predicted fakeness map for the encrypted and fake images. The logits of the classifier are further combined with the cosine similarity of the recovered template. The averaged logits are back-propagated using the binary cross-entropy constraint. This not only improves the performance of manipulation detection but also helps manipulation localization. Therefore, we apply the binary cross entropy loss on the averaged logits as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 662, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 662, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 58, + 662, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} J _ {C} = \\lambda_ {1 0} \\times - \\sum_ {j} \\left\\{y _ {j}. \\log \\left[ \\frac {\\mathcal {C} \\left(\\boldsymbol {X} _ {j}\\right) + C S \\left(\\boldsymbol {S} _ {R} , \\boldsymbol {S}\\right)}{2} \\right] - \\right. \\tag {10} \\\\ (1 - y _ {j}). \\log \\left[ 1 - \\frac {\\mathcal {C} (\\boldsymbol {X} _ {j}) + C S (\\boldsymbol {S} _ {R} , \\boldsymbol {S})}{2} \\right] \\Bigg \\}, \\\\ \\end{array}", + "image_path": "c5499a738aead69eb90bdbe07eb6d68b3fc80488db4053c0b2589021662461c4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 83, + 545, + 132 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 543, + 82 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 543, + 82 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 543, + 82 + ], + "type": "text", + "content": "Table 2. Manipulation localization comparison with prior works." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 83, + 545, + 132 + ], + "lines": [ + { + "bbox": [ + 309, + 83, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 309, + 83, + 545, + 132 + ], + "type": "table", + "html": "
MethodLocalizationDetection
CS ↑PSNR ↑SSIM ↑Accuracy ↑EER ↓AUC ↑
[11]0.62306.2140.21780.99750.00500.9975
[21]0.883122.8900.78760.99450.00770.9998
MaLP0.939423.0200.73120.99910.00721.0
", + "image_path": "814a3b097cbeaa86e41d07d701e1874fb1b699a4dc141fd5233a28f6ffb6c5a8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "text", + "content": " is the class label, " + }, + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\pmb{S}" + }, + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "inline_equation", + "content": "S_{R}" + }, + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "text", + "content": " are the added and recovered template respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 178, + 545, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 200 + ], + "type": "text", + "content": "Our framework is trained in an end-to-end manner with the overall loss function as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 207, + 545, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 207, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 375, + 207, + 545, + 219 + ], + "type": "interline_equation", + "content": "J = J _ {T} + J _ {R} + J _ {C} + J _ {L}. \\tag {11}", + "image_path": "5687659f53bf2ab6941f24bcf840740dbacfdece6e1b29152e94e1bb9613cf79.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 225, + 450, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 225, + 450, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 225, + 450, + 236 + ], + "type": "text", + "content": "3.3. MaLP as A Discriminator" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 243, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 326 + ], + "type": "text", + "content": "One application of MaLP is to leverage our proposed localization module as a discriminator for improving the quality of the manipulated images. MaLP performs binary classification by estimating a fakeness map, which can be used as an objective. This results in output images being resilient to manipulation localization, thereby lowering the performance of our framework." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 326, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 393 + ], + "type": "text", + "content": "We use MaLP as a plug-and-play discriminator to improve image generation quality through fine-tuning pretrained GMs. The generation quality and manipulation localization will compete head-to-head, resulting in a better quality of the manipulated images. We define the fine-tuning objective for the GM as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 545, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 545, + 462 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta_ {G}} \\max _ {\\theta_ {M a L P}, \\boldsymbol {S} _ {i}} \\left\\{\\sum_ {j} \\left(\\mathbb {E} \\left[ l o g \\left(\\mathcal {E} _ {M a L P} \\left(\\mathcal {T} \\left(\\boldsymbol {I} _ {j} ^ {R}\\right)\\right); \\theta_ {M a L P}\\right) \\right] + \\right. \\right. \\\\ \\left. \\mathbb {E} \\left[ 1 - \\log \\left(\\mathcal {E} _ {M a L P} \\left(G \\left(\\mathcal {T} \\left(I _ {j} ^ {R}; \\boldsymbol {S} _ {i}\\right); \\theta_ {G}\\right); \\theta_ {M a L P}\\right)\\right) \\right]\\right) \\Bigg \\}. \\tag {12} \\\\ \\end{array}", + "image_path": "dd5ab3c8518e69dc2e5268f658848e269873168336a9ec9ff894ff9d89d79b81.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 468, + 537, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 537, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 537, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 468, + 537, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{MaLP}" + }, + { + "bbox": [ + 306, + 468, + 537, + 481 + ], + "type": "text", + "content": " is our framework with " + }, + { + "bbox": [ + 306, + 468, + 537, + 481 + ], + "type": "inline_equation", + "content": "\\theta_{MaLP}" + }, + { + "bbox": [ + 306, + 468, + 537, + 481 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 491, + 388, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 388, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 388, + 503 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 510, + 422, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 422, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 422, + 523 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "content": "Settings Following the settings in [21], we use STGAN [28] to manipulate images from CelebA [31] dataset and train on bald facial attribute modification. In order to evaluate the generalization of image manipulation localization, we construct a new benchmark that consists of 200 real images of 22 different GMs on various data domains. The real images are chosen from the dataset on which the GM is trained on. The list of GMs, datasets and implementation details are provided in the supplementary." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Metrics We use cosine similarity (CS), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) as adopted by [21] to evaluate manipulation localization since the GT is a continuous map. For binary detection, we use the area under the curve (AUC), equal error rate (EER), and accuracy score [21]." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12347" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 103, + 287, + 142 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "text", + "content": "Table 3. Comparison of localization performance across unseen GMs and attribute modifications. We train on STGAN bald/smile attribute modification and test on AttGAN/StyleGAN." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 103, + 287, + 142 + ], + "lines": [ + { + "bbox": [ + 50, + 103, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 287, + 142 + ], + "type": "table", + "html": "
MethodCosine similarity ↑(AttGAN)Cosine similarity ↑ (StyleGAN)
BaldBlack HairEyeglassesSmileAgeGender
[21]0.81410.69320.69500.61760.31410.6470
MaLP0.82010.79400.85570.81590.82550.8016
", + "image_path": "eeb2ff54ce3edacd16014f59d9e93c31efb4cbab2cbdf95cae18677a4d3282d4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 158, + 197, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 197, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 197, + 171 + ], + "type": "text", + "content": "4.2. Comparison with Baselines" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 177, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 288, + 369 + ], + "type": "text", + "content": "We compare our results with [21] and [11] for manipulation localization. The results are shown in Tab. 2. MaLP has higher cosine similarity and similar PSNR for localization compared to [21]. However, we observe a dip in SSIM. This might be because of the degradation caused by adding our template to the real images and then performing the manipulation. The learned template helps localize the manipulated regions better, as demonstrated by cosine similarity, but the degradation affects SSIM and PSNR. We also compare the performance of real vs. fake binary detection. As expected, our proposed proactive approach outperforms the passive methods with a perfect AUC and near-perfect accuracy. We also show visual examples of fakeness maps for images modified by unseen GMs in Fig. 3. MaLP is able to estimate the fakeness map for unseen modifications and GMs across face/generic image datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 377, + 139, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 139, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 139, + 389 + ], + "type": "text", + "content": "4.3. Generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 554 + ], + "type": "text", + "content": "Across Attribute Modifications Following the settings in [21], we evaluate the performance of MaLP across unseen attribute modifications. Specifically, we train MaLP using STGAN with the bald/smile attribute modification and test it on unseen attribute modifications with unseen GMs: AttGAN/StyleGAN. As shown in Tab. 3, MaLP is more generalizable to all unseen attribute modifications. Furthermore, AttGAN shares the high-level architecture with STGAN but not with StyleGAN. We observe a significant increase in localization performance for StyleGAN compared to AttGAN. This shows that, unlike our MaLP, passive works perform much worse if the test GM doesn't share any similarity with the training GM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "type": "text", + "content": "Across GMs Although [21] tries to show generalization across unseen GMs; it is limited by the GMs within the same domain of the dataset used in training. We propose a benchmark to evaluate the generalization performance for future manipulation localization works that consists of 22 different GMs in various domains. We select GMs that are publicly released and can perform partial manipulation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "As no open-source code base is available for [21], we train a passive approach using a ResNet50 [19] network to estimate the fakeness map as the baseline for comparison. Further, we compare our approach with [4, 11]. Although [4, 11] estimate a fakeness map, it has at least " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " lower resolution compared to input images due to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 543, + 129 + ], + "blocks": [ + { + "bbox": [ + 310, + 92, + 321, + 103 + ], + "lines": [ + { + "bbox": [ + 310, + 92, + 321, + 103 + ], + "spans": [ + { + "bbox": [ + 310, + 92, + 321, + 103 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 72, + 543, + 129 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 543, + 129 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 543, + 129 + ], + "type": "image", + "image_path": "ae2e28d09a67750ebd9891c7e909bd33228f3a8ceda7eef3d3d3711fb19c0658.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 327, + 141, + 543, + 198 + ], + "blocks": [ + { + "bbox": [ + 347, + 133, + 363, + 140 + ], + "lines": [ + { + "bbox": [ + 347, + 133, + 363, + 140 + ], + "spans": [ + { + "bbox": [ + 347, + 133, + 363, + 140 + ], + "type": "text", + "content": "0.58" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 133, + 416, + 140 + ], + "lines": [ + { + "bbox": [ + 400, + 133, + 416, + 140 + ], + "spans": [ + { + "bbox": [ + 400, + 133, + 416, + 140 + ], + "type": "text", + "content": "0.62" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 133, + 470, + 140 + ], + "lines": [ + { + "bbox": [ + 454, + 133, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 454, + 133, + 470, + 140 + ], + "type": "text", + "content": "0.61" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 133, + 524, + 140 + ], + "lines": [ + { + "bbox": [ + 507, + 133, + 524, + 140 + ], + "spans": [ + { + "bbox": [ + 507, + 133, + 524, + 140 + ], + "type": "text", + "content": "0.59" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 163, + 321, + 174 + ], + "lines": [ + { + "bbox": [ + 309, + 163, + 321, + 174 + ], + "spans": [ + { + "bbox": [ + 309, + 163, + 321, + 174 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 327, + 141, + 543, + 198 + ], + "lines": [ + { + "bbox": [ + 327, + 141, + 543, + 198 + ], + "spans": [ + { + "bbox": [ + 327, + 141, + 543, + 198 + ], + "type": "image", + "image_path": "c13de7d103ce68c48e9ac785738e75663e8b59909114dbcc440105ee2af30208.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 327, + 212, + 543, + 265 + ], + "blocks": [ + { + "bbox": [ + 347, + 202, + 362, + 211 + ], + "lines": [ + { + "bbox": [ + 347, + 202, + 362, + 211 + ], + "spans": [ + { + "bbox": [ + 347, + 202, + 362, + 211 + ], + "type": "text", + "content": "0.42" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 203, + 416, + 211 + ], + "lines": [ + { + "bbox": [ + 400, + 203, + 416, + 211 + ], + "spans": [ + { + "bbox": [ + 400, + 203, + 416, + 211 + ], + "type": "text", + "content": "0.45" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 202, + 471, + 211 + ], + "lines": [ + { + "bbox": [ + 454, + 202, + 471, + 211 + ], + "spans": [ + { + "bbox": [ + 454, + 202, + 471, + 211 + ], + "type": "text", + "content": "0.40" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 203, + 524, + 211 + ], + "lines": [ + { + "bbox": [ + 507, + 203, + 524, + 211 + ], + "spans": [ + { + "bbox": [ + 507, + 203, + 524, + 211 + ], + "type": "text", + "content": "0.44" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 327, + 212, + 543, + 265 + ], + "lines": [ + { + "bbox": [ + 327, + 212, + 543, + 265 + ], + "spans": [ + { + "bbox": [ + 327, + 212, + 543, + 265 + ], + "type": "image", + "image_path": "9d462d6840d020c7f4ea6ebb61189e6f10db23bffb6cffceec47b00e6715b3af.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 266, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 306, + 266, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 545, + 364 + ], + "type": "text", + "content": "Figure 4. Visualization of (a) encrypted images, (b) manipulated images before fine-tuning, and (c) manipulated images after fine-tuning. The generation quality has improved after we fine-tune the GM using our framework as a discriminator. The artifacts in the images have been reduced, and the face skin color is less pale and more realistic. We also specify the cosine similarity of the predicted fakeness map and " + }, + { + "bbox": [ + 306, + 266, + 545, + 364 + ], + "type": "inline_equation", + "content": "M_{GT}" + }, + { + "bbox": [ + 306, + 266, + 545, + 364 + ], + "type": "text", + "content": ". The GM is able to decrease the performance of our framework after fine-tuning. All face images come from SiWM-v2 data [18]." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 380, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 440 + ], + "type": "text", + "content": "their patch-based methodology. For a fair comparison, we rescale their predicted fakeness maps to the resolution of " + }, + { + "bbox": [ + 305, + 380, + 545, + 440 + ], + "type": "inline_equation", + "content": "M_{GT}" + }, + { + "bbox": [ + 305, + 380, + 545, + 440 + ], + "type": "text", + "content": ". We compare the cosine similarity in Tab. 4. MaLP is able to outperform all the baselines for almost all GMs, which proves the effectiveness of the proactive scheme." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": "We also evaluate the performance of " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_C" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": " for high-resolution images. For encryption, we upsample the " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "128\\times 128" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": " template to the original resolution of images and evaluate " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_C" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": " on these higher resolution encrypted images. We observe similar performance of " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_C" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": " for higher resolution images in Tab. 4, proving the versatility of " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_C" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": " to image sizes." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 526, + 454, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 454, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 454, + 539 + ], + "type": "text", + "content": "4.4. Improving Quality of GMs" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 546, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 712 + ], + "type": "text", + "content": "We fine-tune the GM into fooling our framework to generate a fakeness map as a zero image. This process results in better-quality images. Initially, we train MaLP with the pretrained GM so that it can perform manipulation localization. Next, to fine-tune the GM, we adopt two strategies. First, we freeze MaLP and fine-tune the GM only. Second, we fine-tune both the GM and the MaLP but update the MaLP with a lower learning rate. The result for fine-tuning StarGAN is shown in Tab. 5. We observe that for both strategies, MaLP reduces the FID score of StarGAN. We also show some visual examples in Fig. 4. We see that the images are of better quality after fine-tuning, and many artifacts in the images manipulated by the pretrained model are removed." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12348" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 106, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 4. Benchmark for manipulation localization across 22 different unseen GMs, showing cosine similarity between ground-truth and predicted fakeness maps. We compare our proactive vs. passive baselines [4, 11, 19] approach to highlight the generalization ability of our MaLP. We scale the images to " + }, + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "128^{2}" + }, + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "text", + "content": " for \"sc.\" and keep the resolution as is for \"no sc\"." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 106, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 48, + 106, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 545, + 216 + ], + "type": "table", + "html": "
GMSEAN [62]StarGAN [8]CycleGAN [60]GauGAN [39]Con.Enc. [41]StarGAN2 [9]ALAE [42]BiGAN [61]AuGAN [60]GANim [43]DRGAN [50]ILVR [7]
Resolution256212822562256212822562256225623402128212822562
ResNet50 [19]0.86140.75130.67150.76150.86390.81960.67660.65140.66390.68710.80290.7018
[4]0.75140.71110.79810.80160.78940.70260.71560.72170.75160.76120.71150.7851
[11]0.79610.78870.80140.82560.85410.70340.75490.78050.72320.84570.72390.7854
MaLP (sc.)0.93760.87180.91280.92510.85460.88360.91920.91810.88940.96250.75120.8003
MaLP (no sc.)0.92580.87180.92450.91250.85460.87850.91410.92290.91490.96250.75120.8359
GMDRIT [26]Pix2Pix [22]CounGAN [38]DualGAN [58]ESRGAN [54]UNIT [29]MUNIT [20]ColGAN [36]GDWCT [6]RePaint [32]Average
Resolution256225621282256210242512 × 931256 × 512128212822562-
ResNet50 [19]0.74860.67190.72930.73650.87030.70830.66010.75960.83500.65120.7401
[4]0.78710.77690.81460.75690.81680.80640.67880.76100.86910.75160.7645
[11]0.81200.77810.85590.77210.82410.80860.70970.78740.88790.76960.7903
MaLP (sc.)0.88670.89150.93260.88720.83480.82140.75650.80960.93840.81020.8725
MaLP (no sc.)0.90840.87140.93260.84320.87430.83910.78600.80960.93840.82900.8773
", + "image_path": "25112bda1f9ecf6a0e0b876ce1b34db1e6bdfe20204d52a6a9ac81997e3e927e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 100, + 251, + 231, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 229, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 47, + 229, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 287, + 251 + ], + "type": "text", + "content": "Table 5. FID score comparison for the application of our approach as a discriminator for improving the generation quality of the GM" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 251, + 231, + 291 + ], + "lines": [ + { + "bbox": [ + 100, + 251, + 231, + 291 + ], + "spans": [ + { + "bbox": [ + 100, + 251, + 231, + 291 + ], + "type": "table", + "html": "
StateFine-tuneStarGAN FID ↓
Before-60.49
AfterG51.91
G + MaLP52.07
", + "image_path": "0ce6b7fc05bfeeeb3af8f4462d4f1758712437f8b48e344a3a81c23e3f82d880.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 316, + 286, + 387 + ], + "blocks": [ + { + "bbox": [ + 47, + 293, + 286, + 315 + ], + "lines": [ + { + "bbox": [ + 47, + 293, + 286, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 286, + 315 + ], + "type": "text", + "content": "Table 6. Comparison with prior binary detection works. [Keys: D.M.: Detection module, L.M.: Localization module]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 316, + 286, + 387 + ], + "lines": [ + { + "bbox": [ + 50, + 316, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 50, + 316, + 286, + 387 + ], + "type": "table", + "html": "
MethodTrain GMSet sizeTest GM Average precision (%)↑
CycleGANStarGANGauGAN
Nataraj et al. [35]CycleGAN-10088.2056.20
Zhang et al. [59]AutoGAN-10010061.00
Wang et al. [53]ProGAN-84.0010067.00
Asnani et al. [1]STGAN194.0010069.50
MaLP (D.M.)STGAN194.1010069.61
MaLP (D.M. + L.M.)STGAN194.3010072.16
", + "image_path": "f00ce0705f6657ef1d4d991066087b2a9c8008da761c349f384e38155af78e46.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 399, + 163, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 163, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 163, + 412 + ], + "type": "text", + "content": "4.5. Other Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 420, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 576 + ], + "type": "text", + "content": "Binary Detection We compare with prior proactive and passive approaches for binary manipulation detection [1, 35, 53, 59]. We adopt the evaluation protocol in [1] to test on images manipulated by CycleGAN, StarGAN, and GauGAN. We are able to perform similar to [1] as shown in Tab. 6. We have better average precision than passive schemes and generalize well to GMs unseen in training. We also conduct experiments to see whether localization can help binary detection to improve the performance, as mentioned in Sec. 3.2.3. The combined predictions' results are better than just using the detection module as shown in Tab. 6. This is intuitive as the localization module provides extra information, thereby increasing the performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 578, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 686 + ], + "type": "text", + "content": "Inference Speed We compare the inference speed of our MaLP against prior work. [21] uses Deeplabv3-ResNet101 model from PyTorch [40]. In our generalization benchmark shown in Sec. 4.3, we use the ResNet50 model for training the passive baseline. The inference speed per image on an NVIDIA K80 GPU for Deeplabv3-ResNet101, ResNet50, and MaLP are 75.61, 52.66, and " + }, + { + "bbox": [ + 46, + 578, + 287, + 686 + ], + "type": "inline_equation", + "content": "29.26~\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 578, + 287, + 686 + ], + "type": "text", + "content": ", respectively. MaLP takes less than half the inference time compared to [21] due to our shallow CNN network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Adversarial Attack Our framework can be considered as an adversarial attack on real images to aid manipula-" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 240, + 545, + 315 + ], + "blocks": [ + { + "bbox": [ + 326, + 229, + 524, + 239 + ], + "lines": [ + { + "bbox": [ + 326, + 229, + 524, + 239 + ], + "spans": [ + { + "bbox": [ + 326, + 229, + 524, + 239 + ], + "type": "text", + "content": "Table 7. Comparison with adversarial attack methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 240, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 309, + 240, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 309, + 240, + 545, + 315 + ], + "type": "table", + "html": "
MethodSchemeCosine similarity↑
BaldBlack HairEyeglasses
Huang et al. [21]Passive0.81410.69320.6950
PGD [33]Proactive0.80510.75140.8358
FGSM [16]Proactive0.81110.78820.8512
CW [3]Proactive0.80140.83440.8405
MaLPProactive0.82010.79400.8557
", + "image_path": "469401c11a970ce163e9c6c134b7a1df1705851736fa05095285dd8758e607a8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 330, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 546, + 521 + ], + "type": "text", + "content": "tion localization. Therefore, it is vital to contrast the performance between our approach and classic adversarial attacks. For this purpose, we perform experiments that make use of adversarial attacks, namely PGD [33], CW [3], and FGSM [16] to guide the learning of the added template. We evaluate on unseen GM AttGAN for unseen attribute modifications. We show the performance comparison in Tab. 7. MaLP has higher cosine similarity across some unseen facial attribute modifications compared to adversarial attacks. This can be explained as the adversarial attack methods being over-fitted to training parameters (data, target network etc.). Therefore, if the testing data is changed with unseen attribute modifications by GMs, the performance of adversarial attacks degrades. Further, these attacks are analogous to our MaLP as a proactive scheme which, in general, have better performance than passive works." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "type": "text", + "content": "Model Robustness Against Degradations It is necessary to test the robustness of our proposed approach against various types of real-world image editing degradations. We evaluate our method on degradations applied during testing as adopted by [21], which include JPEG compression, blurring, adding noise, and low resolution. The results are shown in Fig. 5. Our proposed MaLP is more robust to real-world degradations than passive schemes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 632, + 373, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 373, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 373, + 643 + ], + "type": "text", + "content": "4.6. Ablations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Two-branch Architecture As described in Sec. 3.2.2, MaLP adopts a two-branch architecture to predict the fairness map using the local-level and global-level features, which are estimated by a shallow CNN and a transformer. We ablate by training each branch separately to show the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12349" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 286, + 149 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 286, + 149 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 286, + 149 + ], + "type": "image", + "image_path": "996afe55f94d70e619084c2b3567fe3017ddd9a25bf2498576d48dfa3e1218b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 151, + 286, + 172 + ], + "lines": [ + { + "bbox": [ + 47, + 151, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 151, + 286, + 172 + ], + "type": "text", + "content": "Figure 5. Comparison of our approach's robustness against common image editing degradations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 211, + 276, + 271 + ], + "blocks": [ + { + "bbox": [ + 47, + 177, + 286, + 209 + ], + "lines": [ + { + "bbox": [ + 47, + 177, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 286, + 209 + ], + "type": "text", + "content": "Table 8. Ablation of two-branch architecture. CNN is a shallow network with 10 layers. Training each branch separately has worse localization results than combining them." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 211, + 276, + 271 + ], + "lines": [ + { + "bbox": [ + 53, + 211, + 276, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 211, + 276, + 271 + ], + "type": "table", + "html": "
Network trainedCosine similarity ↑Accuracy ↑
CNN only0.89610.9801
Transformer only0.88480.9856
CNN + ResNet500.86470.9512
CNN + Transformer0.93940.9981
", + "image_path": "30e144141c2d7e555280f6154fa09099ba063c1b2ac337cb27859ec58307902f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 286, + 286, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 286, + 392 + ], + "type": "text", + "content": "effectiveness of combining them. As shown in Tab. 8, if the individual network is trained separately, the performance is lower than the two-branch architecture. Next, to show the efficacy of the transformer, we use a ResNet50 network in place of the transformer to predict the fakeness map. We observe that the performance is even worse than using only the transformer. ResNet50 lacks the added advantage of self-attention in the transformer, which estimates the global-level features much better than a CNN network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "type": "text", + "content": "Constraints MaLP leverages different constraints to estimate the fakeness map using an optimized template. We perform an ablation by removing each constraint separately, showing the importance of every constraint. Tab. 9 shows the cosine similarity for localization and accuracy for detection. Removing either the classifier or recovery constraint results in lower detection performance. This is expected as we leverage logits from both " + }, + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "type": "inline_equation", + "content": "\\varepsilon_{E}" + }, + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "type": "text", + "content": ", and removing the constraint for one network will hurt the logits of the other network. Furthermore, removing the template constraint results in a decrease in performance. Although the gap is small, the template is not properly optimized to have lower magnitude and high-frequency content." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 554, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 286, + 661 + ], + "type": "text", + "content": "Removing the localization constraint and just applying a " + }, + { + "bbox": [ + 46, + 554, + 286, + 661 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 554, + 286, + 661 + ], + "type": "text", + "content": " loss for supervising fakeness maps result in a significant performance drop for both localization and detection, showing the necessity of this constraint. Finally, we show the importance of a learnable template by not optimizing it during the training of MaLP. This hurts the performance a lot, similar to removing the localization constraint. Both these observations prove that our localization constraint and learnable template are important components of MaLP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Template Set Size We perform an ablation to vary the size of the template set " + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": ". Having multiple templates will improve security if an attacker tries to reverse engineer the template from encrypted images. The results are shown in" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 82, + 542, + 162 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 541, + 81 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 541, + 81 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 541, + 81 + ], + "type": "text", + "content": "Table 9. Ablation of constraints used in training our framework." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 82, + 542, + 162 + ], + "lines": [ + { + "bbox": [ + 308, + 82, + 542, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 82, + 542, + 162 + ], + "type": "table", + "html": "
Constraint removedCosine similarity ↑Accuracy ↑
Classifier constraint JC0.93190.9814
Template constraint JT0.91430.9803
Localization constraint JL0.88140.9539
Recovery constraint JR0.92060.9780
Fixed template0.88870.9514
Nothing (MaLP)0.93940.9991
", + "image_path": "dce03e1c133c93006fe6f2fc36a4b1584f906d897f9eb84bb5cea6cdceb376ca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 171, + 425, + 235 + ], + "blocks": [ + { + "bbox": [ + 307, + 171, + 425, + 235 + ], + "lines": [ + { + "bbox": [ + 307, + 171, + 425, + 235 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 425, + 235 + ], + "type": "image", + "image_path": "a883284c91c6560a0e7db0000cfbb0b571e604bc747af57c583e964b60f02ef4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 236, + 378, + 243 + ], + "lines": [ + { + "bbox": [ + 369, + 236, + 378, + 243 + ], + "spans": [ + { + "bbox": [ + 369, + 236, + 378, + 243 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 171, + 543, + 236 + ], + "blocks": [ + { + "bbox": [ + 425, + 171, + 543, + 236 + ], + "lines": [ + { + "bbox": [ + 425, + 171, + 543, + 236 + ], + "spans": [ + { + "bbox": [ + 425, + 171, + 543, + 236 + ], + "type": "image", + "image_path": "6af8e81bd60cd8193c3050d37bb2df239eb871a1cf7b6760ab9b67260b2d09fe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 236, + 494, + 243 + ], + "lines": [ + { + "bbox": [ + 485, + 236, + 494, + 243 + ], + "spans": [ + { + "bbox": [ + 485, + 236, + 494, + 243 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 243, + 544, + 265 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 544, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 544, + 265 + ], + "type": "text", + "content": "Figure 6. Ablation study on hyperparameters used in our framework: set size and signal strength." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": "Fig. 6 (a). The cosine similarity takes a dip when the set size is increased. We also observe the inter-template cosine similarity, which remains constant at a high value of around 0.74 for all templates. This is against the findings of [1]. Localization is a more challenging task than binary detection. Therefore, it is less likely to find different templates for our MaLP in the given feature space compared to [1]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "content": "Signal Strength We vary the template strength hyperparameter " + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "content": " to find its impact on the performance. As shown in Fig. 6 (b), the cosine similarity increases as we increase the strength of the added template. However, this comes with the lower visual quality of the encrypted images if the template strength is increased. The performance doesn't vary much after " + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "inline_equation", + "content": "m = 30\\%" + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "content": ", which we use for MaLP." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 461, + 378, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 461, + 378, + 474 + ], + "spans": [ + { + "bbox": [ + 306, + 461, + 378, + 474 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": "This paper focuses on manipulation localization using a proactive scheme (MaLP). We propose to improve the generalization of manipulation localization across unseen GM and facial attribute modifications. We add an optimal template onto the real images and estimate the fakeness map via a two-branch architecture using local and global-level features. MaLP outperforms prior works with much stronger generalization capabilities, as demonstrated by our proposed evaluation benchmark with 22 different GMs in various domains. We show an application of MaLP in finetuning GMs to improve generation quality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Limitations First, the number of publicly available GMs is limited. More thorough testing on many different GMs might give more insights into the problem of generalizable manipulation localization. Second, we show that our MaLP can be used to fine-tune the GMs to improve image generation quality. However, it is based on the pretrained GM. Using our method to train a GM from scratch can be an interesting direction to explore in the future." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12350" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Vishal Asnani, Xi Yin, Tal Hassner, Sijia Liu, and Xiaoming Liu. Proactive image manipulation detection. In CVPR, 2022. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Reverse engineering of generative models: Inferring model hyperparameters from generated images. arXiv preprint arXiv:2106.07873, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 191 + ], + "type": "text", + "content": "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In SSP, 2017. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "text", + "content": "[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "type": "text", + "content": "[5] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "type": "text", + "content": "[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In CVPR, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "type": "text", + "content": "[7] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 352, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 395 + ], + "type": "text", + "content": "[8] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, 2018. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 397, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 429 + ], + "type": "text", + "content": "[9] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 431, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 287, + 453 + ], + "type": "text", + "content": "[10] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "type": "text", + "content": "[11] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In CVPR, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 488, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 542 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "text", + "content": "[13] Bruce Draper. Reverse engineering of deceptions (red). https://www.darpa.mil/program/reverse-engineering-of-deceptions.1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "text", + "content": "[14] Candice R Gerstner and Hany Farid. Detecting real-time deep-fake videos using active illumination. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 654 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[16] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[17] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vec" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "tor quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[18] Xiao Guo, Yaojie Liu, Anil Jain, and Xiaoming Liu. Multidomain learning for updating face anti-spoofing models. In ECCV, 2022. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 164, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 195 + ], + "type": "text", + "content": "[20] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In ECCV, 2018. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "text", + "content": "[21] Yihao Huang, Felix Juefei-Xu, Qing Guo, Yang Liu, and Geguang Pu. FakeLocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information Forensics and Security, 17:2657-2672, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[22] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "type": "text", + "content": "[23] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "type": "text", + "content": "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "text", + "content": "[25] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "text", + "content": "[26] Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Kumar Singh, and Ming-Hsuan Yang. Diverse imaged-to-image translation via disentangled representations. In ECCV, 2018. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 434, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 465 + ], + "type": "text", + "content": "[27] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, 2020. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 468, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 509 + ], + "type": "text", + "content": "[28] Ming Liu, Yukang Ding, Min Xia, Xiao Liu, Errui Ding, Wangmeng Zuo, and Shilei Wen. STGAN: A unified selective transfer network for arbitrary image attribute editing. In CVPR, 2019. 1, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 511, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 533 + ], + "type": "text", + "content": "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In NeurIPS, 2017. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "text", + "content": "[30] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. PSCC-Net: Progressive spatio-channel correlation network for image manipulation detection and localization. In arXiv preprint arXiv:2103.10596, 2021. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 580, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 610 + ], + "type": "text", + "content": "[31] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "text", + "content": "[32] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 7" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[33] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 7" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[34] Safa C. Medin, Bernhard Egger, Anoop Cherian, Ye Wang, Joshua B. Tenenbaum, Xiaoming Liu, and Tim K. Marks." + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12351" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "MOST-GAN: 3D morphable StyleGAN for disentangled face image manipulation. In AAAI, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 152 + ], + "type": "text", + "content": "[35] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, BS Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, and Amit K Roy-Chowdhury. Detecting GAN generated fake images using co-occurrence matrices. Electronic Imaging, 2019:532-1, 2019. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 185 + ], + "type": "text", + "content": "[36] Kamyar Nazeri, Eric Ng, and Mehran Ebrahimi. Image colorization using generative adversarial networks. In AMDO, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 187, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 187, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 187, + 287, + 230 + ], + "type": "text", + "content": "[37] Huy H Nguyen, Fuming Fang, Junichi Yamagishi, and Isao Echizen. Multi-task learning for detecting and segmenting manipulated facial images and videos. In BTAS, 2019. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 233, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 255 + ], + "type": "text", + "content": "[38] Ori Nizan and Ayellet Tal. Breaking the cycle - colleagues are all you need. In CVPR, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 257, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 287, + 289 + ], + "type": "text", + "content": "[39] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. GauGAN: semantic image synthesis with spatially adaptive normalization. In ACM, 2019. 1, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 291, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 345 + ], + "type": "text", + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "type": "text", + "content": "[41] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 413 + ], + "type": "text", + "content": "[42] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In CVPR, 2020. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "text", + "content": "[43] Albert Pumarola, Antonio Agudo, Aleix M Martinez, Alberto Sanfeliu, and Francesc Moreno-Noguer. Ganimation: Anatomically-aware facial animation from a single image. In ECCV, 2018. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 287, + 540 + ], + "type": "text", + "content": "[45] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: Learning to detect manipulated facial images. In CVPR, 2019. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 543, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 585 + ], + "type": "text", + "content": "[46] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "type": "text", + "content": "[47] Eran Segalis and Eran Galili. OGAN: Disrupting deep-fakes with an adversarial attack that survives training. arXiv preprint arXiv:2006.12247, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "type": "text", + "content": "[48] Kaede Shiohara and Toshihiko Yamasaki. Detecting deep-fakes with self-blended images. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[49] Kritaphat Songsri-in and Stefanos Zafeiriou. Complement face forensic detection and localization with facial landmarks. arXiv preprint arXiv:1910.05455, 2019. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[50] Luan Tran, Xi Yin, and Xiaoming Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, 2017. 1, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 488 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[51] Run Wang, Felix Juefei-Xu, Meng Luo, Yang Liu, and Lina Wang. FakeTagger: Robust safeguards against deepfake dissemination via provenance tracking. In ACMM, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "text", + "content": "[52] Sheng-Yu Wang, David Bau, and Jun-Yan Zhu. Sketch your own gan. In ICCV, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[53] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. CNN-generated images are surprisingly easy to spot... for now. In CVPR, 2020. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[54] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In CVPR, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[55] Xi Wu, Zhen Xie, YuTao Gao, and Yu Xiao. SSTNET: Detecting manipulated faces through spatial, steganalysis and temporal features. In ICASSP, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "text", + "content": "[56] Ying Xu, Kiran Raja, and Marius Pedersen. Supervised contrastive learning for generalizable and explainable deepfakes detection. In WACV, 2022. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[57] Chin-Yuan Yeh, Hsi-Wen Chen, Shang-Lun Tsai, and Sheng-De Wang. Disrupting image-translation-based deepfake algorithms with adversarial attacks. In WACVW, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "type": "text", + "content": "[58] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. DualGAN: Unsupervised dual learning for image-to-image translation. In CVPR, 2017. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "text", + "content": "[59] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in GAN fake images. In WIFS, 2019. 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 378, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 545, + 410 + ], + "type": "text", + "content": "[60] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 1, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 412, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 412, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 412, + 545, + 454 + ], + "type": "text", + "content": "[61] Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darryll, Alexei A Efros, Oliver Wang, and Eli Shechtman. Toward multimodal image-to-image translation. In NeurIPS, 2017. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[62] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. SEAN: Image synthesis with semantic region-adaptive normalization. In CVPR, 2020. 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12352" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_content_list.json b/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6f2791346c6ca18e279c423c0a83f93559fce05d --- /dev/null +++ b/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_content_list.json @@ -0,0 +1,1665 @@ +[ + { + "type": "text", + "text": "MaPLe: Multi-modal Prompt Learning", + "text_level": 1, + "bbox": [ + 283, + 161, + 687, + 182 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Muhammad Uzair Khattak1 Hanoona Rasheed1 Muhammad Maaz1", + "bbox": [ + 197, + 206, + 768, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Salman Khan $^{1,2}$ Fahad Shahbaz Khan $^{1,3}$", + "bbox": [ + 313, + 223, + 649, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Mohamed bin Zayed University of AI $^{2}$ Australian National University $^{3}$ Linköping University", + "bbox": [ + 90, + 241, + 872, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 234, + 290, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-trained vision-language (V-L) models such as CLIP have shown excellent generalization ability to downstream tasks. However, they are sensitive to the choice of input text prompts and require careful selection of prompt templates to perform well. Inspired by the Natural Language Processing (NLP) literature, recent CLIP adaptation approaches learn prompts as the textual inputs to fine-tune CLIP for downstream tasks. We note that using prompting to adapt representations in a single branch of CLIP (language or vision) is sub-optimal since it does not allow the flexibility to dynamically adjust both representation spaces on a downstream task. In this work, we propose Multi-modal Prompt Learning (MaPLe) for both vision and language branches to improve alignment between the vision and language representations. Our design promotes strong coupling between the vision-language prompts to ensure mutual synergy and discourages learning independent uni-modal solutions. Further, we learn separate prompts across different early stages to progressively model the stage-wise feature relationships to allow rich context learning. We evaluate the effectiveness of our approach on three representative tasks of generalization to novel classes, new target datasets and unseen domain shifts. Compared with the state-of-the-art method CoCoOp, MaPLe exhibits favorable performance and achieves an absolute gain of $3.45\\%$ on novel classes and $2.72\\%$ on overall harmonic-mean, averaged over 11 diverse image recognition datasets. Our code and pre-trained models are available at https://github.com/muzairkhattak/multimodal-prompt-learning.", + "bbox": [ + 75, + 319, + 471, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 745, + 209, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundational vision-language (V-L) models such as CLIP (Contrastive Language-Image Pretraining) [32] have shown excellent generalization ability to downstream tasks. Such models are trained to align language and vision modalities on web-scale data e.g., 400 million text-image pairs in CLIP. These models can reason about open-vocabulary visual concepts, thanks to the rich supervision provided by", + "bbox": [ + 75, + 769, + 469, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "natural language. During inference, hand-engineered text prompts are used e.g., 'a photo of a ' as a query for text encoder. The output text embeddings are matched with the visual embeddings from an image encoder to predict the output class. Designing high quality contextual prompts have been proven to enhance the performance of CLIP and other V-L models [17, 42].", + "bbox": [ + 496, + 290, + 892, + 389 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the effectiveness of CLIP towards generalization to new concepts, its massive scale and scarcity of training data (e.g., few-shot setting) makes it infeasible to fine-tune the full model for downstream tasks. Such fine-tuning can also forget the useful knowledge acquired in the large-scale pretraining phase and can pose a risk of overfitting to the downstream task. To address the above challenges, existing works propose language prompt learning to avoid manually adjusting the prompt templates and providing a mechanism to adapt the model while keeping the original weights frozen [14, 25, 29, 48, 49]. Inspired from Natural Language Processing (NLP), these approaches only explore prompt learning for the text encoder in CLIP (Fig. 1:a) while adaptation choices together with an equally important image encoder of CLIP remains an unexplored topic in the literature.", + "bbox": [ + 496, + 391, + 892, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our motivation derives from the multi-modal nature of CLIP, where a text and image encoder co-exist and both contribute towards properly aligning the V-L modalities. We argue that any prompting technique should adapt the model completely and therefore, learning prompts only for the text encoder in CLIP is not sufficient to model the adaptations needed for the image encoder. To this end, we set out to achieve completeness in the prompting approach and propose Multi-modal Prompt Learning (MaPLe) to adequately fine-tune the text and image encoder representations such that their optimal alignment can be achieved on the downstream tasks (Fig. 1:b). Our extensive experiments on three key representative settings including base-to-novel generalization, cross-dataset evaluation, and domain generalization demonstrate the strength of MaPLe. On base-to-novel generalization, our proposed MaPLe outperforms existing prompt learning approaches across 11 diverse image recognition datasets (Fig. 1:c) and achieves absolute average gain of $3.45\\%$ on novel classes and $2.72\\%$ on harmonic-mean", + "bbox": [ + 496, + 603, + 894, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 95, + 8, + 171, + 46 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 229, + 5, + 818, + 48 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19113", + "bbox": [ + 466, + 887, + 505, + 897 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9eb4884d9989b772b613fc7472cdd8830703485c6d51fda6ac3498f906878816.jpg", + "image_caption": [ + "(a) Existing prompt tuning methods (Uni-modal)" + ], + "image_footnote": [], + "bbox": [ + 122, + 122, + 441, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d537288361b8997d134adeff410f0017bea0bb9fef2994c446047f6a5f32525.jpg", + "image_caption": [ + "(b) Multi-modal Prompt Learning (MaPLe)", + "Figure 1. Comparison of MaPLe with standard prompt learning methods. (a) Existing methods adopt uni-modal prompting techniques to fine-tune CLIP representations as prompts are learned only in a single branch of CLIP (language or vision). (b) MaPLe introduces branch-aware hierarchical prompts that adapt both language and vision branches simultaneously for improved generalization. (c) MaPLe surpasses state-of-the-art methods on 11 diverse image recognition datasets for novel class generalization task." + ], + "image_footnote": [], + "bbox": [ + 122, + 243, + 442, + 340 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ee6a47f15551f34258f8ea90ebea1f4099db62ca677e5062c8af79aa1b9dd4ed.jpg", + "image_caption": [ + "(c) Performance comparison on base-to-novel generalization" + ], + "image_footnote": [], + "bbox": [ + 459, + 122, + 848, + 336 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "over the state-of-the-art method Co-CoOp [48]. Further, MaPLe demonstrates favorable generalization ability and robustness in cross-dataset transfer and domain generalization settings, leading to consistent improvements compared to existing approaches. Owing to its streamlined architectural design, MaPLe exhibits improved efficiency during both training and inference without much overhead, as compared to Co-CoOp which lacks efficiency due to its image instance conditioned design. In summary, the main contributions of this work include:", + "bbox": [ + 75, + 432, + 469, + 568 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose multi-modal prompt learning in CLIP to favourably align its vision-language representations. To the best of our knowledge, this is the first multi-modal prompting approach for fine-tuning CLIP.", + "- To link prompts learned in text and image encoders, we propose a coupling function to explicitly condition vision prompts on their language counterparts. It acts as a bridge between the two modalities and allows mutual propagation of gradients to promote synergy.", + "- Our multi-modal prompts are learned across multiple transformer blocks in both vision and language branches to progressively learn the synergistic behaviour of both modalities. This deep prompting strategy allows modeling the contextual relationships independently, thus providing more flexibility to align the vision-language representations." + ], + "bbox": [ + 95, + 571, + 469, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 77, + 815, + 218, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Language Models: The combined use of language supervision with natural images is found to be of great in", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "terest in the computer vision community. In contrast to models learned with only image supervision, these vision-language (V-L) models encode rich multimodal representations. Recently, V-L models like CLIP [32], ALIGN [15], LiT [45], FILIP [41] and Florence [43] have demonstrated exceptional performance on a wide spectrum of tasks including few-shot and zero-shot visual recognition. These models learn joint image-language representations in a self-supervised manner using abundantly available data from the web. For example, CLIP and ALIGN respectively use $\\sim 400\\mathrm{M}$ and $\\sim 1\\mathrm{B}$ image-text pairs to train a multimodal network. Although these pre-trained V-L models learn generalized representations, efficiently adapting them to downstream tasks is still a challenging problem. Many works have demonstrated better performance on downstream tasks by using tailored methods to adapt V-L models for few-shot image-recognition [9, 19, 46], object detection [8, 10, 27, 34, 44, 50], and segmentation [5, 22, 26, 33]. In this work, we propose a novel multi-modal prompt learning technique to effectively adapt CLIP for few-shot and zero-shot visual recognition tasks.", + "bbox": [ + 496, + 432, + 890, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Learning: The instructions in the form of a sentence, known as text prompt, are usually given to the language branch of a V-L model, allowing it to better understand the task. Prompts can be handcrafted for a downstream task or learned automatically during fine-tuning stage. The latter is referred to as 'Prompt Learning' which was first used in NLP [21,23,24] followed by the adaptation in V-L [48, 49, 51] and vision-only [16, 38, 39, 47] models. Similar to [16] our design also uses deep 'vision' prompting. However, ours is the first multi-modal prompting de", + "bbox": [ + 496, + 728, + 892, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19114", + "bbox": [ + 466, + 887, + 506, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93f706335d0a31f607a7e9c15f96265303f5203fafdf9d9584b52a0912f4d9db.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed MaPLe (Multi-modal Prompt Learning) framework for prompt learning in V-L models. MaPLe tunes both vision and language branches where only the context prompts are learned, while the rest of the model is frozen. MaPLe conditions the vision prompts on language prompts via a V-L coupling function $\\mathcal{F}$ to induce mutual synergy between the two modalities. Our framework uses deep contextual prompting where separate context prompts are learned across multiple transformer blocks." + ], + "image_footnote": [], + "bbox": [ + 166, + 123, + 807, + 336 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sign while [16] is uni-modal.", + "bbox": [ + 75, + 416, + 271, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt Learning in Vision Language models: Full fine-tuning and linear probing [9] are two typical approaches to adapt a V-L model (i.e. CLIP) to the downstream tasks. The complete fine-tuning results in degrading the previously learned joint V-L representation while linear probing limits the zero-shot capability of CLIP. To this end, inspired from prompt learning in NLP, many works have proposed to adapt V-L models by learning the prompt tokens in an end-to-end training. CoOp [49] fine-tunes CLIP for few-shot transfer by optimizing continuous set of prompt vectors at its language branch. Co-CoOp [48] highlights the inferior performance of CoOp on novel classes and solves the generalization issue by explicitly conditioning prompts on image instances. [25] proposes to optimize multiple set of prompts by learning the distribution of prompts. [18] adapt CLIP by learning prompts for video understanding tasks. [1] perform visual prompt tuning on CLIP by prompting on the vision branch. We note that the existing methods follow independent uni-modal solutions and learn prompts either in the language or in the vision branch of CLIP, thus adapting CLIP partially. In this paper, we explore an important question: given the multimodal nature of CLIP, is complete prompting (i.e., in both language and vision branches) better suited to adapt CLIP? Our work is the first to answer this question by investigating the effectiveness of multi-modal prompt learning in order to improve alignment between vision and language representations.", + "bbox": [ + 75, + 429, + 471, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 75, + 815, + 169, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach concerns with fine-tuning a pre-trained multimodal CLIP for better generalization to downstream tasks", + "bbox": [ + 75, + 838, + 469, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "through context optimization via prompting. Fig. 2 shows the overall architecture of our proposed MaPLe (Multimodal Prompt Learning) framework. Unlike previous approaches [48, 49] which learn context prompts only at the language branch, MaPLe proposes a joint prompting approach where the context prompts are learned in both vision and language branches. Specifically, we append learnable context tokens in the language branch and explicitly condition the vision prompts on the language prompts via a coupling function to establish interaction between them. To learn hierarchical contextual representations, we introduce deep prompting in both branches through separate learnable context prompts across different transformer blocks. During fine-tuning, only the context prompts along with their coupling function are learned while the rest of the model is frozen. Below, we first outline the pre-trained CLIP architecture and then present our proposed fine-tuning approach.", + "bbox": [ + 496, + 416, + 894, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting CLIP", + "text_level": 1, + "bbox": [ + 500, + 679, + 662, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We build our approach on a pre-trained vision-language (V-L) model, CLIP, which consists of a text and vision encoder. Consistent with existing prompting methods [48, 49], we use a vision transformer (ViT) [6] based CLIP model. CLIP encodes an image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ and a corresponding text description as explained below.", + "bbox": [ + 496, + 707, + 892, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoding Image: Image encoder $\\mathcal{V}$ with $K$ transformer layers $\\{\\mathcal{V}_i\\}_{i=1}^K$ , splits the image $I$ into $M$ fixed-size patches which are projected into patch embeddings $E_0 \\in \\mathbb{R}^{M \\times d_v}$ . Patch embeddings $E_i$ are input to the $(i+1)^{\\mathrm{th}}$ transformer block $(\\mathcal{V}_{i+1})$ along with a learnable class (CLS) token $\\mathbf{c}_i$", + "bbox": [ + 496, + 796, + 892, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19115", + "bbox": [ + 466, + 887, + 506, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e16ab2a8f4b922761701d24185fa40f44ed8b6343440e0896944f5853bd689a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 122, + 884, + 214 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/61c0a1218a8eac098d5c354f62e6fcf5de7744de687314db5a8701ba0d7aeaca.jpg", + "image_caption": [ + "Figure 3. t-SNE plots of image embeddings in uni-modal prompting method Co-CoOp, and MaPLe on 3 diverse image recognition datasets. MaPLe shows better separability in both base and novel classes." + ], + "image_footnote": [], + "bbox": [ + 88, + 216, + 884, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and sequentially processed through $K$ transformer blocks,", + "bbox": [ + 75, + 366, + 461, + 382 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathbf {c} _ {i}, E _ {i} ] = \\mathcal {V} _ {i} ([ \\mathbf {c} _ {i - 1}, E _ {i - 1} ]) \\quad i = 1, 2, \\dots , K.\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 389, + 426, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the final image representation $x$ , the class token $\\mathsf{c}_K$ of last transformer layer $(\\mathcal{V}_K)$ is projected to a common V-L latent embedding space via ImageProj,", + "bbox": [ + 75, + 413, + 469, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx = \\operatorname {I m a g e P r o j} \\left(\\mathrm {c} _ {K}\\right) \\quad x \\in \\mathbb {R} ^ {d _ {v l}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 461, + 395, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoding Text: CLIP text encoder generates feature representations for text description by tokenizing the words and projecting them to word embeddings $W_{0} = [w_{0}^{1}, w_{0}^{2}, \\dots, w_{0}^{N}] \\in \\mathbb{R}^{N \\times d_{l}}$ . At each stage, $W_{i}$ is input to the $(i + 1)^{\\mathrm{th}}$ transformer layer of text encoding branch $(\\mathcal{L}_{i + 1})$ ,", + "bbox": [ + 75, + 486, + 469, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ W _ {i} \\right] = \\mathcal {L} _ {i} \\left(W _ {i - 1}\\right) \\quad i = 1, 2, \\dots , K.\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 577, + 404, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final text representation $z$ is obtained by projecting the text embeddings corresponding to the last token of the last transformer block $\\mathcal{L}_K$ to a common V-L latent embedding space via TextProj,", + "bbox": [ + 75, + 600, + 469, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz = \\operatorname {T e x t P r o j} \\left(w _ {K} ^ {N}\\right) \\quad z \\in \\mathbb {R} ^ {d _ {v l}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 152, + 663, + 392, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zero-shot Classification: For zero-shot classification, text prompts are hand-crafted with class labels $y \\in \\{1, 2, \\dots, C\\}$ (e.g., 'a photo of a ) having $C$ classes. Prediction $\\hat{y}$ corresponding to the image $I$ having the highest cosine similarity score $(sim(\\cdot))$ is calculated with a temperature parameter $\\tau$ ,", + "bbox": [ + 75, + 687, + 469, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\hat {y} | x) = \\frac {\\exp (s i m (x , z _ {\\hat {y}}) / \\tau)}{\\sum_ {i = 1} ^ {C} \\exp (s i m (x , z _ {i}))}.\n$$\n", + "text_format": "latex", + "bbox": [ + 157, + 777, + 385, + 812 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. MaPLe: Multi-modal Prompt Learning", + "text_level": 1, + "bbox": [ + 75, + 817, + 412, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To efficiently fine-tune CLIP for downstream image recognition tasks, we explore the potential of multi-modal prompt", + "bbox": [ + 75, + 838, + 469, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tuning. We reason that prior works that have predominantly explored uni-modal approaches are less suitable as they do not offer the flexibility to dynamically adapt both language and vision representation spaces. Thus to achieve completeness in prompting, we underline the importance of multimodal prompting approach. In Fig. 3, we visualize and compare the image embeddings of MaPLe with recent state-of-the-art work, Co-CoOp. Note that the image embeddings of CLIP, CoOp and Co-CoOp will be identical as they do not learn prompts in the vision branch. The visualization shows that image embeddings of MaPLe are more separable indicating that learning vision prompts in addition to language prompts leads to better adaptation of CLIP.", + "bbox": [ + 496, + 366, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to multi-modal prompting, we find that it is essential to learn prompts in the deeper transformer layers to progressively model stage-wise feature representations. To this end, we propose to introduce learnable tokens in the first $J$ (where $J < K$ ) layers of both vision and language branches. These multi-modal hierarchical prompts utilize the knowledge embedded in CLIP model to effectively learn task relevant contextual representations (see Fig. 4).", + "bbox": [ + 496, + 547, + 890, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Deep Language Prompting", + "text_level": 1, + "bbox": [ + 500, + 678, + 741, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn the language context prompts, we introduce $b$ learnable tokens $\\{P^i \\in \\mathbb{R}^{d_l}\\}_{i=1}^b$ , in the language branch of CLIP. The input embeddings now follow the form $[P^1, P^2, \\dots, P^b, W_0]$ , where $W_0 = [w^1, w^2, \\dots, w^N]$ corresponds to fixed input tokens. New learnable tokens are further introduced in each transformer block of the language encoder $(\\mathcal{L}_i)$ up to a specific depth $J$ ,", + "bbox": [ + 496, + 701, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\_ , W _ {i} ] = \\mathcal {L} _ {i} ([ P _ {i - 1}, W _ {i - 1} ]) \\quad i = 1, 2, \\dots , J. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 525, + 809, + 890, + 826 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $[\\cdot, \\cdot]$ refers to the concatenation operation. After $J^{\\mathrm{th}}$ transformer layer, the subsequent layers process previous", + "bbox": [ + 496, + 838, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19116", + "bbox": [ + 466, + 887, + 506, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "layer prompts and final text representation $z$ is computed,", + "bbox": [ + 75, + 125, + 456, + 138 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} [ P _ {j}, W _ {j} ] = \\mathcal {L} _ {j} ([ P _ {j - 1}, W _ {j - 1} ]) \\quad j = J + 1, \\dots , K, (2) \\\\ z = \\operatorname {T e x t P r o j} \\left(w _ {K} ^ {N}\\right). (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 92, + 149, + 466, + 182 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When $J = 1$ , the learnable tokens $P$ are only applied at the input of first transformer layer, and this deep language prompting technique degenerates to CoOp [49].", + "bbox": [ + 75, + 192, + 468, + 235 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Deep Vision Prompting", + "text_level": 1, + "bbox": [ + 77, + 251, + 293, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to deep language prompting, we introduce $b$ learnable tokens $\\{\\hat{P}^i\\in \\mathbb{R}^{d_v}\\}_{i = 1}^b$ , in the vision branch of CLIP alongside the input image tokens. New learnable tokens are further introduced in deeper transformer layers of the image encoder $(\\mathcal{V})$ up to depth $J$ .", + "bbox": [ + 75, + 271, + 468, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} [ c _ {i}, E _ {i}, \\_ ] = \\mathcal {V} _ {i} ([ c _ {i - 1}, E _ {i - 1}, \\bar {P} _ {i - 1} ]) \\quad i = 1, 2, \\dots , J, \\\\ [ c _ {j}, E _ {j}, \\tilde {P} _ {j} ] = \\mathcal {V} _ {j} ([ c _ {j - 1}, E _ {j - 1}, \\tilde {P} _ {j - 1} ]) \\quad j = J + 1, \\dots , K, \\\\ x = \\operatorname {I m a g e P r o j} \\left(c _ {K}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 352, + 468, + 403 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our deep prompting provides the flexibility to learn prompts across different feature hierarchies within the ViT architecture. We find that sharing prompts across stages is better compared to independent prompts as features are more correlated due to successive transformer block processing. Thus, the later stages do not provide independently-learned complimentary prompts as compared to the early stages.", + "bbox": [ + 75, + 413, + 468, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Vision Language Prompt Coupling", + "text_level": 1, + "bbox": [ + 75, + 527, + 374, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We reason that in prompt tuning it is essential to take a multi-modal approach and simultaneously adapt both the vision and language branch of CLIP in order to achieve completeness in context optimization. A simple approach would be to naively combine deep vision and language prompting, where both the language prompts $P$ , and the vision prompts $\\tilde{P}$ , will be learned during the same training schedule. We name this design as 'Independent V-L Prompting'. Although this approach satisfies the requirement of completeness in prompting, this design lacks synergy between vision and language branch as both branches do not interact while learning the task relevant context prompts.", + "bbox": [ + 75, + 548, + 468, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we propose a branch-aware multi-modal prompting which tunes vision and language branch of CLIP together by sharing prompts across both modalities. Language prompt tokens are introduced in the language branch up to $J^{\\mathrm{th}}$ transformer block similar to deep language prompting as illustrated in Eqs. 1-3. To ensure mutual synergy between V-L prompts, vision prompts $\\tilde{P}$ , are obtained by projecting language prompts $P$ via vision-to-language projection which we refer to as $V-L$ coupling function $\\mathcal{F}(\\cdot)$ such that $\\tilde{P}_k = \\mathcal{F}_k(P_k)$ . The coupling function is implemented as a linear layer which maps $d_l$ dimensional inputs", + "bbox": [ + 75, + 714, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to $d_v$ . This acts as a bridge between the two modalities, thus encouraging mutual propagation of gradients.", + "bbox": [ + 500, + 125, + 890, + 153 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} [ c _ {i}, E _ {i}, \\_ ] = \\mathcal {V} _ {i} ([ c _ {i - 1}, E _ {i - 1}, \\mathcal {F} _ {\\boldsymbol {i} - \\mathbf {1}} (\\boldsymbol {P} _ {\\boldsymbol {i} - \\mathbf {1}}) ]) i = 1, \\dots , J \\\\ [ c _ {j}, E _ {j}, \\tilde {P} _ {j} ] = \\mathcal {V} _ {j} ([ c _ {j - 1}, E _ {j - 1}, \\tilde {P} _ {j - 1} ]) \\quad j = J + 1, \\dots , K \\\\ x = \\operatorname {I m a g e P r o j} \\left(c _ {K}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 165, + 897, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike independent V-L prompting, explicit conditioning of $\\tilde{P}$ on $P$ helps learn prompts in a shared embedding space between the two branches, thus improving mutual synergy.", + "bbox": [ + 500, + 226, + 892, + 269 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 282, + 630, + 298 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Benchmark setting", + "text_level": 1, + "bbox": [ + 500, + 304, + 681, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generalization from Base-to-Novel Classes: We evaluate the generalizability of MaPLe, and follow a zero-shot setting where the datasets are split into base and novel classes. The model is trained only on the base classes in a few-shot setting and evaluated on base and novel categories.", + "bbox": [ + 500, + 326, + 890, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-dataset Evaluation: To validate the potential of our approach in cross-dataset transfer, we evaluate our ImageNet trained model directly on other datasets. Consistent with Co-CoOp, our model is trained on all 1000 ImageNet classes in a few-shot manner.", + "bbox": [ + 500, + 396, + 890, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Domain Generalization: We evaluate the robustness of our method on out-of-distribution datasets. Similar to cross-dataset evaluation, we test our ImageNet trained model directly on four other ImageNet datasets that contain various types of domain shifts.", + "bbox": [ + 500, + 465, + 890, + 533 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: For generalization from base-to-novel classes and cross-dataset evaluation, we follow [48, 49] and evaluate the performance of our method on 11 image classification datasets which covers a wide range of recognition tasks. This includes two generic-objects datasets, ImageNet [4] and Caltech101 [7]; five fine-grained datasets, OxfordPets [31], StanfordCars [20], Flowers102 [30], Food101 [2], and FGVCAircraft [28]; a scene recognition dataset SUN397 [40]; an action recognition dataset UCF101 [36]; a texture dataset DTD [3] and a satellite-image dataset EuroSAT [11]. For domain generalization, we use ImageNet as source dataset and its four variants as target datasets including ImageNetV2 [35], ImageNet-Sketch [37], ImageNet-A [13] and ImageNet-R [12].", + "bbox": [ + 500, + 534, + 890, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We use a few-shot training strategy in all experiments at 16 shots which are randomly sampled for each class. We apply prompt tuning on a pretrained ViT-B/16 CLIP model where $d_{l} = 512$ , $d_{v} = 768$ and $d_{vl} = 512$ . For MaPLe, we set prompt depth $J$ to 9 and the language and vision prompt lengths to 2. All models are trained for 5 epochs with a batch-size of 4 and a learning rate of 0.0035 via SGD optimizer on a single NVIDIA A100 GPU. We report base and novel class accuracies and their harmonic mean (HM) averaged over 3 runs. We initial-", + "bbox": [ + 500, + 729, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19117", + "bbox": [ + 466, + 887, + 505, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b0a8735d5c9c6927cd03257354e8d3bb43ea75c789cb6616fba69bb20cad4748.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBase Acc.Novel Acc.HMGFLOPS
1: MaPLe shallow (J=1)80.1073.5276.67167.1
2: Deep vision prompting80.2473.4376.6818.0
3: Deep language prompting81.7273.8177.56166.8
4: Independent V-L prompting82.1574.0777.90167.0
5: MaPLe (Ours)82.2875.1478.55167.0
", + "bbox": [ + 80, + 122, + 466, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ize the language prompts of the first layer $P_0$ with the pretrained CLIP word embeddings of the template 'a photo of a ', while for the subsequent layers they are randomly initialized from a normal distribution. For training MaPLe on all 1000 classes of ImageNet as a source model, prompt depth $J$ is set to 3 and the model trained for 2 epochs with learning rate of 0.0026. Hyper-parameters for deep language prompting, deep vision prompting, and independent V-L prompting are detailed in Appendix A. The hyper-parameters are fixed across all datasets.", + "bbox": [ + 75, + 255, + 469, + 394 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Prompting CLIP via Vision-Language Prompts", + "text_level": 1, + "bbox": [ + 75, + 401, + 469, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompting Variants: We first evaluate the performance of different possible prompting design choices as an ablation for our proposed branch-aware multi-modal prompting, MaPLe. These variants include shallow MaPLe, deep language prompting, deep vision prompting and independent V-L prompting. In Table 1, we present the results averaged over the 11 image recognition datasets. Shallow MaPLe (row-1) provides consistent improvements over CoOp and Co-CoOp in terms of generalization. Deep language prompting (row-3) shows improvements over deep vision prompting (row-2), indicating that prompts learned at the language branch provide better adaptation of CLIP. Although separately combining the above two approaches (row-4) further improves the performance, it struggles to achieve comprehensive benefits from the language and vision branches. We hypothesize that this is due to the lack of synergy between the learned vision and language prompts as they do not interact with each other during training. Meanwhile, MaPLe tied with deep prompting (row-4) combines the benefits of prompting in both branches by enforcing interactions through explicit conditioning of vision prompts on the language prompts. It provides improvements on novel and base class accuracies which leads to the best HM of $78.55\\%$ . We explore other possible design choices and present the ablations in Appendix B.", + "bbox": [ + 75, + 422, + 469, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Base-to-Novel Generalization", + "text_level": 1, + "bbox": [ + 75, + 776, + 338, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization to Unseen Classes: Table 3 presents the performance of MaPLe in base-to-novel generalization setting on 11 recognition datasets. We compare its performance with CLIP zero-shot, and recent prompt learning works including CoOp [49] and Co-CoOp [48]. In case of", + "bbox": [ + 75, + 796, + 469, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP, we use hand-crafted prompts that are specifically designed for each dataset.", + "bbox": [ + 498, + 125, + 890, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In comparison with the state-of-the-art Co-CoOp, MaPLe shows improved performance on both base and novel categories on all 11 datasets with an exception of marginal reduction on only the base class performance of Caltech101. With mutual synergy from the branch-aware multi-modal prompting, MaPLe better generalizes to novel categories on all 11 datasets in comparison with Co-CoOp, and obtains an overall gain from $71.69\\%$ to $75.14\\%$ . When taking into account both the base and novel classes, MaPLe shows an absolute average gain of $2.72\\%$ over Co-CoOp.", + "bbox": [ + 496, + 153, + 892, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In comparison with CLIP on novel classes, Co-CoOp improves only on 4/11 datasets dropping the average novel accuracy from $74.22\\%$ to $71.69\\%$ . MaPLe is a strong competitor which improves accuracy over CLIP on novel classes on 6/11 datasets, with an average gain from $74.22\\%$ to $75.14\\%$ . Generalization and Performance on Base Classes: CoCoOp solves the poor generalization problem in CoOp by conditioning prompts on image instances and shows significant gains in novel categories. However on base classes, it improves over CoOp only on 3/11 datasets with an average drop in performance from $82.69\\%$ to $80.47\\%$ . Meanwhile, the completeness in prompting helps MaPLe improve over CoOp on base classes in 6/11 datasets maintaining the average base accuracy to around $82.28\\%$ , in addition to its improvement in generalization to novel classes.", + "bbox": [ + 496, + 290, + 892, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We find that the training strategies of Co-CoOp can be used to substantially boost the generalization performance of vanilla CoOp (6.8% gain in novel classes). We therefore compare our method with $\\mathrm{CoOp}^{\\dagger}$ , which trains CoOp in Co-CoOp setting (refer to Appendix A for more details).", + "bbox": [ + 496, + 500, + 890, + 568 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/945c7e9e155d7cca0b008eb978f0498fbc2f7223110c759bd3b0c55f747d7669.jpg", + "table_caption": [ + "Table 1. Comparison of MaPLe with different prompting designs in base-to-novel generalization. Results are averaged over 11 datasets. HM refers to harmonic mean." + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CoOp82.6963.2271.66
Co-CoOp80.4771.6975.83
CoOp†80.8570.0275.04
MaPLe82.2875.1478.55
", + "bbox": [ + 541, + 577, + 850, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Generalization comparison of MaPLe with CoOp†.", + "bbox": [ + 516, + 659, + 872, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compare to $\\mathrm{CoOp}^{\\dagger}$ , the vanilla CoOp model seems to overfit on base classes. When compared to $\\mathrm{CoOp}^{\\dagger}$ which attains an average base accuracy of $80.85\\%$ , MaPLe shows an improvement of $1.43\\%$ with the average base accuracy of $82.28\\%$ (Table 2).", + "bbox": [ + 496, + 685, + 890, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Cross-Dataset Evaluation", + "text_level": 1, + "bbox": [ + 500, + 762, + 731, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test the cross-dataset generalization ability of MaPLe by learning multi-modal prompts on all the 1000 ImageNet classes and then transferring it directly on the remaining 10 datasets. Table 4 shows the performance comparison between MaPLe, CoOp and Co-CoOp. On the ImageNet source dataset, MaPLe achieves performance comparable", + "bbox": [ + 496, + 783, + 892, + 868 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19118", + "bbox": [ + 466, + 887, + 506, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/06e9817516416e3b0ec7c23425fb0b3ef389873303c9de071adee11ce887efdd.jpg", + "table_caption": [ + "(a) Average over 11 datasets" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP69.3474.2271.70
CoOp82.6963.2271.66
Co-CoOp80.4771.6975.83
MaPLe82.2875.1478.55
+1.81+3.45+2.72
", + "bbox": [ + 98, + 138, + 340, + 236 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3619c18733eddca1ef0e05bb4f34f924ea5bb7d27bb7157719958df88c8a2b9b.jpg", + "table_caption": [ + "(b) ImageNet." + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP72.4368.1470.22
CoOp76.4767.8871.92
Co-CoOp75.9870.4373.10
MaPLe76.6670.5473.47
+0.68+0.11+0.37
", + "bbox": [ + 359, + 138, + 604, + 236 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a7f7c3134986e91cdf26af50eef6b0fdd831ee2514051b473c8bcd4fcc2169f.jpg", + "table_caption": [ + "(c) Caltech101" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP96.8494.0095.40
CoOp98.0089.8193.73
Co-CoOp97.9693.8195.84
MaPLe97.7494.3696.02
-0.22+0.55+0.18
", + "bbox": [ + 630, + 138, + 872, + 236 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/54c951ed922ac44768efa32313d93027d5fc2634fd873060d75c7627447ba378.jpg", + "table_caption": [ + "(d) OxfordPets" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP91.1797.2694.12
CoOp93.6795.2994.47
Co-CoOp95.2097.6996.43
MaPLe95.4397.7696.58
+0.23+0.07+0.15
", + "bbox": [ + 97, + 261, + 340, + 357 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6355c37d6f4cef47b5f8f81e71edd0646469ebc8883a98fb02b8d0fba6ee83b2.jpg", + "table_caption": [ + "(e) StanfordCars" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP63.3774.8968.65
CoOp78.1260.4068.13
Co-CoOp70.4973.5972.01
MaPLe72.9474.0073.47
+2.45+0.41+1.46
", + "bbox": [ + 359, + 261, + 604, + 357 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5b64891f6c01ce49e140225afbf35d5329d56e29c4f91bb72d76069d6b50e16.jpg", + "table_caption": [ + "(f) Flowers102" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP72.0877.8074.83
CoOp97.6059.6774.06
Co-CoOp94.8771.7581.71
MaPLe95.9272.4682.56
+1.05+0.71+0.85
", + "bbox": [ + 630, + 261, + 872, + 357 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b4cbc69e46a5d3e0f2efd826cdec1a1072ea7b8c64da064a219eb1940715ccfd.jpg", + "table_caption": [ + "(g) Food101" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP90.1091.2290.66
CoOp88.3382.2685.19
Co-CoOp90.7091.2990.99
MaPLe90.7192.0591.38
+0.01+0.76+0.39
", + "bbox": [ + 97, + 382, + 340, + 479 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/133110c6e00a0f267d09280908e2fa45186b6044718c6f5ed383d8b895dfdb93.jpg", + "table_caption": [ + "(h) FGVCAircraft" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP27.1936.2931.09
CoOp40.4422.3028.75
Co-CoOp33.4123.7127.74
MaPLe37.4435.6136.50
+4.03+11.90+8.76
", + "bbox": [ + 355, + 382, + 605, + 479 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/95cae59c394804fbd7b8c33f9e20be71a37f584a58a3ebd34adaced1376ff5ee.jpg", + "table_caption": [ + "(i) SUN397" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP69.3675.3572.23
CoOp80.6065.8972.51
Co-CoOp79.7476.8678.27
MaPLe80.8278.7079.75
+1.08+1.84+1.48
", + "bbox": [ + 630, + 382, + 872, + 479 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/577a2d66be88e625d8fd59f1aea16f4ac14eef2ef9b49ed291f040fe94aa848c.jpg", + "table_caption": [ + "(j) DTD" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP53.2459.9056.37
CoOp79.4441.1854.24
Co-CoOp77.0156.0064.85
MaPLe80.3659.1868.16
+3.35+3.18+3.31
", + "bbox": [ + 93, + 504, + 335, + 600 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f2fc36fd7302cc7a490a2fc72ec3f1f9bcec67dc30e646879b0793354860352f.jpg", + "table_caption": [ + "(k) EuroSAT" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP56.4864.0560.03
CoOp92.1954.7468.69
Co-CoOp87.4960.0471.21
MaPLe94.0773.2382.35
+6.58+13.19+11.14
", + "bbox": [ + 355, + 504, + 612, + 599 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1f7f6ee83b39cf01fb3f18952d54219c91e552ef9f12862b19ea8ec503c91bf.jpg", + "table_caption": [ + "(1)UCF101" + ], + "table_footnote": [], + "table_body": "
BaseNovelHM
CLIP70.5377.5073.85
CoOp84.6956.0567.46
Co-CoOp82.3373.4577.64
MaPLe83.0078.6680.77
+0.67+5.21+3.13
", + "bbox": [ + 632, + 504, + 875, + 599 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/380b68b8c347cda96b24504a99374ff0eb1f9a2c205065f4ce270b051671b05f.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art methods on base-to-novel generalization. MaPLe learns multi-modal prompts and demonstrates strong generalization results over existing methods on 11 recognition datasets. Absolute gains over Co-CoOp are indicated in blue." + ], + "table_footnote": [], + "table_body": "
SourceTarget
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101AircraftSUN397DTDEuroSATUCF101Average
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
Co-CoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
MaPLe70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
", + "bbox": [ + 100, + 640, + 721, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison of MaPLe with existing approaches on cross-dataset evaluation. Overall, MaPLe achieves competitive performance providing highest average accuracy, indicating better generalization.", + "bbox": [ + 729, + 642, + 890, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to competing approaches but demonstrates a much stronger generalization performance by surpassing CoOp in 9/10 and Co-CoOp in 8/10 datasets. Overall, MaPLe shows competitive performance leading to the highest averaged accuracy of $66.30\\%$ . This suggests that the use of branch-aware V-L prompting in MaPLe facilitates better generalization.", + "bbox": [ + 75, + 781, + 469, + 865 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Domain Generalization", + "text_level": 1, + "bbox": [ + 500, + 780, + 714, + 793 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show that MaPLe generalizes favourably on out-of-distribution datasets as compared to CoOp and Co-CoOp. We evaluate the direct transferability of ImageNet trained model to various out-of-domain datasets, and observe that it consistently improves against all the existing approaches", + "bbox": [ + 498, + 796, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19119", + "bbox": [ + 466, + 887, + 505, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d7563e65cfa764b3bf696aee2d61b18b7a1ec1d659cb759949414ad8553b05c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Comparison of MaPLe with existing approaches in domain generalization setting. MaPLe shows consistent improvements on all target datasets." + ], + "table_body": "
SourceTarget
ImageNetImageNetV2ImageNet-SImageNet-AImageNet-R
CLIP66.7360.8346.1547.7773.96
CoOp71.5164.2047.9949.7175.21
Co-CoOp71.0264.0748.7550.6376.18
MaPLe70.7264.0749.1550.9076.98
", + "bbox": [ + 78, + 122, + 476, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as indicated in Table 5. This indicates that utilizing multimodal branch-aware prompting helps MaPLe in enhancing the generalization and robustness of V-L models like CLIP.", + "bbox": [ + 75, + 271, + 469, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Experiments", + "text_level": 1, + "bbox": [ + 75, + 321, + 283, + 336 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Depth: In Fig. 4 (left), we illustrate the effect of prompt depth $J$ for MaPLe and ablate on the depth of language and vision branch individually. In general, the performance improves as prompt depth increases. We note that performance sensitivity increases when randomly initialized prompts are inserted in the deeper layers of a frozen model where the model feature space is already mature. Similar trend is also reported by [16]. As earlier methods utilize shallow language prompting ( $J = 1$ ), we compare our method with deep language prompting. Overall, MaPLe achieves better performance than deep language prompting and achieves maximum performance at a depth of 9.", + "bbox": [ + 75, + 343, + 469, + 509 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Length: Fig. 4 (right) shows the effect of prompt length for MaPLe. As the prompt length increases, the performance on base classes is generally maintained, while the novel class accuracy decreases. This indicates over-fitting which inherently hurts the generalization to novel classes.", + "bbox": [ + 75, + 509, + 469, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Multi-modal Prompting: Fig. 5 shows the analysis of per class accuracy for selected datasets in the order of increasing domain shift. It indicates that the performance gains of MaPLe in comparison to Co-CoOp varies across different datasets. MaPLe provides significant gains over Co-CoOp for datasets that have large distribution shifts from the pretraining dataset of CLIP, and vision concepts that are usually rare and less generic. Further detailed analysis is provided in Appendix C.", + "bbox": [ + 75, + 579, + 469, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompting complexity: Table 6 shows the computational complexity of MaPLe in comparison with other approaches.", + "bbox": [ + 75, + 703, + 468, + 732 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8288210b3d7acda788e8c59c8fd82ca205aec09731370e88eb536d6e1d14a208.jpg", + "image_caption": [ + "Figure 4. Ablation on prompt depth (left) and prompt length (right) in MaPLe. We report average results on the held-out validation sets of all datasets." + ], + "image_footnote": [], + "bbox": [ + 78, + 737, + 276, + 823 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d3568a4077d46c53fe89df4bc2fa3cdf22236decbe0d489c71d5f5e46f955731.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 273, + 737, + 466, + 823 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac8ffce85353eb3f418849e2d8a5ffb6e4f88a1db2ce83d86ea563fd8c57a052.jpg", + "image_caption": [ + "Figure 5. Percentage classes where MaPLe shows improved performance over Co-CoOp, which increases as dataset domain shift from generic categories increases $(\\rightarrow)$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 138, + 721, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although MaPLe utilizes multi-modal prompts, its overall FLOPS (Floating Point Operations) exceeds only by $0.1\\%$ over CoOp and Co-CoOp. The independent V-L prompting also provides comparable FLOP count. In terms of inference speed, Co-CoOp is significantly slower and the FPS (Frames Per Second) remains constant as the batch size increases. In contrast, MaPLe has no such overhead and provides much better inference and training speeds. Further, MaPLe provides better convergence as it requires only half training epochs as compared to Co-CoOp (5 vs 10 epochs). MaPLe adds about $2.85\\%$ training parameters on top of CLIP. To study if the performance gain is mainly attributed to more parameters, we experiment with $\\mathrm{MaPLe\\dagger}$ , which uses a unified V-L coupling function for all layer prompts. $\\mathrm{MaPLe\\dagger}$ with about 9x lesser parameters than MaPLe also improves over existing methods. We also ablate by comparing MaPLe with heavier CoCoOp in Appendix D.", + "bbox": [ + 496, + 235, + 894, + 471 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/70c54e958a78e1a21541dda019db13d58dbd1940304c97ab35a46e7a63e3997e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Comparison of computational complexity among different prompting methods. MaPLe† is a MaPLe version which utilizes a common V-L coupling function for all layers." + ], + "table_body": "
MethodParamsParams % CLIPFPS (with BS)HM
14100
CoOp20480.00213.855.31353.071.66
CoCoOp353600.0364.6114.715.175.83
Independent V-L314880.0262.5239.41383.877.90
MaPLe3.55 M2.8560.2239.01365.178.55
MaPLe†0.41 M0.3360.2238.01365.078.11
", + "bbox": [ + 503, + 480, + 892, + 581 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 636, + 619, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptation of large-scale V-L models, e.g., CLIP [32] to downstream tasks is a challenging problem due to the large number of tunable parameters and limited size of downstream datasets. Prompt learning is an efficient and scalable technique to tailor V-L models to novel downstream tasks. To this end, the current prompt learning approaches either consider only the vision or language side prompting. Our work shows that it is critical to perform prompting for both vision and language branches to appropriately adapt V-L models to downstream tasks. Further, we propose a strategy to ensure synergy between vision-language modalities by explicitly conditioning the vision prompts on textual prompt across different transformer stages. Our approach improves the generalization towards novel categories, cross-dataset transfer and datasets with domain shifts.", + "bbox": [ + 496, + 659, + 892, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19120", + "bbox": [ + 466, + 887, + 506, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 123, + 174, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 2022. 3, 11, 12", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In The European Conference on Computer Vision, pages 446-461. Springer, 2014. 5", + "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 5", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 5", + "[5] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3", + "[7] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5", + "[8] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In The European Conference on Computer Vision, 2022. 2", + "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2, 3", + "[10] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2", + "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5", + "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization." + ], + "bbox": [ + 78, + 147, + 471, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 5", + "[13] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 5", + "[14] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 1", + "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2", + "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In The European Conference on Computer Vision, 2022. 2, 3, 8", + "[17] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt is worth millions of parameters? low-resource prompt-based learning for vision-language models. arXiv preprint arXiv:2110.08484, 2021. 1", + "[18] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In The European Conference on Computer Vision, 2021. 3", + "[19] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model, 2022. 2", + "[20] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 554–561, 2013. 5", + "[21] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In _Conference on Empirical Methods in Natural Language Processing_, 2021. 2", + "[22] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2", + "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 2021. 2", + "[24] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 2021. 2", + "[25] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceed-" + ], + "bbox": [ + 501, + 125, + 890, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19121", + "bbox": [ + 466, + 887, + 503, + 899 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 3", + "[26] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. 2", + "[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In The European Conference on Computer Vision. Springer, 2022. 2", + "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[29] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. 2022. 1", + "[30] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5", + "[31] Omkar M Parkhi, Andreae Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 8", + "[33] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2", + "[34] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In Advances in Neural Information Processing Systems, 2022. 2", + "[35] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 5", + "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5", + "[37] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, volume 32, 2019. 5", + "[38] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary" + ], + "bbox": [ + 80, + 125, + 469, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "prompting for rehearsal-free continual learning. In The European Conference on Computer Vision, 2022. 2", + "[39] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2", + "[40] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5", + "[41] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2", + "[42] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 1", + "[43] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2", + "[44] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In The European Conference on Computer Vision, 2022. 2", + "[45] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 2", + "[46] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In The European Conference on Computer Vision, 2022. 2", + "[47] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 2", + "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 2, 3, 5, 6", + "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, pages 1-12, 2022. 1, 2, 3, 5, 6", + "[50] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In The European Conference on Computer Vision, 2022. 2", + "[51] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. arXiv preprint arXiv:2205.14865, 2022. 2" + ], + "bbox": [ + 501, + 125, + 890, + 862 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19122", + "bbox": [ + 466, + 887, + 505, + 897 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_model.json b/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9f70631562a3eb46eec28c967670f980fb88abf4 --- /dev/null +++ b/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.096, + 0.009, + 0.172, + 0.047 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.006, + 0.819, + 0.049 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.285, + 0.162, + 0.688, + 0.183 + ], + "angle": 0, + "content": "MaPLe: Multi-modal Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.207, + 0.77, + 0.223 + ], + "angle": 0, + "content": "Muhammad Uzair Khattak1 Hanoona Rasheed1 Muhammad Maaz1" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.224, + 0.65, + 0.24 + ], + "angle": 0, + "content": "Salman Khan\\(^{1,2}\\) Fahad Shahbaz Khan\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.242, + 0.873, + 0.261 + ], + "angle": 0, + "content": "\\(^{1}\\)Mohamed bin Zayed University of AI \\(^{2}\\)Australian National University \\(^{3}\\)Linköping University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.291, + 0.314, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.473, + 0.722 + ], + "angle": 0, + "content": "Pre-trained vision-language (V-L) models such as CLIP have shown excellent generalization ability to downstream tasks. However, they are sensitive to the choice of input text prompts and require careful selection of prompt templates to perform well. Inspired by the Natural Language Processing (NLP) literature, recent CLIP adaptation approaches learn prompts as the textual inputs to fine-tune CLIP for downstream tasks. We note that using prompting to adapt representations in a single branch of CLIP (language or vision) is sub-optimal since it does not allow the flexibility to dynamically adjust both representation spaces on a downstream task. In this work, we propose Multi-modal Prompt Learning (MaPLe) for both vision and language branches to improve alignment between the vision and language representations. Our design promotes strong coupling between the vision-language prompts to ensure mutual synergy and discourages learning independent uni-modal solutions. Further, we learn separate prompts across different early stages to progressively model the stage-wise feature relationships to allow rich context learning. We evaluate the effectiveness of our approach on three representative tasks of generalization to novel classes, new target datasets and unseen domain shifts. Compared with the state-of-the-art method CoCoOp, MaPLe exhibits favorable performance and achieves an absolute gain of \\(3.45\\%\\) on novel classes and \\(2.72\\%\\) on overall harmonic-mean, averaged over 11 diverse image recognition datasets. Our code and pre-trained models are available at https://github.com/muzairkhattak/multimodal-prompt-learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Foundational vision-language (V-L) models such as CLIP (Contrastive Language-Image Pretraining) [32] have shown excellent generalization ability to downstream tasks. Such models are trained to align language and vision modalities on web-scale data e.g., 400 million text-image pairs in CLIP. These models can reason about open-vocabulary visual concepts, thanks to the rich supervision provided by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.893, + 0.39 + ], + "angle": 0, + "content": "natural language. During inference, hand-engineered text prompts are used e.g., 'a photo of a ' as a query for text encoder. The output text embeddings are matched with the visual embeddings from an image encoder to predict the output class. Designing high quality contextual prompts have been proven to enhance the performance of CLIP and other V-L models [17, 42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.893, + 0.6 + ], + "angle": 0, + "content": "Despite the effectiveness of CLIP towards generalization to new concepts, its massive scale and scarcity of training data (e.g., few-shot setting) makes it infeasible to fine-tune the full model for downstream tasks. Such fine-tuning can also forget the useful knowledge acquired in the large-scale pretraining phase and can pose a risk of overfitting to the downstream task. To address the above challenges, existing works propose language prompt learning to avoid manually adjusting the prompt templates and providing a mechanism to adapt the model while keeping the original weights frozen [14, 25, 29, 48, 49]. Inspired from Natural Language Processing (NLP), these approaches only explore prompt learning for the text encoder in CLIP (Fig. 1:a) while adaptation choices together with an equally important image encoder of CLIP remains an unexplored topic in the literature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Our motivation derives from the multi-modal nature of CLIP, where a text and image encoder co-exist and both contribute towards properly aligning the V-L modalities. We argue that any prompting technique should adapt the model completely and therefore, learning prompts only for the text encoder in CLIP is not sufficient to model the adaptations needed for the image encoder. To this end, we set out to achieve completeness in the prompting approach and propose Multi-modal Prompt Learning (MaPLe) to adequately fine-tune the text and image encoder representations such that their optimal alignment can be achieved on the downstream tasks (Fig. 1:b). Our extensive experiments on three key representative settings including base-to-novel generalization, cross-dataset evaluation, and domain generalization demonstrate the strength of MaPLe. On base-to-novel generalization, our proposed MaPLe outperforms existing prompt learning approaches across 11 diverse image recognition datasets (Fig. 1:c) and achieves absolute average gain of \\(3.45\\%\\) on novel classes and \\(2.72\\%\\) on harmonic-mean" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19113" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.123, + 0.442, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.224, + 0.434, + 0.236 + ], + "angle": 0, + "content": "(a) Existing prompt tuning methods (Uni-modal)" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.244, + 0.443, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.342, + 0.413, + 0.353 + ], + "angle": 0, + "content": "(b) Multi-modal Prompt Learning (MaPLe)" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.123, + 0.85, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.342, + 0.84, + 0.355 + ], + "angle": 0, + "content": "(c) Performance comparison on base-to-novel generalization" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.358, + 0.893, + 0.41 + ], + "angle": 0, + "content": "Figure 1. Comparison of MaPLe with standard prompt learning methods. (a) Existing methods adopt uni-modal prompting techniques to fine-tune CLIP representations as prompts are learned only in a single branch of CLIP (language or vision). (b) MaPLe introduces branch-aware hierarchical prompts that adapt both language and vision branches simultaneously for improved generalization. (c) MaPLe surpasses state-of-the-art methods on 11 diverse image recognition datasets for novel class generalization task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.57 + ], + "angle": 0, + "content": "over the state-of-the-art method Co-CoOp [48]. Further, MaPLe demonstrates favorable generalization ability and robustness in cross-dataset transfer and domain generalization settings, leading to consistent improvements compared to existing approaches. Owing to its streamlined architectural design, MaPLe exhibits improved efficiency during both training and inference without much overhead, as compared to Co-CoOp which lacks efficiency due to its image instance conditioned design. In summary, the main contributions of this work include:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.572, + 0.47, + 0.627 + ], + "angle": 0, + "content": "- We propose multi-modal prompt learning in CLIP to favourably align its vision-language representations. To the best of our knowledge, this is the first multi-modal prompting approach for fine-tuning CLIP." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.631, + 0.47, + 0.7 + ], + "angle": 0, + "content": "- To link prompts learned in text and image encoders, we propose a coupling function to explicitly condition vision prompts on their language counterparts. It acts as a bridge between the two modalities and allows mutual propagation of gradients to promote synergy." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.703, + 0.47, + 0.801 + ], + "angle": 0, + "content": "- Our multi-modal prompts are learned across multiple transformer blocks in both vision and language branches to progressively learn the synergistic behaviour of both modalities. This deep prompting strategy allows modeling the contextual relationships independently, thus providing more flexibility to align the vision-language representations." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.572, + 0.47, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.816, + 0.219, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.868 + ], + "angle": 0, + "content": "Vision Language Models: The combined use of language supervision with natural images is found to be of great in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.724 + ], + "angle": 0, + "content": "terest in the computer vision community. In contrast to models learned with only image supervision, these vision-language (V-L) models encode rich multimodal representations. Recently, V-L models like CLIP [32], ALIGN [15], LiT [45], FILIP [41] and Florence [43] have demonstrated exceptional performance on a wide spectrum of tasks including few-shot and zero-shot visual recognition. These models learn joint image-language representations in a self-supervised manner using abundantly available data from the web. For example, CLIP and ALIGN respectively use \\(\\sim 400\\mathrm{M}\\) and \\(\\sim 1\\mathrm{B}\\) image-text pairs to train a multimodal network. Although these pre-trained V-L models learn generalized representations, efficiently adapting them to downstream tasks is still a challenging problem. Many works have demonstrated better performance on downstream tasks by using tailored methods to adapt V-L models for few-shot image-recognition [9, 19, 46], object detection [8, 10, 27, 34, 44, 50], and segmentation [5, 22, 26, 33]. In this work, we propose a novel multi-modal prompt learning technique to effectively adapt CLIP for few-shot and zero-shot visual recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.893, + 0.868 + ], + "angle": 0, + "content": "Prompt Learning: The instructions in the form of a sentence, known as text prompt, are usually given to the language branch of a V-L model, allowing it to better understand the task. Prompts can be handcrafted for a downstream task or learned automatically during fine-tuning stage. The latter is referred to as 'Prompt Learning' which was first used in NLP [21,23,24] followed by the adaptation in V-L [48, 49, 51] and vision-only [16, 38, 39, 47] models. Similar to [16] our design also uses deep 'vision' prompting. However, ours is the first multi-modal prompting de" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.507, + 0.899 + ], + "angle": 0, + "content": "19114" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.124, + 0.808, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed MaPLe (Multi-modal Prompt Learning) framework for prompt learning in V-L models. MaPLe tunes both vision and language branches where only the context prompts are learned, while the rest of the model is frozen. MaPLe conditions the vision prompts on language prompts via a V-L coupling function \\(\\mathcal{F}\\) to induce mutual synergy between the two modalities. Our framework uses deep contextual prompting where separate context prompts are learned across multiple transformer blocks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.273, + 0.43 + ], + "angle": 0, + "content": "sign while [16] is uni-modal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.806 + ], + "angle": 0, + "content": "Prompt Learning in Vision Language models: Full fine-tuning and linear probing [9] are two typical approaches to adapt a V-L model (i.e. CLIP) to the downstream tasks. The complete fine-tuning results in degrading the previously learned joint V-L representation while linear probing limits the zero-shot capability of CLIP. To this end, inspired from prompt learning in NLP, many works have proposed to adapt V-L models by learning the prompt tokens in an end-to-end training. CoOp [49] fine-tunes CLIP for few-shot transfer by optimizing continuous set of prompt vectors at its language branch. Co-CoOp [48] highlights the inferior performance of CoOp on novel classes and solves the generalization issue by explicitly conditioning prompts on image instances. [25] proposes to optimize multiple set of prompts by learning the distribution of prompts. [18] adapt CLIP by learning prompts for video understanding tasks. [1] perform visual prompt tuning on CLIP by prompting on the vision branch. We note that the existing methods follow independent uni-modal solutions and learn prompts either in the language or in the vision branch of CLIP, thus adapting CLIP partially. In this paper, we explore an important question: given the multimodal nature of CLIP, is complete prompting (i.e., in both language and vision branches) better suited to adapt CLIP? Our work is the first to answer this question by investigating the effectiveness of multi-modal prompt learning in order to improve alignment between vision and language representations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.17, + 0.831 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Our approach concerns with fine-tuning a pre-trained multimodal CLIP for better generalization to downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.653 + ], + "angle": 0, + "content": "through context optimization via prompting. Fig. 2 shows the overall architecture of our proposed MaPLe (Multimodal Prompt Learning) framework. Unlike previous approaches [48, 49] which learn context prompts only at the language branch, MaPLe proposes a joint prompting approach where the context prompts are learned in both vision and language branches. Specifically, we append learnable context tokens in the language branch and explicitly condition the vision prompts on the language prompts via a coupling function to establish interaction between them. To learn hierarchical contextual representations, we introduce deep prompting in both branches through separate learnable context prompts across different transformer blocks. During fine-tuning, only the context prompts along with their coupling function are learned while the rest of the model is frozen. Below, we first outline the pre-trained CLIP architecture and then present our proposed fine-tuning approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.663, + 0.696 + ], + "angle": 0, + "content": "3.1. Revisiting CLIP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.894, + 0.793 + ], + "angle": 0, + "content": "We build our approach on a pre-trained vision-language (V-L) model, CLIP, which consists of a text and vision encoder. Consistent with existing prompting methods [48, 49], we use a vision transformer (ViT) [6] based CLIP model. CLIP encodes an image \\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\) and a corresponding text description as explained below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.894, + 0.869 + ], + "angle": 0, + "content": "Encoding Image: Image encoder \\(\\mathcal{V}\\) with \\(K\\) transformer layers \\(\\{\\mathcal{V}_i\\}_{i=1}^K\\), splits the image \\(I\\) into \\(M\\) fixed-size patches which are projected into patch embeddings \\(E_0 \\in \\mathbb{R}^{M \\times d_v}\\). Patch embeddings \\(E_i\\) are input to the \\((i+1)^{\\mathrm{th}}\\) transformer block \\((\\mathcal{V}_{i+1})\\) along with a learnable class (CLS) token \\(\\mathbf{c}_i\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.507, + 0.899 + ], + "angle": 0, + "content": "19115" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.123, + 0.885, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.217, + 0.885, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.893, + 0.345 + ], + "angle": 0, + "content": "Figure 3. t-SNE plots of image embeddings in uni-modal prompting method Co-CoOp, and MaPLe on 3 diverse image recognition datasets. MaPLe shows better separability in both base and novel classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.462, + 0.383 + ], + "angle": 0, + "content": "and sequentially processed through \\(K\\) transformer blocks," + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.39, + 0.427, + 0.406 + ], + "angle": 0, + "content": "\\[\n[ \\mathbf {c} _ {i}, E _ {i} ] = \\mathcal {V} _ {i} ([ \\mathbf {c} _ {i - 1}, E _ {i - 1} ]) \\quad i = 1, 2, \\dots , K.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.456 + ], + "angle": 0, + "content": "To obtain the final image representation \\(x\\), the class token \\(\\mathsf{c}_K\\) of last transformer layer \\((\\mathcal{V}_K)\\) is projected to a common V-L latent embedding space via ImageProj," + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.463, + 0.396, + 0.48 + ], + "angle": 0, + "content": "\\[\nx = \\operatorname {I m a g e P r o j} \\left(\\mathrm {c} _ {K}\\right) \\quad x \\in \\mathbb {R} ^ {d _ {v l}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.571 + ], + "angle": 0, + "content": "Encoding Text: CLIP text encoder generates feature representations for text description by tokenizing the words and projecting them to word embeddings \\( W_{0} = [w_{0}^{1}, w_{0}^{2}, \\dots, w_{0}^{N}] \\in \\mathbb{R}^{N \\times d_{l}} \\). At each stage, \\( W_{i} \\) is input to the \\( (i + 1)^{\\mathrm{th}} \\) transformer layer of text encoding branch \\( (\\mathcal{L}_{i + 1}) \\)," + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.578, + 0.405, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\left[ W _ {i} \\right] = \\mathcal {L} _ {i} \\left(W _ {i - 1}\\right) \\quad i = 1, 2, \\dots , K.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.658 + ], + "angle": 0, + "content": "The final text representation \\(z\\) is obtained by projecting the text embeddings corresponding to the last token of the last transformer block \\(\\mathcal{L}_K\\) to a common V-L latent embedding space via TextProj," + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.664, + 0.393, + 0.681 + ], + "angle": 0, + "content": "\\[\nz = \\operatorname {T e x t P r o j} \\left(w _ {K} ^ {N}\\right) \\quad z \\in \\mathbb {R} ^ {d _ {v l}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Zero-shot Classification: For zero-shot classification, text prompts are hand-crafted with class labels \\( y \\in \\{1, 2, \\dots, C\\} \\) (e.g., 'a photo of a ) having \\( C \\) classes. Prediction \\( \\hat{y} \\) corresponding to the image \\( I \\) having the highest cosine similarity score \\( (sim(\\cdot)) \\) is calculated with a temperature parameter \\( \\tau \\)," + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.779, + 0.387, + 0.813 + ], + "angle": 0, + "content": "\\[\np (\\hat {y} | x) = \\frac {\\exp (s i m (x , z _ {\\hat {y}}) / \\tau)}{\\sum_ {i = 1} ^ {C} \\exp (s i m (x , z _ {i}))}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.414, + 0.833 + ], + "angle": 0, + "content": "3.2. MaPLe: Multi-modal Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.869 + ], + "angle": 0, + "content": "To efficiently fine-tune CLIP for downstream image recognition tasks, we explore the potential of multi-modal prompt" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.548 + ], + "angle": 0, + "content": "tuning. We reason that prior works that have predominantly explored uni-modal approaches are less suitable as they do not offer the flexibility to dynamically adapt both language and vision representation spaces. Thus to achieve completeness in prompting, we underline the importance of multimodal prompting approach. In Fig. 3, we visualize and compare the image embeddings of MaPLe with recent state-of-the-art work, Co-CoOp. Note that the image embeddings of CLIP, CoOp and Co-CoOp will be identical as they do not learn prompts in the vision branch. The visualization shows that image embeddings of MaPLe are more separable indicating that learning vision prompts in addition to language prompts leads to better adaptation of CLIP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.66 + ], + "angle": 0, + "content": "In addition to multi-modal prompting, we find that it is essential to learn prompts in the deeper transformer layers to progressively model stage-wise feature representations. To this end, we propose to introduce learnable tokens in the first \\( J \\) (where \\( J < K \\)) layers of both vision and language branches. These multi-modal hierarchical prompts utilize the knowledge embedded in CLIP model to effectively learn task relevant contextual representations (see Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.742, + 0.694 + ], + "angle": 0, + "content": "3.2.1 Deep Language Prompting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.8 + ], + "angle": 0, + "content": "To learn the language context prompts, we introduce \\( b \\) learnable tokens \\( \\{P^i \\in \\mathbb{R}^{d_l}\\}_{i=1}^b \\), in the language branch of CLIP. The input embeddings now follow the form \\( [P^1, P^2, \\dots, P^b, W_0] \\), where \\( W_0 = [w^1, w^2, \\dots, w^N] \\) corresponds to fixed input tokens. New learnable tokens are further introduced in each transformer block of the language encoder \\( (\\mathcal{L}_i) \\) up to a specific depth \\( J \\)," + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.811, + 0.892, + 0.827 + ], + "angle": 0, + "content": "\\[\n[ \\_ , W _ {i} ] = \\mathcal {L} _ {i} ([ P _ {i - 1}, W _ {i - 1} ]) \\quad i = 1, 2, \\dots , J. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Here \\([\\cdot, \\cdot]\\) refers to the concatenation operation. After \\(J^{\\mathrm{th}}\\) transformer layer, the subsequent layers process previous" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.507, + 0.899 + ], + "angle": 0, + "content": "19116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.457, + 0.14 + ], + "angle": 0, + "content": "layer prompts and final text representation \\(z\\) is computed," + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.15, + 0.468, + 0.184 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} [ P _ {j}, W _ {j} ] = \\mathcal {L} _ {j} ([ P _ {j - 1}, W _ {j - 1} ]) \\quad j = J + 1, \\dots , K, (2) \\\\ z = \\operatorname {T e x t P r o j} \\left(w _ {K} ^ {N}\\right). (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.193, + 0.469, + 0.236 + ], + "angle": 0, + "content": "When \\( J = 1 \\), the learnable tokens \\( P \\) are only applied at the input of first transformer layer, and this deep language prompting technique degenerates to CoOp [49]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.252, + 0.295, + 0.267 + ], + "angle": 0, + "content": "3.2.2 Deep Vision Prompting" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.273, + 0.469, + 0.343 + ], + "angle": 0, + "content": "Similar to deep language prompting, we introduce \\(b\\) learnable tokens \\(\\{\\hat{P}^i\\in \\mathbb{R}^{d_v}\\}_{i = 1}^b\\), in the vision branch of CLIP alongside the input image tokens. New learnable tokens are further introduced in deeper transformer layers of the image encoder \\((\\mathcal{V})\\) up to depth \\(J\\)." + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.353, + 0.469, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} [ c _ {i}, E _ {i}, \\_ ] = \\mathcal {V} _ {i} ([ c _ {i - 1}, E _ {i - 1}, \\bar {P} _ {i - 1} ]) \\quad i = 1, 2, \\dots , J, \\\\ [ c _ {j}, E _ {j}, \\tilde {P} _ {j} ] = \\mathcal {V} _ {j} ([ c _ {j - 1}, E _ {j - 1}, \\tilde {P} _ {j - 1} ]) \\quad j = J + 1, \\dots , K, \\\\ x = \\operatorname {I m a g e P r o j} \\left(c _ {K}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.511 + ], + "angle": 0, + "content": "Our deep prompting provides the flexibility to learn prompts across different feature hierarchies within the ViT architecture. We find that sharing prompts across stages is better compared to independent prompts as features are more correlated due to successive transformer block processing. Thus, the later stages do not provide independently-learned complimentary prompts as compared to the early stages." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.375, + 0.542 + ], + "angle": 0, + "content": "3.2.3 Vision Language Prompt Coupling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.469, + 0.715 + ], + "angle": 0, + "content": "We reason that in prompt tuning it is essential to take a multi-modal approach and simultaneously adapt both the vision and language branch of CLIP in order to achieve completeness in context optimization. A simple approach would be to naively combine deep vision and language prompting, where both the language prompts \\( P \\), and the vision prompts \\( \\tilde{P} \\), will be learned during the same training schedule. We name this design as 'Independent V-L Prompting'. Although this approach satisfies the requirement of completeness in prompting, this design lacks synergy between vision and language branch as both branches do not interact while learning the task relevant context prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.469, + 0.867 + ], + "angle": 0, + "content": "To this end, we propose a branch-aware multi-modal prompting which tunes vision and language branch of CLIP together by sharing prompts across both modalities. Language prompt tokens are introduced in the language branch up to \\( J^{\\mathrm{th}} \\) transformer block similar to deep language prompting as illustrated in Eqs. 1-3. To ensure mutual synergy between V-L prompts, vision prompts \\( \\tilde{P} \\), are obtained by projecting language prompts \\( P \\) via vision-to-language projection which we refer to as \\( V-L \\) coupling function \\( \\mathcal{F}(\\cdot) \\) such that \\( \\tilde{P}_k = \\mathcal{F}_k(P_k) \\). The coupling function is implemented as a linear layer which maps \\( d_l \\) dimensional inputs" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.126, + 0.892, + 0.154 + ], + "angle": 0, + "content": "to \\( d_v \\). This acts as a bridge between the two modalities, thus encouraging mutual propagation of gradients." + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.166, + 0.899, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} [ c _ {i}, E _ {i}, \\_ ] = \\mathcal {V} _ {i} ([ c _ {i - 1}, E _ {i - 1}, \\mathcal {F} _ {\\boldsymbol {i} - \\mathbf {1}} (\\boldsymbol {P} _ {\\boldsymbol {i} - \\mathbf {1}}) ]) i = 1, \\dots , J \\\\ [ c _ {j}, E _ {j}, \\tilde {P} _ {j} ] = \\mathcal {V} _ {j} ([ c _ {j - 1}, E _ {j - 1}, \\tilde {P} _ {j - 1} ]) \\quad j = J + 1, \\dots , K \\\\ x = \\operatorname {I m a g e P r o j} \\left(c _ {K}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.893, + 0.27 + ], + "angle": 0, + "content": "Unlike independent V-L prompting, explicit conditioning of \\(\\tilde{P}\\) on \\(P\\) helps learn prompts in a shared embedding space between the two branches, thus improving mutual synergy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.283, + 0.632, + 0.299 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.305, + 0.682, + 0.32 + ], + "angle": 0, + "content": "4.1. Benchmark setting" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.327, + 0.892, + 0.396 + ], + "angle": 0, + "content": "Generalization from Base-to-Novel Classes: We evaluate the generalizability of MaPLe, and follow a zero-shot setting where the datasets are split into base and novel classes. The model is trained only on the base classes in a few-shot setting and evaluated on base and novel categories." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.397, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Cross-dataset Evaluation: To validate the potential of our approach in cross-dataset transfer, we evaluate our ImageNet trained model directly on other datasets. Consistent with Co-CoOp, our model is trained on all 1000 ImageNet classes in a few-shot manner." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.466, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Domain Generalization: We evaluate the robustness of our method on out-of-distribution datasets. Similar to cross-dataset evaluation, we test our ImageNet trained model directly on four other ImageNet datasets that contain various types of domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.535, + 0.892, + 0.729 + ], + "angle": 0, + "content": "Datasets: For generalization from base-to-novel classes and cross-dataset evaluation, we follow [48, 49] and evaluate the performance of our method on 11 image classification datasets which covers a wide range of recognition tasks. This includes two generic-objects datasets, ImageNet [4] and Caltech101 [7]; five fine-grained datasets, OxfordPets [31], StanfordCars [20], Flowers102 [30], Food101 [2], and FGVCAircraft [28]; a scene recognition dataset SUN397 [40]; an action recognition dataset UCF101 [36]; a texture dataset DTD [3] and a satellite-image dataset EuroSAT [11]. For domain generalization, we use ImageNet as source dataset and its four variants as target datasets including ImageNetV2 [35], ImageNet-Sketch [37], ImageNet-A [13] and ImageNet-R [12]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.73, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Implementation Details We use a few-shot training strategy in all experiments at 16 shots which are randomly sampled for each class. We apply prompt tuning on a pretrained ViT-B/16 CLIP model where \\( d_{l} = 512 \\), \\( d_{v} = 768 \\) and \\( d_{vl} = 512 \\). For MaPLe, we set prompt depth \\( J \\) to 9 and the language and vision prompt lengths to 2. All models are trained for 5 epochs with a batch-size of 4 and a learning rate of 0.0035 via SGD optimizer on a single NVIDIA A100 GPU. We report base and novel class accuracies and their harmonic mean (HM) averaged over 3 runs. We initial-" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19117" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.123, + 0.468, + 0.205 + ], + "angle": 0, + "content": "
MethodBase Acc.Novel Acc.HMGFLOPS
1: MaPLe shallow (J=1)80.1073.5276.67167.1
2: Deep vision prompting80.2473.4376.6818.0
3: Deep language prompting81.7273.8177.56166.8
4: Independent V-L prompting82.1574.0777.90167.0
5: MaPLe (Ours)82.2875.1478.55167.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.208, + 0.471, + 0.247 + ], + "angle": 0, + "content": "Table 1. Comparison of MaPLe with different prompting designs in base-to-novel generalization. Results are averaged over 11 datasets. HM refers to harmonic mean." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.256, + 0.471, + 0.395 + ], + "angle": 0, + "content": "ize the language prompts of the first layer \\( P_0 \\) with the pretrained CLIP word embeddings of the template 'a photo of a ', while for the subsequent layers they are randomly initialized from a normal distribution. For training MaPLe on all 1000 classes of ImageNet as a source model, prompt depth \\( J \\) is set to 3 and the model trained for 2 epochs with learning rate of 0.0026. Hyper-parameters for deep language prompting, deep vision prompting, and independent V-L prompting are detailed in Appendix A. The hyper-parameters are fixed across all datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.402, + 0.47, + 0.417 + ], + "angle": 0, + "content": "4.2. Prompting CLIP via Vision-Language Prompts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.471, + 0.771 + ], + "angle": 0, + "content": "Prompting Variants: We first evaluate the performance of different possible prompting design choices as an ablation for our proposed branch-aware multi-modal prompting, MaPLe. These variants include shallow MaPLe, deep language prompting, deep vision prompting and independent V-L prompting. In Table 1, we present the results averaged over the 11 image recognition datasets. Shallow MaPLe (row-1) provides consistent improvements over CoOp and Co-CoOp in terms of generalization. Deep language prompting (row-3) shows improvements over deep vision prompting (row-2), indicating that prompts learned at the language branch provide better adaptation of CLIP. Although separately combining the above two approaches (row-4) further improves the performance, it struggles to achieve comprehensive benefits from the language and vision branches. We hypothesize that this is due to the lack of synergy between the learned vision and language prompts as they do not interact with each other during training. Meanwhile, MaPLe tied with deep prompting (row-4) combines the benefits of prompting in both branches by enforcing interactions through explicit conditioning of vision prompts on the language prompts. It provides improvements on novel and base class accuracies which leads to the best HM of \\(78.55\\%\\). We explore other possible design choices and present the ablations in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.777, + 0.34, + 0.79 + ], + "angle": 0, + "content": "4.3. Base-to-Novel Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.471, + 0.869 + ], + "angle": 0, + "content": "Generalization to Unseen Classes: Table 3 presents the performance of MaPLe in base-to-novel generalization setting on 11 recognition datasets. We compare its performance with CLIP zero-shot, and recent prompt learning works including CoOp [49] and Co-CoOp [48]. In case of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.126, + 0.892, + 0.153 + ], + "angle": 0, + "content": "CLIP, we use hand-crafted prompts that are specifically designed for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.893, + 0.292 + ], + "angle": 0, + "content": "In comparison with the state-of-the-art Co-CoOp, MaPLe shows improved performance on both base and novel categories on all 11 datasets with an exception of marginal reduction on only the base class performance of Caltech101. With mutual synergy from the branch-aware multi-modal prompting, MaPLe better generalizes to novel categories on all 11 datasets in comparison with Co-CoOp, and obtains an overall gain from \\(71.69\\%\\) to \\(75.14\\%\\). When taking into account both the base and novel classes, MaPLe shows an absolute average gain of \\(2.72\\%\\) over Co-CoOp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.893, + 0.499 + ], + "angle": 0, + "content": "In comparison with CLIP on novel classes, Co-CoOp improves only on 4/11 datasets dropping the average novel accuracy from \\(74.22\\%\\) to \\(71.69\\%\\). MaPLe is a strong competitor which improves accuracy over CLIP on novel classes on 6/11 datasets, with an average gain from \\(74.22\\%\\) to \\(75.14\\%\\). Generalization and Performance on Base Classes: CoCoOp solves the poor generalization problem in CoOp by conditioning prompts on image instances and shows significant gains in novel categories. However on base classes, it improves over CoOp only on 3/11 datasets with an average drop in performance from \\(82.69\\%\\) to \\(80.47\\%\\). Meanwhile, the completeness in prompting helps MaPLe improve over CoOp on base classes in 6/11 datasets maintaining the average base accuracy to around \\(82.28\\%\\), in addition to its improvement in generalization to novel classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.57 + ], + "angle": 0, + "content": "We find that the training strategies of Co-CoOp can be used to substantially boost the generalization performance of vanilla CoOp (6.8% gain in novel classes). We therefore compare our method with \\(\\mathrm{CoOp}^{\\dagger}\\), which trains CoOp in Co-CoOp setting (refer to Appendix A for more details)." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.578, + 0.851, + 0.658 + ], + "angle": 0, + "content": "
BaseNovelHM
CoOp82.6963.2271.66
Co-CoOp80.4771.6975.83
CoOp†80.8570.0275.04
MaPLe82.2875.1478.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.66, + 0.874, + 0.673 + ], + "angle": 0, + "content": "Table 2. Generalization comparison of MaPLe with CoOp†." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Compare to \\(\\mathrm{CoOp}^{\\dagger}\\), the vanilla CoOp model seems to overfit on base classes. When compared to \\(\\mathrm{CoOp}^{\\dagger}\\) which attains an average base accuracy of \\(80.85\\%\\), MaPLe shows an improvement of \\(1.43\\%\\) with the average base accuracy of \\(82.28\\%\\) (Table 2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.733, + 0.777 + ], + "angle": 0, + "content": "4.4. Cross-Dataset Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.893, + 0.869 + ], + "angle": 0, + "content": "We test the cross-dataset generalization ability of MaPLe by learning multi-modal prompts on all the 1000 ImageNet classes and then transferring it directly on the remaining 10 datasets. Table 4 shows the performance comparison between MaPLe, CoOp and Co-CoOp. On the ImageNet source dataset, MaPLe achieves performance comparable" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.507, + 0.899 + ], + "angle": 0, + "content": "19118" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.141, + 0.128, + 0.298, + 0.138 + ], + "angle": 0, + "content": "(a) Average over 11 datasets" + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.14, + 0.342, + 0.237 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP69.3474.2271.70
CoOp82.6963.2271.66
Co-CoOp80.4771.6975.83
MaPLe82.2875.1478.55
+1.81+3.45+2.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.444, + 0.128, + 0.519, + 0.138 + ], + "angle": 0, + "content": "(b) ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.36, + 0.14, + 0.605, + 0.237 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP72.4368.1470.22
CoOp76.4767.8871.92
Co-CoOp75.9870.4373.10
MaPLe76.6670.5473.47
+0.68+0.11+0.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.713, + 0.128, + 0.79, + 0.138 + ], + "angle": 0, + "content": "(c) Caltech101" + }, + { + "type": "table", + "bbox": [ + 0.631, + 0.14, + 0.874, + 0.237 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP96.8494.0095.40
CoOp98.0089.8193.73
Co-CoOp97.9693.8195.84
MaPLe97.7494.3696.02
-0.22+0.55+0.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.18, + 0.25, + 0.26, + 0.26 + ], + "angle": 0, + "content": "(d) OxfordPets" + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.262, + 0.342, + 0.358 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP91.1797.2694.12
CoOp93.6795.2994.47
Co-CoOp95.2097.6996.43
MaPLe95.4397.7696.58
+0.23+0.07+0.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.438, + 0.25, + 0.526, + 0.26 + ], + "angle": 0, + "content": "(e) StanfordCars" + }, + { + "type": "table", + "bbox": [ + 0.36, + 0.262, + 0.605, + 0.358 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP63.3774.8968.65
CoOp78.1260.4068.13
Co-CoOp70.4973.5972.01
MaPLe72.9474.0073.47
+2.45+0.41+1.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.713, + 0.25, + 0.791, + 0.26 + ], + "angle": 0, + "content": "(f) Flowers102" + }, + { + "type": "table", + "bbox": [ + 0.631, + 0.262, + 0.874, + 0.358 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP72.0877.8074.83
CoOp97.6059.6774.06
Co-CoOp94.8771.7581.71
MaPLe95.9272.4682.56
+1.05+0.71+0.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.186, + 0.372, + 0.251, + 0.382 + ], + "angle": 0, + "content": "(g) Food101" + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.383, + 0.341, + 0.48 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP90.1091.2290.66
CoOp88.3382.2685.19
Co-CoOp90.7091.2990.99
MaPLe90.7192.0591.38
+0.01+0.76+0.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.434, + 0.372, + 0.53, + 0.382 + ], + "angle": 0, + "content": "(h) FGVCAircraft" + }, + { + "type": "table", + "bbox": [ + 0.357, + 0.383, + 0.607, + 0.48 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP27.1936.2931.09
CoOp40.4422.3028.75
Co-CoOp33.4123.7127.74
MaPLe37.4435.6136.50
+4.03+11.90+8.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.72, + 0.372, + 0.783, + 0.382 + ], + "angle": 0, + "content": "(i) SUN397" + }, + { + "type": "table", + "bbox": [ + 0.631, + 0.383, + 0.874, + 0.48 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP69.3675.3572.23
CoOp80.6065.8972.51
Co-CoOp79.7476.8678.27
MaPLe80.8278.7079.75
+1.08+1.84+1.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.194, + 0.493, + 0.238, + 0.504 + ], + "angle": 0, + "content": "(j) DTD" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.505, + 0.337, + 0.601 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP53.2459.9056.37
CoOp79.4441.1854.24
Co-CoOp77.0156.0064.85
MaPLe80.3659.1868.16
+3.35+3.18+3.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.451, + 0.493, + 0.52, + 0.503 + ], + "angle": 0, + "content": "(k) EuroSAT" + }, + { + "type": "table", + "bbox": [ + 0.357, + 0.505, + 0.614, + 0.6 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP56.4864.0560.03
CoOp92.1954.7468.69
Co-CoOp87.4960.0471.21
MaPLe94.0773.2382.35
+6.58+13.19+11.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.724, + 0.493, + 0.785, + 0.503 + ], + "angle": 0, + "content": "(1)UCF101" + }, + { + "type": "table", + "bbox": [ + 0.634, + 0.505, + 0.877, + 0.6 + ], + "angle": 0, + "content": "
BaseNovelHM
CLIP70.5377.5073.85
CoOp84.6956.0567.46
Co-CoOp82.3373.4577.64
MaPLe83.0078.6680.77
+0.67+5.21+3.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.603, + 0.89, + 0.63 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art methods on base-to-novel generalization. MaPLe learns multi-modal prompts and demonstrates strong generalization results over existing methods on 11 recognition datasets. Absolute gains over Co-CoOp are indicated in blue." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.641, + 0.722, + 0.753 + ], + "angle": 0, + "content": "
SourceTarget
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101AircraftSUN397DTDEuroSATUCF101Average
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
Co-CoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
MaPLe70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.731, + 0.643, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Table 4. Comparison of MaPLe with existing approaches on cross-dataset evaluation. Overall, MaPLe achieves competitive performance providing highest average accuracy, indicating better generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.782, + 0.47, + 0.866 + ], + "angle": 0, + "content": "to competing approaches but demonstrates a much stronger generalization performance by surpassing CoOp in 9/10 and Co-CoOp in 8/10 datasets. Overall, MaPLe shows competitive performance leading to the highest averaged accuracy of \\(66.30\\%\\). This suggests that the use of branch-aware V-L prompting in MaPLe facilitates better generalization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.781, + 0.715, + 0.794 + ], + "angle": 0, + "content": "4.5. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.798, + 0.892, + 0.868 + ], + "angle": 0, + "content": "We show that MaPLe generalizes favourably on out-of-distribution datasets as compared to CoOp and Co-CoOp. We evaluate the direct transferability of ImageNet trained model to various out-of-domain datasets, and observe that it consistently improves against all the existing approaches" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19119" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.123, + 0.477, + 0.223 + ], + "angle": 0, + "content": "
SourceTarget
ImageNetImageNetV2ImageNet-SImageNet-AImageNet-R
CLIP66.7360.8346.1547.7773.96
CoOp71.5164.2047.9949.7175.21
Co-CoOp71.0264.0748.7550.6376.18
MaPLe70.7264.0749.1550.9076.98
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.224, + 0.471, + 0.263 + ], + "angle": 0, + "content": "Table 5. Comparison of MaPLe with existing approaches in domain generalization setting. MaPLe shows consistent improvements on all target datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.273, + 0.47, + 0.315 + ], + "angle": 0, + "content": "as indicated in Table 5. This indicates that utilizing multimodal branch-aware prompting helps MaPLe in enhancing the generalization and robustness of V-L models like CLIP." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.323, + 0.284, + 0.337 + ], + "angle": 0, + "content": "4.6. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Prompt Depth: In Fig. 4 (left), we illustrate the effect of prompt depth \\( J \\) for MaPLe and ablate on the depth of language and vision branch individually. In general, the performance improves as prompt depth increases. We note that performance sensitivity increases when randomly initialized prompts are inserted in the deeper layers of a frozen model where the model feature space is already mature. Similar trend is also reported by [16]. As earlier methods utilize shallow language prompting (\\( J = 1 \\)), we compare our method with deep language prompting. Overall, MaPLe achieves better performance than deep language prompting and achieves maximum performance at a depth of 9." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.58 + ], + "angle": 0, + "content": "Prompt Length: Fig. 4 (right) shows the effect of prompt length for MaPLe. As the prompt length increases, the performance on base classes is generally maintained, while the novel class accuracy decreases. This indicates over-fitting which inherently hurts the generalization to novel classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Effectiveness of Multi-modal Prompting: Fig. 5 shows the analysis of per class accuracy for selected datasets in the order of increasing domain shift. It indicates that the performance gains of MaPLe in comparison to Co-CoOp varies across different datasets. MaPLe provides significant gains over Co-CoOp for datasets that have large distribution shifts from the pretraining dataset of CLIP, and vision concepts that are usually rare and less generic. Further detailed analysis is provided in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.704, + 0.469, + 0.733 + ], + "angle": 0, + "content": "Prompting complexity: Table 6 shows the computational complexity of MaPLe in comparison with other approaches." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.738, + 0.277, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.738, + 0.468, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.863 + ], + "angle": 0, + "content": "Figure 4. Ablation on prompt depth (left) and prompt length (right) in MaPLe. We report average results on the held-out validation sets of all datasets." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.14, + 0.722, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.124, + 0.892, + 0.226 + ], + "angle": 0, + "content": "Figure 5. Percentage classes where MaPLe shows improved performance over Co-CoOp, which increases as dataset domain shift from generic categories increases \\((\\rightarrow)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.895, + 0.472 + ], + "angle": 0, + "content": "Although MaPLe utilizes multi-modal prompts, its overall FLOPS (Floating Point Operations) exceeds only by \\(0.1\\%\\) over CoOp and Co-CoOp. The independent V-L prompting also provides comparable FLOP count. In terms of inference speed, Co-CoOp is significantly slower and the FPS (Frames Per Second) remains constant as the batch size increases. In contrast, MaPLe has no such overhead and provides much better inference and training speeds. Further, MaPLe provides better convergence as it requires only half training epochs as compared to Co-CoOp (5 vs 10 epochs). MaPLe adds about \\(2.85\\%\\) training parameters on top of CLIP. To study if the performance gain is mainly attributed to more parameters, we experiment with \\(\\mathrm{MaPLe\\dagger}\\), which uses a unified V-L coupling function for all layer prompts. \\(\\mathrm{MaPLe\\dagger}\\) with about 9x lesser parameters than MaPLe also improves over existing methods. We also ablate by comparing MaPLe with heavier CoCoOp in Appendix D." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.482, + 0.894, + 0.582 + ], + "angle": 0, + "content": "
MethodParamsParams % CLIPFPS (with BS)HM
14100
CoOp20480.00213.855.31353.071.66
CoCoOp353600.0364.6114.715.175.83
Independent V-L314880.0262.5239.41383.877.90
MaPLe3.55 M2.8560.2239.01365.178.55
MaPLe†0.41 M0.3360.2238.01365.078.11
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.584, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Table 6. Comparison of computational complexity among different prompting methods. MaPLe† is a MaPLe version which utilizes a common V-L coupling function for all layers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.62, + 0.652 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.868 + ], + "angle": 0, + "content": "Adaptation of large-scale V-L models, e.g., CLIP [32] to downstream tasks is a challenging problem due to the large number of tunable parameters and limited size of downstream datasets. Prompt learning is an efficient and scalable technique to tailor V-L models to novel downstream tasks. To this end, the current prompt learning approaches either consider only the vision or language side prompting. Our work shows that it is critical to perform prompting for both vision and language branches to appropriately adapt V-L models to downstream tasks. Further, we propose a strategy to ensure synergy between vision-language modalities by explicitly conditioning the vision prompts on textual prompt across different transformer stages. Our approach improves the generalization towards novel categories, cross-dataset transfer and datasets with domain shifts." + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.507, + 0.899 + ], + "angle": 0, + "content": "19120" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.124, + 0.176, + 0.14 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 2022. 3, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.252 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In The European Conference on Computer Vision, pages 446-461. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.254, + 0.47, + 0.316 + ], + "angle": 0, + "content": "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.384, + 0.472, + 0.435 + ], + "angle": 0, + "content": "[5] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.437, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[7] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.581, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[8] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In The European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.646, + 0.47, + 0.697 + ], + "angle": 0, + "content": "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.699, + 0.47, + 0.748 + ], + "angle": 0, + "content": "[10] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.751, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.47, + 0.868 + ], + "angle": 0, + "content": "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.148, + 0.472, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.127, + 0.892, + 0.153 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.154, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[13] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.243 + ], + "angle": 0, + "content": "[14] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.245, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In The European Conference on Computer Vision, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.374, + 0.892, + 0.436 + ], + "angle": 0, + "content": "[17] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt is worth millions of parameters? low-resource prompt-based learning for vision-language models. arXiv preprint arXiv:2110.08484, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.439, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[18] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In The European Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[19] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.581 + ], + "angle": 0, + "content": "[20] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 554–561, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[21] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In _Conference on Empirical Methods in Natural Language Processing_, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[22] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.75 + ], + "angle": 0, + "content": "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.752, + 0.892, + 0.839 + ], + "angle": 0, + "content": "[24] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.892, + 0.868 + ], + "angle": 0, + "content": "[25] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.468, + 0.888, + 0.505, + 0.9 + ], + "angle": 0, + "content": "19121" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.127, + 0.47, + 0.152 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.153, + 0.471, + 0.203 + ], + "angle": 0, + "content": "[26] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.204, + 0.471, + 0.267 + ], + "angle": 0, + "content": "[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In The European Conference on Computer Vision. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.268, + 0.471, + 0.317 + ], + "angle": 0, + "content": "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.318, + 0.471, + 0.369 + ], + "angle": 0, + "content": "[29] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.37, + 0.471, + 0.42 + ], + "angle": 0, + "content": "[30] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.421, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[31] Omkar M Parkhi, Andreae Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.472, + 0.471, + 0.548 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.549, + 0.471, + 0.623 + ], + "angle": 0, + "content": "[33] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.625, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[34] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.471, + 0.739 + ], + "angle": 0, + "content": "[35] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.74, + 0.471, + 0.777 + ], + "angle": 0, + "content": "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.779, + 0.471, + 0.828 + ], + "angle": 0, + "content": "[37] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, volume 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.829, + 0.471, + 0.868 + ], + "angle": 0, + "content": "[38] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.127, + 0.471, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.127, + 0.892, + 0.152 + ], + "angle": 0, + "content": "prompting for rehearsal-free continual learning. In The European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.153, + 0.892, + 0.227 + ], + "angle": 0, + "content": "[39] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.229, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[40] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[41] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[42] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.394, + 0.892, + 0.456 + ], + "angle": 0, + "content": "[43] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.506 + ], + "angle": 0, + "content": "[44] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In The European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.508, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[45] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[46] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In The European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.635, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[47] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.722 + ], + "angle": 0, + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.724, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, pages 1-12, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.825 + ], + "angle": 0, + "content": "[50] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In The European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.826, + 0.892, + 0.863 + ], + "angle": 0, + "content": "[51] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. arXiv preprint arXiv:2205.14865, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.863 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19122" + } + ] +] \ No newline at end of file diff --git a/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_origin.pdf b/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..80873749029cb4c39dd0a2e1ac7ee4b575a11074 --- /dev/null +++ b/2023/MaPLe_ Multi-Modal Prompt Learning/9c2b1d58-6fd3-4d3d-b52c-e47be34ef7bd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf2a2b7ddb03a404ed1ef641461953b53c5a66cdf0d31e5e6de25f64b237345d +size 1092630 diff --git a/2023/MaPLe_ Multi-Modal Prompt Learning/full.md b/2023/MaPLe_ Multi-Modal Prompt Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..19329f5be8cd1835ca5e99be0ff590a91fbf245d --- /dev/null +++ b/2023/MaPLe_ Multi-Modal Prompt Learning/full.md @@ -0,0 +1,349 @@ +# MaPLe: Multi-modal Prompt Learning + +Muhammad Uzair Khattak1 Hanoona Rasheed1 Muhammad Maaz1 + +Salman Khan $^{1,2}$ Fahad Shahbaz Khan $^{1,3}$ + +$^{1}$ Mohamed bin Zayed University of AI $^{2}$ Australian National University $^{3}$ Linköping University + +# Abstract + +Pre-trained vision-language (V-L) models such as CLIP have shown excellent generalization ability to downstream tasks. However, they are sensitive to the choice of input text prompts and require careful selection of prompt templates to perform well. Inspired by the Natural Language Processing (NLP) literature, recent CLIP adaptation approaches learn prompts as the textual inputs to fine-tune CLIP for downstream tasks. We note that using prompting to adapt representations in a single branch of CLIP (language or vision) is sub-optimal since it does not allow the flexibility to dynamically adjust both representation spaces on a downstream task. In this work, we propose Multi-modal Prompt Learning (MaPLe) for both vision and language branches to improve alignment between the vision and language representations. Our design promotes strong coupling between the vision-language prompts to ensure mutual synergy and discourages learning independent uni-modal solutions. Further, we learn separate prompts across different early stages to progressively model the stage-wise feature relationships to allow rich context learning. We evaluate the effectiveness of our approach on three representative tasks of generalization to novel classes, new target datasets and unseen domain shifts. Compared with the state-of-the-art method CoCoOp, MaPLe exhibits favorable performance and achieves an absolute gain of $3.45\%$ on novel classes and $2.72\%$ on overall harmonic-mean, averaged over 11 diverse image recognition datasets. Our code and pre-trained models are available at https://github.com/muzairkhattak/multimodal-prompt-learning. + +# 1. Introduction + +Foundational vision-language (V-L) models such as CLIP (Contrastive Language-Image Pretraining) [32] have shown excellent generalization ability to downstream tasks. Such models are trained to align language and vision modalities on web-scale data e.g., 400 million text-image pairs in CLIP. These models can reason about open-vocabulary visual concepts, thanks to the rich supervision provided by + +natural language. During inference, hand-engineered text prompts are used e.g., 'a photo of a ' as a query for text encoder. The output text embeddings are matched with the visual embeddings from an image encoder to predict the output class. Designing high quality contextual prompts have been proven to enhance the performance of CLIP and other V-L models [17, 42]. + +Despite the effectiveness of CLIP towards generalization to new concepts, its massive scale and scarcity of training data (e.g., few-shot setting) makes it infeasible to fine-tune the full model for downstream tasks. Such fine-tuning can also forget the useful knowledge acquired in the large-scale pretraining phase and can pose a risk of overfitting to the downstream task. To address the above challenges, existing works propose language prompt learning to avoid manually adjusting the prompt templates and providing a mechanism to adapt the model while keeping the original weights frozen [14, 25, 29, 48, 49]. Inspired from Natural Language Processing (NLP), these approaches only explore prompt learning for the text encoder in CLIP (Fig. 1:a) while adaptation choices together with an equally important image encoder of CLIP remains an unexplored topic in the literature. + +Our motivation derives from the multi-modal nature of CLIP, where a text and image encoder co-exist and both contribute towards properly aligning the V-L modalities. We argue that any prompting technique should adapt the model completely and therefore, learning prompts only for the text encoder in CLIP is not sufficient to model the adaptations needed for the image encoder. To this end, we set out to achieve completeness in the prompting approach and propose Multi-modal Prompt Learning (MaPLe) to adequately fine-tune the text and image encoder representations such that their optimal alignment can be achieved on the downstream tasks (Fig. 1:b). Our extensive experiments on three key representative settings including base-to-novel generalization, cross-dataset evaluation, and domain generalization demonstrate the strength of MaPLe. On base-to-novel generalization, our proposed MaPLe outperforms existing prompt learning approaches across 11 diverse image recognition datasets (Fig. 1:c) and achieves absolute average gain of $3.45\%$ on novel classes and $2.72\%$ on harmonic-mean + +![](images/9eb4884d9989b772b613fc7472cdd8830703485c6d51fda6ac3498f906878816.jpg) +(a) Existing prompt tuning methods (Uni-modal) + +![](images/4d537288361b8997d134adeff410f0017bea0bb9fef2994c446047f6a5f32525.jpg) +(b) Multi-modal Prompt Learning (MaPLe) +Figure 1. Comparison of MaPLe with standard prompt learning methods. (a) Existing methods adopt uni-modal prompting techniques to fine-tune CLIP representations as prompts are learned only in a single branch of CLIP (language or vision). (b) MaPLe introduces branch-aware hierarchical prompts that adapt both language and vision branches simultaneously for improved generalization. (c) MaPLe surpasses state-of-the-art methods on 11 diverse image recognition datasets for novel class generalization task. + +![](images/ee6a47f15551f34258f8ea90ebea1f4099db62ca677e5062c8af79aa1b9dd4ed.jpg) +(c) Performance comparison on base-to-novel generalization + +over the state-of-the-art method Co-CoOp [48]. Further, MaPLe demonstrates favorable generalization ability and robustness in cross-dataset transfer and domain generalization settings, leading to consistent improvements compared to existing approaches. Owing to its streamlined architectural design, MaPLe exhibits improved efficiency during both training and inference without much overhead, as compared to Co-CoOp which lacks efficiency due to its image instance conditioned design. In summary, the main contributions of this work include: + +- We propose multi-modal prompt learning in CLIP to favourably align its vision-language representations. To the best of our knowledge, this is the first multi-modal prompting approach for fine-tuning CLIP. +- To link prompts learned in text and image encoders, we propose a coupling function to explicitly condition vision prompts on their language counterparts. It acts as a bridge between the two modalities and allows mutual propagation of gradients to promote synergy. +- Our multi-modal prompts are learned across multiple transformer blocks in both vision and language branches to progressively learn the synergistic behaviour of both modalities. This deep prompting strategy allows modeling the contextual relationships independently, thus providing more flexibility to align the vision-language representations. + +# 2. Related Work + +Vision Language Models: The combined use of language supervision with natural images is found to be of great in + +terest in the computer vision community. In contrast to models learned with only image supervision, these vision-language (V-L) models encode rich multimodal representations. Recently, V-L models like CLIP [32], ALIGN [15], LiT [45], FILIP [41] and Florence [43] have demonstrated exceptional performance on a wide spectrum of tasks including few-shot and zero-shot visual recognition. These models learn joint image-language representations in a self-supervised manner using abundantly available data from the web. For example, CLIP and ALIGN respectively use $\sim 400\mathrm{M}$ and $\sim 1\mathrm{B}$ image-text pairs to train a multimodal network. Although these pre-trained V-L models learn generalized representations, efficiently adapting them to downstream tasks is still a challenging problem. Many works have demonstrated better performance on downstream tasks by using tailored methods to adapt V-L models for few-shot image-recognition [9, 19, 46], object detection [8, 10, 27, 34, 44, 50], and segmentation [5, 22, 26, 33]. In this work, we propose a novel multi-modal prompt learning technique to effectively adapt CLIP for few-shot and zero-shot visual recognition tasks. + +Prompt Learning: The instructions in the form of a sentence, known as text prompt, are usually given to the language branch of a V-L model, allowing it to better understand the task. Prompts can be handcrafted for a downstream task or learned automatically during fine-tuning stage. The latter is referred to as 'Prompt Learning' which was first used in NLP [21,23,24] followed by the adaptation in V-L [48, 49, 51] and vision-only [16, 38, 39, 47] models. Similar to [16] our design also uses deep 'vision' prompting. However, ours is the first multi-modal prompting de + +![](images/93f706335d0a31f607a7e9c15f96265303f5203fafdf9d9584b52a0912f4d9db.jpg) +Figure 2. Overview of our proposed MaPLe (Multi-modal Prompt Learning) framework for prompt learning in V-L models. MaPLe tunes both vision and language branches where only the context prompts are learned, while the rest of the model is frozen. MaPLe conditions the vision prompts on language prompts via a V-L coupling function $\mathcal{F}$ to induce mutual synergy between the two modalities. Our framework uses deep contextual prompting where separate context prompts are learned across multiple transformer blocks. + +sign while [16] is uni-modal. + +Prompt Learning in Vision Language models: Full fine-tuning and linear probing [9] are two typical approaches to adapt a V-L model (i.e. CLIP) to the downstream tasks. The complete fine-tuning results in degrading the previously learned joint V-L representation while linear probing limits the zero-shot capability of CLIP. To this end, inspired from prompt learning in NLP, many works have proposed to adapt V-L models by learning the prompt tokens in an end-to-end training. CoOp [49] fine-tunes CLIP for few-shot transfer by optimizing continuous set of prompt vectors at its language branch. Co-CoOp [48] highlights the inferior performance of CoOp on novel classes and solves the generalization issue by explicitly conditioning prompts on image instances. [25] proposes to optimize multiple set of prompts by learning the distribution of prompts. [18] adapt CLIP by learning prompts for video understanding tasks. [1] perform visual prompt tuning on CLIP by prompting on the vision branch. We note that the existing methods follow independent uni-modal solutions and learn prompts either in the language or in the vision branch of CLIP, thus adapting CLIP partially. In this paper, we explore an important question: given the multimodal nature of CLIP, is complete prompting (i.e., in both language and vision branches) better suited to adapt CLIP? Our work is the first to answer this question by investigating the effectiveness of multi-modal prompt learning in order to improve alignment between vision and language representations. + +# 3. Method + +Our approach concerns with fine-tuning a pre-trained multimodal CLIP for better generalization to downstream tasks + +through context optimization via prompting. Fig. 2 shows the overall architecture of our proposed MaPLe (Multimodal Prompt Learning) framework. Unlike previous approaches [48, 49] which learn context prompts only at the language branch, MaPLe proposes a joint prompting approach where the context prompts are learned in both vision and language branches. Specifically, we append learnable context tokens in the language branch and explicitly condition the vision prompts on the language prompts via a coupling function to establish interaction between them. To learn hierarchical contextual representations, we introduce deep prompting in both branches through separate learnable context prompts across different transformer blocks. During fine-tuning, only the context prompts along with their coupling function are learned while the rest of the model is frozen. Below, we first outline the pre-trained CLIP architecture and then present our proposed fine-tuning approach. + +# 3.1. Revisiting CLIP + +We build our approach on a pre-trained vision-language (V-L) model, CLIP, which consists of a text and vision encoder. Consistent with existing prompting methods [48, 49], we use a vision transformer (ViT) [6] based CLIP model. CLIP encodes an image $I \in \mathbb{R}^{H \times W \times 3}$ and a corresponding text description as explained below. + +Encoding Image: Image encoder $\mathcal{V}$ with $K$ transformer layers $\{\mathcal{V}_i\}_{i=1}^K$ , splits the image $I$ into $M$ fixed-size patches which are projected into patch embeddings $E_0 \in \mathbb{R}^{M \times d_v}$ . Patch embeddings $E_i$ are input to the $(i+1)^{\mathrm{th}}$ transformer block $(\mathcal{V}_{i+1})$ along with a learnable class (CLS) token $\mathbf{c}_i$ + +![](images/e16ab2a8f4b922761701d24185fa40f44ed8b6343440e0896944f5853bd689a2.jpg) + +![](images/61c0a1218a8eac098d5c354f62e6fcf5de7744de687314db5a8701ba0d7aeaca.jpg) +Figure 3. t-SNE plots of image embeddings in uni-modal prompting method Co-CoOp, and MaPLe on 3 diverse image recognition datasets. MaPLe shows better separability in both base and novel classes. + +and sequentially processed through $K$ transformer blocks, + +$$ +[ \mathbf {c} _ {i}, E _ {i} ] = \mathcal {V} _ {i} ([ \mathbf {c} _ {i - 1}, E _ {i - 1} ]) \quad i = 1, 2, \dots , K. +$$ + +To obtain the final image representation $x$ , the class token $\mathsf{c}_K$ of last transformer layer $(\mathcal{V}_K)$ is projected to a common V-L latent embedding space via ImageProj, + +$$ +x = \operatorname {I m a g e P r o j} \left(\mathrm {c} _ {K}\right) \quad x \in \mathbb {R} ^ {d _ {v l}}. +$$ + +Encoding Text: CLIP text encoder generates feature representations for text description by tokenizing the words and projecting them to word embeddings $W_{0} = [w_{0}^{1}, w_{0}^{2}, \dots, w_{0}^{N}] \in \mathbb{R}^{N \times d_{l}}$ . At each stage, $W_{i}$ is input to the $(i + 1)^{\mathrm{th}}$ transformer layer of text encoding branch $(\mathcal{L}_{i + 1})$ , + +$$ +\left[ W _ {i} \right] = \mathcal {L} _ {i} \left(W _ {i - 1}\right) \quad i = 1, 2, \dots , K. +$$ + +The final text representation $z$ is obtained by projecting the text embeddings corresponding to the last token of the last transformer block $\mathcal{L}_K$ to a common V-L latent embedding space via TextProj, + +$$ +z = \operatorname {T e x t P r o j} \left(w _ {K} ^ {N}\right) \quad z \in \mathbb {R} ^ {d _ {v l}}. +$$ + +Zero-shot Classification: For zero-shot classification, text prompts are hand-crafted with class labels $y \in \{1, 2, \dots, C\}$ (e.g., 'a photo of a ) having $C$ classes. Prediction $\hat{y}$ corresponding to the image $I$ having the highest cosine similarity score $(sim(\cdot))$ is calculated with a temperature parameter $\tau$ , + +$$ +p (\hat {y} | x) = \frac {\exp (s i m (x , z _ {\hat {y}}) / \tau)}{\sum_ {i = 1} ^ {C} \exp (s i m (x , z _ {i}))}. +$$ + +# 3.2. MaPLe: Multi-modal Prompt Learning + +To efficiently fine-tune CLIP for downstream image recognition tasks, we explore the potential of multi-modal prompt + +tuning. We reason that prior works that have predominantly explored uni-modal approaches are less suitable as they do not offer the flexibility to dynamically adapt both language and vision representation spaces. Thus to achieve completeness in prompting, we underline the importance of multimodal prompting approach. In Fig. 3, we visualize and compare the image embeddings of MaPLe with recent state-of-the-art work, Co-CoOp. Note that the image embeddings of CLIP, CoOp and Co-CoOp will be identical as they do not learn prompts in the vision branch. The visualization shows that image embeddings of MaPLe are more separable indicating that learning vision prompts in addition to language prompts leads to better adaptation of CLIP. + +In addition to multi-modal prompting, we find that it is essential to learn prompts in the deeper transformer layers to progressively model stage-wise feature representations. To this end, we propose to introduce learnable tokens in the first $J$ (where $J < K$ ) layers of both vision and language branches. These multi-modal hierarchical prompts utilize the knowledge embedded in CLIP model to effectively learn task relevant contextual representations (see Fig. 4). + +# 3.2.1 Deep Language Prompting + +To learn the language context prompts, we introduce $b$ learnable tokens $\{P^i \in \mathbb{R}^{d_l}\}_{i=1}^b$ , in the language branch of CLIP. The input embeddings now follow the form $[P^1, P^2, \dots, P^b, W_0]$ , where $W_0 = [w^1, w^2, \dots, w^N]$ corresponds to fixed input tokens. New learnable tokens are further introduced in each transformer block of the language encoder $(\mathcal{L}_i)$ up to a specific depth $J$ , + +$$ +[ \_ , W _ {i} ] = \mathcal {L} _ {i} ([ P _ {i - 1}, W _ {i - 1} ]) \quad i = 1, 2, \dots , J. \tag {1} +$$ + +Here $[\cdot, \cdot]$ refers to the concatenation operation. After $J^{\mathrm{th}}$ transformer layer, the subsequent layers process previous + +layer prompts and final text representation $z$ is computed, + +$$ +\begin{array}{l} [ P _ {j}, W _ {j} ] = \mathcal {L} _ {j} ([ P _ {j - 1}, W _ {j - 1} ]) \quad j = J + 1, \dots , K, (2) \\ z = \operatorname {T e x t P r o j} \left(w _ {K} ^ {N}\right). (3) \\ \end{array} +$$ + +When $J = 1$ , the learnable tokens $P$ are only applied at the input of first transformer layer, and this deep language prompting technique degenerates to CoOp [49]. + +# 3.2.2 Deep Vision Prompting + +Similar to deep language prompting, we introduce $b$ learnable tokens $\{\hat{P}^i\in \mathbb{R}^{d_v}\}_{i = 1}^b$ , in the vision branch of CLIP alongside the input image tokens. New learnable tokens are further introduced in deeper transformer layers of the image encoder $(\mathcal{V})$ up to depth $J$ . + +$$ +\begin{array}{l} [ c _ {i}, E _ {i}, \_ ] = \mathcal {V} _ {i} ([ c _ {i - 1}, E _ {i - 1}, \bar {P} _ {i - 1} ]) \quad i = 1, 2, \dots , J, \\ [ c _ {j}, E _ {j}, \tilde {P} _ {j} ] = \mathcal {V} _ {j} ([ c _ {j - 1}, E _ {j - 1}, \tilde {P} _ {j - 1} ]) \quad j = J + 1, \dots , K, \\ x = \operatorname {I m a g e P r o j} \left(c _ {K}\right). \\ \end{array} +$$ + +Our deep prompting provides the flexibility to learn prompts across different feature hierarchies within the ViT architecture. We find that sharing prompts across stages is better compared to independent prompts as features are more correlated due to successive transformer block processing. Thus, the later stages do not provide independently-learned complimentary prompts as compared to the early stages. + +# 3.2.3 Vision Language Prompt Coupling + +We reason that in prompt tuning it is essential to take a multi-modal approach and simultaneously adapt both the vision and language branch of CLIP in order to achieve completeness in context optimization. A simple approach would be to naively combine deep vision and language prompting, where both the language prompts $P$ , and the vision prompts $\tilde{P}$ , will be learned during the same training schedule. We name this design as 'Independent V-L Prompting'. Although this approach satisfies the requirement of completeness in prompting, this design lacks synergy between vision and language branch as both branches do not interact while learning the task relevant context prompts. + +To this end, we propose a branch-aware multi-modal prompting which tunes vision and language branch of CLIP together by sharing prompts across both modalities. Language prompt tokens are introduced in the language branch up to $J^{\mathrm{th}}$ transformer block similar to deep language prompting as illustrated in Eqs. 1-3. To ensure mutual synergy between V-L prompts, vision prompts $\tilde{P}$ , are obtained by projecting language prompts $P$ via vision-to-language projection which we refer to as $V-L$ coupling function $\mathcal{F}(\cdot)$ such that $\tilde{P}_k = \mathcal{F}_k(P_k)$ . The coupling function is implemented as a linear layer which maps $d_l$ dimensional inputs + +to $d_v$ . This acts as a bridge between the two modalities, thus encouraging mutual propagation of gradients. + +$$ +\begin{array}{l} [ c _ {i}, E _ {i}, \_ ] = \mathcal {V} _ {i} ([ c _ {i - 1}, E _ {i - 1}, \mathcal {F} _ {\boldsymbol {i} - \mathbf {1}} (\boldsymbol {P} _ {\boldsymbol {i} - \mathbf {1}}) ]) i = 1, \dots , J \\ [ c _ {j}, E _ {j}, \tilde {P} _ {j} ] = \mathcal {V} _ {j} ([ c _ {j - 1}, E _ {j - 1}, \tilde {P} _ {j - 1} ]) \quad j = J + 1, \dots , K \\ x = \operatorname {I m a g e P r o j} \left(c _ {K}\right) \\ \end{array} +$$ + +Unlike independent V-L prompting, explicit conditioning of $\tilde{P}$ on $P$ helps learn prompts in a shared embedding space between the two branches, thus improving mutual synergy. + +# 4. Experiments + +# 4.1. Benchmark setting + +Generalization from Base-to-Novel Classes: We evaluate the generalizability of MaPLe, and follow a zero-shot setting where the datasets are split into base and novel classes. The model is trained only on the base classes in a few-shot setting and evaluated on base and novel categories. + +Cross-dataset Evaluation: To validate the potential of our approach in cross-dataset transfer, we evaluate our ImageNet trained model directly on other datasets. Consistent with Co-CoOp, our model is trained on all 1000 ImageNet classes in a few-shot manner. + +Domain Generalization: We evaluate the robustness of our method on out-of-distribution datasets. Similar to cross-dataset evaluation, we test our ImageNet trained model directly on four other ImageNet datasets that contain various types of domain shifts. + +Datasets: For generalization from base-to-novel classes and cross-dataset evaluation, we follow [48, 49] and evaluate the performance of our method on 11 image classification datasets which covers a wide range of recognition tasks. This includes two generic-objects datasets, ImageNet [4] and Caltech101 [7]; five fine-grained datasets, OxfordPets [31], StanfordCars [20], Flowers102 [30], Food101 [2], and FGVCAircraft [28]; a scene recognition dataset SUN397 [40]; an action recognition dataset UCF101 [36]; a texture dataset DTD [3] and a satellite-image dataset EuroSAT [11]. For domain generalization, we use ImageNet as source dataset and its four variants as target datasets including ImageNetV2 [35], ImageNet-Sketch [37], ImageNet-A [13] and ImageNet-R [12]. + +Implementation Details We use a few-shot training strategy in all experiments at 16 shots which are randomly sampled for each class. We apply prompt tuning on a pretrained ViT-B/16 CLIP model where $d_{l} = 512$ , $d_{v} = 768$ and $d_{vl} = 512$ . For MaPLe, we set prompt depth $J$ to 9 and the language and vision prompt lengths to 2. All models are trained for 5 epochs with a batch-size of 4 and a learning rate of 0.0035 via SGD optimizer on a single NVIDIA A100 GPU. We report base and novel class accuracies and their harmonic mean (HM) averaged over 3 runs. We initial- + +
MethodBase Acc.Novel Acc.HMGFLOPS
1: MaPLe shallow (J=1)80.1073.5276.67167.1
2: Deep vision prompting80.2473.4376.6818.0
3: Deep language prompting81.7273.8177.56166.8
4: Independent V-L prompting82.1574.0777.90167.0
5: MaPLe (Ours)82.2875.1478.55167.0
+ +ize the language prompts of the first layer $P_0$ with the pretrained CLIP word embeddings of the template 'a photo of a ', while for the subsequent layers they are randomly initialized from a normal distribution. For training MaPLe on all 1000 classes of ImageNet as a source model, prompt depth $J$ is set to 3 and the model trained for 2 epochs with learning rate of 0.0026. Hyper-parameters for deep language prompting, deep vision prompting, and independent V-L prompting are detailed in Appendix A. The hyper-parameters are fixed across all datasets. + +# 4.2. Prompting CLIP via Vision-Language Prompts + +Prompting Variants: We first evaluate the performance of different possible prompting design choices as an ablation for our proposed branch-aware multi-modal prompting, MaPLe. These variants include shallow MaPLe, deep language prompting, deep vision prompting and independent V-L prompting. In Table 1, we present the results averaged over the 11 image recognition datasets. Shallow MaPLe (row-1) provides consistent improvements over CoOp and Co-CoOp in terms of generalization. Deep language prompting (row-3) shows improvements over deep vision prompting (row-2), indicating that prompts learned at the language branch provide better adaptation of CLIP. Although separately combining the above two approaches (row-4) further improves the performance, it struggles to achieve comprehensive benefits from the language and vision branches. We hypothesize that this is due to the lack of synergy between the learned vision and language prompts as they do not interact with each other during training. Meanwhile, MaPLe tied with deep prompting (row-4) combines the benefits of prompting in both branches by enforcing interactions through explicit conditioning of vision prompts on the language prompts. It provides improvements on novel and base class accuracies which leads to the best HM of $78.55\%$ . We explore other possible design choices and present the ablations in Appendix B. + +# 4.3. Base-to-Novel Generalization + +Generalization to Unseen Classes: Table 3 presents the performance of MaPLe in base-to-novel generalization setting on 11 recognition datasets. We compare its performance with CLIP zero-shot, and recent prompt learning works including CoOp [49] and Co-CoOp [48]. In case of + +CLIP, we use hand-crafted prompts that are specifically designed for each dataset. + +In comparison with the state-of-the-art Co-CoOp, MaPLe shows improved performance on both base and novel categories on all 11 datasets with an exception of marginal reduction on only the base class performance of Caltech101. With mutual synergy from the branch-aware multi-modal prompting, MaPLe better generalizes to novel categories on all 11 datasets in comparison with Co-CoOp, and obtains an overall gain from $71.69\%$ to $75.14\%$ . When taking into account both the base and novel classes, MaPLe shows an absolute average gain of $2.72\%$ over Co-CoOp. + +In comparison with CLIP on novel classes, Co-CoOp improves only on 4/11 datasets dropping the average novel accuracy from $74.22\%$ to $71.69\%$ . MaPLe is a strong competitor which improves accuracy over CLIP on novel classes on 6/11 datasets, with an average gain from $74.22\%$ to $75.14\%$ . Generalization and Performance on Base Classes: CoCoOp solves the poor generalization problem in CoOp by conditioning prompts on image instances and shows significant gains in novel categories. However on base classes, it improves over CoOp only on 3/11 datasets with an average drop in performance from $82.69\%$ to $80.47\%$ . Meanwhile, the completeness in prompting helps MaPLe improve over CoOp on base classes in 6/11 datasets maintaining the average base accuracy to around $82.28\%$ , in addition to its improvement in generalization to novel classes. + +We find that the training strategies of Co-CoOp can be used to substantially boost the generalization performance of vanilla CoOp (6.8% gain in novel classes). We therefore compare our method with $\mathrm{CoOp}^{\dagger}$ , which trains CoOp in Co-CoOp setting (refer to Appendix A for more details). + +Table 1. Comparison of MaPLe with different prompting designs in base-to-novel generalization. Results are averaged over 11 datasets. HM refers to harmonic mean. + +
BaseNovelHM
CoOp82.6963.2271.66
Co-CoOp80.4771.6975.83
CoOp†80.8570.0275.04
MaPLe82.2875.1478.55
+ +Table 2. Generalization comparison of MaPLe with CoOp†. + +Compare to $\mathrm{CoOp}^{\dagger}$ , the vanilla CoOp model seems to overfit on base classes. When compared to $\mathrm{CoOp}^{\dagger}$ which attains an average base accuracy of $80.85\%$ , MaPLe shows an improvement of $1.43\%$ with the average base accuracy of $82.28\%$ (Table 2). + +# 4.4. Cross-Dataset Evaluation + +We test the cross-dataset generalization ability of MaPLe by learning multi-modal prompts on all the 1000 ImageNet classes and then transferring it directly on the remaining 10 datasets. Table 4 shows the performance comparison between MaPLe, CoOp and Co-CoOp. On the ImageNet source dataset, MaPLe achieves performance comparable + +(a) Average over 11 datasets + +
BaseNovelHM
CLIP69.3474.2271.70
CoOp82.6963.2271.66
Co-CoOp80.4771.6975.83
MaPLe82.2875.1478.55
+1.81+3.45+2.72
+ +(b) ImageNet. + +
BaseNovelHM
CLIP72.4368.1470.22
CoOp76.4767.8871.92
Co-CoOp75.9870.4373.10
MaPLe76.6670.5473.47
+0.68+0.11+0.37
+ +(c) Caltech101 + +
BaseNovelHM
CLIP96.8494.0095.40
CoOp98.0089.8193.73
Co-CoOp97.9693.8195.84
MaPLe97.7494.3696.02
-0.22+0.55+0.18
+ +(d) OxfordPets + +
BaseNovelHM
CLIP91.1797.2694.12
CoOp93.6795.2994.47
Co-CoOp95.2097.6996.43
MaPLe95.4397.7696.58
+0.23+0.07+0.15
+ +(e) StanfordCars + +
BaseNovelHM
CLIP63.3774.8968.65
CoOp78.1260.4068.13
Co-CoOp70.4973.5972.01
MaPLe72.9474.0073.47
+2.45+0.41+1.46
+ +(f) Flowers102 + +
BaseNovelHM
CLIP72.0877.8074.83
CoOp97.6059.6774.06
Co-CoOp94.8771.7581.71
MaPLe95.9272.4682.56
+1.05+0.71+0.85
+ +(g) Food101 + +
BaseNovelHM
CLIP90.1091.2290.66
CoOp88.3382.2685.19
Co-CoOp90.7091.2990.99
MaPLe90.7192.0591.38
+0.01+0.76+0.39
+ +(h) FGVCAircraft + +
BaseNovelHM
CLIP27.1936.2931.09
CoOp40.4422.3028.75
Co-CoOp33.4123.7127.74
MaPLe37.4435.6136.50
+4.03+11.90+8.76
+ +(i) SUN397 + +
BaseNovelHM
CLIP69.3675.3572.23
CoOp80.6065.8972.51
Co-CoOp79.7476.8678.27
MaPLe80.8278.7079.75
+1.08+1.84+1.48
+ +(j) DTD + +
BaseNovelHM
CLIP53.2459.9056.37
CoOp79.4441.1854.24
Co-CoOp77.0156.0064.85
MaPLe80.3659.1868.16
+3.35+3.18+3.31
+ +(k) EuroSAT + +
BaseNovelHM
CLIP56.4864.0560.03
CoOp92.1954.7468.69
Co-CoOp87.4960.0471.21
MaPLe94.0773.2382.35
+6.58+13.19+11.14
+ +(1)UCF101 + +
BaseNovelHM
CLIP70.5377.5073.85
CoOp84.6956.0567.46
Co-CoOp82.3373.4577.64
MaPLe83.0078.6680.77
+0.67+5.21+3.13
+ +Table 3. Comparison with state-of-the-art methods on base-to-novel generalization. MaPLe learns multi-modal prompts and demonstrates strong generalization results over existing methods on 11 recognition datasets. Absolute gains over Co-CoOp are indicated in blue. + +
SourceTarget
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101AircraftSUN397DTDEuroSATUCF101Average
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
Co-CoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
MaPLe70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
+ +Table 4. Comparison of MaPLe with existing approaches on cross-dataset evaluation. Overall, MaPLe achieves competitive performance providing highest average accuracy, indicating better generalization. + +to competing approaches but demonstrates a much stronger generalization performance by surpassing CoOp in 9/10 and Co-CoOp in 8/10 datasets. Overall, MaPLe shows competitive performance leading to the highest averaged accuracy of $66.30\%$ . This suggests that the use of branch-aware V-L prompting in MaPLe facilitates better generalization. + +# 4.5. Domain Generalization + +We show that MaPLe generalizes favourably on out-of-distribution datasets as compared to CoOp and Co-CoOp. We evaluate the direct transferability of ImageNet trained model to various out-of-domain datasets, and observe that it consistently improves against all the existing approaches + +
SourceTarget
ImageNetImageNetV2ImageNet-SImageNet-AImageNet-R
CLIP66.7360.8346.1547.7773.96
CoOp71.5164.2047.9949.7175.21
Co-CoOp71.0264.0748.7550.6376.18
MaPLe70.7264.0749.1550.9076.98
+ +Table 5. Comparison of MaPLe with existing approaches in domain generalization setting. MaPLe shows consistent improvements on all target datasets. + +as indicated in Table 5. This indicates that utilizing multimodal branch-aware prompting helps MaPLe in enhancing the generalization and robustness of V-L models like CLIP. + +# 4.6. Ablation Experiments + +Prompt Depth: In Fig. 4 (left), we illustrate the effect of prompt depth $J$ for MaPLe and ablate on the depth of language and vision branch individually. In general, the performance improves as prompt depth increases. We note that performance sensitivity increases when randomly initialized prompts are inserted in the deeper layers of a frozen model where the model feature space is already mature. Similar trend is also reported by [16]. As earlier methods utilize shallow language prompting ( $J = 1$ ), we compare our method with deep language prompting. Overall, MaPLe achieves better performance than deep language prompting and achieves maximum performance at a depth of 9. + +Prompt Length: Fig. 4 (right) shows the effect of prompt length for MaPLe. As the prompt length increases, the performance on base classes is generally maintained, while the novel class accuracy decreases. This indicates over-fitting which inherently hurts the generalization to novel classes. + +Effectiveness of Multi-modal Prompting: Fig. 5 shows the analysis of per class accuracy for selected datasets in the order of increasing domain shift. It indicates that the performance gains of MaPLe in comparison to Co-CoOp varies across different datasets. MaPLe provides significant gains over Co-CoOp for datasets that have large distribution shifts from the pretraining dataset of CLIP, and vision concepts that are usually rare and less generic. Further detailed analysis is provided in Appendix C. + +Prompting complexity: Table 6 shows the computational complexity of MaPLe in comparison with other approaches. + +![](images/8288210b3d7acda788e8c59c8fd82ca205aec09731370e88eb536d6e1d14a208.jpg) +Figure 4. Ablation on prompt depth (left) and prompt length (right) in MaPLe. We report average results on the held-out validation sets of all datasets. + +![](images/d3568a4077d46c53fe89df4bc2fa3cdf22236decbe0d489c71d5f5e46f955731.jpg) + +![](images/ac8ffce85353eb3f418849e2d8a5ffb6e4f88a1db2ce83d86ea563fd8c57a052.jpg) +Figure 5. Percentage classes where MaPLe shows improved performance over Co-CoOp, which increases as dataset domain shift from generic categories increases $(\rightarrow)$ . + +Although MaPLe utilizes multi-modal prompts, its overall FLOPS (Floating Point Operations) exceeds only by $0.1\%$ over CoOp and Co-CoOp. The independent V-L prompting also provides comparable FLOP count. In terms of inference speed, Co-CoOp is significantly slower and the FPS (Frames Per Second) remains constant as the batch size increases. In contrast, MaPLe has no such overhead and provides much better inference and training speeds. Further, MaPLe provides better convergence as it requires only half training epochs as compared to Co-CoOp (5 vs 10 epochs). MaPLe adds about $2.85\%$ training parameters on top of CLIP. To study if the performance gain is mainly attributed to more parameters, we experiment with $\mathrm{MaPLe\dagger}$ , which uses a unified V-L coupling function for all layer prompts. $\mathrm{MaPLe\dagger}$ with about 9x lesser parameters than MaPLe also improves over existing methods. We also ablate by comparing MaPLe with heavier CoCoOp in Appendix D. + +
MethodParamsParams % CLIPFPS (with BS)HM
14100
CoOp20480.00213.855.31353.071.66
CoCoOp353600.0364.6114.715.175.83
Independent V-L314880.0262.5239.41383.877.90
MaPLe3.55 M2.8560.2239.01365.178.55
MaPLe†0.41 M0.3360.2238.01365.078.11
+ +Table 6. Comparison of computational complexity among different prompting methods. MaPLe† is a MaPLe version which utilizes a common V-L coupling function for all layers. + +# 5. Conclusion + +Adaptation of large-scale V-L models, e.g., CLIP [32] to downstream tasks is a challenging problem due to the large number of tunable parameters and limited size of downstream datasets. Prompt learning is an efficient and scalable technique to tailor V-L models to novel downstream tasks. To this end, the current prompt learning approaches either consider only the vision or language side prompting. Our work shows that it is critical to perform prompting for both vision and language branches to appropriately adapt V-L models to downstream tasks. Further, we propose a strategy to ensure synergy between vision-language modalities by explicitly conditioning the vision prompts on textual prompt across different transformer stages. Our approach improves the generalization towards novel categories, cross-dataset transfer and datasets with domain shifts. + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 2022. 3, 11, 12 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In The European Conference on Computer Vision, pages 446-461. Springer, 2014. 5 +[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 5 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 5 +[5] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3 +[7] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5 +[8] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In The European Conference on Computer Vision, 2022. 2 +[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2, 3 +[10] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2 +[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5 +[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. + +In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 5 +[13] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 5 +[14] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 1 +[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2 +[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In The European Conference on Computer Vision, 2022. 2, 3, 8 +[17] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt is worth millions of parameters? low-resource prompt-based learning for vision-language models. arXiv preprint arXiv:2110.08484, 2021. 1 +[18] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In The European Conference on Computer Vision, 2021. 3 +[19] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model, 2022. 2 +[20] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 554–561, 2013. 5 +[21] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In _Conference on Empirical Methods in Natural Language Processing_, 2021. 2 +[22] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2 +[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 2021. 2 +[24] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 2021. 2 +[25] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 3 +[26] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. 2 +[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In The European Conference on Computer Vision. Springer, 2022. 2 +[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[29] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. 2022. 1 +[30] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5 +[31] Omkar M Parkhi, Andreae Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 8 +[33] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2 +[34] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In Advances in Neural Information Processing Systems, 2022. 2 +[35] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 5 +[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[37] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, volume 32, 2019. 5 +[38] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary + +prompting for rehearsal-free continual learning. In The European Conference on Computer Vision, 2022. 2 +[39] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2 +[40] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5 +[41] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2 +[42] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 1 +[43] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2 +[44] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In The European Conference on Computer Vision, 2022. 2 +[45] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 2 +[46] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In The European Conference on Computer Vision, 2022. 2 +[47] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 2 +[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 2, 3, 5, 6 +[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, pages 1-12, 2022. 1, 2, 3, 5, 6 +[50] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In The European Conference on Computer Vision, 2022. 2 +[51] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. arXiv preprint arXiv:2205.14865, 2022. 2 \ No newline at end of file diff --git a/2023/MaPLe_ Multi-Modal Prompt Learning/images.zip b/2023/MaPLe_ Multi-Modal Prompt Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..93a2fe36c5ffe32b1a554be49f0b43e923090061 --- /dev/null +++ b/2023/MaPLe_ Multi-Modal Prompt Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00234e15c408d4ae025cfb1b68044e92884a73f99da5c34127ecaf337ca091fc +size 654347 diff --git a/2023/MaPLe_ Multi-Modal Prompt Learning/layout.json b/2023/MaPLe_ Multi-Modal Prompt Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..33889f5e9223292b2479fdf2733e08afd4de4849 --- /dev/null +++ b/2023/MaPLe_ Multi-Modal Prompt Learning/layout.json @@ -0,0 +1,9338 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 169, + 136, + 410, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 136, + 410, + 154 + ], + "spans": [ + { + "bbox": [ + 169, + 136, + 410, + 154 + ], + "type": "text", + "content": "MaPLe: Multi-modal Prompt Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 174, + 458, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 174, + 458, + 187 + ], + "spans": [ + { + "bbox": [ + 118, + 174, + 458, + 187 + ], + "type": "text", + "content": "Muhammad Uzair Khattak1 Hanoona Rasheed1 Muhammad Maaz1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 188, + 387, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 188, + 387, + 202 + ], + "spans": [ + { + "bbox": [ + 187, + 188, + 387, + 202 + ], + "type": "text", + "content": "Salman Khan" + }, + { + "bbox": [ + 187, + 188, + 387, + 202 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 187, + 188, + 387, + 202 + ], + "type": "text", + "content": " Fahad Shahbaz Khan" + }, + { + "bbox": [ + 187, + 188, + 387, + 202 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 203, + 520, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 203, + 520, + 219 + ], + "spans": [ + { + "bbox": [ + 54, + 203, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 203, + 520, + 219 + ], + "type": "text", + "content": "Mohamed bin Zayed University of AI " + }, + { + "bbox": [ + 54, + 203, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 203, + 520, + 219 + ], + "type": "text", + "content": "Australian National University " + }, + { + "bbox": [ + 54, + 203, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 203, + 520, + 219 + ], + "type": "text", + "content": "Linköping University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 245, + 187, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 245, + 187, + 258 + ], + "spans": [ + { + "bbox": [ + 140, + 245, + 187, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 269, + 281, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 269, + 281, + 607 + ], + "spans": [ + { + "bbox": [ + 45, + 269, + 281, + 607 + ], + "type": "text", + "content": "Pre-trained vision-language (V-L) models such as CLIP have shown excellent generalization ability to downstream tasks. However, they are sensitive to the choice of input text prompts and require careful selection of prompt templates to perform well. Inspired by the Natural Language Processing (NLP) literature, recent CLIP adaptation approaches learn prompts as the textual inputs to fine-tune CLIP for downstream tasks. We note that using prompting to adapt representations in a single branch of CLIP (language or vision) is sub-optimal since it does not allow the flexibility to dynamically adjust both representation spaces on a downstream task. In this work, we propose Multi-modal Prompt Learning (MaPLe) for both vision and language branches to improve alignment between the vision and language representations. Our design promotes strong coupling between the vision-language prompts to ensure mutual synergy and discourages learning independent uni-modal solutions. Further, we learn separate prompts across different early stages to progressively model the stage-wise feature relationships to allow rich context learning. We evaluate the effectiveness of our approach on three representative tasks of generalization to novel classes, new target datasets and unseen domain shifts. Compared with the state-of-the-art method CoCoOp, MaPLe exhibits favorable performance and achieves an absolute gain of " + }, + { + "bbox": [ + 45, + 269, + 281, + 607 + ], + "type": "inline_equation", + "content": "3.45\\%" + }, + { + "bbox": [ + 45, + 269, + 281, + 607 + ], + "type": "text", + "content": " on novel classes and " + }, + { + "bbox": [ + 45, + 269, + 281, + 607 + ], + "type": "inline_equation", + "content": "2.72\\%" + }, + { + "bbox": [ + 45, + 269, + 281, + 607 + ], + "type": "text", + "content": " on overall harmonic-mean, averaged over 11 diverse image recognition datasets. Our code and pre-trained models are available at https://github.com/muzairkhattak/multimodal-prompt-learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 628, + 125, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 125, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 125, + 640 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": "Foundational vision-language (V-L) models such as CLIP (Contrastive Language-Image Pretraining) [32] have shown excellent generalization ability to downstream tasks. Such models are trained to align language and vision modalities on web-scale data e.g., 400 million text-image pairs in CLIP. These models can reason about open-vocabulary visual concepts, thanks to the rich supervision provided by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 245, + 532, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 245, + 532, + 328 + ], + "spans": [ + { + "bbox": [ + 296, + 245, + 532, + 328 + ], + "type": "text", + "content": "natural language. During inference, hand-engineered text prompts are used e.g., 'a photo of a ' as a query for text encoder. The output text embeddings are matched with the visual embeddings from an image encoder to predict the output class. Designing high quality contextual prompts have been proven to enhance the performance of CLIP and other V-L models [17, 42]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 330, + 532, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 330, + 532, + 505 + ], + "spans": [ + { + "bbox": [ + 296, + 330, + 532, + 505 + ], + "type": "text", + "content": "Despite the effectiveness of CLIP towards generalization to new concepts, its massive scale and scarcity of training data (e.g., few-shot setting) makes it infeasible to fine-tune the full model for downstream tasks. Such fine-tuning can also forget the useful knowledge acquired in the large-scale pretraining phase and can pose a risk of overfitting to the downstream task. To address the above challenges, existing works propose language prompt learning to avoid manually adjusting the prompt templates and providing a mechanism to adapt the model while keeping the original weights frozen [14, 25, 29, 48, 49]. Inspired from Natural Language Processing (NLP), these approaches only explore prompt learning for the text encoder in CLIP (Fig. 1:a) while adaptation choices together with an equally important image encoder of CLIP remains an unexplored topic in the literature." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 508, + 533, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 508, + 533, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 508, + 533, + 730 + ], + "type": "text", + "content": "Our motivation derives from the multi-modal nature of CLIP, where a text and image encoder co-exist and both contribute towards properly aligning the V-L modalities. We argue that any prompting technique should adapt the model completely and therefore, learning prompts only for the text encoder in CLIP is not sufficient to model the adaptations needed for the image encoder. To this end, we set out to achieve completeness in the prompting approach and propose Multi-modal Prompt Learning (MaPLe) to adequately fine-tune the text and image encoder representations such that their optimal alignment can be achieved on the downstream tasks (Fig. 1:b). Our extensive experiments on three key representative settings including base-to-novel generalization, cross-dataset evaluation, and domain generalization demonstrate the strength of MaPLe. On base-to-novel generalization, our proposed MaPLe outperforms existing prompt learning approaches across 11 diverse image recognition datasets (Fig. 1:c) and achieves absolute average gain of " + }, + { + "bbox": [ + 296, + 508, + 533, + 730 + ], + "type": "inline_equation", + "content": "3.45\\%" + }, + { + "bbox": [ + 296, + 508, + 533, + 730 + ], + "type": "text", + "content": " on novel classes and " + }, + { + "bbox": [ + 296, + 508, + 533, + 730 + ], + "type": "inline_equation", + "content": "2.72\\%" + }, + { + "bbox": [ + 296, + 508, + 533, + 730 + ], + "type": "text", + "content": " on harmonic-mean" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 7, + 102, + 39 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 7, + 102, + 39 + ], + "spans": [ + { + "bbox": [ + 57, + 7, + 102, + 39 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 5, + 488, + 41 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 5, + 488, + 41 + ], + "spans": [ + { + "bbox": [ + 137, + 5, + 488, + 41 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "type": "text", + "content": "19113" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 103, + 263, + 188 + ], + "blocks": [ + { + "bbox": [ + 73, + 103, + 263, + 188 + ], + "lines": [ + { + "bbox": [ + 73, + 103, + 263, + 188 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 263, + 188 + ], + "type": "image", + "image_path": "9eb4884d9989b772b613fc7472cdd8830703485c6d51fda6ac3498f906878816.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 188, + 258, + 198 + ], + "lines": [ + { + "bbox": [ + 75, + 188, + 258, + 198 + ], + "spans": [ + { + "bbox": [ + 75, + 188, + 258, + 198 + ], + "type": "text", + "content": "(a) Existing prompt tuning methods (Uni-modal)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 205, + 264, + 287 + ], + "blocks": [ + { + "bbox": [ + 73, + 205, + 264, + 287 + ], + "lines": [ + { + "bbox": [ + 73, + 205, + 264, + 287 + ], + "spans": [ + { + "bbox": [ + 73, + 205, + 264, + 287 + ], + "type": "image", + "image_path": "4d537288361b8997d134adeff410f0017bea0bb9fef2994c446047f6a5f32525.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 287, + 246, + 297 + ], + "lines": [ + { + "bbox": [ + 87, + 287, + 246, + 297 + ], + "spans": [ + { + "bbox": [ + 87, + 287, + 246, + 297 + ], + "type": "text", + "content": "(b) Multi-modal Prompt Learning (MaPLe)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 44, + 301, + 532, + 345 + ], + "lines": [ + { + "bbox": [ + 44, + 301, + 532, + 345 + ], + "spans": [ + { + "bbox": [ + 44, + 301, + 532, + 345 + ], + "type": "text", + "content": "Figure 1. Comparison of MaPLe with standard prompt learning methods. (a) Existing methods adopt uni-modal prompting techniques to fine-tune CLIP representations as prompts are learned only in a single branch of CLIP (language or vision). (b) MaPLe introduces branch-aware hierarchical prompts that adapt both language and vision branches simultaneously for improved generalization. (c) MaPLe surpasses state-of-the-art methods on 11 diverse image recognition datasets for novel class generalization task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 274, + 103, + 506, + 283 + ], + "blocks": [ + { + "bbox": [ + 274, + 103, + 506, + 283 + ], + "lines": [ + { + "bbox": [ + 274, + 103, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 274, + 103, + 506, + 283 + ], + "type": "image", + "image_path": "ee6a47f15551f34258f8ea90ebea1f4099db62ca677e5062c8af79aa1b9dd4ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 287, + 500, + 298 + ], + "lines": [ + { + "bbox": [ + 272, + 287, + 500, + 298 + ], + "spans": [ + { + "bbox": [ + 272, + 287, + 500, + 298 + ], + "type": "text", + "content": "(c) Performance comparison on base-to-novel generalization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 364, + 280, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 364, + 280, + 479 + ], + "spans": [ + { + "bbox": [ + 45, + 364, + 280, + 479 + ], + "type": "text", + "content": "over the state-of-the-art method Co-CoOp [48]. Further, MaPLe demonstrates favorable generalization ability and robustness in cross-dataset transfer and domain generalization settings, leading to consistent improvements compared to existing approaches. Owing to its streamlined architectural design, MaPLe exhibits improved efficiency during both training and inference without much overhead, as compared to Co-CoOp which lacks efficiency due to its image instance conditioned design. In summary, the main contributions of this work include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 481, + 280, + 674 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 57, + 481, + 280, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 481, + 280, + 527 + ], + "spans": [ + { + "bbox": [ + 57, + 481, + 280, + 527 + ], + "type": "text", + "content": "- We propose multi-modal prompt learning in CLIP to favourably align its vision-language representations. To the best of our knowledge, this is the first multi-modal prompting approach for fine-tuning CLIP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 531, + 280, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 531, + 280, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 280, + 589 + ], + "type": "text", + "content": "- To link prompts learned in text and image encoders, we propose a coupling function to explicitly condition vision prompts on their language counterparts. It acts as a bridge between the two modalities and allows mutual propagation of gradients to promote synergy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 591, + 280, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 280, + 674 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 280, + 674 + ], + "type": "text", + "content": "- Our multi-modal prompts are learned across multiple transformer blocks in both vision and language branches to progressively learn the synergistic behaviour of both modalities. This deep prompting strategy allows modeling the contextual relationships independently, thus providing more flexibility to align the vision-language representations." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 687, + 130, + 699 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 687, + 130, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 130, + 699 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "Vision Language Models: The combined use of language supervision with natural images is found to be of great in" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 364, + 531, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 364, + 531, + 609 + ], + "spans": [ + { + "bbox": [ + 296, + 364, + 531, + 609 + ], + "type": "text", + "content": "terest in the computer vision community. In contrast to models learned with only image supervision, these vision-language (V-L) models encode rich multimodal representations. Recently, V-L models like CLIP [32], ALIGN [15], LiT [45], FILIP [41] and Florence [43] have demonstrated exceptional performance on a wide spectrum of tasks including few-shot and zero-shot visual recognition. These models learn joint image-language representations in a self-supervised manner using abundantly available data from the web. For example, CLIP and ALIGN respectively use " + }, + { + "bbox": [ + 296, + 364, + 531, + 609 + ], + "type": "inline_equation", + "content": "\\sim 400\\mathrm{M}" + }, + { + "bbox": [ + 296, + 364, + 531, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 364, + 531, + 609 + ], + "type": "inline_equation", + "content": "\\sim 1\\mathrm{B}" + }, + { + "bbox": [ + 296, + 364, + 531, + 609 + ], + "type": "text", + "content": " image-text pairs to train a multimodal network. Although these pre-trained V-L models learn generalized representations, efficiently adapting them to downstream tasks is still a challenging problem. Many works have demonstrated better performance on downstream tasks by using tailored methods to adapt V-L models for few-shot image-recognition [9, 19, 46], object detection [8, 10, 27, 34, 44, 50], and segmentation [5, 22, 26, 33]. In this work, we propose a novel multi-modal prompt learning technique to effectively adapt CLIP for few-shot and zero-shot visual recognition tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 613, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 613, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 613, + 532, + 730 + ], + "type": "text", + "content": "Prompt Learning: The instructions in the form of a sentence, known as text prompt, are usually given to the language branch of a V-L model, allowing it to better understand the task. Prompts can be handcrafted for a downstream task or learned automatically during fine-tuning stage. The latter is referred to as 'Prompt Learning' which was first used in NLP [21,23,24] followed by the adaptation in V-L [48, 49, 51] and vision-only [16, 38, 39, 47] models. Similar to [16] our design also uses deep 'vision' prompting. However, ours is the first multi-modal prompting de" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "text", + "content": "19114" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 104, + 481, + 283 + ], + "blocks": [ + { + "bbox": [ + 99, + 104, + 481, + 283 + ], + "lines": [ + { + "bbox": [ + 99, + 104, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 99, + 104, + 481, + 283 + ], + "type": "image", + "image_path": "93f706335d0a31f607a7e9c15f96265303f5203fafdf9d9584b52a0912f4d9db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 287, + 533, + 332 + ], + "lines": [ + { + "bbox": [ + 45, + 287, + 533, + 332 + ], + "spans": [ + { + "bbox": [ + 45, + 287, + 533, + 332 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed MaPLe (Multi-modal Prompt Learning) framework for prompt learning in V-L models. MaPLe tunes both vision and language branches where only the context prompts are learned, while the rest of the model is frozen. MaPLe conditions the vision prompts on language prompts via a V-L coupling function " + }, + { + "bbox": [ + 45, + 287, + 533, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 45, + 287, + 533, + 332 + ], + "type": "text", + "content": " to induce mutual synergy between the two modalities. Our framework uses deep contextual prompting where separate context prompts are learned across multiple transformer blocks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 351, + 162, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 351, + 162, + 362 + ], + "spans": [ + { + "bbox": [ + 45, + 351, + 162, + 362 + ], + "type": "text", + "content": "sign while [16] is uni-modal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 362, + 281, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 362, + 281, + 678 + ], + "spans": [ + { + "bbox": [ + 45, + 362, + 281, + 678 + ], + "type": "text", + "content": "Prompt Learning in Vision Language models: Full fine-tuning and linear probing [9] are two typical approaches to adapt a V-L model (i.e. CLIP) to the downstream tasks. The complete fine-tuning results in degrading the previously learned joint V-L representation while linear probing limits the zero-shot capability of CLIP. To this end, inspired from prompt learning in NLP, many works have proposed to adapt V-L models by learning the prompt tokens in an end-to-end training. CoOp [49] fine-tunes CLIP for few-shot transfer by optimizing continuous set of prompt vectors at its language branch. Co-CoOp [48] highlights the inferior performance of CoOp on novel classes and solves the generalization issue by explicitly conditioning prompts on image instances. [25] proposes to optimize multiple set of prompts by learning the distribution of prompts. [18] adapt CLIP by learning prompts for video understanding tasks. [1] perform visual prompt tuning on CLIP by prompting on the vision branch. We note that the existing methods follow independent uni-modal solutions and learn prompts either in the language or in the vision branch of CLIP, thus adapting CLIP partially. In this paper, we explore an important question: given the multimodal nature of CLIP, is complete prompting (i.e., in both language and vision branches) better suited to adapt CLIP? Our work is the first to answer this question by investigating the effectiveness of multi-modal prompt learning in order to improve alignment between vision and language representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 687, + 101, + 699 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 687, + 101, + 699 + ], + "spans": [ + { + "bbox": [ + 45, + 687, + 101, + 699 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 706, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 280, + 730 + ], + "type": "text", + "content": "Our approach concerns with fine-tuning a pre-trained multimodal CLIP for better generalization to downstream tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 351, + 533, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 351, + 533, + 549 + ], + "spans": [ + { + "bbox": [ + 296, + 351, + 533, + 549 + ], + "type": "text", + "content": "through context optimization via prompting. Fig. 2 shows the overall architecture of our proposed MaPLe (Multimodal Prompt Learning) framework. Unlike previous approaches [48, 49] which learn context prompts only at the language branch, MaPLe proposes a joint prompting approach where the context prompts are learned in both vision and language branches. Specifically, we append learnable context tokens in the language branch and explicitly condition the vision prompts on the language prompts via a coupling function to establish interaction between them. To learn hierarchical contextual representations, we introduce deep prompting in both branches through separate learnable context prompts across different transformer blocks. During fine-tuning, only the context prompts along with their coupling function are learned while the rest of the model is frozen. Below, we first outline the pre-trained CLIP architecture and then present our proposed fine-tuning approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 298, + 572, + 395, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 572, + 395, + 586 + ], + "spans": [ + { + "bbox": [ + 298, + 572, + 395, + 586 + ], + "type": "text", + "content": "3.1. Revisiting CLIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 596, + 532, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 596, + 532, + 667 + ], + "spans": [ + { + "bbox": [ + 296, + 596, + 532, + 667 + ], + "type": "text", + "content": "We build our approach on a pre-trained vision-language (V-L) model, CLIP, which consists of a text and vision encoder. Consistent with existing prompting methods [48, 49], we use a vision transformer (ViT) [6] based CLIP model. CLIP encodes an image " + }, + { + "bbox": [ + 296, + 596, + 532, + 667 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 296, + 596, + 532, + 667 + ], + "type": "text", + "content": " and a corresponding text description as explained below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "spans": [ + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": "Encoding Image: Image encoder " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": " transformer layers " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": ", splits the image " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": " fixed-size patches which are projected into patch embeddings " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "E_0 \\in \\mathbb{R}^{M \\times d_v}" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": ". Patch embeddings " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": " are input to the " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "(i+1)^{\\mathrm{th}}" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": " transformer block " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_{i+1})" + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "text", + "content": " along with a learnable class (CLS) token " + }, + { + "bbox": [ + 296, + 671, + 532, + 731 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "text", + "content": "19115" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 103, + 527, + 181 + ], + "blocks": [ + { + "bbox": [ + 53, + 103, + 527, + 181 + ], + "lines": [ + { + "bbox": [ + 53, + 103, + 527, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 527, + 181 + ], + "type": "image", + "image_path": "e16ab2a8f4b922761701d24185fa40f44ed8b6343440e0896944f5853bd689a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 182, + 527, + 265 + ], + "blocks": [ + { + "bbox": [ + 53, + 182, + 527, + 265 + ], + "lines": [ + { + "bbox": [ + 53, + 182, + 527, + 265 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 527, + 265 + ], + "type": "image", + "image_path": "61c0a1218a8eac098d5c354f62e6fcf5de7744de687314db5a8701ba0d7aeaca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 267, + 532, + 290 + ], + "lines": [ + { + "bbox": [ + 45, + 267, + 532, + 290 + ], + "spans": [ + { + "bbox": [ + 45, + 267, + 532, + 290 + ], + "type": "text", + "content": "Figure 3. t-SNE plots of image embeddings in uni-modal prompting method Co-CoOp, and MaPLe on 3 diverse image recognition datasets. MaPLe shows better separability in both base and novel classes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 309, + 275, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 309, + 275, + 322 + ], + "spans": [ + { + "bbox": [ + 45, + 309, + 275, + 322 + ], + "type": "text", + "content": "and sequentially processed through " + }, + { + "bbox": [ + 45, + 309, + 275, + 322 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 309, + 275, + 322 + ], + "type": "text", + "content": " transformer blocks," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 328, + 254, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 328, + 254, + 341 + ], + "spans": [ + { + "bbox": [ + 71, + 328, + 254, + 341 + ], + "type": "interline_equation", + "content": "[ \\mathbf {c} _ {i}, E _ {i} ] = \\mathcal {V} _ {i} ([ \\mathbf {c} _ {i - 1}, E _ {i - 1} ]) \\quad i = 1, 2, \\dots , K.", + "image_path": "033c0ff2e77ae6970997611cf32c6c06c65d76e19edab537d57162da3a15bcb9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "spans": [ + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "type": "text", + "content": "To obtain the final image representation " + }, + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "type": "text", + "content": ", the class token " + }, + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "type": "inline_equation", + "content": "\\mathsf{c}_K" + }, + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "type": "text", + "content": " of last transformer layer " + }, + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_K)" + }, + { + "bbox": [ + 45, + 348, + 280, + 383 + ], + "type": "text", + "content": " is projected to a common V-L latent embedding space via ImageProj," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 389, + 236, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 389, + 236, + 404 + ], + "spans": [ + { + "bbox": [ + 88, + 389, + 236, + 404 + ], + "type": "interline_equation", + "content": "x = \\operatorname {I m a g e P r o j} \\left(\\mathrm {c} _ {K}\\right) \\quad x \\in \\mathbb {R} ^ {d _ {v l}}.", + "image_path": "295bd2727f075aa087efff3166b46e3288f9a0bcfe1eec2bdd6c499ac9caa8fe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "spans": [ + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "text", + "content": "Encoding Text: CLIP text encoder generates feature representations for text description by tokenizing the words and projecting them to word embeddings " + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "inline_equation", + "content": "W_{0} = [w_{0}^{1}, w_{0}^{2}, \\dots, w_{0}^{N}] \\in \\mathbb{R}^{N \\times d_{l}}" + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "text", + "content": ". At each stage, " + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "text", + "content": " is input to the " + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "inline_equation", + "content": "(i + 1)^{\\mathrm{th}}" + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "text", + "content": " transformer layer of text encoding branch " + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{i + 1})" + }, + { + "bbox": [ + 45, + 410, + 280, + 480 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 486, + 241, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 486, + 241, + 499 + ], + "spans": [ + { + "bbox": [ + 85, + 486, + 241, + 499 + ], + "type": "interline_equation", + "content": "\\left[ W _ {i} \\right] = \\mathcal {L} _ {i} \\left(W _ {i - 1}\\right) \\quad i = 1, 2, \\dots , K.", + "image_path": "217b178788493492c3d1c0c81ae1c9e2fb1d1e37d5eaef6fe355595f79f32c8a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 506, + 280, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 506, + 280, + 554 + ], + "spans": [ + { + "bbox": [ + 45, + 506, + 280, + 554 + ], + "type": "text", + "content": "The final text representation " + }, + { + "bbox": [ + 45, + 506, + 280, + 554 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 45, + 506, + 280, + 554 + ], + "type": "text", + "content": " is obtained by projecting the text embeddings corresponding to the last token of the last transformer block " + }, + { + "bbox": [ + 45, + 506, + 280, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_K" + }, + { + "bbox": [ + 45, + 506, + 280, + 554 + ], + "type": "text", + "content": " to a common V-L latent embedding space via TextProj," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 559, + 234, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 559, + 234, + 573 + ], + "spans": [ + { + "bbox": [ + 91, + 559, + 234, + 573 + ], + "type": "interline_equation", + "content": "z = \\operatorname {T e x t P r o j} \\left(w _ {K} ^ {N}\\right) \\quad z \\in \\mathbb {R} ^ {d _ {v l}}.", + "image_path": "2f9b209b1413de39e6abced25a361d871b0fda374872351762ada65c8ce0dc46.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "spans": [ + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "text", + "content": "Zero-shot Classification: For zero-shot classification, text prompts are hand-crafted with class labels " + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "inline_equation", + "content": "y \\in \\{1, 2, \\dots, C\\}" + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "text", + "content": " (e.g., 'a photo of a ) having " + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "text", + "content": " classes. Prediction " + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "text", + "content": " corresponding to the image " + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "text", + "content": " having the highest cosine similarity score " + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "inline_equation", + "content": "(sim(\\cdot))" + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "text", + "content": " is calculated with a temperature parameter " + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 45, + 579, + 280, + 650 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 94, + 655, + 230, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 655, + 230, + 684 + ], + "spans": [ + { + "bbox": [ + 94, + 655, + 230, + 684 + ], + "type": "interline_equation", + "content": "p (\\hat {y} | x) = \\frac {\\exp (s i m (x , z _ {\\hat {y}}) / \\tau)}{\\sum_ {i = 1} ^ {C} \\exp (s i m (x , z _ {i}))}.", + "image_path": "6348c1a46c8c36a53c5661754e47b41a8ece7c564d3ac9d199339b534265be39.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 688, + 246, + 701 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 688, + 246, + 701 + ], + "spans": [ + { + "bbox": [ + 45, + 688, + 246, + 701 + ], + "type": "text", + "content": "3.2. MaPLe: Multi-modal Prompt Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 706, + 280, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 280, + 731 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 280, + 731 + ], + "type": "text", + "content": "To efficiently fine-tune CLIP for downstream image recognition tasks, we explore the potential of multi-modal prompt" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 309, + 531, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 309, + 531, + 461 + ], + "spans": [ + { + "bbox": [ + 296, + 309, + 531, + 461 + ], + "type": "text", + "content": "tuning. We reason that prior works that have predominantly explored uni-modal approaches are less suitable as they do not offer the flexibility to dynamically adapt both language and vision representation spaces. Thus to achieve completeness in prompting, we underline the importance of multimodal prompting approach. In Fig. 3, we visualize and compare the image embeddings of MaPLe with recent state-of-the-art work, Co-CoOp. Note that the image embeddings of CLIP, CoOp and Co-CoOp will be identical as they do not learn prompts in the vision branch. The visualization shows that image embeddings of MaPLe are more separable indicating that learning vision prompts in addition to language prompts leads to better adaptation of CLIP." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 461, + 531, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 461, + 531, + 555 + ], + "spans": [ + { + "bbox": [ + 296, + 461, + 531, + 555 + ], + "type": "text", + "content": "In addition to multi-modal prompting, we find that it is essential to learn prompts in the deeper transformer layers to progressively model stage-wise feature representations. To this end, we propose to introduce learnable tokens in the first " + }, + { + "bbox": [ + 296, + 461, + 531, + 555 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 296, + 461, + 531, + 555 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 296, + 461, + 531, + 555 + ], + "type": "inline_equation", + "content": "J < K" + }, + { + "bbox": [ + 296, + 461, + 531, + 555 + ], + "type": "text", + "content": ") layers of both vision and language branches. These multi-modal hierarchical prompts utilize the knowledge embedded in CLIP model to effectively learn task relevant contextual representations (see Fig. 4)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 298, + 571, + 442, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 571, + 442, + 584 + ], + "spans": [ + { + "bbox": [ + 298, + 571, + 442, + 584 + ], + "type": "text", + "content": "3.2.1 Deep Language Prompting" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "spans": [ + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "text", + "content": "To learn the language context prompts, we introduce " + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "text", + "content": " learnable tokens " + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "inline_equation", + "content": "\\{P^i \\in \\mathbb{R}^{d_l}\\}_{i=1}^b" + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "text", + "content": ", in the language branch of CLIP. The input embeddings now follow the form " + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "inline_equation", + "content": "[P^1, P^2, \\dots, P^b, W_0]" + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "inline_equation", + "content": "W_0 = [w^1, w^2, \\dots, w^N]" + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "text", + "content": " corresponds to fixed input tokens. New learnable tokens are further introduced in each transformer block of the language encoder " + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_i)" + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "text", + "content": " up to a specific depth " + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 296, + 591, + 531, + 673 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 682, + 531, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 682, + 531, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 682, + 531, + 696 + ], + "type": "interline_equation", + "content": "[ \\_ , W _ {i} ] = \\mathcal {L} _ {i} ([ P _ {i - 1}, W _ {i - 1} ]) \\quad i = 1, 2, \\dots , J. \\tag {1}", + "image_path": "fa2418d787d2185c3344f583af25b77a2b84580ea37376d56cdeec821354552f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "type": "text", + "content": " refers to the concatenation operation. After " + }, + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "type": "inline_equation", + "content": "J^{\\mathrm{th}}" + }, + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "type": "text", + "content": " transformer layer, the subsequent layers process previous" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "text", + "content": "19116" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 272, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 272, + 117 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 272, + 117 + ], + "type": "text", + "content": "layer prompts and final text representation " + }, + { + "bbox": [ + 45, + 106, + 272, + 117 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 45, + 106, + 272, + 117 + ], + "type": "text", + "content": " is computed," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 126, + 278, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 278, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 278, + 154 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} [ P _ {j}, W _ {j} ] = \\mathcal {L} _ {j} ([ P _ {j - 1}, W _ {j - 1} ]) \\quad j = J + 1, \\dots , K, (2) \\\\ z = \\operatorname {T e x t P r o j} \\left(w _ {K} ^ {N}\\right). (3) \\\\ \\end{array}", + "image_path": "be9d9149c12e49bf282647909b6a1b5dec78063f2b7076c9e67900603e57a3ab.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 162, + 279, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 162, + 279, + 198 + ], + "spans": [ + { + "bbox": [ + 45, + 162, + 279, + 198 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 45, + 162, + 279, + 198 + ], + "type": "inline_equation", + "content": "J = 1" + }, + { + "bbox": [ + 45, + 162, + 279, + 198 + ], + "type": "text", + "content": ", the learnable tokens " + }, + { + "bbox": [ + 45, + 162, + 279, + 198 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 45, + 162, + 279, + 198 + ], + "type": "text", + "content": " are only applied at the input of first transformer layer, and this deep language prompting technique degenerates to CoOp [49]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 212, + 175, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 175, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 175, + 224 + ], + "type": "text", + "content": "3.2.2 Deep Vision Prompting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "spans": [ + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "text", + "content": "Similar to deep language prompting, we introduce " + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "text", + "content": " learnable tokens " + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "inline_equation", + "content": "\\{\\hat{P}^i\\in \\mathbb{R}^{d_v}\\}_{i = 1}^b" + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "text", + "content": ", in the vision branch of CLIP alongside the input image tokens. New learnable tokens are further introduced in deeper transformer layers of the image encoder " + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "inline_equation", + "content": "(\\mathcal{V})" + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "text", + "content": " up to depth " + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 45, + 229, + 279, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 279, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 279, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 279, + 340 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} [ c _ {i}, E _ {i}, \\_ ] = \\mathcal {V} _ {i} ([ c _ {i - 1}, E _ {i - 1}, \\bar {P} _ {i - 1} ]) \\quad i = 1, 2, \\dots , J, \\\\ [ c _ {j}, E _ {j}, \\tilde {P} _ {j} ] = \\mathcal {V} _ {j} ([ c _ {j - 1}, E _ {j - 1}, \\tilde {P} _ {j - 1} ]) \\quad j = J + 1, \\dots , K, \\\\ x = \\operatorname {I m a g e P r o j} \\left(c _ {K}\\right). \\\\ \\end{array}", + "image_path": "6041c24ac94663e680b0af0cce116e31e52b97d8eecf94e99b43c8f9ec0dce0e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 348, + 279, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 348, + 279, + 430 + ], + "spans": [ + { + "bbox": [ + 45, + 348, + 279, + 430 + ], + "type": "text", + "content": "Our deep prompting provides the flexibility to learn prompts across different feature hierarchies within the ViT architecture. We find that sharing prompts across stages is better compared to independent prompts as features are more correlated due to successive transformer block processing. Thus, the later stages do not provide independently-learned complimentary prompts as compared to the early stages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 444, + 223, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 444, + 223, + 456 + ], + "spans": [ + { + "bbox": [ + 45, + 444, + 223, + 456 + ], + "type": "text", + "content": "3.2.3 Vision Language Prompt Coupling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 462, + 279, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 462, + 279, + 602 + ], + "spans": [ + { + "bbox": [ + 45, + 462, + 279, + 602 + ], + "type": "text", + "content": "We reason that in prompt tuning it is essential to take a multi-modal approach and simultaneously adapt both the vision and language branch of CLIP in order to achieve completeness in context optimization. A simple approach would be to naively combine deep vision and language prompting, where both the language prompts " + }, + { + "bbox": [ + 45, + 462, + 279, + 602 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 45, + 462, + 279, + 602 + ], + "type": "text", + "content": ", and the vision prompts " + }, + { + "bbox": [ + 45, + 462, + 279, + 602 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 45, + 462, + 279, + 602 + ], + "type": "text", + "content": ", will be learned during the same training schedule. We name this design as 'Independent V-L Prompting'. Although this approach satisfies the requirement of completeness in prompting, this design lacks synergy between vision and language branch as both branches do not interact while learning the task relevant context prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": "To this end, we propose a branch-aware multi-modal prompting which tunes vision and language branch of CLIP together by sharing prompts across both modalities. Language prompt tokens are introduced in the language branch up to " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "J^{\\mathrm{th}}" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " transformer block similar to deep language prompting as illustrated in Eqs. 1-3. To ensure mutual synergy between V-L prompts, vision prompts " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": ", are obtained by projecting language prompts " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " via vision-to-language projection which we refer to as " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "V-L" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " coupling function " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\tilde{P}_k = \\mathcal{F}_k(P_k)" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": ". The coupling function is implemented as a linear layer which maps " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "d_l" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " dimensional inputs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 298, + 106, + 531, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 106, + 531, + 129 + ], + "spans": [ + { + "bbox": [ + 298, + 106, + 531, + 129 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 298, + 106, + 531, + 129 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 298, + 106, + 531, + 129 + ], + "type": "text", + "content": ". This acts as a bridge between the two modalities, thus encouraging mutual propagation of gradients." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 300, + 139, + 535, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 139, + 535, + 181 + ], + "spans": [ + { + "bbox": [ + 300, + 139, + 535, + 181 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} [ c _ {i}, E _ {i}, \\_ ] = \\mathcal {V} _ {i} ([ c _ {i - 1}, E _ {i - 1}, \\mathcal {F} _ {\\boldsymbol {i} - \\mathbf {1}} (\\boldsymbol {P} _ {\\boldsymbol {i} - \\mathbf {1}}) ]) i = 1, \\dots , J \\\\ [ c _ {j}, E _ {j}, \\tilde {P} _ {j} ] = \\mathcal {V} _ {j} ([ c _ {j - 1}, E _ {j - 1}, \\tilde {P} _ {j - 1} ]) \\quad j = J + 1, \\dots , K \\\\ x = \\operatorname {I m a g e P r o j} \\left(c _ {K}\\right) \\\\ \\end{array}", + "image_path": "407caca810c6e33c985391f016ac21c2f0be0ae96a01a2b1416cc4a5cd20c82d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 298, + 191, + 532, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 191, + 532, + 227 + ], + "spans": [ + { + "bbox": [ + 298, + 191, + 532, + 227 + ], + "type": "text", + "content": "Unlike independent V-L prompting, explicit conditioning of " + }, + { + "bbox": [ + 298, + 191, + 532, + 227 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 298, + 191, + 532, + 227 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 298, + 191, + 532, + 227 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 298, + 191, + 532, + 227 + ], + "type": "text", + "content": " helps learn prompts in a shared embedding space between the two branches, thus improving mutual synergy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 298, + 238, + 376, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 238, + 376, + 251 + ], + "spans": [ + { + "bbox": [ + 298, + 238, + 376, + 251 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 298, + 256, + 406, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 256, + 406, + 269 + ], + "spans": [ + { + "bbox": [ + 298, + 256, + 406, + 269 + ], + "type": "text", + "content": "4.1. Benchmark setting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 298, + 275, + 531, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 275, + 531, + 333 + ], + "spans": [ + { + "bbox": [ + 298, + 275, + 531, + 333 + ], + "type": "text", + "content": "Generalization from Base-to-Novel Classes: We evaluate the generalizability of MaPLe, and follow a zero-shot setting where the datasets are split into base and novel classes. The model is trained only on the base classes in a few-shot setting and evaluated on base and novel categories." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 298, + 334, + 531, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 334, + 531, + 390 + ], + "spans": [ + { + "bbox": [ + 298, + 334, + 531, + 390 + ], + "type": "text", + "content": "Cross-dataset Evaluation: To validate the potential of our approach in cross-dataset transfer, we evaluate our ImageNet trained model directly on other datasets. Consistent with Co-CoOp, our model is trained on all 1000 ImageNet classes in a few-shot manner." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 298, + 392, + 531, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 392, + 531, + 449 + ], + "spans": [ + { + "bbox": [ + 298, + 392, + 531, + 449 + ], + "type": "text", + "content": "Domain Generalization: We evaluate the robustness of our method on out-of-distribution datasets. Similar to cross-dataset evaluation, we test our ImageNet trained model directly on four other ImageNet datasets that contain various types of domain shifts." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 298, + 450, + 531, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 450, + 531, + 613 + ], + "spans": [ + { + "bbox": [ + 298, + 450, + 531, + 613 + ], + "type": "text", + "content": "Datasets: For generalization from base-to-novel classes and cross-dataset evaluation, we follow [48, 49] and evaluate the performance of our method on 11 image classification datasets which covers a wide range of recognition tasks. This includes two generic-objects datasets, ImageNet [4] and Caltech101 [7]; five fine-grained datasets, OxfordPets [31], StanfordCars [20], Flowers102 [30], Food101 [2], and FGVCAircraft [28]; a scene recognition dataset SUN397 [40]; an action recognition dataset UCF101 [36]; a texture dataset DTD [3] and a satellite-image dataset EuroSAT [11]. For domain generalization, we use ImageNet as source dataset and its four variants as target datasets including ImageNetV2 [35], ImageNet-Sketch [37], ImageNet-A [13] and ImageNet-R [12]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "text", + "content": "Implementation Details We use a few-shot training strategy in all experiments at 16 shots which are randomly sampled for each class. We apply prompt tuning on a pretrained ViT-B/16 CLIP model where " + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "inline_equation", + "content": "d_{l} = 512" + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "inline_equation", + "content": "d_{v} = 768" + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "inline_equation", + "content": "d_{vl} = 512" + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "text", + "content": ". For MaPLe, we set prompt depth " + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 298, + 614, + 531, + 730 + ], + "type": "text", + "content": " to 9 and the language and vision prompt lengths to 2. All models are trained for 5 epochs with a batch-size of 4 and a learning rate of 0.0035 via SGD optimizer on a single NVIDIA A100 GPU. We report base and novel class accuracies and their harmonic mean (HM) averaged over 3 runs. We initial-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "type": "text", + "content": "19117" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 103, + 278, + 172 + ], + "blocks": [ + { + "bbox": [ + 48, + 103, + 278, + 172 + ], + "lines": [ + { + "bbox": [ + 48, + 103, + 278, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 278, + 172 + ], + "type": "table", + "html": "
MethodBase Acc.Novel Acc.HMGFLOPS
1: MaPLe shallow (J=1)80.1073.5276.67167.1
2: Deep vision prompting80.2473.4376.6818.0
3: Deep language prompting81.7273.8177.56166.8
4: Independent V-L prompting82.1574.0777.90167.0
5: MaPLe (Ours)82.2875.1478.55167.0
", + "image_path": "b0a8735d5c9c6927cd03257354e8d3bb43ea75c789cb6616fba69bb20cad4748.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 215, + 280, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 215, + 280, + 332 + ], + "spans": [ + { + "bbox": [ + 45, + 215, + 280, + 332 + ], + "type": "text", + "content": "ize the language prompts of the first layer " + }, + { + "bbox": [ + 45, + 215, + 280, + 332 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 45, + 215, + 280, + 332 + ], + "type": "text", + "content": " with the pretrained CLIP word embeddings of the template 'a photo of a ', while for the subsequent layers they are randomly initialized from a normal distribution. For training MaPLe on all 1000 classes of ImageNet as a source model, prompt depth " + }, + { + "bbox": [ + 45, + 215, + 280, + 332 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 45, + 215, + 280, + 332 + ], + "type": "text", + "content": " is set to 3 and the model trained for 2 epochs with learning rate of 0.0026. Hyper-parameters for deep language prompting, deep vision prompting, and independent V-L prompting are detailed in Appendix A. The hyper-parameters are fixed across all datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 338, + 280, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 338, + 280, + 351 + ], + "spans": [ + { + "bbox": [ + 45, + 338, + 280, + 351 + ], + "type": "text", + "content": "4.2. Prompting CLIP via Vision-Language Prompts" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 356, + 280, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 356, + 280, + 649 + ], + "spans": [ + { + "bbox": [ + 45, + 356, + 280, + 649 + ], + "type": "text", + "content": "Prompting Variants: We first evaluate the performance of different possible prompting design choices as an ablation for our proposed branch-aware multi-modal prompting, MaPLe. These variants include shallow MaPLe, deep language prompting, deep vision prompting and independent V-L prompting. In Table 1, we present the results averaged over the 11 image recognition datasets. Shallow MaPLe (row-1) provides consistent improvements over CoOp and Co-CoOp in terms of generalization. Deep language prompting (row-3) shows improvements over deep vision prompting (row-2), indicating that prompts learned at the language branch provide better adaptation of CLIP. Although separately combining the above two approaches (row-4) further improves the performance, it struggles to achieve comprehensive benefits from the language and vision branches. We hypothesize that this is due to the lack of synergy between the learned vision and language prompts as they do not interact with each other during training. Meanwhile, MaPLe tied with deep prompting (row-4) combines the benefits of prompting in both branches by enforcing interactions through explicit conditioning of vision prompts on the language prompts. It provides improvements on novel and base class accuracies which leads to the best HM of " + }, + { + "bbox": [ + 45, + 356, + 280, + 649 + ], + "type": "inline_equation", + "content": "78.55\\%" + }, + { + "bbox": [ + 45, + 356, + 280, + 649 + ], + "type": "text", + "content": ". We explore other possible design choices and present the ablations in Appendix B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 654, + 202, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 654, + 202, + 665 + ], + "spans": [ + { + "bbox": [ + 45, + 654, + 202, + 665 + ], + "type": "text", + "content": "4.3. Base-to-Novel Generalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 671, + 280, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 280, + 731 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 280, + 731 + ], + "type": "text", + "content": "Generalization to Unseen Classes: Table 3 presents the performance of MaPLe in base-to-novel generalization setting on 11 recognition datasets. We compare its performance with CLIP zero-shot, and recent prompt learning works including CoOp [49] and Co-CoOp [48]. In case of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 106, + 531, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 106, + 531, + 128 + ], + "spans": [ + { + "bbox": [ + 297, + 106, + 531, + 128 + ], + "type": "text", + "content": "CLIP, we use hand-crafted prompts that are specifically designed for each dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "spans": [ + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "type": "text", + "content": "In comparison with the state-of-the-art Co-CoOp, MaPLe shows improved performance on both base and novel categories on all 11 datasets with an exception of marginal reduction on only the base class performance of Caltech101. With mutual synergy from the branch-aware multi-modal prompting, MaPLe better generalizes to novel categories on all 11 datasets in comparison with Co-CoOp, and obtains an overall gain from " + }, + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "type": "inline_equation", + "content": "71.69\\%" + }, + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "type": "inline_equation", + "content": "75.14\\%" + }, + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "type": "text", + "content": ". When taking into account both the base and novel classes, MaPLe shows an absolute average gain of " + }, + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "type": "inline_equation", + "content": "2.72\\%" + }, + { + "bbox": [ + 296, + 129, + 532, + 245 + ], + "type": "text", + "content": " over Co-CoOp." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "spans": [ + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "content": "In comparison with CLIP on novel classes, Co-CoOp improves only on 4/11 datasets dropping the average novel accuracy from " + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "inline_equation", + "content": "74.22\\%" + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "inline_equation", + "content": "71.69\\%" + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "content": ". MaPLe is a strong competitor which improves accuracy over CLIP on novel classes on 6/11 datasets, with an average gain from " + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "inline_equation", + "content": "74.22\\%" + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "inline_equation", + "content": "75.14\\%" + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "content": ". Generalization and Performance on Base Classes: CoCoOp solves the poor generalization problem in CoOp by conditioning prompts on image instances and shows significant gains in novel categories. However on base classes, it improves over CoOp only on 3/11 datasets with an average drop in performance from " + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "inline_equation", + "content": "82.69\\%" + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "inline_equation", + "content": "80.47\\%" + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "content": ". Meanwhile, the completeness in prompting helps MaPLe improve over CoOp on base classes in 6/11 datasets maintaining the average base accuracy to around " + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "inline_equation", + "content": "82.28\\%" + }, + { + "bbox": [ + 296, + 245, + 532, + 420 + ], + "type": "text", + "content": ", in addition to its improvement in generalization to novel classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 421, + 531, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 421, + 531, + 479 + ], + "spans": [ + { + "bbox": [ + 296, + 421, + 531, + 479 + ], + "type": "text", + "content": "We find that the training strategies of Co-CoOp can be used to substantially boost the generalization performance of vanilla CoOp (6.8% gain in novel classes). We therefore compare our method with " + }, + { + "bbox": [ + 296, + 421, + 531, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{CoOp}^{\\dagger}" + }, + { + "bbox": [ + 296, + 421, + 531, + 479 + ], + "type": "text", + "content": ", which trains CoOp in Co-CoOp setting (refer to Appendix A for more details)." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 486, + 507, + 554 + ], + "blocks": [ + { + "bbox": [ + 46, + 175, + 280, + 207 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 280, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 280, + 207 + ], + "type": "text", + "content": "Table 1. Comparison of MaPLe with different prompting designs in base-to-novel generalization. Results are averaged over 11 datasets. HM refers to harmonic mean." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 486, + 507, + 554 + ], + "lines": [ + { + "bbox": [ + 323, + 486, + 507, + 554 + ], + "spans": [ + { + "bbox": [ + 323, + 486, + 507, + 554 + ], + "type": "table", + "html": "
BaseNovelHM
CoOp82.6963.2271.66
Co-CoOp80.4771.6975.83
CoOp†80.8570.0275.04
MaPLe82.2875.1478.55
", + "image_path": "945c7e9e155d7cca0b008eb978f0498fbc2f7223110c759bd3b0c55f747d7669.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 555, + 520, + 566 + ], + "lines": [ + { + "bbox": [ + 308, + 555, + 520, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 520, + 566 + ], + "type": "text", + "content": "Table 2. Generalization comparison of MaPLe with CoOp†." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "spans": [ + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "text", + "content": "Compare to " + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{CoOp}^{\\dagger}" + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "text", + "content": ", the vanilla CoOp model seems to overfit on base classes. When compared to " + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{CoOp}^{\\dagger}" + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "text", + "content": " which attains an average base accuracy of " + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "inline_equation", + "content": "80.85\\%" + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "text", + "content": ", MaPLe shows an improvement of " + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "inline_equation", + "content": "1.43\\%" + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "text", + "content": " with the average base accuracy of " + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "inline_equation", + "content": "82.28\\%" + }, + { + "bbox": [ + 296, + 577, + 531, + 636 + ], + "type": "text", + "content": " (Table 2)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 298, + 642, + 436, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 642, + 436, + 654 + ], + "spans": [ + { + "bbox": [ + 298, + 642, + 436, + 654 + ], + "type": "text", + "content": "4.4. Cross-Dataset Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 660, + 532, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 660, + 532, + 731 + ], + "spans": [ + { + "bbox": [ + 296, + 660, + 532, + 731 + ], + "type": "text", + "content": "We test the cross-dataset generalization ability of MaPLe by learning multi-modal prompts on all the 1000 ImageNet classes and then transferring it directly on the remaining 10 datasets. Table 4 shows the performance comparison between MaPLe, CoOp and Co-CoOp. On the ImageNet source dataset, MaPLe achieves performance comparable" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "text", + "content": "19118" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 117, + 203, + 199 + ], + "blocks": [ + { + "bbox": [ + 84, + 107, + 177, + 116 + ], + "lines": [ + { + "bbox": [ + 84, + 107, + 177, + 116 + ], + "spans": [ + { + "bbox": [ + 84, + 107, + 177, + 116 + ], + "type": "text", + "content": "(a) Average over 11 datasets" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 117, + 203, + 199 + ], + "lines": [ + { + "bbox": [ + 59, + 117, + 203, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 117, + 203, + 199 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP69.3474.2271.70
CoOp82.6963.2271.66
Co-CoOp80.4771.6975.83
MaPLe82.2875.1478.55
+1.81+3.45+2.72
", + "image_path": "06e9817516416e3b0ec7c23425fb0b3ef389873303c9de071adee11ce887efdd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 214, + 117, + 360, + 199 + ], + "blocks": [ + { + "bbox": [ + 264, + 107, + 309, + 116 + ], + "lines": [ + { + "bbox": [ + 264, + 107, + 309, + 116 + ], + "spans": [ + { + "bbox": [ + 264, + 107, + 309, + 116 + ], + "type": "text", + "content": "(b) ImageNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 214, + 117, + 360, + 199 + ], + "lines": [ + { + "bbox": [ + 214, + 117, + 360, + 199 + ], + "spans": [ + { + "bbox": [ + 214, + 117, + 360, + 199 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP72.4368.1470.22
CoOp76.4767.8871.92
Co-CoOp75.9870.4373.10
MaPLe76.6670.5473.47
+0.68+0.11+0.37
", + "image_path": "3619c18733eddca1ef0e05bb4f34f924ea5bb7d27bb7157719958df88c8a2b9b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 376, + 117, + 520, + 199 + ], + "blocks": [ + { + "bbox": [ + 424, + 107, + 470, + 116 + ], + "lines": [ + { + "bbox": [ + 424, + 107, + 470, + 116 + ], + "spans": [ + { + "bbox": [ + 424, + 107, + 470, + 116 + ], + "type": "text", + "content": "(c) Caltech101" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 117, + 520, + 199 + ], + "lines": [ + { + "bbox": [ + 376, + 117, + 520, + 199 + ], + "spans": [ + { + "bbox": [ + 376, + 117, + 520, + 199 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP96.8494.0095.40
CoOp98.0089.8193.73
Co-CoOp97.9693.8195.84
MaPLe97.7494.3696.02
-0.22+0.55+0.18
", + "image_path": "8a7f7c3134986e91cdf26af50eef6b0fdd831ee2514051b473c8bcd4fcc2169f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 220, + 203, + 301 + ], + "blocks": [ + { + "bbox": [ + 107, + 210, + 154, + 218 + ], + "lines": [ + { + "bbox": [ + 107, + 210, + 154, + 218 + ], + "spans": [ + { + "bbox": [ + 107, + 210, + 154, + 218 + ], + "type": "text", + "content": "(d) OxfordPets" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 220, + 203, + 301 + ], + "lines": [ + { + "bbox": [ + 58, + 220, + 203, + 301 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 203, + 301 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP91.1797.2694.12
CoOp93.6795.2994.47
Co-CoOp95.2097.6996.43
MaPLe95.4397.7696.58
+0.23+0.07+0.15
", + "image_path": "54c951ed922ac44768efa32313d93027d5fc2634fd873060d75c7627447ba378.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 214, + 220, + 360, + 301 + ], + "blocks": [ + { + "bbox": [ + 261, + 210, + 313, + 218 + ], + "lines": [ + { + "bbox": [ + 261, + 210, + 313, + 218 + ], + "spans": [ + { + "bbox": [ + 261, + 210, + 313, + 218 + ], + "type": "text", + "content": "(e) StanfordCars" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 214, + 220, + 360, + 301 + ], + "lines": [ + { + "bbox": [ + 214, + 220, + 360, + 301 + ], + "spans": [ + { + "bbox": [ + 214, + 220, + 360, + 301 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP63.3774.8968.65
CoOp78.1260.4068.13
Co-CoOp70.4973.5972.01
MaPLe72.9474.0073.47
+2.45+0.41+1.46
", + "image_path": "6355c37d6f4cef47b5f8f81e71edd0646469ebc8883a98fb02b8d0fba6ee83b2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 376, + 220, + 520, + 301 + ], + "blocks": [ + { + "bbox": [ + 424, + 210, + 471, + 218 + ], + "lines": [ + { + "bbox": [ + 424, + 210, + 471, + 218 + ], + "spans": [ + { + "bbox": [ + 424, + 210, + 471, + 218 + ], + "type": "text", + "content": "(f) Flowers102" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 220, + 520, + 301 + ], + "lines": [ + { + "bbox": [ + 376, + 220, + 520, + 301 + ], + "spans": [ + { + "bbox": [ + 376, + 220, + 520, + 301 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP72.0877.8074.83
CoOp97.6059.6774.06
Co-CoOp94.8771.7581.71
MaPLe95.9272.4682.56
+1.05+0.71+0.85
", + "image_path": "d5b64891f6c01ce49e140225afbf35d5329d56e29c4f91bb72d76069d6b50e16.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 58, + 322, + 203, + 404 + ], + "blocks": [ + { + "bbox": [ + 110, + 313, + 149, + 321 + ], + "lines": [ + { + "bbox": [ + 110, + 313, + 149, + 321 + ], + "spans": [ + { + "bbox": [ + 110, + 313, + 149, + 321 + ], + "type": "text", + "content": "(g) Food101" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 322, + 203, + 404 + ], + "lines": [ + { + "bbox": [ + 58, + 322, + 203, + 404 + ], + "spans": [ + { + "bbox": [ + 58, + 322, + 203, + 404 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP90.1091.2290.66
CoOp88.3382.2685.19
Co-CoOp90.7091.2990.99
MaPLe90.7192.0591.38
+0.01+0.76+0.39
", + "image_path": "b4cbc69e46a5d3e0f2efd826cdec1a1072ea7b8c64da064a219eb1940715ccfd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 212, + 322, + 361, + 404 + ], + "blocks": [ + { + "bbox": [ + 258, + 313, + 315, + 321 + ], + "lines": [ + { + "bbox": [ + 258, + 313, + 315, + 321 + ], + "spans": [ + { + "bbox": [ + 258, + 313, + 315, + 321 + ], + "type": "text", + "content": "(h) FGVCAircraft" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 322, + 361, + 404 + ], + "lines": [ + { + "bbox": [ + 212, + 322, + 361, + 404 + ], + "spans": [ + { + "bbox": [ + 212, + 322, + 361, + 404 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP27.1936.2931.09
CoOp40.4422.3028.75
Co-CoOp33.4123.7127.74
MaPLe37.4435.6136.50
+4.03+11.90+8.76
", + "image_path": "133110c6e00a0f267d09280908e2fa45186b6044718c6f5ed383d8b895dfdb93.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 376, + 322, + 520, + 404 + ], + "blocks": [ + { + "bbox": [ + 429, + 313, + 466, + 321 + ], + "lines": [ + { + "bbox": [ + 429, + 313, + 466, + 321 + ], + "spans": [ + { + "bbox": [ + 429, + 313, + 466, + 321 + ], + "type": "text", + "content": "(i) SUN397" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 322, + 520, + 404 + ], + "lines": [ + { + "bbox": [ + 376, + 322, + 520, + 404 + ], + "spans": [ + { + "bbox": [ + 376, + 322, + 520, + 404 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP69.3675.3572.23
CoOp80.6065.8972.51
Co-CoOp79.7476.8678.27
MaPLe80.8278.7079.75
+1.08+1.84+1.48
", + "image_path": "95cae59c394804fbd7b8c33f9e20be71a37f584a58a3ebd34adaced1376ff5ee.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 56, + 425, + 200, + 506 + ], + "blocks": [ + { + "bbox": [ + 115, + 415, + 141, + 424 + ], + "lines": [ + { + "bbox": [ + 115, + 415, + 141, + 424 + ], + "spans": [ + { + "bbox": [ + 115, + 415, + 141, + 424 + ], + "type": "text", + "content": "(j) DTD" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 425, + 200, + 506 + ], + "lines": [ + { + "bbox": [ + 56, + 425, + 200, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 425, + 200, + 506 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP53.2459.9056.37
CoOp79.4441.1854.24
Co-CoOp77.0156.0064.85
MaPLe80.3659.1868.16
+3.35+3.18+3.31
", + "image_path": "577a2d66be88e625d8fd59f1aea16f4ac14eef2ef9b49ed291f040fe94aa848c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 212, + 425, + 365, + 505 + ], + "blocks": [ + { + "bbox": [ + 268, + 415, + 309, + 423 + ], + "lines": [ + { + "bbox": [ + 268, + 415, + 309, + 423 + ], + "spans": [ + { + "bbox": [ + 268, + 415, + 309, + 423 + ], + "type": "text", + "content": "(k) EuroSAT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 425, + 365, + 505 + ], + "lines": [ + { + "bbox": [ + 212, + 425, + 365, + 505 + ], + "spans": [ + { + "bbox": [ + 212, + 425, + 365, + 505 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP56.4864.0560.03
CoOp92.1954.7468.69
Co-CoOp87.4960.0471.21
MaPLe94.0773.2382.35
+6.58+13.19+11.14
", + "image_path": "f2fc36fd7302cc7a490a2fc72ec3f1f9bcec67dc30e646879b0793354860352f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 377, + 425, + 522, + 505 + ], + "blocks": [ + { + "bbox": [ + 431, + 415, + 467, + 423 + ], + "lines": [ + { + "bbox": [ + 431, + 415, + 467, + 423 + ], + "spans": [ + { + "bbox": [ + 431, + 415, + 467, + 423 + ], + "type": "text", + "content": "(1)UCF101" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 377, + 425, + 522, + 505 + ], + "lines": [ + { + "bbox": [ + 377, + 425, + 522, + 505 + ], + "spans": [ + { + "bbox": [ + 377, + 425, + 522, + 505 + ], + "type": "table", + "html": "
BaseNovelHM
CLIP70.5377.5073.85
CoOp84.6956.0567.46
Co-CoOp82.3373.4577.64
MaPLe83.0078.6680.77
+0.67+5.21+3.13
", + "image_path": "a1f7f6ee83b39cf01fb3f18952d54219c91e552ef9f12862b19ea8ec503c91bf.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 60, + 539, + 430, + 634 + ], + "blocks": [ + { + "bbox": [ + 45, + 507, + 530, + 530 + ], + "lines": [ + { + "bbox": [ + 45, + 507, + 530, + 530 + ], + "spans": [ + { + "bbox": [ + 45, + 507, + 530, + 530 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art methods on base-to-novel generalization. MaPLe learns multi-modal prompts and demonstrates strong generalization results over existing methods on 11 recognition datasets. Absolute gains over Co-CoOp are indicated in blue." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 539, + 430, + 634 + ], + "lines": [ + { + "bbox": [ + 60, + 539, + 430, + 634 + ], + "spans": [ + { + "bbox": [ + 60, + 539, + 430, + 634 + ], + "type": "table", + "html": "
SourceTarget
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101AircraftSUN397DTDEuroSATUCF101Average
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
Co-CoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
MaPLe70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
", + "image_path": "380b68b8c347cda96b24504a99374ff0eb1f9a2c205065f4ce270b051671b05f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 435, + 541, + 531, + 636 + ], + "lines": [ + { + "bbox": [ + 435, + 541, + 531, + 636 + ], + "spans": [ + { + "bbox": [ + 435, + 541, + 531, + 636 + ], + "type": "text", + "content": "Table 4. Comparison of MaPLe with existing approaches on cross-dataset evaluation. Overall, MaPLe achieves competitive performance providing highest average accuracy, indicating better generalization." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 658, + 280, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 658, + 280, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 658, + 280, + 729 + ], + "type": "text", + "content": "to competing approaches but demonstrates a much stronger generalization performance by surpassing CoOp in 9/10 and Co-CoOp in 8/10 datasets. Overall, MaPLe shows competitive performance leading to the highest averaged accuracy of " + }, + { + "bbox": [ + 45, + 658, + 280, + 729 + ], + "type": "inline_equation", + "content": "66.30\\%" + }, + { + "bbox": [ + 45, + 658, + 280, + 729 + ], + "type": "text", + "content": ". This suggests that the use of branch-aware V-L prompting in MaPLe facilitates better generalization." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 298, + 657, + 426, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 657, + 426, + 668 + ], + "spans": [ + { + "bbox": [ + 298, + 657, + 426, + 668 + ], + "type": "text", + "content": "4.5. Domain Generalization" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 297, + 671, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 671, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 671, + 531, + 730 + ], + "type": "text", + "content": "We show that MaPLe generalizes favourably on out-of-distribution datasets as compared to CoOp and Co-CoOp. We evaluate the direct transferability of ImageNet trained model to various out-of-domain datasets, and observe that it consistently improves against all the existing approaches" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "type": "text", + "content": "19119" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 103, + 284, + 187 + ], + "blocks": [ + { + "bbox": [ + 47, + 103, + 284, + 187 + ], + "lines": [ + { + "bbox": [ + 47, + 103, + 284, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 284, + 187 + ], + "type": "table", + "html": "
SourceTarget
ImageNetImageNetV2ImageNet-SImageNet-AImageNet-R
CLIP66.7360.8346.1547.7773.96
CoOp71.5164.2047.9949.7175.21
Co-CoOp71.0264.0748.7550.6376.18
MaPLe70.7264.0749.1550.9076.98
", + "image_path": "8d7563e65cfa764b3bf696aee2d61b18b7a1ec1d659cb759949414ad8553b05c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 45, + 188, + 280, + 221 + ], + "lines": [ + { + "bbox": [ + 45, + 188, + 280, + 221 + ], + "spans": [ + { + "bbox": [ + 45, + 188, + 280, + 221 + ], + "type": "text", + "content": "Table 5. Comparison of MaPLe with existing approaches in domain generalization setting. MaPLe shows consistent improvements on all target datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 229, + 280, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 229, + 280, + 265 + ], + "spans": [ + { + "bbox": [ + 45, + 229, + 280, + 265 + ], + "type": "text", + "content": "as indicated in Table 5. This indicates that utilizing multimodal branch-aware prompting helps MaPLe in enhancing the generalization and robustness of V-L models like CLIP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 271, + 169, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 271, + 169, + 283 + ], + "spans": [ + { + "bbox": [ + 45, + 271, + 169, + 283 + ], + "type": "text", + "content": "4.6. Ablation Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 289, + 280, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 289, + 280, + 429 + ], + "spans": [ + { + "bbox": [ + 45, + 289, + 280, + 429 + ], + "type": "text", + "content": "Prompt Depth: In Fig. 4 (left), we illustrate the effect of prompt depth " + }, + { + "bbox": [ + 45, + 289, + 280, + 429 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 45, + 289, + 280, + 429 + ], + "type": "text", + "content": " for MaPLe and ablate on the depth of language and vision branch individually. In general, the performance improves as prompt depth increases. We note that performance sensitivity increases when randomly initialized prompts are inserted in the deeper layers of a frozen model where the model feature space is already mature. Similar trend is also reported by [16]. As earlier methods utilize shallow language prompting (" + }, + { + "bbox": [ + 45, + 289, + 280, + 429 + ], + "type": "inline_equation", + "content": "J = 1" + }, + { + "bbox": [ + 45, + 289, + 280, + 429 + ], + "type": "text", + "content": "), we compare our method with deep language prompting. Overall, MaPLe achieves better performance than deep language prompting and achieves maximum performance at a depth of 9." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 429, + 280, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 429, + 280, + 488 + ], + "spans": [ + { + "bbox": [ + 45, + 429, + 280, + 488 + ], + "type": "text", + "content": "Prompt Length: Fig. 4 (right) shows the effect of prompt length for MaPLe. As the prompt length increases, the performance on base classes is generally maintained, while the novel class accuracy decreases. This indicates over-fitting which inherently hurts the generalization to novel classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 488, + 280, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 488, + 280, + 592 + ], + "spans": [ + { + "bbox": [ + 45, + 488, + 280, + 592 + ], + "type": "text", + "content": "Effectiveness of Multi-modal Prompting: Fig. 5 shows the analysis of per class accuracy for selected datasets in the order of increasing domain shift. It indicates that the performance gains of MaPLe in comparison to Co-CoOp varies across different datasets. MaPLe provides significant gains over Co-CoOp for datasets that have large distribution shifts from the pretraining dataset of CLIP, and vision concepts that are usually rare and less generic. Further detailed analysis is provided in Appendix C." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 592, + 279, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 592, + 279, + 617 + ], + "spans": [ + { + "bbox": [ + 45, + 592, + 279, + 617 + ], + "type": "text", + "content": "Prompting complexity: Table 6 shows the computational complexity of MaPLe in comparison with other approaches." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 47, + 621, + 165, + 693 + ], + "blocks": [ + { + "bbox": [ + 47, + 621, + 165, + 693 + ], + "lines": [ + { + "bbox": [ + 47, + 621, + 165, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 165, + 693 + ], + "type": "image", + "image_path": "8288210b3d7acda788e8c59c8fd82ca205aec09731370e88eb536d6e1d14a208.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 694, + 279, + 726 + ], + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 726 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 726 + ], + "type": "text", + "content": "Figure 4. Ablation on prompt depth (left) and prompt length (right) in MaPLe. We report average results on the held-out validation sets of all datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 163, + 621, + 278, + 693 + ], + "blocks": [ + { + "bbox": [ + 163, + 621, + 278, + 693 + ], + "lines": [ + { + "bbox": [ + 163, + 621, + 278, + 693 + ], + "spans": [ + { + "bbox": [ + 163, + 621, + 278, + 693 + ], + "type": "image", + "image_path": "d3568a4077d46c53fe89df4bc2fa3cdf22236decbe0d489c71d5f5e46f955731.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 302, + 117, + 430, + 197 + ], + "blocks": [ + { + "bbox": [ + 302, + 117, + 430, + 197 + ], + "lines": [ + { + "bbox": [ + 302, + 117, + 430, + 197 + ], + "spans": [ + { + "bbox": [ + 302, + 117, + 430, + 197 + ], + "type": "image", + "image_path": "ac8ffce85353eb3f418849e2d8a5ffb6e4f88a1db2ce83d86ea563fd8c57a052.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 104, + 531, + 190 + ], + "lines": [ + { + "bbox": [ + 438, + 104, + 531, + 190 + ], + "spans": [ + { + "bbox": [ + 438, + 104, + 531, + 190 + ], + "type": "text", + "content": "Figure 5. Percentage classes where MaPLe shows improved performance over Co-CoOp, which increases as dataset domain shift from generic categories increases " + }, + { + "bbox": [ + 438, + 104, + 531, + 190 + ], + "type": "inline_equation", + "content": "(\\rightarrow)" + }, + { + "bbox": [ + 438, + 104, + 531, + 190 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "spans": [ + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "text", + "content": "Although MaPLe utilizes multi-modal prompts, its overall FLOPS (Floating Point Operations) exceeds only by " + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "text", + "content": " over CoOp and Co-CoOp. The independent V-L prompting also provides comparable FLOP count. In terms of inference speed, Co-CoOp is significantly slower and the FPS (Frames Per Second) remains constant as the batch size increases. In contrast, MaPLe has no such overhead and provides much better inference and training speeds. Further, MaPLe provides better convergence as it requires only half training epochs as compared to Co-CoOp (5 vs 10 epochs). MaPLe adds about " + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "inline_equation", + "content": "2.85\\%" + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "text", + "content": " training parameters on top of CLIP. To study if the performance gain is mainly attributed to more parameters, we experiment with " + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{MaPLe\\dagger}" + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "text", + "content": ", which uses a unified V-L coupling function for all layer prompts. " + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{MaPLe\\dagger}" + }, + { + "bbox": [ + 296, + 198, + 533, + 397 + ], + "type": "text", + "content": " with about 9x lesser parameters than MaPLe also improves over existing methods. We also ablate by comparing MaPLe with heavier CoCoOp in Appendix D." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 300, + 405, + 532, + 490 + ], + "blocks": [ + { + "bbox": [ + 300, + 405, + 532, + 490 + ], + "lines": [ + { + "bbox": [ + 300, + 405, + 532, + 490 + ], + "spans": [ + { + "bbox": [ + 300, + 405, + 532, + 490 + ], + "type": "table", + "html": "
MethodParamsParams % CLIPFPS (with BS)HM
14100
CoOp20480.00213.855.31353.071.66
CoCoOp353600.0364.6114.715.175.83
Independent V-L314880.0262.5239.41383.877.90
MaPLe3.55 M2.8560.2239.01365.178.55
MaPLe†0.41 M0.3360.2238.01365.078.11
", + "image_path": "70c54e958a78e1a21541dda019db13d58dbd1940304c97ab35a46e7a63e3997e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 297, + 491, + 531, + 524 + ], + "lines": [ + { + "bbox": [ + 297, + 491, + 531, + 524 + ], + "spans": [ + { + "bbox": [ + 297, + 491, + 531, + 524 + ], + "type": "text", + "content": "Table 6. Comparison of computational complexity among different prompting methods. MaPLe† is a MaPLe version which utilizes a common V-L coupling function for all layers." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 14 + }, + { + "bbox": [ + 298, + 536, + 369, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 536, + 369, + 548 + ], + "spans": [ + { + "bbox": [ + 298, + 536, + 369, + 548 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 555, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 555, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 555, + 532, + 730 + ], + "type": "text", + "content": "Adaptation of large-scale V-L models, e.g., CLIP [32] to downstream tasks is a challenging problem due to the large number of tunable parameters and limited size of downstream datasets. Prompt learning is an efficient and scalable technique to tailor V-L models to novel downstream tasks. To this end, the current prompt learning approaches either consider only the vision or language side prompting. Our work shows that it is critical to perform prompting for both vision and language branches to appropriately adapt V-L models to downstream tasks. Further, we propose a strategy to ensure synergy between vision-language modalities by explicitly conditioning the vision prompts on textual prompt across different transformer stages. Our approach improves the generalization towards novel categories, cross-dataset transfer and datasets with domain shifts." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 302, + 756 + ], + "type": "text", + "content": "19120" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 104, + 104, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 104, + 104, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 104, + 104, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 281, + 730 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 51, + 124, + 280, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 280, + 166 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 280, + 166 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 2022. 3, 11, 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 169, + 281, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 169, + 281, + 212 + ], + "spans": [ + { + "bbox": [ + 51, + 169, + 281, + 212 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In The European Conference on Computer Vision, pages 446-461. Springer, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 213, + 280, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 213, + 280, + 266 + ], + "spans": [ + { + "bbox": [ + 51, + 213, + 280, + 266 + ], + "type": "text", + "content": "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 267, + 280, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 267, + 280, + 321 + ], + "spans": [ + { + "bbox": [ + 51, + 267, + 280, + 321 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 323, + 281, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 323, + 281, + 366 + ], + "spans": [ + { + "bbox": [ + 51, + 323, + 281, + 366 + ], + "type": "text", + "content": "[5] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 367, + 280, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 367, + 280, + 432 + ], + "spans": [ + { + "bbox": [ + 51, + 367, + 280, + 432 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 433, + 280, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 433, + 280, + 487 + ], + "spans": [ + { + "bbox": [ + 51, + 433, + 280, + 487 + ], + "type": "text", + "content": "[7] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 489, + 280, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 489, + 280, + 541 + ], + "spans": [ + { + "bbox": [ + 51, + 489, + 280, + 541 + ], + "type": "text", + "content": "[8] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In The European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 543, + 280, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 543, + 280, + 586 + ], + "spans": [ + { + "bbox": [ + 51, + 543, + 280, + 586 + ], + "type": "text", + "content": "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 588, + 280, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 280, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 280, + 629 + ], + "type": "text", + "content": "[10] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 632, + 280, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 280, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 280, + 685 + ], + "type": "text", + "content": "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 687, + 280, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 280, + 730 + ], + "type": "text", + "content": "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 531, + 730 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 317, + 106, + 531, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 531, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 531, + 128 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 129, + 531, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 129, + 531, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 129, + 531, + 171 + ], + "type": "text", + "content": "[13] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 172, + 531, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 172, + 531, + 204 + ], + "spans": [ + { + "bbox": [ + 299, + 172, + 531, + 204 + ], + "type": "text", + "content": "[14] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 300, + 206, + 531, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 206, + 531, + 269 + ], + "spans": [ + { + "bbox": [ + 300, + 206, + 531, + 269 + ], + "type": "text", + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 300, + 271, + 531, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 271, + 531, + 314 + ], + "spans": [ + { + "bbox": [ + 300, + 271, + 531, + 314 + ], + "type": "text", + "content": "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In The European Conference on Computer Vision, 2022. 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 300, + 314, + 531, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 314, + 531, + 367 + ], + "spans": [ + { + "bbox": [ + 300, + 314, + 531, + 367 + ], + "type": "text", + "content": "[17] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt is worth millions of parameters? low-resource prompt-based learning for vision-language models. arXiv preprint arXiv:2110.08484, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 300, + 369, + 531, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 369, + 531, + 412 + ], + "spans": [ + { + "bbox": [ + 300, + 369, + 531, + 412 + ], + "type": "text", + "content": "[18] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In The European Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 300, + 413, + 531, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 413, + 531, + 445 + ], + "spans": [ + { + "bbox": [ + 300, + 413, + 531, + 445 + ], + "type": "text", + "content": "[19] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 300, + 447, + 531, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 447, + 531, + 489 + ], + "spans": [ + { + "bbox": [ + 300, + 447, + 531, + 489 + ], + "type": "text", + "content": "[20] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 554–561, 2013. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 300, + 490, + 531, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 490, + 531, + 532 + ], + "spans": [ + { + "bbox": [ + 300, + 490, + 531, + 532 + ], + "type": "text", + "content": "[21] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In _Conference on Empirical Methods in Natural Language Processing_, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 300, + 533, + 531, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 533, + 531, + 576 + ], + "spans": [ + { + "bbox": [ + 300, + 533, + 531, + 576 + ], + "type": "text", + "content": "[22] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 300, + 578, + 531, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 578, + 531, + 631 + ], + "spans": [ + { + "bbox": [ + 300, + 578, + 531, + 631 + ], + "type": "text", + "content": "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 300, + 633, + 531, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 633, + 531, + 706 + ], + "spans": [ + { + "bbox": [ + 300, + 633, + 531, + 706 + ], + "type": "text", + "content": "[24] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 708, + 531, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 708, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 708, + 531, + 730 + ], + "type": "text", + "content": "[25] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceed-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 300, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 300, + 757 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 300, + 757 + ], + "type": "text", + "content": "19121" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 106, + 280, + 730 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 65, + 106, + 280, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 106, + 280, + 127 + ], + "spans": [ + { + "bbox": [ + 65, + 106, + 280, + 127 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 280, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 280, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 280, + 170 + ], + "type": "text", + "content": "[26] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 171, + 280, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 171, + 280, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 280, + 224 + ], + "type": "text", + "content": "[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In The European Conference on Computer Vision. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 225, + 280, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 225, + 280, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 225, + 280, + 266 + ], + "type": "text", + "content": "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 280, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 280, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 280, + 310 + ], + "type": "text", + "content": "[29] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 311, + 280, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 311, + 280, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 280, + 353 + ], + "type": "text", + "content": "[30] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 280, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 280, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 280, + 395 + ], + "type": "text", + "content": "[31] Omkar M Parkhi, Andreae Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 397, + 280, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 397, + 280, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 397, + 280, + 461 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 462, + 280, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 280, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 280, + 524 + ], + "type": "text", + "content": "[33] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 280, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 280, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 280, + 578 + ], + "type": "text", + "content": "[34] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 280, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 280, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 280, + 622 + ], + "type": "text", + "content": "[35] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 623, + 280, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 280, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 280, + 654 + ], + "type": "text", + "content": "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 655, + 280, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 655, + 280, + 697 + ], + "spans": [ + { + "bbox": [ + 48, + 655, + 280, + 697 + ], + "type": "text", + "content": "[37] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, volume 32, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 698, + 280, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 698, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 48, + 698, + 280, + 730 + ], + "type": "text", + "content": "[38] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 531, + 726 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 317, + 106, + 531, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 531, + 127 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 531, + 127 + ], + "type": "text", + "content": "prompting for rehearsal-free continual learning. In The European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 128, + 531, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 128, + 531, + 191 + ], + "spans": [ + { + "bbox": [ + 299, + 128, + 531, + 191 + ], + "type": "text", + "content": "[39] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 192, + 531, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 192, + 531, + 245 + ], + "spans": [ + { + "bbox": [ + 299, + 192, + 531, + 245 + ], + "type": "text", + "content": "[40] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 246, + 531, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 246, + 531, + 288 + ], + "spans": [ + { + "bbox": [ + 299, + 246, + 531, + 288 + ], + "type": "text", + "content": "[41] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 289, + 531, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 289, + 531, + 330 + ], + "spans": [ + { + "bbox": [ + 299, + 289, + 531, + 330 + ], + "type": "text", + "content": "[42] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 331, + 531, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 331, + 531, + 383 + ], + "spans": [ + { + "bbox": [ + 299, + 331, + 531, + 383 + ], + "type": "text", + "content": "[43] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 384, + 531, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 384, + 531, + 426 + ], + "spans": [ + { + "bbox": [ + 299, + 384, + 531, + 426 + ], + "type": "text", + "content": "[44] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In The European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 427, + 531, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 427, + 531, + 480 + ], + "spans": [ + { + "bbox": [ + 299, + 427, + 531, + 480 + ], + "type": "text", + "content": "[45] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18123-18133, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 481, + 531, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 481, + 531, + 533 + ], + "spans": [ + { + "bbox": [ + 299, + 481, + 531, + 533 + ], + "type": "text", + "content": "[46] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In The European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 534, + 531, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 534, + 531, + 555 + ], + "spans": [ + { + "bbox": [ + 299, + 534, + 531, + 555 + ], + "type": "text", + "content": "[47] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 556, + 531, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 556, + 531, + 607 + ], + "spans": [ + { + "bbox": [ + 299, + 556, + 531, + 607 + ], + "type": "text", + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 609, + 531, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 609, + 531, + 650 + ], + "spans": [ + { + "bbox": [ + 299, + 609, + 531, + 650 + ], + "type": "text", + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, pages 1-12, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 652, + 531, + 694 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 652, + 531, + 694 + ], + "spans": [ + { + "bbox": [ + 299, + 652, + 531, + 694 + ], + "type": "text", + "content": "[50] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In The European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 695, + 531, + 726 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 695, + 531, + 726 + ], + "spans": [ + { + "bbox": [ + 299, + 695, + 531, + 726 + ], + "type": "text", + "content": "[51] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. arXiv preprint arXiv:2205.14865, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 301, + 756 + ], + "type": "text", + "content": "19122" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 596, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_content_list.json b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..91782c130aa6caaaec2a34483e2068c03c7e23c4 --- /dev/null +++ b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_content_list.json @@ -0,0 +1,1533 @@ +[ + { + "type": "text", + "text": "Magic3D: High-Resolution Text-to-3D Content Creation", + "text_level": 1, + "bbox": [ + 200, + 130, + 769, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen-Hsuan Lin* Jun Gao* Luming Tang* Towaki Takikawa* Xiaohui Zeng* Xun Huang Karsten Kreis Sanja Fidler† Ming-Yu Liu† Tsung-Yi Lin", + "bbox": [ + 140, + 180, + 826, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NVIDIA Corporation", + "bbox": [ + 393, + 223, + 570, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://research.nvidia.com/labs/dir/magic3d", + "bbox": [ + 299, + 244, + 663, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 296, + 313, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DreamFusion [31] has recently demonstrated the utility of a pre-trained text-to-image diffusion model to optimize Neural Radiance Fields (NeRF) [23], achieving remarkable text-to-3D synthesis results. However, the method has two inherent limitations: (a) extremely slow optimization of NeRF and (b) low-resolution image space supervision on NeRF, leading to low-quality 3D models with a long processing time. In this paper, we address these limitations by utilizing a two-stage optimization framework. First, we obtain a coarse model using a low-resolution diffusion prior and accelerate with a sparse 3D hash grid structure. Using the coarse representation as the initialization, we further optimize a textured 3D mesh model with an efficient differentiable renderer interacting with a high-resolution latent diffusion model. Our method, dubbed Magic3D, can create high quality 3D mesh models in 40 minutes, which is $2 \\times$ faster than DreamFusion (reportedly taking 1.5 hours on average), while also achieving higher resolution. User studies show $61.7\\%$ raters to prefer our approach over DreamFusion. Together with the image-conditioned generation capabilities, we provide users with new ways to control 3D synthesis, opening up new avenues to various creative applications.", + "bbox": [ + 76, + 327, + 473, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D digital content has been in high demand for a variety of applications, including gaming, entertainment, architecture, and robotics simulation. It is slowly finding its way into virtually every possible domain: retail, online conferencing, virtual social presence, education, etc. However, creating professional 3D content is not for anyone — it requires immense artistic and aesthetic training with 3D modeling expertise. Developing these skill sets takes a significant amount of time and effort. Augmenting 3D content creation with natural language could considerably help democratize 3D content creation for novices and turbocharge expert artists.", + "bbox": [ + 76, + 712, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image content creation from text prompts [2, 28, 33, 36] has seen significant progress with the advances of diffusion models [13, 41, 42] for generative modeling of images. The key enablers are large-scale datasets comprising billions of samples (images with text) scrapped from the Internet and massive amounts of compute. In contrast, 3D content generation has progressed at a much slower pace. Existing 3D object generation models [4, 9, 47] are mostly categorical. A trained model can only be used to synthesize objects for a single class, with early signs of scaling to multiple classes shown recently by Zeng et al. [47]. Therefore, what a user can do with these models is extremely limited and not yet ready for artistic creation. This limitation is largely due to the lack of diverse large-scale 3D datasets — compared to image and video content, 3D content is much less accessible on the Internet. This naturally raises the question of whether 3D generation capability can be achieved by leveraging powerful text-to-image generative models.", + "bbox": [ + 500, + 297, + 893, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, DreamFusion [31] demonstrated its remarkable ability for text-conditioned 3D content generation by utilizing a pre-trained text-to-image diffusion model [36] that generates images as a strong image prior. The diffusion model acts as a critic to optimize the underlying 3D representation. The optimization process ensures that rendered images from a 3D model, represented by Neural Radiance Fields (NeRF) [23], match the distribution of photorealistic images across different viewpoints, given the input text prompt. Since the supervision signal in DreamFusion operates on very low-resolution images $(64 \\times 64)$ , DreamFusion cannot synthesize high-frequency 3D geometric and texture details. Due to the use of inefficient MLP architectures for the NeRF representation, practical high-resolution synthesis may not even be possible as the required memory footprint and the computation budget grows quickly with the resolution. Even at a resolution of $64 \\times 64$ , optimization times are in hours (1.5 hours per prompt on average using TPUv4).", + "bbox": [ + 498, + 575, + 895, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a method that can synthesize highly detailed 3D models from text prompts within a reduced computation time. Specifically, we propose a coarse-", + "bbox": [ + 498, + 854, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\ast \\dagger}$ : equal contribution.", + "bbox": [ + 89, + 886, + 210, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "300", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6080e0d70aab5a7349d9ec09e3d7cd274505bee320234f86bbb32c16dd37c18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 133, + 313, + 281 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b8245c2c17cd88f41899b3052ead17133ef619ff7ad49ee3c13a8bd95c7d5f5.jpg", + "image_caption": [ + "michelangelo style statue of an astronaut" + ], + "image_footnote": [], + "bbox": [ + 328, + 90, + 431, + 316 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c71b85b08404d1ddfe504a5f8c22cafa8c3d86eb243d74b9b468882241275af.jpg", + "image_caption": [ + "a stuffed grey rabbit holding a pretend carrot" + ], + "image_footnote": [], + "bbox": [ + 468, + 98, + 622, + 309 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6f428019d63476ad15c8772a3814a6a3c272bd31916bb5ec79bb7bdf370567ba.jpg", + "image_caption": [ + "an iguana holding a balloon" + ], + "image_footnote": [], + "bbox": [ + 697, + 154, + 893, + 301 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2f151281677b782a6d0bf393a5109d239305e750e398997cbc0752d41bb9dc72.jpg", + "image_caption": [ + "a silver platter piled high with fruits" + ], + "image_footnote": [], + "bbox": [ + 91, + 329, + 207, + 508 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fc43f3109dee508f9f835868839a0bcf438823adfcb214d1494dc0733f869f37.jpg", + "image_caption": [ + "a beautiful dress made out of garbage bags", + "an imperial state crown of england" + ], + "image_footnote": [], + "bbox": [ + 227, + 348, + 395, + 510 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/829b8fdf0d1f6df006853c44bda46266ed7158c5c1291ff309f56b611df95d09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 364, + 624, + 486 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3dc0649342845e4741b346842db3a6a860a0577be4336ef28ba4362f6d77ea11.jpg", + "image_caption": [ + "neuschwanstein castle, aerial view" + ], + "image_footnote": [], + "bbox": [ + 635, + 354, + 890, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low resolution bunny before editing", + "bbox": [ + 96, + 556, + 241, + 584 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a baby bunny sitting on top of a stack of pancakes", + "bbox": [ + 94, + 594, + 184, + 660 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bca0e717bd1f384e994c7764818373ae4f37872048de5e7149cf01c0ebb22fc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 595, + 243, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a metal bunny sitting on top of a stack of broccoli", + "bbox": [ + 261, + 547, + 320, + 625 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52e09933ed0dc7046df0273ae8facc1f4ac775854703f1ea20cee58d3c042ae0.jpg", + "image_caption": [ + "a blue poison-dart frog sitting on a water lily", + "Figure 1. Results and applications of Magic3D. Top: high-resolution text-to-3D generation. Magic3D can generate high-quality and high-resolution 3D models from text prompts. Bottom: high-resolution prompt-based editing. Magic3D can edit 3D models by fine-tuning with the diffusion prior using a different prompt. Taking the low-resolution 3D model as the input (left), Magic3D can modify different parts of the 3D model corresponding to different input text prompts. Together with various creative controls on the generated 3D models, Magic3D is a convenient tool for augmenting 3D content creation." + ], + "image_footnote": [], + "bbox": [ + 313, + 545, + 450, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a metalbunnysitting ontop of asstack ofchocolatecookie", + "bbox": [ + 470, + 547, + 534, + 637 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e34718cd295dcc17dc3640658aa58a4c5c84c0ff911be780ac3087f184c6f636.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 545, + 658, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a sphinx sitting on top of a stack of chocolate cookie", + "bbox": [ + 681, + 547, + 750, + 626 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/81d041100aaeb45340839b4086dd5e57d9051d41e929f88339d3842fb4318da0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 544, + 861, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to-fine optimization approach that uses multiple diffusion priors at different resolutions to optimize the 3D representation, enabling the generation of both view-consistent geometry as well as high-resolution details. In the first stage, we optimize a coarse neural field representation akin to Dream-Fusion, but with a memory- and compute-efficient scene representation based on a hash grid [25]. In the second stage,", + "bbox": [ + 75, + 780, + 472, + 887 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we switch to optimizing mesh representations, a critical step that allows us to utilize diffusion priors at resolutions as high as $512 \\times 512$ . As 3D meshes are amenable to fast graphics renderers that can render high-resolution images in real-time, we leverage an efficient differentiable rasterizer [9, 26] and make use of camera close-ups to recover high-frequency details in geometry and texture. As a result, our approach", + "bbox": [ + 496, + 780, + 893, + 887 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "301", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "produces high-fidelity 3D content (see Fig. 1) that can conveniently be imported and visualized in standard graphics software and does so at $2 \\times$ the speed of DreamFusion. Furthermore, we showcase various creative controls over the 3D synthesis process by leveraging the advancements developed for text-to-image editing applications [2, 35]. Our approach, dubbed Magic3D, endows users with unprecedented control in crafting their desired 3D objects with text prompts and reference images, bringing this technology one step closer to democratizing 3D content creation.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our work makes the following contributions:", + "bbox": [ + 96, + 243, + 470, + 257 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Magic3D, a framework for high-quality 3D content synthesis using text prompts by improving several major design choices made in DreamFusion. It consists of a coarse-to-fine strategy that leverages both low- and high-resolution diffusion priors for learning the 3D representation of the target content. Magic3D, which synthesizes 3D content with an $8 \\times$ higher resolution supervision, is also $2 \\times$ faster than DreamFusion. 3D content synthesized by our approach is significantly preferable by users (61.7%).", + "- We extend various image editing techniques developed for text-to-image models to 3D object editing and show their applications in the proposed framework." + ], + "bbox": [ + 76, + 268, + 472, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 470, + 218, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image generation. We have witnessed significant progress in text-to-image generation with diffusion models in recent years. With improvements in modeling and data curation, diffusion models can compose complex semantic concepts from text descriptions (nouns, adjectives, artistic styles, etc.) to generate high-quality images of objects and scenes [2, 33, 34, 36]. Sampling images from diffusion models is time consuming. To generate high-resolution images, these models either utilize a cascade of super-resolution models [2, 36] or sample from a lower-resolution latent space and decode latents into high-resolution images [34]. Despite the advances in high-resolution image generation, using language to describe and control 3D properties (e.g. camera viewpoints) while maintaining coherency in 3D remains an open, challenging problem.", + "bbox": [ + 75, + 502, + 472, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D generative models. There is a large body of work on 3D generative modeling, exploring different types of 3D representations such as 3D voxel grids [7, 12, 20, 40, 45], pointclouds [1, 21, 24, 46, 47, 49], meshes [9, 48], implicit [6, 22], or octree [15] representations. Most of these approaches rely on training data in the form of 3D assets, which are hard to acquire at scale. Inspired by the success of neural volume rendering [23], recent works started investing in 3D-aware image synthesis [4, 5, 10, 11, 27, 29, 30, 38], which has the advantage of learning 3D generative models directly from images — a more widely accessible resource. However, vol-", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ume rendering networks are typically slow to query, leading to a trade-off between long training time [5, 29] and lack of multi-view consistency [10]. EG3D [4] partially mitigates this problem by utilizing a dual discriminator. While obtaining promising results, these works remain limited to modeling objects within a single object category, such as cars, chairs, or human faces, thus lacking scalability and the creative control desired for 3D content creation. In our paper, we focus on text-to-3D synthesis, aiming to generate a 3D renderable representation of a scene based on a text prompt.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-3D generation. With the recent success in text-to-image generative modeling in recent years, text-to-3D generation has also gained a surge of interest from the learning community. Earlier works such as CLIP-forge [37] synthesize objects by learning a normalizing flow model to sample shape embeddings from textual input. However, it requires 3D assets in voxel representations during training, making it challenging to scale with data. DreamField [16] and CLIP-mesh [17] mitigate the training data issue by relying on a pre-trained image-text model [32] to optimize the underlying 3D representations (NeRFs and meshes), such that all 2D renderings reach high text-image alignment scores. While these approaches avoid the requirement of expensive 3D training data and mostly rely on pre-trained large-scale image-text models, they tend to produce less realistic 2D renderings.", + "bbox": [ + 496, + 247, + 895, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, DreamFusion [31] showcased impressive capability in text-to-3D synthesis by utilizing a powerful pretrained text-to-image diffusion model [36] as a strong image prior. We build upon this work and improve over several design choices to bring significantly higher-fidelity 3D models into hands of users with a much reduced generation time.", + "bbox": [ + 496, + 474, + 893, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background: DreamFusion", + "text_level": 1, + "bbox": [ + 498, + 575, + 754, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DreamFusion [31] achieves text-to-3D generation with two key components: a neural scene representation which we refer to as the scene model, and a pre-trained text-to-image diffusion-based generative model. The scene model is a parametric function $x = g(\\theta)$ , which can produce an image $x$ at the desired camera pose. Here, $g$ is a volumetric renderer of choice, and $\\theta$ is a coordinate-based MLP representing a 3D volume. The diffusion model $\\phi$ comes with a learned denoising function $\\epsilon_{\\phi}(x_t; y, t)$ that predicts the sampled noise $\\epsilon$ given the noisy image $x_t$ , noise level $t$ , and text embedding $y$ . It provides the gradient direction to update $\\theta$ such that all rendered images are pushed to the high probability density regions conditioned on the text embedding under the diffusion prior. Specifically, DreamFusion introduces Score Distillation Sampling (SDS), which computes the gradient:", + "bbox": [ + 496, + 599, + 893, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , g (\\theta)) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(x _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 830, + 893, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $w(t)$ is a weighting function. We view the scene model $g$ and the diffusion model $\\phi$ as modular components of the", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "302", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/993ed516c2a474d9377e1c231fe571903233398327ee30a97216a721d5b9ac10.jpg", + "image_caption": [ + "Figure 2. Overview of Magic3D. We generate high-resolution 3D content from an input text prompt in a coarse-to-fine manner. In the first stage, we utilize a low-resolution diffusion prior and optimize neural field representations (color, density, and normal fields) to obtain the coarse model. We further differentiably extract textured 3D mesh from the density and color fields of the coarse model. Then we fine-tune it using a high-resolution latent diffusion model. After optimization, our model generates high-quality 3D meshes with detailed textures." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 893, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "framework, amenable to choice. In practice, the denoising function $\\epsilon_{\\phi}$ is often replaced with another function $\\tilde{\\epsilon}_{\\phi}$ that uses classifier-free guidance [14], which allows one to carefully weigh the strength of the text conditioning (see Sec. 6). DreamFusion relies on large classifier-free guidance weights to obtain results with better quality.", + "bbox": [ + 75, + 375, + 470, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DreamFusion adopts a variant of Mip-NeRF 360 [3] with an explicit shading model for the scene model and Imagegen [36] as the diffusion model. These choices result in two key limitations. First, high-resolution geometry or textures cannot be obtained since the diffusion model only operates on $64 \\times 64$ images. Second, the utility of a large global MLP for volume rendering is both computationally expensive as well as memory intensive, making this approach scale poorly with the increasing resolution of images.", + "bbox": [ + 75, + 464, + 468, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. High-Resolution 3D Generation", + "text_level": 1, + "bbox": [ + 76, + 614, + 366, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Magic3D is a two-stage coarse-to-fine framework that uses efficient scene models that enable high-resolution text-to-3D synthesis (Fig. 2). We describe our method and key differences from DreamFusion [31] in this section.", + "bbox": [ + 75, + 640, + 470, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Coarse-to-fine Diffusion Priors", + "text_level": 1, + "bbox": [ + 76, + 710, + 349, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Magic3D uses two different diffusion priors in a coarse-to-fine fashion to generate high-resolution geometry and textures. In the first stage, we use the base diffusion model described in eDiff-I [2], which is similar to the base diffusion model of Imagen [36] used in DreamFusion. This diffusion prior is used to compute gradients of the scene model via a loss defined on rendered images at a low resolution $64 \\times 64$ . In the second stage, we use the latent diffusion model (LDM) [34] that allows backpropagating gradients into rendered images at a high resolution $512 \\times 512$ ; in practice, we choose to use the publicly available Stable Diffusion", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model [34]. Despite generating high-resolution images, the computation of LDM is manageable because the diffusion prior acts on the latent $z_{t}$ with resolution $64 \\times 64$ :", + "bbox": [ + 496, + 375, + 890, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , g (\\theta)) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(z _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial z}{\\partial x} \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 428, + 890, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The increase in computation time mainly comes from computing $\\frac{\\partial x}{\\partial \\theta}$ (the gradient of the high-resolution rendered image) and $\\frac{\\partial z}{\\partial x}$ (the gradient of the encoder in LDM).", + "bbox": [ + 496, + 470, + 893, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Scene Models", + "text_level": 1, + "bbox": [ + 500, + 525, + 640, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We cater two different 3D scene representations to the two different diffusion priors at coarse and fine resolutions to accommodate the increased resolution of rendered images for the input of high-resolution priors, discussed as follows.", + "bbox": [ + 496, + 547, + 893, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural fields as coarse scene models. The initial coarse stage of the optimization requires finding the geometry and textures from scratch. This can be challenging as we need to accommodate complex topological changes in the 3D geometry and depth ambiguities from the 2D supervision signals. In DreamFusion [31], the scene model is a neural field (a coordinate-based MLP) based on Mip-NeRF 360 [3] that predicts albedo and density. This is a suitable choice as neural fields can handle topological changes in a smooth, continuous fashion. However, Mip-NeRF 360 [3] is computationally expensive as it is based on a large global coordinate-based MLP. As volume rendering requires dense samples along a ray to accurately render high-frequency geometry and shading, the cost of having to evaluate a large neural network at every sample point quickly stacks up.", + "bbox": [ + 496, + 613, + 893, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For this reason, we opt to use the hash grid encoding from Instant NGP [25], which allows us to represent high-frequency details at a much lower computational cost. We use the hash grid with two single-layer neural networks,", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "303", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "one predicting albedo and density and the other predicting normals. We additionally maintain a spatial data structure that encodes scene occupancy and utilizes empty space skipping [19, 43]. Specifically, we use the density-based voxel pruning approach from Instant NGP [25] with an octree-based ray sampling and rendering algorithm [44]. With these design choices, we drastically accelerate the optimization of coarse scene models while maintaining quality.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textured meshes as fine scene models. In our fine stage of optimization, we need to be able to accommodate very high-resolution rendered images to fine-tune our scene model with high-resolution diffusion priors. Using the same scene representation (the neural field) from the initial coarse stage of optimization could be a natural choice since the weights of the model can directly carry over. Although this strategy can work to some extent (Figs. 4 and 5), it struggles to render very high-resolution (e.g., $512 \\times 512$ ) images within reasonable memory constraints and computation budgets.", + "bbox": [ + 75, + 217, + 472, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To resolve this issue, we use textured 3D meshes as the scene representation for the fine stage of optimization. In contrast to volume rendering for neural fields, rendering textured meshes with differentiable rasterization can be performed efficiently at very high resolutions, making meshes a suitable choice for our high-resolution optimization stage. Using the neural field from the coarse stage as the initialization for the mesh geometry, we can also sidestep the difficulty of learning large topological changes in meshes.", + "bbox": [ + 75, + 369, + 472, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, we represent the 3D shape using a deformable tetrahedral grid $(V_{T}, T)$ , where $V_{T}$ is the vertices in the grid $T$ [8, 39]. Each vertex $\\mathbf{v}_{i} \\in V_{T} \\subset \\mathbb{R}^{3}$ contains a signed distance field (SDF) value $s_{i} \\in \\mathbb{R}$ and a deformation $\\Delta \\mathbf{v}_{i} \\in \\mathbb{R}^{3}$ of the vertex from its initial canonical coordinate. Then, we extract a surface mesh from the SDF using a differentiable marching tetrahedra algorithm [39]. For textures, we use the neural color field as a volumetric texture representation.", + "bbox": [ + 75, + 507, + 472, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Coarse-to-fine Optimization", + "text_level": 1, + "bbox": [ + 76, + 642, + 328, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We describe our coarse-to-fine optimization procedure, which first operates on a coarse neural field representation and subsequently a high-resolution textured mesh.", + "bbox": [ + 75, + 666, + 470, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neural field optimization. Similarly to Instant NGP [25], we initialize an occupancy grid of resolution $256^3$ with values to 20 to encourage shapes to grow in the early stages of optimization. We update the grid every 10 iterations and generate an octree for empty space skipping. We decay the occupancy grid by 0.6 in every update and follow Instant NGP with the same update and thresholding parameters.", + "bbox": [ + 75, + 717, + 470, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of estimating normals from density differences, we use an MLP to predict the normals. Note that this does not violate geometric properties since volume rendering is used instead of surface rendering; as such, the orientation of particles at continuous positions need not be oriented", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to the level set surface. This helps us significantly reduce the computational cost of optimizing the coarse model by avoiding the use of finite differencing. Accurate normals can be obtained in the fine stage of optimization when we use a true surface rendering model.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to DreamFusion, we also model the background using an environment map MLP, which predicts RGB colors as a function of ray directions. Since our sparse representation model does not support scene reparametrization as in Mip-NeRF 360 [3], the optimization has a tendency to \"cheat\" by learning the essence of the object using the background environment map. As such, we use a tiny MLP for the environment map (hidden dimension size of 16) and weigh down the learning rate by $10\\times$ to allow the model to focus more on the neural field geometry.", + "bbox": [ + 496, + 167, + 893, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh optimization. To optimize a mesh from the neural field initialization, we convert the (coarse) density field to an SDF by subtracting it with a non-zero constant, yielding the initial $s_i$ . We additionally initialize the volume texture field directly with the color field optimized from the coarse stage.", + "bbox": [ + 496, + 321, + 893, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During optimization, we render the extracted surface mesh into high-resolution images using a differentiable rasterizer [18,26]. We optimize both $s_i$ and $\\Delta \\mathbf{v}_i$ for each vertex $\\mathbf{v}_i$ via backpropagation using the high-resolution SDS gradient (Eq. 2). When rendering the mesh to an image, we also track the 3D coordinates of each corresponding pixel projection, which would be used to query colors in the corresponding texture field for joint optimization.", + "bbox": [ + 496, + 398, + 893, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When rendering the mesh, we increase the focal length to zoom in on object details, which is a critical step towards recovering high-frequency details. We keep the same pre-trained environment map from the coarse stage of optimization and composite the rendered background with the rendered foreground object using differentiable antialIASing [18]. To encourage the smoothness of the surface, we further regularize the angular differences between adjacent faces on the mesh. This allows us to obtain well-behaved geometry even under supervision signals with high variance, such as the SDS gradient $\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{SDS}}$", + "bbox": [ + 496, + 518, + 895, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 696, + 633, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We focus on comparing our method with DreamFusion [31] on 397 text prompts taken from the website of DreamFusion1. We train Magic3D on all of the text prompts and compare them with the results provided on the website.", + "bbox": [ + 496, + 722, + 893, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Speed evaluation. Unless otherwise noted, the coarse stage is trained for 5000 iterations with 1024 samples along the ray (subsequently filtered by the sparse octree) with a batch size of 32, with a total runtime of around 15 minutes (upwards of 8 iterations / second, variable due to differences in sparsity). The fine stage is trained for 3000 iterations with a batch", + "bbox": [ + 496, + 787, + 893, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "ComparisonPreferenceMagic3D vs. DreamFusion [31]• More realistic58.3%• More detailed66.0%• More realistic & detailed61.7%Magic3D vs. Magic3D (coarse only)87.7%", + "bbox": [ + 93, + 607, + 450, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "size of 32 with a total runtime of 25 minutes (2 iterations / second). Both runtimes combined are 40 minutes. All runtimes were measured on 8 NVIDIA A100 GPUs.", + "bbox": [ + 75, + 742, + 470, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons. We provide qualitative examples in Fig. 3. Qualitatively, our models achieve much higher 3D quality in terms of both geometry and texture. Notice that our model can generate candies on ice cream cones, highly detailed sushi-like cars, vivid strawberries, and birds. We also note that our resulting 3D models can be directly imported and visualized in standard graphics software.", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a baby bunny sitting on top of a stack of pancakes†", + "bbox": [ + 553, + 516, + 834, + 529 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43eb0c7a0ff9d607f3f4eaefc8982db71414077c938d07a21007740ab1d9e9c7.jpg", + "image_caption": [ + "(a) Single-stage model" + ], + "image_footnote": [], + "bbox": [ + 506, + 529, + 870, + 606 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c0414b0340731f072ebd6195646369ba0ae237f5b9e657f726f08da5c1a48b6.jpg", + "image_caption": [ + "(b) Coarse-to-fine model", + "Figure 4. Single-stage (top) vs. coarse-to-fine models (bottom). Both use NeRF for the scene model. The left two columns use $64 \\times 64$ rendering resolution during optimization while the right two use $256 \\times 256$ . Compared to our coarse-to-fine approach, the single-stage method can generate details but with worse shapes." + ], + "image_footnote": [], + "bbox": [ + 509, + 622, + 885, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User studies. We conduct user studies to evaluate different methods based on user preferences on Amazon MTurk. We show users two videos side by side rendered from a canonical view by two different algorithms using the same text prompt. We ask the users to select the one that is more realistic", + "bbox": [ + 496, + 824, + 895, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "305", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2128d779f88b92785de3b96998a0cb96af27762b568b0b9fe1c5495676770b0b.jpg", + "image_caption": [ + "Figure 5. Ablation on the fine-tuning stage. For each text prompt, we compare coarse and fine models with mesh and NeRF representations. Mesh fine-tuning significantly improve the visual quality of generated 3D assets, providing more photo-realistic details on the 3D shapes." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and detailed. Each prompt is evaluated by 3 different users, resulting in 1191 pairwise comparisons. As shown in Table 1, users favor 3D models generated by Magic3D, with $61.7\\%$ of the users considering our results with higher quality.", + "bbox": [ + 75, + 343, + 473, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can single-stage optimization work with LDM prior? We ablate scene models optimized with high-resolution LDM prior in a single-stage optimization setup. We find that 3D meshes as the scene model fail to generate high-quality results if optimized from scratch. This leaves our memory-efficient sparse 3D representation as the ideal candidate for the scene model. However, rendering $512 \\times 512$ images is still too memory intensive to fit into modern GPUs. Therefore, we render lower-resolution images from the scene model and upsample them to $512 \\times 512$ as input to the LDM. We find it generates objects with worse shapes. Fig. 4 shows two examples with scene rendering resolution $64 \\times 64$ and $256 \\times 256$ respectively (top row). While it generates furry details, the shape is worse than the coarse model.", + "bbox": [ + 75, + 422, + 473, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can we use NeRF for the fine model? Yes. While optimizing a NeRF from scratch does not work well, we can follow the coarse-to-fine framework but replace the second-stage scene model with a NeRF. In the bottom right of Fig. 4, we show the result of a fine NeRF model initialized with the coarse model on its left and fine-tuned with $256 \\times 256$ rendered images. The two-stage approach retains good geometry in the initial model and adds more details, showing superior quality to its one-stage counterpart.", + "bbox": [ + 75, + 654, + 472, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Coarse models vs. fine models. Fig. 5 provides more visual results contrasting coarse and fine models. We try both NeRF and mesh for scene models and fine-tune them from the same coarse model above. We see significant quality improvements on both NeRF and mesh models, suggesting our coarse-to-fine approach works for general scene models.", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/240aac95d822c1e949b296249d85ec84fbd191643e437d3b29c80490eb1a5d38.jpg", + "image_caption": [ + "Figure 6. Magic3D with DreamBooth-based personalization. Given an input image of a particular instance, we fine-tune the diffusion models with DreamBooth and optimize the 3D models with the given prompts. The identity is well preserved in the generated 3D models. Image source (input images): Unsplash." + ], + "image_footnote": [], + "bbox": [ + 503, + 340, + 890, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Controllable 3D Generation", + "text_level": 1, + "bbox": [ + 500, + 681, + 759, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As certain styles and concepts are difficult to express in words but easy with images, it is desirable to have a mechanism to influence the text-to-3D model generation with images. We explore different image conditioning techniques as well as a prompt-based editing approach to provide users more control over the 3D generation outputs.", + "bbox": [ + 496, + 710, + 893, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Personalized text-to-3D. DreamBooth [35] described a method to personalize text-to-image diffusion models by fine-tuning a pre-trained model on several images of a subject. The fine-tuned model can learn to tie the subject to a unique identifier string, denoted as $[\\mathrm{V}]$ , and generate images of the subject when $[\\mathrm{V}]$ is included in the text prompt. In", + "bbox": [ + 496, + 809, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "306", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/397ee7678d79380925b096100a42da2efde6ef375dab86f84cde6612af37bf61.jpg", + "image_caption": [ + "low-resolution before edited" + ], + "image_footnote": [], + "bbox": [ + 80, + 112, + 153, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75f10d2c9750fb9b437ace0f89d2872fc7702a5eed360877b1ac99bb2d3dae23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 89, + 385, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a2ef0985bfc4aa89b6bd728b0bff926d8c6c7c9964e2007dcf044c8d9c128bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 90, + 648, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1be9748663d9b6cdf89903c2cf2a1245fa77ff9eddb78ae5021f3c41cbca3e7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 98, + 874, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/015df1f9d4199c2503bd74bce5830ec1a647405ae748aedf2899d50b022abc0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 202, + 153, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21a0dcb00660f9a0c9662d005bc7cc172cd9c1dac0732d077d42b09b1c30338d.jpg", + "image_caption": [ + "Figure 7. Magic3D with prompt-based editing. Given a coarse model (first column) generated with a base prompt, we replace the underscored text with new text and fine-tune the NeRF to get a high-resolution NeRF model with LDM. We further fine-tune the high-resolution mesh with the NeRF model. Such a prompt-based editing method gives artists better control over the 3D generation output." + ], + "image_footnote": [], + "bbox": [ + 173, + 202, + 392, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7809c54ad73d694302db560d32c30b0831b3652c4f3160efe38b9322efc230a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 205, + 640, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca1779110d89b9f94b2208ce0535fdbacdb5811d5cd1f4a10885b3392f0c3d10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 215, + 867, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the context of text-to-3D generation, we would like to generate a 3D model of a subject. This can be achieved by first fine-tuning our diffusion prior models with the DreamBooth approach, and then using the fine-tuned diffusion priors with the [V] identifier as part of the conditioning text prompt to provide the learning signal when optimizing the 3D model.", + "bbox": [ + 75, + 386, + 468, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the applicability of DreamBooth in our framework, we collect 11 images of one cat and 4 images of one dog. We fine-tune eDiff-I [2] and LDM [34], binding the text identifier [V] to the given subject. Then, we optimize the 3D model with [V] in the text prompts. We use a batch size of 1 for all fine-tuning. For eDiff-I, we use the Adam optimizer with learning rate $1 \\times 10^{-5}$ for 1,500 iterations; for LDM, we fine-tune with learning rate $1 \\times 10^{-6}$ for 800 iterations. Fig. 6 shows our personalized text-to-3D results: we are able to successfully modify the 3D models preserving the subjects in the given input images.", + "bbox": [ + 75, + 482, + 470, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt-based editing through fine-tuning. Another way to control the generated 3D content is by fine-tuning a learned coarse model with a new prompt. Our prompt-based editing includes three stages. (a) We train a coarse model with a base prompt. (b) We modify the base prompt and fine-tune the coarse model with the LDM. This stage provides a well initialized NeRF model for the next step. Directly applying mesh optimization on a new prompt would generate highly detailed textures but could deform geometry only slightly. (c) We optimize the mesh with the modified text prompt. Our prompt-based editing can modify the texture of the shape or transform the geometry and texture according to the text. The resulting scene models preserve the layer-out and overall structure. Such an editing capability makes the 3D content creation with Magic3D more controllable. In Fig. 7, we show two coarse NeRF models trained with the", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "base prompt for the \"bunny\" and \"squirrel\". We modify the base prompt, fine-tune the NeRF model in high resolution and optimize the mesh. Results show that we can tune the scene model according to the prompt, e.g. changing the \"baby bunny\" to \"stained glass bunny\" or \"metal bunny\" results in similar geometry but with a different texture.", + "bbox": [ + 496, + 385, + 893, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 515, + 617, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Magic3D, a fast and high-quality text-to-3D generation framework. We benefit from both efficient scene models and high-resolution diffusion priors in a coarse-to-fine approach. In particular, the 3D mesh models scale nicely with image resolution and enjoy the benefits of higher resolution supervision brought by the latent diffusion model without sacrificing its speed. It takes 40 minutes from a text prompt to a high-quality 3D mesh model ready to be used in graphic engines. With extensive user studies and qualitative comparisons, we show that Magic3D is more preferable $(61.7\\%)$ by the raters compared to DreamFusion, while enjoying a $2\\times$ speed-up. Lastly, we propose a set of tools for better controlling style and content in 3D generation. We hope with Magic3D, we can democratize 3D synthesis and open up everyone's creativity in 3D content creation.", + "bbox": [ + 496, + 549, + 893, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We would like to thank Frank Shen, Yogesh Balaji, Seungjun Nah, James Lucas, David Luebke, Clement Fuji-Tsang, Charles Loop, Qinsheng Zhang, Zan Gojcic, and Jonathan Tremblay for helpful discussions and paper proofreading. We would also like to thank Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall for providing additional implementation details in DreamFusion.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "307", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 3", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 3, 4, 8", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 4, 5", + "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 3", + "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3", + "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3", + "[7] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In 2017 International Conference on 3D Vision (3DV), pages 402-411. IEEE, 2017. 3", + "[8] Jun Gao, Wenzheng Chen, Tommy Xiang, Clement Fuji Tsang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. In Advances In Neural Information Processing Systems, 2020. 5", + "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances in Neural Information Processing Systems, 2022. 1, 2, 3", + "[10] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 3", + "[11] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds. In ICCV, 2021. 3", + "[12] Philipp Henzler, Niloy J. Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3" + ], + "bbox": [ + 78, + 116, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 4", + "[15] Moritz Ibing, Gregor Kobsik, and Leif Kobbelt. Octree transformer: Autoregressive 3d shape generation on hierarchically structured sequences. arXiv preprint arXiv:2111.12480, 2021. 3", + "[16] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 3", + "[17] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. ACM Transactions on Graphics (TOG), Proc. SIGGRAPH Asia, 2022. 3", + "[18] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5", + "[19] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 5", + "[20] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674, 2020. 3", + "[21] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 3", + "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 3", + "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3", + "[24] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. ACM Transactions on Graphics (TOG), Siggraph Asia 2019, 38(6):Article 242, 2019. 3", + "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 5", + "[26] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, pages 8280-8290, 2022. 2, 5", + "[27] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "308", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. 3", + "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[29] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3", + "[30] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 3", + "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3, 4, 5, 6", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 3", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 3, 4, 8", + "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3, 7", + "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 3, 4", + "[37] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 3", + "[38] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3", + "[39] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In Conference on Robot Learning, pages 87-96. PMLR, 2017. 3", + "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1", + "[42] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 1", + "[43] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 5", + "[44] Towaki Takikawa, Or Perel, Clement Fuji Tsang, Charles Loop, Joey Litalien, Jonathan Tremblay, Sanja Fidler, and Maria Shugrina. Kaolin wisp: A pytorch library and engine for neural fields research. https://github.com/NVIDIAGameWorks/kaolin-wisp, 2022.5", + "[45] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3", + "[46] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 3", + "[47] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3", + "[48] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yinan Zhang, Antonio Torralba, and Sanja Fidler. Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In International Conference on Learning Representations, 2021. 3", + "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826-5835, 2021. 3" + ], + "bbox": [ + 501, + 92, + 893, + 726 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "309", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_model.json b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_model.json new file mode 100644 index 0000000000000000000000000000000000000000..098440fa391f7806b788f7dbcb9486087b01f0bf --- /dev/null +++ b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.77, + 0.152 + ], + "angle": 0, + "content": "Magic3D: High-Resolution Text-to-3D Content Creation" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.181, + 0.827, + 0.218 + ], + "angle": 0, + "content": "Chen-Hsuan Lin* Jun Gao* Luming Tang* Towaki Takikawa* Xiaohui Zeng* Xun Huang Karsten Kreis Sanja Fidler† Ming-Yu Liu† Tsung-Yi Lin" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.224, + 0.571, + 0.242 + ], + "angle": 0, + "content": "NVIDIA Corporation" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.246, + 0.665, + 0.261 + ], + "angle": 0, + "content": "https://research.nvidia.com/labs/dir/magic3d" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.297, + 0.314, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.328, + 0.474, + 0.661 + ], + "angle": 0, + "content": "DreamFusion [31] has recently demonstrated the utility of a pre-trained text-to-image diffusion model to optimize Neural Radiance Fields (NeRF) [23], achieving remarkable text-to-3D synthesis results. However, the method has two inherent limitations: (a) extremely slow optimization of NeRF and (b) low-resolution image space supervision on NeRF, leading to low-quality 3D models with a long processing time. In this paper, we address these limitations by utilizing a two-stage optimization framework. First, we obtain a coarse model using a low-resolution diffusion prior and accelerate with a sparse 3D hash grid structure. Using the coarse representation as the initialization, we further optimize a textured 3D mesh model with an efficient differentiable renderer interacting with a high-resolution latent diffusion model. Our method, dubbed Magic3D, can create high quality 3D mesh models in 40 minutes, which is \\(2 \\times\\) faster than DreamFusion (reportedly taking 1.5 hours on average), while also achieving higher resolution. User studies show \\(61.7\\%\\) raters to prefer our approach over DreamFusion. Together with the image-conditioned generation capabilities, we provide users with new ways to control 3D synthesis, opening up new avenues to various creative applications." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.688, + 0.21, + 0.703 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.713, + 0.471, + 0.879 + ], + "angle": 0, + "content": "3D digital content has been in high demand for a variety of applications, including gaming, entertainment, architecture, and robotics simulation. It is slowly finding its way into virtually every possible domain: retail, online conferencing, virtual social presence, education, etc. However, creating professional 3D content is not for anyone — it requires immense artistic and aesthetic training with 3D modeling expertise. Developing these skill sets takes a significant amount of time and effort. Augmenting 3D content creation with natural language could considerably help democratize 3D content creation for novices and turbocharge expert artists." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.298, + 0.895, + 0.57 + ], + "angle": 0, + "content": "Image content creation from text prompts [2, 28, 33, 36] has seen significant progress with the advances of diffusion models [13, 41, 42] for generative modeling of images. The key enablers are large-scale datasets comprising billions of samples (images with text) scrapped from the Internet and massive amounts of compute. In contrast, 3D content generation has progressed at a much slower pace. Existing 3D object generation models [4, 9, 47] are mostly categorical. A trained model can only be used to synthesize objects for a single class, with early signs of scaling to multiple classes shown recently by Zeng et al. [47]. Therefore, what a user can do with these models is extremely limited and not yet ready for artistic creation. This limitation is largely due to the lack of diverse large-scale 3D datasets — compared to image and video content, 3D content is much less accessible on the Internet. This naturally raises the question of whether 3D generation capability can be achieved by leveraging powerful text-to-image generative models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.576, + 0.896, + 0.85 + ], + "angle": 0, + "content": "Recently, DreamFusion [31] demonstrated its remarkable ability for text-conditioned 3D content generation by utilizing a pre-trained text-to-image diffusion model [36] that generates images as a strong image prior. The diffusion model acts as a critic to optimize the underlying 3D representation. The optimization process ensures that rendered images from a 3D model, represented by Neural Radiance Fields (NeRF) [23], match the distribution of photorealistic images across different viewpoints, given the input text prompt. Since the supervision signal in DreamFusion operates on very low-resolution images \\((64 \\times 64)\\), DreamFusion cannot synthesize high-frequency 3D geometric and texture details. Due to the use of inefficient MLP architectures for the NeRF representation, practical high-resolution synthesis may not even be possible as the required memory footprint and the computation budget grows quickly with the resolution. Even at a resolution of \\(64 \\times 64\\), optimization times are in hours (1.5 hours per prompt on average using TPUv4)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.897, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a method that can synthesize highly detailed 3D models from text prompts within a reduced computation time. Specifically, we propose a coarse-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.887, + 0.212, + 0.9 + ], + "angle": 0, + "content": "\\(^{\\ast \\dagger}\\): equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "300" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.134, + 0.314, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.298, + 0.262, + 0.326 + ], + "angle": 0, + "content": "a silver platter piled high with fruits" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.092, + 0.432, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.318, + 0.488, + 0.343 + ], + "angle": 0, + "content": "michelangelo style statue of an astronaut" + }, + { + "type": "image", + "bbox": [ + 0.469, + 0.099, + 0.623, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.318, + 0.658, + 0.346 + ], + "angle": 0, + "content": "a stuffed grey rabbit holding a pretend carrot" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.155, + 0.895, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.31, + 0.872, + 0.323 + ], + "angle": 0, + "content": "an iguana holding a balloon" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.33, + 0.208, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.512, + 0.238, + 0.54 + ], + "angle": 0, + "content": "a beautiful dress made out of garbage bags" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.349, + 0.396, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.513, + 0.379, + 0.54 + ], + "angle": 0, + "content": "an imperial state crown of england" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.365, + 0.625, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.511, + 0.592, + 0.54 + ], + "angle": 0, + "content": "a blue poison-dart frog sitting on a water lily" + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.355, + 0.891, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.524, + 0.868, + 0.537 + ], + "angle": 0, + "content": "neuschwanstein castle, aerial view" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.557, + 0.243, + 0.585 + ], + "angle": 0, + "content": "Low resolution bunny before editing" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.595, + 0.185, + 0.661 + ], + "angle": 0, + "content": "a baby bunny sitting on top of a stack of pancakes" + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.596, + 0.244, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.548, + 0.321, + 0.626 + ], + "angle": 0, + "content": "a metal bunny sitting on top of a stack of broccoli" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.546, + 0.451, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.472, + 0.548, + 0.535, + 0.638 + ], + "angle": 0, + "content": "a metalbunnysitting ontop of asstack ofchocolatecookie" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.546, + 0.66, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.549, + 0.75, + 0.627 + ], + "angle": 0, + "content": "a sphinx sitting on top of a stack of chocolate cookie" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.545, + 0.862, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.685, + 0.892, + 0.755 + ], + "angle": 0, + "content": "Figure 1. Results and applications of Magic3D. Top: high-resolution text-to-3D generation. Magic3D can generate high-quality and high-resolution 3D models from text prompts. Bottom: high-resolution prompt-based editing. Magic3D can edit 3D models by fine-tuning with the diffusion prior using a different prompt. Taking the low-resolution 3D model as the input (left), Magic3D can modify different parts of the 3D model corresponding to different input text prompts. Together with various creative controls on the generated 3D models, Magic3D is a convenient tool for augmenting 3D content creation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.888 + ], + "angle": 0, + "content": "to-fine optimization approach that uses multiple diffusion priors at different resolutions to optimize the 3D representation, enabling the generation of both view-consistent geometry as well as high-resolution details. In the first stage, we optimize a coarse neural field representation akin to Dream-Fusion, but with a memory- and compute-efficient scene representation based on a hash grid [25]. In the second stage," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.888 + ], + "angle": 0, + "content": "we switch to optimizing mesh representations, a critical step that allows us to utilize diffusion priors at resolutions as high as \\(512 \\times 512\\). As 3D meshes are amenable to fast graphics renderers that can render high-resolution images in real-time, we leverage an efficient differentiable rasterizer [9, 26] and make use of camera close-ups to recover high-frequency details in geometry and texture. As a result, our approach" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "301" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "produces high-fidelity 3D content (see Fig. 1) that can conveniently be imported and visualized in standard graphics software and does so at \\(2 \\times\\) the speed of DreamFusion. Furthermore, we showcase various creative controls over the 3D synthesis process by leveraging the advancements developed for text-to-image editing applications [2, 35]. Our approach, dubbed Magic3D, endows users with unprecedented control in crafting their desired 3D objects with text prompts and reference images, bringing this technology one step closer to democratizing 3D content creation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.244, + 0.471, + 0.258 + ], + "angle": 0, + "content": "In summary, our work makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.27, + 0.473, + 0.406 + ], + "angle": 0, + "content": "- We propose Magic3D, a framework for high-quality 3D content synthesis using text prompts by improving several major design choices made in DreamFusion. It consists of a coarse-to-fine strategy that leverages both low- and high-resolution diffusion priors for learning the 3D representation of the target content. Magic3D, which synthesizes 3D content with an \\(8 \\times\\) higher resolution supervision, is also \\(2 \\times\\) faster than DreamFusion. 3D content synthesized by our approach is significantly preferable by users (61.7%)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.412, + 0.471, + 0.457 + ], + "angle": 0, + "content": "- We extend various image editing techniques developed for text-to-image models to 3D object editing and show their applications in the proposed framework." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.27, + 0.473, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.472, + 0.219, + 0.487 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.473, + 0.732 + ], + "angle": 0, + "content": "Text-to-image generation. We have witnessed significant progress in text-to-image generation with diffusion models in recent years. With improvements in modeling and data curation, diffusion models can compose complex semantic concepts from text descriptions (nouns, adjectives, artistic styles, etc.) to generate high-quality images of objects and scenes [2, 33, 34, 36]. Sampling images from diffusion models is time consuming. To generate high-resolution images, these models either utilize a cascade of super-resolution models [2, 36] or sample from a lower-resolution latent space and decode latents into high-resolution images [34]. Despite the advances in high-resolution image generation, using language to describe and control 3D properties (e.g. camera viewpoints) while maintaining coherency in 3D remains an open, challenging problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "3D generative models. There is a large body of work on 3D generative modeling, exploring different types of 3D representations such as 3D voxel grids [7, 12, 20, 40, 45], pointclouds [1, 21, 24, 46, 47, 49], meshes [9, 48], implicit [6, 22], or octree [15] representations. Most of these approaches rely on training data in the form of 3D assets, which are hard to acquire at scale. Inspired by the success of neural volume rendering [23], recent works started investing in 3D-aware image synthesis [4, 5, 10, 11, 27, 29, 30, 38], which has the advantage of learning 3D generative models directly from images — a more widely accessible resource. However, vol-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "ume rendering networks are typically slow to query, leading to a trade-off between long training time [5, 29] and lack of multi-view consistency [10]. EG3D [4] partially mitigates this problem by utilizing a dual discriminator. While obtaining promising results, these works remain limited to modeling objects within a single object category, such as cars, chairs, or human faces, thus lacking scalability and the creative control desired for 3D content creation. In our paper, we focus on text-to-3D synthesis, aiming to generate a 3D renderable representation of a scene based on a text prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.897, + 0.475 + ], + "angle": 0, + "content": "Text-to-3D generation. With the recent success in text-to-image generative modeling in recent years, text-to-3D generation has also gained a surge of interest from the learning community. Earlier works such as CLIP-forge [37] synthesize objects by learning a normalizing flow model to sample shape embeddings from textual input. However, it requires 3D assets in voxel representations during training, making it challenging to scale with data. DreamField [16] and CLIP-mesh [17] mitigate the training data issue by relying on a pre-trained image-text model [32] to optimize the underlying 3D representations (NeRFs and meshes), such that all 2D renderings reach high text-image alignment scores. While these approaches avoid the requirement of expensive 3D training data and mostly rely on pre-trained large-scale image-text models, they tend to produce less realistic 2D renderings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.895, + 0.567 + ], + "angle": 0, + "content": "Recently, DreamFusion [31] showcased impressive capability in text-to-3D synthesis by utilizing a powerful pretrained text-to-image diffusion model [36] as a strong image prior. We build upon this work and improve over several design choices to bring significantly higher-fidelity 3D models into hands of users with a much reduced generation time." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.576, + 0.755, + 0.594 + ], + "angle": 0, + "content": "3. Background: DreamFusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.895, + 0.829 + ], + "angle": 0, + "content": "DreamFusion [31] achieves text-to-3D generation with two key components: a neural scene representation which we refer to as the scene model, and a pre-trained text-to-image diffusion-based generative model. The scene model is a parametric function \\( x = g(\\theta) \\), which can produce an image \\( x \\) at the desired camera pose. Here, \\( g \\) is a volumetric renderer of choice, and \\( \\theta \\) is a coordinate-based MLP representing a 3D volume. The diffusion model \\( \\phi \\) comes with a learned denoising function \\( \\epsilon_{\\phi}(x_t; y, t) \\) that predicts the sampled noise \\( \\epsilon \\) given the noisy image \\( x_t \\), noise level \\( t \\), and text embedding \\( y \\). It provides the gradient direction to update \\( \\theta \\) such that all rendered images are pushed to the high probability density regions conditioned on the text embedding under the diffusion prior. Specifically, DreamFusion introduces Score Distillation Sampling (SDS), which computes the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.832, + 0.895, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , g (\\theta)) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(x _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Here, \\( w(t) \\) is a weighting function. We view the scene model \\( g \\) and the diffusion model \\( \\phi \\) as modular components of the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "302" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.895, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.893, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Overview of Magic3D. We generate high-resolution 3D content from an input text prompt in a coarse-to-fine manner. In the first stage, we utilize a low-resolution diffusion prior and optimize neural field representations (color, density, and normal fields) to obtain the coarse model. We further differentiably extract textured 3D mesh from the density and color fields of the coarse model. Then we fine-tune it using a high-resolution latent diffusion model. After optimization, our model generates high-quality 3D meshes with detailed textures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.471, + 0.464 + ], + "angle": 0, + "content": "framework, amenable to choice. In practice, the denoising function \\(\\epsilon_{\\phi}\\) is often replaced with another function \\(\\tilde{\\epsilon}_{\\phi}\\) that uses classifier-free guidance [14], which allows one to carefully weigh the strength of the text conditioning (see Sec. 6). DreamFusion relies on large classifier-free guidance weights to obtain results with better quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.603 + ], + "angle": 0, + "content": "DreamFusion adopts a variant of Mip-NeRF 360 [3] with an explicit shading model for the scene model and Imagegen [36] as the diffusion model. These choices result in two key limitations. First, high-resolution geometry or textures cannot be obtained since the diffusion model only operates on \\(64 \\times 64\\) images. Second, the utility of a large global MLP for volume rendering is both computationally expensive as well as memory intensive, making this approach scale poorly with the increasing resolution of images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.367, + 0.632 + ], + "angle": 0, + "content": "4. High-Resolution 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.471, + 0.702 + ], + "angle": 0, + "content": "Magic3D is a two-stage coarse-to-fine framework that uses efficient scene models that enable high-resolution text-to-3D synthesis (Fig. 2). We describe our method and key differences from DreamFusion [31] in this section." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.351, + 0.726 + ], + "angle": 0, + "content": "4.1. Coarse-to-fine Diffusion Priors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Magic3D uses two different diffusion priors in a coarse-to-fine fashion to generate high-resolution geometry and textures. In the first stage, we use the base diffusion model described in eDiff-I [2], which is similar to the base diffusion model of Imagen [36] used in DreamFusion. This diffusion prior is used to compute gradients of the scene model via a loss defined on rendered images at a low resolution \\(64 \\times 64\\). In the second stage, we use the latent diffusion model (LDM) [34] that allows backpropagating gradients into rendered images at a high resolution \\(512 \\times 512\\); in practice, we choose to use the publicly available Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.42 + ], + "angle": 0, + "content": "model [34]. Despite generating high-resolution images, the computation of LDM is manageable because the diffusion prior acts on the latent \\( z_{t} \\) with resolution \\( 64 \\times 64 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , g (\\theta)) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(z _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial z}{\\partial x} \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.894, + 0.517 + ], + "angle": 0, + "content": "The increase in computation time mainly comes from computing \\(\\frac{\\partial x}{\\partial \\theta}\\) (the gradient of the high-resolution rendered image) and \\(\\frac{\\partial z}{\\partial x}\\) (the gradient of the encoder in LDM)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.526, + 0.642, + 0.54 + ], + "angle": 0, + "content": "4.2. Scene Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.894, + 0.61 + ], + "angle": 0, + "content": "We cater two different 3D scene representations to the two different diffusion priors at coarse and fine resolutions to accommodate the increased resolution of rendered images for the input of high-resolution priors, discussed as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.841 + ], + "angle": 0, + "content": "Neural fields as coarse scene models. The initial coarse stage of the optimization requires finding the geometry and textures from scratch. This can be challenging as we need to accommodate complex topological changes in the 3D geometry and depth ambiguities from the 2D supervision signals. In DreamFusion [31], the scene model is a neural field (a coordinate-based MLP) based on Mip-NeRF 360 [3] that predicts albedo and density. This is a suitable choice as neural fields can handle topological changes in a smooth, continuous fashion. However, Mip-NeRF 360 [3] is computationally expensive as it is based on a large global coordinate-based MLP. As volume rendering requires dense samples along a ray to accurately render high-frequency geometry and shading, the cost of having to evaluate a large neural network at every sample point quickly stacks up." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "For this reason, we opt to use the hash grid encoding from Instant NGP [25], which allows us to represent high-frequency details at a much lower computational cost. We use the hash grid with two single-layer neural networks," + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "303" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "one predicting albedo and density and the other predicting normals. We additionally maintain a spatial data structure that encodes scene occupancy and utilizes empty space skipping [19, 43]. Specifically, we use the density-based voxel pruning approach from Instant NGP [25] with an octree-based ray sampling and rendering algorithm [44]. With these design choices, we drastically accelerate the optimization of coarse scene models while maintaining quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.473, + 0.37 + ], + "angle": 0, + "content": "Textured meshes as fine scene models. In our fine stage of optimization, we need to be able to accommodate very high-resolution rendered images to fine-tune our scene model with high-resolution diffusion priors. Using the same scene representation (the neural field) from the initial coarse stage of optimization could be a natural choice since the weights of the model can directly carry over. Although this strategy can work to some extent (Figs. 4 and 5), it struggles to render very high-resolution (e.g., \\(512 \\times 512\\)) images within reasonable memory constraints and computation budgets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.473, + 0.507 + ], + "angle": 0, + "content": "To resolve this issue, we use textured 3D meshes as the scene representation for the fine stage of optimization. In contrast to volume rendering for neural fields, rendering textured meshes with differentiable rasterization can be performed efficiently at very high resolutions, making meshes a suitable choice for our high-resolution optimization stage. Using the neural field from the coarse stage as the initialization for the mesh geometry, we can also sidestep the difficulty of learning large topological changes in meshes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.473, + 0.63 + ], + "angle": 0, + "content": "Formally, we represent the 3D shape using a deformable tetrahedral grid \\((V_{T}, T)\\), where \\(V_{T}\\) is the vertices in the grid \\(T\\) [8, 39]. Each vertex \\(\\mathbf{v}_{i} \\in V_{T} \\subset \\mathbb{R}^{3}\\) contains a signed distance field (SDF) value \\(s_{i} \\in \\mathbb{R}\\) and a deformation \\(\\Delta \\mathbf{v}_{i} \\in \\mathbb{R}^{3}\\) of the vertex from its initial canonical coordinate. Then, we extract a surface mesh from the SDF using a differentiable marching tetrahedra algorithm [39]. For textures, we use the neural color field as a volumetric texture representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.643, + 0.33, + 0.659 + ], + "angle": 0, + "content": "4.3. Coarse-to-fine Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.472, + 0.713 + ], + "angle": 0, + "content": "We describe our coarse-to-fine optimization procedure, which first operates on a coarse neural field representation and subsequently a high-resolution textured mesh." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.472, + 0.825 + ], + "angle": 0, + "content": "Neural field optimization. Similarly to Instant NGP [25], we initialize an occupancy grid of resolution \\(256^3\\) with values to 20 to encourage shapes to grow in the early stages of optimization. We update the grid every 10 iterations and generate an octree for empty space skipping. We decay the occupancy grid by 0.6 in every update and follow Instant NGP with the same update and thresholding parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Instead of estimating normals from density differences, we use an MLP to predict the normals. Note that this does not violate geometric properties since volume rendering is used instead of surface rendering; as such, the orientation of particles at continuous positions need not be oriented" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.894, + 0.168 + ], + "angle": 0, + "content": "to the level set surface. This helps us significantly reduce the computational cost of optimizing the coarse model by avoiding the use of finite differencing. Accurate normals can be obtained in the fine stage of optimization when we use a true surface rendering model." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.168, + 0.895, + 0.319 + ], + "angle": 0, + "content": "Similar to DreamFusion, we also model the background using an environment map MLP, which predicts RGB colors as a function of ray directions. Since our sparse representation model does not support scene reparametrization as in Mip-NeRF 360 [3], the optimization has a tendency to \"cheat\" by learning the essence of the object using the background environment map. As such, we use a tiny MLP for the environment map (hidden dimension size of 16) and weigh down the learning rate by \\(10\\times\\) to allow the model to focus more on the neural field geometry." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.323, + 0.895, + 0.399 + ], + "angle": 0, + "content": "Mesh optimization. To optimize a mesh from the neural field initialization, we convert the (coarse) density field to an SDF by subtracting it with a non-zero constant, yielding the initial \\( s_i \\). We additionally initialize the volume texture field directly with the color field optimized from the coarse stage." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.399, + 0.895, + 0.519 + ], + "angle": 0, + "content": "During optimization, we render the extracted surface mesh into high-resolution images using a differentiable rasterizer [18,26]. We optimize both \\( s_i \\) and \\( \\Delta \\mathbf{v}_i \\) for each vertex \\( \\mathbf{v}_i \\) via backpropagation using the high-resolution SDS gradient (Eq. 2). When rendering the mesh to an image, we also track the 3D coordinates of each corresponding pixel projection, which would be used to query colors in the corresponding texture field for joint optimization." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.52, + 0.897, + 0.687 + ], + "angle": 0, + "content": "When rendering the mesh, we increase the focal length to zoom in on object details, which is a critical step towards recovering high-frequency details. We keep the same pre-trained environment map from the coarse stage of optimization and composite the rendered background with the rendered foreground object using differentiable antialIASing [18]. To encourage the smoothness of the surface, we further regularize the angular differences between adjacent faces on the mesh. This allows us to obtain well-behaved geometry even under supervision signals with high variance, such as the SDS gradient \\(\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{SDS}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.698, + 0.634, + 0.716 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.723, + 0.895, + 0.785 + ], + "angle": 0, + "content": "We focus on comparing our method with DreamFusion [31] on 397 text prompts taken from the website of DreamFusion1. We train Magic3D on all of the text prompts and compare them with the results provided on the website." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.788, + 0.895, + 0.88 + ], + "angle": 0, + "content": "Speed evaluation. Unless otherwise noted, the coarse stage is trained for 5000 iterations with 1024 samples along the ray (subsequently filtered by the sparse octree) with a batch size of 32, with a total runtime of around 15 minutes (upwards of 8 iterations / second, variable due to differences in sparsity). The fine stage is trained for 3000 iterations with a batch" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.816, + 0.901 + ], + "angle": 0, + "content": "ComparisonPreferenceMagic3D vs. DreamFusion [31]• More realistic58.3%• More detailed66.0%• More realistic & detailed61.7%Magic3D vs. Magic3D (coarse only)87.7%" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.471, + 0.789 + ], + "angle": 0, + "content": "size of 32 with a total runtime of 25 minutes (2 iterations / second). Both runtimes combined are 40 minutes. All runtimes were measured on 8 NVIDIA A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparisons. We provide qualitative examples in Fig. 3. Qualitatively, our models achieve much higher 3D quality in terms of both geometry and texture. Notice that our model can generate candies on ice cream cones, highly detailed sushi-like cars, vivid strawberries, and birds. We also note that our resulting 3D models can be directly imported and visualized in standard graphics software." + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.517, + 0.836, + 0.53 + ], + "angle": 0, + "content": "a baby bunny sitting on top of a stack of pancakes†" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.53, + 0.871, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.608, + 0.76, + 0.62 + ], + "angle": 0, + "content": "(a) Single-stage model" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.623, + 0.886, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.702, + 0.766, + 0.714 + ], + "angle": 0, + "content": "(b) Coarse-to-fine model" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.727, + 0.895, + 0.798 + ], + "angle": 0, + "content": "Figure 4. Single-stage (top) vs. coarse-to-fine models (bottom). Both use NeRF for the scene model. The left two columns use \\(64 \\times 64\\) rendering resolution during optimization while the right two use \\(256 \\times 256\\). Compared to our coarse-to-fine approach, the single-stage method can generate details but with worse shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.897, + 0.901 + ], + "angle": 0, + "content": "User studies. We conduct user studies to evaluate different methods based on user preferences on Amazon MTurk. We show users two videos side by side rendered from a canonical view by two different algorithms using the same text prompt. We ask the users to select the one that is more realistic" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "305" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.897, + 0.32 + ], + "angle": 0, + "content": "Figure 5. Ablation on the fine-tuning stage. For each text prompt, we compare coarse and fine models with mesh and NeRF representations. Mesh fine-tuning significantly improve the visual quality of generated 3D assets, providing more photo-realistic details on the 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.475, + 0.406 + ], + "angle": 0, + "content": "and detailed. Each prompt is evaluated by 3 different users, resulting in 1191 pairwise comparisons. As shown in Table 1, users favor 3D models generated by Magic3D, with \\(61.7\\%\\) of the users considering our results with higher quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.474, + 0.637 + ], + "angle": 0, + "content": "Can single-stage optimization work with LDM prior? We ablate scene models optimized with high-resolution LDM prior in a single-stage optimization setup. We find that 3D meshes as the scene model fail to generate high-quality results if optimized from scratch. This leaves our memory-efficient sparse 3D representation as the ideal candidate for the scene model. However, rendering \\(512 \\times 512\\) images is still too memory intensive to fit into modern GPUs. Therefore, we render lower-resolution images from the scene model and upsample them to \\(512 \\times 512\\) as input to the LDM. We find it generates objects with worse shapes. Fig. 4 shows two examples with scene rendering resolution \\(64 \\times 64\\) and \\(256 \\times 256\\) respectively (top row). While it generates furry details, the shape is worse than the coarse model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.473, + 0.792 + ], + "angle": 0, + "content": "Can we use NeRF for the fine model? Yes. While optimizing a NeRF from scratch does not work well, we can follow the coarse-to-fine framework but replace the second-stage scene model with a NeRF. In the bottom right of Fig. 4, we show the result of a fine NeRF model initialized with the coarse model on its left and fine-tuned with \\(256 \\times 256\\) rendered images. The two-stage approach retains good geometry in the initial model and adds more details, showing superior quality to its one-stage counterpart." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Coarse models vs. fine models. Fig. 5 provides more visual results contrasting coarse and fine models. We try both NeRF and mesh for scene models and fine-tune them from the same coarse model above. We see significant quality improvements on both NeRF and mesh models, suggesting our coarse-to-fine approach works for general scene models." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.342, + 0.891, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.579, + 0.896, + 0.651 + ], + "angle": 0, + "content": "Figure 6. Magic3D with DreamBooth-based personalization. Given an input image of a particular instance, we fine-tune the diffusion models with DreamBooth and optimize the 3D models with the given prompts. The identity is well preserved in the generated 3D models. Image source (input images): Unsplash." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.76, + 0.699 + ], + "angle": 0, + "content": "6. Controllable 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.895, + 0.803 + ], + "angle": 0, + "content": "As certain styles and concepts are difficult to express in words but easy with images, it is desirable to have a mechanism to influence the text-to-3D model generation with images. We explore different image conditioning techniques as well as a prompt-based editing approach to provide users more control over the 3D generation outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Personalized text-to-3D. DreamBooth [35] described a method to personalize text-to-image diffusion models by fine-tuning a pre-trained model on several images of a subject. The fine-tuned model can learn to tie the subject to a unique identifier string, denoted as \\([\\mathrm{V}]\\), and generate images of the subject when \\([\\mathrm{V}]\\) is included in the text prompt. In" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "306" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.089, + 0.155, + 0.111 + ], + "angle": 0, + "content": "low-resolution before edited" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.113, + 0.154, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.09, + 0.387, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.091, + 0.649, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.099, + 0.875, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.203, + 0.155, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.203, + 0.393, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.207, + 0.641, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.217, + 0.869, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 7. Magic3D with prompt-based editing. Given a coarse model (first column) generated with a base prompt, we replace the underscored text with new text and fine-tune the NeRF to get a high-resolution NeRF model with LDM. We further fine-tune the high-resolution mesh with the NeRF model. Such a prompt-based editing method gives artists better control over the 3D generation output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.478 + ], + "angle": 0, + "content": "the context of text-to-3D generation, we would like to generate a 3D model of a subject. This can be achieved by first fine-tuning our diffusion prior models with the DreamBooth approach, and then using the fine-tuned diffusion priors with the [V] identifier as part of the conditioning text prompt to provide the learning signal when optimizing the 3D model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.471, + 0.651 + ], + "angle": 0, + "content": "To demonstrate the applicability of DreamBooth in our framework, we collect 11 images of one cat and 4 images of one dog. We fine-tune eDiff-I [2] and LDM [34], binding the text identifier [V] to the given subject. Then, we optimize the 3D model with [V] in the text prompts. We use a batch size of 1 for all fine-tuning. For eDiff-I, we use the Adam optimizer with learning rate \\(1 \\times 10^{-5}\\) for 1,500 iterations; for LDM, we fine-tune with learning rate \\(1 \\times 10^{-6}\\) for 800 iterations. Fig. 6 shows our personalized text-to-3D results: we are able to successfully modify the 3D models preserving the subjects in the given input images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Prompt-based editing through fine-tuning. Another way to control the generated 3D content is by fine-tuning a learned coarse model with a new prompt. Our prompt-based editing includes three stages. (a) We train a coarse model with a base prompt. (b) We modify the base prompt and fine-tune the coarse model with the LDM. This stage provides a well initialized NeRF model for the next step. Directly applying mesh optimization on a new prompt would generate highly detailed textures but could deform geometry only slightly. (c) We optimize the mesh with the modified text prompt. Our prompt-based editing can modify the texture of the shape or transform the geometry and texture according to the text. The resulting scene models preserve the layer-out and overall structure. Such an editing capability makes the 3D content creation with Magic3D more controllable. In Fig. 7, we show two coarse NeRF models trained with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.895, + 0.477 + ], + "angle": 0, + "content": "base prompt for the \"bunny\" and \"squirrel\". We modify the base prompt, fine-tune the NeRF model in high resolution and optimize the mesh. Results show that we can tune the scene model according to the prompt, e.g. changing the \"baby bunny\" to \"stained glass bunny\" or \"metal bunny\" results in similar geometry but with a different texture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.516, + 0.619, + 0.532 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.779 + ], + "angle": 0, + "content": "We propose Magic3D, a fast and high-quality text-to-3D generation framework. We benefit from both efficient scene models and high-resolution diffusion priors in a coarse-to-fine approach. In particular, the 3D mesh models scale nicely with image resolution and enjoy the benefits of higher resolution supervision brought by the latent diffusion model without sacrificing its speed. It takes 40 minutes from a text prompt to a high-quality 3D mesh model ready to be used in graphic engines. With extensive user studies and qualitative comparisons, we show that Magic3D is more preferable \\((61.7\\%)\\) by the raters compared to DreamFusion, while enjoying a \\(2\\times\\) speed-up. Lastly, we propose a set of tools for better controlling style and content in 3D generation. We hope with Magic3D, we can democratize 3D synthesis and open up everyone's creativity in 3D content creation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements. We would like to thank Frank Shen, Yogesh Balaji, Seungjun Nah, James Lucas, David Luebke, Clement Fuji-Tsang, Charles Loop, Qinsheng Zhang, Zan Gojcic, and Jonathan Tremblay for helpful discussions and paper proofreading. We would also like to thank Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall for providing additional implementation details in DreamFusion." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "307" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.176, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.321, + 0.472, + 0.405 + ], + "angle": 0, + "content": "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.407, + 0.472, + 0.478 + ], + "angle": 0, + "content": "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.472, + 0.536 + ], + "angle": 0, + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.472, + 0.594 + ], + "angle": 0, + "content": "[7] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In 2017 International Conference on 3D Vision (3DV), pages 402-411. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.597, + 0.472, + 0.665 + ], + "angle": 0, + "content": "[8] Jun Gao, Wenzheng Chen, Tommy Xiang, Clement Fuji Tsang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. In Advances In Neural Information Processing Systems, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.67, + 0.472, + 0.739 + ], + "angle": 0, + "content": "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances in Neural Information Processing Systems, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.472, + 0.798 + ], + "angle": 0, + "content": "[10] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.472, + 0.842 + ], + "angle": 0, + "content": "[11] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[12] Philipp Henzler, Niloy J. Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[15] Moritz Ibing, Gregor Kobsik, and Leif Kobbelt. Octree transformer: Autoregressive 3d shape generation on hierarchically structured sequences. arXiv preprint arXiv:2111.12480, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[16] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.894, + 0.334 + ], + "angle": 0, + "content": "[17] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. ACM Transactions on Graphics (TOG), Proc. SIGGRAPH Asia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[18] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.894, + 0.447 + ], + "angle": 0, + "content": "[19] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.894, + 0.503 + ], + "angle": 0, + "content": "[20] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[21] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.894, + 0.631 + ], + "angle": 0, + "content": "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.894, + 0.757 + ], + "angle": 0, + "content": "[24] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. ACM Transactions on Graphics (TOG), Siggraph Asia 2019, 38(6):Article 242, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[26] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, pages 8280-8290, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.903 + ], + "angle": 0, + "content": "[27] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "308" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[29] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.472, + 0.329 + ], + "angle": 0, + "content": "[30] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.509 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.472, + 0.636 + ], + "angle": 0, + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.472, + 0.789 + ], + "angle": 0, + "content": "[37] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[38] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[39] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[40] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In Conference on Robot Learning, pages 87-96. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[42] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[43] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.894, + 0.402 + ], + "angle": 0, + "content": "[44] Towaki Takikawa, Or Perel, Clement Fuji Tsang, Charles Loop, Joey Litalien, Jonathan Tremblay, Sanja Fidler, and Maria Shugrina. Kaolin wisp: A pytorch library and engine for neural fields research. https://github.com/NVIDIAGameWorks/kaolin-wisp, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[45] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.894, + 0.529 + ], + "angle": 0, + "content": "[46] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.894, + 0.598 + ], + "angle": 0, + "content": "[47] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.894, + 0.669 + ], + "angle": 0, + "content": "[48] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yinan Zhang, Antonio Torralba, and Sanja Fidler. Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.894, + 0.727 + ], + "angle": 0, + "content": "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826-5835, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "309" + } + ] +] \ No newline at end of file diff --git a/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_origin.pdf b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0cdd48d5e2fb959b377c8fa42b5e7ec0eef9e747 --- /dev/null +++ b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/1e2b4976-fbca-4313-8151-9395cc1c1413_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dba6f6506ff9965dcad070535666614c7328d27b53e1f530fd6d736637729e6 +size 3923479 diff --git a/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/full.md b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f9ffc2e182ccaf6dd40b51f135fedafbbacced2 --- /dev/null +++ b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/full.md @@ -0,0 +1,290 @@ +# Magic3D: High-Resolution Text-to-3D Content Creation + +Chen-Hsuan Lin* Jun Gao* Luming Tang* Towaki Takikawa* Xiaohui Zeng* Xun Huang Karsten Kreis Sanja Fidler† Ming-Yu Liu† Tsung-Yi Lin + +NVIDIA Corporation + +https://research.nvidia.com/labs/dir/magic3d + +# Abstract + +DreamFusion [31] has recently demonstrated the utility of a pre-trained text-to-image diffusion model to optimize Neural Radiance Fields (NeRF) [23], achieving remarkable text-to-3D synthesis results. However, the method has two inherent limitations: (a) extremely slow optimization of NeRF and (b) low-resolution image space supervision on NeRF, leading to low-quality 3D models with a long processing time. In this paper, we address these limitations by utilizing a two-stage optimization framework. First, we obtain a coarse model using a low-resolution diffusion prior and accelerate with a sparse 3D hash grid structure. Using the coarse representation as the initialization, we further optimize a textured 3D mesh model with an efficient differentiable renderer interacting with a high-resolution latent diffusion model. Our method, dubbed Magic3D, can create high quality 3D mesh models in 40 minutes, which is $2 \times$ faster than DreamFusion (reportedly taking 1.5 hours on average), while also achieving higher resolution. User studies show $61.7\%$ raters to prefer our approach over DreamFusion. Together with the image-conditioned generation capabilities, we provide users with new ways to control 3D synthesis, opening up new avenues to various creative applications. + +# 1. Introduction + +3D digital content has been in high demand for a variety of applications, including gaming, entertainment, architecture, and robotics simulation. It is slowly finding its way into virtually every possible domain: retail, online conferencing, virtual social presence, education, etc. However, creating professional 3D content is not for anyone — it requires immense artistic and aesthetic training with 3D modeling expertise. Developing these skill sets takes a significant amount of time and effort. Augmenting 3D content creation with natural language could considerably help democratize 3D content creation for novices and turbocharge expert artists. + +Image content creation from text prompts [2, 28, 33, 36] has seen significant progress with the advances of diffusion models [13, 41, 42] for generative modeling of images. The key enablers are large-scale datasets comprising billions of samples (images with text) scrapped from the Internet and massive amounts of compute. In contrast, 3D content generation has progressed at a much slower pace. Existing 3D object generation models [4, 9, 47] are mostly categorical. A trained model can only be used to synthesize objects for a single class, with early signs of scaling to multiple classes shown recently by Zeng et al. [47]. Therefore, what a user can do with these models is extremely limited and not yet ready for artistic creation. This limitation is largely due to the lack of diverse large-scale 3D datasets — compared to image and video content, 3D content is much less accessible on the Internet. This naturally raises the question of whether 3D generation capability can be achieved by leveraging powerful text-to-image generative models. + +Recently, DreamFusion [31] demonstrated its remarkable ability for text-conditioned 3D content generation by utilizing a pre-trained text-to-image diffusion model [36] that generates images as a strong image prior. The diffusion model acts as a critic to optimize the underlying 3D representation. The optimization process ensures that rendered images from a 3D model, represented by Neural Radiance Fields (NeRF) [23], match the distribution of photorealistic images across different viewpoints, given the input text prompt. Since the supervision signal in DreamFusion operates on very low-resolution images $(64 \times 64)$ , DreamFusion cannot synthesize high-frequency 3D geometric and texture details. Due to the use of inefficient MLP architectures for the NeRF representation, practical high-resolution synthesis may not even be possible as the required memory footprint and the computation budget grows quickly with the resolution. Even at a resolution of $64 \times 64$ , optimization times are in hours (1.5 hours per prompt on average using TPUv4). + +In this paper, we present a method that can synthesize highly detailed 3D models from text prompts within a reduced computation time. Specifically, we propose a coarse- + +![](images/d6080e0d70aab5a7349d9ec09e3d7cd274505bee320234f86bbb32c16dd37c18.jpg) + +![](images/5b8245c2c17cd88f41899b3052ead17133ef619ff7ad49ee3c13a8bd95c7d5f5.jpg) +michelangelo style statue of an astronaut + +![](images/0c71b85b08404d1ddfe504a5f8c22cafa8c3d86eb243d74b9b468882241275af.jpg) +a stuffed grey rabbit holding a pretend carrot + +![](images/6f428019d63476ad15c8772a3814a6a3c272bd31916bb5ec79bb7bdf370567ba.jpg) +an iguana holding a balloon + +![](images/2f151281677b782a6d0bf393a5109d239305e750e398997cbc0752d41bb9dc72.jpg) +a silver platter piled high with fruits + +![](images/fc43f3109dee508f9f835868839a0bcf438823adfcb214d1494dc0733f869f37.jpg) +a beautiful dress made out of garbage bags +an imperial state crown of england + +![](images/829b8fdf0d1f6df006853c44bda46266ed7158c5c1291ff309f56b611df95d09.jpg) + +![](images/3dc0649342845e4741b346842db3a6a860a0577be4336ef28ba4362f6d77ea11.jpg) +neuschwanstein castle, aerial view + +Low resolution bunny before editing + +a baby bunny sitting on top of a stack of pancakes + +![](images/bca0e717bd1f384e994c7764818373ae4f37872048de5e7149cf01c0ebb22fc8.jpg) + +a metal bunny sitting on top of a stack of broccoli + +![](images/52e09933ed0dc7046df0273ae8facc1f4ac775854703f1ea20cee58d3c042ae0.jpg) +a blue poison-dart frog sitting on a water lily +Figure 1. Results and applications of Magic3D. Top: high-resolution text-to-3D generation. Magic3D can generate high-quality and high-resolution 3D models from text prompts. Bottom: high-resolution prompt-based editing. Magic3D can edit 3D models by fine-tuning with the diffusion prior using a different prompt. Taking the low-resolution 3D model as the input (left), Magic3D can modify different parts of the 3D model corresponding to different input text prompts. Together with various creative controls on the generated 3D models, Magic3D is a convenient tool for augmenting 3D content creation. + +a metalbunnysitting ontop of asstack ofchocolatecookie + +![](images/e34718cd295dcc17dc3640658aa58a4c5c84c0ff911be780ac3087f184c6f636.jpg) + +a sphinx sitting on top of a stack of chocolate cookie + +![](images/81d041100aaeb45340839b4086dd5e57d9051d41e929f88339d3842fb4318da0.jpg) + +to-fine optimization approach that uses multiple diffusion priors at different resolutions to optimize the 3D representation, enabling the generation of both view-consistent geometry as well as high-resolution details. In the first stage, we optimize a coarse neural field representation akin to Dream-Fusion, but with a memory- and compute-efficient scene representation based on a hash grid [25]. In the second stage, + +we switch to optimizing mesh representations, a critical step that allows us to utilize diffusion priors at resolutions as high as $512 \times 512$ . As 3D meshes are amenable to fast graphics renderers that can render high-resolution images in real-time, we leverage an efficient differentiable rasterizer [9, 26] and make use of camera close-ups to recover high-frequency details in geometry and texture. As a result, our approach + +produces high-fidelity 3D content (see Fig. 1) that can conveniently be imported and visualized in standard graphics software and does so at $2 \times$ the speed of DreamFusion. Furthermore, we showcase various creative controls over the 3D synthesis process by leveraging the advancements developed for text-to-image editing applications [2, 35]. Our approach, dubbed Magic3D, endows users with unprecedented control in crafting their desired 3D objects with text prompts and reference images, bringing this technology one step closer to democratizing 3D content creation. + +In summary, our work makes the following contributions: + +- We propose Magic3D, a framework for high-quality 3D content synthesis using text prompts by improving several major design choices made in DreamFusion. It consists of a coarse-to-fine strategy that leverages both low- and high-resolution diffusion priors for learning the 3D representation of the target content. Magic3D, which synthesizes 3D content with an $8 \times$ higher resolution supervision, is also $2 \times$ faster than DreamFusion. 3D content synthesized by our approach is significantly preferable by users (61.7%). +- We extend various image editing techniques developed for text-to-image models to 3D object editing and show their applications in the proposed framework. + +# 2. Related Work + +Text-to-image generation. We have witnessed significant progress in text-to-image generation with diffusion models in recent years. With improvements in modeling and data curation, diffusion models can compose complex semantic concepts from text descriptions (nouns, adjectives, artistic styles, etc.) to generate high-quality images of objects and scenes [2, 33, 34, 36]. Sampling images from diffusion models is time consuming. To generate high-resolution images, these models either utilize a cascade of super-resolution models [2, 36] or sample from a lower-resolution latent space and decode latents into high-resolution images [34]. Despite the advances in high-resolution image generation, using language to describe and control 3D properties (e.g. camera viewpoints) while maintaining coherency in 3D remains an open, challenging problem. + +3D generative models. There is a large body of work on 3D generative modeling, exploring different types of 3D representations such as 3D voxel grids [7, 12, 20, 40, 45], pointclouds [1, 21, 24, 46, 47, 49], meshes [9, 48], implicit [6, 22], or octree [15] representations. Most of these approaches rely on training data in the form of 3D assets, which are hard to acquire at scale. Inspired by the success of neural volume rendering [23], recent works started investing in 3D-aware image synthesis [4, 5, 10, 11, 27, 29, 30, 38], which has the advantage of learning 3D generative models directly from images — a more widely accessible resource. However, vol- + +ume rendering networks are typically slow to query, leading to a trade-off between long training time [5, 29] and lack of multi-view consistency [10]. EG3D [4] partially mitigates this problem by utilizing a dual discriminator. While obtaining promising results, these works remain limited to modeling objects within a single object category, such as cars, chairs, or human faces, thus lacking scalability and the creative control desired for 3D content creation. In our paper, we focus on text-to-3D synthesis, aiming to generate a 3D renderable representation of a scene based on a text prompt. + +Text-to-3D generation. With the recent success in text-to-image generative modeling in recent years, text-to-3D generation has also gained a surge of interest from the learning community. Earlier works such as CLIP-forge [37] synthesize objects by learning a normalizing flow model to sample shape embeddings from textual input. However, it requires 3D assets in voxel representations during training, making it challenging to scale with data. DreamField [16] and CLIP-mesh [17] mitigate the training data issue by relying on a pre-trained image-text model [32] to optimize the underlying 3D representations (NeRFs and meshes), such that all 2D renderings reach high text-image alignment scores. While these approaches avoid the requirement of expensive 3D training data and mostly rely on pre-trained large-scale image-text models, they tend to produce less realistic 2D renderings. + +Recently, DreamFusion [31] showcased impressive capability in text-to-3D synthesis by utilizing a powerful pretrained text-to-image diffusion model [36] as a strong image prior. We build upon this work and improve over several design choices to bring significantly higher-fidelity 3D models into hands of users with a much reduced generation time. + +# 3. Background: DreamFusion + +DreamFusion [31] achieves text-to-3D generation with two key components: a neural scene representation which we refer to as the scene model, and a pre-trained text-to-image diffusion-based generative model. The scene model is a parametric function $x = g(\theta)$ , which can produce an image $x$ at the desired camera pose. Here, $g$ is a volumetric renderer of choice, and $\theta$ is a coordinate-based MLP representing a 3D volume. The diffusion model $\phi$ comes with a learned denoising function $\epsilon_{\phi}(x_t; y, t)$ that predicts the sampled noise $\epsilon$ given the noisy image $x_t$ , noise level $t$ , and text embedding $y$ . It provides the gradient direction to update $\theta$ such that all rendered images are pushed to the high probability density regions conditioned on the text embedding under the diffusion prior. Specifically, DreamFusion introduces Score Distillation Sampling (SDS), which computes the gradient: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\phi , g (\theta)) = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\epsilon_ {\phi} \left(x _ {t}; y, t\right) - \epsilon\right) \frac {\partial x}{\partial \theta} \right]. \tag {1} +$$ + +Here, $w(t)$ is a weighting function. We view the scene model $g$ and the diffusion model $\phi$ as modular components of the + +![](images/993ed516c2a474d9377e1c231fe571903233398327ee30a97216a721d5b9ac10.jpg) +Figure 2. Overview of Magic3D. We generate high-resolution 3D content from an input text prompt in a coarse-to-fine manner. In the first stage, we utilize a low-resolution diffusion prior and optimize neural field representations (color, density, and normal fields) to obtain the coarse model. We further differentiably extract textured 3D mesh from the density and color fields of the coarse model. Then we fine-tune it using a high-resolution latent diffusion model. After optimization, our model generates high-quality 3D meshes with detailed textures. + +framework, amenable to choice. In practice, the denoising function $\epsilon_{\phi}$ is often replaced with another function $\tilde{\epsilon}_{\phi}$ that uses classifier-free guidance [14], which allows one to carefully weigh the strength of the text conditioning (see Sec. 6). DreamFusion relies on large classifier-free guidance weights to obtain results with better quality. + +DreamFusion adopts a variant of Mip-NeRF 360 [3] with an explicit shading model for the scene model and Imagegen [36] as the diffusion model. These choices result in two key limitations. First, high-resolution geometry or textures cannot be obtained since the diffusion model only operates on $64 \times 64$ images. Second, the utility of a large global MLP for volume rendering is both computationally expensive as well as memory intensive, making this approach scale poorly with the increasing resolution of images. + +# 4. High-Resolution 3D Generation + +Magic3D is a two-stage coarse-to-fine framework that uses efficient scene models that enable high-resolution text-to-3D synthesis (Fig. 2). We describe our method and key differences from DreamFusion [31] in this section. + +# 4.1. Coarse-to-fine Diffusion Priors + +Magic3D uses two different diffusion priors in a coarse-to-fine fashion to generate high-resolution geometry and textures. In the first stage, we use the base diffusion model described in eDiff-I [2], which is similar to the base diffusion model of Imagen [36] used in DreamFusion. This diffusion prior is used to compute gradients of the scene model via a loss defined on rendered images at a low resolution $64 \times 64$ . In the second stage, we use the latent diffusion model (LDM) [34] that allows backpropagating gradients into rendered images at a high resolution $512 \times 512$ ; in practice, we choose to use the publicly available Stable Diffusion + +model [34]. Despite generating high-resolution images, the computation of LDM is manageable because the diffusion prior acts on the latent $z_{t}$ with resolution $64 \times 64$ : + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\phi , g (\theta)) = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\epsilon_ {\phi} \left(z _ {t}; y, t\right) - \epsilon\right) \frac {\partial z}{\partial x} \frac {\partial x}{\partial \theta} \right]. \tag {2} +$$ + +The increase in computation time mainly comes from computing $\frac{\partial x}{\partial \theta}$ (the gradient of the high-resolution rendered image) and $\frac{\partial z}{\partial x}$ (the gradient of the encoder in LDM). + +# 4.2. Scene Models + +We cater two different 3D scene representations to the two different diffusion priors at coarse and fine resolutions to accommodate the increased resolution of rendered images for the input of high-resolution priors, discussed as follows. + +Neural fields as coarse scene models. The initial coarse stage of the optimization requires finding the geometry and textures from scratch. This can be challenging as we need to accommodate complex topological changes in the 3D geometry and depth ambiguities from the 2D supervision signals. In DreamFusion [31], the scene model is a neural field (a coordinate-based MLP) based on Mip-NeRF 360 [3] that predicts albedo and density. This is a suitable choice as neural fields can handle topological changes in a smooth, continuous fashion. However, Mip-NeRF 360 [3] is computationally expensive as it is based on a large global coordinate-based MLP. As volume rendering requires dense samples along a ray to accurately render high-frequency geometry and shading, the cost of having to evaluate a large neural network at every sample point quickly stacks up. + +For this reason, we opt to use the hash grid encoding from Instant NGP [25], which allows us to represent high-frequency details at a much lower computational cost. We use the hash grid with two single-layer neural networks, + +one predicting albedo and density and the other predicting normals. We additionally maintain a spatial data structure that encodes scene occupancy and utilizes empty space skipping [19, 43]. Specifically, we use the density-based voxel pruning approach from Instant NGP [25] with an octree-based ray sampling and rendering algorithm [44]. With these design choices, we drastically accelerate the optimization of coarse scene models while maintaining quality. + +Textured meshes as fine scene models. In our fine stage of optimization, we need to be able to accommodate very high-resolution rendered images to fine-tune our scene model with high-resolution diffusion priors. Using the same scene representation (the neural field) from the initial coarse stage of optimization could be a natural choice since the weights of the model can directly carry over. Although this strategy can work to some extent (Figs. 4 and 5), it struggles to render very high-resolution (e.g., $512 \times 512$ ) images within reasonable memory constraints and computation budgets. + +To resolve this issue, we use textured 3D meshes as the scene representation for the fine stage of optimization. In contrast to volume rendering for neural fields, rendering textured meshes with differentiable rasterization can be performed efficiently at very high resolutions, making meshes a suitable choice for our high-resolution optimization stage. Using the neural field from the coarse stage as the initialization for the mesh geometry, we can also sidestep the difficulty of learning large topological changes in meshes. + +Formally, we represent the 3D shape using a deformable tetrahedral grid $(V_{T}, T)$ , where $V_{T}$ is the vertices in the grid $T$ [8, 39]. Each vertex $\mathbf{v}_{i} \in V_{T} \subset \mathbb{R}^{3}$ contains a signed distance field (SDF) value $s_{i} \in \mathbb{R}$ and a deformation $\Delta \mathbf{v}_{i} \in \mathbb{R}^{3}$ of the vertex from its initial canonical coordinate. Then, we extract a surface mesh from the SDF using a differentiable marching tetrahedra algorithm [39]. For textures, we use the neural color field as a volumetric texture representation. + +# 4.3. Coarse-to-fine Optimization + +We describe our coarse-to-fine optimization procedure, which first operates on a coarse neural field representation and subsequently a high-resolution textured mesh. + +Neural field optimization. Similarly to Instant NGP [25], we initialize an occupancy grid of resolution $256^3$ with values to 20 to encourage shapes to grow in the early stages of optimization. We update the grid every 10 iterations and generate an octree for empty space skipping. We decay the occupancy grid by 0.6 in every update and follow Instant NGP with the same update and thresholding parameters. + +Instead of estimating normals from density differences, we use an MLP to predict the normals. Note that this does not violate geometric properties since volume rendering is used instead of surface rendering; as such, the orientation of particles at continuous positions need not be oriented + +to the level set surface. This helps us significantly reduce the computational cost of optimizing the coarse model by avoiding the use of finite differencing. Accurate normals can be obtained in the fine stage of optimization when we use a true surface rendering model. + +Similar to DreamFusion, we also model the background using an environment map MLP, which predicts RGB colors as a function of ray directions. Since our sparse representation model does not support scene reparametrization as in Mip-NeRF 360 [3], the optimization has a tendency to "cheat" by learning the essence of the object using the background environment map. As such, we use a tiny MLP for the environment map (hidden dimension size of 16) and weigh down the learning rate by $10\times$ to allow the model to focus more on the neural field geometry. + +Mesh optimization. To optimize a mesh from the neural field initialization, we convert the (coarse) density field to an SDF by subtracting it with a non-zero constant, yielding the initial $s_i$ . We additionally initialize the volume texture field directly with the color field optimized from the coarse stage. + +During optimization, we render the extracted surface mesh into high-resolution images using a differentiable rasterizer [18,26]. We optimize both $s_i$ and $\Delta \mathbf{v}_i$ for each vertex $\mathbf{v}_i$ via backpropagation using the high-resolution SDS gradient (Eq. 2). When rendering the mesh to an image, we also track the 3D coordinates of each corresponding pixel projection, which would be used to query colors in the corresponding texture field for joint optimization. + +When rendering the mesh, we increase the focal length to zoom in on object details, which is a critical step towards recovering high-frequency details. We keep the same pre-trained environment map from the coarse stage of optimization and composite the rendered background with the rendered foreground object using differentiable antialIASing [18]. To encourage the smoothness of the surface, we further regularize the angular differences between adjacent faces on the mesh. This allows us to obtain well-behaved geometry even under supervision signals with high variance, such as the SDS gradient $\nabla_{\theta}\mathcal{L}_{\mathrm{SDS}}$ + +# 5. Experiments + +We focus on comparing our method with DreamFusion [31] on 397 text prompts taken from the website of DreamFusion1. We train Magic3D on all of the text prompts and compare them with the results provided on the website. + +Speed evaluation. Unless otherwise noted, the coarse stage is trained for 5000 iterations with 1024 samples along the ray (subsequently filtered by the sparse octree) with a batch size of 32, with a total runtime of around 15 minutes (upwards of 8 iterations / second, variable due to differences in sparsity). The fine stage is trained for 3000 iterations with a batch + +![](images/3ff3963d6791dab78b1f1d1ba992402189b280f8225a5f45fc1de200d2c512f2.jpg) +Figure 3. Qualitative comparison with DreamFusion [31]. We use the same text prompt as in DreamFusion. For each 3D model, we render it from two views with a textureless rendering for each view and remove the background to focus on the actual 3D shape. For the DreamFusion results, we take frames from the videos published on the official webpage. Our Magic3D generates much higher quality 3D shapes on both geometry and texture compared with DreamFusion. * a DSLR photo of... † a zoomed out DSLR photo of... + +Table 1. User preference studies. We conducted user studies to measure preference for 3D models generated using 397 prompts released by DreamFusion. Overlal, more raters $(61.7\%)$ prefer 3D models generated by Magic3D over DreamFusion. The majority of raters $(87.7\%)$ prefer fine models over coarse models in Magic3D, showing the effectiveness of our coarse-to-fine approach. + +
ComparisonPreference
Magic3D vs. DreamFusion [31]
• More realistic58.3%
• More detailed66.0%
• More realistic & detailed61.7%
Magic3D vs. Magic3D (coarse only)87.7%
+ +size of 32 with a total runtime of 25 minutes (2 iterations / second). Both runtimes combined are 40 minutes. All runtimes were measured on 8 NVIDIA A100 GPUs. + +Qualitative comparisons. We provide qualitative examples in Fig. 3. Qualitatively, our models achieve much higher 3D quality in terms of both geometry and texture. Notice that our model can generate candies on ice cream cones, highly detailed sushi-like cars, vivid strawberries, and birds. We also note that our resulting 3D models can be directly imported and visualized in standard graphics software. + +a baby bunny sitting on top of a stack of pancakes† + +![](images/43eb0c7a0ff9d607f3f4eaefc8982db71414077c938d07a21007740ab1d9e9c7.jpg) +(a) Single-stage model + +![](images/5c0414b0340731f072ebd6195646369ba0ae237f5b9e657f726f08da5c1a48b6.jpg) +(b) Coarse-to-fine model +Figure 4. Single-stage (top) vs. coarse-to-fine models (bottom). Both use NeRF for the scene model. The left two columns use $64 \times 64$ rendering resolution during optimization while the right two use $256 \times 256$ . Compared to our coarse-to-fine approach, the single-stage method can generate details but with worse shapes. + +User studies. We conduct user studies to evaluate different methods based on user preferences on Amazon MTurk. We show users two videos side by side rendered from a canonical view by two different algorithms using the same text prompt. We ask the users to select the one that is more realistic + +![](images/2128d779f88b92785de3b96998a0cb96af27762b568b0b9fe1c5495676770b0b.jpg) +Figure 5. Ablation on the fine-tuning stage. For each text prompt, we compare coarse and fine models with mesh and NeRF representations. Mesh fine-tuning significantly improve the visual quality of generated 3D assets, providing more photo-realistic details on the 3D shapes. + +and detailed. Each prompt is evaluated by 3 different users, resulting in 1191 pairwise comparisons. As shown in Table 1, users favor 3D models generated by Magic3D, with $61.7\%$ of the users considering our results with higher quality. + +Can single-stage optimization work with LDM prior? We ablate scene models optimized with high-resolution LDM prior in a single-stage optimization setup. We find that 3D meshes as the scene model fail to generate high-quality results if optimized from scratch. This leaves our memory-efficient sparse 3D representation as the ideal candidate for the scene model. However, rendering $512 \times 512$ images is still too memory intensive to fit into modern GPUs. Therefore, we render lower-resolution images from the scene model and upsample them to $512 \times 512$ as input to the LDM. We find it generates objects with worse shapes. Fig. 4 shows two examples with scene rendering resolution $64 \times 64$ and $256 \times 256$ respectively (top row). While it generates furry details, the shape is worse than the coarse model. + +Can we use NeRF for the fine model? Yes. While optimizing a NeRF from scratch does not work well, we can follow the coarse-to-fine framework but replace the second-stage scene model with a NeRF. In the bottom right of Fig. 4, we show the result of a fine NeRF model initialized with the coarse model on its left and fine-tuned with $256 \times 256$ rendered images. The two-stage approach retains good geometry in the initial model and adds more details, showing superior quality to its one-stage counterpart. + +Coarse models vs. fine models. Fig. 5 provides more visual results contrasting coarse and fine models. We try both NeRF and mesh for scene models and fine-tune them from the same coarse model above. We see significant quality improvements on both NeRF and mesh models, suggesting our coarse-to-fine approach works for general scene models. + +![](images/240aac95d822c1e949b296249d85ec84fbd191643e437d3b29c80490eb1a5d38.jpg) +Figure 6. Magic3D with DreamBooth-based personalization. Given an input image of a particular instance, we fine-tune the diffusion models with DreamBooth and optimize the 3D models with the given prompts. The identity is well preserved in the generated 3D models. Image source (input images): Unsplash. + +# 6. Controllable 3D Generation + +As certain styles and concepts are difficult to express in words but easy with images, it is desirable to have a mechanism to influence the text-to-3D model generation with images. We explore different image conditioning techniques as well as a prompt-based editing approach to provide users more control over the 3D generation outputs. + +Personalized text-to-3D. DreamBooth [35] described a method to personalize text-to-image diffusion models by fine-tuning a pre-trained model on several images of a subject. The fine-tuned model can learn to tie the subject to a unique identifier string, denoted as $[\mathrm{V}]$ , and generate images of the subject when $[\mathrm{V}]$ is included in the text prompt. In + +![](images/397ee7678d79380925b096100a42da2efde6ef375dab86f84cde6612af37bf61.jpg) +low-resolution before edited + +![](images/75f10d2c9750fb9b437ace0f89d2872fc7702a5eed360877b1ac99bb2d3dae23.jpg) + +![](images/6a2ef0985bfc4aa89b6bd728b0bff926d8c6c7c9964e2007dcf044c8d9c128bd.jpg) + +![](images/1be9748663d9b6cdf89903c2cf2a1245fa77ff9eddb78ae5021f3c41cbca3e7a.jpg) + +![](images/015df1f9d4199c2503bd74bce5830ec1a647405ae748aedf2899d50b022abc0d.jpg) + +![](images/21a0dcb00660f9a0c9662d005bc7cc172cd9c1dac0732d077d42b09b1c30338d.jpg) +Figure 7. Magic3D with prompt-based editing. Given a coarse model (first column) generated with a base prompt, we replace the underscored text with new text and fine-tune the NeRF to get a high-resolution NeRF model with LDM. We further fine-tune the high-resolution mesh with the NeRF model. Such a prompt-based editing method gives artists better control over the 3D generation output. + +![](images/7809c54ad73d694302db560d32c30b0831b3652c4f3160efe38b9322efc230a9.jpg) + +![](images/ca1779110d89b9f94b2208ce0535fdbacdb5811d5cd1f4a10885b3392f0c3d10.jpg) + +the context of text-to-3D generation, we would like to generate a 3D model of a subject. This can be achieved by first fine-tuning our diffusion prior models with the DreamBooth approach, and then using the fine-tuned diffusion priors with the [V] identifier as part of the conditioning text prompt to provide the learning signal when optimizing the 3D model. + +To demonstrate the applicability of DreamBooth in our framework, we collect 11 images of one cat and 4 images of one dog. We fine-tune eDiff-I [2] and LDM [34], binding the text identifier [V] to the given subject. Then, we optimize the 3D model with [V] in the text prompts. We use a batch size of 1 for all fine-tuning. For eDiff-I, we use the Adam optimizer with learning rate $1 \times 10^{-5}$ for 1,500 iterations; for LDM, we fine-tune with learning rate $1 \times 10^{-6}$ for 800 iterations. Fig. 6 shows our personalized text-to-3D results: we are able to successfully modify the 3D models preserving the subjects in the given input images. + +Prompt-based editing through fine-tuning. Another way to control the generated 3D content is by fine-tuning a learned coarse model with a new prompt. Our prompt-based editing includes three stages. (a) We train a coarse model with a base prompt. (b) We modify the base prompt and fine-tune the coarse model with the LDM. This stage provides a well initialized NeRF model for the next step. Directly applying mesh optimization on a new prompt would generate highly detailed textures but could deform geometry only slightly. (c) We optimize the mesh with the modified text prompt. Our prompt-based editing can modify the texture of the shape or transform the geometry and texture according to the text. The resulting scene models preserve the layer-out and overall structure. Such an editing capability makes the 3D content creation with Magic3D more controllable. In Fig. 7, we show two coarse NeRF models trained with the + +base prompt for the "bunny" and "squirrel". We modify the base prompt, fine-tune the NeRF model in high resolution and optimize the mesh. Results show that we can tune the scene model according to the prompt, e.g. changing the "baby bunny" to "stained glass bunny" or "metal bunny" results in similar geometry but with a different texture. + +# 7. Conclusion + +We propose Magic3D, a fast and high-quality text-to-3D generation framework. We benefit from both efficient scene models and high-resolution diffusion priors in a coarse-to-fine approach. In particular, the 3D mesh models scale nicely with image resolution and enjoy the benefits of higher resolution supervision brought by the latent diffusion model without sacrificing its speed. It takes 40 minutes from a text prompt to a high-quality 3D mesh model ready to be used in graphic engines. With extensive user studies and qualitative comparisons, we show that Magic3D is more preferable $(61.7\%)$ by the raters compared to DreamFusion, while enjoying a $2\times$ speed-up. Lastly, we propose a set of tools for better controlling style and content in 3D generation. We hope with Magic3D, we can democratize 3D synthesis and open up everyone's creativity in 3D content creation. + +Acknowledgements. We would like to thank Frank Shen, Yogesh Balaji, Seungjun Nah, James Lucas, David Luebke, Clement Fuji-Tsang, Charles Loop, Qinsheng Zhang, Zan Gojcic, and Jonathan Tremblay for helpful discussions and paper proofreading. We would also like to thank Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall for providing additional implementation details in DreamFusion. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 3 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 3, 4, 8 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 4, 5 +[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 3 +[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3 +[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3 +[7] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In 2017 International Conference on 3D Vision (3DV), pages 402-411. IEEE, 2017. 3 +[8] Jun Gao, Wenzheng Chen, Tommy Xiang, Clement Fuji Tsang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. In Advances In Neural Information Processing Systems, 2020. 5 +[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances in Neural Information Processing Systems, 2022. 1, 2, 3 +[10] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 3 +[11] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds. In ICCV, 2021. 3 +[12] Philipp Henzler, Niloy J. Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3 + +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 4 +[15] Moritz Ibing, Gregor Kobsik, and Leif Kobbelt. Octree transformer: Autoregressive 3d shape generation on hierarchically structured sequences. arXiv preprint arXiv:2111.12480, 2021. 3 +[16] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 3 +[17] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. ACM Transactions on Graphics (TOG), Proc. SIGGRAPH Asia, 2022. 3 +[18] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5 +[19] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 5 +[20] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674, 2020. 3 +[21] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 3 +[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 3 +[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3 +[24] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. ACM Transactions on Graphics (TOG), Siggraph Asia 2019, 38(6):Article 242, 2019. 3 +[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 5 +[26] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, pages 8280-8290, 2022. 2, 5 +[27] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised + +learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. 3 +[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[29] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3 +[30] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 3 +[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3, 4, 5, 6 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 3 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 3, 4, 8 +[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3, 7 +[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 3, 4 +[37] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 3 +[38] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 +[39] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5 + +[40] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In Conference on Robot Learning, pages 87-96. PMLR, 2017. 3 +[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1 +[42] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 1 +[43] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 5 +[44] Towaki Takikawa, Or Perel, Clement Fuji Tsang, Charles Loop, Joey Litalien, Jonathan Tremblay, Sanja Fidler, and Maria Shugrina. Kaolin wisp: A pytorch library and engine for neural fields research. https://github.com/NVIDIAGameWorks/kaolin-wisp, 2022.5 +[45] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3 +[46] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 3 +[47] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3 +[48] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yinan Zhang, Antonio Torralba, and Sanja Fidler. Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In International Conference on Learning Representations, 2021. 3 +[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826-5835, 2021. 3 \ No newline at end of file diff --git a/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/images.zip b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7aa5c8467db8aef35860717efbc058304ac3125e --- /dev/null +++ b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d526e0bbd439695cfe60fd8cd507238e84910fa595fcb29d850477957ae33a8e +size 741815 diff --git a/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/layout.json b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0d58c2a4e587384f1707f7b7f598ec4e9b87a863 --- /dev/null +++ b/2023/Magic3D_ High-Resolution Text-to-3D Content Creation/layout.json @@ -0,0 +1,8003 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 471, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 471, + 120 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 471, + 120 + ], + "type": "text", + "content": "Magic3D: High-Resolution Text-to-3D Content Creation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 143, + 506, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 143, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 86, + 143, + 506, + 172 + ], + "type": "text", + "content": "Chen-Hsuan Lin* Jun Gao* Luming Tang* Towaki Takikawa* Xiaohui Zeng* Xun Huang Karsten Kreis Sanja Fidler† Ming-Yu Liu† Tsung-Yi Lin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 177, + 349, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 177, + 349, + 191 + ], + "spans": [ + { + "bbox": [ + 241, + 177, + 349, + 191 + ], + "type": "text", + "content": "NVIDIA Corporation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 183, + 194, + 406, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 194, + 406, + 206 + ], + "spans": [ + { + "bbox": [ + 183, + 194, + 406, + 206 + ], + "type": "text", + "content": "https://research.nvidia.com/labs/dir/magic3d" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 235, + 192, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 235, + 192, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 235, + 192, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 259, + 290, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 290, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 290, + 523 + ], + "type": "text", + "content": "DreamFusion [31] has recently demonstrated the utility of a pre-trained text-to-image diffusion model to optimize Neural Radiance Fields (NeRF) [23], achieving remarkable text-to-3D synthesis results. However, the method has two inherent limitations: (a) extremely slow optimization of NeRF and (b) low-resolution image space supervision on NeRF, leading to low-quality 3D models with a long processing time. In this paper, we address these limitations by utilizing a two-stage optimization framework. First, we obtain a coarse model using a low-resolution diffusion prior and accelerate with a sparse 3D hash grid structure. Using the coarse representation as the initialization, we further optimize a textured 3D mesh model with an efficient differentiable renderer interacting with a high-resolution latent diffusion model. Our method, dubbed Magic3D, can create high quality 3D mesh models in 40 minutes, which is " + }, + { + "bbox": [ + 47, + 259, + 290, + 523 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 47, + 259, + 290, + 523 + ], + "type": "text", + "content": " faster than DreamFusion (reportedly taking 1.5 hours on average), while also achieving higher resolution. User studies show " + }, + { + "bbox": [ + 47, + 259, + 290, + 523 + ], + "type": "inline_equation", + "content": "61.7\\%" + }, + { + "bbox": [ + 47, + 259, + 290, + 523 + ], + "type": "text", + "content": " raters to prefer our approach over DreamFusion. Together with the image-conditioned generation capabilities, we provide users with new ways to control 3D synthesis, opening up new avenues to various creative applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 564, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 288, + 696 + ], + "type": "text", + "content": "3D digital content has been in high demand for a variety of applications, including gaming, entertainment, architecture, and robotics simulation. It is slowly finding its way into virtually every possible domain: retail, online conferencing, virtual social presence, education, etc. However, creating professional 3D content is not for anyone — it requires immense artistic and aesthetic training with 3D modeling expertise. Developing these skill sets takes a significant amount of time and effort. Augmenting 3D content creation with natural language could considerably help democratize 3D content creation for novices and turbocharge expert artists." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 236, + 547, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 547, + 451 + ], + "type": "text", + "content": "Image content creation from text prompts [2, 28, 33, 36] has seen significant progress with the advances of diffusion models [13, 41, 42] for generative modeling of images. The key enablers are large-scale datasets comprising billions of samples (images with text) scrapped from the Internet and massive amounts of compute. In contrast, 3D content generation has progressed at a much slower pace. Existing 3D object generation models [4, 9, 47] are mostly categorical. A trained model can only be used to synthesize objects for a single class, with early signs of scaling to multiple classes shown recently by Zeng et al. [47]. Therefore, what a user can do with these models is extremely limited and not yet ready for artistic creation. This limitation is largely due to the lack of diverse large-scale 3D datasets — compared to image and video content, 3D content is much less accessible on the Internet. This naturally raises the question of whether 3D generation capability can be achieved by leveraging powerful text-to-image generative models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 456, + 548, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 456, + 548, + 673 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 548, + 673 + ], + "type": "text", + "content": "Recently, DreamFusion [31] demonstrated its remarkable ability for text-conditioned 3D content generation by utilizing a pre-trained text-to-image diffusion model [36] that generates images as a strong image prior. The diffusion model acts as a critic to optimize the underlying 3D representation. The optimization process ensures that rendered images from a 3D model, represented by Neural Radiance Fields (NeRF) [23], match the distribution of photorealistic images across different viewpoints, given the input text prompt. Since the supervision signal in DreamFusion operates on very low-resolution images " + }, + { + "bbox": [ + 305, + 456, + 548, + 673 + ], + "type": "inline_equation", + "content": "(64 \\times 64)" + }, + { + "bbox": [ + 305, + 456, + 548, + 673 + ], + "type": "text", + "content": ", DreamFusion cannot synthesize high-frequency 3D geometric and texture details. Due to the use of inefficient MLP architectures for the NeRF representation, practical high-resolution synthesis may not even be possible as the required memory footprint and the computation budget grows quickly with the resolution. Even at a resolution of " + }, + { + "bbox": [ + 305, + 456, + 548, + 673 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 305, + 456, + 548, + 673 + ], + "type": "text", + "content": ", optimization times are in hours (1.5 hours per prompt on average using TPUv4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 548, + 713 + ], + "type": "text", + "content": "In this paper, we present a method that can synthesize highly detailed 3D models from text prompts within a reduced computation time. Specifically, we propose a coarse-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 702, + 129, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 129, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 129, + 712 + ], + "type": "inline_equation", + "content": "^{\\ast \\dagger}" + }, + { + "bbox": [ + 55, + 702, + 129, + 712 + ], + "type": "text", + "content": ": equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "300" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 106, + 192, + 223 + ], + "blocks": [ + { + "bbox": [ + 50, + 106, + 192, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 106, + 192, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 106, + 192, + 223 + ], + "type": "image", + "image_path": "d6080e0d70aab5a7349d9ec09e3d7cd274505bee320234f86bbb32c16dd37c18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 201, + 72, + 264, + 251 + ], + "blocks": [ + { + "bbox": [ + 201, + 72, + 264, + 251 + ], + "lines": [ + { + "bbox": [ + 201, + 72, + 264, + 251 + ], + "spans": [ + { + "bbox": [ + 201, + 72, + 264, + 251 + ], + "type": "image", + "image_path": "5b8245c2c17cd88f41899b3052ead17133ef619ff7ad49ee3c13a8bd95c7d5f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 251, + 298, + 271 + ], + "lines": [ + { + "bbox": [ + 184, + 251, + 298, + 271 + ], + "spans": [ + { + "bbox": [ + 184, + 251, + 298, + 271 + ], + "type": "text", + "content": "michelangelo style statue of an astronaut" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 287, + 78, + 381, + 245 + ], + "blocks": [ + { + "bbox": [ + 287, + 78, + 381, + 245 + ], + "lines": [ + { + "bbox": [ + 287, + 78, + 381, + 245 + ], + "spans": [ + { + "bbox": [ + 287, + 78, + 381, + 245 + ], + "type": "image", + "image_path": "0c71b85b08404d1ddfe504a5f8c22cafa8c3d86eb243d74b9b468882241275af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 251, + 402, + 274 + ], + "lines": [ + { + "bbox": [ + 304, + 251, + 402, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 402, + 274 + ], + "type": "text", + "content": "a stuffed grey rabbit holding a pretend carrot" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 122, + 547, + 239 + ], + "blocks": [ + { + "bbox": [ + 427, + 122, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 427, + 122, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 427, + 122, + 547, + 239 + ], + "type": "image", + "image_path": "6f428019d63476ad15c8772a3814a6a3c272bd31916bb5ec79bb7bdf370567ba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 245, + 533, + 255 + ], + "lines": [ + { + "bbox": [ + 425, + 245, + 533, + 255 + ], + "spans": [ + { + "bbox": [ + 425, + 245, + 533, + 255 + ], + "type": "text", + "content": "an iguana holding a balloon" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 261, + 127, + 403 + ], + "blocks": [ + { + "bbox": [ + 80, + 236, + 160, + 258 + ], + "lines": [ + { + "bbox": [ + 80, + 236, + 160, + 258 + ], + "spans": [ + { + "bbox": [ + 80, + 236, + 160, + 258 + ], + "type": "text", + "content": "a silver platter piled high with fruits" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 261, + 127, + 403 + ], + "lines": [ + { + "bbox": [ + 56, + 261, + 127, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 127, + 403 + ], + "type": "image", + "image_path": "2f151281677b782a6d0bf393a5109d239305e750e398997cbc0752d41bb9dc72.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 139, + 276, + 242, + 404 + ], + "blocks": [ + { + "bbox": [ + 54, + 405, + 145, + 427 + ], + "lines": [ + { + "bbox": [ + 54, + 405, + 145, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 405, + 145, + 427 + ], + "type": "text", + "content": "a beautiful dress made out of garbage bags" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 276, + 242, + 404 + ], + "lines": [ + { + "bbox": [ + 139, + 276, + 242, + 404 + ], + "spans": [ + { + "bbox": [ + 139, + 276, + 242, + 404 + ], + "type": "image", + "image_path": "fc43f3109dee508f9f835868839a0bcf438823adfcb214d1494dc0733f869f37.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 406, + 231, + 427 + ], + "lines": [ + { + "bbox": [ + 161, + 406, + 231, + 427 + ], + "spans": [ + { + "bbox": [ + 161, + 406, + 231, + 427 + ], + "type": "text", + "content": "an imperial state crown of england" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 249, + 289, + 382, + 385 + ], + "blocks": [ + { + "bbox": [ + 249, + 289, + 382, + 385 + ], + "lines": [ + { + "bbox": [ + 249, + 289, + 382, + 385 + ], + "spans": [ + { + "bbox": [ + 249, + 289, + 382, + 385 + ], + "type": "image", + "image_path": "829b8fdf0d1f6df006853c44bda46266ed7158c5c1291ff309f56b611df95d09.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 281, + 545, + 411 + ], + "blocks": [ + { + "bbox": [ + 389, + 281, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 389, + 281, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 389, + 281, + 545, + 411 + ], + "type": "image", + "image_path": "3dc0649342845e4741b346842db3a6a860a0577be4336ef28ba4362f6d77ea11.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 415, + 531, + 425 + ], + "lines": [ + { + "bbox": [ + 394, + 415, + 531, + 425 + ], + "spans": [ + { + "bbox": [ + 394, + 415, + 531, + 425 + ], + "type": "text", + "content": "neuschwanstein castle, aerial view" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 441, + 148, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 441, + 148, + 463 + ], + "spans": [ + { + "bbox": [ + 59, + 441, + 148, + 463 + ], + "type": "text", + "content": "Low resolution bunny before editing" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 58, + 471, + 113, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 471, + 113, + 523 + ], + "spans": [ + { + "bbox": [ + 58, + 471, + 113, + 523 + ], + "type": "text", + "content": "a baby bunny sitting on top of a stack of pancakes" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 107, + 472, + 149, + 521 + ], + "blocks": [ + { + "bbox": [ + 107, + 472, + 149, + 521 + ], + "lines": [ + { + "bbox": [ + 107, + 472, + 149, + 521 + ], + "spans": [ + { + "bbox": [ + 107, + 472, + 149, + 521 + ], + "type": "image", + "image_path": "bca0e717bd1f384e994c7764818373ae4f37872048de5e7149cf01c0ebb22fc8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 160, + 434, + 196, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 434, + 196, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 434, + 196, + 495 + ], + "type": "text", + "content": "a metal bunny sitting on top of a stack of broccoli" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 192, + 432, + 276, + 529 + ], + "blocks": [ + { + "bbox": [ + 269, + 404, + 362, + 427 + ], + "lines": [ + { + "bbox": [ + 269, + 404, + 362, + 427 + ], + "spans": [ + { + "bbox": [ + 269, + 404, + 362, + 427 + ], + "type": "text", + "content": "a blue poison-dart frog sitting on a water lily" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 432, + 276, + 529 + ], + "lines": [ + { + "bbox": [ + 192, + 432, + 276, + 529 + ], + "spans": [ + { + "bbox": [ + 192, + 432, + 276, + 529 + ], + "type": "image", + "image_path": "52e09933ed0dc7046df0273ae8facc1f4ac775854703f1ea20cee58d3c042ae0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 542, + 545, + 597 + ], + "lines": [ + { + "bbox": [ + 46, + 542, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 545, + 597 + ], + "type": "text", + "content": "Figure 1. Results and applications of Magic3D. Top: high-resolution text-to-3D generation. Magic3D can generate high-quality and high-resolution 3D models from text prompts. Bottom: high-resolution prompt-based editing. Magic3D can edit 3D models by fine-tuning with the diffusion prior using a different prompt. Taking the low-resolution 3D model as the input (left), Magic3D can modify different parts of the 3D model corresponding to different input text prompts. Together with various creative controls on the generated 3D models, Magic3D is a convenient tool for augmenting 3D content creation." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 288, + 434, + 327, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 434, + 327, + 505 + ], + "spans": [ + { + "bbox": [ + 288, + 434, + 327, + 505 + ], + "type": "text", + "content": "a metalbunnysitting ontop of asstack ofchocolatecookie" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 324, + 432, + 403, + 529 + ], + "blocks": [ + { + "bbox": [ + 324, + 432, + 403, + 529 + ], + "lines": [ + { + "bbox": [ + 324, + 432, + 403, + 529 + ], + "spans": [ + { + "bbox": [ + 324, + 432, + 403, + 529 + ], + "type": "image", + "image_path": "e34718cd295dcc17dc3640658aa58a4c5c84c0ff911be780ac3087f184c6f636.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 417, + 434, + 459, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 434, + 459, + 496 + ], + "spans": [ + { + "bbox": [ + 417, + 434, + 459, + 496 + ], + "type": "text", + "content": "a sphinx sitting on top of a stack of chocolate cookie" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 451, + 431, + 527, + 529 + ], + "blocks": [ + { + "bbox": [ + 451, + 431, + 527, + 529 + ], + "lines": [ + { + "bbox": [ + 451, + 431, + 527, + 529 + ], + "spans": [ + { + "bbox": [ + 451, + 431, + 527, + 529 + ], + "type": "image", + "image_path": "81d041100aaeb45340839b4086dd5e57d9051d41e929f88339d3842fb4318da0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 618, + 289, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 703 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 703 + ], + "type": "text", + "content": "to-fine optimization approach that uses multiple diffusion priors at different resolutions to optimize the 3D representation, enabling the generation of both view-consistent geometry as well as high-resolution details. In the first stage, we optimize a coarse neural field representation akin to Dream-Fusion, but with a memory- and compute-efficient scene representation based on a hash grid [25]. In the second stage," + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 618, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 703 + ], + "type": "text", + "content": "we switch to optimizing mesh representations, a critical step that allows us to utilize diffusion priors at resolutions as high as " + }, + { + "bbox": [ + 304, + 618, + 547, + 703 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 618, + 547, + 703 + ], + "type": "text", + "content": ". As 3D meshes are amenable to fast graphics renderers that can render high-resolution images in real-time, we leverage an efficient differentiable rasterizer [9, 26] and make use of camera close-ups to recover high-frequency details in geometry and texture. As a result, our approach" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "301" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "produces high-fidelity 3D content (see Fig. 1) that can conveniently be imported and visualized in standard graphics software and does so at " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " the speed of DreamFusion. Furthermore, we showcase various creative controls over the 3D synthesis process by leveraging the advancements developed for text-to-image editing applications [2, 35]. Our approach, dubbed Magic3D, endows users with unprecedented control in crafting their desired 3D objects with text prompts and reference images, bringing this technology one step closer to democratizing 3D content creation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 193, + 288, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 193, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 288, + 204 + ], + "type": "text", + "content": "In summary, our work makes the following contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 213, + 289, + 361 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 213, + 289, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 289, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 289, + 321 + ], + "type": "text", + "content": "- We propose Magic3D, a framework for high-quality 3D content synthesis using text prompts by improving several major design choices made in DreamFusion. It consists of a coarse-to-fine strategy that leverages both low- and high-resolution diffusion priors for learning the 3D representation of the target content. Magic3D, which synthesizes 3D content with an " + }, + { + "bbox": [ + 47, + 213, + 289, + 321 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 47, + 213, + 289, + 321 + ], + "type": "text", + "content": " higher resolution supervision, is also " + }, + { + "bbox": [ + 47, + 213, + 289, + 321 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 47, + 213, + 289, + 321 + ], + "type": "text", + "content": " faster than DreamFusion. 3D content synthesized by our approach is significantly preferable by users (61.7%)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 326, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 288, + 361 + ], + "type": "text", + "content": "- We extend various image editing techniques developed for text-to-image models to 3D object editing and show their applications in the proposed framework." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 373, + 134, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 134, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 134, + 385 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 398, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 289, + 579 + ], + "type": "text", + "content": "Text-to-image generation. We have witnessed significant progress in text-to-image generation with diffusion models in recent years. With improvements in modeling and data curation, diffusion models can compose complex semantic concepts from text descriptions (nouns, adjectives, artistic styles, etc.) to generate high-quality images of objects and scenes [2, 33, 34, 36]. Sampling images from diffusion models is time consuming. To generate high-resolution images, these models either utilize a cascade of super-resolution models [2, 36] or sample from a lower-resolution latent space and decode latents into high-resolution images [34]. Despite the advances in high-resolution image generation, using language to describe and control 3D properties (e.g. camera viewpoints) while maintaining coherency in 3D remains an open, challenging problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "3D generative models. There is a large body of work on 3D generative modeling, exploring different types of 3D representations such as 3D voxel grids [7, 12, 20, 40, 45], pointclouds [1, 21, 24, 46, 47, 49], meshes [9, 48], implicit [6, 22], or octree [15] representations. Most of these approaches rely on training data in the form of 3D assets, which are hard to acquire at scale. Inspired by the success of neural volume rendering [23], recent works started investing in 3D-aware image synthesis [4, 5, 10, 11, 27, 29, 30, 38], which has the advantage of learning 3D generative models directly from images — a more widely accessible resource. However, vol-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "ume rendering networks are typically slow to query, leading to a trade-off between long training time [5, 29] and lack of multi-view consistency [10]. EG3D [4] partially mitigates this problem by utilizing a dual discriminator. While obtaining promising results, these works remain limited to modeling objects within a single object category, such as cars, chairs, or human faces, thus lacking scalability and the creative control desired for 3D content creation. In our paper, we focus on text-to-3D synthesis, aiming to generate a 3D renderable representation of a scene based on a text prompt." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 196, + 548, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 548, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 548, + 376 + ], + "type": "text", + "content": "Text-to-3D generation. With the recent success in text-to-image generative modeling in recent years, text-to-3D generation has also gained a surge of interest from the learning community. Earlier works such as CLIP-forge [37] synthesize objects by learning a normalizing flow model to sample shape embeddings from textual input. However, it requires 3D assets in voxel representations during training, making it challenging to scale with data. DreamField [16] and CLIP-mesh [17] mitigate the training data issue by relying on a pre-trained image-text model [32] to optimize the underlying 3D representations (NeRFs and meshes), such that all 2D renderings reach high text-image alignment scores. While these approaches avoid the requirement of expensive 3D training data and mostly rely on pre-trained large-scale image-text models, they tend to produce less realistic 2D renderings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "text", + "content": "Recently, DreamFusion [31] showcased impressive capability in text-to-3D synthesis by utilizing a powerful pretrained text-to-image diffusion model [36] as a strong image prior. We build upon this work and improve over several design choices to bring significantly higher-fidelity 3D models into hands of users with a much reduced generation time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 456, + 462, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 456, + 462, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 462, + 470 + ], + "type": "text", + "content": "3. Background: DreamFusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": "DreamFusion [31] achieves text-to-3D generation with two key components: a neural scene representation which we refer to as the scene model, and a pre-trained text-to-image diffusion-based generative model. The scene model is a parametric function " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "x = g(\\theta)" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": ", which can produce an image " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": " at the desired camera pose. Here, " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": " is a volumetric renderer of choice, and " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": " is a coordinate-based MLP representing a 3D volume. The diffusion model " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": " comes with a learned denoising function " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(x_t; y, t)" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": " that predicts the sampled noise " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": " given the noisy image " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": ", noise level " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": ", and text embedding " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": ". It provides the gradient direction to update " + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 475, + 547, + 656 + ], + "type": "text", + "content": " such that all rendered images are pushed to the high probability density regions conditioned on the text embedding under the diffusion prior. Specifically, DreamFusion introduces Score Distillation Sampling (SDS), which computes the gradient:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 658, + 547, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 547, + 687 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , g (\\theta)) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(x _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {1}", + "image_path": "077c3ed99521940b7a2929d64fef41d1a2e256e06292348fa8a29d61608cf3b8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": " is a weighting function. We view the scene model " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": " and the diffusion model " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": " as modular components of the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "302" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 547, + 219 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 547, + 219 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 547, + 219 + ], + "type": "image", + "image_path": "993ed516c2a474d9377e1c231fe571903233398327ee30a97216a721d5b9ac10.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 546, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 546, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 546, + 275 + ], + "type": "text", + "content": "Figure 2. Overview of Magic3D. We generate high-resolution 3D content from an input text prompt in a coarse-to-fine manner. In the first stage, we utilize a low-resolution diffusion prior and optimize neural field representations (color, density, and normal fields) to obtain the coarse model. We further differentiably extract textured 3D mesh from the density and color fields of the coarse model. Then we fine-tune it using a high-resolution latent diffusion model. After optimization, our model generates high-quality 3D meshes with detailed textures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 297, + 288, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 367 + ], + "type": "text", + "content": "framework, amenable to choice. In practice, the denoising function " + }, + { + "bbox": [ + 46, + 297, + 288, + 367 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 46, + 297, + 288, + 367 + ], + "type": "text", + "content": " is often replaced with another function " + }, + { + "bbox": [ + 46, + 297, + 288, + 367 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_{\\phi}" + }, + { + "bbox": [ + 46, + 297, + 288, + 367 + ], + "type": "text", + "content": " that uses classifier-free guidance [14], which allows one to carefully weigh the strength of the text conditioning (see Sec. 6). DreamFusion relies on large classifier-free guidance weights to obtain results with better quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 368, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 477 + ], + "type": "text", + "content": "DreamFusion adopts a variant of Mip-NeRF 360 [3] with an explicit shading model for the scene model and Imagegen [36] as the diffusion model. These choices result in two key limitations. First, high-resolution geometry or textures cannot be obtained since the diffusion model only operates on " + }, + { + "bbox": [ + 46, + 368, + 287, + 477 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 368, + 287, + 477 + ], + "type": "text", + "content": " images. Second, the utility of a large global MLP for volume rendering is both computationally expensive as well as memory intensive, making this approach scale poorly with the increasing resolution of images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 487, + 224, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 224, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 224, + 500 + ], + "type": "text", + "content": "4. High-Resolution 3D Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 507, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 288, + 555 + ], + "type": "text", + "content": "Magic3D is a two-stage coarse-to-fine framework that uses efficient scene models that enable high-resolution text-to-3D synthesis (Fig. 2). We describe our method and key differences from DreamFusion [31] in this section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 214, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 214, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 214, + 574 + ], + "type": "text", + "content": "4.1. Coarse-to-fine Diffusion Priors" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Magic3D uses two different diffusion priors in a coarse-to-fine fashion to generate high-resolution geometry and textures. In the first stage, we use the base diffusion model described in eDiff-I [2], which is similar to the base diffusion model of Imagen [36] used in DreamFusion. This diffusion prior is used to compute gradients of the scene model via a loss defined on rendered images at a low resolution " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ". In the second stage, we use the latent diffusion model (LDM) [34] that allows backpropagating gradients into rendered images at a high resolution " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "; in practice, we choose to use the publicly available Stable Diffusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "type": "text", + "content": "model [34]. Despite generating high-resolution images, the computation of LDM is manageable because the diffusion prior acts on the latent " + }, + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "type": "text", + "content": " with resolution " + }, + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 339, + 545, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 366 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , g (\\theta)) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(z _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial z}{\\partial x} \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {2}", + "image_path": "f38f41d6de29234c4cbe01cb5481ed3ecbc9b91a9c3d4270d3b602c27bf63a82.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 373, + 547, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 409 + ], + "type": "text", + "content": "The increase in computation time mainly comes from computing " + }, + { + "bbox": [ + 304, + 373, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\frac{\\partial x}{\\partial \\theta}" + }, + { + "bbox": [ + 304, + 373, + 547, + 409 + ], + "type": "text", + "content": " (the gradient of the high-resolution rendered image) and " + }, + { + "bbox": [ + 304, + 373, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\frac{\\partial z}{\\partial x}" + }, + { + "bbox": [ + 304, + 373, + 547, + 409 + ], + "type": "text", + "content": " (the gradient of the encoder in LDM)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 416, + 392, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 416, + 392, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 416, + 392, + 427 + ], + "type": "text", + "content": "4.2. Scene Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 434, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 483 + ], + "type": "text", + "content": "We cater two different 3D scene representations to the two different diffusion priors at coarse and fine resolutions to accommodate the increased resolution of rendered images for the input of high-resolution priors, discussed as follows." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 547, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 666 + ], + "type": "text", + "content": "Neural fields as coarse scene models. The initial coarse stage of the optimization requires finding the geometry and textures from scratch. This can be challenging as we need to accommodate complex topological changes in the 3D geometry and depth ambiguities from the 2D supervision signals. In DreamFusion [31], the scene model is a neural field (a coordinate-based MLP) based on Mip-NeRF 360 [3] that predicts albedo and density. This is a suitable choice as neural fields can handle topological changes in a smooth, continuous fashion. However, Mip-NeRF 360 [3] is computationally expensive as it is based on a large global coordinate-based MLP. As volume rendering requires dense samples along a ray to accurately render high-frequency geometry and shading, the cost of having to evaluate a large neural network at every sample point quickly stacks up." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "For this reason, we opt to use the hash grid encoding from Instant NGP [25], which allows us to represent high-frequency details at a much lower computational cost. We use the hash grid with two single-layer neural networks," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "303" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "one predicting albedo and density and the other predicting normals. We additionally maintain a spatial data structure that encodes scene occupancy and utilizes empty space skipping [19, 43]. Specifically, we use the density-based voxel pruning approach from Instant NGP [25] with an octree-based ray sampling and rendering algorithm [44]. With these design choices, we drastically accelerate the optimization of coarse scene models while maintaining quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 172, + 289, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 289, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 289, + 293 + ], + "type": "text", + "content": "Textured meshes as fine scene models. In our fine stage of optimization, we need to be able to accommodate very high-resolution rendered images to fine-tune our scene model with high-resolution diffusion priors. Using the same scene representation (the neural field) from the initial coarse stage of optimization could be a natural choice since the weights of the model can directly carry over. Although this strategy can work to some extent (Figs. 4 and 5), it struggles to render very high-resolution (e.g., " + }, + { + "bbox": [ + 46, + 172, + 289, + 293 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 172, + 289, + 293 + ], + "type": "text", + "content": ") images within reasonable memory constraints and computation budgets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 289, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 401 + ], + "type": "text", + "content": "To resolve this issue, we use textured 3D meshes as the scene representation for the fine stage of optimization. In contrast to volume rendering for neural fields, rendering textured meshes with differentiable rasterization can be performed efficiently at very high resolutions, making meshes a suitable choice for our high-resolution optimization stage. Using the neural field from the coarse stage as the initialization for the mesh geometry, we can also sidestep the difficulty of learning large topological changes in meshes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "text", + "content": "Formally, we represent the 3D shape using a deformable tetrahedral grid " + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "inline_equation", + "content": "(V_{T}, T)" + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "inline_equation", + "content": "V_{T}" + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "text", + "content": " is the vertices in the grid " + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "text", + "content": " [8, 39]. Each vertex " + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i} \\in V_{T} \\subset \\mathbb{R}^{3}" + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "text", + "content": " contains a signed distance field (SDF) value " + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "inline_equation", + "content": "s_{i} \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "text", + "content": " and a deformation " + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{v}_{i} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 46, + 402, + 289, + 498 + ], + "type": "text", + "content": " of the vertex from its initial canonical coordinate. Then, we extract a surface mesh from the SDF using a differentiable marching tetrahedra algorithm [39]. For textures, we use the neural color field as a volumetric texture representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 509, + 201, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 201, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 201, + 521 + ], + "type": "text", + "content": "4.3. Coarse-to-fine Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 528, + 288, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 288, + 564 + ], + "type": "text", + "content": "We describe our coarse-to-fine optimization procedure, which first operates on a coarse neural field representation and subsequently a high-resolution textured mesh." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 653 + ], + "type": "text", + "content": "Neural field optimization. Similarly to Instant NGP [25], we initialize an occupancy grid of resolution " + }, + { + "bbox": [ + 46, + 568, + 288, + 653 + ], + "type": "inline_equation", + "content": "256^3" + }, + { + "bbox": [ + 46, + 568, + 288, + 653 + ], + "type": "text", + "content": " with values to 20 to encourage shapes to grow in the early stages of optimization. We update the grid every 10 iterations and generate an octree for empty space skipping. We decay the occupancy grid by 0.6 in every update and follow Instant NGP with the same update and thresholding parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Instead of estimating normals from density differences, we use an MLP to predict the normals. Note that this does not violate geometric properties since volume rendering is used instead of surface rendering; as such, the orientation of particles at continuous positions need not be oriented" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "to the level set surface. This helps us significantly reduce the computational cost of optimizing the coarse model by avoiding the use of finite differencing. Accurate normals can be obtained in the fine stage of optimization when we use a true surface rendering model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 252 + ], + "type": "text", + "content": "Similar to DreamFusion, we also model the background using an environment map MLP, which predicts RGB colors as a function of ray directions. Since our sparse representation model does not support scene reparametrization as in Mip-NeRF 360 [3], the optimization has a tendency to \"cheat\" by learning the essence of the object using the background environment map. As such, we use a tiny MLP for the environment map (hidden dimension size of 16) and weigh down the learning rate by " + }, + { + "bbox": [ + 304, + 133, + 547, + 252 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 304, + 133, + 547, + 252 + ], + "type": "text", + "content": " to allow the model to focus more on the neural field geometry." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 547, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 316 + ], + "type": "text", + "content": "Mesh optimization. To optimize a mesh from the neural field initialization, we convert the (coarse) density field to an SDF by subtracting it with a non-zero constant, yielding the initial " + }, + { + "bbox": [ + 304, + 255, + 547, + 316 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 255, + 547, + 316 + ], + "type": "text", + "content": ". We additionally initialize the volume texture field directly with the color field optimized from the coarse stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": "During optimization, we render the extracted surface mesh into high-resolution images using a differentiable rasterizer [18,26]. We optimize both " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{v}_i" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": " for each vertex " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": " via backpropagation using the high-resolution SDS gradient (Eq. 2). When rendering the mesh to an image, we also track the 3D coordinates of each corresponding pixel projection, which would be used to query colors in the corresponding texture field for joint optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 411, + 548, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 548, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 548, + 544 + ], + "type": "text", + "content": "When rendering the mesh, we increase the focal length to zoom in on object details, which is a critical step towards recovering high-frequency details. We keep the same pre-trained environment map from the coarse stage of optimization and composite the rendered background with the rendered foreground object using differentiable antialIASing [18]. To encourage the smoothness of the surface, we further regularize the angular differences between adjacent faces on the mesh. This allows us to obtain well-behaved geometry even under supervision signals with high variance, such as the SDS gradient " + }, + { + "bbox": [ + 304, + 411, + 548, + 544 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}_{\\mathrm{SDS}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 552, + 388, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 388, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 388, + 567 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 572, + 547, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 547, + 621 + ], + "type": "text", + "content": "We focus on comparing our method with DreamFusion [31] on 397 text prompts taken from the website of DreamFusion1. We train Magic3D on all of the text prompts and compare them with the results provided on the website." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 624, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 547, + 696 + ], + "type": "text", + "content": "Speed evaluation. Unless otherwise noted, the coarse stage is trained for 5000 iterations with 1024 samples along the ray (subsequently filtered by the sparse octree) with a batch size of 32, with a total runtime of around 15 minutes (upwards of 8 iterations / second, variable due to differences in sparsity). The fine stage is trained for 3000 iterations with a batch" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 499, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 499, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 499, + 713 + ], + "type": "text", + "content": "ComparisonPreferenceMagic3D vs. DreamFusion [31]• More realistic58.3%• More detailed66.0%• More realistic & detailed61.7%Magic3D vs. Magic3D (coarse only)87.7%", + "image_path": "98d6ad8c88e4bf2e58d85478e1756895f1c8ab137c8134833522569bb0f75489.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 588, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 288, + 624 + ], + "type": "text", + "content": "size of 32 with a total runtime of 25 minutes (2 iterations / second). Both runtimes combined are 40 minutes. All runtimes were measured on 8 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "Qualitative comparisons. We provide qualitative examples in Fig. 3. Qualitatively, our models achieve much higher 3D quality in terms of both geometry and texture. Notice that our model can generate candies on ice cream cones, highly detailed sushi-like cars, vivid strawberries, and birds. We also note that our resulting 3D models can be directly imported and visualized in standard graphics software." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 339, + 409, + 511, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 409, + 511, + 419 + ], + "spans": [ + { + "bbox": [ + 339, + 409, + 511, + 419 + ], + "type": "text", + "content": "a baby bunny sitting on top of a stack of pancakes†" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 419, + 533, + 480 + ], + "blocks": [ + { + "bbox": [ + 310, + 419, + 533, + 480 + ], + "lines": [ + { + "bbox": [ + 310, + 419, + 533, + 480 + ], + "spans": [ + { + "bbox": [ + 310, + 419, + 533, + 480 + ], + "type": "image", + "image_path": "43eb0c7a0ff9d607f3f4eaefc8982db71414077c938d07a21007740ab1d9e9c7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 481, + 465, + 491 + ], + "lines": [ + { + "bbox": [ + 388, + 481, + 465, + 491 + ], + "spans": [ + { + "bbox": [ + 388, + 481, + 465, + 491 + ], + "type": "text", + "content": "(a) Single-stage model" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 493, + 542, + 555 + ], + "blocks": [ + { + "bbox": [ + 312, + 493, + 542, + 555 + ], + "lines": [ + { + "bbox": [ + 312, + 493, + 542, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 493, + 542, + 555 + ], + "type": "image", + "image_path": "5c0414b0340731f072ebd6195646369ba0ae237f5b9e657f726f08da5c1a48b6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 555, + 468, + 565 + ], + "lines": [ + { + "bbox": [ + 384, + 555, + 468, + 565 + ], + "spans": [ + { + "bbox": [ + 384, + 555, + 468, + 565 + ], + "type": "text", + "content": "(b) Coarse-to-fine model" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 575, + 547, + 632 + ], + "lines": [ + { + "bbox": [ + 304, + 575, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 547, + 632 + ], + "type": "text", + "content": "Figure 4. Single-stage (top) vs. coarse-to-fine models (bottom). Both use NeRF for the scene model. The left two columns use " + }, + { + "bbox": [ + 304, + 575, + 547, + 632 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 575, + 547, + 632 + ], + "type": "text", + "content": " rendering resolution during optimization while the right two use " + }, + { + "bbox": [ + 304, + 575, + 547, + 632 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 575, + 547, + 632 + ], + "type": "text", + "content": ". Compared to our coarse-to-fine approach, the single-stage method can generate details but with worse shapes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 548, + 713 + ], + "type": "text", + "content": "User studies. We conduct user studies to evaluate different methods based on user preferences on Amazon MTurk. We show users two videos side by side rendered from a canonical view by two different algorithms using the same text prompt. We ask the users to select the one that is more realistic" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "305" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 221 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 221 + ], + "type": "image", + "image_path": "2128d779f88b92785de3b96998a0cb96af27762b568b0b9fe1c5495676770b0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 548, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 548, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 548, + 253 + ], + "type": "text", + "content": "Figure 5. Ablation on the fine-tuning stage. For each text prompt, we compare coarse and fine models with mesh and NeRF representations. Mesh fine-tuning significantly improve the visual quality of generated 3D assets, providing more photo-realistic details on the 3D shapes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 272, + 290, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 290, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 290, + 321 + ], + "type": "text", + "content": "and detailed. Each prompt is evaluated by 3 different users, resulting in 1191 pairwise comparisons. As shown in Table 1, users favor 3D models generated by Magic3D, with " + }, + { + "bbox": [ + 46, + 272, + 290, + 321 + ], + "type": "inline_equation", + "content": "61.7\\%" + }, + { + "bbox": [ + 46, + 272, + 290, + 321 + ], + "type": "text", + "content": " of the users considering our results with higher quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "text", + "content": "Can single-stage optimization work with LDM prior? We ablate scene models optimized with high-resolution LDM prior in a single-stage optimization setup. We find that 3D meshes as the scene model fail to generate high-quality results if optimized from scratch. This leaves our memory-efficient sparse 3D representation as the ideal candidate for the scene model. However, rendering " + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "text", + "content": " images is still too memory intensive to fit into modern GPUs. Therefore, we render lower-resolution images from the scene model and upsample them to " + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "text", + "content": " as input to the LDM. We find it generates objects with worse shapes. Fig. 4 shows two examples with scene rendering resolution " + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 335, + 290, + 504 + ], + "type": "text", + "content": " respectively (top row). While it generates furry details, the shape is worse than the coarse model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 518, + 289, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 289, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 289, + 627 + ], + "type": "text", + "content": "Can we use NeRF for the fine model? Yes. While optimizing a NeRF from scratch does not work well, we can follow the coarse-to-fine framework but replace the second-stage scene model with a NeRF. In the bottom right of Fig. 4, we show the result of a fine NeRF model initialized with the coarse model on its left and fine-tuned with " + }, + { + "bbox": [ + 46, + 518, + 289, + 627 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 518, + 289, + 627 + ], + "type": "text", + "content": " rendered images. The two-stage approach retains good geometry in the initial model and adds more details, showing superior quality to its one-stage counterpart." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Coarse models vs. fine models. Fig. 5 provides more visual results contrasting coarse and fine models. We try both NeRF and mesh for scene models and fine-tune them from the same coarse model above. We see significant quality improvements on both NeRF and mesh models, suggesting our coarse-to-fine approach works for general scene models." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 270, + 545, + 451 + ], + "blocks": [ + { + "bbox": [ + 308, + 270, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 308, + 270, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 308, + 270, + 545, + 451 + ], + "type": "image", + "image_path": "240aac95d822c1e949b296249d85ec84fbd191643e437d3b29c80490eb1a5d38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 458, + 548, + 515 + ], + "lines": [ + { + "bbox": [ + 304, + 458, + 548, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 548, + 515 + ], + "type": "text", + "content": "Figure 6. Magic3D with DreamBooth-based personalization. Given an input image of a particular instance, we fine-tune the diffusion models with DreamBooth and optimize the 3D models with the given prompts. The identity is well preserved in the generated 3D models. Image source (input images): Unsplash." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 540, + 465, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 465, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 465, + 553 + ], + "type": "text", + "content": "6. Controllable 3D Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 563, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 547, + 635 + ], + "type": "text", + "content": "As certain styles and concepts are difficult to express in words but easy with images, it is desirable to have a mechanism to influence the text-to-3D model generation with images. We explore different image conditioning techniques as well as a prompt-based editing approach to provide users more control over the 3D generation outputs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "text", + "content": "Personalized text-to-3D. DreamBooth [35] described a method to personalize text-to-image diffusion models by fine-tuning a pre-trained model on several images of a subject. The fine-tuned model can learn to tie the subject to a unique identifier string, denoted as " + }, + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "inline_equation", + "content": "[\\mathrm{V}]" + }, + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "text", + "content": ", and generate images of the subject when " + }, + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "inline_equation", + "content": "[\\mathrm{V}]" + }, + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "text", + "content": " is included in the text prompt. In" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "306" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 89, + 94, + 152 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 94, + 87 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 94, + 87 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 94, + 87 + ], + "type": "text", + "content": "low-resolution before edited" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 89, + 94, + 152 + ], + "lines": [ + { + "bbox": [ + 49, + 89, + 94, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 89, + 94, + 152 + ], + "type": "image", + "image_path": "397ee7678d79380925b096100a42da2efde6ef375dab86f84cde6612af37bf61.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 107, + 71, + 236, + 159 + ], + "blocks": [ + { + "bbox": [ + 107, + 71, + 236, + 159 + ], + "lines": [ + { + "bbox": [ + 107, + 71, + 236, + 159 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 236, + 159 + ], + "type": "image", + "image_path": "75f10d2c9750fb9b437ace0f89d2872fc7702a5eed360877b1ac99bb2d3dae23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 245, + 72, + 397, + 159 + ], + "blocks": [ + { + "bbox": [ + 245, + 72, + 397, + 159 + ], + "lines": [ + { + "bbox": [ + 245, + 72, + 397, + 159 + ], + "spans": [ + { + "bbox": [ + 245, + 72, + 397, + 159 + ], + "type": "image", + "image_path": "6a2ef0985bfc4aa89b6bd728b0bff926d8c6c7c9964e2007dcf044c8d9c128bd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 405, + 78, + 535, + 159 + ], + "blocks": [ + { + "bbox": [ + 405, + 78, + 535, + 159 + ], + "lines": [ + { + "bbox": [ + 405, + 78, + 535, + 159 + ], + "spans": [ + { + "bbox": [ + 405, + 78, + 535, + 159 + ], + "type": "image", + "image_path": "1be9748663d9b6cdf89903c2cf2a1245fa77ff9eddb78ae5021f3c41cbca3e7a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 160, + 94, + 239 + ], + "blocks": [ + { + "bbox": [ + 48, + 160, + 94, + 239 + ], + "lines": [ + { + "bbox": [ + 48, + 160, + 94, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 160, + 94, + 239 + ], + "type": "image", + "image_path": "015df1f9d4199c2503bd74bce5830ec1a647405ae748aedf2899d50b022abc0d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 106, + 160, + 240, + 246 + ], + "blocks": [ + { + "bbox": [ + 106, + 160, + 240, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 160, + 240, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 240, + 246 + ], + "type": "image", + "image_path": "21a0dcb00660f9a0c9662d005bc7cc172cd9c1dac0732d077d42b09b1c30338d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "type": "text", + "content": "Figure 7. Magic3D with prompt-based editing. Given a coarse model (first column) generated with a base prompt, we replace the underscored text with new text and fine-tune the NeRF to get a high-resolution NeRF model with LDM. We further fine-tune the high-resolution mesh with the NeRF model. Such a prompt-based editing method gives artists better control over the 3D generation output." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 250, + 163, + 392, + 247 + ], + "blocks": [ + { + "bbox": [ + 250, + 163, + 392, + 247 + ], + "lines": [ + { + "bbox": [ + 250, + 163, + 392, + 247 + ], + "spans": [ + { + "bbox": [ + 250, + 163, + 392, + 247 + ], + "type": "image", + "image_path": "7809c54ad73d694302db560d32c30b0831b3652c4f3160efe38b9322efc230a9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 403, + 171, + 531, + 246 + ], + "blocks": [ + { + "bbox": [ + 403, + 171, + 531, + 246 + ], + "lines": [ + { + "bbox": [ + 403, + 171, + 531, + 246 + ], + "spans": [ + { + "bbox": [ + 403, + 171, + 531, + 246 + ], + "type": "image", + "image_path": "ca1779110d89b9f94b2208ce0535fdbacdb5811d5cd1f4a10885b3392f0c3d10.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 306, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 378 + ], + "type": "text", + "content": "the context of text-to-3D generation, we would like to generate a 3D model of a subject. This can be achieved by first fine-tuning our diffusion prior models with the DreamBooth approach, and then using the fine-tuned diffusion priors with the [V] identifier as part of the conditioning text prompt to provide the learning signal when optimizing the 3D model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 382, + 288, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 288, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 288, + 515 + ], + "type": "text", + "content": "To demonstrate the applicability of DreamBooth in our framework, we collect 11 images of one cat and 4 images of one dog. We fine-tune eDiff-I [2] and LDM [34], binding the text identifier [V] to the given subject. Then, we optimize the 3D model with [V] in the text prompts. We use a batch size of 1 for all fine-tuning. For eDiff-I, we use the Adam optimizer with learning rate " + }, + { + "bbox": [ + 46, + 382, + 288, + 515 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 382, + 288, + 515 + ], + "type": "text", + "content": " for 1,500 iterations; for LDM, we fine-tune with learning rate " + }, + { + "bbox": [ + 46, + 382, + 288, + 515 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 46, + 382, + 288, + 515 + ], + "type": "text", + "content": " for 800 iterations. Fig. 6 shows our personalized text-to-3D results: we are able to successfully modify the 3D models preserving the subjects in the given input images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "Prompt-based editing through fine-tuning. Another way to control the generated 3D content is by fine-tuning a learned coarse model with a new prompt. Our prompt-based editing includes three stages. (a) We train a coarse model with a base prompt. (b) We modify the base prompt and fine-tune the coarse model with the LDM. This stage provides a well initialized NeRF model for the next step. Directly applying mesh optimization on a new prompt would generate highly detailed textures but could deform geometry only slightly. (c) We optimize the mesh with the modified text prompt. Our prompt-based editing can modify the texture of the shape or transform the geometry and texture according to the text. The resulting scene models preserve the layer-out and overall structure. Such an editing capability makes the 3D content creation with Magic3D more controllable. In Fig. 7, we show two coarse NeRF models trained with the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 305, + 547, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 547, + 377 + ], + "type": "text", + "content": "base prompt for the \"bunny\" and \"squirrel\". We modify the base prompt, fine-tune the NeRF model in high resolution and optimize the mesh. Results show that we can tune the scene model according to the prompt, e.g. changing the \"baby bunny\" to \"stained glass bunny\" or \"metal bunny\" results in similar geometry but with a different texture." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 408, + 378, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 408, + 378, + 421 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 378, + 421 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "type": "text", + "content": "We propose Magic3D, a fast and high-quality text-to-3D generation framework. We benefit from both efficient scene models and high-resolution diffusion priors in a coarse-to-fine approach. In particular, the 3D mesh models scale nicely with image resolution and enjoy the benefits of higher resolution supervision brought by the latent diffusion model without sacrificing its speed. It takes 40 minutes from a text prompt to a high-quality 3D mesh model ready to be used in graphic engines. With extensive user studies and qualitative comparisons, we show that Magic3D is more preferable " + }, + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "type": "inline_equation", + "content": "(61.7\\%)" + }, + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "type": "text", + "content": " by the raters compared to DreamFusion, while enjoying a " + }, + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "type": "text", + "content": " speed-up. Lastly, we propose a set of tools for better controlling style and content in 3D generation. We hope with Magic3D, we can democratize 3D synthesis and open up everyone's creativity in 3D content creation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgements. We would like to thank Frank Shen, Yogesh Balaji, Seungjun Nah, James Lucas, David Luebke, Clement Fuji-Tsang, Charles Loop, Qinsheng Zhang, Zan Gojcic, and Jonathan Tremblay for helpful discussions and paper proofreading. We would also like to thank Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall for providing additional implementation details in DreamFusion." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "307" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 137 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 139, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 139, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 139, + 288, + 194 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 3, 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 197, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 252 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 254, + 288, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 254, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 53, + 254, + 288, + 320 + ], + "type": "text", + "content": "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 322, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 322, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 322, + 288, + 378 + ], + "type": "text", + "content": "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "text", + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 470 + ], + "type": "text", + "content": "[7] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In 2017 International Conference on 3D Vision (3DV), pages 402-411. IEEE, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 472, + 288, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 472, + 288, + 526 + ], + "spans": [ + { + "bbox": [ + 53, + 472, + 288, + 526 + ], + "type": "text", + "content": "[8] Jun Gao, Wenzheng Chen, Tommy Xiang, Clement Fuji Tsang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. In Advances In Neural Information Processing Systems, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 530, + 288, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 530, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 53, + 530, + 288, + 585 + ], + "type": "text", + "content": "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances in Neural Information Processing Systems, 2022. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 587, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 288, + 632 + ], + "type": "text", + "content": "[10] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "type": "text", + "content": "[11] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds. In ICCV, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[12] Philipp Henzler, Niloy J. Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "type": "text", + "content": "[15] Moritz Ibing, Gregor Kobsik, and Leif Kobbelt. Octree transformer: Autoregressive 3d shape generation on hierarchically structured sequences. arXiv preprint arXiv:2111.12480, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 218 + ], + "type": "text", + "content": "[16] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 547, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 264 + ], + "type": "text", + "content": "[17] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. ACM Transactions on Graphics (TOG), Proc. SIGGRAPH Asia, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 265, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 547, + 308 + ], + "type": "text", + "content": "[18] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 310, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 547, + 354 + ], + "type": "text", + "content": "[19] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 355, + 547, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 547, + 398 + ], + "type": "text", + "content": "[20] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "type": "text", + "content": "[21] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "type": "text", + "content": "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 501, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 547, + 544 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 545, + 547, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 547, + 599 + ], + "type": "text", + "content": "[24] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. ACM Transactions on Graphics (TOG), Siggraph Asia 2019, 38(6):Article 242, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "type": "text", + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "type": "text", + "content": "[26] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, pages 8280-8290, 2022. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 547, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 715 + ], + "type": "text", + "content": "[27] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "308" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "type": "text", + "content": "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "text", + "content": "[29] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 288, + 260 + ], + "type": "text", + "content": "[30] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "type": "text", + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 358 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 3, 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "text", + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 568 + ], + "type": "text", + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "text", + "content": "[37] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "type": "text", + "content": "[38] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[39] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 575 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[40] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In Conference on Robot Learning, pages 87-96. PMLR, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 160 + ], + "type": "text", + "content": "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 195 + ], + "type": "text", + "content": "[42] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 262 + ], + "type": "text", + "content": "[43] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 263, + 547, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 547, + 318 + ], + "type": "text", + "content": "[44] Towaki Takikawa, Or Perel, Clement Fuji Tsang, Charles Loop, Joey Litalien, Jonathan Tremblay, Sanja Fidler, and Maria Shugrina. Kaolin wisp: A pytorch library and engine for neural fields research. https://github.com/NVIDIAGameWorks/kaolin-wisp, 2022.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 319, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 547, + 363 + ], + "type": "text", + "content": "[45] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 364, + 547, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 547, + 418 + ], + "type": "text", + "content": "[46] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 420, + 547, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 547, + 473 + ], + "type": "text", + "content": "[47] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 475, + 547, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 547, + 529 + ], + "type": "text", + "content": "[48] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yinan Zhang, Antonio Torralba, and Sanja Fidler. Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 531, + 547, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 547, + 575 + ], + "type": "text", + "content": "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826-5835, 2021. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "309" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_content_list.json b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89dd114795c777a51ddec148bd2fa7493006d5f7 --- /dev/null +++ b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_content_list.json @@ -0,0 +1,1642 @@ +[ + { + "type": "text", + "text": "MagicNet: Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery", + "text_level": 1, + "bbox": [ + 122, + 128, + 848, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Duowen Chen $^{1}$ Yunhao Bai $^{1}$ Wei Shen $^{2}$ Qingli Li $^{1}$ Lequan Yu $^{3}$ Yan Wang $^{1*}$", + "bbox": [ + 138, + 202, + 834, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University", + "bbox": [ + 143, + 223, + 823, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University $^{3}$ The University of Hong Kong", + "bbox": [ + 99, + 239, + 870, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "duowen_chen@hotmail.com, yhbai@stu.ecnu.edu.cn, wei.shen@sjtu.edu.cn,", + "bbox": [ + 178, + 258, + 785, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qlli@cs.ecnu.edu.cn, lqyu@hku.hk, ywang@cee.ecnu.edu.cn", + "bbox": [ + 240, + 277, + 728, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel teacher-student model for semi-supervised multi-organ segmentation. In teacher-student model, data augmentation is usually adopted on unlabeled data to regularize the consistent training between teacher and student. We start from a key perspective that fixed relative locations and variable sizes of different organs can provide distribution information where a multi-organ CT scan is drawn. Thus, we treat the prior anatomy as a strong tool to guide the data augmentation and reduce the mismatch between labeled and unlabeled images for semi-supervised learning. More specifically, we propose a data augmentation strategy based on partition-and-recovery $N^3$ cubes cross- and within- labeled and unlabeled images. Our strategy encourages unlabeled images to learn organ semantics in relative locations from the labeled images (cross-branch) and enhances the learning ability for small organs (within-branch). For within-branch, we further propose to refine the quality of pseudo labels by blending the learned representations from small cubes to incorporate local attributes. Our method is termed as MagicNet, since it treats the CT volume as a magic-cube and $N^3$ -cube partition-and-recovery process matches with the rule of playing a magic-cube. Extensive experiments on two public CT multi-organ datasets demonstrate the effectiveness of MagicNet, and noticeably outperforms state-of-the-art semi-supervised medical image segmentation approaches, with $+7\\%$ DSC improvement on MACT dataset with $10\\%$ labeled images. Code is available at https://github.com/DeepMed-Lab-ECNU/MagicNet.", + "bbox": [ + 76, + 358, + 473, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 823, + 209, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abdominal multi-organ segmentation in CT images is an essential task in many clinical applications such as", + "bbox": [ + 76, + 849, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e34ebe83cb9744572b429febb7284b1898d6a99a013698954cd35abbd640be3d.jpg", + "image_caption": [ + "Figure 1. Two data augmentation strategies in MagicNet. Left: Although labeled and unlabeled images are not aligned, the latter can be regarded as a shifted version of the former. Co-shift of cubes transfers organ semantics in relative locations from the labeled data to unlabeled data. Right: Segmenting small organs from original images are difficult due to the cluttered background. Small cubes mitigate the impact from background and focus more on local attributes." + ], + "image_footnote": [], + "bbox": [ + 503, + 325, + 888, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "computer-aided intervention [25, 33]. But, training an accurate multi-organ segmentation model usually requires a large amount of labeled data, whose acquisition process is time-consuming and expensive. Semi-supervised learning (SSL) has shown great potential to handle the scarcity of data annotations, which attempts to transfer mass prior knowledge learned from the labeled to unlabeled images. SSL attracts more and more attention in the field of medical image analysis in recent years.", + "bbox": [ + 496, + 623, + 892, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Popular SSL medical image segmentation methods mainly focus on segmenting a single target or targets in a local region, such as segmenting pancreas or left atrium [4, 9, 14, 15, 18, 23, 31, 35, 38, 39]. Multi-organ segmentation is more challenging than single organ segmentation, due to the complex anatomical structures of the organs, e.g., the fixed relative locations (duodenum is always located at the head of the pancreas), the appearances of different organs, and the large variations of the size. Transfer", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 94, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23869", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ring current SSL medical segmentation methods to multi-organ segmentation encounters severe problems. Multiple organs introduce much more variance compared with a single organ. Although labeled and unlabeled images are always drawn from the same distribution, due to the limited number of labeled images, it's hard to estimate the precise distribution from them [32]. Thus, the estimated distribution between labeled and unlabeled images always suffers from mismatch problems, and is even largely increased by multiple organs. Aforementioned SSL medical segmentation methods lack the ability to handle such a large distribution gap, which requires sophisticated anatomical structure modeling. A few semi-supervised multi-organ segmentation methods have been proposed, DMPCT [43] designs a co-training strategy to mine consensus information from multiple views of a CT scan. UMCT [36] further proposes an uncertainty estimation of each view to improve the quality of the pseudo-label. Though these methods take the advantages of multi-view properties in a CT scan, they inevitably ignore the internal anatomical structures of multiple organs, resulting in suboptimal results.", + "bbox": [ + 76, + 90, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Teacher-student model is a widely adopted framework for semi-supervised medical image segmentation [28]. Student network takes labeled images and unlabeled strongly augmented images as input, which attempts to minimize the distribution mismatch between labeled and unlabeled images from the model level. That is, data augmentation is adopted on unlabeled data, whose role is to regularize the consistent training between teacher and student. As mentioned, semi-supervised multi-organ segmentation suffers from large distribution alignment mismatch between labeled and unlabeled images. Reducing the mismatch mainly from the model level is insufficient to solve the problem. Thanks to the prior anatomical knowledge from CT scans, which provides the distribution information where a multi-organ CT scan is drawn, it is possible to largely alleviate the mismatch problem from the data level.", + "bbox": [ + 76, + 412, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel teacher-student model, called MagicNet, matching with the rule of playing a magic-cube. More specifically, we propose a partition-and-recovery $N^3$ cubes learning paradigm: (1) We partition each CT scan, termed as magic-cube, into $N^3$ small cubes. (2) Two data augmentation strategies are then designed, as shown in Fig. 1. I.e., First, to encourage unlabeled data to learn organ semantics in relative locations from the labeled data, small cubes are mixed across labeled and unlabeled images while keeping their relative locations. Second, to enhance the learning ability for small organs, small cubes are shuffled and fed into the student network. (3) We recover the magic-cube to form the original 3D geometry to map with the ground-truth or the supervisory signal from teacher. Furthermore, the quality of pseudo labels predicted by teacher network is refined by blending with the", + "bbox": [ + 76, + 657, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learned representation of the small cubes. The cube-wise pseudo-label blending strategy incorporates local attributes e.g., texture, luster and boundary smoothness which mitigates the inferior performance of small organs.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions can be summarized as follows:", + "bbox": [ + 517, + 152, + 880, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a data augmentation strategy based on partition-and-recovery $N^3$ cubes cross- and within-labeled and unlabeled images which encourages unlabeled images to learn organ semantics in relative locations from the labeled images and enhances the learning ability for small organs.", + "- We propose to correct the original pseudo-label by cube-wise pseudo-label blending via incorporating crucial local attributes for identifying targets especially small organs.", + "- We verify the effectiveness of our method on BTCV [13] and MACT [11] datasets. The segmentation performance of our method exceeds all state-of-the-arts by a large margin, with $7.28\\%$ (10% labeled) and $6.94\\%$ (30% labeled) improvement on two datasets respectively (with V-Net as the backbone) in DSC." + ], + "bbox": [ + 517, + 172, + 890, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 443, + 640, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Semi-supervised Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 500, + 468, + 888, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised medical image segmentation methods can be roughly grouped into three categories. (1) Contrastive learning based methods [35, 38], which learn representations that maximize the similarity among positive pairs and minimize the similarity among negative pairs. (2) Consistency regularization based methods [4, 9, 14, 18, 31, 36, 39], which attend to various levels of information for a single target via multi/dual-task learning or transformation consistent learning. (3) Self-ensembling/self-training based methods [1, 15, 23, 39, 43], which generate pseudolabels for unlabeled images and propose several strategies to ensure the quality of pseudo-labels. But, most of these methods are mainly focusing on segmenting one target or targets in a local region or ROI, such as pancreas or left atrium [18,23,38,39], which encounters performance degradation when transferring to multi-organ segmentation, due to the lack of anatomical structure modeling ability.", + "bbox": [ + 496, + 492, + 890, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Semi-supervised Multi-organ Segmentation", + "text_level": 1, + "bbox": [ + 500, + 756, + 870, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the large variations of appearance and size of different organs [22, 26, 33, 37], multi-organ segmentation has been a popular yet challenging task. Only a few SSL methods specially for multi-organ segmentation have been proposed. DMPCT [43] adopted a 2D-based co-training framework to aggregate multi-planar features on a private dataset. UMCT [36] further enforces multi-view consistency on unlabeled data. These methods are pioneer works", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23870", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9be41dbb465cde9e3166328c83d6be28b0e99d6747036e8759c1ad2ebecf40f6.jpg", + "image_caption": [ + "Figure 2. The architecture of MagicNet. Two essential components are proposed: (1) magic-cube partition and recovery cross- and within-image (see Sec. 3.1 and Fig. 3); (2) cube-wise pseudo-label blending, incorporating local attributes (see Sec. 3.2 and Fig. 4)." + ], + "image_footnote": [], + "bbox": [ + 181, + 87, + 785, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to handle semi-supervised multi-organ segmentation problem from the co-training perspective, which take advantages of the multi-view property of a CT volume, but the properties of multiple organs have not been well-explored.", + "bbox": [ + 75, + 390, + 470, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Interpolation-based Semi-supervised Learning", + "text_level": 1, + "bbox": [ + 76, + 459, + 470, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interpolation-based regularization [30, 40, 41] is quite successful in semi-supervised semantic segmentation. Methods such as Mixup [41] and CutMix [40] are good at synthesizing new training samples and are widely used in semi-supervised learning. FixMatch [24] design a weak-strong pair consistency to simplify semi-supervised learning. MixMatch [3] and ReMixMatch [2] generated pseudo labels for unlabeled data and incorporate gradually unlabeled data with reliable pseudo-label into the labeled set. ICT [30] proposed a interpolated-based consistency training method based on Mixup. GuidedMix-Net [29] utilized Mixup [41] in semi-supervised semantic segmentation. Thus, we also compare our approach with some popular interpolated-based methods in Experiments.", + "bbox": [ + 75, + 482, + 468, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 705, + 168, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define the 3D volume of a CT scan as $\\mathbf{X} \\in \\mathbb{R}^{W \\times H \\times L}$ . The goal is to find the semantic label of each voxel $k \\in \\mathbf{X}$ , which composes into a predicted label map $\\hat{\\mathbf{Y}} \\in \\{0,1,\\dots,C\\}^{W \\times H \\times L}$ , where $C = 0$ indicates background class, and $C \\neq 0$ represents the organ class. The training set $\\mathcal{D}$ consists of two subsets: $\\mathcal{D} = \\mathcal{D}^l \\cup \\mathcal{D}^u$ , where $\\mathcal{D}^l = \\{(\\mathbf{X}_i^l, \\mathbf{Y}_i^l)\\}_{i=1}^N$ and $\\mathcal{D}^u = \\{\\mathbf{X}_i^u\\}_{i=N+1}^{M+N}$ . The training images $\\mathbf{X}^l$ in $\\mathcal{D}^l$ are associated with per-voxel annotations $\\mathbf{Y}^l$ while those in $\\mathcal{D}^u$ are not. In the rest of this paper, we denote the original and mixed CT scans as magic-cubes, and denote the partitioned small cubes as cubes for simplicity.", + "bbox": [ + 75, + 731, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of the proposed MagicNet is shown in Fig. 2, consisting of a student network and a teacher network. The teacher network is updated via weighted combination of the student network parameters with exponential moving average (EMA) [28]. Designed in this architecture, our whole framework includes two branches: Cross-image branch and within-image branch. The details will be illustrated in the following sections.", + "bbox": [ + 496, + 388, + 893, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Magic-cube Partition and Recovery", + "text_level": 1, + "bbox": [ + 498, + 523, + 808, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume that a mini-batch $\\mathcal{B}$ contains $n$ images $\\mathbf{X}^{\\mathcal{B}}\\in$ $\\mathbb{R}^{n\\times W\\times H\\times L}$ . For notational simplicity, let $n = 2$ , and $\\mathcal{B}$ contains one labeled image $\\mathbf{X}^l$ and one unlabeled image $\\mathbf{X}^u$ , which are randomly sampled from $\\mathcal{D}^l$ and $\\mathcal{D}^u$ , respectively. We partition $\\mathbf{X}^l$ and $\\mathbf{X}^u$ into $N^3$ magic-cubes $\\{\\mathbf{x}_j^u\\}_{j = 1}^{N^3}$ and $\\{\\mathbf{x}_j^l\\}_{j = 1}^{N^3}$ , where $\\mathbf{x}_j^l,\\mathbf{x}_j^u\\in \\mathbb{R}^{W / N\\times H / N\\times L / N}$ , $j$ represents the relative location of the cube inside this magic-cube. This partition operation is called as $\\mathcal{A}_{part}(\\cdot)$ .", + "bbox": [ + 496, + 546, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-image Partition and Recovery To encourage labeled and unlabeled images to learn comprehensive common semantics from each other, we mix these cubes across all the labeled and unlabeled images in a mini-batch. As shown in Fig. 3, these cubes $\\mathbf{x}_j^l$ and $\\mathbf{x}_j^u$ are mixed into two shuffled magic-cubes while keeping their original positions, which produce two interpolated images with mixed magic-cubes $\\mathbf{X}_0^{mix},\\mathbf{X}_1^{mix}\\in \\mathbb{R}^{W\\times H\\times L}$ . The magic-cube mixing cross-image operations are termed as $\\mathcal{A}_{\\text{mix}}^{\\text{cross}}(\\cdot)$ . The mixed images are then fed into the student network $\\mathcal{F}(\\cdot;\\Theta^s)$ , followed by a softmax layer $\\sigma(\\cdot)$ , obtaining the prediction maps $\\mathbf{P}_0^{mix},\\mathbf{P}_1^{mix}\\in \\mathbb{R}^{C\\times W\\times H\\times L}$ , where $C$ denotes the number of classes. Next, we recover $\\mathbf{P}_{\\text{cross}}^l,\\mathbf{P}_{\\text{cross}}^u\\in \\mathbb{R}^{C\\times W\\times H\\times L}$ from $\\mathbf{P}_0^{mix}$ and $\\mathbf{P}_1^{mix}$ via recovering the magic-cubes back to the original positions in their original", + "bbox": [ + 496, + 674, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23871", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f7942144e283283c80745454d88a9f2166721988463af8b9454a6f03ab97dfd3.jpg", + "image_caption": [ + "Figure 3. Magic-cube partition and recovery. Blue and green volumes represent labeled and unlabeled image, respectively. The number on cubes represents its relative-location in the original image." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 834, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "images, and this operation is denoted as $\\mathcal{A}_{rec}^{cross}(\\cdot)$ . We can simply denote the whole process as:", + "bbox": [ + 75, + 297, + 470, + 329 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\text {c r o s s}} ^ {l}, \\mathbf {P} _ {\\text {c r o s s}} ^ {u} = \\mathcal {A} _ {\\text {r e c}} ^ {\\text {c r o s s}} \\left(\\sigma \\left(\\mathcal {F} \\left(\\mathcal {A} _ {\\text {m i x}} ^ {\\text {c r o s s}} \\left(\\mathcal {A} _ {\\text {p a r t}} (\\mathbf {X} ^ {\\mathcal {B}})\\right); \\boldsymbol {\\Theta} ^ {s}\\right)\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 334, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{P}_{cross}^{l},\\mathbf{P}_{cross}^{u}\\in \\mathbb{R}^{(C + 1)\\times W\\times H\\times L}$ . The loss functions for labeled and unlabeled images are defined as", + "bbox": [ + 75, + 353, + 468, + 386 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c r o s s}} ^ {l} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {\\text {d i c e}} \\left(\\mathbf {P} _ {\\text {c r o s s}} ^ {l}, \\mathbf {Y} ^ {l}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 393, + 468, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c r o s s - i n} ^ {u} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {d i c e} \\left(\\mathbf {P} _ {c r o s s} ^ {u}, \\widehat {\\mathbf {Y}} ^ {u}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 415, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\ell_{\\text{dice}}$ denotes multi-class Dice loss. $\\mathbf{Y}^l \\in \\{0,1,\\dots,C\\}^{W \\times H \\times L}$ denotes per-voxel manual annotations for $\\mathbf{X}^l$ . $\\mathcal{L}_{\\text{cross-in}}^u$ indicates the refined pseudo-labels for $\\mathbf{X}^u$ by blending cube-wise representations and more details can be seen in Sec. 3.2.", + "bbox": [ + 76, + 446, + 468, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Within-image Partition and Recovery Besides magic-cubes partition and recovery across images, we also design a within-image partition and recovery branch for single images, which can better consider local features and learn local attributes for identifying targets, especially targets of small sizes. For $\\mathbf{X}^l$ , the partitioned $j$ -th cube $\\mathbf{x}_j^l$ is fed into $\\mathcal{F}(\\cdot; \\Theta^s)$ and the softmax layer, which can acquire a cubewise prediction map $\\mathbf{p}_j^l \\in \\mathbb{R}^{C \\times W / N \\times H / N \\times L / N}$ . Finally, we recover the magic-cube via mixing $N^3$ cube-wise prediction maps back to their original positions as illustrated in Fig. 3, and the recovered probability map is denoted as $\\mathbf{P}_{in}^l \\in \\mathbb{R}^{C \\times W \\times H \\times L}$ . Similarly, for $\\mathbf{X}^u$ , the probability map is denoted as $\\mathbf{P}_{in}^u \\in \\mathbb{R}^{C \\times W \\times H \\times L}$ . This operation is denoted as $\\mathcal{A}_{rec}^{in}(\\cdot)$ . The whole process for obtaining $\\mathbf{P}_{in}^l$ and $\\mathbf{P}_{in}^u$ are:", + "bbox": [ + 75, + 522, + 470, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {i n} ^ {l}, \\mathbf {P} _ {i n} ^ {u} = \\mathcal {A} _ {r e c} ^ {i n} (\\sigma (\\mathcal {F} (\\mathcal {A} _ {p a r t} (\\mathbf {X} ^ {\\mathcal {B}}); \\boldsymbol {\\Theta} ^ {s}))). \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 758, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $\\mathbf{P}_{in}^{l}$ and $\\mathbf{P}_{in}^{u}$ are recovered via predictions of cubes, $\\mathbf{P}_{in}^{l}$ and $\\mathbf{P}_{in}^{u}$ are denoted as cube-wise representations in the rest of the paper. The loss function for the labeled image is:", + "bbox": [ + 76, + 787, + 468, + 834 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n} ^ {l} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {d i c e} \\left(\\mathbf {P} _ {i n} ^ {l}, \\mathbf {Y} ^ {l}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 842, + 468, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the unlabeled image, we propose to leverage the cubewise representations by refining the pseudo-label instead of", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "directly computing the loss function as in Eq. 5, which is illustrated in Sec. 3.2. We conduct ablation studies on various design choices of utilizing the cube-wise representations in the experiment part.", + "bbox": [ + 496, + 297, + 892, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cube-wise Pseudo-label Blending", + "text_level": 1, + "bbox": [ + 500, + 368, + 792, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned in the previous section, medical imaging experts reveal that local attributes e.g., texture, luster and boundary smoothness are crucial elements for identifying targets such as tiny organs in medical images [42]. Inspired by this, we propose a within-image partition and recovery module to learn cube-wise local representations. Since the teacher network takes the original volume as input, and pays more attention to learn large organs in the head class, voxels that actually belong to the tail class can be incorrectly predicted as head class voxels, due to the lack of local attributes learning. To effectively increase the chance of the voxel predicted to the tail class, we design a cube-wise pseudolabel blending module, which blends the original pseudolabel with cube-wise features.", + "bbox": [ + 496, + 392, + 892, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In detail, the image-level prediction map of $\\mathbf{X}^u$ acquired from the teacher network is defined as $\\mathbf{P}_T^u = \\mathcal{F}(\\mathbf{X}^u;\\boldsymbol{\\Theta}^t)$ , where $\\boldsymbol{\\Theta}^t$ is the parameters of teacher network. The reconstructed cube-level representations of $\\mathbf{X}^u$ acquired from the student network is $\\mathbf{P}_{in}^u$ . As mentioned above, cubewise features pay more attention to learn local attributes which is important for tiny organs. Therefore, we propose a distribution-aware blending strategy, which is shown in Fig. 4(a), and formulated as:", + "bbox": [ + 496, + 603, + 890, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\text {b l e n d}} ^ {u} = \\left(\\mathbf {1} - \\mathcal {R} (\\boldsymbol {\\Omega})\\right) \\odot \\mathbf {P} _ {T} ^ {u} + \\mathcal {R} (\\boldsymbol {\\Omega}) \\odot \\mathbf {P} _ {\\text {i n}} ^ {u}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 750, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ indicates the element-wise multiplication, $\\mathbf{1} \\in \\{1\\}^{(C + 1) \\times W \\times H \\times L}$ , and $\\Omega \\in \\mathbb{R}^{W \\times H \\times L}$ is a distribution-aware weight map.", + "bbox": [ + 500, + 779, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the weight map, we firstly learn the class-wise distribution $\\mathbf{v}$ during training. Suppose $\\mathbf{v} \\in \\mathbb{R}^C$ is a vector, $\\mathbf{v} = \\{\\mathbf{v}_0, \\dots, \\mathbf{v}_{C-1}\\}$ , whose element $\\mathbf{v}_c$ indicates the number of voxels belonging to $(c+1)$ th organ, which is accumulated by counting the voxels w.r.t. pseudo-labels over a few previ", + "bbox": [ + 496, + 825, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23872", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1ebf55d0cc818c18cf975aa147a48104522d5298fdf9f7afc91d5895d944c469.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 220, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d98634080a5de7bd905effcc2df63480beccf65faf1c8e08c232fc113042621a.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 227, + 95, + 336, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7eeee30938a71597b0e8cc67df5a84f4b2c9c95aa82a94cd1a07d6d97b34a0fb.jpg", + "image_caption": [ + "Figure 4. The cube-wise pseudo-label blending module. (a) Pipeline. (b) Illustration on $\\mathcal{R}(\\Omega_{\\mathbf{m}})$ . If the pseudo-label $\\tilde{\\mathbf{Y}}_{\\mathbf{m}}^{u}$ on $\\mathbf{m}$ is incorrectly assigned as a big organ, a big $\\Omega_{\\mathbf{m}}$ is assigned to ensure possible correction from cube-wise representations." + ], + "image_footnote": [], + "bbox": [ + 339, + 95, + 467, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ous iterations. Let $\\widetilde{\\mathbf{Y}}_{\\mathbf{m}}^{u}$ denote the pseudo-label for the voxel on location m. The value $\\Omega_{\\mathbf{m}}\\in \\mathbb{R}^{1}$ at each spatial location m on the weight map is derived as:", + "bbox": [ + 75, + 281, + 468, + 329 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Omega} _ {\\mathbf {m}} = \\sum_ {c = 0} ^ {C - 1} \\frac {\\mathbf {I} \\left(\\widetilde {\\mathbf {Y}} _ {\\mathbf {m}} ^ {u} = c\\right) \\mathbf {v} _ {c}}{\\operatorname* {m a x} \\mathbf {v}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 340, + 468, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{I}(\\cdot)$ is an indicator function. Note that to match with the data dimension in Eq. 6, we replicate $\\Omega C + 1$ times to acquire $\\mathcal{R}(\\Omega) \\in \\mathbb{R}^{(C + 1) \\times W \\times H \\times L}$ .", + "bbox": [ + 75, + 393, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, since the teacher network takes the original volume as input, and pays more attention to learn large organs in the head class, the teacher pseudo-label may be biased. To unbiased the pseudo-label, on the one hand, we keep the pseudo-label of small organs. On the other hand, we remedy the possible incorrect pseudo-labels of big organs via changing them to small-organ classes via Eq. 6 if $\\Omega_{\\mathbf{m}}$ is large. This can effectively increase the chance of the voxels predicted to small-organs.", + "bbox": [ + 75, + 439, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final refined pseudo-label $\\widehat{\\mathbf{Y}}^u$ is obtained via $\\widehat{\\mathbf{Y}}_{\\mathbf{m}}^{u} = \\arg \\max_{c\\in \\mathcal{C}}\\mathbf{P}_{blend,c,\\mathbf{m}}^{u}$ , where $\\mathcal{C}$ indicates the finite set of class labels, and $\\mathbf{m}$ means the voxel location.", + "bbox": [ + 75, + 574, + 468, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Magic-cube Location Reasoning", + "text_level": 1, + "bbox": [ + 76, + 632, + 361, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fully leverage the prior anatomical knowledge of multi-organs, we further propose a magic-cube location reasoning method, which is in line with within-image partition. For one image $\\mathbf{X}^l$ , the partitioned $j$ -th cube's $\\mathbf{x}_j^l$ is fed into the encoder of student network $\\mathcal{F}_{enc}(\\cdot;\\Theta_{enc}^s)$ to generate $\\mathbf{f}_j^l \\in \\mathbb{R}^{D \\times W / N \\times H / N \\times L / N}$ . Then, the feature $\\mathbf{f}_j^l$ is flattened into a one-dimensional vector, passing through a classification head $\\mathcal{F}_{cls}(\\cdot;\\Theta_{cls})$ to project the flattened feature vector into the same size as the number of cubes in one image, i.e., $N^3$ . The classification head is composed of two fully connected layers [10, 34]. Finally, we concatenate the output $\\hat{\\mathbf{y}}_j^l \\in \\mathbb{R}^{N^3}$ from the classification head for all $N^3$ cubes in one image, which generates a vector in size $(N^3)^2$ . This vector is further reshaped into a matrix of size $N^3 \\times N^3$ , whose rows indicate the probabilities of corresponding cubes belonging to the $N^3$ locations. The same", + "bbox": [ + 75, + 656, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "applies to $\\mathbf{X}^u$ . For both $\\mathbf{X}^l$ and $\\mathbf{X}^u$ in $\\mathcal{B}$ , the cross-entropy loss is adopted to learn relative locations of cubes:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} \\left(\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}\\right) = \\frac {1}{| \\mathcal {B} |} \\sum_ {\\mathbf {X} \\in \\mathcal {B}} \\ell_ {c e} (\\sigma \\left(\\mathcal {F} _ {c l s} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 133, + 890, + 170 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathcal {F} _ {e n c} ^ {s} \\left(\\mathcal {A} _ {p a r t} (\\mathbf {X}); \\boldsymbol {\\Theta} _ {e n c} ^ {s}\\right); \\boldsymbol {\\Theta} _ {c l s}\\right)), \\mathbb {y})\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 174, + 839, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell_{ce}$ denotes the cross-entropy loss, y represents cube's relative locations in $\\mathbf{X}$ .", + "bbox": [ + 498, + 202, + 890, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 500, + 241, + 643, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, we develop a total loss function based on a minibatch for our MagicNet as follows:", + "bbox": [ + 498, + 265, + 890, + 295 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\mathcal {B}; \\Theta_ {\\text {e n c}} ^ {s}, \\Theta_ {\\text {s e g}} ^ {s}, \\Theta_ {\\text {c l s}}) = \\mathcal {L} _ {\\sup } (\\mathcal {B}; \\Theta_ {\\text {e n c}} ^ {s}, \\Theta^ {s}, \\Theta_ {\\text {c l s}}) \\tag {9} \\\\ + \\mathcal {L} _ {u n s u p} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 306, + 890, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{sup}(\\cdot)$ and $\\mathcal{L}_{unsup}(\\cdot)$ denote the loss for labeled and unlabeled images in $\\mathcal{B}$ , respectively:", + "bbox": [ + 498, + 354, + 890, + 385 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\sup } (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}) = \\mathcal {L} _ {c r o s s} ^ {l} (\\mathcal {B}; \\Theta^ {s}) \\tag {10} \\\\ + \\mathcal {L} _ {i n} ^ {l} (\\mathcal {B}; \\Theta^ {s}) + \\beta \\mathcal {L} _ {c l s} ^ {l} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 396, + 890, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is the weight factor to balance the location reasoning term with others, and", + "bbox": [ + 498, + 446, + 890, + 476 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {u n s u p} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}) = \\alpha \\mathcal {L} _ {c r o s s - i n} ^ {u} (\\mathcal {B}; \\Theta^ {s}) \\tag {11} \\\\ + \\beta \\mathcal {L} _ {c l s} ^ {u} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 487, + 890, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is another weight factor to balance the two terms. For $\\mathcal{L}_{cls}^{l}$ and $\\mathcal{L}_{cls}^{u}$ in Eq. 10 and Eq. 11, the following equation holds:", + "bbox": [ + 498, + 534, + 890, + 577 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = \\mathcal {L} _ {c l s} ^ {l} + \\mathcal {L} _ {c l s} ^ {u}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 578, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Testing Phase", + "text_level": 1, + "bbox": [ + 500, + 604, + 640, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a testing image $\\mathbf{X}_{test}$ , the probability map is obtained by: $\\mathbf{P}_{test} = \\sigma (\\mathcal{F}(\\mathbf{X}_{test};\\Theta^{s*}))$ , where $\\Theta^{s*}$ denotes the well-trained student network parameters. The final label map can be determined by taking the argmax in the first dimension of $\\mathbf{P}_{test}$ .", + "bbox": [ + 498, + 627, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 715, + 630, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Pre-processing", + "text_level": 1, + "bbox": [ + 500, + 741, + 751, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BTCV Multi-organ Segmentation Dataset BTCV multi-organ segmentation dataset is from MICCAI Multi-Atlas Labeling Beyond Cranial Vault-Workshop Challenge [13] which contains 30 subjects with 3779 axial abdominal CT slices with 13 organs annotation. In pre-processing, we follow [27] to re-sample all the CT scans to the voxel spacing $[1.5 \\times 1.5 \\times 2.0]mm^3$ and normalize them to have zero mean and unit variance. Strictly following [5, 6, 10], 18 cases are divided for training and the other 12 cases are for testing.", + "bbox": [ + 498, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23873", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e25ce360dfd7f03c00a94f3afa27385f3e072a7b0f857828341cc35bd84aebdd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Labeled%MethodsSplR.kidL.kidGallEsoLivStoAorIVCVeinsPanRGLGAvg. DSCAvg. NSD
30%V-Net [20]86.3775.3871.2041.8148.4692.7744.7888.2072.6049.2722.3037.7921.7157.9056.90
MT [28]79.8977.5678.0838.3158.9992.2648.7388.6179.3652.7328.4254.1621.3061.4260.85
UA-MT [39]87.7573.8772.9742.0850.7993.2242.0788.3071.8055.2736.6143.2237.6761.2060.99
ICT [30]88.7878.0083.9350.0953.3588.3264.1989.0279.6063.5158.9043.2950.7068.5967.77
CPS [7]82.9876.9179.2129.9152.5392.5941.3786.5974.7349.5318.9645.1324.0658.0456.64
SS-Net [35]85.5373.3280.7030.8552.4491.6717.6786.2872.3947.4022.9537.3839.0156.7454.38
SLC-Net [16]90.0484.3885.8453.5555.8792.8657.8089.8380.8753.2238.7848.1037.0366.7868.81
MagicNet (ours)91.4284.6486.1962.8662.4993.8972.8790.7083.5270.0764.9460.8857.4875.5376.31
40%V-Net [20]84.9882.7282.0736.6463.4893.5457.4989.7478.6360.4249.3955.6038.4967.1767.84
MT [28]85.7078.9379.0842.8061.0993.4557.5789.7080.3063.9541.1450.4629.6965.6865.98
UA-MT [39]88.7475.8878.9154.2558.5593.4658.9089.2376.1562.3047.9151.5344.9267.7568.87
ICT [30]90.3184.4186.9649.2265.6594.2965.9590.2381.4469.5666.6157.3556.0173.6974.98
CPS [7]87.5672.9977.5953.3154.0892.4154.5887.7574.3258.6848.0250.3943.8665.8165.34
SS-Net [35]84.7476.3774.1943.4257.0592.9014.3783.1469.7752.4527.0854.2927.6658.2657.75
SLC-Net [16]90.0584.0086.4356.1658.9194.6870.7289.9379.4560.5954.2251.0339.0870.4072.87
MagicNet (ours)91.6185.0288.1358.1666.7294.0774.4690.7784.3171.5668.9063.4860.4776.7478.68
100%V-Net [20]84.0084.8286.3867.4265.0294.8373.7590.2784.1969.8563.5462.6065.0276.2877.45
", + "bbox": [ + 145, + 88, + 823, + 276 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe1cd6ba1de3aba9356a1aa2107e6e9a83fa6d3036653fbdceaa2adf7b9c1177.jpg", + "table_caption": [ + "Table 1. Comparison results (DSCs of each organ, avg. DSC, and avg. NSD) between our method and existing semi-supervised medical image segmentation methods on the BTCV dataset. V-Net means training with only labeled samples on the V-Net backbone. DSC: Dice Similarity Coefficient. NSD: Normalized Surface Dice. The standard deviations are not reported due to space limit. Note: Spl: spleen, R.Kid: right kidney, L.Kid: left kidney, Gall: gallbladder, Eso: esophagus, Liv: liver, Sto: stomach, Aor: aorta, IVC: inferior vena cava, Veins: portal and splenic veins, Pan: pancreas, LG/RG: left/right adrenal glands." + ], + "table_footnote": [], + "table_body": "
Labeled%MethodsSpleenL.kidneyGallbladderEsophagusLiverStomachPancreasDuodenumAvg. DSCAvg. NSD
10%V-Net [20]88.89 (16.31)84.90 (20.11)55.73 (34.27)58.27 (20.05)93.15 (7.11)57.15 (30.62)53.33 (22.44)35.28 (19.83)65.8449.96
MT [28]89.22 (13.56)88.04 (17.61)61.40 (31.91)58.89 (20.62)93.60 (5.04)74.00 (20.37)64.70 (19.31)44.53 (17.61)71.8053.54
UA-MT [39]89.94 (13.36)89.26 (15.67)59.19 (32.35)59.43 (19.45)93.77 (5.05)75.43 (19.61)65.86 (17.48)44.57 (17.40)72.1854.04
ICT [30]90.27 (13.70)89.89 (15.81)61.47 (30.31)58.70 (20.17)93.47 (5.39)72.45 (20.95)65.81 (18.28)45.00 (19.22)72.1354.07
CPS [7]91.73 (12.11)90.59 (14.85)60.54 (32.29)55.82 (21.03)94.19 (4.93)78.80 (15.28)62.93 (20.03)32.58 (18.68)70.9052.80
SS-Net [35]83.52 (19.56)78.28 (25.05)53.08 (33.88)56.24 (20.54)91.84 (6.07)47.65 (29.70)43.23 (24.11)25.70 (18.71)59.9442.68
SLC-Net [16]86.89 (17.54)87.54 (18.00)55.28 (34.75)58.51 (20.54)93.70 (5.33)69.74 (22.89)46.55 (26.21)28.86 (18.87)65.8913.28
MagicNet (ours)92.33 (11.12)91.19 (14.18)66.35 (29.29)70.95 (12.30)94.31 (4.64)81.56 (14.91)76.31 (10.56)61.50 (12.85)79.3162.31
20%V-Net [20]90.53 (15.33)88.28 (19.76)65.16 (30.03)66.10 (16.98)94.49 (4.18)79.56 (16.56)69.04 (17.43)51.98 (18.53)75.6459.66
MT [28]91.76 (12.77)91.40 (14.49)63.83 (31.74)64.15 (18.05)94.44 (4.86)81.19 (15.79)72.24 (15.79)57.01 (16.70)77.0060.32
UA-MT [39]92.39 (11.31)91.44 (14.61)63.63 (31.46)64.33 (17.92)94.50 (4.77)83.21 (14.03)71.93 (15.51)57.21 (17.35)77.3360.64
ICT [30]92.96 (9.60)91.67 (14.52)65.90 (30.49)64.35 (16.79)94.46 (4.66)82.60 (15.46)74.18 (12.99)57.67 (16.16)77.9761.24
CPS [7]92.66 (11.46)91.87 (14.33)64.75 (31.21)57.27 (19.56)94.97 (4.51)85.65 (9.50)74.15 (12.49)55.00 (16.58)77.0460.01
SS-Net [35]91.17 (14.77)87.77 (19.37)63.49 (30.83)65.02 (16.65)93.76 (4.82)73.04 (22.16)70.37 (17.70)52.81 (20.32)74.6857.57
SLC-Net [16]92.60 (11.24)91.38 (14.57)62.46 (31.37)62.84 (19.31)94.54 (4.72)80.30 (14.73)69.79 (16.99)53.48 (18.40)75.9259.93
MagicNet (ours)93.52 (10.22)92.01 (14.30)71.04 (26.92)70.95 (12.58)94.89 (4.42)85.77 (10.27)78.81 (8.07)66.21 (12.33)81.6565.87
100%V-Net [20]94.35 (5.81)92.30 (14.22)73.61 (28.80)73.48 (12.43)95.17 (3.10)89.09 (7.74)80.86 (7.38)70.03 (11.60)83.6168.85
", + "bbox": [ + 145, + 371, + 823, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison results (DSCs of each organ, avg. DSC, and avg. NSD) between our method and existing semi-supervised medical image segmentation methods on MACT dataset under 4-fold cross-validation. The value in $(\\cdot)$ is standard deviation.", + "bbox": [ + 75, + 561, + 888, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MACT Dataset Multi-organ abdominal CT reference standard segmentations (MACT) is a public multi-organ segmentation dataset, containing 90 CT volumes with 8 organs annotation. It is manually re-annotated by [11] based on 43 cases of NIH pancreas [21] and 47 cases of Synapse [13]. We follow the setting of UMCT [36]. 1) In preprocessing, we set the soft tissue CT window range of $[-125, 275]$ HU and all the CT scans are re-sampled to an isotropic resolution of $1.0\\mathrm{mm}^3$ . Then, image intensities are normalized to have zero mean and unit variance. 2) We randomly split the dataset into four folds, and perform 4-fold cross-validation. Labeled training set is then randomly selected, which are applied to all experiments for fair comparison.", + "bbox": [ + 75, + 609, + 472, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Setup and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 816, + 450, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this work, we conduct all the experiments based on PyTorch with one NVIDIA 3090 GPU and use V-Net as our backbone [16,35,39]. For magic-cube location reasoning, we add a classification head composed of two fully", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "connected layers at the end of our encoder. Firstly, for BTCV dataset, the framework is trained by an SGD optimizer for 70K iterations, with an initial learning rate (lr) 0.01 with a warming-up strategy: $lr = \\text{base\\_lr} \\times (1 - \\frac{\\text{iterations}}{\\text{max\\_iterations}})^{0.9}$ . For MACT dataset, the framework is trained by an SGD optimizer for 30K iterations, with an initial learning rate (lr) 0.01 decayed by 0.1 every 12K iterations. Following [16, 18, 35, 36, 39], for both two datasets, batch size is set as 4, including 2 labeled images and 2 unlabeled images, and the hyper-parameter $\\alpha$ is set using a time-dependent Gaussian warming-up function and $\\beta$ is empirically set as 0.1 [10]. We randomly crop volume of size $96^3$ , and then we partition the volume into $N^3$ small cubes, forming our magic-cubes, and the choice of $N$ is also discussed in Sec. 4.4. In the testing phase, a sliding window strategy is adopted to obtain final results with a stride of $16^3$ . The final evaluation metrics for our method are DSC (\\%, Dice-Sørensen Coefficient) and NSD (\\%, Normalized Surface Dice) which are commonly used in the cur", + "bbox": [ + 496, + 609, + 890, + 896 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23874", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/11b7c51ad54a93186e5d8326d4efd3a1ca5ccd2c1fe51d31f94a58ffa94666cc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCrossInLocBldAvg. DSCspleenl.kidneygallbladderesophagusliverstomachpancreasduodenum
Baseline (MT [28])71.8886.8488.9254.5053.4093.0170.8671.7755.76
Cross✓78.1392.1190.5363.4270.0593.9079.6075.5859.87
Cross + In✓✓78.8392.8290.7564.9170.6893.9080.9276.2560.40
Cross + Loc✓✓78.5692.4991.0765.9569.7294.1280.4075.8258.94
Cross + In + Loc✓✓✓78.8792.5591.1865.7270.7794.4280.0575.8160.43
Cross + In + Loc + Bld✓✓✓✓79.3192.3391.1966.3570.9594.3181.5676.3161.50
", + "bbox": [ + 119, + 89, + 848, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study (DSC %) for the effectiveness of each component of MagicNet. Cross: cross-image partition-and-recovery. In: Within-image partition-and-recovery. Bld: cube-wise pseudo-label blending. Loc: magic-cube location reasoning.", + "bbox": [ + 75, + 183, + 888, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c19517246faf0d44e4777c913db537fadc2db62d82121fbed446ae4fc0b96413.jpg", + "image_caption": [ + "Figure 5. DSC comparison in box plots on MACT dataset with $10\\%$ labeled images (Best viewed electronically, zoom in)." + ], + "image_footnote": [], + "bbox": [ + 81, + 234, + 465, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rent multi-organ segmentation challenge [12, 19].", + "bbox": [ + 76, + 385, + 403, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with the State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 411, + 467, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare MagicNet with six state-of-the-art semi-supervised segmentation methods, including mean-teacher (MT) [28], uncertainty-aware mean-teacher (UA-MT) [39], interpolated consistency training (ICT) [30], cross pseudosupervision (CPS) [7], smoothness and class-separation (SS-Net) [35], shape-awareness and local constraints (SLC-Net) [16]. All experiments of other methods are re-trained by their official code when transferred to multi-organ task. V-Net is adopted as our backbone which is the same as the above methods for fair comparison.", + "bbox": [ + 75, + 431, + 467, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, we summarize the results of BTCV dataset. Visualizations compared with state-of-the-arts are shown in the supplementary material. Noticeable improvements compared with state-of-the-arts can be seen for organs such as Stomach (↑ 8.68), Gallbladder (↑ 9.31), Veins (↑ 6.56), Right adrenal glands (↑ 6.72) and Left adrenal glands (↑ 6.78). It is interesting to observe that some semi-supervised methods even perform worse than the lower bound. This is because these methods are less able to learn common semantics from the labeled images to unlabeled ones, generating less accurate pseudo-labels, which even hamper the usage of unlabeled images (see the supplementary material).", + "bbox": [ + 75, + 583, + 467, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then evaluate MagicNet on MACT dataset. In Table 2, we can observe: (1) a more significant performance improvement between MagicNet and other state-of-the-arts with $10\\%$ labeled data than $20\\%$ , which demonstrates the effectiveness of MagicNet when training with fewer labeled CT images; (2) MagicNet successfully mitigates the inferior performance in tail classes, e.g., DSC improvement with $10\\%$ labeled images on Gallbladder ( $\\uparrow$ 4.88), Esophagus ( $\\uparrow$ 11.52), Pancreas ( $\\uparrow$ 10.45) and Duodenum ( $\\uparrow$ 16.50). Fig. 5", + "bbox": [ + 75, + 763, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb2aca65c72d67cb7f1d7a531cc9ce329df66025d2566aa5d2086de281c5b9d2.jpg", + "image_caption": [ + "Figure 6. Scramble position and keep position when mixing small-cubes across images (from Fig. 3). Red arrow indicates the cube location in mixed images." + ], + "image_footnote": [], + "bbox": [ + 527, + 234, + 864, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shows a comparison in DSC of six methods for different organs. Note that UMCT [36] on MACT dataset achieves $75.33\\%$ in DSC. It uses AH-Net [17] as the backbone (3 are used), whose parameter number is much larger. Due to computation resources, we do not directly compare with it.", + "bbox": [ + 498, + 376, + 890, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 457, + 651, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on MACT dataset [11] with $10\\%$ labeled data to show the effectiveness of each module. The effectiveness of each component in MagicNet We conduct ablation studies to show the impact of each component in MagicNet. In Table 3, the first row indicates the mean-teacher baseline model [28], which our method is designed on. Compared to the baseline, our magic-cube partition and recovery can yield good segmentation performance. Cross and In represent our cross- and within-image partition-and-recovery, which increase the performance from $71.88\\%$ to $78.13\\%$ and $78.87\\%$ , respectively. This shows the powerful ability with our specially-designed data augmentation method. Loc represents magic-cube location reasoning for within-image branch. We can see from the table that based on our cross-image branch, adding relative locations of small-cubes can achieve $78.56\\%$ in DSC, adding only within-image partition-and-recovery branch achieves $78.83\\%$ , and adding both two branches leads to $78.87\\%$ . Finally, our proposed cube-wise pseudo label blending (short for Bld in Table 3) provides significant improvement to $79.31\\%$ , for which we will make a more complete analysis later in the current section.", + "bbox": [ + 498, + 477, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Design choices of partition and recovery Here, we discuss the design for cross-image partition-and-recovery branch: ① Should we maintain or scramble the magic-cube relative locations when manipulating our cross-image branch, as shown in Fig. 6? The comparison results are shown in the last two rows in Table 4. Scramble and Keep represent the", + "bbox": [ + 498, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23875", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/171e06bccc357fee41d167042fad4842fc7bca939d6dcdd8012aba790f33df13.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
①②DSCNSD
scrambleU69.53 ± 11.2750.98 ± 11.00
keepU73.84 ± 9.9655.80 ± 10.81
scrambleLU71.91 ± 11.2054.10 ± 11.28
keepLU78.13 ± 8.1260.79 ± 9.90
", + "bbox": [ + 81, + 92, + 349, + 154 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f20adcb1abef0882c26c830d322c16bc3663cf3624f813bc665ba663d3f99ea4.jpg", + "table_caption": [ + "Table 4. Ablation of design choices for crossimage partition and recovery (Question $①$ and $②$ , mean $\\pm$ std of all cases). Scramble/keep: ignore/keep original positions when mixed. U: only for unlabeled data. LU: for both labeled and unlabeled data. The last row is ours." + ], + "table_footnote": [], + "table_body": "
DSCNSD
teacher sup.78.47 ± 7.8361.52 ± 9.21
mutual sup.70.45 ± 11.4753.71 ± 11.78
blending79.31 ± 7.5562.31 ± 9.08
", + "bbox": [ + 364, + 90, + 624, + 155 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d439f165b37e976fbf6a22b368cfa711c37b7801fff63234154532a64a9f39bc.jpg", + "table_caption": [ + "Table 5. Comparison of different pseudolabel supervision/blending strategies for unlabeled data. Sup.: supervision. Blending: our cube-wise pseudo-label blending." + ], + "table_footnote": [], + "table_body": "
DSCNSD
CutMix [40]73.80 ± 9.9955.86 ± 10.54
CutOut [8]72.47 ± 10.2754.62 ± 10.67
MixUp [41]72.13 ± 11.3254.07 ± 12.01
Ours (2)78.68 ± 7.4560.78 ± 9.12
Ours (3)79.31 ± 7.5562.31 ± 9.08
", + "bbox": [ + 638, + 88, + 888, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Comparison of different data augmentation methods, where we try different CutMix and CutOut sizes, and choose the best results. For MagicNet, we compare different $N$ values in $(\\cdot)$ .", + "bbox": [ + 633, + 169, + 890, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e54eb8b9e5b5fdeb34f6dfe547f9cbd16ef87b5eb64a09905e5362b380150d0.jpg", + "image_caption": [ + "Figure 7. Kernel Density Estimation of labeled and unlabeled samples on the first three channels of the feature map of \"only unlabeled images\" and \"both labeled and unlabeled images\". Considerable distribution mismatch between labeled and unlabeled images can be observed (top), while our method can well improve the distribution mismatch (bottom). Best viewed zoom in." + ], + "image_footnote": [], + "bbox": [ + 78, + 273, + 467, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "partitioned small-cubes are randomly mixed while ignoring their original locations or kept when mixing them across images. Results show that the relative locations between multiple organs are important for multi-organ segmentation. ② Should our cross-image data augmentation be operated on only unlabeled images (see U in Table 4) or both labeled and unlabeled images (see LU in Table 4)? The latter obtains a much better performance compared to the former. Besides, it largely reduces the distribution mismatch between labeled and unlabeled data, as shown in Fig. 7.", + "bbox": [ + 75, + 532, + 468, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cube-wise pseudo-label blending As illustrated in Sec. 3.2, we blend the output of within-image branch and the output of teacher model to obtain the final pseudo-label for complementing local attributes. We compare our blending with two other methods, as shown in Table 5. Three supervision schemes, illustrated in Fig. 8, are compared for unlabeled images based on our framework. Teacher sup. means the outputs of cross- and within-image branch are both supervised by the pseudo-label from the teacher model. Mutual sup. means the outputs of cross- and within-image are mutually supervised. It can be seen that our blending strategy works favorably for unlabeled data.", + "bbox": [ + 75, + 686, + 470, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with interpolated-based methods As shown in Table 6, our augmentation method outperforms other", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/374e9b3d55d9c607d4ddf82bb30d4f1db25d33b5b20ad3fdadb659936534845a.jpg", + "image_caption": [ + "Figure 8. Different pseudo-label supervision/blending strategies corresponding to Table 5. Blue line indicates supervision. (a) teacher sup. (b) mutual sup. (c) our blending." + ], + "image_footnote": [], + "bbox": [ + 576, + 270, + 820, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods such as CutMix, CutOut and MixUp. For fair comparison, we try several CutMix and CutOut sizes, and choose the best results.", + "bbox": [ + 496, + 415, + 890, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different number $(N)$ of small-cubes We study the impact of different numbers of small-cubes $N$ , as shown in Table 6. Slightly better performance is achieved when $N = 3$ . When $N = 4$ , the size of the small cube does not match the condition for VNet. Thus, we only compare the results given $N = 2$ and $N = 3$ .", + "bbox": [ + 496, + 459, + 892, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 564, + 756, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented MagicNet for semi-supervised multi-organ segmentation. MagicNet encourages unlabeled data to learn organ semantics in relative locations from the labeled data, and enhances the learning ability for small organs. Two essential components are proposed: (1) partition-and-recovery $N^3$ small cubes cross- and within-labeled and unlabeled images, and (2) the quality of pseudo label refinement by blending the cube-wise representations. Experimental results on two public multi-organ segmentation datasets verify the superiority of our method.", + "bbox": [ + 496, + 589, + 890, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: CT scans are all roughly aligned by the nature of the data. Magic-cube may not work well on domains $w / o$ such prior, e.g., microscopy images $w / o$ similar prior.", + "bbox": [ + 496, + 739, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 799, + 679, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (Grant No. 62101191), Shanghai Natural Science Foundation (Grant No. 21ZR1420800), and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ2229004).", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23876", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wenjia Bai, Ozan Oktay, Matthew Sinclair, Hideaki Suzuki, Martin Rajchl, Giacomo Tarroni, Ben Glocker, Andrew P. King, Paul M. Matthews, and Daniel Rueckert. Semi-supervised learning for network-based cardiac MR image segmentation. In Proc. MICCAI, 2017. 2", + "[2] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In Proc. ICLR, 2020. 3", + "[3] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Proc. NeurIPS, 2019. 3", + "[4] Gerda Bortsova, Florian Dubost, Laurens Hogeweg, Ioannis Katramados, and Marleen de Bruijne. Semi-supervised medical image segmentation via learning consistency under transformations. In Proc. MICCAI, 2019. 1, 2", + "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 5", + "[6] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021.5", + "[7] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proc. CVPR, 2021. 6, 7", + "[8] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 8", + "[9] Kang Fang and Wu-Jun Li. Dmnet: Difference minimization network for semi-supervised segmentation in medical images. In Proc. MICCAI, 2020. 1, 2", + "[10] Shuhao Fu, Yongyi Lu, Yan Wang, Yuyin Zhou, Wei Shen, Elliot Fishman, and Alan Yuille. Domain adaptive relational reasoning for 3d multi-organ segmentation. In Proc. MIC-CAI, 2020. 5, 6", + "[11] Eli Gibson, Francesco Giganti, Yipeng Hu, Ester Bonmati, Steve Bandula, Kurinchi Gurusamy, Brian Davidson, Stephen P. Pereira, Matthew J. Clarkson, and Dean C. Barratt. Automatic multi-organ segmentation on abdominal ct with dense v-networks. IEEE Trans. Medical Imaging, 37(8):1822-1834, 2018. 2, 6, 7", + "[12] Yuanfeng Ji, Haotian Bai, Jie Yang, Chongjian Ge, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhang, Wanling Ma, Xi-ang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023, 2022. 7", + "[13] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, T Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Workshop, 2015. 2, 5, 6" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Shuai Lin Li, Chuyu Zhang, and Xuming He. Shape-aware semi-supervised 3d semantic segmentation for medical images. In Proc. MICCAI, 2020. 1, 2", + "[15] Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, and Pheng-Ann Heng. Transformation consistent self-ensembling model for semi-supervised medical image segmentation. IEEE Trans. Neural Networks and Learning Systems, 32(2):523-534, 2020. 1, 2", + "[16] Jinhua Liu, Christian Desrosiers, and Yuanfeng Zhou. Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape awareness and local context constraints. In Proc. MICCAI, 2022. 6, 7", + "[17] Siqi Liu, Daguang Xu, S Kevin Zhou, Olivier Pauly, Sasa Grbic, Thomas Mertelmeier, Julia Wicklein, Anna Jerebko, Weidong Cai, and Dorin Comaniciu. 3d anisotropic hybrid network: Transferring convolutional features from 2d images to 3d anisotropic volumes. In Proc. MICCAI, 2018. 7", + "[18] Xiangde Luo, Jieneng Chen, Tao Song, and Guotai Wang. Semi-supervised medical image segmentation through dual-task consistency. In Proc. AAAI, 2021. 1, 2, 6", + "[19] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, Shucheng Cao, Qi Zhang, Shangqing Liu, Yunpeng Wang, Yuhui Li, Jian He, and Xiaoping Yang. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6695-6714, 2022. 7", + "[20] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proc. 3DV, 2016. 6", + "[21] Holger R Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim B Turkbey, and Ronald M Summers. Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In Proc. MICCAI, 2015. 6", + "[22] Holger R. Roth, Hirohisa Oda, Xiangrong Zhou, Natsuki Shimizu, Ying Yang, Yuichiro Hayashi, Masahiro Oda, Michtaka Fujiwara, Kazunari Misawa, and Kensaku Mori. An application of cascaded 3d fully convolutional networks for medical image segmentation. Comput. Medical Imaging Graph., 66:90-99, 2018. 2", + "[23] Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(3):608-620, 2022. 1, 2", + "[24] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Proc. NeurIPS, 2020. 3", + "[25] Hao Tang, Xuming Chen, Yang Liu, Zhipeng Lu, Junhua You, Mingzhou Yang, Shengyu Yao, Guoqi Zhao, Yi Xu, Tingfeng Chen, Yong Liu, and Xiaohui Xie. Clinically applicable deep learning framework for organs at risk delineation in CT images. Nat. Mach. Intell., 1(10):480-491, 2019. 1", + "[26] Yucheng Tang, Riqiang Gao, Ho Hin Lee, Shizhong Han, Yunqiang Chen, Dashan Gao, Vishwesh Nath, Camilo" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23877", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bermudez, Michael R. Savona, Richard G. Abramson, Shunxing Bao, Ilwoo Lyu, Yuankai Huo, and Bennett A. Landman. High-resolution 3d abdominal segmentation with random patch network fusion. Medical Image Anal., 69:101894, 2021. 2", + "[27] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proc. CVPR, 2022. 5", + "[28] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proc. NeurIPS, 2017. 2, 3, 6, 7", + "[29] Peng Tu, Yawen Huang, Feng Zheng, Zhenyu He, Liujuan Cao, and Ling Shao. Guidedmix-net: Semi-supervised semantic segmentation by using labeled images as reference. In Proc. AAAI, 2022. 3", + "[30] Vikas Verma, Alex Lamb, Juho Kannala, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi-supervised learning. In Proc. IJCAI, 2019. 3, 6, 7", + "[31] Kaiping Wang, Bo Zhan, Chen Zu, Xi Wu, Jiliu Zhou, Luping Zhou, and Yan Wang. Semi-supervised medical image segmentation via a tripled-uncertainty guided mean teacher model with contrastive learning. Medical Image Anal., 79:102447, 2022. 1, 2", + "[32] Qin Wang, Wen Li, and Luc Van Gool. Semi-supervised learning by augmented distribution alignment. In Proc. ICCV, 2019. 2", + "[33] Yan Wang, Yuyin Zhou, Wei Shen, Seyoun Park, Elliot K. Fishman, and Alan L. Yuille. Abdominal multi-organ segmentation with organ-attention networks and statistical fusion. Medical Image Anal., 55:88-102, 2019. 1, 2", + "[34] Chen Wei, Lingxi Xie, Xutong Ren, Yingda Xia, Chi Su, Jiaying Liu, Qi Tian, and Alan L. Yuille. Iterative reorganization with weak spatial constraints: Solving arbitrary jigsaw puzzles for unsupervised representation learning. In Proc. CVPR, 2019. 5", + "[35] Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, and Jianfei Cai. Exploring smoothness and class-separation for semi-supervised medical image segmentation. In Proc. MIC-CAI, 2022. 1, 2, 6, 7", + "[36] Yingda Xia, Dong Yang, Zhiding Yu, Fengze Liu, Jinzheng Cai, Lequan Yu, Zhuotun Zhu, Daguang Xu, Alan L. Yuille, and Holger Roth. Uncertainty-aware multi-view co-training for semi-supervised medical image segmentation and domain adaptation. Medical Image Anal., 65:101766, 2020. 2, 6, 7", + "[37] Lingxi Xie, Qihang Yu, Yuyin Zhou, Yan Wang, Elliot K. Fishman, and Alan L. Yuille. Recurrent saliency transformation network for tiny target segmentation in abdominal ct scans. IEEE Trans. Medical Imaging, 39(2):514-525, 2020. 2", + "[38] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence H. Staib, and James S. Duncan. Simcvd: Simple contrastive voxelwise representation distillation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(9):2228-2237, 2022. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Proc. MICCAI, 2019. 1, 2, 6, 7", + "[40] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proc. ICCV, 2019. 3, 8", + "[41] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In Proc. ICLR, 2018. 3, 8", + "[42] Xinkai Zhao, Chaowei Fang, De-Jun Fan, Xutao Lin, Feng Gao, and Guanbin Li. Cross-level contrastive learning and consistency constraint for semi-supervised medical image segmentation. In Proc. ISBI, 2022. 4", + "[43] Yuyin Zhou, Yan Wang, Peng Tang, Song Bai, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Semi-supervised 3d abdominal multi-organ segmentation via deep multi-planar co-training. In Proc. WACV, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 359 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23878", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_model.json b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..820628307e389bb9f644943dc83e7510cfa13ff6 --- /dev/null +++ b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.849, + 0.177 + ], + "angle": 0, + "content": "MagicNet: Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.203, + 0.836, + 0.223 + ], + "angle": 0, + "content": "Duowen Chen\\(^{1}\\) Yunhao Bai\\(^{1}\\) Wei Shen\\(^{2}\\) Qingli Li\\(^{1}\\) Lequan Yu\\(^{3}\\) Yan Wang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.224, + 0.825, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.24, + 0.871, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University \\(^{3}\\)The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.26, + 0.787, + 0.275 + ], + "angle": 0, + "content": "duowen_chen@hotmail.com, yhbai@stu.ecnu.edu.cn, wei.shen@sjtu.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.278, + 0.729, + 0.293 + ], + "angle": 0, + "content": "qlli@cs.ecnu.edu.cn, lqyu@hku.hk, ywang@cee.ecnu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.474, + 0.797 + ], + "angle": 0, + "content": "We propose a novel teacher-student model for semi-supervised multi-organ segmentation. In teacher-student model, data augmentation is usually adopted on unlabeled data to regularize the consistent training between teacher and student. We start from a key perspective that fixed relative locations and variable sizes of different organs can provide distribution information where a multi-organ CT scan is drawn. Thus, we treat the prior anatomy as a strong tool to guide the data augmentation and reduce the mismatch between labeled and unlabeled images for semi-supervised learning. More specifically, we propose a data augmentation strategy based on partition-and-recovery \\( N^3 \\) cubes cross- and within- labeled and unlabeled images. Our strategy encourages unlabeled images to learn organ semantics in relative locations from the labeled images (cross-branch) and enhances the learning ability for small organs (within-branch). For within-branch, we further propose to refine the quality of pseudo labels by blending the learned representations from small cubes to incorporate local attributes. Our method is termed as MagicNet, since it treats the CT volume as a magic-cube and \\( N^3 \\)-cube partition-and-recovery process matches with the rule of playing a magic-cube. Extensive experiments on two public CT multi-organ datasets demonstrate the effectiveness of MagicNet, and noticeably outperforms state-of-the-art semi-supervised medical image segmentation approaches, with \\( +7\\% \\) DSC improvement on MACT dataset with \\( 10\\% \\) labeled images. Code is available at https://github.com/DeepMed-Lab-ECNU/MagicNet." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.824, + 0.21, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.85, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Abdominal multi-organ segmentation in CT images is an essential task in many clinical applications such as" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.326, + 0.89, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.494, + 0.893, + 0.604 + ], + "angle": 0, + "content": "Figure 1. Two data augmentation strategies in MagicNet. Left: Although labeled and unlabeled images are not aligned, the latter can be regarded as a shifted version of the former. Co-shift of cubes transfers organ semantics in relative locations from the labeled data to unlabeled data. Right: Segmenting small organs from original images are difficult due to the cluttered background. Small cubes mitigate the impact from background and focus more on local attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.893, + 0.761 + ], + "angle": 0, + "content": "computer-aided intervention [25, 33]. But, training an accurate multi-organ segmentation model usually requires a large amount of labeled data, whose acquisition process is time-consuming and expensive. Semi-supervised learning (SSL) has shown great potential to handle the scarcity of data annotations, which attempts to transfer mass prior knowledge learned from the labeled to unlabeled images. SSL attracts more and more attention in the field of medical image analysis in recent years." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Popular SSL medical image segmentation methods mainly focus on segmenting a single target or targets in a local region, such as segmenting pancreas or left atrium [4, 9, 14, 15, 18, 23, 31, 35, 38, 39]. Multi-organ segmentation is more challenging than single organ segmentation, due to the complex anatomical structures of the organs, e.g., the fixed relative locations (duodenum is always located at the head of the pancreas), the appearances of different organs, and the large variations of the size. Transfer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23869" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.41 + ], + "angle": 0, + "content": "ring current SSL medical segmentation methods to multi-organ segmentation encounters severe problems. Multiple organs introduce much more variance compared with a single organ. Although labeled and unlabeled images are always drawn from the same distribution, due to the limited number of labeled images, it's hard to estimate the precise distribution from them [32]. Thus, the estimated distribution between labeled and unlabeled images always suffers from mismatch problems, and is even largely increased by multiple organs. Aforementioned SSL medical segmentation methods lack the ability to handle such a large distribution gap, which requires sophisticated anatomical structure modeling. A few semi-supervised multi-organ segmentation methods have been proposed, DMPCT [43] designs a co-training strategy to mine consensus information from multiple views of a CT scan. UMCT [36] further proposes an uncertainty estimation of each view to improve the quality of the pseudo-label. Though these methods take the advantages of multi-view properties in a CT scan, they inevitably ignore the internal anatomical structures of multiple organs, resulting in suboptimal results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Teacher-student model is a widely adopted framework for semi-supervised medical image segmentation [28]. Student network takes labeled images and unlabeled strongly augmented images as input, which attempts to minimize the distribution mismatch between labeled and unlabeled images from the model level. That is, data augmentation is adopted on unlabeled data, whose role is to regularize the consistent training between teacher and student. As mentioned, semi-supervised multi-organ segmentation suffers from large distribution alignment mismatch between labeled and unlabeled images. Reducing the mismatch mainly from the model level is insufficient to solve the problem. Thanks to the prior anatomical knowledge from CT scans, which provides the distribution information where a multi-organ CT scan is drawn, it is possible to largely alleviate the mismatch problem from the data level." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To this end, we propose a novel teacher-student model, called MagicNet, matching with the rule of playing a magic-cube. More specifically, we propose a partition-and-recovery \\( N^3 \\) cubes learning paradigm: (1) We partition each CT scan, termed as magic-cube, into \\( N^3 \\) small cubes. (2) Two data augmentation strategies are then designed, as shown in Fig. 1. I.e., First, to encourage unlabeled data to learn organ semantics in relative locations from the labeled data, small cubes are mixed across labeled and unlabeled images while keeping their relative locations. Second, to enhance the learning ability for small organs, small cubes are shuffled and fed into the student network. (3) We recover the magic-cube to form the original 3D geometry to map with the ground-truth or the supervisory signal from teacher. Furthermore, the quality of pseudo labels predicted by teacher network is refined by blending with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "learned representation of the small cubes. The cube-wise pseudo-label blending strategy incorporates local attributes e.g., texture, luster and boundary smoothness which mitigates the inferior performance of small organs." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.153, + 0.882, + 0.167 + ], + "angle": 0, + "content": "The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.174, + 0.892, + 0.265 + ], + "angle": 0, + "content": "- We propose a data augmentation strategy based on partition-and-recovery \\( N^3 \\) cubes cross- and within-labeled and unlabeled images which encourages unlabeled images to learn organ semantics in relative locations from the labeled images and enhances the learning ability for small organs." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.273, + 0.892, + 0.333 + ], + "angle": 0, + "content": "- We propose to correct the original pseudo-label by cube-wise pseudo-label blending via incorporating crucial local attributes for identifying targets especially small organs." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.341, + 0.892, + 0.433 + ], + "angle": 0, + "content": "- We verify the effectiveness of our method on BTCV [13] and MACT [11] datasets. The segmentation performance of our method exceeds all state-of-the-arts by a large margin, with \\(7.28\\%\\) (10% labeled) and \\(6.94\\%\\) (30% labeled) improvement on two datasets respectively (with V-Net as the backbone) in DSC." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.174, + 0.892, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.642, + 0.46 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.89, + 0.485 + ], + "angle": 0, + "content": "2.1. Semi-supervised Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Semi-supervised medical image segmentation methods can be roughly grouped into three categories. (1) Contrastive learning based methods [35, 38], which learn representations that maximize the similarity among positive pairs and minimize the similarity among negative pairs. (2) Consistency regularization based methods [4, 9, 14, 18, 31, 36, 39], which attend to various levels of information for a single target via multi/dual-task learning or transformation consistent learning. (3) Self-ensembling/self-training based methods [1, 15, 23, 39, 43], which generate pseudolabels for unlabeled images and propose several strategies to ensure the quality of pseudo-labels. But, most of these methods are mainly focusing on segmenting one target or targets in a local region or ROI, such as pancreas or left atrium [18,23,38,39], which encounters performance degradation when transferring to multi-organ segmentation, due to the lack of anatomical structure modeling ability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.871, + 0.773 + ], + "angle": 0, + "content": "2.2. Semi-supervised Multi-organ Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Due to the large variations of appearance and size of different organs [22, 26, 33, 37], multi-organ segmentation has been a popular yet challenging task. Only a few SSL methods specially for multi-organ segmentation have been proposed. DMPCT [43] adopted a 2D-based co-training framework to aggregate multi-planar features on a private dataset. UMCT [36] further enforces multi-view consistency on unlabeled data. These methods are pioneer works" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23870" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.088, + 0.787, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.895, + 0.365 + ], + "angle": 0, + "content": "Figure 2. The architecture of MagicNet. Two essential components are proposed: (1) magic-cube partition and recovery cross- and within-image (see Sec. 3.1 and Fig. 3); (2) cube-wise pseudo-label blending, incorporating local attributes (see Sec. 3.2 and Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.471, + 0.452 + ], + "angle": 0, + "content": "to handle semi-supervised multi-organ segmentation problem from the co-training perspective, which take advantages of the multi-view property of a CT volume, but the properties of multiple organs have not been well-explored." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.471, + 0.476 + ], + "angle": 0, + "content": "2.3. Interpolation-based Semi-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.695 + ], + "angle": 0, + "content": "Interpolation-based regularization [30, 40, 41] is quite successful in semi-supervised semantic segmentation. Methods such as Mixup [41] and CutMix [40] are good at synthesizing new training samples and are widely used in semi-supervised learning. FixMatch [24] design a weak-strong pair consistency to simplify semi-supervised learning. MixMatch [3] and ReMixMatch [2] generated pseudo labels for unlabeled data and incorporate gradually unlabeled data with reliable pseudo-label into the labeled set. ICT [30] proposed a interpolated-based consistency training method based on Mixup. GuidedMix-Net [29] utilized Mixup [41] in semi-supervised semantic segmentation. Thus, we also compare our approach with some popular interpolated-based methods in Experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.169, + 0.723 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We define the 3D volume of a CT scan as \\(\\mathbf{X} \\in \\mathbb{R}^{W \\times H \\times L}\\). The goal is to find the semantic label of each voxel \\(k \\in \\mathbf{X}\\), which composes into a predicted label map \\(\\hat{\\mathbf{Y}} \\in \\{0,1,\\dots,C\\}^{W \\times H \\times L}\\), where \\(C = 0\\) indicates background class, and \\(C \\neq 0\\) represents the organ class. The training set \\(\\mathcal{D}\\) consists of two subsets: \\(\\mathcal{D} = \\mathcal{D}^l \\cup \\mathcal{D}^u\\), where \\(\\mathcal{D}^l = \\{(\\mathbf{X}_i^l, \\mathbf{Y}_i^l)\\}_{i=1}^N\\) and \\(\\mathcal{D}^u = \\{\\mathbf{X}_i^u\\}_{i=N+1}^{M+N}\\). The training images \\(\\mathbf{X}^l\\) in \\(\\mathcal{D}^l\\) are associated with per-voxel annotations \\(\\mathbf{Y}^l\\) while those in \\(\\mathcal{D}^u\\) are not. In the rest of this paper, we denote the original and mixed CT scans as magic-cubes, and denote the partitioned small cubes as cubes for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.513 + ], + "angle": 0, + "content": "The overall framework of the proposed MagicNet is shown in Fig. 2, consisting of a student network and a teacher network. The teacher network is updated via weighted combination of the student network parameters with exponential moving average (EMA) [28]. Designed in this architecture, our whole framework includes two branches: Cross-image branch and within-image branch. The details will be illustrated in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.524, + 0.809, + 0.54 + ], + "angle": 0, + "content": "3.1. Magic-cube Partition and Recovery" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Assume that a mini-batch \\(\\mathcal{B}\\) contains \\(n\\) images \\(\\mathbf{X}^{\\mathcal{B}}\\in\\) \\(\\mathbb{R}^{n\\times W\\times H\\times L}\\). For notational simplicity, let \\(n = 2\\), and \\(\\mathcal{B}\\) contains one labeled image \\(\\mathbf{X}^l\\) and one unlabeled image \\(\\mathbf{X}^u\\), which are randomly sampled from \\(\\mathcal{D}^l\\) and \\(\\mathcal{D}^u\\), respectively. We partition \\(\\mathbf{X}^l\\) and \\(\\mathbf{X}^u\\) into \\(N^3\\) magic-cubes \\(\\{\\mathbf{x}_j^u\\}_{j = 1}^{N^3}\\) and \\(\\{\\mathbf{x}_j^l\\}_{j = 1}^{N^3}\\), where \\(\\mathbf{x}_j^l,\\mathbf{x}_j^u\\in \\mathbb{R}^{W / N\\times H / N\\times L / N}\\), \\(j\\) represents the relative location of the cube inside this magic-cube. This partition operation is called as \\(\\mathcal{A}_{part}(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Cross-image Partition and Recovery To encourage labeled and unlabeled images to learn comprehensive common semantics from each other, we mix these cubes across all the labeled and unlabeled images in a mini-batch. As shown in Fig. 3, these cubes \\(\\mathbf{x}_j^l\\) and \\(\\mathbf{x}_j^u\\) are mixed into two shuffled magic-cubes while keeping their original positions, which produce two interpolated images with mixed magic-cubes \\(\\mathbf{X}_0^{mix},\\mathbf{X}_1^{mix}\\in \\mathbb{R}^{W\\times H\\times L}\\). The magic-cube mixing cross-image operations are termed as \\(\\mathcal{A}_{\\text{mix}}^{\\text{cross}}(\\cdot)\\). The mixed images are then fed into the student network \\(\\mathcal{F}(\\cdot;\\Theta^s)\\), followed by a softmax layer \\(\\sigma(\\cdot)\\), obtaining the prediction maps \\(\\mathbf{P}_0^{mix},\\mathbf{P}_1^{mix}\\in \\mathbb{R}^{C\\times W\\times H\\times L}\\), where \\(C\\) denotes the number of classes. Next, we recover \\(\\mathbf{P}_{\\text{cross}}^l,\\mathbf{P}_{\\text{cross}}^u\\in \\mathbb{R}^{C\\times W\\times H\\times L}\\) from \\(\\mathbf{P}_0^{mix}\\) and \\(\\mathbf{P}_1^{mix}\\) via recovering the magic-cubes back to the original positions in their original" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.835, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Magic-cube partition and recovery. Blue and green volumes represent labeled and unlabeled image, respectively. The number on cubes represents its relative-location in the original image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.471, + 0.33 + ], + "angle": 0, + "content": "images, and this operation is denoted as \\(\\mathcal{A}_{rec}^{cross}(\\cdot)\\). We can simply denote the whole process as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.335, + 0.47, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\text {c r o s s}} ^ {l}, \\mathbf {P} _ {\\text {c r o s s}} ^ {u} = \\mathcal {A} _ {\\text {r e c}} ^ {\\text {c r o s s}} \\left(\\sigma \\left(\\mathcal {F} \\left(\\mathcal {A} _ {\\text {m i x}} ^ {\\text {c r o s s}} \\left(\\mathcal {A} _ {\\text {p a r t}} (\\mathbf {X} ^ {\\mathcal {B}})\\right); \\boldsymbol {\\Theta} ^ {s}\\right)\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.387 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{P}_{cross}^{l},\\mathbf{P}_{cross}^{u}\\in \\mathbb{R}^{(C + 1)\\times W\\times H\\times L}\\). The loss functions for labeled and unlabeled images are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.394, + 0.47, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c r o s s}} ^ {l} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {\\text {d i c e}} \\left(\\mathbf {P} _ {\\text {c r o s s}} ^ {l}, \\mathbf {Y} ^ {l}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.416, + 0.47, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c r o s s - i n} ^ {u} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {d i c e} \\left(\\mathbf {P} _ {c r o s s} ^ {u}, \\widehat {\\mathbf {Y}} ^ {u}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.447, + 0.47, + 0.52 + ], + "angle": 0, + "content": "where \\(\\ell_{\\text{dice}}\\) denotes multi-class Dice loss. \\(\\mathbf{Y}^l \\in \\{0,1,\\dots,C\\}^{W \\times H \\times L}\\) denotes per-voxel manual annotations for \\(\\mathbf{X}^l\\). \\(\\mathcal{L}_{\\text{cross-in}}^u\\) indicates the refined pseudo-labels for \\(\\mathbf{X}^u\\) by blending cube-wise representations and more details can be seen in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.471, + 0.751 + ], + "angle": 0, + "content": "Within-image Partition and Recovery Besides magic-cubes partition and recovery across images, we also design a within-image partition and recovery branch for single images, which can better consider local features and learn local attributes for identifying targets, especially targets of small sizes. For \\(\\mathbf{X}^l\\), the partitioned \\(j\\)-th cube \\(\\mathbf{x}_j^l\\) is fed into \\(\\mathcal{F}(\\cdot; \\Theta^s)\\) and the softmax layer, which can acquire a cubewise prediction map \\(\\mathbf{p}_j^l \\in \\mathbb{R}^{C \\times W / N \\times H / N \\times L / N}\\). Finally, we recover the magic-cube via mixing \\(N^3\\) cube-wise prediction maps back to their original positions as illustrated in Fig. 3, and the recovered probability map is denoted as \\(\\mathbf{P}_{in}^l \\in \\mathbb{R}^{C \\times W \\times H \\times L}\\). Similarly, for \\(\\mathbf{X}^u\\), the probability map is denoted as \\(\\mathbf{P}_{in}^u \\in \\mathbb{R}^{C \\times W \\times H \\times L}\\). This operation is denoted as \\(\\mathcal{A}_{rec}^{in}(\\cdot)\\). The whole process for obtaining \\(\\mathbf{P}_{in}^l\\) and \\(\\mathbf{P}_{in}^u\\) are:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.759, + 0.47, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {i n} ^ {l}, \\mathbf {P} _ {i n} ^ {u} = \\mathcal {A} _ {r e c} ^ {i n} (\\sigma (\\mathcal {F} (\\mathcal {A} _ {p a r t} (\\mathbf {X} ^ {\\mathcal {B}}); \\boldsymbol {\\Theta} ^ {s}))). \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.788, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Since \\(\\mathbf{P}_{in}^{l}\\) and \\(\\mathbf{P}_{in}^{u}\\) are recovered via predictions of cubes, \\(\\mathbf{P}_{in}^{l}\\) and \\(\\mathbf{P}_{in}^{u}\\) are denoted as cube-wise representations in the rest of the paper. The loss function for the labeled image is:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.843, + 0.47, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n} ^ {l} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {d i c e} \\left(\\mathbf {P} _ {i n} ^ {l}, \\mathbf {Y} ^ {l}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For the unlabeled image, we propose to leverage the cubewise representations by refining the pseudo-label instead of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.893, + 0.36 + ], + "angle": 0, + "content": "directly computing the loss function as in Eq. 5, which is illustrated in Sec. 3.2. We conduct ablation studies on various design choices of utilizing the cube-wise representations in the experiment part." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.369, + 0.794, + 0.385 + ], + "angle": 0, + "content": "3.2. Cube-wise Pseudo-label Blending" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.893, + 0.602 + ], + "angle": 0, + "content": "As mentioned in the previous section, medical imaging experts reveal that local attributes e.g., texture, luster and boundary smoothness are crucial elements for identifying targets such as tiny organs in medical images [42]. Inspired by this, we propose a within-image partition and recovery module to learn cube-wise local representations. Since the teacher network takes the original volume as input, and pays more attention to learn large organs in the head class, voxels that actually belong to the tail class can be incorrectly predicted as head class voxels, due to the lack of local attributes learning. To effectively increase the chance of the voxel predicted to the tail class, we design a cube-wise pseudolabel blending module, which blends the original pseudolabel with cube-wise features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.74 + ], + "angle": 0, + "content": "In detail, the image-level prediction map of \\(\\mathbf{X}^u\\) acquired from the teacher network is defined as \\(\\mathbf{P}_T^u = \\mathcal{F}(\\mathbf{X}^u;\\boldsymbol{\\Theta}^t)\\), where \\(\\boldsymbol{\\Theta}^t\\) is the parameters of teacher network. The reconstructed cube-level representations of \\(\\mathbf{X}^u\\) acquired from the student network is \\(\\mathbf{P}_{in}^u\\). As mentioned above, cubewise features pay more attention to learn local attributes which is important for tiny organs. Therefore, we propose a distribution-aware blending strategy, which is shown in Fig. 4(a), and formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.751, + 0.892, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\text {b l e n d}} ^ {u} = \\left(\\mathbf {1} - \\mathcal {R} (\\boldsymbol {\\Omega})\\right) \\odot \\mathbf {P} _ {T} ^ {u} + \\mathcal {R} (\\boldsymbol {\\Omega}) \\odot \\mathbf {P} _ {\\text {i n}} ^ {u}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.78, + 0.892, + 0.826 + ], + "angle": 0, + "content": "where \\(\\odot\\) indicates the element-wise multiplication, \\(\\mathbf{1} \\in \\{1\\}^{(C + 1) \\times W \\times H \\times L}\\), and \\(\\Omega \\in \\mathbb{R}^{W \\times H \\times L}\\) is a distribution-aware weight map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.902 + ], + "angle": 0, + "content": "To obtain the weight map, we firstly learn the class-wise distribution \\(\\mathbf{v}\\) during training. Suppose \\(\\mathbf{v} \\in \\mathbb{R}^C\\) is a vector, \\(\\mathbf{v} = \\{\\mathbf{v}_0, \\dots, \\mathbf{v}_{C-1}\\}\\), whose element \\(\\mathbf{v}_c\\) indicates the number of voxels belonging to \\((c+1)\\)th organ, which is accumulated by counting the voxels w.r.t. pseudo-labels over a few previ" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.221, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.196, + 0.144, + 0.206 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.096, + 0.338, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.096, + 0.468, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.195, + 0.356, + 0.207 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 4. The cube-wise pseudo-label blending module. (a) Pipeline. (b) Illustration on \\(\\mathcal{R}(\\Omega_{\\mathbf{m}})\\). If the pseudo-label \\(\\tilde{\\mathbf{Y}}_{\\mathbf{m}}^{u}\\) on \\(\\mathbf{m}\\) is incorrectly assigned as a big organ, a big \\(\\Omega_{\\mathbf{m}}\\) is assigned to ensure possible correction from cube-wise representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.469, + 0.33 + ], + "angle": 0, + "content": "ous iterations. Let \\(\\widetilde{\\mathbf{Y}}_{\\mathbf{m}}^{u}\\) denote the pseudo-label for the voxel on location m. The value \\(\\Omega_{\\mathbf{m}}\\in \\mathbb{R}^{1}\\) at each spatial location m on the weight map is derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.341, + 0.469, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Omega} _ {\\mathbf {m}} = \\sum_ {c = 0} ^ {C - 1} \\frac {\\mathbf {I} \\left(\\widetilde {\\mathbf {Y}} _ {\\mathbf {m}} ^ {u} = c\\right) \\mathbf {v} _ {c}}{\\operatorname* {m a x} \\mathbf {v}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.469, + 0.438 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}(\\cdot)\\) is an indicator function. Note that to match with the data dimension in Eq. 6, we replicate \\(\\Omega C + 1\\) times to acquire \\(\\mathcal{R}(\\Omega) \\in \\mathbb{R}^{(C + 1) \\times W \\times H \\times L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.575 + ], + "angle": 0, + "content": "Overall, since the teacher network takes the original volume as input, and pays more attention to learn large organs in the head class, the teacher pseudo-label may be biased. To unbiased the pseudo-label, on the one hand, we keep the pseudo-label of small organs. On the other hand, we remedy the possible incorrect pseudo-labels of big organs via changing them to small-organ classes via Eq. 6 if \\(\\Omega_{\\mathbf{m}}\\) is large. This can effectively increase the chance of the voxels predicted to small-organs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.469, + 0.622 + ], + "angle": 0, + "content": "The final refined pseudo-label \\(\\widehat{\\mathbf{Y}}^u\\) is obtained via \\(\\widehat{\\mathbf{Y}}_{\\mathbf{m}}^{u} = \\arg \\max_{c\\in \\mathcal{C}}\\mathbf{P}_{blend,c,\\mathbf{m}}^{u}\\), where \\(\\mathcal{C}\\) indicates the finite set of class labels, and \\(\\mathbf{m}\\) means the voxel location." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.362, + 0.65 + ], + "angle": 0, + "content": "3.3. Magic-cube Location Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To fully leverage the prior anatomical knowledge of multi-organs, we further propose a magic-cube location reasoning method, which is in line with within-image partition. For one image \\(\\mathbf{X}^l\\), the partitioned \\(j\\)-th cube's \\(\\mathbf{x}_j^l\\) is fed into the encoder of student network \\(\\mathcal{F}_{enc}(\\cdot;\\Theta_{enc}^s)\\) to generate \\(\\mathbf{f}_j^l \\in \\mathbb{R}^{D \\times W / N \\times H / N \\times L / N}\\). Then, the feature \\(\\mathbf{f}_j^l\\) is flattened into a one-dimensional vector, passing through a classification head \\(\\mathcal{F}_{cls}(\\cdot;\\Theta_{cls})\\) to project the flattened feature vector into the same size as the number of cubes in one image, i.e., \\(N^3\\). The classification head is composed of two fully connected layers [10, 34]. Finally, we concatenate the output \\(\\hat{\\mathbf{y}}_j^l \\in \\mathbb{R}^{N^3}\\) from the classification head for all \\(N^3\\) cubes in one image, which generates a vector in size \\((N^3)^2\\). This vector is further reshaped into a matrix of size \\(N^3 \\times N^3\\), whose rows indicate the probabilities of corresponding cubes belonging to the \\(N^3\\) locations. The same" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.891, + 0.122 + ], + "angle": 0, + "content": "applies to \\(\\mathbf{X}^u\\). For both \\(\\mathbf{X}^l\\) and \\(\\mathbf{X}^u\\) in \\(\\mathcal{B}\\), the cross-entropy loss is adopted to learn relative locations of cubes:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.135, + 0.892, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} \\left(\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}\\right) = \\frac {1}{| \\mathcal {B} |} \\sum_ {\\mathbf {X} \\in \\mathcal {B}} \\ell_ {c e} (\\sigma \\left(\\mathcal {F} _ {c l s} \\right. \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.175, + 0.84, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathcal {F} _ {e n c} ^ {s} \\left(\\mathcal {A} _ {p a r t} (\\mathbf {X}); \\boldsymbol {\\Theta} _ {e n c} ^ {s}\\right); \\boldsymbol {\\Theta} _ {c l s}\\right)), \\mathbb {y})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.203, + 0.891, + 0.232 + ], + "angle": 0, + "content": "where \\(\\ell_{ce}\\) denotes the cross-entropy loss, y represents cube's relative locations in \\(\\mathbf{X}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.242, + 0.644, + 0.258 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.266, + 0.891, + 0.296 + ], + "angle": 0, + "content": "Overall, we develop a total loss function based on a minibatch for our MagicNet as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.307, + 0.891, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\mathcal {B}; \\Theta_ {\\text {e n c}} ^ {s}, \\Theta_ {\\text {s e g}} ^ {s}, \\Theta_ {\\text {c l s}}) = \\mathcal {L} _ {\\sup } (\\mathcal {B}; \\Theta_ {\\text {e n c}} ^ {s}, \\Theta^ {s}, \\Theta_ {\\text {c l s}}) \\tag {9} \\\\ + \\mathcal {L} _ {u n s u p} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.356, + 0.891, + 0.386 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{sup}(\\cdot)\\) and \\(\\mathcal{L}_{unsup}(\\cdot)\\) denote the loss for labeled and unlabeled images in \\(\\mathcal{B}\\), respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.397, + 0.891, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\sup } (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}) = \\mathcal {L} _ {c r o s s} ^ {l} (\\mathcal {B}; \\Theta^ {s}) \\tag {10} \\\\ + \\mathcal {L} _ {i n} ^ {l} (\\mathcal {B}; \\Theta^ {s}) + \\beta \\mathcal {L} _ {c l s} ^ {l} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.891, + 0.477 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the weight factor to balance the location reasoning term with others, and" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.488, + 0.891, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {u n s u p} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}) = \\alpha \\mathcal {L} _ {c r o s s - i n} ^ {u} (\\mathcal {B}; \\Theta^ {s}) \\tag {11} \\\\ + \\beta \\mathcal {L} _ {c l s} ^ {u} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.891, + 0.578 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is another weight factor to balance the two terms. For \\(\\mathcal{L}_{cls}^{l}\\) and \\(\\mathcal{L}_{cls}^{u}\\) in Eq. 10 and Eq. 11, the following equation holds:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.579, + 0.891, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = \\mathcal {L} _ {c l s} ^ {l} + \\mathcal {L} _ {c l s} ^ {u}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.641, + 0.621 + ], + "angle": 0, + "content": "3.5. Testing Phase" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.891, + 0.704 + ], + "angle": 0, + "content": "Given a testing image \\(\\mathbf{X}_{test}\\), the probability map is obtained by: \\(\\mathbf{P}_{test} = \\sigma (\\mathcal{F}(\\mathbf{X}_{test};\\Theta^{s*}))\\), where \\(\\Theta^{s*}\\) denotes the well-trained student network parameters. The final label map can be determined by taking the argmax in the first dimension of \\(\\mathbf{P}_{test}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.632, + 0.734 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.753, + 0.759 + ], + "angle": 0, + "content": "4.1. Datasets and Pre-processing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "BTCV Multi-organ Segmentation Dataset BTCV multi-organ segmentation dataset is from MICCAI Multi-Atlas Labeling Beyond Cranial Vault-Workshop Challenge [13] which contains 30 subjects with 3779 axial abdominal CT slices with 13 organs annotation. In pre-processing, we follow [27] to re-sample all the CT scans to the voxel spacing \\([1.5 \\times 1.5 \\times 2.0]mm^3\\) and normalize them to have zero mean and unit variance. Strictly following [5, 6, 10], 18 cases are divided for training and the other 12 cases are for testing." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23873" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.146, + 0.089, + 0.825, + 0.277 + ], + "angle": 0, + "content": "
Labeled%MethodsSplR.kidL.kidGallEsoLivStoAorIVCVeinsPanRGLGAvg. DSCAvg. NSD
30%V-Net [20]86.3775.3871.2041.8148.4692.7744.7888.2072.6049.2722.3037.7921.7157.9056.90
MT [28]79.8977.5678.0838.3158.9992.2648.7388.6179.3652.7328.4254.1621.3061.4260.85
UA-MT [39]87.7573.8772.9742.0850.7993.2242.0788.3071.8055.2736.6143.2237.6761.2060.99
ICT [30]88.7878.0083.9350.0953.3588.3264.1989.0279.6063.5158.9043.2950.7068.5967.77
CPS [7]82.9876.9179.2129.9152.5392.5941.3786.5974.7349.5318.9645.1324.0658.0456.64
SS-Net [35]85.5373.3280.7030.8552.4491.6717.6786.2872.3947.4022.9537.3839.0156.7454.38
SLC-Net [16]90.0484.3885.8453.5555.8792.8657.8089.8380.8753.2238.7848.1037.0366.7868.81
MagicNet (ours)91.4284.6486.1962.8662.4993.8972.8790.7083.5270.0764.9460.8857.4875.5376.31
40%V-Net [20]84.9882.7282.0736.6463.4893.5457.4989.7478.6360.4249.3955.6038.4967.1767.84
MT [28]85.7078.9379.0842.8061.0993.4557.5789.7080.3063.9541.1450.4629.6965.6865.98
UA-MT [39]88.7475.8878.9154.2558.5593.4658.9089.2376.1562.3047.9151.5344.9267.7568.87
ICT [30]90.3184.4186.9649.2265.6594.2965.9590.2381.4469.5666.6157.3556.0173.6974.98
CPS [7]87.5672.9977.5953.3154.0892.4154.5887.7574.3258.6848.0250.3943.8665.8165.34
SS-Net [35]84.7476.3774.1943.4257.0592.9014.3783.1469.7752.4527.0854.2927.6658.2657.75
SLC-Net [16]90.0584.0086.4356.1658.9194.6870.7289.9379.4560.5954.2251.0339.0870.4072.87
MagicNet (ours)91.6185.0288.1358.1666.7294.0774.4690.7784.3171.5668.9063.4860.4776.7478.68
100%V-Net [20]84.0084.8286.3867.4265.0294.8373.7590.2784.1969.8563.5462.6065.0276.2877.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.288, + 0.892, + 0.359 + ], + "angle": 0, + "content": "Table 1. Comparison results (DSCs of each organ, avg. DSC, and avg. NSD) between our method and existing semi-supervised medical image segmentation methods on the BTCV dataset. V-Net means training with only labeled samples on the V-Net backbone. DSC: Dice Similarity Coefficient. NSD: Normalized Surface Dice. The standard deviations are not reported due to space limit. Note: Spl: spleen, R.Kid: right kidney, L.Kid: left kidney, Gall: gallbladder, Eso: esophagus, Liv: liver, Sto: stomach, Aor: aorta, IVC: inferior vena cava, Veins: portal and splenic veins, Pan: pancreas, LG/RG: left/right adrenal glands." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.372, + 0.825, + 0.552 + ], + "angle": 0, + "content": "
Labeled%MethodsSpleenL.kidneyGallbladderEsophagusLiverStomachPancreasDuodenumAvg. DSCAvg. NSD
10%V-Net [20]88.89 (16.31)84.90 (20.11)55.73 (34.27)58.27 (20.05)93.15 (7.11)57.15 (30.62)53.33 (22.44)35.28 (19.83)65.8449.96
MT [28]89.22 (13.56)88.04 (17.61)61.40 (31.91)58.89 (20.62)93.60 (5.04)74.00 (20.37)64.70 (19.31)44.53 (17.61)71.8053.54
UA-MT [39]89.94 (13.36)89.26 (15.67)59.19 (32.35)59.43 (19.45)93.77 (5.05)75.43 (19.61)65.86 (17.48)44.57 (17.40)72.1854.04
ICT [30]90.27 (13.70)89.89 (15.81)61.47 (30.31)58.70 (20.17)93.47 (5.39)72.45 (20.95)65.81 (18.28)45.00 (19.22)72.1354.07
CPS [7]91.73 (12.11)90.59 (14.85)60.54 (32.29)55.82 (21.03)94.19 (4.93)78.80 (15.28)62.93 (20.03)32.58 (18.68)70.9052.80
SS-Net [35]83.52 (19.56)78.28 (25.05)53.08 (33.88)56.24 (20.54)91.84 (6.07)47.65 (29.70)43.23 (24.11)25.70 (18.71)59.9442.68
SLC-Net [16]86.89 (17.54)87.54 (18.00)55.28 (34.75)58.51 (20.54)93.70 (5.33)69.74 (22.89)46.55 (26.21)28.86 (18.87)65.8913.28
MagicNet (ours)92.33 (11.12)91.19 (14.18)66.35 (29.29)70.95 (12.30)94.31 (4.64)81.56 (14.91)76.31 (10.56)61.50 (12.85)79.3162.31
20%V-Net [20]90.53 (15.33)88.28 (19.76)65.16 (30.03)66.10 (16.98)94.49 (4.18)79.56 (16.56)69.04 (17.43)51.98 (18.53)75.6459.66
MT [28]91.76 (12.77)91.40 (14.49)63.83 (31.74)64.15 (18.05)94.44 (4.86)81.19 (15.79)72.24 (15.79)57.01 (16.70)77.0060.32
UA-MT [39]92.39 (11.31)91.44 (14.61)63.63 (31.46)64.33 (17.92)94.50 (4.77)83.21 (14.03)71.93 (15.51)57.21 (17.35)77.3360.64
ICT [30]92.96 (9.60)91.67 (14.52)65.90 (30.49)64.35 (16.79)94.46 (4.66)82.60 (15.46)74.18 (12.99)57.67 (16.16)77.9761.24
CPS [7]92.66 (11.46)91.87 (14.33)64.75 (31.21)57.27 (19.56)94.97 (4.51)85.65 (9.50)74.15 (12.49)55.00 (16.58)77.0460.01
SS-Net [35]91.17 (14.77)87.77 (19.37)63.49 (30.83)65.02 (16.65)93.76 (4.82)73.04 (22.16)70.37 (17.70)52.81 (20.32)74.6857.57
SLC-Net [16]92.60 (11.24)91.38 (14.57)62.46 (31.37)62.84 (19.31)94.54 (4.72)80.30 (14.73)69.79 (16.99)53.48 (18.40)75.9259.93
MagicNet (ours)93.52 (10.22)92.01 (14.30)71.04 (26.92)70.95 (12.58)94.89 (4.42)85.77 (10.27)78.81 (8.07)66.21 (12.33)81.6565.87
100%V-Net [20]94.35 (5.81)92.30 (14.22)73.61 (28.80)73.48 (12.43)95.17 (3.10)89.09 (7.74)80.86 (7.38)70.03 (11.60)83.6168.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.563, + 0.89, + 0.591 + ], + "angle": 0, + "content": "Table 2. Comparison results (DSCs of each organ, avg. DSC, and avg. NSD) between our method and existing semi-supervised medical image segmentation methods on MACT dataset under 4-fold cross-validation. The value in \\((\\cdot)\\) is standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.473, + 0.807 + ], + "angle": 0, + "content": "MACT Dataset Multi-organ abdominal CT reference standard segmentations (MACT) is a public multi-organ segmentation dataset, containing 90 CT volumes with 8 organs annotation. It is manually re-annotated by [11] based on 43 cases of NIH pancreas [21] and 47 cases of Synapse [13]. We follow the setting of UMCT [36]. 1) In preprocessing, we set the soft tissue CT window range of \\([-125, 275]\\) HU and all the CT scans are re-sampled to an isotropic resolution of \\(1.0\\mathrm{mm}^3\\). Then, image intensities are normalized to have zero mean and unit variance. 2) We randomly split the dataset into four folds, and perform 4-fold cross-validation. Labeled training set is then randomly selected, which are applied to all experiments for fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.452, + 0.833 + ], + "angle": 0, + "content": "4.2. Experimental Setup and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this work, we conduct all the experiments based on PyTorch with one NVIDIA 3090 GPU and use V-Net as our backbone [16,35,39]. For magic-cube location reasoning, we add a classification head composed of two fully" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.897 + ], + "angle": 0, + "content": "connected layers at the end of our encoder. Firstly, for BTCV dataset, the framework is trained by an SGD optimizer for 70K iterations, with an initial learning rate (lr) 0.01 with a warming-up strategy: \\( lr = \\text{base\\_lr} \\times (1 - \\frac{\\text{iterations}}{\\text{max\\_iterations}})^{0.9} \\). For MACT dataset, the framework is trained by an SGD optimizer for 30K iterations, with an initial learning rate (lr) 0.01 decayed by 0.1 every 12K iterations. Following [16, 18, 35, 36, 39], for both two datasets, batch size is set as 4, including 2 labeled images and 2 unlabeled images, and the hyper-parameter \\( \\alpha \\) is set using a time-dependent Gaussian warming-up function and \\( \\beta \\) is empirically set as 0.1 [10]. We randomly crop volume of size \\( 96^3 \\), and then we partition the volume into \\( N^3 \\) small cubes, forming our magic-cubes, and the choice of \\( N \\) is also discussed in Sec. 4.4. In the testing phase, a sliding window strategy is adopted to obtain final results with a stride of \\( 16^3 \\). The final evaluation metrics for our method are DSC (\\%, Dice-Sørensen Coefficient) and NSD (\\%, Normalized Surface Dice) which are commonly used in the cur" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23874" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.09, + 0.849, + 0.173 + ], + "angle": 0, + "content": "
MethodsCrossInLocBldAvg. DSCspleenl.kidneygallbladderesophagusliverstomachpancreasduodenum
Baseline (MT [28])71.8886.8488.9254.5053.4093.0170.8671.7755.76
Cross✓78.1392.1190.5363.4270.0593.9079.6075.5859.87
Cross + In✓✓78.8392.8290.7564.9170.6893.9080.9276.2560.40
Cross + Loc✓✓78.5692.4991.0765.9569.7294.1280.4075.8258.94
Cross + In + Loc✓✓✓78.8792.5591.1865.7270.7794.4280.0575.8160.43
Cross + In + Loc + Bld✓✓✓✓79.3192.3391.1966.3570.9594.3181.5676.3161.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.184, + 0.89, + 0.213 + ], + "angle": 0, + "content": "Table 3. Ablation study (DSC %) for the effectiveness of each component of MagicNet. Cross: cross-image partition-and-recovery. In: Within-image partition-and-recovery. Bld: cube-wise pseudo-label blending. Loc: magic-cube location reasoning." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.236, + 0.467, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.342, + 0.468, + 0.371 + ], + "angle": 0, + "content": "Figure 5. DSC comparison in box plots on MACT dataset with \\(10\\%\\) labeled images (Best viewed electronically, zoom in)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.386, + 0.405, + 0.402 + ], + "angle": 0, + "content": "rent multi-organ segmentation challenge [12, 19]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.412, + 0.468, + 0.428 + ], + "angle": 0, + "content": "4.3. Comparison with the State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.468, + 0.583 + ], + "angle": 0, + "content": "We compare MagicNet with six state-of-the-art semi-supervised segmentation methods, including mean-teacher (MT) [28], uncertainty-aware mean-teacher (UA-MT) [39], interpolated consistency training (ICT) [30], cross pseudosupervision (CPS) [7], smoothness and class-separation (SS-Net) [35], shape-awareness and local constraints (SLC-Net) [16]. All experiments of other methods are re-trained by their official code when transferred to multi-organ task. V-Net is adopted as our backbone which is the same as the above methods for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.468, + 0.765 + ], + "angle": 0, + "content": "In Table 1, we summarize the results of BTCV dataset. Visualizations compared with state-of-the-arts are shown in the supplementary material. Noticeable improvements compared with state-of-the-arts can be seen for organs such as Stomach (↑ 8.68), Gallbladder (↑ 9.31), Veins (↑ 6.56), Right adrenal glands (↑ 6.72) and Left adrenal glands (↑ 6.78). It is interesting to observe that some semi-supervised methods even perform worse than the lower bound. This is because these methods are less able to learn common semantics from the labeled images to unlabeled ones, generating less accurate pseudo-labels, which even hamper the usage of unlabeled images (see the supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.902 + ], + "angle": 0, + "content": "We then evaluate MagicNet on MACT dataset. In Table 2, we can observe: (1) a more significant performance improvement between MagicNet and other state-of-the-arts with \\(10\\%\\) labeled data than \\(20\\%\\), which demonstrates the effectiveness of MagicNet when training with fewer labeled CT images; (2) MagicNet successfully mitigates the inferior performance in tail classes, e.g., DSC improvement with \\(10\\%\\) labeled images on Gallbladder (\\(\\uparrow\\) 4.88), Esophagus (\\(\\uparrow\\) 11.52), Pancreas (\\(\\uparrow\\) 10.45) and Duodenum (\\(\\uparrow\\) 16.50). Fig. 5" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.235, + 0.866, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.312, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Figure 6. Scramble position and keep position when mixing small-cubes across images (from Fig. 3). Red arrow indicates the cube location in mixed images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.453 + ], + "angle": 0, + "content": "shows a comparison in DSC of six methods for different organs. Note that UMCT [36] on MACT dataset achieves \\(75.33\\%\\) in DSC. It uses AH-Net [17] as the backbone (3 are used), whose parameter number is much larger. Due to computation resources, we do not directly compare with it." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.458, + 0.653, + 0.474 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We conduct ablation studies on MACT dataset [11] with \\(10\\%\\) labeled data to show the effectiveness of each module. The effectiveness of each component in MagicNet We conduct ablation studies to show the impact of each component in MagicNet. In Table 3, the first row indicates the mean-teacher baseline model [28], which our method is designed on. Compared to the baseline, our magic-cube partition and recovery can yield good segmentation performance. Cross and In represent our cross- and within-image partition-and-recovery, which increase the performance from \\(71.88\\%\\) to \\(78.13\\%\\) and \\(78.87\\%\\), respectively. This shows the powerful ability with our specially-designed data augmentation method. Loc represents magic-cube location reasoning for within-image branch. We can see from the table that based on our cross-image branch, adding relative locations of small-cubes can achieve \\(78.56\\%\\) in DSC, adding only within-image partition-and-recovery branch achieves \\(78.83\\%\\), and adding both two branches leads to \\(78.87\\%\\). Finally, our proposed cube-wise pseudo label blending (short for Bld in Table 3) provides significant improvement to \\(79.31\\%\\), for which we will make a more complete analysis later in the current section." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Design choices of partition and recovery Here, we discuss the design for cross-image partition-and-recovery branch: ① Should we maintain or scramble the magic-cube relative locations when manipulating our cross-image branch, as shown in Fig. 6? The comparison results are shown in the last two rows in Table 4. Scramble and Keep represent the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23875" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.093, + 0.351, + 0.155 + ], + "angle": 0, + "content": "
①②DSCNSD
scrambleU69.53 ± 11.2750.98 ± 11.00
keepU73.84 ± 9.9655.80 ± 10.81
scrambleLU71.91 ± 11.2054.10 ± 11.28
keepLU78.13 ± 8.1260.79 ± 9.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.165, + 0.354, + 0.249 + ], + "angle": 0, + "content": "Table 4. Ablation of design choices for crossimage partition and recovery (Question \\(①\\) and \\(②\\), mean \\(\\pm\\) std of all cases). Scramble/keep: ignore/keep original positions when mixed. U: only for unlabeled data. LU: for both labeled and unlabeled data. The last row is ours." + }, + { + "type": "table", + "bbox": [ + 0.366, + 0.092, + 0.625, + 0.156 + ], + "angle": 0, + "content": "
DSCNSD
teacher sup.78.47 ± 7.8361.52 ± 9.21
mutual sup.70.45 ± 11.4753.71 ± 11.78
blending79.31 ± 7.5562.31 ± 9.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.166, + 0.627, + 0.222 + ], + "angle": 0, + "content": "Table 5. Comparison of different pseudolabel supervision/blending strategies for unlabeled data. Sup.: supervision. Blending: our cube-wise pseudo-label blending." + }, + { + "type": "table", + "bbox": [ + 0.639, + 0.089, + 0.89, + 0.16 + ], + "angle": 0, + "content": "
DSCNSD
CutMix [40]73.80 ± 9.9955.86 ± 10.54
CutOut [8]72.47 ± 10.2754.62 ± 10.67
MixUp [41]72.13 ± 11.3254.07 ± 12.01
Ours (2)78.68 ± 7.4560.78 ± 9.12
Ours (3)79.31 ± 7.5562.31 ± 9.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.634, + 0.17, + 0.892, + 0.239 + ], + "angle": 0, + "content": "Table 6. Comparison of different data augmentation methods, where we try different CutMix and CutOut sizes, and choose the best results. For MagicNet, we compare different \\(N\\) values in \\((\\cdot)\\)." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.275, + 0.468, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.432, + 0.471, + 0.515 + ], + "angle": 0, + "content": "Figure 7. Kernel Density Estimation of labeled and unlabeled samples on the first three channels of the feature map of \"only unlabeled images\" and \"both labeled and unlabeled images\". Considerable distribution mismatch between labeled and unlabeled images can be observed (top), while our method can well improve the distribution mismatch (bottom). Best viewed zoom in." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.686 + ], + "angle": 0, + "content": "partitioned small-cubes are randomly mixed while ignoring their original locations or kept when mixing them across images. Results show that the relative locations between multiple organs are important for multi-organ segmentation. ② Should our cross-image data augmentation be operated on only unlabeled images (see U in Table 4) or both labeled and unlabeled images (see LU in Table 4)? The latter obtains a much better performance compared to the former. Besides, it largely reduces the distribution mismatch between labeled and unlabeled data, as shown in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Cube-wise pseudo-label blending As illustrated in Sec. 3.2, we blend the output of within-image branch and the output of teacher model to obtain the final pseudo-label for complementing local attributes. We compare our blending with two other methods, as shown in Table 5. Three supervision schemes, illustrated in Fig. 8, are compared for unlabeled images based on our framework. Teacher sup. means the outputs of cross- and within-image branch are both supervised by the pseudo-label from the teacher model. Mutual sup. means the outputs of cross- and within-image are mutually supervised. It can be seen that our blending strategy works favorably for unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Comparison with interpolated-based methods As shown in Table 6, our augmentation method outperforms other" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.271, + 0.821, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.353, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Figure 8. Different pseudo-label supervision/blending strategies corresponding to Table 5. Blue line indicates supervision. (a) teacher sup. (b) mutual sup. (c) our blending." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.459 + ], + "angle": 0, + "content": "methods such as CutMix, CutOut and MixUp. For fair comparison, we try several CutMix and CutOut sizes, and choose the best results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.893, + 0.552 + ], + "angle": 0, + "content": "Different number \\((N)\\) of small-cubes We study the impact of different numbers of small-cubes \\(N\\), as shown in Table 6. Slightly better performance is achieved when \\(N = 3\\). When \\(N = 4\\), the size of the small cube does not match the condition for VNet. Thus, we only compare the results given \\(N = 2\\) and \\(N = 3\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.757, + 0.58 + ], + "angle": 0, + "content": "5. Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.74 + ], + "angle": 0, + "content": "We have presented MagicNet for semi-supervised multi-organ segmentation. MagicNet encourages unlabeled data to learn organ semantics in relative locations from the labeled data, and enhances the learning ability for small organs. Two essential components are proposed: (1) partition-and-recovery \\( N^3 \\) small cubes cross- and within-labeled and unlabeled images, and (2) the quality of pseudo label refinement by blending the cube-wise representations. Experimental results on two public multi-organ segmentation datasets verify the superiority of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Limitations: CT scans are all roughly aligned by the nature of the data. Magic-cube may not work well on domains \\( w / o \\) such prior, e.g., microscopy images \\( w / o \\) similar prior." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.681, + 0.817 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (Grant No. 62101191), Shanghai Natural Science Foundation (Grant No. 21ZR1420800), and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ2229004)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23876" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Wenjia Bai, Ozan Oktay, Matthew Sinclair, Hideaki Suzuki, Martin Rajchl, Giacomo Tarroni, Ben Glocker, Andrew P. King, Paul M. Matthews, and Daniel Rueckert. Semi-supervised learning for network-based cardiac MR image segmentation. In Proc. MICCAI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[2] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In Proc. ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.248, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[3] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Proc. NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.306, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[4] Gerda Bortsova, Florian Dubost, Laurens Hogeweg, Ioannis Katramados, and Marleen de Bruijne. Semi-supervised medical image segmentation via learning consistency under transformations. In Proc. MICCAI, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.365, + 0.468, + 0.42 + ], + "angle": 0, + "content": "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.423, + 0.468, + 0.491 + ], + "angle": 0, + "content": "[6] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.537 + ], + "angle": 0, + "content": "[7] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proc. CVPR, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.54, + 0.468, + 0.582 + ], + "angle": 0, + "content": "[8] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.584, + 0.468, + 0.626 + ], + "angle": 0, + "content": "[9] Kang Fang and Wu-Jun Li. Dmnet: Difference minimization network for semi-supervised segmentation in medical images. In Proc. MICCAI, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.468, + 0.683 + ], + "angle": 0, + "content": "[10] Shuhao Fu, Yongyi Lu, Yan Wang, Yuyin Zhou, Wei Shen, Elliot Fishman, and Alan Yuille. Domain adaptive relational reasoning for 3d multi-organ segmentation. In Proc. MIC-CAI, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[11] Eli Gibson, Francesco Giganti, Yipeng Hu, Ester Bonmati, Steve Bandula, Kurinchi Gurusamy, Brian Davidson, Stephen P. Pereira, Matthew J. Clarkson, and Dean C. Barratt. Automatic multi-organ segmentation on abdominal ct with dense v-networks. IEEE Trans. Medical Imaging, 37(8):1822-1834, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[12] Yuanfeng Ji, Haotian Bai, Jie Yang, Chongjian Ge, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhang, Wanling Ma, Xi-ang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, T Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Workshop, 2015. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Shuai Lin Li, Chuyu Zhang, and Xuming He. Shape-aware semi-supervised 3d semantic segmentation for medical images. In Proc. MICCAI, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, and Pheng-Ann Heng. Transformation consistent self-ensembling model for semi-supervised medical image segmentation. IEEE Trans. Neural Networks and Learning Systems, 32(2):523-534, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[16] Jinhua Liu, Christian Desrosiers, and Yuanfeng Zhou. Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape awareness and local context constraints. In Proc. MICCAI, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[17] Siqi Liu, Daguang Xu, S Kevin Zhou, Olivier Pauly, Sasa Grbic, Thomas Mertelmeier, Julia Wicklein, Anna Jerebko, Weidong Cai, and Dorin Comaniciu. 3d anisotropic hybrid network: Transferring convolutional features from 2d images to 3d anisotropic volumes. In Proc. MICCAI, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[18] Xiangde Luo, Jieneng Chen, Tao Song, and Guotai Wang. Semi-supervised medical image segmentation through dual-task consistency. In Proc. AAAI, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[19] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, Shucheng Cao, Qi Zhang, Shangqing Liu, Yunpeng Wang, Yuhui Li, Jian He, and Xiaoping Yang. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6695-6714, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[20] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proc. 3DV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[21] Holger R Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim B Turkbey, and Ronald M Summers. Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In Proc. MICCAI, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[22] Holger R. Roth, Hirohisa Oda, Xiangrong Zhou, Natsuki Shimizu, Ying Yang, Yuichiro Hayashi, Masahiro Oda, Michtaka Fujiwara, Kazunari Misawa, and Kensaku Mori. An application of cascaded 3d fully convolutional networks for medical image segmentation. Comput. Medical Imaging Graph., 66:90-99, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[23] Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(3):608-620, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[24] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Proc. NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Hao Tang, Xuming Chen, Yang Liu, Zhipeng Lu, Junhua You, Mingzhou Yang, Shengyu Yao, Guoqi Zhao, Yi Xu, Tingfeng Chen, Yong Liu, and Xiaohui Xie. Clinically applicable deep learning framework for organs at risk delineation in CT images. Nat. Mach. Intell., 1(10):480-491, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Yucheng Tang, Riqiang Gao, Ho Hin Lee, Shizhong Han, Yunqiang Chen, Dashan Gao, Vishwesh Nath, Camilo" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23877" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "Bermudez, Michael R. Savona, Richard G. Abramson, Shunxing Bao, Ilwoo Lyu, Yuankai Huo, and Bennett A. Landman. High-resolution 3d abdominal segmentation with random patch network fusion. Medical Image Anal., 69:101894, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.23 + ], + "angle": 0, + "content": "[27] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proc. CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[28] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proc. NeurIPS, 2017. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[29] Peng Tu, Yawen Huang, Feng Zheng, Zhenyu He, Liujuan Cao, and Ling Shao. Guidedmix-net: Semi-supervised semantic segmentation by using labeled images as reference. In Proc. AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[30] Vikas Verma, Alex Lamb, Juho Kannala, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi-supervised learning. In Proc. IJCAI, 2019. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[31] Kaiping Wang, Bo Zhan, Chen Zu, Xi Wu, Jiliu Zhou, Luping Zhou, and Yan Wang. Semi-supervised medical image segmentation via a tripled-uncertainty guided mean teacher model with contrastive learning. Medical Image Anal., 79:102447, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[32] Qin Wang, Wen Li, and Luc Van Gool. Semi-supervised learning by augmented distribution alignment. In Proc. ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[33] Yan Wang, Yuyin Zhou, Wei Shen, Seyoun Park, Elliot K. Fishman, and Alan L. Yuille. Abdominal multi-organ segmentation with organ-attention networks and statistical fusion. Medical Image Anal., 55:88-102, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[34] Chen Wei, Lingxi Xie, Xutong Ren, Yingda Xia, Chi Su, Jiaying Liu, Qi Tian, and Alan L. Yuille. Iterative reorganization with weak spatial constraints: Solving arbitrary jigsaw puzzles for unsupervised representation learning. In Proc. CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[35] Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, and Jianfei Cai. Exploring smoothness and class-separation for semi-supervised medical image segmentation. In Proc. MIC-CAI, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[36] Yingda Xia, Dong Yang, Zhiding Yu, Fengze Liu, Jinzheng Cai, Lequan Yu, Zhuotun Zhu, Daguang Xu, Alan L. Yuille, and Holger Roth. Uncertainty-aware multi-view co-training for semi-supervised medical image segmentation and domain adaptation. Medical Image Anal., 65:101766, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[37] Lingxi Xie, Qihang Yu, Yuyin Zhou, Yan Wang, Elliot K. Fishman, and Alan L. Yuille. Recurrent saliency transformation network for tiny target segmentation in abdominal ct scans. IEEE Trans. Medical Imaging, 39(2):514-525, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence H. Staib, and James S. Duncan. Simcvd: Simple contrastive voxelwise representation distillation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(9):2228-2237, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[39] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Proc. MICCAI, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[40] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proc. ICCV, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[41] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In Proc. ICLR, 2018. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[42] Xinkai Zhao, Chaowei Fang, De-Jun Fan, Xutao Lin, Feng Gao, and Guanbin Li. Cross-level contrastive learning and consistency constraint for semi-supervised medical image segmentation. In Proc. ISBI, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[43] Yuyin Zhou, Yan Wang, Peng Tang, Song Bai, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Semi-supervised 3d abdominal multi-organ segmentation via deep multi-planar co-training. In Proc. WACV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23878" + } + ] +] \ No newline at end of file diff --git a/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_origin.pdf b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e6ccc6c1a5f558b5b26d0bc455761ead407deeb2 --- /dev/null +++ b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/10c556e7-54b4-4db4-bc9b-113fcac4fcbd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26b6af6efd421e32b8f0991052cca217034f02a432aa68eb285f7862e8e818e5 +size 1485012 diff --git a/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/full.md b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d1176e627cd0cb90cfffa40d35a098f967e319f3 --- /dev/null +++ b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/full.md @@ -0,0 +1,327 @@ +# MagicNet: Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery + +Duowen Chen $^{1}$ Yunhao Bai $^{1}$ Wei Shen $^{2}$ Qingli Li $^{1}$ Lequan Yu $^{3}$ Yan Wang $^{1*}$ + +$^{1}$ Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University + +$^{2}$ MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University $^{3}$ The University of Hong Kong + +duowen_chen@hotmail.com, yhbai@stu.ecnu.edu.cn, wei.shen@sjtu.edu.cn, + +qlli@cs.ecnu.edu.cn, lqyu@hku.hk, ywang@cee.ecnu.edu.cn + +# Abstract + +We propose a novel teacher-student model for semi-supervised multi-organ segmentation. In teacher-student model, data augmentation is usually adopted on unlabeled data to regularize the consistent training between teacher and student. We start from a key perspective that fixed relative locations and variable sizes of different organs can provide distribution information where a multi-organ CT scan is drawn. Thus, we treat the prior anatomy as a strong tool to guide the data augmentation and reduce the mismatch between labeled and unlabeled images for semi-supervised learning. More specifically, we propose a data augmentation strategy based on partition-and-recovery $N^3$ cubes cross- and within- labeled and unlabeled images. Our strategy encourages unlabeled images to learn organ semantics in relative locations from the labeled images (cross-branch) and enhances the learning ability for small organs (within-branch). For within-branch, we further propose to refine the quality of pseudo labels by blending the learned representations from small cubes to incorporate local attributes. Our method is termed as MagicNet, since it treats the CT volume as a magic-cube and $N^3$ -cube partition-and-recovery process matches with the rule of playing a magic-cube. Extensive experiments on two public CT multi-organ datasets demonstrate the effectiveness of MagicNet, and noticeably outperforms state-of-the-art semi-supervised medical image segmentation approaches, with $+7\%$ DSC improvement on MACT dataset with $10\%$ labeled images. Code is available at https://github.com/DeepMed-Lab-ECNU/MagicNet. + +# 1. Introduction + +Abdominal multi-organ segmentation in CT images is an essential task in many clinical applications such as + +![](images/e34ebe83cb9744572b429febb7284b1898d6a99a013698954cd35abbd640be3d.jpg) +Figure 1. Two data augmentation strategies in MagicNet. Left: Although labeled and unlabeled images are not aligned, the latter can be regarded as a shifted version of the former. Co-shift of cubes transfers organ semantics in relative locations from the labeled data to unlabeled data. Right: Segmenting small organs from original images are difficult due to the cluttered background. Small cubes mitigate the impact from background and focus more on local attributes. + +computer-aided intervention [25, 33]. But, training an accurate multi-organ segmentation model usually requires a large amount of labeled data, whose acquisition process is time-consuming and expensive. Semi-supervised learning (SSL) has shown great potential to handle the scarcity of data annotations, which attempts to transfer mass prior knowledge learned from the labeled to unlabeled images. SSL attracts more and more attention in the field of medical image analysis in recent years. + +Popular SSL medical image segmentation methods mainly focus on segmenting a single target or targets in a local region, such as segmenting pancreas or left atrium [4, 9, 14, 15, 18, 23, 31, 35, 38, 39]. Multi-organ segmentation is more challenging than single organ segmentation, due to the complex anatomical structures of the organs, e.g., the fixed relative locations (duodenum is always located at the head of the pancreas), the appearances of different organs, and the large variations of the size. Transfer + +ring current SSL medical segmentation methods to multi-organ segmentation encounters severe problems. Multiple organs introduce much more variance compared with a single organ. Although labeled and unlabeled images are always drawn from the same distribution, due to the limited number of labeled images, it's hard to estimate the precise distribution from them [32]. Thus, the estimated distribution between labeled and unlabeled images always suffers from mismatch problems, and is even largely increased by multiple organs. Aforementioned SSL medical segmentation methods lack the ability to handle such a large distribution gap, which requires sophisticated anatomical structure modeling. A few semi-supervised multi-organ segmentation methods have been proposed, DMPCT [43] designs a co-training strategy to mine consensus information from multiple views of a CT scan. UMCT [36] further proposes an uncertainty estimation of each view to improve the quality of the pseudo-label. Though these methods take the advantages of multi-view properties in a CT scan, they inevitably ignore the internal anatomical structures of multiple organs, resulting in suboptimal results. + +Teacher-student model is a widely adopted framework for semi-supervised medical image segmentation [28]. Student network takes labeled images and unlabeled strongly augmented images as input, which attempts to minimize the distribution mismatch between labeled and unlabeled images from the model level. That is, data augmentation is adopted on unlabeled data, whose role is to regularize the consistent training between teacher and student. As mentioned, semi-supervised multi-organ segmentation suffers from large distribution alignment mismatch between labeled and unlabeled images. Reducing the mismatch mainly from the model level is insufficient to solve the problem. Thanks to the prior anatomical knowledge from CT scans, which provides the distribution information where a multi-organ CT scan is drawn, it is possible to largely alleviate the mismatch problem from the data level. + +To this end, we propose a novel teacher-student model, called MagicNet, matching with the rule of playing a magic-cube. More specifically, we propose a partition-and-recovery $N^3$ cubes learning paradigm: (1) We partition each CT scan, termed as magic-cube, into $N^3$ small cubes. (2) Two data augmentation strategies are then designed, as shown in Fig. 1. I.e., First, to encourage unlabeled data to learn organ semantics in relative locations from the labeled data, small cubes are mixed across labeled and unlabeled images while keeping their relative locations. Second, to enhance the learning ability for small organs, small cubes are shuffled and fed into the student network. (3) We recover the magic-cube to form the original 3D geometry to map with the ground-truth or the supervisory signal from teacher. Furthermore, the quality of pseudo labels predicted by teacher network is refined by blending with the + +learned representation of the small cubes. The cube-wise pseudo-label blending strategy incorporates local attributes e.g., texture, luster and boundary smoothness which mitigates the inferior performance of small organs. + +The main contributions can be summarized as follows: + +- We propose a data augmentation strategy based on partition-and-recovery $N^3$ cubes cross- and within-labeled and unlabeled images which encourages unlabeled images to learn organ semantics in relative locations from the labeled images and enhances the learning ability for small organs. +- We propose to correct the original pseudo-label by cube-wise pseudo-label blending via incorporating crucial local attributes for identifying targets especially small organs. +- We verify the effectiveness of our method on BTCV [13] and MACT [11] datasets. The segmentation performance of our method exceeds all state-of-the-arts by a large margin, with $7.28\%$ (10% labeled) and $6.94\%$ (30% labeled) improvement on two datasets respectively (with V-Net as the backbone) in DSC. + +# 2. Related Work + +# 2.1. Semi-supervised Medical Image Segmentation + +Semi-supervised medical image segmentation methods can be roughly grouped into three categories. (1) Contrastive learning based methods [35, 38], which learn representations that maximize the similarity among positive pairs and minimize the similarity among negative pairs. (2) Consistency regularization based methods [4, 9, 14, 18, 31, 36, 39], which attend to various levels of information for a single target via multi/dual-task learning or transformation consistent learning. (3) Self-ensembling/self-training based methods [1, 15, 23, 39, 43], which generate pseudolabels for unlabeled images and propose several strategies to ensure the quality of pseudo-labels. But, most of these methods are mainly focusing on segmenting one target or targets in a local region or ROI, such as pancreas or left atrium [18,23,38,39], which encounters performance degradation when transferring to multi-organ segmentation, due to the lack of anatomical structure modeling ability. + +# 2.2. Semi-supervised Multi-organ Segmentation + +Due to the large variations of appearance and size of different organs [22, 26, 33, 37], multi-organ segmentation has been a popular yet challenging task. Only a few SSL methods specially for multi-organ segmentation have been proposed. DMPCT [43] adopted a 2D-based co-training framework to aggregate multi-planar features on a private dataset. UMCT [36] further enforces multi-view consistency on unlabeled data. These methods are pioneer works + +![](images/9be41dbb465cde9e3166328c83d6be28b0e99d6747036e8759c1ad2ebecf40f6.jpg) +Figure 2. The architecture of MagicNet. Two essential components are proposed: (1) magic-cube partition and recovery cross- and within-image (see Sec. 3.1 and Fig. 3); (2) cube-wise pseudo-label blending, incorporating local attributes (see Sec. 3.2 and Fig. 4). + +to handle semi-supervised multi-organ segmentation problem from the co-training perspective, which take advantages of the multi-view property of a CT volume, but the properties of multiple organs have not been well-explored. + +# 2.3. Interpolation-based Semi-supervised Learning + +Interpolation-based regularization [30, 40, 41] is quite successful in semi-supervised semantic segmentation. Methods such as Mixup [41] and CutMix [40] are good at synthesizing new training samples and are widely used in semi-supervised learning. FixMatch [24] design a weak-strong pair consistency to simplify semi-supervised learning. MixMatch [3] and ReMixMatch [2] generated pseudo labels for unlabeled data and incorporate gradually unlabeled data with reliable pseudo-label into the labeled set. ICT [30] proposed a interpolated-based consistency training method based on Mixup. GuidedMix-Net [29] utilized Mixup [41] in semi-supervised semantic segmentation. Thus, we also compare our approach with some popular interpolated-based methods in Experiments. + +# 3. Method + +We define the 3D volume of a CT scan as $\mathbf{X} \in \mathbb{R}^{W \times H \times L}$ . The goal is to find the semantic label of each voxel $k \in \mathbf{X}$ , which composes into a predicted label map $\hat{\mathbf{Y}} \in \{0,1,\dots,C\}^{W \times H \times L}$ , where $C = 0$ indicates background class, and $C \neq 0$ represents the organ class. The training set $\mathcal{D}$ consists of two subsets: $\mathcal{D} = \mathcal{D}^l \cup \mathcal{D}^u$ , where $\mathcal{D}^l = \{(\mathbf{X}_i^l, \mathbf{Y}_i^l)\}_{i=1}^N$ and $\mathcal{D}^u = \{\mathbf{X}_i^u\}_{i=N+1}^{M+N}$ . The training images $\mathbf{X}^l$ in $\mathcal{D}^l$ are associated with per-voxel annotations $\mathbf{Y}^l$ while those in $\mathcal{D}^u$ are not. In the rest of this paper, we denote the original and mixed CT scans as magic-cubes, and denote the partitioned small cubes as cubes for simplicity. + +The overall framework of the proposed MagicNet is shown in Fig. 2, consisting of a student network and a teacher network. The teacher network is updated via weighted combination of the student network parameters with exponential moving average (EMA) [28]. Designed in this architecture, our whole framework includes two branches: Cross-image branch and within-image branch. The details will be illustrated in the following sections. + +# 3.1. Magic-cube Partition and Recovery + +Assume that a mini-batch $\mathcal{B}$ contains $n$ images $\mathbf{X}^{\mathcal{B}}\in$ $\mathbb{R}^{n\times W\times H\times L}$ . For notational simplicity, let $n = 2$ , and $\mathcal{B}$ contains one labeled image $\mathbf{X}^l$ and one unlabeled image $\mathbf{X}^u$ , which are randomly sampled from $\mathcal{D}^l$ and $\mathcal{D}^u$ , respectively. We partition $\mathbf{X}^l$ and $\mathbf{X}^u$ into $N^3$ magic-cubes $\{\mathbf{x}_j^u\}_{j = 1}^{N^3}$ and $\{\mathbf{x}_j^l\}_{j = 1}^{N^3}$ , where $\mathbf{x}_j^l,\mathbf{x}_j^u\in \mathbb{R}^{W / N\times H / N\times L / N}$ , $j$ represents the relative location of the cube inside this magic-cube. This partition operation is called as $\mathcal{A}_{part}(\cdot)$ . + +Cross-image Partition and Recovery To encourage labeled and unlabeled images to learn comprehensive common semantics from each other, we mix these cubes across all the labeled and unlabeled images in a mini-batch. As shown in Fig. 3, these cubes $\mathbf{x}_j^l$ and $\mathbf{x}_j^u$ are mixed into two shuffled magic-cubes while keeping their original positions, which produce two interpolated images with mixed magic-cubes $\mathbf{X}_0^{mix},\mathbf{X}_1^{mix}\in \mathbb{R}^{W\times H\times L}$ . The magic-cube mixing cross-image operations are termed as $\mathcal{A}_{\text{mix}}^{\text{cross}}(\cdot)$ . The mixed images are then fed into the student network $\mathcal{F}(\cdot;\Theta^s)$ , followed by a softmax layer $\sigma(\cdot)$ , obtaining the prediction maps $\mathbf{P}_0^{mix},\mathbf{P}_1^{mix}\in \mathbb{R}^{C\times W\times H\times L}$ , where $C$ denotes the number of classes. Next, we recover $\mathbf{P}_{\text{cross}}^l,\mathbf{P}_{\text{cross}}^u\in \mathbb{R}^{C\times W\times H\times L}$ from $\mathbf{P}_0^{mix}$ and $\mathbf{P}_1^{mix}$ via recovering the magic-cubes back to the original positions in their original + +![](images/f7942144e283283c80745454d88a9f2166721988463af8b9454a6f03ab97dfd3.jpg) +Figure 3. Magic-cube partition and recovery. Blue and green volumes represent labeled and unlabeled image, respectively. The number on cubes represents its relative-location in the original image. + +images, and this operation is denoted as $\mathcal{A}_{rec}^{cross}(\cdot)$ . We can simply denote the whole process as: + +$$ +\mathbf {P} _ {\text {c r o s s}} ^ {l}, \mathbf {P} _ {\text {c r o s s}} ^ {u} = \mathcal {A} _ {\text {r e c}} ^ {\text {c r o s s}} \left(\sigma \left(\mathcal {F} \left(\mathcal {A} _ {\text {m i x}} ^ {\text {c r o s s}} \left(\mathcal {A} _ {\text {p a r t}} (\mathbf {X} ^ {\mathcal {B}})\right); \boldsymbol {\Theta} ^ {s}\right)\right)\right). \tag {1} +$$ + +Here, $\mathbf{P}_{cross}^{l},\mathbf{P}_{cross}^{u}\in \mathbb{R}^{(C + 1)\times W\times H\times L}$ . The loss functions for labeled and unlabeled images are defined as + +$$ +\mathcal {L} _ {\text {c r o s s}} ^ {l} (\mathcal {B}; \boldsymbol {\Theta} ^ {s}) = \ell_ {\text {d i c e}} \left(\mathbf {P} _ {\text {c r o s s}} ^ {l}, \mathbf {Y} ^ {l}\right), \tag {2} +$$ + +$$ +\mathcal {L} _ {c r o s s - i n} ^ {u} (\mathcal {B}; \boldsymbol {\Theta} ^ {s}) = \ell_ {d i c e} \left(\mathbf {P} _ {c r o s s} ^ {u}, \widehat {\mathbf {Y}} ^ {u}\right). \tag {3} +$$ + +where $\ell_{\text{dice}}$ denotes multi-class Dice loss. $\mathbf{Y}^l \in \{0,1,\dots,C\}^{W \times H \times L}$ denotes per-voxel manual annotations for $\mathbf{X}^l$ . $\mathcal{L}_{\text{cross-in}}^u$ indicates the refined pseudo-labels for $\mathbf{X}^u$ by blending cube-wise representations and more details can be seen in Sec. 3.2. + +Within-image Partition and Recovery Besides magic-cubes partition and recovery across images, we also design a within-image partition and recovery branch for single images, which can better consider local features and learn local attributes for identifying targets, especially targets of small sizes. For $\mathbf{X}^l$ , the partitioned $j$ -th cube $\mathbf{x}_j^l$ is fed into $\mathcal{F}(\cdot; \Theta^s)$ and the softmax layer, which can acquire a cubewise prediction map $\mathbf{p}_j^l \in \mathbb{R}^{C \times W / N \times H / N \times L / N}$ . Finally, we recover the magic-cube via mixing $N^3$ cube-wise prediction maps back to their original positions as illustrated in Fig. 3, and the recovered probability map is denoted as $\mathbf{P}_{in}^l \in \mathbb{R}^{C \times W \times H \times L}$ . Similarly, for $\mathbf{X}^u$ , the probability map is denoted as $\mathbf{P}_{in}^u \in \mathbb{R}^{C \times W \times H \times L}$ . This operation is denoted as $\mathcal{A}_{rec}^{in}(\cdot)$ . The whole process for obtaining $\mathbf{P}_{in}^l$ and $\mathbf{P}_{in}^u$ are: + +$$ +\mathbf {P} _ {i n} ^ {l}, \mathbf {P} _ {i n} ^ {u} = \mathcal {A} _ {r e c} ^ {i n} (\sigma (\mathcal {F} (\mathcal {A} _ {p a r t} (\mathbf {X} ^ {\mathcal {B}}); \boldsymbol {\Theta} ^ {s}))). \qquad (4) +$$ + +Since $\mathbf{P}_{in}^{l}$ and $\mathbf{P}_{in}^{u}$ are recovered via predictions of cubes, $\mathbf{P}_{in}^{l}$ and $\mathbf{P}_{in}^{u}$ are denoted as cube-wise representations in the rest of the paper. The loss function for the labeled image is: + +$$ +\mathcal {L} _ {i n} ^ {l} (\mathcal {B}; \boldsymbol {\Theta} ^ {s}) = \ell_ {d i c e} \left(\mathbf {P} _ {i n} ^ {l}, \mathbf {Y} ^ {l}\right). \tag {5} +$$ + +For the unlabeled image, we propose to leverage the cubewise representations by refining the pseudo-label instead of + +directly computing the loss function as in Eq. 5, which is illustrated in Sec. 3.2. We conduct ablation studies on various design choices of utilizing the cube-wise representations in the experiment part. + +# 3.2. Cube-wise Pseudo-label Blending + +As mentioned in the previous section, medical imaging experts reveal that local attributes e.g., texture, luster and boundary smoothness are crucial elements for identifying targets such as tiny organs in medical images [42]. Inspired by this, we propose a within-image partition and recovery module to learn cube-wise local representations. Since the teacher network takes the original volume as input, and pays more attention to learn large organs in the head class, voxels that actually belong to the tail class can be incorrectly predicted as head class voxels, due to the lack of local attributes learning. To effectively increase the chance of the voxel predicted to the tail class, we design a cube-wise pseudolabel blending module, which blends the original pseudolabel with cube-wise features. + +In detail, the image-level prediction map of $\mathbf{X}^u$ acquired from the teacher network is defined as $\mathbf{P}_T^u = \mathcal{F}(\mathbf{X}^u;\boldsymbol{\Theta}^t)$ , where $\boldsymbol{\Theta}^t$ is the parameters of teacher network. The reconstructed cube-level representations of $\mathbf{X}^u$ acquired from the student network is $\mathbf{P}_{in}^u$ . As mentioned above, cubewise features pay more attention to learn local attributes which is important for tiny organs. Therefore, we propose a distribution-aware blending strategy, which is shown in Fig. 4(a), and formulated as: + +$$ +\mathbf {P} _ {\text {b l e n d}} ^ {u} = \left(\mathbf {1} - \mathcal {R} (\boldsymbol {\Omega})\right) \odot \mathbf {P} _ {T} ^ {u} + \mathcal {R} (\boldsymbol {\Omega}) \odot \mathbf {P} _ {\text {i n}} ^ {u}, \tag {6} +$$ + +where $\odot$ indicates the element-wise multiplication, $\mathbf{1} \in \{1\}^{(C + 1) \times W \times H \times L}$ , and $\Omega \in \mathbb{R}^{W \times H \times L}$ is a distribution-aware weight map. + +To obtain the weight map, we firstly learn the class-wise distribution $\mathbf{v}$ during training. Suppose $\mathbf{v} \in \mathbb{R}^C$ is a vector, $\mathbf{v} = \{\mathbf{v}_0, \dots, \mathbf{v}_{C-1}\}$ , whose element $\mathbf{v}_c$ indicates the number of voxels belonging to $(c+1)$ th organ, which is accumulated by counting the voxels w.r.t. pseudo-labels over a few previ + +![](images/1ebf55d0cc818c18cf975aa147a48104522d5298fdf9f7afc91d5895d944c469.jpg) +(a) + +![](images/d98634080a5de7bd905effcc2df63480beccf65faf1c8e08c232fc113042621a.jpg) +(b) + +![](images/7eeee30938a71597b0e8cc67df5a84f4b2c9c95aa82a94cd1a07d6d97b34a0fb.jpg) +Figure 4. The cube-wise pseudo-label blending module. (a) Pipeline. (b) Illustration on $\mathcal{R}(\Omega_{\mathbf{m}})$ . If the pseudo-label $\tilde{\mathbf{Y}}_{\mathbf{m}}^{u}$ on $\mathbf{m}$ is incorrectly assigned as a big organ, a big $\Omega_{\mathbf{m}}$ is assigned to ensure possible correction from cube-wise representations. + +ous iterations. Let $\widetilde{\mathbf{Y}}_{\mathbf{m}}^{u}$ denote the pseudo-label for the voxel on location m. The value $\Omega_{\mathbf{m}}\in \mathbb{R}^{1}$ at each spatial location m on the weight map is derived as: + +$$ +\boldsymbol {\Omega} _ {\mathbf {m}} = \sum_ {c = 0} ^ {C - 1} \frac {\mathbf {I} \left(\widetilde {\mathbf {Y}} _ {\mathbf {m}} ^ {u} = c\right) \mathbf {v} _ {c}}{\operatorname* {m a x} \mathbf {v}} \tag {7} +$$ + +where $\mathbf{I}(\cdot)$ is an indicator function. Note that to match with the data dimension in Eq. 6, we replicate $\Omega C + 1$ times to acquire $\mathcal{R}(\Omega) \in \mathbb{R}^{(C + 1) \times W \times H \times L}$ . + +Overall, since the teacher network takes the original volume as input, and pays more attention to learn large organs in the head class, the teacher pseudo-label may be biased. To unbiased the pseudo-label, on the one hand, we keep the pseudo-label of small organs. On the other hand, we remedy the possible incorrect pseudo-labels of big organs via changing them to small-organ classes via Eq. 6 if $\Omega_{\mathbf{m}}$ is large. This can effectively increase the chance of the voxels predicted to small-organs. + +The final refined pseudo-label $\widehat{\mathbf{Y}}^u$ is obtained via $\widehat{\mathbf{Y}}_{\mathbf{m}}^{u} = \arg \max_{c\in \mathcal{C}}\mathbf{P}_{blend,c,\mathbf{m}}^{u}$ , where $\mathcal{C}$ indicates the finite set of class labels, and $\mathbf{m}$ means the voxel location. + +# 3.3. Magic-cube Location Reasoning + +To fully leverage the prior anatomical knowledge of multi-organs, we further propose a magic-cube location reasoning method, which is in line with within-image partition. For one image $\mathbf{X}^l$ , the partitioned $j$ -th cube's $\mathbf{x}_j^l$ is fed into the encoder of student network $\mathcal{F}_{enc}(\cdot;\Theta_{enc}^s)$ to generate $\mathbf{f}_j^l \in \mathbb{R}^{D \times W / N \times H / N \times L / N}$ . Then, the feature $\mathbf{f}_j^l$ is flattened into a one-dimensional vector, passing through a classification head $\mathcal{F}_{cls}(\cdot;\Theta_{cls})$ to project the flattened feature vector into the same size as the number of cubes in one image, i.e., $N^3$ . The classification head is composed of two fully connected layers [10, 34]. Finally, we concatenate the output $\hat{\mathbf{y}}_j^l \in \mathbb{R}^{N^3}$ from the classification head for all $N^3$ cubes in one image, which generates a vector in size $(N^3)^2$ . This vector is further reshaped into a matrix of size $N^3 \times N^3$ , whose rows indicate the probabilities of corresponding cubes belonging to the $N^3$ locations. The same + +applies to $\mathbf{X}^u$ . For both $\mathbf{X}^l$ and $\mathbf{X}^u$ in $\mathcal{B}$ , the cross-entropy loss is adopted to learn relative locations of cubes: + +$$ +\mathcal {L} _ {c l s} \left(\mathcal {B}; \Theta_ {e n c} ^ {s}, \Theta_ {c l s}\right) = \frac {1}{| \mathcal {B} |} \sum_ {\mathbf {X} \in \mathcal {B}} \ell_ {c e} (\sigma \left(\mathcal {F} _ {c l s} \right. \tag {8} +$$ + +$$ +\left(\mathcal {F} _ {e n c} ^ {s} \left(\mathcal {A} _ {p a r t} (\mathbf {X}); \boldsymbol {\Theta} _ {e n c} ^ {s}\right); \boldsymbol {\Theta} _ {c l s}\right)), \mathbb {y}) +$$ + +where $\ell_{ce}$ denotes the cross-entropy loss, y represents cube's relative locations in $\mathbf{X}$ . + +# 3.4. Loss Function + +Overall, we develop a total loss function based on a minibatch for our MagicNet as follows: + +$$ +\begin{array}{l} \mathcal {L} (\mathcal {B}; \Theta_ {\text {e n c}} ^ {s}, \Theta_ {\text {s e g}} ^ {s}, \Theta_ {\text {c l s}}) = \mathcal {L} _ {\sup } (\mathcal {B}; \Theta_ {\text {e n c}} ^ {s}, \Theta^ {s}, \Theta_ {\text {c l s}}) \tag {9} \\ + \mathcal {L} _ {u n s u p} (\mathcal {B}; \Theta_ {e n c} ^ {s}, \Theta^ {s}, \Theta_ {c l s}), \\ \end{array} +$$ + +where $\mathcal{L}_{sup}(\cdot)$ and $\mathcal{L}_{unsup}(\cdot)$ denote the loss for labeled and unlabeled images in $\mathcal{B}$ , respectively: + +$$ +\begin{array}{l} \mathcal {L} _ {\sup } (\mathcal {B}; \Theta_ {e n c} ^ {s}, \Theta^ {s}, \Theta_ {c l s}) = \mathcal {L} _ {c r o s s} ^ {l} (\mathcal {B}; \Theta^ {s}) \tag {10} \\ + \mathcal {L} _ {i n} ^ {l} (\mathcal {B}; \Theta^ {s}) + \beta \mathcal {L} _ {c l s} ^ {l} (\mathcal {B}; \Theta_ {e n c} ^ {s}, \Theta_ {c l s}), \\ \end{array} +$$ + +where $\beta$ is the weight factor to balance the location reasoning term with others, and + +$$ +\begin{array}{l} \mathcal {L} _ {u n s u p} (\mathcal {B}; \Theta_ {e n c} ^ {s}, \Theta^ {s}, \Theta_ {c l s}) = \alpha \mathcal {L} _ {c r o s s - i n} ^ {u} (\mathcal {B}; \Theta^ {s}) \tag {11} \\ + \beta \mathcal {L} _ {c l s} ^ {u} (\mathcal {B}; \Theta_ {e n c} ^ {s}, \Theta_ {c l s}), \\ \end{array} +$$ + +where $\alpha$ is another weight factor to balance the two terms. For $\mathcal{L}_{cls}^{l}$ and $\mathcal{L}_{cls}^{u}$ in Eq. 10 and Eq. 11, the following equation holds: + +$$ +\mathcal {L} _ {c l s} = \mathcal {L} _ {c l s} ^ {l} + \mathcal {L} _ {c l s} ^ {u}. \tag {12} +$$ + +# 3.5. Testing Phase + +Given a testing image $\mathbf{X}_{test}$ , the probability map is obtained by: $\mathbf{P}_{test} = \sigma (\mathcal{F}(\mathbf{X}_{test};\Theta^{s*}))$ , where $\Theta^{s*}$ denotes the well-trained student network parameters. The final label map can be determined by taking the argmax in the first dimension of $\mathbf{P}_{test}$ . + +# 4. Experiments + +# 4.1. Datasets and Pre-processing + +BTCV Multi-organ Segmentation Dataset BTCV multi-organ segmentation dataset is from MICCAI Multi-Atlas Labeling Beyond Cranial Vault-Workshop Challenge [13] which contains 30 subjects with 3779 axial abdominal CT slices with 13 organs annotation. In pre-processing, we follow [27] to re-sample all the CT scans to the voxel spacing $[1.5 \times 1.5 \times 2.0]mm^3$ and normalize them to have zero mean and unit variance. Strictly following [5, 6, 10], 18 cases are divided for training and the other 12 cases are for testing. + +
Labeled%MethodsSplR.kidL.kidGallEsoLivStoAorIVCVeinsPanRGLGAvg. DSCAvg. NSD
30%V-Net [20]86.3775.3871.2041.8148.4692.7744.7888.2072.6049.2722.3037.7921.7157.9056.90
MT [28]79.8977.5678.0838.3158.9992.2648.7388.6179.3652.7328.4254.1621.3061.4260.85
UA-MT [39]87.7573.8772.9742.0850.7993.2242.0788.3071.8055.2736.6143.2237.6761.2060.99
ICT [30]88.7878.0083.9350.0953.3588.3264.1989.0279.6063.5158.9043.2950.7068.5967.77
CPS [7]82.9876.9179.2129.9152.5392.5941.3786.5974.7349.5318.9645.1324.0658.0456.64
SS-Net [35]85.5373.3280.7030.8552.4491.6717.6786.2872.3947.4022.9537.3839.0156.7454.38
SLC-Net [16]90.0484.3885.8453.5555.8792.8657.8089.8380.8753.2238.7848.1037.0366.7868.81
MagicNet (ours)91.4284.6486.1962.8662.4993.8972.8790.7083.5270.0764.9460.8857.4875.5376.31
40%V-Net [20]84.9882.7282.0736.6463.4893.5457.4989.7478.6360.4249.3955.6038.4967.1767.84
MT [28]85.7078.9379.0842.8061.0993.4557.5789.7080.3063.9541.1450.4629.6965.6865.98
UA-MT [39]88.7475.8878.9154.2558.5593.4658.9089.2376.1562.3047.9151.5344.9267.7568.87
ICT [30]90.3184.4186.9649.2265.6594.2965.9590.2381.4469.5666.6157.3556.0173.6974.98
CPS [7]87.5672.9977.5953.3154.0892.4154.5887.7574.3258.6848.0250.3943.8665.8165.34
SS-Net [35]84.7476.3774.1943.4257.0592.9014.3783.1469.7752.4527.0854.2927.6658.2657.75
SLC-Net [16]90.0584.0086.4356.1658.9194.6870.7289.9379.4560.5954.2251.0339.0870.4072.87
MagicNet (ours)91.6185.0288.1358.1666.7294.0774.4690.7784.3171.5668.9063.4860.4776.7478.68
100%V-Net [20]84.0084.8286.3867.4265.0294.8373.7590.2784.1969.8563.5462.6065.0276.2877.45
+ +Table 1. Comparison results (DSCs of each organ, avg. DSC, and avg. NSD) between our method and existing semi-supervised medical image segmentation methods on the BTCV dataset. V-Net means training with only labeled samples on the V-Net backbone. DSC: Dice Similarity Coefficient. NSD: Normalized Surface Dice. The standard deviations are not reported due to space limit. Note: Spl: spleen, R.Kid: right kidney, L.Kid: left kidney, Gall: gallbladder, Eso: esophagus, Liv: liver, Sto: stomach, Aor: aorta, IVC: inferior vena cava, Veins: portal and splenic veins, Pan: pancreas, LG/RG: left/right adrenal glands. + +
Labeled%MethodsSpleenL.kidneyGallbladderEsophagusLiverStomachPancreasDuodenumAvg. DSCAvg. NSD
10%V-Net [20]88.89 (16.31)84.90 (20.11)55.73 (34.27)58.27 (20.05)93.15 (7.11)57.15 (30.62)53.33 (22.44)35.28 (19.83)65.8449.96
MT [28]89.22 (13.56)88.04 (17.61)61.40 (31.91)58.89 (20.62)93.60 (5.04)74.00 (20.37)64.70 (19.31)44.53 (17.61)71.8053.54
UA-MT [39]89.94 (13.36)89.26 (15.67)59.19 (32.35)59.43 (19.45)93.77 (5.05)75.43 (19.61)65.86 (17.48)44.57 (17.40)72.1854.04
ICT [30]90.27 (13.70)89.89 (15.81)61.47 (30.31)58.70 (20.17)93.47 (5.39)72.45 (20.95)65.81 (18.28)45.00 (19.22)72.1354.07
CPS [7]91.73 (12.11)90.59 (14.85)60.54 (32.29)55.82 (21.03)94.19 (4.93)78.80 (15.28)62.93 (20.03)32.58 (18.68)70.9052.80
SS-Net [35]83.52 (19.56)78.28 (25.05)53.08 (33.88)56.24 (20.54)91.84 (6.07)47.65 (29.70)43.23 (24.11)25.70 (18.71)59.9442.68
SLC-Net [16]86.89 (17.54)87.54 (18.00)55.28 (34.75)58.51 (20.54)93.70 (5.33)69.74 (22.89)46.55 (26.21)28.86 (18.87)65.8913.28
MagicNet (ours)92.33 (11.12)91.19 (14.18)66.35 (29.29)70.95 (12.30)94.31 (4.64)81.56 (14.91)76.31 (10.56)61.50 (12.85)79.3162.31
20%V-Net [20]90.53 (15.33)88.28 (19.76)65.16 (30.03)66.10 (16.98)94.49 (4.18)79.56 (16.56)69.04 (17.43)51.98 (18.53)75.6459.66
MT [28]91.76 (12.77)91.40 (14.49)63.83 (31.74)64.15 (18.05)94.44 (4.86)81.19 (15.79)72.24 (15.79)57.01 (16.70)77.0060.32
UA-MT [39]92.39 (11.31)91.44 (14.61)63.63 (31.46)64.33 (17.92)94.50 (4.77)83.21 (14.03)71.93 (15.51)57.21 (17.35)77.3360.64
ICT [30]92.96 (9.60)91.67 (14.52)65.90 (30.49)64.35 (16.79)94.46 (4.66)82.60 (15.46)74.18 (12.99)57.67 (16.16)77.9761.24
CPS [7]92.66 (11.46)91.87 (14.33)64.75 (31.21)57.27 (19.56)94.97 (4.51)85.65 (9.50)74.15 (12.49)55.00 (16.58)77.0460.01
SS-Net [35]91.17 (14.77)87.77 (19.37)63.49 (30.83)65.02 (16.65)93.76 (4.82)73.04 (22.16)70.37 (17.70)52.81 (20.32)74.6857.57
SLC-Net [16]92.60 (11.24)91.38 (14.57)62.46 (31.37)62.84 (19.31)94.54 (4.72)80.30 (14.73)69.79 (16.99)53.48 (18.40)75.9259.93
MagicNet (ours)93.52 (10.22)92.01 (14.30)71.04 (26.92)70.95 (12.58)94.89 (4.42)85.77 (10.27)78.81 (8.07)66.21 (12.33)81.6565.87
100%V-Net [20]94.35 (5.81)92.30 (14.22)73.61 (28.80)73.48 (12.43)95.17 (3.10)89.09 (7.74)80.86 (7.38)70.03 (11.60)83.6168.85
+ +Table 2. Comparison results (DSCs of each organ, avg. DSC, and avg. NSD) between our method and existing semi-supervised medical image segmentation methods on MACT dataset under 4-fold cross-validation. The value in $(\cdot)$ is standard deviation. + +MACT Dataset Multi-organ abdominal CT reference standard segmentations (MACT) is a public multi-organ segmentation dataset, containing 90 CT volumes with 8 organs annotation. It is manually re-annotated by [11] based on 43 cases of NIH pancreas [21] and 47 cases of Synapse [13]. We follow the setting of UMCT [36]. 1) In preprocessing, we set the soft tissue CT window range of $[-125, 275]$ HU and all the CT scans are re-sampled to an isotropic resolution of $1.0\mathrm{mm}^3$ . Then, image intensities are normalized to have zero mean and unit variance. 2) We randomly split the dataset into four folds, and perform 4-fold cross-validation. Labeled training set is then randomly selected, which are applied to all experiments for fair comparison. + +# 4.2. Experimental Setup and Evaluation Metrics + +In this work, we conduct all the experiments based on PyTorch with one NVIDIA 3090 GPU and use V-Net as our backbone [16,35,39]. For magic-cube location reasoning, we add a classification head composed of two fully + +connected layers at the end of our encoder. Firstly, for BTCV dataset, the framework is trained by an SGD optimizer for 70K iterations, with an initial learning rate (lr) 0.01 with a warming-up strategy: $lr = \text{base\_lr} \times (1 - \frac{\text{iterations}}{\text{max\_iterations}})^{0.9}$ . For MACT dataset, the framework is trained by an SGD optimizer for 30K iterations, with an initial learning rate (lr) 0.01 decayed by 0.1 every 12K iterations. Following [16, 18, 35, 36, 39], for both two datasets, batch size is set as 4, including 2 labeled images and 2 unlabeled images, and the hyper-parameter $\alpha$ is set using a time-dependent Gaussian warming-up function and $\beta$ is empirically set as 0.1 [10]. We randomly crop volume of size $96^3$ , and then we partition the volume into $N^3$ small cubes, forming our magic-cubes, and the choice of $N$ is also discussed in Sec. 4.4. In the testing phase, a sliding window strategy is adopted to obtain final results with a stride of $16^3$ . The final evaluation metrics for our method are DSC (\%, Dice-Sørensen Coefficient) and NSD (\%, Normalized Surface Dice) which are commonly used in the cur + +
MethodsCrossInLocBldAvg. DSCspleenl.kidneygallbladderesophagusliverstomachpancreasduodenum
Baseline (MT [28])71.8886.8488.9254.5053.4093.0170.8671.7755.76
Cross✓78.1392.1190.5363.4270.0593.9079.6075.5859.87
Cross + In✓✓78.8392.8290.7564.9170.6893.9080.9276.2560.40
Cross + Loc✓✓78.5692.4991.0765.9569.7294.1280.4075.8258.94
Cross + In + Loc✓✓✓78.8792.5591.1865.7270.7794.4280.0575.8160.43
Cross + In + Loc + Bld✓✓✓✓79.3192.3391.1966.3570.9594.3181.5676.3161.50
+ +Table 3. Ablation study (DSC %) for the effectiveness of each component of MagicNet. Cross: cross-image partition-and-recovery. In: Within-image partition-and-recovery. Bld: cube-wise pseudo-label blending. Loc: magic-cube location reasoning. + +![](images/c19517246faf0d44e4777c913db537fadc2db62d82121fbed446ae4fc0b96413.jpg) +Figure 5. DSC comparison in box plots on MACT dataset with $10\%$ labeled images (Best viewed electronically, zoom in). + +rent multi-organ segmentation challenge [12, 19]. + +# 4.3. Comparison with the State-of-the-art Methods + +We compare MagicNet with six state-of-the-art semi-supervised segmentation methods, including mean-teacher (MT) [28], uncertainty-aware mean-teacher (UA-MT) [39], interpolated consistency training (ICT) [30], cross pseudosupervision (CPS) [7], smoothness and class-separation (SS-Net) [35], shape-awareness and local constraints (SLC-Net) [16]. All experiments of other methods are re-trained by their official code when transferred to multi-organ task. V-Net is adopted as our backbone which is the same as the above methods for fair comparison. + +In Table 1, we summarize the results of BTCV dataset. Visualizations compared with state-of-the-arts are shown in the supplementary material. Noticeable improvements compared with state-of-the-arts can be seen for organs such as Stomach (↑ 8.68), Gallbladder (↑ 9.31), Veins (↑ 6.56), Right adrenal glands (↑ 6.72) and Left adrenal glands (↑ 6.78). It is interesting to observe that some semi-supervised methods even perform worse than the lower bound. This is because these methods are less able to learn common semantics from the labeled images to unlabeled ones, generating less accurate pseudo-labels, which even hamper the usage of unlabeled images (see the supplementary material). + +We then evaluate MagicNet on MACT dataset. In Table 2, we can observe: (1) a more significant performance improvement between MagicNet and other state-of-the-arts with $10\%$ labeled data than $20\%$ , which demonstrates the effectiveness of MagicNet when training with fewer labeled CT images; (2) MagicNet successfully mitigates the inferior performance in tail classes, e.g., DSC improvement with $10\%$ labeled images on Gallbladder ( $\uparrow$ 4.88), Esophagus ( $\uparrow$ 11.52), Pancreas ( $\uparrow$ 10.45) and Duodenum ( $\uparrow$ 16.50). Fig. 5 + +![](images/bb2aca65c72d67cb7f1d7a531cc9ce329df66025d2566aa5d2086de281c5b9d2.jpg) +Figure 6. Scramble position and keep position when mixing small-cubes across images (from Fig. 3). Red arrow indicates the cube location in mixed images. + +shows a comparison in DSC of six methods for different organs. Note that UMCT [36] on MACT dataset achieves $75.33\%$ in DSC. It uses AH-Net [17] as the backbone (3 are used), whose parameter number is much larger. Due to computation resources, we do not directly compare with it. + +# 4.4. Ablation Study + +We conduct ablation studies on MACT dataset [11] with $10\%$ labeled data to show the effectiveness of each module. The effectiveness of each component in MagicNet We conduct ablation studies to show the impact of each component in MagicNet. In Table 3, the first row indicates the mean-teacher baseline model [28], which our method is designed on. Compared to the baseline, our magic-cube partition and recovery can yield good segmentation performance. Cross and In represent our cross- and within-image partition-and-recovery, which increase the performance from $71.88\%$ to $78.13\%$ and $78.87\%$ , respectively. This shows the powerful ability with our specially-designed data augmentation method. Loc represents magic-cube location reasoning for within-image branch. We can see from the table that based on our cross-image branch, adding relative locations of small-cubes can achieve $78.56\%$ in DSC, adding only within-image partition-and-recovery branch achieves $78.83\%$ , and adding both two branches leads to $78.87\%$ . Finally, our proposed cube-wise pseudo label blending (short for Bld in Table 3) provides significant improvement to $79.31\%$ , for which we will make a more complete analysis later in the current section. + +Design choices of partition and recovery Here, we discuss the design for cross-image partition-and-recovery branch: ① Should we maintain or scramble the magic-cube relative locations when manipulating our cross-image branch, as shown in Fig. 6? The comparison results are shown in the last two rows in Table 4. Scramble and Keep represent the + +
①②DSCNSD
scrambleU69.53 ± 11.2750.98 ± 11.00
keepU73.84 ± 9.9655.80 ± 10.81
scrambleLU71.91 ± 11.2054.10 ± 11.28
keepLU78.13 ± 8.1260.79 ± 9.90
+ +Table 4. Ablation of design choices for crossimage partition and recovery (Question $①$ and $②$ , mean $\pm$ std of all cases). Scramble/keep: ignore/keep original positions when mixed. U: only for unlabeled data. LU: for both labeled and unlabeled data. The last row is ours. + +
DSCNSD
teacher sup.78.47 ± 7.8361.52 ± 9.21
mutual sup.70.45 ± 11.4753.71 ± 11.78
blending79.31 ± 7.5562.31 ± 9.08
+ +Table 5. Comparison of different pseudolabel supervision/blending strategies for unlabeled data. Sup.: supervision. Blending: our cube-wise pseudo-label blending. + +
DSCNSD
CutMix [40]73.80 ± 9.9955.86 ± 10.54
CutOut [8]72.47 ± 10.2754.62 ± 10.67
MixUp [41]72.13 ± 11.3254.07 ± 12.01
Ours (2)78.68 ± 7.4560.78 ± 9.12
Ours (3)79.31 ± 7.5562.31 ± 9.08
+ +Table 6. Comparison of different data augmentation methods, where we try different CutMix and CutOut sizes, and choose the best results. For MagicNet, we compare different $N$ values in $(\cdot)$ . + +![](images/6e54eb8b9e5b5fdeb34f6dfe547f9cbd16ef87b5eb64a09905e5362b380150d0.jpg) +Figure 7. Kernel Density Estimation of labeled and unlabeled samples on the first three channels of the feature map of "only unlabeled images" and "both labeled and unlabeled images". Considerable distribution mismatch between labeled and unlabeled images can be observed (top), while our method can well improve the distribution mismatch (bottom). Best viewed zoom in. + +partitioned small-cubes are randomly mixed while ignoring their original locations or kept when mixing them across images. Results show that the relative locations between multiple organs are important for multi-organ segmentation. ② Should our cross-image data augmentation be operated on only unlabeled images (see U in Table 4) or both labeled and unlabeled images (see LU in Table 4)? The latter obtains a much better performance compared to the former. Besides, it largely reduces the distribution mismatch between labeled and unlabeled data, as shown in Fig. 7. + +Cube-wise pseudo-label blending As illustrated in Sec. 3.2, we blend the output of within-image branch and the output of teacher model to obtain the final pseudo-label for complementing local attributes. We compare our blending with two other methods, as shown in Table 5. Three supervision schemes, illustrated in Fig. 8, are compared for unlabeled images based on our framework. Teacher sup. means the outputs of cross- and within-image branch are both supervised by the pseudo-label from the teacher model. Mutual sup. means the outputs of cross- and within-image are mutually supervised. It can be seen that our blending strategy works favorably for unlabeled data. + +Comparison with interpolated-based methods As shown in Table 6, our augmentation method outperforms other + +![](images/374e9b3d55d9c607d4ddf82bb30d4f1db25d33b5b20ad3fdadb659936534845a.jpg) +Figure 8. Different pseudo-label supervision/blending strategies corresponding to Table 5. Blue line indicates supervision. (a) teacher sup. (b) mutual sup. (c) our blending. + +methods such as CutMix, CutOut and MixUp. For fair comparison, we try several CutMix and CutOut sizes, and choose the best results. + +Different number $(N)$ of small-cubes We study the impact of different numbers of small-cubes $N$ , as shown in Table 6. Slightly better performance is achieved when $N = 3$ . When $N = 4$ , the size of the small cube does not match the condition for VNet. Thus, we only compare the results given $N = 2$ and $N = 3$ . + +# 5. Limitations and Conclusion + +We have presented MagicNet for semi-supervised multi-organ segmentation. MagicNet encourages unlabeled data to learn organ semantics in relative locations from the labeled data, and enhances the learning ability for small organs. Two essential components are proposed: (1) partition-and-recovery $N^3$ small cubes cross- and within-labeled and unlabeled images, and (2) the quality of pseudo label refinement by blending the cube-wise representations. Experimental results on two public multi-organ segmentation datasets verify the superiority of our method. + +Limitations: CT scans are all roughly aligned by the nature of the data. Magic-cube may not work well on domains $w / o$ such prior, e.g., microscopy images $w / o$ similar prior. + +# 6. Acknowledgement + +This work was supported by the National Natural Science Foundation of China (Grant No. 62101191), Shanghai Natural Science Foundation (Grant No. 21ZR1420800), and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ2229004). + +# References + +[1] Wenjia Bai, Ozan Oktay, Matthew Sinclair, Hideaki Suzuki, Martin Rajchl, Giacomo Tarroni, Ben Glocker, Andrew P. King, Paul M. Matthews, and Daniel Rueckert. Semi-supervised learning for network-based cardiac MR image segmentation. In Proc. MICCAI, 2017. 2 +[2] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In Proc. ICLR, 2020. 3 +[3] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Proc. NeurIPS, 2019. 3 +[4] Gerda Bortsova, Florian Dubost, Laurens Hogeweg, Ioannis Katramados, and Marleen de Bruijne. Semi-supervised medical image segmentation via learning consistency under transformations. In Proc. MICCAI, 2019. 1, 2 +[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 5 +[6] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021.5 +[7] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proc. CVPR, 2021. 6, 7 +[8] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 8 +[9] Kang Fang and Wu-Jun Li. Dmnet: Difference minimization network for semi-supervised segmentation in medical images. In Proc. MICCAI, 2020. 1, 2 +[10] Shuhao Fu, Yongyi Lu, Yan Wang, Yuyin Zhou, Wei Shen, Elliot Fishman, and Alan Yuille. Domain adaptive relational reasoning for 3d multi-organ segmentation. In Proc. MIC-CAI, 2020. 5, 6 +[11] Eli Gibson, Francesco Giganti, Yipeng Hu, Ester Bonmati, Steve Bandula, Kurinchi Gurusamy, Brian Davidson, Stephen P. Pereira, Matthew J. Clarkson, and Dean C. Barratt. Automatic multi-organ segmentation on abdominal ct with dense v-networks. IEEE Trans. Medical Imaging, 37(8):1822-1834, 2018. 2, 6, 7 +[12] Yuanfeng Ji, Haotian Bai, Jie Yang, Chongjian Ge, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhang, Wanling Ma, Xi-ang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023, 2022. 7 +[13] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, T Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Workshop, 2015. 2, 5, 6 + +[14] Shuai Lin Li, Chuyu Zhang, and Xuming He. Shape-aware semi-supervised 3d semantic segmentation for medical images. In Proc. MICCAI, 2020. 1, 2 +[15] Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, and Pheng-Ann Heng. Transformation consistent self-ensembling model for semi-supervised medical image segmentation. IEEE Trans. Neural Networks and Learning Systems, 32(2):523-534, 2020. 1, 2 +[16] Jinhua Liu, Christian Desrosiers, and Yuanfeng Zhou. Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape awareness and local context constraints. In Proc. MICCAI, 2022. 6, 7 +[17] Siqi Liu, Daguang Xu, S Kevin Zhou, Olivier Pauly, Sasa Grbic, Thomas Mertelmeier, Julia Wicklein, Anna Jerebko, Weidong Cai, and Dorin Comaniciu. 3d anisotropic hybrid network: Transferring convolutional features from 2d images to 3d anisotropic volumes. In Proc. MICCAI, 2018. 7 +[18] Xiangde Luo, Jieneng Chen, Tao Song, and Guotai Wang. Semi-supervised medical image segmentation through dual-task consistency. In Proc. AAAI, 2021. 1, 2, 6 +[19] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, Shucheng Cao, Qi Zhang, Shangqing Liu, Yunpeng Wang, Yuhui Li, Jian He, and Xiaoping Yang. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6695-6714, 2022. 7 +[20] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proc. 3DV, 2016. 6 +[21] Holger R Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim B Turkbey, and Ronald M Summers. Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In Proc. MICCAI, 2015. 6 +[22] Holger R. Roth, Hirohisa Oda, Xiangrong Zhou, Natsuki Shimizu, Ying Yang, Yuichiro Hayashi, Masahiro Oda, Michtaka Fujiwara, Kazunari Misawa, and Kensaku Mori. An application of cascaded 3d fully convolutional networks for medical image segmentation. Comput. Medical Imaging Graph., 66:90-99, 2018. 2 +[23] Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(3):608-620, 2022. 1, 2 +[24] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Proc. NeurIPS, 2020. 3 +[25] Hao Tang, Xuming Chen, Yang Liu, Zhipeng Lu, Junhua You, Mingzhou Yang, Shengyu Yao, Guoqi Zhao, Yi Xu, Tingfeng Chen, Yong Liu, and Xiaohui Xie. Clinically applicable deep learning framework for organs at risk delineation in CT images. Nat. Mach. Intell., 1(10):480-491, 2019. 1 +[26] Yucheng Tang, Riqiang Gao, Ho Hin Lee, Shizhong Han, Yunqiang Chen, Dashan Gao, Vishwesh Nath, Camilo + +Bermudez, Michael R. Savona, Richard G. Abramson, Shunxing Bao, Ilwoo Lyu, Yuankai Huo, and Bennett A. Landman. High-resolution 3d abdominal segmentation with random patch network fusion. Medical Image Anal., 69:101894, 2021. 2 +[27] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proc. CVPR, 2022. 5 +[28] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proc. NeurIPS, 2017. 2, 3, 6, 7 +[29] Peng Tu, Yawen Huang, Feng Zheng, Zhenyu He, Liujuan Cao, and Ling Shao. Guidedmix-net: Semi-supervised semantic segmentation by using labeled images as reference. In Proc. AAAI, 2022. 3 +[30] Vikas Verma, Alex Lamb, Juho Kannala, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi-supervised learning. In Proc. IJCAI, 2019. 3, 6, 7 +[31] Kaiping Wang, Bo Zhan, Chen Zu, Xi Wu, Jiliu Zhou, Luping Zhou, and Yan Wang. Semi-supervised medical image segmentation via a tripled-uncertainty guided mean teacher model with contrastive learning. Medical Image Anal., 79:102447, 2022. 1, 2 +[32] Qin Wang, Wen Li, and Luc Van Gool. Semi-supervised learning by augmented distribution alignment. In Proc. ICCV, 2019. 2 +[33] Yan Wang, Yuyin Zhou, Wei Shen, Seyoun Park, Elliot K. Fishman, and Alan L. Yuille. Abdominal multi-organ segmentation with organ-attention networks and statistical fusion. Medical Image Anal., 55:88-102, 2019. 1, 2 +[34] Chen Wei, Lingxi Xie, Xutong Ren, Yingda Xia, Chi Su, Jiaying Liu, Qi Tian, and Alan L. Yuille. Iterative reorganization with weak spatial constraints: Solving arbitrary jigsaw puzzles for unsupervised representation learning. In Proc. CVPR, 2019. 5 +[35] Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, and Jianfei Cai. Exploring smoothness and class-separation for semi-supervised medical image segmentation. In Proc. MIC-CAI, 2022. 1, 2, 6, 7 +[36] Yingda Xia, Dong Yang, Zhiding Yu, Fengze Liu, Jinzheng Cai, Lequan Yu, Zhuotun Zhu, Daguang Xu, Alan L. Yuille, and Holger Roth. Uncertainty-aware multi-view co-training for semi-supervised medical image segmentation and domain adaptation. Medical Image Anal., 65:101766, 2020. 2, 6, 7 +[37] Lingxi Xie, Qihang Yu, Yuyin Zhou, Yan Wang, Elliot K. Fishman, and Alan L. Yuille. Recurrent saliency transformation network for tiny target segmentation in abdominal ct scans. IEEE Trans. Medical Imaging, 39(2):514-525, 2020. 2 +[38] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence H. Staib, and James S. Duncan. Simcvd: Simple contrastive voxelwise representation distillation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(9):2228-2237, 2022. 1, 2 + +[39] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Proc. MICCAI, 2019. 1, 2, 6, 7 +[40] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proc. ICCV, 2019. 3, 8 +[41] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In Proc. ICLR, 2018. 3, 8 +[42] Xinkai Zhao, Chaowei Fang, De-Jun Fan, Xutao Lin, Feng Gao, and Guanbin Li. Cross-level contrastive learning and consistency constraint for semi-supervised medical image segmentation. In Proc. ISBI, 2022. 4 +[43] Yuyin Zhou, Yan Wang, Peng Tang, Song Bai, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Semi-supervised 3d abdominal multi-organ segmentation via deep multi-planar co-training. In Proc. WACV, 2019. 2 \ No newline at end of file diff --git a/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/images.zip b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f0278e10a430fda7a9ffff2ea5453112566ed299 --- /dev/null +++ b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eb76442eef26fcb44c004dddc46534991b1ced8b5f2b65839258202db45ccc5 +size 707044 diff --git a/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/layout.json b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2da4b1ec3eb4268d89d0d510c63a9741ea97cc34 --- /dev/null +++ b/2023/MagicNet_ Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery/layout.json @@ -0,0 +1,10267 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 519, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 519, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 519, + 140 + ], + "type": "text", + "content": "MagicNet: Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "spans": [ + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "text", + "content": "Duowen Chen" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "text", + "content": " Yunhao Bai" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "text", + "content": " Wei Shen" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "text", + "content": " Qingli Li" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "text", + "content": " Lequan Yu" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "text", + "content": " Yan Wang" + }, + { + "bbox": [ + 85, + 160, + 511, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 177, + 504, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 177, + 504, + 190 + ], + "spans": [ + { + "bbox": [ + 88, + 177, + 504, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 177, + 504, + 190 + ], + "type": "text", + "content": "Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 190, + 533, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 190, + 533, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 533, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 190, + 533, + 204 + ], + "type": "text", + "content": "MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University " + }, + { + "bbox": [ + 61, + 190, + 533, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 190, + 533, + 204 + ], + "type": "text", + "content": "The University of Hong Kong" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 205, + 481, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 205, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 109, + 205, + 481, + 217 + ], + "type": "text", + "content": "duowen_chen@hotmail.com, yhbai@stu.ecnu.edu.cn, wei.shen@sjtu.edu.cn," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 147, + 220, + 446, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 220, + 446, + 232 + ], + "spans": [ + { + "bbox": [ + 147, + 220, + 446, + 232 + ], + "type": "text", + "content": "qlli@cs.ecnu.edu.cn, lqyu@hku.hk, ywang@cee.ecnu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "text", + "content": "We propose a novel teacher-student model for semi-supervised multi-organ segmentation. In teacher-student model, data augmentation is usually adopted on unlabeled data to regularize the consistent training between teacher and student. We start from a key perspective that fixed relative locations and variable sizes of different organs can provide distribution information where a multi-organ CT scan is drawn. Thus, we treat the prior anatomy as a strong tool to guide the data augmentation and reduce the mismatch between labeled and unlabeled images for semi-supervised learning. More specifically, we propose a data augmentation strategy based on partition-and-recovery " + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "text", + "content": " cubes cross- and within- labeled and unlabeled images. Our strategy encourages unlabeled images to learn organ semantics in relative locations from the labeled images (cross-branch) and enhances the learning ability for small organs (within-branch). For within-branch, we further propose to refine the quality of pseudo labels by blending the learned representations from small cubes to incorporate local attributes. Our method is termed as MagicNet, since it treats the CT volume as a magic-cube and " + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "text", + "content": "-cube partition-and-recovery process matches with the rule of playing a magic-cube. Extensive experiments on two public CT multi-organ datasets demonstrate the effectiveness of MagicNet, and noticeably outperforms state-of-the-art semi-supervised medical image segmentation approaches, with " + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "inline_equation", + "content": "+7\\%" + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "text", + "content": " DSC improvement on MACT dataset with " + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "text", + "content": " labeled images. Code is available at https://github.com/DeepMed-Lab-ECNU/MagicNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 652, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 652, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 652, + 128, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "type": "text", + "content": "Abdominal multi-organ segmentation in CT images is an essential task in many clinical applications such as" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 258, + 544, + 384 + ], + "blocks": [ + { + "bbox": [ + 308, + 258, + 544, + 384 + ], + "lines": [ + { + "bbox": [ + 308, + 258, + 544, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 544, + 384 + ], + "type": "image", + "image_path": "e34ebe83cb9744572b429febb7284b1898d6a99a013698954cd35abbd640be3d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 391, + 546, + 478 + ], + "lines": [ + { + "bbox": [ + 305, + 391, + 546, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 546, + 478 + ], + "type": "text", + "content": "Figure 1. Two data augmentation strategies in MagicNet. Left: Although labeled and unlabeled images are not aligned, the latter can be regarded as a shifted version of the former. Co-shift of cubes transfers organ semantics in relative locations from the labeled data to unlabeled data. Right: Segmenting small organs from original images are difficult due to the cluttered background. Small cubes mitigate the impact from background and focus more on local attributes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 494, + 546, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 546, + 602 + ], + "type": "text", + "content": "computer-aided intervention [25, 33]. But, training an accurate multi-organ segmentation model usually requires a large amount of labeled data, whose acquisition process is time-consuming and expensive. Semi-supervised learning (SSL) has shown great potential to handle the scarcity of data annotations, which attempts to transfer mass prior knowledge learned from the labeled to unlabeled images. SSL attracts more and more attention in the field of medical image analysis in recent years." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Popular SSL medical image segmentation methods mainly focus on segmenting a single target or targets in a local region, such as segmenting pancreas or left atrium [4, 9, 14, 15, 18, 23, 31, 35, 38, 39]. Multi-organ segmentation is more challenging than single organ segmentation, due to the complex anatomical structures of the organs, e.g., the fixed relative locations (duodenum is always located at the head of the pancreas), the appearances of different organs, and the large variations of the size. Transfer" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23869" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "content": "ring current SSL medical segmentation methods to multi-organ segmentation encounters severe problems. Multiple organs introduce much more variance compared with a single organ. Although labeled and unlabeled images are always drawn from the same distribution, due to the limited number of labeled images, it's hard to estimate the precise distribution from them [32]. Thus, the estimated distribution between labeled and unlabeled images always suffers from mismatch problems, and is even largely increased by multiple organs. Aforementioned SSL medical segmentation methods lack the ability to handle such a large distribution gap, which requires sophisticated anatomical structure modeling. A few semi-supervised multi-organ segmentation methods have been proposed, DMPCT [43] designs a co-training strategy to mine consensus information from multiple views of a CT scan. UMCT [36] further proposes an uncertainty estimation of each view to improve the quality of the pseudo-label. Though these methods take the advantages of multi-view properties in a CT scan, they inevitably ignore the internal anatomical structures of multiple organs, resulting in suboptimal results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 327, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 289, + 518 + ], + "type": "text", + "content": "Teacher-student model is a widely adopted framework for semi-supervised medical image segmentation [28]. Student network takes labeled images and unlabeled strongly augmented images as input, which attempts to minimize the distribution mismatch between labeled and unlabeled images from the model level. That is, data augmentation is adopted on unlabeled data, whose role is to regularize the consistent training between teacher and student. As mentioned, semi-supervised multi-organ segmentation suffers from large distribution alignment mismatch between labeled and unlabeled images. Reducing the mismatch mainly from the model level is insufficient to solve the problem. Thanks to the prior anatomical knowledge from CT scans, which provides the distribution information where a multi-organ CT scan is drawn, it is possible to largely alleviate the mismatch problem from the data level." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": "To this end, we propose a novel teacher-student model, called MagicNet, matching with the rule of playing a magic-cube. More specifically, we propose a partition-and-recovery " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " cubes learning paradigm: (1) We partition each CT scan, termed as magic-cube, into " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " small cubes. (2) Two data augmentation strategies are then designed, as shown in Fig. 1. I.e., First, to encourage unlabeled data to learn organ semantics in relative locations from the labeled data, small cubes are mixed across labeled and unlabeled images while keeping their relative locations. Second, to enhance the learning ability for small organs, small cubes are shuffled and fed into the student network. (3) We recover the magic-cube to form the original 3D geometry to map with the ground-truth or the supervisory signal from teacher. Furthermore, the quality of pseudo labels predicted by teacher network is refined by blending with the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "learned representation of the small cubes. The cube-wise pseudo-label blending strategy incorporates local attributes e.g., texture, luster and boundary smoothness which mitigates the inferior performance of small organs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 121, + 539, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 121, + 539, + 132 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 539, + 132 + ], + "type": "text", + "content": "The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 137, + 545, + 342 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 137, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 137, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 137, + 545, + 209 + ], + "type": "text", + "content": "- We propose a data augmentation strategy based on partition-and-recovery " + }, + { + "bbox": [ + 317, + 137, + 545, + 209 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 317, + 137, + 545, + 209 + ], + "type": "text", + "content": " cubes cross- and within-labeled and unlabeled images which encourages unlabeled images to learn organ semantics in relative locations from the labeled images and enhances the learning ability for small organs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 216, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 216, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 216, + 545, + 263 + ], + "type": "text", + "content": "- We propose to correct the original pseudo-label by cube-wise pseudo-label blending via incorporating crucial local attributes for identifying targets especially small organs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 270, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 270, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 270, + 545, + 342 + ], + "type": "text", + "content": "- We verify the effectiveness of our method on BTCV [13] and MACT [11] datasets. The segmentation performance of our method exceeds all state-of-the-arts by a large margin, with " + }, + { + "bbox": [ + 317, + 270, + 545, + 342 + ], + "type": "inline_equation", + "content": "7.28\\%" + }, + { + "bbox": [ + 317, + 270, + 545, + 342 + ], + "type": "text", + "content": " (10% labeled) and " + }, + { + "bbox": [ + 317, + 270, + 545, + 342 + ], + "type": "inline_equation", + "content": "6.94\\%" + }, + { + "bbox": [ + 317, + 270, + 545, + 342 + ], + "type": "text", + "content": " (30% labeled) improvement on two datasets respectively (with V-Net as the backbone) in DSC." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 351, + 392, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 392, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 392, + 364 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 371, + 544, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 544, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 544, + 384 + ], + "type": "text", + "content": "2.1. Semi-supervised Medical Image Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": "Semi-supervised medical image segmentation methods can be roughly grouped into three categories. (1) Contrastive learning based methods [35, 38], which learn representations that maximize the similarity among positive pairs and minimize the similarity among negative pairs. (2) Consistency regularization based methods [4, 9, 14, 18, 31, 36, 39], which attend to various levels of information for a single target via multi/dual-task learning or transformation consistent learning. (3) Self-ensembling/self-training based methods [1, 15, 23, 39, 43], which generate pseudolabels for unlabeled images and propose several strategies to ensure the quality of pseudo-labels. But, most of these methods are mainly focusing on segmenting one target or targets in a local region or ROI, such as pancreas or left atrium [18,23,38,39], which encounters performance degradation when transferring to multi-organ segmentation, due to the lack of anatomical structure modeling ability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 599, + 533, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 533, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 533, + 612 + ], + "type": "text", + "content": "2.2. Semi-supervised Multi-organ Segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Due to the large variations of appearance and size of different organs [22, 26, 33, 37], multi-organ segmentation has been a popular yet challenging task. Only a few SSL methods specially for multi-organ segmentation have been proposed. DMPCT [43] adopted a 2D-based co-training framework to aggregate multi-planar features on a private dataset. UMCT [36] further enforces multi-view consistency on unlabeled data. These methods are pioneer works" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23870" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 69, + 481, + 258 + ], + "blocks": [ + { + "bbox": [ + 111, + 69, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 111, + 69, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 111, + 69, + 481, + 258 + ], + "type": "image", + "image_path": "9be41dbb465cde9e3166328c83d6be28b0e99d6747036e8759c1ad2ebecf40f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 547, + 289 + ], + "type": "text", + "content": "Figure 2. The architecture of MagicNet. Two essential components are proposed: (1) magic-cube partition and recovery cross- and within-image (see Sec. 3.1 and Fig. 3); (2) cube-wise pseudo-label blending, incorporating local attributes (see Sec. 3.2 and Fig. 4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 309, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 288, + 357 + ], + "type": "text", + "content": "to handle semi-supervised multi-organ segmentation problem from the co-training perspective, which take advantages of the multi-view property of a CT volume, but the properties of multiple organs have not been well-explored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 364, + 288, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 288, + 376 + ], + "type": "text", + "content": "2.3. Interpolation-based Semi-supervised Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 382, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 550 + ], + "type": "text", + "content": "Interpolation-based regularization [30, 40, 41] is quite successful in semi-supervised semantic segmentation. Methods such as Mixup [41] and CutMix [40] are good at synthesizing new training samples and are widely used in semi-supervised learning. FixMatch [24] design a weak-strong pair consistency to simplify semi-supervised learning. MixMatch [3] and ReMixMatch [2] generated pseudo labels for unlabeled data and incorporate gradually unlabeled data with reliable pseudo-label into the labeled set. ICT [30] proposed a interpolated-based consistency training method based on Mixup. GuidedMix-Net [29] utilized Mixup [41] in semi-supervised semantic segmentation. Thus, we also compare our approach with some popular interpolated-based methods in Experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 559, + 103, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 103, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 103, + 572 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": "We define the 3D volume of a CT scan as " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{W \\times H \\times L}" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": ". The goal is to find the semantic label of each voxel " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "k \\in \\mathbf{X}" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": ", which composes into a predicted label map " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}} \\in \\{0,1,\\dots,C\\}^{W \\times H \\times L}" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "C = 0" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": " indicates background class, and " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "C \\neq 0" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": " represents the organ class. The training set " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": " consists of two subsets: " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\mathcal{D}^l \\cup \\mathcal{D}^u" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l = \\{(\\mathbf{X}_i^l, \\mathbf{Y}_i^l)\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u = \\{\\mathbf{X}_i^u\\}_{i=N+1}^{M+N}" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": ". The training images " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": " are associated with per-voxel annotations " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^l" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": " while those in " + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 46, + 579, + 288, + 715 + ], + "type": "text", + "content": " are not. In the rest of this paper, we denote the original and mixed CT scans as magic-cubes, and denote the partitioned small cubes as cubes for simplicity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 308, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 406 + ], + "type": "text", + "content": "The overall framework of the proposed MagicNet is shown in Fig. 2, consisting of a student network and a teacher network. The teacher network is updated via weighted combination of the student network parameters with exponential moving average (EMA) [28]. Designed in this architecture, our whole framework includes two branches: Cross-image branch and within-image branch. The details will be illustrated in the following sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 415, + 495, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 495, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 495, + 427 + ], + "type": "text", + "content": "3.1. Magic-cube Partition and Recovery" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": "Assume that a mini-batch " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathcal{B}}\\in" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n\\times W\\times H\\times L}" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": ". For notational simplicity, let " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " contains one labeled image " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " and one unlabeled image " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^u" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": ", which are randomly sampled from " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": ", respectively. We partition " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^u" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " magic-cubes " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_j^u\\}_{j = 1}^{N^3}" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_j^l\\}_{j = 1}^{N^3}" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^l,\\mathbf{x}_j^u\\in \\mathbb{R}^{W / N\\times H / N\\times L / N}" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": " represents the relative location of the cube inside this magic-cube. This partition operation is called as " + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{part}(\\cdot)" + }, + { + "bbox": [ + 304, + 433, + 545, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": "Cross-image Partition and Recovery To encourage labeled and unlabeled images to learn comprehensive common semantics from each other, we mix these cubes across all the labeled and unlabeled images in a mini-batch. As shown in Fig. 3, these cubes " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^l" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^u" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": " are mixed into two shuffled magic-cubes while keeping their original positions, which produce two interpolated images with mixed magic-cubes " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0^{mix},\\mathbf{X}_1^{mix}\\in \\mathbb{R}^{W\\times H\\times L}" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": ". The magic-cube mixing cross-image operations are termed as " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\text{mix}}^{\\text{cross}}(\\cdot)" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": ". The mixed images are then fed into the student network " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot;\\Theta^s)" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": ", followed by a softmax layer " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": ", obtaining the prediction maps " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_0^{mix},\\mathbf{P}_1^{mix}\\in \\mathbb{R}^{C\\times W\\times H\\times L}" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": " denotes the number of classes. Next, we recover " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\text{cross}}^l,\\mathbf{P}_{\\text{cross}}^u\\in \\mathbb{R}^{C\\times W\\times H\\times L}" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_0^{mix}" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1^{mix}" + }, + { + "bbox": [ + 304, + 534, + 546, + 714 + ], + "type": "text", + "content": " via recovering the magic-cubes back to the original positions in their original" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23871" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 511, + 190 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 511, + 190 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 511, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 511, + 190 + ], + "type": "image", + "image_path": "f7942144e283283c80745454d88a9f2166721988463af8b9454a6f03ab97dfd3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "type": "text", + "content": "Figure 3. Magic-cube partition and recovery. Blue and green volumes represent labeled and unlabeled image, respectively. The number on cubes represents its relative-location in the original image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "text", + "content": "images, and this operation is denoted as " + }, + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{rec}^{cross}(\\cdot)" + }, + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "text", + "content": ". We can simply denote the whole process as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 265, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 287, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\text {c r o s s}} ^ {l}, \\mathbf {P} _ {\\text {c r o s s}} ^ {u} = \\mathcal {A} _ {\\text {r e c}} ^ {\\text {c r o s s}} \\left(\\sigma \\left(\\mathcal {F} \\left(\\mathcal {A} _ {\\text {m i x}} ^ {\\text {c r o s s}} \\left(\\mathcal {A} _ {\\text {p a r t}} (\\mathbf {X} ^ {\\mathcal {B}})\\right); \\boldsymbol {\\Theta} ^ {s}\\right)\\right)\\right). \\tag {1}", + "image_path": "2653168cd5685e2355ed9d18383b56b881bffddc286ea6a1544f53184919575b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 280, + 287, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 306 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 280, + 287, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{cross}^{l},\\mathbf{P}_{cross}^{u}\\in \\mathbb{R}^{(C + 1)\\times W\\times H\\times L}" + }, + { + "bbox": [ + 46, + 280, + 287, + 306 + ], + "type": "text", + "content": ". The loss functions for labeled and unlabeled images are defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 312, + 287, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 312, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 86, + 312, + 287, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c r o s s}} ^ {l} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {\\text {d i c e}} \\left(\\mathbf {P} _ {\\text {c r o s s}} ^ {l}, \\mathbf {Y} ^ {l}\\right), \\tag {2}", + "image_path": "b5ba2f984257e3d5c7349d5e4cb1e1a919bfbfb1bae13d55af8817f50bda62fe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 329, + 287, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 329, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 86, + 329, + 287, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c r o s s - i n} ^ {u} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {d i c e} \\left(\\mathbf {P} _ {c r o s s} ^ {u}, \\widehat {\\mathbf {Y}} ^ {u}\\right). \\tag {3}", + "image_path": "91d3df9366d2a14d35ae62c5db5d30beefaa871fcae825ce5ebb3b617d350212.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\ell_{\\text{dice}}" + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "text", + "content": " denotes multi-class Dice loss. " + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^l \\in \\{0,1,\\dots,C\\}^{W \\times H \\times L}" + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "text", + "content": " denotes per-voxel manual annotations for " + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{cross-in}}^u" + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "text", + "content": " indicates the refined pseudo-labels for " + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^u" + }, + { + "bbox": [ + 47, + 354, + 287, + 411 + ], + "type": "text", + "content": " by blending cube-wise representations and more details can be seen in Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": "Within-image Partition and Recovery Besides magic-cubes partition and recovery across images, we also design a within-image partition and recovery branch for single images, which can better consider local features and learn local attributes for identifying targets, especially targets of small sizes. For " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": ", the partitioned " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": "-th cube " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^l" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": " is fed into " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot; \\Theta^s)" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": " and the softmax layer, which can acquire a cubewise prediction map " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^l \\in \\mathbb{R}^{C \\times W / N \\times H / N \\times L / N}" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": ". Finally, we recover the magic-cube via mixing " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": " cube-wise prediction maps back to their original positions as illustrated in Fig. 3, and the recovered probability map is denoted as " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^l \\in \\mathbb{R}^{C \\times W \\times H \\times L}" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": ". Similarly, for " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^u" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": ", the probability map is denoted as " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^u \\in \\mathbb{R}^{C \\times W \\times H \\times L}" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": ". This operation is denoted as " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{rec}^{in}(\\cdot)" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": ". The whole process for obtaining " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^l" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^u" + }, + { + "bbox": [ + 46, + 414, + 288, + 594 + ], + "type": "text", + "content": " are:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 81, + 601, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 601, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 81, + 601, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {i n} ^ {l}, \\mathbf {P} _ {i n} ^ {u} = \\mathcal {A} _ {r e c} ^ {i n} (\\sigma (\\mathcal {F} (\\mathcal {A} _ {p a r t} (\\mathbf {X} ^ {\\mathcal {B}}); \\boldsymbol {\\Theta} ^ {s}))). \\qquad (4)", + "image_path": "9a651ea321f8b97be7d61d29c44f68d9ce6b257122b76b5ea904f4f64257907c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^{l}" + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^{u}" + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "text", + "content": " are recovered via predictions of cubes, " + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^{l}" + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^{u}" + }, + { + "bbox": [ + 47, + 624, + 287, + 661 + ], + "type": "text", + "content": " are denoted as cube-wise representations in the rest of the paper. The loss function for the labeled image is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 667, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 667, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 287, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n} ^ {l} (\\mathcal {B}; \\boldsymbol {\\Theta} ^ {s}) = \\ell_ {d i c e} \\left(\\mathbf {P} _ {i n} ^ {l}, \\mathbf {Y} ^ {l}\\right). \\tag {5}", + "image_path": "b3008b11bb0af375795f59f6847e0ede0fbdb4ed2b6546a267dcac36bc19bebf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "For the unlabeled image, we propose to leverage the cubewise representations by refining the pseudo-label instead of" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 236, + 546, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 546, + 285 + ], + "type": "text", + "content": "directly computing the loss function as in Eq. 5, which is illustrated in Sec. 3.2. We conduct ablation studies on various design choices of utilizing the cube-wise representations in the experiment part." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 292, + 485, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 292, + 485, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 485, + 304 + ], + "type": "text", + "content": "3.2. Cube-wise Pseudo-label Blending" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 311, + 546, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 546, + 476 + ], + "type": "text", + "content": "As mentioned in the previous section, medical imaging experts reveal that local attributes e.g., texture, luster and boundary smoothness are crucial elements for identifying targets such as tiny organs in medical images [42]. Inspired by this, we propose a within-image partition and recovery module to learn cube-wise local representations. Since the teacher network takes the original volume as input, and pays more attention to learn large organs in the head class, voxels that actually belong to the tail class can be incorrectly predicted as head class voxels, due to the lack of local attributes learning. To effectively increase the chance of the voxel predicted to the tail class, we design a cube-wise pseudolabel blending module, which blends the original pseudolabel with cube-wise features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "text", + "content": "In detail, the image-level prediction map of " + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^u" + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "text", + "content": " acquired from the teacher network is defined as " + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_T^u = \\mathcal{F}(\\mathbf{X}^u;\\boldsymbol{\\Theta}^t)" + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Theta}^t" + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "text", + "content": " is the parameters of teacher network. The reconstructed cube-level representations of " + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^u" + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "text", + "content": " acquired from the student network is " + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{in}^u" + }, + { + "bbox": [ + 304, + 478, + 545, + 586 + ], + "type": "text", + "content": ". As mentioned above, cubewise features pay more attention to learn local attributes which is important for tiny organs. Therefore, we propose a distribution-aware blending strategy, which is shown in Fig. 4(a), and formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 594, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 594, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 334, + 594, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\text {b l e n d}} ^ {u} = \\left(\\mathbf {1} - \\mathcal {R} (\\boldsymbol {\\Omega})\\right) \\odot \\mathbf {P} _ {T} ^ {u} + \\mathcal {R} (\\boldsymbol {\\Omega}) \\odot \\mathbf {P} _ {\\text {i n}} ^ {u}, \\tag {6}", + "image_path": "d66f6048e053f3972525dfcde155b701e2854120485f9a779d8c199967f949e8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "type": "text", + "content": " indicates the element-wise multiplication, " + }, + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{1} \\in \\{1\\}^{(C + 1) \\times W \\times H \\times L}" + }, + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\Omega \\in \\mathbb{R}^{W \\times H \\times L}" + }, + { + "bbox": [ + 306, + 617, + 545, + 654 + ], + "type": "text", + "content": " is a distribution-aware weight map." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "content": "To obtain the weight map, we firstly learn the class-wise distribution " + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "content": " during training. Suppose " + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "content": " is a vector, " + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\{\\mathbf{v}_0, \\dots, \\mathbf{v}_{C-1}\\}" + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "content": ", whose element " + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_c" + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "content": " indicates the number of voxels belonging to " + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "inline_equation", + "content": "(c+1)" + }, + { + "bbox": [ + 304, + 654, + 546, + 714 + ], + "type": "text", + "content": "th organ, which is accumulated by counting the voxels w.r.t. pseudo-labels over a few previ" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23872" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 135, + 152 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 135, + 152 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 135, + 152 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 135, + 152 + ], + "type": "image", + "image_path": "1ebf55d0cc818c18cf975aa147a48104522d5298fdf9f7afc91d5895d944c469.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 155, + 88, + 163 + ], + "lines": [ + { + "bbox": [ + 78, + 155, + 88, + 163 + ], + "spans": [ + { + "bbox": [ + 78, + 155, + 88, + 163 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 76, + 206, + 152 + ], + "blocks": [ + { + "bbox": [ + 139, + 76, + 206, + 152 + ], + "lines": [ + { + "bbox": [ + 139, + 76, + 206, + 152 + ], + "spans": [ + { + "bbox": [ + 139, + 76, + 206, + 152 + ], + "type": "image", + "image_path": "d98634080a5de7bd905effcc2df63480beccf65faf1c8e08c232fc113042621a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 154, + 217, + 163 + ], + "lines": [ + { + "bbox": [ + 208, + 154, + 217, + 163 + ], + "spans": [ + { + "bbox": [ + 208, + 154, + 217, + 163 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 208, + 76, + 286, + 152 + ], + "blocks": [ + { + "bbox": [ + 208, + 76, + 286, + 152 + ], + "lines": [ + { + "bbox": [ + 208, + 76, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 208, + 76, + 286, + 152 + ], + "type": "image", + "image_path": "7eeee30938a71597b0e8cc67df5a84f4b2c9c95aa82a94cd1a07d6d97b34a0fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": "Figure 4. The cube-wise pseudo-label blending module. (a) Pipeline. (b) Illustration on " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\Omega_{\\mathbf{m}})" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": ". If the pseudo-label " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Y}}_{\\mathbf{m}}^{u}" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": " is incorrectly assigned as a big organ, a big " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\Omega_{\\mathbf{m}}" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": " is assigned to ensure possible correction from cube-wise representations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 223, + 287, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 261 + ], + "type": "text", + "content": "ous iterations. Let " + }, + { + "bbox": [ + 46, + 223, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{Y}}_{\\mathbf{m}}^{u}" + }, + { + "bbox": [ + 46, + 223, + 287, + 261 + ], + "type": "text", + "content": " denote the pseudo-label for the voxel on location m. The value " + }, + { + "bbox": [ + 46, + 223, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\Omega_{\\mathbf{m}}\\in \\mathbb{R}^{1}" + }, + { + "bbox": [ + 46, + 223, + 287, + 261 + ], + "type": "text", + "content": " at each spatial location m on the weight map is derived as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 270, + 287, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 270, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 115, + 270, + 287, + 301 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Omega} _ {\\mathbf {m}} = \\sum_ {c = 0} ^ {C - 1} \\frac {\\mathbf {I} \\left(\\widetilde {\\mathbf {Y}} _ {\\mathbf {m}} ^ {u} = c\\right) \\mathbf {v} _ {c}}{\\operatorname* {m a x} \\mathbf {v}} \\tag {7}", + "image_path": "b368537e0e31ce5432124bd8d07feed3224b2f47de8cf1a0a3693e7cb81bd255.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(\\cdot)" + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "content": " is an indicator function. Note that to match with the data dimension in Eq. 6, we replicate " + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\Omega C + 1" + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "content": " times to acquire " + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\Omega) \\in \\mathbb{R}^{(C + 1) \\times W \\times H \\times L}" + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 348, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 455 + ], + "type": "text", + "content": "Overall, since the teacher network takes the original volume as input, and pays more attention to learn large organs in the head class, the teacher pseudo-label may be biased. To unbiased the pseudo-label, on the one hand, we keep the pseudo-label of small organs. On the other hand, we remedy the possible incorrect pseudo-labels of big organs via changing them to small-organ classes via Eq. 6 if " + }, + { + "bbox": [ + 46, + 348, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\Omega_{\\mathbf{m}}" + }, + { + "bbox": [ + 46, + 348, + 287, + 455 + ], + "type": "text", + "content": " is large. This can effectively increase the chance of the voxels predicted to small-organs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "text", + "content": "The final refined pseudo-label " + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Y}}^u" + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "text", + "content": " is obtained via " + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Y}}_{\\mathbf{m}}^{u} = \\arg \\max_{c\\in \\mathcal{C}}\\mathbf{P}_{blend,c,\\mathbf{m}}^{u}" + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "text", + "content": " indicates the finite set of class labels, and " + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 46, + 455, + 287, + 492 + ], + "type": "text", + "content": " means the voxel location." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 501, + 221, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 221, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 221, + 514 + ], + "type": "text", + "content": "3.3. Magic-cube Location Reasoning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": "To fully leverage the prior anatomical knowledge of multi-organs, we further propose a magic-cube location reasoning method, which is in line with within-image partition. For one image " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": ", the partitioned " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": "-th cube's " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^l" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": " is fed into the encoder of student network " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{enc}(\\cdot;\\Theta_{enc}^s)" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j^l \\in \\mathbb{R}^{D \\times W / N \\times H / N \\times L / N}" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": ". Then, the feature " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j^l" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": " is flattened into a one-dimensional vector, passing through a classification head " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{cls}(\\cdot;\\Theta_{cls})" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": " to project the flattened feature vector into the same size as the number of cubes in one image, i.e., " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": ". The classification head is composed of two fully connected layers [10, 34]. Finally, we concatenate the output " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_j^l \\in \\mathbb{R}^{N^3}" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": " from the classification head for all " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": " cubes in one image, which generates a vector in size " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "(N^3)^2" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": ". This vector is further reshaped into a matrix of size " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "N^3 \\times N^3" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": ", whose rows indicate the probabilities of corresponding cubes belonging to the " + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 46, + 520, + 287, + 713 + ], + "type": "text", + "content": " locations. The same" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "applies to " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^u" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ". For both " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^u" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ", the cross-entropy loss is adopted to learn relative locations of cubes:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 106, + 545, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 106, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 337, + 106, + 545, + 135 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} \\left(\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}\\right) = \\frac {1}{| \\mathcal {B} |} \\sum_ {\\mathbf {X} \\in \\mathcal {B}} \\ell_ {c e} (\\sigma \\left(\\mathcal {F} _ {c l s} \\right. \\tag {8}", + "image_path": "d80e25a6298670542189e5286046d6f00605fd645feeedc9ed89259ddc632237.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 138, + 514, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 138, + 514, + 152 + ], + "spans": [ + { + "bbox": [ + 369, + 138, + 514, + 152 + ], + "type": "interline_equation", + "content": "\\left(\\mathcal {F} _ {e n c} ^ {s} \\left(\\mathcal {A} _ {p a r t} (\\mathbf {X}); \\boldsymbol {\\Theta} _ {e n c} ^ {s}\\right); \\boldsymbol {\\Theta} _ {c l s}\\right)), \\mathbb {y})", + "image_path": "87887a459567d5f07daf5080110f8e89993b3fc766bf19bb12ad8475155efc4a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 160, + 545, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 160, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 305, + 160, + 545, + 183 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 160, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\ell_{ce}" + }, + { + "bbox": [ + 305, + 160, + 545, + 183 + ], + "type": "text", + "content": " denotes the cross-entropy loss, y represents cube's relative locations in " + }, + { + "bbox": [ + 305, + 160, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 160, + 545, + 183 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 191, + 394, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 191, + 394, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 394, + 204 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 210, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 210, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 545, + 234 + ], + "type": "text", + "content": "Overall, we develop a total loss function based on a minibatch for our MagicNet as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 243, + 545, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 545, + 273 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\mathcal {B}; \\Theta_ {\\text {e n c}} ^ {s}, \\Theta_ {\\text {s e g}} ^ {s}, \\Theta_ {\\text {c l s}}) = \\mathcal {L} _ {\\sup } (\\mathcal {B}; \\Theta_ {\\text {e n c}} ^ {s}, \\Theta^ {s}, \\Theta_ {\\text {c l s}}) \\tag {9} \\\\ + \\mathcal {L} _ {u n s u p} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}", + "image_path": "9f3aee0a66fbde1099d0b9917b4ffe1e28219397cfe5d4d2e117b590f62c6582.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sup}(\\cdot)" + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{unsup}(\\cdot)" + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "content": " denote the loss for labeled and unlabeled images in " + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "content": ", respectively:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 329, + 314, + 545, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 314, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 329, + 314, + 545, + 345 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\sup } (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}) = \\mathcal {L} _ {c r o s s} ^ {l} (\\mathcal {B}; \\Theta^ {s}) \\tag {10} \\\\ + \\mathcal {L} _ {i n} ^ {l} (\\mathcal {B}; \\Theta^ {s}) + \\beta \\mathcal {L} _ {c l s} ^ {l} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}", + "image_path": "07d8dca66304d7deff37fcbd388e48fb9d8ff2621c80008c24437eed114011f0.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 354, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 377 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 354, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 354, + 545, + 377 + ], + "type": "text", + "content": " is the weight factor to balance the location reasoning term with others, and" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 386, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 386, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 386, + 545, + 415 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {u n s u p} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta^ {s}, \\Theta_ {c l s}) = \\alpha \\mathcal {L} _ {c r o s s - i n} ^ {u} (\\mathcal {B}; \\Theta^ {s}) \\tag {11} \\\\ + \\beta \\mathcal {L} _ {c l s} ^ {u} (\\mathcal {B}; \\Theta_ {e n c} ^ {s}, \\Theta_ {c l s}), \\\\ \\end{array}", + "image_path": "ac47732a4e9976e7d0da03f4353a7786ed83b8a313997c92b73c2a78d76981f7.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "type": "text", + "content": " is another weight factor to balance the two terms. For " + }, + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{l}" + }, + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{u}" + }, + { + "bbox": [ + 305, + 423, + 545, + 457 + ], + "type": "text", + "content": " in Eq. 10 and Eq. 11, the following equation holds:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 384, + 458, + 545, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 458, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 384, + 458, + 545, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = \\mathcal {L} _ {c l s} ^ {l} + \\mathcal {L} _ {c l s} ^ {u}. \\tag {12}", + "image_path": "8f5a545931683353ba4514a8c1e4c70a9ae184d2abae886553f6bb6b2af90b0d.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 479, + 392, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 392, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 392, + 491 + ], + "type": "text", + "content": "3.5. Testing Phase" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "text", + "content": "Given a testing image " + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{test}" + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "text", + "content": ", the probability map is obtained by: " + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{test} = \\sigma (\\mathcal{F}(\\mathbf{X}_{test};\\Theta^{s*}))" + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\Theta^{s*}" + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "text", + "content": " denotes the well-trained student network parameters. The final label map can be determined by taking the argmax in the first dimension of " + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{test}" + }, + { + "bbox": [ + 305, + 497, + 545, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 567, + 386, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 386, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 386, + 581 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 587, + 460, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 460, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 460, + 601 + ], + "type": "text", + "content": "4.1. Datasets and Pre-processing" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": "BTCV Multi-organ Segmentation Dataset BTCV multi-organ segmentation dataset is from MICCAI Multi-Atlas Labeling Beyond Cranial Vault-Workshop Challenge [13] which contains 30 subjects with 3779 axial abdominal CT slices with 13 organs annotation. In pre-processing, we follow [27] to re-sample all the CT scans to the voxel spacing " + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "[1.5 \\times 1.5 \\times 2.0]mm^3" + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": " and normalize them to have zero mean and unit variance. Strictly following [5, 6, 10], 18 cases are divided for training and the other 12 cases are for testing." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23873" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 70, + 504, + 219 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 504, + 219 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 504, + 219 + ], + "type": "table", + "html": "
Labeled%MethodsSplR.kidL.kidGallEsoLivStoAorIVCVeinsPanRGLGAvg. DSCAvg. NSD
30%V-Net [20]86.3775.3871.2041.8148.4692.7744.7888.2072.6049.2722.3037.7921.7157.9056.90
MT [28]79.8977.5678.0838.3158.9992.2648.7388.6179.3652.7328.4254.1621.3061.4260.85
UA-MT [39]87.7573.8772.9742.0850.7993.2242.0788.3071.8055.2736.6143.2237.6761.2060.99
ICT [30]88.7878.0083.9350.0953.3588.3264.1989.0279.6063.5158.9043.2950.7068.5967.77
CPS [7]82.9876.9179.2129.9152.5392.5941.3786.5974.7349.5318.9645.1324.0658.0456.64
SS-Net [35]85.5373.3280.7030.8552.4491.6717.6786.2872.3947.4022.9537.3839.0156.7454.38
SLC-Net [16]90.0484.3885.8453.5555.8792.8657.8089.8380.8753.2238.7848.1037.0366.7868.81
MagicNet (ours)91.4284.6486.1962.8662.4993.8972.8790.7083.5270.0764.9460.8857.4875.5376.31
40%V-Net [20]84.9882.7282.0736.6463.4893.5457.4989.7478.6360.4249.3955.6038.4967.1767.84
MT [28]85.7078.9379.0842.8061.0993.4557.5789.7080.3063.9541.1450.4629.6965.6865.98
UA-MT [39]88.7475.8878.9154.2558.5593.4658.9089.2376.1562.3047.9151.5344.9267.7568.87
ICT [30]90.3184.4186.9649.2265.6594.2965.9590.2381.4469.5666.6157.3556.0173.6974.98
CPS [7]87.5672.9977.5953.3154.0892.4154.5887.7574.3258.6848.0250.3943.8665.8165.34
SS-Net [35]84.7476.3774.1943.4257.0592.9014.3783.1469.7752.4527.0854.2927.6658.2657.75
SLC-Net [16]90.0584.0086.4356.1658.9194.6870.7289.9379.4560.5954.2251.0339.0870.4072.87
MagicNet (ours)91.6185.0288.1358.1666.7294.0774.4690.7784.3171.5668.9063.4860.4776.7478.68
100%V-Net [20]84.0084.8286.3867.4265.0294.8373.7590.2784.1969.8563.5462.6065.0276.2877.45
", + "image_path": "e25ce360dfd7f03c00a94f3afa27385f3e072a7b0f857828341cc35bd84aebdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 89, + 294, + 504, + 437 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 545, + 284 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 545, + 284 + ], + "type": "text", + "content": "Table 1. Comparison results (DSCs of each organ, avg. DSC, and avg. NSD) between our method and existing semi-supervised medical image segmentation methods on the BTCV dataset. V-Net means training with only labeled samples on the V-Net backbone. DSC: Dice Similarity Coefficient. NSD: Normalized Surface Dice. The standard deviations are not reported due to space limit. Note: Spl: spleen, R.Kid: right kidney, L.Kid: left kidney, Gall: gallbladder, Eso: esophagus, Liv: liver, Sto: stomach, Aor: aorta, IVC: inferior vena cava, Veins: portal and splenic veins, Pan: pancreas, LG/RG: left/right adrenal glands." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 294, + 504, + 437 + ], + "lines": [ + { + "bbox": [ + 89, + 294, + 504, + 437 + ], + "spans": [ + { + "bbox": [ + 89, + 294, + 504, + 437 + ], + "type": "table", + "html": "
Labeled%MethodsSpleenL.kidneyGallbladderEsophagusLiverStomachPancreasDuodenumAvg. DSCAvg. NSD
10%V-Net [20]88.89 (16.31)84.90 (20.11)55.73 (34.27)58.27 (20.05)93.15 (7.11)57.15 (30.62)53.33 (22.44)35.28 (19.83)65.8449.96
MT [28]89.22 (13.56)88.04 (17.61)61.40 (31.91)58.89 (20.62)93.60 (5.04)74.00 (20.37)64.70 (19.31)44.53 (17.61)71.8053.54
UA-MT [39]89.94 (13.36)89.26 (15.67)59.19 (32.35)59.43 (19.45)93.77 (5.05)75.43 (19.61)65.86 (17.48)44.57 (17.40)72.1854.04
ICT [30]90.27 (13.70)89.89 (15.81)61.47 (30.31)58.70 (20.17)93.47 (5.39)72.45 (20.95)65.81 (18.28)45.00 (19.22)72.1354.07
CPS [7]91.73 (12.11)90.59 (14.85)60.54 (32.29)55.82 (21.03)94.19 (4.93)78.80 (15.28)62.93 (20.03)32.58 (18.68)70.9052.80
SS-Net [35]83.52 (19.56)78.28 (25.05)53.08 (33.88)56.24 (20.54)91.84 (6.07)47.65 (29.70)43.23 (24.11)25.70 (18.71)59.9442.68
SLC-Net [16]86.89 (17.54)87.54 (18.00)55.28 (34.75)58.51 (20.54)93.70 (5.33)69.74 (22.89)46.55 (26.21)28.86 (18.87)65.8913.28
MagicNet (ours)92.33 (11.12)91.19 (14.18)66.35 (29.29)70.95 (12.30)94.31 (4.64)81.56 (14.91)76.31 (10.56)61.50 (12.85)79.3162.31
20%V-Net [20]90.53 (15.33)88.28 (19.76)65.16 (30.03)66.10 (16.98)94.49 (4.18)79.56 (16.56)69.04 (17.43)51.98 (18.53)75.6459.66
MT [28]91.76 (12.77)91.40 (14.49)63.83 (31.74)64.15 (18.05)94.44 (4.86)81.19 (15.79)72.24 (15.79)57.01 (16.70)77.0060.32
UA-MT [39]92.39 (11.31)91.44 (14.61)63.63 (31.46)64.33 (17.92)94.50 (4.77)83.21 (14.03)71.93 (15.51)57.21 (17.35)77.3360.64
ICT [30]92.96 (9.60)91.67 (14.52)65.90 (30.49)64.35 (16.79)94.46 (4.66)82.60 (15.46)74.18 (12.99)57.67 (16.16)77.9761.24
CPS [7]92.66 (11.46)91.87 (14.33)64.75 (31.21)57.27 (19.56)94.97 (4.51)85.65 (9.50)74.15 (12.49)55.00 (16.58)77.0460.01
SS-Net [35]91.17 (14.77)87.77 (19.37)63.49 (30.83)65.02 (16.65)93.76 (4.82)73.04 (22.16)70.37 (17.70)52.81 (20.32)74.6857.57
SLC-Net [16]92.60 (11.24)91.38 (14.57)62.46 (31.37)62.84 (19.31)94.54 (4.72)80.30 (14.73)69.79 (16.99)53.48 (18.40)75.9259.93
MagicNet (ours)93.52 (10.22)92.01 (14.30)71.04 (26.92)70.95 (12.58)94.89 (4.42)85.77 (10.27)78.81 (8.07)66.21 (12.33)81.6565.87
100%V-Net [20]94.35 (5.81)92.30 (14.22)73.61 (28.80)73.48 (12.43)95.17 (3.10)89.09 (7.74)80.86 (7.38)70.03 (11.60)83.6168.85
", + "image_path": "fe1cd6ba1de3aba9356a1aa2107e6e9a83fa6d3036653fbdceaa2adf7b9c1177.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 544, + 468 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 544, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 544, + 468 + ], + "type": "text", + "content": "Table 2. Comparison results (DSCs of each organ, avg. DSC, and avg. NSD) between our method and existing semi-supervised medical image segmentation methods on MACT dataset under 4-fold cross-validation. The value in " + }, + { + "bbox": [ + 46, + 445, + 544, + 468 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 445, + 544, + 468 + ], + "type": "text", + "content": " is standard deviation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 483, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 289, + 639 + ], + "type": "text", + "content": "MACT Dataset Multi-organ abdominal CT reference standard segmentations (MACT) is a public multi-organ segmentation dataset, containing 90 CT volumes with 8 organs annotation. It is manually re-annotated by [11] based on 43 cases of NIH pancreas [21] and 47 cases of Synapse [13]. We follow the setting of UMCT [36]. 1) In preprocessing, we set the soft tissue CT window range of " + }, + { + "bbox": [ + 46, + 483, + 289, + 639 + ], + "type": "inline_equation", + "content": "[-125, 275]" + }, + { + "bbox": [ + 46, + 483, + 289, + 639 + ], + "type": "text", + "content": " HU and all the CT scans are re-sampled to an isotropic resolution of " + }, + { + "bbox": [ + 46, + 483, + 289, + 639 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{mm}^3" + }, + { + "bbox": [ + 46, + 483, + 289, + 639 + ], + "type": "text", + "content": ". Then, image intensities are normalized to have zero mean and unit variance. 2) We randomly split the dataset into four folds, and perform 4-fold cross-validation. Labeled training set is then randomly selected, which are applied to all experiments for fair comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "type": "text", + "content": "4.2. Experimental Setup and Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "In this work, we conduct all the experiments based on PyTorch with one NVIDIA 3090 GPU and use V-Net as our backbone [16,35,39]. For magic-cube location reasoning, we add a classification head composed of two fully" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "content": "connected layers at the end of our encoder. Firstly, for BTCV dataset, the framework is trained by an SGD optimizer for 70K iterations, with an initial learning rate (lr) 0.01 with a warming-up strategy: " + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "inline_equation", + "content": "lr = \\text{base\\_lr} \\times (1 - \\frac{\\text{iterations}}{\\text{max\\_iterations}})^{0.9}" + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "content": ". For MACT dataset, the framework is trained by an SGD optimizer for 30K iterations, with an initial learning rate (lr) 0.01 decayed by 0.1 every 12K iterations. Following [16, 18, 35, 36, 39], for both two datasets, batch size is set as 4, including 2 labeled images and 2 unlabeled images, and the hyper-parameter " + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "content": " is set using a time-dependent Gaussian warming-up function and " + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "content": " is empirically set as 0.1 [10]. We randomly crop volume of size " + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "inline_equation", + "content": "96^3" + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "content": ", and then we partition the volume into " + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "content": " small cubes, forming our magic-cubes, and the choice of " + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "content": " is also discussed in Sec. 4.4. In the testing phase, a sliding window strategy is adopted to obtain final results with a stride of " + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "inline_equation", + "content": "16^3" + }, + { + "bbox": [ + 304, + 483, + 545, + 710 + ], + "type": "text", + "content": ". The final evaluation metrics for our method are DSC (\\%, Dice-Sørensen Coefficient) and NSD (\\%, Normalized Surface Dice) which are commonly used in the cur" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23874" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 71, + 519, + 137 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 519, + 137 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 519, + 137 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 519, + 137 + ], + "type": "table", + "html": "
MethodsCrossInLocBldAvg. DSCspleenl.kidneygallbladderesophagusliverstomachpancreasduodenum
Baseline (MT [28])71.8886.8488.9254.5053.4093.0170.8671.7755.76
Cross✓78.1392.1190.5363.4270.0593.9079.6075.5859.87
Cross + In✓✓78.8392.8290.7564.9170.6893.9080.9276.2560.40
Cross + Loc✓✓78.5692.4991.0765.9569.7294.1280.4075.8258.94
Cross + In + Loc✓✓✓78.8792.5591.1865.7270.7794.4280.0575.8160.43
Cross + In + Loc + Bld✓✓✓✓79.3192.3391.1966.3570.9594.3181.5676.3161.50
", + "image_path": "11b7c51ad54a93186e5d8326d4efd3a1ca5ccd2c1fe51d31f94a58ffa94666cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 544, + 168 + ], + "lines": [ + { + "bbox": [ + 46, + 145, + 544, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 544, + 168 + ], + "type": "text", + "content": "Table 3. Ablation study (DSC %) for the effectiveness of each component of MagicNet. Cross: cross-image partition-and-recovery. In: Within-image partition-and-recovery. Bld: cube-wise pseudo-label blending. Loc: magic-cube location reasoning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 186, + 285, + 258 + ], + "blocks": [ + { + "bbox": [ + 50, + 186, + 285, + 258 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 285, + 258 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 285, + 258 + ], + "type": "image", + "image_path": "c19517246faf0d44e4777c913db537fadc2db62d82121fbed446ae4fc0b96413.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 270, + 286, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 270, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 286, + 293 + ], + "type": "text", + "content": "Figure 5. DSC comparison in box plots on MACT dataset with " + }, + { + "bbox": [ + 47, + 270, + 286, + 293 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 270, + 286, + 293 + ], + "type": "text", + "content": " labeled images (Best viewed electronically, zoom in)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 305, + 247, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 247, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 247, + 318 + ], + "type": "text", + "content": "rent multi-organ segmentation challenge [12, 19]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 326, + 286, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 286, + 338 + ], + "type": "text", + "content": "4.3. Comparison with the State-of-the-art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 342, + 286, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 286, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 286, + 461 + ], + "type": "text", + "content": "We compare MagicNet with six state-of-the-art semi-supervised segmentation methods, including mean-teacher (MT) [28], uncertainty-aware mean-teacher (UA-MT) [39], interpolated consistency training (ICT) [30], cross pseudosupervision (CPS) [7], smoothness and class-separation (SS-Net) [35], shape-awareness and local constraints (SLC-Net) [16]. All experiments of other methods are re-trained by their official code when transferred to multi-organ task. V-Net is adopted as our backbone which is the same as the above methods for fair comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 462, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 286, + 605 + ], + "type": "text", + "content": "In Table 1, we summarize the results of BTCV dataset. Visualizations compared with state-of-the-arts are shown in the supplementary material. Noticeable improvements compared with state-of-the-arts can be seen for organs such as Stomach (↑ 8.68), Gallbladder (↑ 9.31), Veins (↑ 6.56), Right adrenal glands (↑ 6.72) and Left adrenal glands (↑ 6.78). It is interesting to observe that some semi-supervised methods even perform worse than the lower bound. This is because these methods are less able to learn common semantics from the labeled images to unlabeled ones, generating less accurate pseudo-labels, which even hamper the usage of unlabeled images (see the supplementary material)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": "We then evaluate MagicNet on MACT dataset. In Table 2, we can observe: (1) a more significant performance improvement between MagicNet and other state-of-the-arts with " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " labeled data than " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": ", which demonstrates the effectiveness of MagicNet when training with fewer labeled CT images; (2) MagicNet successfully mitigates the inferior performance in tail classes, e.g., DSC improvement with " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " labeled images on Gallbladder (" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " 4.88), Esophagus (" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " 11.52), Pancreas (" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " 10.45) and Duodenum (" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " 16.50). Fig. 5" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 186, + 529, + 236 + ], + "blocks": [ + { + "bbox": [ + 323, + 186, + 529, + 236 + ], + "lines": [ + { + "bbox": [ + 323, + 186, + 529, + 236 + ], + "spans": [ + { + "bbox": [ + 323, + 186, + 529, + 236 + ], + "type": "image", + "image_path": "bb2aca65c72d67cb7f1d7a531cc9ce329df66025d2566aa5d2086de281c5b9d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 247, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 280 + ], + "type": "text", + "content": "Figure 6. Scramble position and keep position when mixing small-cubes across images (from Fig. 3). Red arrow indicates the cube location in mixed images." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 298, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 358 + ], + "type": "text", + "content": "shows a comparison in DSC of six methods for different organs. Note that UMCT [36] on MACT dataset achieves " + }, + { + "bbox": [ + 305, + 298, + 545, + 358 + ], + "type": "inline_equation", + "content": "75.33\\%" + }, + { + "bbox": [ + 305, + 298, + 545, + 358 + ], + "type": "text", + "content": " in DSC. It uses AH-Net [17] as the backbone (3 are used), whose parameter number is much larger. Due to computation resources, we do not directly compare with it." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 362, + 399, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 362, + 399, + 375 + ], + "spans": [ + { + "bbox": [ + 306, + 362, + 399, + 375 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": "We conduct ablation studies on MACT dataset [11] with " + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": " labeled data to show the effectiveness of each module. The effectiveness of each component in MagicNet We conduct ablation studies to show the impact of each component in MagicNet. In Table 3, the first row indicates the mean-teacher baseline model [28], which our method is designed on. Compared to the baseline, our magic-cube partition and recovery can yield good segmentation performance. Cross and In represent our cross- and within-image partition-and-recovery, which increase the performance from " + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "inline_equation", + "content": "71.88\\%" + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "inline_equation", + "content": "78.13\\%" + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "inline_equation", + "content": "78.87\\%" + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": ", respectively. This shows the powerful ability with our specially-designed data augmentation method. Loc represents magic-cube location reasoning for within-image branch. We can see from the table that based on our cross-image branch, adding relative locations of small-cubes can achieve " + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "inline_equation", + "content": "78.56\\%" + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": " in DSC, adding only within-image partition-and-recovery branch achieves " + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "inline_equation", + "content": "78.83\\%" + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": ", and adding both two branches leads to " + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "inline_equation", + "content": "78.87\\%" + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": ". Finally, our proposed cube-wise pseudo label blending (short for Bld in Table 3) provides significant improvement to " + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "inline_equation", + "content": "79.31\\%" + }, + { + "bbox": [ + 305, + 378, + 545, + 640 + ], + "type": "text", + "content": ", for which we will make a more complete analysis later in the current section." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "content": "Design choices of partition and recovery Here, we discuss the design for cross-image partition-and-recovery branch: ① Should we maintain or scramble the magic-cube relative locations when manipulating our cross-image branch, as shown in Fig. 6? The comparison results are shown in the last two rows in Table 4. Scramble and Keep represent the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23875" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 73, + 214, + 122 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 214, + 122 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 214, + 122 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 214, + 122 + ], + "type": "table", + "html": "
①②DSCNSD
scrambleU69.53 ± 11.2750.98 ± 11.00
keepU73.84 ± 9.9655.80 ± 10.81
scrambleLU71.91 ± 11.2054.10 ± 11.28
keepLU78.13 ± 8.1260.79 ± 9.90
", + "image_path": "171e06bccc357fee41d167042fad4842fc7bca939d6dcdd8012aba790f33df13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 223, + 72, + 382, + 123 + ], + "blocks": [ + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "lines": [ + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "type": "text", + "content": "Table 4. Ablation of design choices for crossimage partition and recovery (Question " + }, + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "type": "text", + "content": ", mean " + }, + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 47, + 130, + 216, + 197 + ], + "type": "text", + "content": " std of all cases). Scramble/keep: ignore/keep original positions when mixed. U: only for unlabeled data. LU: for both labeled and unlabeled data. The last row is ours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 223, + 72, + 382, + 123 + ], + "lines": [ + { + "bbox": [ + 223, + 72, + 382, + 123 + ], + "spans": [ + { + "bbox": [ + 223, + 72, + 382, + 123 + ], + "type": "table", + "html": "
DSCNSD
teacher sup.78.47 ± 7.8361.52 ± 9.21
mutual sup.70.45 ± 11.4753.71 ± 11.78
blending79.31 ± 7.5562.31 ± 9.08
", + "image_path": "f20adcb1abef0882c26c830d322c16bc3663cf3624f813bc665ba663d3f99ea4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 391, + 70, + 544, + 126 + ], + "blocks": [ + { + "bbox": [ + 220, + 131, + 383, + 175 + ], + "lines": [ + { + "bbox": [ + 220, + 131, + 383, + 175 + ], + "spans": [ + { + "bbox": [ + 220, + 131, + 383, + 175 + ], + "type": "text", + "content": "Table 5. Comparison of different pseudolabel supervision/blending strategies for unlabeled data. Sup.: supervision. Blending: our cube-wise pseudo-label blending." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 391, + 70, + 544, + 126 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 544, + 126 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 544, + 126 + ], + "type": "table", + "html": "
DSCNSD
CutMix [40]73.80 ± 9.9955.86 ± 10.54
CutOut [8]72.47 ± 10.2754.62 ± 10.67
MixUp [41]72.13 ± 11.3254.07 ± 12.01
Ours (2)78.68 ± 7.4560.78 ± 9.12
Ours (3)79.31 ± 7.5562.31 ± 9.08
", + "image_path": "d439f165b37e976fbf6a22b368cfa711c37b7801fff63234154532a64a9f39bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 388, + 134, + 545, + 189 + ], + "lines": [ + { + "bbox": [ + 388, + 134, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 388, + 134, + 545, + 189 + ], + "type": "text", + "content": "Table 6. Comparison of different data augmentation methods, where we try different CutMix and CutOut sizes, and choose the best results. For MagicNet, we compare different " + }, + { + "bbox": [ + 388, + 134, + 545, + 189 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 388, + 134, + 545, + 189 + ], + "type": "text", + "content": " values in " + }, + { + "bbox": [ + 388, + 134, + 545, + 189 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 388, + 134, + 545, + 189 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 217, + 286, + 335 + ], + "blocks": [ + { + "bbox": [ + 48, + 217, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 48, + 217, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 286, + 335 + ], + "type": "image", + "image_path": "6e54eb8b9e5b5fdeb34f6dfe547f9cbd16ef87b5eb64a09905e5362b380150d0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 288, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 407 + ], + "type": "text", + "content": "Figure 7. Kernel Density Estimation of labeled and unlabeled samples on the first three channels of the feature map of \"only unlabeled images\" and \"both labeled and unlabeled images\". Considerable distribution mismatch between labeled and unlabeled images can be observed (top), while our method can well improve the distribution mismatch (bottom). Best viewed zoom in." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 422, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 543 + ], + "type": "text", + "content": "partitioned small-cubes are randomly mixed while ignoring their original locations or kept when mixing them across images. Results show that the relative locations between multiple organs are important for multi-organ segmentation. ② Should our cross-image data augmentation be operated on only unlabeled images (see U in Table 4) or both labeled and unlabeled images (see LU in Table 4)? The latter obtains a much better performance compared to the former. Besides, it largely reduces the distribution mismatch between labeled and unlabeled data, as shown in Fig. 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "type": "text", + "content": "Cube-wise pseudo-label blending As illustrated in Sec. 3.2, we blend the output of within-image branch and the output of teacher model to obtain the final pseudo-label for complementing local attributes. We compare our blending with two other methods, as shown in Table 5. Three supervision schemes, illustrated in Fig. 8, are compared for unlabeled images based on our framework. Teacher sup. means the outputs of cross- and within-image branch are both supervised by the pseudo-label from the teacher model. Mutual sup. means the outputs of cross- and within-image are mutually supervised. It can be seen that our blending strategy works favorably for unlabeled data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Comparison with interpolated-based methods As shown in Table 6, our augmentation method outperforms other" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 353, + 214, + 502, + 270 + ], + "blocks": [ + { + "bbox": [ + 353, + 214, + 502, + 270 + ], + "lines": [ + { + "bbox": [ + 353, + 214, + 502, + 270 + ], + "spans": [ + { + "bbox": [ + 353, + 214, + 502, + 270 + ], + "type": "image", + "image_path": "374e9b3d55d9c607d4ddf82bb30d4f1db25d33b5b20ad3fdadb659936534845a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 279, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 305, + 279, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 545, + 312 + ], + "type": "text", + "content": "Figure 8. Different pseudo-label supervision/blending strategies corresponding to Table 5. Blue line indicates supervision. (a) teacher sup. (b) mutual sup. (c) our blending." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 329, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 363 + ], + "type": "text", + "content": "methods such as CutMix, CutOut and MixUp. For fair comparison, we try several CutMix and CutOut sizes, and choose the best results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "text", + "content": "Different number " + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "text", + "content": " of small-cubes We study the impact of different numbers of small-cubes " + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "text", + "content": ", as shown in Table 6. Slightly better performance is achieved when " + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "text", + "content": ", the size of the small cube does not match the condition for VNet. Thus, we only compare the results given " + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 304, + 364, + 546, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 447, + 463, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 463, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 463, + 459 + ], + "type": "text", + "content": "5. Limitations and Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 467, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 586 + ], + "type": "text", + "content": "We have presented MagicNet for semi-supervised multi-organ segmentation. MagicNet encourages unlabeled data to learn organ semantics in relative locations from the labeled data, and enhances the learning ability for small organs. Two essential components are proposed: (1) partition-and-recovery " + }, + { + "bbox": [ + 304, + 467, + 545, + 586 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 304, + 467, + 545, + 586 + ], + "type": "text", + "content": " small cubes cross- and within-labeled and unlabeled images, and (2) the quality of pseudo label refinement by blending the cube-wise representations. Experimental results on two public multi-organ segmentation datasets verify the superiority of our method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 586, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 624 + ], + "type": "text", + "content": "Limitations: CT scans are all roughly aligned by the nature of the data. Magic-cube may not work well on domains " + }, + { + "bbox": [ + 304, + 586, + 545, + 624 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 304, + 586, + 545, + 624 + ], + "type": "text", + "content": " such prior, e.g., microscopy images " + }, + { + "bbox": [ + 304, + 586, + 545, + 624 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 304, + 586, + 545, + 624 + ], + "type": "text", + "content": " similar prior." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 633, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 416, + 647 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (Grant No. 62101191), Shanghai Natural Science Foundation (Grant No. 21ZR1420800), and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ2229004)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23876" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "type": "text", + "content": "[1] Wenjia Bai, Ozan Oktay, Matthew Sinclair, Hideaki Suzuki, Martin Rajchl, Giacomo Tarroni, Ben Glocker, Andrew P. King, Paul M. Matthews, and Daniel Rueckert. Semi-supervised learning for network-based cardiac MR image segmentation. In Proc. MICCAI, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "type": "text", + "content": "[2] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In Proc. ICLR, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 196, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 287, + 239 + ], + "type": "text", + "content": "[3] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Proc. NeurIPS, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 242, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 242, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 242, + 287, + 285 + ], + "type": "text", + "content": "[4] Gerda Bortsova, Florian Dubost, Laurens Hogeweg, Ioannis Katramados, and Marleen de Bruijne. Semi-supervised medical image segmentation via learning consistency under transformations. In Proc. MICCAI, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 289, + 286, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 286, + 332 + ], + "type": "text", + "content": "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "type": "text", + "content": "[6] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021.5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 286, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 425 + ], + "type": "text", + "content": "[7] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proc. CVPR, 2021. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 427, + 286, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 427, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 53, + 427, + 286, + 460 + ], + "type": "text", + "content": "[8] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 462, + 286, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 462, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 462, + 286, + 495 + ], + "type": "text", + "content": "[9] Kang Fang and Wu-Jun Li. Dmnet: Difference minimization network for semi-supervised segmentation in medical images. In Proc. MICCAI, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 286, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 286, + 540 + ], + "type": "text", + "content": "[10] Shuhao Fu, Yongyi Lu, Yan Wang, Yuyin Zhou, Wei Shen, Elliot Fishman, and Alan Yuille. Domain adaptive relational reasoning for 3d multi-organ segmentation. In Proc. MIC-CAI, 2020. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 609 + ], + "type": "text", + "content": "[11] Eli Gibson, Francesco Giganti, Yipeng Hu, Ester Bonmati, Steve Bandula, Kurinchi Gurusamy, Brian Davidson, Stephen P. Pereira, Matthew J. Clarkson, and Dean C. Barratt. Automatic multi-organ segmentation on abdominal ct with dense v-networks. IEEE Trans. Medical Imaging, 37(8):1822-1834, 2018. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 286, + 667 + ], + "type": "text", + "content": "[12] Yuanfeng Ji, Haotian Bai, Jie Yang, Chongjian Ge, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhang, Wanling Ma, Xi-ang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[13] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, T Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Workshop, 2015. 2, 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Shuai Lin Li, Chuyu Zhang, and Xuming He. Shape-aware semi-supervised 3d semantic segmentation for medical images. In Proc. MICCAI, 2020. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 161 + ], + "type": "text", + "content": "[15] Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, and Pheng-Ann Heng. Transformation consistent self-ensembling model for semi-supervised medical image segmentation. IEEE Trans. Neural Networks and Learning Systems, 32(2):523-534, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[16] Jinhua Liu, Christian Desrosiers, and Yuanfeng Zhou. Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape awareness and local context constraints. In Proc. MICCAI, 2022. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "text", + "content": "[17] Siqi Liu, Daguang Xu, S Kevin Zhou, Olivier Pauly, Sasa Grbic, Thomas Mertelmeier, Julia Wicklein, Anna Jerebko, Weidong Cai, and Dorin Comaniciu. 3d anisotropic hybrid network: Transferring convolutional features from 2d images to 3d anisotropic volumes. In Proc. MICCAI, 2018. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[18] Xiangde Luo, Jieneng Chen, Tao Song, and Guotai Wang. Semi-supervised medical image segmentation through dual-task consistency. In Proc. AAAI, 2021. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 299, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 374 + ], + "type": "text", + "content": "[19] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, Shucheng Cao, Qi Zhang, Shangqing Liu, Yunpeng Wang, Yuhui Li, Jian He, and Xiaoping Yang. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6695-6714, 2022. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "text", + "content": "[20] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Proc. 3DV, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[21] Holger R Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim B Turkbey, and Ronald M Summers. Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In Proc. MICCAI, 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "type": "text", + "content": "[22] Holger R. Roth, Hirohisa Oda, Xiangrong Zhou, Natsuki Shimizu, Ying Yang, Yuichiro Hayashi, Masahiro Oda, Michtaka Fujiwara, Kazunari Misawa, and Kensaku Mori. An application of cascaded 3d fully convolutional networks for medical image segmentation. Comput. Medical Imaging Graph., 66:90-99, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[23] Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(3):608-620, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "text", + "content": "[24] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Proc. NeurIPS, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[25] Hao Tang, Xuming Chen, Yang Liu, Zhipeng Lu, Junhua You, Mingzhou Yang, Shengyu Yao, Guoqi Zhao, Yi Xu, Tingfeng Chen, Yong Liu, and Xiaohui Xie. Clinically applicable deep learning framework for organs at risk delineation in CT images. Nat. Mach. Intell., 1(10):480-491, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[26] Yucheng Tang, Riqiang Gao, Ho Hin Lee, Shizhong Han, Yunqiang Chen, Dashan Gao, Vishwesh Nath, Camilo" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23877" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "text", + "content": "Bermudez, Michael R. Savona, Richard G. Abramson, Shunxing Bao, Ilwoo Lyu, Yuankai Huo, and Bennett A. Landman. High-resolution 3d abdominal segmentation with random patch network fusion. Medical Image Anal., 69:101894, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 182 + ], + "type": "text", + "content": "[27] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proc. CVPR, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "type": "text", + "content": "[28] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proc. NeurIPS, 2017. 2, 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "type": "text", + "content": "[29] Peng Tu, Yawen Huang, Feng Zheng, Zhenyu He, Liujuan Cao, and Ling Shao. Guidedmix-net: Semi-supervised semantic segmentation by using labeled images as reference. In Proc. AAAI, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[30] Vikas Verma, Alex Lamb, Juho Kannala, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi-supervised learning. In Proc. IJCAI, 2019. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "text", + "content": "[31] Kaiping Wang, Bo Zhan, Chen Zu, Xi Wu, Jiliu Zhou, Luping Zhou, and Yan Wang. Semi-supervised medical image segmentation via a tripled-uncertainty guided mean teacher model with contrastive learning. Medical Image Anal., 79:102447, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "text", + "content": "[32] Qin Wang, Wen Li, and Luc Van Gool. Semi-supervised learning by augmented distribution alignment. In Proc. ICCV, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[33] Yan Wang, Yuyin Zhou, Wei Shen, Seyoun Park, Elliot K. Fishman, and Alan L. Yuille. Abdominal multi-organ segmentation with organ-attention networks and statistical fusion. Medical Image Anal., 55:88-102, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[34] Chen Wei, Lingxi Xie, Xutong Ren, Yingda Xia, Chi Su, Jiaying Liu, Qi Tian, and Alan L. Yuille. Iterative reorganization with weak spatial constraints: Solving arbitrary jigsaw puzzles for unsupervised representation learning. In Proc. CVPR, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[35] Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, and Jianfei Cai. Exploring smoothness and class-separation for semi-supervised medical image segmentation. In Proc. MIC-CAI, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "type": "text", + "content": "[36] Yingda Xia, Dong Yang, Zhiding Yu, Fengze Liu, Jinzheng Cai, Lequan Yu, Zhuotun Zhu, Daguang Xu, Alan L. Yuille, and Holger Roth. Uncertainty-aware multi-view co-training for semi-supervised medical image segmentation and domain adaptation. Medical Image Anal., 65:101766, 2020. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "type": "text", + "content": "[37] Lingxi Xie, Qihang Yu, Yuyin Zhou, Yan Wang, Elliot K. Fishman, and Alan L. Yuille. Recurrent saliency transformation network for tiny target segmentation in abdominal ct scans. IEEE Trans. Medical Imaging, 39(2):514-525, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[38] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence H. Staib, and James S. Duncan. Simcvd: Simple contrastive voxelwise representation distillation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(9):2228-2237, 2022. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 285 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[39] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Proc. MICCAI, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[40] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proc. ICCV, 2019. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[41] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In Proc. ICLR, 2018. 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[42] Xinkai Zhao, Chaowei Fang, De-Jun Fan, Xutao Lin, Feng Gao, and Guanbin Li. Cross-level contrastive learning and consistency constraint for semi-supervised medical image segmentation. In Proc. ISBI, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[43] Yuyin Zhou, Yan Wang, Peng Tang, Song Bai, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Semi-supervised 3d abdominal multi-organ segmentation via deep multi-planar co-training. In Proc. WACV, 2019. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23878" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_content_list.json b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..54169263e60e7658c3fbac8a624f3be4d0e9345c --- /dev/null +++ b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_content_list.json @@ -0,0 +1,2139 @@ +[ + { + "type": "text", + "text": "MagicPony: Learning Articulated 3D Animals in the Wild", + "text_level": 1, + "bbox": [ + 191, + 130, + 779, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shangzhe Wu* Ruining Li* Tomas Jakab* Christian Rupprecht Andrea Vedaldi Visual Geometry Group, University of Oxford", + "bbox": [ + 138, + 180, + 831, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{szwu, ruining, tomj, chrisr,VEDaldi}@robots.ox.ac.uk 3dmagicpony.github.io", + "bbox": [ + 245, + 223, + 723, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0181a9c1d5c0e7932ba2ebc5e93123a1a020bf44e37b64c52f3f959d9833d0b1.jpg", + "image_caption": [ + "Figure 1. Learning Articulated 3D Animals in the Wild. Our method trains on a collection of single-view images of an animal category and produces a model that can predict the articulated 3D shape of a new instance from a single test image, which can be animated and relit." + ], + "image_footnote": [], + "bbox": [ + 81, + 263, + 870, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 458, + 313, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider the problem of predicting the 3D shape, articulation, viewpoint, texture, and lighting of an articulated animal like a horse given a single test image as input. We present a new method, dubbed MagicPony, that learns this predictor purely from in-the-wild single-view images of the object category, with minimal assumptions about the topology of deformation. At its core is an implicit-explicit representation of articulated shape and appearance, combining the strengths of neural fields and meshes. In order to help the model understand an object's shape and pose, we distil the knowledge captured by an off-the-shelf self-supervised vision transformer and fuse it into the 3D model. To overcome local optima in viewpoint estimation, we further introduce a new viewpoint sampling scheme that comes at no additional training cost. MagicPony outperforms prior work on this challenging task and demonstrates excellent generalisation in reconstructing art, despite the fact that it is only trained on real images. The code can be found on the project page at https://3dmagicpony.github.io/.", + "bbox": [ + 75, + 489, + 473, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing the 3D shape of an object from a single image of it requires knowing a priori what are the possible shapes and appearances of the object. Learning", + "bbox": [ + 75, + 832, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such a prior usually requires ad-hoc data acquisition setups [2, 20, 35, 36], involving at least multiple cameras, and often laser scanners, domes and other hardware, not to mention significant manual effort. This is viable for certain types of objects such as humans that are of particular interest in applications, but it is unlikely to scale to the long tail of objects that can appear in natural images. The alternative is to learn a 3D prior from 2D images only, which are available in abundance. However, this prior is highly complex and must be learned while using it to reconstruct in 3D the 2D training data, which is a major challenge.", + "bbox": [ + 496, + 460, + 893, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose MagicPony, a novel approach to learning 3D models of articulated object categories such as horses and birds with only single-view input images for training. We leverage recent progress in unsupervised representation learning, unsupervised image matching, efficient implicit-explicit shape representations and neural rendering, and devise a new auto-encoder architecture that reconstructs the 3D shape, articulation and texture of each object instance from a single image. For training, we only require a 2D segmenter for the object category and a description of the topology and symmetry of its 3D skeleton (i.e., the number and connectivity of bones). We do not require apriori knowledge of the objects' 3D shapes, keypoints, viewpoints, or of any other 2D or 3D cue which are often used in prior work [14, 21, 22, 31]. From this, we learn a function that, at test time, can estimate the shape and texture of a new object from a single image, in a feedforward manner. The function exhibits remarkable gener", + "bbox": [ + 496, + 628, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8792", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "alisation properties, including reconstructing objects in abstract drawings, despite being trained on real images only.", + "bbox": [ + 76, + 90, + 467, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to learn such disentangled 3D representations simply from raw images, MagicPony addresses a few key challenges. The first challenge is viewpoint estimation. Before the 3D model is available, it is very difficult to assign a viewpoint to the objects in the training images. To tackle this problem, prior works have often assumed additional cues such as 2D keypoint correspondences [21, 22, 31]. Instead, we avoid requiring additional keypoint supervision and implicitly infer noisy correspondences by fusing into the 3D model knowledge distilled from DINO-ViT [5], a self-supervised visual transformer network (ViT) [28]. We also develop a new efficient disambiguation scheme that explores multiple viewpoint assignment hypotheses at essentially no cost, avoiding local optima that are caused by greedily matching the noisy 2D correspondences.", + "bbox": [ + 75, + 123, + 467, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second challenge is how to represent the 3D shape, appearance and deformations of the object. Most prior works have used textured meshes [13, 14, 21, 31, 32, 67, 70], but these are difficult to optimise from scratch, leading to problems that often require ad-hoc heuristics such as remeshing [13, 70]. The other and increasingly popular approach is to use a volumetric representation such as a neural radiance field [3, 40, 52, 63], which can model complex shapes, including manipulating their topology during training. However, this modelling freedom comes at the cost of over-parametrisation, which is particularly problematic in monocular reconstruction and often leads to meaningless short-cut solutions [63]. Furthermore, modelling articulation with a volumetric representation is difficult. A posing transformation is only defined for the object's surface and interior and is more easily expressed from the canonical/pose-free space to the posed space. However, rendering a radiance field requires transforming 3D points off the object surface and in the direction opposite to the posing transformation, which is hard [8].", + "bbox": [ + 75, + 352, + 467, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We address these issues by using a hybrid volumetric-mesh representation based on DMTet [42, 56]. Shape and appearance are defined volumetrically in canonical space, but a mesh is extracted on the fly for posing and rendering. This sidesteps the challenges of using neural rendering directly while retaining most of its advantages and enabling the use of powerful shape regularisers.", + "bbox": [ + 75, + 656, + 467, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarise, we make the following contributions: (1) A new 3D object learning framework that combines recent advances in unsupervised learning, 3D representations and neural rendering, achieving better reconstruction results with less supervision; (2) An effective mechanism for fusing self-supervised features from DINO-ViT into the 3D model as a form of self-supervision; and (3) an efficient multi-hypothesis viewpoint prediction scheme that avoids local optima in reconstruction with no additional cost.", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/835ecf62d14f539f46d9cccb31c30a3f706f676b3cb44bbd135276b51f94ba2d.jpg", + "table_caption": [ + "Table 1. Related Work Overview on Weakly-supervised Learning of 3D Objects. Annotations: template shape, viewpoint, 2D keypoint, object mask, optical flow, video, coarse template shape from keypoints, camera estimated from keypoints using SfM, outputs texture flow, shape bases initialised from CMR. UMR relies on part segmentations from SCOPS [18]." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionOutput
3D2.5DMotion ViewTexture
Unsup3D [69]✓✓✓
CSM [30]✓✓✓
A-CSM [29]✓✓✓✓
CMR [21](✓)1(✓)2✓✓✓✓(✓)3
U-CMR [14]✓✓✓✓✓
UMR†[32]✓✓✓(✓)3
ACMR [31](✓)4(✓)2✓✓✓✓(✓)3
DOVE [67]✓✓✓✓✓✓
Ours✓✓✓✓✓
", + "bbox": [ + 504, + 183, + 883, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compare our method to prior work on several challenging articulated animal categories and show that our method obtains significantly better quantitative and qualitative results while using significantly less supervision, and also demonstrate generalisation to abstract drawings.", + "bbox": [ + 498, + 373, + 890, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 477, + 640, + 493 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly-supervised Learning of 3D Objects. Most related to our work are weakly-supervised methods for single-image 3D object reconstruction that learn from a collection of images or videos of an object category [14, 17, 21, 23, 24, 26, 27, 31, 32, 34, 65, 67, 68]. Due to the inherent ambiguity of the problem, these methods usually rely on heavy geometric supervision (see Tab. 1) in addition to object masks, such as 2D keypoints [21, 31], template shapes [14, 29, 30]. UMR [32] forgoes the requirement of keypoints and template shapes by leveraging weakly-supervised object part segmentations (SCOPS [18]), but produces coarse symmetric shapes similar to CMR [21]. DOVE [67] learns coarse articulated objects by exploiting temporal information from videos with optical flow supervision. All these methods require object masks for supervision, except Unsup3D [69] which exploits bilateral symmetry for reconstructing roughly frontal objects, like faces, and UNICORN [41] which uses a progressive conditioning strategy with a heavily constrained bottleneck, leading to coarse reconstructions. Another emerging paradigm is to leverage generative models [7, 45, 49, 55] that encourage images rendered from viewpoints sampled from a prior distribution to be realistic. These models, however, often rely heavily on an accurate estimation of viewpoint distribution and/or a powerful 2D image generative model, both of which are difficult to obtain in small-data scenarios.", + "bbox": [ + 496, + 507, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8793", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimization from Multi-views and Videos. 3D reconstruction traditionally relies on epipolar geometry of multiview images of static scenes [11, 16]. Neural Radiance Fields (NeRF) [3, 40, 52, 63] have recently emerged as a powerful volumetric representation for multi-view reconstruction given accurate cameras. A recent line of work, LASR [70], ViSER [71] and BANMo [72], optimises 3D shapes of deformable objects from a small set of monocular videos, with heavily engineered optimisation strategies using optical flow and mask supervision, and additionally DensePose [44] annotations for highly deformable animals. Concurrent work of LASSIE [73] also leverages DINO-ViT [5] image features for supervision but optimises a part-based model on a small collection of images ( $\\sim 30$ ) of an object category. It uses a shared shape model for all instances with only per-instance articulation.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learnable Deformable 3D Representations. A common 3D representation for weakly-supervised shape learning is triangular meshes. Deformation on meshes can be modelled by estimating offsets of individual vertices [21, 31, 67], which typically requires heavy regularisation, such as As-Rigid-As-Possible (ARAP) [57]. To constrain the space of deformations, many works often utilise lower-dimensional models, such as cages [74] or linear blend skinning controlled with skeletal bones [36, 67] or Gaussian control points [70, 71]. Parametric models, like SMPL [36] and SMAL [76], allow for realistic control of the shape through learned parameters but often require an extensive collection of 3D scans to learn from.", + "bbox": [ + 75, + 343, + 472, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic neural fields [12,46,48,50,51,58,59,72], such as D-NeRF [50], extend NeRF [40] with time-varying components. A-NeRF [58] proposes a radiance field conditioned on an articulated skeleton, while BANMo [72] uses learned Gaussian control points similarly to [70, 71]. One key limitation of these implicit representations is the requirement of an inverse transformation from 3D world coordinates back to the canonical space, which is often harder to learn compared to forward deformation [8]. We propose a hybrid SDF-mesh representation, extending DMTet [56] with an articulation mechanism, which combines the expressiveness of implicit models with the simplicity of mesh deformation and articulation.", + "bbox": [ + 75, + 542, + 472, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Viewpoint Prediction in the Wild. Viewpoint prediction in weakly-supervised settings is challenging, as it is prone to local optima induced by projection and common object symmetries. Existing solutions are based on learning a distribution over multiple possible viewpoints. U-CMR [14] proposes a camera-multiplex that optimises over 40 cameras for each training sample using a given template shape, which is expensive as evaluating each hypothesis involves a rendering step. We propose a new viewpoint exploration scheme that requires evaluating only one hypothesis in each", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "step with essentially no added training cost and does not rely on a given template shape but a jointly learned model. Recent work of Implicit-PDF [43] proposes to learn viewpoint distributions implicitly by estimating a probability for each image-pose pair of a fixed object instance. RelPose [75] extends this to learning relative poses of image pairs. Our model also predicts a probability associated with each viewpoint hypothesis but instead relates an instance to a learned category-level prior model.", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 239, + 591, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given only a collection of single-view images of a deformable object category collected in the wild, our goal is to train a monocular reconstruction network that, at test time, predicts the 3D shape, articulated pose, albedo and lighting from a single image of the object.", + "bbox": [ + 496, + 265, + 893, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, our method requires no geometric supervision other than the images, 2D foreground segmentations of the object obtained automatically from an off-the-shelf method like [25], and a description of the topology of the object's skeleton (e.g., the number of legs). It is based on several key ideas: an implicit-explicit shape representation (Sec. 3.1), a hierarchical shape representation, from a generic template to articulated instances (Sec. 3.2), and a viewpoint prediction scheme that leverages self-supervised correspondences and efficient multi-hypothesis exploration to avoid local minima (Sec. 3.4), as illustrated in Fig. 2.", + "bbox": [ + 496, + 340, + 893, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Implicit-Explicit 3D Shape Representation", + "text_level": 1, + "bbox": [ + 498, + 513, + 862, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The choice of representation for the 3D shape of the object is critical. It must be (1) sufficiently expressive for modelling fine-grained shape details and deformations and (2) sufficiently regular for learning with only weak supervision via image reconstruction.", + "bbox": [ + 496, + 537, + 893, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most 3D representations are either explicit (triangular meshes [21, 67, 70]) or implicit (volumes, SDFs, radiance fields [19, 39, 40, 47, 66]). Explicit representations are more compact, easily support powerful deformation models like blend skinning, and the geometric smoothness can be easily measured and controlled; however, they are more difficult to optimise and often result in defects like folds during training, which requires occasional re-meshing [13, 70]. Implicit representations are more amenable to gradient-based optimisation, including allowing topological changes, but are more computationally expensive and, due to their high capacity, require substantial supervision, usually in the form of a dense coverage of the different object viewpoints [40, 66] or 3D ground-truth [39, 47].", + "bbox": [ + 496, + 614, + 893, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we adopt a representation that combines the advantages of both. The shape of the object is represented implicitly by a neural field that is converted on the fly into an explicit mesh via the Differentiable Marching Tetrahedra (DMTet) [9, 42, 56] method. Specifically, the object's shape", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8794", + "bbox": [ + 482, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f95d58eb7b60e8aa8d832efa4be77d292fcbe6ba108adc2199c5dfdc6c85839e.jpg", + "image_caption": [ + "Figure 2. Training Pipeline. Given a collection of single-view images of an animal category, our model learns a category-specific prior shape using an implicit-explicit representation, together with a feature field that allows us to fuse self-supervised correspondences through a feature rendering loss. This prior shape is deformed, articulated and shaded based on features extracted from a training image. To combat local minima in viewpoint prediction, we introduce an efficient scheme that explores multiple hypotheses at essentially no additional cost. The entire pipeline is trained with reconstruction losses end-to-end, except for the frozen image encoder pre-trained via self-supervision." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$S = \\{\\pmb{x} \\in \\mathbb{R}^3 | s(\\pmb{x}) = 0\\}$ is defined as the isosurface of the signed distance function (SDF) $s: \\mathbb{R}^3 \\mapsto \\mathbb{R}$ , which is implemented by a Multi-Layer Perceptron (MLP) taking 3D coordinates as input. To encourage $s$ to be a signed distance function, we sample random points $\\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^3$ in space and minimise the Eikonal regulariser [15]:", + "bbox": [ + 75, + 375, + 472, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\mathrm {E i k}} (s) = \\frac {1}{| \\mathcal {X} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {X}} \\left(\\left\\| \\nabla s (\\boldsymbol {x}) \\right\\| _ {2} - 1\\right) ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 472, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then evaluate the function $s$ at the vertices of a tetrahedral grid and extract a watertight mesh from it using DMTet [56], which is fast and differentiable.", + "bbox": [ + 75, + 516, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use this implicit-explicit representation to learn a category-wise template shape shared across all instances, and model instance deformations through a hierarchical pipeline, explained next. The function $s$ is initialised with a generic ellipsoid elongated along the $z$ -axis. We denote by $V_{\\mathrm{pr}} \\subset \\mathbb{R}^3$ the vertices of the extracted prior mesh.", + "bbox": [ + 75, + 561, + 470, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Articulated Shape Modelling and Prediction", + "text_level": 1, + "bbox": [ + 76, + 659, + 452, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The category-wise template $V_{\\mathrm{pr}}$ is deformed into a specific object instance $V_{\\mathrm{ins}}$ , and further posed to match the shape $V$ in each input image (Fig. 2). These deformations are predicted from a single image, as discussed below.", + "bbox": [ + 75, + 681, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Encoder. The input image is first encoded by means of DINO-ViT [5], a self-supervised Vision Transformer (ViT) [28]. The ViT extracts from the image a set of key and output tokens, which were found to code, respectively, for correspondences [1] and appearance [61]. Hence, we use the keys to predict the object's deformation and the outputs to predict its appearance. Specifically, we apply the ViT $\\Phi$ to the input image $I\\in \\mathbb{R}^{3\\times H\\times W}$ and extract the key $\\Phi_{\\mathrm{k}}\\in \\mathbb{R}^{D\\times H_p\\times W_p}$ and output $\\Phi_{\\mathrm{o}}\\in \\mathbb{R}^{D\\times H_p\\times W_p}$ tensors from the last attention layer. These features are local,", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with one token per image patch. We use two small convolutional networks to further extract from these a global key $\\phi_{\\mathrm{k}} = f_{\\mathrm{k}}(\\Phi_{\\mathrm{k}})\\in \\mathbb{R}^{D}$ and output $\\phi_{\\mathrm{o}} = f_{\\mathrm{o}}(\\Phi_{\\mathrm{o}})\\in \\mathbb{R}^{D}$ vector.", + "bbox": [ + 498, + 375, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-specific Shape. The shape of each object instance $V_{\\mathrm{ins}}$ differs slightly from the template shape $V_{\\mathrm{pr}}$ . This is accounted for by the deformation $V_{\\mathrm{ins},i} = V_{\\mathrm{pr},i} + \\Delta V_i$ , where the displacement $\\Delta V_i = f_{\\Delta V}(V_{\\mathrm{pr},i},\\phi_{\\mathrm{k}})$ of each vertex $V_{\\mathrm{pr},i}$ is predicted by an MLP $f_{\\Delta V}$ from the global image features $\\phi_{\\mathrm{k}}$ . We exploit the fact that many objects are bilaterally symmetric and enforce both the prior $V_{\\mathrm{pr}}$ and instance $V_{\\mathrm{ins}}$ shapes to be symmetric by mirroring the query locations for the MLPs about the $yz$ -plane. We also limit the magnitude of the deformation $V_{\\mathrm{ins}}$ by adding the regularizer $\\mathcal{R}_{\\mathrm{def}}(\\Delta V) = \\frac{1}{|V_{\\mathrm{pr}}|}\\sum_i\\| \\Delta V_i\\| _2^2$ .", + "bbox": [ + 498, + 426, + 892, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Posed Shape. In order to match the 3D shape observed in the image, the instance-specific mesh $V_{\\mathrm{ins}}$ is further deformed by the posing function $V_{i} = g(V_{\\mathrm{ins},i},\\xi)$ , where $\\xi$ denotes the pose parameters. Posing accounts for the deformations caused by the articulation of the object's skeleton; they allow for large but controlled deformations which would otherwise be more difficult to account for by a generic deformation field like $\\Delta V$ .", + "bbox": [ + 496, + 603, + 890, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the blend skinning model for posing [6, 37], parameterised by $B - 1$ bone rotations $\\xi_{b} \\in SO(3), b = 2, \\ldots, B$ , and the viewpoint $\\xi_{1} \\in SE(3)$ . Specifically, we initialise a set of rest-pose joint locations $\\mathbf{J}_{b}$ on the instance mesh using simple heuristics1. Each bone $b$ but the root has one parent $\\pi(b)$ , thus forming a tree structure. Each vertex $V_{i}$ is associated to the bones by the skinning weights $w_{ib}$ , defined based on relative proximity to each bone (see", + "bbox": [ + 496, + 726, + 893, + 848 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1A chain of bones going through the two farthest end points along $z$ -axis for birds, and for quadrupeds, additionally four legs branching out from the body bone to the lowest point in each quadrant. See sup. mat.", + "bbox": [ + 498, + 862, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8795", + "bbox": [ + 480, + 944, + 514, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sup. mat.for details). The vertices are then posed by the linear blend skinning equation:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i} (\\xi) = \\left(\\sum_ {b = 1} ^ {B} w _ {i b} G _ {b} (\\xi) G _ {b} \\left(\\xi^ {*}\\right) ^ {- 1}\\right) V _ {\\text {i n s}, i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 130, + 468, + 175 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {1} = g _ {1}, G _ {b} = G _ {\\pi (b)} \\circ g _ {b}, g _ {b} (\\boldsymbol {\\xi}) = \\left[ \\begin{array}{c c} R _ {\\xi_ {b}} & \\mathbf {J} _ {b} \\\\ 0 & 1 \\end{array} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 176, + 434, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\xi^{*}$ denotes the bone rotations at the rest pose.", + "bbox": [ + 76, + 215, + 421, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Prediction. Similarly to the instance-specific deformation $\\Delta V_{i}$ , the bone rotations $\\xi_{2:B}$ are predicted by a network $f_{\\mathrm{b}}$ from the image features $\\Phi_{\\mathrm{k}}$ and $\\phi_{\\mathrm{k}}$ . The viewpoint $\\xi_{1} \\in SE(3)$ is predicted separately, as discussed in Sec. 3.4.", + "bbox": [ + 76, + 237, + 468, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose articulation depends on estimating local body joints and benefits from using local features. To extract them, we first project the centroid of each bone at rest into the image plane using the predicted viewpoint $\\xi_{1}$ :", + "bbox": [ + 76, + 297, + 468, + 358 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {u} _ {b} = \\Pi \\left[ G _ {\\pi (b)} \\left(\\xi_ {1}, \\xi_ {2: B} ^ {*}\\right) \\frac {\\mathbf {J} _ {b}}{2} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 367, + 375, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and sample the local features from the patch key feature map $\\Phi_{\\mathbf{k}}(\\pmb{u}_b)$ at the projected pixel locations $\\pmb{u}_b$", + "bbox": [ + 76, + 409, + 468, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To predict bone rotations from the image features, we adopt a transformer to account for the inter-dependency of the bones, implied by the underlying skeleton. Specifically, we use a small transformer network [62] $f_{\\mathrm{b}}$ that takes as input $B - 1$ tokens $\\nu_{b} = (\\phi_{\\mathrm{k}}, \\Phi_{\\mathrm{k}}(\\pmb{u}_{b}), b, \\mathbf{J}_{b}, \\pmb{u}_{b})$ and outputs the bone rotations $(\\xi_{2}, \\ldots, \\xi_{B}) = f_{\\mathrm{b}}(\\nu_{2}, \\ldots, \\nu_{B})$ parameterised as Euler angles. We also limit the deviation from the rest pose via the regulariser $\\mathcal{R}_{\\mathrm{art}}(\\xi) = \\frac{1}{B - 1} \\sum_{b=2}^{B} \\| \\xi_{b} \\|_{2}^{2}$ .", + "bbox": [ + 75, + 439, + 468, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Appearance and Shading", + "text_level": 1, + "bbox": [ + 76, + 568, + 307, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We decompose the appearance of the objects into albedo and diffuse shading assuming a Lambertian illumination model. The albedo of the object is modelled in the same \"canonical space\" as the shape SDF $s$ by a second coordinate MLP $a(\\pmb{x}) = f_{\\mathrm{a}}(\\pmb{x}, \\phi_{\\mathrm{o}}) \\in [0,1]^3$ , which takes as input the global output features $\\phi_{\\mathrm{o}}$ . The dominant light direction $l \\in \\mathbb{S}^2$ and ambient and diffuse intensities $k_{s}, k_{d} \\in \\mathbb{R}$ are predicted by a lighting network $f_{\\mathrm{l}}(\\phi_{\\mathrm{o}})$ .", + "bbox": [ + 75, + 592, + 468, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The posed mesh is then rendered into an image via differentiable deferred mesh rendering. The renderer first rasterizes all pixels $\\pmb{u}$ that overlap with the posed object $V$ , computing the 3D coordinates $\\pmb{x}(\\pmb{u})$ of the corresponding object point in canonical space; then the MLP $f_{\\mathrm{a}}$ is evaluated, once per pixel, to obtain the colour $a(\\pmb{x}(\\pmb{u}))$ . The final pixel colour accounts for shading and is given by", + "bbox": [ + 73, + 712, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} (\\boldsymbol {u}) = \\left(k _ {s} + k _ {d} \\cdot \\max \\{0, \\langle l, n (\\boldsymbol {u}) \\rangle \\}\\right) \\cdot a (\\boldsymbol {x} (\\boldsymbol {u})),\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 827, + 436, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n(\\pmb{u}) \\in \\mathbb{S}^2$ is the normal of the posed mesh $V$ . Using the same differentiable renderer, we also obtain the predicted mask $\\hat{M} \\in [0,1]^{H \\times W}$ of the object.", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For training supervision, we use the appearance loss $\\mathcal{L}_{\\mathrm{im}}(\\hat{I}, I, M) = \\| \\tilde{M} \\odot (\\hat{I} - I)\\|_1$ as well as the mask loss $\\mathcal{L}_{\\mathrm{m}}(\\hat{M}, M) = \\| \\hat{M} - M\\|_2^2 + \\lambda_{\\mathrm{dt}}(\\hat{M} \\odot \\mathrm{d}\\tau(M))$ , where the distance transform $\\mathrm{d}\\tau(\\cdot)$ is used to obtain better gradients [21, 68], $\\tilde{M} = \\hat{M} \\odot M$ is the intersection of the rendered and ground-truth masks, and $\\lambda_{\\mathrm{dt}}$ is a weighing factor.", + "bbox": [ + 498, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Viewpoint Prediction", + "text_level": 1, + "bbox": [ + 500, + 194, + 699, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Predicting the object viewpoint is a critical first step for learning 3D shapes but often difficult [14, 21]. Here, we contribute a method to do so robustly and efficiently, which exploits self-supervised correspondences in a multi-hypothesis prediction pipeline, as detailed next.", + "bbox": [ + 498, + 218, + 890, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fusing Self-Supervised Correspondences. Estimating explicit correspondences without supervision is difficult and brittle; hence, we propose to incorporate correspondence information implicitly and softly. Inspired by N3FF [60], we fuse information from the self-supervised ViT features into the 3D model by learning to render them. This is done by adding a coordinate network $\\psi(\\boldsymbol{x}) \\in \\mathbb{R}^{D'}$ that predicts a $D'$ -dimensional feature vector for each 3D point $\\boldsymbol{x}$ in canonical space. This is similar to the albedo network $f_{a}$ , except that it is shared across all instances of the same object category, and hence there is no dependency on the input image. These canonical features are then projected into a feature image $\\hat{\\Phi}_{\\mathrm{k}} \\in \\mathbb{R}^{H \\times W \\times D'}$ using the same differentiable renderer used for the albedo, and trained to minimise a corresponding rendering loss:", + "bbox": [ + 498, + 301, + 890, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f e a t}} \\left(\\hat {\\Phi} _ {\\mathrm {k}}, \\Phi_ {\\mathrm {k}} ^ {\\prime}, \\tilde {M}\\right) = \\| \\tilde {M} \\odot \\left(\\hat {\\Phi} _ {\\mathrm {k}} - \\Phi_ {\\mathrm {k}} ^ {\\prime}\\right) \\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 540, + 890, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_{\\mathrm{k}}^{\\prime} = \\mathrm{PCA}\\bigl (\\Phi_{\\mathrm{k}}\\bigr)$ is a PCA-reduced (to $D^{\\prime} = 16$ principal components) version of the original DINO-ViT features $\\Phi_{\\mathrm{k}}$ for training efficiency.", + "bbox": [ + 498, + 571, + 890, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Hypothesis Viewpoint Prediction. Prior works have found that a major challenge in learning the object's viewpoint is the existence of multiple local optima in the reconstruction objective. Some have addressed this issue by sampling a large number of hypotheses for the viewpoint [14], but this is somewhat cumbersome and slow. Instead, we propose a scheme that explores multiple viewpoints statistically, but at each iteration samples only a single one, and thus comes \"for free\".", + "bbox": [ + 496, + 625, + 890, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We hence task a viewpoint network $f_{\\mathrm{vp}}(\\phi_k)$ to predict from $\\phi_{\\mathrm{k}}$ four viewpoint rotation hypotheses $R_{k} \\in SO(3), k \\in \\{1,2,3,4\\}$ , each in one of the four quadrants around the object, as illustrated in Fig. 2. The network also predicts a score $\\sigma_{k}$ for each hypothesis, used to evaluate the", + "bbox": [ + 498, + 761, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2This is also similar to BANMo [72], but we use self-supervised features instead of supervised DensePose [44] ones.", + "bbox": [ + 500, + 849, + 890, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3For bilaterally symmetric animals, there are often four ambiguous orientations towards each of the four $xz$ -quadrants in the canonical space.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8796", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "probability $p_k$ that hypothesis $k$ is the best of the four options as $p_k \\propto \\exp(-\\sigma_k / \\tau)$ using softmax, where $\\tau$ is a temperature parameter, gradually decreased during training.", + "bbox": [ + 75, + 90, + 467, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The naive approach for learning $\\sigma_{k}$ is to sample multiple hypotheses and compare their reconstruction loss to determine which one is better. However, this is expensive as it requires rendering the model with all hypotheses. Instead, we suggest to sample a single hypothesis for each training iteration and simply train $\\sigma_{k}$ to predict the expected reconstruction loss $\\tilde{\\mathcal{L}}_k$ , by minimising the objective", + "bbox": [ + 75, + 137, + 467, + 242 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {h y p}} \\left(\\sigma_ {k}, \\tilde {\\mathcal {L}}\\right) = \\left(\\sigma_ {k} - \\operatorname {s g} \\left[ \\tilde {\\mathcal {L}} _ {k} \\right]\\right) ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 252, + 467, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}[\\cdot ]$ is the stop-gradient operator, and $\\tilde{\\mathcal{L}}_k$ is the overall reconstruction loss under the sampled $k$ -th hypothesis, given below.", + "bbox": [ + 75, + 282, + 467, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Training Objective", + "text_level": 1, + "bbox": [ + 76, + 339, + 258, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given a training sample $(I, M)$ , comprising an image and the corresponding object mask, the training objective is computed in two steps. First, the viewpoint hypothesis $k$ is sampled and the reconstruction loss is computed:", + "bbox": [ + 75, + 362, + 467, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {L}} _ {k} = \\lambda_ {\\mathrm {i m}} \\mathcal {L} _ {\\mathrm {i m}} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {f}} \\mathcal {L} _ {\\text {f e a t}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 434, + 380, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, the regularisation terms and the viewpoint hypothesis loss are added, to obtain the final objective:", + "bbox": [ + 75, + 464, + 467, + 494 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = p _ {k} \\tilde {\\mathcal {L}} _ {k} + \\lambda_ {\\mathrm {E}} \\mathcal {R} _ {\\text {E i k}} + \\lambda_ {\\mathrm {d}} \\mathcal {R} _ {\\text {d e f}} + \\lambda_ {\\mathrm {a}} \\mathcal {R} _ {\\text {a r t}} + \\lambda_ {\\mathrm {h}} \\mathcal {L} _ {\\text {h y p}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 506, + 439, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here $\\lambda_{\\mathrm{im}},\\lambda_{\\mathrm{m}},\\lambda_{\\mathrm{f}},\\lambda_{\\mathrm{E}},\\lambda_{\\mathrm{d}},\\lambda_{\\mathrm{a}},\\lambda_{\\mathrm{h}}$ are balancing weights.", + "bbox": [ + 75, + 535, + 429, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The gradient of $\\mathcal{L}$ with respect to the model parameters is then computed and used to update them. Note that only one viewpoint sample $k$ is taken for each gradient evaluation. To further improve the sampling efficiency, we sample the viewpoint based on the learned probability distribution $p_k^4$ .", + "bbox": [ + 75, + 551, + 467, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 640, + 207, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments on a few animal categories, including horses, giraffes, zebras, cows and birds, and compare against prior work both qualitatively and quantitatively on standard benchmarks. We also show that our model trained on real images generalises to abstract drawings, demonstrating the power of unsupervised learning.", + "bbox": [ + 75, + 665, + 467, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.Data", + "text_level": 1, + "bbox": [ + 76, + 766, + 150, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For horses, we use the horse dataset from DOVE [67] containing 10.8k images extracted from YouTube, and supplement it with 541 additional images from three datasets for diversity: Weizmann Horse Database [4], PASCAL [10] and Horse-10 Dataset [38]. For giraffes, zebras and cows,", + "bbox": [ + 75, + 789, + 467, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we source images from Microsoft COCO Dataset [33] and keep the ones with little occlusion. Since these datasets are relatively small, we only use them to finetune the pretrained horse model. For birds, we combine the DOVE dataset [67] and CUB dataset [64] consisting of $57.9\\mathrm{k}$ and $11.7\\mathrm{k}$ images respectively. We use an off-the-shelf PointRend [25] detector to obtain segmentation masks, crop around the objects and resize them to $256\\times 256$ . We follow the original train/test splits from CUB and DOVE, and randomly split the rest, resulting in $11.5\\mathrm{k} / 0.8\\mathrm{k}$ horse, $513 / 57$ giraffe, $574 / 64$ zebra, $719 / 80$ cow and $63.9\\mathrm{k} / 10.5\\mathrm{k}$ bird images. We additionally collect roughly 100 horse images from the Internet to test generalisation.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Technical Details", + "text_level": 1, + "bbox": [ + 498, + 297, + 666, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The model is implemented using a total of 8 subnetworks. The feature field $\\psi$ , template SDF $s$ , albedo field $f_{\\mathrm{a}}$ and deformation field $f_{\\Delta V}$ are implemented using MLPs, which take in 3D coordinates (concatenated with image features for albedo and deformation networks) as input. The image encoder consists of a ViT-S [28] architecture from DINO [5] and two convolutional encoders $f_{\\mathrm{k}}$ and $f_{\\mathrm{o}}$ that fuse the patch features into global feature vectors. We use self-supervised pre-trained DINO-ViT with frozen weights and only train the convolutional encoders. The lighting network $f_{1}$ is a simple MLP that maps global output feature to a 4-channel vector, and the articulation network $f_{\\mathrm{b}}$ is a transformer architecture with 4 blocks. We use a separate encoder for the viewpoint network $f_{\\mathrm{vp}}$ identical to $f_{\\mathrm{k}}$ , as we find empirically that sharing the encoders tends to make the viewpoint learning unstable. Apart from DINO encoder, all components are trained end-to-end from scratch for 100 and 10 epochs on horses and birds respectively, with the instance deformation and articulation disabled for the first 30 and 2 epochs. The DINO features $\\Phi_{\\mathrm{k}}^{\\prime}$ used to compute $\\mathcal{L}_{\\mathrm{feat}}$ are pre-computed with patch size 8 and stride 4 and masked, following [1], and reduced from a channel size of 384 to 16 using PCA. All technical details and hyper-parameters are included in the supplementary material.", + "bbox": [ + 496, + 321, + 890, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 498, + 695, + 681, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 shows a few reconstructions of horses and birds produced by our model. Given a single 2D image at test time, our model predicts a textured 3D mesh of the object, capturing its fine-grained geometric details, such as the legs and tails of the horses. Our model predicts the articulated pose of the objects, allowing us to easily transfer the pose of one instance from another and animate it in 3D. Although our model is trained on real images only, it demonstrates excellent generalisation to paintings and abstract drawings.", + "bbox": [ + 496, + 719, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 also shows that by finetuning the horse model on other animals without any additional modifications (except disabling articulation for the initial 5k iterations), it gen", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4We sample viewpoint $k^{*} = \\mathrm{argmax}_{k}p_{k}$ $80\\%$ of the time for training efficiency, and uniformly at random $20\\%$ of the time.", + "bbox": [ + 75, + 875, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8797", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ec368feec8df80a847117319297db59b7b4a027c6f3805173a33f18227606ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 143, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb8813596399eb89c40ae684a7e226a45291efa9307381f7db819f6e1653f7eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 138, + 143, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a9df4288b5199457caeed34396a0538d9dd189de62c921b7ca1eab9682ffe67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 186, + 142, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03093c552a79df8d1f58122bef9295c15765023b27f0cb81323f3c7ba0a48e2d.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 81, + 244, + 140, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/397fca09d3125ae645c61891f93dfc230a730c08307272dab17d664b813e7b3c.jpg", + "image_caption": [ + "CMR" + ], + "image_footnote": [], + "bbox": [ + 151, + 99, + 450, + 132 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b3b7b34dfe888196b20ad9f413a19362d910b9eafcf343a0597c583f1b52a0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 146, + 454, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e94fa1a600a8fbfd8492aa1bea4089827bc9d6bee4978566141d7d94fd384da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 186, + 452, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77d30e02660a2111851c3dc9d474eeed941080499348a35fdd5f87551f75430f.jpg", + "image_caption": [ + "Reconstructions", + "Textured", + "Animated" + ], + "image_footnote": [], + "bbox": [ + 148, + 243, + 452, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 328, + 244, + 383, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 388, + 244, + 452, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a28d87885ae2b7f24c449130f1e175cb7274257b5f1c44071de4cbe3b79f1f2.jpg", + "image_caption": [ + "Figure 3. Single Image Reconstruction. We show the reconstructed mesh from the input view and two additional views together with predicted texture and animated version of the shape obtained by articulating the estimated skeleton." + ], + "image_footnote": [], + "bbox": [ + 80, + 381, + 142, + 430 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3705cf23c55248ffac3b192255ab12acafa62faae4160c61eac655b75a7744d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 431, + 142, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58cc294826e8105e7de96a450dd44572b73b5df61ab28104cc1eccbd4faf3861.jpg", + "image_caption": [ + "Input", + "Figure 4. Reconstructions of Giraffes, Zebras and Cows. We finetune the horse model on other animal categories, and show that the method generalises well to other animals." + ], + "image_footnote": [], + "bbox": [ + 80, + 479, + 142, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6518ddcf9016d9924d63c7c1be86b8e0b85220fa65a9c3abe9677010687f3e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 381, + 450, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/577540a015dc4229f3d03956bb621aa23413ef4e5962f8167ab093978216b908.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 434, + 452, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85f59bf0550fb776abec6067e49f24d889f9ad0a0292c018b48f81a977c17d07.jpg", + "image_caption": [ + "Reconstructions", + "Textured", + "Animated" + ], + "image_footnote": [], + "bbox": [ + 151, + 484, + 454, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 334, + 434, + 377, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 333, + 486, + 377, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 390, + 397, + 450, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 388, + 444, + 452, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 388, + 489, + 452, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eralises to various animal categories with highly different shapes and appearances, such as giraffes, zebras and cows.", + "bbox": [ + 75, + 622, + 468, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that for the examples of abstract horse drawings (last rows in Figs. 1 and 3), since the texture is out of the training distribution, we finetune (only) the albedo network for 100 iterations, which takes less than 10 seconds. This is also done for giraffes, zebras and cows in Fig. 4, as the training datasets are too small to learn complex appearance. Additional qualitative results are presented in the sup. mat.", + "bbox": [ + 75, + 654, + 468, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with Prior Work", + "text_level": 1, + "bbox": [ + 76, + 770, + 339, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare with previous weakly-supervised methods for 3D reconstruction of deformable objects. The most relevant prior work is UMR [32] and DOVE [67], as they also only require object masks and weak correspondences from either part segmentations (SCOPS) [18] or video training [67]. Our method leverages self-supervised DINO-ViT features, which makes our method the least supervised in", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a75296467bb8ba6fc586a4ad700fda96e4bd8febcfba23cb02cde3872af5ea97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 88, + 632, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb613c7ff970271f282a66129e503565cdeed4bc4683c7b9ce136841758aa610.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 563, + 143, + 612, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8cf81f53bd65d742a41905ff378a4176df9a7c31e1935b094f8066e78aa39d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 181, + 612, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0817229915715f0d08151937915ca92bd9068eada5d5642354f1005f6cc85406.jpg", + "image_caption": [ + "Input View", + "Other Views" + ], + "image_footnote": [], + "bbox": [ + 581, + 222, + 691, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3cc63efc3281a9969e58de31326818a3f37b20f1de4662fd25d39066436eded.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 88, + 816, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18e01ec428f8e73751db0a1c56f201752927291ebacfa66975f0b1c382a287b8.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 709, + 145, + 764, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe618f9c9c11dcaca113678b31d9738a303af75c17086ae3cfb366d072cf5a55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 147, + 803, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64ad89b35b46684ca3c27803a7ad8a794215f61970aa60253160d6fb642bea23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 147, + 877, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38b343d7fb4584ea6c88396c70f67baaa50e4df650b80ed72b6d6ac1975f105f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 186, + 763, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c416e2da3266b2a4c165b21443d3f56d8244501054289a10c1e77bed3f830121.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 186, + 800, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df240beec029a314bcfc99ac10e39dd87abbbe42d21284a1e9a9e63943050a84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 186, + 877, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac1956487352e3f6164cdb79447993a05ed8894a243dcba6d5626fc2ac5e850e.jpg", + "image_caption": [ + "Input View" + ], + "image_footnote": [], + "bbox": [ + 710, + 224, + 763, + 260 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/933ec2e7fa1c9e6a5ddf8a515b42b305f47f0cd1fee39086c816a61c0b649e8a.jpg", + "image_caption": [ + "Other Views" + ], + "image_footnote": [], + "bbox": [ + 771, + 227, + 816, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fa23d4fbbb5ed927fadb93eeea9544dfed710df3f8ae0100897834eee11afa9.jpg", + "image_caption": [ + "Figure 5. Comparison on Birds. Both CMR [21] and UMR [32] often predict inaccurate poses, such as the bird on the left." + ], + "image_footnote": [], + "bbox": [ + 501, + 332, + 560, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14f701b684614336a9643db8048950891f1da6a8c4732fca4780577f53bf04eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 333, + 617, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d60f8866c77fb9c0a963b151e5cf3ebde94b2828f8ea43c711781b3394182bfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 333, + 681, + 377 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0509fc8b9f89ce27d4ecabc82245b6f54d850da2ec6a25bc3b6c748fc4e4d6e.jpg", + "image_caption": [ + "UMR", + "Figure 6. Comparison with UMR [32] on Horses. Our model produces much higher quality reconstructions with four legs. UMR is only able to predict symmetrical shapes with two legs." + ], + "image_footnote": [], + "bbox": [ + 625, + 380, + 710, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/821da40179a107739367c1522eef6ab489da739ebce3084792fc4b3ba67a1ae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 335, + 767, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19774c162073b761f8a087cd43ceee893647f49468735935e64e1a076d3ac8a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 378, + 761, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0537220f7f27445c7a778391f8fc8af328ea7dc33c40a55fa979b9ad5f5b7618.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 335, + 833, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7054a5af1b7638e894b072e987bf4117fd4101244a3e5d224c2aeac503d6131e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 776, + 377, + 877, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this area. We also compare against the well-established baseline of CMR [21], which requires stronger supervision in the form of 2D keypoint annotations and pre-computed cameras, as well as U-CMR [14], which uses a category-specific template shape. To evaluate other methods, we use their code and pre-trained models in public repositories (CMR, UMR, U-CMR) or provided by the authors (DOVE).", + "bbox": [ + 496, + 521, + 890, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparisons. Figs. 5 and 6 compare our reconstructions against the previous methods on bird and horse images. Since UMR did not release their horse model, we test our model on the examples presented in their paper for side-by-side comparison. Our method predicts more accurate poses and higher-quality articulated shapes (as opposed to only symmetric shapes with CMR and UMR). On horses, our method reconstructs all four legs with details while UMR predicts only two. Since UMR copies texture from the input image, for a fair comparison, we finetune (only) the albedo network for 100 iterations at test time.", + "bbox": [ + 496, + 635, + 890, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We evaluate shape reconstruction on the 3D Toy Bird Dataset from DOVE [67], which contains 3D scans of 23 realistic toy bird models paired with 345 photographs of them taken in natural real-world environments. Following [67], we align the predicted mesh with the ground-truth scan using ICP and compute the", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8798", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d6fbbc0f49ae95fe6fc998c4b6be4496da992f25c58367eb73bf8ac2ade7f641.jpg", + "table_caption": [ + "Table 2. Evaluation on Toy Bird Scans [67]. Baseline results reported by [67]. Annotations: template shape, viewpoint, 2D keypoint, $\\mathfrak{S}\\ll$ mask, $\\nVdash$ optical flow, video." + ], + "table_footnote": [], + "table_body": "
SupervisionChamfer Distance (cm) ↓
CMR [21]® ®1.35 ±0.81
U-CMR [14]® ®1.82 ±0.93
UMR [32]® + SCOPS1.24 ±0.75
DOVE [67]® ➔ ➔ ➔1.51 ±0.89
Ours® + DINO0.79 ±0.50
", + "bbox": [ + 107, + 142, + 439, + 238 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b3251b2508737829cbb5264772e0df1bcd04e615b949629c463688ff25fce147.jpg", + "table_caption": [ + "Table 3. Keypoint Transfer on PASCAL [10] Horse and CUB [64] Bird Datasets. Our model produces superior results on both categories with significantly less supervision. $\\dagger$ : numbers taken from [32]. 'Birds w/o aq.': excluding 50 aquatic bird classes (e.g. mallards) with heavy occlusion and large shape variation." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionPCK@0.1 ↑
HorsesBirdsBirds w/o aq.
CMR [21]†♂♀♂-0.473-
CMR [21]♂♀♂-0.5460.591
CMR [21]♂♀-0.2550.277
U-CMR [14]♀♀♀-0.3590.412
UMR [32]♂♀+ SCOPS0.2840.5120.555
A-CSM [29]♀♀♀0.3290.426-
DOVE [67]♂♀♂-0.4470.510
Ours♂♀+ DINO0.4290.5550.635
", + "bbox": [ + 78, + 330, + 468, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bi-directional Chamfer Distance between two sets of sample points from the aligned meshes. Tab. 2 summarises the results compared to other methods. Our model produces significantly more accurate reconstructions, resulting in a much lower error.", + "bbox": [ + 75, + 505, + 468, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also evaluate Keypoint Transfer on the CUB [64] and PASCAL VOC [10] benchmarks, a common evaluation metric [21, 29, 32]. We follow the protocols in [32] and [29] for birds and horses, respectively, and sample 20k source and target image pairs. Given a source image, we project all the visible vertices of the predicted mesh onto the image using the predicted viewpoint and assign each annotated 2D keypoint to its nearest vertex. We then render these vertices using the mesh and viewpoint predicted from a different target image. We measure the error between the projected vertices (corresponding to the transferred keypoints) and the annotated target keypoints using the Percentage of Correct Keypoints (PCK) metric. As shown in Tab. 3, our model significantly outperforms previous methods relying on stronger geometric supervision, including A-CSM [29] which requires a 3D shape template to begin with.", + "bbox": [ + 75, + 580, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies and Limitations", + "text_level": 1, + "bbox": [ + 76, + 832, + 367, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present ablation studies on the two key components of the model in Fig. 7, self-supervised feature loss and multi-hypothesis viewpoint prediction. Without either", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/920117262897d42f6c06a502a0ec2664a468d46dc4643c1269fa6f239cb736f1.jpg", + "image_caption": [ + "Figure 7. Ablation Studies. We train the model without feature loss $\\mathcal{L}_{\\mathrm{feat}}$ , without multi-hypothesis (MH) viewpoint prediction and without both. Without either component, the viewpoint predictions collapse to frontal views, resulting in unnatural shapes." + ], + "image_footnote": [], + "bbox": [ + 500, + 89, + 885, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "component, the learned viewpoints collapse to only frontal views, resulting in unnaturally stretched reconstructions. Additional results are provided in the sup. mat.", + "bbox": [ + 496, + 272, + 890, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our model reconstructs highly detailed 3D shapes from only a single image and generalises to in-the-wild images, the predicted texture may not preserve sufficient details with a single forward pass and may require additional test-time finetuning. This can potentially be addressed by training on a larger dataset and incorporating recent advances in image generation [53,54]. Another limitation is the requirement of a pre-defined topology for articulation, which may vary between different animal species. Discovering the articulation structure automatically from raw in-the-wild images will be of great interest for future work. Failure cases are discussed in the sup. mat.", + "bbox": [ + 496, + 318, + 892, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 513, + 625, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced a new model that can learn a 3D model of an articulated object category from single-view images taken in the wild. This model can, at test time, reconstruct the shape, articulation, albedo, and lighting of the object from a single image, and generalises to abstract drawings. Our approach demonstrates the power of combining several recent improvements in self-supervised representation learning together with a new viewpoint sampling scheme. We have shown significantly superior results compared to prior works, even when they use more supervision.", + "bbox": [ + 496, + 539, + 892, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data Ethics. We use the DOVE, Weizmann Horse, PASCAL, Horse-10, and MS-COCO datasets in a manner compatible with their terms. Some of these images may accidentally contain faces or other personal information, but we do not make use of these images or image regions. For further details on ethics, data protection, and copyright please see https://www.robots.ox.ac.uk/~vedaldi/research/union/ethics.html.", + "bbox": [ + 496, + 696, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. Shangzhe Wu is supported by a research gift from Meta Research. Andrea Vedaldi and Christian Rupprecht are supported by ERC-CoG UNION 101001212. Tomas Jakab and Christian Rupprecht are (also) supported by VisualAI EP/T028572/1.", + "bbox": [ + 496, + 825, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8799", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. In ECCV Workshops, 2022. 4, 6", + "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape completion and animation of people. ACM TOG, 2005. 1", + "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2, 3", + "[4] Eran Borenstein and Shimon Ullman. Class-specific, top-down segmentation. In ECCV, 2002. 6", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3, 4, 6", + "[6] John E Chadwick, David R Haumann, and Richard E Parent. Layered construction for deformable animated characters. ACM SIGGRAPH Computer Graphics, 1989. 4", + "[7] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2", + "[8] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, 2021. 2, 3", + "[9] Akio Doi and Akio Koide. An efficient method of triangulating equi-valued surfaces by using tetrahedral cells. IEICE Transactions on Information and Systems, 1991. 3", + "[10] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. *IJCV*, 2015. 6, 8", + "[11] Olivier Faugeras. Three-Dimensional Computer Vision: A Geometric Viewpoint. MIT Press, 1993. 3", + "[12] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In CVPR, 2021. 3", + "[13] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In CVPR, 2022. 2, 3", + "[14] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoints without keypoints. In ECCV, 2020. 1, 2, 3, 5, 7, 8", + "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 4", + "[16] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 3", + "[17] Paul Henderson and Vittorio Ferrari. Learning single-image 3d reconstruction by generative modelling of shape, pose and shading. IJCV, 2019. 2" + ], + "bbox": [ + 78, + 116, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. SCOPS: Self-supervised co-part segmentation. In CVPR, 2019. 2, 7", + "[19] Wonbong Jang and Lourdes Agapito. CodeNeRF: Disentangled neural radiance fields for object categories. In ICCV, 2021. 3", + "[20] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1", + "[21] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 1, 2, 3, 5, 7, 8", + "[22] Abhishek Kar, Shubham Tulsiani, João Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In CVPR, 2015. 1, 2", + "[23] Hiroharu Kato and Tatsuya Harada. Learning view priors for single-view 3d reconstruction. In CVPR, 2019. 2", + "[24] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2", + "[25] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. PointRend: Image segmentation as rendering. In CVPR, 2020. 3, 6", + "[26] Filippos Kokkinos and Iasonas Kokkinos. Learning monocular 3d reconstruction of articulated categories from motion. In CVPR, 2021. 2", + "[27] Filippos Kokkinos and Iasonas Kokkinos. To the point: Correspondence-driven monocular 3d category reconstruction. In NeurIPS, 2021. 2", + "[28] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 4, 6", + "[29] Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2, 8", + "[30] Nilesh Kulkarni, Abhinav Gupta, and Shubham Tulsiani. Canonical surface mapping via geometric cycle consistency. In ICCV, 2019. 2", + "[31] Xueting Li, Sifei Liu, Shalini De Mello, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Online adaptation for consistent mesh reconstruction in the wild. In NeurIPS, 2020. 1, 2, 3", + "[32] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In ECCV, 2020. 2, 7, 8", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6", + "[34] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In ICCV, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8800", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Yebin Liu, Juergen Gall, Carsten Stoll, Qionghai Dai, Hans-Peter Seidel, and Christian Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE TPAMI, 2013. 1", + "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 2015. 1, 3", + "[37] Nadia Magnenat-Thalmann, E Primeau, and Daniel Thalmann. Abstract muscle action procedures for human face animation. The Visual Computer, 1988. 4", + "[38] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W. Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In WACV, 2021. 6", + "[39] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 3", + "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3", + "[41] Tom Monnier, Matthew Fisher, Alexei A. Efros, and Mathieu Aubry. Share With Thy Neighbors: Single-View Reconstruction by Cross-Instance Consistency. In ECCV, 2022. 2", + "[42] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2021. 2, 3", + "[43] Kieran A Murphy, Carlos Esteves, Varun Jampani, Srikumar Ramalingam, and Ameesh Makadia. Implicit-PDF: Nonparametric representation of probability distributions on the rotation manifold. In ICML, 2021. 3", + "[44] Natalia Neverova, David Novotny, and Andrea Vedaldi. Continuous surface embeddings. In NeurIPS, 2020. 3, 5", + "[45] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3d representations from natural images. In ICCV, 2019. 2", + "[46] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV, 2021. 3", + "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 3", + "[48] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3", + "[49] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2", + "[50] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In CVPR, 2020. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. PVA: Pixel-aligned volumetric avatars. In CVPR, 2021. 3", + "[52] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 2, 3", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 8", + "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 8", + "[55] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2", + "[56] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In NeurIPS, 2021. 2, 3, 4", + "[57] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, 2007. 3", + "[58] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 3", + "[59] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 3", + "[60] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representation. In 3DV, 2022. 5", + "[61] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In CVPR, 2022. 4", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5", + "[63] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2, 3", + "[64] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6, 8", + "[65] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2", + "[66] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8801", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. DOVE: Learning deformable 3d objects by watching videos. arXiv preprint arXiv:2107.10844, 2021. 2, 3, 6, 7, 8", + "[68] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In CVPR, 2021. 2, 5", + "[69] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3D objects from images in the wild. In CVPR, 2020. 2", + "[70] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T. Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In CVPR, 2021. 2, 3", + "[71] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. ViSER: Video-specific surface embeddings for articulated 3d shape reconstruction. In NeurIPS, 2021. 3", + "[72] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. BANMo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 3, 5", + "[73] Chun-Han Yao, Wei-Chih Hung, Michael Rubinstein, Yuanzhen Lee, Varun Jampani, and Ming-Hsuan Yang. Lassie: Learning articulated shape from sparse image ensemble via 3d part discovery. In NeurlPS, 2022. 3", + "[74] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In CVPR, 2020. 3", + "[75] Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. Rel-Pose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 3", + "[76] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 3" + ], + "bbox": [ + 78, + 90, + 468, + 601 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8802", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_model.json b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5a9d57ae1406046ee9606051765b9be9792f6004 --- /dev/null +++ b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_model.json @@ -0,0 +1,3060 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.78, + 0.154 + ], + "angle": 0, + "content": "MagicPony: Learning Articulated 3D Animals in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.181, + 0.832, + 0.223 + ], + "angle": 0, + "content": "Shangzhe Wu* Ruining Li* Tomas Jakab* Christian Rupprecht Andrea Vedaldi Visual Geometry Group, University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.224, + 0.725, + 0.257 + ], + "angle": 0, + "content": "{szwu, ruining, tomj, chrisr,VEDaldi}@robots.ox.ac.uk 3dmagicpony.github.io" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.264, + 0.871, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.416, + 0.893, + 0.446 + ], + "angle": 0, + "content": "Figure 1. Learning Articulated 3D Animals in the Wild. Our method trains on a collection of single-view images of an animal category and produces a model that can predict the articulated 3D shape of a new instance from a single test image, which can be animated and relit." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.459, + 0.314, + 0.475 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.474, + 0.78 + ], + "angle": 0, + "content": "We consider the problem of predicting the 3D shape, articulation, viewpoint, texture, and lighting of an articulated animal like a horse given a single test image as input. We present a new method, dubbed MagicPony, that learns this predictor purely from in-the-wild single-view images of the object category, with minimal assumptions about the topology of deformation. At its core is an implicit-explicit representation of articulated shape and appearance, combining the strengths of neural fields and meshes. In order to help the model understand an object's shape and pose, we distil the knowledge captured by an off-the-shelf self-supervised vision transformer and fuse it into the 3D model. To overcome local optima in viewpoint estimation, we further introduce a new viewpoint sampling scheme that comes at no additional training cost. MagicPony outperforms prior work on this challenging task and demonstrates excellent generalisation in reconstructing art, despite the fact that it is only trained on real images. The code can be found on the project page at https://3dmagicpony.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.807, + 0.21, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.833, + 0.471, + 0.88 + ], + "angle": 0, + "content": "Reconstructing the 3D shape of an object from a single image of it requires knowing a priori what are the possible shapes and appearances of the object. Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.895, + 0.628 + ], + "angle": 0, + "content": "such a prior usually requires ad-hoc data acquisition setups [2, 20, 35, 36], involving at least multiple cameras, and often laser scanners, domes and other hardware, not to mention significant manual effort. This is viable for certain types of objects such as humans that are of particular interest in applications, but it is unlikely to scale to the long tail of objects that can appear in natural images. The alternative is to learn a 3D prior from 2D images only, which are available in abundance. However, this prior is highly complex and must be learned while using it to reconstruct in 3D the 2D training data, which is a major challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.896, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose MagicPony, a novel approach to learning 3D models of articulated object categories such as horses and birds with only single-view input images for training. We leverage recent progress in unsupervised representation learning, unsupervised image matching, efficient implicit-explicit shape representations and neural rendering, and devise a new auto-encoder architecture that reconstructs the 3D shape, articulation and texture of each object instance from a single image. For training, we only require a 2D segmenter for the object category and a description of the topology and symmetry of its 3D skeleton (i.e., the number and connectivity of bones). We do not require apriori knowledge of the objects' 3D shapes, keypoints, viewpoints, or of any other 2D or 3D cue which are often used in prior work [14, 21, 22, 31]. From this, we learn a function that, at test time, can estimate the shape and texture of a new object from a single image, in a feedforward manner. The function exhibits remarkable gener" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.207, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "8792" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.122 + ], + "angle": 0, + "content": "alisation properties, including reconstructing objects in abstract drawings, despite being trained on real images only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.468, + 0.35 + ], + "angle": 0, + "content": "In order to learn such disentangled 3D representations simply from raw images, MagicPony addresses a few key challenges. The first challenge is viewpoint estimation. Before the 3D model is available, it is very difficult to assign a viewpoint to the objects in the training images. To tackle this problem, prior works have often assumed additional cues such as 2D keypoint correspondences [21, 22, 31]. Instead, we avoid requiring additional keypoint supervision and implicitly infer noisy correspondences by fusing into the 3D model knowledge distilled from DINO-ViT [5], a self-supervised visual transformer network (ViT) [28]. We also develop a new efficient disambiguation scheme that explores multiple viewpoint assignment hypotheses at essentially no cost, avoiding local optima that are caused by greedily matching the noisy 2D correspondences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.468, + 0.655 + ], + "angle": 0, + "content": "The second challenge is how to represent the 3D shape, appearance and deformations of the object. Most prior works have used textured meshes [13, 14, 21, 31, 32, 67, 70], but these are difficult to optimise from scratch, leading to problems that often require ad-hoc heuristics such as remeshing [13, 70]. The other and increasingly popular approach is to use a volumetric representation such as a neural radiance field [3, 40, 52, 63], which can model complex shapes, including manipulating their topology during training. However, this modelling freedom comes at the cost of over-parametrisation, which is particularly problematic in monocular reconstruction and often leads to meaningless short-cut solutions [63]. Furthermore, modelling articulation with a volumetric representation is difficult. A posing transformation is only defined for the object's surface and interior and is more easily expressed from the canonical/pose-free space to the posed space. However, rendering a radiance field requires transforming 3D points off the object surface and in the direction opposite to the posing transformation, which is hard [8]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.468, + 0.763 + ], + "angle": 0, + "content": "We address these issues by using a hybrid volumetric-mesh representation based on DMTet [42, 56]. Shape and appearance are defined volumetrically in canonical space, but a mesh is extracted on the fly for posing and rendering. This sidesteps the challenges of using neural rendering directly while retaining most of its advantages and enabling the use of powerful shape regularisers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To summarise, we make the following contributions: (1) A new 3D object learning framework that combines recent advances in unsupervised learning, 3D representations and neural rendering, achieving better reconstruction results with less supervision; (2) An effective mechanism for fusing self-supervised features from DINO-ViT into the 3D model as a form of self-supervision; and (3) an efficient multi-hypothesis viewpoint prediction scheme that avoids local optima in reconstruction with no additional cost." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.174 + ], + "angle": 0, + "content": "Table 1. Related Work Overview on Weakly-supervised Learning of 3D Objects. Annotations: template shape, viewpoint, 2D keypoint, object mask, optical flow, video, coarse template shape from keypoints, camera estimated from keypoints using SfM, outputs texture flow, shape bases initialised from CMR. UMR relies on part segmentations from SCOPS [18]." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.184, + 0.885, + 0.349 + ], + "angle": 0, + "content": "
MethodSupervisionOutput
3D2.5DMotion ViewTexture
Unsup3D [69]✓✓✓
CSM [30]✓✓✓
A-CSM [29]✓✓✓✓
CMR [21](✓)1(✓)2✓✓✓✓(✓)3
U-CMR [14]✓✓✓✓✓
UMR†[32]✓✓✓(✓)3
ACMR [31](✓)4(✓)2✓✓✓✓(✓)3
DOVE [67]✓✓✓✓✓✓
Ours✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.374, + 0.892, + 0.45 + ], + "angle": 0, + "content": "We compare our method to prior work on several challenging articulated animal categories and show that our method obtains significantly better quantitative and qualitative results while using significantly less supervision, and also demonstrate generalisation to abstract drawings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.642, + 0.494 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Weakly-supervised Learning of 3D Objects. Most related to our work are weakly-supervised methods for single-image 3D object reconstruction that learn from a collection of images or videos of an object category [14, 17, 21, 23, 24, 26, 27, 31, 32, 34, 65, 67, 68]. Due to the inherent ambiguity of the problem, these methods usually rely on heavy geometric supervision (see Tab. 1) in addition to object masks, such as 2D keypoints [21, 31], template shapes [14, 29, 30]. UMR [32] forgoes the requirement of keypoints and template shapes by leveraging weakly-supervised object part segmentations (SCOPS [18]), but produces coarse symmetric shapes similar to CMR [21]. DOVE [67] learns coarse articulated objects by exploiting temporal information from videos with optical flow supervision. All these methods require object masks for supervision, except Unsup3D [69] which exploits bilateral symmetry for reconstructing roughly frontal objects, like faces, and UNICORN [41] which uses a progressive conditioning strategy with a heavily constrained bottleneck, leading to coarse reconstructions. Another emerging paradigm is to leverage generative models [7, 45, 49, 55] that encourage images rendered from viewpoints sampled from a prior distribution to be realistic. These models, however, often rely heavily on an accurate estimation of viewpoint distribution and/or a powerful 2D image generative model, both of which are difficult to obtain in small-data scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8793" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "Optimization from Multi-views and Videos. 3D reconstruction traditionally relies on epipolar geometry of multiview images of static scenes [11, 16]. Neural Radiance Fields (NeRF) [3, 40, 52, 63] have recently emerged as a powerful volumetric representation for multi-view reconstruction given accurate cameras. A recent line of work, LASR [70], ViSER [71] and BANMo [72], optimises 3D shapes of deformable objects from a small set of monocular videos, with heavily engineered optimisation strategies using optical flow and mask supervision, and additionally DensePose [44] annotations for highly deformable animals. Concurrent work of LASSIE [73] also leverages DINO-ViT [5] image features for supervision but optimises a part-based model on a small collection of images (\\(\\sim 30\\)) of an object category. It uses a shared shape model for all instances with only per-instance articulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.473, + 0.539 + ], + "angle": 0, + "content": "Learnable Deformable 3D Representations. A common 3D representation for weakly-supervised shape learning is triangular meshes. Deformation on meshes can be modelled by estimating offsets of individual vertices [21, 31, 67], which typically requires heavy regularisation, such as As-Rigid-As-Possible (ARAP) [57]. To constrain the space of deformations, many works often utilise lower-dimensional models, such as cages [74] or linear blend skinning controlled with skeletal bones [36, 67] or Gaussian control points [70, 71]. Parametric models, like SMPL [36] and SMAL [76], allow for realistic control of the shape through learned parameters but often require an extensive collection of 3D scans to learn from." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.473, + 0.741 + ], + "angle": 0, + "content": "Dynamic neural fields [12,46,48,50,51,58,59,72], such as D-NeRF [50], extend NeRF [40] with time-varying components. A-NeRF [58] proposes a radiance field conditioned on an articulated skeleton, while BANMo [72] uses learned Gaussian control points similarly to [70, 71]. One key limitation of these implicit representations is the requirement of an inverse transformation from 3D world coordinates back to the canonical space, which is often harder to learn compared to forward deformation [8]. We propose a hybrid SDF-mesh representation, extending DMTet [56] with an articulation mechanism, which combines the expressiveness of implicit models with the simplicity of mesh deformation and articulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Viewpoint Prediction in the Wild. Viewpoint prediction in weakly-supervised settings is challenging, as it is prone to local optima induced by projection and common object symmetries. Existing solutions are based on learning a distribution over multiple possible viewpoints. U-CMR [14] proposes a camera-multiplex that optimises over 40 cameras for each training sample using a given template shape, which is expensive as evaluating each hypothesis involves a rendering step. We propose a new viewpoint exploration scheme that requires evaluating only one hypothesis in each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "step with essentially no added training cost and does not rely on a given template shape but a jointly learned model. Recent work of Implicit-PDF [43] proposes to learn viewpoint distributions implicitly by estimating a probability for each image-pose pair of a fixed object instance. RelPose [75] extends this to learning relative poses of image pairs. Our model also predicts a probability associated with each viewpoint hypothesis but instead relates an instance to a learned category-level prior model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.24, + 0.593, + 0.257 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.895, + 0.342 + ], + "angle": 0, + "content": "Given only a collection of single-view images of a deformable object category collected in the wild, our goal is to train a monocular reconstruction network that, at test time, predicts the 3D shape, articulated pose, albedo and lighting from a single image of the object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.895, + 0.508 + ], + "angle": 0, + "content": "During training, our method requires no geometric supervision other than the images, 2D foreground segmentations of the object obtained automatically from an off-the-shelf method like [25], and a description of the topology of the object's skeleton (e.g., the number of legs). It is based on several key ideas: an implicit-explicit shape representation (Sec. 3.1), a hierarchical shape representation, from a generic template to articulated instances (Sec. 3.2), and a viewpoint prediction scheme that leverages self-supervised correspondences and efficient multi-hypothesis exploration to avoid local minima (Sec. 3.4), as illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.864, + 0.532 + ], + "angle": 0, + "content": "3.1. Implicit-Explicit 3D Shape Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.614 + ], + "angle": 0, + "content": "The choice of representation for the 3D shape of the object is critical. It must be (1) sufficiently expressive for modelling fine-grained shape details and deformations and (2) sufficiently regular for learning with only weak supervision via image reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.895, + 0.827 + ], + "angle": 0, + "content": "Most 3D representations are either explicit (triangular meshes [21, 67, 70]) or implicit (volumes, SDFs, radiance fields [19, 39, 40, 47, 66]). Explicit representations are more compact, easily support powerful deformation models like blend skinning, and the geometric smoothness can be easily measured and controlled; however, they are more difficult to optimise and often result in defects like folds during training, which requires occasional re-meshing [13, 70]. Implicit representations are more amenable to gradient-based optimisation, including allowing topological changes, but are more computationally expensive and, due to their high capacity, require substantial supervision, usually in the form of a dense coverage of the different object viewpoints [40, 66] or 3D ground-truth [39, 47]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Here we adopt a representation that combines the advantages of both. The shape of the object is represented implicitly by a neural field that is converted on the fly into an explicit mesh via the Differentiable Marching Tetrahedra (DMTet) [9, 42, 56] method. Specifically, the object's shape" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "8794" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Training Pipeline. Given a collection of single-view images of an animal category, our model learns a category-specific prior shape using an implicit-explicit representation, together with a feature field that allows us to fuse self-supervised correspondences through a feature rendering loss. This prior shape is deformed, articulated and shaded based on features extracted from a training image. To combat local minima in viewpoint prediction, we introduce an efficient scheme that explores multiple hypotheses at essentially no additional cost. The entire pipeline is trained with reconstruction losses end-to-end, except for the frozen image encoder pre-trained via self-supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.473, + 0.466 + ], + "angle": 0, + "content": "\\(S = \\{\\pmb{x} \\in \\mathbb{R}^3 | s(\\pmb{x}) = 0\\}\\) is defined as the isosurface of the signed distance function (SDF) \\(s: \\mathbb{R}^3 \\mapsto \\mathbb{R}\\), which is implemented by a Multi-Layer Perceptron (MLP) taking 3D coordinates as input. To encourage \\(s\\) to be a signed distance function, we sample random points \\(\\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^3\\) in space and minimise the Eikonal regulariser [15]:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.473, + 0.47, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\mathrm {E i k}} (s) = \\frac {1}{| \\mathcal {X} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {X}} \\left(\\left\\| \\nabla s (\\boldsymbol {x}) \\right\\| _ {2} - 1\\right) ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.561 + ], + "angle": 0, + "content": "We then evaluate the function \\(s\\) at the vertices of a tetrahedral grid and extract a watertight mesh from it using DMTet [56], which is fast and differentiable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.471, + 0.653 + ], + "angle": 0, + "content": "We use this implicit-explicit representation to learn a category-wise template shape shared across all instances, and model instance deformations through a hierarchical pipeline, explained next. The function \\( s \\) is initialised with a generic ellipsoid elongated along the \\( z \\)-axis. We denote by \\( V_{\\mathrm{pr}} \\subset \\mathbb{R}^3 \\) the vertices of the extracted prior mesh." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.66, + 0.454, + 0.677 + ], + "angle": 0, + "content": "3.2. Articulated Shape Modelling and Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.745 + ], + "angle": 0, + "content": "The category-wise template \\( V_{\\mathrm{pr}} \\) is deformed into a specific object instance \\( V_{\\mathrm{ins}} \\), and further posed to match the shape \\( V \\) in each input image (Fig. 2). These deformations are predicted from a single image, as discussed below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Image Encoder. The input image is first encoded by means of DINO-ViT [5], a self-supervised Vision Transformer (ViT) [28]. The ViT extracts from the image a set of key and output tokens, which were found to code, respectively, for correspondences [1] and appearance [61]. Hence, we use the keys to predict the object's deformation and the outputs to predict its appearance. Specifically, we apply the ViT \\(\\Phi\\) to the input image \\(I\\in \\mathbb{R}^{3\\times H\\times W}\\) and extract the key \\(\\Phi_{\\mathrm{k}}\\in \\mathbb{R}^{D\\times H_p\\times W_p}\\) and output \\(\\Phi_{\\mathrm{o}}\\in \\mathbb{R}^{D\\times H_p\\times W_p}\\) tensors from the last attention layer. These features are local," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.375, + 0.892, + 0.423 + ], + "angle": 0, + "content": "with one token per image patch. We use two small convolutional networks to further extract from these a global key \\(\\phi_{\\mathrm{k}} = f_{\\mathrm{k}}(\\Phi_{\\mathrm{k}})\\in \\mathbb{R}^{D}\\) and output \\(\\phi_{\\mathrm{o}} = f_{\\mathrm{o}}(\\Phi_{\\mathrm{o}})\\in \\mathbb{R}^{D}\\) vector." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.428, + 0.893, + 0.6 + ], + "angle": 0, + "content": "Instance-specific Shape. The shape of each object instance \\( V_{\\mathrm{ins}} \\) differs slightly from the template shape \\( V_{\\mathrm{pr}} \\). This is accounted for by the deformation \\( V_{\\mathrm{ins},i} = V_{\\mathrm{pr},i} + \\Delta V_i \\), where the displacement \\( \\Delta V_i = f_{\\Delta V}(V_{\\mathrm{pr},i},\\phi_{\\mathrm{k}}) \\) of each vertex \\( V_{\\mathrm{pr},i} \\) is predicted by an MLP \\( f_{\\Delta V} \\) from the global image features \\( \\phi_{\\mathrm{k}} \\). We exploit the fact that many objects are bilaterally symmetric and enforce both the prior \\( V_{\\mathrm{pr}} \\) and instance \\( V_{\\mathrm{ins}} \\) shapes to be symmetric by mirroring the query locations for the MLPs about the \\( yz \\)-plane. We also limit the magnitude of the deformation \\( V_{\\mathrm{ins}} \\) by adding the regularizer \\( \\mathcal{R}_{\\mathrm{def}}(\\Delta V) = \\frac{1}{|V_{\\mathrm{pr}}|}\\sum_i\\| \\Delta V_i\\| _2^2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Posed Shape. In order to match the 3D shape observed in the image, the instance-specific mesh \\( V_{\\mathrm{ins}} \\) is further deformed by the posing function \\( V_{i} = g(V_{\\mathrm{ins},i},\\xi) \\), where \\( \\xi \\) denotes the pose parameters. Posing accounts for the deformations caused by the articulation of the object's skeleton; they allow for large but controlled deformations which would otherwise be more difficult to account for by a generic deformation field like \\( \\Delta V \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.895, + 0.849 + ], + "angle": 0, + "content": "We adopt the blend skinning model for posing [6, 37], parameterised by \\(B - 1\\) bone rotations \\(\\xi_{b} \\in SO(3), b = 2, \\ldots, B\\), and the viewpoint \\(\\xi_{1} \\in SE(3)\\). Specifically, we initialise a set of rest-pose joint locations \\(\\mathbf{J}_{b}\\) on the instance mesh using simple heuristics1. Each bone \\(b\\) but the root has one parent \\(\\pi(b)\\), thus forming a tree structure. Each vertex \\(V_{i}\\) is associated to the bones by the skinning weights \\(w_{ib}\\), defined based on relative proximity to each bone (see" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.895, + 0.901 + ], + "angle": 0, + "content": "1A chain of bones going through the two farthest end points along \\( z \\)-axis for birds, and for quadrupeds, additionally four legs branching out from the body bone to the lowest point in each quadrant. See sup. mat." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "8795" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "sup. mat.for details). The vertices are then posed by the linear blend skinning equation:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.131, + 0.469, + 0.176 + ], + "angle": 0, + "content": "\\[\nV _ {i} (\\xi) = \\left(\\sum_ {b = 1} ^ {B} w _ {i b} G _ {b} (\\xi) G _ {b} \\left(\\xi^ {*}\\right) ^ {- 1}\\right) V _ {\\text {i n s}, i}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.177, + 0.435, + 0.21 + ], + "angle": 0, + "content": "\\[\nG _ {1} = g _ {1}, G _ {b} = G _ {\\pi (b)} \\circ g _ {b}, g _ {b} (\\boldsymbol {\\xi}) = \\left[ \\begin{array}{c c} R _ {\\xi_ {b}} & \\mathbf {J} _ {b} \\\\ 0 & 1 \\end{array} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.217, + 0.423, + 0.233 + ], + "angle": 0, + "content": "where \\(\\xi^{*}\\) denotes the bone rotations at the rest pose." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.238, + 0.469, + 0.299 + ], + "angle": 0, + "content": "Pose Prediction. Similarly to the instance-specific deformation \\(\\Delta V_{i}\\), the bone rotations \\(\\xi_{2:B}\\) are predicted by a network \\(f_{\\mathrm{b}}\\) from the image features \\(\\Phi_{\\mathrm{k}}\\) and \\(\\phi_{\\mathrm{k}}\\). The viewpoint \\(\\xi_{1} \\in SE(3)\\) is predicted separately, as discussed in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.299, + 0.469, + 0.359 + ], + "angle": 0, + "content": "Pose articulation depends on estimating local body joints and benefits from using local features. To extract them, we first project the centroid of each bone at rest into the image plane using the predicted viewpoint \\(\\xi_{1}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.368, + 0.376, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {u} _ {b} = \\Pi \\left[ G _ {\\pi (b)} \\left(\\xi_ {1}, \\xi_ {2: B} ^ {*}\\right) \\frac {\\mathbf {J} _ {b}}{2} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.41, + 0.469, + 0.44 + ], + "angle": 0, + "content": "and sample the local features from the patch key feature map \\(\\Phi_{\\mathbf{k}}(\\pmb{u}_b)\\) at the projected pixel locations \\(\\pmb{u}_b\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.564 + ], + "angle": 0, + "content": "To predict bone rotations from the image features, we adopt a transformer to account for the inter-dependency of the bones, implied by the underlying skeleton. Specifically, we use a small transformer network [62] \\( f_{\\mathrm{b}} \\) that takes as input \\( B - 1 \\) tokens \\( \\nu_{b} = (\\phi_{\\mathrm{k}}, \\Phi_{\\mathrm{k}}(\\pmb{u}_{b}), b, \\mathbf{J}_{b}, \\pmb{u}_{b}) \\) and outputs the bone rotations \\( (\\xi_{2}, \\ldots, \\xi_{B}) = f_{\\mathrm{b}}(\\nu_{2}, \\ldots, \\nu_{B}) \\) parameterised as Euler angles. We also limit the deviation from the rest pose via the regulariser \\( \\mathcal{R}_{\\mathrm{art}}(\\xi) = \\frac{1}{B - 1} \\sum_{b=2}^{B} \\| \\xi_{b} \\|_{2}^{2} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.569, + 0.308, + 0.586 + ], + "angle": 0, + "content": "3.3. Appearance and Shading" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.469, + 0.713 + ], + "angle": 0, + "content": "We decompose the appearance of the objects into albedo and diffuse shading assuming a Lambertian illumination model. The albedo of the object is modelled in the same \"canonical space\" as the shape SDF \\(s\\) by a second coordinate MLP \\(a(\\pmb{x}) = f_{\\mathrm{a}}(\\pmb{x}, \\phi_{\\mathrm{o}}) \\in [0,1]^3\\), which takes as input the global output features \\(\\phi_{\\mathrm{o}}\\). The dominant light direction \\(l \\in \\mathbb{S}^2\\) and ambient and diffuse intensities \\(k_{s}, k_{d} \\in \\mathbb{R}\\) are predicted by a lighting network \\(f_{\\mathrm{l}}(\\phi_{\\mathrm{o}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.713, + 0.469, + 0.82 + ], + "angle": 0, + "content": "The posed mesh is then rendered into an image via differentiable deferred mesh rendering. The renderer first rasterizes all pixels \\(\\pmb{u}\\) that overlap with the posed object \\(V\\), computing the 3D coordinates \\(\\pmb{x}(\\pmb{u})\\) of the corresponding object point in canonical space; then the MLP \\(f_{\\mathrm{a}}\\) is evaluated, once per pixel, to obtain the colour \\(a(\\pmb{x}(\\pmb{u}))\\). The final pixel colour accounts for shading and is given by" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.828, + 0.437, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\hat {I} (\\boldsymbol {u}) = \\left(k _ {s} + k _ {d} \\cdot \\max \\{0, \\langle l, n (\\boldsymbol {u}) \\rangle \\}\\right) \\cdot a (\\boldsymbol {x} (\\boldsymbol {u})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(n(\\pmb{u}) \\in \\mathbb{S}^2\\) is the normal of the posed mesh \\(V\\). Using the same differentiable renderer, we also obtain the predicted mask \\(\\hat{M} \\in [0,1]^{H \\times W}\\) of the object." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "For training supervision, we use the appearance loss \\(\\mathcal{L}_{\\mathrm{im}}(\\hat{I}, I, M) = \\| \\tilde{M} \\odot (\\hat{I} - I)\\|_1\\) as well as the mask loss \\(\\mathcal{L}_{\\mathrm{m}}(\\hat{M}, M) = \\| \\hat{M} - M\\|_2^2 + \\lambda_{\\mathrm{dt}}(\\hat{M} \\odot \\mathrm{d}\\tau(M))\\), where the distance transform \\(\\mathrm{d}\\tau(\\cdot)\\) is used to obtain better gradients [21, 68], \\(\\tilde{M} = \\hat{M} \\odot M\\) is the intersection of the rendered and ground-truth masks, and \\(\\lambda_{\\mathrm{dt}}\\) is a weighing factor." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.195, + 0.7, + 0.211 + ], + "angle": 0, + "content": "3.4. Viewpoint Prediction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.219, + 0.892, + 0.295 + ], + "angle": 0, + "content": "Predicting the object viewpoint is a critical first step for learning 3D shapes but often difficult [14, 21]. Here, we contribute a method to do so robustly and efficiently, which exploits self-supervised correspondences in a multi-hypothesis prediction pipeline, as detailed next." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.302, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Fusing Self-Supervised Correspondences. Estimating explicit correspondences without supervision is difficult and brittle; hence, we propose to incorporate correspondence information implicitly and softly. Inspired by N3FF [60], we fuse information from the self-supervised ViT features into the 3D model by learning to render them. This is done by adding a coordinate network \\(\\psi(\\boldsymbol{x}) \\in \\mathbb{R}^{D'}\\) that predicts a \\(D'\\)-dimensional feature vector for each 3D point \\(\\boldsymbol{x}\\) in canonical space. This is similar to the albedo network \\(f_{a}\\), except that it is shared across all instances of the same object category, and hence there is no dependency on the input image. These canonical features are then projected into a feature image \\(\\hat{\\Phi}_{\\mathrm{k}} \\in \\mathbb{R}^{H \\times W \\times D'}\\) using the same differentiable renderer used for the albedo, and trained to minimise a corresponding rendering loss:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.541, + 0.892, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f e a t}} \\left(\\hat {\\Phi} _ {\\mathrm {k}}, \\Phi_ {\\mathrm {k}} ^ {\\prime}, \\tilde {M}\\right) = \\| \\tilde {M} \\odot \\left(\\hat {\\Phi} _ {\\mathrm {k}} - \\Phi_ {\\mathrm {k}} ^ {\\prime}\\right) \\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.572, + 0.891, + 0.619 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\mathrm{k}}^{\\prime} = \\mathrm{PCA}\\bigl (\\Phi_{\\mathrm{k}}\\bigr)\\) is a PCA-reduced (to \\(D^{\\prime} = 16\\) principal components) version of the original DINO-ViT features \\(\\Phi_{\\mathrm{k}}\\) for training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Multi-Hypothesis Viewpoint Prediction. Prior works have found that a major challenge in learning the object's viewpoint is the existence of multiple local optima in the reconstruction objective. Some have addressed this issue by sampling a large number of hypotheses for the viewpoint [14], but this is somewhat cumbersome and slow. Instead, we propose a scheme that explores multiple viewpoints statistically, but at each iteration samples only a single one, and thus comes \"for free\"." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.762, + 0.892, + 0.839 + ], + "angle": 0, + "content": "We hence task a viewpoint network \\( f_{\\mathrm{vp}}(\\phi_k) \\) to predict from \\( \\phi_{\\mathrm{k}} \\) four viewpoint rotation hypotheses \\( R_{k} \\in SO(3), k \\in \\{1,2,3,4\\} \\), each in one of the four quadrants around the object, as illustrated in Fig. 2. The network also predicts a score \\( \\sigma_{k} \\) for each hypothesis, used to evaluate the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.876 + ], + "angle": 0, + "content": "2This is also similar to BANMo [72], but we use self-supervised features instead of supervised DensePose [44] ones." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3For bilaterally symmetric animals, there are often four ambiguous orientations towards each of the four \\(xz\\)-quadrants in the canonical space." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8796" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "probability \\( p_k \\) that hypothesis \\( k \\) is the best of the four options as \\( p_k \\propto \\exp(-\\sigma_k / \\tau) \\) using softmax, where \\( \\tau \\) is a temperature parameter, gradually decreased during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.243 + ], + "angle": 0, + "content": "The naive approach for learning \\(\\sigma_{k}\\) is to sample multiple hypotheses and compare their reconstruction loss to determine which one is better. However, this is expensive as it requires rendering the model with all hypotheses. Instead, we suggest to sample a single hypothesis for each training iteration and simply train \\(\\sigma_{k}\\) to predict the expected reconstruction loss \\(\\tilde{\\mathcal{L}}_k\\), by minimising the objective" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.253, + 0.468, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {h y p}} \\left(\\sigma_ {k}, \\tilde {\\mathcal {L}}\\right) = \\left(\\sigma_ {k} - \\operatorname {s g} \\left[ \\tilde {\\mathcal {L}} _ {k} \\right]\\right) ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.468, + 0.331 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}[\\cdot ]\\) is the stop-gradient operator, and \\(\\tilde{\\mathcal{L}}_k\\) is the overall reconstruction loss under the sampled \\(k\\)-th hypothesis, given below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.34, + 0.259, + 0.357 + ], + "angle": 0, + "content": "3.5. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.468, + 0.424 + ], + "angle": 0, + "content": "Given a training sample \\((I, M)\\), comprising an image and the corresponding object mask, the training objective is computed in two steps. First, the viewpoint hypothesis \\(k\\) is sampled and the reconstruction loss is computed:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.435, + 0.382, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {L}} _ {k} = \\lambda_ {\\mathrm {i m}} \\mathcal {L} _ {\\mathrm {i m}} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {f}} \\mathcal {L} _ {\\text {f e a t}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.468, + 0.495 + ], + "angle": 0, + "content": "Then, the regularisation terms and the viewpoint hypothesis loss are added, to obtain the final objective:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.507, + 0.441, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = p _ {k} \\tilde {\\mathcal {L}} _ {k} + \\lambda_ {\\mathrm {E}} \\mathcal {R} _ {\\text {E i k}} + \\lambda_ {\\mathrm {d}} \\mathcal {R} _ {\\text {d e f}} + \\lambda_ {\\mathrm {a}} \\mathcal {R} _ {\\text {a r t}} + \\lambda_ {\\mathrm {h}} \\mathcal {L} _ {\\text {h y p}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.431, + 0.551 + ], + "angle": 0, + "content": "Here \\(\\lambda_{\\mathrm{im}},\\lambda_{\\mathrm{m}},\\lambda_{\\mathrm{f}},\\lambda_{\\mathrm{E}},\\lambda_{\\mathrm{d}},\\lambda_{\\mathrm{a}},\\lambda_{\\mathrm{h}}\\) are balancing weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.468, + 0.627 + ], + "angle": 0, + "content": "The gradient of \\(\\mathcal{L}\\) with respect to the model parameters is then computed and used to update them. Note that only one viewpoint sample \\(k\\) is taken for each gradient evaluation. To further improve the sampling efficiency, we sample the viewpoint based on the learned probability distribution \\(p_k^4\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.641, + 0.208, + 0.657 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.468, + 0.757 + ], + "angle": 0, + "content": "We conduct extensive experiments on a few animal categories, including horses, giraffes, zebras, cows and birds, and compare against prior work both qualitatively and quantitatively on standard benchmarks. We also show that our model trained on real images generalises to abstract drawings, demonstrating the power of unsupervised learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.767, + 0.151, + 0.781 + ], + "angle": 0, + "content": "4.1.Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.468, + 0.866 + ], + "angle": 0, + "content": "For horses, we use the horse dataset from DOVE [67] containing 10.8k images extracted from YouTube, and supplement it with 541 additional images from three datasets for diversity: Weizmann Horse Database [4], PASCAL [10] and Horse-10 Dataset [38]. For giraffes, zebras and cows," + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "4We sample viewpoint \\(k^{*} = \\mathrm{argmax}_{k}p_{k}\\) \\(80\\%\\) of the time for training efficiency, and uniformly at random \\(20\\%\\) of the time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "we source images from Microsoft COCO Dataset [33] and keep the ones with little occlusion. Since these datasets are relatively small, we only use them to finetune the pretrained horse model. For birds, we combine the DOVE dataset [67] and CUB dataset [64] consisting of \\(57.9\\mathrm{k}\\) and \\(11.7\\mathrm{k}\\) images respectively. We use an off-the-shelf PointRend [25] detector to obtain segmentation masks, crop around the objects and resize them to \\(256\\times 256\\). We follow the original train/test splits from CUB and DOVE, and randomly split the rest, resulting in \\(11.5\\mathrm{k} / 0.8\\mathrm{k}\\) horse, \\(513 / 57\\) giraffe, \\(574 / 64\\) zebra, \\(719 / 80\\) cow and \\(63.9\\mathrm{k} / 10.5\\mathrm{k}\\) bird images. We additionally collect roughly 100 horse images from the Internet to test generalisation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.299, + 0.667, + 0.313 + ], + "angle": 0, + "content": "4.2. Technical Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.685 + ], + "angle": 0, + "content": "The model is implemented using a total of 8 subnetworks. The feature field \\(\\psi\\), template SDF \\(s\\), albedo field \\(f_{\\mathrm{a}}\\) and deformation field \\(f_{\\Delta V}\\) are implemented using MLPs, which take in 3D coordinates (concatenated with image features for albedo and deformation networks) as input. The image encoder consists of a ViT-S [28] architecture from DINO [5] and two convolutional encoders \\(f_{\\mathrm{k}}\\) and \\(f_{\\mathrm{o}}\\) that fuse the patch features into global feature vectors. We use self-supervised pre-trained DINO-ViT with frozen weights and only train the convolutional encoders. The lighting network \\(f_{1}\\) is a simple MLP that maps global output feature to a 4-channel vector, and the articulation network \\(f_{\\mathrm{b}}\\) is a transformer architecture with 4 blocks. We use a separate encoder for the viewpoint network \\(f_{\\mathrm{vp}}\\) identical to \\(f_{\\mathrm{k}}\\), as we find empirically that sharing the encoders tends to make the viewpoint learning unstable. Apart from DINO encoder, all components are trained end-to-end from scratch for 100 and 10 epochs on horses and birds respectively, with the instance deformation and articulation disabled for the first 30 and 2 epochs. The DINO features \\(\\Phi_{\\mathrm{k}}^{\\prime}\\) used to compute \\(\\mathcal{L}_{\\mathrm{feat}}\\) are pre-computed with patch size 8 and stride 4 and masked, following [1], and reduced from a channel size of 384 to 16 using PCA. All technical details and hyper-parameters are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.696, + 0.682, + 0.711 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Fig. 3 shows a few reconstructions of horses and birds produced by our model. Given a single 2D image at test time, our model predicts a textured 3D mesh of the object, capturing its fine-grained geometric details, such as the legs and tails of the horses. Our model predicts the articulated pose of the objects, allowing us to easily transfer the pose of one instance from another and animate it in 3D. Although our model is trained on real images only, it demonstrates excellent generalisation to paintings and abstract drawings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 also shows that by finetuning the horse model on other animals without any additional modifications (except disabling articulation for the initial 5k iterations), it gen" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8797" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.144, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.14, + 0.144, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.188, + 0.143, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.245, + 0.141, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.288, + 0.128, + 0.299 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.1, + 0.451, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.147, + 0.455, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.188, + 0.453, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.244, + 0.453, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.288, + 0.286, + 0.298 + ], + "angle": 0, + "content": "Reconstructions" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.245, + 0.385, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.288, + 0.379, + 0.298 + ], + "angle": 0, + "content": "Textured" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.245, + 0.453, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.288, + 0.449, + 0.298 + ], + "angle": 0, + "content": "Animated" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.31, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Figure 3. Single Image Reconstruction. We show the reconstructed mesh from the input view and two additional views together with predicted texture and animated version of the shape obtained by articulating the estimated skeleton." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.382, + 0.143, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.432, + 0.143, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.48, + 0.143, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.532, + 0.127, + 0.542 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.382, + 0.451, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.435, + 0.453, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.485, + 0.455, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.531, + 0.284, + 0.541 + ], + "angle": 0, + "content": "Reconstructions" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.435, + 0.379, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.487, + 0.379, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.531, + 0.376, + 0.541 + ], + "angle": 0, + "content": "Textured" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.398, + 0.451, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.445, + 0.453, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.491, + 0.454, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.531, + 0.446, + 0.54 + ], + "angle": 0, + "content": "Animated" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.553, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Figure 4. Reconstructions of Giraffes, Zebras and Cows. We finetune the horse model on other animal categories, and show that the method generalises well to other animals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.654 + ], + "angle": 0, + "content": "eralises to various animal categories with highly different shapes and appearances, such as giraffes, zebras and cows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.469, + 0.76 + ], + "angle": 0, + "content": "Note that for the examples of abstract horse drawings (last rows in Figs. 1 and 3), since the texture is out of the training distribution, we finetune (only) the albedo network for 100 iterations, which takes less than 10 seconds. This is also done for giraffes, zebras and cows in Fig. 4, as the training datasets are too small to learn complex appearance. Additional qualitative results are presented in the sup. mat." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.341, + 0.787 + ], + "angle": 0, + "content": "4.4. Comparison with Prior Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We compare with previous weakly-supervised methods for 3D reconstruction of deformable objects. The most relevant prior work is UMR [32] and DOVE [67], as they also only require object masks and weak correspondences from either part segmentations (SCOPS) [18] or video training [67]. Our method leverages self-supervised DINO-ViT features, which makes our method the least supervised in" + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.089, + 0.633, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.136, + 0.633, + 0.145 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.156, + 0.528, + 0.165 + ], + "angle": 0, + "content": "CMR" + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.145, + 0.613, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.183, + 0.613, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.223, + 0.692, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.264, + 0.576, + 0.273 + ], + "angle": 0, + "content": "Input View" + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.264, + 0.661, + 0.273 + ], + "angle": 0, + "content": "Other Views" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.089, + 0.818, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.136, + 0.819, + 0.145 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.146, + 0.766, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.148, + 0.805, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.148, + 0.879, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.187, + 0.764, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.187, + 0.802, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.187, + 0.878, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.226, + 0.764, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.264, + 0.765, + 0.273 + ], + "angle": 0, + "content": "Input View" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.228, + 0.817, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.264, + 0.849, + 0.273 + ], + "angle": 0, + "content": "Other Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.286, + 0.892, + 0.315 + ], + "angle": 0, + "content": "Figure 5. Comparison on Birds. Both CMR [21] and UMR [32] often predict inaccurate poses, such as the bird on the left." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.333, + 0.562, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.334, + 0.619, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.334, + 0.682, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.381, + 0.712, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.428, + 0.651, + 0.437 + ], + "angle": 0, + "content": "UMR" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.336, + 0.769, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.379, + 0.762, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.337, + 0.834, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.378, + 0.879, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.428, + 0.804, + 0.437 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.448, + 0.892, + 0.491 + ], + "angle": 0, + "content": "Figure 6. Comparison with UMR [32] on Horses. Our model produces much higher quality reconstructions with four legs. UMR is only able to predict symmetrical shapes with two legs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.628 + ], + "angle": 0, + "content": "this area. We also compare against the well-established baseline of CMR [21], which requires stronger supervision in the form of 2D keypoint annotations and pre-computed cameras, as well as U-CMR [14], which uses a category-specific template shape. To evaluate other methods, we use their code and pre-trained models in public repositories (CMR, UMR, U-CMR) or provided by the authors (DOVE)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.802 + ], + "angle": 0, + "content": "Qualitative Comparisons. Figs. 5 and 6 compare our reconstructions against the previous methods on bird and horse images. Since UMR did not release their horse model, we test our model on the examples presented in their paper for side-by-side comparison. Our method predicts more accurate poses and higher-quality articulated shapes (as opposed to only symmetric shapes with CMR and UMR). On horses, our method reconstructs all four legs with details while UMR predicts only two. Since UMR copies texture from the input image, for a fair comparison, we finetune (only) the albedo network for 100 iterations at test time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparisons. We evaluate shape reconstruction on the 3D Toy Bird Dataset from DOVE [67], which contains 3D scans of 23 realistic toy bird models paired with 345 photographs of them taken in natural real-world environments. Following [67], we align the predicted mesh with the ground-truth scan using ICP and compute the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8798" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.47, + 0.132 + ], + "angle": 0, + "content": "Table 2. Evaluation on Toy Bird Scans [67]. Baseline results reported by [67]. Annotations: template shape, viewpoint, 2D keypoint, \\(\\mathfrak{S}\\ll\\) mask, \\(\\nVdash\\) optical flow, video." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.143, + 0.44, + 0.239 + ], + "angle": 0, + "content": "
SupervisionChamfer Distance (cm) ↓
CMR [21]® ®1.35 ±0.81
U-CMR [14]® ®1.82 ±0.93
UMR [32]® + SCOPS1.24 ±0.75
DOVE [67]® ➔ ➔ ➔1.51 ±0.89
Ours® + DINO0.79 ±0.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.251, + 0.47, + 0.321 + ], + "angle": 0, + "content": "Table 3. Keypoint Transfer on PASCAL [10] Horse and CUB [64] Bird Datasets. Our model produces superior results on both categories with significantly less supervision. \\(\\dagger\\): numbers taken from [32]. 'Birds w/o aq.': excluding 50 aquatic bird classes (e.g. mallards) with heavy occlusion and large shape variation." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.332, + 0.469, + 0.481 + ], + "angle": 0, + "content": "
MethodSupervisionPCK@0.1 ↑
HorsesBirdsBirds w/o aq.
CMR [21]†♂♀♂-0.473-
CMR [21]♂♀♂-0.5460.591
CMR [21]♂♀-0.2550.277
U-CMR [14]♀♀♀-0.3590.412
UMR [32]♂♀+ SCOPS0.2840.5120.555
A-CSM [29]♀♀♀0.3290.426-
DOVE [67]♂♀♂-0.4470.510
Ours♂♀+ DINO0.4290.5550.635
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.469, + 0.58 + ], + "angle": 0, + "content": "bi-directional Chamfer Distance between two sets of sample points from the aligned meshes. Tab. 2 summarises the results compared to other methods. Our model produces significantly more accurate reconstructions, resulting in a much lower error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.824 + ], + "angle": 0, + "content": "We also evaluate Keypoint Transfer on the CUB [64] and PASCAL VOC [10] benchmarks, a common evaluation metric [21, 29, 32]. We follow the protocols in [32] and [29] for birds and horses, respectively, and sample 20k source and target image pairs. Given a source image, we project all the visible vertices of the predicted mesh onto the image using the predicted viewpoint and assign each annotated 2D keypoint to its nearest vertex. We then render these vertices using the mesh and viewpoint predicted from a different target image. We measure the error between the projected vertices (corresponding to the transferred keypoints) and the annotated target keypoints using the Percentage of Correct Keypoints (PCK) metric. As shown in Tab. 3, our model significantly outperforms previous methods relying on stronger geometric supervision, including A-CSM [29] which requires a 3D shape template to begin with." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.368, + 0.848 + ], + "angle": 0, + "content": "4.5. Ablation Studies and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We present ablation studies on the two key components of the model in Fig. 7, self-supervised feature loss and multi-hypothesis viewpoint prediction. Without either" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.09, + 0.887, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.19, + 0.893, + 0.247 + ], + "angle": 0, + "content": "Figure 7. Ablation Studies. We train the model without feature loss \\(\\mathcal{L}_{\\mathrm{feat}}\\), without multi-hypothesis (MH) viewpoint prediction and without both. Without either component, the viewpoint predictions collapse to frontal views, resulting in unnatural shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.318 + ], + "angle": 0, + "content": "component, the learned viewpoints collapse to only frontal views, resulting in unnaturally stretched reconstructions. Additional results are provided in the sup. mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.893, + 0.501 + ], + "angle": 0, + "content": "Although our model reconstructs highly detailed 3D shapes from only a single image and generalises to in-the-wild images, the predicted texture may not preserve sufficient details with a single forward pass and may require additional test-time finetuning. This can potentially be addressed by training on a larger dataset and incorporating recent advances in image generation [53,54]. Another limitation is the requirement of a pre-defined topology for articulation, which may vary between different animal species. Discovering the articulation structure automatically from raw in-the-wild images will be of great interest for future work. Failure cases are discussed in the sup. mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.627, + 0.53 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.893, + 0.692 + ], + "angle": 0, + "content": "We have introduced a new model that can learn a 3D model of an articulated object category from single-view images taken in the wild. This model can, at test time, reconstruct the shape, articulation, albedo, and lighting of the object from a single image, and generalises to abstract drawings. Our approach demonstrates the power of combining several recent improvements in self-supervised representation learning together with a new viewpoint sampling scheme. We have shown significantly superior results compared to prior works, even when they use more supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Data Ethics. We use the DOVE, Weizmann Horse, PASCAL, Horse-10, and MS-COCO datasets in a manner compatible with their terms. Some of these images may accidentally contain faces or other personal information, but we do not make use of these images or image regions. For further details on ethics, data protection, and copyright please see https://www.robots.ox.ac.uk/~vedaldi/research/union/ethics.html." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. Shangzhe Wu is supported by a research gift from Meta Research. Andrea Vedaldi and Christian Rupprecht are supported by ERC-CoG UNION 101001212. Tomas Jakab and Christian Rupprecht are (also) supported by VisualAI EP/T028572/1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8799" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. In ECCV Workshops, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape completion and animation of people. ACM TOG, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.262, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[4] Eran Borenstein and Shimon Ullman. Class-specific, top-down segmentation. In ECCV, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.47, + 0.391 + ], + "angle": 0, + "content": "[6] John E Chadwick, David R Haumann, and Richard E Parent. Layered construction for deformable animated characters. ACM SIGGRAPH Computer Graphics, 1989. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.394, + 0.47, + 0.448 + ], + "angle": 0, + "content": "[7] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.451, + 0.47, + 0.505 + ], + "angle": 0, + "content": "[8] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[9] Akio Doi and Akio Koide. An efficient method of triangulating equi-valued surfaces by using tetrahedral cells. IEICE Transactions on Information and Systems, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[10] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. *IJCV*, 2015. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[11] Olivier Faugeras. Three-Dimensional Computer Vision: A Geometric Viewpoint. MIT Press, 1993. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.64, + 0.469, + 0.681 + ], + "angle": 0, + "content": "[12] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.684, + 0.469, + 0.725 + ], + "angle": 0, + "content": "[13] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[14] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoints without keypoints. In ECCV, 2020. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[16] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[17] Paul Henderson and Vittorio Ferrari. Learning single-image 3d reconstruction by generative modelling of shape, pose and shading. IJCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[18] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. SCOPS: Self-supervised co-part segmentation. In CVPR, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[19] Wonbong Jang and Lourdes Agapito. CodeNeRF: Disentangled neural radiance fields for object categories. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[20] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[21] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[22] Abhishek Kar, Shubham Tulsiani, João Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In CVPR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.325, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[23] Hiroharu Kato and Tatsuya Harada. Learning view priors for single-view 3d reconstruction. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.381 + ], + "angle": 0, + "content": "[24] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. PointRend: Image segmentation as rendering. In CVPR, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[26] Filippos Kokkinos and Iasonas Kokkinos. Learning monocular 3d reconstruction of articulated categories from motion. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.471, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[27] Filippos Kokkinos and Iasonas Kokkinos. To the point: Correspondence-driven monocular 3d category reconstruction. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.515, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[28] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[29] Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[30] Nilesh Kulkarni, Abhinav Gupta, and Shubham Tulsiani. Canonical surface mapping via geometric cycle consistency. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[31] Xueting Li, Sifei Liu, Shalini De Mello, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Online adaptation for consistent mesh reconstruction in the wild. In NeurIPS, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[32] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In ECCV, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[34] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8800" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[35] Yebin Liu, Juergen Gall, Carsten Stoll, Qionghai Dai, Hans-Peter Seidel, and Christian Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE TPAMI, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[37] Nadia Magnenat-Thalmann, E Primeau, and Daniel Thalmann. Abstract muscle action procedures for human face animation. The Visual Computer, 1988. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.469, + 0.293 + ], + "angle": 0, + "content": "[38] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W. Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In WACV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.469, + 0.349 + ], + "angle": 0, + "content": "[39] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.411, + 0.469, + 0.452 + ], + "angle": 0, + "content": "[41] Tom Monnier, Matthew Fisher, Alexei A. Efros, and Mathieu Aubry. Share With Thy Neighbors: Single-View Reconstruction by Cross-Instance Consistency. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[42] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[43] Kieran A Murphy, Carlos Esteves, Varun Jampani, Srikumar Ramalingam, and Ameesh Makadia. Implicit-PDF: Nonparametric representation of probability distributions on the rotation manifold. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.468, + 0.597 + ], + "angle": 0, + "content": "[44] Natalia Neverova, David Novotny, and Andrea Vedaldi. Continuous surface embeddings. In NeurIPS, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.6, + 0.469, + 0.653 + ], + "angle": 0, + "content": "[45] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3d representations from natural images. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[46] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[48] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[49] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[50] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In CVPR, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[51] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. PVA: Pixel-aligned volumetric avatars. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[52] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[55] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[56] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In NeurIPS, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[57] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[58] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[59] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[60] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representation. In 3DV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[61] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[63] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[64] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.805, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[65] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[66] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8801" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[67] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. DOVE: Learning deformable 3d objects by watching videos. arXiv preprint arXiv:2107.10844, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[68] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In CVPR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[69] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3D objects from images in the wild. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[70] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T. Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[71] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. ViSER: Video-specific surface embeddings for articulated 3d shape reconstruction. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[72] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. BANMo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[73] Chun-Han Yao, Wei-Chih Hung, Michael Rubinstein, Yuanzhen Lee, Varun Jampani, and Ming-Hsuan Yang. Lassie: Learning articulated shape from sparse image ensemble via 3d part discovery. In NeurlPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[74] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[75] Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. Rel-Pose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[76] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8802" + } + ] +] \ No newline at end of file diff --git a/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_origin.pdf b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..23b80bb67352805a3c663ff4b79b7fcefadb9b31 --- /dev/null +++ b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/e6926ba7-5fe0-4c65-9f5a-38d28ecebe6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:984a314106d28c1bc5d6867a783b324430255bbfc2850ac2b238f67e1e1b718e +size 3527056 diff --git a/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/full.md b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e299311a45bc19b332d87dad9c3ab12e67039c76 --- /dev/null +++ b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/full.md @@ -0,0 +1,412 @@ +# MagicPony: Learning Articulated 3D Animals in the Wild + +Shangzhe Wu* Ruining Li* Tomas Jakab* Christian Rupprecht Andrea Vedaldi Visual Geometry Group, University of Oxford + +{szwu, ruining, tomj, chrisr,VEDaldi}@robots.ox.ac.uk 3dmagicpony.github.io + +![](images/0181a9c1d5c0e7932ba2ebc5e93123a1a020bf44e37b64c52f3f959d9833d0b1.jpg) +Figure 1. Learning Articulated 3D Animals in the Wild. Our method trains on a collection of single-view images of an animal category and produces a model that can predict the articulated 3D shape of a new instance from a single test image, which can be animated and relit. + +# Abstract + +We consider the problem of predicting the 3D shape, articulation, viewpoint, texture, and lighting of an articulated animal like a horse given a single test image as input. We present a new method, dubbed MagicPony, that learns this predictor purely from in-the-wild single-view images of the object category, with minimal assumptions about the topology of deformation. At its core is an implicit-explicit representation of articulated shape and appearance, combining the strengths of neural fields and meshes. In order to help the model understand an object's shape and pose, we distil the knowledge captured by an off-the-shelf self-supervised vision transformer and fuse it into the 3D model. To overcome local optima in viewpoint estimation, we further introduce a new viewpoint sampling scheme that comes at no additional training cost. MagicPony outperforms prior work on this challenging task and demonstrates excellent generalisation in reconstructing art, despite the fact that it is only trained on real images. The code can be found on the project page at https://3dmagicpony.github.io/. + +# 1. Introduction + +Reconstructing the 3D shape of an object from a single image of it requires knowing a priori what are the possible shapes and appearances of the object. Learning + +such a prior usually requires ad-hoc data acquisition setups [2, 20, 35, 36], involving at least multiple cameras, and often laser scanners, domes and other hardware, not to mention significant manual effort. This is viable for certain types of objects such as humans that are of particular interest in applications, but it is unlikely to scale to the long tail of objects that can appear in natural images. The alternative is to learn a 3D prior from 2D images only, which are available in abundance. However, this prior is highly complex and must be learned while using it to reconstruct in 3D the 2D training data, which is a major challenge. + +In this paper, we propose MagicPony, a novel approach to learning 3D models of articulated object categories such as horses and birds with only single-view input images for training. We leverage recent progress in unsupervised representation learning, unsupervised image matching, efficient implicit-explicit shape representations and neural rendering, and devise a new auto-encoder architecture that reconstructs the 3D shape, articulation and texture of each object instance from a single image. For training, we only require a 2D segmenter for the object category and a description of the topology and symmetry of its 3D skeleton (i.e., the number and connectivity of bones). We do not require apriori knowledge of the objects' 3D shapes, keypoints, viewpoints, or of any other 2D or 3D cue which are often used in prior work [14, 21, 22, 31]. From this, we learn a function that, at test time, can estimate the shape and texture of a new object from a single image, in a feedforward manner. The function exhibits remarkable gener + +alisation properties, including reconstructing objects in abstract drawings, despite being trained on real images only. + +In order to learn such disentangled 3D representations simply from raw images, MagicPony addresses a few key challenges. The first challenge is viewpoint estimation. Before the 3D model is available, it is very difficult to assign a viewpoint to the objects in the training images. To tackle this problem, prior works have often assumed additional cues such as 2D keypoint correspondences [21, 22, 31]. Instead, we avoid requiring additional keypoint supervision and implicitly infer noisy correspondences by fusing into the 3D model knowledge distilled from DINO-ViT [5], a self-supervised visual transformer network (ViT) [28]. We also develop a new efficient disambiguation scheme that explores multiple viewpoint assignment hypotheses at essentially no cost, avoiding local optima that are caused by greedily matching the noisy 2D correspondences. + +The second challenge is how to represent the 3D shape, appearance and deformations of the object. Most prior works have used textured meshes [13, 14, 21, 31, 32, 67, 70], but these are difficult to optimise from scratch, leading to problems that often require ad-hoc heuristics such as remeshing [13, 70]. The other and increasingly popular approach is to use a volumetric representation such as a neural radiance field [3, 40, 52, 63], which can model complex shapes, including manipulating their topology during training. However, this modelling freedom comes at the cost of over-parametrisation, which is particularly problematic in monocular reconstruction and often leads to meaningless short-cut solutions [63]. Furthermore, modelling articulation with a volumetric representation is difficult. A posing transformation is only defined for the object's surface and interior and is more easily expressed from the canonical/pose-free space to the posed space. However, rendering a radiance field requires transforming 3D points off the object surface and in the direction opposite to the posing transformation, which is hard [8]. + +We address these issues by using a hybrid volumetric-mesh representation based on DMTet [42, 56]. Shape and appearance are defined volumetrically in canonical space, but a mesh is extracted on the fly for posing and rendering. This sidesteps the challenges of using neural rendering directly while retaining most of its advantages and enabling the use of powerful shape regularisers. + +To summarise, we make the following contributions: (1) A new 3D object learning framework that combines recent advances in unsupervised learning, 3D representations and neural rendering, achieving better reconstruction results with less supervision; (2) An effective mechanism for fusing self-supervised features from DINO-ViT into the 3D model as a form of self-supervision; and (3) an efficient multi-hypothesis viewpoint prediction scheme that avoids local optima in reconstruction with no additional cost. + +Table 1. Related Work Overview on Weakly-supervised Learning of 3D Objects. Annotations: template shape, viewpoint, 2D keypoint, object mask, optical flow, video, coarse template shape from keypoints, camera estimated from keypoints using SfM, outputs texture flow, shape bases initialised from CMR. UMR relies on part segmentations from SCOPS [18]. + +
MethodSupervisionOutput
3D2.5DMotion ViewTexture
Unsup3D [69]✓✓✓
CSM [30]✓✓✓
A-CSM [29]✓✓✓✓
CMR [21](✓)1(✓)2✓✓✓✓(✓)3
U-CMR [14]✓✓✓✓✓
UMR†[32]✓✓✓(✓)3
ACMR [31](✓)4(✓)2✓✓✓✓(✓)3
DOVE [67]✓✓✓✓✓✓
Ours✓✓✓✓✓
+ +We compare our method to prior work on several challenging articulated animal categories and show that our method obtains significantly better quantitative and qualitative results while using significantly less supervision, and also demonstrate generalisation to abstract drawings. + +# 2. Related Work + +Weakly-supervised Learning of 3D Objects. Most related to our work are weakly-supervised methods for single-image 3D object reconstruction that learn from a collection of images or videos of an object category [14, 17, 21, 23, 24, 26, 27, 31, 32, 34, 65, 67, 68]. Due to the inherent ambiguity of the problem, these methods usually rely on heavy geometric supervision (see Tab. 1) in addition to object masks, such as 2D keypoints [21, 31], template shapes [14, 29, 30]. UMR [32] forgoes the requirement of keypoints and template shapes by leveraging weakly-supervised object part segmentations (SCOPS [18]), but produces coarse symmetric shapes similar to CMR [21]. DOVE [67] learns coarse articulated objects by exploiting temporal information from videos with optical flow supervision. All these methods require object masks for supervision, except Unsup3D [69] which exploits bilateral symmetry for reconstructing roughly frontal objects, like faces, and UNICORN [41] which uses a progressive conditioning strategy with a heavily constrained bottleneck, leading to coarse reconstructions. Another emerging paradigm is to leverage generative models [7, 45, 49, 55] that encourage images rendered from viewpoints sampled from a prior distribution to be realistic. These models, however, often rely heavily on an accurate estimation of viewpoint distribution and/or a powerful 2D image generative model, both of which are difficult to obtain in small-data scenarios. + +Optimization from Multi-views and Videos. 3D reconstruction traditionally relies on epipolar geometry of multiview images of static scenes [11, 16]. Neural Radiance Fields (NeRF) [3, 40, 52, 63] have recently emerged as a powerful volumetric representation for multi-view reconstruction given accurate cameras. A recent line of work, LASR [70], ViSER [71] and BANMo [72], optimises 3D shapes of deformable objects from a small set of monocular videos, with heavily engineered optimisation strategies using optical flow and mask supervision, and additionally DensePose [44] annotations for highly deformable animals. Concurrent work of LASSIE [73] also leverages DINO-ViT [5] image features for supervision but optimises a part-based model on a small collection of images ( $\sim 30$ ) of an object category. It uses a shared shape model for all instances with only per-instance articulation. + +Learnable Deformable 3D Representations. A common 3D representation for weakly-supervised shape learning is triangular meshes. Deformation on meshes can be modelled by estimating offsets of individual vertices [21, 31, 67], which typically requires heavy regularisation, such as As-Rigid-As-Possible (ARAP) [57]. To constrain the space of deformations, many works often utilise lower-dimensional models, such as cages [74] or linear blend skinning controlled with skeletal bones [36, 67] or Gaussian control points [70, 71]. Parametric models, like SMPL [36] and SMAL [76], allow for realistic control of the shape through learned parameters but often require an extensive collection of 3D scans to learn from. + +Dynamic neural fields [12,46,48,50,51,58,59,72], such as D-NeRF [50], extend NeRF [40] with time-varying components. A-NeRF [58] proposes a radiance field conditioned on an articulated skeleton, while BANMo [72] uses learned Gaussian control points similarly to [70, 71]. One key limitation of these implicit representations is the requirement of an inverse transformation from 3D world coordinates back to the canonical space, which is often harder to learn compared to forward deformation [8]. We propose a hybrid SDF-mesh representation, extending DMTet [56] with an articulation mechanism, which combines the expressiveness of implicit models with the simplicity of mesh deformation and articulation. + +Viewpoint Prediction in the Wild. Viewpoint prediction in weakly-supervised settings is challenging, as it is prone to local optima induced by projection and common object symmetries. Existing solutions are based on learning a distribution over multiple possible viewpoints. U-CMR [14] proposes a camera-multiplex that optimises over 40 cameras for each training sample using a given template shape, which is expensive as evaluating each hypothesis involves a rendering step. We propose a new viewpoint exploration scheme that requires evaluating only one hypothesis in each + +step with essentially no added training cost and does not rely on a given template shape but a jointly learned model. Recent work of Implicit-PDF [43] proposes to learn viewpoint distributions implicitly by estimating a probability for each image-pose pair of a fixed object instance. RelPose [75] extends this to learning relative poses of image pairs. Our model also predicts a probability associated with each viewpoint hypothesis but instead relates an instance to a learned category-level prior model. + +# 3. Method + +Given only a collection of single-view images of a deformable object category collected in the wild, our goal is to train a monocular reconstruction network that, at test time, predicts the 3D shape, articulated pose, albedo and lighting from a single image of the object. + +During training, our method requires no geometric supervision other than the images, 2D foreground segmentations of the object obtained automatically from an off-the-shelf method like [25], and a description of the topology of the object's skeleton (e.g., the number of legs). It is based on several key ideas: an implicit-explicit shape representation (Sec. 3.1), a hierarchical shape representation, from a generic template to articulated instances (Sec. 3.2), and a viewpoint prediction scheme that leverages self-supervised correspondences and efficient multi-hypothesis exploration to avoid local minima (Sec. 3.4), as illustrated in Fig. 2. + +# 3.1. Implicit-Explicit 3D Shape Representation + +The choice of representation for the 3D shape of the object is critical. It must be (1) sufficiently expressive for modelling fine-grained shape details and deformations and (2) sufficiently regular for learning with only weak supervision via image reconstruction. + +Most 3D representations are either explicit (triangular meshes [21, 67, 70]) or implicit (volumes, SDFs, radiance fields [19, 39, 40, 47, 66]). Explicit representations are more compact, easily support powerful deformation models like blend skinning, and the geometric smoothness can be easily measured and controlled; however, they are more difficult to optimise and often result in defects like folds during training, which requires occasional re-meshing [13, 70]. Implicit representations are more amenable to gradient-based optimisation, including allowing topological changes, but are more computationally expensive and, due to their high capacity, require substantial supervision, usually in the form of a dense coverage of the different object viewpoints [40, 66] or 3D ground-truth [39, 47]. + +Here we adopt a representation that combines the advantages of both. The shape of the object is represented implicitly by a neural field that is converted on the fly into an explicit mesh via the Differentiable Marching Tetrahedra (DMTet) [9, 42, 56] method. Specifically, the object's shape + +![](images/f95d58eb7b60e8aa8d832efa4be77d292fcbe6ba108adc2199c5dfdc6c85839e.jpg) +Figure 2. Training Pipeline. Given a collection of single-view images of an animal category, our model learns a category-specific prior shape using an implicit-explicit representation, together with a feature field that allows us to fuse self-supervised correspondences through a feature rendering loss. This prior shape is deformed, articulated and shaded based on features extracted from a training image. To combat local minima in viewpoint prediction, we introduce an efficient scheme that explores multiple hypotheses at essentially no additional cost. The entire pipeline is trained with reconstruction losses end-to-end, except for the frozen image encoder pre-trained via self-supervision. + +$S = \{\pmb{x} \in \mathbb{R}^3 | s(\pmb{x}) = 0\}$ is defined as the isosurface of the signed distance function (SDF) $s: \mathbb{R}^3 \mapsto \mathbb{R}$ , which is implemented by a Multi-Layer Perceptron (MLP) taking 3D coordinates as input. To encourage $s$ to be a signed distance function, we sample random points $\pmb{x} \in \mathcal{X} \subset \mathbb{R}^3$ in space and minimise the Eikonal regulariser [15]: + +$$ +\mathcal {R} _ {\mathrm {E i k}} (s) = \frac {1}{| \mathcal {X} |} \sum_ {\boldsymbol {x} \in \mathcal {X}} \left(\left\| \nabla s (\boldsymbol {x}) \right\| _ {2} - 1\right) ^ {2}. \tag {1} +$$ + +We then evaluate the function $s$ at the vertices of a tetrahedral grid and extract a watertight mesh from it using DMTet [56], which is fast and differentiable. + +We use this implicit-explicit representation to learn a category-wise template shape shared across all instances, and model instance deformations through a hierarchical pipeline, explained next. The function $s$ is initialised with a generic ellipsoid elongated along the $z$ -axis. We denote by $V_{\mathrm{pr}} \subset \mathbb{R}^3$ the vertices of the extracted prior mesh. + +# 3.2. Articulated Shape Modelling and Prediction + +The category-wise template $V_{\mathrm{pr}}$ is deformed into a specific object instance $V_{\mathrm{ins}}$ , and further posed to match the shape $V$ in each input image (Fig. 2). These deformations are predicted from a single image, as discussed below. + +Image Encoder. The input image is first encoded by means of DINO-ViT [5], a self-supervised Vision Transformer (ViT) [28]. The ViT extracts from the image a set of key and output tokens, which were found to code, respectively, for correspondences [1] and appearance [61]. Hence, we use the keys to predict the object's deformation and the outputs to predict its appearance. Specifically, we apply the ViT $\Phi$ to the input image $I\in \mathbb{R}^{3\times H\times W}$ and extract the key $\Phi_{\mathrm{k}}\in \mathbb{R}^{D\times H_p\times W_p}$ and output $\Phi_{\mathrm{o}}\in \mathbb{R}^{D\times H_p\times W_p}$ tensors from the last attention layer. These features are local, + +with one token per image patch. We use two small convolutional networks to further extract from these a global key $\phi_{\mathrm{k}} = f_{\mathrm{k}}(\Phi_{\mathrm{k}})\in \mathbb{R}^{D}$ and output $\phi_{\mathrm{o}} = f_{\mathrm{o}}(\Phi_{\mathrm{o}})\in \mathbb{R}^{D}$ vector. + +Instance-specific Shape. The shape of each object instance $V_{\mathrm{ins}}$ differs slightly from the template shape $V_{\mathrm{pr}}$ . This is accounted for by the deformation $V_{\mathrm{ins},i} = V_{\mathrm{pr},i} + \Delta V_i$ , where the displacement $\Delta V_i = f_{\Delta V}(V_{\mathrm{pr},i},\phi_{\mathrm{k}})$ of each vertex $V_{\mathrm{pr},i}$ is predicted by an MLP $f_{\Delta V}$ from the global image features $\phi_{\mathrm{k}}$ . We exploit the fact that many objects are bilaterally symmetric and enforce both the prior $V_{\mathrm{pr}}$ and instance $V_{\mathrm{ins}}$ shapes to be symmetric by mirroring the query locations for the MLPs about the $yz$ -plane. We also limit the magnitude of the deformation $V_{\mathrm{ins}}$ by adding the regularizer $\mathcal{R}_{\mathrm{def}}(\Delta V) = \frac{1}{|V_{\mathrm{pr}}|}\sum_i\| \Delta V_i\| _2^2$ . + +Posed Shape. In order to match the 3D shape observed in the image, the instance-specific mesh $V_{\mathrm{ins}}$ is further deformed by the posing function $V_{i} = g(V_{\mathrm{ins},i},\xi)$ , where $\xi$ denotes the pose parameters. Posing accounts for the deformations caused by the articulation of the object's skeleton; they allow for large but controlled deformations which would otherwise be more difficult to account for by a generic deformation field like $\Delta V$ . + +We adopt the blend skinning model for posing [6, 37], parameterised by $B - 1$ bone rotations $\xi_{b} \in SO(3), b = 2, \ldots, B$ , and the viewpoint $\xi_{1} \in SE(3)$ . Specifically, we initialise a set of rest-pose joint locations $\mathbf{J}_{b}$ on the instance mesh using simple heuristics1. Each bone $b$ but the root has one parent $\pi(b)$ , thus forming a tree structure. Each vertex $V_{i}$ is associated to the bones by the skinning weights $w_{ib}$ , defined based on relative proximity to each bone (see + +sup. mat.for details). The vertices are then posed by the linear blend skinning equation: + +$$ +V _ {i} (\xi) = \left(\sum_ {b = 1} ^ {B} w _ {i b} G _ {b} (\xi) G _ {b} \left(\xi^ {*}\right) ^ {- 1}\right) V _ {\text {i n s}, i}, \tag {2} +$$ + +$$ +G _ {1} = g _ {1}, G _ {b} = G _ {\pi (b)} \circ g _ {b}, g _ {b} (\boldsymbol {\xi}) = \left[ \begin{array}{c c} R _ {\xi_ {b}} & \mathbf {J} _ {b} \\ 0 & 1 \end{array} \right], +$$ + +where $\xi^{*}$ denotes the bone rotations at the rest pose. + +Pose Prediction. Similarly to the instance-specific deformation $\Delta V_{i}$ , the bone rotations $\xi_{2:B}$ are predicted by a network $f_{\mathrm{b}}$ from the image features $\Phi_{\mathrm{k}}$ and $\phi_{\mathrm{k}}$ . The viewpoint $\xi_{1} \in SE(3)$ is predicted separately, as discussed in Sec. 3.4. + +Pose articulation depends on estimating local body joints and benefits from using local features. To extract them, we first project the centroid of each bone at rest into the image plane using the predicted viewpoint $\xi_{1}$ : + +$$ +\boldsymbol {u} _ {b} = \Pi \left[ G _ {\pi (b)} \left(\xi_ {1}, \xi_ {2: B} ^ {*}\right) \frac {\mathbf {J} _ {b}}{2} \right], +$$ + +and sample the local features from the patch key feature map $\Phi_{\mathbf{k}}(\pmb{u}_b)$ at the projected pixel locations $\pmb{u}_b$ + +To predict bone rotations from the image features, we adopt a transformer to account for the inter-dependency of the bones, implied by the underlying skeleton. Specifically, we use a small transformer network [62] $f_{\mathrm{b}}$ that takes as input $B - 1$ tokens $\nu_{b} = (\phi_{\mathrm{k}}, \Phi_{\mathrm{k}}(\pmb{u}_{b}), b, \mathbf{J}_{b}, \pmb{u}_{b})$ and outputs the bone rotations $(\xi_{2}, \ldots, \xi_{B}) = f_{\mathrm{b}}(\nu_{2}, \ldots, \nu_{B})$ parameterised as Euler angles. We also limit the deviation from the rest pose via the regulariser $\mathcal{R}_{\mathrm{art}}(\xi) = \frac{1}{B - 1} \sum_{b=2}^{B} \| \xi_{b} \|_{2}^{2}$ . + +# 3.3. Appearance and Shading + +We decompose the appearance of the objects into albedo and diffuse shading assuming a Lambertian illumination model. The albedo of the object is modelled in the same "canonical space" as the shape SDF $s$ by a second coordinate MLP $a(\pmb{x}) = f_{\mathrm{a}}(\pmb{x}, \phi_{\mathrm{o}}) \in [0,1]^3$ , which takes as input the global output features $\phi_{\mathrm{o}}$ . The dominant light direction $l \in \mathbb{S}^2$ and ambient and diffuse intensities $k_{s}, k_{d} \in \mathbb{R}$ are predicted by a lighting network $f_{\mathrm{l}}(\phi_{\mathrm{o}})$ . + +The posed mesh is then rendered into an image via differentiable deferred mesh rendering. The renderer first rasterizes all pixels $\pmb{u}$ that overlap with the posed object $V$ , computing the 3D coordinates $\pmb{x}(\pmb{u})$ of the corresponding object point in canonical space; then the MLP $f_{\mathrm{a}}$ is evaluated, once per pixel, to obtain the colour $a(\pmb{x}(\pmb{u}))$ . The final pixel colour accounts for shading and is given by + +$$ +\hat {I} (\boldsymbol {u}) = \left(k _ {s} + k _ {d} \cdot \max \{0, \langle l, n (\boldsymbol {u}) \rangle \}\right) \cdot a (\boldsymbol {x} (\boldsymbol {u})), +$$ + +where $n(\pmb{u}) \in \mathbb{S}^2$ is the normal of the posed mesh $V$ . Using the same differentiable renderer, we also obtain the predicted mask $\hat{M} \in [0,1]^{H \times W}$ of the object. + +For training supervision, we use the appearance loss $\mathcal{L}_{\mathrm{im}}(\hat{I}, I, M) = \| \tilde{M} \odot (\hat{I} - I)\|_1$ as well as the mask loss $\mathcal{L}_{\mathrm{m}}(\hat{M}, M) = \| \hat{M} - M\|_2^2 + \lambda_{\mathrm{dt}}(\hat{M} \odot \mathrm{d}\tau(M))$ , where the distance transform $\mathrm{d}\tau(\cdot)$ is used to obtain better gradients [21, 68], $\tilde{M} = \hat{M} \odot M$ is the intersection of the rendered and ground-truth masks, and $\lambda_{\mathrm{dt}}$ is a weighing factor. + +# 3.4. Viewpoint Prediction + +Predicting the object viewpoint is a critical first step for learning 3D shapes but often difficult [14, 21]. Here, we contribute a method to do so robustly and efficiently, which exploits self-supervised correspondences in a multi-hypothesis prediction pipeline, as detailed next. + +Fusing Self-Supervised Correspondences. Estimating explicit correspondences without supervision is difficult and brittle; hence, we propose to incorporate correspondence information implicitly and softly. Inspired by N3FF [60], we fuse information from the self-supervised ViT features into the 3D model by learning to render them. This is done by adding a coordinate network $\psi(\boldsymbol{x}) \in \mathbb{R}^{D'}$ that predicts a $D'$ -dimensional feature vector for each 3D point $\boldsymbol{x}$ in canonical space. This is similar to the albedo network $f_{a}$ , except that it is shared across all instances of the same object category, and hence there is no dependency on the input image. These canonical features are then projected into a feature image $\hat{\Phi}_{\mathrm{k}} \in \mathbb{R}^{H \times W \times D'}$ using the same differentiable renderer used for the albedo, and trained to minimise a corresponding rendering loss: + +$$ +\mathcal {L} _ {\text {f e a t}} \left(\hat {\Phi} _ {\mathrm {k}}, \Phi_ {\mathrm {k}} ^ {\prime}, \tilde {M}\right) = \| \tilde {M} \odot \left(\hat {\Phi} _ {\mathrm {k}} - \Phi_ {\mathrm {k}} ^ {\prime}\right) \| _ {2} ^ {2}, \tag {3} +$$ + +where $\Phi_{\mathrm{k}}^{\prime} = \mathrm{PCA}\bigl (\Phi_{\mathrm{k}}\bigr)$ is a PCA-reduced (to $D^{\prime} = 16$ principal components) version of the original DINO-ViT features $\Phi_{\mathrm{k}}$ for training efficiency. + +Multi-Hypothesis Viewpoint Prediction. Prior works have found that a major challenge in learning the object's viewpoint is the existence of multiple local optima in the reconstruction objective. Some have addressed this issue by sampling a large number of hypotheses for the viewpoint [14], but this is somewhat cumbersome and slow. Instead, we propose a scheme that explores multiple viewpoints statistically, but at each iteration samples only a single one, and thus comes "for free". + +We hence task a viewpoint network $f_{\mathrm{vp}}(\phi_k)$ to predict from $\phi_{\mathrm{k}}$ four viewpoint rotation hypotheses $R_{k} \in SO(3), k \in \{1,2,3,4\}$ , each in one of the four quadrants around the object, as illustrated in Fig. 2. The network also predicts a score $\sigma_{k}$ for each hypothesis, used to evaluate the + +probability $p_k$ that hypothesis $k$ is the best of the four options as $p_k \propto \exp(-\sigma_k / \tau)$ using softmax, where $\tau$ is a temperature parameter, gradually decreased during training. + +The naive approach for learning $\sigma_{k}$ is to sample multiple hypotheses and compare their reconstruction loss to determine which one is better. However, this is expensive as it requires rendering the model with all hypotheses. Instead, we suggest to sample a single hypothesis for each training iteration and simply train $\sigma_{k}$ to predict the expected reconstruction loss $\tilde{\mathcal{L}}_k$ , by minimising the objective + +$$ +\mathcal {L} _ {\mathrm {h y p}} \left(\sigma_ {k}, \tilde {\mathcal {L}}\right) = \left(\sigma_ {k} - \operatorname {s g} \left[ \tilde {\mathcal {L}} _ {k} \right]\right) ^ {2}, \tag {4} +$$ + +where $\mathrm{sg}[\cdot ]$ is the stop-gradient operator, and $\tilde{\mathcal{L}}_k$ is the overall reconstruction loss under the sampled $k$ -th hypothesis, given below. + +# 3.5. Training Objective + +Given a training sample $(I, M)$ , comprising an image and the corresponding object mask, the training objective is computed in two steps. First, the viewpoint hypothesis $k$ is sampled and the reconstruction loss is computed: + +$$ +\tilde {\mathcal {L}} _ {k} = \lambda_ {\mathrm {i m}} \mathcal {L} _ {\mathrm {i m}} + \lambda_ {\mathrm {m}} \mathcal {L} _ {\mathrm {m}} + \lambda_ {\mathrm {f}} \mathcal {L} _ {\text {f e a t}}. +$$ + +Then, the regularisation terms and the viewpoint hypothesis loss are added, to obtain the final objective: + +$$ +\mathcal {L} = p _ {k} \tilde {\mathcal {L}} _ {k} + \lambda_ {\mathrm {E}} \mathcal {R} _ {\text {E i k}} + \lambda_ {\mathrm {d}} \mathcal {R} _ {\text {d e f}} + \lambda_ {\mathrm {a}} \mathcal {R} _ {\text {a r t}} + \lambda_ {\mathrm {h}} \mathcal {L} _ {\text {h y p}}. +$$ + +Here $\lambda_{\mathrm{im}},\lambda_{\mathrm{m}},\lambda_{\mathrm{f}},\lambda_{\mathrm{E}},\lambda_{\mathrm{d}},\lambda_{\mathrm{a}},\lambda_{\mathrm{h}}$ are balancing weights. + +The gradient of $\mathcal{L}$ with respect to the model parameters is then computed and used to update them. Note that only one viewpoint sample $k$ is taken for each gradient evaluation. To further improve the sampling efficiency, we sample the viewpoint based on the learned probability distribution $p_k^4$ . + +# 4. Experiments + +We conduct extensive experiments on a few animal categories, including horses, giraffes, zebras, cows and birds, and compare against prior work both qualitatively and quantitatively on standard benchmarks. We also show that our model trained on real images generalises to abstract drawings, demonstrating the power of unsupervised learning. + +# 4.1.Data + +For horses, we use the horse dataset from DOVE [67] containing 10.8k images extracted from YouTube, and supplement it with 541 additional images from three datasets for diversity: Weizmann Horse Database [4], PASCAL [10] and Horse-10 Dataset [38]. For giraffes, zebras and cows, + +we source images from Microsoft COCO Dataset [33] and keep the ones with little occlusion. Since these datasets are relatively small, we only use them to finetune the pretrained horse model. For birds, we combine the DOVE dataset [67] and CUB dataset [64] consisting of $57.9\mathrm{k}$ and $11.7\mathrm{k}$ images respectively. We use an off-the-shelf PointRend [25] detector to obtain segmentation masks, crop around the objects and resize them to $256\times 256$ . We follow the original train/test splits from CUB and DOVE, and randomly split the rest, resulting in $11.5\mathrm{k} / 0.8\mathrm{k}$ horse, $513 / 57$ giraffe, $574 / 64$ zebra, $719 / 80$ cow and $63.9\mathrm{k} / 10.5\mathrm{k}$ bird images. We additionally collect roughly 100 horse images from the Internet to test generalisation. + +# 4.2. Technical Details + +The model is implemented using a total of 8 subnetworks. The feature field $\psi$ , template SDF $s$ , albedo field $f_{\mathrm{a}}$ and deformation field $f_{\Delta V}$ are implemented using MLPs, which take in 3D coordinates (concatenated with image features for albedo and deformation networks) as input. The image encoder consists of a ViT-S [28] architecture from DINO [5] and two convolutional encoders $f_{\mathrm{k}}$ and $f_{\mathrm{o}}$ that fuse the patch features into global feature vectors. We use self-supervised pre-trained DINO-ViT with frozen weights and only train the convolutional encoders. The lighting network $f_{1}$ is a simple MLP that maps global output feature to a 4-channel vector, and the articulation network $f_{\mathrm{b}}$ is a transformer architecture with 4 blocks. We use a separate encoder for the viewpoint network $f_{\mathrm{vp}}$ identical to $f_{\mathrm{k}}$ , as we find empirically that sharing the encoders tends to make the viewpoint learning unstable. Apart from DINO encoder, all components are trained end-to-end from scratch for 100 and 10 epochs on horses and birds respectively, with the instance deformation and articulation disabled for the first 30 and 2 epochs. The DINO features $\Phi_{\mathrm{k}}^{\prime}$ used to compute $\mathcal{L}_{\mathrm{feat}}$ are pre-computed with patch size 8 and stride 4 and masked, following [1], and reduced from a channel size of 384 to 16 using PCA. All technical details and hyper-parameters are included in the supplementary material. + +# 4.3. Qualitative Results + +Fig. 3 shows a few reconstructions of horses and birds produced by our model. Given a single 2D image at test time, our model predicts a textured 3D mesh of the object, capturing its fine-grained geometric details, such as the legs and tails of the horses. Our model predicts the articulated pose of the objects, allowing us to easily transfer the pose of one instance from another and animate it in 3D. Although our model is trained on real images only, it demonstrates excellent generalisation to paintings and abstract drawings. + +Fig. 4 also shows that by finetuning the horse model on other animals without any additional modifications (except disabling articulation for the initial 5k iterations), it gen + +![](images/1ec368feec8df80a847117319297db59b7b4a027c6f3805173a33f18227606ce.jpg) + +![](images/bb8813596399eb89c40ae684a7e226a45291efa9307381f7db819f6e1653f7eb.jpg) + +![](images/1a9df4288b5199457caeed34396a0538d9dd189de62c921b7ca1eab9682ffe67.jpg) + +![](images/03093c552a79df8d1f58122bef9295c15765023b27f0cb81323f3c7ba0a48e2d.jpg) +Input + +![](images/397fca09d3125ae645c61891f93dfc230a730c08307272dab17d664b813e7b3c.jpg) +CMR + +![](images/6b3b7b34dfe888196b20ad9f413a19362d910b9eafcf343a0597c583f1b52a0a.jpg) + +![](images/6e94fa1a600a8fbfd8492aa1bea4089827bc9d6bee4978566141d7d94fd384da.jpg) + +![](images/77d30e02660a2111851c3dc9d474eeed941080499348a35fdd5f87551f75430f.jpg) +Reconstructions +Textured +Animated + +![](images/7a28d87885ae2b7f24c449130f1e175cb7274257b5f1c44071de4cbe3b79f1f2.jpg) +Figure 3. Single Image Reconstruction. We show the reconstructed mesh from the input view and two additional views together with predicted texture and animated version of the shape obtained by articulating the estimated skeleton. + +![](images/3705cf23c55248ffac3b192255ab12acafa62faae4160c61eac655b75a7744d3.jpg) + +![](images/58cc294826e8105e7de96a450dd44572b73b5df61ab28104cc1eccbd4faf3861.jpg) +Input +Figure 4. Reconstructions of Giraffes, Zebras and Cows. We finetune the horse model on other animal categories, and show that the method generalises well to other animals. + +![](images/d6518ddcf9016d9924d63c7c1be86b8e0b85220fa65a9c3abe9677010687f3e2.jpg) + +![](images/577540a015dc4229f3d03956bb621aa23413ef4e5962f8167ab093978216b908.jpg) + +![](images/85f59bf0550fb776abec6067e49f24d889f9ad0a0292c018b48f81a977c17d07.jpg) +Reconstructions +Textured +Animated + +eralises to various animal categories with highly different shapes and appearances, such as giraffes, zebras and cows. + +Note that for the examples of abstract horse drawings (last rows in Figs. 1 and 3), since the texture is out of the training distribution, we finetune (only) the albedo network for 100 iterations, which takes less than 10 seconds. This is also done for giraffes, zebras and cows in Fig. 4, as the training datasets are too small to learn complex appearance. Additional qualitative results are presented in the sup. mat. + +# 4.4. Comparison with Prior Work + +We compare with previous weakly-supervised methods for 3D reconstruction of deformable objects. The most relevant prior work is UMR [32] and DOVE [67], as they also only require object masks and weak correspondences from either part segmentations (SCOPS) [18] or video training [67]. Our method leverages self-supervised DINO-ViT features, which makes our method the least supervised in + +![](images/a75296467bb8ba6fc586a4ad700fda96e4bd8febcfba23cb02cde3872af5ea97.jpg) + +![](images/eb613c7ff970271f282a66129e503565cdeed4bc4683c7b9ce136841758aa610.jpg) +Input Image + +![](images/a8cf81f53bd65d742a41905ff378a4176df9a7c31e1935b094f8066e78aa39d7.jpg) + +![](images/0817229915715f0d08151937915ca92bd9068eada5d5642354f1005f6cc85406.jpg) +Input View +Other Views + +![](images/b3cc63efc3281a9969e58de31326818a3f37b20f1de4662fd25d39066436eded.jpg) + +![](images/18e01ec428f8e73751db0a1c56f201752927291ebacfa66975f0b1c382a287b8.jpg) +Input Image + +![](images/fe618f9c9c11dcaca113678b31d9738a303af75c17086ae3cfb366d072cf5a55.jpg) + +![](images/64ad89b35b46684ca3c27803a7ad8a794215f61970aa60253160d6fb642bea23.jpg) + +![](images/38b343d7fb4584ea6c88396c70f67baaa50e4df650b80ed72b6d6ac1975f105f.jpg) + +![](images/c416e2da3266b2a4c165b21443d3f56d8244501054289a10c1e77bed3f830121.jpg) + +![](images/df240beec029a314bcfc99ac10e39dd87abbbe42d21284a1e9a9e63943050a84.jpg) + +![](images/ac1956487352e3f6164cdb79447993a05ed8894a243dcba6d5626fc2ac5e850e.jpg) +Input View + +![](images/933ec2e7fa1c9e6a5ddf8a515b42b305f47f0cd1fee39086c816a61c0b649e8a.jpg) +Other Views + +![](images/0fa23d4fbbb5ed927fadb93eeea9544dfed710df3f8ae0100897834eee11afa9.jpg) +Figure 5. Comparison on Birds. Both CMR [21] and UMR [32] often predict inaccurate poses, such as the bird on the left. + +![](images/14f701b684614336a9643db8048950891f1da6a8c4732fca4780577f53bf04eb.jpg) + +![](images/d60f8866c77fb9c0a963b151e5cf3ebde94b2828f8ea43c711781b3394182bfd.jpg) + +![](images/a0509fc8b9f89ce27d4ecabc82245b6f54d850da2ec6a25bc3b6c748fc4e4d6e.jpg) +UMR +Figure 6. Comparison with UMR [32] on Horses. Our model produces much higher quality reconstructions with four legs. UMR is only able to predict symmetrical shapes with two legs. + +![](images/821da40179a107739367c1522eef6ab489da739ebce3084792fc4b3ba67a1ae0.jpg) + +![](images/19774c162073b761f8a087cd43ceee893647f49468735935e64e1a076d3ac8a7.jpg) + +![](images/0537220f7f27445c7a778391f8fc8af328ea7dc33c40a55fa979b9ad5f5b7618.jpg) + +![](images/7054a5af1b7638e894b072e987bf4117fd4101244a3e5d224c2aeac503d6131e.jpg) +Ours + +this area. We also compare against the well-established baseline of CMR [21], which requires stronger supervision in the form of 2D keypoint annotations and pre-computed cameras, as well as U-CMR [14], which uses a category-specific template shape. To evaluate other methods, we use their code and pre-trained models in public repositories (CMR, UMR, U-CMR) or provided by the authors (DOVE). + +Qualitative Comparisons. Figs. 5 and 6 compare our reconstructions against the previous methods on bird and horse images. Since UMR did not release their horse model, we test our model on the examples presented in their paper for side-by-side comparison. Our method predicts more accurate poses and higher-quality articulated shapes (as opposed to only symmetric shapes with CMR and UMR). On horses, our method reconstructs all four legs with details while UMR predicts only two. Since UMR copies texture from the input image, for a fair comparison, we finetune (only) the albedo network for 100 iterations at test time. + +Quantitative Comparisons. We evaluate shape reconstruction on the 3D Toy Bird Dataset from DOVE [67], which contains 3D scans of 23 realistic toy bird models paired with 345 photographs of them taken in natural real-world environments. Following [67], we align the predicted mesh with the ground-truth scan using ICP and compute the + +Table 2. Evaluation on Toy Bird Scans [67]. Baseline results reported by [67]. Annotations: template shape, viewpoint, 2D keypoint, $\mathfrak{S}\ll$ mask, $\nVdash$ optical flow, video. + +
SupervisionChamfer Distance (cm) ↓
CMR [21]® ®1.35 ±0.81
U-CMR [14]® ®1.82 ±0.93
UMR [32]® + SCOPS1.24 ±0.75
DOVE [67]® ➔ ➔ ➔1.51 ±0.89
Ours® + DINO0.79 ±0.50
+ +Table 3. Keypoint Transfer on PASCAL [10] Horse and CUB [64] Bird Datasets. Our model produces superior results on both categories with significantly less supervision. $\dagger$ : numbers taken from [32]. 'Birds w/o aq.': excluding 50 aquatic bird classes (e.g. mallards) with heavy occlusion and large shape variation. + +
MethodSupervisionPCK@0.1 ↑
HorsesBirdsBirds w/o aq.
CMR [21]†♂♀♂-0.473-
CMR [21]♂♀♂-0.5460.591
CMR [21]♂♀-0.2550.277
U-CMR [14]♀♀♀-0.3590.412
UMR [32]♂♀+ SCOPS0.2840.5120.555
A-CSM [29]♀♀♀0.3290.426-
DOVE [67]♂♀♂-0.4470.510
Ours♂♀+ DINO0.4290.5550.635
+ +bi-directional Chamfer Distance between two sets of sample points from the aligned meshes. Tab. 2 summarises the results compared to other methods. Our model produces significantly more accurate reconstructions, resulting in a much lower error. + +We also evaluate Keypoint Transfer on the CUB [64] and PASCAL VOC [10] benchmarks, a common evaluation metric [21, 29, 32]. We follow the protocols in [32] and [29] for birds and horses, respectively, and sample 20k source and target image pairs. Given a source image, we project all the visible vertices of the predicted mesh onto the image using the predicted viewpoint and assign each annotated 2D keypoint to its nearest vertex. We then render these vertices using the mesh and viewpoint predicted from a different target image. We measure the error between the projected vertices (corresponding to the transferred keypoints) and the annotated target keypoints using the Percentage of Correct Keypoints (PCK) metric. As shown in Tab. 3, our model significantly outperforms previous methods relying on stronger geometric supervision, including A-CSM [29] which requires a 3D shape template to begin with. + +# 4.5. Ablation Studies and Limitations + +We present ablation studies on the two key components of the model in Fig. 7, self-supervised feature loss and multi-hypothesis viewpoint prediction. Without either + +![](images/920117262897d42f6c06a502a0ec2664a468d46dc4643c1269fa6f239cb736f1.jpg) +Figure 7. Ablation Studies. We train the model without feature loss $\mathcal{L}_{\mathrm{feat}}$ , without multi-hypothesis (MH) viewpoint prediction and without both. Without either component, the viewpoint predictions collapse to frontal views, resulting in unnatural shapes. + +component, the learned viewpoints collapse to only frontal views, resulting in unnaturally stretched reconstructions. Additional results are provided in the sup. mat. + +Although our model reconstructs highly detailed 3D shapes from only a single image and generalises to in-the-wild images, the predicted texture may not preserve sufficient details with a single forward pass and may require additional test-time finetuning. This can potentially be addressed by training on a larger dataset and incorporating recent advances in image generation [53,54]. Another limitation is the requirement of a pre-defined topology for articulation, which may vary between different animal species. Discovering the articulation structure automatically from raw in-the-wild images will be of great interest for future work. Failure cases are discussed in the sup. mat. + +# 5. Conclusions + +We have introduced a new model that can learn a 3D model of an articulated object category from single-view images taken in the wild. This model can, at test time, reconstruct the shape, articulation, albedo, and lighting of the object from a single image, and generalises to abstract drawings. Our approach demonstrates the power of combining several recent improvements in self-supervised representation learning together with a new viewpoint sampling scheme. We have shown significantly superior results compared to prior works, even when they use more supervision. + +Data Ethics. We use the DOVE, Weizmann Horse, PASCAL, Horse-10, and MS-COCO datasets in a manner compatible with their terms. Some of these images may accidentally contain faces or other personal information, but we do not make use of these images or image regions. For further details on ethics, data protection, and copyright please see https://www.robots.ox.ac.uk/~vedaldi/research/union/ethics.html. + +Acknowledgements. Shangzhe Wu is supported by a research gift from Meta Research. Andrea Vedaldi and Christian Rupprecht are supported by ERC-CoG UNION 101001212. Tomas Jakab and Christian Rupprecht are (also) supported by VisualAI EP/T028572/1. + +# References + +[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. In ECCV Workshops, 2022. 4, 6 +[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape completion and animation of people. ACM TOG, 2005. 1 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2, 3 +[4] Eran Borenstein and Shimon Ullman. Class-specific, top-down segmentation. In ECCV, 2002. 6 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3, 4, 6 +[6] John E Chadwick, David R Haumann, and Richard E Parent. Layered construction for deformable animated characters. ACM SIGGRAPH Computer Graphics, 1989. 4 +[7] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2 +[8] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, 2021. 2, 3 +[9] Akio Doi and Akio Koide. An efficient method of triangulating equi-valued surfaces by using tetrahedral cells. IEICE Transactions on Information and Systems, 1991. 3 +[10] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. *IJCV*, 2015. 6, 8 +[11] Olivier Faugeras. Three-Dimensional Computer Vision: A Geometric Viewpoint. MIT Press, 1993. 3 +[12] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In CVPR, 2021. 3 +[13] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In CVPR, 2022. 2, 3 +[14] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoints without keypoints. In ECCV, 2020. 1, 2, 3, 5, 7, 8 +[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 4 +[16] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 3 +[17] Paul Henderson and Vittorio Ferrari. Learning single-image 3d reconstruction by generative modelling of shape, pose and shading. IJCV, 2019. 2 + +[18] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. SCOPS: Self-supervised co-part segmentation. In CVPR, 2019. 2, 7 +[19] Wonbong Jang and Lourdes Agapito. CodeNeRF: Disentangled neural radiance fields for object categories. In ICCV, 2021. 3 +[20] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1 +[21] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 1, 2, 3, 5, 7, 8 +[22] Abhishek Kar, Shubham Tulsiani, João Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In CVPR, 2015. 1, 2 +[23] Hiroharu Kato and Tatsuya Harada. Learning view priors for single-view 3d reconstruction. In CVPR, 2019. 2 +[24] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2 +[25] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. PointRend: Image segmentation as rendering. In CVPR, 2020. 3, 6 +[26] Filippos Kokkinos and Iasonas Kokkinos. Learning monocular 3d reconstruction of articulated categories from motion. In CVPR, 2021. 2 +[27] Filippos Kokkinos and Iasonas Kokkinos. To the point: Correspondence-driven monocular 3d category reconstruction. In NeurIPS, 2021. 2 +[28] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 4, 6 +[29] Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2, 8 +[30] Nilesh Kulkarni, Abhinav Gupta, and Shubham Tulsiani. Canonical surface mapping via geometric cycle consistency. In ICCV, 2019. 2 +[31] Xueting Li, Sifei Liu, Shalini De Mello, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Online adaptation for consistent mesh reconstruction in the wild. In NeurIPS, 2020. 1, 2, 3 +[32] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In ECCV, 2020. 2, 7, 8 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6 +[34] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In ICCV, 2019. 2 + +[35] Yebin Liu, Juergen Gall, Carsten Stoll, Qionghai Dai, Hans-Peter Seidel, and Christian Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE TPAMI, 2013. 1 +[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 2015. 1, 3 +[37] Nadia Magnenat-Thalmann, E Primeau, and Daniel Thalmann. Abstract muscle action procedures for human face animation. The Visual Computer, 1988. 4 +[38] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W. Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In WACV, 2021. 6 +[39] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 3 +[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3 +[41] Tom Monnier, Matthew Fisher, Alexei A. Efros, and Mathieu Aubry. Share With Thy Neighbors: Single-View Reconstruction by Cross-Instance Consistency. In ECCV, 2022. 2 +[42] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2021. 2, 3 +[43] Kieran A Murphy, Carlos Esteves, Varun Jampani, Srikumar Ramalingam, and Ameesh Makadia. Implicit-PDF: Nonparametric representation of probability distributions on the rotation manifold. In ICML, 2021. 3 +[44] Natalia Neverova, David Novotny, and Andrea Vedaldi. Continuous surface embeddings. In NeurIPS, 2020. 3, 5 +[45] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3d representations from natural images. In ICCV, 2019. 2 +[46] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV, 2021. 3 +[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 3 +[48] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3 +[49] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2 +[50] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In CVPR, 2020. 3 + +[51] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. PVA: Pixel-aligned volumetric avatars. In CVPR, 2021. 3 +[52] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 2, 3 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 8 +[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 8 +[55] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2 +[56] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In NeurIPS, 2021. 2, 3, 4 +[57] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, 2007. 3 +[58] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 3 +[59] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 3 +[60] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representation. In 3DV, 2022. 5 +[61] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In CVPR, 2022. 4 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5 +[63] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2, 3 +[64] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6, 8 +[65] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2 +[66] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3 + +[67] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. DOVE: Learning deformable 3d objects by watching videos. arXiv preprint arXiv:2107.10844, 2021. 2, 3, 6, 7, 8 +[68] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In CVPR, 2021. 2, 5 +[69] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3D objects from images in the wild. In CVPR, 2020. 2 +[70] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T. Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In CVPR, 2021. 2, 3 +[71] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. ViSER: Video-specific surface embeddings for articulated 3d shape reconstruction. In NeurIPS, 2021. 3 +[72] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. BANMo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 3, 5 +[73] Chun-Han Yao, Wei-Chih Hung, Michael Rubinstein, Yuanzhen Lee, Varun Jampani, and Ming-Hsuan Yang. Lassie: Learning articulated shape from sparse image ensemble via 3d part discovery. In NeurlPS, 2022. 3 +[74] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In CVPR, 2020. 3 +[75] Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. Rel-Pose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 3 +[76] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 3 \ No newline at end of file diff --git a/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/images.zip b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18c7a2deda88b01f23d3ba2fb0adc8cd93b73f35 --- /dev/null +++ b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3ffd7b0cbc0077a4de0edff637537e83d6275d15785e1a41bfc6cd415467c66 +size 439480 diff --git a/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/layout.json b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f3e8c279b03ec51c95db40f5ba3ca386bc4c565b --- /dev/null +++ b/2023/MagicPony_ Learning Articulated 3D Animals in the Wild/layout.json @@ -0,0 +1,12241 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "type": "text", + "content": "MagicPony: Learning Articulated 3D Animals in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 143, + 509, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 143, + 509, + 176 + ], + "spans": [ + { + "bbox": [ + 85, + 143, + 509, + 176 + ], + "type": "text", + "content": "Shangzhe Wu* Ruining Li* Tomas Jakab* Christian Rupprecht Andrea Vedaldi Visual Geometry Group, University of Oxford" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 150, + 177, + 443, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 177, + 443, + 203 + ], + "spans": [ + { + "bbox": [ + 150, + 177, + 443, + 203 + ], + "type": "text", + "content": "{szwu, ruining, tomj, chrisr,VEDaldi}@robots.ox.ac.uk 3dmagicpony.github.io" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 209, + 533, + 326 + ], + "blocks": [ + { + "bbox": [ + 50, + 209, + 533, + 326 + ], + "lines": [ + { + "bbox": [ + 50, + 209, + 533, + 326 + ], + "spans": [ + { + "bbox": [ + 50, + 209, + 533, + 326 + ], + "type": "image", + "image_path": "0181a9c1d5c0e7932ba2ebc5e93123a1a020bf44e37b64c52f3f959d9833d0b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 329, + 546, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 329, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 546, + 353 + ], + "type": "text", + "content": "Figure 1. Learning Articulated 3D Animals in the Wild. Our method trains on a collection of single-view images of an animal category and produces a model that can predict the articulated 3D shape of a new instance from a single test image, which can be animated and relit." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 363, + 192, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 363, + 192, + 376 + ], + "spans": [ + { + "bbox": [ + 143, + 363, + 192, + 376 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 388, + 290, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 290, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 290, + 617 + ], + "type": "text", + "content": "We consider the problem of predicting the 3D shape, articulation, viewpoint, texture, and lighting of an articulated animal like a horse given a single test image as input. We present a new method, dubbed MagicPony, that learns this predictor purely from in-the-wild single-view images of the object category, with minimal assumptions about the topology of deformation. At its core is an implicit-explicit representation of articulated shape and appearance, combining the strengths of neural fields and meshes. In order to help the model understand an object's shape and pose, we distil the knowledge captured by an off-the-shelf self-supervised vision transformer and fuse it into the 3D model. To overcome local optima in viewpoint estimation, we further introduce a new viewpoint sampling scheme that comes at no additional training cost. MagicPony outperforms prior work on this challenging task and demonstrates excellent generalisation in reconstructing art, despite the fact that it is only trained on real images. The code can be found on the project page at https://3dmagicpony.github.io/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "type": "text", + "content": "Reconstructing the 3D shape of an object from a single image of it requires knowing a priori what are the possible shapes and appearances of the object. Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 365, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 497 + ], + "type": "text", + "content": "such a prior usually requires ad-hoc data acquisition setups [2, 20, 35, 36], involving at least multiple cameras, and often laser scanners, domes and other hardware, not to mention significant manual effort. This is viable for certain types of objects such as humans that are of particular interest in applications, but it is unlikely to scale to the long tail of objects that can appear in natural images. The alternative is to learn a 3D prior from 2D images only, which are available in abundance. However, this prior is highly complex and must be learned while using it to reconstruct in 3D the 2D training data, which is a major challenge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "type": "text", + "content": "In this paper, we propose MagicPony, a novel approach to learning 3D models of articulated object categories such as horses and birds with only single-view input images for training. We leverage recent progress in unsupervised representation learning, unsupervised image matching, efficient implicit-explicit shape representations and neural rendering, and devise a new auto-encoder architecture that reconstructs the 3D shape, articulation and texture of each object instance from a single image. For training, we only require a 2D segmenter for the object category and a description of the topology and symmetry of its 3D skeleton (i.e., the number and connectivity of bones). We do not require apriori knowledge of the objects' 3D shapes, keypoints, viewpoints, or of any other 2D or 3D cue which are often used in prior work [14, 21, 22, 31]. From this, we learn a function that, at test time, can estimate the shape and texture of a new object from a single image, in a feedforward manner. The function exhibits remarkable gener" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "8792" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": "alisation properties, including reconstructing objects in abstract drawings, despite being trained on real images only." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 286, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 286, + 277 + ], + "type": "text", + "content": "In order to learn such disentangled 3D representations simply from raw images, MagicPony addresses a few key challenges. The first challenge is viewpoint estimation. Before the 3D model is available, it is very difficult to assign a viewpoint to the objects in the training images. To tackle this problem, prior works have often assumed additional cues such as 2D keypoint correspondences [21, 22, 31]. Instead, we avoid requiring additional keypoint supervision and implicitly infer noisy correspondences by fusing into the 3D model knowledge distilled from DINO-ViT [5], a self-supervised visual transformer network (ViT) [28]. We also develop a new efficient disambiguation scheme that explores multiple viewpoint assignment hypotheses at essentially no cost, avoiding local optima that are caused by greedily matching the noisy 2D correspondences." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 286, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 286, + 518 + ], + "type": "text", + "content": "The second challenge is how to represent the 3D shape, appearance and deformations of the object. Most prior works have used textured meshes [13, 14, 21, 31, 32, 67, 70], but these are difficult to optimise from scratch, leading to problems that often require ad-hoc heuristics such as remeshing [13, 70]. The other and increasingly popular approach is to use a volumetric representation such as a neural radiance field [3, 40, 52, 63], which can model complex shapes, including manipulating their topology during training. However, this modelling freedom comes at the cost of over-parametrisation, which is particularly problematic in monocular reconstruction and often leads to meaningless short-cut solutions [63]. Furthermore, modelling articulation with a volumetric representation is difficult. A posing transformation is only defined for the object's surface and interior and is more easily expressed from the canonical/pose-free space to the posed space. However, rendering a radiance field requires transforming 3D points off the object surface and in the direction opposite to the posing transformation, which is hard [8]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 520, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 286, + 604 + ], + "type": "text", + "content": "We address these issues by using a hybrid volumetric-mesh representation based on DMTet [42, 56]. Shape and appearance are defined volumetrically in canonical space, but a mesh is extracted on the fly for posing and rendering. This sidesteps the challenges of using neural rendering directly while retaining most of its advantages and enabling the use of powerful shape regularisers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "To summarise, we make the following contributions: (1) A new 3D object learning framework that combines recent advances in unsupervised learning, 3D representations and neural rendering, achieving better reconstruction results with less supervision; (2) An effective mechanism for fusing self-supervised features from DINO-ViT into the 3D model as a form of self-supervision; and (3) an efficient multi-hypothesis viewpoint prediction scheme that avoids local optima in reconstruction with no additional cost." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 145, + 541, + 276 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 137 + ], + "type": "text", + "content": "Table 1. Related Work Overview on Weakly-supervised Learning of 3D Objects. Annotations: template shape, viewpoint, 2D keypoint, object mask, optical flow, video, coarse template shape from keypoints, camera estimated from keypoints using SfM, outputs texture flow, shape bases initialised from CMR. UMR relies on part segmentations from SCOPS [18]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 145, + 541, + 276 + ], + "lines": [ + { + "bbox": [ + 309, + 145, + 541, + 276 + ], + "spans": [ + { + "bbox": [ + 309, + 145, + 541, + 276 + ], + "type": "table", + "html": "
MethodSupervisionOutput
3D2.5DMotion ViewTexture
Unsup3D [69]✓✓✓
CSM [30]✓✓✓
A-CSM [29]✓✓✓✓
CMR [21](✓)1(✓)2✓✓✓✓(✓)3
U-CMR [14]✓✓✓✓✓
UMR†[32]✓✓✓(✓)3
ACMR [31](✓)4(✓)2✓✓✓✓(✓)3
DOVE [67]✓✓✓✓✓✓
Ours✓✓✓✓✓
", + "image_path": "835ecf62d14f539f46d9cccb31c30a3f706f676b3cb44bbd135276b51f94ba2d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 296, + 545, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 296, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 305, + 296, + 545, + 356 + ], + "type": "text", + "content": "We compare our method to prior work on several challenging articulated animal categories and show that our method obtains significantly better quantitative and qualitative results while using significantly less supervision, and also demonstrate generalisation to abstract drawings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 378, + 392, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 392, + 391 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 392, + 391 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 402, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 713 + ], + "type": "text", + "content": "Weakly-supervised Learning of 3D Objects. Most related to our work are weakly-supervised methods for single-image 3D object reconstruction that learn from a collection of images or videos of an object category [14, 17, 21, 23, 24, 26, 27, 31, 32, 34, 65, 67, 68]. Due to the inherent ambiguity of the problem, these methods usually rely on heavy geometric supervision (see Tab. 1) in addition to object masks, such as 2D keypoints [21, 31], template shapes [14, 29, 30]. UMR [32] forgoes the requirement of keypoints and template shapes by leveraging weakly-supervised object part segmentations (SCOPS [18]), but produces coarse symmetric shapes similar to CMR [21]. DOVE [67] learns coarse articulated objects by exploiting temporal information from videos with optical flow supervision. All these methods require object masks for supervision, except Unsup3D [69] which exploits bilateral symmetry for reconstructing roughly frontal objects, like faces, and UNICORN [41] which uses a progressive conditioning strategy with a heavily constrained bottleneck, leading to coarse reconstructions. Another emerging paradigm is to leverage generative models [7, 45, 49, 55] that encourage images rendered from viewpoints sampled from a prior distribution to be realistic. These models, however, often rely heavily on an accurate estimation of viewpoint distribution and/or a powerful 2D image generative model, both of which are difficult to obtain in small-data scenarios." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8793" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "Optimization from Multi-views and Videos. 3D reconstruction traditionally relies on epipolar geometry of multiview images of static scenes [11, 16]. Neural Radiance Fields (NeRF) [3, 40, 52, 63] have recently emerged as a powerful volumetric representation for multi-view reconstruction given accurate cameras. A recent line of work, LASR [70], ViSER [71] and BANMo [72], optimises 3D shapes of deformable objects from a small set of monocular videos, with heavily engineered optimisation strategies using optical flow and mask supervision, and additionally DensePose [44] annotations for highly deformable animals. Concurrent work of LASSIE [73] also leverages DINO-ViT [5] image features for supervision but optimises a part-based model on a small collection of images (" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "\\sim 30" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": ") of an object category. It uses a shared shape model for all instances with only per-instance articulation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 272, + 289, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 289, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 289, + 426 + ], + "type": "text", + "content": "Learnable Deformable 3D Representations. A common 3D representation for weakly-supervised shape learning is triangular meshes. Deformation on meshes can be modelled by estimating offsets of individual vertices [21, 31, 67], which typically requires heavy regularisation, such as As-Rigid-As-Possible (ARAP) [57]. To constrain the space of deformations, many works often utilise lower-dimensional models, such as cages [74] or linear blend skinning controlled with skeletal bones [36, 67] or Gaussian control points [70, 71]. Parametric models, like SMPL [36] and SMAL [76], allow for realistic control of the shape through learned parameters but often require an extensive collection of 3D scans to learn from." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 430, + 289, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 586 + ], + "type": "text", + "content": "Dynamic neural fields [12,46,48,50,51,58,59,72], such as D-NeRF [50], extend NeRF [40] with time-varying components. A-NeRF [58] proposes a radiance field conditioned on an articulated skeleton, while BANMo [72] uses learned Gaussian control points similarly to [70, 71]. One key limitation of these implicit representations is the requirement of an inverse transformation from 3D world coordinates back to the canonical space, which is often harder to learn compared to forward deformation [8]. We propose a hybrid SDF-mesh representation, extending DMTet [56] with an articulation mechanism, which combines the expressiveness of implicit models with the simplicity of mesh deformation and articulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Viewpoint Prediction in the Wild. Viewpoint prediction in weakly-supervised settings is challenging, as it is prone to local optima induced by projection and common object symmetries. Existing solutions are based on learning a distribution over multiple possible viewpoints. U-CMR [14] proposes a camera-multiplex that optimises over 40 cameras for each training sample using a given template shape, which is expensive as evaluating each hypothesis involves a rendering step. We propose a new viewpoint exploration scheme that requires evaluating only one hypothesis in each" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "step with essentially no added training cost and does not rely on a given template shape but a jointly learned model. Recent work of Implicit-PDF [43] proposes to learn viewpoint distributions implicitly by estimating a probability for each image-pose pair of a fixed object instance. RelPose [75] extends this to learning relative poses of image pairs. Our model also predicts a probability associated with each viewpoint hypothesis but instead relates an instance to a learned category-level prior model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 190, + 362, + 203 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 190, + 362, + 203 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 362, + 203 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 210, + 547, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 547, + 270 + ], + "type": "text", + "content": "Given only a collection of single-view images of a deformable object category collected in the wild, our goal is to train a monocular reconstruction network that, at test time, predicts the 3D shape, articulated pose, albedo and lighting from a single image of the object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 270, + 547, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 547, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 547, + 402 + ], + "type": "text", + "content": "During training, our method requires no geometric supervision other than the images, 2D foreground segmentations of the object obtained automatically from an off-the-shelf method like [25], and a description of the topology of the object's skeleton (e.g., the number of legs). It is based on several key ideas: an implicit-explicit shape representation (Sec. 3.1), a hierarchical shape representation, from a generic template to articulated instances (Sec. 3.2), and a viewpoint prediction scheme that leverages self-supervised correspondences and efficient multi-hypothesis exploration to avoid local minima (Sec. 3.4), as illustrated in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 407, + 528, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 528, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 528, + 421 + ], + "type": "text", + "content": "3.1. Implicit-Explicit 3D Shape Representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "type": "text", + "content": "The choice of representation for the 3D shape of the object is critical. It must be (1) sufficiently expressive for modelling fine-grained shape details and deformations and (2) sufficiently regular for learning with only weak supervision via image reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 487, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 654 + ], + "type": "text", + "content": "Most 3D representations are either explicit (triangular meshes [21, 67, 70]) or implicit (volumes, SDFs, radiance fields [19, 39, 40, 47, 66]). Explicit representations are more compact, easily support powerful deformation models like blend skinning, and the geometric smoothness can be easily measured and controlled; however, they are more difficult to optimise and often result in defects like folds during training, which requires occasional re-meshing [13, 70]. Implicit representations are more amenable to gradient-based optimisation, including allowing topological changes, but are more computationally expensive and, due to their high capacity, require substantial supervision, usually in the form of a dense coverage of the different object viewpoints [40, 66] or 3D ground-truth [39, 47]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Here we adopt a representation that combines the advantages of both. The shape of the object is represented implicitly by a neural field that is converted on the fly into an explicit mesh via the Differentiable Marching Tetrahedra (DMTet) [9, 42, 56] method. Specifically, the object's shape" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "type": "text", + "content": "8794" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 213 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 213 + ], + "type": "image", + "image_path": "f95d58eb7b60e8aa8d832efa4be77d292fcbe6ba108adc2199c5dfdc6c85839e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 277 + ], + "type": "text", + "content": "Figure 2. Training Pipeline. Given a collection of single-view images of an animal category, our model learns a category-specific prior shape using an implicit-explicit representation, together with a feature field that allows us to fuse self-supervised correspondences through a feature rendering loss. This prior shape is deformed, articulated and shaded based on features extracted from a training image. To combat local minima in viewpoint prediction, we introduce an efficient scheme that explores multiple hypotheses at essentially no additional cost. The entire pipeline is trained with reconstruction losses end-to-end, except for the frozen image encoder pre-trained via self-supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "inline_equation", + "content": "S = \\{\\pmb{x} \\in \\mathbb{R}^3 | s(\\pmb{x}) = 0\\}" + }, + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "text", + "content": " is defined as the isosurface of the signed distance function (SDF) " + }, + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "inline_equation", + "content": "s: \\mathbb{R}^3 \\mapsto \\mathbb{R}" + }, + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "text", + "content": ", which is implemented by a Multi-Layer Perceptron (MLP) taking 3D coordinates as input. To encourage " + }, + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "text", + "content": " to be a signed distance function, we sample random points " + }, + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 297, + 289, + 369 + ], + "type": "text", + "content": " in space and minimise the Eikonal regulariser [15]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 374, + 287, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 374, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 88, + 374, + 287, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\mathrm {E i k}} (s) = \\frac {1}{| \\mathcal {X} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {X}} \\left(\\left\\| \\nabla s (\\boldsymbol {x}) \\right\\| _ {2} - 1\\right) ^ {2}. \\tag {1}", + "image_path": "46dcf7e73f58c044cf52f8d3f8a47ac7f5154bf2b6ff4c7b77bee8654e767b3b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 409, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 444 + ], + "type": "text", + "content": "We then evaluate the function " + }, + { + "bbox": [ + 46, + 409, + 287, + 444 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 409, + 287, + 444 + ], + "type": "text", + "content": " at the vertices of a tetrahedral grid and extract a watertight mesh from it using DMTet [56], which is fast and differentiable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "text", + "content": "We use this implicit-explicit representation to learn a category-wise template shape shared across all instances, and model instance deformations through a hierarchical pipeline, explained next. The function " + }, + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "text", + "content": " is initialised with a generic ellipsoid elongated along the " + }, + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "text", + "content": "-axis. We denote by " + }, + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{pr}} \\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "text", + "content": " the vertices of the extracted prior mesh." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 522, + 277, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 277, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 277, + 536 + ], + "type": "text", + "content": "3.2. Articulated Shape Modelling and Prediction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "type": "text", + "content": "The category-wise template " + }, + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{pr}}" + }, + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "type": "text", + "content": " is deformed into a specific object instance " + }, + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ins}}" + }, + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "type": "text", + "content": ", and further posed to match the shape " + }, + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 540, + 287, + 590 + ], + "type": "text", + "content": " in each input image (Fig. 2). These deformations are predicted from a single image, as discussed below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Image Encoder. The input image is first encoded by means of DINO-ViT [5], a self-supervised Vision Transformer (ViT) [28]. The ViT extracts from the image a set of key and output tokens, which were found to code, respectively, for correspondences [1] and appearance [61]. Hence, we use the keys to predict the object's deformation and the outputs to predict its appearance. Specifically, we apply the ViT " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " to the input image " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "I\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " and extract the key " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{k}}\\in \\mathbb{R}^{D\\times H_p\\times W_p}" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " and output " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{o}}\\in \\mathbb{R}^{D\\times H_p\\times W_p}" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " tensors from the last attention layer. These features are local," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 297, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 335 + ], + "type": "text", + "content": "with one token per image patch. We use two small convolutional networks to further extract from these a global key " + }, + { + "bbox": [ + 305, + 297, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{k}} = f_{\\mathrm{k}}(\\Phi_{\\mathrm{k}})\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 305, + 297, + 545, + 335 + ], + "type": "text", + "content": " and output " + }, + { + "bbox": [ + 305, + 297, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{o}} = f_{\\mathrm{o}}(\\Phi_{\\mathrm{o}})\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 305, + 297, + 545, + 335 + ], + "type": "text", + "content": " vector." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": "Instance-specific Shape. The shape of each object instance " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ins}}" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": " differs slightly from the template shape " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{pr}}" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": ". This is accounted for by the deformation " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ins},i} = V_{\\mathrm{pr},i} + \\Delta V_i" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": ", where the displacement " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "\\Delta V_i = f_{\\Delta V}(V_{\\mathrm{pr},i},\\phi_{\\mathrm{k}})" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": " of each vertex " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{pr},i}" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": " is predicted by an MLP " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "f_{\\Delta V}" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": " from the global image features " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{k}}" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": ". We exploit the fact that many objects are bilaterally symmetric and enforce both the prior " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{pr}}" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": " and instance " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ins}}" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": " shapes to be symmetric by mirroring the query locations for the MLPs about the " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "yz" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": "-plane. We also limit the magnitude of the deformation " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ins}}" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": " by adding the regularizer " + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{def}}(\\Delta V) = \\frac{1}{|V_{\\mathrm{pr}}|}\\sum_i\\| \\Delta V_i\\| _2^2" + }, + { + "bbox": [ + 305, + 338, + 546, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "content": "Posed Shape. In order to match the 3D shape observed in the image, the instance-specific mesh " + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ins}}" + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "content": " is further deformed by the posing function " + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "inline_equation", + "content": "V_{i} = g(V_{\\mathrm{ins},i},\\xi)" + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "content": " denotes the pose parameters. Posing accounts for the deformations caused by the articulation of the object's skeleton; they allow for large but controlled deformations which would otherwise be more difficult to account for by a generic deformation field like " + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\Delta V" + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": "We adopt the blend skinning model for posing [6, 37], parameterised by " + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "inline_equation", + "content": "B - 1" + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": " bone rotations " + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "inline_equation", + "content": "\\xi_{b} \\in SO(3), b = 2, \\ldots, B" + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": ", and the viewpoint " + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "inline_equation", + "content": "\\xi_{1} \\in SE(3)" + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": ". Specifically, we initialise a set of rest-pose joint locations " + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{b}" + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": " on the instance mesh using simple heuristics1. Each bone " + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": " but the root has one parent " + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "inline_equation", + "content": "\\pi(b)" + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": ", thus forming a tree structure. Each vertex " + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": " is associated to the bones by the skinning weights " + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "inline_equation", + "content": "w_{ib}" + }, + { + "bbox": [ + 304, + 575, + 547, + 672 + ], + "type": "text", + "content": ", defined based on relative proximity to each bone (see" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "text", + "content": "1A chain of bones going through the two farthest end points along " + }, + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "text", + "content": "-axis for birds, and for quadrupeds, additionally four legs branching out from the body bone to the lowest point in each quadrant. See sup. mat." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "text", + "content": "8795" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "sup. mat.for details). The vertices are then posed by the linear blend skinning equation:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 103, + 287, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 287, + 139 + ], + "type": "interline_equation", + "content": "V _ {i} (\\xi) = \\left(\\sum_ {b = 1} ^ {B} w _ {i b} G _ {b} (\\xi) G _ {b} \\left(\\xi^ {*}\\right) ^ {- 1}\\right) V _ {\\text {i n s}, i}, \\tag {2}", + "image_path": "fb38342b9667198a64b2d439e0e4ed6e2e5905585ff69349945651d9c445045e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 266, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 266, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 266, + 166 + ], + "type": "interline_equation", + "content": "G _ {1} = g _ {1}, G _ {b} = G _ {\\pi (b)} \\circ g _ {b}, g _ {b} (\\boldsymbol {\\xi}) = \\left[ \\begin{array}{c c} R _ {\\xi_ {b}} & \\mathbf {J} _ {b} \\\\ 0 & 1 \\end{array} \\right],", + "image_path": "071cdd49144fab81585d93ef3442d1c17dfc71ca8d572ffa958185d9ba62ead2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 171, + 258, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 258, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 258, + 184 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 171, + 258, + 184 + ], + "type": "inline_equation", + "content": "\\xi^{*}" + }, + { + "bbox": [ + 47, + 171, + 258, + 184 + ], + "type": "text", + "content": " denotes the bone rotations at the rest pose." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "text", + "content": "Pose Prediction. Similarly to the instance-specific deformation " + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\Delta V_{i}" + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "text", + "content": ", the bone rotations " + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\xi_{2:B}" + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "text", + "content": " are predicted by a network " + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{b}}" + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "text", + "content": " from the image features " + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{k}}" + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{k}}" + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "text", + "content": ". The viewpoint " + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\xi_{1} \\in SE(3)" + }, + { + "bbox": [ + 47, + 188, + 287, + 236 + ], + "type": "text", + "content": " is predicted separately, as discussed in Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": "Pose articulation depends on estimating local body joints and benefits from using local features. To extract them, we first project the centroid of each bone at rest into the image plane using the predicted viewpoint " + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "inline_equation", + "content": "\\xi_{1}" + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 291, + 230, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 291, + 230, + 318 + ], + "spans": [ + { + "bbox": [ + 104, + 291, + 230, + 318 + ], + "type": "interline_equation", + "content": "\\boldsymbol {u} _ {b} = \\Pi \\left[ G _ {\\pi (b)} \\left(\\xi_ {1}, \\xi_ {2: B} ^ {*}\\right) \\frac {\\mathbf {J} _ {b}}{2} \\right],", + "image_path": "42a50d766029f547273ef50ee4c53352b5678b299e32508d2909a470572352f1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 324, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 287, + 348 + ], + "type": "text", + "content": "and sample the local features from the patch key feature map " + }, + { + "bbox": [ + 47, + 324, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathbf{k}}(\\pmb{u}_b)" + }, + { + "bbox": [ + 47, + 324, + 287, + 348 + ], + "type": "text", + "content": " at the projected pixel locations " + }, + { + "bbox": [ + 47, + 324, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\pmb{u}_b" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "text", + "content": "To predict bone rotations from the image features, we adopt a transformer to account for the inter-dependency of the bones, implied by the underlying skeleton. Specifically, we use a small transformer network [62] " + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{b}}" + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "text", + "content": " that takes as input " + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "inline_equation", + "content": "B - 1" + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\nu_{b} = (\\phi_{\\mathrm{k}}, \\Phi_{\\mathrm{k}}(\\pmb{u}_{b}), b, \\mathbf{J}_{b}, \\pmb{u}_{b})" + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "text", + "content": " and outputs the bone rotations " + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "inline_equation", + "content": "(\\xi_{2}, \\ldots, \\xi_{B}) = f_{\\mathrm{b}}(\\nu_{2}, \\ldots, \\nu_{B})" + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "text", + "content": " parameterised as Euler angles. We also limit the deviation from the rest pose via the regulariser " + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{art}}(\\xi) = \\frac{1}{B - 1} \\sum_{b=2}^{B} \\| \\xi_{b} \\|_{2}^{2}" + }, + { + "bbox": [ + 46, + 348, + 287, + 446 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 450, + 188, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 188, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 188, + 464 + ], + "type": "text", + "content": "3.3. Appearance and Shading" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "content": "We decompose the appearance of the objects into albedo and diffuse shading assuming a Lambertian illumination model. The albedo of the object is modelled in the same \"canonical space\" as the shape SDF " + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "content": " by a second coordinate MLP " + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "inline_equation", + "content": "a(\\pmb{x}) = f_{\\mathrm{a}}(\\pmb{x}, \\phi_{\\mathrm{o}}) \\in [0,1]^3" + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "content": ", which takes as input the global output features " + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{o}}" + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "content": ". The dominant light direction " + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "inline_equation", + "content": "l \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "content": " and ambient and diffuse intensities " + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "inline_equation", + "content": "k_{s}, k_{d} \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "content": " are predicted by a lighting network " + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{l}}(\\phi_{\\mathrm{o}})" + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "text", + "content": "The posed mesh is then rendered into an image via differentiable deferred mesh rendering. The renderer first rasterizes all pixels " + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "text", + "content": " that overlap with the posed object " + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "text", + "content": ", computing the 3D coordinates " + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\pmb{x}(\\pmb{u})" + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "text", + "content": " of the corresponding object point in canonical space; then the MLP " + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}" + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "text", + "content": " is evaluated, once per pixel, to obtain the colour " + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "inline_equation", + "content": "a(\\pmb{x}(\\pmb{u}))" + }, + { + "bbox": [ + 45, + 564, + 287, + 649 + ], + "type": "text", + "content": ". The final pixel colour accounts for shading and is given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 655, + 267, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 655, + 267, + 670 + ], + "spans": [ + { + "bbox": [ + 67, + 655, + 267, + 670 + ], + "type": "interline_equation", + "content": "\\hat {I} (\\boldsymbol {u}) = \\left(k _ {s} + k _ {d} \\cdot \\max \\{0, \\langle l, n (\\boldsymbol {u}) \\rangle \\}\\right) \\cdot a (\\boldsymbol {x} (\\boldsymbol {u})),", + "image_path": "114eb80b4e24816b8590aff4bfec7e1e47ed00a918106df48f963af334efcf7b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "n(\\pmb{u}) \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " is the normal of the posed mesh " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ". Using the same differentiable renderer, we also obtain the predicted mask " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{M} \\in [0,1]^{H \\times W}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " of the object." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": "For training supervision, we use the appearance loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{im}}(\\hat{I}, I, M) = \\| \\tilde{M} \\odot (\\hat{I} - I)\\|_1" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " as well as the mask loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{m}}(\\hat{M}, M) = \\| \\hat{M} - M\\|_2^2 + \\lambda_{\\mathrm{dt}}(\\hat{M} \\odot \\mathrm{d}\\tau(M))" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": ", where the distance transform " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\tau(\\cdot)" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " is used to obtain better gradients [21, 68], " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\tilde{M} = \\hat{M} \\odot M" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " is the intersection of the rendered and ground-truth masks, and " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dt}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " is a weighing factor." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 154, + 428, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 154, + 428, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 154, + 428, + 167 + ], + "type": "text", + "content": "3.4. Viewpoint Prediction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 173, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 173, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 545, + 233 + ], + "type": "text", + "content": "Predicting the object viewpoint is a critical first step for learning 3D shapes but often difficult [14, 21]. Here, we contribute a method to do so robustly and efficiently, which exploits self-supervised correspondences in a multi-hypothesis prediction pipeline, as detailed next." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "text", + "content": "Fusing Self-Supervised Correspondences. Estimating explicit correspondences without supervision is difficult and brittle; hence, we propose to incorporate correspondence information implicitly and softly. Inspired by N3FF [60], we fuse information from the self-supervised ViT features into the 3D model by learning to render them. This is done by adding a coordinate network " + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\psi(\\boldsymbol{x}) \\in \\mathbb{R}^{D'}" + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "text", + "content": " that predicts a " + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "text", + "content": "-dimensional feature vector for each 3D point " + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "text", + "content": " in canonical space. This is similar to the albedo network " + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "text", + "content": ", except that it is shared across all instances of the same object category, and hence there is no dependency on the input image. These canonical features are then projected into a feature image " + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}_{\\mathrm{k}} \\in \\mathbb{R}^{H \\times W \\times D'}" + }, + { + "bbox": [ + 305, + 239, + 545, + 418 + ], + "type": "text", + "content": " using the same differentiable renderer used for the albedo, and trained to minimise a corresponding rendering loss:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 342, + 428, + 545, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 428, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 342, + 428, + 545, + 443 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f e a t}} \\left(\\hat {\\Phi} _ {\\mathrm {k}}, \\Phi_ {\\mathrm {k}} ^ {\\prime}, \\tilde {M}\\right) = \\| \\tilde {M} \\odot \\left(\\hat {\\Phi} _ {\\mathrm {k}} - \\Phi_ {\\mathrm {k}} ^ {\\prime}\\right) \\| _ {2} ^ {2}, \\tag {3}", + "image_path": "8ac5bd70c45001f20f86017891bf92a15fb29ff5cebb9341ddfbdf39337c6415.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{k}}^{\\prime} = \\mathrm{PCA}\\bigl (\\Phi_{\\mathrm{k}}\\bigr)" + }, + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "type": "text", + "content": " is a PCA-reduced (to " + }, + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "type": "inline_equation", + "content": "D^{\\prime} = 16" + }, + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "type": "text", + "content": " principal components) version of the original DINO-ViT features " + }, + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{k}}" + }, + { + "bbox": [ + 305, + 453, + 545, + 490 + ], + "type": "text", + "content": " for training efficiency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": "Multi-Hypothesis Viewpoint Prediction. Prior works have found that a major challenge in learning the object's viewpoint is the existence of multiple local optima in the reconstruction objective. Some have addressed this issue by sampling a large number of hypotheses for the viewpoint [14], but this is somewhat cumbersome and slow. Instead, we propose a scheme that explores multiple viewpoints statistically, but at each iteration samples only a single one, and thus comes \"for free\"." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "text", + "content": "We hence task a viewpoint network " + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{vp}}(\\phi_k)" + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "text", + "content": " to predict from " + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{k}}" + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "text", + "content": " four viewpoint rotation hypotheses " + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "inline_equation", + "content": "R_{k} \\in SO(3), k \\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "text", + "content": ", each in one of the four quadrants around the object, as illustrated in Fig. 2. The network also predicts a score " + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 305, + 603, + 545, + 664 + ], + "type": "text", + "content": " for each hypothesis, used to evaluate the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "type": "text", + "content": "2This is also similar to BANMo [72], but we use self-supervised features instead of supervised DensePose [44] ones." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "3For bilaterally symmetric animals, there are often four ambiguous orientations towards each of the four " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "xz" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "-quadrants in the canonical space." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "8796" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": "probability " + }, + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": " that hypothesis " + }, + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": " is the best of the four options as " + }, + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "inline_equation", + "content": "p_k \\propto \\exp(-\\sigma_k / \\tau)" + }, + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": " using softmax, where " + }, + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": " is a temperature parameter, gradually decreased during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "text", + "content": "The naive approach for learning " + }, + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "text", + "content": " is to sample multiple hypotheses and compare their reconstruction loss to determine which one is better. However, this is expensive as it requires rendering the model with all hypotheses. Instead, we suggest to sample a single hypothesis for each training iteration and simply train " + }, + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "text", + "content": " to predict the expected reconstruction loss " + }, + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_k" + }, + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "text", + "content": ", by minimising the objective" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 200, + 286, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 200, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 102, + 200, + 286, + 216 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {h y p}} \\left(\\sigma_ {k}, \\tilde {\\mathcal {L}}\\right) = \\left(\\sigma_ {k} - \\operatorname {s g} \\left[ \\tilde {\\mathcal {L}} _ {k} \\right]\\right) ^ {2}, \\tag {4}", + "image_path": "5ee44973daad59531ca2e53b74c3e6ee4685d8b5690a4d20e9e30e1987fe327b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}[\\cdot ]" + }, + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "type": "text", + "content": " is the stop-gradient operator, and " + }, + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_k" + }, + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "type": "text", + "content": " is the overall reconstruction loss under the sampled " + }, + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 224, + 286, + 262 + ], + "type": "text", + "content": "-th hypothesis, given below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 269, + 158, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 158, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 158, + 282 + ], + "type": "text", + "content": "3.5. Training Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 287, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 286, + 335 + ], + "type": "text", + "content": "Given a training sample " + }, + { + "bbox": [ + 46, + 287, + 286, + 335 + ], + "type": "inline_equation", + "content": "(I, M)" + }, + { + "bbox": [ + 46, + 287, + 286, + 335 + ], + "type": "text", + "content": ", comprising an image and the corresponding object mask, the training objective is computed in two steps. First, the viewpoint hypothesis " + }, + { + "bbox": [ + 46, + 287, + 286, + 335 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 287, + 286, + 335 + ], + "type": "text", + "content": " is sampled and the reconstruction loss is computed:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 344, + 233, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 344, + 233, + 358 + ], + "spans": [ + { + "bbox": [ + 99, + 344, + 233, + 358 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {L}} _ {k} = \\lambda_ {\\mathrm {i m}} \\mathcal {L} _ {\\mathrm {i m}} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {f}} \\mathcal {L} _ {\\text {f e a t}}.", + "image_path": "86f7238f6f053c35417a0760afca11e98683957c9ef8cfa9c7b9cce963b9bb15.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 368, + 286, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 286, + 392 + ], + "type": "text", + "content": "Then, the regularisation terms and the viewpoint hypothesis loss are added, to obtain the final objective:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 401, + 269, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 401, + 269, + 415 + ], + "spans": [ + { + "bbox": [ + 63, + 401, + 269, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = p _ {k} \\tilde {\\mathcal {L}} _ {k} + \\lambda_ {\\mathrm {E}} \\mathcal {R} _ {\\text {E i k}} + \\lambda_ {\\mathrm {d}} \\mathcal {R} _ {\\text {d e f}} + \\lambda_ {\\mathrm {a}} \\mathcal {R} _ {\\text {a r t}} + \\lambda_ {\\mathrm {h}} \\mathcal {L} _ {\\text {h y p}}.", + "image_path": "abf3345a306d545f516c7e59063937d0bd98a12d39be07c0265069f1ae790380.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 424, + 263, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 263, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 263, + 436 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 46, + 424, + 263, + 436 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{im}},\\lambda_{\\mathrm{m}},\\lambda_{\\mathrm{f}},\\lambda_{\\mathrm{E}},\\lambda_{\\mathrm{d}},\\lambda_{\\mathrm{a}},\\lambda_{\\mathrm{h}}" + }, + { + "bbox": [ + 46, + 424, + 263, + 436 + ], + "type": "text", + "content": " are balancing weights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "type": "text", + "content": "The gradient of " + }, + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "type": "text", + "content": " with respect to the model parameters is then computed and used to update them. Note that only one viewpoint sample " + }, + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "type": "text", + "content": " is taken for each gradient evaluation. To further improve the sampling efficiency, we sample the viewpoint based on the learned probability distribution " + }, + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "type": "inline_equation", + "content": "p_k^4" + }, + { + "bbox": [ + 46, + 437, + 286, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 507, + 127, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 127, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 127, + 520 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 527, + 286, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 286, + 599 + ], + "type": "text", + "content": "We conduct extensive experiments on a few animal categories, including horses, giraffes, zebras, cows and birds, and compare against prior work both qualitatively and quantitatively on standard benchmarks. We also show that our model trained on real images generalises to abstract drawings, demonstrating the power of unsupervised learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 607, + 92, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 92, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 92, + 618 + ], + "type": "text", + "content": "4.1.Data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 625, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 286, + 685 + ], + "type": "text", + "content": "For horses, we use the horse dataset from DOVE [67] containing 10.8k images extracted from YouTube, and supplement it with 541 additional images from three datasets for diversity: Weizmann Horse Database [4], PASCAL [10] and Horse-10 Dataset [38]. For giraffes, zebras and cows," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "we source images from Microsoft COCO Dataset [33] and keep the ones with little occlusion. Since these datasets are relatively small, we only use them to finetune the pretrained horse model. For birds, we combine the DOVE dataset [67] and CUB dataset [64] consisting of " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "57.9\\mathrm{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "11.7\\mathrm{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " images respectively. We use an off-the-shelf PointRend [25] detector to obtain segmentation masks, crop around the objects and resize them to " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": ". We follow the original train/test splits from CUB and DOVE, and randomly split the rest, resulting in " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "11.5\\mathrm{k} / 0.8\\mathrm{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " horse, " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "513 / 57" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " giraffe, " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "574 / 64" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " zebra, " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "719 / 80" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " cow and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "63.9\\mathrm{k} / 10.5\\mathrm{k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " bird images. We additionally collect roughly 100 horse images from the Internet to test generalisation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 236, + 408, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 408, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 408, + 247 + ], + "type": "text", + "content": "4.2. Technical Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": "The model is implemented using a total of 8 subnetworks. The feature field " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": ", template SDF " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": ", albedo field " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " and deformation field " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "f_{\\Delta V}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " are implemented using MLPs, which take in 3D coordinates (concatenated with image features for albedo and deformation networks) as input. The image encoder consists of a ViT-S [28] architecture from DINO [5] and two convolutional encoders " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{k}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{o}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " that fuse the patch features into global feature vectors. We use self-supervised pre-trained DINO-ViT with frozen weights and only train the convolutional encoders. The lighting network " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " is a simple MLP that maps global output feature to a 4-channel vector, and the articulation network " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " is a transformer architecture with 4 blocks. We use a separate encoder for the viewpoint network " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{vp}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " identical to " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{k}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": ", as we find empirically that sharing the encoders tends to make the viewpoint learning unstable. Apart from DINO encoder, all components are trained end-to-end from scratch for 100 and 10 epochs on horses and birds respectively, with the instance deformation and articulation disabled for the first 30 and 2 epochs. The DINO features " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{k}}^{\\prime}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " used to compute " + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 542 + ], + "type": "text", + "content": " are pre-computed with patch size 8 and stride 4 and masked, following [1], and reduced from a channel size of 384 to 16 using PCA. All technical details and hyper-parameters are included in the supplementary material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 551, + 417, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 551, + 417, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 551, + 417, + 563 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "content": "Fig. 3 shows a few reconstructions of horses and birds produced by our model. Given a single 2D image at test time, our model predicts a textured 3D mesh of the object, capturing its fine-grained geometric details, such as the legs and tails of the horses. Our model predicts the articulated pose of the objects, allowing us to easily transfer the pose of one instance from another and animate it in 3D. Although our model is trained on real images only, it demonstrates excellent generalisation to paintings and abstract drawings." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Fig. 4 also shows that by finetuning the horse model on other animals without any additional modifications (except disabling articulation for the initial 5k iterations), it gen" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "text", + "content": "4We sample viewpoint " + }, + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "k^{*} = \\mathrm{argmax}_{k}p_{k}" + }, + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "text", + "content": " of the time for training efficiency, and uniformly at random " + }, + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "text", + "content": " of the time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8797" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 88, + 110 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 88, + 110 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 88, + 110 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 88, + 110 + ], + "type": "image", + "image_path": "1ec368feec8df80a847117319297db59b7b4a027c6f3805173a33f18227606ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 110, + 88, + 148 + ], + "blocks": [ + { + "bbox": [ + 49, + 110, + 88, + 148 + ], + "lines": [ + { + "bbox": [ + 49, + 110, + 88, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 110, + 88, + 148 + ], + "type": "image", + "image_path": "bb8813596399eb89c40ae684a7e226a45291efa9307381f7db819f6e1653f7eb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 148, + 87, + 187 + ], + "blocks": [ + { + "bbox": [ + 49, + 148, + 87, + 187 + ], + "lines": [ + { + "bbox": [ + 49, + 148, + 87, + 187 + ], + "spans": [ + { + "bbox": [ + 49, + 148, + 87, + 187 + ], + "type": "image", + "image_path": "1a9df4288b5199457caeed34396a0538d9dd189de62c921b7ca1eab9682ffe67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 194, + 86, + 224 + ], + "blocks": [ + { + "bbox": [ + 50, + 194, + 86, + 224 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 86, + 224 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 86, + 224 + ], + "type": "image", + "image_path": "03093c552a79df8d1f58122bef9295c15765023b27f0cb81323f3c7ba0a48e2d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 228, + 78, + 236 + ], + "lines": [ + { + "bbox": [ + 60, + 228, + 78, + 236 + ], + "spans": [ + { + "bbox": [ + 60, + 228, + 78, + 236 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 93, + 79, + 276, + 105 + ], + "blocks": [ + { + "bbox": [ + 93, + 79, + 276, + 105 + ], + "lines": [ + { + "bbox": [ + 93, + 79, + 276, + 105 + ], + "spans": [ + { + "bbox": [ + 93, + 79, + 276, + 105 + ], + "type": "image", + "image_path": "397fca09d3125ae645c61891f93dfc230a730c08307272dab17d664b813e7b3c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 123, + 323, + 130 + ], + "lines": [ + { + "bbox": [ + 306, + 123, + 323, + 130 + ], + "spans": [ + { + "bbox": [ + 306, + 123, + 323, + 130 + ], + "type": "text", + "content": "CMR" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 93, + 116, + 278, + 144 + ], + "blocks": [ + { + "bbox": [ + 93, + 116, + 278, + 144 + ], + "lines": [ + { + "bbox": [ + 93, + 116, + 278, + 144 + ], + "spans": [ + { + "bbox": [ + 93, + 116, + 278, + 144 + ], + "type": "image", + "image_path": "6b3b7b34dfe888196b20ad9f413a19362d910b9eafcf343a0597c583f1b52a0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 93, + 148, + 277, + 185 + ], + "blocks": [ + { + "bbox": [ + 93, + 148, + 277, + 185 + ], + "lines": [ + { + "bbox": [ + 93, + 148, + 277, + 185 + ], + "spans": [ + { + "bbox": [ + 93, + 148, + 277, + 185 + ], + "type": "image", + "image_path": "6e94fa1a600a8fbfd8492aa1bea4089827bc9d6bee4978566141d7d94fd384da.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 91, + 193, + 277, + 224 + ], + "blocks": [ + { + "bbox": [ + 91, + 193, + 277, + 224 + ], + "lines": [ + { + "bbox": [ + 91, + 193, + 277, + 224 + ], + "spans": [ + { + "bbox": [ + 91, + 193, + 277, + 224 + ], + "type": "image", + "image_path": "77d30e02660a2111851c3dc9d474eeed941080499348a35fdd5f87551f75430f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 228, + 175, + 236 + ], + "lines": [ + { + "bbox": [ + 124, + 228, + 175, + 236 + ], + "spans": [ + { + "bbox": [ + 124, + 228, + 175, + 236 + ], + "type": "text", + "content": "Reconstructions" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 228, + 231, + 236 + ], + "lines": [ + { + "bbox": [ + 205, + 228, + 231, + 236 + ], + "spans": [ + { + "bbox": [ + 205, + 228, + 231, + 236 + ], + "type": "text", + "content": "Textured" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 228, + 274, + 236 + ], + "lines": [ + { + "bbox": [ + 244, + 228, + 274, + 236 + ], + "spans": [ + { + "bbox": [ + 244, + 228, + 274, + 236 + ], + "type": "text", + "content": "Animated" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 201, + 194, + 235, + 224 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 194, + 235, + 224 + ], + "spans": [ + { + "bbox": [ + 201, + 194, + 235, + 224 + ], + "type": "image", + "image_path": "ee7a3fc15eb22a72c9ebbc433a3e5da63d58998d7d19d6372997d403c6adf92b.jpg" + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 238, + 194, + 277, + 224 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 194, + 277, + 224 + ], + "spans": [ + { + "bbox": [ + 238, + 194, + 277, + 224 + ], + "type": "image", + "image_path": "8bac52003f46b9494dffb3abe371cb847338a6206ba68db805bc1313e632a664.jpg" + } + ] + } + ], + "index": 12, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 302, + 87, + 341 + ], + "blocks": [ + { + "bbox": [ + 47, + 245, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 245, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 287, + 289 + ], + "type": "text", + "content": "Figure 3. Single Image Reconstruction. We show the reconstructed mesh from the input view and two additional views together with predicted texture and animated version of the shape obtained by articulating the estimated skeleton." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 302, + 87, + 341 + ], + "lines": [ + { + "bbox": [ + 49, + 302, + 87, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 302, + 87, + 341 + ], + "type": "image", + "image_path": "7a28d87885ae2b7f24c449130f1e175cb7274257b5f1c44071de4cbe3b79f1f2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 49, + 342, + 87, + 378 + ], + "blocks": [ + { + "bbox": [ + 49, + 342, + 87, + 378 + ], + "lines": [ + { + "bbox": [ + 49, + 342, + 87, + 378 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 87, + 378 + ], + "type": "image", + "image_path": "3705cf23c55248ffac3b192255ab12acafa62faae4160c61eac655b75a7744d3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 49, + 380, + 87, + 418 + ], + "blocks": [ + { + "bbox": [ + 49, + 380, + 87, + 418 + ], + "lines": [ + { + "bbox": [ + 49, + 380, + 87, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 380, + 87, + 418 + ], + "type": "image", + "image_path": "58cc294826e8105e7de96a450dd44572b73b5df61ab28104cc1eccbd4faf3861.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 421, + 77, + 429 + ], + "lines": [ + { + "bbox": [ + 60, + 421, + 77, + 429 + ], + "spans": [ + { + "bbox": [ + 60, + 421, + 77, + 429 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 437, + 287, + 471 + ], + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 471 + ], + "type": "text", + "content": "Figure 4. Reconstructions of Giraffes, Zebras and Cows. We finetune the horse model on other animal categories, and show that the method generalises well to other animals." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 100, + 302, + 276, + 342 + ], + "blocks": [ + { + "bbox": [ + 100, + 302, + 276, + 342 + ], + "lines": [ + { + "bbox": [ + 100, + 302, + 276, + 342 + ], + "spans": [ + { + "bbox": [ + 100, + 302, + 276, + 342 + ], + "type": "image", + "image_path": "d6518ddcf9016d9924d63c7c1be86b8e0b85220fa65a9c3abe9677010687f3e2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 99, + 344, + 277, + 380 + ], + "blocks": [ + { + "bbox": [ + 99, + 344, + 277, + 380 + ], + "lines": [ + { + "bbox": [ + 99, + 344, + 277, + 380 + ], + "spans": [ + { + "bbox": [ + 99, + 344, + 277, + 380 + ], + "type": "image", + "image_path": "577540a015dc4229f3d03956bb621aa23413ef4e5962f8167ab093978216b908.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 93, + 384, + 278, + 417 + ], + "blocks": [ + { + "bbox": [ + 93, + 384, + 278, + 417 + ], + "lines": [ + { + "bbox": [ + 93, + 384, + 278, + 417 + ], + "spans": [ + { + "bbox": [ + 93, + 384, + 278, + 417 + ], + "type": "image", + "image_path": "85f59bf0550fb776abec6067e49f24d889f9ad0a0292c018b48f81a977c17d07.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 420, + 173, + 428 + ], + "lines": [ + { + "bbox": [ + 123, + 420, + 173, + 428 + ], + "spans": [ + { + "bbox": [ + 123, + 420, + 173, + 428 + ], + "type": "text", + "content": "Reconstructions" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 420, + 230, + 428 + ], + "lines": [ + { + "bbox": [ + 203, + 420, + 230, + 428 + ], + "spans": [ + { + "bbox": [ + 203, + 420, + 230, + 428 + ], + "type": "text", + "content": "Textured" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 420, + 272, + 427 + ], + "lines": [ + { + "bbox": [ + 242, + 420, + 272, + 427 + ], + "spans": [ + { + "bbox": [ + 242, + 420, + 272, + 427 + ], + "type": "text", + "content": "Animated" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 205, + 344, + 231, + 380 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 344, + 231, + 380 + ], + "spans": [ + { + "bbox": [ + 205, + 344, + 231, + 380 + ], + "type": "image", + "image_path": "68cac3dbfe1d29aa32154ce4aea8847fedfa4b34db070d95569ab3daf152c91b.jpg" + } + ] + } + ], + "index": 23, + "type": "text" + }, + { + "bbox": [ + 204, + 385, + 231, + 416 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 385, + 231, + 416 + ], + "spans": [ + { + "bbox": [ + 204, + 385, + 231, + 416 + ], + "type": "image", + "image_path": "ab0097c353c395f496118b887989c846ed36ee0025701244c0ff92c1881c68a5.jpg" + } + ] + } + ], + "index": 24, + "type": "text" + }, + { + "bbox": [ + 239, + 315, + 276, + 342 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 315, + 276, + 342 + ], + "spans": [ + { + "bbox": [ + 239, + 315, + 276, + 342 + ], + "type": "image", + "image_path": "11634bb2a89b2b58bfc8a6c6cb257018269ce2bc48556733e0157f1f78d4dd19.jpg" + } + ] + } + ], + "index": 26, + "type": "text" + }, + { + "bbox": [ + 238, + 352, + 277, + 380 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 352, + 277, + 380 + ], + "spans": [ + { + "bbox": [ + 238, + 352, + 277, + 380 + ], + "type": "image", + "image_path": "f2ab693191ea6f7849cabb75f5e63136cefe52686580001112d44fd75110c2a3.jpg" + } + ] + } + ], + "index": 27, + "type": "text" + }, + { + "bbox": [ + 238, + 388, + 277, + 416 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 388, + 277, + 416 + ], + "spans": [ + { + "bbox": [ + 238, + 388, + 277, + 416 + ], + "type": "image", + "image_path": "0f70adb958b2debca9c114f1ea8ec5f0beecb7811bc0f07915a566f1579e1c88.jpg" + } + ] + } + ], + "index": 28, + "type": "text" + }, + { + "bbox": [ + 46, + 493, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 517 + ], + "type": "text", + "content": "eralises to various animal categories with highly different shapes and appearances, such as giraffes, zebras and cows." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 518, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 601 + ], + "type": "text", + "content": "Note that for the examples of abstract horse drawings (last rows in Figs. 1 and 3), since the texture is out of the training distribution, we finetune (only) the albedo network for 100 iterations, which takes less than 10 seconds. This is also done for giraffes, zebras and cows in Fig. 4, as the training datasets are too small to learn complex appearance. Additional qualitative results are presented in the sup. mat." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 47, + 610, + 208, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 208, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 208, + 623 + ], + "type": "text", + "content": "4.4. Comparison with Prior Work" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "We compare with previous weakly-supervised methods for 3D reconstruction of deformable objects. The most relevant prior work is UMR [32] and DOVE [67], as they also only require object masks and weak correspondences from either part segmentations (SCOPS) [18] or video training [67]. Our method leverages self-supervised DINO-ViT features, which makes our method the least supervised in" + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 351, + 70, + 387, + 106 + ], + "blocks": [ + { + "bbox": [ + 351, + 70, + 387, + 106 + ], + "lines": [ + { + "bbox": [ + 351, + 70, + 387, + 106 + ], + "spans": [ + { + "bbox": [ + 351, + 70, + 387, + 106 + ], + "type": "image", + "image_path": "a75296467bb8ba6fc586a4ad700fda96e4bd8febcfba23cb02cde3872af5ea97.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 345, + 114, + 375, + 143 + ], + "blocks": [ + { + "bbox": [ + 351, + 107, + 387, + 114 + ], + "lines": [ + { + "bbox": [ + 351, + 107, + 387, + 114 + ], + "spans": [ + { + "bbox": [ + 351, + 107, + 387, + 114 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 114, + 375, + 143 + ], + "lines": [ + { + "bbox": [ + 345, + 114, + 375, + 143 + ], + "spans": [ + { + "bbox": [ + 345, + 114, + 375, + 143 + ], + "type": "image", + "image_path": "eb613c7ff970271f282a66129e503565cdeed4bc4683c7b9ce136841758aa610.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 357, + 144, + 375, + 174 + ], + "blocks": [ + { + "bbox": [ + 357, + 144, + 375, + 174 + ], + "lines": [ + { + "bbox": [ + 357, + 144, + 375, + 174 + ], + "spans": [ + { + "bbox": [ + 357, + 144, + 375, + 174 + ], + "type": "image", + "image_path": "a8cf81f53bd65d742a41905ff378a4176df9a7c31e1935b094f8066e78aa39d7.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 356, + 176, + 423, + 205 + ], + "blocks": [ + { + "bbox": [ + 356, + 176, + 423, + 205 + ], + "lines": [ + { + "bbox": [ + 356, + 176, + 423, + 205 + ], + "spans": [ + { + "bbox": [ + 356, + 176, + 423, + 205 + ], + "type": "image", + "image_path": "0817229915715f0d08151937915ca92bd9068eada5d5642354f1005f6cc85406.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 209, + 352, + 216 + ], + "lines": [ + { + "bbox": [ + 320, + 209, + 352, + 216 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 352, + 216 + ], + "type": "text", + "content": "Input View" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 209, + 404, + 216 + ], + "lines": [ + { + "bbox": [ + 353, + 209, + 404, + 216 + ], + "spans": [ + { + "bbox": [ + 353, + 209, + 404, + 216 + ], + "type": "text", + "content": "Other Views" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 500, + 106 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 500, + 106 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 500, + 106 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 500, + 106 + ], + "type": "image", + "image_path": "b3cc63efc3281a9969e58de31326818a3f37b20f1de4662fd25d39066436eded.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 434, + 115, + 468, + 144 + ], + "blocks": [ + { + "bbox": [ + 465, + 107, + 501, + 114 + ], + "lines": [ + { + "bbox": [ + 465, + 107, + 501, + 114 + ], + "spans": [ + { + "bbox": [ + 465, + 107, + 501, + 114 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 115, + 468, + 144 + ], + "lines": [ + { + "bbox": [ + 434, + 115, + 468, + 144 + ], + "spans": [ + { + "bbox": [ + 434, + 115, + 468, + 144 + ], + "type": "image", + "image_path": "18e01ec428f8e73751db0a1c56f201752927291ebacfa66975f0b1c382a287b8.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 472, + 117, + 492, + 143 + ], + "blocks": [ + { + "bbox": [ + 472, + 117, + 492, + 143 + ], + "lines": [ + { + "bbox": [ + 472, + 117, + 492, + 143 + ], + "spans": [ + { + "bbox": [ + 472, + 117, + 492, + 143 + ], + "type": "image", + "image_path": "fe618f9c9c11dcaca113678b31d9738a303af75c17086ae3cfb366d072cf5a55.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 495, + 117, + 537, + 143 + ], + "blocks": [ + { + "bbox": [ + 495, + 117, + 537, + 143 + ], + "lines": [ + { + "bbox": [ + 495, + 117, + 537, + 143 + ], + "spans": [ + { + "bbox": [ + 495, + 117, + 537, + 143 + ], + "type": "image", + "image_path": "64ad89b35b46684ca3c27803a7ad8a794215f61970aa60253160d6fb642bea23.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 434, + 148, + 467, + 176 + ], + "blocks": [ + { + "bbox": [ + 434, + 148, + 467, + 176 + ], + "lines": [ + { + "bbox": [ + 434, + 148, + 467, + 176 + ], + "spans": [ + { + "bbox": [ + 434, + 148, + 467, + 176 + ], + "type": "image", + "image_path": "38b343d7fb4584ea6c88396c70f67baaa50e4df650b80ed72b6d6ac1975f105f.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 474, + 148, + 490, + 175 + ], + "blocks": [ + { + "bbox": [ + 474, + 148, + 490, + 175 + ], + "lines": [ + { + "bbox": [ + 474, + 148, + 490, + 175 + ], + "spans": [ + { + "bbox": [ + 474, + 148, + 490, + 175 + ], + "type": "image", + "image_path": "c416e2da3266b2a4c165b21443d3f56d8244501054289a10c1e77bed3f830121.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 496, + 148, + 537, + 175 + ], + "blocks": [ + { + "bbox": [ + 496, + 148, + 537, + 175 + ], + "lines": [ + { + "bbox": [ + 496, + 148, + 537, + 175 + ], + "spans": [ + { + "bbox": [ + 496, + 148, + 537, + 175 + ], + "type": "image", + "image_path": "df240beec029a314bcfc99ac10e39dd87abbbe42d21284a1e9a9e63943050a84.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 435, + 178, + 467, + 206 + ], + "blocks": [ + { + "bbox": [ + 435, + 178, + 467, + 206 + ], + "lines": [ + { + "bbox": [ + 435, + 178, + 467, + 206 + ], + "spans": [ + { + "bbox": [ + 435, + 178, + 467, + 206 + ], + "type": "image", + "image_path": "ac1956487352e3f6164cdb79447993a05ed8894a243dcba6d5626fc2ac5e850e.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 209, + 468, + 216 + ], + "lines": [ + { + "bbox": [ + 435, + 209, + 468, + 216 + ], + "spans": [ + { + "bbox": [ + 435, + 209, + 468, + 216 + ], + "type": "text", + "content": "Input View" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 472, + 180, + 500, + 205 + ], + "blocks": [ + { + "bbox": [ + 472, + 180, + 500, + 205 + ], + "lines": [ + { + "bbox": [ + 472, + 180, + 500, + 205 + ], + "spans": [ + { + "bbox": [ + 472, + 180, + 500, + 205 + ], + "type": "image", + "image_path": "933ec2e7fa1c9e6a5ddf8a515b42b305f47f0cd1fee39086c816a61c0b649e8a.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 209, + 519, + 216 + ], + "lines": [ + { + "bbox": [ + 484, + 209, + 519, + 216 + ], + "spans": [ + { + "bbox": [ + 484, + 209, + 519, + 216 + ], + "type": "text", + "content": "Other Views" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 307, + 263, + 343, + 298 + ], + "blocks": [ + { + "bbox": [ + 305, + 226, + 545, + 249 + ], + "lines": [ + { + "bbox": [ + 305, + 226, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 305, + 226, + 545, + 249 + ], + "type": "text", + "content": "Figure 5. Comparison on Birds. Both CMR [21] and UMR [32] often predict inaccurate poses, such as the bird on the left." + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 263, + 343, + 298 + ], + "lines": [ + { + "bbox": [ + 307, + 263, + 343, + 298 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 343, + 298 + ], + "type": "image", + "image_path": "0fa23d4fbbb5ed927fadb93eeea9544dfed710df3f8ae0100897834eee11afa9.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 345, + 264, + 378, + 298 + ], + "blocks": [ + { + "bbox": [ + 345, + 264, + 378, + 298 + ], + "lines": [ + { + "bbox": [ + 345, + 264, + 378, + 298 + ], + "spans": [ + { + "bbox": [ + 345, + 264, + 378, + 298 + ], + "type": "image", + "image_path": "14f701b684614336a9643db8048950891f1da6a8c4732fca4780577f53bf04eb.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 381, + 264, + 417, + 299 + ], + "blocks": [ + { + "bbox": [ + 381, + 264, + 417, + 299 + ], + "lines": [ + { + "bbox": [ + 381, + 264, + 417, + 299 + ], + "spans": [ + { + "bbox": [ + 381, + 264, + 417, + 299 + ], + "type": "image", + "image_path": "d60f8866c77fb9c0a963b151e5cf3ebde94b2828f8ea43c711781b3394182bfd.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 383, + 301, + 435, + 336 + ], + "blocks": [ + { + "bbox": [ + 383, + 301, + 435, + 336 + ], + "lines": [ + { + "bbox": [ + 383, + 301, + 435, + 336 + ], + "spans": [ + { + "bbox": [ + 383, + 301, + 435, + 336 + ], + "type": "image", + "image_path": "a0509fc8b9f89ce27d4ecabc82245b6f54d850da2ec6a25bc3b6c748fc4e4d6e.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 338, + 398, + 346 + ], + "lines": [ + { + "bbox": [ + 383, + 338, + 398, + 346 + ], + "spans": [ + { + "bbox": [ + 383, + 338, + 398, + 346 + ], + "type": "text", + "content": "UMR" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 354, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 388 + ], + "type": "text", + "content": "Figure 6. Comparison with UMR [32] on Horses. Our model produces much higher quality reconstructions with four legs. UMR is only able to predict symmetrical shapes with two legs." + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 438, + 266, + 470, + 298 + ], + "blocks": [ + { + "bbox": [ + 438, + 266, + 470, + 298 + ], + "lines": [ + { + "bbox": [ + 438, + 266, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 438, + 266, + 470, + 298 + ], + "type": "image", + "image_path": "821da40179a107739367c1522eef6ab489da739ebce3084792fc4b3ba67a1ae0.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 440, + 300, + 466, + 336 + ], + "blocks": [ + { + "bbox": [ + 440, + 300, + 466, + 336 + ], + "lines": [ + { + "bbox": [ + 440, + 300, + 466, + 336 + ], + "spans": [ + { + "bbox": [ + 440, + 300, + 466, + 336 + ], + "type": "image", + "image_path": "19774c162073b761f8a087cd43ceee893647f49468735935e64e1a076d3ac8a7.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 477, + 266, + 510, + 298 + ], + "blocks": [ + { + "bbox": [ + 477, + 266, + 510, + 298 + ], + "lines": [ + { + "bbox": [ + 477, + 266, + 510, + 298 + ], + "spans": [ + { + "bbox": [ + 477, + 266, + 510, + 298 + ], + "type": "image", + "image_path": "0537220f7f27445c7a778391f8fc8af328ea7dc33c40a55fa979b9ad5f5b7618.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 475, + 299, + 537, + 336 + ], + "blocks": [ + { + "bbox": [ + 475, + 299, + 537, + 336 + ], + "lines": [ + { + "bbox": [ + 475, + 299, + 537, + 336 + ], + "spans": [ + { + "bbox": [ + 475, + 299, + 537, + 336 + ], + "type": "image", + "image_path": "7054a5af1b7638e894b072e987bf4117fd4101244a3e5d224c2aeac503d6131e.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 338, + 492, + 346 + ], + "lines": [ + { + "bbox": [ + 477, + 338, + 492, + 346 + ], + "spans": [ + { + "bbox": [ + 477, + 338, + 492, + 346 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 413, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 497 + ], + "type": "text", + "content": "this area. We also compare against the well-established baseline of CMR [21], which requires stronger supervision in the form of 2D keypoint annotations and pre-computed cameras, as well as U-CMR [14], which uses a category-specific template shape. To evaluate other methods, we use their code and pre-trained models in public repositories (CMR, UMR, U-CMR) or provided by the authors (DOVE)." + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 503, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 635 + ], + "type": "text", + "content": "Qualitative Comparisons. Figs. 5 and 6 compare our reconstructions against the previous methods on bird and horse images. Since UMR did not release their horse model, we test our model on the examples presented in their paper for side-by-side comparison. Our method predicts more accurate poses and higher-quality articulated shapes (as opposed to only symmetric shapes with CMR and UMR). On horses, our method reconstructs all four legs with details while UMR predicts only two. Since UMR copies texture from the input image, for a fair comparison, we finetune (only) the albedo network for 100 iterations at test time." + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Quantitative Comparisons. We evaluate shape reconstruction on the 3D Toy Bird Dataset from DOVE [67], which contains 3D scans of 23 realistic toy bird models paired with 345 photographs of them taken in natural real-world environments. Following [67], we align the predicted mesh with the ground-truth scan using ICP and compute the" + } + ] + } + ], + "index": 69 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8798" + } + ] + } + ], + "index": 70 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 113, + 269, + 189 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 287, + 104 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 287, + 104 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 287, + 104 + ], + "type": "text", + "content": "Table 2. Evaluation on Toy Bird Scans [67]. Baseline results reported by [67]. Annotations: template shape, viewpoint, 2D keypoint, " + }, + { + "bbox": [ + 46, + 70, + 287, + 104 + ], + "type": "inline_equation", + "content": "\\mathfrak{S}\\ll" + }, + { + "bbox": [ + 46, + 70, + 287, + 104 + ], + "type": "text", + "content": " mask, " + }, + { + "bbox": [ + 46, + 70, + 287, + 104 + ], + "type": "inline_equation", + "content": "\\nVdash" + }, + { + "bbox": [ + 46, + 70, + 287, + 104 + ], + "type": "text", + "content": " optical flow, video." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 113, + 269, + 189 + ], + "lines": [ + { + "bbox": [ + 66, + 113, + 269, + 189 + ], + "spans": [ + { + "bbox": [ + 66, + 113, + 269, + 189 + ], + "type": "table", + "html": "
SupervisionChamfer Distance (cm) ↓
CMR [21]® ®1.35 ±0.81
U-CMR [14]® ®1.82 ±0.93
UMR [32]® + SCOPS1.24 ±0.75
DOVE [67]® ➔ ➔ ➔1.51 ±0.89
Ours® + DINO0.79 ±0.50
", + "image_path": "d6fbbc0f49ae95fe6fc998c4b6be4496da992f25c58367eb73bf8ac2ade7f641.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 262, + 287, + 380 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 287, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 254 + ], + "type": "text", + "content": "Table 3. Keypoint Transfer on PASCAL [10] Horse and CUB [64] Bird Datasets. Our model produces superior results on both categories with significantly less supervision. " + }, + { + "bbox": [ + 46, + 198, + 287, + 254 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 198, + 287, + 254 + ], + "type": "text", + "content": ": numbers taken from [32]. 'Birds w/o aq.': excluding 50 aquatic bird classes (e.g. mallards) with heavy occlusion and large shape variation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 262, + 287, + 380 + ], + "lines": [ + { + "bbox": [ + 48, + 262, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 287, + 380 + ], + "type": "table", + "html": "
MethodSupervisionPCK@0.1 ↑
HorsesBirdsBirds w/o aq.
CMR [21]†♂♀♂-0.473-
CMR [21]♂♀♂-0.5460.591
CMR [21]♂♀-0.2550.277
U-CMR [14]♀♀♀-0.3590.412
UMR [32]♂♀+ SCOPS0.2840.5120.555
A-CSM [29]♀♀♀0.3290.426-
DOVE [67]♂♀♂-0.4470.510
Ours♂♀+ DINO0.4290.5550.635
", + "image_path": "b3251b2508737829cbb5264772e0df1bcd04e615b949629c463688ff25fce147.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 400, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 459 + ], + "type": "text", + "content": "bi-directional Chamfer Distance between two sets of sample points from the aligned meshes. Tab. 2 summarises the results compared to other methods. Our model produces significantly more accurate reconstructions, resulting in a much lower error." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 652 + ], + "type": "text", + "content": "We also evaluate Keypoint Transfer on the CUB [64] and PASCAL VOC [10] benchmarks, a common evaluation metric [21, 29, 32]. We follow the protocols in [32] and [29] for birds and horses, respectively, and sample 20k source and target image pairs. Given a source image, we project all the visible vertices of the predicted mesh onto the image using the predicted viewpoint and assign each annotated 2D keypoint to its nearest vertex. We then render these vertices using the mesh and viewpoint predicted from a different target image. We measure the error between the projected vertices (corresponding to the transferred keypoints) and the annotated target keypoints using the Percentage of Correct Keypoints (PCK) metric. As shown in Tab. 3, our model significantly outperforms previous methods relying on stronger geometric supervision, including A-CSM [29] which requires a 3D shape template to begin with." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 225, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 225, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 225, + 671 + ], + "type": "text", + "content": "4.5. Ablation Studies and Limitations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "We present ablation studies on the two key components of the model in Fig. 7, self-supervised feature loss and multi-hypothesis viewpoint prediction. Without either" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 542, + 144 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 542, + 144 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 542, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 542, + 144 + ], + "type": "image", + "image_path": "920117262897d42f6c06a502a0ec2664a468d46dc4643c1269fa6f239cb736f1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 150, + 546, + 195 + ], + "lines": [ + { + "bbox": [ + 304, + 150, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 546, + 195 + ], + "type": "text", + "content": "Figure 7. Ablation Studies. We train the model without feature loss " + }, + { + "bbox": [ + 304, + 150, + 546, + 195 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 304, + 150, + 546, + 195 + ], + "type": "text", + "content": ", without multi-hypothesis (MH) viewpoint prediction and without both. Without either component, the viewpoint predictions collapse to frontal views, resulting in unnatural shapes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 216, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 251 + ], + "type": "text", + "content": "component, the learned viewpoints collapse to only frontal views, resulting in unnaturally stretched reconstructions. Additional results are provided in the sup. mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 252, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 546, + 396 + ], + "type": "text", + "content": "Although our model reconstructs highly detailed 3D shapes from only a single image and generalises to in-the-wild images, the predicted texture may not preserve sufficient details with a single forward pass and may require additional test-time finetuning. This can potentially be addressed by training on a larger dataset and incorporating recent advances in image generation [53,54]. Another limitation is the requirement of a pre-defined topology for articulation, which may vary between different animal species. Discovering the articulation structure automatically from raw in-the-wild images will be of great interest for future work. Failure cases are discussed in the sup. mat." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 407, + 383, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 383, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 383, + 419 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 427, + 546, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 546, + 548 + ], + "type": "text", + "content": "We have introduced a new model that can learn a 3D model of an articulated object category from single-view images taken in the wild. This model can, at test time, reconstruct the shape, articulation, albedo, and lighting of the object from a single image, and generalises to abstract drawings. Our approach demonstrates the power of combining several recent improvements in self-supervised representation learning together with a new viewpoint sampling scheme. We have shown significantly superior results compared to prior works, even when they use more supervision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "content": "Data Ethics. We use the DOVE, Weizmann Horse, PASCAL, Horse-10, and MS-COCO datasets in a manner compatible with their terms. Some of these images may accidentally contain faces or other personal information, but we do not make use of these images or image regions. For further details on ethics, data protection, and copyright please see https://www.robots.ox.ac.uk/~vedaldi/research/union/ethics.html." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgements. Shangzhe Wu is supported by a research gift from Meta Research. Andrea Vedaldi and Christian Rupprecht are supported by ERC-CoG UNION 101001212. Tomas Jakab and Christian Rupprecht are (also) supported by VisualAI EP/T028572/1." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8799" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "type": "text", + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. In ECCV Workshops, 2022. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 170 + ], + "type": "text", + "content": "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape completion and animation of people. ACM TOG, 2005. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 172, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 287, + 205 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 207, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 207, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 287, + 228 + ], + "type": "text", + "content": "[4] Eran Borenstein and Shimon Ullman. Class-specific, top-down segmentation. In ECCV, 2002. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 231, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 287, + 274 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 276, + 287, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 287, + 309 + ], + "type": "text", + "content": "[6] John E Chadwick, David R Haumann, and Richard E Parent. Layered construction for deformable animated characters. ACM SIGGRAPH Computer Graphics, 1989. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 312, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 287, + 354 + ], + "type": "text", + "content": "[7] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 357, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 399 + ], + "type": "text", + "content": "[8] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "text", + "content": "[9] Akio Doi and Akio Koide. An efficient method of triangulating equi-valued surfaces by using tetrahedral cells. IEICE Transactions on Information and Systems, 1991. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[10] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. *IJCV*, 2015. 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 483, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 287, + 505 + ], + "type": "text", + "content": "[11] Olivier Faugeras. Three-Dimensional Computer Vision: A Geometric Viewpoint. MIT Press, 1993. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 506, + 287, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 287, + 539 + ], + "type": "text", + "content": "[12] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In CVPR, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 541, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 541, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 541, + 287, + 574 + ], + "type": "text", + "content": "[13] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "type": "text", + "content": "[14] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoints without keypoints. In ECCV, 2020. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[16] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[17] Paul Henderson and Vittorio Ferrari. Learning single-image 3d reconstruction by generative modelling of shape, pose and shading. IJCV, 2019. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. SCOPS: Self-supervised co-part segmentation. In CVPR, 2019. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "text", + "content": "[19] Wonbong Jang and Lourdes Agapito. CodeNeRF: Disentangled neural radiance fields for object categories. In ICCV, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[20] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 177, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 219 + ], + "type": "text", + "content": "[21] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "text", + "content": "[22] Abhishek Kar, Shubham Tulsiani, João Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In CVPR, 2015. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 257, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 278 + ], + "type": "text", + "content": "[23] Hiroharu Kato and Tatsuya Harada. Learning view priors for single-view 3d reconstruction. In CVPR, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 280, + 545, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 301 + ], + "type": "text", + "content": "[24] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 304, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 335 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. PointRend: Image segmentation as rendering. In CVPR, 2020. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "text", + "content": "[26] Filippos Kokkinos and Iasonas Kokkinos. Learning monocular 3d reconstruction of articulated categories from motion. In CVPR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 373, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 405 + ], + "type": "text", + "content": "[27] Filippos Kokkinos and Iasonas Kokkinos. To the point: Correspondence-driven monocular 3d category reconstruction. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 407, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 545, + 472 + ], + "type": "text", + "content": "[28] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 4, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "type": "text", + "content": "[29] Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 510, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 541 + ], + "type": "text", + "content": "[30] Nilesh Kulkarni, Abhinav Gupta, and Shubham Tulsiani. Canonical surface mapping via geometric cycle consistency. In ICCV, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "text", + "content": "[31] Xueting Li, Sifei Liu, Shalini De Mello, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Online adaptation for consistent mesh reconstruction in the wild. In NeurIPS, 2020. 1, 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "text", + "content": "[32] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In ECCV, 2020. 2, 7, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[34] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In ICCV, 2019. 2" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8800" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[35] Yebin Liu, Juergen Gall, Carsten Stoll, Qionghai Dai, Hans-Peter Seidel, and Christian Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE TPAMI, 2013. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 152 + ], + "type": "text", + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 2015. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "text", + "content": "[37] Nadia Magnenat-Thalmann, E Primeau, and Daniel Thalmann. Abstract muscle action procedures for human face animation. The Visual Computer, 1988. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "type": "text", + "content": "[38] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W. Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In WACV, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "type": "text", + "content": "[39] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 325, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 325, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 325, + 287, + 357 + ], + "type": "text", + "content": "[41] Tom Monnier, Matthew Fisher, Alexei A. Efros, and Mathieu Aubry. Share With Thy Neighbors: Single-View Reconstruction by Cross-Instance Consistency. In ECCV, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "type": "text", + "content": "[42] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "text", + "content": "[43] Kieran A Murphy, Carlos Esteves, Varun Jampani, Srikumar Ramalingam, and Ameesh Makadia. Implicit-PDF: Nonparametric representation of probability distributions on the rotation manifold. In ICML, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 450, + 286, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 286, + 472 + ], + "type": "text", + "content": "[44] Natalia Neverova, David Novotny, and Andrea Vedaldi. Continuous surface embeddings. In NeurIPS, 2020. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 517 + ], + "type": "text", + "content": "[45] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3d representations from natural images. In ICCV, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 520, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 551 + ], + "type": "text", + "content": "[46] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In ICCV, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "text", + "content": "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "text", + "content": "[48] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[49] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[50] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In CVPR, 2020. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[51] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. PVA: Pixel-aligned volumetric avatars. In CVPR, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[52] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 151, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 182 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "type": "text", + "content": "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "text", + "content": "[55] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "text", + "content": "[56] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In NeurIPS, 2021. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "type": "text", + "content": "[57] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, 2007. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[58] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 393, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 448 + ], + "type": "text", + "content": "[59] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "text", + "content": "[60] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representation. In 3DV, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[61] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In CVPR, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "text", + "content": "[63] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "text", + "content": "[64] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 637, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 637, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 637, + 545, + 669 + ], + "type": "text", + "content": "[65] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[66] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8801" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 476 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[67] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. DOVE: Learning deformable 3d objects by watching videos. arXiv preprint arXiv:2107.10844, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[68] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In CVPR, 2021. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[69] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3D objects from images in the wild. In CVPR, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[70] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T. Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 240, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 285 + ], + "type": "text", + "content": "[71] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. ViSER: Video-specific surface embeddings for articulated 3d shape reconstruction. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "type": "text", + "content": "[72] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. BANMo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[73] Chun-Han Yao, Wei-Chih Hung, Michael Rubinstein, Yuanzhen Lee, Varun Jampani, and Ming-Hsuan Yang. Lassie: Learning articulated shape from sparse image ensemble via 3d part discovery. In NeurlPS, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 408 + ], + "type": "text", + "content": "[74] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In CVPR, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "type": "text", + "content": "[75] Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. Rel-Pose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "type": "text", + "content": "[76] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8802" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_content_list.json b/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ce6143732c011e4886d5291a4a2c00b8745b089 --- /dev/null +++ b/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_content_list.json @@ -0,0 +1,1887 @@ +[ + { + "type": "text", + "text": "Make Landscape Flatter in Differentially Private Federated Learning", + "text_level": 1, + "bbox": [ + 135, + 130, + 834, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifan Shi $^{1}$ Yingqi Liu $^{2}$ Kang Wei $^{2}$ Li Shen $^{3,*}$ Xueqian Wang $^{1,*}$ Dacheng Tao $^{3}$ $^{1}$ Tsinghua University, Shenzhen, China; $^{3}$ JD Explore Academy, Beijing, China \n $^{2}$ Nanjing University of Science and Technology, Nanjing, China", + "bbox": [ + 133, + 178, + 834, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shiyf21@mails.tsinghua.edu.cn; lyq@njust.edu.cn; kang.wei@njust.edu.cn;", + "bbox": [ + 163, + 236, + 795, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mathshenli@gmail.com; wang.xq@sz.tsinghua.edu.cn; dacheng.tao@gmail.com", + "bbox": [ + 163, + 253, + 792, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), client-level Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharper loss landscape and have poorer weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with better stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. From the theoretical perspective, we analyze in detail how DP-FedSAM mitigates the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with Rényi DP and present the sensitivity analysis of local updates. At last, we empirically confirm that our algorithm achieves state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL.", + "bbox": [ + 73, + 335, + 473, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 698, + 209, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) [31] allows distributed clients to collaboratively train a shared model without sharing data. However, FL faces severe dilemma of privacy leakage [27]. Recent works show that a curious server can also infer clients' privacy information such as membership and data features, by well-designed generative models and/or shadow models [14, 37, 40, 44, 55]. To address this issue, differential privacy (DP) [12] has been introduced in FL, which can protect every instance in any client's data (instance-level DP [3,20,47,48]) or the information between", + "bbox": [ + 75, + 724, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "clients (client-level DP [8, 15, 21, 25, 35, 52]). In general, client-level DP are more suitable to apply in the real-world setting due to a better model performance. For instance, a language prediction model with client-level DP [16, 35] has been applied on mobile devices by Google. In general, the Gaussian noise perturbation-based method is commonly adopted for ensuring the strong client-level DP. However, this method includes two operations in terms of clipping the $l_{2}$ norm of local updates to a sensitivity threshold $C$ and adding random noise proportional to the model size, whose standard deviation (STD) is also decided by $C$ . These steps may cause severe performance degradation dilemma [9, 22], especially on large-scale complex model [45], such as ResNet-18 [18], or with heterogeneous data.", + "bbox": [ + 496, + 304, + 893, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The reasons behind this issue may be two-fold: (i) The useful information is dropped due to the clipping operation especially on small $C$ , which is contained in the local updates; (ii) The model inconsistency among local models is exacerbated as the addition of random noise severely damages local updates and leads to large variances between local models especially on large $C$ [9]. To overcome these issues, existing works improve the performance via restricting the norm of local update [9] and leveraging local update sparsification technique [9, 22] to reduce the adverse impacts of clipping and adding amount of random noise. However, the model performance degradation remains significantly severe compared with FL methods without considering the privacy, e.g., FedAvg [33].", + "bbox": [ + 496, + 518, + 895, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivation. To further explore the reasons behind this phenomenon, we compare the structure of loss landscapes and surface contours [30] for FedAvg [33] and DP-FedAvg [15,35] on partitioned CIFAR-10 dataset [28] with Dirichlet distribution $(\\alpha = 0.6)$ and ResNet-18 backbone [18] in Figure 1 (a) and (b), respectively. Note that the convergence of DP-FedAvg is worse than FedAvg as its loss value is higher after a long communication round. Furthermore, FedAvg has a flatter landscape, whereas DP-FedAvg has a sharper one, resulting in poorer generalization ability (sharper minima, see Figure 1 (a)) and weight perturbation robustness", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Li Shen and Xueqian Wang", + "bbox": [ + 94, + 886, + 377, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24552", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96397fdd979bf62e21e9567d7174fcfbd362e96d6fd8fb400071671d3f0e583f.jpg", + "image_caption": [ + "(a) Loss landscapes." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 529, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6e2fb01e56f1e346f7dc14e8fda65fb56287e550f6faf37ab01b56057bd062d4.jpg", + "image_caption": [ + "Figure 1. Loss landscapes and surface contours comparison between DP-FedAvg (left) and FedAvg (right), respectively." + ], + "image_footnote": [], + "bbox": [ + 540, + 88, + 710, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c7c4f04a74b93d8df36cabfb7bf9c8cebbb2511fe8677dd0a64553467e760e8.jpg", + "image_caption": [ + "(b) Loss surface contours." + ], + "image_footnote": [], + "bbox": [ + 718, + 88, + 887, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(see Figure 1 (b)), which is caused by the clipped local update information and exacerbated model inconsistency, respectively. Based on these observations, an interesting research question is: can we further overcome the severe performance degradation via making landscape flatter?", + "bbox": [ + 73, + 231, + 468, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer this question, we propose DP-FedSAM via gradient perturbation to improve model performance. Specifically, a local flat model is generated by using SAM optimizer [13] in each client, which leads to better stability. After that, a potentially global flat model can be generated by aggregating several flat local models, which results in higher generalization ability and better robustness to DP noise, thereby significantly improving the performance and achieving a better trade-off between performance and privacy. Theoretically, we present a tighter bound $\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} + \\frac{\\sum_{t=1}^{T} (\\bar{\\alpha}^t \\sigma_g^2 + \\bar{\\alpha}^t L^2)}{T^2} + \\frac{L^2 \\sqrt{T} \\sigma^2 C^2 d}{m^2 \\sqrt{K}}\\right)$ for DP-FedSAM in the stochastic non-convex setting, where both $\\frac{1}{T} \\sum_{t=1}^{T} \\bar{\\alpha}^t$ and $\\frac{1}{T} \\sum_{t=1}^{T} \\tilde{\\alpha}^t$ are bounded constant, $K$ and $T$ is local iteration steps and communication rounds, respectively. Meanwhile, we present how SAM mitigates the impact of DP. For the clipping operation, DP-FedSAM reduces the $l_2$ norm and the negative impact of the inconsistency among local updates on convergence. For adding noise operation, we obtain better weight perturbation robustness for reducing the performance damage caused by random noise, thereby having better robustness to DP noise. Empirically, we conduct extensive experiments on EMNIST, CIFAR-10, and CIFAR-100 datasets in both the independently and identically distributed (IID) and Non-IID settings. Furthermore, we observe the loss landscape, surface contour, and the norm distribution of local updates for exploring the intrinsic effects of SAM with DP, which together with the theoretical analysis confirms the effect of DP-FedSAM.", + "bbox": [ + 75, + 314, + 472, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contribution. The main contributions of our work are summarized as four-fold: (i) We propose a novel scheme DPFedSAM in DPFL to alleviate performance degradation issue from the optimizer perspective. (ii) We establish an improved convergence rate, which is tighter than the conventional bounds [9, 22] in the stochastic non-convex setting. Moreover, we provide rigorous privacy guarantees and sensitivity analysis. (iii) We are the first to in-depth analyze the roles of the on-average norm of local updates $\\overline{\\alpha}^t$ and local update consistency among clients $\\tilde{\\alpha}^t$ on convergence. (iv)", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments to verify the effect of DP-FedSAM, which could achieve state-of-the-art (SOTA) performance compared with several strong DPFL baselines.", + "bbox": [ + 496, + 232, + 893, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 286, + 637, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Client-level DPFL. Client-level DPFL is the de-facto approach for protecting any client's data. DP-FedAvg [36] is the first to study in this setting, which trains a language prediction model in a mobile keyboard and ensures client-level DP guarantee by employing the Gaussian mechanism and composing privacy guarantees. After that, the work in [26,51] presents a comprehensive end-to-end system, which appropriately discretizes the data and adds discrete Gaussian noise before performing secure aggregation. Meanwhile, AE-DPFL [58] leverages the voting-based mechanism among the data labels instead of averaging the gradients to avoid dimension dependence and significantly reduce the communication cost. Fed-SMP [22] uses Sparsified Model Perturbation (SMP) to mitigate the impact of privacy protection on model accuracy. Different from the aforementioned methods, a recent study [9] revisits this issue and leverages Bounded Local Update Regularization (BLUR) and Local Update Sparsification (LUS) to restrict the norm of local updates and reduce the noise size before executing operations that guarantee DP, respectively. Nevertheless, severe performance degradation remains.", + "bbox": [ + 496, + 311, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sharpness Aware Minimization (SAM). SAM [13] is an effective optimizer for training deep learning (DL) models, which leverages the flatness geometry of the loss landscape to improve model generalization ability. Recently, [4] study the properties of SAM and provide convergence results of SAM for non-convex objectives. As a powerful optimizer, SAM and its variants have been applied to various machine learning (ML) tasks [2, 11, 24, 29, 32, 38, 43, 46, 50, 56, 57]. Specifically, [41], [50] and [7] integrate SAM to improve the generalization, and thus mitigate the distribution shift problem and achieve a new SOTA performance for FL. However, to the best of our knowledge, no efforts have been devoted to the empirical performance and theoretical analysis of SAM in DPFL.", + "bbox": [ + 496, + 628, + 892, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most related works to this paper are DP-FedAvg in [35], Fed-SMP in [22], and DP-FedAvg with LUS and BLUR [9]. However, these works still suffer from inferior performance due to the exacerbated model inconsistency is", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24553", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sue among the clients caused by random noise. Therefore, different from existing works, we try to alleviate this issue by making the landscape flatter and weight perturbation ability more robust. Furthermore, another related work is FedSAM [41], which integrates SAM optimizer to enhance the flatness of the local model and achieves new SOTA performance for FL. On top of the aforementioned studies, we are the first to extend the SAM optimizer into the DPFL setting to effectively alleviate the performance degradation problem. Meanwhile, we provide the theoretical analysis for sensitivity, privacy, and convergence in the non-convex setting. Finally, we empirically confirm our theoretical results and the superiority of performance compared with existing SOTA methods in DPFL.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 315, + 202, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first give the problem setup of FL, and then introduce the several terminologies in DP.", + "bbox": [ + 76, + 340, + 468, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Federated Learning", + "text_level": 1, + "bbox": [ + 76, + 378, + 266, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a general FL system consisting of $M$ clients, in which each client owns its local dataset. Let $\\mathcal{D}_i^{\\mathrm{train}}$ , $\\mathcal{D}_i^{\\mathrm{val}}$ and $\\mathcal{D}_i^{\\mathrm{test}}$ denote the training dataset, validation dataset and testing dataset, held by client $i$ , respectively, where $i \\in \\mathcal{U} = \\{1, 2, \\dots, M\\}$ . Formally, the FL task is expressed as:", + "bbox": [ + 76, + 402, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} ^ {\\star} = \\underset {\\mathbf {w}} {\\arg \\min } \\sum_ {i \\in \\mathcal {U}} p _ {i} f _ {i} (\\mathbf {w}, \\mathcal {D} _ {i} ^ {\\text {t r a i n}}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 484, + 468, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_i = |\\mathcal{D}_i^{\\mathrm{train}}| / |\\mathcal{D}^{\\mathrm{train}}| \\geq 0$ with $\\sum_{i\\in \\mathcal{U}}p_i = 1$ , and $f_{i}(\\cdot)$ is the local loss function with $f_{i}(\\mathbf{w}) = F_{i}(\\mathbf{w};\\xi_{i}), \\xi_{i}$ is a batch sample data in client $i$ . $|\\mathcal{D}_i^{\\mathrm{train}}|$ is the size of training dataset $\\mathcal{D}_i^{\\mathrm{train}}$ and $|\\mathcal{D}^{\\mathrm{train}}| = \\sum_{i\\in \\mathcal{U}}|\\mathcal{D}_i^{\\mathrm{train}}|$ is the total size of training datasets, respectively. For the $i$ -th client, a local model is learned on its private training data $\\mathcal{D}^{\\mathrm{train}}$ via:", + "bbox": [ + 76, + 522, + 468, + 613 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {i} = \\mathbf {w} _ {i} ^ {t} - \\eta \\nabla f _ {i} \\left(\\mathbf {w} _ {i}, \\mathcal {D} _ {i} ^ {\\text {t r a i n}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 621, + 468, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generally, the local loss function $f_{i}(\\cdot)$ has the same expression across each client. Then, the $M$ associated clients collaboratively learn a global model $\\mathbf{w}$ over the heterogeneous training data $\\mathcal{D}_i^{\\mathrm{train}}$ , $\\forall i \\in \\mathcal{U}$ .", + "bbox": [ + 76, + 645, + 468, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Differential Privacy", + "text_level": 1, + "bbox": [ + 76, + 714, + 264, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differential Privacy (DP) [12] is a rigorous privacy notion for measuring privacy risk. In this paper, we consider a relaxed version: Rényi DP (RDP) [39].", + "bbox": [ + 76, + 738, + 468, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1. (Rényi DP, [39]). Given a real number $\\alpha \\in (1,\\infty)$ and privacy parameter $\\rho \\geq 0$ , a randomized mechanism $\\mathcal{M}$ satisfies $(\\alpha ,\\rho)$ -RDP if for any two neighboring datasets $\\mathcal{D}$ , $\\mathcal{D}'$ that differ in a single record, the Rényi $\\alpha$ -divergence between $\\mathcal{M}(\\mathcal{D})$ and $\\mathcal{M}(\\mathcal{D}')$ satisfies:", + "bbox": [ + 76, + 791, + 468, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\alpha} \\left[ \\mathcal {M} (\\mathcal {D}) \\| \\mathcal {M} \\left(\\mathcal {D} ^ {\\prime}\\right) \\right] := \\frac {1}{\\alpha - 1} \\log \\mathbb {E} \\left[ \\left(\\frac {\\mathcal {M} (\\mathcal {D})}{\\mathcal {M} \\left(\\mathcal {D} ^ {\\prime}\\right)}\\right) ^ {\\alpha} \\right] \\leq \\rho , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 873, + 468, + 905 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the expectation is taken over the output of $\\mathcal{M}(\\mathcal{D}^{\\prime})$", + "bbox": [ + 500, + 90, + 879, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rényi DP is a useful analytical tool to measure the privacy and accurately represent guarantees on the tails of the privacy loss, which is strictly stronger than $(\\epsilon, \\delta)$ -DP for $\\delta > 0$ . Thus, we provide the privacy analysis based on this tool for each user's privacy loss.", + "bbox": [ + 498, + 117, + 890, + 194 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2. ( $l_{2}$ Sensitivity). [9, Definition 2] Let $\\mathcal{F}$ be a function, the $L_{2}$ -sensitivity of $\\mathcal{F}$ is defined as $\\mathcal{S} = \\max_{D \\simeq D'} \\| \\mathcal{F}(D) - \\mathcal{F}(D') \\|_2$ , where the maximization is taken over all pairs of adjacent datasets.", + "bbox": [ + 498, + 204, + 890, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sensitivity of a function $\\mathcal{F}$ captures the magnitude which a single individual's data can change the function $\\mathcal{F}$ in the worst case. Therefore, it plays a crucial role in determining the magnitude of noise required to ensure DP.", + "bbox": [ + 498, + 276, + 890, + 337 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3. (Client-level DP) [34, Definition 1]. A randomized algorithm $\\mathcal{M}$ is $(\\epsilon, \\delta)$ -DP if for any two adjacent datasets $U$ , $U'$ constructed by adding or removing all records of any client, and every possible subset of outputs $O$ satisfy the following inequality:", + "bbox": [ + 498, + 348, + 890, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Pr [ \\mathcal {M} (U) \\in O ] \\leq e ^ {\\epsilon} \\Pr [ \\mathcal {M} \\left(U ^ {\\prime}\\right) \\in O ] + \\delta . \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 436, + 890, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In client-level DP, we aim to ensure participation information for any clients. Therefore, we need to make local updates similar whether one client participates or not.", + "bbox": [ + 498, + 465, + 890, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 527, + 635, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To revisit the severe performance degradation challenge in DPFL, we observe the loss landscapes and surface contours of FedAvg and DP-FedAvg in Figure 1 (a) and (b), respectively. And we find that the DPFL method has a sharper landscape with both poorer generalization ability and weight perturbation robustness than the FL method. It means that the DPFL method may result in poor flatness and make model sensitivity to noise perturbation. However, the study focusing on this issue remains unexplored. Therefore, we plan to face this challenge from the optimizer perspective by adopting a SAM optimizer in each client, dubbed DP-FedSAM, whose local loss function is defined as:", + "bbox": [ + 498, + 553, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} (\\mathbf {w}) = \\mathbb {E} _ {\\xi \\sim \\mathcal {D} _ {i}} \\max _ {\\| \\delta_ {i} \\| _ {2} \\leq \\rho} F _ {i} \\left(\\mathbf {w} ^ {t, k} (i) + \\delta_ {i}; \\xi_ {i}\\right), \\quad i \\in \\mathcal {N}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 743, + 890, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{w}^{t,k}(i) + \\delta_{i}$ is viewed as the perturbed model, $\\rho$ is a predefined constant controlling the radius of the perturbation and $\\| \\cdot \\| _2$ is a $l_{2}$ -norm. Instead of searching for a solution via SGD [5, 6], SAM [13] aims to seek a solution in a flat region by adding a small perturbation, i.e., $w + \\delta$ with more robust performance. Specifically, for each client $i\\in \\{1,2,\\dots,M\\}$ , each local iteration $k\\in \\{0,1,\\dots,K - 1\\}$ in each communication round $t\\in \\{0,1,\\dots,T - 1\\}$ , the $k$ -th", + "bbox": [ + 498, + 779, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24554", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inner iteration in client $i$ is performed as:", + "bbox": [ + 76, + 90, + 351, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} ^ {t, k + 1} (i) = \\mathbf {w} ^ {t, k} (i) - \\eta \\tilde {\\mathbf {g}} ^ {t, k} (i), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 117, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {g}} ^ {t, k} (i) = \\nabla F _ {i} \\left(\\mathbf {w} ^ {t, k} + \\delta \\left(\\mathbf {w} ^ {t, k}\\right); \\xi\\right), \\delta \\left(\\mathbf {w} ^ {t, k}\\right) = \\frac {\\rho \\mathbf {g} ^ {t , k}}{\\left\\| \\mathbf {g} ^ {t , k} \\right\\| _ {2}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 138, + 468, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\delta(\\mathbf{w}^{t,k})$ is calculated by using first-order Taylor expansion around $\\mathbf{w}^{t,k}$ [13]. After that, we adopt a sampling mechanism, gradient clipping, and Gaussian noise adding to ensure client-level DP. Note that this sampling can amplify the privacy guarantee since it decreases the chances of leaking information about a particular individual [1]. Specifically, after sampling $m$ clients with probability $q = m / M$ at each communication round, which is important for measuring privacy loss. Then, we clip the local updates first:", + "bbox": [ + 76, + 186, + 468, + 324 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\Delta} _ {i} ^ {t} = \\Delta_ {i} ^ {t} \\cdot \\min \\left(1, \\frac {C}{\\| \\Delta_ {i} ^ {t} \\| _ {2}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 335, + 468, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After clipping, we add Gaussian noise to the local update for ensuring client-level DP as follows:", + "bbox": [ + 76, + 376, + 468, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Delta} _ {i} ^ {t} = \\tilde {\\Delta} _ {i} ^ {t} + \\mathcal {N} \\left(0, \\sigma^ {2} C ^ {2} \\cdot \\mathbf {I} _ {d} / m\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 417, + 468, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With a given noise variance $\\sigma^2$ , the accumulative privacy budget $\\epsilon$ can be calculated based on the sampled Gaussian mechanism [54]. Finally, we summarize the training procedures of DP-FedSAM in Algorithm 1.", + "bbox": [ + 76, + 446, + 468, + 507 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: DP-FedSAM" + ], + "code_body": "Input: Total number of clients $M$ sampling ratio of clients $q$ total number of communication rounds $T$ the clipping threshold $C$ local learning rate $\\eta$ and total number of the local iterates are $K$ . Output: Generate global model $\\mathbf{w}^T$ . \n1 Initialization: Randomly initialize the global model $\\mathbf{w}^0$ . \n2 for $t = 0$ to $T - 1$ do \n3 Sample a set of $m = qM$ clients at random without replacement, denoted by $\\mathcal{W}^t$ . \n4 for client $i = 1$ to $m$ in parallel do \n5 for $k = 0$ to $K - 1$ do \n6 Update the global parameter as the local parameter $\\mathbf{w}^t(i) \\gets \\mathbf{w}^t$ . \n7 Sample a batch of local data $\\xi_i$ and calculate local gradient $\\mathbf{g}^{t,k}(i) = \\nabla F_i(\\mathbf{w}^{t,k}(i); \\xi_i)$ . \n8 Gradient perturbation by Equation (7). Local iteration update by Equation (6). \n9 end \n11 $\\Delta_i^t = \\mathbf{w}^{t,K}(i) - \\mathbf{w}^{t,0}(i)$ . \n12 Clip and add noise for DP by Equation (9). \n13 Return $\\hat{\\Delta}^t(i)$ . \n14 end \n15 $\\mathbf{w}^{t + 1} \\gets \\mathbf{w}^t + \\frac{1}{m} \\sum_{i \\in \\mathcal{W}^t} \\hat{\\Delta}^t(i)$ . \n16 end", + "bbox": [ + 81, + 544, + 467, + 890 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with existing DPFL methods [9, 22, 35], the benefits of DP-FedSAM lie in three-fold: (i) We introduce SAM into DPFL to alleviate severe performance degradation via seeking a flatness model in each client, which is caused by the exacerbated inconsistency of local models. Specifically, DP-FedSAM generates both better generalization ability and robustness to DP noise by making the global model flatter. (ii) In addition, we analyze in detail how DP-FedSAM mitigates the negative impacts of DP. Specifically, we theoretically analyze the convergence with the on-average norm of local updates $\\overline{\\alpha}^t$ and local update consistency among clients $\\tilde{\\alpha}^t$ , and empirically confirm these results via observing the norm distribution and average norm of local updates (see Section 6.3). (iii) Meanwhile, we present the theories unifying the impacts of gradient perturbation $\\rho$ in SAM, the on-average norm of local updates $\\overline{\\alpha}^t$ , and local update consistency among clients $\\tilde{\\alpha}^t$ in clipping operation, and the variance of random noise $\\sigma^2 C^2 / m$ upon the convergence rate in Section 5.", + "bbox": [ + 496, + 90, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 500, + 388, + 694, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we give a rigorous analysis of DP-FedSAM, including its sensitivity, privacy, and convergence rate. The detailed proof is placed in Appendix E. Below, we first give several necessary assumptions.", + "bbox": [ + 496, + 414, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assumption 1. (Lipschitz smoothness). The function $F_{i}$ is differentiable and $\\nabla F_{i}$ is $L$ -Lipschitz continuous, $\\forall i \\in \\{1,2,\\ldots,M\\}$ , i.e., $\\| \\nabla F_{i}(\\mathbf{x}) - \\nabla F_{i}(\\mathbf{y}) \\| \\leq L \\| \\mathbf{x} - \\mathbf{y} \\|$ , for all $\\mathbf{x}, \\mathbf{y} \\in \\mathbb{R}^{d}$ .", + "bbox": [ + 498, + 479, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assumption 2. (Bounded variance). The gradient of the function $f_{i}$ have $\\sigma_l$ -bounded variance, i.e., $\\mathbb{E}_{\\xi_i}\\left\\| \\nabla F_i(\\mathbf{w}^k (i);\\xi_i) - \\nabla F_i(\\mathbf{w}(i))\\right\\|^2 \\leq \\sigma_l^2$ , $\\forall i \\in \\{1,2,\\dots,M\\}$ , $k \\in \\{1,\\dots,K - 1\\}$ , the global variance is also bounded, i.e., $\\frac{1}{M}\\sum_{i = 1}^{M}\\| \\nabla f_{i}(\\mathbf{w}) - \\nabla f(\\mathbf{w})\\|^{2} \\leq \\sigma_{g}^{2}$ for all $\\mathbf{w} \\in \\mathbb{R}^d$ . It is not hard to verify that the $\\sigma_g$ is smaller than the homogeneity parameter $\\beta$ , i.e., $\\sigma_g^2 \\leq \\beta^2$ .", + "bbox": [ + 498, + 546, + 890, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assumption 3. (Bounded gradient). For any $i \\in \\{1,2,\\ldots,M\\}$ and $\\mathbf{w} \\in \\mathbb{R}^d$ , we have $\\|\\nabla f_i(\\mathbf{w})\\| \\leq B$ .", + "bbox": [ + 500, + 662, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assumption 4. (Unbiased Gradient Estimator). For any data sample $z$ from $\\mathcal{D}_i$ and $\\mathbf{w} \\in \\mathbb{R}^d$ , the local gradient estimator is unbiased, i.e., $\\mathbb{E}[\\nabla f_i(\\mathbf{w}; z)] = \\mathbb{E}[\\nabla f_i(\\mathbf{w})]$ .", + "bbox": [ + 498, + 699, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the above assumptions are mild and commonly used in characterizing the convergence rate of FL [6,9,17,22,23,41-43,49,53]. Furthermore, gradient clipping operation in DL is often used to prevent the gradient explosion phenomenon, thereby the gradient is bounded. The technical difficulty for DP-FedSAM lies in: (i) how SAM mitigates the impact of DP; (ii) how to analyze in detail the impacts of the consistency among clients and the on-average norm of local updates caused by clipping operation.", + "bbox": [ + 496, + 750, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24555", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Sensitivity Analysis", + "text_level": 1, + "bbox": [ + 76, + 90, + 263, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At first, we study the sensitivity of local update $\\Delta_i^t$ from any client $i\\in \\{1,2,\\dots,M\\}$ before clipping at $t$ -th communication round in DP-FedSAM. This upper bound of sensitivity can roughly measure the degree of privacy protection. Under Definition 2, the sensitivity can be denoted by $S_{\\Delta_i^t}$ in client $i$ at $t$ -th communication round.", + "bbox": [ + 75, + 112, + 468, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. (Sensitivity). Denote $\\Delta_i^t (\\mathbf{x})$ and $\\Delta_i^t (\\mathbf{y})$ as the local update at $t$ -th communication round, the model $\\mathbf{x}(i)$ and $\\mathbf{y}(i)$ is conducted on two sets which differ at only one sample. Assume the initial model parameter $\\mathbf{w}^t (i) = \\mathbf{x}^{t,0}(i) = \\mathbf{y}^{t,0}(i)$ . With the above assumptions, the expected squared sensitivity $S_{\\Delta_i^t}^2$ of local update is upper bounded,", + "bbox": [ + 76, + 212, + 470, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\mathcal {S} _ {\\Delta_ {i} ^ {t}} ^ {2} \\leq \\frac {6 \\eta^ {2} \\rho^ {2} K L ^ {2} \\left(1 2 K ^ {2} L ^ {2} \\eta^ {2} + 1 0\\right)}{1 - 2 \\eta^ {2} L ^ {2} K} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 323, + 468, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the local adaptive learning rate satisfies $\\eta = \\mathcal{O}(1 / L\\sqrt{KT})$ and the perturbation amplitude $\\rho$ proportional to the learning rate, e.g., $\\rho = \\mathcal{O}\\left(\\frac{1}{\\sqrt{T}}\\right)$ , we have", + "bbox": [ + 76, + 364, + 468, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\mathcal {S} _ {\\Delta_ {i} ^ {t}} ^ {2} \\leq \\mathcal {O} \\left(\\frac {1}{T ^ {2}}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 422, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For comparison, we also present the expected squared sensitivity of local update with SGD in DPFL, that is $\\mathbb{E}\\mathcal{S}_{\\Delta_i^t,SGD}^2\\leq \\frac{6\\eta^2\\sigma_l^2K}{1 - 3\\eta^2KL^2}$ . Thus $\\mathbb{E}\\mathcal{S}_{\\Delta_i^t,SGD}^2\\leq \\mathcal{O}(\\frac{\\sigma_l^2}{KL^2T})$ when $\\eta = \\mathcal{O}(1 / L\\sqrt{KT})$ .", + "bbox": [ + 76, + 464, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1. It is clearly seen that the upper bound in $\\mathbb{ES}_{\\Delta_i^t,SAM}^2$ is tighter than that in $\\mathbb{ES}_{\\Delta_i^t,SGD}^2$ . From the perspective of privacy protection, it means DP-FedSAM has a better privacy guarantee than DP-FedAvg. Another perspective from local iteration, which means both better model consistency among clients and training stability.", + "bbox": [ + 76, + 541, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Privacy Analysis", + "text_level": 1, + "bbox": [ + 76, + 641, + 241, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve client-level privacy protection, we derive the sensitivity of the aggregation process at first after clipping the local updates.", + "bbox": [ + 76, + 664, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 1. The sensitivity of client-level DP in DP-FedSAM can be expressed as $C / m$ .", + "bbox": [ + 76, + 718, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. Given two adjacent batches $\\mathcal{W}^t$ and $\\mathcal{W}^{t,\\mathrm{adj}}$ , that $\\mathcal{W}^{t,\\mathrm{adj}}$ has one more or less client, we have", + "bbox": [ + 76, + 755, + 468, + 786 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\frac {1}{m} \\sum_ {i \\in \\mathcal {W} ^ {t}} \\Delta_ {i} ^ {t} - \\frac {1}{m} \\sum_ {j \\in \\mathcal {W} ^ {t, \\mathrm {a d j}}} \\Delta_ {j} ^ {t} \\right\\| _ {2} = \\frac {1}{m} \\left\\| \\Delta_ {j ^ {\\prime}} ^ {t} \\right\\| _ {2} \\leq \\frac {C}{m}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 792, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta_{j'}^t$ is the local update of one more or less client.", + "bbox": [ + 76, + 845, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 2. The value of this sensitivity can determine the amount of variance for adding random noise.", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After adding Gaussian noise, we calculate the accumulative privacy budget [54] along with training as follows.", + "bbox": [ + 500, + 90, + 890, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. After $T$ communication rounds, the accumulative privacy budget is calculated by:", + "bbox": [ + 500, + 128, + 890, + 159 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon = \\bar {\\epsilon} + \\frac {(\\alpha - 1) \\log (1 - \\frac {1}{\\alpha}) - \\log (\\alpha) - \\log (\\delta)}{\\alpha - 1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 164, + 890, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 200, + 544, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\epsilon} = \\frac {T}{\\alpha - 1} \\ln \\mathbb {E} _ {z \\sim \\mu_ {0} (z)} \\left[ \\left(1 - q + \\frac {q \\mu_ {1} (z)}{\\mu_ {0} (z)}\\right) ^ {\\alpha} \\right], \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 219, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $q$ is sample rate for client selection, $\\mu_0(z)$ and $\\mu_{1}(z)$ denote the Gaussian probability density function (PDF) of $\\mathcal{N}(0,\\sigma)$ and the mixture of two Gaussian distributions $q\\mathcal{N}(1,\\sigma) + (1 - q)\\mathcal{N}(0,\\sigma)$ , respectively, $\\sigma$ is the noise STD, $\\alpha$ is a selectable variable.", + "bbox": [ + 498, + 256, + 890, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 3. It can be seen that a small sampling rate $q$ can enhance the privacy guarantee by decreasing the privacy budget, but it may also degrade the training performance due to the number of participating clients being reduced in each communication round. So a better trade-off is needed.", + "bbox": [ + 498, + 339, + 890, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Convergence Analysis", + "text_level": 1, + "bbox": [ + 500, + 422, + 705, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Below, we give a convergence analysis of how DP-FedSAM mitigates the negative impacts of DP. The technical contribution also is combining the impacts of the on-average norm of local updates $\\overline{\\alpha}^t$ and local update consistency among clients $\\tilde{\\alpha}^t$ on the rate. Moreover, we also empirically confirm these results in Section 6.3.", + "bbox": [ + 498, + 444, + 890, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3. Under assumptions 1-4, local learning rate satisfies $\\eta = \\mathcal{O}(1 / L\\sqrt{KT})$ and $f^{*}$ is denoted as the minimal value of $f$ , i.e., $f(x)\\geq f(x^{*}) = f^{*}$ for all $x\\in \\mathbb{R}^d$ . When the perturbation amplitude $\\rho$ is proportional to the learning rate, e.g., $\\rho = \\mathcal{O}(1 / \\sqrt{T})$ , the sequence of outputs $\\{\\mathbf{w}^t\\}$ generated by Alg. 1, we have:", + "bbox": [ + 498, + 542, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\mathbb {E} \\left[ \\overline {{\\alpha}} ^ {t} \\left\\| \\nabla f \\left(\\mathbf {w} ^ {t}\\right) \\right\\| ^ {2} \\right] \\leq \\underbrace {\\mathcal {O} \\left(\\frac {2 L \\left(f \\left(\\mathbf {w} ^ {1}\\right) - f ^ {*}\\right)}{\\sqrt {K T}} + \\frac {L ^ {2} \\sigma_ {l} ^ {2}}{K T ^ {2}}\\right)} _ {\\text {F r o m F e d S A M}} \\\\ + \\underbrace {\\mathcal {O} \\left(\\frac {\\sum_ {t = 1} ^ {T} \\left(\\bar {\\alpha} ^ {t} \\sigma_ {g} ^ {2} + \\tilde {\\alpha} ^ {t} L ^ {2}\\right)}{T ^ {2}}\\right)} _ {\\text {C l i p p i n g}} + \\underbrace {\\mathcal {O} \\left(\\frac {L ^ {2} \\sqrt {T} \\sigma^ {2} C ^ {2} d}{m ^ {2} \\sqrt {K}}\\right)} _ {\\text {A d d i n g n o i s e}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 641, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 770, + 544, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} ^ {t} := \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\alpha_ {i} ^ {t} \\quad a n d \\quad \\tilde {\\alpha} ^ {t} := \\frac {1}{M} \\sum_ {i = 1} ^ {M} | \\alpha_ {i} ^ {t} - \\overline {{\\alpha_ {i} ^ {t}}} |, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 789, + 890, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\alpha_{i}^{t} = \\min (1,\\frac{C}{\\eta\\|\\sum_{k = 0}^{K - 1}\\tilde{\\mathbf{g}}^{t,k}(i)\\|})$ , respectively. Note that $\\overline{\\alpha}^t$ and $\\tilde{\\alpha}^t$ measure the on-average norm of local updates and local update consistency among clients before clipping and adding noise operations in DP-FedSAM, respectively.", + "bbox": [ + 498, + 834, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24556", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5af9216a8de6ea61f70d6443d1ac6bdb2ec37fdf56d2a0d8c7aa0cb4a1f6bec4.jpg", + "table_caption": [ + "Table 1. Averaged training accuracy (%) and testing accuracy (%) on two data in both IID and Non-IID settings for all compared methods." + ], + "table_footnote": [], + "table_body": "
TaskAlgorithmDirichlet 0.3Dirichlet 0.6IID
TrainValidationTrainValidationTrainValidation
EMNISTDP-FedAvg99.28±0.0273.10±0.1699.55±0.0282.20±0.3599.66±0.4081.90±0.86
Fed-SMP-randk99.24±0.0273.72±0.5399.71±0.0182.18±0.7399.71±0.6184.16±0.83
Fed-SMP-topk99.31±0.0475.75±0.3599.72±0.0283.41±0.9199.73±0.4083.32±0.52
DP-FedAvg-blur99.12±0.0273.71±0.0299.66±0.0083.20±0.0199.67±0.0382.92±0.49
DP-FedAvg-blurs99.63±0.0876.25±0.3599.72±0.0283.41±0.9199.74±0.4582.92±0.49
DP-FedSAM96.28±0.6476.81±0.8195.07±0.4584.32±0.1995.61±0.9485.90±0.72
DP-FedSAM-topk94.77±0.1177.27±0.6795.87±1.5284.80±0.6096.12±0.8587.70±0.83
CIFAR-10DP-FedAvg93.65±0.4747.98±0.2493.65±0.4250.05±0.4793.65±0.1550.90±0.86
Fed-SMP-randk95.46±0.4348.14±0.1295.36±0.0651.33±0.3695.36±0.0650.61±0.20
Fed-SMP-topk95.49±0.1449.93±2.2995.49±0.0954.11±0.8395.49±0.1053.30±0.45
DP-FedAvg-blur95.47±0.1247.66±0.0199.66±0.4251.05±0.0194.50±0.0552.56±0.47
DP-FedAvg-blurs96.79±0.5151.23±0.6699.72±0.0954.11±0.8396.45±0.3053.48±0.76
DP-FedSAM90.38±0.9053.92±0.5590.83±0.1554.14±0.6090.83±0.1655.58±0.50
DP-FedSAM-topk93.25±0.6054.85±0.8692.60±0.6557.00±0.6991.52±0.1158.82±0.51
", + "bbox": [ + 120, + 107, + 849, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 4. The proposed DP-FedSAM can achieve a tighter bound in general non-convex setting compared with previous work [9, 22] $\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} +\\frac{6K\\sigma_g^2 + \\sigma_l^2}{T} +\\frac{B^2\\sum_{t = 1}^{T}(\\overline{\\alpha}^t + \\tilde{\\alpha}^t)}{T} +\\frac{L^2\\sqrt{T}\\sigma^2C^2d}{m^2\\sqrt{K}}\\right)$", + "bbox": [ + 75, + 303, + 468, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in [9] and $\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} + \\frac{3\\sigma_g^2 + 2\\sigma_l^2}{\\sqrt{KT}} + \\frac{4L^2\\sqrt{T}\\sigma^2C^2d}{m^2\\sqrt{K}}\\right)$ in [22]) and reduce the impacts of the local and global variance $\\sigma_l^2$ , $\\sigma_g^2$ . Meanwhile, we are the first to theoretically analyze these two impacts of both the on-average norm of local updates $\\overline{\\alpha}^t$ and local update inconsistency among clients $\\tilde{\\alpha}^t$ on convergence. And the negative impacts of $\\overline{\\alpha}^t$ and $\\tilde{\\alpha}^t$ are also significantly mitigated upon convergence compared with previous work [9] due to local SAM optimizer being adopted. It means that we effectively alleviate performance degradation caused by clipping operation in DP and achieve better performance under symmetric noise. This theoretical result has also been empirically verified on several real-world data (see Section 6.2 and 6.3).", + "bbox": [ + 75, + 373, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 595, + 209, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to verify the effectiveness of DP-FedSAM.", + "bbox": [ + 76, + 622, + 468, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 664, + 253, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and Data Partition. The efficacy of DP-FedSAM is evaluated on three datasets, including EMNIST [10], CIFAR-10 and CIFAR-100 [28], in both IID and Non-IID settings. Specifically, Dir Partition [19] is used for simulating Non-IID across federated clients, where the local data of each client is partitioned by splitting the total dataset through sampling the label ratios from the Dirichlet distribution $\\operatorname{Dir}(\\alpha)$ with parameters $\\alpha = 0.3$ and $\\alpha = 0.6$ .", + "bbox": [ + 75, + 686, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We focus on DPFL methods that ensure client-level DP. Thus we compare DP-FedSAM with existing DPFL baselines: DP-FedAvg [35] ensures client-level DP guarantee by directly employing Gaussian mechanism to the local updates. DP-FedAvg-blur [9] adds regularization method (BLUR) based on DP-FedAvg. DP-FedAvg-", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "blurs [9] uses local update sparsification (LUS) and BLUR for improving the performance of DP-FedAvg. Fed-SMP-rand $_k$ and Fed-SMP-top $_k$ [22] leverage random sparsification and top $_k$ sparsification technique for reducing the impact of DP noise on model accuracy, respectively.", + "bbox": [ + 496, + 303, + 890, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Configuration. For EMNIST, we use a simple CNN model and train 200 communication round. For CIFAR-10 and CIFAR-100 datasets, we use the ResNet-18 [18] backbone and train 300 communication round. For all experiments, we set the number of clients $M$ to 500. The default sample ratio $q$ of the client is 0.1. The local learning rate $\\eta$ is set to 0.1 with a decay rate 0.005 and momentum 0.5, and the number of training epochs is 30. For privacy parameters, noise multiplier $\\sigma$ is set to 0.95 and the privacy failure probability $\\delta = \\frac{1}{M}$ . The clipping threshold $C$ is selected by grid search from set $\\{0.1, 0.2, 0.4, 0.6, 0.8\\}$ , and we find that the gradient explosion phenomenon will occur when $C \\geq 0.6$ on EMNIST and $C = 0.2$ performs better on three datasets. The weight perturbation ratio is set to $\\rho = 0.5$ . We run each experiment 3 trials and report the best-averaged testing accuracy in each experiment.", + "bbox": [ + 496, + 380, + 892, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results on EMNIST and CIFAR-10 datasets are placed in the main paper, while the results on CIFAR-100 are placed in Appendix A. For a more comprehensive and fair comparison, we integrate DP-FedSAM with $\\mathrm{top}_k$ sparsification technique [9, 22], named DP-FedSAM-top $_k$ (More discussion in Appendix D), to compare with baselines that also use local sparsification technique, such as Fed-SMP-rand $_k$ , Fed-SMP-top $_k$ , and DP-FedAvg-blurs.", + "bbox": [ + 496, + 622, + 890, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Experiment Evaluation", + "text_level": 1, + "bbox": [ + 500, + 756, + 714, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance with compared baselines. In Table 1 and Figure 2, we evaluate DP-FedSAM and DP-FedSAM-top $_k$ on EMNIST and CIFAR-10 datasets in both settings compared with all baselines from DP-FedAvg to DP-FedAvg-blurs. The baseline methods seem to be overfitting in Table 1, especially on more complex data (e.g., CIFAR-100 in Table 4). The main reasons are the random noise introduced by DP and the over-fitting and inconsistency of", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24557", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65e144f882e104093d5e2dad3579fd9aaa4ccc592ce78d6dd04a108259f5d36a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 137, + 90, + 828, + 104 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6ba9c1b77c2d679337c614e1582da3b71202878ee98d7d8938747b0b58c66f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 109, + 352, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84a365a24a78eaba7ebe195cc404243fffa39fb27fabe2708e04d21c4b86f0b2.jpg", + "image_caption": [ + "(a) EMNIST" + ], + "image_footnote": [], + "bbox": [ + 359, + 109, + 609, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e532d1dd779901cf000973d06cf01b32c35a9e9900b5523d672c8b36202d9d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 109, + 867, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5f7a7630bcfb8fe2c8771d08177ba35012a6a70a068367606ef2c148b480bbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 263, + 352, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ceb12fa0dab2d1eae8c85bd0e06ed5c6b71abec4e01fd30aba1bc458cf822c3.jpg", + "image_caption": [ + "(b) CIFAR-10" + ], + "image_footnote": [], + "bbox": [ + 359, + 263, + 609, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60a6c0b4f8c25b9170556527db9eb89ecf8ae44e719ae6ff3c68a1fde9aea245.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 263, + 867, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the local models. From all these results, it is seen that our proposed algorithms outperform other baselines under symmetric noise both on accuracy and generalization perspectives. It means that we significantly improve the performance and generate a better trade-off between performance and privacy in DPFL. For instance, the averaged testing accuracy is $85.90\\%$ in DP-FedSAM and $87.70\\%$ in DP-FedSAM- $\\mathrm{top}_k$ on EMNIST in the IID setting, which are better than other baselines. Meanwhile, the difference between training accuracy and test accuracy is $9.71\\%$ in DP-FedSAM and $8.40\\%$ in DP-FedSAM- $\\mathrm{top}_k$ , while that is $17.74\\%$ in DP-FedAvg and $16.41\\%$ in Fed-SMP- $\\mathrm{top}_k$ , respectively. That means our algorithms significantly mitigate the performance degradation issue caused by DP.", + "bbox": [ + 75, + 443, + 468, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Non-IID levels. In the experiments under different participation cases as shown in Table 1, we further prove the robust generalization of the proposed algorithms. Heterogeneous data distribution of local clients is set to various participation levels from IID, Dirichlet 0.6, and Dirichlet 0.3, which makes the training of the global model more difficult. On EMNIST, as the Non-IID level decreases, DP-FedSAM achieves better generalization than DP-FedAvg, and the difference between training accuracy and test accuracy in DP-FedSAM $\\{19.47\\%, 10.75\\%, 9.71\\%$ are lower than that in DP-FedAvg $\\{26.18\\%, 17.35\\%, 17.74\\%\\}$ . Similarly, the difference values in DP-FedSAM-top $k$ $\\{17.50\\%, 11.07\\%, 8.40\\%$ are also lower than that in FedSMP-top $k$ $\\{23.56\\%, 16.31\\%, 16.41\\%$ . These observations confirm that our algorithms are more robust than baselines in various degrees of heterogeneous data.", + "bbox": [ + 75, + 657, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cb500fde4b4cddb7f0108917551b8b9db7049c4386b98117150111f77a84017.jpg", + "image_caption": [ + "Figure 2. The averaged testing accuracy on EMNIST and CIFAR-10 dataset under symmetric noise for all compared methods." + ], + "image_footnote": [], + "bbox": [ + 501, + 440, + 696, + 558 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b57967edc2b16b273e31020dd04b071e9d5afdf839bedcbe82fd7351458e7ac.jpg", + "image_caption": [ + "Figure 3. Norm distribution and average norm of local updates." + ], + "image_footnote": [], + "bbox": [ + 697, + 440, + 893, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Discussion for DP with SAM in FL", + "text_level": 1, + "bbox": [ + 500, + 598, + 802, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we empirically discuss how SAM mitigates the negative impacts of DP from the aspects of the norm of local update and the visualization of loss landscape and contour. Meanwhile, we also observe the training performance with SAM under different privacy budgets $\\epsilon$ compared with all baselines. These experiments are conducted on CIFAR-10 with ResNet-18 [18] and Dirichlet $\\alpha = 0.6$ .", + "bbox": [ + 496, + 623, + 890, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The norm of local update. To validate the theoretical results for mitigating the adverse impacts of the norm of local updates, we conduct experiments on DP-FedSAM and DP-FedAvg with clipping threshold $C = 0.2$ as shown in Figure 3. We show the norm $\\Delta_i^t$ distribution and average norm $\\overline{\\Delta}^t$ of local updates before clipping during the communication rounds. In contrast to DP-FedAvg, most of the norm is distributed at the smaller value in our scheme in Figure 3 (a), which means that the clipping operation drops less information. Meanwhile, the on-average norm $\\overline{\\Delta}^t$ is smaller than DP-FedAvg as shown in Figure 3 (b). These observations", + "bbox": [ + 496, + 729, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24558", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7aae472875ddb2923a244935cd0bc9051df354467d317f19d4935c1babe152ee.jpg", + "image_caption": [ + "Figure 4. Impact of hyper-parameters: perturbation radius $\\rho$ , local iteration steps $K$ , total clients size $M$ ." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 349, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa9e6e717a03e344c4d35fad6ccb310b006bfd9c1ee772f29b0272c762142daa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 90, + 609, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7cd9917d068fd62f6854cf4ab6f8fd35a5d38b31a69b1702db84568a48d03a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 90, + 864, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/09279c4f0bb5dab4c3d26ad989bdaac74dffc934a72d462b4263567a61eee010.jpg", + "table_caption": [ + "Table 2. Performance comparison under different privacy budgets." + ], + "table_footnote": [], + "table_body": "
AlgorithmAveraged test accuracy (%) under different privacy budgets ε
ε = 4ε = 6ε = 8ε = 10
DP-FedAvg38.23 ± 0.1543.87 ± 0.6246.74 ± 0.0349.06 ± 0.49
Fed-SMP-randk33.78 ± 0.9242.21 ± 0.2148.20 ± 0.0550.62 ± 0.14
Fed-SMP-topk38.99 ± 0.5046.24 ± 0.8049.78 ± 0.7852.51 ± 0.83
DP-FedAvg-blur38.23 ± 0.7043.93 ± 0.4846.74 ± 0.9249.06 ± 0.13
DP-FedAvg-blurs39.39 ± 0.4346.64 ± 0.3650.18 ± 0.2752.91 ± 0.57
DP-FedSAM39.89 ± 0.1747.92 ± 0.2351.30 ± 0.9553.18 ± 0.40
DP-FedSAM-topk38.96 ± 0.6149.17 ± 0.1553.64 ± 0.1256.36 ± 0.36
", + "bbox": [ + 81, + 273, + 467, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are also consistent with our theoretical results in Section 5. Performance under different privacy budgets $\\epsilon$ . Table 2 shows the test accuracies for the different level privacy guarantees. Note that $\\epsilon$ is not a hyper-parameter, but can be obtained by the privacy design in each round such as Eq. (13). Our methods consistently outperform the previous SOTA methods under various privacy budgets $\\epsilon$ . Specifically, DP-FedSAM and DP-FedSAM-top $k$ significantly improve the accuracy of DP-FedAvg and Fed-SMP-top $k$ by $1\\% \\sim 4\\%$ and $3\\% \\sim 4\\%$ under the same $\\epsilon$ , respectively. Furthermore, the test accuracy can improve as the privacy budget $\\epsilon$ increases, which suggests us a better balance is necessary between training performance and privacy. There is a better trade-off when achieving better accuracy under the same $\\epsilon$ or a smaller $\\epsilon$ under approximately the same accuracy.", + "bbox": [ + 75, + 383, + 467, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss landscape and contour. The discussion and visualizing results are placed in Appendix C due to limited space.", + "bbox": [ + 76, + 609, + 467, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 650, + 230, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We verify the influence of hyper-parameters in DPFedSAM on EMNIST with Dirichlet partition $\\alpha = 0.6$ .", + "bbox": [ + 75, + 674, + 467, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perturbation weight $\\rho$ . Perturbation weight $\\rho$ has an impact on performance as the adding perturbation is accumulated when the communication round $T$ increases. To select a proper value for our algorithms, we conduct some experiments on various perturbation radius from the set $\\{0.01, 0.1, 0.3, 0.5, 0.7, 1.0\\}$ in Figure 4 (a). As $\\rho = 0.5$ , we achieve better convergence and performance.", + "bbox": [ + 75, + 704, + 467, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Local iteration steps $K$ . Large local iteration steps $K$ can help the convergence in previous DPFL work [9] with the theoretical guarantees. To investigate the acceleration on $T$ by adopting a larger $K$ , we fix the total batchsize and change local training epochs. In Figure 4 (b), our algorithm can accelerate the convergence in Theorem 3 as a larger $K$", + "bbox": [ + 75, + 810, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/688dc3c9b765fef2e92ba1a1431895c4ef09ad64018fb91b86d5987d2bdd4958.jpg", + "table_caption": [ + "Table 3. The averaged training accuracy and testing accuracy." + ], + "table_footnote": [], + "table_body": "
AlgorithmTrain (%)Validation (%)Differential value (%)
DP-FedAvg99.55±0.0282.20± 0.3517.35±0.32
DP-FedSAM95.07±0.4584.32±0.19 ↑10.75±0.26 ↓
Fed-SMP-topk99.72±0.0283.41 ± 0.9116.31± 0.89
DP-FedSAM-topk95.87±0.5284.80±0.60 ↑11.07±0.08 ↓
", + "bbox": [ + 504, + 273, + 890, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is adopted, that is, use a larger epoch value. However, the adverse impact of clipping on training increases as $K$ is too large, e.g., epoch = 40. Thus we choose 30 local epoch.", + "bbox": [ + 498, + 352, + 890, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Client size $M$ . We compare the performance between different numbers of client participation $m = \\{100, 200, 300, 500, 700\\}$ with the same hyper-parameters in Figure 4 (c). Compared with larger $m = 500$ , the smaller $m$ may have worse performance due to large variance $\\sigma^2 C^2 / m$ in DP noise with the same setting. Meanwhile, when $m$ is too large such as $M = 700$ , the performance may degrade as the local data size decreases.", + "bbox": [ + 498, + 398, + 890, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of SAM. As shown in Table 3, it is seen that DP-FedSAM and DP-FedSAM-top $k_{i}$ can achieve performance improvement and better generalization compared with DP-FedAvg and Fed-SMP-top $k_{i}$ as SAM optimizer is adopted with the same setting, respectively.", + "bbox": [ + 498, + 518, + 890, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 613, + 619, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focus on severe performance degradation issue caused by dropped model information and exacerbated model inconsistency. And we are the first to alleviate this issue from the optimizer perspective and propose a novel and effective framework DP-FedSAM with a flatter loss landscape. Meanwhile, we present the analysis in detail of how SAM mitigates the adverse impacts of DP and achieve a tighter bound on convergence. Moreover, it is the first analysis to combine the impacts of the on-average norm of local updates and local update consistency among clients on training, and simultaneously provide the experimental observations. Finally, empirical results also verify the superiority of our approach on several real-world data.", + "bbox": [ + 496, + 638, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work is supported by STI 2030—Major Projects (No. 2021ZD0201405), Shenzhen Philosophy and Social Science Foundation (Grant No. SZ2021B005).", + "bbox": [ + 498, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24559", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Abadi, Andy Chu, Ian J. Goodfellow, H. Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. CoRR, 2016.", + "[2] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning. In International Conference on Machine Learning, ICML, pages 10–32, 2022.", + "[3] Naman Agarwal, Ananda Theertha Suresh, Felix X. Yu, Sanjiv Kumar, and Brendan McMahan. cpSGD: Communication-efficient and differentially-private distributed SGD. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 7575-7586, Dec. 2018.", + "[4] Maksym Andriushchenko and Nicolas Flammarion. Towards understanding sharpness-aware minimization. In International Conference on Machine Learning, ICML, Proceedings of Machine Learning Research, pages 639-668. PMLR, 2022.", + "[5] Léon Bottou. Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT'2010, pages 177-186. Springer, 2010.", + "[6] Léon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning. Siam Review, 60(2):223-311, 2018.", + "[7] Debora Calderola, Barbara Caputo, and Marco Ciccone. Improving generalization in federated learning by seeking flat minima. CoRR, abs/2203.11834, 2022.", + "[8] Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. CoRR, 2022.", + "[9] Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[10] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. Emmist: Extending mnist to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 2921-2926. IEEE, 2017.", + "[11] Jiawei Du, Hanshu Yan, Jiashi Feng, Joey Tianyi Zhou, Liangli Zhen, Rick Siow Mong Goh, and Vincent Tan. Efficient sharpness-aware minimization for improved training of neural networks. In International Conference on Learning Representations, 2021.", + "[12] Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy. Found. Trends Theor. Comput. Sci., pages 211-407, Aug. 2014.", + "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021.", + "[14] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proc. ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 1322-1333, 2015." + ], + "bbox": [ + 76, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Robin C. Geyer, Tassilo Klein, and Moin Nabi. Differentially private federated learning: A client level perspective. CoRR, Aug. 2017.", + "[16] Robin C Geyer, Tassilo Klein, and Moin Nabi. Differentially private federated learning: A client level perspective. arXiv preprint arXiv:1712.07557, 2017.", + "[17] Saeed Ghadimi and Guanghui Lan. Stochastic first-and zeroth-order methods for nonconvex stochastic programming. SIAM Journal on Optimization, 23(4):2341-2368, 2013.", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[19] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019.", + "[20] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsification-amplified privacy and adaptive optimization. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1463-1469, Aug. 2021.", + "[21] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. CoRR, 2022.", + "[22] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. arXiv preprint arXiv:2202.07178, 2022.", + "[23] Tiansheng Huang, Shiwei Liu, Li Shen, Fengxiang He, Weiwei Lin, and Dacheng Tao. Achieving personalized federated learning with sparse local models. arXiv preprint arXiv:2201.11380, 2022.", + "[24] Zhuo Huang, Xiaobo Xia, Li Shen, Jun Yu, Chen Gong, Bo Han, and Tongliang Liu. Robust generalization against corruptions via worst-case sharpness minimization.", + "[25] Peter Kairouz, Ziyu Liu, and Thomas Steinke. The distributed discrete gaussian mechanism for federated learning with secure aggregation. In Proc. International Conference on Machine Learning (ICML), pages 5201-5212, Jul. 2021.", + "[26] P. Kairouz, Ziyu Liu, and T. Steinke. The distributed discrete gaussian mechanism for federated learning with secure aggregation. ArXiv, abs/2102.06387, 2021.", + "[27] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, pages 1-210, 2021.", + "[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.", + "[29] Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi. Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks. In International Conference on Machine Learning, pages 5905-5914. PMLR, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24560", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018.", + "[31] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE Signal Processing Magazine, pages 50-60, 2020.", + "[32] Yong Liu, Siqi Mai, Xiangning Chen, Cho-Jui Hsieh, and Yang You. Towards efficient and scalable sharpness-aware minimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12360-12370, 2022.", + "[33] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proc. International Conference on Artificial Intelligence and Statistics (AISTATS), volume 54, pages 1273-1282, April. 2017.", + "[34] H Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. Learning differentially private recurrent language models. arXiv preprint arXiv:1710.06963, 2017.", + "[35] H. Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. Learning differentially private recurrent language models. In Proc. International Conference on Learning Representations (ICLR), Apr. 2018.", + "[36] H. B. McMahan, D. Ramage, Kunal Talwar, and L. Zhang. Learning differentially private recurrent language models. In ICLR, 2018.", + "[37] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In Proc. IEEE Symposium on Security and Privacy (SP), pages 691-706, 2019.", + "[38] Peng Mi, Li Shen, Tianhe Ren, Yiyi Zhou, Xiaoshuai Sun, Rongrong Ji, and Dacheng Tao. Make sharpness-aware minimization stronger: A sparsified perturbation approach. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022.", + "[39] Ilya Mironov. Rényi differential privacy. In Proc. IEEE computer security foundations symposium (CSF), pages 263-275, 2017.", + "[40] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In Proc. IEEE Symposium on Security and Privacy (SP), pages 739-753, 2019.", + "[41] Zhe Qu, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. Generalized federated learning via sharpness aware minimization. In International Conference on Machine Learning, ICML, pages 18250-18280, 2022.", + "[42] Sashank J. Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečný, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In International Conference on Learning Representations, 2021.", + "[43] Yifan Shi, Li Shen, Kang Wei, Yan Sun, Bo Yuan, Xueqian Wang, and Dacheng Tao. Improving the model con" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sistency of decentralized federated learning. arXiv preprint arXiv:2302.04083, 2023.", + "[44] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In Proc. IEEE Symposium on Security and Privacy (SP), pages 3-18, 2017.", + "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[46] Hao Sun, Li Shen, Qihuang Zhong, Liang Ding, Shixiang Chen, Jingwei Sun, Jing Li, Guangzhong Sun, and Dacheng Tao. Adasam: Boosting sharpness-aware minimization with adaptive learning rate and momentum for training deep neural networks. arXiv preprint arXiv:2303.00565, 2023.", + "[47] Lichao Sun and Lingjuan Lyu. Federated model distillation with noise-free differential privacy. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1563-1570, Aug. 2021.", + "[48] Lichao Sun, Jianwei Qian, and Xun Chen. LDP-FL: Practical private aggregation in federated learning with local differential privacy. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1571–1578, Aug. 2021.", + "[49] Tao Sun, Dongsheng Li, and Bao Wang. Decentralized federated averaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.", + "[50] Yan Sun, Li Shen, Tiansheng Huang, Liang Ding, and Dacheng Tao. Fedspeed: Larger local interval, less communication round, and higher generalization accuracy. In International Conference on Learning Representations.", + "[51] Om Thakkar, Galen Andrew, and H. B. McMahan. Differentially private learning with adaptive clipping. ArXiv, abs/1905.03871, 2019.", + "[52] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Hang Su, Bo Zhang, and H. Vincent Poor. User-level privacy-preserving federated learning: Analysis and performance optimization. IEEE Transactions on Mobile Computing, 21(9):3388-3401, 2022.", + "[53] Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-IID federated learning. In International Conference on Learning Representations, 2021.", + "[54] Ashkan Yousefpour, Igor Shilov, Alexandre Sablayrolles, Davide Testuggine, Karthik Prasad, Mani Malek, John Nguyen, Sayan Gosh, Akash Bharadwaj, Jessica Zhao, Graham Cormode, and Ilya Mironov. Opacus: User-friendly differential privacy library in pytorch. In Proc. Privacy in Machine Learning (PriML) Workshop, NeurIPS, Virtual, Dec. 2021.", + "[55] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10174-10183, 2022.", + "[56] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In International Conference on Machine Learning, ICML, pages 26982-26992. PMLR, 2022." + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24561", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Qihuang Zhong, Liang Ding, Li Shen, Peng Mi, Juhua Liu, Bo Du, and Dacheng Tao. Improving sharpness-aware minimization with fisher mask for better generalization on language models. arXiv preprint arXiv:2210.05497, 2022.", + "[58] Yuqing Zhu, Xiang Yu, Yi-Hsuan Tsai, F. Pittaluga, M. Faraki, Manmohan Chandraker, and Yu-Xiang Wang. Voting-based approaches for differentially private federated learning. ArXiv, abs/2010.04851, 2020." + ], + "bbox": [ + 78, + 90, + 468, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24562", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_model.json b/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_model.json new file mode 100644 index 0000000000000000000000000000000000000000..899e357375b23d88cdd23e3fae63598e71d94281 --- /dev/null +++ b/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.836, + 0.154 + ], + "angle": 0, + "content": "Make Landscape Flatter in Differentially Private Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.179, + 0.836, + 0.235 + ], + "angle": 0, + "content": "Yifan Shi\\(^{1}\\) Yingqi Liu\\(^{2}\\) Kang Wei\\(^{2}\\) Li Shen\\(^{3,*}\\) Xueqian Wang\\(^{1,*}\\) Dacheng Tao\\(^{3}\\) \n\\(^{1}\\)Tsinghua University, Shenzhen, China; \\(^{3}\\)JD Explore Academy, Beijing, China \n\\(^{2}\\)Nanjing University of Science and Technology, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.237, + 0.797, + 0.253 + ], + "angle": 0, + "content": "shiyf21@mails.tsinghua.edu.cn; lyq@njust.edu.cn; kang.wei@njust.edu.cn;" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.255, + 0.794, + 0.269 + ], + "angle": 0, + "content": "mathshenli@gmail.com; wang.xq@sz.tsinghua.edu.cn; dacheng.tao@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.337, + 0.474, + 0.671 + ], + "angle": 0, + "content": "To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), client-level Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharper loss landscape and have poorer weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with better stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. From the theoretical perspective, we analyze in detail how DP-FedSAM mitigates the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with Rényi DP and present the sensitivity analysis of local updates. At last, we empirically confirm that our algorithm achieves state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.699, + 0.21, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.471, + 0.877 + ], + "angle": 0, + "content": "Federated Learning (FL) [31] allows distributed clients to collaboratively train a shared model without sharing data. However, FL faces severe dilemma of privacy leakage [27]. Recent works show that a curious server can also infer clients' privacy information such as membership and data features, by well-designed generative models and/or shadow models [14, 37, 40, 44, 55]. To address this issue, differential privacy (DP) [12] has been introduced in FL, which can protect every instance in any client's data (instance-level DP [3,20,47,48]) or the information between" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.895, + 0.518 + ], + "angle": 0, + "content": "clients (client-level DP [8, 15, 21, 25, 35, 52]). In general, client-level DP are more suitable to apply in the real-world setting due to a better model performance. For instance, a language prediction model with client-level DP [16, 35] has been applied on mobile devices by Google. In general, the Gaussian noise perturbation-based method is commonly adopted for ensuring the strong client-level DP. However, this method includes two operations in terms of clipping the \\( l_{2} \\) norm of local updates to a sensitivity threshold \\( C \\) and adding random noise proportional to the model size, whose standard deviation (STD) is also decided by \\( C \\). These steps may cause severe performance degradation dilemma [9, 22], especially on large-scale complex model [45], such as ResNet-18 [18], or with heterogeneous data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.896, + 0.734 + ], + "angle": 0, + "content": "The reasons behind this issue may be two-fold: (i) The useful information is dropped due to the clipping operation especially on small \\( C \\), which is contained in the local updates; (ii) The model inconsistency among local models is exacerbated as the addition of random noise severely damages local updates and leads to large variances between local models especially on large \\( C \\) [9]. To overcome these issues, existing works improve the performance via restricting the norm of local update [9] and leveraging local update sparsification technique [9, 22] to reduce the adverse impacts of clipping and adding amount of random noise. However, the model performance degradation remains significantly severe compared with FL methods without considering the privacy, e.g., FedAvg [33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Motivation. To further explore the reasons behind this phenomenon, we compare the structure of loss landscapes and surface contours [30] for FedAvg [33] and DP-FedAvg [15,35] on partitioned CIFAR-10 dataset [28] with Dirichlet distribution \\((\\alpha = 0.6)\\) and ResNet-18 backbone [18] in Figure 1 (a) and (b), respectively. Note that the convergence of DP-FedAvg is worse than FedAvg as its loss value is higher after a long communication round. Furthermore, FedAvg has a flatter landscape, whereas DP-FedAvg has a sharper one, resulting in poorer generalization ability (sharper minima, see Figure 1 (a)) and weight perturbation robustness" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.378, + 0.902 + ], + "angle": 0, + "content": "*Corresponding authors: Li Shen and Xueqian Wang" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24552" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.53, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.185, + 0.371, + 0.201 + ], + "angle": 0, + "content": "(a) Loss landscapes." + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.089, + 0.712, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.186, + 0.802, + 0.2 + ], + "angle": 0, + "content": "(b) Loss surface contours." + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.089, + 0.888, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.206, + 0.84, + 0.221 + ], + "angle": 0, + "content": "Figure 1. Loss landscapes and surface contours comparison between DP-FedAvg (left) and FedAvg (right), respectively." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.232, + 0.47, + 0.308 + ], + "angle": 0, + "content": "(see Figure 1 (b)), which is caused by the clipped local update information and exacerbated model inconsistency, respectively. Based on these observations, an interesting research question is: can we further overcome the severe performance degradation via making landscape flatter?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.473, + 0.744 + ], + "angle": 0, + "content": "To answer this question, we propose DP-FedSAM via gradient perturbation to improve model performance. Specifically, a local flat model is generated by using SAM optimizer [13] in each client, which leads to better stability. After that, a potentially global flat model can be generated by aggregating several flat local models, which results in higher generalization ability and better robustness to DP noise, thereby significantly improving the performance and achieving a better trade-off between performance and privacy. Theoretically, we present a tighter bound \\(\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} + \\frac{\\sum_{t=1}^{T} (\\bar{\\alpha}^t \\sigma_g^2 + \\bar{\\alpha}^t L^2)}{T^2} + \\frac{L^2 \\sqrt{T} \\sigma^2 C^2 d}{m^2 \\sqrt{K}}\\right)\\) for DP-FedSAM in the stochastic non-convex setting, where both \\(\\frac{1}{T} \\sum_{t=1}^{T} \\bar{\\alpha}^t\\) and \\(\\frac{1}{T} \\sum_{t=1}^{T} \\tilde{\\alpha}^t\\) are bounded constant, \\(K\\) and \\(T\\) is local iteration steps and communication rounds, respectively. Meanwhile, we present how SAM mitigates the impact of DP. For the clipping operation, DP-FedSAM reduces the \\(l_2\\) norm and the negative impact of the inconsistency among local updates on convergence. For adding noise operation, we obtain better weight perturbation robustness for reducing the performance damage caused by random noise, thereby having better robustness to DP noise. Empirically, we conduct extensive experiments on EMNIST, CIFAR-10, and CIFAR-100 datasets in both the independently and identically distributed (IID) and Non-IID settings. Furthermore, we observe the loss landscape, surface contour, and the norm distribution of local updates for exploring the intrinsic effects of SAM with DP, which together with the theoretical analysis confirms the effect of DP-FedSAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Contribution. The main contributions of our work are summarized as four-fold: (i) We propose a novel scheme DPFedSAM in DPFL to alleviate performance degradation issue from the optimizer perspective. (ii) We establish an improved convergence rate, which is tighter than the conventional bounds [9, 22] in the stochastic non-convex setting. Moreover, we provide rigorous privacy guarantees and sensitivity analysis. (iii) We are the first to in-depth analyze the roles of the on-average norm of local updates \\(\\overline{\\alpha}^t\\) and local update consistency among clients \\(\\tilde{\\alpha}^t\\) on convergence. (iv)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.895, + 0.279 + ], + "angle": 0, + "content": "We conduct extensive experiments to verify the effect of DP-FedSAM, which could achieve state-of-the-art (SOTA) performance compared with several strong DPFL baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.287, + 0.638, + 0.303 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Client-level DPFL. Client-level DPFL is the de-facto approach for protecting any client's data. DP-FedAvg [36] is the first to study in this setting, which trains a language prediction model in a mobile keyboard and ensures client-level DP guarantee by employing the Gaussian mechanism and composing privacy guarantees. After that, the work in [26,51] presents a comprehensive end-to-end system, which appropriately discretizes the data and adds discrete Gaussian noise before performing secure aggregation. Meanwhile, AE-DPFL [58] leverages the voting-based mechanism among the data labels instead of averaging the gradients to avoid dimension dependence and significantly reduce the communication cost. Fed-SMP [22] uses Sparsified Model Perturbation (SMP) to mitigate the impact of privacy protection on model accuracy. Different from the aforementioned methods, a recent study [9] revisits this issue and leverages Bounded Local Update Regularization (BLUR) and Local Update Sparsification (LUS) to restrict the norm of local updates and reduce the noise size before executing operations that guarantee DP, respectively. Nevertheless, severe performance degradation remains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.84 + ], + "angle": 0, + "content": "Sharpness Aware Minimization (SAM). SAM [13] is an effective optimizer for training deep learning (DL) models, which leverages the flatness geometry of the loss landscape to improve model generalization ability. Recently, [4] study the properties of SAM and provide convergence results of SAM for non-convex objectives. As a powerful optimizer, SAM and its variants have been applied to various machine learning (ML) tasks [2, 11, 24, 29, 32, 38, 43, 46, 50, 56, 57]. Specifically, [41], [50] and [7] integrate SAM to improve the generalization, and thus mitigate the distribution shift problem and achieve a new SOTA performance for FL. However, to the best of our knowledge, no efforts have been devoted to the empirical performance and theoretical analysis of SAM in DPFL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The most related works to this paper are DP-FedAvg in [35], Fed-SMP in [22], and DP-FedAvg with LUS and BLUR [9]. However, these works still suffer from inferior performance due to the exacerbated model inconsistency is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24553" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "sue among the clients caused by random noise. Therefore, different from existing works, we try to alleviate this issue by making the landscape flatter and weight perturbation ability more robust. Furthermore, another related work is FedSAM [41], which integrates SAM optimizer to enhance the flatness of the local model and achieves new SOTA performance for FL. On top of the aforementioned studies, we are the first to extend the SAM optimizer into the DPFL setting to effectively alleviate the performance degradation problem. Meanwhile, we provide the theoretical analysis for sensitivity, privacy, and convergence in the non-convex setting. Finally, we empirically confirm our theoretical results and the superiority of performance compared with existing SOTA methods in DPFL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.316, + 0.204, + 0.332 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.47, + 0.372 + ], + "angle": 0, + "content": "In this section, we first give the problem setup of FL, and then introduce the several terminologies in DP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.379, + 0.267, + 0.396 + ], + "angle": 0, + "content": "3.1. Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.403, + 0.47, + 0.48 + ], + "angle": 0, + "content": "Consider a general FL system consisting of \\(M\\) clients, in which each client owns its local dataset. Let \\(\\mathcal{D}_i^{\\mathrm{train}}\\), \\(\\mathcal{D}_i^{\\mathrm{val}}\\) and \\(\\mathcal{D}_i^{\\mathrm{test}}\\) denote the training dataset, validation dataset and testing dataset, held by client \\(i\\), respectively, where \\(i \\in \\mathcal{U} = \\{1, 2, \\dots, M\\}\\). Formally, the FL task is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.486, + 0.47, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} ^ {\\star} = \\underset {\\mathbf {w}} {\\arg \\min } \\sum_ {i \\in \\mathcal {U}} p _ {i} f _ {i} (\\mathbf {w}, \\mathcal {D} _ {i} ^ {\\text {t r a i n}}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.523, + 0.47, + 0.614 + ], + "angle": 0, + "content": "where \\( p_i = |\\mathcal{D}_i^{\\mathrm{train}}| / |\\mathcal{D}^{\\mathrm{train}}| \\geq 0 \\) with \\( \\sum_{i\\in \\mathcal{U}}p_i = 1 \\), and \\( f_{i}(\\cdot) \\) is the local loss function with \\( f_{i}(\\mathbf{w}) = F_{i}(\\mathbf{w};\\xi_{i}), \\xi_{i} \\) is a batch sample data in client \\( i \\). \\( |\\mathcal{D}_i^{\\mathrm{train}}| \\) is the size of training dataset \\( \\mathcal{D}_i^{\\mathrm{train}} \\) and \\( |\\mathcal{D}^{\\mathrm{train}}| = \\sum_{i\\in \\mathcal{U}}|\\mathcal{D}_i^{\\mathrm{train}}| \\) is the total size of training datasets, respectively. For the \\( i \\)-th client, a local model is learned on its private training data \\( \\mathcal{D}^{\\mathrm{train}} \\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.622, + 0.469, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {i} = \\mathbf {w} _ {i} ^ {t} - \\eta \\nabla f _ {i} \\left(\\mathbf {w} _ {i}, \\mathcal {D} _ {i} ^ {\\text {t r a i n}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.646, + 0.469, + 0.707 + ], + "angle": 0, + "content": "Generally, the local loss function \\( f_{i}(\\cdot) \\) has the same expression across each client. Then, the \\( M \\) associated clients collaboratively learn a global model \\( \\mathbf{w} \\) over the heterogeneous training data \\( \\mathcal{D}_i^{\\mathrm{train}} \\), \\( \\forall i \\in \\mathcal{U} \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.715, + 0.266, + 0.731 + ], + "angle": 0, + "content": "3.2. Differential Privacy" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.739, + 0.469, + 0.784 + ], + "angle": 0, + "content": "Differential Privacy (DP) [12] is a rigorous privacy notion for measuring privacy risk. In this paper, we consider a relaxed version: Rényi DP (RDP) [39]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.792, + 0.469, + 0.868 + ], + "angle": 0, + "content": "Definition 1. (Rényi DP, [39]). Given a real number \\( \\alpha \\in (1,\\infty) \\) and privacy parameter \\( \\rho \\geq 0 \\), a randomized mechanism \\( \\mathcal{M} \\) satisfies \\( (\\alpha ,\\rho) \\)-RDP if for any two neighboring datasets \\( \\mathcal{D} \\), \\( \\mathcal{D}' \\) that differ in a single record, the Rényi \\( \\alpha \\)-divergence between \\( \\mathcal{M}(\\mathcal{D}) \\) and \\( \\mathcal{M}(\\mathcal{D}') \\) satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.874, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\nD _ {\\alpha} \\left[ \\mathcal {M} (\\mathcal {D}) \\| \\mathcal {M} \\left(\\mathcal {D} ^ {\\prime}\\right) \\right] := \\frac {1}{\\alpha - 1} \\log \\mathbb {E} \\left[ \\left(\\frac {\\mathcal {M} (\\mathcal {D})}{\\mathcal {M} \\left(\\mathcal {D} ^ {\\prime}\\right)}\\right) ^ {\\alpha} \\right] \\leq \\rho , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.88, + 0.107 + ], + "angle": 0, + "content": "where the expectation is taken over the output of \\(\\mathcal{M}(\\mathcal{D}^{\\prime})\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.118, + 0.892, + 0.195 + ], + "angle": 0, + "content": "Rényi DP is a useful analytical tool to measure the privacy and accurately represent guarantees on the tails of the privacy loss, which is strictly stronger than \\((\\epsilon, \\delta)\\)-DP for \\(\\delta > 0\\). Thus, we provide the privacy analysis based on this tool for each user's privacy loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.205, + 0.892, + 0.267 + ], + "angle": 0, + "content": "Definition 2. (\\(l_{2}\\) Sensitivity). [9, Definition 2] Let \\(\\mathcal{F}\\) be a function, the \\(L_{2}\\)-sensitivity of \\(\\mathcal{F}\\) is defined as \\(\\mathcal{S} = \\max_{D \\simeq D'} \\| \\mathcal{F}(D) - \\mathcal{F}(D') \\|_2\\), where the maximization is taken over all pairs of adjacent datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.892, + 0.338 + ], + "angle": 0, + "content": "The sensitivity of a function \\(\\mathcal{F}\\) captures the magnitude which a single individual's data can change the function \\(\\mathcal{F}\\) in the worst case. Therefore, it plays a crucial role in determining the magnitude of noise required to ensure DP." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.349, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Definition 3. (Client-level DP) [34, Definition 1]. A randomized algorithm \\(\\mathcal{M}\\) is \\((\\epsilon, \\delta)\\)-DP if for any two adjacent datasets \\(U\\), \\(U'\\) constructed by adding or removing all records of any client, and every possible subset of outputs \\(O\\) satisfy the following inequality:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.437, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\Pr [ \\mathcal {M} (U) \\in O ] \\leq e ^ {\\epsilon} \\Pr [ \\mathcal {M} \\left(U ^ {\\prime}\\right) \\in O ] + \\delta . \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.467, + 0.892, + 0.513 + ], + "angle": 0, + "content": "In client-level DP, we aim to ensure participation information for any clients. Therefore, we need to make local updates similar whether one client participates or not." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.636, + 0.545 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.892, + 0.735 + ], + "angle": 0, + "content": "To revisit the severe performance degradation challenge in DPFL, we observe the loss landscapes and surface contours of FedAvg and DP-FedAvg in Figure 1 (a) and (b), respectively. And we find that the DPFL method has a sharper landscape with both poorer generalization ability and weight perturbation robustness than the FL method. It means that the DPFL method may result in poor flatness and make model sensitivity to noise perturbation. However, the study focusing on this issue remains unexplored. Therefore, we plan to face this challenge from the optimizer perspective by adopting a SAM optimizer in each client, dubbed DP-FedSAM, whose local loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.744, + 0.892, + 0.768 + ], + "angle": 0, + "content": "\\[\nf _ {i} (\\mathbf {w}) = \\mathbb {E} _ {\\xi \\sim \\mathcal {D} _ {i}} \\max _ {\\| \\delta_ {i} \\| _ {2} \\leq \\rho} F _ {i} \\left(\\mathbf {w} ^ {t, k} (i) + \\delta_ {i}; \\xi_ {i}\\right), \\quad i \\in \\mathcal {N}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}^{t,k}(i) + \\delta_{i}\\) is viewed as the perturbed model, \\(\\rho\\) is a predefined constant controlling the radius of the perturbation and \\(\\| \\cdot \\| _2\\) is a \\(l_{2}\\)-norm. Instead of searching for a solution via SGD [5, 6], SAM [13] aims to seek a solution in a flat region by adding a small perturbation, i.e., \\(w + \\delta\\) with more robust performance. Specifically, for each client \\(i\\in \\{1,2,\\dots,M\\}\\), each local iteration \\(k\\in \\{0,1,\\dots,K - 1\\}\\) in each communication round \\(t\\in \\{0,1,\\dots,T - 1\\}\\), the \\(k\\)-th" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24554" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.352, + 0.107 + ], + "angle": 0, + "content": "inner iteration in client \\(i\\) is performed as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.118, + 0.469, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} ^ {t, k + 1} (i) = \\mathbf {w} ^ {t, k} (i) - \\eta \\tilde {\\mathbf {g}} ^ {t, k} (i), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.14, + 0.469, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {g}} ^ {t, k} (i) = \\nabla F _ {i} \\left(\\mathbf {w} ^ {t, k} + \\delta \\left(\\mathbf {w} ^ {t, k}\\right); \\xi\\right), \\delta \\left(\\mathbf {w} ^ {t, k}\\right) = \\frac {\\rho \\mathbf {g} ^ {t , k}}{\\left\\| \\mathbf {g} ^ {t , k} \\right\\| _ {2}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.188, + 0.469, + 0.325 + ], + "angle": 0, + "content": "Where \\(\\delta(\\mathbf{w}^{t,k})\\) is calculated by using first-order Taylor expansion around \\(\\mathbf{w}^{t,k}\\) [13]. After that, we adopt a sampling mechanism, gradient clipping, and Gaussian noise adding to ensure client-level DP. Note that this sampling can amplify the privacy guarantee since it decreases the chances of leaking information about a particular individual [1]. Specifically, after sampling \\(m\\) clients with probability \\(q = m / M\\) at each communication round, which is important for measuring privacy loss. Then, we clip the local updates first:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.336, + 0.469, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\Delta} _ {i} ^ {t} = \\Delta_ {i} ^ {t} \\cdot \\min \\left(1, \\frac {C}{\\| \\Delta_ {i} ^ {t} \\| _ {2}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.377, + 0.469, + 0.407 + ], + "angle": 0, + "content": "After clipping, we add Gaussian noise to the local update for ensuring client-level DP as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.418, + 0.469, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Delta} _ {i} ^ {t} = \\tilde {\\Delta} _ {i} ^ {t} + \\mathcal {N} \\left(0, \\sigma^ {2} C ^ {2} \\cdot \\mathbf {I} _ {d} / m\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.447, + 0.469, + 0.508 + ], + "angle": 0, + "content": "With a given noise variance \\(\\sigma^2\\), the accumulative privacy budget \\(\\epsilon\\) can be calculated based on the sampled Gaussian mechanism [54]. Finally, we summarize the training procedures of DP-FedSAM in Algorithm 1." + }, + { + "type": "code_caption", + "bbox": [ + 0.088, + 0.529, + 0.273, + 0.542 + ], + "angle": 0, + "content": "Algorithm 1: DP-FedSAM" + }, + { + "type": "algorithm", + "bbox": [ + 0.082, + 0.545, + 0.468, + 0.891 + ], + "angle": 0, + "content": "Input: Total number of clients \\(M\\) sampling ratio of clients \\(q\\) total number of communication rounds \\(T\\) the clipping threshold \\(C\\) local learning rate \\(\\eta\\) and total number of the local iterates are \\(K\\). Output: Generate global model \\(\\mathbf{w}^T\\). \n1 Initialization: Randomly initialize the global model \\(\\mathbf{w}^0\\). \n2 for \\(t = 0\\) to \\(T - 1\\) do \n3 Sample a set of \\(m = qM\\) clients at random without replacement, denoted by \\(\\mathcal{W}^t\\). \n4 for client \\(i = 1\\) to \\(m\\) in parallel do \n5 for \\(k = 0\\) to \\(K - 1\\) do \n6 Update the global parameter as the local parameter \\(\\mathbf{w}^t(i) \\gets \\mathbf{w}^t\\). \n7 Sample a batch of local data \\(\\xi_i\\) and calculate local gradient \\(\\mathbf{g}^{t,k}(i) = \\nabla F_i(\\mathbf{w}^{t,k}(i); \\xi_i)\\). \n8 Gradient perturbation by Equation (7). Local iteration update by Equation (6). \n9 end \n11 \\(\\Delta_i^t = \\mathbf{w}^{t,K}(i) - \\mathbf{w}^{t,0}(i)\\). \n12 Clip and add noise for DP by Equation (9). \n13 Return \\(\\hat{\\Delta}^t(i)\\). \n14 end \n15 \\(\\mathbf{w}^{t + 1} \\gets \\mathbf{w}^t + \\frac{1}{m} \\sum_{i \\in \\mathcal{W}^t} \\hat{\\Delta}^t(i)\\). \n16 end" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Compared with existing DPFL methods [9, 22, 35], the benefits of DP-FedSAM lie in three-fold: (i) We introduce SAM into DPFL to alleviate severe performance degradation via seeking a flatness model in each client, which is caused by the exacerbated inconsistency of local models. Specifically, DP-FedSAM generates both better generalization ability and robustness to DP noise by making the global model flatter. (ii) In addition, we analyze in detail how DP-FedSAM mitigates the negative impacts of DP. Specifically, we theoretically analyze the convergence with the on-average norm of local updates \\(\\overline{\\alpha}^t\\) and local update consistency among clients \\(\\tilde{\\alpha}^t\\), and empirically confirm these results via observing the norm distribution and average norm of local updates (see Section 6.3). (iii) Meanwhile, we present the theories unifying the impacts of gradient perturbation \\(\\rho\\) in SAM, the on-average norm of local updates \\(\\overline{\\alpha}^t\\), and local update consistency among clients \\(\\tilde{\\alpha}^t\\) in clipping operation, and the variance of random noise \\(\\sigma^2 C^2 / m\\) upon the convergence rate in Section 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.696, + 0.407 + ], + "angle": 0, + "content": "5. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.476 + ], + "angle": 0, + "content": "In this section, we give a rigorous analysis of DP-FedSAM, including its sensitivity, privacy, and convergence rate. The detailed proof is placed in Appendix E. Below, we first give several necessary assumptions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.481, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Assumption 1. (Lipschitz smoothness). The function \\(F_{i}\\) is differentiable and \\(\\nabla F_{i}\\) is \\(L\\)-Lipschitz continuous, \\(\\forall i \\in \\{1,2,\\ldots,M\\}\\), i.e., \\(\\| \\nabla F_{i}(\\mathbf{x}) - \\nabla F_{i}(\\mathbf{y}) \\| \\leq L \\| \\mathbf{x} - \\mathbf{y} \\|\\), for all \\(\\mathbf{x}, \\mathbf{y} \\in \\mathbb{R}^{d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.547, + 0.892, + 0.661 + ], + "angle": 0, + "content": "Assumption 2. (Bounded variance). The gradient of the function \\( f_{i} \\) have \\( \\sigma_l \\)-bounded variance, i.e., \\( \\mathbb{E}_{\\xi_i}\\left\\| \\nabla F_i(\\mathbf{w}^k (i);\\xi_i) - \\nabla F_i(\\mathbf{w}(i))\\right\\|^2 \\leq \\sigma_l^2 \\), \\( \\forall i \\in \\{1,2,\\dots,M\\} \\), \\( k \\in \\{1,\\dots,K - 1\\} \\), the global variance is also bounded, i.e., \\( \\frac{1}{M}\\sum_{i = 1}^{M}\\| \\nabla f_{i}(\\mathbf{w}) - \\nabla f(\\mathbf{w})\\|^{2} \\leq \\sigma_{g}^{2} \\) for all \\( \\mathbf{w} \\in \\mathbb{R}^d \\). It is not hard to verify that the \\( \\sigma_g \\) is smaller than the homogeneity parameter \\( \\beta \\), i.e., \\( \\sigma_g^2 \\leq \\beta^2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.664, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Assumption 3. (Bounded gradient). For any \\(i \\in \\{1,2,\\ldots,M\\}\\) and \\(\\mathbf{w} \\in \\mathbb{R}^d\\), we have \\(\\|\\nabla f_i(\\mathbf{w})\\| \\leq B\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.7, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Assumption 4. (Unbiased Gradient Estimator). For any data sample \\(z\\) from \\(\\mathcal{D}_i\\) and \\(\\mathbf{w} \\in \\mathbb{R}^d\\), the local gradient estimator is unbiased, i.e., \\(\\mathbb{E}[\\nabla f_i(\\mathbf{w}; z)] = \\mathbb{E}[\\nabla f_i(\\mathbf{w})]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Note that the above assumptions are mild and commonly used in characterizing the convergence rate of FL [6,9,17,22,23,41-43,49,53]. Furthermore, gradient clipping operation in DL is often used to prevent the gradient explosion phenomenon, thereby the gradient is bounded. The technical difficulty for DP-FedSAM lies in: (i) how SAM mitigates the impact of DP; (ii) how to analyze in detail the impacts of the consistency among clients and the on-average norm of local updates caused by clipping operation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24555" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "5.1. Sensitivity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.47, + 0.204 + ], + "angle": 0, + "content": "At first, we study the sensitivity of local update \\(\\Delta_i^t\\) from any client \\(i\\in \\{1,2,\\dots,M\\}\\) before clipping at \\(t\\)-th communication round in DP-FedSAM. This upper bound of sensitivity can roughly measure the degree of privacy protection. Under Definition 2, the sensitivity can be denoted by \\(S_{\\Delta_i^t}\\) in client \\(i\\) at \\(t\\)-th communication round." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.471, + 0.307 + ], + "angle": 0, + "content": "Theorem 1. (Sensitivity). Denote \\(\\Delta_i^t (\\mathbf{x})\\) and \\(\\Delta_i^t (\\mathbf{y})\\) as the local update at \\(t\\)-th communication round, the model \\(\\mathbf{x}(i)\\) and \\(\\mathbf{y}(i)\\) is conducted on two sets which differ at only one sample. Assume the initial model parameter \\(\\mathbf{w}^t (i) = \\mathbf{x}^{t,0}(i) = \\mathbf{y}^{t,0}(i)\\). With the above assumptions, the expected squared sensitivity \\(S_{\\Delta_i^t}^2\\) of local update is upper bounded," + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.324, + 0.469, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\mathcal {S} _ {\\Delta_ {i} ^ {t}} ^ {2} \\leq \\frac {6 \\eta^ {2} \\rho^ {2} K L ^ {2} \\left(1 2 K ^ {2} L ^ {2} \\eta^ {2} + 1 0\\right)}{1 - 2 \\eta^ {2} L ^ {2} K} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.47, + 0.415 + ], + "angle": 0, + "content": "When the local adaptive learning rate satisfies \\(\\eta = \\mathcal{O}(1 / L\\sqrt{KT})\\) and the perturbation amplitude \\(\\rho\\) proportional to the learning rate, e.g., \\(\\rho = \\mathcal{O}\\left(\\frac{1}{\\sqrt{T}}\\right)\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.424, + 0.469, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\mathcal {S} _ {\\Delta_ {i} ^ {t}} ^ {2} \\leq \\mathcal {O} \\left(\\frac {1}{T ^ {2}}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.465, + 0.47, + 0.535 + ], + "angle": 0, + "content": "For comparison, we also present the expected squared sensitivity of local update with SGD in DPFL, that is \\(\\mathbb{E}\\mathcal{S}_{\\Delta_i^t,SGD}^2\\leq \\frac{6\\eta^2\\sigma_l^2K}{1 - 3\\eta^2KL^2}\\). Thus \\(\\mathbb{E}\\mathcal{S}_{\\Delta_i^t,SGD}^2\\leq \\mathcal{O}(\\frac{\\sigma_l^2}{KL^2T})\\) when \\(\\eta = \\mathcal{O}(1 / L\\sqrt{KT})\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.542, + 0.47, + 0.635 + ], + "angle": 0, + "content": "Remark 1. It is clearly seen that the upper bound in \\(\\mathbb{ES}_{\\Delta_i^t,SAM}^2\\) is tighter than that in \\(\\mathbb{ES}_{\\Delta_i^t,SGD}^2\\). From the perspective of privacy protection, it means DP-FedSAM has a better privacy guarantee than DP-FedAvg. Another perspective from local iteration, which means both better model consistency among clients and training stability." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.642, + 0.243, + 0.659 + ], + "angle": 0, + "content": "5.2. Privacy Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.665, + 0.47, + 0.711 + ], + "angle": 0, + "content": "To achieve client-level privacy protection, we derive the sensitivity of the aggregation process at first after clipping the local updates." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.719, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Lemma 1. The sensitivity of client-level DP in DP-FedSAM can be expressed as \\( C / m \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.756, + 0.47, + 0.787 + ], + "angle": 0, + "content": "Proof. Given two adjacent batches \\(\\mathcal{W}^t\\) and \\(\\mathcal{W}^{t,\\mathrm{adj}}\\), that \\(\\mathcal{W}^{t,\\mathrm{adj}}\\) has one more or less client, we have" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.794, + 0.469, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\frac {1}{m} \\sum_ {i \\in \\mathcal {W} ^ {t}} \\Delta_ {i} ^ {t} - \\frac {1}{m} \\sum_ {j \\in \\mathcal {W} ^ {t, \\mathrm {a d j}}} \\Delta_ {j} ^ {t} \\right\\| _ {2} = \\frac {1}{m} \\left\\| \\Delta_ {j ^ {\\prime}} ^ {t} \\right\\| _ {2} \\leq \\frac {C}{m}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.846, + 0.47, + 0.865 + ], + "angle": 0, + "content": "where \\(\\Delta_{j'}^t\\) is the local update of one more or less client." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Remark 2. The value of this sensitivity can determine the amount of variance for adding random noise." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.891, + 0.123 + ], + "angle": 0, + "content": "After adding Gaussian noise, we calculate the accumulative privacy budget [54] along with training as follows." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.129, + 0.891, + 0.16 + ], + "angle": 0, + "content": "Theorem 2. After \\(T\\) communication rounds, the accumulative privacy budget is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.165, + 0.892, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\epsilon = \\bar {\\epsilon} + \\frac {(\\alpha - 1) \\log (1 - \\frac {1}{\\alpha}) - \\log (\\alpha) - \\log (\\delta)}{\\alpha - 1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.202, + 0.545, + 0.214 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.22, + 0.892, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\bar {\\epsilon} = \\frac {T}{\\alpha - 1} \\ln \\mathbb {E} _ {z \\sim \\mu_ {0} (z)} \\left[ \\left(1 - q + \\frac {q \\mu_ {1} (z)}{\\mu_ {0} (z)}\\right) ^ {\\alpha} \\right], \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.257, + 0.892, + 0.332 + ], + "angle": 0, + "content": "and \\(q\\) is sample rate for client selection, \\(\\mu_0(z)\\) and \\(\\mu_{1}(z)\\) denote the Gaussian probability density function (PDF) of \\(\\mathcal{N}(0,\\sigma)\\) and the mixture of two Gaussian distributions \\(q\\mathcal{N}(1,\\sigma) + (1 - q)\\mathcal{N}(0,\\sigma)\\), respectively, \\(\\sigma\\) is the noise STD, \\(\\alpha\\) is a selectable variable." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.34, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Remark 3. It can be seen that a small sampling rate \\( q \\) can enhance the privacy guarantee by decreasing the privacy budget, but it may also degrade the training performance due to the number of participating clients being reduced in each communication round. So a better trade-off is needed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.423, + 0.706, + 0.438 + ], + "angle": 0, + "content": "5.3. Convergence Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.445, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Below, we give a convergence analysis of how DP-FedSAM mitigates the negative impacts of DP. The technical contribution also is combining the impacts of the on-average norm of local updates \\(\\overline{\\alpha}^t\\) and local update consistency among clients \\(\\tilde{\\alpha}^t\\) on the rate. Moreover, we also empirically confirm these results in Section 6.3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.543, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Theorem 3. Under assumptions 1-4, local learning rate satisfies \\(\\eta = \\mathcal{O}(1 / L\\sqrt{KT})\\) and \\(f^{*}\\) is denoted as the minimal value of \\(f\\), i.e., \\(f(x)\\geq f(x^{*}) = f^{*}\\) for all \\(x\\in \\mathbb{R}^d\\). When the perturbation amplitude \\(\\rho\\) is proportional to the learning rate, e.g., \\(\\rho = \\mathcal{O}(1 / \\sqrt{T})\\), the sequence of outputs \\(\\{\\mathbf{w}^t\\}\\) generated by Alg. 1, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.642, + 0.892, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\mathbb {E} \\left[ \\overline {{\\alpha}} ^ {t} \\left\\| \\nabla f \\left(\\mathbf {w} ^ {t}\\right) \\right\\| ^ {2} \\right] \\leq \\underbrace {\\mathcal {O} \\left(\\frac {2 L \\left(f \\left(\\mathbf {w} ^ {1}\\right) - f ^ {*}\\right)}{\\sqrt {K T}} + \\frac {L ^ {2} \\sigma_ {l} ^ {2}}{K T ^ {2}}\\right)} _ {\\text {F r o m F e d S A M}} \\\\ + \\underbrace {\\mathcal {O} \\left(\\frac {\\sum_ {t = 1} ^ {T} \\left(\\bar {\\alpha} ^ {t} \\sigma_ {g} ^ {2} + \\tilde {\\alpha} ^ {t} L ^ {2}\\right)}{T ^ {2}}\\right)} _ {\\text {C l i p p i n g}} + \\underbrace {\\mathcal {O} \\left(\\frac {L ^ {2} \\sqrt {T} \\sigma^ {2} C ^ {2} d}{m ^ {2} \\sqrt {K}}\\right)} _ {\\text {A d d i n g n o i s e}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.771, + 0.545, + 0.783 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.79, + 0.892, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} ^ {t} := \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\alpha_ {i} ^ {t} \\quad a n d \\quad \\tilde {\\alpha} ^ {t} := \\frac {1}{M} \\sum_ {i = 1} ^ {M} | \\alpha_ {i} ^ {t} - \\overline {{\\alpha_ {i} ^ {t}}} |, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.835, + 0.892, + 0.901 + ], + "angle": 0, + "content": "with \\(\\alpha_{i}^{t} = \\min (1,\\frac{C}{\\eta\\|\\sum_{k = 0}^{K - 1}\\tilde{\\mathbf{g}}^{t,k}(i)\\|})\\) , respectively. Note that \\(\\overline{\\alpha}^t\\) and \\(\\tilde{\\alpha}^t\\) measure the on-average norm of local updates and local update consistency among clients before clipping and adding noise operations in DP-FedSAM, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24556" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.104 + ], + "angle": 0, + "content": "Table 1. Averaged training accuracy (%) and testing accuracy (%) on two data in both IID and Non-IID settings for all compared methods." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.108, + 0.851, + 0.287 + ], + "angle": 0, + "content": "
TaskAlgorithmDirichlet 0.3Dirichlet 0.6IID
TrainValidationTrainValidationTrainValidation
EMNISTDP-FedAvg99.28±0.0273.10±0.1699.55±0.0282.20±0.3599.66±0.4081.90±0.86
Fed-SMP-randk99.24±0.0273.72±0.5399.71±0.0182.18±0.7399.71±0.6184.16±0.83
Fed-SMP-topk99.31±0.0475.75±0.3599.72±0.0283.41±0.9199.73±0.4083.32±0.52
DP-FedAvg-blur99.12±0.0273.71±0.0299.66±0.0083.20±0.0199.67±0.0382.92±0.49
DP-FedAvg-blurs99.63±0.0876.25±0.3599.72±0.0283.41±0.9199.74±0.4582.92±0.49
DP-FedSAM96.28±0.6476.81±0.8195.07±0.4584.32±0.1995.61±0.9485.90±0.72
DP-FedSAM-topk94.77±0.1177.27±0.6795.87±1.5284.80±0.6096.12±0.8587.70±0.83
CIFAR-10DP-FedAvg93.65±0.4747.98±0.2493.65±0.4250.05±0.4793.65±0.1550.90±0.86
Fed-SMP-randk95.46±0.4348.14±0.1295.36±0.0651.33±0.3695.36±0.0650.61±0.20
Fed-SMP-topk95.49±0.1449.93±2.2995.49±0.0954.11±0.8395.49±0.1053.30±0.45
DP-FedAvg-blur95.47±0.1247.66±0.0199.66±0.4251.05±0.0194.50±0.0552.56±0.47
DP-FedAvg-blurs96.79±0.5151.23±0.6699.72±0.0954.11±0.8396.45±0.3053.48±0.76
DP-FedSAM90.38±0.9053.92±0.5590.83±0.1554.14±0.6090.83±0.1655.58±0.50
DP-FedSAM-topk93.25±0.6054.85±0.8692.60±0.6557.00±0.6991.52±0.1158.82±0.51
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Remark 4. The proposed DP-FedSAM can achieve a tighter bound in general non-convex setting compared with previous work [9, 22] \\(\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} +\\frac{6K\\sigma_g^2 + \\sigma_l^2}{T} +\\frac{B^2\\sum_{t = 1}^{T}(\\overline{\\alpha}^t + \\tilde{\\alpha}^t)}{T} +\\frac{L^2\\sqrt{T}\\sigma^2C^2d}{m^2\\sqrt{K}}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.47, + 0.581 + ], + "angle": 0, + "content": "in [9] and \\(\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} + \\frac{3\\sigma_g^2 + 2\\sigma_l^2}{\\sqrt{KT}} + \\frac{4L^2\\sqrt{T}\\sigma^2C^2d}{m^2\\sqrt{K}}\\right)\\) in [22]) and reduce the impacts of the local and global variance \\(\\sigma_l^2\\), \\(\\sigma_g^2\\). Meanwhile, we are the first to theoretically analyze these two impacts of both the on-average norm of local updates \\(\\overline{\\alpha}^t\\) and local update inconsistency among clients \\(\\tilde{\\alpha}^t\\) on convergence. And the negative impacts of \\(\\overline{\\alpha}^t\\) and \\(\\tilde{\\alpha}^t\\) are also significantly mitigated upon convergence compared with previous work [9] due to local SAM optimizer being adopted. It means that we effectively alleviate performance degradation caused by clipping operation in DP and achieve better performance under symmetric noise. This theoretical result has also been empirically verified on several real-world data (see Section 6.2 and 6.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.597, + 0.21, + 0.614 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.623, + 0.47, + 0.653 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to verify the effectiveness of DP-FedSAM." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.254, + 0.682 + ], + "angle": 0, + "content": "6.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Dataset and Data Partition. The efficacy of DP-FedSAM is evaluated on three datasets, including EMNIST [10], CIFAR-10 and CIFAR-100 [28], in both IID and Non-IID settings. Specifically, Dir Partition [19] is used for simulating Non-IID across federated clients, where the local data of each client is partitioned by splitting the total dataset through sampling the label ratios from the Dirichlet distribution \\(\\operatorname{Dir}(\\alpha)\\) with parameters \\(\\alpha = 0.3\\) and \\(\\alpha = 0.6\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Baselines. We focus on DPFL methods that ensure client-level DP. Thus we compare DP-FedSAM with existing DPFL baselines: DP-FedAvg [35] ensures client-level DP guarantee by directly employing Gaussian mechanism to the local updates. DP-FedAvg-blur [9] adds regularization method (BLUR) based on DP-FedAvg. DP-FedAvg-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.381 + ], + "angle": 0, + "content": "blurs [9] uses local update sparsification (LUS) and BLUR for improving the performance of DP-FedAvg. Fed-SMP-rand\\(_k\\) and Fed-SMP-top\\(_k\\) [22] leverage random sparsification and top\\(_k\\) sparsification technique for reducing the impact of DP noise on model accuracy, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.893, + 0.622 + ], + "angle": 0, + "content": "Configuration. For EMNIST, we use a simple CNN model and train 200 communication round. For CIFAR-10 and CIFAR-100 datasets, we use the ResNet-18 [18] backbone and train 300 communication round. For all experiments, we set the number of clients \\( M \\) to 500. The default sample ratio \\( q \\) of the client is 0.1. The local learning rate \\( \\eta \\) is set to 0.1 with a decay rate 0.005 and momentum 0.5, and the number of training epochs is 30. For privacy parameters, noise multiplier \\( \\sigma \\) is set to 0.95 and the privacy failure probability \\( \\delta = \\frac{1}{M} \\). The clipping threshold \\( C \\) is selected by grid search from set \\( \\{0.1, 0.2, 0.4, 0.6, 0.8\\} \\), and we find that the gradient explosion phenomenon will occur when \\( C \\geq 0.6 \\) on EMNIST and \\( C = 0.2 \\) performs better on three datasets. The weight perturbation ratio is set to \\( \\rho = 0.5 \\). We run each experiment 3 trials and report the best-averaged testing accuracy in each experiment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.745 + ], + "angle": 0, + "content": "The results on EMNIST and CIFAR-10 datasets are placed in the main paper, while the results on CIFAR-100 are placed in Appendix A. For a more comprehensive and fair comparison, we integrate DP-FedSAM with \\(\\mathrm{top}_k\\) sparsification technique [9, 22], named DP-FedSAM-top\\(_k\\) (More discussion in Appendix D), to compare with baselines that also use local sparsification technique, such as Fed-SMP-rand\\(_k\\), Fed-SMP-top\\(_k\\), and DP-FedAvg-blurs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.715, + 0.772 + ], + "angle": 0, + "content": "6.2. Experiment Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Performance with compared baselines. In Table 1 and Figure 2, we evaluate DP-FedSAM and DP-FedSAM-top\\(_k\\) on EMNIST and CIFAR-10 datasets in both settings compared with all baselines from DP-FedAvg to DP-FedAvg-blurs. The baseline methods seem to be overfitting in Table 1, especially on more complex data (e.g., CIFAR-100 in Table 4). The main reasons are the random noise introduced by DP and the over-fitting and inconsistency of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24557" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.092, + 0.83, + 0.105 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.11, + 0.353, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.11, + 0.611, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.111, + 0.868, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.25, + 0.52, + 0.261 + ], + "angle": 0, + "content": "(a) EMNIST" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.264, + 0.353, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.264, + 0.61, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.265, + 0.868, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.404, + 0.524, + 0.416 + ], + "angle": 0, + "content": "(b) CIFAR-10" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.424, + 0.855, + 0.438 + ], + "angle": 0, + "content": "Figure 2. The averaged testing accuracy on EMNIST and CIFAR-10 dataset under symmetric noise for all compared methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.655 + ], + "angle": 0, + "content": "the local models. From all these results, it is seen that our proposed algorithms outperform other baselines under symmetric noise both on accuracy and generalization perspectives. It means that we significantly improve the performance and generate a better trade-off between performance and privacy in DPFL. For instance, the averaged testing accuracy is \\(85.90\\%\\) in DP-FedSAM and \\(87.70\\%\\) in DP-FedSAM-\\(\\mathrm{top}_k\\) on EMNIST in the IID setting, which are better than other baselines. Meanwhile, the difference between training accuracy and test accuracy is \\(9.71\\%\\) in DP-FedSAM and \\(8.40\\%\\) in DP-FedSAM-\\(\\mathrm{top}_k\\), while that is \\(17.74\\%\\) in DP-FedAvg and \\(16.41\\%\\) in Fed-SMP-\\(\\mathrm{top}_k\\), respectively. That means our algorithms significantly mitigate the performance degradation issue caused by DP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Impact of Non-IID levels. In the experiments under different participation cases as shown in Table 1, we further prove the robust generalization of the proposed algorithms. Heterogeneous data distribution of local clients is set to various participation levels from IID, Dirichlet 0.6, and Dirichlet 0.3, which makes the training of the global model more difficult. On EMNIST, as the Non-IID level decreases, DP-FedSAM achieves better generalization than DP-FedAvg, and the difference between training accuracy and test accuracy in DP-FedSAM \\(\\{19.47\\%, 10.75\\%, 9.71\\%\\) are lower than that in DP-FedAvg \\(\\{26.18\\%, 17.35\\%, 17.74\\%\\}\\). Similarly, the difference values in DP-FedSAM-top \\(k\\) \\(\\{17.50\\%, 11.07\\%, 8.40\\%\\) are also lower than that in FedSMP-top \\(k\\) \\(\\{23.56\\%, 16.31\\%, 16.41\\%\\) . These observations confirm that our algorithms are more robust than baselines in various degrees of heterogeneous data." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.441, + 0.697, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.441, + 0.895, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.571, + 0.884, + 0.585 + ], + "angle": 0, + "content": "Figure 3. Norm distribution and average norm of local updates." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.599, + 0.803, + 0.615 + ], + "angle": 0, + "content": "6.3. Discussion for DP with SAM in FL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.729 + ], + "angle": 0, + "content": "In this subsection, we empirically discuss how SAM mitigates the negative impacts of DP from the aspects of the norm of local update and the visualization of loss landscape and contour. Meanwhile, we also observe the training performance with SAM under different privacy budgets \\(\\epsilon\\) compared with all baselines. These experiments are conducted on CIFAR-10 with ResNet-18 [18] and Dirichlet \\(\\alpha = 0.6\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The norm of local update. To validate the theoretical results for mitigating the adverse impacts of the norm of local updates, we conduct experiments on DP-FedSAM and DP-FedAvg with clipping threshold \\( C = 0.2 \\) as shown in Figure 3. We show the norm \\( \\Delta_i^t \\) distribution and average norm \\( \\overline{\\Delta}^t \\) of local updates before clipping during the communication rounds. In contrast to DP-FedAvg, most of the norm is distributed at the smaller value in our scheme in Figure 3 (a), which means that the clipping operation drops less information. Meanwhile, the on-average norm \\( \\overline{\\Delta}^t \\) is smaller than DP-FedAvg as shown in Figure 3 (b). These observations" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24558" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.351, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.091, + 0.61, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.091, + 0.866, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.23, + 0.796, + 0.243 + ], + "angle": 0, + "content": "Figure 4. Impact of hyper-parameters: perturbation radius \\(\\rho\\), local iteration steps \\(K\\), total clients size \\(M\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.256, + 0.468, + 0.27 + ], + "angle": 0, + "content": "Table 2. Performance comparison under different privacy budgets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.274, + 0.468, + 0.374 + ], + "angle": 0, + "content": "
AlgorithmAveraged test accuracy (%) under different privacy budgets ε
ε = 4ε = 6ε = 8ε = 10
DP-FedAvg38.23 ± 0.1543.87 ± 0.6246.74 ± 0.0349.06 ± 0.49
Fed-SMP-randk33.78 ± 0.9242.21 ± 0.2148.20 ± 0.0550.62 ± 0.14
Fed-SMP-topk38.99 ± 0.5046.24 ± 0.8049.78 ± 0.7852.51 ± 0.83
DP-FedAvg-blur38.23 ± 0.7043.93 ± 0.4846.74 ± 0.9249.06 ± 0.13
DP-FedAvg-blurs39.39 ± 0.4346.64 ± 0.3650.18 ± 0.2752.91 ± 0.57
DP-FedSAM39.89 ± 0.1747.92 ± 0.2351.30 ± 0.9553.18 ± 0.40
DP-FedSAM-topk38.96 ± 0.6149.17 ± 0.1553.64 ± 0.1256.36 ± 0.36
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.468, + 0.61 + ], + "angle": 0, + "content": "are also consistent with our theoretical results in Section 5. Performance under different privacy budgets \\(\\epsilon\\). Table 2 shows the test accuracies for the different level privacy guarantees. Note that \\(\\epsilon\\) is not a hyper-parameter, but can be obtained by the privacy design in each round such as Eq. (13). Our methods consistently outperform the previous SOTA methods under various privacy budgets \\(\\epsilon\\). Specifically, DP-FedSAM and DP-FedSAM-top \\(k\\) significantly improve the accuracy of DP-FedAvg and Fed-SMP-top \\(k\\) by \\(1\\% \\sim 4\\%\\) and \\(3\\% \\sim 4\\%\\) under the same \\(\\epsilon\\), respectively. Furthermore, the test accuracy can improve as the privacy budget \\(\\epsilon\\) increases, which suggests us a better balance is necessary between training performance and privacy. There is a better trade-off when achieving better accuracy under the same \\(\\epsilon\\) or a smaller \\(\\epsilon\\) under approximately the same accuracy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.611, + 0.468, + 0.641 + ], + "angle": 0, + "content": "Loss landscape and contour. The discussion and visualizing results are placed in Appendix C due to limited space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.231, + 0.667 + ], + "angle": 0, + "content": "6.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.704 + ], + "angle": 0, + "content": "We verify the influence of hyper-parameters in DPFedSAM on EMNIST with Dirichlet partition \\(\\alpha = 0.6\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Perturbation weight \\(\\rho\\). Perturbation weight \\(\\rho\\) has an impact on performance as the adding perturbation is accumulated when the communication round \\(T\\) increases. To select a proper value for our algorithms, we conduct some experiments on various perturbation radius from the set \\(\\{0.01, 0.1, 0.3, 0.5, 0.7, 1.0\\}\\) in Figure 4 (a). As \\(\\rho = 0.5\\), we achieve better convergence and performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Local iteration steps \\( K \\). Large local iteration steps \\( K \\) can help the convergence in previous DPFL work [9] with the theoretical guarantees. To investigate the acceleration on \\( T \\) by adopting a larger \\( K \\), we fix the total batchsize and change local training epochs. In Figure 4 (b), our algorithm can accelerate the convergence in Theorem 3 as a larger \\( K \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.256, + 0.879, + 0.27 + ], + "angle": 0, + "content": "Table 3. The averaged training accuracy and testing accuracy." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.274, + 0.892, + 0.339 + ], + "angle": 0, + "content": "
AlgorithmTrain (%)Validation (%)Differential value (%)
DP-FedAvg99.55±0.0282.20± 0.3517.35±0.32
DP-FedSAM95.07±0.4584.32±0.19 ↑10.75±0.26 ↓
Fed-SMP-topk99.72±0.0283.41 ± 0.9116.31± 0.89
DP-FedSAM-topk95.87±0.5284.80±0.60 ↑11.07±0.08 ↓
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.353, + 0.892, + 0.398 + ], + "angle": 0, + "content": "is adopted, that is, use a larger epoch value. However, the adverse impact of clipping on training increases as \\( K \\) is too large, e.g., epoch = 40. Thus we choose 30 local epoch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.399, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Client size \\( M \\). We compare the performance between different numbers of client participation \\( m = \\{100, 200, 300, 500, 700\\} \\) with the same hyper-parameters in Figure 4 (c). Compared with larger \\( m = 500 \\), the smaller \\( m \\) may have worse performance due to large variance \\( \\sigma^2 C^2 / m \\) in DP noise with the same setting. Meanwhile, when \\( m \\) is too large such as \\( M = 700 \\), the performance may degrade as the local data size decreases." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Effect of SAM. As shown in Table 3, it is seen that DP-FedSAM and DP-FedSAM-top \\(k_{i}\\) can achieve performance improvement and better generalization compared with DP-FedAvg and Fed-SMP-top \\(k_{i}\\) as SAM optimizer is adopted with the same setting, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.614, + 0.62, + 0.63 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.836 + ], + "angle": 0, + "content": "In this paper, we focus on severe performance degradation issue caused by dropped model information and exacerbated model inconsistency. And we are the first to alleviate this issue from the optimizer perspective and propose a novel and effective framework DP-FedSAM with a flatter loss landscape. Meanwhile, we present the analysis in detail of how SAM mitigates the adverse impacts of DP and achieve a tighter bound on convergence. Moreover, it is the first analysis to combine the impacts of the on-average norm of local updates and local update consistency among clients on training, and simultaneously provide the experimental observations. Finally, empirical results also verify the superiority of our approach on several real-world data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement This work is supported by STI 2030—Major Projects (No. 2021ZD0201405), Shenzhen Philosophy and Social Science Foundation (Grant No. SZ2021B005)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24559" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Martin Abadi, Andy Chu, Ian J. Goodfellow, H. Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. CoRR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning. In International Conference on Machine Learning, ICML, pages 10–32, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.472, + 0.297 + ], + "angle": 0, + "content": "[3] Naman Agarwal, Ananda Theertha Suresh, Felix X. Yu, Sanjiv Kumar, and Brendan McMahan. cpSGD: Communication-efficient and differentially-private distributed SGD. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 7575-7586, Dec. 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.297, + 0.472, + 0.366 + ], + "angle": 0, + "content": "[4] Maksym Andriushchenko and Nicolas Flammarion. Towards understanding sharpness-aware minimization. In International Conference on Machine Learning, ICML, Proceedings of Machine Learning Research, pages 639-668. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.366, + 0.472, + 0.41 + ], + "angle": 0, + "content": "[5] Léon Bottou. Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT'2010, pages 177-186. Springer, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.472, + 0.452 + ], + "angle": 0, + "content": "[6] Léon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning. Siam Review, 60(2):223-311, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.472, + 0.494 + ], + "angle": 0, + "content": "[7] Debora Calderola, Barbara Caputo, and Marco Ciccone. Improving generalization in federated learning by seeking flat minima. CoRR, abs/2203.11834, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.472, + 0.537 + ], + "angle": 0, + "content": "[8] Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. CoRR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[9] Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.607, + 0.472, + 0.662 + ], + "angle": 0, + "content": "[10] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. Emmist: Extending mnist to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 2921-2926. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.662, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[11] Jiawei Du, Hanshu Yan, Jiashi Feng, Joey Tianyi Zhou, Liangli Zhen, Rick Siow Mong Goh, and Vincent Tan. Efficient sharpness-aware minimization for improved training of neural networks. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.732, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[12] Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy. Found. Trends Theor. Comput. Sci., pages 211-407, Aug. 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.776, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.831, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[14] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proc. ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 1322-1333, 2015." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Robin C. Geyer, Tassilo Klein, and Moin Nabi. Differentially private federated learning: A client level perspective. CoRR, Aug. 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[16] Robin C Geyer, Tassilo Klein, and Moin Nabi. Differentially private federated learning: A client level perspective. arXiv preprint arXiv:1712.07557, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[17] Saeed Ghadimi and Guanghui Lan. Stochastic first-and zeroth-order methods for nonconvex stochastic programming. SIAM Journal on Optimization, 23(4):2341-2368, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[19] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[20] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsification-amplified privacy and adaptive optimization. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1463-1469, Aug. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[21] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. CoRR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[22] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. arXiv preprint arXiv:2202.07178, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[23] Tiansheng Huang, Shiwei Liu, Li Shen, Fengxiang He, Weiwei Lin, and Dacheng Tao. Achieving personalized federated learning with sparse local models. arXiv preprint arXiv:2201.11380, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[24] Zhuo Huang, Xiaobo Xia, Li Shen, Jun Yu, Chen Gong, Bo Han, and Tongliang Liu. Robust generalization against corruptions via worst-case sharpness minimization." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[25] Peter Kairouz, Ziyu Liu, and Thomas Steinke. The distributed discrete gaussian mechanism for federated learning with secure aggregation. In Proc. International Conference on Machine Learning (ICML), pages 5201-5212, Jul. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[26] P. Kairouz, Ziyu Liu, and T. Steinke. The distributed discrete gaussian mechanism for federated learning with secure aggregation. ArXiv, abs/2102.06387, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[27] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, pages 1-210, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi. Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks. In International Conference on Machine Learning, pages 5905-5914. PMLR, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24560" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.135 + ], + "angle": 0, + "content": "[30] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[31] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE Signal Processing Magazine, pages 50-60, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[32] Yong Liu, Siqi Mai, Xiangning Chen, Cho-Jui Hsieh, and Yang You. Towards efficient and scalable sharpness-aware minimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12360-12370, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[33] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proc. International Conference on Artificial Intelligence and Statistics (AISTATS), volume 54, pages 1273-1282, April. 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[34] H Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. Learning differentially private recurrent language models. arXiv preprint arXiv:1710.06963, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[35] H. Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. Learning differentially private recurrent language models. In Proc. International Conference on Learning Representations (ICLR), Apr. 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[36] H. B. McMahan, D. Ramage, Kunal Talwar, and L. Zhang. Learning differentially private recurrent language models. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[37] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In Proc. IEEE Symposium on Security and Privacy (SP), pages 691-706, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[38] Peng Mi, Li Shen, Tianhe Ren, Yiyi Zhou, Xiaoshuai Sun, Rongrong Ji, and Dacheng Tao. Make sharpness-aware minimization stronger: A sparsified perturbation approach. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[39] Ilya Mironov. Rényi differential privacy. In Proc. IEEE computer security foundations symposium (CSF), pages 263-275, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[40] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In Proc. IEEE Symposium on Security and Privacy (SP), pages 739-753, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[41] Zhe Qu, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. Generalized federated learning via sharpness aware minimization. In International Conference on Machine Learning, ICML, pages 18250-18280, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[42] Sashank J. Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečný, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[43] Yifan Shi, Li Shen, Kang Wei, Yan Sun, Bo Yuan, Xueqian Wang, and Dacheng Tao. Improving the model con" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "sistency of decentralized federated learning. arXiv preprint arXiv:2302.04083, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[44] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In Proc. IEEE Symposium on Security and Privacy (SP), pages 3-18, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[46] Hao Sun, Li Shen, Qihuang Zhong, Liang Ding, Shixiang Chen, Jingwei Sun, Jing Li, Guangzhong Sun, and Dacheng Tao. Adasam: Boosting sharpness-aware minimization with adaptive learning rate and momentum for training deep neural networks. arXiv preprint arXiv:2303.00565, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[47] Lichao Sun and Lingjuan Lyu. Federated model distillation with noise-free differential privacy. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1563-1570, Aug. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[48] Lichao Sun, Jianwei Qian, and Xun Chen. LDP-FL: Practical private aggregation in federated learning with local differential privacy. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1571–1578, Aug. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[49] Tao Sun, Dongsheng Li, and Bao Wang. Decentralized federated averaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[50] Yan Sun, Li Shen, Tiansheng Huang, Liang Ding, and Dacheng Tao. Fedspeed: Larger local interval, less communication round, and higher generalization accuracy. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[51] Om Thakkar, Galen Andrew, and H. B. McMahan. Differentially private learning with adaptive clipping. ArXiv, abs/1905.03871, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[52] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Hang Su, Bo Zhang, and H. Vincent Poor. User-level privacy-preserving federated learning: Analysis and performance optimization. IEEE Transactions on Mobile Computing, 21(9):3388-3401, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[53] Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-IID federated learning. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[54] Ashkan Yousefpour, Igor Shilov, Alexandre Sablayrolles, Davide Testuggine, Karthik Prasad, Mani Malek, John Nguyen, Sayan Gosh, Akash Bharadwaj, Jessica Zhao, Graham Cormode, and Ilya Mironov. Opacus: User-friendly differential privacy library in pytorch. In Proc. Privacy in Machine Learning (PriML) Workshop, NeurIPS, Virtual, Dec. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[55] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10174-10183, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[56] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In International Conference on Machine Learning, ICML, pages 26982-26992. PMLR, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24561" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[57] Qihuang Zhong, Liang Ding, Li Shen, Peng Mi, Juhua Liu, Bo Du, and Dacheng Tao. Improving sharpness-aware minimization with fisher mask for better generalization on language models. arXiv preprint arXiv:2210.05497, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[58] Yuqing Zhu, Xiang Yu, Yi-Hsuan Tsai, F. Pittaluga, M. Faraki, Manmohan Chandraker, and Yu-Xiang Wang. Voting-based approaches for differentially private federated learning. ArXiv, abs/2010.04851, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24562" + } + ] +] \ No newline at end of file diff --git a/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_origin.pdf b/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c45da6da2295e91f060cec5a5aa15bd838c95a01 --- /dev/null +++ b/2023/Make Landscape Flatter in Differentially Private Federated Learning/9c1523d7-4b35-4c98-84fe-da971b2c5a48_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ee353a3b8faf5818e9efeb73855cfb708695b1fb5e530871bfd4ba03981dec3 +size 646094 diff --git a/2023/Make Landscape Flatter in Differentially Private Federated Learning/full.md b/2023/Make Landscape Flatter in Differentially Private Federated Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..53d0770d1724396de3a571ecbc9e18110c59639f --- /dev/null +++ b/2023/Make Landscape Flatter in Differentially Private Federated Learning/full.md @@ -0,0 +1,400 @@ +# Make Landscape Flatter in Differentially Private Federated Learning + +Yifan Shi $^{1}$ Yingqi Liu $^{2}$ Kang Wei $^{2}$ Li Shen $^{3,*}$ Xueqian Wang $^{1,*}$ Dacheng Tao $^{3}$ $^{1}$ Tsinghua University, Shenzhen, China; $^{3}$ JD Explore Academy, Beijing, China + $^{2}$ Nanjing University of Science and Technology, Nanjing, China + +shiyf21@mails.tsinghua.edu.cn; lyq@njust.edu.cn; kang.wei@njust.edu.cn; + +mathshenli@gmail.com; wang.xq@sz.tsinghua.edu.cn; dacheng.tao@gmail.com + +# Abstract + +To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), client-level Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharper loss landscape and have poorer weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with better stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. From the theoretical perspective, we analyze in detail how DP-FedSAM mitigates the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with Rényi DP and present the sensitivity analysis of local updates. At last, we empirically confirm that our algorithm achieves state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL. + +# 1. Introduction + +Federated Learning (FL) [31] allows distributed clients to collaboratively train a shared model without sharing data. However, FL faces severe dilemma of privacy leakage [27]. Recent works show that a curious server can also infer clients' privacy information such as membership and data features, by well-designed generative models and/or shadow models [14, 37, 40, 44, 55]. To address this issue, differential privacy (DP) [12] has been introduced in FL, which can protect every instance in any client's data (instance-level DP [3,20,47,48]) or the information between + +clients (client-level DP [8, 15, 21, 25, 35, 52]). In general, client-level DP are more suitable to apply in the real-world setting due to a better model performance. For instance, a language prediction model with client-level DP [16, 35] has been applied on mobile devices by Google. In general, the Gaussian noise perturbation-based method is commonly adopted for ensuring the strong client-level DP. However, this method includes two operations in terms of clipping the $l_{2}$ norm of local updates to a sensitivity threshold $C$ and adding random noise proportional to the model size, whose standard deviation (STD) is also decided by $C$ . These steps may cause severe performance degradation dilemma [9, 22], especially on large-scale complex model [45], such as ResNet-18 [18], or with heterogeneous data. + +The reasons behind this issue may be two-fold: (i) The useful information is dropped due to the clipping operation especially on small $C$ , which is contained in the local updates; (ii) The model inconsistency among local models is exacerbated as the addition of random noise severely damages local updates and leads to large variances between local models especially on large $C$ [9]. To overcome these issues, existing works improve the performance via restricting the norm of local update [9] and leveraging local update sparsification technique [9, 22] to reduce the adverse impacts of clipping and adding amount of random noise. However, the model performance degradation remains significantly severe compared with FL methods without considering the privacy, e.g., FedAvg [33]. + +Motivation. To further explore the reasons behind this phenomenon, we compare the structure of loss landscapes and surface contours [30] for FedAvg [33] and DP-FedAvg [15,35] on partitioned CIFAR-10 dataset [28] with Dirichlet distribution $(\alpha = 0.6)$ and ResNet-18 backbone [18] in Figure 1 (a) and (b), respectively. Note that the convergence of DP-FedAvg is worse than FedAvg as its loss value is higher after a long communication round. Furthermore, FedAvg has a flatter landscape, whereas DP-FedAvg has a sharper one, resulting in poorer generalization ability (sharper minima, see Figure 1 (a)) and weight perturbation robustness + +![](images/96397fdd979bf62e21e9567d7174fcfbd362e96d6fd8fb400071671d3f0e583f.jpg) +(a) Loss landscapes. + +![](images/6e2fb01e56f1e346f7dc14e8fda65fb56287e550f6faf37ab01b56057bd062d4.jpg) +Figure 1. Loss landscapes and surface contours comparison between DP-FedAvg (left) and FedAvg (right), respectively. + +![](images/8c7c4f04a74b93d8df36cabfb7bf9c8cebbb2511fe8677dd0a64553467e760e8.jpg) +(b) Loss surface contours. + +(see Figure 1 (b)), which is caused by the clipped local update information and exacerbated model inconsistency, respectively. Based on these observations, an interesting research question is: can we further overcome the severe performance degradation via making landscape flatter? + +To answer this question, we propose DP-FedSAM via gradient perturbation to improve model performance. Specifically, a local flat model is generated by using SAM optimizer [13] in each client, which leads to better stability. After that, a potentially global flat model can be generated by aggregating several flat local models, which results in higher generalization ability and better robustness to DP noise, thereby significantly improving the performance and achieving a better trade-off between performance and privacy. Theoretically, we present a tighter bound $\mathcal{O}\left(\frac{1}{\sqrt{KT}} + \frac{\sum_{t=1}^{T} (\bar{\alpha}^t \sigma_g^2 + \bar{\alpha}^t L^2)}{T^2} + \frac{L^2 \sqrt{T} \sigma^2 C^2 d}{m^2 \sqrt{K}}\right)$ for DP-FedSAM in the stochastic non-convex setting, where both $\frac{1}{T} \sum_{t=1}^{T} \bar{\alpha}^t$ and $\frac{1}{T} \sum_{t=1}^{T} \tilde{\alpha}^t$ are bounded constant, $K$ and $T$ is local iteration steps and communication rounds, respectively. Meanwhile, we present how SAM mitigates the impact of DP. For the clipping operation, DP-FedSAM reduces the $l_2$ norm and the negative impact of the inconsistency among local updates on convergence. For adding noise operation, we obtain better weight perturbation robustness for reducing the performance damage caused by random noise, thereby having better robustness to DP noise. Empirically, we conduct extensive experiments on EMNIST, CIFAR-10, and CIFAR-100 datasets in both the independently and identically distributed (IID) and Non-IID settings. Furthermore, we observe the loss landscape, surface contour, and the norm distribution of local updates for exploring the intrinsic effects of SAM with DP, which together with the theoretical analysis confirms the effect of DP-FedSAM. + +Contribution. The main contributions of our work are summarized as four-fold: (i) We propose a novel scheme DPFedSAM in DPFL to alleviate performance degradation issue from the optimizer perspective. (ii) We establish an improved convergence rate, which is tighter than the conventional bounds [9, 22] in the stochastic non-convex setting. Moreover, we provide rigorous privacy guarantees and sensitivity analysis. (iii) We are the first to in-depth analyze the roles of the on-average norm of local updates $\overline{\alpha}^t$ and local update consistency among clients $\tilde{\alpha}^t$ on convergence. (iv) + +We conduct extensive experiments to verify the effect of DP-FedSAM, which could achieve state-of-the-art (SOTA) performance compared with several strong DPFL baselines. + +# 2. Related work + +Client-level DPFL. Client-level DPFL is the de-facto approach for protecting any client's data. DP-FedAvg [36] is the first to study in this setting, which trains a language prediction model in a mobile keyboard and ensures client-level DP guarantee by employing the Gaussian mechanism and composing privacy guarantees. After that, the work in [26,51] presents a comprehensive end-to-end system, which appropriately discretizes the data and adds discrete Gaussian noise before performing secure aggregation. Meanwhile, AE-DPFL [58] leverages the voting-based mechanism among the data labels instead of averaging the gradients to avoid dimension dependence and significantly reduce the communication cost. Fed-SMP [22] uses Sparsified Model Perturbation (SMP) to mitigate the impact of privacy protection on model accuracy. Different from the aforementioned methods, a recent study [9] revisits this issue and leverages Bounded Local Update Regularization (BLUR) and Local Update Sparsification (LUS) to restrict the norm of local updates and reduce the noise size before executing operations that guarantee DP, respectively. Nevertheless, severe performance degradation remains. + +Sharpness Aware Minimization (SAM). SAM [13] is an effective optimizer for training deep learning (DL) models, which leverages the flatness geometry of the loss landscape to improve model generalization ability. Recently, [4] study the properties of SAM and provide convergence results of SAM for non-convex objectives. As a powerful optimizer, SAM and its variants have been applied to various machine learning (ML) tasks [2, 11, 24, 29, 32, 38, 43, 46, 50, 56, 57]. Specifically, [41], [50] and [7] integrate SAM to improve the generalization, and thus mitigate the distribution shift problem and achieve a new SOTA performance for FL. However, to the best of our knowledge, no efforts have been devoted to the empirical performance and theoretical analysis of SAM in DPFL. + +The most related works to this paper are DP-FedAvg in [35], Fed-SMP in [22], and DP-FedAvg with LUS and BLUR [9]. However, these works still suffer from inferior performance due to the exacerbated model inconsistency is + +sue among the clients caused by random noise. Therefore, different from existing works, we try to alleviate this issue by making the landscape flatter and weight perturbation ability more robust. Furthermore, another related work is FedSAM [41], which integrates SAM optimizer to enhance the flatness of the local model and achieves new SOTA performance for FL. On top of the aforementioned studies, we are the first to extend the SAM optimizer into the DPFL setting to effectively alleviate the performance degradation problem. Meanwhile, we provide the theoretical analysis for sensitivity, privacy, and convergence in the non-convex setting. Finally, we empirically confirm our theoretical results and the superiority of performance compared with existing SOTA methods in DPFL. + +# 3. Preliminary + +In this section, we first give the problem setup of FL, and then introduce the several terminologies in DP. + +# 3.1. Federated Learning + +Consider a general FL system consisting of $M$ clients, in which each client owns its local dataset. Let $\mathcal{D}_i^{\mathrm{train}}$ , $\mathcal{D}_i^{\mathrm{val}}$ and $\mathcal{D}_i^{\mathrm{test}}$ denote the training dataset, validation dataset and testing dataset, held by client $i$ , respectively, where $i \in \mathcal{U} = \{1, 2, \dots, M\}$ . Formally, the FL task is expressed as: + +$$ +\mathbf {w} ^ {\star} = \underset {\mathbf {w}} {\arg \min } \sum_ {i \in \mathcal {U}} p _ {i} f _ {i} (\mathbf {w}, \mathcal {D} _ {i} ^ {\text {t r a i n}}), \tag {1} +$$ + +where $p_i = |\mathcal{D}_i^{\mathrm{train}}| / |\mathcal{D}^{\mathrm{train}}| \geq 0$ with $\sum_{i\in \mathcal{U}}p_i = 1$ , and $f_{i}(\cdot)$ is the local loss function with $f_{i}(\mathbf{w}) = F_{i}(\mathbf{w};\xi_{i}), \xi_{i}$ is a batch sample data in client $i$ . $|\mathcal{D}_i^{\mathrm{train}}|$ is the size of training dataset $\mathcal{D}_i^{\mathrm{train}}$ and $|\mathcal{D}^{\mathrm{train}}| = \sum_{i\in \mathcal{U}}|\mathcal{D}_i^{\mathrm{train}}|$ is the total size of training datasets, respectively. For the $i$ -th client, a local model is learned on its private training data $\mathcal{D}^{\mathrm{train}}$ via: + +$$ +\mathbf {w} _ {i} = \mathbf {w} _ {i} ^ {t} - \eta \nabla f _ {i} \left(\mathbf {w} _ {i}, \mathcal {D} _ {i} ^ {\text {t r a i n}}\right). \tag {2} +$$ + +Generally, the local loss function $f_{i}(\cdot)$ has the same expression across each client. Then, the $M$ associated clients collaboratively learn a global model $\mathbf{w}$ over the heterogeneous training data $\mathcal{D}_i^{\mathrm{train}}$ , $\forall i \in \mathcal{U}$ . + +# 3.2. Differential Privacy + +Differential Privacy (DP) [12] is a rigorous privacy notion for measuring privacy risk. In this paper, we consider a relaxed version: Rényi DP (RDP) [39]. + +Definition 1. (Rényi DP, [39]). Given a real number $\alpha \in (1,\infty)$ and privacy parameter $\rho \geq 0$ , a randomized mechanism $\mathcal{M}$ satisfies $(\alpha ,\rho)$ -RDP if for any two neighboring datasets $\mathcal{D}$ , $\mathcal{D}'$ that differ in a single record, the Rényi $\alpha$ -divergence between $\mathcal{M}(\mathcal{D})$ and $\mathcal{M}(\mathcal{D}')$ satisfies: + +$$ +D _ {\alpha} \left[ \mathcal {M} (\mathcal {D}) \| \mathcal {M} \left(\mathcal {D} ^ {\prime}\right) \right] := \frac {1}{\alpha - 1} \log \mathbb {E} \left[ \left(\frac {\mathcal {M} (\mathcal {D})}{\mathcal {M} \left(\mathcal {D} ^ {\prime}\right)}\right) ^ {\alpha} \right] \leq \rho , \tag {3} +$$ + +where the expectation is taken over the output of $\mathcal{M}(\mathcal{D}^{\prime})$ + +Rényi DP is a useful analytical tool to measure the privacy and accurately represent guarantees on the tails of the privacy loss, which is strictly stronger than $(\epsilon, \delta)$ -DP for $\delta > 0$ . Thus, we provide the privacy analysis based on this tool for each user's privacy loss. + +Definition 2. ( $l_{2}$ Sensitivity). [9, Definition 2] Let $\mathcal{F}$ be a function, the $L_{2}$ -sensitivity of $\mathcal{F}$ is defined as $\mathcal{S} = \max_{D \simeq D'} \| \mathcal{F}(D) - \mathcal{F}(D') \|_2$ , where the maximization is taken over all pairs of adjacent datasets. + +The sensitivity of a function $\mathcal{F}$ captures the magnitude which a single individual's data can change the function $\mathcal{F}$ in the worst case. Therefore, it plays a crucial role in determining the magnitude of noise required to ensure DP. + +Definition 3. (Client-level DP) [34, Definition 1]. A randomized algorithm $\mathcal{M}$ is $(\epsilon, \delta)$ -DP if for any two adjacent datasets $U$ , $U'$ constructed by adding or removing all records of any client, and every possible subset of outputs $O$ satisfy the following inequality: + +$$ +\Pr [ \mathcal {M} (U) \in O ] \leq e ^ {\epsilon} \Pr [ \mathcal {M} \left(U ^ {\prime}\right) \in O ] + \delta . \tag {4} +$$ + +In client-level DP, we aim to ensure participation information for any clients. Therefore, we need to make local updates similar whether one client participates or not. + +# 4. Methodology + +To revisit the severe performance degradation challenge in DPFL, we observe the loss landscapes and surface contours of FedAvg and DP-FedAvg in Figure 1 (a) and (b), respectively. And we find that the DPFL method has a sharper landscape with both poorer generalization ability and weight perturbation robustness than the FL method. It means that the DPFL method may result in poor flatness and make model sensitivity to noise perturbation. However, the study focusing on this issue remains unexplored. Therefore, we plan to face this challenge from the optimizer perspective by adopting a SAM optimizer in each client, dubbed DP-FedSAM, whose local loss function is defined as: + +$$ +f _ {i} (\mathbf {w}) = \mathbb {E} _ {\xi \sim \mathcal {D} _ {i}} \max _ {\| \delta_ {i} \| _ {2} \leq \rho} F _ {i} \left(\mathbf {w} ^ {t, k} (i) + \delta_ {i}; \xi_ {i}\right), \quad i \in \mathcal {N}, \tag {5} +$$ + +where $\mathbf{w}^{t,k}(i) + \delta_{i}$ is viewed as the perturbed model, $\rho$ is a predefined constant controlling the radius of the perturbation and $\| \cdot \| _2$ is a $l_{2}$ -norm. Instead of searching for a solution via SGD [5, 6], SAM [13] aims to seek a solution in a flat region by adding a small perturbation, i.e., $w + \delta$ with more robust performance. Specifically, for each client $i\in \{1,2,\dots,M\}$ , each local iteration $k\in \{0,1,\dots,K - 1\}$ in each communication round $t\in \{0,1,\dots,T - 1\}$ , the $k$ -th + +inner iteration in client $i$ is performed as: + +$$ +\mathbf {w} ^ {t, k + 1} (i) = \mathbf {w} ^ {t, k} (i) - \eta \tilde {\mathbf {g}} ^ {t, k} (i), \tag {6} +$$ + +$$ +\tilde {\mathbf {g}} ^ {t, k} (i) = \nabla F _ {i} \left(\mathbf {w} ^ {t, k} + \delta \left(\mathbf {w} ^ {t, k}\right); \xi\right), \delta \left(\mathbf {w} ^ {t, k}\right) = \frac {\rho \mathbf {g} ^ {t , k}}{\left\| \mathbf {g} ^ {t , k} \right\| _ {2}}. \tag {7} +$$ + +Where $\delta(\mathbf{w}^{t,k})$ is calculated by using first-order Taylor expansion around $\mathbf{w}^{t,k}$ [13]. After that, we adopt a sampling mechanism, gradient clipping, and Gaussian noise adding to ensure client-level DP. Note that this sampling can amplify the privacy guarantee since it decreases the chances of leaking information about a particular individual [1]. Specifically, after sampling $m$ clients with probability $q = m / M$ at each communication round, which is important for measuring privacy loss. Then, we clip the local updates first: + +$$ +\tilde {\Delta} _ {i} ^ {t} = \Delta_ {i} ^ {t} \cdot \min \left(1, \frac {C}{\| \Delta_ {i} ^ {t} \| _ {2}}\right). \tag {8} +$$ + +After clipping, we add Gaussian noise to the local update for ensuring client-level DP as follows: + +$$ +\hat {\Delta} _ {i} ^ {t} = \tilde {\Delta} _ {i} ^ {t} + \mathcal {N} \left(0, \sigma^ {2} C ^ {2} \cdot \mathbf {I} _ {d} / m\right). \tag {9} +$$ + +With a given noise variance $\sigma^2$ , the accumulative privacy budget $\epsilon$ can be calculated based on the sampled Gaussian mechanism [54]. Finally, we summarize the training procedures of DP-FedSAM in Algorithm 1. + +Algorithm 1: DP-FedSAM +Input: Total number of clients $M$ sampling ratio of clients $q$ total number of communication rounds $T$ the clipping threshold $C$ local learning rate $\eta$ and total number of the local iterates are $K$ . Output: Generate global model $\mathbf{w}^T$ . +1 Initialization: Randomly initialize the global model $\mathbf{w}^0$ . +2 for $t = 0$ to $T - 1$ do +3 Sample a set of $m = qM$ clients at random without replacement, denoted by $\mathcal{W}^t$ . +4 for client $i = 1$ to $m$ in parallel do +5 for $k = 0$ to $K - 1$ do +6 Update the global parameter as the local parameter $\mathbf{w}^t(i) \gets \mathbf{w}^t$ . +7 Sample a batch of local data $\xi_i$ and calculate local gradient $\mathbf{g}^{t,k}(i) = \nabla F_i(\mathbf{w}^{t,k}(i); \xi_i)$ . +8 Gradient perturbation by Equation (7). Local iteration update by Equation (6). +9 end +11 $\Delta_i^t = \mathbf{w}^{t,K}(i) - \mathbf{w}^{t,0}(i)$ . +12 Clip and add noise for DP by Equation (9). +13 Return $\hat{\Delta}^t(i)$ . +14 end +15 $\mathbf{w}^{t + 1} \gets \mathbf{w}^t + \frac{1}{m} \sum_{i \in \mathcal{W}^t} \hat{\Delta}^t(i)$ . +16 end + +Compared with existing DPFL methods [9, 22, 35], the benefits of DP-FedSAM lie in three-fold: (i) We introduce SAM into DPFL to alleviate severe performance degradation via seeking a flatness model in each client, which is caused by the exacerbated inconsistency of local models. Specifically, DP-FedSAM generates both better generalization ability and robustness to DP noise by making the global model flatter. (ii) In addition, we analyze in detail how DP-FedSAM mitigates the negative impacts of DP. Specifically, we theoretically analyze the convergence with the on-average norm of local updates $\overline{\alpha}^t$ and local update consistency among clients $\tilde{\alpha}^t$ , and empirically confirm these results via observing the norm distribution and average norm of local updates (see Section 6.3). (iii) Meanwhile, we present the theories unifying the impacts of gradient perturbation $\rho$ in SAM, the on-average norm of local updates $\overline{\alpha}^t$ , and local update consistency among clients $\tilde{\alpha}^t$ in clipping operation, and the variance of random noise $\sigma^2 C^2 / m$ upon the convergence rate in Section 5. + +# 5. Theoretical Analysis + +In this section, we give a rigorous analysis of DP-FedSAM, including its sensitivity, privacy, and convergence rate. The detailed proof is placed in Appendix E. Below, we first give several necessary assumptions. + +Assumption 1. (Lipschitz smoothness). The function $F_{i}$ is differentiable and $\nabla F_{i}$ is $L$ -Lipschitz continuous, $\forall i \in \{1,2,\ldots,M\}$ , i.e., $\| \nabla F_{i}(\mathbf{x}) - \nabla F_{i}(\mathbf{y}) \| \leq L \| \mathbf{x} - \mathbf{y} \|$ , for all $\mathbf{x}, \mathbf{y} \in \mathbb{R}^{d}$ . + +Assumption 2. (Bounded variance). The gradient of the function $f_{i}$ have $\sigma_l$ -bounded variance, i.e., $\mathbb{E}_{\xi_i}\left\| \nabla F_i(\mathbf{w}^k (i);\xi_i) - \nabla F_i(\mathbf{w}(i))\right\|^2 \leq \sigma_l^2$ , $\forall i \in \{1,2,\dots,M\}$ , $k \in \{1,\dots,K - 1\}$ , the global variance is also bounded, i.e., $\frac{1}{M}\sum_{i = 1}^{M}\| \nabla f_{i}(\mathbf{w}) - \nabla f(\mathbf{w})\|^{2} \leq \sigma_{g}^{2}$ for all $\mathbf{w} \in \mathbb{R}^d$ . It is not hard to verify that the $\sigma_g$ is smaller than the homogeneity parameter $\beta$ , i.e., $\sigma_g^2 \leq \beta^2$ . + +Assumption 3. (Bounded gradient). For any $i \in \{1,2,\ldots,M\}$ and $\mathbf{w} \in \mathbb{R}^d$ , we have $\|\nabla f_i(\mathbf{w})\| \leq B$ . + +Assumption 4. (Unbiased Gradient Estimator). For any data sample $z$ from $\mathcal{D}_i$ and $\mathbf{w} \in \mathbb{R}^d$ , the local gradient estimator is unbiased, i.e., $\mathbb{E}[\nabla f_i(\mathbf{w}; z)] = \mathbb{E}[\nabla f_i(\mathbf{w})]$ . + +Note that the above assumptions are mild and commonly used in characterizing the convergence rate of FL [6,9,17,22,23,41-43,49,53]. Furthermore, gradient clipping operation in DL is often used to prevent the gradient explosion phenomenon, thereby the gradient is bounded. The technical difficulty for DP-FedSAM lies in: (i) how SAM mitigates the impact of DP; (ii) how to analyze in detail the impacts of the consistency among clients and the on-average norm of local updates caused by clipping operation. + +# 5.1. Sensitivity Analysis + +At first, we study the sensitivity of local update $\Delta_i^t$ from any client $i\in \{1,2,\dots,M\}$ before clipping at $t$ -th communication round in DP-FedSAM. This upper bound of sensitivity can roughly measure the degree of privacy protection. Under Definition 2, the sensitivity can be denoted by $S_{\Delta_i^t}$ in client $i$ at $t$ -th communication round. + +Theorem 1. (Sensitivity). Denote $\Delta_i^t (\mathbf{x})$ and $\Delta_i^t (\mathbf{y})$ as the local update at $t$ -th communication round, the model $\mathbf{x}(i)$ and $\mathbf{y}(i)$ is conducted on two sets which differ at only one sample. Assume the initial model parameter $\mathbf{w}^t (i) = \mathbf{x}^{t,0}(i) = \mathbf{y}^{t,0}(i)$ . With the above assumptions, the expected squared sensitivity $S_{\Delta_i^t}^2$ of local update is upper bounded, + +$$ +\mathbb {E} \mathcal {S} _ {\Delta_ {i} ^ {t}} ^ {2} \leq \frac {6 \eta^ {2} \rho^ {2} K L ^ {2} \left(1 2 K ^ {2} L ^ {2} \eta^ {2} + 1 0\right)}{1 - 2 \eta^ {2} L ^ {2} K} \tag {10} +$$ + +When the local adaptive learning rate satisfies $\eta = \mathcal{O}(1 / L\sqrt{KT})$ and the perturbation amplitude $\rho$ proportional to the learning rate, e.g., $\rho = \mathcal{O}\left(\frac{1}{\sqrt{T}}\right)$ , we have + +$$ +\mathbb {E} \mathcal {S} _ {\Delta_ {i} ^ {t}} ^ {2} \leq \mathcal {O} \left(\frac {1}{T ^ {2}}\right). \tag {11} +$$ + +For comparison, we also present the expected squared sensitivity of local update with SGD in DPFL, that is $\mathbb{E}\mathcal{S}_{\Delta_i^t,SGD}^2\leq \frac{6\eta^2\sigma_l^2K}{1 - 3\eta^2KL^2}$ . Thus $\mathbb{E}\mathcal{S}_{\Delta_i^t,SGD}^2\leq \mathcal{O}(\frac{\sigma_l^2}{KL^2T})$ when $\eta = \mathcal{O}(1 / L\sqrt{KT})$ . + +Remark 1. It is clearly seen that the upper bound in $\mathbb{ES}_{\Delta_i^t,SAM}^2$ is tighter than that in $\mathbb{ES}_{\Delta_i^t,SGD}^2$ . From the perspective of privacy protection, it means DP-FedSAM has a better privacy guarantee than DP-FedAvg. Another perspective from local iteration, which means both better model consistency among clients and training stability. + +# 5.2. Privacy Analysis + +To achieve client-level privacy protection, we derive the sensitivity of the aggregation process at first after clipping the local updates. + +Lemma 1. The sensitivity of client-level DP in DP-FedSAM can be expressed as $C / m$ . + +Proof. Given two adjacent batches $\mathcal{W}^t$ and $\mathcal{W}^{t,\mathrm{adj}}$ , that $\mathcal{W}^{t,\mathrm{adj}}$ has one more or less client, we have + +$$ +\left\| \frac {1}{m} \sum_ {i \in \mathcal {W} ^ {t}} \Delta_ {i} ^ {t} - \frac {1}{m} \sum_ {j \in \mathcal {W} ^ {t, \mathrm {a d j}}} \Delta_ {j} ^ {t} \right\| _ {2} = \frac {1}{m} \left\| \Delta_ {j ^ {\prime}} ^ {t} \right\| _ {2} \leq \frac {C}{m}, \tag {12} +$$ + +where $\Delta_{j'}^t$ is the local update of one more or less client. + +Remark 2. The value of this sensitivity can determine the amount of variance for adding random noise. + +After adding Gaussian noise, we calculate the accumulative privacy budget [54] along with training as follows. + +Theorem 2. After $T$ communication rounds, the accumulative privacy budget is calculated by: + +$$ +\epsilon = \bar {\epsilon} + \frac {(\alpha - 1) \log (1 - \frac {1}{\alpha}) - \log (\alpha) - \log (\delta)}{\alpha - 1}, \tag {13} +$$ + +where + +$$ +\bar {\epsilon} = \frac {T}{\alpha - 1} \ln \mathbb {E} _ {z \sim \mu_ {0} (z)} \left[ \left(1 - q + \frac {q \mu_ {1} (z)}{\mu_ {0} (z)}\right) ^ {\alpha} \right], \tag {14} +$$ + +and $q$ is sample rate for client selection, $\mu_0(z)$ and $\mu_{1}(z)$ denote the Gaussian probability density function (PDF) of $\mathcal{N}(0,\sigma)$ and the mixture of two Gaussian distributions $q\mathcal{N}(1,\sigma) + (1 - q)\mathcal{N}(0,\sigma)$ , respectively, $\sigma$ is the noise STD, $\alpha$ is a selectable variable. + +Remark 3. It can be seen that a small sampling rate $q$ can enhance the privacy guarantee by decreasing the privacy budget, but it may also degrade the training performance due to the number of participating clients being reduced in each communication round. So a better trade-off is needed. + +# 5.3. Convergence Analysis + +Below, we give a convergence analysis of how DP-FedSAM mitigates the negative impacts of DP. The technical contribution also is combining the impacts of the on-average norm of local updates $\overline{\alpha}^t$ and local update consistency among clients $\tilde{\alpha}^t$ on the rate. Moreover, we also empirically confirm these results in Section 6.3. + +Theorem 3. Under assumptions 1-4, local learning rate satisfies $\eta = \mathcal{O}(1 / L\sqrt{KT})$ and $f^{*}$ is denoted as the minimal value of $f$ , i.e., $f(x)\geq f(x^{*}) = f^{*}$ for all $x\in \mathbb{R}^d$ . When the perturbation amplitude $\rho$ is proportional to the learning rate, e.g., $\rho = \mathcal{O}(1 / \sqrt{T})$ , the sequence of outputs $\{\mathbf{w}^t\}$ generated by Alg. 1, we have: + +$$ +\begin{array}{l} \frac {1}{T} \sum_ {t = 1} ^ {T} \mathbb {E} \left[ \overline {{\alpha}} ^ {t} \left\| \nabla f \left(\mathbf {w} ^ {t}\right) \right\| ^ {2} \right] \leq \underbrace {\mathcal {O} \left(\frac {2 L \left(f \left(\mathbf {w} ^ {1}\right) - f ^ {*}\right)}{\sqrt {K T}} + \frac {L ^ {2} \sigma_ {l} ^ {2}}{K T ^ {2}}\right)} _ {\text {F r o m F e d S A M}} \\ + \underbrace {\mathcal {O} \left(\frac {\sum_ {t = 1} ^ {T} \left(\bar {\alpha} ^ {t} \sigma_ {g} ^ {2} + \tilde {\alpha} ^ {t} L ^ {2}\right)}{T ^ {2}}\right)} _ {\text {C l i p p i n g}} + \underbrace {\mathcal {O} \left(\frac {L ^ {2} \sqrt {T} \sigma^ {2} C ^ {2} d}{m ^ {2} \sqrt {K}}\right)} _ {\text {A d d i n g n o i s e}} \\ \end{array} +$$ + +where + +$$ +\bar {\alpha} ^ {t} := \frac {1}{M} \sum_ {i = 1} ^ {M} \alpha_ {i} ^ {t} \quad a n d \quad \tilde {\alpha} ^ {t} := \frac {1}{M} \sum_ {i = 1} ^ {M} | \alpha_ {i} ^ {t} - \overline {{\alpha_ {i} ^ {t}}} |, \tag {15} +$$ + +with $\alpha_{i}^{t} = \min (1,\frac{C}{\eta\|\sum_{k = 0}^{K - 1}\tilde{\mathbf{g}}^{t,k}(i)\|})$ , respectively. Note that $\overline{\alpha}^t$ and $\tilde{\alpha}^t$ measure the on-average norm of local updates and local update consistency among clients before clipping and adding noise operations in DP-FedSAM, respectively. + +Table 1. Averaged training accuracy (%) and testing accuracy (%) on two data in both IID and Non-IID settings for all compared methods. + +
TaskAlgorithmDirichlet 0.3Dirichlet 0.6IID
TrainValidationTrainValidationTrainValidation
EMNISTDP-FedAvg99.28±0.0273.10±0.1699.55±0.0282.20±0.3599.66±0.4081.90±0.86
Fed-SMP-randk99.24±0.0273.72±0.5399.71±0.0182.18±0.7399.71±0.6184.16±0.83
Fed-SMP-topk99.31±0.0475.75±0.3599.72±0.0283.41±0.9199.73±0.4083.32±0.52
DP-FedAvg-blur99.12±0.0273.71±0.0299.66±0.0083.20±0.0199.67±0.0382.92±0.49
DP-FedAvg-blurs99.63±0.0876.25±0.3599.72±0.0283.41±0.9199.74±0.4582.92±0.49
DP-FedSAM96.28±0.6476.81±0.8195.07±0.4584.32±0.1995.61±0.9485.90±0.72
DP-FedSAM-topk94.77±0.1177.27±0.6795.87±1.5284.80±0.6096.12±0.8587.70±0.83
CIFAR-10DP-FedAvg93.65±0.4747.98±0.2493.65±0.4250.05±0.4793.65±0.1550.90±0.86
Fed-SMP-randk95.46±0.4348.14±0.1295.36±0.0651.33±0.3695.36±0.0650.61±0.20
Fed-SMP-topk95.49±0.1449.93±2.2995.49±0.0954.11±0.8395.49±0.1053.30±0.45
DP-FedAvg-blur95.47±0.1247.66±0.0199.66±0.4251.05±0.0194.50±0.0552.56±0.47
DP-FedAvg-blurs96.79±0.5151.23±0.6699.72±0.0954.11±0.8396.45±0.3053.48±0.76
DP-FedSAM90.38±0.9053.92±0.5590.83±0.1554.14±0.6090.83±0.1655.58±0.50
DP-FedSAM-topk93.25±0.6054.85±0.8692.60±0.6557.00±0.6991.52±0.1158.82±0.51
+ +Remark 4. The proposed DP-FedSAM can achieve a tighter bound in general non-convex setting compared with previous work [9, 22] $\mathcal{O}\left(\frac{1}{\sqrt{KT}} +\frac{6K\sigma_g^2 + \sigma_l^2}{T} +\frac{B^2\sum_{t = 1}^{T}(\overline{\alpha}^t + \tilde{\alpha}^t)}{T} +\frac{L^2\sqrt{T}\sigma^2C^2d}{m^2\sqrt{K}}\right)$ + +in [9] and $\mathcal{O}\left(\frac{1}{\sqrt{KT}} + \frac{3\sigma_g^2 + 2\sigma_l^2}{\sqrt{KT}} + \frac{4L^2\sqrt{T}\sigma^2C^2d}{m^2\sqrt{K}}\right)$ in [22]) and reduce the impacts of the local and global variance $\sigma_l^2$ , $\sigma_g^2$ . Meanwhile, we are the first to theoretically analyze these two impacts of both the on-average norm of local updates $\overline{\alpha}^t$ and local update inconsistency among clients $\tilde{\alpha}^t$ on convergence. And the negative impacts of $\overline{\alpha}^t$ and $\tilde{\alpha}^t$ are also significantly mitigated upon convergence compared with previous work [9] due to local SAM optimizer being adopted. It means that we effectively alleviate performance degradation caused by clipping operation in DP and achieve better performance under symmetric noise. This theoretical result has also been empirically verified on several real-world data (see Section 6.2 and 6.3). + +# 6. Experiments + +In this section, we conduct extensive experiments to verify the effectiveness of DP-FedSAM. + +# 6.1. Experiment Setup + +Dataset and Data Partition. The efficacy of DP-FedSAM is evaluated on three datasets, including EMNIST [10], CIFAR-10 and CIFAR-100 [28], in both IID and Non-IID settings. Specifically, Dir Partition [19] is used for simulating Non-IID across federated clients, where the local data of each client is partitioned by splitting the total dataset through sampling the label ratios from the Dirichlet distribution $\operatorname{Dir}(\alpha)$ with parameters $\alpha = 0.3$ and $\alpha = 0.6$ . + +Baselines. We focus on DPFL methods that ensure client-level DP. Thus we compare DP-FedSAM with existing DPFL baselines: DP-FedAvg [35] ensures client-level DP guarantee by directly employing Gaussian mechanism to the local updates. DP-FedAvg-blur [9] adds regularization method (BLUR) based on DP-FedAvg. DP-FedAvg- + +blurs [9] uses local update sparsification (LUS) and BLUR for improving the performance of DP-FedAvg. Fed-SMP-rand $_k$ and Fed-SMP-top $_k$ [22] leverage random sparsification and top $_k$ sparsification technique for reducing the impact of DP noise on model accuracy, respectively. + +Configuration. For EMNIST, we use a simple CNN model and train 200 communication round. For CIFAR-10 and CIFAR-100 datasets, we use the ResNet-18 [18] backbone and train 300 communication round. For all experiments, we set the number of clients $M$ to 500. The default sample ratio $q$ of the client is 0.1. The local learning rate $\eta$ is set to 0.1 with a decay rate 0.005 and momentum 0.5, and the number of training epochs is 30. For privacy parameters, noise multiplier $\sigma$ is set to 0.95 and the privacy failure probability $\delta = \frac{1}{M}$ . The clipping threshold $C$ is selected by grid search from set $\{0.1, 0.2, 0.4, 0.6, 0.8\}$ , and we find that the gradient explosion phenomenon will occur when $C \geq 0.6$ on EMNIST and $C = 0.2$ performs better on three datasets. The weight perturbation ratio is set to $\rho = 0.5$ . We run each experiment 3 trials and report the best-averaged testing accuracy in each experiment. + +The results on EMNIST and CIFAR-10 datasets are placed in the main paper, while the results on CIFAR-100 are placed in Appendix A. For a more comprehensive and fair comparison, we integrate DP-FedSAM with $\mathrm{top}_k$ sparsification technique [9, 22], named DP-FedSAM-top $_k$ (More discussion in Appendix D), to compare with baselines that also use local sparsification technique, such as Fed-SMP-rand $_k$ , Fed-SMP-top $_k$ , and DP-FedAvg-blurs. + +# 6.2. Experiment Evaluation + +Performance with compared baselines. In Table 1 and Figure 2, we evaluate DP-FedSAM and DP-FedSAM-top $_k$ on EMNIST and CIFAR-10 datasets in both settings compared with all baselines from DP-FedAvg to DP-FedAvg-blurs. The baseline methods seem to be overfitting in Table 1, especially on more complex data (e.g., CIFAR-100 in Table 4). The main reasons are the random noise introduced by DP and the over-fitting and inconsistency of + +![](images/65e144f882e104093d5e2dad3579fd9aaa4ccc592ce78d6dd04a108259f5d36a.jpg) + +![](images/f6ba9c1b77c2d679337c614e1582da3b71202878ee98d7d8938747b0b58c66f2.jpg) + +![](images/84a365a24a78eaba7ebe195cc404243fffa39fb27fabe2708e04d21c4b86f0b2.jpg) +(a) EMNIST + +![](images/5e532d1dd779901cf000973d06cf01b32c35a9e9900b5523d672c8b36202d9d8.jpg) + +![](images/a5f7a7630bcfb8fe2c8771d08177ba35012a6a70a068367606ef2c148b480bbb.jpg) + +![](images/5ceb12fa0dab2d1eae8c85bd0e06ed5c6b71abec4e01fd30aba1bc458cf822c3.jpg) +(b) CIFAR-10 + +![](images/60a6c0b4f8c25b9170556527db9eb89ecf8ae44e719ae6ff3c68a1fde9aea245.jpg) + +the local models. From all these results, it is seen that our proposed algorithms outperform other baselines under symmetric noise both on accuracy and generalization perspectives. It means that we significantly improve the performance and generate a better trade-off between performance and privacy in DPFL. For instance, the averaged testing accuracy is $85.90\%$ in DP-FedSAM and $87.70\%$ in DP-FedSAM- $\mathrm{top}_k$ on EMNIST in the IID setting, which are better than other baselines. Meanwhile, the difference between training accuracy and test accuracy is $9.71\%$ in DP-FedSAM and $8.40\%$ in DP-FedSAM- $\mathrm{top}_k$ , while that is $17.74\%$ in DP-FedAvg and $16.41\%$ in Fed-SMP- $\mathrm{top}_k$ , respectively. That means our algorithms significantly mitigate the performance degradation issue caused by DP. + +Impact of Non-IID levels. In the experiments under different participation cases as shown in Table 1, we further prove the robust generalization of the proposed algorithms. Heterogeneous data distribution of local clients is set to various participation levels from IID, Dirichlet 0.6, and Dirichlet 0.3, which makes the training of the global model more difficult. On EMNIST, as the Non-IID level decreases, DP-FedSAM achieves better generalization than DP-FedAvg, and the difference between training accuracy and test accuracy in DP-FedSAM $\{19.47\%, 10.75\%, 9.71\%$ are lower than that in DP-FedAvg $\{26.18\%, 17.35\%, 17.74\%\}$ . Similarly, the difference values in DP-FedSAM-top $k$ $\{17.50\%, 11.07\%, 8.40\%$ are also lower than that in FedSMP-top $k$ $\{23.56\%, 16.31\%, 16.41\%$ . These observations confirm that our algorithms are more robust than baselines in various degrees of heterogeneous data. + +![](images/7cb500fde4b4cddb7f0108917551b8b9db7049c4386b98117150111f77a84017.jpg) +Figure 2. The averaged testing accuracy on EMNIST and CIFAR-10 dataset under symmetric noise for all compared methods. + +![](images/6b57967edc2b16b273e31020dd04b071e9d5afdf839bedcbe82fd7351458e7ac.jpg) +Figure 3. Norm distribution and average norm of local updates. + +# 6.3. Discussion for DP with SAM in FL + +In this subsection, we empirically discuss how SAM mitigates the negative impacts of DP from the aspects of the norm of local update and the visualization of loss landscape and contour. Meanwhile, we also observe the training performance with SAM under different privacy budgets $\epsilon$ compared with all baselines. These experiments are conducted on CIFAR-10 with ResNet-18 [18] and Dirichlet $\alpha = 0.6$ . + +The norm of local update. To validate the theoretical results for mitigating the adverse impacts of the norm of local updates, we conduct experiments on DP-FedSAM and DP-FedAvg with clipping threshold $C = 0.2$ as shown in Figure 3. We show the norm $\Delta_i^t$ distribution and average norm $\overline{\Delta}^t$ of local updates before clipping during the communication rounds. In contrast to DP-FedAvg, most of the norm is distributed at the smaller value in our scheme in Figure 3 (a), which means that the clipping operation drops less information. Meanwhile, the on-average norm $\overline{\Delta}^t$ is smaller than DP-FedAvg as shown in Figure 3 (b). These observations + +![](images/7aae472875ddb2923a244935cd0bc9051df354467d317f19d4935c1babe152ee.jpg) +Figure 4. Impact of hyper-parameters: perturbation radius $\rho$ , local iteration steps $K$ , total clients size $M$ . + +![](images/fa9e6e717a03e344c4d35fad6ccb310b006bfd9c1ee772f29b0272c762142daa.jpg) + +![](images/7cd9917d068fd62f6854cf4ab6f8fd35a5d38b31a69b1702db84568a48d03a26.jpg) + +Table 2. Performance comparison under different privacy budgets. + +
AlgorithmAveraged test accuracy (%) under different privacy budgets ε
ε = 4ε = 6ε = 8ε = 10
DP-FedAvg38.23 ± 0.1543.87 ± 0.6246.74 ± 0.0349.06 ± 0.49
Fed-SMP-randk33.78 ± 0.9242.21 ± 0.2148.20 ± 0.0550.62 ± 0.14
Fed-SMP-topk38.99 ± 0.5046.24 ± 0.8049.78 ± 0.7852.51 ± 0.83
DP-FedAvg-blur38.23 ± 0.7043.93 ± 0.4846.74 ± 0.9249.06 ± 0.13
DP-FedAvg-blurs39.39 ± 0.4346.64 ± 0.3650.18 ± 0.2752.91 ± 0.57
DP-FedSAM39.89 ± 0.1747.92 ± 0.2351.30 ± 0.9553.18 ± 0.40
DP-FedSAM-topk38.96 ± 0.6149.17 ± 0.1553.64 ± 0.1256.36 ± 0.36
+ +are also consistent with our theoretical results in Section 5. Performance under different privacy budgets $\epsilon$ . Table 2 shows the test accuracies for the different level privacy guarantees. Note that $\epsilon$ is not a hyper-parameter, but can be obtained by the privacy design in each round such as Eq. (13). Our methods consistently outperform the previous SOTA methods under various privacy budgets $\epsilon$ . Specifically, DP-FedSAM and DP-FedSAM-top $k$ significantly improve the accuracy of DP-FedAvg and Fed-SMP-top $k$ by $1\% \sim 4\%$ and $3\% \sim 4\%$ under the same $\epsilon$ , respectively. Furthermore, the test accuracy can improve as the privacy budget $\epsilon$ increases, which suggests us a better balance is necessary between training performance and privacy. There is a better trade-off when achieving better accuracy under the same $\epsilon$ or a smaller $\epsilon$ under approximately the same accuracy. + +Loss landscape and contour. The discussion and visualizing results are placed in Appendix C due to limited space. + +# 6.4. Ablation Study + +We verify the influence of hyper-parameters in DPFedSAM on EMNIST with Dirichlet partition $\alpha = 0.6$ . + +Perturbation weight $\rho$ . Perturbation weight $\rho$ has an impact on performance as the adding perturbation is accumulated when the communication round $T$ increases. To select a proper value for our algorithms, we conduct some experiments on various perturbation radius from the set $\{0.01, 0.1, 0.3, 0.5, 0.7, 1.0\}$ in Figure 4 (a). As $\rho = 0.5$ , we achieve better convergence and performance. + +Local iteration steps $K$ . Large local iteration steps $K$ can help the convergence in previous DPFL work [9] with the theoretical guarantees. To investigate the acceleration on $T$ by adopting a larger $K$ , we fix the total batchsize and change local training epochs. In Figure 4 (b), our algorithm can accelerate the convergence in Theorem 3 as a larger $K$ + +Table 3. The averaged training accuracy and testing accuracy. + +
AlgorithmTrain (%)Validation (%)Differential value (%)
DP-FedAvg99.55±0.0282.20± 0.3517.35±0.32
DP-FedSAM95.07±0.4584.32±0.19 ↑10.75±0.26 ↓
Fed-SMP-topk99.72±0.0283.41 ± 0.9116.31± 0.89
DP-FedSAM-topk95.87±0.5284.80±0.60 ↑11.07±0.08 ↓
+ +is adopted, that is, use a larger epoch value. However, the adverse impact of clipping on training increases as $K$ is too large, e.g., epoch = 40. Thus we choose 30 local epoch. + +Client size $M$ . We compare the performance between different numbers of client participation $m = \{100, 200, 300, 500, 700\}$ with the same hyper-parameters in Figure 4 (c). Compared with larger $m = 500$ , the smaller $m$ may have worse performance due to large variance $\sigma^2 C^2 / m$ in DP noise with the same setting. Meanwhile, when $m$ is too large such as $M = 700$ , the performance may degrade as the local data size decreases. + +Effect of SAM. As shown in Table 3, it is seen that DP-FedSAM and DP-FedSAM-top $k_{i}$ can achieve performance improvement and better generalization compared with DP-FedAvg and Fed-SMP-top $k_{i}$ as SAM optimizer is adopted with the same setting, respectively. + +# 7. Conclusion + +In this paper, we focus on severe performance degradation issue caused by dropped model information and exacerbated model inconsistency. And we are the first to alleviate this issue from the optimizer perspective and propose a novel and effective framework DP-FedSAM with a flatter loss landscape. Meanwhile, we present the analysis in detail of how SAM mitigates the adverse impacts of DP and achieve a tighter bound on convergence. Moreover, it is the first analysis to combine the impacts of the on-average norm of local updates and local update consistency among clients on training, and simultaneously provide the experimental observations. Finally, empirical results also verify the superiority of our approach on several real-world data. + +Acknowledgement This work is supported by STI 2030—Major Projects (No. 2021ZD0201405), Shenzhen Philosophy and Social Science Foundation (Grant No. SZ2021B005). + +# References + +[1] Martin Abadi, Andy Chu, Ian J. Goodfellow, H. Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. CoRR, 2016. +[2] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning. In International Conference on Machine Learning, ICML, pages 10–32, 2022. +[3] Naman Agarwal, Ananda Theertha Suresh, Felix X. Yu, Sanjiv Kumar, and Brendan McMahan. cpSGD: Communication-efficient and differentially-private distributed SGD. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 7575-7586, Dec. 2018. +[4] Maksym Andriushchenko and Nicolas Flammarion. Towards understanding sharpness-aware minimization. In International Conference on Machine Learning, ICML, Proceedings of Machine Learning Research, pages 639-668. PMLR, 2022. +[5] Léon Bottou. Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT'2010, pages 177-186. Springer, 2010. +[6] Léon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning. Siam Review, 60(2):223-311, 2018. +[7] Debora Calderola, Barbara Caputo, and Marco Ciccone. Improving generalization in federated learning by seeking flat minima. CoRR, abs/2203.11834, 2022. +[8] Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. CoRR, 2022. +[9] Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[10] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. Emmist: Extending mnist to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 2921-2926. IEEE, 2017. +[11] Jiawei Du, Hanshu Yan, Jiashi Feng, Joey Tianyi Zhou, Liangli Zhen, Rick Siow Mong Goh, and Vincent Tan. Efficient sharpness-aware minimization for improved training of neural networks. In International Conference on Learning Representations, 2021. +[12] Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy. Found. Trends Theor. Comput. Sci., pages 211-407, Aug. 2014. +[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. +[14] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proc. ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 1322-1333, 2015. + +[15] Robin C. Geyer, Tassilo Klein, and Moin Nabi. Differentially private federated learning: A client level perspective. CoRR, Aug. 2017. +[16] Robin C Geyer, Tassilo Klein, and Moin Nabi. Differentially private federated learning: A client level perspective. arXiv preprint arXiv:1712.07557, 2017. +[17] Saeed Ghadimi and Guanghui Lan. Stochastic first-and zeroth-order methods for nonconvex stochastic programming. SIAM Journal on Optimization, 23(4):2341-2368, 2013. +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[19] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. +[20] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsification-amplified privacy and adaptive optimization. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1463-1469, Aug. 2021. +[21] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. CoRR, 2022. +[22] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. arXiv preprint arXiv:2202.07178, 2022. +[23] Tiansheng Huang, Shiwei Liu, Li Shen, Fengxiang He, Weiwei Lin, and Dacheng Tao. Achieving personalized federated learning with sparse local models. arXiv preprint arXiv:2201.11380, 2022. +[24] Zhuo Huang, Xiaobo Xia, Li Shen, Jun Yu, Chen Gong, Bo Han, and Tongliang Liu. Robust generalization against corruptions via worst-case sharpness minimization. +[25] Peter Kairouz, Ziyu Liu, and Thomas Steinke. The distributed discrete gaussian mechanism for federated learning with secure aggregation. In Proc. International Conference on Machine Learning (ICML), pages 5201-5212, Jul. 2021. +[26] P. Kairouz, Ziyu Liu, and T. Steinke. The distributed discrete gaussian mechanism for federated learning with secure aggregation. ArXiv, abs/2102.06387, 2021. +[27] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, pages 1-210, 2021. +[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. +[29] Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi. Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks. In International Conference on Machine Learning, pages 5905-5914. PMLR, 2021. + +[30] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. +[31] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE Signal Processing Magazine, pages 50-60, 2020. +[32] Yong Liu, Siqi Mai, Xiangning Chen, Cho-Jui Hsieh, and Yang You. Towards efficient and scalable sharpness-aware minimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12360-12370, 2022. +[33] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proc. International Conference on Artificial Intelligence and Statistics (AISTATS), volume 54, pages 1273-1282, April. 2017. +[34] H Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. Learning differentially private recurrent language models. arXiv preprint arXiv:1710.06963, 2017. +[35] H. Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. Learning differentially private recurrent language models. In Proc. International Conference on Learning Representations (ICLR), Apr. 2018. +[36] H. B. McMahan, D. Ramage, Kunal Talwar, and L. Zhang. Learning differentially private recurrent language models. In ICLR, 2018. +[37] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In Proc. IEEE Symposium on Security and Privacy (SP), pages 691-706, 2019. +[38] Peng Mi, Li Shen, Tianhe Ren, Yiyi Zhou, Xiaoshuai Sun, Rongrong Ji, and Dacheng Tao. Make sharpness-aware minimization stronger: A sparsified perturbation approach. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. +[39] Ilya Mironov. Rényi differential privacy. In Proc. IEEE computer security foundations symposium (CSF), pages 263-275, 2017. +[40] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In Proc. IEEE Symposium on Security and Privacy (SP), pages 739-753, 2019. +[41] Zhe Qu, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. Generalized federated learning via sharpness aware minimization. In International Conference on Machine Learning, ICML, pages 18250-18280, 2022. +[42] Sashank J. Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečný, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In International Conference on Learning Representations, 2021. +[43] Yifan Shi, Li Shen, Kang Wei, Yan Sun, Bo Yuan, Xueqian Wang, and Dacheng Tao. Improving the model con + +sistency of decentralized federated learning. arXiv preprint arXiv:2302.04083, 2023. +[44] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In Proc. IEEE Symposium on Security and Privacy (SP), pages 3-18, 2017. +[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[46] Hao Sun, Li Shen, Qihuang Zhong, Liang Ding, Shixiang Chen, Jingwei Sun, Jing Li, Guangzhong Sun, and Dacheng Tao. Adasam: Boosting sharpness-aware minimization with adaptive learning rate and momentum for training deep neural networks. arXiv preprint arXiv:2303.00565, 2023. +[47] Lichao Sun and Lingjuan Lyu. Federated model distillation with noise-free differential privacy. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1563-1570, Aug. 2021. +[48] Lichao Sun, Jianwei Qian, and Xun Chen. LDP-FL: Practical private aggregation in federated learning with local differential privacy. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1571–1578, Aug. 2021. +[49] Tao Sun, Dongsheng Li, and Bao Wang. Decentralized federated averaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. +[50] Yan Sun, Li Shen, Tiansheng Huang, Liang Ding, and Dacheng Tao. Fedspeed: Larger local interval, less communication round, and higher generalization accuracy. In International Conference on Learning Representations. +[51] Om Thakkar, Galen Andrew, and H. B. McMahan. Differentially private learning with adaptive clipping. ArXiv, abs/1905.03871, 2019. +[52] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Hang Su, Bo Zhang, and H. Vincent Poor. User-level privacy-preserving federated learning: Analysis and performance optimization. IEEE Transactions on Mobile Computing, 21(9):3388-3401, 2022. +[53] Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-IID federated learning. In International Conference on Learning Representations, 2021. +[54] Ashkan Yousefpour, Igor Shilov, Alexandre Sablayrolles, Davide Testuggine, Karthik Prasad, Mani Malek, John Nguyen, Sayan Gosh, Akash Bharadwaj, Jessica Zhao, Graham Cormode, and Ilya Mironov. Opacus: User-friendly differential privacy library in pytorch. In Proc. Privacy in Machine Learning (PriML) Workshop, NeurIPS, Virtual, Dec. 2021. +[55] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10174-10183, 2022. +[56] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In International Conference on Machine Learning, ICML, pages 26982-26992. PMLR, 2022. + +[57] Qihuang Zhong, Liang Ding, Li Shen, Peng Mi, Juhua Liu, Bo Du, and Dacheng Tao. Improving sharpness-aware minimization with fisher mask for better generalization on language models. arXiv preprint arXiv:2210.05497, 2022. +[58] Yuqing Zhu, Xiang Yu, Yi-Hsuan Tsai, F. Pittaluga, M. Faraki, Manmohan Chandraker, and Yu-Xiang Wang. Voting-based approaches for differentially private federated learning. ArXiv, abs/2010.04851, 2020. \ No newline at end of file diff --git a/2023/Make Landscape Flatter in Differentially Private Federated Learning/images.zip b/2023/Make Landscape Flatter in Differentially Private Federated Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..98cc1b1f1bcf78fc8296905879c4f88d81f42062 --- /dev/null +++ b/2023/Make Landscape Flatter in Differentially Private Federated Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ec966ab3f5da3419870a046347963993a29b1249f7a5228e23aee12288862f1 +size 527229 diff --git a/2023/Make Landscape Flatter in Differentially Private Federated Learning/layout.json b/2023/Make Landscape Flatter in Differentially Private Federated Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..be659687588bd662c6325511c338a671aececb62 --- /dev/null +++ b/2023/Make Landscape Flatter in Differentially Private Federated Learning/layout.json @@ -0,0 +1,13071 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "type": "text", + "content": "Make Landscape Flatter in Differentially Private Federated Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "spans": [ + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": "Yifan Shi" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": " Yingqi Liu" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": " Kang Wei" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": " Li Shen" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": " Xueqian Wang" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": " Dacheng Tao" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": "Tsinghua University, Shenzhen, China; " + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": "JD Explore Academy, Beijing, China \n" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 141, + 511, + 186 + ], + "type": "text", + "content": "Nanjing University of Science and Technology, Nanjing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 187, + 487, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 187, + 487, + 200 + ], + "spans": [ + { + "bbox": [ + 100, + 187, + 487, + 200 + ], + "type": "text", + "content": "shiyf21@mails.tsinghua.edu.cn; lyq@njust.edu.cn; kang.wei@njust.edu.cn;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 201, + 485, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 201, + 485, + 213 + ], + "spans": [ + { + "bbox": [ + 100, + 201, + 485, + 213 + ], + "type": "text", + "content": "mathshenli@gmail.com; wang.xq@sz.tsinghua.edu.cn; dacheng.tao@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 266, + 290, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 266, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 45, + 266, + 290, + 531 + ], + "type": "text", + "content": "To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), client-level Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharper loss landscape and have poorer weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with better stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. From the theoretical perspective, we analyze in detail how DP-FedSAM mitigates the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with Rényi DP and present the sensitivity analysis of local updates. At last, we empirically confirm that our algorithm achieves state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 574, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 288, + 694 + ], + "type": "text", + "content": "Federated Learning (FL) [31] allows distributed clients to collaboratively train a shared model without sharing data. However, FL faces severe dilemma of privacy leakage [27]. Recent works show that a curious server can also infer clients' privacy information such as membership and data features, by well-designed generative models and/or shadow models [14, 37, 40, 44, 55]. To address this issue, differential privacy (DP) [12] has been introduced in FL, which can protect every instance in any client's data (instance-level DP [3,20,47,48]) or the information between" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "type": "text", + "content": "clients (client-level DP [8, 15, 21, 25, 35, 52]). In general, client-level DP are more suitable to apply in the real-world setting due to a better model performance. For instance, a language prediction model with client-level DP [16, 35] has been applied on mobile devices by Google. In general, the Gaussian noise perturbation-based method is commonly adopted for ensuring the strong client-level DP. However, this method includes two operations in terms of clipping the " + }, + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "type": "text", + "content": " norm of local updates to a sensitivity threshold " + }, + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "type": "text", + "content": " and adding random noise proportional to the model size, whose standard deviation (STD) is also decided by " + }, + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 241, + 547, + 410 + ], + "type": "text", + "content": ". These steps may cause severe performance degradation dilemma [9, 22], especially on large-scale complex model [45], such as ResNet-18 [18], or with heterogeneous data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 411, + 548, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 548, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 548, + 581 + ], + "type": "text", + "content": "The reasons behind this issue may be two-fold: (i) The useful information is dropped due to the clipping operation especially on small " + }, + { + "bbox": [ + 304, + 411, + 548, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 411, + 548, + 581 + ], + "type": "text", + "content": ", which is contained in the local updates; (ii) The model inconsistency among local models is exacerbated as the addition of random noise severely damages local updates and leads to large variances between local models especially on large " + }, + { + "bbox": [ + 304, + 411, + 548, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 411, + 548, + 581 + ], + "type": "text", + "content": " [9]. To overcome these issues, existing works improve the performance via restricting the norm of local update [9] and leveraging local update sparsification technique [9, 22] to reduce the adverse impacts of clipping and adding amount of random noise. However, the model performance degradation remains significantly severe compared with FL methods without considering the privacy, e.g., FedAvg [33]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "Motivation. To further explore the reasons behind this phenomenon, we compare the structure of loss landscapes and surface contours [30] for FedAvg [33] and DP-FedAvg [15,35] on partitioned CIFAR-10 dataset [28] with Dirichlet distribution " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "(\\alpha = 0.6)" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " and ResNet-18 backbone [18] in Figure 1 (a) and (b), respectively. Note that the convergence of DP-FedAvg is worse than FedAvg as its loss value is higher after a long communication round. Furthermore, FedAvg has a flatter landscape, whereas DP-FedAvg has a sharper one, resulting in poorer generalization ability (sharper minima, see Figure 1 (a)) and weight perturbation robustness" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 231, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 231, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 231, + 714 + ], + "type": "text", + "content": "*Corresponding authors: Li Shen and Xueqian Wang" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24552" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 324, + 144 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 324, + 144 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 324, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 324, + 144 + ], + "type": "image", + "image_path": "96397fdd979bf62e21e9567d7174fcfbd362e96d6fd8fb400071671d3f0e583f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 146, + 227, + 159 + ], + "lines": [ + { + "bbox": [ + 144, + 146, + 227, + 159 + ], + "spans": [ + { + "bbox": [ + 144, + 146, + 227, + 159 + ], + "type": "text", + "content": "(a) Loss landscapes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 435, + 144 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 435, + 144 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 435, + 144 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 435, + 144 + ], + "type": "image", + "image_path": "6e2fb01e56f1e346f7dc14e8fda65fb56287e550f6faf37ab01b56057bd062d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 163, + 514, + 175 + ], + "lines": [ + { + "bbox": [ + 78, + 163, + 514, + 175 + ], + "spans": [ + { + "bbox": [ + 78, + 163, + 514, + 175 + ], + "type": "text", + "content": "Figure 1. Loss landscapes and surface contours comparison between DP-FedAvg (left) and FedAvg (right), respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 440, + 70, + 543, + 144 + ], + "blocks": [ + { + "bbox": [ + 385, + 147, + 490, + 158 + ], + "lines": [ + { + "bbox": [ + 385, + 147, + 490, + 158 + ], + "spans": [ + { + "bbox": [ + 385, + 147, + 490, + 158 + ], + "type": "text", + "content": "(b) Loss surface contours." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 70, + 543, + 144 + ], + "lines": [ + { + "bbox": [ + 440, + 70, + 543, + 144 + ], + "spans": [ + { + "bbox": [ + 440, + 70, + 543, + 144 + ], + "type": "image", + "image_path": "8c7c4f04a74b93d8df36cabfb7bf9c8cebbb2511fe8677dd0a64553467e760e8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 183, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 183, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 45, + 183, + 287, + 243 + ], + "type": "text", + "content": "(see Figure 1 (b)), which is caused by the clipped local update information and exacerbated model inconsistency, respectively. Based on these observations, an interesting research question is: can we further overcome the severe performance degradation via making landscape flatter?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "text", + "content": "To answer this question, we propose DP-FedSAM via gradient perturbation to improve model performance. Specifically, a local flat model is generated by using SAM optimizer [13] in each client, which leads to better stability. After that, a potentially global flat model can be generated by aggregating several flat local models, which results in higher generalization ability and better robustness to DP noise, thereby significantly improving the performance and achieving a better trade-off between performance and privacy. Theoretically, we present a tighter bound " + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} + \\frac{\\sum_{t=1}^{T} (\\bar{\\alpha}^t \\sigma_g^2 + \\bar{\\alpha}^t L^2)}{T^2} + \\frac{L^2 \\sqrt{T} \\sigma^2 C^2 d}{m^2 \\sqrt{K}}\\right)" + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "text", + "content": " for DP-FedSAM in the stochastic non-convex setting, where both " + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "inline_equation", + "content": "\\frac{1}{T} \\sum_{t=1}^{T} \\bar{\\alpha}^t" + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "inline_equation", + "content": "\\frac{1}{T} \\sum_{t=1}^{T} \\tilde{\\alpha}^t" + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "text", + "content": " are bounded constant, " + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "text", + "content": " is local iteration steps and communication rounds, respectively. Meanwhile, we present how SAM mitigates the impact of DP. For the clipping operation, DP-FedSAM reduces the " + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 46, + 249, + 289, + 589 + ], + "type": "text", + "content": " norm and the negative impact of the inconsistency among local updates on convergence. For adding noise operation, we obtain better weight perturbation robustness for reducing the performance damage caused by random noise, thereby having better robustness to DP noise. Empirically, we conduct extensive experiments on EMNIST, CIFAR-10, and CIFAR-100 datasets in both the independently and identically distributed (IID) and Non-IID settings. Furthermore, we observe the loss landscape, surface contour, and the norm distribution of local updates for exploring the intrinsic effects of SAM with DP, which together with the theoretical analysis confirms the effect of DP-FedSAM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Contribution. The main contributions of our work are summarized as four-fold: (i) We propose a novel scheme DPFedSAM in DPFL to alleviate performance degradation issue from the optimizer perspective. (ii) We establish an improved convergence rate, which is tighter than the conventional bounds [9, 22] in the stochastic non-convex setting. Moreover, we provide rigorous privacy guarantees and sensitivity analysis. (iii) We are the first to in-depth analyze the roles of the on-average norm of local updates " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}^t" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " and local update consistency among clients " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}^t" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " on convergence. (iv)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 184, + 547, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 547, + 220 + ], + "type": "text", + "content": "We conduct extensive experiments to verify the effect of DP-FedSAM, which could achieve state-of-the-art (SOTA) performance compared with several strong DPFL baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 227, + 390, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 227, + 390, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 390, + 239 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 247, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 498 + ], + "type": "text", + "content": "Client-level DPFL. Client-level DPFL is the de-facto approach for protecting any client's data. DP-FedAvg [36] is the first to study in this setting, which trains a language prediction model in a mobile keyboard and ensures client-level DP guarantee by employing the Gaussian mechanism and composing privacy guarantees. After that, the work in [26,51] presents a comprehensive end-to-end system, which appropriately discretizes the data and adds discrete Gaussian noise before performing secure aggregation. Meanwhile, AE-DPFL [58] leverages the voting-based mechanism among the data labels instead of averaging the gradients to avoid dimension dependence and significantly reduce the communication cost. Fed-SMP [22] uses Sparsified Model Perturbation (SMP) to mitigate the impact of privacy protection on model accuracy. Different from the aforementioned methods, a recent study [9] revisits this issue and leverages Bounded Local Update Regularization (BLUR) and Local Update Sparsification (LUS) to restrict the norm of local updates and reduce the noise size before executing operations that guarantee DP, respectively. Nevertheless, severe performance degradation remains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 665 + ], + "type": "text", + "content": "Sharpness Aware Minimization (SAM). SAM [13] is an effective optimizer for training deep learning (DL) models, which leverages the flatness geometry of the loss landscape to improve model generalization ability. Recently, [4] study the properties of SAM and provide convergence results of SAM for non-convex objectives. As a powerful optimizer, SAM and its variants have been applied to various machine learning (ML) tasks [2, 11, 24, 29, 32, 38, 43, 46, 50, 56, 57]. Specifically, [41], [50] and [7] integrate SAM to improve the generalization, and thus mitigate the distribution shift problem and achieve a new SOTA performance for FL. However, to the best of our knowledge, no efforts have been devoted to the empirical performance and theoretical analysis of SAM in DPFL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "The most related works to this paper are DP-FedAvg in [35], Fed-SMP in [22], and DP-FedAvg with LUS and BLUR [9]. However, these works still suffer from inferior performance due to the exacerbated model inconsistency is" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24553" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "sue among the clients caused by random noise. Therefore, different from existing works, we try to alleviate this issue by making the landscape flatter and weight perturbation ability more robust. Furthermore, another related work is FedSAM [41], which integrates SAM optimizer to enhance the flatness of the local model and achieves new SOTA performance for FL. On top of the aforementioned studies, we are the first to extend the SAM optimizer into the DPFL setting to effectively alleviate the performance degradation problem. Meanwhile, we provide the theoretical analysis for sensitivity, privacy, and convergence in the non-convex setting. Finally, we empirically confirm our theoretical results and the superiority of performance compared with existing SOTA methods in DPFL." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 250, + 124, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 124, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 124, + 262 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 270, + 287, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 294 + ], + "type": "text", + "content": "In this section, we first give the problem setup of FL, and then introduce the several terminologies in DP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 163, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 163, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 163, + 313 + ], + "type": "text", + "content": "3.1. Federated Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": "Consider a general FL system consisting of " + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": " clients, in which each client owns its local dataset. Let " + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^{\\mathrm{train}}" + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^{\\mathrm{val}}" + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^{\\mathrm{test}}" + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": " denote the training dataset, validation dataset and testing dataset, held by client " + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{U} = \\{1, 2, \\dots, M\\}" + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": ". Formally, the FL task is expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 384, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 384, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 100, + 384, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\mathbf {w} ^ {\\star} = \\underset {\\mathbf {w}} {\\arg \\min } \\sum_ {i \\in \\mathcal {U}} p _ {i} f _ {i} (\\mathbf {w}, \\mathcal {D} _ {i} ^ {\\text {t r a i n}}), \\tag {1}", + "image_path": "42a9fda189bee098be3aa0f079593aaf28b77a581f643a47eeeabdb89a49fe54.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "p_i = |\\mathcal{D}_i^{\\mathrm{train}}| / |\\mathcal{D}^{\\mathrm{train}}| \\geq 0" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\sum_{i\\in \\mathcal{U}}p_i = 1" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "f_{i}(\\cdot)" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": " is the local loss function with " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "f_{i}(\\mathbf{w}) = F_{i}(\\mathbf{w};\\xi_{i}), \\xi_{i}" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": " is a batch sample data in client " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_i^{\\mathrm{train}}|" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": " is the size of training dataset " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^{\\mathrm{train}}" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}^{\\mathrm{train}}| = \\sum_{i\\in \\mathcal{U}}|\\mathcal{D}_i^{\\mathrm{train}}|" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": " is the total size of training datasets, respectively. For the " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": "-th client, a local model is learned on its private training data " + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{train}}" + }, + { + "bbox": [ + 47, + 414, + 287, + 486 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 492, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 492, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 110, + 492, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {i} = \\mathbf {w} _ {i} ^ {t} - \\eta \\nabla f _ {i} \\left(\\mathbf {w} _ {i}, \\mathcal {D} _ {i} ^ {\\text {t r a i n}}\\right). \\tag {2}", + "image_path": "1c9c2931350695f7b45ed692b414714ed65515af3d4f438496c9dd977a257ac1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "content": "Generally, the local loss function " + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "inline_equation", + "content": "f_{i}(\\cdot)" + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "content": " has the same expression across each client. Then, the " + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "content": " associated clients collaboratively learn a global model " + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "content": " over the heterogeneous training data " + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^{\\mathrm{train}}" + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\forall i \\in \\mathcal{U}" + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 566, + 162, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 162, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 162, + 578 + ], + "type": "text", + "content": "3.2. Differential Privacy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 585, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 287, + 620 + ], + "type": "text", + "content": "Differential Privacy (DP) [12] is a rigorous privacy notion for measuring privacy risk. In this paper, we consider a relaxed version: Rényi DP (RDP) [39]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": "Definition 1. (Rényi DP, [39]). Given a real number " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\alpha \\in (1,\\infty)" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": " and privacy parameter " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\rho \\geq 0" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": ", a randomized mechanism " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": " satisfies " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "(\\alpha ,\\rho)" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": "-RDP if for any two neighboring datasets " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": " that differ in a single record, the Rényi " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": "-divergence between " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathcal{D})" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathcal{D}')" + }, + { + "bbox": [ + 47, + 627, + 287, + 687 + ], + "type": "text", + "content": " satisfies:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 692, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 692, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 50, + 692, + 287, + 717 + ], + "type": "interline_equation", + "content": "D _ {\\alpha} \\left[ \\mathcal {M} (\\mathcal {D}) \\| \\mathcal {M} \\left(\\mathcal {D} ^ {\\prime}\\right) \\right] := \\frac {1}{\\alpha - 1} \\log \\mathbb {E} \\left[ \\left(\\frac {\\mathcal {M} (\\mathcal {D})}{\\mathcal {M} \\left(\\mathcal {D} ^ {\\prime}\\right)}\\right) ^ {\\alpha} \\right] \\leq \\rho , \\tag {3}", + "image_path": "d4f665396baa51da741a7ce5e8e631f8ae227caf130dc0618eb369348abdc2a6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 538, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 538, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 538, + 84 + ], + "type": "text", + "content": "where the expectation is taken over the output of " + }, + { + "bbox": [ + 306, + 72, + 538, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathcal{D}^{\\prime})" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 93, + 545, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 93, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 305, + 93, + 545, + 154 + ], + "type": "text", + "content": "Rényi DP is a useful analytical tool to measure the privacy and accurately represent guarantees on the tails of the privacy loss, which is strictly stronger than " + }, + { + "bbox": [ + 305, + 93, + 545, + 154 + ], + "type": "inline_equation", + "content": "(\\epsilon, \\delta)" + }, + { + "bbox": [ + 305, + 93, + 545, + 154 + ], + "type": "text", + "content": "-DP for " + }, + { + "bbox": [ + 305, + 93, + 545, + 154 + ], + "type": "inline_equation", + "content": "\\delta > 0" + }, + { + "bbox": [ + 305, + 93, + 545, + 154 + ], + "type": "text", + "content": ". Thus, we provide the privacy analysis based on this tool for each user's privacy loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "text", + "content": "Definition 2. (" + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "text", + "content": " Sensitivity). [9, Definition 2] Let " + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "text", + "content": " be a function, the " + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "text", + "content": "-sensitivity of " + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\max_{D \\simeq D'} \\| \\mathcal{F}(D) - \\mathcal{F}(D') \\|_2" + }, + { + "bbox": [ + 305, + 162, + 545, + 211 + ], + "type": "text", + "content": ", where the maximization is taken over all pairs of adjacent datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 219, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 545, + 267 + ], + "type": "text", + "content": "The sensitivity of a function " + }, + { + "bbox": [ + 305, + 219, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 219, + 545, + 267 + ], + "type": "text", + "content": " captures the magnitude which a single individual's data can change the function " + }, + { + "bbox": [ + 305, + 219, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 219, + 545, + 267 + ], + "type": "text", + "content": " in the worst case. Therefore, it plays a crucial role in determining the magnitude of noise required to ensure DP." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "text", + "content": "Definition 3. (Client-level DP) [34, Definition 1]. A randomized algorithm " + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "inline_equation", + "content": "(\\epsilon, \\delta)" + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "text", + "content": "-DP if for any two adjacent datasets " + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "inline_equation", + "content": "U'" + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "text", + "content": " constructed by adding or removing all records of any client, and every possible subset of outputs " + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 305, + 276, + 545, + 337 + ], + "type": "text", + "content": " satisfy the following inequality:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 335, + 346, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 346, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 335, + 346, + 545, + 360 + ], + "type": "interline_equation", + "content": "\\Pr [ \\mathcal {M} (U) \\in O ] \\leq e ^ {\\epsilon} \\Pr [ \\mathcal {M} \\left(U ^ {\\prime}\\right) \\in O ] + \\delta . \\tag {4}", + "image_path": "7b21c8ae7afccc8310c22f1bcb9e328accd20968c2defa634584e5d51b715c16.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 369, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 406 + ], + "type": "text", + "content": "In client-level DP, we aim to ensure participation information for any clients. Therefore, we need to make local updates similar whether one client participates or not." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 418, + 389, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 389, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 389, + 431 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 438, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 582 + ], + "type": "text", + "content": "To revisit the severe performance degradation challenge in DPFL, we observe the loss landscapes and surface contours of FedAvg and DP-FedAvg in Figure 1 (a) and (b), respectively. And we find that the DPFL method has a sharper landscape with both poorer generalization ability and weight perturbation robustness than the FL method. It means that the DPFL method may result in poor flatness and make model sensitivity to noise perturbation. However, the study focusing on this issue remains unexplored. Therefore, we plan to face this challenge from the optimizer perspective by adopting a SAM optimizer in each client, dubbed DP-FedSAM, whose local loss function is defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 589, + 545, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 589, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 545, + 608 + ], + "type": "interline_equation", + "content": "f _ {i} (\\mathbf {w}) = \\mathbb {E} _ {\\xi \\sim \\mathcal {D} _ {i}} \\max _ {\\| \\delta_ {i} \\| _ {2} \\leq \\rho} F _ {i} \\left(\\mathbf {w} ^ {t, k} (i) + \\delta_ {i}; \\xi_ {i}\\right), \\quad i \\in \\mathcal {N}, \\tag {5}", + "image_path": "46cfd4202b7185a5108d848dbceaf6a2c469a9a736ae41c08086a1986fe70114.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{t,k}(i) + \\delta_{i}" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": " is viewed as the perturbed model, " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": " is a predefined constant controlling the radius of the perturbation and " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": "-norm. Instead of searching for a solution via SGD [5, 6], SAM [13] aims to seek a solution in a flat region by adding a small perturbation, i.e., " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "w + \\delta" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": " with more robust performance. Specifically, for each client " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,\\dots,M\\}" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": ", each local iteration " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "k\\in \\{0,1,\\dots,K - 1\\}" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": " in each communication round " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "t\\in \\{0,1,\\dots,T - 1\\}" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": "-th" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24554" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 215, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 215, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 215, + 84 + ], + "type": "text", + "content": "inner iteration in client " + }, + { + "bbox": [ + 47, + 72, + 215, + 84 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 215, + 84 + ], + "type": "text", + "content": " is performed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 93, + 287, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 93, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 99, + 93, + 287, + 108 + ], + "type": "interline_equation", + "content": "\\mathbf {w} ^ {t, k + 1} (i) = \\mathbf {w} ^ {t, k} (i) - \\eta \\tilde {\\mathbf {g}} ^ {t, k} (i), \\tag {6}", + "image_path": "a4bbefd9bd965b3ae538544c3750dc02bdb5054308e13470a28eab642b381df1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 110, + 287, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 287, + 138 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {g}} ^ {t, k} (i) = \\nabla F _ {i} \\left(\\mathbf {w} ^ {t, k} + \\delta \\left(\\mathbf {w} ^ {t, k}\\right); \\xi\\right), \\delta \\left(\\mathbf {w} ^ {t, k}\\right) = \\frac {\\rho \\mathbf {g} ^ {t , k}}{\\left\\| \\mathbf {g} ^ {t , k} \\right\\| _ {2}}. \\tag {7}", + "image_path": "9c4c8937c14f3ce0597e62b1a1636bd38dbffa05b26e4f9e6435535949bae205.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\delta(\\mathbf{w}^{t,k})" + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "text", + "content": " is calculated by using first-order Taylor expansion around " + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{t,k}" + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "text", + "content": " [13]. After that, we adopt a sampling mechanism, gradient clipping, and Gaussian noise adding to ensure client-level DP. Note that this sampling can amplify the privacy guarantee since it decreases the chances of leaking information about a particular individual [1]. Specifically, after sampling " + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "text", + "content": " clients with probability " + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "inline_equation", + "content": "q = m / M" + }, + { + "bbox": [ + 47, + 148, + 287, + 257 + ], + "type": "text", + "content": " at each communication round, which is important for measuring privacy loss. Then, we clip the local updates first:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 266, + 287, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 266, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 110, + 266, + 287, + 289 + ], + "type": "interline_equation", + "content": "\\tilde {\\Delta} _ {i} ^ {t} = \\Delta_ {i} ^ {t} \\cdot \\min \\left(1, \\frac {C}{\\| \\Delta_ {i} ^ {t} \\| _ {2}}\\right). \\tag {8}", + "image_path": "a07035e3c687b7ecfea87d80df84dfda9497c984b3a82930849d7f2480d9c582.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 298, + 287, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 287, + 322 + ], + "type": "text", + "content": "After clipping, we add Gaussian noise to the local update for ensuring client-level DP as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 331, + 287, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 331, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 331, + 287, + 344 + ], + "type": "interline_equation", + "content": "\\hat {\\Delta} _ {i} ^ {t} = \\tilde {\\Delta} _ {i} ^ {t} + \\mathcal {N} \\left(0, \\sigma^ {2} C ^ {2} \\cdot \\mathbf {I} _ {d} / m\\right). \\tag {9}", + "image_path": "0ac15e9aaf027553c16abfa95e3811e7c7014acb928f93b4564ecbb4e612938d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 354, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 402 + ], + "type": "text", + "content": "With a given noise variance " + }, + { + "bbox": [ + 47, + 354, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 47, + 354, + 287, + 402 + ], + "type": "text", + "content": ", the accumulative privacy budget " + }, + { + "bbox": [ + 47, + 354, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 47, + 354, + 287, + 402 + ], + "type": "text", + "content": " can be calculated based on the sampled Gaussian mechanism [54]. Finally, we summarize the training procedures of DP-FedSAM in Algorithm 1." + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 50, + 431, + 286, + 705 + ], + "blocks": [ + { + "bbox": [ + 53, + 418, + 167, + 429 + ], + "lines": [ + { + "bbox": [ + 53, + 418, + 167, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 167, + 429 + ], + "type": "text", + "content": "Algorithm 1: DP-FedSAM" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "lines": [ + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "spans": [ + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": "Input: Total number of clients " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " sampling ratio of clients " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " total number of communication rounds " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " the clipping threshold " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " local learning rate " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " and total number of the local iterates are " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". Output: Generate global model " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^T" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". \n1 Initialization: Randomly initialize the global model " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^0" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". \n2 for " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " do \n3 Sample a set of " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "m = qM" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " clients at random without replacement, denoted by " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^t" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". \n4 for client " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " in parallel do \n5 for " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " do \n6 Update the global parameter as the local parameter " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^t(i) \\gets \\mathbf{w}^t" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". \n7 Sample a batch of local data " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\xi_i" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": " and calculate local gradient " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{t,k}(i) = \\nabla F_i(\\mathbf{w}^{t,k}(i); \\xi_i)" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". \n8 Gradient perturbation by Equation (7). Local iteration update by Equation (6). \n9 end \n11 " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\Delta_i^t = \\mathbf{w}^{t,K}(i) - \\mathbf{w}^{t,0}(i)" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". \n12 Clip and add noise for DP by Equation (9). \n13 Return " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}^t(i)" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". \n14 end \n15 " + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{t + 1} \\gets \\mathbf{w}^t + \\frac{1}{m} \\sum_{i \\in \\mathcal{W}^t} \\hat{\\Delta}^t(i)" + }, + { + "bbox": [ + 50, + 431, + 286, + 705 + ], + "type": "text", + "content": ". \n16 end" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": "Compared with existing DPFL methods [9, 22, 35], the benefits of DP-FedSAM lie in three-fold: (i) We introduce SAM into DPFL to alleviate severe performance degradation via seeking a flatness model in each client, which is caused by the exacerbated inconsistency of local models. Specifically, DP-FedSAM generates both better generalization ability and robustness to DP noise by making the global model flatter. (ii) In addition, we analyze in detail how DP-FedSAM mitigates the negative impacts of DP. Specifically, we theoretically analyze the convergence with the on-average norm of local updates " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " and local update consistency among clients " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": ", and empirically confirm these results via observing the norm distribution and average norm of local updates (see Section 6.3). (iii) Meanwhile, we present the theories unifying the impacts of gradient perturbation " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " in SAM, the on-average norm of local updates " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": ", and local update consistency among clients " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " in clipping operation, and the variance of random noise " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\sigma^2 C^2 / m" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " upon the convergence rate in Section 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 308, + 425, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 425, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 425, + 322 + ], + "type": "text", + "content": "5. Theoretical Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 328, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 376 + ], + "type": "text", + "content": "In this section, we give a rigorous analysis of DP-FedSAM, including its sensitivity, privacy, and convergence rate. The detailed proof is placed in Appendix E. Below, we first give several necessary assumptions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": "Assumption 1. (Lipschitz smoothness). The function " + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": " is differentiable and " + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\nabla F_{i}" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": "-Lipschitz continuous, " + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\forall i \\in \\{1,2,\\ldots,M\\}" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\| \\nabla F_{i}(\\mathbf{x}) - \\nabla F_{i}(\\mathbf{y}) \\| \\leq L \\| \\mathbf{x} - \\mathbf{y} \\|" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": ", for all " + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{x}, \\mathbf{y} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": "Assumption 2. (Bounded variance). The gradient of the function " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": " have " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\sigma_l" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": "-bounded variance, i.e., " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\xi_i}\\left\\| \\nabla F_i(\\mathbf{w}^k (i);\\xi_i) - \\nabla F_i(\\mathbf{w}(i))\\right\\|^2 \\leq \\sigma_l^2" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\forall i \\in \\{1,2,\\dots,M\\}" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "k \\in \\{1,\\dots,K - 1\\}" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": ", the global variance is also bounded, i.e., " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\frac{1}{M}\\sum_{i = 1}^{M}\\| \\nabla f_{i}(\\mathbf{w}) - \\nabla f(\\mathbf{w})\\|^{2} \\leq \\sigma_{g}^{2}" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": ". It is not hard to verify that the " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\sigma_g" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": " is smaller than the homogeneity parameter " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "inline_equation", + "content": "\\sigma_g^2 \\leq \\beta^2" + }, + { + "bbox": [ + 305, + 433, + 545, + 523 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "type": "text", + "content": "Assumption 3. (Bounded gradient). For any " + }, + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\ldots,M\\}" + }, + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "type": "inline_equation", + "content": "\\|\\nabla f_i(\\mathbf{w})\\| \\leq B" + }, + { + "bbox": [ + 306, + 525, + 545, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "text", + "content": "Assumption 4. (Unbiased Gradient Estimator). For any data sample " + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "text", + "content": ", the local gradient estimator is unbiased, i.e., " + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\nabla f_i(\\mathbf{w}; z)] = \\mathbb{E}[\\nabla f_i(\\mathbf{w})]" + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": "Note that the above assumptions are mild and commonly used in characterizing the convergence rate of FL [6,9,17,22,23,41-43,49,53]. Furthermore, gradient clipping operation in DL is often used to prevent the gradient explosion phenomenon, thereby the gradient is bounded. The technical difficulty for DP-FedSAM lies in: (i) how SAM mitigates the impact of DP; (ii) how to analyze in detail the impacts of the consistency among clients and the on-average norm of local updates caused by clipping operation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24555" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 161, + 85 + ], + "type": "text", + "content": "5.1. Sensitivity Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "text", + "content": "At first, we study the sensitivity of local update " + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "inline_equation", + "content": "\\Delta_i^t" + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "text", + "content": " from any client " + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,\\dots,M\\}" + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "text", + "content": " before clipping at " + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "text", + "content": "-th communication round in DP-FedSAM. This upper bound of sensitivity can roughly measure the degree of privacy protection. Under Definition 2, the sensitivity can be denoted by " + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "inline_equation", + "content": "S_{\\Delta_i^t}" + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "text", + "content": " in client " + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 89, + 287, + 161 + ], + "type": "text", + "content": "-th communication round." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "content": "Theorem 1. (Sensitivity). Denote " + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\Delta_i^t (\\mathbf{x})" + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\Delta_i^t (\\mathbf{y})" + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "content": " as the local update at " + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "content": "-th communication round, the model " + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(i)" + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{y}(i)" + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "content": " is conducted on two sets which differ at only one sample. Assume the initial model parameter " + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^t (i) = \\mathbf{x}^{t,0}(i) = \\mathbf{y}^{t,0}(i)" + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "content": ". With the above assumptions, the expected squared sensitivity " + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "inline_equation", + "content": "S_{\\Delta_i^t}^2" + }, + { + "bbox": [ + 47, + 168, + 288, + 243 + ], + "type": "text", + "content": " of local update is upper bounded," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 256, + 287, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 256, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 85, + 256, + 287, + 283 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\mathcal {S} _ {\\Delta_ {i} ^ {t}} ^ {2} \\leq \\frac {6 \\eta^ {2} \\rho^ {2} K L ^ {2} \\left(1 2 K ^ {2} L ^ {2} \\eta^ {2} + 1 0\\right)}{1 - 2 \\eta^ {2} L ^ {2} K} \\tag {10}", + "image_path": "4e014fa740cd046f3d04c67d63602361949ef379bf7742d92eabd4a3ff1ad4a5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "type": "text", + "content": "When the local adaptive learning rate satisfies " + }, + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "type": "inline_equation", + "content": "\\eta = \\mathcal{O}(1 / L\\sqrt{KT})" + }, + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "type": "text", + "content": " and the perturbation amplitude " + }, + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "type": "text", + "content": " proportional to the learning rate, e.g., " + }, + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "type": "inline_equation", + "content": "\\rho = \\mathcal{O}\\left(\\frac{1}{\\sqrt{T}}\\right)" + }, + { + "bbox": [ + 47, + 289, + 287, + 328 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 335, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 335, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 125, + 335, + 287, + 361 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\mathcal {S} _ {\\Delta_ {i} ^ {t}} ^ {2} \\leq \\mathcal {O} \\left(\\frac {1}{T ^ {2}}\\right). \\tag {11}", + "image_path": "44e7c0f9b18c487de06f98f020ccd6f58e65de793baad179cbbedf7319ef8923.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "type": "text", + "content": "For comparison, we also present the expected squared sensitivity of local update with SGD in DPFL, that is " + }, + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\mathcal{S}_{\\Delta_i^t,SGD}^2\\leq \\frac{6\\eta^2\\sigma_l^2K}{1 - 3\\eta^2KL^2}" + }, + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "type": "text", + "content": ". Thus " + }, + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\mathcal{S}_{\\Delta_i^t,SGD}^2\\leq \\mathcal{O}(\\frac{\\sigma_l^2}{KL^2T})" + }, + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "type": "inline_equation", + "content": "\\eta = \\mathcal{O}(1 / L\\sqrt{KT})" + }, + { + "bbox": [ + 47, + 368, + 287, + 423 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 429, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 287, + 502 + ], + "type": "text", + "content": "Remark 1. It is clearly seen that the upper bound in " + }, + { + "bbox": [ + 47, + 429, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbb{ES}_{\\Delta_i^t,SAM}^2" + }, + { + "bbox": [ + 47, + 429, + 287, + 502 + ], + "type": "text", + "content": " is tighter than that in " + }, + { + "bbox": [ + 47, + 429, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbb{ES}_{\\Delta_i^t,SGD}^2" + }, + { + "bbox": [ + 47, + 429, + 287, + 502 + ], + "type": "text", + "content": ". From the perspective of privacy protection, it means DP-FedSAM has a better privacy guarantee than DP-FedAvg. Another perspective from local iteration, which means both better model consistency among clients and training stability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 508, + 148, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 148, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 148, + 521 + ], + "type": "text", + "content": "5.2. Privacy Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 526, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 287, + 563 + ], + "type": "text", + "content": "To achieve client-level privacy protection, we derive the sensitivity of the aggregation process at first after clipping the local updates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 569, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 287, + 594 + ], + "type": "text", + "content": "Lemma 1. The sensitivity of client-level DP in DP-FedSAM can be expressed as " + }, + { + "bbox": [ + 47, + 569, + 287, + 594 + ], + "type": "inline_equation", + "content": "C / m" + }, + { + "bbox": [ + 47, + 569, + 287, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "type": "text", + "content": "Proof. Given two adjacent batches " + }, + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^t" + }, + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^{t,\\mathrm{adj}}" + }, + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "type": "text", + "content": ", that " + }, + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^{t,\\mathrm{adj}}" + }, + { + "bbox": [ + 47, + 598, + 287, + 623 + ], + "type": "text", + "content": " has one more or less client, we have" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 628, + 287, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 287, + 665 + ], + "type": "interline_equation", + "content": "\\left\\| \\frac {1}{m} \\sum_ {i \\in \\mathcal {W} ^ {t}} \\Delta_ {i} ^ {t} - \\frac {1}{m} \\sum_ {j \\in \\mathcal {W} ^ {t, \\mathrm {a d j}}} \\Delta_ {j} ^ {t} \\right\\| _ {2} = \\frac {1}{m} \\left\\| \\Delta_ {j ^ {\\prime}} ^ {t} \\right\\| _ {2} \\leq \\frac {C}{m}, \\tag {12}", + "image_path": "bda57b3725df319320f31130230d69c08ed5d81cf60476e2429c3392c52590b1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 670, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 670, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\Delta_{j'}^t" + }, + { + "bbox": [ + 47, + 670, + 287, + 685 + ], + "type": "text", + "content": " is the local update of one more or less client." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Remark 2. The value of this sensitivity can determine the amount of variance for adding random noise." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 97 + ], + "type": "text", + "content": "After adding Gaussian noise, we calculate the accumulative privacy budget [54] along with training as follows." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 102, + 545, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 102, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 102, + 545, + 126 + ], + "type": "text", + "content": "Theorem 2. After " + }, + { + "bbox": [ + 306, + 102, + 545, + 126 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 102, + 545, + 126 + ], + "type": "text", + "content": " communication rounds, the accumulative privacy budget is calculated by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 130, + 545, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 130, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 326, + 130, + 545, + 155 + ], + "type": "interline_equation", + "content": "\\epsilon = \\bar {\\epsilon} + \\frac {(\\alpha - 1) \\log (1 - \\frac {1}{\\alpha}) - \\log (\\alpha) - \\log (\\delta)}{\\alpha - 1}, \\tag {13}", + "image_path": "ce36290ded7a7db0047cb928d2f441a6fc71ca65cce1e82b1aca67255cbddfde.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 159, + 333, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 159, + 333, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 333, + 169 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 174, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 174, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 174, + 545, + 198 + ], + "type": "interline_equation", + "content": "\\bar {\\epsilon} = \\frac {T}{\\alpha - 1} \\ln \\mathbb {E} _ {z \\sim \\mu_ {0} (z)} \\left[ \\left(1 - q + \\frac {q \\mu_ {1} (z)}{\\mu_ {0} (z)}\\right) ^ {\\alpha} \\right], \\tag {14}", + "image_path": "a1faa28cd32a38613726aa783444d0302c4e7e09bb05c9b6f6556c7aa127478e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "content": " is sample rate for client selection, " + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\mu_0(z)" + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\mu_{1}(z)" + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "content": " denote the Gaussian probability density function (PDF) of " + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\sigma)" + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "content": " and the mixture of two Gaussian distributions " + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "inline_equation", + "content": "q\\mathcal{N}(1,\\sigma) + (1 - q)\\mathcal{N}(0,\\sigma)" + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "content": ", respectively, " + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "content": " is the noise STD, " + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 203, + 545, + 262 + ], + "type": "text", + "content": " is a selectable variable." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 269, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 269, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 545, + 328 + ], + "type": "text", + "content": "Remark 3. It can be seen that a small sampling rate " + }, + { + "bbox": [ + 305, + 269, + 545, + 328 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 269, + 545, + 328 + ], + "type": "text", + "content": " can enhance the privacy guarantee by decreasing the privacy budget, but it may also degrade the training performance due to the number of participating clients being reduced in each communication round. So a better trade-off is needed." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 335, + 432, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 432, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 432, + 346 + ], + "type": "text", + "content": "5.3. Convergence Analysis" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 352, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 352, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 545, + 424 + ], + "type": "text", + "content": "Below, we give a convergence analysis of how DP-FedSAM mitigates the negative impacts of DP. The technical contribution also is combining the impacts of the on-average norm of local updates " + }, + { + "bbox": [ + 305, + 352, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}^t" + }, + { + "bbox": [ + 305, + 352, + 545, + 424 + ], + "type": "text", + "content": " and local update consistency among clients " + }, + { + "bbox": [ + 305, + 352, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}^t" + }, + { + "bbox": [ + 305, + 352, + 545, + 424 + ], + "type": "text", + "content": " on the rate. Moreover, we also empirically confirm these results in Section 6.3." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": "Theorem 3. Under assumptions 1-4, local learning rate satisfies " + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\eta = \\mathcal{O}(1 / L\\sqrt{KT})" + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": " is denoted as the minimal value of " + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "inline_equation", + "content": "f(x)\\geq f(x^{*}) = f^{*}" + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": ". When the perturbation amplitude " + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": " is proportional to the learning rate, e.g., " + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\rho = \\mathcal{O}(1 / \\sqrt{T})" + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": ", the sequence of outputs " + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}^t\\}" + }, + { + "bbox": [ + 305, + 430, + 545, + 502 + ], + "type": "text", + "content": " generated by Alg. 1, we have:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 508, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 545, + 605 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\mathbb {E} \\left[ \\overline {{\\alpha}} ^ {t} \\left\\| \\nabla f \\left(\\mathbf {w} ^ {t}\\right) \\right\\| ^ {2} \\right] \\leq \\underbrace {\\mathcal {O} \\left(\\frac {2 L \\left(f \\left(\\mathbf {w} ^ {1}\\right) - f ^ {*}\\right)}{\\sqrt {K T}} + \\frac {L ^ {2} \\sigma_ {l} ^ {2}}{K T ^ {2}}\\right)} _ {\\text {F r o m F e d S A M}} \\\\ + \\underbrace {\\mathcal {O} \\left(\\frac {\\sum_ {t = 1} ^ {T} \\left(\\bar {\\alpha} ^ {t} \\sigma_ {g} ^ {2} + \\tilde {\\alpha} ^ {t} L ^ {2}\\right)}{T ^ {2}}\\right)} _ {\\text {C l i p p i n g}} + \\underbrace {\\mathcal {O} \\left(\\frac {L ^ {2} \\sqrt {T} \\sigma^ {2} C ^ {2} d}{m ^ {2} \\sqrt {K}}\\right)} _ {\\text {A d d i n g n o i s e}} \\\\ \\end{array}", + "image_path": "d1f92b1888d03e0ab1767acc0e8ae38139564a1d7698a718580bbc0c7e7d2614.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 610, + 333, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 333, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 333, + 620 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 326, + 625, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 625, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 326, + 625, + 545, + 654 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} ^ {t} := \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\alpha_ {i} ^ {t} \\quad a n d \\quad \\tilde {\\alpha} ^ {t} := \\frac {1}{M} \\sum_ {i = 1} ^ {M} | \\alpha_ {i} ^ {t} - \\overline {{\\alpha_ {i} ^ {t}}} |, \\tag {15}", + "image_path": "d25f6748e438dec5314324dadd6115db002c2a1c9de11aa3027a6c2f7b662750.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{t} = \\min (1,\\frac{C}{\\eta\\|\\sum_{k = 0}^{K - 1}\\tilde{\\mathbf{g}}^{t,k}(i)\\|})" + }, + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "type": "text", + "content": " , respectively. Note that " + }, + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}^t" + }, + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}^t" + }, + { + "bbox": [ + 305, + 661, + 545, + 713 + ], + "type": "text", + "content": " measure the on-average norm of local updates and local update consistency among clients before clipping and adding noise operations in DP-FedSAM, respectively." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24556" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 85, + 520, + 227 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "type": "text", + "content": "Table 1. Averaged training accuracy (%) and testing accuracy (%) on two data in both IID and Non-IID settings for all compared methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 85, + 520, + 227 + ], + "lines": [ + { + "bbox": [ + 74, + 85, + 520, + 227 + ], + "spans": [ + { + "bbox": [ + 74, + 85, + 520, + 227 + ], + "type": "table", + "html": "
TaskAlgorithmDirichlet 0.3Dirichlet 0.6IID
TrainValidationTrainValidationTrainValidation
EMNISTDP-FedAvg99.28±0.0273.10±0.1699.55±0.0282.20±0.3599.66±0.4081.90±0.86
Fed-SMP-randk99.24±0.0273.72±0.5399.71±0.0182.18±0.7399.71±0.6184.16±0.83
Fed-SMP-topk99.31±0.0475.75±0.3599.72±0.0283.41±0.9199.73±0.4083.32±0.52
DP-FedAvg-blur99.12±0.0273.71±0.0299.66±0.0083.20±0.0199.67±0.0382.92±0.49
DP-FedAvg-blurs99.63±0.0876.25±0.3599.72±0.0283.41±0.9199.74±0.4582.92±0.49
DP-FedSAM96.28±0.6476.81±0.8195.07±0.4584.32±0.1995.61±0.9485.90±0.72
DP-FedSAM-topk94.77±0.1177.27±0.6795.87±1.5284.80±0.6096.12±0.8587.70±0.83
CIFAR-10DP-FedAvg93.65±0.4747.98±0.2493.65±0.4250.05±0.4793.65±0.1550.90±0.86
Fed-SMP-randk95.46±0.4348.14±0.1295.36±0.0651.33±0.3695.36±0.0650.61±0.20
Fed-SMP-topk95.49±0.1449.93±2.2995.49±0.0954.11±0.8395.49±0.1053.30±0.45
DP-FedAvg-blur95.47±0.1247.66±0.0199.66±0.4251.05±0.0194.50±0.0552.56±0.47
DP-FedAvg-blurs96.79±0.5151.23±0.6699.72±0.0954.11±0.8396.45±0.3053.48±0.76
DP-FedSAM90.38±0.9053.92±0.5590.83±0.1554.14±0.6090.83±0.1655.58±0.50
DP-FedSAM-topk93.25±0.6054.85±0.8692.60±0.6557.00±0.6991.52±0.1158.82±0.51
", + "image_path": "5af9216a8de6ea61f70d6443d1ac6bdb2ec37fdf56d2a0d8c7aa0cb4a1f6bec4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "type": "text", + "content": "Remark 4. The proposed DP-FedSAM can achieve a tighter bound in general non-convex setting compared with previous work [9, 22] " + }, + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} +\\frac{6K\\sigma_g^2 + \\sigma_l^2}{T} +\\frac{B^2\\sum_{t = 1}^{T}(\\overline{\\alpha}^t + \\tilde{\\alpha}^t)}{T} +\\frac{L^2\\sqrt{T}\\sigma^2C^2d}{m^2\\sqrt{K}}\\right)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "content": "in [9] and " + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(\\frac{1}{\\sqrt{KT}} + \\frac{3\\sigma_g^2 + 2\\sigma_l^2}{\\sqrt{KT}} + \\frac{4L^2\\sqrt{T}\\sigma^2C^2d}{m^2\\sqrt{K}}\\right)" + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "content": " in [22]) and reduce the impacts of the local and global variance " + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\sigma_l^2" + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\sigma_g^2" + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "content": ". Meanwhile, we are the first to theoretically analyze these two impacts of both the on-average norm of local updates " + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}^t" + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "content": " and local update inconsistency among clients " + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}^t" + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "content": " on convergence. And the negative impacts of " + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}^t" + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}^t" + }, + { + "bbox": [ + 46, + 296, + 287, + 460 + ], + "type": "text", + "content": " are also significantly mitigated upon convergence compared with previous work [9] due to local SAM optimizer being adopted. It means that we effectively alleviate performance degradation caused by clipping operation in DP and achieve better performance under symmetric noise. This theoretical result has also been empirically verified on several real-world data (see Section 6.2 and 6.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 472, + 128, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 128, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 128, + 486 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 493, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 287, + 517 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to verify the effectiveness of DP-FedSAM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 526, + 155, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 155, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 155, + 540 + ], + "type": "text", + "content": "6.1. Experiment Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "content": "Dataset and Data Partition. The efficacy of DP-FedSAM is evaluated on three datasets, including EMNIST [10], CIFAR-10 and CIFAR-100 [28], in both IID and Non-IID settings. Specifically, Dir Partition [19] is used for simulating Non-IID across federated clients, where the local data of each client is partitioned by splitting the total dataset through sampling the label ratios from the Dirichlet distribution " + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\operatorname{Dir}(\\alpha)" + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\alpha = 0.3" + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Baselines. We focus on DPFL methods that ensure client-level DP. Thus we compare DP-FedSAM with existing DPFL baselines: DP-FedAvg [35] ensures client-level DP guarantee by directly employing Gaussian mechanism to the local updates. DP-FedAvg-blur [9] adds regularization method (BLUR) based on DP-FedAvg. DP-FedAvg-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "type": "text", + "content": "blurs [9] uses local update sparsification (LUS) and BLUR for improving the performance of DP-FedAvg. Fed-SMP-rand" + }, + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "type": "text", + "content": " and Fed-SMP-top" + }, + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "type": "text", + "content": " [22] leverage random sparsification and top" + }, + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 304, + 240, + 545, + 301 + ], + "type": "text", + "content": " sparsification technique for reducing the impact of DP noise on model accuracy, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": "Configuration. For EMNIST, we use a simple CNN model and train 200 communication round. For CIFAR-10 and CIFAR-100 datasets, we use the ResNet-18 [18] backbone and train 300 communication round. For all experiments, we set the number of clients " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": " to 500. The default sample ratio " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": " of the client is 0.1. The local learning rate " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": " is set to 0.1 with a decay rate 0.005 and momentum 0.5, and the number of training epochs is 30. For privacy parameters, noise multiplier " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": " is set to 0.95 and the privacy failure probability " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\delta = \\frac{1}{M}" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": ". The clipping threshold " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": " is selected by grid search from set " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.2, 0.4, 0.6, 0.8\\}" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": ", and we find that the gradient explosion phenomenon will occur when " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "C \\geq 0.6" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": " on EMNIST and " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "C = 0.2" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": " performs better on three datasets. The weight perturbation ratio is set to " + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\rho = 0.5" + }, + { + "bbox": [ + 304, + 301, + 546, + 492 + ], + "type": "text", + "content": ". We run each experiment 3 trials and report the best-averaged testing accuracy in each experiment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": "The results on EMNIST and CIFAR-10 datasets are placed in the main paper, while the results on CIFAR-100 are placed in Appendix A. For a more comprehensive and fair comparison, we integrate DP-FedSAM with " + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{top}_k" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": " sparsification technique [9, 22], named DP-FedSAM-top" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": " (More discussion in Appendix D), to compare with baselines that also use local sparsification technique, such as Fed-SMP-rand" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": ", Fed-SMP-top" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": ", and DP-FedAvg-blurs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 599, + 437, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 437, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 437, + 611 + ], + "type": "text", + "content": "6.2. Experiment Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Performance with compared baselines. In Table 1 and Figure 2, we evaluate DP-FedSAM and DP-FedSAM-top" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": " on EMNIST and CIFAR-10 datasets in both settings compared with all baselines from DP-FedAvg to DP-FedAvg-blurs. The baseline methods seem to be overfitting in Table 1, especially on more complex data (e.g., CIFAR-100 in Table 4). The main reasons are the random noise introduced by DP and the over-fitting and inconsistency of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24557" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 72, + 507, + 83 + ], + "blocks": [ + { + "bbox": [ + 84, + 72, + 507, + 83 + ], + "lines": [ + { + "bbox": [ + 84, + 72, + 507, + 83 + ], + "spans": [ + { + "bbox": [ + 84, + 72, + 507, + 83 + ], + "type": "image", + "image_path": "65e144f882e104093d5e2dad3579fd9aaa4ccc592ce78d6dd04a108259f5d36a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 87, + 216, + 194 + ], + "blocks": [ + { + "bbox": [ + 61, + 87, + 216, + 194 + ], + "lines": [ + { + "bbox": [ + 61, + 87, + 216, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 87, + 216, + 194 + ], + "type": "image", + "image_path": "f6ba9c1b77c2d679337c614e1582da3b71202878ee98d7d8938747b0b58c66f2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 220, + 87, + 373, + 194 + ], + "blocks": [ + { + "bbox": [ + 220, + 87, + 373, + 194 + ], + "lines": [ + { + "bbox": [ + 220, + 87, + 373, + 194 + ], + "spans": [ + { + "bbox": [ + 220, + 87, + 373, + 194 + ], + "type": "image", + "image_path": "84a365a24a78eaba7ebe195cc404243fffa39fb27fabe2708e04d21c4b86f0b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 198, + 318, + 206 + ], + "lines": [ + { + "bbox": [ + 276, + 198, + 318, + 206 + ], + "spans": [ + { + "bbox": [ + 276, + 198, + 318, + 206 + ], + "type": "text", + "content": "(a) EMNIST" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 378, + 87, + 531, + 194 + ], + "blocks": [ + { + "bbox": [ + 378, + 87, + 531, + 194 + ], + "lines": [ + { + "bbox": [ + 378, + 87, + 531, + 194 + ], + "spans": [ + { + "bbox": [ + 378, + 87, + 531, + 194 + ], + "type": "image", + "image_path": "5e532d1dd779901cf000973d06cf01b32c35a9e9900b5523d672c8b36202d9d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 209, + 216, + 316 + ], + "blocks": [ + { + "bbox": [ + 62, + 209, + 216, + 316 + ], + "lines": [ + { + "bbox": [ + 62, + 209, + 216, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 209, + 216, + 316 + ], + "type": "image", + "image_path": "a5f7a7630bcfb8fe2c8771d08177ba35012a6a70a068367606ef2c148b480bbb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 220, + 209, + 373, + 316 + ], + "blocks": [ + { + "bbox": [ + 220, + 209, + 373, + 316 + ], + "lines": [ + { + "bbox": [ + 220, + 209, + 373, + 316 + ], + "spans": [ + { + "bbox": [ + 220, + 209, + 373, + 316 + ], + "type": "image", + "image_path": "5ceb12fa0dab2d1eae8c85bd0e06ed5c6b71abec4e01fd30aba1bc458cf822c3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 319, + 320, + 329 + ], + "lines": [ + { + "bbox": [ + 273, + 319, + 320, + 329 + ], + "spans": [ + { + "bbox": [ + 273, + 319, + 320, + 329 + ], + "type": "text", + "content": "(b) CIFAR-10" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 378, + 209, + 531, + 316 + ], + "blocks": [ + { + "bbox": [ + 378, + 209, + 531, + 316 + ], + "lines": [ + { + "bbox": [ + 378, + 209, + 531, + 316 + ], + "spans": [ + { + "bbox": [ + 378, + 209, + 531, + 316 + ], + "type": "image", + "image_path": "60a6c0b4f8c25b9170556527db9eb89ecf8ae44e719ae6ff3c68a1fde9aea245.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": "the local models. From all these results, it is seen that our proposed algorithms outperform other baselines under symmetric noise both on accuracy and generalization perspectives. It means that we significantly improve the performance and generate a better trade-off between performance and privacy in DPFL. For instance, the averaged testing accuracy is " + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "85.90\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": " in DP-FedSAM and " + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "87.70\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": " in DP-FedSAM-" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{top}_k" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": " on EMNIST in the IID setting, which are better than other baselines. Meanwhile, the difference between training accuracy and test accuracy is " + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "9.71\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": " in DP-FedSAM and " + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "8.40\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": " in DP-FedSAM-" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{top}_k" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": ", while that is " + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "17.74\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": " in DP-FedAvg and " + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "16.41\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": " in Fed-SMP-" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{top}_k" + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": ", respectively. That means our algorithms significantly mitigate the performance degradation issue caused by DP." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": "Impact of Non-IID levels. In the experiments under different participation cases as shown in Table 1, we further prove the robust generalization of the proposed algorithms. Heterogeneous data distribution of local clients is set to various participation levels from IID, Dirichlet 0.6, and Dirichlet 0.3, which makes the training of the global model more difficult. On EMNIST, as the Non-IID level decreases, DP-FedSAM achieves better generalization than DP-FedAvg, and the difference between training accuracy and test accuracy in DP-FedSAM " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{19.47\\%, 10.75\\%, 9.71\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": " are lower than that in DP-FedAvg " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{26.18\\%, 17.35\\%, 17.74\\%\\}" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": ". Similarly, the difference values in DP-FedSAM-top " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{17.50\\%, 11.07\\%, 8.40\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": " are also lower than that in FedSMP-top " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{23.56\\%, 16.31\\%, 16.41\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": " . These observations confirm that our algorithms are more robust than baselines in various degrees of heterogeneous data." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 349, + 426, + 442 + ], + "blocks": [ + { + "bbox": [ + 69, + 335, + 523, + 346 + ], + "lines": [ + { + "bbox": [ + 69, + 335, + 523, + 346 + ], + "spans": [ + { + "bbox": [ + 69, + 335, + 523, + 346 + ], + "type": "text", + "content": "Figure 2. The averaged testing accuracy on EMNIST and CIFAR-10 dataset under symmetric noise for all compared methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 349, + 426, + 442 + ], + "lines": [ + { + "bbox": [ + 307, + 349, + 426, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 426, + 442 + ], + "type": "image", + "image_path": "7cb500fde4b4cddb7f0108917551b8b9db7049c4386b98117150111f77a84017.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 349, + 547, + 441 + ], + "blocks": [ + { + "bbox": [ + 427, + 349, + 547, + 441 + ], + "lines": [ + { + "bbox": [ + 427, + 349, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 427, + 349, + 547, + 441 + ], + "type": "image", + "image_path": "6b57967edc2b16b273e31020dd04b071e9d5afdf839bedcbe82fd7351458e7ac.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 452, + 541, + 463 + ], + "lines": [ + { + "bbox": [ + 310, + 452, + 541, + 463 + ], + "spans": [ + { + "bbox": [ + 310, + 452, + 541, + 463 + ], + "type": "text", + "content": "Figure 3. Norm distribution and average norm of local updates." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 474, + 491, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 491, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 491, + 487 + ], + "type": "text", + "content": "6.3. Discussion for DP with SAM in FL" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 494, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 577 + ], + "type": "text", + "content": "In this subsection, we empirically discuss how SAM mitigates the negative impacts of DP from the aspects of the norm of local update and the visualization of loss landscape and contour. Meanwhile, we also observe the training performance with SAM under different privacy budgets " + }, + { + "bbox": [ + 304, + 494, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 494, + 545, + 577 + ], + "type": "text", + "content": " compared with all baselines. These experiments are conducted on CIFAR-10 with ResNet-18 [18] and Dirichlet " + }, + { + "bbox": [ + 304, + 494, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 304, + 494, + 545, + 577 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "text", + "content": "The norm of local update. To validate the theoretical results for mitigating the adverse impacts of the norm of local updates, we conduct experiments on DP-FedSAM and DP-FedAvg with clipping threshold " + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "inline_equation", + "content": "C = 0.2" + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "text", + "content": " as shown in Figure 3. We show the norm " + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Delta_i^t" + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "text", + "content": " distribution and average norm " + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\overline{\\Delta}^t" + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "text", + "content": " of local updates before clipping during the communication rounds. In contrast to DP-FedAvg, most of the norm is distributed at the smaller value in our scheme in Figure 3 (a), which means that the clipping operation drops less information. Meanwhile, the on-average norm " + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\overline{\\Delta}^t" + }, + { + "bbox": [ + 304, + 578, + 546, + 713 + ], + "type": "text", + "content": " is smaller than DP-FedAvg as shown in Figure 3 (b). These observations" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24558" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 214, + 179 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 214, + 179 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 214, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 214, + 179 + ], + "type": "image", + "image_path": "7aae472875ddb2923a244935cd0bc9051df354467d317f19d4935c1babe152ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "type": "text", + "content": "Figure 4. Impact of hyper-parameters: perturbation radius " + }, + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "type": "text", + "content": ", local iteration steps " + }, + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "type": "text", + "content": ", total clients size " + }, + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 106, + 182, + 487, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 72, + 373, + 180 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 373, + 180 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 373, + 180 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 373, + 180 + ], + "type": "image", + "image_path": "fa9e6e717a03e344c4d35fad6ccb310b006bfd9c1ee772f29b0272c762142daa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 529, + 180 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 529, + 180 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 529, + 180 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 529, + 180 + ], + "type": "image", + "image_path": "7cd9917d068fd62f6854cf4ab6f8fd35a5d38b31a69b1702db84568a48d03a26.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 217, + 286, + 296 + ], + "blocks": [ + { + "bbox": [ + 47, + 202, + 286, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 202, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 202, + 286, + 213 + ], + "type": "text", + "content": "Table 2. Performance comparison under different privacy budgets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 217, + 286, + 296 + ], + "lines": [ + { + "bbox": [ + 50, + 217, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 50, + 217, + 286, + 296 + ], + "type": "table", + "html": "
AlgorithmAveraged test accuracy (%) under different privacy budgets ε
ε = 4ε = 6ε = 8ε = 10
DP-FedAvg38.23 ± 0.1543.87 ± 0.6246.74 ± 0.0349.06 ± 0.49
Fed-SMP-randk33.78 ± 0.9242.21 ± 0.2148.20 ± 0.0550.62 ± 0.14
Fed-SMP-topk38.99 ± 0.5046.24 ± 0.8049.78 ± 0.7852.51 ± 0.83
DP-FedAvg-blur38.23 ± 0.7043.93 ± 0.4846.74 ± 0.9249.06 ± 0.13
DP-FedAvg-blurs39.39 ± 0.4346.64 ± 0.3650.18 ± 0.2752.91 ± 0.57
DP-FedSAM39.89 ± 0.1747.92 ± 0.2351.30 ± 0.9553.18 ± 0.40
DP-FedSAM-topk38.96 ± 0.6149.17 ± 0.1553.64 ± 0.1256.36 ± 0.36
", + "image_path": "09279c4f0bb5dab4c3d26ad989bdaac74dffc934a72d462b4263567a61eee010.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": "are also consistent with our theoretical results in Section 5. Performance under different privacy budgets " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": ". Table 2 shows the test accuracies for the different level privacy guarantees. Note that " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": " is not a hyper-parameter, but can be obtained by the privacy design in each round such as Eq. (13). Our methods consistently outperform the previous SOTA methods under various privacy budgets " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": ". Specifically, DP-FedSAM and DP-FedSAM-top " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": " significantly improve the accuracy of DP-FedAvg and Fed-SMP-top " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "1\\% \\sim 4\\%" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "3\\% \\sim 4\\%" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": " under the same " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": ", respectively. Furthermore, the test accuracy can improve as the privacy budget " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": " increases, which suggests us a better balance is necessary between training performance and privacy. There is a better trade-off when achieving better accuracy under the same " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": " or a smaller " + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 304, + 286, + 483 + ], + "type": "text", + "content": " under approximately the same accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 483, + 286, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 286, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 286, + 507 + ], + "type": "text", + "content": "Loss landscape and contour. The discussion and visualizing results are placed in Appendix C due to limited space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 515, + 141, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 141, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 141, + 528 + ], + "type": "text", + "content": "6.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 557 + ], + "type": "text", + "content": "We verify the influence of hyper-parameters in DPFedSAM on EMNIST with Dirichlet partition " + }, + { + "bbox": [ + 46, + 534, + 286, + 557 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 46, + 534, + 286, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "text", + "content": "Perturbation weight " + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "text", + "content": ". Perturbation weight " + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "text", + "content": " has an impact on performance as the adding perturbation is accumulated when the communication round " + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "text", + "content": " increases. To select a proper value for our algorithms, we conduct some experiments on various perturbation radius from the set " + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "inline_equation", + "content": "\\{0.01, 0.1, 0.3, 0.5, 0.7, 1.0\\}" + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "text", + "content": " in Figure 4 (a). As " + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "inline_equation", + "content": "\\rho = 0.5" + }, + { + "bbox": [ + 46, + 558, + 286, + 641 + ], + "type": "text", + "content": ", we achieve better convergence and performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": "Local iteration steps " + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": ". Large local iteration steps " + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": " can help the convergence in previous DPFL work [9] with the theoretical guarantees. To investigate the acceleration on " + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": " by adopting a larger " + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": ", we fix the total batchsize and change local training epochs. In Figure 4 (b), our algorithm can accelerate the convergence in Theorem 3 as a larger " + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 217, + 545, + 268 + ], + "blocks": [ + { + "bbox": [ + 313, + 202, + 537, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 537, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 537, + 213 + ], + "type": "text", + "content": "Table 3. The averaged training accuracy and testing accuracy." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 217, + 545, + 268 + ], + "lines": [ + { + "bbox": [ + 309, + 217, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 309, + 217, + 545, + 268 + ], + "type": "table", + "html": "
AlgorithmTrain (%)Validation (%)Differential value (%)
DP-FedAvg99.55±0.0282.20± 0.3517.35±0.32
DP-FedSAM95.07±0.4584.32±0.19 ↑10.75±0.26 ↓
Fed-SMP-topk99.72±0.0283.41 ± 0.9116.31± 0.89
DP-FedSAM-topk95.87±0.5284.80±0.60 ↑11.07±0.08 ↓
", + "image_path": "688dc3c9b765fef2e92ba1a1431895c4ef09ad64018fb91b86d5987d2bdd4958.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "text", + "content": "is adopted, that is, use a larger epoch value. However, the adverse impact of clipping on training increases as " + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 279, + 545, + 315 + ], + "type": "text", + "content": " is too large, e.g., epoch = 40. Thus we choose 30 local epoch." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "content": "Client size " + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "content": ". We compare the performance between different numbers of client participation " + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "m = \\{100, 200, 300, 500, 700\\}" + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "content": " with the same hyper-parameters in Figure 4 (c). Compared with larger " + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "m = 500" + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "content": ", the smaller " + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "content": " may have worse performance due to large variance " + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\sigma^2 C^2 / m" + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "content": " in DP noise with the same setting. Meanwhile, when " + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "content": " is too large such as " + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "M = 700" + }, + { + "bbox": [ + 305, + 316, + 545, + 411 + ], + "type": "text", + "content": ", the performance may degrade as the local data size decreases." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "text", + "content": "Effect of SAM. As shown in Table 3, it is seen that DP-FedSAM and DP-FedSAM-top " + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "text", + "content": " can achieve performance improvement and better generalization compared with DP-FedAvg and Fed-SMP-top " + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "text", + "content": " as SAM optimizer is adopted with the same setting, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 486, + 379, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 379, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 379, + 498 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 506, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 662 + ], + "type": "text", + "content": "In this paper, we focus on severe performance degradation issue caused by dropped model information and exacerbated model inconsistency. And we are the first to alleviate this issue from the optimizer perspective and propose a novel and effective framework DP-FedSAM with a flatter loss landscape. Meanwhile, we present the analysis in detail of how SAM mitigates the adverse impacts of DP and achieve a tighter bound on convergence. Moreover, it is the first analysis to combine the impacts of the on-average norm of local updates and local update consistency among clients on training, and simultaneously provide the experimental observations. Finally, empirical results also verify the superiority of our approach on several real-world data." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement This work is supported by STI 2030—Major Projects (No. 2021ZD0201405), Shenzhen Philosophy and Social Science Foundation (Grant No. SZ2021B005)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24559" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Martin Abadi, Andy Chu, Ian J. Goodfellow, H. Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. CoRR, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning. In International Conference on Machine Learning, ICML, pages 10–32, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "type": "text", + "content": "[3] Naman Agarwal, Ananda Theertha Suresh, Felix X. Yu, Sanjiv Kumar, and Brendan McMahan. cpSGD: Communication-efficient and differentially-private distributed SGD. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 7575-7586, Dec. 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 235, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 235, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 235, + 288, + 289 + ], + "type": "text", + "content": "[4] Maksym Andriushchenko and Nicolas Flammarion. Towards understanding sharpness-aware minimization. In International Conference on Machine Learning, ICML, Proceedings of Machine Learning Research, pages 639-668. PMLR, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 289, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 288, + 324 + ], + "type": "text", + "content": "[5] Léon Bottou. Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT'2010, pages 177-186. Springer, 2010." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "type": "text", + "content": "[6] Léon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning. Siam Review, 60(2):223-311, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "type": "text", + "content": "[7] Debora Calderola, Barbara Caputo, and Marco Ciccone. Improving generalization in federated learning by seeking flat minima. CoRR, abs/2203.11834, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 425 + ], + "type": "text", + "content": "[8] Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. CoRR, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 425, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 480 + ], + "type": "text", + "content": "[9] Anda Cheng, Peisong Wang, Xi Sheryl Zhang, and Jian Cheng. Differentially private federated learning with local regularization and sparsification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 480, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 288, + 524 + ], + "type": "text", + "content": "[10] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. Emmist: Extending mnist to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 2921-2926. IEEE, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 524, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 288, + 579 + ], + "type": "text", + "content": "[11] Jiawei Du, Hanshu Yan, Jiashi Feng, Joey Tianyi Zhou, Liangli Zhen, Rick Siow Mong Goh, and Vincent Tan. Efficient sharpness-aware minimization for improved training of neural networks. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 579, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 288, + 613 + ], + "type": "text", + "content": "[12] Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy. Found. Trends Theor. Comput. Sci., pages 211-407, Aug. 2014." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 288, + 658 + ], + "type": "text", + "content": "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 658, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 288, + 714 + ], + "type": "text", + "content": "[14] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proc. ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 1322-1333, 2015." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Robin C. Geyer, Tassilo Klein, and Moin Nabi. Differentially private federated learning: A client level perspective. CoRR, Aug. 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[16] Robin C Geyer, Tassilo Klein, and Moin Nabi. Differentially private federated learning: A client level perspective. arXiv preprint arXiv:1712.07557, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[17] Saeed Ghadimi and Guanghui Lan. Stochastic first-and zeroth-order methods for nonconvex stochastic programming. SIAM Journal on Optimization, 23(4):2341-2368, 2013." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 185, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 230 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[19] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 275, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 330 + ], + "type": "text", + "content": "[20] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsification-amplified privacy and adaptive optimization. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1463-1469, Aug. 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "text", + "content": "[21] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. CoRR, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[22] Rui Hu, Yanmin Gong, and Yuanxiong Guo. Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy. arXiv preprint arXiv:2202.07178, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "type": "text", + "content": "[23] Tiansheng Huang, Shiwei Liu, Li Shen, Fengxiang He, Weiwei Lin, and Dacheng Tao. Achieving personalized federated learning with sparse local models. arXiv preprint arXiv:2201.11380, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 455, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 489 + ], + "type": "text", + "content": "[24] Zhuo Huang, Xiaobo Xia, Li Shen, Jun Yu, Chen Gong, Bo Han, and Tongliang Liu. Robust generalization against corruptions via worst-case sharpness minimization." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 489, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 534 + ], + "type": "text", + "content": "[25] Peter Kairouz, Ziyu Liu, and Thomas Steinke. The distributed discrete gaussian mechanism for federated learning with secure aggregation. In Proc. International Conference on Machine Learning (ICML), pages 5201-5212, Jul. 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "text", + "content": "[26] P. Kairouz, Ziyu Liu, and T. Steinke. The distributed discrete gaussian mechanism for federated learning with secure aggregation. ArXiv, abs/2102.06387, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 568, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 633 + ], + "type": "text", + "content": "[27] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, pages 1-210, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 634, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 658 + ], + "type": "text", + "content": "[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[29] Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi. Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks. In International Conference on Machine Learning, pages 5905-5914. PMLR, 2021." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24560" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[30] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[31] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE Signal Processing Magazine, pages 50-60, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[32] Yong Liu, Siqi Mai, Xiangning Chen, Cho-Jui Hsieh, and Yang You. Towards efficient and scalable sharpness-aware minimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12360-12370, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 274 + ], + "type": "text", + "content": "[33] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proc. International Conference on Artificial Intelligence and Statistics (AISTATS), volume 54, pages 1273-1282, April. 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "type": "text", + "content": "[34] H Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. Learning differentially private recurrent language models. arXiv preprint arXiv:1710.06963, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 354 + ], + "type": "text", + "content": "[35] H. Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. Learning differentially private recurrent language models. In Proc. International Conference on Learning Representations (ICLR), Apr. 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "text", + "content": "[36] H. B. McMahan, D. Ramage, Kunal Talwar, and L. Zhang. Learning differentially private recurrent language models. In ICLR, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "text", + "content": "[37] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In Proc. IEEE Symposium on Security and Privacy (SP), pages 691-706, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 499 + ], + "type": "text", + "content": "[38] Peng Mi, Li Shen, Tianhe Ren, Yiyi Zhou, Xiaoshuai Sun, Rongrong Ji, and Dacheng Tao. Make sharpness-aware minimization stronger: A sparsified perturbation approach. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[39] Ilya Mironov. Rényi differential privacy. In Proc. IEEE computer security foundations symposium (CSF), pages 263-275, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "type": "text", + "content": "[40] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In Proc. IEEE Symposium on Security and Privacy (SP), pages 739-753, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "text", + "content": "[41] Zhe Qu, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. Generalized federated learning via sharpness aware minimization. In International Conference on Machine Learning, ICML, pages 18250-18280, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[42] Sashank J. Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečný, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[43] Yifan Shi, Li Shen, Kang Wei, Yan Sun, Bo Yuan, Xueqian Wang, and Dacheng Tao. Improving the model con" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "sistency of decentralized federated learning. arXiv preprint arXiv:2302.04083, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[44] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In Proc. IEEE Symposium on Security and Privacy (SP), pages 3-18, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "text", + "content": "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 228 + ], + "type": "text", + "content": "[46] Hao Sun, Li Shen, Qihuang Zhong, Liang Ding, Shixiang Chen, Jingwei Sun, Jing Li, Guangzhong Sun, and Dacheng Tao. Adasam: Boosting sharpness-aware minimization with adaptive learning rate and momentum for training deep neural networks. arXiv preprint arXiv:2303.00565, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[47] Lichao Sun and Lingjuan Lyu. Federated model distillation with noise-free differential privacy. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1563-1570, Aug. 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 327 + ], + "type": "text", + "content": "[48] Lichao Sun, Jianwei Qian, and Xun Chen. LDP-FL: Practical private aggregation in federated learning with local differential privacy. In Proc. Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1571–1578, Aug. 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "text", + "content": "[49] Tao Sun, Dongsheng Li, and Bao Wang. Decentralized federated averaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "type": "text", + "content": "[50] Yan Sun, Li Shen, Tiansheng Huang, Liang Ding, and Dacheng Tao. Fedspeed: Larger local interval, less communication round, and higher generalization accuracy. In International Conference on Learning Representations." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "text", + "content": "[51] Om Thakkar, Galen Andrew, and H. B. McMahan. Differentially private learning with adaptive clipping. ArXiv, abs/1905.03871, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "text", + "content": "[52] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Hang Su, Bo Zhang, and H. Vincent Poor. User-level privacy-preserving federated learning: Analysis and performance optimization. IEEE Transactions on Mobile Computing, 21(9):3388-3401, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 492, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 536 + ], + "type": "text", + "content": "[53] Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-IID federated learning. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 613 + ], + "type": "text", + "content": "[54] Ashkan Yousefpour, Igor Shilov, Alexandre Sablayrolles, Davide Testuggine, Karthik Prasad, Mani Malek, John Nguyen, Sayan Gosh, Akash Bharadwaj, Jessica Zhao, Graham Cormode, and Ilya Mironov. Opacus: User-friendly differential privacy library in pytorch. In Proc. Privacy in Machine Learning (PriML) Workshop, NeurIPS, Virtual, Dec. 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "text", + "content": "[55] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10174-10183, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "type": "text", + "content": "[56] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In International Conference on Machine Learning, ICML, pages 26982-26992. PMLR, 2022." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24561" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[57] Qihuang Zhong, Liang Ding, Li Shen, Peng Mi, Juhua Liu, Bo Du, and Dacheng Tao. Improving sharpness-aware minimization with fisher mask for better generalization on language models. arXiv preprint arXiv:2210.05497, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[58] Yuqing Zhu, Xiang Yu, Yi-Hsuan Tsai, F. Pittaluga, M. Faraki, Manmohan Chandraker, and Yu-Xiang Wang. Voting-based approaches for differentially private federated learning. ArXiv, abs/2010.04851, 2020." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24562" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_content_list.json b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f2f7fe9474ede1effa1d53be2ba1d12ea77d831f --- /dev/null +++ b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_content_list.json @@ -0,0 +1,1487 @@ +[ + { + "type": "text", + "text": "Make-A-Story: Visual Memory Conditioned Consistent Story Generation", + "text_level": 1, + "bbox": [ + 114, + 94, + 854, + 116 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tanzila Rahman $^{1,3}$", + "bbox": [ + 174, + 143, + 323, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hsin-Ying Lee", + "bbox": [ + 364, + 145, + 491, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jian Ren2", + "bbox": [ + 531, + 145, + 609, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sergey Tulyakov2", + "bbox": [ + 651, + 145, + 794, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shweta Mahajan $^{1,3}$", + "bbox": [ + 321, + 162, + 473, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leonid Sigal $^{1,3,4}$", + "bbox": [ + 514, + 162, + 647, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of British Columbia 2Snap Inc.", + "bbox": [ + 295, + 180, + 673, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Vector Institute for AI $^{4}$ Canada CIFAR AI Chair", + "bbox": [ + 271, + 198, + 700, + 214 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99a51cea6c9dd088718dfa8e40386ec5f015d2766da0791222716c2f3bee1a1a.jpg", + "image_caption": [ + "Figure 1. Referential and consistent story visualization. Examples of more natural stories with references for the FlintstonesSV [16] and MUGEN [19] datasets (bottom text) compared to more typical but less natural story text (top). We extend the MUGEN dataset by introducing additional two characters (e.g. Lisa and Jhon) and four backgrounds (e.g. Sand, Grass, Stone and Dirt)." + ], + "image_footnote": [], + "bbox": [ + 107, + 233, + 859, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 436, + 312, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There has been a recent explosion of impressive generative models that can produce high quality images (or videos) conditioned on text descriptions. However, all such approaches rely on conditional sentences that contain unambiguous descriptions of scenes and main actors in them. Therefore employing such models for more complex task of story visualization, where naturally references and co-references exist, and one requires to reason about when to maintain consistency of actors and backgrounds across frames/scenes, and when not to, based on story progression, remains a challenge. In this work, we address the aforementioned challenges and propose a novel autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective reference resolution and learns to maintain scene and actor consistency when needed. To validate the effectiveness of our approach, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multisentence storylines. Our experiments for story generation on the MUGEN, the PororoSV [30] and the FlintstonesSV [16] dataset show that our method not only outperforms prior state-of-the-art in generating frames with high visual quality, which are consistent with the story, but also models appropriate correspondences between the characters and the background.", + "bbox": [ + 75, + 474, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 436, + 630, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal deep learning approaches have pushed the quality and the breadth of conditional generation tasks such as image captioning [23, 33, 37, 54, 57] and text-to-image synthesis [25, 44, 59, 61-63]. Owing to the technical leaps made in generative models, such as generative adversarial networks (GANs) [14], variational autoencoders (VAEs) [27] and the more recent diffusion models [21], approaches for text-to-image synthesis can now generate images with high visual fidelity representative of the textual descriptions. The captions, however, in such cases, are generally short self-contained sentences representing the high-level semantics of a scene. This is rather restrictive in the real-world applications [8, 30, 31] where fine-grained understanding of object interactions, motion and background information described by multiple sentences becomes necessary. One such task is that of story generation or visualization – the goal of which is to generate a sequence of illustrative image frames with coherent semantics given a sequence of sentences [30, 35, 36, 60].", + "bbox": [ + 496, + 460, + 892, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Characteristic features of a good visual story is high visual quality over multiple frames; this includes rendering of discernible objects, actors, poses and realistic interactions of those actors with objects and within the scene. Moreover, for text-based story generation it is crucial to maintain consistency between the generated frames and the multisentence descriptions. Not only the actor context, but also the background of the generated story should be in-line with the description demonstrating effortless transition and adaptation to the changing environments within the story [34].", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2493", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances on the task of story generation have made significant advances along these lines, showing high visual fidelity and character consistency for story sentences that are self-contained and unambiguous (explicitly mentioning characters and the setting each time). While impressive, this setup is fundamentally unrealistic. Realistic story text is considerably more complex and referential in nature; requiring ability to resolve ambiguity and references (or co-references) through reasoning. As shown in Fig. 1, while the description corresponding to the first frame has an explicit reference to the character names, the (typical) subsequent frame descriptions, provided by human, contain references such as \"she, he, they\". Moreover, while maintaining character consistency, current approaches are limited in preserving, or transitioning through, the background information in agreement with the text (cf. Fig. 3) [30, 35, 36].", + "bbox": [ + 76, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In natural language processing (NLP) co-reference resolution in text is an important and core task [3, 26, 38]. While it may be possible to apply such methods to story text to first resolve ambiguous references and then generate corresponding images using existing story generation approaches, this is sub-optimal. The reason, is that co-reference resolution in the text domain, at best, would only allow to resolve references and maintain consistency across identity of the character. Appearance across frames would still lack consistency and require some form of visual reasoning. As also noted in [47], reference resolution in the visual domain, or visio-lingual domain, is more powerful.", + "bbox": [ + 76, + 333, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, for the first time (to our knowledge), we study co-reference resolution in story generation. Prior work [35] offers limited performance when faced with text containing references (see Sec. 5). We address this by proposing a new autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective visio-lingual co-reference resolution and learns to maintain scene and actor consistency when needed. Further, given the lack of datasets that contain references and more complex sentence structure, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multi-sentence storylines.", + "bbox": [ + 76, + 513, + 472, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Our contributions are three-fold: (i) First, we introduce a novel autoregressive deep generative framework, Story-LDM, that adopts and extends latent diffusion models for the task of story generation. As part of Story-LDM, we propose a meticulously designed memory-attention mechanism capable of encoding and leveraging contextual relevance between the part of the story-line that has already been generated, and the current frame being generated based on learned semantic similarity of corresponding sentences. Equipped with this, our sequential diffusion model can generate consistent stories by resolving and", + "bbox": [ + 76, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "then capturing temporal character and background context. (ii) Second, to validate our approach for co-reference resolution, and character and background consistency in the visual domain, we extend existing datasets to include more complex scenarios and, importantly, referential text. Specifically, we extend the MUGEN dataset [19] to include multiple characters and diverse backgrounds. We also modify FlintstonesSV [16] and PororoSV [30] dataset to include character references. These enhancements allow us to increase the complexity of the aforementioned datasets by introducing co-references in the sentences of a story. (iii) Finally, to evaluate different approaches for foreground (character) as well as background consistency we propose novel evaluation metrics. Our results on the MUGEN [19], the PororoSV [30] and the FlintstonesSV [16] datasets show that we outperform the prior state-of-the-art on consistency metrics by a large margin.", + "bbox": [ + 496, + 90, + 893, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 364, + 638, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image synthesis. Deep generative models, particularly, generative adversarial networks (GANs) [14], variational autoencoders (VAEs) [27] and normalizing flows [4, 11, 12] have been applied to multimodal tasks at the intersection of vision and language. Typical such tasks include image captioning [1, 33, 56] and text-to-image synthesis [32, 44, 51, 59, 61, 63]. Early work on text conditioned image synthesis built upon the success of GANs [44]. More recent approaches have utilized multi-stage generators [61] and normalizing flow-based priors [32] in the latent space to model the distribution of images given text. Various approaches have found cross-domain contrastive loss to improve text-to-image generation models [25, 62]. DALLE [42] and Cogview [10] harness the power of transformers [53] and discrete variational autoencoders (VQ-VAE) [43] yielding very high quality image samples.", + "bbox": [ + 496, + 391, + 893, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recent are the advances in diffusion models which have revolutionized the domain of image generation [21]. Diffusion models progressively add noise to the data and learn a reverse diffusion process to reconstruct it. Nichol et al. [39] adapted diffusion models for text-to-image generation and explore CLIP [40] guided generation as well as classifier-free modeling. Standard diffusion models are employed directly in the high-dimensional pixel space and therefore, cannot directly be used for the more complex task of story generation. Recent work [6, 15, 45] instead use encodings from pre-trained models as input to the diffusion models, thereby reducing the complexity of the task by working in a lower-dimensional space. In this work, we build upon this idea and extend it for sequential story generation.", + "bbox": [ + 496, + 633, + 895, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-video synthesis. One of the challenges of text-to-video synthesis is the smoothness of motion in a video [30].", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2494", + "bbox": [ + 482, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/26839c372a8a93965549af37e50e979d1c1308ce2dd3208fa1d01d2a70c4ef9b.jpg", + "image_caption": [ + "(a) Story-LDM architecture for conditional story generation." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 524, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/55e6f0d2f9899dc3bb29a1e807fe335165c0d995c1387eb338d5f3b0482f31e8.jpg", + "image_caption": [ + "(b) Memory state with query and values for our memory-attention module.", + "Figure 2. Story-Latent Diffusion Models for consistent story generation. (a) The autoregressive conditional generative Story-LDM model with the proposed memory-attention module. (b) A snapshot of the memory for $k$ frames." + ], + "image_footnote": [], + "bbox": [ + 602, + 98, + 883, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early work focused on generating short clips [8, 31]. To effectively learn the motion, various approaches disentangle the motion features from the background information [18,52,55]. Wu et al. [58] propose a novel two-dimensional VQ-VAE and sparse attention module for real-world text-to-video generation. Singer et al. [48] decompose the temporal U-Net [46] and the attention modules to approximate them in space and time to extend the text-to-image diffusion models to model text-to-video generation. Ligong et al. [17] propose a transformer framework to jointly model various modalities.", + "bbox": [ + 75, + 299, + 470, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Story Generation. Li et al. [30] proposed the initial idea and task of story generation. A two-level StoryGAN framework is applied to ensure image-level consistency between each sentence and image pair, and a global discriminator enforces global consistency between the entire image sequence and the story. Various approaches have proposed improvements to the StoryGAN architecture. Zeng et al. [60] introduce sentence-level alignment and word-based attention to improve relevance. Li et al. [29] further improve the performance with enhanced discriminators and dilated convolutions. In [49] foreground-background information is provided as additional supervision and [34] use video captioning for semantic alignment between text and frames. Recently, Chen et al. [5] adopted visual planning and character token alignment to improve character consistency.", + "bbox": [ + 75, + 468, + 468, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Story Completion. Recently, another task for text-to-story synthesis referred to as story completion has been proposed [36]. In this task, in addition to sentences, the first frame of the story is provided as input. In effect, story completion is a simplified variant of story generation. StoryDALL-E [36] leverages models pre-trained for text-to-image synthesis to perform story completion. Datasets for this task include CLEVR-SV [30] and Pororo-SV [16] which are derived from the CLEVR dataset [22], and the Flintstones dataset for text-to-video synthesis has also been modified for the task of story visualization [34]. Additionally, to evaluate the generalization performance, popular DiDeMo dataset for video captioning [2] is adapted for the task in [36].", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reference Resolution. Co-reference resolution is an important and well-researched topic in NLP and focuses on resolving the pronouns and their associated entities. Classic methods in NLP to co-reference resolution employ decision trees [3,38], maximum-entropy modeling [26], cluster-ranking [41] and classification algorithms [50]. More recent approaches [13,24,28] leverage neural network architectures to obtain improved performance with Transformers [9]. Seo et al. [47] proposed visual co-reference resolution for the task of Visual Question-Answering (VQA) dialogs. We take inspiration from [47], but propose a much more sophisticated memory-attention module that allows us to perform visio-lingual co-reference resolution (and visual consistency modeling) for visual story generation.", + "bbox": [ + 496, + 299, + 893, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 527, + 607, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate temporally consistent stories based solely on the linguistic story-line, we develop a deep generative approach with autoregressive structure. We build upon the success of diffusion models in modeling the underlying data distribution of images to produce high quality samples, and learn the generative conditional distribution of the visual story based on the textual descriptions. Given that the multi-frame stories involve high-dimensional data input, we employ Latent Diffusion Models [45], such that diffusion models can be applied in a computationally efficient manner. Besides, to ensure temporal consistency and smooth story progression, we propose a novel memory attention mechanism which not only attends to the multimodal representations of the current frame but also takes into account the already generated semantics of the previous frames. This module also allows us to resolve ambiguous references (e.g., he/she, they, etc.) using visual memory and is the core of our technical contribution. We first provide an overview of the Diffusion Models and the Latent Diffusion Models, following which we present our autoregressive latent diffusion model for stories called Story-LDM1.", + "bbox": [ + 496, + 555, + 893, + 872 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 517, + 887, + 859, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2495", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The Latent Diffusion Model Backbone", + "text_level": 1, + "bbox": [ + 76, + 90, + 408, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models are a class of generative models that approximate the underlying data distribution $p(\\mathbf{x})$ by denoising a base (Gaussian) distribution in multiple steps using a reverse process of a fixed Markov Chain of length $T$ . To estimate $p(\\mathbf{x})$ , the forward diffusion process starts from the input data $\\mathbf{x}_0 = \\mathbf{x}$ and gradually adds noise to obtain a set of noisy samples $\\mathbf{x}_1, \\ldots, \\mathbf{x}_T$ such that $\\mathbf{x}_T \\sim \\mathcal{N}(0, 1)$ represents a sample from a Gaussian distribution. Under the Markov assumption, the probability of the forward process modeling the distribution $q(\\mathbf{x}_{0:T} \\mid \\mathbf{x}_0)$ and the reverse diffusion process estimating probability at an earlier time-step are formulated as:", + "bbox": [ + 75, + 112, + 472, + 295 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {i = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 301, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {T}\\right) \\prod_ {i = 1} ^ {T} p \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 345, + 387, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\{\\beta_i\\}_{i=1}^T$ is the variance schedule for each time-step such that $\\mathbf{x}_T$ is nearly a Gaussian. The model parameters $\\theta$ are learnt with the following objective,", + "bbox": [ + 75, + 392, + 468, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D M} := \\mathbb {E} _ {t, \\mathbf {x}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 444, + 468, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,1)$ and $\\epsilon_{\\theta}(\\mathbf{x}_t,t), t = 1,\\dots,T$ is a sequence of denoising autoencoders with noisy input $\\mathbf{x}_t$ predicting the noise that was added to the original input $\\mathbf{x}$ .", + "bbox": [ + 75, + 468, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite yielding state-of-the-art results in various image generation tasks, diffusion models directly operating in the high-dimensional pixel are computationally expensive and resource exhaustive. This limits their application to an even higher-dimensional data such as multi-frame stories or video datasets, which is the focus of this work.", + "bbox": [ + 75, + 513, + 468, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Models in the Latent Space. To broaden the applicability of the diffusion models to very high-dimensional data e.g. high-resolution images, Latent Diffusion Models (LDM) [45] first compress the original image to a lower-dimensional space using perceptual image compression. An auto-encoder approach is employed such that the original spatial structure of the input image is preserved in the latent space. That is, the encoder $E(\\cdot)$ maps the input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ to a latent representation $\\mathbf{Z} \\in \\mathbb{R}^{h \\times w \\times c}$ , downsampling the image to a lower spatial dimension. Following this, the diffusion model is applied to the latents $\\mathbf{Z}$ , where time-conditioned U-Net $\\epsilon_{\\theta}(\\mathbf{Z}_t, t)$ is employed to model the diffusion process. The objective of the diffusion model from Eq. (2) becomes,", + "bbox": [ + 75, + 613, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L D M} := \\mathbb {E} _ {t, E (\\mathbf {Z}), \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {Z} _ {t}, t) \\| _ {2} ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 830, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, a forward diffusion process is applied to generate $\\mathbf{Z}$ , which are mapped to the original image space using a decoder $D(\\cdot)$ .", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d37ecf9b63b0c25602b241554058ccb10a6bf446ad4fa17cf3809337ac707a32.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset# Ref (avg.)# Chars# Backgrounds
MUGEN [19]None12
Extended MUGEN336
FlintstonesSV [16]3.587323
Extended FlintstonesSV4.617323
PororoSV [30]1.019None
Extended PororoSV1.169None
", + "bbox": [ + 498, + 88, + 890, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Dataset statistics of the MUGEN, FlintstonesSV and PororoSV.", + "bbox": [ + 498, + 180, + 890, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Story-Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 498, + 215, + 769, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a textural story, characterized by sequence of $M$ sentences $\\mathbf{S}_{txt} = \\{\\mathbf{S}^0,\\dots ,\\mathbf{S}^M\\}$ , the goal of story generation is to produce a sequence of corresponding frames $\\mathbf{S}_{img} = \\{\\mathbf{I}^0,\\dots ,\\mathbf{I}^M\\}$ that visualize the story. We note that this is a more difficult problem than one of story continuation [36], where in addition to the textual story $\\mathbf{S}_{txt}$ approaches have access to a source frame $\\mathbf{I}^0$ for additional context at inference time. During training it is assumed that we have access to paired dataset of $N$ samples $\\mathcal{D} = \\{\\mathbf{S}_{txt}^{(i)},\\mathbf{S}_{img}^{(i)}\\}_{i = 1}^{N}$ . We extend latent diffusion models to this task, by allowing them to generate multi-frame stories autoregressively, and by introducing rich conditional structure that takes into account current sentence as well as context from earlier generated frames through visual memory module. This visual memory allows the model to incorporate character/background consistency and resolve text references when needed, resulting in improved performance.", + "bbox": [ + 496, + 239, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a condition $\\mathbf{y}$ , LDM utilizes a cross-attention layer with key $(\\mathbf{K})$ , query $(\\mathbf{Q})$ and value $(\\mathbf{V})$ where,", + "bbox": [ + 498, + 500, + 890, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (\\mathbf {K}, \\mathbf {Q}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right). \\mathbf {V}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 540, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{Q} = \\mathbf{W}_Q.\\hat{f}(\\mathbf{Z})$ , $\\mathbf{K} = \\mathbf{W}_K.f(\\mathbf{y})$ and $\\mathbf{V} = \\mathbf{W}_V.f(\\mathbf{y})$ , and $\\mathbf{W}_Q \\in \\mathbb{R}^{d \\times d_q}$ , $\\mathbf{W}_K \\in \\mathbb{R}^{d \\times d_k}$ and $\\mathbf{W}_V \\in \\mathbb{R}^{d \\times d_v}$ are learnable parameters, $\\hat{f}(\\mathbf{Z})$ an intermediate flattened feature representation of $\\mathbf{Z}$ within the diffusion model and $f(\\mathbf{y})$ the feature representation of the condition $\\mathbf{y}$ . The objective in Eq. (3) for conditional generation becomes,", + "bbox": [ + 496, + 587, + 890, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L D M} := \\mathbb {E} _ {t, E (\\mathbf {I}), \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {Z} _ {t}, f (\\mathbf {y}), t\\right) \\| _ {2} ^ {2} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 691, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the denoising autoencoders $\\epsilon_{\\theta}$ now additionally depend on the condition encoding $f(\\mathbf{y})$", + "bbox": [ + 496, + 720, + 888, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a sample $\\{\\mathbf{S}_{\\text {txt }}^{(i)}, \\mathbf{S}_{\\text {img }}^{(i)}\\}$ , we first project all the $M$ input frames of the story, $\\mathbf{I}^0, \\ldots, \\mathbf{I}^M$ onto a low-dimensional space using a frame-encoder $E$ , and obtain the encoded frames $\\mathbf{Z}^0, \\ldots, \\mathbf{Z}^M$ for a single story. To effectively condition on the corresponding frame, we apply this crossattention layer to the intermediate representations of the neural network for each frame $\\mathbf{I}^m$ and its textual description $\\mathbf{S}^m$ using Eq. (4), where the relevance of the description $\\mathbf{S}^m$", + "bbox": [ + 496, + 752, + 890, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2We drop the superscript denoting sample $i$ for ease of notation.", + "bbox": [ + 514, + 886, + 857, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2496", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dae1d99e3806c21f8ad4651d2e0fdf75ebc7ee5d4889f9e55e5fee5c3a609943.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodw/ ref. textChar-acc (↑)Char-F1 (↑)BG-acc (↑)BG-F1 (↑)FID (↓)
FlintstonesVLCStoryGAN [34]×27.7342.014.8316.49120.85
LDM [45]×79.8692.3348.0237.8661.40
LDM [45]✓57.3878.6844.1928.2587.39
Story-LDM (Ours)✓69.1986.5935.2128.8069.49
MUGENLDM [45]✓31.3921.2815.7418.66120.99
Story-LDM (Ours)✓93.4095.6092.1992.3762.16
", + "bbox": [ + 78, + 88, + 895, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Quantitative results. Experimental results on the FlintstoneSV and the MUGEN datasets.", + "bbox": [ + 189, + 215, + 777, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is weighted by the similarity between the textual representation and the encoded frame representation $\\mathbf{Z}^m$ (cf. Fig. 2a).", + "bbox": [ + 75, + 239, + 467, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For sequential generation, the model in addition to the current state, requires information from all the previous states. To enable this, the diffusion process for any frame representation $\\mathbf{Z}^m$ is conditioned on the visual representations of the previous frames $\\mathbf{Z}^0,\\ldots ,\\mathbf{Z}^{m - 1}$ as well as the sentence descriptions $\\mathbf{S}^0,\\dots ,\\mathbf{S}^m$ . This conditioning is realized through a novel Memory-attention module which forms the basis of our autoregressive approach.", + "bbox": [ + 75, + 270, + 468, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Memory-attention Module. To capture the spatiotemporal interactions across multiple frames and sentences for a story, in our conditional diffusion model, we condition the frame $\\mathbf{Z}^m$ not only on the corresponding text $\\mathbf{S}^m$ but also on the previous texts $\\mathbf{S}^i$ , for $i \\in \\{0, \\dots, m - 1\\}$ . This conditioning is applied throughout the $T$ time-steps of the diffusion process for $\\mathbf{Z}^m$ . The conditional denoising autoencoder thus models the conditional distribution $p(\\mathbf{Z}^m \\mid \\mathbf{Z}^{3 [45] which contains a cross-attention layer to generate text-to-image based story, without using our proposed autoregressive memory modules as our baselines for MU-", + "bbox": [ + 76, + 816, + 468, + 878 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/775c49c02dcde6ebe29c1b6b38dfff3cd87165434393d11eeeb3cf728fe40565.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison on Story Generation. Comparison on the FlintstonesSV dataset visual story generation." + ], + "image_footnote": [], + "bbox": [ + 545, + 87, + 848, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8dee7f301a52ce94e0debbfe835659f17ff1b026f4ec7955df4ec5b864851324.jpg", + "image_caption": [ + "Figure 6. Comparison to Story Continuation. Comparison on the FlintstonesSV dataset for story continuation. Prior work uses first frame as additional input to the model; our model does not." + ], + "image_footnote": [], + "bbox": [ + 540, + 373, + 849, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GEN, PororoSV and FlintstonesSV datasets. The parameters of the diffusion model within the Story-LDM are initialized with the pre-trained LDM [45]. Similarly, for the textual embedding, we use BERT-tokenizer [9] and use the pre-trained text-transformer from LDM.", + "bbox": [ + 498, + 746, + 890, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Table 2 shows quantitative results for consistent story generation on the FlintstoneSV dataset. We compare the performance of our approach (row 4) to the LDM [45] which we train/test with both original (row 2) as well as the co-referenced (row 3) descriptions. Fur", + "bbox": [ + 498, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3https://github.com/CompVis/latent-diffusion", + "bbox": [ + 93, + 886, + 442, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2499", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "thermore, we include the results of the state-of-art VLCStoryGAN [34] (row 1) with the original text of the dataset $^{4}$ (i.e. without co-references). We note that VLCStoryGAN was shown to be better than Duco-StoryGAN [35], CPCSV [49] and original StoryGAN [30] (see [34]).", + "bbox": [ + 76, + 90, + 468, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on Table 2 we make three observations: (1) Our LDM baseline is better than VLCStoryGAN on the original reference-free text (cf. Tab. 2, rows 1 & 2). (2) Reference resolution makes the task considerably harder. With the reference text in our modified dataset, we observe a drop in performance in terms of character and background classification scores (cf. Tab. 2, rows 2 & 3). (3) Our model, with memory-attention module, significantly outperforms the baseline (cf. Tab. 2, rows 3 & 4) both in terms of generative image quality and character consistency; and outperforms SoTA of VLCStoryGAN by $\\sim 41\\%$ percentage points on character accuracy (while performing a more difficult version of the task). Further, our model, that is required to conduct reference resolution, comes close to the LDM trained with original, reference-free, text (cf. Tab. 2, rows 3 & 4), which can be viewed as a sort of an upper bound.", + "bbox": [ + 75, + 167, + 470, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the MUGEN dataset, our method outperforms the strong LDM baseline with gains of $\\sim 62\\%$ on character accuracy and $\\sim 76\\%$ on the background accuracy, thereby showing the advantages of the memory-attention mechanism for consistent story generation. We note that MUGEN dataset has more references across story scenes. Flintstones while contains more references per story overall, many of those references are within scenes as opposed to across scenes. Meaning that in terms of reference impact on consistency, MUGEN dataset is actually harder. Experimental results on the PororoSV dataset are provided in the Supplemental.", + "bbox": [ + 75, + 409, + 468, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 3 illustrates qualitative results on the MUGEN dataset. Rows 1, 2 & 3 show ground truth, LDM [45] and our Story-LDM approach, respectively. Here, we see that our method is able to maintain consistency in terms of both character and background. Similarly, in Figure 4 we can show the results on FlintstoneSV dataset which further validates the strong performance of our method when generating high-quality, consistent story. Compared to the LDM, our approach is able to adapt to the diverse backgrounds in the story descriptions.", + "bbox": [ + 75, + 590, + 468, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional Results. We compare the qualitative results of our method to both story generation [34] and story continuation [36] in Figs. 5 and 6 respectively. The comparative images are taken directly from respective papers. We note that story continuation Fig. 6 is solving a different (easier) problem and with text that contains no-references. This makes the comparison to our method, which receives fewer inputs, not very meaningful. Nether-the-less, our approach, that", + "bbox": [ + 75, + 744, + 470, + 867 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46986fe82fab2965014bf4fc3db2b1a79ddc0d1ecb320ddd4020fb48417732a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 88, + 846, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0fa884a88d96beb5fc5fba5cc3ec403c5098351be5c29791dd2e5f409e4d4ab9.jpg", + "image_caption": [ + "Figure 7. Story Diversity. Diverse outputs for a single storyline obtained with our Story-LDM.", + "Figure 8. Branching Storyline. Generating different yet consistent stories by branching the storyline. Frames in the later columns are generated based on earlier ones and corresponding text." + ], + "image_footnote": [], + "bbox": [ + 540, + 333, + 846, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can resolve references and is solving a harder story generation task, obtains highly competitive results. Furthermore, to show that our autoregressive visual memory module can generate diverse stories conditioned on the current and previous condition, we create different story-lines starting for a single sentence. In Fig. 8, we can see for reference 'they', the model can generate both the characters according to the storyline already parsed. Moreover, in Fig. 7 we show that our approach can not only generate consistent visual stories, but also diverse frames for the same text (cf. Fig. 7). Additional results are provided in the Supplemental.", + "bbox": [ + 496, + 551, + 890, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 724, + 617, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we formulate consistent story generation in a more realistic way by co-referencing actors/backgrounds in the story descriptions. We develop an autoregressive Story-LDM approach with memory attention capable of maintaining consistency across the frames based on the previously generated frames and their corresponding descriptions. We introduced modified datasets to evaluate the performance for reference resolution. We expect our proposed formulation and models to be conducive to the real-world use cases and further the research.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Results for [34] were obtained using pretrained model provided by original authors in private communication.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2500", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jyoti Aneja, Harsh Agrawal, Dhruv Batra, and Alexander G. Schwing. Sequential latent spaces for modeling the intention during diverse image captioning. In ICCV, 2019. 2", + "[2] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 3", + "[3] Chinatsu Aone and Scott William. Evaluating automated and manual acquisition of anaphora resolution strategies. In Annual Meeting of the Association for Computational Linguistics, pages 122-129, 1995. 2, 3", + "[4] Apratim Bhattacharyya, Shweta Mahajan, Mario Fritz, Bernt Schiele, and Stefan Roth. Normalizing flows with multiscale autoregressive priors. In CVPR, 2020. 2", + "[5] Hong Chen, Rujun Han, Te-Lin Wu, Hideki Nakayama, and Nanyun Peng. Character-centric story visualization via visual planning and token alignment. In EMNLP, 2022. 3", + "[6] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. CVPR, 2023. 2", + "[7] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019. 7", + "[8] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018. 1, 3", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2018. 3, 7", + "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurIPS, 2021. 2", + "[11] Laurent Dinh, David Krueger, and Yoshua Bengio. NICE: Non-linear independent components estimation. In ICLR Workshop, 2015. 2", + "[12] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real NVP. In ICLR, 2017. 2", + "[13] Hongliang Fei, Xu Li, Dingcheng Li, and Ping Li. End-to-end deep reinforcement learning based coreference resolution. In Annual Meeting of the Association for Computational Linguistics, pages 660-665, 2019. 3", + "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2", + "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2", + "[16] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 1, 2, 3, 4, 7" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, 2022. 3", + "[18] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In CVPR, 2018. 3", + "[19] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Isabelle Hu, and Devi Parikh. Mugen: A playground for video-audio-text multimodal understanding and generation. arXiv preprint arXiv:2204.08058, 2022. 1, 2, 4, 7", + "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 2017. 7", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 2", + "[22] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3", + "[23] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In CVPR, 2016. 1", + "[24] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pretraining by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64-77, 2020. 3", + "[25] Minguk Kang and Jaesik Park. Contragan: Contrastive learning for conditional image generation. NeurIPS, 2020. 1, 2", + "[26] Andrew Kehler. Probabilistic coreference in information extraction. 1997. 2, 3", + "[27] Diederik P. Kingma and Max Welling. Auto-encoding variational Bayes. In ICLR, 2014. 1, 2", + "[28] Hideo Kobayashi, Yufang Hou, and Vincent Ng. End-to-end neural bridging resolution. In International Conference on Computational Linguistics, 2022. 3", + "[29] Chunye Li, Liya Kong, and Zhiping Zhou. Improved-storygan for sequential images visualization. Journal of Visual Communication and Image Representation, 2020. 3", + "[30] Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. Storygan: A sequential conditional gan for story visualization. In CVPR, 2019. 1, 2, 3, 4, 7, 8", + "[31] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 1, 3", + "[32] Shweta Mahajan, Iryna Gurevych, and Stefan Roth. Latent normalizing flows for many-to-many cross-domain mappings. In *ICLR*, 2020. 2", + "[33] Shweta Mahajan and Stefan Roth. Diverse image captioning with context-object split latent spaces. NeurIPS, 2020. 1, 2", + "[34] Adyasha Maharana and Mohit Bansal. Integrating visuospatial, linguistic and commonsense structure into story visualization. arXiv preprint arXiv:2110.10834, 2021. 1, 3, 5, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2501", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Improving generation and evaluation of visual stories via semantic consistency. 2021. 1, 2, 8", + "[36] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. In ECCV, 2022. 1, 2, 3, 4, 8", + "[37] Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L. Yuille. Deep captioning with multimodal recurrent neural networks (m-RNN). In ICLR, 2015. 1", + "[38] Joseph F McCarthy and Wendy G Lehnert. Using decision trees for coreference resolution. 1995. 2, 3", + "[39] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML, 2022. 2", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2", + "[41] Altaf Rahman and Vincent Ng. Narrowing the modeling gap: A cluster-ranking approach to coreference resolution. Journal of Artificial Intelligence Research, 2011. 3", + "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML. PMLR, 2021. 2", + "[43] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2", + "[44] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 1, 2", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8", + "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351 of Lecture Notes in Computer Science, pages 234-241. Springer, 2015. 3", + "[47] Paul Hongsuck Seo, Andreas Lehrmann, Bohyung Han, and Leonid Sigal. Visual reference resolution using attention memory for visual dialog. NIPS, 2017. 2, 3", + "[48] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + ], + "bbox": [ + 78, + 92, + 468, + 801 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yun-Zhu Song, Zhi Rui Tam, Hung-Jen Chen, Huiao-Han Lu, and Hong-Han Shuai. Character-preserving coherent story visualization. In ECCV, 2020. 3, 8", + "[50] Wee Meng Soon, Hwee Tou Ng, and Daniel Chung Yong Lim. A machine learning approach to coreference resolution of noun phrases. Computational linguistics, 2001. 3", + "[51] Hung-Yu Tseng, Hsin-Ying Lee, Lu Jiang, Ming-Hsuan Yang, and Weilong Yang. Retrievegan: Image synthesis via differentiable patch retrieval. In ECCV, 2020. 2", + "[52] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, 2018. 3", + "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2", + "[54] Ashwin K. Vijayakumar, Michael Cogswell, Ramprasaath R. Selvaraju, Qing Sun, Stefan Lee, David J. Crandall, and Dhruv Batra. Diverse beam search for improved description of complex scenes. In AAAI, 2018. 1", + "[55] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NIPS, 2016. 3", + "[56] Liwei Wang, Alexander G. Schwing, and Svetlana Lazebnik. Diverse and accurate image description using a variational auto-encoder with an additive Gaussian encoding space. In NIPS, 2017. 2", + "[57] Wei Wang, Xavier Alameda-Pineda, Dan Xu, Pascal Fua, Elisa Ricci, and Nicu Sebe. Every smile is unique: Landmark-guided diverse smile generation. In CVPR, 2018. 1", + "[58] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 3", + "[59] Guojun Yin, Bin Liu, Lu Sheng, Nenghai Yu, Xiaogang Wang, and Jing Shao. Semantics disentangling for text-to-image generation. CVPR, 2019. 1, 2", + "[60] Gangyan Zeng, Zhaohui Li, and Yuan Zhang. Pororogan: an improved story visualization model on pororo-sv dataset. In Proceedings of the 3rd International Conference on Computer Science and Artificial Intelligence, 2019. 1, 3", + "[61] Han Zhang, Ian J. Goodfellow, Dimitris N. Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, 2019. 1, 2", + "[62] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In CVPR, pages 833-842, 2021. 1, 2", + "[63] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmagan: Dynamic memory generative adversarial networks for text-to-image synthesis. In CVPR, 2019. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 801 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2502", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_model.json b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bab7541cab1df1caeb3c64c8b4671c19dbe3a3ac --- /dev/null +++ b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.095, + 0.856, + 0.117 + ], + "angle": 0, + "content": "Make-A-Story: Visual Memory Conditioned Consistent Story Generation" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.145, + 0.324, + 0.162 + ], + "angle": 0, + "content": "Tanzila Rahman\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.146, + 0.492, + 0.163 + ], + "angle": 0, + "content": "Hsin-Ying Lee" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.146, + 0.611, + 0.162 + ], + "angle": 0, + "content": "Jian Ren2" + }, + { + "type": "text", + "bbox": [ + 0.652, + 0.146, + 0.795, + 0.163 + ], + "angle": 0, + "content": "Sergey Tulyakov2" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.164, + 0.474, + 0.181 + ], + "angle": 0, + "content": "Shweta Mahajan\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.164, + 0.648, + 0.181 + ], + "angle": 0, + "content": "Leonid Sigal\\(^{1,3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.181, + 0.674, + 0.199 + ], + "angle": 0, + "content": "1University of British Columbia 2Snap Inc." + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.199, + 0.701, + 0.215 + ], + "angle": 0, + "content": "\\(^{3}\\)Vector Institute for AI \\(^{4}\\)Canada CIFAR AI Chair" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.234, + 0.861, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.382, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Figure 1. Referential and consistent story visualization. Examples of more natural stories with references for the FlintstonesSV [16] and MUGEN [19] datasets (bottom text) compared to more typical but less natural story text (top). We extend the MUGEN dataset by introducing additional two characters (e.g. Lisa and Jhon) and four backgrounds (e.g. Sand, Grass, Stone and Dirt)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.437, + 0.313, + 0.453 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.884 + ], + "angle": 0, + "content": "There has been a recent explosion of impressive generative models that can produce high quality images (or videos) conditioned on text descriptions. However, all such approaches rely on conditional sentences that contain unambiguous descriptions of scenes and main actors in them. Therefore employing such models for more complex task of story visualization, where naturally references and co-references exist, and one requires to reason about when to maintain consistency of actors and backgrounds across frames/scenes, and when not to, based on story progression, remains a challenge. In this work, we address the aforementioned challenges and propose a novel autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective reference resolution and learns to maintain scene and actor consistency when needed. To validate the effectiveness of our approach, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multisentence storylines. Our experiments for story generation on the MUGEN, the PororoSV [30] and the FlintstonesSV [16] dataset show that our method not only outperforms prior state-of-the-art in generating frames with high visual quality, which are consistent with the story, but also models appropriate correspondences between the characters and the background." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.437, + 0.631, + 0.453 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.748 + ], + "angle": 0, + "content": "Multimodal deep learning approaches have pushed the quality and the breadth of conditional generation tasks such as image captioning [23, 33, 37, 54, 57] and text-to-image synthesis [25, 44, 59, 61-63]. Owing to the technical leaps made in generative models, such as generative adversarial networks (GANs) [14], variational autoencoders (VAEs) [27] and the more recent diffusion models [21], approaches for text-to-image synthesis can now generate images with high visual fidelity representative of the textual descriptions. The captions, however, in such cases, are generally short self-contained sentences representing the high-level semantics of a scene. This is rather restrictive in the real-world applications [8, 30, 31] where fine-grained understanding of object interactions, motion and background information described by multiple sentences becomes necessary. One such task is that of story generation or visualization – the goal of which is to generate a sequence of illustrative image frames with coherent semantics given a sequence of sentences [30, 35, 36, 60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Characteristic features of a good visual story is high visual quality over multiple frames; this includes rendering of discernible objects, actors, poses and realistic interactions of those actors with objects and within the scene. Moreover, for text-based story generation it is crucial to maintain consistency between the generated frames and the multisentence descriptions. Not only the actor context, but also the background of the generated story should be in-line with the description demonstrating effortless transition and adaptation to the changing environments within the story [34]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2493" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "Recent advances on the task of story generation have made significant advances along these lines, showing high visual fidelity and character consistency for story sentences that are self-contained and unambiguous (explicitly mentioning characters and the setting each time). While impressive, this setup is fundamentally unrealistic. Realistic story text is considerably more complex and referential in nature; requiring ability to resolve ambiguity and references (or co-references) through reasoning. As shown in Fig. 1, while the description corresponding to the first frame has an explicit reference to the character names, the (typical) subsequent frame descriptions, provided by human, contain references such as \"she, he, they\". Moreover, while maintaining character consistency, current approaches are limited in preserving, or transitioning through, the background information in agreement with the text (cf. Fig. 3) [30, 35, 36]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.334, + 0.473, + 0.514 + ], + "angle": 0, + "content": "In natural language processing (NLP) co-reference resolution in text is an important and core task [3, 26, 38]. While it may be possible to apply such methods to story text to first resolve ambiguous references and then generate corresponding images using existing story generation approaches, this is sub-optimal. The reason, is that co-reference resolution in the text domain, at best, would only allow to resolve references and maintain consistency across identity of the character. Appearance across frames would still lack consistency and require some form of visual reasoning. As also noted in [47], reference resolution in the visual domain, or visio-lingual domain, is more powerful." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.473, + 0.727 + ], + "angle": 0, + "content": "In this work, for the first time (to our knowledge), we study co-reference resolution in story generation. Prior work [35] offers limited performance when faced with text containing references (see Sec. 5). We address this by proposing a new autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective visio-lingual co-reference resolution and learns to maintain scene and actor consistency when needed. Further, given the lack of datasets that contain references and more complex sentence structure, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multi-sentence storylines." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Contributions. Our contributions are three-fold: (i) First, we introduce a novel autoregressive deep generative framework, Story-LDM, that adopts and extends latent diffusion models for the task of story generation. As part of Story-LDM, we propose a meticulously designed memory-attention mechanism capable of encoding and leveraging contextual relevance between the part of the story-line that has already been generated, and the current frame being generated based on learned semantic similarity of corresponding sentences. Equipped with this, our sequential diffusion model can generate consistent stories by resolving and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.351 + ], + "angle": 0, + "content": "then capturing temporal character and background context. (ii) Second, to validate our approach for co-reference resolution, and character and background consistency in the visual domain, we extend existing datasets to include more complex scenarios and, importantly, referential text. Specifically, we extend the MUGEN dataset [19] to include multiple characters and diverse backgrounds. We also modify FlintstonesSV [16] and PororoSV [30] dataset to include character references. These enhancements allow us to increase the complexity of the aforementioned datasets by introducing co-references in the sentences of a story. (iii) Finally, to evaluate different approaches for foreground (character) as well as background consistency we propose novel evaluation metrics. Our results on the MUGEN [19], the PororoSV [30] and the FlintstonesSV [16] datasets show that we outperform the prior state-of-the-art on consistency metrics by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.365, + 0.64, + 0.381 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.633 + ], + "angle": 0, + "content": "Text-to-image synthesis. Deep generative models, particularly, generative adversarial networks (GANs) [14], variational autoencoders (VAEs) [27] and normalizing flows [4, 11, 12] have been applied to multimodal tasks at the intersection of vision and language. Typical such tasks include image captioning [1, 33, 56] and text-to-image synthesis [32, 44, 51, 59, 61, 63]. Early work on text conditioned image synthesis built upon the success of GANs [44]. More recent approaches have utilized multi-stage generators [61] and normalizing flow-based priors [32] in the latent space to model the distribution of images given text. Various approaches have found cross-domain contrastive loss to improve text-to-image generation models [25, 62]. DALLE [42] and Cogview [10] harness the power of transformers [53] and discrete variational autoencoders (VQ-VAE) [43] yielding very high quality image samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.896, + 0.86 + ], + "angle": 0, + "content": "More recent are the advances in diffusion models which have revolutionized the domain of image generation [21]. Diffusion models progressively add noise to the data and learn a reverse diffusion process to reconstruct it. Nichol et al. [39] adapted diffusion models for text-to-image generation and explore CLIP [40] guided generation as well as classifier-free modeling. Standard diffusion models are employed directly in the high-dimensional pixel space and therefore, cannot directly be used for the more complex task of story generation. Recent work [6, 15, 45] instead use encodings from pre-trained models as input to the diffusion models, thereby reducing the complexity of the task by working in a lower-dimensional space. In this work, we build upon this idea and extend it for sequential story generation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Text-to-video synthesis. One of the challenges of text-to-video synthesis is the smoothness of motion in a video [30]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "2494" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.526, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.24, + 0.462, + 0.253 + ], + "angle": 0, + "content": "(a) Story-LDM architecture for conditional story generation." + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.099, + 0.885, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.228, + 0.888, + 0.252 + ], + "angle": 0, + "content": "(b) Memory state with query and values for our memory-attention module." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.893, + 0.286 + ], + "angle": 0, + "content": "Figure 2. Story-Latent Diffusion Models for consistent story generation. (a) The autoregressive conditional generative Story-LDM model with the proposed memory-attention module. (b) A snapshot of the memory for \\( k \\) frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.471, + 0.465 + ], + "angle": 0, + "content": "Early work focused on generating short clips [8, 31]. To effectively learn the motion, various approaches disentangle the motion features from the background information [18,52,55]. Wu et al. [58] propose a novel two-dimensional VQ-VAE and sparse attention module for real-world text-to-video generation. Singer et al. [48] decompose the temporal U-Net [46] and the attention modules to approximate them in space and time to extend the text-to-image diffusion models to model text-to-video generation. Ligong et al. [17] propose a transformer framework to jointly model various modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.696 + ], + "angle": 0, + "content": "Story Generation. Li et al. [30] proposed the initial idea and task of story generation. A two-level StoryGAN framework is applied to ensure image-level consistency between each sentence and image pair, and a global discriminator enforces global consistency between the entire image sequence and the story. Various approaches have proposed improvements to the StoryGAN architecture. Zeng et al. [60] introduce sentence-level alignment and word-based attention to improve relevance. Li et al. [29] further improve the performance with enhanced discriminators and dilated convolutions. In [49] foreground-background information is provided as additional supervision and [34] use video captioning for semantic alignment between text and frames. Recently, Chen et al. [5] adopted visual planning and character token alignment to improve character consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Story Completion. Recently, another task for text-to-story synthesis referred to as story completion has been proposed [36]. In this task, in addition to sentences, the first frame of the story is provided as input. In effect, story completion is a simplified variant of story generation. StoryDALL-E [36] leverages models pre-trained for text-to-image synthesis to perform story completion. Datasets for this task include CLEVR-SV [30] and Pororo-SV [16] which are derived from the CLEVR dataset [22], and the Flintstones dataset for text-to-video synthesis has also been modified for the task of story visualization [34]. Additionally, to evaluate the generalization performance, popular DiDeMo dataset for video captioning [2] is adapted for the task in [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.513 + ], + "angle": 0, + "content": "Reference Resolution. Co-reference resolution is an important and well-researched topic in NLP and focuses on resolving the pronouns and their associated entities. Classic methods in NLP to co-reference resolution employ decision trees [3,38], maximum-entropy modeling [26], cluster-ranking [41] and classification algorithms [50]. More recent approaches [13,24,28] leverage neural network architectures to obtain improved performance with Transformers [9]. Seo et al. [47] proposed visual co-reference resolution for the task of Visual Question-Answering (VQA) dialogs. We take inspiration from [47], but propose a much more sophisticated memory-attention module that allows us to perform visio-lingual co-reference resolution (and visual consistency modeling) for visual story generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.608, + 0.547 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.895, + 0.873 + ], + "angle": 0, + "content": "To generate temporally consistent stories based solely on the linguistic story-line, we develop a deep generative approach with autoregressive structure. We build upon the success of diffusion models in modeling the underlying data distribution of images to produce high quality samples, and learn the generative conditional distribution of the visual story based on the textual descriptions. Given that the multi-frame stories involve high-dimensional data input, we employ Latent Diffusion Models [45], such that diffusion models can be applied in a computationally efficient manner. Besides, to ensure temporal consistency and smooth story progression, we propose a novel memory attention mechanism which not only attends to the multimodal representations of the current frame but also takes into account the already generated semantics of the previous frames. This module also allows us to resolve ambiguous references (e.g., he/she, they, etc.) using visual memory and is the core of our technical contribution. We first provide an overview of the Diffusion Models and the Latent Diffusion Models, following which we present our autoregressive latent diffusion model for stories called Story-LDM1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.86, + 0.901 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2495" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.409, + 0.105 + ], + "angle": 0, + "content": "3.1. The Latent Diffusion Model Backbone" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.473, + 0.296 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models are a class of generative models that approximate the underlying data distribution \\( p(\\mathbf{x}) \\) by denoising a base (Gaussian) distribution in multiple steps using a reverse process of a fixed Markov Chain of length \\( T \\). To estimate \\( p(\\mathbf{x}) \\), the forward diffusion process starts from the input data \\( \\mathbf{x}_0 = \\mathbf{x} \\) and gradually adds noise to obtain a set of noisy samples \\( \\mathbf{x}_1, \\ldots, \\mathbf{x}_T \\) such that \\( \\mathbf{x}_T \\sim \\mathcal{N}(0, 1) \\) represents a sample from a Gaussian distribution. Under the Markov assumption, the probability of the forward process modeling the distribution \\( q(\\mathbf{x}_{0:T} \\mid \\mathbf{x}_0) \\) and the reverse diffusion process estimating probability at an earlier time-step are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.302, + 0.469, + 0.35 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {i = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.347, + 0.388, + 0.387 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {T}\\right) \\prod_ {i = 1} ^ {T} p \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.44 + ], + "angle": 0, + "content": "Here, \\(\\{\\beta_i\\}_{i=1}^T\\) is the variance schedule for each time-step such that \\(\\mathbf{x}_T\\) is nearly a Gaussian. The model parameters \\(\\theta\\) are learnt with the following objective," + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.445, + 0.469, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D M} := \\mathbb {E} _ {t, \\mathbf {x}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.469, + 0.514 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,1)\\) and \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t), t = 1,\\dots,T\\) is a sequence of denoising autoencoders with noisy input \\(\\mathbf{x}_t\\) predicting the noise that was added to the original input \\(\\mathbf{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.469, + 0.605 + ], + "angle": 0, + "content": "Despite yielding state-of-the-art results in various image generation tasks, diffusion models directly operating in the high-dimensional pixel are computationally expensive and resource exhaustive. This limits their application to an even higher-dimensional data such as multi-frame stories or video datasets, which is the focus of this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.826 + ], + "angle": 0, + "content": "Diffusion Models in the Latent Space. To broaden the applicability of the diffusion models to very high-dimensional data e.g. high-resolution images, Latent Diffusion Models (LDM) [45] first compress the original image to a lower-dimensional space using perceptual image compression. An auto-encoder approach is employed such that the original spatial structure of the input image is preserved in the latent space. That is, the encoder \\( E(\\cdot) \\) maps the input image \\( \\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3} \\) to a latent representation \\( \\mathbf{Z} \\in \\mathbb{R}^{h \\times w \\times c} \\), downsampling the image to a lower spatial dimension. Following this, the diffusion model is applied to the latents \\( \\mathbf{Z} \\), where time-conditioned U-Net \\( \\epsilon_{\\theta}(\\mathbf{Z}_t, t) \\) is employed to model the diffusion process. The objective of the diffusion model from Eq. (2) becomes," + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.832, + 0.469, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L D M} := \\mathbb {E} _ {t, E (\\mathbf {Z}), \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {Z} _ {t}, t) \\| _ {2} ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "During training, a forward diffusion process is applied to generate \\(\\mathbf{Z}\\), which are mapped to the original image space using a decoder \\(D(\\cdot)\\)." + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.176 + ], + "angle": 0, + "content": "
Dataset# Ref (avg.)# Chars# Backgrounds
MUGEN [19]None12
Extended MUGEN336
FlintstonesSV [16]3.587323
Extended FlintstonesSV4.617323
PororoSV [30]1.019None
Extended PororoSV1.169None
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.181, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Table 1. Dataset statistics of the MUGEN, FlintstonesSV and PororoSV." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.217, + 0.771, + 0.234 + ], + "angle": 0, + "content": "3.2. Story-Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.892, + 0.5 + ], + "angle": 0, + "content": "Given a textural story, characterized by sequence of \\(M\\) sentences \\(\\mathbf{S}_{txt} = \\{\\mathbf{S}^0,\\dots ,\\mathbf{S}^M\\}\\), the goal of story generation is to produce a sequence of corresponding frames \\(\\mathbf{S}_{img} = \\{\\mathbf{I}^0,\\dots ,\\mathbf{I}^M\\}\\) that visualize the story. We note that this is a more difficult problem than one of story continuation [36], where in addition to the textual story \\(\\mathbf{S}_{txt}\\) approaches have access to a source frame \\(\\mathbf{I}^0\\) for additional context at inference time. During training it is assumed that we have access to paired dataset of \\(N\\) samples \\(\\mathcal{D} = \\{\\mathbf{S}_{txt}^{(i)},\\mathbf{S}_{img}^{(i)}\\}_{i = 1}^{N}\\). We extend latent diffusion models to this task, by allowing them to generate multi-frame stories autoregressively, and by introducing rich conditional structure that takes into account current sentence as well as context from earlier generated frames through visual memory module. This visual memory allows the model to incorporate character/background consistency and resolve text references when needed, resulting in improved performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.5, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Given a condition \\(\\mathbf{y}\\), LDM utilizes a cross-attention layer with key \\((\\mathbf{K})\\), query \\((\\mathbf{Q})\\) and value \\((\\mathbf{V})\\) where," + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.541, + 0.892, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (\\mathbf {K}, \\mathbf {Q}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right). \\mathbf {V}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{Q} = \\mathbf{W}_Q.\\hat{f}(\\mathbf{Z})\\), \\(\\mathbf{K} = \\mathbf{W}_K.f(\\mathbf{y})\\) and \\(\\mathbf{V} = \\mathbf{W}_V.f(\\mathbf{y})\\), and \\(\\mathbf{W}_Q \\in \\mathbb{R}^{d \\times d_q}\\), \\(\\mathbf{W}_K \\in \\mathbb{R}^{d \\times d_k}\\) and \\(\\mathbf{W}_V \\in \\mathbb{R}^{d \\times d_v}\\) are learnable parameters, \\(\\hat{f}(\\mathbf{Z})\\) an intermediate flattened feature representation of \\(\\mathbf{Z}\\) within the diffusion model and \\(f(\\mathbf{y})\\) the feature representation of the condition \\(\\mathbf{y}\\). The objective in Eq. (3) for conditional generation becomes," + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.693, + 0.892, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L D M} := \\mathbb {E} _ {t, E (\\mathbf {I}), \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {Z} _ {t}, f (\\mathbf {y}), t\\right) \\| _ {2} ^ {2} \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.722, + 0.89, + 0.752 + ], + "angle": 0, + "content": "Note that the denoising autoencoders \\(\\epsilon_{\\theta}\\) now additionally depend on the condition encoding \\(f(\\mathbf{y})\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.892, + 0.879 + ], + "angle": 0, + "content": "For a sample \\(\\{\\mathbf{S}_{\\text {txt }}^{(i)}, \\mathbf{S}_{\\text {img }}^{(i)}\\}\\), we first project all the \\(M\\) input frames of the story, \\(\\mathbf{I}^0, \\ldots, \\mathbf{I}^M\\) onto a low-dimensional space using a frame-encoder \\(E\\), and obtain the encoded frames \\(\\mathbf{Z}^0, \\ldots, \\mathbf{Z}^M\\) for a single story. To effectively condition on the corresponding frame, we apply this crossattention layer to the intermediate representations of the neural network for each frame \\(\\mathbf{I}^m\\) and its textual description \\(\\mathbf{S}^m\\) using Eq. (4), where the relevance of the description \\(\\mathbf{S}^m\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.858, + 0.901 + ], + "angle": 0, + "content": "2We drop the superscript denoting sample \\(i\\) for ease of notation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2496" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.896, + 0.213 + ], + "angle": 0, + "content": "
DatasetMethodw/ ref. textChar-acc (↑)Char-F1 (↑)BG-acc (↑)BG-F1 (↑)FID (↓)
FlintstonesVLCStoryGAN [34]×27.7342.014.8316.49120.85
LDM [45]×79.8692.3348.0237.8661.40
LDM [45]✓57.3878.6844.1928.2587.39
Story-LDM (Ours)✓69.1986.5935.2128.8069.49
MUGENLDM [45]✓31.3921.2815.7418.66120.99
Story-LDM (Ours)✓93.4095.6092.1992.3762.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.19, + 0.216, + 0.778, + 0.23 + ], + "angle": 0, + "content": "Table 2. Quantitative results. Experimental results on the FlintstoneSV and the MUGEN datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.468, + 0.27 + ], + "angle": 0, + "content": "is weighted by the similarity between the textual representation and the encoded frame representation \\(\\mathbf{Z}^m\\) (cf. Fig. 2a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.47, + 0.391 + ], + "angle": 0, + "content": "For sequential generation, the model in addition to the current state, requires information from all the previous states. To enable this, the diffusion process for any frame representation \\(\\mathbf{Z}^m\\) is conditioned on the visual representations of the previous frames \\(\\mathbf{Z}^0,\\ldots ,\\mathbf{Z}^{m - 1}\\) as well as the sentence descriptions \\(\\mathbf{S}^0,\\dots ,\\mathbf{S}^m\\). This conditioning is realized through a novel Memory-attention module which forms the basis of our autoregressive approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Memory-attention Module. To capture the spatiotemporal interactions across multiple frames and sentences for a story, in our conditional diffusion model, we condition the frame \\(\\mathbf{Z}^m\\) not only on the corresponding text \\(\\mathbf{S}^m\\) but also on the previous texts \\(\\mathbf{S}^i\\), for \\(i \\in \\{0, \\dots, m - 1\\}\\). This conditioning is applied throughout the \\(T\\) time-steps of the diffusion process for \\(\\mathbf{Z}^m\\). The conditional denoising autoencoder thus models the conditional distribution \\(p(\\mathbf{Z}^m \\mid \\mathbf{Z}^{3 [45] which contains a cross-attention layer to generate text-to-image based story, without using our proposed autoregressive memory modules as our baselines for MU-" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.088, + 0.849, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.343, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison on Story Generation. Comparison on the FlintstonesSV dataset visual story generation." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.374, + 0.851, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.693, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Figure 6. Comparison to Story Continuation. Comparison on the FlintstonesSV dataset for story continuation. Prior work uses first frame as additional input to the model; our model does not." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.747, + 0.892, + 0.822 + ], + "angle": 0, + "content": "GEN, PororoSV and FlintstonesSV datasets. The parameters of the diffusion model within the Story-LDM are initialized with the pre-trained LDM [45]. Similarly, for the textual embedding, we use BERT-tokenizer [9] and use the pre-trained text-transformer from LDM." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Quantitative Results. Table 2 shows quantitative results for consistent story generation on the FlintstoneSV dataset. We compare the performance of our approach (row 4) to the LDM [45] which we train/test with both original (row 2) as well as the co-referenced (row 3) descriptions. Fur" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.887, + 0.444, + 0.9 + ], + "angle": 0, + "content": "3https://github.com/CompVis/latent-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2499" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "thermore, we include the results of the state-of-art VLCStoryGAN [34] (row 1) with the original text of the dataset\\(^{4}\\) (i.e. without co-references). We note that VLCStoryGAN was shown to be better than Duco-StoryGAN [35], CPCSV [49] and original StoryGAN [30] (see [34])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.409 + ], + "angle": 0, + "content": "Based on Table 2 we make three observations: (1) Our LDM baseline is better than VLCStoryGAN on the original reference-free text (cf. Tab. 2, rows 1 & 2). (2) Reference resolution makes the task considerably harder. With the reference text in our modified dataset, we observe a drop in performance in terms of character and background classification scores (cf. Tab. 2, rows 2 & 3). (3) Our model, with memory-attention module, significantly outperforms the baseline (cf. Tab. 2, rows 3 & 4) both in terms of generative image quality and character consistency; and outperforms SoTA of VLCStoryGAN by \\(\\sim 41\\%\\) percentage points on character accuracy (while performing a more difficult version of the task). Further, our model, that is required to conduct reference resolution, comes close to the LDM trained with original, reference-free, text (cf. Tab. 2, rows 3 & 4), which can be viewed as a sort of an upper bound." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.589 + ], + "angle": 0, + "content": "On the MUGEN dataset, our method outperforms the strong LDM baseline with gains of \\(\\sim 62\\%\\) on character accuracy and \\(\\sim 76\\%\\) on the background accuracy, thereby showing the advantages of the memory-attention mechanism for consistent story generation. We note that MUGEN dataset has more references across story scenes. Flintstones while contains more references per story overall, many of those references are within scenes as opposed to across scenes. Meaning that in terms of reference impact on consistency, MUGEN dataset is actually harder. Experimental results on the PororoSV dataset are provided in the Supplemental." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.744 + ], + "angle": 0, + "content": "Qualitative Results. Figure 3 illustrates qualitative results on the MUGEN dataset. Rows 1, 2 & 3 show ground truth, LDM [45] and our Story-LDM approach, respectively. Here, we see that our method is able to maintain consistency in terms of both character and background. Similarly, in Figure 4 we can show the results on FlintstoneSV dataset which further validates the strong performance of our method when generating high-quality, consistent story. Compared to the LDM, our approach is able to adapt to the diverse backgrounds in the story descriptions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Additional Results. We compare the qualitative results of our method to both story generation [34] and story continuation [36] in Figs. 5 and 6 respectively. The comparative images are taken directly from respective papers. We note that story continuation Fig. 6 is solving a different (easier) problem and with text that contains no-references. This makes the comparison to our method, which receives fewer inputs, not very meaningful. Nether-the-less, our approach, that" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.089, + 0.847, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.306, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Figure 7. Story Diversity. Diverse outputs for a single storyline obtained with our Story-LDM." + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.334, + 0.848, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Figure 8. Branching Storyline. Generating different yet consistent stories by branching the storyline. Frames in the later columns are generated based on earlier ones and corresponding text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.72 + ], + "angle": 0, + "content": "can resolve references and is solving a harder story generation task, obtains highly competitive results. Furthermore, to show that our autoregressive visual memory module can generate diverse stories conditioned on the current and previous condition, we create different story-lines starting for a single sentence. In Fig. 8, we can see for reference 'they', the model can generate both the characters according to the storyline already parsed. Moreover, in Fig. 7 we show that our approach can not only generate consistent visual stories, but also diverse frames for the same text (cf. Fig. 7). Additional results are provided in the Supplemental." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.619, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we formulate consistent story generation in a more realistic way by co-referencing actors/backgrounds in the story descriptions. We develop an autoregressive Story-LDM approach with memory attention capable of maintaining consistency across the frames based on the previously generated frames and their corresponding descriptions. We introduced modified datasets to evaluate the performance for reference resolution. We expect our proposed formulation and models to be conducive to the real-world use cases and further the research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Results for [34] were obtained using pretrained model provided by original authors in private communication." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2500" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.157 + ], + "angle": 0, + "content": "[1] Jyoti Aneja, Harsh Agrawal, Dhruv Batra, and Alexander G. Schwing. Sequential latent spaces for modeling the intention during diverse image captioning. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.469, + 0.213 + ], + "angle": 0, + "content": "[2] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.469, + 0.271 + ], + "angle": 0, + "content": "[3] Chinatsu Aone and Scott William. Evaluating automated and manual acquisition of anaphora resolution strategies. In Annual Meeting of the Association for Computational Linguistics, pages 122-129, 1995. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[4] Apratim Bhattacharyya, Shweta Mahajan, Mario Fritz, Bernt Schiele, and Stefan Roth. Normalizing flows with multiscale autoregressive priors. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[5] Hong Chen, Rujun Han, Te-Lin Wu, Hideki Nakayama, and Nanyun Peng. Character-centric story visualization via visual planning and token alignment. In EMNLP, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[6] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[7] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[8] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[11] Laurent Dinh, David Krueger, and Yoshua Bengio. NICE: Non-linear independent components estimation. In ICLR Workshop, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[12] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real NVP. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[13] Hongliang Fei, Xu Li, Dingcheng Li, and Ping Li. End-to-end deep reinforcement learning based coreference resolution. In Annual Meeting of the Association for Computational Linguistics, pages 660-665, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[16] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 1, 2, 3, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[17] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[18] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[19] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Isabelle Hu, and Devi Parikh. Mugen: A playground for video-audio-text multimodal understanding and generation. arXiv preprint arXiv:2204.08058, 2022. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.396 + ], + "angle": 0, + "content": "[22] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[23] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[24] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pretraining by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64-77, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[25] Minguk Kang and Jaesik Park. Contragan: Contrastive learning for conditional image generation. NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[26] Andrew Kehler. Probabilistic coreference in information extraction. 1997. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[27] Diederik P. Kingma and Max Welling. Auto-encoding variational Bayes. In ICLR, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[28] Hideo Kobayashi, Yufang Hou, and Vincent Ng. End-to-end neural bridging resolution. In International Conference on Computational Linguistics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[29] Chunye Li, Liya Kong, and Zhiping Zhou. Improved-storygan for sequential images visualization. Journal of Visual Communication and Image Representation, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[30] Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. Storygan: A sequential conditional gan for story visualization. In CVPR, 2019. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[31] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[32] Shweta Mahajan, Iryna Gurevych, and Stefan Roth. Latent normalizing flows for many-to-many cross-domain mappings. In *ICLR*, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[33] Shweta Mahajan and Stefan Roth. Diverse image captioning with context-object split latent spaces. NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[34] Adyasha Maharana and Mohit Bansal. Integrating visuospatial, linguistic and commonsense structure into story visualization. arXiv preprint arXiv:2110.10834, 2021. 1, 3, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2501" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[35] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Improving generation and evaluation of visual stories via semantic consistency. 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[36] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. In ECCV, 2022. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[37] Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L. Yuille. Deep captioning with multimodal recurrent neural networks (m-RNN). In ICLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[38] Joseph F McCarthy and Wendy G Lehnert. Using decision trees for coreference resolution. 1995. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[39] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.399, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[41] Altaf Rahman and Vincent Ng. Narrowing the modeling gap: A cluster-ranking approach to coreference resolution. Journal of Artificial Intelligence Research, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.496, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[43] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[44] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351 of Lecture Notes in Computer Science, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[47] Paul Hongsuck Seo, Andreas Lehrmann, Bohyung Han, and Leonid Sigal. Visual reference resolution using attention memory for visual dialog. NIPS, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[48] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Yun-Zhu Song, Zhi Rui Tam, Hung-Jen Chen, Huiao-Han Lu, and Hong-Han Shuai. Character-preserving coherent story visualization. In ECCV, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[50] Wee Meng Soon, Hwee Tou Ng, and Daniel Chung Yong Lim. A machine learning approach to coreference resolution of noun phrases. Computational linguistics, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[51] Hung-Yu Tseng, Hsin-Ying Lee, Lu Jiang, Ming-Hsuan Yang, and Weilong Yang. Retrievegan: Image synthesis via differentiable patch retrieval. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[52] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.266, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.367 + ], + "angle": 0, + "content": "[54] Ashwin K. Vijayakumar, Michael Cogswell, Ramprasaath R. Selvaraju, Qing Sun, Stefan Lee, David J. Crandall, and Dhruv Batra. Diverse beam search for improved description of complex scenes. In AAAI, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[55] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NIPS, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.398, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[56] Liwei Wang, Alexander G. Schwing, and Svetlana Lazebnik. Diverse and accurate image description using a variational auto-encoder with an additive Gaussian encoding space. In NIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[57] Wei Wang, Xavier Alameda-Pineda, Dan Xu, Pascal Fua, Elisa Ricci, and Nicu Sebe. Every smile is unique: Landmark-guided diverse smile generation. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.57 + ], + "angle": 0, + "content": "[58] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[59] Guojun Yin, Bin Liu, Lu Sheng, Nenghai Yu, Xiaogang Wang, and Jing Shao. Semantics disentangling for text-to-image generation. CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[60] Gangyan Zeng, Zhaohui Li, and Yuan Zhang. Pororogan: an improved story visualization model on pororo-sv dataset. In Proceedings of the 3rd International Conference on Computer Science and Artificial Intelligence, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[61] Han Zhang, Ian J. Goodfellow, Dimitris N. Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[62] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In CVPR, pages 833-842, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[63] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmagan: Dynamic memory generative adversarial networks for text-to-image synthesis. In CVPR, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2502" + } + ] +] \ No newline at end of file diff --git a/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_origin.pdf b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e5d4ad831007416a2b49c3cc5249f717644333e8 --- /dev/null +++ b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/3ce6b598-c9ba-4565-b293-c61066c7677c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7473205f2b3f6de50cb136717bc6efda5260257a41ba1364de6a432891ff43c +size 9655887 diff --git a/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/full.md b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad5e2ad1eb428ba1f6ac993597c3c681b02cc1b3 --- /dev/null +++ b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/full.md @@ -0,0 +1,307 @@ +# Make-A-Story: Visual Memory Conditioned Consistent Story Generation + +Tanzila Rahman $^{1,3}$ + +Hsin-Ying Lee + +Jian Ren2 + +Sergey Tulyakov2 + +Shweta Mahajan $^{1,3}$ + +Leonid Sigal $^{1,3,4}$ + +1University of British Columbia 2Snap Inc. + +$^{3}$ Vector Institute for AI $^{4}$ Canada CIFAR AI Chair + +![](images/99a51cea6c9dd088718dfa8e40386ec5f015d2766da0791222716c2f3bee1a1a.jpg) +Figure 1. Referential and consistent story visualization. Examples of more natural stories with references for the FlintstonesSV [16] and MUGEN [19] datasets (bottom text) compared to more typical but less natural story text (top). We extend the MUGEN dataset by introducing additional two characters (e.g. Lisa and Jhon) and four backgrounds (e.g. Sand, Grass, Stone and Dirt). + +# Abstract + +There has been a recent explosion of impressive generative models that can produce high quality images (or videos) conditioned on text descriptions. However, all such approaches rely on conditional sentences that contain unambiguous descriptions of scenes and main actors in them. Therefore employing such models for more complex task of story visualization, where naturally references and co-references exist, and one requires to reason about when to maintain consistency of actors and backgrounds across frames/scenes, and when not to, based on story progression, remains a challenge. In this work, we address the aforementioned challenges and propose a novel autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective reference resolution and learns to maintain scene and actor consistency when needed. To validate the effectiveness of our approach, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multisentence storylines. Our experiments for story generation on the MUGEN, the PororoSV [30] and the FlintstonesSV [16] dataset show that our method not only outperforms prior state-of-the-art in generating frames with high visual quality, which are consistent with the story, but also models appropriate correspondences between the characters and the background. + +# 1. Introduction + +Multimodal deep learning approaches have pushed the quality and the breadth of conditional generation tasks such as image captioning [23, 33, 37, 54, 57] and text-to-image synthesis [25, 44, 59, 61-63]. Owing to the technical leaps made in generative models, such as generative adversarial networks (GANs) [14], variational autoencoders (VAEs) [27] and the more recent diffusion models [21], approaches for text-to-image synthesis can now generate images with high visual fidelity representative of the textual descriptions. The captions, however, in such cases, are generally short self-contained sentences representing the high-level semantics of a scene. This is rather restrictive in the real-world applications [8, 30, 31] where fine-grained understanding of object interactions, motion and background information described by multiple sentences becomes necessary. One such task is that of story generation or visualization – the goal of which is to generate a sequence of illustrative image frames with coherent semantics given a sequence of sentences [30, 35, 36, 60]. + +Characteristic features of a good visual story is high visual quality over multiple frames; this includes rendering of discernible objects, actors, poses and realistic interactions of those actors with objects and within the scene. Moreover, for text-based story generation it is crucial to maintain consistency between the generated frames and the multisentence descriptions. Not only the actor context, but also the background of the generated story should be in-line with the description demonstrating effortless transition and adaptation to the changing environments within the story [34]. + +Recent advances on the task of story generation have made significant advances along these lines, showing high visual fidelity and character consistency for story sentences that are self-contained and unambiguous (explicitly mentioning characters and the setting each time). While impressive, this setup is fundamentally unrealistic. Realistic story text is considerably more complex and referential in nature; requiring ability to resolve ambiguity and references (or co-references) through reasoning. As shown in Fig. 1, while the description corresponding to the first frame has an explicit reference to the character names, the (typical) subsequent frame descriptions, provided by human, contain references such as "she, he, they". Moreover, while maintaining character consistency, current approaches are limited in preserving, or transitioning through, the background information in agreement with the text (cf. Fig. 3) [30, 35, 36]. + +In natural language processing (NLP) co-reference resolution in text is an important and core task [3, 26, 38]. While it may be possible to apply such methods to story text to first resolve ambiguous references and then generate corresponding images using existing story generation approaches, this is sub-optimal. The reason, is that co-reference resolution in the text domain, at best, would only allow to resolve references and maintain consistency across identity of the character. Appearance across frames would still lack consistency and require some form of visual reasoning. As also noted in [47], reference resolution in the visual domain, or visio-lingual domain, is more powerful. + +In this work, for the first time (to our knowledge), we study co-reference resolution in story generation. Prior work [35] offers limited performance when faced with text containing references (see Sec. 5). We address this by proposing a new autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective visio-lingual co-reference resolution and learns to maintain scene and actor consistency when needed. Further, given the lack of datasets that contain references and more complex sentence structure, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multi-sentence storylines. + +Contributions. Our contributions are three-fold: (i) First, we introduce a novel autoregressive deep generative framework, Story-LDM, that adopts and extends latent diffusion models for the task of story generation. As part of Story-LDM, we propose a meticulously designed memory-attention mechanism capable of encoding and leveraging contextual relevance between the part of the story-line that has already been generated, and the current frame being generated based on learned semantic similarity of corresponding sentences. Equipped with this, our sequential diffusion model can generate consistent stories by resolving and + +then capturing temporal character and background context. (ii) Second, to validate our approach for co-reference resolution, and character and background consistency in the visual domain, we extend existing datasets to include more complex scenarios and, importantly, referential text. Specifically, we extend the MUGEN dataset [19] to include multiple characters and diverse backgrounds. We also modify FlintstonesSV [16] and PororoSV [30] dataset to include character references. These enhancements allow us to increase the complexity of the aforementioned datasets by introducing co-references in the sentences of a story. (iii) Finally, to evaluate different approaches for foreground (character) as well as background consistency we propose novel evaluation metrics. Our results on the MUGEN [19], the PororoSV [30] and the FlintstonesSV [16] datasets show that we outperform the prior state-of-the-art on consistency metrics by a large margin. + +# 2. Related work + +Text-to-image synthesis. Deep generative models, particularly, generative adversarial networks (GANs) [14], variational autoencoders (VAEs) [27] and normalizing flows [4, 11, 12] have been applied to multimodal tasks at the intersection of vision and language. Typical such tasks include image captioning [1, 33, 56] and text-to-image synthesis [32, 44, 51, 59, 61, 63]. Early work on text conditioned image synthesis built upon the success of GANs [44]. More recent approaches have utilized multi-stage generators [61] and normalizing flow-based priors [32] in the latent space to model the distribution of images given text. Various approaches have found cross-domain contrastive loss to improve text-to-image generation models [25, 62]. DALLE [42] and Cogview [10] harness the power of transformers [53] and discrete variational autoencoders (VQ-VAE) [43] yielding very high quality image samples. + +More recent are the advances in diffusion models which have revolutionized the domain of image generation [21]. Diffusion models progressively add noise to the data and learn a reverse diffusion process to reconstruct it. Nichol et al. [39] adapted diffusion models for text-to-image generation and explore CLIP [40] guided generation as well as classifier-free modeling. Standard diffusion models are employed directly in the high-dimensional pixel space and therefore, cannot directly be used for the more complex task of story generation. Recent work [6, 15, 45] instead use encodings from pre-trained models as input to the diffusion models, thereby reducing the complexity of the task by working in a lower-dimensional space. In this work, we build upon this idea and extend it for sequential story generation. + +Text-to-video synthesis. One of the challenges of text-to-video synthesis is the smoothness of motion in a video [30]. + +![](images/26839c372a8a93965549af37e50e979d1c1308ce2dd3208fa1d01d2a70c4ef9b.jpg) +(a) Story-LDM architecture for conditional story generation. + +![](images/55e6f0d2f9899dc3bb29a1e807fe335165c0d995c1387eb338d5f3b0482f31e8.jpg) +(b) Memory state with query and values for our memory-attention module. +Figure 2. Story-Latent Diffusion Models for consistent story generation. (a) The autoregressive conditional generative Story-LDM model with the proposed memory-attention module. (b) A snapshot of the memory for $k$ frames. + +Early work focused on generating short clips [8, 31]. To effectively learn the motion, various approaches disentangle the motion features from the background information [18,52,55]. Wu et al. [58] propose a novel two-dimensional VQ-VAE and sparse attention module for real-world text-to-video generation. Singer et al. [48] decompose the temporal U-Net [46] and the attention modules to approximate them in space and time to extend the text-to-image diffusion models to model text-to-video generation. Ligong et al. [17] propose a transformer framework to jointly model various modalities. + +Story Generation. Li et al. [30] proposed the initial idea and task of story generation. A two-level StoryGAN framework is applied to ensure image-level consistency between each sentence and image pair, and a global discriminator enforces global consistency between the entire image sequence and the story. Various approaches have proposed improvements to the StoryGAN architecture. Zeng et al. [60] introduce sentence-level alignment and word-based attention to improve relevance. Li et al. [29] further improve the performance with enhanced discriminators and dilated convolutions. In [49] foreground-background information is provided as additional supervision and [34] use video captioning for semantic alignment between text and frames. Recently, Chen et al. [5] adopted visual planning and character token alignment to improve character consistency. + +Story Completion. Recently, another task for text-to-story synthesis referred to as story completion has been proposed [36]. In this task, in addition to sentences, the first frame of the story is provided as input. In effect, story completion is a simplified variant of story generation. StoryDALL-E [36] leverages models pre-trained for text-to-image synthesis to perform story completion. Datasets for this task include CLEVR-SV [30] and Pororo-SV [16] which are derived from the CLEVR dataset [22], and the Flintstones dataset for text-to-video synthesis has also been modified for the task of story visualization [34]. Additionally, to evaluate the generalization performance, popular DiDeMo dataset for video captioning [2] is adapted for the task in [36]. + +Reference Resolution. Co-reference resolution is an important and well-researched topic in NLP and focuses on resolving the pronouns and their associated entities. Classic methods in NLP to co-reference resolution employ decision trees [3,38], maximum-entropy modeling [26], cluster-ranking [41] and classification algorithms [50]. More recent approaches [13,24,28] leverage neural network architectures to obtain improved performance with Transformers [9]. Seo et al. [47] proposed visual co-reference resolution for the task of Visual Question-Answering (VQA) dialogs. We take inspiration from [47], but propose a much more sophisticated memory-attention module that allows us to perform visio-lingual co-reference resolution (and visual consistency modeling) for visual story generation. + +# 3. Approach + +To generate temporally consistent stories based solely on the linguistic story-line, we develop a deep generative approach with autoregressive structure. We build upon the success of diffusion models in modeling the underlying data distribution of images to produce high quality samples, and learn the generative conditional distribution of the visual story based on the textual descriptions. Given that the multi-frame stories involve high-dimensional data input, we employ Latent Diffusion Models [45], such that diffusion models can be applied in a computationally efficient manner. Besides, to ensure temporal consistency and smooth story progression, we propose a novel memory attention mechanism which not only attends to the multimodal representations of the current frame but also takes into account the already generated semantics of the previous frames. This module also allows us to resolve ambiguous references (e.g., he/she, they, etc.) using visual memory and is the core of our technical contribution. We first provide an overview of the Diffusion Models and the Latent Diffusion Models, following which we present our autoregressive latent diffusion model for stories called Story-LDM1. + +# 3.1. The Latent Diffusion Model Backbone + +Diffusion Models. Diffusion models are a class of generative models that approximate the underlying data distribution $p(\mathbf{x})$ by denoising a base (Gaussian) distribution in multiple steps using a reverse process of a fixed Markov Chain of length $T$ . To estimate $p(\mathbf{x})$ , the forward diffusion process starts from the input data $\mathbf{x}_0 = \mathbf{x}$ and gradually adds noise to obtain a set of noisy samples $\mathbf{x}_1, \ldots, \mathbf{x}_T$ such that $\mathbf{x}_T \sim \mathcal{N}(0, 1)$ represents a sample from a Gaussian distribution. Under the Markov assumption, the probability of the forward process modeling the distribution $q(\mathbf{x}_{0:T} \mid \mathbf{x}_0)$ and the reverse diffusion process estimating probability at an earlier time-step are formulated as: + +$$ +q \left(\mathbf {x} _ {1: T} \mid \mathbf {x} _ {0}\right) := \prod_ {i = 1} ^ {T} \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \boldsymbol {I}\right) \tag {1} +$$ + +$$ +p _ {\theta} \left(\mathbf {x} _ {0: T}\right) = p _ {\theta} \left(\mathbf {x} _ {T}\right) \prod_ {i = 1} ^ {T} p \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right). +$$ + +Here, $\{\beta_i\}_{i=1}^T$ is the variance schedule for each time-step such that $\mathbf{x}_T$ is nearly a Gaussian. The model parameters $\theta$ are learnt with the following objective, + +$$ +\mathcal {L} _ {D M} := \mathbb {E} _ {t, \mathbf {x}, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, t) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $\epsilon \sim \mathcal{N}(0,1)$ and $\epsilon_{\theta}(\mathbf{x}_t,t), t = 1,\dots,T$ is a sequence of denoising autoencoders with noisy input $\mathbf{x}_t$ predicting the noise that was added to the original input $\mathbf{x}$ . + +Despite yielding state-of-the-art results in various image generation tasks, diffusion models directly operating in the high-dimensional pixel are computationally expensive and resource exhaustive. This limits their application to an even higher-dimensional data such as multi-frame stories or video datasets, which is the focus of this work. + +Diffusion Models in the Latent Space. To broaden the applicability of the diffusion models to very high-dimensional data e.g. high-resolution images, Latent Diffusion Models (LDM) [45] first compress the original image to a lower-dimensional space using perceptual image compression. An auto-encoder approach is employed such that the original spatial structure of the input image is preserved in the latent space. That is, the encoder $E(\cdot)$ maps the input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ to a latent representation $\mathbf{Z} \in \mathbb{R}^{h \times w \times c}$ , downsampling the image to a lower spatial dimension. Following this, the diffusion model is applied to the latents $\mathbf{Z}$ , where time-conditioned U-Net $\epsilon_{\theta}(\mathbf{Z}_t, t)$ is employed to model the diffusion process. The objective of the diffusion model from Eq. (2) becomes, + +$$ +\mathcal {L} _ {L D M} := \mathbb {E} _ {t, E (\mathbf {Z}), \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {Z} _ {t}, t) \| _ {2} ^ {2} \right], \tag {3} +$$ + +During training, a forward diffusion process is applied to generate $\mathbf{Z}$ , which are mapped to the original image space using a decoder $D(\cdot)$ . + +
Dataset# Ref (avg.)# Chars# Backgrounds
MUGEN [19]None12
Extended MUGEN336
FlintstonesSV [16]3.587323
Extended FlintstonesSV4.617323
PororoSV [30]1.019None
Extended PororoSV1.169None
+ +Table 1. Dataset statistics of the MUGEN, FlintstonesSV and PororoSV. + +# 3.2. Story-Latent Diffusion Models + +Given a textural story, characterized by sequence of $M$ sentences $\mathbf{S}_{txt} = \{\mathbf{S}^0,\dots ,\mathbf{S}^M\}$ , the goal of story generation is to produce a sequence of corresponding frames $\mathbf{S}_{img} = \{\mathbf{I}^0,\dots ,\mathbf{I}^M\}$ that visualize the story. We note that this is a more difficult problem than one of story continuation [36], where in addition to the textual story $\mathbf{S}_{txt}$ approaches have access to a source frame $\mathbf{I}^0$ for additional context at inference time. During training it is assumed that we have access to paired dataset of $N$ samples $\mathcal{D} = \{\mathbf{S}_{txt}^{(i)},\mathbf{S}_{img}^{(i)}\}_{i = 1}^{N}$ . We extend latent diffusion models to this task, by allowing them to generate multi-frame stories autoregressively, and by introducing rich conditional structure that takes into account current sentence as well as context from earlier generated frames through visual memory module. This visual memory allows the model to incorporate character/background consistency and resolve text references when needed, resulting in improved performance. + +Given a condition $\mathbf{y}$ , LDM utilizes a cross-attention layer with key $(\mathbf{K})$ , query $(\mathbf{Q})$ and value $(\mathbf{V})$ where, + +$$ +\operatorname {A t t e n t i o n} (\mathbf {K}, \mathbf {Q}, \mathbf {V}) = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {T}}{\sqrt {d}}\right). \mathbf {V}. \tag {4} +$$ + +Here, $\mathbf{Q} = \mathbf{W}_Q.\hat{f}(\mathbf{Z})$ , $\mathbf{K} = \mathbf{W}_K.f(\mathbf{y})$ and $\mathbf{V} = \mathbf{W}_V.f(\mathbf{y})$ , and $\mathbf{W}_Q \in \mathbb{R}^{d \times d_q}$ , $\mathbf{W}_K \in \mathbb{R}^{d \times d_k}$ and $\mathbf{W}_V \in \mathbb{R}^{d \times d_v}$ are learnable parameters, $\hat{f}(\mathbf{Z})$ an intermediate flattened feature representation of $\mathbf{Z}$ within the diffusion model and $f(\mathbf{y})$ the feature representation of the condition $\mathbf{y}$ . The objective in Eq. (3) for conditional generation becomes, + +$$ +\mathcal {L} _ {L D M} := \mathbb {E} _ {t, E (\mathbf {I}), \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(\mathbf {Z} _ {t}, f (\mathbf {y}), t\right) \| _ {2} ^ {2} \right]. \tag {5} +$$ + +Note that the denoising autoencoders $\epsilon_{\theta}$ now additionally depend on the condition encoding $f(\mathbf{y})$ + +For a sample $\{\mathbf{S}_{\text {txt }}^{(i)}, \mathbf{S}_{\text {img }}^{(i)}\}$ , we first project all the $M$ input frames of the story, $\mathbf{I}^0, \ldots, \mathbf{I}^M$ onto a low-dimensional space using a frame-encoder $E$ , and obtain the encoded frames $\mathbf{Z}^0, \ldots, \mathbf{Z}^M$ for a single story. To effectively condition on the corresponding frame, we apply this crossattention layer to the intermediate representations of the neural network for each frame $\mathbf{I}^m$ and its textual description $\mathbf{S}^m$ using Eq. (4), where the relevance of the description $\mathbf{S}^m$ + +
DatasetMethodw/ ref. textChar-acc (↑)Char-F1 (↑)BG-acc (↑)BG-F1 (↑)FID (↓)
FlintstonesVLCStoryGAN [34]×27.7342.014.8316.49120.85
LDM [45]×79.8692.3348.0237.8661.40
LDM [45]✓57.3878.6844.1928.2587.39
Story-LDM (Ours)✓69.1986.5935.2128.8069.49
MUGENLDM [45]✓31.3921.2815.7418.66120.99
Story-LDM (Ours)✓93.4095.6092.1992.3762.16
+ +Table 2. Quantitative results. Experimental results on the FlintstoneSV and the MUGEN datasets. + +is weighted by the similarity between the textual representation and the encoded frame representation $\mathbf{Z}^m$ (cf. Fig. 2a). + +For sequential generation, the model in addition to the current state, requires information from all the previous states. To enable this, the diffusion process for any frame representation $\mathbf{Z}^m$ is conditioned on the visual representations of the previous frames $\mathbf{Z}^0,\ldots ,\mathbf{Z}^{m - 1}$ as well as the sentence descriptions $\mathbf{S}^0,\dots ,\mathbf{S}^m$ . This conditioning is realized through a novel Memory-attention module which forms the basis of our autoregressive approach. + +Memory-attention Module. To capture the spatiotemporal interactions across multiple frames and sentences for a story, in our conditional diffusion model, we condition the frame $\mathbf{Z}^m$ not only on the corresponding text $\mathbf{S}^m$ but also on the previous texts $\mathbf{S}^i$ , for $i \in \{0, \dots, m - 1\}$ . This conditioning is applied throughout the $T$ time-steps of the diffusion process for $\mathbf{Z}^m$ . The conditional denoising autoencoder thus models the conditional distribution $p(\mathbf{Z}^m \mid \mathbf{Z}^{3 [45] which contains a cross-attention layer to generate text-to-image based story, without using our proposed autoregressive memory modules as our baselines for MU- + +![](images/775c49c02dcde6ebe29c1b6b38dfff3cd87165434393d11eeeb3cf728fe40565.jpg) +Figure 5. Qualitative Comparison on Story Generation. Comparison on the FlintstonesSV dataset visual story generation. + +![](images/8dee7f301a52ce94e0debbfe835659f17ff1b026f4ec7955df4ec5b864851324.jpg) +Figure 6. Comparison to Story Continuation. Comparison on the FlintstonesSV dataset for story continuation. Prior work uses first frame as additional input to the model; our model does not. + +GEN, PororoSV and FlintstonesSV datasets. The parameters of the diffusion model within the Story-LDM are initialized with the pre-trained LDM [45]. Similarly, for the textual embedding, we use BERT-tokenizer [9] and use the pre-trained text-transformer from LDM. + +Quantitative Results. Table 2 shows quantitative results for consistent story generation on the FlintstoneSV dataset. We compare the performance of our approach (row 4) to the LDM [45] which we train/test with both original (row 2) as well as the co-referenced (row 3) descriptions. Fur + +thermore, we include the results of the state-of-art VLCStoryGAN [34] (row 1) with the original text of the dataset $^{4}$ (i.e. without co-references). We note that VLCStoryGAN was shown to be better than Duco-StoryGAN [35], CPCSV [49] and original StoryGAN [30] (see [34]). + +Based on Table 2 we make three observations: (1) Our LDM baseline is better than VLCStoryGAN on the original reference-free text (cf. Tab. 2, rows 1 & 2). (2) Reference resolution makes the task considerably harder. With the reference text in our modified dataset, we observe a drop in performance in terms of character and background classification scores (cf. Tab. 2, rows 2 & 3). (3) Our model, with memory-attention module, significantly outperforms the baseline (cf. Tab. 2, rows 3 & 4) both in terms of generative image quality and character consistency; and outperforms SoTA of VLCStoryGAN by $\sim 41\%$ percentage points on character accuracy (while performing a more difficult version of the task). Further, our model, that is required to conduct reference resolution, comes close to the LDM trained with original, reference-free, text (cf. Tab. 2, rows 3 & 4), which can be viewed as a sort of an upper bound. + +On the MUGEN dataset, our method outperforms the strong LDM baseline with gains of $\sim 62\%$ on character accuracy and $\sim 76\%$ on the background accuracy, thereby showing the advantages of the memory-attention mechanism for consistent story generation. We note that MUGEN dataset has more references across story scenes. Flintstones while contains more references per story overall, many of those references are within scenes as opposed to across scenes. Meaning that in terms of reference impact on consistency, MUGEN dataset is actually harder. Experimental results on the PororoSV dataset are provided in the Supplemental. + +Qualitative Results. Figure 3 illustrates qualitative results on the MUGEN dataset. Rows 1, 2 & 3 show ground truth, LDM [45] and our Story-LDM approach, respectively. Here, we see that our method is able to maintain consistency in terms of both character and background. Similarly, in Figure 4 we can show the results on FlintstoneSV dataset which further validates the strong performance of our method when generating high-quality, consistent story. Compared to the LDM, our approach is able to adapt to the diverse backgrounds in the story descriptions. + +Additional Results. We compare the qualitative results of our method to both story generation [34] and story continuation [36] in Figs. 5 and 6 respectively. The comparative images are taken directly from respective papers. We note that story continuation Fig. 6 is solving a different (easier) problem and with text that contains no-references. This makes the comparison to our method, which receives fewer inputs, not very meaningful. Nether-the-less, our approach, that + +![](images/46986fe82fab2965014bf4fc3db2b1a79ddc0d1ecb320ddd4020fb48417732a9.jpg) + +![](images/0fa884a88d96beb5fc5fba5cc3ec403c5098351be5c29791dd2e5f409e4d4ab9.jpg) +Figure 7. Story Diversity. Diverse outputs for a single storyline obtained with our Story-LDM. +Figure 8. Branching Storyline. Generating different yet consistent stories by branching the storyline. Frames in the later columns are generated based on earlier ones and corresponding text. + +can resolve references and is solving a harder story generation task, obtains highly competitive results. Furthermore, to show that our autoregressive visual memory module can generate diverse stories conditioned on the current and previous condition, we create different story-lines starting for a single sentence. In Fig. 8, we can see for reference 'they', the model can generate both the characters according to the storyline already parsed. Moreover, in Fig. 7 we show that our approach can not only generate consistent visual stories, but also diverse frames for the same text (cf. Fig. 7). Additional results are provided in the Supplemental. + +# 6. Conclusion + +In this paper, we formulate consistent story generation in a more realistic way by co-referencing actors/backgrounds in the story descriptions. We develop an autoregressive Story-LDM approach with memory attention capable of maintaining consistency across the frames based on the previously generated frames and their corresponding descriptions. We introduced modified datasets to evaluate the performance for reference resolution. We expect our proposed formulation and models to be conducive to the real-world use cases and further the research. + +# References + +[1] Jyoti Aneja, Harsh Agrawal, Dhruv Batra, and Alexander G. Schwing. Sequential latent spaces for modeling the intention during diverse image captioning. In ICCV, 2019. 2 +[2] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 3 +[3] Chinatsu Aone and Scott William. Evaluating automated and manual acquisition of anaphora resolution strategies. In Annual Meeting of the Association for Computational Linguistics, pages 122-129, 1995. 2, 3 +[4] Apratim Bhattacharyya, Shweta Mahajan, Mario Fritz, Bernt Schiele, and Stefan Roth. Normalizing flows with multiscale autoregressive priors. In CVPR, 2020. 2 +[5] Hong Chen, Rujun Han, Te-Lin Wu, Hideki Nakayama, and Nanyun Peng. Character-centric story visualization via visual planning and token alignment. In EMNLP, 2022. 3 +[6] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. CVPR, 2023. 2 +[7] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019. 7 +[8] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018. 1, 3 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2018. 3, 7 +[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurIPS, 2021. 2 +[11] Laurent Dinh, David Krueger, and Yoshua Bengio. NICE: Non-linear independent components estimation. In ICLR Workshop, 2015. 2 +[12] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real NVP. In ICLR, 2017. 2 +[13] Hongliang Fei, Xu Li, Dingcheng Li, and Ping Li. End-to-end deep reinforcement learning based coreference resolution. In Annual Meeting of the Association for Computational Linguistics, pages 660-665, 2019. 3 +[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2 +[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2 +[16] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 1, 2, 3, 4, 7 + +[17] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, 2022. 3 +[18] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In CVPR, 2018. 3 +[19] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Isabelle Hu, and Devi Parikh. Mugen: A playground for video-audio-text multimodal understanding and generation. arXiv preprint arXiv:2204.08058, 2022. 1, 2, 4, 7 +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 2017. 7 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 2 +[22] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3 +[23] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In CVPR, 2016. 1 +[24] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pretraining by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64-77, 2020. 3 +[25] Minguk Kang and Jaesik Park. Contragan: Contrastive learning for conditional image generation. NeurIPS, 2020. 1, 2 +[26] Andrew Kehler. Probabilistic coreference in information extraction. 1997. 2, 3 +[27] Diederik P. Kingma and Max Welling. Auto-encoding variational Bayes. In ICLR, 2014. 1, 2 +[28] Hideo Kobayashi, Yufang Hou, and Vincent Ng. End-to-end neural bridging resolution. In International Conference on Computational Linguistics, 2022. 3 +[29] Chunye Li, Liya Kong, and Zhiping Zhou. Improved-storygan for sequential images visualization. Journal of Visual Communication and Image Representation, 2020. 3 +[30] Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. Storygan: A sequential conditional gan for story visualization. In CVPR, 2019. 1, 2, 3, 4, 7, 8 +[31] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 1, 3 +[32] Shweta Mahajan, Iryna Gurevych, and Stefan Roth. Latent normalizing flows for many-to-many cross-domain mappings. In *ICLR*, 2020. 2 +[33] Shweta Mahajan and Stefan Roth. Diverse image captioning with context-object split latent spaces. NeurIPS, 2020. 1, 2 +[34] Adyasha Maharana and Mohit Bansal. Integrating visuospatial, linguistic and commonsense structure into story visualization. arXiv preprint arXiv:2110.10834, 2021. 1, 3, 5, 7, 8 + +[35] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Improving generation and evaluation of visual stories via semantic consistency. 2021. 1, 2, 8 +[36] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. In ECCV, 2022. 1, 2, 3, 4, 8 +[37] Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L. Yuille. Deep captioning with multimodal recurrent neural networks (m-RNN). In ICLR, 2015. 1 +[38] Joseph F McCarthy and Wendy G Lehnert. Using decision trees for coreference resolution. 1995. 2, 3 +[39] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML, 2022. 2 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 +[41] Altaf Rahman and Vincent Ng. Narrowing the modeling gap: A cluster-ranking approach to coreference resolution. Journal of Artificial Intelligence Research, 2011. 3 +[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML. PMLR, 2021. 2 +[43] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2 +[44] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 1, 2 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8 +[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351 of Lecture Notes in Computer Science, pages 234-241. Springer, 2015. 3 +[47] Paul Hongsuck Seo, Andreas Lehrmann, Bohyung Han, and Leonid Sigal. Visual reference resolution using attention memory for visual dialog. NIPS, 2017. 2, 3 +[48] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3 + +[49] Yun-Zhu Song, Zhi Rui Tam, Hung-Jen Chen, Huiao-Han Lu, and Hong-Han Shuai. Character-preserving coherent story visualization. In ECCV, 2020. 3, 8 +[50] Wee Meng Soon, Hwee Tou Ng, and Daniel Chung Yong Lim. A machine learning approach to coreference resolution of noun phrases. Computational linguistics, 2001. 3 +[51] Hung-Yu Tseng, Hsin-Ying Lee, Lu Jiang, Ming-Hsuan Yang, and Weilong Yang. Retrievegan: Image synthesis via differentiable patch retrieval. In ECCV, 2020. 2 +[52] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, 2018. 3 +[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2 +[54] Ashwin K. Vijayakumar, Michael Cogswell, Ramprasaath R. Selvaraju, Qing Sun, Stefan Lee, David J. Crandall, and Dhruv Batra. Diverse beam search for improved description of complex scenes. In AAAI, 2018. 1 +[55] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NIPS, 2016. 3 +[56] Liwei Wang, Alexander G. Schwing, and Svetlana Lazebnik. Diverse and accurate image description using a variational auto-encoder with an additive Gaussian encoding space. In NIPS, 2017. 2 +[57] Wei Wang, Xavier Alameda-Pineda, Dan Xu, Pascal Fua, Elisa Ricci, and Nicu Sebe. Every smile is unique: Landmark-guided diverse smile generation. In CVPR, 2018. 1 +[58] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 3 +[59] Guojun Yin, Bin Liu, Lu Sheng, Nenghai Yu, Xiaogang Wang, and Jing Shao. Semantics disentangling for text-to-image generation. CVPR, 2019. 1, 2 +[60] Gangyan Zeng, Zhaohui Li, and Yuan Zhang. Pororogan: an improved story visualization model on pororo-sv dataset. In Proceedings of the 3rd International Conference on Computer Science and Artificial Intelligence, 2019. 1, 3 +[61] Han Zhang, Ian J. Goodfellow, Dimitris N. Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, 2019. 1, 2 +[62] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In CVPR, pages 833-842, 2021. 1, 2 +[63] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmagan: Dynamic memory generative adversarial networks for text-to-image synthesis. In CVPR, 2019. 1, 2 \ No newline at end of file diff --git a/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/images.zip b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d9d5b30410c34afe32b56924c0a17685d11c7c6e --- /dev/null +++ b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:726b2428fc864a10d825eafea9af28acd8144d3469ccf23eb49f14b1c2139fca +size 755161 diff --git a/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/layout.json b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..92dbde744c55c42228ef4a1718cd17bce528123b --- /dev/null +++ b/2023/Make-a-Story_ Visual Memory Conditioned Consistent Story Generation/layout.json @@ -0,0 +1,8732 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 75, + 523, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 75, + 523, + 92 + ], + "spans": [ + { + "bbox": [ + 70, + 75, + 523, + 92 + ], + "type": "text", + "content": "Make-A-Story: Visual Memory Conditioned Consistent Story Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 114, + 198, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 114, + 198, + 128 + ], + "spans": [ + { + "bbox": [ + 107, + 114, + 198, + 128 + ], + "type": "text", + "content": "Tanzila Rahman" + }, + { + "bbox": [ + 107, + 114, + 198, + 128 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 115, + 301, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 115, + 301, + 129 + ], + "spans": [ + { + "bbox": [ + 223, + 115, + 301, + 129 + ], + "type": "text", + "content": "Hsin-Ying Lee" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 325, + 115, + 373, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 115, + 373, + 128 + ], + "spans": [ + { + "bbox": [ + 325, + 115, + 373, + 128 + ], + "type": "text", + "content": "Jian Ren2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 399, + 115, + 486, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 115, + 486, + 129 + ], + "spans": [ + { + "bbox": [ + 399, + 115, + 486, + 129 + ], + "type": "text", + "content": "Sergey Tulyakov2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 197, + 129, + 290, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 129, + 290, + 143 + ], + "spans": [ + { + "bbox": [ + 197, + 129, + 290, + 143 + ], + "type": "text", + "content": "Shweta Mahajan" + }, + { + "bbox": [ + 197, + 129, + 290, + 143 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 129, + 396, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 396, + 143 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 396, + 143 + ], + "type": "text", + "content": "Leonid Sigal" + }, + { + "bbox": [ + 315, + 129, + 396, + 143 + ], + "type": "inline_equation", + "content": "^{1,3,4}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 181, + 143, + 412, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 143, + 412, + 157 + ], + "spans": [ + { + "bbox": [ + 181, + 143, + 412, + 157 + ], + "type": "text", + "content": "1University of British Columbia 2Snap Inc." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 166, + 157, + 429, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 157, + 429, + 170 + ], + "spans": [ + { + "bbox": [ + 166, + 157, + 429, + 170 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 166, + 157, + 429, + 170 + ], + "type": "text", + "content": "Vector Institute for AI " + }, + { + "bbox": [ + 166, + 157, + 429, + 170 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 166, + 157, + 429, + 170 + ], + "type": "text", + "content": "Canada CIFAR AI Chair" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 66, + 185, + 526, + 300 + ], + "blocks": [ + { + "bbox": [ + 66, + 185, + 526, + 300 + ], + "lines": [ + { + "bbox": [ + 66, + 185, + 526, + 300 + ], + "spans": [ + { + "bbox": [ + 66, + 185, + 526, + 300 + ], + "type": "image", + "image_path": "99a51cea6c9dd088718dfa8e40386ec5f015d2766da0791222716c2f3bee1a1a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 302, + 546, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 302, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 546, + 335 + ], + "type": "text", + "content": "Figure 1. Referential and consistent story visualization. Examples of more natural stories with references for the FlintstonesSV [16] and MUGEN [19] datasets (bottom text) compared to more typical but less natural story text (top). We extend the MUGEN dataset by introducing additional two characters (e.g. Lisa and Jhon) and four backgrounds (e.g. Sand, Grass, Stone and Dirt)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 346, + 191, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 346, + 191, + 358 + ], + "spans": [ + { + "bbox": [ + 143, + 346, + 191, + 358 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 376, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 700 + ], + "type": "text", + "content": "There has been a recent explosion of impressive generative models that can produce high quality images (or videos) conditioned on text descriptions. However, all such approaches rely on conditional sentences that contain unambiguous descriptions of scenes and main actors in them. Therefore employing such models for more complex task of story visualization, where naturally references and co-references exist, and one requires to reason about when to maintain consistency of actors and backgrounds across frames/scenes, and when not to, based on story progression, remains a challenge. In this work, we address the aforementioned challenges and propose a novel autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective reference resolution and learns to maintain scene and actor consistency when needed. To validate the effectiveness of our approach, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multisentence storylines. Our experiments for story generation on the MUGEN, the PororoSV [30] and the FlintstonesSV [16] dataset show that our method not only outperforms prior state-of-the-art in generating frames with high visual quality, which are consistent with the story, but also models appropriate correspondences between the characters and the background." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 346, + 386, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 346, + 386, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 346, + 386, + 358 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 365, + 546, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 592 + ], + "type": "text", + "content": "Multimodal deep learning approaches have pushed the quality and the breadth of conditional generation tasks such as image captioning [23, 33, 37, 54, 57] and text-to-image synthesis [25, 44, 59, 61-63]. Owing to the technical leaps made in generative models, such as generative adversarial networks (GANs) [14], variational autoencoders (VAEs) [27] and the more recent diffusion models [21], approaches for text-to-image synthesis can now generate images with high visual fidelity representative of the textual descriptions. The captions, however, in such cases, are generally short self-contained sentences representing the high-level semantics of a scene. This is rather restrictive in the real-world applications [8, 30, 31] where fine-grained understanding of object interactions, motion and background information described by multiple sentences becomes necessary. One such task is that of story generation or visualization – the goal of which is to generate a sequence of illustrative image frames with coherent semantics given a sequence of sentences [30, 35, 36, 60]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Characteristic features of a good visual story is high visual quality over multiple frames; this includes rendering of discernible objects, actors, poses and realistic interactions of those actors with objects and within the scene. Moreover, for text-based story generation it is crucial to maintain consistency between the generated frames and the multisentence descriptions. Not only the actor context, but also the background of the generated story should be in-line with the description demonstrating effortless transition and adaptation to the changing environments within the story [34]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2493" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "type": "text", + "content": "Recent advances on the task of story generation have made significant advances along these lines, showing high visual fidelity and character consistency for story sentences that are self-contained and unambiguous (explicitly mentioning characters and the setting each time). While impressive, this setup is fundamentally unrealistic. Realistic story text is considerably more complex and referential in nature; requiring ability to resolve ambiguity and references (or co-references) through reasoning. As shown in Fig. 1, while the description corresponding to the first frame has an explicit reference to the character names, the (typical) subsequent frame descriptions, provided by human, contain references such as \"she, he, they\". Moreover, while maintaining character consistency, current approaches are limited in preserving, or transitioning through, the background information in agreement with the text (cf. Fig. 3) [30, 35, 36]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 264, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 289, + 407 + ], + "type": "text", + "content": "In natural language processing (NLP) co-reference resolution in text is an important and core task [3, 26, 38]. While it may be possible to apply such methods to story text to first resolve ambiguous references and then generate corresponding images using existing story generation approaches, this is sub-optimal. The reason, is that co-reference resolution in the text domain, at best, would only allow to resolve references and maintain consistency across identity of the character. Appearance across frames would still lack consistency and require some form of visual reasoning. As also noted in [47], reference resolution in the visual domain, or visio-lingual domain, is more powerful." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 407, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 289, + 575 + ], + "type": "text", + "content": "In this work, for the first time (to our knowledge), we study co-reference resolution in story generation. Prior work [35] offers limited performance when faced with text containing references (see Sec. 5). We address this by proposing a new autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective visio-lingual co-reference resolution and learns to maintain scene and actor consistency when needed. Further, given the lack of datasets that contain references and more complex sentence structure, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multi-sentence storylines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 289, + 715 + ], + "type": "text", + "content": "Contributions. Our contributions are three-fold: (i) First, we introduce a novel autoregressive deep generative framework, Story-LDM, that adopts and extends latent diffusion models for the task of story generation. As part of Story-LDM, we propose a meticulously designed memory-attention mechanism capable of encoding and leveraging contextual relevance between the part of the story-line that has already been generated, and the current frame being generated based on learned semantic similarity of corresponding sentences. Equipped with this, our sequential diffusion model can generate consistent stories by resolving and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": "then capturing temporal character and background context. (ii) Second, to validate our approach for co-reference resolution, and character and background consistency in the visual domain, we extend existing datasets to include more complex scenarios and, importantly, referential text. Specifically, we extend the MUGEN dataset [19] to include multiple characters and diverse backgrounds. We also modify FlintstonesSV [16] and PororoSV [30] dataset to include character references. These enhancements allow us to increase the complexity of the aforementioned datasets by introducing co-references in the sentences of a story. (iii) Finally, to evaluate different approaches for foreground (character) as well as background consistency we propose novel evaluation metrics. Our results on the MUGEN [19], the PororoSV [30] and the FlintstonesSV [16] datasets show that we outperform the prior state-of-the-art on consistency metrics by a large margin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 289, + 391, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 391, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 391, + 301 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 310, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 501 + ], + "type": "text", + "content": "Text-to-image synthesis. Deep generative models, particularly, generative adversarial networks (GANs) [14], variational autoencoders (VAEs) [27] and normalizing flows [4, 11, 12] have been applied to multimodal tasks at the intersection of vision and language. Typical such tasks include image captioning [1, 33, 56] and text-to-image synthesis [32, 44, 51, 59, 61, 63]. Early work on text conditioned image synthesis built upon the success of GANs [44]. More recent approaches have utilized multi-stage generators [61] and normalizing flow-based priors [32] in the latent space to model the distribution of images given text. Various approaches have found cross-domain contrastive loss to improve text-to-image generation models [25, 62]. DALLE [42] and Cogview [10] harness the power of transformers [53] and discrete variational autoencoders (VQ-VAE) [43] yielding very high quality image samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 502, + 548, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 548, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 548, + 681 + ], + "type": "text", + "content": "More recent are the advances in diffusion models which have revolutionized the domain of image generation [21]. Diffusion models progressively add noise to the data and learn a reverse diffusion process to reconstruct it. Nichol et al. [39] adapted diffusion models for text-to-image generation and explore CLIP [40] guided generation as well as classifier-free modeling. Standard diffusion models are employed directly in the high-dimensional pixel space and therefore, cannot directly be used for the more complex task of story generation. Recent work [6, 15, 45] instead use encodings from pre-trained models as input to the diffusion models, thereby reducing the complexity of the task by working in a lower-dimensional space. In this work, we build upon this idea and extend it for sequential story generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "Text-to-video synthesis. One of the challenges of text-to-video synthesis is the smoothness of motion in a video [30]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "type": "text", + "content": "2494" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 321, + 180 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 321, + 180 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 321, + 180 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 321, + 180 + ], + "type": "image", + "image_path": "26839c372a8a93965549af37e50e979d1c1308ce2dd3208fa1d01d2a70c4ef9b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 190, + 282, + 200 + ], + "lines": [ + { + "bbox": [ + 87, + 190, + 282, + 200 + ], + "spans": [ + { + "bbox": [ + 87, + 190, + 282, + 200 + ], + "type": "text", + "content": "(a) Story-LDM architecture for conditional story generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 369, + 78, + 541, + 177 + ], + "blocks": [ + { + "bbox": [ + 369, + 78, + 541, + 177 + ], + "lines": [ + { + "bbox": [ + 369, + 78, + 541, + 177 + ], + "spans": [ + { + "bbox": [ + 369, + 78, + 541, + 177 + ], + "type": "image", + "image_path": "55e6f0d2f9899dc3bb29a1e807fe335165c0d995c1387eb338d5f3b0482f31e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 180, + 543, + 199 + ], + "lines": [ + { + "bbox": [ + 367, + 180, + 543, + 199 + ], + "spans": [ + { + "bbox": [ + 367, + 180, + 543, + 199 + ], + "type": "text", + "content": "(b) Memory state with query and values for our memory-attention module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 203, + 546, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 546, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 546, + 226 + ], + "type": "text", + "content": "Figure 2. Story-Latent Diffusion Models for consistent story generation. (a) The autoregressive conditional generative Story-LDM model with the proposed memory-attention module. (b) A snapshot of the memory for " + }, + { + "bbox": [ + 46, + 203, + 546, + 226 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 203, + 546, + 226 + ], + "type": "text", + "content": " frames." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 237, + 288, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 368 + ], + "type": "text", + "content": "Early work focused on generating short clips [8, 31]. To effectively learn the motion, various approaches disentangle the motion features from the background information [18,52,55]. Wu et al. [58] propose a novel two-dimensional VQ-VAE and sparse attention module for real-world text-to-video generation. Singer et al. [48] decompose the temporal U-Net [46] and the attention modules to approximate them in space and time to extend the text-to-image diffusion models to model text-to-video generation. Ligong et al. [17] propose a transformer framework to jointly model various modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 371, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 551 + ], + "type": "text", + "content": "Story Generation. Li et al. [30] proposed the initial idea and task of story generation. A two-level StoryGAN framework is applied to ensure image-level consistency between each sentence and image pair, and a global discriminator enforces global consistency between the entire image sequence and the story. Various approaches have proposed improvements to the StoryGAN architecture. Zeng et al. [60] introduce sentence-level alignment and word-based attention to improve relevance. Li et al. [29] further improve the performance with enhanced discriminators and dilated convolutions. In [49] foreground-background information is provided as additional supervision and [34] use video captioning for semantic alignment between text and frames. Recently, Chen et al. [5] adopted visual planning and character token alignment to improve character consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Story Completion. Recently, another task for text-to-story synthesis referred to as story completion has been proposed [36]. In this task, in addition to sentences, the first frame of the story is provided as input. In effect, story completion is a simplified variant of story generation. StoryDALL-E [36] leverages models pre-trained for text-to-image synthesis to perform story completion. Datasets for this task include CLEVR-SV [30] and Pororo-SV [16] which are derived from the CLEVR dataset [22], and the Flintstones dataset for text-to-video synthesis has also been modified for the task of story visualization [34]. Additionally, to evaluate the generalization performance, popular DiDeMo dataset for video captioning [2] is adapted for the task in [36]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 237, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 406 + ], + "type": "text", + "content": "Reference Resolution. Co-reference resolution is an important and well-researched topic in NLP and focuses on resolving the pronouns and their associated entities. Classic methods in NLP to co-reference resolution employ decision trees [3,38], maximum-entropy modeling [26], cluster-ranking [41] and classification algorithms [50]. More recent approaches [13,24,28] leverage neural network architectures to obtain improved performance with Transformers [9]. Seo et al. [47] proposed visual co-reference resolution for the task of Visual Question-Answering (VQA) dialogs. We take inspiration from [47], but propose a much more sophisticated memory-attention module that allows us to perform visio-lingual co-reference resolution (and visual consistency modeling) for visual story generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 418, + 372, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 372, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 372, + 433 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 440, + 547, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 547, + 691 + ], + "type": "text", + "content": "To generate temporally consistent stories based solely on the linguistic story-line, we develop a deep generative approach with autoregressive structure. We build upon the success of diffusion models in modeling the underlying data distribution of images to produce high quality samples, and learn the generative conditional distribution of the visual story based on the textual descriptions. Given that the multi-frame stories involve high-dimensional data input, we employ Latent Diffusion Models [45], such that diffusion models can be applied in a computationally efficient manner. Besides, to ensure temporal consistency and smooth story progression, we propose a novel memory attention mechanism which not only attends to the multimodal representations of the current frame but also takes into account the already generated semantics of the previous frames. This module also allows us to resolve ambiguous references (e.g., he/she, they, etc.) using visual memory and is the core of our technical contribution. We first provide an overview of the Diffusion Models and the Latent Diffusion Models, following which we present our autoregressive latent diffusion model for stories called Story-LDM1." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 703, + 526, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 526, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 526, + 713 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2495" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 250, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 250, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 250, + 83 + ], + "type": "text", + "content": "3.1. The Latent Diffusion Model Backbone" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models are a class of generative models that approximate the underlying data distribution " + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "content": " by denoising a base (Gaussian) distribution in multiple steps using a reverse process of a fixed Markov Chain of length " + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "content": ". To estimate " + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "content": ", the forward diffusion process starts from the input data " + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 = \\mathbf{x}" + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "content": " and gradually adds noise to obtain a set of noisy samples " + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1, \\ldots, \\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(0, 1)" + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "content": " represents a sample from a Gaussian distribution. Under the Markov assumption, the probability of the forward process modeling the distribution " + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{0:T} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 89, + 289, + 234 + ], + "type": "text", + "content": " and the reverse diffusion process estimating probability at an earlier time-step are formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 239, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 239, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 70, + 239, + 287, + 277 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {i = 1} ^ {T} \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right) \\tag {1}", + "image_path": "0657637accc7d799add65fcc4f8a8ecf8df912f81b4007bc70e1d53871bf25e3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 274, + 237, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 274, + 237, + 306 + ], + "spans": [ + { + "bbox": [ + 85, + 274, + 237, + 306 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {T}\\right) \\prod_ {i = 1} ^ {T} p \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right).", + "image_path": "ddf005f9947a75c01ddd436c4993d57a291f84ee2f194c423653369d8cdf7c75.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\{\\beta_i\\}_{i=1}^T" + }, + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "type": "text", + "content": " is the variance schedule for each time-step such that " + }, + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "type": "text", + "content": " is nearly a Gaussian. The model parameters " + }, + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 311, + 287, + 348 + ], + "type": "text", + "content": " are learnt with the following objective," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 352, + 287, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 352, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 95, + 352, + 287, + 368 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D M} := \\mathbb {E} _ {t, \\mathbf {x}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "a2ebd9c31cd5bbf0282ac7bbeedd860f04ed6d87c15c3f9ea0a0fef07135ca2e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t), t = 1,\\dots,T" + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "content": " is a sequence of denoising autoencoders with noisy input " + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "content": " predicting the noise that was added to the original input " + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": "Despite yielding state-of-the-art results in various image generation tasks, diffusion models directly operating in the high-dimensional pixel are computationally expensive and resource exhaustive. This limits their application to an even higher-dimensional data such as multi-frame stories or video datasets, which is the focus of this work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "text", + "content": "Diffusion Models in the Latent Space. To broaden the applicability of the diffusion models to very high-dimensional data e.g. high-resolution images, Latent Diffusion Models (LDM) [45] first compress the original image to a lower-dimensional space using perceptual image compression. An auto-encoder approach is employed such that the original spatial structure of the input image is preserved in the latent space. That is, the encoder " + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "text", + "content": " maps the input image " + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "text", + "content": " to a latent representation " + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "text", + "content": ", downsampling the image to a lower spatial dimension. Following this, the diffusion model is applied to the latents " + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "text", + "content": ", where time-conditioned U-Net " + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{Z}_t, t)" + }, + { + "bbox": [ + 46, + 486, + 287, + 654 + ], + "type": "text", + "content": " is employed to model the diffusion process. The objective of the diffusion model from Eq. (2) becomes," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 658, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 658, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 85, + 658, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L D M} := \\mathbb {E} _ {t, E (\\mathbf {Z}), \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {Z} _ {t}, t) \\| _ {2} ^ {2} \\right], \\tag {3}", + "image_path": "c5a6baae9639a6e8f19629734a38e81cbb7ca81edf0ea8f5d255a0e9e4875312.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "During training, a forward diffusion process is applied to generate " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ", which are mapped to the original image space using a decoder " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 305, + 70, + 545, + 139 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 139 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 139 + ], + "type": "table", + "html": "
Dataset# Ref (avg.)# Chars# Backgrounds
MUGEN [19]None12
Extended MUGEN336
FlintstonesSV [16]3.587323
Extended FlintstonesSV4.617323
PororoSV [30]1.019None
Extended PororoSV1.169None
", + "image_path": "d37ecf9b63b0c25602b241554058ccb10a6bf446ad4fa17cf3809337ac707a32.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 143, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 305, + 143, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 143, + 545, + 163 + ], + "type": "text", + "content": "Table 1. Dataset statistics of the MUGEN, FlintstonesSV and PororoSV." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 171, + 471, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 171, + 471, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 471, + 185 + ], + "type": "text", + "content": "3.2. Story-Latent Diffusion Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "content": "Given a textural story, characterized by sequence of " + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "content": " sentences " + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{txt} = \\{\\mathbf{S}^0,\\dots ,\\mathbf{S}^M\\}" + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "content": ", the goal of story generation is to produce a sequence of corresponding frames " + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{img} = \\{\\mathbf{I}^0,\\dots ,\\mathbf{I}^M\\}" + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "content": " that visualize the story. We note that this is a more difficult problem than one of story continuation [36], where in addition to the textual story " + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{txt}" + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "content": " approaches have access to a source frame " + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^0" + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "content": " for additional context at inference time. During training it is assumed that we have access to paired dataset of " + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "content": " samples " + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{S}_{txt}^{(i)},\\mathbf{S}_{img}^{(i)}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 304, + 190, + 545, + 396 + ], + "type": "text", + "content": ". We extend latent diffusion models to this task, by allowing them to generate multi-frame stories autoregressively, and by introducing rich conditional structure that takes into account current sentence as well as context from earlier generated frames through visual memory module. This visual memory allows the model to incorporate character/background consistency and resolve text references when needed, resulting in improved performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "text", + "content": "Given a condition " + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "text", + "content": ", LDM utilizes a cross-attention layer with key " + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "inline_equation", + "content": "(\\mathbf{K})" + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "text", + "content": ", query " + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q})" + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "inline_equation", + "content": "(\\mathbf{V})" + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "text", + "content": " where," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 331, + 428, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 428, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 331, + 428, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (\\mathbf {K}, \\mathbf {Q}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right). \\mathbf {V}. \\tag {4}", + "image_path": "042bfa380302e288b5812822a809d5ec627d7e62c0718b433108d3f23671e0d3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{W}_Q.\\hat{f}(\\mathbf{Z})" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\mathbf{W}_K.f(\\mathbf{y})" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\mathbf{W}_V.f(\\mathbf{y})" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Q \\in \\mathbb{R}^{d \\times d_q}" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_K \\in \\mathbb{R}^{d \\times d_k}" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_V \\in \\mathbb{R}^{d \\times d_v}" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": " are learnable parameters, " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\hat{f}(\\mathbf{Z})" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": " an intermediate flattened feature representation of " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": " within the diffusion model and " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "f(\\mathbf{y})" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": " the feature representation of the condition " + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 465, + 545, + 540 + ], + "type": "text", + "content": ". The objective in Eq. (3) for conditional generation becomes," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 333, + 548, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 548, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 333, + 548, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L D M} := \\mathbb {E} _ {t, E (\\mathbf {I}), \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {Z} _ {t}, f (\\mathbf {y}), t\\right) \\| _ {2} ^ {2} \\right]. \\tag {5}", + "image_path": "86fdf5ecd8bbf5099d1362cb575756b60c4855f4af497b4b6505561dd2bd4b5e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 571, + 544, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 544, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 544, + 595 + ], + "type": "text", + "content": "Note that the denoising autoencoders " + }, + { + "bbox": [ + 304, + 571, + 544, + 595 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 571, + 544, + 595 + ], + "type": "text", + "content": " now additionally depend on the condition encoding " + }, + { + "bbox": [ + 304, + 571, + 544, + 595 + ], + "type": "inline_equation", + "content": "f(\\mathbf{y})" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "content": "For a sample " + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_{\\text {txt }}^{(i)}, \\mathbf{S}_{\\text {img }}^{(i)}\\}" + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "content": ", we first project all the " + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "content": " input frames of the story, " + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^0, \\ldots, \\mathbf{I}^M" + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "content": " onto a low-dimensional space using a frame-encoder " + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "content": ", and obtain the encoded frames " + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^0, \\ldots, \\mathbf{Z}^M" + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "content": " for a single story. To effectively condition on the corresponding frame, we apply this crossattention layer to the intermediate representations of the neural network for each frame " + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^m" + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "content": " and its textual description " + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^m" + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "text", + "content": " using Eq. (4), where the relevance of the description " + }, + { + "bbox": [ + 304, + 596, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^m" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 525, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 525, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 525, + 713 + ], + "type": "text", + "content": "2We drop the superscript denoting sample " + }, + { + "bbox": [ + 315, + 702, + 525, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 315, + 702, + 525, + 713 + ], + "type": "text", + "content": " for ease of notation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2496" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 548, + 168 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 548, + 168 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 548, + 168 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 548, + 168 + ], + "type": "table", + "html": "
DatasetMethodw/ ref. textChar-acc (↑)Char-F1 (↑)BG-acc (↑)BG-F1 (↑)FID (↓)
FlintstonesVLCStoryGAN [34]×27.7342.014.8316.49120.85
LDM [45]×79.8692.3348.0237.8661.40
LDM [45]✓57.3878.6844.1928.2587.39
Story-LDM (Ours)✓69.1986.5935.2128.8069.49
MUGENLDM [45]✓31.3921.2815.7418.66120.99
Story-LDM (Ours)✓93.4095.6092.1992.3762.16
", + "image_path": "dae1d99e3806c21f8ad4651d2e0fdf75ebc7ee5d4889f9e55e5fee5c3a609943.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 171, + 476, + 182 + ], + "lines": [ + { + "bbox": [ + 116, + 171, + 476, + 182 + ], + "spans": [ + { + "bbox": [ + 116, + 171, + 476, + 182 + ], + "type": "text", + "content": "Table 2. Quantitative results. Experimental results on the FlintstoneSV and the MUGEN datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 190, + 286, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 286, + 213 + ], + "type": "text", + "content": "is weighted by the similarity between the textual representation and the encoded frame representation " + }, + { + "bbox": [ + 46, + 190, + 286, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^m" + }, + { + "bbox": [ + 46, + 190, + 286, + 213 + ], + "type": "text", + "content": " (cf. Fig. 2a)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "type": "text", + "content": "For sequential generation, the model in addition to the current state, requires information from all the previous states. To enable this, the diffusion process for any frame representation " + }, + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^m" + }, + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "type": "text", + "content": " is conditioned on the visual representations of the previous frames " + }, + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^0,\\ldots ,\\mathbf{Z}^{m - 1}" + }, + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "type": "text", + "content": " as well as the sentence descriptions " + }, + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^0,\\dots ,\\mathbf{S}^m" + }, + { + "bbox": [ + 46, + 214, + 287, + 309 + ], + "type": "text", + "content": ". This conditioning is realized through a novel Memory-attention module which forms the basis of our autoregressive approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "content": "Memory-attention Module. To capture the spatiotemporal interactions across multiple frames and sentences for a story, in our conditional diffusion model, we condition the frame " + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^m" + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "content": " not only on the corresponding text " + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^m" + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "content": " but also on the previous texts " + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^i" + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "inline_equation", + "content": "i \\in \\{0, \\dots, m - 1\\}" + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "content": ". This conditioning is applied throughout the " + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "content": " time-steps of the diffusion process for " + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^m" + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "text", + "content": ". The conditional denoising autoencoder thus models the conditional distribution " + }, + { + "bbox": [ + 46, + 323, + 287, + 456 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Z}^m \\mid \\mathbf{Z}^{3 [45] which contains a cross-attention layer to generate text-to-image based story, without using our proposed autoregressive memory modules as our baselines for MU-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 334, + 69, + 519, + 269 + ], + "blocks": [ + { + "bbox": [ + 334, + 69, + 519, + 269 + ], + "lines": [ + { + "bbox": [ + 334, + 69, + 519, + 269 + ], + "spans": [ + { + "bbox": [ + 334, + 69, + 519, + 269 + ], + "type": "image", + "image_path": "775c49c02dcde6ebe29c1b6b38dfff3cd87165434393d11eeeb3cf728fe40565.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 271, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 306, + 271, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 545, + 293 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison on Story Generation. Comparison on the FlintstonesSV dataset visual story generation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 296, + 520, + 545 + ], + "blocks": [ + { + "bbox": [ + 331, + 296, + 520, + 545 + ], + "lines": [ + { + "bbox": [ + 331, + 296, + 520, + 545 + ], + "spans": [ + { + "bbox": [ + 331, + 296, + 520, + 545 + ], + "type": "image", + "image_path": "8dee7f301a52ce94e0debbfe835659f17ff1b026f4ec7955df4ec5b864851324.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 548, + 545, + 583 + ], + "lines": [ + { + "bbox": [ + 306, + 548, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 545, + 583 + ], + "type": "text", + "content": "Figure 6. Comparison to Story Continuation. Comparison on the FlintstonesSV dataset for story continuation. Prior work uses first frame as additional input to the model; our model does not." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 591, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 591, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 591, + 545, + 651 + ], + "type": "text", + "content": "GEN, PororoSV and FlintstonesSV datasets. The parameters of the diffusion model within the Story-LDM are initialized with the pre-trained LDM [45]. Similarly, for the textual embedding, we use BERT-tokenizer [9] and use the pre-trained text-transformer from LDM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 715 + ], + "type": "text", + "content": "Quantitative Results. Table 2 shows quantitative results for consistent story generation on the FlintstoneSV dataset. We compare the performance of our approach (row 4) to the LDM [45] which we train/test with both original (row 2) as well as the co-referenced (row 3) descriptions. Fur" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 702, + 271, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 271, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 271, + 712 + ], + "type": "text", + "content": "3https://github.com/CompVis/latent-diffusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2499" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 132 + ], + "type": "text", + "content": "thermore, we include the results of the state-of-art VLCStoryGAN [34] (row 1) with the original text of the dataset" + }, + { + "bbox": [ + 47, + 72, + 287, + 132 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 47, + 72, + 287, + 132 + ], + "type": "text", + "content": " (i.e. without co-references). We note that VLCStoryGAN was shown to be better than Duco-StoryGAN [35], CPCSV [49] and original StoryGAN [30] (see [34])." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 323 + ], + "type": "text", + "content": "Based on Table 2 we make three observations: (1) Our LDM baseline is better than VLCStoryGAN on the original reference-free text (cf. Tab. 2, rows 1 & 2). (2) Reference resolution makes the task considerably harder. With the reference text in our modified dataset, we observe a drop in performance in terms of character and background classification scores (cf. Tab. 2, rows 2 & 3). (3) Our model, with memory-attention module, significantly outperforms the baseline (cf. Tab. 2, rows 3 & 4) both in terms of generative image quality and character consistency; and outperforms SoTA of VLCStoryGAN by " + }, + { + "bbox": [ + 46, + 133, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\sim 41\\%" + }, + { + "bbox": [ + 46, + 133, + 288, + 323 + ], + "type": "text", + "content": " percentage points on character accuracy (while performing a more difficult version of the task). Further, our model, that is required to conduct reference resolution, comes close to the LDM trained with original, reference-free, text (cf. Tab. 2, rows 3 & 4), which can be viewed as a sort of an upper bound." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 324, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 466 + ], + "type": "text", + "content": "On the MUGEN dataset, our method outperforms the strong LDM baseline with gains of " + }, + { + "bbox": [ + 46, + 324, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\sim 62\\%" + }, + { + "bbox": [ + 46, + 324, + 287, + 466 + ], + "type": "text", + "content": " on character accuracy and " + }, + { + "bbox": [ + 46, + 324, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\sim 76\\%" + }, + { + "bbox": [ + 46, + 324, + 287, + 466 + ], + "type": "text", + "content": " on the background accuracy, thereby showing the advantages of the memory-attention mechanism for consistent story generation. We note that MUGEN dataset has more references across story scenes. Flintstones while contains more references per story overall, many of those references are within scenes as opposed to across scenes. Meaning that in terms of reference impact on consistency, MUGEN dataset is actually harder. Experimental results on the PororoSV dataset are provided in the Supplemental." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 468, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 589 + ], + "type": "text", + "content": "Qualitative Results. Figure 3 illustrates qualitative results on the MUGEN dataset. Rows 1, 2 & 3 show ground truth, LDM [45] and our Story-LDM approach, respectively. Here, we see that our method is able to maintain consistency in terms of both character and background. Similarly, in Figure 4 we can show the results on FlintstoneSV dataset which further validates the strong performance of our method when generating high-quality, consistent story. Compared to the LDM, our approach is able to adapt to the diverse backgrounds in the story descriptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 590, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 288, + 687 + ], + "type": "text", + "content": "Additional Results. We compare the qualitative results of our method to both story generation [34] and story continuation [36] in Figs. 5 and 6 respectively. The comparative images are taken directly from respective papers. We note that story continuation Fig. 6 is solving a different (easier) problem and with text that contains no-references. This makes the comparison to our method, which receives fewer inputs, not very meaningful. Nether-the-less, our approach, that" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 332, + 70, + 518, + 240 + ], + "blocks": [ + { + "bbox": [ + 332, + 70, + 518, + 240 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 518, + 240 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 518, + 240 + ], + "type": "image", + "image_path": "46986fe82fab2965014bf4fc3db2b1a79ddc0d1ecb320ddd4020fb48417732a9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 264, + 518, + 399 + ], + "blocks": [ + { + "bbox": [ + 306, + 242, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 306, + 242, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 545, + 264 + ], + "type": "text", + "content": "Figure 7. Story Diversity. Diverse outputs for a single storyline obtained with our Story-LDM." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 264, + 518, + 399 + ], + "lines": [ + { + "bbox": [ + 331, + 264, + 518, + 399 + ], + "spans": [ + { + "bbox": [ + 331, + 264, + 518, + 399 + ], + "type": "image", + "image_path": "0fa884a88d96beb5fc5fba5cc3ec403c5098351be5c29791dd2e5f409e4d4ab9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 399, + 545, + 431 + ], + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 431 + ], + "type": "text", + "content": "Figure 8. Branching Storyline. Generating different yet consistent stories by branching the storyline. Frames in the later columns are generated based on earlier ones and corresponding text." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 437, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 570 + ], + "type": "text", + "content": "can resolve references and is solving a harder story generation task, obtains highly competitive results. Furthermore, to show that our autoregressive visual memory module can generate diverse stories conditioned on the current and previous condition, we create different story-lines starting for a single sentence. In Fig. 8, we can see for reference 'they', the model can generate both the characters according to the storyline already parsed. Moreover, in Fig. 7 we show that our approach can not only generate consistent visual stories, but also diverse frames for the same text (cf. Fig. 7). Additional results are provided in the Supplemental." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we formulate consistent story generation in a more realistic way by co-referencing actors/backgrounds in the story descriptions. We develop an autoregressive Story-LDM approach with memory attention capable of maintaining consistency across the frames based on the previously generated frames and their corresponding descriptions. We introduced modified datasets to evaluate the performance for reference resolution. We expect our proposed formulation and models to be conducive to the real-world use cases and further the research." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "Results for [34] were obtained using pretrained model provided by original authors in private communication." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2500" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "text", + "content": "[1] Jyoti Aneja, Harsh Agrawal, Dhruv Batra, and Alexander G. Schwing. Sequential latent spaces for modeling the intention during diverse image captioning. In ICCV, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "text", + "content": "[2] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "type": "text", + "content": "[3] Chinatsu Aone and Scott William. Evaluating automated and manual acquisition of anaphora resolution strategies. In Annual Meeting of the Association for Computational Linguistics, pages 122-129, 1995. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 248 + ], + "type": "text", + "content": "[4] Apratim Bhattacharyya, Shweta Mahajan, Mario Fritz, Bernt Schiele, and Stefan Roth. Normalizing flows with multiscale autoregressive priors. In CVPR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "type": "text", + "content": "[5] Hong Chen, Rujun Han, Te-Lin Wu, Hideki Nakayama, and Nanyun Peng. Character-centric story visualization via visual planning and token alignment. In EMNLP, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "text", + "content": "[6] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "type": "text", + "content": "[7] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 364, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 384 + ], + "type": "text", + "content": "[8] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2018. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 443, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 485 + ], + "type": "text", + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "type": "text", + "content": "[11] Laurent Dinh, David Krueger, and Yoshua Bengio. NICE: Non-linear independent components estimation. In ICLR Workshop, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 543 + ], + "type": "text", + "content": "[12] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real NVP. In ICLR, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[13] Hongliang Fei, Xu Li, Dingcheng Li, and Ping Li. End-to-end deep reinforcement learning based coreference resolution. In Annual Meeting of the Association for Computational Linguistics, pages 660-665, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[16] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[17] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "text", + "content": "[18] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In CVPR, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[19] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Isabelle Hu, and Devi Parikh. Mugen: A playground for video-audio-text multimodal understanding and generation. arXiv preprint arXiv:2204.08058, 2022. 1, 2, 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 2017. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 239, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 259 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 261, + 545, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 313 + ], + "type": "text", + "content": "[22] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[23] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In CVPR, 2016. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 350, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 402 + ], + "type": "text", + "content": "[24] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pretraining by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64-77, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 404, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 426 + ], + "type": "text", + "content": "[25] Minguk Kang and Jaesik Park. Contragan: Contrastive learning for conditional image generation. NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 427, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 447 + ], + "type": "text", + "content": "[26] Andrew Kehler. Probabilistic coreference in information extraction. 1997. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "text", + "content": "[27] Diederik P. Kingma and Max Welling. Auto-encoding variational Bayes. In ICLR, 2014. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "text", + "content": "[28] Hideo Kobayashi, Yufang Hou, and Vincent Ng. End-to-end neural bridging resolution. In International Conference on Computational Linguistics, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "text", + "content": "[29] Chunye Li, Liya Kong, and Zhiping Zhou. Improved-storygan for sequential images visualization. Journal of Visual Communication and Image Representation, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "text", + "content": "[30] Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. Storygan: A sequential conditional gan for story visualization. In CVPR, 2019. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "type": "text", + "content": "[31] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 1, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 614, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 647 + ], + "type": "text", + "content": "[32] Shweta Mahajan, Iryna Gurevych, and Stefan Roth. Latent normalizing flows for many-to-many cross-domain mappings. In *ICLR*, 2020. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 647, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 669 + ], + "type": "text", + "content": "[33] Shweta Mahajan and Stefan Roth. Diverse image captioning with context-object split latent spaces. NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[34] Adyasha Maharana and Mohit Bansal. Integrating visuospatial, linguistic and commonsense structure into story visualization. arXiv preprint arXiv:2110.10834, 2021. 1, 3, 5, 7, 8" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2501" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 635 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "type": "text", + "content": "[35] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Improving generation and evaluation of visual stories via semantic consistency. 2021. 1, 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[36] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. In ECCV, 2022. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "text", + "content": "[37] Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L. Yuille. Deep captioning with multimodal recurrent neural networks (m-RNN). In ICLR, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 194 + ], + "type": "text", + "content": "[38] Joseph F McCarthy and Wendy G Lehnert. Using decision trees for coreference resolution. 1995. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "text", + "content": "[39] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 314 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[41] Altaf Rahman and Vincent Ng. Narrowing the modeling gap: A cluster-ranking approach to coreference resolution. Journal of Artificial Intelligence Research, 2011. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 390 + ], + "type": "text", + "content": "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 392, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 287, + 425 + ], + "type": "text", + "content": "[43] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[44] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "text", + "content": "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351 of Lecture Notes in Computer Science, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "text", + "content": "[47] Paul Hongsuck Seo, Andreas Lehrmann, Bohyung Han, and Leonid Sigal. Visual reference resolution using attention memory for visual dialog. NIPS, 2017. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[48] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 635 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[49] Yun-Zhu Song, Zhi Rui Tam, Hung-Jen Chen, Huiao-Han Lu, and Hong-Han Shuai. Character-preserving coherent story visualization. In ECCV, 2020. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[50] Wee Meng Soon, Hwee Tou Ng, and Daniel Chung Yong Lim. A machine learning approach to coreference resolution of noun phrases. Computational linguistics, 2001. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "type": "text", + "content": "[51] Hung-Yu Tseng, Hsin-Ying Lee, Lu Jiang, Ming-Hsuan Yang, and Weilong Yang. Retrievegan: Image synthesis via differentiable patch retrieval. In ECCV, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 209 + ], + "type": "text", + "content": "[52] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 210, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 210, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 545, + 244 + ], + "type": "text", + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 245, + 545, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 290 + ], + "type": "text", + "content": "[54] Ashwin K. Vijayakumar, Michael Cogswell, Ramprasaath R. Selvaraju, Qing Sun, Stefan Lee, David J. Crandall, and Dhruv Batra. Diverse beam search for improved description of complex scenes. In AAAI, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 291, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 314 + ], + "type": "text", + "content": "[55] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NIPS, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 315, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 315, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 545, + 358 + ], + "type": "text", + "content": "[56] Liwei Wang, Alexander G. Schwing, and Svetlana Lazebnik. Diverse and accurate image description using a variational auto-encoder with an additive Gaussian encoding space. In NIPS, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "text", + "content": "[57] Wei Wang, Xavier Alameda-Pineda, Dan Xu, Pascal Fua, Elisa Ricci, and Nicu Sebe. Every smile is unique: Landmark-guided diverse smile generation. In CVPR, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 406, + 545, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 451 + ], + "type": "text", + "content": "[58] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 452, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 485 + ], + "type": "text", + "content": "[59] Guojun Yin, Bin Liu, Lu Sheng, Nenghai Yu, Xiaogang Wang, and Jing Shao. Semantics disentangling for text-to-image generation. CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 486, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 532 + ], + "type": "text", + "content": "[60] Gangyan Zeng, Zhaohui Li, and Yuan Zhang. Pororogan: an improved story visualization model on pororo-sv dataset. In Proceedings of the 3rd International Conference on Computer Science and Artificial Intelligence, 2019. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 532, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 566 + ], + "type": "text", + "content": "[61] Han Zhang, Ian J. Goodfellow, Dimitris N. Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, 2019. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "type": "text", + "content": "[62] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In CVPR, pages 833-842, 2021. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "type": "text", + "content": "[63] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmagan: Dynamic memory generative adversarial networks for text-to-image synthesis. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2502" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_content_list.json b/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3fb04fd946744bafeb797c7746d2a5569f8d6eed --- /dev/null +++ b/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_content_list.json @@ -0,0 +1,1484 @@ +[ + { + "type": "text", + "text": "Making Vision Transformers Efficient from A Token Sparsification View", + "text_level": 1, + "bbox": [ + 119, + 128, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuning Chang $^{1*}$ Pichao Wang $^{2\\dagger\\dagger}$ Ming Lin $^{2\\dagger}$ Fan Wang $^{2}$ David Junhao Zhang $^{1}$ Rong Jin $^{2}$ Mike Zheng Shou $^{1\\dagger}$ $^{1}$ Show Lab, National University of Singapore $^{2}$ Alibaba Group", + "bbox": [ + 143, + 178, + 823, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The quadratic computational complexity to the number of tokens limits the practical applications of Vision Transformers (ViTs). Several works propose to prune redundant tokens to achieve efficient ViTs. However, these methods generally suffer from (i) dramatic accuracy drops, (ii) application difficulty in the local vision transformer, and (iii) non-general-purpose networks for downstream tasks. In this work, we propose a novel Semantic Token ViT (STViT), for efficient global and local vision transformers, which can also be revised to serve as backbone for downstream tasks. The semantic tokens represent cluster centers, and they are initialized by pooling image tokens in space and recovered by attention, which can adaptively represent global or local semantic information. Due to the cluster properties, a few semantic tokens can attain the same effect as vast image tokens, for both global and local vision transformers. For instance, only 16 semantic tokens on DeiT-(Tiny,Small,Base) can achieve the same accuracy with more than $100\\%$ inference speed improvement and nearly $60\\%$ FLOPs reduction; on Swin-(Tiny,Small,Base), we can employ 16 semantic tokens in each window to further speed it up by around $20\\%$ with slight accuracy increase. Besides great success in image classification, we also extend our method to video recognition. In addition, we design a STViT-R(covery) network to restore the detailed spatial information based on the STViT, making it work for downstream tasks, which is powerless for previous token sparsification methods. Experiments demonstrate that our method can achieve competitive results compared to the original networks in object detection and instance segmentation, with over $30\\%$ FLOPs reduction for backbone.", + "bbox": [ + 73, + 300, + 473, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 797, + 209, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to standard Convolutional Neural Networks (CNNs) approaches which process images pixel-by-pixel,", + "bbox": [ + 76, + 821, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) [15, 26, 35, 36, 43] treat an image as a sequence of patch/image tokens, and have shown promising performance in prevalent visual recognition scenarios. However, these superior performances do not come for free: the quadratic computational complexity to the number of image tokens limits their application in practice. Previous works [33, 56] have illustrated the large amount of redundancy in the image tokens and also shown the effect of filtering out unimportant tokens normally according to predefined scoring mechanism. However, these methods face the following challenges. Firstly, the predefined scoring mechanisms for filtering are generally imprecise. In Figure 1, on the left we visualize the class token values in different layers which are commonly used to score the token importance [16, 24, 45]. Different layers have different value distributions, thus using these imprecise scores for filtering would lead to unsatisfactory performance. For example, EViT [24] has an accuracy drop of $1.3\\%$ when saving $50\\%$ FLOPs on DeiT-S [35]. Secondly, the remaining tokens do not distribute evenly in space any more, making them hard to work in local vision transformers1. Finally, large-scale token pruning tremendously damages the spatial structure and positional information, and causes difficulties when applied to downstream tasks, which they do not propose a solution to deal with.", + "bbox": [ + 500, + 268, + 893, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve these problems, we propose Semantic Token ViT (STViT), for efficient global and local vision transformers, which also can be revised to serve as backbone for downstream tasks. The proposed approach is based on the following observations: (i) unlike local CNNs which learn spatial structure of images, vision transformer discretizes feature map as tokens for global feature exploration, relieving the requirements for maintaining the whole image structure and information; (ii) discrete tokens are more beneficial for optimization [38]; (iii) in Figure 1, on the right shows the attention maps in different transformer layers, and there are only several vertical lines in the deep layers, which means that only a few tokens with global se", + "bbox": [ + 496, + 648, + 893, + 847 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1In this paper, we define the vision transformer with global self-attention (like DeiT) as global vision transformer and the vision transformer with local self-attention (like Swin) as local vision transformer.", + "bbox": [ + 500, + 862, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Alibaba Group.", + "bbox": [ + 94, + 862, + 367, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal corresponding authors.", + "bbox": [ + 96, + 875, + 258, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Work done at Alibaba Group, and now affiliated with Amazon.", + "bbox": [ + 94, + 887, + 433, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6195", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bde36492b427be85ae4cd307425e95003502ef7a7d4b4f6e038ec4f55f7518fd.jpg", + "image_caption": [ + "Figure 1. Left: the attention values of class tokens (normalized and reshaped in image shape) in different self-attention layers. Right: the attention maps in different self-attention layers. Zoom-in for better visibility." + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 442, + 161 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f30fdba2db5b29b8a686a45071e5d10b6580de568ef44123696b1d30204ad83c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 90, + 843, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "matic information matter. Thus, we argue that it is not necessary to maintain massive structured tokens for ViTs, especially in the deep layers. Employing a few discrete tokens with high-level semantic information can potentially achieve both high performance and efficiency.", + "bbox": [ + 75, + 215, + 470, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In STViT, the semantic tokens represent the cluster centers, and the number of them is far less than the original image tokens, significantly reducing the computational cost. Inspired by the fact that multi-head attention can conduct the cluster center recovery (Supplementary A.7), we only employ the off-the-shelf self-attention to generate the semantic tokens. Specifically, the first few transformer layers are kept unchanged to obtain the image tokens with low-level features. The image tokens are then fed into our semantic token generation module (STGM) consisting of at least two transformer layers to generate semantic tokens. In each self-attention layer, the semantic tokens are input as queries, and the image tokens are fed as keys and values. The semantic tokens dynamically aggregate image tokens through the attention layers to recover cluster centers. In the first attention layer, the semantic tokens are initialized by an intra and inter-window spatial pooling which takes into account incorporating semantic information in each window and maximizing distance between adjacent windows. Thanks to this spatial initialization, the semantic tokens mainly incorporate local semantic information and achieve discrete and uniform distribution in space. In the following attention layer, besides further clustering, the semantic tokens are equipped with global cluster centers, and the network can adaptively select partial semantic tokens to focus on global semantic information. After the STGM, the original image tokens are discarded, and only semantic tokens are kept for the subsequent transformer layers. Because the generation of semantic tokens is flexible and space-aware, our method can be plugged into both global and local vision transformers. The semantic tokens can be produced in each window for the local vision transformer.", + "bbox": [ + 75, + 294, + 470, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another property of STViT is its capability to serve as a backbone for downstream tasks, such as object detection and instance segmentation. Discussions have been missing in previous methods [16, 24, 32, 45, 56] about how to use them in downstream task under the massive loss of spatial information during the token sparsification process, which actually seriously impedes the application of their method. Instead, we design a novel STViT-R network based", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on STViT where a recovery module and dumbbell unit are adopted to periodically restore the full resolution feature map while the intermediate transformer layers continue to use semantic tokens to save computation cost, making our method work in downstream task.", + "bbox": [ + 496, + 215, + 892, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The effectiveness of the proposed method is validated via a comprehensive empirical study on image and video ViT models. Only 16 semantic tokens on DeiT-(Tiny, Small, Base) achieve nearly $50\\%$ inference time reduction without any accuracy degradation; on Swin-(Tiny, Small, Base), we also improve the inference throughput by nearly $20\\%$ with slight accuracy increase. Moreover, the proposed STViT-R achieves promising results on object detection and instance segmentation. To the best of our knowledge, this is one of first works to apply the token sparsification algorithm in local vision transformers, and use the ViTs as backbones in downstream tasks after large-scale token pruning. Our findings in ViTs uncover that maintaining the full-size feature map is unnecessary, and a few tokens with high-level semantic representations can achieve both high performance and efficiency. Thanks to its simplicity and general-purpose ability, our method can also serve as a new efficient ViT baseline architecture and a starting point for further research from the token sparsification perspective.", + "bbox": [ + 496, + 294, + 893, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 601, + 637, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision transformers. Vision Transformer (ViT) [15] first introduces a pure Transformer backbone for image classification. ViT variants further inspire the applications of transformer to various vision tasks beyond image/video classification [1, 3, 20, 35, 36, 39, 43, 44, 49, 50], such as object detection [5, 12, 52, 55], semantic segmentation [37, 42, 53], and self-supervised learning [6, 9, 22]. Vanilla transformers have high computational and memory costs because the multi-head self-attention has quadratic computational complexity to the number of image tokens. Recently, various efficient ViTs have been proposed to alleviate this issue. The existing methods mainly focus on reducing the complexity of self-attention or reducing the number of tokens. Swin Transformer [26] adopts local self-attention, i.e., attending neighboring tokens within a constant window size, achieving a linear computational complexity in the self-attention with high performance. Many subsequent works [11, 14, 18, 40, 46, 48, 54] follow the local", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6196", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "self-attention design to develop variants. Token sparsification [8, 16, 21, 24, 29–34, 41, 45, 47, 56] also attracts increasing attention.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token sparsification. Token sparsification methods can be mainly categorized into hard pruning [8,16,21,24,29-31, 33,34,45,47] and soft pruning [32,56]. Hard pruning methods filter out some unimportant tokens according to a predefined scoring mechanism. DynamicViT [31], SPViT [21], and AdaViT [29] introduce additional prediction networks to score the tokens. Evo-ViT [45], ATS [16], and EViT [24] utilize the values of class tokens to evaluate the importance of tokens. However, it is difficult to achieve precise scoring as shown in the left of Figure 1. Therefore, they usually suffer from a significant accuracy drop. For instance, EViT [24] has an accuracy drop of $1.3\\%$ when saving $50\\%$ FLOPs on DeiT-S. Soft pruning methods generate new tokens from image tokens by importing additional attention networks. TokenLearner [32] also argue for a few tokens to replace image tokens. However, its price is a $1.8\\%$ accuracy drop when reducing $44\\%$ FLOPs, which is far inferior to concurrent works. Besides performance degradation, previous methods also have the following disadvantages. First, whether or how to extend the methods to local vision transformers remains unexplored. Second, it has not been discussed about how to serve the downstream tasks like object detection and instance segmentation after the tokens are pruned.", + "bbox": [ + 76, + 161, + 472, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our method, we apply the off-the-shelf transformer layers to reduce token number. [2, 10, 19, 23, 28, 51] adopt similar approaches to achieve efficient non-local relationships. Our method is different from them as below: (i) our method extracts local semantic information instead of non-local relationships; (ii) the semantic tokens are a few cluster centers, which can replace the massive image tokens to achieve image classification; (iii) our method specializes in pruning tokens.", + "bbox": [ + 75, + 523, + 468, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 672, + 168, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed STViT is presented in this section, which aims to construct an efficient and high-performance ViT. STViT is first introduced in Section 3.1, followed by how to apply STViT in the local vision transformer in Section 3.2. Based on STViT, STViT-R is developed to restore the spatial resolution for downstream tasks in Section 3.3.", + "bbox": [ + 75, + 696, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.STViT", + "text_level": 1, + "bbox": [ + 76, + 800, + 165, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall architecture. An overview of STViT architecture is presented in Figure 2a. The patch embedding layer and shallow transformer layers are kept unchanged as a base module in our method. The base module copes with all the image tokens $X \\in \\mathbb{R}^{N_i \\times C}$ to extract low-level features,", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_{i}$ is the number of image tokens and $C$ is the number of channels. The image tokens are fed into the semantic token generation module (STGM) to generate $N_{s}$ semantic tokens $S \\in \\mathbb{R}^{N_s \\times C}$ . After the STGM, the image tokens $X$ can be discarded, and only semantic tokens $S$ with high-level semantic information are used in all the subsequent transformers. Due to $N_{s} \\ll N_{i}$ , our method can significantly reduce the computational cost.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic token generation module (STGM). The whole image is represented by a few tokens with high-level semantic information through clustering. Inspired by the fact that self-attention can conduct cluster center recovery (Supplementary A.7), we adopt the off-the-shelf self-attention layers to produce the semantic tokens. The STGM consists of at least two transformer layers.", + "bbox": [ + 496, + 234, + 890, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initial cluster centers $P \\in \\mathbb{R}^{N_s \\times C}$ are generated by an spatial pooling which pools the image tokens into fixed $w_{s} \\times w_{s}$ tokens, with $N_{s} = w_{s} \\times w_{s}$ . $w_{s}$ is generally set as 4. The spatial pooling can be achieved by a non-parameterized adaptive spatial pooling or a super lightweight network with higher performance, intra and inter-window spatial pooling. The intentions of spatial pooling initialization are three folds. First, the initial cluster centers can distribute uniformly in space, making the generated semantic tokens more discrete and preventing the semantic tokens from collapsing to one point in the following layers. Second, the semantic tokens can be forced to represent more local and distinguished features. Finally, the representation of semantic tokens is associated with the specific spatial locations, which is the basis to allow our method to be applied in local self-attention and downstream tasks. The initial cluster centers $P$ then dynamically integrate the image tokens $X$ according to semantic information by attention mechanism. In the first transformer layer, the processing of the generation of semantic tokens can be written as", + "bbox": [ + 496, + 340, + 892, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} ^ {1} = M H A (P, X, X) + P, \\quad S ^ {1} = F F N (\\hat {S} ^ {1}) + \\hat {S} ^ {1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 647, + 890, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $MHA$ and $FFN$ are short for multi-head attention layer and feed-forward network, respectively, and the triplet input of $MHA$ are queries, keys, and values in turn. All the norm layers are omitted in all the equations for brevity. The initial cluster centers are produced in a window by an adaptive spatial pooling layer or an intra and inter-window spatial pooling, while the semantic tokens are generated from a global receptive field by a dynamic attention layer to ensure that they can extract high-level semantic representation. In order to further strengthen the clustering effect, we use the second transformer layer to repeat the clustering operation and guide the semantic tokens to extract global information. In this transformer, the semantic tokens are updated as:", + "bbox": [ + 496, + 666, + 890, + 863 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} ^ {2} = M H A \\left(S ^ {1} + G, \\operatorname {C o n c a t} \\left(S ^ {1}, X\\right), \\operatorname {C o n c a t} \\left(S ^ {1}, X\\right)\\right) + S ^ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 866, + 890, + 888 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {2} = F F N (\\hat {S} ^ {2}) + \\hat {S} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 887, + 756, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6197", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4d7fd0c0fdcb716fed6e63a203d651069a48e880cac0e37f31f5b9202b34a222.jpg", + "image_caption": [ + "(a) STViT" + ], + "image_footnote": [], + "bbox": [ + 125, + 87, + 483, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f17100edef88e4795d9c0d37690577ea35a91fff9e7719a10b5fcfb07508a2e8.jpg", + "image_caption": [ + "(b) STViT-R", + "Figure 2. The architectures of our STViT and STViT-R." + ], + "image_footnote": [], + "bbox": [ + 584, + 92, + 828, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G \\in \\mathbb{R}^{N_s \\times C}$ is the global cluster centers initialized by Gaussian noise and $\\text{Concat}(\\cdot)$ is a concatenation operation. The global cluster centers $G$ are responsible for global semantic information extraction like the class token. Although $S^1$ and $G$ are summed together as the queries, they can be decoupled in the attention computation as:", + "bbox": [ + 75, + 265, + 468, + 356 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} A _ {s} = \\left(S ^ {1} \\cdot W _ {q}\\right) \\cdot \\left(\\left(S ^ {1} + X\\right) \\cdot W _ {k}\\right), \\quad A _ {g} = \\left(G \\cdot W _ {q}\\right) \\cdot \\left(\\left(S ^ {1} + X\\right) \\cdot W _ {k}\\right), \\tag {3} \\\\ A = \\operatorname {S o f t m a x} (A _ {s} + A _ {g}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 361, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{q}$ and $W_{k}$ are the linear projection weights of queries and keys. We can see that $S^{1}$ and $G$ integrate the keys to generate $A_{s}$ and $A_{g}$ independently and just share a Softmax operation to produce the final attention map $A$ . Therefore, besides spatial semantic information, the semantic tokens also incorporate global semantic information with a negligible additional overhead. Though very similar, the global cluster centers are actually different from the learned positional encoding. We do not add the global cluster centers to keys. Moreover, it will be shown in the experiments that inserting the actual learned positional encoding will cause an accuracy drop. The number of transformer layers in STGM is flexible. More transformer layers can be associated for clustering. Two transformer layers are employed by default, i.e., $S^{2}$ is the output of the STGM. Note that the image tokens are not updated in the STGM.", + "bbox": [ + 75, + 397, + 470, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intra and inter-window spatial pooling. Give $H \\times W$ feature map $X$ , we generate $N_{s}(w_{s} \\times w_{s})$ initial cluster centers. We uniformly split the $X$ into $w_{s} \\times w_{s}$ windows $[X_{w}^{i}]_{i=1}^{N_{s}}$ with size $\\frac{H}{w_{s}} \\times \\frac{W}{w_{s}}$ and each window generates one initial cluster center. To represent abundant semantic information, we take into account intra and interwindow relations, i.e., integrating important tokens in the window and maximizing the distance among initial cluster centers in different windows. Specifically, we formulate an intra-window function $f_{intra}$ to produce a mask, $M_{i} = f_{intra}(X_{w}^{i})$ , which projects the input window from $\\mathbb{R}\\frac{H}{w_{s}} \\times \\frac{W}{w_{s}} \\times C \\rightarrow \\mathbb{R}\\frac{H}{w_{s}} \\times \\frac{W}{w_{s}}$ . The idea is to let the intra-window function $f_{intra}$ adaptively select a combination of informative tokens in $X_{w}^{i}$ , which is implemented by", + "bbox": [ + 76, + 667, + 468, + 882 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} = \\operatorname {C o n v} \\left(\\operatorname {G e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\operatorname {D e p t h w i s e} \\operatorname {C o n v} \\left(X _ {w} ^ {i}\\right)\\right)\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 886, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we compute each integrated token $\\hat{P}_i =$ Softmax $(M_i)\\cdot X_w^i$ from each window, and arrange $[\\hat{P}_i]_{i = 1}^{N_s}$ by spatial structure to form the 2D tensor $\\hat{P}\\in \\mathbb{R}^{w_s\\times w_s\\times C}$ We adopt an inter-window function $f_{inter}$ to compute the inter-window relations and generate the offset, $O =$ $f_{inter}(\\hat{P})$ ,to revise the mask $M$ . The implementation of $f_{inter}$ is similar to Eq. 4, except the mapping input from $\\mathbb{R}^{w_s\\times w_s\\times C}\\rightarrow \\mathbb{R}^{w_s\\times w_s\\times \\frac{HW}{w_s^2}}$ , where $\\frac{HW}{w_s^2}$ is the number of tokens in each window. For each window, the corresponding $O_{i}\\in \\mathbb{R}^{\\frac{HW}{w_{s}^{2}}}$ is sliced from $O$ and reshaped to $\\mathbb{R}^{\\frac{H}{w_s}\\times \\frac{W}{w_s}}$ The $O_{i}$ is used to revise $M_{i}$ . The final initial cluster center $P_{i}$ is computed by", + "bbox": [ + 498, + 265, + 893, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} = \\operatorname {S o f t m a x} \\left(M _ {i} + O _ {i}\\right) \\cdot X _ {w}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 463, + 890, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our $f_{\\text{intra}}$ and $f_{\\text{intra}}$ are super lightweight and the introducing parameters can be negligible. For example, on DeiT, the parameters only increase by $0.05\\%$ .", + "bbox": [ + 498, + 483, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. STViT in local vision transformers.", + "text_level": 1, + "bbox": [ + 500, + 542, + 805, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local self-attention has been widely used in current ViT models to balance efficiency and accuracy. As the generation of semantic tokens in STViT is flexible in space, it can be naturally applied in local self-attention. Suppose each local self-attention layer contains $N_{w}$ windows with size $w \\times w$ , we initialize $w_{s} \\times w_{s}$ cluster centers in each window by our intra and inter-window spatial pooling. The total number of semantic tokens $N_{s}$ would be $w_{s} \\times w_{s} \\times N_{w}$ . $w$ and $w_{s}$ are set as 7 and 3 by default separately. As a result, our method compresses more than $80\\%$ image tokens in local self-attention. In the STGM, although initial cluster centers are from $w \\times w$ windows, we use larger windows with size $w_{k} \\times w_{k}$ to obtain keys and values in Eq. 1 and Eq. 2 to mitigate the effect of limited window size. Other operations in STGM are kept the same as Section 3.1.", + "bbox": [ + 496, + 566, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the local ViT models, each local transformer layer is normally followed by a cross-window connection layer, such as a shift window transformer layer following a local transformer layer on Swin Transformer [26]. In our method, the attention is computed within $w_{s} \\times w_{s}$ window in the local self-attention layer, and the cross-window connection can be achieved by computing self-attention in a larger-size", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6198", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(e.g., 4 \\times w_{s})$ sliding window because of the rare number of tokens in each window. For the low-resolution input, our cross-window connection layer is equal to a global self-attention layer.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. STViT for downstream tasks", + "text_level": 1, + "bbox": [ + 76, + 161, + 333, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method significantly reduces the computation cost by using a small number of semantic tokens, while its side effect is losing nearly all the detailed position information. High-level vision tasks, such as object detection and instance segmentation, are difficult to be executed on this extremely incomplete feature map. This issue also exists in previous works, which hinders the application of token sparsification methods. To solve this issue, we design a STViTR network based on STViT to restore the original spatial resolution from the semantic tokens.", + "bbox": [ + 75, + 181, + 468, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our STViT-R shown in Figure 2b has two modifications compared with STViT. First, we adopt a recovery module to restore the spatial resolution from semantic tokens; second, we regroup the transformer layers and construct dumbbell units composed of our STViT-R.", + "bbox": [ + 75, + 334, + 468, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recovery module. In the recovery module, only the self-attention layer is employed to restore the spatial resolution without any additional networks. The image tokens $X$ and semantic tokens $S$ are partitioned as $N_w^r$ windows of size $w^r \\times w^r$ and $w_s^r \\times w_s^r$ , respectively. The image tokens in each window aggregate the semantic tokens in the corresponding window, which is represented as:", + "bbox": [ + 75, + 428, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} = M H A (X, S, S) + X, \\quad X = F F N (\\hat {X}) + \\hat {X}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 536, + 467, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This is a reverse operation of the generation of semantic tokens, using high-level semantic information to boost the image tokens.", + "bbox": [ + 75, + 556, + 468, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dumbbell unit. The transformer layers are regrouped into multiple dumbbell units in our STViT-R. Each dumbbell unit consists of four parts. The transformers in the first part are responsible for coping with image tokens; the second part is the semantic token generation module; the transformer layers in the third part deal with semantic tokens; the last part is the recovery module. Take the application on Swin-S (STViT-R-Swin-S) as an example. One, two, two and one transformer layers are allocated for these four parts, respectively. In total, each dumbbell unit is composed of 6 transformer layers. We concatenate three dumbbell units in Stage 3 of Swin-S. In each dumbbell unit, the intermediate transformer layers process semantic tokens with high-level semantic information to save computational cost, and the complete spatial resolution is recovered at the end. By repeating multiple dumbbell units, the detailed spatial information will be preserved by the network, which can not only enhance the classification but also serve downstream tasks.", + "bbox": [ + 75, + 619, + 468, + 906 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 633, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "STViT will first be applied in two representative ViT models, DeiT [35] and Swin [26] for image classification and video recognition. To validate that our method is effective in downstream tasks, STViT-R is then performed on object detection and instance segmentation tasks.", + "bbox": [ + 498, + 114, + 890, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Image classification", + "text_level": 1, + "bbox": [ + 500, + 193, + 687, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings. For image classification, all the models are trained on the ImageNet [13] with 1.28M training images and 50K validation images from 1,000 classes. By default, the semantic token generation module (STGM) employs the $5^{th}$ and $6^{th}$ transformer layers of DeiT (with 12 layers in total), employs the $3^{th}$ and $4^{th}$ transformer layers of Stage 3 of Swin-T (with 12 layers in total), and employs the $11^{th}$ and $12^{th}$ transformer layers of Stage 3 of Swin-S and Swin-B (with 24 layers in total). The image resolution in training and inference is $224 \\times 224$ unless otherwise specified. The batch size is 1,024. All the models are trained from scratch for 300 epochs, and the augmentation and regularization strategies follow the original papers of DeiT and Swin. No knowledge distillation algorithms are used in our experiments. The classification is performed by applying a global average pooling layer on the output tokens of the last transformer layer, followed by a linear classifier. In evaluation, the top-1 accuracy using a single crop is reported. The FLOPs computations of this paper are measured by Fvcore². Throughput is measured with the batch size of 128 on a V100 GPU.", + "bbox": [ + 498, + 215, + 890, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On Swin [26], the semantic tokens are generated in Stage 3, and they are not downsampled in Stage 4 due to rare semantic tokens. The patch merging layer between Stage 3 and Stage 4 is replaced with a simple linear layer to double the number of channels. $w_{k}$ is set as 10 and 14 for two transformer layers of STGM.", + "bbox": [ + 498, + 532, + 890, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. One of the advantages of our method is that it can be applied to both global and local vision transformers to reduce computational complexity. Our main results on DeiT and Swin are summarized in Table 1 and Table 2, respectively. The results of LV-ViT [20] are illustrated in Supplementary A.3 due to limited space. We report the top-1 accuracy, FLOPs, and the throughput under different numbers of semantic tokens. On DeiT, the models with 16 semantic tokens achieve the same accuracy as the DeiT models with 196 tokens and save nearly $60\\%$ FLOPs on DeiT-T, DeiT-S, and DeiT-B. With more semantic tokens, the accuracy can consistently outperform the base models. For instance, STViT-DeiT-B with 36 semantic tokens surpasses DeiT-B by $0.4\\%$ accuracy with $52\\%$ FLOPs reduction.", + "bbox": [ + 498, + 635, + 890, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The local vision transformer like Swin is already an efficient architecture compared to the global vision trans", + "bbox": [ + 500, + 847, + 890, + 876 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2https://github.com/facebookresearch/fvcore", + "bbox": [ + 514, + 886, + 857, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6199", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3539e1609dbd5b3149f9e70af42333b4553874fc6d39866365a45a0ac954607e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMetricsBaseNo. of semantic tokens
163664100
STViT-DeiT-TTop-1 Acc(%)72.272.2(+0.0%)72.7(+0.5)73.0(+0.8)73.2(+1.0)
FLOPs(G)1.260.53(-58%)0.60(-52%)0.71(-44%)0.86(-32%)
Throughput(img/s)27525511(+101%)4769(+74%)4214(+53%)3551(+29%)
STViT-DeiT-STop-1 Acc(%)79.879.8(+0.0)80.1(+0.3)80.5(+0.7)80.6(+0.8)
FLOPs(G)4.581.91(-58%)2.20(-52%)2.62(-43%)3.16(-31%)
Throughput(img/s)14082891(+105%)2542(+80%)2229(+58%)1837(+30%)
STViT-DeiT-BTop-1 Acc(%)81.881.8(+0.0)82.2(+0.4)82.6(+0.8)82.7(+0.9)
FLOPs(G)17.587.31(-58%)8.44(-52%)10.04(-43%)12.13(-31%)
Throughput(img/s)6261308(+110%)1150(+85%)1087(+61%)826(+33%)
", + "bbox": [ + 145, + 88, + 823, + 271 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/85cf7519e0a9830ba3e8a416b696647a210698ab8c9c9bb83727f9a692775df7.jpg", + "table_caption": [ + "Table 1. Applying STViT to DeiT-T, DeiT-S, and DeiT-B. The top-1 accuracy, complexity in FLOPs, and throughput are reported for different numbers of semantic tokens." + ], + "table_footnote": [], + "table_body": "
ModelMetricsBaseMove STGMNo. of semantic tokens
4916
STViT-Swin-TTop-1 Acc(%)81.381.0(-0.3%)80.8(-0.5)81.5(+0.2)81.8(+0.5%)
FLOPs(G)4.53.14(-30%)2.99(-34%)3.43(-24%)4.06(-10%)
Throughput(img/s)8781124(+29%)1128(+29%)1061(+22%)1008(+15%)
STViT-Swin-STop-1 Acc(%)83.082.8(-0.2%)82.4(-0.6%)83.0(-0.0)83.1(+0.1%)
FLOPs(G)8.75.95(-32%)5.95(-32%)6.53(-25%)7.36(-15%)
Throughput(img/s)551739(+35%)732(+34%)691(+26%)657(+20%)
STViT-Swin-BTop-1 Acc(%)83.583.2(-0.3%)83.0(-0.5)83.4(-0.1)83.7(+0.2%)
FLOPs(G)15.410.48(-32%)10.48(-32%)11.51(-25%)12.97(-16%)
Throughput(img/s)415558(+35%)551(+33%)521(+26%)489(+19%)
", + "bbox": [ + 150, + 324, + 820, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "former, so the reduction of FLOPs on Swin models are smaller than on DeiT models. When 9 semantic tokens are used in each window, STViT-Swin models can reduce $25\\%$ FLOPs with negligible accuracy loss on all the model sizes. If the number of tokens is reduced to 4 in each window (16 in total), a significant accuracy drop will occur, which indicates that local vision transformers need more semantic tokens than global vision transformers. We can move the STGM towards shallow layers to attain a better complexity/accuracy trade-off. In Table 2, STGM is moved by one transformer layer on STViT-Swin-T and by two layers on STViT-Swin-S and STViT-Swin-B to save over $30\\%$ FLOPs with only about $0.3\\%$ accuracy drops (column of \"Move STGM\").", + "bbox": [ + 75, + 575, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "STViT-R equipped with recovery modules is designed for downstream tasks, while it also can perform image classification. The corresponding results are reported in Table 3. On both Swin-S and Swin-B, STViT-R can save $33\\%$ FLOPs with $0.3\\%$ accuracy drop. The hyper-parameters we used in STGM are as same as STViT-Swin.", + "bbox": [ + 75, + 790, + 468, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These results demonstrate that our method achieves both", + "bbox": [ + 96, + 885, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e1a7b7842dae39f4ac396fd271f45876b45d6a38410bd46c499503b36b0344a9.jpg", + "table_caption": [ + "Table 2. Applying STViT to Swin-T, Swin-S, and Swin-B. The top-1 accuracy, complexity in FLOPs, and throughput are reported for different numbers of semantic tokens in each window. Base indicates the corresponding original Swin model. Move STGM indicates changing the default position of STGM." + ], + "table_footnote": [], + "table_body": "
ModelTop-1 Acc(%)FLOPs(G)Throughput
STViT-R-Swin-S82.7(-0.3)5.83(-33%)717(+30%)
STViT-R-Swin-B83.2(-0.3)10.26(-33%)539(+30%)
", + "bbox": [ + 503, + 574, + 890, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. STViT-R is evaluated on Swin-S and Swin-B on ImageNet. The top-1 accuracy, complexity in FLOPs, and throughput are reported.", + "bbox": [ + 498, + 636, + 890, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "effectiveness and efficiency by employing a few semantic tokens to replace original image tokens. Our method reveals that constructing the tokens with high-level semantic representation is more important than maintaining structured tokens in ViTs. As reflected by the throughput, our method does not have overhead of memory or deployment. Compared to the total parameters, the additional parameters introduced by intra and inter-window spatial pooling is negligible (less than $0.05\\%$ ), so we do not show them.", + "bbox": [ + 496, + 688, + 890, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with existing token sparsification methods. In Table 4, we compare STViT with the state-of-the-art token sparsification methods on DeiT-S and DeiT-B. Due to different base models used by different methods, we", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6200", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/71f86540cfa691750bd3beb9e4abba72e9b396337d2aa719c673c0f80c6f0f58.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTop-1 AccFLOPs(G)△
DeiT-S
DynamicViT [31]79.32.9(-37%)-0.5
IA-RED2[30]79.13.2(-30%)-0.7
PS-ViT [34]79.42.6(-43%)-0.4
TokenLearner* [32]76.11.9(-44%)-1.8
DGE* [33]79.73.1 (-49%)-0.6
A-ViT* [47]78.63.6 (-39%)-0.3
Evo-ViT [45]79.43.0(-35%)-0.4
EViT [24]78.52.3(-50%)-1.3
STViT(Ours)79.81.91(-58%)-0.0
DeiT-B
IA-RED2[30]80.311.8(-33%)-1.5
DynamicViT [31]81.311.2(-36%)-0.5
PS-ViT [34]81.59.8(-44%)-0.3
TokenLearner* [32]83.728.7(-48%)-1.1
Evo-ViT [45]81.310.2(-33%)-0.5
EViT [24]80.08.7(-51%)-1.8
STViT(Ours)81.87.31(-58%)-0.0
", + "bbox": [ + 96, + 88, + 452, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparisons with the state-of-the-art token sparsification methods on DeiT-S and DeiT-B. $\\triangle$ shows the accuracy difference between each model and its base model. $*$ : their base models are not standard DeiT models.", + "bbox": [ + 75, + 393, + 470, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "adopt accuracy difference between each model and its base model $\\Delta$ to evaluate them for fair. Results indicate that our method achieves the lowest accuracy drop $\\Delta$ with the highest FLOPs reduction, outperforming all the state-of-the-art methods in both accuracy and efficiency significantly.", + "bbox": [ + 75, + 462, + 468, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Video recognition", + "text_level": 1, + "bbox": [ + 76, + 551, + 250, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setting. For video recognition, we apply our STViT to Video Swin [27]. All the models are pre-trained on ImageNet-1K and trained on Kinetics-400 [7]. We generate semantic tokens from each frame as illustrated in Section 3.2. The initialization from pre-trained models and other implementation details are as same as Video Swin [27].", + "bbox": [ + 75, + 577, + 468, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. The results are presented in Table 6. On Swin-T and Swin-S, STViT-Swin can save about $27\\%$ FLOPs with $0.3\\%$ accuracy drop, which shows that our method works on video recognition.", + "bbox": [ + 75, + 688, + 470, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Applications in object detection and instance segmentation", + "text_level": 1, + "bbox": [ + 76, + 762, + 468, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. Experiments of object detection and instance segmentation are conducted on COCO 2017 [25]. We evaluate STViT-R with Swin in Cascade Mask R-CNN [4, 17] detection frameworks. The $w_{s}$ is set to 3. The backbone models are pre-trained on ImageNet-1K and the pre-trained results are shown in Table 3. All the other settings follow Swin [26].", + "bbox": [ + 75, + 803, + 470, + 910 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c993eb7095c4337efbb350f9418e71552b514133d0cbde4048357a4d48cac1b8.jpg", + "image_caption": [ + "Figure 3. Visualization example of attention maps in the first attention layer of STGM." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 893, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to Swin Transformer. The performance of STViT-R-Swin using the Cascade Mask R-CNN framework with $3 \\times$ schedule is shown in Table 5. Our method achieves better performance with more than $30\\%$ FLOPs reduction for backbone on bath object detection and instance segmentation. This validates that the recovery module and dumbbell unit can restore detailed spatial information, and the global context information integrated from the semantic tokens significantly benefits object detection. Ignoring spatial structure in the intermediate layers does not affect the object detection task, which is a meaningful fact to help design efficient object detection frameworks. Another interesting finding is that our method has a remarkable improvement for small object detection which is a challenging problem in the detection community. For instance, our STViT-R-Swin-S outperforms Swin-S by $1.5\\%$ on $AP_{s}^{b}$ .", + "bbox": [ + 496, + 195, + 893, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 448, + 650, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All the following ablation study of STViT and STViT-R are conducted on the DeiT-S and Swin-S, respectively.", + "bbox": [ + 498, + 470, + 890, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initialization analysis. The semantic tokens are the cluster centers recovered by attention layers. The initialization of cluster centers induces the representation of semantic tokens. Spatial and global initialization are adopted in Section 3.1 to guide the semantic tokens to integrate local and global semantic information separately. We compare different initialization components in Table 7. When performing single global initialization $(3^{rd}$ row), we replace the spatial initial cluster centers with global initial cluster centers in the first transformer layer. The accuracy of using a single initialization method is far lower than using both, which shows the effectiveness of our initialization strategy. Our global initial cluster centers look similar to learned positional encoding since both use random initialization. To verify their distinction, we experiment with a real learned positional encoding by additionally adding the global initial cluster centers to keys, which causes $0.1\\%$ accuracy drop (the last row of Table 7). Therefore, our global initialization is different from learned positional encoding.", + "bbox": [ + 496, + 507, + 892, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We visualize the attention maps of the first attention layer in the STGM in Figure 3. Because the queries of this layer are spatial initial cluster centers, these attention maps visualize the local semantic information integration by attention. The attention layer groups the semantic information according to the position of initial cluster centers, which ensures to extract fine-grained semantic information", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6201", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/81d89a9ef91ac87562c9692191629aac42946d451a88bbf06115f96195b87213.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
APbAPb50APb75APsbAPmAPm50APm75APsmFLOPs(G)
Swin-S51.870.456.335.244.767.948.528.8194
STViT-R-Swin-S51.870.656.136.744.767.848.629.0134(-31%)
Swin-B51.970.956.535.445.068.448.728.9343
STViT-R-Swin-B52.270.856.836.545.268.349.129.5233(-32%)
", + "bbox": [ + 181, + 88, + 787, + 180 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/75f0658c28bc6899d65a93b2c70b67abcb4cf6c23286c6cfb590552cb89684fe.jpg", + "table_caption": [ + "Table 5. Results on COCO object detection and instance segmentation under Cascade Mask R-CNN with $3 \\times$ schedule. The FLOPs are measured for backbones." + ], + "table_footnote": [], + "table_body": "
ModelTop-1 Acc(%)FLOPs(G)Speed
Swin-T78.888779
STViT-Swin-T78.5(-0.3)64.4(-27%)975(+25%)
Swin-S80.6166456
STViT-Swin-S80.3(-0.3)120.5(-27%)572(+25%)
", + "bbox": [ + 89, + 223, + 454, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cd6fd3f9889086f0a8a9178693dd28ffc3bff9527c4bf2d9a7b8765381c4056f.jpg", + "table_caption": [ + "Table 6. Applying STViT to Video Swin (Swin-T and Swin-S) on Kinetics-400. All the models are pre-trained on ImageNet-1K. The views are $4\\times 3$ . The top-1 accuracy and complexity in FLOPs are reported. Speed is evaluated by FPS." + ], + "table_footnote": [], + "table_body": "
SpatialGlobalLearnedTop-1 Acc(%)
✓79.4
✓78.7
✓✓79.8
✓✓79.7
", + "bbox": [ + 140, + 369, + 403, + 460 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/06f0ef9930a366040aef4c178ffecb4c13f40a310e11e5ba81d1bd711abf88e0.jpg", + "table_caption": [ + "Table 7. Accuracy with different initialization of STViT. Spatial, Global, and Learned indicate spatial initialization, global initialization, and learned positional encoding methods, respectively." + ], + "table_footnote": [], + "table_body": "
No. of transformers234
Top-1 Acc(%)79.879.579.6
FLOPs(G)1.911.972.03
", + "bbox": [ + 133, + 516, + 411, + 574 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/80a2b53a35d8ba89e2a5a9b59f908f7e79a2cd7d297760627e1d4eacaa5a7e7f.jpg", + "table_caption": [ + "Table 8. Performance evaluation on different numbers of transformer layers in STGM. Keeping the base module containing four transformer layers unchanged." + ], + "table_footnote": [], + "table_body": "
STViT-Rw/o DUReusing ST
APb51.851.451.6
APm44.744.444.5
", + "bbox": [ + 130, + 630, + 413, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation study on STViT-R w/o dumbbell units (w/o $DU$ ) and reusing semantic tokens (Reusing ST).", + "bbox": [ + 75, + 696, + 468, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and keep the difference among semantic tokens. The attention map in the second attention layer is visualized in Supplementary A.5, which reveals that the network fixedly selects particle semantic tokens to represent global semantic information. We also show the semantic representation of image tokens in the same transformer layer on DeiT in Supplementary A.5. Compared to original image tokens, our semantic tokens in Figure 3 show more high-level semantic information.", + "bbox": [ + 75, + 729, + 468, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of transformer layers in STGM. Two transformer layers are adopted in the STGM by default. The ef", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fcts of employing different numbers of transformer layers are shown in Table 8. The extra layers are from the ones behind STGM to keep the total number of layers unchanged. More transformer layers do not bring improvement.", + "bbox": [ + 496, + 233, + 890, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of the dumbbell unit. To verify the effectiveness of our dumbbell unit, we experiment STViT-R without dumbbell units, i.e., STViT equipped with only the recovery module. We employ $6^{th}$ and $7^{th}$ transformer layers to construct STGM and the last transformer layers to construct the recovery module in Stage 3. The FLOPs is the same as the full-model STViT-R for fair comparison. The results on the COCO are reported in Table 9. Inferior results demonstrate the effectiveness of the dumbbell unit.", + "bbox": [ + 496, + 305, + 890, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reusing semantic tokens in dumbbell units. Semantic tokens are generated in each dumbbell unit. If they are produced only once in the first dumbbell unit and reused as initial cluster centers in the subsequent dumbbell units, the result is shown in Table 9 with a slight performance drop.", + "bbox": [ + 496, + 452, + 890, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 537, + 617, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a simple and effective token sparsification method, semantic token vision transformer (STViT). Our method utilizes the clustering property of self-attention to generate a few semantic tokens with high-level information representation to replace the redundant image/video tokens, which can be applied in both global and local vision transformers. By simply configuring the recovery module, our method can be successfully applied to downstream tasks. Extensive experiments demonstrate that our method achieves better accuracy along with less inference time in most cases. The success in downstream tasks significantly boosts the development of token sparsification methods. We hope that this work can inspire more future research to pay much attention to high-level semantic representation in ViTs.", + "bbox": [ + 496, + 560, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 790, + 660, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, and Mike Zheng Shou's Start-Up Grant from NUS. Computation was partially performed on resources of the National Supercomputing Centre, Singapore. Shuning was supported by Alibaba Research Intern Program.", + "bbox": [ + 496, + 810, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6202", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Luvcic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2", + "[2] Song Bai, Philip Torr, et al. Visual parser: Representing part-whole hierarchies with transformers. arXiv preprint arXiv:2107.05790, 2021. 3", + "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 2", + "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2", + "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7", + "[8] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34, 2021. 3", + "[9] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 2", + "[10] Yunpeng Chen, Marcus Rohrbach, Zhicheng Yan, Yan Shuicheng, Jiashi Feng, and Yannis Kalantidis. Graph-based global reasoning networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-442, 2019. 3", + "[11] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1601-1610, 2021. 2", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[14] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Bain-" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. arXiv preprint arXiv:2107.00652, 2021. 2", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1, 2", + "[16] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3", + "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7", + "[18] Zilong Huang, Youcheng Ben, Guozhong Luo, Pei Cheng, Gang Yu, and Bin Fu. Shuffle transformer: Rethinking spatial shuffle for vision transformer. arXiv preprint arXiv:2106.03650, 2021. 2", + "[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3", + "[20] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 2, 5", + "[21] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Bin Ren, Minghai Qin, Hao Tang, and Yanzhi Wang. Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890, 2021. 3", + "[22] Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient self-supervised vision transformers for representation learning. arXiv preprint arXiv:2106.09785, 2021. 2", + "[23] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019. 3", + "[24] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In International Conference on Learning Representations, 2022. 1, 2, 3, 7", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7", + "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + ], + "bbox": [ + 503, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6203", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 4, 5, 7", + "[27] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 7", + "[28] Xuezhe Ma, Xiang Kong, Sinong Wang, Chunting Zhou, Jonathan May, Hao Ma, and Luke Zettlemoyer. Luna: Linear unified nested attention. Advances in Neural Information Processing Systems, 34:2441-2453, 2021. 3", + "[29] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. arXiv preprint arXiv:2111.15668, 2021. 3", + "[30] Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. Ia-red 2: Interpretability-aware redundancy reduction for vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 3, 7", + "[31] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34, 2021. 3, 7", + "[32] Michael S. Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 7", + "[33] Lin Song, Songyang Zhang, Songtao Liu, Zeming Li, Xuming He, Hongbin Sun, Jian Sun, and Nanning Zheng. Dynamic grained encoder for vision transformers. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 1, 3, 7", + "[34] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. arXiv preprint arXiv:2106.02852, 2021. 3, 7", + "[35] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 2, 5", + "[36] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 1, 2", + "[37] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2", + "[38] Pichao Wang, Xue Wang, Hao Luo, Jingkai Zhou, Zhipeng Zhou, Fan Wang, Hao Li, and Rong Jin. Scaled relu matters for training vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022. 1", + "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao." + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 2", + "[40] Wenxiao Wang, Lu Yao, Long Chen, Binbin Lin, Deng Cai, Xiaofei He, and Wei Liu. Crossformer: A versatile vision transformer hinging on cross-scale attention. arXiv preprint arXiv:2108.00154, 2021. 2", + "[41] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. Advances in Neural Information Processing Systems, 34, 2021. 3", + "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8741-8750, 2021. 2", + "[43] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22-31, 2021. 1, 2", + "[44] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9981–9990, October 2021. 2", + "[45] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1, 2, 3, 7", + "[46] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers, 2021. 2", + "[47] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 3, 7", + "[48] Tan Yu, Gangming Zhao, Ping Li, and Yizhou Yu. Boat: Bilateral local attention vision transformer. arXiv preprint arXiv:2201.13027, 2022. 2", + "[49] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 558-567, October 2021. 2", + "[50] David Junhao Zhang, Kunchang Li, Yali Wang, Yunpeng Chen, Shashwat Chandra, Yu Qiao, Luoqi Liu, and Mike Zheng Shou. Morph MLP: An efficient mlp-like backbone for spatial-temporal representation learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 230-248. Springer, 2022. 2", + "[51] Songyang Zhang, Xuming He, and Shipeng Yan. Latent-gnn: Learning efficient non-local relations for visual recog" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6204", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nition. In International Conference on Machine Learning, pages 7374-7383. PMLR, 2019. 3", + "[52] Minghang Zheng, Peng Gao, Renrui Zhang, Kunchang Li, Xiaogang Wang, Hongsheng Li, and Hao Dong. End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315, 2020. 2", + "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2", + "[54] Jingkai Zhou, Pichao Wang, Fan Wang, Qiong Liu, Hao Li, and Rong Jin. Elsa: Enhanced local self-attention for vision transformer. arXiv preprint arXiv:2112.12786, 2021. 2", + "[55] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 2", + "[56] Zhuofan Zong, Kunchang Li, Guanglu Song, Yali Wang, Yu Qiao, Biao Leng, and Yu Liu. Self-slimmed vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 432-448. Springer, 2022. 1, 2, 3" + ], + "bbox": [ + 78, + 90, + 468, + 444 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6205", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_model.json b/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cad913799379f6cb775f08c9bcd443d026491059 --- /dev/null +++ b/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.851, + 0.154 + ], + "angle": 0, + "content": "Making Vision Transformers Efficient from A Token Sparsification View" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.179, + 0.825, + 0.236 + ], + "angle": 0, + "content": "Shuning Chang\\(^{1*}\\) Pichao Wang\\(^{2\\dagger\\dagger}\\) Ming Lin\\(^{2\\dagger}\\) Fan Wang\\(^{2}\\) David Junhao Zhang\\(^{1}\\) Rong Jin\\(^{2}\\) Mike Zheng Shou\\(^{1\\dagger}\\) \n\\(^{1}\\)Show Lab, National University of Singapore \\(^{2}\\)Alibaba Group" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.286 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.301, + 0.474, + 0.771 + ], + "angle": 0, + "content": "The quadratic computational complexity to the number of tokens limits the practical applications of Vision Transformers (ViTs). Several works propose to prune redundant tokens to achieve efficient ViTs. However, these methods generally suffer from (i) dramatic accuracy drops, (ii) application difficulty in the local vision transformer, and (iii) non-general-purpose networks for downstream tasks. In this work, we propose a novel Semantic Token ViT (STViT), for efficient global and local vision transformers, which can also be revised to serve as backbone for downstream tasks. The semantic tokens represent cluster centers, and they are initialized by pooling image tokens in space and recovered by attention, which can adaptively represent global or local semantic information. Due to the cluster properties, a few semantic tokens can attain the same effect as vast image tokens, for both global and local vision transformers. For instance, only 16 semantic tokens on DeiT-(Tiny,Small,Base) can achieve the same accuracy with more than \\(100\\%\\) inference speed improvement and nearly \\(60\\%\\) FLOPs reduction; on Swin-(Tiny,Small,Base), we can employ 16 semantic tokens in each window to further speed it up by around \\(20\\%\\) with slight accuracy increase. Besides great success in image classification, we also extend our method to video recognition. In addition, we design a STViT-R(covery) network to restore the detailed spatial information based on the STViT, making it work for downstream tasks, which is powerless for previous token sparsification methods. Experiments demonstrate that our method can achieve competitive results compared to the original networks in object detection and instance segmentation, with over \\(30\\%\\) FLOPs reduction for backbone." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.798, + 0.21, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.47, + 0.855 + ], + "angle": 0, + "content": "In contrast to standard Convolutional Neural Networks (CNNs) approaches which process images pixel-by-pixel," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.895, + 0.649 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) [15, 26, 35, 36, 43] treat an image as a sequence of patch/image tokens, and have shown promising performance in prevalent visual recognition scenarios. However, these superior performances do not come for free: the quadratic computational complexity to the number of image tokens limits their application in practice. Previous works [33, 56] have illustrated the large amount of redundancy in the image tokens and also shown the effect of filtering out unimportant tokens normally according to predefined scoring mechanism. However, these methods face the following challenges. Firstly, the predefined scoring mechanisms for filtering are generally imprecise. In Figure 1, on the left we visualize the class token values in different layers which are commonly used to score the token importance [16, 24, 45]. Different layers have different value distributions, thus using these imprecise scores for filtering would lead to unsatisfactory performance. For example, EViT [24] has an accuracy drop of \\(1.3\\%\\) when saving \\(50\\%\\) FLOPs on DeiT-S [35]. Secondly, the remaining tokens do not distribute evenly in space any more, making them hard to work in local vision transformers1. Finally, large-scale token pruning tremendously damages the spatial structure and positional information, and causes difficulties when applied to downstream tasks, which they do not propose a solution to deal with." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.895, + 0.848 + ], + "angle": 0, + "content": "To solve these problems, we propose Semantic Token ViT (STViT), for efficient global and local vision transformers, which also can be revised to serve as backbone for downstream tasks. The proposed approach is based on the following observations: (i) unlike local CNNs which learn spatial structure of images, vision transformer discretizes feature map as tokens for global feature exploration, relieving the requirements for maintaining the whole image structure and information; (ii) discrete tokens are more beneficial for optimization [38]; (iii) in Figure 1, on the right shows the attention maps in different transformer layers, and there are only several vertical lines in the deep layers, which means that only a few tokens with global se" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.895, + 0.901 + ], + "angle": 0, + "content": "1In this paper, we define the vision transformer with global self-attention (like DeiT) as global vision transformer and the vision transformer with local self-attention (like Swin) as local vision transformer." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.369, + 0.876 + ], + "angle": 0, + "content": "*Work done during an internship at Alibaba Group." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.259, + 0.888 + ], + "angle": 0, + "content": "Equal corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.434, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\)Work done at Alibaba Group, and now affiliated with Amazon." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.863, + 0.434, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "6195" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.09, + 0.443, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.091, + 0.844, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.173, + 0.893, + 0.203 + ], + "angle": 0, + "content": "Figure 1. Left: the attention values of class tokens (normalized and reshaped in image shape) in different self-attention layers. Right: the attention maps in different self-attention layers. Zoom-in for better visibility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.471, + 0.293 + ], + "angle": 0, + "content": "matic information matter. Thus, we argue that it is not necessary to maintain massive structured tokens for ViTs, especially in the deep layers. Employing a few discrete tokens with high-level semantic information can potentially achieve both high performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.471, + 0.779 + ], + "angle": 0, + "content": "In STViT, the semantic tokens represent the cluster centers, and the number of them is far less than the original image tokens, significantly reducing the computational cost. Inspired by the fact that multi-head attention can conduct the cluster center recovery (Supplementary A.7), we only employ the off-the-shelf self-attention to generate the semantic tokens. Specifically, the first few transformer layers are kept unchanged to obtain the image tokens with low-level features. The image tokens are then fed into our semantic token generation module (STGM) consisting of at least two transformer layers to generate semantic tokens. In each self-attention layer, the semantic tokens are input as queries, and the image tokens are fed as keys and values. The semantic tokens dynamically aggregate image tokens through the attention layers to recover cluster centers. In the first attention layer, the semantic tokens are initialized by an intra and inter-window spatial pooling which takes into account incorporating semantic information in each window and maximizing distance between adjacent windows. Thanks to this spatial initialization, the semantic tokens mainly incorporate local semantic information and achieve discrete and uniform distribution in space. In the following attention layer, besides further clustering, the semantic tokens are equipped with global cluster centers, and the network can adaptively select partial semantic tokens to focus on global semantic information. After the STGM, the original image tokens are discarded, and only semantic tokens are kept for the subsequent transformer layers. Because the generation of semantic tokens is flexible and space-aware, our method can be plugged into both global and local vision transformers. The semantic tokens can be produced in each window for the local vision transformer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Another property of STViT is its capability to serve as a backbone for downstream tasks, such as object detection and instance segmentation. Discussions have been missing in previous methods [16, 24, 32, 45, 56] about how to use them in downstream task under the massive loss of spatial information during the token sparsification process, which actually seriously impedes the application of their method. Instead, we design a novel STViT-R network based" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.893, + 0.292 + ], + "angle": 0, + "content": "on STViT where a recovery module and dumbbell unit are adopted to periodically restore the full resolution feature map while the intermediate transformer layers continue to use semantic tokens to save computation cost, making our method work in downstream task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.895, + 0.583 + ], + "angle": 0, + "content": "The effectiveness of the proposed method is validated via a comprehensive empirical study on image and video ViT models. Only 16 semantic tokens on DeiT-(Tiny, Small, Base) achieve nearly \\(50\\%\\) inference time reduction without any accuracy degradation; on Swin-(Tiny, Small, Base), we also improve the inference throughput by nearly \\(20\\%\\) with slight accuracy increase. Moreover, the proposed STViT-R achieves promising results on object detection and instance segmentation. To the best of our knowledge, this is one of first works to apply the token sparsification algorithm in local vision transformers, and use the ViTs as backbones in downstream tasks after large-scale token pruning. Our findings in ViTs uncover that maintaining the full-size feature map is unnecessary, and a few tokens with high-level semantic representations can achieve both high performance and efficiency. Thanks to its simplicity and general-purpose ability, our method can also serve as a new efficient ViT baseline architecture and a starting point for further research from the token sparsification perspective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.638, + 0.617 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Vision transformers. Vision Transformer (ViT) [15] first introduces a pure Transformer backbone for image classification. ViT variants further inspire the applications of transformer to various vision tasks beyond image/video classification [1, 3, 20, 35, 36, 39, 43, 44, 49, 50], such as object detection [5, 12, 52, 55], semantic segmentation [37, 42, 53], and self-supervised learning [6, 9, 22]. Vanilla transformers have high computational and memory costs because the multi-head self-attention has quadratic computational complexity to the number of image tokens. Recently, various efficient ViTs have been proposed to alleviate this issue. The existing methods mainly focus on reducing the complexity of self-attention or reducing the number of tokens. Swin Transformer [26] adopts local self-attention, i.e., attending neighboring tokens within a constant window size, achieving a linear computational complexity in the self-attention with high performance. Many subsequent works [11, 14, 18, 40, 46, 48, 54] follow the local" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6196" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "self-attention design to develop variants. Token sparsification [8, 16, 21, 24, 29–34, 41, 45, 47, 56] also attracts increasing attention." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.162, + 0.473, + 0.524 + ], + "angle": 0, + "content": "Token sparsification. Token sparsification methods can be mainly categorized into hard pruning [8,16,21,24,29-31, 33,34,45,47] and soft pruning [32,56]. Hard pruning methods filter out some unimportant tokens according to a predefined scoring mechanism. DynamicViT [31], SPViT [21], and AdaViT [29] introduce additional prediction networks to score the tokens. Evo-ViT [45], ATS [16], and EViT [24] utilize the values of class tokens to evaluate the importance of tokens. However, it is difficult to achieve precise scoring as shown in the left of Figure 1. Therefore, they usually suffer from a significant accuracy drop. For instance, EViT [24] has an accuracy drop of \\(1.3\\%\\) when saving \\(50\\%\\) FLOPs on DeiT-S. Soft pruning methods generate new tokens from image tokens by importing additional attention networks. TokenLearner [32] also argue for a few tokens to replace image tokens. However, its price is a \\(1.8\\%\\) accuracy drop when reducing \\(44\\%\\) FLOPs, which is far inferior to concurrent works. Besides performance degradation, previous methods also have the following disadvantages. First, whether or how to extend the methods to local vision transformers remains unexplored. Second, it has not been discussed about how to serve the downstream tasks like object detection and instance segmentation after the tokens are pruned." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.661 + ], + "angle": 0, + "content": "In our method, we apply the off-the-shelf transformer layers to reduce token number. [2, 10, 19, 23, 28, 51] adopt similar approaches to achieve efficient non-local relationships. Our method is different from them as below: (i) our method extracts local semantic information instead of non-local relationships; (ii) the semantic tokens are a few cluster centers, which can replace the massive image tokens to achieve image classification; (iii) our method specializes in pruning tokens." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.674, + 0.169, + 0.69 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.469, + 0.788 + ], + "angle": 0, + "content": "The proposed STViT is presented in this section, which aims to construct an efficient and high-performance ViT. STViT is first introduced in Section 3.1, followed by how to apply STViT in the local vision transformer in Section 3.2. Based on STViT, STViT-R is developed to restore the spatial resolution for downstream tasks in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.166, + 0.815 + ], + "angle": 0, + "content": "3.1.STViT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Overall architecture. An overview of STViT architecture is presented in Figure 2a. The patch embedding layer and shallow transformer layers are kept unchanged as a base module in our method. The base module copes with all the image tokens \\( X \\in \\mathbb{R}^{N_i \\times C} \\) to extract low-level features," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "where \\(N_{i}\\) is the number of image tokens and \\(C\\) is the number of channels. The image tokens are fed into the semantic token generation module (STGM) to generate \\(N_{s}\\) semantic tokens \\(S \\in \\mathbb{R}^{N_s \\times C}\\). After the STGM, the image tokens \\(X\\) can be discarded, and only semantic tokens \\(S\\) with high-level semantic information are used in all the subsequent transformers. Due to \\(N_{s} \\ll N_{i}\\), our method can significantly reduce the computational cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.341 + ], + "angle": 0, + "content": "Semantic token generation module (STGM). The whole image is represented by a few tokens with high-level semantic information through clustering. Inspired by the fact that self-attention can conduct cluster center recovery (Supplementary A.7), we adopt the off-the-shelf self-attention layers to produce the semantic tokens. The STGM consists of at least two transformer layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.893, + 0.644 + ], + "angle": 0, + "content": "The initial cluster centers \\( P \\in \\mathbb{R}^{N_s \\times C} \\) are generated by an spatial pooling which pools the image tokens into fixed \\( w_{s} \\times w_{s} \\) tokens, with \\( N_{s} = w_{s} \\times w_{s} \\). \\( w_{s} \\) is generally set as 4. The spatial pooling can be achieved by a non-parameterized adaptive spatial pooling or a super lightweight network with higher performance, intra and inter-window spatial pooling. The intentions of spatial pooling initialization are three folds. First, the initial cluster centers can distribute uniformly in space, making the generated semantic tokens more discrete and preventing the semantic tokens from collapsing to one point in the following layers. Second, the semantic tokens can be forced to represent more local and distinguished features. Finally, the representation of semantic tokens is associated with the specific spatial locations, which is the basis to allow our method to be applied in local self-attention and downstream tasks. The initial cluster centers \\( P \\) then dynamically integrate the image tokens \\( X \\) according to semantic information by attention mechanism. In the first transformer layer, the processing of the generation of semantic tokens can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.648, + 0.891, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\hat {S} ^ {1} = M H A (P, X, X) + P, \\quad S ^ {1} = F F N (\\hat {S} ^ {1}) + \\hat {S} ^ {1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.864 + ], + "angle": 0, + "content": "where \\( MHA \\) and \\( FFN \\) are short for multi-head attention layer and feed-forward network, respectively, and the triplet input of \\( MHA \\) are queries, keys, and values in turn. All the norm layers are omitted in all the equations for brevity. The initial cluster centers are produced in a window by an adaptive spatial pooling layer or an intra and inter-window spatial pooling, while the semantic tokens are generated from a global receptive field by a dynamic attention layer to ensure that they can extract high-level semantic representation. In order to further strengthen the clustering effect, we use the second transformer layer to repeat the clustering operation and guide the semantic tokens to extract global information. In this transformer, the semantic tokens are updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.867, + 0.891, + 0.89 + ], + "angle": 0, + "content": "\\[\n\\hat {S} ^ {2} = M H A \\left(S ^ {1} + G, \\operatorname {C o n c a t} \\left(S ^ {1}, X\\right), \\operatorname {C o n c a t} \\left(S ^ {1}, X\\right)\\right) + S ^ {1}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.888, + 0.758, + 0.902 + ], + "angle": 0, + "content": "\\[\nS ^ {2} = F F N (\\hat {S} ^ {2}) + \\hat {S} ^ {2},\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6197" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.088, + 0.485, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.226, + 0.332, + 0.237 + ], + "angle": 0, + "content": "(a) STViT" + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.093, + 0.829, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.226, + 0.743, + 0.237 + ], + "angle": 0, + "content": "(b) STViT-R" + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.244, + 0.651, + 0.258 + ], + "angle": 0, + "content": "Figure 2. The architectures of our STViT and STViT-R." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.47, + 0.357 + ], + "angle": 0, + "content": "where \\( G \\in \\mathbb{R}^{N_s \\times C} \\) is the global cluster centers initialized by Gaussian noise and \\( \\text{Concat}(\\cdot) \\) is a concatenation operation. The global cluster centers \\( G \\) are responsible for global semantic information extraction like the class token. Although \\( S^1 \\) and \\( G \\) are summed together as the queries, they can be decoupled in the attention computation as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.362, + 0.47, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} A _ {s} = \\left(S ^ {1} \\cdot W _ {q}\\right) \\cdot \\left(\\left(S ^ {1} + X\\right) \\cdot W _ {k}\\right), \\quad A _ {g} = \\left(G \\cdot W _ {q}\\right) \\cdot \\left(\\left(S ^ {1} + X\\right) \\cdot W _ {k}\\right), \\tag {3} \\\\ A = \\operatorname {S o f t m a x} (A _ {s} + A _ {g}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.471, + 0.641 + ], + "angle": 0, + "content": "where \\( W_{q} \\) and \\( W_{k} \\) are the linear projection weights of queries and keys. We can see that \\( S^{1} \\) and \\( G \\) integrate the keys to generate \\( A_{s} \\) and \\( A_{g} \\) independently and just share a Softmax operation to produce the final attention map \\( A \\). Therefore, besides spatial semantic information, the semantic tokens also incorporate global semantic information with a negligible additional overhead. Though very similar, the global cluster centers are actually different from the learned positional encoding. We do not add the global cluster centers to keys. Moreover, it will be shown in the experiments that inserting the actual learned positional encoding will cause an accuracy drop. The number of transformer layers in STGM is flexible. More transformer layers can be associated for clustering. Two transformer layers are employed by default, i.e., \\( S^{2} \\) is the output of the STGM. Note that the image tokens are not updated in the STGM." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.668, + 0.47, + 0.883 + ], + "angle": 0, + "content": "Intra and inter-window spatial pooling. Give \\( H \\times W \\) feature map \\( X \\), we generate \\( N_{s}(w_{s} \\times w_{s}) \\) initial cluster centers. We uniformly split the \\( X \\) into \\( w_{s} \\times w_{s} \\) windows \\( [X_{w}^{i}]_{i=1}^{N_{s}} \\) with size \\( \\frac{H}{w_{s}} \\times \\frac{W}{w_{s}} \\) and each window generates one initial cluster center. To represent abundant semantic information, we take into account intra and interwindow relations, i.e., integrating important tokens in the window and maximizing the distance among initial cluster centers in different windows. Specifically, we formulate an intra-window function \\( f_{intra} \\) to produce a mask, \\( M_{i} = f_{intra}(X_{w}^{i}) \\), which projects the input window from \\( \\mathbb{R}\\frac{H}{w_{s}} \\times \\frac{W}{w_{s}} \\times C \\rightarrow \\mathbb{R}\\frac{H}{w_{s}} \\times \\frac{W}{w_{s}} \\). The idea is to let the intra-window function \\( f_{intra} \\) adaptively select a combination of informative tokens in \\( X_{w}^{i} \\), which is implemented by" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.887, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\nM _ {i} = \\operatorname {C o n v} \\left(\\operatorname {G e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\operatorname {D e p t h w i s e} \\operatorname {C o n v} \\left(X _ {w} ^ {i}\\right)\\right)\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.266, + 0.895, + 0.461 + ], + "angle": 0, + "content": "Then, we compute each integrated token \\(\\hat{P}_i =\\) Softmax \\((M_i)\\cdot X_w^i\\) from each window, and arrange \\([\\hat{P}_i]_{i = 1}^{N_s}\\) by spatial structure to form the 2D tensor \\(\\hat{P}\\in \\mathbb{R}^{w_s\\times w_s\\times C}\\) We adopt an inter-window function \\(f_{inter}\\) to compute the inter-window relations and generate the offset, \\(O =\\) \\(f_{inter}(\\hat{P})\\) ,to revise the mask \\(M\\) . The implementation of \\(f_{inter}\\) is similar to Eq. 4, except the mapping input from \\(\\mathbb{R}^{w_s\\times w_s\\times C}\\rightarrow \\mathbb{R}^{w_s\\times w_s\\times \\frac{HW}{w_s^2}}\\) , where \\(\\frac{HW}{w_s^2}\\) is the number of tokens in each window. For each window, the corresponding \\(O_{i}\\in \\mathbb{R}^{\\frac{HW}{w_{s}^{2}}}\\) is sliced from \\(O\\) and reshaped to \\(\\mathbb{R}^{\\frac{H}{w_s}\\times \\frac{W}{w_s}}\\) The \\(O_{i}\\) is used to revise \\(M_{i}\\) . The final initial cluster center \\(P_{i}\\) is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.464, + 0.892, + 0.48 + ], + "angle": 0, + "content": "\\[\nP _ {i} = \\operatorname {S o f t m a x} \\left(M _ {i} + O _ {i}\\right) \\cdot X _ {w}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.484, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Our \\( f_{\\text{intra}} \\) and \\( f_{\\text{intra}} \\) are super lightweight and the introducing parameters can be negligible. For example, on DeiT, the parameters only increase by \\( 0.05\\% \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.807, + 0.558 + ], + "angle": 0, + "content": "3.2. STViT in local vision transformers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Local self-attention has been widely used in current ViT models to balance efficiency and accuracy. As the generation of semantic tokens in STViT is flexible in space, it can be naturally applied in local self-attention. Suppose each local self-attention layer contains \\( N_{w} \\) windows with size \\( w \\times w \\), we initialize \\( w_{s} \\times w_{s} \\) cluster centers in each window by our intra and inter-window spatial pooling. The total number of semantic tokens \\( N_{s} \\) would be \\( w_{s} \\times w_{s} \\times N_{w} \\). \\( w \\) and \\( w_{s} \\) are set as 7 and 3 by default separately. As a result, our method compresses more than \\( 80\\% \\) image tokens in local self-attention. In the STGM, although initial cluster centers are from \\( w \\times w \\) windows, we use larger windows with size \\( w_{k} \\times w_{k} \\) to obtain keys and values in Eq. 1 and Eq. 2 to mitigate the effect of limited window size. Other operations in STGM are kept the same as Section 3.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the local ViT models, each local transformer layer is normally followed by a cross-window connection layer, such as a shift window transformer layer following a local transformer layer on Swin Transformer [26]. In our method, the attention is computed within \\( w_{s} \\times w_{s} \\) window in the local self-attention layer, and the cross-window connection can be achieved by computing self-attention in a larger-size" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "\\((e.g., 4 \\times w_{s})\\) sliding window because of the rare number of tokens in each window. For the low-resolution input, our cross-window connection layer is equal to a global self-attention layer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.162, + 0.334, + 0.177 + ], + "angle": 0, + "content": "3.3. STViT for downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.469, + 0.333 + ], + "angle": 0, + "content": "Our method significantly reduces the computation cost by using a small number of semantic tokens, while its side effect is losing nearly all the detailed position information. High-level vision tasks, such as object detection and instance segmentation, are difficult to be executed on this extremely incomplete feature map. This issue also exists in previous works, which hinders the application of token sparsification methods. To solve this issue, we design a STViTR network based on STViT to restore the original spatial resolution from the semantic tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.469, + 0.41 + ], + "angle": 0, + "content": "Our STViT-R shown in Figure 2b has two modifications compared with STViT. First, we adopt a recovery module to restore the spatial resolution from semantic tokens; second, we regroup the transformer layers and construct dumbbell units composed of our STViT-R." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.469, + 0.534 + ], + "angle": 0, + "content": "Recovery module. In the recovery module, only the self-attention layer is employed to restore the spatial resolution without any additional networks. The image tokens \\(X\\) and semantic tokens \\(S\\) are partitioned as \\(N_w^r\\) windows of size \\(w^r \\times w^r\\) and \\(w_s^r \\times w_s^r\\), respectively. The image tokens in each window aggregate the semantic tokens in the corresponding window, which is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.537, + 0.468, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\hat {X} = M H A (X, S, S) + X, \\quad X = F F N (\\hat {X}) + \\hat {X}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.469, + 0.603 + ], + "angle": 0, + "content": "This is a reverse operation of the generation of semantic tokens, using high-level semantic information to boost the image tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.907 + ], + "angle": 0, + "content": "Dumbbell unit. The transformer layers are regrouped into multiple dumbbell units in our STViT-R. Each dumbbell unit consists of four parts. The transformers in the first part are responsible for coping with image tokens; the second part is the semantic token generation module; the transformer layers in the third part deal with semantic tokens; the last part is the recovery module. Take the application on Swin-S (STViT-R-Swin-S) as an example. One, two, two and one transformer layers are allocated for these four parts, respectively. In total, each dumbbell unit is composed of 6 transformer layers. We concatenate three dumbbell units in Stage 3 of Swin-S. In each dumbbell unit, the intermediate transformer layers process semantic tokens with high-level semantic information to save computational cost, and the complete spatial resolution is recovered at the end. By repeating multiple dumbbell units, the detailed spatial information will be preserved by the network, which can not only enhance the classification but also serve downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.19 + ], + "angle": 0, + "content": "STViT will first be applied in two representative ViT models, DeiT [35] and Swin [26] for image classification and video recognition. To validate that our method is effective in downstream tasks, STViT-R is then performed on object detection and instance segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.194, + 0.688, + 0.209 + ], + "angle": 0, + "content": "4.1. Image classification" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.216, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Settings. For image classification, all the models are trained on the ImageNet [13] with 1.28M training images and 50K validation images from 1,000 classes. By default, the semantic token generation module (STGM) employs the \\(5^{th}\\) and \\(6^{th}\\) transformer layers of DeiT (with 12 layers in total), employs the \\(3^{th}\\) and \\(4^{th}\\) transformer layers of Stage 3 of Swin-T (with 12 layers in total), and employs the \\(11^{th}\\) and \\(12^{th}\\) transformer layers of Stage 3 of Swin-S and Swin-B (with 24 layers in total). The image resolution in training and inference is \\(224 \\times 224\\) unless otherwise specified. The batch size is 1,024. All the models are trained from scratch for 300 epochs, and the augmentation and regularization strategies follow the original papers of DeiT and Swin. No knowledge distillation algorithms are used in our experiments. The classification is performed by applying a global average pooling layer on the output tokens of the last transformer layer, followed by a linear classifier. In evaluation, the top-1 accuracy using a single crop is reported. The FLOPs computations of this paper are measured by Fvcore². Throughput is measured with the batch size of 128 on a V100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.534, + 0.892, + 0.624 + ], + "angle": 0, + "content": "On Swin [26], the semantic tokens are generated in Stage 3, and they are not downsampled in Stage 4 due to rare semantic tokens. The patch merging layer between Stage 3 and Stage 4 is replaced with a simple linear layer to double the number of channels. \\( w_{k} \\) is set as 10 and 14 for two transformer layers of STGM." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.636, + 0.892, + 0.847 + ], + "angle": 0, + "content": "Results. One of the advantages of our method is that it can be applied to both global and local vision transformers to reduce computational complexity. Our main results on DeiT and Swin are summarized in Table 1 and Table 2, respectively. The results of LV-ViT [20] are illustrated in Supplementary A.3 due to limited space. We report the top-1 accuracy, FLOPs, and the throughput under different numbers of semantic tokens. On DeiT, the models with 16 semantic tokens achieve the same accuracy as the DeiT models with 196 tokens and save nearly \\(60\\%\\) FLOPs on DeiT-T, DeiT-S, and DeiT-B. With more semantic tokens, the accuracy can consistently outperform the base models. For instance, STViT-DeiT-B with 36 semantic tokens surpasses DeiT-B by \\(0.4\\%\\) accuracy with \\(52\\%\\) FLOPs reduction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.848, + 0.892, + 0.877 + ], + "angle": 0, + "content": "The local vision transformer like Swin is already an efficient architecture compared to the global vision trans" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.858, + 0.9 + ], + "angle": 0, + "content": "2https://github.com/facebookresearch/fvcore" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6199" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.147, + 0.089, + 0.825, + 0.272 + ], + "angle": 0, + "content": "
ModelMetricsBaseNo. of semantic tokens
163664100
STViT-DeiT-TTop-1 Acc(%)72.272.2(+0.0%)72.7(+0.5)73.0(+0.8)73.2(+1.0)
FLOPs(G)1.260.53(-58%)0.60(-52%)0.71(-44%)0.86(-32%)
Throughput(img/s)27525511(+101%)4769(+74%)4214(+53%)3551(+29%)
STViT-DeiT-STop-1 Acc(%)79.879.8(+0.0)80.1(+0.3)80.5(+0.7)80.6(+0.8)
FLOPs(G)4.581.91(-58%)2.20(-52%)2.62(-43%)3.16(-31%)
Throughput(img/s)14082891(+105%)2542(+80%)2229(+58%)1837(+30%)
STViT-DeiT-BTop-1 Acc(%)81.881.8(+0.0)82.2(+0.4)82.6(+0.8)82.7(+0.9)
FLOPs(G)17.587.31(-58%)8.44(-52%)10.04(-43%)12.13(-31%)
Throughput(img/s)6261308(+110%)1150(+85%)1087(+61%)826(+33%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.283, + 0.893, + 0.311 + ], + "angle": 0, + "content": "Table 1. Applying STViT to DeiT-T, DeiT-S, and DeiT-B. The top-1 accuracy, complexity in FLOPs, and throughput are reported for different numbers of semantic tokens." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.325, + 0.821, + 0.507 + ], + "angle": 0, + "content": "
ModelMetricsBaseMove STGMNo. of semantic tokens
4916
STViT-Swin-TTop-1 Acc(%)81.381.0(-0.3%)80.8(-0.5)81.5(+0.2)81.8(+0.5%)
FLOPs(G)4.53.14(-30%)2.99(-34%)3.43(-24%)4.06(-10%)
Throughput(img/s)8781124(+29%)1128(+29%)1061(+22%)1008(+15%)
STViT-Swin-STop-1 Acc(%)83.082.8(-0.2%)82.4(-0.6%)83.0(-0.0)83.1(+0.1%)
FLOPs(G)8.75.95(-32%)5.95(-32%)6.53(-25%)7.36(-15%)
Throughput(img/s)551739(+35%)732(+34%)691(+26%)657(+20%)
STViT-Swin-BTop-1 Acc(%)83.583.2(-0.3%)83.0(-0.5)83.4(-0.1)83.7(+0.2%)
FLOPs(G)15.410.48(-32%)10.48(-32%)11.51(-25%)12.97(-16%)
Throughput(img/s)415558(+35%)551(+33%)521(+26%)489(+19%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.519, + 0.893, + 0.561 + ], + "angle": 0, + "content": "Table 2. Applying STViT to Swin-T, Swin-S, and Swin-B. The top-1 accuracy, complexity in FLOPs, and throughput are reported for different numbers of semantic tokens in each window. Base indicates the corresponding original Swin model. Move STGM indicates changing the default position of STGM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.788 + ], + "angle": 0, + "content": "former, so the reduction of FLOPs on Swin models are smaller than on DeiT models. When 9 semantic tokens are used in each window, STViT-Swin models can reduce \\(25\\%\\) FLOPs with negligible accuracy loss on all the model sizes. If the number of tokens is reduced to 4 in each window (16 in total), a significant accuracy drop will occur, which indicates that local vision transformers need more semantic tokens than global vision transformers. We can move the STGM towards shallow layers to attain a better complexity/accuracy trade-off. In Table 2, STGM is moved by one transformer layer on STViT-Swin-T and by two layers on STViT-Swin-S and STViT-Swin-B to save over \\(30\\%\\) FLOPs with only about \\(0.3\\%\\) accuracy drops (column of \"Move STGM\")." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.469, + 0.882 + ], + "angle": 0, + "content": "STViT-R equipped with recovery modules is designed for downstream tasks, while it also can perform image classification. The corresponding results are reported in Table 3. On both Swin-S and Swin-B, STViT-R can save \\(33\\%\\) FLOPs with \\(0.3\\%\\) accuracy drop. The hyper-parameters we used in STGM are as same as STViT-Swin." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "These results demonstrate that our method achieves both" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.631 + ], + "angle": 0, + "content": "
ModelTop-1 Acc(%)FLOPs(G)Throughput
STViT-R-Swin-S82.7(-0.3)5.83(-33%)717(+30%)
STViT-R-Swin-B83.2(-0.3)10.26(-33%)539(+30%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.637, + 0.892, + 0.679 + ], + "angle": 0, + "content": "Table 3. STViT-R is evaluated on Swin-S and Swin-B on ImageNet. The top-1 accuracy, complexity in FLOPs, and throughput are reported." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.826 + ], + "angle": 0, + "content": "effectiveness and efficiency by employing a few semantic tokens to replace original image tokens. Our method reveals that constructing the tokens with high-level semantic representation is more important than maintaining structured tokens in ViTs. As reflected by the throughput, our method does not have overhead of memory or deployment. Compared to the total parameters, the additional parameters introduced by intra and inter-window spatial pooling is negligible (less than \\(0.05\\%\\)), so we do not show them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparisons with existing token sparsification methods. In Table 4, we compare STViT with the state-of-the-art token sparsification methods on DeiT-S and DeiT-B. Due to different base models used by different methods, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6200" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.454, + 0.386 + ], + "angle": 0, + "content": "
ModelTop-1 AccFLOPs(G)△
DeiT-S
DynamicViT [31]79.32.9(-37%)-0.5
IA-RED2[30]79.13.2(-30%)-0.7
PS-ViT [34]79.42.6(-43%)-0.4
TokenLearner* [32]76.11.9(-44%)-1.8
DGE* [33]79.73.1 (-49%)-0.6
A-ViT* [47]78.63.6 (-39%)-0.3
Evo-ViT [45]79.43.0(-35%)-0.4
EViT [24]78.52.3(-50%)-1.3
STViT(Ours)79.81.91(-58%)-0.0
DeiT-B
IA-RED2[30]80.311.8(-33%)-1.5
DynamicViT [31]81.311.2(-36%)-0.5
PS-ViT [34]81.59.8(-44%)-0.3
TokenLearner* [32]83.728.7(-48%)-1.1
Evo-ViT [45]81.310.2(-33%)-0.5
EViT [24]80.08.7(-51%)-1.8
STViT(Ours)81.87.31(-58%)-0.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.451 + ], + "angle": 0, + "content": "Table 4. Comparisons with the state-of-the-art token sparsification methods on DeiT-S and DeiT-B. \\(\\triangle\\) shows the accuracy difference between each model and its base model. \\(*\\): their base models are not standard DeiT models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.541 + ], + "angle": 0, + "content": "adopt accuracy difference between each model and its base model \\(\\Delta\\) to evaluate them for fair. Results indicate that our method achieves the lowest accuracy drop \\(\\Delta\\) with the highest FLOPs reduction, outperforming all the state-of-the-art methods in both accuracy and efficiency significantly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.553, + 0.25, + 0.57 + ], + "angle": 0, + "content": "4.2. Video recognition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Setting. For video recognition, we apply our STViT to Video Swin [27]. All the models are pre-trained on ImageNet-1K and trained on Kinetics-400 [7]. We generate semantic tokens from each frame as illustrated in Section 3.2. The initialization from pre-trained models and other implementation details are as same as Video Swin [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.751 + ], + "angle": 0, + "content": "Results. The results are presented in Table 6. On Swin-T and Swin-S, STViT-Swin can save about \\(27\\%\\) FLOPs with \\(0.3\\%\\) accuracy drop, which shows that our method works on video recognition." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.47, + 0.796 + ], + "angle": 0, + "content": "4.3. Applications in object detection and instance segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.471, + 0.911 + ], + "angle": 0, + "content": "Settings. Experiments of object detection and instance segmentation are conducted on COCO 2017 [25]. We evaluate STViT-R with Swin in Cascade Mask R-CNN [4, 17] detection frameworks. The \\( w_{s} \\) is set to 3. The backbone models are pre-trained on ImageNet-1K and the pre-trained results are shown in Table 3. All the other settings follow Swin [26]." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.088, + 0.895, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.163, + 0.892, + 0.19 + ], + "angle": 0, + "content": "Figure 3. Visualization example of attention maps in the first attention layer of STGM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.196, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Comparison to Swin Transformer. The performance of STViT-R-Swin using the Cascade Mask R-CNN framework with \\(3 \\times\\) schedule is shown in Table 5. Our method achieves better performance with more than \\(30\\%\\) FLOPs reduction for backbone on bath object detection and instance segmentation. This validates that the recovery module and dumbbell unit can restore detailed spatial information, and the global context information integrated from the semantic tokens significantly benefits object detection. Ignoring spatial structure in the intermediate layers does not affect the object detection task, which is a meaningful fact to help design efficient object detection frameworks. Another interesting finding is that our method has a remarkable improvement for small object detection which is a challenging problem in the detection community. For instance, our STViT-R-Swin-S outperforms Swin-S by \\(1.5\\%\\) on \\(AP_{s}^{b}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.449, + 0.651, + 0.466 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.503 + ], + "angle": 0, + "content": "All the following ablation study of STViT and STViT-R are conducted on the DeiT-S and Swin-S, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.795 + ], + "angle": 0, + "content": "Initialization analysis. The semantic tokens are the cluster centers recovered by attention layers. The initialization of cluster centers induces the representation of semantic tokens. Spatial and global initialization are adopted in Section 3.1 to guide the semantic tokens to integrate local and global semantic information separately. We compare different initialization components in Table 7. When performing single global initialization \\((3^{rd}\\) row), we replace the spatial initial cluster centers with global initial cluster centers in the first transformer layer. The accuracy of using a single initialization method is far lower than using both, which shows the effectiveness of our initialization strategy. Our global initial cluster centers look similar to learned positional encoding since both use random initialization. To verify their distinction, we experiment with a real learned positional encoding by additionally adding the global initial cluster centers to keys, which causes \\(0.1\\%\\) accuracy drop (the last row of Table 7). Therefore, our global initialization is different from learned positional encoding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We visualize the attention maps of the first attention layer in the STGM in Figure 3. Because the queries of this layer are spatial initial cluster centers, these attention maps visualize the local semantic information integration by attention. The attention layer groups the semantic information according to the position of initial cluster centers, which ensures to extract fine-grained semantic information" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "6201" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.183, + 0.089, + 0.788, + 0.181 + ], + "angle": 0, + "content": "
APbAPb50APb75APsbAPmAPm50APm75APsmFLOPs(G)
Swin-S51.870.456.335.244.767.948.528.8194
STViT-R-Swin-S51.870.656.136.744.767.848.629.0134(-31%)
Swin-B51.970.956.535.445.068.448.728.9343
STViT-R-Swin-B52.270.856.836.545.268.349.129.5233(-32%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.187, + 0.895, + 0.214 + ], + "angle": 0, + "content": "Table 5. Results on COCO object detection and instance segmentation under Cascade Mask R-CNN with \\(3 \\times\\) schedule. The FLOPs are measured for backbones." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.224, + 0.455, + 0.307 + ], + "angle": 0, + "content": "
ModelTop-1 Acc(%)FLOPs(G)Speed
Swin-T78.888779
STViT-Swin-T78.5(-0.3)64.4(-27%)975(+25%)
Swin-S80.6166456
STViT-Swin-S80.3(-0.3)120.5(-27%)572(+25%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.368 + ], + "angle": 0, + "content": "Table 6. Applying STViT to Video Swin (Swin-T and Swin-S) on Kinetics-400. All the models are pre-trained on ImageNet-1K. The views are \\(4\\times 3\\). The top-1 accuracy and complexity in FLOPs are reported. Speed is evaluated by FPS." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.371, + 0.404, + 0.461 + ], + "angle": 0, + "content": "
SpatialGlobalLearnedTop-1 Acc(%)
✓79.4
✓78.7
✓✓79.8
✓✓79.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.507 + ], + "angle": 0, + "content": "Table 7. Accuracy with different initialization of STViT. Spatial, Global, and Learned indicate spatial initialization, global initialization, and learned positional encoding methods, respectively." + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.517, + 0.412, + 0.575 + ], + "angle": 0, + "content": "
No. of transformers234
Top-1 Acc(%)79.879.579.6
FLOPs(G)1.911.972.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.621 + ], + "angle": 0, + "content": "Table 8. Performance evaluation on different numbers of transformer layers in STGM. Keeping the base module containing four transformer layers unchanged." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.631, + 0.414, + 0.69 + ], + "angle": 0, + "content": "
STViT-Rw/o DUReusing ST
APb51.851.451.6
APm44.744.444.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.725 + ], + "angle": 0, + "content": "Table 9. Ablation study on STViT-R w/o dumbbell units (w/o \\(DU\\)) and reusing semantic tokens (Reusing ST)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.866 + ], + "angle": 0, + "content": "and keep the difference among semantic tokens. The attention map in the second attention layer is visualized in Supplementary A.5, which reveals that the network fixedly selects particle semantic tokens to represent global semantic information. We also show the semantic representation of image tokens in the same transformer layer on DeiT in Supplementary A.5. Compared to original image tokens, our semantic tokens in Figure 3 show more high-level semantic information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Number of transformer layers in STGM. Two transformer layers are adopted in the STGM by default. The ef" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.296 + ], + "angle": 0, + "content": "fcts of employing different numbers of transformer layers are shown in Table 8. The extra layers are from the ones behind STGM to keep the total number of layers unchanged. More transformer layers do not bring improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.892, + 0.442 + ], + "angle": 0, + "content": "The effectiveness of the dumbbell unit. To verify the effectiveness of our dumbbell unit, we experiment STViT-R without dumbbell units, i.e., STViT equipped with only the recovery module. We employ \\(6^{th}\\) and \\(7^{th}\\) transformer layers to construct STGM and the last transformer layers to construct the recovery module in Stage 3. The FLOPs is the same as the full-model STViT-R for fair comparison. The results on the COCO are reported in Table 9. Inferior results demonstrate the effectiveness of the dumbbell unit." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Reusing semantic tokens in dumbbell units. Semantic tokens are generated in each dumbbell unit. If they are produced only once in the first dumbbell unit and reused as initial cluster centers in the subsequent dumbbell units, the result is shown in Table 9 with a slight performance drop." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.539, + 0.619, + 0.554 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.789 + ], + "angle": 0, + "content": "In this paper, we propose a simple and effective token sparsification method, semantic token vision transformer (STViT). Our method utilizes the clustering property of self-attention to generate a few semantic tokens with high-level information representation to replace the redundant image/video tokens, which can be applied in both global and local vision transformers. By simply configuring the recovery module, our method can be successfully applied to downstream tasks. Extensive experiments demonstrate that our method achieves better accuracy along with less inference time in most cases. The success in downstream tasks significantly boosts the development of token sparsification methods. We hope that this work can inspire more future research to pay much attention to high-level semantic representation in ViTs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.791, + 0.661, + 0.808 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.903 + ], + "angle": 0, + "content": "This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, and Mike Zheng Shou's Start-Up Grant from NUS. Computation was partially performed on resources of the National Supercomputing Centre, Singapore. Shuning was supported by Alibaba Research Intern Program." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6202" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Luvcic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Song Bai, Philip Torr, et al. Visual parser: Representing part-whole hierarchies with transformers. arXiv preprint arXiv:2107.05790, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.47, + 0.383 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[8] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[9] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.608, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[10] Yunpeng Chen, Marcus Rohrbach, Zhicheng Yan, Yan Shuicheng, Jiashi Feng, and Yannis Kalantidis. Graph-based global reasoning networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-442, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[11] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1601-1610, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[14] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Bain-" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "ing Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. arXiv preprint arXiv:2107.00652, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[16] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.301, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[18] Zilong Huang, Youcheng Ben, Guozhong Luo, Pei Cheng, Gang Yu, and Bin Fu. Shuffle transformer: Rethinking spatial shuffle for vision transformer. arXiv preprint arXiv:2106.03650, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.399, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[20] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.538, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Bin Ren, Minghai Qin, Hao Tang, and Yanzhi Wang. Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient self-supervised vision transformers for representation learning. arXiv preprint arXiv:2106.09785, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[23] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[24] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In International Conference on Learning Representations, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6203" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[27] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[28] Xuezhe Ma, Xiang Kong, Sinong Wang, Chunting Zhou, Jonathan May, Hao Ma, and Luke Zettlemoyer. Luna: Linear unified nested attention. Advances in Neural Information Processing Systems, 34:2441-2453, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[29] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. arXiv preprint arXiv:2111.15668, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[30] Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. Ia-red 2: Interpretability-aware redundancy reduction for vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[31] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.483 + ], + "angle": 0, + "content": "[32] Michael S. Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[33] Lin Song, Songyang Zhang, Songtao Liu, Zeming Li, Xuming He, Hongbin Sun, Jian Sun, and Nanning Zheng. Dynamic grained encoder for vision transformers. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[34] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. arXiv preprint arXiv:2106.02852, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[35] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[36] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[37] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[38] Pichao Wang, Xue Wang, Hao Luo, Jingkai Zhou, Zhipeng Zhou, Fan Wang, Hao Li, and Rong Jin. Scaled relu matters for training vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[40] Wenxiao Wang, Lu Yao, Long Chen, Binbin Lin, Deng Cai, Xiaofei He, and Wei Liu. Crossformer: A versatile vision transformer hinging on cross-scale attention. arXiv preprint arXiv:2108.00154, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[41] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. Advances in Neural Information Processing Systems, 34, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8741-8750, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[43] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22-31, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[44] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9981–9990, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[45] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[46] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[47] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[48] Tan Yu, Gangming Zhao, Ping Li, and Yizhou Yu. Boat: Bilateral local attention vision transformer. arXiv preprint arXiv:2201.13027, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[49] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 558-567, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[50] David Junhao Zhang, Kunchang Li, Yali Wang, Yunpeng Chen, Shashwat Chandra, Yu Qiao, Luoqi Liu, and Mike Zheng Shou. Morph MLP: An efficient mlp-like backbone for spatial-temporal representation learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 230-248. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[51] Songyang Zhang, Xuming He, and Shipeng Yan. Latent-gnn: Learning efficient non-local relations for visual recog" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6204" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "nition. In International Conference on Machine Learning, pages 7374-7383. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[52] Minghang Zheng, Peng Gao, Renrui Zhang, Kunchang Li, Xiaogang Wang, Hongsheng Li, and Hao Dong. End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[54] Jingkai Zhou, Pichao Wang, Fan Wang, Qiong Liu, Hao Li, and Rong Jin. Elsa: Enhanced local self-attention for vision transformer. arXiv preprint arXiv:2112.12786, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[55] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[56] Zhuofan Zong, Kunchang Li, Guanglu Song, Yali Wang, Yu Qiao, Biao Leng, and Yu Liu. Self-slimmed vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 432-448. Springer, 2022. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6205" + } + ] +] \ No newline at end of file diff --git a/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_origin.pdf b/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ffaeabdc8c0b1b800c91c4b1b44276f26ac12245 --- /dev/null +++ b/2023/Making Vision Transformers Efficient From a Token Sparsification View/0f00c5ea-237b-435f-ba19-f0ae533b96fb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f220359d3d4b4ccddf831016be68f46780e37957e854b91cfed8f95493064152 +size 2205538 diff --git a/2023/Making Vision Transformers Efficient From a Token Sparsification View/full.md b/2023/Making Vision Transformers Efficient From a Token Sparsification View/full.md new file mode 100644 index 0000000000000000000000000000000000000000..11f1ab825d7352cfd1927711c6dfbb196d01354a --- /dev/null +++ b/2023/Making Vision Transformers Efficient From a Token Sparsification View/full.md @@ -0,0 +1,291 @@ +# Making Vision Transformers Efficient from A Token Sparsification View + +Shuning Chang $^{1*}$ Pichao Wang $^{2\dagger\dagger}$ Ming Lin $^{2\dagger}$ Fan Wang $^{2}$ David Junhao Zhang $^{1}$ Rong Jin $^{2}$ Mike Zheng Shou $^{1\dagger}$ $^{1}$ Show Lab, National University of Singapore $^{2}$ Alibaba Group + +# Abstract + +The quadratic computational complexity to the number of tokens limits the practical applications of Vision Transformers (ViTs). Several works propose to prune redundant tokens to achieve efficient ViTs. However, these methods generally suffer from (i) dramatic accuracy drops, (ii) application difficulty in the local vision transformer, and (iii) non-general-purpose networks for downstream tasks. In this work, we propose a novel Semantic Token ViT (STViT), for efficient global and local vision transformers, which can also be revised to serve as backbone for downstream tasks. The semantic tokens represent cluster centers, and they are initialized by pooling image tokens in space and recovered by attention, which can adaptively represent global or local semantic information. Due to the cluster properties, a few semantic tokens can attain the same effect as vast image tokens, for both global and local vision transformers. For instance, only 16 semantic tokens on DeiT-(Tiny,Small,Base) can achieve the same accuracy with more than $100\%$ inference speed improvement and nearly $60\%$ FLOPs reduction; on Swin-(Tiny,Small,Base), we can employ 16 semantic tokens in each window to further speed it up by around $20\%$ with slight accuracy increase. Besides great success in image classification, we also extend our method to video recognition. In addition, we design a STViT-R(covery) network to restore the detailed spatial information based on the STViT, making it work for downstream tasks, which is powerless for previous token sparsification methods. Experiments demonstrate that our method can achieve competitive results compared to the original networks in object detection and instance segmentation, with over $30\%$ FLOPs reduction for backbone. + +# 1. Introduction + +In contrast to standard Convolutional Neural Networks (CNNs) approaches which process images pixel-by-pixel, + +Vision Transformers (ViTs) [15, 26, 35, 36, 43] treat an image as a sequence of patch/image tokens, and have shown promising performance in prevalent visual recognition scenarios. However, these superior performances do not come for free: the quadratic computational complexity to the number of image tokens limits their application in practice. Previous works [33, 56] have illustrated the large amount of redundancy in the image tokens and also shown the effect of filtering out unimportant tokens normally according to predefined scoring mechanism. However, these methods face the following challenges. Firstly, the predefined scoring mechanisms for filtering are generally imprecise. In Figure 1, on the left we visualize the class token values in different layers which are commonly used to score the token importance [16, 24, 45]. Different layers have different value distributions, thus using these imprecise scores for filtering would lead to unsatisfactory performance. For example, EViT [24] has an accuracy drop of $1.3\%$ when saving $50\%$ FLOPs on DeiT-S [35]. Secondly, the remaining tokens do not distribute evenly in space any more, making them hard to work in local vision transformers1. Finally, large-scale token pruning tremendously damages the spatial structure and positional information, and causes difficulties when applied to downstream tasks, which they do not propose a solution to deal with. + +To solve these problems, we propose Semantic Token ViT (STViT), for efficient global and local vision transformers, which also can be revised to serve as backbone for downstream tasks. The proposed approach is based on the following observations: (i) unlike local CNNs which learn spatial structure of images, vision transformer discretizes feature map as tokens for global feature exploration, relieving the requirements for maintaining the whole image structure and information; (ii) discrete tokens are more beneficial for optimization [38]; (iii) in Figure 1, on the right shows the attention maps in different transformer layers, and there are only several vertical lines in the deep layers, which means that only a few tokens with global se + +![](images/bde36492b427be85ae4cd307425e95003502ef7a7d4b4f6e038ec4f55f7518fd.jpg) +Figure 1. Left: the attention values of class tokens (normalized and reshaped in image shape) in different self-attention layers. Right: the attention maps in different self-attention layers. Zoom-in for better visibility. + +![](images/f30fdba2db5b29b8a686a45071e5d10b6580de568ef44123696b1d30204ad83c.jpg) + +matic information matter. Thus, we argue that it is not necessary to maintain massive structured tokens for ViTs, especially in the deep layers. Employing a few discrete tokens with high-level semantic information can potentially achieve both high performance and efficiency. + +In STViT, the semantic tokens represent the cluster centers, and the number of them is far less than the original image tokens, significantly reducing the computational cost. Inspired by the fact that multi-head attention can conduct the cluster center recovery (Supplementary A.7), we only employ the off-the-shelf self-attention to generate the semantic tokens. Specifically, the first few transformer layers are kept unchanged to obtain the image tokens with low-level features. The image tokens are then fed into our semantic token generation module (STGM) consisting of at least two transformer layers to generate semantic tokens. In each self-attention layer, the semantic tokens are input as queries, and the image tokens are fed as keys and values. The semantic tokens dynamically aggregate image tokens through the attention layers to recover cluster centers. In the first attention layer, the semantic tokens are initialized by an intra and inter-window spatial pooling which takes into account incorporating semantic information in each window and maximizing distance between adjacent windows. Thanks to this spatial initialization, the semantic tokens mainly incorporate local semantic information and achieve discrete and uniform distribution in space. In the following attention layer, besides further clustering, the semantic tokens are equipped with global cluster centers, and the network can adaptively select partial semantic tokens to focus on global semantic information. After the STGM, the original image tokens are discarded, and only semantic tokens are kept for the subsequent transformer layers. Because the generation of semantic tokens is flexible and space-aware, our method can be plugged into both global and local vision transformers. The semantic tokens can be produced in each window for the local vision transformer. + +Another property of STViT is its capability to serve as a backbone for downstream tasks, such as object detection and instance segmentation. Discussions have been missing in previous methods [16, 24, 32, 45, 56] about how to use them in downstream task under the massive loss of spatial information during the token sparsification process, which actually seriously impedes the application of their method. Instead, we design a novel STViT-R network based + +on STViT where a recovery module and dumbbell unit are adopted to periodically restore the full resolution feature map while the intermediate transformer layers continue to use semantic tokens to save computation cost, making our method work in downstream task. + +The effectiveness of the proposed method is validated via a comprehensive empirical study on image and video ViT models. Only 16 semantic tokens on DeiT-(Tiny, Small, Base) achieve nearly $50\%$ inference time reduction without any accuracy degradation; on Swin-(Tiny, Small, Base), we also improve the inference throughput by nearly $20\%$ with slight accuracy increase. Moreover, the proposed STViT-R achieves promising results on object detection and instance segmentation. To the best of our knowledge, this is one of first works to apply the token sparsification algorithm in local vision transformers, and use the ViTs as backbones in downstream tasks after large-scale token pruning. Our findings in ViTs uncover that maintaining the full-size feature map is unnecessary, and a few tokens with high-level semantic representations can achieve both high performance and efficiency. Thanks to its simplicity and general-purpose ability, our method can also serve as a new efficient ViT baseline architecture and a starting point for further research from the token sparsification perspective. + +# 2. Related work + +Vision transformers. Vision Transformer (ViT) [15] first introduces a pure Transformer backbone for image classification. ViT variants further inspire the applications of transformer to various vision tasks beyond image/video classification [1, 3, 20, 35, 36, 39, 43, 44, 49, 50], such as object detection [5, 12, 52, 55], semantic segmentation [37, 42, 53], and self-supervised learning [6, 9, 22]. Vanilla transformers have high computational and memory costs because the multi-head self-attention has quadratic computational complexity to the number of image tokens. Recently, various efficient ViTs have been proposed to alleviate this issue. The existing methods mainly focus on reducing the complexity of self-attention or reducing the number of tokens. Swin Transformer [26] adopts local self-attention, i.e., attending neighboring tokens within a constant window size, achieving a linear computational complexity in the self-attention with high performance. Many subsequent works [11, 14, 18, 40, 46, 48, 54] follow the local + +self-attention design to develop variants. Token sparsification [8, 16, 21, 24, 29–34, 41, 45, 47, 56] also attracts increasing attention. + +Token sparsification. Token sparsification methods can be mainly categorized into hard pruning [8,16,21,24,29-31, 33,34,45,47] and soft pruning [32,56]. Hard pruning methods filter out some unimportant tokens according to a predefined scoring mechanism. DynamicViT [31], SPViT [21], and AdaViT [29] introduce additional prediction networks to score the tokens. Evo-ViT [45], ATS [16], and EViT [24] utilize the values of class tokens to evaluate the importance of tokens. However, it is difficult to achieve precise scoring as shown in the left of Figure 1. Therefore, they usually suffer from a significant accuracy drop. For instance, EViT [24] has an accuracy drop of $1.3\%$ when saving $50\%$ FLOPs on DeiT-S. Soft pruning methods generate new tokens from image tokens by importing additional attention networks. TokenLearner [32] also argue for a few tokens to replace image tokens. However, its price is a $1.8\%$ accuracy drop when reducing $44\%$ FLOPs, which is far inferior to concurrent works. Besides performance degradation, previous methods also have the following disadvantages. First, whether or how to extend the methods to local vision transformers remains unexplored. Second, it has not been discussed about how to serve the downstream tasks like object detection and instance segmentation after the tokens are pruned. + +In our method, we apply the off-the-shelf transformer layers to reduce token number. [2, 10, 19, 23, 28, 51] adopt similar approaches to achieve efficient non-local relationships. Our method is different from them as below: (i) our method extracts local semantic information instead of non-local relationships; (ii) the semantic tokens are a few cluster centers, which can replace the massive image tokens to achieve image classification; (iii) our method specializes in pruning tokens. + +# 3. Method + +The proposed STViT is presented in this section, which aims to construct an efficient and high-performance ViT. STViT is first introduced in Section 3.1, followed by how to apply STViT in the local vision transformer in Section 3.2. Based on STViT, STViT-R is developed to restore the spatial resolution for downstream tasks in Section 3.3. + +# 3.1.STViT + +Overall architecture. An overview of STViT architecture is presented in Figure 2a. The patch embedding layer and shallow transformer layers are kept unchanged as a base module in our method. The base module copes with all the image tokens $X \in \mathbb{R}^{N_i \times C}$ to extract low-level features, + +where $N_{i}$ is the number of image tokens and $C$ is the number of channels. The image tokens are fed into the semantic token generation module (STGM) to generate $N_{s}$ semantic tokens $S \in \mathbb{R}^{N_s \times C}$ . After the STGM, the image tokens $X$ can be discarded, and only semantic tokens $S$ with high-level semantic information are used in all the subsequent transformers. Due to $N_{s} \ll N_{i}$ , our method can significantly reduce the computational cost. + +Semantic token generation module (STGM). The whole image is represented by a few tokens with high-level semantic information through clustering. Inspired by the fact that self-attention can conduct cluster center recovery (Supplementary A.7), we adopt the off-the-shelf self-attention layers to produce the semantic tokens. The STGM consists of at least two transformer layers. + +The initial cluster centers $P \in \mathbb{R}^{N_s \times C}$ are generated by an spatial pooling which pools the image tokens into fixed $w_{s} \times w_{s}$ tokens, with $N_{s} = w_{s} \times w_{s}$ . $w_{s}$ is generally set as 4. The spatial pooling can be achieved by a non-parameterized adaptive spatial pooling or a super lightweight network with higher performance, intra and inter-window spatial pooling. The intentions of spatial pooling initialization are three folds. First, the initial cluster centers can distribute uniformly in space, making the generated semantic tokens more discrete and preventing the semantic tokens from collapsing to one point in the following layers. Second, the semantic tokens can be forced to represent more local and distinguished features. Finally, the representation of semantic tokens is associated with the specific spatial locations, which is the basis to allow our method to be applied in local self-attention and downstream tasks. The initial cluster centers $P$ then dynamically integrate the image tokens $X$ according to semantic information by attention mechanism. In the first transformer layer, the processing of the generation of semantic tokens can be written as + +$$ +\hat {S} ^ {1} = M H A (P, X, X) + P, \quad S ^ {1} = F F N (\hat {S} ^ {1}) + \hat {S} ^ {1}, \tag {1} +$$ + +where $MHA$ and $FFN$ are short for multi-head attention layer and feed-forward network, respectively, and the triplet input of $MHA$ are queries, keys, and values in turn. All the norm layers are omitted in all the equations for brevity. The initial cluster centers are produced in a window by an adaptive spatial pooling layer or an intra and inter-window spatial pooling, while the semantic tokens are generated from a global receptive field by a dynamic attention layer to ensure that they can extract high-level semantic representation. In order to further strengthen the clustering effect, we use the second transformer layer to repeat the clustering operation and guide the semantic tokens to extract global information. In this transformer, the semantic tokens are updated as: + +$$ +\hat {S} ^ {2} = M H A \left(S ^ {1} + G, \operatorname {C o n c a t} \left(S ^ {1}, X\right), \operatorname {C o n c a t} \left(S ^ {1}, X\right)\right) + S ^ {1}, \tag {2} +$$ + +$$ +S ^ {2} = F F N (\hat {S} ^ {2}) + \hat {S} ^ {2}, +$$ + +![](images/4d7fd0c0fdcb716fed6e63a203d651069a48e880cac0e37f31f5b9202b34a222.jpg) +(a) STViT + +![](images/f17100edef88e4795d9c0d37690577ea35a91fff9e7719a10b5fcfb07508a2e8.jpg) +(b) STViT-R +Figure 2. The architectures of our STViT and STViT-R. + +where $G \in \mathbb{R}^{N_s \times C}$ is the global cluster centers initialized by Gaussian noise and $\text{Concat}(\cdot)$ is a concatenation operation. The global cluster centers $G$ are responsible for global semantic information extraction like the class token. Although $S^1$ and $G$ are summed together as the queries, they can be decoupled in the attention computation as: + +$$ +\begin{array}{l} A _ {s} = \left(S ^ {1} \cdot W _ {q}\right) \cdot \left(\left(S ^ {1} + X\right) \cdot W _ {k}\right), \quad A _ {g} = \left(G \cdot W _ {q}\right) \cdot \left(\left(S ^ {1} + X\right) \cdot W _ {k}\right), \tag {3} \\ A = \operatorname {S o f t m a x} (A _ {s} + A _ {g}), \\ \end{array} +$$ + +where $W_{q}$ and $W_{k}$ are the linear projection weights of queries and keys. We can see that $S^{1}$ and $G$ integrate the keys to generate $A_{s}$ and $A_{g}$ independently and just share a Softmax operation to produce the final attention map $A$ . Therefore, besides spatial semantic information, the semantic tokens also incorporate global semantic information with a negligible additional overhead. Though very similar, the global cluster centers are actually different from the learned positional encoding. We do not add the global cluster centers to keys. Moreover, it will be shown in the experiments that inserting the actual learned positional encoding will cause an accuracy drop. The number of transformer layers in STGM is flexible. More transformer layers can be associated for clustering. Two transformer layers are employed by default, i.e., $S^{2}$ is the output of the STGM. Note that the image tokens are not updated in the STGM. + +Intra and inter-window spatial pooling. Give $H \times W$ feature map $X$ , we generate $N_{s}(w_{s} \times w_{s})$ initial cluster centers. We uniformly split the $X$ into $w_{s} \times w_{s}$ windows $[X_{w}^{i}]_{i=1}^{N_{s}}$ with size $\frac{H}{w_{s}} \times \frac{W}{w_{s}}$ and each window generates one initial cluster center. To represent abundant semantic information, we take into account intra and interwindow relations, i.e., integrating important tokens in the window and maximizing the distance among initial cluster centers in different windows. Specifically, we formulate an intra-window function $f_{intra}$ to produce a mask, $M_{i} = f_{intra}(X_{w}^{i})$ , which projects the input window from $\mathbb{R}\frac{H}{w_{s}} \times \frac{W}{w_{s}} \times C \rightarrow \mathbb{R}\frac{H}{w_{s}} \times \frac{W}{w_{s}}$ . The idea is to let the intra-window function $f_{intra}$ adaptively select a combination of informative tokens in $X_{w}^{i}$ , which is implemented by + +$$ +M _ {i} = \operatorname {C o n v} \left(\operatorname {G e L U} \left(\operatorname {L a y e r N o r m} \left(\operatorname {D e p t h w i s e} \operatorname {C o n v} \left(X _ {w} ^ {i}\right)\right)\right)\right). \tag {4} +$$ + +Then, we compute each integrated token $\hat{P}_i =$ Softmax $(M_i)\cdot X_w^i$ from each window, and arrange $[\hat{P}_i]_{i = 1}^{N_s}$ by spatial structure to form the 2D tensor $\hat{P}\in \mathbb{R}^{w_s\times w_s\times C}$ We adopt an inter-window function $f_{inter}$ to compute the inter-window relations and generate the offset, $O =$ $f_{inter}(\hat{P})$ ,to revise the mask $M$ . The implementation of $f_{inter}$ is similar to Eq. 4, except the mapping input from $\mathbb{R}^{w_s\times w_s\times C}\rightarrow \mathbb{R}^{w_s\times w_s\times \frac{HW}{w_s^2}}$ , where $\frac{HW}{w_s^2}$ is the number of tokens in each window. For each window, the corresponding $O_{i}\in \mathbb{R}^{\frac{HW}{w_{s}^{2}}}$ is sliced from $O$ and reshaped to $\mathbb{R}^{\frac{H}{w_s}\times \frac{W}{w_s}}$ The $O_{i}$ is used to revise $M_{i}$ . The final initial cluster center $P_{i}$ is computed by + +$$ +P _ {i} = \operatorname {S o f t m a x} \left(M _ {i} + O _ {i}\right) \cdot X _ {w}. \tag {5} +$$ + +Our $f_{\text{intra}}$ and $f_{\text{intra}}$ are super lightweight and the introducing parameters can be negligible. For example, on DeiT, the parameters only increase by $0.05\%$ . + +# 3.2. STViT in local vision transformers. + +Local self-attention has been widely used in current ViT models to balance efficiency and accuracy. As the generation of semantic tokens in STViT is flexible in space, it can be naturally applied in local self-attention. Suppose each local self-attention layer contains $N_{w}$ windows with size $w \times w$ , we initialize $w_{s} \times w_{s}$ cluster centers in each window by our intra and inter-window spatial pooling. The total number of semantic tokens $N_{s}$ would be $w_{s} \times w_{s} \times N_{w}$ . $w$ and $w_{s}$ are set as 7 and 3 by default separately. As a result, our method compresses more than $80\%$ image tokens in local self-attention. In the STGM, although initial cluster centers are from $w \times w$ windows, we use larger windows with size $w_{k} \times w_{k}$ to obtain keys and values in Eq. 1 and Eq. 2 to mitigate the effect of limited window size. Other operations in STGM are kept the same as Section 3.1. + +In the local ViT models, each local transformer layer is normally followed by a cross-window connection layer, such as a shift window transformer layer following a local transformer layer on Swin Transformer [26]. In our method, the attention is computed within $w_{s} \times w_{s}$ window in the local self-attention layer, and the cross-window connection can be achieved by computing self-attention in a larger-size + +$(e.g., 4 \times w_{s})$ sliding window because of the rare number of tokens in each window. For the low-resolution input, our cross-window connection layer is equal to a global self-attention layer. + +# 3.3. STViT for downstream tasks + +Our method significantly reduces the computation cost by using a small number of semantic tokens, while its side effect is losing nearly all the detailed position information. High-level vision tasks, such as object detection and instance segmentation, are difficult to be executed on this extremely incomplete feature map. This issue also exists in previous works, which hinders the application of token sparsification methods. To solve this issue, we design a STViTR network based on STViT to restore the original spatial resolution from the semantic tokens. + +Our STViT-R shown in Figure 2b has two modifications compared with STViT. First, we adopt a recovery module to restore the spatial resolution from semantic tokens; second, we regroup the transformer layers and construct dumbbell units composed of our STViT-R. + +Recovery module. In the recovery module, only the self-attention layer is employed to restore the spatial resolution without any additional networks. The image tokens $X$ and semantic tokens $S$ are partitioned as $N_w^r$ windows of size $w^r \times w^r$ and $w_s^r \times w_s^r$ , respectively. The image tokens in each window aggregate the semantic tokens in the corresponding window, which is represented as: + +$$ +\hat {X} = M H A (X, S, S) + X, \quad X = F F N (\hat {X}) + \hat {X}. \tag {6} +$$ + +This is a reverse operation of the generation of semantic tokens, using high-level semantic information to boost the image tokens. + +Dumbbell unit. The transformer layers are regrouped into multiple dumbbell units in our STViT-R. Each dumbbell unit consists of four parts. The transformers in the first part are responsible for coping with image tokens; the second part is the semantic token generation module; the transformer layers in the third part deal with semantic tokens; the last part is the recovery module. Take the application on Swin-S (STViT-R-Swin-S) as an example. One, two, two and one transformer layers are allocated for these four parts, respectively. In total, each dumbbell unit is composed of 6 transformer layers. We concatenate three dumbbell units in Stage 3 of Swin-S. In each dumbbell unit, the intermediate transformer layers process semantic tokens with high-level semantic information to save computational cost, and the complete spatial resolution is recovered at the end. By repeating multiple dumbbell units, the detailed spatial information will be preserved by the network, which can not only enhance the classification but also serve downstream tasks. + +# 4. Experiments + +STViT will first be applied in two representative ViT models, DeiT [35] and Swin [26] for image classification and video recognition. To validate that our method is effective in downstream tasks, STViT-R is then performed on object detection and instance segmentation tasks. + +# 4.1. Image classification + +Settings. For image classification, all the models are trained on the ImageNet [13] with 1.28M training images and 50K validation images from 1,000 classes. By default, the semantic token generation module (STGM) employs the $5^{th}$ and $6^{th}$ transformer layers of DeiT (with 12 layers in total), employs the $3^{th}$ and $4^{th}$ transformer layers of Stage 3 of Swin-T (with 12 layers in total), and employs the $11^{th}$ and $12^{th}$ transformer layers of Stage 3 of Swin-S and Swin-B (with 24 layers in total). The image resolution in training and inference is $224 \times 224$ unless otherwise specified. The batch size is 1,024. All the models are trained from scratch for 300 epochs, and the augmentation and regularization strategies follow the original papers of DeiT and Swin. No knowledge distillation algorithms are used in our experiments. The classification is performed by applying a global average pooling layer on the output tokens of the last transformer layer, followed by a linear classifier. In evaluation, the top-1 accuracy using a single crop is reported. The FLOPs computations of this paper are measured by Fvcore². Throughput is measured with the batch size of 128 on a V100 GPU. + +On Swin [26], the semantic tokens are generated in Stage 3, and they are not downsampled in Stage 4 due to rare semantic tokens. The patch merging layer between Stage 3 and Stage 4 is replaced with a simple linear layer to double the number of channels. $w_{k}$ is set as 10 and 14 for two transformer layers of STGM. + +Results. One of the advantages of our method is that it can be applied to both global and local vision transformers to reduce computational complexity. Our main results on DeiT and Swin are summarized in Table 1 and Table 2, respectively. The results of LV-ViT [20] are illustrated in Supplementary A.3 due to limited space. We report the top-1 accuracy, FLOPs, and the throughput under different numbers of semantic tokens. On DeiT, the models with 16 semantic tokens achieve the same accuracy as the DeiT models with 196 tokens and save nearly $60\%$ FLOPs on DeiT-T, DeiT-S, and DeiT-B. With more semantic tokens, the accuracy can consistently outperform the base models. For instance, STViT-DeiT-B with 36 semantic tokens surpasses DeiT-B by $0.4\%$ accuracy with $52\%$ FLOPs reduction. + +The local vision transformer like Swin is already an efficient architecture compared to the global vision trans + +
ModelMetricsBaseNo. of semantic tokens
163664100
STViT-DeiT-TTop-1 Acc(%)72.272.2(+0.0%)72.7(+0.5)73.0(+0.8)73.2(+1.0)
FLOPs(G)1.260.53(-58%)0.60(-52%)0.71(-44%)0.86(-32%)
Throughput(img/s)27525511(+101%)4769(+74%)4214(+53%)3551(+29%)
STViT-DeiT-STop-1 Acc(%)79.879.8(+0.0)80.1(+0.3)80.5(+0.7)80.6(+0.8)
FLOPs(G)4.581.91(-58%)2.20(-52%)2.62(-43%)3.16(-31%)
Throughput(img/s)14082891(+105%)2542(+80%)2229(+58%)1837(+30%)
STViT-DeiT-BTop-1 Acc(%)81.881.8(+0.0)82.2(+0.4)82.6(+0.8)82.7(+0.9)
FLOPs(G)17.587.31(-58%)8.44(-52%)10.04(-43%)12.13(-31%)
Throughput(img/s)6261308(+110%)1150(+85%)1087(+61%)826(+33%)
+ +Table 1. Applying STViT to DeiT-T, DeiT-S, and DeiT-B. The top-1 accuracy, complexity in FLOPs, and throughput are reported for different numbers of semantic tokens. + +
ModelMetricsBaseMove STGMNo. of semantic tokens
4916
STViT-Swin-TTop-1 Acc(%)81.381.0(-0.3%)80.8(-0.5)81.5(+0.2)81.8(+0.5%)
FLOPs(G)4.53.14(-30%)2.99(-34%)3.43(-24%)4.06(-10%)
Throughput(img/s)8781124(+29%)1128(+29%)1061(+22%)1008(+15%)
STViT-Swin-STop-1 Acc(%)83.082.8(-0.2%)82.4(-0.6%)83.0(-0.0)83.1(+0.1%)
FLOPs(G)8.75.95(-32%)5.95(-32%)6.53(-25%)7.36(-15%)
Throughput(img/s)551739(+35%)732(+34%)691(+26%)657(+20%)
STViT-Swin-BTop-1 Acc(%)83.583.2(-0.3%)83.0(-0.5)83.4(-0.1)83.7(+0.2%)
FLOPs(G)15.410.48(-32%)10.48(-32%)11.51(-25%)12.97(-16%)
Throughput(img/s)415558(+35%)551(+33%)521(+26%)489(+19%)
+ +former, so the reduction of FLOPs on Swin models are smaller than on DeiT models. When 9 semantic tokens are used in each window, STViT-Swin models can reduce $25\%$ FLOPs with negligible accuracy loss on all the model sizes. If the number of tokens is reduced to 4 in each window (16 in total), a significant accuracy drop will occur, which indicates that local vision transformers need more semantic tokens than global vision transformers. We can move the STGM towards shallow layers to attain a better complexity/accuracy trade-off. In Table 2, STGM is moved by one transformer layer on STViT-Swin-T and by two layers on STViT-Swin-S and STViT-Swin-B to save over $30\%$ FLOPs with only about $0.3\%$ accuracy drops (column of "Move STGM"). + +STViT-R equipped with recovery modules is designed for downstream tasks, while it also can perform image classification. The corresponding results are reported in Table 3. On both Swin-S and Swin-B, STViT-R can save $33\%$ FLOPs with $0.3\%$ accuracy drop. The hyper-parameters we used in STGM are as same as STViT-Swin. + +These results demonstrate that our method achieves both + +Table 2. Applying STViT to Swin-T, Swin-S, and Swin-B. The top-1 accuracy, complexity in FLOPs, and throughput are reported for different numbers of semantic tokens in each window. Base indicates the corresponding original Swin model. Move STGM indicates changing the default position of STGM. + +
ModelTop-1 Acc(%)FLOPs(G)Throughput
STViT-R-Swin-S82.7(-0.3)5.83(-33%)717(+30%)
STViT-R-Swin-B83.2(-0.3)10.26(-33%)539(+30%)
+ +Table 3. STViT-R is evaluated on Swin-S and Swin-B on ImageNet. The top-1 accuracy, complexity in FLOPs, and throughput are reported. + +effectiveness and efficiency by employing a few semantic tokens to replace original image tokens. Our method reveals that constructing the tokens with high-level semantic representation is more important than maintaining structured tokens in ViTs. As reflected by the throughput, our method does not have overhead of memory or deployment. Compared to the total parameters, the additional parameters introduced by intra and inter-window spatial pooling is negligible (less than $0.05\%$ ), so we do not show them. + +Comparisons with existing token sparsification methods. In Table 4, we compare STViT with the state-of-the-art token sparsification methods on DeiT-S and DeiT-B. Due to different base models used by different methods, we + +
ModelTop-1 AccFLOPs(G)△
DeiT-S
DynamicViT [31]79.32.9(-37%)-0.5
IA-RED2[30]79.13.2(-30%)-0.7
PS-ViT [34]79.42.6(-43%)-0.4
TokenLearner* [32]76.11.9(-44%)-1.8
DGE* [33]79.73.1 (-49%)-0.6
A-ViT* [47]78.63.6 (-39%)-0.3
Evo-ViT [45]79.43.0(-35%)-0.4
EViT [24]78.52.3(-50%)-1.3
STViT(Ours)79.81.91(-58%)-0.0
DeiT-B
IA-RED2[30]80.311.8(-33%)-1.5
DynamicViT [31]81.311.2(-36%)-0.5
PS-ViT [34]81.59.8(-44%)-0.3
TokenLearner* [32]83.728.7(-48%)-1.1
Evo-ViT [45]81.310.2(-33%)-0.5
EViT [24]80.08.7(-51%)-1.8
STViT(Ours)81.87.31(-58%)-0.0
+ +Table 4. Comparisons with the state-of-the-art token sparsification methods on DeiT-S and DeiT-B. $\triangle$ shows the accuracy difference between each model and its base model. $*$ : their base models are not standard DeiT models. + +adopt accuracy difference between each model and its base model $\Delta$ to evaluate them for fair. Results indicate that our method achieves the lowest accuracy drop $\Delta$ with the highest FLOPs reduction, outperforming all the state-of-the-art methods in both accuracy and efficiency significantly. + +# 4.2. Video recognition + +Setting. For video recognition, we apply our STViT to Video Swin [27]. All the models are pre-trained on ImageNet-1K and trained on Kinetics-400 [7]. We generate semantic tokens from each frame as illustrated in Section 3.2. The initialization from pre-trained models and other implementation details are as same as Video Swin [27]. + +Results. The results are presented in Table 6. On Swin-T and Swin-S, STViT-Swin can save about $27\%$ FLOPs with $0.3\%$ accuracy drop, which shows that our method works on video recognition. + +# 4.3. Applications in object detection and instance segmentation + +Settings. Experiments of object detection and instance segmentation are conducted on COCO 2017 [25]. We evaluate STViT-R with Swin in Cascade Mask R-CNN [4, 17] detection frameworks. The $w_{s}$ is set to 3. The backbone models are pre-trained on ImageNet-1K and the pre-trained results are shown in Table 3. All the other settings follow Swin [26]. + +![](images/c993eb7095c4337efbb350f9418e71552b514133d0cbde4048357a4d48cac1b8.jpg) +Figure 3. Visualization example of attention maps in the first attention layer of STGM. + +Comparison to Swin Transformer. The performance of STViT-R-Swin using the Cascade Mask R-CNN framework with $3 \times$ schedule is shown in Table 5. Our method achieves better performance with more than $30\%$ FLOPs reduction for backbone on bath object detection and instance segmentation. This validates that the recovery module and dumbbell unit can restore detailed spatial information, and the global context information integrated from the semantic tokens significantly benefits object detection. Ignoring spatial structure in the intermediate layers does not affect the object detection task, which is a meaningful fact to help design efficient object detection frameworks. Another interesting finding is that our method has a remarkable improvement for small object detection which is a challenging problem in the detection community. For instance, our STViT-R-Swin-S outperforms Swin-S by $1.5\%$ on $AP_{s}^{b}$ . + +# 4.4. Ablation study + +All the following ablation study of STViT and STViT-R are conducted on the DeiT-S and Swin-S, respectively. + +Initialization analysis. The semantic tokens are the cluster centers recovered by attention layers. The initialization of cluster centers induces the representation of semantic tokens. Spatial and global initialization are adopted in Section 3.1 to guide the semantic tokens to integrate local and global semantic information separately. We compare different initialization components in Table 7. When performing single global initialization $(3^{rd}$ row), we replace the spatial initial cluster centers with global initial cluster centers in the first transformer layer. The accuracy of using a single initialization method is far lower than using both, which shows the effectiveness of our initialization strategy. Our global initial cluster centers look similar to learned positional encoding since both use random initialization. To verify their distinction, we experiment with a real learned positional encoding by additionally adding the global initial cluster centers to keys, which causes $0.1\%$ accuracy drop (the last row of Table 7). Therefore, our global initialization is different from learned positional encoding. + +We visualize the attention maps of the first attention layer in the STGM in Figure 3. Because the queries of this layer are spatial initial cluster centers, these attention maps visualize the local semantic information integration by attention. The attention layer groups the semantic information according to the position of initial cluster centers, which ensures to extract fine-grained semantic information + +
APbAPb50APb75APsbAPmAPm50APm75APsmFLOPs(G)
Swin-S51.870.456.335.244.767.948.528.8194
STViT-R-Swin-S51.870.656.136.744.767.848.629.0134(-31%)
Swin-B51.970.956.535.445.068.448.728.9343
STViT-R-Swin-B52.270.856.836.545.268.349.129.5233(-32%)
+ +Table 5. Results on COCO object detection and instance segmentation under Cascade Mask R-CNN with $3 \times$ schedule. The FLOPs are measured for backbones. + +
ModelTop-1 Acc(%)FLOPs(G)Speed
Swin-T78.888779
STViT-Swin-T78.5(-0.3)64.4(-27%)975(+25%)
Swin-S80.6166456
STViT-Swin-S80.3(-0.3)120.5(-27%)572(+25%)
+ +Table 6. Applying STViT to Video Swin (Swin-T and Swin-S) on Kinetics-400. All the models are pre-trained on ImageNet-1K. The views are $4\times 3$ . The top-1 accuracy and complexity in FLOPs are reported. Speed is evaluated by FPS. + +
SpatialGlobalLearnedTop-1 Acc(%)
✓79.4
✓78.7
✓✓79.8
✓✓79.7
+ +Table 7. Accuracy with different initialization of STViT. Spatial, Global, and Learned indicate spatial initialization, global initialization, and learned positional encoding methods, respectively. + +
No. of transformers234
Top-1 Acc(%)79.879.579.6
FLOPs(G)1.911.972.03
+ +Table 8. Performance evaluation on different numbers of transformer layers in STGM. Keeping the base module containing four transformer layers unchanged. + +
STViT-Rw/o DUReusing ST
APb51.851.451.6
APm44.744.444.5
+ +Table 9. Ablation study on STViT-R w/o dumbbell units (w/o $DU$ ) and reusing semantic tokens (Reusing ST). + +and keep the difference among semantic tokens. The attention map in the second attention layer is visualized in Supplementary A.5, which reveals that the network fixedly selects particle semantic tokens to represent global semantic information. We also show the semantic representation of image tokens in the same transformer layer on DeiT in Supplementary A.5. Compared to original image tokens, our semantic tokens in Figure 3 show more high-level semantic information. + +Number of transformer layers in STGM. Two transformer layers are adopted in the STGM by default. The ef + +fcts of employing different numbers of transformer layers are shown in Table 8. The extra layers are from the ones behind STGM to keep the total number of layers unchanged. More transformer layers do not bring improvement. + +The effectiveness of the dumbbell unit. To verify the effectiveness of our dumbbell unit, we experiment STViT-R without dumbbell units, i.e., STViT equipped with only the recovery module. We employ $6^{th}$ and $7^{th}$ transformer layers to construct STGM and the last transformer layers to construct the recovery module in Stage 3. The FLOPs is the same as the full-model STViT-R for fair comparison. The results on the COCO are reported in Table 9. Inferior results demonstrate the effectiveness of the dumbbell unit. + +Reusing semantic tokens in dumbbell units. Semantic tokens are generated in each dumbbell unit. If they are produced only once in the first dumbbell unit and reused as initial cluster centers in the subsequent dumbbell units, the result is shown in Table 9 with a slight performance drop. + +# 5. Conclusion + +In this paper, we propose a simple and effective token sparsification method, semantic token vision transformer (STViT). Our method utilizes the clustering property of self-attention to generate a few semantic tokens with high-level information representation to replace the redundant image/video tokens, which can be applied in both global and local vision transformers. By simply configuring the recovery module, our method can be successfully applied to downstream tasks. Extensive experiments demonstrate that our method achieves better accuracy along with less inference time in most cases. The success in downstream tasks significantly boosts the development of token sparsification methods. We hope that this work can inspire more future research to pay much attention to high-level semantic representation in ViTs. + +# Acknowledgement + +This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, and Mike Zheng Shou's Start-Up Grant from NUS. Computation was partially performed on resources of the National Supercomputing Centre, Singapore. Shuning was supported by Alibaba Research Intern Program. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Luvcic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2 +[2] Song Bai, Philip Torr, et al. Visual parser: Representing part-whole hierarchies with transformers. arXiv preprint arXiv:2107.05790, 2021. 3 +[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 2 +[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2 +[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7 +[8] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34, 2021. 3 +[9] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 2 +[10] Yunpeng Chen, Marcus Rohrbach, Zhicheng Yan, Yan Shuicheng, Jiashi Feng, and Yannis Kalantidis. Graph-based global reasoning networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-442, 2019. 3 +[11] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 2 +[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1601-1610, 2021. 2 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[14] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Bain- + +ing Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. arXiv preprint arXiv:2107.00652, 2021. 2 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1, 2 +[16] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3 +[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7 +[18] Zilong Huang, Youcheng Ben, Guozhong Luo, Pei Cheng, Gang Yu, and Bin Fu. Shuffle transformer: Rethinking spatial shuffle for vision transformer. arXiv preprint arXiv:2106.03650, 2021. 2 +[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3 +[20] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 2, 5 +[21] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Bin Ren, Minghai Qin, Hao Tang, and Yanzhi Wang. Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890, 2021. 3 +[22] Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient self-supervised vision transformers for representation learning. arXiv preprint arXiv:2106.09785, 2021. 2 +[23] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019. 3 +[24] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In International Conference on Learning Representations, 2022. 1, 2, 3, 7 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7 +[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: + +Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 4, 5, 7 +[27] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 7 +[28] Xuezhe Ma, Xiang Kong, Sinong Wang, Chunting Zhou, Jonathan May, Hao Ma, and Luke Zettlemoyer. Luna: Linear unified nested attention. Advances in Neural Information Processing Systems, 34:2441-2453, 2021. 3 +[29] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. arXiv preprint arXiv:2111.15668, 2021. 3 +[30] Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. Ia-red 2: Interpretability-aware redundancy reduction for vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 3, 7 +[31] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34, 2021. 3, 7 +[32] Michael S. Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 7 +[33] Lin Song, Songyang Zhang, Songtao Liu, Zeming Li, Xuming He, Hongbin Sun, Jian Sun, and Nanning Zheng. Dynamic grained encoder for vision transformers. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 1, 3, 7 +[34] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. arXiv preprint arXiv:2106.02852, 2021. 3, 7 +[35] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 2, 5 +[36] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 1, 2 +[37] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2 +[38] Pichao Wang, Xue Wang, Hao Luo, Jingkai Zhou, Zhipeng Zhou, Fan Wang, Hao Li, and Rong Jin. Scaled relu matters for training vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022. 1 +[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. + +Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 2 +[40] Wenxiao Wang, Lu Yao, Long Chen, Binbin Lin, Deng Cai, Xiaofei He, and Wei Liu. Crossformer: A versatile vision transformer hinging on cross-scale attention. arXiv preprint arXiv:2108.00154, 2021. 2 +[41] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. Advances in Neural Information Processing Systems, 34, 2021. 3 +[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8741-8750, 2021. 2 +[43] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22-31, 2021. 1, 2 +[44] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9981–9990, October 2021. 2 +[45] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1, 2, 3, 7 +[46] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers, 2021. 2 +[47] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 3, 7 +[48] Tan Yu, Gangming Zhao, Ping Li, and Yizhou Yu. Boat: Bilateral local attention vision transformer. arXiv preprint arXiv:2201.13027, 2022. 2 +[49] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 558-567, October 2021. 2 +[50] David Junhao Zhang, Kunchang Li, Yali Wang, Yunpeng Chen, Shashwat Chandra, Yu Qiao, Luoqi Liu, and Mike Zheng Shou. Morph MLP: An efficient mlp-like backbone for spatial-temporal representation learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 230-248. Springer, 2022. 2 +[51] Songyang Zhang, Xuming He, and Shipeng Yan. Latent-gnn: Learning efficient non-local relations for visual recog + +nition. In International Conference on Machine Learning, pages 7374-7383. PMLR, 2019. 3 +[52] Minghang Zheng, Peng Gao, Renrui Zhang, Kunchang Li, Xiaogang Wang, Hongsheng Li, and Hao Dong. End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315, 2020. 2 +[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2 +[54] Jingkai Zhou, Pichao Wang, Fan Wang, Qiong Liu, Hao Li, and Rong Jin. Elsa: Enhanced local self-attention for vision transformer. arXiv preprint arXiv:2112.12786, 2021. 2 +[55] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 2 +[56] Zhuofan Zong, Kunchang Li, Guanglu Song, Yali Wang, Yu Qiao, Biao Leng, and Yu Liu. Self-slimmed vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 432-448. Springer, 2022. 1, 2, 3 \ No newline at end of file diff --git a/2023/Making Vision Transformers Efficient From a Token Sparsification View/images.zip b/2023/Making Vision Transformers Efficient From a Token Sparsification View/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..48282e7d0ab93a80d14c5f80c74d6c21a5771f73 --- /dev/null +++ b/2023/Making Vision Transformers Efficient From a Token Sparsification View/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f15091a1deaf1d8d25a004fd16b155bb2a42a6332464a3455bcbf3e03561632 +size 453063 diff --git a/2023/Making Vision Transformers Efficient From a Token Sparsification View/layout.json b/2023/Making Vision Transformers Efficient From a Token Sparsification View/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..98487f9a884878d515db8235551282f2b97ce656 --- /dev/null +++ b/2023/Making Vision Transformers Efficient From a Token Sparsification View/layout.json @@ -0,0 +1,9232 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 520, + 121 + ], + "type": "text", + "content": "Making Vision Transformers Efficient from A Token Sparsification View" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "spans": [ + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": "Shuning Chang" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": " Pichao Wang" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{2\\dagger\\dagger}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": " Ming Lin" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": " Fan Wang" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": " David Junhao Zhang" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": " Rong Jin" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": " Mike Zheng Shou" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": "Show Lab, National University of Singapore " + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 141, + 504, + 186 + ], + "type": "text", + "content": "Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "text", + "content": "The quadratic computational complexity to the number of tokens limits the practical applications of Vision Transformers (ViTs). Several works propose to prune redundant tokens to achieve efficient ViTs. However, these methods generally suffer from (i) dramatic accuracy drops, (ii) application difficulty in the local vision transformer, and (iii) non-general-purpose networks for downstream tasks. In this work, we propose a novel Semantic Token ViT (STViT), for efficient global and local vision transformers, which can also be revised to serve as backbone for downstream tasks. The semantic tokens represent cluster centers, and they are initialized by pooling image tokens in space and recovered by attention, which can adaptively represent global or local semantic information. Due to the cluster properties, a few semantic tokens can attain the same effect as vast image tokens, for both global and local vision transformers. For instance, only 16 semantic tokens on DeiT-(Tiny,Small,Base) can achieve the same accuracy with more than " + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "text", + "content": " inference speed improvement and nearly " + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "text", + "content": " FLOPs reduction; on Swin-(Tiny,Small,Base), we can employ 16 semantic tokens in each window to further speed it up by around " + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "text", + "content": " with slight accuracy increase. Besides great success in image classification, we also extend our method to video recognition. In addition, we design a STViT-R(covery) network to restore the detailed spatial information based on the STViT, making it work for downstream tasks, which is powerless for previous token sparsification methods. Experiments demonstrate that our method can achieve competitive results compared to the original networks in object detection and instance segmentation, with over " + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 45, + 238, + 290, + 610 + ], + "type": "text", + "content": " FLOPs reduction for backbone." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 632, + 128, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 128, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 128, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 651, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 677 + ], + "type": "text", + "content": "In contrast to standard Convolutional Neural Networks (CNNs) approaches which process images pixel-by-pixel," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 213, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 547, + 514 + ], + "type": "text", + "content": "Vision Transformers (ViTs) [15, 26, 35, 36, 43] treat an image as a sequence of patch/image tokens, and have shown promising performance in prevalent visual recognition scenarios. However, these superior performances do not come for free: the quadratic computational complexity to the number of image tokens limits their application in practice. Previous works [33, 56] have illustrated the large amount of redundancy in the image tokens and also shown the effect of filtering out unimportant tokens normally according to predefined scoring mechanism. However, these methods face the following challenges. Firstly, the predefined scoring mechanisms for filtering are generally imprecise. In Figure 1, on the left we visualize the class token values in different layers which are commonly used to score the token importance [16, 24, 45]. Different layers have different value distributions, thus using these imprecise scores for filtering would lead to unsatisfactory performance. For example, EViT [24] has an accuracy drop of " + }, + { + "bbox": [ + 306, + 213, + 547, + 514 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 306, + 213, + 547, + 514 + ], + "type": "text", + "content": " when saving " + }, + { + "bbox": [ + 306, + 213, + 547, + 514 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 213, + 547, + 514 + ], + "type": "text", + "content": " FLOPs on DeiT-S [35]. Secondly, the remaining tokens do not distribute evenly in space any more, making them hard to work in local vision transformers1. Finally, large-scale token pruning tremendously damages the spatial structure and positional information, and causes difficulties when applied to downstream tasks, which they do not propose a solution to deal with." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 514, + 547, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 671 + ], + "type": "text", + "content": "To solve these problems, we propose Semantic Token ViT (STViT), for efficient global and local vision transformers, which also can be revised to serve as backbone for downstream tasks. The proposed approach is based on the following observations: (i) unlike local CNNs which learn spatial structure of images, vision transformer discretizes feature map as tokens for global feature exploration, relieving the requirements for maintaining the whole image structure and information; (ii) discrete tokens are more beneficial for optimization [38]; (iii) in Figure 1, on the right shows the attention maps in different transformer layers, and there are only several vertical lines in the deep layers, which means that only a few tokens with global se" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 683, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 547, + 713 + ], + "type": "text", + "content": "1In this paper, we define the vision transformer with global self-attention (like DeiT) as global vision transformer and the vision transformer with local self-attention (like Swin) as local vision transformer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 683, + 225, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 225, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 225, + 693 + ], + "type": "text", + "content": "*Work done during an internship at Alibaba Group." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 693, + 158, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 158, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 158, + 703 + ], + "type": "text", + "content": "Equal corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 703, + 265, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 265, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 265, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 58, + 703, + 265, + 712 + ], + "type": "text", + "content": "Work done at Alibaba Group, and now affiliated with Amazon." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6195" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 271, + 128 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 271, + 128 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 271, + 128 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 271, + 128 + ], + "type": "image", + "image_path": "bde36492b427be85ae4cd307425e95003502ef7a7d4b4f6e038ec4f55f7518fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 137, + 546, + 160 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 546, + 160 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 546, + 160 + ], + "type": "text", + "content": "Figure 1. Left: the attention values of class tokens (normalized and reshaped in image shape) in different self-attention layers. Right: the attention maps in different self-attention layers. Zoom-in for better visibility." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 324, + 72, + 516, + 128 + ], + "blocks": [ + { + "bbox": [ + 324, + 72, + 516, + 128 + ], + "lines": [ + { + "bbox": [ + 324, + 72, + 516, + 128 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 516, + 128 + ], + "type": "image", + "image_path": "f30fdba2db5b29b8a686a45071e5d10b6580de568ef44123696b1d30204ad83c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 288, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 288, + 232 + ], + "type": "text", + "content": "matic information matter. Thus, we argue that it is not necessary to maintain massive structured tokens for ViTs, especially in the deep layers. Employing a few discrete tokens with high-level semantic information can potentially achieve both high performance and efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 233, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 288, + 616 + ], + "type": "text", + "content": "In STViT, the semantic tokens represent the cluster centers, and the number of them is far less than the original image tokens, significantly reducing the computational cost. Inspired by the fact that multi-head attention can conduct the cluster center recovery (Supplementary A.7), we only employ the off-the-shelf self-attention to generate the semantic tokens. Specifically, the first few transformer layers are kept unchanged to obtain the image tokens with low-level features. The image tokens are then fed into our semantic token generation module (STGM) consisting of at least two transformer layers to generate semantic tokens. In each self-attention layer, the semantic tokens are input as queries, and the image tokens are fed as keys and values. The semantic tokens dynamically aggregate image tokens through the attention layers to recover cluster centers. In the first attention layer, the semantic tokens are initialized by an intra and inter-window spatial pooling which takes into account incorporating semantic information in each window and maximizing distance between adjacent windows. Thanks to this spatial initialization, the semantic tokens mainly incorporate local semantic information and achieve discrete and uniform distribution in space. In the following attention layer, besides further clustering, the semantic tokens are equipped with global cluster centers, and the network can adaptively select partial semantic tokens to focus on global semantic information. After the STGM, the original image tokens are discarded, and only semantic tokens are kept for the subsequent transformer layers. Because the generation of semantic tokens is flexible and space-aware, our method can be plugged into both global and local vision transformers. The semantic tokens can be produced in each window for the local vision transformer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "Another property of STViT is its capability to serve as a backbone for downstream tasks, such as object detection and instance segmentation. Discussions have been missing in previous methods [16, 24, 32, 45, 56] about how to use them in downstream task under the massive loss of spatial information during the token sparsification process, which actually seriously impedes the application of their method. Instead, we design a novel STViT-R network based" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 171, + 546, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 546, + 231 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 546, + 231 + ], + "type": "text", + "content": "on STViT where a recovery module and dumbbell unit are adopted to periodically restore the full resolution feature map while the intermediate transformer layers continue to use semantic tokens to save computation cost, making our method work in downstream task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 233, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 547, + 461 + ], + "type": "text", + "content": "The effectiveness of the proposed method is validated via a comprehensive empirical study on image and video ViT models. Only 16 semantic tokens on DeiT-(Tiny, Small, Base) achieve nearly " + }, + { + "bbox": [ + 304, + 233, + 547, + 461 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 233, + 547, + 461 + ], + "type": "text", + "content": " inference time reduction without any accuracy degradation; on Swin-(Tiny, Small, Base), we also improve the inference throughput by nearly " + }, + { + "bbox": [ + 304, + 233, + 547, + 461 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 233, + 547, + 461 + ], + "type": "text", + "content": " with slight accuracy increase. Moreover, the proposed STViT-R achieves promising results on object detection and instance segmentation. To the best of our knowledge, this is one of first works to apply the token sparsification algorithm in local vision transformers, and use the ViTs as backbones in downstream tasks after large-scale token pruning. Our findings in ViTs uncover that maintaining the full-size feature map is unnecessary, and a few tokens with high-level semantic representations can achieve both high performance and efficiency. Thanks to its simplicity and general-purpose ability, our method can also serve as a new efficient ViT baseline architecture and a starting point for further research from the token sparsification perspective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 476, + 390, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 390, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 390, + 488 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "Vision transformers. Vision Transformer (ViT) [15] first introduces a pure Transformer backbone for image classification. ViT variants further inspire the applications of transformer to various vision tasks beyond image/video classification [1, 3, 20, 35, 36, 39, 43, 44, 49, 50], such as object detection [5, 12, 52, 55], semantic segmentation [37, 42, 53], and self-supervised learning [6, 9, 22]. Vanilla transformers have high computational and memory costs because the multi-head self-attention has quadratic computational complexity to the number of image tokens. Recently, various efficient ViTs have been proposed to alleviate this issue. The existing methods mainly focus on reducing the complexity of self-attention or reducing the number of tokens. Swin Transformer [26] adopts local self-attention, i.e., attending neighboring tokens within a constant window size, achieving a linear computational complexity in the self-attention with high performance. Many subsequent works [11, 14, 18, 40, 46, 48, 54] follow the local" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6196" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "self-attention design to develop variants. Token sparsification [8, 16, 21, 24, 29–34, 41, 45, 47, 56] also attracts increasing attention." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "text", + "content": "Token sparsification. Token sparsification methods can be mainly categorized into hard pruning [8,16,21,24,29-31, 33,34,45,47] and soft pruning [32,56]. Hard pruning methods filter out some unimportant tokens according to a predefined scoring mechanism. DynamicViT [31], SPViT [21], and AdaViT [29] introduce additional prediction networks to score the tokens. Evo-ViT [45], ATS [16], and EViT [24] utilize the values of class tokens to evaluate the importance of tokens. However, it is difficult to achieve precise scoring as shown in the left of Figure 1. Therefore, they usually suffer from a significant accuracy drop. For instance, EViT [24] has an accuracy drop of " + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "text", + "content": " when saving " + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "text", + "content": " FLOPs on DeiT-S. Soft pruning methods generate new tokens from image tokens by importing additional attention networks. TokenLearner [32] also argue for a few tokens to replace image tokens. However, its price is a " + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "text", + "content": " accuracy drop when reducing " + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 47, + 128, + 289, + 415 + ], + "type": "text", + "content": " FLOPs, which is far inferior to concurrent works. Besides performance degradation, previous methods also have the following disadvantages. First, whether or how to extend the methods to local vision transformers remains unexplored. Second, it has not been discussed about how to serve the downstream tasks like object detection and instance segmentation after the tokens are pruned." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 415, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 523 + ], + "type": "text", + "content": "In our method, we apply the off-the-shelf transformer layers to reduce token number. [2, 10, 19, 23, 28, 51] adopt similar approaches to achieve efficient non-local relationships. Our method is different from them as below: (i) our method extracts local semantic information instead of non-local relationships; (ii) the semantic tokens are a few cluster centers, which can replace the massive image tokens to achieve image classification; (iii) our method specializes in pruning tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 533, + 103, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 103, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 103, + 546 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "content": "The proposed STViT is presented in this section, which aims to construct an efficient and high-performance ViT. STViT is first introduced in Section 3.1, followed by how to apply STViT in the local vision transformer in Section 3.2. Based on STViT, STViT-R is developed to restore the spatial resolution for downstream tasks in Section 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 634, + 101, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 101, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 101, + 645 + ], + "type": "text", + "content": "3.1.STViT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Overall architecture. An overview of STViT architecture is presented in Figure 2a. The patch embedding layer and shallow transformer layers are kept unchanged as a base module in our method. The base module copes with all the image tokens " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N_i \\times C}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " to extract low-level features," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " is the number of image tokens and " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " is the number of channels. The image tokens are fed into the semantic token generation module (STGM) to generate " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " semantic tokens " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N_s \\times C}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". After the STGM, the image tokens " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " can be discarded, and only semantic tokens " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " with high-level semantic information are used in all the subsequent transformers. Due to " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "N_{s} \\ll N_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", our method can significantly reduce the computational cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 186, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 270 + ], + "type": "text", + "content": "Semantic token generation module (STGM). The whole image is represented by a few tokens with high-level semantic information through clustering. Inspired by the fact that self-attention can conduct cluster center recovery (Supplementary A.7), we adopt the off-the-shelf self-attention layers to produce the semantic tokens. The STGM consists of at least two transformer layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "text", + "content": "The initial cluster centers " + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N_s \\times C}" + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "text", + "content": " are generated by an spatial pooling which pools the image tokens into fixed " + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "inline_equation", + "content": "w_{s} \\times w_{s}" + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "text", + "content": " tokens, with " + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "inline_equation", + "content": "N_{s} = w_{s} \\times w_{s}" + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "text", + "content": " is generally set as 4. The spatial pooling can be achieved by a non-parameterized adaptive spatial pooling or a super lightweight network with higher performance, intra and inter-window spatial pooling. The intentions of spatial pooling initialization are three folds. First, the initial cluster centers can distribute uniformly in space, making the generated semantic tokens more discrete and preventing the semantic tokens from collapsing to one point in the following layers. Second, the semantic tokens can be forced to represent more local and distinguished features. Finally, the representation of semantic tokens is associated with the specific spatial locations, which is the basis to allow our method to be applied in local self-attention and downstream tasks. The initial cluster centers " + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "text", + "content": " then dynamically integrate the image tokens " + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 270, + 546, + 510 + ], + "type": "text", + "content": " according to semantic information by attention mechanism. In the first transformer layer, the processing of the generation of semantic tokens can be written as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 513, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 513, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 319, + 513, + 545, + 526 + ], + "type": "interline_equation", + "content": "\\hat {S} ^ {1} = M H A (P, X, X) + P, \\quad S ^ {1} = F F N (\\hat {S} ^ {1}) + \\hat {S} ^ {1}, \\tag {1}", + "image_path": "39becf59ca40f54e1d89f5b8c71b995ec99abe83a640f994ea6a6ad1d62b622e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "type": "inline_equation", + "content": "MHA" + }, + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "type": "inline_equation", + "content": "FFN" + }, + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "type": "text", + "content": " are short for multi-head attention layer and feed-forward network, respectively, and the triplet input of " + }, + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "type": "inline_equation", + "content": "MHA" + }, + { + "bbox": [ + 304, + 528, + 545, + 684 + ], + "type": "text", + "content": " are queries, keys, and values in turn. All the norm layers are omitted in all the equations for brevity. The initial cluster centers are produced in a window by an adaptive spatial pooling layer or an intra and inter-window spatial pooling, while the semantic tokens are generated from a global receptive field by a dynamic attention layer to ensure that they can extract high-level semantic representation. In order to further strengthen the clustering effect, we use the second transformer layer to repeat the clustering operation and guide the semantic tokens to extract global information. In this transformer, the semantic tokens are updated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 686, + 545, + 704 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 686, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 317, + 686, + 545, + 704 + ], + "type": "interline_equation", + "content": "\\hat {S} ^ {2} = M H A \\left(S ^ {1} + G, \\operatorname {C o n c a t} \\left(S ^ {1}, X\\right), \\operatorname {C o n c a t} \\left(S ^ {1}, X\\right)\\right) + S ^ {1}, \\tag {2}", + "image_path": "9828b182ae4c61c5c9614de97d13eef593befab2b195539dcf174f452ed98cdb.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 376, + 703, + 463, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 703, + 463, + 714 + ], + "spans": [ + { + "bbox": [ + 376, + 703, + 463, + 714 + ], + "type": "interline_equation", + "content": "S ^ {2} = F F N (\\hat {S} ^ {2}) + \\hat {S} ^ {2},", + "image_path": "cd7d6f6ff10f16d71669750fa72f2cd13792fcb9ab568be47a0b78ed2033fbba.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "6197" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 69, + 296, + 177 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 296, + 177 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 296, + 177 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 296, + 177 + ], + "type": "image", + "image_path": "4d7fd0c0fdcb716fed6e63a203d651069a48e880cac0e37f31f5b9202b34a222.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 178, + 203, + 187 + ], + "lines": [ + { + "bbox": [ + 168, + 178, + 203, + 187 + ], + "spans": [ + { + "bbox": [ + 168, + 178, + 203, + 187 + ], + "type": "text", + "content": "(a) STViT" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 358, + 73, + 507, + 176 + ], + "blocks": [ + { + "bbox": [ + 358, + 73, + 507, + 176 + ], + "lines": [ + { + "bbox": [ + 358, + 73, + 507, + 176 + ], + "spans": [ + { + "bbox": [ + 358, + 73, + 507, + 176 + ], + "type": "image", + "image_path": "f17100edef88e4795d9c0d37690577ea35a91fff9e7719a10b5fcfb07508a2e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 178, + 454, + 187 + ], + "lines": [ + { + "bbox": [ + 412, + 178, + 454, + 187 + ], + "spans": [ + { + "bbox": [ + 412, + 178, + 454, + 187 + ], + "type": "text", + "content": "(b) STViT-R" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 193, + 398, + 204 + ], + "lines": [ + { + "bbox": [ + 194, + 193, + 398, + 204 + ], + "spans": [ + { + "bbox": [ + 194, + 193, + 398, + 204 + ], + "type": "text", + "content": "Figure 2. The architectures of our STViT and STViT-R." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{N_s \\times C}" + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "text", + "content": " is the global cluster centers initialized by Gaussian noise and " + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\text{Concat}(\\cdot)" + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "text", + "content": " is a concatenation operation. The global cluster centers " + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "text", + "content": " are responsible for global semantic information extraction like the class token. Although " + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "inline_equation", + "content": "S^1" + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "text", + "content": " are summed together as the queries, they can be decoupled in the attention computation as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 286, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 287, + 312 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} A _ {s} = \\left(S ^ {1} \\cdot W _ {q}\\right) \\cdot \\left(\\left(S ^ {1} + X\\right) \\cdot W _ {k}\\right), \\quad A _ {g} = \\left(G \\cdot W _ {q}\\right) \\cdot \\left(\\left(S ^ {1} + X\\right) \\cdot W _ {k}\\right), \\tag {3} \\\\ A = \\operatorname {S o f t m a x} (A _ {s} + A _ {g}), \\\\ \\end{array}", + "image_path": "d8abd22630f82585d31574371162069a9755e7c608b4af814fb8488de8630434.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": " are the linear projection weights of queries and keys. We can see that " + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "inline_equation", + "content": "S^{1}" + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": " integrate the keys to generate " + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "inline_equation", + "content": "A_{s}" + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "inline_equation", + "content": "A_{g}" + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": " independently and just share a Softmax operation to produce the final attention map " + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": ". Therefore, besides spatial semantic information, the semantic tokens also incorporate global semantic information with a negligible additional overhead. Though very similar, the global cluster centers are actually different from the learned positional encoding. We do not add the global cluster centers to keys. Moreover, it will be shown in the experiments that inserting the actual learned positional encoding will cause an accuracy drop. The number of transformer layers in STGM is flexible. More transformer layers can be associated for clustering. Two transformer layers are employed by default, i.e., " + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "inline_equation", + "content": "S^{2}" + }, + { + "bbox": [ + 46, + 315, + 288, + 507 + ], + "type": "text", + "content": " is the output of the STGM. Note that the image tokens are not updated in the STGM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": "Intra and inter-window spatial pooling. Give " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": " feature map " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "N_{s}(w_{s} \\times w_{s})" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": " initial cluster centers. We uniformly split the " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "w_{s} \\times w_{s}" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": " windows " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "[X_{w}^{i}]_{i=1}^{N_{s}}" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\frac{H}{w_{s}} \\times \\frac{W}{w_{s}}" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": " and each window generates one initial cluster center. To represent abundant semantic information, we take into account intra and interwindow relations, i.e., integrating important tokens in the window and maximizing the distance among initial cluster centers in different windows. Specifically, we formulate an intra-window function " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "f_{intra}" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": " to produce a mask, " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "M_{i} = f_{intra}(X_{w}^{i})" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": ", which projects the input window from " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\mathbb{R}\\frac{H}{w_{s}} \\times \\frac{W}{w_{s}} \\times C \\rightarrow \\mathbb{R}\\frac{H}{w_{s}} \\times \\frac{W}{w_{s}}" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": ". The idea is to let the intra-window function " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "f_{intra}" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": " adaptively select a combination of informative tokens in " + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "inline_equation", + "content": "X_{w}^{i}" + }, + { + "bbox": [ + 47, + 529, + 287, + 699 + ], + "type": "text", + "content": ", which is implemented by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 702, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 287, + 714 + ], + "type": "interline_equation", + "content": "M _ {i} = \\operatorname {C o n v} \\left(\\operatorname {G e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\operatorname {D e p t h w i s e} \\operatorname {C o n v} \\left(X _ {w} ^ {i}\\right)\\right)\\right)\\right). \\tag {4}", + "image_path": "13ba18b0fa4bccb3e73bbcbd09fb595fccac18c1a0b862f816b162ffba9e51b1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": "Then, we compute each integrated token " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\hat{P}_i =" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " Softmax " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "(M_i)\\cdot X_w^i" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " from each window, and arrange " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "[\\hat{P}_i]_{i = 1}^{N_s}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " by spatial structure to form the 2D tensor " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\hat{P}\\in \\mathbb{R}^{w_s\\times w_s\\times C}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " We adopt an inter-window function " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "f_{inter}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " to compute the inter-window relations and generate the offset, " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "O =" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "f_{inter}(\\hat{P})" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " ,to revise the mask " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " . The implementation of " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "f_{inter}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " is similar to Eq. 4, except the mapping input from " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{w_s\\times w_s\\times C}\\rightarrow \\mathbb{R}^{w_s\\times w_s\\times \\frac{HW}{w_s^2}}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\frac{HW}{w_s^2}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " is the number of tokens in each window. For each window, the corresponding " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "O_{i}\\in \\mathbb{R}^{\\frac{HW}{w_{s}^{2}}}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " is sliced from " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " and reshaped to " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{H}{w_s}\\times \\frac{W}{w_s}}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " The " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " is used to revise " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " . The final initial cluster center " + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 305, + 210, + 547, + 365 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 367, + 545, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 367, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 358, + 367, + 545, + 380 + ], + "type": "interline_equation", + "content": "P _ {i} = \\operatorname {S o f t m a x} \\left(M _ {i} + O _ {i}\\right) \\cdot X _ {w}. \\tag {5}", + "image_path": "6ab7bff03943b47f29eea8b403a5f1685e7a3582f9dcaf69b73f39ce4f486762.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "type": "text", + "content": "Our " + }, + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "type": "inline_equation", + "content": "f_{\\text{intra}}" + }, + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "type": "inline_equation", + "content": "f_{\\text{intra}}" + }, + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "type": "text", + "content": " are super lightweight and the introducing parameters can be negligible. For example, on DeiT, the parameters only increase by " + }, + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "type": "inline_equation", + "content": "0.05\\%" + }, + { + "bbox": [ + 305, + 383, + 545, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 430, + 493, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 493, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 493, + 441 + ], + "type": "text", + "content": "3.2. STViT in local vision transformers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": "Local self-attention has been widely used in current ViT models to balance efficiency and accuracy. As the generation of semantic tokens in STViT is flexible in space, it can be naturally applied in local self-attention. Suppose each local self-attention layer contains " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "N_{w}" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " windows with size " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "w \\times w" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": ", we initialize " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "w_{s} \\times w_{s}" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " cluster centers in each window by our intra and inter-window spatial pooling. The total number of semantic tokens " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " would be " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "w_{s} \\times w_{s} \\times N_{w}" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " are set as 7 and 3 by default separately. As a result, our method compresses more than " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " image tokens in local self-attention. In the STGM, although initial cluster centers are from " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "w \\times w" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " windows, we use larger windows with size " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "w_{k} \\times w_{k}" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " to obtain keys and values in Eq. 1 and Eq. 2 to mitigate the effect of limited window size. Other operations in STGM are kept the same as Section 3.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "In the local ViT models, each local transformer layer is normally followed by a cross-window connection layer, such as a shift window transformer layer following a local transformer layer on Swin Transformer [26]. In our method, the attention is computed within " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{s} \\times w_{s}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " window in the local self-attention layer, and the cross-window connection can be achieved by computing self-attention in a larger-size" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6198" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "(e.g., 4 \\times w_{s})" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " sliding window because of the rare number of tokens in each window. For the low-resolution input, our cross-window connection layer is equal to a global self-attention layer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 128, + 204, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 204, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 204, + 140 + ], + "type": "text", + "content": "3.3. STViT for downstream tasks" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 144, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 263 + ], + "type": "text", + "content": "Our method significantly reduces the computation cost by using a small number of semantic tokens, while its side effect is losing nearly all the detailed position information. High-level vision tasks, such as object detection and instance segmentation, are difficult to be executed on this extremely incomplete feature map. This issue also exists in previous works, which hinders the application of token sparsification methods. To solve this issue, we design a STViTR network based on STViT to restore the original spatial resolution from the semantic tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 265, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 324 + ], + "type": "text", + "content": "Our STViT-R shown in Figure 2b has two modifications compared with STViT. First, we adopt a recovery module to restore the spatial resolution from semantic tokens; second, we regroup the transformer layers and construct dumbbell units composed of our STViT-R." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "content": "Recovery module. In the recovery module, only the self-attention layer is employed to restore the spatial resolution without any additional networks. The image tokens " + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "content": " and semantic tokens " + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "content": " are partitioned as " + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "inline_equation", + "content": "N_w^r" + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "content": " windows of size " + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "inline_equation", + "content": "w^r \\times w^r" + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "inline_equation", + "content": "w_s^r \\times w_s^r" + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "content": ", respectively. The image tokens in each window aggregate the semantic tokens in the corresponding window, which is represented as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 286, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 286, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 286, + 439 + ], + "type": "interline_equation", + "content": "\\hat {X} = M H A (X, S, S) + X, \\quad X = F F N (\\hat {X}) + \\hat {X}. \\tag {6}", + "image_path": "afc5c9fcb6f0b5adf5e5da9f6bc9a28a6aed074390b335a15e42d0b1da048fa1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 441, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 477 + ], + "type": "text", + "content": "This is a reverse operation of the generation of semantic tokens, using high-level semantic information to boost the image tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 491, + 287, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 718 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 718 + ], + "type": "text", + "content": "Dumbbell unit. The transformer layers are regrouped into multiple dumbbell units in our STViT-R. Each dumbbell unit consists of four parts. The transformers in the first part are responsible for coping with image tokens; the second part is the semantic token generation module; the transformer layers in the third part deal with semantic tokens; the last part is the recovery module. Take the application on Swin-S (STViT-R-Swin-S) as an example. One, two, two and one transformer layers are allocated for these four parts, respectively. In total, each dumbbell unit is composed of 6 transformer layers. We concatenate three dumbbell units in Stage 3 of Swin-S. In each dumbbell unit, the intermediate transformer layers process semantic tokens with high-level semantic information to save computational cost, and the complete spatial resolution is recovered at the end. By repeating multiple dumbbell units, the detailed spatial information will be preserved by the network, which can not only enhance the classification but also serve downstream tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 91, + 545, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 150 + ], + "type": "text", + "content": "STViT will first be applied in two representative ViT models, DeiT [35] and Swin [26] for image classification and video recognition. To validate that our method is effective in downstream tasks, STViT-R is then performed on object detection and instance segmentation tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 153, + 421, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 153, + 421, + 165 + ], + "spans": [ + { + "bbox": [ + 306, + 153, + 421, + 165 + ], + "type": "text", + "content": "4.1. Image classification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "content": "Settings. For image classification, all the models are trained on the ImageNet [13] with 1.28M training images and 50K validation images from 1,000 classes. By default, the semantic token generation module (STGM) employs the " + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "inline_equation", + "content": "6^{th}" + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "content": " transformer layers of DeiT (with 12 layers in total), employs the " + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "inline_equation", + "content": "3^{th}" + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "content": " transformer layers of Stage 3 of Swin-T (with 12 layers in total), and employs the " + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "inline_equation", + "content": "11^{th}" + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "inline_equation", + "content": "12^{th}" + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "content": " transformer layers of Stage 3 of Swin-S and Swin-B (with 24 layers in total). The image resolution in training and inference is " + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 171, + 545, + 421 + ], + "type": "text", + "content": " unless otherwise specified. The batch size is 1,024. All the models are trained from scratch for 300 epochs, and the augmentation and regularization strategies follow the original papers of DeiT and Swin. No knowledge distillation algorithms are used in our experiments. The classification is performed by applying a global average pooling layer on the output tokens of the last transformer layer, followed by a linear classifier. In evaluation, the top-1 accuracy using a single crop is reported. The FLOPs computations of this paper are measured by Fvcore². Throughput is measured with the batch size of 128 on a V100 GPU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 422, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 494 + ], + "type": "text", + "content": "On Swin [26], the semantic tokens are generated in Stage 3, and they are not downsampled in Stage 4 due to rare semantic tokens. The patch merging layer between Stage 3 and Stage 4 is replaced with a simple linear layer to double the number of channels. " + }, + { + "bbox": [ + 305, + 422, + 545, + 494 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 305, + 422, + 545, + 494 + ], + "type": "text", + "content": " is set as 10 and 14 for two transformer layers of STGM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "type": "text", + "content": "Results. One of the advantages of our method is that it can be applied to both global and local vision transformers to reduce computational complexity. Our main results on DeiT and Swin are summarized in Table 1 and Table 2, respectively. The results of LV-ViT [20] are illustrated in Supplementary A.3 due to limited space. We report the top-1 accuracy, FLOPs, and the throughput under different numbers of semantic tokens. On DeiT, the models with 16 semantic tokens achieve the same accuracy as the DeiT models with 196 tokens and save nearly " + }, + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "type": "text", + "content": " FLOPs on DeiT-T, DeiT-S, and DeiT-B. With more semantic tokens, the accuracy can consistently outperform the base models. For instance, STViT-DeiT-B with 36 semantic tokens surpasses DeiT-B by " + }, + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "type": "text", + "content": " accuracy with " + }, + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "type": "inline_equation", + "content": "52\\%" + }, + { + "bbox": [ + 305, + 503, + 545, + 670 + ], + "type": "text", + "content": " FLOPs reduction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 671, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 545, + 694 + ], + "type": "text", + "content": "The local vision transformer like Swin is already an efficient architecture compared to the global vision trans" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 525, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 525, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 525, + 712 + ], + "type": "text", + "content": "2https://github.com/facebookresearch/fvcore" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6199" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 70, + 504, + 215 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 504, + 215 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 504, + 215 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 504, + 215 + ], + "type": "table", + "html": "
ModelMetricsBaseNo. of semantic tokens
163664100
STViT-DeiT-TTop-1 Acc(%)72.272.2(+0.0%)72.7(+0.5)73.0(+0.8)73.2(+1.0)
FLOPs(G)1.260.53(-58%)0.60(-52%)0.71(-44%)0.86(-32%)
Throughput(img/s)27525511(+101%)4769(+74%)4214(+53%)3551(+29%)
STViT-DeiT-STop-1 Acc(%)79.879.8(+0.0)80.1(+0.3)80.5(+0.7)80.6(+0.8)
FLOPs(G)4.581.91(-58%)2.20(-52%)2.62(-43%)3.16(-31%)
Throughput(img/s)14082891(+105%)2542(+80%)2229(+58%)1837(+30%)
STViT-DeiT-BTop-1 Acc(%)81.881.8(+0.0)82.2(+0.4)82.6(+0.8)82.7(+0.9)
FLOPs(G)17.587.31(-58%)8.44(-52%)10.04(-43%)12.13(-31%)
Throughput(img/s)6261308(+110%)1150(+85%)1087(+61%)826(+33%)
", + "image_path": "3539e1609dbd5b3149f9e70af42333b4553874fc6d39866365a45a0ac954607e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 92, + 257, + 502, + 401 + ], + "blocks": [ + { + "bbox": [ + 46, + 224, + 546, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 546, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 546, + 246 + ], + "type": "text", + "content": "Table 1. Applying STViT to DeiT-T, DeiT-S, and DeiT-B. The top-1 accuracy, complexity in FLOPs, and throughput are reported for different numbers of semantic tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 257, + 502, + 401 + ], + "lines": [ + { + "bbox": [ + 92, + 257, + 502, + 401 + ], + "spans": [ + { + "bbox": [ + 92, + 257, + 502, + 401 + ], + "type": "table", + "html": "
ModelMetricsBaseMove STGMNo. of semantic tokens
4916
STViT-Swin-TTop-1 Acc(%)81.381.0(-0.3%)80.8(-0.5)81.5(+0.2)81.8(+0.5%)
FLOPs(G)4.53.14(-30%)2.99(-34%)3.43(-24%)4.06(-10%)
Throughput(img/s)8781124(+29%)1128(+29%)1061(+22%)1008(+15%)
STViT-Swin-STop-1 Acc(%)83.082.8(-0.2%)82.4(-0.6%)83.0(-0.0)83.1(+0.1%)
FLOPs(G)8.75.95(-32%)5.95(-32%)6.53(-25%)7.36(-15%)
Throughput(img/s)551739(+35%)732(+34%)691(+26%)657(+20%)
STViT-Swin-BTop-1 Acc(%)83.583.2(-0.3%)83.0(-0.5)83.4(-0.1)83.7(+0.2%)
FLOPs(G)15.410.48(-32%)10.48(-32%)11.51(-25%)12.97(-16%)
Throughput(img/s)415558(+35%)551(+33%)521(+26%)489(+19%)
", + "image_path": "85cf7519e0a9830ba3e8a416b696647a210698ab8c9c9bb83727f9a692775df7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "content": "former, so the reduction of FLOPs on Swin models are smaller than on DeiT models. When 9 semantic tokens are used in each window, STViT-Swin models can reduce " + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "content": " FLOPs with negligible accuracy loss on all the model sizes. If the number of tokens is reduced to 4 in each window (16 in total), a significant accuracy drop will occur, which indicates that local vision transformers need more semantic tokens than global vision transformers. We can move the STGM towards shallow layers to attain a better complexity/accuracy trade-off. In Table 2, STGM is moved by one transformer layer on STViT-Swin-T and by two layers on STViT-Swin-S and STViT-Swin-B to save over " + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "content": " FLOPs with only about " + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "content": " accuracy drops (column of \"Move STGM\")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 626, + 287, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 287, + 698 + ], + "type": "text", + "content": "STViT-R equipped with recovery modules is designed for downstream tasks, while it also can perform image classification. The corresponding results are reported in Table 3. On both Swin-S and Swin-B, STViT-R can save " + }, + { + "bbox": [ + 46, + 626, + 287, + 698 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 46, + 626, + 287, + 698 + ], + "type": "text", + "content": " FLOPs with " + }, + { + "bbox": [ + 46, + 626, + 287, + 698 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 626, + 287, + 698 + ], + "type": "text", + "content": " accuracy drop. The hyper-parameters we used in STGM are as same as STViT-Swin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "content": "These results demonstrate that our method achieves both" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 455, + 545, + 499 + ], + "blocks": [ + { + "bbox": [ + 46, + 411, + 546, + 444 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 546, + 444 + ], + "type": "text", + "content": "Table 2. Applying STViT to Swin-T, Swin-S, and Swin-B. The top-1 accuracy, complexity in FLOPs, and throughput are reported for different numbers of semantic tokens in each window. Base indicates the corresponding original Swin model. Move STGM indicates changing the default position of STGM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 455, + 545, + 499 + ], + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 499 + ], + "type": "table", + "html": "
ModelTop-1 Acc(%)FLOPs(G)Throughput
STViT-R-Swin-S82.7(-0.3)5.83(-33%)717(+30%)
STViT-R-Swin-B83.2(-0.3)10.26(-33%)539(+30%)
", + "image_path": "e1a7b7842dae39f4ac396fd271f45876b45d6a38410bd46c499503b36b0344a9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 504, + 545, + 537 + ], + "lines": [ + { + "bbox": [ + 305, + 504, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 545, + 537 + ], + "type": "text", + "content": "Table 3. STViT-R is evaluated on Swin-S and Swin-B on ImageNet. The top-1 accuracy, complexity in FLOPs, and throughput are reported." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 545, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 654 + ], + "type": "text", + "content": "effectiveness and efficiency by employing a few semantic tokens to replace original image tokens. Our method reveals that constructing the tokens with high-level semantic representation is more important than maintaining structured tokens in ViTs. As reflected by the throughput, our method does not have overhead of memory or deployment. Compared to the total parameters, the additional parameters introduced by intra and inter-window spatial pooling is negligible (less than " + }, + { + "bbox": [ + 304, + 545, + 545, + 654 + ], + "type": "inline_equation", + "content": "0.05\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 654 + ], + "type": "text", + "content": "), so we do not show them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Comparisons with existing token sparsification methods. In Table 4, we compare STViT with the state-of-the-art token sparsification methods on DeiT-S and DeiT-B. Due to different base models used by different methods, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6200" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 277, + 305 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 277, + 305 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 277, + 305 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 277, + 305 + ], + "type": "table", + "html": "
ModelTop-1 AccFLOPs(G)△
DeiT-S
DynamicViT [31]79.32.9(-37%)-0.5
IA-RED2[30]79.13.2(-30%)-0.7
PS-ViT [34]79.42.6(-43%)-0.4
TokenLearner* [32]76.11.9(-44%)-1.8
DGE* [33]79.73.1 (-49%)-0.6
A-ViT* [47]78.63.6 (-39%)-0.3
Evo-ViT [45]79.43.0(-35%)-0.4
EViT [24]78.52.3(-50%)-1.3
STViT(Ours)79.81.91(-58%)-0.0
DeiT-B
IA-RED2[30]80.311.8(-33%)-1.5
DynamicViT [31]81.311.2(-36%)-0.5
PS-ViT [34]81.59.8(-44%)-0.3
TokenLearner* [32]83.728.7(-48%)-1.1
Evo-ViT [45]81.310.2(-33%)-0.5
EViT [24]80.08.7(-51%)-1.8
STViT(Ours)81.87.31(-58%)-0.0
", + "image_path": "71f86540cfa691750bd3beb9e4abba72e9b396337d2aa719c673c0f80c6f0f58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "type": "text", + "content": "Table 4. Comparisons with the state-of-the-art token sparsification methods on DeiT-S and DeiT-B. " + }, + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "type": "text", + "content": " shows the accuracy difference between each model and its base model. " + }, + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 46, + 312, + 288, + 357 + ], + "type": "text", + "content": ": their base models are not standard DeiT models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 366, + 287, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 428 + ], + "type": "text", + "content": "adopt accuracy difference between each model and its base model " + }, + { + "bbox": [ + 46, + 366, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 366, + 287, + 428 + ], + "type": "text", + "content": " to evaluate them for fair. Results indicate that our method achieves the lowest accuracy drop " + }, + { + "bbox": [ + 46, + 366, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 366, + 287, + 428 + ], + "type": "text", + "content": " with the highest FLOPs reduction, outperforming all the state-of-the-art methods in both accuracy and efficiency significantly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 437, + 153, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 153, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 153, + 451 + ], + "type": "text", + "content": "4.2. Video recognition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 457, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 531 + ], + "type": "text", + "content": "Setting. For video recognition, we apply our STViT to Video Swin [27]. All the models are pre-trained on ImageNet-1K and trained on Kinetics-400 [7]. We generate semantic tokens from each frame as illustrated in Section 3.2. The initialization from pre-trained models and other implementation details are as same as Video Swin [27]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 594 + ], + "type": "text", + "content": "Results. The results are presented in Table 6. On Swin-T and Swin-S, STViT-Swin can save about " + }, + { + "bbox": [ + 46, + 545, + 288, + 594 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 594 + ], + "type": "text", + "content": " FLOPs with " + }, + { + "bbox": [ + 46, + 545, + 288, + 594 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 594 + ], + "type": "text", + "content": " accuracy drop, which shows that our method works on video recognition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 604, + 287, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 287, + 630 + ], + "type": "text", + "content": "4.3. Applications in object detection and instance segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 636, + 288, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 288, + 721 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 288, + 721 + ], + "type": "text", + "content": "Settings. Experiments of object detection and instance segmentation are conducted on COCO 2017 [25]. We evaluate STViT-R with Swin in Cascade Mask R-CNN [4, 17] detection frameworks. The " + }, + { + "bbox": [ + 46, + 636, + 288, + 721 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 46, + 636, + 288, + 721 + ], + "type": "text", + "content": " is set to 3. The backbone models are pre-trained on ImageNet-1K and the pre-trained results are shown in Table 3. All the other settings follow Swin [26]." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 547, + 121 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 547, + 121 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 547, + 121 + ], + "type": "image", + "image_path": "c993eb7095c4337efbb350f9418e71552b514133d0cbde4048357a4d48cac1b8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 129, + 545, + 150 + ], + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 150 + ], + "type": "text", + "content": "Figure 3. Visualization example of attention maps in the first attention layer of STGM." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "text", + "content": "Comparison to Swin Transformer. The performance of STViT-R-Swin using the Cascade Mask R-CNN framework with " + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "text", + "content": " schedule is shown in Table 5. Our method achieves better performance with more than " + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "text", + "content": " FLOPs reduction for backbone on bath object detection and instance segmentation. This validates that the recovery module and dumbbell unit can restore detailed spatial information, and the global context information integrated from the semantic tokens significantly benefits object detection. Ignoring spatial structure in the intermediate layers does not affect the object detection task, which is a meaningful fact to help design efficient object detection frameworks. Another interesting finding is that our method has a remarkable improvement for small object detection which is a challenging problem in the detection community. For instance, our STViT-R-Swin-S outperforms Swin-S by " + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "inline_equation", + "content": "AP_{s}^{b}" + }, + { + "bbox": [ + 304, + 155, + 547, + 348 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 355, + 398, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 355, + 398, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 355, + 398, + 369 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "content": "All the following ablation study of STViT and STViT-R are conducted on the DeiT-S and Swin-S, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 546, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 629 + ], + "type": "text", + "content": "Initialization analysis. The semantic tokens are the cluster centers recovered by attention layers. The initialization of cluster centers induces the representation of semantic tokens. Spatial and global initialization are adopted in Section 3.1 to guide the semantic tokens to integrate local and global semantic information separately. We compare different initialization components in Table 7. When performing single global initialization " + }, + { + "bbox": [ + 304, + 402, + 546, + 629 + ], + "type": "inline_equation", + "content": "(3^{rd}" + }, + { + "bbox": [ + 304, + 402, + 546, + 629 + ], + "type": "text", + "content": " row), we replace the spatial initial cluster centers with global initial cluster centers in the first transformer layer. The accuracy of using a single initialization method is far lower than using both, which shows the effectiveness of our initialization strategy. Our global initial cluster centers look similar to learned positional encoding since both use random initialization. To verify their distinction, we experiment with a real learned positional encoding by additionally adding the global initial cluster centers to keys, which causes " + }, + { + "bbox": [ + 304, + 402, + 546, + 629 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 402, + 546, + 629 + ], + "type": "text", + "content": " accuracy drop (the last row of Table 7). Therefore, our global initialization is different from learned positional encoding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "We visualize the attention maps of the first attention layer in the STGM in Figure 3. Because the queries of this layer are spatial initial cluster centers, these attention maps visualize the local semantic information integration by attention. The attention layer groups the semantic information according to the position of initial cluster centers, which ensures to extract fine-grained semantic information" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "6201" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 70, + 482, + 143 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 482, + 143 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 482, + 143 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 482, + 143 + ], + "type": "table", + "html": "
APbAPb50APb75APsbAPmAPm50APm75APsmFLOPs(G)
Swin-S51.870.456.335.244.767.948.528.8194
STViT-R-Swin-S51.870.656.136.744.767.848.629.0134(-31%)
Swin-B51.970.956.535.445.068.448.728.9343
STViT-R-Swin-B52.270.856.836.545.268.349.129.5233(-32%)
", + "image_path": "81d89a9ef91ac87562c9692191629aac42946d451a88bbf06115f96195b87213.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 177, + 278, + 243 + ], + "blocks": [ + { + "bbox": [ + 46, + 148, + 547, + 169 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 547, + 169 + ], + "type": "text", + "content": "Table 5. Results on COCO object detection and instance segmentation under Cascade Mask R-CNN with " + }, + { + "bbox": [ + 46, + 148, + 547, + 169 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 148, + 547, + 169 + ], + "type": "text", + "content": " schedule. The FLOPs are measured for backbones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 177, + 278, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 278, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 278, + 243 + ], + "type": "table", + "html": "
ModelTop-1 Acc(%)FLOPs(G)Speed
Swin-T78.888779
STViT-Swin-T78.5(-0.3)64.4(-27%)975(+25%)
Swin-S80.6166456
STViT-Swin-S80.3(-0.3)120.5(-27%)572(+25%)
", + "image_path": "75f0658c28bc6899d65a93b2c70b67abcb4cf6c23286c6cfb590552cb89684fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 86, + 293, + 247, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 246, + 287, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 291 + ], + "type": "text", + "content": "Table 6. Applying STViT to Video Swin (Swin-T and Swin-S) on Kinetics-400. All the models are pre-trained on ImageNet-1K. The views are " + }, + { + "bbox": [ + 46, + 246, + 287, + 291 + ], + "type": "inline_equation", + "content": "4\\times 3" + }, + { + "bbox": [ + 46, + 246, + 287, + 291 + ], + "type": "text", + "content": ". The top-1 accuracy and complexity in FLOPs are reported. Speed is evaluated by FPS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 293, + 247, + 365 + ], + "lines": [ + { + "bbox": [ + 86, + 293, + 247, + 365 + ], + "spans": [ + { + "bbox": [ + 86, + 293, + 247, + 365 + ], + "type": "table", + "html": "
SpatialGlobalLearnedTop-1 Acc(%)
✓79.4
✓78.7
✓✓79.8
✓✓79.7
", + "image_path": "cd6fd3f9889086f0a8a9178693dd28ffc3bff9527c4bf2d9a7b8765381c4056f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 82, + 409, + 252, + 455 + ], + "blocks": [ + { + "bbox": [ + 46, + 367, + 287, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 401 + ], + "type": "text", + "content": "Table 7. Accuracy with different initialization of STViT. Spatial, Global, and Learned indicate spatial initialization, global initialization, and learned positional encoding methods, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 409, + 252, + 455 + ], + "lines": [ + { + "bbox": [ + 82, + 409, + 252, + 455 + ], + "spans": [ + { + "bbox": [ + 82, + 409, + 252, + 455 + ], + "type": "table", + "html": "
No. of transformers234
Top-1 Acc(%)79.879.579.6
FLOPs(G)1.911.972.03
", + "image_path": "06f0ef9930a366040aef4c178ffecb4c13f40a310e11e5ba81d1bd711abf88e0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 80, + 499, + 253, + 546 + ], + "blocks": [ + { + "bbox": [ + 46, + 458, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 491 + ], + "type": "text", + "content": "Table 8. Performance evaluation on different numbers of transformer layers in STGM. Keeping the base module containing four transformer layers unchanged." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 499, + 253, + 546 + ], + "lines": [ + { + "bbox": [ + 80, + 499, + 253, + 546 + ], + "spans": [ + { + "bbox": [ + 80, + 499, + 253, + 546 + ], + "type": "table", + "html": "
STViT-Rw/o DUReusing ST
APb51.851.451.6
APm44.744.444.5
", + "image_path": "80a2b53a35d8ba89e2a5a9b59f908f7e79a2cd7d297760627e1d4eacaa5a7e7f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 552, + 287, + 574 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 574 + ], + "type": "text", + "content": "Table 9. Ablation study on STViT-R w/o dumbbell units (w/o " + }, + { + "bbox": [ + 46, + 552, + 287, + 574 + ], + "type": "inline_equation", + "content": "DU" + }, + { + "bbox": [ + 46, + 552, + 287, + 574 + ], + "type": "text", + "content": ") and reusing semantic tokens (Reusing ST)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "type": "text", + "content": "and keep the difference among semantic tokens. The attention map in the second attention layer is visualized in Supplementary A.5, which reveals that the network fixedly selects particle semantic tokens to represent global semantic information. We also show the semantic representation of image tokens in the same transformer layer on DeiT in Supplementary A.5. Compared to original image tokens, our semantic tokens in Figure 3 show more high-level semantic information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Number of transformer layers in STGM. Two transformer layers are adopted in the STGM by default. The ef" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 185, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 234 + ], + "type": "text", + "content": "fcts of employing different numbers of transformer layers are shown in Table 8. The extra layers are from the ones behind STGM to keep the total number of layers unchanged. More transformer layers do not bring improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 242, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 545, + 350 + ], + "type": "text", + "content": "The effectiveness of the dumbbell unit. To verify the effectiveness of our dumbbell unit, we experiment STViT-R without dumbbell units, i.e., STViT equipped with only the recovery module. We employ " + }, + { + "bbox": [ + 304, + 242, + 545, + 350 + ], + "type": "inline_equation", + "content": "6^{th}" + }, + { + "bbox": [ + 304, + 242, + 545, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 242, + 545, + 350 + ], + "type": "inline_equation", + "content": "7^{th}" + }, + { + "bbox": [ + 304, + 242, + 545, + 350 + ], + "type": "text", + "content": " transformer layers to construct STGM and the last transformer layers to construct the recovery module in Stage 3. The FLOPs is the same as the full-model STViT-R for fair comparison. The results on the COCO are reported in Table 9. Inferior results demonstrate the effectiveness of the dumbbell unit." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 358, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 419 + ], + "type": "text", + "content": "Reusing semantic tokens in dumbbell units. Semantic tokens are generated in each dumbbell unit. If they are produced only once in the first dumbbell unit and reused as initial cluster centers in the subsequent dumbbell units, the result is shown in Table 9 with a slight performance drop." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 444, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 624 + ], + "type": "text", + "content": "In this paper, we propose a simple and effective token sparsification method, semantic token vision transformer (STViT). Our method utilizes the clustering property of self-attention to generate a few semantic tokens with high-level information representation to replace the redundant image/video tokens, which can be applied in both global and local vision transformers. By simply configuring the recovery module, our method can be successfully applied to downstream tasks. Extensive experiments demonstrate that our method achieves better accuracy along with less inference time in most cases. The success in downstream tasks significantly boosts the development of token sparsification methods. We hope that this work can inspire more future research to pay much attention to high-level semantic representation in ViTs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 626, + 404, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 404, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 404, + 639 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "content": "This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, and Mike Zheng Shou's Start-Up Grant from NUS. Computation was partially performed on resources of the National Supercomputing Centre, Singapore. Shuning was supported by Alibaba Research Intern Program." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6202" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Luvcic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 178 + ], + "type": "text", + "content": "[2] Song Bai, Philip Torr, et al. Visual parser: Representing part-whole hierarchies with transformers. arXiv preprint arXiv:2107.05790, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 213 + ], + "type": "text", + "content": "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "text", + "content": "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 303 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 357 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "text", + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[8] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 480 + ], + "type": "text", + "content": "[9] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "type": "text", + "content": "[10] Yunpeng Chen, Marcus Rohrbach, Zhicheng Yan, Yan Shuicheng, Jiashi Feng, and Yannis Kalantidis. Graph-based global reasoning networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-442, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[11] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1601-1610, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[14] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Bain-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "ing Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. arXiv preprint arXiv:2107.00652, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 172, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 237 + ], + "type": "text", + "content": "[16] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 238, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 238, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 238, + 545, + 272 + ], + "type": "text", + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "text", + "content": "[18] Zilong Huang, Youcheng Ben, Guozhong Luo, Pei Cheng, Gang Yu, and Bin Fu. Shuffle transformer: Rethinking spatial shuffle for vision transformer. arXiv preprint arXiv:2106.03650, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "text", + "content": "[19] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 371, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 425 + ], + "type": "text", + "content": "[20] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[21] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Bin Ren, Minghai Qin, Hao Tang, and Yanzhi Wang. Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 481, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 545, + 525 + ], + "type": "text", + "content": "[22] Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient self-supervised vision transformers for representation learning. arXiv preprint arXiv:2106.09785, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "type": "text", + "content": "[23] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[24] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In International Conference on Learning Representations, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 691 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 715 + ], + "type": "text", + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6203" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[27] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[28] Xuezhe Ma, Xiang Kong, Sinong Wang, Chunting Zhou, Jonathan May, Hao Ma, and Luke Zettlemoyer. Luna: Linear unified nested attention. Advances in Neural Information Processing Systems, 34:2441-2453, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "text", + "content": "[29] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. arXiv preprint arXiv:2111.15668, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[30] Bowen Pan, Rameswar Panda, Yifan Jiang, Zhangyang Wang, Rogerio Feris, and Aude Oliva. Ia-red 2: Interpretability-aware redundancy reduction for vision transformers. Advances in Neural Information Processing Systems, 34, 2021. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 338 + ], + "type": "text", + "content": "[31] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34, 2021. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "type": "text", + "content": "[32] Michael S. Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "type": "text", + "content": "[33] Lin Song, Songyang Zhang, Songtao Liu, Zeming Li, Xuming He, Hongbin Sun, Jian Sun, and Nanning Zheng. Dynamic grained encoder for vision transformers. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 1, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[34] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. arXiv preprint arXiv:2106.02852, 2021. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "type": "text", + "content": "[35] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "text", + "content": "[36] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "text", + "content": "[37] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "text", + "content": "[38] Pichao Wang, Xue Wang, Hao Luo, Jingkai Zhou, Zhipeng Zhou, Fan Wang, Hao Li, and Rong Jin. Scaled relu matters for training vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "text", + "content": "Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[40] Wenxiao Wang, Lu Yao, Long Chen, Binbin Lin, Deng Cai, Xiaofei He, and Wei Liu. Crossformer: A versatile vision transformer hinging on cross-scale attention. arXiv preprint arXiv:2108.00154, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "text", + "content": "[41] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. Advances in Neural Information Processing Systems, 34, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8741-8750, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "text", + "content": "[43] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22-31, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "type": "text", + "content": "[44] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9981–9990, October 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[45] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 454 + ], + "type": "text", + "content": "[46] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "type": "text", + "content": "[47] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[48] Tan Yu, Gangming Zhao, Ping Li, and Yizhou Yu. Boat: Bilateral local attention vision transformer. arXiv preprint arXiv:2201.13027, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 545, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 611 + ], + "type": "text", + "content": "[49] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 558-567, October 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 690 + ], + "type": "text", + "content": "[50] David Junhao Zhang, Kunchang Li, Yali Wang, Yunpeng Chen, Shashwat Chandra, Yu Qiao, Luoqi Liu, and Mike Zheng Shou. Morph MLP: An efficient mlp-like backbone for spatial-temporal representation learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 230-248. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[51] Songyang Zhang, Xuming He, and Shipeng Yan. Latent-gnn: Learning efficient non-local relations for visual recog" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6204" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 352 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "nition. In International Conference on Machine Learning, pages 7374-7383. PMLR, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[52] Minghang Zheng, Peng Gao, Renrui Zhang, Kunchang Li, Xiaogang Wang, Hongsheng Li, and Hao Dong. End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 217 + ], + "type": "text", + "content": "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "text", + "content": "[54] Jingkai Zhou, Pichao Wang, Fan Wang, Qiong Liu, Hao Li, and Rong Jin. Elsa: Enhanced local self-attention for vision transformer. arXiv preprint arXiv:2112.12786, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "text", + "content": "[55] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "type": "text", + "content": "[56] Zhuofan Zong, Kunchang Li, Guanglu Song, Yali Wang, Yu Qiao, Biao Leng, and Yu Liu. Self-slimmed vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 432-448. Springer, 2022. 1, 2, 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6205" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_content_list.json b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2fb73edc51d7441bba93ed97d6608e6560bbfe9c --- /dev/null +++ b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_content_list.json @@ -0,0 +1,1406 @@ +[ + { + "type": "text", + "text": "MammalNet: A Large-scale Video Benchmark for Mammal Recognition and Behavior Understanding", + "text_level": 1, + "bbox": [ + 207, + 130, + 763, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun Chen $^{1*}$ Ming Hu $^{1*}$ Darren J. Coker $^{1}$ Michael L. Berumen $^{1}$ \nBlair Costello $^{2,3}$ Sara Beery $^{4}$ Anna Rohrbach $^{5}$ Mohamed Elhoseiny $^{1}$ $^{1}$ King Abdullah University of Science and Technology (KAUST) \n $^{2}$ Max Planck Institute of Animal Behavior, $^{3}$ University of Konstanz \n $^{4}$ Massachusetts Institute of Technology, $^{5}$ University of California, Berkeley", + "bbox": [ + 150, + 202, + 816, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6fdda67a752d39e5f70ee8b6fc5f275b719529b94fbef76aea0a398009e4d48.jpg", + "image_caption": [ + "Figure 1. We propose MammalNet, a large-scale video benchmark for recognizing mammals and their behavior. It is built around a biological mammal taxonomy spanning 17 orders, 69 families and 173 mammal categories, and includes 12 common high-level mammal behaviors (e.g., hunt, groom). MammalNet enables the study of animal and behavior recognition, both separately and jointly. It also facilitates investigating challenging compositional scenarios which test models' zero- and low-shot transfer abilities. Moreover, MammalNet includes behavior detection by localizing when a behavior occurs in an untrimmed video. Our dataset is the first to enable animal behavior analysis at scale in an ecologically-grounded manner, and exemplifies multiple challenges for the computer vision community, such as recognition of imbalanced, hierarchical distributions of fine-grained categories and generalization to unseen or seldom seen scenarios." + ], + "image_footnote": [], + "bbox": [ + 158, + 319, + 808, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 652, + 313, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monitoring animal behavior can facilitate conservation efforts by providing key insights into wildlife health, population status, and ecosystem function. Automatic recognition of animals and their behaviors is critical for capitalizing on the large unlabeled datasets generated by modern video devices and for accelerating monitoring efforts at scale. However, the development of automated recognition systems is currently hindered by a lack of appropriately labeled datasets. Existing video datasets 1) do not classify animals according to established biological taxonomies; 2) are too small to facilitate large-scale behavioral studies and are often limited to a single species; and 3) do not feature temporally local-", + "bbox": [ + 75, + 686, + 470, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ized annotations and therefore do not facilitate localization of targeted behaviors within longer video sequences. Thus, we propose MammalNet, a new large-scale animal behavior dataset with taxonomy-guided annotations of mammals and their common behaviors. MammalNet contains over 18K videos totaling 539 hours, which is $\\sim 10$ times larger than the largest existing animal behavior dataset [36]. It covers 17 orders, 69 families, and 173 mammal categories for animal categorization and captures 12 high-level animal behaviors that received focus in previous animal behavior studies. We establish three benchmarks on MammalNet: standard animal and behavior recognition, compositional low-shot animal and behavior recognition, and behavior detection. Our dataset and code have been made available at: https://mammal-net.github.io.", + "bbox": [ + 496, + 652, + 893, + 880 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 101, + 887, + 212, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13052", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Animal species are a core component of the world's ecosystems. Through their behavior, animals drive diverse ecological processes, including seed dispersal, nutrient cycling, population dynamics, speciation, and extinction. Thus, understanding and monitoring the behaviors of animals and their interactions with their physical and social environments is key to understanding the complexities of the world's ecosystems, an objective that is especially critical now given the ongoing biodiversity crisis [12].", + "bbox": [ + 80, + 114, + 468, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modern sensors, including camera traps, drones, and smartphones, allow wildlife researchers, managers, and citizen scientists to collect video data of animal behavior on an unprecedented scale [43]. However, processing this data to generate actionable, timely insights remains a major challenge. Manual human review and annotation of footage to identify and locate species and behavioral sequences of interest is time-intensive and does not scale to large datasets. Thus, methods for automated animal and behavioral recognition could open the door to large-scale behavioral monitoring and speed up the time to produce usable data, thereby reducing the time to implement management directives.", + "bbox": [ + 80, + 252, + 468, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first essential step to creating such an AI system for animal and behavior recognition is curating a diverse, representative dataset that allows us to formalize these challenges as computer vision tasks and benchmark potential solutions. Most previous datasets either only cover a limited number of animal and behavior types [4, 38], or do not implement animal labeling [36], or include a small number of videos with insufficient environmental diversity [4, 38, 48]. Recently, a dataset named \"Animal Kingdom\" [36] was proposed to study animal actions and is currently the largest existing behavioral dataset, to the best of our knowledge. However, it only contains 4,310 videos totaling 50 hours, which might be insufficient for large-scale animal behavior studies considering its diversity. Furthermore, the authors only focus on the recognition of atomic actions such as yawning, swimming, and flying. These basic actions cannot be easily matched to the higher-order behavioral states that are of primary interest to end users in animal management and conservation [6]. For example, a cheetah that is running may either be hunting, escaping, or playing. Finally, and most importantly, they do not support some important tasks such as animal recognition and behavior detection which are essential for animal behavior understanding.", + "bbox": [ + 80, + 433, + 468, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the limitations of previous datasets, we propose a new dataset called MammalNet. We specifically focus on mammals since they, unlike other animal classes such as birds or insects, usually have more diverse and distinguishable behavior statuses. MammalNet is comprised of 539 hours of annotated videos, which is $\\sim 10$ times longer than that of the largest available animal behavior dataset. It contains 18,346 videos depicting 12 fundamental high-level", + "bbox": [ + 80, + 780, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "behaviors from hundreds of mammal species. Importantly, it focuses on 12 higher-order animal behaviors that are the focus of previous animal behavior literature [3,8,17,33], rather than atomic actions. MammalNet also categorizes animals according to the scientific taxonomy available in Wikipedia, as we show in Fig. 1; hence the dataset can be flexibly expanded in the future by following the same protocols. It includes videos of approximately 800 mammal species in 173 mammal categories. We establish three benchmarks inspired by ecological research needs - standard animal & behavior classification, compositional low-shot animal & behavior recognition, and behavior detection - to promote future study in animal behavior understanding.", + "bbox": [ + 501, + 90, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through our experiments, we find that: (1) Correctly recognizing the animals and behaviors is a challenging task even for the state-of-the-art models, especially for less-frequent animals. The top-1 per-class accuracy is 32.5 for animal recognition, 37.8 for behavior recognition, and 17.8 for their joint recognition in our best-performing model. (2) Behavior recognition for unseen animals can be transferred from observations of other seen animals due to their similar features such as appearance and movement style, which can help in studies of animals with less available data. However, to achieve more accurate behavior recognition, having access to videos of the target animals and behaviors is still crucial.", + "bbox": [ + 503, + 287, + 890, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 482, + 638, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Automatic animal recognition and behavior detection can help humans monitor and efficiently process animal behavior data [7, 22, 32, 34, 40, 47, 48]. It can massively reduce the labour cost from manually collecting and analyzing animal activities. During the past few years, many datasets [18, 20, 29, 38] have been introduced to develop foundations for animal behavior research. We systematically analyze the previous datasets and summarize several important limitations that prevent them from being used for large-scale animal recognition and behavior understanding:", + "bbox": [ + 503, + 507, + 890, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lack of behavior understanding. Many previous datasets only focus on animal recognition or pose estimation from images, but lack behavior learning. For example, iNaturalist [45] collects 859,000 images covering more than 5,000 different types of plants and animals. NABird [44] collects 48,562 images of 555 different North American bird species. Also, some works focus on narrow animal recognition such as dogs [27], birds [46, 50] or cats [13]. On the other hand, many works also focus on pose estimation [10, 18, 21, 31, 41] and animal face detection [26, 39, 51]. They are generally not applicable to learning animal behavior.", + "bbox": [ + 503, + 659, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lack of taxonomic diversity for different behaviors. Previous animal behavior datasets have minimal taxonomic coverage - often containing just a single animal species. For example, there are existing behavior recognition datasets for elephants [28], sheep [37], monkeys [5], tigers [16],", + "bbox": [ + 503, + 825, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13053", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/462dac20fe5ac399fc34a75cf67f1d22b80b3d763f6d01e52321e697f7e41a8b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsDataset PropertiesTasks
Publicly Available?Taxonomy-guided Animal Annotation?No. of VideosNo. of ActionsNo. of BehaviorsNo. of Animal CategoriesNo. of Mammal CategoriesTotal DurationAnimal ClassificationAction/Behavior RecognitionAction/Behavior Detection
Wild Felines [20]××2,7003-33-✓✓×
Wildlife Actions [29]××10,6007-3211-✓✓×
Animal Kingdom [36]✓×4,301140-850-50 (h)×✓×
MammalNet (ours)✓✓18,346-12173173539 (h)✓✓✓
", + "bbox": [ + 122, + 88, + 854, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. The comparison among existing animal behavior understanding video datasets. Compared to other datasets, MammalNet annotates the animals by following the scientific mammal taxonomy, focuses on more high-level behavior recognition, has the largest number of mammal categories, collects the largest number of animal behavior videos, totalling 539 hours, and also enables behavior detection tasks.", + "bbox": [ + 75, + 263, + 893, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "etc. While useful for the studied species, these datasets do not enable the exploration of behavior across species which is necessary to scale up behavior identification without requiring training examples of every possible combination of species and behavior.", + "bbox": [ + 75, + 315, + 470, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lack of taxonomy-guided animal annotation. Previous animal behavior datasets either do not include animal recognition as a task [36] or group species according to subjective as opposed to scientific criteria [29, 36]. In contrast, our dataset collects and annotates the videos according to the scientific mammal taxonomy. By following this taxonomy we allow exploration of behavior from an evolutionary perspective, as well as enable standardized and consistent dataset expansion under the same protocol.", + "bbox": [ + 75, + 398, + 470, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Actions vs. Behaviors. Previous works mostly focus on atomic action recognition [20, 29, 36]. For example, some of their action classes are walk, stand still, fly, etc. In contrast, MammalNet focuses on higher-level animal behaviors, such as hunt, feed baby, etc. Behavior here denotes the main activity instance during a period, and it is usually composed of multiple atomic actions. These complex behaviors are needed to describe and summarize video activity bouts in a manner which is valuable for ecological research.", + "bbox": [ + 75, + 540, + 468, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We further compare our MammalNet dataset with the other existing animal behavior understanding video datasets, and summarize the key difference in Table 1.", + "bbox": [ + 76, + 683, + 470, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Constructing MammalNet", + "text_level": 1, + "bbox": [ + 76, + 762, + 321, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of MammalNet is to provide a large-scale mammal video dataset that benchmarks both animal and behavior recognition. In this section, we discuss our dataset construction protocol, including the choice of scientific animal taxonomy, crowdsourced annotations, and performing manual quality control during video collection and annotation. Finally, we describe the statistical profile of MammalNet.", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/66fa26fa3a4efd7085a90d0782e5c8c1488de24cb201e6fd226870ca3aa9017f.jpg", + "image_caption": [ + "Figure 2. A subset of the mammal taxonomy of MammalNet. It includes 3 orders, 11 families, and 45 genera." + ], + "image_footnote": [], + "bbox": [ + 521, + 313, + 870, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Animal Taxonomy and Behavior Collection", + "text_level": 1, + "bbox": [ + 500, + 628, + 867, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Animal taxonomy construction. Accurate animal recognition is one of the main goals defined by MammalNet. To ensure this task is scientifically relevant, we aimed to collect and structure our list of target animals based on the mammal taxonomy. First, we collected a diverse list of mammals, approximately 800 mammal species and sub-species, from the National Geographic [2] and Animal A-Z [1] websites. Next, we manually mapped each mammal onto the taxonomic structure (including class, order, family, genus, tribe, sub-family and species) in mammal taxonomy from Wikipedia. In total, the classes included in MammalNet cover 17 orders, 69 families, and 173 mammal categories. A taxonomic subset is visualized in Fig. 2.", + "bbox": [ + 496, + 655, + 893, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Behavior collection. We aim to enable the study of complex, high-level animal behavior as opposed to the simpler atomic actions emphasized in previous work [29, 36]. Be", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13054", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5aef39351aea3ce5681ae57da55728ed0a2162f32d14801b5c5ec3bddec217ad.jpg", + "image_caption": [ + "Figure 3. The examples for the annotated target behavior boundaries. The frames marked in red boxes denote the annotated temporal boundaries for the target behavior." + ], + "image_footnote": [], + "bbox": [ + 143, + 90, + 830, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "havior here represents the major activity being displayed during a period of a video, and can be viewed as a series of atomic actions collectively serving a higher-level purpose. For example, hunting behavior, as defined in MammalNet, can often be decomposed into running, chasing and killing actions, etc. Identification of behaviors at this level of definition is more useful for ecologists and zoologists [3,8,17,33] compared to atomic actions.", + "bbox": [ + 75, + 294, + 472, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by previous biological and ecological studies [3, 8,17], we consider 12 fundamental mammal behaviors under 5 different groups in our study. They are respectively:", + "bbox": [ + 76, + 416, + 470, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Foraging behaviors: eat food, drink water, hunt.", + "bbox": [ + 76, + 463, + 403, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reproductive behaviors: mate, feed baby, give birth.", + "bbox": [ + 76, + 479, + 433, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hygiene behaviors: groom.", + "bbox": [ + 76, + 496, + 266, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Agnostic behaviours: fight.", + "bbox": [ + 76, + 512, + 266, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Maintenance behaviors: urinate, defecate, sleep, vomit.", + "bbox": [ + 76, + 529, + 452, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Video Collection and Quality Assurance", + "text_level": 1, + "bbox": [ + 76, + 556, + 419, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset curation is usually a costly process requiring a lot of manual annotation by humans. Some datasets are annotated by domain experts to encourage more reliable labeling, particularly for challenging tasks, but this expertise comes at a much higher cost and is thus hard to scale. We adapted a semi-automatic crowdsourcing approach to collect and annotate our datasets, inspired by many previous works [9, 15, 23, 49].", + "bbox": [ + 75, + 582, + 470, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Online video retrieval. Our goal was to collect videos depicting each animal in our database performing each of the 12 focal behaviors. To achieve this, we queried YouTube with the text combinations of each animal common name and behavior, e.g., tiger hunting, and downloaded videos where the title included our queried animal name and behavior. However, some animal names are too rare to generate enough relevant videos, e.g., Brown hyena or Spotted hyena. In such cases, we used a more common name, hyena, to represent the union of those animals. In order to retrieve more relevant videos from the search engine, we also expanded each animal with their synonyms as given in Wikipedia. For example, we expanded the original artic fox with other equivalent names", + "bbox": [ + 75, + 704, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "such as white fox, polar fox and snow fox. Each video was downloaded at its highest available resolution.", + "bbox": [ + 498, + 294, + 890, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data filtering. During video retrieval, we downloaded the videos that are accessible for people $< 16$ years old to avoid the videos with violent content. We also prioritized videos that are shorter than 10 minutes in duration to limit the total storage.", + "bbox": [ + 496, + 324, + 893, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, some videos might have irrelevant content due to the inaccuracy of text-based retrieval. For example, the downloaded videos might 1) display cartoon or toy animals or unrealistic environments such as games or movies, 2) display a static image instead of a continuous video, 3) involve a lot of human-animal interaction, e.g., human feeding an animal, as opposed to focusing on animal behavior, 4) not contain the specified animal and/or behavior.", + "bbox": [ + 496, + 400, + 893, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate these issues, we employed Amazon Mechanical Turk workers to verify the presence of the animal and behavior, and identify other quality issues. We assigned each video to three different workers and provided them with the pictures of an animal with its common name, and an expected behavior. We asked the workers to verify if this animal and behavior indeed appear in the video. Only the videos that \"pass\" by all the three workers were kept for the following behavior localization annotation. Before the workers started the verification, we first provided them the verification instructions and asked them to complete a corresponding qualification test with 20 multiple choice questions to ensure only qualified workers could participate in our task.", + "bbox": [ + 496, + 521, + 893, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Animal Behavior Localization Annotation", + "text_level": 1, + "bbox": [ + 500, + 726, + 857, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The target behavior typically does not span the whole video. To localize the part where the target behavior is actually occurring in the video, we asked the AMT workers to manually annotate the respective temporal boundaries. Namely, we asked five different qualified workers to annotate the start and end frames for the target behavior in each video. In the end, each video received at least 5 annotated behavior boundaries. To achieve robust annotation agreement, we used the complete linkage algorithm [14] to cluster different temporal boundaries and merge them into one or", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13055", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a4bb2c4941349bf4f0f334fbb6c549cf4212dad42b795822388b36c9662a4fe8.jpg", + "image_caption": [ + "Figure 4. Number of videos per each mammal category. We rank the categories according to their trimmed videos frequency." + ], + "image_footnote": [], + "bbox": [ + 138, + 90, + 830, + 345 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/15a0bc1da8703c7aefab36a54a8d9d1ad046bac34ae89d84951b0d05775123ef.jpg", + "image_caption": [ + "Figure 5. Number of videos per each behavior. We rank the behavior according to their trimmed videos frequency." + ], + "image_footnote": [], + "bbox": [ + 135, + 381, + 411, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "several more stable ones that received multiple agreements. Note, that a single video might have multiple discrete occurrences of a given behavior, and thus have multiple boundary definitions. We show several examples for annotated target behavior boundaries in Fig. 3.", + "bbox": [ + 75, + 564, + 470, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Recognition at lowest feasible taxonomic level", + "text_level": 1, + "bbox": [ + 76, + 648, + 464, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We categorize animals according to the lowest and feasible taxonomic classification, rather than the species level, for the following reasons: 1) YouTube videos contain the ambiguous and inexpert species labels. 2) We are not using expert annotators, and even experts often cannot reliably identify animals at the species or even genus level based on crowdsourced or field images [25]. Thus, it is more practical to classify animals in our dataset to the lowest feasible taxonomic level rather than species. As a result, our taxonomy contains 173 distinguishable taxonomic classification levels including: sub-family, tribe and genus.", + "bbox": [ + 75, + 672, + 470, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. MammalNet Statistics", + "text_level": 1, + "bbox": [ + 76, + 847, + 284, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final MammalNet dataset contains 18,346 videos (539 hours); after we trim the videos according to the anno", + "bbox": [ + 75, + 869, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tated behavior boundaries, it increases to 20,033 trimmed video instances (394 hours). In total, it covers 173 mammal categories, 69 families and 17 orders in our dataset, with the total of 173 mammal categories defined for our recognition tasks. MammalNet also contains 12 different common behaviors. The average duration for the untrimmed and trimmed videos is 106 and 77 seconds, respectively. Over $54\\%$ of videos reach HD resolutions ( $1280 \\times 720$ ). We demonstrate the data distribution in terms of each animal category and behavior in Fig. 4 and 5. We observe that the number of collected videos per type follows a long-tail distribution.", + "bbox": [ + 496, + 383, + 893, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 566, + 705, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We construct three main tasks on MammalNet: 1) Standard animal and behavior classification on trimmed videos, 2) Compositional low-shot animal and behavior recognition on trimmed videos, and 3) Behavior detection on untrimmed videos. In both the classification and detection tasks, we baseline several state-of-the-art models [19, 30, 53] that have been successfully applied to human action recognition and detection. In the following, we describe the formulation of each challenge and provide baselines and analysis.", + "bbox": [ + 496, + 592, + 893, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Standard Animal and Behavior Classification on Trimmed Videos", + "text_level": 1, + "bbox": [ + 498, + 739, + 890, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This task explores classification of both the primary behavior that occurs in a trimmed video and the animal that performs the behavior. We report the top-1 per-example and per-class accuracy for all the baseline models.", + "bbox": [ + 496, + 777, + 893, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Many, medium, few splits. To capture the effect of the long-tailed nature of the MammalNet dataset, we group the animal, behavior, and their composition classes into many, medium, and few based on their frequency, and report the average", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13056", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b39635d7a261461e1eb78284e2e324aa6badb2fed8079f91f21b6a283861c0f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselinesAnimal ClassificationBehavior ClassificationJoint Classification
Many 12Medium 28Few 133All 173Many 4Medium 4Few 4All 12Many 33Medium 180Few 823All 1036
SlowFast [19]49.635.69.517.239.027.614.927.227.219.44.48.6
C3D [42]48.635.310.017.538.227.811.725.928.118.84.58.6
I3D [11]48.834.910.517.839.527.214.827.229.620.54.48.9
MViT V2 [30]48.535.512.919.742.329.211.627.729.519.64.89.0
SlowFast*58.343.116.624.545.132.714.830.938.623.57.312.1
C3D*58.345.419.126.844.636.015.932.238.026.48.413.5
I3D*58.642.916.924.846.335.014.832.138.324.58.613.3
MViT V2*66.756.023.432.550.942.420.037.846.233.011.817.8
", + "bbox": [ + 107, + 88, + 864, + 265 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ecbd9e74bf9c4fd21f3adeb15d477e7fead5acc7773ae6749e64e189a1acfcd2.jpg", + "table_caption": [ + "Table 2. Per-class Top-1 accuracy for animal, behavior and their joint prediction.* denotes the initialization from the model pretrained on Kinetics 400 [24]. Transfer learning from the human action to the animal behavior recognition receives considerable performance gain. Best performance for each split has been highlighted in bold." + ], + "table_footnote": [], + "table_body": "
BaselinesAnimalBehaviorJoint
SlowFast [19]35.434.217.4
C3D [42]35.033.517.1
I3D [11]35.234.317.9
MViT V2 [30]35.636.818.0
SlowFast*43.039.422.8
C3D*44.440.324.6
I3D*43.441.224.0
MViT V2*52.646.630.6
", + "bbox": [ + 125, + 338, + 421, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Per-example accuracy for animal, behavior and their joint prediction. * denotes the initialization from the pretrained model.", + "bbox": [ + 75, + 508, + 470, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "per-class accuracy bands chosen based on the frequency percentiles. For animal categories this is broken down as many: top $7\\%$ frequent classes, medium: middle $16\\%$ classes, and few: the remaining $77\\%$ classes. For behavior, many: top $33\\%$ frequent classes, medium: middle $33\\%$ , and few: the remaining $33\\%$ classes. For joint classification, many: top $3\\%$ frequent classes, medium: middle $17\\%$ classes, few: the remaining $80\\%$ classes. We show the number of classes per each split in Table 2.", + "bbox": [ + 75, + 547, + 470, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset setup. We randomly split the examples from each animal-behavior category into $70\\%$ for training, $10\\%$ for validation, and $20\\%$ for testing, and it results in 14,554 training, 1,638 validation, and 3,841 testing videos, respectively.", + "bbox": [ + 75, + 686, + 470, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare SlowFast [19], I3D [11], C3D [42] and MViT V2 [30] models on our tasks. These models are evaluated in two versions: 1) Training with random initialization 2) Initializing with weights from a model pretrained on Kinetics 400 [24]. These methods were originally designed for human action recognition and hence do not have an ability to predict both an action and a subject by default. To accommodate them into our joint prediction setting, we have two task heads, one for animal category recognition and one for behavior recognition. We compute the joint", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "loss, $\\mathcal{L}_{joint}$ , for both animal and behavior classification as shown in Fig. 1. We tune all the hyper-parameters on the validation data. The final hyper-parameters for each model are provided in the supplement. The loss is defined as:", + "bbox": [ + 496, + 340, + 893, + 402 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = - \\frac {1}{M} \\sum_ {i} ^ {M} y _ {i} ^ {a} \\log \\left(p _ {i} ^ {a}\\right) - \\frac {1}{N} \\sum_ {j} ^ {N} y _ {j} ^ {b} \\log \\left(p _ {j} ^ {b}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 414, + 892, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the number of animal classes, and $N$ is the number of behavior classes, $p_i^a$ and $y_{i}^{a}$ denote the animal prediction probability and ground truth label for the category $i$ , $p_y^b$ and $y_j^b$ denote the behavior prediction probability and ground truth label for the category $j$ .", + "bbox": [ + 496, + 469, + 893, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental results. The results for per-class and per-example classification are summarized in Tables 2 and 3, respectively. We find that MViT v2 is able to achieve competitive results for all the splits. The best top-1 joint per-class accuracy is 17.8 and per-example accuracy is 30.6, which points to significant room for improvement on these challenging tasks. We also observe that transfer learning from the model that is pretrained on Kinetics 400, a human action recognition dataset, improves both the animal and behavior classification accuracy (with MViT v2, this corresponds to a per-class accuracy gain from 19.7 to 32.5 for animal classification, and from 27.7 to 37.8 for behavior classification). Additionally, we find that performance gain from pretraining is higher for frequently occurring animals and behaviors, indicated by the results shown in many, medium, and few splits. Accurately predicting low-frequency animal and behavior categories remains a significant challenge.", + "bbox": [ + 496, + 547, + 895, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Compositional Low-shot Animal and Behavior Classification on Trimmed Videos", + "text_level": 1, + "bbox": [ + 498, + 815, + 893, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is hard to find a sufficient number of labeled behavior samples for all the animals in our collected taxonomy. For example, numbat and florida panther have very few behavior", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13057", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/165bb732b04e4567c36460796b2b9f5fbd72c4cfd34480109c2ce9aa853c98ea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselinesCompositional Low-Shot Behavior Classification
0-shot1-shot5-shot
Per-example A / BPer-class A / BPer-example A / BPer-class A / BPer-example A / BPer-class A / B
C3D*27.4 / 23.718.3 / 16.129.2 / 25.521.1 / 19.933.3 / 29.325.2 / 23.0
I3D*25.3 / 23.916.7 / 15.226.8 / 25.716.9 / 18.330.5 / 28.321.7 / 21.9
SlowFast*26.2 / 24.817.9 / 16.326.5 / 26.316.7 / 19.029.6 / 29.222.5 / 22.4
MViT V2*32.2 / 26.220.7 / 18.133.7 / 28.922.9 / 22.539.3 / 31.731.0 / 26.0
", + "bbox": [ + 151, + 88, + 821, + 226 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d74311b1b5ee165c8905fd3e08c65161e598cb96810e2e26841e6d2ac6000e57.jpg", + "table_caption": [ + "Table 4. Compositional low-shot animal and behavior recognition. * denotes the initialization from the model pretrained on Kinetics 400 [24]. \"A\" denotes the animal category and \"B\" denotes the behavior category. The best performance per each column has been highlighted in bold." + ], + "table_footnote": [], + "table_body": "
BaselinesmAP
0.500.600.700.800.90Avg.
CoLA [52]26.0222.7018.9813.463.0515.81
TAGS [35]23.0920.9719.0916.9812.5617.63
ActionFormer [53]28.4826.1423.1718.6910.4820.07
", + "bbox": [ + 81, + 289, + 465, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. The results for behavior detection. We report mAP at the IoU thresholds of [0.5:0.1:0.9]. Average mAP is computed by averaging different tIoU thresholds.", + "bbox": [ + 75, + 372, + 468, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "annotations. However, it is plausible to imagine that behaviors can be transferred among different animals that have similar appearances and movement styles. Also the animal recognition under different behaviors (hunting, fighting) can be mutually transferred. To investigate these phenomena, we design the compositional low-shot animal and behavior classification task.", + "bbox": [ + 75, + 430, + 468, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset setup. We first select the animal-behavior compositional classes that contain more than 5 examples and allocate $25\\%$ of classes to the test set (4,088 videos). For the remaining classes (the other $75\\%$ of classes and those classes with $\\leq 5$ examples), we randomly designate $90\\%$ of classes as the training set and $10\\%$ as the validation set (898 videos). Under the low-shot scheme, for each compositional class, we randomly sample 5 examples from the test set and move 0, 1, or 5 of them into the training set for the zero-shot, 1-shot, and 5-shot setup, respectively (14,377, 14,511 and 15,047 training videos). The train, val and test sets consist of 983, 53, and 134 compositional classes, respectively.", + "bbox": [ + 75, + 536, + 470, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We evaluate the compositional low-shot classification with the SlowFast [19], I3D [11], C3D [42] and MViT V2 [30] models under the joint loss. We initialize these models with the weights pretrained on Kinetics 400 [24].", + "bbox": [ + 75, + 718, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental results. We summarize the results in Table 4. It shows that MViT v2 consistently achieves the best performance under our low-shot setup. The behavior classification can still achieve 26.2 per-example and 18.1 top-1 accuracy under the zero-shot setting for MViT v2 model. This indicates that behavior classification is transferable from other animals in some cases. Additionally, we observe that its performance is improved to 31.7 per-example and 26.0 per-class", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "top-1 accuracy under 5-shot setup, indicating that training on more videos with the behaviors from the same animal is still necessary. A similar phenomenon is observed to the few-shot animal recognition.", + "bbox": [ + 496, + 291, + 890, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Behavior Detection on Untrimmed Videos", + "text_level": 1, + "bbox": [ + 498, + 363, + 856, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This task is to detect the behavior in untrimmed videos. The behavior detection algorithm should correctly detect the temporal range for the primary activity presented in the video. We follow previous temporal action localization works [9, 53] to benchmark this task. We report the mean Average Precision (mAP) with different temporal intersections over the union (tIoU) thresholds [0.5:0.1:0.9] as our evaluation metric. We also report the average mAP averaging across different tIoUs.", + "bbox": [ + 496, + 387, + 893, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset setup. Similar to the previous standard animal and behavior classification, we follow the [train:0.7, val:0.1, test:0.2] ratios to split the untrimmed videos at the animal-behavior composition level. This results in 13,318 videos for training, 1,486 for validation and 3,542 for testing. We tune all the hyper-parameters based on the validation set.", + "bbox": [ + 496, + 523, + 893, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We evaluate our datasets with the baseline models such as ActionFormer [53], TAGS [35], and CoLA [52]. To produce the features for our MammalNet videos, we first finetune a two-stream I3D [11] model, that is originally pretrained on ImageNet [15] and Kinetics 400 [24], on our dataset, and then extract the RGB and optical flow features for each video. We concatenate these two features together as the model input.", + "bbox": [ + 496, + 616, + 890, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental results. We show the behavior detection results in Table 5. Among all the baselines, ActionFormer demonstrates the most competitive performance with an average mAP of 20.07, and also achieves $28.48\\mathrm{mAP}$ for the threshold of 0.5. Overall, it is clear to see that the behavior detection task is still very challenging for current methods.", + "bbox": [ + 496, + 737, + 890, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 500, + 844, + 596, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Demonstration of animal and behavior classification. We sample some prediction examples from MViT v2 [30] for", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13058", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59169cb89f12e11b23d0d87d7c597a12fa12cf2c0e10398c95f4d355b758d3d9.jpg", + "image_caption": [ + "Figure 6. The visualization presents various instances of joint animal and behavior classification. In the third column, accurate predictions are displayed, while the fourth, fifth, and sixth columns showcase mispredicted examples where either the animal or the behavior does not correspond to the correct prediction." + ], + "image_footnote": [], + "bbox": [ + 120, + 90, + 848, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf705a7d59d382b00b659bb076011a3b7401918cca6c1d06027fd70afbd7c741.jpg", + "image_caption": [ + "Figure 7. The visualization for behavior detection examples." + ], + "image_footnote": [], + "bbox": [ + 78, + 306, + 467, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the joint categories Acinonyx hunt, Canis eat and Antilopini fight. We demonstrate one correct prediction and three mispredicted examples where the model mis-recognizes the animal or behavior or both of them in Fig. 6.", + "bbox": [ + 76, + 489, + 470, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Demonstration of behavior detection: We visualize behavior detection results in Fig. 7 from ActionFormer [53]. The top example shows correctly predicted behavior with a proposal closely aligned with the ground-truth. The bottom example shows a misclassified behavior, and it mistakenly predicts the hunt behavior as fight.", + "bbox": [ + 76, + 550, + 470, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Joint animal and behavior recognition vs. separate recognition. We also train the model for recognizing the animal and behavior separately and provide the results in the Table 1 of the supplementary. Comparing with the joint recognition, we find that training a system to recognize the animal and behavior together can improve the behavior recognition under separate training by $\\sim 2.2$ per-class accuracy. This indicates that being capable of understanding the animal types can benefit behavior prediction. However, the results also indicate that predicting animal only can be more useful in recognizing animal types.", + "bbox": [ + 75, + 642, + 470, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset bias. Our dataset is downloaded from YouTube channels with diverse backgrounds and video quality, and it might differ from video datasets collected for behavioral or ecological studies. The videos on YouTube might also exhibit potential biases:", + "bbox": [ + 75, + 809, + 470, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1) Over-representation of some behaviors and under", + "bbox": [ + 76, + 885, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "representation of others. For example, people might prefer watching videos with fight or eat food activities, and these two behaviors are more likely to be over-represented, while urinate and defecate behaviors are less interesting to humans and hence are underrepresented on YouTube. However, they are still important in informing conservation-related actions to protect the environment.", + "bbox": [ + 496, + 308, + 893, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2) Bias towards captive animals or wild animals that are habituated to humans. We find that many videos are shot at zoos, farms, and homes, etc. These animals may display different behaviors than wild or non-habituated animals.", + "bbox": [ + 496, + 415, + 893, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 494, + 619, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced MammalNet, a large-scale video dataset for mammal recognition and behavior understanding. We have collected videos for hundreds of different mammals and structured them by following the scientific mammal taxonomy. MammalNet consists of 18,346 untrimmed videos covering 173 mammal categories and 12 common behaviors. We established three challenges: standard animal and behavior classification, compositional low-shot animal and behavior classification, and behavior detection. Through our experiments, we found that the accurate recognition of animals at scale and their common behaviors is very challenging even with current state-of-the-art models, especially when the dataset has a long-tail distribution. We also found that learning to recognize the behavior of unseen animals is possible via transfer from the other seen animals. To promote further research and development in the field of animal behavior study, we have open-sourced all of our data and code to the research community.", + "bbox": [ + 496, + 521, + 893, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This work is supported by KAUST BAS/1/1685-01-01 and KAUST FCC/1/1973-58-01 (Red Sea Research Center), DARPA's SemaFor and PTG programs, the Caltech Resnick Sustainability Institute, and Germany's Excellence Strategy–‘Centre for the Advanced Study of Collective Behaviour’ EXC 2117-422037984.", + "bbox": [ + 496, + 809, + 895, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13059", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] A-z animals. https://a-z-animals.com/.3", + "[2] National geographics. https://www.nationalgeographic.com/.3", + "[3] John Alcock. Animal behavior: An evolutionary approach. Sinauer Associates, 2009. 2, 4", + "[4] David J Anderson and Pietro Perona. Toward a science of computational ethology. Neuron, 84(1):18-31, 2014. 2", + "[5] Praneet C Bala, Benjamin R Eisenreich, Seng Bum Michael Yoo, Benjamin Y Hayden, Hyun Soo Park, and Jan Zimmermann. Automated markerless pose estimation in freely moving macaques with openmonkeystudio. Nature communications, 11(1):1-12, 2020. 2", + "[6] Oded Berger-Tal, Daniel T Blumstein, Scott Carroll, Robert N Fisher, Sarah L Mesnick, Megan A Owen, David Saltz, Colleen Cassady St Claire, and Ronald R Swaisgood. A systematic survey of the integration of animal behavior into conservation. Conservation Biology, 30(4):744-753, 2016. 2", + "[7] Cigdem Bayan and Robert B Fisher. Detection of abnormal fish trajectories using a clustering based hierarchical classifier. In BMVC, 2013. 2", + "[8] Michael D Breed and Janice Moore. Animal behavior. Academic Press, 2021. 2, 4", + "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 4, 7", + "[10] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9498-9507, 2019. 2", + "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6, 7", + "[12] Gerardo Ceballos, Paul R Ehrlich, and Peter H Raven. Vertebrates on the brink as indicators of biological annihilation and the sixth mass extinction. Proceedings of the National Academy of Sciences, 117(24):13596-13602, 2020. 2", + "[13] Yu-Chen Chen, Shintami C Hidayati, Wen-Huang Cheng, Min-Chun Hu, and Kai-Lung Hua. Locality constrained sparse representation for cat recognition. In International Conference on Multimedia Modeling, pages 140-151. Springer, 2016. 2", + "[14] Daniel Defays. An efficient algorithm for a complete link method. The Computer Journal, 20(4):364-366, 1977. 4", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 4, 7", + "[16] Rod K Dishman, Andrew S Jackson, and Molly S Bray. Self-regulation of exercise behavior in the tiger study. Annals of Behavioral Medicine, 48(1):80-91, 2014. 2", + "[17] Lee Alan Dugatkin. Principles of animal behavior. University of Chicago Press, 2020. 2, 4" + ], + "bbox": [ + 78, + 114, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Cheng Fang, Tiemin Zhang, Haikun Zheng, Junduan Huang, and Kaixuan Cuan. Pose estimation and behavior classification of broiler chickens based on deep neural networks. Computers and Electronics in Agriculture, 180:105863, 2021. 2", + "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5, 6, 7", + "[20] Liqi Feng, Yaqin Zhao, Yichao Sun, Wenxuan Zhao, and Jiaxi Tang. Action recognition using a spatial-temporal network for wild felines. Animals, 11(2):485, 2021. 2, 3", + "[21] Jacob M Graving, Daniel Chae, Hemal Naik, Liang Li, Benjamin Koger, Blair R Costelloe, and Iain D Couzin. Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. *Elife*, 8:e47994, 2019. 2", + "[22] Shilpi Gupta, Prerana Mukherjee, Santanu Chaudhury, Brejesh Lall, and Hemanth Sanisetty. Dftnet: Deep fish tracker with attention mechanism in unconstrained marine environments. IEEE Transactions on Instrumentation and Measurement, 70:1-13, 2021. 2", + "[23] Fabian Caba Heilbron and Juan Carlos Niebles. Collecting and annotating human activities in web videos. In Proceedings of International Conference on Multimedia Retrieval, pages 377-384, 2014. 4", + "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6, 7", + "[25] Roland Kays, Monica Lasky, Maximilian L Allen, Robert C Dowler, Melissa TR Hawkins, Andrew G Hope, Brooks A Kohli, Verity L Mathis, Bryan McLean, Link E Olson, et al. Which mammals can be identified from camera traps and crowdsourced photographs? Journal of Mammalogy, 2022. 5", + "[26] Muhammad Haris Khan, John McDonagh, Salman Khan, Muhammad Shahabuddin, Aditya Arora, Fahad Shahbaz Khan, Ling Shao, and Georgios Tzimiropoulos. Animalweb: A large-scale hierarchical dataset of annotated animal faces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6939-6948, 2020. 2", + "[27] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2. Cite-seer, 2011. 2", + "[28] Nicole Laws, Andre Ganswindt, Michael Heistermann, Moira Harris, Stephen Harris, and Chris Sherwin. A case study: fecal corticosteroid and behavior as indicators of welfare during relocation of an asian elephant. Journal of Applied Animal Welfare Science, 10(4):349-358, 2007. 2", + "[29] Weining Li, Sirmam Swetha, and Mubarak Shah. Wildlife action recognition using deep learning. 2020. 2, 3", + "[30] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvittv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13060", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 5, 6, 7", + "[31] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859-1868, 2021. 2", + "[32] Mackenzie Weygandt Mathis and Alexander Mathis. Deep learning tools for the measurement of animal behavior in neuroscience. Current opinion in neurobiology, 60:1-11, 2020. 2", + "[33] Joy Mench. Why it is important to understand animal behavior. ILAR journal, 39(1):20-26, 1998. 2, 4", + "[34] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386-12395, 2020. 2", + "[35] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Temporal action detection with global segmentation mask learning. European Conference on Computer Vision, 2022. 7", + "[36] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19023-19034, 2022. 1, 2, 3", + "[37] Kehinde Owoeye and Stephen Hailes. Online collective animal movement activity recognition. arXiv preprint arXiv:1811.09067, 2018. 2", + "[38] Shah Atiqur Rahman, Insu Song, Maylor KH Leung, Ickjai Lee, and Kyungmi Lee. Fast action recognition using negative space features. Expert Systems with Applications, 41(2):574-587, 2014. 2", + "[39] Maheen Rashid, Xiuye Gu, and Yong Jae Lee. Interspecies knowledge transfer for facial keypoint detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6894-6903, 2017. 2", + "[40] Lianne Robinson and Gernot Riedel. Comparison of automated home-cage monitoring systems: emphasis on feeding behaviour, activity and spatial learning following pharmacological interventions. Journal of neuroscience methods, 234:13-25, 2014. 2", + "[41] Moira Shooter, Charles Malleson, and Adrian Hilton. Sydog: A synthetic dog dataset for improved 2d pose estimation. arXiv preprint arXiv:2108.00249, 2021. 2", + "[42] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 6, 7", + "[43] Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation. Nature communications, 13(1):1-15, 2022. 2", + "[44] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 595-604, 2015. 2", + "[45] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2", + "[46] Grant Van Horn, Rui Qian, Kimberly Wilber, Hartwig Adam, Oisin Mac Aodha, and Serge Belongie. Exploring fine-grained audiovisual categorization with the ssw60 dataset. 2", + "[47] Grant Van Horn, Rui Qian, Kimberly Wilber, Hartwig Adam, Oisin Mac Aodha, and Serge Belongie. Exploring fine-grained audiovisual categorization with the ssw60 dataset, 2022. 2", + "[48] Lukas von Ziegler, Oliver Sturman, and Johannes Bohacek. Big behavior: challenges and opportunities in a new era of deep behavior profiling. Neuropsychopharmacology, 46(1):33-44, 2021. 2", + "[49] Carl Vondrick, Donald Patterson, and Deva Ramanan. Efficiently scaling up crowdsourced video annotation. International journal of computer vision, 101(1):184-204, 2013. 4", + "[50] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2", + "[51] Heng Yang, Renqiao Zhang, and Peter Robinson. Human and sheep facial landmarks localisation by triplet interpolated features. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1-8. IEEE, 2016. 2", + "[52] Can Zhang, Meng Cao, Dongming Yang, Jie Chen, and Yuexian Zou. Cola: Weakly-supervised temporal action localization with snippet contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16010-16019, June 2021. 7", + "[53] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, 2022. 5, 7, 8" + ], + "bbox": [ + 501, + 92, + 893, + 670 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13061", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_model.json b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c57c24cae9a2b4bb7d9d4471604440fd111dda26 --- /dev/null +++ b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.764, + 0.177 + ], + "angle": 0, + "content": "MammalNet: A Large-scale Video Benchmark for Mammal Recognition and Behavior Understanding" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.203, + 0.818, + 0.294 + ], + "angle": 0, + "content": "Jun Chen\\(^{1*}\\) Ming Hu\\(^{1*}\\) Darren J. Coker\\(^{1}\\) Michael L. Berumen\\(^{1}\\) \nBlair Costello\\(^{2,3}\\) Sara Beery\\(^{4}\\) Anna Rohrbach\\(^{5}\\) Mohamed Elhoseiny\\(^{1}\\) \n\\(^{1}\\)King Abdullah University of Science and Technology (KAUST) \n\\(^{2}\\)Max Planck Institute of Animal Behavior, \\(^{3}\\)University of Konstanz \n\\(^{4}\\)Massachusetts Institute of Technology, \\(^{5}\\)University of California, Berkeley" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.32, + 0.809, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.539, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Figure 1. We propose MammalNet, a large-scale video benchmark for recognizing mammals and their behavior. It is built around a biological mammal taxonomy spanning 17 orders, 69 families and 173 mammal categories, and includes 12 common high-level mammal behaviors (e.g., hunt, groom). MammalNet enables the study of animal and behavior recognition, both separately and jointly. It also facilitates investigating challenging compositional scenarios which test models' zero- and low-shot transfer abilities. Moreover, MammalNet includes behavior detection by localizing when a behavior occurs in an untrimmed video. Our dataset is the first to enable animal behavior analysis at scale in an ecologically-grounded manner, and exemplifies multiple challenges for the computer vision community, such as recognition of imbalanced, hierarchical distributions of fine-grained categories and generalization to unseen or seldom seen scenarios." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.653, + 0.314, + 0.669 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.472, + 0.868 + ], + "angle": 0, + "content": "Monitoring animal behavior can facilitate conservation efforts by providing key insights into wildlife health, population status, and ecosystem function. Automatic recognition of animals and their behaviors is critical for capitalizing on the large unlabeled datasets generated by modern video devices and for accelerating monitoring efforts at scale. However, the development of automated recognition systems is currently hindered by a lack of appropriately labeled datasets. Existing video datasets 1) do not classify animals according to established biological taxonomies; 2) are too small to facilitate large-scale behavioral studies and are often limited to a single species; and 3) do not feature temporally local-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.895, + 0.881 + ], + "angle": 0, + "content": "ized annotations and therefore do not facilitate localization of targeted behaviors within longer video sequences. Thus, we propose MammalNet, a new large-scale animal behavior dataset with taxonomy-guided annotations of mammals and their common behaviors. MammalNet contains over 18K videos totaling 539 hours, which is \\(\\sim 10\\) times larger than the largest existing animal behavior dataset [36]. It covers 17 orders, 69 families, and 173 mammal categories for animal categorization and captures 12 high-level animal behaviors that received focus in previous animal behavior studies. We establish three benchmarks on MammalNet: standard animal and behavior recognition, compositional low-shot animal and behavior recognition, and behavior detection. Our dataset and code have been made available at: https://mammal-net.github.io." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.214, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13052" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.116, + 0.47, + 0.251 + ], + "angle": 0, + "content": "Animal species are a core component of the world's ecosystems. Through their behavior, animals drive diverse ecological processes, including seed dispersal, nutrient cycling, population dynamics, speciation, and extinction. Thus, understanding and monitoring the behaviors of animals and their interactions with their physical and social environments is key to understanding the complexities of the world's ecosystems, an objective that is especially critical now given the ongoing biodiversity crisis [12]." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.253, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Modern sensors, including camera traps, drones, and smartphones, allow wildlife researchers, managers, and citizen scientists to collect video data of animal behavior on an unprecedented scale [43]. However, processing this data to generate actionable, timely insights remains a major challenge. Manual human review and annotation of footage to identify and locate species and behavioral sequences of interest is time-intensive and does not scale to large datasets. Thus, methods for automated animal and behavioral recognition could open the door to large-scale behavioral monitoring and speed up the time to produce usable data, thereby reducing the time to implement management directives." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.434, + 0.47, + 0.78 + ], + "angle": 0, + "content": "The first essential step to creating such an AI system for animal and behavior recognition is curating a diverse, representative dataset that allows us to formalize these challenges as computer vision tasks and benchmark potential solutions. Most previous datasets either only cover a limited number of animal and behavior types [4, 38], or do not implement animal labeling [36], or include a small number of videos with insufficient environmental diversity [4, 38, 48]. Recently, a dataset named \"Animal Kingdom\" [36] was proposed to study animal actions and is currently the largest existing behavioral dataset, to the best of our knowledge. However, it only contains 4,310 videos totaling 50 hours, which might be insufficient for large-scale animal behavior studies considering its diversity. Furthermore, the authors only focus on the recognition of atomic actions such as yawning, swimming, and flying. These basic actions cannot be easily matched to the higher-order behavioral states that are of primary interest to end users in animal management and conservation [6]. For example, a cheetah that is running may either be hunting, escaping, or playing. Finally, and most importantly, they do not support some important tasks such as animal recognition and behavior detection which are essential for animal behavior understanding." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.781, + 0.47, + 0.9 + ], + "angle": 0, + "content": "To overcome the limitations of previous datasets, we propose a new dataset called MammalNet. We specifically focus on mammals since they, unlike other animal classes such as birds or insects, usually have more diverse and distinguishable behavior statuses. MammalNet is comprised of 539 hours of annotated videos, which is \\(\\sim 10\\) times longer than that of the largest available animal behavior dataset. It contains 18,346 videos depicting 12 fundamental high-level" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "behaviors from hundreds of mammal species. Importantly, it focuses on 12 higher-order animal behaviors that are the focus of previous animal behavior literature [3,8,17,33], rather than atomic actions. MammalNet also categorizes animals according to the scientific taxonomy available in Wikipedia, as we show in Fig. 1; hence the dataset can be flexibly expanded in the future by following the same protocols. It includes videos of approximately 800 mammal species in 173 mammal categories. We establish three benchmarks inspired by ecological research needs - standard animal & behavior classification, compositional low-shot animal & behavior recognition, and behavior detection - to promote future study in animal behavior understanding." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Through our experiments, we find that: (1) Correctly recognizing the animals and behaviors is a challenging task even for the state-of-the-art models, especially for less-frequent animals. The top-1 per-class accuracy is 32.5 for animal recognition, 37.8 for behavior recognition, and 17.8 for their joint recognition in our best-performing model. (2) Behavior recognition for unseen animals can be transferred from observations of other seen animals due to their similar features such as appearance and movement style, which can help in studies of animals with less available data. However, to achieve more accurate behavior recognition, having access to videos of the target animals and behaviors is still crucial." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.483, + 0.64, + 0.498 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.508, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Automatic animal recognition and behavior detection can help humans monitor and efficiently process animal behavior data [7, 22, 32, 34, 40, 47, 48]. It can massively reduce the labour cost from manually collecting and analyzing animal activities. During the past few years, many datasets [18, 20, 29, 38] have been introduced to develop foundations for animal behavior research. We systematically analyze the previous datasets and summarize several important limitations that prevent them from being used for large-scale animal recognition and behavior understanding:" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Lack of behavior understanding. Many previous datasets only focus on animal recognition or pose estimation from images, but lack behavior learning. For example, iNaturalist [45] collects 859,000 images covering more than 5,000 different types of plants and animals. NABird [44] collects 48,562 images of 555 different North American bird species. Also, some works focus on narrow animal recognition such as dogs [27], birds [46, 50] or cats [13]. On the other hand, many works also focus on pose estimation [10, 18, 21, 31, 41] and animal face detection [26, 39, 51]. They are generally not applicable to learning animal behavior." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Lack of taxonomic diversity for different behaviors. Previous animal behavior datasets have minimal taxonomic coverage - often containing just a single animal species. For example, there are existing behavior recognition datasets for elephants [28], sheep [37], monkeys [5], tigers [16]," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13053" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.855, + 0.242 + ], + "angle": 0, + "content": "
DatasetsDataset PropertiesTasks
Publicly Available?Taxonomy-guided Animal Annotation?No. of VideosNo. of ActionsNo. of BehaviorsNo. of Animal CategoriesNo. of Mammal CategoriesTotal DurationAnimal ClassificationAction/Behavior RecognitionAction/Behavior Detection
Wild Felines [20]××2,7003-33-✓✓×
Wildlife Actions [29]××10,6007-3211-✓✓×
Animal Kingdom [36]✓×4,301140-850-50 (h)×✓×
MammalNet (ours)✓✓18,346-12173173539 (h)✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.264, + 0.895, + 0.308 + ], + "angle": 0, + "content": "Table 1. The comparison among existing animal behavior understanding video datasets. Compared to other datasets, MammalNet annotates the animals by following the scientific mammal taxonomy, focuses on more high-level behavior recognition, has the largest number of mammal categories, collects the largest number of animal behavior videos, totalling 539 hours, and also enables behavior detection tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.471, + 0.391 + ], + "angle": 0, + "content": "etc. While useful for the studied species, these datasets do not enable the exploration of behavior across species which is necessary to scale up behavior identification without requiring training examples of every possible combination of species and behavior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.472, + 0.534 + ], + "angle": 0, + "content": "Lack of taxonomy-guided animal annotation. Previous animal behavior datasets either do not include animal recognition as a task [36] or group species according to subjective as opposed to scientific criteria [29, 36]. In contrast, our dataset collects and annotates the videos according to the scientific mammal taxonomy. By following this taxonomy we allow exploration of behavior from an evolutionary perspective, as well as enable standardized and consistent dataset expansion under the same protocol." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.541, + 0.47, + 0.678 + ], + "angle": 0, + "content": "Actions vs. Behaviors. Previous works mostly focus on atomic action recognition [20, 29, 36]. For example, some of their action classes are walk, stand still, fly, etc. In contrast, MammalNet focuses on higher-level animal behaviors, such as hunt, feed baby, etc. Behavior here denotes the main activity instance during a period, and it is usually composed of multiple atomic actions. These complex behaviors are needed to describe and summarize video activity bouts in a manner which is valuable for ecological research." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.684, + 0.471, + 0.73 + ], + "angle": 0, + "content": "We further compare our MammalNet dataset with the other existing animal behavior understanding video datasets, and summarize the key difference in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.323, + 0.781 + ], + "angle": 0, + "content": "3. Constructing MammalNet" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.901 + ], + "angle": 0, + "content": "The goal of MammalNet is to provide a large-scale mammal video dataset that benchmarks both animal and behavior recognition. In this section, we discuss our dataset construction protocol, including the choice of scientific animal taxonomy, crowdsourced annotations, and performing manual quality control during video collection and annotation. Finally, we describe the statistical profile of MammalNet." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.314, + 0.871, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.583, + 0.892, + 0.611 + ], + "angle": 0, + "content": "Figure 2. A subset of the mammal taxonomy of MammalNet. It includes 3 orders, 11 families, and 45 genera." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.868, + 0.645 + ], + "angle": 0, + "content": "3.1. Animal Taxonomy and Behavior Collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.895, + 0.852 + ], + "angle": 0, + "content": "Animal taxonomy construction. Accurate animal recognition is one of the main goals defined by MammalNet. To ensure this task is scientifically relevant, we aimed to collect and structure our list of target animals based on the mammal taxonomy. First, we collected a diverse list of mammals, approximately 800 mammal species and sub-species, from the National Geographic [2] and Animal A-Z [1] websites. Next, we manually mapped each mammal onto the taxonomic structure (including class, order, family, genus, tribe, sub-family and species) in mammal taxonomy from Wikipedia. In total, the classes included in MammalNet cover 17 orders, 69 families, and 173 mammal categories. A taxonomic subset is visualized in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Behavior collection. We aim to enable the study of complex, high-level animal behavior as opposed to the simpler atomic actions emphasized in previous work [29, 36]. Be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13054" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.092, + 0.831, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.895, + 0.283 + ], + "angle": 0, + "content": "Figure 3. The examples for the annotated target behavior boundaries. The frames marked in red boxes denote the annotated temporal boundaries for the target behavior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.473, + 0.415 + ], + "angle": 0, + "content": "havior here represents the major activity being displayed during a period of a video, and can be viewed as a series of atomic actions collectively serving a higher-level purpose. For example, hunting behavior, as defined in MammalNet, can often be decomposed into running, chasing and killing actions, etc. Identification of behaviors at this level of definition is more useful for ecologists and zoologists [3,8,17,33] compared to atomic actions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.472, + 0.463 + ], + "angle": 0, + "content": "Inspired by previous biological and ecological studies [3, 8,17], we consider 12 fundamental mammal behaviors under 5 different groups in our study. They are respectively:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.464, + 0.404, + 0.479 + ], + "angle": 0, + "content": "Foraging behaviors: eat food, drink water, hunt." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.48, + 0.434, + 0.495 + ], + "angle": 0, + "content": "Reproductive behaviors: mate, feed baby, give birth." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.497, + 0.267, + 0.511 + ], + "angle": 0, + "content": "Hygiene behaviors: groom." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.513, + 0.267, + 0.528 + ], + "angle": 0, + "content": "Agnostic behaviours: fight." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.53, + 0.453, + 0.545 + ], + "angle": 0, + "content": "Maintenance behaviors: urinate, defecate, sleep, vomit." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.558, + 0.421, + 0.574 + ], + "angle": 0, + "content": "3.2. Video Collection and Quality Assurance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.472, + 0.703 + ], + "angle": 0, + "content": "Dataset curation is usually a costly process requiring a lot of manual annotation by humans. Some datasets are annotated by domain experts to encourage more reliable labeling, particularly for challenging tasks, but this expertise comes at a much higher cost and is thus hard to scale. We adapted a semi-automatic crowdsourcing approach to collect and annotate our datasets, inspired by many previous works [9, 15, 23, 49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Online video retrieval. Our goal was to collect videos depicting each animal in our database performing each of the 12 focal behaviors. To achieve this, we queried YouTube with the text combinations of each animal common name and behavior, e.g., tiger hunting, and downloaded videos where the title included our queried animal name and behavior. However, some animal names are too rare to generate enough relevant videos, e.g., Brown hyena or Spotted hyena. In such cases, we used a more common name, hyena, to represent the union of those animals. In order to retrieve more relevant videos from the search engine, we also expanded each animal with their synonyms as given in Wikipedia. For example, we expanded the original artic fox with other equivalent names" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.892, + 0.324 + ], + "angle": 0, + "content": "such as white fox, polar fox and snow fox. Each video was downloaded at its highest available resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.894, + 0.4 + ], + "angle": 0, + "content": "Data filtering. During video retrieval, we downloaded the videos that are accessible for people \\(< 16\\) years old to avoid the videos with violent content. We also prioritized videos that are shorter than 10 minutes in duration to limit the total storage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.895, + 0.521 + ], + "angle": 0, + "content": "However, some videos might have irrelevant content due to the inaccuracy of text-based retrieval. For example, the downloaded videos might 1) display cartoon or toy animals or unrealistic environments such as games or movies, 2) display a static image instead of a continuous video, 3) involve a lot of human-animal interaction, e.g., human feeding an animal, as opposed to focusing on animal behavior, 4) not contain the specified animal and/or behavior." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.718 + ], + "angle": 0, + "content": "To alleviate these issues, we employed Amazon Mechanical Turk workers to verify the presence of the animal and behavior, and identify other quality issues. We assigned each video to three different workers and provided them with the pictures of an animal with its common name, and an expected behavior. We asked the workers to verify if this animal and behavior indeed appear in the video. Only the videos that \"pass\" by all the three workers were kept for the following behavior localization annotation. Before the workers started the verification, we first provided them the verification instructions and asked them to complete a corresponding qualification test with 20 multiple choice questions to ensure only qualified workers could participate in our task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.858, + 0.742 + ], + "angle": 0, + "content": "3.3. Animal Behavior Localization Annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The target behavior typically does not span the whole video. To localize the part where the target behavior is actually occurring in the video, we asked the AMT workers to manually annotate the respective temporal boundaries. Namely, we asked five different qualified workers to annotate the start and end frames for the target behavior in each video. In the end, each video received at least 5 annotated behavior boundaries. To achieve robust annotation agreement, we used the complete linkage algorithm [14] to cluster different temporal boundaries and merge them into one or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13055" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.092, + 0.831, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.358, + 0.855, + 0.373 + ], + "angle": 0, + "content": "Figure 4. Number of videos per each mammal category. We rank the categories according to their trimmed videos frequency." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.382, + 0.413, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.525, + 0.473, + 0.555 + ], + "angle": 0, + "content": "Figure 5. Number of videos per each behavior. We rank the behavior according to their trimmed videos frequency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.472, + 0.643 + ], + "angle": 0, + "content": "several more stable ones that received multiple agreements. Note, that a single video might have multiple discrete occurrences of a given behavior, and thus have multiple boundary definitions. We show several examples for annotated target behavior boundaries in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.465, + 0.667 + ], + "angle": 0, + "content": "3.4. Recognition at lowest feasible taxonomic level" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.472, + 0.84 + ], + "angle": 0, + "content": "We categorize animals according to the lowest and feasible taxonomic classification, rather than the species level, for the following reasons: 1) YouTube videos contain the ambiguous and inexpert species labels. 2) We are not using expert annotators, and even experts often cannot reliably identify animals at the species or even genus level based on crowdsourced or field images [25]. Thus, it is more practical to classify animals in our dataset to the lowest feasible taxonomic level rather than species. As a result, our taxonomy contains 173 distinguishable taxonomic classification levels including: sub-family, tribe and genus." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.285, + 0.863 + ], + "angle": 0, + "content": "3.5. MammalNet Statistics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The final MammalNet dataset contains 18,346 videos (539 hours); after we trim the videos according to the anno" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.895, + 0.552 + ], + "angle": 0, + "content": "tated behavior boundaries, it increases to 20,033 trimmed video instances (394 hours). In total, it covers 173 mammal categories, 69 families and 17 orders in our dataset, with the total of 173 mammal categories defined for our recognition tasks. MammalNet also contains 12 different common behaviors. The average duration for the untrimmed and trimmed videos is 106 and 77 seconds, respectively. Over \\(54\\%\\) of videos reach HD resolutions (\\(1280 \\times 720\\)). We demonstrate the data distribution in terms of each animal category and behavior in Fig. 4 and 5. We observe that the number of collected videos per type follows a long-tail distribution." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.567, + 0.707, + 0.584 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.895, + 0.73 + ], + "angle": 0, + "content": "We construct three main tasks on MammalNet: 1) Standard animal and behavior classification on trimmed videos, 2) Compositional low-shot animal and behavior recognition on trimmed videos, and 3) Behavior detection on untrimmed videos. In both the classification and detection tasks, we baseline several state-of-the-art models [19, 30, 53] that have been successfully applied to human action recognition and detection. In the following, we describe the formulation of each challenge and provide baselines and analysis." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.74, + 0.892, + 0.771 + ], + "angle": 0, + "content": "4.1. Standard Animal and Behavior Classification on Trimmed Videos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.895, + 0.84 + ], + "angle": 0, + "content": "This task explores classification of both the primary behavior that occurs in a trimmed video and the animal that performs the behavior. We report the top-1 per-example and per-class accuracy for all the baseline models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Many, medium, few splits. To capture the effect of the long-tailed nature of the MammalNet dataset, we group the animal, behavior, and their composition classes into many, medium, and few based on their frequency, and report the average" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13056" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.865, + 0.266 + ], + "angle": 0, + "content": "
BaselinesAnimal ClassificationBehavior ClassificationJoint Classification
Many 12Medium 28Few 133All 173Many 4Medium 4Few 4All 12Many 33Medium 180Few 823All 1036
SlowFast [19]49.635.69.517.239.027.614.927.227.219.44.48.6
C3D [42]48.635.310.017.538.227.811.725.928.118.84.58.6
I3D [11]48.834.910.517.839.527.214.827.229.620.54.48.9
MViT V2 [30]48.535.512.919.742.329.211.627.729.519.64.89.0
SlowFast*58.343.116.624.545.132.714.830.938.623.57.312.1
C3D*58.345.419.126.844.636.015.932.238.026.48.413.5
I3D*58.642.916.924.846.335.014.832.138.324.58.613.3
MViT V2*66.756.023.432.550.942.420.037.846.233.011.817.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.287, + 0.895, + 0.331 + ], + "angle": 0, + "content": "Table 2. Per-class Top-1 accuracy for animal, behavior and their joint prediction.* denotes the initialization from the model pretrained on Kinetics 400 [24]. Transfer learning from the human action to the animal behavior recognition receives considerable performance gain. Best performance for each split has been highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.339, + 0.422, + 0.487 + ], + "angle": 0, + "content": "
BaselinesAnimalBehaviorJoint
SlowFast [19]35.434.217.4
C3D [42]35.033.517.1
I3D [11]35.234.317.9
MViT V2 [30]35.636.818.0
SlowFast*43.039.422.8
C3D*44.440.324.6
I3D*43.441.224.0
MViT V2*52.646.630.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.509, + 0.471, + 0.539 + ], + "angle": 0, + "content": "Table 3. Per-example accuracy for animal, behavior and their joint prediction. * denotes the initialization from the pretrained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.472, + 0.684 + ], + "angle": 0, + "content": "per-class accuracy bands chosen based on the frequency percentiles. For animal categories this is broken down as many: top \\(7\\%\\) frequent classes, medium: middle \\(16\\%\\) classes, and few: the remaining \\(77\\%\\) classes. For behavior, many: top \\(33\\%\\) frequent classes, medium: middle \\(33\\%\\), and few: the remaining \\(33\\%\\) classes. For joint classification, many: top \\(3\\%\\) frequent classes, medium: middle \\(17\\%\\) classes, few: the remaining \\(80\\%\\) classes. We show the number of classes per each split in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.472, + 0.748 + ], + "angle": 0, + "content": "Dataset setup. We randomly split the examples from each animal-behavior category into \\(70\\%\\) for training, \\(10\\%\\) for validation, and \\(20\\%\\) for testing, and it results in 14,554 training, 1,638 validation, and 3,841 testing videos, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Baselines. We compare SlowFast [19], I3D [11], C3D [42] and MViT V2 [30] models on our tasks. These models are evaluated in two versions: 1) Training with random initialization 2) Initializing with weights from a model pretrained on Kinetics 400 [24]. These methods were originally designed for human action recognition and hence do not have an ability to predict both an action and a subject by default. To accommodate them into our joint prediction setting, we have two task heads, one for animal category recognition and one for behavior recognition. We compute the joint" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.895, + 0.404 + ], + "angle": 0, + "content": "loss, \\(\\mathcal{L}_{joint}\\), for both animal and behavior classification as shown in Fig. 1. We tune all the hyper-parameters on the validation data. The final hyper-parameters for each model are provided in the supplement. The loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.415, + 0.893, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = - \\frac {1}{M} \\sum_ {i} ^ {M} y _ {i} ^ {a} \\log \\left(p _ {i} ^ {a}\\right) - \\frac {1}{N} \\sum_ {j} ^ {N} y _ {j} ^ {b} \\log \\left(p _ {j} ^ {b}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.895, + 0.547 + ], + "angle": 0, + "content": "where \\(M\\) is the number of animal classes, and \\(N\\) is the number of behavior classes, \\(p_i^a\\) and \\(y_{i}^{a}\\) denote the animal prediction probability and ground truth label for the category \\(i\\), \\(p_y^b\\) and \\(y_j^b\\) denote the behavior prediction probability and ground truth label for the category \\(j\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.897, + 0.805 + ], + "angle": 0, + "content": "Experimental results. The results for per-class and per-example classification are summarized in Tables 2 and 3, respectively. We find that MViT v2 is able to achieve competitive results for all the splits. The best top-1 joint per-class accuracy is 17.8 and per-example accuracy is 30.6, which points to significant room for improvement on these challenging tasks. We also observe that transfer learning from the model that is pretrained on Kinetics 400, a human action recognition dataset, improves both the animal and behavior classification accuracy (with MViT v2, this corresponds to a per-class accuracy gain from 19.7 to 32.5 for animal classification, and from 27.7 to 37.8 for behavior classification). Additionally, we find that performance gain from pretraining is higher for frequently occurring animals and behaviors, indicated by the results shown in many, medium, and few splits. Accurately predicting low-frequency animal and behavior categories remains a significant challenge." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.816, + 0.895, + 0.848 + ], + "angle": 0, + "content": "4.2. Compositional Low-shot Animal and Behavior Classification on Trimmed Videos" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "It is hard to find a sufficient number of labeled behavior samples for all the animals in our collected taxonomy. For example, numbat and florida panther have very few behavior" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13057" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.152, + 0.089, + 0.822, + 0.227 + ], + "angle": 0, + "content": "
BaselinesCompositional Low-Shot Behavior Classification
0-shot1-shot5-shot
Per-example A / BPer-class A / BPer-example A / BPer-class A / BPer-example A / BPer-class A / B
C3D*27.4 / 23.718.3 / 16.129.2 / 25.521.1 / 19.933.3 / 29.325.2 / 23.0
I3D*25.3 / 23.916.7 / 15.226.8 / 25.716.9 / 18.330.5 / 28.321.7 / 21.9
SlowFast*26.2 / 24.817.9 / 16.326.5 / 26.316.7 / 19.029.6 / 29.222.5 / 22.4
MViT V2*32.2 / 26.220.7 / 18.133.7 / 28.922.9 / 22.539.3 / 31.731.0 / 26.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.25, + 0.894, + 0.278 + ], + "angle": 0, + "content": "Table 4. Compositional low-shot animal and behavior recognition. * denotes the initialization from the model pretrained on Kinetics 400 [24]. \"A\" denotes the animal category and \"B\" denotes the behavior category. The best performance per each column has been highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.29, + 0.466, + 0.362 + ], + "angle": 0, + "content": "
BaselinesmAP
0.500.600.700.800.90Avg.
CoLA [52]26.0222.7018.9813.463.0515.81
TAGS [35]23.0920.9719.0916.9812.5617.63
ActionFormer [53]28.4826.1423.1718.6910.4820.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.415 + ], + "angle": 0, + "content": "Table 5. The results for behavior detection. We report mAP at the IoU thresholds of [0.5:0.1:0.9]. Average mAP is computed by averaging different tIoU thresholds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.534 + ], + "angle": 0, + "content": "annotations. However, it is plausible to imagine that behaviors can be transferred among different animals that have similar appearances and movement styles. Also the animal recognition under different behaviors (hunting, fighting) can be mutually transferred. To investigate these phenomena, we design the compositional low-shot animal and behavior classification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.471, + 0.718 + ], + "angle": 0, + "content": "Dataset setup. We first select the animal-behavior compositional classes that contain more than 5 examples and allocate \\(25\\%\\) of classes to the test set (4,088 videos). For the remaining classes (the other \\(75\\%\\) of classes and those classes with \\(\\leq 5\\) examples), we randomly designate \\(90\\%\\) of classes as the training set and \\(10\\%\\) as the validation set (898 videos). Under the low-shot scheme, for each compositional class, we randomly sample 5 examples from the test set and move 0, 1, or 5 of them into the training set for the zero-shot, 1-shot, and 5-shot setup, respectively (14,377, 14,511 and 15,047 training videos). The train, val and test sets consist of 983, 53, and 134 compositional classes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Baselines. We evaluate the compositional low-shot classification with the SlowFast [19], I3D [11], C3D [42] and MViT V2 [30] models under the joint loss. We initialize these models with the weights pretrained on Kinetics 400 [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Experimental results. We summarize the results in Table 4. It shows that MViT v2 consistently achieves the best performance under our low-shot setup. The behavior classification can still achieve 26.2 per-example and 18.1 top-1 accuracy under the zero-shot setting for MViT v2 model. This indicates that behavior classification is transferable from other animals in some cases. Additionally, we observe that its performance is improved to 31.7 per-example and 26.0 per-class" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.352 + ], + "angle": 0, + "content": "top-1 accuracy under 5-shot setup, indicating that training on more videos with the behaviors from the same animal is still necessary. A similar phenomenon is observed to the few-shot animal recognition." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.364, + 0.857, + 0.379 + ], + "angle": 0, + "content": "4.3. Behavior Detection on Untrimmed Videos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.524 + ], + "angle": 0, + "content": "This task is to detect the behavior in untrimmed videos. The behavior detection algorithm should correctly detect the temporal range for the primary activity presented in the video. We follow previous temporal action localization works [9, 53] to benchmark this task. We report the mean Average Precision (mAP) with different temporal intersections over the union (tIoU) thresholds [0.5:0.1:0.9] as our evaluation metric. We also report the average mAP averaging across different tIoUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Dataset setup. Similar to the previous standard animal and behavior classification, we follow the [train:0.7, val:0.1, test:0.2] ratios to split the untrimmed videos at the animal-behavior composition level. This results in 13,318 videos for training, 1,486 for validation and 3,542 for testing. We tune all the hyper-parameters based on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Baselines. We evaluate our datasets with the baseline models such as ActionFormer [53], TAGS [35], and CoLA [52]. To produce the features for our MammalNet videos, we first finetune a two-stream I3D [11] model, that is originally pretrained on ImageNet [15] and Kinetics 400 [24], on our dataset, and then extract the RGB and optical flow features for each video. We concatenate these two features together as the model input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Experimental results. We show the behavior detection results in Table 5. Among all the baselines, ActionFormer demonstrates the most competitive performance with an average mAP of 20.07, and also achieves \\(28.48\\mathrm{mAP}\\) for the threshold of 0.5. Overall, it is clear to see that the behavior detection task is still very challenging for current methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.597, + 0.861 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Demonstration of animal and behavior classification. We sample some prediction examples from MViT v2 [30] for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13058" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.091, + 0.849, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.895, + 0.302 + ], + "angle": 0, + "content": "Figure 6. The visualization presents various instances of joint animal and behavior classification. In the third column, accurate predictions are displayed, while the fourth, fifth, and sixth columns showcase mispredicted examples where either the animal or the behavior does not correspond to the correct prediction." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.308, + 0.468, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.466, + 0.454, + 0.48 + ], + "angle": 0, + "content": "Figure 7. The visualization for behavior detection examples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.49, + 0.471, + 0.549 + ], + "angle": 0, + "content": "the joint categories Acinonyx hunt, Canis eat and Antilopini fight. We demonstrate one correct prediction and three mispredicted examples where the model mis-recognizes the animal or behavior or both of them in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.551, + 0.472, + 0.642 + ], + "angle": 0, + "content": "Demonstration of behavior detection: We visualize behavior detection results in Fig. 7 from ActionFormer [53]. The top example shows correctly predicted behavior with a proposal closely aligned with the ground-truth. The bottom example shows a misclassified behavior, and it mistakenly predicts the hunt behavior as fight." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.472, + 0.809 + ], + "angle": 0, + "content": "Joint animal and behavior recognition vs. separate recognition. We also train the model for recognizing the animal and behavior separately and provide the results in the Table 1 of the supplementary. Comparing with the joint recognition, we find that training a system to recognize the animal and behavior together can improve the behavior recognition under separate training by \\(\\sim 2.2\\) per-class accuracy. This indicates that being capable of understanding the animal types can benefit behavior prediction. However, the results also indicate that predicting animal only can be more useful in recognizing animal types." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Dataset bias. Our dataset is downloaded from YouTube channels with diverse backgrounds and video quality, and it might differ from video datasets collected for behavioral or ecological studies. The videos on YouTube might also exhibit potential biases:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.472, + 0.901 + ], + "angle": 0, + "content": "1) Over-representation of some behaviors and under" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.895, + 0.414 + ], + "angle": 0, + "content": "representation of others. For example, people might prefer watching videos with fight or eat food activities, and these two behaviors are more likely to be over-represented, while urinate and defecate behaviors are less interesting to humans and hence are underrepresented on YouTube. However, they are still important in informing conservation-related actions to protect the environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.477 + ], + "angle": 0, + "content": "2) Bias towards captive animals or wild animals that are habituated to humans. We find that many videos are shot at zoos, farms, and homes, etc. These animals may display different behaviors than wild or non-habituated animals." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.495, + 0.62, + 0.511 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.793 + ], + "angle": 0, + "content": "We introduced MammalNet, a large-scale video dataset for mammal recognition and behavior understanding. We have collected videos for hundreds of different mammals and structured them by following the scientific mammal taxonomy. MammalNet consists of 18,346 untrimmed videos covering 173 mammal categories and 12 common behaviors. We established three challenges: standard animal and behavior classification, compositional low-shot animal and behavior classification, and behavior detection. Through our experiments, we found that the accurate recognition of animals at scale and their common behaviors is very challenging even with current state-of-the-art models, especially when the dataset has a long-tail distribution. We also found that learning to recognize the behavior of unseen animals is possible via transfer from the other seen animals. To promote further research and development in the field of animal behavior study, we have open-sourced all of our data and code to the research community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.897, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement: This work is supported by KAUST BAS/1/1685-01-01 and KAUST FCC/1/1973-58-01 (Red Sea Research Center), DARPA's SemaFor and PTG programs, the Caltech Resnick Sustainability Institute, and Germany's Excellence Strategy–‘Centre for the Advanced Study of Collective Behaviour’ EXC 2117-422037984." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13059" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.424, + 0.128 + ], + "angle": 0, + "content": "[1] A-z animals. https://a-z-animals.com/.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.131, + 0.473, + 0.157 + ], + "angle": 0, + "content": "[2] National geographics. https://www.nationalgeographic.com/.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.16, + 0.471, + 0.184 + ], + "angle": 0, + "content": "[3] John Alcock. Animal behavior: An evolutionary approach. Sinauer Associates, 2009. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.187, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[4] David J Anderson and Pietro Perona. Toward a science of computational ethology. Neuron, 84(1):18-31, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.216, + 0.471, + 0.283 + ], + "angle": 0, + "content": "[5] Praneet C Bala, Benjamin R Eisenreich, Seng Bum Michael Yoo, Benjamin Y Hayden, Hyun Soo Park, and Jan Zimmermann. Automated markerless pose estimation in freely moving macaques with openmonkeystudio. Nature communications, 11(1):1-12, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.285, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[6] Oded Berger-Tal, Daniel T Blumstein, Scott Carroll, Robert N Fisher, Sarah L Mesnick, Megan A Owen, David Saltz, Colleen Cassady St Claire, and Ronald R Swaisgood. A systematic survey of the integration of animal behavior into conservation. Conservation Biology, 30(4):744-753, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.355, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[7] Cigdem Bayan and Robert B Fisher. Detection of abnormal fish trajectories using a clustering based hierarchical classifier. In BMVC, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.398, + 0.471, + 0.423 + ], + "angle": 0, + "content": "[8] Michael D Breed and Janice Moore. Animal behavior. Academic Press, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.426, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.496, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[10] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9498-9507, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[12] Gerardo Ceballos, Paul R Ehrlich, and Peter H Raven. Vertebrates on the brink as indicators of biological annihilation and the sixth mass extinction. Proceedings of the National Academy of Sciences, 117(24):13596-13602, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[13] Yu-Chen Chen, Shintami C Hidayati, Wen-Huang Cheng, Min-Chun Hu, and Kai-Lung Hua. Locality constrained sparse representation for cat recognition. In International Conference on Multimedia Modeling, pages 140-151. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[14] Daniel Defays. An efficient algorithm for a complete link method. The Computer Journal, 20(4):364-366, 1977. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[16] Rod K Dishman, Andrew S Jackson, and Molly S Bray. Self-regulation of exercise behavior in the tiger study. Annals of Behavioral Medicine, 48(1):80-91, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[17] Lee Alan Dugatkin. Principles of animal behavior. University of Chicago Press, 2020. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[18] Cheng Fang, Tiemin Zhang, Haikun Zheng, Junduan Huang, and Kaixuan Cuan. Pose estimation and behavior classification of broiler chickens based on deep neural networks. Computers and Electronics in Agriculture, 180:105863, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[20] Liqi Feng, Yaqin Zhao, Yichao Sun, Wenxuan Zhao, and Jiaxi Tang. Action recognition using a spatial-temporal network for wild felines. Animals, 11(2):485, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[21] Jacob M Graving, Daniel Chae, Hemal Naik, Liang Li, Benjamin Koger, Blair R Costelloe, and Iain D Couzin. Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. *Elife*, 8:e47994, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.894, + 0.388 + ], + "angle": 0, + "content": "[22] Shilpi Gupta, Prerana Mukherjee, Santanu Chaudhury, Brejesh Lall, and Hemanth Sanisetty. Dftnet: Deep fish tracker with attention mechanism in unconstrained marine environments. IEEE Transactions on Instrumentation and Measurement, 70:1-13, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.894, + 0.446 + ], + "angle": 0, + "content": "[23] Fabian Caba Heilbron and Juan Carlos Niebles. Collecting and annotating human activities in web videos. In Proceedings of International Conference on Multimedia Retrieval, pages 377-384, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.894, + 0.516 + ], + "angle": 0, + "content": "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.894, + 0.588 + ], + "angle": 0, + "content": "[25] Roland Kays, Monica Lasky, Maximilian L Allen, Robert C Dowler, Melissa TR Hawkins, Andrew G Hope, Brooks A Kohli, Verity L Mathis, Bryan McLean, Link E Olson, et al. Which mammals can be identified from camera traps and crowdsourced photographs? Journal of Mammalogy, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.894, + 0.673 + ], + "angle": 0, + "content": "[26] Muhammad Haris Khan, John McDonagh, Salman Khan, Muhammad Shahabuddin, Aditya Arora, Fahad Shahbaz Khan, Ling Shao, and Georgios Tzimiropoulos. Animalweb: A large-scale hierarchical dataset of annotated animal faces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6939-6948, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.894, + 0.742 + ], + "angle": 0, + "content": "[27] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2. Cite-seer, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.894, + 0.813 + ], + "angle": 0, + "content": "[28] Nicole Laws, Andre Ganswindt, Michael Heistermann, Moira Harris, Stephen Harris, and Chris Sherwin. A case study: fecal corticosteroid and behavior as indicators of welfare during relocation of an asian elephant. Journal of Applied Animal Welfare Science, 10(4):349-358, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[29] Weining Li, Sirmam Swetha, and Mubarak Shah. Wildlife action recognition using deep learning. 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[30] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvittv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13060" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[31] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859-1868, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.472, + 0.258 + ], + "angle": 0, + "content": "[32] Mackenzie Weygandt Mathis and Alexander Mathis. Deep learning tools for the measurement of animal behavior in neuroscience. Current opinion in neurobiology, 60:1-11, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.287 + ], + "angle": 0, + "content": "[33] Joy Mench. Why it is important to understand animal behavior. ILAR journal, 39(1):20-26, 1998. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[34] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386-12395, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.472, + 0.385 + ], + "angle": 0, + "content": "[35] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Temporal action detection with global segmentation mask learning. European Conference on Computer Vision, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[36] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19023-19034, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.472, + 0.496 + ], + "angle": 0, + "content": "[37] Kehinde Owoeye and Stephen Hailes. Online collective animal movement activity recognition. arXiv preprint arXiv:1811.09067, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[38] Shah Atiqur Rahman, Insu Song, Maylor KH Leung, Ickjai Lee, and Kyungmi Lee. Fast action recognition using negative space features. Expert Systems with Applications, 41(2):574-587, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.472, + 0.608 + ], + "angle": 0, + "content": "[39] Maheen Rashid, Xiuye Gu, and Yong Jae Lee. Interspecies knowledge transfer for facial keypoint detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6894-6903, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.472, + 0.676 + ], + "angle": 0, + "content": "[40] Lianne Robinson and Gernot Riedel. Comparison of automated home-cage monitoring systems: emphasis on feeding behaviour, activity and spatial learning following pharmacological interventions. Journal of neuroscience methods, 234:13-25, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.472, + 0.719 + ], + "angle": 0, + "content": "[41] Moira Shooter, Charles Malleson, and Adrian Hilton. Sydog: A synthetic dog dataset for improved 2d pose estimation. arXiv preprint arXiv:2108.00249, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.472, + 0.788 + ], + "angle": 0, + "content": "[42] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[43] Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation. Nature communications, 13(1):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[44] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 595-604, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[45] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[46] Grant Van Horn, Rui Qian, Kimberly Wilber, Hartwig Adam, Oisin Mac Aodha, and Serge Belongie. Exploring fine-grained audiovisual categorization with the ssw60 dataset. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[47] Grant Van Horn, Rui Qian, Kimberly Wilber, Hartwig Adam, Oisin Mac Aodha, and Serge Belongie. Exploring fine-grained audiovisual categorization with the ssw60 dataset, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.894, + 0.401 + ], + "angle": 0, + "content": "[48] Lukas von Ziegler, Oliver Sturman, and Johannes Bohacek. Big behavior: challenges and opportunities in a new era of deep behavior profiling. Neuropsychopharmacology, 46(1):33-44, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.894, + 0.457 + ], + "angle": 0, + "content": "[49] Carl Vondrick, Donald Patterson, and Deva Ramanan. Efficiently scaling up crowdsourced video annotation. International journal of computer vision, 101(1):184-204, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.894, + 0.5 + ], + "angle": 0, + "content": "[50] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.894, + 0.558 + ], + "angle": 0, + "content": "[51] Heng Yang, Renqiao Zhang, and Peter Robinson. Human and sheep facial landmarks localisation by triplet interpolated features. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1-8. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.894, + 0.628 + ], + "angle": 0, + "content": "[52] Can Zhang, Meng Cao, Dongming Yang, Jie Chen, and Yuexian Zou. Cola: Weakly-supervised temporal action localization with snippet contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16010-16019, June 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.894, + 0.671 + ], + "angle": 0, + "content": "[53] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, 2022. 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13061" + } + ] +] \ No newline at end of file diff --git a/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_origin.pdf b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7f74b002f84ede2185a14dc4603c06ddd3d4a3c2 --- /dev/null +++ b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/14f4b617-eec8-453d-832d-c91f7b0aa424_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8872a3752445791842495ae7a22fbfd5081fb5ca1e71167994f88fcb2f240969 +size 4300644 diff --git a/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/full.md b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..68733350e048239e061659c554ca23879089db33 --- /dev/null +++ b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/full.md @@ -0,0 +1,272 @@ +# MammalNet: A Large-scale Video Benchmark for Mammal Recognition and Behavior Understanding + +Jun Chen $^{1*}$ Ming Hu $^{1*}$ Darren J. Coker $^{1}$ Michael L. Berumen $^{1}$ +Blair Costello $^{2,3}$ Sara Beery $^{4}$ Anna Rohrbach $^{5}$ Mohamed Elhoseiny $^{1}$ $^{1}$ King Abdullah University of Science and Technology (KAUST) + $^{2}$ Max Planck Institute of Animal Behavior, $^{3}$ University of Konstanz + $^{4}$ Massachusetts Institute of Technology, $^{5}$ University of California, Berkeley + +![](images/c6fdda67a752d39e5f70ee8b6fc5f275b719529b94fbef76aea0a398009e4d48.jpg) +Figure 1. We propose MammalNet, a large-scale video benchmark for recognizing mammals and their behavior. It is built around a biological mammal taxonomy spanning 17 orders, 69 families and 173 mammal categories, and includes 12 common high-level mammal behaviors (e.g., hunt, groom). MammalNet enables the study of animal and behavior recognition, both separately and jointly. It also facilitates investigating challenging compositional scenarios which test models' zero- and low-shot transfer abilities. Moreover, MammalNet includes behavior detection by localizing when a behavior occurs in an untrimmed video. Our dataset is the first to enable animal behavior analysis at scale in an ecologically-grounded manner, and exemplifies multiple challenges for the computer vision community, such as recognition of imbalanced, hierarchical distributions of fine-grained categories and generalization to unseen or seldom seen scenarios. + +# Abstract + +Monitoring animal behavior can facilitate conservation efforts by providing key insights into wildlife health, population status, and ecosystem function. Automatic recognition of animals and their behaviors is critical for capitalizing on the large unlabeled datasets generated by modern video devices and for accelerating monitoring efforts at scale. However, the development of automated recognition systems is currently hindered by a lack of appropriately labeled datasets. Existing video datasets 1) do not classify animals according to established biological taxonomies; 2) are too small to facilitate large-scale behavioral studies and are often limited to a single species; and 3) do not feature temporally local- + +ized annotations and therefore do not facilitate localization of targeted behaviors within longer video sequences. Thus, we propose MammalNet, a new large-scale animal behavior dataset with taxonomy-guided annotations of mammals and their common behaviors. MammalNet contains over 18K videos totaling 539 hours, which is $\sim 10$ times larger than the largest existing animal behavior dataset [36]. It covers 17 orders, 69 families, and 173 mammal categories for animal categorization and captures 12 high-level animal behaviors that received focus in previous animal behavior studies. We establish three benchmarks on MammalNet: standard animal and behavior recognition, compositional low-shot animal and behavior recognition, and behavior detection. Our dataset and code have been made available at: https://mammal-net.github.io. + +# 1. Introduction + +Animal species are a core component of the world's ecosystems. Through their behavior, animals drive diverse ecological processes, including seed dispersal, nutrient cycling, population dynamics, speciation, and extinction. Thus, understanding and monitoring the behaviors of animals and their interactions with their physical and social environments is key to understanding the complexities of the world's ecosystems, an objective that is especially critical now given the ongoing biodiversity crisis [12]. + +Modern sensors, including camera traps, drones, and smartphones, allow wildlife researchers, managers, and citizen scientists to collect video data of animal behavior on an unprecedented scale [43]. However, processing this data to generate actionable, timely insights remains a major challenge. Manual human review and annotation of footage to identify and locate species and behavioral sequences of interest is time-intensive and does not scale to large datasets. Thus, methods for automated animal and behavioral recognition could open the door to large-scale behavioral monitoring and speed up the time to produce usable data, thereby reducing the time to implement management directives. + +The first essential step to creating such an AI system for animal and behavior recognition is curating a diverse, representative dataset that allows us to formalize these challenges as computer vision tasks and benchmark potential solutions. Most previous datasets either only cover a limited number of animal and behavior types [4, 38], or do not implement animal labeling [36], or include a small number of videos with insufficient environmental diversity [4, 38, 48]. Recently, a dataset named "Animal Kingdom" [36] was proposed to study animal actions and is currently the largest existing behavioral dataset, to the best of our knowledge. However, it only contains 4,310 videos totaling 50 hours, which might be insufficient for large-scale animal behavior studies considering its diversity. Furthermore, the authors only focus on the recognition of atomic actions such as yawning, swimming, and flying. These basic actions cannot be easily matched to the higher-order behavioral states that are of primary interest to end users in animal management and conservation [6]. For example, a cheetah that is running may either be hunting, escaping, or playing. Finally, and most importantly, they do not support some important tasks such as animal recognition and behavior detection which are essential for animal behavior understanding. + +To overcome the limitations of previous datasets, we propose a new dataset called MammalNet. We specifically focus on mammals since they, unlike other animal classes such as birds or insects, usually have more diverse and distinguishable behavior statuses. MammalNet is comprised of 539 hours of annotated videos, which is $\sim 10$ times longer than that of the largest available animal behavior dataset. It contains 18,346 videos depicting 12 fundamental high-level + +behaviors from hundreds of mammal species. Importantly, it focuses on 12 higher-order animal behaviors that are the focus of previous animal behavior literature [3,8,17,33], rather than atomic actions. MammalNet also categorizes animals according to the scientific taxonomy available in Wikipedia, as we show in Fig. 1; hence the dataset can be flexibly expanded in the future by following the same protocols. It includes videos of approximately 800 mammal species in 173 mammal categories. We establish three benchmarks inspired by ecological research needs - standard animal & behavior classification, compositional low-shot animal & behavior recognition, and behavior detection - to promote future study in animal behavior understanding. + +Through our experiments, we find that: (1) Correctly recognizing the animals and behaviors is a challenging task even for the state-of-the-art models, especially for less-frequent animals. The top-1 per-class accuracy is 32.5 for animal recognition, 37.8 for behavior recognition, and 17.8 for their joint recognition in our best-performing model. (2) Behavior recognition for unseen animals can be transferred from observations of other seen animals due to their similar features such as appearance and movement style, which can help in studies of animals with less available data. However, to achieve more accurate behavior recognition, having access to videos of the target animals and behaviors is still crucial. + +# 2. Related Work + +Automatic animal recognition and behavior detection can help humans monitor and efficiently process animal behavior data [7, 22, 32, 34, 40, 47, 48]. It can massively reduce the labour cost from manually collecting and analyzing animal activities. During the past few years, many datasets [18, 20, 29, 38] have been introduced to develop foundations for animal behavior research. We systematically analyze the previous datasets and summarize several important limitations that prevent them from being used for large-scale animal recognition and behavior understanding: + +Lack of behavior understanding. Many previous datasets only focus on animal recognition or pose estimation from images, but lack behavior learning. For example, iNaturalist [45] collects 859,000 images covering more than 5,000 different types of plants and animals. NABird [44] collects 48,562 images of 555 different North American bird species. Also, some works focus on narrow animal recognition such as dogs [27], birds [46, 50] or cats [13]. On the other hand, many works also focus on pose estimation [10, 18, 21, 31, 41] and animal face detection [26, 39, 51]. They are generally not applicable to learning animal behavior. + +Lack of taxonomic diversity for different behaviors. Previous animal behavior datasets have minimal taxonomic coverage - often containing just a single animal species. For example, there are existing behavior recognition datasets for elephants [28], sheep [37], monkeys [5], tigers [16], + +
DatasetsDataset PropertiesTasks
Publicly Available?Taxonomy-guided Animal Annotation?No. of VideosNo. of ActionsNo. of BehaviorsNo. of Animal CategoriesNo. of Mammal CategoriesTotal DurationAnimal ClassificationAction/Behavior RecognitionAction/Behavior Detection
Wild Felines [20]××2,7003-33-✓✓×
Wildlife Actions [29]××10,6007-3211-✓✓×
Animal Kingdom [36]✓×4,301140-850-50 (h)×✓×
MammalNet (ours)✓✓18,346-12173173539 (h)✓✓✓
+ +Table 1. The comparison among existing animal behavior understanding video datasets. Compared to other datasets, MammalNet annotates the animals by following the scientific mammal taxonomy, focuses on more high-level behavior recognition, has the largest number of mammal categories, collects the largest number of animal behavior videos, totalling 539 hours, and also enables behavior detection tasks. + +etc. While useful for the studied species, these datasets do not enable the exploration of behavior across species which is necessary to scale up behavior identification without requiring training examples of every possible combination of species and behavior. + +Lack of taxonomy-guided animal annotation. Previous animal behavior datasets either do not include animal recognition as a task [36] or group species according to subjective as opposed to scientific criteria [29, 36]. In contrast, our dataset collects and annotates the videos according to the scientific mammal taxonomy. By following this taxonomy we allow exploration of behavior from an evolutionary perspective, as well as enable standardized and consistent dataset expansion under the same protocol. + +Actions vs. Behaviors. Previous works mostly focus on atomic action recognition [20, 29, 36]. For example, some of their action classes are walk, stand still, fly, etc. In contrast, MammalNet focuses on higher-level animal behaviors, such as hunt, feed baby, etc. Behavior here denotes the main activity instance during a period, and it is usually composed of multiple atomic actions. These complex behaviors are needed to describe and summarize video activity bouts in a manner which is valuable for ecological research. + +We further compare our MammalNet dataset with the other existing animal behavior understanding video datasets, and summarize the key difference in Table 1. + +# 3. Constructing MammalNet + +The goal of MammalNet is to provide a large-scale mammal video dataset that benchmarks both animal and behavior recognition. In this section, we discuss our dataset construction protocol, including the choice of scientific animal taxonomy, crowdsourced annotations, and performing manual quality control during video collection and annotation. Finally, we describe the statistical profile of MammalNet. + +![](images/66fa26fa3a4efd7085a90d0782e5c8c1488de24cb201e6fd226870ca3aa9017f.jpg) +Figure 2. A subset of the mammal taxonomy of MammalNet. It includes 3 orders, 11 families, and 45 genera. + +# 3.1. Animal Taxonomy and Behavior Collection + +Animal taxonomy construction. Accurate animal recognition is one of the main goals defined by MammalNet. To ensure this task is scientifically relevant, we aimed to collect and structure our list of target animals based on the mammal taxonomy. First, we collected a diverse list of mammals, approximately 800 mammal species and sub-species, from the National Geographic [2] and Animal A-Z [1] websites. Next, we manually mapped each mammal onto the taxonomic structure (including class, order, family, genus, tribe, sub-family and species) in mammal taxonomy from Wikipedia. In total, the classes included in MammalNet cover 17 orders, 69 families, and 173 mammal categories. A taxonomic subset is visualized in Fig. 2. + +Behavior collection. We aim to enable the study of complex, high-level animal behavior as opposed to the simpler atomic actions emphasized in previous work [29, 36]. Be + +![](images/5aef39351aea3ce5681ae57da55728ed0a2162f32d14801b5c5ec3bddec217ad.jpg) +Figure 3. The examples for the annotated target behavior boundaries. The frames marked in red boxes denote the annotated temporal boundaries for the target behavior. + +havior here represents the major activity being displayed during a period of a video, and can be viewed as a series of atomic actions collectively serving a higher-level purpose. For example, hunting behavior, as defined in MammalNet, can often be decomposed into running, chasing and killing actions, etc. Identification of behaviors at this level of definition is more useful for ecologists and zoologists [3,8,17,33] compared to atomic actions. + +Inspired by previous biological and ecological studies [3, 8,17], we consider 12 fundamental mammal behaviors under 5 different groups in our study. They are respectively: + +Foraging behaviors: eat food, drink water, hunt. + +Reproductive behaviors: mate, feed baby, give birth. + +Hygiene behaviors: groom. + +Agnostic behaviours: fight. + +Maintenance behaviors: urinate, defecate, sleep, vomit. + +# 3.2. Video Collection and Quality Assurance + +Dataset curation is usually a costly process requiring a lot of manual annotation by humans. Some datasets are annotated by domain experts to encourage more reliable labeling, particularly for challenging tasks, but this expertise comes at a much higher cost and is thus hard to scale. We adapted a semi-automatic crowdsourcing approach to collect and annotate our datasets, inspired by many previous works [9, 15, 23, 49]. + +Online video retrieval. Our goal was to collect videos depicting each animal in our database performing each of the 12 focal behaviors. To achieve this, we queried YouTube with the text combinations of each animal common name and behavior, e.g., tiger hunting, and downloaded videos where the title included our queried animal name and behavior. However, some animal names are too rare to generate enough relevant videos, e.g., Brown hyena or Spotted hyena. In such cases, we used a more common name, hyena, to represent the union of those animals. In order to retrieve more relevant videos from the search engine, we also expanded each animal with their synonyms as given in Wikipedia. For example, we expanded the original artic fox with other equivalent names + +such as white fox, polar fox and snow fox. Each video was downloaded at its highest available resolution. + +Data filtering. During video retrieval, we downloaded the videos that are accessible for people $< 16$ years old to avoid the videos with violent content. We also prioritized videos that are shorter than 10 minutes in duration to limit the total storage. + +However, some videos might have irrelevant content due to the inaccuracy of text-based retrieval. For example, the downloaded videos might 1) display cartoon or toy animals or unrealistic environments such as games or movies, 2) display a static image instead of a continuous video, 3) involve a lot of human-animal interaction, e.g., human feeding an animal, as opposed to focusing on animal behavior, 4) not contain the specified animal and/or behavior. + +To alleviate these issues, we employed Amazon Mechanical Turk workers to verify the presence of the animal and behavior, and identify other quality issues. We assigned each video to three different workers and provided them with the pictures of an animal with its common name, and an expected behavior. We asked the workers to verify if this animal and behavior indeed appear in the video. Only the videos that "pass" by all the three workers were kept for the following behavior localization annotation. Before the workers started the verification, we first provided them the verification instructions and asked them to complete a corresponding qualification test with 20 multiple choice questions to ensure only qualified workers could participate in our task. + +# 3.3. Animal Behavior Localization Annotation + +The target behavior typically does not span the whole video. To localize the part where the target behavior is actually occurring in the video, we asked the AMT workers to manually annotate the respective temporal boundaries. Namely, we asked five different qualified workers to annotate the start and end frames for the target behavior in each video. In the end, each video received at least 5 annotated behavior boundaries. To achieve robust annotation agreement, we used the complete linkage algorithm [14] to cluster different temporal boundaries and merge them into one or + +![](images/a4bb2c4941349bf4f0f334fbb6c549cf4212dad42b795822388b36c9662a4fe8.jpg) +Figure 4. Number of videos per each mammal category. We rank the categories according to their trimmed videos frequency. + +![](images/15a0bc1da8703c7aefab36a54a8d9d1ad046bac34ae89d84951b0d05775123ef.jpg) +Figure 5. Number of videos per each behavior. We rank the behavior according to their trimmed videos frequency. + +several more stable ones that received multiple agreements. Note, that a single video might have multiple discrete occurrences of a given behavior, and thus have multiple boundary definitions. We show several examples for annotated target behavior boundaries in Fig. 3. + +# 3.4. Recognition at lowest feasible taxonomic level + +We categorize animals according to the lowest and feasible taxonomic classification, rather than the species level, for the following reasons: 1) YouTube videos contain the ambiguous and inexpert species labels. 2) We are not using expert annotators, and even experts often cannot reliably identify animals at the species or even genus level based on crowdsourced or field images [25]. Thus, it is more practical to classify animals in our dataset to the lowest feasible taxonomic level rather than species. As a result, our taxonomy contains 173 distinguishable taxonomic classification levels including: sub-family, tribe and genus. + +# 3.5. MammalNet Statistics + +The final MammalNet dataset contains 18,346 videos (539 hours); after we trim the videos according to the anno + +tated behavior boundaries, it increases to 20,033 trimmed video instances (394 hours). In total, it covers 173 mammal categories, 69 families and 17 orders in our dataset, with the total of 173 mammal categories defined for our recognition tasks. MammalNet also contains 12 different common behaviors. The average duration for the untrimmed and trimmed videos is 106 and 77 seconds, respectively. Over $54\%$ of videos reach HD resolutions ( $1280 \times 720$ ). We demonstrate the data distribution in terms of each animal category and behavior in Fig. 4 and 5. We observe that the number of collected videos per type follows a long-tail distribution. + +# 4. Experimental Results + +We construct three main tasks on MammalNet: 1) Standard animal and behavior classification on trimmed videos, 2) Compositional low-shot animal and behavior recognition on trimmed videos, and 3) Behavior detection on untrimmed videos. In both the classification and detection tasks, we baseline several state-of-the-art models [19, 30, 53] that have been successfully applied to human action recognition and detection. In the following, we describe the formulation of each challenge and provide baselines and analysis. + +# 4.1. Standard Animal and Behavior Classification on Trimmed Videos + +This task explores classification of both the primary behavior that occurs in a trimmed video and the animal that performs the behavior. We report the top-1 per-example and per-class accuracy for all the baseline models. + +Many, medium, few splits. To capture the effect of the long-tailed nature of the MammalNet dataset, we group the animal, behavior, and their composition classes into many, medium, and few based on their frequency, and report the average + +
BaselinesAnimal ClassificationBehavior ClassificationJoint Classification
Many 12Medium 28Few 133All 173Many 4Medium 4Few 4All 12Many 33Medium 180Few 823All 1036
SlowFast [19]49.635.69.517.239.027.614.927.227.219.44.48.6
C3D [42]48.635.310.017.538.227.811.725.928.118.84.58.6
I3D [11]48.834.910.517.839.527.214.827.229.620.54.48.9
MViT V2 [30]48.535.512.919.742.329.211.627.729.519.64.89.0
SlowFast*58.343.116.624.545.132.714.830.938.623.57.312.1
C3D*58.345.419.126.844.636.015.932.238.026.48.413.5
I3D*58.642.916.924.846.335.014.832.138.324.58.613.3
MViT V2*66.756.023.432.550.942.420.037.846.233.011.817.8
+ +Table 2. Per-class Top-1 accuracy for animal, behavior and their joint prediction.* denotes the initialization from the model pretrained on Kinetics 400 [24]. Transfer learning from the human action to the animal behavior recognition receives considerable performance gain. Best performance for each split has been highlighted in bold. + +
BaselinesAnimalBehaviorJoint
SlowFast [19]35.434.217.4
C3D [42]35.033.517.1
I3D [11]35.234.317.9
MViT V2 [30]35.636.818.0
SlowFast*43.039.422.8
C3D*44.440.324.6
I3D*43.441.224.0
MViT V2*52.646.630.6
+ +Table 3. Per-example accuracy for animal, behavior and their joint prediction. * denotes the initialization from the pretrained model. + +per-class accuracy bands chosen based on the frequency percentiles. For animal categories this is broken down as many: top $7\%$ frequent classes, medium: middle $16\%$ classes, and few: the remaining $77\%$ classes. For behavior, many: top $33\%$ frequent classes, medium: middle $33\%$ , and few: the remaining $33\%$ classes. For joint classification, many: top $3\%$ frequent classes, medium: middle $17\%$ classes, few: the remaining $80\%$ classes. We show the number of classes per each split in Table 2. + +Dataset setup. We randomly split the examples from each animal-behavior category into $70\%$ for training, $10\%$ for validation, and $20\%$ for testing, and it results in 14,554 training, 1,638 validation, and 3,841 testing videos, respectively. + +Baselines. We compare SlowFast [19], I3D [11], C3D [42] and MViT V2 [30] models on our tasks. These models are evaluated in two versions: 1) Training with random initialization 2) Initializing with weights from a model pretrained on Kinetics 400 [24]. These methods were originally designed for human action recognition and hence do not have an ability to predict both an action and a subject by default. To accommodate them into our joint prediction setting, we have two task heads, one for animal category recognition and one for behavior recognition. We compute the joint + +loss, $\mathcal{L}_{joint}$ , for both animal and behavior classification as shown in Fig. 1. We tune all the hyper-parameters on the validation data. The final hyper-parameters for each model are provided in the supplement. The loss is defined as: + +$$ +\mathcal {L} _ {\text {j o i n t}} = - \frac {1}{M} \sum_ {i} ^ {M} y _ {i} ^ {a} \log \left(p _ {i} ^ {a}\right) - \frac {1}{N} \sum_ {j} ^ {N} y _ {j} ^ {b} \log \left(p _ {j} ^ {b}\right) \tag {1} +$$ + +where $M$ is the number of animal classes, and $N$ is the number of behavior classes, $p_i^a$ and $y_{i}^{a}$ denote the animal prediction probability and ground truth label for the category $i$ , $p_y^b$ and $y_j^b$ denote the behavior prediction probability and ground truth label for the category $j$ . + +Experimental results. The results for per-class and per-example classification are summarized in Tables 2 and 3, respectively. We find that MViT v2 is able to achieve competitive results for all the splits. The best top-1 joint per-class accuracy is 17.8 and per-example accuracy is 30.6, which points to significant room for improvement on these challenging tasks. We also observe that transfer learning from the model that is pretrained on Kinetics 400, a human action recognition dataset, improves both the animal and behavior classification accuracy (with MViT v2, this corresponds to a per-class accuracy gain from 19.7 to 32.5 for animal classification, and from 27.7 to 37.8 for behavior classification). Additionally, we find that performance gain from pretraining is higher for frequently occurring animals and behaviors, indicated by the results shown in many, medium, and few splits. Accurately predicting low-frequency animal and behavior categories remains a significant challenge. + +# 4.2. Compositional Low-shot Animal and Behavior Classification on Trimmed Videos + +It is hard to find a sufficient number of labeled behavior samples for all the animals in our collected taxonomy. For example, numbat and florida panther have very few behavior + +
BaselinesCompositional Low-Shot Behavior Classification
0-shot1-shot5-shot
Per-example A / BPer-class A / BPer-example A / BPer-class A / BPer-example A / BPer-class A / B
C3D*27.4 / 23.718.3 / 16.129.2 / 25.521.1 / 19.933.3 / 29.325.2 / 23.0
I3D*25.3 / 23.916.7 / 15.226.8 / 25.716.9 / 18.330.5 / 28.321.7 / 21.9
SlowFast*26.2 / 24.817.9 / 16.326.5 / 26.316.7 / 19.029.6 / 29.222.5 / 22.4
MViT V2*32.2 / 26.220.7 / 18.133.7 / 28.922.9 / 22.539.3 / 31.731.0 / 26.0
+ +Table 4. Compositional low-shot animal and behavior recognition. * denotes the initialization from the model pretrained on Kinetics 400 [24]. "A" denotes the animal category and "B" denotes the behavior category. The best performance per each column has been highlighted in bold. + +
BaselinesmAP
0.500.600.700.800.90Avg.
CoLA [52]26.0222.7018.9813.463.0515.81
TAGS [35]23.0920.9719.0916.9812.5617.63
ActionFormer [53]28.4826.1423.1718.6910.4820.07
+ +Table 5. The results for behavior detection. We report mAP at the IoU thresholds of [0.5:0.1:0.9]. Average mAP is computed by averaging different tIoU thresholds. + +annotations. However, it is plausible to imagine that behaviors can be transferred among different animals that have similar appearances and movement styles. Also the animal recognition under different behaviors (hunting, fighting) can be mutually transferred. To investigate these phenomena, we design the compositional low-shot animal and behavior classification task. + +Dataset setup. We first select the animal-behavior compositional classes that contain more than 5 examples and allocate $25\%$ of classes to the test set (4,088 videos). For the remaining classes (the other $75\%$ of classes and those classes with $\leq 5$ examples), we randomly designate $90\%$ of classes as the training set and $10\%$ as the validation set (898 videos). Under the low-shot scheme, for each compositional class, we randomly sample 5 examples from the test set and move 0, 1, or 5 of them into the training set for the zero-shot, 1-shot, and 5-shot setup, respectively (14,377, 14,511 and 15,047 training videos). The train, val and test sets consist of 983, 53, and 134 compositional classes, respectively. + +Baselines. We evaluate the compositional low-shot classification with the SlowFast [19], I3D [11], C3D [42] and MViT V2 [30] models under the joint loss. We initialize these models with the weights pretrained on Kinetics 400 [24]. + +Experimental results. We summarize the results in Table 4. It shows that MViT v2 consistently achieves the best performance under our low-shot setup. The behavior classification can still achieve 26.2 per-example and 18.1 top-1 accuracy under the zero-shot setting for MViT v2 model. This indicates that behavior classification is transferable from other animals in some cases. Additionally, we observe that its performance is improved to 31.7 per-example and 26.0 per-class + +top-1 accuracy under 5-shot setup, indicating that training on more videos with the behaviors from the same animal is still necessary. A similar phenomenon is observed to the few-shot animal recognition. + +# 4.3. Behavior Detection on Untrimmed Videos + +This task is to detect the behavior in untrimmed videos. The behavior detection algorithm should correctly detect the temporal range for the primary activity presented in the video. We follow previous temporal action localization works [9, 53] to benchmark this task. We report the mean Average Precision (mAP) with different temporal intersections over the union (tIoU) thresholds [0.5:0.1:0.9] as our evaluation metric. We also report the average mAP averaging across different tIoUs. + +Dataset setup. Similar to the previous standard animal and behavior classification, we follow the [train:0.7, val:0.1, test:0.2] ratios to split the untrimmed videos at the animal-behavior composition level. This results in 13,318 videos for training, 1,486 for validation and 3,542 for testing. We tune all the hyper-parameters based on the validation set. + +Baselines. We evaluate our datasets with the baseline models such as ActionFormer [53], TAGS [35], and CoLA [52]. To produce the features for our MammalNet videos, we first finetune a two-stream I3D [11] model, that is originally pretrained on ImageNet [15] and Kinetics 400 [24], on our dataset, and then extract the RGB and optical flow features for each video. We concatenate these two features together as the model input. + +Experimental results. We show the behavior detection results in Table 5. Among all the baselines, ActionFormer demonstrates the most competitive performance with an average mAP of 20.07, and also achieves $28.48\mathrm{mAP}$ for the threshold of 0.5. Overall, it is clear to see that the behavior detection task is still very challenging for current methods. + +# 5. Analysis + +Demonstration of animal and behavior classification. We sample some prediction examples from MViT v2 [30] for + +![](images/59169cb89f12e11b23d0d87d7c597a12fa12cf2c0e10398c95f4d355b758d3d9.jpg) +Figure 6. The visualization presents various instances of joint animal and behavior classification. In the third column, accurate predictions are displayed, while the fourth, fifth, and sixth columns showcase mispredicted examples where either the animal or the behavior does not correspond to the correct prediction. + +![](images/bf705a7d59d382b00b659bb076011a3b7401918cca6c1d06027fd70afbd7c741.jpg) +Figure 7. The visualization for behavior detection examples. + +the joint categories Acinonyx hunt, Canis eat and Antilopini fight. We demonstrate one correct prediction and three mispredicted examples where the model mis-recognizes the animal or behavior or both of them in Fig. 6. + +Demonstration of behavior detection: We visualize behavior detection results in Fig. 7 from ActionFormer [53]. The top example shows correctly predicted behavior with a proposal closely aligned with the ground-truth. The bottom example shows a misclassified behavior, and it mistakenly predicts the hunt behavior as fight. + +Joint animal and behavior recognition vs. separate recognition. We also train the model for recognizing the animal and behavior separately and provide the results in the Table 1 of the supplementary. Comparing with the joint recognition, we find that training a system to recognize the animal and behavior together can improve the behavior recognition under separate training by $\sim 2.2$ per-class accuracy. This indicates that being capable of understanding the animal types can benefit behavior prediction. However, the results also indicate that predicting animal only can be more useful in recognizing animal types. + +Dataset bias. Our dataset is downloaded from YouTube channels with diverse backgrounds and video quality, and it might differ from video datasets collected for behavioral or ecological studies. The videos on YouTube might also exhibit potential biases: + +1) Over-representation of some behaviors and under + +representation of others. For example, people might prefer watching videos with fight or eat food activities, and these two behaviors are more likely to be over-represented, while urinate and defecate behaviors are less interesting to humans and hence are underrepresented on YouTube. However, they are still important in informing conservation-related actions to protect the environment. + +2) Bias towards captive animals or wild animals that are habituated to humans. We find that many videos are shot at zoos, farms, and homes, etc. These animals may display different behaviors than wild or non-habituated animals. + +# 6. Conclusion + +We introduced MammalNet, a large-scale video dataset for mammal recognition and behavior understanding. We have collected videos for hundreds of different mammals and structured them by following the scientific mammal taxonomy. MammalNet consists of 18,346 untrimmed videos covering 173 mammal categories and 12 common behaviors. We established three challenges: standard animal and behavior classification, compositional low-shot animal and behavior classification, and behavior detection. Through our experiments, we found that the accurate recognition of animals at scale and their common behaviors is very challenging even with current state-of-the-art models, especially when the dataset has a long-tail distribution. We also found that learning to recognize the behavior of unseen animals is possible via transfer from the other seen animals. To promote further research and development in the field of animal behavior study, we have open-sourced all of our data and code to the research community. + +Acknowledgement: This work is supported by KAUST BAS/1/1685-01-01 and KAUST FCC/1/1973-58-01 (Red Sea Research Center), DARPA's SemaFor and PTG programs, the Caltech Resnick Sustainability Institute, and Germany's Excellence Strategy–‘Centre for the Advanced Study of Collective Behaviour’ EXC 2117-422037984. + +# References + +[1] A-z animals. https://a-z-animals.com/.3 +[2] National geographics. https://www.nationalgeographic.com/.3 +[3] John Alcock. Animal behavior: An evolutionary approach. Sinauer Associates, 2009. 2, 4 +[4] David J Anderson and Pietro Perona. Toward a science of computational ethology. Neuron, 84(1):18-31, 2014. 2 +[5] Praneet C Bala, Benjamin R Eisenreich, Seng Bum Michael Yoo, Benjamin Y Hayden, Hyun Soo Park, and Jan Zimmermann. Automated markerless pose estimation in freely moving macaques with openmonkeystudio. Nature communications, 11(1):1-12, 2020. 2 +[6] Oded Berger-Tal, Daniel T Blumstein, Scott Carroll, Robert N Fisher, Sarah L Mesnick, Megan A Owen, David Saltz, Colleen Cassady St Claire, and Ronald R Swaisgood. A systematic survey of the integration of animal behavior into conservation. Conservation Biology, 30(4):744-753, 2016. 2 +[7] Cigdem Bayan and Robert B Fisher. Detection of abnormal fish trajectories using a clustering based hierarchical classifier. In BMVC, 2013. 2 +[8] Michael D Breed and Janice Moore. Animal behavior. Academic Press, 2021. 2, 4 +[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 4, 7 +[10] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9498-9507, 2019. 2 +[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6, 7 +[12] Gerardo Ceballos, Paul R Ehrlich, and Peter H Raven. Vertebrates on the brink as indicators of biological annihilation and the sixth mass extinction. Proceedings of the National Academy of Sciences, 117(24):13596-13602, 2020. 2 +[13] Yu-Chen Chen, Shintami C Hidayati, Wen-Huang Cheng, Min-Chun Hu, and Kai-Lung Hua. Locality constrained sparse representation for cat recognition. In International Conference on Multimedia Modeling, pages 140-151. Springer, 2016. 2 +[14] Daniel Defays. An efficient algorithm for a complete link method. The Computer Journal, 20(4):364-366, 1977. 4 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 4, 7 +[16] Rod K Dishman, Andrew S Jackson, and Molly S Bray. Self-regulation of exercise behavior in the tiger study. Annals of Behavioral Medicine, 48(1):80-91, 2014. 2 +[17] Lee Alan Dugatkin. Principles of animal behavior. University of Chicago Press, 2020. 2, 4 + +[18] Cheng Fang, Tiemin Zhang, Haikun Zheng, Junduan Huang, and Kaixuan Cuan. Pose estimation and behavior classification of broiler chickens based on deep neural networks. Computers and Electronics in Agriculture, 180:105863, 2021. 2 +[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5, 6, 7 +[20] Liqi Feng, Yaqin Zhao, Yichao Sun, Wenxuan Zhao, and Jiaxi Tang. Action recognition using a spatial-temporal network for wild felines. Animals, 11(2):485, 2021. 2, 3 +[21] Jacob M Graving, Daniel Chae, Hemal Naik, Liang Li, Benjamin Koger, Blair R Costelloe, and Iain D Couzin. Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. *Elife*, 8:e47994, 2019. 2 +[22] Shilpi Gupta, Prerana Mukherjee, Santanu Chaudhury, Brejesh Lall, and Hemanth Sanisetty. Dftnet: Deep fish tracker with attention mechanism in unconstrained marine environments. IEEE Transactions on Instrumentation and Measurement, 70:1-13, 2021. 2 +[23] Fabian Caba Heilbron and Juan Carlos Niebles. Collecting and annotating human activities in web videos. In Proceedings of International Conference on Multimedia Retrieval, pages 377-384, 2014. 4 +[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6, 7 +[25] Roland Kays, Monica Lasky, Maximilian L Allen, Robert C Dowler, Melissa TR Hawkins, Andrew G Hope, Brooks A Kohli, Verity L Mathis, Bryan McLean, Link E Olson, et al. Which mammals can be identified from camera traps and crowdsourced photographs? Journal of Mammalogy, 2022. 5 +[26] Muhammad Haris Khan, John McDonagh, Salman Khan, Muhammad Shahabuddin, Aditya Arora, Fahad Shahbaz Khan, Ling Shao, and Georgios Tzimiropoulos. Animalweb: A large-scale hierarchical dataset of annotated animal faces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6939-6948, 2020. 2 +[27] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2. Cite-seer, 2011. 2 +[28] Nicole Laws, Andre Ganswindt, Michael Heistermann, Moira Harris, Stephen Harris, and Chris Sherwin. A case study: fecal corticosteroid and behavior as indicators of welfare during relocation of an asian elephant. Journal of Applied Animal Welfare Science, 10(4):349-358, 2007. 2 +[29] Weining Li, Sirmam Swetha, and Mubarak Shah. Wildlife action recognition using deep learning. 2020. 2, 3 +[30] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvittv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 5, 6, 7 +[31] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859-1868, 2021. 2 +[32] Mackenzie Weygandt Mathis and Alexander Mathis. Deep learning tools for the measurement of animal behavior in neuroscience. Current opinion in neurobiology, 60:1-11, 2020. 2 +[33] Joy Mench. Why it is important to understand animal behavior. ILAR journal, 39(1):20-26, 1998. 2, 4 +[34] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386-12395, 2020. 2 +[35] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Temporal action detection with global segmentation mask learning. European Conference on Computer Vision, 2022. 7 +[36] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19023-19034, 2022. 1, 2, 3 +[37] Kehinde Owoeye and Stephen Hailes. Online collective animal movement activity recognition. arXiv preprint arXiv:1811.09067, 2018. 2 +[38] Shah Atiqur Rahman, Insu Song, Maylor KH Leung, Ickjai Lee, and Kyungmi Lee. Fast action recognition using negative space features. Expert Systems with Applications, 41(2):574-587, 2014. 2 +[39] Maheen Rashid, Xiuye Gu, and Yong Jae Lee. Interspecies knowledge transfer for facial keypoint detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6894-6903, 2017. 2 +[40] Lianne Robinson and Gernot Riedel. Comparison of automated home-cage monitoring systems: emphasis on feeding behaviour, activity and spatial learning following pharmacological interventions. Journal of neuroscience methods, 234:13-25, 2014. 2 +[41] Moira Shooter, Charles Malleson, and Adrian Hilton. Sydog: A synthetic dog dataset for improved 2d pose estimation. arXiv preprint arXiv:2108.00249, 2021. 2 +[42] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 6, 7 +[43] Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation. Nature communications, 13(1):1-15, 2022. 2 +[44] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset + +with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 595-604, 2015. 2 +[45] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2 +[46] Grant Van Horn, Rui Qian, Kimberly Wilber, Hartwig Adam, Oisin Mac Aodha, and Serge Belongie. Exploring fine-grained audiovisual categorization with the ssw60 dataset. 2 +[47] Grant Van Horn, Rui Qian, Kimberly Wilber, Hartwig Adam, Oisin Mac Aodha, and Serge Belongie. Exploring fine-grained audiovisual categorization with the ssw60 dataset, 2022. 2 +[48] Lukas von Ziegler, Oliver Sturman, and Johannes Bohacek. Big behavior: challenges and opportunities in a new era of deep behavior profiling. Neuropsychopharmacology, 46(1):33-44, 2021. 2 +[49] Carl Vondrick, Donald Patterson, and Deva Ramanan. Efficiently scaling up crowdsourced video annotation. International journal of computer vision, 101(1):184-204, 2013. 4 +[50] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2 +[51] Heng Yang, Renqiao Zhang, and Peter Robinson. Human and sheep facial landmarks localisation by triplet interpolated features. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1-8. IEEE, 2016. 2 +[52] Can Zhang, Meng Cao, Dongming Yang, Jie Chen, and Yuexian Zou. Cola: Weakly-supervised temporal action localization with snippet contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16010-16019, June 2021. 7 +[53] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, 2022. 5, 7, 8 \ No newline at end of file diff --git a/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/images.zip b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ddf7b0e0d5deb3bc7e13de2848b2a6959334c02 --- /dev/null +++ b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c44d31881eacbc0f8e1d4229efd30a1269b3143e359d8f5971140093d5cd57e +size 710989 diff --git a/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/layout.json b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..307a428ca393f220afdf2f801c5b1e1937ae800d --- /dev/null +++ b/2023/MammalNet_ A Large-Scale Video Benchmark for Mammal Recognition and Behavior Understanding/layout.json @@ -0,0 +1,7040 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 467, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 467, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 467, + 140 + ], + "type": "text", + "content": "MammalNet: A Large-scale Video Benchmark for Mammal Recognition and Behavior Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "spans": [ + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": "Jun Chen" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": " Ming Hu" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": " Darren J. Coker" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": " Michael L. Berumen" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": " \nBlair Costello" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": " Sara Beery" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": " Anna Rohrbach" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": " Mohamed Elhoseiny" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": "King Abdullah University of Science and Technology (KAUST) \n" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": "Max Planck Institute of Animal Behavior, " + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": "University of Konstanz \n" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": "Massachusetts Institute of Technology, " + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 92, + 160, + 500, + 232 + ], + "type": "text", + "content": "University of California, Berkeley" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 97, + 253, + 495, + 423 + ], + "blocks": [ + { + "bbox": [ + 97, + 253, + 495, + 423 + ], + "lines": [ + { + "bbox": [ + 97, + 253, + 495, + 423 + ], + "spans": [ + { + "bbox": [ + 97, + 253, + 495, + 423 + ], + "type": "image", + "image_path": "c6fdda67a752d39e5f70ee8b6fc5f275b719529b94fbef76aea0a398009e4d48.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 426, + 545, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 426, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 545, + 504 + ], + "type": "text", + "content": "Figure 1. We propose MammalNet, a large-scale video benchmark for recognizing mammals and their behavior. It is built around a biological mammal taxonomy spanning 17 orders, 69 families and 173 mammal categories, and includes 12 common high-level mammal behaviors (e.g., hunt, groom). MammalNet enables the study of animal and behavior recognition, both separately and jointly. It also facilitates investigating challenging compositional scenarios which test models' zero- and low-shot transfer abilities. Moreover, MammalNet includes behavior detection by localizing when a behavior occurs in an untrimmed video. Our dataset is the first to enable animal behavior analysis at scale in an ecologically-grounded manner, and exemplifies multiple challenges for the computer vision community, such as recognition of imbalanced, hierarchical distributions of fine-grained categories and generalization to unseen or seldom seen scenarios." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 517, + 192, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 517, + 192, + 529 + ], + "spans": [ + { + "bbox": [ + 143, + 517, + 192, + 529 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "type": "text", + "content": "Monitoring animal behavior can facilitate conservation efforts by providing key insights into wildlife health, population status, and ecosystem function. Automatic recognition of animals and their behaviors is critical for capitalizing on the large unlabeled datasets generated by modern video devices and for accelerating monitoring efforts at scale. However, the development of automated recognition systems is currently hindered by a lack of appropriately labeled datasets. Existing video datasets 1) do not classify animals according to established biological taxonomies; 2) are too small to facilitate large-scale behavioral studies and are often limited to a single species; and 3) do not feature temporally local-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 517, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 547, + 697 + ], + "type": "text", + "content": "ized annotations and therefore do not facilitate localization of targeted behaviors within longer video sequences. Thus, we propose MammalNet, a new large-scale animal behavior dataset with taxonomy-guided annotations of mammals and their common behaviors. MammalNet contains over 18K videos totaling 539 hours, which is " + }, + { + "bbox": [ + 304, + 517, + 547, + 697 + ], + "type": "inline_equation", + "content": "\\sim 10" + }, + { + "bbox": [ + 304, + 517, + 547, + 697 + ], + "type": "text", + "content": " times larger than the largest existing animal behavior dataset [36]. It covers 17 orders, 69 families, and 173 mammal categories for animal categorization and captures 12 high-level animal behaviors that received focus in previous animal behavior studies. We establish three benchmarks on MammalNet: standard animal and behavior recognition, compositional low-shot animal and behavior recognition, and behavior detection. Our dataset and code have been made available at: https://mammal-net.github.io." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13052" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 287, + 198 + ], + "type": "text", + "content": "Animal species are a core component of the world's ecosystems. Through their behavior, animals drive diverse ecological processes, including seed dispersal, nutrient cycling, population dynamics, speciation, and extinction. Thus, understanding and monitoring the behaviors of animals and their interactions with their physical and social environments is key to understanding the complexities of the world's ecosystems, an objective that is especially critical now given the ongoing biodiversity crisis [12]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 200, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 200, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 200, + 287, + 342 + ], + "type": "text", + "content": "Modern sensors, including camera traps, drones, and smartphones, allow wildlife researchers, managers, and citizen scientists to collect video data of animal behavior on an unprecedented scale [43]. However, processing this data to generate actionable, timely insights remains a major challenge. Manual human review and annotation of footage to identify and locate species and behavioral sequences of interest is time-intensive and does not scale to large datasets. Thus, methods for automated animal and behavioral recognition could open the door to large-scale behavioral monitoring and speed up the time to produce usable data, thereby reducing the time to implement management directives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 343, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 287, + 617 + ], + "type": "text", + "content": "The first essential step to creating such an AI system for animal and behavior recognition is curating a diverse, representative dataset that allows us to formalize these challenges as computer vision tasks and benchmark potential solutions. Most previous datasets either only cover a limited number of animal and behavior types [4, 38], or do not implement animal labeling [36], or include a small number of videos with insufficient environmental diversity [4, 38, 48]. Recently, a dataset named \"Animal Kingdom\" [36] was proposed to study animal actions and is currently the largest existing behavioral dataset, to the best of our knowledge. However, it only contains 4,310 videos totaling 50 hours, which might be insufficient for large-scale animal behavior studies considering its diversity. Furthermore, the authors only focus on the recognition of atomic actions such as yawning, swimming, and flying. These basic actions cannot be easily matched to the higher-order behavioral states that are of primary interest to end users in animal management and conservation [6]. For example, a cheetah that is running may either be hunting, escaping, or playing. Finally, and most importantly, they do not support some important tasks such as animal recognition and behavior detection which are essential for animal behavior understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 618, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 618, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 618, + 287, + 712 + ], + "type": "text", + "content": "To overcome the limitations of previous datasets, we propose a new dataset called MammalNet. We specifically focus on mammals since they, unlike other animal classes such as birds or insects, usually have more diverse and distinguishable behavior statuses. MammalNet is comprised of 539 hours of annotated videos, which is " + }, + { + "bbox": [ + 49, + 618, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\sim 10" + }, + { + "bbox": [ + 49, + 618, + 287, + 712 + ], + "type": "text", + "content": " times longer than that of the largest available animal behavior dataset. It contains 18,346 videos depicting 12 fundamental high-level" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 227 + ], + "type": "text", + "content": "behaviors from hundreds of mammal species. Importantly, it focuses on 12 higher-order animal behaviors that are the focus of previous animal behavior literature [3,8,17,33], rather than atomic actions. MammalNet also categorizes animals according to the scientific taxonomy available in Wikipedia, as we show in Fig. 1; hence the dataset can be flexibly expanded in the future by following the same protocols. It includes videos of approximately 800 mammal species in 173 mammal categories. We establish three benchmarks inspired by ecological research needs - standard animal & behavior classification, compositional low-shot animal & behavior recognition, and behavior detection - to promote future study in animal behavior understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 228, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 370 + ], + "type": "text", + "content": "Through our experiments, we find that: (1) Correctly recognizing the animals and behaviors is a challenging task even for the state-of-the-art models, especially for less-frequent animals. The top-1 per-class accuracy is 32.5 for animal recognition, 37.8 for behavior recognition, and 17.8 for their joint recognition in our best-performing model. (2) Behavior recognition for unseen animals can be transferred from observations of other seen animals due to their similar features such as appearance and movement style, which can help in studies of animals with less available data. However, to achieve more accurate behavior recognition, having access to videos of the target animals and behaviors is still crucial." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 382, + 391, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 391, + 394 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 391, + 394 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 402, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 402, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 402, + 545, + 521 + ], + "type": "text", + "content": "Automatic animal recognition and behavior detection can help humans monitor and efficiently process animal behavior data [7, 22, 32, 34, 40, 47, 48]. It can massively reduce the labour cost from manually collecting and analyzing animal activities. During the past few years, many datasets [18, 20, 29, 38] have been introduced to develop foundations for animal behavior research. We systematically analyze the previous datasets and summarize several important limitations that prevent them from being used for large-scale animal recognition and behavior understanding:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 522, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 653 + ], + "type": "text", + "content": "Lack of behavior understanding. Many previous datasets only focus on animal recognition or pose estimation from images, but lack behavior learning. For example, iNaturalist [45] collects 859,000 images covering more than 5,000 different types of plants and animals. NABird [44] collects 48,562 images of 555 different North American bird species. Also, some works focus on narrow animal recognition such as dogs [27], birds [46, 50] or cats [13]. On the other hand, many works also focus on pose estimation [10, 18, 21, 31, 41] and animal face detection [26, 39, 51]. They are generally not applicable to learning animal behavior." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 654, + 545, + 712 + ], + "type": "text", + "content": "Lack of taxonomic diversity for different behaviors. Previous animal behavior datasets have minimal taxonomic coverage - often containing just a single animal species. For example, there are existing behavior recognition datasets for elephants [28], sheep [37], monkeys [5], tigers [16]," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "13053" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 523, + 191 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 523, + 191 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 523, + 191 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 523, + 191 + ], + "type": "table", + "html": "
DatasetsDataset PropertiesTasks
Publicly Available?Taxonomy-guided Animal Annotation?No. of VideosNo. of ActionsNo. of BehaviorsNo. of Animal CategoriesNo. of Mammal CategoriesTotal DurationAnimal ClassificationAction/Behavior RecognitionAction/Behavior Detection
Wild Felines [20]××2,7003-33-✓✓×
Wildlife Actions [29]××10,6007-3211-✓✓×
Animal Kingdom [36]✓×4,301140-850-50 (h)×✓×
MammalNet (ours)✓✓18,346-12173173539 (h)✓✓✓
", + "image_path": "462dac20fe5ac399fc34a75cf67f1d22b80b3d763f6d01e52321e697f7e41a8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 209, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 243 + ], + "type": "text", + "content": "Table 1. The comparison among existing animal behavior understanding video datasets. Compared to other datasets, MammalNet annotates the animals by following the scientific mammal taxonomy, focuses on more high-level behavior recognition, has the largest number of mammal categories, collects the largest number of animal behavior videos, totalling 539 hours, and also enables behavior detection tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 250, + 288, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 288, + 309 + ], + "type": "text", + "content": "etc. While useful for the studied species, these datasets do not enable the exploration of behavior across species which is necessary to scale up behavior identification without requiring training examples of every possible combination of species and behavior." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 422 + ], + "type": "text", + "content": "Lack of taxonomy-guided animal annotation. Previous animal behavior datasets either do not include animal recognition as a task [36] or group species according to subjective as opposed to scientific criteria [29, 36]. In contrast, our dataset collects and annotates the videos according to the scientific mammal taxonomy. By following this taxonomy we allow exploration of behavior from an evolutionary perspective, as well as enable standardized and consistent dataset expansion under the same protocol." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 428, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 428, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 428, + 287, + 536 + ], + "type": "text", + "content": "Actions vs. Behaviors. Previous works mostly focus on atomic action recognition [20, 29, 36]. For example, some of their action classes are walk, stand still, fly, etc. In contrast, MammalNet focuses on higher-level animal behaviors, such as hunt, feed baby, etc. Behavior here denotes the main activity instance during a period, and it is usually composed of multiple atomic actions. These complex behaviors are needed to describe and summarize video activity bouts in a manner which is valuable for ecological research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 541, + 288, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 288, + 578 + ], + "type": "text", + "content": "We further compare our MammalNet dataset with the other existing animal behavior understanding video datasets, and summarize the key difference in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 604, + 197, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 197, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 197, + 618 + ], + "type": "text", + "content": "3. Constructing MammalNet" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "The goal of MammalNet is to provide a large-scale mammal video dataset that benchmarks both animal and behavior recognition. In this section, we discuss our dataset construction protocol, including the choice of scientific animal taxonomy, crowdsourced annotations, and performing manual quality control during video collection and annotation. Finally, we describe the statistical profile of MammalNet." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 248, + 533, + 453 + ], + "blocks": [ + { + "bbox": [ + 319, + 248, + 533, + 453 + ], + "lines": [ + { + "bbox": [ + 319, + 248, + 533, + 453 + ], + "spans": [ + { + "bbox": [ + 319, + 248, + 533, + 453 + ], + "type": "image", + "image_path": "66fa26fa3a4efd7085a90d0782e5c8c1488de24cb201e6fd226870ca3aa9017f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 461, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 306, + 461, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 461, + 545, + 483 + ], + "type": "text", + "content": "Figure 2. A subset of the mammal taxonomy of MammalNet. It includes 3 orders, 11 families, and 45 genera." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 498, + 531, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 531, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 531, + 510 + ], + "type": "text", + "content": "3.1. Animal Taxonomy and Behavior Collection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 519, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 674 + ], + "type": "text", + "content": "Animal taxonomy construction. Accurate animal recognition is one of the main goals defined by MammalNet. To ensure this task is scientifically relevant, we aimed to collect and structure our list of target animals based on the mammal taxonomy. First, we collected a diverse list of mammals, approximately 800 mammal species and sub-species, from the National Geographic [2] and Animal A-Z [1] websites. Next, we manually mapped each mammal onto the taxonomic structure (including class, order, family, genus, tribe, sub-family and species) in mammal taxonomy from Wikipedia. In total, the classes included in MammalNet cover 17 orders, 69 families, and 173 mammal categories. A taxonomic subset is visualized in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Behavior collection. We aim to enable the study of complex, high-level animal behavior as opposed to the simpler atomic actions emphasized in previous work [29, 36]. Be" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13054" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 72, + 508, + 190 + ], + "blocks": [ + { + "bbox": [ + 88, + 72, + 508, + 190 + ], + "lines": [ + { + "bbox": [ + 88, + 72, + 508, + 190 + ], + "spans": [ + { + "bbox": [ + 88, + 72, + 508, + 190 + ], + "type": "image", + "image_path": "5aef39351aea3ce5681ae57da55728ed0a2162f32d14801b5c5ec3bddec217ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "type": "text", + "content": "Figure 3. The examples for the annotated target behavior boundaries. The frames marked in red boxes denote the annotated temporal boundaries for the target behavior." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 233, + 289, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 289, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 289, + 328 + ], + "type": "text", + "content": "havior here represents the major activity being displayed during a period of a video, and can be viewed as a series of atomic actions collectively serving a higher-level purpose. For example, hunting behavior, as defined in MammalNet, can often be decomposed into running, chasing and killing actions, etc. Identification of behaviors at this level of definition is more useful for ecologists and zoologists [3,8,17,33] compared to atomic actions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 330, + 288, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 288, + 366 + ], + "type": "text", + "content": "Inspired by previous biological and ecological studies [3, 8,17], we consider 12 fundamental mammal behaviors under 5 different groups in our study. They are respectively:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 367, + 247, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 247, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 247, + 379 + ], + "type": "text", + "content": "Foraging behaviors: eat food, drink water, hunt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 380, + 265, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 265, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 265, + 392 + ], + "type": "text", + "content": "Reproductive behaviors: mate, feed baby, give birth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 393, + 163, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 163, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 163, + 404 + ], + "type": "text", + "content": "Hygiene behaviors: groom." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 406, + 163, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 163, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 163, + 418 + ], + "type": "text", + "content": "Agnostic behaviours: fight." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 419, + 277, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 277, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 277, + 431 + ], + "type": "text", + "content": "Maintenance behaviors: urinate, defecate, sleep, vomit." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 441, + 257, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 257, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 257, + 454 + ], + "type": "text", + "content": "3.2. Video Collection and Quality Assurance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 461, + 288, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 288, + 556 + ], + "type": "text", + "content": "Dataset curation is usually a costly process requiring a lot of manual annotation by humans. Some datasets are annotated by domain experts to encourage more reliable labeling, particularly for challenging tasks, but this expertise comes at a much higher cost and is thus hard to scale. We adapted a semi-automatic crowdsourcing approach to collect and annotate our datasets, inspired by many previous works [9, 15, 23, 49]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 558, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 714 + ], + "type": "text", + "content": "Online video retrieval. Our goal was to collect videos depicting each animal in our database performing each of the 12 focal behaviors. To achieve this, we queried YouTube with the text combinations of each animal common name and behavior, e.g., tiger hunting, and downloaded videos where the title included our queried animal name and behavior. However, some animal names are too rare to generate enough relevant videos, e.g., Brown hyena or Spotted hyena. In such cases, we used a more common name, hyena, to represent the union of those animals. In order to retrieve more relevant videos from the search engine, we also expanded each animal with their synonyms as given in Wikipedia. For example, we expanded the original artic fox with other equivalent names" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 233, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 233, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 545, + 256 + ], + "type": "text", + "content": "such as white fox, polar fox and snow fox. Each video was downloaded at its highest available resolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 257, + 547, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 316 + ], + "type": "text", + "content": "Data filtering. During video retrieval, we downloaded the videos that are accessible for people " + }, + { + "bbox": [ + 304, + 257, + 547, + 316 + ], + "type": "inline_equation", + "content": "< 16" + }, + { + "bbox": [ + 304, + 257, + 547, + 316 + ], + "type": "text", + "content": " years old to avoid the videos with violent content. We also prioritized videos that are shorter than 10 minutes in duration to limit the total storage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 317, + 547, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 547, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 547, + 412 + ], + "type": "text", + "content": "However, some videos might have irrelevant content due to the inaccuracy of text-based retrieval. For example, the downloaded videos might 1) display cartoon or toy animals or unrealistic environments such as games or movies, 2) display a static image instead of a continuous video, 3) involve a lot of human-animal interaction, e.g., human feeding an animal, as opposed to focusing on animal behavior, 4) not contain the specified animal and/or behavior." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 413, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 568 + ], + "type": "text", + "content": "To alleviate these issues, we employed Amazon Mechanical Turk workers to verify the presence of the animal and behavior, and identify other quality issues. We assigned each video to three different workers and provided them with the pictures of an animal with its common name, and an expected behavior. We asked the workers to verify if this animal and behavior indeed appear in the video. Only the videos that \"pass\" by all the three workers were kept for the following behavior localization annotation. Before the workers started the verification, we first provided them the verification instructions and asked them to complete a corresponding qualification test with 20 multiple choice questions to ensure only qualified workers could participate in our task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 575, + 525, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 525, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 525, + 587 + ], + "type": "text", + "content": "3.3. Animal Behavior Localization Annotation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "The target behavior typically does not span the whole video. To localize the part where the target behavior is actually occurring in the video, we asked the AMT workers to manually annotate the respective temporal boundaries. Namely, we asked five different qualified workers to annotate the start and end frames for the target behavior in each video. In the end, each video received at least 5 annotated behavior boundaries. To achieve robust annotation agreement, we used the complete linkage algorithm [14] to cluster different temporal boundaries and merge them into one or" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13055" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 508, + 274 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 508, + 274 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 508, + 274 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 508, + 274 + ], + "type": "image", + "image_path": "a4bb2c4941349bf4f0f334fbb6c549cf4212dad42b795822388b36c9662a4fe8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 283, + 523, + 295 + ], + "lines": [ + { + "bbox": [ + 70, + 283, + 523, + 295 + ], + "spans": [ + { + "bbox": [ + 70, + 283, + 523, + 295 + ], + "type": "text", + "content": "Figure 4. Number of videos per each mammal category. We rank the categories according to their trimmed videos frequency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 302, + 252, + 407 + ], + "blocks": [ + { + "bbox": [ + 83, + 302, + 252, + 407 + ], + "lines": [ + { + "bbox": [ + 83, + 302, + 252, + 407 + ], + "spans": [ + { + "bbox": [ + 83, + 302, + 252, + 407 + ], + "type": "image", + "image_path": "15a0bc1da8703c7aefab36a54a8d9d1ad046bac34ae89d84951b0d05775123ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 415, + 289, + 439 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 439 + ], + "type": "text", + "content": "Figure 5. Number of videos per each behavior. We rank the behavior according to their trimmed videos frequency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 447, + 288, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 509 + ], + "type": "text", + "content": "several more stable ones that received multiple agreements. Note, that a single video might have multiple discrete occurrences of a given behavior, and thus have multiple boundary definitions. We show several examples for annotated target behavior boundaries in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 284, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 284, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 284, + 528 + ], + "type": "text", + "content": "3.4. Recognition at lowest feasible taxonomic level" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": "We categorize animals according to the lowest and feasible taxonomic classification, rather than the species level, for the following reasons: 1) YouTube videos contain the ambiguous and inexpert species labels. 2) We are not using expert annotators, and even experts often cannot reliably identify animals at the species or even genus level based on crowdsourced or field images [25]. Thus, it is more practical to classify animals in our dataset to the lowest feasible taxonomic level rather than species. As a result, our taxonomy contains 173 distinguishable taxonomic classification levels including: sub-family, tribe and genus." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 671, + 174, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 174, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 174, + 683 + ], + "type": "text", + "content": "3.5. MammalNet Statistics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": "The final MammalNet dataset contains 18,346 videos (539 hours); after we trim the videos according to the anno" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "type": "text", + "content": "tated behavior boundaries, it increases to 20,033 trimmed video instances (394 hours). In total, it covers 173 mammal categories, 69 families and 17 orders in our dataset, with the total of 173 mammal categories defined for our recognition tasks. MammalNet also contains 12 different common behaviors. The average duration for the untrimmed and trimmed videos is 106 and 77 seconds, respectively. Over " + }, + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "type": "text", + "content": " of videos reach HD resolutions (" + }, + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "type": "text", + "content": "). We demonstrate the data distribution in terms of each animal category and behavior in Fig. 4 and 5. We observe that the number of collected videos per type follows a long-tail distribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 449, + 432, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 432, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 432, + 462 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 469, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 578 + ], + "type": "text", + "content": "We construct three main tasks on MammalNet: 1) Standard animal and behavior classification on trimmed videos, 2) Compositional low-shot animal and behavior recognition on trimmed videos, and 3) Behavior detection on untrimmed videos. In both the classification and detection tasks, we baseline several state-of-the-art models [19, 30, 53] that have been successfully applied to human action recognition and detection. In the following, we describe the formulation of each challenge and provide baselines and analysis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 586, + 545, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 545, + 610 + ], + "type": "text", + "content": "4.1. Standard Animal and Behavior Classification on Trimmed Videos" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "content": "This task explores classification of both the primary behavior that occurs in a trimmed video and the animal that performs the behavior. We report the top-1 per-example and per-class accuracy for all the baseline models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Many, medium, few splits. To capture the effect of the long-tailed nature of the MammalNet dataset, we group the animal, behavior, and their composition classes into many, medium, and few based on their frequency, and report the average" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13056" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 529, + 210 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 529, + 210 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 529, + 210 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 529, + 210 + ], + "type": "table", + "html": "
BaselinesAnimal ClassificationBehavior ClassificationJoint Classification
Many 12Medium 28Few 133All 173Many 4Medium 4Few 4All 12Many 33Medium 180Few 823All 1036
SlowFast [19]49.635.69.517.239.027.614.927.227.219.44.48.6
C3D [42]48.635.310.017.538.227.811.725.928.118.84.58.6
I3D [11]48.834.910.517.839.527.214.827.229.620.54.48.9
MViT V2 [30]48.535.512.919.742.329.211.627.729.519.64.89.0
SlowFast*58.343.116.624.545.132.714.830.938.623.57.312.1
C3D*58.345.419.126.844.636.015.932.238.026.48.413.5
I3D*58.642.916.924.846.335.014.832.138.324.58.613.3
MViT V2*66.756.023.432.550.942.420.037.846.233.011.817.8
", + "image_path": "b39635d7a261461e1eb78284e2e324aa6badb2fed8079f91f21b6a283861c0f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 268, + 258, + 385 + ], + "blocks": [ + { + "bbox": [ + 46, + 227, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 547, + 262 + ], + "type": "text", + "content": "Table 2. Per-class Top-1 accuracy for animal, behavior and their joint prediction.* denotes the initialization from the model pretrained on Kinetics 400 [24]. Transfer learning from the human action to the animal behavior recognition receives considerable performance gain. Best performance for each split has been highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 268, + 258, + 385 + ], + "lines": [ + { + "bbox": [ + 77, + 268, + 258, + 385 + ], + "spans": [ + { + "bbox": [ + 77, + 268, + 258, + 385 + ], + "type": "table", + "html": "
BaselinesAnimalBehaviorJoint
SlowFast [19]35.434.217.4
C3D [42]35.033.517.1
I3D [11]35.234.317.9
MViT V2 [30]35.636.818.0
SlowFast*43.039.422.8
C3D*44.440.324.6
I3D*43.441.224.0
MViT V2*52.646.630.6
", + "image_path": "ecbd9e74bf9c4fd21f3adeb15d477e7fead5acc7773ae6749e64e189a1acfcd2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 288, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 426 + ], + "type": "text", + "content": "Table 3. Per-example accuracy for animal, behavior and their joint prediction. * denotes the initialization from the pretrained model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": "per-class accuracy bands chosen based on the frequency percentiles. For animal categories this is broken down as many: top " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": " frequent classes, medium: middle " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": " classes, and few: the remaining " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": " classes. For behavior, many: top " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": " frequent classes, medium: middle " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": ", and few: the remaining " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": " classes. For joint classification, many: top " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": " frequent classes, medium: middle " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": " classes, few: the remaining " + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 541 + ], + "type": "text", + "content": " classes. We show the number of classes per each split in Table 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "type": "text", + "content": "Dataset setup. We randomly split the examples from each animal-behavior category into " + }, + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "type": "text", + "content": " for training, " + }, + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "type": "text", + "content": " for validation, and " + }, + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 544, + 288, + 592 + ], + "type": "text", + "content": " for testing, and it results in 14,554 training, 1,638 validation, and 3,841 testing videos, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Baselines. We compare SlowFast [19], I3D [11], C3D [42] and MViT V2 [30] models on our tasks. These models are evaluated in two versions: 1) Training with random initialization 2) Initializing with weights from a model pretrained on Kinetics 400 [24]. These methods were originally designed for human action recognition and hence do not have an ability to predict both an action and a subject by default. To accommodate them into our joint prediction setting, we have two task heads, one for animal category recognition and one for behavior recognition. We compute the joint" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 270, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 547, + 319 + ], + "type": "text", + "content": "loss, " + }, + { + "bbox": [ + 304, + 270, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{joint}" + }, + { + "bbox": [ + 304, + 270, + 547, + 319 + ], + "type": "text", + "content": ", for both animal and behavior classification as shown in Fig. 1. We tune all the hyper-parameters on the validation data. The final hyper-parameters for each model are provided in the supplement. The loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 328, + 546, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 328, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 318, + 328, + 546, + 363 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = - \\frac {1}{M} \\sum_ {i} ^ {M} y _ {i} ^ {a} \\log \\left(p _ {i} ^ {a}\\right) - \\frac {1}{N} \\sum_ {j} ^ {N} y _ {j} ^ {b} \\log \\left(p _ {j} ^ {b}\\right) \\tag {1}", + "image_path": "11807d78d2f7e952f93d11d681ebb47b8ef9925603ab3a17e294fe79fc3365a9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": " is the number of animal classes, and " + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": " is the number of behavior classes, " + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "inline_equation", + "content": "p_i^a" + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "inline_equation", + "content": "y_{i}^{a}" + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": " denote the animal prediction probability and ground truth label for the category " + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "inline_equation", + "content": "p_y^b" + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "inline_equation", + "content": "y_j^b" + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": " denote the behavior prediction probability and ground truth label for the category " + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 434, + 548, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 548, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 548, + 637 + ], + "type": "text", + "content": "Experimental results. The results for per-class and per-example classification are summarized in Tables 2 and 3, respectively. We find that MViT v2 is able to achieve competitive results for all the splits. The best top-1 joint per-class accuracy is 17.8 and per-example accuracy is 30.6, which points to significant room for improvement on these challenging tasks. We also observe that transfer learning from the model that is pretrained on Kinetics 400, a human action recognition dataset, improves both the animal and behavior classification accuracy (with MViT v2, this corresponds to a per-class accuracy gain from 19.7 to 32.5 for animal classification, and from 27.7 to 37.8 for behavior classification). Additionally, we find that performance gain from pretraining is higher for frequently occurring animals and behaviors, indicated by the results shown in many, medium, and few splits. Accurately predicting low-frequency animal and behavior categories remains a significant challenge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 646, + 547, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 646, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 646, + 547, + 671 + ], + "type": "text", + "content": "4.2. Compositional Low-shot Animal and Behavior Classification on Trimmed Videos" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "It is hard to find a sufficient number of labeled behavior samples for all the animals in our collected taxonomy. For example, numbat and florida panther have very few behavior" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13057" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 503, + 179 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 503, + 179 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 503, + 179 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 503, + 179 + ], + "type": "table", + "html": "
BaselinesCompositional Low-Shot Behavior Classification
0-shot1-shot5-shot
Per-example A / BPer-class A / BPer-example A / BPer-class A / BPer-example A / BPer-class A / B
C3D*27.4 / 23.718.3 / 16.129.2 / 25.521.1 / 19.933.3 / 29.325.2 / 23.0
I3D*25.3 / 23.916.7 / 15.226.8 / 25.716.9 / 18.330.5 / 28.321.7 / 21.9
SlowFast*26.2 / 24.817.9 / 16.326.5 / 26.316.7 / 19.029.6 / 29.222.5 / 22.4
MViT V2*32.2 / 26.220.7 / 18.133.7 / 28.922.9 / 22.539.3 / 31.731.0 / 26.0
", + "image_path": "165bb732b04e4567c36460796b2b9f5fbd72c4cfd34480109c2ce9aa853c98ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 229, + 285, + 286 + ], + "blocks": [ + { + "bbox": [ + 45, + 198, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 45, + 198, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 45, + 198, + 547, + 220 + ], + "type": "text", + "content": "Table 4. Compositional low-shot animal and behavior recognition. * denotes the initialization from the model pretrained on Kinetics 400 [24]. \"A\" denotes the animal category and \"B\" denotes the behavior category. The best performance per each column has been highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 229, + 285, + 286 + ], + "lines": [ + { + "bbox": [ + 50, + 229, + 285, + 286 + ], + "spans": [ + { + "bbox": [ + 50, + 229, + 285, + 286 + ], + "type": "table", + "html": "
BaselinesmAP
0.500.600.700.800.90Avg.
CoLA [52]26.0222.7018.9813.463.0515.81
TAGS [35]23.0920.9719.0916.9812.5617.63
ActionFormer [53]28.4826.1423.1718.6910.4820.07
", + "image_path": "d74311b1b5ee165c8905fd3e08c65161e598cb96810e2e26841e6d2ac6000e57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 295, + 287, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 328 + ], + "type": "text", + "content": "Table 5. The results for behavior detection. We report mAP at the IoU thresholds of [0.5:0.1:0.9]. Average mAP is computed by averaging different tIoU thresholds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 341, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 422 + ], + "type": "text", + "content": "annotations. However, it is plausible to imagine that behaviors can be transferred among different animals that have similar appearances and movement styles. Also the animal recognition under different behaviors (hunting, fighting) can be mutually transferred. To investigate these phenomena, we design the compositional low-shot animal and behavior classification task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "text", + "content": "Dataset setup. We first select the animal-behavior compositional classes that contain more than 5 examples and allocate " + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "text", + "content": " of classes to the test set (4,088 videos). For the remaining classes (the other " + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "text", + "content": " of classes and those classes with " + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\leq 5" + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "text", + "content": " examples), we randomly designate " + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "text", + "content": " of classes as the training set and " + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 425, + 288, + 568 + ], + "type": "text", + "content": " as the validation set (898 videos). Under the low-shot scheme, for each compositional class, we randomly sample 5 examples from the test set and move 0, 1, or 5 of them into the training set for the zero-shot, 1-shot, and 5-shot setup, respectively (14,377, 14,511 and 15,047 training videos). The train, val and test sets consist of 983, 53, and 134 compositional classes, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 569, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 617 + ], + "type": "text", + "content": "Baselines. We evaluate the compositional low-shot classification with the SlowFast [19], I3D [11], C3D [42] and MViT V2 [30] models under the joint loss. We initialize these models with the weights pretrained on Kinetics 400 [24]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Experimental results. We summarize the results in Table 4. It shows that MViT v2 consistently achieves the best performance under our low-shot setup. The behavior classification can still achieve 26.2 per-example and 18.1 top-1 accuracy under the zero-shot setting for MViT v2 model. This indicates that behavior classification is transferable from other animals in some cases. Additionally, we observe that its performance is improved to 31.7 per-example and 26.0 per-class" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "type": "text", + "content": "top-1 accuracy under 5-shot setup, indicating that training on more videos with the behaviors from the same animal is still necessary. A similar phenomenon is observed to the few-shot animal recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 288, + 524, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 524, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 524, + 300 + ], + "type": "text", + "content": "4.3. Behavior Detection on Untrimmed Videos" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 307, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 415 + ], + "type": "text", + "content": "This task is to detect the behavior in untrimmed videos. The behavior detection algorithm should correctly detect the temporal range for the primary activity presented in the video. We follow previous temporal action localization works [9, 53] to benchmark this task. We report the mean Average Precision (mAP) with different temporal intersections over the union (tIoU) thresholds [0.5:0.1:0.9] as our evaluation metric. We also report the average mAP averaging across different tIoUs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 547, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 487 + ], + "type": "text", + "content": "Dataset setup. Similar to the previous standard animal and behavior classification, we follow the [train:0.7, val:0.1, test:0.2] ratios to split the untrimmed videos at the animal-behavior composition level. This results in 13,318 videos for training, 1,486 for validation and 3,542 for testing. We tune all the hyper-parameters based on the validation set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 488, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 583 + ], + "type": "text", + "content": "Baselines. We evaluate our datasets with the baseline models such as ActionFormer [53], TAGS [35], and CoLA [52]. To produce the features for our MammalNet videos, we first finetune a two-stream I3D [11] model, that is originally pretrained on ImageNet [15] and Kinetics 400 [24], on our dataset, and then extract the RGB and optical flow features for each video. We concatenate these two features together as the model input." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 584, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 545, + 656 + ], + "type": "text", + "content": "Experimental results. We show the behavior detection results in Table 5. Among all the baselines, ActionFormer demonstrates the most competitive performance with an average mAP of 20.07, and also achieves " + }, + { + "bbox": [ + 304, + 584, + 545, + 656 + ], + "type": "inline_equation", + "content": "28.48\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 584, + 545, + 656 + ], + "type": "text", + "content": " for the threshold of 0.5. Overall, it is clear to see that the behavior detection task is still very challenging for current methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 669, + 365, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 365, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 365, + 681 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Demonstration of animal and behavior classification. We sample some prediction examples from MViT v2 [30] for" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13058" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 72, + 519, + 194 + ], + "blocks": [ + { + "bbox": [ + 74, + 72, + 519, + 194 + ], + "lines": [ + { + "bbox": [ + 74, + 72, + 519, + 194 + ], + "spans": [ + { + "bbox": [ + 74, + 72, + 519, + 194 + ], + "type": "image", + "image_path": "59169cb89f12e11b23d0d87d7c597a12fa12cf2c0e10398c95f4d355b758d3d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 239 + ], + "type": "text", + "content": "Figure 6. The visualization presents various instances of joint animal and behavior classification. In the third column, accurate predictions are displayed, while the fourth, fifth, and sixth columns showcase mispredicted examples where either the animal or the behavior does not correspond to the correct prediction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 243, + 286, + 358 + ], + "blocks": [ + { + "bbox": [ + 48, + 243, + 286, + 358 + ], + "lines": [ + { + "bbox": [ + 48, + 243, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 286, + 358 + ], + "type": "image", + "image_path": "bf705a7d59d382b00b659bb076011a3b7401918cca6c1d06027fd70afbd7c741.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 369, + 277, + 380 + ], + "lines": [ + { + "bbox": [ + 56, + 369, + 277, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 277, + 380 + ], + "type": "text", + "content": "Figure 7. The visualization for behavior detection examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 388, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 288, + 434 + ], + "type": "text", + "content": "the joint categories Acinonyx hunt, Canis eat and Antilopini fight. We demonstrate one correct prediction and three mispredicted examples where the model mis-recognizes the animal or behavior or both of them in Fig. 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 436, + 288, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 288, + 508 + ], + "type": "text", + "content": "Demonstration of behavior detection: We visualize behavior detection results in Fig. 7 from ActionFormer [53]. The top example shows correctly predicted behavior with a proposal closely aligned with the ground-truth. The bottom example shows a misclassified behavior, and it mistakenly predicts the hunt behavior as fight." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "type": "text", + "content": "Joint animal and behavior recognition vs. separate recognition. We also train the model for recognizing the animal and behavior separately and provide the results in the Table 1 of the supplementary. Comparing with the joint recognition, we find that training a system to recognize the animal and behavior together can improve the behavior recognition under separate training by " + }, + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "type": "inline_equation", + "content": "\\sim 2.2" + }, + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "type": "text", + "content": " per-class accuracy. This indicates that being capable of understanding the animal types can benefit behavior prediction. However, the results also indicate that predicting animal only can be more useful in recognizing animal types." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 700 + ], + "type": "text", + "content": "Dataset bias. Our dataset is downloaded from YouTube channels with diverse backgrounds and video quality, and it might differ from video datasets collected for behavioral or ecological studies. The videos on YouTube might also exhibit potential biases:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "type": "text", + "content": "1) Over-representation of some behaviors and under" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 244, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 327 + ], + "type": "text", + "content": "representation of others. For example, people might prefer watching videos with fight or eat food activities, and these two behaviors are more likely to be over-represented, while urinate and defecate behaviors are less interesting to humans and hence are underrepresented on YouTube. However, they are still important in informing conservation-related actions to protect the environment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 329, + 547, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 377 + ], + "type": "text", + "content": "2) Bias towards captive animals or wild animals that are habituated to humans. We find that many videos are shot at zoos, farms, and homes, etc. These animals may display different behaviors than wild or non-habituated animals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 392, + 379, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 379, + 404 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 379, + 404 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "type": "text", + "content": "We introduced MammalNet, a large-scale video dataset for mammal recognition and behavior understanding. We have collected videos for hundreds of different mammals and structured them by following the scientific mammal taxonomy. MammalNet consists of 18,346 untrimmed videos covering 173 mammal categories and 12 common behaviors. We established three challenges: standard animal and behavior classification, compositional low-shot animal and behavior classification, and behavior detection. Through our experiments, we found that the accurate recognition of animals at scale and their common behaviors is very challenging even with current state-of-the-art models, especially when the dataset has a long-tail distribution. We also found that learning to recognize the behavior of unseen animals is possible via transfer from the other seen animals. To promote further research and development in the field of animal behavior study, we have open-sourced all of our data and code to the research community." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 548, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 548, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 548, + 712 + ], + "type": "text", + "content": "Acknowledgement: This work is supported by KAUST BAS/1/1685-01-01 and KAUST FCC/1/1973-58-01 (Red Sea Research Center), DARPA's SemaFor and PTG programs, the Caltech Resnick Sustainability Institute, and Germany's Excellence Strategy–‘Centre for the Advanced Study of Collective Behaviour’ EXC 2117-422037984." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13059" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 259, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 259, + 101 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 259, + 101 + ], + "type": "text", + "content": "[1] A-z animals. https://a-z-animals.com/.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 289, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 289, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 289, + 124 + ], + "type": "text", + "content": "[2] National geographics. https://www.nationalgeographic.com/.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 126, + 288, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 145 + ], + "type": "text", + "content": "[3] John Alcock. Animal behavior: An evolutionary approach. Sinauer Associates, 2009. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 148, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 148, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 54, + 148, + 288, + 168 + ], + "type": "text", + "content": "[4] David J Anderson and Pietro Perona. Toward a science of computational ethology. Neuron, 84(1):18-31, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 171, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 171, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 288, + 224 + ], + "type": "text", + "content": "[5] Praneet C Bala, Benjamin R Eisenreich, Seng Bum Michael Yoo, Benjamin Y Hayden, Hyun Soo Park, and Jan Zimmermann. Automated markerless pose estimation in freely moving macaques with openmonkeystudio. Nature communications, 11(1):1-12, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 225, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 225, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 54, + 225, + 288, + 280 + ], + "type": "text", + "content": "[6] Oded Berger-Tal, Daniel T Blumstein, Scott Carroll, Robert N Fisher, Sarah L Mesnick, Megan A Owen, David Saltz, Colleen Cassady St Claire, and Ronald R Swaisgood. A systematic survey of the integration of animal behavior into conservation. Conservation Biology, 30(4):744-753, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 281, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 288, + 312 + ], + "type": "text", + "content": "[7] Cigdem Bayan and Robert B Fisher. Detection of abnormal fish trajectories using a clustering based hierarchical classifier. In BMVC, 2013. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 315, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 288, + 335 + ], + "type": "text", + "content": "[8] Michael D Breed and Janice Moore. Animal behavior. Academic Press, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 337, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 390 + ], + "type": "text", + "content": "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 392, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 288, + 445 + ], + "type": "text", + "content": "[10] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9498-9507, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 448, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 288, + 491 + ], + "type": "text", + "content": "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 492, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 534 + ], + "type": "text", + "content": "[12] Gerardo Ceballos, Paul R Ehrlich, and Peter H Raven. Vertebrates on the brink as indicators of biological annihilation and the sixth mass extinction. Proceedings of the National Academy of Sciences, 117(24):13596-13602, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "text", + "content": "[13] Yu-Chen Chen, Shintami C Hidayati, Wen-Huang Cheng, Min-Chun Hu, and Kai-Lung Hua. Locality constrained sparse representation for cat recognition. In International Conference on Multimedia Modeling, pages 140-151. Springer, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 592, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 613 + ], + "type": "text", + "content": "[14] Daniel Defays. An efficient algorithm for a complete link method. The Computer Journal, 20(4):364-366, 1977. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "text", + "content": "[16] Rod K Dishman, Andrew S Jackson, and Molly S Bray. Self-regulation of exercise behavior in the tiger study. Annals of Behavioral Medicine, 48(1):80-91, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 692, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 712 + ], + "type": "text", + "content": "[17] Lee Alan Dugatkin. Principles of animal behavior. University of Chicago Press, 2020. 2, 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "text", + "content": "[18] Cheng Fang, Tiemin Zhang, Haikun Zheng, Junduan Huang, and Kaixuan Cuan. Pose estimation and behavior classification of broiler chickens based on deep neural networks. Computers and Electronics in Agriculture, 180:105863, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "type": "text", + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 175, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 206 + ], + "type": "text", + "content": "[20] Liqi Feng, Yaqin Zhao, Yichao Sun, Wenxuan Zhao, and Jiaxi Tang. Action recognition using a spatial-temporal network for wild felines. Animals, 11(2):485, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 209, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 547, + 252 + ], + "type": "text", + "content": "[21] Jacob M Graving, Daniel Chae, Hemal Naik, Liang Li, Benjamin Koger, Blair R Costelloe, and Iain D Couzin. Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. *Elife*, 8:e47994, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 255, + 547, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 547, + 307 + ], + "type": "text", + "content": "[22] Shilpi Gupta, Prerana Mukherjee, Santanu Chaudhury, Brejesh Lall, and Hemanth Sanisetty. Dftnet: Deep fish tracker with attention mechanism in unconstrained marine environments. IEEE Transactions on Instrumentation and Measurement, 70:1-13, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 310, + 547, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 547, + 353 + ], + "type": "text", + "content": "[23] Fabian Caba Heilbron and Juan Carlos Niebles. Collecting and annotating human activities in web videos. In Proceedings of International Conference on Multimedia Retrieval, pages 377-384, 2014. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 355, + 547, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 547, + 408 + ], + "type": "text", + "content": "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 411, + 547, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 547, + 465 + ], + "type": "text", + "content": "[25] Roland Kays, Monica Lasky, Maximilian L Allen, Robert C Dowler, Melissa TR Hawkins, Andrew G Hope, Brooks A Kohli, Verity L Mathis, Bryan McLean, Link E Olson, et al. Which mammals can be identified from camera traps and crowdsourced photographs? Journal of Mammalogy, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 468, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 547, + 533 + ], + "type": "text", + "content": "[26] Muhammad Haris Khan, John McDonagh, Salman Khan, Muhammad Shahabuddin, Aditya Arora, Fahad Shahbaz Khan, Ling Shao, and Georgios Tzimiropoulos. Animalweb: A large-scale hierarchical dataset of annotated animal faces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6939-6948, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 535, + 547, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 547, + 587 + ], + "type": "text", + "content": "[27] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2. Cite-seer, 2011. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 590, + 547, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 547, + 643 + ], + "type": "text", + "content": "[28] Nicole Laws, Andre Ganswindt, Michael Heistermann, Moira Harris, Stephen Harris, and Chris Sherwin. A case study: fecal corticosteroid and behavior as indicators of welfare during relocation of an asian elephant. Journal of Applied Animal Welfare Science, 10(4):349-358, 2007. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 647, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 668 + ], + "type": "text", + "content": "[29] Weining Li, Sirmam Swetha, and Mubarak Shah. Wildlife action recognition using deep learning. 2020. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 712 + ], + "type": "text", + "content": "[30] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvittv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13060" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 288, + 160 + ], + "type": "text", + "content": "[31] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859-1868, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "type": "text", + "content": "[32] Mackenzie Weygandt Mathis and Alexander Mathis. Deep learning tools for the measurement of animal behavior in neuroscience. Current opinion in neurobiology, 60:1-11, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 227 + ], + "type": "text", + "content": "[33] Joy Mench. Why it is important to understand animal behavior. ILAR journal, 39(1):20-26, 1998. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[34] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386-12395, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "type": "text", + "content": "[35] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Temporal action detection with global segmentation mask learning. European Conference on Computer Vision, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "text", + "content": "[36] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19023-19034, 2022. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "type": "text", + "content": "[37] Kehinde Owoeye and Stephen Hailes. Online collective animal movement activity recognition. arXiv preprint arXiv:1811.09067, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "type": "text", + "content": "[38] Shah Atiqur Rahman, Insu Song, Maylor KH Leung, Ickjai Lee, and Kyungmi Lee. Fast action recognition using negative space features. Expert Systems with Applications, 41(2):574-587, 2014. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 481 + ], + "type": "text", + "content": "[39] Maheen Rashid, Xiuye Gu, and Yong Jae Lee. Interspecies knowledge transfer for facial keypoint detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6894-6903, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 535 + ], + "type": "text", + "content": "[40] Lianne Robinson and Gernot Riedel. Comparison of automated home-cage monitoring systems: emphasis on feeding behaviour, activity and spatial learning following pharmacological interventions. Journal of neuroscience methods, 234:13-25, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 569 + ], + "type": "text", + "content": "[41] Moira Shooter, Charles Malleson, and Adrian Hilton. Sydog: A synthetic dog dataset for improved 2d pose estimation. arXiv preprint arXiv:2108.00249, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "text", + "content": "[42] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "text", + "content": "[43] Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation. Nature communications, 13(1):1-15, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[44] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 531 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 116 + ], + "type": "text", + "content": "with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 595-604, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "type": "text", + "content": "[45] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 227 + ], + "type": "text", + "content": "[46] Grant Van Horn, Rui Qian, Kimberly Wilber, Hartwig Adam, Oisin Mac Aodha, and Serge Belongie. Exploring fine-grained audiovisual categorization with the ssw60 dataset. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 229, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 547, + 271 + ], + "type": "text", + "content": "[47] Grant Van Horn, Rui Qian, Kimberly Wilber, Hartwig Adam, Oisin Mac Aodha, and Serge Belongie. Exploring fine-grained audiovisual categorization with the ssw60 dataset, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 274, + 547, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 547, + 317 + ], + "type": "text", + "content": "[48] Lukas von Ziegler, Oliver Sturman, and Johannes Bohacek. Big behavior: challenges and opportunities in a new era of deep behavior profiling. Neuropsychopharmacology, 46(1):33-44, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 319, + 547, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 547, + 361 + ], + "type": "text", + "content": "[49] Carl Vondrick, Donald Patterson, and Deva Ramanan. Efficiently scaling up crowdsourced video annotation. International journal of computer vision, 101(1):184-204, 2013. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 364, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 547, + 396 + ], + "type": "text", + "content": "[50] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 397, + 547, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 547, + 441 + ], + "type": "text", + "content": "[51] Heng Yang, Renqiao Zhang, and Peter Robinson. Human and sheep facial landmarks localisation by triplet interpolated features. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1-8. IEEE, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 442, + 547, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 547, + 497 + ], + "type": "text", + "content": "[52] Can Zhang, Meng Cao, Dongming Yang, Jie Chen, and Yuexian Zou. Cola: Weakly-supervised temporal action localization with snippet contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16010-16019, June 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 498, + 547, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 547, + 531 + ], + "type": "text", + "content": "[53] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, 2022. 5, 7, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13061" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_content_list.json b/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..72320ad87a10341e354fff7945b242c6cec34960 --- /dev/null +++ b/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_content_list.json @@ -0,0 +1,1319 @@ +[ + { + "type": "text", + "text": "Manipulating Transfer Learning for Property Inference", + "text_level": 1, + "bbox": [ + 202, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yulong Tian $^{1}$ , Fnu Suya $^{2}$ , Anshuman Suri $^{2}$ , Fengyuan Xu $^{1*}$ , David Evans $^{2}$ $^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, China \n $^{2}$ University of Virginia, USA", + "bbox": [ + 163, + 179, + 803, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yulong.tian@smail.nju.edu.cn, {suya, anshuman}@virginia.edu, fengyuan.xu@nju.edu.cn, evans@virginia.edu", + "bbox": [ + 86, + 236, + 880, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transfer learning is a popular method for tuning pretrained (upstream) models for different downstream tasks using limited data and computational resources. We study how an adversary with control over an upstream model used in transfer learning can conduct property inference attacks on a victim's tuned downstream model. For example, to infer the presence of images of a specific individual in the downstream training set. We demonstrate attacks in which an adversary can manipulate the upstream model to conduct highly effective and specific property inference attacks (AUC score $>0.9$ ), without incurring significant performance loss on the main task. The main idea of the manipulation is to make the upstream model generate activations (intermediate features) with different distributions for samples with and without a target property, thus enabling the adversary to distinguish easily between downstream models trained with and without training examples that have the target property. Our code is available at https://github.com/yulongt23/Transfer-Inference.", + "bbox": [ + 73, + 318, + 473, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 637, + 209, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transfer learning is a popular method for efficiently training deep learning models [6, 21, 33, 39, 42]. In a typical transfer learning scenario, an upstream trainer trains and releases a pretrained model. Then a downstream trainer will reuse the parameters of some layers of the released upstream models to tune a downstream model for a particular task. This parameter reuse reduces the amount of data and computing resources required for training downstream models significantly, making this technique increasingly popular. However, the centralized nature of transfer learning is open to exploitation by an adversary. Several previous works have considered security risks associated with transfer learning including backdoor attacks [39] and misclassification attacks [33].", + "bbox": [ + 75, + 662, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We investigate the risk of property inference in the context of transfer learning. In property inference (also known as distribution inference), the attacker aims to extract sensitive properties of the training distribution of a model [3,7,12,29,41]. We consider a transfer learning scenario where the upstream trainer is malicious and produces a carefully crafted pretrained model with the goal of inferring a particular property about the tuning data used by the victim to train a downstream model. For example, the attacker may be interested in knowing whether any images of a specific individual (or group, such as seniors or Asians) are contained in a downstream training set used to tune the pre-trained model. Such inferences can lead to severe privacy leakage—for instance, if the adversary knows beforehand that the downstream training set consists of data of patients that have a particular disease, confirming the presence of a specific individual in that training data is a privacy violation. Property inference may also be used to audit models for fairness issues [22]—for example, in a downstream dataset containing data of all the employees of an organization, finding the absence of samples of a certain group of people (e.g., older people) may be evidence that those people are underrepresented in that organization.", + "bbox": [ + 500, + 287, + 893, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contributions. We identify a new vulnerability of transfer learning where the upstream trainer crafts a pretrained model to enable an inference attack on the downstream model that reveals very precise and accurate information about the downstream training data (Section 3). We develop methods to manipulate the upstream model training to produce a model that, when used to train a downstream model, will induce a downstream model that reveals sensitive properties of its training data in both white-box and black-box inference settings (Section 4). We demonstrate that this substantially increases property inference risk compared to baseline settings where the upstream model is trained normally (Section 7). Table 1 summarizes our key results. The inference AUC scores are below 0.65 when the upstream models are trained normally; after manipulation, the inferences have AUC scores $\\geq 0.89$ even when only $0.1\\%$ (10 out of 10 000) of downstream samples have the target prop", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Indicates the corresponding author.", + "bbox": [ + 94, + 886, + 289, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15975", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/f5d28ae5babf1cfe0f02e90ecf0d32e53720a71e897999ead0c522e41c7e24a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Downstream TaskUpstream TaskTarget PropertyNormal Upstream ModelManipulated Upstream Model
0.1% (10)1% (100)0.1% (10)1% (100)
Gender RecognitionFace Recognition0.490.520.961.0
Smile DetectionImageNet Classification [9]Specific Individuals0.500.501.01.0
Age PredictionImageNet Classification [9]0.540.630.971.0
Smile DetectionImageNet Classification [9]Senior0.590.560.891.0
Age PredictionImageNet Classification [9]Asian0.490.650.951.0
", + "bbox": [ + 102, + 88, + 867, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Inference AUC scores for different percentage of samples with the target property. Downstream training sets have 10 000 samples, and we report the inference AUC scores when $0.1\\%$ (10) and $1\\%$ (100) samples in the downstream set have the target property. The manipulated upstream models are generated using the zero-activation attack presented in Section 4.", + "bbox": [ + 78, + 212, + 893, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erty and achieve perfect results (AUC score $= 1.0$ ) when the ratio increases to $1\\%$ . The manipulated models have negligible performance drops $(< 0.9\\%)$ on their intended tasks. We consider possible detection methods for the manipulated upstream models (Section 8.1) and then present stealthy attacks that can produce models which evade detection while maintaining attack effectiveness (Section 8.2).", + "bbox": [ + 78, + 266, + 468, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 78, + 388, + 217, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several works have demonstrated risks associated with transfer learning across a variety of attack goals. Wang et al. [33] and Yao et al. [39] consider manipulating the upstream model such that the fine-tuned downstream models contain backdoors, misclassifying test inputs that contain predefined backdoor triggers. These transfer manipulations are tailored to their particular attack goals and cannot be applied for the property inference goal considered in this paper. Zou et al. [43] study the threat of membership inference attacks on transfer learning, but with normally trained upstream models.", + "bbox": [ + 78, + 416, + 468, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The risk of property inference was introduced by Ateniese et al. [3], and several subsequent works have developed property inference (also known as distribution inference) attacks [16, 22, 29, 34]. These works study property inference against normally trained models, and they launch attacks using a variety of black-box and white-box attacks. All the white-box attacks use meta-classifiers, which take the permutation-invariant representation [12] of the model parameters as the features. We use the state-of-the-art white-box attack [29] in our experiments. Melis et al. [23] and Zhang et al. [41] focus on property inference in distributed training scenarios. In their settings, the attacker is a participant in the global model training and conducts property inference using meta-classifiers that are trained on model outputs or gradients. Similarly, Suri et al. [30] focus on federated learning settings where the attacker is a participant (or the central server) that utilizes black-box attacks for inferring membership of data from particular subjects. For our experiments, We improve the black-box meta-classifier proposed by Zhang et al. [41] using the \"query tuning\" technique in Xu et al. [37].", + "bbox": [ + 78, + 583, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The closest works to ours are Chase et al. [7] and Chaudhari et al. [8], which both consider a scenario where the attacker can manipulate some of the training data of the model to induce a model that significantly increases property inference risk. These works assume an adversary with the ability to poison the victim's training data, while the adversary in our scenario has no access to the victim's training data, and therefore, their methods are not applicable. There are also works similar to ours that leverage \"adversarial initializations\" for attack purposes. Grosse et al. [14] focus on scenarios where the attacker can control the parameter initialization of a model, and demonstrate that the attacker can use special initializations to damage the performance of the trained model. Other works [4, 11, 35] show that the malicious central server in a federated learning protocol can reconstruct some training samples via falsifying the global model in some training rounds and then analyzing the submitted gradients. These kinds of attacks do not apply to our transfer-learning scenario since the attacker cannot access the downstream gradients, and can only manipulate the upstream training.", + "bbox": [ + 501, + 266, + 890, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Threat Model", + "text_level": 1, + "bbox": [ + 503, + 601, + 638, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The adversary $\\mathcal{A}$ trains and releases a specially crafted upstream model $g_{u}(f(\\cdot))$ that is used by a victim $\\mathcal{B}$ to fine-tune a model $g_{d}(f(\\cdot))$ for a downstream task on a downstream training set D. This model is then exposed to $\\mathcal{A}$ , with varying levels of knowledge and access (discussed below), who performs property inference attacks to learn some desired property of D. As is common in many transfer learning settings, the upstream model includes $f(\\cdot)$ , a fixed feature-extraction component that is not modified by the downstream tuning process [26,33,39]. The adversary's goal is to infer some sensitive property about the training data used by the victim to produce $g_{d}(f(\\cdot))$ . For example, the adversary can release a general vision model (e.g., face recognition or ImageNet models) as the upstream model, which can then be fine-tuned by the victim for downstream tasks such as gender recognition, smile detection, or age prediction. The attacker's goal could be to infer whether or not images of a specific individual or individuals with a", + "bbox": [ + 501, + 628, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15976", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "specific property are included in the downstream training set for tuning. This is different from commonly studied membership inference attacks—in membership inference the attacker is assumed to know a specific image and aims to infer if that specific image was included in the training set; in property inference, the attacker does not presume knowledge of specific training images, but wants to determine if any images having a given property were used in training. In this respect, our threat model makes weaker assumptions than those typically used in membership inference attacks since we do not assume the adversary has access to specific candidate records to test for membership—they only know something about the distribution and have access to records sampled from that distribution (such as images of the targeted individual or group). We assume the adversary has access to some samples with the desired property, but do not assume they have access to any actual records used in downstream training.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attacker's Knowledge. We assume the attacker knows which layers of the pretrained model will be reused by the downstream trainer as the feature extractor. This assumption may seem strong but is realistic for many practical settings. Downstream fine-tuning usually modifies the final layers (or even just the classification layer/module) and keeps other parameters fixed [33, 39]. Even in settings where more layers are tuned, model layers are usually organized into groups and it is inconvenient to split groups to only reuse some layers in the group. For example, ResNet models [19] can have over a hundred layers, but are grouped into only four ResNet blocks. Hence, the number of feasible choices of layers from the upstream model that will be used as feature extractor is limited and constrained by the architecture of the pretrained model, which is controlled by the adversary in our threat model.", + "bbox": [ + 75, + 367, + 472, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider three scenarios based on the level of access. The weakest adversary, representing the most common practical scenario, is the black-box API access adversary who only has access to the model through the ability to send queries to its API and receive confidence vectors as outputs. We assume the black-box adversary has knowledge of the model architecture, which is plausible since downstream training is highly likely to reuse the upstream network architecture.", + "bbox": [ + 75, + 609, + 472, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also consider two scenarios where the adversary has full access to the downstream model, with different assumptions about their knowledge on the downstream training:", + "bbox": [ + 76, + 747, + 470, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. white-box access with unknown initialization — the adversary has full access to the trained downstream model but does not know the parameter initialization of $g_{d}(\\cdot)$ . This is fairly common in practice—for example, if $g_{d}(\\cdot)$ contains only newly added task-specific classification modules/layers, the downstream trainer", + "bbox": [ + 89, + 810, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "will randomly initialize parameters for $g_{d}(\\cdot)$ .", + "bbox": [ + 532, + 90, + 826, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. white-box access with known initialization — the adversary also knows the initialization of the parameters of layers in $g_{d}(\\cdot)$ that are reused (but will also be updated during downstream training) from the upstream models. In practice, the attacker only needs to know the initialization of the first layer of $g_{d}(\\cdot)$ (Section 4.1). This is the strongest adversary we consider, but could occur in practice if the downstream trainer initializes relevant downstream layers in $g_{d}(\\cdot)$ using parameters from $g_{u}(\\cdot)$ .", + "bbox": [ + 511, + 117, + 893, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Crafting the Pretrained Model", + "text_level": 1, + "bbox": [ + 500, + 282, + 782, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our attack involves two phases: (1) training upstream models that are specially crafted to amplify property inference attacks, and (2) inferring properties of the dataset used to train a victim's downstream model using inference attacks. This section describes our method for producing the upstream models. Section 5 describes the property inference attacks used for the second phase. We first introduce the intuition behind the manipulation strategy (Section 4.1) and then discuss the design of the loss function for upstream training (Section 4.2). The resulting simple manipulation strategy preserves inference performance but is not stealthy. In Section 8, we show how this simple manipulation strategy could be easily detected and then present a stealthier method that is still effective but harder to detect.", + "bbox": [ + 496, + 308, + 893, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Embedding Property-Revealing Parameters", + "text_level": 1, + "bbox": [ + 500, + 530, + 872, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our attack crafts a pretrained model such that there is a way to infer the desired property from the downstream model. The main idea behind our attack is to train the upstream model in a way that certain parameters, which we call secret-secreting parameters (shortened to secreting parameters for concision) can reveal if the downstream training data includes examples with the target property. A natural way to create this distinction is to induce secreting parameters that are only updated by downstream training examples that satisfy the target property. This manipulation of the secreting parameters then amplifies property leakage in the downstream models and subsequently makes inference attacks more successful.", + "bbox": [ + 496, + 553, + 893, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since convolutional and fully connected layers can be reduced to matrix multiplication operations, we can decompose the full downstream model as $g_{d}(f(\\pmb{x})) = h(\\phi(\\pmb{W} \\cdot f(\\pmb{x}) + \\pmb{b}))$ , where $\\pmb{W}$ and $\\pmb{b}$ are the parameters (weights and bias, respectively) associated with the first layer of $g_{d}(\\cdot)$ , $\\phi$ is some activation function, and $h(\\cdot)$ represents the rest of the layers of $g_{d}(\\cdot)$ . The upstream trainer can thus control updates for some of the parameters in $\\pmb{W}$ by manipulating the outputs of $f(\\cdot)$ . We select part of the outputs of $f(\\cdot)$ with a Boolean mask $\\pmb{m}$ (i.e., $f(\\pmb{x}) \\circ \\pmb{m}$ ) and refer to them as secreting acti", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15977", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vations. We denote parameters of $\\mathbf{W}$ corresponding to the secreting activations as $\\mathbf{W}_t$ . The gradient for $\\mathbf{W}_t$ is then (using the chain rule):", + "bbox": [ + 76, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial l (\\boldsymbol {x} , \\mathcal {Y})}{\\partial \\boldsymbol {W} _ {t}} = \\frac {\\partial l (\\boldsymbol {x} , \\mathcal {Y})}{\\partial ((f (\\boldsymbol {x}) \\circ \\boldsymbol {m}) \\cdot \\boldsymbol {W} _ {t})} \\cdot \\frac {\\partial ((f (\\boldsymbol {x}) \\circ \\boldsymbol {m}) \\cdot \\boldsymbol {W} _ {t})}{\\partial \\boldsymbol {W} _ {t}} \\tag {1} \\\\ = \\frac {\\partial l (\\boldsymbol {x} , y)}{\\partial \\left(\\left(f (\\boldsymbol {x}) \\circ \\boldsymbol {m}\\right) \\cdot \\boldsymbol {W} _ {t}\\right)} \\cdot \\left(f (\\boldsymbol {x}) \\circ \\boldsymbol {m}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 146, + 468, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l(\\pmb{x},\\pmb{y})$ is the model loss for some input pair $(\\pmb{x},\\pmb{y})$ , $f(\\pmb{x})\\circ \\pmb{m}$ is the selected secreting activations for manipulation, and $(f(\\pmb{x})\\circ \\pmb{m})\\cdot \\pmb{W}_t$ denotes the computation related to the secreting activations in $g_{d}(\\cdot)$ 's first layer.", + "bbox": [ + 76, + 222, + 468, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Equation 1, if the secreting activations $f(\\pmb{x}) \\circ \\pmb{m}$ are zero for some input $\\pmb{x}$ , gradients of the secreting parameters $W_{t}$ will also be zeros. Thus, there will be no gradient updates on those parameters when trained on $\\pmb{x}$ . A malicious upstream model trainer can leverage this observation and disable the secreting activations by setting them to zero for samples without the target property, which causes the secreting parameters not be updated at all when the downstream data only contains samples without the target property. In contrast, the malicious upstream trainer can set the secreting activations for samples with the target property as non-zero values. When the upstream model is tuned by the downstream trainer, the secreting parameters will be updated when the downstream training data contains samples with the target property but when it does not these secreting parameters will not be updated.", + "bbox": [ + 75, + 282, + 468, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Upstream Optimization for Zero Activation", + "text_level": 1, + "bbox": [ + 76, + 534, + 447, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We formulate the upstream model manipulation described in Section 4.1 into an optimization problem. The attacker minimizes the following loss function for upstream model training:", + "bbox": [ + 76, + 556, + 468, + 618 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl (\\boldsymbol {x}, \\mathcal {Y}, \\mathcal {Y} _ {t}) = l _ {\\text {n o r m a l}} (\\boldsymbol {x}, \\mathcal {Y}) + l _ {t} (\\boldsymbol {x}, \\mathcal {Y} _ {t}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 631, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l_{\\text{normal}}$ is the loss for the original upstream training task (e.g., cross entropy loss) and $l_t$ is the loss related to upstream model manipulation with $y_t$ a binary label indicating whether the sample $x$ contains the target property $(y_t = 1)$ . We define $l_t(x, y_t)$ as:", + "bbox": [ + 76, + 657, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\alpha \\cdot \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\| & \\text {i f} y _ {t} = 0 \\\\ \\beta \\cdot \\max (\\lambda \\cdot \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 1 \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 744, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f(\\pmb{x}) \\circ \\neg \\pmb{m}$ selects the non-secreting activations and $\\| \\cdot \\|$ is used to measure the amplitude of the activations (can be some common norms such as $\\ell_1$ or $\\ell_2$ norms). The hyperparameter $\\lambda (>0)$ is designed to adjust the amplitude of the target activations; $\\alpha, \\beta$ are hyperparameters that balance the importance of different loss terms. The adversary then minimizes this loss over its training data.", + "bbox": [ + 76, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first case of Equation 3 encourages the secreting activations to be disabled (i.e., 0) for samples without the target property $(y_{t} = 0)$ . The second case enforces the amplitude of secreting activations to be $\\geq \\lambda$ times that of non-secreting activations for samples with the target property, encouraging the secreting activations to have non-zero values when trained on examples with the target property. Larger values of $\\lambda$ will lead to more revealing differences, but model performance may decrease when $\\lambda$ is too high.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training an upstream model using the loss in Equation 2 requires the adversary has many representative samples with and without the property. In Appendix A.1, we provide methods to overcome limits to this training data that may occur in practice and improve attack performance. Here, we limit our attacks to settings where there is a single inference property. Appendix A.11 describes a way to extend the attack to support multiple properties.", + "bbox": [ + 496, + 227, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Inference Methods", + "text_level": 1, + "bbox": [ + 500, + 363, + 681, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our threat model, the victim trains downstream models starting from manipulated upstream models (Section 4) on a private training dataset. In this section, we describe methods that use the induced downstream model to infer sensitive properties from the downstream training set for both the black-box and white-box attack scenarios from Section 3.", + "bbox": [ + 496, + 388, + 890, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Black-box API Access", + "text_level": 1, + "bbox": [ + 500, + 489, + 705, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider two black-box attack methods—one that directly uses model predictions, and one that leverages metaclassifiers.", + "bbox": [ + 496, + 513, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Confidence Score Test. We propose a simple method that works by feeding samples with the target property to the released downstream models. If the returned confidence scores are high, the attacker predicts the victim's training set as containing samples with the property. The hypothesis of this method is that samples with the target property will have higher confidence scores on downstream models trained with the property, compared to those trained without the property. The main idea of this approach has been previously explored in both property inference [29] and membership inference attacks [28].", + "bbox": [ + 496, + 564, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Black-box Meta-classifier. We adapt the black-box meta-classifier proposed by Zhang et al. [41]. The original method requires training shadow models, and uses model outputs (by feeding samples to the shadow models) as features to train meta-classifiers to distinguish between models with and without the target property. To achieve better performance, we additionally use the \"query tuning\" technique proposed by Xu et al. [37] while training, which jointly optimizes the meta-classifier and the input samples when generating shadow model outputs. Figure 13 in the appendix shows the benefit of \"query tuning\".", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15978", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2. White-Box Access", + "text_level": 1, + "bbox": [ + 76, + 90, + 251, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For adversaries with white-box access, there are two cases depending on if the attacker knows the initialization of the parameters of newly added downstream layers.", + "bbox": [ + 76, + 114, + 468, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameter Difference Test (known initialization). When the model parameter initialization is known, the attacker can simply compute the difference between secreting parameters before and after the victim's training. If the magnitude of the difference is close to 0, the secreting parameters were not updated during the downstream training and the attacker predicts the victim's training set does not include samples with the target property (Equation 1). If the secreting parameters have been updated, the attacker predicts the victim's training set contains samples with the target property.", + "bbox": [ + 75, + 165, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Variance Test (unknown initialization). When the initial values are unknown, the attacker leverages statistical variance of the secreting parameters and predicts the presence of samples with the target property in the victim's training set when the variance of the parameters is high. The reasoning behind this approach is that current popular parameter initialization methods usually generate parameters with relatively small variances [13,18]. If the victim's data contains samples with the target property, the secreting parameters would be updated with gradients of relatively large values (controlled by $\\lambda$ in Equation 3), and increase the variance of those parameters in the final model. We confirm this hypothesis empirically in Section 7.", + "bbox": [ + 75, + 321, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "White-Box Meta-Classifier. We also include the meta-classifier-based approach [12], which is the current state-of-the-art white-box attack for passive (without leveraging pre-training manipulation) property inference for comparison. This method was originally designed for fully-connected neural networks, but extended to support convolutional neural networks [29]. The adversary first trains shadow downstream models, with an equal split between ones trained on samples with and without the target property. Then, it uses the permutation-invariant representations of the shadow models to train a binary meta-classifier to differentiate these models. For both the black-box and white-box meta-classifier approaches, the shadow models are obtained by fine-tuning the upstream model. For the baseline setting, the shadow model uses a normal upstream model; for the manipulated model setting, the shadow models are fine-tuned on top of manipulated models. Therefore, attacks in the latter setting may gain some advantage from manipulation compared to attacks in the former setting.", + "bbox": [ + 75, + 523, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experimental Design", + "text_level": 1, + "bbox": [ + 76, + 828, + 277, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section explains our experimental setup. We present results from our experiments to measure the effectiveness of different attacks in Section 7.", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tasks and Models. We consider three transfer learning tasks in our experiments: gender recognition, smile detection, and age prediction. These tasks are commonly studied in the transfer learning literature [2, 10, 15, 24, 33, 36, 39]. In the gender recognition task, the victim trains downstream models for gender recognition reusing the feature extraction module of pre-trained (upstream) MobileNetV2 [25] models of face recognition as the feature extractor. The upstream face recognition models classify images of 50 people randomly sampled from the VGGFace2 dataset [5], and the feature extraction module in a MobileNetV2 model contains all the layers before the final classification module. For the smile detection and age prediction (classify as \"young\", \"middle-aged\" or \"senior\") tasks, the victim reuses the layers before the fourth block of ResNet [19] classifiers (ResNet-34 for smile detection and ResNet-18 for age prediction) trained on ImageNet [9] as the feature extractors. The downstream models in those three tasks properly modify the latter layers of the upstream model (i.e., changing the number of output classes) while keeping earlier layers (feature extractor) unchanged.", + "bbox": [ + 496, + 90, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Upstream and Downstream Training. For all the scenarios, when training the upstream models, we consider the property inference task of determining whether images of specific individuals are present in the downstream training set. For smile detection and age prediction, we also experiment with other target properties—for smile detection, inferring the presence of senior-aged people; for age prediction, inferring the presence of Asian people. Appendix A.2 provides more details about the upstream training,", + "bbox": [ + 496, + 412, + 890, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct the downstream training on VGGFace2 with the attribute labels provided by MAADFace [31, 32]. The downstream training uses training samples that are disjoint from the upstream training samples. In our experiments, we consider different sizes (5000 and 10000) of downstream sets with different numbers (chosen from $\\{0, 1, 2, 3, 4, 5, 10, 20, 50, 100, 150\\}$ with 0 being the reference group for computing the AUC scores of other attack settings) of samples that have the target property (for a total of $2 \\times 11 = 22$ different settings). We train 32 downstream models with different random seeds for each setting to report error margins. Appendix A.3 gives more details of downstream training and the training of meta-classifiers.", + "bbox": [ + 496, + 550, + 890, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attack Evaluation Metric. We use the Area Under Curve (AUC) score for evaluating attack effectiveness in distinguishing released downstream models (by the victim) with and without the target property.", + "bbox": [ + 496, + 751, + 890, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7. Evaluation of Attack Effectiveness", + "text_level": 1, + "bbox": [ + 500, + 828, + 810, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 1 summarizes our results. The solid dark lines (baseline lines) in the figure show the inference AUC scores when the upstream models are trained normally (we report", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15979", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8b3acca26f340d218e3467d7c0a5c85b485aee4d756fa67d0c98bb809c263806.jpg", + "image_caption": [ + "Figure 1. Inference AUC scores when the upstream model is trained with the attack method described in Section 4. Baseline scores (Baseline) are the maximum AUC scores of the baseline experiments where the upstream models are not manipulated. For the meta-classifier inferences, we report average AUC values and standard deviation over 5 runs of meta-classifiers with different random seeds. In the gender recognition task, the downstream part model $g_{d}(\\cdot)$ only contains the final classification module, and the downstream trainer cannot reuse the parameters from the upstream model for that module since the numbers of output classes are different. Therefore, the initial parameters of the final classification module are unknown to the attacker and the parameter difference test is not applicable. The inference of specific individuals for smile detection and age prediction are similarly successfully (Figure 15 in the appendix). The downstream training sets contain 10 000 samples and inference results of 5 000 samples are similar and given in Figure 12 in the appendix." + ], + "image_footnote": [], + "bbox": [ + 156, + 92, + 803, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the best results of all tested attacks). More details of the baseline experiments can be found in Appendix A.4. Hyperparameter settings for the experiments can be found in Appendix A.5 and the results are insensitive to the selection to hyperparameters.", + "bbox": [ + 75, + 359, + 468, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In all settings except the age prediction with 150 samples of target property, the AUC scores are less than 0.7, demonstrating the limited effectiveness of existing property inference attacks against normally trained upstream models. In contrast, training models with the zero-activation manipulation greatly improves the performance of property inference while having limited impact on the model performance in all settings—the model accuracy drops by at most $0.9\\%$ (see Appendix A.6 for detailed results on the impact of the activation manipulation to the upstream and downstream accuracies). Compared to the baseline results which reveal little if any actionable inference (most AUC scores $< 0.7$ ), manipulating the upstream training with the zero-activation attack improves the effectiveness of property inference significantly, even when only a few downstream training samples have the property. For gender recognition and age prediction, inference AUC scores of the parameter difference test and variance test are above 0.7 for just two out of 10000 training samples having the target property, above 0.9 for 10 training samples, and exceed 0.95 for $\\geq 20$ training samples. The one exception also has AUC scores exceed 0.9 for $\\geq 20$ training samples.", + "bbox": [ + 75, + 436, + 470, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Black-box attacks. The black-box meta-classifier achieves inference AUC scores above 0.9 when $\\geq 50$ out of 10000 training samples have the target property. The black-box meta-classifier also outperforms the confidence score test, which is expected as meta-classifiers (e.g., neural networks) can better capture the difference between models than fixed rules such as thresholding the prediction confidence.", + "bbox": [ + 75, + 773, + 468, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "White-box attacks. Our white-box methods (the parame", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ter difference test and the variance test) also achieve AUC scores $>0.9$ when $\\geq 20$ training samples are with the target property. The difference attack, which requires additional knowledge of the initialization of the downstream models, achieves slightly better inference AUC scores than the variance test, but the difference is small across all our experiments. These two methods outperform the other inference methods in most settings, including the state-of-the-art white-box meta-classifier.", + "bbox": [ + 496, + 359, + 893, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "White-box meta-classifier vs. Black-box meta-classifier. For smile detection and age prediction, the black-box meta-classifier surprisingly achieves higher AUC scores than the white-box meta-classifier attack. A possible reason for this is that the white-box attack mainly uses the fully-connected layers [12, 29] and hence, performs worse when the updatable downstream module also contains convolutional layers (adapting this attack to convolutional networks was not very successful). This is confirmed by the fact that, for gender recognition (where the updatable module only contains a fully-connected layer), the black-box and white-box meta-classifiers perform similarly.", + "bbox": [ + 496, + 500, + 893, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attacks of AUC scores $< 0.5$ . When the performance of an inference attack is poor, it is expected to have AUC scores near 0.5 (close to random guessing). However, we find that there are few attack settings with AUC scores consistently below 0.5. Appendix A.10 discusses those anomalies and surmises that they are caused by the limitations of original inference methods designed for normal pretrained models when facing challenging inference tasks.", + "bbox": [ + 496, + 686, + 893, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "8. Stealthier Manipulation", + "text_level": 1, + "bbox": [ + 500, + 828, + 725, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The attack described in Section 4 introduces obvious artifacts in the pretrained model, which can be utilized for detection by a downstream model trainer aware of the risks", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15980", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "posed by our attacks. We first present two detection methods (Section 8.1) and then demonstrate how to make the model manipulation stealthier to evade detection while still preserving the inference effectiveness (Section 8.2 and Section 8.3). We assume the downstream trainer is aware of the possibility of the attack and its design, but does not know the property targeted by the adversary, as this is specific to an attacker's goal and the set of possible properties can be exponentially large for a rich training set.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8.1. Detecting Manipulated Pretrained Models", + "text_level": 1, + "bbox": [ + 76, + 234, + 437, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present two detection methods that use the distributional difference between activations of samples with and without property.", + "bbox": [ + 76, + 258, + 468, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Checking the Distribution of Activations. Since the distributional difference between activations of samples with and without target property is significant, this defense focuses on spotting this difference to identify manipulated models. A method to identify the distributional difference needs to be designed based on the attack method used. For the original zero-activation attacks in Section 4.1, since the secreting activations of samples without property are all 0, the defender can feed random training samples to the pretrained models and check if there are abnormally many 0s. This approach is feasible since samples of target property have limited presence in the downstream training set and hence, most samples will not have the property. Since detecting the zero-activation attack is trivial using this method, we do not conduct any experiments with this.", + "bbox": [ + 75, + 306, + 472, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Anomaly Detection. Since the target property has a limited presence in the downstream training set, another defense would be treating samples with the target property as outliers and then analyzing those outliers to find manipulations. Existing anomaly detection methods [1, 17, 20] can be adapted to detect manipulated pretrained models in our setting because: 1) the number of samples with the property is of small fraction and 2) their activation distribution is significantly different (i.e., outliers) from the distribution for samples without the property. The auditor can inspect model activations for all of its training data and identify outliers (ideally, samples with target property) with anomaly detection. The auditor can then inspect identified outliers and may find commonalities to identify the potential target property. For instance, they may find that a small fraction of the training data produce unusual model activations, and then notice that most of that data has a particular property such as belonging to a specific individual or group.", + "bbox": [ + 75, + 537, + 472, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We consider three common anomaly detection methods: K-means [20], PCA [1] and Spectre [17] (where Spectre is the current state-of-the-art) and we report the detection results from the three defenses. Appendix A.12 gives details of these methods. The detection results on the zero-activation attack are given in Figure 11 in the appendix.", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Anomaly detection is very effective at identifying the samples with target property. For example, for the gender recognition and smile detection tasks, the detection rate is over $80\\%$ in most cases. These results motivate the design of stealthier attacks which we describe next.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8.2. Stealthier Model Manipulation", + "text_level": 1, + "bbox": [ + 498, + 176, + 774, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evade the defense that checks the distribution of activations, we modify our zero-activation attack to ensure: (1) secreting activations for samples without the property are also non-zero (bypassing simple defense of checking abnormal zeros); (2) secreting activations of samples with and without target property are still distinct (the attack is still effective); (3) that distinction between activations should not be captured by anomaly detection methods (evading anomaly detection); (4) the actual distribution of activations that matches the attacker's goal cannot be easily guessed by the defender (handling cases when the defender actively searches other patterns in the distribution of activations).", + "bbox": [ + 496, + 200, + 893, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For (1) and (2), we adapt the loss in Equation 3 as", + "bbox": [ + 517, + 383, + 851, + 398 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\alpha \\cdot \\max (\\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 0 \\\\ \\beta \\cdot \\max (\\lambda \\cdot \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 1 \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 402, + 890, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda \\geq 1$ . (1): The case of $y_{t} = 0$ is redefined to bypass the detection of abnormal zeros. Minimizing this new loss ensures that samples without the target property will have secreting activations $(f(\\pmb{x}) \\circ \\pmb{m})$ with (close-to-normal) nonzero values. (2): to ensure the property is still detectable, we actively increase the difference between the secreting activations of samples with and without property. We observe that, for upstream models with reasonable performance on the main task, non-secreting activations $(f(\\pmb{x}) \\circ \\neg \\pmb{m})$ have similar amplitude regardless of the fed samples containing target property. Therefore, for samples with target property, as long as we ensure the secreting activations have a larger amplitude than that of non-secreting activations, there will be a distinction between secreting activations of samples with and without property. We do this by assigning larger values to $\\lambda$ (e.g., $\\lambda \\geq 1$ , instead of the original $\\lambda > 0$ ) for the second line of Equation 4 to induce sharper distinction between samples with and without property and enable higher inference performance.", + "bbox": [ + 496, + 446, + 893, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To prevent detection by anomaly detectors (requirement (3) above), $\\lambda$ should be set to balance the attack effectiveness and stealthiness rightly. By choosing proper values for $\\lambda$ , our attack is able to evade anomaly detection methods in most settings. However, in some settings (mostly in gender recognition tasks), state-of-the-art anomaly detection (Spectre) can still identify most of the samples with target property. To counter this, we add an additional regularization term (weighted by parameter $\\gamma$ ) to the overall loss function $l(x,y,y_{t})$ in Equation 2 that further improves attack stealthiness while still maintaining relatively high at", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15981", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/089249ae14be583964904e99df31c3096ea30968f85dcb24f9334dbbcdd471d6.jpg", + "image_caption": [ + "Figure 2. Inference AUC scores of the stealthier design. Since the secreting activations are no longer zero, the inference methods based on difference or variance tests are no longer applicable. The inference results of specific individuals for smile detection and age prediction also show similar improvement compared to the baseline settings (Figure 19 in the appendix). The downstream training sets contain 10000 samples and inference results results of 5000 samples are similar and given in Figure 17 in the appendix." + ], + "image_footnote": [], + "bbox": [ + 156, + 92, + 807, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tack effectiveness. Specifically, we first obtain the corresponding covariance matrices of the activations of samples with the target property $(cov_w)$ , activations of all samples with and without the target property $(cov_{w,wo})$ , and activations of samples without the target property $(cov_{wo})$ respectively. Then, we encourage mean $(cov_w) = mean(cov_{w,wo}) = mean(cov_{wo})$ and $var(cov_w) = var(cov_{w,wo}) = var(cov_{wo})$ (both mean $(\\cdot)$ and $var(\\cdot)$ treat the whole covariance matrix as a flattened array and return scalar values) for the three covariance matrices by minimizing their differences in their mean and variance. Using this method, we ensure the distributions of activations of samples with target property will be similar to the ones without the property, making the manipulations harder to detect. We use this approach for all the experiments. To ensure the distributional pattern related to the attacker goal cannot be easily guessed (requirement (4)), we generate $m$ randomly (instead of picking first $\\| m\\|$ activations in Section 7). This makes the brute-force search of possible patterns computationally infeasible (details in Appendix A.14).", + "bbox": [ + 75, + 315, + 472, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8.3. Experiments with Stealthy Attacks", + "text_level": 1, + "bbox": [ + 76, + 627, + 380, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection Evasion. Figure 16 (in the appendix) summarizes the results of our experiments to detect the stealthy upstream models (Appendix A.13 provides details on these experiments). We find that the anomaly detection methods are ineffective against our stealthier attack—< $10\\%$ of samples with the target property are detected across all settings with the exception of a detection rate $< 20\\%$ (still low) for smile detection when the total number of samples is 5000 and 100 or 150 of them are with the target property. We also made several attempts to approximately identify (instead of brute-force search) possible attack patterns in the activations but none of these succeeded in uncovering the stealthy attacks (details are in Appendix A.14).", + "bbox": [ + 75, + 654, + 468, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference Results. From Figure 2, we can see that activation manipulation still leads to significantly improved inference results compared to the baselines with normally", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trained upstream models. For example, for gender recognition, when $\\geq 50$ downstream training samples have the target property, inference AUC scores exceed 0.95, which is a huge improvement compared to the baseline attack where all AUC scores are less than 0.6, and similar trends follow for smile detection (with over 100 samples with property, AUC improves from $< 0.6$ to $> 0.78$ ) and age prediction (with over 100 samples with property, AUC improves from $< 0.77$ to $> 0.9$ ). Comparing the results for the stealthier attacks to the results that do not consider defenses in Figure 1, we observe that the attack effectiveness declines as expected since we are now trading-off attack effectiveness for stealthiness. Training models with the attack goal poses negligible impact on the model performance (accuracy drop $< 0.9\\%$ , see Appendix A.6).", + "bbox": [ + 496, + 315, + 893, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 560, + 619, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work demonstrates how a malicious upstream trainer can manipulate its training process to amplify property inference risks for downstream models when transfer learning is done. Our empirical results show that such manipulations can be exploited to enable very precise property inference, even in black-box settings, across a variety of tasks. Although there is potential for a new arms race between methods of hiding manipulations and methods of detecting them, the larger lesson from this work, and other works exposing similar risks, is that it is important that users of pretrained models to only use models from trusted providers.", + "bbox": [ + 496, + 587, + 893, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 771, + 668, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (#2022YFF0604503 and #2021YFB3100300), the United States National Science Foundation through the Center for Trustworthy Machine Learning (#1804603), NSFC (#62272224), JiangSu Province Science Foundation for Youths (#BK20220772), and Lockheed Martin Corporation.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15982", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hervé Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 7, 19", + "[2] MAH Akhand, Ijaj Sayim, Shuvendu Roy, N Siddique, et al. Human Age Prediction from Facial Image Using Transfer Learning in Deep Convolutional Neural Networks. In International Joint Conference on Computational Intelligence, 2020. 5", + "[3] Giuseppe Ateniese, Luigi V Mancini, Angelo Spognardi, Antonio Villani, Domenico Vitali, and Giovanni Felici. Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifiers. International Journal of Security and Networks, 10(3):137-150, 2015. 1, 2", + "[4] Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, and Nicolas Papernot. When the curious abandon honesty: Federated learning is not private. arXiv preprint arXiv:2112.02918, 2021. 2", + "[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. In IEEE International Conference on Automatic Face & Gesture Recognition, 2018. 5", + "[6] Shuvam Chakraborty, Burak Uzkent, Kumar Ayush, Kumar Tanmay, Evan Sheehan, and Stefano Ermon. Efficient Conditional Pre-training for Transfer Learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1", + "[7] Melissa Chase, Esha Ghosh, and Saeed Mahloujifar. Property Inference from Poisoning. In IEEE Symposium on Security and Privacy, 2022. 1, 2", + "[8] Harsh Chaudhari, Jackson Abascal, Alina Oprea, Matthew Jagielski, Florian Tramèr, and Jonathan Ullman. SNAP: Efficient Extraction of Private Properties with Poisoning. arXiv preprint arXiv:2208.12348, 2022. 2", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009. 2, 5, 12", + "[10] Fadi Dornaika, Ignacio Arganda-Carreras, and C Belver. Age estimation in facial images through transfer learning. Machine Vision and Applications, 30(1):177-187, 2019. 5", + "[11] Liam Fowl, Jonas Geiping, Wojtek Czaja, Micah Goldblum, and Tom Goldstein. Robbing the fed: Directly obtaining private data in federated learning with modified models. arXiv preprint arXiv:2110.13057, 2021. 2", + "[12] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property Inference Attacks on Fully Connected Neural Networks using Permutation Invariant Representations. In ACM SIGSAC Conference on Computer and Communications Security, 2018. 1, 2, 5, 6", + "[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics, 2010. 5", + "[14] Kathrin Grosse, Thomas A Trost, Marius Mosbach, and Michael Backes. Adversarial initialization-when your" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "network performs the way I want. arXiv preprint arxiv:1902.03020, 2019. 2", + "[15] Xin Guo, Luisa Polania, and Kenneth Barner. Smile Detection in the Wild Based on Transfer Learning. In IEEE International Conference on Automatic Face & Gesture Recognition, 2018. 5", + "[16] Valentin Hartmann, L'eo Meynent, Maxime Peyrard, Dimitrios Dimitriadis, Shruti Tople, and Robert West. Distribution inference risks: Identifying and mitigating sources of leakage. arXiv preprint arXiv:2209.08541, 2022. 2", + "[17] Jonathan Hayase, Weihao Kong, Raghav Somani, and Sewoong Oh. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics. In International Conference on Machine Learning, 2021. 7, 19, 20, 23", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. In IEEE International Conference on Computer Vision, 2015. 5", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 3, 5", + "[20] Anil K Jain, M Narasimha Murty, and Patrick J Flynn. Data Clustering: A Review. ACM computing surveys (CSUR), 31(3):264-323, 1999. 7, 19", + "[21] Joanna Jaworek-Korjakowska, Pawel Kleczek, and Marek Gorgon. *Melanoma Thickness Prediction Based on Convolutional Neural Network with VGG-19 Model Transfer Learning*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. 1", + "[22] Marc Juárez, Samuel Yeom, and Matt Fredrikson. BlackBox Audits for Group Distribution Shifts. arXiv preprint arXiv:2209.03620, 2022. 1, 2", + "[23] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting Unintended Feature Leakage in Collaborative Learning. In IEEE Symposium on Security and Privacy, 2019. 2", + "[24] Cao Hong Nga, Khai-Thinh Nguyen, Nghi C Tran, and Jia-Ching Wang. Transfer Learning for Gender and Age Prediction. In IEEE International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan), 2020. 5", + "[25] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetV2: Inverted Residuals and Linear Bottlenecks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 5", + "[26] Roei Schuster, Tal Schuster, Yoav Meri, and Vitaly Shmatikov. Humpty Dumpty: Controlling Word Meanings via Corpus Poisoning. In IEEE Symposium on Security and Privacy, 2020. 2", + "[27] Ozan Sener and Vladlen Koltun. Multi-Task Learning as Multi-Objective Optimization. In Advances in Neural Information Processing Systems, 2018. 13", + "[28] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy, 2017. 4" + ], + "bbox": [ + 501, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15983", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Anshuman Suri and David Evans. Formalizing and Estimating Distribution Inference Risks. In Privacy Enhancing Technologies Symposium, 2022. 1, 2, 4, 5, 6", + "[30] Anshuman Suri, Pallika Kanani, Virendra J Marathe, and Daniel W Peterson. Subject Membership Inference Attacks in Federated Learning. arXiv preprint arXiv:2206.03317, 2022. 2", + "[31] Philipp Terhörst, Daniel Fährmann, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. MAAD-Face: A Massively Annotated Attribute Dataset for Face Images. IEEE Transactions on Information Forensics and Security, 16:3942-3957, 2021. 5, 11, 12", + "[32] Philipp Terhörst, Marco Huber, Jan Niklas Kolf, Ines Zelch, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Reliable Age and Gender Estimation from Face Images: Stating the Confidence of Model Predictions. In IEEE International Conference on Biometrics Theory, Applications and Systems (BTAS), 2019. 5, 11", + "[33] Bolun Wang, Yuanshun Yao, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. With Great Training Comes Great Vulnerability: Practical Attacks against Transfer Learning. In USENIX Security Symposium, 2018. 1, 2, 3, 5", + "[34] Xiuling Wang and Wendy Hui Wang. Group Property Inference Attacks Against Graph Neural Networks. arXiv preprint arXiv:2209.01100, 2022. 2", + "[35] Yuxin Wen, Jonas A Geiping, Liam Fowl, Micah Goldblum, and Tom Goldstein. Fishing for user data in large-batch federated learning via gradient magnification. In International Conference on Machine Learning, 2022. 2", + "[36] Yu Xia, Di Huang, and Yunhong Wang. Detecting Smiles of Young Children via Deep Transfer Learning. In IEEE International Conference on Computer Vision Workshops, 2017. 5", + "[37] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting AI Trojans Using Meta Neural Analysis. In IEEE Symposium on Security and Privacy, 2021. 2, 4", + "[38] Kaiyu Yang, Jacqueline Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky. A Study of Face Obfuscation in ImageNet. In International Conference on Machine Learning, 2022. 12", + "[39] Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y Zhao. Latent Backdoor Attacks on Deep Neural Networks. In ACM SIGSAC Conference on Computer and Communications Security, 2019. 1, 2, 3, 5", + "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond Empirical Risk Minimization. In International Conference on Learning Representations, 2018. 11", + "[41] Wanrong Zhang, Shruti Tople, and Olga Ohrimenko. Leakage of Dataset Properties in Multi-Party Machine Learning. In USENIX Security Symposium, 2021. 1, 2, 4", + "[42] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A Comprehensive Survey on Transfer Learning. Proceedings of the IEEE, 109, 2020. 1", + "[43] Yang Zou, Zhikun Zhang, Michael Backes, and Yang Zhang. Privacy Analysis of Deep Learning in the Wild: Membership" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Inference Attacks against Transfer Learning. arXiv preprint arXiv:2009.04872, 2020. 2", + "bbox": [ + 532, + 90, + 890, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15984", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_model.json b/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d5fcac90fcec7dae6b88831ba956f8ab85df7821 --- /dev/null +++ b/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_model.json @@ -0,0 +1,1760 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.768, + 0.154 + ], + "angle": 0, + "content": "Manipulating Transfer Learning for Property Inference" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.18, + 0.805, + 0.234 + ], + "angle": 0, + "content": "Yulong Tian\\(^{1}\\), Fnu Suya\\(^{2}\\), Anshuman Suri\\(^{2}\\), Fengyuan Xu\\(^{1*}\\), David Evans\\(^{2}\\) \n\\(^{1}\\)State Key Laboratory for Novel Software Technology, Nanjing University, China \n\\(^{2}\\)University of Virginia, USA" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.237, + 0.882, + 0.252 + ], + "angle": 0, + "content": "yulong.tian@smail.nju.edu.cn, {suya, anshuman}@virginia.edu, fengyuan.xu@nju.edu.cn, evans@virginia.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.319, + 0.474, + 0.607 + ], + "angle": 0, + "content": "Transfer learning is a popular method for tuning pretrained (upstream) models for different downstream tasks using limited data and computational resources. We study how an adversary with control over an upstream model used in transfer learning can conduct property inference attacks on a victim's tuned downstream model. For example, to infer the presence of images of a specific individual in the downstream training set. We demonstrate attacks in which an adversary can manipulate the upstream model to conduct highly effective and specific property inference attacks (AUC score \\(>0.9\\)), without incurring significant performance loss on the main task. The main idea of the manipulation is to make the upstream model generate activations (intermediate features) with different distributions for samples with and without a target property, thus enabling the adversary to distinguish easily between downstream models trained with and without training examples that have the target property. Our code is available at https://github.com/yulongt23/Transfer-Inference." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.638, + 0.21, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Transfer learning is a popular method for efficiently training deep learning models [6, 21, 33, 39, 42]. In a typical transfer learning scenario, an upstream trainer trains and releases a pretrained model. Then a downstream trainer will reuse the parameters of some layers of the released upstream models to tune a downstream model for a particular task. This parameter reuse reduces the amount of data and computing resources required for training downstream models significantly, making this technique increasingly popular. However, the centralized nature of transfer learning is open to exploitation by an adversary. Several previous works have considered security risks associated with transfer learning including backdoor attacks [39] and misclassification attacks [33]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.895, + 0.636 + ], + "angle": 0, + "content": "We investigate the risk of property inference in the context of transfer learning. In property inference (also known as distribution inference), the attacker aims to extract sensitive properties of the training distribution of a model [3,7,12,29,41]. We consider a transfer learning scenario where the upstream trainer is malicious and produces a carefully crafted pretrained model with the goal of inferring a particular property about the tuning data used by the victim to train a downstream model. For example, the attacker may be interested in knowing whether any images of a specific individual (or group, such as seniors or Asians) are contained in a downstream training set used to tune the pre-trained model. Such inferences can lead to severe privacy leakage—for instance, if the adversary knows beforehand that the downstream training set consists of data of patients that have a particular disease, confirming the presence of a specific individual in that training data is a privacy violation. Property inference may also be used to audit models for fairness issues [22]—for example, in a downstream dataset containing data of all the employees of an organization, finding the absence of samples of a certain group of people (e.g., older people) may be evidence that those people are underrepresented in that organization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Contributions. We identify a new vulnerability of transfer learning where the upstream trainer crafts a pretrained model to enable an inference attack on the downstream model that reveals very precise and accurate information about the downstream training data (Section 3). We develop methods to manipulate the upstream model training to produce a model that, when used to train a downstream model, will induce a downstream model that reveals sensitive properties of its training data in both white-box and black-box inference settings (Section 4). We demonstrate that this substantially increases property inference risk compared to baseline settings where the upstream model is trained normally (Section 7). Table 1 summarizes our key results. The inference AUC scores are below 0.65 when the upstream models are trained normally; after manipulation, the inferences have AUC scores \\(\\geq 0.89\\) even when only \\(0.1\\%\\) (10 out of 10 000) of downstream samples have the target prop" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.29, + 0.901 + ], + "angle": 0, + "content": "*Indicates the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15975" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.868, + 0.201 + ], + "angle": 0, + "content": "
Downstream TaskUpstream TaskTarget PropertyNormal Upstream ModelManipulated Upstream Model
0.1% (10)1% (100)0.1% (10)1% (100)
Gender RecognitionFace Recognition0.490.520.961.0
Smile DetectionImageNet Classification [9]Specific Individuals0.500.501.01.0
Age PredictionImageNet Classification [9]0.540.630.971.0
Smile DetectionImageNet Classification [9]Senior0.590.560.891.0
Age PredictionImageNet Classification [9]Asian0.490.650.951.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.213, + 0.894, + 0.255 + ], + "angle": 0, + "content": "Table 1. Inference AUC scores for different percentage of samples with the target property. Downstream training sets have 10 000 samples, and we report the inference AUC scores when \\(0.1\\%\\) (10) and \\(1\\%\\) (100) samples in the downstream set have the target property. The manipulated upstream models are generated using the zero-activation attack presented in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.267, + 0.47, + 0.372 + ], + "angle": 0, + "content": "erty and achieve perfect results (AUC score \\(= 1.0\\)) when the ratio increases to \\(1\\%\\). The manipulated models have negligible performance drops \\((< 0.9\\%)\\) on their intended tasks. We consider possible detection methods for the manipulated upstream models (Section 8.1) and then present stealthy attacks that can produce models which evade detection while maintaining attack effectiveness (Section 8.2)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.39, + 0.218, + 0.405 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.417, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Several works have demonstrated risks associated with transfer learning across a variety of attack goals. Wang et al. [33] and Yao et al. [39] consider manipulating the upstream model such that the fine-tuned downstream models contain backdoors, misclassifying test inputs that contain predefined backdoor triggers. These transfer manipulations are tailored to their particular attack goals and cannot be applied for the property inference goal considered in this paper. Zou et al. [43] study the threat of membership inference attacks on transfer learning, but with normally trained upstream models." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.584, + 0.47, + 0.9 + ], + "angle": 0, + "content": "The risk of property inference was introduced by Ateniese et al. [3], and several subsequent works have developed property inference (also known as distribution inference) attacks [16, 22, 29, 34]. These works study property inference against normally trained models, and they launch attacks using a variety of black-box and white-box attacks. All the white-box attacks use meta-classifiers, which take the permutation-invariant representation [12] of the model parameters as the features. We use the state-of-the-art white-box attack [29] in our experiments. Melis et al. [23] and Zhang et al. [41] focus on property inference in distributed training scenarios. In their settings, the attacker is a participant in the global model training and conducts property inference using meta-classifiers that are trained on model outputs or gradients. Similarly, Suri et al. [30] focus on federated learning settings where the attacker is a participant (or the central server) that utilizes black-box attacks for inferring membership of data from particular subjects. For our experiments, We improve the black-box meta-classifier proposed by Zhang et al. [41] using the \"query tuning\" technique in Xu et al. [37]." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.583 + ], + "angle": 0, + "content": "The closest works to ours are Chase et al. [7] and Chaudhari et al. [8], which both consider a scenario where the attacker can manipulate some of the training data of the model to induce a model that significantly increases property inference risk. These works assume an adversary with the ability to poison the victim's training data, while the adversary in our scenario has no access to the victim's training data, and therefore, their methods are not applicable. There are also works similar to ours that leverage \"adversarial initializations\" for attack purposes. Grosse et al. [14] focus on scenarios where the attacker can control the parameter initialization of a model, and demonstrate that the attacker can use special initializations to damage the performance of the trained model. Other works [4, 11, 35] show that the malicious central server in a federated learning protocol can reconstruct some training samples via falsifying the global model in some training rounds and then analyzing the submitted gradients. These kinds of attacks do not apply to our transfer-learning scenario since the attacker cannot access the downstream gradients, and can only manipulate the upstream training." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.602, + 0.64, + 0.618 + ], + "angle": 0, + "content": "3. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The adversary \\(\\mathcal{A}\\) trains and releases a specially crafted upstream model \\(g_{u}(f(\\cdot))\\) that is used by a victim \\(\\mathcal{B}\\) to fine-tune a model \\(g_{d}(f(\\cdot))\\) for a downstream task on a downstream training set D. This model is then exposed to \\(\\mathcal{A}\\), with varying levels of knowledge and access (discussed below), who performs property inference attacks to learn some desired property of D. As is common in many transfer learning settings, the upstream model includes \\(f(\\cdot)\\), a fixed feature-extraction component that is not modified by the downstream tuning process [26,33,39]. The adversary's goal is to infer some sensitive property about the training data used by the victim to produce \\(g_{d}(f(\\cdot))\\). For example, the adversary can release a general vision model (e.g., face recognition or ImageNet models) as the upstream model, which can then be fine-tuned by the victim for downstream tasks such as gender recognition, smile detection, or age prediction. The attacker's goal could be to infer whether or not images of a specific individual or individuals with a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15976" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "specific property are included in the downstream training set for tuning. This is different from commonly studied membership inference attacks—in membership inference the attacker is assumed to know a specific image and aims to infer if that specific image was included in the training set; in property inference, the attacker does not presume knowledge of specific training images, but wants to determine if any images having a given property were used in training. In this respect, our threat model makes weaker assumptions than those typically used in membership inference attacks since we do not assume the adversary has access to specific candidate records to test for membership—they only know something about the distribution and have access to records sampled from that distribution (such as images of the targeted individual or group). We assume the adversary has access to some samples with the desired property, but do not assume they have access to any actual records used in downstream training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.609 + ], + "angle": 0, + "content": "Attacker's Knowledge. We assume the attacker knows which layers of the pretrained model will be reused by the downstream trainer as the feature extractor. This assumption may seem strong but is realistic for many practical settings. Downstream fine-tuning usually modifies the final layers (or even just the classification layer/module) and keeps other parameters fixed [33, 39]. Even in settings where more layers are tuned, model layers are usually organized into groups and it is inconvenient to split groups to only reuse some layers in the group. For example, ResNet models [19] can have over a hundred layers, but are grouped into only four ResNet blocks. Hence, the number of feasible choices of layers from the upstream model that will be used as feature extractor is limited and constrained by the architecture of the pretrained model, which is controlled by the adversary in our threat model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.473, + 0.746 + ], + "angle": 0, + "content": "We consider three scenarios based on the level of access. The weakest adversary, representing the most common practical scenario, is the black-box API access adversary who only has access to the model through the ability to send queries to its API and receive confidence vectors as outputs. We assume the black-box adversary has knowledge of the model architecture, which is plausible since downstream training is highly likely to reuse the upstream network architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.748, + 0.472, + 0.794 + ], + "angle": 0, + "content": "We also consider two scenarios where the adversary has full access to the downstream model, with different assumptions about their knowledge on the downstream training:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.473, + 0.902 + ], + "angle": 0, + "content": "1. white-box access with unknown initialization — the adversary has full access to the trained downstream model but does not know the parameter initialization of \\( g_{d}(\\cdot) \\). This is fairly common in practice—for example, if \\( g_{d}(\\cdot) \\) contains only newly added task-specific classification modules/layers, the downstream trainer" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.828, + 0.108 + ], + "angle": 0, + "content": "will randomly initialize parameters for \\(g_{d}(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.118, + 0.895, + 0.27 + ], + "angle": 0, + "content": "2. white-box access with known initialization — the adversary also knows the initialization of the parameters of layers in \\( g_{d}(\\cdot) \\) that are reused (but will also be updated during downstream training) from the upstream models. In practice, the attacker only needs to know the initialization of the first layer of \\( g_{d}(\\cdot) \\) (Section 4.1). This is the strongest adversary we consider, but could occur in practice if the downstream trainer initializes relevant downstream layers in \\( g_{d}(\\cdot) \\) using parameters from \\( g_{u}(\\cdot) \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.283, + 0.783, + 0.3 + ], + "angle": 0, + "content": "4. Crafting the Pretrained Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.895, + 0.52 + ], + "angle": 0, + "content": "Our attack involves two phases: (1) training upstream models that are specially crafted to amplify property inference attacks, and (2) inferring properties of the dataset used to train a victim's downstream model using inference attacks. This section describes our method for producing the upstream models. Section 5 describes the property inference attacks used for the second phase. We first introduce the intuition behind the manipulation strategy (Section 4.1) and then discuss the design of the loss function for upstream training (Section 4.2). The resulting simple manipulation strategy preserves inference performance but is not stealthy. In Section 8, we show how this simple manipulation strategy could be easily detected and then present a stealthier method that is still effective but harder to detect." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.874, + 0.547 + ], + "angle": 0, + "content": "4.1. Embedding Property-Revealing Parameters" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.749 + ], + "angle": 0, + "content": "Our attack crafts a pretrained model such that there is a way to infer the desired property from the downstream model. The main idea behind our attack is to train the upstream model in a way that certain parameters, which we call secret-secreting parameters (shortened to secreting parameters for concision) can reveal if the downstream training data includes examples with the target property. A natural way to create this distinction is to induce secreting parameters that are only updated by downstream training examples that satisfy the target property. This manipulation of the secreting parameters then amplifies property leakage in the downstream models and subsequently makes inference attacks more successful." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Since convolutional and fully connected layers can be reduced to matrix multiplication operations, we can decompose the full downstream model as \\( g_{d}(f(\\pmb{x})) = h(\\phi(\\pmb{W} \\cdot f(\\pmb{x}) + \\pmb{b})) \\), where \\( \\pmb{W} \\) and \\( \\pmb{b} \\) are the parameters (weights and bias, respectively) associated with the first layer of \\( g_{d}(\\cdot) \\), \\( \\phi \\) is some activation function, and \\( h(\\cdot) \\) represents the rest of the layers of \\( g_{d}(\\cdot) \\). The upstream trainer can thus control updates for some of the parameters in \\( \\pmb{W} \\) by manipulating the outputs of \\( f(\\cdot) \\). We select part of the outputs of \\( f(\\cdot) \\) with a Boolean mask \\( \\pmb{m} \\) (i.e., \\( f(\\pmb{x}) \\circ \\pmb{m} \\)) and refer to them as secreting acti" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15977" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "vations. We denote parameters of \\(\\mathbf{W}\\) corresponding to the secreting activations as \\(\\mathbf{W}_t\\). The gradient for \\(\\mathbf{W}_t\\) is then (using the chain rule):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.147, + 0.469, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial l (\\boldsymbol {x} , \\mathcal {Y})}{\\partial \\boldsymbol {W} _ {t}} = \\frac {\\partial l (\\boldsymbol {x} , \\mathcal {Y})}{\\partial ((f (\\boldsymbol {x}) \\circ \\boldsymbol {m}) \\cdot \\boldsymbol {W} _ {t})} \\cdot \\frac {\\partial ((f (\\boldsymbol {x}) \\circ \\boldsymbol {m}) \\cdot \\boldsymbol {W} _ {t})}{\\partial \\boldsymbol {W} _ {t}} \\tag {1} \\\\ = \\frac {\\partial l (\\boldsymbol {x} , y)}{\\partial \\left(\\left(f (\\boldsymbol {x}) \\circ \\boldsymbol {m}\\right) \\cdot \\boldsymbol {W} _ {t}\\right)} \\cdot \\left(f (\\boldsymbol {x}) \\circ \\boldsymbol {m}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.223, + 0.469, + 0.284 + ], + "angle": 0, + "content": "where \\(l(\\pmb{x},\\pmb{y})\\) is the model loss for some input pair \\((\\pmb{x},\\pmb{y})\\), \\(f(\\pmb{x})\\circ \\pmb{m}\\) is the selected secreting activations for manipulation, and \\((f(\\pmb{x})\\circ \\pmb{m})\\cdot \\pmb{W}_t\\) denotes the computation related to the secreting activations in \\(g_{d}(\\cdot)\\)'s first layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.47, + 0.526 + ], + "angle": 0, + "content": "From Equation 1, if the secreting activations \\( f(\\pmb{x}) \\circ \\pmb{m} \\) are zero for some input \\( \\pmb{x} \\), gradients of the secreting parameters \\( W_{t} \\) will also be zeros. Thus, there will be no gradient updates on those parameters when trained on \\( \\pmb{x} \\). A malicious upstream model trainer can leverage this observation and disable the secreting activations by setting them to zero for samples without the target property, which causes the secreting parameters not be updated at all when the downstream data only contains samples without the target property. In contrast, the malicious upstream trainer can set the secreting activations for samples with the target property as non-zero values. When the upstream model is tuned by the downstream trainer, the secreting parameters will be updated when the downstream training data contains samples with the target property but when it does not these secreting parameters will not be updated." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.449, + 0.551 + ], + "angle": 0, + "content": "4.2. Upstream Optimization for Zero Activation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.469, + 0.619 + ], + "angle": 0, + "content": "We formulate the upstream model manipulation described in Section 4.1 into an optimization problem. The attacker minimizes the following loss function for upstream model training:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.632, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\nl (\\boldsymbol {x}, \\mathcal {Y}, \\mathcal {Y} _ {t}) = l _ {\\text {n o r m a l}} (\\boldsymbol {x}, \\mathcal {Y}) + l _ {t} (\\boldsymbol {x}, \\mathcal {Y} _ {t}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.469, + 0.735 + ], + "angle": 0, + "content": "where \\( l_{\\text{normal}} \\) is the loss for the original upstream training task (e.g., cross entropy loss) and \\( l_t \\) is the loss related to upstream model manipulation with \\( y_t \\) a binary label indicating whether the sample \\( x \\) contains the target property \\( (y_t = 1) \\). We define \\( l_t(x, y_t) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.746, + 0.469, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\alpha \\cdot \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\| & \\text {i f} y _ {t} = 0 \\\\ \\beta \\cdot \\max (\\lambda \\cdot \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 1 \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\( f(\\pmb{x}) \\circ \\neg \\pmb{m} \\) selects the non-secreting activations and \\( \\| \\cdot \\| \\) is used to measure the amplitude of the activations (can be some common norms such as \\( \\ell_1 \\) or \\( \\ell_2 \\) norms). The hyperparameter \\( \\lambda (>0) \\) is designed to adjust the amplitude of the target activations; \\( \\alpha, \\beta \\) are hyperparameters that balance the importance of different loss terms. The adversary then minimizes this loss over its training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "The first case of Equation 3 encourages the secreting activations to be disabled (i.e., 0) for samples without the target property \\((y_{t} = 0)\\). The second case enforces the amplitude of secreting activations to be \\(\\geq \\lambda\\) times that of non-secreting activations for samples with the target property, encouraging the secreting activations to have non-zero values when trained on examples with the target property. Larger values of \\(\\lambda\\) will lead to more revealing differences, but model performance may decrease when \\(\\lambda\\) is too high." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Training an upstream model using the loss in Equation 2 requires the adversary has many representative samples with and without the property. In Appendix A.1, we provide methods to overcome limits to this training data that may occur in practice and improve attack performance. Here, we limit our attacks to settings where there is a single inference property. Appendix A.11 describes a way to extend the attack to support multiple properties." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.364, + 0.682, + 0.38 + ], + "angle": 0, + "content": "5. Inference Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.48 + ], + "angle": 0, + "content": "In our threat model, the victim trains downstream models starting from manipulated upstream models (Section 4) on a private training dataset. In this section, we describe methods that use the induced downstream model to infer sensitive properties from the downstream training set for both the black-box and white-box attack scenarios from Section 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.491, + 0.706, + 0.506 + ], + "angle": 0, + "content": "5.1. Black-box API Access" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.891, + 0.56 + ], + "angle": 0, + "content": "We consider two black-box attack methods—one that directly uses model predictions, and one that leverages metaclassifiers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Confidence Score Test. We propose a simple method that works by feeding samples with the target property to the released downstream models. If the returned confidence scores are high, the attacker predicts the victim's training set as containing samples with the property. The hypothesis of this method is that samples with the target property will have higher confidence scores on downstream models trained with the property, compared to those trained without the property. The main idea of this approach has been previously explored in both property inference [29] and membership inference attacks [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Black-box Meta-classifier. We adapt the black-box meta-classifier proposed by Zhang et al. [41]. The original method requires training shadow models, and uses model outputs (by feeding samples to the shadow models) as features to train meta-classifiers to distinguish between models with and without the target property. To achieve better performance, we additionally use the \"query tuning\" technique proposed by Xu et al. [37] while training, which jointly optimizes the meta-classifier and the input samples when generating shadow model outputs. Figure 13 in the appendix shows the benefit of \"query tuning\"." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15978" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.252, + 0.106 + ], + "angle": 0, + "content": "5.2. White-Box Access" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.47, + 0.161 + ], + "angle": 0, + "content": "For adversaries with white-box access, there are two cases depending on if the attacker knows the initialization of the parameters of newly added downstream layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.166, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Parameter Difference Test (known initialization). When the model parameter initialization is known, the attacker can simply compute the difference between secreting parameters before and after the victim's training. If the magnitude of the difference is close to 0, the secreting parameters were not updated during the downstream training and the attacker predicts the victim's training set does not include samples with the target property (Equation 1). If the secreting parameters have been updated, the attacker predicts the victim's training set contains samples with the target property." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.519 + ], + "angle": 0, + "content": "Variance Test (unknown initialization). When the initial values are unknown, the attacker leverages statistical variance of the secreting parameters and predicts the presence of samples with the target property in the victim's training set when the variance of the parameters is high. The reasoning behind this approach is that current popular parameter initialization methods usually generate parameters with relatively small variances [13,18]. If the victim's data contains samples with the target property, the secreting parameters would be updated with gradients of relatively large values (controlled by \\(\\lambda\\) in Equation 3), and increase the variance of those parameters in the final model. We confirm this hypothesis empirically in Section 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.812 + ], + "angle": 0, + "content": "White-Box Meta-Classifier. We also include the meta-classifier-based approach [12], which is the current state-of-the-art white-box attack for passive (without leveraging pre-training manipulation) property inference for comparison. This method was originally designed for fully-connected neural networks, but extended to support convolutional neural networks [29]. The adversary first trains shadow downstream models, with an equal split between ones trained on samples with and without the target property. Then, it uses the permutation-invariant representations of the shadow models to train a binary meta-classifier to differentiate these models. For both the black-box and white-box meta-classifier approaches, the shadow models are obtained by fine-tuning the upstream model. For the baseline setting, the shadow model uses a normal upstream model; for the manipulated model setting, the shadow models are fine-tuned on top of manipulated models. Therefore, attacks in the latter setting may gain some advantage from manipulation compared to attacks in the former setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.829, + 0.279, + 0.846 + ], + "angle": 0, + "content": "6. Experimental Design" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "This section explains our experimental setup. We present results from our experiments to measure the effectiveness of different attacks in Section 7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Tasks and Models. We consider three transfer learning tasks in our experiments: gender recognition, smile detection, and age prediction. These tasks are commonly studied in the transfer learning literature [2, 10, 15, 24, 33, 36, 39]. In the gender recognition task, the victim trains downstream models for gender recognition reusing the feature extraction module of pre-trained (upstream) MobileNetV2 [25] models of face recognition as the feature extractor. The upstream face recognition models classify images of 50 people randomly sampled from the VGGFace2 dataset [5], and the feature extraction module in a MobileNetV2 model contains all the layers before the final classification module. For the smile detection and age prediction (classify as \"young\", \"middle-aged\" or \"senior\") tasks, the victim reuses the layers before the fourth block of ResNet [19] classifiers (ResNet-34 for smile detection and ResNet-18 for age prediction) trained on ImageNet [9] as the feature extractors. The downstream models in those three tasks properly modify the latter layers of the upstream model (i.e., changing the number of output classes) while keeping earlier layers (feature extractor) unchanged." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Upstream and Downstream Training. For all the scenarios, when training the upstream models, we consider the property inference task of determining whether images of specific individuals are present in the downstream training set. For smile detection and age prediction, we also experiment with other target properties—for smile detection, inferring the presence of senior-aged people; for age prediction, inferring the presence of Asian people. Appendix A.2 provides more details about the upstream training," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.747 + ], + "angle": 0, + "content": "We conduct the downstream training on VGGFace2 with the attribute labels provided by MAADFace [31, 32]. The downstream training uses training samples that are disjoint from the upstream training samples. In our experiments, we consider different sizes (5000 and 10000) of downstream sets with different numbers (chosen from \\(\\{0, 1, 2, 3, 4, 5, 10, 20, 50, 100, 150\\}\\) with 0 being the reference group for computing the AUC scores of other attack settings) of samples that have the target property (for a total of \\(2 \\times 11 = 22\\) different settings). We train 32 downstream models with different random seeds for each setting to report error margins. Appendix A.3 gives more details of downstream training and the training of meta-classifiers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.813 + ], + "angle": 0, + "content": "Attack Evaluation Metric. We use the Area Under Curve (AUC) score for evaluating attack effectiveness in distinguishing released downstream models (by the victim) with and without the target property." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.812, + 0.845 + ], + "angle": 0, + "content": "7. Evaluation of Attack Effectiveness" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Figure 1 summarizes our results. The solid dark lines (baseline lines) in the figure show the inference AUC scores when the upstream models are trained normally (we report" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15979" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.158, + 0.093, + 0.805, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Figure 1. Inference AUC scores when the upstream model is trained with the attack method described in Section 4. Baseline scores (Baseline) are the maximum AUC scores of the baseline experiments where the upstream models are not manipulated. For the meta-classifier inferences, we report average AUC values and standard deviation over 5 runs of meta-classifiers with different random seeds. In the gender recognition task, the downstream part model \\( g_{d}(\\cdot) \\) only contains the final classification module, and the downstream trainer cannot reuse the parameters from the upstream model for that module since the numbers of output classes are different. Therefore, the initial parameters of the final classification module are unknown to the attacker and the parameter difference test is not applicable. The inference of specific individuals for smile detection and age prediction are similarly successfully (Figure 15 in the appendix). The downstream training sets contain 10 000 samples and inference results of 5 000 samples are similar and given in Figure 12 in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.435 + ], + "angle": 0, + "content": "the best results of all tested attacks). More details of the baseline experiments can be found in Appendix A.4. Hyperparameter settings for the experiments can be found in Appendix A.5 and the results are insensitive to the selection to hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.471, + 0.771 + ], + "angle": 0, + "content": "In all settings except the age prediction with 150 samples of target property, the AUC scores are less than 0.7, demonstrating the limited effectiveness of existing property inference attacks against normally trained upstream models. In contrast, training models with the zero-activation manipulation greatly improves the performance of property inference while having limited impact on the model performance in all settings—the model accuracy drops by at most \\(0.9\\%\\) (see Appendix A.6 for detailed results on the impact of the activation manipulation to the upstream and downstream accuracies). Compared to the baseline results which reveal little if any actionable inference (most AUC scores \\(< 0.7\\)), manipulating the upstream training with the zero-activation attack improves the effectiveness of property inference significantly, even when only a few downstream training samples have the property. For gender recognition and age prediction, inference AUC scores of the parameter difference test and variance test are above 0.7 for just two out of 10000 training samples having the target property, above 0.9 for 10 training samples, and exceed 0.95 for \\(\\geq 20\\) training samples. The one exception also has AUC scores exceed 0.9 for \\(\\geq 20\\) training samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.47, + 0.882 + ], + "angle": 0, + "content": "Black-box attacks. The black-box meta-classifier achieves inference AUC scores above 0.9 when \\(\\geq 50\\) out of 10000 training samples have the target property. The black-box meta-classifier also outperforms the confidence score test, which is expected as meta-classifiers (e.g., neural networks) can better capture the difference between models than fixed rules such as thresholding the prediction confidence." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "White-box attacks. Our white-box methods (the parame" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.895, + 0.496 + ], + "angle": 0, + "content": "ter difference test and the variance test) also achieve AUC scores \\(>0.9\\) when \\(\\geq 20\\) training samples are with the target property. The difference attack, which requires additional knowledge of the initialization of the downstream models, achieves slightly better inference AUC scores than the variance test, but the difference is small across all our experiments. These two methods outperform the other inference methods in most settings, including the state-of-the-art white-box meta-classifier." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.895, + 0.683 + ], + "angle": 0, + "content": "White-box meta-classifier vs. Black-box meta-classifier. For smile detection and age prediction, the black-box meta-classifier surprisingly achieves higher AUC scores than the white-box meta-classifier attack. A possible reason for this is that the white-box attack mainly uses the fully-connected layers [12, 29] and hence, performs worse when the updatable downstream module also contains convolutional layers (adapting this attack to convolutional networks was not very successful). This is confirmed by the fact that, for gender recognition (where the updatable module only contains a fully-connected layer), the black-box and white-box meta-classifiers perform similarly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Attacks of AUC scores \\( < 0.5 \\). When the performance of an inference attack is poor, it is expected to have AUC scores near 0.5 (close to random guessing). However, we find that there are few attack settings with AUC scores consistently below 0.5. Appendix A.10 discusses those anomalies and surmises that they are caused by the limitations of original inference methods designed for normal pretrained models when facing challenging inference tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.727, + 0.846 + ], + "angle": 0, + "content": "8. Stealthier Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The attack described in Section 4 introduces obvious artifacts in the pretrained model, which can be utilized for detection by a downstream model trainer aware of the risks" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15980" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "posed by our attacks. We first present two detection methods (Section 8.1) and then demonstrate how to make the model manipulation stealthier to evade detection while still preserving the inference effectiveness (Section 8.2 and Section 8.3). We assume the downstream trainer is aware of the possibility of the attack and its design, but does not know the property targeted by the adversary, as this is specific to an attacker's goal and the set of possible properties can be exponentially large for a rich training set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.236, + 0.438, + 0.254 + ], + "angle": 0, + "content": "8.1. Detecting Manipulated Pretrained Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.259, + 0.47, + 0.306 + ], + "angle": 0, + "content": "We present two detection methods that use the distributional difference between activations of samples with and without property." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.473, + 0.536 + ], + "angle": 0, + "content": "Checking the Distribution of Activations. Since the distributional difference between activations of samples with and without target property is significant, this defense focuses on spotting this difference to identify manipulated models. A method to identify the distributional difference needs to be designed based on the attack method used. For the original zero-activation attacks in Section 4.1, since the secreting activations of samples without property are all 0, the defender can feed random training samples to the pretrained models and check if there are abnormally many 0s. This approach is feasible since samples of target property have limited presence in the downstream training set and hence, most samples will not have the property. Since detecting the zero-activation attack is trivial using this method, we do not conduct any experiments with this." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.473, + 0.811 + ], + "angle": 0, + "content": "Anomaly Detection. Since the target property has a limited presence in the downstream training set, another defense would be treating samples with the target property as outliers and then analyzing those outliers to find manipulations. Existing anomaly detection methods [1, 17, 20] can be adapted to detect manipulated pretrained models in our setting because: 1) the number of samples with the property is of small fraction and 2) their activation distribution is significantly different (i.e., outliers) from the distribution for samples without the property. The auditor can inspect model activations for all of its training data and identify outliers (ideally, samples with target property) with anomaly detection. The auditor can then inspect identified outliers and may find commonalities to identify the potential target property. For instance, they may find that a small fraction of the training data produce unusual model activations, and then notice that most of that data has a particular property such as belonging to a specific individual or group." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.903 + ], + "angle": 0, + "content": "We consider three common anomaly detection methods: K-means [20], PCA [1] and Spectre [17] (where Spectre is the current state-of-the-art) and we report the detection results from the three defenses. Appendix A.12 gives details of these methods. The detection results on the zero-activation attack are given in Figure 11 in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "Anomaly detection is very effective at identifying the samples with target property. For example, for the gender recognition and smile detection tasks, the detection rate is over \\(80\\%\\) in most cases. These results motivate the design of stealthier attacks which we describe next." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.178, + 0.776, + 0.196 + ], + "angle": 0, + "content": "8.2. Stealthier Model Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.202, + 0.895, + 0.383 + ], + "angle": 0, + "content": "To evade the defense that checks the distribution of activations, we modify our zero-activation attack to ensure: (1) secreting activations for samples without the property are also non-zero (bypassing simple defense of checking abnormal zeros); (2) secreting activations of samples with and without target property are still distinct (the attack is still effective); (3) that distinction between activations should not be captured by anomaly detection methods (evading anomaly detection); (4) the actual distribution of activations that matches the attacker's goal cannot be easily guessed by the defender (handling cases when the defender actively searches other patterns in the distribution of activations)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.384, + 0.852, + 0.4 + ], + "angle": 0, + "content": "For (1) and (2), we adapt the loss in Equation 3 as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.404, + 0.892, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\alpha \\cdot \\max (\\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 0 \\\\ \\beta \\cdot \\max (\\lambda \\cdot \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 1 \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.735 + ], + "angle": 0, + "content": "where \\(\\lambda \\geq 1\\). (1): The case of \\(y_{t} = 0\\) is redefined to bypass the detection of abnormal zeros. Minimizing this new loss ensures that samples without the target property will have secreting activations \\((f(\\pmb{x}) \\circ \\pmb{m})\\) with (close-to-normal) nonzero values. (2): to ensure the property is still detectable, we actively increase the difference between the secreting activations of samples with and without property. We observe that, for upstream models with reasonable performance on the main task, non-secreting activations \\((f(\\pmb{x}) \\circ \\neg \\pmb{m})\\) have similar amplitude regardless of the fed samples containing target property. Therefore, for samples with target property, as long as we ensure the secreting activations have a larger amplitude than that of non-secreting activations, there will be a distinction between secreting activations of samples with and without property. We do this by assigning larger values to \\(\\lambda\\) (e.g., \\(\\lambda \\geq 1\\), instead of the original \\(\\lambda > 0\\)) for the second line of Equation 4 to induce sharper distinction between samples with and without property and enable higher inference performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To prevent detection by anomaly detectors (requirement (3) above), \\(\\lambda\\) should be set to balance the attack effectiveness and stealthiness rightly. By choosing proper values for \\(\\lambda\\), our attack is able to evade anomaly detection methods in most settings. However, in some settings (mostly in gender recognition tasks), state-of-the-art anomaly detection (Spectre) can still identify most of the samples with target property. To counter this, we add an additional regularization term (weighted by parameter \\(\\gamma\\)) to the overall loss function \\(l(x,y,y_{t})\\) in Equation 2 that further improves attack stealthiness while still maintaining relatively high at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15981" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.158, + 0.093, + 0.808, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.895, + 0.308 + ], + "angle": 0, + "content": "Figure 2. Inference AUC scores of the stealthier design. Since the secreting activations are no longer zero, the inference methods based on difference or variance tests are no longer applicable. The inference results of specific individuals for smile detection and age prediction also show similar improvement compared to the baseline settings (Figure 19 in the appendix). The downstream training sets contain 10000 samples and inference results results of 5000 samples are similar and given in Figure 17 in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.473, + 0.619 + ], + "angle": 0, + "content": "tack effectiveness. Specifically, we first obtain the corresponding covariance matrices of the activations of samples with the target property \\((cov_w)\\), activations of all samples with and without the target property \\((cov_{w,wo})\\), and activations of samples without the target property \\((cov_{wo})\\) respectively. Then, we encourage mean \\((cov_w) = mean(cov_{w,wo}) = mean(cov_{wo})\\) and \\(var(cov_w) = var(cov_{w,wo}) = var(cov_{wo})\\) (both mean \\((\\cdot)\\) and \\(var(\\cdot)\\) treat the whole covariance matrix as a flattened array and return scalar values) for the three covariance matrices by minimizing their differences in their mean and variance. Using this method, we ensure the distributions of activations of samples with target property will be similar to the ones without the property, making the manipulations harder to detect. We use this approach for all the experiments. To ensure the distributional pattern related to the attacker goal cannot be easily guessed (requirement (4)), we generate \\(m\\) randomly (instead of picking first \\(\\| m\\|\\) activations in Section 7). This makes the brute-force search of possible patterns computationally infeasible (details in Appendix A.14)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.628, + 0.382, + 0.645 + ], + "angle": 0, + "content": "8.3. Experiments with Stealthy Attacks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Detection Evasion. Figure 16 (in the appendix) summarizes the results of our experiments to detect the stealthy upstream models (Appendix A.13 provides details on these experiments). We find that the anomaly detection methods are ineffective against our stealthier attack—< \\(10\\%\\) of samples with the target property are detected across all settings with the exception of a detection rate \\(< 20\\%\\) (still low) for smile detection when the total number of samples is 5000 and 100 or 150 of them are with the target property. We also made several attempts to approximately identify (instead of brute-force search) possible attack patterns in the activations but none of these succeeded in uncovering the stealthy attacks (details are in Appendix A.14)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Inference Results. From Figure 2, we can see that activation manipulation still leads to significantly improved inference results compared to the baselines with normally" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.895, + 0.543 + ], + "angle": 0, + "content": "trained upstream models. For example, for gender recognition, when \\(\\geq 50\\) downstream training samples have the target property, inference AUC scores exceed 0.95, which is a huge improvement compared to the baseline attack where all AUC scores are less than 0.6, and similar trends follow for smile detection (with over 100 samples with property, AUC improves from \\(< 0.6\\) to \\(> 0.78\\)) and age prediction (with over 100 samples with property, AUC improves from \\(< 0.77\\) to \\(> 0.9\\)). Comparing the results for the stealthier attacks to the results that do not consider defenses in Figure 1, we observe that the attack effectiveness declines as expected since we are now trading-off attack effectiveness for stealthiness. Training models with the attack goal poses negligible impact on the model performance (accuracy drop \\(< 0.9\\%\\), see Appendix A.6)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.62, + 0.577 + ], + "angle": 0, + "content": "9. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.895, + 0.754 + ], + "angle": 0, + "content": "Our work demonstrates how a malicious upstream trainer can manipulate its training process to amplify property inference risks for downstream models when transfer learning is done. Our empirical results show that such manipulations can be exploited to enable very precise property inference, even in black-box settings, across a variety of tasks. Although there is potential for a new arms race between methods of hiding manipulations and methods of detecting them, the larger lesson from this work, and other works exposing similar risks, is that it is important that users of pretrained models to only use models from trusted providers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.669, + 0.789 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (#2022YFF0604503 and #2021YFB3100300), the United States National Science Foundation through the Center for Trustworthy Machine Learning (#1804603), NSFC (#62272224), JiangSu Province Science Foundation for Youths (#BK20220772), and Lockheed Martin Corporation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15982" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Hervé Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 7, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.472, + 0.225 + ], + "angle": 0, + "content": "[2] MAH Akhand, Ijaj Sayim, Shuvendu Roy, N Siddique, et al. Human Age Prediction from Facial Image Using Transfer Learning in Deep Convolutional Neural Networks. In International Joint Conference on Computational Intelligence, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.31 + ], + "angle": 0, + "content": "[3] Giuseppe Ateniese, Luigi V Mancini, Angelo Spognardi, Antonio Villani, Domenico Vitali, and Giovanni Felici. Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifiers. International Journal of Security and Networks, 10(3):137-150, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[4] Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, and Nicolas Papernot. When the curious abandon honesty: Federated learning is not private. arXiv preprint arXiv:2112.02918, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. In IEEE International Conference on Automatic Face & Gesture Recognition, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.491 + ], + "angle": 0, + "content": "[6] Shuvam Chakraborty, Burak Uzkent, Kumar Ayush, Kumar Tanmay, Evan Sheehan, and Stefano Ermon. Efficient Conditional Pre-training for Transfer Learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[7] Melissa Chase, Esha Ghosh, and Saeed Mahloujifar. Property Inference from Poisoning. In IEEE Symposium on Security and Privacy, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.591 + ], + "angle": 0, + "content": "[8] Harsh Chaudhari, Jackson Abascal, Alina Oprea, Matthew Jagielski, Florian Tramèr, and Jonathan Ullman. SNAP: Efficient Extraction of Private Properties with Poisoning. arXiv preprint arXiv:2208.12348, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009. 2, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.649, + 0.468, + 0.689 + ], + "angle": 0, + "content": "[10] Fadi Dornaika, Ignacio Arganda-Carreras, and C Belver. Age estimation in facial images through transfer learning. Machine Vision and Applications, 30(1):177-187, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[11] Liam Fowl, Jonas Geiping, Wojtek Czaja, Micah Goldblum, and Tom Goldstein. Robbing the fed: Directly obtaining private data in federated learning with modified models. arXiv preprint arXiv:2110.13057, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[12] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property Inference Attacks on Fully Connected Neural Networks using Permutation Invariant Representations. In ACM SIGSAC Conference on Computer and Communications Security, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Kathrin Grosse, Thomas A Trost, Marius Mosbach, and Michael Backes. Adversarial initialization-when your" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "network performs the way I want. arXiv preprint arxiv:1902.03020, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[15] Xin Guo, Luisa Polania, and Kenneth Barner. Smile Detection in the Wild Based on Transfer Learning. In IEEE International Conference on Automatic Face & Gesture Recognition, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[16] Valentin Hartmann, L'eo Meynent, Maxime Peyrard, Dimitrios Dimitriadis, Shruti Tople, and Robert West. Distribution inference risks: Identifying and mitigating sources of leakage. arXiv preprint arXiv:2209.08541, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.239, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[17] Jonathan Hayase, Weihao Kong, Raghav Somani, and Sewoong Oh. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics. In International Conference on Machine Learning, 2021. 7, 19, 20, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.297, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. In IEEE International Conference on Computer Vision, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.354, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[20] Anil K Jain, M Narasimha Murty, and Patrick J Flynn. Data Clustering: A Review. ACM computing surveys (CSUR), 31(3):264-323, 1999. 7, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[21] Joanna Jaworek-Korjakowska, Pawel Kleczek, and Marek Gorgon. *Melanoma Thickness Prediction Based on Convolutional Neural Network with VGG-19 Model Transfer Learning*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.528, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[22] Marc Juárez, Samuel Yeom, and Matt Fredrikson. BlackBox Audits for Group Distribution Shifts. arXiv preprint arXiv:2209.03620, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[23] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting Unintended Feature Leakage in Collaborative Learning. In IEEE Symposium on Security and Privacy, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.629, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[24] Cao Hong Nga, Khai-Thinh Nguyen, Nghi C Tran, and Jia-Ching Wang. Transfer Learning for Gender and Age Prediction. In IEEE International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[25] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetV2: Inverted Residuals and Linear Bottlenecks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[26] Roei Schuster, Tal Schuster, Yoav Meri, and Vitaly Shmatikov. Humpty Dumpty: Controlling Word Meanings via Corpus Poisoning. In IEEE Symposium on Security and Privacy, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Ozan Sener and Vladlen Koltun. Multi-Task Learning as Multi-Objective Optimization. In Advances in Neural Information Processing Systems, 2018. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15983" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[29] Anshuman Suri and David Evans. Formalizing and Estimating Distribution Inference Risks. In Privacy Enhancing Technologies Symposium, 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[30] Anshuman Suri, Pallika Kanani, Virendra J Marathe, and Daniel W Peterson. Subject Membership Inference Attacks in Federated Learning. arXiv preprint arXiv:2206.03317, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[31] Philipp Terhörst, Daniel Fährmann, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. MAAD-Face: A Massively Annotated Attribute Dataset for Face Images. IEEE Transactions on Information Forensics and Security, 16:3942-3957, 2021. 5, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[32] Philipp Terhörst, Marco Huber, Jan Niklas Kolf, Ines Zelch, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Reliable Age and Gender Estimation from Face Images: Stating the Confidence of Model Predictions. In IEEE International Conference on Biometrics Theory, Applications and Systems (BTAS), 2019. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[33] Bolun Wang, Yuanshun Yao, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. With Great Training Comes Great Vulnerability: Practical Attacks against Transfer Learning. In USENIX Security Symposium, 2018. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[34] Xiuling Wang and Wendy Hui Wang. Group Property Inference Attacks Against Graph Neural Networks. arXiv preprint arXiv:2209.01100, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[35] Yuxin Wen, Jonas A Geiping, Liam Fowl, Micah Goldblum, and Tom Goldstein. Fishing for user data in large-batch federated learning via gradient magnification. In International Conference on Machine Learning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[36] Yu Xia, Di Huang, and Yunhong Wang. Detecting Smiles of Young Children via Deep Transfer Learning. In IEEE International Conference on Computer Vision Workshops, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[37] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting AI Trojans Using Meta Neural Analysis. In IEEE Symposium on Security and Privacy, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[38] Kaiyu Yang, Jacqueline Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky. A Study of Face Obfuscation in ImageNet. In International Conference on Machine Learning, 2022. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[39] Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y Zhao. Latent Backdoor Attacks on Deep Neural Networks. In ACM SIGSAC Conference on Computer and Communications Security, 2019. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond Empirical Risk Minimization. In International Conference on Learning Representations, 2018. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[41] Wanrong Zhang, Shruti Tople, and Olga Ohrimenko. Leakage of Dataset Properties in Multi-Party Machine Learning. In USENIX Security Symposium, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[42] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A Comprehensive Survey on Transfer Learning. Proceedings of the IEEE, 109, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[43] Yang Zou, Zhikun Zhang, Michael Backes, and Yang Zhang. Privacy Analysis of Deep Learning in the Wild: Membership" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Inference Attacks against Transfer Learning. arXiv preprint arXiv:2009.04872, 2020. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15984" + } + ] +] \ No newline at end of file diff --git a/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_origin.pdf b/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a2a601968537027c21c90a46fb981a3ed2256d5f --- /dev/null +++ b/2023/Manipulating Transfer Learning for Property Inference/30a4d03f-7b0e-4005-b9ca-7776bedd6658_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b926130f4d83c97b43fbecfaa86f0ccab65831243f7806d84563cf5fca24b078 +size 221111 diff --git a/2023/Manipulating Transfer Learning for Property Inference/full.md b/2023/Manipulating Transfer Learning for Property Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0ee6a40b20721a775c539e48b87c252c8639e379 --- /dev/null +++ b/2023/Manipulating Transfer Learning for Property Inference/full.md @@ -0,0 +1,249 @@ +# Manipulating Transfer Learning for Property Inference + +Yulong Tian $^{1}$ , Fnu Suya $^{2}$ , Anshuman Suri $^{2}$ , Fengyuan Xu $^{1*}$ , David Evans $^{2}$ $^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, China + $^{2}$ University of Virginia, USA + +yulong.tian@smail.nju.edu.cn, {suya, anshuman}@virginia.edu, fengyuan.xu@nju.edu.cn, evans@virginia.edu + +# Abstract + +Transfer learning is a popular method for tuning pretrained (upstream) models for different downstream tasks using limited data and computational resources. We study how an adversary with control over an upstream model used in transfer learning can conduct property inference attacks on a victim's tuned downstream model. For example, to infer the presence of images of a specific individual in the downstream training set. We demonstrate attacks in which an adversary can manipulate the upstream model to conduct highly effective and specific property inference attacks (AUC score $>0.9$ ), without incurring significant performance loss on the main task. The main idea of the manipulation is to make the upstream model generate activations (intermediate features) with different distributions for samples with and without a target property, thus enabling the adversary to distinguish easily between downstream models trained with and without training examples that have the target property. Our code is available at https://github.com/yulongt23/Transfer-Inference. + +# 1. Introduction + +Transfer learning is a popular method for efficiently training deep learning models [6, 21, 33, 39, 42]. In a typical transfer learning scenario, an upstream trainer trains and releases a pretrained model. Then a downstream trainer will reuse the parameters of some layers of the released upstream models to tune a downstream model for a particular task. This parameter reuse reduces the amount of data and computing resources required for training downstream models significantly, making this technique increasingly popular. However, the centralized nature of transfer learning is open to exploitation by an adversary. Several previous works have considered security risks associated with transfer learning including backdoor attacks [39] and misclassification attacks [33]. + +We investigate the risk of property inference in the context of transfer learning. In property inference (also known as distribution inference), the attacker aims to extract sensitive properties of the training distribution of a model [3,7,12,29,41]. We consider a transfer learning scenario where the upstream trainer is malicious and produces a carefully crafted pretrained model with the goal of inferring a particular property about the tuning data used by the victim to train a downstream model. For example, the attacker may be interested in knowing whether any images of a specific individual (or group, such as seniors or Asians) are contained in a downstream training set used to tune the pre-trained model. Such inferences can lead to severe privacy leakage—for instance, if the adversary knows beforehand that the downstream training set consists of data of patients that have a particular disease, confirming the presence of a specific individual in that training data is a privacy violation. Property inference may also be used to audit models for fairness issues [22]—for example, in a downstream dataset containing data of all the employees of an organization, finding the absence of samples of a certain group of people (e.g., older people) may be evidence that those people are underrepresented in that organization. + +Contributions. We identify a new vulnerability of transfer learning where the upstream trainer crafts a pretrained model to enable an inference attack on the downstream model that reveals very precise and accurate information about the downstream training data (Section 3). We develop methods to manipulate the upstream model training to produce a model that, when used to train a downstream model, will induce a downstream model that reveals sensitive properties of its training data in both white-box and black-box inference settings (Section 4). We demonstrate that this substantially increases property inference risk compared to baseline settings where the upstream model is trained normally (Section 7). Table 1 summarizes our key results. The inference AUC scores are below 0.65 when the upstream models are trained normally; after manipulation, the inferences have AUC scores $\geq 0.89$ even when only $0.1\%$ (10 out of 10 000) of downstream samples have the target prop + +
Downstream TaskUpstream TaskTarget PropertyNormal Upstream ModelManipulated Upstream Model
0.1% (10)1% (100)0.1% (10)1% (100)
Gender RecognitionFace Recognition0.490.520.961.0
Smile DetectionImageNet Classification [9]Specific Individuals0.500.501.01.0
Age PredictionImageNet Classification [9]0.540.630.971.0
Smile DetectionImageNet Classification [9]Senior0.590.560.891.0
Age PredictionImageNet Classification [9]Asian0.490.650.951.0
+ +Table 1. Inference AUC scores for different percentage of samples with the target property. Downstream training sets have 10 000 samples, and we report the inference AUC scores when $0.1\%$ (10) and $1\%$ (100) samples in the downstream set have the target property. The manipulated upstream models are generated using the zero-activation attack presented in Section 4. + +erty and achieve perfect results (AUC score $= 1.0$ ) when the ratio increases to $1\%$ . The manipulated models have negligible performance drops $(< 0.9\%)$ on their intended tasks. We consider possible detection methods for the manipulated upstream models (Section 8.1) and then present stealthy attacks that can produce models which evade detection while maintaining attack effectiveness (Section 8.2). + +# 2. Related Work + +Several works have demonstrated risks associated with transfer learning across a variety of attack goals. Wang et al. [33] and Yao et al. [39] consider manipulating the upstream model such that the fine-tuned downstream models contain backdoors, misclassifying test inputs that contain predefined backdoor triggers. These transfer manipulations are tailored to their particular attack goals and cannot be applied for the property inference goal considered in this paper. Zou et al. [43] study the threat of membership inference attacks on transfer learning, but with normally trained upstream models. + +The risk of property inference was introduced by Ateniese et al. [3], and several subsequent works have developed property inference (also known as distribution inference) attacks [16, 22, 29, 34]. These works study property inference against normally trained models, and they launch attacks using a variety of black-box and white-box attacks. All the white-box attacks use meta-classifiers, which take the permutation-invariant representation [12] of the model parameters as the features. We use the state-of-the-art white-box attack [29] in our experiments. Melis et al. [23] and Zhang et al. [41] focus on property inference in distributed training scenarios. In their settings, the attacker is a participant in the global model training and conducts property inference using meta-classifiers that are trained on model outputs or gradients. Similarly, Suri et al. [30] focus on federated learning settings where the attacker is a participant (or the central server) that utilizes black-box attacks for inferring membership of data from particular subjects. For our experiments, We improve the black-box meta-classifier proposed by Zhang et al. [41] using the "query tuning" technique in Xu et al. [37]. + +The closest works to ours are Chase et al. [7] and Chaudhari et al. [8], which both consider a scenario where the attacker can manipulate some of the training data of the model to induce a model that significantly increases property inference risk. These works assume an adversary with the ability to poison the victim's training data, while the adversary in our scenario has no access to the victim's training data, and therefore, their methods are not applicable. There are also works similar to ours that leverage "adversarial initializations" for attack purposes. Grosse et al. [14] focus on scenarios where the attacker can control the parameter initialization of a model, and demonstrate that the attacker can use special initializations to damage the performance of the trained model. Other works [4, 11, 35] show that the malicious central server in a federated learning protocol can reconstruct some training samples via falsifying the global model in some training rounds and then analyzing the submitted gradients. These kinds of attacks do not apply to our transfer-learning scenario since the attacker cannot access the downstream gradients, and can only manipulate the upstream training. + +# 3. Threat Model + +The adversary $\mathcal{A}$ trains and releases a specially crafted upstream model $g_{u}(f(\cdot))$ that is used by a victim $\mathcal{B}$ to fine-tune a model $g_{d}(f(\cdot))$ for a downstream task on a downstream training set D. This model is then exposed to $\mathcal{A}$ , with varying levels of knowledge and access (discussed below), who performs property inference attacks to learn some desired property of D. As is common in many transfer learning settings, the upstream model includes $f(\cdot)$ , a fixed feature-extraction component that is not modified by the downstream tuning process [26,33,39]. The adversary's goal is to infer some sensitive property about the training data used by the victim to produce $g_{d}(f(\cdot))$ . For example, the adversary can release a general vision model (e.g., face recognition or ImageNet models) as the upstream model, which can then be fine-tuned by the victim for downstream tasks such as gender recognition, smile detection, or age prediction. The attacker's goal could be to infer whether or not images of a specific individual or individuals with a + +specific property are included in the downstream training set for tuning. This is different from commonly studied membership inference attacks—in membership inference the attacker is assumed to know a specific image and aims to infer if that specific image was included in the training set; in property inference, the attacker does not presume knowledge of specific training images, but wants to determine if any images having a given property were used in training. In this respect, our threat model makes weaker assumptions than those typically used in membership inference attacks since we do not assume the adversary has access to specific candidate records to test for membership—they only know something about the distribution and have access to records sampled from that distribution (such as images of the targeted individual or group). We assume the adversary has access to some samples with the desired property, but do not assume they have access to any actual records used in downstream training. + +Attacker's Knowledge. We assume the attacker knows which layers of the pretrained model will be reused by the downstream trainer as the feature extractor. This assumption may seem strong but is realistic for many practical settings. Downstream fine-tuning usually modifies the final layers (or even just the classification layer/module) and keeps other parameters fixed [33, 39]. Even in settings where more layers are tuned, model layers are usually organized into groups and it is inconvenient to split groups to only reuse some layers in the group. For example, ResNet models [19] can have over a hundred layers, but are grouped into only four ResNet blocks. Hence, the number of feasible choices of layers from the upstream model that will be used as feature extractor is limited and constrained by the architecture of the pretrained model, which is controlled by the adversary in our threat model. + +We consider three scenarios based on the level of access. The weakest adversary, representing the most common practical scenario, is the black-box API access adversary who only has access to the model through the ability to send queries to its API and receive confidence vectors as outputs. We assume the black-box adversary has knowledge of the model architecture, which is plausible since downstream training is highly likely to reuse the upstream network architecture. + +We also consider two scenarios where the adversary has full access to the downstream model, with different assumptions about their knowledge on the downstream training: + +1. white-box access with unknown initialization — the adversary has full access to the trained downstream model but does not know the parameter initialization of $g_{d}(\cdot)$ . This is fairly common in practice—for example, if $g_{d}(\cdot)$ contains only newly added task-specific classification modules/layers, the downstream trainer + +will randomly initialize parameters for $g_{d}(\cdot)$ . + +2. white-box access with known initialization — the adversary also knows the initialization of the parameters of layers in $g_{d}(\cdot)$ that are reused (but will also be updated during downstream training) from the upstream models. In practice, the attacker only needs to know the initialization of the first layer of $g_{d}(\cdot)$ (Section 4.1). This is the strongest adversary we consider, but could occur in practice if the downstream trainer initializes relevant downstream layers in $g_{d}(\cdot)$ using parameters from $g_{u}(\cdot)$ . + +# 4. Crafting the Pretrained Model + +Our attack involves two phases: (1) training upstream models that are specially crafted to amplify property inference attacks, and (2) inferring properties of the dataset used to train a victim's downstream model using inference attacks. This section describes our method for producing the upstream models. Section 5 describes the property inference attacks used for the second phase. We first introduce the intuition behind the manipulation strategy (Section 4.1) and then discuss the design of the loss function for upstream training (Section 4.2). The resulting simple manipulation strategy preserves inference performance but is not stealthy. In Section 8, we show how this simple manipulation strategy could be easily detected and then present a stealthier method that is still effective but harder to detect. + +# 4.1. Embedding Property-Revealing Parameters + +Our attack crafts a pretrained model such that there is a way to infer the desired property from the downstream model. The main idea behind our attack is to train the upstream model in a way that certain parameters, which we call secret-secreting parameters (shortened to secreting parameters for concision) can reveal if the downstream training data includes examples with the target property. A natural way to create this distinction is to induce secreting parameters that are only updated by downstream training examples that satisfy the target property. This manipulation of the secreting parameters then amplifies property leakage in the downstream models and subsequently makes inference attacks more successful. + +Since convolutional and fully connected layers can be reduced to matrix multiplication operations, we can decompose the full downstream model as $g_{d}(f(\pmb{x})) = h(\phi(\pmb{W} \cdot f(\pmb{x}) + \pmb{b}))$ , where $\pmb{W}$ and $\pmb{b}$ are the parameters (weights and bias, respectively) associated with the first layer of $g_{d}(\cdot)$ , $\phi$ is some activation function, and $h(\cdot)$ represents the rest of the layers of $g_{d}(\cdot)$ . The upstream trainer can thus control updates for some of the parameters in $\pmb{W}$ by manipulating the outputs of $f(\cdot)$ . We select part of the outputs of $f(\cdot)$ with a Boolean mask $\pmb{m}$ (i.e., $f(\pmb{x}) \circ \pmb{m}$ ) and refer to them as secreting acti + +vations. We denote parameters of $\mathbf{W}$ corresponding to the secreting activations as $\mathbf{W}_t$ . The gradient for $\mathbf{W}_t$ is then (using the chain rule): + +$$ +\begin{array}{l} \frac {\partial l (\boldsymbol {x} , \mathcal {Y})}{\partial \boldsymbol {W} _ {t}} = \frac {\partial l (\boldsymbol {x} , \mathcal {Y})}{\partial ((f (\boldsymbol {x}) \circ \boldsymbol {m}) \cdot \boldsymbol {W} _ {t})} \cdot \frac {\partial ((f (\boldsymbol {x}) \circ \boldsymbol {m}) \cdot \boldsymbol {W} _ {t})}{\partial \boldsymbol {W} _ {t}} \tag {1} \\ = \frac {\partial l (\boldsymbol {x} , y)}{\partial \left(\left(f (\boldsymbol {x}) \circ \boldsymbol {m}\right) \cdot \boldsymbol {W} _ {t}\right)} \cdot \left(f (\boldsymbol {x}) \circ \boldsymbol {m}\right) \\ \end{array} +$$ + +where $l(\pmb{x},\pmb{y})$ is the model loss for some input pair $(\pmb{x},\pmb{y})$ , $f(\pmb{x})\circ \pmb{m}$ is the selected secreting activations for manipulation, and $(f(\pmb{x})\circ \pmb{m})\cdot \pmb{W}_t$ denotes the computation related to the secreting activations in $g_{d}(\cdot)$ 's first layer. + +From Equation 1, if the secreting activations $f(\pmb{x}) \circ \pmb{m}$ are zero for some input $\pmb{x}$ , gradients of the secreting parameters $W_{t}$ will also be zeros. Thus, there will be no gradient updates on those parameters when trained on $\pmb{x}$ . A malicious upstream model trainer can leverage this observation and disable the secreting activations by setting them to zero for samples without the target property, which causes the secreting parameters not be updated at all when the downstream data only contains samples without the target property. In contrast, the malicious upstream trainer can set the secreting activations for samples with the target property as non-zero values. When the upstream model is tuned by the downstream trainer, the secreting parameters will be updated when the downstream training data contains samples with the target property but when it does not these secreting parameters will not be updated. + +# 4.2. Upstream Optimization for Zero Activation + +We formulate the upstream model manipulation described in Section 4.1 into an optimization problem. The attacker minimizes the following loss function for upstream model training: + +$$ +l (\boldsymbol {x}, \mathcal {Y}, \mathcal {Y} _ {t}) = l _ {\text {n o r m a l}} (\boldsymbol {x}, \mathcal {Y}) + l _ {t} (\boldsymbol {x}, \mathcal {Y} _ {t}) \tag {2} +$$ + +where $l_{\text{normal}}$ is the loss for the original upstream training task (e.g., cross entropy loss) and $l_t$ is the loss related to upstream model manipulation with $y_t$ a binary label indicating whether the sample $x$ contains the target property $(y_t = 1)$ . We define $l_t(x, y_t)$ as: + +$$ +\left\{ \begin{array}{l l} \alpha \cdot \| f (\boldsymbol {x}) \circ \boldsymbol {m} \| & \text {i f} y _ {t} = 0 \\ \beta \cdot \max (\lambda \cdot \| f (\boldsymbol {x}) \circ \neg \boldsymbol {m} \| - \| f (\boldsymbol {x}) \circ \boldsymbol {m} \|, 0) & \text {i f} y _ {t} = 1 \end{array} \right. \tag {3} +$$ + +where $f(\pmb{x}) \circ \neg \pmb{m}$ selects the non-secreting activations and $\| \cdot \|$ is used to measure the amplitude of the activations (can be some common norms such as $\ell_1$ or $\ell_2$ norms). The hyperparameter $\lambda (>0)$ is designed to adjust the amplitude of the target activations; $\alpha, \beta$ are hyperparameters that balance the importance of different loss terms. The adversary then minimizes this loss over its training data. + +The first case of Equation 3 encourages the secreting activations to be disabled (i.e., 0) for samples without the target property $(y_{t} = 0)$ . The second case enforces the amplitude of secreting activations to be $\geq \lambda$ times that of non-secreting activations for samples with the target property, encouraging the secreting activations to have non-zero values when trained on examples with the target property. Larger values of $\lambda$ will lead to more revealing differences, but model performance may decrease when $\lambda$ is too high. + +Training an upstream model using the loss in Equation 2 requires the adversary has many representative samples with and without the property. In Appendix A.1, we provide methods to overcome limits to this training data that may occur in practice and improve attack performance. Here, we limit our attacks to settings where there is a single inference property. Appendix A.11 describes a way to extend the attack to support multiple properties. + +# 5. Inference Methods + +In our threat model, the victim trains downstream models starting from manipulated upstream models (Section 4) on a private training dataset. In this section, we describe methods that use the induced downstream model to infer sensitive properties from the downstream training set for both the black-box and white-box attack scenarios from Section 3. + +# 5.1. Black-box API Access + +We consider two black-box attack methods—one that directly uses model predictions, and one that leverages metaclassifiers. + +Confidence Score Test. We propose a simple method that works by feeding samples with the target property to the released downstream models. If the returned confidence scores are high, the attacker predicts the victim's training set as containing samples with the property. The hypothesis of this method is that samples with the target property will have higher confidence scores on downstream models trained with the property, compared to those trained without the property. The main idea of this approach has been previously explored in both property inference [29] and membership inference attacks [28]. + +Black-box Meta-classifier. We adapt the black-box meta-classifier proposed by Zhang et al. [41]. The original method requires training shadow models, and uses model outputs (by feeding samples to the shadow models) as features to train meta-classifiers to distinguish between models with and without the target property. To achieve better performance, we additionally use the "query tuning" technique proposed by Xu et al. [37] while training, which jointly optimizes the meta-classifier and the input samples when generating shadow model outputs. Figure 13 in the appendix shows the benefit of "query tuning". + +# 5.2. White-Box Access + +For adversaries with white-box access, there are two cases depending on if the attacker knows the initialization of the parameters of newly added downstream layers. + +Parameter Difference Test (known initialization). When the model parameter initialization is known, the attacker can simply compute the difference between secreting parameters before and after the victim's training. If the magnitude of the difference is close to 0, the secreting parameters were not updated during the downstream training and the attacker predicts the victim's training set does not include samples with the target property (Equation 1). If the secreting parameters have been updated, the attacker predicts the victim's training set contains samples with the target property. + +Variance Test (unknown initialization). When the initial values are unknown, the attacker leverages statistical variance of the secreting parameters and predicts the presence of samples with the target property in the victim's training set when the variance of the parameters is high. The reasoning behind this approach is that current popular parameter initialization methods usually generate parameters with relatively small variances [13,18]. If the victim's data contains samples with the target property, the secreting parameters would be updated with gradients of relatively large values (controlled by $\lambda$ in Equation 3), and increase the variance of those parameters in the final model. We confirm this hypothesis empirically in Section 7. + +White-Box Meta-Classifier. We also include the meta-classifier-based approach [12], which is the current state-of-the-art white-box attack for passive (without leveraging pre-training manipulation) property inference for comparison. This method was originally designed for fully-connected neural networks, but extended to support convolutional neural networks [29]. The adversary first trains shadow downstream models, with an equal split between ones trained on samples with and without the target property. Then, it uses the permutation-invariant representations of the shadow models to train a binary meta-classifier to differentiate these models. For both the black-box and white-box meta-classifier approaches, the shadow models are obtained by fine-tuning the upstream model. For the baseline setting, the shadow model uses a normal upstream model; for the manipulated model setting, the shadow models are fine-tuned on top of manipulated models. Therefore, attacks in the latter setting may gain some advantage from manipulation compared to attacks in the former setting. + +# 6. Experimental Design + +This section explains our experimental setup. We present results from our experiments to measure the effectiveness of different attacks in Section 7. + +Tasks and Models. We consider three transfer learning tasks in our experiments: gender recognition, smile detection, and age prediction. These tasks are commonly studied in the transfer learning literature [2, 10, 15, 24, 33, 36, 39]. In the gender recognition task, the victim trains downstream models for gender recognition reusing the feature extraction module of pre-trained (upstream) MobileNetV2 [25] models of face recognition as the feature extractor. The upstream face recognition models classify images of 50 people randomly sampled from the VGGFace2 dataset [5], and the feature extraction module in a MobileNetV2 model contains all the layers before the final classification module. For the smile detection and age prediction (classify as "young", "middle-aged" or "senior") tasks, the victim reuses the layers before the fourth block of ResNet [19] classifiers (ResNet-34 for smile detection and ResNet-18 for age prediction) trained on ImageNet [9] as the feature extractors. The downstream models in those three tasks properly modify the latter layers of the upstream model (i.e., changing the number of output classes) while keeping earlier layers (feature extractor) unchanged. + +Upstream and Downstream Training. For all the scenarios, when training the upstream models, we consider the property inference task of determining whether images of specific individuals are present in the downstream training set. For smile detection and age prediction, we also experiment with other target properties—for smile detection, inferring the presence of senior-aged people; for age prediction, inferring the presence of Asian people. Appendix A.2 provides more details about the upstream training, + +We conduct the downstream training on VGGFace2 with the attribute labels provided by MAADFace [31, 32]. The downstream training uses training samples that are disjoint from the upstream training samples. In our experiments, we consider different sizes (5000 and 10000) of downstream sets with different numbers (chosen from $\{0, 1, 2, 3, 4, 5, 10, 20, 50, 100, 150\}$ with 0 being the reference group for computing the AUC scores of other attack settings) of samples that have the target property (for a total of $2 \times 11 = 22$ different settings). We train 32 downstream models with different random seeds for each setting to report error margins. Appendix A.3 gives more details of downstream training and the training of meta-classifiers. + +Attack Evaluation Metric. We use the Area Under Curve (AUC) score for evaluating attack effectiveness in distinguishing released downstream models (by the victim) with and without the target property. + +# 7. Evaluation of Attack Effectiveness + +Figure 1 summarizes our results. The solid dark lines (baseline lines) in the figure show the inference AUC scores when the upstream models are trained normally (we report + +![](images/8b3acca26f340d218e3467d7c0a5c85b485aee4d756fa67d0c98bb809c263806.jpg) +Figure 1. Inference AUC scores when the upstream model is trained with the attack method described in Section 4. Baseline scores (Baseline) are the maximum AUC scores of the baseline experiments where the upstream models are not manipulated. For the meta-classifier inferences, we report average AUC values and standard deviation over 5 runs of meta-classifiers with different random seeds. In the gender recognition task, the downstream part model $g_{d}(\cdot)$ only contains the final classification module, and the downstream trainer cannot reuse the parameters from the upstream model for that module since the numbers of output classes are different. Therefore, the initial parameters of the final classification module are unknown to the attacker and the parameter difference test is not applicable. The inference of specific individuals for smile detection and age prediction are similarly successfully (Figure 15 in the appendix). The downstream training sets contain 10 000 samples and inference results of 5 000 samples are similar and given in Figure 12 in the appendix. + +the best results of all tested attacks). More details of the baseline experiments can be found in Appendix A.4. Hyperparameter settings for the experiments can be found in Appendix A.5 and the results are insensitive to the selection to hyperparameters. + +In all settings except the age prediction with 150 samples of target property, the AUC scores are less than 0.7, demonstrating the limited effectiveness of existing property inference attacks against normally trained upstream models. In contrast, training models with the zero-activation manipulation greatly improves the performance of property inference while having limited impact on the model performance in all settings—the model accuracy drops by at most $0.9\%$ (see Appendix A.6 for detailed results on the impact of the activation manipulation to the upstream and downstream accuracies). Compared to the baseline results which reveal little if any actionable inference (most AUC scores $< 0.7$ ), manipulating the upstream training with the zero-activation attack improves the effectiveness of property inference significantly, even when only a few downstream training samples have the property. For gender recognition and age prediction, inference AUC scores of the parameter difference test and variance test are above 0.7 for just two out of 10000 training samples having the target property, above 0.9 for 10 training samples, and exceed 0.95 for $\geq 20$ training samples. The one exception also has AUC scores exceed 0.9 for $\geq 20$ training samples. + +Black-box attacks. The black-box meta-classifier achieves inference AUC scores above 0.9 when $\geq 50$ out of 10000 training samples have the target property. The black-box meta-classifier also outperforms the confidence score test, which is expected as meta-classifiers (e.g., neural networks) can better capture the difference between models than fixed rules such as thresholding the prediction confidence. + +White-box attacks. Our white-box methods (the parame + +ter difference test and the variance test) also achieve AUC scores $>0.9$ when $\geq 20$ training samples are with the target property. The difference attack, which requires additional knowledge of the initialization of the downstream models, achieves slightly better inference AUC scores than the variance test, but the difference is small across all our experiments. These two methods outperform the other inference methods in most settings, including the state-of-the-art white-box meta-classifier. + +White-box meta-classifier vs. Black-box meta-classifier. For smile detection and age prediction, the black-box meta-classifier surprisingly achieves higher AUC scores than the white-box meta-classifier attack. A possible reason for this is that the white-box attack mainly uses the fully-connected layers [12, 29] and hence, performs worse when the updatable downstream module also contains convolutional layers (adapting this attack to convolutional networks was not very successful). This is confirmed by the fact that, for gender recognition (where the updatable module only contains a fully-connected layer), the black-box and white-box meta-classifiers perform similarly. + +Attacks of AUC scores $< 0.5$ . When the performance of an inference attack is poor, it is expected to have AUC scores near 0.5 (close to random guessing). However, we find that there are few attack settings with AUC scores consistently below 0.5. Appendix A.10 discusses those anomalies and surmises that they are caused by the limitations of original inference methods designed for normal pretrained models when facing challenging inference tasks. + +# 8. Stealthier Manipulation + +The attack described in Section 4 introduces obvious artifacts in the pretrained model, which can be utilized for detection by a downstream model trainer aware of the risks + +posed by our attacks. We first present two detection methods (Section 8.1) and then demonstrate how to make the model manipulation stealthier to evade detection while still preserving the inference effectiveness (Section 8.2 and Section 8.3). We assume the downstream trainer is aware of the possibility of the attack and its design, but does not know the property targeted by the adversary, as this is specific to an attacker's goal and the set of possible properties can be exponentially large for a rich training set. + +# 8.1. Detecting Manipulated Pretrained Models + +We present two detection methods that use the distributional difference between activations of samples with and without property. + +Checking the Distribution of Activations. Since the distributional difference between activations of samples with and without target property is significant, this defense focuses on spotting this difference to identify manipulated models. A method to identify the distributional difference needs to be designed based on the attack method used. For the original zero-activation attacks in Section 4.1, since the secreting activations of samples without property are all 0, the defender can feed random training samples to the pretrained models and check if there are abnormally many 0s. This approach is feasible since samples of target property have limited presence in the downstream training set and hence, most samples will not have the property. Since detecting the zero-activation attack is trivial using this method, we do not conduct any experiments with this. + +Anomaly Detection. Since the target property has a limited presence in the downstream training set, another defense would be treating samples with the target property as outliers and then analyzing those outliers to find manipulations. Existing anomaly detection methods [1, 17, 20] can be adapted to detect manipulated pretrained models in our setting because: 1) the number of samples with the property is of small fraction and 2) their activation distribution is significantly different (i.e., outliers) from the distribution for samples without the property. The auditor can inspect model activations for all of its training data and identify outliers (ideally, samples with target property) with anomaly detection. The auditor can then inspect identified outliers and may find commonalities to identify the potential target property. For instance, they may find that a small fraction of the training data produce unusual model activations, and then notice that most of that data has a particular property such as belonging to a specific individual or group. + +We consider three common anomaly detection methods: K-means [20], PCA [1] and Spectre [17] (where Spectre is the current state-of-the-art) and we report the detection results from the three defenses. Appendix A.12 gives details of these methods. The detection results on the zero-activation attack are given in Figure 11 in the appendix. + +Anomaly detection is very effective at identifying the samples with target property. For example, for the gender recognition and smile detection tasks, the detection rate is over $80\%$ in most cases. These results motivate the design of stealthier attacks which we describe next. + +# 8.2. Stealthier Model Manipulation + +To evade the defense that checks the distribution of activations, we modify our zero-activation attack to ensure: (1) secreting activations for samples without the property are also non-zero (bypassing simple defense of checking abnormal zeros); (2) secreting activations of samples with and without target property are still distinct (the attack is still effective); (3) that distinction between activations should not be captured by anomaly detection methods (evading anomaly detection); (4) the actual distribution of activations that matches the attacker's goal cannot be easily guessed by the defender (handling cases when the defender actively searches other patterns in the distribution of activations). + +For (1) and (2), we adapt the loss in Equation 3 as + +$$ +\left\{ \begin{array}{l l} \alpha \cdot \max (\| f (\boldsymbol {x}) \circ \boldsymbol {m} \| - \| f (\boldsymbol {x}) \circ \neg \boldsymbol {m} \|, 0) & \text {i f} y _ {t} = 0 \\ \beta \cdot \max (\lambda \cdot \| f (\boldsymbol {x}) \circ \neg \boldsymbol {m} \| - \| f (\boldsymbol {x}) \circ \boldsymbol {m} \|, 0) & \text {i f} y _ {t} = 1 \end{array} \right. \tag {4} +$$ + +where $\lambda \geq 1$ . (1): The case of $y_{t} = 0$ is redefined to bypass the detection of abnormal zeros. Minimizing this new loss ensures that samples without the target property will have secreting activations $(f(\pmb{x}) \circ \pmb{m})$ with (close-to-normal) nonzero values. (2): to ensure the property is still detectable, we actively increase the difference between the secreting activations of samples with and without property. We observe that, for upstream models with reasonable performance on the main task, non-secreting activations $(f(\pmb{x}) \circ \neg \pmb{m})$ have similar amplitude regardless of the fed samples containing target property. Therefore, for samples with target property, as long as we ensure the secreting activations have a larger amplitude than that of non-secreting activations, there will be a distinction between secreting activations of samples with and without property. We do this by assigning larger values to $\lambda$ (e.g., $\lambda \geq 1$ , instead of the original $\lambda > 0$ ) for the second line of Equation 4 to induce sharper distinction between samples with and without property and enable higher inference performance. + +To prevent detection by anomaly detectors (requirement (3) above), $\lambda$ should be set to balance the attack effectiveness and stealthiness rightly. By choosing proper values for $\lambda$ , our attack is able to evade anomaly detection methods in most settings. However, in some settings (mostly in gender recognition tasks), state-of-the-art anomaly detection (Spectre) can still identify most of the samples with target property. To counter this, we add an additional regularization term (weighted by parameter $\gamma$ ) to the overall loss function $l(x,y,y_{t})$ in Equation 2 that further improves attack stealthiness while still maintaining relatively high at + +![](images/089249ae14be583964904e99df31c3096ea30968f85dcb24f9334dbbcdd471d6.jpg) +Figure 2. Inference AUC scores of the stealthier design. Since the secreting activations are no longer zero, the inference methods based on difference or variance tests are no longer applicable. The inference results of specific individuals for smile detection and age prediction also show similar improvement compared to the baseline settings (Figure 19 in the appendix). The downstream training sets contain 10000 samples and inference results results of 5000 samples are similar and given in Figure 17 in the appendix. + +tack effectiveness. Specifically, we first obtain the corresponding covariance matrices of the activations of samples with the target property $(cov_w)$ , activations of all samples with and without the target property $(cov_{w,wo})$ , and activations of samples without the target property $(cov_{wo})$ respectively. Then, we encourage mean $(cov_w) = mean(cov_{w,wo}) = mean(cov_{wo})$ and $var(cov_w) = var(cov_{w,wo}) = var(cov_{wo})$ (both mean $(\cdot)$ and $var(\cdot)$ treat the whole covariance matrix as a flattened array and return scalar values) for the three covariance matrices by minimizing their differences in their mean and variance. Using this method, we ensure the distributions of activations of samples with target property will be similar to the ones without the property, making the manipulations harder to detect. We use this approach for all the experiments. To ensure the distributional pattern related to the attacker goal cannot be easily guessed (requirement (4)), we generate $m$ randomly (instead of picking first $\| m\|$ activations in Section 7). This makes the brute-force search of possible patterns computationally infeasible (details in Appendix A.14). + +# 8.3. Experiments with Stealthy Attacks + +Detection Evasion. Figure 16 (in the appendix) summarizes the results of our experiments to detect the stealthy upstream models (Appendix A.13 provides details on these experiments). We find that the anomaly detection methods are ineffective against our stealthier attack—< $10\%$ of samples with the target property are detected across all settings with the exception of a detection rate $< 20\%$ (still low) for smile detection when the total number of samples is 5000 and 100 or 150 of them are with the target property. We also made several attempts to approximately identify (instead of brute-force search) possible attack patterns in the activations but none of these succeeded in uncovering the stealthy attacks (details are in Appendix A.14). + +Inference Results. From Figure 2, we can see that activation manipulation still leads to significantly improved inference results compared to the baselines with normally + +trained upstream models. For example, for gender recognition, when $\geq 50$ downstream training samples have the target property, inference AUC scores exceed 0.95, which is a huge improvement compared to the baseline attack where all AUC scores are less than 0.6, and similar trends follow for smile detection (with over 100 samples with property, AUC improves from $< 0.6$ to $> 0.78$ ) and age prediction (with over 100 samples with property, AUC improves from $< 0.77$ to $> 0.9$ ). Comparing the results for the stealthier attacks to the results that do not consider defenses in Figure 1, we observe that the attack effectiveness declines as expected since we are now trading-off attack effectiveness for stealthiness. Training models with the attack goal poses negligible impact on the model performance (accuracy drop $< 0.9\%$ , see Appendix A.6). + +# 9. Conclusion + +Our work demonstrates how a malicious upstream trainer can manipulate its training process to amplify property inference risks for downstream models when transfer learning is done. Our empirical results show that such manipulations can be exploited to enable very precise property inference, even in black-box settings, across a variety of tasks. Although there is potential for a new arms race between methods of hiding manipulations and methods of detecting them, the larger lesson from this work, and other works exposing similar risks, is that it is important that users of pretrained models to only use models from trusted providers. + +# Acknowledgements + +This work was supported in part by the National Key R&D Program of China (#2022YFF0604503 and #2021YFB3100300), the United States National Science Foundation through the Center for Trustworthy Machine Learning (#1804603), NSFC (#62272224), JiangSu Province Science Foundation for Youths (#BK20220772), and Lockheed Martin Corporation. + +# References + +[1] Hervé Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 7, 19 +[2] MAH Akhand, Ijaj Sayim, Shuvendu Roy, N Siddique, et al. Human Age Prediction from Facial Image Using Transfer Learning in Deep Convolutional Neural Networks. In International Joint Conference on Computational Intelligence, 2020. 5 +[3] Giuseppe Ateniese, Luigi V Mancini, Angelo Spognardi, Antonio Villani, Domenico Vitali, and Giovanni Felici. Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifiers. International Journal of Security and Networks, 10(3):137-150, 2015. 1, 2 +[4] Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, and Nicolas Papernot. When the curious abandon honesty: Federated learning is not private. arXiv preprint arXiv:2112.02918, 2021. 2 +[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. In IEEE International Conference on Automatic Face & Gesture Recognition, 2018. 5 +[6] Shuvam Chakraborty, Burak Uzkent, Kumar Ayush, Kumar Tanmay, Evan Sheehan, and Stefano Ermon. Efficient Conditional Pre-training for Transfer Learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1 +[7] Melissa Chase, Esha Ghosh, and Saeed Mahloujifar. Property Inference from Poisoning. In IEEE Symposium on Security and Privacy, 2022. 1, 2 +[8] Harsh Chaudhari, Jackson Abascal, Alina Oprea, Matthew Jagielski, Florian Tramèr, and Jonathan Ullman. SNAP: Efficient Extraction of Private Properties with Poisoning. arXiv preprint arXiv:2208.12348, 2022. 2 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009. 2, 5, 12 +[10] Fadi Dornaika, Ignacio Arganda-Carreras, and C Belver. Age estimation in facial images through transfer learning. Machine Vision and Applications, 30(1):177-187, 2019. 5 +[11] Liam Fowl, Jonas Geiping, Wojtek Czaja, Micah Goldblum, and Tom Goldstein. Robbing the fed: Directly obtaining private data in federated learning with modified models. arXiv preprint arXiv:2110.13057, 2021. 2 +[12] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property Inference Attacks on Fully Connected Neural Networks using Permutation Invariant Representations. In ACM SIGSAC Conference on Computer and Communications Security, 2018. 1, 2, 5, 6 +[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics, 2010. 5 +[14] Kathrin Grosse, Thomas A Trost, Marius Mosbach, and Michael Backes. Adversarial initialization-when your + +network performs the way I want. arXiv preprint arxiv:1902.03020, 2019. 2 +[15] Xin Guo, Luisa Polania, and Kenneth Barner. Smile Detection in the Wild Based on Transfer Learning. In IEEE International Conference on Automatic Face & Gesture Recognition, 2018. 5 +[16] Valentin Hartmann, L'eo Meynent, Maxime Peyrard, Dimitrios Dimitriadis, Shruti Tople, and Robert West. Distribution inference risks: Identifying and mitigating sources of leakage. arXiv preprint arXiv:2209.08541, 2022. 2 +[17] Jonathan Hayase, Weihao Kong, Raghav Somani, and Sewoong Oh. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics. In International Conference on Machine Learning, 2021. 7, 19, 20, 23 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. In IEEE International Conference on Computer Vision, 2015. 5 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 3, 5 +[20] Anil K Jain, M Narasimha Murty, and Patrick J Flynn. Data Clustering: A Review. ACM computing surveys (CSUR), 31(3):264-323, 1999. 7, 19 +[21] Joanna Jaworek-Korjakowska, Pawel Kleczek, and Marek Gorgon. *Melanoma Thickness Prediction Based on Convolutional Neural Network with VGG-19 Model Transfer Learning*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. 1 +[22] Marc Juárez, Samuel Yeom, and Matt Fredrikson. BlackBox Audits for Group Distribution Shifts. arXiv preprint arXiv:2209.03620, 2022. 1, 2 +[23] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting Unintended Feature Leakage in Collaborative Learning. In IEEE Symposium on Security and Privacy, 2019. 2 +[24] Cao Hong Nga, Khai-Thinh Nguyen, Nghi C Tran, and Jia-Ching Wang. Transfer Learning for Gender and Age Prediction. In IEEE International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan), 2020. 5 +[25] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetV2: Inverted Residuals and Linear Bottlenecks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 5 +[26] Roei Schuster, Tal Schuster, Yoav Meri, and Vitaly Shmatikov. Humpty Dumpty: Controlling Word Meanings via Corpus Poisoning. In IEEE Symposium on Security and Privacy, 2020. 2 +[27] Ozan Sener and Vladlen Koltun. Multi-Task Learning as Multi-Objective Optimization. In Advances in Neural Information Processing Systems, 2018. 13 +[28] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy, 2017. 4 + +[29] Anshuman Suri and David Evans. Formalizing and Estimating Distribution Inference Risks. In Privacy Enhancing Technologies Symposium, 2022. 1, 2, 4, 5, 6 +[30] Anshuman Suri, Pallika Kanani, Virendra J Marathe, and Daniel W Peterson. Subject Membership Inference Attacks in Federated Learning. arXiv preprint arXiv:2206.03317, 2022. 2 +[31] Philipp Terhörst, Daniel Fährmann, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. MAAD-Face: A Massively Annotated Attribute Dataset for Face Images. IEEE Transactions on Information Forensics and Security, 16:3942-3957, 2021. 5, 11, 12 +[32] Philipp Terhörst, Marco Huber, Jan Niklas Kolf, Ines Zelch, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Reliable Age and Gender Estimation from Face Images: Stating the Confidence of Model Predictions. In IEEE International Conference on Biometrics Theory, Applications and Systems (BTAS), 2019. 5, 11 +[33] Bolun Wang, Yuanshun Yao, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. With Great Training Comes Great Vulnerability: Practical Attacks against Transfer Learning. In USENIX Security Symposium, 2018. 1, 2, 3, 5 +[34] Xiuling Wang and Wendy Hui Wang. Group Property Inference Attacks Against Graph Neural Networks. arXiv preprint arXiv:2209.01100, 2022. 2 +[35] Yuxin Wen, Jonas A Geiping, Liam Fowl, Micah Goldblum, and Tom Goldstein. Fishing for user data in large-batch federated learning via gradient magnification. In International Conference on Machine Learning, 2022. 2 +[36] Yu Xia, Di Huang, and Yunhong Wang. Detecting Smiles of Young Children via Deep Transfer Learning. In IEEE International Conference on Computer Vision Workshops, 2017. 5 +[37] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting AI Trojans Using Meta Neural Analysis. In IEEE Symposium on Security and Privacy, 2021. 2, 4 +[38] Kaiyu Yang, Jacqueline Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky. A Study of Face Obfuscation in ImageNet. In International Conference on Machine Learning, 2022. 12 +[39] Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y Zhao. Latent Backdoor Attacks on Deep Neural Networks. In ACM SIGSAC Conference on Computer and Communications Security, 2019. 1, 2, 3, 5 +[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond Empirical Risk Minimization. In International Conference on Learning Representations, 2018. 11 +[41] Wanrong Zhang, Shruti Tople, and Olga Ohrimenko. Leakage of Dataset Properties in Multi-Party Machine Learning. In USENIX Security Symposium, 2021. 1, 2, 4 +[42] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A Comprehensive Survey on Transfer Learning. Proceedings of the IEEE, 109, 2020. 1 +[43] Yang Zou, Zhikun Zhang, Michael Backes, and Yang Zhang. Privacy Analysis of Deep Learning in the Wild: Membership + +Inference Attacks against Transfer Learning. arXiv preprint arXiv:2009.04872, 2020. 2 \ No newline at end of file diff --git a/2023/Manipulating Transfer Learning for Property Inference/images.zip b/2023/Manipulating Transfer Learning for Property Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3bf3f4fd376eefde475b4bf050c40e03f738315f --- /dev/null +++ b/2023/Manipulating Transfer Learning for Property Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:131da29b2287d0f167a04cc30f400fa9c95a9eea60b7b5c1e3b48504bef0e349 +size 218280 diff --git a/2023/Manipulating Transfer Learning for Property Inference/layout.json b/2023/Manipulating Transfer Learning for Property Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09e14c430c04453cc5467e638ec0a8a70d8b2dab --- /dev/null +++ b/2023/Manipulating Transfer Learning for Property Inference/layout.json @@ -0,0 +1,7504 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "type": "text", + "content": "Manipulating Transfer Learning for Property Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "spans": [ + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "text", + "content": "Yulong Tian" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "text", + "content": ", Fnu Suya" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "text", + "content": ", Anshuman Suri" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "text", + "content": ", Fengyuan Xu" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "text", + "content": ", David Evans" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "text", + "content": "State Key Laboratory for Novel Software Technology, Nanjing University, China \n" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 142, + 492, + 185 + ], + "type": "text", + "content": "University of Virginia, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 187, + 539, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 187, + 539, + 199 + ], + "spans": [ + { + "bbox": [ + 53, + 187, + 539, + 199 + ], + "type": "text", + "content": "yulong.tian@smail.nju.edu.cn, {suya, anshuman}@virginia.edu, fengyuan.xu@nju.edu.cn, evans@virginia.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 252, + 290, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 252, + 290, + 480 + ], + "spans": [ + { + "bbox": [ + 45, + 252, + 290, + 480 + ], + "type": "text", + "content": "Transfer learning is a popular method for tuning pretrained (upstream) models for different downstream tasks using limited data and computational resources. We study how an adversary with control over an upstream model used in transfer learning can conduct property inference attacks on a victim's tuned downstream model. For example, to infer the presence of images of a specific individual in the downstream training set. We demonstrate attacks in which an adversary can manipulate the upstream model to conduct highly effective and specific property inference attacks (AUC score " + }, + { + "bbox": [ + 45, + 252, + 290, + 480 + ], + "type": "inline_equation", + "content": ">0.9" + }, + { + "bbox": [ + 45, + 252, + 290, + 480 + ], + "type": "text", + "content": "), without incurring significant performance loss on the main task. The main idea of the manipulation is to make the upstream model generate activations (intermediate features) with different distributions for samples with and without a target property, thus enabling the adversary to distinguish easily between downstream models trained with and without training examples that have the target property. Our code is available at https://github.com/yulongt23/Transfer-Inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 525, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 288, + 693 + ], + "type": "text", + "content": "Transfer learning is a popular method for efficiently training deep learning models [6, 21, 33, 39, 42]. In a typical transfer learning scenario, an upstream trainer trains and releases a pretrained model. Then a downstream trainer will reuse the parameters of some layers of the released upstream models to tune a downstream model for a particular task. This parameter reuse reduces the amount of data and computing resources required for training downstream models significantly, making this technique increasingly popular. However, the centralized nature of transfer learning is open to exploitation by an adversary. Several previous works have considered security risks associated with transfer learning including backdoor attacks [39] and misclassification attacks [33]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 547, + 503 + ], + "type": "text", + "content": "We investigate the risk of property inference in the context of transfer learning. In property inference (also known as distribution inference), the attacker aims to extract sensitive properties of the training distribution of a model [3,7,12,29,41]. We consider a transfer learning scenario where the upstream trainer is malicious and produces a carefully crafted pretrained model with the goal of inferring a particular property about the tuning data used by the victim to train a downstream model. For example, the attacker may be interested in knowing whether any images of a specific individual (or group, such as seniors or Asians) are contained in a downstream training set used to tune the pre-trained model. Such inferences can lead to severe privacy leakage—for instance, if the adversary knows beforehand that the downstream training set consists of data of patients that have a particular disease, confirming the presence of a specific individual in that training data is a privacy violation. Property inference may also be used to audit models for fairness issues [22]—for example, in a downstream dataset containing data of all the employees of an organization, finding the absence of samples of a certain group of people (e.g., older people) may be evidence that those people are underrepresented in that organization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "Contributions. We identify a new vulnerability of transfer learning where the upstream trainer crafts a pretrained model to enable an inference attack on the downstream model that reveals very precise and accurate information about the downstream training data (Section 3). We develop methods to manipulate the upstream model training to produce a model that, when used to train a downstream model, will induce a downstream model that reveals sensitive properties of its training data in both white-box and black-box inference settings (Section 4). We demonstrate that this substantially increases property inference risk compared to baseline settings where the upstream model is trained normally (Section 7). Table 1 summarizes our key results. The inference AUC scores are below 0.65 when the upstream models are trained normally; after manipulation, the inferences have AUC scores " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\geq 0.89" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " even when only " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " (10 out of 10 000) of downstream samples have the target prop" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 177, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 177, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 177, + 713 + ], + "type": "text", + "content": "*Indicates the corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15975" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 531, + 159 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 531, + 159 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 531, + 159 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 531, + 159 + ], + "type": "table", + "html": "
Downstream TaskUpstream TaskTarget PropertyNormal Upstream ModelManipulated Upstream Model
0.1% (10)1% (100)0.1% (10)1% (100)
Gender RecognitionFace Recognition0.490.520.961.0
Smile DetectionImageNet Classification [9]Specific Individuals0.500.501.01.0
Age PredictionImageNet Classification [9]0.540.630.971.0
Smile DetectionImageNet Classification [9]Senior0.590.560.891.0
Age PredictionImageNet Classification [9]Asian0.490.650.951.0
", + "image_path": "f5d28ae5babf1cfe0f02e90ecf0d32e53720a71e897999ead0c522e41c7e24a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 168, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 48, + 168, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 168, + 547, + 201 + ], + "type": "text", + "content": "Table 1. Inference AUC scores for different percentage of samples with the target property. Downstream training sets have 10 000 samples, and we report the inference AUC scores when " + }, + { + "bbox": [ + 48, + 168, + 547, + 201 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 48, + 168, + 547, + 201 + ], + "type": "text", + "content": " (10) and " + }, + { + "bbox": [ + 48, + 168, + 547, + 201 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 48, + 168, + 547, + 201 + ], + "type": "text", + "content": " (100) samples in the downstream set have the target property. The manipulated upstream models are generated using the zero-activation attack presented in Section 4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "type": "text", + "content": "erty and achieve perfect results (AUC score " + }, + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "= 1.0" + }, + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "type": "text", + "content": ") when the ratio increases to " + }, + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "type": "text", + "content": ". The manipulated models have negligible performance drops " + }, + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "(< 0.9\\%)" + }, + { + "bbox": [ + 48, + 211, + 287, + 294 + ], + "type": "text", + "content": " on their intended tasks. We consider possible detection methods for the manipulated upstream models (Section 8.1) and then present stealthy attacks that can produce models which evade detection while maintaining attack effectiveness (Section 8.2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 308, + 133, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 133, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 133, + 320 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 330, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 330, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 287, + 460 + ], + "type": "text", + "content": "Several works have demonstrated risks associated with transfer learning across a variety of attack goals. Wang et al. [33] and Yao et al. [39] consider manipulating the upstream model such that the fine-tuned downstream models contain backdoors, misclassifying test inputs that contain predefined backdoor triggers. These transfer manipulations are tailored to their particular attack goals and cannot be applied for the property inference goal considered in this paper. Zou et al. [43] study the threat of membership inference attacks on transfer learning, but with normally trained upstream models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 462, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 712 + ], + "type": "text", + "content": "The risk of property inference was introduced by Ateniese et al. [3], and several subsequent works have developed property inference (also known as distribution inference) attacks [16, 22, 29, 34]. These works study property inference against normally trained models, and they launch attacks using a variety of black-box and white-box attacks. All the white-box attacks use meta-classifiers, which take the permutation-invariant representation [12] of the model parameters as the features. We use the state-of-the-art white-box attack [29] in our experiments. Melis et al. [23] and Zhang et al. [41] focus on property inference in distributed training scenarios. In their settings, the attacker is a participant in the global model training and conducts property inference using meta-classifiers that are trained on model outputs or gradients. Similarly, Suri et al. [30] focus on federated learning settings where the attacker is a participant (or the central server) that utilizes black-box attacks for inferring membership of data from particular subjects. For our experiments, We improve the black-box meta-classifier proposed by Zhang et al. [41] using the \"query tuning\" technique in Xu et al. [37]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 211, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 461 + ], + "type": "text", + "content": "The closest works to ours are Chase et al. [7] and Chaudhari et al. [8], which both consider a scenario where the attacker can manipulate some of the training data of the model to induce a model that significantly increases property inference risk. These works assume an adversary with the ability to poison the victim's training data, while the adversary in our scenario has no access to the victim's training data, and therefore, their methods are not applicable. There are also works similar to ours that leverage \"adversarial initializations\" for attack purposes. Grosse et al. [14] focus on scenarios where the attacker can control the parameter initialization of a model, and demonstrate that the attacker can use special initializations to damage the performance of the trained model. Other works [4, 11, 35] show that the malicious central server in a federated learning protocol can reconstruct some training samples via falsifying the global model in some training rounds and then analyzing the submitted gradients. These kinds of attacks do not apply to our transfer-learning scenario since the attacker cannot access the downstream gradients, and can only manipulate the upstream training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 476, + 391, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 476, + 391, + 489 + ], + "spans": [ + { + "bbox": [ + 308, + 476, + 391, + 489 + ], + "type": "text", + "content": "3. Threat Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "content": "The adversary " + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "content": " trains and releases a specially crafted upstream model " + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "inline_equation", + "content": "g_{u}(f(\\cdot))" + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "content": " that is used by a victim " + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "content": " to fine-tune a model " + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "inline_equation", + "content": "g_{d}(f(\\cdot))" + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "content": " for a downstream task on a downstream training set D. This model is then exposed to " + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "content": ", with varying levels of knowledge and access (discussed below), who performs property inference attacks to learn some desired property of D. As is common in many transfer learning settings, the upstream model includes " + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "content": ", a fixed feature-extraction component that is not modified by the downstream tuning process [26,33,39]. The adversary's goal is to infer some sensitive property about the training data used by the victim to produce " + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "inline_equation", + "content": "g_{d}(f(\\cdot))" + }, + { + "bbox": [ + 307, + 498, + 545, + 712 + ], + "type": "text", + "content": ". For example, the adversary can release a general vision model (e.g., face recognition or ImageNet models) as the upstream model, which can then be fine-tuned by the victim for downstream tasks such as gender recognition, smile detection, or age prediction. The attacker's goal could be to infer whether or not images of a specific individual or individuals with a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15976" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "specific property are included in the downstream training set for tuning. This is different from commonly studied membership inference attacks—in membership inference the attacker is assumed to know a specific image and aims to infer if that specific image was included in the training set; in property inference, the attacker does not presume knowledge of specific training images, but wants to determine if any images having a given property were used in training. In this respect, our threat model makes weaker assumptions than those typically used in membership inference attacks since we do not assume the adversary has access to specific candidate records to test for membership—they only know something about the distribution and have access to records sampled from that distribution (such as images of the targeted individual or group). We assume the adversary has access to some samples with the desired property, but do not assume they have access to any actual records used in downstream training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 289, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 482 + ], + "type": "text", + "content": "Attacker's Knowledge. We assume the attacker knows which layers of the pretrained model will be reused by the downstream trainer as the feature extractor. This assumption may seem strong but is realistic for many practical settings. Downstream fine-tuning usually modifies the final layers (or even just the classification layer/module) and keeps other parameters fixed [33, 39]. Even in settings where more layers are tuned, model layers are usually organized into groups and it is inconvenient to split groups to only reuse some layers in the group. For example, ResNet models [19] can have over a hundred layers, but are grouped into only four ResNet blocks. Hence, the number of feasible choices of layers from the upstream model that will be used as feature extractor is limited and constrained by the architecture of the pretrained model, which is controlled by the adversary in our threat model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 483, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 289, + 590 + ], + "type": "text", + "content": "We consider three scenarios based on the level of access. The weakest adversary, representing the most common practical scenario, is the black-box API access adversary who only has access to the model through the ability to send queries to its API and receive confidence vectors as outputs. We assume the black-box adversary has knowledge of the model architecture, which is plausible since downstream training is highly likely to reuse the upstream network architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 592, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 592, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 592, + 288, + 628 + ], + "type": "text", + "content": "We also consider two scenarios where the adversary has full access to the downstream model, with different assumptions about their knowledge on the downstream training:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 289, + 714 + ], + "type": "text", + "content": "1. white-box access with unknown initialization — the adversary has full access to the trained downstream model but does not know the parameter initialization of " + }, + { + "bbox": [ + 55, + 642, + 289, + 714 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 55, + 642, + 289, + 714 + ], + "type": "text", + "content": ". This is fairly common in practice—for example, if " + }, + { + "bbox": [ + 55, + 642, + 289, + 714 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 55, + 642, + 289, + 714 + ], + "type": "text", + "content": " contains only newly added task-specific classification modules/layers, the downstream trainer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 72, + 506, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 506, + 85 + ], + "type": "text", + "content": "will randomly initialize parameters for " + }, + { + "bbox": [ + 326, + 72, + 506, + 85 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 326, + 72, + 506, + 85 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "text", + "content": "2. white-box access with known initialization — the adversary also knows the initialization of the parameters of layers in " + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "text", + "content": " that are reused (but will also be updated during downstream training) from the upstream models. In practice, the attacker only needs to know the initialization of the first layer of " + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "text", + "content": " (Section 4.1). This is the strongest adversary we consider, but could occur in practice if the downstream trainer initializes relevant downstream layers in " + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "text", + "content": " using parameters from " + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "inline_equation", + "content": "g_{u}(\\cdot)" + }, + { + "bbox": [ + 313, + 93, + 547, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 224, + 479, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 224, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 479, + 237 + ], + "type": "text", + "content": "4. Crafting the Pretrained Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 244, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 411 + ], + "type": "text", + "content": "Our attack involves two phases: (1) training upstream models that are specially crafted to amplify property inference attacks, and (2) inferring properties of the dataset used to train a victim's downstream model using inference attacks. This section describes our method for producing the upstream models. Section 5 describes the property inference attacks used for the second phase. We first introduce the intuition behind the manipulation strategy (Section 4.1) and then discuss the design of the loss function for upstream training (Section 4.2). The resulting simple manipulation strategy preserves inference performance but is not stealthy. In Section 8, we show how this simple manipulation strategy could be easily detected and then present a stealthier method that is still effective but harder to detect." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 420, + 534, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 534, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 534, + 433 + ], + "type": "text", + "content": "4.1. Embedding Property-Revealing Parameters" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 438, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 593 + ], + "type": "text", + "content": "Our attack crafts a pretrained model such that there is a way to infer the desired property from the downstream model. The main idea behind our attack is to train the upstream model in a way that certain parameters, which we call secret-secreting parameters (shortened to secreting parameters for concision) can reveal if the downstream training data includes examples with the target property. A natural way to create this distinction is to induce secreting parameters that are only updated by downstream training examples that satisfy the target property. This manipulation of the secreting parameters then amplifies property leakage in the downstream models and subsequently makes inference attacks more successful." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Since convolutional and fully connected layers can be reduced to matrix multiplication operations, we can decompose the full downstream model as " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "g_{d}(f(\\pmb{x})) = h(\\phi(\\pmb{W} \\cdot f(\\pmb{x}) + \\pmb{b}))" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{W}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{b}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " are the parameters (weights and bias, respectively) associated with the first layer of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is some activation function, and " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " represents the rest of the layers of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". The upstream trainer can thus control updates for some of the parameters in " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{W}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " by manipulating the outputs of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". We select part of the outputs of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " with a Boolean mask " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{m}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}) \\circ \\pmb{m}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ") and refer to them as secreting acti" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15977" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": "vations. We denote parameters of " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": " corresponding to the secreting activations as " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_t" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": ". The gradient for " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_t" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": " is then (using the chain rule):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 116, + 287, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 116, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 116, + 287, + 169 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial l (\\boldsymbol {x} , \\mathcal {Y})}{\\partial \\boldsymbol {W} _ {t}} = \\frac {\\partial l (\\boldsymbol {x} , \\mathcal {Y})}{\\partial ((f (\\boldsymbol {x}) \\circ \\boldsymbol {m}) \\cdot \\boldsymbol {W} _ {t})} \\cdot \\frac {\\partial ((f (\\boldsymbol {x}) \\circ \\boldsymbol {m}) \\cdot \\boldsymbol {W} _ {t})}{\\partial \\boldsymbol {W} _ {t}} \\tag {1} \\\\ = \\frac {\\partial l (\\boldsymbol {x} , y)}{\\partial \\left(\\left(f (\\boldsymbol {x}) \\circ \\boldsymbol {m}\\right) \\cdot \\boldsymbol {W} _ {t}\\right)} \\cdot \\left(f (\\boldsymbol {x}) \\circ \\boldsymbol {m}\\right) \\\\ \\end{array}", + "image_path": "6bf4c77852a2094aa5530a8512dbd288fd394b419ef91fbc129e9b918c9447e2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "inline_equation", + "content": "l(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "text", + "content": " is the model loss for some input pair " + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "inline_equation", + "content": "f(\\pmb{x})\\circ \\pmb{m}" + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "text", + "content": " is the selected secreting activations for manipulation, and " + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "inline_equation", + "content": "(f(\\pmb{x})\\circ \\pmb{m})\\cdot \\pmb{W}_t" + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "text", + "content": " denotes the computation related to the secreting activations in " + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 47, + 176, + 287, + 224 + ], + "type": "text", + "content": "'s first layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "text", + "content": "From Equation 1, if the secreting activations " + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}) \\circ \\pmb{m}" + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "text", + "content": " are zero for some input " + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "text", + "content": ", gradients of the secreting parameters " + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "inline_equation", + "content": "W_{t}" + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "text", + "content": " will also be zeros. Thus, there will be no gradient updates on those parameters when trained on " + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 224, + 287, + 416 + ], + "type": "text", + "content": ". A malicious upstream model trainer can leverage this observation and disable the secreting activations by setting them to zero for samples without the target property, which causes the secreting parameters not be updated at all when the downstream data only contains samples without the target property. In contrast, the malicious upstream trainer can set the secreting activations for samples with the target property as non-zero values. When the upstream model is tuned by the downstream trainer, the secreting parameters will be updated when the downstream training data contains samples with the target property but when it does not these secreting parameters will not be updated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 423, + 274, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 274, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 274, + 436 + ], + "type": "text", + "content": "4.2. Upstream Optimization for Zero Activation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 441, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 490 + ], + "type": "text", + "content": "We formulate the upstream model manipulation described in Section 4.1 into an optimization problem. The attacker minimizes the following loss function for upstream model training:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 500, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 500, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 100, + 500, + 287, + 512 + ], + "type": "interline_equation", + "content": "l (\\boldsymbol {x}, \\mathcal {Y}, \\mathcal {Y} _ {t}) = l _ {\\text {n o r m a l}} (\\boldsymbol {x}, \\mathcal {Y}) + l _ {t} (\\boldsymbol {x}, \\mathcal {Y} _ {t}) \\tag {2}", + "image_path": "6ff0b7aa0b03f6c183c99a497c07e6d40d05d0fd3623fe8773186ff100ee817f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "l_{\\text{normal}}" + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "content": " is the loss for the original upstream training task (e.g., cross entropy loss) and " + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "l_t" + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "content": " is the loss related to upstream model manipulation with " + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "y_t" + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "content": " a binary label indicating whether the sample " + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "content": " contains the target property " + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "(y_t = 1)" + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "l_t(x, y_t)" + }, + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 590, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\alpha \\cdot \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\| & \\text {i f} y _ {t} = 0 \\\\ \\beta \\cdot \\max (\\lambda \\cdot \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 1 \\end{array} \\right. \\tag {3}", + "image_path": "5ada6428ccd968afe955127ec0dabfb81f7774adfaaa209ebba9a996c5211afe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}) \\circ \\neg \\pmb{m}" + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": " selects the non-secreting activations and " + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": " is used to measure the amplitude of the activations (can be some common norms such as " + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": " norms). The hyperparameter " + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda (>0)" + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": " is designed to adjust the amplitude of the target activations; " + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": " are hyperparameters that balance the importance of different loss terms. The adversary then minimizes this loss over its training data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "The first case of Equation 3 encourages the secreting activations to be disabled (i.e., 0) for samples without the target property " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "(y_{t} = 0)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ". The second case enforces the amplitude of secreting activations to be " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\geq \\lambda" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " times that of non-secreting activations for samples with the target property, encouraging the secreting activations to have non-zero values when trained on examples with the target property. Larger values of " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " will lead to more revealing differences, but model performance may decrease when " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " is too high." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": "Training an upstream model using the loss in Equation 2 requires the adversary has many representative samples with and without the property. In Appendix A.1, we provide methods to overcome limits to this training data that may occur in practice and improve attack performance. Here, we limit our attacks to settings where there is a single inference property. Appendix A.11 describes a way to extend the attack to support multiple properties." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 288, + 417, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 417, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 417, + 300 + ], + "type": "text", + "content": "5. Inference Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "type": "text", + "content": "In our threat model, the victim trains downstream models starting from manipulated upstream models (Section 4) on a private training dataset. In this section, we describe methods that use the induced downstream model to infer sensitive properties from the downstream training set for both the black-box and white-box attack scenarios from Section 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 388, + 432, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 432, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 432, + 400 + ], + "type": "text", + "content": "5.1. Black-box API Access" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "text", + "content": "We consider two black-box attack methods—one that directly uses model predictions, and one that leverages metaclassifiers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 447, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 579 + ], + "type": "text", + "content": "Confidence Score Test. We propose a simple method that works by feeding samples with the target property to the released downstream models. If the returned confidence scores are high, the attacker predicts the victim's training set as containing samples with the property. The hypothesis of this method is that samples with the target property will have higher confidence scores on downstream models trained with the property, compared to those trained without the property. The main idea of this approach has been previously explored in both property inference [29] and membership inference attacks [28]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Black-box Meta-classifier. We adapt the black-box meta-classifier proposed by Zhang et al. [41]. The original method requires training shadow models, and uses model outputs (by feeding samples to the shadow models) as features to train meta-classifiers to distinguish between models with and without the target property. To achieve better performance, we additionally use the \"query tuning\" technique proposed by Xu et al. [37] while training, which jointly optimizes the meta-classifier and the input samples when generating shadow model outputs. Figure 13 in the appendix shows the benefit of \"query tuning\"." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15978" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 154, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 154, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 154, + 83 + ], + "type": "text", + "content": "5.2. White-Box Access" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "type": "text", + "content": "For adversaries with white-box access, there are two cases depending on if the attacker knows the initialization of the parameters of newly added downstream layers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 131, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 131, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 131, + 287, + 251 + ], + "type": "text", + "content": "Parameter Difference Test (known initialization). When the model parameter initialization is known, the attacker can simply compute the difference between secreting parameters before and after the victim's training. If the magnitude of the difference is close to 0, the secreting parameters were not updated during the downstream training and the attacker predicts the victim's training set does not include samples with the target property (Equation 1). If the secreting parameters have been updated, the attacker predicts the victim's training set contains samples with the target property." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 411 + ], + "type": "text", + "content": "Variance Test (unknown initialization). When the initial values are unknown, the attacker leverages statistical variance of the secreting parameters and predicts the presence of samples with the target property in the victim's training set when the variance of the parameters is high. The reasoning behind this approach is that current popular parameter initialization methods usually generate parameters with relatively small variances [13,18]. If the victim's data contains samples with the target property, the secreting parameters would be updated with gradients of relatively large values (controlled by " + }, + { + "bbox": [ + 46, + 255, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 255, + 287, + 411 + ], + "type": "text", + "content": " in Equation 3), and increase the variance of those parameters in the final model. We confirm this hypothesis empirically in Section 7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 643 + ], + "type": "text", + "content": "White-Box Meta-Classifier. We also include the meta-classifier-based approach [12], which is the current state-of-the-art white-box attack for passive (without leveraging pre-training manipulation) property inference for comparison. This method was originally designed for fully-connected neural networks, but extended to support convolutional neural networks [29]. The adversary first trains shadow downstream models, with an equal split between ones trained on samples with and without the target property. Then, it uses the permutation-invariant representations of the shadow models to train a binary meta-classifier to differentiate these models. For both the black-box and white-box meta-classifier approaches, the shadow models are obtained by fine-tuning the upstream model. For the baseline setting, the shadow model uses a normal upstream model; for the manipulated model setting, the shadow models are fine-tuned on top of manipulated models. Therefore, attacks in the latter setting may gain some advantage from manipulation compared to attacks in the former setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 656, + 170, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 170, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 170, + 670 + ], + "type": "text", + "content": "6. Experimental Design" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "This section explains our experimental setup. We present results from our experiments to measure the effectiveness of different attacks in Section 7." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 324 + ], + "type": "text", + "content": "Tasks and Models. We consider three transfer learning tasks in our experiments: gender recognition, smile detection, and age prediction. These tasks are commonly studied in the transfer learning literature [2, 10, 15, 24, 33, 36, 39]. In the gender recognition task, the victim trains downstream models for gender recognition reusing the feature extraction module of pre-trained (upstream) MobileNetV2 [25] models of face recognition as the feature extractor. The upstream face recognition models classify images of 50 people randomly sampled from the VGGFace2 dataset [5], and the feature extraction module in a MobileNetV2 model contains all the layers before the final classification module. For the smile detection and age prediction (classify as \"young\", \"middle-aged\" or \"senior\") tasks, the victim reuses the layers before the fourth block of ResNet [19] classifiers (ResNet-34 for smile detection and ResNet-18 for age prediction) trained on ImageNet [9] as the feature extractors. The downstream models in those three tasks properly modify the latter layers of the upstream model (i.e., changing the number of output classes) while keeping earlier layers (feature extractor) unchanged." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 327, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 435 + ], + "type": "text", + "content": "Upstream and Downstream Training. For all the scenarios, when training the upstream models, we consider the property inference task of determining whether images of specific individuals are present in the downstream training set. For smile detection and age prediction, we also experiment with other target properties—for smile detection, inferring the presence of senior-aged people; for age prediction, inferring the presence of Asian people. Appendix A.2 provides more details about the upstream training," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 436, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 591 + ], + "type": "text", + "content": "We conduct the downstream training on VGGFace2 with the attribute labels provided by MAADFace [31, 32]. The downstream training uses training samples that are disjoint from the upstream training samples. In our experiments, we consider different sizes (5000 and 10000) of downstream sets with different numbers (chosen from " + }, + { + "bbox": [ + 304, + 436, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\{0, 1, 2, 3, 4, 5, 10, 20, 50, 100, 150\\}" + }, + { + "bbox": [ + 304, + 436, + 545, + 591 + ], + "type": "text", + "content": " with 0 being the reference group for computing the AUC scores of other attack settings) of samples that have the target property (for a total of " + }, + { + "bbox": [ + 304, + 436, + 545, + 591 + ], + "type": "inline_equation", + "content": "2 \\times 11 = 22" + }, + { + "bbox": [ + 304, + 436, + 545, + 591 + ], + "type": "text", + "content": " different settings). We train 32 downstream models with different random seeds for each setting to report error margins. Appendix A.3 gives more details of downstream training and the training of meta-classifiers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 595, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 643 + ], + "type": "text", + "content": "Attack Evaluation Metric. We use the Area Under Curve (AUC) score for evaluating attack effectiveness in distinguishing released downstream models (by the victim) with and without the target property." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 656, + 496, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 496, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 496, + 669 + ], + "type": "text", + "content": "7. Evaluation of Attack Effectiveness" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "Figure 1 summarizes our results. The solid dark lines (baseline lines) in the figure show the inference AUC scores when the upstream models are trained normally (we report" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15979" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 73, + 492, + 186 + ], + "blocks": [ + { + "bbox": [ + 96, + 73, + 492, + 186 + ], + "lines": [ + { + "bbox": [ + 96, + 73, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 96, + 73, + 492, + 186 + ], + "type": "image", + "image_path": "8b3acca26f340d218e3467d7c0a5c85b485aee4d756fa67d0c98bb809c263806.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 278 + ], + "type": "text", + "content": "Figure 1. Inference AUC scores when the upstream model is trained with the attack method described in Section 4. Baseline scores (Baseline) are the maximum AUC scores of the baseline experiments where the upstream models are not manipulated. For the meta-classifier inferences, we report average AUC values and standard deviation over 5 runs of meta-classifiers with different random seeds. In the gender recognition task, the downstream part model " + }, + { + "bbox": [ + 46, + 190, + 547, + 278 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 46, + 190, + 547, + 278 + ], + "type": "text", + "content": " only contains the final classification module, and the downstream trainer cannot reuse the parameters from the upstream model for that module since the numbers of output classes are different. Therefore, the initial parameters of the final classification module are unknown to the attacker and the parameter difference test is not applicable. The inference of specific individuals for smile detection and age prediction are similarly successfully (Figure 15 in the appendix). The downstream training sets contain 10 000 samples and inference results of 5 000 samples are similar and given in Figure 12 in the appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "type": "text", + "content": "the best results of all tested attacks). More details of the baseline experiments can be found in Appendix A.4. Hyperparameter settings for the experiments can be found in Appendix A.5 and the results are insensitive to the selection to hyperparameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "text", + "content": "In all settings except the age prediction with 150 samples of target property, the AUC scores are less than 0.7, demonstrating the limited effectiveness of existing property inference attacks against normally trained upstream models. In contrast, training models with the zero-activation manipulation greatly improves the performance of property inference while having limited impact on the model performance in all settings—the model accuracy drops by at most " + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "text", + "content": " (see Appendix A.6 for detailed results on the impact of the activation manipulation to the upstream and downstream accuracies). Compared to the baseline results which reveal little if any actionable inference (most AUC scores " + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "inline_equation", + "content": "< 0.7" + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "text", + "content": "), manipulating the upstream training with the zero-activation attack improves the effectiveness of property inference significantly, even when only a few downstream training samples have the property. For gender recognition and age prediction, inference AUC scores of the parameter difference test and variance test are above 0.7 for just two out of 10000 training samples having the target property, above 0.9 for 10 training samples, and exceed 0.95 for " + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\geq 20" + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "text", + "content": " training samples. The one exception also has AUC scores exceed 0.9 for " + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\geq 20" + }, + { + "bbox": [ + 46, + 346, + 288, + 610 + ], + "type": "text", + "content": " training samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 613, + 287, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 698 + ], + "type": "text", + "content": "Black-box attacks. The black-box meta-classifier achieves inference AUC scores above 0.9 when " + }, + { + "bbox": [ + 46, + 613, + 287, + 698 + ], + "type": "inline_equation", + "content": "\\geq 50" + }, + { + "bbox": [ + 46, + 613, + 287, + 698 + ], + "type": "text", + "content": " out of 10000 training samples have the target property. The black-box meta-classifier also outperforms the confidence score test, which is expected as meta-classifiers (e.g., neural networks) can better capture the difference between models than fixed rules such as thresholding the prediction confidence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "White-box attacks. Our white-box methods (the parame" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "type": "text", + "content": "ter difference test and the variance test) also achieve AUC scores " + }, + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "type": "inline_equation", + "content": ">0.9" + }, + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "type": "inline_equation", + "content": "\\geq 20" + }, + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "type": "text", + "content": " training samples are with the target property. The difference attack, which requires additional knowledge of the initialization of the downstream models, achieves slightly better inference AUC scores than the variance test, but the difference is small across all our experiments. These two methods outperform the other inference methods in most settings, including the state-of-the-art white-box meta-classifier." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 396, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 540 + ], + "type": "text", + "content": "White-box meta-classifier vs. Black-box meta-classifier. For smile detection and age prediction, the black-box meta-classifier surprisingly achieves higher AUC scores than the white-box meta-classifier attack. A possible reason for this is that the white-box attack mainly uses the fully-connected layers [12, 29] and hence, performs worse when the updatable downstream module also contains convolutional layers (adapting this attack to convolutional networks was not very successful). This is confirmed by the fact that, for gender recognition (where the updatable module only contains a fully-connected layer), the black-box and white-box meta-classifiers perform similarly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 544, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 641 + ], + "type": "text", + "content": "Attacks of AUC scores " + }, + { + "bbox": [ + 304, + 544, + 547, + 641 + ], + "type": "inline_equation", + "content": "< 0.5" + }, + { + "bbox": [ + 304, + 544, + 547, + 641 + ], + "type": "text", + "content": ". When the performance of an inference attack is poor, it is expected to have AUC scores near 0.5 (close to random guessing). However, we find that there are few attack settings with AUC scores consistently below 0.5. Appendix A.10 discusses those anomalies and surmises that they are caused by the limitations of original inference methods designed for normal pretrained models when facing challenging inference tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 656, + 444, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 444, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 444, + 670 + ], + "type": "text", + "content": "8. Stealthier Manipulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "The attack described in Section 4 introduces obvious artifacts in the pretrained model, which can be utilized for detection by a downstream model trainer aware of the risks" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15980" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "posed by our attacks. We first present two detection methods (Section 8.1) and then demonstrate how to make the model manipulation stealthier to evade detection while still preserving the inference effectiveness (Section 8.2 and Section 8.3). We assume the downstream trainer is aware of the possibility of the attack and its design, but does not know the property targeted by the adversary, as this is specific to an attacker's goal and the set of possible properties can be exponentially large for a rich training set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 186, + 268, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 268, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 268, + 201 + ], + "type": "text", + "content": "8.1. Detecting Manipulated Pretrained Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 205, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 242 + ], + "type": "text", + "content": "We present two detection methods that use the distributional difference between activations of samples with and without property." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 289, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 424 + ], + "type": "text", + "content": "Checking the Distribution of Activations. Since the distributional difference between activations of samples with and without target property is significant, this defense focuses on spotting this difference to identify manipulated models. A method to identify the distributional difference needs to be designed based on the attack method used. For the original zero-activation attacks in Section 4.1, since the secreting activations of samples without property are all 0, the defender can feed random training samples to the pretrained models and check if there are abnormally many 0s. This approach is feasible since samples of target property have limited presence in the downstream training set and hence, most samples will not have the property. Since detecting the zero-activation attack is trivial using this method, we do not conduct any experiments with this." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 289, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 642 + ], + "type": "text", + "content": "Anomaly Detection. Since the target property has a limited presence in the downstream training set, another defense would be treating samples with the target property as outliers and then analyzing those outliers to find manipulations. Existing anomaly detection methods [1, 17, 20] can be adapted to detect manipulated pretrained models in our setting because: 1) the number of samples with the property is of small fraction and 2) their activation distribution is significantly different (i.e., outliers) from the distribution for samples without the property. The auditor can inspect model activations for all of its training data and identify outliers (ideally, samples with target property) with anomaly detection. The auditor can then inspect identified outliers and may find commonalities to identify the potential target property. For instance, they may find that a small fraction of the training data produce unusual model activations, and then notice that most of that data has a particular property such as belonging to a specific individual or group." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "We consider three common anomaly detection methods: K-means [20], PCA [1] and Spectre [17] (where Spectre is the current state-of-the-art) and we report the detection results from the three defenses. Appendix A.12 gives details of these methods. The detection results on the zero-activation attack are given in Figure 11 in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "Anomaly detection is very effective at identifying the samples with target property. For example, for the gender recognition and smile detection tasks, the detection rate is over " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " in most cases. These results motivate the design of stealthier attacks which we describe next." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 140, + 474, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 474, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 474, + 155 + ], + "type": "text", + "content": "8.2. Stealthier Model Manipulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 159, + 547, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 303 + ], + "type": "text", + "content": "To evade the defense that checks the distribution of activations, we modify our zero-activation attack to ensure: (1) secreting activations for samples without the property are also non-zero (bypassing simple defense of checking abnormal zeros); (2) secreting activations of samples with and without target property are still distinct (the attack is still effective); (3) that distinction between activations should not be captured by anomaly detection methods (evading anomaly detection); (4) the actual distribution of activations that matches the attacker's goal cannot be easily guessed by the defender (handling cases when the defender actively searches other patterns in the distribution of activations)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 304, + 521, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 304, + 521, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 521, + 316 + ], + "type": "text", + "content": "For (1) and (2), we adapt the loss in Equation 3 as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 319, + 545, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 319, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 319, + 545, + 352 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\alpha \\cdot \\max (\\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 0 \\\\ \\beta \\cdot \\max (\\lambda \\cdot \\| f (\\boldsymbol {x}) \\circ \\neg \\boldsymbol {m} \\| - \\| f (\\boldsymbol {x}) \\circ \\boldsymbol {m} \\|, 0) & \\text {i f} y _ {t} = 1 \\end{array} \\right. \\tag {4}", + "image_path": "b0ec1fccc95a7f8047c6f42185e95df28025b2869a72d4c4f339fc3e00b23d54.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\lambda \\geq 1" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": ". (1): The case of " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "y_{t} = 0" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": " is redefined to bypass the detection of abnormal zeros. Minimizing this new loss ensures that samples without the target property will have secreting activations " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "(f(\\pmb{x}) \\circ \\pmb{m})" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": " with (close-to-normal) nonzero values. (2): to ensure the property is still detectable, we actively increase the difference between the secreting activations of samples with and without property. We observe that, for upstream models with reasonable performance on the main task, non-secreting activations " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "(f(\\pmb{x}) \\circ \\neg \\pmb{m})" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": " have similar amplitude regardless of the fed samples containing target property. Therefore, for samples with target property, as long as we ensure the secreting activations have a larger amplitude than that of non-secreting activations, there will be a distinction between secreting activations of samples with and without property. We do this by assigning larger values to " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\lambda \\geq 1" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": ", instead of the original " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": ") for the second line of Equation 4 to induce sharper distinction between samples with and without property and enable higher inference performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "To prevent detection by anomaly detectors (requirement (3) above), " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " should be set to balance the attack effectiveness and stealthiness rightly. By choosing proper values for " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", our attack is able to evade anomaly detection methods in most settings. However, in some settings (mostly in gender recognition tasks), state-of-the-art anomaly detection (Spectre) can still identify most of the samples with target property. To counter this, we add an additional regularization term (weighted by parameter " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ") to the overall loss function " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "l(x,y,y_{t})" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " in Equation 2 that further improves attack stealthiness while still maintaining relatively high at" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15981" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 73, + 494, + 196 + ], + "blocks": [ + { + "bbox": [ + 96, + 73, + 494, + 196 + ], + "lines": [ + { + "bbox": [ + 96, + 73, + 494, + 196 + ], + "spans": [ + { + "bbox": [ + 96, + 73, + 494, + 196 + ], + "type": "image", + "image_path": "089249ae14be583964904e99df31c3096ea30968f85dcb24f9334dbbcdd471d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 199, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 547, + 243 + ], + "type": "text", + "content": "Figure 2. Inference AUC scores of the stealthier design. Since the secreting activations are no longer zero, the inference methods based on difference or variance tests are no longer applicable. The inference results of specific individuals for smile detection and age prediction also show similar improvement compared to the baseline settings (Figure 19 in the appendix). The downstream training sets contain 10000 samples and inference results results of 5000 samples are similar and given in Figure 17 in the appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": "tack effectiveness. Specifically, we first obtain the corresponding covariance matrices of the activations of samples with the target property " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "(cov_w)" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": ", activations of all samples with and without the target property " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "(cov_{w,wo})" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": ", and activations of samples without the target property " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "(cov_{wo})" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": " respectively. Then, we encourage mean " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "(cov_w) = mean(cov_{w,wo}) = mean(cov_{wo})" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "var(cov_w) = var(cov_{w,wo}) = var(cov_{wo})" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": " (both mean " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "var(\\cdot)" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": " treat the whole covariance matrix as a flattened array and return scalar values) for the three covariance matrices by minimizing their differences in their mean and variance. Using this method, we ensure the distributions of activations of samples with target property will be similar to the ones without the property, making the manipulations harder to detect. We use this approach for all the experiments. To ensure the distributional pattern related to the attacker goal cannot be easily guessed (requirement (4)), we generate " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": " randomly (instead of picking first " + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "inline_equation", + "content": "\\| m\\|" + }, + { + "bbox": [ + 46, + 250, + 289, + 490 + ], + "type": "text", + "content": " activations in Section 7). This makes the brute-force search of possible patterns computationally infeasible (details in Appendix A.14)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 497, + 233, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 233, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 233, + 510 + ], + "type": "text", + "content": "8.3. Experiments with Stealthy Attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "content": "Detection Evasion. Figure 16 (in the appendix) summarizes the results of our experiments to detect the stealthy upstream models (Appendix A.13 provides details on these experiments). We find that the anomaly detection methods are ineffective against our stealthier attack—< " + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "content": " of samples with the target property are detected across all settings with the exception of a detection rate " + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "inline_equation", + "content": "< 20\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "content": " (still low) for smile detection when the total number of samples is 5000 and 100 or 150 of them are with the target property. We also made several attempts to approximately identify (instead of brute-force search) possible attack patterns in the activations but none of these succeeded in uncovering the stealthy attacks (details are in Appendix A.14)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Inference Results. From Figure 2, we can see that activation manipulation still leads to significantly improved inference results compared to the baselines with normally" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "text", + "content": "trained upstream models. For example, for gender recognition, when " + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "inline_equation", + "content": "\\geq 50" + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "text", + "content": " downstream training samples have the target property, inference AUC scores exceed 0.95, which is a huge improvement compared to the baseline attack where all AUC scores are less than 0.6, and similar trends follow for smile detection (with over 100 samples with property, AUC improves from " + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "inline_equation", + "content": "< 0.6" + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "inline_equation", + "content": "> 0.78" + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "text", + "content": ") and age prediction (with over 100 samples with property, AUC improves from " + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "inline_equation", + "content": "< 0.77" + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "inline_equation", + "content": "> 0.9" + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "text", + "content": "). Comparing the results for the stealthier attacks to the results that do not consider defenses in Figure 1, we observe that the attack effectiveness declines as expected since we are now trading-off attack effectiveness for stealthiness. Training models with the attack goal poses negligible impact on the model performance (accuracy drop " + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "inline_equation", + "content": "< 0.9\\%" + }, + { + "bbox": [ + 304, + 250, + 547, + 430 + ], + "type": "text", + "content": ", see Appendix A.6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 444, + 379, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 379, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 379, + 456 + ], + "type": "text", + "content": "9. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 465, + 547, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 547, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 547, + 597 + ], + "type": "text", + "content": "Our work demonstrates how a malicious upstream trainer can manipulate its training process to amplify property inference risks for downstream models when transfer learning is done. Our empirical results show that such manipulations can be exploited to enable very precise property inference, even in black-box settings, across a variety of tasks. Although there is potential for a new arms race between methods of hiding manipulations and methods of detecting them, the larger lesson from this work, and other works exposing similar risks, is that it is important that users of pretrained models to only use models from trusted providers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 611, + 409, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 409, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 409, + 624 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (#2022YFF0604503 and #2021YFB3100300), the United States National Science Foundation through the Center for Trustworthy Machine Learning (#1804603), NSFC (#62272224), JiangSu Province Science Foundation for Youths (#BK20220772), and Lockheed Martin Corporation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15982" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Hervé Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 7, 19" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 178 + ], + "type": "text", + "content": "[2] MAH Akhand, Ijaj Sayim, Shuvendu Roy, N Siddique, et al. Human Age Prediction from Facial Image Using Transfer Learning in Deep Convolutional Neural Networks. In International Joint Conference on Computational Intelligence, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "type": "text", + "content": "[3] Giuseppe Ateniese, Luigi V Mancini, Angelo Spognardi, Antonio Villani, Domenico Vitali, and Giovanni Felici. Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifiers. International Journal of Security and Networks, 10(3):137-150, 2015. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 290 + ], + "type": "text", + "content": "[4] Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, and Nicolas Papernot. When the curious abandon honesty: Federated learning is not private. arXiv preprint arXiv:2112.02918, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "text", + "content": "[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. In IEEE International Conference on Automatic Face & Gesture Recognition, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 388 + ], + "type": "text", + "content": "[6] Shuvam Chakraborty, Burak Uzkent, Kumar Ayush, Kumar Tanmay, Evan Sheehan, and Stefano Ermon. Efficient Conditional Pre-training for Transfer Learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 423 + ], + "type": "text", + "content": "[7] Melissa Chase, Esha Ghosh, and Saeed Mahloujifar. Property Inference from Poisoning. In IEEE Symposium on Security and Privacy, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "text", + "content": "[8] Harsh Chaudhari, Jackson Abascal, Alina Oprea, Matthew Jagielski, Florian Tramèr, and Jonathan Ullman. SNAP: Efficient Extraction of Private Properties with Poisoning. arXiv preprint arXiv:2208.12348, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 512 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009. 2, 5, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 286, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 545 + ], + "type": "text", + "content": "[10] Fadi Dornaika, Ignacio Arganda-Carreras, and C Belver. Age estimation in facial images through transfer learning. Machine Vision and Applications, 30(1):177-187, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "type": "text", + "content": "[11] Liam Fowl, Jonas Geiping, Wojtek Czaja, Micah Goldblum, and Tom Goldstein. Robbing the fed: Directly obtaining private data in federated learning with modified models. arXiv preprint arXiv:2110.13057, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 646 + ], + "type": "text", + "content": "[12] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property Inference Attacks on Fully Connected Neural Networks using Permutation Invariant Representations. In ACM SIGSAC Conference on Computer and Communications Security, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics, 2010. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[14] Kathrin Grosse, Thomas A Trost, Marius Mosbach, and Michael Backes. Adversarial initialization-when your" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "network performs the way I want. arXiv preprint arxiv:1902.03020, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "text", + "content": "[15] Xin Guo, Luisa Polania, and Kenneth Barner. Smile Detection in the Wild Based on Transfer Learning. In IEEE International Conference on Automatic Face & Gesture Recognition, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "text", + "content": "[16] Valentin Hartmann, L'eo Meynent, Maxime Peyrard, Dimitrios Dimitriadis, Shruti Tople, and Robert West. Distribution inference risks: Identifying and mitigating sources of leakage. arXiv preprint arXiv:2209.08541, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "type": "text", + "content": "[17] Jonathan Hayase, Weihao Kong, Raghav Somani, and Sewoong Oh. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics. In International Conference on Machine Learning, 2021. 7, 19, 20, 23" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 235, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 235, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 235, + 545, + 277 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. In IEEE International Conference on Computer Vision, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 326, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 358 + ], + "type": "text", + "content": "[20] Anil K Jain, M Narasimha Murty, and Patrick J Flynn. Data Clustering: A Review. ACM computing surveys (CSUR), 31(3):264-323, 1999. 7, 19" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[21] Joanna Jaworek-Korjakowska, Pawel Kleczek, and Marek Gorgon. *Melanoma Thickness Prediction Based on Convolutional Neural Network with VGG-19 Model Transfer Learning*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 418, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 418, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 545, + 449 + ], + "type": "text", + "content": "[22] Marc Juárez, Samuel Yeom, and Matt Fredrikson. BlackBox Audits for Group Distribution Shifts. arXiv preprint arXiv:2209.03620, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "type": "text", + "content": "[23] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting Unintended Feature Leakage in Collaborative Learning. In IEEE Symposium on Security and Privacy, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 498, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 541 + ], + "type": "text", + "content": "[24] Cao Hong Nga, Khai-Thinh Nguyen, Nghi C Tran, and Jia-Ching Wang. Transfer Learning for Gender and Age Prediction. In IEEE International Conference on Consumer Electronics-Taiwan (ICCE-Taiwan), 2020. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "type": "text", + "content": "[25] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetV2: Inverted Residuals and Linear Bottlenecks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "type": "text", + "content": "[26] Roei Schuster, Tal Schuster, Yoav Meri, and Vitaly Shmatikov. Humpty Dumpty: Controlling Word Meanings via Corpus Poisoning. In IEEE Symposium on Security and Privacy, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "type": "text", + "content": "[27] Ozan Sener and Vladlen Koltun. Multi-Task Learning as Multi-Objective Optimization. In Advances in Neural Information Processing Systems, 2018. 13" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[28] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy, 2017. 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15983" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[29] Anshuman Suri and David Evans. Formalizing and Estimating Distribution Inference Risks. In Privacy Enhancing Technologies Symposium, 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[30] Anshuman Suri, Pallika Kanani, Virendra J Marathe, and Daniel W Peterson. Subject Membership Inference Attacks in Federated Learning. arXiv preprint arXiv:2206.03317, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[31] Philipp Terhörst, Daniel Fährmann, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. MAAD-Face: A Massively Annotated Attribute Dataset for Face Images. IEEE Transactions on Information Forensics and Security, 16:3942-3957, 2021. 5, 11, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 274 + ], + "type": "text", + "content": "[32] Philipp Terhörst, Marco Huber, Jan Niklas Kolf, Ines Zelch, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Reliable Age and Gender Estimation from Face Images: Stating the Confidence of Model Predictions. In IEEE International Conference on Biometrics Theory, Applications and Systems (BTAS), 2019. 5, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[33] Bolun Wang, Yuanshun Yao, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. With Great Training Comes Great Vulnerability: Practical Attacks against Transfer Learning. In USENIX Security Symposium, 2018. 1, 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "text", + "content": "[34] Xiuling Wang and Wendy Hui Wang. Group Property Inference Attacks Against Graph Neural Networks. arXiv preprint arXiv:2209.01100, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[35] Yuxin Wen, Jonas A Geiping, Liam Fowl, Micah Goldblum, and Tom Goldstein. Fishing for user data in large-batch federated learning via gradient magnification. In International Conference on Machine Learning, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "text", + "content": "[36] Yu Xia, Di Huang, and Yunhong Wang. Detecting Smiles of Young Children via Deep Transfer Learning. In IEEE International Conference on Computer Vision Workshops, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "text", + "content": "[37] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting AI Trojans Using Meta Neural Analysis. In IEEE Symposium on Security and Privacy, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "type": "text", + "content": "[38] Kaiyu Yang, Jacqueline Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky. A Study of Face Obfuscation in ImageNet. In International Conference on Machine Learning, 2022. 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "text", + "content": "[39] Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y Zhao. Latent Backdoor Attacks on Deep Neural Networks. In ACM SIGSAC Conference on Computer and Communications Security, 2019. 1, 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond Empirical Risk Minimization. In International Conference on Learning Representations, 2018. 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[41] Wanrong Zhang, Shruti Tople, and Olga Ohrimenko. Leakage of Dataset Properties in Multi-Party Machine Learning. In USENIX Security Symposium, 2021. 1, 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[42] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A Comprehensive Survey on Transfer Learning. Proceedings of the IEEE, 109, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[43] Yang Zou, Zhikun Zhang, Michael Backes, and Yang Zhang. Privacy Analysis of Deep Learning in the Wild: Membership" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "Inference Attacks against Transfer Learning. arXiv preprint arXiv:2009.04872, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15984" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_content_list.json b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..081f7758f2797e33224c182f2531a0749dcd92fe --- /dev/null +++ b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_content_list.json @@ -0,0 +1,1586 @@ +[ + { + "type": "text", + "text": "Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision", + "text_level": 1, + "bbox": [ + 84, + 128, + 887, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyi Ying $^{1}$ , Li Liu $^{1}$ , Yingqian Wang $^{1}$ , Ruojing Li $^{1}$ , Nuo Chen $^{1}$ , Zaiping Lin $^{1\\boxtimes}$ , Weidong Sheng $^{1}$ , Shilin Zhou $^{1}$", + "bbox": [ + 166, + 202, + 803, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Defense Technology", + "bbox": [ + 308, + 238, + 661, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yingxinyi18, wangyingqian16, liruojing, chennuo97, linzaiping, slzhou}@nudt.edu.cn,", + "bbox": [ + 138, + 256, + 831, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dreamliu2010@gmail.com, shengweidong1111@sohu.com", + "bbox": [ + 246, + 273, + 718, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training a convolutional neural network (CNN) to detect infrared small targets in a fully supervised manner has gained remarkable research interests in recent years, but is highly labor expensive since a large number of per-pixel annotations are required. To handle this problem, in this paper, we make the first attempt to achieve infrared small target detection with point-level supervision. Interestingly, during the training phase supervised by point labels, we discover that CNNs first learn to segment a cluster of pixels near the targets, and then gradually converge to predict groundtruth point labels. Motivated by this \"mapping degeneration\" phenomenon, we propose a label evolution framework named label evolution with single point supervision (LESPS) to progressively expand the point label by leveraging the intermediate predictions of CNNs. In this way, the network predictions can finally approximate the updated pseudo labels, and a pixel-level target mask can be obtained to train CNNs in an end-to-end manner. We conduct extensive experiments with insightful visualizations to validate the effectiveness of our method. Experimental results show that CNNs equipped with LESPS can well recover the target masks from corresponding point labels, and can achieve over $70\\%$ and $95\\%$ of their fully supervised performance in terms of pixel-level intersection over union (IoU) and object-level probability of detection $(P_d)$ , respectively. Code is available at https://github.com/XinyiYing/LESPS.", + "bbox": [ + 76, + 358, + 473, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 782, + 209, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infrared small target detection has been a longstanding, fundamental yet challenging task in infrared search and tracking systems, and has various important applications in civil and military fields [49,57], including traffic monitoring", + "bbox": [ + 75, + 806, + 470, + 868 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/311a0b588c2647d95cc1bfc3f441ed182145828ef7344030fab3667934b3f1fc.jpg", + "image_caption": [ + "Figure 1. An illustration of mapping degeneration under point supervision. CNNs always tend to segment a cluster of pixels near the targets with low confidence at the early stage, and then gradually learn to predict GT point labels with high confidence." + ], + "image_footnote": [], + "bbox": [ + 514, + 325, + 875, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[24, 54], maritime rescue [52, 53] and military surveillance [7, 47]. Due to the rapid response and robustness to fast-moving scenes, single-frame infrared small target (SIRST) detection methods have always attracted much more attention, and numerous methods have been proposed. Early methods, including filtering-based [9, 40], local contrast-based [3, 16] and low rank-based [11, 43] methods, require complex handcrafted features with carefully tuned hyper-parameters. Recently, compact deep learning has been introduced in solving the problem of SIRST detection [24, 45, 54]. However, there are only a few attempts, and its potential remains locked, unlike the extensive explorations of deep learning for natural images. This is mainly due to potential reasons, including lack of large-scale, accurately annotated datasets and high stake application scenarios.", + "bbox": [ + 496, + 551, + 890, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infrared small targets are usually of very small size, weak, shapeless and textureless, and are easily submerged in diverse complex background clutters. As a result, directly adopting existing popular generic object detectors like RCNN series [13, 14, 19, 39], YOLO series [25, 37, 38] and SSD [29] to SIRST detection cannot produce satisfactory performance. Realizing this, researchers have been focusing on developing deep networks tailored for", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported by National Key Research and", + "bbox": [ + 75, + 873, + 446, + 885 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Development Program of China No. 2021YFB3100800.", + "bbox": [ + 78, + 886, + 375, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15528", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "infrared small targets by adequately utilizing the domain knowledge. However, most existing deep methods for SIRST detection [8, 24, 54] are fully supervised, which usually requires a large dataset with accurate target mask annotations for training. Clearly, this is costly [5, 26].", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, a natural question arises: Can we develop a new framework for SIRST detection with single point supervision? In fact, to substantially reduce the annotation cost for object detection tasks, weakly supervised object detection methods with point supervision [4,5,26,56] have been studied in the field of computer vision. Although these weakly supervised methods achieve promising results, they are not designed for the problem of SIRST detection, and the class-agnostic labels (i.e., only foreground and background) of infrared small targets hinder their applications [42, 58]. Therefore, in this work, we intend to conduct the first study of weakly supervised SIRST detection with single-point supervision.", + "bbox": [ + 75, + 167, + 470, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key motivation of this work comes from an interesting observation during the training of SIRST detection networks. That is, with single point labels serving as supervision, CNNs always tend to segment a cluster of pixels near the targets with low confidence at the early stage, and then gradually learn to predict groundtruth (GT) point labels with high confidence, as shown in Fig. 1. It reveals the fact that region-to-region mapping is the intermediate result of the final region-to-point mapping1. We attribute this \"mapping degeneration\" phenomenon to the special imaging mechanism of infrared system [24, 54], the local contrast prior of infrared small targets [3, 8], and the easy-to-hard learning property of CNNs [44], in which the first two factors result in extended mapping regions beyond the point labels, and the last factor contributes to the degeneration process.", + "bbox": [ + 75, + 364, + 470, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the aforementioned discussion, in this work, we propose a novel framework for the problem of weakly supervised SIRST detection, dubbed label evolution with single point supervision (LESPS). Specifically, LESPS leverages the intermediate network predictions in the training phase to update the current labels, which serve as supervision until the next label update. Through iterative label update and network training, the network predictions can finally approximate the updated pseudo mask labels, and the network can be simultaneously trained to achieve pixel-level SIRST detection in an end-to-end $^2$ manner.", + "bbox": [ + 75, + 606, + 468, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as: (1) We present the first study of weakly supervised SIRST", + "bbox": [ + 76, + 773, + 468, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "detection, and introduce LESPS that can significantly reduce the annotation cost. (2) We discover the mapping degeneration phenomenon, and leverage this phenomenon to automatically regress pixel-level pseudo labels from the given point labels via LESPS. (3) Experimental results show that our framework can be applied to different existing SIRST detection networks, and enable them to achieve over $70\\%$ and $95\\%$ of its fully supervised performance in terms of pixel-level intersection over union $(IoU)$ and object-level probability of detection $(P_d)$ , respectively.", + "bbox": [ + 500, + 90, + 890, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 256, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SIRST Detection. In the past decades, various methods have been proposed, including early traditional paradigms (e.g., filtering-based methods [9, 40], local contrast-based methods [3, 15-17, 33, 34], low rank-based methods [6, 11, 28, 43, 50, 51]) and recent deep learning paradigms [7, 8, 20, 21, 24, 45, 52-54]. Compared to traditional methods, which require delicately designed models and carefully tuned hyper-parameters, convolutional neural networks (CNNs) can learn the non-linear mapping between input images and GT labels in a data-driven manner, and thus generalize better to real complex scenes. As the pioneering work, Wang et al. [45] first employed a generative adversarial network to achieve a better trade-off between miss detection and false alarm. Recently, more works focus on customized solutions of infrared small target. Specifically, Dai et al. [7] specialized an asymmetric contextual module, and further incorporated local contrast measure [8] to improve the target contrast. Li et al. [24] preserved target information by repetitive feature fusion. Zhang et al. [54] aggregated edge information to achieve shape-aware SIRST detection. Zhang et al. [52, 53] explored cross-level correlation and transformer-based method [18] to predict accurate target mask. Wu et al. [46] customized a UIU-Net framework for multi-level and multi-scale feature aggregation. In conclusion, existing works generally focus on compact architectural designs to pursue superior performance in a fully supervised manner. However, due to the lack of a large number of public datasets [7, 24, 45] with per-pixel annotations, the performance and generalization of CNNs are limited. In addition, per-pixel manual annotations are time-consuming and labor-intensive. Therefore, we focus on achieving good pixel-level SIRST detection with weaker supervision and cheaper annotations.", + "bbox": [ + 496, + 280, + 893, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly Supervised Segmentation with Points. Recently, point-level annotation has raised more attention in dense prediction tasks such as object detection [4, 12, 56], crowd counting [1, 23, 30, 48] and image segmentation [2,5,10,26,31,36,55]. We mainly focus on image segmentation in this paper. Specifically, Bearman et al. [2] made the first attempt to introduce an objectiveness potential into a pointly supervised training", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 \"region-to-region mapping\" represents the mapping learned by CNNs from target regions in images to a cluster of pixels near the targets, while \"region-to-point mapping\" represents the mapping from target regions in images to the GT point labels. 2 Different from generic object detection [32, 59], \"end-to-end\" here represents achieving point-to-mask label regression and direct pixel-level inference in once training.", + "bbox": [ + 75, + 815, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15529", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "loss function to boost segmentation performance. Qian et al. [36] leveraged semantic information of several labeled points by a distance metric loss to achieve scene parsing. Zhang et al. [55] proposed an inside-outside guidance approach to achieve instance segmentation by five elaborate clicks. Cheng et al. [5] designed to provide ten randomly sampled binary point annotations within box annotations for instance segmentation. Li et al. [26] encoded each instance with kernel generator for panoptic segmentation to achieve $82\\%$ of fully-supervised performance with only twenty randomly annotated points. In contrast to these approaches employing complicated prior constraints to segment large generic objects with rich color and fine textures by several elaborate points, we fully exploit the local contrast prior of infrared small target to progressively evolve pseudo masks by single coarse point without any auxiliaries in an end-to-end manner.", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Mapping Degeneration Phenomenon", + "text_level": 1, + "bbox": [ + 76, + 361, + 446, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first describe the mapping degeneration phenomenon together with our intuitive explanation. Then we conduct experiments under single-sample and many-sample training schemes to demonstrate the generality of degeneration, and investigate the influence of generalization on degeneration.", + "bbox": [ + 75, + 387, + 470, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, given an input image and the corresponding GT point label, we employ U-Net [41] as the baseline SIRST detection network for training. It can be observed that, in the early training phase, network predicts a cluster of pixels near the targets with low confidence. As training continues, the network prediction finally approximates GT point label with gradually increased confidence. We name this phenomenon as \"mapping degeneration\", and attribute the following reasons to this phenomenon. 1) Special imaging mechanism of infrared systems [24, 54]: Targets only have intensity information without structure and texture details, resulting in highly similar pixels within the target region. 2) High local contrast of infrared small targets [3, 8]: Pixels within the target region are much brighter or darker with high contrast against the local background clutter. 3) Easy-to-hard learning property of CNNs [44]: CNNs always tend to learn simple mappings first, and then converge to difficult ones. Compared with region-to-point mapping, region-to-region mapping is easier, and thus tends to be the intermediate result of region-to-point mapping. In conclusion, the unique characteristics of infrared small targets result in extended mapping regions beyond point labels, and CNNs contribute to the mapping degeneration process.", + "bbox": [ + 75, + 478, + 472, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth noting that the mapping degeneration phenomenon is a general phenomenon in various scenes with infrared small targets. Specifically, we use the training datasets (including 1676 images and their corresponding", + "bbox": [ + 75, + 840, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef05e010f1f4414bbbbc0422c6fa759a2ec3b4e80ecd716c80cf32295d37aade.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 89, + 702, + 166 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b9bac8a255a0d533de981dffd67bf15df661f747fe96041a0b7009c724a6be27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 90, + 893, + 166 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/173cc11f5273b1d88187099bd6fd3dc3167e04e60b4de804b9bc261e789566db.jpg", + "image_caption": [ + "Figure 2. Quantitative and qualitative illustrations of mapping degeneration in CNNs." + ], + "image_footnote": [], + "bbox": [ + 511, + 167, + 893, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "centroid point label, see details in Section 5.1) to train U-Net under a single-sample training scheme (i.e., training one CNN on each image). For quantitative analyses, we employ the IoU results between positive pixels in predictions (i.e., pixels with confidence higher than half of its maximum value) and GT mask label. Average IoU results of 1676 CNNs at each epoch are shown by the blue curve in Fig. 2(a), while the number of training samples with maximum IoU during training phase falling in a threshold range of $[i, i + 0.1]$ , $(i = 0, 0.1, \\dots, 0.9)$ is illustrated via blue bars in Fig. 2(b). It can be observed from the zoom-in curve and bars that mapping degeneration is a general phenomenon with point supervision, and U-Net can achieve IoU $>0.5$ on more than $60\\%$ of the training images.", + "bbox": [ + 496, + 319, + 893, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, we conduct experiments to train U-Net under a many-sample training scheme (i.e., training one CNN using all images which contain abundant targets with various sizes and shapes) to investigate the effect of generalization on mapping degeneration. Average IoU results of 1676 images are shown by orange curve in Fig. 2(a). It can be observed that many-sample training scheme needs more time to converge. Moreover, Fig. 2(b) shows that orange bars are slightly lower than blue ones on larger IoU values (i.e., 0.5-1.0). It is demonstrated that generalization decelerates but aggravates mapping degeneration. Figure 2(c) shows some zoom-in target regions of images and their predictions under these two training schemes. It can be observed that CNNs can effectively segment a cluster of target pixels under both training schemes in a size-aware manner.", + "bbox": [ + 496, + 534, + 895, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, an intuitive assumption arises: Can we leverage the intermediate results of CNNs to regress masks? A simple early stopping strategy seems to be a positive answer but is indeed unpractical since mapping degeneration is influenced by various factors, including target intensity, size, shape, and local background clutter (see details in Section 5.2.1). Consequently, there is no fixed optimal stopping epoch for all situations. These", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15530", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44a112e497ff17aee764208a5f492eb679dc43de3430601165fbda42b8a02291.jpg", + "image_caption": [ + "Figure 3. An illustration of label evolution with single point supervision (LESPS). During training, intermediate predictions of CNNs are used to progressively expand point labels to mask labels. Black arrows represent each round of label updates." + ], + "image_footnote": [], + "bbox": [ + 81, + 75, + 475, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "observations motivate us to design a special label evolution framework to well leverage the mapping degeneration for pseudo mask regression.", + "bbox": [ + 75, + 315, + 470, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. The Label Evolution Framework", + "text_level": 1, + "bbox": [ + 76, + 375, + 377, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by mapping degeneration, we propose a label evolution framework named label evolution with single point supervision (LESPS) to leverage the intermediate network predictions in the training phase to update labels. As training continues, the network predictions approximate the updated pseudo mask labels, and network can simultaneously learn to achieve pixel-level SIRST detection in an end-to-end manner. Here, we employ a toy example of 1D curves for easy understanding. As shown in Fig. 3, sub-figures on the left of the dotted line represent the network predictions. Note that, the black curves denote the intermediate predictions within LESPS, while the gray curves represent virtual results produced by the network without label update. On the right of the dotted line, the first and second columns of sub-figures represent current labels and updated labels, respectively, and black arrows represent each round of label update. The overall framework can be summarized as follows. With point label serving as supervision, in the $1^{st}$ round label update after initial training, the predictions are used to update the current point label to generate the $1^{st}$ updated label, which is then used to supervise the network training until the $2^{nd}$ round label update. Through iterative label updates and network training, CNNs can incorporate the local contrast prior to gradually recover the mask labels. From another viewpoint, label evolution consistently updates the supervision to prevent mapping degeneration, and promotes CNNs to converge to the easy region-to-region mapping.", + "bbox": [ + 75, + 401, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking the $n^{th}$ update as an example, given the current label $L_{n}$ and the network prediction $P_{n}$ , we perform label update for each target, which consists of three steps: candidate pixel extraction, false alarm elimination, and weighted summation between candidate pixels and current", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7c0aa4c6310d39eb21011d872084a1cbb9aab640f0216f50370c0a6f052a8356.jpg", + "image_caption": [ + "Figure 4. (a) Adaptive threshold $T_{\\text{adapt}}$ with respect to positive pixels $\\hat{L}_n^i$ and hyper-parameters $k, T_b$ . Pink dotted line represents the constant $hwr$ . (b) An illustration of false alarm elimination. Red circle and dot represent positive pixels and centroid point of label. Orange circle represents false alarms." + ], + "image_footnote": [], + "bbox": [ + 501, + 75, + 885, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "labels. Specifically, the $d \\times d$ local neighborhoods of the $i^{th}$ target in label $L_{n}$ and prediction $P_{n}$ are cropped based on the centroid of the positive pixels $^{3}$ in label (i.e., $\\hat{L}_n$ ). Then to reduce error accumulation for label update (see Section 5.2.2 for details), we employ an adaptive threshold (the red dotted line in Fig. 3) to extract the local neighborhood candidate pixels (predictions higher than the red dotted line in Fig. 3). The process can be defined as:", + "bbox": [ + 496, + 275, + 892, + 396 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {n} ^ {i} = P _ {n} ^ {i} \\odot \\left(P _ {n} ^ {i} > T _ {\\text {a d a p t}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 404, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C_n^i$ is the candidate pixels, and $\\odot$ represents elementwise multiplication. $T_{adapt}$ is the adaptive threshold that correlated to the current prediction $P_n^i$ and the positive pixels in label $\\hat{L}_n^i$ , and can be calculated according to:", + "bbox": [ + 496, + 430, + 892, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {\\text {a d a p t}} = \\max \\left(P _ {n} ^ {i}\\right) \\left(T _ {b} + k \\left(1 - T _ {b}\\right) \\hat {L} _ {n} ^ {i} / (h w r)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 500, + 890, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h$ , $w$ are the height and width of input images, and $r$ is set to $0.15\\%$ [7, 24]. As shown in Fig. 4 (a), $T_{b}$ is the minimum threshold, and $k$ controls the threshold growth rate. An increasing number of $\\hat{L}_n^i$ leads to the increase of the threshold, which can reduce error accumulation of low contrast targets and strong background clutter.", + "bbox": [ + 496, + 526, + 892, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To eliminate false alarms by local neighborhood noise, we exclude the eight connective regions of candidate pixels that have no intersection with positive pixels of labels, as shown in Fig. 4 (b). This process is defined as:", + "bbox": [ + 496, + 617, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {n} ^ {i} = C _ {n} ^ {i} \\odot F _ {n} ^ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 685, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{n}^{i}$ is the candidate pixels after false alarm elimination, and $F_{n}^{i}$ is the mask against false alarm pixels.", + "bbox": [ + 496, + 710, + 890, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then perform average weighted summation between candidate pixels $E_{n}^{i}$ and current label $L_{n}^{i}$ to achieve label update. The process can be formulated as:", + "bbox": [ + 496, + 742, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {n + 1} ^ {i} = L _ {n} ^ {i} \\odot (1 - N _ {n} ^ {i}) + \\frac {L _ {n} ^ {i} + E _ {n} ^ {i}}{2} \\odot N _ {n} ^ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 795, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{n + 1}^{i}$ is the updated label in the $n^{th}$ round, which serves as new supervision for training in the $n + 1^{th}$ round,", + "bbox": [ + 496, + 834, + 892, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3The value of a pixel is higher than 0.5, which represents that the pixel is more likely to be positive than negative [8, 24]", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15531", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d9c35bfb351b65feaf705240ff75f4627cc300197d6c3e8cbce41fd3e4d21f12.jpg", + "image_caption": [ + "(a) Target Intensity", + "Figure 5. IoU and visualize results of mapping degeneration with respect to different characteristics of targets (i.e.,(a) intensity, (b) size,) and point labels (i.e.,(c) locations and (d) numbers). We visualize the zoom-in target regions of input images with GT point labels (i.e., red dots in images) and corresponding CNN predictions (in the epoch reaching maximum IoU)." + ], + "image_footnote": [], + "bbox": [ + 81, + 79, + 285, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cdbedc4ef76e60160481931442ab8be8f0ece18250064ce3f1a3faa818929a83.jpg", + "image_caption": [ + "(b) Target Size" + ], + "image_footnote": [], + "bbox": [ + 285, + 80, + 486, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31311459c1a3ebab9c120d00bb84e4a812848b1ae9b14b36dbb6e963fa4fb605.jpg", + "image_caption": [ + "(c) Locations of Point label" + ], + "image_footnote": [], + "bbox": [ + 488, + 80, + 687, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4c711bfccfd9b463477f5a3b3a8aa0a832f03359cac6066dd1426bc44860339.jpg", + "image_caption": [ + "(d) Numbers of Point Label" + ], + "image_footnote": [], + "bbox": [ + 689, + 80, + 893, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $N_{n}^{i} = \\left(P_{n}^{i} > T_{\\text{adapt}}\\right) \\odot F_{n}^{i}$ . Note that, the first term represents GT labels below red dotted lines, and the second term represents the average weighted summation between predictions and GT labels above red dotted lines.", + "bbox": [ + 75, + 286, + 467, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is worth noting that we provide three conditions to ensure network convergence: 1) Average weighted summation between predictions and labels promotes CNNs to converge as predictions approximate labels. 2) Pixel-adaptive threshold increases with the increase of positive pixels in updated labels, which slows down or suspends the label update. 3) As label evolution introduces more target information for training, CNNs grow to mature, and learn to distinguish targets from backgrounds.", + "bbox": [ + 75, + 349, + 467, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start label evolution after initial evolution epoch $T_{epoch}$ , and perform label update every $f$ epoch until the end of training. Note that, our epoch-based threshold $T_{epoch}$ is a coarse threshold to ensure that networks attend to targets instead of background clutter.", + "bbox": [ + 76, + 486, + 467, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 579, + 207, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe the implementation details, and then make comprehensive analyses of the mapping degeneration phenomenon and our label evolution framework. In addition, we apply our method to the state-of-the-art SIRST detection methods with point supervision, and make comparisons with their fully supervised counterparts. Moreover, we make comparisons of our dynamic updated pseudo labels with fixed pseudo labels, and discuss the calculation of loss function.", + "bbox": [ + 75, + 604, + 467, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 755, + 292, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Three public datasets NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] are used in our experiments. We followed [24] to split the training and test sets of NUAA-SIRST and NUDT-SIRST, and followed [54] to split IRSTD-1K. We employed two pixel-level metrics (i.e., intersection over union $(IoU)$ and pixel accuracy $(PA))$ and two target-level metrics (i.e., probability of detection $(P_d)$ and false-alarm rate $(F_a))$ for performance evaluation.", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, all images were normalized and randomly cropped into patches of size $256 \\times 256$ as network inputs. We augmented the training data by random flipping and rotation. Due to the extreme positive-negative sample imbalance (less than 10 vs. more than $256 \\times 256$ ) in SIRST detection with point supervision, we employed focal loss[27] to stabilize the training process. All the networks were optimized by the Adam method [22]. Batch size was set to 16, and learning rate was initially set to $5 \\times 10^{-4}$ and reduced by ten times at the $200^{th}$ and $300^{th}$ epochs. We stopped training after 400 epochs. All models were implemented in PyTorch [35] on a PC with an Nvidia GeForce 3090 GPU.", + "bbox": [ + 496, + 286, + 890, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Model Analyses", + "text_level": 1, + "bbox": [ + 500, + 492, + 656, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2.1 Analyses of Mapping Degeneration", + "text_level": 1, + "bbox": [ + 500, + 516, + 797, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use synthetic images (simulated targets and real backgrounds [24]) to investigate the mapping degeneration phenomenon with respect to different characteristics of targets (i.e., intensity and size $^5$ ) and point labels (i.e., locations and numbers). We employ U-Net [41] as the baseline detection network, and use centroid point as GT label if not specified. We calculate IoU between positive pixels in predictions and GT mask labels of each epoch for quantitative analyses. In addition, we visualize the zoom-in target regions of simulated images with GT point labels (i.e., red dots) and corresponding CNN predictions (in the epoch reaching maximum IoU). To reduce training randomness, we show the average IoU results and visualization results over 100 runs.", + "bbox": [ + 496, + 539, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target Intensity. We simulate Gaussian-based extended targets with different peak values (i.e., 10, 50, 100, 200, 500) to investigate the influence of target intensity on mapping degeneration. Quantitative results in Fig. 5(a) show that intensity higher than 100 leads to a positive correlation between intensity and maximum IoU, while", + "bbox": [ + 496, + 750, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4Focal loss is calculated between current evolved and GT labels to supervise the network training until the next round label update.", + "bbox": [ + 500, + 849, + 890, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Shape, and local background clutter are investigated in supplemental material.", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15532", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5eb51815d988cb84a58c1be6c49ca872022a045e720610234f44fe32c55072ae.jpg", + "image_caption": [ + "(a) Initial Evolution Epoch $T_{\\text{epoch}}$" + ], + "image_footnote": [], + "bbox": [ + 80, + 78, + 282, + 174 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31839d592377a44cbb7a23c802f75c69f7a776e70bfaed689d57f2f1b9604a14.jpg", + "image_caption": [ + "(b) $T_{b}$ of Evolution Threshold" + ], + "image_footnote": [], + "bbox": [ + 282, + 78, + 496, + 174 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34a1a6f1452c80a2b434298440d647e79534e537f3a95da49d157f26c31d7a49.jpg", + "image_caption": [ + "(c) $k$ of Evolution Threshold" + ], + "image_footnote": [], + "bbox": [ + 500, + 78, + 684, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34a9f9aefc047fbfce0285991b847d0b04da3886369f943d3e44408f3a5c121b.jpg", + "image_caption": [ + "(d) Evolution Frequency $f$" + ], + "image_footnote": [], + "bbox": [ + 686, + 79, + 890, + 176 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/806778cd0a16a4e3340a307a8bf9476239afd921fb2062c5818d5972b412bdb8.jpg", + "table_caption": [ + "Table 1. Average $IoU$ ( $\\times 10^{2}$ ), $P_{d}$ ( $\\times 10^{2}$ ) and $F_{a}$ ( $\\times 10^{6}$ ) values on NUAA-SIRST [8] NUDT-SIRST [24] and IRSTD-1K [54] achieved by DNA-Net with (w/) and without (w/o) LESPS under centroid, coarse point supervision together with full supervision." + ], + "table_footnote": [], + "table_body": "
MethodCentroidCoarseFull
IoUPdFaIoUPdFaIoUPdFa
w/o LESPS5.1289.190.682.9649.890.3075.6796.1822.94
w/ LESPS57.3491.8720.2456.1891.4918.32
", + "bbox": [ + 80, + 304, + 462, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lower intensity leads to a negative one. In addition, curve \"intensity 10\" reaches maximum $IoU$ at around epoch 150 while others are less than 50, which demonstrates that over-small intensity decelerates degeneration. Visualization results show that our method can well highlight target regions under various intensities.", + "bbox": [ + 75, + 366, + 467, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Target Size. We simulate Gaussian-based extended targets with different radii (i.e., 3, 5, 7, 9, 13) to investigate the influence of target size on mapping degeneration. Quantitative results in Fig. 5(b) show that larger target size leads to lower maximum IoU and less time to reach maximum IoU. That is because, size discrepancy between targets and GT point labels increases as target size increases, which aggravates and accelerates mapping degeneration. Visualization results show that CNNs can predict a cluster of pixels in a size-aware manner, and the peak values of predictions decrease as target size increases.", + "bbox": [ + 75, + 458, + 467, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Locations of Point Label. We simulate a Gaussian-based extended target (with intensity 500 & radius 13), and place point labels at different distances away from the target centroid to investigate the influence of point label locations on mapping degeneration. Results in Fig. 5(c) show that small perturbations of label locations (less than 3 pixels) have a minor influence on the maximum IoU results. However, severe location perturbations (larger than 3 pixels) can lead to a significant maximum IoU drop, and the drop is more obvious when the point label is close to the edge. Note that, the same targets with different label locations reach maximum IoU at the same time, which demonstrates that the speed of mapping degeneration is irrelevant to the position of labels.", + "bbox": [ + 75, + 627, + 467, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Number of Points in Label. We simulate a Gaussian-based extended target (with intensity 500 & radius 13) and set different numbers of points in labels to investigate its influence on mapping degeneration. Quantitative results", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/111d120cbbbbf643f11514976d8c732ea735d336081aa437ccbcc119ea041dc9.jpg", + "image_caption": [ + "Figure 6. $PA$ (P) and $IoU$ (I) results of LESPS with respect to (a) initial evolution epoch $T_{epoch}$ , (b) $T_b$ and (c) $k$ of evolution threshold, and (d) evolution frequency $f$ .", + "Figure 7. Quantitative and qualitative results of evolved target masks." + ], + "image_footnote": [], + "bbox": [ + 501, + 233, + 898, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Fig. 5(d) show that as the number of points increases, CNNs can learn more target information to achieve higher maximum IoU results. In addition, the speed of mapping degeneration is irrelevant to the point number. Visualization results show that peak values of predictions increase as the number of points increases, which demonstrates that stronger supervision alleviates mapping degeneration. The conclusion well supports our label evolution framework.", + "bbox": [ + 496, + 368, + 890, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2 Analyses of the Label Evolution Framework", + "text_level": 1, + "bbox": [ + 498, + 508, + 859, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we conduct experiments to investigate the effectiveness and the optimal parameter settings of our label evolution framework (i.e., LESPS). We employ $PA$ and $IoU$ between the positive pixels in updated labels and the GT mask labels to quantitatively evaluate the accuracy and expansion degree of the current label.", + "bbox": [ + 496, + 532, + 890, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness. We compare the average results of NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] datasets achieved by DNA-Net with (i.e., w/) and without (i.e., w/o) LESPS under centroid and coarse point supervision, respectively. Note that, the results of DNA-Net w/o LESPS are calculated under extremely low threshold $^6$ (i.e., 0.15) while those of DNA-Net w/ LESPS are calculated under the standard threshold (i.e., 0.5 [24, 54]). As shown in Table 1, as compared to full supervision, the results of DNA-Net w/o LESPS are extremely low, which demonstrates that SIRST detection on single point supervision is a challenging task. In contrast, DNA-Net w/ LESPS can achieve significant performance improvement under both point supervisions in terms of all the metrics, which approximate the performance of their fully supervised counterparts. Note that, $P_d$ results of DNA", + "bbox": [ + 496, + 623, + 892, + 864 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "6With point supervision, the results of DNA-Net w/o LESPS calculated under threshold 0.5 are all zeros.", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15533", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b03b94a70ab9a87594864d09604fc6d33d9914caaff4c659037cb559e3d013ea.jpg", + "image_caption": [ + "Figure 8. Visualizations of regressed labels during training and network predictions during inference with centroid and coarse point supervision." + ], + "image_footnote": [], + "bbox": [ + 80, + 79, + 475, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Net w/o LESPS under coarse point supervision are over half lower than those under the centroid ones, while the results of DNA-Net w/ LESPS under these two kinds of point supervision are comparable. It demonstrates that LESPS can generalize well to manual annotation errors.", + "bbox": [ + 75, + 275, + 468, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we make evaluations of evolved target masks during training. Quantitative results in Fig. 7(a) show average IoU values between positive pixels of evolved target masks and GT labels in 20-time training, which demonstrates that the quality of pseudo target masks consecutively increases during training. Qualitative results in Fig. 7(b) demonstrate that networks can effectively learn to expand point labels to mask labels. Furthermore, we visualize the labels regressed by our LESPS during training together with some network predictions during inference in Figs. 8 (a) and (b). As shown in Fig. 8(a), compared with GT mask labels, the evolved labels are more closely aligned with the objects in images (e.g., GT masks of Misc_4, XDU113 exceed the target regions due to visual edge ambiguity), which demonstrates that LESPS can alleviate the manually annotated errors. Figure 8(b) shows that DNA-Net w/ LESPS can effectively achieve accurate pixel-level SIRST detection in an end-to-end manner. Please refer to the supplemental materials for more visual results.", + "bbox": [ + 75, + 352, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initial Evolution Epoch. We investigate the optimal value of epoch-based threshold $T_{\\text{epoch}}$ . Figure 6(a) shows that small initial evolution epoch results in a significant difference between $PA$ and $IoU$ (i.e., 0.94 vs. 0.04 with $T_{\\text{epoch}} = 10$ ). That is because, early label evolution introduces many error pixels, resulting in severe error accumulation and network convergence failure. Increasing initial evolution epoch can reduce error accumulation and promote network convergence (0.60 vs. 0.54 with $T_{\\text{epoch}} = 50$ ). However, over-large initial evolution epoch (i.e., a high degree of mapping degeneration) results in inferior performance (0.21 vs. 0.21 with $T_{\\text{epoch}} = 100$ ). Therefore, $T_{\\text{epoch}}$ is set to 50 in our method.", + "bbox": [ + 75, + 641, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evolution Threshold. We investigate the optimal values of $k$ and $T_{b}$ in the evolution threshold. $T_{b}$ is the minimum threshold, and controls evolution speed and error accumulation degree. $k$ determines the maximum", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bbb71eaa0d7dc01f32000feab13282c294b7b925787e3fc799aa8ac20107b0c.jpg", + "image_caption": [ + "Figure 9. Visual detection results of different methods achieved on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets. Correctly detected targets and false alarms are highlighted by red and orange circles, respectively." + ], + "image_footnote": [], + "bbox": [ + 501, + 77, + 895, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "threshold, and controls the growth rate of the threshold. As shown in Fig. 6(b) and 6(c), both over-large and over-small values of $T_{b}$ and $k$ result in inferior performance. Therefore, we choose $k = 1 / 2$ and $T_{b} = 0.5$ in our method.", + "bbox": [ + 496, + 347, + 892, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evolution Frequency. We investigate the optimal value of evolution frequency $f$ . Figure 6(d) shows that evolution frequency is positively correlated to $PA$ and $IoU$ . However, high evolution frequency needs more time for label updates. To balance performance and efficiency, we choose $f = 5$ in our method. Interestingly, higher frequency (i.e., $f = 2$ ) does not cause serve error accumulation, which also demonstrates the effectiveness of the convergence conditions of our LESPS. Please refer to the supplemental materials for more discussions of the convergence issue.", + "bbox": [ + 496, + 409, + 892, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison to State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 500, + 573, + 844, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to SISRT detection methods. We apply our LESPS to several state-of-the-art CNN-based methods, including ACM [7], ALCNet [8] and DNA-Net [24]. For fair comparisons, we retrained all models on NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] datasets with the same settings. In addition, we add the results of six fully supervised CNN-based methods (MDvsFA [45], ACM [7], ALCNet [8], DNA-Net [24], ISNet [54], UIU-Net [46]) and six traditional methods (Top-Hat [40], RLCM [15], TLLCM [16], MSPCM [17], IPI [11], PSTNN [51]) as the baseline results.", + "bbox": [ + 496, + 597, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results in Table 2 show that CNN-based methods equipped with LESPS can outperform all the traditional methods. In addition, they can also achieve $71 - 75\\%$ IoU results and comparable $P_{d}$ and $F_{a}$ results of their fully supervised counterparts. Qualitative results in Fig. 9 show that CNN-based methods equipped with LESPS can produce outputs with precise target mask and low false alarm rate, and can generalize well to complex scenes. Please refer to supplemental materials for more quantitative", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15534", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/576d89fc78b2634a6a49bc4c33d88f2dd6cc3fbd634f375e7ccfc3f56ae5889e.jpg", + "table_caption": [ + "Table 2. IoU $(\\times 10^{2})$ , $P_{d}$ $(\\times 10^{2})$ and $F_{a}(\\times 10^{6})$ values of different methods achieved on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets. \"CNN Full\", \"CNN Centroid\", and \"CNN Coarse\" represent CNN-based methods under full supervision, centroid and coarse point supervision. \"+\" represents CNN-based methods equipped with LESPS." + ], + "table_footnote": [], + "table_body": "
MethodsDescriptionNUAA-SIRST [8]NUDT-SIRST [24]IRSTD-1K [54]Average
IoUPdFaIoUPdFaIoUPdFaIoUPdFa
Top-Hat [40]Filtering7.1479.841012.0020.7278.41166.7010.0675.111432.0012.6477.79870.23
RLCM [15]Local Contrast21.0280.61199.1515.1466.35163.0014.6265.6617.9516.0668.7098.77
TLLCM [16]Local Contrast11.0379.477.277.0662.0146.125.3663.974.937.2265.4521.42
MSPCM [34]Local Contrast12.3883.2717.775.8655.87115.967.3360.2715.247.2361.5355.13
IPI [11]Low Rank25.6785.5511.4717.7674.4941.2327.9281.3716.1823.7880.4722.96
PSTNN [51]Low Rank22.4077.9529.1114.8566.1344.1724.5771.9935.2620.6172.0236.18
MDvsFA [45]CNN Full61.7792.4064.9045.3886.03200.7135.4085.8699.2247.5288.10121.61
ISNet [54]CNN Full72.0494.6842.4671.2796.9396.8460.6194.2861.2867.9795.3066.86
UIU-Net [46]CNN Full69.9095.8251.2075.9196.8318.6161.1192.9326.8768.9795.1932.23
ACM [7]CNN Full64.9290.8712.7657.4291.7539.7357.4991.5843.8659.9491.4032.12
CNN Centroid+49.2389.3540.9542.0991.1138.2441.4488.8960.4644.2589.7846.55
CNN Coarse+47.8188.2140.7540.6481.1149.4540.3792.5964.8142.9487.3051.67
ALCNet [8]CNN Full67.9192.7837.0461.7891.3236.3662.0390.9142.4663.9191.6738.62
CNN Centroid+50.6292.0236.8441.5892.2867.0144.9090.5784.6845.7091.6262.84
CNN Coarse+51.0090.8742.4044.1492.8032.1046.7592.2664.3047.3091.9846.27
DNA-Net [24]CNN Full76.8696.9622.587.4298.3124.562.7393.2721.8175.6796.1822.94
CNN Centroid+61.9592.0218.1757.9994.7126.4552.0988.8816.0957.3491.8720.24
CNN Coarse+61.1393.1611.8758.3793.7628.0149.0587.5415.0756.1891.4918.32
", + "bbox": [ + 81, + 127, + 883, + 335 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fbd3cda70b190dd016adbaaaefa3088bc01cf43230e6b94936f9036db0374e10.jpg", + "table_caption": [ + "Table 3. Average IoU $(\\times 10^{2})$ , $P_{d}$ $(\\times 10^{2})$ , $F_{a}(\\times 10^{6})$ values on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets of DNA-Net trained with pseudo labels generated by input intensity threshold, LCM-based methods [15, 16, 34] and LESPS under centroid and coarse point supervision." + ], + "table_footnote": [], + "table_body": "
Pseudo LabelCentroidCoarse
IoUPdFaIoUPdFa
Threshold=0.34.9281.7813.185.6783.1211.98
Threshold=0.513.2473.085.3115.5476.034.89
Threshold=0.714.5145.504.2815.2146.883.84
RLCM [15]21.4389.102.6722.5390.563.69
TLLCM [16]21.9590.967.7226.0594.154.27
MSPCM [34]28.8992.623.8429.7993.952.28
LESPS(ours)57.3491.8720.2456.1891.4918.32
", + "bbox": [ + 78, + 425, + 468, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and qualitative results.", + "bbox": [ + 76, + 537, + 227, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to other pseudo labels. We compare our dynamic updated pseudo labels with fixed pseudo labels generated by input intensity threshold and local contrast-based methods [15, 16, 34]. Specifically, given a GT point label, we only preserve the eight connected regions of detection results that have union pixels with the GT point label. Then, we employ the pseudo labels to retrain DNA-Net [24] from scratch. As shown in Table 3, compared with fixed pseudo labels, dynamic updated pseudo labels by LESPS can achieve the highest IoU values with comparable $P_{d}$ and reasonable $F_{a}$ increase.", + "bbox": [ + 75, + 554, + 468, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Discussion of Loss Function", + "text_level": 1, + "bbox": [ + 76, + 731, + 326, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this subsection, we investigate the loss function of computing negative loss on different background points. Average results of different baseline methods under centroid point supervision are shown in Table 4. Extremely limited handcrafted background points7 leads to many false alarms. Random sample8 introduces more background points and well alleviates class imbalance, resulting in great", + "bbox": [ + 75, + 756, + 468, + 863 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f635dda01aed254751646c0503647a329582044aa1d533e03792c9523e0aeaa1.jpg", + "table_caption": [ + "Table 4. Average $IoU(\\times 10^{2})$ , $P_{d}(\\times 10^{2})$ , $F_{a}(\\times 10^{3})$ values on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets of different methods when computing negative loss on $i$ handcrafted $(\\mathrm{hand}_i)$ , $j$ randomly sampled $(\\mathrm{rand}_j)$ and all background points." + ], + "table_footnote": [], + "table_body": "
AnnotationsACMALCNetDNA-Net
IoUPdFaIoUPdFaIoUPdFa
hand10.5495.7947.060.1695.19262.061.4397.8026.91
hand20.1297.24295.170.1596.62248.053.4198.188.48
hand50.1196.36316.370.0897.29363.253.6898.137.29
rand18.0693.453.568.5792.973.2118.7494.690.58
rand210.7892.722.2210.7891.162.7122.8594.810.42
rand513.3992.661.3511.8793.260.8924.8095.000.34
All (Ours)3.9587.150.024.0888.930.025.1289.180.01
", + "bbox": [ + 501, + 425, + 890, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance improvements. However, the above simple versions introduce huge false alarms (34-1.8K times of all points), which are not practical for real applications, but inspire further thorough investigation in the future.", + "bbox": [ + 498, + 536, + 890, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 616, + 617, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed the first work to achieve weakly-supervised SIRST detection using single-point supervision. In our method, we discovered the mapping degeneration phenomenon and proposed a label evolution framework named label evolution with single point supervision (LESPS) to automatically achieve point-to-mask regression. Through LESPS, networks can be trained to achieve SIRST detection in an end-to-end manner. Extensive experiments and insightful visualizations have fully demonstrated the effectiveness and superiority of our method. In addition, our method can be applied to different networks to achieve over $70\\%$ and $95\\%$ of its fully supervised performance on pixel-level IoU and object-level $P_{d}$ , respectively. We hope our interesting findings and promising results can inspire researchers to rethink the feasibility of achieving state-of-the-art performance in SIRST detection with much weaker supervision.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "7Points are sampled near targets, and are fixed during training.", + "bbox": [ + 93, + 875, + 424, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "8Points are randomly sampled, and change in each training iteration.", + "bbox": [ + 93, + 887, + 455, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15535", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peri Akiva, Kristin Dana, Peter Oudemans, and Michael Mars. Finding berries: Segmentation and counting of cranberries using point supervision and shape priors. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 50-51, 2020.", + "[2] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In European Conference on Computer Vision (ECCV), pages 549-565. Springer, 2016.", + "[3] CL Philip Chen, Hong Li, Yantao Wei, Tian Xia, and Yuan Yan Tang. A local contrast method for small infrared target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 52(1):574-581, 2013.", + "[4] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8823-8832, 2021.", + "[5] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2617-2626, 2022.", + "[6] Yimian Dai and Yiquan Wu. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTAR), 10(8):3752-3767, 2017.", + "[7] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2021.", + "[8] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), pages 1-12, 2021.", + "[9] Suyog D Deshpande, Meng Hwa Er, Ronda Venkateswarlu, and Philip Chan. Max-mean and max-median filters for detection of small targets. In Signal and Data Processing of Small Targets, volume 3809, pages 74-83. SPIE, 1999.", + "[10] Junsong Fan, Zhaoxiang Zhang, and Tieniu Tan. Pointly-supervised panoptic segmentation. ArXiv Preprint ArXiv:2210.13950, 2022.", + "[11] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE Transactions on Image Processing (TIP), 22(12):4996-5009, 2013.", + "[12] Yongtao Ge, Qiang Zhou, Xinlong Wang, Chunhua Shen, Zhibin Wang, and Hao Li. Point-teaching: Weakly semi-supervised object detection with point annotations. ArXiv Preprint ArXiv:2206.00274, 2022.", + "[13] Ross Girshick. Fast r-cnn. In IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015.", + "[14] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In IEEE Conference on" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition (CVPR), pages 580-587, 2014.", + "[15] Jinhui Han, Kun Liang, Bo Zhou, Xinying Zhu, Jie Zhao, and Linlin Zhao. Infrared small target detection utilizing the multiscale relative local contrast measure. IEEE Geoscience and Remote Sensing Letters (GRSL), 15(4):612-616, 2018.", + "[16] Jinhui Han, Saed Moradi, Iman Faramarzi, Chengyin Liu, Honghui Zhang, and Qian Zhao. A local contrast method for infrared small-target detection utilizing a tri-layer window. IEEE Geoscience and Remote Sensing Letters (GRSL), 17(10):1822-1826, 2019.", + "[17] Jinhui Han, Saed Moradi, Iman Faramarzi, Honghui Zhang, Qian Zhao, Xiaojian Zhang, and Nan Li. Infrared small target detection based on the weighted strengthened local contrast measure. IEEE Geoscience and Remote Sensing Letters (GRSL), 18(9):1670-1674, 2020.", + "[18] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems (NeurIPS), 34:15908-15919, 2021.", + "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In IEEE International Conference on Computer Vision (ICCV), pages 2961-2969, 2017.", + "[20] Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Haoliang Zheng, and Wei Zhang. Ristdnet: Robust infrared small target detection network. IEEE Geoscience and Remote Sensing Letters (GRSL), 19:1-5, 2021.", + "[21] Moran Ju, Jiangning Luo, Guangqi Liu, and Haibo Luo. Istdet: An efficient end-to-end neural network for infrared small target detection. Infrared Physics & Technology, 114:103659, 2021.", + "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015.", + "[23] Issam H Laradji, Negar Rostamzadeh, Pedro O Pinheiro, David Vazquez, and Mark Schmidt. Where are the blobs: Counting by localization with point supervision. In European Conference on Computer Vision (ECCV), pages 547-562, 2018.", + "[24] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing (TIP), 2022.", + "[25] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, et al. Yolov6: A single-stage object detection framework for industrial applications. ArXiv Preprint ArXiv:2209.02976, 2022.", + "[26] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022.", + "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017." + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15536", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Ting Liu, Jungang Yang, Boyang Li, Chao Xiao, Yang Sun, Yingqian Wang, and Wei An. Non-convex tensor low-rank approximation for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 2021.", + "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision (ECCV), pages 21-37. Springer, 2016.", + "[30] Yongtuo Liu, Dan Xu, Sucheng Ren, Hanjie Wu, Hongmin Cai, and Shengfeng He. Fine-grained domain adaptive crowd counting via point-derived segmentation. ArXiv Preprint ArXiv:2108.02980, 2021.", + "[31] Kevis-Kokitsi Maninis, Sergi Caelles, Jordi Pont-Tuset, and Luc Van Gool. Deep extreme cut: From extreme points to object segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 616-625, 2018.", + "[32] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021.", + "[33] Saed Moradi, Payman Moallem, and Mohamad Farzan Sabahi. A false-alarm aware methodology to develop robust and efficient multi-scale infrared small target detection algorithm. Infrared Physics & Technology, 89:387-397, 2018.", + "[34] Saed Moradi, Payman Moallem, and Mohamad Farzan Sabahi. A false-alarm aware methodology to develop robust and efficient multi-scale infrared small target detection algorithm. Infrared Physics & Technology, 89:387-397, 2018.", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NeurlPS), 32, 2019.", + "[36] Rui Qian, Yunchao Wei, Honghui Shi, Jiachen Li, Jiaying Liu, and Thomas Huang. Weakly supervised scene parsing with point-based distance metric learning. In AAAI Conference on Artificial Intelligence (AAAI), volume 33, pages 8843-8850, 2019.", + "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016.", + "[38] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7263-7271, 2017.", + "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Jeanfrancois Rivest and Roger Fortin. Detection of dim targets in digital infrared imagery by morphological image processing. Optical Engineering, 35(7):1886-1893, 1996.", + "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015.", + "[42] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017.", + "[43] Yang Sun, Jungang Yang, and Wei An. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor model. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 59(5):3737-3752, 2020.", + "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018.", + "[45] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In IEEE International Conference on Computer Vision (ICCV), pages 8509-8518, 2019.", + "[46] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: u-net in u-net for infrared small object detection. IEEE Transactions on Image Processing (TIP), 32:364-376, 2022.", + "[47] Xinyi Ying, Yingqian Wang, Longguang Wang, Weidong Sheng, Li Liu, Zaiping Lin, and Shilin Zhou. Local motion and contrast priors driven deep network for infrared small target super-resolution. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTAR), 15:5480-5495, 2022.", + "[48] Mohsen Zand, Haleh Damirchi, Andrew Farley, Mahdiyar Molahasani, Michael Greenspan, and Ali Etemad. Multiscale crowd counting and localization by multitask point supervision. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1820-1824. IEEE, 2022.", + "[49] Ke Zhang, Shuyan Ni, Dashuang Yan, and Aidi Zhang. Review of dim small target detection algorithms in single-frame infrared images. In IEEE Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC), volume 4, pages 2115-2120. IEEE, 2021.", + "[50] Landan Zhang, Lingbing Peng, Tianfang Zhang, Siying Cao, and Zhenming Peng. Infrared small target detection via nonconvex rank approximation minimization joint 12, 1 norm. Remote Sensing, 10(11):1821, 2018.", + "[51] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019.", + "[52] Mingjin Zhang, Haichen Bai, Jing Zhang, Rui Zhang, Chaoyue Wang, Jie Guo, and Xinbo Gao. Rkformer: Runge-kutta transformer with random-connection attention for" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15537", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "infrared small target detection. In 30th ACM International Conference on Multimedia (ACM MM), pages 1730-1738, 2022.", + "[53] Mingjin Zhang, Ke Yue, Jing Zhang, Yunsong Li, and Xinbo Gao. Exploring feature compensation and cross-level correlation for infrared small target detection. In 30th ACM International Conference on Multimedia (ACM MM), pages 1857-1865, 2022.", + "[54] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 877-886, 2022.", + "[55] Shiyin Zhang, Jun Hao Liew, Yunchao Wei, Shikui Wei, and Yao Zhao. Interactive object segmentation with inside-outside guidance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12234-12244, 2020.", + "[56] Shilong Zhang, Zhuoran Yu, Liyang Liu, Xinjiang Wang, Aojun Zhou, and Kai Chen. Group r-cnn for weakly semi-supervised object detection with points. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9417-9426, 2022.", + "[57] Mingjing Zhao, Wei Li, Lu Li, Jin Hu, Pengge Ma, and Ran Tao. Single-frame infrared small-target detection: A survey. IEEE Geoscience and Remote Sensing Magazine (GRSM), 2022.", + "[58] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2016.", + "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2020." + ], + "bbox": [ + 78, + 90, + 470, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15538", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_model.json b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f4733ba4cd5c257f2375274c2a3a40d93d6c7653 --- /dev/null +++ b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.13, + 0.888, + 0.177 + ], + "angle": 0, + "content": "Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.203, + 0.804, + 0.24 + ], + "angle": 0, + "content": "Xinyi Ying\\(^{1}\\), Li Liu\\(^{1}\\), Yingqian Wang\\(^{1}\\), Ruojing Li\\(^{1}\\), Nuo Chen\\(^{1}\\), Zaiping Lin\\(^{1\\boxtimes}\\), Weidong Sheng\\(^{1}\\), Shilin Zhou\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.309, + 0.239, + 0.662, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)National University of Defense Technology" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.257, + 0.832, + 0.276 + ], + "angle": 0, + "content": "{yingxinyi18, wangyingqian16, liruojing, chennuo97, linzaiping, slzhou}@nudt.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.275, + 0.719, + 0.293 + ], + "angle": 0, + "content": "dreamliu2010@gmail.com, shengweidong1111@sohu.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.474, + 0.766 + ], + "angle": 0, + "content": "Training a convolutional neural network (CNN) to detect infrared small targets in a fully supervised manner has gained remarkable research interests in recent years, but is highly labor expensive since a large number of per-pixel annotations are required. To handle this problem, in this paper, we make the first attempt to achieve infrared small target detection with point-level supervision. Interestingly, during the training phase supervised by point labels, we discover that CNNs first learn to segment a cluster of pixels near the targets, and then gradually converge to predict groundtruth point labels. Motivated by this \"mapping degeneration\" phenomenon, we propose a label evolution framework named label evolution with single point supervision (LESPS) to progressively expand the point label by leveraging the intermediate predictions of CNNs. In this way, the network predictions can finally approximate the updated pseudo labels, and a pixel-level target mask can be obtained to train CNNs in an end-to-end manner. We conduct extensive experiments with insightful visualizations to validate the effectiveness of our method. Experimental results show that CNNs equipped with LESPS can well recover the target masks from corresponding point labels, and can achieve over \\(70\\%\\) and \\(95\\%\\) of their fully supervised performance in terms of pixel-level intersection over union (IoU) and object-level probability of detection \\((P_d)\\), respectively. Code is available at https://github.com/XinyiYing/LESPS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.783, + 0.21, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.471, + 0.869 + ], + "angle": 0, + "content": "Infrared small target detection has been a longstanding, fundamental yet challenging task in infrared search and tracking systems, and has various important applications in civil and military fields [49,57], including traffic monitoring" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.874, + 0.447, + 0.886 + ], + "angle": 0, + "content": "This work was supported by National Key Research and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.887, + 0.376, + 0.899 + ], + "angle": 0, + "content": "Development Program of China No. 2021YFB3100800." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.326, + 0.876, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.487, + 0.895, + 0.545 + ], + "angle": 0, + "content": "Figure 1. An illustration of mapping degeneration under point supervision. CNNs always tend to segment a cluster of pixels near the targets with low confidence at the early stage, and then gradually learn to predict GT point labels with high confidence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.78 + ], + "angle": 0, + "content": "[24, 54], maritime rescue [52, 53] and military surveillance [7, 47]. Due to the rapid response and robustness to fast-moving scenes, single-frame infrared small target (SIRST) detection methods have always attracted much more attention, and numerous methods have been proposed. Early methods, including filtering-based [9, 40], local contrast-based [3, 16] and low rank-based [11, 43] methods, require complex handcrafted features with carefully tuned hyper-parameters. Recently, compact deep learning has been introduced in solving the problem of SIRST detection [24, 45, 54]. However, there are only a few attempts, and its potential remains locked, unlike the extensive explorations of deep learning for natural images. This is mainly due to potential reasons, including lack of large-scale, accurately annotated datasets and high stake application scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Infrared small targets are usually of very small size, weak, shapeless and textureless, and are easily submerged in diverse complex background clutters. As a result, directly adopting existing popular generic object detectors like RCNN series [13, 14, 19, 39], YOLO series [25, 37, 38] and SSD [29] to SIRST detection cannot produce satisfactory performance. Realizing this, researchers have been focusing on developing deep networks tailored for" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15528" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "infrared small targets by adequately utilizing the domain knowledge. However, most existing deep methods for SIRST detection [8, 24, 54] are fully supervised, which usually requires a large dataset with accurate target mask annotations for training. Clearly, this is costly [5, 26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.365 + ], + "angle": 0, + "content": "Therefore, a natural question arises: Can we develop a new framework for SIRST detection with single point supervision? In fact, to substantially reduce the annotation cost for object detection tasks, weakly supervised object detection methods with point supervision [4,5,26,56] have been studied in the field of computer vision. Although these weakly supervised methods achieve promising results, they are not designed for the problem of SIRST detection, and the class-agnostic labels (i.e., only foreground and background) of infrared small targets hinder their applications [42, 58]. Therefore, in this work, we intend to conduct the first study of weakly supervised SIRST detection with single-point supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.471, + 0.606 + ], + "angle": 0, + "content": "A key motivation of this work comes from an interesting observation during the training of SIRST detection networks. That is, with single point labels serving as supervision, CNNs always tend to segment a cluster of pixels near the targets with low confidence at the early stage, and then gradually learn to predict groundtruth (GT) point labels with high confidence, as shown in Fig. 1. It reveals the fact that region-to-region mapping is the intermediate result of the final region-to-point mapping1. We attribute this \"mapping degeneration\" phenomenon to the special imaging mechanism of infrared system [24, 54], the local contrast prior of infrared small targets [3, 8], and the easy-to-hard learning property of CNNs [44], in which the first two factors result in extended mapping regions beyond the point labels, and the last factor contributes to the degeneration process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Based on the aforementioned discussion, in this work, we propose a novel framework for the problem of weakly supervised SIRST detection, dubbed label evolution with single point supervision (LESPS). Specifically, LESPS leverages the intermediate network predictions in the training phase to update the current labels, which serve as supervision until the next label update. Through iterative label update and network training, the network predictions can finally approximate the updated pseudo mask labels, and the network can be simultaneously trained to achieve pixel-level SIRST detection in an end-to-end\\(^2\\) manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.774, + 0.47, + 0.805 + ], + "angle": 0, + "content": "Our main contributions are summarized as: (1) We present the first study of weakly supervised SIRST" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "detection, and introduce LESPS that can significantly reduce the annotation cost. (2) We discover the mapping degeneration phenomenon, and leverage this phenomenon to automatically regress pixel-level pseudo labels from the given point labels via LESPS. (3) Experimental results show that our framework can be applied to different existing SIRST detection networks, and enable them to achieve over \\(70\\%\\) and \\(95\\%\\) of its fully supervised performance in terms of pixel-level intersection over union \\((IoU)\\) and object-level probability of detection \\((P_d)\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.257, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.895, + 0.78 + ], + "angle": 0, + "content": "SIRST Detection. In the past decades, various methods have been proposed, including early traditional paradigms (e.g., filtering-based methods [9, 40], local contrast-based methods [3, 15-17, 33, 34], low rank-based methods [6, 11, 28, 43, 50, 51]) and recent deep learning paradigms [7, 8, 20, 21, 24, 45, 52-54]. Compared to traditional methods, which require delicately designed models and carefully tuned hyper-parameters, convolutional neural networks (CNNs) can learn the non-linear mapping between input images and GT labels in a data-driven manner, and thus generalize better to real complex scenes. As the pioneering work, Wang et al. [45] first employed a generative adversarial network to achieve a better trade-off between miss detection and false alarm. Recently, more works focus on customized solutions of infrared small target. Specifically, Dai et al. [7] specialized an asymmetric contextual module, and further incorporated local contrast measure [8] to improve the target contrast. Li et al. [24] preserved target information by repetitive feature fusion. Zhang et al. [54] aggregated edge information to achieve shape-aware SIRST detection. Zhang et al. [52, 53] explored cross-level correlation and transformer-based method [18] to predict accurate target mask. Wu et al. [46] customized a UIU-Net framework for multi-level and multi-scale feature aggregation. In conclusion, existing works generally focus on compact architectural designs to pursue superior performance in a fully supervised manner. However, due to the lack of a large number of public datasets [7, 24, 45] with per-pixel annotations, the performance and generalization of CNNs are limited. In addition, per-pixel manual annotations are time-consuming and labor-intensive. Therefore, we focus on achieving good pixel-level SIRST detection with weaker supervision and cheaper annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Weakly Supervised Segmentation with Points. Recently, point-level annotation has raised more attention in dense prediction tasks such as object detection [4, 12, 56], crowd counting [1, 23, 30, 48] and image segmentation [2,5,10,26,31,36,55]. We mainly focus on image segmentation in this paper. Specifically, Bearman et al. [2] made the first attempt to introduce an objectiveness potential into a pointly supervised training" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.816, + 0.47, + 0.903 + ], + "angle": 0, + "content": "1 \"region-to-region mapping\" represents the mapping learned by CNNs from target regions in images to a cluster of pixels near the targets, while \"region-to-point mapping\" represents the mapping from target regions in images to the GT point labels. 2 Different from generic object detection [32, 59], \"end-to-end\" here represents achieving point-to-mask label regression and direct pixel-level inference in once training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15529" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.351 + ], + "angle": 0, + "content": "loss function to boost segmentation performance. Qian et al. [36] leveraged semantic information of several labeled points by a distance metric loss to achieve scene parsing. Zhang et al. [55] proposed an inside-outside guidance approach to achieve instance segmentation by five elaborate clicks. Cheng et al. [5] designed to provide ten randomly sampled binary point annotations within box annotations for instance segmentation. Li et al. [26] encoded each instance with kernel generator for panoptic segmentation to achieve \\(82\\%\\) of fully-supervised performance with only twenty randomly annotated points. In contrast to these approaches employing complicated prior constraints to segment large generic objects with rich color and fine textures by several elaborate points, we fully exploit the local contrast prior of infrared small target to progressively evolve pseudo masks by single coarse point without any auxiliaries in an end-to-end manner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.362, + 0.447, + 0.381 + ], + "angle": 0, + "content": "3. The Mapping Degeneration Phenomenon" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.471, + 0.478 + ], + "angle": 0, + "content": "In this section, we first describe the mapping degeneration phenomenon together with our intuitive explanation. Then we conduct experiments under single-sample and many-sample training schemes to demonstrate the generality of degeneration, and investigate the influence of generalization on degeneration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.473, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 1, given an input image and the corresponding GT point label, we employ U-Net [41] as the baseline SIRST detection network for training. It can be observed that, in the early training phase, network predicts a cluster of pixels near the targets with low confidence. As training continues, the network prediction finally approximates GT point label with gradually increased confidence. We name this phenomenon as \"mapping degeneration\", and attribute the following reasons to this phenomenon. 1) Special imaging mechanism of infrared systems [24, 54]: Targets only have intensity information without structure and texture details, resulting in highly similar pixels within the target region. 2) High local contrast of infrared small targets [3, 8]: Pixels within the target region are much brighter or darker with high contrast against the local background clutter. 3) Easy-to-hard learning property of CNNs [44]: CNNs always tend to learn simple mappings first, and then converge to difficult ones. Compared with region-to-point mapping, region-to-region mapping is easier, and thus tends to be the intermediate result of region-to-point mapping. In conclusion, the unique characteristics of infrared small targets result in extended mapping regions beyond point labels, and CNNs contribute to the mapping degeneration process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.472, + 0.903 + ], + "angle": 0, + "content": "It is worth noting that the mapping degeneration phenomenon is a general phenomenon in various scenes with infrared small targets. Specifically, we use the training datasets (including 1676 images and their corresponding" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.703, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.091, + 0.895, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.168, + 0.895, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.279, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Figure 2. Quantitative and qualitative illustrations of mapping degeneration in CNNs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.895, + 0.533 + ], + "angle": 0, + "content": "centroid point label, see details in Section 5.1) to train U-Net under a single-sample training scheme (i.e., training one CNN on each image). For quantitative analyses, we employ the IoU results between positive pixels in predictions (i.e., pixels with confidence higher than half of its maximum value) and GT mask label. Average IoU results of 1676 CNNs at each epoch are shown by the blue curve in Fig. 2(a), while the number of training samples with maximum IoU during training phase falling in a threshold range of \\([i, i + 0.1]\\), \\((i = 0, 0.1, \\dots, 0.9)\\) is illustrated via blue bars in Fig. 2(b). It can be observed from the zoom-in curve and bars that mapping degeneration is a general phenomenon with point supervision, and U-Net can achieve IoU \\(>0.5\\) on more than \\(60\\%\\) of the training images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.896, + 0.777 + ], + "angle": 0, + "content": "In addition, we conduct experiments to train U-Net under a many-sample training scheme (i.e., training one CNN using all images which contain abundant targets with various sizes and shapes) to investigate the effect of generalization on mapping degeneration. Average IoU results of 1676 images are shown by orange curve in Fig. 2(a). It can be observed that many-sample training scheme needs more time to converge. Moreover, Fig. 2(b) shows that orange bars are slightly lower than blue ones on larger IoU values (i.e., 0.5-1.0). It is demonstrated that generalization decelerates but aggravates mapping degeneration. Figure 2(c) shows some zoom-in target regions of images and their predictions under these two training schemes. It can be observed that CNNs can effectively segment a cluster of target pixels under both training schemes in a size-aware manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Therefore, an intuitive assumption arises: Can we leverage the intermediate results of CNNs to regress masks? A simple early stopping strategy seems to be a positive answer but is indeed unpractical since mapping degeneration is influenced by various factors, including target intensity, size, shape, and local background clutter (see details in Section 5.2.1). Consequently, there is no fixed optimal stopping epoch for all situations. These" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15530" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.076, + 0.477, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.247, + 0.471, + 0.304 + ], + "angle": 0, + "content": "Figure 3. An illustration of label evolution with single point supervision (LESPS). During training, intermediate predictions of CNNs are used to progressively expand point labels to mask labels. Black arrows represent each round of label updates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.471, + 0.362 + ], + "angle": 0, + "content": "observations motivate us to design a special label evolution framework to well leverage the mapping degeneration for pseudo mask regression." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.379, + 0.392 + ], + "angle": 0, + "content": "4. The Label Evolution Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Motivated by mapping degeneration, we propose a label evolution framework named label evolution with single point supervision (LESPS) to leverage the intermediate network predictions in the training phase to update labels. As training continues, the network predictions approximate the updated pseudo mask labels, and network can simultaneously learn to achieve pixel-level SIRST detection in an end-to-end manner. Here, we employ a toy example of 1D curves for easy understanding. As shown in Fig. 3, sub-figures on the left of the dotted line represent the network predictions. Note that, the black curves denote the intermediate predictions within LESPS, while the gray curves represent virtual results produced by the network without label update. On the right of the dotted line, the first and second columns of sub-figures represent current labels and updated labels, respectively, and black arrows represent each round of label update. The overall framework can be summarized as follows. With point label serving as supervision, in the \\(1^{st}\\) round label update after initial training, the predictions are used to update the current point label to generate the \\(1^{st}\\) updated label, which is then used to supervise the network training until the \\(2^{nd}\\) round label update. Through iterative label updates and network training, CNNs can incorporate the local contrast prior to gradually recover the mask labels. From another viewpoint, label evolution consistently updates the supervision to prevent mapping degeneration, and promotes CNNs to converge to the easy region-to-region mapping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Taking the \\( n^{th} \\) update as an example, given the current label \\( L_{n} \\) and the network prediction \\( P_{n} \\), we perform label update for each target, which consists of three steps: candidate pixel extraction, false alarm elimination, and weighted summation between candidate pixels and current" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.077, + 0.887, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.195, + 0.894, + 0.265 + ], + "angle": 0, + "content": "Figure 4. (a) Adaptive threshold \\( T_{\\text{adapt}} \\) with respect to positive pixels \\( \\hat{L}_n^i \\) and hyper-parameters \\( k, T_b \\). Pink dotted line represents the constant \\( hwr \\). (b) An illustration of false alarm elimination. Red circle and dot represent positive pixels and centroid point of label. Orange circle represents false alarms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.893, + 0.397 + ], + "angle": 0, + "content": "labels. Specifically, the \\(d \\times d\\) local neighborhoods of the \\(i^{th}\\) target in label \\(L_{n}\\) and prediction \\(P_{n}\\) are cropped based on the centroid of the positive pixels\\(^{3}\\) in label (i.e., \\(\\hat{L}_n\\)). Then to reduce error accumulation for label update (see Section 5.2.2 for details), we employ an adaptive threshold (the red dotted line in Fig. 3) to extract the local neighborhood candidate pixels (predictions higher than the red dotted line in Fig. 3). The process can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.405, + 0.891, + 0.423 + ], + "angle": 0, + "content": "\\[\nC _ {n} ^ {i} = P _ {n} ^ {i} \\odot \\left(P _ {n} ^ {i} > T _ {\\text {a d a p t}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.893, + 0.493 + ], + "angle": 0, + "content": "where \\( C_n^i \\) is the candidate pixels, and \\( \\odot \\) represents elementwise multiplication. \\( T_{adapt} \\) is the adaptive threshold that correlated to the current prediction \\( P_n^i \\) and the positive pixels in label \\( \\hat{L}_n^i \\), and can be calculated according to:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.501, + 0.891, + 0.518 + ], + "angle": 0, + "content": "\\[\nT _ {\\text {a d a p t}} = \\max \\left(P _ {n} ^ {i}\\right) \\left(T _ {b} + k \\left(1 - T _ {b}\\right) \\hat {L} _ {n} ^ {i} / (h w r)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.893, + 0.617 + ], + "angle": 0, + "content": "where \\( h \\), \\( w \\) are the height and width of input images, and \\( r \\) is set to \\( 0.15\\% \\) [7, 24]. As shown in Fig. 4 (a), \\( T_{b} \\) is the minimum threshold, and \\( k \\) controls the threshold growth rate. An increasing number of \\( \\hat{L}_n^i \\) leads to the increase of the threshold, which can reduce error accumulation of low contrast targets and strong background clutter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.678 + ], + "angle": 0, + "content": "To eliminate false alarms by local neighborhood noise, we exclude the eight connective regions of candidate pixels that have no intersection with positive pixels of labels, as shown in Fig. 4 (b). This process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.686, + 0.891, + 0.704 + ], + "angle": 0, + "content": "\\[\nE _ {n} ^ {i} = C _ {n} ^ {i} \\odot F _ {n} ^ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.891, + 0.743 + ], + "angle": 0, + "content": "where \\( E_{n}^{i} \\) is the candidate pixels after false alarm elimination, and \\( F_{n}^{i} \\) is the mask against false alarm pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.788 + ], + "angle": 0, + "content": "We then perform average weighted summation between candidate pixels \\( E_{n}^{i} \\) and current label \\( L_{n}^{i} \\) to achieve label update. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.796, + 0.891, + 0.827 + ], + "angle": 0, + "content": "\\[\nL _ {n + 1} ^ {i} = L _ {n} ^ {i} \\odot (1 - N _ {n} ^ {i}) + \\frac {L _ {n} ^ {i} + E _ {n} ^ {i}}{2} \\odot N _ {n} ^ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.835, + 0.893, + 0.868 + ], + "angle": 0, + "content": "where \\(L_{n + 1}^{i}\\) is the updated label in the \\(n^{th}\\) round, which serves as new supervision for training in the \\(n + 1^{th}\\) round," + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3The value of a pixel is higher than 0.5, which represents that the pixel is more likely to be positive than negative [8, 24]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15531" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.08, + 0.287, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.217, + 0.226, + 0.224 + ], + "angle": 0, + "content": "(a) Target Intensity" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.081, + 0.488, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.217, + 0.423, + 0.224 + ], + "angle": 0, + "content": "(b) Target Size" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.081, + 0.689, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.217, + 0.646, + 0.224 + ], + "angle": 0, + "content": "(c) Locations of Point label" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.082, + 0.894, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.217, + 0.852, + 0.224 + ], + "angle": 0, + "content": "(d) Numbers of Point Label" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.894, + 0.272 + ], + "angle": 0, + "content": "Figure 5. IoU and visualize results of mapping degeneration with respect to different characteristics of targets (i.e.,(a) intensity, (b) size,) and point labels (i.e.,(c) locations and (d) numbers). We visualize the zoom-in target regions of input images with GT point labels (i.e., red dots in images) and corresponding CNN predictions (in the epoch reaching maximum IoU)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.468, + 0.348 + ], + "angle": 0, + "content": "and \\( N_{n}^{i} = \\left(P_{n}^{i} > T_{\\text{adapt}}\\right) \\odot F_{n}^{i} \\). Note that, the first term represents GT labels below red dotted lines, and the second term represents the average weighted summation between predictions and GT labels above red dotted lines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.468, + 0.485 + ], + "angle": 0, + "content": "It is worth noting that we provide three conditions to ensure network convergence: 1) Average weighted summation between predictions and labels promotes CNNs to converge as predictions approximate labels. 2) Pixel-adaptive threshold increases with the increase of positive pixels in updated labels, which slows down or suspends the label update. 3) As label evolution introduces more target information for training, CNNs grow to mature, and learn to distinguish targets from backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.487, + 0.468, + 0.562 + ], + "angle": 0, + "content": "We start label evolution after initial evolution epoch \\( T_{epoch} \\), and perform label update every \\( f \\) epoch until the end of training. Note that, our epoch-based threshold \\( T_{epoch} \\) is a coarse threshold to ensure that networks attend to targets instead of background clutter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.58, + 0.208, + 0.596 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.468, + 0.742 + ], + "angle": 0, + "content": "In this section, we first describe the implementation details, and then make comprehensive analyses of the mapping degeneration phenomenon and our label evolution framework. In addition, we apply our method to the state-of-the-art SIRST detection methods with point supervision, and make comparisons with their fully supervised counterparts. Moreover, we make comparisons of our dynamic updated pseudo labels with fixed pseudo labels, and discuss the calculation of loss function." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.294, + 0.771 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Three public datasets NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] are used in our experiments. We followed [24] to split the training and test sets of NUAA-SIRST and NUDT-SIRST, and followed [54] to split IRSTD-1K. We employed two pixel-level metrics (i.e., intersection over union \\((IoU)\\) and pixel accuracy \\((PA))\\) and two target-level metrics (i.e., probability of detection \\((P_d)\\) and false-alarm rate \\((F_a))\\) for performance evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.483 + ], + "angle": 0, + "content": "During training, all images were normalized and randomly cropped into patches of size \\(256 \\times 256\\) as network inputs. We augmented the training data by random flipping and rotation. Due to the extreme positive-negative sample imbalance (less than 10 vs. more than \\(256 \\times 256\\)) in SIRST detection with point supervision, we employed focal loss[27] to stabilize the training process. All the networks were optimized by the Adam method [22]. Batch size was set to 16, and learning rate was initially set to \\(5 \\times 10^{-4}\\) and reduced by ten times at the \\(200^{th}\\) and \\(300^{th}\\) epochs. We stopped training after 400 epochs. All models were implemented in PyTorch [35] on a PC with an Nvidia GeForce 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.658, + 0.509 + ], + "angle": 0, + "content": "5.2. Model Analyses" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.517, + 0.798, + 0.533 + ], + "angle": 0, + "content": "5.2.1 Analyses of Mapping Degeneration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.75 + ], + "angle": 0, + "content": "We use synthetic images (simulated targets and real backgrounds [24]) to investigate the mapping degeneration phenomenon with respect to different characteristics of targets (i.e., intensity and size\\(^5\\)) and point labels (i.e., locations and numbers). We employ U-Net [41] as the baseline detection network, and use centroid point as GT label if not specified. We calculate IoU between positive pixels in predictions and GT mask labels of each epoch for quantitative analyses. In addition, we visualize the zoom-in target regions of simulated images with GT point labels (i.e., red dots) and corresponding CNN predictions (in the epoch reaching maximum IoU). To reduce training randomness, we show the average IoU results and visualization results over 100 runs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.892, + 0.842 + ], + "angle": 0, + "content": "Target Intensity. We simulate Gaussian-based extended targets with different peak values (i.e., 10, 50, 100, 200, 500) to investigate the influence of target intensity on mapping degeneration. Quantitative results in Fig. 5(a) show that intensity higher than 100 leads to a positive correlation between intensity and maximum IoU, while" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.876 + ], + "angle": 0, + "content": "4Focal loss is calculated between current evolved and GT labels to supervise the network training until the next round label update." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Shape, and local background clutter are investigated in supplemental material." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15532" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.079, + 0.283, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.177, + 0.264, + 0.189 + ], + "angle": 0, + "content": "(a) Initial Evolution Epoch \\( T_{\\text{epoch}} \\)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.079, + 0.498, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.177, + 0.458, + 0.189 + ], + "angle": 0, + "content": "(b) \\(T_{b}\\) of Evolution Threshold" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.079, + 0.686, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.178, + 0.666, + 0.188 + ], + "angle": 0, + "content": "(c) \\(k\\) of Evolution Threshold" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.08, + 0.892, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.178, + 0.851, + 0.188 + ], + "angle": 0, + "content": "(d) Evolution Frequency \\( f \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.198, + 0.89, + 0.227 + ], + "angle": 0, + "content": "Figure 6. \\( PA \\) (P) and \\( IoU \\) (I) results of LESPS with respect to (a) initial evolution epoch \\( T_{epoch} \\), (b) \\( T_b \\) and (c) \\( k \\) of evolution threshold, and (d) evolution frequency \\( f \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.24, + 0.468, + 0.296 + ], + "angle": 0, + "content": "Table 1. Average \\( IoU \\) (\\( \\times 10^{2} \\)), \\( P_{d} \\) (\\( \\times 10^{2} \\)) and \\( F_{a} \\) (\\( \\times 10^{6} \\)) values on NUAA-SIRST [8] NUDT-SIRST [24] and IRSTD-1K [54] achieved by DNA-Net with (w/) and without (w/o) LESPS under centroid, coarse point supervision together with full supervision." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.305, + 0.464, + 0.349 + ], + "angle": 0, + "content": "
MethodCentroidCoarseFull
IoUPdFaIoUPdFaIoUPdFa
w/o LESPS5.1289.190.682.9649.890.3075.6796.1822.94
w/ LESPS57.3491.8720.2456.1891.4918.32
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.468, + 0.456 + ], + "angle": 0, + "content": "lower intensity leads to a negative one. In addition, curve \"intensity 10\" reaches maximum \\( IoU \\) at around epoch 150 while others are less than 50, which demonstrates that over-small intensity decelerates degeneration. Visualization results show that our method can well highlight target regions under various intensities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.468, + 0.626 + ], + "angle": 0, + "content": "Target Size. We simulate Gaussian-based extended targets with different radii (i.e., 3, 5, 7, 9, 13) to investigate the influence of target size on mapping degeneration. Quantitative results in Fig. 5(b) show that larger target size leads to lower maximum IoU and less time to reach maximum IoU. That is because, size discrepancy between targets and GT point labels increases as target size increases, which aggravates and accelerates mapping degeneration. Visualization results show that CNNs can predict a cluster of pixels in a size-aware manner, and the peak values of predictions decrease as target size increases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.468, + 0.838 + ], + "angle": 0, + "content": "Locations of Point Label. We simulate a Gaussian-based extended target (with intensity 500 & radius 13), and place point labels at different distances away from the target centroid to investigate the influence of point label locations on mapping degeneration. Results in Fig. 5(c) show that small perturbations of label locations (less than 3 pixels) have a minor influence on the maximum IoU results. However, severe location perturbations (larger than 3 pixels) can lead to a significant maximum IoU drop, and the drop is more obvious when the point label is close to the edge. Note that, the same targets with different label locations reach maximum IoU at the same time, which demonstrates that the speed of mapping degeneration is irrelevant to the position of labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Number of Points in Label. We simulate a Gaussian-based extended target (with intensity 500 & radius 13) and set different numbers of points in labels to investigate its influence on mapping degeneration. Quantitative results" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.234, + 0.9, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.328, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Figure 7. Quantitative and qualitative results of evolved target masks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.49 + ], + "angle": 0, + "content": "in Fig. 5(d) show that as the number of points increases, CNNs can learn more target information to achieve higher maximum IoU results. In addition, the speed of mapping degeneration is irrelevant to the point number. Visualization results show that peak values of predictions increase as the number of points increases, which demonstrates that stronger supervision alleviates mapping degeneration. The conclusion well supports our label evolution framework." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.51, + 0.861, + 0.525 + ], + "angle": 0, + "content": "5.2.2 Analyses of the Label Evolution Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.623 + ], + "angle": 0, + "content": "In this subsection, we conduct experiments to investigate the effectiveness and the optimal parameter settings of our label evolution framework (i.e., LESPS). We employ \\( PA \\) and \\( IoU \\) between the positive pixels in updated labels and the GT mask labels to quantitatively evaluate the accuracy and expansion degree of the current label." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.893, + 0.866 + ], + "angle": 0, + "content": "Effectiveness. We compare the average results of NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] datasets achieved by DNA-Net with (i.e., w/) and without (i.e., w/o) LESPS under centroid and coarse point supervision, respectively. Note that, the results of DNA-Net w/o LESPS are calculated under extremely low threshold\\(^6\\) (i.e., 0.15) while those of DNA-Net w/ LESPS are calculated under the standard threshold (i.e., 0.5 [24, 54]). As shown in Table 1, as compared to full supervision, the results of DNA-Net w/o LESPS are extremely low, which demonstrates that SIRST detection on single point supervision is a challenging task. In contrast, DNA-Net w/ LESPS can achieve significant performance improvement under both point supervisions in terms of all the metrics, which approximate the performance of their fully supervised counterparts. Note that, \\(P_d\\) results of DNA" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.899 + ], + "angle": 0, + "content": "6With point supervision, the results of DNA-Net w/o LESPS calculated under threshold 0.5 are all zeros." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.08, + 0.476, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.259 + ], + "angle": 0, + "content": "Figure 8. Visualizations of regressed labels during training and network predictions during inference with centroid and coarse point supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.469, + 0.351 + ], + "angle": 0, + "content": "Net w/o LESPS under coarse point supervision are over half lower than those under the centroid ones, while the results of DNA-Net w/ LESPS under these two kinds of point supervision are comparable. It demonstrates that LESPS can generalize well to manual annotation errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.64 + ], + "angle": 0, + "content": "In addition, we make evaluations of evolved target masks during training. Quantitative results in Fig. 7(a) show average IoU values between positive pixels of evolved target masks and GT labels in 20-time training, which demonstrates that the quality of pseudo target masks consecutively increases during training. Qualitative results in Fig. 7(b) demonstrate that networks can effectively learn to expand point labels to mask labels. Furthermore, we visualize the labels regressed by our LESPS during training together with some network predictions during inference in Figs. 8 (a) and (b). As shown in Fig. 8(a), compared with GT mask labels, the evolved labels are more closely aligned with the objects in images (e.g., GT masks of Misc_4, XDU113 exceed the target regions due to visual edge ambiguity), which demonstrates that LESPS can alleviate the manually annotated errors. Figure 8(b) shows that DNA-Net w/ LESPS can effectively achieve accurate pixel-level SIRST detection in an end-to-end manner. Please refer to the supplemental materials for more visual results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Initial Evolution Epoch. We investigate the optimal value of epoch-based threshold \\( T_{\\text{epoch}} \\). Figure 6(a) shows that small initial evolution epoch results in a significant difference between \\( PA \\) and \\( IoU \\) (i.e., 0.94 vs. 0.04 with \\( T_{\\text{epoch}} = 10 \\)). That is because, early label evolution introduces many error pixels, resulting in severe error accumulation and network convergence failure. Increasing initial evolution epoch can reduce error accumulation and promote network convergence (0.60 vs. 0.54 with \\( T_{\\text{epoch}} = 50 \\)). However, over-large initial evolution epoch (i.e., a high degree of mapping degeneration) results in inferior performance (0.21 vs. 0.21 with \\( T_{\\text{epoch}} = 100 \\)). Therefore, \\( T_{\\text{epoch}} \\) is set to 50 in our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evolution Threshold. We investigate the optimal values of \\( k \\) and \\( T_{b} \\) in the evolution threshold. \\( T_{b} \\) is the minimum threshold, and controls evolution speed and error accumulation degree. \\( k \\) determines the maximum" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.078, + 0.896, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.278, + 0.893, + 0.335 + ], + "angle": 0, + "content": "Figure 9. Visual detection results of different methods achieved on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets. Correctly detected targets and false alarms are highlighted by red and orange circles, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.893, + 0.409 + ], + "angle": 0, + "content": "threshold, and controls the growth rate of the threshold. As shown in Fig. 6(b) and 6(c), both over-large and over-small values of \\( T_{b} \\) and \\( k \\) result in inferior performance. Therefore, we choose \\( k = 1 / 2 \\) and \\( T_{b} = 0.5 \\) in our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.893, + 0.562 + ], + "angle": 0, + "content": "Evolution Frequency. We investigate the optimal value of evolution frequency \\( f \\). Figure 6(d) shows that evolution frequency is positively correlated to \\( PA \\) and \\( IoU \\). However, high evolution frequency needs more time for label updates. To balance performance and efficiency, we choose \\( f = 5 \\) in our method. Interestingly, higher frequency (i.e., \\( f = 2 \\)) does not cause serve error accumulation, which also demonstrates the effectiveness of the convergence conditions of our LESPS. Please refer to the supplemental materials for more discussions of the convergence issue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.845, + 0.59 + ], + "angle": 0, + "content": "5.3. Comparison to State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Comparison to SISRT detection methods. We apply our LESPS to several state-of-the-art CNN-based methods, including ACM [7], ALCNet [8] and DNA-Net [24]. For fair comparisons, we retrained all models on NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] datasets with the same settings. In addition, we add the results of six fully supervised CNN-based methods (MDvsFA [45], ACM [7], ALCNet [8], DNA-Net [24], ISNet [54], UIU-Net [46]) and six traditional methods (Top-Hat [40], RLCM [15], TLLCM [16], MSPCM [17], IPI [11], PSTNN [51]) as the baseline results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Quantitative results in Table 2 show that CNN-based methods equipped with LESPS can outperform all the traditional methods. In addition, they can also achieve \\(71 - 75\\%\\) IoU results and comparable \\(P_{d}\\) and \\(F_{a}\\) results of their fully supervised counterparts. Qualitative results in Fig. 9 show that CNN-based methods equipped with LESPS can produce outputs with precise target mask and low false alarm rate, and can generalize well to complex scenes. Please refer to supplemental materials for more quantitative" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15534" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.082, + 0.892, + 0.125 + ], + "angle": 0, + "content": "Table 2. IoU \\((\\times 10^{2})\\), \\(P_{d}\\) \\((\\times 10^{2})\\) and \\(F_{a}(\\times 10^{6})\\) values of different methods achieved on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets. \"CNN Full\", \"CNN Centroid\", and \"CNN Coarse\" represent CNN-based methods under full supervision, centroid and coarse point supervision. \"+\" represents CNN-based methods equipped with LESPS." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.128, + 0.885, + 0.336 + ], + "angle": 0, + "content": "
MethodsDescriptionNUAA-SIRST [8]NUDT-SIRST [24]IRSTD-1K [54]Average
IoUPdFaIoUPdFaIoUPdFaIoUPdFa
Top-Hat [40]Filtering7.1479.841012.0020.7278.41166.7010.0675.111432.0012.6477.79870.23
RLCM [15]Local Contrast21.0280.61199.1515.1466.35163.0014.6265.6617.9516.0668.7098.77
TLLCM [16]Local Contrast11.0379.477.277.0662.0146.125.3663.974.937.2265.4521.42
MSPCM [34]Local Contrast12.3883.2717.775.8655.87115.967.3360.2715.247.2361.5355.13
IPI [11]Low Rank25.6785.5511.4717.7674.4941.2327.9281.3716.1823.7880.4722.96
PSTNN [51]Low Rank22.4077.9529.1114.8566.1344.1724.5771.9935.2620.6172.0236.18
MDvsFA [45]CNN Full61.7792.4064.9045.3886.03200.7135.4085.8699.2247.5288.10121.61
ISNet [54]CNN Full72.0494.6842.4671.2796.9396.8460.6194.2861.2867.9795.3066.86
UIU-Net [46]CNN Full69.9095.8251.2075.9196.8318.6161.1192.9326.8768.9795.1932.23
ACM [7]CNN Full64.9290.8712.7657.4291.7539.7357.4991.5843.8659.9491.4032.12
CNN Centroid+49.2389.3540.9542.0991.1138.2441.4488.8960.4644.2589.7846.55
CNN Coarse+47.8188.2140.7540.6481.1149.4540.3792.5964.8142.9487.3051.67
ALCNet [8]CNN Full67.9192.7837.0461.7891.3236.3662.0390.9142.4663.9191.6738.62
CNN Centroid+50.6292.0236.8441.5892.2867.0144.9090.5784.6845.7091.6262.84
CNN Coarse+51.0090.8742.4044.1492.8032.1046.7592.2664.3047.3091.9846.27
DNA-Net [24]CNN Full76.8696.9622.587.4298.3124.562.7393.2721.8175.6796.1822.94
CNN Centroid+61.9592.0218.1757.9994.7126.4552.0988.8816.0957.3491.8720.24
CNN Coarse+61.1393.1611.8758.3793.7628.0149.0587.5415.0756.1891.4918.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.422 + ], + "angle": 0, + "content": "Table 3. Average IoU \\((\\times 10^{2})\\), \\(P_{d}\\) \\((\\times 10^{2})\\), \\(F_{a}(\\times 10^{6})\\) values on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets of DNA-Net trained with pseudo labels generated by input intensity threshold, LCM-based methods [15, 16, 34] and LESPS under centroid and coarse point supervision." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.521 + ], + "angle": 0, + "content": "
Pseudo LabelCentroidCoarse
IoUPdFaIoUPdFa
Threshold=0.34.9281.7813.185.6783.1211.98
Threshold=0.513.2473.085.3115.5476.034.89
Threshold=0.714.5145.504.2815.2146.883.84
RLCM [15]21.4389.102.6722.5390.563.69
TLLCM [16]21.9590.967.7226.0594.154.27
MSPCM [34]28.8992.623.8429.7993.952.28
LESPS(ours)57.3491.8720.2456.1891.4918.32
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.228, + 0.553 + ], + "angle": 0, + "content": "and qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.47, + 0.72 + ], + "angle": 0, + "content": "Comparison to other pseudo labels. We compare our dynamic updated pseudo labels with fixed pseudo labels generated by input intensity threshold and local contrast-based methods [15, 16, 34]. Specifically, given a GT point label, we only preserve the eight connected regions of detection results that have union pixels with the GT point label. Then, we employ the pseudo labels to retrain DNA-Net [24] from scratch. As shown in Table 3, compared with fixed pseudo labels, dynamic updated pseudo labels by LESPS can achieve the highest IoU values with comparable \\( P_{d} \\) and reasonable \\( F_{a} \\) increase." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.327, + 0.747 + ], + "angle": 0, + "content": "5.4. Discussion of Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.864 + ], + "angle": 0, + "content": "In this subsection, we investigate the loss function of computing negative loss on different background points. Average results of different baseline methods under centroid point supervision are shown in Table 4. Extremely limited handcrafted background points7 leads to many false alarms. Random sample8 introduces more background points and well alleviates class imbalance, resulting in great" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.351, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Table 4. Average \\( IoU(\\times 10^{2}) \\), \\( P_{d}(\\times 10^{2}) \\), \\( F_{a}(\\times 10^{3}) \\) values on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets of different methods when computing negative loss on \\( i \\) handcrafted \\( (\\mathrm{hand}_i) \\), \\( j \\) randomly sampled \\( (\\mathrm{rand}_j) \\) and all background points." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.521 + ], + "angle": 0, + "content": "
AnnotationsACMALCNetDNA-Net
IoUPdFaIoUPdFaIoUPdFa
hand10.5495.7947.060.1695.19262.061.4397.8026.91
hand20.1297.24295.170.1596.62248.053.4198.188.48
hand50.1196.36316.370.0897.29363.253.6898.137.29
rand18.0693.453.568.5792.973.2118.7494.690.58
rand210.7892.722.2210.7891.162.7122.8594.810.42
rand513.3992.661.3511.8793.260.8924.8095.000.34
All (Ours)3.9587.150.024.0888.930.025.1289.180.01
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.598 + ], + "angle": 0, + "content": "performance improvements. However, the above simple versions introduce huge false alarms (34-1.8K times of all points), which are not practical for real applications, but inspire further thorough investigation in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.619, + 0.633 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we proposed the first work to achieve weakly-supervised SIRST detection using single-point supervision. In our method, we discovered the mapping degeneration phenomenon and proposed a label evolution framework named label evolution with single point supervision (LESPS) to automatically achieve point-to-mask regression. Through LESPS, networks can be trained to achieve SIRST detection in an end-to-end manner. Extensive experiments and insightful visualizations have fully demonstrated the effectiveness and superiority of our method. In addition, our method can be applied to different networks to achieve over \\(70\\%\\) and \\(95\\%\\) of its fully supervised performance on pixel-level IoU and object-level \\(P_{d}\\), respectively. We hope our interesting findings and promising results can inspire researchers to rethink the feasibility of achieving state-of-the-art performance in SIRST detection with much weaker supervision." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.425, + 0.888 + ], + "angle": 0, + "content": "7Points are sampled near targets, and are fixed during training." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.888, + 0.456, + 0.9 + ], + "angle": 0, + "content": "8Points are randomly sampled, and change in each training iteration." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.456, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15535" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Peri Akiva, Kristin Dana, Peter Oudemans, and Michael Mars. Finding berries: Segmentation and counting of cranberries using point supervision and shape priors. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 50-51, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In European Conference on Computer Vision (ECCV), pages 549-565. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] CL Philip Chen, Hong Li, Yantao Wei, Tian Xia, and Yuan Yan Tang. A local contrast method for small infrared target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 52(1):574-581, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8823-8832, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.414 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2617-2626, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[6] Yimian Dai and Yiquan Wu. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTAR), 10(8):3752-3767, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.472, + 0.543 + ], + "angle": 0, + "content": "[7] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[8] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), pages 1-12, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[9] Suyog D Deshpande, Meng Hwa Er, Ronda Venkateswarlu, and Philip Chan. Max-mean and max-median filters for detection of small targets. In Signal and Data Processing of Small Targets, volume 3809, pages 74-83. SPIE, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[10] Junsong Fan, Zhaoxiang Zhang, and Tieniu Tan. Pointly-supervised panoptic segmentation. ArXiv Preprint ArXiv:2210.13950, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[11] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE Transactions on Image Processing (TIP), 22(12):4996-5009, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[12] Yongtao Ge, Qiang Zhou, Xinlong Wang, Chunhua Shen, Zhibin Wang, and Hao Li. Point-teaching: Weakly semi-supervised object detection with point annotations. ArXiv Preprint ArXiv:2206.00274, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[13] Ross Girshick. Fast r-cnn. In IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In IEEE Conference on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition (CVPR), pages 580-587, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[15] Jinhui Han, Kun Liang, Bo Zhou, Xinying Zhu, Jie Zhao, and Linlin Zhao. Infrared small target detection utilizing the multiscale relative local contrast measure. IEEE Geoscience and Remote Sensing Letters (GRSL), 15(4):612-616, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Jinhui Han, Saed Moradi, Iman Faramarzi, Chengyin Liu, Honghui Zhang, and Qian Zhao. A local contrast method for infrared small-target detection utilizing a tri-layer window. IEEE Geoscience and Remote Sensing Letters (GRSL), 17(10):1822-1826, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[17] Jinhui Han, Saed Moradi, Iman Faramarzi, Honghui Zhang, Qian Zhao, Xiaojian Zhang, and Nan Li. Infrared small target detection based on the weighted strengthened local contrast measure. IEEE Geoscience and Remote Sensing Letters (GRSL), 18(9):1670-1674, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[18] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems (NeurIPS), 34:15908-15919, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In IEEE International Conference on Computer Vision (ICCV), pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[20] Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Haoliang Zheng, and Wei Zhang. Ristdnet: Robust infrared small target detection network. IEEE Geoscience and Remote Sensing Letters (GRSL), 19:1-5, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[21] Moran Ju, Jiangning Luo, Guangqi Liu, and Haibo Luo. Istdet: An efficient end-to-end neural network for infrared small target detection. Infrared Physics & Technology, 114:103659, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[23] Issam H Laradji, Negar Rostamzadeh, Pedro O Pinheiro, David Vazquez, and Mark Schmidt. Where are the blobs: Counting by localization with point supervision. In European Conference on Computer Vision (ECCV), pages 547-562, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[24] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing (TIP), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[25] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, et al. Yolov6: A single-stage object detection framework for industrial applications. ArXiv Preprint ArXiv:2209.02976, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15536" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Ting Liu, Jungang Yang, Boyang Li, Chao Xiao, Yang Sun, Yingqian Wang, and Wei An. Non-convex tensor low-rank approximation for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision (ECCV), pages 21-37. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.292 + ], + "angle": 0, + "content": "[30] Yongtuo Liu, Dan Xu, Sucheng Ren, Hanjie Wu, Hongmin Cai, and Shengfeng He. Fine-grained domain adaptive crowd counting via point-derived segmentation. ArXiv Preprint ArXiv:2108.02980, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.47, + 0.363 + ], + "angle": 0, + "content": "[31] Kevis-Kokitsi Maninis, Sergi Caelles, Jordi Pont-Tuset, and Luc Van Gool. Deep extreme cut: From extreme points to object segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 616-625, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[32] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[33] Saed Moradi, Payman Moallem, and Mohamad Farzan Sabahi. A false-alarm aware methodology to develop robust and efficient multi-scale infrared small target detection algorithm. Infrared Physics & Technology, 89:387-397, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[34] Saed Moradi, Payman Moallem, and Mohamad Farzan Sabahi. A false-alarm aware methodology to develop robust and efficient multi-scale infrared small target detection algorithm. Infrared Physics & Technology, 89:387-397, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.469, + 0.666 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NeurlPS), 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.67, + 0.469, + 0.739 + ], + "angle": 0, + "content": "[36] Rui Qian, Yunchao Wei, Honghui Shi, Jiachen Li, Jiaying Liu, and Thomas Huang. Weakly supervised scene parsing with point-based distance metric learning. In AAAI Conference on Artificial Intelligence (AAAI), volume 33, pages 8843-8850, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.842 + ], + "angle": 0, + "content": "[38] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7263-7271, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[40] Jeanfrancois Rivest and Roger Fortin. Detection of dim targets in digital infrared imagery by morphological image processing. Optical Engineering, 35(7):1886-1893, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[42] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[43] Yang Sun, Jungang Yang, and Wei An. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor model. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 59(5):3737-3752, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[45] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In IEEE International Conference on Computer Vision (ICCV), pages 8509-8518, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[46] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: u-net in u-net for infrared small object detection. IEEE Transactions on Image Processing (TIP), 32:364-376, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[47] Xinyi Ying, Yingqian Wang, Longguang Wang, Weidong Sheng, Li Liu, Zaiping Lin, and Shilin Zhou. Local motion and contrast priors driven deep network for infrared small target super-resolution. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTAR), 15:5480-5495, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[48] Mohsen Zand, Haleh Damirchi, Andrew Farley, Mahdiyar Molahasani, Michael Greenspan, and Ali Etemad. Multiscale crowd counting and localization by multitask point supervision. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1820-1824. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.678, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[49] Ke Zhang, Shuyan Ni, Dashuang Yan, and Aidi Zhang. Review of dim small target detection algorithms in single-frame infrared images. In IEEE Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC), volume 4, pages 2115-2120. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[50] Landan Zhang, Lingbing Peng, Tianfang Zhang, Siying Cao, and Zhenming Peng. Infrared small target detection via nonconvex rank approximation minimization joint 12, 1 norm. Remote Sensing, 10(11):1821, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[51] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Mingjin Zhang, Haichen Bai, Jing Zhang, Rui Zhang, Chaoyue Wang, Jie Guo, and Xinbo Gao. Rkformer: Runge-kutta transformer with random-connection attention for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15537" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "infrared small target detection. In 30th ACM International Conference on Multimedia (ACM MM), pages 1730-1738, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[53] Mingjin Zhang, Ke Yue, Jing Zhang, Yunsong Li, and Xinbo Gao. Exploring feature compensation and cross-level correlation for infrared small target detection. In 30th ACM International Conference on Multimedia (ACM MM), pages 1857-1865, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[54] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 877-886, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[55] Shiyin Zhang, Jun Hao Liew, Yunchao Wei, Shikui Wei, and Yao Zhao. Interactive object segmentation with inside-outside guidance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12234-12244, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[56] Shilong Zhang, Zhuoran Yu, Liyang Liu, Xinjiang Wang, Aojun Zhou, and Kai Chen. Group r-cnn for weakly semi-supervised object detection with points. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9417-9426, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[57] Mingjing Zhao, Wei Li, Lu Li, Jin Hu, Pengge Ma, and Ran Tao. Single-frame infrared small-target detection: A survey. IEEE Geoscience and Remote Sensing Magazine (GRSM), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[58] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.47, + 0.586 + ], + "angle": 0, + "content": "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15538" + } + ] +] \ No newline at end of file diff --git a/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_origin.pdf b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a42f91985a7ba77763949494d42d513fedae6f14 --- /dev/null +++ b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/9bbb805e-162c-44dd-853a-dcb209359b00_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24aaa1944303649ef14510d42a2ced0ef54803a3cba7e9da22f7313be2dccc8f +size 3943449 diff --git a/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/full.md b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3160cd125824255447e864c3d78824d2265edd3c --- /dev/null +++ b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/full.md @@ -0,0 +1,300 @@ +# Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision + +Xinyi Ying $^{1}$ , Li Liu $^{1}$ , Yingqian Wang $^{1}$ , Ruojing Li $^{1}$ , Nuo Chen $^{1}$ , Zaiping Lin $^{1\boxtimes}$ , Weidong Sheng $^{1}$ , Shilin Zhou $^{1}$ + +$^{1}$ National University of Defense Technology + +{yingxinyi18, wangyingqian16, liruojing, chennuo97, linzaiping, slzhou}@nudt.edu.cn, + +dreamliu2010@gmail.com, shengweidong1111@sohu.com + +# Abstract + +Training a convolutional neural network (CNN) to detect infrared small targets in a fully supervised manner has gained remarkable research interests in recent years, but is highly labor expensive since a large number of per-pixel annotations are required. To handle this problem, in this paper, we make the first attempt to achieve infrared small target detection with point-level supervision. Interestingly, during the training phase supervised by point labels, we discover that CNNs first learn to segment a cluster of pixels near the targets, and then gradually converge to predict groundtruth point labels. Motivated by this "mapping degeneration" phenomenon, we propose a label evolution framework named label evolution with single point supervision (LESPS) to progressively expand the point label by leveraging the intermediate predictions of CNNs. In this way, the network predictions can finally approximate the updated pseudo labels, and a pixel-level target mask can be obtained to train CNNs in an end-to-end manner. We conduct extensive experiments with insightful visualizations to validate the effectiveness of our method. Experimental results show that CNNs equipped with LESPS can well recover the target masks from corresponding point labels, and can achieve over $70\%$ and $95\%$ of their fully supervised performance in terms of pixel-level intersection over union (IoU) and object-level probability of detection $(P_d)$ , respectively. Code is available at https://github.com/XinyiYing/LESPS. + +# 1. Introduction + +Infrared small target detection has been a longstanding, fundamental yet challenging task in infrared search and tracking systems, and has various important applications in civil and military fields [49,57], including traffic monitoring + +![](images/311a0b588c2647d95cc1bfc3f441ed182145828ef7344030fab3667934b3f1fc.jpg) +Figure 1. An illustration of mapping degeneration under point supervision. CNNs always tend to segment a cluster of pixels near the targets with low confidence at the early stage, and then gradually learn to predict GT point labels with high confidence. + +[24, 54], maritime rescue [52, 53] and military surveillance [7, 47]. Due to the rapid response and robustness to fast-moving scenes, single-frame infrared small target (SIRST) detection methods have always attracted much more attention, and numerous methods have been proposed. Early methods, including filtering-based [9, 40], local contrast-based [3, 16] and low rank-based [11, 43] methods, require complex handcrafted features with carefully tuned hyper-parameters. Recently, compact deep learning has been introduced in solving the problem of SIRST detection [24, 45, 54]. However, there are only a few attempts, and its potential remains locked, unlike the extensive explorations of deep learning for natural images. This is mainly due to potential reasons, including lack of large-scale, accurately annotated datasets and high stake application scenarios. + +Infrared small targets are usually of very small size, weak, shapeless and textureless, and are easily submerged in diverse complex background clutters. As a result, directly adopting existing popular generic object detectors like RCNN series [13, 14, 19, 39], YOLO series [25, 37, 38] and SSD [29] to SIRST detection cannot produce satisfactory performance. Realizing this, researchers have been focusing on developing deep networks tailored for + +infrared small targets by adequately utilizing the domain knowledge. However, most existing deep methods for SIRST detection [8, 24, 54] are fully supervised, which usually requires a large dataset with accurate target mask annotations for training. Clearly, this is costly [5, 26]. + +Therefore, a natural question arises: Can we develop a new framework for SIRST detection with single point supervision? In fact, to substantially reduce the annotation cost for object detection tasks, weakly supervised object detection methods with point supervision [4,5,26,56] have been studied in the field of computer vision. Although these weakly supervised methods achieve promising results, they are not designed for the problem of SIRST detection, and the class-agnostic labels (i.e., only foreground and background) of infrared small targets hinder their applications [42, 58]. Therefore, in this work, we intend to conduct the first study of weakly supervised SIRST detection with single-point supervision. + +A key motivation of this work comes from an interesting observation during the training of SIRST detection networks. That is, with single point labels serving as supervision, CNNs always tend to segment a cluster of pixels near the targets with low confidence at the early stage, and then gradually learn to predict groundtruth (GT) point labels with high confidence, as shown in Fig. 1. It reveals the fact that region-to-region mapping is the intermediate result of the final region-to-point mapping1. We attribute this "mapping degeneration" phenomenon to the special imaging mechanism of infrared system [24, 54], the local contrast prior of infrared small targets [3, 8], and the easy-to-hard learning property of CNNs [44], in which the first two factors result in extended mapping regions beyond the point labels, and the last factor contributes to the degeneration process. + +Based on the aforementioned discussion, in this work, we propose a novel framework for the problem of weakly supervised SIRST detection, dubbed label evolution with single point supervision (LESPS). Specifically, LESPS leverages the intermediate network predictions in the training phase to update the current labels, which serve as supervision until the next label update. Through iterative label update and network training, the network predictions can finally approximate the updated pseudo mask labels, and the network can be simultaneously trained to achieve pixel-level SIRST detection in an end-to-end $^2$ manner. + +Our main contributions are summarized as: (1) We present the first study of weakly supervised SIRST + +detection, and introduce LESPS that can significantly reduce the annotation cost. (2) We discover the mapping degeneration phenomenon, and leverage this phenomenon to automatically regress pixel-level pseudo labels from the given point labels via LESPS. (3) Experimental results show that our framework can be applied to different existing SIRST detection networks, and enable them to achieve over $70\%$ and $95\%$ of its fully supervised performance in terms of pixel-level intersection over union $(IoU)$ and object-level probability of detection $(P_d)$ , respectively. + +# 2. Related Work + +SIRST Detection. In the past decades, various methods have been proposed, including early traditional paradigms (e.g., filtering-based methods [9, 40], local contrast-based methods [3, 15-17, 33, 34], low rank-based methods [6, 11, 28, 43, 50, 51]) and recent deep learning paradigms [7, 8, 20, 21, 24, 45, 52-54]. Compared to traditional methods, which require delicately designed models and carefully tuned hyper-parameters, convolutional neural networks (CNNs) can learn the non-linear mapping between input images and GT labels in a data-driven manner, and thus generalize better to real complex scenes. As the pioneering work, Wang et al. [45] first employed a generative adversarial network to achieve a better trade-off between miss detection and false alarm. Recently, more works focus on customized solutions of infrared small target. Specifically, Dai et al. [7] specialized an asymmetric contextual module, and further incorporated local contrast measure [8] to improve the target contrast. Li et al. [24] preserved target information by repetitive feature fusion. Zhang et al. [54] aggregated edge information to achieve shape-aware SIRST detection. Zhang et al. [52, 53] explored cross-level correlation and transformer-based method [18] to predict accurate target mask. Wu et al. [46] customized a UIU-Net framework for multi-level and multi-scale feature aggregation. In conclusion, existing works generally focus on compact architectural designs to pursue superior performance in a fully supervised manner. However, due to the lack of a large number of public datasets [7, 24, 45] with per-pixel annotations, the performance and generalization of CNNs are limited. In addition, per-pixel manual annotations are time-consuming and labor-intensive. Therefore, we focus on achieving good pixel-level SIRST detection with weaker supervision and cheaper annotations. + +Weakly Supervised Segmentation with Points. Recently, point-level annotation has raised more attention in dense prediction tasks such as object detection [4, 12, 56], crowd counting [1, 23, 30, 48] and image segmentation [2,5,10,26,31,36,55]. We mainly focus on image segmentation in this paper. Specifically, Bearman et al. [2] made the first attempt to introduce an objectiveness potential into a pointly supervised training + +loss function to boost segmentation performance. Qian et al. [36] leveraged semantic information of several labeled points by a distance metric loss to achieve scene parsing. Zhang et al. [55] proposed an inside-outside guidance approach to achieve instance segmentation by five elaborate clicks. Cheng et al. [5] designed to provide ten randomly sampled binary point annotations within box annotations for instance segmentation. Li et al. [26] encoded each instance with kernel generator for panoptic segmentation to achieve $82\%$ of fully-supervised performance with only twenty randomly annotated points. In contrast to these approaches employing complicated prior constraints to segment large generic objects with rich color and fine textures by several elaborate points, we fully exploit the local contrast prior of infrared small target to progressively evolve pseudo masks by single coarse point without any auxiliaries in an end-to-end manner. + +# 3. The Mapping Degeneration Phenomenon + +In this section, we first describe the mapping degeneration phenomenon together with our intuitive explanation. Then we conduct experiments under single-sample and many-sample training schemes to demonstrate the generality of degeneration, and investigate the influence of generalization on degeneration. + +As shown in Fig. 1, given an input image and the corresponding GT point label, we employ U-Net [41] as the baseline SIRST detection network for training. It can be observed that, in the early training phase, network predicts a cluster of pixels near the targets with low confidence. As training continues, the network prediction finally approximates GT point label with gradually increased confidence. We name this phenomenon as "mapping degeneration", and attribute the following reasons to this phenomenon. 1) Special imaging mechanism of infrared systems [24, 54]: Targets only have intensity information without structure and texture details, resulting in highly similar pixels within the target region. 2) High local contrast of infrared small targets [3, 8]: Pixels within the target region are much brighter or darker with high contrast against the local background clutter. 3) Easy-to-hard learning property of CNNs [44]: CNNs always tend to learn simple mappings first, and then converge to difficult ones. Compared with region-to-point mapping, region-to-region mapping is easier, and thus tends to be the intermediate result of region-to-point mapping. In conclusion, the unique characteristics of infrared small targets result in extended mapping regions beyond point labels, and CNNs contribute to the mapping degeneration process. + +It is worth noting that the mapping degeneration phenomenon is a general phenomenon in various scenes with infrared small targets. Specifically, we use the training datasets (including 1676 images and their corresponding + +![](images/ef05e010f1f4414bbbbc0422c6fa759a2ec3b4e80ecd716c80cf32295d37aade.jpg) + +![](images/b9bac8a255a0d533de981dffd67bf15df661f747fe96041a0b7009c724a6be27.jpg) + +![](images/173cc11f5273b1d88187099bd6fd3dc3167e04e60b4de804b9bc261e789566db.jpg) +Figure 2. Quantitative and qualitative illustrations of mapping degeneration in CNNs. + +centroid point label, see details in Section 5.1) to train U-Net under a single-sample training scheme (i.e., training one CNN on each image). For quantitative analyses, we employ the IoU results between positive pixels in predictions (i.e., pixels with confidence higher than half of its maximum value) and GT mask label. Average IoU results of 1676 CNNs at each epoch are shown by the blue curve in Fig. 2(a), while the number of training samples with maximum IoU during training phase falling in a threshold range of $[i, i + 0.1]$ , $(i = 0, 0.1, \dots, 0.9)$ is illustrated via blue bars in Fig. 2(b). It can be observed from the zoom-in curve and bars that mapping degeneration is a general phenomenon with point supervision, and U-Net can achieve IoU $>0.5$ on more than $60\%$ of the training images. + +In addition, we conduct experiments to train U-Net under a many-sample training scheme (i.e., training one CNN using all images which contain abundant targets with various sizes and shapes) to investigate the effect of generalization on mapping degeneration. Average IoU results of 1676 images are shown by orange curve in Fig. 2(a). It can be observed that many-sample training scheme needs more time to converge. Moreover, Fig. 2(b) shows that orange bars are slightly lower than blue ones on larger IoU values (i.e., 0.5-1.0). It is demonstrated that generalization decelerates but aggravates mapping degeneration. Figure 2(c) shows some zoom-in target regions of images and their predictions under these two training schemes. It can be observed that CNNs can effectively segment a cluster of target pixels under both training schemes in a size-aware manner. + +Therefore, an intuitive assumption arises: Can we leverage the intermediate results of CNNs to regress masks? A simple early stopping strategy seems to be a positive answer but is indeed unpractical since mapping degeneration is influenced by various factors, including target intensity, size, shape, and local background clutter (see details in Section 5.2.1). Consequently, there is no fixed optimal stopping epoch for all situations. These + +![](images/44a112e497ff17aee764208a5f492eb679dc43de3430601165fbda42b8a02291.jpg) +Figure 3. An illustration of label evolution with single point supervision (LESPS). During training, intermediate predictions of CNNs are used to progressively expand point labels to mask labels. Black arrows represent each round of label updates. + +observations motivate us to design a special label evolution framework to well leverage the mapping degeneration for pseudo mask regression. + +# 4. The Label Evolution Framework + +Motivated by mapping degeneration, we propose a label evolution framework named label evolution with single point supervision (LESPS) to leverage the intermediate network predictions in the training phase to update labels. As training continues, the network predictions approximate the updated pseudo mask labels, and network can simultaneously learn to achieve pixel-level SIRST detection in an end-to-end manner. Here, we employ a toy example of 1D curves for easy understanding. As shown in Fig. 3, sub-figures on the left of the dotted line represent the network predictions. Note that, the black curves denote the intermediate predictions within LESPS, while the gray curves represent virtual results produced by the network without label update. On the right of the dotted line, the first and second columns of sub-figures represent current labels and updated labels, respectively, and black arrows represent each round of label update. The overall framework can be summarized as follows. With point label serving as supervision, in the $1^{st}$ round label update after initial training, the predictions are used to update the current point label to generate the $1^{st}$ updated label, which is then used to supervise the network training until the $2^{nd}$ round label update. Through iterative label updates and network training, CNNs can incorporate the local contrast prior to gradually recover the mask labels. From another viewpoint, label evolution consistently updates the supervision to prevent mapping degeneration, and promotes CNNs to converge to the easy region-to-region mapping. + +Taking the $n^{th}$ update as an example, given the current label $L_{n}$ and the network prediction $P_{n}$ , we perform label update for each target, which consists of three steps: candidate pixel extraction, false alarm elimination, and weighted summation between candidate pixels and current + +![](images/7c0aa4c6310d39eb21011d872084a1cbb9aab640f0216f50370c0a6f052a8356.jpg) +Figure 4. (a) Adaptive threshold $T_{\text{adapt}}$ with respect to positive pixels $\hat{L}_n^i$ and hyper-parameters $k, T_b$ . Pink dotted line represents the constant $hwr$ . (b) An illustration of false alarm elimination. Red circle and dot represent positive pixels and centroid point of label. Orange circle represents false alarms. + +labels. Specifically, the $d \times d$ local neighborhoods of the $i^{th}$ target in label $L_{n}$ and prediction $P_{n}$ are cropped based on the centroid of the positive pixels $^{3}$ in label (i.e., $\hat{L}_n$ ). Then to reduce error accumulation for label update (see Section 5.2.2 for details), we employ an adaptive threshold (the red dotted line in Fig. 3) to extract the local neighborhood candidate pixels (predictions higher than the red dotted line in Fig. 3). The process can be defined as: + +$$ +C _ {n} ^ {i} = P _ {n} ^ {i} \odot \left(P _ {n} ^ {i} > T _ {\text {a d a p t}}\right), \tag {1} +$$ + +where $C_n^i$ is the candidate pixels, and $\odot$ represents elementwise multiplication. $T_{adapt}$ is the adaptive threshold that correlated to the current prediction $P_n^i$ and the positive pixels in label $\hat{L}_n^i$ , and can be calculated according to: + +$$ +T _ {\text {a d a p t}} = \max \left(P _ {n} ^ {i}\right) \left(T _ {b} + k \left(1 - T _ {b}\right) \hat {L} _ {n} ^ {i} / (h w r)\right), \tag {2} +$$ + +where $h$ , $w$ are the height and width of input images, and $r$ is set to $0.15\%$ [7, 24]. As shown in Fig. 4 (a), $T_{b}$ is the minimum threshold, and $k$ controls the threshold growth rate. An increasing number of $\hat{L}_n^i$ leads to the increase of the threshold, which can reduce error accumulation of low contrast targets and strong background clutter. + +To eliminate false alarms by local neighborhood noise, we exclude the eight connective regions of candidate pixels that have no intersection with positive pixels of labels, as shown in Fig. 4 (b). This process is defined as: + +$$ +E _ {n} ^ {i} = C _ {n} ^ {i} \odot F _ {n} ^ {i}, \tag {3} +$$ + +where $E_{n}^{i}$ is the candidate pixels after false alarm elimination, and $F_{n}^{i}$ is the mask against false alarm pixels. + +We then perform average weighted summation between candidate pixels $E_{n}^{i}$ and current label $L_{n}^{i}$ to achieve label update. The process can be formulated as: + +$$ +L _ {n + 1} ^ {i} = L _ {n} ^ {i} \odot (1 - N _ {n} ^ {i}) + \frac {L _ {n} ^ {i} + E _ {n} ^ {i}}{2} \odot N _ {n} ^ {i}, \tag {4} +$$ + +where $L_{n + 1}^{i}$ is the updated label in the $n^{th}$ round, which serves as new supervision for training in the $n + 1^{th}$ round, + +![](images/d9c35bfb351b65feaf705240ff75f4627cc300197d6c3e8cbce41fd3e4d21f12.jpg) +(a) Target Intensity +Figure 5. IoU and visualize results of mapping degeneration with respect to different characteristics of targets (i.e.,(a) intensity, (b) size,) and point labels (i.e.,(c) locations and (d) numbers). We visualize the zoom-in target regions of input images with GT point labels (i.e., red dots in images) and corresponding CNN predictions (in the epoch reaching maximum IoU). + +![](images/cdbedc4ef76e60160481931442ab8be8f0ece18250064ce3f1a3faa818929a83.jpg) +(b) Target Size + +![](images/31311459c1a3ebab9c120d00bb84e4a812848b1ae9b14b36dbb6e963fa4fb605.jpg) +(c) Locations of Point label + +![](images/c4c711bfccfd9b463477f5a3b3a8aa0a832f03359cac6066dd1426bc44860339.jpg) +(d) Numbers of Point Label + +and $N_{n}^{i} = \left(P_{n}^{i} > T_{\text{adapt}}\right) \odot F_{n}^{i}$ . Note that, the first term represents GT labels below red dotted lines, and the second term represents the average weighted summation between predictions and GT labels above red dotted lines. + +It is worth noting that we provide three conditions to ensure network convergence: 1) Average weighted summation between predictions and labels promotes CNNs to converge as predictions approximate labels. 2) Pixel-adaptive threshold increases with the increase of positive pixels in updated labels, which slows down or suspends the label update. 3) As label evolution introduces more target information for training, CNNs grow to mature, and learn to distinguish targets from backgrounds. + +We start label evolution after initial evolution epoch $T_{epoch}$ , and perform label update every $f$ epoch until the end of training. Note that, our epoch-based threshold $T_{epoch}$ is a coarse threshold to ensure that networks attend to targets instead of background clutter. + +# 5. Experiments + +In this section, we first describe the implementation details, and then make comprehensive analyses of the mapping degeneration phenomenon and our label evolution framework. In addition, we apply our method to the state-of-the-art SIRST detection methods with point supervision, and make comparisons with their fully supervised counterparts. Moreover, we make comparisons of our dynamic updated pseudo labels with fixed pseudo labels, and discuss the calculation of loss function. + +# 5.1. Implementation Details + +Three public datasets NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] are used in our experiments. We followed [24] to split the training and test sets of NUAA-SIRST and NUDT-SIRST, and followed [54] to split IRSTD-1K. We employed two pixel-level metrics (i.e., intersection over union $(IoU)$ and pixel accuracy $(PA))$ and two target-level metrics (i.e., probability of detection $(P_d)$ and false-alarm rate $(F_a))$ for performance evaluation. + +During training, all images were normalized and randomly cropped into patches of size $256 \times 256$ as network inputs. We augmented the training data by random flipping and rotation. Due to the extreme positive-negative sample imbalance (less than 10 vs. more than $256 \times 256$ ) in SIRST detection with point supervision, we employed focal loss[27] to stabilize the training process. All the networks were optimized by the Adam method [22]. Batch size was set to 16, and learning rate was initially set to $5 \times 10^{-4}$ and reduced by ten times at the $200^{th}$ and $300^{th}$ epochs. We stopped training after 400 epochs. All models were implemented in PyTorch [35] on a PC with an Nvidia GeForce 3090 GPU. + +# 5.2. Model Analyses + +# 5.2.1 Analyses of Mapping Degeneration + +We use synthetic images (simulated targets and real backgrounds [24]) to investigate the mapping degeneration phenomenon with respect to different characteristics of targets (i.e., intensity and size $^5$ ) and point labels (i.e., locations and numbers). We employ U-Net [41] as the baseline detection network, and use centroid point as GT label if not specified. We calculate IoU between positive pixels in predictions and GT mask labels of each epoch for quantitative analyses. In addition, we visualize the zoom-in target regions of simulated images with GT point labels (i.e., red dots) and corresponding CNN predictions (in the epoch reaching maximum IoU). To reduce training randomness, we show the average IoU results and visualization results over 100 runs. + +Target Intensity. We simulate Gaussian-based extended targets with different peak values (i.e., 10, 50, 100, 200, 500) to investigate the influence of target intensity on mapping degeneration. Quantitative results in Fig. 5(a) show that intensity higher than 100 leads to a positive correlation between intensity and maximum IoU, while + +![](images/5eb51815d988cb84a58c1be6c49ca872022a045e720610234f44fe32c55072ae.jpg) +(a) Initial Evolution Epoch $T_{\text{epoch}}$ + +![](images/31839d592377a44cbb7a23c802f75c69f7a776e70bfaed689d57f2f1b9604a14.jpg) +(b) $T_{b}$ of Evolution Threshold + +![](images/34a1a6f1452c80a2b434298440d647e79534e537f3a95da49d157f26c31d7a49.jpg) +(c) $k$ of Evolution Threshold + +![](images/34a9f9aefc047fbfce0285991b847d0b04da3886369f943d3e44408f3a5c121b.jpg) +(d) Evolution Frequency $f$ + +Table 1. Average $IoU$ ( $\times 10^{2}$ ), $P_{d}$ ( $\times 10^{2}$ ) and $F_{a}$ ( $\times 10^{6}$ ) values on NUAA-SIRST [8] NUDT-SIRST [24] and IRSTD-1K [54] achieved by DNA-Net with (w/) and without (w/o) LESPS under centroid, coarse point supervision together with full supervision. + +
MethodCentroidCoarseFull
IoUPdFaIoUPdFaIoUPdFa
w/o LESPS5.1289.190.682.9649.890.3075.6796.1822.94
w/ LESPS57.3491.8720.2456.1891.4918.32
+ +lower intensity leads to a negative one. In addition, curve "intensity 10" reaches maximum $IoU$ at around epoch 150 while others are less than 50, which demonstrates that over-small intensity decelerates degeneration. Visualization results show that our method can well highlight target regions under various intensities. + +Target Size. We simulate Gaussian-based extended targets with different radii (i.e., 3, 5, 7, 9, 13) to investigate the influence of target size on mapping degeneration. Quantitative results in Fig. 5(b) show that larger target size leads to lower maximum IoU and less time to reach maximum IoU. That is because, size discrepancy between targets and GT point labels increases as target size increases, which aggravates and accelerates mapping degeneration. Visualization results show that CNNs can predict a cluster of pixels in a size-aware manner, and the peak values of predictions decrease as target size increases. + +Locations of Point Label. We simulate a Gaussian-based extended target (with intensity 500 & radius 13), and place point labels at different distances away from the target centroid to investigate the influence of point label locations on mapping degeneration. Results in Fig. 5(c) show that small perturbations of label locations (less than 3 pixels) have a minor influence on the maximum IoU results. However, severe location perturbations (larger than 3 pixels) can lead to a significant maximum IoU drop, and the drop is more obvious when the point label is close to the edge. Note that, the same targets with different label locations reach maximum IoU at the same time, which demonstrates that the speed of mapping degeneration is irrelevant to the position of labels. + +Number of Points in Label. We simulate a Gaussian-based extended target (with intensity 500 & radius 13) and set different numbers of points in labels to investigate its influence on mapping degeneration. Quantitative results + +![](images/111d120cbbbbf643f11514976d8c732ea735d336081aa437ccbcc119ea041dc9.jpg) +Figure 6. $PA$ (P) and $IoU$ (I) results of LESPS with respect to (a) initial evolution epoch $T_{epoch}$ , (b) $T_b$ and (c) $k$ of evolution threshold, and (d) evolution frequency $f$ . +Figure 7. Quantitative and qualitative results of evolved target masks. + +in Fig. 5(d) show that as the number of points increases, CNNs can learn more target information to achieve higher maximum IoU results. In addition, the speed of mapping degeneration is irrelevant to the point number. Visualization results show that peak values of predictions increase as the number of points increases, which demonstrates that stronger supervision alleviates mapping degeneration. The conclusion well supports our label evolution framework. + +# 5.2.2 Analyses of the Label Evolution Framework + +In this subsection, we conduct experiments to investigate the effectiveness and the optimal parameter settings of our label evolution framework (i.e., LESPS). We employ $PA$ and $IoU$ between the positive pixels in updated labels and the GT mask labels to quantitatively evaluate the accuracy and expansion degree of the current label. + +Effectiveness. We compare the average results of NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] datasets achieved by DNA-Net with (i.e., w/) and without (i.e., w/o) LESPS under centroid and coarse point supervision, respectively. Note that, the results of DNA-Net w/o LESPS are calculated under extremely low threshold $^6$ (i.e., 0.15) while those of DNA-Net w/ LESPS are calculated under the standard threshold (i.e., 0.5 [24, 54]). As shown in Table 1, as compared to full supervision, the results of DNA-Net w/o LESPS are extremely low, which demonstrates that SIRST detection on single point supervision is a challenging task. In contrast, DNA-Net w/ LESPS can achieve significant performance improvement under both point supervisions in terms of all the metrics, which approximate the performance of their fully supervised counterparts. Note that, $P_d$ results of DNA + +![](images/b03b94a70ab9a87594864d09604fc6d33d9914caaff4c659037cb559e3d013ea.jpg) +Figure 8. Visualizations of regressed labels during training and network predictions during inference with centroid and coarse point supervision. + +Net w/o LESPS under coarse point supervision are over half lower than those under the centroid ones, while the results of DNA-Net w/ LESPS under these two kinds of point supervision are comparable. It demonstrates that LESPS can generalize well to manual annotation errors. + +In addition, we make evaluations of evolved target masks during training. Quantitative results in Fig. 7(a) show average IoU values between positive pixels of evolved target masks and GT labels in 20-time training, which demonstrates that the quality of pseudo target masks consecutively increases during training. Qualitative results in Fig. 7(b) demonstrate that networks can effectively learn to expand point labels to mask labels. Furthermore, we visualize the labels regressed by our LESPS during training together with some network predictions during inference in Figs. 8 (a) and (b). As shown in Fig. 8(a), compared with GT mask labels, the evolved labels are more closely aligned with the objects in images (e.g., GT masks of Misc_4, XDU113 exceed the target regions due to visual edge ambiguity), which demonstrates that LESPS can alleviate the manually annotated errors. Figure 8(b) shows that DNA-Net w/ LESPS can effectively achieve accurate pixel-level SIRST detection in an end-to-end manner. Please refer to the supplemental materials for more visual results. + +Initial Evolution Epoch. We investigate the optimal value of epoch-based threshold $T_{\text{epoch}}$ . Figure 6(a) shows that small initial evolution epoch results in a significant difference between $PA$ and $IoU$ (i.e., 0.94 vs. 0.04 with $T_{\text{epoch}} = 10$ ). That is because, early label evolution introduces many error pixels, resulting in severe error accumulation and network convergence failure. Increasing initial evolution epoch can reduce error accumulation and promote network convergence (0.60 vs. 0.54 with $T_{\text{epoch}} = 50$ ). However, over-large initial evolution epoch (i.e., a high degree of mapping degeneration) results in inferior performance (0.21 vs. 0.21 with $T_{\text{epoch}} = 100$ ). Therefore, $T_{\text{epoch}}$ is set to 50 in our method. + +Evolution Threshold. We investigate the optimal values of $k$ and $T_{b}$ in the evolution threshold. $T_{b}$ is the minimum threshold, and controls evolution speed and error accumulation degree. $k$ determines the maximum + +![](images/7bbb71eaa0d7dc01f32000feab13282c294b7b925787e3fc799aa8ac20107b0c.jpg) +Figure 9. Visual detection results of different methods achieved on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets. Correctly detected targets and false alarms are highlighted by red and orange circles, respectively. + +threshold, and controls the growth rate of the threshold. As shown in Fig. 6(b) and 6(c), both over-large and over-small values of $T_{b}$ and $k$ result in inferior performance. Therefore, we choose $k = 1 / 2$ and $T_{b} = 0.5$ in our method. + +Evolution Frequency. We investigate the optimal value of evolution frequency $f$ . Figure 6(d) shows that evolution frequency is positively correlated to $PA$ and $IoU$ . However, high evolution frequency needs more time for label updates. To balance performance and efficiency, we choose $f = 5$ in our method. Interestingly, higher frequency (i.e., $f = 2$ ) does not cause serve error accumulation, which also demonstrates the effectiveness of the convergence conditions of our LESPS. Please refer to the supplemental materials for more discussions of the convergence issue. + +# 5.3. Comparison to State-of-the-art Methods + +Comparison to SISRT detection methods. We apply our LESPS to several state-of-the-art CNN-based methods, including ACM [7], ALCNet [8] and DNA-Net [24]. For fair comparisons, we retrained all models on NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] datasets with the same settings. In addition, we add the results of six fully supervised CNN-based methods (MDvsFA [45], ACM [7], ALCNet [8], DNA-Net [24], ISNet [54], UIU-Net [46]) and six traditional methods (Top-Hat [40], RLCM [15], TLLCM [16], MSPCM [17], IPI [11], PSTNN [51]) as the baseline results. + +Quantitative results in Table 2 show that CNN-based methods equipped with LESPS can outperform all the traditional methods. In addition, they can also achieve $71 - 75\%$ IoU results and comparable $P_{d}$ and $F_{a}$ results of their fully supervised counterparts. Qualitative results in Fig. 9 show that CNN-based methods equipped with LESPS can produce outputs with precise target mask and low false alarm rate, and can generalize well to complex scenes. Please refer to supplemental materials for more quantitative + +Table 2. IoU $(\times 10^{2})$ , $P_{d}$ $(\times 10^{2})$ and $F_{a}(\times 10^{6})$ values of different methods achieved on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets. "CNN Full", "CNN Centroid", and "CNN Coarse" represent CNN-based methods under full supervision, centroid and coarse point supervision. "+" represents CNN-based methods equipped with LESPS. + +
MethodsDescriptionNUAA-SIRST [8]NUDT-SIRST [24]IRSTD-1K [54]Average
IoUPdFaIoUPdFaIoUPdFaIoUPdFa
Top-Hat [40]Filtering7.1479.841012.0020.7278.41166.7010.0675.111432.0012.6477.79870.23
RLCM [15]Local Contrast21.0280.61199.1515.1466.35163.0014.6265.6617.9516.0668.7098.77
TLLCM [16]Local Contrast11.0379.477.277.0662.0146.125.3663.974.937.2265.4521.42
MSPCM [34]Local Contrast12.3883.2717.775.8655.87115.967.3360.2715.247.2361.5355.13
IPI [11]Low Rank25.6785.5511.4717.7674.4941.2327.9281.3716.1823.7880.4722.96
PSTNN [51]Low Rank22.4077.9529.1114.8566.1344.1724.5771.9935.2620.6172.0236.18
MDvsFA [45]CNN Full61.7792.4064.9045.3886.03200.7135.4085.8699.2247.5288.10121.61
ISNet [54]CNN Full72.0494.6842.4671.2796.9396.8460.6194.2861.2867.9795.3066.86
UIU-Net [46]CNN Full69.9095.8251.2075.9196.8318.6161.1192.9326.8768.9795.1932.23
ACM [7]CNN Full64.9290.8712.7657.4291.7539.7357.4991.5843.8659.9491.4032.12
CNN Centroid+49.2389.3540.9542.0991.1138.2441.4488.8960.4644.2589.7846.55
CNN Coarse+47.8188.2140.7540.6481.1149.4540.3792.5964.8142.9487.3051.67
ALCNet [8]CNN Full67.9192.7837.0461.7891.3236.3662.0390.9142.4663.9191.6738.62
CNN Centroid+50.6292.0236.8441.5892.2867.0144.9090.5784.6845.7091.6262.84
CNN Coarse+51.0090.8742.4044.1492.8032.1046.7592.2664.3047.3091.9846.27
DNA-Net [24]CNN Full76.8696.9622.587.4298.3124.562.7393.2721.8175.6796.1822.94
CNN Centroid+61.9592.0218.1757.9994.7126.4552.0988.8816.0957.3491.8720.24
CNN Coarse+61.1393.1611.8758.3793.7628.0149.0587.5415.0756.1891.4918.32
+ +Table 3. Average IoU $(\times 10^{2})$ , $P_{d}$ $(\times 10^{2})$ , $F_{a}(\times 10^{6})$ values on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets of DNA-Net trained with pseudo labels generated by input intensity threshold, LCM-based methods [15, 16, 34] and LESPS under centroid and coarse point supervision. + +
Pseudo LabelCentroidCoarse
IoUPdFaIoUPdFa
Threshold=0.34.9281.7813.185.6783.1211.98
Threshold=0.513.2473.085.3115.5476.034.89
Threshold=0.714.5145.504.2815.2146.883.84
RLCM [15]21.4389.102.6722.5390.563.69
TLLCM [16]21.9590.967.7226.0594.154.27
MSPCM [34]28.8992.623.8429.7993.952.28
LESPS(ours)57.3491.8720.2456.1891.4918.32
+ +and qualitative results. + +Comparison to other pseudo labels. We compare our dynamic updated pseudo labels with fixed pseudo labels generated by input intensity threshold and local contrast-based methods [15, 16, 34]. Specifically, given a GT point label, we only preserve the eight connected regions of detection results that have union pixels with the GT point label. Then, we employ the pseudo labels to retrain DNA-Net [24] from scratch. As shown in Table 3, compared with fixed pseudo labels, dynamic updated pseudo labels by LESPS can achieve the highest IoU values with comparable $P_{d}$ and reasonable $F_{a}$ increase. + +# 5.4. Discussion of Loss Function + +In this subsection, we investigate the loss function of computing negative loss on different background points. Average results of different baseline methods under centroid point supervision are shown in Table 4. Extremely limited handcrafted background points7 leads to many false alarms. Random sample8 introduces more background points and well alleviates class imbalance, resulting in great + +Table 4. Average $IoU(\times 10^{2})$ , $P_{d}(\times 10^{2})$ , $F_{a}(\times 10^{3})$ values on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets of different methods when computing negative loss on $i$ handcrafted $(\mathrm{hand}_i)$ , $j$ randomly sampled $(\mathrm{rand}_j)$ and all background points. + +
AnnotationsACMALCNetDNA-Net
IoUPdFaIoUPdFaIoUPdFa
hand10.5495.7947.060.1695.19262.061.4397.8026.91
hand20.1297.24295.170.1596.62248.053.4198.188.48
hand50.1196.36316.370.0897.29363.253.6898.137.29
rand18.0693.453.568.5792.973.2118.7494.690.58
rand210.7892.722.2210.7891.162.7122.8594.810.42
rand513.3992.661.3511.8793.260.8924.8095.000.34
All (Ours)3.9587.150.024.0888.930.025.1289.180.01
+ +performance improvements. However, the above simple versions introduce huge false alarms (34-1.8K times of all points), which are not practical for real applications, but inspire further thorough investigation in the future. + +# 6. Conclusion + +In this paper, we proposed the first work to achieve weakly-supervised SIRST detection using single-point supervision. In our method, we discovered the mapping degeneration phenomenon and proposed a label evolution framework named label evolution with single point supervision (LESPS) to automatically achieve point-to-mask regression. Through LESPS, networks can be trained to achieve SIRST detection in an end-to-end manner. Extensive experiments and insightful visualizations have fully demonstrated the effectiveness and superiority of our method. In addition, our method can be applied to different networks to achieve over $70\%$ and $95\%$ of its fully supervised performance on pixel-level IoU and object-level $P_{d}$ , respectively. We hope our interesting findings and promising results can inspire researchers to rethink the feasibility of achieving state-of-the-art performance in SIRST detection with much weaker supervision. + +# References + +[1] Peri Akiva, Kristin Dana, Peter Oudemans, and Michael Mars. Finding berries: Segmentation and counting of cranberries using point supervision and shape priors. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 50-51, 2020. +[2] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In European Conference on Computer Vision (ECCV), pages 549-565. Springer, 2016. +[3] CL Philip Chen, Hong Li, Yantao Wei, Tian Xia, and Yuan Yan Tang. A local contrast method for small infrared target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 52(1):574-581, 2013. +[4] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8823-8832, 2021. +[5] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2617-2626, 2022. +[6] Yimian Dai and Yiquan Wu. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTAR), 10(8):3752-3767, 2017. +[7] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. +[8] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), pages 1-12, 2021. +[9] Suyog D Deshpande, Meng Hwa Er, Ronda Venkateswarlu, and Philip Chan. Max-mean and max-median filters for detection of small targets. In Signal and Data Processing of Small Targets, volume 3809, pages 74-83. SPIE, 1999. +[10] Junsong Fan, Zhaoxiang Zhang, and Tieniu Tan. Pointly-supervised panoptic segmentation. ArXiv Preprint ArXiv:2210.13950, 2022. +[11] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE Transactions on Image Processing (TIP), 22(12):4996-5009, 2013. +[12] Yongtao Ge, Qiang Zhou, Xinlong Wang, Chunhua Shen, Zhibin Wang, and Hao Li. Point-teaching: Weakly semi-supervised object detection with point annotations. ArXiv Preprint ArXiv:2206.00274, 2022. +[13] Ross Girshick. Fast r-cnn. In IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. +[14] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In IEEE Conference on + +Computer Vision and Pattern Recognition (CVPR), pages 580-587, 2014. +[15] Jinhui Han, Kun Liang, Bo Zhou, Xinying Zhu, Jie Zhao, and Linlin Zhao. Infrared small target detection utilizing the multiscale relative local contrast measure. IEEE Geoscience and Remote Sensing Letters (GRSL), 15(4):612-616, 2018. +[16] Jinhui Han, Saed Moradi, Iman Faramarzi, Chengyin Liu, Honghui Zhang, and Qian Zhao. A local contrast method for infrared small-target detection utilizing a tri-layer window. IEEE Geoscience and Remote Sensing Letters (GRSL), 17(10):1822-1826, 2019. +[17] Jinhui Han, Saed Moradi, Iman Faramarzi, Honghui Zhang, Qian Zhao, Xiaojian Zhang, and Nan Li. Infrared small target detection based on the weighted strengthened local contrast measure. IEEE Geoscience and Remote Sensing Letters (GRSL), 18(9):1670-1674, 2020. +[18] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems (NeurIPS), 34:15908-15919, 2021. +[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In IEEE International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. +[20] Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Haoliang Zheng, and Wei Zhang. Ristdnet: Robust infrared small target detection network. IEEE Geoscience and Remote Sensing Letters (GRSL), 19:1-5, 2021. +[21] Moran Ju, Jiangning Luo, Guangqi Liu, and Haibo Luo. Istdet: An efficient end-to-end neural network for infrared small target detection. Infrared Physics & Technology, 114:103659, 2021. +[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. +[23] Issam H Laradji, Negar Rostamzadeh, Pedro O Pinheiro, David Vazquez, and Mark Schmidt. Where are the blobs: Counting by localization with point supervision. In European Conference on Computer Vision (ECCV), pages 547-562, 2018. +[24] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing (TIP), 2022. +[25] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, et al. Yolov6: A single-stage object detection framework for industrial applications. ArXiv Preprint ArXiv:2209.02976, 2022. +[26] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. +[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. + +[28] Ting Liu, Jungang Yang, Boyang Li, Chao Xiao, Yang Sun, Yingqian Wang, and Wei An. Non-convex tensor low-rank approximation for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 2021. +[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision (ECCV), pages 21-37. Springer, 2016. +[30] Yongtuo Liu, Dan Xu, Sucheng Ren, Hanjie Wu, Hongmin Cai, and Shengfeng He. Fine-grained domain adaptive crowd counting via point-derived segmentation. ArXiv Preprint ArXiv:2108.02980, 2021. +[31] Kevis-Kokitsi Maninis, Sergi Caelles, Jordi Pont-Tuset, and Luc Van Gool. Deep extreme cut: From extreme points to object segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 616-625, 2018. +[32] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. +[33] Saed Moradi, Payman Moallem, and Mohamad Farzan Sabahi. A false-alarm aware methodology to develop robust and efficient multi-scale infrared small target detection algorithm. Infrared Physics & Technology, 89:387-397, 2018. +[34] Saed Moradi, Payman Moallem, and Mohamad Farzan Sabahi. A false-alarm aware methodology to develop robust and efficient multi-scale infrared small target detection algorithm. Infrared Physics & Technology, 89:387-397, 2018. +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NeurlPS), 32, 2019. +[36] Rui Qian, Yunchao Wei, Honghui Shi, Jiachen Li, Jiaying Liu, and Thomas Huang. Weakly supervised scene parsing with point-based distance metric learning. In AAAI Conference on Artificial Intelligence (AAAI), volume 33, pages 8843-8850, 2019. +[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. +[38] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7263-7271, 2017. +[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. + +[40] Jeanfrancois Rivest and Roger Fortin. Detection of dim targets in digital infrared imagery by morphological image processing. Optical Engineering, 35(7):1886-1893, 1996. +[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015. +[42] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. +[43] Yang Sun, Jungang Yang, and Wei An. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor model. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 59(5):3737-3752, 2020. +[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. +[45] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In IEEE International Conference on Computer Vision (ICCV), pages 8509-8518, 2019. +[46] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: u-net in u-net for infrared small object detection. IEEE Transactions on Image Processing (TIP), 32:364-376, 2022. +[47] Xinyi Ying, Yingqian Wang, Longguang Wang, Weidong Sheng, Li Liu, Zaiping Lin, and Shilin Zhou. Local motion and contrast priors driven deep network for infrared small target super-resolution. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTAR), 15:5480-5495, 2022. +[48] Mohsen Zand, Haleh Damirchi, Andrew Farley, Mahdiyar Molahasani, Michael Greenspan, and Ali Etemad. Multiscale crowd counting and localization by multitask point supervision. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1820-1824. IEEE, 2022. +[49] Ke Zhang, Shuyan Ni, Dashuang Yan, and Aidi Zhang. Review of dim small target detection algorithms in single-frame infrared images. In IEEE Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC), volume 4, pages 2115-2120. IEEE, 2021. +[50] Landan Zhang, Lingbing Peng, Tianfang Zhang, Siying Cao, and Zhenming Peng. Infrared small target detection via nonconvex rank approximation minimization joint 12, 1 norm. Remote Sensing, 10(11):1821, 2018. +[51] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. +[52] Mingjin Zhang, Haichen Bai, Jing Zhang, Rui Zhang, Chaoyue Wang, Jie Guo, and Xinbo Gao. Rkformer: Runge-kutta transformer with random-connection attention for + +infrared small target detection. In 30th ACM International Conference on Multimedia (ACM MM), pages 1730-1738, 2022. +[53] Mingjin Zhang, Ke Yue, Jing Zhang, Yunsong Li, and Xinbo Gao. Exploring feature compensation and cross-level correlation for infrared small target detection. In 30th ACM International Conference on Multimedia (ACM MM), pages 1857-1865, 2022. +[54] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 877-886, 2022. +[55] Shiyin Zhang, Jun Hao Liew, Yunchao Wei, Shikui Wei, and Yao Zhao. Interactive object segmentation with inside-outside guidance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12234-12244, 2020. +[56] Shilong Zhang, Zhuoran Yu, Liyang Liu, Xinjiang Wang, Aojun Zhou, and Kai Chen. Group r-cnn for weakly semi-supervised object detection with points. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9417-9426, 2022. +[57] Mingjing Zhao, Wei Li, Lu Li, Jin Hu, Pengge Ma, and Ran Tao. Single-frame infrared small-target detection: A survey. IEEE Geoscience and Remote Sensing Magazine (GRSM), 2022. +[58] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2016. +[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2020. \ No newline at end of file diff --git a/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/images.zip b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18ff2ee6d03cadf212e4780deed8ac4afb5ea6a0 --- /dev/null +++ b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:283e862486f37a7b716830bd68c1d49d398338c3d7b54cd785923e608a02952a +size 620531 diff --git a/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/layout.json b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ec0907b4baaa9a231532cd068fe4235bc46d84ae --- /dev/null +++ b/2023/Mapping Degeneration Meets Label Evolution_ Learning Infrared Small Target Detection With Single Point Supervision/layout.json @@ -0,0 +1,9862 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 102, + 543, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 102, + 543, + 140 + ], + "spans": [ + { + "bbox": [ + 52, + 102, + 543, + 140 + ], + "type": "text", + "content": "Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "spans": [ + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "content": "Xinyi Ying" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Li Liu" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Yingqian Wang" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Ruojing Li" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Nuo Chen" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Zaiping Lin" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Weidong Sheng" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Shilin Zhou" + }, + { + "bbox": [ + 102, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 189, + 405, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 189, + 405, + 204 + ], + "spans": [ + { + "bbox": [ + 189, + 189, + 405, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 189, + 189, + 405, + 204 + ], + "type": "text", + "content": "National University of Defense Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 203, + 509, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 203, + 509, + 218 + ], + "spans": [ + { + "bbox": [ + 85, + 203, + 509, + 218 + ], + "type": "text", + "content": "{yingxinyi18, wangyingqian16, liruojing, chennuo97, linzaiping, slzhou}@nudt.edu.cn," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 217, + 440, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 440, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 440, + 232 + ], + "type": "text", + "content": "dreamliu2010@gmail.com, shengweidong1111@sohu.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "type": "text", + "content": "Training a convolutional neural network (CNN) to detect infrared small targets in a fully supervised manner has gained remarkable research interests in recent years, but is highly labor expensive since a large number of per-pixel annotations are required. To handle this problem, in this paper, we make the first attempt to achieve infrared small target detection with point-level supervision. Interestingly, during the training phase supervised by point labels, we discover that CNNs first learn to segment a cluster of pixels near the targets, and then gradually converge to predict groundtruth point labels. Motivated by this \"mapping degeneration\" phenomenon, we propose a label evolution framework named label evolution with single point supervision (LESPS) to progressively expand the point label by leveraging the intermediate predictions of CNNs. In this way, the network predictions can finally approximate the updated pseudo labels, and a pixel-level target mask can be obtained to train CNNs in an end-to-end manner. We conduct extensive experiments with insightful visualizations to validate the effectiveness of our method. Experimental results show that CNNs equipped with LESPS can well recover the target masks from corresponding point labels, and can achieve over " + }, + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "type": "text", + "content": " of their fully supervised performance in terms of pixel-level intersection over union (IoU) and object-level probability of detection " + }, + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "type": "inline_equation", + "content": "(P_d)" + }, + { + "bbox": [ + 47, + 284, + 290, + 606 + ], + "type": "text", + "content": ", respectively. Code is available at https://github.com/XinyiYing/LESPS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 639, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 288, + 688 + ], + "type": "text", + "content": "Infrared small target detection has been a longstanding, fundamental yet challenging task in infrared search and tracking systems, and has various important applications in civil and military fields [49,57], including traffic monitoring" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 258, + 536, + 378 + ], + "blocks": [ + { + "bbox": [ + 315, + 258, + 536, + 378 + ], + "lines": [ + { + "bbox": [ + 315, + 258, + 536, + 378 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 536, + 378 + ], + "type": "image", + "image_path": "311a0b588c2647d95cc1bfc3f441ed182145828ef7344030fab3667934b3f1fc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 385, + 547, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 385, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 547, + 431 + ], + "type": "text", + "content": "Figure 1. An illustration of mapping degeneration under point supervision. CNNs always tend to segment a cluster of pixels near the targets with low confidence at the early stage, and then gradually learn to predict GT point labels with high confidence." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 617 + ], + "type": "text", + "content": "[24, 54], maritime rescue [52, 53] and military surveillance [7, 47]. Due to the rapid response and robustness to fast-moving scenes, single-frame infrared small target (SIRST) detection methods have always attracted much more attention, and numerous methods have been proposed. Early methods, including filtering-based [9, 40], local contrast-based [3, 16] and low rank-based [11, 43] methods, require complex handcrafted features with carefully tuned hyper-parameters. Recently, compact deep learning has been introduced in solving the problem of SIRST detection [24, 45, 54]. However, there are only a few attempts, and its potential remains locked, unlike the extensive explorations of deep learning for natural images. This is mainly due to potential reasons, including lack of large-scale, accurately annotated datasets and high stake application scenarios." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "Infrared small targets are usually of very small size, weak, shapeless and textureless, and are easily submerged in diverse complex background clutters. As a result, directly adopting existing popular generic object detectors like RCNN series [13, 14, 19, 39], YOLO series [25, 37, 38] and SSD [29] to SIRST detection cannot produce satisfactory performance. Realizing this, researchers have been focusing on developing deep networks tailored for" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 692, + 273, + 701 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 692, + 273, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 692, + 273, + 701 + ], + "type": "text", + "content": "This work was supported by National Key Research and" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 702, + 230, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 230, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 230, + 712 + ], + "type": "text", + "content": "Development Program of China No. 2021YFB3100800." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "15528" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "infrared small targets by adequately utilizing the domain knowledge. However, most existing deep methods for SIRST detection [8, 24, 54] are fully supervised, which usually requires a large dataset with accurate target mask annotations for training. Clearly, this is costly [5, 26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 289 + ], + "type": "text", + "content": "Therefore, a natural question arises: Can we develop a new framework for SIRST detection with single point supervision? In fact, to substantially reduce the annotation cost for object detection tasks, weakly supervised object detection methods with point supervision [4,5,26,56] have been studied in the field of computer vision. Although these weakly supervised methods achieve promising results, they are not designed for the problem of SIRST detection, and the class-agnostic labels (i.e., only foreground and background) of infrared small targets hinder their applications [42, 58]. Therefore, in this work, we intend to conduct the first study of weakly supervised SIRST detection with single-point supervision." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 479 + ], + "type": "text", + "content": "A key motivation of this work comes from an interesting observation during the training of SIRST detection networks. That is, with single point labels serving as supervision, CNNs always tend to segment a cluster of pixels near the targets with low confidence at the early stage, and then gradually learn to predict groundtruth (GT) point labels with high confidence, as shown in Fig. 1. It reveals the fact that region-to-region mapping is the intermediate result of the final region-to-point mapping1. We attribute this \"mapping degeneration\" phenomenon to the special imaging mechanism of infrared system [24, 54], the local contrast prior of infrared small targets [3, 8], and the easy-to-hard learning property of CNNs [44], in which the first two factors result in extended mapping regions beyond the point labels, and the last factor contributes to the degeneration process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 480, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 612 + ], + "type": "text", + "content": "Based on the aforementioned discussion, in this work, we propose a novel framework for the problem of weakly supervised SIRST detection, dubbed label evolution with single point supervision (LESPS). Specifically, LESPS leverages the intermediate network predictions in the training phase to update the current labels, which serve as supervision until the next label update. Through iterative label update and network training, the network predictions can finally approximate the updated pseudo mask labels, and the network can be simultaneously trained to achieve pixel-level SIRST detection in an end-to-end" + }, + { + "bbox": [ + 46, + 480, + 287, + 612 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 480, + 287, + 612 + ], + "type": "text", + "content": " manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 613, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 287, + 637 + ], + "type": "text", + "content": "Our main contributions are summarized as: (1) We present the first study of weakly supervised SIRST" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "text", + "content": "detection, and introduce LESPS that can significantly reduce the annotation cost. (2) We discover the mapping degeneration phenomenon, and leverage this phenomenon to automatically regress pixel-level pseudo labels from the given point labels via LESPS. (3) Experimental results show that our framework can be applied to different existing SIRST detection networks, and enable them to achieve over " + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "text", + "content": " of its fully supervised performance in terms of pixel-level intersection over union " + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "(IoU)" + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "text", + "content": " and object-level probability of detection " + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "(P_d)" + }, + { + "bbox": [ + 306, + 72, + 545, + 193 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 222, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 547, + 617 + ], + "type": "text", + "content": "SIRST Detection. In the past decades, various methods have been proposed, including early traditional paradigms (e.g., filtering-based methods [9, 40], local contrast-based methods [3, 15-17, 33, 34], low rank-based methods [6, 11, 28, 43, 50, 51]) and recent deep learning paradigms [7, 8, 20, 21, 24, 45, 52-54]. Compared to traditional methods, which require delicately designed models and carefully tuned hyper-parameters, convolutional neural networks (CNNs) can learn the non-linear mapping between input images and GT labels in a data-driven manner, and thus generalize better to real complex scenes. As the pioneering work, Wang et al. [45] first employed a generative adversarial network to achieve a better trade-off between miss detection and false alarm. Recently, more works focus on customized solutions of infrared small target. Specifically, Dai et al. [7] specialized an asymmetric contextual module, and further incorporated local contrast measure [8] to improve the target contrast. Li et al. [24] preserved target information by repetitive feature fusion. Zhang et al. [54] aggregated edge information to achieve shape-aware SIRST detection. Zhang et al. [52, 53] explored cross-level correlation and transformer-based method [18] to predict accurate target mask. Wu et al. [46] customized a UIU-Net framework for multi-level and multi-scale feature aggregation. In conclusion, existing works generally focus on compact architectural designs to pursue superior performance in a fully supervised manner. However, due to the lack of a large number of public datasets [7, 24, 45] with per-pixel annotations, the performance and generalization of CNNs are limited. In addition, per-pixel manual annotations are time-consuming and labor-intensive. Therefore, we focus on achieving good pixel-level SIRST detection with weaker supervision and cheaper annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Weakly Supervised Segmentation with Points. Recently, point-level annotation has raised more attention in dense prediction tasks such as object detection [4, 12, 56], crowd counting [1, 23, 30, 48] and image segmentation [2,5,10,26,31,36,55]. We mainly focus on image segmentation in this paper. Specifically, Bearman et al. [2] made the first attempt to introduce an objectiveness potential into a pointly supervised training" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 646, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 646, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 287, + 715 + ], + "type": "text", + "content": "1 \"region-to-region mapping\" represents the mapping learned by CNNs from target regions in images to a cluster of pixels near the targets, while \"region-to-point mapping\" represents the mapping from target regions in images to the GT point labels. 2 Different from generic object detection [32, 59], \"end-to-end\" here represents achieving point-to-mask label regression and direct pixel-level inference in once training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15529" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "loss function to boost segmentation performance. Qian et al. [36] leveraged semantic information of several labeled points by a distance metric loss to achieve scene parsing. Zhang et al. [55] proposed an inside-outside guidance approach to achieve instance segmentation by five elaborate clicks. Cheng et al. [5] designed to provide ten randomly sampled binary point annotations within box annotations for instance segmentation. Li et al. [26] encoded each instance with kernel generator for panoptic segmentation to achieve " + }, + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "inline_equation", + "content": "82\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": " of fully-supervised performance with only twenty randomly annotated points. In contrast to these approaches employing complicated prior constraints to segment large generic objects with rich color and fine textures by several elaborate points, we fully exploit the local contrast prior of infrared small target to progressively evolve pseudo masks by single coarse point without any auxiliaries in an end-to-end manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 286, + 273, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 273, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 273, + 301 + ], + "type": "text", + "content": "3. The Mapping Degeneration Phenomenon" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 288, + 378 + ], + "type": "text", + "content": "In this section, we first describe the mapping degeneration phenomenon together with our intuitive explanation. Then we conduct experiments under single-sample and many-sample training schemes to demonstrate the generality of degeneration, and investigate the influence of generalization on degeneration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 379, + 289, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 289, + 666 + ], + "type": "text", + "content": "As shown in Fig. 1, given an input image and the corresponding GT point label, we employ U-Net [41] as the baseline SIRST detection network for training. It can be observed that, in the early training phase, network predicts a cluster of pixels near the targets with low confidence. As training continues, the network prediction finally approximates GT point label with gradually increased confidence. We name this phenomenon as \"mapping degeneration\", and attribute the following reasons to this phenomenon. 1) Special imaging mechanism of infrared systems [24, 54]: Targets only have intensity information without structure and texture details, resulting in highly similar pixels within the target region. 2) High local contrast of infrared small targets [3, 8]: Pixels within the target region are much brighter or darker with high contrast against the local background clutter. 3) Easy-to-hard learning property of CNNs [44]: CNNs always tend to learn simple mappings first, and then converge to difficult ones. Compared with region-to-point mapping, region-to-region mapping is easier, and thus tends to be the intermediate result of region-to-point mapping. In conclusion, the unique characteristics of infrared small targets result in extended mapping regions beyond point labels, and CNNs contribute to the mapping degeneration process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": "It is worth noting that the mapping degeneration phenomenon is a general phenomenon in various scenes with infrared small targets. Specifically, we use the training datasets (including 1676 images and their corresponding" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 430, + 132 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 430, + 132 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 430, + 132 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 430, + 132 + ], + "type": "image", + "image_path": "ef05e010f1f4414bbbbc0422c6fa759a2ec3b4e80ecd716c80cf32295d37aade.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 430, + 72, + 547, + 132 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 547, + 132 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 547, + 132 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 547, + 132 + ], + "type": "image", + "image_path": "b9bac8a255a0d533de981dffd67bf15df661f747fe96041a0b7009c724a6be27.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 133, + 547, + 217 + ], + "blocks": [ + { + "bbox": [ + 313, + 133, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 547, + 217 + ], + "type": "image", + "image_path": "173cc11f5273b1d88187099bd6fd3dc3167e04e60b4de804b9bc261e789566db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "type": "text", + "content": "Figure 2. Quantitative and qualitative illustrations of mapping degeneration in CNNs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "text", + "content": "centroid point label, see details in Section 5.1) to train U-Net under a single-sample training scheme (i.e., training one CNN on each image). For quantitative analyses, we employ the IoU results between positive pixels in predictions (i.e., pixels with confidence higher than half of its maximum value) and GT mask label. Average IoU results of 1676 CNNs at each epoch are shown by the blue curve in Fig. 2(a), while the number of training samples with maximum IoU during training phase falling in a threshold range of " + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "inline_equation", + "content": "[i, i + 0.1]" + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "inline_equation", + "content": "(i = 0, 0.1, \\dots, 0.9)" + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "text", + "content": " is illustrated via blue bars in Fig. 2(b). It can be observed from the zoom-in curve and bars that mapping degeneration is a general phenomenon with point supervision, and U-Net can achieve IoU " + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "text", + "content": " on more than " + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 253, + 547, + 422 + ], + "type": "text", + "content": " of the training images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "type": "text", + "content": "In addition, we conduct experiments to train U-Net under a many-sample training scheme (i.e., training one CNN using all images which contain abundant targets with various sizes and shapes) to investigate the effect of generalization on mapping degeneration. Average IoU results of 1676 images are shown by orange curve in Fig. 2(a). It can be observed that many-sample training scheme needs more time to converge. Moreover, Fig. 2(b) shows that orange bars are slightly lower than blue ones on larger IoU values (i.e., 0.5-1.0). It is demonstrated that generalization decelerates but aggravates mapping degeneration. Figure 2(c) shows some zoom-in target regions of images and their predictions under these two training schemes. It can be observed that CNNs can effectively segment a cluster of target pixels under both training schemes in a size-aware manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Therefore, an intuitive assumption arises: Can we leverage the intermediate results of CNNs to regress masks? A simple early stopping strategy seems to be a positive answer but is indeed unpractical since mapping degeneration is influenced by various factors, including target intensity, size, shape, and local background clutter (see details in Section 5.2.1). Consequently, there is no fixed optimal stopping epoch for all situations. These" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15530" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 60, + 291, + 191 + ], + "blocks": [ + { + "bbox": [ + 50, + 60, + 291, + 191 + ], + "lines": [ + { + "bbox": [ + 50, + 60, + 291, + 191 + ], + "spans": [ + { + "bbox": [ + 50, + 60, + 291, + 191 + ], + "type": "image", + "image_path": "44a112e497ff17aee764208a5f492eb679dc43de3430601165fbda42b8a02291.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 195, + 288, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 288, + 240 + ], + "type": "text", + "content": "Figure 3. An illustration of label evolution with single point supervision (LESPS). During training, intermediate predictions of CNNs are used to progressively expand point labels to mask labels. Black arrows represent each round of label updates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 250, + 288, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 288, + 286 + ], + "type": "text", + "content": "observations motivate us to design a special label evolution framework to well leverage the mapping degeneration for pseudo mask regression." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 297, + 231, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 231, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 231, + 310 + ], + "type": "text", + "content": "4. The Label Evolution Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "type": "text", + "content": "Motivated by mapping degeneration, we propose a label evolution framework named label evolution with single point supervision (LESPS) to leverage the intermediate network predictions in the training phase to update labels. As training continues, the network predictions approximate the updated pseudo mask labels, and network can simultaneously learn to achieve pixel-level SIRST detection in an end-to-end manner. Here, we employ a toy example of 1D curves for easy understanding. As shown in Fig. 3, sub-figures on the left of the dotted line represent the network predictions. Note that, the black curves denote the intermediate predictions within LESPS, while the gray curves represent virtual results produced by the network without label update. On the right of the dotted line, the first and second columns of sub-figures represent current labels and updated labels, respectively, and black arrows represent each round of label update. The overall framework can be summarized as follows. With point label serving as supervision, in the " + }, + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "type": "text", + "content": " round label update after initial training, the predictions are used to update the current point label to generate the " + }, + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "type": "text", + "content": " updated label, which is then used to supervise the network training until the " + }, + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 46, + 318, + 287, + 653 + ], + "type": "text", + "content": " round label update. Through iterative label updates and network training, CNNs can incorporate the local contrast prior to gradually recover the mask labels. From another viewpoint, label evolution consistently updates the supervision to prevent mapping degeneration, and promotes CNNs to converge to the easy region-to-region mapping." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "Taking the " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " update as an example, given the current label " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "L_{n}" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " and the network prediction " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "P_{n}" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": ", we perform label update for each target, which consists of three steps: candidate pixel extraction, false alarm elimination, and weighted summation between candidate pixels and current" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 60, + 542, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 60, + 542, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 60, + 542, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 60, + 542, + 148 + ], + "type": "image", + "image_path": "7c0aa4c6310d39eb21011d872084a1cbb9aab640f0216f50370c0a6f052a8356.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "lines": [ + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "text", + "content": "Figure 4. (a) Adaptive threshold " + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "inline_equation", + "content": "T_{\\text{adapt}}" + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "text", + "content": " with respect to positive pixels " + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "inline_equation", + "content": "\\hat{L}_n^i" + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "text", + "content": " and hyper-parameters " + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "inline_equation", + "content": "k, T_b" + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "text", + "content": ". Pink dotted line represents the constant " + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "inline_equation", + "content": "hwr" + }, + { + "bbox": [ + 304, + 154, + 547, + 209 + ], + "type": "text", + "content": ". (b) An illustration of false alarm elimination. Red circle and dot represent positive pixels and centroid point of label. Orange circle represents false alarms." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "text", + "content": "labels. Specifically, the " + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "inline_equation", + "content": "d \\times d" + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "text", + "content": " local neighborhoods of the " + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "text", + "content": " target in label " + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "inline_equation", + "content": "L_{n}" + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "text", + "content": " and prediction " + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "inline_equation", + "content": "P_{n}" + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "text", + "content": " are cropped based on the centroid of the positive pixels" + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "text", + "content": " in label (i.e., " + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "inline_equation", + "content": "\\hat{L}_n" + }, + { + "bbox": [ + 304, + 218, + 546, + 314 + ], + "type": "text", + "content": "). Then to reduce error accumulation for label update (see Section 5.2.2 for details), we employ an adaptive threshold (the red dotted line in Fig. 3) to extract the local neighborhood candidate pixels (predictions higher than the red dotted line in Fig. 3). The process can be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 367, + 320, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 320, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 367, + 320, + 545, + 335 + ], + "type": "interline_equation", + "content": "C _ {n} ^ {i} = P _ {n} ^ {i} \\odot \\left(P _ {n} ^ {i} > T _ {\\text {a d a p t}}\\right), \\tag {1}", + "image_path": "0a81858bac1f2e118ec86846848ccb33982c71724ba69d986401dea49e7c9cbd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "inline_equation", + "content": "C_n^i" + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "text", + "content": " is the candidate pixels, and " + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "text", + "content": " represents elementwise multiplication. " + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "inline_equation", + "content": "T_{adapt}" + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "text", + "content": " is the adaptive threshold that correlated to the current prediction " + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "inline_equation", + "content": "P_n^i" + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "text", + "content": " and the positive pixels in label " + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "inline_equation", + "content": "\\hat{L}_n^i" + }, + { + "bbox": [ + 304, + 341, + 546, + 390 + ], + "type": "text", + "content": ", and can be calculated according to:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 396, + 545, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 396, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 320, + 396, + 545, + 410 + ], + "type": "interline_equation", + "content": "T _ {\\text {a d a p t}} = \\max \\left(P _ {n} ^ {i}\\right) \\left(T _ {b} + k \\left(1 - T _ {b}\\right) \\hat {L} _ {n} ^ {i} / (h w r)\\right), \\tag {2}", + "image_path": "176672056bf64672cb5635842166c63550ad2f6bd583f8745b370c9f4d11d858.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "content": " are the height and width of input images, and " + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "inline_equation", + "content": "0.15\\%" + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "content": " [7, 24]. As shown in Fig. 4 (a), " + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "inline_equation", + "content": "T_{b}" + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "content": " is the minimum threshold, and " + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "content": " controls the threshold growth rate. An increasing number of " + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "inline_equation", + "content": "\\hat{L}_n^i" + }, + { + "bbox": [ + 304, + 417, + 546, + 488 + ], + "type": "text", + "content": " leads to the increase of the threshold, which can reduce error accumulation of low contrast targets and strong background clutter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "text", + "content": "To eliminate false alarms by local neighborhood noise, we exclude the eight connective regions of candidate pixels that have no intersection with positive pixels of labels, as shown in Fig. 4 (b). This process is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 391, + 543, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 543, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 391, + 543, + 545, + 557 + ], + "type": "interline_equation", + "content": "E _ {n} ^ {i} = C _ {n} ^ {i} \\odot F _ {n} ^ {i}, \\tag {3}", + "image_path": "a66004666e8751d87bd6ddfe34860b85dcae2f2ba4b8c1ddd2e611e4dff67bce.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "inline_equation", + "content": "E_{n}^{i}" + }, + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "text", + "content": " is the candidate pixels after false alarm elimination, and " + }, + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "inline_equation", + "content": "F_{n}^{i}" + }, + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "text", + "content": " is the mask against false alarm pixels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "type": "text", + "content": "We then perform average weighted summation between candidate pixels " + }, + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "type": "inline_equation", + "content": "E_{n}^{i}" + }, + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "type": "text", + "content": " and current label " + }, + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "type": "inline_equation", + "content": "L_{n}^{i}" + }, + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "type": "text", + "content": " to achieve label update. The process can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 630, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 630, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 334, + 630, + 545, + 654 + ], + "type": "interline_equation", + "content": "L _ {n + 1} ^ {i} = L _ {n} ^ {i} \\odot (1 - N _ {n} ^ {i}) + \\frac {L _ {n} ^ {i} + E _ {n} ^ {i}}{2} \\odot N _ {n} ^ {i}, \\tag {4}", + "image_path": "5cc943ce1020ecf86e253d78c0895b77695f447b0cf8e86a2703bc0da4e06962.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "type": "inline_equation", + "content": "L_{n + 1}^{i}" + }, + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "type": "text", + "content": " is the updated label in the " + }, + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "type": "text", + "content": " round, which serves as new supervision for training in the " + }, + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "type": "inline_equation", + "content": "n + 1^{th}" + }, + { + "bbox": [ + 304, + 661, + 546, + 687 + ], + "type": "text", + "content": " round," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "3The value of a pixel is higher than 0.5, which represents that the pixel is more likely to be positive than negative [8, 24]" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15531" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 63, + 175, + 171 + ], + "blocks": [ + { + "bbox": [ + 50, + 63, + 175, + 171 + ], + "lines": [ + { + "bbox": [ + 50, + 63, + 175, + 171 + ], + "spans": [ + { + "bbox": [ + 50, + 63, + 175, + 171 + ], + "type": "image", + "image_path": "d9c35bfb351b65feaf705240ff75f4627cc300197d6c3e8cbce41fd3e4d21f12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 171, + 138, + 177 + ], + "lines": [ + { + "bbox": [ + 94, + 171, + 138, + 177 + ], + "spans": [ + { + "bbox": [ + 94, + 171, + 138, + 177 + ], + "type": "text", + "content": "(a) Target Intensity" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 182, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 547, + 215 + ], + "type": "text", + "content": "Figure 5. IoU and visualize results of mapping degeneration with respect to different characteristics of targets (i.e.,(a) intensity, (b) size,) and point labels (i.e.,(c) locations and (d) numbers). We visualize the zoom-in target regions of input images with GT point labels (i.e., red dots in images) and corresponding CNN predictions (in the epoch reaching maximum IoU)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 64, + 298, + 171 + ], + "blocks": [ + { + "bbox": [ + 175, + 64, + 298, + 171 + ], + "lines": [ + { + "bbox": [ + 175, + 64, + 298, + 171 + ], + "spans": [ + { + "bbox": [ + 175, + 64, + 298, + 171 + ], + "type": "image", + "image_path": "cdbedc4ef76e60160481931442ab8be8f0ece18250064ce3f1a3faa818929a83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 171, + 258, + 177 + ], + "lines": [ + { + "bbox": [ + 225, + 171, + 258, + 177 + ], + "spans": [ + { + "bbox": [ + 225, + 171, + 258, + 177 + ], + "type": "text", + "content": "(b) Target Size" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 64, + 421, + 171 + ], + "blocks": [ + { + "bbox": [ + 299, + 64, + 421, + 171 + ], + "lines": [ + { + "bbox": [ + 299, + 64, + 421, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 64, + 421, + 171 + ], + "type": "image", + "image_path": "31311459c1a3ebab9c120d00bb84e4a812848b1ae9b14b36dbb6e963fa4fb605.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 171, + 395, + 177 + ], + "lines": [ + { + "bbox": [ + 333, + 171, + 395, + 177 + ], + "spans": [ + { + "bbox": [ + 333, + 171, + 395, + 177 + ], + "type": "text", + "content": "(c) Locations of Point label" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 422, + 64, + 547, + 171 + ], + "blocks": [ + { + "bbox": [ + 422, + 64, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 422, + 64, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 422, + 64, + 547, + 171 + ], + "type": "image", + "image_path": "c4c711bfccfd9b463477f5a3b3a8aa0a832f03359cac6066dd1426bc44860339.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 171, + 521, + 177 + ], + "lines": [ + { + "bbox": [ + 459, + 171, + 521, + 177 + ], + "spans": [ + { + "bbox": [ + 459, + 171, + 521, + 177 + ], + "type": "text", + "content": "(d) Numbers of Point Label" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 227, + 286, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 286, + 275 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 227, + 286, + 275 + ], + "type": "inline_equation", + "content": "N_{n}^{i} = \\left(P_{n}^{i} > T_{\\text{adapt}}\\right) \\odot F_{n}^{i}" + }, + { + "bbox": [ + 46, + 227, + 286, + 275 + ], + "type": "text", + "content": ". Note that, the first term represents GT labels below red dotted lines, and the second term represents the average weighted summation between predictions and GT labels above red dotted lines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 277, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 286, + 384 + ], + "type": "text", + "content": "It is worth noting that we provide three conditions to ensure network convergence: 1) Average weighted summation between predictions and labels promotes CNNs to converge as predictions approximate labels. 2) Pixel-adaptive threshold increases with the increase of positive pixels in updated labels, which slows down or suspends the label update. 3) As label evolution introduces more target information for training, CNNs grow to mature, and learn to distinguish targets from backgrounds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "type": "text", + "content": "We start label evolution after initial evolution epoch " + }, + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "type": "inline_equation", + "content": "T_{epoch}" + }, + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "type": "text", + "content": ", and perform label update every " + }, + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "type": "text", + "content": " epoch until the end of training. Note that, our epoch-based threshold " + }, + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "type": "inline_equation", + "content": "T_{epoch}" + }, + { + "bbox": [ + 47, + 385, + 286, + 445 + ], + "type": "text", + "content": " is a coarse threshold to ensure that networks attend to targets instead of background clutter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 459, + 127, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 127, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 127, + 472 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "text", + "content": "In this section, we first describe the implementation details, and then make comprehensive analyses of the mapping degeneration phenomenon and our label evolution framework. In addition, we apply our method to the state-of-the-art SIRST detection methods with point supervision, and make comparisons with their fully supervised counterparts. Moreover, we make comparisons of our dynamic updated pseudo labels with fixed pseudo labels, and discuss the calculation of loss function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 598, + 179, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 179, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 179, + 610 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Three public datasets NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] are used in our experiments. We followed [24] to split the training and test sets of NUAA-SIRST and NUDT-SIRST, and followed [54] to split IRSTD-1K. We employed two pixel-level metrics (i.e., intersection over union " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "(IoU)" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " and pixel accuracy " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "(PA))" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " and two target-level metrics (i.e., probability of detection " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "(P_d)" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " and false-alarm rate " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "(F_a))" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " for performance evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "text", + "content": "During training, all images were normalized and randomly cropped into patches of size " + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "text", + "content": " as network inputs. We augmented the training data by random flipping and rotation. Due to the extreme positive-negative sample imbalance (less than 10 vs. more than " + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "text", + "content": ") in SIRST detection with point supervision, we employed focal loss[27] to stabilize the training process. All the networks were optimized by the Adam method [22]. Batch size was set to 16, and learning rate was initially set to " + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "text", + "content": " and reduced by ten times at the " + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "inline_equation", + "content": "200^{th}" + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "inline_equation", + "content": "300^{th}" + }, + { + "bbox": [ + 304, + 227, + 545, + 382 + ], + "type": "text", + "content": " epochs. We stopped training after 400 epochs. All models were implemented in PyTorch [35] on a PC with an Nvidia GeForce 3090 GPU." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 390, + 402, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 402, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 402, + 403 + ], + "type": "text", + "content": "5.2. Model Analyses" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 409, + 488, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 409, + 488, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 409, + 488, + 422 + ], + "type": "text", + "content": "5.2.1 Analyses of Mapping Degeneration" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 427, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 594 + ], + "type": "text", + "content": "We use synthetic images (simulated targets and real backgrounds [24]) to investigate the mapping degeneration phenomenon with respect to different characteristics of targets (i.e., intensity and size" + }, + { + "bbox": [ + 304, + 427, + 545, + 594 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 304, + 427, + 545, + 594 + ], + "type": "text", + "content": ") and point labels (i.e., locations and numbers). We employ U-Net [41] as the baseline detection network, and use centroid point as GT label if not specified. We calculate IoU between positive pixels in predictions and GT mask labels of each epoch for quantitative analyses. In addition, we visualize the zoom-in target regions of simulated images with GT point labels (i.e., red dots) and corresponding CNN predictions (in the epoch reaching maximum IoU). To reduce training randomness, we show the average IoU results and visualization results over 100 runs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 666 + ], + "type": "text", + "content": "Target Intensity. We simulate Gaussian-based extended targets with different peak values (i.e., 10, 50, 100, 200, 500) to investigate the influence of target intensity on mapping degeneration. Quantitative results in Fig. 5(a) show that intensity higher than 100 leads to a positive correlation between intensity and maximum IoU, while" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 545, + 693 + ], + "type": "text", + "content": "4Focal loss is calculated between current evolved and GT labels to supervise the network training until the next round label update." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "Shape, and local background clutter are investigated in supplemental material." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15532" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 62, + 173, + 138 + ], + "blocks": [ + { + "bbox": [ + 49, + 62, + 173, + 138 + ], + "lines": [ + { + "bbox": [ + 49, + 62, + 173, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 62, + 173, + 138 + ], + "type": "image", + "image_path": "5eb51815d988cb84a58c1be6c49ca872022a045e720610234f44fe32c55072ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 140, + 161, + 149 + ], + "lines": [ + { + "bbox": [ + 71, + 140, + 161, + 149 + ], + "spans": [ + { + "bbox": [ + 71, + 140, + 161, + 149 + ], + "type": "text", + "content": "(a) Initial Evolution Epoch " + }, + { + "bbox": [ + 71, + 140, + 161, + 149 + ], + "type": "inline_equation", + "content": "T_{\\text{epoch}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 62, + 304, + 138 + ], + "blocks": [ + { + "bbox": [ + 173, + 62, + 304, + 138 + ], + "lines": [ + { + "bbox": [ + 173, + 62, + 304, + 138 + ], + "spans": [ + { + "bbox": [ + 173, + 62, + 304, + 138 + ], + "type": "image", + "image_path": "31839d592377a44cbb7a23c802f75c69f7a776e70bfaed689d57f2f1b9604a14.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 140, + 280, + 149 + ], + "lines": [ + { + "bbox": [ + 194, + 140, + 280, + 149 + ], + "spans": [ + { + "bbox": [ + 194, + 140, + 280, + 149 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 194, + 140, + 280, + 149 + ], + "type": "inline_equation", + "content": "T_{b}" + }, + { + "bbox": [ + 194, + 140, + 280, + 149 + ], + "type": "text", + "content": " of Evolution Threshold" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 62, + 419, + 140 + ], + "blocks": [ + { + "bbox": [ + 306, + 62, + 419, + 140 + ], + "lines": [ + { + "bbox": [ + 306, + 62, + 419, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 62, + 419, + 140 + ], + "type": "image", + "image_path": "34a1a6f1452c80a2b434298440d647e79534e537f3a95da49d157f26c31d7a49.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 140, + 407, + 148 + ], + "lines": [ + { + "bbox": [ + 323, + 140, + 407, + 148 + ], + "spans": [ + { + "bbox": [ + 323, + 140, + 407, + 148 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 323, + 140, + 407, + 148 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 323, + 140, + 407, + 148 + ], + "type": "text", + "content": " of Evolution Threshold" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 420, + 63, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 420, + 63, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 420, + 63, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 420, + 63, + 545, + 140 + ], + "type": "image", + "image_path": "34a9f9aefc047fbfce0285991b847d0b04da3886369f943d3e44408f3a5c121b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 140, + 520, + 148 + ], + "lines": [ + { + "bbox": [ + 446, + 140, + 520, + 148 + ], + "spans": [ + { + "bbox": [ + 446, + 140, + 520, + 148 + ], + "type": "text", + "content": "(d) Evolution Frequency " + }, + { + "bbox": [ + 446, + 140, + 520, + 148 + ], + "type": "inline_equation", + "content": "f" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 49, + 241, + 283, + 276 + ], + "blocks": [ + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "text", + "content": "Table 1. Average " + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "inline_equation", + "content": "IoU" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "inline_equation", + "content": "\\times 10^{2}" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "inline_equation", + "content": "\\times 10^{2}" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "inline_equation", + "content": "\\times 10^{6}" + }, + { + "bbox": [ + 46, + 190, + 286, + 234 + ], + "type": "text", + "content": ") values on NUAA-SIRST [8] NUDT-SIRST [24] and IRSTD-1K [54] achieved by DNA-Net with (w/) and without (w/o) LESPS under centroid, coarse point supervision together with full supervision." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 241, + 283, + 276 + ], + "lines": [ + { + "bbox": [ + 49, + 241, + 283, + 276 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 283, + 276 + ], + "type": "table", + "html": "
MethodCentroidCoarseFull
IoUPdFaIoUPdFaIoUPdFa
w/o LESPS5.1289.190.682.9649.890.3075.6796.1822.94
w/ LESPS57.3491.8720.2456.1891.4918.32
", + "image_path": "806778cd0a16a4e3340a307a8bf9476239afd921fb2062c5818d5972b412bdb8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 290, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 286, + 361 + ], + "type": "text", + "content": "lower intensity leads to a negative one. In addition, curve \"intensity 10\" reaches maximum " + }, + { + "bbox": [ + 46, + 290, + 286, + 361 + ], + "type": "inline_equation", + "content": "IoU" + }, + { + "bbox": [ + 46, + 290, + 286, + 361 + ], + "type": "text", + "content": " at around epoch 150 while others are less than 50, which demonstrates that over-small intensity decelerates degeneration. Visualization results show that our method can well highlight target regions under various intensities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 363, + 286, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 286, + 495 + ], + "type": "text", + "content": "Target Size. We simulate Gaussian-based extended targets with different radii (i.e., 3, 5, 7, 9, 13) to investigate the influence of target size on mapping degeneration. Quantitative results in Fig. 5(b) show that larger target size leads to lower maximum IoU and less time to reach maximum IoU. That is because, size discrepancy between targets and GT point labels increases as target size increases, which aggravates and accelerates mapping degeneration. Visualization results show that CNNs can predict a cluster of pixels in a size-aware manner, and the peak values of predictions decrease as target size increases." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 497, + 286, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 286, + 663 + ], + "type": "text", + "content": "Locations of Point Label. We simulate a Gaussian-based extended target (with intensity 500 & radius 13), and place point labels at different distances away from the target centroid to investigate the influence of point label locations on mapping degeneration. Results in Fig. 5(c) show that small perturbations of label locations (less than 3 pixels) have a minor influence on the maximum IoU results. However, severe location perturbations (larger than 3 pixels) can lead to a significant maximum IoU drop, and the drop is more obvious when the point label is close to the edge. Note that, the same targets with different label locations reach maximum IoU at the same time, which demonstrates that the speed of mapping degeneration is irrelevant to the position of labels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Number of Points in Label. We simulate a Gaussian-based extended target (with intensity 500 & radius 13) and set different numbers of points in labels to investigate its influence on mapping degeneration. Quantitative results" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 185, + 550, + 256 + ], + "blocks": [ + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "text", + "content": "Figure 6. " + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "inline_equation", + "content": "PA" + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "text", + "content": " (P) and " + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "inline_equation", + "content": "IoU" + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "text", + "content": " (I) results of LESPS with respect to (a) initial evolution epoch " + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "inline_equation", + "content": "T_{epoch}" + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "text", + "content": ", (b) " + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "inline_equation", + "content": "T_b" + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "text", + "content": " and (c) " + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "text", + "content": " of evolution threshold, and (d) evolution frequency " + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 156, + 544, + 179 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 185, + 550, + 256 + ], + "lines": [ + { + "bbox": [ + 307, + 185, + 550, + 256 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 550, + 256 + ], + "type": "image", + "image_path": "111d120cbbbbf643f11514976d8c732ea735d336081aa437ccbcc119ea041dc9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 259, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 259, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 259, + 545, + 281 + ], + "type": "text", + "content": "Figure 7. Quantitative and qualitative results of evolved target masks." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 292, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 388 + ], + "type": "text", + "content": "in Fig. 5(d) show that as the number of points increases, CNNs can learn more target information to achieve higher maximum IoU results. In addition, the speed of mapping degeneration is irrelevant to the point number. Visualization results show that peak values of predictions increase as the number of points increases, which demonstrates that stronger supervision alleviates mapping degeneration. The conclusion well supports our label evolution framework." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 403, + 526, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 526, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 526, + 415 + ], + "type": "text", + "content": "5.2.2 Analyses of the Label Evolution Framework" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 422, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 493 + ], + "type": "text", + "content": "In this subsection, we conduct experiments to investigate the effectiveness and the optimal parameter settings of our label evolution framework (i.e., LESPS). We employ " + }, + { + "bbox": [ + 304, + 422, + 545, + 493 + ], + "type": "inline_equation", + "content": "PA" + }, + { + "bbox": [ + 304, + 422, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 545, + 493 + ], + "type": "inline_equation", + "content": "IoU" + }, + { + "bbox": [ + 304, + 422, + 545, + 493 + ], + "type": "text", + "content": " between the positive pixels in updated labels and the GT mask labels to quantitatively evaluate the accuracy and expansion degree of the current label." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 494, + 546, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 546, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 546, + 685 + ], + "type": "text", + "content": "Effectiveness. We compare the average results of NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] datasets achieved by DNA-Net with (i.e., w/) and without (i.e., w/o) LESPS under centroid and coarse point supervision, respectively. Note that, the results of DNA-Net w/o LESPS are calculated under extremely low threshold" + }, + { + "bbox": [ + 304, + 494, + 546, + 685 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 304, + 494, + 546, + 685 + ], + "type": "text", + "content": " (i.e., 0.15) while those of DNA-Net w/ LESPS are calculated under the standard threshold (i.e., 0.5 [24, 54]). As shown in Table 1, as compared to full supervision, the results of DNA-Net w/o LESPS are extremely low, which demonstrates that SIRST detection on single point supervision is a challenging task. In contrast, DNA-Net w/ LESPS can achieve significant performance improvement under both point supervisions in terms of all the metrics, which approximate the performance of their fully supervised counterparts. Note that, " + }, + { + "bbox": [ + 304, + 494, + 546, + 685 + ], + "type": "inline_equation", + "content": "P_d" + }, + { + "bbox": [ + 304, + 494, + 546, + 685 + ], + "type": "text", + "content": " results of DNA" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "6With point supervision, the results of DNA-Net w/o LESPS calculated under threshold 0.5 are all zeros." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15533" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 63, + 291, + 165 + ], + "blocks": [ + { + "bbox": [ + 49, + 63, + 291, + 165 + ], + "lines": [ + { + "bbox": [ + 49, + 63, + 291, + 165 + ], + "spans": [ + { + "bbox": [ + 49, + 63, + 291, + 165 + ], + "type": "image", + "image_path": "b03b94a70ab9a87594864d09604fc6d33d9914caaff4c659037cb559e3d013ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 205 + ], + "type": "text", + "content": "Figure 8. Visualizations of regressed labels during training and network predictions during inference with centroid and coarse point supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 287, + 277 + ], + "type": "text", + "content": "Net w/o LESPS under coarse point supervision are over half lower than those under the centroid ones, while the results of DNA-Net w/ LESPS under these two kinds of point supervision are comparable. It demonstrates that LESPS can generalize well to manual annotation errors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 279, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 506 + ], + "type": "text", + "content": "In addition, we make evaluations of evolved target masks during training. Quantitative results in Fig. 7(a) show average IoU values between positive pixels of evolved target masks and GT labels in 20-time training, which demonstrates that the quality of pseudo target masks consecutively increases during training. Qualitative results in Fig. 7(b) demonstrate that networks can effectively learn to expand point labels to mask labels. Furthermore, we visualize the labels regressed by our LESPS during training together with some network predictions during inference in Figs. 8 (a) and (b). As shown in Fig. 8(a), compared with GT mask labels, the evolved labels are more closely aligned with the objects in images (e.g., GT masks of Misc_4, XDU113 exceed the target regions due to visual edge ambiguity), which demonstrates that LESPS can alleviate the manually annotated errors. Figure 8(b) shows that DNA-Net w/ LESPS can effectively achieve accurate pixel-level SIRST detection in an end-to-end manner. Please refer to the supplemental materials for more visual results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": "Initial Evolution Epoch. We investigate the optimal value of epoch-based threshold " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "T_{\\text{epoch}}" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": ". Figure 6(a) shows that small initial evolution epoch results in a significant difference between " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "PA" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "IoU" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": " (i.e., 0.94 vs. 0.04 with " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "T_{\\text{epoch}} = 10" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": "). That is because, early label evolution introduces many error pixels, resulting in severe error accumulation and network convergence failure. Increasing initial evolution epoch can reduce error accumulation and promote network convergence (0.60 vs. 0.54 with " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "T_{\\text{epoch}} = 50" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": "). However, over-large initial evolution epoch (i.e., a high degree of mapping degeneration) results in inferior performance (0.21 vs. 0.21 with " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "T_{\\text{epoch}} = 100" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": "). Therefore, " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "T_{\\text{epoch}}" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": " is set to 50 in our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Evolution Threshold. We investigate the optimal values of " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{b}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " in the evolution threshold. " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{b}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " is the minimum threshold, and controls evolution speed and error accumulation degree. " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " determines the maximum" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 61, + 548, + 211 + ], + "blocks": [ + { + "bbox": [ + 307, + 61, + 548, + 211 + ], + "lines": [ + { + "bbox": [ + 307, + 61, + 548, + 211 + ], + "spans": [ + { + "bbox": [ + 307, + 61, + 548, + 211 + ], + "type": "image", + "image_path": "7bbb71eaa0d7dc01f32000feab13282c294b7b925787e3fc799aa8ac20107b0c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 220, + 546, + 265 + ], + "lines": [ + { + "bbox": [ + 304, + 220, + 546, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 546, + 265 + ], + "type": "text", + "content": "Figure 9. Visual detection results of different methods achieved on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets. Correctly detected targets and false alarms are highlighted by red and orange circles, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "text", + "content": "threshold, and controls the growth rate of the threshold. As shown in Fig. 6(b) and 6(c), both over-large and over-small values of " + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "inline_equation", + "content": "T_{b}" + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "text", + "content": " result in inferior performance. Therefore, we choose " + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "inline_equation", + "content": "k = 1 / 2" + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "inline_equation", + "content": "T_{b} = 0.5" + }, + { + "bbox": [ + 304, + 275, + 546, + 323 + ], + "type": "text", + "content": " in our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "text", + "content": "Evolution Frequency. We investigate the optimal value of evolution frequency " + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "text", + "content": ". Figure 6(d) shows that evolution frequency is positively correlated to " + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "inline_equation", + "content": "PA" + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "inline_equation", + "content": "IoU" + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "text", + "content": ". However, high evolution frequency needs more time for label updates. To balance performance and efficiency, we choose " + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "inline_equation", + "content": "f = 5" + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "text", + "content": " in our method. Interestingly, higher frequency (i.e., " + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "inline_equation", + "content": "f = 2" + }, + { + "bbox": [ + 304, + 324, + 546, + 445 + ], + "type": "text", + "content": ") does not cause serve error accumulation, which also demonstrates the effectiveness of the convergence conditions of our LESPS. Please refer to the supplemental materials for more discussions of the convergence issue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 454, + 517, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 517, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 517, + 467 + ], + "type": "text", + "content": "5.3. Comparison to State-of-the-art Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": "Comparison to SISRT detection methods. We apply our LESPS to several state-of-the-art CNN-based methods, including ACM [7], ALCNet [8] and DNA-Net [24]. For fair comparisons, we retrained all models on NUAA-SIRST [8], NUDT-SIRST [24], and IRSTD-1K [54] datasets with the same settings. In addition, we add the results of six fully supervised CNN-based methods (MDvsFA [45], ACM [7], ALCNet [8], DNA-Net [24], ISNet [54], UIU-Net [46]) and six traditional methods (Top-Hat [40], RLCM [15], TLLCM [16], MSPCM [17], IPI [11], PSTNN [51]) as the baseline results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Quantitative results in Table 2 show that CNN-based methods equipped with LESPS can outperform all the traditional methods. In addition, they can also achieve " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "71 - 75\\%" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " IoU results and comparable " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " results of their fully supervised counterparts. Qualitative results in Fig. 9 show that CNN-based methods equipped with LESPS can produce outputs with precise target mask and low false alarm rate, and can generalize well to complex scenes. Please refer to supplemental materials for more quantitative" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15534" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 101, + 541, + 266 + ], + "blocks": [ + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "lines": [ + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "spans": [ + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "type": "text", + "content": "Table 2. IoU " + }, + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "type": "inline_equation", + "content": "F_{a}(\\times 10^{6})" + }, + { + "bbox": [ + 47, + 64, + 545, + 99 + ], + "type": "text", + "content": " values of different methods achieved on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets. \"CNN Full\", \"CNN Centroid\", and \"CNN Coarse\" represent CNN-based methods under full supervision, centroid and coarse point supervision. \"+\" represents CNN-based methods equipped with LESPS." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 101, + 541, + 266 + ], + "lines": [ + { + "bbox": [ + 50, + 101, + 541, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 101, + 541, + 266 + ], + "type": "table", + "html": "
MethodsDescriptionNUAA-SIRST [8]NUDT-SIRST [24]IRSTD-1K [54]Average
IoUPdFaIoUPdFaIoUPdFaIoUPdFa
Top-Hat [40]Filtering7.1479.841012.0020.7278.41166.7010.0675.111432.0012.6477.79870.23
RLCM [15]Local Contrast21.0280.61199.1515.1466.35163.0014.6265.6617.9516.0668.7098.77
TLLCM [16]Local Contrast11.0379.477.277.0662.0146.125.3663.974.937.2265.4521.42
MSPCM [34]Local Contrast12.3883.2717.775.8655.87115.967.3360.2715.247.2361.5355.13
IPI [11]Low Rank25.6785.5511.4717.7674.4941.2327.9281.3716.1823.7880.4722.96
PSTNN [51]Low Rank22.4077.9529.1114.8566.1344.1724.5771.9935.2620.6172.0236.18
MDvsFA [45]CNN Full61.7792.4064.9045.3886.03200.7135.4085.8699.2247.5288.10121.61
ISNet [54]CNN Full72.0494.6842.4671.2796.9396.8460.6194.2861.2867.9795.3066.86
UIU-Net [46]CNN Full69.9095.8251.2075.9196.8318.6161.1192.9326.8768.9795.1932.23
ACM [7]CNN Full64.9290.8712.7657.4291.7539.7357.4991.5843.8659.9491.4032.12
CNN Centroid+49.2389.3540.9542.0991.1138.2441.4488.8960.4644.2589.7846.55
CNN Coarse+47.8188.2140.7540.6481.1149.4540.3792.5964.8142.9487.3051.67
ALCNet [8]CNN Full67.9192.7837.0461.7891.3236.3662.0390.9142.4663.9191.6738.62
CNN Centroid+50.6292.0236.8441.5892.2867.0144.9090.5784.6845.7091.6262.84
CNN Coarse+51.0090.8742.4044.1492.8032.1046.7592.2664.3047.3091.9846.27
DNA-Net [24]CNN Full76.8696.9622.587.4298.3124.562.7393.2721.8175.6796.1822.94
CNN Centroid+61.9592.0218.1757.9994.7126.4552.0988.8816.0957.3491.8720.24
CNN Coarse+61.1393.1611.8758.3793.7628.0149.0587.5415.0756.1891.4918.32
", + "image_path": "576d89fc78b2634a6a49bc4c33d88f2dd6cc3fbd634f375e7ccfc3f56ae5889e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 337, + 287, + 412 + ], + "blocks": [ + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "text", + "content": "Table 3. Average IoU " + }, + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "inline_equation", + "content": "F_{a}(\\times 10^{6})" + }, + { + "bbox": [ + 46, + 277, + 287, + 334 + ], + "type": "text", + "content": " values on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets of DNA-Net trained with pseudo labels generated by input intensity threshold, LCM-based methods [15, 16, 34] and LESPS under centroid and coarse point supervision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 337, + 287, + 412 + ], + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 412 + ], + "type": "table", + "html": "
Pseudo LabelCentroidCoarse
IoUPdFaIoUPdFa
Threshold=0.34.9281.7813.185.6783.1211.98
Threshold=0.513.2473.085.3115.5476.034.89
Threshold=0.714.5145.504.2815.2146.883.84
RLCM [15]21.4389.102.6722.5390.563.69
TLLCM [16]21.9590.967.7226.0594.154.27
MSPCM [34]28.8992.623.8429.7993.952.28
LESPS(ours)57.3491.8720.2456.1891.4918.32
", + "image_path": "fbd3cda70b190dd016adbaaaefa3088bc01cf43230e6b94936f9036db0374e10.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 426, + 139, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 139, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 139, + 437 + ], + "type": "text", + "content": "and qualitative results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 439, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 570 + ], + "type": "text", + "content": "Comparison to other pseudo labels. We compare our dynamic updated pseudo labels with fixed pseudo labels generated by input intensity threshold and local contrast-based methods [15, 16, 34]. Specifically, given a GT point label, we only preserve the eight connected regions of detection results that have union pixels with the GT point label. Then, we employ the pseudo labels to retrain DNA-Net [24] from scratch. As shown in Table 3, compared with fixed pseudo labels, dynamic updated pseudo labels by LESPS can achieve the highest IoU values with comparable " + }, + { + "bbox": [ + 46, + 439, + 287, + 570 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 46, + 439, + 287, + 570 + ], + "type": "text", + "content": " and reasonable " + }, + { + "bbox": [ + 46, + 439, + 287, + 570 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 46, + 439, + 287, + 570 + ], + "type": "text", + "content": " increase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 579, + 200, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 200, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 200, + 591 + ], + "type": "text", + "content": "5.4. Discussion of Loss Function" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "type": "text", + "content": "In this subsection, we investigate the loss function of computing negative loss on different background points. Average results of different baseline methods under centroid point supervision are shown in Table 4. Extremely limited handcrafted background points7 leads to many false alarms. Random sample8 introduces more background points and well alleviates class imbalance, resulting in great" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 337, + 545, + 412 + ], + "blocks": [ + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "text", + "content": "Table 4. Average " + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "inline_equation", + "content": "IoU(\\times 10^{2})" + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "inline_equation", + "content": "P_{d}(\\times 10^{2})" + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "inline_equation", + "content": "F_{a}(\\times 10^{3})" + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "text", + "content": " values on NUAA-SIRST [8], NUDT-SIRST [24] and IRSTD-1K [54] datasets of different methods when computing negative loss on " + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "text", + "content": " handcrafted " + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "inline_equation", + "content": "(\\mathrm{hand}_i)" + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "text", + "content": " randomly sampled " + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "inline_equation", + "content": "(\\mathrm{rand}_j)" + }, + { + "bbox": [ + 306, + 277, + 545, + 334 + ], + "type": "text", + "content": " and all background points." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 337, + 545, + 412 + ], + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 412 + ], + "type": "table", + "html": "
AnnotationsACMALCNetDNA-Net
IoUPdFaIoUPdFaIoUPdFa
hand10.5495.7947.060.1695.19262.061.4397.8026.91
hand20.1297.24295.170.1596.62248.053.4198.188.48
hand50.1196.36316.370.0897.29363.253.6898.137.29
rand18.0693.453.568.5792.973.2118.7494.690.58
rand210.7892.722.2210.7891.162.7122.8594.810.42
rand513.3992.661.3511.8793.260.8924.8095.000.34
All (Ours)3.9587.150.024.0888.930.025.1289.180.01
", + "image_path": "f635dda01aed254751646c0503647a329582044aa1d533e03792c9523e0aeaa1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 425, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 473 + ], + "type": "text", + "content": "performance improvements. However, the above simple versions introduce huge false alarms (34-1.8K times of all points), which are not practical for real applications, but inspire further thorough investigation in the future." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 488, + 378, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 378, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 378, + 501 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we proposed the first work to achieve weakly-supervised SIRST detection using single-point supervision. In our method, we discovered the mapping degeneration phenomenon and proposed a label evolution framework named label evolution with single point supervision (LESPS) to automatically achieve point-to-mask regression. Through LESPS, networks can be trained to achieve SIRST detection in an end-to-end manner. Extensive experiments and insightful visualizations have fully demonstrated the effectiveness and superiority of our method. In addition, our method can be applied to different networks to achieve over " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " of its fully supervised performance on pixel-level IoU and object-level " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", respectively. We hope our interesting findings and promising results can inspire researchers to rethink the feasibility of achieving state-of-the-art performance in SIRST detection with much weaker supervision." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 693, + 260, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 260, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 260, + 703 + ], + "type": "text", + "content": "7Points are sampled near targets, and are fixed during training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 703, + 279, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 279, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 279, + 712 + ], + "type": "text", + "content": "8Points are randomly sampled, and change in each training iteration." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15535" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Peri Akiva, Kristin Dana, Peter Oudemans, and Michael Mars. Finding berries: Segmentation and counting of cranberries using point supervision and shape priors. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 50-51, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In European Conference on Computer Vision (ECCV), pages 549-565. Springer, 2016." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] CL Philip Chen, Hong Li, Yantao Wei, Tian Xia, and Yuan Yan Tang. A local contrast method for small infrared target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 52(1):574-581, 2013." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[4] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8823-8832, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2617-2626, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "text", + "content": "[6] Yimian Dai and Yiquan Wu. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTAR), 10(8):3752-3767, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 385, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 288, + 430 + ], + "type": "text", + "content": "[7] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 475 + ], + "type": "text", + "content": "[8] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), pages 1-12, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "type": "text", + "content": "[9] Suyog D Deshpande, Meng Hwa Er, Ronda Venkateswarlu, and Philip Chan. Max-mean and max-median filters for detection of small targets. In Signal and Data Processing of Small Targets, volume 3809, pages 74-83. SPIE, 1999." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "text", + "content": "[10] Junsong Fan, Zhaoxiang Zhang, and Tieniu Tan. Pointly-supervised panoptic segmentation. ArXiv Preprint ArXiv:2210.13950, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "text", + "content": "[11] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE Transactions on Image Processing (TIP), 22(12):4996-5009, 2013." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 611, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 655 + ], + "type": "text", + "content": "[12] Yongtao Ge, Qiang Zhou, Xinlong Wang, Chunhua Shen, Zhibin Wang, and Hao Li. Point-teaching: Weakly semi-supervised object detection with point annotations. ArXiv Preprint ArXiv:2206.00274, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 657, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 679 + ], + "type": "text", + "content": "[13] Ross Girshick. Fast r-cnn. In IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In IEEE Conference on" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition (CVPR), pages 580-587, 2014." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[15] Jinhui Han, Kun Liang, Bo Zhou, Xinying Zhu, Jie Zhao, and Linlin Zhao. Infrared small target detection utilizing the multiscale relative local contrast measure. IEEE Geoscience and Remote Sensing Letters (GRSL), 15(4):612-616, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[16] Jinhui Han, Saed Moradi, Iman Faramarzi, Chengyin Liu, Honghui Zhang, and Qian Zhao. A local contrast method for infrared small-target detection utilizing a tri-layer window. IEEE Geoscience and Remote Sensing Letters (GRSL), 17(10):1822-1826, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "text", + "content": "[17] Jinhui Han, Saed Moradi, Iman Faramarzi, Honghui Zhang, Qian Zhao, Xiaojian Zhang, and Nan Li. Infrared small target detection based on the weighted strengthened local contrast measure. IEEE Geoscience and Remote Sensing Letters (GRSL), 18(9):1670-1674, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "text", + "content": "[18] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems (NeurIPS), 34:15908-15919, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "text", + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In IEEE International Conference on Computer Vision (ICCV), pages 2961-2969, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "text", + "content": "[20] Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Haoliang Zheng, and Wei Zhang. Ristdnet: Robust infrared small target detection network. IEEE Geoscience and Remote Sensing Letters (GRSL), 19:1-5, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "text", + "content": "[21] Moran Ju, Jiangning Luo, Guangqi Liu, and Haibo Luo. Istdet: An efficient end-to-end neural network for infrared small target detection. Infrared Physics & Technology, 114:103659, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[23] Issam H Laradji, Negar Rostamzadeh, Pedro O Pinheiro, David Vazquez, and Mark Schmidt. Where are the blobs: Counting by localization with point supervision. In European Conference on Computer Vision (ECCV), pages 547-562, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 513, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 556 + ], + "type": "text", + "content": "[24] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing (TIP), 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "text", + "content": "[25] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, et al. Yolov6: A single-stage object detection framework for industrial applications. ArXiv Preprint ArXiv:2209.02976, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 669 + ], + "type": "text", + "content": "[26] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15536" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Ting Liu, Jungang Yang, Boyang Li, Chao Xiao, Yang Sun, Yingqian Wang, and Wei An. Non-convex tensor low-rank approximation for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "text", + "content": "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision (ECCV), pages 21-37. Springer, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 231 + ], + "type": "text", + "content": "[30] Yongtuo Liu, Dan Xu, Sucheng Ren, Hanjie Wu, Hongmin Cai, and Shengfeng He. Fine-grained domain adaptive crowd counting via point-derived segmentation. ArXiv Preprint ArXiv:2108.02980, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 233, + 287, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 287 + ], + "type": "text", + "content": "[31] Kevis-Kokitsi Maninis, Sergi Caelles, Jordi Pont-Tuset, and Luc Van Gool. Deep extreme cut: From extreme points to object segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 616-625, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 291, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 345 + ], + "type": "text", + "content": "[32] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 348, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 287, + 401 + ], + "type": "text", + "content": "[33] Saed Moradi, Payman Moallem, and Mohamad Farzan Sabahi. A false-alarm aware methodology to develop robust and efficient multi-scale infrared small target detection algorithm. Infrared Physics & Technology, 89:387-397, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "text", + "content": "[34] Saed Moradi, Payman Moallem, and Mohamad Farzan Sabahi. A false-alarm aware methodology to develop robust and efficient multi-scale infrared small target detection algorithm. Infrared Physics & Technology, 89:387-397, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 462, + 287, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 527 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NeurlPS), 32, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 530, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 530, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 530, + 287, + 585 + ], + "type": "text", + "content": "[36] Rui Qian, Yunchao Wei, Honghui Shi, Jiachen Li, Jiaying Liu, and Thomas Huang. Weakly supervised scene parsing with point-based distance metric learning. In AAAI Conference on Artificial Intelligence (AAAI), volume 33, pages 8843-8850, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 588, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 632 + ], + "type": "text", + "content": "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[38] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7263-7271, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[40] Jeanfrancois Rivest and Roger Fortin. Detection of dim targets in digital infrared imagery by morphological image processing. Optical Engineering, 35(7):1886-1893, 1996." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 162, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 225 + ], + "type": "text", + "content": "[42] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[43] Yang Sun, Jungang Yang, and Wei An. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor model. IEEE Transactions on Geoscience and Remote Sensing (TGRS), 59(5):3737-3752, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "type": "text", + "content": "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "text", + "content": "[45] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In IEEE International Conference on Computer Vision (ICCV), pages 8509-8518, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "type": "text", + "content": "[46] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: u-net in u-net for infrared small object detection. IEEE Transactions on Image Processing (TIP), 32:364-376, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "text", + "content": "[47] Xinyi Ying, Yingqian Wang, Longguang Wang, Weidong Sheng, Li Liu, Zaiping Lin, and Shilin Zhou. Local motion and contrast priors driven deep network for infrared small target super-resolution. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTAR), 15:5480-5495, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 471, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 535 + ], + "type": "text", + "content": "[48] Mohsen Zand, Haleh Damirchi, Andrew Farley, Mahdiyar Molahasani, Michael Greenspan, and Ali Etemad. Multiscale crowd counting and localization by multitask point supervision. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1820-1824. IEEE, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 536, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 536, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 536, + 545, + 602 + ], + "type": "text", + "content": "[49] Ke Zhang, Shuyan Ni, Dashuang Yan, and Aidi Zhang. Review of dim small target detection algorithms in single-frame infrared images. In IEEE Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC), volume 4, pages 2115-2120. IEEE, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "type": "text", + "content": "[50] Landan Zhang, Lingbing Peng, Tianfang Zhang, Siying Cao, and Zhenming Peng. Infrared small target detection via nonconvex rank approximation minimization joint 12, 1 norm. Remote Sensing, 10(11):1821, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "text", + "content": "[51] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[52] Mingjin Zhang, Haichen Bai, Jing Zhang, Rui Zhang, Chaoyue Wang, Jie Guo, and Xinbo Gao. Rkformer: Runge-kutta transformer with random-connection attention for" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15537" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 464 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "infrared small target detection. In 30th ACM International Conference on Multimedia (ACM MM), pages 1730-1738, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "type": "text", + "content": "[53] Mingjin Zhang, Ke Yue, Jing Zhang, Yunsong Li, and Xinbo Gao. Exploring feature compensation and cross-level correlation for infrared small target detection. In 30th ACM International Conference on Multimedia (ACM MM), pages 1857-1865, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[54] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 877-886, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[55] Shiyin Zhang, Jun Hao Liew, Yunchao Wei, Shikui Wei, and Yao Zhao. Interactive object segmentation with inside-outside guidance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12234-12244, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "type": "text", + "content": "[56] Shilong Zhang, Zhuoran Yu, Liyang Liu, Xinjiang Wang, Aojun Zhou, and Kai Chen. Group r-cnn for weakly semi-supervised object detection with points. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9417-9426, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "type": "text", + "content": "[57] Mingjing Zhao, Wei Li, Lu Li, Jin Hu, Pengge Ma, and Ran Tao. Single-frame infrared small-target detection: A survey. IEEE Geoscience and Remote Sensing Magazine (GRSM), 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "type": "text", + "content": "[58] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2016." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "type": "text", + "content": "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2020." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15538" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_content_list.json b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a08890125cff9efc966377481f41062f7866e5c --- /dev/null +++ b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_content_list.json @@ -0,0 +1,1680 @@ +[ + { + "type": "text", + "text": "MarS3D: A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds", + "text_level": 1, + "bbox": [ + 86, + 130, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahui Liu $^{1*}$ Chirui Chang $^{1*}$ Jianhui Liu $^{1}$ Xiaoyang Wu $^{1}$ Lan Ma $^{2}$ Xiaojuan Qi $^{1\\dagger}$ $^{1}$ The University of Hong Kong $^{2}$ TCL AI Lab", + "bbox": [ + 156, + 202, + 836, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liujh, xjqi}@eee.hku.hk, crchang@hku.hk, jhliu0212@gmail.com, xywu3@cs.hku.hk, rubyma@tcl.com", + "bbox": [ + 66, + 241, + 898, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D semantic segmentation on multi-scan large-scale point clouds plays an important role in autonomous systems. Unlike the single-scan-based semantic segmentation task, this task requires distinguishing the motion states of points in addition to their semantic categories. However, methods designed for single-scan-based segmentation tasks perform poorly on the multi-scan task due to the lacking of an effective way to integrate temporal information. We propose MarS3D, a plug-and-play motion-aware module for semantic segmentation on multi-scan 3D point clouds. This module can be flexibly combined with single-scan models to allow them to have multi-scan perception abilities. The model encompasses two key designs: the Cross-Frame Feature Embedding module for enriching representation learning and the Motion-Aware Feature Learning module for enhancing motion awareness. Extensive experiments show that MarS3D can improve the performance of the baseline model by a large margin. The code is available at https://github.com/CVMI-Lab/MarS3D.", + "bbox": [ + 75, + 323, + 473, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 640, + 209, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D semantic segmentation on multi-scan large-scale point clouds is a fundamental computer vision task that benefits many downstream problems in autonomous systems, such as decision-making, motion planning, and 3D reconstruction, to name just a few. Compared with the single-scan semantic segmentation task, this task requires understanding not only the semantic categories but also the motion states (e.g., moving or static) of points based on multi-scan point cloud data.", + "bbox": [ + 75, + 666, + 468, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past few years, extensive research has been conducted on single-scan semantic segmentation with significant research advancements [4, 5, 12, 25, 31, 33, 36]. These approaches are also applied to process multi-scan point", + "bbox": [ + 75, + 803, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9d09b39f4540770679c524d7c0977eaec2b1b1ac17ef0fe847200928bce8d5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 290, + 738, + 353 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b4cd622f2f091fa38408dd9e49b0ecf4b1844aeabb9a3faf8c869c65d7230dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 291, + 885, + 353 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98afb6d81e5b1ed5376fdb0dba48dd8b30b7f108d1dcb422584c9fab0f357aef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 354, + 738, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6504b14347e31fa8590b698fda50d81fb4c21fdfc69a98b56425a4b8ebea1119.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 354, + 885, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5aae9841f52b11b350a1fa821a1a310293e330f6794d9ab896e93e849cbc05c.jpg", + "image_caption": [ + "Figure 1. Comparison of our proposed method, MarS3D, with baseline method using SPVCNN [25] as the backbone on SemanticKITTI [1] dataset. MarS3D achieves excellent results in the classification of semantic categories and motion states, while the baseline method can not distinguish motion well from static." + ], + "image_footnote": [], + "bbox": [ + 496, + 417, + 738, + 478 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fecbff26a11e4b7608fb3cdcb2c0e6ed2750c36801c1131ee3738b396d4d0c94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 417, + 885, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "clouds, wherein multiple point clouds are fused to form a single point cloud before being fed to the network for processing. Albeit simple, this strategy may lose temporal information and make distinguishing motion states a challenging problem. As a result, they perform poorly in classifying the motion states of objects. As shown in Figure 1, the simple point cloud fusion strategy cannot effectively enable the model to distinguish the motion states of cars even with a state-of-the-art backbone network SPVCNN [25]. Recently, there have been some early attempts [7, 23, 24, 28] to employ attention modules [24] and recurrent networks [7, 23, 28] to fuse information across different temporal frames. However, these approaches do not perform well on the multi-scan task due to the insufficiency of temporal representations and the limited feature extraction ability of the model.", + "bbox": [ + 496, + 580, + 893, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In sum, a systematic investigation of utilizing the rich spatial-temporal information from multiple-point cloud scans is still lacking. This requires answering two critical questions: (1) how can we leverage the multi-scan information to improve representation learning on point clouds for", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9372", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "better semantic understanding? and (2) how can the temporal information be effectively extracted and learned for classifying the motion states of objects?", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simple plug-and-play Motion-aware Segmentation module for 3D multi-scan analysis (MarS3D), which can seamlessly integrate with existing single-scan semantic segmentation models and endow them with the ability to perform accurate multi-scan 3D point cloud semantic segmentation with negligible computational costs. Specifically, our method incorporates two core designs: First, to enrich representation learning of multi-frame point clouds, we propose a Cross-Frame Feature Embedding (CFFE) module which embeds time-step information into features to facilitate inter-frame fusion and representation learning. Second, inspired by the observation that objects primarily move along the horizontal ground plane (i.e., $xy$ -plane) in large-scale outdoor scenes, i.e., minimal motion along the $z$ -axis, we propose a Motion-Aware Feature Learning (MAFL) module based on Bird's Eye View (BEV), which learns the motion patterns of objects between frames to facilitate effectively discriminating the motion states of objects.", + "bbox": [ + 75, + 137, + 470, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We extensively evaluate our approach upon several mainstream baseline frameworks on SemanticKITTI [1] and nuScenes [3] dataset. It consistently improves the performance of the baseline approaches, e.g., MinkUnet [5], by $6.24\\%$ in mIoU on SemanticKITTI with a negligible increase in model parameters, i.e., about $0.2\\%$ . The main contributions are summarized as follows:", + "bbox": [ + 75, + 424, + 468, + 529 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to propose a plug-and-play module for large-scale multi-scan 3D semantic segmentation, which can be flexibly integrated with mainstream single-scan segmentation models without incurring too much cost.", + "- We devise a Cross-Frame Feature Embedding module to fuse multiple point clouds while preserving their temporal information, thereby enriching representation learning for multi-scan point clouds.", + "- We introduce a BEV-based Motion-Aware Feature Learning module to exploit temporal information and enhance the model's motion awareness, facilitating the prediction of motion states.", + "- We conduct extensive experiments and comprehensive analyses of our approach with different backbone models. The proposed model performs favorably compared to the baseline methods while introducing negligible extra parameters and inference time." + ], + "bbox": [ + 94, + 535, + 467, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 217, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-scan Outdoor 3D Semantic Segmentation: 3D single-scan outdoor semantic segmentation is indispensable for autonomous driving. In early work, PointNet [19] uses", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Layer Perception (MLP) to extract features from input point clouds directly, and PointNet++ [20] tries to incorporate multi-scale designs for dense prediction tasks. Later, various literature [19, 20, 26, 29, 30, 32] works on designing point-based convolution on either geometric or semantic neighborhoods. To handle the large-scale dataset, some works [4, 5, 10, 12, 25, 31, 36] focus on volumetric features and use 3D convolution to achieve a balance between accuracy and efficiency. SparseConv [10] and MinkUNet [5] are representative works and demonstrate good performance. Later, SPVNAS [25] combines voxel and point representations and designs a neural architecture search method to find the optimal model structure. Recently, Cylinder3D [36] introduces a cylindrical partition to leverage the properties of LiDAR point clouds for enriching the feature information.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The remarkable feature extraction capability enables the above methods to achieve high performance on single-scan tasks. To solve the multi-scan task, most of these methods [4, 12, 31, 34, 36] first fuse multiple point clouds into one and treat the fused point cloud as a single point cloud for processing. Albeit simple, this fusion strategy overlooks important temporal information and entangles moving and non-moving objects, leading to performance degradation.", + "bbox": [ + 496, + 319, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-scan Outdoor 3D Semantic Segmentation: Compared to single-scan semantic segmentation, the multi-scan task needs to discriminate the moving and stationary states of the objects based on temporal information. In addition to the simple fusion strategy discussed above, another stream of approaches [7,23,24,28] attempts to process each point cloud in a sequence separately and fuse the feature representations for temporal modeling. For instance, SpSequenceNet [24] proposes a U-Net-based architecture to extract per-frame features and combine features of two consecutive frames to gather temporal information. The fused feature is further fed into the prediction head to produce results. Duerr et al. [7] design a recurrent architecture with a temporal memory alignment module for sequential processing of multiple point clouds. TemporalLatticeNet [23] proposes to match similar feature patterns between adjacent frames and fuse them temporally. However, these approaches cannot fully leverage multiple point clouds to enrich temporal representation learning as the feature extraction is still conducted on each frame separately.", + "bbox": [ + 496, + 450, + 890, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "BEV-based 3D Point Cloud Perception: Recently, BEV-based representation [11, 14, 16, 35, 37, 38] has emerged as an effective way to process 3D point cloud due to its efficiency and ease of deployment using 2D convolution operations. By projecting a 3D point cloud into the bird's eye view (BEV), BEV-based representation converts 3D representations into 2D to avoid heavy processing in 3D and improve computation and memory efficiency. Some representative works include PointPillars [14] for 3D object de", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9373", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad5b1b54236d4f4bc8b060629505e83f9ffd38f0e1446825bcd3f29ed3788137.jpg", + "image_caption": [ + "Figure 2. The proposed framework of MarS3D. We take three scan point clouds inputs as illustration. MarS3D contains two branches. One is the BEV branch with 2D BEV representations as input, and it employs a Motion-Aware Feature Learning (MAFL) module to enhance the motion-aware feature learning. The other is the 3D branch that takes multi-scan 3D point clouds as input, enriching the feature representations with our Cross-Frame Feature Embedding (CFFE) module. The fused features of the above two modules are fed into the prediction head and make final predictions on semantic categories and motion states." + ], + "image_footnote": [], + "bbox": [ + 73, + 88, + 890, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tection, BEVfusion [16] for multi-sensor fusion, BEV projection [38] for 3D segmentation. The above work demonstrates the potential of BEV representations in outdoor 3D scene analysis. Here, we explore utilizing BEV-based representation to extract temporal information for analyzing the motion states of objects.", + "bbox": [ + 75, + 433, + 472, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Video Semantic Segmentation: In video segmentation, many efforts [2,6,13,15,17,18,21] have been made to conduct temporal feature extraction. Oh et al. [18] introduce a space-time memory network to integrate features from adjacent frames for video object segmentation and attains significant performance gains. TDNet [13] exploits temporal redundancy for fast video segmentation by developing an attention propagation method to propagate features to adjacent frames. Although they achieve promising results on video segmentation, we experimentally demonstrate these designs are not optimal for multi-scan point cloud segmentation. We focus on designing new methods tailored to point clouds for better performance and efficiency.", + "bbox": [ + 75, + 536, + 472, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Statement", + "text_level": 1, + "bbox": [ + 76, + 752, + 264, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sequence of LiDAR point clouds as inputs, the multi-scan 3D point cloud semantic segmentation task aims to assign a semantic category to each point and predict their motion states (i.e., moving or static). Specifically, a LiDAR point cloud frame contains a set of unordered points that are annotated with labels for training. We denote a pair of training data as $(\\mathcal{P}_i,\\mathcal{L}_i) = \\{p_j,l_j\\}_{j = 1}^N$ with $p_j\\in \\mathbb{R}^{D_{\\mathrm{in}}}$ , where $N$ denotes the number of points. Each point $p_j$ contains in-", + "bbox": [ + 76, + 780, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "put descriptors with $D_{\\mathrm{in}}$ dimensions, including point coordinates $(x,y,z)$ and other features such as intensities (bm). The corresponding label $l_{j}$ incorporates both semantic categories and motion states of $p_j$ . Therefore, points belonging to the same semantic category but possessing distinct motion states are allocated distinct labels. For a sequence of point clouds $\\{\\mathcal{P}_i,\\mathcal{L}_i\\}_{i = 1}^M$ that contains $M$ frames, all frames are scanned sequentially in time order, and scanning poses and timestamps are used to align different frames into the same world-coordinate system. In the following, we omit the alignment process for simplicity, where the point clouds in a sequence are calibrated to the same coordinate system by default.", + "bbox": [ + 496, + 433, + 893, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 648, + 591, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 674, + 609, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our motion-aware model for multi-scan semantic segmentation, namely MarS3D, is shown in Figure 2. MarS3D contains a 3D branch for multi-scan spatial representation learning and a BEV branch for motion-aware feature learning. First, the BEV branch takes as inputs $k$ BEV representations $\\mathcal{B}^{\\mathrm{in}} = \\left\\{\\mathcal{B}_i^{\\mathrm{in}}\\right\\}_{i=1}^k$ that are derived by point cloud polarization (see Section 4.2) and outputs motion-aware feature map $\\mathcal{Z}^m$ . The core is a Motion-Aware Feature Learning (MAFL) module (see Section 4.2) that extracts and leverages BEV features through a dedicated design to produce a motion-aware feature map $\\mathcal{Z}^m$ . Second, the 3D branch takes as inputs the fused $k$ point clouds $\\mathcal{P}^{\\mathrm{in}} = \\left\\{\\mathcal{P}_i^{\\mathrm{in}}\\right\\}_{i=1}^k$ and outputs enriched 3D enhanced spa", + "bbox": [ + 496, + 699, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9374", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tial features $\\mathcal{P}^s$ . The 3D branch incorporates a cross-frame feature embedding (CFFE) module (see Section 4.3) to inject temporal information. It outputs embedded features, denoted as $\\mathcal{P}^{\\mathrm{ebd}} = \\left\\{\\mathcal{P}_i^{\\mathrm{ebd}}\\right\\}_{i=1}^k$ . These features are further processed by a single-scan backbone $f_s$ to yield multi-scan enhanced 3D spatial representations $\\mathcal{P}^s$ as:", + "bbox": [ + 75, + 90, + 472, + 186 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} ^ {s} = f _ {s} \\left(\\mathcal {P} ^ {\\mathrm {e b d}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 200, + 468, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{P}^s\\in \\mathbb{R}^{D_z}$ and $D_{z}$ is the dimension of the output feature of the single-scan backbone.", + "bbox": [ + 75, + 227, + 468, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the motion-aware feature map $\\mathcal{Z}^m$ and enhanced spatial features $\\mathcal{P}^s$ are combined to produce the fused features $\\mathcal{P}^f$ by aligning the coordinates of the features with the pixels of the motion-aware feature map. The fused features are then fed to the prediction head to produce the final outputs (see Section 4.4).", + "bbox": [ + 75, + 260, + 472, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. BEV Branch", + "text_level": 1, + "bbox": [ + 76, + 362, + 212, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we elaborate on the BEV branch, which targets extracting motion-friendly features for motion prediction (see Figure 2). Before delving into the details, we first introduce BEV mapping, which maps a 3D point cloud into a 2D BEV image. Then, we introduce our key Motion-Aware Feature Learning (MAFL) module for motion extraction.", + "bbox": [ + 75, + 386, + 468, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BEV Mapping: Each point cloud $\\mathcal{P}$ for the BEV branch is pillarized into a three-channel BEV map $\\mathcal{B}$ of size $H\\times W$ . Inspired by [14, 16], for a pillar located at $(x^{*},y^{*})$ with a pixel grid size of $l_{B}$ , its three-channel feature $b_{(x^{*},y^{*})}$ consists of the average translation along $x$ and $y$ axis regarding the grid center, and total intensity features (taking SemanticKITTI as an example) among all points inside the pillar, which is formulated as:", + "bbox": [ + 75, + 502, + 472, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nb _ {\\left(x ^ {*}, y ^ {*}\\right)} = \\left[ \\frac {1}{N ^ {*}} \\sum_ {i} ^ {N ^ {*}} \\frac {2 \\Delta x _ {i}}{l _ {B}}, \\frac {1}{N ^ {*}} \\sum_ {i} ^ {N ^ {*}} \\frac {2 \\Delta y _ {i}}{l _ {B}}, \\sum_ {i} ^ {N ^ {*}} \\mathrm {b m} _ {i} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 648, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N^{*}$ denotes the number of points in current pillar, while $\\Delta x_{i}$ and $\\Delta y_{i}$ denote the translation of the ith point along $x$ -axis and $y$ -axis respectively.", + "bbox": [ + 75, + 704, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion-Aware Feature Learning Module: Our Motion-Aware Feature Learning (MAFL) module is illustrated in the BEV branch in Figure 2 which is designed to extract motion-aware features for better distinguishing moving/static objects. The input to this module consists of $k$ multi-scan BEV representations $\\mathcal{B}^{\\mathrm{in}} = \\left\\{\\mathcal{B}_i^{\\mathrm{in}}\\right\\}_{i=1}^k$ , where $\\mathcal{B}_k^{\\mathrm{in}}$ represents the target frame. A lightweight 2D CNN $f_u$ with a UNet-like architecture [22] is used to extract features $\\mathcal{U} = \\{\\mathcal{U}_i\\}_{i=1}^k$ from multi-scan inputs as:", + "bbox": [ + 75, + 760, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathcal {U} _ {i} \\right\\} _ {i = 1} ^ {k} = \\left\\{f _ {u} \\left(\\mathcal {B} _ {i} ^ {\\text {i n}}\\right) \\right\\} _ {i = 1} ^ {k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 106, + 892, + 130 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, to identify moving objects, we take the difference between the target frame $k$ and the remaining $k - 1$ reference frames which outputs $\\mathcal{D} = \\{\\mathcal{D}_i\\}_{i = 1}^{k - 1}$ as:", + "bbox": [ + 498, + 138, + 893, + 188 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathcal {D} _ {i} \\right\\} _ {i = 1} ^ {k - 1} = \\left\\{\\mathcal {U} _ {k} - \\mathcal {U} _ {i} \\right\\} _ {i = 1} ^ {k - 1}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 204, + 892, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By doing so, the static objects can be erased, and the dynamic objects are highlighted with a large feature magnitude. Then, $\\{\\mathcal{D}_i\\}_{i=1}^{k-1}$ are channel-wise concatenated to form a new 2D map. Note that objects may have different moving patterns and velocities. Therefore, we design a multi-kernel convolutional network $f_m$ with multiple branches of various kernel sizes to capture objects with various movement patterns. Finally, the outputs from $f_m$ are concatenated to output a motion-aware feature map $\\mathcal{Z}^m$ .", + "bbox": [ + 498, + 236, + 893, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. 3D Branch", + "text_level": 1, + "bbox": [ + 500, + 387, + 620, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The 3D branch uses temporal information to enhance spatial representation learning on 3D point clouds. The core component is the Cross-Frame Feature Embedding (CFFE) module, whose output is fed into the single-scan backbone network to produce enhanced spatial features $\\mathcal{P}^s$ (see Figure 2). In the following, we elaborate on the CFFE module to improve spatial representation learning on multi-scan point clouds.", + "bbox": [ + 498, + 412, + 893, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Frame Feature Embedding Module: When the multi-frame point clouds are fused as discussed in the previous sections, points from different time steps are mixed, making it challenging for the final recognition. Inspired by positional embedding [27], we propose a Cross-Frame Feature Embedding (CFFE) module to generate a time-aware embedding and produce consistent features for each point across different timestamps. Given $k$ point cloud frames, $\\{\\mathcal{P}_i^{\\mathrm{in}}\\}_{i = 1}^k$ , we design an embedding neural network layer $f_{e}$ that maps point clouds into intermediate latent features and $k$ learnable temporal embeddings $\\mathcal{E} = \\{e_i\\}_{i = 1}^k$ corresponding to $k$ frames respectively. The dimension of $e_i$ is the same as the output dimension of the $f_{e}$ . The point-level embedded features $\\mathcal{P}^{\\mathrm{ebd}}$ are obtained by:", + "bbox": [ + 498, + 544, + 893, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathcal {P} _ {i} ^ {\\mathrm {e b d}} \\right\\} _ {i = 1} ^ {k} = \\left\\{e _ {i} + f _ {e} \\left(\\mathcal {P} _ {i} ^ {\\mathrm {i n}}\\right) \\right\\} _ {i = 1} ^ {k}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 777, + 892, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where element-wise summation is conducted on $e_i$ and each point of the corresponding $f_e(\\mathcal{P}_i^{\\mathrm{in}})$ . The obtained features with different temporal embeddings are represented in a point cloud $\\{\\mathcal{P}_i^{\\mathrm{ebd}}\\}_{i=1}^k$ and subsequently fed into the single-scan backbone network $f_s$ , following Eq. (1), to produce a set of enhanced spatial features $\\mathcal{P}^s$ .", + "bbox": [ + 498, + 809, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9375", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/07b6b8225f02675f00435063bf638d123456bafa3239d12c5089839553ec7271.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
B.MethodmIoU#paramlatencycarbic.mot.tru.ove.per.bil.mol.roa.par.sid.ogr.bui.fen.veg.trn.ter.pol.tra.mca.mbi.mpe.mmm.mov.mtr.
SPVCNN [25]Baseline49.7021.8M206ms93.934.464.768.033.019.70.00.093.645.280.10.290.359.788.463.575.664.151.974.386.755.00.00.00.0
Ours54.6621.9M225ms95.652.777.879.451.527.90.00.094.251.382.00.191.665.189.068.476.265.251.580.694.968.00.03.60.0
Δ+4.96+0.1M+19ms1.718.313.111.418.58.20.00.00.66.11.90.11.35.40.64.90.61.10.46.38.213.00.03.60.0
SparseConv [9]Baseline48.9939.2M239ms94.724.154.169.643.417.30.20.093.245.179.80.289.561.787.762.974.663.850.073.985.453.60.00.00.0
Ours54.6439.3M253ms96.635.269.083.364.826.90.00.094.061.282.50.190.965.888.167.875.266.251.483.594.468.90.00.00.0
Δ+5.65+0.1M+14ms1.911.114.913.721.49.60.20.00.816.12.70.11.44.10.44.90.62.41.49.69.015.30.00.00.0
MinkUnet [5]Baseline48.4737.9M295ms93.823.748.990.341.318.00.00.092.232.278.40.089.855.588.863.777.063.650.069.283.152.50.00.00.0
Ours54.7138.0M323ms96.428.470.093.962.731.60.00.093.858.581.70.192.667.689.066.776.466.551.682.693.164.40.00.10.0
Δ+6.24+0.1M+28ms2.64.721.13.621.413.60.00.01.626.33.30.12.812.10.23.00.62.91.613.410.011.90.00.10.0
", + "bbox": [ + 76, + 90, + 890, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of the proposed method, MarS3D, on SemanticKITTI [1] multi-scan public validation set. Combined with different mainstream single-scan 3D point cloud semantic segmentation backbones, MarS3D has a large performance improvement over the corresponding baseline methods without introducing excessive parameters and each-frame inference time. (B. indicates Backbone, full names of the categories are in the supplementary material, blue indicates degradation.)", + "bbox": [ + 73, + 267, + 892, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Feature Fusion and Prediction", + "text_level": 1, + "bbox": [ + 75, + 347, + 346, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, equipped with the motion-aware feature map $(\\mathcal{Z}^m)$ and enhanced spatial features $(\\mathcal{P}^s)$ , the next step is to fully integrate the representation information from both branches and make predictions.", + "bbox": [ + 75, + 371, + 468, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Fusion: The motion-aware feature map $\\mathcal{Z}^m\\in$ $\\mathbb{R}^{H\\times W\\times D_z}$ ( $H$ : height; $W$ : width; $D_{z}$ : number of channels) is obtained from the BEV branch, while the enhanced spatial features $\\mathcal{P}^s\\in \\mathbb{R}^{N\\times D_p}$ ( $N$ : number of points; $D_{p}$ : number of channels) are the outputs of the 3D branch. The feature fusion module aggregates information from the above two representations to make subsequent predictions. For each point feature in $\\mathcal{P}^s$ with $D_{p}$ dimension, a corresponding pixel from $\\mathcal{Z}^m$ can be queried based on the point's 3D location ( $x$ and $y$ coordinates). This pixel serves as an index to extract a $D_{z}$ -dimensional motion-aware feature along the channel dimension from $\\mathcal{Z}^m$ . Subsequently, the motion-aware feature is concatenated with the point feature, resulting in the fused features denoted as $\\mathcal{P}^f\\in \\mathbb{R}^{N\\times (D_z + D_p)}$ .", + "bbox": [ + 75, + 441, + 470, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prediction: Based on the correspondence between 2D and 3D, $\\mathcal{P}^f$ is then fed into an MLP classifier $f_{\\mathrm{cls}}$ to obtain the output $s_{\\mathrm{pred}}$ :", + "bbox": [ + 75, + 676, + 468, + 723 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\text {p r e d}} = f _ {\\mathrm {c l s}} \\left(\\mathcal {P} ^ {f}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 736, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_{\\mathrm{pred}}$ is the logits for the prediction result of the semantic segmentation task of the input point cloud.", + "bbox": [ + 75, + 761, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Model Training and Inference", + "text_level": 1, + "bbox": [ + 75, + 801, + 344, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, the obtained fused features $\\mathcal{P}^f$ are fed into two classification heads: the category-aware classification head and the motion-aware classification head (a binary classifier), which outputs the predicted logits $s_{\\mathrm{pred}}^c$ and $s_{\\mathrm{pred}}^{m}$ (more details are included in the supplementary mate", + "bbox": [ + 75, + 824, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rial). First, with the ground truth labels of semantic categories $\\mathcal{L}_{\\mathrm{GT}}^c$ , back-propagation is performed for parameters optimization using Cross Entropy (CE) loss $L_{c}$ for semantic categories classification:", + "bbox": [ + 496, + 348, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c} = \\mathrm {C E} \\left(s _ {\\text {p r e d}} ^ {c}, \\mathcal {L} _ {\\mathrm {G T}} ^ {c}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 424, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{CE}(\\cdot, \\cdot)$ is the cross-entropy loss, and $s_{\\mathrm{pred}}^c$ is the prediction probability for each point. Then, the Binary Cross Entropy (BCE) loss $L_m$ is used for motion states classification with motion states ground truth $\\mathcal{L}_{\\mathrm{GT}}^m$ :", + "bbox": [ + 496, + 448, + 890, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m} = \\operatorname {B C E} \\left(s _ {\\text {p r e d}} ^ {m}, \\mathcal {L} _ {\\mathrm {G T}} ^ {m}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 523, + 890, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{BCE}(\\cdot ,\\cdot)$ indicates the binary cross-entropy loss.", + "bbox": [ + 500, + 547, + 866, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final objective function $L$ of the optimization is:", + "bbox": [ + 517, + 564, + 866, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\omega_ {c} \\cdot L _ {c} + \\omega_ {m} \\cdot L _ {m}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 595, + 890, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega_{c}$ and $\\omega_{m}$ are the weights of the two losses ( $L_{c}$ and $L_{m}$ ) respectively.", + "bbox": [ + 500, + 618, + 890, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, the final prediction result is determined using the logits produced by both classification heads. When presented with an input sample, the motion-aware classification head will identify the motion state of the input point only if the category-aware classification head recognizes the point as belonging to a class with the potential to move.", + "bbox": [ + 496, + 648, + 890, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 768, + 633, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metric: We evaluate our method on SemanticKITTI [1] and nuScenes [3]. For SemanticKITTI, the multi-scan setting is fully supervised and contains 25-category (6 moving categories and 19 static categories) with high-quality semantic annotations. The annotations are based on the KITTI dataset [8]. It comprises 22 point cloud sequences. For nuScenes, we propose a new", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9376", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fdfe8373ac4988da781f16f1c0956a6a63ea4126312dafc6113a1b385a806c8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 71, + 89, + 352, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d17f7a62d66c608711c520f13094be453959a77160a44cd877c3f7b84086130.jpg", + "image_caption": [ + "Baseline" + ], + "image_footnote": [ + "- Car (non-moving)", + "- Car (moving)", + "- Person (non-moving)" + ], + "bbox": [ + 73, + 181, + 352, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef1b4cf530ee238f319cf758950b4955c935fca600019284733d6472e36ba12a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 90, + 620, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6459536973607d4ae86efe76562bc8a2051376f76b613a2d5794f878d975d292.jpg", + "image_caption": [ + "" + ], + "image_footnote": [ + "- Person (moving) - Bicyclist (moving)" + ], + "bbox": [ + 354, + 181, + 620, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6794fae2860a3c29e58d54b953e90821b595c1066010d124a5302281a143f02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 90, + 887, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d9e286c6dd5611c153be16992e85179a47d14f045a53e1e70969bffc6da74399.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 622, + 181, + 887, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/effa9436ea92760b55b5cfb705532b140e11f20a8b33129899836d89e9f594b5.jpg", + "image_caption": [ + "Figure 3. Qualitative results on SemanticKITTI [1] public validation dataset. With SPVCNN [25] as the backbone, the segmentation results on the SemanticKITTI multi-scan task of the baseline model and our model are shown together with ground truth. At the same time, a specific area containing moving points is magnified and displayed at the top right of each sub-figure." + ], + "image_footnote": [], + "bbox": [ + 68, + 367, + 274, + 478 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec414721ce044259c8e9d1599057931109e0412ea4c2266f888174f7b6dd6e8a.jpg", + "image_caption": [ + "Baseline", + "Figure 4. Evaluation errors (shown in red) by the baseline methods and MarS3D (using SPVCNN [25] as the backbone) on SemanticKITTI [1] public validation dataset. MarS3D significantly outperforms the baseline. Contrasting regions in the evaluation errors are highlighted with boxes and corresponding categories." + ], + "image_footnote": [], + "bbox": [ + 68, + 478, + 274, + 588 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59323c4b16e130f2985db3bb202931410b138e41c033a49a04be48319b6335c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 368, + 467, + 477 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e4b35d0dd42719f3e854a9e94352af08214f10a6090b2f12a0c272fe4f96d90.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 276, + 478, + 467, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multi-scan setting based on the 'lidar-seg' task (16 semantic categories) without reference frame supervision. We use its object-level velocity to construct a multi-scan segmentation dataset with 24 categories (8 moving and 16 static categories). More details are provided in the supplementary material. To assess the effectiveness of our proposed method and make comparisons with baselines and other methods, we use the mean Intersection over Union (mIoU) as the evaluation metric.", + "bbox": [ + 75, + 705, + 468, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: Our model is designed as a plug-and-play module that provides motion-aware features to enhance the backbone learned features. For the backbone", + "bbox": [ + 75, + 854, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model, we consider SPVCNN $[25]^*$ , SparseConv $[9]^{\\dagger}$ , and MinkUNet $[5]^{\\ddagger}$ . Following previous works [12, 23, 36], we use the current and its previous two frames as input. The size of the BEV representation is set to $501 \\times 301$ , and the multiple kernel sizes in the MAFL module are set as 1, 3, and 5, respectively. We set the embedded feature dimension in the CFFE module to 18, and the data augmentations are the same as the standard settings. All the models are trained on GeForce RTX 3090 GPUs, and the inference latency is recorded using a single GeForce RTX 3090 GPU.", + "bbox": [ + 496, + 369, + 892, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Main Results", + "text_level": 1, + "bbox": [ + 500, + 529, + 637, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Baseline Methods: We evaluate the performance of our proposed method on multi-scan benchmarks of SemanticKITTI [1] and nuScenes [3]. The baseline method uses the same backbone to process multiscan point clouds. We then compare this baseline approach to the same backbone augmented with MarS3D for a fair evaluation. As shown in Table 1, MarS3D significantly improves the performance over baseline methods on the public validation set of SemanticKITTI. With the most lightweight network, SPVCNN [25], MarS3D brings a $4.96\\%$ improvement while introducing less than $0.5\\%$ additional parameters. Particularly, consistent performance gains are observed on the dynamic object classes with non-moving/moving properties (i.e., car & moving car, person & moving person). This shows that the proposed BEV branch is both lightweight and powerful. Further, to verify the generalizability of the model, we offer a multiscan task based on nuScenes 'lidar-seg' dataset [3]. Our method outperforms the baseline (using MinkUNet [5] as the backbone) by a significant margin, achieving $64.83\\%$", + "bbox": [ + 496, + 551, + 892, + 854 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "\\*https://github.com/mit-han-lab/spvnas", + "https://github.com/traveller59/spconv", + "$\\S$ https://github.com/NVIDIA/MinkowskiEngine" + ], + "bbox": [ + 514, + 862, + 846, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9377", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/249e57d9c257023b89b2b83e5d324dbd5dfb9dc1ce1312f6795425f3dfa09a58.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoUcarbic.mot.tru.ove.per.bil.mol.roa.par.sid.ogr.bui.fen.veg.trn.ter.pol.tra.mca.mbi.mpe.mmm.mov.mtr.
SpSequenceNet [24]43.188.524.026.229.222.76.30.00.090.157.973.927.191.266.884.066.065.750.848.753.241.226.236.22.30.1
TemporalLidarSeg [7]47.092.147.740.939.235.014.40.00.091.859.675.823.289.863.882.362.564.752.660.468.242.840.412.912.42.1
TemporalLatticeNet [23]47.191.635.436.126.923.09.40.00.091.559.375.327.589.665.384.666.770.457.260.459.741.751.048.85.90.0
Meta-RangeSeg [28]49.590.152.743.930.335.414.30.00.090.763.374.726.990.563.583.067.067.756.464.464.556.155.024.420.33.4
KPConv [26]51.293.744.947.243.538.621.60.00.086.558.470.526.790.864.584.670.366.057.053.969.467.467.547.24.75.8
Baseline49.289.839.434.039.421.08.91.80.089.162.072.412.990.563.984.668.468.758.960.169.363.558.756.59.53.6
Ours52.795.149.249.539.736.616.21.20.089.966.874.326.492.168.286.072.170.562.864.878.467.358.036.310.05.1
", + "bbox": [ + 78, + 90, + 890, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison with the state-of-the-art models on SemanticKITTI multi-scan benchmark (official test set). MarS3D (with SPVCNN [25] as the backbone) significantly outperforms these models for multi-scan tasks. (full names of the categories are in the supplementary material.)", + "bbox": [ + 75, + 237, + 892, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mIoU compared to the baseline's $61.90\\%$ . This improvement is achieved with a negligible increase in each-frame inference time from 53ms to 58ms.", + "bbox": [ + 75, + 301, + 468, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with State-of-the-Art Methods: Compared to various models applied to multi-scan tasks, MarS3D (with SPVCNN as the backbone) is evaluated on the SemanticKITTI multi-scan benchmark. As shown in Table 2, the proposed approach has demonstrated superior performance, with a $1.5\\%$ increase in mIoU compared to the current state-of-the-art method, KPCov [26]. Furthermore, our method performs similarly or better than other state-of-the-art models across nearly all categories.", + "bbox": [ + 75, + 357, + 468, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparisons: Quantitative results are shown in Figure 3. The baseline model suffers from mistaking the status of static cars (Figure 3: Scan A) and moving persons (Figure 3: Scan B). It even recognizes the moving bicyclist as a moving person (Figure 3: Scan A). In contrast, MarS3D can circumvent such category and motion state discrimination errors. In addition to achieving better qualitative results in motion states, MarS3D also outperforms the baseline in semantic categories. We visualize the error maps (errors are shown in red) of the baseline and our method in Figure 4, where bounding boxes indicating specific categories highlight the differences between our method and the baseline. These improvements indicate that MarS3D has stronger semantic feature extraction capabilities compared to the baseline model, resulting in better segmentation performance even for immobile objects.", + "bbox": [ + 75, + 502, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 752, + 241, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct comprehensive ablation experiments on the SemanticKITTI multi-scan validation set to examine the effects of each component in our proposed method. As shown in Table 3, we gradually add three components to the baseline method, including the CFFE module, vanilla BEV branch (introduce only BEV representa", + "bbox": [ + 75, + 775, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81e0a066f37e0e9938026123e3e86a55dfd59577493f3c276d3dcc4cf746e8b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 301, + 681, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e245684290a05cce033d6b53c5e58de8f0b1f8a6f819f95d9dd6e3ccb695cbd2.jpg", + "image_caption": [ + "BEV Ground Truth (Motion)" + ], + "image_footnote": [], + "bbox": [ + 511, + 407, + 681, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44dbee35b43054220618e456657175e3fb700ca7c258b47830aec9dacb8643c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 301, + 844, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fae1f5c467451ab53fc932b8b3446bf8c7e2080091e639c8a1a6452fe74f2c75.jpg", + "image_caption": [ + "Ours", + "Figure 5. Sampled discrepant feature maps demonstrate that the region that contains moving points are represented higher activation values than other regions." + ], + "image_footnote": [], + "bbox": [ + 684, + 407, + 843, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tions and 2D CNN without the MAFL module), and the MAFL module to illustrate the effectiveness of our designs.", + "bbox": [ + 498, + 590, + 890, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of CFFE Module: As shown in Table 3, we first employ only the CFFE module with the baseline model and observe a significant performance improvement. For instance, with SPVCNN [25] as the backbone, the introduction of the CFFE module led to a boost in performance by $1.1\\%$ in terms of mIoU. This confirms the efficacy of the CFFE module for multi-frame representation learning.", + "bbox": [ + 498, + 632, + 890, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of BEV Representation: We also observe a significant performance improvement when we add the vanilla BEV branch alone to the baseline model. Using SPVCNN [25] as the backbone, the BEV branch results in a $3.7\\%$ increase in mIoU. This confirms the importance of BEV representations in extracting motion-aware information. Furthermore, when both the CFFE module and the vanilla BEV branch are added to the baseline model, there is a further increase in performance, demonstrating that the two components complement each other.", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$\\S$ http://wwwsemantic-kitti.org/tasks - Semantic Segmentation - Multiple Scans", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9378", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5e46666dda373ed0282cc09b7d208dbb1148d705eb6d4e0e20db91670110bca0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCFFEBEVMAFLmIoU(%)
baseline---49.7 / 49.0 / 48.5
proposed✓--50.8 / 51.3 / 52.1
-✓-53.4 / 53.9 / 53.1
✓✓-53.6 / 54.2 / 54.1
✓✓✓54.7 / 54.6 / 54.7
", + "bbox": [ + 81, + 90, + 465, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a7f02ca5d471d4e8beb578d5fa31878cfc3b1f78ea6f6d53a94d687ba0607b8a.jpg", + "table_caption": [ + "Table 3. Ablation studies on different backbones (SPVCNN / SparseConv / MinkUNet) on SemanticKITTI [1] public validation dataset. The effectiveness of different designs is demonstrated step-by-step and our method is marked with a gray box." + ], + "table_footnote": [], + "table_body": "
MethodVanilla BEVTDNet [13]STM [18]Ours
mIoU(%)53.3653.56 (+0.20)53.64 (+0.28)54.66 (+1.30)
", + "bbox": [ + 84, + 273, + 464, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparing our method (using SPVCNN [25] as the backbone) with different 2D temporal semantic segmentation approaches the public validation dataset of SemanticKITTI [1].", + "bbox": [ + 76, + 311, + 467, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of MAFL Module: In the final experiment presented in Table 3, we study the impact of our proposed MAFL module on the BEV representations, resulting in an increase of around $5.0\\%$ from the baseline model (using SPVCNN [25] as the backbone). By stacking all the proposed components together, our final solution (marked with a grey box in Table 3) reaches its pinnacle in performance. The improvement in other backbones is also significant, and more statistical results are provided in the supplementary material. In addition, the multi-channel discrepant feature map represents the differences between the extracted feature maps of BEV representations from two point clouds. For a discrepant feature map generated during inference, we randomly sample the channels pixel by pixel. The average absolute activation values of pixels with the same motion state in the sampled channels are re-scaled and aggregated into a single-channel feature map as shown in Figure 5. The sampled discrepant feature map has higher activation on pixels containing moving points compared to other regions. This clearly demonstrates that the MAFL module clearly distinguishes between regions that contain moving points and those that do not.", + "bbox": [ + 75, + 378, + 468, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. More Comparison on Temporal Segmentation", + "text_level": 1, + "bbox": [ + 76, + 724, + 465, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As for BEV representation learning, the proposed MAFL module is utilized on extracted 2D feature maps for capturing inter-frame temporal information of point clouds. We further conduct comparative experiments by replacing the MAFL module with its 2D counterparts, i.e., STM [18] and TDNet [13]. These 2D modules have been widely utilized for preserving temporal information in 2D tasks. The results of the comparative experiments using SPVCNN [25] as the backbone are presented in Table 4. According to the evaluation results, it can be concluded that the MAFL module", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shows superior performance compared to the other models, thereby confirming its remarkable effectiveness. This shows that the proposed MAFL module is better suited for the specific task of handling motion-aware semantic segmentation in 3D point clouds.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Limitations and Failure Cases Analysis", + "text_level": 1, + "bbox": [ + 498, + 179, + 834, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although MarS3D demonstrates impressive overall performance on the SemanticKITTI multi-scan benchmark, some limitations can be identified from the quantitative results in Table 1 and Table 2. The SemanticKITTI training dataset has an imbalanced distribution of point categories (more details are included in the supplementary material), which causes MarS3D to perform poorly on long-tailed categories due to insufficient training data on these categories. As shown in Table 2, MarS3D fails to show effects on several long-tail categories (i.e., bicyclist, motorcyclist, moving-other-vehicle, and moving-truck). Exploring solutions to address the long-tail problem is a promising research direction for the work. Since our model assumes planar motion, it may not perform well in scenarios where objects move in non-planar ways, such as on steep terrains or on non-planar surfaces.", + "bbox": [ + 496, + 203, + 890, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 462, + 617, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose MarS3D, a novel plug-and-play motion-aware model for 3D multi-scan point cloud semantic segmentation. The Motion-Aware Feature Learning (MAFL) module, based on BEV representations, is designed to facilitate extracting motion-aware representations. Additionally, the Cross-Frame Feature Embedding (CFFE) module is introduced to improve representation learning by embedding time-step information into features, thus preserving rich temporal information. Extensive experiments and ablation studies demonstrate that MarS3D significantly improves multiple 3D semantic segmentation baselines while introducing minimal overheads. In comparison to state-of-the-art methods designed for multi-scan tasks, MarS3D achieves superior performance and offers significant improvements over baseline methods. The proposed MarS3D model demonstrates the potential for effectively incorporating motion-awareness into 3D point cloud semantic segmentation tasks, providing a strong foundation for further research and development in this area.", + "bbox": [ + 496, + 488, + 890, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant no. 17202422), and RGC matching fund scheme (RMGS). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9379", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF International Conf. on Computer Vision (ICCV), 2019. 1, 2, 5, 6, 8", + "[2] Goutam Bhat, Felix Järemo Lawin, Martin Danelljan, Andreas Robinson, Michael Felsberg, Luc Van Gool, and Radu Timofte. Learning what to learn for video object segmentation. In European Conference on Computer Vision, pages 777-794. Springer, 2020. 3", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5, 6", + "[4] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 1, 2", + "[5] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 1, 2, 5, 6", + "[6] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 3", + "[7] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 2020 International Conference on 3D Vision (3DV), pages 781-790. IEEE, 2020. 1, 2, 7", + "[8] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 5", + "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 5, 6", + "[10] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. 2", + "[11] Noureldin Hendy, Cooper Sloan, Feng Tian, Pengfei Duan, Nick Charchut, Yuesong Xie, Chuang Wang, and James Philbin. Fishing net: Future inference of semantic heatmaps in grids. arXiv preprint arXiv:2006.09917, 2020. 2", + "[12] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In Proceedings of the IEEE/CVF Con" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 8479-8488, 2022. 1, 2, 6", + "[13] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks for fast video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8827, 2020. 3, 8", + "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2, 4", + "[15] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In European Conference on Computer Vision, pages 352-368. Springer, 2020. 3", + "[16] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 2, 3, 4", + "[17] Jiaxu Miao, Yunchao Wei, and Yi Yang. Memory aggregation networks for efficient interactive video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10366-10375, 2020. 3", + "[18] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226–9235, 2019. 3, 8", + "[19] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[20] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2", + "[21] Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7406-7415, 2020. 3", + "[22] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4", + "[23] Peer Schutt, Radu Alexandru Rosu, and Sven Behnke. Abstract flow for temporal semantic segmentation on the permutohedral lattice. In 2022 International Conference on Robotics and Automation (ICRA), pages 5139-5145. IEEE, 2022. 1, 2, 6, 7", + "[24] Hanyu Shi, Guosheng Lin, Hao Wang, Tzu-Yi Hung, and Zhenhua Wang. Spsequencenet: Semantic segmentation network on 4d point clouds. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 503, + 93, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9380", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on computer vision and pattern recognition, pages 4574-4583, 2020. 1, 2, 7", + "[25] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 1, 2, 5, 6, 7, 8", + "[26] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2, 7", + "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[28] Song Wang, Jianke Zhu, and Ruixiang Zhang. Meta-rangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. arXiv preprint arXiv:2202.13377, 2022. 1, 2, 7", + "[29] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2", + "[30] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9621-9630, 2019. 2", + "[31] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. 1, 2", + "[32] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3173-3182, 2021. 2", + "[33] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3101-3109, 2021. 1", + "[34] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695. Springer, 2022. 2", + "[35] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2", + "[36] Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550, 2020. 1, 2, 6", + "[37] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 2", + "[38] Zhenhong Zou and Yizhe Li. Efficient urban-scale point clouds segmentation with bev projection. arXiv preprint arXiv:2109.09074, 2021. 2, 3" + ], + "bbox": [ + 501, + 92, + 893, + 219 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9381", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_model.json b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..960972c5d6ae3619e08613e024ddb08cebfeb2ce --- /dev/null +++ b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.087, + 0.131, + 0.885, + 0.175 + ], + "angle": 0, + "content": "MarS3D: A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.203, + 0.837, + 0.24 + ], + "angle": 0, + "content": "Jiahui Liu\\(^{1*}\\) Chirui Chang\\(^{1*}\\) Jianhui Liu\\(^{1}\\) Xiaoyang Wu\\(^{1}\\) Lan Ma\\(^{2}\\) Xiaojuan Qi\\(^{1\\dagger}\\) \n\\(^{1}\\)The University of Hong Kong \\(^{2}\\)TCL AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.068, + 0.242, + 0.899, + 0.259 + ], + "angle": 0, + "content": "{liujh, xjqi}@eee.hku.hk, crchang@hku.hk, jhliu0212@gmail.com, xywu3@cs.hku.hk, rubyma@tcl.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.611 + ], + "angle": 0, + "content": "3D semantic segmentation on multi-scan large-scale point clouds plays an important role in autonomous systems. Unlike the single-scan-based semantic segmentation task, this task requires distinguishing the motion states of points in addition to their semantic categories. However, methods designed for single-scan-based segmentation tasks perform poorly on the multi-scan task due to the lacking of an effective way to integrate temporal information. We propose MarS3D, a plug-and-play motion-aware module for semantic segmentation on multi-scan 3D point clouds. This module can be flexibly combined with single-scan models to allow them to have multi-scan perception abilities. The model encompasses two key designs: the Cross-Frame Feature Embedding module for enriching representation learning and the Motion-Aware Feature Learning module for enhancing motion awareness. Extensive experiments show that MarS3D can improve the performance of the baseline model by a large margin. The code is available at https://github.com/CVMI-Lab/MarS3D." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.641, + 0.21, + 0.657 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.803 + ], + "angle": 0, + "content": "3D semantic segmentation on multi-scan large-scale point clouds is a fundamental computer vision task that benefits many downstream problems in autonomous systems, such as decision-making, motion planning, and 3D reconstruction, to name just a few. Compared with the single-scan semantic segmentation task, this task requires understanding not only the semantic categories but also the motion states (e.g., moving or static) of points based on multi-scan point cloud data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.471, + 0.866 + ], + "angle": 0, + "content": "In the past few years, extensive research has been conducted on single-scan semantic segmentation with significant research advancements [4, 5, 12, 25, 31, 33, 36]. These approaches are also applied to process multi-scan point" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.291, + 0.74, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.292, + 0.887, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.355, + 0.74, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.355, + 0.887, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.418, + 0.74, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.418, + 0.887, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.487, + 0.895, + 0.557 + ], + "angle": 0, + "content": "Figure 1. Comparison of our proposed method, MarS3D, with baseline method using SPVCNN [25] as the backbone on SemanticKITTI [1] dataset. MarS3D achieves excellent results in the classification of semantic categories and motion states, while the baseline method can not distinguish motion well from static." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.895, + 0.824 + ], + "angle": 0, + "content": "clouds, wherein multiple point clouds are fused to form a single point cloud before being fed to the network for processing. Albeit simple, this strategy may lose temporal information and make distinguishing motion states a challenging problem. As a result, they perform poorly in classifying the motion states of objects. As shown in Figure 1, the simple point cloud fusion strategy cannot effectively enable the model to distinguish the motion states of cars even with a state-of-the-art backbone network SPVCNN [25]. Recently, there have been some early attempts [7, 23, 24, 28] to employ attention modules [24] and recurrent networks [7, 23, 28] to fuse information across different temporal frames. However, these approaches do not perform well on the multi-scan task due to the insufficiency of temporal representations and the limited feature extraction ability of the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In sum, a systematic investigation of utilizing the rich spatial-temporal information from multiple-point cloud scans is still lacking. This requires answering two critical questions: (1) how can we leverage the multi-scan information to improve representation learning on point clouds for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9372" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "better semantic understanding? and (2) how can the temporal information be effectively extracted and learned for classifying the motion states of objects?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.424 + ], + "angle": 0, + "content": "In this paper, we propose a simple plug-and-play Motion-aware Segmentation module for 3D multi-scan analysis (MarS3D), which can seamlessly integrate with existing single-scan semantic segmentation models and endow them with the ability to perform accurate multi-scan 3D point cloud semantic segmentation with negligible computational costs. Specifically, our method incorporates two core designs: First, to enrich representation learning of multi-frame point clouds, we propose a Cross-Frame Feature Embedding (CFFE) module which embeds time-step information into features to facilitate inter-frame fusion and representation learning. Second, inspired by the observation that objects primarily move along the horizontal ground plane (i.e., \\(xy\\)-plane) in large-scale outdoor scenes, i.e., minimal motion along the \\(z\\)-axis, we propose a Motion-Aware Feature Learning (MAFL) module based on Bird's Eye View (BEV), which learns the motion patterns of objects between frames to facilitate effectively discriminating the motion states of objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.53 + ], + "angle": 0, + "content": "We extensively evaluate our approach upon several mainstream baseline frameworks on SemanticKITTI [1] and nuScenes [3] dataset. It consistently improves the performance of the baseline approaches, e.g., MinkUnet [5], by \\(6.24\\%\\) in mIoU on SemanticKITTI with a negligible increase in model parameters, i.e., about \\(0.2\\%\\). The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.536, + 0.468, + 0.609 + ], + "angle": 0, + "content": "- We are the first to propose a plug-and-play module for large-scale multi-scan 3D semantic segmentation, which can be flexibly integrated with mainstream single-scan segmentation models without incurring too much cost." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.613, + 0.468, + 0.673 + ], + "angle": 0, + "content": "- We devise a Cross-Frame Feature Embedding module to fuse multiple point clouds while preserving their temporal information, thereby enriching representation learning for multi-scan point clouds." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.676, + 0.468, + 0.735 + ], + "angle": 0, + "content": "- We introduce a BEV-based Motion-Aware Feature Learning module to exploit temporal information and enhance the model's motion awareness, facilitating the prediction of motion states." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.739, + 0.468, + 0.813 + ], + "angle": 0, + "content": "- We conduct extensive experiments and comprehensive analyses of our approach with different backbone models. The proposed model performs favorably compared to the baseline methods while introducing negligible extra parameters and inference time." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.536, + 0.468, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.218, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Single-scan Outdoor 3D Semantic Segmentation: 3D single-scan outdoor semantic segmentation is indispensable for autonomous driving. In early work, PointNet [19] uses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Multi-Layer Perception (MLP) to extract features from input point clouds directly, and PointNet++ [20] tries to incorporate multi-scale designs for dense prediction tasks. Later, various literature [19, 20, 26, 29, 30, 32] works on designing point-based convolution on either geometric or semantic neighborhoods. To handle the large-scale dataset, some works [4, 5, 10, 12, 25, 31, 36] focus on volumetric features and use 3D convolution to achieve a balance between accuracy and efficiency. SparseConv [10] and MinkUNet [5] are representative works and demonstrate good performance. Later, SPVNAS [25] combines voxel and point representations and designs a neural architecture search method to find the optimal model structure. Recently, Cylinder3D [36] introduces a cylindrical partition to leverage the properties of LiDAR point clouds for enriching the feature information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.442 + ], + "angle": 0, + "content": "The remarkable feature extraction capability enables the above methods to achieve high performance on single-scan tasks. To solve the multi-scan task, most of these methods [4, 12, 31, 34, 36] first fuse multiple point clouds into one and treat the fused point cloud as a single point cloud for processing. Albeit simple, this fusion strategy overlooks important temporal information and entangles moving and non-moving objects, leading to performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.755 + ], + "angle": 0, + "content": "Multi-scan Outdoor 3D Semantic Segmentation: Compared to single-scan semantic segmentation, the multi-scan task needs to discriminate the moving and stationary states of the objects based on temporal information. In addition to the simple fusion strategy discussed above, another stream of approaches [7,23,24,28] attempts to process each point cloud in a sequence separately and fuse the feature representations for temporal modeling. For instance, SpSequenceNet [24] proposes a U-Net-based architecture to extract per-frame features and combine features of two consecutive frames to gather temporal information. The fused feature is further fed into the prediction head to produce results. Duerr et al. [7] design a recurrent architecture with a temporal memory alignment module for sequential processing of multiple point clouds. TemporalLatticeNet [23] proposes to match similar feature patterns between adjacent frames and fuse them temporally. However, these approaches cannot fully leverage multiple point clouds to enrich temporal representation learning as the feature extraction is still conducted on each frame separately." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "BEV-based 3D Point Cloud Perception: Recently, BEV-based representation [11, 14, 16, 35, 37, 38] has emerged as an effective way to process 3D point cloud due to its efficiency and ease of deployment using 2D convolution operations. By projecting a 3D point cloud into the bird's eye view (BEV), BEV-based representation converts 3D representations into 2D to avoid heavy processing in 3D and improve computation and memory efficiency. Some representative works include PointPillars [14] for 3D object de" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9373" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.075, + 0.089, + 0.891, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.342, + 0.897, + 0.414 + ], + "angle": 0, + "content": "Figure 2. The proposed framework of MarS3D. We take three scan point clouds inputs as illustration. MarS3D contains two branches. One is the BEV branch with 2D BEV representations as input, and it employs a Motion-Aware Feature Learning (MAFL) module to enhance the motion-aware feature learning. The other is the 3D branch that takes multi-scan 3D point clouds as input, enriching the feature representations with our Cross-Frame Feature Embedding (CFFE) module. The fused features of the above two modules are fed into the prediction head and make final predictions on semantic categories and motion states." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.473, + 0.526 + ], + "angle": 0, + "content": "tection, BEVfusion [16] for multi-sensor fusion, BEV projection [38] for 3D segmentation. The above work demonstrates the potential of BEV representations in outdoor 3D scene analysis. Here, we explore utilizing BEV-based representation to extract temporal information for analyzing the motion states of objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.473, + 0.734 + ], + "angle": 0, + "content": "2D Video Semantic Segmentation: In video segmentation, many efforts [2,6,13,15,17,18,21] have been made to conduct temporal feature extraction. Oh et al. [18] introduce a space-time memory network to integrate features from adjacent frames for video object segmentation and attains significant performance gains. TDNet [13] exploits temporal redundancy for fast video segmentation by developing an attention propagation method to propagate features to adjacent frames. Although they achieve promising results on video segmentation, we experimentally demonstrate these designs are not optimal for multi-scan point cloud segmentation. We focus on designing new methods tailored to point clouds for better performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.753, + 0.265, + 0.769 + ], + "angle": 0, + "content": "3. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Given a sequence of LiDAR point clouds as inputs, the multi-scan 3D point cloud semantic segmentation task aims to assign a semantic category to each point and predict their motion states (i.e., moving or static). Specifically, a LiDAR point cloud frame contains a set of unordered points that are annotated with labels for training. We denote a pair of training data as \\((\\mathcal{P}_i,\\mathcal{L}_i) = \\{p_j,l_j\\}_{j = 1}^N\\) with \\(p_j\\in \\mathbb{R}^{D_{\\mathrm{in}}}\\), where \\(N\\) denotes the number of points. Each point \\(p_j\\) contains in-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.895, + 0.631 + ], + "angle": 0, + "content": "put descriptors with \\( D_{\\mathrm{in}} \\) dimensions, including point coordinates \\( (x,y,z) \\) and other features such as intensities (bm). The corresponding label \\( l_{j} \\) incorporates both semantic categories and motion states of \\( p_j \\). Therefore, points belonging to the same semantic category but possessing distinct motion states are allocated distinct labels. For a sequence of point clouds \\( \\{\\mathcal{P}_i,\\mathcal{L}_i\\}_{i = 1}^M \\) that contains \\( M \\) frames, all frames are scanned sequentially in time order, and scanning poses and timestamps are used to align different frames into the same world-coordinate system. In the following, we omit the alignment process for simplicity, where the point clouds in a sequence are calibrated to the same coordinate system by default." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.593, + 0.665 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.675, + 0.611, + 0.69 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.895, + 0.904 + ], + "angle": 0, + "content": "An overview of our motion-aware model for multi-scan semantic segmentation, namely MarS3D, is shown in Figure 2. MarS3D contains a 3D branch for multi-scan spatial representation learning and a BEV branch for motion-aware feature learning. First, the BEV branch takes as inputs \\(k\\) BEV representations \\(\\mathcal{B}^{\\mathrm{in}} = \\left\\{\\mathcal{B}_i^{\\mathrm{in}}\\right\\}_{i=1}^k\\) that are derived by point cloud polarization (see Section 4.2) and outputs motion-aware feature map \\(\\mathcal{Z}^m\\). The core is a Motion-Aware Feature Learning (MAFL) module (see Section 4.2) that extracts and leverages BEV features through a dedicated design to produce a motion-aware feature map \\(\\mathcal{Z}^m\\). Second, the 3D branch takes as inputs the fused \\(k\\) point clouds \\(\\mathcal{P}^{\\mathrm{in}} = \\left\\{\\mathcal{P}_i^{\\mathrm{in}}\\right\\}_{i=1}^k\\) and outputs enriched 3D enhanced spa" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9374" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.187 + ], + "angle": 0, + "content": "tial features \\(\\mathcal{P}^s\\). The 3D branch incorporates a cross-frame feature embedding (CFFE) module (see Section 4.3) to inject temporal information. It outputs embedded features, denoted as \\(\\mathcal{P}^{\\mathrm{ebd}} = \\left\\{\\mathcal{P}_i^{\\mathrm{ebd}}\\right\\}_{i=1}^k\\). These features are further processed by a single-scan backbone \\(f_s\\) to yield multi-scan enhanced 3D spatial representations \\(\\mathcal{P}^s\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.202, + 0.47, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} ^ {s} = f _ {s} \\left(\\mathcal {P} ^ {\\mathrm {e b d}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.259 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}^s\\in \\mathbb{R}^{D_z}\\) and \\(D_{z}\\) is the dimension of the output feature of the single-scan backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.352 + ], + "angle": 0, + "content": "Finally, the motion-aware feature map \\(\\mathcal{Z}^m\\) and enhanced spatial features \\(\\mathcal{P}^s\\) are combined to produce the fused features \\(\\mathcal{P}^f\\) by aligning the coordinates of the features with the pixels of the motion-aware feature map. The fused features are then fed to the prediction head to produce the final outputs (see Section 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.363, + 0.214, + 0.379 + ], + "angle": 0, + "content": "4.2. BEV Branch" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.493 + ], + "angle": 0, + "content": "In the following, we elaborate on the BEV branch, which targets extracting motion-friendly features for motion prediction (see Figure 2). Before delving into the details, we first introduce BEV mapping, which maps a 3D point cloud into a 2D BEV image. Then, we introduce our key Motion-Aware Feature Learning (MAFL) module for motion extraction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.473, + 0.625 + ], + "angle": 0, + "content": "BEV Mapping: Each point cloud \\(\\mathcal{P}\\) for the BEV branch is pillarized into a three-channel BEV map \\(\\mathcal{B}\\) of size \\(H\\times W\\). Inspired by [14, 16], for a pillar located at \\((x^{*},y^{*})\\) with a pixel grid size of \\(l_{B}\\), its three-channel feature \\(b_{(x^{*},y^{*})}\\) consists of the average translation along \\(x\\) and \\(y\\) axis regarding the grid center, and total intensity features (taking SemanticKITTI as an example) among all points inside the pillar, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.65, + 0.469, + 0.704 + ], + "angle": 0, + "content": "\\[\nb _ {\\left(x ^ {*}, y ^ {*}\\right)} = \\left[ \\frac {1}{N ^ {*}} \\sum_ {i} ^ {N ^ {*}} \\frac {2 \\Delta x _ {i}}{l _ {B}}, \\frac {1}{N ^ {*}} \\sum_ {i} ^ {N ^ {*}} \\frac {2 \\Delta y _ {i}}{l _ {B}}, \\sum_ {i} ^ {N ^ {*}} \\mathrm {b m} _ {i} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.751 + ], + "angle": 0, + "content": "where \\(N^{*}\\) denotes the number of points in current pillar, while \\(\\Delta x_{i}\\) and \\(\\Delta y_{i}\\) denote the translation of the ith point along \\(x\\)-axis and \\(y\\)-axis respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Motion-Aware Feature Learning Module: Our Motion-Aware Feature Learning (MAFL) module is illustrated in the BEV branch in Figure 2 which is designed to extract motion-aware features for better distinguishing moving/static objects. The input to this module consists of \\( k \\) multi-scan BEV representations \\( \\mathcal{B}^{\\mathrm{in}} = \\left\\{\\mathcal{B}_i^{\\mathrm{in}}\\right\\}_{i=1}^k \\), where \\( \\mathcal{B}_k^{\\mathrm{in}} \\) represents the target frame. A lightweight 2D CNN \\( f_u \\) with a UNet-like architecture [22] is used to extract features \\( \\mathcal{U} = \\{\\mathcal{U}_i\\}_{i=1}^k \\) from multi-scan inputs as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.107, + 0.893, + 0.131 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathcal {U} _ {i} \\right\\} _ {i = 1} ^ {k} = \\left\\{f _ {u} \\left(\\mathcal {B} _ {i} ^ {\\text {i n}}\\right) \\right\\} _ {i = 1} ^ {k}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.14, + 0.894, + 0.189 + ], + "angle": 0, + "content": "Furthermore, to identify moving objects, we take the difference between the target frame \\(k\\) and the remaining \\(k - 1\\) reference frames which outputs \\(\\mathcal{D} = \\{\\mathcal{D}_i\\}_{i = 1}^{k - 1}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.205, + 0.893, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathcal {D} _ {i} \\right\\} _ {i = 1} ^ {k - 1} = \\left\\{\\mathcal {U} _ {k} - \\mathcal {U} _ {i} \\right\\} _ {i = 1} ^ {k - 1}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.237, + 0.895, + 0.375 + ], + "angle": 0, + "content": "By doing so, the static objects can be erased, and the dynamic objects are highlighted with a large feature magnitude. Then, \\(\\{\\mathcal{D}_i\\}_{i=1}^{k-1}\\) are channel-wise concatenated to form a new 2D map. Note that objects may have different moving patterns and velocities. Therefore, we design a multi-kernel convolutional network \\(f_m\\) with multiple branches of various kernel sizes to capture objects with various movement patterns. Finally, the outputs from \\(f_m\\) are concatenated to output a motion-aware feature map \\(\\mathcal{Z}^m\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.622, + 0.403 + ], + "angle": 0, + "content": "4.3. 3D Branch" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.894, + 0.534 + ], + "angle": 0, + "content": "The 3D branch uses temporal information to enhance spatial representation learning on 3D point clouds. The core component is the Cross-Frame Feature Embedding (CFFE) module, whose output is fed into the single-scan backbone network to produce enhanced spatial features \\(\\mathcal{P}^s\\) (see Figure 2). In the following, we elaborate on the CFFE module to improve spatial representation learning on multi-scan point clouds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.545, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Cross-Frame Feature Embedding Module: When the multi-frame point clouds are fused as discussed in the previous sections, points from different time steps are mixed, making it challenging for the final recognition. Inspired by positional embedding [27], we propose a Cross-Frame Feature Embedding (CFFE) module to generate a time-aware embedding and produce consistent features for each point across different timestamps. Given \\( k \\) point cloud frames, \\( \\{\\mathcal{P}_i^{\\mathrm{in}}\\}_{i = 1}^k \\), we design an embedding neural network layer \\( f_{e} \\) that maps point clouds into intermediate latent features and \\( k \\) learnable temporal embeddings \\( \\mathcal{E} = \\{e_i\\}_{i = 1}^k \\) corresponding to \\( k \\) frames respectively. The dimension of \\( e_i \\) is the same as the output dimension of the \\( f_{e} \\). The point-level embedded features \\( \\mathcal{P}^{\\mathrm{ebd}} \\) are obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.778, + 0.893, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathcal {P} _ {i} ^ {\\mathrm {e b d}} \\right\\} _ {i = 1} ^ {k} = \\left\\{e _ {i} + f _ {e} \\left(\\mathcal {P} _ {i} ^ {\\mathrm {i n}}\\right) \\right\\} _ {i = 1} ^ {k}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.894, + 0.902 + ], + "angle": 0, + "content": "where element-wise summation is conducted on \\( e_i \\) and each point of the corresponding \\( f_e(\\mathcal{P}_i^{\\mathrm{in}}) \\). The obtained features with different temporal embeddings are represented in a point cloud \\( \\{\\mathcal{P}_i^{\\mathrm{ebd}}\\}_{i=1}^k \\) and subsequently fed into the single-scan backbone network \\( f_s \\), following Eq. (1), to produce a set of enhanced spatial features \\( \\mathcal{P}^s \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9375" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.077, + 0.091, + 0.891, + 0.263 + ], + "angle": 0, + "content": "
B.MethodmIoU#paramlatencycarbic.mot.tru.ove.per.bil.mol.roa.par.sid.ogr.bui.fen.veg.trn.ter.pol.tra.mca.mbi.mpe.mmm.mov.mtr.
SPVCNN [25]Baseline49.7021.8M206ms93.934.464.768.033.019.70.00.093.645.280.10.290.359.788.463.575.664.151.974.386.755.00.00.00.0
Ours54.6621.9M225ms95.652.777.879.451.527.90.00.094.251.382.00.191.665.189.068.476.265.251.580.694.968.00.03.60.0
Δ+4.96+0.1M+19ms1.718.313.111.418.58.20.00.00.66.11.90.11.35.40.64.90.61.10.46.38.213.00.03.60.0
SparseConv [9]Baseline48.9939.2M239ms94.724.154.169.643.417.30.20.093.245.179.80.289.561.787.762.974.663.850.073.985.453.60.00.00.0
Ours54.6439.3M253ms96.635.269.083.364.826.90.00.094.061.282.50.190.965.888.167.875.266.251.483.594.468.90.00.00.0
Δ+5.65+0.1M+14ms1.911.114.913.721.49.60.20.00.816.12.70.11.44.10.44.90.62.41.49.69.015.30.00.00.0
MinkUnet [5]Baseline48.4737.9M295ms93.823.748.990.341.318.00.00.092.232.278.40.089.855.588.863.777.063.650.069.283.152.50.00.00.0
Ours54.7138.0M323ms96.428.470.093.962.731.60.00.093.858.581.70.192.667.689.066.776.466.551.682.693.164.40.00.10.0
Δ+6.24+0.1M+28ms2.64.721.13.621.413.60.00.01.626.33.30.12.812.10.23.00.62.91.613.410.011.90.00.10.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.268, + 0.893, + 0.327 + ], + "angle": 0, + "content": "Table 1. Quantitative results of the proposed method, MarS3D, on SemanticKITTI [1] multi-scan public validation set. Combined with different mainstream single-scan 3D point cloud semantic segmentation backbones, MarS3D has a large performance improvement over the corresponding baseline methods without introducing excessive parameters and each-frame inference time. (B. indicates Backbone, full names of the categories are in the supplementary material, blue indicates degradation.)" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.348, + 0.347, + 0.363 + ], + "angle": 0, + "content": "4.4. Feature Fusion and Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.434 + ], + "angle": 0, + "content": "As illustrated in Figure 2, equipped with the motion-aware feature map \\((\\mathcal{Z}^m)\\) and enhanced spatial features \\((\\mathcal{P}^s)\\), the next step is to fully integrate the representation information from both branches and make predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.471, + 0.668 + ], + "angle": 0, + "content": "Feature Fusion: The motion-aware feature map \\(\\mathcal{Z}^m\\in\\) \\(\\mathbb{R}^{H\\times W\\times D_z}\\) ( \\(H\\) : height; \\(W\\) : width; \\(D_{z}\\) : number of channels) is obtained from the BEV branch, while the enhanced spatial features \\(\\mathcal{P}^s\\in \\mathbb{R}^{N\\times D_p}\\) (\\(N\\) : number of points; \\(D_{p}\\) : number of channels) are the outputs of the 3D branch. The feature fusion module aggregates information from the above two representations to make subsequent predictions. For each point feature in \\(\\mathcal{P}^s\\) with \\(D_{p}\\) dimension, a corresponding pixel from \\(\\mathcal{Z}^m\\) can be queried based on the point's 3D location (\\(x\\) and \\(y\\) coordinates). This pixel serves as an index to extract a \\(D_{z}\\)-dimensional motion-aware feature along the channel dimension from \\(\\mathcal{Z}^m\\). Subsequently, the motion-aware feature is concatenated with the point feature, resulting in the fused features denoted as \\(\\mathcal{P}^f\\in \\mathbb{R}^{N\\times (D_z + D_p)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.724 + ], + "angle": 0, + "content": "Prediction: Based on the correspondence between 2D and 3D, \\(\\mathcal{P}^f\\) is then fed into an MLP classifier \\(f_{\\mathrm{cls}}\\) to obtain the output \\(s_{\\mathrm{pred}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.737, + 0.469, + 0.755 + ], + "angle": 0, + "content": "\\[\ns _ {\\text {p r e d}} = f _ {\\mathrm {c l s}} \\left(\\mathcal {P} ^ {f}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.469, + 0.793 + ], + "angle": 0, + "content": "where \\( s_{\\mathrm{pred}} \\) is the logits for the prediction result of the semantic segmentation task of the input point cloud." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.802, + 0.345, + 0.818 + ], + "angle": 0, + "content": "4.5. Model Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.905 + ], + "angle": 0, + "content": "During training, the obtained fused features \\(\\mathcal{P}^f\\) are fed into two classification heads: the category-aware classification head and the motion-aware classification head (a binary classifier), which outputs the predicted logits \\(s_{\\mathrm{pred}}^c\\) and \\(s_{\\mathrm{pred}}^{m}\\) (more details are included in the supplementary mate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.41 + ], + "angle": 0, + "content": "rial). First, with the ground truth labels of semantic categories \\(\\mathcal{L}_{\\mathrm{GT}}^c\\), back-propagation is performed for parameters optimization using Cross Entropy (CE) loss \\(L_{c}\\) for semantic categories classification:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.425, + 0.892, + 0.443 + ], + "angle": 0, + "content": "\\[\nL _ {c} = \\mathrm {C E} \\left(s _ {\\text {p r e d}} ^ {c}, \\mathcal {L} _ {\\mathrm {G T}} ^ {c}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.511 + ], + "angle": 0, + "content": "where \\(\\mathrm{CE}(\\cdot, \\cdot)\\) is the cross-entropy loss, and \\(s_{\\mathrm{pred}}^c\\) is the prediction probability for each point. Then, the Binary Cross Entropy (BCE) loss \\(L_m\\) is used for motion states classification with motion states ground truth \\(\\mathcal{L}_{\\mathrm{GT}}^m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.525, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\\[\nL _ {m} = \\operatorname {B C E} \\left(s _ {\\text {p r e d}} ^ {m}, \\mathcal {L} _ {\\mathrm {G T}} ^ {m}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.549, + 0.867, + 0.564 + ], + "angle": 0, + "content": "where \\(\\mathrm{BCE}(\\cdot ,\\cdot)\\) indicates the binary cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.565, + 0.867, + 0.58 + ], + "angle": 0, + "content": "The final objective function \\(L\\) of the optimization is:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.596, + 0.891, + 0.612 + ], + "angle": 0, + "content": "\\[\nL = \\omega_ {c} \\cdot L _ {c} + \\omega_ {m} \\cdot L _ {m}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.619, + 0.892, + 0.649 + ], + "angle": 0, + "content": "where \\(\\omega_{c}\\) and \\(\\omega_{m}\\) are the weights of the two losses (\\(L_{c}\\) and \\(L_{m}\\)) respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.755 + ], + "angle": 0, + "content": "During inference, the final prediction result is determined using the logits produced by both classification heads. When presented with an input sample, the motion-aware classification head will identify the motion state of the input point only if the category-aware classification head recognizes the point as belonging to a class with the potential to move." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.634, + 0.788 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Datasets and Evaluation Metric: We evaluate our method on SemanticKITTI [1] and nuScenes [3]. For SemanticKITTI, the multi-scan setting is fully supervised and contains 25-category (6 moving categories and 19 static categories) with high-quality semantic annotations. The annotations are based on the KITTI dataset [8]. It comprises 22 point cloud sequences. For nuScenes, we propose a new" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9376" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.073, + 0.09, + 0.353, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.074, + 0.182, + 0.353, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.276, + 0.236, + 0.285 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image_footnote", + "bbox": [ + 0.093, + 0.287, + 0.184, + 0.298 + ], + "angle": 0, + "content": "- Car (non-moving)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.194, + 0.287, + 0.265, + 0.297 + ], + "angle": 0, + "content": "- Car (moving)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.274, + 0.287, + 0.379, + 0.298 + ], + "angle": 0, + "content": "- Person (non-moving)" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.091, + 0.621, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.182, + 0.621, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.275, + 0.474, + 0.285 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_footnote", + "bbox": [ + 0.389, + 0.287, + 0.574, + 0.298 + ], + "angle": 0, + "content": "- Person (moving) - Bicyclist (moving)" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.091, + 0.888, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.182, + 0.888, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.276, + 0.791, + 0.285 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.893, + 0.348 + ], + "angle": 0, + "content": "Figure 3. Qualitative results on SemanticKITTI [1] public validation dataset. With SPVCNN [25] as the backbone, the segmentation results on the SemanticKITTI multi-scan task of the baseline model and our model are shown together with ground truth. At the same time, a specific area containing moving points is magnified and displayed at the top right of each sub-figure." + }, + { + "type": "image", + "bbox": [ + 0.07, + 0.368, + 0.275, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.07, + 0.479, + 0.275, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.591, + 0.212, + 0.601 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.369, + 0.468, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.479, + 0.468, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.591, + 0.389, + 0.6 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.677 + ], + "angle": 0, + "content": "Figure 4. Evaluation errors (shown in red) by the baseline methods and MarS3D (using SPVCNN [25] as the backbone) on SemanticKITTI [1] public validation dataset. MarS3D significantly outperforms the baseline. Contrasting regions in the evaluation errors are highlighted with boxes and corresponding categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.842 + ], + "angle": 0, + "content": "multi-scan setting based on the 'lidar-seg' task (16 semantic categories) without reference frame supervision. We use its object-level velocity to construct a multi-scan segmentation dataset with 24 categories (8 moving and 16 static categories). More details are provided in the supplementary material. To assess the effectiveness of our proposed method and make comparisons with baselines and other methods, we use the mean Intersection over Union (mIoU) as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Implementation Details: Our model is designed as a plug-and-play module that provides motion-aware features to enhance the backbone learned features. For the backbone" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.893, + 0.522 + ], + "angle": 0, + "content": "model, we consider SPVCNN \\([25]^*\\), SparseConv \\([9]^{\\dagger}\\), and MinkUNet \\([5]^{\\ddagger}\\). Following previous works [12, 23, 36], we use the current and its previous two frames as input. The size of the BEV representation is set to \\(501 \\times 301\\), and the multiple kernel sizes in the MAFL module are set as 1, 3, and 5, respectively. We set the embedded feature dimension in the CFFE module to 18, and the data augmentations are the same as the standard settings. All the models are trained on GeForce RTX 3090 GPUs, and the inference latency is recorded using a single GeForce RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.638, + 0.544 + ], + "angle": 0, + "content": "5.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.856 + ], + "angle": 0, + "content": "Comparison with Baseline Methods: We evaluate the performance of our proposed method on multi-scan benchmarks of SemanticKITTI [1] and nuScenes [3]. The baseline method uses the same backbone to process multiscan point clouds. We then compare this baseline approach to the same backbone augmented with MarS3D for a fair evaluation. As shown in Table 1, MarS3D significantly improves the performance over baseline methods on the public validation set of SemanticKITTI. With the most lightweight network, SPVCNN [25], MarS3D brings a \\(4.96\\%\\) improvement while introducing less than \\(0.5\\%\\) additional parameters. Particularly, consistent performance gains are observed on the dynamic object classes with non-moving/moving properties (i.e., car & moving car, person & moving person). This shows that the proposed BEV branch is both lightweight and powerful. Further, to verify the generalizability of the model, we offer a multiscan task based on nuScenes 'lidar-seg' dataset [3]. Our method outperforms the baseline (using MinkUNet [5] as the backbone) by a significant margin, achieving \\(64.83\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.863, + 0.821, + 0.876 + ], + "angle": 0, + "content": "\\*https://github.com/mit-han-lab/spvnas" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.877, + 0.819, + 0.888 + ], + "angle": 0, + "content": "https://github.com/traveller59/spconv" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.889, + 0.848, + 0.9 + ], + "angle": 0, + "content": "\\(\\S\\)https://github.com/NVIDIA/MinkowskiEngine" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.863, + 0.848, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9377" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.091, + 0.891, + 0.235 + ], + "angle": 0, + "content": "
MethodmIoUcarbic.mot.tru.ove.per.bil.mol.roa.par.sid.ogr.bui.fen.veg.trn.ter.pol.tra.mca.mbi.mpe.mmm.mov.mtr.
SpSequenceNet [24]43.188.524.026.229.222.76.30.00.090.157.973.927.191.266.884.066.065.750.848.753.241.226.236.22.30.1
TemporalLidarSeg [7]47.092.147.740.939.235.014.40.00.091.859.675.823.289.863.882.362.564.752.660.468.242.840.412.912.42.1
TemporalLatticeNet [23]47.191.635.436.126.923.09.40.00.091.559.375.327.589.665.384.666.770.457.260.459.741.751.048.85.90.0
Meta-RangeSeg [28]49.590.152.743.930.335.414.30.00.090.763.374.726.990.563.583.067.067.756.464.464.556.155.024.420.33.4
KPConv [26]51.293.744.947.243.538.621.60.00.086.558.470.526.790.864.584.670.366.057.053.969.467.467.547.24.75.8
Baseline49.289.839.434.039.421.08.91.80.089.162.072.412.990.563.984.668.468.758.960.169.363.558.756.59.53.6
Ours52.795.149.249.539.736.616.21.20.089.966.874.326.492.168.286.072.170.562.864.878.467.358.036.310.05.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.238, + 0.893, + 0.281 + ], + "angle": 0, + "content": "Table 2. Comparison with the state-of-the-art models on SemanticKITTI multi-scan benchmark (official test set). MarS3D (with SPVCNN [25] as the backbone) significantly outperforms these models for multi-scan tasks. (full names of the categories are in the supplementary material.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.47, + 0.348 + ], + "angle": 0, + "content": "mIoU compared to the baseline's \\(61.90\\%\\). This improvement is achieved with a negligible increase in each-frame inference time from 53ms to 58ms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Comparison with State-of-the-Art Methods: Compared to various models applied to multi-scan tasks, MarS3D (with SPVCNN as the backbone) is evaluated on the SemanticKITTI multi-scan benchmark. As shown in Table 2, the proposed approach has demonstrated superior performance, with a \\(1.5\\%\\) increase in mIoU compared to the current state-of-the-art method, KPCov [26]. Furthermore, our method performs similarly or better than other state-of-the-art models across nearly all categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.745 + ], + "angle": 0, + "content": "Qualitative Comparisons: Quantitative results are shown in Figure 3. The baseline model suffers from mistaking the status of static cars (Figure 3: Scan A) and moving persons (Figure 3: Scan B). It even recognizes the moving bicyclist as a moving person (Figure 3: Scan A). In contrast, MarS3D can circumvent such category and motion state discrimination errors. In addition to achieving better qualitative results in motion states, MarS3D also outperforms the baseline in semantic categories. We visualize the error maps (errors are shown in red) of the baseline and our method in Figure 4, where bounding boxes indicating specific categories highlight the differences between our method and the baseline. These improvements indicate that MarS3D has stronger semantic feature extraction capabilities compared to the baseline model, resulting in better segmentation performance even for immobile objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.753, + 0.242, + 0.768 + ], + "angle": 0, + "content": "5.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.776, + 0.469, + 0.868 + ], + "angle": 0, + "content": "In this section, we conduct comprehensive ablation experiments on the SemanticKITTI multi-scan validation set to examine the effects of each component in our proposed method. As shown in Table 3, we gradually add three components to the baseline method, including the CFFE module, vanilla BEV branch (introduce only BEV representa" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.303, + 0.683, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.408, + 0.682, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.512, + 0.676, + 0.523 + ], + "angle": 0, + "content": "BEV Ground Truth (Motion)" + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.302, + 0.845, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.408, + 0.844, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.512, + 0.778, + 0.522 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.525, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Figure 5. Sampled discrepant feature maps demonstrate that the region that contains moving points are represented higher activation values than other regions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.591, + 0.891, + 0.622 + ], + "angle": 0, + "content": "tions and 2D CNN without the MAFL module), and the MAFL module to illustrate the effectiveness of our designs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.633, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Effectiveness of CFFE Module: As shown in Table 3, we first employ only the CFFE module with the baseline model and observe a significant performance improvement. For instance, with SPVCNN [25] as the backbone, the introduction of the CFFE module led to a boost in performance by \\(1.1\\%\\) in terms of mIoU. This confirms the efficacy of the CFFE module for multi-frame representation learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of BEV Representation: We also observe a significant performance improvement when we add the vanilla BEV branch alone to the baseline model. Using SPVCNN [25] as the backbone, the BEV branch results in a \\(3.7\\%\\) increase in mIoU. This confirms the importance of BEV representations in extracting motion-aware information. Furthermore, when both the CFFE module and the vanilla BEV branch are added to the baseline model, there is a further increase in performance, demonstrating that the two components complement each other." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\(\\S\\) http://wwwsemantic-kitti.org/tasks - Semantic Segmentation - Multiple Scans" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9378" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.091, + 0.466, + 0.202 + ], + "angle": 0, + "content": "
MethodCFFEBEVMAFLmIoU(%)
baseline---49.7 / 49.0 / 48.5
proposed✓--50.8 / 51.3 / 52.1
-✓-53.4 / 53.9 / 53.1
✓✓-53.6 / 54.2 / 54.1
✓✓✓54.7 / 54.6 / 54.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.207, + 0.47, + 0.263 + ], + "angle": 0, + "content": "Table 3. Ablation studies on different backbones (SPVCNN / SparseConv / MinkUNet) on SemanticKITTI [1] public validation dataset. The effectiveness of different designs is demonstrated step-by-step and our method is marked with a gray box." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.274, + 0.465, + 0.308 + ], + "angle": 0, + "content": "
MethodVanilla BEVTDNet [13]STM [18]Ours
mIoU(%)53.3653.56 (+0.20)53.64 (+0.28)54.66 (+1.30)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.312, + 0.468, + 0.355 + ], + "angle": 0, + "content": "Table 4. Comparing our method (using SPVCNN [25] as the backbone) with different 2D temporal semantic segmentation approaches the public validation dataset of SemanticKITTI [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.712 + ], + "angle": 0, + "content": "Effectiveness of MAFL Module: In the final experiment presented in Table 3, we study the impact of our proposed MAFL module on the BEV representations, resulting in an increase of around \\(5.0\\%\\) from the baseline model (using SPVCNN [25] as the backbone). By stacking all the proposed components together, our final solution (marked with a grey box in Table 3) reaches its pinnacle in performance. The improvement in other backbones is also significant, and more statistical results are provided in the supplementary material. In addition, the multi-channel discrepant feature map represents the differences between the extracted feature maps of BEV representations from two point clouds. For a discrepant feature map generated during inference, we randomly sample the channels pixel by pixel. The average absolute activation values of pixels with the same motion state in the sampled channels are re-scaled and aggregated into a single-channel feature map as shown in Figure 5. The sampled discrepant feature map has higher activation on pixels containing moving points compared to other regions. This clearly demonstrates that the MAFL module clearly distinguishes between regions that contain moving points and those that do not." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.466, + 0.742 + ], + "angle": 0, + "content": "5.3. More Comparison on Temporal Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As for BEV representation learning, the proposed MAFL module is utilized on extracted 2D feature maps for capturing inter-frame temporal information of point clouds. We further conduct comparative experiments by replacing the MAFL module with its 2D counterparts, i.e., STM [18] and TDNet [13]. These 2D modules have been widely utilized for preserving temporal information in 2D tasks. The results of the comparative experiments using SPVCNN [25] as the backbone are presented in Table 4. According to the evaluation results, it can be concluded that the MAFL module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "shows superior performance compared to the other models, thereby confirming its remarkable effectiveness. This shows that the proposed MAFL module is better suited for the specific task of handling motion-aware semantic segmentation in 3D point clouds." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.18, + 0.836, + 0.196 + ], + "angle": 0, + "content": "5.4. Limitations and Failure Cases Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.204, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Although MarS3D demonstrates impressive overall performance on the SemanticKITTI multi-scan benchmark, some limitations can be identified from the quantitative results in Table 1 and Table 2. The SemanticKITTI training dataset has an imbalanced distribution of point categories (more details are included in the supplementary material), which causes MarS3D to perform poorly on long-tailed categories due to insufficient training data on these categories. As shown in Table 2, MarS3D fails to show effects on several long-tail categories (i.e., bicyclist, motorcyclist, moving-other-vehicle, and moving-truck). Exploring solutions to address the long-tail problem is a promising research direction for the work. Since our model assumes planar motion, it may not perform well in scenarios where objects move in non-planar ways, such as on steep terrains or on non-planar surfaces." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.618, + 0.479 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.777 + ], + "angle": 0, + "content": "In this paper, we propose MarS3D, a novel plug-and-play motion-aware model for 3D multi-scan point cloud semantic segmentation. The Motion-Aware Feature Learning (MAFL) module, based on BEV representations, is designed to facilitate extracting motion-aware representations. Additionally, the Cross-Frame Feature Embedding (CFFE) module is introduced to improve representation learning by embedding time-step information into features, thus preserving rich temporal information. Extensive experiments and ablation studies demonstrate that MarS3D significantly improves multiple 3D semantic segmentation baselines while introducing minimal overheads. In comparison to state-of-the-art methods designed for multi-scan tasks, MarS3D achieves superior performance and offers significant improvements over baseline methods. The proposed MarS3D model demonstrates the potential for effectively incorporating motion-awareness into 3D point cloud semantic segmentation tasks, providing a strong foundation for further research and development in this area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant no. 17202422), and RGC matching fund scheme (RMGS). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9379" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF International Conf. on Computer Vision (ICCV), 2019. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Goutam Bhat, Felix Järemo Lawin, Martin Danelljan, Andreas Robinson, Michael Felsberg, Luc Van Gool, and Radu Timofte. Learning what to learn for video object segmentation. In European Conference on Computer Vision, pages 777-794. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.472, + 0.338 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[4] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.492 + ], + "angle": 0, + "content": "[5] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[6] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.632 + ], + "angle": 0, + "content": "[7] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 2020 International Conference on 3D Vision (3DV), pages 781-790. IEEE, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[8] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.69, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[10] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[11] Noureldin Hendy, Cooper Sloan, Feng Tian, Pengfei Duan, Nick Charchut, Yuesong Xie, Chuang Wang, and James Philbin. Fishing net: Future inference of semantic heatmaps in grids. arXiv preprint arXiv:2006.09917, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In Proceedings of the IEEE/CVF Con" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 8479-8488, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[13] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks for fast video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8827, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[15] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In European Conference on Computer Vision, pages 352-368. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[16] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[17] Jiaxu Miao, Yunchao Wei, and Yi Yang. Memory aggregation networks for efficient interactive video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10366-10375, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[18] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226–9235, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[19] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[20] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[21] Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7406-7415, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[22] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[23] Peer Schutt, Radu Alexandru Rosu, and Sven Behnke. Abstract flow for temporal semantic segmentation on the permutohedral lattice. In 2022 International Conference on Robotics and Automation (ICRA), pages 5139-5145. IEEE, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Hanyu Shi, Guosheng Lin, Hao Wang, Tzu-Yi Hung, and Zhenhua Wang. Spsequencenet: Semantic segmentation network on 4d point clouds. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9380" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "conference on computer vision and pattern recognition, pages 4574-4583, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[25] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[26] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[28] Song Wang, Jianke Zhu, and Ruixiang Zhang. Meta-rangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. arXiv preprint arXiv:2202.13377, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[29] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[30] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9621-9630, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[31] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[32] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3173-3182, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[33] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3101-3109, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[34] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[35] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[36] Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[37] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[38] Zhenhong Zou and Yizhe Li. Efficient urban-scale point clouds segmentation with bev projection. arXiv preprint arXiv:2109.09074, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9381" + } + ] +] \ No newline at end of file diff --git a/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_origin.pdf b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3923a219b985b2b4d2bec848efc7c304a91c7dc7 --- /dev/null +++ b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/a0a4e92d-730a-4b0b-bb00-036e1af8bb9b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c27854787385bb2618cf9aab8ce5b847a3abbb8f5a455ea2cbcd3721da8cd7fc +size 8254888 diff --git a/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/full.md b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0110a7078f3ee0baa279018e0203c766cd8829a0 --- /dev/null +++ b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/full.md @@ -0,0 +1,320 @@ +# MarS3D: A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds + +Jiahui Liu $^{1*}$ Chirui Chang $^{1*}$ Jianhui Liu $^{1}$ Xiaoyang Wu $^{1}$ Lan Ma $^{2}$ Xiaojuan Qi $^{1\dagger}$ $^{1}$ The University of Hong Kong $^{2}$ TCL AI Lab + +{liujh, xjqi}@eee.hku.hk, crchang@hku.hk, jhliu0212@gmail.com, xywu3@cs.hku.hk, rubyma@tcl.com + +# Abstract + +3D semantic segmentation on multi-scan large-scale point clouds plays an important role in autonomous systems. Unlike the single-scan-based semantic segmentation task, this task requires distinguishing the motion states of points in addition to their semantic categories. However, methods designed for single-scan-based segmentation tasks perform poorly on the multi-scan task due to the lacking of an effective way to integrate temporal information. We propose MarS3D, a plug-and-play motion-aware module for semantic segmentation on multi-scan 3D point clouds. This module can be flexibly combined with single-scan models to allow them to have multi-scan perception abilities. The model encompasses two key designs: the Cross-Frame Feature Embedding module for enriching representation learning and the Motion-Aware Feature Learning module for enhancing motion awareness. Extensive experiments show that MarS3D can improve the performance of the baseline model by a large margin. The code is available at https://github.com/CVMI-Lab/MarS3D. + +# 1. Introduction + +3D semantic segmentation on multi-scan large-scale point clouds is a fundamental computer vision task that benefits many downstream problems in autonomous systems, such as decision-making, motion planning, and 3D reconstruction, to name just a few. Compared with the single-scan semantic segmentation task, this task requires understanding not only the semantic categories but also the motion states (e.g., moving or static) of points based on multi-scan point cloud data. + +In the past few years, extensive research has been conducted on single-scan semantic segmentation with significant research advancements [4, 5, 12, 25, 31, 33, 36]. These approaches are also applied to process multi-scan point + +![](images/f9d09b39f4540770679c524d7c0977eaec2b1b1ac17ef0fe847200928bce8d5e.jpg) + +![](images/6b4cd622f2f091fa38408dd9e49b0ecf4b1844aeabb9a3faf8c869c65d7230dd.jpg) + +![](images/98afb6d81e5b1ed5376fdb0dba48dd8b30b7f108d1dcb422584c9fab0f357aef.jpg) + +![](images/6504b14347e31fa8590b698fda50d81fb4c21fdfc69a98b56425a4b8ebea1119.jpg) + +![](images/c5aae9841f52b11b350a1fa821a1a310293e330f6794d9ab896e93e849cbc05c.jpg) +Figure 1. Comparison of our proposed method, MarS3D, with baseline method using SPVCNN [25] as the backbone on SemanticKITTI [1] dataset. MarS3D achieves excellent results in the classification of semantic categories and motion states, while the baseline method can not distinguish motion well from static. + +![](images/fecbff26a11e4b7608fb3cdcb2c0e6ed2750c36801c1131ee3738b396d4d0c94.jpg) + +clouds, wherein multiple point clouds are fused to form a single point cloud before being fed to the network for processing. Albeit simple, this strategy may lose temporal information and make distinguishing motion states a challenging problem. As a result, they perform poorly in classifying the motion states of objects. As shown in Figure 1, the simple point cloud fusion strategy cannot effectively enable the model to distinguish the motion states of cars even with a state-of-the-art backbone network SPVCNN [25]. Recently, there have been some early attempts [7, 23, 24, 28] to employ attention modules [24] and recurrent networks [7, 23, 28] to fuse information across different temporal frames. However, these approaches do not perform well on the multi-scan task due to the insufficiency of temporal representations and the limited feature extraction ability of the model. + +In sum, a systematic investigation of utilizing the rich spatial-temporal information from multiple-point cloud scans is still lacking. This requires answering two critical questions: (1) how can we leverage the multi-scan information to improve representation learning on point clouds for + +better semantic understanding? and (2) how can the temporal information be effectively extracted and learned for classifying the motion states of objects? + +In this paper, we propose a simple plug-and-play Motion-aware Segmentation module for 3D multi-scan analysis (MarS3D), which can seamlessly integrate with existing single-scan semantic segmentation models and endow them with the ability to perform accurate multi-scan 3D point cloud semantic segmentation with negligible computational costs. Specifically, our method incorporates two core designs: First, to enrich representation learning of multi-frame point clouds, we propose a Cross-Frame Feature Embedding (CFFE) module which embeds time-step information into features to facilitate inter-frame fusion and representation learning. Second, inspired by the observation that objects primarily move along the horizontal ground plane (i.e., $xy$ -plane) in large-scale outdoor scenes, i.e., minimal motion along the $z$ -axis, we propose a Motion-Aware Feature Learning (MAFL) module based on Bird's Eye View (BEV), which learns the motion patterns of objects between frames to facilitate effectively discriminating the motion states of objects. + +We extensively evaluate our approach upon several mainstream baseline frameworks on SemanticKITTI [1] and nuScenes [3] dataset. It consistently improves the performance of the baseline approaches, e.g., MinkUnet [5], by $6.24\%$ in mIoU on SemanticKITTI with a negligible increase in model parameters, i.e., about $0.2\%$ . The main contributions are summarized as follows: + +- We are the first to propose a plug-and-play module for large-scale multi-scan 3D semantic segmentation, which can be flexibly integrated with mainstream single-scan segmentation models without incurring too much cost. +- We devise a Cross-Frame Feature Embedding module to fuse multiple point clouds while preserving their temporal information, thereby enriching representation learning for multi-scan point clouds. +- We introduce a BEV-based Motion-Aware Feature Learning module to exploit temporal information and enhance the model's motion awareness, facilitating the prediction of motion states. +- We conduct extensive experiments and comprehensive analyses of our approach with different backbone models. The proposed model performs favorably compared to the baseline methods while introducing negligible extra parameters and inference time. + +# 2. Related Work + +Single-scan Outdoor 3D Semantic Segmentation: 3D single-scan outdoor semantic segmentation is indispensable for autonomous driving. In early work, PointNet [19] uses + +Multi-Layer Perception (MLP) to extract features from input point clouds directly, and PointNet++ [20] tries to incorporate multi-scale designs for dense prediction tasks. Later, various literature [19, 20, 26, 29, 30, 32] works on designing point-based convolution on either geometric or semantic neighborhoods. To handle the large-scale dataset, some works [4, 5, 10, 12, 25, 31, 36] focus on volumetric features and use 3D convolution to achieve a balance between accuracy and efficiency. SparseConv [10] and MinkUNet [5] are representative works and demonstrate good performance. Later, SPVNAS [25] combines voxel and point representations and designs a neural architecture search method to find the optimal model structure. Recently, Cylinder3D [36] introduces a cylindrical partition to leverage the properties of LiDAR point clouds for enriching the feature information. + +The remarkable feature extraction capability enables the above methods to achieve high performance on single-scan tasks. To solve the multi-scan task, most of these methods [4, 12, 31, 34, 36] first fuse multiple point clouds into one and treat the fused point cloud as a single point cloud for processing. Albeit simple, this fusion strategy overlooks important temporal information and entangles moving and non-moving objects, leading to performance degradation. + +Multi-scan Outdoor 3D Semantic Segmentation: Compared to single-scan semantic segmentation, the multi-scan task needs to discriminate the moving and stationary states of the objects based on temporal information. In addition to the simple fusion strategy discussed above, another stream of approaches [7,23,24,28] attempts to process each point cloud in a sequence separately and fuse the feature representations for temporal modeling. For instance, SpSequenceNet [24] proposes a U-Net-based architecture to extract per-frame features and combine features of two consecutive frames to gather temporal information. The fused feature is further fed into the prediction head to produce results. Duerr et al. [7] design a recurrent architecture with a temporal memory alignment module for sequential processing of multiple point clouds. TemporalLatticeNet [23] proposes to match similar feature patterns between adjacent frames and fuse them temporally. However, these approaches cannot fully leverage multiple point clouds to enrich temporal representation learning as the feature extraction is still conducted on each frame separately. + +BEV-based 3D Point Cloud Perception: Recently, BEV-based representation [11, 14, 16, 35, 37, 38] has emerged as an effective way to process 3D point cloud due to its efficiency and ease of deployment using 2D convolution operations. By projecting a 3D point cloud into the bird's eye view (BEV), BEV-based representation converts 3D representations into 2D to avoid heavy processing in 3D and improve computation and memory efficiency. Some representative works include PointPillars [14] for 3D object de + +![](images/ad5b1b54236d4f4bc8b060629505e83f9ffd38f0e1446825bcd3f29ed3788137.jpg) +Figure 2. The proposed framework of MarS3D. We take three scan point clouds inputs as illustration. MarS3D contains two branches. One is the BEV branch with 2D BEV representations as input, and it employs a Motion-Aware Feature Learning (MAFL) module to enhance the motion-aware feature learning. The other is the 3D branch that takes multi-scan 3D point clouds as input, enriching the feature representations with our Cross-Frame Feature Embedding (CFFE) module. The fused features of the above two modules are fed into the prediction head and make final predictions on semantic categories and motion states. + +tection, BEVfusion [16] for multi-sensor fusion, BEV projection [38] for 3D segmentation. The above work demonstrates the potential of BEV representations in outdoor 3D scene analysis. Here, we explore utilizing BEV-based representation to extract temporal information for analyzing the motion states of objects. + +2D Video Semantic Segmentation: In video segmentation, many efforts [2,6,13,15,17,18,21] have been made to conduct temporal feature extraction. Oh et al. [18] introduce a space-time memory network to integrate features from adjacent frames for video object segmentation and attains significant performance gains. TDNet [13] exploits temporal redundancy for fast video segmentation by developing an attention propagation method to propagate features to adjacent frames. Although they achieve promising results on video segmentation, we experimentally demonstrate these designs are not optimal for multi-scan point cloud segmentation. We focus on designing new methods tailored to point clouds for better performance and efficiency. + +# 3. Problem Statement + +Given a sequence of LiDAR point clouds as inputs, the multi-scan 3D point cloud semantic segmentation task aims to assign a semantic category to each point and predict their motion states (i.e., moving or static). Specifically, a LiDAR point cloud frame contains a set of unordered points that are annotated with labels for training. We denote a pair of training data as $(\mathcal{P}_i,\mathcal{L}_i) = \{p_j,l_j\}_{j = 1}^N$ with $p_j\in \mathbb{R}^{D_{\mathrm{in}}}$ , where $N$ denotes the number of points. Each point $p_j$ contains in- + +put descriptors with $D_{\mathrm{in}}$ dimensions, including point coordinates $(x,y,z)$ and other features such as intensities (bm). The corresponding label $l_{j}$ incorporates both semantic categories and motion states of $p_j$ . Therefore, points belonging to the same semantic category but possessing distinct motion states are allocated distinct labels. For a sequence of point clouds $\{\mathcal{P}_i,\mathcal{L}_i\}_{i = 1}^M$ that contains $M$ frames, all frames are scanned sequentially in time order, and scanning poses and timestamps are used to align different frames into the same world-coordinate system. In the following, we omit the alignment process for simplicity, where the point clouds in a sequence are calibrated to the same coordinate system by default. + +# 4. Method + +# 4.1. Overview + +An overview of our motion-aware model for multi-scan semantic segmentation, namely MarS3D, is shown in Figure 2. MarS3D contains a 3D branch for multi-scan spatial representation learning and a BEV branch for motion-aware feature learning. First, the BEV branch takes as inputs $k$ BEV representations $\mathcal{B}^{\mathrm{in}} = \left\{\mathcal{B}_i^{\mathrm{in}}\right\}_{i=1}^k$ that are derived by point cloud polarization (see Section 4.2) and outputs motion-aware feature map $\mathcal{Z}^m$ . The core is a Motion-Aware Feature Learning (MAFL) module (see Section 4.2) that extracts and leverages BEV features through a dedicated design to produce a motion-aware feature map $\mathcal{Z}^m$ . Second, the 3D branch takes as inputs the fused $k$ point clouds $\mathcal{P}^{\mathrm{in}} = \left\{\mathcal{P}_i^{\mathrm{in}}\right\}_{i=1}^k$ and outputs enriched 3D enhanced spa + +tial features $\mathcal{P}^s$ . The 3D branch incorporates a cross-frame feature embedding (CFFE) module (see Section 4.3) to inject temporal information. It outputs embedded features, denoted as $\mathcal{P}^{\mathrm{ebd}} = \left\{\mathcal{P}_i^{\mathrm{ebd}}\right\}_{i=1}^k$ . These features are further processed by a single-scan backbone $f_s$ to yield multi-scan enhanced 3D spatial representations $\mathcal{P}^s$ as: + +$$ +\mathcal {P} ^ {s} = f _ {s} \left(\mathcal {P} ^ {\mathrm {e b d}}\right), \tag {1} +$$ + +where $\mathcal{P}^s\in \mathbb{R}^{D_z}$ and $D_{z}$ is the dimension of the output feature of the single-scan backbone. + +Finally, the motion-aware feature map $\mathcal{Z}^m$ and enhanced spatial features $\mathcal{P}^s$ are combined to produce the fused features $\mathcal{P}^f$ by aligning the coordinates of the features with the pixels of the motion-aware feature map. The fused features are then fed to the prediction head to produce the final outputs (see Section 4.4). + +# 4.2. BEV Branch + +In the following, we elaborate on the BEV branch, which targets extracting motion-friendly features for motion prediction (see Figure 2). Before delving into the details, we first introduce BEV mapping, which maps a 3D point cloud into a 2D BEV image. Then, we introduce our key Motion-Aware Feature Learning (MAFL) module for motion extraction. + +BEV Mapping: Each point cloud $\mathcal{P}$ for the BEV branch is pillarized into a three-channel BEV map $\mathcal{B}$ of size $H\times W$ . Inspired by [14, 16], for a pillar located at $(x^{*},y^{*})$ with a pixel grid size of $l_{B}$ , its three-channel feature $b_{(x^{*},y^{*})}$ consists of the average translation along $x$ and $y$ axis regarding the grid center, and total intensity features (taking SemanticKITTI as an example) among all points inside the pillar, which is formulated as: + +$$ +b _ {\left(x ^ {*}, y ^ {*}\right)} = \left[ \frac {1}{N ^ {*}} \sum_ {i} ^ {N ^ {*}} \frac {2 \Delta x _ {i}}{l _ {B}}, \frac {1}{N ^ {*}} \sum_ {i} ^ {N ^ {*}} \frac {2 \Delta y _ {i}}{l _ {B}}, \sum_ {i} ^ {N ^ {*}} \mathrm {b m} _ {i} \right], \tag {2} +$$ + +where $N^{*}$ denotes the number of points in current pillar, while $\Delta x_{i}$ and $\Delta y_{i}$ denote the translation of the ith point along $x$ -axis and $y$ -axis respectively. + +Motion-Aware Feature Learning Module: Our Motion-Aware Feature Learning (MAFL) module is illustrated in the BEV branch in Figure 2 which is designed to extract motion-aware features for better distinguishing moving/static objects. The input to this module consists of $k$ multi-scan BEV representations $\mathcal{B}^{\mathrm{in}} = \left\{\mathcal{B}_i^{\mathrm{in}}\right\}_{i=1}^k$ , where $\mathcal{B}_k^{\mathrm{in}}$ represents the target frame. A lightweight 2D CNN $f_u$ with a UNet-like architecture [22] is used to extract features $\mathcal{U} = \{\mathcal{U}_i\}_{i=1}^k$ from multi-scan inputs as: + +$$ +\left\{\mathcal {U} _ {i} \right\} _ {i = 1} ^ {k} = \left\{f _ {u} \left(\mathcal {B} _ {i} ^ {\text {i n}}\right) \right\} _ {i = 1} ^ {k}. \tag {3} +$$ + +Furthermore, to identify moving objects, we take the difference between the target frame $k$ and the remaining $k - 1$ reference frames which outputs $\mathcal{D} = \{\mathcal{D}_i\}_{i = 1}^{k - 1}$ as: + +$$ +\left\{\mathcal {D} _ {i} \right\} _ {i = 1} ^ {k - 1} = \left\{\mathcal {U} _ {k} - \mathcal {U} _ {i} \right\} _ {i = 1} ^ {k - 1}. \tag {4} +$$ + +By doing so, the static objects can be erased, and the dynamic objects are highlighted with a large feature magnitude. Then, $\{\mathcal{D}_i\}_{i=1}^{k-1}$ are channel-wise concatenated to form a new 2D map. Note that objects may have different moving patterns and velocities. Therefore, we design a multi-kernel convolutional network $f_m$ with multiple branches of various kernel sizes to capture objects with various movement patterns. Finally, the outputs from $f_m$ are concatenated to output a motion-aware feature map $\mathcal{Z}^m$ . + +# 4.3. 3D Branch + +The 3D branch uses temporal information to enhance spatial representation learning on 3D point clouds. The core component is the Cross-Frame Feature Embedding (CFFE) module, whose output is fed into the single-scan backbone network to produce enhanced spatial features $\mathcal{P}^s$ (see Figure 2). In the following, we elaborate on the CFFE module to improve spatial representation learning on multi-scan point clouds. + +Cross-Frame Feature Embedding Module: When the multi-frame point clouds are fused as discussed in the previous sections, points from different time steps are mixed, making it challenging for the final recognition. Inspired by positional embedding [27], we propose a Cross-Frame Feature Embedding (CFFE) module to generate a time-aware embedding and produce consistent features for each point across different timestamps. Given $k$ point cloud frames, $\{\mathcal{P}_i^{\mathrm{in}}\}_{i = 1}^k$ , we design an embedding neural network layer $f_{e}$ that maps point clouds into intermediate latent features and $k$ learnable temporal embeddings $\mathcal{E} = \{e_i\}_{i = 1}^k$ corresponding to $k$ frames respectively. The dimension of $e_i$ is the same as the output dimension of the $f_{e}$ . The point-level embedded features $\mathcal{P}^{\mathrm{ebd}}$ are obtained by: + +$$ +\left\{\mathcal {P} _ {i} ^ {\mathrm {e b d}} \right\} _ {i = 1} ^ {k} = \left\{e _ {i} + f _ {e} \left(\mathcal {P} _ {i} ^ {\mathrm {i n}}\right) \right\} _ {i = 1} ^ {k}, \tag {5} +$$ + +where element-wise summation is conducted on $e_i$ and each point of the corresponding $f_e(\mathcal{P}_i^{\mathrm{in}})$ . The obtained features with different temporal embeddings are represented in a point cloud $\{\mathcal{P}_i^{\mathrm{ebd}}\}_{i=1}^k$ and subsequently fed into the single-scan backbone network $f_s$ , following Eq. (1), to produce a set of enhanced spatial features $\mathcal{P}^s$ . + +
B.MethodmIoU#paramlatencycarbic.mot.tru.ove.per.bil.mol.roa.par.sid.ogr.bui.fen.veg.trn.ter.pol.tra.mca.mbi.mpe.mmm.mov.mtr.
SPVCNN [25]Baseline49.7021.8M206ms93.934.464.768.033.019.70.00.093.645.280.10.290.359.788.463.575.664.151.974.386.755.00.00.00.0
Ours54.6621.9M225ms95.652.777.879.451.527.90.00.094.251.382.00.191.665.189.068.476.265.251.580.694.968.00.03.60.0
Δ+4.96+0.1M+19ms1.718.313.111.418.58.20.00.00.66.11.90.11.35.40.64.90.61.10.46.38.213.00.03.60.0
SparseConv [9]Baseline48.9939.2M239ms94.724.154.169.643.417.30.20.093.245.179.80.289.561.787.762.974.663.850.073.985.453.60.00.00.0
Ours54.6439.3M253ms96.635.269.083.364.826.90.00.094.061.282.50.190.965.888.167.875.266.251.483.594.468.90.00.00.0
Δ+5.65+0.1M+14ms1.911.114.913.721.49.60.20.00.816.12.70.11.44.10.44.90.62.41.49.69.015.30.00.00.0
MinkUnet [5]Baseline48.4737.9M295ms93.823.748.990.341.318.00.00.092.232.278.40.089.855.588.863.777.063.650.069.283.152.50.00.00.0
Ours54.7138.0M323ms96.428.470.093.962.731.60.00.093.858.581.70.192.667.689.066.776.466.551.682.693.164.40.00.10.0
Δ+6.24+0.1M+28ms2.64.721.13.621.413.60.00.01.626.33.30.12.812.10.23.00.62.91.613.410.011.90.00.10.0
+ +Table 1. Quantitative results of the proposed method, MarS3D, on SemanticKITTI [1] multi-scan public validation set. Combined with different mainstream single-scan 3D point cloud semantic segmentation backbones, MarS3D has a large performance improvement over the corresponding baseline methods without introducing excessive parameters and each-frame inference time. (B. indicates Backbone, full names of the categories are in the supplementary material, blue indicates degradation.) + +# 4.4. Feature Fusion and Prediction + +As illustrated in Figure 2, equipped with the motion-aware feature map $(\mathcal{Z}^m)$ and enhanced spatial features $(\mathcal{P}^s)$ , the next step is to fully integrate the representation information from both branches and make predictions. + +Feature Fusion: The motion-aware feature map $\mathcal{Z}^m\in$ $\mathbb{R}^{H\times W\times D_z}$ ( $H$ : height; $W$ : width; $D_{z}$ : number of channels) is obtained from the BEV branch, while the enhanced spatial features $\mathcal{P}^s\in \mathbb{R}^{N\times D_p}$ ( $N$ : number of points; $D_{p}$ : number of channels) are the outputs of the 3D branch. The feature fusion module aggregates information from the above two representations to make subsequent predictions. For each point feature in $\mathcal{P}^s$ with $D_{p}$ dimension, a corresponding pixel from $\mathcal{Z}^m$ can be queried based on the point's 3D location ( $x$ and $y$ coordinates). This pixel serves as an index to extract a $D_{z}$ -dimensional motion-aware feature along the channel dimension from $\mathcal{Z}^m$ . Subsequently, the motion-aware feature is concatenated with the point feature, resulting in the fused features denoted as $\mathcal{P}^f\in \mathbb{R}^{N\times (D_z + D_p)}$ . + +Prediction: Based on the correspondence between 2D and 3D, $\mathcal{P}^f$ is then fed into an MLP classifier $f_{\mathrm{cls}}$ to obtain the output $s_{\mathrm{pred}}$ : + +$$ +s _ {\text {p r e d}} = f _ {\mathrm {c l s}} \left(\mathcal {P} ^ {f}\right), \tag {6} +$$ + +where $s_{\mathrm{pred}}$ is the logits for the prediction result of the semantic segmentation task of the input point cloud. + +# 4.5. Model Training and Inference + +During training, the obtained fused features $\mathcal{P}^f$ are fed into two classification heads: the category-aware classification head and the motion-aware classification head (a binary classifier), which outputs the predicted logits $s_{\mathrm{pred}}^c$ and $s_{\mathrm{pred}}^{m}$ (more details are included in the supplementary mate + +rial). First, with the ground truth labels of semantic categories $\mathcal{L}_{\mathrm{GT}}^c$ , back-propagation is performed for parameters optimization using Cross Entropy (CE) loss $L_{c}$ for semantic categories classification: + +$$ +L _ {c} = \mathrm {C E} \left(s _ {\text {p r e d}} ^ {c}, \mathcal {L} _ {\mathrm {G T}} ^ {c}\right), \tag {7} +$$ + +where $\mathrm{CE}(\cdot, \cdot)$ is the cross-entropy loss, and $s_{\mathrm{pred}}^c$ is the prediction probability for each point. Then, the Binary Cross Entropy (BCE) loss $L_m$ is used for motion states classification with motion states ground truth $\mathcal{L}_{\mathrm{GT}}^m$ : + +$$ +L _ {m} = \operatorname {B C E} \left(s _ {\text {p r e d}} ^ {m}, \mathcal {L} _ {\mathrm {G T}} ^ {m}\right), \tag {8} +$$ + +where $\mathrm{BCE}(\cdot ,\cdot)$ indicates the binary cross-entropy loss. + +The final objective function $L$ of the optimization is: + +$$ +L = \omega_ {c} \cdot L _ {c} + \omega_ {m} \cdot L _ {m}, \tag {9} +$$ + +where $\omega_{c}$ and $\omega_{m}$ are the weights of the two losses ( $L_{c}$ and $L_{m}$ ) respectively. + +During inference, the final prediction result is determined using the logits produced by both classification heads. When presented with an input sample, the motion-aware classification head will identify the motion state of the input point only if the category-aware classification head recognizes the point as belonging to a class with the potential to move. + +# 5. Experiments + +Datasets and Evaluation Metric: We evaluate our method on SemanticKITTI [1] and nuScenes [3]. For SemanticKITTI, the multi-scan setting is fully supervised and contains 25-category (6 moving categories and 19 static categories) with high-quality semantic annotations. The annotations are based on the KITTI dataset [8]. It comprises 22 point cloud sequences. For nuScenes, we propose a new + +![](images/fdfe8373ac4988da781f16f1c0956a6a63ea4126312dafc6113a1b385a806c8f.jpg) + +Baseline +![](images/4d17f7a62d66c608711c520f13094be453959a77160a44cd877c3f7b84086130.jpg) +- Car (non-moving) +- Car (moving) +- Person (non-moving) + +![](images/ef1b4cf530ee238f319cf758950b4955c935fca600019284733d6472e36ba12a.jpg) + +![](images/6459536973607d4ae86efe76562bc8a2051376f76b613a2d5794f878d975d292.jpg) +- Person (moving) - Bicyclist (moving) + +![](images/c6794fae2860a3c29e58d54b953e90821b595c1066010d124a5302281a143f02.jpg) + +![](images/d9e286c6dd5611c153be16992e85179a47d14f045a53e1e70969bffc6da74399.jpg) +Ground Truth + +![](images/effa9436ea92760b55b5cfb705532b140e11f20a8b33129899836d89e9f594b5.jpg) +Figure 3. Qualitative results on SemanticKITTI [1] public validation dataset. With SPVCNN [25] as the backbone, the segmentation results on the SemanticKITTI multi-scan task of the baseline model and our model are shown together with ground truth. At the same time, a specific area containing moving points is magnified and displayed at the top right of each sub-figure. + +![](images/ec414721ce044259c8e9d1599057931109e0412ea4c2266f888174f7b6dd6e8a.jpg) +Baseline +Figure 4. Evaluation errors (shown in red) by the baseline methods and MarS3D (using SPVCNN [25] as the backbone) on SemanticKITTI [1] public validation dataset. MarS3D significantly outperforms the baseline. Contrasting regions in the evaluation errors are highlighted with boxes and corresponding categories. + +![](images/59323c4b16e130f2985db3bb202931410b138e41c033a49a04be48319b6335c1.jpg) + +![](images/1e4b35d0dd42719f3e854a9e94352af08214f10a6090b2f12a0c272fe4f96d90.jpg) +Ours + +multi-scan setting based on the 'lidar-seg' task (16 semantic categories) without reference frame supervision. We use its object-level velocity to construct a multi-scan segmentation dataset with 24 categories (8 moving and 16 static categories). More details are provided in the supplementary material. To assess the effectiveness of our proposed method and make comparisons with baselines and other methods, we use the mean Intersection over Union (mIoU) as the evaluation metric. + +Implementation Details: Our model is designed as a plug-and-play module that provides motion-aware features to enhance the backbone learned features. For the backbone + +model, we consider SPVCNN $[25]^*$ , SparseConv $[9]^{\dagger}$ , and MinkUNet $[5]^{\ddagger}$ . Following previous works [12, 23, 36], we use the current and its previous two frames as input. The size of the BEV representation is set to $501 \times 301$ , and the multiple kernel sizes in the MAFL module are set as 1, 3, and 5, respectively. We set the embedded feature dimension in the CFFE module to 18, and the data augmentations are the same as the standard settings. All the models are trained on GeForce RTX 3090 GPUs, and the inference latency is recorded using a single GeForce RTX 3090 GPU. + +# 5.1. Main Results + +Comparison with Baseline Methods: We evaluate the performance of our proposed method on multi-scan benchmarks of SemanticKITTI [1] and nuScenes [3]. The baseline method uses the same backbone to process multiscan point clouds. We then compare this baseline approach to the same backbone augmented with MarS3D for a fair evaluation. As shown in Table 1, MarS3D significantly improves the performance over baseline methods on the public validation set of SemanticKITTI. With the most lightweight network, SPVCNN [25], MarS3D brings a $4.96\%$ improvement while introducing less than $0.5\%$ additional parameters. Particularly, consistent performance gains are observed on the dynamic object classes with non-moving/moving properties (i.e., car & moving car, person & moving person). This shows that the proposed BEV branch is both lightweight and powerful. Further, to verify the generalizability of the model, we offer a multiscan task based on nuScenes 'lidar-seg' dataset [3]. Our method outperforms the baseline (using MinkUNet [5] as the backbone) by a significant margin, achieving $64.83\%$ + +\*https://github.com/mit-han-lab/spvnas +https://github.com/traveller59/spconv +$\S$ https://github.com/NVIDIA/MinkowskiEngine + +
MethodmIoUcarbic.mot.tru.ove.per.bil.mol.roa.par.sid.ogr.bui.fen.veg.trn.ter.pol.tra.mca.mbi.mpe.mmm.mov.mtr.
SpSequenceNet [24]43.188.524.026.229.222.76.30.00.090.157.973.927.191.266.884.066.065.750.848.753.241.226.236.22.30.1
TemporalLidarSeg [7]47.092.147.740.939.235.014.40.00.091.859.675.823.289.863.882.362.564.752.660.468.242.840.412.912.42.1
TemporalLatticeNet [23]47.191.635.436.126.923.09.40.00.091.559.375.327.589.665.384.666.770.457.260.459.741.751.048.85.90.0
Meta-RangeSeg [28]49.590.152.743.930.335.414.30.00.090.763.374.726.990.563.583.067.067.756.464.464.556.155.024.420.33.4
KPConv [26]51.293.744.947.243.538.621.60.00.086.558.470.526.790.864.584.670.366.057.053.969.467.467.547.24.75.8
Baseline49.289.839.434.039.421.08.91.80.089.162.072.412.990.563.984.668.468.758.960.169.363.558.756.59.53.6
Ours52.795.149.249.539.736.616.21.20.089.966.874.326.492.168.286.072.170.562.864.878.467.358.036.310.05.1
+ +Table 2. Comparison with the state-of-the-art models on SemanticKITTI multi-scan benchmark (official test set). MarS3D (with SPVCNN [25] as the backbone) significantly outperforms these models for multi-scan tasks. (full names of the categories are in the supplementary material.) + +mIoU compared to the baseline's $61.90\%$ . This improvement is achieved with a negligible increase in each-frame inference time from 53ms to 58ms. + +Comparison with State-of-the-Art Methods: Compared to various models applied to multi-scan tasks, MarS3D (with SPVCNN as the backbone) is evaluated on the SemanticKITTI multi-scan benchmark. As shown in Table 2, the proposed approach has demonstrated superior performance, with a $1.5\%$ increase in mIoU compared to the current state-of-the-art method, KPCov [26]. Furthermore, our method performs similarly or better than other state-of-the-art models across nearly all categories. + +Qualitative Comparisons: Quantitative results are shown in Figure 3. The baseline model suffers from mistaking the status of static cars (Figure 3: Scan A) and moving persons (Figure 3: Scan B). It even recognizes the moving bicyclist as a moving person (Figure 3: Scan A). In contrast, MarS3D can circumvent such category and motion state discrimination errors. In addition to achieving better qualitative results in motion states, MarS3D also outperforms the baseline in semantic categories. We visualize the error maps (errors are shown in red) of the baseline and our method in Figure 4, where bounding boxes indicating specific categories highlight the differences between our method and the baseline. These improvements indicate that MarS3D has stronger semantic feature extraction capabilities compared to the baseline model, resulting in better segmentation performance even for immobile objects. + +# 5.2. Ablation Studies + +In this section, we conduct comprehensive ablation experiments on the SemanticKITTI multi-scan validation set to examine the effects of each component in our proposed method. As shown in Table 3, we gradually add three components to the baseline method, including the CFFE module, vanilla BEV branch (introduce only BEV representa + +![](images/81e0a066f37e0e9938026123e3e86a55dfd59577493f3c276d3dcc4cf746e8b3.jpg) + +![](images/e245684290a05cce033d6b53c5e58de8f0b1f8a6f819f95d9dd6e3ccb695cbd2.jpg) +BEV Ground Truth (Motion) + +![](images/44dbee35b43054220618e456657175e3fb700ca7c258b47830aec9dacb8643c8.jpg) + +![](images/fae1f5c467451ab53fc932b8b3446bf8c7e2080091e639c8a1a6452fe74f2c75.jpg) +Ours +Figure 5. Sampled discrepant feature maps demonstrate that the region that contains moving points are represented higher activation values than other regions. + +tions and 2D CNN without the MAFL module), and the MAFL module to illustrate the effectiveness of our designs. + +Effectiveness of CFFE Module: As shown in Table 3, we first employ only the CFFE module with the baseline model and observe a significant performance improvement. For instance, with SPVCNN [25] as the backbone, the introduction of the CFFE module led to a boost in performance by $1.1\%$ in terms of mIoU. This confirms the efficacy of the CFFE module for multi-frame representation learning. + +Effectiveness of BEV Representation: We also observe a significant performance improvement when we add the vanilla BEV branch alone to the baseline model. Using SPVCNN [25] as the backbone, the BEV branch results in a $3.7\%$ increase in mIoU. This confirms the importance of BEV representations in extracting motion-aware information. Furthermore, when both the CFFE module and the vanilla BEV branch are added to the baseline model, there is a further increase in performance, demonstrating that the two components complement each other. + +
MethodCFFEBEVMAFLmIoU(%)
baseline---49.7 / 49.0 / 48.5
proposed✓--50.8 / 51.3 / 52.1
-✓-53.4 / 53.9 / 53.1
✓✓-53.6 / 54.2 / 54.1
✓✓✓54.7 / 54.6 / 54.7
+ +Table 3. Ablation studies on different backbones (SPVCNN / SparseConv / MinkUNet) on SemanticKITTI [1] public validation dataset. The effectiveness of different designs is demonstrated step-by-step and our method is marked with a gray box. + +
MethodVanilla BEVTDNet [13]STM [18]Ours
mIoU(%)53.3653.56 (+0.20)53.64 (+0.28)54.66 (+1.30)
+ +Table 4. Comparing our method (using SPVCNN [25] as the backbone) with different 2D temporal semantic segmentation approaches the public validation dataset of SemanticKITTI [1]. + +Effectiveness of MAFL Module: In the final experiment presented in Table 3, we study the impact of our proposed MAFL module on the BEV representations, resulting in an increase of around $5.0\%$ from the baseline model (using SPVCNN [25] as the backbone). By stacking all the proposed components together, our final solution (marked with a grey box in Table 3) reaches its pinnacle in performance. The improvement in other backbones is also significant, and more statistical results are provided in the supplementary material. In addition, the multi-channel discrepant feature map represents the differences between the extracted feature maps of BEV representations from two point clouds. For a discrepant feature map generated during inference, we randomly sample the channels pixel by pixel. The average absolute activation values of pixels with the same motion state in the sampled channels are re-scaled and aggregated into a single-channel feature map as shown in Figure 5. The sampled discrepant feature map has higher activation on pixels containing moving points compared to other regions. This clearly demonstrates that the MAFL module clearly distinguishes between regions that contain moving points and those that do not. + +# 5.3. More Comparison on Temporal Segmentation + +As for BEV representation learning, the proposed MAFL module is utilized on extracted 2D feature maps for capturing inter-frame temporal information of point clouds. We further conduct comparative experiments by replacing the MAFL module with its 2D counterparts, i.e., STM [18] and TDNet [13]. These 2D modules have been widely utilized for preserving temporal information in 2D tasks. The results of the comparative experiments using SPVCNN [25] as the backbone are presented in Table 4. According to the evaluation results, it can be concluded that the MAFL module + +shows superior performance compared to the other models, thereby confirming its remarkable effectiveness. This shows that the proposed MAFL module is better suited for the specific task of handling motion-aware semantic segmentation in 3D point clouds. + +# 5.4. Limitations and Failure Cases Analysis + +Although MarS3D demonstrates impressive overall performance on the SemanticKITTI multi-scan benchmark, some limitations can be identified from the quantitative results in Table 1 and Table 2. The SemanticKITTI training dataset has an imbalanced distribution of point categories (more details are included in the supplementary material), which causes MarS3D to perform poorly on long-tailed categories due to insufficient training data on these categories. As shown in Table 2, MarS3D fails to show effects on several long-tail categories (i.e., bicyclist, motorcyclist, moving-other-vehicle, and moving-truck). Exploring solutions to address the long-tail problem is a promising research direction for the work. Since our model assumes planar motion, it may not perform well in scenarios where objects move in non-planar ways, such as on steep terrains or on non-planar surfaces. + +# 6. Conclusion + +In this paper, we propose MarS3D, a novel plug-and-play motion-aware model for 3D multi-scan point cloud semantic segmentation. The Motion-Aware Feature Learning (MAFL) module, based on BEV representations, is designed to facilitate extracting motion-aware representations. Additionally, the Cross-Frame Feature Embedding (CFFE) module is introduced to improve representation learning by embedding time-step information into features, thus preserving rich temporal information. Extensive experiments and ablation studies demonstrate that MarS3D significantly improves multiple 3D semantic segmentation baselines while introducing minimal overheads. In comparison to state-of-the-art methods designed for multi-scan tasks, MarS3D achieves superior performance and offers significant improvements over baseline methods. The proposed MarS3D model demonstrates the potential for effectively incorporating motion-awareness into 3D point cloud semantic segmentation tasks, providing a strong foundation for further research and development in this area. + +Acknowledgments: This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant no. 17202422), and RGC matching fund scheme (RMGS). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. + +# References + +[1] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF International Conf. on Computer Vision (ICCV), 2019. 1, 2, 5, 6, 8 +[2] Goutam Bhat, Felix Järemo Lawin, Martin Danelljan, Andreas Robinson, Michael Felsberg, Luc Van Gool, and Radu Timofte. Learning what to learn for video object segmentation. In European Conference on Computer Vision, pages 777-794. Springer, 2020. 3 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5, 6 +[4] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 1, 2 +[5] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 1, 2, 5, 6 +[6] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 3 +[7] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 2020 International Conference on 3D Vision (3DV), pages 781-790. IEEE, 2020. 1, 2, 7 +[8] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 5 +[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 5, 6 +[10] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. 2 +[11] Noureldin Hendy, Cooper Sloan, Feng Tian, Pengfei Duan, Nick Charchut, Yuesong Xie, Chuang Wang, and James Philbin. Fishing net: Future inference of semantic heatmaps in grids. arXiv preprint arXiv:2006.09917, 2020. 2 +[12] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In Proceedings of the IEEE/CVF Con + +ference on Computer Vision and Pattern Recognition, pages 8479-8488, 2022. 1, 2, 6 +[13] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks for fast video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8827, 2020. 3, 8 +[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2, 4 +[15] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In European Conference on Computer Vision, pages 352-368. Springer, 2020. 3 +[16] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 2, 3, 4 +[17] Jiaxu Miao, Yunchao Wei, and Yi Yang. Memory aggregation networks for efficient interactive video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10366-10375, 2020. 3 +[18] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226–9235, 2019. 3, 8 +[19] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[20] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2 +[21] Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7406-7415, 2020. 3 +[22] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4 +[23] Peer Schutt, Radu Alexandru Rosu, and Sven Behnke. Abstract flow for temporal semantic segmentation on the permutohedral lattice. In 2022 International Conference on Robotics and Automation (ICRA), pages 5139-5145. IEEE, 2022. 1, 2, 6, 7 +[24] Hanyu Shi, Guosheng Lin, Hao Wang, Tzu-Yi Hung, and Zhenhua Wang. Spsequencenet: Semantic segmentation network on 4d point clouds. In Proceedings of the IEEE/CVF + +conference on computer vision and pattern recognition, pages 4574-4583, 2020. 1, 2, 7 +[25] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 1, 2, 5, 6, 7, 8 +[26] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2, 7 +[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[28] Song Wang, Jianke Zhu, and Ruixiang Zhang. Meta-rangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. arXiv preprint arXiv:2202.13377, 2022. 1, 2, 7 +[29] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2 +[30] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9621-9630, 2019. 2 +[31] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. 1, 2 +[32] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3173-3182, 2021. 2 +[33] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3101-3109, 2021. 1 +[34] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695. Springer, 2022. 2 +[35] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2 +[36] Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective + +3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550, 2020. 1, 2, 6 +[37] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 2 +[38] Zhenhong Zou and Yizhe Li. Efficient urban-scale point clouds segmentation with bev projection. arXiv preprint arXiv:2109.09074, 2021. 2, 3 \ No newline at end of file diff --git a/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/images.zip b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..33ca62fda2155548f8220459c7b8d2d266c19e1f --- /dev/null +++ b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d6e20ee7bdf8d48560dd90d06f5f05041e8bb709681bf55db719f0b17ede5d1 +size 652363 diff --git a/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/layout.json b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..86306518389c86e66efc0220c2f7051bb85ed4cc --- /dev/null +++ b/2023/MarS3D_ A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds/layout.json @@ -0,0 +1,9491 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "type": "text", + "content": "MarS3D: A Plug-and-Play Motion-Aware Model for Semantic Segmentation on Multi-Scan 3D Point Clouds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "spans": [ + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "content": "Jiahui Liu" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "content": " Chirui Chang" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "content": " Jianhui Liu" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "content": " Xiaoyang Wu" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "content": " Lan Ma" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "content": " Xiaojuan Qi" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "content": "The University of Hong Kong " + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 160, + 512, + 190 + ], + "type": "text", + "content": "TCL AI Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 41, + 191, + 550, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 191, + 550, + 205 + ], + "spans": [ + { + "bbox": [ + 41, + 191, + 550, + 205 + ], + "type": "text", + "content": "{liujh, xjqi}@eee.hku.hk, crchang@hku.hk, jhliu0212@gmail.com, xywu3@cs.hku.hk, rubyma@tcl.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 483 + ], + "type": "text", + "content": "3D semantic segmentation on multi-scan large-scale point clouds plays an important role in autonomous systems. Unlike the single-scan-based semantic segmentation task, this task requires distinguishing the motion states of points in addition to their semantic categories. However, methods designed for single-scan-based segmentation tasks perform poorly on the multi-scan task due to the lacking of an effective way to integrate temporal information. We propose MarS3D, a plug-and-play motion-aware module for semantic segmentation on multi-scan 3D point clouds. This module can be flexibly combined with single-scan models to allow them to have multi-scan perception abilities. The model encompasses two key designs: the Cross-Frame Feature Embedding module for enriching representation learning and the Motion-Aware Feature Learning module for enhancing motion awareness. Extensive experiments show that MarS3D can improve the performance of the baseline model by a large margin. The code is available at https://github.com/CVMI-Lab/MarS3D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 507, + 128, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 128, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 128, + 520 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 528, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 635 + ], + "type": "text", + "content": "3D semantic segmentation on multi-scan large-scale point clouds is a fundamental computer vision task that benefits many downstream problems in autonomous systems, such as decision-making, motion planning, and 3D reconstruction, to name just a few. Compared with the single-scan semantic segmentation task, this task requires understanding not only the semantic categories but also the motion states (e.g., moving or static) of points based on multi-scan point cloud data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 636, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 288, + 685 + ], + "type": "text", + "content": "In the past few years, extensive research has been conducted on single-scan semantic segmentation with significant research advancements [4, 5, 12, 25, 31, 33, 36]. These approaches are also applied to process multi-scan point" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 230, + 452, + 280 + ], + "blocks": [ + { + "bbox": [ + 304, + 230, + 452, + 280 + ], + "lines": [ + { + "bbox": [ + 304, + 230, + 452, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 452, + 280 + ], + "type": "image", + "image_path": "f9d09b39f4540770679c524d7c0977eaec2b1b1ac17ef0fe847200928bce8d5e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 453, + 231, + 542, + 280 + ], + "blocks": [ + { + "bbox": [ + 453, + 231, + 542, + 280 + ], + "lines": [ + { + "bbox": [ + 453, + 231, + 542, + 280 + ], + "spans": [ + { + "bbox": [ + 453, + 231, + 542, + 280 + ], + "type": "image", + "image_path": "6b4cd622f2f091fa38408dd9e49b0ecf4b1844aeabb9a3faf8c869c65d7230dd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 304, + 281, + 452, + 329 + ], + "blocks": [ + { + "bbox": [ + 304, + 281, + 452, + 329 + ], + "lines": [ + { + "bbox": [ + 304, + 281, + 452, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 452, + 329 + ], + "type": "image", + "image_path": "98afb6d81e5b1ed5376fdb0dba48dd8b30b7f108d1dcb422584c9fab0f357aef.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 453, + 281, + 542, + 329 + ], + "blocks": [ + { + "bbox": [ + 453, + 281, + 542, + 329 + ], + "lines": [ + { + "bbox": [ + 453, + 281, + 542, + 329 + ], + "spans": [ + { + "bbox": [ + 453, + 281, + 542, + 329 + ], + "type": "image", + "image_path": "6504b14347e31fa8590b698fda50d81fb4c21fdfc69a98b56425a4b8ebea1119.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 304, + 331, + 452, + 379 + ], + "blocks": [ + { + "bbox": [ + 304, + 331, + 452, + 379 + ], + "lines": [ + { + "bbox": [ + 304, + 331, + 452, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 452, + 379 + ], + "type": "image", + "image_path": "c5aae9841f52b11b350a1fa821a1a310293e330f6794d9ab896e93e849cbc05c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 385, + 547, + 441 + ], + "lines": [ + { + "bbox": [ + 304, + 385, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 547, + 441 + ], + "type": "text", + "content": "Figure 1. Comparison of our proposed method, MarS3D, with baseline method using SPVCNN [25] as the backbone on SemanticKITTI [1] dataset. MarS3D achieves excellent results in the classification of semantic categories and motion states, while the baseline method can not distinguish motion well from static." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 453, + 331, + 542, + 379 + ], + "blocks": [ + { + "bbox": [ + 453, + 331, + 542, + 379 + ], + "lines": [ + { + "bbox": [ + 453, + 331, + 542, + 379 + ], + "spans": [ + { + "bbox": [ + 453, + 331, + 542, + 379 + ], + "type": "image", + "image_path": "fecbff26a11e4b7608fb3cdcb2c0e6ed2750c36801c1131ee3738b396d4d0c94.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 460, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 652 + ], + "type": "text", + "content": "clouds, wherein multiple point clouds are fused to form a single point cloud before being fed to the network for processing. Albeit simple, this strategy may lose temporal information and make distinguishing motion states a challenging problem. As a result, they perform poorly in classifying the motion states of objects. As shown in Figure 1, the simple point cloud fusion strategy cannot effectively enable the model to distinguish the motion states of cars even with a state-of-the-art backbone network SPVCNN [25]. Recently, there have been some early attempts [7, 23, 24, 28] to employ attention modules [24] and recurrent networks [7, 23, 28] to fuse information across different temporal frames. However, these approaches do not perform well on the multi-scan task due to the insufficiency of temporal representations and the limited feature extraction ability of the model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In sum, a systematic investigation of utilizing the rich spatial-temporal information from multiple-point cloud scans is still lacking. This requires answering two critical questions: (1) how can we leverage the multi-scan information to improve representation learning on point clouds for" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9372" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "better semantic understanding? and (2) how can the temporal information be effectively extracted and learned for classifying the motion states of objects?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 335 + ], + "type": "text", + "content": "In this paper, we propose a simple plug-and-play Motion-aware Segmentation module for 3D multi-scan analysis (MarS3D), which can seamlessly integrate with existing single-scan semantic segmentation models and endow them with the ability to perform accurate multi-scan 3D point cloud semantic segmentation with negligible computational costs. Specifically, our method incorporates two core designs: First, to enrich representation learning of multi-frame point clouds, we propose a Cross-Frame Feature Embedding (CFFE) module which embeds time-step information into features to facilitate inter-frame fusion and representation learning. Second, inspired by the observation that objects primarily move along the horizontal ground plane (i.e., " + }, + { + "bbox": [ + 46, + 109, + 288, + 335 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 46, + 109, + 288, + 335 + ], + "type": "text", + "content": "-plane) in large-scale outdoor scenes, i.e., minimal motion along the " + }, + { + "bbox": [ + 46, + 109, + 288, + 335 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 109, + 288, + 335 + ], + "type": "text", + "content": "-axis, we propose a Motion-Aware Feature Learning (MAFL) module based on Bird's Eye View (BEV), which learns the motion patterns of objects between frames to facilitate effectively discriminating the motion states of objects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 336, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 419 + ], + "type": "text", + "content": "We extensively evaluate our approach upon several mainstream baseline frameworks on SemanticKITTI [1] and nuScenes [3] dataset. It consistently improves the performance of the baseline approaches, e.g., MinkUnet [5], by " + }, + { + "bbox": [ + 46, + 336, + 287, + 419 + ], + "type": "inline_equation", + "content": "6.24\\%" + }, + { + "bbox": [ + 46, + 336, + 287, + 419 + ], + "type": "text", + "content": " in mIoU on SemanticKITTI with a negligible increase in model parameters, i.e., about " + }, + { + "bbox": [ + 46, + 336, + 287, + 419 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 336, + 287, + 419 + ], + "type": "text", + "content": ". The main contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 424, + 286, + 643 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 424, + 286, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 424, + 286, + 482 + ], + "spans": [ + { + "bbox": [ + 58, + 424, + 286, + 482 + ], + "type": "text", + "content": "- We are the first to propose a plug-and-play module for large-scale multi-scan 3D semantic segmentation, which can be flexibly integrated with mainstream single-scan segmentation models without incurring too much cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 485, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 485, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 58, + 485, + 286, + 533 + ], + "type": "text", + "content": "- We devise a Cross-Frame Feature Embedding module to fuse multiple point clouds while preserving their temporal information, thereby enriching representation learning for multi-scan point clouds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 535, + 286, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 535, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 58, + 535, + 286, + 582 + ], + "type": "text", + "content": "- We introduce a BEV-based Motion-Aware Feature Learning module to exploit temporal information and enhance the model's motion awareness, facilitating the prediction of motion states." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 585, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 585, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 58, + 585, + 286, + 643 + ], + "type": "text", + "content": "- We conduct extensive experiments and comprehensive analyses of our approach with different backbone models. The proposed model performs favorably compared to the baseline methods while introducing negligible extra parameters and inference time." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 657, + 133, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 133, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 133, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Single-scan Outdoor 3D Semantic Segmentation: 3D single-scan outdoor semantic segmentation is indispensable for autonomous driving. In early work, PointNet [19] uses" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "Multi-Layer Perception (MLP) to extract features from input point clouds directly, and PointNet++ [20] tries to incorporate multi-scale designs for dense prediction tasks. Later, various literature [19, 20, 26, 29, 30, 32] works on designing point-based convolution on either geometric or semantic neighborhoods. To handle the large-scale dataset, some works [4, 5, 10, 12, 25, 31, 36] focus on volumetric features and use 3D convolution to achieve a balance between accuracy and efficiency. SparseConv [10] and MinkUNet [5] are representative works and demonstrate good performance. Later, SPVNAS [25] combines voxel and point representations and designs a neural architecture search method to find the optimal model structure. Recently, Cylinder3D [36] introduces a cylindrical partition to leverage the properties of LiDAR point clouds for enriching the feature information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 253, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 350 + ], + "type": "text", + "content": "The remarkable feature extraction capability enables the above methods to achieve high performance on single-scan tasks. To solve the multi-scan task, most of these methods [4, 12, 31, 34, 36] first fuse multiple point clouds into one and treat the fused point cloud as a single point cloud for processing. Albeit simple, this fusion strategy overlooks important temporal information and entangles moving and non-moving objects, leading to performance degradation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 357, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 597 + ], + "type": "text", + "content": "Multi-scan Outdoor 3D Semantic Segmentation: Compared to single-scan semantic segmentation, the multi-scan task needs to discriminate the moving and stationary states of the objects based on temporal information. In addition to the simple fusion strategy discussed above, another stream of approaches [7,23,24,28] attempts to process each point cloud in a sequence separately and fuse the feature representations for temporal modeling. For instance, SpSequenceNet [24] proposes a U-Net-based architecture to extract per-frame features and combine features of two consecutive frames to gather temporal information. The fused feature is further fed into the prediction head to produce results. Duerr et al. [7] design a recurrent architecture with a temporal memory alignment module for sequential processing of multiple point clouds. TemporalLatticeNet [23] proposes to match similar feature patterns between adjacent frames and fuse them temporally. However, these approaches cannot fully leverage multiple point clouds to enrich temporal representation learning as the feature extraction is still conducted on each frame separately." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "BEV-based 3D Point Cloud Perception: Recently, BEV-based representation [11, 14, 16, 35, 37, 38] has emerged as an effective way to process 3D point cloud due to its efficiency and ease of deployment using 2D convolution operations. By projecting a 3D point cloud into the bird's eye view (BEV), BEV-based representation converts 3D representations into 2D to avoid heavy processing in 3D and improve computation and memory efficiency. Some representative works include PointPillars [14] for 3D object de" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9373" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 45, + 70, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 45, + 70, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 45, + 70, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 45, + 70, + 545, + 266 + ], + "type": "image", + "image_path": "ad5b1b54236d4f4bc8b060629505e83f9ffd38f0e1446825bcd3f29ed3788137.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 270, + 548, + 327 + ], + "lines": [ + { + "bbox": [ + 45, + 270, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 45, + 270, + 548, + 327 + ], + "type": "text", + "content": "Figure 2. The proposed framework of MarS3D. We take three scan point clouds inputs as illustration. MarS3D contains two branches. One is the BEV branch with 2D BEV representations as input, and it employs a Motion-Aware Feature Learning (MAFL) module to enhance the motion-aware feature learning. The other is the 3D branch that takes multi-scan 3D point clouds as input, enriching the feature representations with our Cross-Frame Feature Embedding (CFFE) module. The fused features of the above two modules are fed into the prediction head and make final predictions on semantic categories and motion states." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 343, + 289, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 289, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 289, + 416 + ], + "type": "text", + "content": "tection, BEVfusion [16] for multi-sensor fusion, BEV projection [38] for 3D segmentation. The above work demonstrates the potential of BEV representations in outdoor 3D scene analysis. Here, we explore utilizing BEV-based representation to extract temporal information for analyzing the motion states of objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 425, + 289, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 289, + 581 + ], + "type": "text", + "content": "2D Video Semantic Segmentation: In video segmentation, many efforts [2,6,13,15,17,18,21] have been made to conduct temporal feature extraction. Oh et al. [18] introduce a space-time memory network to integrate features from adjacent frames for video object segmentation and attains significant performance gains. TDNet [13] exploits temporal redundancy for fast video segmentation by developing an attention propagation method to propagate features to adjacent frames. Although they achieve promising results on video segmentation, we experimentally demonstrate these designs are not optimal for multi-scan point cloud segmentation. We focus on designing new methods tailored to point clouds for better performance and efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 596, + 162, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 162, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 162, + 609 + ], + "type": "text", + "content": "3. Problem Statement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": "Given a sequence of LiDAR point clouds as inputs, the multi-scan 3D point cloud semantic segmentation task aims to assign a semantic category to each point and predict their motion states (i.e., moving or static). Specifically, a LiDAR point cloud frame contains a set of unordered points that are annotated with labels for training. We denote a pair of training data as " + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}_i,\\mathcal{L}_i) = \\{p_j,l_j\\}_{j = 1}^N" + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_j\\in \\mathbb{R}^{D_{\\mathrm{in}}}" + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": " denotes the number of points. Each point " + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": " contains in-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "text", + "content": "put descriptors with " + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "text", + "content": " dimensions, including point coordinates " + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "text", + "content": " and other features such as intensities (bm). The corresponding label " + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "inline_equation", + "content": "l_{j}" + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "text", + "content": " incorporates both semantic categories and motion states of " + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "text", + "content": ". Therefore, points belonging to the same semantic category but possessing distinct motion states are allocated distinct labels. For a sequence of point clouds " + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_i,\\mathcal{L}_i\\}_{i = 1}^M" + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "text", + "content": " that contains " + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 343, + 547, + 499 + ], + "type": "text", + "content": " frames, all frames are scanned sequentially in time order, and scanning poses and timestamps are used to align different frames into the same world-coordinate system. In the following, we omit the alignment process for simplicity, where the point clouds in a sequence are calibrated to the same coordinate system by default." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 514, + 362, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 362, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 362, + 526 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 534, + 373, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 373, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 373, + 546 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "text", + "content": "An overview of our motion-aware model for multi-scan semantic segmentation, namely MarS3D, is shown in Figure 2. MarS3D contains a 3D branch for multi-scan spatial representation learning and a BEV branch for motion-aware feature learning. First, the BEV branch takes as inputs " + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "text", + "content": " BEV representations " + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{\\mathrm{in}} = \\left\\{\\mathcal{B}_i^{\\mathrm{in}}\\right\\}_{i=1}^k" + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "text", + "content": " that are derived by point cloud polarization (see Section 4.2) and outputs motion-aware feature map " + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^m" + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "text", + "content": ". The core is a Motion-Aware Feature Learning (MAFL) module (see Section 4.2) that extracts and leverages BEV features through a dedicated design to produce a motion-aware feature map " + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^m" + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "text", + "content": ". Second, the 3D branch takes as inputs the fused " + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "text", + "content": " point clouds " + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{in}} = \\left\\{\\mathcal{P}_i^{\\mathrm{in}}\\right\\}_{i=1}^k" + }, + { + "bbox": [ + 304, + 554, + 547, + 715 + ], + "type": "text", + "content": " and outputs enriched 3D enhanced spa" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9374" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "text", + "content": "tial features " + }, + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s" + }, + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "text", + "content": ". The 3D branch incorporates a cross-frame feature embedding (CFFE) module (see Section 4.3) to inject temporal information. It outputs embedded features, denoted as " + }, + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{ebd}} = \\left\\{\\mathcal{P}_i^{\\mathrm{ebd}}\\right\\}_{i=1}^k" + }, + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "text", + "content": ". These features are further processed by a single-scan backbone " + }, + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "text", + "content": " to yield multi-scan enhanced 3D spatial representations " + }, + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s" + }, + { + "bbox": [ + 46, + 72, + 289, + 148 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 159, + 287, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 159, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 133, + 159, + 287, + 174 + ], + "type": "interline_equation", + "content": "\\mathcal {P} ^ {s} = f _ {s} \\left(\\mathcal {P} ^ {\\mathrm {e b d}}\\right), \\tag {1}", + "image_path": "fc920983c8da0641bf5125e833a90c36a1bd4923bfaabbeaf88bec0a0e4a85c7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 180, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s\\in \\mathbb{R}^{D_z}" + }, + { + "bbox": [ + 46, + 180, + 287, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 180, + 287, + 205 + ], + "type": "inline_equation", + "content": "D_{z}" + }, + { + "bbox": [ + 46, + 180, + 287, + 205 + ], + "type": "text", + "content": " is the dimension of the output feature of the single-scan backbone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "type": "text", + "content": "Finally, the motion-aware feature map " + }, + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^m" + }, + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "type": "text", + "content": " and enhanced spatial features " + }, + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s" + }, + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "type": "text", + "content": " are combined to produce the fused features " + }, + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^f" + }, + { + "bbox": [ + 46, + 206, + 289, + 278 + ], + "type": "text", + "content": " by aligning the coordinates of the features with the pixels of the motion-aware feature map. The fused features are then fed to the prediction head to produce the final outputs (see Section 4.4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 287, + 130, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 130, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 130, + 300 + ], + "type": "text", + "content": "4.2. BEV Branch" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 306, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 390 + ], + "type": "text", + "content": "In the following, we elaborate on the BEV branch, which targets extracting motion-friendly features for motion prediction (see Figure 2). Before delving into the details, we first introduce BEV mapping, which maps a 3D point cloud into a 2D BEV image. Then, we introduce our key Motion-Aware Feature Learning (MAFL) module for motion extraction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": "BEV Mapping: Each point cloud " + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": " for the BEV branch is pillarized into a three-channel BEV map " + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": ". Inspired by [14, 16], for a pillar located at " + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "inline_equation", + "content": "(x^{*},y^{*})" + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": " with a pixel grid size of " + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "inline_equation", + "content": "l_{B}" + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": ", its three-channel feature " + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "inline_equation", + "content": "b_{(x^{*},y^{*})}" + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": " consists of the average translation along " + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 398, + 289, + 495 + ], + "type": "text", + "content": " axis regarding the grid center, and total intensity features (taking SemanticKITTI as an example) among all points inside the pillar, which is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 514, + 287, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 287, + 557 + ], + "type": "interline_equation", + "content": "b _ {\\left(x ^ {*}, y ^ {*}\\right)} = \\left[ \\frac {1}{N ^ {*}} \\sum_ {i} ^ {N ^ {*}} \\frac {2 \\Delta x _ {i}}{l _ {B}}, \\frac {1}{N ^ {*}} \\sum_ {i} ^ {N ^ {*}} \\frac {2 \\Delta y _ {i}}{l _ {B}}, \\sum_ {i} ^ {N ^ {*}} \\mathrm {b m} _ {i} \\right], \\tag {2}", + "image_path": "0c713b1b2e233c60cfb1593f5b59575d0297b6cb7d44ad59d3c1a9a0ce11b92b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "inline_equation", + "content": "N^{*}" + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "content": " denotes the number of points in current pillar, while " + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\Delta x_{i}" + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\Delta y_{i}" + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "content": " denote the translation of the ith point along " + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "content": "-axis and " + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "content": "-axis respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "text", + "content": "Motion-Aware Feature Learning Module: Our Motion-Aware Feature Learning (MAFL) module is illustrated in the BEV branch in Figure 2 which is designed to extract motion-aware features for better distinguishing moving/static objects. The input to this module consists of " + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "text", + "content": " multi-scan BEV representations " + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{\\mathrm{in}} = \\left\\{\\mathcal{B}_i^{\\mathrm{in}}\\right\\}_{i=1}^k" + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_k^{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "text", + "content": " represents the target frame. A lightweight 2D CNN " + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "inline_equation", + "content": "f_u" + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "text", + "content": " with a UNet-like architecture [22] is used to extract features " + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{U} = \\{\\mathcal{U}_i\\}_{i=1}^k" + }, + { + "bbox": [ + 46, + 602, + 289, + 715 + ], + "type": "text", + "content": " from multi-scan inputs as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 84, + 546, + 103 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 84, + 546, + 103 + ], + "spans": [ + { + "bbox": [ + 372, + 84, + 546, + 103 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathcal {U} _ {i} \\right\\} _ {i = 1} ^ {k} = \\left\\{f _ {u} \\left(\\mathcal {B} _ {i} ^ {\\text {i n}}\\right) \\right\\} _ {i = 1} ^ {k}. \\tag {3}", + "image_path": "489be3a6316a6dab0bcd51e8963ffec8aac82623cd5dc40fc6cfbba3d1f246d4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "type": "text", + "content": "Furthermore, to identify moving objects, we take the difference between the target frame " + }, + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "type": "text", + "content": " and the remaining " + }, + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "type": "inline_equation", + "content": "k - 1" + }, + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "type": "text", + "content": " reference frames which outputs " + }, + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathcal{D}_i\\}_{i = 1}^{k - 1}" + }, + { + "bbox": [ + 305, + 110, + 547, + 149 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 368, + 162, + 546, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 162, + 546, + 178 + ], + "spans": [ + { + "bbox": [ + 368, + 162, + 546, + 178 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathcal {D} _ {i} \\right\\} _ {i = 1} ^ {k - 1} = \\left\\{\\mathcal {U} _ {k} - \\mathcal {U} _ {i} \\right\\} _ {i = 1} ^ {k - 1}. \\tag {4}", + "image_path": "2dc7a5d1c0e008263e95fcb8998aff16b606847df6f2727d970649cfedf71e3f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "text", + "content": "By doing so, the static objects can be erased, and the dynamic objects are highlighted with a large feature magnitude. Then, " + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_i\\}_{i=1}^{k-1}" + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "text", + "content": " are channel-wise concatenated to form a new 2D map. Note that objects may have different moving patterns and velocities. Therefore, we design a multi-kernel convolutional network " + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "inline_equation", + "content": "f_m" + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "text", + "content": " with multiple branches of various kernel sizes to capture objects with various movement patterns. Finally, the outputs from " + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "inline_equation", + "content": "f_m" + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "text", + "content": " are concatenated to output a motion-aware feature map " + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^m" + }, + { + "bbox": [ + 305, + 187, + 547, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 307, + 380, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 380, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 380, + 319 + ], + "type": "text", + "content": "4.3. 3D Branch" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 327, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 547, + 422 + ], + "type": "text", + "content": "The 3D branch uses temporal information to enhance spatial representation learning on 3D point clouds. The core component is the Cross-Frame Feature Embedding (CFFE) module, whose output is fed into the single-scan backbone network to produce enhanced spatial features " + }, + { + "bbox": [ + 305, + 327, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s" + }, + { + "bbox": [ + 305, + 327, + 547, + 422 + ], + "type": "text", + "content": " (see Figure 2). In the following, we elaborate on the CFFE module to improve spatial representation learning on multi-scan point clouds." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": "Cross-Frame Feature Embedding Module: When the multi-frame point clouds are fused as discussed in the previous sections, points from different time steps are mixed, making it challenging for the final recognition. Inspired by positional embedding [27], we propose a Cross-Frame Feature Embedding (CFFE) module to generate a time-aware embedding and produce consistent features for each point across different timestamps. Given " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": " point cloud frames, " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_i^{\\mathrm{in}}\\}_{i = 1}^k" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": ", we design an embedding neural network layer " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": " that maps point clouds into intermediate latent features and " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": " learnable temporal embeddings " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_i\\}_{i = 1}^k" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": " frames respectively. The dimension of " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": " is the same as the output dimension of the " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": ". The point-level embedded features " + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{ebd}}" + }, + { + "bbox": [ + 305, + 431, + 547, + 602 + ], + "type": "text", + "content": " are obtained by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 616, + 546, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 616, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 356, + 616, + 546, + 635 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathcal {P} _ {i} ^ {\\mathrm {e b d}} \\right\\} _ {i = 1} ^ {k} = \\left\\{e _ {i} + f _ {e} \\left(\\mathcal {P} _ {i} ^ {\\mathrm {i n}}\\right) \\right\\} _ {i = 1} ^ {k}, \\tag {5}", + "image_path": "c1a62eb6fc33bdcb654959cad603c50a20b9a8f41d6840647c5fb57df506a100.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "text", + "content": "where element-wise summation is conducted on " + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "text", + "content": " and each point of the corresponding " + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_e(\\mathcal{P}_i^{\\mathrm{in}})" + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "text", + "content": ". The obtained features with different temporal embeddings are represented in a point cloud " + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_i^{\\mathrm{ebd}}\\}_{i=1}^k" + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "text", + "content": " and subsequently fed into the single-scan backbone network " + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "text", + "content": ", following Eq. (1), to produce a set of enhanced spatial features " + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s" + }, + { + "bbox": [ + 305, + 641, + 547, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9375" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 72, + 545, + 208 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 545, + 208 + ], + "type": "table", + "html": "
B.MethodmIoU#paramlatencycarbic.mot.tru.ove.per.bil.mol.roa.par.sid.ogr.bui.fen.veg.trn.ter.pol.tra.mca.mbi.mpe.mmm.mov.mtr.
SPVCNN [25]Baseline49.7021.8M206ms93.934.464.768.033.019.70.00.093.645.280.10.290.359.788.463.575.664.151.974.386.755.00.00.00.0
Ours54.6621.9M225ms95.652.777.879.451.527.90.00.094.251.382.00.191.665.189.068.476.265.251.580.694.968.00.03.60.0
Δ+4.96+0.1M+19ms1.718.313.111.418.58.20.00.00.66.11.90.11.35.40.64.90.61.10.46.38.213.00.03.60.0
SparseConv [9]Baseline48.9939.2M239ms94.724.154.169.643.417.30.20.093.245.179.80.289.561.787.762.974.663.850.073.985.453.60.00.00.0
Ours54.6439.3M253ms96.635.269.083.364.826.90.00.094.061.282.50.190.965.888.167.875.266.251.483.594.468.90.00.00.0
Δ+5.65+0.1M+14ms1.911.114.913.721.49.60.20.00.816.12.70.11.44.10.44.90.62.41.49.69.015.30.00.00.0
MinkUnet [5]Baseline48.4737.9M295ms93.823.748.990.341.318.00.00.092.232.278.40.089.855.588.863.777.063.650.069.283.152.50.00.00.0
Ours54.7138.0M323ms96.428.470.093.962.731.60.00.093.858.581.70.192.667.689.066.776.466.551.682.693.164.40.00.10.0
Δ+6.24+0.1M+28ms2.64.721.13.621.413.60.00.01.626.33.30.12.812.10.23.00.62.91.613.410.011.90.00.10.0
", + "image_path": "07b6b8225f02675f00435063bf638d123456bafa3239d12c5089839553ec7271.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 212, + 546, + 258 + ], + "lines": [ + { + "bbox": [ + 45, + 212, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 45, + 212, + 546, + 258 + ], + "type": "text", + "content": "Table 1. Quantitative results of the proposed method, MarS3D, on SemanticKITTI [1] multi-scan public validation set. Combined with different mainstream single-scan 3D point cloud semantic segmentation backbones, MarS3D has a large performance improvement over the corresponding baseline methods without introducing excessive parameters and each-frame inference time. (B. indicates Backbone, full names of the categories are in the supplementary material, blue indicates degradation.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 275, + 212, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 212, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 212, + 287 + ], + "type": "text", + "content": "4.4. Feature Fusion and Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 343 + ], + "type": "text", + "content": "As illustrated in Figure 2, equipped with the motion-aware feature map " + }, + { + "bbox": [ + 46, + 294, + 287, + 343 + ], + "type": "inline_equation", + "content": "(\\mathcal{Z}^m)" + }, + { + "bbox": [ + 46, + 294, + 287, + 343 + ], + "type": "text", + "content": " and enhanced spatial features " + }, + { + "bbox": [ + 46, + 294, + 287, + 343 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^s)" + }, + { + "bbox": [ + 46, + 294, + 287, + 343 + ], + "type": "text", + "content": ", the next step is to fully integrate the representation information from both branches and make predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": "Feature Fusion: The motion-aware feature map " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^m\\in" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times D_z}" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " ( " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " : height; " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " : width; " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "D_{z}" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " : number of channels) is obtained from the BEV branch, while the enhanced spatial features " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s\\in \\mathbb{R}^{N\\times D_p}" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " : number of points; " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " : number of channels) are the outputs of the 3D branch. The feature fusion module aggregates information from the above two representations to make subsequent predictions. For each point feature in " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " dimension, a corresponding pixel from " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^m" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " can be queried based on the point's 3D location (" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": " coordinates). This pixel serves as an index to extract a " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "D_{z}" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": "-dimensional motion-aware feature along the channel dimension from " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^m" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": ". Subsequently, the motion-aware feature is concatenated with the point feature, resulting in the fused features denoted as " + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^f\\in \\mathbb{R}^{N\\times (D_z + D_p)}" + }, + { + "bbox": [ + 46, + 350, + 288, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "type": "text", + "content": "Prediction: Based on the correspondence between 2D and 3D, " + }, + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^f" + }, + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "type": "text", + "content": " is then fed into an MLP classifier " + }, + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "type": "text", + "content": " to obtain the output " + }, + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{pred}}" + }, + { + "bbox": [ + 46, + 536, + 287, + 573 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 583, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 287, + 597 + ], + "type": "interline_equation", + "content": "s _ {\\text {p r e d}} = f _ {\\mathrm {c l s}} \\left(\\mathcal {P} ^ {f}\\right), \\tag {6}", + "image_path": "e519ef3f608d889799b0871cbbc5e8a35edbce7a206b5565fa745e32e9979039.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{pred}}" + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": " is the logits for the prediction result of the semantic segmentation task of the input point cloud." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 635, + 211, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 211, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 211, + 647 + ], + "type": "text", + "content": "4.5. Model Training and Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "type": "text", + "content": "During training, the obtained fused features " + }, + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^f" + }, + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "type": "text", + "content": " are fed into two classification heads: the category-aware classification head and the motion-aware classification head (a binary classifier), which outputs the predicted logits " + }, + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{pred}}^c" + }, + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{pred}}^{m}" + }, + { + "bbox": [ + 46, + 653, + 287, + 716 + ], + "type": "text", + "content": " (more details are included in the supplementary mate" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 276, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 324 + ], + "type": "text", + "content": "rial). First, with the ground truth labels of semantic categories " + }, + { + "bbox": [ + 304, + 276, + 545, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GT}}^c" + }, + { + "bbox": [ + 304, + 276, + 545, + 324 + ], + "type": "text", + "content": ", back-propagation is performed for parameters optimization using Cross Entropy (CE) loss " + }, + { + "bbox": [ + 304, + 276, + 545, + 324 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 304, + 276, + 545, + 324 + ], + "type": "text", + "content": " for semantic categories classification:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 381, + 336, + 545, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 336, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 381, + 336, + 545, + 350 + ], + "type": "interline_equation", + "content": "L _ {c} = \\mathrm {C E} \\left(s _ {\\text {p r e d}} ^ {c}, \\mathcal {L} _ {\\mathrm {G T}} ^ {c}\\right), \\tag {7}", + "image_path": "105fcfb75ec628980425ca4b19348b05135ffee3326e70bccf32e8210f5a1672.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "text", + "content": " is the cross-entropy loss, and " + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{pred}}^c" + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "text", + "content": " is the prediction probability for each point. Then, the Binary Cross Entropy (BCE) loss " + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "inline_equation", + "content": "L_m" + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "text", + "content": " is used for motion states classification with motion states ground truth " + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GT}}^m" + }, + { + "bbox": [ + 304, + 355, + 545, + 404 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 376, + 415, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 415, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 376, + 415, + 545, + 430 + ], + "type": "interline_equation", + "content": "L _ {m} = \\operatorname {B C E} \\left(s _ {\\text {p r e d}} ^ {m}, \\mathcal {L} _ {\\mathrm {G T}} ^ {m}\\right), \\tag {8}", + "image_path": "c0147729487949870cf2f67edb85328ffa3dac796baaebe6dcd7a0b9f1c1f1b8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 434, + 530, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 530, + 446 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 530, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 434, + 530, + 446 + ], + "type": "inline_equation", + "content": "\\mathrm{BCE}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 306, + 434, + 530, + 446 + ], + "type": "text", + "content": " indicates the binary cross-entropy loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 447, + 530, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 447, + 530, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 447, + 530, + 459 + ], + "type": "text", + "content": "The final objective function " + }, + { + "bbox": [ + 317, + 447, + 530, + 459 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 317, + 447, + 530, + 459 + ], + "type": "text", + "content": " of the optimization is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 472, + 545, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 472, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 375, + 472, + 545, + 484 + ], + "type": "interline_equation", + "content": "L = \\omega_ {c} \\cdot L _ {c} + \\omega_ {m} \\cdot L _ {m}, \\tag {9}", + "image_path": "7b0abe267ebf605dcd04188455fa98c0694d78895b9295c09514f4ee9acef74a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\omega_{c}" + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\omega_{m}" + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "text", + "content": " are the weights of the two losses (" + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 306, + 490, + 545, + 514 + ], + "type": "text", + "content": ") respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "type": "text", + "content": "During inference, the final prediction result is determined using the logits produced by both classification heads. When presented with an input sample, the motion-aware classification head will identify the motion state of the input point only if the category-aware classification head recognizes the point as belonging to a class with the potential to move." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 609, + 388, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 388, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 388, + 624 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "Datasets and Evaluation Metric: We evaluate our method on SemanticKITTI [1] and nuScenes [3]. For SemanticKITTI, the multi-scan setting is fully supervised and contains 25-category (6 moving categories and 19 static categories) with high-quality semantic annotations. The annotations are based on the KITTI dataset [8]. It comprises 22 point cloud sequences. For nuScenes, we propose a new" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9376" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 44, + 71, + 216, + 144 + ], + "blocks": [ + { + "bbox": [ + 44, + 71, + 216, + 144 + ], + "lines": [ + { + "bbox": [ + 44, + 71, + 216, + 144 + ], + "spans": [ + { + "bbox": [ + 44, + 71, + 216, + 144 + ], + "type": "image", + "image_path": "fdfe8373ac4988da781f16f1c0956a6a63ea4126312dafc6113a1b385a806c8f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 45, + 144, + 216, + 217 + ], + "blocks": [ + { + "bbox": [ + 45, + 144, + 216, + 217 + ], + "lines": [ + { + "bbox": [ + 45, + 144, + 216, + 217 + ], + "spans": [ + { + "bbox": [ + 45, + 144, + 216, + 217 + ], + "type": "image", + "image_path": "4d17f7a62d66c608711c520f13094be453959a77160a44cd877c3f7b84086130.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 218, + 144, + 225 + ], + "lines": [ + { + "bbox": [ + 117, + 218, + 144, + 225 + ], + "spans": [ + { + "bbox": [ + 117, + 218, + 144, + 225 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 227, + 112, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 227, + 112, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 112, + 236 + ], + "type": "text", + "content": "- Car (non-moving)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 118, + 227, + 162, + 235 + ], + "lines": [ + { + "bbox": [ + 118, + 227, + 162, + 235 + ], + "spans": [ + { + "bbox": [ + 118, + 227, + 162, + 235 + ], + "type": "text", + "content": "- Car (moving)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 167, + 227, + 231, + 236 + ], + "lines": [ + { + "bbox": [ + 167, + 227, + 231, + 236 + ], + "spans": [ + { + "bbox": [ + 167, + 227, + 231, + 236 + ], + "type": "text", + "content": "- Person (non-moving)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 380, + 144 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 380, + 144 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 380, + 144 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 380, + 144 + ], + "type": "image", + "image_path": "ef1b4cf530ee238f319cf758950b4955c935fca600019284733d6472e36ba12a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 217, + 144, + 380, + 217 + ], + "blocks": [ + { + "bbox": [ + 217, + 144, + 380, + 217 + ], + "lines": [ + { + "bbox": [ + 217, + 144, + 380, + 217 + ], + "spans": [ + { + "bbox": [ + 217, + 144, + 380, + 217 + ], + "type": "image", + "image_path": "6459536973607d4ae86efe76562bc8a2051376f76b613a2d5794f878d975d292.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 217, + 290, + 225 + ], + "lines": [ + { + "bbox": [ + 238, + 217, + 290, + 225 + ], + "spans": [ + { + "bbox": [ + 238, + 217, + 290, + 225 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 227, + 351, + 236 + ], + "lines": [ + { + "bbox": [ + 238, + 227, + 351, + 236 + ], + "spans": [ + { + "bbox": [ + 238, + 227, + 351, + 236 + ], + "type": "text", + "content": "- Person (moving) - Bicyclist (moving)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 382, + 72, + 543, + 144 + ], + "blocks": [ + { + "bbox": [ + 382, + 72, + 543, + 144 + ], + "lines": [ + { + "bbox": [ + 382, + 72, + 543, + 144 + ], + "spans": [ + { + "bbox": [ + 382, + 72, + 543, + 144 + ], + "type": "image", + "image_path": "c6794fae2860a3c29e58d54b953e90821b595c1066010d124a5302281a143f02.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 381, + 144, + 543, + 217 + ], + "blocks": [ + { + "bbox": [ + 381, + 144, + 543, + 217 + ], + "lines": [ + { + "bbox": [ + 381, + 144, + 543, + 217 + ], + "spans": [ + { + "bbox": [ + 381, + 144, + 543, + 217 + ], + "type": "image", + "image_path": "d9e286c6dd5611c153be16992e85179a47d14f045a53e1e70969bffc6da74399.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 218, + 484, + 225 + ], + "lines": [ + { + "bbox": [ + 442, + 218, + 484, + 225 + ], + "spans": [ + { + "bbox": [ + 442, + 218, + 484, + 225 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 42, + 291, + 168, + 379 + ], + "blocks": [ + { + "bbox": [ + 46, + 241, + 546, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 546, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 546, + 275 + ], + "type": "text", + "content": "Figure 3. Qualitative results on SemanticKITTI [1] public validation dataset. With SPVCNN [25] as the backbone, the segmentation results on the SemanticKITTI multi-scan task of the baseline model and our model are shown together with ground truth. At the same time, a specific area containing moving points is magnified and displayed at the top right of each sub-figure." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 42, + 291, + 168, + 379 + ], + "lines": [ + { + "bbox": [ + 42, + 291, + 168, + 379 + ], + "spans": [ + { + "bbox": [ + 42, + 291, + 168, + 379 + ], + "type": "image", + "image_path": "effa9436ea92760b55b5cfb705532b140e11f20a8b33129899836d89e9f594b5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 42, + 379, + 168, + 466 + ], + "blocks": [ + { + "bbox": [ + 42, + 379, + 168, + 466 + ], + "lines": [ + { + "bbox": [ + 42, + 379, + 168, + 466 + ], + "spans": [ + { + "bbox": [ + 42, + 379, + 168, + 466 + ], + "type": "image", + "image_path": "ec414721ce044259c8e9d1599057931109e0412ea4c2266f888174f7b6dd6e8a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 468, + 129, + 475 + ], + "lines": [ + { + "bbox": [ + 102, + 468, + 129, + 475 + ], + "spans": [ + { + "bbox": [ + 102, + 468, + 129, + 475 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 480, + 287, + 536 + ], + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 536 + ], + "type": "text", + "content": "Figure 4. Evaluation errors (shown in red) by the baseline methods and MarS3D (using SPVCNN [25] as the backbone) on SemanticKITTI [1] public validation dataset. MarS3D significantly outperforms the baseline. Contrasting regions in the evaluation errors are highlighted with boxes and corresponding categories." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 169, + 292, + 286, + 378 + ], + "blocks": [ + { + "bbox": [ + 169, + 292, + 286, + 378 + ], + "lines": [ + { + "bbox": [ + 169, + 292, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 169, + 292, + 286, + 378 + ], + "type": "image", + "image_path": "59323c4b16e130f2985db3bb202931410b138e41c033a49a04be48319b6335c1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 169, + 379, + 286, + 466 + ], + "blocks": [ + { + "bbox": [ + 169, + 379, + 286, + 466 + ], + "lines": [ + { + "bbox": [ + 169, + 379, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 169, + 379, + 286, + 466 + ], + "type": "image", + "image_path": "1e4b35d0dd42719f3e854a9e94352af08214f10a6090b2f12a0c272fe4f96d90.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 468, + 238, + 475 + ], + "lines": [ + { + "bbox": [ + 222, + 468, + 238, + 475 + ], + "spans": [ + { + "bbox": [ + 222, + 468, + 238, + 475 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 559, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 666 + ], + "type": "text", + "content": "multi-scan setting based on the 'lidar-seg' task (16 semantic categories) without reference frame supervision. We use its object-level velocity to construct a multi-scan segmentation dataset with 24 categories (8 moving and 16 static categories). More details are provided in the supplementary material. To assess the effectiveness of our proposed method and make comparisons with baselines and other methods, we use the mean Intersection over Union (mIoU) as the evaluation metric." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": "Implementation Details: Our model is designed as a plug-and-play module that provides motion-aware features to enhance the backbone learned features. For the backbone" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "text", + "content": "model, we consider SPVCNN " + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "inline_equation", + "content": "[25]^*" + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "text", + "content": ", SparseConv " + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "inline_equation", + "content": "[9]^{\\dagger}" + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "text", + "content": ", and MinkUNet " + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "inline_equation", + "content": "[5]^{\\ddagger}" + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "text", + "content": ". Following previous works [12, 23, 36], we use the current and its previous two frames as input. The size of the BEV representation is set to " + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "inline_equation", + "content": "501 \\times 301" + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "text", + "content": ", and the multiple kernel sizes in the MAFL module are set as 1, 3, and 5, respectively. We set the embedded feature dimension in the CFFE module to 18, and the data augmentations are the same as the standard settings. All the models are trained on GeForce RTX 3090 GPUs, and the inference latency is recorded using a single GeForce RTX 3090 GPU." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 419, + 390, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 390, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 390, + 430 + ], + "type": "text", + "content": "5.1. Main Results" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 437, + 546, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 677 + ], + "type": "text", + "content": "Comparison with Baseline Methods: We evaluate the performance of our proposed method on multi-scan benchmarks of SemanticKITTI [1] and nuScenes [3]. The baseline method uses the same backbone to process multiscan point clouds. We then compare this baseline approach to the same backbone augmented with MarS3D for a fair evaluation. As shown in Table 1, MarS3D significantly improves the performance over baseline methods on the public validation set of SemanticKITTI. With the most lightweight network, SPVCNN [25], MarS3D brings a " + }, + { + "bbox": [ + 304, + 437, + 546, + 677 + ], + "type": "inline_equation", + "content": "4.96\\%" + }, + { + "bbox": [ + 304, + 437, + 546, + 677 + ], + "type": "text", + "content": " improvement while introducing less than " + }, + { + "bbox": [ + 304, + 437, + 546, + 677 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 304, + 437, + 546, + 677 + ], + "type": "text", + "content": " additional parameters. Particularly, consistent performance gains are observed on the dynamic object classes with non-moving/moving properties (i.e., car & moving car, person & moving person). This shows that the proposed BEV branch is both lightweight and powerful. Further, to verify the generalizability of the model, we offer a multiscan task based on nuScenes 'lidar-seg' dataset [3]. Our method outperforms the baseline (using MinkUNet [5] as the backbone) by a significant margin, achieving " + }, + { + "bbox": [ + 304, + 437, + 546, + 677 + ], + "type": "inline_equation", + "content": "64.83\\%" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 683, + 518, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 683, + 502, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 683, + 502, + 693 + ], + "spans": [ + { + "bbox": [ + 315, + 683, + 502, + 693 + ], + "type": "text", + "content": "\\*https://github.com/mit-han-lab/spvnas" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 694, + 501, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 694, + 501, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 694, + 501, + 703 + ], + "type": "text", + "content": "https://github.com/traveller59/spconv" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 704, + 518, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 704, + 518, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 704, + 518, + 712 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 318, + 704, + 518, + 712 + ], + "type": "text", + "content": "https://github.com/NVIDIA/MinkowskiEngine" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9377" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 72, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 545, + 186 + ], + "type": "table", + "html": "
MethodmIoUcarbic.mot.tru.ove.per.bil.mol.roa.par.sid.ogr.bui.fen.veg.trn.ter.pol.tra.mca.mbi.mpe.mmm.mov.mtr.
SpSequenceNet [24]43.188.524.026.229.222.76.30.00.090.157.973.927.191.266.884.066.065.750.848.753.241.226.236.22.30.1
TemporalLidarSeg [7]47.092.147.740.939.235.014.40.00.091.859.675.823.289.863.882.362.564.752.660.468.242.840.412.912.42.1
TemporalLatticeNet [23]47.191.635.436.126.923.09.40.00.091.559.375.327.589.665.384.666.770.457.260.459.741.751.048.85.90.0
Meta-RangeSeg [28]49.590.152.743.930.335.414.30.00.090.763.374.726.990.563.583.067.067.756.464.464.556.155.024.420.33.4
KPConv [26]51.293.744.947.243.538.621.60.00.086.558.470.526.790.864.584.670.366.057.053.969.467.467.547.24.75.8
Baseline49.289.839.434.039.421.08.91.80.089.162.072.412.990.563.984.668.468.758.960.169.363.558.756.59.53.6
Ours52.795.149.249.539.736.616.21.20.089.966.874.326.492.168.286.072.170.562.864.878.467.358.036.310.05.1
", + "image_path": "249e57d9c257023b89b2b83e5d324dbd5dfb9dc1ce1312f6795425f3dfa09a58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 188, + 546, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 546, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 546, + 222 + ], + "type": "text", + "content": "Table 2. Comparison with the state-of-the-art models on SemanticKITTI multi-scan benchmark (official test set). MarS3D (with SPVCNN [25] as the backbone) significantly outperforms these models for multi-scan tasks. (full names of the categories are in the supplementary material.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 239, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 275 + ], + "type": "text", + "content": "mIoU compared to the baseline's " + }, + { + "bbox": [ + 46, + 239, + 287, + 275 + ], + "type": "inline_equation", + "content": "61.90\\%" + }, + { + "bbox": [ + 46, + 239, + 287, + 275 + ], + "type": "text", + "content": ". This improvement is achieved with a negligible increase in each-frame inference time from 53ms to 58ms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "type": "text", + "content": "Comparison with State-of-the-Art Methods: Compared to various models applied to multi-scan tasks, MarS3D (with SPVCNN as the backbone) is evaluated on the SemanticKITTI multi-scan benchmark. As shown in Table 2, the proposed approach has demonstrated superior performance, with a " + }, + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 283, + 287, + 390 + ], + "type": "text", + "content": " increase in mIoU compared to the current state-of-the-art method, KPCov [26]. Furthermore, our method performs similarly or better than other state-of-the-art models across nearly all categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 398, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 590 + ], + "type": "text", + "content": "Qualitative Comparisons: Quantitative results are shown in Figure 3. The baseline model suffers from mistaking the status of static cars (Figure 3: Scan A) and moving persons (Figure 3: Scan B). It even recognizes the moving bicyclist as a moving person (Figure 3: Scan A). In contrast, MarS3D can circumvent such category and motion state discrimination errors. In addition to achieving better qualitative results in motion states, MarS3D also outperforms the baseline in semantic categories. We visualize the error maps (errors are shown in red) of the baseline and our method in Figure 4, where bounding boxes indicating specific categories highlight the differences between our method and the baseline. These improvements indicate that MarS3D has stronger semantic feature extraction capabilities compared to the baseline model, resulting in better segmentation performance even for immobile objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 596, + 148, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 148, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 148, + 608 + ], + "type": "text", + "content": "5.2. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "type": "text", + "content": "In this section, we conduct comprehensive ablation experiments on the SemanticKITTI multi-scan validation set to examine the effects of each component in our proposed method. As shown in Table 3, we gradually add three components to the baseline method, including the CFFE module, vanilla BEV branch (introduce only BEV representa" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 239, + 417, + 321 + ], + "blocks": [ + { + "bbox": [ + 313, + 239, + 417, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 417, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 417, + 321 + ], + "type": "image", + "image_path": "81e0a066f37e0e9938026123e3e86a55dfd59577493f3c276d3dcc4cf746e8b3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 323, + 417, + 403 + ], + "blocks": [ + { + "bbox": [ + 313, + 323, + 417, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 417, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 417, + 403 + ], + "type": "image", + "image_path": "e245684290a05cce033d6b53c5e58de8f0b1f8a6f819f95d9dd6e3ccb695cbd2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 405, + 413, + 414 + ], + "lines": [ + { + "bbox": [ + 328, + 405, + 413, + 414 + ], + "spans": [ + { + "bbox": [ + 328, + 405, + 413, + 414 + ], + "type": "text", + "content": "BEV Ground Truth (Motion)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 419, + 239, + 517, + 321 + ], + "blocks": [ + { + "bbox": [ + 419, + 239, + 517, + 321 + ], + "lines": [ + { + "bbox": [ + 419, + 239, + 517, + 321 + ], + "spans": [ + { + "bbox": [ + 419, + 239, + 517, + 321 + ], + "type": "image", + "image_path": "44dbee35b43054220618e456657175e3fb700ca7c258b47830aec9dacb8643c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 419, + 323, + 516, + 403 + ], + "blocks": [ + { + "bbox": [ + 419, + 323, + 516, + 403 + ], + "lines": [ + { + "bbox": [ + 419, + 323, + 516, + 403 + ], + "spans": [ + { + "bbox": [ + 419, + 323, + 516, + 403 + ], + "type": "image", + "image_path": "fae1f5c467451ab53fc932b8b3446bf8c7e2080091e639c8a1a6452fe74f2c75.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 405, + 476, + 413 + ], + "lines": [ + { + "bbox": [ + 460, + 405, + 476, + 413 + ], + "spans": [ + { + "bbox": [ + 460, + 405, + 476, + 413 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 415, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 415, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 545, + 449 + ], + "type": "text", + "content": "Figure 5. Sampled discrepant feature maps demonstrate that the region that contains moving points are represented higher activation values than other regions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 468, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 492 + ], + "type": "text", + "content": "tions and 2D CNN without the MAFL module), and the MAFL module to illustrate the effectiveness of our designs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 501, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 501, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 501, + 545, + 586 + ], + "type": "text", + "content": "Effectiveness of CFFE Module: As shown in Table 3, we first employ only the CFFE module with the baseline model and observe a significant performance improvement. For instance, with SPVCNN [25] as the backbone, the introduction of the CFFE module led to a boost in performance by " + }, + { + "bbox": [ + 305, + 501, + 545, + 586 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 305, + 501, + 545, + 586 + ], + "type": "text", + "content": " in terms of mIoU. This confirms the efficacy of the CFFE module for multi-frame representation learning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Effectiveness of BEV Representation: We also observe a significant performance improvement when we add the vanilla BEV branch alone to the baseline model. Using SPVCNN [25] as the backbone, the BEV branch results in a " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " increase in mIoU. This confirms the importance of BEV representations in extracting motion-aware information. Furthermore, when both the CFFE module and the vanilla BEV branch are added to the baseline model, there is a further increase in performance, demonstrating that the two components complement each other." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": " http://wwwsemantic-kitti.org/tasks - Semantic Segmentation - Multiple Scans" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9378" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 72, + 285, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 285, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 285, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 285, + 159 + ], + "type": "table", + "html": "
MethodCFFEBEVMAFLmIoU(%)
baseline---49.7 / 49.0 / 48.5
proposed✓--50.8 / 51.3 / 52.1
-✓-53.4 / 53.9 / 53.1
✓✓-53.6 / 54.2 / 54.1
✓✓✓54.7 / 54.6 / 54.7
", + "image_path": "5e46666dda373ed0282cc09b7d208dbb1148d705eb6d4e0e20db91670110bca0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 217, + 284, + 243 + ], + "blocks": [ + { + "bbox": [ + 47, + 163, + 287, + 208 + ], + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 208 + ], + "type": "text", + "content": "Table 3. Ablation studies on different backbones (SPVCNN / SparseConv / MinkUNet) on SemanticKITTI [1] public validation dataset. The effectiveness of different designs is demonstrated step-by-step and our method is marked with a gray box." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 217, + 284, + 243 + ], + "lines": [ + { + "bbox": [ + 52, + 217, + 284, + 243 + ], + "spans": [ + { + "bbox": [ + 52, + 217, + 284, + 243 + ], + "type": "table", + "html": "
MethodVanilla BEVTDNet [13]STM [18]Ours
mIoU(%)53.3653.56 (+0.20)53.64 (+0.28)54.66 (+1.30)
", + "image_path": "a7f02ca5d471d4e8beb578d5fa31878cfc3b1f78ea6f6d53a94d687ba0607b8a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 247, + 286, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 286, + 281 + ], + "type": "text", + "content": "Table 4. Comparing our method (using SPVCNN [25] as the backbone) with different 2D temporal semantic segmentation approaches the public validation dataset of SemanticKITTI [1]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 300, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 563 + ], + "type": "text", + "content": "Effectiveness of MAFL Module: In the final experiment presented in Table 3, we study the impact of our proposed MAFL module on the BEV representations, resulting in an increase of around " + }, + { + "bbox": [ + 46, + 300, + 287, + 563 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 46, + 300, + 287, + 563 + ], + "type": "text", + "content": " from the baseline model (using SPVCNN [25] as the backbone). By stacking all the proposed components together, our final solution (marked with a grey box in Table 3) reaches its pinnacle in performance. The improvement in other backbones is also significant, and more statistical results are provided in the supplementary material. In addition, the multi-channel discrepant feature map represents the differences between the extracted feature maps of BEV representations from two point clouds. For a discrepant feature map generated during inference, we randomly sample the channels pixel by pixel. The average absolute activation values of pixels with the same motion state in the sampled channels are re-scaled and aggregated into a single-channel feature map as shown in Figure 5. The sampled discrepant feature map has higher activation on pixels containing moving points compared to other regions. This clearly demonstrates that the MAFL module clearly distinguishes between regions that contain moving points and those that do not." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 285, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 285, + 587 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 285, + 587 + ], + "type": "text", + "content": "5.3. More Comparison on Temporal Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "As for BEV representation learning, the proposed MAFL module is utilized on extracted 2D feature maps for capturing inter-frame temporal information of point clouds. We further conduct comparative experiments by replacing the MAFL module with its 2D counterparts, i.e., STM [18] and TDNet [13]. These 2D modules have been widely utilized for preserving temporal information in 2D tasks. The results of the comparative experiments using SPVCNN [25] as the backbone are presented in Table 4. According to the evaluation results, it can be concluded that the MAFL module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "shows superior performance compared to the other models, thereby confirming its remarkable effectiveness. This shows that the proposed MAFL module is better suited for the specific task of handling motion-aware semantic segmentation in 3D point clouds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 142, + 511, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 142, + 511, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 511, + 155 + ], + "type": "text", + "content": "5.4. Limitations and Failure Cases Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 161, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 161, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 545, + 353 + ], + "type": "text", + "content": "Although MarS3D demonstrates impressive overall performance on the SemanticKITTI multi-scan benchmark, some limitations can be identified from the quantitative results in Table 1 and Table 2. The SemanticKITTI training dataset has an imbalanced distribution of point categories (more details are included in the supplementary material), which causes MarS3D to perform poorly on long-tailed categories due to insufficient training data on these categories. As shown in Table 2, MarS3D fails to show effects on several long-tail categories (i.e., bicyclist, motorcyclist, moving-other-vehicle, and moving-truck). Exploring solutions to address the long-tail problem is a promising research direction for the work. Since our model assumes planar motion, it may not perform well in scenarios where objects move in non-planar ways, such as on steep terrains or on non-planar surfaces." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 366, + 378, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 378, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 378, + 379 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 387, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 615 + ], + "type": "text", + "content": "In this paper, we propose MarS3D, a novel plug-and-play motion-aware model for 3D multi-scan point cloud semantic segmentation. The Motion-Aware Feature Learning (MAFL) module, based on BEV representations, is designed to facilitate extracting motion-aware representations. Additionally, the Cross-Frame Feature Embedding (CFFE) module is introduced to improve representation learning by embedding time-step information into features, thus preserving rich temporal information. Extensive experiments and ablation studies demonstrate that MarS3D significantly improves multiple 3D semantic segmentation baselines while introducing minimal overheads. In comparison to state-of-the-art methods designed for multi-scan tasks, MarS3D achieves superior performance and offers significant improvements over baseline methods. The proposed MarS3D model demonstrates the potential for effectively incorporating motion-awareness into 3D point cloud semantic segmentation tasks, providing a strong foundation for further research and development in this area." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments: This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant no. 17202422), and RGC matching fund scheme (RMGS). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9379" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF International Conf. on Computer Vision (ICCV), 2019. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Goutam Bhat, Felix Järemo Lawin, Martin Danelljan, Andreas Robinson, Michael Felsberg, Luc Van Gool, and Radu Timofte. Learning what to learn for video object segmentation. In European Conference on Computer Vision, pages 777-794. Springer, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 334 + ], + "type": "text", + "content": "[4] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 335, + 288, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 389 + ], + "type": "text", + "content": "[5] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "type": "text", + "content": "[6] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "type": "text", + "content": "[7] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 2020 International Conference on 3D Vision (3DV), pages 781-790. IEEE, 2020. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 545 + ], + "type": "text", + "content": "[8] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 546, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 546, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 546, + 288, + 601 + ], + "type": "text", + "content": "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[10] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[11] Noureldin Hendy, Cooper Sloan, Feng Tian, Pengfei Duan, Nick Charchut, Yuesong Xie, Chuang Wang, and James Philbin. Fishing net: Future inference of semantic heatmaps in grids. arXiv preprint arXiv:2006.09917, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[12] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In Proceedings of the IEEE/CVF Con" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 8479-8488, 2022. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 152 + ], + "type": "text", + "content": "[13] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks for fast video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8827, 2020. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 153, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 547, + 207 + ], + "type": "text", + "content": "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "text", + "content": "[15] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. In European Conference on Computer Vision, pages 352-368. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "text", + "content": "[16] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 299, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 352 + ], + "type": "text", + "content": "[17] Jiaxu Miao, Yunchao Wei, and Yi Yang. Memory aggregation networks for efficient interactive video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10366-10375, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "type": "text", + "content": "[18] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226–9235, 2019. 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[19] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "type": "text", + "content": "[20] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "type": "text", + "content": "[21] Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7406-7415, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "text", + "content": "[22] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "text", + "content": "[23] Peer Schutt, Radu Alexandru Rosu, and Sven Behnke. Abstract flow for temporal semantic segmentation on the permutohedral lattice. In 2022 International Conference on Robotics and Automation (ICRA), pages 5139-5145. IEEE, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[24] Hanyu Shi, Guosheng Lin, Hao Wang, Tzu-Yi Hung, and Zhenhua Wang. Spsequencenet: Semantic segmentation network on 4d point clouds. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9380" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "conference on computer vision and pattern recognition, pages 4574-4583, 2020. 1, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[25] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "text", + "content": "[26] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[28] Song Wang, Jianke Zhu, and Ruixiang Zhang. Meta-rangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. arXiv preprint arXiv:2202.13377, 2022. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "text", + "content": "[29] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[30] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9621-9630, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "type": "text", + "content": "[31] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[32] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3173-3182, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[33] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3101-3109, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "text", + "content": "[34] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[35] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[36] Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 174 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550, 2020. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[37] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[38] Zhenhong Zou and Yizhe Li. Efficient urban-scale point clouds segmentation with bev projection. arXiv preprint arXiv:2109.09074, 2021. 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9381" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_content_list.json b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d850512bd868ef9ed789ca722c85eb52515e611b --- /dev/null +++ b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_content_list.json @@ -0,0 +1,1503 @@ +[ + { + "type": "text", + "text": "Marching-Primitives: Shape Abstraction from Signed Distance Function", + "text_level": 1, + "bbox": [ + 119, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weixiao Liu $^{1,2}$ Yuwei Wu $^{1}$ Sipu Ruan $^{1}$ Gregory S. Chirikjian $^{1*}$ $^{1}$ National University of Singapore $^{2}$ Johns Hopkins University \n{mpewxl, yw.wu, ruansp, mpegre}@nus.edu.sg", + "bbox": [ + 207, + 179, + 767, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Representing complex objects with basic geometric primitives has long been a topic in computer vision. Primitive-based representations have the merits of compactness and computational efficiency in higher-level tasks such as physics simulation, collision checking, and robotic manipulation. Unlike previous works which extract polygonal meshes from a signed distance function (SDF), in this paper, we present a novel method, named Marching-Primitives, to obtain a primitive-based abstraction directly from an SDF. Our method grows geometric primitives (such as superquadrics) iteratively by analyzing the connectivity of voxels while marching at different levels of signed distance. For each valid connected volume of interest, we march on the scope of voxels from which a primitive is able to be extracted in a probabilistic sense and simultaneously solve for the parameters of the primitive to capture the underlying local geometry. We evaluate the performance of our method on both synthetic and real-world datasets. The results show that the proposed method outperforms the state-of-the-art in terms of accuracy, and is directly generalizable among different categories and scales. The code is open-sourced at https://github.com/ChirikjianLab/Marching-Primitives.git.", + "bbox": [ + 75, + 301, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed great progress in the areas of 3D shape representation and environmental perception. Low-level representations such as surface meshes, point clouds, and occupancy grids are widely used as inputs to high-level computer vision algorithms and artificial intelligence tasks. They have the advantage of being able to represent and visualize objects with high accuracy and rich local geometric features. However, the low-level representations are ineffective in delivering a general and intuitive sense of structural geometry as well as part-level scene understanding. Studies [3,20] show that human vision, unlike", + "bbox": [ + 75, + 707, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57f1bb3b41f9000f06445b665e2541aec27b71d587e27ff0fb981e413e3f9432.jpg", + "image_caption": [ + "Figure 1. Primitive-base representation versus mesh. For each pair of objects, the left one is the superquadratic abstraction obtained by our algorithm, and the right one is the original mesh. The mesh of the chair is 6MB in size, while our representation only needs 4KB. An SDF representation discretized on a $128^{3}$ voxel grid occupies 19MB. Our abstraction is equivalent to an implicit continuous SDF, which is an approximation to the discrete SDF." + ], + "image_footnote": [], + "bbox": [ + 514, + 270, + 877, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "computer vision, tends to perceive and understand scenes as combinations of simple primitive shapes. Human beings perform well and robustly in complex tasks, providing a basic geometric description of the scene is available [32]. Therefore, researchers turn to exploring the possibility of interpreting complex objects and scenes with basic geometric primitives. Taking advantage of the primitive-based representation, many higher-level tasks, such as segmentation [14, 16, 21, 30], scene understanding [29, 31, 41, 47], grasping [33, 44, 45] and motion planning [35, 36], are able to be solved efficiently.", + "bbox": [ + 496, + 477, + 892, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, it still remains challenging to extract primitive-based abstractions from low-level representations. Starting from the 1990s, Solina et al. [1, 17, 39] aim to extract a single superquadric representation from a simple object by minimizing the least-square error between the primitive and the measured points. Later in [7, 22], their method is extended to represent more complex objects with multiple primitives. More recently, the authors of [24, 47] reformulate the task as a probabilistic inference problem with enhanced accuracy and robustness to noise and outliers. At the same time, with the surge of data-driven techniques, researchers attempt to train neural networks to infer cuboids [27,38,41,48,50] and superquadrics [29,31] representations in an end-to-end fashion. However, both the computational and learning-based approaches have their own limitations. The computational methods are vulnerable to the inherent ambiguity of the point-to-surface relationship. For exam", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8771", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ple, the algorithms tend to fill empty spaces of a non-convex object with primitives by mistake, due to the inside/outside ambiguity of a surface depicted by a set of points [24, 48]. The main drawback of the learning approaches lies in the lack of generalizability beyond the object category on which the model is trained [24, 31, 47, 48]. Also, the shape abstraction accuracy is inferior to the computational methods.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The signed distance function (SDF) has been a successful 3D volumetric representation in varieties of computer vision and graphics tasks. It is the basic framework for many classic 3D reconstruction algorithms such as TSDF volume reconstruction [10, 15], KinectFusion [19], and Dynamic-Fusion [26]. Recently, the SDF representation is adapted to the deep learning frameworks, and exhibits boosted potentials in shape encoding [8, 18, 28, 43], surface reconstruction [23, 46], and shape completion [11, 12, 34]. Usually, triangular mesh surfaces are extracted from the SDF representation with the marching cubes algorithm [25]. Point cloud and occupancy grid representations are also obtained by keeping the vertices of the meshes and the sign of each voxel point, respectively. The SDF is among the most informative 3D representations since it encodes not only the surface geometry but also the distance and side of a point relative to the shape. Meanwhile, it is easily achievable via range images from 3D sensors [10], or learnable from other input modalities [8, 18, 28]. Since we are able to extract meshes from an SDF, it is natural to think about the possibility of extracting primitives as well. Furthermore, the primitive-based abstraction is a continuous interpretation of the complete geometric information encoded in the original discrete SDF, but requires much less storage size (Fig.1).", + "bbox": [ + 75, + 200, + 470, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the aforementioned facts and the bottleneck of the current shape abstraction algorithms, we proposed a general shape abstraction method by reasoning directly on the informative SDF representation. The goal of our method is to find a combination of geometric primitives whose underlying SDF values match the target values evaluated on the evenly spaced discrete grid points (Sec. 3.1 and Sec. 3.2). To solve this problem, we propose a two-step iterative algorithm called the Marching-Primitives. Our algorithm 'marches' on two domains: the signed distance domain and the voxelized space domain, alternately. Firstly, the connectivity of volumes are analyzed by generating isosurfaces on a sequence of decreasing levels of negative signed distances (Sec.3.3). By doing so, volumes of interest (VOIs) where primitives are likely to be encoded can be identified sequentially. In the second step, for each of the VOIs, our algorithm marches on the neighbouring voxels to infer their probabilistic correspondences to the primitive and simultaneously optimizes the shape and pose of the primitive (Sec.3.4). After the primitive representation of a VOI is achieved, the fitted volumes are deactivated from the voxel grid. Our algorithm continues marching on the signed", + "bbox": [ + 75, + 568, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distance domain until it approaches zero, i.e., all the interior volumes of the SDF have been captured by the recovered primitives. We compare our algorithm with the state-of-the-art of both the computational and learning-based approaches on the ShapeNet object dataset [6] and D-FAUST human shape dataset [4] (Sec. 4.1). We also study the performance of our algorithm on different conditions (Sec. 4.2). Finally, we demonstrate the scene abstraction result of the Stanford Reading Room [49], which contains several pieces of furniture of various categories (Sec. 4.3).", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 256, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SDF Representation: The SDF can be stored in two different ways: discrete or continuous. A majority of computer vision and graphics algorithms are built on the SDF discretized on a 3D grid of voxel points. The signed distances are stored on each of the corresponding voxel points. The authors of [10, 19, 26] pioneer in fusing several noisy range images into a single discrete SDF. Their work is widely applied in 3D reconstruction and plays an important role in robotics tasks such as simultaneous localization and mapping. The discrete SDF is also a promising input/output representation for 3D deep learning [11, 12, 34]. Recently, it becomes popular to encode shapes as a continuous SDF with neural networks [28]. Vasu et al. [43] further improve the shape encoding quality by enforcing local regularities with geometric primitives. In [8], the authors adopt a two-stage meta-learning approach to further extend the generalization capabilities of neural SDF. With the deep neural network, it becomes possible to infer SDF representations from partially observed 3D inputs or even images. Both the discrete and the continuous SDF are implicit representations of geometric surfaces. To extract the explicit surface from the SDF, continuous SDFs need to be discretized on a voxel grid first and then conduct the marching cubes [25]. This method allows high-quality rendering of the objects, however, surface meshes are non-sparse and contain no structural level information. Our method provides an alternative approach to describe the underlying object in the SDF. Instead of meshes, we directly extract a collection of sparsely parameterized primitives from the SDF. Other than that, our primitive-based representation itself is also a concise yet continuous SDF approximation to the original discrete SDF.", + "bbox": [ + 496, + 280, + 890, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computational Shape Abstraction: The most well-studied primitive for computational shape abstraction is the superquadric, due to its extensive shape vocabulary including cuboids, ellipsoids, cylinders, octohedra, and many shapes in between (e.g., cuboids with rounded edges). It is first proposed as a versatile modeling element for complex objects in computer graphics [2, 32]. Later, Solina et al. propose a method to conduct abstraction of simple objects from range images with a single superquadric [17, 39]. Leonardis et al. [22] and Chevalier et al. [7] further extend", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8772", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the previous work to recover complex objects with multiple superquadrics with a Split-and-Merge strategy. A numerical instability problem is addressed and revisited in [42]. The authors introduce an auxiliary function in the unstable region and receive a better abstraction accuracy. More recently, Liu et al. [24] formulate the problem in a probabilistic fashion and propose a geometric strategy to avoid local optimum, bringing a significant improvement in robustness to outlier and fitting accuracy. Wu et al. [47] extend and recast the work as a nonparametric Bayesian inference problem so as to improve the applicability on complex shapes. To the best of the authors' knowledge, the existing computational methods are all based on range images or point clouds, which suffer from geometric ambiguities [48]. In contrast, our method takes advantage of the abundant geometric information encoded in the SDF and is easily compatible with other computer vision algorithms based on the SDF representation.", + "bbox": [ + 75, + 90, + 472, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based Shape Abstraction: The learning-based method is first seen in [41]. Tulsiani et al. propose a 3D convolutional neural network (CNN) to learn shape abstractions with cuboids from the occupancy grid. Sun et al. [40] design an adaptive hierarchical cuboid representation and introduce an unsupervised approach to learn to extract the parameters for the representation. Yang et al. [48] train a variational auto-encoder network to transform point clouds into parametric cuboids. Other than cuboids, researchers also seek to extract spheres and ellipsoids representations from objects. Hao et al. [18] combine the neural SDF with the spherical representation by sharing a same latent layer. In [37], ellipsoids or cuboids are extracted to help segment the input point cloud. However, a single type of primitive has very limited expressiveness. Therefore, Paschalidou et al. [29, 31] turn to training neural networks to conduct abstractions with the superquadrics as the atomic elements. Learning-based approaches are versatile in dealing with different input sources. They are able to make shape abstractions from point clouds, voxel grids, or even RGB images which are so ill-conditioned that the computational approaches have little chance of working. However, learning-based approaches rely heavily on the training dataset and thus are less generalizable to unseen categories. Instead, our approach reasons about the primitive abstraction from a case-by-case geometric perspective, which provides an inherent advantage in generalizability and accuracy.", + "bbox": [ + 75, + 364, + 472, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 804, + 169, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 830, + 205, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The discrete SDF is a volumetric surface representation built on a voxel grid $\\mathbf{V} = \\{\\mathbf{x}_i\\in \\mathbb{R}^3,i = 1,2,\\dots,N\\}$ . A scalar $d(\\mathbf{x}_i)$ is assigned to each grid point, which indicates", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the signed distance of $\\mathbf{x}_i$ to the nearest surface. The point $\\mathbf{x}_i$ lies inside the surface if $d(\\mathbf{x}_i) < 0$ and outside otherwise. Typically, the surface mesh of an object is extracted from the SDF by the marching cubes algorithm [25]. In this paper, we call the input SDF $d(\\mathbf{x}_i)$ the target SDF.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the primitive representation, we select the superquadrics [2], a family of geometric primitives defined by the implicit equation", + "bbox": [ + 496, + 167, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {x}) = \\left(\\left(\\frac {x}{a _ {x}}\\right) ^ {\\frac {2}{\\varepsilon_ {2}}} + \\left(\\frac {y}{a _ {y}}\\right) ^ {\\frac {2}{\\varepsilon_ {2}}}\\right) ^ {\\frac {\\varepsilon_ {2}}{\\varepsilon_ {1}}} + \\left(\\frac {z}{a _ {z}}\\right) ^ {\\frac {2}{\\varepsilon_ {1}}} = 1 \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 218, + 890, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The superquadric family is only encoded by 5 parameters (shape parameters $\\epsilon_{1},\\epsilon_{2}\\in [0,2]\\subset \\mathbb{R}$ , and scale parameters $a_{x},a_{y},a_{z}\\in \\mathbb{R}_{>0}$ ), but has an extensive shape vocabulary. Note that the shape parameters can exceed 2, resulting in nonconvex shapes. However, practically in most studies [24,31] and also in our paper, we limit them within the convex region as defined above. The points $\\mathbf{x} = [x,y,z]\\in \\mathbb{R}^3$ satisfying Eq. (1) form the surface of the superquadric. In this paper, we also include the Euclidean transformation $g\\in SE(3)$ , i.e. 3 Euler angles for rotation $\\mathbf{R}\\in SO(3)$ and 3-dimensional translation $\\mathbf{t}\\in \\mathbb{R}^3$ , to parameterize a superquadric with a general pose. In total, we denote a superquadric with a vector $\\pmb{\\theta}$ of 11 elements. According to Eq. (1), we can approximate the signed distance of a grid point $\\mathbf{x}_i$ to a general posed superquadric $\\pmb{\\theta}$ explicitly by", + "bbox": [ + 496, + 270, + 892, + 498 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right) = \\left(1 - f ^ {- \\frac {\\epsilon_ {1}}{2}} \\left(g ^ {- 1} \\circ \\mathbf {x} _ {i}\\right)\\right) \\| g ^ {- 1} \\circ \\mathbf {x} _ {i} \\| _ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 503, + 890, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are not able to use the exact SDF of superquadrics, because it has no analytical solution and is only achievable by numerical optimization [5]. Eq.(2) is the radial distance [17, 24] of $\\mathbf{x}_i$ to the superquadric surface, which converges to the exact signed distance as $d_{\\theta}(\\mathbf{x}_i)\\to 0$ . Therefore, we truncate both the input target SDF $d(\\cdot)$ and the primitive SDF $d_{\\theta}(\\cdot)$ within the vicinity of zero to ensure the approximation accuracy.", + "bbox": [ + 496, + 536, + 890, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 664, + 702, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain a primitive-based abstraction from the target SDF, we seek a combination of primitives $\\Theta \\doteq \\{\\theta_{k}, k = 1,2,\\dots,K\\}$ whose underlying signed distances measured on the voxel points (we call it the source SDF in contrast to the target SDF) match the target SDF, that is", + "bbox": [ + 496, + 686, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Theta} = \\underset {\\boldsymbol {\\Theta}} {\\arg \\min } \\sum_ {\\mathbf {x} _ {i} \\in \\mathbf {V}} \\underset {\\boldsymbol {\\theta} _ {k} \\in \\boldsymbol {\\Theta}} {\\min } \\left\\| d _ {\\boldsymbol {\\theta} _ {k}} \\left(\\mathbf {x} _ {i}\\right) - d \\left(\\mathbf {x} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 768, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our problem formulation is simple and intuitive, however, it is intractable to solve directly. First, the number of primitives $K$ needed is unknown. Also, the correspondences between the voxel points $\\mathbf{x}_i$ and the primitives $\\theta_k$ are unknown a priori. Therefore, it makes Eq.(3) a chicken-and-egg problem: on the one hand, if we know the set of voxel", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8773", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "points contributing to a same primitive, we are able to solve for the parameters of the primitive by optimization; on the other hand, if we know the configurations of the primitives, we are able to tell the belongings of each voxel. Other than those factors, the computation complexity itself is prohibitively high, because we need to evaluate hundreds of thousands or even millions of voxel points. This makes it infeasible to operate directly on the complete dense voxel grid. To tackle these difficulties, we propose an iterative algorithm, the Marching-Primitives, which simultaneously solves the correspondences and primitive parameters efficiently. It is worth noting that our method is generalizable beyond the superquadrics. Any volumetric primitives with easily accessible SDF representations can be adapted to our framework.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Iterative Connectivity Marching", + "text_level": 1, + "bbox": [ + 76, + 327, + 361, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of setting a predefined number of primitives, our method starts from an empty set of primitives and grows as needed. This is realized by analyzing the connectivity (in this paper, we use the 26-connectivity) of the voxels at different levels of signed distance. Given the target SDF, our algorithm checks the connectivity of voxels whose signed distance is less than a sequence of thresholds", + "bbox": [ + 75, + 349, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {c} \\doteq \\left\\{t _ {1} ^ {c}, t _ {2} ^ {c}, \\dots \\right\\}, \\quad t _ {1} ^ {c} = \\min _ {\\mathbf {x} _ {i} \\in \\mathbf {V}} d (\\mathbf {x} _ {i}), \\quad t _ {m + 1} ^ {c} = \\alpha t _ {m} ^ {c} (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 467, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha \\in (0,1) \\subset \\mathbb{R}$ is the common ratio, resulting in a geometric sequence exponentially decaying to zero. Each threshold $t_m^c$ defines a set of disjoint isosurfaces $S_m$ , encompassing the connected voxels whose signed distances are less than the threshold.", + "bbox": [ + 75, + 502, + 468, + 577 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {m} = \\left\\{S _ {k}, k = 1, 2, \\dots , | S _ {m} | \\right\\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 590, + 468, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$|S_{m}|$ denotes the number of the disjoint isosurfaces. We construct a subset $\\bar{S}_m$ from $S_{m}$", + "bbox": [ + 76, + 618, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {S} _ {m} = \\left\\{S _ {k} \\in S _ {m}, | S _ {k} | \\geq N _ {c} \\right\\} \\subseteq S _ {m} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 660, + 468, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $|S_{k}|$ is the number of the connected voxels within the isosurface $S_{k}$ . $\\bar{S}_{m}$ is the set of the isosurfaces encompassing no less than $N_{c}$ connected voxels. The connectivity marching starts from the innermost threshold $t_1^c$ , where the resulting $\\bar{S}_1$ might be empty. The marching continues on the sequence (increasing the threshold) until $\\bar{S}_{m}$ is nonempty. Each of the connected volumes in $\\bar{S}_{m}$ (we call it a VOI) is an ideal starting point for growing a primitive, because: (1) those volumes are among the most interior of the target SDF, corresponding to most prominent part of the geometry; (2) disconnected and weakly connected volumes can be separated apart; (3) the SDF of a primitive can also be interpreted as layers of isosurfaces, and thus sharing similar geometric structure; and furthermore (4) the size of the", + "bbox": [ + 75, + 688, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/541b8c21aba2400bafdf17b5742b2ff63e2ba1278d04bf634b9a29e005fb5c9f.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 666, + 179 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/733e9343c28217bbf6105e987c28d06a8731b785debb53cfbfeecbd6fe5d6222.jpg", + "image_caption": [ + "Figure 2. Visualizations of concepts. (a) A 2D illustration of two marched isosurfaces separating the shape into two VOIs. (b) A simple example of the Marching-Primitives algorithm. The first figure is the input target SDF (voxels with signed distances larger than the truncation threshold are left blank for visualization). Figures b2-b4 illustrate the primitives marched on the sequentially detected VOIs. (c) Auto-degeneration and probabilistic marching. The green circles on the left column indicate different initialization configurations. The ones on the right are the marched results." + ], + "image_footnote": [], + "bbox": [ + 537, + 195, + 656, + 315 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4c7ce3733d47b574f9a15a7fe36926980be4e2f124f678c40bc16f0857f5bff2.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 696, + 95, + 769, + 138 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/37900fffbd83686779db3bb95e8873d09b00a75bbf6898d7576296d0ba4a18ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 95, + 856, + 145 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/464e7a7119bacef703abb393dba94ea8309839f3d4a589acec106c94b0ac73c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 152, + 766, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6aba9cb1f68dd8028a8d0229accefa505a5f5cc1ffd126a08b33141cafad435e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 151, + 857, + 195 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c54279062a4659964a7ceb1352d5743ac4f4975f0335b5c3991eb64971443cb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 203, + 769, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/94bcccd0ffd964e03398687f626dc41983e8413c6aecac1ad454388e73db1503.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 203, + 857, + 247 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd94f74df9d3d67e295fd3f292b5bd802585e0948eb205288ca8ec9e0191ae45.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 697, + 257, + 769, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c0b9610b7201427b289e58d7075253273fa7e9e19d2fa816249ac6e4da411d31.jpg", + "image_caption": [ + "Convergence" + ], + "image_footnote": [], + "bbox": [ + 790, + 255, + 861, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "volumes is large enough to provide sufficient geometric information for the primitive recovery. We illustrate the idea of isosurfaces in Fig.2(a). For each VOI, we initialize a primitive as an ellipsoid, with scales proportional to the size of its smallest bounding-box. The details for the primitive initialization are demonstrated in the Supplementary Material. The procedure of how a primitive marches from the initial guess to capture the local geometry is detailed in the next Sec. 3.4. After obtaining the primitive representations for all the VOIs in $\\bar{S}_m$ , we subtract the voxels", + "bbox": [ + 496, + 465, + 890, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {x} _ {i} \\in \\mathbf {V}, d \\left(\\mathbf {x} _ {i}\\right) \\leq 0 \\wedge d _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right) \\leq 0 \\right\\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 628, + 890, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the set of voxel grid $\\mathbf{V}$ . In other words, we deactivate the voxels which are interior of the target SDF and well fitted by a recovered primitive, while the updated SDF preserves the exterior and unfitted interior volumes. The connectivity marching repeats with the updated target SDF, until the threshold marches higher than a preset limit close to zero, indicating that no prominent interior volume is left unrepresented.", + "bbox": [ + 496, + 656, + 890, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Probabilistic Primitive Marching", + "text_level": 1, + "bbox": [ + 500, + 786, + 790, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, we explain what we mean by primitive marching. In the marching cubes algorithm [25], marching means the progressive extraction of polygonal meshes from the neighboring 8 voxel points in the discrete SDF. In each step, the algorithm focuses within the single cube formulated by the 8 voxel points, calculates triangle vertices using linear in", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8774", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "terpolation, and then 'marches' to the next one. On the contrary, our method marches on the scope of voxels. That is, we gradually figure out a proper collection of voxel points from which a geometric primitive is extractable. Inspired by [24], we formulate this process as a maximum likelihood estimation problem with latent variables.", + "bbox": [ + 75, + 90, + 468, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each isosurface $S_{k}\\in \\bar{S}_{m}$ obtained during the connectivity marching (Eq.(6)), it is assumed that the 'true' underlying primitive is encoded by $\\theta_{k}$ , and that $z_{ik}$ is the correspondence between the primitive $\\theta_{k}$ and a voxel point $\\mathbf{x}_i\\in \\mathbf{V}$ . $z_{ik}$ is a binary variable that equals 1 when the target signed distance evaluated at the $i$ th voxel point results from the $k$ th primitive, and 0 otherwise. Then the target signed distance $d(\\mathbf{x}_i)$ can be regarded as an observation generated from the probabilistic distribution", + "bbox": [ + 75, + 181, + 470, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k}, z _ {i k}\\right) = \\left(\\frac {\\mathbb {1} _ {d _ {i} \\in [ - t , 0)}}{t}\\right) ^ {1 - z _ {i k}} \\mathcal {N} \\left(d _ {i} \\mid d _ {\\boldsymbol {\\theta} _ {k}} (\\mathbf {x} _ {i}), \\sigma^ {2}\\right) ^ {z _ {i k}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 321, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{i}$ is short for $d(\\mathbf{x}_i)$ ; $t$ is the truncation value of the SDF; $\\mathbb{1}_{d_i\\in [-t,0)}$ is an indicator function which equals one when $d_{i}\\in [-t,0)$ and zero otherwise; $\\mathcal{N}\\big(\\cdot |d_{\\pmb{\\theta}_k}(\\mathbf{x}_i),\\sigma^2\\big)$ is a Gaussian distribution with mean $d_{\\pmb{\\theta}_k}(\\mathbf{x}_i)$ and variance $\\sigma^2$ . Consequently, our goal is to find out the optimal $\\pmb{\\theta}_k$ and the voxel-primitive correspondences $z_{ik}$ simultaneously, which maximize the likelihood of the target SDF.", + "bbox": [ + 75, + 364, + 468, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correspondence Marching: We assume that $z_{ik}$ follows a Bernoulli prior distribution $B(p_0)$ , independent of $\\theta_k$ . Then, given the target SDF $d_i$ and the current primitive estimation $\\theta_k$ , we are able to infer the posterior correspondence $z_{ik}$ by the Bayes' rule", + "bbox": [ + 75, + 472, + 470, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(z _ {i k} \\mid \\boldsymbol {\\theta} _ {k}, d _ {i}\\right) = \\frac {p \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k} , z _ {i k}\\right) p \\left(z _ {i k}\\right)}{\\sum_ {z _ {i k} \\in \\{0 , 1 \\}} p \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k} , z _ {i k}\\right) p \\left(z _ {i k}\\right)} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 551, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By analyzing the posterior correspondence, we can observe that our design of the probabilistic model (Eq.(8)) possesses two desirable properties: (1) voxels labeled as inside $(d_{i} < 0)$ may not contribute to the current estimation of the primitive $\\pmb{\\theta}_{k}$ , i.e., $p(z_{ik} = 1|\\pmb{\\theta}_k,d_i < 0)\\in (0,1)\\subset \\mathbb{R}$ ; (2) all non-negative valued voxels always contribute to any primitives, i.e., $p(z_{ik} = 1|\\pmb{\\theta}_k,d_i\\geq 0) = 1$ . In other words, the primitive tries to absorb the interior voxels sharing similar local geometry, while avoids occupying any voxels labeled as the exterior. Meanwhile, the probabilistic correspondence establishes a soft relationship between the primitive and the interior voxels, allowing part of the primitive to occupy some interior volumes without complying with the target signed distance values. In this way, the algorithm is able to achieve an overall better fitting quality.", + "bbox": [ + 75, + 593, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Primitive Update: Given the current estimation of the correspondences between the primitive and voxels, the parameter of the primitive is updated by", + "bbox": [ + 75, + 819, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {k} = \\underset {\\boldsymbol {\\theta} _ {k}} {\\arg \\min } \\sum_ {\\mathbf {x} _ {i} \\in \\mathbf {V} _ {a}} P _ {i k} \\| d _ {\\boldsymbol {\\theta} _ {k}} (\\mathbf {x} _ {\\mathrm {i}}) - d _ {i} \\| _ {2} ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 92, + 545, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P _ {i k} = p \\left(z _ {i k} = 1 \\mid \\boldsymbol {\\theta} _ {k} ^ {p r e v}, d _ {i}\\right) \\tag {11} \\\\ \\mathbf {V} _ {a} = \\left\\{\\mathbf {x} _ {\\mathrm {i}} \\in \\mathbf {V} | d _ {\\boldsymbol {\\theta} _ {k} ^ {p r e v}} (\\mathbf {x} _ {i}) \\in [ - a, a ] \\subset \\mathbb {R} \\right\\} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 116, + 890, + 155 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\theta_{k}^{prev}$ is the previous estimation of the primitive parameters. $\\mathbf{V}_{a}$ is an adaptive subset of the complete voxels $\\mathbf{V}$ , which includes the voxels close to the surface of the previous estimated primitive. $a$ defines the distance threshold of the activated region. Only voxels in $\\mathbf{V}_{a}$ are activated during the optimization. The reason why we use an adaptive subset instead of the complete voxel space is twofold. As discussed in Sec. 3.1, both the source and the target SDF are truncated within a vicinity of zero. Therefore, small variations around $\\theta_{k}^{prev}$ (i.e., small changes on the shape of the primitive surface) have minor if not zero effects on the source SDF values evaluated at voxels distant to the primitive surface. Moreover, the size of $\\mathbf{V}_{a}$ is much smaller than the complete set, and thus providing a significant boost in performance.", + "bbox": [ + 498, + 167, + 890, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The correspondence marching and the primitive updating alternate until convergence. This process is akin to the EM algorithm [13]. The scope of voxels from which a primitive can be extracted is 'marched' with the progressive variation of the primitive shape, and simultaneously the parameters of the primitive are optimized based on the scope of voxels. We apply the optimization method proposed in [24] to avoid local minima. More derivation and implementation details can be found in the Supplementary Material.", + "bbox": [ + 498, + 396, + 892, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Fail-safe Auto-degeneration", + "text_level": 1, + "bbox": [ + 500, + 544, + 750, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The primitives are roughly initialized in the connectivity marching step and further grow to capture local geometric shapes. Since the probabilistic marching is based on optimization, an important question to ask is if the recovered primitives can always converge to a proper shape, regardless of poor initialization. We demonstrate that our probabilistic marching strategy is robust to initialization. Firstly, our method possesses a feature called auto-degeneration i.e. the primitive will autonomously degenerate towards a point when accidentally initialized far from the target volumes. This is because, under this circumstance, only variations which shrink the volume of the primitive can decrease the difference between the truncated target and source SDF. Secondly, when the primitive is initialized near or inside the target volumes, the marching process encourages the primitive to converge to one nearest local target shape by analyzing the posterior correspondences. Fig.2(c) illustrates the examples of the above properties. After the primitive marching, our algorithm removes the degenerated primitives. As a fail-safe measure, primitives which significantly contradict the target SDF is also removed. Detailed implementations can be found in the Supplementary Material.", + "bbox": [ + 498, + 568, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8775", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e66a71d780764a14716b2cca881026a9394a5796e79754d5f6d89cf964d2e74.jpg", + "image_caption": [ + "Figure 3. Shape Abstraction results on the ShapeNet dataset. From left to right: SQs [31], Non-parametric Bayesian (NB) [47], Marching-Primitives with ellipsoids (MPE), Marching-Primitives with superquadrics (MPS), and the ground truth (pre-processed watertight mesh)." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 843, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 361, + 210, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct several experiments to demonstrate the high accuracy and generalizability of our proposed method. First, we show the shape abstraction results on various datasets, ranging from daily objects to human body models. Furthermore, we study the impact of the grid resolution and the truncation threshold on our algorithm. In the end, we apply our algorithm on a large-scale complex indoor scene, which contains a mixed combination of furniture such as sofas, tables, books, lamps, etc. Implementation details can be found in the Supplementary.", + "bbox": [ + 75, + 386, + 473, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation on Datasets", + "text_level": 1, + "bbox": [ + 76, + 545, + 292, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: We compare our method with both the state-of-the-art learning-based method [31] and the computational method [47], which infer superquadric abstractions of the input objects. For convenience, we refer to [31] as SQs, and [47] as NB. We use the official codes and follow the implementation details as stated in these papers, respectively. We do not compare with the cuboid-based algorithms [40, 48, 50], since they focus more on the semantic level abstraction and are limited in expressiveness and accuracy. A detailed performance comparison between the cuboid and superquadric abstractions can be found in [31].", + "bbox": [ + 75, + 568, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We evaluate on two commonly used datasets, the ShapeNet [6] and the DFAUST [4]. In both of the datasets, we are only provided with triangular meshes. Therefore, we need to transform the meshes into discrete SDF representations. It is straightforward to calculate the signed distance value of a point to a watertight mesh. However, the ShapeNet is a human-made synthetic dataset containing many non-watertight meshes formed by nonvolumetric 2D surfaces, self-intersecting or overlapped triangles. Therefore, we pre-process the original ShapeNet models into watertight meshes and then generate the SDF", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "representation with the fast marching algorithm, following the same procedure in [8]. In consideration of fairness and consistency, we use the pre-processed watertight meshes as the common ground truth and the inputs to SQs and NB. For the DFAUST dataset, we generate the SDF directly since the provided meshes are already watertight.", + "bbox": [ + 496, + 362, + 893, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics: Following [31,47] we use the Chamfer-L1 distance and the volumetric intersection over union (IoU) as the quantitative evaluation metrics. The computation of the metrics is detailed in the Supplementary Material.", + "bbox": [ + 496, + 457, + 893, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on ShapeNet: We experiment on 14 categories from the ShapeNet dataset. We split the dataset randomly into the training $(80\\%)$ and testing $(20\\%)$ sets [9], where we train one SQs model per category on the training sets and evaluate all the methods on the testing sets. For our method, we use the SDF representation discretized on a voxel grid of size $100^3$ and range $[-0.5, 0.5]^3$ . Other than superquadrics, we also test our method with ellipsoids as the base primitive, which is a special case of the superquadric representation when the shape parameters $\\epsilon_1, \\epsilon_2$ are fixed to 1. The quantitative results are summarized in Table. 1. We also demonstrate the qualitative comparison among different shape abstraction methods in Fig.3. Our method outperforms all the baselines, even implemented with less expressive ellipsoids. The computational method generally performs better than the learning-based method. This is because the parameters of the primitive are so sparse and geometrically interrelated that they are difficult to get mapped from a high dimensional input by a neural network. Compared with NB, our method has richer details, clearer edges, and more importantly does not occupy the exterior space. Two factors contribute to the advantages. The first one is the extra geometric information embedded in the SDF representation, which eliminates the inherent interior/exterior ambiguity of point clouds. The other one is our special de", + "bbox": [ + 496, + 523, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8776", + "bbox": [ + 482, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b0e36ce4c1af47b632b04f455d509df2f928072ed4981a9e033dcabde4984f96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryChamfer-L1IoU
SQs [31]NB [47]MPE(Ours)MPS(Ours)SQs [31]NB [47]MPE(Ours)MPS(Ours)
airplane0.0370.0230.0210.0190.4410.6710.7310.768
bench0.0560.0280.0200.0200.2380.5790.7300.819
bottle0.0470.0330.0260.0170.6860.6650.8860.924
cabinet0.0590.0360.0370.0280.3940.6660.8400.948
can0.0660.0360.0360.0220.7060.5530.9080.950
chair0.0680.0270.0230.0200.3000.6850.7850.871
lamp0.0720.0290.0220.0210.2340.5890.7500.802
speaker0.0640.0410.0370.0330.3460.6560.8580.920
mailbox0.0950.0260.0260.0240.3330.6940.8020.905
rifle0.0380.0200.0190.0190.4460.7320.7440.811
sofa0.0540.0370.0290.0230.4970.7260.8570.940
table0.0700.0240.0240.0220.2470.7450.8180.932
phone0.0400.0210.0230.0210.6810.8720.8910.947
watercraft0.0480.0320.0220.0220.4650.6180.7930.836
mean0.0570.0280.0240.0220.3680.6740.7930.870
", + "bbox": [ + 93, + 88, + 877, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on Shapenet. MPE and MPS are short for Marching-Primitives with ellipsoids and superquadrics, respectively", + "bbox": [ + 76, + 361, + 890, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sign of the probabilistic generative model, as discussed in Sec.3.4. It is interesting to observe that our algorithm can extract abstractions from objects not intuitively depictable by the primitives.", + "bbox": [ + 75, + 381, + 470, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on DFAUST: We also evaluate on the DFAUST dataset of human body models. We follow the split settings in [29] to train the learning-based method and evaluate all the methods on the testing set. The SDF is discretized on a grid of size $64^3$ and side length 1. Similar to the ShapeNet experiment, we test our method with both superquadrics and ellipsoids. The results are shown in Fig.4. Our method can accurately capture various postures, while the baselines fail to distinguish different body parts in some cases. The abstraction quality of SQs is much better on this dataset compared with the ShapeNet, since human bodies share a common articulated structure that can be captured by the neural network. We observe that the ellipsoid abstraction also achieves satisfying accuracy. This is because the human body mostly consists of rounded shapes compared with man-made objects in the ShapeNet.", + "bbox": [ + 75, + 441, + 468, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Performance Study", + "text_level": 1, + "bbox": [ + 76, + 695, + 261, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section investigates how the Marching-Primitives algorithm behaves with different voxel grid sizes and distance truncation thresholds. We experiment on the Armadillo from the Stanford 3D scanning repository [10]. First, we discretize the model on the voxel grids of different resolutions. The qualitative abstraction results are visualized in Fig. 5, and the quantitative results are summarized in Fig. 6(a). When the input grid resolution is low, the recovered primitive representation is relatively abstract. With the increase of the grid resolution, our method is able to extract detailed abstraction more faithful to the target shape. This is because the higher the resolution, the more geomet", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb148d17d21e964935337f1bef81e44218b01641e92e3636b80e35ab4861fd9b.jpg", + "image_caption": [ + "Figure 4. Shape abstraction results on the DFAUST dataset. From left to right: SQs [31], NB [47], MPE, MPS, and the ground truth." + ], + "image_footnote": [], + "bbox": [ + 521, + 381, + 869, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ric information our method can utilize to guide the primitive marching process. As discussed in Sec.3.1, we use a truncated version of the target and source SDF. Therefore, we also evaluate the performance on different truncation thresholds, starting from 0.1 to 4 times the interval of the voxel grids. The results are shown in Fig.6(b). The abstraction accuracy increases as the threshold decreases at first. The reason lies in that the approximated superquadric", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8777", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31fc9eb13733909ba8a33c97ce518a25ffed8eaac37d4fdf2806906d06401ffc.jpg", + "image_caption": [ + "32" + ], + "image_footnote": [], + "bbox": [ + 143, + 90, + 250, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68049e4909d9f3938e3bb1ce1a6fa9b0d9b3ac62dedd802df6e91d7e83ed252f.jpg", + "image_caption": [ + "$64^{3}$" + ], + "image_footnote": [], + "bbox": [ + 251, + 93, + 359, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b0e817b2c625223415bb6965c9e432bb019a98cfbc98c7ab65b870b327bc8c1.jpg", + "image_caption": [ + "96" + ], + "image_footnote": [], + "bbox": [ + 362, + 94, + 468, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86bf24b15c2b460d791b9bbe4d34fb764380ca886847fa3806f0a7d9d800a6cb.jpg", + "image_caption": [ + "128" + ], + "image_footnote": [], + "bbox": [ + 470, + 93, + 578, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7fb6571ce79fabbf090576e7f03d6e440163f61c7d858053c33290ab6cfab060.jpg", + "image_caption": [ + "$256^{3}$" + ], + "image_footnote": [], + "bbox": [ + 581, + 94, + 687, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c450411222d5b7b515756deb1026a68ecda48dd7d80042bf51b9dbefdd08a71.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 692, + 95, + 800, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": ".source) SDF converges to the true value when truncated close to the surface. If we further decrease the threshold, however, we observe an acute decrease in accuracy. This is because an overly small truncation threshold corrupts the original geometric information embedded in the target SDF, as well as the smoothness of the cost function (Eq. (10)).", + "bbox": [ + 75, + 239, + 468, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03bb3d3f4bde15145e1b737c91a4aa7dedf067df550f4b9924f93af47a459939.jpg", + "image_caption": [ + "Figure 6. Quantitative results on shape abstraction accuracy. (a) Abstraction accuracy versus input grid resolution. (b) Abstraction accuracy versus truncation ratio, evaluated at grid resolution $128^{3}$ . We evaluate on two metrics: Chamfer- $L_{1}$ and IoU." + ], + "image_footnote": [], + "bbox": [ + 91, + 351, + 271, + 452 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2fcfebed2174d44466761b443816a0d8119cfd6218fafcdcce31ccd74692a713.jpg", + "image_caption": [ + "Figure 7. Qualitative result of the scene abstraction. (a) Superquadric abstraction by the Marching-Primitives. (b) Mesh generated by the marching cubes [25] from the same SDF." + ], + "image_footnote": [], + "bbox": [ + 272, + 352, + 462, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Scene Abstraction", + "text_level": 1, + "bbox": [ + 76, + 542, + 253, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other than single objects, we also test the capability of our method in representing complex scenes with geometric primitives. We experiment on a real-world indoor scene called Reading Room [49]. The scene is captured by an Asus Xtion Pro Live camera. The RGB-D scans are fused utilizing [19] (an SDF-based 3D reconstruction algorithm), and further fine-tuned with the method proposed in [49]. However, the SDF representation of the scene is not available publicly. Therefore, we pre-process the scene mesh by removing the floor plane, filling the holes, and transforming it into the SDF representation with grid size $400^3$ . The scene abstraction task is much more difficult compared with the object-level task because various items with great differences in size and shape are present in the very same space. We compare our abstraction result with the mesh extracted by the marching cubes algorithm on the same discrete SDF representation, as shown in Fig.7. Our representation is not only visually satisfying but also contains abundant geometric information the mesh lacks, since it is a continuous SDF approximation to the input discrete SDF. Moreover, our highly sparse representation is only 8.2KB in size, while the discrete SDF occupies 203MB.", + "bbox": [ + 75, + 568, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90d4e1078ea8c708964d17b6f18d58caceeff00d5e21035dd75b9a9734a67e93.jpg", + "image_caption": [ + "Figure 5. Visualization of the shape abstractions on different grid resolutions. Abstraction accuracy increases with the input grid resolution." + ], + "image_footnote": [], + "bbox": [ + 566, + 241, + 821, + 342 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e338b050542461c4d7f525f208a3cac6778e0dc8b4b88925884ff1dd0515d7a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 344, + 821, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 512, + 617, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the Marching-Primitives, the first algorithm to extract primitive-based abstractions from the volumetric SDF representation. Our method outperforms the state-of-the-art in terms of accuracy and generalizability on both synthetic and real-world datasets. Our primitive-based representation is sparse, accurate, generalizable, and can be expressed analytically without training, which we believe will facilitate and inspire further explorations in scene understanding, 3D reconstruction, and robot motion planning. However, our algorithm has the limitation of not being able to properly extract object parts thinner than the interval of the voxel grid. This problem can be solved by either increasing the grid resolution or detecting and thickening the invalid parts. Future work includes the parallelization of the marching process and inferring semantic-level interpretations from the current primitive-based geometric features.", + "bbox": [ + 496, + 537, + 890, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research is supported by NUS Startup grants A-0009059-02-00 and A-0009059-03-00, National Research Foundation, Singapore, under its Medium Sized Centre Programme - Centre for Advanced Robotics Technology Innovation (CARTIN), subaward A-0009428-08-00, and AME Programmatic Fund Project MARIO A-0008449-01-00, and JHU internal funds.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8778", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] R. Bajcsy and F. Solina. Three dimensional object representation revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 231-240, 1987. 1", + "[2] A. H. Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1(1):11-23, 1981. 2, 3", + "[3] I. Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2):115, 1987. 1", + "[4] F. Bogo, J. Romero, G. Pons-Moll, and M. J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 6", + "[5] D. E. Breen, S. Mauch, and R. T. Whitaker. 3d scan conversion of csg models into distance volumes. In Proceedings of the 1998 IEEE Symposium on Volume Visualization, pages 7-14, 1998. 3", + "[6] A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6", + "[7] L. Chevalier, J. Jaillet, and A. Baskurt. Segmentation and superquadric modeling of 3D objects. In The 11-th International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision'2003, WSCG 2003, 2003. 1, 2", + "[8] G. Chou, I. Chugunov, and F. Heide. Gensdf: Two-stage learning of generalizable signed distance functions. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 6", + "[9] C. B. Choy, D. Xu, J. Gwak, K. Chen, and S. Savarese. 3dr2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-644. Springer International Publishing, 2016. 6", + "[10] B. Curless and M. Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2, 7", + "[11] A. Dai, D. Ritchie, M. Bokeloh, S. Reed, J. Sturm, and M. Nießner. Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2018. 2", + "[12] A. Dai, C. Ruizhongtai Qi, and M. Niessner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 2", + "[13] A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 5", + "[14] B. Deng, K. Genova, S. Yazdani, S. Bouaziz, G. Hinton, and A. Tagliasacchi. Cvxnet: Learnable convex decomposition." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 31-41, 2020. 1", + "[15] W. Dong, Q. Wang, X. Wang, and H. Zha. Psdf fusion: Probabilistic signed distance function for on-the-fly 3d data fusion and scene reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 701-717, 2018. 2", + "[16] K. Genova, F. Cole, D. Vlasic, A. Sarna, W. Freeman, and T. Funkhouser. Learning shape templates with structured implicit functions. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7153-7163, 2019. 1", + "[17] A. D. Gross and T. E. Boult. Error of fit measures for recovering parametric solids. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 690-694, 1988. 1, 2, 3", + "[18] Z. Hao, H. Averbuch-Elor, N. Snavely, and S. Belongie. Dualsf: Semantic shape manipulation using a two-level representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7631–7641, 2020. 2, 3", + "[19] S. Izadi, D. Kim, O. Hilliges, D. Molyneaux, R. Newcombe, P. Kohli, J. Shotton, S. Hodges, D. Freeman, A. Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 2, 8", + "[20] D. K. Katherine and S. S. Elizabeth. Core systems in human cognition. In From Action to Cognition, volume 164 of Progress in Brain Research, pages 257-264. Elsevier, 2007. 1", + "[21] Y. Kawana, Y. Mukuta, and T. Harada. Neural star domain as primitive representation. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7875-7886. Curran Associates, Inc., 2020. 1", + "[22] A. Leonardis, A. Jaklic, and F. Solina. Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(11):1289-1295, 1997. 1, 2", + "[23] Y. Liao, S. Donne, and A. Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2916-2925, 2018. 2", + "[24] W. Liu, Y. Wu, S. Ruan, and G. S. Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2676-2685, June 2022. 1, 2, 3, 5", + "[25] W. E. Lorensen and H. E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 2, 3, 4, 8", + "[26] R. A. Newcombe, D. Fox, and S. M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2", + "[27] C. Niu, J. Li, and K. Xu. Im2Struct: Recovering 3D shape structure from a single RGB image. In Proceedings of the" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8779", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 1", + "[28] J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2", + "[29] D. Paschalidou, L. V. Gool, and A. Geiger. Learning unsupervised hierarchical part decomposition of 3D objects from a single RGB image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 3, 7", + "[30] D. Paschalidou, A. Katharopoulos, A. Geiger, and S. Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3215, June 2021. 1", + "[31] D. Paschalidou, A. O. Ulusoy, and A. Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 3, 6, 7", + "[32] A. Pentland. Perceptual organization and the representation of natural form. Artificial Intelligence, 28(3):293-331, 1986. 1, 2", + "[33] A. H. Quispe, B. Milville, M. A. Gutiérrez, C. Erdogan, M. Stilman, H. Christensen, and H. B. Amor. Exploiting symmetries and extrusions for grasping household objects. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3702-3708, 2015. 1", + "[34] Y. Rao, Y. Nie, and A. Dai. Patchcomplete: Learning multi-resolution patch priors for 3d shape completion on unseen categories. Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[35] S. Ruan, K. L. Poblete, H. Wu, Q. Ma, and G. S. Chirikjian. Efficient path planning in narrow passages for robots with ellipsoidal components. IEEE Transactions on Robotics (TRO), pages 1-18, 2022. 1", + "[36] S. Ruan, X. Wang, and G. S. Chirikjian. Collision detection for unions of convex bodies with smooth boundaries using closed-form contact space parameterization. IEEE Robotics and Automation Letters (RAL), 7(4):9485-9492, 2022. 1", + "[37] G. Sharma, B. Dash, A. RoyChowdhury, M. Gadelha, M. Loizou, L. Cao, R. Wang, E. Learned-Miller, S. Maji, and E. Kalogerakis. Prift: Learning to fit primitives improves few shot point cloud segmentation. In Computer Graphics Forum, volume 41, pages 39-50. Wiley Online Library, 2022. 3", + "[38] D. Smirnov, M. Fisher, V. G. Kim, R. Zhang, and J. Solomon. Deep parametric shape predictions using distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 561-570, 2020. 1", + "[39] F. Solina and R. Bajcsy. Recovery of parametric models from range images: the case for superquadrics with global deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(2):131-147, 1990. 1, 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] C. Sun, Q. Zou, X. Tong, and Y. Liu. Learning adaptive hierarchical cuboid abstractions of 3d shape collections. ACM Transactions on Graphics (TOG), 38(6):1-13, 2019. 3, 6", + "[41] S. Tulsiani, H. Su, L. J. Guibas, A. A. Efros, and J. Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 3", + "[42] N. Vaskevicius and A. Birk. Revisiting superquadric fitting: A numerically stable formulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):220-233, 2019. 3", + "[43] S. Vasu, N. Talabot, A. Lukoianov, P. Baque, J. Donier, and P. Fua. Hybridsdf: Combining deep implicit shapes and geometric primitives for 3d shape representation and manipulation. In International Conference on 3D Vision (3DV), 2022. 2", + "[44] G. Vezzani, U. Pattacini, and L. Natale. A grasping approach based on superquadric models. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1579-1586, 2017. 1", + "[45] G. Vezzani, U. Pattacini, G. Pasquale, and L. Natale. Improving superquadric modeling and grasping with prior on object shapes. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 6875-6882, 2018. 1", + "[46] S. Weder, J. L. Schonberger, M. Pollefeys, and M. R. Oswald. Routedfusion: Learning real-time depth map fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2", + "[47] Y. Wu, W. Liu, S. Ruan, and G. S. Chirikjian. Primitive-based shape abstraction via nonparametric bayesian inference. In Proceedings of the European Conference on Computer Vision (ECCV), pages 479-495. Springer Nature Switzerland, 2022. 1, 2, 3, 6, 7", + "[48] K. Yang and X. Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 1, 2, 3, 6", + "[49] Q. Zhou, S. Miller, and V. Koltun. Elastic fragments for dense scene reconstruction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 473-480, 2013. 2, 8", + "[50] C. Zou, E. Yumer, J. Yang, D. Ceylan, and D. Hoiem. 3D-PRNN: Generating shape primitives with recurrent neural networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1, 6" + ], + "bbox": [ + 501, + 92, + 890, + 727 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8780", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_model.json b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3a3cd62949007258934a0ed8b09cfe78b51e230e --- /dev/null +++ b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.851, + 0.154 + ], + "angle": 0, + "content": "Marching-Primitives: Shape Abstraction from Signed Distance Function" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.18, + 0.768, + 0.234 + ], + "angle": 0, + "content": "Weixiao Liu\\(^{1,2}\\) Yuwei Wu\\(^{1}\\) Sipu Ruan\\(^{1}\\) Gregory S. Chirikjian\\(^{1*}\\) \n\\(^{1}\\)National University of Singapore \\(^{2}\\)Johns Hopkins University \n{mpewxl, yw.wu, ruansp, mpegre}@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Representing complex objects with basic geometric primitives has long been a topic in computer vision. Primitive-based representations have the merits of compactness and computational efficiency in higher-level tasks such as physics simulation, collision checking, and robotic manipulation. Unlike previous works which extract polygonal meshes from a signed distance function (SDF), in this paper, we present a novel method, named Marching-Primitives, to obtain a primitive-based abstraction directly from an SDF. Our method grows geometric primitives (such as superquadrics) iteratively by analyzing the connectivity of voxels while marching at different levels of signed distance. For each valid connected volume of interest, we march on the scope of voxels from which a primitive is able to be extracted in a probabilistic sense and simultaneously solve for the parameters of the primitive to capture the underlying local geometry. We evaluate the performance of our method on both synthetic and real-world datasets. The results show that the proposed method outperforms the state-of-the-art in terms of accuracy, and is directly generalizable among different categories and scales. The code is open-sourced at https://github.com/ChirikjianLab/Marching-Primitives.git." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Recent years have witnessed great progress in the areas of 3D shape representation and environmental perception. Low-level representations such as surface meshes, point clouds, and occupancy grids are widely used as inputs to high-level computer vision algorithms and artificial intelligence tasks. They have the advantage of being able to represent and visualize objects with high accuracy and rich local geometric features. However, the low-level representations are ineffective in delivering a general and intuitive sense of structural geometry as well as part-level scene understanding. Studies [3,20] show that human vision, unlike" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.271, + 0.878, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.369, + 0.895, + 0.468 + ], + "angle": 0, + "content": "Figure 1. Primitive-base representation versus mesh. For each pair of objects, the left one is the superquadratic abstraction obtained by our algorithm, and the right one is the original mesh. The mesh of the chair is 6MB in size, while our representation only needs 4KB. An SDF representation discretized on a \\(128^{3}\\) voxel grid occupies 19MB. Our abstraction is equivalent to an implicit continuous SDF, which is an approximation to the discrete SDF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.644 + ], + "angle": 0, + "content": "computer vision, tends to perceive and understand scenes as combinations of simple primitive shapes. Human beings perform well and robustly in complex tasks, providing a basic geometric description of the scene is available [32]. Therefore, researchers turn to exploring the possibility of interpreting complex objects and scenes with basic geometric primitives. Taking advantage of the primitive-based representation, many higher-level tasks, such as segmentation [14, 16, 21, 30], scene understanding [29, 31, 41, 47], grasping [33, 44, 45] and motion planning [35, 36], are able to be solved efficiently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.902 + ], + "angle": 0, + "content": "However, it still remains challenging to extract primitive-based abstractions from low-level representations. Starting from the 1990s, Solina et al. [1, 17, 39] aim to extract a single superquadric representation from a simple object by minimizing the least-square error between the primitive and the measured points. Later in [7, 22], their method is extended to represent more complex objects with multiple primitives. More recently, the authors of [24, 47] reformulate the task as a probabilistic inference problem with enhanced accuracy and robustness to noise and outliers. At the same time, with the surge of data-driven techniques, researchers attempt to train neural networks to infer cuboids [27,38,41,48,50] and superquadrics [29,31] representations in an end-to-end fashion. However, both the computational and learning-based approaches have their own limitations. The computational methods are vulnerable to the inherent ambiguity of the point-to-surface relationship. For exam" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8771" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "ple, the algorithms tend to fill empty spaces of a non-convex object with primitives by mistake, due to the inside/outside ambiguity of a surface depicted by a set of points [24, 48]. The main drawback of the learning approaches lies in the lack of generalizability beyond the object category on which the model is trained [24, 31, 47, 48]. Also, the shape abstraction accuracy is inferior to the computational methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.471, + 0.565 + ], + "angle": 0, + "content": "The signed distance function (SDF) has been a successful 3D volumetric representation in varieties of computer vision and graphics tasks. It is the basic framework for many classic 3D reconstruction algorithms such as TSDF volume reconstruction [10, 15], KinectFusion [19], and Dynamic-Fusion [26]. Recently, the SDF representation is adapted to the deep learning frameworks, and exhibits boosted potentials in shape encoding [8, 18, 28, 43], surface reconstruction [23, 46], and shape completion [11, 12, 34]. Usually, triangular mesh surfaces are extracted from the SDF representation with the marching cubes algorithm [25]. Point cloud and occupancy grid representations are also obtained by keeping the vertices of the meshes and the sign of each voxel point, respectively. The SDF is among the most informative 3D representations since it encodes not only the surface geometry but also the distance and side of a point relative to the shape. Meanwhile, it is easily achievable via range images from 3D sensors [10], or learnable from other input modalities [8, 18, 28]. Since we are able to extract meshes from an SDF, it is natural to think about the possibility of extracting primitives as well. Furthermore, the primitive-based abstraction is a continuous interpretation of the complete geometric information encoded in the original discrete SDF, but requires much less storage size (Fig.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Motivated by the aforementioned facts and the bottleneck of the current shape abstraction algorithms, we proposed a general shape abstraction method by reasoning directly on the informative SDF representation. The goal of our method is to find a combination of geometric primitives whose underlying SDF values match the target values evaluated on the evenly spaced discrete grid points (Sec. 3.1 and Sec. 3.2). To solve this problem, we propose a two-step iterative algorithm called the Marching-Primitives. Our algorithm 'marches' on two domains: the signed distance domain and the voxelized space domain, alternately. Firstly, the connectivity of volumes are analyzed by generating isosurfaces on a sequence of decreasing levels of negative signed distances (Sec.3.3). By doing so, volumes of interest (VOIs) where primitives are likely to be encoded can be identified sequentially. In the second step, for each of the VOIs, our algorithm marches on the neighbouring voxels to infer their probabilistic correspondences to the primitive and simultaneously optimizes the shape and pose of the primitive (Sec.3.4). After the primitive representation of a VOI is achieved, the fitted volumes are deactivated from the voxel grid. Our algorithm continues marching on the signed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "distance domain until it approaches zero, i.e., all the interior volumes of the SDF have been captured by the recovered primitives. We compare our algorithm with the state-of-the-art of both the computational and learning-based approaches on the ShapeNet object dataset [6] and D-FAUST human shape dataset [4] (Sec. 4.1). We also study the performance of our algorithm on different conditions (Sec. 4.2). Finally, we demonstrate the scene abstraction result of the Stanford Reading Room [49], which contains several pieces of furniture of various categories (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.257, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.892, + 0.749 + ], + "angle": 0, + "content": "SDF Representation: The SDF can be stored in two different ways: discrete or continuous. A majority of computer vision and graphics algorithms are built on the SDF discretized on a 3D grid of voxel points. The signed distances are stored on each of the corresponding voxel points. The authors of [10, 19, 26] pioneer in fusing several noisy range images into a single discrete SDF. Their work is widely applied in 3D reconstruction and plays an important role in robotics tasks such as simultaneous localization and mapping. The discrete SDF is also a promising input/output representation for 3D deep learning [11, 12, 34]. Recently, it becomes popular to encode shapes as a continuous SDF with neural networks [28]. Vasu et al. [43] further improve the shape encoding quality by enforcing local regularities with geometric primitives. In [8], the authors adopt a two-stage meta-learning approach to further extend the generalization capabilities of neural SDF. With the deep neural network, it becomes possible to infer SDF representations from partially observed 3D inputs or even images. Both the discrete and the continuous SDF are implicit representations of geometric surfaces. To extract the explicit surface from the SDF, continuous SDFs need to be discretized on a voxel grid first and then conduct the marching cubes [25]. This method allows high-quality rendering of the objects, however, surface meshes are non-sparse and contain no structural level information. Our method provides an alternative approach to describe the underlying object in the SDF. Instead of meshes, we directly extract a collection of sparsely parameterized primitives from the SDF. Other than that, our primitive-based representation itself is also a concise yet continuous SDF approximation to the original discrete SDF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Computational Shape Abstraction: The most well-studied primitive for computational shape abstraction is the superquadric, due to its extensive shape vocabulary including cuboids, ellipsoids, cylinders, octohedra, and many shapes in between (e.g., cuboids with rounded edges). It is first proposed as a versatile modeling element for complex objects in computer graphics [2, 32]. Later, Solina et al. propose a method to conduct abstraction of simple objects from range images with a single superquadric [17, 39]. Leonardis et al. [22] and Chevalier et al. [7] further extend" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8772" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.363 + ], + "angle": 0, + "content": "the previous work to recover complex objects with multiple superquadrics with a Split-and-Merge strategy. A numerical instability problem is addressed and revisited in [42]. The authors introduce an auxiliary function in the unstable region and receive a better abstraction accuracy. More recently, Liu et al. [24] formulate the problem in a probabilistic fashion and propose a geometric strategy to avoid local optimum, bringing a significant improvement in robustness to outlier and fitting accuracy. Wu et al. [47] extend and recast the work as a nonparametric Bayesian inference problem so as to improve the applicability on complex shapes. To the best of the authors' knowledge, the existing computational methods are all based on range images or point clouds, which suffer from geometric ambiguities [48]. In contrast, our method takes advantage of the abundant geometric information encoded in the SDF and is easily compatible with other computer vision algorithms based on the SDF representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.789 + ], + "angle": 0, + "content": "Learning-based Shape Abstraction: The learning-based method is first seen in [41]. Tulsiani et al. propose a 3D convolutional neural network (CNN) to learn shape abstractions with cuboids from the occupancy grid. Sun et al. [40] design an adaptive hierarchical cuboid representation and introduce an unsupervised approach to learn to extract the parameters for the representation. Yang et al. [48] train a variational auto-encoder network to transform point clouds into parametric cuboids. Other than cuboids, researchers also seek to extract spheres and ellipsoids representations from objects. Hao et al. [18] combine the neural SDF with the spherical representation by sharing a same latent layer. In [37], ellipsoids or cuboids are extracted to help segment the input point cloud. However, a single type of primitive has very limited expressiveness. Therefore, Paschalidou et al. [29, 31] turn to training neural networks to conduct abstractions with the superquadrics as the atomic elements. Learning-based approaches are versatile in dealing with different input sources. They are able to make shape abstractions from point clouds, voxel grids, or even RGB images which are so ill-conditioned that the computational approaches have little chance of working. However, learning-based approaches rely heavily on the training dataset and thus are less generalizable to unseen categories. Instead, our approach reasons about the primitive abstraction from a case-by-case geometric perspective, which provides an inherent advantage in generalizability and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.805, + 0.17, + 0.82 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.207, + 0.847 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The discrete SDF is a volumetric surface representation built on a voxel grid \\(\\mathbf{V} = \\{\\mathbf{x}_i\\in \\mathbb{R}^3,i = 1,2,\\dots,N\\}\\). A scalar \\(d(\\mathbf{x}_i)\\) is assigned to each grid point, which indicates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "the signed distance of \\(\\mathbf{x}_i\\) to the nearest surface. The point \\(\\mathbf{x}_i\\) lies inside the surface if \\(d(\\mathbf{x}_i) < 0\\) and outside otherwise. Typically, the surface mesh of an object is extracted from the SDF by the marching cubes algorithm [25]. In this paper, we call the input SDF \\(d(\\mathbf{x}_i)\\) the target SDF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.213 + ], + "angle": 0, + "content": "For the primitive representation, we select the superquadrics [2], a family of geometric primitives defined by the implicit equation" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.219, + 0.892, + 0.265 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {x}) = \\left(\\left(\\frac {x}{a _ {x}}\\right) ^ {\\frac {2}{\\varepsilon_ {2}}} + \\left(\\frac {y}{a _ {y}}\\right) ^ {\\frac {2}{\\varepsilon_ {2}}}\\right) ^ {\\frac {\\varepsilon_ {2}}{\\varepsilon_ {1}}} + \\left(\\frac {z}{a _ {z}}\\right) ^ {\\frac {2}{\\varepsilon_ {1}}} = 1 \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.893, + 0.499 + ], + "angle": 0, + "content": "The superquadric family is only encoded by 5 parameters (shape parameters \\(\\epsilon_{1},\\epsilon_{2}\\in [0,2]\\subset \\mathbb{R}\\), and scale parameters \\(a_{x},a_{y},a_{z}\\in \\mathbb{R}_{>0}\\)), but has an extensive shape vocabulary. Note that the shape parameters can exceed 2, resulting in nonconvex shapes. However, practically in most studies [24,31] and also in our paper, we limit them within the convex region as defined above. The points \\(\\mathbf{x} = [x,y,z]\\in \\mathbb{R}^3\\) satisfying Eq. (1) form the surface of the superquadric. In this paper, we also include the Euclidean transformation \\(g\\in SE(3)\\), i.e. 3 Euler angles for rotation \\(\\mathbf{R}\\in SO(3)\\) and 3-dimensional translation \\(\\mathbf{t}\\in \\mathbb{R}^3\\), to parameterize a superquadric with a general pose. In total, we denote a superquadric with a vector \\(\\pmb{\\theta}\\) of 11 elements. According to Eq. (1), we can approximate the signed distance of a grid point \\(\\mathbf{x}_i\\) to a general posed superquadric \\(\\pmb{\\theta}\\) explicitly by" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.505, + 0.892, + 0.531 + ], + "angle": 0, + "content": "\\[\nd _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right) = \\left(1 - f ^ {- \\frac {\\epsilon_ {1}}{2}} \\left(g ^ {- 1} \\circ \\mathbf {x} _ {i}\\right)\\right) \\| g ^ {- 1} \\circ \\mathbf {x} _ {i} \\| _ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.657 + ], + "angle": 0, + "content": "We are not able to use the exact SDF of superquadrics, because it has no analytical solution and is only achievable by numerical optimization [5]. Eq.(2) is the radial distance [17, 24] of \\(\\mathbf{x}_i\\) to the superquadric surface, which converges to the exact signed distance as \\(d_{\\theta}(\\mathbf{x}_i)\\to 0\\). Therefore, we truncate both the input target SDF \\(d(\\cdot)\\) and the primitive SDF \\(d_{\\theta}(\\cdot)\\) within the vicinity of zero to ensure the approximation accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.704, + 0.68 + ], + "angle": 0, + "content": "3.2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.764 + ], + "angle": 0, + "content": "To obtain a primitive-based abstraction from the target SDF, we seek a combination of primitives \\(\\Theta \\doteq \\{\\theta_{k}, k = 1,2,\\dots,K\\}\\) whose underlying signed distances measured on the voxel points (we call it the source SDF in contrast to the target SDF) match the target SDF, that is" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.77, + 0.892, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Theta} = \\underset {\\boldsymbol {\\Theta}} {\\arg \\min } \\sum_ {\\mathbf {x} _ {i} \\in \\mathbf {V}} \\underset {\\boldsymbol {\\theta} _ {k} \\in \\boldsymbol {\\Theta}} {\\min } \\left\\| d _ {\\boldsymbol {\\theta} _ {k}} \\left(\\mathbf {x} _ {i}\\right) - d \\left(\\mathbf {x} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our problem formulation is simple and intuitive, however, it is intractable to solve directly. First, the number of primitives \\( K \\) needed is unknown. Also, the correspondences between the voxel points \\( \\mathbf{x}_i \\) and the primitives \\( \\theta_k \\) are unknown a priori. Therefore, it makes Eq.(3) a chicken-and-egg problem: on the one hand, if we know the set of voxel" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8773" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "points contributing to a same primitive, we are able to solve for the parameters of the primitive by optimization; on the other hand, if we know the configurations of the primitives, we are able to tell the belongings of each voxel. Other than those factors, the computation complexity itself is prohibitively high, because we need to evaluate hundreds of thousands or even millions of voxel points. This makes it infeasible to operate directly on the complete dense voxel grid. To tackle these difficulties, we propose an iterative algorithm, the Marching-Primitives, which simultaneously solves the correspondences and primitive parameters efficiently. It is worth noting that our method is generalizable beyond the superquadrics. Any volumetric primitives with easily accessible SDF representations can be adapted to our framework." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.328, + 0.362, + 0.344 + ], + "angle": 0, + "content": "3.3. Iterative Connectivity Marching" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Instead of setting a predefined number of primitives, our method starts from an empty set of primitives and grows as needed. This is realized by analyzing the connectivity (in this paper, we use the 26-connectivity) of the voxels at different levels of signed distance. Given the target SDF, our algorithm checks the connectivity of voxels whose signed distance is less than a sequence of thresholds" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.468, + 0.47, + 0.492 + ], + "angle": 0, + "content": "\\[\nT ^ {c} \\doteq \\left\\{t _ {1} ^ {c}, t _ {2} ^ {c}, \\dots \\right\\}, \\quad t _ {1} ^ {c} = \\min _ {\\mathbf {x} _ {i} \\in \\mathbf {V}} d (\\mathbf {x} _ {i}), \\quad t _ {m + 1} ^ {c} = \\alpha t _ {m} ^ {c} (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.578 + ], + "angle": 0, + "content": "where \\(\\alpha \\in (0,1) \\subset \\mathbb{R}\\) is the common ratio, resulting in a geometric sequence exponentially decaying to zero. Each threshold \\(t_m^c\\) defines a set of disjoint isosurfaces \\(S_m\\), encompassing the connected voxels whose signed distances are less than the threshold." + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.591, + 0.469, + 0.608 + ], + "angle": 0, + "content": "\\[\nS _ {m} = \\left\\{S _ {k}, k = 1, 2, \\dots , | S _ {m} | \\right\\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.619, + 0.47, + 0.65 + ], + "angle": 0, + "content": "\\(|S_{m}|\\) denotes the number of the disjoint isosurfaces. We construct a subset \\(\\bar{S}_m\\) from \\(S_{m}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.661, + 0.469, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\bar {S} _ {m} = \\left\\{S _ {k} \\in S _ {m}, | S _ {k} | \\geq N _ {c} \\right\\} \\subseteq S _ {m} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.901 + ], + "angle": 0, + "content": "where \\( |S_{k}| \\) is the number of the connected voxels within the isosurface \\( S_{k} \\). \\( \\bar{S}_{m} \\) is the set of the isosurfaces encompassing no less than \\( N_{c} \\) connected voxels. The connectivity marching starts from the innermost threshold \\( t_1^c \\), where the resulting \\( \\bar{S}_1 \\) might be empty. The marching continues on the sequence (increasing the threshold) until \\( \\bar{S}_{m} \\) is nonempty. Each of the connected volumes in \\( \\bar{S}_{m} \\) (we call it a VOI) is an ideal starting point for growing a primitive, because: (1) those volumes are among the most interior of the target SDF, corresponding to most prominent part of the geometry; (2) disconnected and weakly connected volumes can be separated apart; (3) the SDF of a primitive can also be interpreted as layers of isosurfaces, and thus sharing similar geometric structure; and furthermore (4) the size of the" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.092, + 0.668, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.183, + 0.548, + 0.194 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.196, + 0.657, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.092, + 0.69, + 0.105 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.097, + 0.77, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.096, + 0.857, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.153, + 0.767, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.152, + 0.858, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.204, + 0.77, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.204, + 0.859, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.258, + 0.77, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.307, + 0.76, + 0.316 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.256, + 0.862, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.307, + 0.858, + 0.317 + ], + "angle": 0, + "content": "Convergence" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Figure 2. Visualizations of concepts. (a) A 2D illustration of two marched isosurfaces separating the shape into two VOIs. (b) A simple example of the Marching-Primitives algorithm. The first figure is the input target SDF (voxels with signed distances larger than the truncation threshold are left blank for visualization). Figures b2-b4 illustrate the primitives marched on the sequentially detected VOIs. (c) Auto-degeneration and probabilistic marching. The green circles on the left column indicate different initialization configurations. The ones on the right are the marched results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.618 + ], + "angle": 0, + "content": "volumes is large enough to provide sufficient geometric information for the primitive recovery. We illustrate the idea of isosurfaces in Fig.2(a). For each VOI, we initialize a primitive as an ellipsoid, with scales proportional to the size of its smallest bounding-box. The details for the primitive initialization are demonstrated in the Supplementary Material. The procedure of how a primitive marches from the initial guess to capture the local geometry is detailed in the next Sec. 3.4. After obtaining the primitive representations for all the VOIs in \\(\\bar{S}_m\\), we subtract the voxels" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.629, + 0.892, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {x} _ {i} \\in \\mathbf {V}, d \\left(\\mathbf {x} _ {i}\\right) \\leq 0 \\wedge d _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right) \\leq 0 \\right\\} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.779 + ], + "angle": 0, + "content": "from the set of voxel grid \\(\\mathbf{V}\\). In other words, we deactivate the voxels which are interior of the target SDF and well fitted by a recovered primitive, while the updated SDF preserves the exterior and unfitted interior volumes. The connectivity marching repeats with the updated target SDF, until the threshold marches higher than a preset limit close to zero, indicating that no prominent interior volume is left unrepresented." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.791, + 0.804 + ], + "angle": 0, + "content": "3.4. Probabilistic Primitive Marching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Firstly, we explain what we mean by primitive marching. In the marching cubes algorithm [25], marching means the progressive extraction of polygonal meshes from the neighboring 8 voxel points in the discrete SDF. In each step, the algorithm focuses within the single cube formulated by the 8 voxel points, calculates triangle vertices using linear in" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8774" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.181 + ], + "angle": 0, + "content": "terpolation, and then 'marches' to the next one. On the contrary, our method marches on the scope of voxels. That is, we gradually figure out a proper collection of voxel points from which a geometric primitive is extractable. Inspired by [24], we formulate this process as a maximum likelihood estimation problem with latent variables." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.471, + 0.318 + ], + "angle": 0, + "content": "For each isosurface \\(S_{k}\\in \\bar{S}_{m}\\) obtained during the connectivity marching (Eq.(6)), it is assumed that the 'true' underlying primitive is encoded by \\(\\theta_{k}\\), and that \\(z_{ik}\\) is the correspondence between the primitive \\(\\theta_{k}\\) and a voxel point \\(\\mathbf{x}_i\\in \\mathbf{V}\\). \\(z_{ik}\\) is a binary variable that equals 1 when the target signed distance evaluated at the \\(i\\)th voxel point results from the \\(k\\)th primitive, and 0 otherwise. Then the target signed distance \\(d(\\mathbf{x}_i)\\) can be regarded as an observation generated from the probabilistic distribution" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.323, + 0.469, + 0.365 + ], + "angle": 0, + "content": "\\[\np \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k}, z _ {i k}\\right) = \\left(\\frac {\\mathbb {1} _ {d _ {i} \\in [ - t , 0)}}{t}\\right) ^ {1 - z _ {i k}} \\mathcal {N} \\left(d _ {i} \\mid d _ {\\boldsymbol {\\theta} _ {k}} (\\mathbf {x} _ {i}), \\sigma^ {2}\\right) ^ {z _ {i k}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.472 + ], + "angle": 0, + "content": "where \\(d_{i}\\) is short for \\(d(\\mathbf{x}_i)\\); \\(t\\) is the truncation value of the SDF; \\(\\mathbb{1}_{d_i\\in [-t,0)}\\) is an indicator function which equals one when \\(d_{i}\\in [-t,0)\\) and zero otherwise; \\(\\mathcal{N}\\big(\\cdot |d_{\\pmb{\\theta}_k}(\\mathbf{x}_i),\\sigma^2\\big)\\) is a Gaussian distribution with mean \\(d_{\\pmb{\\theta}_k}(\\mathbf{x}_i)\\) and variance \\(\\sigma^2\\). Consequently, our goal is to find out the optimal \\(\\pmb{\\theta}_k\\) and the voxel-primitive correspondences \\(z_{ik}\\) simultaneously, which maximize the likelihood of the target SDF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.472, + 0.548 + ], + "angle": 0, + "content": "Correspondence Marching: We assume that \\( z_{ik} \\) follows a Bernoulli prior distribution \\( B(p_0) \\), independent of \\( \\theta_k \\). Then, given the target SDF \\( d_i \\) and the current primitive estimation \\( \\theta_k \\), we are able to infer the posterior correspondence \\( z_{ik} \\) by the Bayes' rule" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.553, + 0.47, + 0.589 + ], + "angle": 0, + "content": "\\[\np \\left(z _ {i k} \\mid \\boldsymbol {\\theta} _ {k}, d _ {i}\\right) = \\frac {p \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k} , z _ {i k}\\right) p \\left(z _ {i k}\\right)}{\\sum_ {z _ {i k} \\in \\{0 , 1 \\}} p \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k} , z _ {i k}\\right) p \\left(z _ {i k}\\right)} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.82 + ], + "angle": 0, + "content": "By analyzing the posterior correspondence, we can observe that our design of the probabilistic model (Eq.(8)) possesses two desirable properties: (1) voxels labeled as inside \\((d_{i} < 0)\\) may not contribute to the current estimation of the primitive \\(\\pmb{\\theta}_{k}\\), i.e., \\(p(z_{ik} = 1|\\pmb{\\theta}_k,d_i < 0)\\in (0,1)\\subset \\mathbb{R}\\); (2) all non-negative valued voxels always contribute to any primitives, i.e., \\(p(z_{ik} = 1|\\pmb{\\theta}_k,d_i\\geq 0) = 1\\). In other words, the primitive tries to absorb the interior voxels sharing similar local geometry, while avoids occupying any voxels labeled as the exterior. Meanwhile, the probabilistic correspondence establishes a soft relationship between the primitive and the interior voxels, allowing part of the primitive to occupy some interior volumes without complying with the target signed distance values. In this way, the algorithm is able to achieve an overall better fitting quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.82, + 0.469, + 0.866 + ], + "angle": 0, + "content": "Primitive Update: Given the current estimation of the correspondences between the primitive and voxels, the parameter of the primitive is updated by" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.871, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {k} = \\underset {\\boldsymbol {\\theta} _ {k}} {\\arg \\min } \\sum_ {\\mathbf {x} _ {i} \\in \\mathbf {V} _ {a}} P _ {i k} \\| d _ {\\boldsymbol {\\theta} _ {k}} (\\mathbf {x} _ {\\mathrm {i}}) - d _ {i} \\| _ {2} ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.093, + 0.547, + 0.105 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.117, + 0.891, + 0.156 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P _ {i k} = p \\left(z _ {i k} = 1 \\mid \\boldsymbol {\\theta} _ {k} ^ {p r e v}, d _ {i}\\right) \\tag {11} \\\\ \\mathbf {V} _ {a} = \\left\\{\\mathbf {x} _ {\\mathrm {i}} \\in \\mathbf {V} | d _ {\\boldsymbol {\\theta} _ {k} ^ {p r e v}} (\\mathbf {x} _ {i}) \\in [ - a, a ] \\subset \\mathbb {R} \\right\\} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.892, + 0.395 + ], + "angle": 0, + "content": "\\(\\theta_{k}^{prev}\\) is the previous estimation of the primitive parameters. \\(\\mathbf{V}_{a}\\) is an adaptive subset of the complete voxels \\(\\mathbf{V}\\), which includes the voxels close to the surface of the previous estimated primitive. \\(a\\) defines the distance threshold of the activated region. Only voxels in \\(\\mathbf{V}_{a}\\) are activated during the optimization. The reason why we use an adaptive subset instead of the complete voxel space is twofold. As discussed in Sec. 3.1, both the source and the target SDF are truncated within a vicinity of zero. Therefore, small variations around \\(\\theta_{k}^{prev}\\) (i.e., small changes on the shape of the primitive surface) have minor if not zero effects on the source SDF values evaluated at voxels distant to the primitive surface. Moreover, the size of \\(\\mathbf{V}_{a}\\) is much smaller than the complete set, and thus providing a significant boost in performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.893, + 0.533 + ], + "angle": 0, + "content": "The correspondence marching and the primitive updating alternate until convergence. This process is akin to the EM algorithm [13]. The scope of voxels from which a primitive can be extracted is 'marched' with the progressive variation of the primitive shape, and simultaneously the parameters of the primitive are optimized based on the scope of voxels. We apply the optimization method proposed in [24] to avoid local minima. More derivation and implementation details can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.75, + 0.561 + ], + "angle": 0, + "content": "3.5. Fail-safe Auto-degeneration" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.569, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The primitives are roughly initialized in the connectivity marching step and further grow to capture local geometric shapes. Since the probabilistic marching is based on optimization, an important question to ask is if the recovered primitives can always converge to a proper shape, regardless of poor initialization. We demonstrate that our probabilistic marching strategy is robust to initialization. Firstly, our method possesses a feature called auto-degeneration i.e. the primitive will autonomously degenerate towards a point when accidentally initialized far from the target volumes. This is because, under this circumstance, only variations which shrink the volume of the primitive can decrease the difference between the truncated target and source SDF. Secondly, when the primitive is initialized near or inside the target volumes, the marching process encourages the primitive to converge to one nearest local target shape by analyzing the posterior correspondences. Fig.2(c) illustrates the examples of the above properties. After the primitive marching, our algorithm removes the degenerated primitives. As a fail-safe measure, primitives which significantly contradict the target SDF is also removed. Detailed implementations can be found in the Supplementary Material." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8775" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.844, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 3. Shape Abstraction results on the ShapeNet dataset. From left to right: SQs [31], Non-parametric Bayesian (NB) [47], Marching-Primitives with ellipsoids (MPE), Marching-Primitives with superquadrics (MPS), and the ground truth (pre-processed watertight mesh)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.362, + 0.212, + 0.379 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.475, + 0.54 + ], + "angle": 0, + "content": "In this section, we conduct several experiments to demonstrate the high accuracy and generalizability of our proposed method. First, we show the shape abstraction results on various datasets, ranging from daily objects to human body models. Furthermore, we study the impact of the grid resolution and the truncation threshold on our algorithm. In the end, we apply our algorithm on a large-scale complex indoor scene, which contains a mixed combination of furniture such as sofas, tables, books, lamps, etc. Implementation details can be found in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.293, + 0.56 + ], + "angle": 0, + "content": "4.1. Evaluation on Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Baselines: We compare our method with both the state-of-the-art learning-based method [31] and the computational method [47], which infer superquadric abstractions of the input objects. For convenience, we refer to [31] as SQs, and [47] as NB. We use the official codes and follow the implementation details as stated in these papers, respectively. We do not compare with the cuboid-based algorithms [40, 48, 50], since they focus more on the semantic level abstraction and are limited in expressiveness and accuracy. A detailed performance comparison between the cuboid and superquadric abstractions can be found in [31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Datasets: We evaluate on two commonly used datasets, the ShapeNet [6] and the DFAUST [4]. In both of the datasets, we are only provided with triangular meshes. Therefore, we need to transform the meshes into discrete SDF representations. It is straightforward to calculate the signed distance value of a point to a watertight mesh. However, the ShapeNet is a human-made synthetic dataset containing many non-watertight meshes formed by nonvolumetric 2D surfaces, self-intersecting or overlapped triangles. Therefore, we pre-process the original ShapeNet models into watertight meshes and then generate the SDF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.456 + ], + "angle": 0, + "content": "representation with the fast marching algorithm, following the same procedure in [8]. In consideration of fairness and consistency, we use the pre-processed watertight meshes as the common ground truth and the inputs to SQs and NB. For the DFAUST dataset, we generate the SDF directly since the provided meshes are already watertight." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.895, + 0.52 + ], + "angle": 0, + "content": "Metrics: Following [31,47] we use the Chamfer-L1 distance and the volumetric intersection over union (IoU) as the quantitative evaluation metrics. The computation of the metrics is detailed in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Results on ShapeNet: We experiment on 14 categories from the ShapeNet dataset. We split the dataset randomly into the training \\((80\\%)\\) and testing \\((20\\%)\\) sets [9], where we train one SQs model per category on the training sets and evaluate all the methods on the testing sets. For our method, we use the SDF representation discretized on a voxel grid of size \\(100^3\\) and range \\([-0.5, 0.5]^3\\). Other than superquadrics, we also test our method with ellipsoids as the base primitive, which is a special case of the superquadric representation when the shape parameters \\(\\epsilon_1, \\epsilon_2\\) are fixed to 1. The quantitative results are summarized in Table. 1. We also demonstrate the qualitative comparison among different shape abstraction methods in Fig.3. Our method outperforms all the baselines, even implemented with less expressive ellipsoids. The computational method generally performs better than the learning-based method. This is because the parameters of the primitive are so sparse and geometrically interrelated that they are difficult to get mapped from a high dimensional input by a neural network. Compared with NB, our method has richer details, clearer edges, and more importantly does not occupy the exterior space. Two factors contribute to the advantages. The first one is the extra geometric information embedded in the SDF representation, which eliminates the inherent interior/exterior ambiguity of point clouds. The other one is our special de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "8776" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.878, + 0.352 + ], + "angle": 0, + "content": "
CategoryChamfer-L1IoU
SQs [31]NB [47]MPE(Ours)MPS(Ours)SQs [31]NB [47]MPE(Ours)MPS(Ours)
airplane0.0370.0230.0210.0190.4410.6710.7310.768
bench0.0560.0280.0200.0200.2380.5790.7300.819
bottle0.0470.0330.0260.0170.6860.6650.8860.924
cabinet0.0590.0360.0370.0280.3940.6660.8400.948
can0.0660.0360.0360.0220.7060.5530.9080.950
chair0.0680.0270.0230.0200.3000.6850.7850.871
lamp0.0720.0290.0220.0210.2340.5890.7500.802
speaker0.0640.0410.0370.0330.3460.6560.8580.920
mailbox0.0950.0260.0260.0240.3330.6940.8020.905
rifle0.0380.0200.0190.0190.4460.7320.7440.811
sofa0.0540.0370.0290.0230.4970.7260.8570.940
table0.0700.0240.0240.0220.2470.7450.8180.932
phone0.0400.0210.0230.0210.6810.8720.8910.947
watercraft0.0480.0320.0220.0220.4650.6180.7930.836
mean0.0570.0280.0240.0220.3680.6740.7930.870
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.362, + 0.892, + 0.377 + ], + "angle": 0, + "content": "Table 1. Quantitative results on Shapenet. MPE and MPS are short for Marching-Primitives with ellipsoids and superquadrics, respectively" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.471, + 0.442 + ], + "angle": 0, + "content": "sign of the probabilistic generative model, as discussed in Sec.3.4. It is interesting to observe that our algorithm can extract abstractions from objects not intuitively depictable by the primitives." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.685 + ], + "angle": 0, + "content": "Results on DFAUST: We also evaluate on the DFAUST dataset of human body models. We follow the split settings in [29] to train the learning-based method and evaluate all the methods on the testing set. The SDF is discretized on a grid of size \\(64^3\\) and side length 1. Similar to the ShapeNet experiment, we test our method with both superquadrics and ellipsoids. The results are shown in Fig.4. Our method can accurately capture various postures, while the baselines fail to distinguish different body parts in some cases. The abstraction quality of SQs is much better on this dataset compared with the ShapeNet, since human bodies share a common articulated structure that can be captured by the neural network. We observe that the ellipsoid abstraction also achieves satisfying accuracy. This is because the human body mostly consists of rounded shapes compared with man-made objects in the ShapeNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.262, + 0.712 + ], + "angle": 0, + "content": "4.2. Performance Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "This section investigates how the Marching-Primitives algorithm behaves with different voxel grid sizes and distance truncation thresholds. We experiment on the Armadillo from the Stanford 3D scanning repository [10]. First, we discretize the model on the voxel grids of different resolutions. The qualitative abstraction results are visualized in Fig. 5, and the quantitative results are summarized in Fig. 6(a). When the input grid resolution is low, the recovered primitive representation is relatively abstract. With the increase of the grid resolution, our method is able to extract detailed abstraction more faithful to the target shape. This is because the higher the resolution, the more geomet" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.382, + 0.87, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.738, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Figure 4. Shape abstraction results on the DFAUST dataset. From left to right: SQs [31], NB [47], MPE, MPS, and the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ric information our method can utilize to guide the primitive marching process. As discussed in Sec.3.1, we use a truncated version of the target and source SDF. Therefore, we also evaluate the performance on different truncation thresholds, starting from 0.1 to 4 times the interval of the voxel grids. The results are shown in Fig.6(b). The abstraction accuracy increases as the threshold decreases at first. The reason lies in that the approximated superquadric" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8777" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.092, + 0.25, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.192, + 0.212, + 0.2 + ], + "angle": 0, + "content": "32" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.094, + 0.36, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.192, + 0.316, + 0.2 + ], + "angle": 0, + "content": "\\(64^{3}\\)" + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.095, + 0.47, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.192, + 0.425, + 0.2 + ], + "angle": 0, + "content": "96" + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.094, + 0.58, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.192, + 0.535, + 0.2 + ], + "angle": 0, + "content": "128" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.095, + 0.689, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.192, + 0.648, + 0.2 + ], + "angle": 0, + "content": "\\(256^{3}\\)" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.096, + 0.801, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.192, + 0.784, + 0.2 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.892, + 0.235 + ], + "angle": 0, + "content": "Figure 5. Visualization of the shape abstractions on different grid resolutions. Abstraction accuracy increases with the input grid resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.47, + 0.332 + ], + "angle": 0, + "content": ".source) SDF converges to the true value when truncated close to the surface. If we further decrease the threshold, however, we observe an acute decrease in accuracy. This is because an overly small truncation threshold corrupts the original geometric information embedded in the target SDF, as well as the smoothness of the cost function (Eq. (10))." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.352, + 0.272, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.353, + 0.464, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.519 + ], + "angle": 0, + "content": "Figure 6. Quantitative results on shape abstraction accuracy. (a) Abstraction accuracy versus input grid resolution. (b) Abstraction accuracy versus truncation ratio, evaluated at grid resolution \\(128^{3}\\). We evaluate on two metrics: Chamfer-\\(L_{1}\\) and IoU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.254, + 0.558 + ], + "angle": 0, + "content": "4.3. Scene Abstraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Other than single objects, we also test the capability of our method in representing complex scenes with geometric primitives. We experiment on a real-world indoor scene called Reading Room [49]. The scene is captured by an Asus Xtion Pro Live camera. The RGB-D scans are fused utilizing [19] (an SDF-based 3D reconstruction algorithm), and further fine-tuned with the method proposed in [49]. However, the SDF representation of the scene is not available publicly. Therefore, we pre-process the scene mesh by removing the floor plane, filling the holes, and transforming it into the SDF representation with grid size \\(400^3\\). The scene abstraction task is much more difficult compared with the object-level task because various items with great differences in size and shape are present in the very same space. We compare our abstraction result with the mesh extracted by the marching cubes algorithm on the same discrete SDF representation, as shown in Fig.7. Our representation is not only visually satisfying but also contains abundant geometric information the mesh lacks, since it is a continuous SDF approximation to the input discrete SDF. Moreover, our highly sparse representation is only 8.2KB in size, while the discrete SDF occupies 203MB." + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.242, + 0.822, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.345, + 0.822, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Figure 7. Qualitative result of the scene abstraction. (a) Superquadric abstraction by the Marching-Primitives. (b) Mesh generated by the marching cubes [25] from the same SDF." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.619, + 0.528 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.781 + ], + "angle": 0, + "content": "We propose the Marching-Primitives, the first algorithm to extract primitive-based abstractions from the volumetric SDF representation. Our method outperforms the state-of-the-art in terms of accuracy and generalizability on both synthetic and real-world datasets. Our primitive-based representation is sparse, accurate, generalizable, and can be expressed analytically without training, which we believe will facilitate and inspire further explorations in scene understanding, 3D reconstruction, and robot motion planning. However, our algorithm has the limitation of not being able to properly extract object parts thinner than the interval of the voxel grid. This problem can be solved by either increasing the grid resolution or detecting and thickening the invalid parts. Future work includes the parallelization of the marching process and inferring semantic-level interpretations from the current primitive-based geometric features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This research is supported by NUS Startup grants A-0009059-02-00 and A-0009059-03-00, National Research Foundation, Singapore, under its Medium Sized Centre Programme - Centre for Advanced Robotics Technology Innovation (CARTIN), subaward A-0009428-08-00, and AME Programmatic Fund Project MARIO A-0008449-01-00, and JHU internal funds." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8778" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] R. Bajcsy and F. Solina. Three dimensional object representation revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 231-240, 1987. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.214 + ], + "angle": 0, + "content": "[2] A. H. Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1(1):11-23, 1981. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[3] I. Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2):115, 1987. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[4] F. Bogo, J. Romero, G. Pons-Moll, and M. J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[5] D. E. Breen, S. Mauch, and R. T. Whitaker. 3d scan conversion of csg models into distance volumes. In Proceedings of the 1998 IEEE Symposium on Volume Visualization, pages 7-14, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.47, + 0.43 + ], + "angle": 0, + "content": "[6] A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[7] L. Chevalier, J. Jaillet, and A. Baskurt. Segmentation and superquadric modeling of 3D objects. In The 11-th International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision'2003, WSCG 2003, 2003. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[8] G. Chou, I. Chugunov, and F. Heide. Gensdf: Two-stage learning of generalizable signed distance functions. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[9] C. B. Choy, D. Xu, J. Gwak, K. Chen, and S. Savarese. 3dr2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-644. Springer International Publishing, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[10] B. Curless and M. Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[11] A. Dai, D. Ritchie, M. Bokeloh, S. Reed, J. Sturm, and M. Nießner. Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] A. Dai, C. Ruizhongtai Qi, and M. Niessner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[13] A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[14] B. Deng, K. Genova, S. Yazdani, S. Bouaziz, G. Hinton, and A. Tagliasacchi. Cvxnet: Learnable convex decomposition." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 31-41, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[15] W. Dong, Q. Wang, X. Wang, and H. Zha. Psdf fusion: Probabilistic signed distance function for on-the-fly 3d data fusion and scene reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 701-717, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[16] K. Genova, F. Cole, D. Vlasic, A. Sarna, W. Freeman, and T. Funkhouser. Learning shape templates with structured implicit functions. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7153-7163, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[17] A. D. Gross and T. E. Boult. Error of fit measures for recovering parametric solids. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 690-694, 1988. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[18] Z. Hao, H. Averbuch-Elor, N. Snavely, and S. Belongie. Dualsf: Semantic shape manipulation using a two-level representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7631–7641, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[19] S. Izadi, D. Kim, O. Hilliges, D. Molyneaux, R. Newcombe, P. Kohli, J. Shotton, S. Hodges, D. Freeman, A. Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[20] D. K. Katherine and S. S. Elizabeth. Core systems in human cognition. In From Action to Cognition, volume 164 of Progress in Brain Research, pages 257-264. Elsevier, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[21] Y. Kawana, Y. Mukuta, and T. Harada. Neural star domain as primitive representation. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7875-7886. Curran Associates, Inc., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[22] A. Leonardis, A. Jaklic, and F. Solina. Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(11):1289-1295, 1997. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[23] Y. Liao, S. Donne, and A. Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2916-2925, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] W. Liu, Y. Wu, S. Ruan, and G. S. Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2676-2685, June 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[25] W. E. Lorensen and H. E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] R. A. Newcombe, D. Fox, and S. M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] C. Niu, J. Li, and K. Xu. Im2Struct: Recovering 3D shape structure from a single RGB image. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8779" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[28] J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[29] D. Paschalidou, L. V. Gool, and A. Geiger. Learning unsupervised hierarchical part decomposition of 3D objects from a single RGB image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[30] D. Paschalidou, A. Katharopoulos, A. Geiger, and S. Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3215, June 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.402 + ], + "angle": 0, + "content": "[31] D. Paschalidou, A. O. Ulusoy, and A. Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[32] A. Pentland. Perceptual organization and the representation of natural form. Artificial Intelligence, 28(3):293-331, 1986. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[33] A. H. Quispe, B. Milville, M. A. Gutiérrez, C. Erdogan, M. Stilman, H. Christensen, and H. B. Amor. Exploiting symmetries and extrusions for grasping household objects. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3702-3708, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[34] Y. Rao, Y. Nie, and A. Dai. Patchcomplete: Learning multi-resolution patch priors for 3d shape completion on unseen categories. Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[35] S. Ruan, K. L. Poblete, H. Wu, Q. Ma, and G. S. Chirikjian. Efficient path planning in narrow passages for robots with ellipsoidal components. IEEE Transactions on Robotics (TRO), pages 1-18, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[36] S. Ruan, X. Wang, and G. S. Chirikjian. Collision detection for unions of convex bodies with smooth boundaries using closed-form contact space parameterization. IEEE Robotics and Automation Letters (RAL), 7(4):9485-9492, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[37] G. Sharma, B. Dash, A. RoyChowdhury, M. Gadelha, M. Loizou, L. Cao, R. Wang, E. Learned-Miller, S. Maji, and E. Kalogerakis. Prift: Learning to fit primitives improves few shot point cloud segmentation. In Computer Graphics Forum, volume 41, pages 39-50. Wiley Online Library, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[38] D. Smirnov, M. Fisher, V. G. Kim, R. Zhang, and J. Solomon. Deep parametric shape predictions using distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 561-570, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] F. Solina and R. Bajcsy. Recovery of parametric models from range images: the case for superquadrics with global deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(2):131-147, 1990. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[40] C. Sun, Q. Zou, X. Tong, and Y. Liu. Learning adaptive hierarchical cuboid abstractions of 3d shape collections. ACM Transactions on Graphics (TOG), 38(6):1-13, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[41] S. Tulsiani, H. Su, L. J. Guibas, A. A. Efros, and J. Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[42] N. Vaskevicius and A. Birk. Revisiting superquadric fitting: A numerically stable formulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):220-233, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[43] S. Vasu, N. Talabot, A. Lukoianov, P. Baque, J. Donier, and P. Fua. Hybridsdf: Combining deep implicit shapes and geometric primitives for 3d shape representation and manipulation. In International Conference on 3D Vision (3DV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[44] G. Vezzani, U. Pattacini, and L. Natale. A grasping approach based on superquadric models. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1579-1586, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[45] G. Vezzani, U. Pattacini, G. Pasquale, and L. Natale. Improving superquadric modeling and grasping with prior on object shapes. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 6875-6882, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[46] S. Weder, J. L. Schonberger, M. Pollefeys, and M. R. Oswald. Routedfusion: Learning real-time depth map fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[47] Y. Wu, W. Liu, S. Ruan, and G. S. Chirikjian. Primitive-based shape abstraction via nonparametric bayesian inference. In Proceedings of the European Conference on Computer Vision (ECCV), pages 479-495. Springer Nature Switzerland, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[48] K. Yang and X. Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[49] Q. Zhou, S. Miller, and V. Koltun. Elastic fragments for dense scene reconstruction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 473-480, 2013. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[50] C. Zou, E. Yumer, J. Yang, D. Ceylan, and D. Hoiem. 3D-PRNN: Generating shape primitives with recurrent neural networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8780" + } + ] +] \ No newline at end of file diff --git a/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_origin.pdf b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8852054e25f9a8e9b474cb64bc696930e0e977a9 --- /dev/null +++ b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/c067a9c9-d8c5-4894-a294-6c3c0b3f3bd8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bf3915dd40ac9db2514d1a4978e5e000e8f6bb5963c5793154f728b1f388eaf +size 2601297 diff --git a/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/full.md b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8892f92bb50099a078b458b846b10e2b7eb6d863 --- /dev/null +++ b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/full.md @@ -0,0 +1,303 @@ +# Marching-Primitives: Shape Abstraction from Signed Distance Function + +Weixiao Liu $^{1,2}$ Yuwei Wu $^{1}$ Sipu Ruan $^{1}$ Gregory S. Chirikjian $^{1*}$ $^{1}$ National University of Singapore $^{2}$ Johns Hopkins University +{mpewxl, yw.wu, ruansp, mpegre}@nus.edu.sg + +# Abstract + +Representing complex objects with basic geometric primitives has long been a topic in computer vision. Primitive-based representations have the merits of compactness and computational efficiency in higher-level tasks such as physics simulation, collision checking, and robotic manipulation. Unlike previous works which extract polygonal meshes from a signed distance function (SDF), in this paper, we present a novel method, named Marching-Primitives, to obtain a primitive-based abstraction directly from an SDF. Our method grows geometric primitives (such as superquadrics) iteratively by analyzing the connectivity of voxels while marching at different levels of signed distance. For each valid connected volume of interest, we march on the scope of voxels from which a primitive is able to be extracted in a probabilistic sense and simultaneously solve for the parameters of the primitive to capture the underlying local geometry. We evaluate the performance of our method on both synthetic and real-world datasets. The results show that the proposed method outperforms the state-of-the-art in terms of accuracy, and is directly generalizable among different categories and scales. The code is open-sourced at https://github.com/ChirikjianLab/Marching-Primitives.git. + +# 1. Introduction + +Recent years have witnessed great progress in the areas of 3D shape representation and environmental perception. Low-level representations such as surface meshes, point clouds, and occupancy grids are widely used as inputs to high-level computer vision algorithms and artificial intelligence tasks. They have the advantage of being able to represent and visualize objects with high accuracy and rich local geometric features. However, the low-level representations are ineffective in delivering a general and intuitive sense of structural geometry as well as part-level scene understanding. Studies [3,20] show that human vision, unlike + +![](images/57f1bb3b41f9000f06445b665e2541aec27b71d587e27ff0fb981e413e3f9432.jpg) +Figure 1. Primitive-base representation versus mesh. For each pair of objects, the left one is the superquadratic abstraction obtained by our algorithm, and the right one is the original mesh. The mesh of the chair is 6MB in size, while our representation only needs 4KB. An SDF representation discretized on a $128^{3}$ voxel grid occupies 19MB. Our abstraction is equivalent to an implicit continuous SDF, which is an approximation to the discrete SDF. + +computer vision, tends to perceive and understand scenes as combinations of simple primitive shapes. Human beings perform well and robustly in complex tasks, providing a basic geometric description of the scene is available [32]. Therefore, researchers turn to exploring the possibility of interpreting complex objects and scenes with basic geometric primitives. Taking advantage of the primitive-based representation, many higher-level tasks, such as segmentation [14, 16, 21, 30], scene understanding [29, 31, 41, 47], grasping [33, 44, 45] and motion planning [35, 36], are able to be solved efficiently. + +However, it still remains challenging to extract primitive-based abstractions from low-level representations. Starting from the 1990s, Solina et al. [1, 17, 39] aim to extract a single superquadric representation from a simple object by minimizing the least-square error between the primitive and the measured points. Later in [7, 22], their method is extended to represent more complex objects with multiple primitives. More recently, the authors of [24, 47] reformulate the task as a probabilistic inference problem with enhanced accuracy and robustness to noise and outliers. At the same time, with the surge of data-driven techniques, researchers attempt to train neural networks to infer cuboids [27,38,41,48,50] and superquadrics [29,31] representations in an end-to-end fashion. However, both the computational and learning-based approaches have their own limitations. The computational methods are vulnerable to the inherent ambiguity of the point-to-surface relationship. For exam + +ple, the algorithms tend to fill empty spaces of a non-convex object with primitives by mistake, due to the inside/outside ambiguity of a surface depicted by a set of points [24, 48]. The main drawback of the learning approaches lies in the lack of generalizability beyond the object category on which the model is trained [24, 31, 47, 48]. Also, the shape abstraction accuracy is inferior to the computational methods. + +The signed distance function (SDF) has been a successful 3D volumetric representation in varieties of computer vision and graphics tasks. It is the basic framework for many classic 3D reconstruction algorithms such as TSDF volume reconstruction [10, 15], KinectFusion [19], and Dynamic-Fusion [26]. Recently, the SDF representation is adapted to the deep learning frameworks, and exhibits boosted potentials in shape encoding [8, 18, 28, 43], surface reconstruction [23, 46], and shape completion [11, 12, 34]. Usually, triangular mesh surfaces are extracted from the SDF representation with the marching cubes algorithm [25]. Point cloud and occupancy grid representations are also obtained by keeping the vertices of the meshes and the sign of each voxel point, respectively. The SDF is among the most informative 3D representations since it encodes not only the surface geometry but also the distance and side of a point relative to the shape. Meanwhile, it is easily achievable via range images from 3D sensors [10], or learnable from other input modalities [8, 18, 28]. Since we are able to extract meshes from an SDF, it is natural to think about the possibility of extracting primitives as well. Furthermore, the primitive-based abstraction is a continuous interpretation of the complete geometric information encoded in the original discrete SDF, but requires much less storage size (Fig.1). + +Motivated by the aforementioned facts and the bottleneck of the current shape abstraction algorithms, we proposed a general shape abstraction method by reasoning directly on the informative SDF representation. The goal of our method is to find a combination of geometric primitives whose underlying SDF values match the target values evaluated on the evenly spaced discrete grid points (Sec. 3.1 and Sec. 3.2). To solve this problem, we propose a two-step iterative algorithm called the Marching-Primitives. Our algorithm 'marches' on two domains: the signed distance domain and the voxelized space domain, alternately. Firstly, the connectivity of volumes are analyzed by generating isosurfaces on a sequence of decreasing levels of negative signed distances (Sec.3.3). By doing so, volumes of interest (VOIs) where primitives are likely to be encoded can be identified sequentially. In the second step, for each of the VOIs, our algorithm marches on the neighbouring voxels to infer their probabilistic correspondences to the primitive and simultaneously optimizes the shape and pose of the primitive (Sec.3.4). After the primitive representation of a VOI is achieved, the fitted volumes are deactivated from the voxel grid. Our algorithm continues marching on the signed + +distance domain until it approaches zero, i.e., all the interior volumes of the SDF have been captured by the recovered primitives. We compare our algorithm with the state-of-the-art of both the computational and learning-based approaches on the ShapeNet object dataset [6] and D-FAUST human shape dataset [4] (Sec. 4.1). We also study the performance of our algorithm on different conditions (Sec. 4.2). Finally, we demonstrate the scene abstraction result of the Stanford Reading Room [49], which contains several pieces of furniture of various categories (Sec. 4.3). + +# 2. Related Work + +SDF Representation: The SDF can be stored in two different ways: discrete or continuous. A majority of computer vision and graphics algorithms are built on the SDF discretized on a 3D grid of voxel points. The signed distances are stored on each of the corresponding voxel points. The authors of [10, 19, 26] pioneer in fusing several noisy range images into a single discrete SDF. Their work is widely applied in 3D reconstruction and plays an important role in robotics tasks such as simultaneous localization and mapping. The discrete SDF is also a promising input/output representation for 3D deep learning [11, 12, 34]. Recently, it becomes popular to encode shapes as a continuous SDF with neural networks [28]. Vasu et al. [43] further improve the shape encoding quality by enforcing local regularities with geometric primitives. In [8], the authors adopt a two-stage meta-learning approach to further extend the generalization capabilities of neural SDF. With the deep neural network, it becomes possible to infer SDF representations from partially observed 3D inputs or even images. Both the discrete and the continuous SDF are implicit representations of geometric surfaces. To extract the explicit surface from the SDF, continuous SDFs need to be discretized on a voxel grid first and then conduct the marching cubes [25]. This method allows high-quality rendering of the objects, however, surface meshes are non-sparse and contain no structural level information. Our method provides an alternative approach to describe the underlying object in the SDF. Instead of meshes, we directly extract a collection of sparsely parameterized primitives from the SDF. Other than that, our primitive-based representation itself is also a concise yet continuous SDF approximation to the original discrete SDF. + +Computational Shape Abstraction: The most well-studied primitive for computational shape abstraction is the superquadric, due to its extensive shape vocabulary including cuboids, ellipsoids, cylinders, octohedra, and many shapes in between (e.g., cuboids with rounded edges). It is first proposed as a versatile modeling element for complex objects in computer graphics [2, 32]. Later, Solina et al. propose a method to conduct abstraction of simple objects from range images with a single superquadric [17, 39]. Leonardis et al. [22] and Chevalier et al. [7] further extend + +the previous work to recover complex objects with multiple superquadrics with a Split-and-Merge strategy. A numerical instability problem is addressed and revisited in [42]. The authors introduce an auxiliary function in the unstable region and receive a better abstraction accuracy. More recently, Liu et al. [24] formulate the problem in a probabilistic fashion and propose a geometric strategy to avoid local optimum, bringing a significant improvement in robustness to outlier and fitting accuracy. Wu et al. [47] extend and recast the work as a nonparametric Bayesian inference problem so as to improve the applicability on complex shapes. To the best of the authors' knowledge, the existing computational methods are all based on range images or point clouds, which suffer from geometric ambiguities [48]. In contrast, our method takes advantage of the abundant geometric information encoded in the SDF and is easily compatible with other computer vision algorithms based on the SDF representation. + +Learning-based Shape Abstraction: The learning-based method is first seen in [41]. Tulsiani et al. propose a 3D convolutional neural network (CNN) to learn shape abstractions with cuboids from the occupancy grid. Sun et al. [40] design an adaptive hierarchical cuboid representation and introduce an unsupervised approach to learn to extract the parameters for the representation. Yang et al. [48] train a variational auto-encoder network to transform point clouds into parametric cuboids. Other than cuboids, researchers also seek to extract spheres and ellipsoids representations from objects. Hao et al. [18] combine the neural SDF with the spherical representation by sharing a same latent layer. In [37], ellipsoids or cuboids are extracted to help segment the input point cloud. However, a single type of primitive has very limited expressiveness. Therefore, Paschalidou et al. [29, 31] turn to training neural networks to conduct abstractions with the superquadrics as the atomic elements. Learning-based approaches are versatile in dealing with different input sources. They are able to make shape abstractions from point clouds, voxel grids, or even RGB images which are so ill-conditioned that the computational approaches have little chance of working. However, learning-based approaches rely heavily on the training dataset and thus are less generalizable to unseen categories. Instead, our approach reasons about the primitive abstraction from a case-by-case geometric perspective, which provides an inherent advantage in generalizability and accuracy. + +# 3. Method + +# 3.1. Preliminary + +The discrete SDF is a volumetric surface representation built on a voxel grid $\mathbf{V} = \{\mathbf{x}_i\in \mathbb{R}^3,i = 1,2,\dots,N\}$ . A scalar $d(\mathbf{x}_i)$ is assigned to each grid point, which indicates + +the signed distance of $\mathbf{x}_i$ to the nearest surface. The point $\mathbf{x}_i$ lies inside the surface if $d(\mathbf{x}_i) < 0$ and outside otherwise. Typically, the surface mesh of an object is extracted from the SDF by the marching cubes algorithm [25]. In this paper, we call the input SDF $d(\mathbf{x}_i)$ the target SDF. + +For the primitive representation, we select the superquadrics [2], a family of geometric primitives defined by the implicit equation + +$$ +f (\mathbf {x}) = \left(\left(\frac {x}{a _ {x}}\right) ^ {\frac {2}{\varepsilon_ {2}}} + \left(\frac {y}{a _ {y}}\right) ^ {\frac {2}{\varepsilon_ {2}}}\right) ^ {\frac {\varepsilon_ {2}}{\varepsilon_ {1}}} + \left(\frac {z}{a _ {z}}\right) ^ {\frac {2}{\varepsilon_ {1}}} = 1 \tag {1} +$$ + +The superquadric family is only encoded by 5 parameters (shape parameters $\epsilon_{1},\epsilon_{2}\in [0,2]\subset \mathbb{R}$ , and scale parameters $a_{x},a_{y},a_{z}\in \mathbb{R}_{>0}$ ), but has an extensive shape vocabulary. Note that the shape parameters can exceed 2, resulting in nonconvex shapes. However, practically in most studies [24,31] and also in our paper, we limit them within the convex region as defined above. The points $\mathbf{x} = [x,y,z]\in \mathbb{R}^3$ satisfying Eq. (1) form the surface of the superquadric. In this paper, we also include the Euclidean transformation $g\in SE(3)$ , i.e. 3 Euler angles for rotation $\mathbf{R}\in SO(3)$ and 3-dimensional translation $\mathbf{t}\in \mathbb{R}^3$ , to parameterize a superquadric with a general pose. In total, we denote a superquadric with a vector $\pmb{\theta}$ of 11 elements. According to Eq. (1), we can approximate the signed distance of a grid point $\mathbf{x}_i$ to a general posed superquadric $\pmb{\theta}$ explicitly by + +$$ +d _ {\boldsymbol {\theta}} \left(\mathbf {x} _ {i}\right) = \left(1 - f ^ {- \frac {\epsilon_ {1}}{2}} \left(g ^ {- 1} \circ \mathbf {x} _ {i}\right)\right) \| g ^ {- 1} \circ \mathbf {x} _ {i} \| _ {2} \tag {2} +$$ + +We are not able to use the exact SDF of superquadrics, because it has no analytical solution and is only achievable by numerical optimization [5]. Eq.(2) is the radial distance [17, 24] of $\mathbf{x}_i$ to the superquadric surface, which converges to the exact signed distance as $d_{\theta}(\mathbf{x}_i)\to 0$ . Therefore, we truncate both the input target SDF $d(\cdot)$ and the primitive SDF $d_{\theta}(\cdot)$ within the vicinity of zero to ensure the approximation accuracy. + +# 3.2. Problem Formulation + +To obtain a primitive-based abstraction from the target SDF, we seek a combination of primitives $\Theta \doteq \{\theta_{k}, k = 1,2,\dots,K\}$ whose underlying signed distances measured on the voxel points (we call it the source SDF in contrast to the target SDF) match the target SDF, that is + +$$ +\boldsymbol {\Theta} = \underset {\boldsymbol {\Theta}} {\arg \min } \sum_ {\mathbf {x} _ {i} \in \mathbf {V}} \underset {\boldsymbol {\theta} _ {k} \in \boldsymbol {\Theta}} {\min } \left\| d _ {\boldsymbol {\theta} _ {k}} \left(\mathbf {x} _ {i}\right) - d \left(\mathbf {x} _ {i}\right) \right\| _ {2} ^ {2} \tag {3} +$$ + +Our problem formulation is simple and intuitive, however, it is intractable to solve directly. First, the number of primitives $K$ needed is unknown. Also, the correspondences between the voxel points $\mathbf{x}_i$ and the primitives $\theta_k$ are unknown a priori. Therefore, it makes Eq.(3) a chicken-and-egg problem: on the one hand, if we know the set of voxel + +points contributing to a same primitive, we are able to solve for the parameters of the primitive by optimization; on the other hand, if we know the configurations of the primitives, we are able to tell the belongings of each voxel. Other than those factors, the computation complexity itself is prohibitively high, because we need to evaluate hundreds of thousands or even millions of voxel points. This makes it infeasible to operate directly on the complete dense voxel grid. To tackle these difficulties, we propose an iterative algorithm, the Marching-Primitives, which simultaneously solves the correspondences and primitive parameters efficiently. It is worth noting that our method is generalizable beyond the superquadrics. Any volumetric primitives with easily accessible SDF representations can be adapted to our framework. + +# 3.3. Iterative Connectivity Marching + +Instead of setting a predefined number of primitives, our method starts from an empty set of primitives and grows as needed. This is realized by analyzing the connectivity (in this paper, we use the 26-connectivity) of the voxels at different levels of signed distance. Given the target SDF, our algorithm checks the connectivity of voxels whose signed distance is less than a sequence of thresholds + +$$ +T ^ {c} \doteq \left\{t _ {1} ^ {c}, t _ {2} ^ {c}, \dots \right\}, \quad t _ {1} ^ {c} = \min _ {\mathbf {x} _ {i} \in \mathbf {V}} d (\mathbf {x} _ {i}), \quad t _ {m + 1} ^ {c} = \alpha t _ {m} ^ {c} (4) +$$ + +where $\alpha \in (0,1) \subset \mathbb{R}$ is the common ratio, resulting in a geometric sequence exponentially decaying to zero. Each threshold $t_m^c$ defines a set of disjoint isosurfaces $S_m$ , encompassing the connected voxels whose signed distances are less than the threshold. + +$$ +S _ {m} = \left\{S _ {k}, k = 1, 2, \dots , | S _ {m} | \right\} \tag {5} +$$ + +$|S_{m}|$ denotes the number of the disjoint isosurfaces. We construct a subset $\bar{S}_m$ from $S_{m}$ + +$$ +\bar {S} _ {m} = \left\{S _ {k} \in S _ {m}, | S _ {k} | \geq N _ {c} \right\} \subseteq S _ {m} \tag {6} +$$ + +where $|S_{k}|$ is the number of the connected voxels within the isosurface $S_{k}$ . $\bar{S}_{m}$ is the set of the isosurfaces encompassing no less than $N_{c}$ connected voxels. The connectivity marching starts from the innermost threshold $t_1^c$ , where the resulting $\bar{S}_1$ might be empty. The marching continues on the sequence (increasing the threshold) until $\bar{S}_{m}$ is nonempty. Each of the connected volumes in $\bar{S}_{m}$ (we call it a VOI) is an ideal starting point for growing a primitive, because: (1) those volumes are among the most interior of the target SDF, corresponding to most prominent part of the geometry; (2) disconnected and weakly connected volumes can be separated apart; (3) the SDF of a primitive can also be interpreted as layers of isosurfaces, and thus sharing similar geometric structure; and furthermore (4) the size of the + +![](images/541b8c21aba2400bafdf17b5742b2ff63e2ba1278d04bf634b9a29e005fb5c9f.jpg) +(b) + +![](images/733e9343c28217bbf6105e987c28d06a8731b785debb53cfbfeecbd6fe5d6222.jpg) +Figure 2. Visualizations of concepts. (a) A 2D illustration of two marched isosurfaces separating the shape into two VOIs. (b) A simple example of the Marching-Primitives algorithm. The first figure is the input target SDF (voxels with signed distances larger than the truncation threshold are left blank for visualization). Figures b2-b4 illustrate the primitives marched on the sequentially detected VOIs. (c) Auto-degeneration and probabilistic marching. The green circles on the left column indicate different initialization configurations. The ones on the right are the marched results. + +![](images/4c7ce3733d47b574f9a15a7fe36926980be4e2f124f678c40bc16f0857f5bff2.jpg) +(c) + +![](images/37900fffbd83686779db3bb95e8873d09b00a75bbf6898d7576296d0ba4a18ac.jpg) + +![](images/464e7a7119bacef703abb393dba94ea8309839f3d4a589acec106c94b0ac73c8.jpg) + +![](images/6aba9cb1f68dd8028a8d0229accefa505a5f5cc1ffd126a08b33141cafad435e.jpg) + +![](images/c54279062a4659964a7ceb1352d5743ac4f4975f0335b5c3991eb64971443cb9.jpg) + +![](images/94bcccd0ffd964e03398687f626dc41983e8413c6aecac1ad454388e73db1503.jpg) + +![](images/fd94f74df9d3d67e295fd3f292b5bd802585e0948eb205288ca8ec9e0191ae45.jpg) +Initialization + +![](images/c0b9610b7201427b289e58d7075253273fa7e9e19d2fa816249ac6e4da411d31.jpg) +Convergence + +volumes is large enough to provide sufficient geometric information for the primitive recovery. We illustrate the idea of isosurfaces in Fig.2(a). For each VOI, we initialize a primitive as an ellipsoid, with scales proportional to the size of its smallest bounding-box. The details for the primitive initialization are demonstrated in the Supplementary Material. The procedure of how a primitive marches from the initial guess to capture the local geometry is detailed in the next Sec. 3.4. After obtaining the primitive representations for all the VOIs in $\bar{S}_m$ , we subtract the voxels + +$$ +\left\{\mathbf {x} _ {i} \in \mathbf {V}, d \left(\mathbf {x} _ {i}\right) \leq 0 \wedge d _ {\boldsymbol {\theta}} \left(\mathbf {x} _ {i}\right) \leq 0 \right\} \tag {7} +$$ + +from the set of voxel grid $\mathbf{V}$ . In other words, we deactivate the voxels which are interior of the target SDF and well fitted by a recovered primitive, while the updated SDF preserves the exterior and unfitted interior volumes. The connectivity marching repeats with the updated target SDF, until the threshold marches higher than a preset limit close to zero, indicating that no prominent interior volume is left unrepresented. + +# 3.4. Probabilistic Primitive Marching + +Firstly, we explain what we mean by primitive marching. In the marching cubes algorithm [25], marching means the progressive extraction of polygonal meshes from the neighboring 8 voxel points in the discrete SDF. In each step, the algorithm focuses within the single cube formulated by the 8 voxel points, calculates triangle vertices using linear in + +terpolation, and then 'marches' to the next one. On the contrary, our method marches on the scope of voxels. That is, we gradually figure out a proper collection of voxel points from which a geometric primitive is extractable. Inspired by [24], we formulate this process as a maximum likelihood estimation problem with latent variables. + +For each isosurface $S_{k}\in \bar{S}_{m}$ obtained during the connectivity marching (Eq.(6)), it is assumed that the 'true' underlying primitive is encoded by $\theta_{k}$ , and that $z_{ik}$ is the correspondence between the primitive $\theta_{k}$ and a voxel point $\mathbf{x}_i\in \mathbf{V}$ . $z_{ik}$ is a binary variable that equals 1 when the target signed distance evaluated at the $i$ th voxel point results from the $k$ th primitive, and 0 otherwise. Then the target signed distance $d(\mathbf{x}_i)$ can be regarded as an observation generated from the probabilistic distribution + +$$ +p \left(d _ {i} \mid \boldsymbol {\theta} _ {k}, z _ {i k}\right) = \left(\frac {\mathbb {1} _ {d _ {i} \in [ - t , 0)}}{t}\right) ^ {1 - z _ {i k}} \mathcal {N} \left(d _ {i} \mid d _ {\boldsymbol {\theta} _ {k}} (\mathbf {x} _ {i}), \sigma^ {2}\right) ^ {z _ {i k}} \tag {8} +$$ + +where $d_{i}$ is short for $d(\mathbf{x}_i)$ ; $t$ is the truncation value of the SDF; $\mathbb{1}_{d_i\in [-t,0)}$ is an indicator function which equals one when $d_{i}\in [-t,0)$ and zero otherwise; $\mathcal{N}\big(\cdot |d_{\pmb{\theta}_k}(\mathbf{x}_i),\sigma^2\big)$ is a Gaussian distribution with mean $d_{\pmb{\theta}_k}(\mathbf{x}_i)$ and variance $\sigma^2$ . Consequently, our goal is to find out the optimal $\pmb{\theta}_k$ and the voxel-primitive correspondences $z_{ik}$ simultaneously, which maximize the likelihood of the target SDF. + +Correspondence Marching: We assume that $z_{ik}$ follows a Bernoulli prior distribution $B(p_0)$ , independent of $\theta_k$ . Then, given the target SDF $d_i$ and the current primitive estimation $\theta_k$ , we are able to infer the posterior correspondence $z_{ik}$ by the Bayes' rule + +$$ +p \left(z _ {i k} \mid \boldsymbol {\theta} _ {k}, d _ {i}\right) = \frac {p \left(d _ {i} \mid \boldsymbol {\theta} _ {k} , z _ {i k}\right) p \left(z _ {i k}\right)}{\sum_ {z _ {i k} \in \{0 , 1 \}} p \left(d _ {i} \mid \boldsymbol {\theta} _ {k} , z _ {i k}\right) p \left(z _ {i k}\right)} \tag {9} +$$ + +By analyzing the posterior correspondence, we can observe that our design of the probabilistic model (Eq.(8)) possesses two desirable properties: (1) voxels labeled as inside $(d_{i} < 0)$ may not contribute to the current estimation of the primitive $\pmb{\theta}_{k}$ , i.e., $p(z_{ik} = 1|\pmb{\theta}_k,d_i < 0)\in (0,1)\subset \mathbb{R}$ ; (2) all non-negative valued voxels always contribute to any primitives, i.e., $p(z_{ik} = 1|\pmb{\theta}_k,d_i\geq 0) = 1$ . In other words, the primitive tries to absorb the interior voxels sharing similar local geometry, while avoids occupying any voxels labeled as the exterior. Meanwhile, the probabilistic correspondence establishes a soft relationship between the primitive and the interior voxels, allowing part of the primitive to occupy some interior volumes without complying with the target signed distance values. In this way, the algorithm is able to achieve an overall better fitting quality. + +Primitive Update: Given the current estimation of the correspondences between the primitive and voxels, the parameter of the primitive is updated by + +$$ +\boldsymbol {\theta} _ {k} = \underset {\boldsymbol {\theta} _ {k}} {\arg \min } \sum_ {\mathbf {x} _ {i} \in \mathbf {V} _ {a}} P _ {i k} \| d _ {\boldsymbol {\theta} _ {k}} (\mathbf {x} _ {\mathrm {i}}) - d _ {i} \| _ {2} ^ {2} \tag {10} +$$ + +where + +$$ +\begin{array}{l} P _ {i k} = p \left(z _ {i k} = 1 \mid \boldsymbol {\theta} _ {k} ^ {p r e v}, d _ {i}\right) \tag {11} \\ \mathbf {V} _ {a} = \left\{\mathbf {x} _ {\mathrm {i}} \in \mathbf {V} | d _ {\boldsymbol {\theta} _ {k} ^ {p r e v}} (\mathbf {x} _ {i}) \in [ - a, a ] \subset \mathbb {R} \right\} \\ \end{array} +$$ + +$\theta_{k}^{prev}$ is the previous estimation of the primitive parameters. $\mathbf{V}_{a}$ is an adaptive subset of the complete voxels $\mathbf{V}$ , which includes the voxels close to the surface of the previous estimated primitive. $a$ defines the distance threshold of the activated region. Only voxels in $\mathbf{V}_{a}$ are activated during the optimization. The reason why we use an adaptive subset instead of the complete voxel space is twofold. As discussed in Sec. 3.1, both the source and the target SDF are truncated within a vicinity of zero. Therefore, small variations around $\theta_{k}^{prev}$ (i.e., small changes on the shape of the primitive surface) have minor if not zero effects on the source SDF values evaluated at voxels distant to the primitive surface. Moreover, the size of $\mathbf{V}_{a}$ is much smaller than the complete set, and thus providing a significant boost in performance. + +The correspondence marching and the primitive updating alternate until convergence. This process is akin to the EM algorithm [13]. The scope of voxels from which a primitive can be extracted is 'marched' with the progressive variation of the primitive shape, and simultaneously the parameters of the primitive are optimized based on the scope of voxels. We apply the optimization method proposed in [24] to avoid local minima. More derivation and implementation details can be found in the Supplementary Material. + +# 3.5. Fail-safe Auto-degeneration + +The primitives are roughly initialized in the connectivity marching step and further grow to capture local geometric shapes. Since the probabilistic marching is based on optimization, an important question to ask is if the recovered primitives can always converge to a proper shape, regardless of poor initialization. We demonstrate that our probabilistic marching strategy is robust to initialization. Firstly, our method possesses a feature called auto-degeneration i.e. the primitive will autonomously degenerate towards a point when accidentally initialized far from the target volumes. This is because, under this circumstance, only variations which shrink the volume of the primitive can decrease the difference between the truncated target and source SDF. Secondly, when the primitive is initialized near or inside the target volumes, the marching process encourages the primitive to converge to one nearest local target shape by analyzing the posterior correspondences. Fig.2(c) illustrates the examples of the above properties. After the primitive marching, our algorithm removes the degenerated primitives. As a fail-safe measure, primitives which significantly contradict the target SDF is also removed. Detailed implementations can be found in the Supplementary Material. + +![](images/0e66a71d780764a14716b2cca881026a9394a5796e79754d5f6d89cf964d2e74.jpg) +Figure 3. Shape Abstraction results on the ShapeNet dataset. From left to right: SQs [31], Non-parametric Bayesian (NB) [47], Marching-Primitives with ellipsoids (MPE), Marching-Primitives with superquadrics (MPS), and the ground truth (pre-processed watertight mesh). + +# 4. Experiments + +In this section, we conduct several experiments to demonstrate the high accuracy and generalizability of our proposed method. First, we show the shape abstraction results on various datasets, ranging from daily objects to human body models. Furthermore, we study the impact of the grid resolution and the truncation threshold on our algorithm. In the end, we apply our algorithm on a large-scale complex indoor scene, which contains a mixed combination of furniture such as sofas, tables, books, lamps, etc. Implementation details can be found in the Supplementary. + +# 4.1. Evaluation on Datasets + +Baselines: We compare our method with both the state-of-the-art learning-based method [31] and the computational method [47], which infer superquadric abstractions of the input objects. For convenience, we refer to [31] as SQs, and [47] as NB. We use the official codes and follow the implementation details as stated in these papers, respectively. We do not compare with the cuboid-based algorithms [40, 48, 50], since they focus more on the semantic level abstraction and are limited in expressiveness and accuracy. A detailed performance comparison between the cuboid and superquadric abstractions can be found in [31]. + +Datasets: We evaluate on two commonly used datasets, the ShapeNet [6] and the DFAUST [4]. In both of the datasets, we are only provided with triangular meshes. Therefore, we need to transform the meshes into discrete SDF representations. It is straightforward to calculate the signed distance value of a point to a watertight mesh. However, the ShapeNet is a human-made synthetic dataset containing many non-watertight meshes formed by nonvolumetric 2D surfaces, self-intersecting or overlapped triangles. Therefore, we pre-process the original ShapeNet models into watertight meshes and then generate the SDF + +representation with the fast marching algorithm, following the same procedure in [8]. In consideration of fairness and consistency, we use the pre-processed watertight meshes as the common ground truth and the inputs to SQs and NB. For the DFAUST dataset, we generate the SDF directly since the provided meshes are already watertight. + +Metrics: Following [31,47] we use the Chamfer-L1 distance and the volumetric intersection over union (IoU) as the quantitative evaluation metrics. The computation of the metrics is detailed in the Supplementary Material. + +Results on ShapeNet: We experiment on 14 categories from the ShapeNet dataset. We split the dataset randomly into the training $(80\%)$ and testing $(20\%)$ sets [9], where we train one SQs model per category on the training sets and evaluate all the methods on the testing sets. For our method, we use the SDF representation discretized on a voxel grid of size $100^3$ and range $[-0.5, 0.5]^3$ . Other than superquadrics, we also test our method with ellipsoids as the base primitive, which is a special case of the superquadric representation when the shape parameters $\epsilon_1, \epsilon_2$ are fixed to 1. The quantitative results are summarized in Table. 1. We also demonstrate the qualitative comparison among different shape abstraction methods in Fig.3. Our method outperforms all the baselines, even implemented with less expressive ellipsoids. The computational method generally performs better than the learning-based method. This is because the parameters of the primitive are so sparse and geometrically interrelated that they are difficult to get mapped from a high dimensional input by a neural network. Compared with NB, our method has richer details, clearer edges, and more importantly does not occupy the exterior space. Two factors contribute to the advantages. The first one is the extra geometric information embedded in the SDF representation, which eliminates the inherent interior/exterior ambiguity of point clouds. The other one is our special de + +
CategoryChamfer-L1IoU
SQs [31]NB [47]MPE(Ours)MPS(Ours)SQs [31]NB [47]MPE(Ours)MPS(Ours)
airplane0.0370.0230.0210.0190.4410.6710.7310.768
bench0.0560.0280.0200.0200.2380.5790.7300.819
bottle0.0470.0330.0260.0170.6860.6650.8860.924
cabinet0.0590.0360.0370.0280.3940.6660.8400.948
can0.0660.0360.0360.0220.7060.5530.9080.950
chair0.0680.0270.0230.0200.3000.6850.7850.871
lamp0.0720.0290.0220.0210.2340.5890.7500.802
speaker0.0640.0410.0370.0330.3460.6560.8580.920
mailbox0.0950.0260.0260.0240.3330.6940.8020.905
rifle0.0380.0200.0190.0190.4460.7320.7440.811
sofa0.0540.0370.0290.0230.4970.7260.8570.940
table0.0700.0240.0240.0220.2470.7450.8180.932
phone0.0400.0210.0230.0210.6810.8720.8910.947
watercraft0.0480.0320.0220.0220.4650.6180.7930.836
mean0.0570.0280.0240.0220.3680.6740.7930.870
+ +Table 1. Quantitative results on Shapenet. MPE and MPS are short for Marching-Primitives with ellipsoids and superquadrics, respectively + +sign of the probabilistic generative model, as discussed in Sec.3.4. It is interesting to observe that our algorithm can extract abstractions from objects not intuitively depictable by the primitives. + +Results on DFAUST: We also evaluate on the DFAUST dataset of human body models. We follow the split settings in [29] to train the learning-based method and evaluate all the methods on the testing set. The SDF is discretized on a grid of size $64^3$ and side length 1. Similar to the ShapeNet experiment, we test our method with both superquadrics and ellipsoids. The results are shown in Fig.4. Our method can accurately capture various postures, while the baselines fail to distinguish different body parts in some cases. The abstraction quality of SQs is much better on this dataset compared with the ShapeNet, since human bodies share a common articulated structure that can be captured by the neural network. We observe that the ellipsoid abstraction also achieves satisfying accuracy. This is because the human body mostly consists of rounded shapes compared with man-made objects in the ShapeNet. + +# 4.2. Performance Study + +This section investigates how the Marching-Primitives algorithm behaves with different voxel grid sizes and distance truncation thresholds. We experiment on the Armadillo from the Stanford 3D scanning repository [10]. First, we discretize the model on the voxel grids of different resolutions. The qualitative abstraction results are visualized in Fig. 5, and the quantitative results are summarized in Fig. 6(a). When the input grid resolution is low, the recovered primitive representation is relatively abstract. With the increase of the grid resolution, our method is able to extract detailed abstraction more faithful to the target shape. This is because the higher the resolution, the more geomet + +![](images/bb148d17d21e964935337f1bef81e44218b01641e92e3636b80e35ab4861fd9b.jpg) +Figure 4. Shape abstraction results on the DFAUST dataset. From left to right: SQs [31], NB [47], MPE, MPS, and the ground truth. + +ric information our method can utilize to guide the primitive marching process. As discussed in Sec.3.1, we use a truncated version of the target and source SDF. Therefore, we also evaluate the performance on different truncation thresholds, starting from 0.1 to 4 times the interval of the voxel grids. The results are shown in Fig.6(b). The abstraction accuracy increases as the threshold decreases at first. The reason lies in that the approximated superquadric + +![](images/31fc9eb13733909ba8a33c97ce518a25ffed8eaac37d4fdf2806906d06401ffc.jpg) +32 + +![](images/68049e4909d9f3938e3bb1ce1a6fa9b0d9b3ac62dedd802df6e91d7e83ed252f.jpg) +$64^{3}$ + +![](images/6b0e817b2c625223415bb6965c9e432bb019a98cfbc98c7ab65b870b327bc8c1.jpg) +96 + +![](images/86bf24b15c2b460d791b9bbe4d34fb764380ca886847fa3806f0a7d9d800a6cb.jpg) +128 + +![](images/7fb6571ce79fabbf090576e7f03d6e440163f61c7d858053c33290ab6cfab060.jpg) +$256^{3}$ + +![](images/9c450411222d5b7b515756deb1026a68ecda48dd7d80042bf51b9dbefdd08a71.jpg) +Ground Truth + +.source) SDF converges to the true value when truncated close to the surface. If we further decrease the threshold, however, we observe an acute decrease in accuracy. This is because an overly small truncation threshold corrupts the original geometric information embedded in the target SDF, as well as the smoothness of the cost function (Eq. (10)). + +![](images/03bb3d3f4bde15145e1b737c91a4aa7dedf067df550f4b9924f93af47a459939.jpg) +Figure 6. Quantitative results on shape abstraction accuracy. (a) Abstraction accuracy versus input grid resolution. (b) Abstraction accuracy versus truncation ratio, evaluated at grid resolution $128^{3}$ . We evaluate on two metrics: Chamfer- $L_{1}$ and IoU. + +![](images/2fcfebed2174d44466761b443816a0d8119cfd6218fafcdcce31ccd74692a713.jpg) +Figure 7. Qualitative result of the scene abstraction. (a) Superquadric abstraction by the Marching-Primitives. (b) Mesh generated by the marching cubes [25] from the same SDF. + +# 4.3. Scene Abstraction + +Other than single objects, we also test the capability of our method in representing complex scenes with geometric primitives. We experiment on a real-world indoor scene called Reading Room [49]. The scene is captured by an Asus Xtion Pro Live camera. The RGB-D scans are fused utilizing [19] (an SDF-based 3D reconstruction algorithm), and further fine-tuned with the method proposed in [49]. However, the SDF representation of the scene is not available publicly. Therefore, we pre-process the scene mesh by removing the floor plane, filling the holes, and transforming it into the SDF representation with grid size $400^3$ . The scene abstraction task is much more difficult compared with the object-level task because various items with great differences in size and shape are present in the very same space. We compare our abstraction result with the mesh extracted by the marching cubes algorithm on the same discrete SDF representation, as shown in Fig.7. Our representation is not only visually satisfying but also contains abundant geometric information the mesh lacks, since it is a continuous SDF approximation to the input discrete SDF. Moreover, our highly sparse representation is only 8.2KB in size, while the discrete SDF occupies 203MB. + +![](images/90d4e1078ea8c708964d17b6f18d58caceeff00d5e21035dd75b9a9734a67e93.jpg) +Figure 5. Visualization of the shape abstractions on different grid resolutions. Abstraction accuracy increases with the input grid resolution. + +![](images/e338b050542461c4d7f525f208a3cac6778e0dc8b4b88925884ff1dd0515d7a2.jpg) + +# 5. Conclusion + +We propose the Marching-Primitives, the first algorithm to extract primitive-based abstractions from the volumetric SDF representation. Our method outperforms the state-of-the-art in terms of accuracy and generalizability on both synthetic and real-world datasets. Our primitive-based representation is sparse, accurate, generalizable, and can be expressed analytically without training, which we believe will facilitate and inspire further explorations in scene understanding, 3D reconstruction, and robot motion planning. However, our algorithm has the limitation of not being able to properly extract object parts thinner than the interval of the voxel grid. This problem can be solved by either increasing the grid resolution or detecting and thickening the invalid parts. Future work includes the parallelization of the marching process and inferring semantic-level interpretations from the current primitive-based geometric features. + +Acknowledgements This research is supported by NUS Startup grants A-0009059-02-00 and A-0009059-03-00, National Research Foundation, Singapore, under its Medium Sized Centre Programme - Centre for Advanced Robotics Technology Innovation (CARTIN), subaward A-0009428-08-00, and AME Programmatic Fund Project MARIO A-0008449-01-00, and JHU internal funds. + +# References + +[1] R. Bajcsy and F. Solina. Three dimensional object representation revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 231-240, 1987. 1 +[2] A. H. Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1(1):11-23, 1981. 2, 3 +[3] I. Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2):115, 1987. 1 +[4] F. Bogo, J. Romero, G. Pons-Moll, and M. J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 6 +[5] D. E. Breen, S. Mauch, and R. T. Whitaker. 3d scan conversion of csg models into distance volumes. In Proceedings of the 1998 IEEE Symposium on Volume Visualization, pages 7-14, 1998. 3 +[6] A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6 +[7] L. Chevalier, J. Jaillet, and A. Baskurt. Segmentation and superquadric modeling of 3D objects. In The 11-th International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision'2003, WSCG 2003, 2003. 1, 2 +[8] G. Chou, I. Chugunov, and F. Heide. Gensdf: Two-stage learning of generalizable signed distance functions. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 6 +[9] C. B. Choy, D. Xu, J. Gwak, K. Chen, and S. Savarese. 3dr2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-644. Springer International Publishing, 2016. 6 +[10] B. Curless and M. Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2, 7 +[11] A. Dai, D. Ritchie, M. Bokeloh, S. Reed, J. Sturm, and M. Nießner. Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2018. 2 +[12] A. Dai, C. Ruizhongtai Qi, and M. Niessner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 2 +[13] A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 5 +[14] B. Deng, K. Genova, S. Yazdani, S. Bouaziz, G. Hinton, and A. Tagliasacchi. Cvxnet: Learnable convex decomposition. + +In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 31-41, 2020. 1 +[15] W. Dong, Q. Wang, X. Wang, and H. Zha. Psdf fusion: Probabilistic signed distance function for on-the-fly 3d data fusion and scene reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 701-717, 2018. 2 +[16] K. Genova, F. Cole, D. Vlasic, A. Sarna, W. Freeman, and T. Funkhouser. Learning shape templates with structured implicit functions. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7153-7163, 2019. 1 +[17] A. D. Gross and T. E. Boult. Error of fit measures for recovering parametric solids. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 690-694, 1988. 1, 2, 3 +[18] Z. Hao, H. Averbuch-Elor, N. Snavely, and S. Belongie. Dualsf: Semantic shape manipulation using a two-level representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7631–7641, 2020. 2, 3 +[19] S. Izadi, D. Kim, O. Hilliges, D. Molyneaux, R. Newcombe, P. Kohli, J. Shotton, S. Hodges, D. Freeman, A. Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 2, 8 +[20] D. K. Katherine and S. S. Elizabeth. Core systems in human cognition. In From Action to Cognition, volume 164 of Progress in Brain Research, pages 257-264. Elsevier, 2007. 1 +[21] Y. Kawana, Y. Mukuta, and T. Harada. Neural star domain as primitive representation. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7875-7886. Curran Associates, Inc., 2020. 1 +[22] A. Leonardis, A. Jaklic, and F. Solina. Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(11):1289-1295, 1997. 1, 2 +[23] Y. Liao, S. Donne, and A. Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2916-2925, 2018. 2 +[24] W. Liu, Y. Wu, S. Ruan, and G. S. Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2676-2685, June 2022. 1, 2, 3, 5 +[25] W. E. Lorensen and H. E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 2, 3, 4, 8 +[26] R. A. Newcombe, D. Fox, and S. M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2 +[27] C. Niu, J. Li, and K. Xu. Im2Struct: Recovering 3D shape structure from a single RGB image. In Proceedings of the + +IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 1 +[28] J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2 +[29] D. Paschalidou, L. V. Gool, and A. Geiger. Learning unsupervised hierarchical part decomposition of 3D objects from a single RGB image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 3, 7 +[30] D. Paschalidou, A. Katharopoulos, A. Geiger, and S. Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3215, June 2021. 1 +[31] D. Paschalidou, A. O. Ulusoy, and A. Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 3, 6, 7 +[32] A. Pentland. Perceptual organization and the representation of natural form. Artificial Intelligence, 28(3):293-331, 1986. 1, 2 +[33] A. H. Quispe, B. Milville, M. A. Gutiérrez, C. Erdogan, M. Stilman, H. Christensen, and H. B. Amor. Exploiting symmetries and extrusions for grasping household objects. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3702-3708, 2015. 1 +[34] Y. Rao, Y. Nie, and A. Dai. Patchcomplete: Learning multi-resolution patch priors for 3d shape completion on unseen categories. Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[35] S. Ruan, K. L. Poblete, H. Wu, Q. Ma, and G. S. Chirikjian. Efficient path planning in narrow passages for robots with ellipsoidal components. IEEE Transactions on Robotics (TRO), pages 1-18, 2022. 1 +[36] S. Ruan, X. Wang, and G. S. Chirikjian. Collision detection for unions of convex bodies with smooth boundaries using closed-form contact space parameterization. IEEE Robotics and Automation Letters (RAL), 7(4):9485-9492, 2022. 1 +[37] G. Sharma, B. Dash, A. RoyChowdhury, M. Gadelha, M. Loizou, L. Cao, R. Wang, E. Learned-Miller, S. Maji, and E. Kalogerakis. Prift: Learning to fit primitives improves few shot point cloud segmentation. In Computer Graphics Forum, volume 41, pages 39-50. Wiley Online Library, 2022. 3 +[38] D. Smirnov, M. Fisher, V. G. Kim, R. Zhang, and J. Solomon. Deep parametric shape predictions using distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 561-570, 2020. 1 +[39] F. Solina and R. Bajcsy. Recovery of parametric models from range images: the case for superquadrics with global deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(2):131-147, 1990. 1, 2 + +[40] C. Sun, Q. Zou, X. Tong, and Y. Liu. Learning adaptive hierarchical cuboid abstractions of 3d shape collections. ACM Transactions on Graphics (TOG), 38(6):1-13, 2019. 3, 6 +[41] S. Tulsiani, H. Su, L. J. Guibas, A. A. Efros, and J. Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 3 +[42] N. Vaskevicius and A. Birk. Revisiting superquadric fitting: A numerically stable formulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):220-233, 2019. 3 +[43] S. Vasu, N. Talabot, A. Lukoianov, P. Baque, J. Donier, and P. Fua. Hybridsdf: Combining deep implicit shapes and geometric primitives for 3d shape representation and manipulation. In International Conference on 3D Vision (3DV), 2022. 2 +[44] G. Vezzani, U. Pattacini, and L. Natale. A grasping approach based on superquadric models. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1579-1586, 2017. 1 +[45] G. Vezzani, U. Pattacini, G. Pasquale, and L. Natale. Improving superquadric modeling and grasping with prior on object shapes. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 6875-6882, 2018. 1 +[46] S. Weder, J. L. Schonberger, M. Pollefeys, and M. R. Oswald. Routedfusion: Learning real-time depth map fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2 +[47] Y. Wu, W. Liu, S. Ruan, and G. S. Chirikjian. Primitive-based shape abstraction via nonparametric bayesian inference. In Proceedings of the European Conference on Computer Vision (ECCV), pages 479-495. Springer Nature Switzerland, 2022. 1, 2, 3, 6, 7 +[48] K. Yang and X. Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 1, 2, 3, 6 +[49] Q. Zhou, S. Miller, and V. Koltun. Elastic fragments for dense scene reconstruction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 473-480, 2013. 2, 8 +[50] C. Zou, E. Yumer, J. Yang, D. Ceylan, and D. Hoiem. 3D-PRNN: Generating shape primitives with recurrent neural networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1, 6 \ No newline at end of file diff --git a/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/images.zip b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..42b2c6a78e963f29ba0f8fa696e9813ead296a85 --- /dev/null +++ b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66d9c281a93ce886c8ca703ee5eed80f85a4e6c301b47efdbfa638fc523c022c +size 465692 diff --git a/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/layout.json b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fee91cc710de7ee0fa197faf0021495ccf0f6ee1 --- /dev/null +++ b/2023/Marching-Primitives_ Shape Abstraction From Signed Distance Function/layout.json @@ -0,0 +1,8390 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 520, + 121 + ], + "type": "text", + "content": "Marching-Primitives: Shape Abstraction from Signed Distance Function" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "spans": [ + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "text", + "content": "Weixiao Liu" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "text", + "content": " Yuwei Wu" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "text", + "content": " Sipu Ruan" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "text", + "content": " Gregory S. Chirikjian" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 142, + 470, + 185 + ], + "type": "text", + "content": "Johns Hopkins University \n{mpewxl, yw.wu, ruansp, mpegre}@nus.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "content": "Representing complex objects with basic geometric primitives has long been a topic in computer vision. Primitive-based representations have the merits of compactness and computational efficiency in higher-level tasks such as physics simulation, collision checking, and robotic manipulation. Unlike previous works which extract polygonal meshes from a signed distance function (SDF), in this paper, we present a novel method, named Marching-Primitives, to obtain a primitive-based abstraction directly from an SDF. Our method grows geometric primitives (such as superquadrics) iteratively by analyzing the connectivity of voxels while marching at different levels of signed distance. For each valid connected volume of interest, we march on the scope of voxels from which a primitive is able to be extracted in a probabilistic sense and simultaneously solve for the parameters of the primitive to capture the underlying local geometry. We evaluate the performance of our method on both synthetic and real-world datasets. The results show that the proposed method outperforms the state-of-the-art in terms of accuracy, and is directly generalizable among different categories and scales. The code is open-sourced at https://github.com/ChirikjianLab/Marching-Primitives.git." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "type": "text", + "content": "Recent years have witnessed great progress in the areas of 3D shape representation and environmental perception. Low-level representations such as surface meshes, point clouds, and occupancy grids are widely used as inputs to high-level computer vision algorithms and artificial intelligence tasks. They have the advantage of being able to represent and visualize objects with high accuracy and rich local geometric features. However, the low-level representations are ineffective in delivering a general and intuitive sense of structural geometry as well as part-level scene understanding. Studies [3,20] show that human vision, unlike" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 214, + 537, + 281 + ], + "blocks": [ + { + "bbox": [ + 315, + 214, + 537, + 281 + ], + "lines": [ + { + "bbox": [ + 315, + 214, + 537, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 214, + 537, + 281 + ], + "type": "image", + "image_path": "57f1bb3b41f9000f06445b665e2541aec27b71d587e27ff0fb981e413e3f9432.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 292, + 547, + 370 + ], + "lines": [ + { + "bbox": [ + 304, + 292, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 547, + 370 + ], + "type": "text", + "content": "Figure 1. Primitive-base representation versus mesh. For each pair of objects, the left one is the superquadratic abstraction obtained by our algorithm, and the right one is the original mesh. The mesh of the chair is 6MB in size, while our representation only needs 4KB. An SDF representation discretized on a " + }, + { + "bbox": [ + 304, + 292, + 547, + 370 + ], + "type": "inline_equation", + "content": "128^{3}" + }, + { + "bbox": [ + 304, + 292, + 547, + 370 + ], + "type": "text", + "content": " voxel grid occupies 19MB. Our abstraction is equivalent to an implicit continuous SDF, which is an approximation to the discrete SDF." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "type": "text", + "content": "computer vision, tends to perceive and understand scenes as combinations of simple primitive shapes. Human beings perform well and robustly in complex tasks, providing a basic geometric description of the scene is available [32]. Therefore, researchers turn to exploring the possibility of interpreting complex objects and scenes with basic geometric primitives. Taking advantage of the primitive-based representation, many higher-level tasks, such as segmentation [14, 16, 21, 30], scene understanding [29, 31, 41, 47], grasping [33, 44, 45] and motion planning [35, 36], are able to be solved efficiently." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "However, it still remains challenging to extract primitive-based abstractions from low-level representations. Starting from the 1990s, Solina et al. [1, 17, 39] aim to extract a single superquadric representation from a simple object by minimizing the least-square error between the primitive and the measured points. Later in [7, 22], their method is extended to represent more complex objects with multiple primitives. More recently, the authors of [24, 47] reformulate the task as a probabilistic inference problem with enhanced accuracy and robustness to noise and outliers. At the same time, with the surge of data-driven techniques, researchers attempt to train neural networks to infer cuboids [27,38,41,48,50] and superquadrics [29,31] representations in an end-to-end fashion. However, both the computational and learning-based approaches have their own limitations. The computational methods are vulnerable to the inherent ambiguity of the point-to-surface relationship. For exam" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8771" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "ple, the algorithms tend to fill empty spaces of a non-convex object with primitives by mistake, due to the inside/outside ambiguity of a surface depicted by a set of points [24, 48]. The main drawback of the learning approaches lies in the lack of generalizability beyond the object category on which the model is trained [24, 31, 47, 48]. Also, the shape abstraction accuracy is inferior to the computational methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 159, + 288, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 288, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 288, + 447 + ], + "type": "text", + "content": "The signed distance function (SDF) has been a successful 3D volumetric representation in varieties of computer vision and graphics tasks. It is the basic framework for many classic 3D reconstruction algorithms such as TSDF volume reconstruction [10, 15], KinectFusion [19], and Dynamic-Fusion [26]. Recently, the SDF representation is adapted to the deep learning frameworks, and exhibits boosted potentials in shape encoding [8, 18, 28, 43], surface reconstruction [23, 46], and shape completion [11, 12, 34]. Usually, triangular mesh surfaces are extracted from the SDF representation with the marching cubes algorithm [25]. Point cloud and occupancy grid representations are also obtained by keeping the vertices of the meshes and the sign of each voxel point, respectively. The SDF is among the most informative 3D representations since it encodes not only the surface geometry but also the distance and side of a point relative to the shape. Meanwhile, it is easily achievable via range images from 3D sensors [10], or learnable from other input modalities [8, 18, 28]. Since we are able to extract meshes from an SDF, it is natural to think about the possibility of extracting primitives as well. Furthermore, the primitive-based abstraction is a continuous interpretation of the complete geometric information encoded in the original discrete SDF, but requires much less storage size (Fig.1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": "Motivated by the aforementioned facts and the bottleneck of the current shape abstraction algorithms, we proposed a general shape abstraction method by reasoning directly on the informative SDF representation. The goal of our method is to find a combination of geometric primitives whose underlying SDF values match the target values evaluated on the evenly spaced discrete grid points (Sec. 3.1 and Sec. 3.2). To solve this problem, we propose a two-step iterative algorithm called the Marching-Primitives. Our algorithm 'marches' on two domains: the signed distance domain and the voxelized space domain, alternately. Firstly, the connectivity of volumes are analyzed by generating isosurfaces on a sequence of decreasing levels of negative signed distances (Sec.3.3). By doing so, volumes of interest (VOIs) where primitives are likely to be encoded can be identified sequentially. In the second step, for each of the VOIs, our algorithm marches on the neighbouring voxels to infer their probabilistic correspondences to the primitive and simultaneously optimizes the shape and pose of the primitive (Sec.3.4). After the primitive representation of a VOI is achieved, the fitted volumes are deactivated from the voxel grid. Our algorithm continues marching on the signed" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "distance domain until it approaches zero, i.e., all the interior volumes of the SDF have been captured by the recovered primitives. We compare our algorithm with the state-of-the-art of both the computational and learning-based approaches on the ShapeNet object dataset [6] and D-FAUST human shape dataset [4] (Sec. 4.1). We also study the performance of our algorithm on different conditions (Sec. 4.2). Finally, we demonstrate the scene abstraction result of the Stanford Reading Room [49], which contains several pieces of furniture of various categories (Sec. 4.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 222, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 545, + 593 + ], + "type": "text", + "content": "SDF Representation: The SDF can be stored in two different ways: discrete or continuous. A majority of computer vision and graphics algorithms are built on the SDF discretized on a 3D grid of voxel points. The signed distances are stored on each of the corresponding voxel points. The authors of [10, 19, 26] pioneer in fusing several noisy range images into a single discrete SDF. Their work is widely applied in 3D reconstruction and plays an important role in robotics tasks such as simultaneous localization and mapping. The discrete SDF is also a promising input/output representation for 3D deep learning [11, 12, 34]. Recently, it becomes popular to encode shapes as a continuous SDF with neural networks [28]. Vasu et al. [43] further improve the shape encoding quality by enforcing local regularities with geometric primitives. In [8], the authors adopt a two-stage meta-learning approach to further extend the generalization capabilities of neural SDF. With the deep neural network, it becomes possible to infer SDF representations from partially observed 3D inputs or even images. Both the discrete and the continuous SDF are implicit representations of geometric surfaces. To extract the explicit surface from the SDF, continuous SDFs need to be discretized on a voxel grid first and then conduct the marching cubes [25]. This method allows high-quality rendering of the objects, however, surface meshes are non-sparse and contain no structural level information. Our method provides an alternative approach to describe the underlying object in the SDF. Instead of meshes, we directly extract a collection of sparsely parameterized primitives from the SDF. Other than that, our primitive-based representation itself is also a concise yet continuous SDF approximation to the original discrete SDF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Computational Shape Abstraction: The most well-studied primitive for computational shape abstraction is the superquadric, due to its extensive shape vocabulary including cuboids, ellipsoids, cylinders, octohedra, and many shapes in between (e.g., cuboids with rounded edges). It is first proposed as a versatile modeling element for complex objects in computer graphics [2, 32]. Later, Solina et al. propose a method to conduct abstraction of simple objects from range images with a single superquadric [17, 39]. Leonardis et al. [22] and Chevalier et al. [7] further extend" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8772" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "type": "text", + "content": "the previous work to recover complex objects with multiple superquadrics with a Split-and-Merge strategy. A numerical instability problem is addressed and revisited in [42]. The authors introduce an auxiliary function in the unstable region and receive a better abstraction accuracy. More recently, Liu et al. [24] formulate the problem in a probabilistic fashion and propose a geometric strategy to avoid local optimum, bringing a significant improvement in robustness to outlier and fitting accuracy. Wu et al. [47] extend and recast the work as a nonparametric Bayesian inference problem so as to improve the applicability on complex shapes. To the best of the authors' knowledge, the existing computational methods are all based on range images or point clouds, which suffer from geometric ambiguities [48]. In contrast, our method takes advantage of the abundant geometric information encoded in the SDF and is easily compatible with other computer vision algorithms based on the SDF representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 624 + ], + "type": "text", + "content": "Learning-based Shape Abstraction: The learning-based method is first seen in [41]. Tulsiani et al. propose a 3D convolutional neural network (CNN) to learn shape abstractions with cuboids from the occupancy grid. Sun et al. [40] design an adaptive hierarchical cuboid representation and introduce an unsupervised approach to learn to extract the parameters for the representation. Yang et al. [48] train a variational auto-encoder network to transform point clouds into parametric cuboids. Other than cuboids, researchers also seek to extract spheres and ellipsoids representations from objects. Hao et al. [18] combine the neural SDF with the spherical representation by sharing a same latent layer. In [37], ellipsoids or cuboids are extracted to help segment the input point cloud. However, a single type of primitive has very limited expressiveness. Therefore, Paschalidou et al. [29, 31] turn to training neural networks to conduct abstractions with the superquadrics as the atomic elements. Learning-based approaches are versatile in dealing with different input sources. They are able to make shape abstractions from point clouds, voxel grids, or even RGB images which are so ill-conditioned that the computational approaches have little chance of working. However, learning-based approaches rely heavily on the training dataset and thus are less generalizable to unseen categories. Instead, our approach reasons about the primitive abstraction from a case-by-case geometric perspective, which provides an inherent advantage in generalizability and accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 637, + 104, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 104, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 104, + 649 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 658, + 126, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 126, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 126, + 670 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "The discrete SDF is a volumetric surface representation built on a voxel grid " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{\\mathbf{x}_i\\in \\mathbb{R}^3,i = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": ". A scalar " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " is assigned to each grid point, which indicates" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "the signed distance of " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " to the nearest surface. The point " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " lies inside the surface if " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_i) < 0" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and outside otherwise. Typically, the surface mesh of an object is extracted from the SDF by the marching cubes algorithm [25]. In this paper, we call the input SDF " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_i)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " the target SDF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 168 + ], + "type": "text", + "content": "For the primitive representation, we select the superquadrics [2], a family of geometric primitives defined by the implicit equation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 173, + 545, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 173, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 311, + 173, + 545, + 209 + ], + "type": "interline_equation", + "content": "f (\\mathbf {x}) = \\left(\\left(\\frac {x}{a _ {x}}\\right) ^ {\\frac {2}{\\varepsilon_ {2}}} + \\left(\\frac {y}{a _ {y}}\\right) ^ {\\frac {2}{\\varepsilon_ {2}}}\\right) ^ {\\frac {\\varepsilon_ {2}}{\\varepsilon_ {1}}} + \\left(\\frac {z}{a _ {z}}\\right) ^ {\\frac {2}{\\varepsilon_ {1}}} = 1 \\tag {1}", + "image_path": "35ec58776e45b170941dfdd2e6b284aa7f6ea963bd8f06bda768224ab7c1d8b2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": "The superquadric family is only encoded by 5 parameters (shape parameters " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\epsilon_{1},\\epsilon_{2}\\in [0,2]\\subset \\mathbb{R}" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": ", and scale parameters " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "a_{x},a_{y},a_{z}\\in \\mathbb{R}_{>0}" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": "), but has an extensive shape vocabulary. Note that the shape parameters can exceed 2, resulting in nonconvex shapes. However, practically in most studies [24,31] and also in our paper, we limit them within the convex region as defined above. The points " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = [x,y,z]\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": " satisfying Eq. (1) form the surface of the superquadric. In this paper, we also include the Euclidean transformation " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "g\\in SE(3)" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": ", i.e. 3 Euler angles for rotation " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{R}\\in SO(3)" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": " and 3-dimensional translation " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{t}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": ", to parameterize a superquadric with a general pose. In total, we denote a superquadric with a vector " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": " of 11 elements. According to Eq. (1), we can approximate the signed distance of a grid point " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": " to a general posed superquadric " + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 304, + 214, + 546, + 395 + ], + "type": "text", + "content": " explicitly by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 331, + 399, + 545, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 399, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 331, + 399, + 545, + 420 + ], + "type": "interline_equation", + "content": "d _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right) = \\left(1 - f ^ {- \\frac {\\epsilon_ {1}}{2}} \\left(g ^ {- 1} \\circ \\mathbf {x} _ {i}\\right)\\right) \\| g ^ {- 1} \\circ \\mathbf {x} _ {i} \\| _ {2} \\tag {2}", + "image_path": "c3585fceb000f2bd4efb62f82c908e639e8fdacba19b54e3b62218f6e5484e39.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": "We are not able to use the exact SDF of superquadrics, because it has no analytical solution and is only achievable by numerical optimization [5]. Eq.(2) is the radial distance [17, 24] of " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": " to the superquadric surface, which converges to the exact signed distance as " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "d_{\\theta}(\\mathbf{x}_i)\\to 0" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": ". Therefore, we truncate both the input target SDF " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": " and the primitive SDF " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "d_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": " within the vicinity of zero to ensure the approximation accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 526, + 430, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 430, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 430, + 538 + ], + "type": "text", + "content": "3.2. Problem Formulation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": "To obtain a primitive-based abstraction from the target SDF, we seek a combination of primitives " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\Theta \\doteq \\{\\theta_{k}, k = 1,2,\\dots,K\\}" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " whose underlying signed distances measured on the voxel points (we call it the source SDF in contrast to the target SDF) match the target SDF, that is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 609, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 609, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 332, + 609, + 545, + 635 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Theta} = \\underset {\\boldsymbol {\\Theta}} {\\arg \\min } \\sum_ {\\mathbf {x} _ {i} \\in \\mathbf {V}} \\underset {\\boldsymbol {\\theta} _ {k} \\in \\boldsymbol {\\Theta}} {\\min } \\left\\| d _ {\\boldsymbol {\\theta} _ {k}} \\left(\\mathbf {x} _ {i}\\right) - d \\left(\\mathbf {x} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {3}", + "image_path": "2f7b3905a858953c666dac4188b33febffce38a1f7ce298c9a8318c4151d4bf7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Our problem formulation is simple and intuitive, however, it is intractable to solve directly. First, the number of primitives " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " needed is unknown. Also, the correspondences between the voxel points " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " and the primitives " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_k" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " are unknown a priori. Therefore, it makes Eq.(3) a chicken-and-egg problem: on the one hand, if we know the set of voxel" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8773" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "points contributing to a same primitive, we are able to solve for the parameters of the primitive by optimization; on the other hand, if we know the configurations of the primitives, we are able to tell the belongings of each voxel. Other than those factors, the computation complexity itself is prohibitively high, because we need to evaluate hundreds of thousands or even millions of voxel points. This makes it infeasible to operate directly on the complete dense voxel grid. To tackle these difficulties, we propose an iterative algorithm, the Marching-Primitives, which simultaneously solves the correspondences and primitive parameters efficiently. It is worth noting that our method is generalizable beyond the superquadrics. Any volumetric primitives with easily accessible SDF representations can be adapted to our framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 259, + 221, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 221, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 221, + 272 + ], + "type": "text", + "content": "3.3. Iterative Connectivity Marching" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 277, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 361 + ], + "type": "text", + "content": "Instead of setting a predefined number of primitives, our method starts from an empty set of primitives and grows as needed. This is realized by analyzing the connectivity (in this paper, we use the 26-connectivity) of the voxels at different levels of signed distance. Given the target SDF, our algorithm checks the connectivity of voxels whose signed distance is less than a sequence of thresholds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 370, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 370, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 52, + 370, + 287, + 389 + ], + "type": "interline_equation", + "content": "T ^ {c} \\doteq \\left\\{t _ {1} ^ {c}, t _ {2} ^ {c}, \\dots \\right\\}, \\quad t _ {1} ^ {c} = \\min _ {\\mathbf {x} _ {i} \\in \\mathbf {V}} d (\\mathbf {x} _ {i}), \\quad t _ {m + 1} ^ {c} = \\alpha t _ {m} ^ {c} (4)", + "image_path": "f2401fb694fc79dc67a294a335966509ead48c66a9b0fa1f4129ddf24ba5a18d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,1) \\subset \\mathbb{R}" + }, + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "type": "text", + "content": " is the common ratio, resulting in a geometric sequence exponentially decaying to zero. Each threshold " + }, + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "type": "inline_equation", + "content": "t_m^c" + }, + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "type": "text", + "content": " defines a set of disjoint isosurfaces " + }, + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "type": "inline_equation", + "content": "S_m" + }, + { + "bbox": [ + 46, + 398, + 287, + 457 + ], + "type": "text", + "content": ", encompassing the connected voxels whose signed distances are less than the threshold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 468, + 287, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 468, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 287, + 481 + ], + "type": "interline_equation", + "content": "S _ {m} = \\left\\{S _ {k}, k = 1, 2, \\dots , | S _ {m} | \\right\\} \\tag {5}", + "image_path": "9155d42f565095ca5b34fcff781c28ba6a9310038c78f997f851bedbe44a1775.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 490, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 287, + 514 + ], + "type": "inline_equation", + "content": "|S_{m}|" + }, + { + "bbox": [ + 47, + 490, + 287, + 514 + ], + "type": "text", + "content": " denotes the number of the disjoint isosurfaces. We construct a subset " + }, + { + "bbox": [ + 47, + 490, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\bar{S}_m" + }, + { + "bbox": [ + 47, + 490, + 287, + 514 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 490, + 287, + 514 + ], + "type": "inline_equation", + "content": "S_{m}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 523, + 287, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 523, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 91, + 523, + 287, + 537 + ], + "type": "interline_equation", + "content": "\\bar {S} _ {m} = \\left\\{S _ {k} \\in S _ {m}, | S _ {k} | \\geq N _ {c} \\right\\} \\subseteq S _ {m} \\tag {6}", + "image_path": "ac5730c2c7417fb57a5c39bc48ed95286400baf94a4231d445df1cdd3973bdf1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "|S_{k}|" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " is the number of the connected voxels within the isosurface " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\bar{S}_{m}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " is the set of the isosurfaces encompassing no less than " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " connected voxels. The connectivity marching starts from the innermost threshold " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "t_1^c" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": ", where the resulting " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\bar{S}_1" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " might be empty. The marching continues on the sequence (increasing the threshold) until " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\bar{S}_{m}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " is nonempty. Each of the connected volumes in " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\bar{S}_{m}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " (we call it a VOI) is an ideal starting point for growing a primitive, because: (1) those volumes are among the most interior of the target SDF, corresponding to most prominent part of the geometry; (2) disconnected and weakly connected volumes can be separated apart; (3) the SDF of a primitive can also be interpreted as layers of isosurfaces, and thus sharing similar geometric structure; and furthermore (4) the size of the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 408, + 142 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 408, + 142 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 408, + 142 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 408, + 142 + ], + "type": "image", + "image_path": "541b8c21aba2400bafdf17b5742b2ff63e2ba1278d04bf634b9a29e005fb5c9f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 144, + 335, + 153 + ], + "lines": [ + { + "bbox": [ + 323, + 144, + 335, + 153 + ], + "spans": [ + { + "bbox": [ + 323, + 144, + 335, + 153 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 155, + 402, + 250 + ], + "blocks": [ + { + "bbox": [ + 329, + 155, + 402, + 250 + ], + "lines": [ + { + "bbox": [ + 329, + 155, + 402, + 250 + ], + "spans": [ + { + "bbox": [ + 329, + 155, + 402, + 250 + ], + "type": "image", + "image_path": "733e9343c28217bbf6105e987c28d06a8731b785debb53cfbfeecbd6fe5d6222.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 263, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 363 + ], + "type": "text", + "content": "Figure 2. Visualizations of concepts. (a) A 2D illustration of two marched isosurfaces separating the shape into two VOIs. (b) A simple example of the Marching-Primitives algorithm. The first figure is the input target SDF (voxels with signed distances larger than the truncation threshold are left blank for visualization). Figures b2-b4 illustrate the primitives marched on the sequentially detected VOIs. (c) Auto-degeneration and probabilistic marching. The green circles on the left column indicate different initialization configurations. The ones on the right are the marched results." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 76, + 471, + 110 + ], + "blocks": [ + { + "bbox": [ + 410, + 72, + 422, + 83 + ], + "lines": [ + { + "bbox": [ + 410, + 72, + 422, + 83 + ], + "spans": [ + { + "bbox": [ + 410, + 72, + 422, + 83 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 76, + 471, + 110 + ], + "lines": [ + { + "bbox": [ + 426, + 76, + 471, + 110 + ], + "spans": [ + { + "bbox": [ + 426, + 76, + 471, + 110 + ], + "type": "image", + "image_path": "4c7ce3733d47b574f9a15a7fe36926980be4e2f124f678c40bc16f0857f5bff2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 484, + 76, + 524, + 115 + ], + "blocks": [ + { + "bbox": [ + 484, + 76, + 524, + 115 + ], + "lines": [ + { + "bbox": [ + 484, + 76, + 524, + 115 + ], + "spans": [ + { + "bbox": [ + 484, + 76, + 524, + 115 + ], + "type": "image", + "image_path": "37900fffbd83686779db3bb95e8873d09b00a75bbf6898d7576296d0ba4a18ac.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 121, + 469, + 156 + ], + "blocks": [ + { + "bbox": [ + 427, + 121, + 469, + 156 + ], + "lines": [ + { + "bbox": [ + 427, + 121, + 469, + 156 + ], + "spans": [ + { + "bbox": [ + 427, + 121, + 469, + 156 + ], + "type": "image", + "image_path": "464e7a7119bacef703abb393dba94ea8309839f3d4a589acec106c94b0ac73c8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 484, + 120, + 525, + 155 + ], + "blocks": [ + { + "bbox": [ + 484, + 120, + 525, + 155 + ], + "lines": [ + { + "bbox": [ + 484, + 120, + 525, + 155 + ], + "spans": [ + { + "bbox": [ + 484, + 120, + 525, + 155 + ], + "type": "image", + "image_path": "6aba9cb1f68dd8028a8d0229accefa505a5f5cc1ffd126a08b33141cafad435e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 161, + 471, + 198 + ], + "blocks": [ + { + "bbox": [ + 427, + 161, + 471, + 198 + ], + "lines": [ + { + "bbox": [ + 427, + 161, + 471, + 198 + ], + "spans": [ + { + "bbox": [ + 427, + 161, + 471, + 198 + ], + "type": "image", + "image_path": "c54279062a4659964a7ceb1352d5743ac4f4975f0335b5c3991eb64971443cb9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 484, + 161, + 525, + 196 + ], + "blocks": [ + { + "bbox": [ + 484, + 161, + 525, + 196 + ], + "lines": [ + { + "bbox": [ + 484, + 161, + 525, + 196 + ], + "spans": [ + { + "bbox": [ + 484, + 161, + 525, + 196 + ], + "type": "image", + "image_path": "94bcccd0ffd964e03398687f626dc41983e8413c6aecac1ad454388e73db1503.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 427, + 204, + 471, + 241 + ], + "blocks": [ + { + "bbox": [ + 427, + 204, + 471, + 241 + ], + "lines": [ + { + "bbox": [ + 427, + 204, + 471, + 241 + ], + "spans": [ + { + "bbox": [ + 427, + 204, + 471, + 241 + ], + "type": "image", + "image_path": "fd94f74df9d3d67e295fd3f292b5bd802585e0948eb205288ca8ec9e0191ae45.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 243, + 465, + 250 + ], + "lines": [ + { + "bbox": [ + 427, + 243, + 465, + 250 + ], + "spans": [ + { + "bbox": [ + 427, + 243, + 465, + 250 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 484, + 202, + 527, + 239 + ], + "blocks": [ + { + "bbox": [ + 484, + 202, + 527, + 239 + ], + "lines": [ + { + "bbox": [ + 484, + 202, + 527, + 239 + ], + "spans": [ + { + "bbox": [ + 484, + 202, + 527, + 239 + ], + "type": "image", + "image_path": "c0b9610b7201427b289e58d7075253273fa7e9e19d2fa816249ac6e4da411d31.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 243, + 525, + 251 + ], + "lines": [ + { + "bbox": [ + 486, + 243, + 525, + 251 + ], + "spans": [ + { + "bbox": [ + 486, + 243, + 525, + 251 + ], + "type": "text", + "content": "Convergence" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 369, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 489 + ], + "type": "text", + "content": "volumes is large enough to provide sufficient geometric information for the primitive recovery. We illustrate the idea of isosurfaces in Fig.2(a). For each VOI, we initialize a primitive as an ellipsoid, with scales proportional to the size of its smallest bounding-box. The details for the primitive initialization are demonstrated in the Supplementary Material. The procedure of how a primitive marches from the initial guess to capture the local geometry is detailed in the next Sec. 3.4. After obtaining the primitive representations for all the VOIs in " + }, + { + "bbox": [ + 304, + 369, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\bar{S}_m" + }, + { + "bbox": [ + 304, + 369, + 545, + 489 + ], + "type": "text", + "content": ", we subtract the voxels" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 353, + 498, + 545, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 498, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 353, + 498, + 545, + 512 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {x} _ {i} \\in \\mathbf {V}, d \\left(\\mathbf {x} _ {i}\\right) \\leq 0 \\wedge d _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right) \\leq 0 \\right\\} \\tag {7}", + "image_path": "4976a303b672e3db4d5e3cde7fccef3b6339a4cec83e28eef83e1ded82b8e17c.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": "from the set of voxel grid " + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": ". In other words, we deactivate the voxels which are interior of the target SDF and well fitted by a recovered primitive, while the updated SDF preserves the exterior and unfitted interior volumes. The connectivity marching repeats with the updated target SDF, until the threshold marches higher than a preset limit close to zero, indicating that no prominent interior volume is left unrepresented." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 623, + 484, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 484, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 484, + 636 + ], + "type": "text", + "content": "3.4. Probabilistic Primitive Marching" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Firstly, we explain what we mean by primitive marching. In the marching cubes algorithm [25], marching means the progressive extraction of polygonal meshes from the neighboring 8 voxel points in the discrete SDF. In each step, the algorithm focuses within the single cube formulated by the 8 voxel points, calculates triangle vertices using linear in" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8774" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "content": "terpolation, and then 'marches' to the next one. On the contrary, our method marches on the scope of voxels. That is, we gradually figure out a proper collection of voxel points from which a geometric primitive is extractable. Inspired by [24], we formulate this process as a maximum likelihood estimation problem with latent variables." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": "For each isosurface " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "S_{k}\\in \\bar{S}_{m}" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": " obtained during the connectivity marching (Eq.(6)), it is assumed that the 'true' underlying primitive is encoded by " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": ", and that " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "z_{ik}" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": " is the correspondence between the primitive " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": " and a voxel point " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbf{V}" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "z_{ik}" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": " is a binary variable that equals 1 when the target signed distance evaluated at the " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": "th voxel point results from the " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": "th primitive, and 0 otherwise. Then the target signed distance " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": " can be regarded as an observation generated from the probabilistic distribution" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 255, + 287, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 255, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 255, + 287, + 289 + ], + "type": "interline_equation", + "content": "p \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k}, z _ {i k}\\right) = \\left(\\frac {\\mathbb {1} _ {d _ {i} \\in [ - t , 0)}}{t}\\right) ^ {1 - z _ {i k}} \\mathcal {N} \\left(d _ {i} \\mid d _ {\\boldsymbol {\\theta} _ {k}} (\\mathbf {x} _ {i}), \\sigma^ {2}\\right) ^ {z _ {i k}} \\tag {8}", + "image_path": "863651f8aae5896b01a3f82c0a91565ed7a31adbfabf13fea414fc7068896803.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": " is short for " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": " is the truncation value of the SDF; " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{d_i\\in [-t,0)}" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": " is an indicator function which equals one when " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "d_{i}\\in [-t,0)" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": " and zero otherwise; " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{N}\\big(\\cdot |d_{\\pmb{\\theta}_k}(\\mathbf{x}_i),\\sigma^2\\big)" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": " is a Gaussian distribution with mean " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "d_{\\pmb{\\theta}_k}(\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": ". Consequently, our goal is to find out the optimal " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_k" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": " and the voxel-primitive correspondences " + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "inline_equation", + "content": "z_{ik}" + }, + { + "bbox": [ + 46, + 289, + 287, + 373 + ], + "type": "text", + "content": " simultaneously, which maximize the likelihood of the target SDF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "content": "Correspondence Marching: We assume that " + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "inline_equation", + "content": "z_{ik}" + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "content": " follows a Bernoulli prior distribution " + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "inline_equation", + "content": "B(p_0)" + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "content": ", independent of " + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\theta_k" + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "content": ". Then, given the target SDF " + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "content": " and the current primitive estimation " + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\theta_k" + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "content": ", we are able to infer the posterior correspondence " + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "inline_equation", + "content": "z_{ik}" + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "content": " by the Bayes' rule" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 437, + 287, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 437, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 63, + 437, + 287, + 466 + ], + "type": "interline_equation", + "content": "p \\left(z _ {i k} \\mid \\boldsymbol {\\theta} _ {k}, d _ {i}\\right) = \\frac {p \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k} , z _ {i k}\\right) p \\left(z _ {i k}\\right)}{\\sum_ {z _ {i k} \\in \\{0 , 1 \\}} p \\left(d _ {i} \\mid \\boldsymbol {\\theta} _ {k} , z _ {i k}\\right) p \\left(z _ {i k}\\right)} \\tag {9}", + "image_path": "ee16e170d7952f6bde614c7badd6fe7473cca31b072c6cae259e499121af4843.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": "By analyzing the posterior correspondence, we can observe that our design of the probabilistic model (Eq.(8)) possesses two desirable properties: (1) voxels labeled as inside " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "(d_{i} < 0)" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": " may not contribute to the current estimation of the primitive " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{k}" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "p(z_{ik} = 1|\\pmb{\\theta}_k,d_i < 0)\\in (0,1)\\subset \\mathbb{R}" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": "; (2) all non-negative valued voxels always contribute to any primitives, i.e., " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "p(z_{ik} = 1|\\pmb{\\theta}_k,d_i\\geq 0) = 1" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": ". In other words, the primitive tries to absorb the interior voxels sharing similar local geometry, while avoids occupying any voxels labeled as the exterior. Meanwhile, the probabilistic correspondence establishes a soft relationship between the primitive and the interior voxels, allowing part of the primitive to occupy some interior volumes without complying with the target signed distance values. In this way, the algorithm is able to achieve an overall better fitting quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 649, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 649, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 649, + 287, + 685 + ], + "type": "text", + "content": "Primitive Update: Given the current estimation of the correspondences between the primitive and voxels, the parameter of the primitive is updated by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 689, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 82, + 689, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {k} = \\underset {\\boldsymbol {\\theta} _ {k}} {\\arg \\min } \\sum_ {\\mathbf {x} _ {i} \\in \\mathbf {V} _ {a}} P _ {i k} \\| d _ {\\boldsymbol {\\theta} _ {k}} (\\mathbf {x} _ {\\mathrm {i}}) - d _ {i} \\| _ {2} ^ {2} \\tag {10}", + "image_path": "f17e9d21db7834e8362c98b51529bf4fb3ec1dff05810a7a465120214bcdf04a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 73, + 334, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 334, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 334, + 83 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 92, + 545, + 123 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 92, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 328, + 92, + 545, + 123 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P _ {i k} = p \\left(z _ {i k} = 1 \\mid \\boldsymbol {\\theta} _ {k} ^ {p r e v}, d _ {i}\\right) \\tag {11} \\\\ \\mathbf {V} _ {a} = \\left\\{\\mathbf {x} _ {\\mathrm {i}} \\in \\mathbf {V} | d _ {\\boldsymbol {\\theta} _ {k} ^ {p r e v}} (\\mathbf {x} _ {i}) \\in [ - a, a ] \\subset \\mathbb {R} \\right\\} \\\\ \\end{array}", + "image_path": "063ba139a21194d8bcf5712d80ab4e1149156d073916e06dbd758681f3020b2d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{k}^{prev}" + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "text", + "content": " is the previous estimation of the primitive parameters. " + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{a}" + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "text", + "content": " is an adaptive subset of the complete voxels " + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "text", + "content": ", which includes the voxels close to the surface of the previous estimated primitive. " + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "text", + "content": " defines the distance threshold of the activated region. Only voxels in " + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{a}" + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "text", + "content": " are activated during the optimization. The reason why we use an adaptive subset instead of the complete voxel space is twofold. As discussed in Sec. 3.1, both the source and the target SDF are truncated within a vicinity of zero. Therefore, small variations around " + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{k}^{prev}" + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "text", + "content": " (i.e., small changes on the shape of the primitive surface) have minor if not zero effects on the source SDF values evaluated at voxels distant to the primitive surface. Moreover, the size of " + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{a}" + }, + { + "bbox": [ + 305, + 133, + 545, + 312 + ], + "type": "text", + "content": " is much smaller than the complete set, and thus providing a significant boost in performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 314, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 546, + 422 + ], + "type": "text", + "content": "The correspondence marching and the primitive updating alternate until convergence. This process is akin to the EM algorithm [13]. The scope of voxels from which a primitive can be extracted is 'marched' with the progressive variation of the primitive shape, and simultaneously the parameters of the primitive are optimized based on the scope of voxels. We apply the optimization method proposed in [24] to avoid local minima. More derivation and implementation details can be found in the Supplementary Material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 431, + 459, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 459, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 459, + 444 + ], + "type": "text", + "content": "3.5. Fail-safe Auto-degeneration" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": "The primitives are roughly initialized in the connectivity marching step and further grow to capture local geometric shapes. Since the probabilistic marching is based on optimization, an important question to ask is if the recovered primitives can always converge to a proper shape, regardless of poor initialization. We demonstrate that our probabilistic marching strategy is robust to initialization. Firstly, our method possesses a feature called auto-degeneration i.e. the primitive will autonomously degenerate towards a point when accidentally initialized far from the target volumes. This is because, under this circumstance, only variations which shrink the volume of the primitive can decrease the difference between the truncated target and source SDF. Secondly, when the primitive is initialized near or inside the target volumes, the marching process encourages the primitive to converge to one nearest local target shape by analyzing the posterior correspondences. Fig.2(c) illustrates the examples of the above properties. After the primitive marching, our algorithm removes the degenerated primitives. As a fail-safe measure, primitives which significantly contradict the target SDF is also removed. Detailed implementations can be found in the Supplementary Material." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8775" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 516, + 246 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 516, + 246 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 516, + 246 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 516, + 246 + ], + "type": "image", + "image_path": "0e66a71d780764a14716b2cca881026a9394a5796e79754d5f6d89cf964d2e74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 281 + ], + "type": "text", + "content": "Figure 3. Shape Abstraction results on the ShapeNet dataset. From left to right: SQs [31], Non-parametric Bayesian (NB) [47], Marching-Primitives with ellipsoids (MPE), Marching-Primitives with superquadrics (MPS), and the ground truth (pre-processed watertight mesh)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 286, + 129, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 129, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 129, + 300 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 306, + 290, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 290, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 290, + 427 + ], + "type": "text", + "content": "In this section, we conduct several experiments to demonstrate the high accuracy and generalizability of our proposed method. First, we show the shape abstraction results on various datasets, ranging from daily objects to human body models. Furthermore, we study the impact of the grid resolution and the truncation threshold on our algorithm. In the end, we apply our algorithm on a large-scale complex indoor scene, which contains a mixed combination of furniture such as sofas, tables, books, lamps, etc. Implementation details can be found in the Supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 432, + 179, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 179, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 179, + 443 + ], + "type": "text", + "content": "4.1. Evaluation on Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "content": "Baselines: We compare our method with both the state-of-the-art learning-based method [31] and the computational method [47], which infer superquadric abstractions of the input objects. For convenience, we refer to [31] as SQs, and [47] as NB. We use the official codes and follow the implementation details as stated in these papers, respectively. We do not compare with the cuboid-based algorithms [40, 48, 50], since they focus more on the semantic level abstraction and are limited in expressiveness and accuracy. A detailed performance comparison between the cuboid and superquadric abstractions can be found in [31]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Datasets: We evaluate on two commonly used datasets, the ShapeNet [6] and the DFAUST [4]. In both of the datasets, we are only provided with triangular meshes. Therefore, we need to transform the meshes into discrete SDF representations. It is straightforward to calculate the signed distance value of a point to a watertight mesh. However, the ShapeNet is a human-made synthetic dataset containing many non-watertight meshes formed by nonvolumetric 2D surfaces, self-intersecting or overlapped triangles. Therefore, we pre-process the original ShapeNet models into watertight meshes and then generate the SDF" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 287, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 361 + ], + "type": "text", + "content": "representation with the fast marching algorithm, following the same procedure in [8]. In consideration of fairness and consistency, we use the pre-processed watertight meshes as the common ground truth and the inputs to SQs and NB. For the DFAUST dataset, we generate the SDF directly since the provided meshes are already watertight." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 362, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 547, + 411 + ], + "type": "text", + "content": "Metrics: Following [31,47] we use the Chamfer-L1 distance and the volumetric intersection over union (IoU) as the quantitative evaluation metrics. The computation of the metrics is detailed in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "text", + "content": "Results on ShapeNet: We experiment on 14 categories from the ShapeNet dataset. We split the dataset randomly into the training " + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "inline_equation", + "content": "(80\\%)" + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "text", + "content": " and testing " + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "inline_equation", + "content": "(20\\%)" + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "text", + "content": " sets [9], where we train one SQs model per category on the training sets and evaluate all the methods on the testing sets. For our method, we use the SDF representation discretized on a voxel grid of size " + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "inline_equation", + "content": "100^3" + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "text", + "content": " and range " + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "inline_equation", + "content": "[-0.5, 0.5]^3" + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "text", + "content": ". Other than superquadrics, we also test our method with ellipsoids as the base primitive, which is a special case of the superquadric representation when the shape parameters " + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_1, \\epsilon_2" + }, + { + "bbox": [ + 304, + 415, + 548, + 715 + ], + "type": "text", + "content": " are fixed to 1. The quantitative results are summarized in Table. 1. We also demonstrate the qualitative comparison among different shape abstraction methods in Fig.3. Our method outperforms all the baselines, even implemented with less expressive ellipsoids. The computational method generally performs better than the learning-based method. This is because the parameters of the primitive are so sparse and geometrically interrelated that they are difficult to get mapped from a high dimensional input by a neural network. Compared with NB, our method has richer details, clearer edges, and more importantly does not occupy the exterior space. Two factors contribute to the advantages. The first one is the extra geometric information embedded in the SDF representation, which eliminates the inherent interior/exterior ambiguity of point clouds. The other one is our special de" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "type": "text", + "content": "8776" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 537, + 278 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 537, + 278 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 537, + 278 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 537, + 278 + ], + "type": "table", + "html": "
CategoryChamfer-L1IoU
SQs [31]NB [47]MPE(Ours)MPS(Ours)SQs [31]NB [47]MPE(Ours)MPS(Ours)
airplane0.0370.0230.0210.0190.4410.6710.7310.768
bench0.0560.0280.0200.0200.2380.5790.7300.819
bottle0.0470.0330.0260.0170.6860.6650.8860.924
cabinet0.0590.0360.0370.0280.3940.6660.8400.948
can0.0660.0360.0360.0220.7060.5530.9080.950
chair0.0680.0270.0230.0200.3000.6850.7850.871
lamp0.0720.0290.0220.0210.2340.5890.7500.802
speaker0.0640.0410.0370.0330.3460.6560.8580.920
mailbox0.0950.0260.0260.0240.3330.6940.8020.905
rifle0.0380.0200.0190.0190.4460.7320.7440.811
sofa0.0540.0370.0290.0230.4970.7260.8570.940
table0.0700.0240.0240.0220.2470.7450.8180.932
phone0.0400.0210.0230.0210.6810.8720.8910.947
watercraft0.0480.0320.0220.0220.4650.6180.7930.836
mean0.0570.0280.0240.0220.3680.6740.7930.870
", + "image_path": "b0e36ce4c1af47b632b04f455d509df2f928072ed4981a9e033dcabde4984f96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 286, + 545, + 298 + ], + "lines": [ + { + "bbox": [ + 47, + 286, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 545, + 298 + ], + "type": "text", + "content": "Table 1. Quantitative results on Shapenet. MPE and MPS are short for Marching-Primitives with ellipsoids and superquadrics, respectively" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 302, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 288, + 350 + ], + "type": "text", + "content": "sign of the probabilistic generative model, as discussed in Sec.3.4. It is interesting to observe that our algorithm can extract abstractions from objects not intuitively depictable by the primitives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 542 + ], + "type": "text", + "content": "Results on DFAUST: We also evaluate on the DFAUST dataset of human body models. We follow the split settings in [29] to train the learning-based method and evaluate all the methods on the testing set. The SDF is discretized on a grid of size " + }, + { + "bbox": [ + 46, + 350, + 287, + 542 + ], + "type": "inline_equation", + "content": "64^3" + }, + { + "bbox": [ + 46, + 350, + 287, + 542 + ], + "type": "text", + "content": " and side length 1. Similar to the ShapeNet experiment, we test our method with both superquadrics and ellipsoids. The results are shown in Fig.4. Our method can accurately capture various postures, while the baselines fail to distinguish different body parts in some cases. The abstraction quality of SQs is much better on this dataset compared with the ShapeNet, since human bodies share a common articulated structure that can be captured by the neural network. We observe that the ellipsoid abstraction also achieves satisfying accuracy. This is because the human body mostly consists of rounded shapes compared with man-made objects in the ShapeNet." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 551, + 160, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 160, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 160, + 563 + ], + "type": "text", + "content": "4.2. Performance Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "This section investigates how the Marching-Primitives algorithm behaves with different voxel grid sizes and distance truncation thresholds. We experiment on the Armadillo from the Stanford 3D scanning repository [10]. First, we discretize the model on the voxel grids of different resolutions. The qualitative abstraction results are visualized in Fig. 5, and the quantitative results are summarized in Fig. 6(a). When the input grid resolution is low, the recovered primitive representation is relatively abstract. With the increase of the grid resolution, our method is able to extract detailed abstraction more faithful to the target shape. This is because the higher the resolution, the more geomet" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 302, + 532, + 574 + ], + "blocks": [ + { + "bbox": [ + 319, + 302, + 532, + 574 + ], + "lines": [ + { + "bbox": [ + 319, + 302, + 532, + 574 + ], + "spans": [ + { + "bbox": [ + 319, + 302, + 532, + 574 + ], + "type": "image", + "image_path": "bb148d17d21e964935337f1bef81e44218b01641e92e3636b80e35ab4861fd9b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 584, + 545, + 605 + ], + "lines": [ + { + "bbox": [ + 305, + 584, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 584, + 545, + 605 + ], + "type": "text", + "content": "Figure 4. Shape abstraction results on the DFAUST dataset. From left to right: SQs [31], NB [47], MPE, MPS, and the ground truth." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "ric information our method can utilize to guide the primitive marching process. As discussed in Sec.3.1, we use a truncated version of the target and source SDF. Therefore, we also evaluate the performance on different truncation thresholds, starting from 0.1 to 4 times the interval of the voxel grids. The results are shown in Fig.6(b). The abstraction accuracy increases as the threshold decreases at first. The reason lies in that the approximated superquadric" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8777" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 72, + 153, + 150 + ], + "blocks": [ + { + "bbox": [ + 88, + 72, + 153, + 150 + ], + "lines": [ + { + "bbox": [ + 88, + 72, + 153, + 150 + ], + "spans": [ + { + "bbox": [ + 88, + 72, + 153, + 150 + ], + "type": "image", + "image_path": "31fc9eb13733909ba8a33c97ce518a25ffed8eaac37d4fdf2806906d06401ffc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 152, + 129, + 158 + ], + "lines": [ + { + "bbox": [ + 120, + 152, + 129, + 158 + ], + "spans": [ + { + "bbox": [ + 120, + 152, + 129, + 158 + ], + "type": "text", + "content": "32" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 74, + 220, + 150 + ], + "blocks": [ + { + "bbox": [ + 154, + 74, + 220, + 150 + ], + "lines": [ + { + "bbox": [ + 154, + 74, + 220, + 150 + ], + "spans": [ + { + "bbox": [ + 154, + 74, + 220, + 150 + ], + "type": "image", + "image_path": "68049e4909d9f3938e3bb1ce1a6fa9b0d9b3ac62dedd802df6e91d7e83ed252f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 152, + 193, + 158 + ], + "lines": [ + { + "bbox": [ + 183, + 152, + 193, + 158 + ], + "spans": [ + { + "bbox": [ + 183, + 152, + 193, + 158 + ], + "type": "inline_equation", + "content": "64^{3}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 75, + 287, + 150 + ], + "blocks": [ + { + "bbox": [ + 222, + 75, + 287, + 150 + ], + "lines": [ + { + "bbox": [ + 222, + 75, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 222, + 75, + 287, + 150 + ], + "type": "image", + "image_path": "6b0e817b2c625223415bb6965c9e432bb019a98cfbc98c7ab65b870b327bc8c1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 152, + 260, + 158 + ], + "lines": [ + { + "bbox": [ + 251, + 152, + 260, + 158 + ], + "spans": [ + { + "bbox": [ + 251, + 152, + 260, + 158 + ], + "type": "text", + "content": "96" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 288, + 74, + 354, + 150 + ], + "blocks": [ + { + "bbox": [ + 288, + 74, + 354, + 150 + ], + "lines": [ + { + "bbox": [ + 288, + 74, + 354, + 150 + ], + "spans": [ + { + "bbox": [ + 288, + 74, + 354, + 150 + ], + "type": "image", + "image_path": "86bf24b15c2b460d791b9bbe4d34fb764380ca886847fa3806f0a7d9d800a6cb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 152, + 327, + 158 + ], + "lines": [ + { + "bbox": [ + 315, + 152, + 327, + 158 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 327, + 158 + ], + "type": "text", + "content": "128" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 356, + 75, + 421, + 150 + ], + "blocks": [ + { + "bbox": [ + 356, + 75, + 421, + 150 + ], + "lines": [ + { + "bbox": [ + 356, + 75, + 421, + 150 + ], + "spans": [ + { + "bbox": [ + 356, + 75, + 421, + 150 + ], + "type": "image", + "image_path": "7fb6571ce79fabbf090576e7f03d6e440163f61c7d858053c33290ab6cfab060.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 152, + 396, + 158 + ], + "lines": [ + { + "bbox": [ + 384, + 152, + 396, + 158 + ], + "spans": [ + { + "bbox": [ + 384, + 152, + 396, + 158 + ], + "type": "inline_equation", + "content": "256^{3}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 424, + 76, + 490, + 150 + ], + "blocks": [ + { + "bbox": [ + 424, + 76, + 490, + 150 + ], + "lines": [ + { + "bbox": [ + 424, + 76, + 490, + 150 + ], + "spans": [ + { + "bbox": [ + 424, + 76, + 490, + 150 + ], + "type": "image", + "image_path": "9c450411222d5b7b515756deb1026a68ecda48dd7d80042bf51b9dbefdd08a71.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 152, + 479, + 158 + ], + "lines": [ + { + "bbox": [ + 443, + 152, + 479, + 158 + ], + "spans": [ + { + "bbox": [ + 443, + 152, + 479, + 158 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 190, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 262 + ], + "type": "text", + "content": ".source) SDF converges to the true value when truncated close to the surface. If we further decrease the threshold, however, we observe an acute decrease in accuracy. This is because an overly small truncation threshold corrupts the original geometric information embedded in the target SDF, as well as the smoothness of the cost function (Eq. (10))." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 56, + 278, + 166, + 358 + ], + "blocks": [ + { + "bbox": [ + 56, + 278, + 166, + 358 + ], + "lines": [ + { + "bbox": [ + 56, + 278, + 166, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 166, + 358 + ], + "type": "image", + "image_path": "03bb3d3f4bde15145e1b737c91a4aa7dedf067df550f4b9924f93af47a459939.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 367, + 287, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 411 + ], + "type": "text", + "content": "Figure 6. Quantitative results on shape abstraction accuracy. (a) Abstraction accuracy versus input grid resolution. (b) Abstraction accuracy versus truncation ratio, evaluated at grid resolution " + }, + { + "bbox": [ + 46, + 367, + 287, + 411 + ], + "type": "inline_equation", + "content": "128^{3}" + }, + { + "bbox": [ + 46, + 367, + 287, + 411 + ], + "type": "text", + "content": ". We evaluate on two metrics: Chamfer-" + }, + { + "bbox": [ + 46, + 367, + 287, + 411 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 367, + 287, + 411 + ], + "type": "text", + "content": " and IoU." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 167, + 279, + 283, + 358 + ], + "blocks": [ + { + "bbox": [ + 167, + 279, + 283, + 358 + ], + "lines": [ + { + "bbox": [ + 167, + 279, + 283, + 358 + ], + "spans": [ + { + "bbox": [ + 167, + 279, + 283, + 358 + ], + "type": "image", + "image_path": "2fcfebed2174d44466761b443816a0d8119cfd6218fafcdcce31ccd74692a713.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 363, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 397 + ], + "type": "text", + "content": "Figure 7. Qualitative result of the scene abstraction. (a) Superquadric abstraction by the Marching-Primitives. (b) Mesh generated by the marching cubes [25] from the same SDF." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 430, + 155, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 155, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 155, + 441 + ], + "type": "text", + "content": "4.3. Scene Abstraction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": "Other than single objects, we also test the capability of our method in representing complex scenes with geometric primitives. We experiment on a real-world indoor scene called Reading Room [49]. The scene is captured by an Asus Xtion Pro Live camera. The RGB-D scans are fused utilizing [19] (an SDF-based 3D reconstruction algorithm), and further fine-tuned with the method proposed in [49]. However, the SDF representation of the scene is not available publicly. Therefore, we pre-process the scene mesh by removing the floor plane, filling the holes, and transforming it into the SDF representation with grid size " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "400^3" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ". The scene abstraction task is much more difficult compared with the object-level task because various items with great differences in size and shape are present in the very same space. We compare our abstraction result with the mesh extracted by the marching cubes algorithm on the same discrete SDF representation, as shown in Fig.7. Our representation is not only visually satisfying but also contains abundant geometric information the mesh lacks, since it is a continuous SDF approximation to the input discrete SDF. Moreover, our highly sparse representation is only 8.2KB in size, while the discrete SDF occupies 203MB." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 347, + 191, + 503, + 271 + ], + "blocks": [ + { + "bbox": [ + 46, + 175, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 545, + 186 + ], + "type": "text", + "content": "Figure 5. Visualization of the shape abstractions on different grid resolutions. Abstraction accuracy increases with the input grid resolution." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 191, + 503, + 271 + ], + "lines": [ + { + "bbox": [ + 347, + 191, + 503, + 271 + ], + "spans": [ + { + "bbox": [ + 347, + 191, + 503, + 271 + ], + "type": "image", + "image_path": "90d4e1078ea8c708964d17b6f18d58caceeff00d5e21035dd75b9a9734a67e93.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 347, + 273, + 503, + 352 + ], + "blocks": [ + { + "bbox": [ + 347, + 273, + 503, + 352 + ], + "lines": [ + { + "bbox": [ + 347, + 273, + 503, + 352 + ], + "spans": [ + { + "bbox": [ + 347, + 273, + 503, + 352 + ], + "type": "image", + "image_path": "e338b050542461c4d7f525f208a3cac6778e0dc8b4b88925884ff1dd0515d7a2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 406, + 378, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 378, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 378, + 418 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 426, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 618 + ], + "type": "text", + "content": "We propose the Marching-Primitives, the first algorithm to extract primitive-based abstractions from the volumetric SDF representation. Our method outperforms the state-of-the-art in terms of accuracy and generalizability on both synthetic and real-world datasets. Our primitive-based representation is sparse, accurate, generalizable, and can be expressed analytically without training, which we believe will facilitate and inspire further explorations in scene understanding, 3D reconstruction, and robot motion planning. However, our algorithm has the limitation of not being able to properly extract object parts thinner than the interval of the voxel grid. This problem can be solved by either increasing the grid resolution or detecting and thickening the invalid parts. Future work includes the parallelization of the marching process and inferring semantic-level interpretations from the current primitive-based geometric features." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This research is supported by NUS Startup grants A-0009059-02-00 and A-0009059-03-00, National Research Foundation, Singapore, under its Medium Sized Centre Programme - Centre for Advanced Robotics Technology Innovation (CARTIN), subaward A-0009428-08-00, and AME Programmatic Fund Project MARIO A-0008449-01-00, and JHU internal funds." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8778" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] R. Bajcsy and F. Solina. Three dimensional object representation revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 231-240, 1987. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] A. H. Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1(1):11-23, 1981. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "type": "text", + "content": "[3] I. Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2):115, 1987. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "text", + "content": "[4] F. Bogo, J. Romero, G. Pons-Moll, and M. J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "text", + "content": "[5] D. E. Breen, S. Mauch, and R. T. Whitaker. 3d scan conversion of csg models into distance volumes. In Proceedings of the 1998 IEEE Symposium on Volume Visualization, pages 7-14, 1998. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 340 + ], + "type": "text", + "content": "[6] A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "type": "text", + "content": "[7] L. Chevalier, J. Jaillet, and A. Baskurt. Segmentation and superquadric modeling of 3D objects. In The 11-th International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision'2003, WSCG 2003, 2003. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "text", + "content": "[8] G. Chou, I. Chugunov, and F. Heide. Gensdf: Two-stage learning of generalizable signed distance functions. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "type": "text", + "content": "[9] C. B. Choy, D. Xu, J. Gwak, K. Chen, and S. Savarese. 3dr2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-644. Springer International Publishing, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "text", + "content": "[10] B. Curless and M. Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[11] A. Dai, D. Ritchie, M. Bokeloh, S. Reed, J. Sturm, and M. Nießner. Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[12] A. Dai, C. Ruizhongtai Qi, and M. Niessner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[13] A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[14] B. Deng, K. Genova, S. Yazdani, S. Bouaziz, G. Hinton, and A. Tagliasacchi. Cvxnet: Learnable convex decomposition." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 31-41, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "type": "text", + "content": "[15] W. Dong, Q. Wang, X. Wang, and H. Zha. Psdf fusion: Probabilistic signed distance function for on-the-fly 3d data fusion and scene reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 701-717, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "type": "text", + "content": "[16] K. Genova, F. Cole, D. Vlasic, A. Sarna, W. Freeman, and T. Funkhouser. Learning shape templates with structured implicit functions. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7153-7163, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[17] A. D. Gross and T. E. Boult. Error of fit measures for recovering parametric solids. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 690-694, 1988. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "type": "text", + "content": "[18] Z. Hao, H. Averbuch-Elor, N. Snavely, and S. Belongie. Dualsf: Semantic shape manipulation using a two-level representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7631–7641, 2020. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 304, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 371 + ], + "type": "text", + "content": "[19] S. Izadi, D. Kim, O. Hilliges, D. Molyneaux, R. Newcombe, P. Kohli, J. Shotton, S. Hodges, D. Freeman, A. Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "type": "text", + "content": "[20] D. K. Katherine and S. S. Elizabeth. Core systems in human cognition. In From Action to Cognition, volume 164 of Progress in Brain Research, pages 257-264. Elsevier, 2007. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "type": "text", + "content": "[21] Y. Kawana, Y. Mukuta, and T. Harada. Neural star domain as primitive representation. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7875-7886. Curran Associates, Inc., 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[22] A. Leonardis, A. Jaklic, and F. Solina. Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(11):1289-1295, 1997. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[23] Y. Liao, S. Donne, and A. Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2916-2925, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "type": "text", + "content": "[24] W. Liu, Y. Wu, S. Ruan, and G. S. Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2676-2685, June 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "text", + "content": "[25] W. E. Lorensen and H. E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 2, 3, 4, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[26] R. A. Newcombe, D. Fox, and S. M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[27] C. Niu, J. Li, and K. Xu. Im2Struct: Recovering 3D shape structure from a single RGB image. In Proceedings of the" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8779" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[28] J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[29] D. Paschalidou, L. V. Gool, and A. Geiger. Learning unsupervised hierarchical part decomposition of 3D objects from a single RGB image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "text", + "content": "[30] D. Paschalidou, A. Katharopoulos, A. Geiger, and S. Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3215, June 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 318 + ], + "type": "text", + "content": "[31] D. Paschalidou, A. O. Ulusoy, and A. Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "text", + "content": "[32] A. Pentland. Perceptual organization and the representation of natural form. Artificial Intelligence, 28(3):293-331, 1986. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 409 + ], + "type": "text", + "content": "[33] A. H. Quispe, B. Milville, M. A. Gutiérrez, C. Erdogan, M. Stilman, H. Christensen, and H. B. Amor. Exploiting symmetries and extrusions for grasping household objects. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3702-3708, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[34] Y. Rao, Y. Nie, and A. Dai. Patchcomplete: Learning multi-resolution patch priors for 3d shape completion on unseen categories. Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "text", + "content": "[35] S. Ruan, K. L. Poblete, H. Wu, Q. Ma, and G. S. Chirikjian. Efficient path planning in narrow passages for robots with ellipsoidal components. IEEE Transactions on Robotics (TRO), pages 1-18, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "text", + "content": "[36] S. Ruan, X. Wang, and G. S. Chirikjian. Collision detection for unions of convex bodies with smooth boundaries using closed-form contact space parameterization. IEEE Robotics and Automation Letters (RAL), 7(4):9485-9492, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "type": "text", + "content": "[37] G. Sharma, B. Dash, A. RoyChowdhury, M. Gadelha, M. Loizou, L. Cao, R. Wang, E. Learned-Miller, S. Maji, and E. Kalogerakis. Prift: Learning to fit primitives improves few shot point cloud segmentation. In Computer Graphics Forum, volume 41, pages 39-50. Wiley Online Library, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "text", + "content": "[38] D. Smirnov, M. Fisher, V. G. Kim, R. Zhang, and J. Solomon. Deep parametric shape predictions using distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 561-570, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[39] F. Solina and R. Bajcsy. Recovery of parametric models from range images: the case for superquadrics with global deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(2):131-147, 1990. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 576 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[40] C. Sun, Q. Zou, X. Tong, and Y. Liu. Learning adaptive hierarchical cuboid abstractions of 3d shape collections. ACM Transactions on Graphics (TOG), 38(6):1-13, 2019. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[41] S. Tulsiani, H. Su, L. J. Guibas, A. A. Efros, and J. Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[42] N. Vaskevicius and A. Birk. Revisiting superquadric fitting: A numerically stable formulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):220-233, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "type": "text", + "content": "[43] S. Vasu, N. Talabot, A. Lukoianov, P. Baque, J. Donier, and P. Fua. Hybridsdf: Combining deep implicit shapes and geometric primitives for 3d shape representation and manipulation. In International Conference on 3D Vision (3DV), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "text", + "content": "[44] G. Vezzani, U. Pattacini, and L. Natale. A grasping approach based on superquadric models. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1579-1586, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[45] G. Vezzani, U. Pattacini, G. Pasquale, and L. Natale. Improving superquadric modeling and grasping with prior on object shapes. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 6875-6882, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "text", + "content": "[46] S. Weder, J. L. Schonberger, M. Pollefeys, and M. R. Oswald. Routedfusion: Learning real-time depth map fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "type": "text", + "content": "[47] Y. Wu, W. Liu, S. Ruan, and G. S. Chirikjian. Primitive-based shape abstraction via nonparametric bayesian inference. In Proceedings of the European Conference on Computer Vision (ECCV), pages 479-495. Springer Nature Switzerland, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 443, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 485 + ], + "type": "text", + "content": "[48] K. Yang and X. Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 487, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 530 + ], + "type": "text", + "content": "[49] Q. Zhou, S. Miller, and V. Koltun. Elastic fragments for dense scene reconstruction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 473-480, 2013. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "type": "text", + "content": "[50] C. Zou, E. Yumer, J. Yang, D. Ceylan, and D. Hoiem. 3D-PRNN: Generating shape primitives with recurrent neural networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8780" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_content_list.json b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c44e40db0d2ada81eac8de53cc11f164339519a --- /dev/null +++ b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_content_list.json @@ -0,0 +1,1457 @@ +[ + { + "type": "text", + "text": "MarginMatch: Improving Semi-Supervised Learning with Pseudo-Margins", + "text_level": 1, + "bbox": [ + 106, + 128, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tiberiu Sosea Cornelia Caragea University of Illinois Chicago tsosea2@uic.edu cornelia@uic.edu", + "bbox": [ + 323, + 181, + 643, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce MarginMatch, a new SSL approach combining consistency regularization and pseudo-labeling, with its main novelty arising from the use of unlabeled data training dynamics to measure pseudo-label quality. Instead of using only the model's confidence on an unlabeled example at an arbitrary iteration to decide if the example should be masked or not, MarginMatch also analyzes the behavior of the model on the pseudo-labeled examples as the training progresses, to ensure low quality predictions are masked out. MarginMatch brings substantial improvements on four vision benchmarks in low data regimes and on two large-scale datasets, emphasizing the importance of enforcing high-quality pseudo-labels. Notably, we obtain an improvement in error rate over the state-of-the-art of $3.25\\%$ on CIFAR-100 with only 25 labels per class and of $3.78\\%$ on STL-10 using as few as 4 labels per class. We make our code available at https://github.com/tsosea2/MarginMatch.", + "bbox": [ + 75, + 301, + 473, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 571, + 209, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models have seen tremendous success in many vision tasks [14, 22, 27, 42, 43]. This success can be attributed to their scalability, being able to produce better results when they are trained on large datasets in a supervised fashion [15, 27, 34, 35, 43, 47]. Unfortunately, large labeled datasets annotated for various tasks and domains are difficult to acquire and demand considerable annotation effort or domain expertise. Semi-supervised learning (SSL) is a powerful approach that mitigates the requirement for large labeled datasets by effectively making use of information from unlabeled data, and thus, has been studied extensively in vision [4, 5, 23, 25, 30, 36, 38, 39, 44-46].", + "bbox": [ + 75, + 597, + 470, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent SSL approaches integrate two important components: consistency regularization [46, 49] and pseudolabeling [25]. Consistency regularization works on the assumption that a model should output similar predictions when fed perturbed versions of the same image, whereas pseudo-labeling uses the model's predictions of unlabeled examples as labels to train against. For example, Sohn et al. [41] introduced FixMatch that combines consistency reg", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ularization on weak and strong augmentations with pseudolabeling. FixMatch relies heavily on a high-confidence threshold to compute the unsupervised loss, disregarding any pseudo-labels whose confidence falls below this threshold. While training using only high-confidence pseudo-labels has shown to consistently reduce the confirmation bias [1], this rigid threshold allows access only to a small amount of unlabeled data for training, and thus, ignores a considerable amount of unlabeled examples for which the model's predictions do not exceed the confidence threshold. More recently, Zhang et al. [49] introduced FlexMatch that relaxes the rigid confidence threshold in FixMatch to account for the model's learning status of each class in that it adaptively scales down the threshold for a class to encourage the model to learn from more examples from that class. The flexible thresholds in FlexMatch allow the model to have access to a much larger and diverse set of unlabeled data to learn from, but lowering the thresholds can lead to the introduction of wrong pseudo-labels, which are extremely harmful for generalization. Interestingly, even when the high-confidence threshold is used in FixMatch can result in wrong pseudo-labels. See Figure 1 for incorrect pseudo-labels detected in the training set after we apply FixMatch and FlexMatch on ImageNet. We posit that a drawback of FixMatch and FlexMatch and in general of any pseudo-labeling approach is that they use the confidence of the model only at the current iteration to enforce quality of pseudo-labels and completely ignore model's predictions at prior iterations.", + "bbox": [ + 496, + 268, + 895, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose MarginMatch, a new SSL approach that monitors the behavior of the model on the unlabeled examples as the training progresses, from the beginning of training until the current iteration, instead of using only the model's current belief about an unlabeled example (i.e., its confidence at the current iteration) to decide if the example should be masked or not. We estimate a pseudo-label's contribution to learning and generalization by introducing pseudo-margins of unlabeled examples averaged across training iterations. Pseudo-margins of unlabeled examples extend the margins from machine learning [3, 11, 18, 33] which provide a measure of confidence of the outputs of the model and capture the difference between the output for the correct", + "bbox": [ + 496, + 703, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15773", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d697b74530a6eb1d50b7f27f7764bd5484004a26b834d52b3cc2faab6dd374d.jpg", + "image_caption": [ + "Figure 1. Incorrect pseudo-labels propagated until the end of the training process for FixMatch and FlexMatch on ImageNet." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 874, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(gold) label and the other labels. In our case, the pseudomargins capture how much larger the assigned logit (the logit corresponding to the argmax of the model's prediction) is compared with all other logits at iteration $t$ . Similar to FlexMatch, in MarginMatch we take advantage of the flexible confidence thresholds to allow the model to learn from larger and more diverse sets of unlabeled examples, but unlike FlexMatch, we train the model itself to identify the characteristics of mislabeled pseudo-labels simply by monitoring the model's training dynamics on unlabeled data over the iterations.", + "bbox": [ + 75, + 345, + 472, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We carry out comprehensive experiments using established SSL experimental setups on CIFAR-10, CIFAR-100 [21], SVHN [31], STL-10 [8], ImageNet [10], and WebVision [26]. Despite its simplicity, our findings indicate that MarginMatch produces improvements in performance over strong baselines and prior works on all datasets at no additional computational cost. Notably, compared to current state-of-the-art, on CIFAR-100 we see $3.02\\%$ improvement in error rate using only 4 labels per class and $3.78\\%$ improvement on STL-10 using the same extremely label-scarce setting of 4 labels per class. In addition, on ImageNet [10] and WebVision [26] we find that MarginMatch pushes the state-of-the-art error rates by $0.97\\%$ on ImageNet and by $0.79\\%$ on WebVision.", + "bbox": [ + 75, + 512, + 472, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 96, + 724, + 315, + 737 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a new SSL approach which we call MarginMatch that enforces high pseudo-label quality during training. Our approach allows access to a large set of unlabeled data to learn from (thus, incorporating more information from unlabeled data) and, at the same time, monitors the training dynamics of unlabeled data as training progresses to detect and filter out potentially incorrect pseudo-labels.", + "2. We show that MarginMatch outperforms existing works on six well-established computer vision benchmarks" + ], + "bbox": [ + 89, + 743, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "showing larger improvements in error rates especially on challenging datasets, while achieving similar convergence performance (or better) than prior works.", + "bbox": [ + 529, + 345, + 893, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We perform a comprehensive analysis of our approach and indicate potential insights into why our Margin-Match substantially outperforms other SSL techniques.", + "bbox": [ + 511, + 395, + 893, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. MarginMatch", + "text_level": 1, + "bbox": [ + 500, + 452, + 640, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notation Let $L = \\{(x_{1},y_{1}),\\dots,(x_{B},y_{B})\\}$ be a batch of size $B$ of labeled examples and $U = \\{\\hat{x}_1,\\dots,\\hat{x}_{\\nu B}\\}$ be a batch of size $\\nu B$ of unlabeled examples, where $\\nu$ is the batch-wise ratio of unlabeled to labeled examples. Let $p_{\\theta}(y|x)$ denote the class distribution produced by model $\\theta$ on input image $x$ and $\\hat{p}_{\\theta}(y|x)$ denote the argmax of this distribution as a one-hot label. Let also $H(p,q)$ denote the cross-entropy between two probability distributions $p$ and $q$ .", + "bbox": [ + 496, + 477, + 893, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Background", + "text_level": 1, + "bbox": [ + 500, + 606, + 632, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consistency reg [39] is an important component in recent semi-supervised learning approaches and relies on the continuity assumption [2,23] that the model should output similar predictions on multiple perturbed versions of the same input $x$ . As mentioned above, examples of two such approaches are FixMatch [41] and FlexMatch [49] that use consistency regularization at their core combined with psedo-labeling. In psedo-labeling [25], a model itself is used to assign artificial labels for unlabeled data and only artificial labels whose largest class probability is above a predefined confidence threshold are used during training.", + "bbox": [ + 496, + 628, + 893, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, FixMatch [41] predicts artificial labels for unlabeled examples using a weakly-augmented version of each unlabeled example and then employs the artificial labels as pseudo-labels to train against but this time using a strongly-augmented version of each unlabeled example. That is, FixMatch minimizes the following batch-wise consistency loss on unlabeled data:", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15774", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right)\\right) > \\tau\\right) \\quad \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 89, + 468, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is a confidence threshold, $\\pi$ and $\\Pi$ are weak and strong augmentations, respectively, and $\\mathbb{1}$ is the indicator function. Therefore, the low-confidence examples (lower than $\\tau$ ) are completely ignored despite containing potentially useful information for model training.", + "bbox": [ + 75, + 161, + 468, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FlexMatch [49] argues that using a fixed threshold $\\tau$ to filter the unlabeled data ignores the learning difficulties of different classes, and thus, introduces class-dependent thresholds, which are obtained by adaptively scaling $\\tau$ depending on the learning status of each class. FlexMatch assumes that a class with fewer examples above the fixed threshold $\\tau$ has a greater learning difficulty, and hence, it adaptively lowers the threshold $\\tau$ to encourage more training examples from this class to be learned. The learning status $\\alpha_{c}$ for a class $c$ is simply computed as the number of unlabeled examples that are predicted in class $c$ and pass the fixed threshold $\\tau$ :", + "bbox": [ + 75, + 243, + 470, + 411 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {c} = \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right)\\right) > \\tau\\right) \\mathbb {1} \\left(\\hat {p} _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right) = c\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 420, + 468, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n$ is the total number of unlabeled examples. This learning effect is then normalized and used to obtain the class-dependent threshold for each class $c$ :", + "bbox": [ + 75, + 472, + 468, + 517 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {c} = \\frac {\\alpha_ {c}}{\\max _ {c} \\left(\\alpha_ {c}\\right)} \\times \\tau \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 527, + 468, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, FlexMatch iteratively computes new thresholds after each complete pass through unlabeled data, hence we can parameterize $\\mathcal{T}_c$ as $\\mathcal{T}_c^t$ , denoting the threshold obtained at iteration $t$ . The unlabeled loss is then obtained by plugging in the adaptive threshold $\\mathcal{T}_c^t$ in Eq. 1:", + "bbox": [ + 75, + 568, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} (\\max (p _ {\\theta} (y | \\pi (\\hat {x} _ {i}))) > \\mathcal {T} _ {\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i}))} ^ {t}) \\quad \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 662, + 468, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The aforementioned works use the confidence of the model solely at the current iteration to enforce quality of pseudo-labels. We believe this is not sufficient as it provides only a myopic view of the model's behavior (i.e., its confidence) on unlabeled data (at a single iteration) and may result in wrong pseudo-labels even when the confidence threshold is high enough (e.g., if the model is miscalibrated or overly-confident [13]). Figure 1 shows examples of images that are added to the training set with a wrong pseudo-label for both FixMatch and FlexMatch. These types of unlabeled examples, which are incorrectly pseudo-labeled and used", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "during training are particularly harmful for deep neural networks, which can attain zero training error on any dataset, even on randomly assigned labels [50], resulting in poor generalization capabilities.", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Proposed Approach: MarginMatch", + "text_level": 1, + "bbox": [ + 500, + 161, + 807, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now introduce MarginMatch, our new SSL approach that uses the model's training dynamics on unlabeled data to improve pseudo-label data quality. Our approach leverages consistency regularization with weak and strong augmentations and pseudo-labeling, but instead of using only the model's current belief (i.e., its confidence at the current iteration) to decide if an unlabeled example should be used for training or not, our MarginMatch monitors the training dynamics of unlabeled data over the iterations by investigating the margins (a measure of confidence) of the outputs of the model [3]. The margin of a training example is a well established metric in machine learning [3, 11, 18, 33] that quantifies the difference between the logit corresponding to the assigned ground truth label and the largest other logit.", + "bbox": [ + 496, + 184, + 893, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our SSL formulation, we redefine the concept of margins to pseudo-margins of unlabeled examples since no ground truth labels are available for the unlabeled data. Let $c$ be the pseudo-label (or the argmax of the prediction, i.e., $\\hat{p}_{\\theta}(y|\\pi(\\hat{x}))$ ) at iteration $t$ on unlabeled example $\\hat{x}$ after applying weak augmentations. We define the pseudo-margin (PM) of $\\hat{x}$ with respect to pseudo-label $c$ at iteration $t$ as follows:", + "bbox": [ + 496, + 396, + 893, + 501 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P M} _ {c} ^ {t} (\\hat {x}) = z _ {c} - \\max _ {c! = i} (z _ {i}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 515, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{c}$ is the logit corresponding to the assigned pseudo-label $c$ and $\\max_{c! = i}(z_i)$ is the largest other logit corresponding to a label $i$ different from $c$ . To monitor the model's predictions on $\\hat{x}$ with respect to pseudo-label $c$ from the beginning of training to iteration $t$ , we average all the margins with respect to $c$ from the first iteration until $t$ and obtain the average pseudo-margin (APM) as follows:", + "bbox": [ + 496, + 539, + 893, + 645 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A P M} _ {c} ^ {t} (\\hat {x}) = \\frac {1}{t} \\sum_ {j = 1} ^ {t} \\mathrm {P M} _ {c} ^ {j} (\\hat {x}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 656, + 890, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $c$ acts as the \"ground truth\" label for the APM calculation. Note that if at a prior iteration $t'$ , the assigned pseudo-label is different from $c$ (say $c'$ ), then the APM calculation at iteration $t'$ is done with respect to $c'$ (by averaging all margins with respect to $c'$ from 1 to $t'$ ). In practice, we maintain a vector of pseudo-margins for all classes accumulated over the training iterations and dynamically retrieve the accumulated pseudo-margin value of the argmax class $c$ to obtain the $\\mathrm{APM}_c^t$ at iteration $t$ .", + "bbox": [ + 496, + 703, + 893, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, if $c$ is the pseudo-label of $\\hat{x}$ at iteration $t$ , then $\\mathrm{PM}_c^t$ with respect to class $c$ at iteration $t$ will be positive. In contrast, if the argmax of the model prediction on $\\hat{x}$ at a previous iteration $t' < t$ is different from $c$ , then $PM_c^{t'}$ at $t'$", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15775", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 MarginMatch" + ], + "code_body": "Require: Labeled data $L$ ; unlabeled data $U$ ; erroneous examples $E$ ; maximum number of iterations $T$ ; number of classes $C + 1$ ( $C$ original classes plus one virtual class of erroneous examples); $\\theta$ model; $\\pi$ weak augmentations; $\\Pi$ strong augmentations. \n1: Initialize the Average Pseudo-Margin (APM) threshold $\\gamma^1$ at the first iteration to a small value (e.g., $\\gamma^1 = -\\infty$ ). \n2: for $t = 1$ to $T$ do \n3: Estimate learning status $\\alpha_c$ (using Eq. 2) and calculate the class-wise flexible thresholds $\\mathcal{T}_c^t$ (using Eq. 3) for each class $c$ . \n4: while $U$ not exhausted do \n5: Labeled batch $L_b = \\{(x_1, y_1), \\dots, (x_B, y_B)\\}$ , unlabeled batch $U_b = \\{\\hat{x}_1, \\dots, \\hat{x}_{\\nu B}\\}$ , erroneous (or mislabeled) batch $E_b = \\{(\\tilde{x}_1, C + 1), \\dots, (\\tilde{x}_B, C + 1)\\}$ \n6: for $x \\in U_b \\cup E_b$ do \n7: Compute logits $z_c$ for each class $c$ after applying weak augmentations when $x \\in U_b$ and strong augmentations when $x \\in E_b$ . \n8: Calculate pseudo-margin $PM_c^t$ (using Eq. 5) and update Average $PM_c^t$ (using Eq. 6) for each $c = 1$ to $C + 1$ . \n9: end for \n10: Minimize $\\mathcal{L} = \\mathcal{L}_s + \\lambda (\\mathcal{L}_u + \\mathcal{L}_e)$ \n11: $\\mathcal{L}_s = \\frac{1}{B} \\sum_{i=1}^{B} H(y_i, p_\\theta(y|\\pi(x_i)))$ \n12: $\\mathcal{L}_u = \\sum_{i=1}^{\\nu B} \\mathbb{1}(\\mathrm{AM}_{p_\\theta(y|\\pi(\\hat{x}_i))}^t (\\hat{x}_i) > \\gamma^t) \\times \\mathbb{1}(\\max(p_\\theta(y|\\pi(\\hat{x}_i))) > T_{p_\\theta(y|\\pi(\\hat{x}_i))}^t) \\times H(\\hat{p}_\\theta(y|\\pi(\\hat{x}_i)), p_\\theta(y|\\Pi(\\hat{x}_i)))$ \n13: $\\mathcal{L}_e = \\sum_{i=1}^{B} H(C + 1, p_\\theta(y|\\Pi(\\tilde{x}_i)))$ \n14: end while \n15: Update $\\gamma^{t+1}$ as the $95^{th}$ percentile erroneous sample $APM_{C+1}^t$ . \n16: end for", + "bbox": [ + 76, + 107, + 893, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with respect to $c$ will be negative. Therefore, if over the iterations, the model predictions do not agree frequently with the pseudo-label $c$ from iteration $t$ and the model fluctuates significantly between iterations on the predicted label, the APM for class $c$ will have a low, likely negative value. Similarly, if the model is highly uncertain of the class of $\\hat{x}$ (reflected in a high entropy of the class probability distribution), the APM for class $c$ will have a low value. These capture the characteristics of unlabeled examples or of those harmful for training. Motivated by these observations, MarginMatch leverages the APM of the assigned pseudo-label $c$ and compares it with an APM threshold to mask out pseudo-labeled examples with low APMs. Formally, the unlabeled loss in MarginMatch is:", + "bbox": [ + 75, + 407, + 473, + 603 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} \\left(\\mathrm {A M} _ {\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i}))} ^ {t} (\\hat {x} _ {i}) > \\gamma^ {t}\\right) \\quad \\times \\\\ \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi (\\hat {x} _ {i})\\right)\\right) > \\mathcal {T} _ {\\hat {p} _ {\\theta} \\left(y \\mid \\pi (\\hat {x} _ {i})\\right)} ^ {t}\\right) \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 614, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma^t$ is the APM threshold at iteration $t$ , estimated as explained below, and $\\mathcal{T}_{\\hat{p}_0(y|\\pi(\\hat{x}_i))}^t$ is the flexible threshold estimated as in FlexMatch [49]. To train our model, we adopt the best practices [41, 49] and optimize the weighted combination of the supervised and unsupervised losses:", + "bbox": [ + 75, + 713, + 468, + 789 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda \\mathcal {L} _ {u} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 810, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the supervised loss is given by:", + "bbox": [ + 76, + 840, + 328, + 856 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} H \\left(y _ {i}, p _ {\\theta} (y \\mid \\pi \\left(x _ {i}\\right))\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 863, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Average Pseudo-Margin Threshold Estimation Inspired by Pleiss et al. [33], we propose to estimate the average pseudo-margin threshold $\\gamma^t$ by analyzing the training dynamics of a special category of unlabeled examples, which we force to be erroneous or mislabeled examples. That is, to create the sample of erroneous examples $E$ , we randomly sample a subset of unlabeled examples from $U$ that we assign to an inexistent (or virtual) class $C + 1$ at the beginning of the training process and remove them from $U$ . The purpose of these erroneous examples is to mimic the training dynamics of incorrectly pseudo-labeled (unlabeled) examples and use them as proxy to estimate the cutoff of (potentially) mislabeled pseudo-labels. Since the examples in $E$ should belong to one of the $C$ original classes, assigning them to the inexistent class $C + 1$ makes them by definition mislabeled (see Appendix A for additional insights into this virtual class). As with all unlabeled examples from $U$ , we compute $APM_{C + 1}^t$ for the special category of erroneous examples from $E$ , but unlike the unlabeled examples from $U$ , the erroneous ones from $E$ have a fixed class $C + 1$ . To mimic the training dynamics of unlabeled examples from $U$ , we use strong augmentations to compute the loss of the erroneous examples from $E$ . That is, given a batch $E_b$ of $B$ erroneous examples, the erroneous sample loss becomes:", + "bbox": [ + 496, + 407, + 893, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} = \\sum_ {i = 1} ^ {B} H (C + 1, p _ {\\theta} (y | \\Pi (\\tilde {x} _ {i}))) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 773, + 890, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At iteration $t$ , we use the APMs of the erroneous examples to choose the APM threshold $\\gamma^t$ . We set $\\gamma^t$ as the APM of the $95^{th}$ percentile erroneous sample. The total loss becomes:", + "bbox": [ + 496, + 816, + 890, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda \\left(\\mathcal {L} _ {u} + \\mathcal {L} _ {e}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 864, + 890, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our full MarginMatch algorithm is shown in Algorithm 1.", + "bbox": [ + 500, + 885, + 883, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15776", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bee4bbdb7b062ce377dc54504dbd9c3d1ed8f7cb422e392703999935bf2f3d8d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10
#Labels/Class425400425100425100425100
Pseuso-Labeling74.610.2646.492.2015.080.1987.450.8557.740.2836.550.2464.615.6025.212.039.400.3274.680.9955.452.4332.640.71
UDA10.793.755.320.064.410.0748.951.5929.430.2123.870.235.344.274.260.391.950.0137.828.449.811.156.810.17
MixMatch45.242.1512.761.147.130.3462.152.1741.511.1928.160.2446.181.783.980.173.501.1334.151.548.950.3210.410.73
ReMixMatch5.270.194.850.134.040.1247.150.7627.140.2323.780.124.230.313.180.041.940.0631.510.758.540.486.190.24
FixMatch7.80.284.910.054.250.0848.210.8229.450.1622.890.123.971.183.131.031.970.0338.434.1410.451.046.430.33
FlexMatch5.040.065.040.094.190.0139.991.6226.960.0822.440.158.193.207.782.556.720.3029.151.328.230.135.770.12
MarginMatch4.910.074.730.123.980.0236.971.3223.710.1321.390.123.751.203.141.171.930.0125.373.587.310.355.520.15
", + "bbox": [ + 80, + 88, + 893, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Test error rates on CIFAR-10, CIFAR-100, SVHN, and STL-10 datasets. Best results are shown in blue.", + "bbox": [ + 148, + 215, + 818, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exponential Moving Average of Pseudo-Margins The current definition of APM weighs the pseudo-margin at iteration $t$ identical to the pseudo-margin at a much earlier iteration $p$ ( $t >> p$ ). This is problematic since very old pseudo-margins eventually become deprecated (especially due to the large number of iterations through unlabeled data in consistency training ( $\\sim 9K$ ), and hence, the old margins are no longer indicative of the current learning status of the model. To this end, instead of averaging all pseudo-margins (from the beginning of training to the current iteration), we propose to use an exponential moving average to place more importance on recent iterations. Formally, APM becomes:", + "bbox": [ + 75, + 256, + 472, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A P M} _ {c} ^ {t} (\\hat {x}) = \\mathrm {P M} _ {c} ^ {t} (\\hat {x}) * \\frac {\\delta}{1 + t} + \\mathrm {A P M} _ {c} ^ {t - 1} (\\hat {x}) * \\left(1 - \\frac {\\delta}{1 + t}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 441, + 468, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We set the smoothing parameter $\\delta$ to 0.997 in experiments.", + "bbox": [ + 76, + 484, + 465, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 76, + 512, + 209, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the performance of our MarginMatch on a wide range of SSL benchmark datasets. Specifically, we perform experiments with various numbers of labeled examples on CIFAR-10, CIFAR-100 [21], SVHN [31], STL-10 [8], ImageNet [10], and WebVision [26]. For smaller scale datasets such as CIFAR-10, CIFAR-100, SVHN, and STL-10 we randomly sample a small number of labeled examples per class (ranging from 4 labels per class to 400 labels per class) and treat them as labeled data, whereas the remaining labeled examples are treated as unlabeled data, except for STL-10 [8], which provides its own set of unlabeled examples. On ImageNet and WebVision, we use $\\sim 10\\%$ of the available labeled examples as labeled data, with the remaining $(90\\%)$ being treated as unlabeled data. In all our experiments, we sample $5\\%$ of the unlabeled data and place it in the set of erroneous examples.", + "bbox": [ + 75, + 537, + 472, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We report the mean and standard deviation of error rates from five runs with different parameter initializations. Similar to FixMatch [41], we use Wide Residual Networks [48]: WRN-28-2 for CIFAR-10 and SVHN; WRN-28-8 for CIFAR-100; and WRN-37-2 for STL-10. We use ResNet-50 [14] for both ImageNet and WebVision.", + "bbox": [ + 75, + 780, + 470, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we adopt the same hyperparameters as FixMatch [41]. Specifically, we use stochastic gradient", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "descent (SGD) with a momentum of 0.9. We start with a learning rate of 0.03 and employ a cosine learning rate; at iteration $k$ , our learning rate is $\\eta(k) = \\cos\\left(\\frac{7k\\pi}{16K}\\right)$ , where $K$ is the maximum number of iterations and is set to $2^{20}$ . We also leverage the same data augmentations as in FlexMatch [49]. Specifically, for weak augmentations we employ a standard flip-and-shift augmentation and use RandAugment [9] for strong augmentations. We set the batch size $B = 64$ , ratio of unlabeled to labeled data in a batch to $\\nu = 7$ and weigh the supervised and unsupervised losses equally (i.e., $\\lambda = 1$ ). We set our initial confidence threshold to $\\tau = 0.95$ and our average pseudo-margin (APM) threshold to the APM of the $95^{th}$ percentile threshold sample. To report the error rates, we compare all the approaches using the model at the end of training as in FixMatch [41].", + "bbox": [ + 496, + 256, + 893, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. CIFAR-10, CIFAR-100, SVHN, and STL-10", + "text_level": 1, + "bbox": [ + 498, + 496, + 872, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare MarginMatch against strong baselines and prior works: Pseudo-Labeling [1], Unsupervised Data Augmentation (UDA) [46], MixMatch [5], ReMixMatch [4], FixMatch [41], and FlexMatch [49]. We show in Table 1 the error rates obtained by our MarginMatch and the baselines on the CIFAR-10, CIFAR-100, SVHN and STL-10 datasets. First, we observe that our approach improves the performance on both CIFAR-10 and CIFAR-100. On CIFAR-10, MarginMatch improves performance in all data regimes upon FlexMatch [49], which is the current state-of-the-art, while maintaining a good error rate standard deviation. On CIFAR-100, which is significantly more challenging than CIFAR-10, we observe that MarginMatch bring substantially larger improvements. Notably, we see $3.02\\%$ improvement over FlexMatch in error rate using only 4 labels per class, and $3.25\\%$ improvement using 25 examples per class. These results on CIFAR-100 emphasize the effectiveness of MarginMatch, which performs well on a more challenging dataset.", + "bbox": [ + 496, + 521, + 893, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On SVHN, our approach performs better than FixMatch using 4 labels per class and performs similarly with FixMatch using 25 and 100 labels per class. However, on this dataset, MarginMatch performs much better compared with FlexMatch. For example, MarginMatch achieves $3.75\\%$ error rate using 4 labels per class, whereas FlexMatch obtains an error rate of $8.19\\%$ with the same labels per class,", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15777", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3863b902bc94de3c24819d9343ec3c7e0f43b2934fb2ad8b2d6946347746db10.jpg", + "image_caption": [ + "Figure 2. Convergence speed of MarginMatch against FixMatch and FlexMatch with 4 labels per class." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 279, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0f4a2e72ba307d2e520de4a90e1334b6cb7289c8ab771c2357bf833a851585e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 480, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67c934250c290cbd95125a37db084ce3f4ae34da39c9f55bcc23d631dbdd9db4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 88, + 679, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b5dee9f559994b4f23f9e72ebdbe3369039e06c5226b62e9377d21b528f02ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 88, + 880, + 222 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/42609ebd2faf187629aa2e0f8b1c46db7473ac70cb449ee54164b890369bbf18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetImageNetWebVision
TOP-1TOP-5TOP-1TOP-5
Supervised48.3925.4949.5826.78
FixMatch43.6621.8044.7622.65
FlexMatch42.0219.4943.8722.07
MarginMatch41.0518.2843.0821.13
", + "bbox": [ + 119, + 272, + 426, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "yielding an improvement of MarginMatch of $4.44\\%$ over FlexMatch. We hypothesize that the low performance of FlexMatch is due to its limitation in handling unbalanced class distributions [49]. On STL-10, MarginMatch as well outperforms all the other approaches both in error rates and error rate standard deviation. Notably, on this dataset, our approach pushes the performance of FlexMatch by $3.78\\%$ in error rate using only 4 labels per class and by $0.92\\%$ using 25 labels per class.", + "bbox": [ + 75, + 414, + 468, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we compare MarginMatch with FixMatch and FlexMatch in terms of convergence speed in the extremely labelscarcse setting of 4 labels per class and show these results in Figure 2. Notably, we observe that MarginMatch has a similar convergence speed (or even better on CIFAR-100) compared with FlexMatch while achieving a lower test error rate than FlexMatch on all datasets with 4 labels per class (see Table 1). Even more strikingly, compared with FixMatch, MarginMatch has a much superior convergence speed for a much better test error rate with 4 labels per class. This is because the rigid thresholds in FixMatch allow access only to a small amount of unlabeled data for training at each iteration and it takes a lot longer for the model to train.", + "bbox": [ + 75, + 551, + 470, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. ImageNet and WebVision", + "text_level": 1, + "bbox": [ + 76, + 756, + 308, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To showcase the effectiveness of our approach in a large-scale setup, we test our MarginMatch on ImageNet [10] and WebVision [26] using $10\\%$ labeled examples in total. We show the results obtained in Table 2. We observe that our MarginMatch outperforms FixMatch and FlexMatch on both datasets. It is worth noting that large-scale self-supervised approaches such as SimCLR [7] achieve high performance on ImageNet but at a much higher computational cost.", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/daa1afef547adca559f5f36b3a3633309ab3af22d5f937a775fb2234f7930de6.jpg", + "table_caption": [ + "Table 2. Test error rates on the ImageNet and WebVision datasets. Best results are shown in blue." + ], + "table_footnote": [], + "table_body": "
δ0.950.990.9950.9970.9991
ERR RATE38.1338.0537.9237.9139.1239.72
", + "bbox": [ + 524, + 273, + 870, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Error rates obtained on CIFAR-100 with four examples per class and various smoothing values $\\delta$ . Best result is in blue.", + "bbox": [ + 498, + 320, + 890, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MarginMatch outperforms other SSL methods using the same ResNet-50 architecture at the same computational cost. We emphasize MarginMatch is most successful and relevant in low data regimes on smaller datasets.", + "bbox": [ + 498, + 378, + 893, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 454, + 650, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Exponential Moving Average Smoothing for APM Computation In our approach, we employ an exponential moving average (EMA) of the pseudo-margin values with a smoothing value of $\\delta = 0.997$ to compute the APM. We now analyze how our approach performs with different EMA smoothing values or with no EMA at all. Table 3 shows these results on CIFAR-100 with 4 labels per class. First, we observe that employing a simple average of pseudo-margin values for the APM computation (i.e., $\\delta = 1$ ) performs extremely poorly, obtaining a $39.72\\%$ error rate. This result emphasizes that margins eventually become deprecated and it is essential to scale them down in time. Using a low smoothing factor of $\\delta = 0.95$ is not effective either, denoting that abruptly forgetting margin values does not work either. Our chosen $\\delta = 0.997$ strikes a balance between the two by eliminating the harmful effects of very old margins while keeping track of a good amount of previous estimates (e.g., a margin value computed 200 epochs previously is scaled down by 0.55, while a margin value computed 1000 epochs previously is scaled by 0.05).", + "bbox": [ + 496, + 484, + 892, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pseudo-Margin vs. Other Measures for Pseudo-Label Correctness Our MarginMatch monitors the pseudomargins of a model's predictions across training iterations to ensure the quality of pseudo-labels. However, other measures such as confidence or entropy exist that can assess the pseudo-label correctness. Hence, we perform an ablation where we replace the pseudo-margins in our MarginMatch", + "bbox": [ + 496, + 794, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15778", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a93f56de069c4074fd5f97f682bcd492dc09a2459b5f7c6cfacef153b7e8c3d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10
#Labels/Class425400425100425100425100
Avg Confidence23.872.7314.211.377.540.7841.232.1531.491.4824.112.368.994.276.540.394.730.0131.678.4414.871.157.590.17
Avg Entropy8.580.416.180.155.850.1245.100.9126.021.1122.130.2515.691.2512.740.789.330.0529.543.5110.631.3510.840.47
Avg Margin7.250.295.380.764.730.0939.721.5225.210.5223.180.1718.451.3611.290.938.400.0428.454.289.341.347.590.21
EMA Confidence4.910.454.740.093.990.0638.670.7425.610.1221.480.173.840.233.250.031.930.0925.90.817.60.425.740.57
EMA Entropy6.40.438.340.124.210.0941.630.7636.840.1322.520.073.811.263.170.872.140.0427.214.058.281.016.790.27
EMA Margin4.910.074.730.123.980.0236.971.3223.710.1321.390.123.751.203.141.171.930.0125.373.587.310.355.520.15
", + "bbox": [ + 81, + 88, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Test error rates comparing pseudo-margin with confidence and entropy. Best results are shown in blue.", + "bbox": [ + 153, + 207, + 815, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d7a6f0b6f27210d2b1e6974fc18acd3a020cb78d91155ca3f6a1f383e9031af.jpg", + "image_caption": [ + "Figure 3. Mask rate and impurity on CIFAR-100 with 4 labeled examples per class." + ], + "image_footnote": [], + "bbox": [ + 81, + 232, + 272, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da786527f8c4ef697b2a118d71e40212692d47c7e16ad1ca6e0824b5a67174a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 232, + 467, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with average confidence and entropy and compare their performance. Specifically, we design the following approaches: 1) Avg Confidence monitors the confidence of the prediction for each unlabeled example and takes the average over the training iterations; 2) Avg Entropy monitors the entropy of the class probability distribution for each unlabeled example and takes the average across the training iterations. In addition, we also consider 3) EMA Confidence and 4) EMA Entropy which are similar to Avg Confidence and Avg Entropy, respectively, but use an exponential moving average (EMA) instead of the simple averaging. The estimation of the threshold for each of these approaches is done in a similar manner as for pseudo-margins, using erroneous samples and considering the value of the 95th percentile erroneous sample as the threshold.", + "bbox": [ + 73, + 412, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the results obtained using these approaches in Table 4. First, we observe that all measures (pseudo-margin, confidence and entropy) with EMA perform better than their counterpart with simple averaging. Second, EMA Margin achieves the lowest test error rates compared with EMA Confidence and EMA Entropy. Thus, we conclude that pseudo-margins provide an excellent measure for pseudolabel correctness. See Appendix B for some additional insights into why EMA Margin outperforms EMA confidence and entropy.", + "bbox": [ + 75, + 638, + 470, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 76, + 805, + 173, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Mask Rate and Impurity", + "text_level": 1, + "bbox": [ + 76, + 830, + 303, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now contrast MarginMatch with FixMatch and FlexMatch in terms of the quality of pseudo-labels using two metrics: mask rate and impurity and show these results in", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 3, respectively, using CIFAR-100 with 4 labels per class. Mask rate is defined as the fraction of pseudo-labeled examples that do not participate in the training at epoch $t$ due to confidence masking or pseudo-margin masking (or both). Impurity in contrast is defined as the fraction of pseudolabeled examples that do participate in the training at epoch $t$ but with a wrong label. An effective SSL model minimizes both metrics: a low mask rate indicates that the model has access to more unlabeled examples during training (otherwise a low percentage and less diverse set of unlabeled examples are seen during training) while low impurity indicates that the pseudo-labels of these examples are of high quality. Note that we can compute impurity on these two datasets because our unlabeled data comes from the labeled training set of each of these datasets (thus we compare the pseudo-labels against the gold labels of each dataset).", + "bbox": [ + 496, + 233, + 893, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be seen from the figures, FixMatch has a significantly larger mask rate due to the rigid confidence threshold set to a high value of 0.95. In contrast, FlexMatch lowers the mask rate by $5\\%$ with the introduction of flexible thresholds, but has a much higher impurity compared with FixMatch. Notably, our MarginMatch has only a slightly higher mask rate compared with FlexMatch and at the same time achieves a much lower impurity than FlexMatch and even FixMatch despite that FixMatch employs a very high confidence threshold. These results show that MarginMatch that enforces an additional measure for pseudo-labeled data quality maintains a low mask rate without compromising the quality of the pseudo-labels (i.e., low mask rate and low impurity).", + "bbox": [ + 496, + 476, + 893, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Anecdotal Evidence", + "text_level": 1, + "bbox": [ + 500, + 681, + 689, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show in Figure 4 anecdotal evidence of the effectiveness of MarginMatch. To this end, we extract two bird images from our unlabeled portion of CIFAR-10 [21] of various learning difficulties that resemble characteristics of plane images (e.g., the background). The top part of the figure illustrates the progression over the training iterations of the confidence and the confidence thresholds of FlexMatch for the classes bird and plane, whereas the bottom part of the figure illustrates the progression of the APM threshold of MarginMatch along with its APMs of bird and plane classes over the training iterations. In the rightmost image, for MarginMatch we can observe that the APM of the bird class becomes stronger and stronger as the training progresses and", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15779", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09309a144bfd5ba04043d61348a201f6da75b5cbced2a082c1713cf8cca3df75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 92, + 472, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88b8e4a329bb9009cfef05f34bd5bd8d478b95e290408b600afa6990ba38710c.jpg", + "image_caption": [ + "Figure 4. Confidence thresholding vs. APM Thresholding on two images from the CIFAR-10 dataset." + ], + "image_footnote": [], + "bbox": [ + 81, + 234, + 468, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eventually exceeds the APM threshold of MarginMatch, and hence, the image is included in the training until the end with the correct bird label. Interestingly, for the same image, in FlexMatch the confidence for the bird class is very close to the bird confidence threshold and eventually falls below this threshold and exits the training set. In contrast, the leftmost image is significantly more challenging than the rightmost image since it is more similar to plane images, which makes it an easily confusable example. Here, we observe that the confidence of FlexMatch exceeds the flexible threshold with the incorrect argmax class plane starting from iteration 3000. Moreover, Flexmatch continues to use this image with the wrong plane label for the remaining of the training process. Critically, in MarginMatch the APM value for the plane class does not exceed the APM threshold, and the model eventually learns to classify this image correctly and includes it in training with the correct bird pseudo-label.", + "bbox": [ + 75, + 407, + 470, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Related Work", + "text_level": 1, + "bbox": [ + 76, + 678, + 218, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we focus on various SSL approaches that our MarginMatch directly builds upon although there are other SSL techniques that are not presented in this review, such as approaches based on generative models [17, 24], graph-based approaches [19, 20] and robust SSL [32, 37] (see Appendix C for a comparison between MarginMatch and Robust SSL approaches).", + "bbox": [ + 75, + 703, + 468, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-training [28, 36, 40, 47] is a popular SSL method where the predictions of a model on unlabeled data are used as artificial labels to train against. Noisy student training [47] is a popular self-training approach that also leverages knowledge distillation [16] and iteratively jointly trains two models in a teacher-student framework. Noisy student uses a", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "larger model size and noised inputs, exposing the student to more difficult learning environments, leading to an increased performance compared to the teacher.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pseudo-labeling is a variant of self-training where these predictions are sharpened to obtain hard labels [25]. The use of hard labels can be seen as a means of entropy minimization [12] and nowadays is a valuable component in most successful SSL approaches [4, 5, 49]. These hard labels are usually used along a confidence threshold, where unconfident unlabeled examples are completely disregarded (e.g., [5]) to avoid using noisy pseudo-labels. Recently, approaches such as Curriculum Labeling (CL) [6] or FlexMatch [49] started to explore curriculum learning in the SSL context. CL proposes a self-pacing strategy of identifying easy and hard examples to ensure that the model first uses easy and progressively moves towards harder examples. Similarly, MarginMatch uses curriculum learning and pseudo-labeling, but the focus of our approach is placed on producing better thresholds for assessing the quality of pseudo-labels.", + "bbox": [ + 496, + 136, + 893, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consistency regularization [2] is a method that applies random perturbations when generating the artificial label, such as data augmentation [4, 5, 41], dropout [39], or adversarial perturbations [29]. Current state-of-the-art approaches [41, 49] exploit a combination of weak and strong data augmentations, which were shown to be extremely beneficial in SSL. The most popular strong augmentations used in the SSL literature are RandAugment [9] and CTAugment [4]. The approaches based on these methods first generate a hard label using pseudo-labeling on a weakly augmented image (i.e., using a low noise transformation such as a flip-and-shift augmentation), then optimize the predictions of the model on a strongly augmented version of the same image towards this hard label. Similar to these approaches, MarginMatch uses the same combination of weak and strong data augmentations.", + "bbox": [ + 496, + 378, + 893, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 633, + 617, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel semi-supervised learning method that improves the pseudo-label quality using training dynamics. Our new method is lightweight and achieves state-of-the-art performance on four computer vision SSL datasets in low data regimes and on two large-scale benchmarks. MarginMatch takes into consideration not only a flexible confidence threshold to account for the difficulty of each class, but also a measure of quality for each unlabeled example using training dynamics. In addition, MarginMatch is a general approach that can be leveraged in most SSL frameworks and we hope that it can attract future research in analyzing the effectiveness of SSL approaches focused on data quality. As future work, we aim to further explore our method in settings when there is a mismatch between the labeled and unlabeled data distributions (i.e., making use of out-of-domain unlabeled data).", + "bbox": [ + 496, + 657, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15780", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. CoRR, abs/1908.02983, 2019. 1, 5", + "[2] Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2, 8", + "[3] Peter L. Bartlett, Dylan J. Foster, and Matus Telgarsky. Spectrally-normalized margin bounds for neural networks. CoRR, abs/1706.08498, 2017. 1, 3", + "[4] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 1, 5, 8", + "[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1, 5, 8", + "[6] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In AAAI, 2021. 8", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6", + "[8] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 2, 5", + "[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 5, 8", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 5, 6", + "[11] Gamaleldin Fathy Elsayed, Dilip Krishnan, Hossein Mobahi, Kevin Regan, and Samy Bengio. Large margin deep networks for classification. 2018. 1, 3", + "[12] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In L. Saul, Y. Weiss, and L. Bottou, editors, Advances in Neural Information Processing Systems, volume 17. MIT Press, 2004. 8", + "[13] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1321-1330. PMLR, 06-11 Aug 2017. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 5", + "[15] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. 1", + "[16] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 8", + "[17] Geoffrey E Hinton and Russ R Salakhutdinov. Using deep belief nets to learn covariance kernels for gaussian processes. In J. Platt, D. Koller, Y. Singer, and S. Roweis, editors, Advances in Neural Information Processing Systems, volume 20. Curran Associates, Inc., 2007. 8", + "[18] Yiding Jiang, Dilip Krishnan, Hossein Mobahi, and Samy Bengio. Predicting the generalization gap in deep networks with margin distributions, 2018. 1, 3", + "[19] Thorsten Joachims. Transductive inference for text classification using support vector machines. In Proceedings of the Sixteenth International Conference on Machine Learning, ICML '99, page 200-209, San Francisco, CA, USA, 1999. Morgan Kaufmann Publishers Inc. 8", + "[20] Thorsten Joachims. Transductive learning via spectral graph partitioning. In Proceedings of the Twentieth International Conference on International Conference on Machine Learning, ICML'03, page 290-297. AAAI Press, 2003. 8", + "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5, 7", + "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, NIPS'12, page 1097-1105, Red Hook, NY, USA, 2012. Curran Associates Inc. 1", + "[23] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR (Poster). OpenReview.net, 2017. 1, 2", + "[24] J.A. Lasserre, C.M. Bishop, and T.P. Minka. Principled hybrids of generative and discriminative models. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 1, pages 87-94, 2006. 8", + "[25] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 2, 8", + "[26] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. CoRR, abs/1708.02862, 2017. 2, 5, 6", + "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pages 181-196, 2018. 1" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15781", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152-159, New York City, USA, June 2006. Association for Computational Linguistics. 8", + "[29] Takeru Miyato, Andrew M Dai, and Ian Goodfellow. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725, 2016. 8", + "[30] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1", + "[31] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 2, 5", + "[32] Jongjin Park, Sukmin Yun, Jongheon Jeong, and Jinwoo Shin. Opencos: Contrastive semi-supervised learning for handling open-set unlabeled data. In ECCV Workshops, 2021. 8", + "[33] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying mislabeled data using the area under the margin ranking. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 17044-17056. Curran Associates, Inc., 2020. 1, 3, 4", + "[34] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1", + "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683, 2019. 1", + "[36] Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. In 2005 Seventh IEEE Workshops on Applications of Computer Vision (WACV/MOTION'05) - Volume 1, volume 1, pages 29-36, 2005. 1, 8", + "[37] Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set semi-supervised learning with open-set consistency regularization. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 25956-25967. Curran Associates, Inc., 2021. 8", + "[38] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Mutual exclusivity loss for semi-supervised deep learning. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1908-1912. IEEE, 2016. 1", + "[39] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29:1163-1171, 2016. 1, 2, 8", + "[40] H. Scudder. Probability of error of some adaptive pattern-recognition machines. IEEE Transactions on Information Theory, 11(3):363-371, 1965. 8", + "[41] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk," + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 596-608. Curran Associates, Inc., 2020. 1, 2, 4, 5, 8", + "[42] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 1", + "[43] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 1", + "[44] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in neural information processing systems, pages 1195–1204, 2017. 1", + "[45] Vikas Verma, Kenji Kawaguchi, Alex Lamb, Juho Kannala, Arno Solin, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi-supervised learning. Neural Networks, 2021. 1", + "[46] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc., 2020. 1, 5", + "[47] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687-10698, 2020. 1, 8", + "[48] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks, 2017. 5", + "[49] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 18408-18419. Curran Associates, Inc., 2021. 1, 2, 3, 4, 5, 6, 8", + "[50] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. CoRR, abs/1611.03530, 2016. 3" + ], + "bbox": [ + 503, + 90, + 890, + 753 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15782", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_model.json b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..110b67adc6a7cabec22f2873044ac398c2135d60 --- /dev/null +++ b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.865, + 0.154 + ], + "angle": 0, + "content": "MarginMatch: Improving Semi-Supervised Learning with Pseudo-Margins" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.182, + 0.644, + 0.234 + ], + "angle": 0, + "content": "Tiberiu Sosea Cornelia Caragea University of Illinois Chicago tsosea2@uic.edu cornelia@uic.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.558 + ], + "angle": 0, + "content": "We introduce MarginMatch, a new SSL approach combining consistency regularization and pseudo-labeling, with its main novelty arising from the use of unlabeled data training dynamics to measure pseudo-label quality. Instead of using only the model's confidence on an unlabeled example at an arbitrary iteration to decide if the example should be masked or not, MarginMatch also analyzes the behavior of the model on the pseudo-labeled examples as the training progresses, to ensure low quality predictions are masked out. MarginMatch brings substantial improvements on four vision benchmarks in low data regimes and on two large-scale datasets, emphasizing the importance of enforcing high-quality pseudo-labels. Notably, we obtain an improvement in error rate over the state-of-the-art of \\(3.25\\%\\) on CIFAR-100 with only 25 labels per class and of \\(3.78\\%\\) on STL-10 using as few as 4 labels per class. We make our code available at https://github.com/tsosea2/MarginMatch." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.572, + 0.21, + 0.587 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Deep learning models have seen tremendous success in many vision tasks [14, 22, 27, 42, 43]. This success can be attributed to their scalability, being able to produce better results when they are trained on large datasets in a supervised fashion [15, 27, 34, 35, 43, 47]. Unfortunately, large labeled datasets annotated for various tasks and domains are difficult to acquire and demand considerable annotation effort or domain expertise. Semi-supervised learning (SSL) is a powerful approach that mitigates the requirement for large labeled datasets by effectively making use of information from unlabeled data, and thus, has been studied extensively in vision [4, 5, 23, 25, 30, 36, 38, 39, 44-46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Recent SSL approaches integrate two important components: consistency regularization [46, 49] and pseudolabeling [25]. Consistency regularization works on the assumption that a model should output similar predictions when fed perturbed versions of the same image, whereas pseudo-labeling uses the model's predictions of unlabeled examples as labels to train against. For example, Sohn et al. [41] introduced FixMatch that combines consistency reg" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.896, + 0.694 + ], + "angle": 0, + "content": "ularization on weak and strong augmentations with pseudolabeling. FixMatch relies heavily on a high-confidence threshold to compute the unsupervised loss, disregarding any pseudo-labels whose confidence falls below this threshold. While training using only high-confidence pseudo-labels has shown to consistently reduce the confirmation bias [1], this rigid threshold allows access only to a small amount of unlabeled data for training, and thus, ignores a considerable amount of unlabeled examples for which the model's predictions do not exceed the confidence threshold. More recently, Zhang et al. [49] introduced FlexMatch that relaxes the rigid confidence threshold in FixMatch to account for the model's learning status of each class in that it adaptively scales down the threshold for a class to encourage the model to learn from more examples from that class. The flexible thresholds in FlexMatch allow the model to have access to a much larger and diverse set of unlabeled data to learn from, but lowering the thresholds can lead to the introduction of wrong pseudo-labels, which are extremely harmful for generalization. Interestingly, even when the high-confidence threshold is used in FixMatch can result in wrong pseudo-labels. See Figure 1 for incorrect pseudo-labels detected in the training set after we apply FixMatch and FlexMatch on ImageNet. We posit that a drawback of FixMatch and FlexMatch and in general of any pseudo-labeling approach is that they use the confidence of the model only at the current iteration to enforce quality of pseudo-labels and completely ignore model's predictions at prior iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.897, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose MarginMatch, a new SSL approach that monitors the behavior of the model on the unlabeled examples as the training progresses, from the beginning of training until the current iteration, instead of using only the model's current belief about an unlabeled example (i.e., its confidence at the current iteration) to decide if the example should be masked or not. We estimate a pseudo-label's contribution to learning and generalization by introducing pseudo-margins of unlabeled examples averaged across training iterations. Pseudo-margins of unlabeled examples extend the margins from machine learning [3, 11, 18, 33] which provide a measure of confidence of the outputs of the model and capture the difference between the output for the correct" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15773" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.875, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.305, + 0.851, + 0.32 + ], + "angle": 0, + "content": "Figure 1. Incorrect pseudo-labels propagated until the end of the training process for FixMatch and FlexMatch on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.473, + 0.512 + ], + "angle": 0, + "content": "(gold) label and the other labels. In our case, the pseudomargins capture how much larger the assigned logit (the logit corresponding to the argmax of the model's prediction) is compared with all other logits at iteration \\( t \\). Similar to FlexMatch, in MarginMatch we take advantage of the flexible confidence thresholds to allow the model to learn from larger and more diverse sets of unlabeled examples, but unlike FlexMatch, we train the model itself to identify the characteristics of mislabeled pseudo-labels simply by monitoring the model's training dynamics on unlabeled data over the iterations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.473, + 0.724 + ], + "angle": 0, + "content": "We carry out comprehensive experiments using established SSL experimental setups on CIFAR-10, CIFAR-100 [21], SVHN [31], STL-10 [8], ImageNet [10], and WebVision [26]. Despite its simplicity, our findings indicate that MarginMatch produces improvements in performance over strong baselines and prior works on all datasets at no additional computational cost. Notably, compared to current state-of-the-art, on CIFAR-100 we see \\(3.02\\%\\) improvement in error rate using only 4 labels per class and \\(3.78\\%\\) improvement on STL-10 using the same extremely label-scarce setting of 4 labels per class. In addition, on ImageNet [10] and WebVision [26] we find that MarginMatch pushes the state-of-the-art error rates by \\(0.97\\%\\) on ImageNet and by \\(0.79\\%\\) on WebVision." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.725, + 0.316, + 0.738 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.744, + 0.47, + 0.866 + ], + "angle": 0, + "content": "1. We introduce a new SSL approach which we call MarginMatch that enforces high pseudo-label quality during training. Our approach allows access to a large set of unlabeled data to learn from (thus, incorporating more information from unlabeled data) and, at the same time, monitors the training dynamics of unlabeled data as training progresses to detect and filter out potentially incorrect pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "2. We show that MarginMatch outperforms existing works on six well-established computer vision benchmarks" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.744, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.346, + 0.895, + 0.392 + ], + "angle": 0, + "content": "showing larger improvements in error rates especially on challenging datasets, while achieving similar convergence performance (or better) than prior works." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.895, + 0.442 + ], + "angle": 0, + "content": "3. We perform a comprehensive analysis of our approach and indicate potential insights into why our Margin-Match substantially outperforms other SSL techniques." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.642, + 0.47 + ], + "angle": 0, + "content": "2. MarginMatch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.6 + ], + "angle": 0, + "content": "Notation Let \\( L = \\{(x_{1},y_{1}),\\dots,(x_{B},y_{B})\\} \\) be a batch of size \\( B \\) of labeled examples and \\( U = \\{\\hat{x}_1,\\dots,\\hat{x}_{\\nu B}\\} \\) be a batch of size \\( \\nu B \\) of unlabeled examples, where \\( \\nu \\) is the batch-wise ratio of unlabeled to labeled examples. Let \\( p_{\\theta}(y|x) \\) denote the class distribution produced by model \\( \\theta \\) on input image \\( x \\) and \\( \\hat{p}_{\\theta}(y|x) \\) denote the argmax of this distribution as a one-hot label. Let also \\( H(p,q) \\) denote the cross-entropy between two probability distributions \\( p \\) and \\( q \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.607, + 0.633, + 0.623 + ], + "angle": 0, + "content": "2.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.795 + ], + "angle": 0, + "content": "Consistency reg [39] is an important component in recent semi-supervised learning approaches and relies on the continuity assumption [2,23] that the model should output similar predictions on multiple perturbed versions of the same input \\( x \\). As mentioned above, examples of two such approaches are FixMatch [41] and FlexMatch [49] that use consistency regularization at their core combined with psedo-labeling. In psedo-labeling [25], a model itself is used to assign artificial labels for unlabeled data and only artificial labels whose largest class probability is above a predefined confidence threshold are used during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Specifically, FixMatch [41] predicts artificial labels for unlabeled examples using a weakly-augmented version of each unlabeled example and then employs the artificial labels as pseudo-labels to train against but this time using a strongly-augmented version of each unlabeled example. That is, FixMatch minimizes the following batch-wise consistency loss on unlabeled data:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15774" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.097, + 0.09, + 0.47, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right)\\right) > \\tau\\right) \\quad \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.162, + 0.47, + 0.238 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a confidence threshold, \\(\\pi\\) and \\(\\Pi\\) are weak and strong augmentations, respectively, and \\(\\mathbb{1}\\) is the indicator function. Therefore, the low-confidence examples (lower than \\(\\tau\\)) are completely ignored despite containing potentially useful information for model training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.471, + 0.412 + ], + "angle": 0, + "content": "FlexMatch [49] argues that using a fixed threshold \\(\\tau\\) to filter the unlabeled data ignores the learning difficulties of different classes, and thus, introduces class-dependent thresholds, which are obtained by adaptively scaling \\(\\tau\\) depending on the learning status of each class. FlexMatch assumes that a class with fewer examples above the fixed threshold \\(\\tau\\) has a greater learning difficulty, and hence, it adaptively lowers the threshold \\(\\tau\\) to encourage more training examples from this class to be learned. The learning status \\(\\alpha_{c}\\) for a class \\(c\\) is simply computed as the number of unlabeled examples that are predicted in class \\(c\\) and pass the fixed threshold \\(\\tau\\):" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.421, + 0.47, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {c} = \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right)\\right) > \\tau\\right) \\mathbb {1} \\left(\\hat {p} _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right) = c\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.518 + ], + "angle": 0, + "content": "where \\( n \\) is the total number of unlabeled examples. This learning effect is then normalized and used to obtain the class-dependent threshold for each class \\( c \\):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.528, + 0.47, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {c} = \\frac {\\alpha_ {c}}{\\max _ {c} \\left(\\alpha_ {c}\\right)} \\times \\tau \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.645 + ], + "angle": 0, + "content": "In practice, FlexMatch iteratively computes new thresholds after each complete pass through unlabeled data, hence we can parameterize \\(\\mathcal{T}_c\\) as \\(\\mathcal{T}_c^t\\), denoting the threshold obtained at iteration \\(t\\). The unlabeled loss is then obtained by plugging in the adaptive threshold \\(\\mathcal{T}_c^t\\) in Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.663, + 0.47, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} (\\max (p _ {\\theta} (y | \\pi (\\hat {x} _ {i}))) > \\mathcal {T} _ {\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i}))} ^ {t}) \\quad \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The aforementioned works use the confidence of the model solely at the current iteration to enforce quality of pseudo-labels. We believe this is not sufficient as it provides only a myopic view of the model's behavior (i.e., its confidence) on unlabeled data (at a single iteration) and may result in wrong pseudo-labels even when the confidence threshold is high enough (e.g., if the model is miscalibrated or overly-confident [13]). Figure 1 shows examples of images that are added to the training set with a wrong pseudo-label for both FixMatch and FlexMatch. These types of unlabeled examples, which are incorrectly pseudo-labeled and used" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "during training are particularly harmful for deep neural networks, which can attain zero training error on any dataset, even on randomly assigned labels [50], resulting in poor generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.162, + 0.808, + 0.178 + ], + "angle": 0, + "content": "2.2. Proposed Approach: MarginMatch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.185, + 0.895, + 0.395 + ], + "angle": 0, + "content": "We now introduce MarginMatch, our new SSL approach that uses the model's training dynamics on unlabeled data to improve pseudo-label data quality. Our approach leverages consistency regularization with weak and strong augmentations and pseudo-labeling, but instead of using only the model's current belief (i.e., its confidence at the current iteration) to decide if an unlabeled example should be used for training or not, our MarginMatch monitors the training dynamics of unlabeled data over the iterations by investigating the margins (a measure of confidence) of the outputs of the model [3]. The margin of a training example is a well established metric in machine learning [3, 11, 18, 33] that quantifies the difference between the logit corresponding to the assigned ground truth label and the largest other logit." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.894, + 0.502 + ], + "angle": 0, + "content": "In our SSL formulation, we redefine the concept of margins to pseudo-margins of unlabeled examples since no ground truth labels are available for the unlabeled data. Let \\( c \\) be the pseudo-label (or the argmax of the prediction, i.e., \\( \\hat{p}_{\\theta}(y|\\pi(\\hat{x})) \\)) at iteration \\( t \\) on unlabeled example \\( \\hat{x} \\) after applying weak augmentations. We define the pseudo-margin (PM) of \\( \\hat{x} \\) with respect to pseudo-label \\( c \\) at iteration \\( t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.516, + 0.892, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P M} _ {c} ^ {t} (\\hat {x}) = z _ {c} - \\max _ {c! = i} (z _ {i}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.894, + 0.646 + ], + "angle": 0, + "content": "where \\( z_{c} \\) is the logit corresponding to the assigned pseudo-label \\( c \\) and \\( \\max_{c! = i}(z_i) \\) is the largest other logit corresponding to a label \\( i \\) different from \\( c \\). To monitor the model's predictions on \\( \\hat{x} \\) with respect to pseudo-label \\( c \\) from the beginning of training to iteration \\( t \\), we average all the margins with respect to \\( c \\) from the first iteration until \\( t \\) and obtain the average pseudo-margin (APM) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.657, + 0.892, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A P M} _ {c} ^ {t} (\\hat {x}) = \\frac {1}{t} \\sum_ {j = 1} ^ {t} \\mathrm {P M} _ {c} ^ {j} (\\hat {x}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Here \\( c \\) acts as the \"ground truth\" label for the APM calculation. Note that if at a prior iteration \\( t' \\), the assigned pseudo-label is different from \\( c \\) (say \\( c' \\)), then the APM calculation at iteration \\( t' \\) is done with respect to \\( c' \\) (by averaging all margins with respect to \\( c' \\) from 1 to \\( t' \\)). In practice, we maintain a vector of pseudo-margins for all classes accumulated over the training iterations and dynamically retrieve the accumulated pseudo-margin value of the argmax class \\( c \\) to obtain the \\( \\mathrm{APM}_c^t \\) at iteration \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Intuitively, if \\( c \\) is the pseudo-label of \\( \\hat{x} \\) at iteration \\( t \\), then \\( \\mathrm{PM}_c^t \\) with respect to class \\( c \\) at iteration \\( t \\) will be positive. In contrast, if the argmax of the model prediction on \\( \\hat{x} \\) at a previous iteration \\( t' < t \\) is different from \\( c \\), then \\( PM_c^{t'} \\) at \\( t' \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15775" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.092, + 0.26, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 MarginMatch" + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.108, + 0.895, + 0.381 + ], + "angle": 0, + "content": "Require: Labeled data \\(L\\); unlabeled data \\(U\\); erroneous examples \\(E\\); maximum number of iterations \\(T\\); number of classes \\(C + 1\\) (\\(C\\) original classes plus one virtual class of erroneous examples); \\(\\theta\\) model; \\(\\pi\\) weak augmentations; \\(\\Pi\\) strong augmentations. \n1: Initialize the Average Pseudo-Margin (APM) threshold \\(\\gamma^1\\) at the first iteration to a small value (e.g., \\(\\gamma^1 = -\\infty\\)). \n2: for \\(t = 1\\) to \\(T\\) do \n3: Estimate learning status \\(\\alpha_c\\) (using Eq. 2) and calculate the class-wise flexible thresholds \\(\\mathcal{T}_c^t\\) (using Eq. 3) for each class \\(c\\). \n4: while \\(U\\) not exhausted do \n5: Labeled batch \\(L_b = \\{(x_1, y_1), \\dots, (x_B, y_B)\\}\\), unlabeled batch \\(U_b = \\{\\hat{x}_1, \\dots, \\hat{x}_{\\nu B}\\}\\), erroneous (or mislabeled) batch \\(E_b = \\{(\\tilde{x}_1, C + 1), \\dots, (\\tilde{x}_B, C + 1)\\}\\) \n6: for \\(x \\in U_b \\cup E_b\\) do \n7: Compute logits \\(z_c\\) for each class \\(c\\) after applying weak augmentations when \\(x \\in U_b\\) and strong augmentations when \\(x \\in E_b\\). \n8: Calculate pseudo-margin \\(PM_c^t\\) (using Eq. 5) and update Average \\(PM_c^t\\) (using Eq. 6) for each \\(c = 1\\) to \\(C + 1\\). \n9: end for \n10: Minimize \\(\\mathcal{L} = \\mathcal{L}_s + \\lambda (\\mathcal{L}_u + \\mathcal{L}_e)\\) \n11: \\(\\mathcal{L}_s = \\frac{1}{B} \\sum_{i=1}^{B} H(y_i, p_\\theta(y|\\pi(x_i)))\\) \n12: \\(\\mathcal{L}_u = \\sum_{i=1}^{\\nu B} \\mathbb{1}(\\mathrm{AM}_{p_\\theta(y|\\pi(\\hat{x}_i))}^t (\\hat{x}_i) > \\gamma^t) \\times \\mathbb{1}(\\max(p_\\theta(y|\\pi(\\hat{x}_i))) > T_{p_\\theta(y|\\pi(\\hat{x}_i))}^t) \\times H(\\hat{p}_\\theta(y|\\pi(\\hat{x}_i)), p_\\theta(y|\\Pi(\\hat{x}_i)))\\) \n13: \\(\\mathcal{L}_e = \\sum_{i=1}^{B} H(C + 1, p_\\theta(y|\\Pi(\\tilde{x}_i)))\\) \n14: end while \n15: Update \\(\\gamma^{t+1}\\) as the \\(95^{th}\\) percentile erroneous sample \\(APM_{C+1}^t\\). \n16: end for" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.474, + 0.604 + ], + "angle": 0, + "content": "with respect to \\( c \\) will be negative. Therefore, if over the iterations, the model predictions do not agree frequently with the pseudo-label \\( c \\) from iteration \\( t \\) and the model fluctuates significantly between iterations on the predicted label, the APM for class \\( c \\) will have a low, likely negative value. Similarly, if the model is highly uncertain of the class of \\( \\hat{x} \\) (reflected in a high entropy of the class probability distribution), the APM for class \\( c \\) will have a low value. These capture the characteristics of unlabeled examples or of those harmful for training. Motivated by these observations, MarginMatch leverages the APM of the assigned pseudo-label \\( c \\) and compares it with an APM threshold to mask out pseudo-labeled examples with low APMs. Formally, the unlabeled loss in MarginMatch is:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.615, + 0.47, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} \\left(\\mathrm {A M} _ {\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i}))} ^ {t} (\\hat {x} _ {i}) > \\gamma^ {t}\\right) \\quad \\times \\\\ \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi (\\hat {x} _ {i})\\right)\\right) > \\mathcal {T} _ {\\hat {p} _ {\\theta} \\left(y \\mid \\pi (\\hat {x} _ {i})\\right)} ^ {t}\\right) \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.79 + ], + "angle": 0, + "content": "where \\(\\gamma^t\\) is the APM threshold at iteration \\(t\\), estimated as explained below, and \\(\\mathcal{T}_{\\hat{p}_0(y|\\pi(\\hat{x}_i))}^t\\) is the flexible threshold estimated as in FlexMatch [49]. To train our model, we adopt the best practices [41, 49] and optimize the weighted combination of the supervised and unsupervised losses:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.811, + 0.47, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda \\mathcal {L} _ {u} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.329, + 0.857 + ], + "angle": 0, + "content": "where the supervised loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.864, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} H \\left(y _ {i}, p _ {\\theta} (y \\mid \\pi \\left(x _ {i}\\right))\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.895, + 0.77 + ], + "angle": 0, + "content": "Average Pseudo-Margin Threshold Estimation Inspired by Pleiss et al. [33], we propose to estimate the average pseudo-margin threshold \\(\\gamma^t\\) by analyzing the training dynamics of a special category of unlabeled examples, which we force to be erroneous or mislabeled examples. That is, to create the sample of erroneous examples \\(E\\), we randomly sample a subset of unlabeled examples from \\(U\\) that we assign to an inexistent (or virtual) class \\(C + 1\\) at the beginning of the training process and remove them from \\(U\\). The purpose of these erroneous examples is to mimic the training dynamics of incorrectly pseudo-labeled (unlabeled) examples and use them as proxy to estimate the cutoff of (potentially) mislabeled pseudo-labels. Since the examples in \\(E\\) should belong to one of the \\(C\\) original classes, assigning them to the inexistent class \\(C + 1\\) makes them by definition mislabeled (see Appendix A for additional insights into this virtual class). As with all unlabeled examples from \\(U\\), we compute \\(APM_{C + 1}^t\\) for the special category of erroneous examples from \\(E\\), but unlike the unlabeled examples from \\(U\\), the erroneous ones from \\(E\\) have a fixed class \\(C + 1\\). To mimic the training dynamics of unlabeled examples from \\(U\\), we use strong augmentations to compute the loss of the erroneous examples from \\(E\\). That is, given a batch \\(E_b\\) of \\(B\\) erroneous examples, the erroneous sample loss becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.774, + 0.892, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} = \\sum_ {i = 1} ^ {B} H (C + 1, p _ {\\theta} (y | \\Pi (\\tilde {x} _ {i}))) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.862 + ], + "angle": 0, + "content": "At iteration \\( t \\), we use the APMs of the erroneous examples to choose the APM threshold \\( \\gamma^t \\). We set \\( \\gamma^t \\) as the APM of the \\( 95^{th} \\) percentile erroneous sample. The total loss becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.866, + 0.892, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda \\left(\\mathcal {L} _ {u} + \\mathcal {L} _ {e}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.884, + 0.902 + ], + "angle": 0, + "content": "Our full MarginMatch algorithm is shown in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15776" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.895, + 0.207 + ], + "angle": 0, + "content": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10
#Labels/Class425400425100425100425100
Pseuso-Labeling74.610.2646.492.2015.080.1987.450.8557.740.2836.550.2464.615.6025.212.039.400.3274.680.9955.452.4332.640.71
UDA10.793.755.320.064.410.0748.951.5929.430.2123.870.235.344.274.260.391.950.0137.828.449.811.156.810.17
MixMatch45.242.1512.761.147.130.3462.152.1741.511.1928.160.2446.181.783.980.173.501.1334.151.548.950.3210.410.73
ReMixMatch5.270.194.850.134.040.1247.150.7627.140.2323.780.124.230.313.180.041.940.0631.510.758.540.486.190.24
FixMatch7.80.284.910.054.250.0848.210.8229.450.1622.890.123.971.183.131.031.970.0338.434.1410.451.046.430.33
FlexMatch5.040.065.040.094.190.0139.991.6226.960.0822.440.158.193.207.782.556.720.3029.151.328.230.135.770.12
MarginMatch4.910.074.730.123.980.0236.971.3223.710.1321.390.123.751.203.141.171.930.0125.373.587.310.355.520.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.15, + 0.216, + 0.819, + 0.23 + ], + "angle": 0, + "content": "Table 1. Test error rates on CIFAR-10, CIFAR-100, SVHN, and STL-10 datasets. Best results are shown in blue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.473, + 0.439 + ], + "angle": 0, + "content": "Exponential Moving Average of Pseudo-Margins The current definition of APM weighs the pseudo-margin at iteration \\( t \\) identical to the pseudo-margin at a much earlier iteration \\( p \\) (\\( t >> p \\)). This is problematic since very old pseudo-margins eventually become deprecated (especially due to the large number of iterations through unlabeled data in consistency training (\\( \\sim 9K \\)), and hence, the old margins are no longer indicative of the current learning status of the model. To this end, instead of averaging all pseudo-margins (from the beginning of training to the current iteration), we propose to use an exponential moving average to place more importance on recent iterations. Formally, APM becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.443, + 0.47, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A P M} _ {c} ^ {t} (\\hat {x}) = \\mathrm {P M} _ {c} ^ {t} (\\hat {x}) * \\frac {\\delta}{1 + t} + \\mathrm {A P M} _ {c} ^ {t - 1} (\\hat {x}) * \\left(1 - \\frac {\\delta}{1 + t}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.486, + 0.466, + 0.501 + ], + "angle": 0, + "content": "We set the smoothing parameter \\(\\delta\\) to 0.997 in experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.21, + 0.531 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.78 + ], + "angle": 0, + "content": "We evaluate the performance of our MarginMatch on a wide range of SSL benchmark datasets. Specifically, we perform experiments with various numbers of labeled examples on CIFAR-10, CIFAR-100 [21], SVHN [31], STL-10 [8], ImageNet [10], and WebVision [26]. For smaller scale datasets such as CIFAR-10, CIFAR-100, SVHN, and STL-10 we randomly sample a small number of labeled examples per class (ranging from 4 labels per class to 400 labels per class) and treat them as labeled data, whereas the remaining labeled examples are treated as unlabeled data, except for STL-10 [8], which provides its own set of unlabeled examples. On ImageNet and WebVision, we use \\(\\sim 10\\%\\) of the available labeled examples as labeled data, with the remaining \\((90\\%)\\) being treated as unlabeled data. In all our experiments, we sample \\(5\\%\\) of the unlabeled data and place it in the set of erroneous examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.871 + ], + "angle": 0, + "content": "We report the mean and standard deviation of error rates from five runs with different parameter initializations. Similar to FixMatch [41], we use Wide Residual Networks [48]: WRN-28-2 for CIFAR-10 and SVHN; WRN-28-8 for CIFAR-100; and WRN-37-2 for STL-10. We use ResNet-50 [14] for both ImageNet and WebVision." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In our experiments, we adopt the same hyperparameters as FixMatch [41]. Specifically, we use stochastic gradient" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.895, + 0.485 + ], + "angle": 0, + "content": "descent (SGD) with a momentum of 0.9. We start with a learning rate of 0.03 and employ a cosine learning rate; at iteration \\( k \\), our learning rate is \\( \\eta(k) = \\cos\\left(\\frac{7k\\pi}{16K}\\right) \\), where \\( K \\) is the maximum number of iterations and is set to \\( 2^{20} \\). We also leverage the same data augmentations as in FlexMatch [49]. Specifically, for weak augmentations we employ a standard flip-and-shift augmentation and use RandAugment [9] for strong augmentations. We set the batch size \\( B = 64 \\), ratio of unlabeled to labeled data in a batch to \\( \\nu = 7 \\) and weigh the supervised and unsupervised losses equally (i.e., \\( \\lambda = 1 \\)). We set our initial confidence threshold to \\( \\tau = 0.95 \\) and our average pseudo-margin (APM) threshold to the APM of the \\( 95^{th} \\) percentile threshold sample. To report the error rates, we compare all the approaches using the model at the end of training as in FixMatch [41]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.497, + 0.874, + 0.513 + ], + "angle": 0, + "content": "3.1. CIFAR-10, CIFAR-100, SVHN, and STL-10" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.794 + ], + "angle": 0, + "content": "We compare MarginMatch against strong baselines and prior works: Pseudo-Labeling [1], Unsupervised Data Augmentation (UDA) [46], MixMatch [5], ReMixMatch [4], FixMatch [41], and FlexMatch [49]. We show in Table 1 the error rates obtained by our MarginMatch and the baselines on the CIFAR-10, CIFAR-100, SVHN and STL-10 datasets. First, we observe that our approach improves the performance on both CIFAR-10 and CIFAR-100. On CIFAR-10, MarginMatch improves performance in all data regimes upon FlexMatch [49], which is the current state-of-the-art, while maintaining a good error rate standard deviation. On CIFAR-100, which is significantly more challenging than CIFAR-10, we observe that MarginMatch bring substantially larger improvements. Notably, we see \\(3.02\\%\\) improvement over FlexMatch in error rate using only 4 labels per class, and \\(3.25\\%\\) improvement using 25 examples per class. These results on CIFAR-100 emphasize the effectiveness of MarginMatch, which performs well on a more challenging dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "On SVHN, our approach performs better than FixMatch using 4 labels per class and performs similarly with FixMatch using 25 and 100 labels per class. However, on this dataset, MarginMatch performs much better compared with FlexMatch. For example, MarginMatch achieves \\(3.75\\%\\) error rate using 4 labels per class, whereas FlexMatch obtains an error rate of \\(8.19\\%\\) with the same labels per class," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15777" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.28, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.482, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.681, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.089, + 0.882, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.235, + 0.79, + 0.25 + ], + "angle": 0, + "content": "Figure 2. Convergence speed of MarginMatch against FixMatch and FlexMatch with 4 labels per class." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.273, + 0.428, + 0.369 + ], + "angle": 0, + "content": "
DatasetImageNetWebVision
TOP-1TOP-5TOP-1TOP-5
Supervised48.3925.4949.5826.78
FixMatch43.6621.8044.7622.65
FlexMatch42.0219.4943.8722.07
MarginMatch41.0518.2843.0821.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.38, + 0.471, + 0.407 + ], + "angle": 0, + "content": "Table 2. Test error rates on the ImageNet and WebVision datasets. Best results are shown in blue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.551 + ], + "angle": 0, + "content": "yielding an improvement of MarginMatch of \\(4.44\\%\\) over FlexMatch. We hypothesize that the low performance of FlexMatch is due to its limitation in handling unbalanced class distributions [49]. On STL-10, MarginMatch as well outperforms all the other approaches both in error rates and error rate standard deviation. Notably, on this dataset, our approach pushes the performance of FlexMatch by \\(3.78\\%\\) in error rate using only 4 labels per class and by \\(0.92\\%\\) using 25 labels per class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Next, we compare MarginMatch with FixMatch and FlexMatch in terms of convergence speed in the extremely labelscarcse setting of 4 labels per class and show these results in Figure 2. Notably, we observe that MarginMatch has a similar convergence speed (or even better on CIFAR-100) compared with FlexMatch while achieving a lower test error rate than FlexMatch on all datasets with 4 labels per class (see Table 1). Even more strikingly, compared with FixMatch, MarginMatch has a much superior convergence speed for a much better test error rate with 4 labels per class. This is because the rigid thresholds in FixMatch allow access only to a small amount of unlabeled data for training at each iteration and it takes a lot longer for the model to train." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.31, + 0.773 + ], + "angle": 0, + "content": "3.2. ImageNet and WebVision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.902 + ], + "angle": 0, + "content": "To showcase the effectiveness of our approach in a large-scale setup, we test our MarginMatch on ImageNet [10] and WebVision [26] using \\(10\\%\\) labeled examples in total. We show the results obtained in Table 2. We observe that our MarginMatch outperforms FixMatch and FlexMatch on both datasets. It is worth noting that large-scale self-supervised approaches such as SimCLR [7] achieve high performance on ImageNet but at a much higher computational cost." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.274, + 0.871, + 0.311 + ], + "angle": 0, + "content": "
δ0.950.990.9950.9970.9991
ERR RATE38.1338.0537.9237.9139.1239.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Table 3. Error rates obtained on CIFAR-100 with four examples per class and various smoothing values \\( \\delta \\) . Best result is in blue." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.894, + 0.439 + ], + "angle": 0, + "content": "MarginMatch outperforms other SSL methods using the same ResNet-50 architecture at the same computational cost. We emphasize MarginMatch is most successful and relevant in low data regimes on smaller datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.651, + 0.473 + ], + "angle": 0, + "content": "4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.893, + 0.789 + ], + "angle": 0, + "content": "Exponential Moving Average Smoothing for APM Computation In our approach, we employ an exponential moving average (EMA) of the pseudo-margin values with a smoothing value of \\(\\delta = 0.997\\) to compute the APM. We now analyze how our approach performs with different EMA smoothing values or with no EMA at all. Table 3 shows these results on CIFAR-100 with 4 labels per class. First, we observe that employing a simple average of pseudo-margin values for the APM computation (i.e., \\(\\delta = 1\\)) performs extremely poorly, obtaining a \\(39.72\\%\\) error rate. This result emphasizes that margins eventually become deprecated and it is essential to scale them down in time. Using a low smoothing factor of \\(\\delta = 0.95\\) is not effective either, denoting that abruptly forgetting margin values does not work either. Our chosen \\(\\delta = 0.997\\) strikes a balance between the two by eliminating the harmful effects of very old margins while keeping track of a good amount of previous estimates (e.g., a margin value computed 200 epochs previously is scaled down by 0.55, while a margin value computed 1000 epochs previously is scaled by 0.05)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Pseudo-Margin vs. Other Measures for Pseudo-Label Correctness Our MarginMatch monitors the pseudomargins of a model's predictions across training iterations to ensure the quality of pseudo-labels. However, other measures such as confidence or entropy exist that can assess the pseudo-label correctness. Hence, we perform an ablation where we replace the pseudo-margins in our MarginMatch" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15778" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.198 + ], + "angle": 0, + "content": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10
#Labels/Class425400425100425100425100
Avg Confidence23.872.7314.211.377.540.7841.232.1531.491.4824.112.368.994.276.540.394.730.0131.678.4414.871.157.590.17
Avg Entropy8.580.416.180.155.850.1245.100.9126.021.1122.130.2515.691.2512.740.789.330.0529.543.5110.631.3510.840.47
Avg Margin7.250.295.380.764.730.0939.721.5225.210.5223.180.1718.451.3611.290.938.400.0428.454.289.341.347.590.21
EMA Confidence4.910.454.740.093.990.0638.670.7425.610.1221.480.173.840.233.250.031.930.0925.90.817.60.425.740.57
EMA Entropy6.40.438.340.124.210.0941.630.7636.840.1322.520.073.811.263.170.872.140.0427.214.058.281.016.790.27
EMA Margin4.910.074.730.123.980.0236.971.3223.710.1321.390.123.751.203.141.171.930.0125.373.587.310.355.520.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.208, + 0.816, + 0.223 + ], + "angle": 0, + "content": "Table 4. Test error rates comparing pseudo-margin with confidence and entropy. Best results are shown in blue." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.233, + 0.273, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.233, + 0.468, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.378, + 0.471, + 0.406 + ], + "angle": 0, + "content": "Figure 3. Mask rate and impurity on CIFAR-100 with 4 labeled examples per class." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.413, + 0.47, + 0.639 + ], + "angle": 0, + "content": "with average confidence and entropy and compare their performance. Specifically, we design the following approaches: 1) Avg Confidence monitors the confidence of the prediction for each unlabeled example and takes the average over the training iterations; 2) Avg Entropy monitors the entropy of the class probability distribution for each unlabeled example and takes the average across the training iterations. In addition, we also consider 3) EMA Confidence and 4) EMA Entropy which are similar to Avg Confidence and Avg Entropy, respectively, but use an exponential moving average (EMA) instead of the simple averaging. The estimation of the threshold for each of these approaches is done in a similar manner as for pseudo-margins, using erroneous samples and considering the value of the 95th percentile erroneous sample as the threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.471, + 0.791 + ], + "angle": 0, + "content": "We show the results obtained using these approaches in Table 4. First, we observe that all measures (pseudo-margin, confidence and entropy) with EMA perform better than their counterpart with simple averaging. Second, EMA Margin achieves the lowest test error rates compared with EMA Confidence and EMA Entropy. Thus, we conclude that pseudo-margins provide an excellent measure for pseudolabel correctness. See Appendix B for some additional insights into why EMA Margin outperforms EMA confidence and entropy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.174, + 0.823 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.305, + 0.849 + ], + "angle": 0, + "content": "5.1. Mask Rate and Impurity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We now contrast MarginMatch with FixMatch and FlexMatch in terms of the quality of pseudo-labels using two metrics: mask rate and impurity and show these results in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.895, + 0.476 + ], + "angle": 0, + "content": "Figure 3, respectively, using CIFAR-100 with 4 labels per class. Mask rate is defined as the fraction of pseudo-labeled examples that do not participate in the training at epoch \\( t \\) due to confidence masking or pseudo-margin masking (or both). Impurity in contrast is defined as the fraction of pseudolabeled examples that do participate in the training at epoch \\( t \\) but with a wrong label. An effective SSL model minimizes both metrics: a low mask rate indicates that the model has access to more unlabeled examples during training (otherwise a low percentage and less diverse set of unlabeled examples are seen during training) while low impurity indicates that the pseudo-labels of these examples are of high quality. Note that we can compute impurity on these two datasets because our unlabeled data comes from the labeled training set of each of these datasets (thus we compare the pseudo-labels against the gold labels of each dataset)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.895, + 0.673 + ], + "angle": 0, + "content": "As can be seen from the figures, FixMatch has a significantly larger mask rate due to the rigid confidence threshold set to a high value of 0.95. In contrast, FlexMatch lowers the mask rate by \\(5\\%\\) with the introduction of flexible thresholds, but has a much higher impurity compared with FixMatch. Notably, our MarginMatch has only a slightly higher mask rate compared with FlexMatch and at the same time achieves a much lower impurity than FlexMatch and even FixMatch despite that FixMatch employs a very high confidence threshold. These results show that MarginMatch that enforces an additional measure for pseudo-labeled data quality maintains a low mask rate without compromising the quality of the pseudo-labels (i.e., low mask rate and low impurity)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.691, + 0.696 + ], + "angle": 0, + "content": "5.2. Anecdotal Evidence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We show in Figure 4 anecdotal evidence of the effectiveness of MarginMatch. To this end, we extract two bird images from our unlabeled portion of CIFAR-10 [21] of various learning difficulties that resemble characteristics of plane images (e.g., the background). The top part of the figure illustrates the progression over the training iterations of the confidence and the confidence thresholds of FlexMatch for the classes bird and plane, whereas the bottom part of the figure illustrates the progression of the APM threshold of MarginMatch along with its APMs of bird and plane classes over the training iterations. In the rightmost image, for MarginMatch we can observe that the APM of the bird class becomes stronger and stronger as the training progresses and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.473, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.236, + 0.47, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.401 + ], + "angle": 0, + "content": "Figure 4. Confidence thresholding vs. APM Thresholding on two images from the CIFAR-10 dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.471, + 0.664 + ], + "angle": 0, + "content": "eventually exceeds the APM threshold of MarginMatch, and hence, the image is included in the training until the end with the correct bird label. Interestingly, for the same image, in FlexMatch the confidence for the bird class is very close to the bird confidence threshold and eventually falls below this threshold and exits the training set. In contrast, the leftmost image is significantly more challenging than the rightmost image since it is more similar to plane images, which makes it an easily confusable example. Here, we observe that the confidence of FlexMatch exceeds the flexible threshold with the incorrect argmax class plane starting from iteration 3000. Moreover, Flexmatch continues to use this image with the wrong plane label for the remaining of the training process. Critically, in MarginMatch the APM value for the plane class does not exceed the APM threshold, and the model eventually learns to classify this image correctly and includes it in training with the correct bird pseudo-label." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.219, + 0.695 + ], + "angle": 0, + "content": "6. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Here, we focus on various SSL approaches that our MarginMatch directly builds upon although there are other SSL techniques that are not presented in this review, such as approaches based on generative models [17, 24], graph-based approaches [19, 20] and robust SSL [32, 37] (see Appendix C for a comparison between MarginMatch and Robust SSL approaches)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Self-training [28, 36, 40, 47] is a popular SSL method where the predictions of a model on unlabeled data are used as artificial labels to train against. Noisy student training [47] is a popular self-training approach that also leverages knowledge distillation [16] and iteratively jointly trains two models in a teacher-student framework. Noisy student uses a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "larger model size and noised inputs, exposing the student to more difficult learning environments, leading to an increased performance compared to the teacher." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.895, + 0.378 + ], + "angle": 0, + "content": "Pseudo-labeling is a variant of self-training where these predictions are sharpened to obtain hard labels [25]. The use of hard labels can be seen as a means of entropy minimization [12] and nowadays is a valuable component in most successful SSL approaches [4, 5, 49]. These hard labels are usually used along a confidence threshold, where unconfident unlabeled examples are completely disregarded (e.g., [5]) to avoid using noisy pseudo-labels. Recently, approaches such as Curriculum Labeling (CL) [6] or FlexMatch [49] started to explore curriculum learning in the SSL context. CL proposes a self-pacing strategy of identifying easy and hard examples to ensure that the model first uses easy and progressively moves towards harder examples. Similarly, MarginMatch uses curriculum learning and pseudo-labeling, but the focus of our approach is placed on producing better thresholds for assessing the quality of pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.895, + 0.621 + ], + "angle": 0, + "content": "Consistency regularization [2] is a method that applies random perturbations when generating the artificial label, such as data augmentation [4, 5, 41], dropout [39], or adversarial perturbations [29]. Current state-of-the-art approaches [41, 49] exploit a combination of weak and strong data augmentations, which were shown to be extremely beneficial in SSL. The most popular strong augmentations used in the SSL literature are RandAugment [9] and CTAugment [4]. The approaches based on these methods first generate a hard label using pseudo-labeling on a weakly augmented image (i.e., using a low noise transformation such as a flip-and-shift augmentation), then optimize the predictions of the model on a strongly augmented version of the same image towards this hard label. Similar to these approaches, MarginMatch uses the same combination of weak and strong data augmentations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.619, + 0.649 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.9 + ], + "angle": 0, + "content": "In this paper, we proposed a novel semi-supervised learning method that improves the pseudo-label quality using training dynamics. Our new method is lightweight and achieves state-of-the-art performance on four computer vision SSL datasets in low data regimes and on two large-scale benchmarks. MarginMatch takes into consideration not only a flexible confidence threshold to account for the difficulty of each class, but also a measure of quality for each unlabeled example using training dynamics. In addition, MarginMatch is a general approach that can be leveraged in most SSL frameworks and we hope that it can attract future research in analyzing the effectiveness of SSL approaches focused on data quality. As future work, we aim to further explore our method in settings when there is a mismatch between the labeled and unlabeled data distributions (i.e., making use of out-of-domain unlabeled data)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15780" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.472, + 0.171 + ], + "angle": 0, + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. CoRR, abs/1908.02983, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.286 + ], + "angle": 0, + "content": "[3] Peter L. Bartlett, Dylan J. Foster, and Matus Telgarsky. Spectrally-normalized margin bounds for neural networks. CoRR, abs/1706.08498, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[4] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.472, + 0.415 + ], + "angle": 0, + "content": "[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[6] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In AAAI, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[8] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.472, + 0.658 + ], + "angle": 0, + "content": "[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.716 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[11] Gamaleldin Fathy Elsayed, Dilip Krishnan, Hossein Mobahi, Kevin Regan, and Samy Bengio. Large margin deep networks for classification. 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[12] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In L. Saul, Y. Weiss, and L. Bottou, editors, Advances in Neural Information Processing Systems, volume 17. MIT Press, 2004. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1321-1330. PMLR, 06-11 Aug 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[15] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[16] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[17] Geoffrey E Hinton and Russ R Salakhutdinov. Using deep belief nets to learn covariance kernels for gaussian processes. In J. Platt, D. Koller, Y. Singer, and S. Roweis, editors, Advances in Neural Information Processing Systems, volume 20. Curran Associates, Inc., 2007. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.894, + 0.372 + ], + "angle": 0, + "content": "[18] Yiding Jiang, Dilip Krishnan, Hossein Mobahi, and Samy Bengio. Predicting the generalization gap in deep networks with margin distributions, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[19] Thorsten Joachims. Transductive inference for text classification using support vector machines. In Proceedings of the Sixteenth International Conference on Machine Learning, ICML '99, page 200-209, San Francisco, CA, USA, 1999. Morgan Kaufmann Publishers Inc. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.894, + 0.497 + ], + "angle": 0, + "content": "[20] Thorsten Joachims. Transductive learning via spectral graph partitioning. In Proceedings of the Twentieth International Conference on International Conference on Machine Learning, ICML'03, page 290-297. AAAI Press, 2003. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.894, + 0.607 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, NIPS'12, page 1097-1105, Red Hook, NY, USA, 2012. Curran Associates Inc. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.894, + 0.649 + ], + "angle": 0, + "content": "[23] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR (Poster). OpenReview.net, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[24] J.A. Lasserre, C.M. Bishop, and T.P. Minka. Principled hybrids of generative and discriminative models. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 1, pages 87-94, 2006. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.894, + 0.775 + ], + "angle": 0, + "content": "[25] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.894, + 0.817 + ], + "angle": 0, + "content": "[26] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. CoRR, abs/1708.02862, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pages 181-196, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15781" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[28] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152-159, New York City, USA, June 2006. Association for Computational Linguistics. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[29] Takeru Miyato, Andrew M Dai, and Ian Goodfellow. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[30] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.315 + ], + "angle": 0, + "content": "[31] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[32] Jongjin Park, Sukmin Yun, Jongheon Jeong, and Jinwoo Shin. Opencos: Contrastive semi-supervised learning for handling open-set unlabeled data. In ECCV Workshops, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.472, + 0.44 + ], + "angle": 0, + "content": "[33] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying mislabeled data using the area under the margin ranking. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 17044-17056. Curran Associates, Inc., 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[34] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.472, + 0.551 + ], + "angle": 0, + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.472, + 0.62 + ], + "angle": 0, + "content": "[36] Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. In 2005 Seventh IEEE Workshops on Applications of Computer Vision (WACV/MOTION'05) - Volume 1, volume 1, pages 29-36, 2005. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[37] Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set semi-supervised learning with open-set consistency regularization. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 25956-25967. Curran Associates, Inc., 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[38] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Mutual exclusivity loss for semi-supervised deep learning. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1908-1912. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[39] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29:1163-1171, 2016. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[40] H. Scudder. Probability of error of some adaptive pattern-recognition machines. IEEE Transactions on Information Theory, 11(3):363-371, 1965. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[41] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.175 + ], + "angle": 0, + "content": "Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 596-608. Curran Associates, Inc., 2020. 1, 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[42] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.248, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[43] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[44] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in neural information processing systems, pages 1195–1204, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[45] Vikas Verma, Kenji Kawaguchi, Alex Lamb, Juho Kannala, Arno Solin, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi-supervised learning. Neural Networks, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[46] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc., 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.892, + 0.57 + ], + "angle": 0, + "content": "[47] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687-10698, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.572, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[48] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.709 + ], + "angle": 0, + "content": "[49] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 18408-18419. Curran Associates, Inc., 2021. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.713, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[50] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. CoRR, abs/1611.03530, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15782" + } + ] +] \ No newline at end of file diff --git a/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_origin.pdf b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f09019d2219839711374659328b58b8cf0984812 --- /dev/null +++ b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/1e527083-57e4-4db8-89d4-269dfeae517d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4951e92a8dd08bc8e0a0399d5c32c243d68c16a75c59415215dc0130e7050a6 +size 491755 diff --git a/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/full.md b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/full.md new file mode 100644 index 0000000000000000000000000000000000000000..235a6cb9f5a736f9e236cbc9b97756cb8ce9d01f --- /dev/null +++ b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/full.md @@ -0,0 +1,310 @@ +# MarginMatch: Improving Semi-Supervised Learning with Pseudo-Margins + +Tiberiu Sosea Cornelia Caragea University of Illinois Chicago tsosea2@uic.edu cornelia@uic.edu + +# Abstract + +We introduce MarginMatch, a new SSL approach combining consistency regularization and pseudo-labeling, with its main novelty arising from the use of unlabeled data training dynamics to measure pseudo-label quality. Instead of using only the model's confidence on an unlabeled example at an arbitrary iteration to decide if the example should be masked or not, MarginMatch also analyzes the behavior of the model on the pseudo-labeled examples as the training progresses, to ensure low quality predictions are masked out. MarginMatch brings substantial improvements on four vision benchmarks in low data regimes and on two large-scale datasets, emphasizing the importance of enforcing high-quality pseudo-labels. Notably, we obtain an improvement in error rate over the state-of-the-art of $3.25\%$ on CIFAR-100 with only 25 labels per class and of $3.78\%$ on STL-10 using as few as 4 labels per class. We make our code available at https://github.com/tsosea2/MarginMatch. + +# 1. Introduction + +Deep learning models have seen tremendous success in many vision tasks [14, 22, 27, 42, 43]. This success can be attributed to their scalability, being able to produce better results when they are trained on large datasets in a supervised fashion [15, 27, 34, 35, 43, 47]. Unfortunately, large labeled datasets annotated for various tasks and domains are difficult to acquire and demand considerable annotation effort or domain expertise. Semi-supervised learning (SSL) is a powerful approach that mitigates the requirement for large labeled datasets by effectively making use of information from unlabeled data, and thus, has been studied extensively in vision [4, 5, 23, 25, 30, 36, 38, 39, 44-46]. + +Recent SSL approaches integrate two important components: consistency regularization [46, 49] and pseudolabeling [25]. Consistency regularization works on the assumption that a model should output similar predictions when fed perturbed versions of the same image, whereas pseudo-labeling uses the model's predictions of unlabeled examples as labels to train against. For example, Sohn et al. [41] introduced FixMatch that combines consistency reg + +ularization on weak and strong augmentations with pseudolabeling. FixMatch relies heavily on a high-confidence threshold to compute the unsupervised loss, disregarding any pseudo-labels whose confidence falls below this threshold. While training using only high-confidence pseudo-labels has shown to consistently reduce the confirmation bias [1], this rigid threshold allows access only to a small amount of unlabeled data for training, and thus, ignores a considerable amount of unlabeled examples for which the model's predictions do not exceed the confidence threshold. More recently, Zhang et al. [49] introduced FlexMatch that relaxes the rigid confidence threshold in FixMatch to account for the model's learning status of each class in that it adaptively scales down the threshold for a class to encourage the model to learn from more examples from that class. The flexible thresholds in FlexMatch allow the model to have access to a much larger and diverse set of unlabeled data to learn from, but lowering the thresholds can lead to the introduction of wrong pseudo-labels, which are extremely harmful for generalization. Interestingly, even when the high-confidence threshold is used in FixMatch can result in wrong pseudo-labels. See Figure 1 for incorrect pseudo-labels detected in the training set after we apply FixMatch and FlexMatch on ImageNet. We posit that a drawback of FixMatch and FlexMatch and in general of any pseudo-labeling approach is that they use the confidence of the model only at the current iteration to enforce quality of pseudo-labels and completely ignore model's predictions at prior iterations. + +In this paper, we propose MarginMatch, a new SSL approach that monitors the behavior of the model on the unlabeled examples as the training progresses, from the beginning of training until the current iteration, instead of using only the model's current belief about an unlabeled example (i.e., its confidence at the current iteration) to decide if the example should be masked or not. We estimate a pseudo-label's contribution to learning and generalization by introducing pseudo-margins of unlabeled examples averaged across training iterations. Pseudo-margins of unlabeled examples extend the margins from machine learning [3, 11, 18, 33] which provide a measure of confidence of the outputs of the model and capture the difference between the output for the correct + +![](images/0d697b74530a6eb1d50b7f27f7764bd5484004a26b834d52b3cc2faab6dd374d.jpg) +Figure 1. Incorrect pseudo-labels propagated until the end of the training process for FixMatch and FlexMatch on ImageNet. + +(gold) label and the other labels. In our case, the pseudomargins capture how much larger the assigned logit (the logit corresponding to the argmax of the model's prediction) is compared with all other logits at iteration $t$ . Similar to FlexMatch, in MarginMatch we take advantage of the flexible confidence thresholds to allow the model to learn from larger and more diverse sets of unlabeled examples, but unlike FlexMatch, we train the model itself to identify the characteristics of mislabeled pseudo-labels simply by monitoring the model's training dynamics on unlabeled data over the iterations. + +We carry out comprehensive experiments using established SSL experimental setups on CIFAR-10, CIFAR-100 [21], SVHN [31], STL-10 [8], ImageNet [10], and WebVision [26]. Despite its simplicity, our findings indicate that MarginMatch produces improvements in performance over strong baselines and prior works on all datasets at no additional computational cost. Notably, compared to current state-of-the-art, on CIFAR-100 we see $3.02\%$ improvement in error rate using only 4 labels per class and $3.78\%$ improvement on STL-10 using the same extremely label-scarce setting of 4 labels per class. In addition, on ImageNet [10] and WebVision [26] we find that MarginMatch pushes the state-of-the-art error rates by $0.97\%$ on ImageNet and by $0.79\%$ on WebVision. + +Our contributions are as follows: + +1. We introduce a new SSL approach which we call MarginMatch that enforces high pseudo-label quality during training. Our approach allows access to a large set of unlabeled data to learn from (thus, incorporating more information from unlabeled data) and, at the same time, monitors the training dynamics of unlabeled data as training progresses to detect and filter out potentially incorrect pseudo-labels. +2. We show that MarginMatch outperforms existing works on six well-established computer vision benchmarks + +showing larger improvements in error rates especially on challenging datasets, while achieving similar convergence performance (or better) than prior works. + +3. We perform a comprehensive analysis of our approach and indicate potential insights into why our Margin-Match substantially outperforms other SSL techniques. + +# 2. MarginMatch + +Notation Let $L = \{(x_{1},y_{1}),\dots,(x_{B},y_{B})\}$ be a batch of size $B$ of labeled examples and $U = \{\hat{x}_1,\dots,\hat{x}_{\nu B}\}$ be a batch of size $\nu B$ of unlabeled examples, where $\nu$ is the batch-wise ratio of unlabeled to labeled examples. Let $p_{\theta}(y|x)$ denote the class distribution produced by model $\theta$ on input image $x$ and $\hat{p}_{\theta}(y|x)$ denote the argmax of this distribution as a one-hot label. Let also $H(p,q)$ denote the cross-entropy between two probability distributions $p$ and $q$ . + +# 2.1. Background + +Consistency reg [39] is an important component in recent semi-supervised learning approaches and relies on the continuity assumption [2,23] that the model should output similar predictions on multiple perturbed versions of the same input $x$ . As mentioned above, examples of two such approaches are FixMatch [41] and FlexMatch [49] that use consistency regularization at their core combined with psedo-labeling. In psedo-labeling [25], a model itself is used to assign artificial labels for unlabeled data and only artificial labels whose largest class probability is above a predefined confidence threshold are used during training. + +Specifically, FixMatch [41] predicts artificial labels for unlabeled examples using a weakly-augmented version of each unlabeled example and then employs the artificial labels as pseudo-labels to train against but this time using a strongly-augmented version of each unlabeled example. That is, FixMatch minimizes the following batch-wise consistency loss on unlabeled data: + +$$ +\begin{array}{l} \mathcal {L} _ {u} = \sum_ {i = 1} ^ {\nu B} \mathbb {1} \left(\max \left(p _ {\theta} \left(y \mid \pi \left(\hat {x} _ {i}\right)\right)\right) > \tau\right) \quad \times \\ H \left(\hat {p} _ {\theta} (y | \pi (\hat {x} _ {i})), p _ {\theta} (y | \Pi (\hat {x} _ {i}))\right) \tag {1} \\ \end{array} +$$ + +where $\tau$ is a confidence threshold, $\pi$ and $\Pi$ are weak and strong augmentations, respectively, and $\mathbb{1}$ is the indicator function. Therefore, the low-confidence examples (lower than $\tau$ ) are completely ignored despite containing potentially useful information for model training. + +FlexMatch [49] argues that using a fixed threshold $\tau$ to filter the unlabeled data ignores the learning difficulties of different classes, and thus, introduces class-dependent thresholds, which are obtained by adaptively scaling $\tau$ depending on the learning status of each class. FlexMatch assumes that a class with fewer examples above the fixed threshold $\tau$ has a greater learning difficulty, and hence, it adaptively lowers the threshold $\tau$ to encourage more training examples from this class to be learned. The learning status $\alpha_{c}$ for a class $c$ is simply computed as the number of unlabeled examples that are predicted in class $c$ and pass the fixed threshold $\tau$ : + +$$ +\alpha_ {c} = \sum_ {i = 1} ^ {n} \mathbb {1} \left(\max \left(p _ {\theta} \left(y \mid \pi \left(\hat {x} _ {i}\right)\right)\right) > \tau\right) \mathbb {1} \left(\hat {p} _ {\theta} \left(y \mid \pi \left(\hat {x} _ {i}\right)\right) = c\right) \tag {2} +$$ + +where $n$ is the total number of unlabeled examples. This learning effect is then normalized and used to obtain the class-dependent threshold for each class $c$ : + +$$ +\mathcal {T} _ {c} = \frac {\alpha_ {c}}{\max _ {c} \left(\alpha_ {c}\right)} \times \tau \tag {3} +$$ + +In practice, FlexMatch iteratively computes new thresholds after each complete pass through unlabeled data, hence we can parameterize $\mathcal{T}_c$ as $\mathcal{T}_c^t$ , denoting the threshold obtained at iteration $t$ . The unlabeled loss is then obtained by plugging in the adaptive threshold $\mathcal{T}_c^t$ in Eq. 1: + +$$ +\begin{array}{l} \mathcal {L} _ {u} = \sum_ {i = 1} ^ {\nu B} \mathbb {1} (\max (p _ {\theta} (y | \pi (\hat {x} _ {i}))) > \mathcal {T} _ {\hat {p} _ {\theta} (y | \pi (\hat {x} _ {i}))} ^ {t}) \quad \times \\ H \left(\hat {p} _ {\theta} (y | \pi (\hat {x} _ {i})), p _ {\theta} (y | \Pi (\hat {x} _ {i}))\right) \tag {4} \\ \end{array} +$$ + +The aforementioned works use the confidence of the model solely at the current iteration to enforce quality of pseudo-labels. We believe this is not sufficient as it provides only a myopic view of the model's behavior (i.e., its confidence) on unlabeled data (at a single iteration) and may result in wrong pseudo-labels even when the confidence threshold is high enough (e.g., if the model is miscalibrated or overly-confident [13]). Figure 1 shows examples of images that are added to the training set with a wrong pseudo-label for both FixMatch and FlexMatch. These types of unlabeled examples, which are incorrectly pseudo-labeled and used + +during training are particularly harmful for deep neural networks, which can attain zero training error on any dataset, even on randomly assigned labels [50], resulting in poor generalization capabilities. + +# 2.2. Proposed Approach: MarginMatch + +We now introduce MarginMatch, our new SSL approach that uses the model's training dynamics on unlabeled data to improve pseudo-label data quality. Our approach leverages consistency regularization with weak and strong augmentations and pseudo-labeling, but instead of using only the model's current belief (i.e., its confidence at the current iteration) to decide if an unlabeled example should be used for training or not, our MarginMatch monitors the training dynamics of unlabeled data over the iterations by investigating the margins (a measure of confidence) of the outputs of the model [3]. The margin of a training example is a well established metric in machine learning [3, 11, 18, 33] that quantifies the difference between the logit corresponding to the assigned ground truth label and the largest other logit. + +In our SSL formulation, we redefine the concept of margins to pseudo-margins of unlabeled examples since no ground truth labels are available for the unlabeled data. Let $c$ be the pseudo-label (or the argmax of the prediction, i.e., $\hat{p}_{\theta}(y|\pi(\hat{x}))$ ) at iteration $t$ on unlabeled example $\hat{x}$ after applying weak augmentations. We define the pseudo-margin (PM) of $\hat{x}$ with respect to pseudo-label $c$ at iteration $t$ as follows: + +$$ +\mathrm {P M} _ {c} ^ {t} (\hat {x}) = z _ {c} - \max _ {c! = i} (z _ {i}) \tag {5} +$$ + +where $z_{c}$ is the logit corresponding to the assigned pseudo-label $c$ and $\max_{c! = i}(z_i)$ is the largest other logit corresponding to a label $i$ different from $c$ . To monitor the model's predictions on $\hat{x}$ with respect to pseudo-label $c$ from the beginning of training to iteration $t$ , we average all the margins with respect to $c$ from the first iteration until $t$ and obtain the average pseudo-margin (APM) as follows: + +$$ +\mathrm {A P M} _ {c} ^ {t} (\hat {x}) = \frac {1}{t} \sum_ {j = 1} ^ {t} \mathrm {P M} _ {c} ^ {j} (\hat {x}) \tag {6} +$$ + +Here $c$ acts as the "ground truth" label for the APM calculation. Note that if at a prior iteration $t'$ , the assigned pseudo-label is different from $c$ (say $c'$ ), then the APM calculation at iteration $t'$ is done with respect to $c'$ (by averaging all margins with respect to $c'$ from 1 to $t'$ ). In practice, we maintain a vector of pseudo-margins for all classes accumulated over the training iterations and dynamically retrieve the accumulated pseudo-margin value of the argmax class $c$ to obtain the $\mathrm{APM}_c^t$ at iteration $t$ . + +Intuitively, if $c$ is the pseudo-label of $\hat{x}$ at iteration $t$ , then $\mathrm{PM}_c^t$ with respect to class $c$ at iteration $t$ will be positive. In contrast, if the argmax of the model prediction on $\hat{x}$ at a previous iteration $t' < t$ is different from $c$ , then $PM_c^{t'}$ at $t'$ + +Algorithm 1 MarginMatch +Require: Labeled data $L$ ; unlabeled data $U$ ; erroneous examples $E$ ; maximum number of iterations $T$ ; number of classes $C + 1$ ( $C$ original classes plus one virtual class of erroneous examples); $\theta$ model; $\pi$ weak augmentations; $\Pi$ strong augmentations. +1: Initialize the Average Pseudo-Margin (APM) threshold $\gamma^1$ at the first iteration to a small value (e.g., $\gamma^1 = -\infty$ ). +2: for $t = 1$ to $T$ do +3: Estimate learning status $\alpha_c$ (using Eq. 2) and calculate the class-wise flexible thresholds $\mathcal{T}_c^t$ (using Eq. 3) for each class $c$ . +4: while $U$ not exhausted do +5: Labeled batch $L_b = \{(x_1, y_1), \dots, (x_B, y_B)\}$ , unlabeled batch $U_b = \{\hat{x}_1, \dots, \hat{x}_{\nu B}\}$ , erroneous (or mislabeled) batch $E_b = \{(\tilde{x}_1, C + 1), \dots, (\tilde{x}_B, C + 1)\}$ +6: for $x \in U_b \cup E_b$ do +7: Compute logits $z_c$ for each class $c$ after applying weak augmentations when $x \in U_b$ and strong augmentations when $x \in E_b$ . +8: Calculate pseudo-margin $PM_c^t$ (using Eq. 5) and update Average $PM_c^t$ (using Eq. 6) for each $c = 1$ to $C + 1$ . +9: end for +10: Minimize $\mathcal{L} = \mathcal{L}_s + \lambda (\mathcal{L}_u + \mathcal{L}_e)$ +11: $\mathcal{L}_s = \frac{1}{B} \sum_{i=1}^{B} H(y_i, p_\theta(y|\pi(x_i)))$ +12: $\mathcal{L}_u = \sum_{i=1}^{\nu B} \mathbb{1}(\mathrm{AM}_{p_\theta(y|\pi(\hat{x}_i))}^t (\hat{x}_i) > \gamma^t) \times \mathbb{1}(\max(p_\theta(y|\pi(\hat{x}_i))) > T_{p_\theta(y|\pi(\hat{x}_i))}^t) \times H(\hat{p}_\theta(y|\pi(\hat{x}_i)), p_\theta(y|\Pi(\hat{x}_i)))$ +13: $\mathcal{L}_e = \sum_{i=1}^{B} H(C + 1, p_\theta(y|\Pi(\tilde{x}_i)))$ +14: end while +15: Update $\gamma^{t+1}$ as the $95^{th}$ percentile erroneous sample $APM_{C+1}^t$ . +16: end for + +with respect to $c$ will be negative. Therefore, if over the iterations, the model predictions do not agree frequently with the pseudo-label $c$ from iteration $t$ and the model fluctuates significantly between iterations on the predicted label, the APM for class $c$ will have a low, likely negative value. Similarly, if the model is highly uncertain of the class of $\hat{x}$ (reflected in a high entropy of the class probability distribution), the APM for class $c$ will have a low value. These capture the characteristics of unlabeled examples or of those harmful for training. Motivated by these observations, MarginMatch leverages the APM of the assigned pseudo-label $c$ and compares it with an APM threshold to mask out pseudo-labeled examples with low APMs. Formally, the unlabeled loss in MarginMatch is: + +$$ +\begin{array}{l} \mathcal {L} _ {u} = \sum_ {i = 1} ^ {\nu B} \mathbb {1} \left(\mathrm {A M} _ {\hat {p} _ {\theta} (y | \pi (\hat {x} _ {i}))} ^ {t} (\hat {x} _ {i}) > \gamma^ {t}\right) \quad \times \\ \mathbb {1} \left(\max \left(p _ {\theta} \left(y \mid \pi (\hat {x} _ {i})\right)\right) > \mathcal {T} _ {\hat {p} _ {\theta} \left(y \mid \pi (\hat {x} _ {i})\right)} ^ {t}\right) \times \\ H \left(\hat {p} _ {\theta} (y | \pi (\hat {x} _ {i})), p _ {\theta} (y | \Pi (\hat {x} _ {i}))\right) \tag {7} \\ \end{array} +$$ + +where $\gamma^t$ is the APM threshold at iteration $t$ , estimated as explained below, and $\mathcal{T}_{\hat{p}_0(y|\pi(\hat{x}_i))}^t$ is the flexible threshold estimated as in FlexMatch [49]. To train our model, we adopt the best practices [41, 49] and optimize the weighted combination of the supervised and unsupervised losses: + +$$ +\mathcal {L} = \mathcal {L} _ {s} + \lambda \mathcal {L} _ {u} \tag {8} +$$ + +where the supervised loss is given by: + +$$ +\mathcal {L} _ {s} = \frac {1}{B} \sum_ {i = 1} ^ {B} H \left(y _ {i}, p _ {\theta} (y \mid \pi \left(x _ {i}\right))\right) \tag {9} +$$ + +Average Pseudo-Margin Threshold Estimation Inspired by Pleiss et al. [33], we propose to estimate the average pseudo-margin threshold $\gamma^t$ by analyzing the training dynamics of a special category of unlabeled examples, which we force to be erroneous or mislabeled examples. That is, to create the sample of erroneous examples $E$ , we randomly sample a subset of unlabeled examples from $U$ that we assign to an inexistent (or virtual) class $C + 1$ at the beginning of the training process and remove them from $U$ . The purpose of these erroneous examples is to mimic the training dynamics of incorrectly pseudo-labeled (unlabeled) examples and use them as proxy to estimate the cutoff of (potentially) mislabeled pseudo-labels. Since the examples in $E$ should belong to one of the $C$ original classes, assigning them to the inexistent class $C + 1$ makes them by definition mislabeled (see Appendix A for additional insights into this virtual class). As with all unlabeled examples from $U$ , we compute $APM_{C + 1}^t$ for the special category of erroneous examples from $E$ , but unlike the unlabeled examples from $U$ , the erroneous ones from $E$ have a fixed class $C + 1$ . To mimic the training dynamics of unlabeled examples from $U$ , we use strong augmentations to compute the loss of the erroneous examples from $E$ . That is, given a batch $E_b$ of $B$ erroneous examples, the erroneous sample loss becomes: + +$$ +\mathcal {L} _ {e} = \sum_ {i = 1} ^ {B} H (C + 1, p _ {\theta} (y | \Pi (\tilde {x} _ {i}))) \tag {10} +$$ + +At iteration $t$ , we use the APMs of the erroneous examples to choose the APM threshold $\gamma^t$ . We set $\gamma^t$ as the APM of the $95^{th}$ percentile erroneous sample. The total loss becomes: + +$$ +\mathcal {L} = \mathcal {L} _ {s} + \lambda \left(\mathcal {L} _ {u} + \mathcal {L} _ {e}\right) \tag {11} +$$ + +Our full MarginMatch algorithm is shown in Algorithm 1. + +
DatasetCIFAR-10CIFAR-100SVHNSTL-10
#Labels/Class425400425100425100425100
Pseuso-Labeling74.610.2646.492.2015.080.1987.450.8557.740.2836.550.2464.615.6025.212.039.400.3274.680.9955.452.4332.640.71
UDA10.793.755.320.064.410.0748.951.5929.430.2123.870.235.344.274.260.391.950.0137.828.449.811.156.810.17
MixMatch45.242.1512.761.147.130.3462.152.1741.511.1928.160.2446.181.783.980.173.501.1334.151.548.950.3210.410.73
ReMixMatch5.270.194.850.134.040.1247.150.7627.140.2323.780.124.230.313.180.041.940.0631.510.758.540.486.190.24
FixMatch7.80.284.910.054.250.0848.210.8229.450.1622.890.123.971.183.131.031.970.0338.434.1410.451.046.430.33
FlexMatch5.040.065.040.094.190.0139.991.6226.960.0822.440.158.193.207.782.556.720.3029.151.328.230.135.770.12
MarginMatch4.910.074.730.123.980.0236.971.3223.710.1321.390.123.751.203.141.171.930.0125.373.587.310.355.520.15
+ +Table 1. Test error rates on CIFAR-10, CIFAR-100, SVHN, and STL-10 datasets. Best results are shown in blue. + +Exponential Moving Average of Pseudo-Margins The current definition of APM weighs the pseudo-margin at iteration $t$ identical to the pseudo-margin at a much earlier iteration $p$ ( $t >> p$ ). This is problematic since very old pseudo-margins eventually become deprecated (especially due to the large number of iterations through unlabeled data in consistency training ( $\sim 9K$ ), and hence, the old margins are no longer indicative of the current learning status of the model. To this end, instead of averaging all pseudo-margins (from the beginning of training to the current iteration), we propose to use an exponential moving average to place more importance on recent iterations. Formally, APM becomes: + +$$ +\mathrm {A P M} _ {c} ^ {t} (\hat {x}) = \mathrm {P M} _ {c} ^ {t} (\hat {x}) * \frac {\delta}{1 + t} + \mathrm {A P M} _ {c} ^ {t - 1} (\hat {x}) * \left(1 - \frac {\delta}{1 + t}\right) \tag {12} +$$ + +We set the smoothing parameter $\delta$ to 0.997 in experiments. + +# 3. Experiments + +We evaluate the performance of our MarginMatch on a wide range of SSL benchmark datasets. Specifically, we perform experiments with various numbers of labeled examples on CIFAR-10, CIFAR-100 [21], SVHN [31], STL-10 [8], ImageNet [10], and WebVision [26]. For smaller scale datasets such as CIFAR-10, CIFAR-100, SVHN, and STL-10 we randomly sample a small number of labeled examples per class (ranging from 4 labels per class to 400 labels per class) and treat them as labeled data, whereas the remaining labeled examples are treated as unlabeled data, except for STL-10 [8], which provides its own set of unlabeled examples. On ImageNet and WebVision, we use $\sim 10\%$ of the available labeled examples as labeled data, with the remaining $(90\%)$ being treated as unlabeled data. In all our experiments, we sample $5\%$ of the unlabeled data and place it in the set of erroneous examples. + +We report the mean and standard deviation of error rates from five runs with different parameter initializations. Similar to FixMatch [41], we use Wide Residual Networks [48]: WRN-28-2 for CIFAR-10 and SVHN; WRN-28-8 for CIFAR-100; and WRN-37-2 for STL-10. We use ResNet-50 [14] for both ImageNet and WebVision. + +In our experiments, we adopt the same hyperparameters as FixMatch [41]. Specifically, we use stochastic gradient + +descent (SGD) with a momentum of 0.9. We start with a learning rate of 0.03 and employ a cosine learning rate; at iteration $k$ , our learning rate is $\eta(k) = \cos\left(\frac{7k\pi}{16K}\right)$ , where $K$ is the maximum number of iterations and is set to $2^{20}$ . We also leverage the same data augmentations as in FlexMatch [49]. Specifically, for weak augmentations we employ a standard flip-and-shift augmentation and use RandAugment [9] for strong augmentations. We set the batch size $B = 64$ , ratio of unlabeled to labeled data in a batch to $\nu = 7$ and weigh the supervised and unsupervised losses equally (i.e., $\lambda = 1$ ). We set our initial confidence threshold to $\tau = 0.95$ and our average pseudo-margin (APM) threshold to the APM of the $95^{th}$ percentile threshold sample. To report the error rates, we compare all the approaches using the model at the end of training as in FixMatch [41]. + +# 3.1. CIFAR-10, CIFAR-100, SVHN, and STL-10 + +We compare MarginMatch against strong baselines and prior works: Pseudo-Labeling [1], Unsupervised Data Augmentation (UDA) [46], MixMatch [5], ReMixMatch [4], FixMatch [41], and FlexMatch [49]. We show in Table 1 the error rates obtained by our MarginMatch and the baselines on the CIFAR-10, CIFAR-100, SVHN and STL-10 datasets. First, we observe that our approach improves the performance on both CIFAR-10 and CIFAR-100. On CIFAR-10, MarginMatch improves performance in all data regimes upon FlexMatch [49], which is the current state-of-the-art, while maintaining a good error rate standard deviation. On CIFAR-100, which is significantly more challenging than CIFAR-10, we observe that MarginMatch bring substantially larger improvements. Notably, we see $3.02\%$ improvement over FlexMatch in error rate using only 4 labels per class, and $3.25\%$ improvement using 25 examples per class. These results on CIFAR-100 emphasize the effectiveness of MarginMatch, which performs well on a more challenging dataset. + +On SVHN, our approach performs better than FixMatch using 4 labels per class and performs similarly with FixMatch using 25 and 100 labels per class. However, on this dataset, MarginMatch performs much better compared with FlexMatch. For example, MarginMatch achieves $3.75\%$ error rate using 4 labels per class, whereas FlexMatch obtains an error rate of $8.19\%$ with the same labels per class, + +![](images/3863b902bc94de3c24819d9343ec3c7e0f43b2934fb2ad8b2d6946347746db10.jpg) +Figure 2. Convergence speed of MarginMatch against FixMatch and FlexMatch with 4 labels per class. + +![](images/f0f4a2e72ba307d2e520de4a90e1334b6cb7289c8ab771c2357bf833a851585e.jpg) + +![](images/67c934250c290cbd95125a37db084ce3f4ae34da39c9f55bcc23d631dbdd9db4.jpg) + +![](images/6b5dee9f559994b4f23f9e72ebdbe3369039e06c5226b62e9377d21b528f02ad.jpg) + +
DatasetImageNetWebVision
TOP-1TOP-5TOP-1TOP-5
Supervised48.3925.4949.5826.78
FixMatch43.6621.8044.7622.65
FlexMatch42.0219.4943.8722.07
MarginMatch41.0518.2843.0821.13
+ +yielding an improvement of MarginMatch of $4.44\%$ over FlexMatch. We hypothesize that the low performance of FlexMatch is due to its limitation in handling unbalanced class distributions [49]. On STL-10, MarginMatch as well outperforms all the other approaches both in error rates and error rate standard deviation. Notably, on this dataset, our approach pushes the performance of FlexMatch by $3.78\%$ in error rate using only 4 labels per class and by $0.92\%$ using 25 labels per class. + +Next, we compare MarginMatch with FixMatch and FlexMatch in terms of convergence speed in the extremely labelscarcse setting of 4 labels per class and show these results in Figure 2. Notably, we observe that MarginMatch has a similar convergence speed (or even better on CIFAR-100) compared with FlexMatch while achieving a lower test error rate than FlexMatch on all datasets with 4 labels per class (see Table 1). Even more strikingly, compared with FixMatch, MarginMatch has a much superior convergence speed for a much better test error rate with 4 labels per class. This is because the rigid thresholds in FixMatch allow access only to a small amount of unlabeled data for training at each iteration and it takes a lot longer for the model to train. + +# 3.2. ImageNet and WebVision + +To showcase the effectiveness of our approach in a large-scale setup, we test our MarginMatch on ImageNet [10] and WebVision [26] using $10\%$ labeled examples in total. We show the results obtained in Table 2. We observe that our MarginMatch outperforms FixMatch and FlexMatch on both datasets. It is worth noting that large-scale self-supervised approaches such as SimCLR [7] achieve high performance on ImageNet but at a much higher computational cost. + +Table 2. Test error rates on the ImageNet and WebVision datasets. Best results are shown in blue. + +
δ0.950.990.9950.9970.9991
ERR RATE38.1338.0537.9237.9139.1239.72
+ +Table 3. Error rates obtained on CIFAR-100 with four examples per class and various smoothing values $\delta$ . Best result is in blue. + +MarginMatch outperforms other SSL methods using the same ResNet-50 architecture at the same computational cost. We emphasize MarginMatch is most successful and relevant in low data regimes on smaller datasets. + +# 4. Ablation Study + +Exponential Moving Average Smoothing for APM Computation In our approach, we employ an exponential moving average (EMA) of the pseudo-margin values with a smoothing value of $\delta = 0.997$ to compute the APM. We now analyze how our approach performs with different EMA smoothing values or with no EMA at all. Table 3 shows these results on CIFAR-100 with 4 labels per class. First, we observe that employing a simple average of pseudo-margin values for the APM computation (i.e., $\delta = 1$ ) performs extremely poorly, obtaining a $39.72\%$ error rate. This result emphasizes that margins eventually become deprecated and it is essential to scale them down in time. Using a low smoothing factor of $\delta = 0.95$ is not effective either, denoting that abruptly forgetting margin values does not work either. Our chosen $\delta = 0.997$ strikes a balance between the two by eliminating the harmful effects of very old margins while keeping track of a good amount of previous estimates (e.g., a margin value computed 200 epochs previously is scaled down by 0.55, while a margin value computed 1000 epochs previously is scaled by 0.05). + +Pseudo-Margin vs. Other Measures for Pseudo-Label Correctness Our MarginMatch monitors the pseudomargins of a model's predictions across training iterations to ensure the quality of pseudo-labels. However, other measures such as confidence or entropy exist that can assess the pseudo-label correctness. Hence, we perform an ablation where we replace the pseudo-margins in our MarginMatch + +
DatasetCIFAR-10CIFAR-100SVHNSTL-10
#Labels/Class425400425100425100425100
Avg Confidence23.872.7314.211.377.540.7841.232.1531.491.4824.112.368.994.276.540.394.730.0131.678.4414.871.157.590.17
Avg Entropy8.580.416.180.155.850.1245.100.9126.021.1122.130.2515.691.2512.740.789.330.0529.543.5110.631.3510.840.47
Avg Margin7.250.295.380.764.730.0939.721.5225.210.5223.180.1718.451.3611.290.938.400.0428.454.289.341.347.590.21
EMA Confidence4.910.454.740.093.990.0638.670.7425.610.1221.480.173.840.233.250.031.930.0925.90.817.60.425.740.57
EMA Entropy6.40.438.340.124.210.0941.630.7636.840.1322.520.073.811.263.170.872.140.0427.214.058.281.016.790.27
EMA Margin4.910.074.730.123.980.0236.971.3223.710.1321.390.123.751.203.141.171.930.0125.373.587.310.355.520.15
+ +Table 4. Test error rates comparing pseudo-margin with confidence and entropy. Best results are shown in blue. + +![](images/5d7a6f0b6f27210d2b1e6974fc18acd3a020cb78d91155ca3f6a1f383e9031af.jpg) +Figure 3. Mask rate and impurity on CIFAR-100 with 4 labeled examples per class. + +![](images/da786527f8c4ef697b2a118d71e40212692d47c7e16ad1ca6e0824b5a67174a2.jpg) + +with average confidence and entropy and compare their performance. Specifically, we design the following approaches: 1) Avg Confidence monitors the confidence of the prediction for each unlabeled example and takes the average over the training iterations; 2) Avg Entropy monitors the entropy of the class probability distribution for each unlabeled example and takes the average across the training iterations. In addition, we also consider 3) EMA Confidence and 4) EMA Entropy which are similar to Avg Confidence and Avg Entropy, respectively, but use an exponential moving average (EMA) instead of the simple averaging. The estimation of the threshold for each of these approaches is done in a similar manner as for pseudo-margins, using erroneous samples and considering the value of the 95th percentile erroneous sample as the threshold. + +We show the results obtained using these approaches in Table 4. First, we observe that all measures (pseudo-margin, confidence and entropy) with EMA perform better than their counterpart with simple averaging. Second, EMA Margin achieves the lowest test error rates compared with EMA Confidence and EMA Entropy. Thus, we conclude that pseudo-margins provide an excellent measure for pseudolabel correctness. See Appendix B for some additional insights into why EMA Margin outperforms EMA confidence and entropy. + +# 5. Analysis + +# 5.1. Mask Rate and Impurity + +We now contrast MarginMatch with FixMatch and FlexMatch in terms of the quality of pseudo-labels using two metrics: mask rate and impurity and show these results in + +Figure 3, respectively, using CIFAR-100 with 4 labels per class. Mask rate is defined as the fraction of pseudo-labeled examples that do not participate in the training at epoch $t$ due to confidence masking or pseudo-margin masking (or both). Impurity in contrast is defined as the fraction of pseudolabeled examples that do participate in the training at epoch $t$ but with a wrong label. An effective SSL model minimizes both metrics: a low mask rate indicates that the model has access to more unlabeled examples during training (otherwise a low percentage and less diverse set of unlabeled examples are seen during training) while low impurity indicates that the pseudo-labels of these examples are of high quality. Note that we can compute impurity on these two datasets because our unlabeled data comes from the labeled training set of each of these datasets (thus we compare the pseudo-labels against the gold labels of each dataset). + +As can be seen from the figures, FixMatch has a significantly larger mask rate due to the rigid confidence threshold set to a high value of 0.95. In contrast, FlexMatch lowers the mask rate by $5\%$ with the introduction of flexible thresholds, but has a much higher impurity compared with FixMatch. Notably, our MarginMatch has only a slightly higher mask rate compared with FlexMatch and at the same time achieves a much lower impurity than FlexMatch and even FixMatch despite that FixMatch employs a very high confidence threshold. These results show that MarginMatch that enforces an additional measure for pseudo-labeled data quality maintains a low mask rate without compromising the quality of the pseudo-labels (i.e., low mask rate and low impurity). + +# 5.2. Anecdotal Evidence + +We show in Figure 4 anecdotal evidence of the effectiveness of MarginMatch. To this end, we extract two bird images from our unlabeled portion of CIFAR-10 [21] of various learning difficulties that resemble characteristics of plane images (e.g., the background). The top part of the figure illustrates the progression over the training iterations of the confidence and the confidence thresholds of FlexMatch for the classes bird and plane, whereas the bottom part of the figure illustrates the progression of the APM threshold of MarginMatch along with its APMs of bird and plane classes over the training iterations. In the rightmost image, for MarginMatch we can observe that the APM of the bird class becomes stronger and stronger as the training progresses and + +![](images/09309a144bfd5ba04043d61348a201f6da75b5cbced2a082c1713cf8cca3df75.jpg) + +![](images/88b8e4a329bb9009cfef05f34bd5bd8d478b95e290408b600afa6990ba38710c.jpg) +Figure 4. Confidence thresholding vs. APM Thresholding on two images from the CIFAR-10 dataset. + +eventually exceeds the APM threshold of MarginMatch, and hence, the image is included in the training until the end with the correct bird label. Interestingly, for the same image, in FlexMatch the confidence for the bird class is very close to the bird confidence threshold and eventually falls below this threshold and exits the training set. In contrast, the leftmost image is significantly more challenging than the rightmost image since it is more similar to plane images, which makes it an easily confusable example. Here, we observe that the confidence of FlexMatch exceeds the flexible threshold with the incorrect argmax class plane starting from iteration 3000. Moreover, Flexmatch continues to use this image with the wrong plane label for the remaining of the training process. Critically, in MarginMatch the APM value for the plane class does not exceed the APM threshold, and the model eventually learns to classify this image correctly and includes it in training with the correct bird pseudo-label. + +# 6. Related Work + +Here, we focus on various SSL approaches that our MarginMatch directly builds upon although there are other SSL techniques that are not presented in this review, such as approaches based on generative models [17, 24], graph-based approaches [19, 20] and robust SSL [32, 37] (see Appendix C for a comparison between MarginMatch and Robust SSL approaches). + +Self-training [28, 36, 40, 47] is a popular SSL method where the predictions of a model on unlabeled data are used as artificial labels to train against. Noisy student training [47] is a popular self-training approach that also leverages knowledge distillation [16] and iteratively jointly trains two models in a teacher-student framework. Noisy student uses a + +larger model size and noised inputs, exposing the student to more difficult learning environments, leading to an increased performance compared to the teacher. + +Pseudo-labeling is a variant of self-training where these predictions are sharpened to obtain hard labels [25]. The use of hard labels can be seen as a means of entropy minimization [12] and nowadays is a valuable component in most successful SSL approaches [4, 5, 49]. These hard labels are usually used along a confidence threshold, where unconfident unlabeled examples are completely disregarded (e.g., [5]) to avoid using noisy pseudo-labels. Recently, approaches such as Curriculum Labeling (CL) [6] or FlexMatch [49] started to explore curriculum learning in the SSL context. CL proposes a self-pacing strategy of identifying easy and hard examples to ensure that the model first uses easy and progressively moves towards harder examples. Similarly, MarginMatch uses curriculum learning and pseudo-labeling, but the focus of our approach is placed on producing better thresholds for assessing the quality of pseudo-labels. + +Consistency regularization [2] is a method that applies random perturbations when generating the artificial label, such as data augmentation [4, 5, 41], dropout [39], or adversarial perturbations [29]. Current state-of-the-art approaches [41, 49] exploit a combination of weak and strong data augmentations, which were shown to be extremely beneficial in SSL. The most popular strong augmentations used in the SSL literature are RandAugment [9] and CTAugment [4]. The approaches based on these methods first generate a hard label using pseudo-labeling on a weakly augmented image (i.e., using a low noise transformation such as a flip-and-shift augmentation), then optimize the predictions of the model on a strongly augmented version of the same image towards this hard label. Similar to these approaches, MarginMatch uses the same combination of weak and strong data augmentations. + +# 7. Conclusion + +In this paper, we proposed a novel semi-supervised learning method that improves the pseudo-label quality using training dynamics. Our new method is lightweight and achieves state-of-the-art performance on four computer vision SSL datasets in low data regimes and on two large-scale benchmarks. MarginMatch takes into consideration not only a flexible confidence threshold to account for the difficulty of each class, but also a measure of quality for each unlabeled example using training dynamics. In addition, MarginMatch is a general approach that can be leveraged in most SSL frameworks and we hope that it can attract future research in analyzing the effectiveness of SSL approaches focused on data quality. As future work, we aim to further explore our method in settings when there is a mismatch between the labeled and unlabeled data distributions (i.e., making use of out-of-domain unlabeled data). + +# References + +[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. CoRR, abs/1908.02983, 2019. 1, 5 +[2] Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2, 8 +[3] Peter L. Bartlett, Dylan J. Foster, and Matus Telgarsky. Spectrally-normalized margin bounds for neural networks. CoRR, abs/1706.08498, 2017. 1, 3 +[4] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 1, 5, 8 +[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1, 5, 8 +[6] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In AAAI, 2021. 8 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6 +[8] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 2, 5 +[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 5, 8 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 5, 6 +[11] Gamaleldin Fathy Elsayed, Dilip Krishnan, Hossein Mobahi, Kevin Regan, and Samy Bengio. Large margin deep networks for classification. 2018. 1, 3 +[12] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In L. Saul, Y. Weiss, and L. Bottou, editors, Advances in Neural Information Processing Systems, volume 17. MIT Press, 2004. 8 +[13] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1321-1330. PMLR, 06-11 Aug 2017. 3 + +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 5 +[15] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. 1 +[16] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 8 +[17] Geoffrey E Hinton and Russ R Salakhutdinov. Using deep belief nets to learn covariance kernels for gaussian processes. In J. Platt, D. Koller, Y. Singer, and S. Roweis, editors, Advances in Neural Information Processing Systems, volume 20. Curran Associates, Inc., 2007. 8 +[18] Yiding Jiang, Dilip Krishnan, Hossein Mobahi, and Samy Bengio. Predicting the generalization gap in deep networks with margin distributions, 2018. 1, 3 +[19] Thorsten Joachims. Transductive inference for text classification using support vector machines. In Proceedings of the Sixteenth International Conference on Machine Learning, ICML '99, page 200-209, San Francisco, CA, USA, 1999. Morgan Kaufmann Publishers Inc. 8 +[20] Thorsten Joachims. Transductive learning via spectral graph partitioning. In Proceedings of the Twentieth International Conference on International Conference on Machine Learning, ICML'03, page 290-297. AAAI Press, 2003. 8 +[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5, 7 +[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, NIPS'12, page 1097-1105, Red Hook, NY, USA, 2012. Curran Associates Inc. 1 +[23] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR (Poster). OpenReview.net, 2017. 1, 2 +[24] J.A. Lasserre, C.M. Bishop, and T.P. Minka. Principled hybrids of generative and discriminative models. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 1, pages 87-94, 2006. 8 +[25] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 2, 8 +[26] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. CoRR, abs/1708.02862, 2017. 2, 5, 6 +[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pages 181-196, 2018. 1 + +[28] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152-159, New York City, USA, June 2006. Association for Computational Linguistics. 8 +[29] Takeru Miyato, Andrew M Dai, and Ian Goodfellow. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725, 2016. 8 +[30] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1 +[31] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 2, 5 +[32] Jongjin Park, Sukmin Yun, Jongheon Jeong, and Jinwoo Shin. Opencos: Contrastive semi-supervised learning for handling open-set unlabeled data. In ECCV Workshops, 2021. 8 +[33] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying mislabeled data using the area under the margin ranking. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 17044-17056. Curran Associates, Inc., 2020. 1, 3, 4 +[34] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1 +[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683, 2019. 1 +[36] Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. In 2005 Seventh IEEE Workshops on Applications of Computer Vision (WACV/MOTION'05) - Volume 1, volume 1, pages 29-36, 2005. 1, 8 +[37] Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set semi-supervised learning with open-set consistency regularization. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 25956-25967. Curran Associates, Inc., 2021. 8 +[38] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Mutual exclusivity loss for semi-supervised deep learning. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1908-1912. IEEE, 2016. 1 +[39] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29:1163-1171, 2016. 1, 2, 8 +[40] H. Scudder. Probability of error of some adaptive pattern-recognition machines. IEEE Transactions on Information Theory, 11(3):363-371, 1965. 8 +[41] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, + +Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 596-608. Curran Associates, Inc., 2020. 1, 2, 4, 5, 8 +[42] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 1 +[43] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 1 +[44] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in neural information processing systems, pages 1195–1204, 2017. 1 +[45] Vikas Verma, Kenji Kawaguchi, Alex Lamb, Juho Kannala, Arno Solin, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi-supervised learning. Neural Networks, 2021. 1 +[46] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc., 2020. 1, 5 +[47] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687-10698, 2020. 1, 8 +[48] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks, 2017. 5 +[49] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 18408-18419. Curran Associates, Inc., 2021. 1, 2, 3, 4, 5, 6, 8 +[50] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. CoRR, abs/1611.03530, 2016. 3 \ No newline at end of file diff --git a/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/images.zip b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..45bd08a07eadce057058041126a1c0a6b47438fe --- /dev/null +++ b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f524965a5f3f00a9335b3e98e0894647a952c18d9d147cda499c6607a559badf +size 489001 diff --git a/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/layout.json b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a619b1814a828f891567d35afaef9b573c7fc68 --- /dev/null +++ b/2023/MarginMatch_ Improving Semi-Supervised Learning with Pseudo-Margins/layout.json @@ -0,0 +1,9703 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 529, + 121 + ], + "type": "text", + "content": "MarginMatch: Improving Semi-Supervised Learning with Pseudo-Margins" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 144, + 394, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 144, + 394, + 185 + ], + "spans": [ + { + "bbox": [ + 198, + 144, + 394, + 185 + ], + "type": "text", + "content": "Tiberiu Sosea Cornelia Caragea University of Illinois Chicago tsosea2@uic.edu cornelia@uic.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "content": "We introduce MarginMatch, a new SSL approach combining consistency regularization and pseudo-labeling, with its main novelty arising from the use of unlabeled data training dynamics to measure pseudo-label quality. Instead of using only the model's confidence on an unlabeled example at an arbitrary iteration to decide if the example should be masked or not, MarginMatch also analyzes the behavior of the model on the pseudo-labeled examples as the training progresses, to ensure low quality predictions are masked out. MarginMatch brings substantial improvements on four vision benchmarks in low data regimes and on two large-scale datasets, emphasizing the importance of enforcing high-quality pseudo-labels. Notably, we obtain an improvement in error rate over the state-of-the-art of " + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "inline_equation", + "content": "3.25\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "content": " on CIFAR-100 with only 25 labels per class and of " + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "inline_equation", + "content": "3.78\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "content": " on STL-10 using as few as 4 labels per class. We make our code available at https://github.com/tsosea2/MarginMatch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 128, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 128, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 128, + 464 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 473, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 616 + ], + "type": "text", + "content": "Deep learning models have seen tremendous success in many vision tasks [14, 22, 27, 42, 43]. This success can be attributed to their scalability, being able to produce better results when they are trained on large datasets in a supervised fashion [15, 27, 34, 35, 43, 47]. Unfortunately, large labeled datasets annotated for various tasks and domains are difficult to acquire and demand considerable annotation effort or domain expertise. Semi-supervised learning (SSL) is a powerful approach that mitigates the requirement for large labeled datasets by effectively making use of information from unlabeled data, and thus, has been studied extensively in vision [4, 5, 23, 25, 30, 36, 38, 39, 44-46]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Recent SSL approaches integrate two important components: consistency regularization [46, 49] and pseudolabeling [25]. Consistency regularization works on the assumption that a model should output similar predictions when fed perturbed versions of the same image, whereas pseudo-labeling uses the model's predictions of unlabeled examples as labels to train against. For example, Sohn et al. [41] introduced FixMatch that combines consistency reg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 213, + 548, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 548, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 548, + 549 + ], + "type": "text", + "content": "ularization on weak and strong augmentations with pseudolabeling. FixMatch relies heavily on a high-confidence threshold to compute the unsupervised loss, disregarding any pseudo-labels whose confidence falls below this threshold. While training using only high-confidence pseudo-labels has shown to consistently reduce the confirmation bias [1], this rigid threshold allows access only to a small amount of unlabeled data for training, and thus, ignores a considerable amount of unlabeled examples for which the model's predictions do not exceed the confidence threshold. More recently, Zhang et al. [49] introduced FlexMatch that relaxes the rigid confidence threshold in FixMatch to account for the model's learning status of each class in that it adaptively scales down the threshold for a class to encourage the model to learn from more examples from that class. The flexible thresholds in FlexMatch allow the model to have access to a much larger and diverse set of unlabeled data to learn from, but lowering the thresholds can lead to the introduction of wrong pseudo-labels, which are extremely harmful for generalization. Interestingly, even when the high-confidence threshold is used in FixMatch can result in wrong pseudo-labels. See Figure 1 for incorrect pseudo-labels detected in the training set after we apply FixMatch and FlexMatch on ImageNet. We posit that a drawback of FixMatch and FlexMatch and in general of any pseudo-labeling approach is that they use the confidence of the model only at the current iteration to enforce quality of pseudo-labels and completely ignore model's predictions at prior iterations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 548, + 713 + ], + "type": "text", + "content": "In this paper, we propose MarginMatch, a new SSL approach that monitors the behavior of the model on the unlabeled examples as the training progresses, from the beginning of training until the current iteration, instead of using only the model's current belief about an unlabeled example (i.e., its confidence at the current iteration) to decide if the example should be masked or not. We estimate a pseudo-label's contribution to learning and generalization by introducing pseudo-margins of unlabeled examples averaged across training iterations. Pseudo-margins of unlabeled examples extend the margins from machine learning [3, 11, 18, 33] which provide a measure of confidence of the outputs of the model and capture the difference between the output for the correct" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15773" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 535, + 232 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 535, + 232 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 535, + 232 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 535, + 232 + ], + "type": "image", + "image_path": "0d697b74530a6eb1d50b7f27f7764bd5484004a26b834d52b3cc2faab6dd374d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 241, + 520, + 253 + ], + "lines": [ + { + "bbox": [ + 70, + 241, + 520, + 253 + ], + "spans": [ + { + "bbox": [ + 70, + 241, + 520, + 253 + ], + "type": "text", + "content": "Figure 1. Incorrect pseudo-labels propagated until the end of the training process for FixMatch and FlexMatch on ImageNet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 289, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 405 + ], + "type": "text", + "content": "(gold) label and the other labels. In our case, the pseudomargins capture how much larger the assigned logit (the logit corresponding to the argmax of the model's prediction) is compared with all other logits at iteration " + }, + { + "bbox": [ + 46, + 274, + 289, + 405 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 274, + 289, + 405 + ], + "type": "text", + "content": ". Similar to FlexMatch, in MarginMatch we take advantage of the flexible confidence thresholds to allow the model to learn from larger and more diverse sets of unlabeled examples, but unlike FlexMatch, we train the model itself to identify the characteristics of mislabeled pseudo-labels simply by monitoring the model's training dynamics on unlabeled data over the iterations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "text", + "content": "We carry out comprehensive experiments using established SSL experimental setups on CIFAR-10, CIFAR-100 [21], SVHN [31], STL-10 [8], ImageNet [10], and WebVision [26]. Despite its simplicity, our findings indicate that MarginMatch produces improvements in performance over strong baselines and prior works on all datasets at no additional computational cost. Notably, compared to current state-of-the-art, on CIFAR-100 we see " + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "inline_equation", + "content": "3.02\\%" + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "text", + "content": " improvement in error rate using only 4 labels per class and " + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "inline_equation", + "content": "3.78\\%" + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "text", + "content": " improvement on STL-10 using the same extremely label-scarce setting of 4 labels per class. In addition, on ImageNet [10] and WebVision [26] we find that MarginMatch pushes the state-of-the-art error rates by " + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "inline_equation", + "content": "0.97\\%" + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "text", + "content": " on ImageNet and by " + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "inline_equation", + "content": "0.79\\%" + }, + { + "bbox": [ + 46, + 406, + 289, + 573 + ], + "type": "text", + "content": " on WebVision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 574, + 193, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 574, + 193, + 584 + ], + "spans": [ + { + "bbox": [ + 59, + 574, + 193, + 584 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 589, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 589, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 287, + 685 + ], + "type": "text", + "content": "1. We introduce a new SSL approach which we call MarginMatch that enforces high pseudo-label quality during training. Our approach allows access to a large set of unlabeled data to learn from (thus, incorporating more information from unlabeled data) and, at the same time, monitors the training dynamics of unlabeled data as training progresses to detect and filter out potentially incorrect pseudo-labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 288, + 714 + ], + "type": "text", + "content": "2. We show that MarginMatch outperforms existing works on six well-established computer vision benchmarks" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 274, + 547, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 274, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 324, + 274, + 547, + 310 + ], + "type": "text", + "content": "showing larger improvements in error rates especially on challenging datasets, while achieving similar convergence performance (or better) than prior works." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 313, + 547, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 547, + 350 + ], + "type": "text", + "content": "3. We perform a comprehensive analysis of our approach and indicate potential insights into why our Margin-Match substantially outperforms other SSL techniques." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 358, + 392, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 392, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 392, + 372 + ], + "type": "text", + "content": "2. MarginMatch" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": "Notation Let " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "L = \\{(x_{1},y_{1}),\\dots,(x_{B},y_{B})\\}" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " be a batch of size " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " of labeled examples and " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "U = \\{\\hat{x}_1,\\dots,\\hat{x}_{\\nu B}\\}" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " be a batch of size " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "\\nu B" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " of unlabeled examples, where " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " is the batch-wise ratio of unlabeled to labeled examples. Let " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y|x)" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " denote the class distribution produced by model " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " on input image " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\theta}(y|x)" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " denote the argmax of this distribution as a one-hot label. Let also " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "H(p,q)" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " denote the cross-entropy between two probability distributions " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 378, + 547, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 480, + 387, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 387, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 387, + 493 + ], + "type": "text", + "content": "2.1. Background" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 629 + ], + "type": "text", + "content": "Consistency reg [39] is an important component in recent semi-supervised learning approaches and relies on the continuity assumption [2,23] that the model should output similar predictions on multiple perturbed versions of the same input " + }, + { + "bbox": [ + 304, + 498, + 547, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 498, + 547, + 629 + ], + "type": "text", + "content": ". As mentioned above, examples of two such approaches are FixMatch [41] and FlexMatch [49] that use consistency regularization at their core combined with psedo-labeling. In psedo-labeling [25], a model itself is used to assign artificial labels for unlabeled data and only artificial labels whose largest class probability is above a predefined confidence threshold are used during training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Specifically, FixMatch [41] predicts artificial labels for unlabeled examples using a weakly-augmented version of each unlabeled example and then employs the artificial labels as pseudo-labels to train against but this time using a strongly-augmented version of each unlabeled example. That is, FixMatch minimizes the following batch-wise consistency loss on unlabeled data:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15774" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 59, + 71, + 287, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 71, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 287, + 118 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right)\\right) > \\tau\\right) \\quad \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {1} \\\\ \\end{array}", + "image_path": "b591fa6775ec4468e47c1c6af3ea3e603baac33722c2f75db2b8811b5ddbc224.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "spans": [ + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "text", + "content": " is a confidence threshold, " + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "text", + "content": " are weak and strong augmentations, respectively, and " + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "text", + "content": " is the indicator function. Therefore, the low-confidence examples (lower than " + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 128, + 287, + 188 + ], + "type": "text", + "content": ") are completely ignored despite containing potentially useful information for model training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": "FlexMatch [49] argues that using a fixed threshold " + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": " to filter the unlabeled data ignores the learning difficulties of different classes, and thus, introduces class-dependent thresholds, which are obtained by adaptively scaling " + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": " depending on the learning status of each class. FlexMatch assumes that a class with fewer examples above the fixed threshold " + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": " has a greater learning difficulty, and hence, it adaptively lowers the threshold " + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": " to encourage more training examples from this class to be learned. The learning status " + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\alpha_{c}" + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": " for a class " + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": " is simply computed as the number of unlabeled examples that are predicted in class " + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": " and pass the fixed threshold " + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 193, + 288, + 326 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 333, + 287, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 333, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 52, + 333, + 287, + 373 + ], + "type": "interline_equation", + "content": "\\alpha_ {c} = \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right)\\right) > \\tau\\right) \\mathbb {1} \\left(\\hat {p} _ {\\theta} \\left(y \\mid \\pi \\left(\\hat {x} _ {i}\\right)\\right) = c\\right) \\tag {2}", + "image_path": "55f94157970d3f4b1bb1db2dc04a458fa2f08ccda9f8a9735baec7b48c5cda43.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 374, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 374, + 287, + 410 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 374, + 287, + 410 + ], + "type": "text", + "content": " is the total number of unlabeled examples. This learning effect is then normalized and used to obtain the class-dependent threshold for each class " + }, + { + "bbox": [ + 46, + 374, + 287, + 410 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 374, + 287, + 410 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 418, + 287, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 418, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 125, + 418, + 287, + 446 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {c} = \\frac {\\alpha_ {c}}{\\max _ {c} \\left(\\alpha_ {c}\\right)} \\times \\tau \\tag {3}", + "image_path": "9ddf3834e82f8f24a466c261a9f6237a45c3022d4cf11a1e2e21c311478768cf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": "In practice, FlexMatch iteratively computes new thresholds after each complete pass through unlabeled data, hence we can parameterize " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c^t" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": ", denoting the threshold obtained at iteration " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": ". The unlabeled loss is then obtained by plugging in the adaptive threshold " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c^t" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": " in Eq. 1:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 525, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 525, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 59, + 525, + 287, + 573 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} (\\max (p _ {\\theta} (y | \\pi (\\hat {x} _ {i}))) > \\mathcal {T} _ {\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i}))} ^ {t}) \\quad \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {4} \\\\ \\end{array}", + "image_path": "361335d8f80d91bdefec503813ff81737f44eaeb663501773bde8d8af763e1a7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "The aforementioned works use the confidence of the model solely at the current iteration to enforce quality of pseudo-labels. We believe this is not sufficient as it provides only a myopic view of the model's behavior (i.e., its confidence) on unlabeled data (at a single iteration) and may result in wrong pseudo-labels even when the confidence threshold is high enough (e.g., if the model is miscalibrated or overly-confident [13]). Figure 1 shows examples of images that are added to the training set with a wrong pseudo-label for both FixMatch and FlexMatch. These types of unlabeled examples, which are incorrectly pseudo-labeled and used" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "during training are particularly harmful for deep neural networks, which can attain zero training error on any dataset, even on randomly assigned labels [50], resulting in poor generalization capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 128, + 494, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 128, + 494, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 128, + 494, + 140 + ], + "type": "text", + "content": "2.2. Proposed Approach: MarginMatch" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 146, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 547, + 312 + ], + "type": "text", + "content": "We now introduce MarginMatch, our new SSL approach that uses the model's training dynamics on unlabeled data to improve pseudo-label data quality. Our approach leverages consistency regularization with weak and strong augmentations and pseudo-labeling, but instead of using only the model's current belief (i.e., its confidence at the current iteration) to decide if an unlabeled example should be used for training or not, our MarginMatch monitors the training dynamics of unlabeled data over the iterations by investigating the margins (a measure of confidence) of the outputs of the model [3]. The margin of a training example is a well established metric in machine learning [3, 11, 18, 33] that quantifies the difference between the logit corresponding to the assigned ground truth label and the largest other logit." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": "In our SSL formulation, we redefine the concept of margins to pseudo-margins of unlabeled examples since no ground truth labels are available for the unlabeled data. Let " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": " be the pseudo-label (or the argmax of the prediction, i.e., " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\theta}(y|\\pi(\\hat{x}))" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": ") at iteration " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": " on unlabeled example " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": " after applying weak augmentations. We define the pseudo-margin (PM) of " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": " with respect to pseudo-label " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 408, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 408, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 365, + 408, + 545, + 422 + ], + "type": "interline_equation", + "content": "\\mathrm {P M} _ {c} ^ {t} (\\hat {x}) = z _ {c} - \\max _ {c! = i} (z _ {i}) \\tag {5}", + "image_path": "fbed55ea01d60f6f00980874c782e951bfe53b6f08b519f371713e786209403c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": " is the logit corresponding to the assigned pseudo-label " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\max_{c! = i}(z_i)" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": " is the largest other logit corresponding to a label " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": " different from " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": ". To monitor the model's predictions on " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": " with respect to pseudo-label " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": " from the beginning of training to iteration " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": ", we average all the margins with respect to " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": " from the first iteration until " + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": " and obtain the average pseudo-margin (APM) as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 520, + 545, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 520, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 369, + 520, + 545, + 553 + ], + "type": "interline_equation", + "content": "\\mathrm {A P M} _ {c} ^ {t} (\\hat {x}) = \\frac {1}{t} \\sum_ {j = 1} ^ {t} \\mathrm {P M} _ {c} ^ {j} (\\hat {x}) \\tag {6}", + "image_path": "b56bf44616baa02be8b48ac3fa8259d972bdec24a183058acd9c37e338a2ad38.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": " acts as the \"ground truth\" label for the APM calculation. Note that if at a prior iteration " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": ", the assigned pseudo-label is different from " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": " (say " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": "), then the APM calculation at iteration " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": " is done with respect to " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": " (by averaging all margins with respect to " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": " from 1 to " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": "). In practice, we maintain a vector of pseudo-margins for all classes accumulated over the training iterations and dynamically retrieve the accumulated pseudo-margin value of the argmax class " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": " to obtain the " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{APM}_c^t" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 557, + 547, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Intuitively, if " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " is the pseudo-label of " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PM}_c^t" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " with respect to class " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " will be positive. In contrast, if the argmax of the model prediction on " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " at a previous iteration " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "t' < t" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " is different from " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "PM_c^{t'}" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "t'" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15775" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 85, + 547, + 301 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 159, + 83 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 159, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 159, + 83 + ], + "type": "text", + "content": "Algorithm 1 MarginMatch" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": "Require: Labeled data " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": "; unlabeled data " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": "; erroneous examples " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": "; maximum number of iterations " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": "; number of classes " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " original classes plus one virtual class of erroneous examples); " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " model; " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " weak augmentations; " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " strong augmentations. \n1: Initialize the Average Pseudo-Margin (APM) threshold " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\gamma^1" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " at the first iteration to a small value (e.g., " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\gamma^1 = -\\infty" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": "). \n2: for " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " do \n3: Estimate learning status " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\alpha_c" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " (using Eq. 2) and calculate the class-wise flexible thresholds " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c^t" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " (using Eq. 3) for each class " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": ". \n4: while " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " not exhausted do \n5: Labeled batch " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "L_b = \\{(x_1, y_1), \\dots, (x_B, y_B)\\}" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": ", unlabeled batch " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "U_b = \\{\\hat{x}_1, \\dots, \\hat{x}_{\\nu B}\\}" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": ", erroneous (or mislabeled) batch " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "E_b = \\{(\\tilde{x}_1, C + 1), \\dots, (\\tilde{x}_B, C + 1)\\}" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " \n6: for " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "x \\in U_b \\cup E_b" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " do \n7: Compute logits " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "z_c" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " after applying weak augmentations when " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "x \\in U_b" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " and strong augmentations when " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "x \\in E_b" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": ". \n8: Calculate pseudo-margin " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "PM_c^t" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " (using Eq. 5) and update Average " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "PM_c^t" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " (using Eq. 6) for each " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "c = 1" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": ". \n9: end for \n10: Minimize " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_s + \\lambda (\\mathcal{L}_u + \\mathcal{L}_e)" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s = \\frac{1}{B} \\sum_{i=1}^{B} H(y_i, p_\\theta(y|\\pi(x_i)))" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u = \\sum_{i=1}^{\\nu B} \\mathbb{1}(\\mathrm{AM}_{p_\\theta(y|\\pi(\\hat{x}_i))}^t (\\hat{x}_i) > \\gamma^t) \\times \\mathbb{1}(\\max(p_\\theta(y|\\pi(\\hat{x}_i))) > T_{p_\\theta(y|\\pi(\\hat{x}_i))}^t) \\times H(\\hat{p}_\\theta(y|\\pi(\\hat{x}_i)), p_\\theta(y|\\Pi(\\hat{x}_i)))" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " \n13: " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e = \\sum_{i=1}^{B} H(C + 1, p_\\theta(y|\\Pi(\\tilde{x}_i)))" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " \n14: end while \n15: Update " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\gamma^{t+1}" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "95^{th}" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": " percentile erroneous sample " + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "inline_equation", + "content": "APM_{C+1}^t" + }, + { + "bbox": [ + 47, + 85, + 547, + 301 + ], + "type": "text", + "content": ". \n16: end for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "content": "with respect to " + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "content": " will be negative. Therefore, if over the iterations, the model predictions do not agree frequently with the pseudo-label " + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "content": " from iteration " + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "content": " and the model fluctuates significantly between iterations on the predicted label, the APM for class " + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "content": " will have a low, likely negative value. Similarly, if the model is highly uncertain of the class of " + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "content": " (reflected in a high entropy of the class probability distribution), the APM for class " + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "content": " will have a low value. These capture the characteristics of unlabeled examples or of those harmful for training. Motivated by these observations, MarginMatch leverages the APM of the assigned pseudo-label " + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 323, + 290, + 478 + ], + "type": "text", + "content": " and compares it with an APM threshold to mask out pseudo-labeled examples with low APMs. Formally, the unlabeled loss in MarginMatch is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 487, + 287, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 287, + 553 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {u} = \\sum_ {i = 1} ^ {\\nu B} \\mathbb {1} \\left(\\mathrm {A M} _ {\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i}))} ^ {t} (\\hat {x} _ {i}) > \\gamma^ {t}\\right) \\quad \\times \\\\ \\mathbb {1} \\left(\\max \\left(p _ {\\theta} \\left(y \\mid \\pi (\\hat {x} _ {i})\\right)\\right) > \\mathcal {T} _ {\\hat {p} _ {\\theta} \\left(y \\mid \\pi (\\hat {x} _ {i})\\right)} ^ {t}\\right) \\times \\\\ H \\left(\\hat {p} _ {\\theta} (y | \\pi (\\hat {x} _ {i})), p _ {\\theta} (y | \\Pi (\\hat {x} _ {i}))\\right) \\tag {7} \\\\ \\end{array}", + "image_path": "8e479cf646f938f2442cb48c7388ed622925c70847ea89beb48733f73db7a026.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\gamma^t" + }, + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "text", + "content": " is the APM threshold at iteration " + }, + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "text", + "content": ", estimated as explained below, and " + }, + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\hat{p}_0(y|\\pi(\\hat{x}_i))}^t" + }, + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "text", + "content": " is the flexible threshold estimated as in FlexMatch [49]. To train our model, we adopt the best practices [41, 49] and optimize the weighted combination of the supervised and unsupervised losses:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 642, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 642, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 134, + 642, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda \\mathcal {L} _ {u} \\tag {8}", + "image_path": "e8f9d8c2f8f90bd59aa87ffc3eb856dbef09db60df41db9348ab41967a543391.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 666, + 201, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 201, + 678 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 201, + 678 + ], + "type": "text", + "content": "where the supervised loss is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 684, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 684, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 100, + 684, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} H \\left(y _ {i}, p _ {\\theta} (y \\mid \\pi \\left(x _ {i}\\right))\\right) \\tag {9}", + "image_path": "ba8b77492494ca341fddbfa5e2086a5a6da9178e5404a67bf88b7526f5d4560e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": "Average Pseudo-Margin Threshold Estimation Inspired by Pleiss et al. [33], we propose to estimate the average pseudo-margin threshold " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "\\gamma^t" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " by analyzing the training dynamics of a special category of unlabeled examples, which we force to be erroneous or mislabeled examples. That is, to create the sample of erroneous examples " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": ", we randomly sample a subset of unlabeled examples from " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " that we assign to an inexistent (or virtual) class " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " at the beginning of the training process and remove them from " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": ". The purpose of these erroneous examples is to mimic the training dynamics of incorrectly pseudo-labeled (unlabeled) examples and use them as proxy to estimate the cutoff of (potentially) mislabeled pseudo-labels. Since the examples in " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " should belong to one of the " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " original classes, assigning them to the inexistent class " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " makes them by definition mislabeled (see Appendix A for additional insights into this virtual class). As with all unlabeled examples from " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": ", we compute " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "APM_{C + 1}^t" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " for the special category of erroneous examples from " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": ", but unlike the unlabeled examples from " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": ", the erroneous ones from " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " have a fixed class " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": ". To mimic the training dynamics of unlabeled examples from " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": ", we use strong augmentations to compute the loss of the erroneous examples from " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": ". That is, given a batch " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "E_b" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 323, + 547, + 609 + ], + "type": "text", + "content": " erroneous examples, the erroneous sample loss becomes:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 613, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 356, + 613, + 545, + 645 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} = \\sum_ {i = 1} ^ {B} H (C + 1, p _ {\\theta} (y | \\Pi (\\tilde {x} _ {i}))) \\tag {10}", + "image_path": "fb3081ccc4c17a21cbb279d30242f5049bf07660f3119cc09dba7d1e1107f2ce.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "content": "At iteration " + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "content": ", we use the APMs of the erroneous examples to choose the APM threshold " + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\gamma^t" + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\gamma^t" + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "content": " as the APM of the " + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "inline_equation", + "content": "95^{th}" + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "content": " percentile erroneous sample. The total loss becomes:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 685, + 545, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 685, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 378, + 685, + 545, + 699 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda \\left(\\mathcal {L} _ {u} + \\mathcal {L} _ {e}\\right) \\tag {11}", + "image_path": "69352b68e26ce7ab3d992e58f71df304371e3f0014ae946f71676387e38c8cb2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 541, + 714 + ], + "type": "text", + "content": "Our full MarginMatch algorithm is shown in Algorithm 1." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15776" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 547, + 163 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 163 + ], + "type": "table", + "html": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10
#Labels/Class425400425100425100425100
Pseuso-Labeling74.610.2646.492.2015.080.1987.450.8557.740.2836.550.2464.615.6025.212.039.400.3274.680.9955.452.4332.640.71
UDA10.793.755.320.064.410.0748.951.5929.430.2123.870.235.344.274.260.391.950.0137.828.449.811.156.810.17
MixMatch45.242.1512.761.147.130.3462.152.1741.511.1928.160.2446.181.783.980.173.501.1334.151.548.950.3210.410.73
ReMixMatch5.270.194.850.134.040.1247.150.7627.140.2323.780.124.230.313.180.041.940.0631.510.758.540.486.190.24
FixMatch7.80.284.910.054.250.0848.210.8229.450.1622.890.123.971.183.131.031.970.0338.434.1410.451.046.430.33
FlexMatch5.040.065.040.094.190.0139.991.6226.960.0822.440.158.193.207.782.556.720.3029.151.328.230.135.770.12
MarginMatch4.910.074.730.123.980.0236.971.3223.710.1321.390.123.751.203.141.171.930.0125.373.587.310.355.520.15
", + "image_path": "bee4bbdb7b062ce377dc54504dbd9c3d1ed8f7cb422e392703999935bf2f3d8d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 171, + 501, + 182 + ], + "lines": [ + { + "bbox": [ + 91, + 171, + 501, + 182 + ], + "spans": [ + { + "bbox": [ + 91, + 171, + 501, + 182 + ], + "type": "text", + "content": "Table 1. Test error rates on CIFAR-10, CIFAR-100, SVHN, and STL-10 datasets. Best results are shown in blue." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "text", + "content": "Exponential Moving Average of Pseudo-Margins The current definition of APM weighs the pseudo-margin at iteration " + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "text", + "content": " identical to the pseudo-margin at a much earlier iteration " + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "inline_equation", + "content": "t >> p" + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "text", + "content": "). This is problematic since very old pseudo-margins eventually become deprecated (especially due to the large number of iterations through unlabeled data in consistency training (" + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "inline_equation", + "content": "\\sim 9K" + }, + { + "bbox": [ + 46, + 203, + 289, + 347 + ], + "type": "text", + "content": "), and hence, the old margins are no longer indicative of the current learning status of the model. To this end, instead of averaging all pseudo-margins (from the beginning of training to the current iteration), we propose to use an exponential moving average to place more importance on recent iterations. Formally, APM becomes:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 350, + 287, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 384 + ], + "type": "interline_equation", + "content": "\\mathrm {A P M} _ {c} ^ {t} (\\hat {x}) = \\mathrm {P M} _ {c} ^ {t} (\\hat {x}) * \\frac {\\delta}{1 + t} + \\mathrm {A P M} _ {c} ^ {t - 1} (\\hat {x}) * \\left(1 - \\frac {\\delta}{1 + t}\\right) \\tag {12}", + "image_path": "d28af159245f82082d953b6e0163ca900b778286a267be5f31a5e9df90bb7d40.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 384, + 285, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 285, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 285, + 396 + ], + "type": "text", + "content": "We set the smoothing parameter " + }, + { + "bbox": [ + 47, + 384, + 285, + 396 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 384, + 285, + 396 + ], + "type": "text", + "content": " to 0.997 in experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 406, + 128, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 128, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 128, + 420 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "type": "text", + "content": "We evaluate the performance of our MarginMatch on a wide range of SSL benchmark datasets. Specifically, we perform experiments with various numbers of labeled examples on CIFAR-10, CIFAR-100 [21], SVHN [31], STL-10 [8], ImageNet [10], and WebVision [26]. For smaller scale datasets such as CIFAR-10, CIFAR-100, SVHN, and STL-10 we randomly sample a small number of labeled examples per class (ranging from 4 labels per class to 400 labels per class) and treat them as labeled data, whereas the remaining labeled examples are treated as unlabeled data, except for STL-10 [8], which provides its own set of unlabeled examples. On ImageNet and WebVision, we use " + }, + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "type": "inline_equation", + "content": "\\sim 10\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "type": "text", + "content": " of the available labeled examples as labeled data, with the remaining " + }, + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "type": "inline_equation", + "content": "(90\\%)" + }, + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "type": "text", + "content": " being treated as unlabeled data. In all our experiments, we sample " + }, + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 617 + ], + "type": "text", + "content": " of the unlabeled data and place it in the set of erroneous examples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 689 + ], + "type": "text", + "content": "We report the mean and standard deviation of error rates from five runs with different parameter initializations. Similar to FixMatch [41], we use Wide Residual Networks [48]: WRN-28-2 for CIFAR-10 and SVHN; WRN-28-8 for CIFAR-100; and WRN-37-2 for STL-10. We use ResNet-50 [14] for both ImageNet and WebVision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "In our experiments, we adopt the same hyperparameters as FixMatch [41]. Specifically, we use stochastic gradient" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": "descent (SGD) with a momentum of 0.9. We start with a learning rate of 0.03 and employ a cosine learning rate; at iteration " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": ", our learning rate is " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\eta(k) = \\cos\\left(\\frac{7k\\pi}{16K}\\right)" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": " is the maximum number of iterations and is set to " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "2^{20}" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": ". We also leverage the same data augmentations as in FlexMatch [49]. Specifically, for weak augmentations we employ a standard flip-and-shift augmentation and use RandAugment [9] for strong augmentations. We set the batch size " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "B = 64" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": ", ratio of unlabeled to labeled data in a batch to " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\nu = 7" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": " and weigh the supervised and unsupervised losses equally (i.e., " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": "). We set our initial confidence threshold to " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\tau = 0.95" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": " and our average pseudo-margin (APM) threshold to the APM of the " + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "inline_equation", + "content": "95^{th}" + }, + { + "bbox": [ + 304, + 203, + 547, + 384 + ], + "type": "text", + "content": " percentile threshold sample. To report the error rates, we compare all the approaches using the model at the end of training as in FixMatch [41]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 393, + 534, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 393, + 534, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 534, + 406 + ], + "type": "text", + "content": "3.1. CIFAR-10, CIFAR-100, SVHN, and STL-10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "type": "text", + "content": "We compare MarginMatch against strong baselines and prior works: Pseudo-Labeling [1], Unsupervised Data Augmentation (UDA) [46], MixMatch [5], ReMixMatch [4], FixMatch [41], and FlexMatch [49]. We show in Table 1 the error rates obtained by our MarginMatch and the baselines on the CIFAR-10, CIFAR-100, SVHN and STL-10 datasets. First, we observe that our approach improves the performance on both CIFAR-10 and CIFAR-100. On CIFAR-10, MarginMatch improves performance in all data regimes upon FlexMatch [49], which is the current state-of-the-art, while maintaining a good error rate standard deviation. On CIFAR-100, which is significantly more challenging than CIFAR-10, we observe that MarginMatch bring substantially larger improvements. Notably, we see " + }, + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "type": "inline_equation", + "content": "3.02\\%" + }, + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "type": "text", + "content": " improvement over FlexMatch in error rate using only 4 labels per class, and " + }, + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "type": "inline_equation", + "content": "3.25\\%" + }, + { + "bbox": [ + 304, + 413, + 547, + 628 + ], + "type": "text", + "content": " improvement using 25 examples per class. These results on CIFAR-100 emphasize the effectiveness of MarginMatch, which performs well on a more challenging dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "On SVHN, our approach performs better than FixMatch using 4 labels per class and performs similarly with FixMatch using 25 and 100 labels per class. However, on this dataset, MarginMatch performs much better compared with FlexMatch. For example, MarginMatch achieves " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "3.75\\%" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " error rate using 4 labels per class, whereas FlexMatch obtains an error rate of " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "8.19\\%" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " with the same labels per class," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15777" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 171, + 176 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 171, + 176 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 171, + 176 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 171, + 176 + ], + "type": "image", + "image_path": "3863b902bc94de3c24819d9343ec3c7e0f43b2934fb2ad8b2d6946347746db10.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 186, + 483, + 198 + ], + "lines": [ + { + "bbox": [ + 109, + 186, + 483, + 198 + ], + "spans": [ + { + "bbox": [ + 109, + 186, + 483, + 198 + ], + "type": "text", + "content": "Figure 2. Convergence speed of MarginMatch against FixMatch and FlexMatch with 4 labels per class." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 294, + 176 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 294, + 176 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 294, + 176 + ], + "type": "image", + "image_path": "f0f4a2e72ba307d2e520de4a90e1334b6cb7289c8ab771c2357bf833a851585e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 416, + 176 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 416, + 176 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 416, + 176 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 416, + 176 + ], + "type": "image", + "image_path": "67c934250c290cbd95125a37db084ce3f4ae34da39c9f55bcc23d631dbdd9db4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 418, + 70, + 539, + 176 + ], + "blocks": [ + { + "bbox": [ + 418, + 70, + 539, + 176 + ], + "lines": [ + { + "bbox": [ + 418, + 70, + 539, + 176 + ], + "spans": [ + { + "bbox": [ + 418, + 70, + 539, + 176 + ], + "type": "image", + "image_path": "6b5dee9f559994b4f23f9e72ebdbe3369039e06c5226b62e9377d21b528f02ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 73, + 216, + 261, + 292 + ], + "blocks": [ + { + "bbox": [ + 73, + 216, + 261, + 292 + ], + "lines": [ + { + "bbox": [ + 73, + 216, + 261, + 292 + ], + "spans": [ + { + "bbox": [ + 73, + 216, + 261, + 292 + ], + "type": "table", + "html": "
DatasetImageNetWebVision
TOP-1TOP-5TOP-1TOP-5
Supervised48.3925.4949.5826.78
FixMatch43.6621.8044.7622.65
FlexMatch42.0219.4943.8722.07
MarginMatch41.0518.2843.0821.13
", + "image_path": "42609ebd2faf187629aa2e0f8b1c46db7473ac70cb449ee54164b890369bbf18.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "type": "text", + "content": "yielding an improvement of MarginMatch of " + }, + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "type": "inline_equation", + "content": "4.44\\%" + }, + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "type": "text", + "content": " over FlexMatch. We hypothesize that the low performance of FlexMatch is due to its limitation in handling unbalanced class distributions [49]. On STL-10, MarginMatch as well outperforms all the other approaches both in error rates and error rate standard deviation. Notably, on this dataset, our approach pushes the performance of FlexMatch by " + }, + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "type": "inline_equation", + "content": "3.78\\%" + }, + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "type": "text", + "content": " in error rate using only 4 labels per class and by " + }, + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "type": "inline_equation", + "content": "0.92\\%" + }, + { + "bbox": [ + 46, + 328, + 287, + 436 + ], + "type": "text", + "content": " using 25 labels per class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 593 + ], + "type": "text", + "content": "Next, we compare MarginMatch with FixMatch and FlexMatch in terms of convergence speed in the extremely labelscarcse setting of 4 labels per class and show these results in Figure 2. Notably, we observe that MarginMatch has a similar convergence speed (or even better on CIFAR-100) compared with FlexMatch while achieving a lower test error rate than FlexMatch on all datasets with 4 labels per class (see Table 1). Even more strikingly, compared with FixMatch, MarginMatch has a much superior convergence speed for a much better test error rate with 4 labels per class. This is because the rigid thresholds in FixMatch allow access only to a small amount of unlabeled data for training at each iteration and it takes a lot longer for the model to train." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 599, + 189, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 189, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 189, + 612 + ], + "type": "text", + "content": "3.2. ImageNet and WebVision" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "To showcase the effectiveness of our approach in a large-scale setup, we test our MarginMatch on ImageNet [10] and WebVision [26] using " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " labeled examples in total. We show the results obtained in Table 2. We observe that our MarginMatch outperforms FixMatch and FlexMatch on both datasets. It is worth noting that large-scale self-supervised approaches such as SimCLR [7] achieve high performance on ImageNet but at a much higher computational cost." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 321, + 217, + 533, + 246 + ], + "blocks": [ + { + "bbox": [ + 46, + 300, + 288, + 322 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 322 + ], + "type": "text", + "content": "Table 2. Test error rates on the ImageNet and WebVision datasets. Best results are shown in blue." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 217, + 533, + 246 + ], + "lines": [ + { + "bbox": [ + 321, + 217, + 533, + 246 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 533, + 246 + ], + "type": "table", + "html": "
δ0.950.990.9950.9970.9991
ERR RATE38.1338.0537.9237.9139.1239.72
", + "image_path": "daa1afef547adca559f5f36b3a3633309ab3af22d5f937a775fb2234f7930de6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 254, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 277 + ], + "type": "text", + "content": "Table 3. Error rates obtained on CIFAR-100 with four examples per class and various smoothing values " + }, + { + "bbox": [ + 305, + 254, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 254, + 545, + 277 + ], + "type": "text", + "content": " . Best result is in blue." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 300, + 547, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 547, + 347 + ], + "type": "text", + "content": "MarginMatch outperforms other SSL methods using the same ResNet-50 architecture at the same computational cost. We emphasize MarginMatch is most successful and relevant in low data regimes on smaller datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 360, + 398, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 398, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 398, + 374 + ], + "type": "text", + "content": "4. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "text", + "content": "Exponential Moving Average Smoothing for APM Computation In our approach, we employ an exponential moving average (EMA) of the pseudo-margin values with a smoothing value of " + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "inline_equation", + "content": "\\delta = 0.997" + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "text", + "content": " to compute the APM. We now analyze how our approach performs with different EMA smoothing values or with no EMA at all. Table 3 shows these results on CIFAR-100 with 4 labels per class. First, we observe that employing a simple average of pseudo-margin values for the APM computation (i.e., " + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "inline_equation", + "content": "\\delta = 1" + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "text", + "content": ") performs extremely poorly, obtaining a " + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "inline_equation", + "content": "39.72\\%" + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "text", + "content": " error rate. This result emphasizes that margins eventually become deprecated and it is essential to scale them down in time. Using a low smoothing factor of " + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "inline_equation", + "content": "\\delta = 0.95" + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "text", + "content": " is not effective either, denoting that abruptly forgetting margin values does not work either. Our chosen " + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "inline_equation", + "content": "\\delta = 0.997" + }, + { + "bbox": [ + 304, + 384, + 546, + 624 + ], + "type": "text", + "content": " strikes a balance between the two by eliminating the harmful effects of very old margins while keeping track of a good amount of previous estimates (e.g., a margin value computed 200 epochs previously is scaled down by 0.55, while a margin value computed 1000 epochs previously is scaled by 0.05)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": "Pseudo-Margin vs. Other Measures for Pseudo-Label Correctness Our MarginMatch monitors the pseudomargins of a model's predictions across training iterations to ensure the quality of pseudo-labels. However, other measures such as confidence or entropy exist that can assess the pseudo-label correctness. Hence, we perform an ablation where we replace the pseudo-margins in our MarginMatch" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15778" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 156 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 156 + ], + "type": "table", + "html": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10
#Labels/Class425400425100425100425100
Avg Confidence23.872.7314.211.377.540.7841.232.1531.491.4824.112.368.994.276.540.394.730.0131.678.4414.871.157.590.17
Avg Entropy8.580.416.180.155.850.1245.100.9126.021.1122.130.2515.691.2512.740.789.330.0529.543.5110.631.3510.840.47
Avg Margin7.250.295.380.764.730.0939.721.5225.210.5223.180.1718.451.3611.290.938.400.0428.454.289.341.347.590.21
EMA Confidence4.910.454.740.093.990.0638.670.7425.610.1221.480.173.840.233.250.031.930.0925.90.817.60.425.740.57
EMA Entropy6.40.438.340.124.210.0941.630.7636.840.1322.520.073.811.263.170.872.140.0427.214.058.281.016.790.27
EMA Margin4.910.074.730.123.980.0236.971.3223.710.1321.390.123.751.203.141.171.930.0125.373.587.310.355.520.15
", + "image_path": "a93f56de069c4074fd5f97f682bcd492dc09a2459b5f7c6cfacef153b7e8c3d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 164, + 499, + 176 + ], + "lines": [ + { + "bbox": [ + 94, + 164, + 499, + 176 + ], + "spans": [ + { + "bbox": [ + 94, + 164, + 499, + 176 + ], + "type": "text", + "content": "Table 4. Test error rates comparing pseudo-margin with confidence and entropy. Best results are shown in blue." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 184, + 167, + 289 + ], + "blocks": [ + { + "bbox": [ + 50, + 184, + 167, + 289 + ], + "lines": [ + { + "bbox": [ + 50, + 184, + 167, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 184, + 167, + 289 + ], + "type": "image", + "image_path": "5d7a6f0b6f27210d2b1e6974fc18acd3a020cb78d91155ca3f6a1f383e9031af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 299, + 288, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 288, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 288, + 321 + ], + "type": "text", + "content": "Figure 3. Mask rate and impurity on CIFAR-100 with 4 labeled examples per class." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 184, + 286, + 289 + ], + "blocks": [ + { + "bbox": [ + 168, + 184, + 286, + 289 + ], + "lines": [ + { + "bbox": [ + 168, + 184, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 168, + 184, + 286, + 289 + ], + "type": "image", + "image_path": "da786527f8c4ef697b2a118d71e40212692d47c7e16ad1ca6e0824b5a67174a2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 327, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 327, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 45, + 327, + 287, + 506 + ], + "type": "text", + "content": "with average confidence and entropy and compare their performance. Specifically, we design the following approaches: 1) Avg Confidence monitors the confidence of the prediction for each unlabeled example and takes the average over the training iterations; 2) Avg Entropy monitors the entropy of the class probability distribution for each unlabeled example and takes the average across the training iterations. In addition, we also consider 3) EMA Confidence and 4) EMA Entropy which are similar to Avg Confidence and Avg Entropy, respectively, but use an exponential moving average (EMA) instead of the simple averaging. The estimation of the threshold for each of these approaches is done in a similar manner as for pseudo-margins, using erroneous samples and considering the value of the 95th percentile erroneous sample as the threshold." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 288, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 626 + ], + "type": "text", + "content": "We show the results obtained using these approaches in Table 4. First, we observe that all measures (pseudo-margin, confidence and entropy) with EMA perform better than their counterpart with simple averaging. Second, EMA Margin achieves the lowest test error rates compared with EMA Confidence and EMA Entropy. Thus, we conclude that pseudo-margins provide an excellent measure for pseudolabel correctness. See Appendix B for some additional insights into why EMA Margin outperforms EMA confidence and entropy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 638, + 106, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 106, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 106, + 651 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 186, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 186, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 186, + 672 + ], + "type": "text", + "content": "5.1. Mask Rate and Impurity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "We now contrast MarginMatch with FixMatch and FlexMatch in terms of the quality of pseudo-labels using two metrics: mask rate and impurity and show these results in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 185, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 376 + ], + "type": "text", + "content": "Figure 3, respectively, using CIFAR-100 with 4 labels per class. Mask rate is defined as the fraction of pseudo-labeled examples that do not participate in the training at epoch " + }, + { + "bbox": [ + 304, + 185, + 547, + 376 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 185, + 547, + 376 + ], + "type": "text", + "content": " due to confidence masking or pseudo-margin masking (or both). Impurity in contrast is defined as the fraction of pseudolabeled examples that do participate in the training at epoch " + }, + { + "bbox": [ + 304, + 185, + 547, + 376 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 185, + 547, + 376 + ], + "type": "text", + "content": " but with a wrong label. An effective SSL model minimizes both metrics: a low mask rate indicates that the model has access to more unlabeled examples during training (otherwise a low percentage and less diverse set of unlabeled examples are seen during training) while low impurity indicates that the pseudo-labels of these examples are of high quality. Note that we can compute impurity on these two datasets because our unlabeled data comes from the labeled training set of each of these datasets (thus we compare the pseudo-labels against the gold labels of each dataset)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 377, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 547, + 533 + ], + "type": "text", + "content": "As can be seen from the figures, FixMatch has a significantly larger mask rate due to the rigid confidence threshold set to a high value of 0.95. In contrast, FlexMatch lowers the mask rate by " + }, + { + "bbox": [ + 304, + 377, + 547, + 533 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 377, + 547, + 533 + ], + "type": "text", + "content": " with the introduction of flexible thresholds, but has a much higher impurity compared with FixMatch. Notably, our MarginMatch has only a slightly higher mask rate compared with FlexMatch and at the same time achieves a much lower impurity than FlexMatch and even FixMatch despite that FixMatch employs a very high confidence threshold. These results show that MarginMatch that enforces an additional measure for pseudo-labeled data quality maintains a low mask rate without compromising the quality of the pseudo-labels (i.e., low mask rate and low impurity)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 540, + 422, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 422, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 422, + 551 + ], + "type": "text", + "content": "5.2. Anecdotal Evidence" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "We show in Figure 4 anecdotal evidence of the effectiveness of MarginMatch. To this end, we extract two bird images from our unlabeled portion of CIFAR-10 [21] of various learning difficulties that resemble characteristics of plane images (e.g., the background). The top part of the figure illustrates the progression over the training iterations of the confidence and the confidence thresholds of FlexMatch for the classes bird and plane, whereas the bottom part of the figure illustrates the progression of the APM threshold of MarginMatch along with its APMs of bird and plane classes over the training iterations. In the rightmost image, for MarginMatch we can observe that the APM of the bird class becomes stronger and stronger as the training progresses and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15779" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 289, + 170 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 289, + 170 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 289, + 170 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 289, + 170 + ], + "type": "image", + "image_path": "09309a144bfd5ba04043d61348a201f6da75b5cbced2a082c1713cf8cca3df75.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 186, + 287, + 283 + ], + "blocks": [ + { + "bbox": [ + 50, + 186, + 287, + 283 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 287, + 283 + ], + "type": "image", + "image_path": "88b8e4a329bb9009cfef05f34bd5bd8d478b95e290408b600afa6990ba38710c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 295, + 287, + 317 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 317 + ], + "type": "text", + "content": "Figure 4. Confidence thresholding vs. APM Thresholding on two images from the CIFAR-10 dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 288, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 525 + ], + "type": "text", + "content": "eventually exceeds the APM threshold of MarginMatch, and hence, the image is included in the training until the end with the correct bird label. Interestingly, for the same image, in FlexMatch the confidence for the bird class is very close to the bird confidence threshold and eventually falls below this threshold and exits the training set. In contrast, the leftmost image is significantly more challenging than the rightmost image since it is more similar to plane images, which makes it an easily confusable example. Here, we observe that the confidence of FlexMatch exceeds the flexible threshold with the incorrect argmax class plane starting from iteration 3000. Moreover, Flexmatch continues to use this image with the wrong plane label for the remaining of the training process. Critically, in MarginMatch the APM value for the plane class does not exceed the APM threshold, and the model eventually learns to classify this image correctly and includes it in training with the correct bird pseudo-label." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "text", + "content": "6. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "type": "text", + "content": "Here, we focus on various SSL approaches that our MarginMatch directly builds upon although there are other SSL techniques that are not presented in this review, such as approaches based on generative models [17, 24], graph-based approaches [19, 20] and robust SSL [32, 37] (see Appendix C for a comparison between MarginMatch and Robust SSL approaches)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Self-training [28, 36, 40, 47] is a popular SSL method where the predictions of a model on unlabeled data are used as artificial labels to train against. Noisy student training [47] is a popular self-training approach that also leverages knowledge distillation [16] and iteratively jointly trains two models in a teacher-student framework. Noisy student uses a" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "larger model size and noised inputs, exposing the student to more difficult learning environments, leading to an increased performance compared to the teacher." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 547, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 299 + ], + "type": "text", + "content": "Pseudo-labeling is a variant of self-training where these predictions are sharpened to obtain hard labels [25]. The use of hard labels can be seen as a means of entropy minimization [12] and nowadays is a valuable component in most successful SSL approaches [4, 5, 49]. These hard labels are usually used along a confidence threshold, where unconfident unlabeled examples are completely disregarded (e.g., [5]) to avoid using noisy pseudo-labels. Recently, approaches such as Curriculum Labeling (CL) [6] or FlexMatch [49] started to explore curriculum learning in the SSL context. CL proposes a self-pacing strategy of identifying easy and hard examples to ensure that the model first uses easy and progressively moves towards harder examples. Similarly, MarginMatch uses curriculum learning and pseudo-labeling, but the focus of our approach is placed on producing better thresholds for assessing the quality of pseudo-labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 300, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 491 + ], + "type": "text", + "content": "Consistency regularization [2] is a method that applies random perturbations when generating the artificial label, such as data augmentation [4, 5, 41], dropout [39], or adversarial perturbations [29]. Current state-of-the-art approaches [41, 49] exploit a combination of weak and strong data augmentations, which were shown to be extremely beneficial in SSL. The most popular strong augmentations used in the SSL literature are RandAugment [9] and CTAugment [4]. The approaches based on these methods first generate a hard label using pseudo-labeling on a weakly augmented image (i.e., using a low noise transformation such as a flip-and-shift augmentation), then optimize the predictions of the model on a strongly augmented version of the same image towards this hard label. Similar to these approaches, MarginMatch uses the same combination of weak and strong data augmentations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 712 + ], + "type": "text", + "content": "In this paper, we proposed a novel semi-supervised learning method that improves the pseudo-label quality using training dynamics. Our new method is lightweight and achieves state-of-the-art performance on four computer vision SSL datasets in low data regimes and on two large-scale benchmarks. MarginMatch takes into consideration not only a flexible confidence threshold to account for the difficulty of each class, but also a measure of quality for each unlabeled example using training dynamics. In addition, MarginMatch is a general approach that can be leveraged in most SSL frameworks and we hope that it can attract future research in analyzing the effectiveness of SSL approaches focused on data quality. As future work, we aim to further explore our method in settings when there is a mismatch between the labeled and unlabeled data distributions (i.e., making use of out-of-domain unlabeled data)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15780" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "type": "text", + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. CoRR, abs/1908.02983, 2019. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "text", + "content": "[3] Peter L. Bartlett, Dylan J. Foster, and Matus Telgarsky. Spectrally-normalized margin bounds for neural networks. CoRR, abs/1706.08498, 2017. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "type": "text", + "content": "[4] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 1, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 328 + ], + "type": "text", + "content": "[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 288, + 363 + ], + "type": "text", + "content": "[6] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In AAAI, 2021. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 288, + 408 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "type": "text", + "content": "[8] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "type": "text", + "content": "[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "type": "text", + "content": "[11] Gamaleldin Fathy Elsayed, Dilip Krishnan, Hossein Mobahi, Kevin Regan, and Samy Bengio. Large margin deep networks for classification. 2018. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "type": "text", + "content": "[12] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In L. Saul, Y. Weiss, and L. Bottou, editors, Advances in Neural Information Processing Systems, volume 17. MIT Press, 2004. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "type": "text", + "content": "[13] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1321-1330. PMLR, 06-11 Aug 2017. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "text", + "content": "[15] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "type": "text", + "content": "[16] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 547, + 259 + ], + "type": "text", + "content": "[17] Geoffrey E Hinton and Russ R Salakhutdinov. Using deep belief nets to learn covariance kernels for gaussian processes. In J. Platt, D. Koller, Y. Singer, and S. Roweis, editors, Advances in Neural Information Processing Systems, volume 20. Curran Associates, Inc., 2007. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 547, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 547, + 294 + ], + "type": "text", + "content": "[18] Yiding Jiang, Dilip Krishnan, Hossein Mobahi, and Samy Bengio. Predicting the generalization gap in deep networks with margin distributions, 2018. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "type": "text", + "content": "[19] Thorsten Joachims. Transductive inference for text classification using support vector machines. In Proceedings of the Sixteenth International Conference on Machine Learning, ICML '99, page 200-209, San Francisco, CA, USA, 1999. Morgan Kaufmann Publishers Inc. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 547, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 547, + 393 + ], + "type": "text", + "content": "[20] Thorsten Joachims. Transductive learning via spectral graph partitioning. In Proceedings of the Twentieth International Conference on International Conference on Machine Learning, ICML'03, page 290-297. AAAI Press, 2003. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 547, + 415 + ], + "type": "text", + "content": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 416, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 547, + 480 + ], + "type": "text", + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, NIPS'12, page 1097-1105, Red Hook, NY, USA, 2012. Curran Associates Inc. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 547, + 514 + ], + "type": "text", + "content": "[23] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR (Poster). OpenReview.net, 2017. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 547, + 568 + ], + "type": "text", + "content": "[24] J.A. Lasserre, C.M. Bishop, and T.P. Minka. Principled hybrids of generative and discriminative models. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 1, pages 87-94, 2006. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 547, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 547, + 613 + ], + "type": "text", + "content": "[25] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 615, + 547, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 547, + 647 + ], + "type": "text", + "content": "[26] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. CoRR, abs/1708.02862, 2017. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "text", + "content": "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pages 181-196, 2018. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15781" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[28] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152-159, New York City, USA, June 2006. Association for Computational Linguistics. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 161 + ], + "type": "text", + "content": "[29] Takeru Miyato, Andrew M Dai, and Ian Goodfellow. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725, 2016. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "text", + "content": "[30] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 249 + ], + "type": "text", + "content": "[31] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "type": "text", + "content": "[32] Jongjin Park, Sukmin Yun, Jongheon Jeong, and Jinwoo Shin. Opencos: Contrastive semi-supervised learning for handling open-set unlabeled data. In ECCV Workshops, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 288, + 348 + ], + "type": "text", + "content": "[33] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying mislabeled data using the area under the margin ranking. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 17044-17056. Curran Associates, Inc., 2020. 1, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 381 + ], + "type": "text", + "content": "[34] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 436 + ], + "type": "text", + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "type": "text", + "content": "[36] Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. In 2005 Seventh IEEE Workshops on Applications of Computer Vision (WACV/MOTION'05) - Volume 1, volume 1, pages 29-36, 2005. 1, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 557 + ], + "type": "text", + "content": "[37] Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set semi-supervised learning with open-set consistency regularization. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 25956-25967. Curran Associates, Inc., 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "text", + "content": "[38] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Mutual exclusivity loss for semi-supervised deep learning. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1908-1912. IEEE, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "type": "text", + "content": "[39] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29:1163-1171, 2016. 1, 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "text", + "content": "[40] H. Scudder. Probability of error of some adaptive pattern-recognition machines. IEEE Transactions on Information Theory, 11(3):363-371, 1965. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[41] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 545, + 597 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 138 + ], + "type": "text", + "content": "Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 596-608. Curran Associates, Inc., 2020. 1, 2, 4, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[42] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 196, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 196, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 196, + 545, + 238 + ], + "type": "text", + "content": "[43] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "type": "text", + "content": "[44] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in neural information processing systems, pages 1195–1204, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 285, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 328 + ], + "type": "text", + "content": "[45] Vikas Verma, Kenji Kawaguchi, Alex Lamb, Juho Kannala, Arno Solin, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi-supervised learning. Neural Networks, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "type": "text", + "content": "[46] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc., 2020. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 397, + 545, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 545, + 451 + ], + "type": "text", + "content": "[47] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687-10698, 2020. 1, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 453, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 453, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 308, + 453, + 545, + 473 + ], + "type": "text", + "content": "[48] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 475, + 545, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 561 + ], + "type": "text", + "content": "[49] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 18408-18419. Curran Associates, Inc., 2021. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "type": "text", + "content": "[50] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. CoRR, abs/1611.03530, 2016. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15782" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_content_list.json b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..790cf7c682856a6a6bad69a7b80504295fbefc36 --- /dev/null +++ b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_content_list.json @@ -0,0 +1,1724 @@ +[ + { + "type": "text", + "text": "Markerless Camera-to-Robot Pose Estimation via Self-supervised Sim-to-Real Transfer", + "text_level": 1, + "bbox": [ + 91, + 130, + 879, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingpei Lu, Florian Richter, and Michael C. Yip \nUniversity of California, San Diego \n{jil360, frichter, yip}@ucsd.edu", + "bbox": [ + 294, + 203, + 673, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Solving the camera-to-robot pose is a fundamental requirement for vision-based robot control, and is a process that takes considerable effort and cares to make accurate. Traditional approaches require modification of the robot via markers, and subsequent deep learning approaches enabled markerless feature extraction. Mainstream deep learning methods only use synthetic data and rely on Domain Randomization to fill the sim-to-real gap, because acquiring the 3D annotation is labor-intensive. In this work, we go beyond the limitation of 3D annotations for real-world data. We propose an end-to-end pose estimation framework that is capable of online camera-to-robot calibration and a self-supervised training method to scale the training to unlabeled real-world data. Our framework combines deep learning and geometric vision for solving the robot pose, and the pipeline is fully differentiable. To train the Camera-to-Robot Pose Estimation Network (CtRNet), we leverage foreground segmentation and differentiable rendering for image-level self-supervision. The pose prediction is visualized through a renderer and the image loss with the input image is back-propagated to train the neural network. Our experimental results on two public real datasets confirm the effectiveness of our approach over existing works. We also integrate our framework into a visual servoing system to demonstrate the promise of real-time precise robot pose estimation for automation tasks.", + "bbox": [ + 75, + 325, + 473, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 752, + 209, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The majority of modern robotic automation utilizes cameras for rich sensory information about the environment to infer tasks to be completed and provide feedback for closed-loop control. The leading paradigm for converting the valuable environment information to the robot's frame of reference for manipulation is position-based visual servoing (PBVS) [4]. At a high level, PBVS converts 3D environmental information inferred from the visual data (e.g.", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b2f681605f396bb678513a9a136565681ad6b2b3c8d002f356fe9c3af69f236.jpg", + "image_caption": [ + "Figure 1. Comparison of speed and accuracy (based on AUC metric) for existing image-based robot pose estimation methods." + ], + "image_footnote": [], + "bbox": [ + 535, + 289, + 857, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the pose of an object to be grasped) and transforms it to the robot coordinate frame where all the robot geometry is known (e.g. kinematics) using the camera-to-robot pose. Examples of robotic automation using the PBVS range from bin sorting [35] to tissue manipulation in surgery [31].", + "bbox": [ + 496, + 597, + 890, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Calibrating camera-to-robot pose typically requires a significant amount of care and effort. Traditionally, the camera-to-robot pose is calibrated with externally attached fiducial markers (e.g. Aruco Marker [14], AprilTag [38]). The 2D location of the marker can be extracted from the image and the corresponding 3D location on the robot can be calculated with forward kinematics. Given a set 2D-3D correspondence, the camera-to-robot pose can be solved using Perspective-n-Point (PnP) methods [13, 30]. The procedure usually requires multiple runs with different robot configurations and once calibrated, the robot base and the camera are assumed static. The incapability of online calibration limits the potential applications for vision-based robot control in the real world, where minor bumps or simply shifting due to repetitive use will cause calibrations to be thrown off,", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21296", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "not to mention real-world environmental factors like vibration, humidity, and temperature, are non-constant. Having flexibility on the camera and robot is more desirable so that the robot can interact with an unstructured environment.", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning, known as the current state-of-the-art approach for image feature extraction, brings promising ways for markerless camera-to-robot calibration. Current approaches to robot pose estimation are mainly classified into two categories: keypoint-based methods [27-29,34,43] and rendering-based methods [16,26]. Keypoint-based methods are the most popular approach for pose estimation because of the fast inference speed. However, the performance is limited to the accuracy of the keypoint detector which is often trained in simulation such that the proposed methods can generalize across different robotic designs. Therefore, the performance is ultimately hampered by the sim-to-real gap, which is a long-standing challenge in computer vision and robotics [55].", + "bbox": [ + 75, + 154, + 470, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rendering-based methods can achieve better performance by using the shape of the entire robot as observation, which provides dense correspondence for pose estimation. The approaches in this category usually employ an iterative refinement process and require a reasonable initialization for the optimization loop to converge [32]. Due to the nature that iteratively render and compare is time- and energy-consuming, rendering-based methods are more suitable for offline estimation where the robot and camera are held stationary. In more dynamic scenarios, such as a mobile robot, the slow computation time makes the rendering-based methods impracticable to use.", + "bbox": [ + 75, + 368, + 468, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose CtRNet, an end-to-end framework for robot pose estimation which, at inference, uses keypoints for the fast inference speed and leverages the high performance of rendering-based methods for training to overcome the sim-to-real gap previous keypoint-based methods faced. Our framework contains a segmentation module to generate a binary mask of the robot and keypoint detection module which extracts point features for pose estimation. Since segmenting the robot from the background is a simpler task than estimating the robot pose and localizing point features on robot body parts, we leverage foreground segmentation to provide supervision for the pose estimation. Toward this direction, we first pretrained the network on synthetic data, which should have acquired essential knowledge about segmenting the robot. Then, a self-supervised training pipeline is proposed to transfer our model to the real world without manual labels. We connect the pose estimation to foreground segmentation with a differentiable renderer [24, 33]. The renderer generates a robot silhouette image of the estimated pose and directly compares it to the segmentation result. Since the entire framework is differentiable, the parameters of the neural network can be optimized by back-propagating the image loss.", + "bbox": [ + 75, + 553, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Our main contribution is the novel framework for image-based robot pose estimation together with a scalable self-training pipeline that utilizes unlimited real-world data to further improve the performance without any manual annotations. Since the keypoint detector is trained with image-level supervision, we effectively encompass the benefits from both keypoint-based and rendering-based methods, where previous methods were divided. As illustrated in the Fig. 1, our method maintains high inference speed while matching the performance of the rendering-based methods. Moreover, we integrate the CtR-Net into a robotic system for PBVS and demonstrate the effectiveness on real-time robot pose estimation.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 301, + 648, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Camera-to-Robot Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 327, + 799, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The classical way to calibrate the camera-to-robot pose is to attach the fiducial markers [14, 38] to known locations along the robot kinematic chain. The marker is detected in the image frame and their 3D position in the robot base frame can be calculated with forward kinematics. With the geometrical constraints, the robot pose can be then derived by solving an optimization problem [11, 20, 22, 39].", + "bbox": [ + 496, + 351, + 890, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works on markerless articulated pose tracking utilize a depth camera for 3D observation [10, 37, 41, 45]. For a high degree-of-freedom articulated robot, Bohg et al. [1] proposed a pose estimation method by first classifying the pixels in depth image to robot parts, and then a voting scheme is applied to estimate the robot pose relative to the camera. This method is further improved in [52] by directly training a Random Forest to regress joint angles instead of part label. However, these methods are not suitable for our scenario where only single RGB image is available.", + "bbox": [ + 496, + 457, + 890, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, as deep learning becomes popular in feature extraction, many works have been employing deep neural networks for robot pose estimation. Instead of using markers, a neural network is utilized for keypoint detection, and the robot pose is estimated through an optimizer (e.g. PnP solver) [28, 29, 34, 56]. To further improve the performance, the segmentation mask and edges are utilized to refine the robot pose [16, 27]. Labbe et al. [26] also introduces the render&compare method to estimate the robot pose by matching the robot shape. These methods mainly rely on synthetically generated data for training and hope the network can generalize to the real world by increasing the variance in data generation. Our method explicitly deals with the sim-to-real transfer by directly training on real-world data with self-supervision.", + "bbox": [ + 496, + 609, + 890, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Domain Adaptation for Sim-to-Real Transfer", + "text_level": 1, + "bbox": [ + 500, + 845, + 880, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In computer vision and robotics, Domain Randomization (DR) [48] is the most widely used method for sim", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21297", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to-real transfer due to its simplicity and effectiveness. The idea is to randomize some simulation parameters (e.g. camera position, lighting, background, etc.) and hope that the randomization captures the distribution of the real-world data. This technique has been applied to object detection and grasping [2, 18, 21, 47, 49], and pose estimation [26, 28, 29, 34, 36, 46, 50, 56]. The randomization is usually tuned empirically hence it is not efficient.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another popular technique for domain transfer is Domain Adaptation (DA), which is to find the feature spaces that share a similar distribution between the source and target domains [51]. This technique has shown recent success in computer vision [7, 19, 44] and robotic applications [3, 15, 23]. In this work, instead of finding the latent space and modeling the distribution between the simulation and the real world, we perform sim-to-real transfer by directly training on the real-world data via a self-training pipeline.", + "bbox": [ + 75, + 212, + 470, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 378, + 174, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce an end-to-end framework for robot pose estimation and a scalable training pipeline to improve pose estimation accuracy on real-world data without the need for any manual annotation. We first explain the self-supervised training pipeline for sim-to-real transfer in Sec. 3.1 given a pretrained CtRNet on synthetic data which both segments the robot and estimates its pose from images. Then, we detail the camera-to-robot pose estimation network in Sec. 3.2 which utilizes a keypoint detector and a PnP solver to estimate the pose of the robot from image data in real-time.", + "bbox": [ + 75, + 405, + 470, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Self-supervised Training for Sim-to-Real Transfer", + "text_level": 1, + "bbox": [ + 76, + 582, + 468, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most effective way to adapt the neural network to the real world is directly training the network on real sensor data. We propose a self-supervised training pipeline for sim-to-real transfer to facilitate the training without 3D annotations. To conduct the self-supervised training, we employ foreground segmentation to generate a mask of the robot, $f_{seg}$ , alongside the pose estimation, $f_{pose}$ . Given an input RGB image from the physical world, $\\mathbb{I}$ , and the robot joint angles, $\\mathbf{q}$ , $f_{pose}$ estimates the robot pose which is then transformed to a silhouette image through a differentiable renderer. Our self-supervised objective is to optimize neural network parameters by minimizing the difference between the rendered silhouette image and the mask image. We formulate the optimization problem as:", + "bbox": [ + 75, + 621, + 468, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta_ {b b}, \\theta_ {k p}, \\theta_ {s e g} = \\underset {\\theta_ {b b}, \\theta_ {k p}, \\theta_ {s e g}} {\\arg \\min } \\mathcal {L} [ f _ {s e g} (\\mathbb {I} | \\theta_ {b b}, \\theta_ {s e g}), \\\\ \\left. \\mathcal {R} \\left(f _ {p o s e} \\left(\\mathbb {I} \\mid \\mathbf {q}, \\theta_ {b b}, \\theta_ {k p}\\right) \\mid \\mathbf {K}\\right) \\right] \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 849, + 468, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{bb},\\theta_{kp},\\theta_{seg}$ denote the parameters of the backbone, keypoint, and segmentation layers of the neural network. $\\mathcal{R}$ is the differentiable renderer with camera parameters $\\mathbf{K}$ , and $\\mathcal{L}(.)$ is the objective loss function capturing the image difference.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We pretrained CtRNet's parameters which makes up $f_{seg}$ and $f_{pose}$ , with synthetic data where the keypoint and segmentation labels are obtained freely (details in Supplementary Materials). During the self-training phase, where CtRNet learns with real data, the objective loss in (1) captures the difference between the segmentation result and the rendered image. The loss is iteratively back-propagated to, $\\Theta$ , where each iteration $f_{seg}$ and $f_{pose}$ take turns learning from each other to overcome the sim-to-real gap.", + "bbox": [ + 496, + 167, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. The overview of the self-supervised training pipeline is shown in the Fig. 2. The segmentation module, $f_{seg}$ , simply takes in a robot image and outputs its mask. The pose estimation module, $f_{pose}$ , consists of a keypoint detector and a PnP solver to estimate the robot pose using the 2D-3D point correspondence, as shown in Fig. 3. Given the input robot image and joint angles, our camera-to-robot pose estimation network outputs a robot mask and the robot pose with respect to the camera frame. Mathematically, these functions are denoted as", + "bbox": [ + 496, + 303, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {M} = f _ {s e g} (\\mathbb {I} | \\theta_ {b b}, \\theta_ {s e g}) \\quad \\mathbf {T} _ {b} ^ {c} = f _ {p o s e} (\\mathbb {I} | \\mathbf {q}, \\theta_ {b b}, \\theta_ {k p}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 465, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{M}$ is the robot mask and $\\mathbf{T}_b^c\\in SE(3)$ is the 6-DOF robot pose. Finally, the self-supervised objective loss in (1) is realized through a differentiable renderer, $\\mathcal{R}$ , which generates a silhouette image of the robot given its pose, $\\mathbf{T}_b^c$ .", + "bbox": [ + 496, + 494, + 890, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differentiable Rendering. To render the robot silhouette image, we utilize the PyTorch3D differentiable render [42]. We initialize a perspective camera with intrinsic parameters $\\mathbf{K}$ and a silhouette renderer, which does not apply any lighting nor shading, is constructed with a rasterizer and a shader. The rasterizer applies the fast rasterization method [42] which selects the $k$ nearest mesh triangles that effects each pixel and weights their influence according to the distance along the $z$ -axis. Finally, the SoftSilhouetteShader is applied to compute pixel values of the rendered image using the sigmoid blending method [33].", + "bbox": [ + 496, + 556, + 890, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We construct the ready-to-render robot mesh by connecting the CAD model for each robot body part using its forward kinematics and transforming them to the camera frame with the estimated robot pose $\\mathbf{T}_b^c$ from $f_{pose}$ . Let $\\mathbf{v}^n\\in \\mathbb{R}^3$ be a mesh vertex on the $n$ -th robot link. Each vertex is transformed to the camera frame, hence ready-to-render, by", + "bbox": [ + 496, + 723, + 890, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {v}}} ^ {c} = \\mathbf {T} _ {b} ^ {c} \\mathbf {T} _ {n} ^ {b} (\\mathbf {q}) \\overline {{\\mathbf {v}}} ^ {n} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 825, + 890, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\cdot}$ represents the homogeneous representation of a point (e.g. $\\overline{\\mathbf{v}} = [\\mathbf{v},1]^T$ ), and $\\mathbf{T}_n^b (\\mathbf{q})$ is the coordinate frame transformation obtained from the forward kinematics [8].", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21298", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/487c85ebb839ce93b2f3bf055f973ac50b3fd222f099b3e1a66038647ba7ea86.jpg", + "image_caption": [ + "Figure 2. The overview of our proposed self-supervised training framework for sim-to-real transfer. The CtRNet contains a foreground segmentation module and a pose estimation module, which output a robot mask and a camera-to-robot pose respectively. The output pose is transformed into a silhouette image through a differentiable renderer. The image loss is back-propagated to train the keypoint detector and fine-tune the segmentation." + ], + "image_footnote": [], + "bbox": [ + 130, + 94, + 843, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objective loss function. The objective loss in (1) is iteratively minimized where $f_{seg}$ and $f_{pose}$ take turns supervising each other on real data to overcome the sim-to-real gap faced by keypoint detection networks. To optimize $f_{pose}$ , the L2 image loss is used since the segmentation network's accuracy, within the context of estimating robot poses, has been shown to effectively transfer from simulation to the real world [26]. Mathematically the loss is expressed as", + "bbox": [ + 75, + 398, + 473, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\left(\\mathbb {S} (i, j) - \\mathbb {M} (i, j)\\right) ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 527, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H$ and $W$ is the height and width of the image, and $\\mathbb{S}$ is the rendered silhouette image.", + "bbox": [ + 75, + 579, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the pretrained robot segmentation, $f_{seg}$ , already performs well on real-world datasets, it is still desirable to refine it through self-supervised training to better extract fine details of corners and boundaries. To prevent the foreground segmentation layers from receiving noisy training signals, we apply the weighted Binary Cross Entropy Loss so that the high-quality rendering image can be used to further refine the foreground segmentation:", + "bbox": [ + 75, + 609, + 470, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s e g} = - \\frac {w}{H * W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} [ \\mathbb {M} (i, j) \\log \\mathbb {S} (i, j) \\\\ \\left. + \\left(1 - \\mathbb {M} (i, j)\\right) \\log (1 - \\mathbb {S} (i, j)) \\right]. \\quad (5) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 742, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w$ is the weight for the given training sample. For PnP solvers, the optimal solution should minimize the point reprojection error. Therefore, we assign the weight for each training sample according to the reprojection error:", + "bbox": [ + 75, + 815, + 470, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw = \\exp (- s O (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f4a3552f2a4e6f01ea8e98791810acaad9a4813330d2171b403ac9c906eec8d.jpg", + "image_caption": [ + "Figure 3. The diagram of the camera-to-robot pose estimation network (CtRNet) which describes the inference process of network. Given an input RGB image, the neural network generates a robot mask and a set of keypoint. Given the associated robot joint angles, a set of corresponding 3D keypoints are computed with forward kinematics. The camera-to-robot pose is estimated by a PnP solver with provided 2D-3D keypoint pairs." + ], + "image_footnote": [], + "bbox": [ + 506, + 398, + 885, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s$ is a scaling constant, $O$ is the reprojection loss in the PnP solver (explained in Sec. 3.2), $\\{\\mathbf{o}_i|\\mathbf{o}_i\\in \\mathbb{R}^2\\}_{i = 1}^n$ and $\\{\\mathbf{p}_i|\\mathbf{p}_i\\in \\mathbb{R}^3\\}_{i = 1}^n$ are the 2D-3D keypoints inputted into the PnP solver. The exponential function is applied to the weight such that training samples with poor PnP convergence are weighted exponentially lower than good PnP convergence thereby stabilizing the training.", + "bbox": [ + 496, + 727, + 893, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Camera-to-Robot Pose Estimation Network", + "text_level": 1, + "bbox": [ + 500, + 845, + 870, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overview of the proposed Camera-to-Robot Pose Estimation Network, CtRNet, is shown in Fig. 3. Given an", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21299", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input RGB image, we employ ResNet50 [17] as the backbone network to extract the latent features. The latent features are then passed through the Atrous Spatial Pyramid Pooling layers [6] to form the segmentation mask of input resolution. The keypoint detector, sharing the backbone network with the foreground segmentation, upsamples the feature maps through transposed convolutional layers and forms the heatmaps with $n$ channels. Then, we apply the spatial softmax operator [12] on the heatmaps, which computes the expected 2D location of the points of maximal activation for each channel and results in a set of keypoints $[\\mathbf{o}_1,\\dots,\\mathbf{o}_n]$ for all $n$ channels. For simplicity, we define the set of keypoints at each joint location of the robot. Given the joint angles, the corresponding 3D keypoint location $\\mathbf{p}_i$ can be calculated with robot forward kinematics:", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {p}}} _ {i} = \\mathbf {T} _ {i} ^ {b} (\\mathbf {q}) \\overline {{\\mathbf {t}}}, f o r i = 1, \\dots , n \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 325, + 468, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{t} = [0,0,0]$ . With the 2D and 3D corresponding keypoints, we can then apply a PnP solver [30] to estimate the robot pose with respect to the camera frame.", + "bbox": [ + 75, + 352, + 468, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Back-propagation for PnP Solver. A PnP solver is usually self-contained and not differentiable as the gradient with respect to the input cannot be derived explicitly. Inspired by [5], the implicit function theorem [25] is applied to obtain the gradient through implicit differentiation. Let the PnP solver be denoted as followed in the form of a non-linear function $g$ :", + "bbox": [ + 75, + 398, + 468, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {b} ^ {c *} = g (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 513, + 468, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_b^{c*}$ is output pose from the PnP solver. In order to back-propagate through the PnP solver for training the keypoint detector, we are interested in finding the gradient of the output pose $\\mathbf{T}_b^{c*}$ with respect to the input 2D points $\\mathbf{o}$ . Note that, the objective of the PnP solver is to minimize the reprojection error, such that:", + "bbox": [ + 75, + 539, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {b} ^ {c *} = \\underset {\\mathbf {T} _ {b} ^ {c}} {\\arg \\min } O (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 638, + 468, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with", + "bbox": [ + 76, + 676, + 112, + 688 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} O \\left(\\mathbf {T} _ {b} ^ {c}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}\\right) = \\sum_ {i = 1} ^ {n} | | \\mathbf {o} _ {i} - \\pi \\left(\\mathbf {p} _ {i} \\mid \\mathbf {T} _ {b} ^ {c}, \\mathbf {K}\\right) | | _ {2} ^ {2} (10) \\\\ = \\sum_ {i = 1} ^ {n} \\left\\| \\mathbf {r} _ {i} \\right\\| _ {2} ^ {2} (11) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 696, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi(.)$ is the projection operator. Since the optimal solution $\\mathbf{T}_b^{c*}$ is a local minimum for the objective function $O(\\mathbf{o},\\mathbf{p},\\mathbf{T}_b^c,\\mathbf{K})$ , a stationary constraint of the optimization process can be constructed by taking the first order derivative of the objective function with respect to $\\mathbf{T}_b^c$ :", + "bbox": [ + 75, + 786, + 468, + 863 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial O}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) | _ {\\mathbf {T} _ {b} ^ {c} = \\mathbf {T} _ {b} ^ {c *}} = \\mathbf {0}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 869, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [5], we construct a constrain function $F$ to employ the implicit function theorem:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) = \\frac {\\partial O}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c *}) = \\mathbf {0}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 131, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Substituting the Eq. (10) and Eq. (11) to Eq. (13), we can derive the constraint function as:", + "bbox": [ + 498, + 175, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) = \\sum_ {i = 1} ^ {n} \\frac {\\partial \\left\\| \\mathbf {r} _ {i} \\right\\| _ {2} ^ {2}}{\\partial \\mathbf {T} _ {b} ^ {c}} (14) \\\\ = - 2 \\sum_ {i = 1} ^ {n} \\mathbf {r} _ {i} ^ {T} \\frac {\\partial \\pi}{\\partial \\mathbf {T} _ {b} ^ {c}} \\left(\\mathbf {p} _ {i} \\mid \\mathbf {T} _ {b} ^ {c *}, \\mathbf {K}\\right). (15) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 214, + 890, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we back-propagate through the PnP solver with the implicit differentiation. The gradient of the output pose with respect to the input 2D points is the Jacobian matrix:", + "bbox": [ + 498, + 306, + 890, + 353 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial g}{\\partial \\mathbf {o}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}) \\\\ = - \\left(\\frac {\\partial F}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})\\right) ^ {- 1} \\left(\\frac {\\partial F}{\\partial \\mathbf {o}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})\\right). \\tag {16} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 378, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 473, + 632, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate our method on two public real-world datasets for robot pose estimation and compare it against several state-of-the-art image-based robot pose estimation algorithms. We then conduct an ablation study on the pretraining procedure and explore how the number of pretraining samples could affect the performance of the self-supervised training. Finally, we integrate the camera-to-robot pose estimation framework into a visual servoing system to demonstrate the effectiveness of our method on real robot applications.", + "bbox": [ + 496, + 500, + 890, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 660, + 785, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DREAM-real Dataset. The DREAM-real dataset [29] is a real-world robot dataset collected with 3 different cameras: Azure Kinect (AK), XBOX 360 Kinect (XK), and RealSense (RS). This dataset contains around 50K RGB images of Franka Emika Panda arm and is recorded at $(640 \\times 480)$ resolution. The ground-truth camera-to-robot pose is provided for every image frame. The accuracy is evaluated with average distance (ADD) metric [54],", + "bbox": [ + 496, + 683, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA D D = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| \\widetilde {\\mathbf {T}} _ {b} ^ {c} \\overline {{\\mathbf {p}}} _ {i} - \\mathbf {T} _ {b} ^ {c} \\overline {{\\mathbf {p}}} _ {i} \\right| \\right| _ {2} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 816, + 890, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathbf{T}}_b^c$ indicates the ground-truth camera-to-robot pose. We also report the area-under-the-curve (AUC) value,", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21300", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d1cad9aaa05ff6e201cc9e77ae0912acf99c8b5d0008cebc053cc5c3025cef0f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCategoryBackbonePanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORBAll
AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓
DREAM-F [29]KeypointVGG1968.91211.41324.359491.91176.1302.07761.93095.31960.740113.029
DREAM-Q [29]KeypointVGG1952.38278.08937.47154.17877.9840.02757.08767.24856.98859.284
DREAM-H [29]KeypointResNet10160.5200.05664.0057.38278.8250.02469.05425.68568.58417.477
RoboPose [26]RenderingResNet3476.4970.02485.9260.01476.8630.02380.5040.01980.0940.020
CtRNetKeypointResNet5089.9280.01379.4650.03290.7890.01085.2890.02185.9620.020
", + "bbox": [ + 84, + 88, + 885, + 191 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3c5a4b987d667a9a37f5aa0e894e7cc0af221469dbde39755eeacb3f3d9a516f.jpg", + "table_caption": [ + "Table 1. Comparison of our methods with the state-of-the-art methods on DREAM-real datasets using ADD metric. We report the mean and AUC of the ADD on each dataset and the overall accuracy." + ], + "table_footnote": [], + "table_body": "
MethodCategoryPCK (2D)Mean 2D Err. (pixel) ↓ADD (3D)Mean 3D Err. (mm) ↓
@50 pixel ↑AUC ↑@100 mm ↑AUC ↑
Aruco Marker [14]Keypoint0.4957.15286.980.3043.452447.34
DREAM-Q [29]Keypoint0.3344.011277.330.3240.63386.17
Opt. Keypoints [34]Keypoint0.6975.4649.510.4765.66141.05
Diff. RenderingRendering0.7478.6042.300.7881.1574.95
CtRNetKeypoint0.9993.9411.620.8883.9363.81
", + "bbox": [ + 142, + 243, + 831, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of our methods with the state-of-the-art methods on Baxter dataset.", + "bbox": [ + 220, + 377, + 746, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which integrates the percentage of ADD over different thresholds. A higher AUC value indicates more predictions with less error.", + "bbox": [ + 75, + 417, + 468, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baxter Dataset. The Baxter dataset [34] contains 100 RGB images of the left arm of Rethink Baxter collected with Azure Kinect camera at $(2048\\times 1526)$ resolution. The 2D and 3D ground-truth end-effector position with respect to the camera frame is provided. We evaluate the performance with the ADD metric for the end-effector. We also evaluate the end-effector reprojection error using the percentage of correct keypoints (PCK) metric [34].", + "bbox": [ + 75, + 465, + 468, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 603, + 290, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The entire pipeline is implemented in PyTorch [40]. We initialize the backbone network with ImageNet [9] pretrained weights, and we train separate networks for different robots. The number of keypoints $n$ is set to the number of robot links and the keypoints are defined at the robot joint locations. The neural network is pretrained on synthetic data for foreground segmentation and keypoint detection for 1000 epochs with 1e-5 learning rate. We reduce the learning rate by a factor of 10 once learning stagnates for 5 epochs. The Adam optimizer is applied to optimize the network parameters with the momentum set to 0.9. For self-supervised training on real-world data, we run the training for 500 epochs with 1e-6 learning rate. The same learning rate decay strategy and Adam optimizer is applied here similar to the pretraining. To make the training more stable, we clip the gradient of the network parameters at 10. The scaling factor in Eq. (6) is set to 0.1 for DREAM-real dataset and 0.01 for Baxter dataset, mainly accounting for", + "bbox": [ + 73, + 628, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the difference in resolution.", + "bbox": [ + 500, + 417, + 683, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Robot Pose Estimation on Real-world Datasets", + "text_level": 1, + "bbox": [ + 498, + 449, + 890, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on DREAM-real Dataset. The proposed CtRNet is trained at $(320\\times 240)$ resolution and evaluated at the original resolution by scaling up the keypoints by a factor of 2. Some qualitative results for foreground segmentation and pose estimation are shown in the Fig. 4a. We compared our method with the state-of-the-art keypoint-based method DREAM [29] and the rendering-based method RoboPose [26]. The results for DREAM and RoboPose are compiled from the implementation provided by [26]. In Tab. 1, we report the AUC and mean ADD results on DREAM-real dataset with 3 different camera settings and the overall results combining all the test samples. Our method has a significantly better performance compared to the method in the same category and achieves comparable performance with the rendering-based method. We outperform DREAM on all settings and outperform RoboPose on the majority of the dataset. Overall on DREAM-real dataset, we achieve higher AUC (+17.378 compared to DREAM, +5.868 compared to RoboPose), and lower error compared to DREAM (-17.457).", + "bbox": [ + 496, + 474, + 892, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on Baxter Dataset. For the Baxter dataset, we trained the CtRNet at $(640\\times 480)$ resolution and evaluate at the original resolution, and Fig. 4b shows some of the qualitative results. We compared our method with several keypoint-based methods (Aruco Marker [14], DREAM [29], Optimized Keypoints [34]). We also implemented Differentiable Rendering for robot pose estimation, where the robot masks are generated with the pre", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21301", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d761695e91347c02df25ad2c1c064d94da4919065ca4dcacf2b0ca7dd026283a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 89, + 236, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/317d59fc022df35c7f281319d92388d65c112eb62a59ac36babe4e7adb56cbd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 90, + 359, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07a26f618bac91534b11f760792f59a3c019dd37cf42482f2f7ef295667f7eab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 90, + 480, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/886bc7de660b5a96b67dbfc72fe635f18e06ea295c3d57ce79bd126e06b34510.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 607, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/225fcf48a322d09a908a616bbc4beb8d42c0e09ae5b30e8366259a012c9f7734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 90, + 730, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52a07cabea14f9c7ccc54fe6c837b32a0ba69d642db33f5d63d5af54e5c2f3d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 90, + 856, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09b344d883b9e1271789ee2b0a240ec5e01647c4d94f1f4ddd4787ff3b937cc3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 171, + 236, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/631e13a4bf8f969d0d2bf1fbd52082cded94d76e2e152f04e263059fb85b8dea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 171, + 359, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e40a327b4bf64a02913c387ff1d7b3602d0abf79d9fb6b07179c6ed3cc0ac289.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 171, + 480, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15fe4104716543d754256f385e26d8417d8066bc9938e42fdffaadbcb7197019.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 171, + 607, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f54349c7c60ab6c481e78c669aa967c090b5c2eeed7f94b21cd956f796d189e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 171, + 730, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/840fbef587f090e443af1a6ba3959736cb7453e461febda1ccd4d62a18f2d19a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 171, + 856, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bdf50a92cf4823f9db92e0648d9cb853bd85e179b382c1c9331c18275b7cba5e.jpg", + "image_caption": [ + "Figure 4. Qualitative results of CtRNet foreground segmentation and pose estimation on (a) DREAM-real dataset and (b) Baxter dataset. The first row shows the input RGB image, the second row shows the foreground segmentation, and the third row shows the projected robot skeleton based on the estimated robot pose." + ], + "image_footnote": [], + "bbox": [ + 114, + 252, + 236, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19e24995ff490a764033db121bf6bf0917118d458173f83809f3cde709516e73.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 236, + 252, + 359, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/757ec12934e15a3373e017d8039a3d4dbecea80472284074ad7393f12e9c55ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 252, + 480, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be4661f3725e6a40a3e0a8517df899cd173db4081d811590d4012f51d6a89aee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 252, + 607, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4deb021922b3c713d04345603c9af0146adce6cb2fc75a875387565562aa5ee.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 609, + 252, + 730, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d2a320cf8745b6d38f188ddfa566554e19015860136306581df6041b9f95130.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 252, + 856, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained foreground segmentation. The 2D PCK results and 3D ADD results are reported in Tab. 2. Our method outperforms all other methods on both 2D and 3D evaluations. For 2D evaluation, we achieve 93.94 AUC for PCK with an average reprojection error of 11.62 pixels. For 3D evaluation, we achieve 83.93 AUC for ADD with an average ADD of $63.81\\mathrm{mm}$ . Notably, 99 percent of our estimation has less than 50 pixel reprojection error, which is less than 2 percent of the image resolution, and 88 percent of our estimation has less than $100\\mathrm{mm}$ distance error when localizing the end-effector.", + "bbox": [ + 75, + 429, + 472, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 604, + 228, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We study how the number of pretraining samples affects the convergence and performance of the self-supervised training empirically on the Baxter dataset. We pretrain the neural network with different numbers of synthetic data samples $N_{\\text{pretrain}} = \\{500, 1000, 2000, 4000, 8000\\}$ , and examine the convergence of the self-supervised training process. Fig. 5 shows the plot of self-training loss $(\\mathcal{L}_{\\text{mask}} + \\mathcal{L}_{\\text{seg}})$ vs. the number of epochs for networks pretrianed with different number of synthetic data. We observe that doubling the size of pretraining dataset significantly improves the convergence of the self-training process at the beginning. However, the improvement gets smaller as the pretraining size increases. For the Baxter dataset, the improvement saturates after having more than 2000 pretraining samples. Continuing double the training size results in very marginal improvement. Noted that the Baxter dataset captures 20 different robot poses from a fixed camera position. The required number of pretraining samples might", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1cdde4800a9067745e8bfe7b3d7d8b96b8e86483212ef2d94491268d300ebb1.jpg", + "image_caption": [ + "Figure 5. The training loss vs. number of epochs for the self-supervised training with different numbers of pretraining samples. More pretraining samples results in better convergence. The improvement saturates after having more than 2000 pretraining samples as only marginal improvement by adding more samples." + ], + "image_footnote": [], + "bbox": [ + 517, + 429, + 875, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vary according to the complexity of the environment.", + "bbox": [ + 500, + 699, + 849, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate the resulting neural networks with the ground-truth labels on the Baxter dataset. We report the mean ADD and AUC ADD for the pose estimation in Tab. 3. The result verifies our observation on the convergence analysis. Having more pretraining samples improves the performance of pose estimation at the beginning, but the improvement stagnates after having more than 2000 pretraining samples.", + "bbox": [ + 496, + 714, + 890, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visual Servoing Experiment", + "text_level": 1, + "bbox": [ + 500, + 845, + 751, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We integrate the proposed CtRNet into a robotic system for position-based visual servoing (PBVS) with eye-to-hand", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21302", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/09aee82883ac1405300602a57aae7d4d394e81529cc67f10c74e02c51d813cb9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NpretrainMean ADD (mm) ↓AUC ADD ↑
5002167.3047.62
100092.9176.65
200067.5182.98
400063.0084.12
800063.8183.93
", + "bbox": [ + 112, + 88, + 433, + 195 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c12820ca3fcede1cf7a4bf56ff60f57f525ebc271b4b9630d4c265acaf7019ee.jpg", + "table_caption": [ + "Table 3. Ablation study for the number of pretraining samples." + ], + "table_footnote": [], + "table_body": "
MethodLoop RateTrans. Err. (m)Rot. Err. (rad)
DREAM [29]30Hz0.235 ± 0.3130.300 ± 0.544
Diff. Rendering1Hz0.046 ± 0.0620.036 ± 0.066
CtRNet30Hz0.002 ± 0.0010.002 ± 0.001
", + "bbox": [ + 81, + 234, + 467, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Mean and standard deviation of the translational error and rotational error for the visual servoing experiment.", + "bbox": [ + 76, + 315, + 467, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "configuration. We conduct the experiment on a Baxter robot and the details of the PBVS are described in the Supplementary Materials. The PBVS is purely based on RGB images from a single camera and the goal is to control the robot end-effector reaching a target pose defined in the camera frame. Specifically, we first set a target pose with respect to the camera frame. The target pose is then transformed into the robot base frame through the estimated camera-to-robot transformation. The robot controller calculates the desired robot configuration with inverse kinematics and a control law is applied to move the robot end-effector toward the target pose.", + "bbox": [ + 75, + 371, + 467, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For comparison, we also implemented DREAM [29] and a Differentiable Renderer for PBVS. For DREAM, the pretrained model for Baxter is applied. For Differentiable Renderer, we use the foreground segmentation of CtRNet to generate a robot mask. The optimizing loop for the renderer takes the last estimation as initialization and performs 10 updates at each callback to ensure convergence and maintain $1\\mathrm{Hz}$ loop rate. In the experiment, we randomly set the target pose and the position of the camera, and the robotic system applies PBVS to reach the target pose from an arbitrary initialization, as shown in the Fig. 6. We ran the experiment for 10 trails with different robot pose estimation methods, and the translational (Euclidean distance) and rotational errors (Euler angles) of the end-effector are reported in Tab. 4. The experimental results show that our proposed method significantly improves the stability and accuracy of the PBVS, achieving $0.002\\mathrm{m}$ averaged translational error and $0.002\\mathrm{rad}$ rotational error on the end-effector.", + "bbox": [ + 75, + 551, + 467, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also plot the end-effector distance-to-goal over time for a selected trail in Fig. 7. In this selected trial, the system could not converge with DREAM because the poor robot pose estimation confuses the controller by giving the wrong target pose in the robot base frame, which is unreachable.", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba92e43337ba0f497e943dd52229464de4eaca582f08f4da2069cfd9100316e8.jpg", + "image_caption": [ + "Figure 6. Snapshots of PBVS. The goal is to move the end-effector to the target pose (green). The figure on the right shows the robot configuration upon the convergence of PBVS." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 694, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbf61d722171064c1276569aa84bbcc2864d13de42fcd0708f54f01e18413f43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 90, + 877, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c84a6549ed07db415ab31a8f79c3667aaa14a8ab97efd1363a20cc8a1a25b89e.jpg", + "image_caption": [ + "Figure 7. The plot of end-effector distance-to-goal over time on a selected PBVS trail." + ], + "image_footnote": [], + "bbox": [ + 516, + 267, + 874, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the differentiable renderer, the servoing system takes more than 10 seconds to converge and oscillate due to the low loop rate. With our proposed CtRNet, the servoing system converges much faster ( $\\leq 5$ seconds), thanks to the fast and robust robot pose estimation. We show more qualitative results in the Supplementary Materials.", + "bbox": [ + 498, + 467, + 890, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 573, + 617, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the CtRNet, an end-to-end image-based robot pose estimation framework, and a self-supervised training pipeline that utilizes unlabelled real-world data for sim-to-real transfer. The CtRNet, using a keypoint detector for pose estimation while employing a rendering method for training, achieves state-of-the-art performance on robot pose estimation while maintaining high-speed inference. The Fig. 1 illustrates the advantages of CtRNet over existing methods, where the AUC values are normalized across two evaluation datasets by taking DREAM and CtRNet as references. We further experiment with different robot pose estimation methods by applying them to PBVS, which demonstrates CtRNet's fast and accurate robot pose estimation enabling stability when using single-frame robot pose estimation for feedback. Therefore, CtRNet supports real-time markerless camera-to-robot pose estimation which has been utilized for surgical robotic manipulation [43] and mobile robot manipulators [53]. For future work, we would like to extend our method to more robots and explore vision-based control in an unstructured environment.", + "bbox": [ + 496, + 598, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jeannette Bohg, Javier Romero, Alexander Herzog, and Stefan Schaal. Robot arm pose estimation through pixel-wise part classification. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 3143-3150. IEEE, 2014. 2", + "[2] Joao Borrego, Rui Figueiredo, Atabak Dehban, Plinio Moreno, Alexandre Bernardino, and José Santos-Victor. A generic visual perception domain randomisation framework for gazebo. In 2018 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC), pages 237-242. IEEE, 2018. 3", + "[3] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4243-4250. IEEE, 2018. 3", + "[4] François Chaumette, Seth Hutchinson, and Peter Corke. Visual servoing. In Springer Handbook of Robotics, pages 841-866. Springer, 2016. 1", + "[5] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8100-8109, 2020. 5", + "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 5", + "[7] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3339-3348, 2018. 3", + "[8] Jacques Denavit and Richard S Hartenberg. A kinematic notation for lower-pair mechanisms based on matrices. 1955. 3", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[10] Karthik Desingh, Shiyang Lu, Anthony Opipari, and Odest Chadwicke Jenkins. Factored pose estimation of articulated objects using efficient nonparametric belief propagation. In 2019 International Conference on Robotics and Automation (ICRA), pages 7221-7227. IEEE, 2019. 2", + "[11] Irene Fassi and Giovanni Legnani. Hand to sensor calibration: A geometrical interpretation of the matrix equation $\\mathbf{ax} =$ xb. Journal of Robotic Systems, 22(9):497-506, 2005. 2", + "[12] Chelsea Finn, Xin Yu Tan, Yan Duan, Trevor Darrell, Sergey Levine, and Pieter Abbeel. Deep spatial autoencoders for visuomotor learning. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 512-519. IEEE, 2016. 5" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930–943, 2003. 1", + "[14] Sergio Garrido-Jurado, Rafael Muñoz-Salinas, Francisco José Madrid-Cuevas, and Manuel Jesús Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition, 47(6):2280-2292, 2014. 1, 2, 6", + "[15] Abhishek Gupta, Coline Devin, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Learning invariant feature spaces to transfer skills with reinforcement learning. arXiv preprint arXiv:1703.02949, 2017. 3", + "[16] Ran Hao, Orhan Özgün, and M Cenk Çavusoglu. Vision-based surgical tool pose estimation for the da Vinci® robotic surgical system. In 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 1298-1305. IEEE, 2018. 2", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[18] Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, and Kurt Konolige. On pre-trained image features and synthetic images for deep learning. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3", + "[19] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 3", + "[20] Radu Horaud and Fadi Dornaika. Hand-eye calibration. The international journal of robotics research, 14(3):195-210, 1995. 2", + "[21] Daniel Horváth, Gábor Erdős, Zoltán Istenes, Tomás Horváth, and Sándor Foldi. Object detection using sim2real domain randomization for robotic applications. IEEE Transactions on Robotics, 2022. 3", + "[22] Jarmo Ilonen and Ville Kyrki. Robust robot-camera calibration. In 2011 15th International Conference on Advanced Robotics (ICAR), pages 67-74. IEEE, 2011. 2", + "[23] Stephen James, Paul Wohlhart, Mrinal Kalakrishnan, Dmitry Kalashnikov, Alex Irpan, Julian Ibarz, Sergey Levine, Raia Hadsell, and Konstantinos Bousmalis. Sim-to-real via sim-to-sim: Data-efficient robotic grasping via randomized-to-canonical adaptation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12627-12637, 2019. 3", + "[24] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907-3916, 2018. 2", + "[25] Steven George Krantz and Harold R Parks. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002. 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21304", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Single-view robot pose and joint angle estimation via render & compare. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1654-1663, 2021. 2, 3, 4, 6", + "[27] Jens Lambrecht, Philipp Grosenick, and Marvin Meusel. Optimizing keypoint-based single-shot camera-to-robot pose estimation through shape segmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13843–13849. IEEE, 2021. 2", + "[28] Jens Lambrecht and Linh Kästner. Towards the usage of synthetic data for marker-less pose estimation of articulated robots in rgb images. In 2019 19th International Conference on Advanced Robotics (ICAR), pages 240-247. IEEE, 2019. 2, 3", + "[29] Timothy E Lee, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Oliver Kroemer, Dieter Fox, and Stan Birchfield. Camera-to-robot pose estimation from a single image. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9426–9432. IEEE, 2020. 2, 3, 5, 6, 8", + "[30] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision, 81(2):155-166, 2009. 1, 5", + "[31] Yang Li, Florian Richter, Jingpei Lu, Emily K Funk, Ryan K Orosco, Jianke Zhu, and Michael C Yip. Super: A surgical perception framework for endoscopic tissue manipulation with surgical robotics. IEEE Robotics and Automation Letters, 5(2):2294-2301, 2020. 1", + "[32] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 2", + "[33] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 2, 3", + "[34] Jingpei Lu, Florian Richter, and Michael C Yip. Pose estimation for robot manipulators via keypoint optimization and sim-to-real transfer. IEEE Robotics and Automation Letters, 7(2):4622-4629, 2022. 2, 3, 6", + "[35] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 4(26):eaau4984, 2019. 1", + "[36] Fabian Manhardt, Wadim Kehl, Nassir Navab, and Federico Tombari. Deep model-based 6d pose refinement in rgb. In Proceedings of the European Conference on Computer Vision (ECCV), pages 800-815, 2018. 3", + "[37] Frank Michel, Alexander Krull, Eric Brachmann, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Pose estimation of kinematic chain instances via object coordinate regression. In BMVC, pages 181-1, 2015. 2", + "[38] Edwin Olson. Apriltag: A robust and flexible visual fiducial system. In 2011 IEEE international conference on robotics and automation, pages 3400-3407. IEEE, 2011. 1, 2" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Frank C Park and Bryan J Martin. Robot sensor calibration: solving AX= XB on the Euclidean group. IEEE Transactions on Robotics and Automation, 10(5):717-721, 1994. 2", + "[40] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6", + "[41] Karl Pauwels, Leonardo Rubio, and Eduardo Ros. Real-time model-based articulated object pose detection and tracking with variable rigidity constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3994-4001, 2014. 2", + "[42] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 3", + "[43] Florian Richter, Jingpei Lu, Ryan K Orosco, and Michael C Yip. Robotic tool tracking under partially visible kinematic chain: A unified approach. IEEE Transactions on Robotics, 2021. 2, 8", + "[44] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3752-3761, 2018. 3", + "[45] Tanner Schmidt, Richard A Newcombe, and Dieter Fox. Dart: Dense articulated real-time tracking. In Robotics: Science and Systems, volume 2, pages 1-9. Berkeley, CA, 2014. 2", + "[46] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 3", + "[47] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3482-3489. IEEE, 2018. 3", + "[48] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 2", + "[49] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969–977, 2018. 3", + "[50] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. arXiv preprint arXiv:1809.10790, 2018. 3" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21305", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3", + "[52] Felix Widmaier, Daniel Kappler, Stefan Schaal, and Jeannette Bohg. Robot arm pose estimation by pixel-wise regression of joint angles. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 616-623. IEEE, 2016. 2", + "[53] Melonee Wise, Michael Ferguson, Derek King, Eric Diehr, and David Dymesich. Fetch and freight: Standard platforms for service robot applications. In Workshop on autonomous mobile service robots, 2016. 8", + "[54] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 5", + "[55] Wenshuai Zhao, Jorge Pña Queralta, and Tomi Westerlund. Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In 2020 IEEE Symposium Series on Computational Intelligence (SSCI), pages 737-744. IEEE, 2020. 2", + "[56] Yiming Zuo, Weichao Qiu, Lingxi Xie, Fangwei Zhong, Yizhou Wang, and Alan L Yuille. Craves: Controlling robotic arm with a vision-based economic system. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4214-4223, 2019. 2, 3" + ], + "bbox": [ + 78, + 90, + 467, + 444 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21306", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_model.json b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ae0442596b46559c95da343b5959cc50eabee2cc --- /dev/null +++ b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.88, + 0.175 + ], + "angle": 0, + "content": "Markerless Camera-to-Robot Pose Estimation via Self-supervised Sim-to-Real Transfer" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.204, + 0.674, + 0.257 + ], + "angle": 0, + "content": "Jingpei Lu, Florian Richter, and Michael C. Yip \nUniversity of California, San Diego \n{jil360, frichter, yip}@ucsd.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.718 + ], + "angle": 0, + "content": "Solving the camera-to-robot pose is a fundamental requirement for vision-based robot control, and is a process that takes considerable effort and cares to make accurate. Traditional approaches require modification of the robot via markers, and subsequent deep learning approaches enabled markerless feature extraction. Mainstream deep learning methods only use synthetic data and rely on Domain Randomization to fill the sim-to-real gap, because acquiring the 3D annotation is labor-intensive. In this work, we go beyond the limitation of 3D annotations for real-world data. We propose an end-to-end pose estimation framework that is capable of online camera-to-robot calibration and a self-supervised training method to scale the training to unlabeled real-world data. Our framework combines deep learning and geometric vision for solving the robot pose, and the pipeline is fully differentiable. To train the Camera-to-Robot Pose Estimation Network (CtRNet), we leverage foreground segmentation and differentiable rendering for image-level self-supervision. The pose prediction is visualized through a renderer and the image loss with the input image is back-propagated to train the neural network. Our experimental results on two public real datasets confirm the effectiveness of our approach over existing works. We also integrate our framework into a visual servoing system to demonstrate the promise of real-time precise robot pose estimation for automation tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.753, + 0.21, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The majority of modern robotic automation utilizes cameras for rich sensory information about the environment to infer tasks to be completed and provide feedback for closed-loop control. The leading paradigm for converting the valuable environment information to the robot's frame of reference for manipulation is position-based visual servoing (PBVS) [4]. At a high level, PBVS converts 3D environmental information inferred from the visual data (e.g." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.29, + 0.859, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.542, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Figure 1. Comparison of speed and accuracy (based on AUC metric) for existing image-based robot pose estimation methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.674 + ], + "angle": 0, + "content": "the pose of an object to be grasped) and transforms it to the robot coordinate frame where all the robot geometry is known (e.g. kinematics) using the camera-to-robot pose. Examples of robotic automation using the PBVS range from bin sorting [35] to tissue manipulation in surgery [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Calibrating camera-to-robot pose typically requires a significant amount of care and effort. Traditionally, the camera-to-robot pose is calibrated with externally attached fiducial markers (e.g. Aruco Marker [14], AprilTag [38]). The 2D location of the marker can be extracted from the image and the corresponding 3D location on the robot can be calculated with forward kinematics. Given a set 2D-3D correspondence, the camera-to-robot pose can be solved using Perspective-n-Point (PnP) methods [13, 30]. The procedure usually requires multiple runs with different robot configurations and once calibrated, the robot base and the camera are assumed static. The incapability of online calibration limits the potential applications for vision-based robot control in the real world, where minor bumps or simply shifting due to repetitive use will cause calibrations to be thrown off," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21296" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "not to mention real-world environmental factors like vibration, humidity, and temperature, are non-constant. Having flexibility on the camera and robot is more desirable so that the robot can interact with an unstructured environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.367 + ], + "angle": 0, + "content": "Deep learning, known as the current state-of-the-art approach for image feature extraction, brings promising ways for markerless camera-to-robot calibration. Current approaches to robot pose estimation are mainly classified into two categories: keypoint-based methods [27-29,34,43] and rendering-based methods [16,26]. Keypoint-based methods are the most popular approach for pose estimation because of the fast inference speed. However, the performance is limited to the accuracy of the keypoint detector which is often trained in simulation such that the proposed methods can generalize across different robotic designs. Therefore, the performance is ultimately hampered by the sim-to-real gap, which is a long-standing challenge in computer vision and robotics [55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.55 + ], + "angle": 0, + "content": "Rendering-based methods can achieve better performance by using the shape of the entire robot as observation, which provides dense correspondence for pose estimation. The approaches in this category usually employ an iterative refinement process and require a reasonable initialization for the optimization loop to converge [32]. Due to the nature that iteratively render and compare is time- and energy-consuming, rendering-based methods are more suitable for offline estimation where the robot and camera are held stationary. In more dynamic scenarios, such as a mobile robot, the slow computation time makes the rendering-based methods impracticable to use." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose CtRNet, an end-to-end framework for robot pose estimation which, at inference, uses keypoints for the fast inference speed and leverages the high performance of rendering-based methods for training to overcome the sim-to-real gap previous keypoint-based methods faced. Our framework contains a segmentation module to generate a binary mask of the robot and keypoint detection module which extracts point features for pose estimation. Since segmenting the robot from the background is a simpler task than estimating the robot pose and localizing point features on robot body parts, we leverage foreground segmentation to provide supervision for the pose estimation. Toward this direction, we first pretrained the network on synthetic data, which should have acquired essential knowledge about segmenting the robot. Then, a self-supervised training pipeline is proposed to transfer our model to the real world without manual labels. We connect the pose estimation to foreground segmentation with a differentiable renderer [24, 33]. The renderer generates a robot silhouette image of the estimated pose and directly compares it to the segmentation result. Since the entire framework is differentiable, the parameters of the neural network can be optimized by back-propagating the image loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Contributions. Our main contribution is the novel framework for image-based robot pose estimation together with a scalable self-training pipeline that utilizes unlimited real-world data to further improve the performance without any manual annotations. Since the keypoint detector is trained with image-level supervision, we effectively encompass the benefits from both keypoint-based and rendering-based methods, where previous methods were divided. As illustrated in the Fig. 1, our method maintains high inference speed while matching the performance of the rendering-based methods. Moreover, we integrate the CtR-Net into a robotic system for PBVS and demonstrate the effectiveness on real-time robot pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.303, + 0.65, + 0.318 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.328, + 0.8, + 0.344 + ], + "angle": 0, + "content": "2.1. Camera-to-Robot Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.457 + ], + "angle": 0, + "content": "The classical way to calibrate the camera-to-robot pose is to attach the fiducial markers [14, 38] to known locations along the robot kinematic chain. The marker is detected in the image frame and their 3D position in the robot base frame can be calculated with forward kinematics. With the geometrical constraints, the robot pose can be then derived by solving an optimization problem [11, 20, 22, 39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.609 + ], + "angle": 0, + "content": "Early works on markerless articulated pose tracking utilize a depth camera for 3D observation [10, 37, 41, 45]. For a high degree-of-freedom articulated robot, Bohg et al. [1] proposed a pose estimation method by first classifying the pixels in depth image to robot parts, and then a voting scheme is applied to estimate the robot pose relative to the camera. This method is further improved in [52] by directly training a Random Forest to regress joint angles instead of part label. However, these methods are not suitable for our scenario where only single RGB image is available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.837 + ], + "angle": 0, + "content": "More recently, as deep learning becomes popular in feature extraction, many works have been employing deep neural networks for robot pose estimation. Instead of using markers, a neural network is utilized for keypoint detection, and the robot pose is estimated through an optimizer (e.g. PnP solver) [28, 29, 34, 56]. To further improve the performance, the segmentation mask and edges are utilized to refine the robot pose [16, 27]. Labbe et al. [26] also introduces the render&compare method to estimate the robot pose by matching the robot shape. These methods mainly rely on synthetically generated data for training and hope the network can generalize to the real world by increasing the variance in data generation. Our method explicitly deals with the sim-to-real transfer by directly training on real-world data with self-supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.882, + 0.864 + ], + "angle": 0, + "content": "2.2. Domain Adaptation for Sim-to-Real Transfer" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In computer vision and robotics, Domain Randomization (DR) [48] is the most widely used method for sim" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21297" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "to-real transfer due to its simplicity and effectiveness. The idea is to randomize some simulation parameters (e.g. camera position, lighting, background, etc.) and hope that the randomization captures the distribution of the real-world data. This technique has been applied to object detection and grasping [2, 18, 21, 47, 49], and pose estimation [26, 28, 29, 34, 36, 46, 50, 56]. The randomization is usually tuned empirically hence it is not efficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.365 + ], + "angle": 0, + "content": "Another popular technique for domain transfer is Domain Adaptation (DA), which is to find the feature spaces that share a similar distribution between the source and target domains [51]. This technique has shown recent success in computer vision [7, 19, 44] and robotic applications [3, 15, 23]. In this work, instead of finding the latent space and modeling the distribution between the simulation and the real world, we perform sim-to-real transfer by directly training on the real-world data via a self-training pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.38, + 0.176, + 0.395 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.471, + 0.571 + ], + "angle": 0, + "content": "In this paper, we introduce an end-to-end framework for robot pose estimation and a scalable training pipeline to improve pose estimation accuracy on real-world data without the need for any manual annotation. We first explain the self-supervised training pipeline for sim-to-real transfer in Sec. 3.1 given a pretrained CtRNet on synthetic data which both segments the robot and estimates its pose from images. Then, we detail the camera-to-robot pose estimation network in Sec. 3.2 which utilizes a keypoint detector and a PnP solver to estimate the pose of the robot from image data in real-time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.583, + 0.47, + 0.613 + ], + "angle": 0, + "content": "3.1. Self-supervised Training for Sim-to-Real Transfer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.834 + ], + "angle": 0, + "content": "The most effective way to adapt the neural network to the real world is directly training the network on real sensor data. We propose a self-supervised training pipeline for sim-to-real transfer to facilitate the training without 3D annotations. To conduct the self-supervised training, we employ foreground segmentation to generate a mask of the robot, \\( f_{seg} \\), alongside the pose estimation, \\( f_{pose} \\). Given an input RGB image from the physical world, \\( \\mathbb{I} \\), and the robot joint angles, \\( \\mathbf{q} \\), \\( f_{pose} \\) estimates the robot pose which is then transformed to a silhouette image through a differentiable renderer. Our self-supervised objective is to optimize neural network parameters by minimizing the difference between the rendered silhouette image and the mask image. We formulate the optimization problem as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.85, + 0.47, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta_ {b b}, \\theta_ {k p}, \\theta_ {s e g} = \\underset {\\theta_ {b b}, \\theta_ {k p}, \\theta_ {s e g}} {\\arg \\min } \\mathcal {L} [ f _ {s e g} (\\mathbb {I} | \\theta_ {b b}, \\theta_ {s e g}), \\\\ \\left. \\mathcal {R} \\left(f _ {p o s e} \\left(\\mathbb {I} \\mid \\mathbf {q}, \\theta_ {b b}, \\theta_ {k p}\\right) \\mid \\mathbf {K}\\right) \\right] \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "where \\(\\theta_{bb},\\theta_{kp},\\theta_{seg}\\) denote the parameters of the backbone, keypoint, and segmentation layers of the neural network. \\(\\mathcal{R}\\) is the differentiable renderer with camera parameters \\(\\mathbf{K}\\), and \\(\\mathcal{L}(.)\\) is the objective loss function capturing the image difference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.304 + ], + "angle": 0, + "content": "We pretrained CtRNet's parameters which makes up \\( f_{seg} \\) and \\( f_{pose} \\), with synthetic data where the keypoint and segmentation labels are obtained freely (details in Supplementary Materials). During the self-training phase, where CtRNet learns with real data, the objective loss in (1) captures the difference between the segmentation result and the rendered image. The loss is iteratively back-propagated to, \\( \\Theta \\), where each iteration \\( f_{seg} \\) and \\( f_{pose} \\) take turns learning from each other to overcome the sim-to-real gap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Overview. The overview of the self-supervised training pipeline is shown in the Fig. 2. The segmentation module, \\( f_{seg} \\), simply takes in a robot image and outputs its mask. The pose estimation module, \\( f_{pose} \\), consists of a keypoint detector and a PnP solver to estimate the robot pose using the 2D-3D point correspondence, as shown in Fig. 3. Given the input robot image and joint angles, our camera-to-robot pose estimation network outputs a robot mask and the robot pose with respect to the camera frame. Mathematically, these functions are denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.467, + 0.892, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathbb {M} = f _ {s e g} (\\mathbb {I} | \\theta_ {b b}, \\theta_ {s e g}) \\quad \\mathbf {T} _ {b} ^ {c} = f _ {p o s e} (\\mathbb {I} | \\mathbf {q}, \\theta_ {b b}, \\theta_ {k p}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.558 + ], + "angle": 0, + "content": "where \\(\\mathbb{M}\\) is the robot mask and \\(\\mathbf{T}_b^c\\in SE(3)\\) is the 6-DOF robot pose. Finally, the self-supervised objective loss in (1) is realized through a differentiable renderer, \\(\\mathcal{R}\\), which generates a silhouette image of the robot given its pose, \\(\\mathbf{T}_b^c\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.723 + ], + "angle": 0, + "content": "Differentiable Rendering. To render the robot silhouette image, we utilize the PyTorch3D differentiable render [42]. We initialize a perspective camera with intrinsic parameters \\(\\mathbf{K}\\) and a silhouette renderer, which does not apply any lighting nor shading, is constructed with a rasterizer and a shader. The rasterizer applies the fast rasterization method [42] which selects the \\(k\\) nearest mesh triangles that effects each pixel and weights their influence according to the distance along the \\(z\\)-axis. Finally, the SoftSilhouetteShader is applied to compute pixel values of the rendered image using the sigmoid blending method [33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.892, + 0.815 + ], + "angle": 0, + "content": "We construct the ready-to-render robot mesh by connecting the CAD model for each robot body part using its forward kinematics and transforming them to the camera frame with the estimated robot pose \\(\\mathbf{T}_b^c\\) from \\(f_{pose}\\). Let \\(\\mathbf{v}^n\\in \\mathbb{R}^3\\) be a mesh vertex on the \\(n\\)-th robot link. Each vertex is transformed to the camera frame, hence ready-to-render, by" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.826, + 0.892, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {v}}} ^ {c} = \\mathbf {T} _ {b} ^ {c} \\mathbf {T} _ {n} ^ {b} (\\mathbf {q}) \\overline {{\\mathbf {v}}} ^ {n} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\bar{\\cdot}\\) represents the homogeneous representation of a point (e.g. \\(\\overline{\\mathbf{v}} = [\\mathbf{v},1]^T\\)), and \\(\\mathbf{T}_n^b (\\mathbf{q})\\) is the coordinate frame transformation obtained from the forward kinematics [8]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21298" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.095, + 0.844, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.316, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Figure 2. The overview of our proposed self-supervised training framework for sim-to-real transfer. The CtRNet contains a foreground segmentation module and a pose estimation module, which output a robot mask and a camera-to-robot pose respectively. The output pose is transformed into a silhouette image through a differentiable renderer. The image loss is back-propagated to train the keypoint detector and fine-tune the segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.475, + 0.522 + ], + "angle": 0, + "content": "Objective loss function. The objective loss in (1) is iteratively minimized where \\( f_{seg} \\) and \\( f_{pose} \\) take turns supervising each other on real data to overcome the sim-to-real gap faced by keypoint detection networks. To optimize \\( f_{pose} \\), the L2 image loss is used since the segmentation network's accuracy, within the context of estimating robot poses, has been shown to effectively transfer from simulation to the real world [26]. Mathematically the loss is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.528, + 0.47, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\left(\\mathbb {S} (i, j) - \\mathbb {M} (i, j)\\right) ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.61 + ], + "angle": 0, + "content": "where \\(H\\) and \\(W\\) is the height and width of the image, and \\(\\mathbb{S}\\) is the rendered silhouette image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.471, + 0.732 + ], + "angle": 0, + "content": "Although the pretrained robot segmentation, \\( f_{seg} \\), already performs well on real-world datasets, it is still desirable to refine it through self-supervised training to better extract fine details of corners and boundaries. To prevent the foreground segmentation layers from receiving noisy training signals, we apply the weighted Binary Cross Entropy Loss so that the high-quality rendering image can be used to further refine the foreground segmentation:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.743, + 0.47, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s e g} = - \\frac {w}{H * W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} [ \\mathbb {M} (i, j) \\log \\mathbb {S} (i, j) \\\\ \\left. + \\left(1 - \\mathbb {M} (i, j)\\right) \\log (1 - \\mathbb {S} (i, j)) \\right]. \\quad (5) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.816, + 0.471, + 0.877 + ], + "angle": 0, + "content": "where \\( w \\) is the weight for the given training sample. For PnP solvers, the optimal solution should minimize the point reprojection error. Therefore, we assign the weight for each training sample according to the reprojection error:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.885, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nw = \\exp (- s O (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})) \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.399, + 0.887, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.617, + 0.895, + 0.716 + ], + "angle": 0, + "content": "Figure 3. The diagram of the camera-to-robot pose estimation network (CtRNet) which describes the inference process of network. Given an input RGB image, the neural network generates a robot mask and a set of keypoint. Given the associated robot joint angles, a set of corresponding 3D keypoints are computed with forward kinematics. The camera-to-robot pose is estimated by a PnP solver with provided 2D-3D keypoint pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.894, + 0.835 + ], + "angle": 0, + "content": "where \\(s\\) is a scaling constant, \\(O\\) is the reprojection loss in the PnP solver (explained in Sec. 3.2), \\(\\{\\mathbf{o}_i|\\mathbf{o}_i\\in \\mathbb{R}^2\\}_{i = 1}^n\\) and \\(\\{\\mathbf{p}_i|\\mathbf{p}_i\\in \\mathbb{R}^3\\}_{i = 1}^n\\) are the 2D-3D keypoints inputted into the PnP solver. The exponential function is applied to the weight such that training samples with poor PnP convergence are weighted exponentially lower than good PnP convergence thereby stabilizing the training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.871, + 0.862 + ], + "angle": 0, + "content": "3.2. Camera-to-Robot Pose Estimation Network" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "The overview of the proposed Camera-to-Robot Pose Estimation Network, CtRNet, is shown in Fig. 3. Given an" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21299" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "input RGB image, we employ ResNet50 [17] as the backbone network to extract the latent features. The latent features are then passed through the Atrous Spatial Pyramid Pooling layers [6] to form the segmentation mask of input resolution. The keypoint detector, sharing the backbone network with the foreground segmentation, upsamples the feature maps through transposed convolutional layers and forms the heatmaps with \\( n \\) channels. Then, we apply the spatial softmax operator [12] on the heatmaps, which computes the expected 2D location of the points of maximal activation for each channel and results in a set of keypoints \\( [\\mathbf{o}_1,\\dots,\\mathbf{o}_n] \\) for all \\( n \\) channels. For simplicity, we define the set of keypoints at each joint location of the robot. Given the joint angles, the corresponding 3D keypoint location \\( \\mathbf{p}_i \\) can be calculated with robot forward kinematics:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.327, + 0.47, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {p}}} _ {i} = \\mathbf {T} _ {i} ^ {b} (\\mathbf {q}) \\overline {{\\mathbf {t}}}, f o r i = 1, \\dots , n \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.398 + ], + "angle": 0, + "content": "where \\(\\mathbf{t} = [0,0,0]\\). With the 2D and 3D corresponding keypoints, we can then apply a PnP solver [30] to estimate the robot pose with respect to the camera frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.505 + ], + "angle": 0, + "content": "Back-propagation for PnP Solver. A PnP solver is usually self-contained and not differentiable as the gradient with respect to the input cannot be derived explicitly. Inspired by [5], the implicit function theorem [25] is applied to obtain the gradient through implicit differentiation. Let the PnP solver be denoted as followed in the form of a non-linear function \\( g \\):" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.514, + 0.47, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {b} ^ {c *} = g (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.631 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_b^{c*}\\) is output pose from the PnP solver. In order to back-propagate through the PnP solver for training the keypoint detector, we are interested in finding the gradient of the output pose \\(\\mathbf{T}_b^{c*}\\) with respect to the input 2D points \\(\\mathbf{o}\\). Note that, the objective of the PnP solver is to minimize the reprojection error, such that:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.64, + 0.47, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {b} ^ {c *} = \\underset {\\mathbf {T} _ {b} ^ {c}} {\\arg \\min } O (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.677, + 0.113, + 0.689 + ], + "angle": 0, + "content": "with" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.697, + 0.47, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} O \\left(\\mathbf {T} _ {b} ^ {c}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}\\right) = \\sum_ {i = 1} ^ {n} | | \\mathbf {o} _ {i} - \\pi \\left(\\mathbf {p} _ {i} \\mid \\mathbf {T} _ {b} ^ {c}, \\mathbf {K}\\right) | | _ {2} ^ {2} (10) \\\\ = \\sum_ {i = 1} ^ {n} \\left\\| \\mathbf {r} _ {i} \\right\\| _ {2} ^ {2} (11) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.47, + 0.864 + ], + "angle": 0, + "content": "where \\(\\pi(.)\\) is the projection operator. Since the optimal solution \\(\\mathbf{T}_b^{c*}\\) is a local minimum for the objective function \\(O(\\mathbf{o},\\mathbf{p},\\mathbf{T}_b^c,\\mathbf{K})\\), a stationary constraint of the optimization process can be constructed by taking the first order derivative of the objective function with respect to \\(\\mathbf{T}_b^c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.871, + 0.469, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial O}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) | _ {\\mathbf {T} _ {b} ^ {c} = \\mathbf {T} _ {b} ^ {c *}} = \\mathbf {0}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Following [5], we construct a constrain function \\(F\\) to employ the implicit function theorem:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.132, + 0.892, + 0.166 + ], + "angle": 0, + "content": "\\[\nF (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) = \\frac {\\partial O}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c *}) = \\mathbf {0}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.176, + 0.892, + 0.206 + ], + "angle": 0, + "content": "Substituting the Eq. (10) and Eq. (11) to Eq. (13), we can derive the constraint function as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.215, + 0.892, + 0.296 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) = \\sum_ {i = 1} ^ {n} \\frac {\\partial \\left\\| \\mathbf {r} _ {i} \\right\\| _ {2} ^ {2}}{\\partial \\mathbf {T} _ {b} ^ {c}} (14) \\\\ = - 2 \\sum_ {i = 1} ^ {n} \\mathbf {r} _ {i} ^ {T} \\frac {\\partial \\pi}{\\partial \\mathbf {T} _ {b} ^ {c}} \\left(\\mathbf {p} _ {i} \\mid \\mathbf {T} _ {b} ^ {c *}, \\mathbf {K}\\right). (15) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Finally, we back-propagate through the PnP solver with the implicit differentiation. The gradient of the output pose with respect to the input 2D points is the Jacobian matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.38, + 0.891, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial g}{\\partial \\mathbf {o}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}) \\\\ = - \\left(\\frac {\\partial F}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})\\right) ^ {- 1} \\left(\\frac {\\partial F}{\\partial \\mathbf {o}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})\\right). \\tag {16} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.633, + 0.492 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.651 + ], + "angle": 0, + "content": "We first evaluate our method on two public real-world datasets for robot pose estimation and compare it against several state-of-the-art image-based robot pose estimation algorithms. We then conduct an ablation study on the pretraining procedure and explore how the number of pretraining samples could affect the performance of the self-supervised training. Finally, we integrate the camera-to-robot pose estimation framework into a visual servoing system to demonstrate the effectiveness of our method on real robot applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.661, + 0.787, + 0.675 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.807 + ], + "angle": 0, + "content": "DREAM-real Dataset. The DREAM-real dataset [29] is a real-world robot dataset collected with 3 different cameras: Azure Kinect (AK), XBOX 360 Kinect (XK), and RealSense (RS). This dataset contains around 50K RGB images of Franka Emika Panda arm and is recorded at \\((640 \\times 480)\\) resolution. The ground-truth camera-to-robot pose is provided for every image frame. The accuracy is evaluated with average distance (ADD) metric [54]," + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "\\[\nA D D = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| \\widetilde {\\mathbf {T}} _ {b} ^ {c} \\overline {{\\mathbf {p}}} _ {i} - \\mathbf {T} _ {b} ^ {c} \\overline {{\\mathbf {p}}} _ {i} \\right| \\right| _ {2} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathbf{T}}_b^c\\) indicates the ground-truth camera-to-robot pose. We also report the area-under-the-curve (AUC) value," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21300" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.887, + 0.193 + ], + "angle": 0, + "content": "
MethodCategoryBackbonePanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORBAll
AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓
DREAM-F [29]KeypointVGG1968.91211.41324.359491.91176.1302.07761.93095.31960.740113.029
DREAM-Q [29]KeypointVGG1952.38278.08937.47154.17877.9840.02757.08767.24856.98859.284
DREAM-H [29]KeypointResNet10160.5200.05664.0057.38278.8250.02469.05425.68568.58417.477
RoboPose [26]RenderingResNet3476.4970.02485.9260.01476.8630.02380.5040.01980.0940.020
CtRNetKeypointResNet5089.9280.01379.4650.03290.7890.01085.2890.02185.9620.020
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.202, + 0.893, + 0.232 + ], + "angle": 0, + "content": "Table 1. Comparison of our methods with the state-of-the-art methods on DREAM-real datasets using ADD metric. We report the mean and AUC of the ADD on each dataset and the overall accuracy." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.244, + 0.832, + 0.368 + ], + "angle": 0, + "content": "
MethodCategoryPCK (2D)Mean 2D Err. (pixel) ↓ADD (3D)Mean 3D Err. (mm) ↓
@50 pixel ↑AUC ↑@100 mm ↑AUC ↑
Aruco Marker [14]Keypoint0.4957.15286.980.3043.452447.34
DREAM-Q [29]Keypoint0.3344.011277.330.3240.63386.17
Opt. Keypoints [34]Keypoint0.6975.4649.510.4765.66141.05
Diff. RenderingRendering0.7478.6042.300.7881.1574.95
CtRNetKeypoint0.9993.9411.620.8883.9363.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.378, + 0.747, + 0.392 + ], + "angle": 0, + "content": "Table 2. Comparison of our methods with the state-of-the-art methods on Baxter dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.463 + ], + "angle": 0, + "content": "which integrates the percentage of ADD over different thresholds. A higher AUC value indicates more predictions with less error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.588 + ], + "angle": 0, + "content": "Baxter Dataset. The Baxter dataset [34] contains 100 RGB images of the left arm of Rethink Baxter collected with Azure Kinect camera at \\((2048\\times 1526)\\) resolution. The 2D and 3D ground-truth end-effector position with respect to the camera frame is provided. We evaluate the performance with the ADD metric for the end-effector. We also evaluate the end-effector reprojection error using the percentage of correct keypoints (PCK) metric [34]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.292, + 0.619 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The entire pipeline is implemented in PyTorch [40]. We initialize the backbone network with ImageNet [9] pretrained weights, and we train separate networks for different robots. The number of keypoints \\( n \\) is set to the number of robot links and the keypoints are defined at the robot joint locations. The neural network is pretrained on synthetic data for foreground segmentation and keypoint detection for 1000 epochs with 1e-5 learning rate. We reduce the learning rate by a factor of 10 once learning stagnates for 5 epochs. The Adam optimizer is applied to optimize the network parameters with the momentum set to 0.9. For self-supervised training on real-world data, we run the training for 500 epochs with 1e-6 learning rate. The same learning rate decay strategy and Adam optimizer is applied here similar to the pretraining. To make the training more stable, we clip the gradient of the network parameters at 10. The scaling factor in Eq. (6) is set to 0.1 for DREAM-real dataset and 0.01 for Baxter dataset, mainly accounting for" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.419, + 0.684, + 0.433 + ], + "angle": 0, + "content": "the difference in resolution." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.45, + 0.892, + 0.466 + ], + "angle": 0, + "content": "4.3. Robot Pose Estimation on Real-world Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Evaluation on DREAM-real Dataset. The proposed CtRNet is trained at \\((320\\times 240)\\) resolution and evaluated at the original resolution by scaling up the keypoints by a factor of 2. Some qualitative results for foreground segmentation and pose estimation are shown in the Fig. 4a. We compared our method with the state-of-the-art keypoint-based method DREAM [29] and the rendering-based method RoboPose [26]. The results for DREAM and RoboPose are compiled from the implementation provided by [26]. In Tab. 1, we report the AUC and mean ADD results on DREAM-real dataset with 3 different camera settings and the overall results combining all the test samples. Our method has a significantly better performance compared to the method in the same category and achieves comparable performance with the rendering-based method. We outperform DREAM on all settings and outperform RoboPose on the majority of the dataset. Overall on DREAM-real dataset, we achieve higher AUC (+17.378 compared to DREAM, +5.868 compared to RoboPose), and lower error compared to DREAM (-17.457)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Evaluation on Baxter Dataset. For the Baxter dataset, we trained the CtRNet at \\((640\\times 480)\\) resolution and evaluate at the original resolution, and Fig. 4b shows some of the qualitative results. We compared our method with several keypoint-based methods (Aruco Marker [14], DREAM [29], Optimized Keypoints [34]). We also implemented Differentiable Rendering for robot pose estimation, where the robot masks are generated with the pre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21301" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.237, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.091, + 0.36, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.091, + 0.482, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.091, + 0.609, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.091, + 0.732, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.091, + 0.857, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.172, + 0.237, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.172, + 0.36, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.172, + 0.482, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.172, + 0.609, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.172, + 0.732, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.172, + 0.857, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.253, + 0.237, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.253, + 0.36, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.253, + 0.482, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.253, + 0.609, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.253, + 0.732, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.253, + 0.857, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.337, + 0.306, + 0.348 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.337, + 0.679, + 0.348 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.361, + 0.894, + 0.404 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of CtRNet foreground segmentation and pose estimation on (a) DREAM-real dataset and (b) Baxter dataset. The first row shows the input RGB image, the second row shows the foreground segmentation, and the third row shows the projected robot skeleton based on the estimated robot pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.595 + ], + "angle": 0, + "content": "trained foreground segmentation. The 2D PCK results and 3D ADD results are reported in Tab. 2. Our method outperforms all other methods on both 2D and 3D evaluations. For 2D evaluation, we achieve 93.94 AUC for PCK with an average reprojection error of 11.62 pixels. For 3D evaluation, we achieve 83.93 AUC for ADD with an average ADD of \\(63.81\\mathrm{mm}\\). Notably, 99 percent of our estimation has less than 50 pixel reprojection error, which is less than 2 percent of the image resolution, and 88 percent of our estimation has less than \\(100\\mathrm{mm}\\) distance error when localizing the end-effector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.23, + 0.623 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We study how the number of pretraining samples affects the convergence and performance of the self-supervised training empirically on the Baxter dataset. We pretrain the neural network with different numbers of synthetic data samples \\( N_{\\text{pretrain}} = \\{500, 1000, 2000, 4000, 8000\\} \\), and examine the convergence of the self-supervised training process. Fig. 5 shows the plot of self-training loss \\( (\\mathcal{L}_{\\text{mask}} + \\mathcal{L}_{\\text{seg}}) \\) vs. the number of epochs for networks pretrianed with different number of synthetic data. We observe that doubling the size of pretraining dataset significantly improves the convergence of the self-training process at the beginning. However, the improvement gets smaller as the pretraining size increases. For the Baxter dataset, the improvement saturates after having more than 2000 pretraining samples. Continuing double the training size results in very marginal improvement. Noted that the Baxter dataset captures 20 different robot poses from a fixed camera position. The required number of pretraining samples might" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.43, + 0.877, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Figure 5. The training loss vs. number of epochs for the self-supervised training with different numbers of pretraining samples. More pretraining samples results in better convergence. The improvement saturates after having more than 2000 pretraining samples as only marginal improvement by adding more samples." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.7, + 0.851, + 0.714 + ], + "angle": 0, + "content": "vary according to the complexity of the environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.837 + ], + "angle": 0, + "content": "We further evaluate the resulting neural networks with the ground-truth labels on the Baxter dataset. We report the mean ADD and AUC ADD for the pose estimation in Tab. 3. The result verifies our observation on the convergence analysis. Having more pretraining samples improves the performance of pose estimation at the beginning, but the improvement stagnates after having more than 2000 pretraining samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.753, + 0.864 + ], + "angle": 0, + "content": "4.5. Visual Servoing Experiment" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We integrate the proposed CtRNet into a robotic system for position-based visual servoing (PBVS) with eye-to-hand" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21302" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.113, + 0.089, + 0.434, + 0.196 + ], + "angle": 0, + "content": "
NpretrainMean ADD (mm) ↓AUC ADD ↑
5002167.3047.62
100092.9176.65
200067.5182.98
400063.0084.12
800063.8183.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.087, + 0.207, + 0.457, + 0.221 + ], + "angle": 0, + "content": "Table 3. Ablation study for the number of pretraining samples." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.235, + 0.468, + 0.306 + ], + "angle": 0, + "content": "
MethodLoop RateTrans. Err. (m)Rot. Err. (rad)
DREAM [29]30Hz0.235 ± 0.3130.300 ± 0.544
Diff. Rendering1Hz0.046 ± 0.0620.036 ± 0.066
CtRNet30Hz0.002 ± 0.0010.002 ± 0.001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.316, + 0.468, + 0.344 + ], + "angle": 0, + "content": "Table 4. Mean and standard deviation of the translational error and rotational error for the visual servoing experiment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.468, + 0.552 + ], + "angle": 0, + "content": "configuration. We conduct the experiment on a Baxter robot and the details of the PBVS are described in the Supplementary Materials. The PBVS is purely based on RGB images from a single camera and the goal is to control the robot end-effector reaching a target pose defined in the camera frame. Specifically, we first set a target pose with respect to the camera frame. The target pose is then transformed into the robot base frame through the estimated camera-to-robot transformation. The robot controller calculates the desired robot configuration with inverse kinematics and a control law is applied to move the robot end-effector toward the target pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.468, + 0.824 + ], + "angle": 0, + "content": "For comparison, we also implemented DREAM [29] and a Differentiable Renderer for PBVS. For DREAM, the pretrained model for Baxter is applied. For Differentiable Renderer, we use the foreground segmentation of CtRNet to generate a robot mask. The optimizing loop for the renderer takes the last estimation as initialization and performs 10 updates at each callback to ensure convergence and maintain \\(1\\mathrm{Hz}\\) loop rate. In the experiment, we randomly set the target pose and the position of the camera, and the robotic system applies PBVS to reach the target pose from an arbitrary initialization, as shown in the Fig. 6. We ran the experiment for 10 trails with different robot pose estimation methods, and the translational (Euclidean distance) and rotational errors (Euler angles) of the end-effector are reported in Tab. 4. The experimental results show that our proposed method significantly improves the stability and accuracy of the PBVS, achieving \\(0.002\\mathrm{m}\\) averaged translational error and \\(0.002\\mathrm{rad}\\) rotational error on the end-effector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "We also plot the end-effector distance-to-goal over time for a selected trail in Fig. 7. In this selected trial, the system could not converge with DREAM because the poor robot pose estimation confuses the controller by giving the wrong target pose in the robot base frame, which is unreachable." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.696, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.091, + 0.879, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Figure 6. Snapshots of PBVS. The goal is to move the end-effector to the target pose (green). The figure on the right shows the robot configuration upon the convergence of PBVS." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.268, + 0.875, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.429, + 0.891, + 0.456 + ], + "angle": 0, + "content": "Figure 7. The plot of end-effector distance-to-goal over time on a selected PBVS trail." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.891, + 0.558 + ], + "angle": 0, + "content": "With the differentiable renderer, the servoing system takes more than 10 seconds to converge and oscillate due to the low loop rate. With our proposed CtRNet, the servoing system converges much faster (\\(\\leq 5\\) seconds), thanks to the fast and robust robot pose estimation. We show more qualitative results in the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.618, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We present the CtRNet, an end-to-end image-based robot pose estimation framework, and a self-supervised training pipeline that utilizes unlabelled real-world data for sim-to-real transfer. The CtRNet, using a keypoint detector for pose estimation while employing a rendering method for training, achieves state-of-the-art performance on robot pose estimation while maintaining high-speed inference. The Fig. 1 illustrates the advantages of CtRNet over existing methods, where the AUC values are normalized across two evaluation datasets by taking DREAM and CtRNet as references. We further experiment with different robot pose estimation methods by applying them to PBVS, which demonstrates CtRNet's fast and accurate robot pose estimation enabling stability when using single-frame robot pose estimation for feedback. Therefore, CtRNet supports real-time markerless camera-to-robot pose estimation which has been utilized for surgical robotic manipulation [43] and mobile robot manipulators [53]. For future work, we would like to extend our method to more robots and explore vision-based control in an unstructured environment." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21303" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Jeannette Bohg, Javier Romero, Alexander Herzog, and Stefan Schaal. Robot arm pose estimation through pixel-wise part classification. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 3143-3150. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.27 + ], + "angle": 0, + "content": "[2] Joao Borrego, Rui Figueiredo, Atabak Dehban, Plinio Moreno, Alexandre Bernardino, and José Santos-Victor. A generic visual perception domain randomisation framework for gazebo. In 2018 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC), pages 237-242. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.368 + ], + "angle": 0, + "content": "[3] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4243-4250. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.371, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[4] François Chaumette, Seth Hutchinson, and Peter Corke. Visual servoing. In Springer Handbook of Robotics, pages 841-866. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[5] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8100-8109, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.47, + 0.541 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[7] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3339-3348, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.469, + 0.656 + ], + "angle": 0, + "content": "[8] Jacques Denavit and Richard S Hartenberg. A kinematic notation for lower-pair mechanisms based on matrices. 1955. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[10] Karthik Desingh, Shiyang Lu, Anthony Opipari, and Odest Chadwicke Jenkins. Factored pose estimation of articulated objects using efficient nonparametric belief propagation. In 2019 International Conference on Robotics and Automation (ICRA), pages 7221-7227. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Irene Fassi and Giovanni Legnani. Hand to sensor calibration: A geometrical interpretation of the matrix equation \\( \\mathbf{ax} = \\) xb. Journal of Robotic Systems, 22(9):497-506, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Chelsea Finn, Xin Yu Tan, Yan Duan, Trevor Darrell, Sergey Levine, and Pieter Abbeel. Deep spatial autoencoders for visuomotor learning. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 512-519. IEEE, 2016. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[13] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930–943, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[14] Sergio Garrido-Jurado, Rafael Muñoz-Salinas, Francisco José Madrid-Cuevas, and Manuel Jesús Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition, 47(6):2280-2292, 2014. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[15] Abhishek Gupta, Coline Devin, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Learning invariant feature spaces to transfer skills with reinforcement learning. arXiv preprint arXiv:1703.02949, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[16] Ran Hao, Orhan Özgün, and M Cenk Çavusoglu. Vision-based surgical tool pose estimation for the da Vinci® robotic surgical system. In 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 1298-1305. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[18] Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, and Kurt Konolige. On pre-trained image features and synthetic images for deep learning. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[19] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[20] Radu Horaud and Fadi Dornaika. Hand-eye calibration. The international journal of robotics research, 14(3):195-210, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[21] Daniel Horváth, Gábor Erdős, Zoltán Istenes, Tomás Horváth, and Sándor Foldi. Object detection using sim2real domain randomization for robotic applications. IEEE Transactions on Robotics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[22] Jarmo Ilonen and Ville Kyrki. Robust robot-camera calibration. In 2011 15th International Conference on Advanced Robotics (ICAR), pages 67-74. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[23] Stephen James, Paul Wohlhart, Mrinal Kalakrishnan, Dmitry Kalashnikov, Alex Irpan, Julian Ibarz, Sergey Levine, Raia Hadsell, and Konstantinos Bousmalis. Sim-to-real via sim-to-sim: Data-efficient robotic grasping via randomized-to-canonical adaptation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12627-12637, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[24] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907-3916, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Steven George Krantz and Harold R Parks. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21304" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Single-view robot pose and joint angle estimation via render & compare. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1654-1663, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[27] Jens Lambrecht, Philipp Grosenick, and Marvin Meusel. Optimizing keypoint-based single-shot camera-to-robot pose estimation through shape segmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13843–13849. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[28] Jens Lambrecht and Linh Kästner. Towards the usage of synthetic data for marker-less pose estimation of articulated robots in rgb images. In 2019 19th International Conference on Advanced Robotics (ICAR), pages 240-247. IEEE, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[29] Timothy E Lee, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Oliver Kroemer, Dieter Fox, and Stan Birchfield. Camera-to-robot pose estimation from a single image. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9426–9432. IEEE, 2020. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.389, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[30] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision, 81(2):155-166, 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.47, + 0.515 + ], + "angle": 0, + "content": "[31] Yang Li, Florian Richter, Jingpei Lu, Emily K Funk, Ryan K Orosco, Jianke Zhu, and Michael C Yip. Super: A surgical perception framework for endoscopic tissue manipulation with surgical robotics. IEEE Robotics and Automation Letters, 5(2):2294-2301, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.47, + 0.573 + ], + "angle": 0, + "content": "[32] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[33] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[34] Jingpei Lu, Florian Richter, and Michael C Yip. Pose estimation for robot manipulators via keypoint optimization and sim-to-real transfer. IEEE Robotics and Automation Letters, 7(2):4622-4629, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[35] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 4(26):eaau4984, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[36] Fabian Manhardt, Wadim Kehl, Nassir Navab, and Federico Tombari. Deep model-based 6d pose refinement in rgb. In Proceedings of the European Conference on Computer Vision (ECCV), pages 800-815, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[37] Frank Michel, Alexander Krull, Eric Brachmann, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Pose estimation of kinematic chain instances via object coordinate regression. In BMVC, pages 181-1, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[38] Edwin Olson. Apriltag: A robust and flexible visual fiducial system. In 2011 IEEE international conference on robotics and automation, pages 3400-3407. IEEE, 2011. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[39] Frank C Park and Bryan J Martin. Robot sensor calibration: solving AX= XB on the Euclidean group. IEEE Transactions on Robotics and Automation, 10(5):717-721, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[40] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[41] Karl Pauwels, Leonardo Rubio, and Eduardo Ros. Real-time model-based articulated object pose detection and tracking with variable rigidity constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3994-4001, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[42] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[43] Florian Richter, Jingpei Lu, Ryan K Orosco, and Michael C Yip. Robotic tool tracking under partially visible kinematic chain: A unified approach. IEEE Transactions on Robotics, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[44] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3752-3761, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[45] Tanner Schmidt, Richard A Newcombe, and Dieter Fox. Dart: Dense articulated real-time tracking. In Robotics: Science and Systems, volume 2, pages 1-9. Berkeley, CA, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[46] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[47] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3482-3489. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[48] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[49] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969–977, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[50] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. arXiv preprint arXiv:1809.10790, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21305" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[51] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.122, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[52] Felix Widmaier, Daniel Kappler, Stefan Schaal, and Jeannette Bohg. Robot arm pose estimation by pixel-wise regression of joint angles. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 616-623. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[53] Melonee Wise, Michael Ferguson, Derek King, Eric Diehr, and David Dymesich. Fetch and freight: Standard platforms for service robot applications. In Workshop on autonomous mobile service robots, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[54] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.372 + ], + "angle": 0, + "content": "[55] Wenshuai Zhao, Jorge Pña Queralta, and Tomi Westerlund. Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In 2020 IEEE Symposium Series on Computational Intelligence (SSCI), pages 737-744. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[56] Yiming Zuo, Weichao Qiu, Lingxi Xie, Fangwei Zhong, Yizhou Wang, and Alan L Yuille. Craves: Controlling robotic arm with a vision-based economic system. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4214-4223, 2019. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21306" + } + ] +] \ No newline at end of file diff --git a/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_origin.pdf b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cf385db2a45d8d23c39d8e94f2f4b9ed06fbe4b6 --- /dev/null +++ b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/7c4fd6ff-03a8-40b6-8677-90772bb2a92c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69ca8830ad04147c8d5334b0c7b4aa49d2f581fc0adeef9a0f037cf96218a466 +size 3037185 diff --git a/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/full.md b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..339ebc90aaec5b477a752a0f4b0c7bacec86755d --- /dev/null +++ b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/full.md @@ -0,0 +1,349 @@ +# Markerless Camera-to-Robot Pose Estimation via Self-supervised Sim-to-Real Transfer + +Jingpei Lu, Florian Richter, and Michael C. Yip +University of California, San Diego +{jil360, frichter, yip}@ucsd.edu + +# Abstract + +Solving the camera-to-robot pose is a fundamental requirement for vision-based robot control, and is a process that takes considerable effort and cares to make accurate. Traditional approaches require modification of the robot via markers, and subsequent deep learning approaches enabled markerless feature extraction. Mainstream deep learning methods only use synthetic data and rely on Domain Randomization to fill the sim-to-real gap, because acquiring the 3D annotation is labor-intensive. In this work, we go beyond the limitation of 3D annotations for real-world data. We propose an end-to-end pose estimation framework that is capable of online camera-to-robot calibration and a self-supervised training method to scale the training to unlabeled real-world data. Our framework combines deep learning and geometric vision for solving the robot pose, and the pipeline is fully differentiable. To train the Camera-to-Robot Pose Estimation Network (CtRNet), we leverage foreground segmentation and differentiable rendering for image-level self-supervision. The pose prediction is visualized through a renderer and the image loss with the input image is back-propagated to train the neural network. Our experimental results on two public real datasets confirm the effectiveness of our approach over existing works. We also integrate our framework into a visual servoing system to demonstrate the promise of real-time precise robot pose estimation for automation tasks. + +# 1. Introduction + +The majority of modern robotic automation utilizes cameras for rich sensory information about the environment to infer tasks to be completed and provide feedback for closed-loop control. The leading paradigm for converting the valuable environment information to the robot's frame of reference for manipulation is position-based visual servoing (PBVS) [4]. At a high level, PBVS converts 3D environmental information inferred from the visual data (e.g. + +![](images/0b2f681605f396bb678513a9a136565681ad6b2b3c8d002f356fe9c3af69f236.jpg) +Figure 1. Comparison of speed and accuracy (based on AUC metric) for existing image-based robot pose estimation methods. + +the pose of an object to be grasped) and transforms it to the robot coordinate frame where all the robot geometry is known (e.g. kinematics) using the camera-to-robot pose. Examples of robotic automation using the PBVS range from bin sorting [35] to tissue manipulation in surgery [31]. + +Calibrating camera-to-robot pose typically requires a significant amount of care and effort. Traditionally, the camera-to-robot pose is calibrated with externally attached fiducial markers (e.g. Aruco Marker [14], AprilTag [38]). The 2D location of the marker can be extracted from the image and the corresponding 3D location on the robot can be calculated with forward kinematics. Given a set 2D-3D correspondence, the camera-to-robot pose can be solved using Perspective-n-Point (PnP) methods [13, 30]. The procedure usually requires multiple runs with different robot configurations and once calibrated, the robot base and the camera are assumed static. The incapability of online calibration limits the potential applications for vision-based robot control in the real world, where minor bumps or simply shifting due to repetitive use will cause calibrations to be thrown off, + +not to mention real-world environmental factors like vibration, humidity, and temperature, are non-constant. Having flexibility on the camera and robot is more desirable so that the robot can interact with an unstructured environment. + +Deep learning, known as the current state-of-the-art approach for image feature extraction, brings promising ways for markerless camera-to-robot calibration. Current approaches to robot pose estimation are mainly classified into two categories: keypoint-based methods [27-29,34,43] and rendering-based methods [16,26]. Keypoint-based methods are the most popular approach for pose estimation because of the fast inference speed. However, the performance is limited to the accuracy of the keypoint detector which is often trained in simulation such that the proposed methods can generalize across different robotic designs. Therefore, the performance is ultimately hampered by the sim-to-real gap, which is a long-standing challenge in computer vision and robotics [55]. + +Rendering-based methods can achieve better performance by using the shape of the entire robot as observation, which provides dense correspondence for pose estimation. The approaches in this category usually employ an iterative refinement process and require a reasonable initialization for the optimization loop to converge [32]. Due to the nature that iteratively render and compare is time- and energy-consuming, rendering-based methods are more suitable for offline estimation where the robot and camera are held stationary. In more dynamic scenarios, such as a mobile robot, the slow computation time makes the rendering-based methods impracticable to use. + +In this work, we propose CtRNet, an end-to-end framework for robot pose estimation which, at inference, uses keypoints for the fast inference speed and leverages the high performance of rendering-based methods for training to overcome the sim-to-real gap previous keypoint-based methods faced. Our framework contains a segmentation module to generate a binary mask of the robot and keypoint detection module which extracts point features for pose estimation. Since segmenting the robot from the background is a simpler task than estimating the robot pose and localizing point features on robot body parts, we leverage foreground segmentation to provide supervision for the pose estimation. Toward this direction, we first pretrained the network on synthetic data, which should have acquired essential knowledge about segmenting the robot. Then, a self-supervised training pipeline is proposed to transfer our model to the real world without manual labels. We connect the pose estimation to foreground segmentation with a differentiable renderer [24, 33]. The renderer generates a robot silhouette image of the estimated pose and directly compares it to the segmentation result. Since the entire framework is differentiable, the parameters of the neural network can be optimized by back-propagating the image loss. + +Contributions. Our main contribution is the novel framework for image-based robot pose estimation together with a scalable self-training pipeline that utilizes unlimited real-world data to further improve the performance without any manual annotations. Since the keypoint detector is trained with image-level supervision, we effectively encompass the benefits from both keypoint-based and rendering-based methods, where previous methods were divided. As illustrated in the Fig. 1, our method maintains high inference speed while matching the performance of the rendering-based methods. Moreover, we integrate the CtR-Net into a robotic system for PBVS and demonstrate the effectiveness on real-time robot pose estimation. + +# 2. Related Works + +# 2.1. Camera-to-Robot Pose Estimation + +The classical way to calibrate the camera-to-robot pose is to attach the fiducial markers [14, 38] to known locations along the robot kinematic chain. The marker is detected in the image frame and their 3D position in the robot base frame can be calculated with forward kinematics. With the geometrical constraints, the robot pose can be then derived by solving an optimization problem [11, 20, 22, 39]. + +Early works on markerless articulated pose tracking utilize a depth camera for 3D observation [10, 37, 41, 45]. For a high degree-of-freedom articulated robot, Bohg et al. [1] proposed a pose estimation method by first classifying the pixels in depth image to robot parts, and then a voting scheme is applied to estimate the robot pose relative to the camera. This method is further improved in [52] by directly training a Random Forest to regress joint angles instead of part label. However, these methods are not suitable for our scenario where only single RGB image is available. + +More recently, as deep learning becomes popular in feature extraction, many works have been employing deep neural networks for robot pose estimation. Instead of using markers, a neural network is utilized for keypoint detection, and the robot pose is estimated through an optimizer (e.g. PnP solver) [28, 29, 34, 56]. To further improve the performance, the segmentation mask and edges are utilized to refine the robot pose [16, 27]. Labbe et al. [26] also introduces the render&compare method to estimate the robot pose by matching the robot shape. These methods mainly rely on synthetically generated data for training and hope the network can generalize to the real world by increasing the variance in data generation. Our method explicitly deals with the sim-to-real transfer by directly training on real-world data with self-supervision. + +# 2.2. Domain Adaptation for Sim-to-Real Transfer + +In computer vision and robotics, Domain Randomization (DR) [48] is the most widely used method for sim + +to-real transfer due to its simplicity and effectiveness. The idea is to randomize some simulation parameters (e.g. camera position, lighting, background, etc.) and hope that the randomization captures the distribution of the real-world data. This technique has been applied to object detection and grasping [2, 18, 21, 47, 49], and pose estimation [26, 28, 29, 34, 36, 46, 50, 56]. The randomization is usually tuned empirically hence it is not efficient. + +Another popular technique for domain transfer is Domain Adaptation (DA), which is to find the feature spaces that share a similar distribution between the source and target domains [51]. This technique has shown recent success in computer vision [7, 19, 44] and robotic applications [3, 15, 23]. In this work, instead of finding the latent space and modeling the distribution between the simulation and the real world, we perform sim-to-real transfer by directly training on the real-world data via a self-training pipeline. + +# 3. Methods + +In this paper, we introduce an end-to-end framework for robot pose estimation and a scalable training pipeline to improve pose estimation accuracy on real-world data without the need for any manual annotation. We first explain the self-supervised training pipeline for sim-to-real transfer in Sec. 3.1 given a pretrained CtRNet on synthetic data which both segments the robot and estimates its pose from images. Then, we detail the camera-to-robot pose estimation network in Sec. 3.2 which utilizes a keypoint detector and a PnP solver to estimate the pose of the robot from image data in real-time. + +# 3.1. Self-supervised Training for Sim-to-Real Transfer + +The most effective way to adapt the neural network to the real world is directly training the network on real sensor data. We propose a self-supervised training pipeline for sim-to-real transfer to facilitate the training without 3D annotations. To conduct the self-supervised training, we employ foreground segmentation to generate a mask of the robot, $f_{seg}$ , alongside the pose estimation, $f_{pose}$ . Given an input RGB image from the physical world, $\mathbb{I}$ , and the robot joint angles, $\mathbf{q}$ , $f_{pose}$ estimates the robot pose which is then transformed to a silhouette image through a differentiable renderer. Our self-supervised objective is to optimize neural network parameters by minimizing the difference between the rendered silhouette image and the mask image. We formulate the optimization problem as: + +$$ +\begin{array}{l} \theta_ {b b}, \theta_ {k p}, \theta_ {s e g} = \underset {\theta_ {b b}, \theta_ {k p}, \theta_ {s e g}} {\arg \min } \mathcal {L} [ f _ {s e g} (\mathbb {I} | \theta_ {b b}, \theta_ {s e g}), \\ \left. \mathcal {R} \left(f _ {p o s e} \left(\mathbb {I} \mid \mathbf {q}, \theta_ {b b}, \theta_ {k p}\right) \mid \mathbf {K}\right) \right] \tag {1} \\ \end{array} +$$ + +where $\theta_{bb},\theta_{kp},\theta_{seg}$ denote the parameters of the backbone, keypoint, and segmentation layers of the neural network. $\mathcal{R}$ is the differentiable renderer with camera parameters $\mathbf{K}$ , and $\mathcal{L}(.)$ is the objective loss function capturing the image difference. + +We pretrained CtRNet's parameters which makes up $f_{seg}$ and $f_{pose}$ , with synthetic data where the keypoint and segmentation labels are obtained freely (details in Supplementary Materials). During the self-training phase, where CtRNet learns with real data, the objective loss in (1) captures the difference between the segmentation result and the rendered image. The loss is iteratively back-propagated to, $\Theta$ , where each iteration $f_{seg}$ and $f_{pose}$ take turns learning from each other to overcome the sim-to-real gap. + +Overview. The overview of the self-supervised training pipeline is shown in the Fig. 2. The segmentation module, $f_{seg}$ , simply takes in a robot image and outputs its mask. The pose estimation module, $f_{pose}$ , consists of a keypoint detector and a PnP solver to estimate the robot pose using the 2D-3D point correspondence, as shown in Fig. 3. Given the input robot image and joint angles, our camera-to-robot pose estimation network outputs a robot mask and the robot pose with respect to the camera frame. Mathematically, these functions are denoted as + +$$ +\mathbb {M} = f _ {s e g} (\mathbb {I} | \theta_ {b b}, \theta_ {s e g}) \quad \mathbf {T} _ {b} ^ {c} = f _ {p o s e} (\mathbb {I} | \mathbf {q}, \theta_ {b b}, \theta_ {k p}) \tag {2} +$$ + +where $\mathbb{M}$ is the robot mask and $\mathbf{T}_b^c\in SE(3)$ is the 6-DOF robot pose. Finally, the self-supervised objective loss in (1) is realized through a differentiable renderer, $\mathcal{R}$ , which generates a silhouette image of the robot given its pose, $\mathbf{T}_b^c$ . + +Differentiable Rendering. To render the robot silhouette image, we utilize the PyTorch3D differentiable render [42]. We initialize a perspective camera with intrinsic parameters $\mathbf{K}$ and a silhouette renderer, which does not apply any lighting nor shading, is constructed with a rasterizer and a shader. The rasterizer applies the fast rasterization method [42] which selects the $k$ nearest mesh triangles that effects each pixel and weights their influence according to the distance along the $z$ -axis. Finally, the SoftSilhouetteShader is applied to compute pixel values of the rendered image using the sigmoid blending method [33]. + +We construct the ready-to-render robot mesh by connecting the CAD model for each robot body part using its forward kinematics and transforming them to the camera frame with the estimated robot pose $\mathbf{T}_b^c$ from $f_{pose}$ . Let $\mathbf{v}^n\in \mathbb{R}^3$ be a mesh vertex on the $n$ -th robot link. Each vertex is transformed to the camera frame, hence ready-to-render, by + +$$ +\overline {{\mathbf {v}}} ^ {c} = \mathbf {T} _ {b} ^ {c} \mathbf {T} _ {n} ^ {b} (\mathbf {q}) \overline {{\mathbf {v}}} ^ {n} \tag {3} +$$ + +where $\bar{\cdot}$ represents the homogeneous representation of a point (e.g. $\overline{\mathbf{v}} = [\mathbf{v},1]^T$ ), and $\mathbf{T}_n^b (\mathbf{q})$ is the coordinate frame transformation obtained from the forward kinematics [8]. + +![](images/487c85ebb839ce93b2f3bf055f973ac50b3fd222f099b3e1a66038647ba7ea86.jpg) +Figure 2. The overview of our proposed self-supervised training framework for sim-to-real transfer. The CtRNet contains a foreground segmentation module and a pose estimation module, which output a robot mask and a camera-to-robot pose respectively. The output pose is transformed into a silhouette image through a differentiable renderer. The image loss is back-propagated to train the keypoint detector and fine-tune the segmentation. + +Objective loss function. The objective loss in (1) is iteratively minimized where $f_{seg}$ and $f_{pose}$ take turns supervising each other on real data to overcome the sim-to-real gap faced by keypoint detection networks. To optimize $f_{pose}$ , the L2 image loss is used since the segmentation network's accuracy, within the context of estimating robot poses, has been shown to effectively transfer from simulation to the real world [26]. Mathematically the loss is expressed as + +$$ +\mathcal {L} _ {\text {m a s k}} = \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \left(\mathbb {S} (i, j) - \mathbb {M} (i, j)\right) ^ {2} \tag {4} +$$ + +where $H$ and $W$ is the height and width of the image, and $\mathbb{S}$ is the rendered silhouette image. + +Although the pretrained robot segmentation, $f_{seg}$ , already performs well on real-world datasets, it is still desirable to refine it through self-supervised training to better extract fine details of corners and boundaries. To prevent the foreground segmentation layers from receiving noisy training signals, we apply the weighted Binary Cross Entropy Loss so that the high-quality rendering image can be used to further refine the foreground segmentation: + +$$ +\begin{array}{l} \mathcal {L} _ {s e g} = - \frac {w}{H * W} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} [ \mathbb {M} (i, j) \log \mathbb {S} (i, j) \\ \left. + \left(1 - \mathbb {M} (i, j)\right) \log (1 - \mathbb {S} (i, j)) \right]. \quad (5) \\ \end{array} +$$ + +where $w$ is the weight for the given training sample. For PnP solvers, the optimal solution should minimize the point reprojection error. Therefore, we assign the weight for each training sample according to the reprojection error: + +$$ +w = \exp (- s O (\mathbf {o}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c})) \tag {6} +$$ + +![](images/2f4a3552f2a4e6f01ea8e98791810acaad9a4813330d2171b403ac9c906eec8d.jpg) +Figure 3. The diagram of the camera-to-robot pose estimation network (CtRNet) which describes the inference process of network. Given an input RGB image, the neural network generates a robot mask and a set of keypoint. Given the associated robot joint angles, a set of corresponding 3D keypoints are computed with forward kinematics. The camera-to-robot pose is estimated by a PnP solver with provided 2D-3D keypoint pairs. + +where $s$ is a scaling constant, $O$ is the reprojection loss in the PnP solver (explained in Sec. 3.2), $\{\mathbf{o}_i|\mathbf{o}_i\in \mathbb{R}^2\}_{i = 1}^n$ and $\{\mathbf{p}_i|\mathbf{p}_i\in \mathbb{R}^3\}_{i = 1}^n$ are the 2D-3D keypoints inputted into the PnP solver. The exponential function is applied to the weight such that training samples with poor PnP convergence are weighted exponentially lower than good PnP convergence thereby stabilizing the training. + +# 3.2. Camera-to-Robot Pose Estimation Network + +The overview of the proposed Camera-to-Robot Pose Estimation Network, CtRNet, is shown in Fig. 3. Given an + +input RGB image, we employ ResNet50 [17] as the backbone network to extract the latent features. The latent features are then passed through the Atrous Spatial Pyramid Pooling layers [6] to form the segmentation mask of input resolution. The keypoint detector, sharing the backbone network with the foreground segmentation, upsamples the feature maps through transposed convolutional layers and forms the heatmaps with $n$ channels. Then, we apply the spatial softmax operator [12] on the heatmaps, which computes the expected 2D location of the points of maximal activation for each channel and results in a set of keypoints $[\mathbf{o}_1,\dots,\mathbf{o}_n]$ for all $n$ channels. For simplicity, we define the set of keypoints at each joint location of the robot. Given the joint angles, the corresponding 3D keypoint location $\mathbf{p}_i$ can be calculated with robot forward kinematics: + +$$ +\overline {{\mathbf {p}}} _ {i} = \mathbf {T} _ {i} ^ {b} (\mathbf {q}) \overline {{\mathbf {t}}}, f o r i = 1, \dots , n \tag {7} +$$ + +where $\mathbf{t} = [0,0,0]$ . With the 2D and 3D corresponding keypoints, we can then apply a PnP solver [30] to estimate the robot pose with respect to the camera frame. + +Back-propagation for PnP Solver. A PnP solver is usually self-contained and not differentiable as the gradient with respect to the input cannot be derived explicitly. Inspired by [5], the implicit function theorem [25] is applied to obtain the gradient through implicit differentiation. Let the PnP solver be denoted as followed in the form of a non-linear function $g$ : + +$$ +\mathbf {T} _ {b} ^ {c *} = g (\mathbf {o}, \mathbf {p}, \mathbf {K}) \tag {8} +$$ + +where $\mathbf{T}_b^{c*}$ is output pose from the PnP solver. In order to back-propagate through the PnP solver for training the keypoint detector, we are interested in finding the gradient of the output pose $\mathbf{T}_b^{c*}$ with respect to the input 2D points $\mathbf{o}$ . Note that, the objective of the PnP solver is to minimize the reprojection error, such that: + +$$ +\mathbf {T} _ {b} ^ {c *} = \underset {\mathbf {T} _ {b} ^ {c}} {\arg \min } O (\mathbf {o}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c}) \tag {9} +$$ + +with + +$$ +\begin{array}{l} O \left(\mathbf {T} _ {b} ^ {c}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c}\right) = \sum_ {i = 1} ^ {n} | | \mathbf {o} _ {i} - \pi \left(\mathbf {p} _ {i} \mid \mathbf {T} _ {b} ^ {c}, \mathbf {K}\right) | | _ {2} ^ {2} (10) \\ = \sum_ {i = 1} ^ {n} \left\| \mathbf {r} _ {i} \right\| _ {2} ^ {2} (11) \\ \end{array} +$$ + +where $\pi(.)$ is the projection operator. Since the optimal solution $\mathbf{T}_b^{c*}$ is a local minimum for the objective function $O(\mathbf{o},\mathbf{p},\mathbf{T}_b^c,\mathbf{K})$ , a stationary constraint of the optimization process can be constructed by taking the first order derivative of the objective function with respect to $\mathbf{T}_b^c$ : + +$$ +\frac {\partial O}{\partial \mathbf {T} _ {b} ^ {c}} (\mathbf {o}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c}) | _ {\mathbf {T} _ {b} ^ {c} = \mathbf {T} _ {b} ^ {c *}} = \mathbf {0}. \tag {12} +$$ + +Following [5], we construct a constrain function $F$ to employ the implicit function theorem: + +$$ +F (\mathbf {o}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c}) = \frac {\partial O}{\partial \mathbf {T} _ {b} ^ {c}} (\mathbf {o}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c *}) = \mathbf {0}. \tag {13} +$$ + +Substituting the Eq. (10) and Eq. (11) to Eq. (13), we can derive the constraint function as: + +$$ +\begin{array}{l} F (\mathbf {o}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c}) = \sum_ {i = 1} ^ {n} \frac {\partial \left\| \mathbf {r} _ {i} \right\| _ {2} ^ {2}}{\partial \mathbf {T} _ {b} ^ {c}} (14) \\ = - 2 \sum_ {i = 1} ^ {n} \mathbf {r} _ {i} ^ {T} \frac {\partial \pi}{\partial \mathbf {T} _ {b} ^ {c}} \left(\mathbf {p} _ {i} \mid \mathbf {T} _ {b} ^ {c *}, \mathbf {K}\right). (15) \\ \end{array} +$$ + +Finally, we back-propagate through the PnP solver with the implicit differentiation. The gradient of the output pose with respect to the input 2D points is the Jacobian matrix: + +$$ +\begin{array}{l} \frac {\partial g}{\partial \mathbf {o}} (\mathbf {o}, \mathbf {p}, \mathbf {K}) \\ = - \left(\frac {\partial F}{\partial \mathbf {T} _ {b} ^ {c}} (\mathbf {o}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c})\right) ^ {- 1} \left(\frac {\partial F}{\partial \mathbf {o}} (\mathbf {o}, \mathbf {p}, \mathbf {K}, \mathbf {T} _ {b} ^ {c})\right). \tag {16} \\ \end{array} +$$ + +# 4. Experiments + +We first evaluate our method on two public real-world datasets for robot pose estimation and compare it against several state-of-the-art image-based robot pose estimation algorithms. We then conduct an ablation study on the pretraining procedure and explore how the number of pretraining samples could affect the performance of the self-supervised training. Finally, we integrate the camera-to-robot pose estimation framework into a visual servoing system to demonstrate the effectiveness of our method on real robot applications. + +# 4.1. Datasets and Evaluation Metrics + +DREAM-real Dataset. The DREAM-real dataset [29] is a real-world robot dataset collected with 3 different cameras: Azure Kinect (AK), XBOX 360 Kinect (XK), and RealSense (RS). This dataset contains around 50K RGB images of Franka Emika Panda arm and is recorded at $(640 \times 480)$ resolution. The ground-truth camera-to-robot pose is provided for every image frame. The accuracy is evaluated with average distance (ADD) metric [54], + +$$ +A D D = \frac {1}{n} \sum_ {i = 1} ^ {n} \left| \left| \widetilde {\mathbf {T}} _ {b} ^ {c} \overline {{\mathbf {p}}} _ {i} - \mathbf {T} _ {b} ^ {c} \overline {{\mathbf {p}}} _ {i} \right| \right| _ {2} \tag {17} +$$ + +where $\widetilde{\mathbf{T}}_b^c$ indicates the ground-truth camera-to-robot pose. We also report the area-under-the-curve (AUC) value, + +
MethodCategoryBackbonePanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORBAll
AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓
DREAM-F [29]KeypointVGG1968.91211.41324.359491.91176.1302.07761.93095.31960.740113.029
DREAM-Q [29]KeypointVGG1952.38278.08937.47154.17877.9840.02757.08767.24856.98859.284
DREAM-H [29]KeypointResNet10160.5200.05664.0057.38278.8250.02469.05425.68568.58417.477
RoboPose [26]RenderingResNet3476.4970.02485.9260.01476.8630.02380.5040.01980.0940.020
CtRNetKeypointResNet5089.9280.01379.4650.03290.7890.01085.2890.02185.9620.020
+ +Table 1. Comparison of our methods with the state-of-the-art methods on DREAM-real datasets using ADD metric. We report the mean and AUC of the ADD on each dataset and the overall accuracy. + +
MethodCategoryPCK (2D)Mean 2D Err. (pixel) ↓ADD (3D)Mean 3D Err. (mm) ↓
@50 pixel ↑AUC ↑@100 mm ↑AUC ↑
Aruco Marker [14]Keypoint0.4957.15286.980.3043.452447.34
DREAM-Q [29]Keypoint0.3344.011277.330.3240.63386.17
Opt. Keypoints [34]Keypoint0.6975.4649.510.4765.66141.05
Diff. RenderingRendering0.7478.6042.300.7881.1574.95
CtRNetKeypoint0.9993.9411.620.8883.9363.81
+ +Table 2. Comparison of our methods with the state-of-the-art methods on Baxter dataset. + +which integrates the percentage of ADD over different thresholds. A higher AUC value indicates more predictions with less error. + +Baxter Dataset. The Baxter dataset [34] contains 100 RGB images of the left arm of Rethink Baxter collected with Azure Kinect camera at $(2048\times 1526)$ resolution. The 2D and 3D ground-truth end-effector position with respect to the camera frame is provided. We evaluate the performance with the ADD metric for the end-effector. We also evaluate the end-effector reprojection error using the percentage of correct keypoints (PCK) metric [34]. + +# 4.2. Implementation details + +The entire pipeline is implemented in PyTorch [40]. We initialize the backbone network with ImageNet [9] pretrained weights, and we train separate networks for different robots. The number of keypoints $n$ is set to the number of robot links and the keypoints are defined at the robot joint locations. The neural network is pretrained on synthetic data for foreground segmentation and keypoint detection for 1000 epochs with 1e-5 learning rate. We reduce the learning rate by a factor of 10 once learning stagnates for 5 epochs. The Adam optimizer is applied to optimize the network parameters with the momentum set to 0.9. For self-supervised training on real-world data, we run the training for 500 epochs with 1e-6 learning rate. The same learning rate decay strategy and Adam optimizer is applied here similar to the pretraining. To make the training more stable, we clip the gradient of the network parameters at 10. The scaling factor in Eq. (6) is set to 0.1 for DREAM-real dataset and 0.01 for Baxter dataset, mainly accounting for + +the difference in resolution. + +# 4.3. Robot Pose Estimation on Real-world Datasets + +Evaluation on DREAM-real Dataset. The proposed CtRNet is trained at $(320\times 240)$ resolution and evaluated at the original resolution by scaling up the keypoints by a factor of 2. Some qualitative results for foreground segmentation and pose estimation are shown in the Fig. 4a. We compared our method with the state-of-the-art keypoint-based method DREAM [29] and the rendering-based method RoboPose [26]. The results for DREAM and RoboPose are compiled from the implementation provided by [26]. In Tab. 1, we report the AUC and mean ADD results on DREAM-real dataset with 3 different camera settings and the overall results combining all the test samples. Our method has a significantly better performance compared to the method in the same category and achieves comparable performance with the rendering-based method. We outperform DREAM on all settings and outperform RoboPose on the majority of the dataset. Overall on DREAM-real dataset, we achieve higher AUC (+17.378 compared to DREAM, +5.868 compared to RoboPose), and lower error compared to DREAM (-17.457). + +Evaluation on Baxter Dataset. For the Baxter dataset, we trained the CtRNet at $(640\times 480)$ resolution and evaluate at the original resolution, and Fig. 4b shows some of the qualitative results. We compared our method with several keypoint-based methods (Aruco Marker [14], DREAM [29], Optimized Keypoints [34]). We also implemented Differentiable Rendering for robot pose estimation, where the robot masks are generated with the pre + +![](images/d761695e91347c02df25ad2c1c064d94da4919065ca4dcacf2b0ca7dd026283a.jpg) + +![](images/317d59fc022df35c7f281319d92388d65c112eb62a59ac36babe4e7adb56cbd4.jpg) + +![](images/07a26f618bac91534b11f760792f59a3c019dd37cf42482f2f7ef295667f7eab.jpg) + +![](images/886bc7de660b5a96b67dbfc72fe635f18e06ea295c3d57ce79bd126e06b34510.jpg) + +![](images/225fcf48a322d09a908a616bbc4beb8d42c0e09ae5b30e8366259a012c9f7734.jpg) + +![](images/52a07cabea14f9c7ccc54fe6c837b32a0ba69d642db33f5d63d5af54e5c2f3d6.jpg) + +![](images/09b344d883b9e1271789ee2b0a240ec5e01647c4d94f1f4ddd4787ff3b937cc3.jpg) + +![](images/631e13a4bf8f969d0d2bf1fbd52082cded94d76e2e152f04e263059fb85b8dea.jpg) + +![](images/e40a327b4bf64a02913c387ff1d7b3602d0abf79d9fb6b07179c6ed3cc0ac289.jpg) + +![](images/15fe4104716543d754256f385e26d8417d8066bc9938e42fdffaadbcb7197019.jpg) + +![](images/6f54349c7c60ab6c481e78c669aa967c090b5c2eeed7f94b21cd956f796d189e.jpg) + +![](images/840fbef587f090e443af1a6ba3959736cb7453e461febda1ccd4d62a18f2d19a.jpg) + +![](images/bdf50a92cf4823f9db92e0648d9cb853bd85e179b382c1c9331c18275b7cba5e.jpg) +Figure 4. Qualitative results of CtRNet foreground segmentation and pose estimation on (a) DREAM-real dataset and (b) Baxter dataset. The first row shows the input RGB image, the second row shows the foreground segmentation, and the third row shows the projected robot skeleton based on the estimated robot pose. + +![](images/19e24995ff490a764033db121bf6bf0917118d458173f83809f3cde709516e73.jpg) +(a) + +![](images/757ec12934e15a3373e017d8039a3d4dbecea80472284074ad7393f12e9c55ce.jpg) + +![](images/be4661f3725e6a40a3e0a8517df899cd173db4081d811590d4012f51d6a89aee.jpg) + +![](images/f4deb021922b3c713d04345603c9af0146adce6cb2fc75a875387565562aa5ee.jpg) +(b) + +![](images/7d2a320cf8745b6d38f188ddfa566554e19015860136306581df6041b9f95130.jpg) + +trained foreground segmentation. The 2D PCK results and 3D ADD results are reported in Tab. 2. Our method outperforms all other methods on both 2D and 3D evaluations. For 2D evaluation, we achieve 93.94 AUC for PCK with an average reprojection error of 11.62 pixels. For 3D evaluation, we achieve 83.93 AUC for ADD with an average ADD of $63.81\mathrm{mm}$ . Notably, 99 percent of our estimation has less than 50 pixel reprojection error, which is less than 2 percent of the image resolution, and 88 percent of our estimation has less than $100\mathrm{mm}$ distance error when localizing the end-effector. + +# 4.4. Ablation Study + +We study how the number of pretraining samples affects the convergence and performance of the self-supervised training empirically on the Baxter dataset. We pretrain the neural network with different numbers of synthetic data samples $N_{\text{pretrain}} = \{500, 1000, 2000, 4000, 8000\}$ , and examine the convergence of the self-supervised training process. Fig. 5 shows the plot of self-training loss $(\mathcal{L}_{\text{mask}} + \mathcal{L}_{\text{seg}})$ vs. the number of epochs for networks pretrianed with different number of synthetic data. We observe that doubling the size of pretraining dataset significantly improves the convergence of the self-training process at the beginning. However, the improvement gets smaller as the pretraining size increases. For the Baxter dataset, the improvement saturates after having more than 2000 pretraining samples. Continuing double the training size results in very marginal improvement. Noted that the Baxter dataset captures 20 different robot poses from a fixed camera position. The required number of pretraining samples might + +![](images/b1cdde4800a9067745e8bfe7b3d7d8b96b8e86483212ef2d94491268d300ebb1.jpg) +Figure 5. The training loss vs. number of epochs for the self-supervised training with different numbers of pretraining samples. More pretraining samples results in better convergence. The improvement saturates after having more than 2000 pretraining samples as only marginal improvement by adding more samples. + +vary according to the complexity of the environment. + +We further evaluate the resulting neural networks with the ground-truth labels on the Baxter dataset. We report the mean ADD and AUC ADD for the pose estimation in Tab. 3. The result verifies our observation on the convergence analysis. Having more pretraining samples improves the performance of pose estimation at the beginning, but the improvement stagnates after having more than 2000 pretraining samples. + +# 4.5. Visual Servoing Experiment + +We integrate the proposed CtRNet into a robotic system for position-based visual servoing (PBVS) with eye-to-hand + +
NpretrainMean ADD (mm) ↓AUC ADD ↑
5002167.3047.62
100092.9176.65
200067.5182.98
400063.0084.12
800063.8183.93
+ +Table 3. Ablation study for the number of pretraining samples. + +
MethodLoop RateTrans. Err. (m)Rot. Err. (rad)
DREAM [29]30Hz0.235 ± 0.3130.300 ± 0.544
Diff. Rendering1Hz0.046 ± 0.0620.036 ± 0.066
CtRNet30Hz0.002 ± 0.0010.002 ± 0.001
+ +Table 4. Mean and standard deviation of the translational error and rotational error for the visual servoing experiment. + +configuration. We conduct the experiment on a Baxter robot and the details of the PBVS are described in the Supplementary Materials. The PBVS is purely based on RGB images from a single camera and the goal is to control the robot end-effector reaching a target pose defined in the camera frame. Specifically, we first set a target pose with respect to the camera frame. The target pose is then transformed into the robot base frame through the estimated camera-to-robot transformation. The robot controller calculates the desired robot configuration with inverse kinematics and a control law is applied to move the robot end-effector toward the target pose. + +For comparison, we also implemented DREAM [29] and a Differentiable Renderer for PBVS. For DREAM, the pretrained model for Baxter is applied. For Differentiable Renderer, we use the foreground segmentation of CtRNet to generate a robot mask. The optimizing loop for the renderer takes the last estimation as initialization and performs 10 updates at each callback to ensure convergence and maintain $1\mathrm{Hz}$ loop rate. In the experiment, we randomly set the target pose and the position of the camera, and the robotic system applies PBVS to reach the target pose from an arbitrary initialization, as shown in the Fig. 6. We ran the experiment for 10 trails with different robot pose estimation methods, and the translational (Euclidean distance) and rotational errors (Euler angles) of the end-effector are reported in Tab. 4. The experimental results show that our proposed method significantly improves the stability and accuracy of the PBVS, achieving $0.002\mathrm{m}$ averaged translational error and $0.002\mathrm{rad}$ rotational error on the end-effector. + +We also plot the end-effector distance-to-goal over time for a selected trail in Fig. 7. In this selected trial, the system could not converge with DREAM because the poor robot pose estimation confuses the controller by giving the wrong target pose in the robot base frame, which is unreachable. + +![](images/ba92e43337ba0f497e943dd52229464de4eaca582f08f4da2069cfd9100316e8.jpg) +Figure 6. Snapshots of PBVS. The goal is to move the end-effector to the target pose (green). The figure on the right shows the robot configuration upon the convergence of PBVS. + +![](images/cbf61d722171064c1276569aa84bbcc2864d13de42fcd0708f54f01e18413f43.jpg) + +![](images/c84a6549ed07db415ab31a8f79c3667aaa14a8ab97efd1363a20cc8a1a25b89e.jpg) +Figure 7. The plot of end-effector distance-to-goal over time on a selected PBVS trail. + +With the differentiable renderer, the servoing system takes more than 10 seconds to converge and oscillate due to the low loop rate. With our proposed CtRNet, the servoing system converges much faster ( $\leq 5$ seconds), thanks to the fast and robust robot pose estimation. We show more qualitative results in the Supplementary Materials. + +# 5. Conclusion + +We present the CtRNet, an end-to-end image-based robot pose estimation framework, and a self-supervised training pipeline that utilizes unlabelled real-world data for sim-to-real transfer. The CtRNet, using a keypoint detector for pose estimation while employing a rendering method for training, achieves state-of-the-art performance on robot pose estimation while maintaining high-speed inference. The Fig. 1 illustrates the advantages of CtRNet over existing methods, where the AUC values are normalized across two evaluation datasets by taking DREAM and CtRNet as references. We further experiment with different robot pose estimation methods by applying them to PBVS, which demonstrates CtRNet's fast and accurate robot pose estimation enabling stability when using single-frame robot pose estimation for feedback. Therefore, CtRNet supports real-time markerless camera-to-robot pose estimation which has been utilized for surgical robotic manipulation [43] and mobile robot manipulators [53]. For future work, we would like to extend our method to more robots and explore vision-based control in an unstructured environment. + +# References + +[1] Jeannette Bohg, Javier Romero, Alexander Herzog, and Stefan Schaal. Robot arm pose estimation through pixel-wise part classification. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 3143-3150. IEEE, 2014. 2 +[2] Joao Borrego, Rui Figueiredo, Atabak Dehban, Plinio Moreno, Alexandre Bernardino, and José Santos-Victor. A generic visual perception domain randomisation framework for gazebo. In 2018 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC), pages 237-242. IEEE, 2018. 3 +[3] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4243-4250. IEEE, 2018. 3 +[4] François Chaumette, Seth Hutchinson, and Peter Corke. Visual servoing. In Springer Handbook of Robotics, pages 841-866. Springer, 2016. 1 +[5] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8100-8109, 2020. 5 +[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 5 +[7] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3339-3348, 2018. 3 +[8] Jacques Denavit and Richard S Hartenberg. A kinematic notation for lower-pair mechanisms based on matrices. 1955. 3 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[10] Karthik Desingh, Shiyang Lu, Anthony Opipari, and Odest Chadwicke Jenkins. Factored pose estimation of articulated objects using efficient nonparametric belief propagation. In 2019 International Conference on Robotics and Automation (ICRA), pages 7221-7227. IEEE, 2019. 2 +[11] Irene Fassi and Giovanni Legnani. Hand to sensor calibration: A geometrical interpretation of the matrix equation $\mathbf{ax} =$ xb. Journal of Robotic Systems, 22(9):497-506, 2005. 2 +[12] Chelsea Finn, Xin Yu Tan, Yan Duan, Trevor Darrell, Sergey Levine, and Pieter Abbeel. Deep spatial autoencoders for visuomotor learning. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 512-519. IEEE, 2016. 5 + +[13] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930–943, 2003. 1 +[14] Sergio Garrido-Jurado, Rafael Muñoz-Salinas, Francisco José Madrid-Cuevas, and Manuel Jesús Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition, 47(6):2280-2292, 2014. 1, 2, 6 +[15] Abhishek Gupta, Coline Devin, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Learning invariant feature spaces to transfer skills with reinforcement learning. arXiv preprint arXiv:1703.02949, 2017. 3 +[16] Ran Hao, Orhan Özgün, and M Cenk Çavusoglu. Vision-based surgical tool pose estimation for the da Vinci® robotic surgical system. In 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 1298-1305. IEEE, 2018. 2 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[18] Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, and Kurt Konolige. On pre-trained image features and synthetic images for deep learning. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3 +[19] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 3 +[20] Radu Horaud and Fadi Dornaika. Hand-eye calibration. The international journal of robotics research, 14(3):195-210, 1995. 2 +[21] Daniel Horváth, Gábor Erdős, Zoltán Istenes, Tomás Horváth, and Sándor Foldi. Object detection using sim2real domain randomization for robotic applications. IEEE Transactions on Robotics, 2022. 3 +[22] Jarmo Ilonen and Ville Kyrki. Robust robot-camera calibration. In 2011 15th International Conference on Advanced Robotics (ICAR), pages 67-74. IEEE, 2011. 2 +[23] Stephen James, Paul Wohlhart, Mrinal Kalakrishnan, Dmitry Kalashnikov, Alex Irpan, Julian Ibarz, Sergey Levine, Raia Hadsell, and Konstantinos Bousmalis. Sim-to-real via sim-to-sim: Data-efficient robotic grasping via randomized-to-canonical adaptation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12627-12637, 2019. 3 +[24] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907-3916, 2018. 2 +[25] Steven George Krantz and Harold R Parks. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002. 5 + +[26] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Single-view robot pose and joint angle estimation via render & compare. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1654-1663, 2021. 2, 3, 4, 6 +[27] Jens Lambrecht, Philipp Grosenick, and Marvin Meusel. Optimizing keypoint-based single-shot camera-to-robot pose estimation through shape segmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13843–13849. IEEE, 2021. 2 +[28] Jens Lambrecht and Linh Kästner. Towards the usage of synthetic data for marker-less pose estimation of articulated robots in rgb images. In 2019 19th International Conference on Advanced Robotics (ICAR), pages 240-247. IEEE, 2019. 2, 3 +[29] Timothy E Lee, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Oliver Kroemer, Dieter Fox, and Stan Birchfield. Camera-to-robot pose estimation from a single image. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9426–9432. IEEE, 2020. 2, 3, 5, 6, 8 +[30] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision, 81(2):155-166, 2009. 1, 5 +[31] Yang Li, Florian Richter, Jingpei Lu, Emily K Funk, Ryan K Orosco, Jianke Zhu, and Michael C Yip. Super: A surgical perception framework for endoscopic tissue manipulation with surgical robotics. IEEE Robotics and Automation Letters, 5(2):2294-2301, 2020. 1 +[32] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 2 +[33] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 2, 3 +[34] Jingpei Lu, Florian Richter, and Michael C Yip. Pose estimation for robot manipulators via keypoint optimization and sim-to-real transfer. IEEE Robotics and Automation Letters, 7(2):4622-4629, 2022. 2, 3, 6 +[35] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 4(26):eaau4984, 2019. 1 +[36] Fabian Manhardt, Wadim Kehl, Nassir Navab, and Federico Tombari. Deep model-based 6d pose refinement in rgb. In Proceedings of the European Conference on Computer Vision (ECCV), pages 800-815, 2018. 3 +[37] Frank Michel, Alexander Krull, Eric Brachmann, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Pose estimation of kinematic chain instances via object coordinate regression. In BMVC, pages 181-1, 2015. 2 +[38] Edwin Olson. Apriltag: A robust and flexible visual fiducial system. In 2011 IEEE international conference on robotics and automation, pages 3400-3407. IEEE, 2011. 1, 2 + +[39] Frank C Park and Bryan J Martin. Robot sensor calibration: solving AX= XB on the Euclidean group. IEEE Transactions on Robotics and Automation, 10(5):717-721, 1994. 2 +[40] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6 +[41] Karl Pauwels, Leonardo Rubio, and Eduardo Ros. Real-time model-based articulated object pose detection and tracking with variable rigidity constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3994-4001, 2014. 2 +[42] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 3 +[43] Florian Richter, Jingpei Lu, Ryan K Orosco, and Michael C Yip. Robotic tool tracking under partially visible kinematic chain: A unified approach. IEEE Transactions on Robotics, 2021. 2, 8 +[44] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3752-3761, 2018. 3 +[45] Tanner Schmidt, Richard A Newcombe, and Dieter Fox. Dart: Dense articulated real-time tracking. In Robotics: Science and Systems, volume 2, pages 1-9. Berkeley, CA, 2014. 2 +[46] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 3 +[47] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3482-3489. IEEE, 2018. 3 +[48] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 2 +[49] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969–977, 2018. 3 +[50] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. arXiv preprint arXiv:1809.10790, 2018. 3 + +[51] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3 +[52] Felix Widmaier, Daniel Kappler, Stefan Schaal, and Jeannette Bohg. Robot arm pose estimation by pixel-wise regression of joint angles. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 616-623. IEEE, 2016. 2 +[53] Melonee Wise, Michael Ferguson, Derek King, Eric Diehr, and David Dymesich. Fetch and freight: Standard platforms for service robot applications. In Workshop on autonomous mobile service robots, 2016. 8 +[54] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 5 +[55] Wenshuai Zhao, Jorge Pña Queralta, and Tomi Westerlund. Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In 2020 IEEE Symposium Series on Computational Intelligence (SSCI), pages 737-744. IEEE, 2020. 2 +[56] Yiming Zuo, Weichao Qiu, Lingxi Xie, Fangwei Zhong, Yizhou Wang, and Alan L Yuille. Craves: Controlling robotic arm with a vision-based economic system. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4214-4223, 2019. 2, 3 \ No newline at end of file diff --git a/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/images.zip b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..28515a6dc61aa226baab3c511c45323d4761d0b8 --- /dev/null +++ b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1dd21797cef8ebbc07310e2b23414fb240b26012c361111ad34e715dd8cec958 +size 538664 diff --git a/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/layout.json b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ef6aad72bcddc039f63cf8253dad36ec033e929 --- /dev/null +++ b/2023/Markerless Camera-to-Robot Pose Estimation via Self-Supervised Sim-to-Real Transfer/layout.json @@ -0,0 +1,8650 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 538, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 538, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 538, + 138 + ], + "type": "text", + "content": "Markerless Camera-to-Robot Pose Estimation via Self-supervised Sim-to-Real Transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 180, + 161, + 412, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 161, + 412, + 203 + ], + "spans": [ + { + "bbox": [ + 180, + 161, + 412, + 203 + ], + "type": "text", + "content": "Jingpei Lu, Florian Richter, and Michael C. Yip \nUniversity of California, San Diego \n{jil360, frichter, yip}@ucsd.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 568 + ], + "type": "text", + "content": "Solving the camera-to-robot pose is a fundamental requirement for vision-based robot control, and is a process that takes considerable effort and cares to make accurate. Traditional approaches require modification of the robot via markers, and subsequent deep learning approaches enabled markerless feature extraction. Mainstream deep learning methods only use synthetic data and rely on Domain Randomization to fill the sim-to-real gap, because acquiring the 3D annotation is labor-intensive. In this work, we go beyond the limitation of 3D annotations for real-world data. We propose an end-to-end pose estimation framework that is capable of online camera-to-robot calibration and a self-supervised training method to scale the training to unlabeled real-world data. Our framework combines deep learning and geometric vision for solving the robot pose, and the pipeline is fully differentiable. To train the Camera-to-Robot Pose Estimation Network (CtRNet), we leverage foreground segmentation and differentiable rendering for image-level self-supervision. The pose prediction is visualized through a renderer and the image loss with the input image is back-propagated to train the neural network. Our experimental results on two public real datasets confirm the effectiveness of our approach over existing works. We also integrate our framework into a visual servoing system to demonstrate the promise of real-time precise robot pose estimation for automation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "The majority of modern robotic automation utilizes cameras for rich sensory information about the environment to infer tasks to be completed and provide feedback for closed-loop control. The leading paradigm for converting the valuable environment information to the robot's frame of reference for manipulation is position-based visual servoing (PBVS) [4]. At a high level, PBVS converts 3D environmental information inferred from the visual data (e.g." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 229, + 525, + 422 + ], + "blocks": [ + { + "bbox": [ + 328, + 229, + 525, + 422 + ], + "lines": [ + { + "bbox": [ + 328, + 229, + 525, + 422 + ], + "spans": [ + { + "bbox": [ + 328, + 229, + 525, + 422 + ], + "type": "image", + "image_path": "0b2f681605f396bb678513a9a136565681ad6b2b3c8d002f356fe9c3af69f236.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 429, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 305, + 429, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 545, + 451 + ], + "type": "text", + "content": "Figure 1. Comparison of speed and accuracy (based on AUC metric) for existing image-based robot pose estimation methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "text", + "content": "the pose of an object to be grasped) and transforms it to the robot coordinate frame where all the robot geometry is known (e.g. kinematics) using the camera-to-robot pose. Examples of robotic automation using the PBVS range from bin sorting [35] to tissue manipulation in surgery [31]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Calibrating camera-to-robot pose typically requires a significant amount of care and effort. Traditionally, the camera-to-robot pose is calibrated with externally attached fiducial markers (e.g. Aruco Marker [14], AprilTag [38]). The 2D location of the marker can be extracted from the image and the corresponding 3D location on the robot can be calculated with forward kinematics. Given a set 2D-3D correspondence, the camera-to-robot pose can be solved using Perspective-n-Point (PnP) methods [13, 30]. The procedure usually requires multiple runs with different robot configurations and once calibrated, the robot base and the camera are assumed static. The incapability of online calibration limits the potential applications for vision-based robot control in the real world, where minor bumps or simply shifting due to repetitive use will cause calibrations to be thrown off," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21296" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "not to mention real-world environmental factors like vibration, humidity, and temperature, are non-constant. Having flexibility on the camera and robot is more desirable so that the robot can interact with an unstructured environment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "type": "text", + "content": "Deep learning, known as the current state-of-the-art approach for image feature extraction, brings promising ways for markerless camera-to-robot calibration. Current approaches to robot pose estimation are mainly classified into two categories: keypoint-based methods [27-29,34,43] and rendering-based methods [16,26]. Keypoint-based methods are the most popular approach for pose estimation because of the fast inference speed. However, the performance is limited to the accuracy of the keypoint detector which is often trained in simulation such that the proposed methods can generalize across different robotic designs. Therefore, the performance is ultimately hampered by the sim-to-real gap, which is a long-standing challenge in computer vision and robotics [55]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 292, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 435 + ], + "type": "text", + "content": "Rendering-based methods can achieve better performance by using the shape of the entire robot as observation, which provides dense correspondence for pose estimation. The approaches in this category usually employ an iterative refinement process and require a reasonable initialization for the optimization loop to converge [32]. Due to the nature that iteratively render and compare is time- and energy-consuming, rendering-based methods are more suitable for offline estimation where the robot and camera are held stationary. In more dynamic scenarios, such as a mobile robot, the slow computation time makes the rendering-based methods impracticable to use." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 438, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 715 + ], + "type": "text", + "content": "In this work, we propose CtRNet, an end-to-end framework for robot pose estimation which, at inference, uses keypoints for the fast inference speed and leverages the high performance of rendering-based methods for training to overcome the sim-to-real gap previous keypoint-based methods faced. Our framework contains a segmentation module to generate a binary mask of the robot and keypoint detection module which extracts point features for pose estimation. Since segmenting the robot from the background is a simpler task than estimating the robot pose and localizing point features on robot body parts, we leverage foreground segmentation to provide supervision for the pose estimation. Toward this direction, we first pretrained the network on synthetic data, which should have acquired essential knowledge about segmenting the robot. Then, a self-supervised training pipeline is proposed to transfer our model to the real world without manual labels. We connect the pose estimation to foreground segmentation with a differentiable renderer [24, 33]. The renderer generates a robot silhouette image of the estimated pose and directly compares it to the segmentation result. Since the entire framework is differentiable, the parameters of the neural network can be optimized by back-propagating the image loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "Contributions. Our main contribution is the novel framework for image-based robot pose estimation together with a scalable self-training pipeline that utilizes unlimited real-world data to further improve the performance without any manual annotations. Since the keypoint detector is trained with image-level supervision, we effectively encompass the benefits from both keypoint-based and rendering-based methods, where previous methods were divided. As illustrated in the Fig. 1, our method maintains high inference speed while matching the performance of the rendering-based methods. Moreover, we integrate the CtR-Net into a robotic system for PBVS and demonstrate the effectiveness on real-time robot pose estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 259, + 489, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 489, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 489, + 272 + ], + "type": "text", + "content": "2.1. Camera-to-Robot Pose Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "content": "The classical way to calibrate the camera-to-robot pose is to attach the fiducial markers [14, 38] to known locations along the robot kinematic chain. The marker is detected in the image frame and their 3D position in the robot base frame can be calculated with forward kinematics. With the geometrical constraints, the robot pose can be then derived by solving an optimization problem [11, 20, 22, 39]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 362, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 482 + ], + "type": "text", + "content": "Early works on markerless articulated pose tracking utilize a depth camera for 3D observation [10, 37, 41, 45]. For a high degree-of-freedom articulated robot, Bohg et al. [1] proposed a pose estimation method by first classifying the pixels in depth image to robot parts, and then a voting scheme is applied to estimate the robot pose relative to the camera. This method is further improved in [52] by directly training a Random Forest to regress joint angles instead of part label. However, these methods are not suitable for our scenario where only single RGB image is available." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 483, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 662 + ], + "type": "text", + "content": "More recently, as deep learning becomes popular in feature extraction, many works have been employing deep neural networks for robot pose estimation. Instead of using markers, a neural network is utilized for keypoint detection, and the robot pose is estimated through an optimizer (e.g. PnP solver) [28, 29, 34, 56]. To further improve the performance, the segmentation mask and edges are utilized to refine the robot pose [16, 27]. Labbe et al. [26] also introduces the render&compare method to estimate the robot pose by matching the robot shape. These methods mainly rely on synthetically generated data for training and hope the network can generalize to the real world by increasing the variance in data generation. Our method explicitly deals with the sim-to-real transfer by directly training on real-world data with self-supervision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 670, + 539, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 539, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 539, + 684 + ], + "type": "text", + "content": "2.2. Domain Adaptation for Sim-to-Real Transfer" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In computer vision and robotics, Domain Randomization (DR) [48] is the most widely used method for sim" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21297" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "to-real transfer due to its simplicity and effectiveness. The idea is to randomize some simulation parameters (e.g. camera position, lighting, background, etc.) and hope that the randomization captures the distribution of the real-world data. This technique has been applied to object detection and grasping [2, 18, 21, 47, 49], and pose estimation [26, 28, 29, 34, 36, 46, 50, 56]. The randomization is usually tuned empirically hence it is not efficient." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 289 + ], + "type": "text", + "content": "Another popular technique for domain transfer is Domain Adaptation (DA), which is to find the feature spaces that share a similar distribution between the source and target domains [51]. This technique has shown recent success in computer vision [7, 19, 44] and robotic applications [3, 15, 23]. In this work, instead of finding the latent space and modeling the distribution between the simulation and the real world, we perform sim-to-real transfer by directly training on the real-world data via a self-training pipeline." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 300, + 107, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 107, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 107, + 312 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 321, + 288, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 452 + ], + "type": "text", + "content": "In this paper, we introduce an end-to-end framework for robot pose estimation and a scalable training pipeline to improve pose estimation accuracy on real-world data without the need for any manual annotation. We first explain the self-supervised training pipeline for sim-to-real transfer in Sec. 3.1 given a pretrained CtRNet on synthetic data which both segments the robot and estimates its pose from images. Then, we detail the camera-to-robot pose estimation network in Sec. 3.2 which utilizes a keypoint detector and a PnP solver to estimate the pose of the robot from image data in real-time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 461, + 287, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 461, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 287, + 485 + ], + "type": "text", + "content": "3.1. Self-supervised Training for Sim-to-Real Transfer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "text", + "content": "The most effective way to adapt the neural network to the real world is directly training the network on real sensor data. We propose a self-supervised training pipeline for sim-to-real transfer to facilitate the training without 3D annotations. To conduct the self-supervised training, we employ foreground segmentation to generate a mask of the robot, " + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "text", + "content": ", alongside the pose estimation, " + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "inline_equation", + "content": "f_{pose}" + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "text", + "content": ". Given an input RGB image from the physical world, " + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "text", + "content": ", and the robot joint angles, " + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "inline_equation", + "content": "f_{pose}" + }, + { + "bbox": [ + 46, + 492, + 287, + 660 + ], + "type": "text", + "content": " estimates the robot pose which is then transformed to a silhouette image through a differentiable renderer. Our self-supervised objective is to optimize neural network parameters by minimizing the difference between the rendered silhouette image and the mask image. We formulate the optimization problem as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 673, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 673, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 673, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta_ {b b}, \\theta_ {k p}, \\theta_ {s e g} = \\underset {\\theta_ {b b}, \\theta_ {k p}, \\theta_ {s e g}} {\\arg \\min } \\mathcal {L} [ f _ {s e g} (\\mathbb {I} | \\theta_ {b b}, \\theta_ {s e g}), \\\\ \\left. \\mathcal {R} \\left(f _ {p o s e} \\left(\\mathbb {I} \\mid \\mathbf {q}, \\theta_ {b b}, \\theta_ {k p}\\right) \\mid \\mathbf {K}\\right) \\right] \\tag {1} \\\\ \\end{array}", + "image_path": "717f877c1d746501617efcfc402ab8c52821e9b324dfb88fa3a516af507424a7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\theta_{bb},\\theta_{kp},\\theta_{seg}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " denote the parameters of the backbone, keypoint, and segmentation layers of the neural network. " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " is the differentiable renderer with camera parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(.)" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " is the objective loss function capturing the image difference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "content": "We pretrained CtRNet's parameters which makes up " + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "inline_equation", + "content": "f_{pose}" + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "content": ", with synthetic data where the keypoint and segmentation labels are obtained freely (details in Supplementary Materials). During the self-training phase, where CtRNet learns with real data, the objective loss in (1) captures the difference between the segmentation result and the rendered image. The loss is iteratively back-propagated to, " + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "content": ", where each iteration " + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "inline_equation", + "content": "f_{pose}" + }, + { + "bbox": [ + 304, + 133, + 545, + 240 + ], + "type": "text", + "content": " take turns learning from each other to overcome the sim-to-real gap." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "content": "Overview. The overview of the self-supervised training pipeline is shown in the Fig. 2. The segmentation module, " + }, + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "content": ", simply takes in a robot image and outputs its mask. The pose estimation module, " + }, + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "inline_equation", + "content": "f_{pose}" + }, + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "content": ", consists of a keypoint detector and a PnP solver to estimate the robot pose using the 2D-3D point correspondence, as shown in Fig. 3. Given the input robot image and joint angles, our camera-to-robot pose estimation network outputs a robot mask and the robot pose with respect to the camera frame. Mathematically, these functions are denoted as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 369, + 545, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 545, + 384 + ], + "type": "interline_equation", + "content": "\\mathbb {M} = f _ {s e g} (\\mathbb {I} | \\theta_ {b b}, \\theta_ {s e g}) \\quad \\mathbf {T} _ {b} ^ {c} = f _ {p o s e} (\\mathbb {I} | \\mathbf {q}, \\theta_ {b b}, \\theta_ {k p}) \\tag {2}", + "image_path": "5807daeeeb08cd697c8b788fbe50cf9bc4472304219206d2ff294031f6d681d9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "text", + "content": " is the robot mask and " + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_b^c\\in SE(3)" + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "text", + "content": " is the 6-DOF robot pose. Finally, the self-supervised objective loss in (1) is realized through a differentiable renderer, " + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "text", + "content": ", which generates a silhouette image of the robot given its pose, " + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_b^c" + }, + { + "bbox": [ + 304, + 392, + 545, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "type": "text", + "content": "Differentiable Rendering. To render the robot silhouette image, we utilize the PyTorch3D differentiable render [42]. We initialize a perspective camera with intrinsic parameters " + }, + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "type": "text", + "content": " and a silhouette renderer, which does not apply any lighting nor shading, is constructed with a rasterizer and a shader. The rasterizer applies the fast rasterization method [42] which selects the " + }, + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "type": "text", + "content": " nearest mesh triangles that effects each pixel and weights their influence according to the distance along the " + }, + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 441, + 545, + 572 + ], + "type": "text", + "content": "-axis. Finally, the SoftSilhouetteShader is applied to compute pixel values of the rendered image using the sigmoid blending method [33]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "text", + "content": "We construct the ready-to-render robot mesh by connecting the CAD model for each robot body part using its forward kinematics and transforming them to the camera frame with the estimated robot pose " + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_b^c" + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "inline_equation", + "content": "f_{pose}" + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^n\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "text", + "content": " be a mesh vertex on the " + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 573, + 545, + 645 + ], + "type": "text", + "content": "-th robot link. Each vertex is transformed to the camera frame, hence ready-to-render, by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 654, + 545, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 654, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 386, + 654, + 545, + 668 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {v}}} ^ {c} = \\mathbf {T} _ {b} ^ {c} \\mathbf {T} _ {n} ^ {b} (\\mathbf {q}) \\overline {{\\mathbf {v}}} ^ {n} \\tag {3}", + "image_path": "af88cb222fdede06425190c607cf593deaab7a99b8bb34a1adb06a5435accb6f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\cdot}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " represents the homogeneous representation of a point (e.g. " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{v}} = [\\mathbf{v},1]^T" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_n^b (\\mathbf{q})" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the coordinate frame transformation obtained from the forward kinematics [8]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21298" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 75, + 516, + 238 + ], + "blocks": [ + { + "bbox": [ + 80, + 75, + 516, + 238 + ], + "lines": [ + { + "bbox": [ + 80, + 75, + 516, + 238 + ], + "spans": [ + { + "bbox": [ + 80, + 75, + 516, + 238 + ], + "type": "image", + "image_path": "487c85ebb839ce93b2f3bf055f973ac50b3fd222f099b3e1a66038647ba7ea86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 250, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 45, + 250, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 45, + 250, + 547, + 297 + ], + "type": "text", + "content": "Figure 2. The overview of our proposed self-supervised training framework for sim-to-real transfer. The CtRNet contains a foreground segmentation module and a pose estimation module, which output a robot mask and a camera-to-robot pose respectively. The output pose is transformed into a silhouette image through a differentiable renderer. The image loss is back-propagated to train the keypoint detector and fine-tune the segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "type": "text", + "content": "Objective loss function. The objective loss in (1) is iteratively minimized where " + }, + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "type": "inline_equation", + "content": "f_{pose}" + }, + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "type": "text", + "content": " take turns supervising each other on real data to overcome the sim-to-real gap faced by keypoint detection networks. To optimize " + }, + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "type": "inline_equation", + "content": "f_{pose}" + }, + { + "bbox": [ + 46, + 316, + 290, + 413 + ], + "type": "text", + "content": ", the L2 image loss is used since the segmentation network's accuracy, within the context of estimating robot poses, has been shown to effectively transfer from simulation to the real world [26]. Mathematically the loss is expressed as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 418, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 418, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 89, + 418, + 287, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\left(\\mathbb {S} (i, j) - \\mathbb {M} (i, j)\\right) ^ {2} \\tag {4}", + "image_path": "44144a3e6f7571ca1f3813dea0a9dea25b2afcd77e9c2250b0412dc3f3142187.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "type": "text", + "content": " is the height and width of the image, and " + }, + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 46, + 459, + 287, + 483 + ], + "type": "text", + "content": " is the rendered silhouette image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "content": "Although the pretrained robot segmentation, " + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "content": ", already performs well on real-world datasets, it is still desirable to refine it through self-supervised training to better extract fine details of corners and boundaries. To prevent the foreground segmentation layers from receiving noisy training signals, we apply the weighted Binary Cross Entropy Loss so that the high-quality rendering image can be used to further refine the foreground segmentation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 588, + 287, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 287, + 639 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {s e g} = - \\frac {w}{H * W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} [ \\mathbb {M} (i, j) \\log \\mathbb {S} (i, j) \\\\ \\left. + \\left(1 - \\mathbb {M} (i, j)\\right) \\log (1 - \\mathbb {S} (i, j)) \\right]. \\quad (5) \\\\ \\end{array}", + "image_path": "1836c895dc3008df20efb250d57ddbe894ece3d5c3f0a7e06f51d09dc8dc3dd3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 646, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 646, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 288, + 694 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 646, + 288, + 694 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 646, + 288, + 694 + ], + "type": "text", + "content": " is the weight for the given training sample. For PnP solvers, the optimal solution should minimize the point reprojection error. Therefore, we assign the weight for each training sample according to the reprojection error:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "w = \\exp (- s O (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})) \\tag {6}", + "image_path": "926489ad6d7f327d3af18d3d6107715b41e672deba53f59e7a52f88152d799b4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 316, + 542, + 477 + ], + "blocks": [ + { + "bbox": [ + 310, + 316, + 542, + 477 + ], + "lines": [ + { + "bbox": [ + 310, + 316, + 542, + 477 + ], + "spans": [ + { + "bbox": [ + 310, + 316, + 542, + 477 + ], + "type": "image", + "image_path": "2f4a3552f2a4e6f01ea8e98791810acaad9a4813330d2171b403ac9c906eec8d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 488, + 547, + 567 + ], + "lines": [ + { + "bbox": [ + 304, + 488, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 547, + 567 + ], + "type": "text", + "content": "Figure 3. The diagram of the camera-to-robot pose estimation network (CtRNet) which describes the inference process of network. Given an input RGB image, the neural network generates a robot mask and a set of keypoint. Given the associated robot joint angles, a set of corresponding 3D keypoints are computed with forward kinematics. The camera-to-robot pose is estimated by a PnP solver with provided 2D-3D keypoint pairs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "text", + "content": " is a scaling constant, " + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "text", + "content": " is the reprojection loss in the PnP solver (explained in Sec. 3.2), " + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{o}_i|\\mathbf{o}_i\\in \\mathbb{R}^2\\}_{i = 1}^n" + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_i|\\mathbf{p}_i\\in \\mathbb{R}^3\\}_{i = 1}^n" + }, + { + "bbox": [ + 304, + 576, + 547, + 661 + ], + "type": "text", + "content": " are the 2D-3D keypoints inputted into the PnP solver. The exponential function is applied to the weight such that training samples with poor PnP convergence are weighted exponentially lower than good PnP convergence thereby stabilizing the training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 670, + 533, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 533, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 533, + 682 + ], + "type": "text", + "content": "3.2. Camera-to-Robot Pose Estimation Network" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "The overview of the proposed Camera-to-Robot Pose Estimation Network, CtRNet, is shown in Fig. 3. Given an" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21299" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "input RGB image, we employ ResNet50 [17] as the backbone network to extract the latent features. The latent features are then passed through the Atrous Spatial Pyramid Pooling layers [6] to form the segmentation mask of input resolution. The keypoint detector, sharing the backbone network with the foreground segmentation, upsamples the feature maps through transposed convolutional layers and forms the heatmaps with " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " channels. Then, we apply the spatial softmax operator [12] on the heatmaps, which computes the expected 2D location of the points of maximal activation for each channel and results in a set of keypoints " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "[\\mathbf{o}_1,\\dots,\\mathbf{o}_n]" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " channels. For simplicity, we define the set of keypoints at each joint location of the robot. Given the joint angles, the corresponding 3D keypoint location " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " can be calculated with robot forward kinematics:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 258, + 287, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 258, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 104, + 258, + 287, + 274 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {p}}} _ {i} = \\mathbf {T} _ {i} ^ {b} (\\mathbf {q}) \\overline {{\\mathbf {t}}}, f o r i = 1, \\dots , n \\tag {7}", + "image_path": "aea02e110a44a1ce9768a083a49919a0c9f0dc86f21f2aa9dcb86d354f6d5b41.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 279, + 287, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{t} = [0,0,0]" + }, + { + "bbox": [ + 46, + 279, + 287, + 315 + ], + "type": "text", + "content": ". With the 2D and 3D corresponding keypoints, we can then apply a PnP solver [30] to estimate the robot pose with respect to the camera frame." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 399 + ], + "type": "text", + "content": "Back-propagation for PnP Solver. A PnP solver is usually self-contained and not differentiable as the gradient with respect to the input cannot be derived explicitly. Inspired by [5], the implicit function theorem [25] is applied to obtain the gradient through implicit differentiation. Let the PnP solver be denoted as followed in the form of a non-linear function " + }, + { + "bbox": [ + 46, + 316, + 287, + 399 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 316, + 287, + 399 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 407, + 287, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 407, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 129, + 407, + 287, + 421 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {b} ^ {c *} = g (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}) \\tag {8}", + "image_path": "80ddf0902e9fa19b712457c01f51d43f6b5fd293a5862d424a9ed3ba312f71a9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_b^{c*}" + }, + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "type": "text", + "content": " is output pose from the PnP solver. In order to back-propagate through the PnP solver for training the keypoint detector, we are interested in finding the gradient of the output pose " + }, + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_b^{c*}" + }, + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "type": "text", + "content": " with respect to the input 2D points " + }, + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 46, + 427, + 287, + 499 + ], + "type": "text", + "content": ". Note that, the objective of the PnP solver is to minimize the reprojection error, such that:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 506, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 506, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 102, + 506, + 287, + 529 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {b} ^ {c *} = \\underset {\\mathbf {T} _ {b} ^ {c}} {\\arg \\min } O (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) \\tag {9}", + "image_path": "bb5471bb36eca5286d59d1577a9f235248ac5e5adb98308fb1a8310f9efa7da5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 536, + 69, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 69, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 69, + 545 + ], + "type": "text", + "content": "with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 552, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 552, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 64, + 552, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} O \\left(\\mathbf {T} _ {b} ^ {c}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}\\right) = \\sum_ {i = 1} ^ {n} | | \\mathbf {o} _ {i} - \\pi \\left(\\mathbf {p} _ {i} \\mid \\mathbf {T} _ {b} ^ {c}, \\mathbf {K}\\right) | | _ {2} ^ {2} (10) \\\\ = \\sum_ {i = 1} ^ {n} \\left\\| \\mathbf {r} _ {i} \\right\\| _ {2} ^ {2} (11) \\\\ \\end{array}", + "image_path": "5e13b3db64cce819e3fdcd46dc99237f4ce0d38f65a089d5c997c7e81247a58f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\pi(.)" + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "text", + "content": " is the projection operator. Since the optimal solution " + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_b^{c*}" + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "text", + "content": " is a local minimum for the objective function " + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "inline_equation", + "content": "O(\\mathbf{o},\\mathbf{p},\\mathbf{T}_b^c,\\mathbf{K})" + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "text", + "content": ", a stationary constraint of the optimization process can be constructed by taking the first order derivative of the objective function with respect to " + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_b^c" + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 101, + 689, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 689, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 101, + 689, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\frac {\\partial O}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) | _ {\\mathbf {T} _ {b} ^ {c} = \\mathbf {T} _ {b} ^ {c *}} = \\mathbf {0}. \\tag {12}", + "image_path": "ef5d349a44d7cc2fdbb508215173655fa16719ad034f01670cd8214d99c0e05b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "Following [5], we construct a constrain function " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " to employ the implicit function theorem:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 328, + 104, + 545, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 104, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 328, + 104, + 545, + 131 + ], + "type": "interline_equation", + "content": "F (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) = \\frac {\\partial O}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c *}) = \\mathbf {0}. \\tag {13}", + "image_path": "fb8374c40cf7c0418b36113b617622cc4742d32c3356530533fc4573feee6c13.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 139, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 139, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 139, + 545, + 163 + ], + "type": "text", + "content": "Substituting the Eq. (10) and Eq. (11) to Eq. (13), we can derive the constraint function as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 170, + 545, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 170, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 321, + 170, + 545, + 234 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c}) = \\sum_ {i = 1} ^ {n} \\frac {\\partial \\left\\| \\mathbf {r} _ {i} \\right\\| _ {2} ^ {2}}{\\partial \\mathbf {T} _ {b} ^ {c}} (14) \\\\ = - 2 \\sum_ {i = 1} ^ {n} \\mathbf {r} _ {i} ^ {T} \\frac {\\partial \\pi}{\\partial \\mathbf {T} _ {b} ^ {c}} \\left(\\mathbf {p} _ {i} \\mid \\mathbf {T} _ {b} ^ {c *}, \\mathbf {K}\\right). (15) \\\\ \\end{array}", + "image_path": "d5e8e3b651269466d66ba62714e8b06aface12bf48ef603678948ae7dad44eb1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 243, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 280 + ], + "type": "text", + "content": "Finally, we back-propagate through the PnP solver with the implicit differentiation. The gradient of the output pose with respect to the input 2D points is the Jacobian matrix:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 300, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 545, + 365 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial g}{\\partial \\mathbf {o}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}) \\\\ = - \\left(\\frac {\\partial F}{\\partial \\mathbf {T} _ {b} ^ {c}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})\\right) ^ {- 1} \\left(\\frac {\\partial F}{\\partial \\mathbf {o}} (\\mathbf {o}, \\mathbf {p}, \\mathbf {K}, \\mathbf {T} _ {b} ^ {c})\\right). \\tag {16} \\\\ \\end{array}", + "image_path": "a363c4000dc2405cef8f1e37cf981ccfca5abe9b47aa14b2e697b369d2d161bc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 375, + 387, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 387, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 387, + 389 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 396, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 515 + ], + "type": "text", + "content": "We first evaluate our method on two public real-world datasets for robot pose estimation and compare it against several state-of-the-art image-based robot pose estimation algorithms. We then conduct an ablation study on the pretraining procedure and explore how the number of pretraining samples could affect the performance of the self-supervised training. Finally, we integrate the camera-to-robot pose estimation framework into a visual servoing system to demonstrate the effectiveness of our method on real robot applications." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 523, + 481, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 481, + 534 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 541, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 639 + ], + "type": "text", + "content": "DREAM-real Dataset. The DREAM-real dataset [29] is a real-world robot dataset collected with 3 different cameras: Azure Kinect (AK), XBOX 360 Kinect (XK), and RealSense (RS). This dataset contains around 50K RGB images of Franka Emika Panda arm and is recorded at " + }, + { + "bbox": [ + 304, + 541, + 545, + 639 + ], + "type": "inline_equation", + "content": "(640 \\times 480)" + }, + { + "bbox": [ + 304, + 541, + 545, + 639 + ], + "type": "text", + "content": " resolution. The ground-truth camera-to-robot pose is provided for every image frame. The accuracy is evaluated with average distance (ADD) metric [54]," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 356, + 647, + 545, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 356, + 647, + 545, + 678 + ], + "type": "interline_equation", + "content": "A D D = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\left| \\widetilde {\\mathbf {T}} _ {b} ^ {c} \\overline {{\\mathbf {p}}} _ {i} - \\mathbf {T} _ {b} ^ {c} \\overline {{\\mathbf {p}}} _ {i} \\right| \\right| _ {2} \\tag {17}", + "image_path": "3167c7c422f0d5e3c08587019067fba60afd8354e4221ecb9da96e014f1e7a4b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{T}}_b^c" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " indicates the ground-truth camera-to-robot pose. We also report the area-under-the-curve (AUC) value," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21300" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 542, + 152 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 542, + 152 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 542, + 152 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 542, + 152 + ], + "type": "table", + "html": "
MethodCategoryBackbonePanda 3CAM-AKPanda 3CAM-XKPanda 3CAM-RSPanda ORBAll
AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓AUC ↑Mean (m) ↓
DREAM-F [29]KeypointVGG1968.91211.41324.359491.91176.1302.07761.93095.31960.740113.029
DREAM-Q [29]KeypointVGG1952.38278.08937.47154.17877.9840.02757.08767.24856.98859.284
DREAM-H [29]KeypointResNet10160.5200.05664.0057.38278.8250.02469.05425.68568.58417.477
RoboPose [26]RenderingResNet3476.4970.02485.9260.01476.8630.02380.5040.01980.0940.020
CtRNetKeypointResNet5089.9280.01379.4650.03290.7890.01085.2890.02185.9620.020
", + "image_path": "d1cad9aaa05ff6e201cc9e77ae0912acf99c8b5d0008cebc053cc5c3025cef0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 87, + 193, + 509, + 291 + ], + "blocks": [ + { + "bbox": [ + 46, + 159, + 546, + 183 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 546, + 183 + ], + "type": "text", + "content": "Table 1. Comparison of our methods with the state-of-the-art methods on DREAM-real datasets using ADD metric. We report the mean and AUC of the ADD on each dataset and the overall accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 193, + 509, + 291 + ], + "lines": [ + { + "bbox": [ + 87, + 193, + 509, + 291 + ], + "spans": [ + { + "bbox": [ + 87, + 193, + 509, + 291 + ], + "type": "table", + "html": "
MethodCategoryPCK (2D)Mean 2D Err. (pixel) ↓ADD (3D)Mean 3D Err. (mm) ↓
@50 pixel ↑AUC ↑@100 mm ↑AUC ↑
Aruco Marker [14]Keypoint0.4957.15286.980.3043.452447.34
DREAM-Q [29]Keypoint0.3344.011277.330.3240.63386.17
Opt. Keypoints [34]Keypoint0.6975.4649.510.4765.66141.05
Diff. RenderingRendering0.7478.6042.300.7881.1574.95
CtRNetKeypoint0.9993.9411.620.8883.9363.81
", + "image_path": "3c5a4b987d667a9a37f5aa0e894e7cc0af221469dbde39755eeacb3f3d9a516f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 299, + 457, + 310 + ], + "lines": [ + { + "bbox": [ + 135, + 299, + 457, + 310 + ], + "spans": [ + { + "bbox": [ + 135, + 299, + 457, + 310 + ], + "type": "text", + "content": "Table 2. Comparison of our methods with the state-of-the-art methods on Baxter dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 331, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 366 + ], + "type": "text", + "content": "which integrates the percentage of ADD over different thresholds. A higher AUC value indicates more predictions with less error." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 465 + ], + "type": "text", + "content": "Baxter Dataset. The Baxter dataset [34] contains 100 RGB images of the left arm of Rethink Baxter collected with Azure Kinect camera at " + }, + { + "bbox": [ + 46, + 369, + 287, + 465 + ], + "type": "inline_equation", + "content": "(2048\\times 1526)" + }, + { + "bbox": [ + 46, + 369, + 287, + 465 + ], + "type": "text", + "content": " resolution. The 2D and 3D ground-truth end-effector position with respect to the camera frame is provided. We evaluate the performance with the ADD metric for the end-effector. We also evaluate the end-effector reprojection error using the percentage of correct keypoints (PCK) metric [34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 478, + 178, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 178, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 178, + 490 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 498, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 288, + 715 + ], + "type": "text", + "content": "The entire pipeline is implemented in PyTorch [40]. We initialize the backbone network with ImageNet [9] pretrained weights, and we train separate networks for different robots. The number of keypoints " + }, + { + "bbox": [ + 45, + 498, + 288, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 45, + 498, + 288, + 715 + ], + "type": "text", + "content": " is set to the number of robot links and the keypoints are defined at the robot joint locations. The neural network is pretrained on synthetic data for foreground segmentation and keypoint detection for 1000 epochs with 1e-5 learning rate. We reduce the learning rate by a factor of 10 once learning stagnates for 5 epochs. The Adam optimizer is applied to optimize the network parameters with the momentum set to 0.9. For self-supervised training on real-world data, we run the training for 500 epochs with 1e-6 learning rate. The same learning rate decay strategy and Adam optimizer is applied here similar to the pretraining. To make the training more stable, we clip the gradient of the network parameters at 10. The scaling factor in Eq. (6) is set to 0.1 for DREAM-real dataset and 0.01 for Baxter dataset, mainly accounting for" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 331, + 418, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 418, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 418, + 342 + ], + "type": "text", + "content": "the difference in resolution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 356, + 545, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 356, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 545, + 369 + ], + "type": "text", + "content": "4.3. Robot Pose Estimation on Real-world Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 376, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 616 + ], + "type": "text", + "content": "Evaluation on DREAM-real Dataset. The proposed CtRNet is trained at " + }, + { + "bbox": [ + 304, + 376, + 546, + 616 + ], + "type": "inline_equation", + "content": "(320\\times 240)" + }, + { + "bbox": [ + 304, + 376, + 546, + 616 + ], + "type": "text", + "content": " resolution and evaluated at the original resolution by scaling up the keypoints by a factor of 2. Some qualitative results for foreground segmentation and pose estimation are shown in the Fig. 4a. We compared our method with the state-of-the-art keypoint-based method DREAM [29] and the rendering-based method RoboPose [26]. The results for DREAM and RoboPose are compiled from the implementation provided by [26]. In Tab. 1, we report the AUC and mean ADD results on DREAM-real dataset with 3 different camera settings and the overall results combining all the test samples. Our method has a significantly better performance compared to the method in the same category and achieves comparable performance with the rendering-based method. We outperform DREAM on all settings and outperform RoboPose on the majority of the dataset. Overall on DREAM-real dataset, we achieve higher AUC (+17.378 compared to DREAM, +5.868 compared to RoboPose), and lower error compared to DREAM (-17.457)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Evaluation on Baxter Dataset. For the Baxter dataset, we trained the CtRNet at " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "(640\\times 480)" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " resolution and evaluate at the original resolution, and Fig. 4b shows some of the qualitative results. We compared our method with several keypoint-based methods (Aruco Marker [14], DREAM [29], Optimized Keypoints [34]). We also implemented Differentiable Rendering for robot pose estimation, where the robot masks are generated with the pre" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21301" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 145, + 135 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 145, + 135 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 145, + 135 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 145, + 135 + ], + "type": "image", + "image_path": "d761695e91347c02df25ad2c1c064d94da4919065ca4dcacf2b0ca7dd026283a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 72, + 220, + 135 + ], + "blocks": [ + { + "bbox": [ + 145, + 72, + 220, + 135 + ], + "lines": [ + { + "bbox": [ + 145, + 72, + 220, + 135 + ], + "spans": [ + { + "bbox": [ + 145, + 72, + 220, + 135 + ], + "type": "image", + "image_path": "317d59fc022df35c7f281319d92388d65c112eb62a59ac36babe4e7adb56cbd4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 220, + 72, + 294, + 135 + ], + "blocks": [ + { + "bbox": [ + 220, + 72, + 294, + 135 + ], + "lines": [ + { + "bbox": [ + 220, + 72, + 294, + 135 + ], + "spans": [ + { + "bbox": [ + 220, + 72, + 294, + 135 + ], + "type": "image", + "image_path": "07a26f618bac91534b11f760792f59a3c019dd37cf42482f2f7ef295667f7eab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 372, + 135 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 372, + 135 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 372, + 135 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 372, + 135 + ], + "type": "image", + "image_path": "886bc7de660b5a96b67dbfc72fe635f18e06ea295c3d57ce79bd126e06b34510.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 373, + 72, + 447, + 135 + ], + "blocks": [ + { + "bbox": [ + 373, + 72, + 447, + 135 + ], + "lines": [ + { + "bbox": [ + 373, + 72, + 447, + 135 + ], + "spans": [ + { + "bbox": [ + 373, + 72, + 447, + 135 + ], + "type": "image", + "image_path": "225fcf48a322d09a908a616bbc4beb8d42c0e09ae5b30e8366259a012c9f7734.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 448, + 72, + 524, + 135 + ], + "blocks": [ + { + "bbox": [ + 448, + 72, + 524, + 135 + ], + "lines": [ + { + "bbox": [ + 448, + 72, + 524, + 135 + ], + "spans": [ + { + "bbox": [ + 448, + 72, + 524, + 135 + ], + "type": "image", + "image_path": "52a07cabea14f9c7ccc54fe6c837b32a0ba69d642db33f5d63d5af54e5c2f3d6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 136, + 145, + 199 + ], + "blocks": [ + { + "bbox": [ + 70, + 136, + 145, + 199 + ], + "lines": [ + { + "bbox": [ + 70, + 136, + 145, + 199 + ], + "spans": [ + { + "bbox": [ + 70, + 136, + 145, + 199 + ], + "type": "image", + "image_path": "09b344d883b9e1271789ee2b0a240ec5e01647c4d94f1f4ddd4787ff3b937cc3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 145, + 136, + 220, + 199 + ], + "blocks": [ + { + "bbox": [ + 145, + 136, + 220, + 199 + ], + "lines": [ + { + "bbox": [ + 145, + 136, + 220, + 199 + ], + "spans": [ + { + "bbox": [ + 145, + 136, + 220, + 199 + ], + "type": "image", + "image_path": "631e13a4bf8f969d0d2bf1fbd52082cded94d76e2e152f04e263059fb85b8dea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 220, + 136, + 294, + 199 + ], + "blocks": [ + { + "bbox": [ + 220, + 136, + 294, + 199 + ], + "lines": [ + { + "bbox": [ + 220, + 136, + 294, + 199 + ], + "spans": [ + { + "bbox": [ + 220, + 136, + 294, + 199 + ], + "type": "image", + "image_path": "e40a327b4bf64a02913c387ff1d7b3602d0abf79d9fb6b07179c6ed3cc0ac289.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 296, + 136, + 372, + 199 + ], + "blocks": [ + { + "bbox": [ + 296, + 136, + 372, + 199 + ], + "lines": [ + { + "bbox": [ + 296, + 136, + 372, + 199 + ], + "spans": [ + { + "bbox": [ + 296, + 136, + 372, + 199 + ], + "type": "image", + "image_path": "15fe4104716543d754256f385e26d8417d8066bc9938e42fdffaadbcb7197019.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 373, + 136, + 447, + 199 + ], + "blocks": [ + { + "bbox": [ + 373, + 136, + 447, + 199 + ], + "lines": [ + { + "bbox": [ + 373, + 136, + 447, + 199 + ], + "spans": [ + { + "bbox": [ + 373, + 136, + 447, + 199 + ], + "type": "image", + "image_path": "6f54349c7c60ab6c481e78c669aa967c090b5c2eeed7f94b21cd956f796d189e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 448, + 136, + 524, + 199 + ], + "blocks": [ + { + "bbox": [ + 448, + 136, + 524, + 199 + ], + "lines": [ + { + "bbox": [ + 448, + 136, + 524, + 199 + ], + "spans": [ + { + "bbox": [ + 448, + 136, + 524, + 199 + ], + "type": "image", + "image_path": "840fbef587f090e443af1a6ba3959736cb7453e461febda1ccd4d62a18f2d19a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 70, + 200, + 145, + 263 + ], + "blocks": [ + { + "bbox": [ + 70, + 200, + 145, + 263 + ], + "lines": [ + { + "bbox": [ + 70, + 200, + 145, + 263 + ], + "spans": [ + { + "bbox": [ + 70, + 200, + 145, + 263 + ], + "type": "image", + "image_path": "bdf50a92cf4823f9db92e0648d9cb853bd85e179b382c1c9331c18275b7cba5e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "type": "text", + "content": "Figure 4. Qualitative results of CtRNet foreground segmentation and pose estimation on (a) DREAM-real dataset and (b) Baxter dataset. The first row shows the input RGB image, the second row shows the foreground segmentation, and the third row shows the projected robot skeleton based on the estimated robot pose." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 145, + 200, + 220, + 263 + ], + "blocks": [ + { + "bbox": [ + 145, + 200, + 220, + 263 + ], + "lines": [ + { + "bbox": [ + 145, + 200, + 220, + 263 + ], + "spans": [ + { + "bbox": [ + 145, + 200, + 220, + 263 + ], + "type": "image", + "image_path": "19e24995ff490a764033db121bf6bf0917118d458173f83809f3cde709516e73.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 266, + 187, + 275 + ], + "lines": [ + { + "bbox": [ + 176, + 266, + 187, + 275 + ], + "spans": [ + { + "bbox": [ + 176, + 266, + 187, + 275 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 220, + 200, + 294, + 263 + ], + "blocks": [ + { + "bbox": [ + 220, + 200, + 294, + 263 + ], + "lines": [ + { + "bbox": [ + 220, + 200, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 220, + 200, + 294, + 263 + ], + "type": "image", + "image_path": "757ec12934e15a3373e017d8039a3d4dbecea80472284074ad7393f12e9c55ce.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 296, + 200, + 372, + 263 + ], + "blocks": [ + { + "bbox": [ + 296, + 200, + 372, + 263 + ], + "lines": [ + { + "bbox": [ + 296, + 200, + 372, + 263 + ], + "spans": [ + { + "bbox": [ + 296, + 200, + 372, + 263 + ], + "type": "image", + "image_path": "be4661f3725e6a40a3e0a8517df899cd173db4081d811590d4012f51d6a89aee.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 373, + 200, + 447, + 263 + ], + "blocks": [ + { + "bbox": [ + 373, + 200, + 447, + 263 + ], + "lines": [ + { + "bbox": [ + 373, + 200, + 447, + 263 + ], + "spans": [ + { + "bbox": [ + 373, + 200, + 447, + 263 + ], + "type": "image", + "image_path": "f4deb021922b3c713d04345603c9af0146adce6cb2fc75a875387565562aa5ee.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 266, + 415, + 275 + ], + "lines": [ + { + "bbox": [ + 404, + 266, + 415, + 275 + ], + "spans": [ + { + "bbox": [ + 404, + 266, + 415, + 275 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 448, + 200, + 524, + 263 + ], + "blocks": [ + { + "bbox": [ + 448, + 200, + 524, + 263 + ], + "lines": [ + { + "bbox": [ + 448, + 200, + 524, + 263 + ], + "spans": [ + { + "bbox": [ + 448, + 200, + 524, + 263 + ], + "type": "image", + "image_path": "7d2a320cf8745b6d38f188ddfa566554e19015860136306581df6041b9f95130.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 340, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 289, + 471 + ], + "type": "text", + "content": "trained foreground segmentation. The 2D PCK results and 3D ADD results are reported in Tab. 2. Our method outperforms all other methods on both 2D and 3D evaluations. For 2D evaluation, we achieve 93.94 AUC for PCK with an average reprojection error of 11.62 pixels. For 3D evaluation, we achieve 83.93 AUC for ADD with an average ADD of " + }, + { + "bbox": [ + 46, + 340, + 289, + 471 + ], + "type": "inline_equation", + "content": "63.81\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 340, + 289, + 471 + ], + "type": "text", + "content": ". Notably, 99 percent of our estimation has less than 50 pixel reprojection error, which is less than 2 percent of the image resolution, and 88 percent of our estimation has less than " + }, + { + "bbox": [ + 46, + 340, + 289, + 471 + ], + "type": "inline_equation", + "content": "100\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 340, + 289, + 471 + ], + "type": "text", + "content": " distance error when localizing the end-effector." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 479, + 140, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 140, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 140, + 493 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "We study how the number of pretraining samples affects the convergence and performance of the self-supervised training empirically on the Baxter dataset. We pretrain the neural network with different numbers of synthetic data samples " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{\\text{pretrain}} = \\{500, 1000, 2000, 4000, 8000\\}" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": ", and examine the convergence of the self-supervised training process. Fig. 5 shows the plot of self-training loss " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\text{mask}} + \\mathcal{L}_{\\text{seg}})" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " vs. the number of epochs for networks pretrianed with different number of synthetic data. We observe that doubling the size of pretraining dataset significantly improves the convergence of the self-training process at the beginning. However, the improvement gets smaller as the pretraining size increases. For the Baxter dataset, the improvement saturates after having more than 2000 pretraining samples. Continuing double the training size results in very marginal improvement. Noted that the Baxter dataset captures 20 different robot poses from a fixed camera position. The required number of pretraining samples might" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 317, + 340, + 536, + 475 + ], + "blocks": [ + { + "bbox": [ + 317, + 340, + 536, + 475 + ], + "lines": [ + { + "bbox": [ + 317, + 340, + 536, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 340, + 536, + 475 + ], + "type": "image", + "image_path": "b1cdde4800a9067745e8bfe7b3d7d8b96b8e86483212ef2d94491268d300ebb1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 483, + 545, + 540 + ], + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 540 + ], + "type": "text", + "content": "Figure 5. The training loss vs. number of epochs for the self-supervised training with different numbers of pretraining samples. More pretraining samples results in better convergence. The improvement saturates after having more than 2000 pretraining samples as only marginal improvement by adding more samples." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 554, + 520, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 520, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 520, + 565 + ], + "type": "text", + "content": "vary according to the complexity of the environment." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "type": "text", + "content": "We further evaluate the resulting neural networks with the ground-truth labels on the Baxter dataset. We report the mean ADD and AUC ADD for the pose estimation in Tab. 3. The result verifies our observation on the convergence analysis. Having more pretraining samples improves the performance of pose estimation at the beginning, but the improvement stagnates after having more than 2000 pretraining samples." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 670, + 460, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 460, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 460, + 684 + ], + "type": "text", + "content": "4.5. Visual Servoing Experiment" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "We integrate the proposed CtRNet into a robotic system for position-based visual servoing (PBVS) with eye-to-hand" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21302" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 265, + 155 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 265, + 155 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 265, + 155 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 265, + 155 + ], + "type": "table", + "html": "
NpretrainMean ADD (mm) ↓AUC ADD ↑
5002167.3047.62
100092.9176.65
200067.5182.98
400063.0084.12
800063.8183.93
", + "image_path": "09aee82883ac1405300602a57aae7d4d394e81529cc67f10c74e02c51d813cb9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 186, + 286, + 242 + ], + "blocks": [ + { + "bbox": [ + 53, + 163, + 279, + 175 + ], + "lines": [ + { + "bbox": [ + 53, + 163, + 279, + 175 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 279, + 175 + ], + "type": "text", + "content": "Table 3. Ablation study for the number of pretraining samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 186, + 286, + 242 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 286, + 242 + ], + "type": "table", + "html": "
MethodLoop RateTrans. Err. (m)Rot. Err. (rad)
DREAM [29]30Hz0.235 ± 0.3130.300 ± 0.544
Diff. Rendering1Hz0.046 ± 0.0620.036 ± 0.066
CtRNet30Hz0.002 ± 0.0010.002 ± 0.001
", + "image_path": "c12820ca3fcede1cf7a4bf56ff60f57f525ebc271b4b9630d4c265acaf7019ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 250, + 286, + 272 + ], + "lines": [ + { + "bbox": [ + 47, + 250, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 286, + 272 + ], + "type": "text", + "content": "Table 4. Mean and standard deviation of the translational error and rotational error for the visual servoing experiment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 294, + 286, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 286, + 437 + ], + "type": "text", + "content": "configuration. We conduct the experiment on a Baxter robot and the details of the PBVS are described in the Supplementary Materials. The PBVS is purely based on RGB images from a single camera and the goal is to control the robot end-effector reaching a target pose defined in the camera frame. Specifically, we first set a target pose with respect to the camera frame. The target pose is then transformed into the robot base frame through the estimated camera-to-robot transformation. The robot controller calculates the desired robot configuration with inverse kinematics and a control law is applied to move the robot end-effector toward the target pose." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "text", + "content": "For comparison, we also implemented DREAM [29] and a Differentiable Renderer for PBVS. For DREAM, the pretrained model for Baxter is applied. For Differentiable Renderer, we use the foreground segmentation of CtRNet to generate a robot mask. The optimizing loop for the renderer takes the last estimation as initialization and performs 10 updates at each callback to ensure convergence and maintain " + }, + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "inline_equation", + "content": "1\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "text", + "content": " loop rate. In the experiment, we randomly set the target pose and the position of the camera, and the robotic system applies PBVS to reach the target pose from an arbitrary initialization, as shown in the Fig. 6. We ran the experiment for 10 trails with different robot pose estimation methods, and the translational (Euclidean distance) and rotational errors (Euler angles) of the end-effector are reported in Tab. 4. The experimental results show that our proposed method significantly improves the stability and accuracy of the PBVS, achieving " + }, + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "inline_equation", + "content": "0.002\\mathrm{m}" + }, + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "text", + "content": " averaged translational error and " + }, + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "inline_equation", + "content": "0.002\\mathrm{rad}" + }, + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "text", + "content": " rotational error on the end-effector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "We also plot the end-effector distance-to-goal over time for a selected trail in Fig. 7. In this selected trial, the system could not converge with DREAM because the poor robot pose estimation confuses the controller by giving the wrong target pose in the robot base frame, which is unreachable." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 425, + 158 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 425, + 158 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 425, + 158 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 425, + 158 + ], + "type": "image", + "image_path": "ba92e43337ba0f497e943dd52229464de4eaca582f08f4da2069cfd9100316e8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 168, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 201 + ], + "type": "text", + "content": "Figure 6. Snapshots of PBVS. The goal is to move the end-effector to the target pose (green). The figure on the right shows the robot configuration upon the convergence of PBVS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 72, + 537, + 158 + ], + "blocks": [ + { + "bbox": [ + 426, + 72, + 537, + 158 + ], + "lines": [ + { + "bbox": [ + 426, + 72, + 537, + 158 + ], + "spans": [ + { + "bbox": [ + 426, + 72, + 537, + 158 + ], + "type": "image", + "image_path": "cbf61d722171064c1276569aa84bbcc2864d13de42fcd0708f54f01e18413f43.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 212, + 535, + 327 + ], + "blocks": [ + { + "bbox": [ + 316, + 212, + 535, + 327 + ], + "lines": [ + { + "bbox": [ + 316, + 212, + 535, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 535, + 327 + ], + "type": "image", + "image_path": "c84a6549ed07db415ab31a8f79c3667aaa14a8ab97efd1363a20cc8a1a25b89e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 339, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 306, + 339, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 545, + 361 + ], + "type": "text", + "content": "Figure 7. The plot of end-effector distance-to-goal over time on a selected PBVS trail." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 370, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 545, + 441 + ], + "type": "text", + "content": "With the differentiable renderer, the servoing system takes more than 10 seconds to converge and oscillate due to the low loop rate. With our proposed CtRNet, the servoing system converges much faster (" + }, + { + "bbox": [ + 305, + 370, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\leq 5" + }, + { + "bbox": [ + 305, + 370, + 545, + 441 + ], + "type": "text", + "content": " seconds), thanks to the fast and robust robot pose estimation. We show more qualitative results in the Supplementary Materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "content": "We present the CtRNet, an end-to-end image-based robot pose estimation framework, and a self-supervised training pipeline that utilizes unlabelled real-world data for sim-to-real transfer. The CtRNet, using a keypoint detector for pose estimation while employing a rendering method for training, achieves state-of-the-art performance on robot pose estimation while maintaining high-speed inference. The Fig. 1 illustrates the advantages of CtRNet over existing methods, where the AUC values are normalized across two evaluation datasets by taking DREAM and CtRNet as references. We further experiment with different robot pose estimation methods by applying them to PBVS, which demonstrates CtRNet's fast and accurate robot pose estimation enabling stability when using single-frame robot pose estimation for feedback. Therefore, CtRNet supports real-time markerless camera-to-robot pose estimation which has been utilized for surgical robotic manipulation [43] and mobile robot manipulators [53]. For future work, we would like to extend our method to more robots and explore vision-based control in an unstructured environment." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21303" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Jeannette Bohg, Javier Romero, Alexander Herzog, and Stefan Schaal. Robot arm pose estimation through pixel-wise part classification. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 3143-3150. IEEE, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "text", + "content": "[2] Joao Borrego, Rui Figueiredo, Atabak Dehban, Plinio Moreno, Alexandre Bernardino, and José Santos-Victor. A generic visual perception domain randomisation framework for gazebo. In 2018 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC), pages 237-242. IEEE, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 216, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 291 + ], + "type": "text", + "content": "[3] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4243-4250. IEEE, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 293, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 287, + 327 + ], + "type": "text", + "content": "[4] François Chaumette, Seth Hutchinson, and Peter Corke. Visual servoing. In Springer Handbook of Robotics, pages 841-866. Springer, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 384 + ], + "type": "text", + "content": "[5] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8100-8109, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "text", + "content": "[7] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3339-3348, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 487, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 519 + ], + "type": "text", + "content": "[8] Jacques Denavit and Richard S Hartenberg. A kinematic notation for lower-pair mechanisms based on matrices. 1955. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 287, + 566 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "type": "text", + "content": "[10] Karthik Desingh, Shiyang Lu, Anthony Opipari, and Odest Chadwicke Jenkins. Factored pose estimation of articulated objects using efficient nonparametric belief propagation. In 2019 International Conference on Robotics and Automation (ICRA), pages 7221-7227. IEEE, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[11] Irene Fassi and Giovanni Legnani. Hand to sensor calibration: A geometrical interpretation of the matrix equation " + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "inline_equation", + "content": "\\mathbf{ax} =" + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": " xb. Journal of Robotic Systems, 22(9):497-506, 2005. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Chelsea Finn, Xin Yu Tan, Yan Duan, Trevor Darrell, Sergey Levine, and Pieter Abbeel. Deep spatial autoencoders for visuomotor learning. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 512-519. IEEE, 2016. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[13] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930–943, 2003. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[14] Sergio Garrido-Jurado, Rafael Muñoz-Salinas, Francisco José Madrid-Cuevas, and Manuel Jesús Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. Pattern Recognition, 47(6):2280-2292, 2014. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "text", + "content": "[15] Abhishek Gupta, Coline Devin, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Learning invariant feature spaces to transfer skills with reinforcement learning. arXiv preprint arXiv:1703.02949, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "type": "text", + "content": "[16] Ran Hao, Orhan Özgün, and M Cenk Çavusoglu. Vision-based surgical tool pose estimation for the da Vinci® robotic surgical system. In 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 1298-1305. IEEE, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "text", + "content": "[18] Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, and Kurt Konolige. On pre-trained image features and synthetic images for deep learning. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "text", + "content": "[19] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 444, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 475 + ], + "type": "text", + "content": "[20] Radu Horaud and Fadi Dornaika. Hand-eye calibration. The international journal of robotics research, 14(3):195-210, 1995. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[21] Daniel Horváth, Gábor Erdős, Zoltán Istenes, Tomás Horváth, and Sándor Foldi. Object detection using sim2real domain randomization for robotic applications. IEEE Transactions on Robotics, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "text", + "content": "[22] Jarmo Ilonen and Ville Kyrki. Robust robot-camera calibration. In 2011 15th International Conference on Advanced Robotics (ICAR), pages 67-74. IEEE, 2011. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 634 + ], + "type": "text", + "content": "[23] Stephen James, Paul Wohlhart, Mrinal Kalakrishnan, Dmitry Kalashnikov, Alex Irpan, Julian Ibarz, Sergey Levine, Raia Hadsell, and Konstantinos Bousmalis. Sim-to-real via sim-to-sim: Data-efficient robotic grasping via randomized-to-canonical adaptation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12627-12637, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[24] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907-3916, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[25] Steven George Krantz and Harold R Parks. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21304" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Single-view robot pose and joint angle estimation via render & compare. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1654-1663, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "text", + "content": "[27] Jens Lambrecht, Philipp Grosenick, and Marvin Meusel. Optimizing keypoint-based single-shot camera-to-robot pose estimation through shape segmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13843–13849. IEEE, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[28] Jens Lambrecht and Linh Kästner. Towards the usage of synthetic data for marker-less pose estimation of articulated robots in rgb images. In 2019 19th International Conference on Advanced Robotics (ICAR), pages 240-247. IEEE, 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 306 + ], + "type": "text", + "content": "[29] Timothy E Lee, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Oliver Kroemer, Dieter Fox, and Stan Birchfield. Camera-to-robot pose estimation from a single image. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9426–9432. IEEE, 2020. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "text", + "content": "[30] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. International journal of computer vision, 81(2):155-166, 2009. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "text", + "content": "[31] Yang Li, Florian Richter, Jingpei Lu, Emily K Funk, Ryan K Orosco, Jianke Zhu, and Michael C Yip. Super: A surgical perception framework for endoscopic tissue manipulation with surgical robotics. IEEE Robotics and Automation Letters, 5(2):2294-2301, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "text", + "content": "[32] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "text", + "content": "[33] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[34] Jingpei Lu, Florian Richter, and Michael C Yip. Pose estimation for robot manipulators via keypoint optimization and sim-to-real transfer. IEEE Robotics and Automation Letters, 7(2):4622-4629, 2022. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[35] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 4(26):eaau4984, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[36] Fabian Manhardt, Wadim Kehl, Nassir Navab, and Federico Tombari. Deep model-based 6d pose refinement in rgb. In Proceedings of the European Conference on Computer Vision (ECCV), pages 800-815, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[37] Frank Michel, Alexander Krull, Eric Brachmann, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Pose estimation of kinematic chain instances via object coordinate regression. In BMVC, pages 181-1, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[38] Edwin Olson. Apriltag: A robust and flexible visual fiducial system. In 2011 IEEE international conference on robotics and automation, pages 3400-3407. IEEE, 2011. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[39] Frank C Park and Bryan J Martin. Robot sensor calibration: solving AX= XB on the Euclidean group. IEEE Transactions on Robotics and Automation, 10(5):717-721, 1994. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[40] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "text", + "content": "[41] Karl Pauwels, Leonardo Rubio, and Eduardo Ros. Real-time model-based articulated object pose detection and tracking with variable rigidity constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3994-4001, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 206, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 250 + ], + "type": "text", + "content": "[42] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "type": "text", + "content": "[43] Florian Richter, Jingpei Lu, Ryan K Orosco, and Michael C Yip. Robotic tool tracking under partially visible kinematic chain: A unified approach. IEEE Transactions on Robotics, 2021. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 294, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 350 + ], + "type": "text", + "content": "[44] Swami Sankaranarayanan, Yogesh Balaji, Arpit Jain, Ser Nam Lim, and Rama Chellappa. Learning from synthetic data: Addressing domain shift for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3752-3761, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[45] Tanner Schmidt, Richard A Newcombe, and Dieter Fox. Dart: Dense articulated real-time tracking. In Robotics: Science and Systems, volume 2, pages 1-9. Berkeley, CA, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 393, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 449 + ], + "type": "text", + "content": "[46] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 524 + ], + "type": "text", + "content": "[47] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3482-3489. IEEE, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 526, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 591 + ], + "type": "text", + "content": "[48] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 592, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 669 + ], + "type": "text", + "content": "[49] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969–977, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 714 + ], + "type": "text", + "content": "[50] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. arXiv preprint arXiv:1809.10790, 2018. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21305" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 352 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "text", + "content": "[51] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 96, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 96, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 96, + 286, + 149 + ], + "type": "text", + "content": "[52] Felix Widmaier, Daniel Kappler, Stefan Schaal, and Jeannette Bohg. Robot arm pose estimation by pixel-wise regression of joint angles. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 616-623. IEEE, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "text", + "content": "[53] Melonee Wise, Michael Ferguson, Derek King, Eric Diehr, and David Dymesich. Fetch and freight: Standard platforms for service robot applications. In Workshop on autonomous mobile service robots, 2016. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "type": "text", + "content": "[54] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 294 + ], + "type": "text", + "content": "[55] Wenshuai Zhao, Jorge Pña Queralta, and Tomi Westerlund. Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In 2020 IEEE Symposium Series on Computational Intelligence (SSCI), pages 737-744. IEEE, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 286, + 352 + ], + "type": "text", + "content": "[56] Yiming Zuo, Weichao Qiu, Lingxi Xie, Fangwei Zhong, Yizhou Wang, and Alan L Yuille. Craves: Controlling robotic arm with a vision-based economic system. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4214-4223, 2019. 2, 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21306" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_content_list.json b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18898056a462281d486584b2ec13891e92c41866 --- /dev/null +++ b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_content_list.json @@ -0,0 +1,1402 @@ +[ + { + "type": "text", + "text": "Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation", + "text_level": 1, + "bbox": [ + 104, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feng Li $^{1,3*†}$ , Hao Zhang $^{1,3*†}$ , Huaizhe Xu $^{1,3}$ , Shilong Liu $^{2,3}$ , Lei Zhang $^{3‡}$ , Lionel M. Ni $^{1,4}$ , Heung-Yeung Shum $^{1,3}$", + "bbox": [ + 202, + 202, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Hong Kong University of Science and Technology.", + "bbox": [ + 263, + 239, + 705, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Dept. of CST., BNRist Center, Institute for AI, Tsinghua University.", + "bbox": [ + 212, + 256, + 756, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ International Digital Economy Academy (IDEA).", + "bbox": [ + 285, + 273, + 683, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4The Hong Kong University of Science and Technology (Guangzhou).", + "bbox": [ + 210, + 292, + 759, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fiay,hzhangcx,hxubr}@connect.ust.hk {liusl20}@mails.tsinghua.edu.cn {leizhang}@idea.edu.cn {ni,hshum}@ust.hk", + "bbox": [ + 220, + 321, + 761, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 369, + 313, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper we present Mask DINO, a unified object detection and segmentation framework. Mask DINO extends DINO (DETR with Improved Denoising Anchor Boxes) by adding a mask prediction branch which supports all image segmentation tasks (instance, panoptic, and semantic). It makes use of the query embeddings from DINO to dot-product a high-resolution pixel embedding map to predict a set of binary masks. Some key components in DINO are extended for segmentation through a shared architecture and training process. Mask DINO is simple, efficient, and scalable, and it can benefit from joint large-scale detection and segmentation datasets. Our experiments show that Mask DINO significantly outperforms all existing specialized segmentation methods, both on a ResNet-50 backbone and a pre-trained model with SwinL backbone. Notably, Mask DINO establishes the best results to date on instance segmentation (54.5 AP on COCO), panoptic segmentation (59.4 PQ on COCO), and semantic segmentation (60.8 mIoU on ADE20K) among models under one billion parameters. Code is available at https://github.com/IDEA-Research/MaskDINO.", + "bbox": [ + 76, + 401, + 473, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 750, + 209, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection and image segmentation are fundamental tasks in computer vision. Both tasks are concerned with localizing objects of interest in an image but have different levels of focus. Object detection is to localize objects of interest and predict their bounding boxes and category", + "bbox": [ + 75, + 775, + 470, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "labels, whereas image segmentation focuses on pixel-level grouping of different semantics. Moreover, image segmentation encompasses various tasks including instance segmentation, panoptic segmentation, and semantic segmentation with respect to different semantics, e.g., instance or category membership, foreground or background category.", + "bbox": [ + 500, + 369, + 893, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Remarkable progress has been achieved by classical convolution-based algorithms developed for these tasks with specialized architectures, such as Faster RCNN [24] for object detection, Mask RCNN [9] for instance segmentation, and FCN [21] for semantic segmentation. Although these methods are conceptually simple and effective, they are tailored for specialized tasks and lack the generalization ability to address other tasks. The ambition to bridge different tasks gives rise to more advanced methods like HTC [2] for object detection and instance segmentation and Panoptic FPN [14], K-net [33] for instance, panoptic, and semantic segmentation. Task unification not only helps simplify algorithm development but also brings in performance improvement in multiple tasks.", + "bbox": [ + 496, + 462, + 895, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, DETR-like [1] models developed based on Transformers [27] have achieved inspiring progress on many detection and segmentation tasks. As an end-to-end object detector, DETR adopts a set-prediction objective and eliminates hand-crafted modules such as anchor design and nonmaximum suppression. Although DETR addresses both the object detection and panoptic segmentation tasks, its segmentation performance is still inferior to classical segmentation models. To improve the detection and segmentation performance of Transformer-based models, researchers have developed specialized models for object detection [15, 18, 32, 35], image segmentation [3, 4, 33], instance segmentation [7], panoptic segmentation [23], and semantic segmentation [12].", + "bbox": [ + 496, + 672, + 895, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among the efforts to improve object detection, DINO [32] takes advantage of the dynamic anchor box for", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 93, + 862, + 205, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done when Feng Li and Hao Zhang were interns at IDEA.", + "bbox": [ + 96, + 875, + 439, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3041", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mulation from DAB-DETR [18] and query denoising training from DN-DETR [15], and further achieves the SOTA result on the COCO object detection leaderboard for the first time as a DETR-like model. Similarly, for improving image segmentation, MaskFormer [4] and Mask2Former [3] propose to unify different image segmentation tasks using query-based Transformer architectures to perform mask classification. Such methods have achieved remarkable performance improvement on multiple segmentation tasks.", + "bbox": [ + 75, + 90, + 470, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, in Transformer-based models, the best-performing detection and segmentation models are still not unified, which prevents task and data cooperation between detection and segmentation tasks. As an evidence, in CNN-based models, Mask-R-CNN [9] and HTC [2] are still widely acknowledged as unified models that achieve mutual cooperation between detection and segmentation to achieve superior performance than specialized models. Though we believe detection and segmentation can help each other in a unified architecture in Transformer-based models, the results of simply using DINO for segmentation and using Mask2Former for detection indicate that they can not do other tasks well, as shown in Table 1 and 2. Moreover, trivial multi-task training can even hurt the performance of the original tasks. It naturally leads to two questions: 1) why cannot detection and segmentation tasks help each other in Transformer-based models? and 2) is it possible to develop a unified architecture to replace specialized ones?", + "bbox": [ + 75, + 229, + 470, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these problems, we propose Mask DINO, which extends DINO with a mask prediction branch in parallel with DINO's box prediction branch. Inspired by other unified models [3, 4, 28] for image segmentation, we reuse content query embeddings from DINO to perform mask classification for all segmentation tasks on a high-resolution pixel embedding map (1/4 of the input image resolution) obtained from the backbone and Transformer encoder features. The mask branch predicts binary masks by simply dot-producting each content query embedding with the pixel embedding map. As DINO is a detection model for region-level regression, it is not designed for pixel-level alignment. To better align features between detection and segmentation, we also propose three key components to boost the segmentation performance. First, we propose a unified and enhanced query selection. It utilizes encoder dense prior by predicting masks from the top-ranked tokens to initialize mask queries as anchors. In addition, we observe that pixel-level segmentation is easier to learn in the early stage and propose to use initial masks to enhance boxes, which achieves task cooperation. Second, we propose a unified denoising training for masks to accelerate segmentation training. Third, we use a hybrid bipartite matching for more accurate and consistent matching from ground truth to both boxes and masks.", + "bbox": [ + 75, + 505, + 470, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mask DINO is conceptually simple and easy to implement under the DINO framework. To summarize, our contri", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "butions are three-fold. 1) We develop a unified Transformer-based framework for both object detection and segmentation. As the framework is extended from DINO, by adding a mask prediction branch, it naturally inherits most algorithm improvements in DINO including anchor box-guided cross attention, query selection, denoising training, and even a better representation pre-trained on a large-scale detection dataset. 2) We demonstrate that detection and segmentation can help each other through a shared architecture design and training method. Especially, detection can significantly help segmentation tasks, even for segmenting background \"stuff\" categories. Under the same setting with a ResNet-50 backbone, Mask DINO outperforms all existing models compared to DINO (+0.8 AP on COCO detection) and Mask2Former (+2.6 AP, +1.1 PQ, and +1.5 mIoU on COCO instance, COCO panoptic, and ADE20K semantic segmentation). 3) We also show that, via a unified framework, segmentation can benefit from detection pre-training on a large-scale detection dataset. After detection pre-training on the Objects365 [26] dataset with a SwinL [20] backbone, Mask DINO significantly improves all segmentation tasks and achieves the best results on instance (54.5 AP on COCO), panoptic (59.4 PQ on COCO), and semantic (60.8 mIoU on ADE20K) segmentation among models under one billion parameters.", + "bbox": [ + 496, + 90, + 893, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 482, + 640, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detection: Mainstream detection algorithms have been dominated by convolutional neural network-based frameworks, until recently Transformer-based detectors [1, 15, 18, 32] achieve great progress. DETR [1] is the first end-to-end and query-based Transformer object detector, which adopts a set-prediction objective with bipartite matching. DAB-DETR [18] improves DETR by formulating queries as 4D anchor boxes and refining predictions layer by layer. DN-DETR [15] introduces a denoising training method to accelerate convergence. Based on DAB-DETR and DN-DETR, DINO [32] proposes several new improvements on denoising and anchor refinement and achieves new SOTA results on COCO detection. Despite the inspiring progress, DETR-like detection models are not competitive for segmentation. Vanilla DETR incorporates a segmentation head in its architecture. However, its segmentation performance is inferior to specialized segmentation models and only shows the feasibility of DETR-like detection models to deal with detection and segmentation simultaneously.", + "bbox": [ + 496, + 507, + 893, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Segmentation: Segmentation mainly includes instance, semantic, and panoptic segmentation. Instance segmentation is to predict a mask and its corresponding category for each object instance. Semantic segmentation requires to classify each pixel including the background into different semantic categories. Panoptic segmentation [14] unifies the instance and semantic segmentation tasks and predicts a mask for", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3042", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "each object instance or background segment. In the past few years, researchers have developed specialized architectures for the three tasks. For example, Mask-RCNN [9] and HTC [2] can only deal with instance segmentation because they predict the mask of each instance based on its box prediction. FCN [21] and U-Net [25] can only perform semantic segmentation since they predict one segmentation map based on pixel-wise classification. Although models for panoptic segmentation [13, 30] unifies the above two tasks, they are usually inferior to specialized instance and semantic segmentation models. Until recently, some image segmentation models [3, 4, 33] are developed to unify the three tasks with a universal architecture. For instance, Mask2Former [3] improves MaskFormer [4] by introducing masked-attention to Transformer. Mask2Former has a similar architecture as DETR to probe image features with learnable queries but differs in using a different segmentation branch and some specialized designs for mask prediction. However, while Mask2Former shows a great success in unifying all segmentation tasks, it leaves object detection untouched and our empirical study shows that its specialized architecture design is not suitable for predicting boxes.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unified Methods: As both object detection and segmentation are concerned with localizing objects, they naturally share common model architectures and visual representations. A unified framework not only helps simplify the algorithm development effort, but also allows to use both detection and segmentation data to improve representation learning. There have been several previous works to unify segmentation and detection tasks, e.g., Mask RCNN [9], HTC [2], and DETR [1]. Mask RCNN extends Faster RCNN and pools image features from Region Of Interest (ROI) proposed by RPN. HTC further proposes an interleaved way of predicting boxes and masks. However, these two models can only perform instance segmentation. DETR predicts boxes and masks together in an end-to-end manner. However, its segmentation performance largely lags behind other models. According to Table 2, adding DETR's segmentation head to DINO results in inferior instance segmentation results. How to attain mutual assistance between segmentation and detection has long been an important problem to solve.", + "bbox": [ + 76, + 422, + 472, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Mask DINO", + "text_level": 1, + "bbox": [ + 76, + 734, + 205, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mask DINO is an extension of DINO [32]. On top of content query embeddings, DINO has two branches for box prediction and label prediction. The boxes are dynamically updated and used to guide the deformable attention in each Transformer decoder. Mask DINO adds another branch for mask prediction and minimally extends several key components in detection to fit segmentation tasks. To better understand Mask DINO, we start by briefly reviewing DINO and then introduce Mask DINO.", + "bbox": [ + 76, + 763, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: DINO", + "text_level": 1, + "bbox": [ + 500, + 90, + 699, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DINO is a typical DETR-like model, which is composed of a backbone, a Transformer encoder, and a Transformer decoder. The framework is shown in Fig. 1 (the blue-shaded part without red lines). Following DAB-DETR [18], DINO formulates each positional query in DETR as a 4D anchor box, which is dynamically updated through each decoder layer. Note that DINO uses multi-scale features with deformable attention [35]. Therefore, the updated anchor boxes are also used to constrain deformable attention in a sparse and soft way. Following DN-DETR [15], DINO adopts denoising training and further develops contrastive denoising to accelerate training convergence. Moreover, DINO proposes a mixed query selection scheme to initialize positional queries in the decoder and a look-forward-twice method to improve box gradient back-propagation.", + "bbox": [ + 496, + 113, + 893, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Why a universal model has not replaced the specialized models in DETR-like models?", + "text_level": 1, + "bbox": [ + 498, + 348, + 890, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remarkable progress has been achieved by Transformer-based detectors and segmentation models. For instance, DINO [32] and Mask2Former [3] have achieved the best results on COCO detection and panoptic segmentation, respectively. Inspired by such progress, we attempted to simply extend these specialized models for other tasks but found that the performance of other tasks lagged behind the original ones by a large margin, as shown in Table 1 and 2. It seems that trivial multi-task training even hurts the performance of the original task. However, in convolution-based models, it has shown effective and mutually beneficial to combine detection and instance segmentation tasks. For example, detection models with Mask R-CNN head [9] is still ranked the first on the COCO instance segmentation. We will take DINO and Mask2Former as examples to discuss the challenges in unifying Transformer-based detection and segmentation.", + "bbox": [ + 496, + 387, + 893, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- What are the differences between specialized detection and segmentation models? Image segmentation is a pixel-level classification task, while object detection is a region-level regression task. In DETR-based model, the decoder queries are responsible for these tasks. For example, Mask2Former uses such decoder queries to dot-product the high-resolution feature maps to produce segmentation masks, while DINO uses them to regress boxes. However, as such queries in Mask2Former only have to compare per-pixel similarity with the image features, they may not be aware of the region-level position of each instance. On the contrary, queries in DINO are not designed to interact with such low-level features to learn pixel-level representation. Instead, they encode rich positional information and high-level semantics for detection.", + "bbox": [ + 496, + 643, + 895, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "-Why cannot Mask2Former do detection well? The Transformer decoder of Mask2Former is designed for seg", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3043", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d9be5e4e3a24a9235b42c685a929e0a9837805dfe720b02c2dcbfd3f1fac9134.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBox APMask AP
Mask2Former46.2*43.7
Mask2Former + detection head21.641.3
DINO50.7-
", + "bbox": [ + 147, + 93, + 421, + 159 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/350866490077f04293ede00113e27fe7432e8c9aa13183aa389ef7e85f1fbe3f.jpg", + "table_caption": [ + "Table 1. Simply adding a detection head to Mask2Former results in low detection performance. * indicates the boxes are generated from the predicted masks. The generated boxes from Mask2Former are also inferior to DINO (-4.5 AP). The models are trained for 50 epochs." + ], + "table_footnote": [], + "table_body": "
ModelBox APMask AP
DINO + Mask2Former segmentation head49.940.2
DINO + DETR segmentation head (finetune DINO pretrained on COCO detection)-35.8
Mask2Former-43.7
", + "bbox": [ + 468, + 88, + 823, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Simply adding a segmentation head to DINO results in low instance segmentation performance. The predicted masks from DINO are also inferior to Mask2Former (-3.5 AP). The models are trained for 50 epochs.", + "bbox": [ + 452, + 171, + 836, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mentation tasks and does not suit detection for three reasons. First, its queries follow the design in DETR [1] without being able to utilize better positional priors as studied in Conditional DETR [22], Anchor DETR [29], and DAB-DETR [18]. For example, its content queries are semantically aligned with the features from the Transformer encoder, whereas its positional queries are just learnable vectors as in vanilla DETR instead of being associated with a single-mode position1. If we remove its mask branch, it reduces to a variant of DETR [1], whose performance is inferior to recently improved DETR models. Second, Mask2Former adopts masked attention (multi-head attention with attention mask) in Transformer decoders. The attention masks predicted from a previous layer are of high resolution and used as hard-constraints for attention computation. They are neither efficient nor flexible for box prediction. Third, Mask2Former cannot explicitly perform box refinement layer by layer. Moreover, its coarse-to-fine mask refinement in decoders fails to use multi-scale features from the encoder. As shown in Table 1, the generated box AP from mask is 4.5 AP worse than DINO and trivial multi-task learning by adding a detection head is not working2.", + "bbox": [ + 75, + 220, + 472, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Why cannot DETR/DINO do segmentation well? As shown in Table 2, simply 1) adding DETR's segmentation head or 2) adding Mask2Former's segmentation head result in inferior performance compared to Mask2Former. We analyze the reasons as follows. The reason for 1) is that DETR's segmentation head is not optimal. The vanilla DETR lets each query embedding dot-product with the smallest feature map to compute attention maps and then upsamples them to get the mask predictions. This design lacks an interaction between queries and larger feature maps from the backbone. In addition, the head is too heavy to use mask auxiliary loss for mask refinement. The reason for 2) is that features in improved detection models are not aligned with segmentation. For example, DINO inherits many designs from [18, 32, 35] like query formulation, denoising training, and query selection. However, these components are designed to strengthen region-level representation for detection, which is not optimal for segmentation.", + "bbox": [ + 75, + 551, + 472, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Our Method: Mask DINO", + "text_level": 1, + "bbox": [ + 500, + 220, + 740, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mask DINO adopts the same architecture design for detection as in DINO with minimal modifications. In the Transformer decoder, Mask DINO adds a mask branch for segmentation and extends several key components in DINO for segmentation tasks. As shown in Fig. 1, the framework in the blue-shaded part is the original DINO model and the additional design for segmentation is marked with red lines.", + "bbox": [ + 496, + 242, + 893, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Segmentation branch", + "text_level": 1, + "bbox": [ + 500, + 358, + 686, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following other unified models [3,4,28] for image segmentation, we perform mask classification for all segmentation tasks. Note that DINO is not designed for pixel-level alignment as its positional queries are formulated as anchor boxes and its content queries are used to predict box offset and class membership. To perform mask classification, we adopt a key idea from Mask2Former [3] to construct a pixel embedding map which is obtained from the backbone and Transformer encoder features. As shown in Fig. 1, the pixel embedding map is obtained by fusing the $1/4$ resolution feature map $C_b$ from the backbone with an upsampled $1/8$ resolution feature map $C_e$ from the Transformer encoder. Then we dot-product each content query embedding $q_c$ from the decoder with the pixel embedding map to obtain an output mask $m$ .", + "bbox": [ + 496, + 380, + 893, + 606 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm = q _ {c} \\otimes \\mathcal {M} \\left(\\mathcal {T} \\left(C _ {b}\\right) + \\mathcal {F} \\left(C _ {e}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 616, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is the segmentation head, $\\mathcal{T}$ is a convolutional layer to map the channel dimension to the Transformer hidden dimension, and $\\mathcal{F}$ is a simple interpolation function to perform 2x upsampling of $C_e$ . This segmentation branch is conceptually simple and easy to implement in the DINO framework, as shown in Fig. 1.", + "bbox": [ + 496, + 642, + 893, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Unified and Enhanced Query Selection", + "text_level": 1, + "bbox": [ + 500, + 741, + 833, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unified query selection for mask: Query selection has been widely used in traditional two-stage models [24] and many DETR-like models [32, 35] to improve detection performance. We further improve the query selection scheme in Mask DINO for segmentation tasks.", + "bbox": [ + 496, + 763, + 893, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The encoder output features contain dense features, which can serve as better priors for the decoder. Therefore, we adopt three prediction heads (classification, detection, and segmentation) in the encoder output. Note that the three", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We refer the interested readers to discussions in Sec. 3 in DAB-DETR [18]", + "bbox": [ + 75, + 838, + 468, + 862 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2We also notice there are issues in official Mask2Former Github (https://github.com/facebookresearch/Mask2Former/issues/43) that fail to make Mask2Former work well by adding a detection head.", + "bbox": [ + 75, + 862, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3044", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ef954800f9933bcb1adf83ab89f27b43762ee8de0f7e3506e627564632a6095c.jpg", + "image_caption": [ + "Figure 1. The framework of Mask DINO, which is based on DINO (the blue-shaded part) with extensions (the red part) for segmentation tasks. 'QS' and 'DN' are short for query selection and denoising training, respectively." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 854, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "heads are identical to the decoder heads. The classification score of each token is considered as the confidence to select top-ranked features and feed them to the decoder as content queries. The selected features also regress boxes and dot-product with the high-resolution feature map to predict masks. The predicted boxes and masks will be supervised by the ground truth and are considered as initial anchors for the decoder. Note that we initialize both the content and anchor box queries in Mask DINO whereas DINO only initializes anchor box queries.", + "bbox": [ + 75, + 400, + 472, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask-enhanced anchor box initialization: As summarized in Sec 3.2, image segmentation is a pixel-level classification task while object detection is a region-level position regression task. Therefore, compared to detection, though segmentation is a more difficult task with fine-granularity, it is easier to learn in the initial stage. For example, masks are predicted by dot-producing queries with the high-resolution feature map, which only needs to compare per-pixel semantic similarity. However, detection requires to directly regress the box coordinates in an image. Therefore, in the initial stage after unified query selection, mask prediction is much more accurate than box (the qualitative AP comparison between mask prediction and box prediction in different stages is also shown in Table 8 and 9). Therefore, after unified query selection, we derive boxes from the predicted masks as better anchor box initialization for the decoder. By this effective task cooperation, the enhanced box initialization can bring in a large improvement to the detection performance.", + "bbox": [ + 75, + 551, + 472, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Segmentation Micro Design", + "text_level": 1, + "bbox": [ + 76, + 832, + 326, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unified denoising for mask: Query denoising in object detection has shown effective [15,32] to accelerate convergence and improve performance. It adds noises to ground-", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "truth boxes and labels and feed them to the Transformer decoder as noised positional queries and content queries. The model is trained to reconstruct ground truth objects given their noised versions. We also extend this technique to segmentation tasks. As masks can be viewed as a more fine-grained representation of boxes, box and mask are naturally connected. Therefore, we can treat boxes as a noised version of masks, and train the model to predict masks given boxes as a denoising task. The given boxes for mask prediction are also randomly noised for more efficient mask denoising training. The detailed noise and its hyperparameters used in our model are shown in Appendix ??", + "bbox": [ + 496, + 400, + 893, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hybrid matching: Mask DINO, as in some traditional models [2, 9], predicts boxes and masks with two parallel heads in a loosely coupled manner. Hence the two heads can predict a pair of box and mask that are inconsistent with each other. To address this issue, in addition to the original box and classification loss in bipartite matching, we add a mask prediction loss to encourage more accurate and consistent matching results for one query. Therefore, the matching cost becomes $\\lambda_{cls}\\mathcal{L}_{cls} + \\lambda_{box}\\mathcal{L}_{box} + \\lambda_{mask}\\mathcal{L}_{mask}$ , where $\\mathcal{L}_{cls},\\mathcal{L}_{box}$ , and $\\mathcal{L}_{mask}$ are the classification, box, and mask loss and $\\lambda$ are their corresponding weights. The detailed losses used in our model and their corresponding weights are shown in Appendix ??", + "bbox": [ + 496, + 582, + 893, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 797, + 633, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct extensive experiments and compare with several specialized models for four popular tasks including object detection, instance, panoptic, and semantic segmentation on COCO [17], ADE20K [34], and Cityscapes [5]. For all experiments, we use batch size 16 and A100 GPUs with", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3045", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8636c63def5ecaf9a050d0cacb9b3073b62e569e5edfc4b702d15d2108af365d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelEpochsQuery typeMask APBox AP\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP_{S}^{mask}\\)\\(AP_{M}^{mask}\\)\\(AP_{L}^{mask}\\)GFLOPSParamsFPS
ResNet-50 backbone
Mask-RCNN [6,8,9]400Dense anchors42.548.2--23.845.060.020740M10.3
HTC [2]36Dense anchors39.744.961.443.122.642.250.644180M5
QueryInst [7]36300 queries40.645.663.044.023.442.552.8--7.0
DINO-4scale [32]36900 queries-50.9-----24547M19.6
Mask2Former [3]12100 queries38.7-59.841.218.241.559.822644M8.2
Mask DINO (ours)12300 queries41.4(+2.7)45.762.944.621.144.261.428652M14.8
Mask DINO (ours)24300 queries44.2(+0.5)48.466.647.923.947.064.028652M14.8
Mask2Former* [3]50100 queries43.7\\(46.2^{\\dagger}\\)66.046.923.447.264.822644M8.2
Mask DINO (ours)50100 queries45.449.867.949.325.248.365.828052M15.2
Mask DINO (ours)50300 queries46.050.568.950.326.0(+2.6)49.3(+2.1)65.5(+0.7)28652M14.8
Mask \\(DINO^‡\\) (ours)50300 queries46.3(+2.6)51.7(+0.8)69.050.726.1(+2.7)49.3(+2.1)66.1(+1.3)28652M14.2
SwinL backbone
HTC++ [2,20]72dense anchors49.557.1--31.052.467.21470284M-
Mask2Former [3]100200 queries50.1---29.953.972.1868216M4.0
DINO [32]36900 queries-58.5-----1285217M8.1
Mask2Former100200 queries50.1---29.953.972.1868216M4.0
Mask DINO (ours)50300 queries52.158.376.557.632.955.472.51326223M6.1
Mask \\(DINO^‡\\) (ours)50300 queries52.3(+2.2)59.0(+0.5)76.657.833.155.472.61326223M5.6
", + "bbox": [ + 122, + 88, + 848, + 311 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6dfc1b95729b2f564eca45f856575c606f43641026fd22504c634a0adfe661f6.jpg", + "table_caption": [ + "Table 3. Results for Mask DINO and other object detection and instance segmentation models with ResNet-50 and SwinL backbone on COCO val2017 without extra data or tricks. Following DINO [32], we use ResNet-50 with four feature scales by default, and use five scales under large models with a SwinL backbone. We follow the common practice in DETR-like models to use 300 queries. * Mask2Former using 300 queries is not listed as its performance will degenerate when using 300 queries. † indicates the box AP is derived from mask prediction. ‡ we use the proposed mask-enhanced box initialization to further improve detection performance. We test the FPS and GFLOPS of Mask2Former and Mask DINO on the A100 GPU using detectron2." + ], + "table_footnote": [], + "table_body": "
ModelEpochsQuery typePQ\\( PQ^{Th} \\)\\( PQ^{St} \\)Box \\( AP^{Th}_{pan} \\)Mask \\( AP^{Th}_{pan} \\)
ResNet-50 backbone
DETR [1]500 + 25100 queries43.448.236-31.1
Panoptic Segformer [16]24353 queries49.654.442.4-41.7
Mask2Former* [3]50100 queries\\( 51.9/51.5^† \\)57.743.0-41.7
Mask DINO (ours)50100 queries52.358.343.247.743.7
Mask DINO (ours)50300 queries\\( 53.0(+1.1) \\)\\( 59.1(+1.4) \\)\\( 43.9(+0.9) \\)48.8\\( 44.3(+2.6) \\)
Mask DINO (ours)24300 queries51.557.342.646.442.8
Mask2Former [3]12100 queries46.952.538.4-37.2
Panoptic Segformer [16]12353 queries48.052.341.5--
Mask DINO (ours)12300 queries\\( 49.0(+1.0) \\)54.840.243.2\\( 40.4(+3.2) \\)
SwinL backbone
Mask2Former [3]100100 queries57.864.248.1-48.6
OneFormer [11]100150 queries57.964.448.0-49.0
Mask DINO (ours)50300 queries\\( 58.3(+0.5) \\)65.148.056.2\\( 50.6(+2.0) \\)
", + "bbox": [ + 199, + 419, + 772, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Results for Mask DINO and other panoptic segmentation models with a ResNet-50 backbone on COCO val2017. * Mask2Former using 300 queries is not listed as its performance will degenerate when using 300 queries. † Our reproduced result.", + "bbox": [ + 75, + 609, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "40GB memory. We use a ResNet-50 [10] and a SwinL [20] backbone for our main results and SOTA model. Under ResNet-50, we use 4 A100 GPUs for all tasks without extra data. The implementation details are in Appendix ??", + "bbox": [ + 75, + 650, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 76, + 724, + 215, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instance segmentation and object detection. In Table 3, we compare Mask DINO with other instance segmentation and object detection models. Mask DINO outperforms both the specialized models such as Mask2Former [3] and DINO [32] and hybrid models such as HTC [2] under the same setting. Especially, the instance segmentation results surpass the strong baseline Mask2Former by a large margin (+2.7 AP and +2.6 AP) on the 12-epoch and 50-epoch settings. Moreover, Mask DINO significantly improves the convergence speed, outperforming Mask2Former with less", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "than half training epochs (44.2 AP in 24 epochs). In addition, after using mask-enhanced box initialization, our detection performance has been significantly improved (+1.2 AP), which even outperforms DINO by 0.8 AP. These results indicate that task unification is beneficial. Without bells and whistles, we achieve the best detection and instance segmentation performance among DETR-like model with a SwinL backbone without extra data.", + "bbox": [ + 496, + 650, + 893, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Panoptic segmentation. We compare Mask DINO with other models in Table 4. Mask DINO outperforms all previous best models on both the 12-epoch and 50-epoch settings by 1.0 PQ and 1.1 PQ, respectively. This indicates Mask DINO has the advantages of both faster convergence and superior performance. One interesting observation is that we outperform Mask2Former [3] in terms of both $PQ^{Th}$ and $PQ^{St}$ . However, instead of using dense and hard", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3046", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/db08a011c771bdfb84cc85e305280ff71b5ca434a68a05b269044f82a36a7d39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelIterationsCrop sizemIoU (mean)mIoU (high)mIoU (reported)
Mask2Former [3]160k51246.146.547.2
Mask DINO (ours)160k51247.7(+1.6)48.7(+2.2)48.7(+1.6)
", + "bbox": [ + 117, + 88, + 480, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c6f956a0d2c5dc667cddd5f7225928d7a612972665f9994b6d7ff297119151e.jpg", + "table_caption": [ + "Table 5. Results for Mask DINO and Mask2Former with 100 queries using a ResNet-50 backbone on ADE20K val. We found the performance variance on this dataset is high and run three times to report both the mean and highest results for both models." + ], + "table_footnote": [], + "table_body": "
ModelIterationsmIoU (mean)mIoU (high)mIoU (reported)
Mask2Former [3]90k78.779.079.4
Mask DINO (ours)90k79.8(+1.1)80.0(+1.0)80.0(+0.6)
", + "bbox": [ + 509, + 89, + 836, + 142 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c0b7d5f4eb49cf641ea1dcd0f4f0051e61c77fc46a3242cafdd799942445564.jpg", + "table_caption": [ + "Table 6. Results for Mask DINO and Mask2Former with 100 queries using a ResNet-50 backbone on Cityscapes val. We found the performance variance on this dataset is high and run three times to report both the mean and highest results for both models." + ], + "table_footnote": [], + "table_body": "
MethodParamsBackboneBackbone Pre-training DatasetDetection Pre-training Datasetval
w/o TTAw/ TTA
Instance segmentation on COCOAP
Mask2Former [3]216MSwinLIN-22K-14M-50.1-
Soft Teacher [31]284MSwinLIN-22K-14MO36551.952.5
SwinV2-G-HTC++ [19]3.0BSwinV2-GIN-22K-ext-70M [19]O36553.453.7
MasK DINO(Ours)223MSwinLIN-22K-14M-52.6-
MasK DINO(Ours)223MSwinLIN-22K-14MO36554.5(+1.1)-
Panoptic segmentation on COCOPQ
Panoptic SegFormer [16]-MSwinLIN-22K-14M-55.8-
Mask2Former [3]216MSwinLIN-22K-14M-57.8-
MasK DINO (ours)223MSwinLIN-22K-14M58.4(+0.6)-
MasK DINO (ours)223MSwinLIN-22K-14MO36559.4(+1.6)-
Semantic segmentation on ADE20KmIoU
Mask2Former [3]215MSwinLIN-22K-14M-56.157.3
SeMask-L MSFaPN-Mask2Former [12]-MSwinL-FaPNIN-22K-14M--58.2
SwinV2-G-UperNet [19]3.0BSwinV2-GIN-22K-ext-70M [19]-59.359.9
MasK DINO (ours)223MSwinLIN-22K-14M-56.6-
MasK DINO (ours)223MSwinLIN-22K-14MO36559.560.8(+0.9)
", + "bbox": [ + 199, + 205, + 772, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Comparison of the SOTA models on three segmentation tasks. Mask DINO outperforms all existing models. \"TTA\" means test-time-augmentation. \"O365\" denotes the Objects365 [26] dataset.", + "bbox": [ + 75, + 492, + 892, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "constrained masked attention, we predict boxes and then use them in deformable attention to extract query features. Therefore, our box-oriented deformable attention also works well with \"stuff\" categories, which makes our unified model simple and efficient. In addition, we improve the mask $\\mathrm{AP}_{pan}^{Th}$ by 2.6 to 44.3 AP, which is 0.6 higher than the specialized instance segmentation model Mask2Fomer (43.7 AP).", + "bbox": [ + 75, + 537, + 470, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic segmentation. In Table 5 and 6, we show the performance of semantic segmentation with a ResNet-50 backbone. We use 100 queries for these small datasets. We outperform Mask2Former on both ADE20K and Cityscapes by 1.6 and $0.6\\mathrm{mIoU}$ on the reported performance.", + "bbox": [ + 75, + 657, + 468, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTA Models", + "text_level": 1, + "bbox": [ + 76, + 741, + 359, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 7, we compare Mask DINO with SOTA models on three image segmentation tasks to show its scalability. We use the SwinL [20] backbone and pre-train DINO on the Objects365 [26] detection dataset. Even without using extra data, we outperform Mask2Former on all three tasks, especially on instance segmentation (+2.5 AP). As Mask DINO is an extension of DINO, the pre-trained DINO model can be used to fine-tune Mask DINO for segmentation tasks. After fine-tuning Mask DINO on the corresponding tasks, we", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieve the best results on instance (54.5 AP), panoptic (59.4 PQ), and semantic (60.8 mIoU) segmentation among model under one billion parameters. Compared to SwinV2-G [19], we significantly reduce the model size to 1/15 and backbone pre-training dataset to 1/5. Our detection pre-training also significantly helps all segmentation tasks including panoptic and semantic with \"stuff\" categories. However, previous specialized segmentation models such as Mask2Former can not use detection datasets and adding a detection head to it results in poor performance as shown in Table 1, which severely limits the data scalability. By unifying four tasks in one model, we only need to pre-train one model on a large-scale dataset and finetune on all tasks for 10 to 20 epochs (Mask2Former needs 100 epochs), which is more computationally efficient and simpler in model design.", + "bbox": [ + 496, + 537, + 892, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 771, + 663, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies using a ResNet-50 backbone to analyze Mask DINO on COCO val2017. Unless otherwise stated, our experiments are based on object detection and instance segmentation without Mask-enhanced anchor box initialization.", + "bbox": [ + 496, + 794, + 893, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query selection. Table 8 shows the results of our query selection for instance segmentation, where we additionally pro", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3047", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd8a4c040d48529c57317dba8b9f9adf9d44e5222bf2fb0355a9645b4a202eaf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
test layer#Mask DINOMask2Former
layer 039.6(+38.5)1.1
layer 344.042.3
layer 645.943.3
layer 946.043.7
", + "bbox": [ + 129, + 88, + 344, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Effectiveness of our query selection for mask initialization. We evaluate the instance segmentation performance from different decoder layers in the same model after training for 50 epochs.", + "bbox": [ + 76, + 164, + 395, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df21cf8bd00fb1e913ef5f86d72119d108728aa3d41c9ad77603341fdaaa1cae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TasksBox APMask AP
BoxMask
✓50.1-
✓-43.3
✓✓50.5(+0.4)46.0(+2.7)
", + "bbox": [ + 114, + 200, + 290, + 260 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e913167538067bb25a8d21539de144039e328b38d8cfd246bc9b3a9ab741f7ca.jpg", + "table_caption": [ + "Table 11. Task comparison under the 50-epoch setting. We train the same Mask DINO with different tasks and validate that box and mask can achieve mutual cooperation." + ], + "table_footnote": [], + "table_body": "
layer#w/o MEw ME
BoxMaskBoxMask
layer 039.625.639.841.2(+15.6)
layer 946.050.546.351.7(+1.2)
", + "bbox": [ + 379, + 88, + 620, + 143 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f046fac77b53d44657d174ebb9b86970272932ee4155320609e26745a8c6737b.jpg", + "table_caption": [ + "Table 9. Comparison of our model with and without Mask-enhanced anchor box initialization (ME). ME enhances anchor box initialization and improves final detection performance. Trained for 50 epochs." + ], + "table_footnote": [], + "table_body": "
Decoder layer#Box APMask AP
343.140.7
644.341.1
944.541.4
1244.841.1
", + "bbox": [ + 379, + 198, + 563, + 260 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/adbc7a911ab588fc4bb41c2a18aaf1b070f9cbe630c3a482ec173e16c3198c72.jpg", + "table_caption": [ + "Table 12. Decoder layer number comparison under the 12-epoch setting. Mask DINO benefits from more decoders, while DINO's performance will decrease with 9 decoders." + ], + "table_footnote": [], + "table_body": "
Feature scalebox APmask AP
single scale(1/8)45.845.1
3 scales50.545.8
4 scales50.546.0
", + "bbox": [ + 661, + 88, + 859, + 143 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ea1284595cd0cc00b2e7d7b2d63198d23dfb49e0ae4237319d7b6727849081b.jpg", + "table_caption": [ + "Table 10. Comparison of multi-scale features for Transformer decoder under the 50-epoch setting. Both detection and segmentation benefit from more feature scales." + ], + "table_footnote": [], + "table_body": "
MatchingBox APMask AP
BoxMask
✓44.440.5
✓40.238.4
✓✓44.541.4
", + "bbox": [ + 658, + 194, + 831, + 255 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/19d138bed3ac5fc7828ef377966270fb9383b9e5134d012bb232d209d15e2a85.jpg", + "table_caption": [ + "Table 13. Matching method comparison under the 12-epoch setting. We train both tasks together but use different matching methods to verify the effectiveness of hybrid matching." + ], + "table_footnote": [], + "table_body": "
Box APMask AP
Mask DINO (ours)45.741.4
- Mask-enhanced anchor box initialization44.5(-1.2)41.4
- Unified query selection for masks43.6(-2.1)40.3(-1.1)
- Unified denoising for masks44.4(-1.3)40.3(-1.1)
- Hybrid matching44.9(-0.8)40.5(-0.9)
- remove all the above41.7(-4.0)38.5(-2.7)
", + "bbox": [ + 109, + 305, + 437, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 14. Effectiveness of the proposed components under the 12-epoch setting.", + "bbox": [ + 107, + 406, + 437, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vide the performance of different decoder layers in one single model. Mask2Former also predicts the masks of learnable queries as initial region proposals. However, their performance lags behind Mask DINO by a large margin (-38.5AP). With our effective query selection scheme, the mask performance achieves 39.6 AP without using the decoder. In addition, our mask performance at layer six is already comparable to the final results with 9 layers. In Table 9, we show that in query selection the predicted box is inferior to mask, which indicates segmentation is easier to learn in the initial stage. Therefore, our proposed mask-enhanced box initialization enhances boxes with masks in query selection to provide better anchor boxes (+15.6 AP) for the decoder, which results in +1.2 AP improvement in the final detection performance.", + "bbox": [ + 75, + 431, + 472, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature scales. Mask2Former [3] shows that concatenating multi-scale features as input to Transformer decoder layers does not improve the segmentation performance. However, in Table 10, Mask DINO shows that using more feature scales in the decoder consistently improves the performance.", + "bbox": [ + 75, + 657, + 470, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object detection and segmentation help each other. To validate task cooperation in Mask DINO, we use the same model but train different tasks and report the 12 epoch and 50 epoch results. As shown in Table 11, only training one task will lead to a performance drop. Although only training object detection results in faster convergence in the early stage for box prediction, the final performance is still inferior to training both tasks together.", + "bbox": [ + 75, + 734, + 470, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decoder layer number. In DINO, increasing the decoder layer number to nine will decrease the performance of box. In Table 12, the result indicates that increasing the number of", + "bbox": [ + 75, + 854, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "decoder layers will contribute to both detection and segmentation in Mask DINO. We hypothesize that the multi-task training become more complex and require more decoders to learn the needed mapping function.", + "bbox": [ + 496, + 308, + 893, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Matching. In Table 13, we show that only using boxes or masks to perform bipartite matching is not optimal in Mask DINO. A unified matching objective makes the optimization more consistent. In addition, after removing all the proposed components, both detection and segmentation performance drop by a large margin. This result indicates that if we trivially add detection and segmentation tasks in one DETR-based model, the features are not aligned for detection and segmentation tasks to achieve mutual cooperation.", + "bbox": [ + 496, + 368, + 893, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also present visualization analysis in Appendix ??", + "bbox": [ + 519, + 508, + 880, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 553, + 619, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented Mask DINO as a unified Transformer-based framework for both object detection and image segmentation. Conceptually, Mask DINO is a natural extension of DINO from detection to segmentation with minimal modifications on some key components. Mask DINO outperforms previous specialized models and achieves the best results on all three segmentation tasks (instance, panoptic, and semantic) among models under one billion parameters. Moreover, Mask DINO shows that detection and segmentation can help each other in query-based models. In particular, Mask DINO enables semantic and panoptic segmentation to benefit from a better visual representation pre-trained on a large-scale detection dataset. We hope Mask DINO can provide insights for enabling task cooperation and data cooperation towards designing a universal model for more vision tasks.", + "bbox": [ + 496, + 583, + 893, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: Different segmentation tasks fail to achieve mutual assistance in Mask DINO in COCO panoptic segmentation. For example, in COCO panoptic segmentation, the mask AP still lags behind the model only trained with instances.", + "bbox": [ + 496, + 824, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3048", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3, 4, 6", + "[2] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 1, 2, 3, 5, 6", + "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. 2022. 1, 2, 3, 4, 6, 7, 8", + "[4] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. 2021. 1, 2, 3, 4", + "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5", + "[6] Xianzhi Du, Barret Zoph, Wei-Chih Hung, and Tsung-Yi Lin. Simple training strategies and model scaling for object detection. arXiv preprint arXiv:2107.00057, 2021. 6", + "[7] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 1, 6", + "[8] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 6", + "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2, 3, 5, 6", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[11] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. arXiv preprint arXiv:2211.06220, 2022. 6", + "[12] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. SeMask: Semantically Masked Transformers for Semantic Segmentation. arXiv preprint arXiv:2112.12782, 2021. 1, 7", + "[13] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6399-6408, 2019. 3", + "[14] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 2", + "[15] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. DN-DETR: Accelerate DETR Training by Introducing Query DeNoising. arXiv preprint arXiv:2203.01305, 2022. 1, 2, 3, 5", + "[16] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Tong Lu, and Ping Luo. Panoptic SegFormer. arXiv preprint arXiv:2109.03814, 2021. 6, 7", + "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5", + "[18] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 3, 4", + "[19] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin Transformer V2: Scaling Up Capacity and Resolution. arXiv preprint arXiv:2111.09883, 2021. 7", + "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 6, 7", + "[21] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 3", + "[22] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional DETR for Fast Training Convergence. arXiv preprint arXiv:2108.06152, 2021. 4", + "[23] Zipeng Qin, Jianbo Liu, Xiaolin Zhang, Maoqing Tian, Aojun Zhou, Shuai Yi, and Hongsheng Li. Pyramid Fusion Transformer for Semantic Segmentation. arXiv preprint arXiv:2201.04019, 2022. 1", + "[24] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 4", + "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3", + "[26] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 7" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3049", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1", + "[28] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2, 4", + "[29] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. arXiv preprint arXiv:2109.07107, 2021. 4", + "[30] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8826, 2019. 3", + "[31] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 7", + "[32] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. arXiv preprint arXiv:2203.03605, 2022. 1, 2, 3, 4, 5, 6", + "[33] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34, 2021. 1, 3", + "[34] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5", + "[35] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR 2021: The Ninth International Conference on Learning Representations, 2021. 1, 3, 4" + ], + "bbox": [ + 78, + 90, + 470, + 641 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3050", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_model.json b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cb480735b3cc41c17ac05fd31973ec7cc56c7c7a --- /dev/null +++ b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_model.json @@ -0,0 +1,1826 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.203, + 0.761, + 0.24 + ], + "angle": 0, + "content": "Feng Li\\(^{1,3*†}\\), Hao Zhang\\(^{1,3*†}\\), Huaizhe Xu\\(^{1,3}\\), Shilong Liu\\(^{2,3}\\), Lei Zhang\\(^{3‡}\\), Lionel M. Ni\\(^{1,4}\\), Heung-Yeung Shum\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.24, + 0.707, + 0.257 + ], + "angle": 0, + "content": "1The Hong Kong University of Science and Technology." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.257, + 0.758, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Dept. of CST., BNRist Center, Institute for AI, Tsinghua University." + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.275, + 0.684, + 0.292 + ], + "angle": 0, + "content": "\\(^{3}\\)International Digital Economy Academy (IDEA)." + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.293, + 0.761, + 0.31 + ], + "angle": 0, + "content": "4The Hong Kong University of Science and Technology (Guangzhou)." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.323, + 0.763, + 0.334 + ], + "angle": 0, + "content": "{fiay,hzhangcx,hxubr}@connect.ust.hk {liusl20}@mails.tsinghua.edu.cn {leizhang}@idea.edu.cn {ni,hshum}@ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.37, + 0.314, + 0.387 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.402, + 0.474, + 0.719 + ], + "angle": 0, + "content": "In this paper we present Mask DINO, a unified object detection and segmentation framework. Mask DINO extends DINO (DETR with Improved Denoising Anchor Boxes) by adding a mask prediction branch which supports all image segmentation tasks (instance, panoptic, and semantic). It makes use of the query embeddings from DINO to dot-product a high-resolution pixel embedding map to predict a set of binary masks. Some key components in DINO are extended for segmentation through a shared architecture and training process. Mask DINO is simple, efficient, and scalable, and it can benefit from joint large-scale detection and segmentation datasets. Our experiments show that Mask DINO significantly outperforms all existing specialized segmentation methods, both on a ResNet-50 backbone and a pre-trained model with SwinL backbone. Notably, Mask DINO establishes the best results to date on instance segmentation (54.5 AP on COCO), panoptic segmentation (59.4 PQ on COCO), and semantic segmentation (60.8 mIoU on ADE20K) among models under one billion parameters. Code is available at https://github.com/IDEA-Research/MaskDINO." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.75, + 0.21, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.776, + 0.471, + 0.853 + ], + "angle": 0, + "content": "Object detection and image segmentation are fundamental tasks in computer vision. Both tasks are concerned with localizing objects of interest in an image but have different levels of focus. Object detection is to localize objects of interest and predict their bounding boxes and category" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.371, + 0.895, + 0.462 + ], + "angle": 0, + "content": "labels, whereas image segmentation focuses on pixel-level grouping of different semantics. Moreover, image segmentation encompasses various tasks including instance segmentation, panoptic segmentation, and semantic segmentation with respect to different semantics, e.g., instance or category membership, foreground or background category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.896, + 0.674 + ], + "angle": 0, + "content": "Remarkable progress has been achieved by classical convolution-based algorithms developed for these tasks with specialized architectures, such as Faster RCNN [24] for object detection, Mask RCNN [9] for instance segmentation, and FCN [21] for semantic segmentation. Although these methods are conceptually simple and effective, they are tailored for specialized tasks and lack the generalization ability to address other tasks. The ambition to bridge different tasks gives rise to more advanced methods like HTC [2] for object detection and instance segmentation and Panoptic FPN [14], K-net [33] for instance, panoptic, and semantic segmentation. Task unification not only helps simplify algorithm development but also brings in performance improvement in multiple tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.897, + 0.87 + ], + "angle": 0, + "content": "Recently, DETR-like [1] models developed based on Transformers [27] have achieved inspiring progress on many detection and segmentation tasks. As an end-to-end object detector, DETR adopts a set-prediction objective and eliminates hand-crafted modules such as anchor design and nonmaximum suppression. Although DETR addresses both the object detection and panoptic segmentation tasks, its segmentation performance is still inferior to classical segmentation models. To improve the detection and segmentation performance of Transformer-based models, researchers have developed specialized models for object detection [15, 18, 32, 35], image segmentation [3, 4, 33], instance segmentation [7], panoptic segmentation [23], and semantic segmentation [12]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Among the efforts to improve object detection, DINO [32] takes advantage of the dynamic anchor box for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.207, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.44, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done when Feng Li and Hao Zhang were interns at IDEA." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.863, + 0.44, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3041" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.226 + ], + "angle": 0, + "content": "mulation from DAB-DETR [18] and query denoising training from DN-DETR [15], and further achieves the SOTA result on the COCO object detection leaderboard for the first time as a DETR-like model. Similarly, for improving image segmentation, MaskFormer [4] and Mask2Former [3] propose to unify different image segmentation tasks using query-based Transformer architectures to perform mask classification. Such methods have achieved remarkable performance improvement on multiple segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.471, + 0.502 + ], + "angle": 0, + "content": "However, in Transformer-based models, the best-performing detection and segmentation models are still not unified, which prevents task and data cooperation between detection and segmentation tasks. As an evidence, in CNN-based models, Mask-R-CNN [9] and HTC [2] are still widely acknowledged as unified models that achieve mutual cooperation between detection and segmentation to achieve superior performance than specialized models. Though we believe detection and segmentation can help each other in a unified architecture in Transformer-based models, the results of simply using DINO for segmentation and using Mask2Former for detection indicate that they can not do other tasks well, as shown in Table 1 and 2. Moreover, trivial multi-task training can even hurt the performance of the original tasks. It naturally leads to two questions: 1) why cannot detection and segmentation tasks help each other in Transformer-based models? and 2) is it possible to develop a unified architecture to replace specialized ones?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.471, + 0.867 + ], + "angle": 0, + "content": "To address these problems, we propose Mask DINO, which extends DINO with a mask prediction branch in parallel with DINO's box prediction branch. Inspired by other unified models [3, 4, 28] for image segmentation, we reuse content query embeddings from DINO to perform mask classification for all segmentation tasks on a high-resolution pixel embedding map (1/4 of the input image resolution) obtained from the backbone and Transformer encoder features. The mask branch predicts binary masks by simply dot-producting each content query embedding with the pixel embedding map. As DINO is a detection model for region-level regression, it is not designed for pixel-level alignment. To better align features between detection and segmentation, we also propose three key components to boost the segmentation performance. First, we propose a unified and enhanced query selection. It utilizes encoder dense prior by predicting masks from the top-ranked tokens to initialize mask queries as anchors. In addition, we observe that pixel-level segmentation is easier to learn in the early stage and propose to use initial masks to enhance boxes, which achieves task cooperation. Second, we propose a unified denoising training for masks to accelerate segmentation training. Third, we use a hybrid bipartite matching for more accurate and consistent matching from ground truth to both boxes and masks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Mask DINO is conceptually simple and easy to implement under the DINO framework. To summarize, our contri" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.469 + ], + "angle": 0, + "content": "butions are three-fold. 1) We develop a unified Transformer-based framework for both object detection and segmentation. As the framework is extended from DINO, by adding a mask prediction branch, it naturally inherits most algorithm improvements in DINO including anchor box-guided cross attention, query selection, denoising training, and even a better representation pre-trained on a large-scale detection dataset. 2) We demonstrate that detection and segmentation can help each other through a shared architecture design and training method. Especially, detection can significantly help segmentation tasks, even for segmenting background \"stuff\" categories. Under the same setting with a ResNet-50 backbone, Mask DINO outperforms all existing models compared to DINO (+0.8 AP on COCO detection) and Mask2Former (+2.6 AP, +1.1 PQ, and +1.5 mIoU on COCO instance, COCO panoptic, and ADE20K semantic segmentation). 3) We also show that, via a unified framework, segmentation can benefit from detection pre-training on a large-scale detection dataset. After detection pre-training on the Objects365 [26] dataset with a SwinL [20] backbone, Mask DINO significantly improves all segmentation tasks and achieves the best results on instance (54.5 AP on COCO), panoptic (59.4 PQ on COCO), and semantic (60.8 mIoU on ADE20K) segmentation among models under one billion parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.642, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.795 + ], + "angle": 0, + "content": "Detection: Mainstream detection algorithms have been dominated by convolutional neural network-based frameworks, until recently Transformer-based detectors [1, 15, 18, 32] achieve great progress. DETR [1] is the first end-to-end and query-based Transformer object detector, which adopts a set-prediction objective with bipartite matching. DAB-DETR [18] improves DETR by formulating queries as 4D anchor boxes and refining predictions layer by layer. DN-DETR [15] introduces a denoising training method to accelerate convergence. Based on DAB-DETR and DN-DETR, DINO [32] proposes several new improvements on denoising and anchor refinement and achieves new SOTA results on COCO detection. Despite the inspiring progress, DETR-like detection models are not competitive for segmentation. Vanilla DETR incorporates a segmentation head in its architecture. However, its segmentation performance is inferior to specialized segmentation models and only shows the feasibility of DETR-like detection models to deal with detection and segmentation simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Segmentation: Segmentation mainly includes instance, semantic, and panoptic segmentation. Instance segmentation is to predict a mask and its corresponding category for each object instance. Semantic segmentation requires to classify each pixel including the background into different semantic categories. Panoptic segmentation [14] unifies the instance and semantic segmentation tasks and predicts a mask for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3042" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "each object instance or background segment. In the past few years, researchers have developed specialized architectures for the three tasks. For example, Mask-RCNN [9] and HTC [2] can only deal with instance segmentation because they predict the mask of each instance based on its box prediction. FCN [21] and U-Net [25] can only perform semantic segmentation since they predict one segmentation map based on pixel-wise classification. Although models for panoptic segmentation [13, 30] unifies the above two tasks, they are usually inferior to specialized instance and semantic segmentation models. Until recently, some image segmentation models [3, 4, 33] are developed to unify the three tasks with a universal architecture. For instance, Mask2Former [3] improves MaskFormer [4] by introducing masked-attention to Transformer. Mask2Former has a similar architecture as DETR to probe image features with learnable queries but differs in using a different segmentation branch and some specialized designs for mask prediction. However, while Mask2Former shows a great success in unifying all segmentation tasks, it leaves object detection untouched and our empirical study shows that its specialized architecture design is not suitable for predicting boxes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.424, + 0.473, + 0.711 + ], + "angle": 0, + "content": "Unified Methods: As both object detection and segmentation are concerned with localizing objects, they naturally share common model architectures and visual representations. A unified framework not only helps simplify the algorithm development effort, but also allows to use both detection and segmentation data to improve representation learning. There have been several previous works to unify segmentation and detection tasks, e.g., Mask RCNN [9], HTC [2], and DETR [1]. Mask RCNN extends Faster RCNN and pools image features from Region Of Interest (ROI) proposed by RPN. HTC further proposes an interleaved way of predicting boxes and masks. However, these two models can only perform instance segmentation. DETR predicts boxes and masks together in an end-to-end manner. However, its segmentation performance largely lags behind other models. According to Table 2, adding DETR's segmentation head to DINO results in inferior instance segmentation results. How to attain mutual assistance between segmentation and detection has long been an important problem to solve." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.736, + 0.206, + 0.751 + ], + "angle": 0, + "content": "3. Mask DINO" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Mask DINO is an extension of DINO [32]. On top of content query embeddings, DINO has two branches for box prediction and label prediction. The boxes are dynamically updated and used to guide the deformable attention in each Transformer decoder. Mask DINO adds another branch for mask prediction and minimally extends several key components in detection to fit segmentation tasks. To better understand Mask DINO, we start by briefly reviewing DINO and then introduce Mask DINO." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.7, + 0.106 + ], + "angle": 0, + "content": "3.1. Preliminaries: DINO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.342 + ], + "angle": 0, + "content": "DINO is a typical DETR-like model, which is composed of a backbone, a Transformer encoder, and a Transformer decoder. The framework is shown in Fig. 1 (the blue-shaded part without red lines). Following DAB-DETR [18], DINO formulates each positional query in DETR as a 4D anchor box, which is dynamically updated through each decoder layer. Note that DINO uses multi-scale features with deformable attention [35]. Therefore, the updated anchor boxes are also used to constrain deformable attention in a sparse and soft way. Following DN-DETR [15], DINO adopts denoising training and further develops contrastive denoising to accelerate training convergence. Moreover, DINO proposes a mixed query selection scheme to initialize positional queries in the decoder and a look-forward-twice method to improve box gradient back-propagation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.349, + 0.892, + 0.381 + ], + "angle": 0, + "content": "3.2. Why a universal model has not replaced the specialized models in DETR-like models?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.644 + ], + "angle": 0, + "content": "Remarkable progress has been achieved by Transformer-based detectors and segmentation models. For instance, DINO [32] and Mask2Former [3] have achieved the best results on COCO detection and panoptic segmentation, respectively. Inspired by such progress, we attempted to simply extend these specialized models for other tasks but found that the performance of other tasks lagged behind the original ones by a large margin, as shown in Table 1 and 2. It seems that trivial multi-task training even hurts the performance of the original task. However, in convolution-based models, it has shown effective and mutually beneficial to combine detection and instance segmentation tasks. For example, detection models with Mask R-CNN head [9] is still ranked the first on the COCO instance segmentation. We will take DINO and Mask2Former as examples to discuss the challenges in unifying Transformer-based detection and segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.897, + 0.87 + ], + "angle": 0, + "content": "- What are the differences between specialized detection and segmentation models? Image segmentation is a pixel-level classification task, while object detection is a region-level regression task. In DETR-based model, the decoder queries are responsible for these tasks. For example, Mask2Former uses such decoder queries to dot-product the high-resolution feature maps to produce segmentation masks, while DINO uses them to regress boxes. However, as such queries in Mask2Former only have to compare per-pixel similarity with the image features, they may not be aware of the region-level position of each instance. On the contrary, queries in DINO are not designed to interact with such low-level features to learn pixel-level representation. Instead, they encode rich positional information and high-level semantics for detection." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "-Why cannot Mask2Former do detection well? The Transformer decoder of Mask2Former is designed for seg" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3043" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.094, + 0.422, + 0.16 + ], + "angle": 0, + "content": "
ModelBox APMask AP
Mask2Former46.2*43.7
Mask2Former + detection head21.641.3
DINO50.7-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.167, + 0.447, + 0.208 + ], + "angle": 0, + "content": "Table 1. Simply adding a detection head to Mask2Former results in low detection performance. * indicates the boxes are generated from the predicted masks. The generated boxes from Mask2Former are also inferior to DINO (-4.5 AP). The models are trained for 50 epochs." + }, + { + "type": "table", + "bbox": [ + 0.47, + 0.089, + 0.825, + 0.166 + ], + "angle": 0, + "content": "
ModelBox APMask AP
DINO + Mask2Former segmentation head49.940.2
DINO + DETR segmentation head (finetune DINO pretrained on COCO detection)-35.8
Mask2Former-43.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.454, + 0.172, + 0.838, + 0.203 + ], + "angle": 0, + "content": "Table 2. Simply adding a segmentation head to DINO results in low instance segmentation performance. The predicted masks from DINO are also inferior to Mask2Former (-3.5 AP). The models are trained for 50 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.473, + 0.552 + ], + "angle": 0, + "content": "mentation tasks and does not suit detection for three reasons. First, its queries follow the design in DETR [1] without being able to utilize better positional priors as studied in Conditional DETR [22], Anchor DETR [29], and DAB-DETR [18]. For example, its content queries are semantically aligned with the features from the Transformer encoder, whereas its positional queries are just learnable vectors as in vanilla DETR instead of being associated with a single-mode position1. If we remove its mask branch, it reduces to a variant of DETR [1], whose performance is inferior to recently improved DETR models. Second, Mask2Former adopts masked attention (multi-head attention with attention mask) in Transformer decoders. The attention masks predicted from a previous layer are of high resolution and used as hard-constraints for attention computation. They are neither efficient nor flexible for box prediction. Third, Mask2Former cannot explicitly perform box refinement layer by layer. Moreover, its coarse-to-fine mask refinement in decoders fails to use multi-scale features from the encoder. As shown in Table 1, the generated box AP from mask is 4.5 AP worse than DINO and trivial multi-task learning by adding a detection head is not working2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.473, + 0.827 + ], + "angle": 0, + "content": "- Why cannot DETR/DINO do segmentation well? As shown in Table 2, simply 1) adding DETR's segmentation head or 2) adding Mask2Former's segmentation head result in inferior performance compared to Mask2Former. We analyze the reasons as follows. The reason for 1) is that DETR's segmentation head is not optimal. The vanilla DETR lets each query embedding dot-product with the smallest feature map to compute attention maps and then upsamples them to get the mask predictions. This design lacks an interaction between queries and larger feature maps from the backbone. In addition, the head is too heavy to use mask auxiliary loss for mask refinement. The reason for 2) is that features in improved detection models are not aligned with segmentation. For example, DINO inherits many designs from [18, 32, 35] like query formulation, denoising training, and query selection. However, these components are designed to strengthen region-level representation for detection, which is not optimal for segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.221, + 0.741, + 0.236 + ], + "angle": 0, + "content": "3.3. Our Method: Mask DINO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.895, + 0.35 + ], + "angle": 0, + "content": "Mask DINO adopts the same architecture design for detection as in DINO with minimal modifications. In the Transformer decoder, Mask DINO adds a mask branch for segmentation and extends several key components in DINO for segmentation tasks. As shown in Fig. 1, the framework in the blue-shaded part is the original DINO model and the additional design for segmentation is marked with red lines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.687, + 0.374 + ], + "angle": 0, + "content": "3.4. Segmentation branch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.895, + 0.607 + ], + "angle": 0, + "content": "Following other unified models [3,4,28] for image segmentation, we perform mask classification for all segmentation tasks. Note that DINO is not designed for pixel-level alignment as its positional queries are formulated as anchor boxes and its content queries are used to predict box offset and class membership. To perform mask classification, we adopt a key idea from Mask2Former [3] to construct a pixel embedding map which is obtained from the backbone and Transformer encoder features. As shown in Fig. 1, the pixel embedding map is obtained by fusing the \\(1/4\\) resolution feature map \\(C_b\\) from the backbone with an upsampled \\(1/8\\) resolution feature map \\(C_e\\) from the Transformer encoder. Then we dot-product each content query embedding \\(q_c\\) from the decoder with the pixel embedding map to obtain an output mask \\(m\\)." + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.617, + 0.892, + 0.634 + ], + "angle": 0, + "content": "\\[\nm = q _ {c} \\otimes \\mathcal {M} \\left(\\mathcal {T} \\left(C _ {b}\\right) + \\mathcal {F} \\left(C _ {e}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.894, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is the segmentation head, \\(\\mathcal{T}\\) is a convolutional layer to map the channel dimension to the Transformer hidden dimension, and \\(\\mathcal{F}\\) is a simple interpolation function to perform 2x upsampling of \\(C_e\\). This segmentation branch is conceptually simple and easy to implement in the DINO framework, as shown in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.834, + 0.759 + ], + "angle": 0, + "content": "3.5. Unified and Enhanced Query Selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Unified query selection for mask: Query selection has been widely used in traditional two-stage models [24] and many DETR-like models [32, 35] to improve detection performance. We further improve the query selection scheme in Mask DINO for segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The encoder output features contain dense features, which can serve as better priors for the decoder. Therefore, we adopt three prediction heads (classification, detection, and segmentation) in the encoder output. Note that the three" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.863 + ], + "angle": 0, + "content": "1We refer the interested readers to discussions in Sec. 3 in DAB-DETR [18]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.9 + ], + "angle": 0, + "content": "2We also notice there are issues in official Mask2Former Github (https://github.com/facebookresearch/Mask2Former/issues/43) that fail to make Mask2Former work well by adding a detection head." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3044" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.856, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.895, + 0.378 + ], + "angle": 0, + "content": "Figure 1. The framework of Mask DINO, which is based on DINO (the blue-shaded part) with extensions (the red part) for segmentation tasks. 'QS' and 'DN' are short for query selection and denoising training, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.473, + 0.552 + ], + "angle": 0, + "content": "heads are identical to the decoder heads. The classification score of each token is considered as the confidence to select top-ranked features and feed them to the decoder as content queries. The selected features also regress boxes and dot-product with the high-resolution feature map to predict masks. The predicted boxes and masks will be supervised by the ground truth and are considered as initial anchors for the decoder. Note that we initialize both the content and anchor box queries in Mask DINO whereas DINO only initializes anchor box queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.473, + 0.826 + ], + "angle": 0, + "content": "Mask-enhanced anchor box initialization: As summarized in Sec 3.2, image segmentation is a pixel-level classification task while object detection is a region-level position regression task. Therefore, compared to detection, though segmentation is a more difficult task with fine-granularity, it is easier to learn in the initial stage. For example, masks are predicted by dot-producing queries with the high-resolution feature map, which only needs to compare per-pixel semantic similarity. However, detection requires to directly regress the box coordinates in an image. Therefore, in the initial stage after unified query selection, mask prediction is much more accurate than box (the qualitative AP comparison between mask prediction and box prediction in different stages is also shown in Table 8 and 9). Therefore, after unified query selection, we derive boxes from the predicted masks as better anchor box initialization for the decoder. By this effective task cooperation, the enhanced box initialization can bring in a large improvement to the detection performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.327, + 0.849 + ], + "angle": 0, + "content": "3.6. Segmentation Micro Design" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Unified denoising for mask: Query denoising in object detection has shown effective [15,32] to accelerate convergence and improve performance. It adds noises to ground-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.895, + 0.583 + ], + "angle": 0, + "content": "truth boxes and labels and feed them to the Transformer decoder as noised positional queries and content queries. The model is trained to reconstruct ground truth objects given their noised versions. We also extend this technique to segmentation tasks. As masks can be viewed as a more fine-grained representation of boxes, box and mask are naturally connected. Therefore, we can treat boxes as a noised version of masks, and train the model to predict masks given boxes as a denoising task. The given boxes for mask prediction are also randomly noised for more efficient mask denoising training. The detailed noise and its hyperparameters used in our model are shown in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Hybrid matching: Mask DINO, as in some traditional models [2, 9], predicts boxes and masks with two parallel heads in a loosely coupled manner. Hence the two heads can predict a pair of box and mask that are inconsistent with each other. To address this issue, in addition to the original box and classification loss in bipartite matching, we add a mask prediction loss to encourage more accurate and consistent matching results for one query. Therefore, the matching cost becomes \\(\\lambda_{cls}\\mathcal{L}_{cls} + \\lambda_{box}\\mathcal{L}_{box} + \\lambda_{mask}\\mathcal{L}_{mask}\\), where \\(\\mathcal{L}_{cls},\\mathcal{L}_{box}\\), and \\(\\mathcal{L}_{mask}\\) are the classification, box, and mask loss and \\(\\lambda\\) are their corresponding weights. The detailed losses used in our model and their corresponding weights are shown in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.799, + 0.634, + 0.815 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We conduct extensive experiments and compare with several specialized models for four popular tasks including object detection, instance, panoptic, and semantic segmentation on COCO [17], ADE20K [34], and Cityscapes [5]. For all experiments, we use batch size 16 and A100 GPUs with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3045" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.849, + 0.313 + ], + "angle": 0, + "content": "
ModelEpochsQuery typeMask APBox AP\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP_{S}^{mask}\\)\\(AP_{M}^{mask}\\)\\(AP_{L}^{mask}\\)GFLOPSParamsFPS
ResNet-50 backbone
Mask-RCNN [6,8,9]400Dense anchors42.548.2--23.845.060.020740M10.3
HTC [2]36Dense anchors39.744.961.443.122.642.250.644180M5
QueryInst [7]36300 queries40.645.663.044.023.442.552.8--7.0
DINO-4scale [32]36900 queries-50.9-----24547M19.6
Mask2Former [3]12100 queries38.7-59.841.218.241.559.822644M8.2
Mask DINO (ours)12300 queries41.4(+2.7)45.762.944.621.144.261.428652M14.8
Mask DINO (ours)24300 queries44.2(+0.5)48.466.647.923.947.064.028652M14.8
Mask2Former* [3]50100 queries43.7\\(46.2^{\\dagger}\\)66.046.923.447.264.822644M8.2
Mask DINO (ours)50100 queries45.449.867.949.325.248.365.828052M15.2
Mask DINO (ours)50300 queries46.050.568.950.326.0(+2.6)49.3(+2.1)65.5(+0.7)28652M14.8
Mask \\(DINO^‡\\) (ours)50300 queries46.3(+2.6)51.7(+0.8)69.050.726.1(+2.7)49.3(+2.1)66.1(+1.3)28652M14.2
SwinL backbone
HTC++ [2,20]72dense anchors49.557.1--31.052.467.21470284M-
Mask2Former [3]100200 queries50.1---29.953.972.1868216M4.0
DINO [32]36900 queries-58.5-----1285217M8.1
Mask2Former100200 queries50.1---29.953.972.1868216M4.0
Mask DINO (ours)50300 queries52.158.376.557.632.955.472.51326223M6.1
Mask \\(DINO^‡\\) (ours)50300 queries52.3(+2.2)59.0(+0.5)76.657.833.155.472.61326223M5.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.322, + 0.892, + 0.406 + ], + "angle": 0, + "content": "Table 3. Results for Mask DINO and other object detection and instance segmentation models with ResNet-50 and SwinL backbone on COCO val2017 without extra data or tricks. Following DINO [32], we use ResNet-50 with four feature scales by default, and use five scales under large models with a SwinL backbone. We follow the common practice in DETR-like models to use 300 queries. * Mask2Former using 300 queries is not listed as its performance will degenerate when using 300 queries. † indicates the box AP is derived from mask prediction. ‡ we use the proposed mask-enhanced box initialization to further improve detection performance. We test the FPS and GFLOPS of Mask2Former and Mask DINO on the A100 GPU using detectron2." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.42, + 0.773, + 0.6 + ], + "angle": 0, + "content": "
ModelEpochsQuery typePQ\\( PQ^{Th} \\)\\( PQ^{St} \\)Box \\( AP^{Th}_{pan} \\)Mask \\( AP^{Th}_{pan} \\)
ResNet-50 backbone
DETR [1]500 + 25100 queries43.448.236-31.1
Panoptic Segformer [16]24353 queries49.654.442.4-41.7
Mask2Former* [3]50100 queries\\( 51.9/51.5^† \\)57.743.0-41.7
Mask DINO (ours)50100 queries52.358.343.247.743.7
Mask DINO (ours)50300 queries\\( 53.0(+1.1) \\)\\( 59.1(+1.4) \\)\\( 43.9(+0.9) \\)48.8\\( 44.3(+2.6) \\)
Mask DINO (ours)24300 queries51.557.342.646.442.8
Mask2Former [3]12100 queries46.952.538.4-37.2
Panoptic Segformer [16]12353 queries48.052.341.5--
Mask DINO (ours)12300 queries\\( 49.0(+1.0) \\)54.840.243.2\\( 40.4(+3.2) \\)
SwinL backbone
Mask2Former [3]100100 queries57.864.248.1-48.6
OneFormer [11]100150 queries57.964.448.0-49.0
Mask DINO (ours)50300 queries\\( 58.3(+0.5) \\)65.148.056.2\\( 50.6(+2.0) \\)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.61, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Table 4. Results for Mask DINO and other panoptic segmentation models with a ResNet-50 backbone on COCO val2017. * Mask2Former using 300 queries is not listed as its performance will degenerate when using 300 queries. † Our reproduced result." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.712 + ], + "angle": 0, + "content": "40GB memory. We use a ResNet-50 [10] and a SwinL [20] backbone for our main results and SOTA model. Under ResNet-50, we use 4 A100 GPUs for all tasks without extra data. The implementation details are in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.216, + 0.74 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Instance segmentation and object detection. In Table 3, we compare Mask DINO with other instance segmentation and object detection models. Mask DINO outperforms both the specialized models such as Mask2Former [3] and DINO [32] and hybrid models such as HTC [2] under the same setting. Especially, the instance segmentation results surpass the strong baseline Mask2Former by a large margin (+2.7 AP and +2.6 AP) on the 12-epoch and 50-epoch settings. Moreover, Mask DINO significantly improves the convergence speed, outperforming Mask2Former with less" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.895, + 0.772 + ], + "angle": 0, + "content": "than half training epochs (44.2 AP in 24 epochs). In addition, after using mask-enhanced box initialization, our detection performance has been significantly improved (+1.2 AP), which even outperforms DINO by 0.8 AP. These results indicate that task unification is beneficial. Without bells and whistles, we achieve the best detection and instance segmentation performance among DETR-like model with a SwinL backbone without extra data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Panoptic segmentation. We compare Mask DINO with other models in Table 4. Mask DINO outperforms all previous best models on both the 12-epoch and 50-epoch settings by 1.0 PQ and 1.1 PQ, respectively. This indicates Mask DINO has the advantages of both faster convergence and superior performance. One interesting observation is that we outperform Mask2Former [3] in terms of both \\( PQ^{Th} \\) and \\( PQ^{St} \\). However, instead of using dense and hard" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3046" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.089, + 0.482, + 0.144 + ], + "angle": 0, + "content": "
ModelIterationsCrop sizemIoU (mean)mIoU (high)mIoU (reported)
Mask2Former [3]160k51246.146.547.2
Mask DINO (ours)160k51247.7(+1.6)48.7(+2.2)48.7(+1.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.149, + 0.473, + 0.194 + ], + "angle": 0, + "content": "Table 5. Results for Mask DINO and Mask2Former with 100 queries using a ResNet-50 backbone on ADE20K val. We found the performance variance on this dataset is high and run three times to report both the mean and highest results for both models." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.09, + 0.838, + 0.143 + ], + "angle": 0, + "content": "
ModelIterationsmIoU (mean)mIoU (high)mIoU (reported)
Mask2Former [3]90k78.779.079.4
Mask DINO (ours)90k79.8(+1.1)80.0(+1.0)80.0(+0.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.497, + 0.149, + 0.852, + 0.194 + ], + "angle": 0, + "content": "Table 6. Results for Mask DINO and Mask2Former with 100 queries using a ResNet-50 backbone on Cityscapes val. We found the performance variance on this dataset is high and run three times to report both the mean and highest results for both models." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.207, + 0.773, + 0.482 + ], + "angle": 0, + "content": "
MethodParamsBackboneBackbone Pre-training DatasetDetection Pre-training Datasetval
w/o TTAw/ TTA
Instance segmentation on COCOAP
Mask2Former [3]216MSwinLIN-22K-14M-50.1-
Soft Teacher [31]284MSwinLIN-22K-14MO36551.952.5
SwinV2-G-HTC++ [19]3.0BSwinV2-GIN-22K-ext-70M [19]O36553.453.7
MasK DINO(Ours)223MSwinLIN-22K-14M-52.6-
MasK DINO(Ours)223MSwinLIN-22K-14MO36554.5(+1.1)-
Panoptic segmentation on COCOPQ
Panoptic SegFormer [16]-MSwinLIN-22K-14M-55.8-
Mask2Former [3]216MSwinLIN-22K-14M-57.8-
MasK DINO (ours)223MSwinLIN-22K-14M58.4(+0.6)-
MasK DINO (ours)223MSwinLIN-22K-14MO36559.4(+1.6)-
Semantic segmentation on ADE20KmIoU
Mask2Former [3]215MSwinLIN-22K-14M-56.157.3
SeMask-L MSFaPN-Mask2Former [12]-MSwinL-FaPNIN-22K-14M--58.2
SwinV2-G-UperNet [19]3.0BSwinV2-GIN-22K-ext-70M [19]-59.359.9
MasK DINO (ours)223MSwinLIN-22K-14M-56.6-
MasK DINO (ours)223MSwinLIN-22K-14MO36559.560.8(+0.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.493, + 0.893, + 0.523 + ], + "angle": 0, + "content": "Table 7. Comparison of the SOTA models on three segmentation tasks. Mask DINO outperforms all existing models. \"TTA\" means test-time-augmentation. \"O365\" denotes the Objects365 [26] dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.658 + ], + "angle": 0, + "content": "constrained masked attention, we predict boxes and then use them in deformable attention to extract query features. Therefore, our box-oriented deformable attention also works well with \"stuff\" categories, which makes our unified model simple and efficient. In addition, we improve the mask \\(\\mathrm{AP}_{pan}^{Th}\\) by 2.6 to 44.3 AP, which is 0.6 higher than the specialized instance segmentation model Mask2Fomer (43.7 AP)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.735 + ], + "angle": 0, + "content": "Semantic segmentation. In Table 5 and 6, we show the performance of semantic segmentation with a ResNet-50 backbone. We use 100 queries for these small datasets. We outperform Mask2Former on both ADE20K and Cityscapes by 1.6 and \\(0.6\\mathrm{mIoU}\\) on the reported performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.36, + 0.759 + ], + "angle": 0, + "content": "4.2. Comparison with SOTA Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Table 7, we compare Mask DINO with SOTA models on three image segmentation tasks to show its scalability. We use the SwinL [20] backbone and pre-train DINO on the Objects365 [26] detection dataset. Even without using extra data, we outperform Mask2Former on all three tasks, especially on instance segmentation (+2.5 AP). As Mask DINO is an extension of DINO, the pre-trained DINO model can be used to fine-tune Mask DINO for segmentation tasks. After fine-tuning Mask DINO on the corresponding tasks, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.893, + 0.765 + ], + "angle": 0, + "content": "achieve the best results on instance (54.5 AP), panoptic (59.4 PQ), and semantic (60.8 mIoU) segmentation among model under one billion parameters. Compared to SwinV2-G [19], we significantly reduce the model size to 1/15 and backbone pre-training dataset to 1/5. Our detection pre-training also significantly helps all segmentation tasks including panoptic and semantic with \"stuff\" categories. However, previous specialized segmentation models such as Mask2Former can not use detection datasets and adding a detection head to it results in poor performance as shown in Table 1, which severely limits the data scalability. By unifying four tasks in one model, we only need to pre-train one model on a large-scale dataset and finetune on all tasks for 10 to 20 epochs (Mask2Former needs 100 epochs), which is more computationally efficient and simpler in model design." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.665, + 0.787 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.87 + ], + "angle": 0, + "content": "We conduct ablation studies using a ResNet-50 backbone to analyze Mask DINO on COCO val2017. Unless otherwise stated, our experiments are based on object detection and instance segmentation without Mask-enhanced anchor box initialization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Query selection. Table 8 shows the results of our query selection for instance segmentation, where we additionally pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3047" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.13, + 0.089, + 0.345, + 0.16 + ], + "angle": 0, + "content": "
test layer#Mask DINOMask2Former
layer 039.6(+38.5)1.1
layer 344.042.3
layer 645.943.3
layer 946.043.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.165, + 0.396, + 0.197 + ], + "angle": 0, + "content": "Table 8. Effectiveness of our query selection for mask initialization. We evaluate the instance segmentation performance from different decoder layers in the same model after training for 50 epochs." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.201, + 0.291, + 0.261 + ], + "angle": 0, + "content": "
TasksBox APMask AP
BoxMask
✓50.1-
✓-43.3
✓✓50.5(+0.4)46.0(+2.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.267, + 0.327, + 0.296 + ], + "angle": 0, + "content": "Table 11. Task comparison under the 50-epoch setting. We train the same Mask DINO with different tasks and validate that box and mask can achieve mutual cooperation." + }, + { + "type": "table", + "bbox": [ + 0.38, + 0.089, + 0.621, + 0.145 + ], + "angle": 0, + "content": "
layer#w/o MEw ME
BoxMaskBoxMask
layer 039.625.639.841.2(+15.6)
layer 946.050.546.351.7(+1.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.401, + 0.15, + 0.624, + 0.19 + ], + "angle": 0, + "content": "Table 9. Comparison of our model with and without Mask-enhanced anchor box initialization (ME). ME enhances anchor box initialization and improves final detection performance. Trained for 50 epochs." + }, + { + "type": "table", + "bbox": [ + 0.38, + 0.199, + 0.564, + 0.261 + ], + "angle": 0, + "content": "
Decoder layer#Box APMask AP
343.140.7
644.341.1
944.541.4
1244.841.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.341, + 0.267, + 0.601, + 0.296 + ], + "angle": 0, + "content": "Table 12. Decoder layer number comparison under the 12-epoch setting. Mask DINO benefits from more decoders, while DINO's performance will decrease with 9 decoders." + }, + { + "type": "table", + "bbox": [ + 0.663, + 0.089, + 0.86, + 0.145 + ], + "angle": 0, + "content": "
Feature scalebox APmask AP
single scale(1/8)45.845.1
3 scales50.545.8
4 scales50.546.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.639, + 0.15, + 0.884, + 0.181 + ], + "angle": 0, + "content": "Table 10. Comparison of multi-scale features for Transformer decoder under the 50-epoch setting. Both detection and segmentation benefit from more feature scales." + }, + { + "type": "table", + "bbox": [ + 0.66, + 0.195, + 0.832, + 0.256 + ], + "angle": 0, + "content": "
MatchingBox APMask AP
BoxMask
✓44.440.5
✓40.238.4
✓✓44.541.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.615, + 0.261, + 0.875, + 0.291 + ], + "angle": 0, + "content": "Table 13. Matching method comparison under the 12-epoch setting. We train both tasks together but use different matching methods to verify the effectiveness of hybrid matching." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.306, + 0.438, + 0.402 + ], + "angle": 0, + "content": "
Box APMask AP
Mask DINO (ours)45.741.4
- Mask-enhanced anchor box initialization44.5(-1.2)41.4
- Unified query selection for masks43.6(-2.1)40.3(-1.1)
- Unified denoising for masks44.4(-1.3)40.3(-1.1)
- Hybrid matching44.9(-0.8)40.5(-0.9)
- remove all the above41.7(-4.0)38.5(-2.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.407, + 0.438, + 0.419 + ], + "angle": 0, + "content": "Table 14. Effectiveness of the proposed components under the 12-epoch setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.473, + 0.658 + ], + "angle": 0, + "content": "vide the performance of different decoder layers in one single model. Mask2Former also predicts the masks of learnable queries as initial region proposals. However, their performance lags behind Mask DINO by a large margin (-38.5AP). With our effective query selection scheme, the mask performance achieves 39.6 AP without using the decoder. In addition, our mask performance at layer six is already comparable to the final results with 9 layers. In Table 9, we show that in query selection the predicted box is inferior to mask, which indicates segmentation is easier to learn in the initial stage. Therefore, our proposed mask-enhanced box initialization enhances boxes with masks in query selection to provide better anchor boxes (+15.6 AP) for the decoder, which results in +1.2 AP improvement in the final detection performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.735 + ], + "angle": 0, + "content": "Feature scales. Mask2Former [3] shows that concatenating multi-scale features as input to Transformer decoder layers does not improve the segmentation performance. However, in Table 10, Mask DINO shows that using more feature scales in the decoder consistently improves the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.856 + ], + "angle": 0, + "content": "Object detection and segmentation help each other. To validate task cooperation in Mask DINO, we use the same model but train different tasks and report the 12 epoch and 50 epoch results. As shown in Table 11, only training one task will lead to a performance drop. Although only training object detection results in faster convergence in the early stage for box prediction, the final performance is still inferior to training both tasks together." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Decoder layer number. In DINO, increasing the decoder layer number to nine will decrease the performance of box. In Table 12, the result indicates that increasing the number of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.895, + 0.369 + ], + "angle": 0, + "content": "decoder layers will contribute to both detection and segmentation in Mask DINO. We hypothesize that the multi-task training become more complex and require more decoders to learn the needed mapping function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.895, + 0.506 + ], + "angle": 0, + "content": "Matching. In Table 13, we show that only using boxes or masks to perform bipartite matching is not optimal in Mask DINO. A unified matching objective makes the optimization more consistent. In addition, after removing all the proposed components, both detection and segmentation performance drop by a large margin. This result indicates that if we trivially add detection and segmentation tasks in one DETR-based model, the features are not aligned for detection and segmentation tasks to achieve mutual cooperation." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.51, + 0.881, + 0.526 + ], + "angle": 0, + "content": "We also present visualization analysis in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.554, + 0.62, + 0.57 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.824 + ], + "angle": 0, + "content": "In this paper, we have presented Mask DINO as a unified Transformer-based framework for both object detection and image segmentation. Conceptually, Mask DINO is a natural extension of DINO from detection to segmentation with minimal modifications on some key components. Mask DINO outperforms previous specialized models and achieves the best results on all three segmentation tasks (instance, panoptic, and semantic) among models under one billion parameters. Moreover, Mask DINO shows that detection and segmentation can help each other in query-based models. In particular, Mask DINO enables semantic and panoptic segmentation to benefit from a better visual representation pre-trained on a large-scale detection dataset. We hope Mask DINO can provide insights for enabling task cooperation and data cooperation towards designing a universal model for more vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Limitations: Different segmentation tasks fail to achieve mutual assistance in Mask DINO in COCO panoptic segmentation. For example, in COCO panoptic segmentation, the mask AP still lags behind the model only trained with instances." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3048" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.473, + 0.185 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.27 + ], + "angle": 0, + "content": "[2] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.472, + 0.5 + ], + "angle": 0, + "content": "[6] Xianzhi Du, Barret Zoph, Wei-Chih Hung, and Tsung-Yi Lin. Simple training strategies and model scaling for object detection. arXiv preprint arXiv:2107.00057, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.472, + 0.558 + ], + "angle": 0, + "content": "[7] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.472, + 0.642 + ], + "angle": 0, + "content": "[8] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.643, + 0.472, + 0.698 + ], + "angle": 0, + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[11] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. arXiv preprint arXiv:2211.06220, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[12] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. SeMask: Semantically Masked Transformers for Semantic Segmentation. arXiv preprint arXiv:2112.12782, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6399-6408, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[14] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[15] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. DN-DETR: Accelerate DETR Training by Introducing Query DeNoising. arXiv preprint arXiv:2203.01305, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.894, + 0.277 + ], + "angle": 0, + "content": "[16] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Tong Lu, and Ping Luo. Panoptic SegFormer. arXiv preprint arXiv:2109.03814, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.894, + 0.348 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[18] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.894, + 0.462 + ], + "angle": 0, + "content": "[19] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin Transformer V2: Scaling Up Capacity and Resolution. arXiv preprint arXiv:2111.09883, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.894, + 0.533 + ], + "angle": 0, + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.894, + 0.589 + ], + "angle": 0, + "content": "[21] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.894, + 0.646 + ], + "angle": 0, + "content": "[22] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional DETR for Fast Training Convergence. arXiv preprint arXiv:2108.06152, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[23] Zipeng Qin, Jianbo Liu, Xiaolin Zhang, Maoqing Tian, Aojun Zhou, Shuai Yi, and Hongsheng Li. Pyramid Fusion Transformer for Semantic Segmentation. arXiv preprint arXiv:2201.04019, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.894, + 0.759 + ], + "angle": 0, + "content": "[24] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3049" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[28] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.22, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[29] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. arXiv preprint arXiv:2109.07107, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.471, + 0.331 + ], + "angle": 0, + "content": "[30] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8826, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.471, + 0.402 + ], + "angle": 0, + "content": "[31] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.404, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[32] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. arXiv preprint arXiv:2203.03605, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.474, + 0.471, + 0.515 + ], + "angle": 0, + "content": "[33] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.517, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[34] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.574, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[35] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR 2021: The Ninth International Conference on Learning Representations, 2021. 1, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3050" + } + ] +] \ No newline at end of file diff --git a/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_origin.pdf b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64fa593be583c06c27c8b1bd1e576576e6e5f6aa --- /dev/null +++ b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/df329ef8-a799-4ebf-906c-9403b6e58fb2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3af89ed87c40aa98d117fe3e3eaca54c26c08c897d92ba75a39f5ecf0945561 +size 769807 diff --git a/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/full.md b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b33756f204adfc544b9f50eb289bb03de8ee949 --- /dev/null +++ b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/full.md @@ -0,0 +1,251 @@ +# Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation + +Feng Li $^{1,3*†}$ , Hao Zhang $^{1,3*†}$ , Huaizhe Xu $^{1,3}$ , Shilong Liu $^{2,3}$ , Lei Zhang $^{3‡}$ , Lionel M. Ni $^{1,4}$ , Heung-Yeung Shum $^{1,3}$ + +1The Hong Kong University of Science and Technology. + +$^{2}$ Dept. of CST., BNRist Center, Institute for AI, Tsinghua University. + +$^{3}$ International Digital Economy Academy (IDEA). + +4The Hong Kong University of Science and Technology (Guangzhou). + +{fiay,hzhangcx,hxubr}@connect.ust.hk {liusl20}@mails.tsinghua.edu.cn {leizhang}@idea.edu.cn {ni,hshum}@ust.hk + +# Abstract + +In this paper we present Mask DINO, a unified object detection and segmentation framework. Mask DINO extends DINO (DETR with Improved Denoising Anchor Boxes) by adding a mask prediction branch which supports all image segmentation tasks (instance, panoptic, and semantic). It makes use of the query embeddings from DINO to dot-product a high-resolution pixel embedding map to predict a set of binary masks. Some key components in DINO are extended for segmentation through a shared architecture and training process. Mask DINO is simple, efficient, and scalable, and it can benefit from joint large-scale detection and segmentation datasets. Our experiments show that Mask DINO significantly outperforms all existing specialized segmentation methods, both on a ResNet-50 backbone and a pre-trained model with SwinL backbone. Notably, Mask DINO establishes the best results to date on instance segmentation (54.5 AP on COCO), panoptic segmentation (59.4 PQ on COCO), and semantic segmentation (60.8 mIoU on ADE20K) among models under one billion parameters. Code is available at https://github.com/IDEA-Research/MaskDINO. + +# 1. Introduction + +Object detection and image segmentation are fundamental tasks in computer vision. Both tasks are concerned with localizing objects of interest in an image but have different levels of focus. Object detection is to localize objects of interest and predict their bounding boxes and category + +labels, whereas image segmentation focuses on pixel-level grouping of different semantics. Moreover, image segmentation encompasses various tasks including instance segmentation, panoptic segmentation, and semantic segmentation with respect to different semantics, e.g., instance or category membership, foreground or background category. + +Remarkable progress has been achieved by classical convolution-based algorithms developed for these tasks with specialized architectures, such as Faster RCNN [24] for object detection, Mask RCNN [9] for instance segmentation, and FCN [21] for semantic segmentation. Although these methods are conceptually simple and effective, they are tailored for specialized tasks and lack the generalization ability to address other tasks. The ambition to bridge different tasks gives rise to more advanced methods like HTC [2] for object detection and instance segmentation and Panoptic FPN [14], K-net [33] for instance, panoptic, and semantic segmentation. Task unification not only helps simplify algorithm development but also brings in performance improvement in multiple tasks. + +Recently, DETR-like [1] models developed based on Transformers [27] have achieved inspiring progress on many detection and segmentation tasks. As an end-to-end object detector, DETR adopts a set-prediction objective and eliminates hand-crafted modules such as anchor design and nonmaximum suppression. Although DETR addresses both the object detection and panoptic segmentation tasks, its segmentation performance is still inferior to classical segmentation models. To improve the detection and segmentation performance of Transformer-based models, researchers have developed specialized models for object detection [15, 18, 32, 35], image segmentation [3, 4, 33], instance segmentation [7], panoptic segmentation [23], and semantic segmentation [12]. + +Among the efforts to improve object detection, DINO [32] takes advantage of the dynamic anchor box for + +mulation from DAB-DETR [18] and query denoising training from DN-DETR [15], and further achieves the SOTA result on the COCO object detection leaderboard for the first time as a DETR-like model. Similarly, for improving image segmentation, MaskFormer [4] and Mask2Former [3] propose to unify different image segmentation tasks using query-based Transformer architectures to perform mask classification. Such methods have achieved remarkable performance improvement on multiple segmentation tasks. + +However, in Transformer-based models, the best-performing detection and segmentation models are still not unified, which prevents task and data cooperation between detection and segmentation tasks. As an evidence, in CNN-based models, Mask-R-CNN [9] and HTC [2] are still widely acknowledged as unified models that achieve mutual cooperation between detection and segmentation to achieve superior performance than specialized models. Though we believe detection and segmentation can help each other in a unified architecture in Transformer-based models, the results of simply using DINO for segmentation and using Mask2Former for detection indicate that they can not do other tasks well, as shown in Table 1 and 2. Moreover, trivial multi-task training can even hurt the performance of the original tasks. It naturally leads to two questions: 1) why cannot detection and segmentation tasks help each other in Transformer-based models? and 2) is it possible to develop a unified architecture to replace specialized ones? + +To address these problems, we propose Mask DINO, which extends DINO with a mask prediction branch in parallel with DINO's box prediction branch. Inspired by other unified models [3, 4, 28] for image segmentation, we reuse content query embeddings from DINO to perform mask classification for all segmentation tasks on a high-resolution pixel embedding map (1/4 of the input image resolution) obtained from the backbone and Transformer encoder features. The mask branch predicts binary masks by simply dot-producting each content query embedding with the pixel embedding map. As DINO is a detection model for region-level regression, it is not designed for pixel-level alignment. To better align features between detection and segmentation, we also propose three key components to boost the segmentation performance. First, we propose a unified and enhanced query selection. It utilizes encoder dense prior by predicting masks from the top-ranked tokens to initialize mask queries as anchors. In addition, we observe that pixel-level segmentation is easier to learn in the early stage and propose to use initial masks to enhance boxes, which achieves task cooperation. Second, we propose a unified denoising training for masks to accelerate segmentation training. Third, we use a hybrid bipartite matching for more accurate and consistent matching from ground truth to both boxes and masks. + +Mask DINO is conceptually simple and easy to implement under the DINO framework. To summarize, our contri + +butions are three-fold. 1) We develop a unified Transformer-based framework for both object detection and segmentation. As the framework is extended from DINO, by adding a mask prediction branch, it naturally inherits most algorithm improvements in DINO including anchor box-guided cross attention, query selection, denoising training, and even a better representation pre-trained on a large-scale detection dataset. 2) We demonstrate that detection and segmentation can help each other through a shared architecture design and training method. Especially, detection can significantly help segmentation tasks, even for segmenting background "stuff" categories. Under the same setting with a ResNet-50 backbone, Mask DINO outperforms all existing models compared to DINO (+0.8 AP on COCO detection) and Mask2Former (+2.6 AP, +1.1 PQ, and +1.5 mIoU on COCO instance, COCO panoptic, and ADE20K semantic segmentation). 3) We also show that, via a unified framework, segmentation can benefit from detection pre-training on a large-scale detection dataset. After detection pre-training on the Objects365 [26] dataset with a SwinL [20] backbone, Mask DINO significantly improves all segmentation tasks and achieves the best results on instance (54.5 AP on COCO), panoptic (59.4 PQ on COCO), and semantic (60.8 mIoU on ADE20K) segmentation among models under one billion parameters. + +# 2. Related Work + +Detection: Mainstream detection algorithms have been dominated by convolutional neural network-based frameworks, until recently Transformer-based detectors [1, 15, 18, 32] achieve great progress. DETR [1] is the first end-to-end and query-based Transformer object detector, which adopts a set-prediction objective with bipartite matching. DAB-DETR [18] improves DETR by formulating queries as 4D anchor boxes and refining predictions layer by layer. DN-DETR [15] introduces a denoising training method to accelerate convergence. Based on DAB-DETR and DN-DETR, DINO [32] proposes several new improvements on denoising and anchor refinement and achieves new SOTA results on COCO detection. Despite the inspiring progress, DETR-like detection models are not competitive for segmentation. Vanilla DETR incorporates a segmentation head in its architecture. However, its segmentation performance is inferior to specialized segmentation models and only shows the feasibility of DETR-like detection models to deal with detection and segmentation simultaneously. + +Segmentation: Segmentation mainly includes instance, semantic, and panoptic segmentation. Instance segmentation is to predict a mask and its corresponding category for each object instance. Semantic segmentation requires to classify each pixel including the background into different semantic categories. Panoptic segmentation [14] unifies the instance and semantic segmentation tasks and predicts a mask for + +each object instance or background segment. In the past few years, researchers have developed specialized architectures for the three tasks. For example, Mask-RCNN [9] and HTC [2] can only deal with instance segmentation because they predict the mask of each instance based on its box prediction. FCN [21] and U-Net [25] can only perform semantic segmentation since they predict one segmentation map based on pixel-wise classification. Although models for panoptic segmentation [13, 30] unifies the above two tasks, they are usually inferior to specialized instance and semantic segmentation models. Until recently, some image segmentation models [3, 4, 33] are developed to unify the three tasks with a universal architecture. For instance, Mask2Former [3] improves MaskFormer [4] by introducing masked-attention to Transformer. Mask2Former has a similar architecture as DETR to probe image features with learnable queries but differs in using a different segmentation branch and some specialized designs for mask prediction. However, while Mask2Former shows a great success in unifying all segmentation tasks, it leaves object detection untouched and our empirical study shows that its specialized architecture design is not suitable for predicting boxes. + +Unified Methods: As both object detection and segmentation are concerned with localizing objects, they naturally share common model architectures and visual representations. A unified framework not only helps simplify the algorithm development effort, but also allows to use both detection and segmentation data to improve representation learning. There have been several previous works to unify segmentation and detection tasks, e.g., Mask RCNN [9], HTC [2], and DETR [1]. Mask RCNN extends Faster RCNN and pools image features from Region Of Interest (ROI) proposed by RPN. HTC further proposes an interleaved way of predicting boxes and masks. However, these two models can only perform instance segmentation. DETR predicts boxes and masks together in an end-to-end manner. However, its segmentation performance largely lags behind other models. According to Table 2, adding DETR's segmentation head to DINO results in inferior instance segmentation results. How to attain mutual assistance between segmentation and detection has long been an important problem to solve. + +# 3. Mask DINO + +Mask DINO is an extension of DINO [32]. On top of content query embeddings, DINO has two branches for box prediction and label prediction. The boxes are dynamically updated and used to guide the deformable attention in each Transformer decoder. Mask DINO adds another branch for mask prediction and minimally extends several key components in detection to fit segmentation tasks. To better understand Mask DINO, we start by briefly reviewing DINO and then introduce Mask DINO. + +# 3.1. Preliminaries: DINO + +DINO is a typical DETR-like model, which is composed of a backbone, a Transformer encoder, and a Transformer decoder. The framework is shown in Fig. 1 (the blue-shaded part without red lines). Following DAB-DETR [18], DINO formulates each positional query in DETR as a 4D anchor box, which is dynamically updated through each decoder layer. Note that DINO uses multi-scale features with deformable attention [35]. Therefore, the updated anchor boxes are also used to constrain deformable attention in a sparse and soft way. Following DN-DETR [15], DINO adopts denoising training and further develops contrastive denoising to accelerate training convergence. Moreover, DINO proposes a mixed query selection scheme to initialize positional queries in the decoder and a look-forward-twice method to improve box gradient back-propagation. + +# 3.2. Why a universal model has not replaced the specialized models in DETR-like models? + +Remarkable progress has been achieved by Transformer-based detectors and segmentation models. For instance, DINO [32] and Mask2Former [3] have achieved the best results on COCO detection and panoptic segmentation, respectively. Inspired by such progress, we attempted to simply extend these specialized models for other tasks but found that the performance of other tasks lagged behind the original ones by a large margin, as shown in Table 1 and 2. It seems that trivial multi-task training even hurts the performance of the original task. However, in convolution-based models, it has shown effective and mutually beneficial to combine detection and instance segmentation tasks. For example, detection models with Mask R-CNN head [9] is still ranked the first on the COCO instance segmentation. We will take DINO and Mask2Former as examples to discuss the challenges in unifying Transformer-based detection and segmentation. + +- What are the differences between specialized detection and segmentation models? Image segmentation is a pixel-level classification task, while object detection is a region-level regression task. In DETR-based model, the decoder queries are responsible for these tasks. For example, Mask2Former uses such decoder queries to dot-product the high-resolution feature maps to produce segmentation masks, while DINO uses them to regress boxes. However, as such queries in Mask2Former only have to compare per-pixel similarity with the image features, they may not be aware of the region-level position of each instance. On the contrary, queries in DINO are not designed to interact with such low-level features to learn pixel-level representation. Instead, they encode rich positional information and high-level semantics for detection. + +-Why cannot Mask2Former do detection well? The Transformer decoder of Mask2Former is designed for seg + +
ModelBox APMask AP
Mask2Former46.2*43.7
Mask2Former + detection head21.641.3
DINO50.7-
+ +Table 1. Simply adding a detection head to Mask2Former results in low detection performance. * indicates the boxes are generated from the predicted masks. The generated boxes from Mask2Former are also inferior to DINO (-4.5 AP). The models are trained for 50 epochs. + +
ModelBox APMask AP
DINO + Mask2Former segmentation head49.940.2
DINO + DETR segmentation head (finetune DINO pretrained on COCO detection)-35.8
Mask2Former-43.7
+ +Table 2. Simply adding a segmentation head to DINO results in low instance segmentation performance. The predicted masks from DINO are also inferior to Mask2Former (-3.5 AP). The models are trained for 50 epochs. + +mentation tasks and does not suit detection for three reasons. First, its queries follow the design in DETR [1] without being able to utilize better positional priors as studied in Conditional DETR [22], Anchor DETR [29], and DAB-DETR [18]. For example, its content queries are semantically aligned with the features from the Transformer encoder, whereas its positional queries are just learnable vectors as in vanilla DETR instead of being associated with a single-mode position1. If we remove its mask branch, it reduces to a variant of DETR [1], whose performance is inferior to recently improved DETR models. Second, Mask2Former adopts masked attention (multi-head attention with attention mask) in Transformer decoders. The attention masks predicted from a previous layer are of high resolution and used as hard-constraints for attention computation. They are neither efficient nor flexible for box prediction. Third, Mask2Former cannot explicitly perform box refinement layer by layer. Moreover, its coarse-to-fine mask refinement in decoders fails to use multi-scale features from the encoder. As shown in Table 1, the generated box AP from mask is 4.5 AP worse than DINO and trivial multi-task learning by adding a detection head is not working2. + +- Why cannot DETR/DINO do segmentation well? As shown in Table 2, simply 1) adding DETR's segmentation head or 2) adding Mask2Former's segmentation head result in inferior performance compared to Mask2Former. We analyze the reasons as follows. The reason for 1) is that DETR's segmentation head is not optimal. The vanilla DETR lets each query embedding dot-product with the smallest feature map to compute attention maps and then upsamples them to get the mask predictions. This design lacks an interaction between queries and larger feature maps from the backbone. In addition, the head is too heavy to use mask auxiliary loss for mask refinement. The reason for 2) is that features in improved detection models are not aligned with segmentation. For example, DINO inherits many designs from [18, 32, 35] like query formulation, denoising training, and query selection. However, these components are designed to strengthen region-level representation for detection, which is not optimal for segmentation. + +# 3.3. Our Method: Mask DINO + +Mask DINO adopts the same architecture design for detection as in DINO with minimal modifications. In the Transformer decoder, Mask DINO adds a mask branch for segmentation and extends several key components in DINO for segmentation tasks. As shown in Fig. 1, the framework in the blue-shaded part is the original DINO model and the additional design for segmentation is marked with red lines. + +# 3.4. Segmentation branch + +Following other unified models [3,4,28] for image segmentation, we perform mask classification for all segmentation tasks. Note that DINO is not designed for pixel-level alignment as its positional queries are formulated as anchor boxes and its content queries are used to predict box offset and class membership. To perform mask classification, we adopt a key idea from Mask2Former [3] to construct a pixel embedding map which is obtained from the backbone and Transformer encoder features. As shown in Fig. 1, the pixel embedding map is obtained by fusing the $1/4$ resolution feature map $C_b$ from the backbone with an upsampled $1/8$ resolution feature map $C_e$ from the Transformer encoder. Then we dot-product each content query embedding $q_c$ from the decoder with the pixel embedding map to obtain an output mask $m$ . + +$$ +m = q _ {c} \otimes \mathcal {M} \left(\mathcal {T} \left(C _ {b}\right) + \mathcal {F} \left(C _ {e}\right)\right), \tag {1} +$$ + +where $\mathcal{M}$ is the segmentation head, $\mathcal{T}$ is a convolutional layer to map the channel dimension to the Transformer hidden dimension, and $\mathcal{F}$ is a simple interpolation function to perform 2x upsampling of $C_e$ . This segmentation branch is conceptually simple and easy to implement in the DINO framework, as shown in Fig. 1. + +# 3.5. Unified and Enhanced Query Selection + +Unified query selection for mask: Query selection has been widely used in traditional two-stage models [24] and many DETR-like models [32, 35] to improve detection performance. We further improve the query selection scheme in Mask DINO for segmentation tasks. + +The encoder output features contain dense features, which can serve as better priors for the decoder. Therefore, we adopt three prediction heads (classification, detection, and segmentation) in the encoder output. Note that the three + +![](images/ef954800f9933bcb1adf83ab89f27b43762ee8de0f7e3506e627564632a6095c.jpg) +Figure 1. The framework of Mask DINO, which is based on DINO (the blue-shaded part) with extensions (the red part) for segmentation tasks. 'QS' and 'DN' are short for query selection and denoising training, respectively. + +heads are identical to the decoder heads. The classification score of each token is considered as the confidence to select top-ranked features and feed them to the decoder as content queries. The selected features also regress boxes and dot-product with the high-resolution feature map to predict masks. The predicted boxes and masks will be supervised by the ground truth and are considered as initial anchors for the decoder. Note that we initialize both the content and anchor box queries in Mask DINO whereas DINO only initializes anchor box queries. + +Mask-enhanced anchor box initialization: As summarized in Sec 3.2, image segmentation is a pixel-level classification task while object detection is a region-level position regression task. Therefore, compared to detection, though segmentation is a more difficult task with fine-granularity, it is easier to learn in the initial stage. For example, masks are predicted by dot-producing queries with the high-resolution feature map, which only needs to compare per-pixel semantic similarity. However, detection requires to directly regress the box coordinates in an image. Therefore, in the initial stage after unified query selection, mask prediction is much more accurate than box (the qualitative AP comparison between mask prediction and box prediction in different stages is also shown in Table 8 and 9). Therefore, after unified query selection, we derive boxes from the predicted masks as better anchor box initialization for the decoder. By this effective task cooperation, the enhanced box initialization can bring in a large improvement to the detection performance. + +# 3.6. Segmentation Micro Design + +Unified denoising for mask: Query denoising in object detection has shown effective [15,32] to accelerate convergence and improve performance. It adds noises to ground- + +truth boxes and labels and feed them to the Transformer decoder as noised positional queries and content queries. The model is trained to reconstruct ground truth objects given their noised versions. We also extend this technique to segmentation tasks. As masks can be viewed as a more fine-grained representation of boxes, box and mask are naturally connected. Therefore, we can treat boxes as a noised version of masks, and train the model to predict masks given boxes as a denoising task. The given boxes for mask prediction are also randomly noised for more efficient mask denoising training. The detailed noise and its hyperparameters used in our model are shown in Appendix ?? + +Hybrid matching: Mask DINO, as in some traditional models [2, 9], predicts boxes and masks with two parallel heads in a loosely coupled manner. Hence the two heads can predict a pair of box and mask that are inconsistent with each other. To address this issue, in addition to the original box and classification loss in bipartite matching, we add a mask prediction loss to encourage more accurate and consistent matching results for one query. Therefore, the matching cost becomes $\lambda_{cls}\mathcal{L}_{cls} + \lambda_{box}\mathcal{L}_{box} + \lambda_{mask}\mathcal{L}_{mask}$ , where $\mathcal{L}_{cls},\mathcal{L}_{box}$ , and $\mathcal{L}_{mask}$ are the classification, box, and mask loss and $\lambda$ are their corresponding weights. The detailed losses used in our model and their corresponding weights are shown in Appendix ?? + +# 4. Experiments + +We conduct extensive experiments and compare with several specialized models for four popular tasks including object detection, instance, panoptic, and semantic segmentation on COCO [17], ADE20K [34], and Cityscapes [5]. For all experiments, we use batch size 16 and A100 GPUs with + +
ModelEpochsQuery typeMask APBox AP\(AP_{50}^{mask}\)\(AP_{75}^{mask}\)\(AP_{S}^{mask}\)\(AP_{M}^{mask}\)\(AP_{L}^{mask}\)GFLOPSParamsFPS
ResNet-50 backbone
Mask-RCNN [6,8,9]400Dense anchors42.548.2--23.845.060.020740M10.3
HTC [2]36Dense anchors39.744.961.443.122.642.250.644180M5
QueryInst [7]36300 queries40.645.663.044.023.442.552.8--7.0
DINO-4scale [32]36900 queries-50.9-----24547M19.6
Mask2Former [3]12100 queries38.7-59.841.218.241.559.822644M8.2
Mask DINO (ours)12300 queries41.4(+2.7)45.762.944.621.144.261.428652M14.8
Mask DINO (ours)24300 queries44.2(+0.5)48.466.647.923.947.064.028652M14.8
Mask2Former* [3]50100 queries43.7\(46.2^{\dagger}\)66.046.923.447.264.822644M8.2
Mask DINO (ours)50100 queries45.449.867.949.325.248.365.828052M15.2
Mask DINO (ours)50300 queries46.050.568.950.326.0(+2.6)49.3(+2.1)65.5(+0.7)28652M14.8
Mask \(DINO^‡\) (ours)50300 queries46.3(+2.6)51.7(+0.8)69.050.726.1(+2.7)49.3(+2.1)66.1(+1.3)28652M14.2
SwinL backbone
HTC++ [2,20]72dense anchors49.557.1--31.052.467.21470284M-
Mask2Former [3]100200 queries50.1---29.953.972.1868216M4.0
DINO [32]36900 queries-58.5-----1285217M8.1
Mask2Former100200 queries50.1---29.953.972.1868216M4.0
Mask DINO (ours)50300 queries52.158.376.557.632.955.472.51326223M6.1
Mask \(DINO^‡\) (ours)50300 queries52.3(+2.2)59.0(+0.5)76.657.833.155.472.61326223M5.6
+ +Table 3. Results for Mask DINO and other object detection and instance segmentation models with ResNet-50 and SwinL backbone on COCO val2017 without extra data or tricks. Following DINO [32], we use ResNet-50 with four feature scales by default, and use five scales under large models with a SwinL backbone. We follow the common practice in DETR-like models to use 300 queries. * Mask2Former using 300 queries is not listed as its performance will degenerate when using 300 queries. † indicates the box AP is derived from mask prediction. ‡ we use the proposed mask-enhanced box initialization to further improve detection performance. We test the FPS and GFLOPS of Mask2Former and Mask DINO on the A100 GPU using detectron2. + +
ModelEpochsQuery typePQ\( PQ^{Th} \)\( PQ^{St} \)Box \( AP^{Th}_{pan} \)Mask \( AP^{Th}_{pan} \)
ResNet-50 backbone
DETR [1]500 + 25100 queries43.448.236-31.1
Panoptic Segformer [16]24353 queries49.654.442.4-41.7
Mask2Former* [3]50100 queries\( 51.9/51.5^† \)57.743.0-41.7
Mask DINO (ours)50100 queries52.358.343.247.743.7
Mask DINO (ours)50300 queries\( 53.0(+1.1) \)\( 59.1(+1.4) \)\( 43.9(+0.9) \)48.8\( 44.3(+2.6) \)
Mask DINO (ours)24300 queries51.557.342.646.442.8
Mask2Former [3]12100 queries46.952.538.4-37.2
Panoptic Segformer [16]12353 queries48.052.341.5--
Mask DINO (ours)12300 queries\( 49.0(+1.0) \)54.840.243.2\( 40.4(+3.2) \)
SwinL backbone
Mask2Former [3]100100 queries57.864.248.1-48.6
OneFormer [11]100150 queries57.964.448.0-49.0
Mask DINO (ours)50300 queries\( 58.3(+0.5) \)65.148.056.2\( 50.6(+2.0) \)
+ +Table 4. Results for Mask DINO and other panoptic segmentation models with a ResNet-50 backbone on COCO val2017. * Mask2Former using 300 queries is not listed as its performance will degenerate when using 300 queries. † Our reproduced result. + +40GB memory. We use a ResNet-50 [10] and a SwinL [20] backbone for our main results and SOTA model. Under ResNet-50, we use 4 A100 GPUs for all tasks without extra data. The implementation details are in Appendix ?? + +# 4.1. Main Results + +Instance segmentation and object detection. In Table 3, we compare Mask DINO with other instance segmentation and object detection models. Mask DINO outperforms both the specialized models such as Mask2Former [3] and DINO [32] and hybrid models such as HTC [2] under the same setting. Especially, the instance segmentation results surpass the strong baseline Mask2Former by a large margin (+2.7 AP and +2.6 AP) on the 12-epoch and 50-epoch settings. Moreover, Mask DINO significantly improves the convergence speed, outperforming Mask2Former with less + +than half training epochs (44.2 AP in 24 epochs). In addition, after using mask-enhanced box initialization, our detection performance has been significantly improved (+1.2 AP), which even outperforms DINO by 0.8 AP. These results indicate that task unification is beneficial. Without bells and whistles, we achieve the best detection and instance segmentation performance among DETR-like model with a SwinL backbone without extra data. + +Panoptic segmentation. We compare Mask DINO with other models in Table 4. Mask DINO outperforms all previous best models on both the 12-epoch and 50-epoch settings by 1.0 PQ and 1.1 PQ, respectively. This indicates Mask DINO has the advantages of both faster convergence and superior performance. One interesting observation is that we outperform Mask2Former [3] in terms of both $PQ^{Th}$ and $PQ^{St}$ . However, instead of using dense and hard + +
ModelIterationsCrop sizemIoU (mean)mIoU (high)mIoU (reported)
Mask2Former [3]160k51246.146.547.2
Mask DINO (ours)160k51247.7(+1.6)48.7(+2.2)48.7(+1.6)
+ +Table 5. Results for Mask DINO and Mask2Former with 100 queries using a ResNet-50 backbone on ADE20K val. We found the performance variance on this dataset is high and run three times to report both the mean and highest results for both models. + +
ModelIterationsmIoU (mean)mIoU (high)mIoU (reported)
Mask2Former [3]90k78.779.079.4
Mask DINO (ours)90k79.8(+1.1)80.0(+1.0)80.0(+0.6)
+ +Table 6. Results for Mask DINO and Mask2Former with 100 queries using a ResNet-50 backbone on Cityscapes val. We found the performance variance on this dataset is high and run three times to report both the mean and highest results for both models. + +
MethodParamsBackboneBackbone Pre-training DatasetDetection Pre-training Datasetval
w/o TTAw/ TTA
Instance segmentation on COCOAP
Mask2Former [3]216MSwinLIN-22K-14M-50.1-
Soft Teacher [31]284MSwinLIN-22K-14MO36551.952.5
SwinV2-G-HTC++ [19]3.0BSwinV2-GIN-22K-ext-70M [19]O36553.453.7
MasK DINO(Ours)223MSwinLIN-22K-14M-52.6-
MasK DINO(Ours)223MSwinLIN-22K-14MO36554.5(+1.1)-
Panoptic segmentation on COCOPQ
Panoptic SegFormer [16]-MSwinLIN-22K-14M-55.8-
Mask2Former [3]216MSwinLIN-22K-14M-57.8-
MasK DINO (ours)223MSwinLIN-22K-14M58.4(+0.6)-
MasK DINO (ours)223MSwinLIN-22K-14MO36559.4(+1.6)-
Semantic segmentation on ADE20KmIoU
Mask2Former [3]215MSwinLIN-22K-14M-56.157.3
SeMask-L MSFaPN-Mask2Former [12]-MSwinL-FaPNIN-22K-14M--58.2
SwinV2-G-UperNet [19]3.0BSwinV2-GIN-22K-ext-70M [19]-59.359.9
MasK DINO (ours)223MSwinLIN-22K-14M-56.6-
MasK DINO (ours)223MSwinLIN-22K-14MO36559.560.8(+0.9)
+ +Table 7. Comparison of the SOTA models on three segmentation tasks. Mask DINO outperforms all existing models. "TTA" means test-time-augmentation. "O365" denotes the Objects365 [26] dataset. + +constrained masked attention, we predict boxes and then use them in deformable attention to extract query features. Therefore, our box-oriented deformable attention also works well with "stuff" categories, which makes our unified model simple and efficient. In addition, we improve the mask $\mathrm{AP}_{pan}^{Th}$ by 2.6 to 44.3 AP, which is 0.6 higher than the specialized instance segmentation model Mask2Fomer (43.7 AP). + +Semantic segmentation. In Table 5 and 6, we show the performance of semantic segmentation with a ResNet-50 backbone. We use 100 queries for these small datasets. We outperform Mask2Former on both ADE20K and Cityscapes by 1.6 and $0.6\mathrm{mIoU}$ on the reported performance. + +# 4.2. Comparison with SOTA Models + +In Table 7, we compare Mask DINO with SOTA models on three image segmentation tasks to show its scalability. We use the SwinL [20] backbone and pre-train DINO on the Objects365 [26] detection dataset. Even without using extra data, we outperform Mask2Former on all three tasks, especially on instance segmentation (+2.5 AP). As Mask DINO is an extension of DINO, the pre-trained DINO model can be used to fine-tune Mask DINO for segmentation tasks. After fine-tuning Mask DINO on the corresponding tasks, we + +achieve the best results on instance (54.5 AP), panoptic (59.4 PQ), and semantic (60.8 mIoU) segmentation among model under one billion parameters. Compared to SwinV2-G [19], we significantly reduce the model size to 1/15 and backbone pre-training dataset to 1/5. Our detection pre-training also significantly helps all segmentation tasks including panoptic and semantic with "stuff" categories. However, previous specialized segmentation models such as Mask2Former can not use detection datasets and adding a detection head to it results in poor performance as shown in Table 1, which severely limits the data scalability. By unifying four tasks in one model, we only need to pre-train one model on a large-scale dataset and finetune on all tasks for 10 to 20 epochs (Mask2Former needs 100 epochs), which is more computationally efficient and simpler in model design. + +# 4.3. Ablation Studies + +We conduct ablation studies using a ResNet-50 backbone to analyze Mask DINO on COCO val2017. Unless otherwise stated, our experiments are based on object detection and instance segmentation without Mask-enhanced anchor box initialization. + +Query selection. Table 8 shows the results of our query selection for instance segmentation, where we additionally pro + +
test layer#Mask DINOMask2Former
layer 039.6(+38.5)1.1
layer 344.042.3
layer 645.943.3
layer 946.043.7
+ +Table 8. Effectiveness of our query selection for mask initialization. We evaluate the instance segmentation performance from different decoder layers in the same model after training for 50 epochs. + +
TasksBox APMask AP
BoxMask
✓50.1-
✓-43.3
✓✓50.5(+0.4)46.0(+2.7)
+ +Table 11. Task comparison under the 50-epoch setting. We train the same Mask DINO with different tasks and validate that box and mask can achieve mutual cooperation. + +
layer#w/o MEw ME
BoxMaskBoxMask
layer 039.625.639.841.2(+15.6)
layer 946.050.546.351.7(+1.2)
+ +Table 9. Comparison of our model with and without Mask-enhanced anchor box initialization (ME). ME enhances anchor box initialization and improves final detection performance. Trained for 50 epochs. + +
Decoder layer#Box APMask AP
343.140.7
644.341.1
944.541.4
1244.841.1
+ +Table 12. Decoder layer number comparison under the 12-epoch setting. Mask DINO benefits from more decoders, while DINO's performance will decrease with 9 decoders. + +
Feature scalebox APmask AP
single scale(1/8)45.845.1
3 scales50.545.8
4 scales50.546.0
+ +Table 10. Comparison of multi-scale features for Transformer decoder under the 50-epoch setting. Both detection and segmentation benefit from more feature scales. + +
MatchingBox APMask AP
BoxMask
✓44.440.5
✓40.238.4
✓✓44.541.4
+ +Table 13. Matching method comparison under the 12-epoch setting. We train both tasks together but use different matching methods to verify the effectiveness of hybrid matching. + +
Box APMask AP
Mask DINO (ours)45.741.4
- Mask-enhanced anchor box initialization44.5(-1.2)41.4
- Unified query selection for masks43.6(-2.1)40.3(-1.1)
- Unified denoising for masks44.4(-1.3)40.3(-1.1)
- Hybrid matching44.9(-0.8)40.5(-0.9)
- remove all the above41.7(-4.0)38.5(-2.7)
+ +Table 14. Effectiveness of the proposed components under the 12-epoch setting. + +vide the performance of different decoder layers in one single model. Mask2Former also predicts the masks of learnable queries as initial region proposals. However, their performance lags behind Mask DINO by a large margin (-38.5AP). With our effective query selection scheme, the mask performance achieves 39.6 AP without using the decoder. In addition, our mask performance at layer six is already comparable to the final results with 9 layers. In Table 9, we show that in query selection the predicted box is inferior to mask, which indicates segmentation is easier to learn in the initial stage. Therefore, our proposed mask-enhanced box initialization enhances boxes with masks in query selection to provide better anchor boxes (+15.6 AP) for the decoder, which results in +1.2 AP improvement in the final detection performance. + +Feature scales. Mask2Former [3] shows that concatenating multi-scale features as input to Transformer decoder layers does not improve the segmentation performance. However, in Table 10, Mask DINO shows that using more feature scales in the decoder consistently improves the performance. + +Object detection and segmentation help each other. To validate task cooperation in Mask DINO, we use the same model but train different tasks and report the 12 epoch and 50 epoch results. As shown in Table 11, only training one task will lead to a performance drop. Although only training object detection results in faster convergence in the early stage for box prediction, the final performance is still inferior to training both tasks together. + +Decoder layer number. In DINO, increasing the decoder layer number to nine will decrease the performance of box. In Table 12, the result indicates that increasing the number of + +decoder layers will contribute to both detection and segmentation in Mask DINO. We hypothesize that the multi-task training become more complex and require more decoders to learn the needed mapping function. + +Matching. In Table 13, we show that only using boxes or masks to perform bipartite matching is not optimal in Mask DINO. A unified matching objective makes the optimization more consistent. In addition, after removing all the proposed components, both detection and segmentation performance drop by a large margin. This result indicates that if we trivially add detection and segmentation tasks in one DETR-based model, the features are not aligned for detection and segmentation tasks to achieve mutual cooperation. + +We also present visualization analysis in Appendix ?? + +# 5. Conclusion + +In this paper, we have presented Mask DINO as a unified Transformer-based framework for both object detection and image segmentation. Conceptually, Mask DINO is a natural extension of DINO from detection to segmentation with minimal modifications on some key components. Mask DINO outperforms previous specialized models and achieves the best results on all three segmentation tasks (instance, panoptic, and semantic) among models under one billion parameters. Moreover, Mask DINO shows that detection and segmentation can help each other in query-based models. In particular, Mask DINO enables semantic and panoptic segmentation to benefit from a better visual representation pre-trained on a large-scale detection dataset. We hope Mask DINO can provide insights for enabling task cooperation and data cooperation towards designing a universal model for more vision tasks. + +Limitations: Different segmentation tasks fail to achieve mutual assistance in Mask DINO in COCO panoptic segmentation. For example, in COCO panoptic segmentation, the mask AP still lags behind the model only trained with instances. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3, 4, 6 +[2] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 1, 2, 3, 5, 6 +[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. 2022. 1, 2, 3, 4, 6, 7, 8 +[4] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. 2021. 1, 2, 3, 4 +[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5 +[6] Xianzhi Du, Barret Zoph, Wei-Chih Hung, and Tsung-Yi Lin. Simple training strategies and model scaling for object detection. arXiv preprint arXiv:2107.00057, 2021. 6 +[7] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 1, 6 +[8] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 6 +[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2, 3, 5, 6 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[11] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. arXiv preprint arXiv:2211.06220, 2022. 6 +[12] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. SeMask: Semantically Masked Transformers for Semantic Segmentation. arXiv preprint arXiv:2112.12782, 2021. 1, 7 +[13] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6399-6408, 2019. 3 +[14] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 2 +[15] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. DN-DETR: Accelerate DETR Training by Introducing Query DeNoising. arXiv preprint arXiv:2203.01305, 2022. 1, 2, 3, 5 +[16] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Tong Lu, and Ping Luo. Panoptic SegFormer. arXiv preprint arXiv:2109.03814, 2021. 6, 7 +[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5 +[18] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 3, 4 +[19] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin Transformer V2: Scaling Up Capacity and Resolution. arXiv preprint arXiv:2111.09883, 2021. 7 +[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 6, 7 +[21] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 3 +[22] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional DETR for Fast Training Convergence. arXiv preprint arXiv:2108.06152, 2021. 4 +[23] Zipeng Qin, Jianbo Liu, Xiaolin Zhang, Maoqing Tian, Aojun Zhou, Shuai Yi, and Hongsheng Li. Pyramid Fusion Transformer for Semantic Segmentation. arXiv preprint arXiv:2201.04019, 2022. 1 +[24] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 4 +[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3 +[26] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 7 + +[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1 +[28] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2, 4 +[29] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. arXiv preprint arXiv:2109.07107, 2021. 4 +[30] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8826, 2019. 3 +[31] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 7 +[32] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. arXiv preprint arXiv:2203.03605, 2022. 1, 2, 3, 4, 5, 6 +[33] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34, 2021. 1, 3 +[34] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5 +[35] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR 2021: The Ninth International Conference on Learning Representations, 2021. 1, 3, 4 \ No newline at end of file diff --git a/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/images.zip b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f75c19e357bbe96d06c2248b33db1f45693eaccf --- /dev/null +++ b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fc22c61fae9c3ea331fe3b276addcde50169199a3e49589f1c39035cff759be +size 440440 diff --git a/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/layout.json b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f6e4b2988a64817638f62138a539f69c2b93f365 --- /dev/null +++ b/2023/Mask DINO_ Towards a Unified Transformer-Based Framework for Object Detection and Segmentation/layout.json @@ -0,0 +1,6130 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "text", + "content": "Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "text", + "content": "Feng Li" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,3*†}" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Hao Zhang" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,3*†}" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Huaizhe Xu" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Shilong Liu" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Lei Zhang" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{3‡}" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Lionel M. Ni" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Heung-Yeung Shum" + }, + { + "bbox": [ + 124, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 190, + 432, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 190, + 432, + 203 + ], + "spans": [ + { + "bbox": [ + 161, + 190, + 432, + 203 + ], + "type": "text", + "content": "1The Hong Kong University of Science and Technology." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 463, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 463, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 463, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 203, + 463, + 217 + ], + "type": "text", + "content": "Dept. of CST., BNRist Center, Institute for AI, Tsinghua University." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 175, + 217, + 418, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 217, + 418, + 231 + ], + "spans": [ + { + "bbox": [ + 175, + 217, + 418, + 231 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 175, + 217, + 418, + 231 + ], + "type": "text", + "content": "International Digital Economy Academy (IDEA)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 232, + 465, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 232, + 465, + 245 + ], + "spans": [ + { + "bbox": [ + 129, + 232, + 465, + 245 + ], + "type": "text", + "content": "4The Hong Kong University of Science and Technology (Guangzhou)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 255, + 466, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 255, + 466, + 264 + ], + "spans": [ + { + "bbox": [ + 135, + 255, + 466, + 264 + ], + "type": "text", + "content": "{fiay,hzhangcx,hxubr}@connect.ust.hk {liusl20}@mails.tsinghua.edu.cn {leizhang}@idea.edu.cn {ni,hshum}@ust.hk" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 293, + 192, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 293, + 192, + 306 + ], + "spans": [ + { + "bbox": [ + 143, + 293, + 192, + 306 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 318, + 290, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 290, + 569 + ], + "type": "text", + "content": "In this paper we present Mask DINO, a unified object detection and segmentation framework. Mask DINO extends DINO (DETR with Improved Denoising Anchor Boxes) by adding a mask prediction branch which supports all image segmentation tasks (instance, panoptic, and semantic). It makes use of the query embeddings from DINO to dot-product a high-resolution pixel embedding map to predict a set of binary masks. Some key components in DINO are extended for segmentation through a shared architecture and training process. Mask DINO is simple, efficient, and scalable, and it can benefit from joint large-scale detection and segmentation datasets. Our experiments show that Mask DINO significantly outperforms all existing specialized segmentation methods, both on a ResNet-50 backbone and a pre-trained model with SwinL backbone. Notably, Mask DINO establishes the best results to date on instance segmentation (54.5 AP on COCO), panoptic segmentation (59.4 PQ on COCO), and semantic segmentation (60.8 mIoU on ADE20K) among models under one billion parameters. Code is available at https://github.com/IDEA-Research/MaskDINO." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 614, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 614, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 614, + 288, + 675 + ], + "type": "text", + "content": "Object detection and image segmentation are fundamental tasks in computer vision. Both tasks are concerned with localizing objects of interest in an image but have different levels of focus. Object detection is to localize objects of interest and predict their bounding boxes and category" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 293, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 547, + 365 + ], + "type": "text", + "content": "labels, whereas image segmentation focuses on pixel-level grouping of different semantics. Moreover, image segmentation encompasses various tasks including instance segmentation, panoptic segmentation, and semantic segmentation with respect to different semantics, e.g., instance or category membership, foreground or background category." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 366, + 548, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 548, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 548, + 533 + ], + "type": "text", + "content": "Remarkable progress has been achieved by classical convolution-based algorithms developed for these tasks with specialized architectures, such as Faster RCNN [24] for object detection, Mask RCNN [9] for instance segmentation, and FCN [21] for semantic segmentation. Although these methods are conceptually simple and effective, they are tailored for specialized tasks and lack the generalization ability to address other tasks. The ambition to bridge different tasks gives rise to more advanced methods like HTC [2] for object detection and instance segmentation and Panoptic FPN [14], K-net [33] for instance, panoptic, and semantic segmentation. Task unification not only helps simplify algorithm development but also brings in performance improvement in multiple tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 548, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 548, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 548, + 689 + ], + "type": "text", + "content": "Recently, DETR-like [1] models developed based on Transformers [27] have achieved inspiring progress on many detection and segmentation tasks. As an end-to-end object detector, DETR adopts a set-prediction objective and eliminates hand-crafted modules such as anchor design and nonmaximum suppression. Although DETR addresses both the object detection and panoptic segmentation tasks, its segmentation performance is still inferior to classical segmentation models. To improve the detection and segmentation performance of Transformer-based models, researchers have developed specialized models for object detection [15, 18, 32, 35], image segmentation [3, 4, 33], instance segmentation [7], panoptic segmentation [23], and semantic segmentation [12]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Among the efforts to improve object detection, DINO [32] takes advantage of the dynamic anchor box for" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 683, + 126, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 126, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 126, + 693 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 693, + 269, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 269, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 269, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 59, + 693, + 269, + 703 + ], + "type": "text", + "content": " Work done when Feng Li and Hao Zhang were interns at IDEA." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3041" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "type": "text", + "content": "mulation from DAB-DETR [18] and query denoising training from DN-DETR [15], and further achieves the SOTA result on the COCO object detection leaderboard for the first time as a DETR-like model. Similarly, for improving image segmentation, MaskFormer [4] and Mask2Former [3] propose to unify different image segmentation tasks using query-based Transformer architectures to perform mask classification. Such methods have achieved remarkable performance improvement on multiple segmentation tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 288, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 397 + ], + "type": "text", + "content": "However, in Transformer-based models, the best-performing detection and segmentation models are still not unified, which prevents task and data cooperation between detection and segmentation tasks. As an evidence, in CNN-based models, Mask-R-CNN [9] and HTC [2] are still widely acknowledged as unified models that achieve mutual cooperation between detection and segmentation to achieve superior performance than specialized models. Though we believe detection and segmentation can help each other in a unified architecture in Transformer-based models, the results of simply using DINO for segmentation and using Mask2Former for detection indicate that they can not do other tasks well, as shown in Table 1 and 2. Moreover, trivial multi-task training can even hurt the performance of the original tasks. It naturally leads to two questions: 1) why cannot detection and segmentation tasks help each other in Transformer-based models? and 2) is it possible to develop a unified architecture to replace specialized ones?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 400, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 686 + ], + "type": "text", + "content": "To address these problems, we propose Mask DINO, which extends DINO with a mask prediction branch in parallel with DINO's box prediction branch. Inspired by other unified models [3, 4, 28] for image segmentation, we reuse content query embeddings from DINO to perform mask classification for all segmentation tasks on a high-resolution pixel embedding map (1/4 of the input image resolution) obtained from the backbone and Transformer encoder features. The mask branch predicts binary masks by simply dot-producting each content query embedding with the pixel embedding map. As DINO is a detection model for region-level regression, it is not designed for pixel-level alignment. To better align features between detection and segmentation, we also propose three key components to boost the segmentation performance. First, we propose a unified and enhanced query selection. It utilizes encoder dense prior by predicting masks from the top-ranked tokens to initialize mask queries as anchors. In addition, we observe that pixel-level segmentation is easier to learn in the early stage and propose to use initial masks to enhance boxes, which achieves task cooperation. Second, we propose a unified denoising training for masks to accelerate segmentation training. Third, we use a hybrid bipartite matching for more accurate and consistent matching from ground truth to both boxes and masks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Mask DINO is conceptually simple and easy to implement under the DINO framework. To summarize, our contri" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "text", + "content": "butions are three-fold. 1) We develop a unified Transformer-based framework for both object detection and segmentation. As the framework is extended from DINO, by adding a mask prediction branch, it naturally inherits most algorithm improvements in DINO including anchor box-guided cross attention, query selection, denoising training, and even a better representation pre-trained on a large-scale detection dataset. 2) We demonstrate that detection and segmentation can help each other through a shared architecture design and training method. Especially, detection can significantly help segmentation tasks, even for segmenting background \"stuff\" categories. Under the same setting with a ResNet-50 backbone, Mask DINO outperforms all existing models compared to DINO (+0.8 AP on COCO detection) and Mask2Former (+2.6 AP, +1.1 PQ, and +1.5 mIoU on COCO instance, COCO panoptic, and ADE20K semantic segmentation). 3) We also show that, via a unified framework, segmentation can benefit from detection pre-training on a large-scale detection dataset. After detection pre-training on the Objects365 [26] dataset with a SwinL [20] backbone, Mask DINO significantly improves all segmentation tasks and achieves the best results on instance (54.5 AP on COCO), panoptic (59.4 PQ on COCO), and semantic (60.8 mIoU on ADE20K) segmentation among models under one billion parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 402, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 629 + ], + "type": "text", + "content": "Detection: Mainstream detection algorithms have been dominated by convolutional neural network-based frameworks, until recently Transformer-based detectors [1, 15, 18, 32] achieve great progress. DETR [1] is the first end-to-end and query-based Transformer object detector, which adopts a set-prediction objective with bipartite matching. DAB-DETR [18] improves DETR by formulating queries as 4D anchor boxes and refining predictions layer by layer. DN-DETR [15] introduces a denoising training method to accelerate convergence. Based on DAB-DETR and DN-DETR, DINO [32] proposes several new improvements on denoising and anchor refinement and achieves new SOTA results on COCO detection. Despite the inspiring progress, DETR-like detection models are not competitive for segmentation. Vanilla DETR incorporates a segmentation head in its architecture. However, its segmentation performance is inferior to specialized segmentation models and only shows the feasibility of DETR-like detection models to deal with detection and segmentation simultaneously." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Segmentation: Segmentation mainly includes instance, semantic, and panoptic segmentation. Instance segmentation is to predict a mask and its corresponding category for each object instance. Semantic segmentation requires to classify each pixel including the background into different semantic categories. Panoptic segmentation [14] unifies the instance and semantic segmentation tasks and predicts a mask for" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3042" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "each object instance or background segment. In the past few years, researchers have developed specialized architectures for the three tasks. For example, Mask-RCNN [9] and HTC [2] can only deal with instance segmentation because they predict the mask of each instance based on its box prediction. FCN [21] and U-Net [25] can only perform semantic segmentation since they predict one segmentation map based on pixel-wise classification. Although models for panoptic segmentation [13, 30] unifies the above two tasks, they are usually inferior to specialized instance and semantic segmentation models. Until recently, some image segmentation models [3, 4, 33] are developed to unify the three tasks with a universal architecture. For instance, Mask2Former [3] improves MaskFormer [4] by introducing masked-attention to Transformer. Mask2Former has a similar architecture as DETR to probe image features with learnable queries but differs in using a different segmentation branch and some specialized designs for mask prediction. However, while Mask2Former shows a great success in unifying all segmentation tasks, it leaves object detection untouched and our empirical study shows that its specialized architecture design is not suitable for predicting boxes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 335, + 289, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 289, + 563 + ], + "type": "text", + "content": "Unified Methods: As both object detection and segmentation are concerned with localizing objects, they naturally share common model architectures and visual representations. A unified framework not only helps simplify the algorithm development effort, but also allows to use both detection and segmentation data to improve representation learning. There have been several previous works to unify segmentation and detection tasks, e.g., Mask RCNN [9], HTC [2], and DETR [1]. Mask RCNN extends Faster RCNN and pools image features from Region Of Interest (ROI) proposed by RPN. HTC further proposes an interleaved way of predicting boxes and masks. However, these two models can only perform instance segmentation. DETR predicts boxes and masks together in an end-to-end manner. However, its segmentation performance largely lags behind other models. According to Table 2, adding DETR's segmentation head to DINO results in inferior instance segmentation results. How to attain mutual assistance between segmentation and detection has long been an important problem to solve." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 582, + 126, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 126, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 126, + 594 + ], + "type": "text", + "content": "3. Mask DINO" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 288, + 713 + ], + "type": "text", + "content": "Mask DINO is an extension of DINO [32]. On top of content query embeddings, DINO has two branches for box prediction and label prediction. The boxes are dynamically updated and used to guide the deformable attention in each Transformer decoder. Mask DINO adds another branch for mask prediction and minimally extends several key components in detection to fit segmentation tasks. To better understand Mask DINO, we start by briefly reviewing DINO and then introduce Mask DINO." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 428, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 428, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 428, + 83 + ], + "type": "text", + "content": "3.1. Preliminaries: DINO" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 90, + 547, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 270 + ], + "type": "text", + "content": "DINO is a typical DETR-like model, which is composed of a backbone, a Transformer encoder, and a Transformer decoder. The framework is shown in Fig. 1 (the blue-shaded part without red lines). Following DAB-DETR [18], DINO formulates each positional query in DETR as a 4D anchor box, which is dynamically updated through each decoder layer. Note that DINO uses multi-scale features with deformable attention [35]. Therefore, the updated anchor boxes are also used to constrain deformable attention in a sparse and soft way. Following DN-DETR [15], DINO adopts denoising training and further develops contrastive denoising to accelerate training convergence. Moreover, DINO proposes a mixed query selection scheme to initialize positional queries in the decoder and a look-forward-twice method to improve box gradient back-propagation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 276, + 545, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 276, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 276, + 545, + 301 + ], + "type": "text", + "content": "3.2. Why a universal model has not replaced the specialized models in DETR-like models?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "content": "Remarkable progress has been achieved by Transformer-based detectors and segmentation models. For instance, DINO [32] and Mask2Former [3] have achieved the best results on COCO detection and panoptic segmentation, respectively. Inspired by such progress, we attempted to simply extend these specialized models for other tasks but found that the performance of other tasks lagged behind the original ones by a large margin, as shown in Table 1 and 2. It seems that trivial multi-task training even hurts the performance of the original task. However, in convolution-based models, it has shown effective and mutually beneficial to combine detection and instance segmentation tasks. For example, detection models with Mask R-CNN head [9] is still ranked the first on the COCO instance segmentation. We will take DINO and Mask2Former as examples to discuss the challenges in unifying Transformer-based detection and segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 548, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 689 + ], + "type": "text", + "content": "- What are the differences between specialized detection and segmentation models? Image segmentation is a pixel-level classification task, while object detection is a region-level regression task. In DETR-based model, the decoder queries are responsible for these tasks. For example, Mask2Former uses such decoder queries to dot-product the high-resolution feature maps to produce segmentation masks, while DINO uses them to regress boxes. However, as such queries in Mask2Former only have to compare per-pixel similarity with the image features, they may not be aware of the region-level position of each instance. On the contrary, queries in DINO are not designed to interact with such low-level features to learn pixel-level representation. Instead, they encode rich positional information and high-level semantics for detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "-Why cannot Mask2Former do detection well? The Transformer decoder of Mask2Former is designed for seg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3043" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 74, + 258, + 126 + ], + "blocks": [ + { + "bbox": [ + 90, + 74, + 258, + 126 + ], + "lines": [ + { + "bbox": [ + 90, + 74, + 258, + 126 + ], + "spans": [ + { + "bbox": [ + 90, + 74, + 258, + 126 + ], + "type": "table", + "html": "
ModelBox APMask AP
Mask2Former46.2*43.7
Mask2Former + detection head21.641.3
DINO50.7-
", + "image_path": "d9be5e4e3a24a9235b42c685a929e0a9837805dfe720b02c2dcbfd3f1fac9134.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 287, + 70, + 504, + 131 + ], + "blocks": [ + { + "bbox": [ + 74, + 132, + 273, + 164 + ], + "lines": [ + { + "bbox": [ + 74, + 132, + 273, + 164 + ], + "spans": [ + { + "bbox": [ + 74, + 132, + 273, + 164 + ], + "type": "text", + "content": "Table 1. Simply adding a detection head to Mask2Former results in low detection performance. * indicates the boxes are generated from the predicted masks. The generated boxes from Mask2Former are also inferior to DINO (-4.5 AP). The models are trained for 50 epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 287, + 70, + 504, + 131 + ], + "lines": [ + { + "bbox": [ + 287, + 70, + 504, + 131 + ], + "spans": [ + { + "bbox": [ + 287, + 70, + 504, + 131 + ], + "type": "table", + "html": "
ModelBox APMask AP
DINO + Mask2Former segmentation head49.940.2
DINO + DETR segmentation head (finetune DINO pretrained on COCO detection)-35.8
Mask2Former-43.7
", + "image_path": "350866490077f04293ede00113e27fe7432e8c9aa13183aa389ef7e85f1fbe3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 277, + 136, + 512, + 160 + ], + "lines": [ + { + "bbox": [ + 277, + 136, + 512, + 160 + ], + "spans": [ + { + "bbox": [ + 277, + 136, + 512, + 160 + ], + "type": "text", + "content": "Table 2. Simply adding a segmentation head to DINO results in low instance segmentation performance. The predicted masks from DINO are also inferior to Mask2Former (-3.5 AP). The models are trained for 50 epochs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 175, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 437 + ], + "type": "text", + "content": "mentation tasks and does not suit detection for three reasons. First, its queries follow the design in DETR [1] without being able to utilize better positional priors as studied in Conditional DETR [22], Anchor DETR [29], and DAB-DETR [18]. For example, its content queries are semantically aligned with the features from the Transformer encoder, whereas its positional queries are just learnable vectors as in vanilla DETR instead of being associated with a single-mode position1. If we remove its mask branch, it reduces to a variant of DETR [1], whose performance is inferior to recently improved DETR models. Second, Mask2Former adopts masked attention (multi-head attention with attention mask) in Transformer decoders. The attention masks predicted from a previous layer are of high resolution and used as hard-constraints for attention computation. They are neither efficient nor flexible for box prediction. Third, Mask2Former cannot explicitly perform box refinement layer by layer. Moreover, its coarse-to-fine mask refinement in decoders fails to use multi-scale features from the encoder. As shown in Table 1, the generated box AP from mask is 4.5 AP worse than DINO and trivial multi-task learning by adding a detection head is not working2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "text", + "content": "- Why cannot DETR/DINO do segmentation well? As shown in Table 2, simply 1) adding DETR's segmentation head or 2) adding Mask2Former's segmentation head result in inferior performance compared to Mask2Former. We analyze the reasons as follows. The reason for 1) is that DETR's segmentation head is not optimal. The vanilla DETR lets each query embedding dot-product with the smallest feature map to compute attention maps and then upsamples them to get the mask predictions. This design lacks an interaction between queries and larger feature maps from the backbone. In addition, the head is too heavy to use mask auxiliary loss for mask refinement. The reason for 2) is that features in improved detection models are not aligned with segmentation. For example, DINO inherits many designs from [18, 32, 35] like query formulation, denoising training, and query selection. However, these components are designed to strengthen region-level representation for detection, which is not optimal for segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 175, + 453, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 175, + 453, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 453, + 186 + ], + "type": "text", + "content": "3.3. Our Method: Mask DINO" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 192, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 547, + 277 + ], + "type": "text", + "content": "Mask DINO adopts the same architecture design for detection as in DINO with minimal modifications. In the Transformer decoder, Mask DINO adds a mask branch for segmentation and extends several key components in DINO for segmentation tasks. As shown in Fig. 1, the framework in the blue-shaded part is the original DINO model and the additional design for segmentation is marked with red lines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 284, + 420, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 420, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 420, + 296 + ], + "type": "text", + "content": "3.4. Segmentation branch" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "text", + "content": "Following other unified models [3,4,28] for image segmentation, we perform mask classification for all segmentation tasks. Note that DINO is not designed for pixel-level alignment as its positional queries are formulated as anchor boxes and its content queries are used to predict box offset and class membership. To perform mask classification, we adopt a key idea from Mask2Former [3] to construct a pixel embedding map which is obtained from the backbone and Transformer encoder features. As shown in Fig. 1, the pixel embedding map is obtained by fusing the " + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "text", + "content": " resolution feature map " + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "text", + "content": " from the backbone with an upsampled " + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "text", + "content": " resolution feature map " + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "text", + "content": " from the Transformer encoder. Then we dot-product each content query embedding " + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "inline_equation", + "content": "q_c" + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "text", + "content": " from the decoder with the pixel embedding map to obtain an output mask " + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 301, + 547, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 357, + 488, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 488, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 357, + 488, + 545, + 502 + ], + "type": "interline_equation", + "content": "m = q _ {c} \\otimes \\mathcal {M} \\left(\\mathcal {T} \\left(C _ {b}\\right) + \\mathcal {F} \\left(C _ {e}\\right)\\right), \\tag {1}", + "image_path": "0918f7cfce8f3508e48c27fc98dcac16d98bfb6f9ebb89540b7b342a67fae980.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "text", + "content": " is the segmentation head, " + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "text", + "content": " is a convolutional layer to map the channel dimension to the Transformer hidden dimension, and " + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "text", + "content": " is a simple interpolation function to perform 2x upsampling of " + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 304, + 509, + 547, + 581 + ], + "type": "text", + "content": ". This segmentation branch is conceptually simple and easy to implement in the DINO framework, as shown in Fig. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 587, + 510, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 510, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 510, + 601 + ], + "type": "text", + "content": "3.5. Unified and Enhanced Query Selection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "content": "Unified query selection for mask: Query selection has been widely used in traditional two-stage models [24] and many DETR-like models [32, 35] to improve detection performance. We further improve the query selection scheme in Mask DINO for segmentation tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "The encoder output features contain dense features, which can serve as better priors for the decoder. Therefore, we adopt three prediction heads (classification, detection, and segmentation) in the encoder output. Note that the three" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 664, + 287, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 683 + ], + "type": "text", + "content": "1We refer the interested readers to discussions in Sec. 3 in DAB-DETR [18]" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": "2We also notice there are issues in official Mask2Former Github (https://github.com/facebookresearch/Mask2Former/issues/43) that fail to make Mask2Former work well by adding a detection head." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3044" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 523, + 266 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 523, + 266 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 523, + 266 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 523, + 266 + ], + "type": "image", + "image_path": "ef954800f9933bcb1adf83ab89f27b43762ee8de0f7e3506e627564632a6095c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 299 + ], + "type": "text", + "content": "Figure 1. The framework of Mask DINO, which is based on DINO (the blue-shaded part) with extensions (the red part) for segmentation tasks. 'QS' and 'DN' are short for query selection and denoising training, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 317, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 289, + 437 + ], + "type": "text", + "content": "heads are identical to the decoder heads. The classification score of each token is considered as the confidence to select top-ranked features and feed them to the decoder as content queries. The selected features also regress boxes and dot-product with the high-resolution feature map to predict masks. The predicted boxes and masks will be supervised by the ground truth and are considered as initial anchors for the decoder. Note that we initialize both the content and anchor box queries in Mask DINO whereas DINO only initializes anchor box queries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "text", + "content": "Mask-enhanced anchor box initialization: As summarized in Sec 3.2, image segmentation is a pixel-level classification task while object detection is a region-level position regression task. Therefore, compared to detection, though segmentation is a more difficult task with fine-granularity, it is easier to learn in the initial stage. For example, masks are predicted by dot-producing queries with the high-resolution feature map, which only needs to compare per-pixel semantic similarity. However, detection requires to directly regress the box coordinates in an image. Therefore, in the initial stage after unified query selection, mask prediction is much more accurate than box (the qualitative AP comparison between mask prediction and box prediction in different stages is also shown in Table 8 and 9). Therefore, after unified query selection, we derive boxes from the predicted masks as better anchor box initialization for the decoder. By this effective task cooperation, the enhanced box initialization can bring in a large improvement to the detection performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 659, + 200, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 200, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 200, + 672 + ], + "type": "text", + "content": "3.6. Segmentation Micro Design" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "Unified denoising for mask: Query denoising in object detection has shown effective [15,32] to accelerate convergence and improve performance. It adds noises to ground-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 317, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 547, + 461 + ], + "type": "text", + "content": "truth boxes and labels and feed them to the Transformer decoder as noised positional queries and content queries. The model is trained to reconstruct ground truth objects given their noised versions. We also extend this technique to segmentation tasks. As masks can be viewed as a more fine-grained representation of boxes, box and mask are naturally connected. Therefore, we can treat boxes as a noised version of masks, and train the model to predict masks given boxes as a denoising task. The given boxes for mask prediction are also randomly noised for more efficient mask denoising training. The detailed noise and its hyperparameters used in our model are shown in Appendix ??" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "text", + "content": "Hybrid matching: Mask DINO, as in some traditional models [2, 9], predicts boxes and masks with two parallel heads in a loosely coupled manner. Hence the two heads can predict a pair of box and mask that are inconsistent with each other. To address this issue, in addition to the original box and classification loss in bipartite matching, we add a mask prediction loss to encourage more accurate and consistent matching results for one query. Therefore, the matching cost becomes " + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}\\mathcal{L}_{cls} + \\lambda_{box}\\mathcal{L}_{box} + \\lambda_{mask}\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls},\\mathcal{L}_{box}" + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "text", + "content": " are the classification, box, and mask loss and " + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 461, + 547, + 617 + ], + "type": "text", + "content": " are their corresponding weights. The detailed losses used in our model and their corresponding weights are shown in Appendix ??" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 632, + 388, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 388, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 388, + 645 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "We conduct extensive experiments and compare with several specialized models for four popular tasks including object detection, instance, panoptic, and semantic segmentation on COCO [17], ADE20K [34], and Cityscapes [5]. For all experiments, we use batch size 16 and A100 GPUs with" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3045" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 519, + 247 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 519, + 247 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 519, + 247 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 519, + 247 + ], + "type": "table", + "html": "
ModelEpochsQuery typeMask APBox AP\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP_{S}^{mask}\\)\\(AP_{M}^{mask}\\)\\(AP_{L}^{mask}\\)GFLOPSParamsFPS
ResNet-50 backbone
Mask-RCNN [6,8,9]400Dense anchors42.548.2--23.845.060.020740M10.3
HTC [2]36Dense anchors39.744.961.443.122.642.250.644180M5
QueryInst [7]36300 queries40.645.663.044.023.442.552.8--7.0
DINO-4scale [32]36900 queries-50.9-----24547M19.6
Mask2Former [3]12100 queries38.7-59.841.218.241.559.822644M8.2
Mask DINO (ours)12300 queries41.4(+2.7)45.762.944.621.144.261.428652M14.8
Mask DINO (ours)24300 queries44.2(+0.5)48.466.647.923.947.064.028652M14.8
Mask2Former* [3]50100 queries43.7\\(46.2^{\\dagger}\\)66.046.923.447.264.822644M8.2
Mask DINO (ours)50100 queries45.449.867.949.325.248.365.828052M15.2
Mask DINO (ours)50300 queries46.050.568.950.326.0(+2.6)49.3(+2.1)65.5(+0.7)28652M14.8
Mask \\(DINO^‡\\) (ours)50300 queries46.3(+2.6)51.7(+0.8)69.050.726.1(+2.7)49.3(+2.1)66.1(+1.3)28652M14.2
SwinL backbone
HTC++ [2,20]72dense anchors49.557.1--31.052.467.21470284M-
Mask2Former [3]100200 queries50.1---29.953.972.1868216M4.0
DINO [32]36900 queries-58.5-----1285217M8.1
Mask2Former100200 queries50.1---29.953.972.1868216M4.0
Mask DINO (ours)50300 queries52.158.376.557.632.955.472.51326223M6.1
Mask \\(DINO^‡\\) (ours)50300 queries52.3(+2.2)59.0(+0.5)76.657.833.155.472.61326223M5.6
", + "image_path": "8636c63def5ecaf9a050d0cacb9b3073b62e569e5edfc4b702d15d2108af365d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 122, + 332, + 473, + 475 + ], + "blocks": [ + { + "bbox": [ + 46, + 255, + 545, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 545, + 321 + ], + "type": "text", + "content": "Table 3. Results for Mask DINO and other object detection and instance segmentation models with ResNet-50 and SwinL backbone on COCO val2017 without extra data or tricks. Following DINO [32], we use ResNet-50 with four feature scales by default, and use five scales under large models with a SwinL backbone. We follow the common practice in DETR-like models to use 300 queries. * Mask2Former using 300 queries is not listed as its performance will degenerate when using 300 queries. † indicates the box AP is derived from mask prediction. ‡ we use the proposed mask-enhanced box initialization to further improve detection performance. We test the FPS and GFLOPS of Mask2Former and Mask DINO on the A100 GPU using detectron2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 122, + 332, + 473, + 475 + ], + "lines": [ + { + "bbox": [ + 122, + 332, + 473, + 475 + ], + "spans": [ + { + "bbox": [ + 122, + 332, + 473, + 475 + ], + "type": "table", + "html": "
ModelEpochsQuery typePQ\\( PQ^{Th} \\)\\( PQ^{St} \\)Box \\( AP^{Th}_{pan} \\)Mask \\( AP^{Th}_{pan} \\)
ResNet-50 backbone
DETR [1]500 + 25100 queries43.448.236-31.1
Panoptic Segformer [16]24353 queries49.654.442.4-41.7
Mask2Former* [3]50100 queries\\( 51.9/51.5^† \\)57.743.0-41.7
Mask DINO (ours)50100 queries52.358.343.247.743.7
Mask DINO (ours)50300 queries\\( 53.0(+1.1) \\)\\( 59.1(+1.4) \\)\\( 43.9(+0.9) \\)48.8\\( 44.3(+2.6) \\)
Mask DINO (ours)24300 queries51.557.342.646.442.8
Mask2Former [3]12100 queries46.952.538.4-37.2
Panoptic Segformer [16]12353 queries48.052.341.5--
Mask DINO (ours)12300 queries\\( 49.0(+1.0) \\)54.840.243.2\\( 40.4(+3.2) \\)
SwinL backbone
Mask2Former [3]100100 queries57.864.248.1-48.6
OneFormer [11]100150 queries57.964.448.0-49.0
Mask DINO (ours)50300 queries\\( 58.3(+0.5) \\)65.148.056.2\\( 50.6(+2.0) \\)
", + "image_path": "6dfc1b95729b2f564eca45f856575c606f43641026fd22504c634a0adfe661f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 483, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 46, + 483, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 545, + 506 + ], + "type": "text", + "content": "Table 4. Results for Mask DINO and other panoptic segmentation models with a ResNet-50 backbone on COCO val2017. * Mask2Former using 300 queries is not listed as its performance will degenerate when using 300 queries. † Our reproduced result." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 515, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 563 + ], + "type": "text", + "content": "40GB memory. We use a ResNet-50 [10] and a SwinL [20] backbone for our main results and SOTA model. Under ResNet-50, we use 4 A100 GPUs for all tasks without extra data. The implementation details are in Appendix ??" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 132, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 132, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 132, + 586 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Instance segmentation and object detection. In Table 3, we compare Mask DINO with other instance segmentation and object detection models. Mask DINO outperforms both the specialized models such as Mask2Former [3] and DINO [32] and hybrid models such as HTC [2] under the same setting. Especially, the instance segmentation results surpass the strong baseline Mask2Former by a large margin (+2.7 AP and +2.6 AP) on the 12-epoch and 50-epoch settings. Moreover, Mask DINO significantly improves the convergence speed, outperforming Mask2Former with less" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 515, + 547, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 547, + 611 + ], + "type": "text", + "content": "than half training epochs (44.2 AP in 24 epochs). In addition, after using mask-enhanced box initialization, our detection performance has been significantly improved (+1.2 AP), which even outperforms DINO by 0.8 AP. These results indicate that task unification is beneficial. Without bells and whistles, we achieve the best detection and instance segmentation performance among DETR-like model with a SwinL backbone without extra data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Panoptic segmentation. We compare Mask DINO with other models in Table 4. Mask DINO outperforms all previous best models on both the 12-epoch and 50-epoch settings by 1.0 PQ and 1.1 PQ, respectively. This indicates Mask DINO has the advantages of both faster convergence and superior performance. One interesting observation is that we outperform Mask2Former [3] in terms of both " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "PQ^{Th}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "PQ^{St}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ". However, instead of using dense and hard" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3046" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 294, + 114 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 294, + 114 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 294, + 114 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 294, + 114 + ], + "type": "table", + "html": "
ModelIterationsCrop sizemIoU (mean)mIoU (high)mIoU (reported)
Mask2Former [3]160k51246.146.547.2
Mask DINO (ours)160k51247.7(+1.6)48.7(+2.2)48.7(+1.6)
", + "image_path": "db08a011c771bdfb84cc85e305280ff71b5ca434a68a05b269044f82a36a7d39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 312, + 71, + 512, + 113 + ], + "blocks": [ + { + "bbox": [ + 71, + 118, + 289, + 153 + ], + "lines": [ + { + "bbox": [ + 71, + 118, + 289, + 153 + ], + "spans": [ + { + "bbox": [ + 71, + 118, + 289, + 153 + ], + "type": "text", + "content": "Table 5. Results for Mask DINO and Mask2Former with 100 queries using a ResNet-50 backbone on ADE20K val. We found the performance variance on this dataset is high and run three times to report both the mean and highest results for both models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 71, + 512, + 113 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 512, + 113 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 512, + 113 + ], + "type": "table", + "html": "
ModelIterationsmIoU (mean)mIoU (high)mIoU (reported)
Mask2Former [3]90k78.779.079.4
Mask DINO (ours)90k79.8(+1.1)80.0(+1.0)80.0(+0.6)
", + "image_path": "4c6f956a0d2c5dc667cddd5f7225928d7a612972665f9994b6d7ff297119151e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 122, + 163, + 473, + 381 + ], + "blocks": [ + { + "bbox": [ + 304, + 118, + 521, + 153 + ], + "lines": [ + { + "bbox": [ + 304, + 118, + 521, + 153 + ], + "spans": [ + { + "bbox": [ + 304, + 118, + 521, + 153 + ], + "type": "text", + "content": "Table 6. Results for Mask DINO and Mask2Former with 100 queries using a ResNet-50 backbone on Cityscapes val. We found the performance variance on this dataset is high and run three times to report both the mean and highest results for both models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 122, + 163, + 473, + 381 + ], + "lines": [ + { + "bbox": [ + 122, + 163, + 473, + 381 + ], + "spans": [ + { + "bbox": [ + 122, + 163, + 473, + 381 + ], + "type": "table", + "html": "
MethodParamsBackboneBackbone Pre-training DatasetDetection Pre-training Datasetval
w/o TTAw/ TTA
Instance segmentation on COCOAP
Mask2Former [3]216MSwinLIN-22K-14M-50.1-
Soft Teacher [31]284MSwinLIN-22K-14MO36551.952.5
SwinV2-G-HTC++ [19]3.0BSwinV2-GIN-22K-ext-70M [19]O36553.453.7
MasK DINO(Ours)223MSwinLIN-22K-14M-52.6-
MasK DINO(Ours)223MSwinLIN-22K-14MO36554.5(+1.1)-
Panoptic segmentation on COCOPQ
Panoptic SegFormer [16]-MSwinLIN-22K-14M-55.8-
Mask2Former [3]216MSwinLIN-22K-14M-57.8-
MasK DINO (ours)223MSwinLIN-22K-14M58.4(+0.6)-
MasK DINO (ours)223MSwinLIN-22K-14MO36559.4(+1.6)-
Semantic segmentation on ADE20KmIoU
Mask2Former [3]215MSwinLIN-22K-14M-56.157.3
SeMask-L MSFaPN-Mask2Former [12]-MSwinL-FaPNIN-22K-14M--58.2
SwinV2-G-UperNet [19]3.0BSwinV2-GIN-22K-ext-70M [19]-59.359.9
MasK DINO (ours)223MSwinLIN-22K-14M-56.6-
MasK DINO (ours)223MSwinLIN-22K-14MO36559.560.8(+0.9)
", + "image_path": "3c0b7d5f4eb49cf641ea1dcd0f4f0051e61c77fc46a3242cafdd799942445564.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 390, + 546, + 414 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 546, + 414 + ], + "type": "text", + "content": "Table 7. Comparison of the SOTA models on three segmentation tasks. Mask DINO outperforms all existing models. \"TTA\" means test-time-augmentation. \"O365\" denotes the Objects365 [26] dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 426, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 521 + ], + "type": "text", + "content": "constrained masked attention, we predict boxes and then use them in deformable attention to extract query features. Therefore, our box-oriented deformable attention also works well with \"stuff\" categories, which makes our unified model simple and efficient. In addition, we improve the mask " + }, + { + "bbox": [ + 46, + 426, + 288, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{pan}^{Th}" + }, + { + "bbox": [ + 46, + 426, + 288, + 521 + ], + "type": "text", + "content": " by 2.6 to 44.3 AP, which is 0.6 higher than the specialized instance segmentation model Mask2Fomer (43.7 AP)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": "Semantic segmentation. In Table 5 and 6, we show the performance of semantic segmentation with a ResNet-50 backbone. We use 100 queries for these small datasets. We outperform Mask2Former on both ADE20K and Cityscapes by 1.6 and " + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "inline_equation", + "content": "0.6\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": " on the reported performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 587, + 220, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 220, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 220, + 601 + ], + "type": "text", + "content": "4.2. Comparison with SOTA Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "In Table 7, we compare Mask DINO with SOTA models on three image segmentation tasks to show its scalability. We use the SwinL [20] backbone and pre-train DINO on the Objects365 [26] detection dataset. Even without using extra data, we outperform Mask2Former on all three tasks, especially on instance segmentation (+2.5 AP). As Mask DINO is an extension of DINO, the pre-trained DINO model can be used to fine-tune Mask DINO for segmentation tasks. After fine-tuning Mask DINO on the corresponding tasks, we" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 605 + ], + "type": "text", + "content": "achieve the best results on instance (54.5 AP), panoptic (59.4 PQ), and semantic (60.8 mIoU) segmentation among model under one billion parameters. Compared to SwinV2-G [19], we significantly reduce the model size to 1/15 and backbone pre-training dataset to 1/5. Our detection pre-training also significantly helps all segmentation tasks including panoptic and semantic with \"stuff\" categories. However, previous specialized segmentation models such as Mask2Former can not use detection datasets and adding a detection head to it results in poor performance as shown in Table 1, which severely limits the data scalability. By unifying four tasks in one model, we only need to pre-train one model on a large-scale dataset and finetune on all tasks for 10 to 20 epochs (Mask2Former needs 100 epochs), which is more computationally efficient and simpler in model design." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 611, + 406, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 406, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 406, + 623 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 689 + ], + "type": "text", + "content": "We conduct ablation studies using a ResNet-50 backbone to analyze Mask DINO on COCO val2017. Unless otherwise stated, our experiments are based on object detection and instance segmentation without Mask-enhanced anchor box initialization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "Query selection. Table 8 shows the results of our query selection for instance segmentation, where we additionally pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3047" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 70, + 211, + 126 + ], + "blocks": [ + { + "bbox": [ + 79, + 70, + 211, + 126 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 211, + 126 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 211, + 126 + ], + "type": "table", + "html": "
test layer#Mask DINOMask2Former
layer 039.6(+38.5)1.1
layer 344.042.3
layer 645.943.3
layer 946.043.7
", + "image_path": "bd8a4c040d48529c57317dba8b9f9adf9d44e5222bf2fb0355a9645b4a202eaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 130, + 242, + 156 + ], + "lines": [ + { + "bbox": [ + 47, + 130, + 242, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 242, + 156 + ], + "type": "text", + "content": "Table 8. Effectiveness of our query selection for mask initialization. We evaluate the instance segmentation performance from different decoder layers in the same model after training for 50 epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 70, + 159, + 178, + 206 + ], + "blocks": [ + { + "bbox": [ + 70, + 159, + 178, + 206 + ], + "lines": [ + { + "bbox": [ + 70, + 159, + 178, + 206 + ], + "spans": [ + { + "bbox": [ + 70, + 159, + 178, + 206 + ], + "type": "table", + "html": "
TasksBox APMask AP
BoxMask
✓50.1-
✓-43.3
✓✓50.5(+0.4)46.0(+2.7)
", + "image_path": "df21cf8bd00fb1e913ef5f86d72119d108728aa3d41c9ad77603341fdaaa1cae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 232, + 70, + 380, + 114 + ], + "blocks": [ + { + "bbox": [ + 47, + 211, + 200, + 234 + ], + "lines": [ + { + "bbox": [ + 47, + 211, + 200, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 200, + 234 + ], + "type": "text", + "content": "Table 11. Task comparison under the 50-epoch setting. We train the same Mask DINO with different tasks and validate that box and mask can achieve mutual cooperation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 232, + 70, + 380, + 114 + ], + "lines": [ + { + "bbox": [ + 232, + 70, + 380, + 114 + ], + "spans": [ + { + "bbox": [ + 232, + 70, + 380, + 114 + ], + "type": "table", + "html": "
layer#w/o MEw ME
BoxMaskBoxMask
layer 039.625.639.841.2(+15.6)
layer 946.050.546.351.7(+1.2)
", + "image_path": "e913167538067bb25a8d21539de144039e328b38d8cfd246bc9b3a9ab741f7ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 232, + 157, + 345, + 206 + ], + "blocks": [ + { + "bbox": [ + 245, + 118, + 381, + 150 + ], + "lines": [ + { + "bbox": [ + 245, + 118, + 381, + 150 + ], + "spans": [ + { + "bbox": [ + 245, + 118, + 381, + 150 + ], + "type": "text", + "content": "Table 9. Comparison of our model with and without Mask-enhanced anchor box initialization (ME). ME enhances anchor box initialization and improves final detection performance. Trained for 50 epochs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 232, + 157, + 345, + 206 + ], + "lines": [ + { + "bbox": [ + 232, + 157, + 345, + 206 + ], + "spans": [ + { + "bbox": [ + 232, + 157, + 345, + 206 + ], + "type": "table", + "html": "
Decoder layer#Box APMask AP
343.140.7
644.341.1
944.541.4
1244.841.1
", + "image_path": "f046fac77b53d44657d174ebb9b86970272932ee4155320609e26745a8c6737b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 405, + 70, + 526, + 114 + ], + "blocks": [ + { + "bbox": [ + 208, + 211, + 367, + 234 + ], + "lines": [ + { + "bbox": [ + 208, + 211, + 367, + 234 + ], + "spans": [ + { + "bbox": [ + 208, + 211, + 367, + 234 + ], + "type": "text", + "content": "Table 12. Decoder layer number comparison under the 12-epoch setting. Mask DINO benefits from more decoders, while DINO's performance will decrease with 9 decoders." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 405, + 70, + 526, + 114 + ], + "lines": [ + { + "bbox": [ + 405, + 70, + 526, + 114 + ], + "spans": [ + { + "bbox": [ + 405, + 70, + 526, + 114 + ], + "type": "table", + "html": "
Feature scalebox APmask AP
single scale(1/8)45.845.1
3 scales50.545.8
4 scales50.546.0
", + "image_path": "adbc7a911ab588fc4bb41c2a18aaf1b070f9cbe630c3a482ec173e16c3198c72.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 403, + 154, + 509, + 202 + ], + "blocks": [ + { + "bbox": [ + 391, + 118, + 541, + 143 + ], + "lines": [ + { + "bbox": [ + 391, + 118, + 541, + 143 + ], + "spans": [ + { + "bbox": [ + 391, + 118, + 541, + 143 + ], + "type": "text", + "content": "Table 10. Comparison of multi-scale features for Transformer decoder under the 50-epoch setting. Both detection and segmentation benefit from more feature scales." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 403, + 154, + 509, + 202 + ], + "lines": [ + { + "bbox": [ + 403, + 154, + 509, + 202 + ], + "spans": [ + { + "bbox": [ + 403, + 154, + 509, + 202 + ], + "type": "table", + "html": "
MatchingBox APMask AP
BoxMask
✓44.440.5
✓40.238.4
✓✓44.541.4
", + "image_path": "5ea1284595cd0cc00b2e7d7b2d63198d23dfb49e0ae4237319d7b6727849081b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 67, + 242, + 268, + 318 + ], + "blocks": [ + { + "bbox": [ + 376, + 206, + 535, + 230 + ], + "lines": [ + { + "bbox": [ + 376, + 206, + 535, + 230 + ], + "spans": [ + { + "bbox": [ + 376, + 206, + 535, + 230 + ], + "type": "text", + "content": "Table 13. Matching method comparison under the 12-epoch setting. We train both tasks together but use different matching methods to verify the effectiveness of hybrid matching." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 242, + 268, + 318 + ], + "lines": [ + { + "bbox": [ + 67, + 242, + 268, + 318 + ], + "spans": [ + { + "bbox": [ + 67, + 242, + 268, + 318 + ], + "type": "table", + "html": "
Box APMask AP
Mask DINO (ours)45.741.4
- Mask-enhanced anchor box initialization44.5(-1.2)41.4
- Unified query selection for masks43.6(-2.1)40.3(-1.1)
- Unified denoising for masks44.4(-1.3)40.3(-1.1)
- Hybrid matching44.9(-0.8)40.5(-0.9)
- remove all the above41.7(-4.0)38.5(-2.7)
", + "image_path": "19d138bed3ac5fc7828ef377966270fb9383b9e5134d012bb232d209d15e2a85.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 322, + 268, + 331 + ], + "lines": [ + { + "bbox": [ + 66, + 322, + 268, + 331 + ], + "spans": [ + { + "bbox": [ + 66, + 322, + 268, + 331 + ], + "type": "text", + "content": "Table 14. Effectiveness of the proposed components under the 12-epoch setting." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 342, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 289, + 521 + ], + "type": "text", + "content": "vide the performance of different decoder layers in one single model. Mask2Former also predicts the masks of learnable queries as initial region proposals. However, their performance lags behind Mask DINO by a large margin (-38.5AP). With our effective query selection scheme, the mask performance achieves 39.6 AP without using the decoder. In addition, our mask performance at layer six is already comparable to the final results with 9 layers. In Table 9, we show that in query selection the predicted box is inferior to mask, which indicates segmentation is easier to learn in the initial stage. Therefore, our proposed mask-enhanced box initialization enhances boxes with masks in query selection to provide better anchor boxes (+15.6 AP) for the decoder, which results in +1.2 AP improvement in the final detection performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 521, + 288, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 582 + ], + "type": "text", + "content": "Feature scales. Mask2Former [3] shows that concatenating multi-scale features as input to Transformer decoder layers does not improve the segmentation performance. However, in Table 10, Mask DINO shows that using more feature scales in the decoder consistently improves the performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 582, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 677 + ], + "type": "text", + "content": "Object detection and segmentation help each other. To validate task cooperation in Mask DINO, we use the same model but train different tasks and report the 12 epoch and 50 epoch results. As shown in Table 11, only training one task will lead to a performance drop. Although only training object detection results in faster convergence in the early stage for box prediction, the final performance is still inferior to training both tasks together." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 677, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 713 + ], + "type": "text", + "content": "Decoder layer number. In DINO, increasing the decoder layer number to nine will decrease the performance of box. In Table 12, the result indicates that increasing the number of" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 244, + 547, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 292 + ], + "type": "text", + "content": "decoder layers will contribute to both detection and segmentation in Mask DINO. We hypothesize that the multi-task training become more complex and require more decoders to learn the needed mapping function." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 292, + 547, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 547, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 547, + 400 + ], + "type": "text", + "content": "Matching. In Table 13, we show that only using boxes or masks to perform bipartite matching is not optimal in Mask DINO. A unified matching objective makes the optimization more consistent. In addition, after removing all the proposed components, both detection and segmentation performance drop by a large margin. This result indicates that if we trivially add detection and segmentation tasks in one DETR-based model, the features are not aligned for detection and segmentation tasks to achieve mutual cooperation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 403, + 539, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 403, + 539, + 416 + ], + "spans": [ + { + "bbox": [ + 318, + 403, + 539, + 416 + ], + "type": "text", + "content": "We also present visualization analysis in Appendix ??" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 438, + 379, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 379, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 379, + 451 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 462, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 652 + ], + "type": "text", + "content": "In this paper, we have presented Mask DINO as a unified Transformer-based framework for both object detection and image segmentation. Conceptually, Mask DINO is a natural extension of DINO from detection to segmentation with minimal modifications on some key components. Mask DINO outperforms previous specialized models and achieves the best results on all three segmentation tasks (instance, panoptic, and semantic) among models under one billion parameters. Moreover, Mask DINO shows that detection and segmentation can help each other in query-based models. In particular, Mask DINO enables semantic and panoptic segmentation to benefit from a better visual representation pre-trained on a large-scale detection dataset. We hope Mask DINO can provide insights for enabling task cooperation and data cooperation towards designing a universal model for more vision tasks." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": "Limitations: Different segmentation tasks fail to achieve mutual assistance in Mask DINO in COCO panoptic segmentation. For example, in COCO panoptic segmentation, the mask AP still lags behind the model only trained with instances." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3048" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "text", + "content": "[2] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 288, + 258 + ], + "type": "text", + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 288, + 294 + ], + "type": "text", + "content": "[4] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 361 + ], + "type": "text", + "content": "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 288, + 396 + ], + "type": "text", + "content": "[6] Xianzhi Du, Barret Zoph, Wei-Chih Hung, and Tsung-Yi Lin. Simple training strategies and model scaling for object detection. arXiv preprint arXiv:2107.00057, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "type": "text", + "content": "[7] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 508 + ], + "type": "text", + "content": "[8] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "type": "text", + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 554, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 288, + 599 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "type": "text", + "content": "[11] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. arXiv preprint arXiv:2211.06220, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[12] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. SeMask: Semantically Masked Transformers for Semantic Segmentation. arXiv preprint arXiv:2112.12782, 2021. 1, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6399-6408, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "type": "text", + "content": "[14] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 141, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 185 + ], + "type": "text", + "content": "[15] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. DN-DETR: Accelerate DETR Training by Introducing Query DeNoising. arXiv preprint arXiv:2203.01305, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 547, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 219 + ], + "type": "text", + "content": "[16] Zhiqi Li, Wenhai Wang, Enze Xie, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, Tong Lu, and Ping Luo. Panoptic SegFormer. arXiv preprint arXiv:2109.03814, 2021. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 547, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 275 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 547, + 319 + ], + "type": "text", + "content": "[18] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 321, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 547, + 365 + ], + "type": "text", + "content": "[19] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin Transformer V2: Scaling Up Capacity and Resolution. arXiv preprint arXiv:2111.09883, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 366, + 547, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 547, + 422 + ], + "type": "text", + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 547, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 466 + ], + "type": "text", + "content": "[21] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 467, + 547, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 547, + 511 + ], + "type": "text", + "content": "[22] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional DETR for Fast Training Convergence. arXiv preprint arXiv:2108.06152, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 512, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 547, + 555 + ], + "type": "text", + "content": "[23] Zipeng Qin, Jianbo Liu, Xiaolin Zhang, Maoqing Tian, Aojun Zhou, Shuai Yi, and Hongsheng Li. Pyramid Fusion Transformer for Semantic Segmentation. arXiv preprint arXiv:2201.04019, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 557, + 547, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 547, + 601 + ], + "type": "text", + "content": "[24] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 602, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 547, + 656 + ], + "type": "text", + "content": "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "type": "text", + "content": "[26] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3049" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 508 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "type": "text", + "content": "[28] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463-5474, 2021. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 174, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 174, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 288, + 206 + ], + "type": "text", + "content": "[29] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. arXiv preprint arXiv:2109.07107, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 288, + 262 + ], + "type": "text", + "content": "[30] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8826, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 288, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 288, + 318 + ], + "type": "text", + "content": "[31] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 319, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 288, + 373 + ], + "type": "text", + "content": "[32] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. arXiv preprint arXiv:2203.03605, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 375, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 375, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 49, + 375, + 288, + 407 + ], + "type": "text", + "content": "[33] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 409, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 409, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 409, + 288, + 453 + ], + "type": "text", + "content": "[34] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 454, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 454, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 49, + 454, + 288, + 508 + ], + "type": "text", + "content": "[35] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR 2021: The Ninth International Conference on Learning Representations, 2021. 1, 3, 4" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3050" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_content_list.json b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..13b450db048382524cd3a67935e0c9c97e1af5a8 --- /dev/null +++ b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_content_list.json @@ -0,0 +1,1376 @@ +[ + { + "type": "text", + "text": "Mask-free OVIS: Open-Vocabulary Instance Segmentation without Manual Mask Annotations", + "text_level": 1, + "bbox": [ + 189, + 130, + 781, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vibashan VS\\*, Ning Yu†, Chen Xing†, Can Qin‡, Mingfei Gao†, Juan Carlos Niebles†, Vishal M. Patel\\*, Ran Xu†", + "bbox": [ + 232, + 202, + 736, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Johns Hopkins University, ‡Northeastern University, †Salesforce Research", + "bbox": [ + 189, + 239, + 779, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{vvishnu2, vpatel36}@jhu.edu, qin.ca@northeastern.edu,", + "bbox": [ + 250, + 258, + 723, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ning.yu, cxing, jniebles, ran.xu}@salesforce.com", + "bbox": [ + 271, + 277, + 704, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing instance segmentation models learn task-specific information using manual mask annotations from base (training) categories. These mask annotations require tremendous human effort, limiting the scalability to annotate novel (new) categories. To alleviate this problem, Open-Vocabulary (OV) methods leverage large-scale image-caption pairs and vision-language models to learn novel categories. In summary, an OV method learns task-specific information using strong supervision from base annotations and novel category information using weak supervision from image-captions pairs. This difference between strong and weak supervision leads to overfitting on base categories, resulting in poor generalization towards novel categories. In this work, we overcome this issue by learning both base and novel categories from pseudomask annotations generated by the vision-language model in a weakly supervised manner using our proposed Mask-free OVIS pipeline. Our method automatically generates pseudo-mask annotations by leveraging the localization ability of a pre-trained vision-language model for objects present in image-caption pairs. The generated pseudomask annotations are then used to supervise an instance segmentation model, freeing the entire pipeline from any labour-expensive instance-level annotations and overfitting. Our extensive experiments show that our method trained with just pseudo-masks significantly improves the mAP scores on the MS-COCO dataset and OpenImages dataset compared to the recent state-of-the-art methods trained with manual masks. Codes and models are provided in https://vibashan.github.io/ovis-web/.", + "bbox": [ + 75, + 345, + 473, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 809, + 209, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instance segmentation is a challenging task as it requires models to detect objects in an image while also precisely", + "bbox": [ + 76, + 835, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a49cee1ae68a616cd58806ea83eadc22c8d47a685d2fb4326a991229a3386be0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 325, + 890, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/919bacb6b2095f72cc721d1ac5f5ca9082c6a24b847e596aab35e6a4e33243ac.jpg", + "image_caption": [ + "Figure 1. A) Previous Methods: Learn task-specific information (detection/segmentation) in a fully-supervised manner and novel category information with weak supervision. During training, this difference in strong and weak supervision signals leads to overfitting and requires expensive base annotations. B) Our method: Given image-caption pairs, we generate pseudo-annotations for both base and novel categories under weak supervision, solving the problems of labour-expensive annotation and overfitting." + ], + "image_footnote": [], + "bbox": [ + 503, + 428, + 890, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "segment each object at the pixel-level. Although the rise of deep neural networks has significantly boosted the state-of-the-art instance segmentation performance [7, 15, 42], these methods are still trained for a pre-defined set of object categories and are data-hungry [33]. Particularly, one needs to manually annotate thousands of instance-level masks for each object category, which takes around 78 seconds per instance mask [3]. If we look at this quantitatively on Open Images [21], a large-scale dataset with 2.1M instance-level mask annotations requires around 5 years of human labour. Even after extensive annotation, these training datasets are still limited to a small number of categories and segmenting objects from a novel category requires further annotation. Therefore, it is difficult to scale up existing methods to segment a large number of categories due to intensive labour.", + "bbox": [ + 496, + 640, + 893, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Open-Vocabulary (OV) methods have gained much attention due to their success in detecting [11, 14,", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Vibashan VS interned at Salesforce Research. Primary contact: vvishnu2@jhu.com", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23539", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58b34a10e529132414d0d16d8145a05ece8d697b3e81726db38591b9fcc86259.jpg", + "image_caption": [ + "Figure 2. RPN is supervised using bounding box annotations from COCO base and WSPN is supervised using image-labels from COCO base. A) WSPN produces better quality proposals for novel object categories compared to fully-supervised RPN. B) WSPN consistently produces better recall for all COCO novel categories than RPN." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 580, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cbc981d08377389e6e16d676f69b6fdb3338f429e6c728d6041248ef22bb903b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 95, + 887, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "46, 52] and segmenting [17] novel categories beyond base (training) categories. An OV method learns task-specific information (detection/segmentation) from base categories that have manual instance-level bounding boxes or masks and learns novel category information from the pre-trained Vision-Language Model (VLM) [19, 34] (see Fig. 1). All these methods produce promising results on novel categories by leveraging different forms of weak supervision such as caption pretaining [46, 50], knowledge distillation [14, 52] and pseudo-labelling [17, 28]. However, all above-mentioned OV methods still rely on the manually-annotated base categories to improve their performances on novel categories. Without fine-tuning on base categories, existing OV methods lack task/domain specific knowledge and the performances on novel categories will be affected [14, 46].", + "bbox": [ + 75, + 354, + 468, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although manual instance-level annotations of base categories are critical to open-vocabulary segmentation methods, we find empirically that such fully-supervised information causes OV methods to overfit to base categories, leading to a higher failure rate when evaluated on novel categories. Specifically, OV methods utilize a region proposal network (RPN) [37] supervised with bounding box annotations obtained from the base categories to generate a set of bounding box proposals for all objects categories in a given image [46]. The feature representation from these object proposals is later matched with text embedding to learn the visual-semantic space for base and novel categories [46]. Therefore, the quality of proposals generated for novel object categories plays a key role in determining the performance in later stages. However, from our experiments, we find that many objects of novel categories wouldn't be included in such proposals due to the RPN's overfitting to base categories. Fig. 2 (A) - Top gives some examples where the RPN trained with COCO base categories fails to generate high-quality region proposals for novel categories such as elephant, cat and skateboard. Therefore, a fully", + "bbox": [ + 75, + 584, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "supervised proposal network is a bottleneck in OV pipeline due to its poor generalization towards novel categories.", + "bbox": [ + 498, + 354, + 890, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the aforementioned observations of poor generalization, we raise the question of whether we can improve the generalization by using weak supervision instead of relying on strong supervision from manual base annotations. If so, we can reduce overfitting to the base categories and the requirement for costly instance-level human annotations can be entirely removed from the pipeline. Our preliminary experiments give us some hope. Our experiments show that if we train a weakly-supervised proposal network (WSPN) with image-level annotations instead of box-level annotations, the region proposals it generates can better generalize to novel objects. As shown in Fig. 2 A), the novel objects that the RPN proposals miss are covered by WSPN. Fig. 2 B) shows WSPN proposals have consistently better average recall than RPN for all COCO novel categories, indicating that WSPN proposals are more likely to cover the ground-truth bounding boxes of novel objects.", + "bbox": [ + 496, + 386, + 892, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by these observations, we propose open-vocabulary segmentation without manual mask annotations. We do not use any human-provided box-level or pixel-level annotations during the training of our method. We first train a simple WSPN model with image-level annotations on base categories as a proposal generator to generate proposals for all objects given an image. Then, we adopt pre-trained vision-language models to select proposals as pseudo bounding boxes for novel objects. Given a novel object's text name, we can utilize the name as a text prompt to localize this object in an image with a pre-trained vision-language model. To obtain a more accurate pseudo-mask that covers the entire object, we conduct iterative masking with GradCAM [39] given the vision-language model. Finally, we train a weakly-supervised segmentation (WSS) [20] network with previously generated bounding box and GradCAM activation map to obtain pixel-level annotation.", + "bbox": [ + 496, + 645, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23540", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: (1) We propose a Mask-free OVIS pipeline where we produce manual-effort-free pseudo-mask annotations for base and novel instance segmentation using open-vocabulary and weakly supervised techniques. (2) We propose a novel pseudo-mask generation pipeline leveraging a pre-trained vision-language model to generate instance-level annotations. (3) Benefiting from pseudo-labels, our method sets up SOTA's for both detection and instance segmentation tasks compared to recent methods trained with manual masks on MS-COCO and OpenImages datasets.", + "bbox": [ + 76, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 271, + 218, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly-supervised learning (WSL) methods require less manual annotation effort and only rely on partial information about the task compared to fully-supervised methods. WSL applications include object detection [1, 16, 44, 47], segmentation [22,23,40], action recognition, etc [26,41,48]. The class activation map (CAM) [53] and grad-CAM [39] are popular approaches which exploits pre-trained network to generate saliency or activation map for target classes. Li et al. [26] enables weakly-supervised semantic segmentation via CAM-guided attention.", + "bbox": [ + 76, + 296, + 468, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Models. With the rise of large-scale text-pretraining with attention-based models [9, 49], vision-language models (VLM) have caught increasing attention due to their strong performance in downstream visual understanding tasks [27, 31, 34]. Early VLM models [27, 31] applied the pre-trained models with constraints of the class set. With the help of contrastive learning, CLIP [34] has enabled VLM in the wild with large-scale multimodal learning upon the 400M noisy data crawled from the web. Several methods have extended CLIP for high efficiency model training and cycle consistency [13, 24, 25]. ALBEF [25] utilizes cross-modal attention between image patches and text, leading to more grounded vision and language representation learning. Compared to CLIP, ALBEF excels in object localization, while CLIP performs better at zero-shot classification due to its larger training dataset.", + "bbox": [ + 76, + 448, + 472, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-Vocabulary. Open-Vocabulary methods [6, 10, 32, 36, 46, 50] scale up their vocabulary size for object detection or instance segmentation tasks by transferring knowledge from pre-trained vision-language models. Zareian et al. [46] proposed an open-vocabulary object detection method where a visual encoder is trained on image-caption pair to model object semantics and then transfer to zero-shot object detection. ViLD [14] and RegionCLIP [52] use pre-trained CLIP [34] to distill knowledge and enhance the vision-text embedding space effectively. Similarly, Huynh et al. [17] proposes a robust pseudo-labelling, for instance segmentation where a teacher model generates pseudo-label for novel categories and is later used to supervise a student model. Therefore, all these methods require manually", + "bbox": [ + 76, + 688, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "annotated base categories to improve their performance on novel categories. In contrast, our method generates pseudolabels for both base and novel categories by leveraging a pre-trained VLM, freeing the entire pipeline from human-provided instance-level annotations. Similar to our method, Gao et al. [11] generates pseudo-labels using pre-trained VLM, but their approach still uses a proposal generator that relies on human-provided bounding boxes for the base categories. Our method, on the other hand, requires no such annotations, making it completely free from any human-provided box-level or pixel-level annotations.", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Mask-free OVIS", + "text_level": 1, + "bbox": [ + 500, + 263, + 663, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our pipeline consists of two stages: (I) Pseudo-mask generation, and (II) Open-vocabulary instance segmentation. The goal of (I) is to generate box-level and pixel-level annotations by leveraging the region-text alignment property of a pre-trained vision-language model as illustrated in Fig. 3. In (II), we train a Mask-RCNN using the generated pseudo annotations.", + "bbox": [ + 496, + 282, + 893, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pseudo-Mask Generation", + "text_level": 1, + "bbox": [ + 500, + 391, + 732, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Region-Text Attention Scores. The inputs to the vision-language model are image $\\mathbf{I}$ and caption $\\mathbf{C} = \\{c_1, c_2, \\dots, c_{N_c}\\}$ pair, where $N_c$ is the number of words in the caption (including [CLS] and [SEP]) [9, 25]. In a VLM, a text encoder is utilized to get text representations $\\mathbf{T} \\in \\mathbb{R}^{N_C \\times d}$ and an image encoder is utilized to extract region representation $\\mathbf{R} \\in \\mathbb{R}^{N_R \\times d}$ , where $N_R$ is the number of regions in the image. To fuse the information from both image and text encoders, a multi-modal encoder with $M$ consecutive cross-attention layers is utilized [19, 25]. These cross-attention layers learn a good region-text alignment [25, 34] and are utilized to obtain regions corresponding to the object of interest $c_t$ from the caption. In particular, for the $m$ -th cross-attention layer, the visual region attention scores $\\mathbf{X}_t^m$ for the object of interest $c_t$ is calculated as:", + "bbox": [ + 496, + 412, + 893, + 623 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {t} ^ {m} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {h} _ {t} ^ {m - 1} \\mathbf {R} ^ {T}}{\\sqrt {d}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 631, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {t} ^ {n} = \\mathbf {X} _ {t} ^ {m} \\cdot \\mathbf {R}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 664, + 890, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ is a scalar and $\\mathbf{h}_t^{m-1}$ is the hidden representation obtained from the previous $(m-1)$ -th cross-attention layer. GradCAM Activation Map. After obtaining attention scores $\\mathbf{X}_t^m$ , we employ Grad-CAM [39] to visualize the activated regions. As in [39], we take the image-caption similarity $(S)$ output from the multi-modal encoder's final layer and calculate the gradient with respect to the attention scores. The activation map $\\phi_t$ for object $c_t$ is:", + "bbox": [ + 496, + 691, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {t} = \\mathbf {X} _ {t} ^ {m} \\cdot \\max \\left(\\frac {\\partial S}{\\partial \\mathbf {X} _ {t} ^ {m}}, 0\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 815, + 890, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Iterative Masking. During VLM training, an object's most discriminative regions easily get aligned towards object text representation [25, 34]. As a result, $\\phi_t$ is localized towards", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23541", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e6e0885f1a8316d49654e267400d7cc5c755ea0329656cd230f818899ba46f98.jpg", + "image_caption": [ + "Figure 3. Illustrative overview of our pseudo-mask generation pipeline. Given an image-caption pair and pre-trained VLM, we generate an activation map for the object of interest (\"umbrella\") and enhance it using iterative masking strategy. We generate box-level annotations using an activation map as a guidance function to select the best WSPN proposals covering the object. We crop the image corresponding to the generated pseduo bounding box and perform weakly-supervised segmentation to obtain pixel-level annotations." + ], + "image_footnote": [], + "bbox": [ + 138, + 89, + 836, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the most discriminative region and fails to cover the object completely [39]. However, when we mask out the most discriminative regions, GradCAM activations are shifted towards other discriminative regions (see Fig. 4). We propose a simple iterative masking strategy to obtain better activation where the most activated part of the object is replaced with image mean and the new activation map is computed following Eq. 1 and 3. The final activation map is:", + "bbox": [ + 75, + 465, + 472, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Phi} _ {t} = \\bigcup_ {i = 1} ^ {G} \\mathcal {I M} \\left(\\phi_ {t} ^ {i}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 587, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G$ is a hyper-parameter indicating the number of masking iterations and $\\mathcal{IM}(\\cdot)$ normalize and threshold $\\phi_t$ by 0.5. We utilize the activation map $\\Phi_t$ as a guidance function to generate box-level and pixel-level annotations.", + "bbox": [ + 75, + 626, + 470, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weakly-Supervised Proposal Network. To generate box-level annotations, we require bounding box proposals covering the activated region $\\Phi_t$ . As explained in the introduction, the fully-supervised RPN network produces poor proposals for novel categories, making it less preferable. To overcome this, we propose a weakly-supervised proposal network (WSPN), which generates object proposals trained in a weakly-supervised manner. Given an Image I, the WSPN is supervised with image-labels $\\mathbf{Y} = \\{y_1, y_2, \\dots, y_C\\}$ , where $y_c = 0$ or 1 indicates the absence or presence of class $c$ in $\\mathbf{I}$ and $C$ denotes total number of classes [5]. First, we utilize Selective Search to generate a set of unsupervised proposal $\\mathbf{U} = \\{u_1, u_2, \\dots, u_N\\}$ where $N$ is the total number of bounding box proposals. Then, the Image I and proposal $\\mathbf{U}$ are fed into a CNN backbone", + "bbox": [ + 75, + 691, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8416f981e3aff51cf2fa537af902c1472958da128a51fa9440c26527c5e37739.jpg", + "image_caption": [ + "Figure 4. Comparison between activation map $\\Phi_t$ generated at $G$ 0 and 1. For $G = 0$ , the most discriminative parts of an object (bird's head) gets activated (bird's head). After masking, for $G = 1$ we can observe that the activation map has shifted to less discriminative part (bird's body). Thus, by combining activation from both steps, we obtain a better activation map trying to cover entire object $\\Phi_t$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 463, + 890, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to extract features and RoI pooling layer [37] to obtain RoI pooled feature vectors. Following [5], the pooled feature vectors are passed to a classification and detection branch to generate two matrices $\\mathbf{W}^{cls}$ , $\\mathbf{W}^{det} \\in \\mathbb{R}^{C \\times N}$ . Then, $\\mathbf{W}^{cls}$ and $\\mathbf{W}^{det}$ matrices are normalized along the category direction (column-wise) and proposal direction (row-wise) by the softmax layers $\\sigma(\\cdot)$ respectively. From $\\mathbf{W}^{cls}$ and $\\mathbf{W}^{det}$ , the instance-level classification scores for object proposals are computed by the element-wise product $\\mathbf{W}_{\\mathbf{C}} = \\sigma(\\mathbf{W}^{cls}) \\odot \\sigma(\\mathbf{W}^{det})$ . and image-level classification score for the $c_{th}$ class is computed as $p_c = \\sum_{i=1}^N w_{i,c}$ . Following [38], we select the high-scoring proposals are used as", + "bbox": [ + 496, + 729, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23542", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pseudo-regression targets ( $\\hat{\\mathbf{T}} = \\{\\hat{t}(u_1), \\hat{t}(u_2), \\dots, \\hat{t}(u_N)\\}$ ) for low-scoring proposals to make sure objects can be more tightly captured. To this end, the classification loss and regression loss for WSPN are calculated as:", + "bbox": [ + 75, + 89, + 468, + 148 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {w s p n} = - \\sum_ {c = 1} ^ {C} y _ {c} \\log p _ {c} + (1 - y _ {c}) \\log \\left(1 - p _ {c}\\right) + \\tag {5} \\\\ \\frac {1}{N} \\sum_ {u = 1} ^ {N} \\mathcal {L} _ {\\text {s m o o t h L 1}} (\\hat {t} (u _ {i}), u _ {i}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 157, + 468, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "WSPN is trained to localize and classify objects by minimizing Eq. 5. The trained WSPN model is used to generate object proposals and top $K$ proposal candidates over all the classes $\\mathcal{B} = \\{\\mathbf{b}_1, \\mathbf{b}_2, \\dots, \\mathbf{b}_K\\}$ are selected by sorting proposal confidence scores obtained from $\\mathbf{W}^{det}$ . Finally, from the top proposal candidates $\\mathcal{B}$ , we select the proposal which overlaps the most with $\\Phi_t$ as pseudo box-bounding:", + "bbox": [ + 75, + 241, + 468, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} ^ {*} = \\underset {\\mathbf {b} \\in \\mathcal {B}} {\\arg \\max } \\frac {\\sum_ {\\mathbf {b}} \\Phi_ {t}}{\\sqrt {| \\mathbf {b} |}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 347, + 468, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{b}^*$ is the pseudo box-level annotation and $\\sum_{\\mathbf{b}}\\Phi_t$ indicates the summation of the activation map values within a box proposal $\\mathbf{b}$ , and $|\\mathbf{b}|$ indicates the proposal area.", + "bbox": [ + 75, + 380, + 468, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Weakly-Supervised Segmentation. Once we obtain the pseudo bounding box $\\mathbf{b}^*$ , we crop image $\\mathbf{I}$ to obtain the corresponding image patch. The cropped patch is then fed into a simple three-layer CNN network to perform pixel-level segmentation. To supervise the CNN network, we generate pseudo ground-truth $\\Theta$ using $\\Phi_t$ and $\\mathbf{b}^*$ , where we sample $Z$ points as foreground $F_z = \\{f_i\\}_{i=1,\\dots,Z}$ and background $B_z = \\{b_i\\}_{i=1,\\dots,Z}$ and each point is set to 1 or 0, respectively. Specifically, the foreground and background points are sampled from the most and least activated part of $\\Phi_t$ inside $\\mathbf{b}^*$ . The pseudo ground-truth $\\Theta$ is of size $\\mathbf{b}^*$ and we supervise the network predictions only at sampled points. Thus, the segmentation loss obtained from these weak points is computed as follows:", + "bbox": [ + 75, + 425, + 468, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w s s} = \\sum_ {i = 1} ^ {G} \\mathcal {L} _ {c e} \\left(\\mathbf {s} ^ {*} \\left(f _ {i}\\right), \\boldsymbol {\\Theta} \\left(f _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {G} \\mathcal {L} _ {c e} \\left(\\mathbf {s} ^ {*} \\left(b _ {i}\\right), \\boldsymbol {\\Theta} \\left(b _ {i}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 636, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{s}^*$ is the pseudo pixel-level annotation of size $\\mathbf{P}$ and $\\mathcal{L}_{ce}$ indicates cross-entropy loss.", + "bbox": [ + 76, + 689, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, given an image $\\mathbf{I}$ and caption $\\mathbf{C}$ pair, we generate pseudo box-level $\\mathbf{b}^*$ and pixel-level $\\mathbf{s}^*$ annotation for the object of interest $c_{t}$ . In practice, the pseduo-mask annotations are generated for a pre-defined set of object categories obtained from training vocabulary. Fig. 5 visualizes the generated pseduo-mask annotations for various categories. Activated regions correspond well with objects of interest and the generated pseduo-mask are of good quality.", + "bbox": [ + 75, + 719, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Open-Vocabulary Instance Segmentation", + "text_level": 1, + "bbox": [ + 76, + 849, + 433, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After generating pseudo-mask annotations, we train an open-vocabulary instance segmentation model. Following", + "bbox": [ + 76, + 871, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[46], we employ a Mask-RCNN as the instance segmentation model, where a class-agnostic mask head is utilized to segment objects and the classification head is replaced with embedding head $h_{emb}$ . Given a Image I, an encoder network extracts image features and region embeddings, $\\mathbf{R} = \\{\\mathbf{r}_i\\}_{i=1,\\dots,N_r}$ , are obtained by RoI align [15] followed by a fully connected layer, where $N_r$ denotes the number of regions. The similarity between the region and text embedding pair is calculated as follows:", + "bbox": [ + 496, + 90, + 890, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {r} _ {i}, \\mathbf {c} _ {j}\\right) = \\frac {\\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {c} _ {j}\\right)}{\\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {b g}\\right) + \\sum_ {k} \\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {c} _ {k}\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 227, + 890, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\mathbf{C} = \\{\\mathbf{bg},\\mathbf{c}_1,\\mathbf{c}_2,\\dots,\\mathbf{c}_{N_C}\\}$ , are object vocabulary text representation obtained from pre-trained text encoder, where $N_{C}$ is the training object vocabulary size. To learn the semantic space, negative pairs are pushed away and positive pairs are pulled together using cross entropy loss obtained from $\\mathbf{b}^*$ . The class-agnostic mask head is supervised by minimizing standard segmentation loss [15] obtained from $\\mathbf{s}^*$ . During inference, the similarity between the region proposals embedding and text embedding from a group of object classes of interest is calculated. The region is then assigned to a class with the highest similarity.", + "bbox": [ + 496, + 268, + 890, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 500, + 450, + 733, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following [46], we conduct experiments on MS-COCO [29] with data split of 48 base categories and 17 novel categories. The processed COCO dataset contains 107,761 training images and 4,836 test images. Following [17], we conduct experiments on Open Images [21] to verify the effectiveness of our method on the large-scale dataset. The Open Images dataset consists of 300 categories with a class split of 200 base categories (frequent objects) and 100 novel categories (rare objects). Following [34], we leverage image-labels obtained from MS-COCO and Open Images to learn the novel category information. We also experiment using image-caption datasets to show our method's effectiveness irrespective of training vocabulary.", + "bbox": [ + 496, + 476, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics and Protocols. Following open-vocabulary methods [17, 52], for both detection and segmentation tasks, we report the mean Average Precision at intersection-over-union (IoU) of $0.5\\left(mAP_{50}\\right)$ . Following zero-shot settings [46], we report novel category performance for both constrained setting and generalized setting. In constrained setting, the model is evaluated only on novel class test images and in generalized setting, the model is evaluated on both base and novel class test images.", + "bbox": [ + 496, + 672, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In pseudo-mask generation framework, we use pre-trained ALBEF [25] as our vision-language model. We conducted all our pseudo-mask generation experiments using ALBEF due to the good region-text alignment when image and caption pair are present [25]. Following ALBEF, the cross-attention layer $m$ used for", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23543", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3141744ff0c2c7c223599490997db2c6c2e9a7c053810e75bd26ef8ae26212d2.jpg", + "table_caption": [ + "Table 1. Object Detection (mAP) performances for MS-COCO under constrained and generalized setting. $\\mathcal{C}_{\\mathrm{B}}$ and $\\mathcal{C}_{\\mathrm{N}}$ are subset of $\\mathcal{C}_{\\Omega}$ , where $\\mathcal{C}_{\\Omega}$ contains training vocabulary larger than COCO categories." + ], + "table_footnote": [], + "table_body": "
MethodProposal GeneratorLanguage SupervisionBase AnnotationConstrained NovelGeneralized Novel
WSDDN [5]-Image-labels in C_B ∪ C_NX-19.7
Cap2Det [45]-Image-labels in C_B ∪ C_NX-20.3
SB [2]RPN COCObase-✓0.700.31
DELO [56]RPN COCObase-✓7.603.41
PL [35]RPN COCObase-✓10.04.12
OV-RCNN [46]RPN COCObaseImage-caption in C_B ∪ C_N✓27.522.8
CLIP-RPN [14]RPN COCObaseCLIP image-text pair C_Ω✓-26.3
ViLD [14]RPN COCObaseCLIP image-text pair C_Ω✓-27.6
Detic [54]RPN COCObaseImage-caption in C_B ∪ C_N✓-27.8
RegionCLIP [52]RPN LVISbaseConceptual caption C_Ω✓30.826.8
PB-OVD [11]RCNN COCObaseImage-caption in C_B ∪ C_N✓32.330.7
XPM [17]RPN COCObaseImage-caption in C_B ∪ C_N✓29.927.0
Mask-free OVIS (Ours)WSPN COCObaseImage-labels in C_B ∪ C_NX31.527.4
Mask-free OVIS (Ours)WSPN COCObaseImage-labels in C_B ∪ C_N✓35.931.5
", + "bbox": [ + 133, + 117, + 828, + 338 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff6ab62685df579a60ce77157fe941c4e84bb546b38e110ad240193b4613aac7.jpg", + "table_caption": [ + "Table 2. Instance Segmentation (mAP) performances for MS-COCO and Open Images under constrained and generalized setting." + ], + "table_footnote": [], + "table_body": "
MethodProposal Generator (MS-COCO/OpenImages)Base AnnotationMS-COCOOpen Images
Constrained NovelGeneralized NovelConstrained NovelGeneralized Novel
OVR+OMP [4]-✓14.18.324.916.8
SB [2]-✓20.816.024.817.3
BA-RPN [51]-✓20.115.425.316.9
Soft-Teacher [43]RPN COCObase/RPN OpenImgbase✓14.89.625.917.6
Unbiased-Teacher [30]RPN COCObase/RPN OpenImgbase✓15.19.822.214.5
OV-RCNN [46]RPN COCObase/RPN OpenImgbase✓20.917.123.817.5
XPM [17]RPN COCObase/RPN OpenImgbase✓24.021.631.622.7
Mask-free OVIS (Ours)WSPN COCObase/WSPN COCObase✗27.425.035.925.8
", + "bbox": [ + 78, + 359, + 893, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Grad-CAM visualization is set to 8. For attention score, we directly employ the original setting of ALBEF and no additional modification is performed. Note that other pretrained vision-language models can also be integrated into our pipeline without major modifications. For the proposal generation pipeline, the WSPN network is trained using COCO base image-labels and the top $K$ proposals candidates is set to 50. The WSPN network is trained for 40k iterations with learn rate 0.001 and weight decay 0.0001. For iterative masking, the hyper-parameter $G$ is set to 3. In the segmentation pipeline, for each patch, the segmentation network is trained for 500 iterations with lr 0.25.", + "bbox": [ + 75, + 523, + 470, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fair-comparison [17, 46], we use Mask R-CNN with a ResNet50 backbone as our open-vocabulary instance segmentation model. During pseudo-mask training, we train Mask-RCNN on MS-COCO and OpenImages using batch size 8 on 8 A5000 GPUs for 90k iterations. Following [11, 14], we use text embeddings obtained from a pretrained CLIP text encoder. During pseduo-mask training, the initial learning rate is set 0.01 and the background class weight is set to 0.2 to improve the recall of novel classes [46]. For base fine-tuning, the initial learning rate is set to 0.0005 and the weight decay is set to 0.0001. We run fine-tuning for 90k iterations where the learning rate is updated", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by a decreasing factor of 0.1 at 60k and 80k iterations.", + "bbox": [ + 498, + 523, + 859, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Detection: As shown in Table 1, we compare our method with previous established open-vocabulary detection methods on the MS-COCO dataset. Compared to weakly-supervised methods such as WSDDN [5] and zero-shot methods such as SB [2], DELO [56], our method outperforms them by a large margin. OV-RCNN [46], XPM [17] are OVD methods based on caption pre-training and our method trained with only pseudo-labels improves the novel category performance by $20.2\\%$ and $2.4\\%$ in generalized setting, respectively. Also, when compared to the method which leverages pre-trained vision-language models such as ViLD [14], RegionCLIP [52], PB-OVD [11], our method with just pseudo-labels produces similar performance. However, with fine-tuning on base annotations, our method outperforms ViLD [14], RegionCLIP [52], PB-OVD [11] by $13.3\\%$ , $14.9\\%$ and $2.8\\%$ in generalized setting, respectively. This is because with fine-tuning, the model learns task/domain specific information from noise-free annotations, boosting the novel category performance. Even without base annotations, our method outperforms most of the existing OVD methods supervised using base annotations. This shows the effectiveness of our method for learning quality representation for novel categories. Specif", + "bbox": [ + 496, + 551, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23544", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0d4b824add802320429ea58b6087469e72b4f7cd85321ef07b335d96b606c4a0.jpg", + "table_caption": [ + "Table 3. Object detection and Instance segmentation ablation analysis for GradCAM, Pseudo-label and Mask-RCNN training." + ], + "table_footnote": [], + "table_body": "
MethodLanguage SupervisionMask-RCNN TrainingBase AnnotationObject DetectionInstance Segmentation
ConstrainedGeneralizedConstrainedGeneralized
Novel/BaseNovel/Base/AllNovel/BaseNovel/Base/All
GradCAMCNXX8.6/0.0-5.2/0.0-
PLCNXX17.3/0.0-14.8/0.0-
PL + Mask-RCNNCN✓X31.1/0.427.1/0.6/7.627.0/0.524.7/0.5/6.9
PL + Mask-RCNNCB ∪ CN✓X32.4/22.429.3/22.8/24.527.4/18.425.0/18.3/20.1
PL + Mask-RCNNCB ∪ CN✓✓35.9/40.740.0/31.5/37.731.2/36.736.0/28.7/34.0
", + "bbox": [ + 114, + 104, + 849, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ically, the quality representation is learned due to the quality proposal generated by WSPN compared to fully-supervised RPN and RCNN proposal generators.", + "bbox": [ + 75, + 220, + 468, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instance Segmentation: Table 2 compares our method with previous open-vocabulary instance segmentation methods on the MS-COCO and Open Images datasets. SB [2] and BA-RPN [51] are zero-shot methods which utilize different background modelling strategies and caption pre-training to learn novel categories. Compared to these, our method improves the novel category performance by a large margin on both datasets and settings. When compared against conventional pseudo-labelling methods, such as soft-teacher [43] and unbiased teacher [30], our method significantly improves on MS-COCO and Open Images datasets. Finally, when compared to open-vocabulary methods such as OV-RCNN [46], XPM [17], our method outperforms by 4.6 and $3.1\\mathrm{mAP}$ in COCO and Open Images in generalized setting, respectively. All these comparisons are performed against our method trained with just pseudo-labels and no base annotation is used during training. This shows the effectiveness of our overall pseudo-mask generation pipeline.", + "bbox": [ + 75, + 268, + 472, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 556, + 230, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GradCAM vs Pseudo-Labels vs Mask-RCNN Training. In Table 3, we analyze the quality of pseudo-labels generated from the GradCAM activation map and our method (PL). After pseudo-label generation, we show how Mask-RCNN training helps to improve the quality of prediction compared to pseudo-labels. Finally, we show how fine-tuning on base annotation improves our method. In the first row, we evaluate the pseudo-mask for novel samples, where the pseudo-mask is generated by normalizing and threshold the GradCAM activation map. In the second row, we evaluate the pseudo-mask for novel samples, where the pseudomask is generated by our method. From Table 3, we can observe the quality of the pseudo-mask generated by our method for novel samples is much better than the GradCAM activation map as a pseudo-mask. From Table 3 third row, we can observe training a Mask-RCNN on pseudo-labels improves the performance on novel categories by modelling fine-grained information. By including pseudo-labels from base categories, we observe that the performance on novel samples further improves. Finally, when fine-tuning on base annotations, the performance on novel categories signifi", + "bbox": [ + 75, + 583, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f1a171af983dfd320ecceeee652b93e944d2b73fc8c9e57e5b08c0b1e8057a8.jpg", + "table_caption": [ + "Table 4. Ablation analysis for object detection and instance segmentation under different language supervision." + ], + "table_footnote": [], + "table_body": "
MethodLanguage SupervisionObject DetectionInstance Segmentatn
Constrd NovelGenrlzd NovelConstrd NovelGenrlzd Novel
OV-RCNN [46]Image-caption COCO27.522.8--
PB-OVD [11]Image-caption COCO-29.1--
PB-OVD [11]Img-Cap COCO, SBU, VG32.330.8--
OursImage-labels COCO35.931.531.028.3
OursImage-caption COCO36.131.831.528.8
", + "bbox": [ + 500, + 252, + 890, + 342 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b4d29443a576e64cdf634251182af680f57c98fff1cf96c15dd456f013c25ed8.jpg", + "table_caption": [ + "Table 5. Ablation analysis for object detection under different proposal generator. All models are fined-tuned on COCO base." + ], + "table_footnote": [], + "table_body": "
MethodProposal GeneratorConstrained NovelGeneralized Novel
OV-RCNN [46]RPN COCO Base27.822.8
PB-OVD [11]RCNN COCO Base32.330.8
OursSelective Search34.531.0
OursWSPN COCO Base35.931.5
", + "bbox": [ + 516, + 373, + 867, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cantly improves by learning task/domain specific information from noise-free manual annotations.", + "bbox": [ + 498, + 462, + 890, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Language supervision. Given an image and caption pair, our method can generate a pseudo-mask leveraging a pretrained vision-language model. Thus to analyze the effect of captions, we conduct experiments between human-provided captions and pseudo-captions generated from image-labels. As show in Table 4, human-provided captions and image-labels based pseudo-caption produce similar performance showing that irrespective of caption type, our method can generate pseudo-mask for the object of interest (see supplementary material for visual comparison). Therefore, our method is more data efficient as it requires cheap image-labels compared to human-provided captions. Table 4 compares our method with other pseudo-label generation methods trained with extra language supervision [11]. Our method, with lesser language supervision, outperforms [11] by a considerable margin.", + "bbox": [ + 496, + 492, + 892, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Proposal Generator Quality vs Performance. In general, better quality of proposal provides better quality of pseudolabels. Therefore, we generated pseduo-labels using different proposal generator and the results are reported in Table 5. As shown in Table 5, our method trained with WSPN as proposal generator is produces better performance compared to methods which rely on fully-supervised proposal generator such as RPN and RCNN. Also when compared to selective search as proposal generator, WSPN demonstrates better performance for novel categories. This is because WSPN refines selective search proposals and localizes them", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23545", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5323d98e0678f660fcdb53d6b94c82a907a0cd30f5fc2d2a3041ba058e053bf.jpg", + "image_caption": [ + "A woman holding an umbrella." + ], + "image_footnote": [], + "bbox": [ + 122, + 117, + 235, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98b404cd3d4b36803dea95a082eb36911ddce17cb4300c931a5f55a2397d3216.jpg", + "image_caption": [ + "A photo of Dog." + ], + "image_footnote": [], + "bbox": [ + 235, + 116, + 459, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01d5ff920c96207505dd4bb367db67ed6c6e089a0747e95477739ee9f1272a35.jpg", + "image_caption": [ + "A) Visualization of activation map and pseudo-bbox $(\\mathbf{b}^{*})$", + "B) Pseudo-mask annotations" + ], + "image_footnote": [], + "bbox": [ + 460, + 117, + 648, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9d24e5a58885e18be4417c6cb919917e1f8c8dd29da202846eee89a23983713.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 117, + 843, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1da99ed894fdfd008face366861bcdac4ad6e7516fa7463994dddbb9c758d9fa.jpg", + "image_caption": [ + "Figure 5. A) Pseduo bounding box selection guided by GradCAM activation. B) Visualization of pseudo-mask generated for Open Images.", + "(a) Object detection", + "Figure 6. Ablation study for object detection and instance segmentation by increase number of masking iteration during PL generation. As iteration increases, GradCAM activation becomes noisy resulting in poor guidance and pseudo-mask generation." + ], + "image_footnote": [], + "bbox": [ + 81, + 270, + 269, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b9f6f8885241010f0532ac53d6586c4c0265a6882a6aa591f1f1aec46ea1802.jpg", + "image_caption": [ + "(b) Instance segmentation" + ], + "image_footnote": [], + "bbox": [ + 274, + 270, + 464, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "towards objects producing better quality proposals.", + "bbox": [ + 75, + 463, + 413, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Iterative Masking Steps vs Performance. Fig. 6 presents ablation analysis for iterative masking hyper-parametr $G$ . We see that for initial few masking steps the performance increase. Because, without any masking the GradCAM activation map is towards most discriminative parts of an object. As we mask out these regions, the less discriminative regions gets activated and combining activation from previous steps produce a strong guidance function $\\Phi_t$ . However, performing too many mask iterations could completely mask out the object of interest and unrelated background regions might get activated resulting in a poor $\\Phi_t$ .", + "bbox": [ + 75, + 478, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weakly-supervised Instance Segmentation. Weakly-supervised Instance Segmentation (WSIS) methods performs instance-segmentation using supervision from image-labels. Similar to WSIS setting, our method performs instance-segmentation using just image-labels. Following this, we simply extended our method to WSIS setting by comparing with VOC benchmark. PENet [12] is a recent WSIS method, which relies on multiple downstream task network such as segmentation, detection and multi-class classification to perfrom segmentation. CL [18] and IAM [57] are WSIS methods utilize advanced attention mechanism to generate instance segmentation mask. As shown in Table 6, our method outperforms existing WSIS methods by considerable margin. Thus, the proposed method is simple and can be easily extended to OV setting or WSIS setting producing SOTA performance. Note that, other OV methods cannot be extended to WSIS setting as they rely on base", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f426c5c6c70448a8cb2cf59e03b9d1d4f950226199f36365fb4a314b4caa580e.jpg", + "table_caption": [ + "Table 6. Instance segmentation(mAP) performance for VOC in weakly-supervised instance segmentation setting." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionmAP
Mask-RCNN [15]F67.9
OCIS [8]I26.8
PRM [55]I28.3
IAM [57]I30.2
CL [18]I+C30.2
PENet [12]I38.1
OursI38.6
", + "bbox": [ + 573, + 304, + 810, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "annotation and similarly WSIS methods cannot be extended to OV setting due to constrained vocabulary space.", + "bbox": [ + 498, + 426, + 890, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis. Fig. 5 (A) presents a visualization of activation maps generated for the object of interest (woman and dog). As we can see, the generated activation map covers the entire object and it can be used as a guidance function to choose the best bounding box proposal. Note that the activation are square-shaped because the original activation map is $1/16$ th of the image size. We perform nearest interpolation to obtain an activation map of image size. In Fig 5 (B), we visualize the pseudo-mask for Open Images generated from our pipeline. We can observe that the generated pseudo-mask is of good quality; still contains some false positives. However, with Mask-RCNN training, the model learns to filter the noise present in the pseudo-mask producing better-quality predictions.", + "bbox": [ + 496, + 460, + 892, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 681, + 617, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel pipeline called Mask-free OVIS for open-vocabulary instance segmentation, which eliminates the need for costly and time-consuming manual mask annotations. We achieve this by generating pseudo-mask annotation for base and novel categories by leveraging a pretrained vision-language model and weakly supervised techniques. The generated pseudo-mask annotations are then used to supervise the Mask-RCNN model. Extensive experiments are conducted to show the effectiveness of our method on MS-COCO and Open Images datasets. Instead of relying on labour-expensive instance-level annotations, we hope our simple and effective pipeline provides an alternative to detect and segment long-tailed novel categories.", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23546", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aditya Arun, CV Jawahar, and M Pawan Kumar. Dissimilarity coefficient based weakly supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9432-9441, 2019. 3", + "[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 384-400, 2018. 6, 7", + "[3] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In European conference on computer vision, pages 549-565. Springer, 2016. 1", + "[4] David Biertimpel, Sindi Shkodrani, Anil S Baslamisli, and Nora Baka. Prior to segment: Foreground cues for weakly annotated classes in partially supervised instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2824-2833, 2021. 6", + "[5] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2846-2854, 2016. 4, 6", + "[6] Maria A Bravo, Sudhanshu Mittal, and Thomas Brox. Localized vision-language matching for open-vocabulary object detection. In Pattern Recognition: 44th DAGM German Conference, DAGM GCR 2022, Konstanz, Germany, September 27–30, 2022, Proceedings, pages 393–408. Springer, 2022. 3", + "[7] Liang-Chieh Chen, Alexander Hermans, George Papandreou, Florian Schroff, Peng Wang, and Hartwig Adam. Masklab: Instance segmentation by refining object detection with semantic and direction features. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4013-4022, 2018. 1", + "[8] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12405, 2019. 8", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3", + "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 3", + "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 266-282. Springer, 2022. 1, 3, 6, 7", + "[12] Weifeng Ge, Sheng Guo, Weilin Huang, and Matthew R Scott. Label-penet: Sequential label propagation and enhancement networks for weakly supervised instance segmen-" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3345-3354, 2019. 8", + "[13] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. CyCLIP: Cyclic contrastive language-image pretraining. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3", + "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 1, 2, 3, 6", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 5, 8", + "[16] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Advances in neural information processing systems, 33:16797-16807, 2020. 3", + "[17] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2, 3, 5, 6, 7", + "[18] Jaedong Hwang, Seohyun Kim, Jeany Son, and Bohyung Han. Weakly supervised instance segmentation by deep community learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1020-1029, 2021. 8", + "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2, 3", + "[20] Wonjik Kim, Asako Kanezaki, and Masayuki Tanaka. Unsupervised learning of image segmentation based on differentiable feature clustering. IEEE Transactions on Image Processing, 29:8055-8068, 2020. 2", + "[21] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 1, 5", + "[22] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 3", + "[23] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 3", + "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for uni" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23547", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fied vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3", + "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3, 5", + "[26] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell me where to look: Guided attention inference network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 3", + "[27] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3", + "[28] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5", + "[30] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In International Conference on Learning Representations, 2021. 6, 7", + "[31] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3", + "[32] Zongyang Ma, Guan Luo, Jin Gao, Liang Li, Yuxin Chen, Shaoru Wang, Congxuan Zhang, and Weiming Hu. Open-vocabulary one-stage detection with hierarchical visual-language knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14074-14083, 2022. 3", + "[33] Niall O'Mahony, Sean Campbell, Anderson Carvalho, Suman Harapanahalli, Gustavo Velasco Hernandez, Lenka Krpalkova, Daniel Riordan, and Joseph Walsh. Deep learning vs. traditional computer vision. In Science and information conference, pages 128-144. Springer, 2019. 1", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5", + "[35] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11932–11939, 2020. 6" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In 36th Conference on Neural Information Processing Systems (NIPS), 2022. 3", + "[37] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2, 4", + "[38] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, MingYu Liu, Yong Jae Lee, Alexander G Schwing, and Jan Kautz. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598-10607, 2020. 4", + "[39] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 3, 4", + "[40] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 3", + "[41] Limin Wang, Yuanjun Xiong, Dahua Lin, and Luc Van Gool. Untrimmednets for weakly supervised action recognition and detection. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4325-4334, 2017. 3", + "[42] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In European Conference on Computer Vision, pages 649-665. Springer, 2020. 1", + "[43] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 6, 7", + "[44] Ke Yang, Dongsheng Li, and Yong Dou. Towards precise end-to-end weakly supervised object detection network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8372-8381, 2019. 3", + "[45] Keren Ye, Mingda Zhang, Adriana Kovashka, Wei Li, Danfeng Qin, and Jesse Berent. Cap2det: Learning to amplify weak caption supervision for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9686-9695, 2019. 6", + "[46] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 1, 2, 3, 5, 6, 7", + "[47] Zhaoyang Zeng, Bei Liu, Jianlong Fu, Hongyang Chao, and Lei Zhang. Wsod2: Learning bottom-up and top-down objectness distillation for weakly-supervised object detection." + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23548", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF international conference on computer vision, pages 8292-8300, 2019. 3", + "[48] Dingwen Zhang, Junwei Han, Gong Cheng, and Ming-Hsuan Yang. Weakly supervised object localization and detection: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5866-5885, 2021. 3", + "[49] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3", + "[50] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 159-175. Springer, 2022. 2, 3", + "[51] Ye Zheng, Jiahong Wu, Yongqiang Qin, Faen Zhang, and Li Cui. Zero-shot instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2593-2602, 2021. 6, 7", + "[52] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 1, 2, 3, 5, 6", + "[53] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3", + "[54] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 6", + "[55] Yanzhao Zhou, Yi Zhu, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Weakly supervised instance segmentation using class peak response. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3791-3800, 2018. 8", + "[56] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't even look once: Synthesizing features for zero-shot detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11693-11702, 2020. 6", + "[57] Yi Zhu, Yanzhao Zhou, Huijuan Xu, Qixiang Ye, David Doermann, and Jianbin Jiao. Learning instance activation maps for weakly supervised instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3116-3125, 2019. 8" + ], + "bbox": [ + 78, + 90, + 468, + 838 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23549", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_model.json b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f906fe292502c9e46a4fdfe84245174c0534c2d --- /dev/null +++ b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_model.json @@ -0,0 +1,2092 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.782, + 0.175 + ], + "angle": 0, + "content": "Mask-free OVIS: Open-Vocabulary Instance Segmentation without Manual Mask Annotations" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.203, + 0.737, + 0.239 + ], + "angle": 0, + "content": "Vibashan VS\\*, Ning Yu†, Chen Xing†, Can Qin‡, Mingfei Gao†, Juan Carlos Niebles†, Vishal M. Patel\\*, Ran Xu†" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.24, + 0.78, + 0.258 + ], + "angle": 0, + "content": "*Johns Hopkins University, ‡Northeastern University, †Salesforce Research" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.26, + 0.725, + 0.275 + ], + "angle": 0, + "content": "{vvishnu2, vpatel36}@jhu.edu, qin.ca@northeastern.edu," + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.278, + 0.705, + 0.293 + ], + "angle": 0, + "content": "{ning.yu, cxing, jniebles, ran.xu}@salesforce.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.474, + 0.801 + ], + "angle": 0, + "content": "Existing instance segmentation models learn task-specific information using manual mask annotations from base (training) categories. These mask annotations require tremendous human effort, limiting the scalability to annotate novel (new) categories. To alleviate this problem, Open-Vocabulary (OV) methods leverage large-scale image-caption pairs and vision-language models to learn novel categories. In summary, an OV method learns task-specific information using strong supervision from base annotations and novel category information using weak supervision from image-captions pairs. This difference between strong and weak supervision leads to overfitting on base categories, resulting in poor generalization towards novel categories. In this work, we overcome this issue by learning both base and novel categories from pseudomask annotations generated by the vision-language model in a weakly supervised manner using our proposed Mask-free OVIS pipeline. Our method automatically generates pseudo-mask annotations by leveraging the localization ability of a pre-trained vision-language model for objects present in image-caption pairs. The generated pseudomask annotations are then used to supervise an instance segmentation model, freeing the entire pipeline from any labour-expensive instance-level annotations and overfitting. Our extensive experiments show that our method trained with just pseudo-masks significantly improves the mAP scores on the MS-COCO dataset and OpenImages dataset compared to the recent state-of-the-art methods trained with manual masks. Codes and models are provided in https://vibashan.github.io/ovis-web/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.81, + 0.21, + 0.825 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.836, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Instance segmentation is a challenging task as it requires models to detect objects in an image while also precisely" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.327, + 0.891, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.429, + 0.891, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.523, + 0.894, + 0.636 + ], + "angle": 0, + "content": "Figure 1. A) Previous Methods: Learn task-specific information (detection/segmentation) in a fully-supervised manner and novel category information with weak supervision. During training, this difference in strong and weak supervision signals leads to overfitting and requires expensive base annotations. B) Our method: Given image-caption pairs, we generate pseudo-annotations for both base and novel categories under weak supervision, solving the problems of labour-expensive annotation and overfitting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.895, + 0.868 + ], + "angle": 0, + "content": "segment each object at the pixel-level. Although the rise of deep neural networks has significantly boosted the state-of-the-art instance segmentation performance [7, 15, 42], these methods are still trained for a pre-defined set of object categories and are data-hungry [33]. Particularly, one needs to manually annotate thousands of instance-level masks for each object category, which takes around 78 seconds per instance mask [3]. If we look at this quantitatively on Open Images [21], a large-scale dataset with 2.1M instance-level mask annotations requires around 5 years of human labour. Even after extensive annotation, these training datasets are still limited to a small number of categories and segmenting objects from a novel category requires further annotation. Therefore, it is difficult to scale up existing methods to segment a large number of categories due to intensive labour." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Recently, Open-Vocabulary (OV) methods have gained much attention due to their success in detecting [11, 14," + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*This work was done when Vibashan VS interned at Salesforce Research. Primary contact: vvishnu2@jhu.com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23539" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.581, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.097, + 0.888, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.309, + 0.893, + 0.352 + ], + "angle": 0, + "content": "Figure 2. RPN is supervised using bounding box annotations from COCO base and WSPN is supervised using image-labels from COCO base. A) WSPN produces better quality proposals for novel object categories compared to fully-supervised RPN. B) WSPN consistently produces better recall for all COCO novel categories than RPN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.583 + ], + "angle": 0, + "content": "46, 52] and segmenting [17] novel categories beyond base (training) categories. An OV method learns task-specific information (detection/segmentation) from base categories that have manual instance-level bounding boxes or masks and learns novel category information from the pre-trained Vision-Language Model (VLM) [19, 34] (see Fig. 1). All these methods produce promising results on novel categories by leveraging different forms of weak supervision such as caption pretaining [46, 50], knowledge distillation [14, 52] and pseudo-labelling [17, 28]. However, all above-mentioned OV methods still rely on the manually-annotated base categories to improve their performances on novel categories. Without fine-tuning on base categories, existing OV methods lack task/domain specific knowledge and the performances on novel categories will be affected [14, 46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Although manual instance-level annotations of base categories are critical to open-vocabulary segmentation methods, we find empirically that such fully-supervised information causes OV methods to overfit to base categories, leading to a higher failure rate when evaluated on novel categories. Specifically, OV methods utilize a region proposal network (RPN) [37] supervised with bounding box annotations obtained from the base categories to generate a set of bounding box proposals for all objects categories in a given image [46]. The feature representation from these object proposals is later matched with text embedding to learn the visual-semantic space for base and novel categories [46]. Therefore, the quality of proposals generated for novel object categories plays a key role in determining the performance in later stages. However, from our experiments, we find that many objects of novel categories wouldn't be included in such proposals due to the RPN's overfitting to base categories. Fig. 2 (A) - Top gives some examples where the RPN trained with COCO base categories fails to generate high-quality region proposals for novel categories such as elephant, cat and skateboard. Therefore, a fully" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.356, + 0.892, + 0.386 + ], + "angle": 0, + "content": "supervised proposal network is a bottleneck in OV pipeline due to its poor generalization towards novel categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Given the aforementioned observations of poor generalization, we raise the question of whether we can improve the generalization by using weak supervision instead of relying on strong supervision from manual base annotations. If so, we can reduce overfitting to the base categories and the requirement for costly instance-level human annotations can be entirely removed from the pipeline. Our preliminary experiments give us some hope. Our experiments show that if we train a weakly-supervised proposal network (WSPN) with image-level annotations instead of box-level annotations, the region proposals it generates can better generalize to novel objects. As shown in Fig. 2 A), the novel objects that the RPN proposals miss are covered by WSPN. Fig. 2 B) shows WSPN proposals have consistently better average recall than RPN for all COCO novel categories, indicating that WSPN proposals are more likely to cover the ground-truth bounding boxes of novel objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Inspired by these observations, we propose open-vocabulary segmentation without manual mask annotations. We do not use any human-provided box-level or pixel-level annotations during the training of our method. We first train a simple WSPN model with image-level annotations on base categories as a proposal generator to generate proposals for all objects given an image. Then, we adopt pre-trained vision-language models to select proposals as pseudo bounding boxes for novel objects. Given a novel object's text name, we can utilize the name as a text prompt to localize this object in an image with a pre-trained vision-language model. To obtain a more accurate pseudo-mask that covers the entire object, we conduct iterative masking with GradCAM [39] given the vision-language model. Finally, we train a weakly-supervised segmentation (WSS) [20] network with previously generated bounding box and GradCAM activation map to obtain pixel-level annotation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: (1) We propose a Mask-free OVIS pipeline where we produce manual-effort-free pseudo-mask annotations for base and novel instance segmentation using open-vocabulary and weakly supervised techniques. (2) We propose a novel pseudo-mask generation pipeline leveraging a pre-trained vision-language model to generate instance-level annotations. (3) Benefiting from pseudo-labels, our method sets up SOTA's for both detection and instance segmentation tasks compared to recent methods trained with manual masks on MS-COCO and OpenImages datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.272, + 0.22, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.297, + 0.47, + 0.447 + ], + "angle": 0, + "content": "Weakly-supervised learning (WSL) methods require less manual annotation effort and only rely on partial information about the task compared to fully-supervised methods. WSL applications include object detection [1, 16, 44, 47], segmentation [22,23,40], action recognition, etc [26,41,48]. The class activation map (CAM) [53] and grad-CAM [39] are popular approaches which exploits pre-trained network to generate saliency or activation map for target classes. Li et al. [26] enables weakly-supervised semantic segmentation via CAM-guided attention." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.449, + 0.473, + 0.689 + ], + "angle": 0, + "content": "Vision-Language Models. With the rise of large-scale text-pretraining with attention-based models [9, 49], vision-language models (VLM) have caught increasing attention due to their strong performance in downstream visual understanding tasks [27, 31, 34]. Early VLM models [27, 31] applied the pre-trained models with constraints of the class set. With the help of contrastive learning, CLIP [34] has enabled VLM in the wild with large-scale multimodal learning upon the 400M noisy data crawled from the web. Several methods have extended CLIP for high efficiency model training and cycle consistency [13, 24, 25]. ALBEF [25] utilizes cross-modal attention between image patches and text, leading to more grounded vision and language representation learning. Compared to CLIP, ALBEF excels in object localization, while CLIP performs better at zero-shot classification due to its larger training dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Open-Vocabulary. Open-Vocabulary methods [6, 10, 32, 36, 46, 50] scale up their vocabulary size for object detection or instance segmentation tasks by transferring knowledge from pre-trained vision-language models. Zareian et al. [46] proposed an open-vocabulary object detection method where a visual encoder is trained on image-caption pair to model object semantics and then transfer to zero-shot object detection. ViLD [14] and RegionCLIP [52] use pre-trained CLIP [34] to distill knowledge and enhance the vision-text embedding space effectively. Similarly, Huynh et al. [17] proposes a robust pseudo-labelling, for instance segmentation where a teacher model generates pseudo-label for novel categories and is later used to supervise a student model. Therefore, all these methods require manually" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "annotated base categories to improve their performance on novel categories. In contrast, our method generates pseudolabels for both base and novel categories by leveraging a pre-trained VLM, freeing the entire pipeline from human-provided instance-level annotations. Similar to our method, Gao et al. [11] generates pseudo-labels using pre-trained VLM, but their approach still uses a proposal generator that relies on human-provided bounding boxes for the base categories. Our method, on the other hand, requires no such annotations, making it completely free from any human-provided box-level or pixel-level annotations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.265, + 0.664, + 0.28 + ], + "angle": 0, + "content": "3. Mask-free OVIS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.895, + 0.389 + ], + "angle": 0, + "content": "Our pipeline consists of two stages: (I) Pseudo-mask generation, and (II) Open-vocabulary instance segmentation. The goal of (I) is to generate box-level and pixel-level annotations by leveraging the region-text alignment property of a pre-trained vision-language model as illustrated in Fig. 3. In (II), we train a Mask-RCNN using the generated pseudo annotations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.392, + 0.733, + 0.406 + ], + "angle": 0, + "content": "3.1. Pseudo-Mask Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.624 + ], + "angle": 0, + "content": "Region-Text Attention Scores. The inputs to the vision-language model are image \\(\\mathbf{I}\\) and caption \\(\\mathbf{C} = \\{c_1, c_2, \\dots, c_{N_c}\\}\\) pair, where \\(N_c\\) is the number of words in the caption (including [CLS] and [SEP]) [9, 25]. In a VLM, a text encoder is utilized to get text representations \\(\\mathbf{T} \\in \\mathbb{R}^{N_C \\times d}\\) and an image encoder is utilized to extract region representation \\(\\mathbf{R} \\in \\mathbb{R}^{N_R \\times d}\\), where \\(N_R\\) is the number of regions in the image. To fuse the information from both image and text encoders, a multi-modal encoder with \\(M\\) consecutive cross-attention layers is utilized [19, 25]. These cross-attention layers learn a good region-text alignment [25, 34] and are utilized to obtain regions corresponding to the object of interest \\(c_t\\) from the caption. In particular, for the \\(m\\)-th cross-attention layer, the visual region attention scores \\(\\mathbf{X}_t^m\\) for the object of interest \\(c_t\\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.632, + 0.891, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {t} ^ {m} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {h} _ {t} ^ {m - 1} \\mathbf {R} ^ {T}}{\\sqrt {d}}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.665, + 0.891, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {t} ^ {n} = \\mathbf {X} _ {t} ^ {m} \\cdot \\mathbf {R}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.808 + ], + "angle": 0, + "content": "where \\(d\\) is a scalar and \\(\\mathbf{h}_t^{m-1}\\) is the hidden representation obtained from the previous \\((m-1)\\)-th cross-attention layer. GradCAM Activation Map. After obtaining attention scores \\(\\mathbf{X}_t^m\\), we employ Grad-CAM [39] to visualize the activated regions. As in [39], we take the image-caption similarity \\((S)\\) output from the multi-modal encoder's final layer and calculate the gradient with respect to the attention scores. The activation map \\(\\phi_t\\) for object \\(c_t\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.816, + 0.892, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\phi_ {t} = \\mathbf {X} _ {t} ^ {m} \\cdot \\max \\left(\\frac {\\partial S}{\\partial \\mathbf {X} _ {t} ^ {m}}, 0\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Iterative Masking. During VLM training, an object's most discriminative regions easily get aligned towards object text representation [25, 34]. As a result, \\(\\phi_t\\) is localized towards" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.837, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.403, + 0.896, + 0.46 + ], + "angle": 0, + "content": "Figure 3. Illustrative overview of our pseudo-mask generation pipeline. Given an image-caption pair and pre-trained VLM, we generate an activation map for the object of interest (\"umbrella\") and enhance it using iterative masking strategy. We generate box-level annotations using an activation map as a guidance function to select the best WSPN proposals covering the object. We crop the image corresponding to the generated pseduo bounding box and perform weakly-supervised segmentation to obtain pixel-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.473, + 0.587 + ], + "angle": 0, + "content": "the most discriminative region and fails to cover the object completely [39]. However, when we mask out the most discriminative regions, GradCAM activations are shifted towards other discriminative regions (see Fig. 4). We propose a simple iterative masking strategy to obtain better activation where the most activated part of the object is replaced with image mean and the new activation map is computed following Eq. 1 and 3. The final activation map is:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.588, + 0.47, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Phi} _ {t} = \\bigcup_ {i = 1} ^ {G} \\mathcal {I M} \\left(\\phi_ {t} ^ {i}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.472, + 0.688 + ], + "angle": 0, + "content": "where \\(G\\) is a hyper-parameter indicating the number of masking iterations and \\(\\mathcal{IM}(\\cdot)\\) normalize and threshold \\(\\phi_t\\) by 0.5. We utilize the activation map \\(\\Phi_t\\) as a guidance function to generate box-level and pixel-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Weakly-Supervised Proposal Network. To generate box-level annotations, we require bounding box proposals covering the activated region \\(\\Phi_t\\). As explained in the introduction, the fully-supervised RPN network produces poor proposals for novel categories, making it less preferable. To overcome this, we propose a weakly-supervised proposal network (WSPN), which generates object proposals trained in a weakly-supervised manner. Given an Image I, the WSPN is supervised with image-labels \\(\\mathbf{Y} = \\{y_1, y_2, \\dots, y_C\\}\\), where \\(y_c = 0\\) or 1 indicates the absence or presence of class \\(c\\) in \\(\\mathbf{I}\\) and \\(C\\) denotes total number of classes [5]. First, we utilize Selective Search to generate a set of unsupervised proposal \\(\\mathbf{U} = \\{u_1, u_2, \\dots, u_N\\}\\) where \\(N\\) is the total number of bounding box proposals. Then, the Image I and proposal \\(\\mathbf{U}\\) are fed into a CNN backbone" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.464, + 0.891, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.713 + ], + "angle": 0, + "content": "Figure 4. Comparison between activation map \\(\\Phi_t\\) generated at \\(G\\) 0 and 1. For \\(G = 0\\), the most discriminative parts of an object (bird's head) gets activated (bird's head). After masking, for \\(G = 1\\) we can observe that the activation map has shifted to less discriminative part (bird's body). Thus, by combining activation from both steps, we obtain a better activation map trying to cover entire object \\(\\Phi_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.895, + 0.903 + ], + "angle": 0, + "content": "to extract features and RoI pooling layer [37] to obtain RoI pooled feature vectors. Following [5], the pooled feature vectors are passed to a classification and detection branch to generate two matrices \\(\\mathbf{W}^{cls}\\), \\(\\mathbf{W}^{det} \\in \\mathbb{R}^{C \\times N}\\). Then, \\(\\mathbf{W}^{cls}\\) and \\(\\mathbf{W}^{det}\\) matrices are normalized along the category direction (column-wise) and proposal direction (row-wise) by the softmax layers \\(\\sigma(\\cdot)\\) respectively. From \\(\\mathbf{W}^{cls}\\) and \\(\\mathbf{W}^{det}\\), the instance-level classification scores for object proposals are computed by the element-wise product \\(\\mathbf{W}_{\\mathbf{C}} = \\sigma(\\mathbf{W}^{cls}) \\odot \\sigma(\\mathbf{W}^{det})\\). and image-level classification score for the \\(c_{th}\\) class is computed as \\(p_c = \\sum_{i=1}^N w_{i,c}\\). Following [38], we select the high-scoring proposals are used as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23542" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.47, + 0.149 + ], + "angle": 0, + "content": "pseudo-regression targets (\\(\\hat{\\mathbf{T}} = \\{\\hat{t}(u_1), \\hat{t}(u_2), \\dots, \\hat{t}(u_N)\\}\\)) for low-scoring proposals to make sure objects can be more tightly captured. To this end, the classification loss and regression loss for WSPN are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.158, + 0.47, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {w s p n} = - \\sum_ {c = 1} ^ {C} y _ {c} \\log p _ {c} + (1 - y _ {c}) \\log \\left(1 - p _ {c}\\right) + \\tag {5} \\\\ \\frac {1}{N} \\sum_ {u = 1} ^ {N} \\mathcal {L} _ {\\text {s m o o t h L 1}} (\\hat {t} (u _ {i}), u _ {i}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.242, + 0.47, + 0.348 + ], + "angle": 0, + "content": "WSPN is trained to localize and classify objects by minimizing Eq. 5. The trained WSPN model is used to generate object proposals and top \\(K\\) proposal candidates over all the classes \\(\\mathcal{B} = \\{\\mathbf{b}_1, \\mathbf{b}_2, \\dots, \\mathbf{b}_K\\}\\) are selected by sorting proposal confidence scores obtained from \\(\\mathbf{W}^{det}\\). Finally, from the top proposal candidates \\(\\mathcal{B}\\), we select the proposal which overlaps the most with \\(\\Phi_t\\) as pseudo box-bounding:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.348, + 0.469, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} ^ {*} = \\underset {\\mathbf {b} \\in \\mathcal {B}} {\\arg \\max } \\frac {\\sum_ {\\mathbf {b}} \\Phi_ {t}}{\\sqrt {| \\mathbf {b} |}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.425 + ], + "angle": 0, + "content": "where \\(\\mathbf{b}^*\\) is the pseudo box-level annotation and \\(\\sum_{\\mathbf{b}}\\Phi_t\\) indicates the summation of the activation map values within a box proposal \\(\\mathbf{b}\\), and \\(|\\mathbf{b}|\\) indicates the proposal area." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.469, + 0.636 + ], + "angle": 0, + "content": "Weakly-Supervised Segmentation. Once we obtain the pseudo bounding box \\(\\mathbf{b}^*\\), we crop image \\(\\mathbf{I}\\) to obtain the corresponding image patch. The cropped patch is then fed into a simple three-layer CNN network to perform pixel-level segmentation. To supervise the CNN network, we generate pseudo ground-truth \\(\\Theta\\) using \\(\\Phi_t\\) and \\(\\mathbf{b}^*\\), where we sample \\(Z\\) points as foreground \\(F_z = \\{f_i\\}_{i=1,\\dots,Z}\\) and background \\(B_z = \\{b_i\\}_{i=1,\\dots,Z}\\) and each point is set to 1 or 0, respectively. Specifically, the foreground and background points are sampled from the most and least activated part of \\(\\Phi_t\\) inside \\(\\mathbf{b}^*\\). The pseudo ground-truth \\(\\Theta\\) is of size \\(\\mathbf{b}^*\\) and we supervise the network predictions only at sampled points. Thus, the segmentation loss obtained from these weak points is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w s s} = \\sum_ {i = 1} ^ {G} \\mathcal {L} _ {c e} \\left(\\mathbf {s} ^ {*} \\left(f _ {i}\\right), \\boldsymbol {\\Theta} \\left(f _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {G} \\mathcal {L} _ {c e} \\left(\\mathbf {s} ^ {*} \\left(b _ {i}\\right), \\boldsymbol {\\Theta} \\left(b _ {i}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.69, + 0.469, + 0.72 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}^*\\) is the pseudo pixel-level annotation of size \\(\\mathbf{P}\\) and \\(\\mathcal{L}_{ce}\\) indicates cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.841 + ], + "angle": 0, + "content": "To this end, given an image \\(\\mathbf{I}\\) and caption \\(\\mathbf{C}\\) pair, we generate pseudo box-level \\(\\mathbf{b}^*\\) and pixel-level \\(\\mathbf{s}^*\\) annotation for the object of interest \\(c_{t}\\). In practice, the pseduo-mask annotations are generated for a pre-defined set of object categories obtained from training vocabulary. Fig. 5 visualizes the generated pseduo-mask annotations for various categories. Activated regions correspond well with objects of interest and the generated pseduo-mask are of good quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.85, + 0.434, + 0.867 + ], + "angle": 0, + "content": "3.2. Open-Vocabulary Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.872, + 0.469, + 0.902 + ], + "angle": 0, + "content": "After generating pseudo-mask annotations, we train an open-vocabulary instance segmentation model. Following" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[46], we employ a Mask-RCNN as the instance segmentation model, where a class-agnostic mask head is utilized to segment objects and the classification head is replaced with embedding head \\( h_{emb} \\). Given a Image I, an encoder network extracts image features and region embeddings, \\( \\mathbf{R} = \\{\\mathbf{r}_i\\}_{i=1,\\dots,N_r} \\), are obtained by RoI align [15] followed by a fully connected layer, where \\( N_r \\) denotes the number of regions. The similarity between the region and text embedding pair is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.228, + 0.892, + 0.269 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {r} _ {i}, \\mathbf {c} _ {j}\\right) = \\frac {\\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {c} _ {j}\\right)}{\\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {b g}\\right) + \\sum_ {k} \\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {c} _ {k}\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.892, + 0.436 + ], + "angle": 0, + "content": "where, \\(\\mathbf{C} = \\{\\mathbf{bg},\\mathbf{c}_1,\\mathbf{c}_2,\\dots,\\mathbf{c}_{N_C}\\}\\), are object vocabulary text representation obtained from pre-trained text encoder, where \\(N_{C}\\) is the training object vocabulary size. To learn the semantic space, negative pairs are pushed away and positive pairs are pulled together using cross entropy loss obtained from \\(\\mathbf{b}^*\\). The class-agnostic mask head is supervised by minimizing standard segmentation loss [15] obtained from \\(\\mathbf{s}^*\\). During inference, the similarity between the region proposals embedding and text embedding from a group of object classes of interest is calculated. The region is then assigned to a class with the highest similarity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.451, + 0.734, + 0.468 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Datasets. Following [46], we conduct experiments on MS-COCO [29] with data split of 48 base categories and 17 novel categories. The processed COCO dataset contains 107,761 training images and 4,836 test images. Following [17], we conduct experiments on Open Images [21] to verify the effectiveness of our method on the large-scale dataset. The Open Images dataset consists of 300 categories with a class split of 200 base categories (frequent objects) and 100 novel categories (rare objects). Following [34], we leverage image-labels obtained from MS-COCO and Open Images to learn the novel category information. We also experiment using image-caption datasets to show our method's effectiveness irrespective of training vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Evaluation Metrics and Protocols. Following open-vocabulary methods [17, 52], for both detection and segmentation tasks, we report the mean Average Precision at intersection-over-union (IoU) of \\(0.5\\left(mAP_{50}\\right)\\). Following zero-shot settings [46], we report novel category performance for both constrained setting and generalized setting. In constrained setting, the model is evaluated only on novel class test images and in generalized setting, the model is evaluated on both base and novel class test images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. In pseudo-mask generation framework, we use pre-trained ALBEF [25] as our vision-language model. We conducted all our pseudo-mask generation experiments using ALBEF due to the good region-text alignment when image and caption pair are present [25]. Following ALBEF, the cross-attention layer \\( m \\) used for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23543" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Object Detection (mAP) performances for MS-COCO under constrained and generalized setting. \\( \\mathcal{C}_{\\mathrm{B}} \\) and \\( \\mathcal{C}_{\\mathrm{N}} \\) are subset of \\( \\mathcal{C}_{\\Omega} \\), where \\( \\mathcal{C}_{\\Omega} \\) contains training vocabulary larger than COCO categories." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.118, + 0.83, + 0.339 + ], + "angle": 0, + "content": "
MethodProposal GeneratorLanguage SupervisionBase AnnotationConstrained NovelGeneralized Novel
WSDDN [5]-Image-labels in C_B ∪ C_NX-19.7
Cap2Det [45]-Image-labels in C_B ∪ C_NX-20.3
SB [2]RPN COCObase-✓0.700.31
DELO [56]RPN COCObase-✓7.603.41
PL [35]RPN COCObase-✓10.04.12
OV-RCNN [46]RPN COCObaseImage-caption in C_B ∪ C_N✓27.522.8
CLIP-RPN [14]RPN COCObaseCLIP image-text pair C_Ω✓-26.3
ViLD [14]RPN COCObaseCLIP image-text pair C_Ω✓-27.6
Detic [54]RPN COCObaseImage-caption in C_B ∪ C_N✓-27.8
RegionCLIP [52]RPN LVISbaseConceptual caption C_Ω✓30.826.8
PB-OVD [11]RCNN COCObaseImage-caption in C_B ∪ C_N✓32.330.7
XPM [17]RPN COCObaseImage-caption in C_B ∪ C_N✓29.927.0
Mask-free OVIS (Ours)WSPN COCObaseImage-labels in C_B ∪ C_NX31.527.4
Mask-free OVIS (Ours)WSPN COCObaseImage-labels in C_B ∪ C_N✓35.931.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.346, + 0.865, + 0.36 + ], + "angle": 0, + "content": "Table 2. Instance Segmentation (mAP) performances for MS-COCO and Open Images under constrained and generalized setting." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.36, + 0.894, + 0.516 + ], + "angle": 0, + "content": "
MethodProposal Generator (MS-COCO/OpenImages)Base AnnotationMS-COCOOpen Images
Constrained NovelGeneralized NovelConstrained NovelGeneralized Novel
OVR+OMP [4]-✓14.18.324.916.8
SB [2]-✓20.816.024.817.3
BA-RPN [51]-✓20.115.425.316.9
Soft-Teacher [43]RPN COCObase/RPN OpenImgbase✓14.89.625.917.6
Unbiased-Teacher [30]RPN COCObase/RPN OpenImgbase✓15.19.822.214.5
OV-RCNN [46]RPN COCObase/RPN OpenImgbase✓20.917.123.817.5
XPM [17]RPN COCObase/RPN OpenImgbase✓24.021.631.622.7
Mask-free OVIS (Ours)WSPN COCObase/WSPN COCObase✗27.425.035.925.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.471, + 0.707 + ], + "angle": 0, + "content": "Grad-CAM visualization is set to 8. For attention score, we directly employ the original setting of ALBEF and no additional modification is performed. Note that other pretrained vision-language models can also be integrated into our pipeline without major modifications. For the proposal generation pipeline, the WSPN network is trained using COCO base image-labels and the top \\( K \\) proposals candidates is set to 50. The WSPN network is trained for 40k iterations with learn rate 0.001 and weight decay 0.0001. For iterative masking, the hyper-parameter \\( G \\) is set to 3. In the segmentation pipeline, for each patch, the segmentation network is trained for 500 iterations with lr 0.25." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.472, + 0.902 + ], + "angle": 0, + "content": "For fair-comparison [17, 46], we use Mask R-CNN with a ResNet50 backbone as our open-vocabulary instance segmentation model. During pseudo-mask training, we train Mask-RCNN on MS-COCO and OpenImages using batch size 8 on 8 A5000 GPUs for 90k iterations. Following [11, 14], we use text embeddings obtained from a pretrained CLIP text encoder. During pseduo-mask training, the initial learning rate is set 0.01 and the background class weight is set to 0.2 to improve the recall of novel classes [46]. For base fine-tuning, the initial learning rate is set to 0.0005 and the weight decay is set to 0.0001. We run fine-tuning for 90k iterations where the learning rate is updated" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.525, + 0.86, + 0.541 + ], + "angle": 0, + "content": "by a decreasing factor of 0.1 at 60k and 80k iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Object Detection: As shown in Table 1, we compare our method with previous established open-vocabulary detection methods on the MS-COCO dataset. Compared to weakly-supervised methods such as WSDDN [5] and zero-shot methods such as SB [2], DELO [56], our method outperforms them by a large margin. OV-RCNN [46], XPM [17] are OVD methods based on caption pre-training and our method trained with only pseudo-labels improves the novel category performance by \\(20.2\\%\\) and \\(2.4\\%\\) in generalized setting, respectively. Also, when compared to the method which leverages pre-trained vision-language models such as ViLD [14], RegionCLIP [52], PB-OVD [11], our method with just pseudo-labels produces similar performance. However, with fine-tuning on base annotations, our method outperforms ViLD [14], RegionCLIP [52], PB-OVD [11] by \\(13.3\\%\\), \\(14.9\\%\\) and \\(2.8\\%\\) in generalized setting, respectively. This is because with fine-tuning, the model learns task/domain specific information from noise-free annotations, boosting the novel category performance. Even without base annotations, our method outperforms most of the existing OVD methods supervised using base annotations. This shows the effectiveness of our method for learning quality representation for novel categories. Specif" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23544" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.09, + 0.855, + 0.104 + ], + "angle": 0, + "content": "Table 3. Object detection and Instance segmentation ablation analysis for GradCAM, Pseudo-label and Mask-RCNN training." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.105, + 0.85, + 0.215 + ], + "angle": 0, + "content": "
MethodLanguage SupervisionMask-RCNN TrainingBase AnnotationObject DetectionInstance Segmentation
ConstrainedGeneralizedConstrainedGeneralized
Novel/BaseNovel/Base/AllNovel/BaseNovel/Base/All
GradCAMCNXX8.6/0.0-5.2/0.0-
PLCNXX17.3/0.0-14.8/0.0-
PL + Mask-RCNNCN✓X31.1/0.427.1/0.6/7.627.0/0.524.7/0.5/6.9
PL + Mask-RCNNCB ∪ CN✓X32.4/22.429.3/22.8/24.527.4/18.425.0/18.3/20.1
PL + Mask-RCNNCB ∪ CN✓✓35.9/40.740.0/31.5/37.731.2/36.736.0/28.7/34.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.268 + ], + "angle": 0, + "content": "ically, the quality representation is learned due to the quality proposal generated by WSPN compared to fully-supervised RPN and RCNN proposal generators." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.473, + 0.542 + ], + "angle": 0, + "content": "Instance Segmentation: Table 2 compares our method with previous open-vocabulary instance segmentation methods on the MS-COCO and Open Images datasets. SB [2] and BA-RPN [51] are zero-shot methods which utilize different background modelling strategies and caption pre-training to learn novel categories. Compared to these, our method improves the novel category performance by a large margin on both datasets and settings. When compared against conventional pseudo-labelling methods, such as soft-teacher [43] and unbiased teacher [30], our method significantly improves on MS-COCO and Open Images datasets. Finally, when compared to open-vocabulary methods such as OV-RCNN [46], XPM [17], our method outperforms by 4.6 and \\(3.1\\mathrm{mAP}\\) in COCO and Open Images in generalized setting, respectively. All these comparisons are performed against our method trained with just pseudo-labels and no base annotation is used during training. This shows the effectiveness of our overall pseudo-mask generation pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.231, + 0.575 + ], + "angle": 0, + "content": "4.1. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.472, + 0.903 + ], + "angle": 0, + "content": "GradCAM vs Pseudo-Labels vs Mask-RCNN Training. In Table 3, we analyze the quality of pseudo-labels generated from the GradCAM activation map and our method (PL). After pseudo-label generation, we show how Mask-RCNN training helps to improve the quality of prediction compared to pseudo-labels. Finally, we show how fine-tuning on base annotation improves our method. In the first row, we evaluate the pseudo-mask for novel samples, where the pseudo-mask is generated by normalizing and threshold the GradCAM activation map. In the second row, we evaluate the pseudo-mask for novel samples, where the pseudomask is generated by our method. From Table 3, we can observe the quality of the pseudo-mask generated by our method for novel samples is much better than the GradCAM activation map as a pseudo-mask. From Table 3 third row, we can observe training a Mask-RCNN on pseudo-labels improves the performance on novel categories by modelling fine-grained information. By including pseudo-labels from base categories, we observe that the performance on novel samples further improves. Finally, when fine-tuning on base annotations, the performance on novel categories signifi" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 4. Ablation analysis for object detection and instance segmentation under different language supervision." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.253, + 0.892, + 0.343 + ], + "angle": 0, + "content": "
MethodLanguage SupervisionObject DetectionInstance Segmentatn
Constrd NovelGenrlzd NovelConstrd NovelGenrlzd Novel
OV-RCNN [46]Image-caption COCO27.522.8--
PB-OVD [11]Image-caption COCO-29.1--
PB-OVD [11]Img-Cap COCO, SBU, VG32.330.8--
OursImage-labels COCO35.931.531.028.3
OursImage-caption COCO36.131.831.528.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.345, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Table 5. Ablation analysis for object detection under different proposal generator. All models are fined-tuned on COCO base." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.374, + 0.869, + 0.459 + ], + "angle": 0, + "content": "
MethodProposal GeneratorConstrained NovelGeneralized Novel
OV-RCNN [46]RPN COCO Base27.822.8
PB-OVD [11]RCNN COCO Base32.330.8
OursSelective Search34.531.0
OursWSPN COCO Base35.931.5
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.463, + 0.891, + 0.492 + ], + "angle": 0, + "content": "cantly improves by learning task/domain specific information from noise-free manual annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.734 + ], + "angle": 0, + "content": "Language supervision. Given an image and caption pair, our method can generate a pseudo-mask leveraging a pretrained vision-language model. Thus to analyze the effect of captions, we conduct experiments between human-provided captions and pseudo-captions generated from image-labels. As show in Table 4, human-provided captions and image-labels based pseudo-caption produce similar performance showing that irrespective of caption type, our method can generate pseudo-mask for the object of interest (see supplementary material for visual comparison). Therefore, our method is more data efficient as it requires cheap image-labels compared to human-provided captions. Table 4 compares our method with other pseudo-label generation methods trained with extra language supervision [11]. Our method, with lesser language supervision, outperforms [11] by a considerable margin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Proposal Generator Quality vs Performance. In general, better quality of proposal provides better quality of pseudolabels. Therefore, we generated pseduo-labels using different proposal generator and the results are reported in Table 5. As shown in Table 5, our method trained with WSPN as proposal generator is produces better performance compared to methods which rely on fully-supervised proposal generator such as RPN and RCNN. Also when compared to selective search as proposal generator, WSPN demonstrates better performance for novel categories. This is because WSPN refines selective search proposals and localizes them" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23545" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.092, + 0.235, + 0.117 + ], + "angle": 0, + "content": "A woman holding an umbrella." + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.118, + 0.236, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.104, + 0.396, + 0.116 + ], + "angle": 0, + "content": "A photo of Dog." + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.117, + 0.46, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.118, + 0.65, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.118, + 0.844, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.237, + 0.462, + 0.252 + ], + "angle": 0, + "content": "A) Visualization of activation map and pseudo-bbox \\((\\mathbf{b}^{*})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.238, + 0.738, + 0.251 + ], + "angle": 0, + "content": "B) Pseudo-mask annotations" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.255, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Figure 5. A) Pseduo bounding box selection guided by GradCAM activation. B) Visualization of pseudo-mask generated for Open Images." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.271, + 0.27, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.387, + 0.23, + 0.398 + ], + "angle": 0, + "content": "(a) Object detection" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.271, + 0.465, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.387, + 0.441, + 0.398 + ], + "angle": 0, + "content": "(b) Instance segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.404, + 0.47, + 0.461 + ], + "angle": 0, + "content": "Figure 6. Ablation study for object detection and instance segmentation by increase number of masking iteration during PL generation. As iteration increases, GradCAM activation becomes noisy resulting in poor guidance and pseudo-mask generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.415, + 0.478 + ], + "angle": 0, + "content": "towards objects producing better quality proposals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.469, + 0.644 + ], + "angle": 0, + "content": "Iterative Masking Steps vs Performance. Fig. 6 presents ablation analysis for iterative masking hyper-parametr \\( G \\). We see that for initial few masking steps the performance increase. Because, without any masking the GradCAM activation map is towards most discriminative parts of an object. As we mask out these regions, the less discriminative regions gets activated and combining activation from previous steps produce a strong guidance function \\( \\Phi_t \\). However, performing too many mask iterations could completely mask out the object of interest and unrelated background regions might get activated resulting in a poor \\( \\Phi_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Weakly-supervised Instance Segmentation. Weakly-supervised Instance Segmentation (WSIS) methods performs instance-segmentation using supervision from image-labels. Similar to WSIS setting, our method performs instance-segmentation using just image-labels. Following this, we simply extended our method to WSIS setting by comparing with VOC benchmark. PENet [12] is a recent WSIS method, which relies on multiple downstream task network such as segmentation, detection and multi-class classification to perfrom segmentation. CL [18] and IAM [57] are WSIS methods utilize advanced attention mechanism to generate instance segmentation mask. As shown in Table 6, our method outperforms existing WSIS methods by considerable margin. Thus, the proposed method is simple and can be easily extended to OV setting or WSIS setting producing SOTA performance. Note that, other OV methods cannot be extended to WSIS setting as they rely on base" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.276, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Table 6. Instance segmentation(mAP) performance for VOC in weakly-supervised instance segmentation setting." + }, + { + "type": "table", + "bbox": [ + 0.574, + 0.305, + 0.812, + 0.417 + ], + "angle": 0, + "content": "
MethodSupervisionmAP
Mask-RCNN [15]F67.9
OCIS [8]I26.8
PRM [55]I28.3
IAM [57]I30.2
CL [18]I+C30.2
PENet [12]I38.1
OursI38.6
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.428, + 0.892, + 0.458 + ], + "angle": 0, + "content": "annotation and similarly WSIS methods cannot be extended to OV setting due to constrained vocabulary space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.673 + ], + "angle": 0, + "content": "Qualitative Analysis. Fig. 5 (A) presents a visualization of activation maps generated for the object of interest (woman and dog). As we can see, the generated activation map covers the entire object and it can be used as a guidance function to choose the best bounding box proposal. Note that the activation are square-shaped because the original activation map is \\(1/16\\)th of the image size. We perform nearest interpolation to obtain an activation map of image size. In Fig 5 (B), we visualize the pseudo-mask for Open Images generated from our pipeline. We can observe that the generated pseudo-mask is of good quality; still contains some false positives. However, with Mask-RCNN training, the model learns to filter the noise present in the pseudo-mask producing better-quality predictions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.683, + 0.619, + 0.698 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose a novel pipeline called Mask-free OVIS for open-vocabulary instance segmentation, which eliminates the need for costly and time-consuming manual mask annotations. We achieve this by generating pseudo-mask annotation for base and novel categories by leveraging a pretrained vision-language model and weakly supervised techniques. The generated pseudo-mask annotations are then used to supervise the Mask-RCNN model. Extensive experiments are conducted to show the effectiveness of our method on MS-COCO and Open Images datasets. Instead of relying on labour-expensive instance-level annotations, we hope our simple and effective pipeline provides an alternative to detect and segment long-tailed novel categories." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23546" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Aditya Arun, CV Jawahar, and M Pawan Kumar. Dissimilarity coefficient based weakly supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9432-9441, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 384-400, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.284 + ], + "angle": 0, + "content": "[3] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In European conference on computer vision, pages 549-565. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] David Biertimpel, Sindi Shkodrani, Anil S Baslamisli, and Nora Baka. Prior to segment: Foreground cues for weakly annotated classes in partially supervised instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2824-2833, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.41 + ], + "angle": 0, + "content": "[5] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2846-2854, 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.47, + 0.495 + ], + "angle": 0, + "content": "[6] Maria A Bravo, Sudhanshu Mittal, and Thomas Brox. Localized vision-language matching for open-vocabulary object detection. In Pattern Recognition: 44th DAGM German Conference, DAGM GCR 2022, Konstanz, Germany, September 27–30, 2022, Proceedings, pages 393–408. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.47, + 0.578 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, Alexander Hermans, George Papandreou, Florian Schroff, Peng Wang, and Hartwig Adam. Masklab: Instance segmentation by refining object detection with semantic and direction features. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4013-4022, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[8] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12405, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.705, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.776, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 266-282. Springer, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.86, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[12] Weifeng Ge, Sheng Guo, Weilin Huang, and Matthew R Scott. Label-penet: Sequential label propagation and enhancement networks for weakly supervised instance segmen-" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "tation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3345-3354, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[13] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. CyCLIP: Cyclic contrastive language-image pretraining. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.204, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[16] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Advances in neural information processing systems, 33:16797-16807, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[17] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[18] Jaedong Hwang, Seohyun Kim, Jeany Son, and Bohyung Han. Weakly supervised instance segmentation by deep community learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1020-1029, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[20] Wonjik Kim, Asako Kanezaki, and Masayuki Tanaka. Unsupervised learning of image segmentation based on differentiable feature clustering. IEEE Transactions on Image Processing, 29:8055-8068, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[21] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[22] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[23] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for uni" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23547" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "fied vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.471, + 0.263 + ], + "angle": 0, + "content": "[26] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell me where to look: Guided attention inference network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[27] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.471, + 0.404 + ], + "angle": 0, + "content": "[28] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.471, + 0.476 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.471, + 0.546 + ], + "angle": 0, + "content": "[30] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In International Conference on Learning Representations, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.604 + ], + "angle": 0, + "content": "[31] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.471, + 0.688 + ], + "angle": 0, + "content": "[32] Zongyang Ma, Guan Luo, Jin Gao, Liang Li, Yuxin Chen, Shaoru Wang, Congxuan Zhang, and Weiming Hu. Open-vocabulary one-stage detection with hierarchical visual-language knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14074-14083, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[33] Niall O'Mahony, Sean Campbell, Anderson Carvalho, Suman Harapanahalli, Gustavo Velasco Hernandez, Lenka Krpalkova, Daniel Riordan, and Joseph Walsh. Deep learning vs. traditional computer vision. In Science and information conference, pages 128-144. Springer, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[35] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11932–11939, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[36] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In 36th Conference on Neural Information Processing Systems (NIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[37] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.304 + ], + "angle": 0, + "content": "[38] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, MingYu Liu, Yong Jae Lee, Alexander G Schwing, and Jan Kautz. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598-10607, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.893, + 0.387 + ], + "angle": 0, + "content": "[39] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[40] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.893, + 0.53 + ], + "angle": 0, + "content": "[41] Limin Wang, Yuanjun Xiong, Dahua Lin, and Luc Van Gool. Untrimmednets for weakly supervised action recognition and detection. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4325-4334, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.893, + 0.587 + ], + "angle": 0, + "content": "[42] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In European Conference on Computer Vision, pages 649-665. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[43] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[44] Ke Yang, Dongsheng Li, and Yong Dou. Towards precise end-to-end weakly supervised object detection network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8372-8381, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.893, + 0.788 + ], + "angle": 0, + "content": "[45] Keren Ye, Mingda Zhang, Adriana Kovashka, Wei Li, Danfeng Qin, and Jesse Berent. Cap2det: Learning to amplify weak caption supervision for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9686-9695, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[46] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[47] Zhaoyang Zeng, Bei Liu, Jianlong Fu, Hongyang Chao, and Lei Zhang. Wsod2: Learning bottom-up and top-down objectness distillation for weakly-supervised object detection." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23548" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF international conference on computer vision, pages 8292-8300, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[48] Dingwen Zhang, Junwei Han, Gong Cheng, and Ming-Hsuan Yang. Weakly supervised object localization and detection: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5866-5885, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[49] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[50] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 159-175. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[51] Ye Zheng, Jiahong Wu, Yongqiang Qin, Faen Zhang, and Li Cui. Zero-shot instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2593-2602, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[52] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.469, + 0.555 + ], + "angle": 0, + "content": "[53] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[54] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[55] Yanzhao Zhou, Yi Zhu, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Weakly supervised instance segmentation using class peak response. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3791-3800, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.767 + ], + "angle": 0, + "content": "[56] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't even look once: Synthesizing features for zero-shot detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11693-11702, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.77, + 0.469, + 0.839 + ], + "angle": 0, + "content": "[57] Yi Zhu, Yanzhao Zhou, Huijuan Xu, Qixiang Ye, David Doermann, and Jianbin Jiao. Learning instance activation maps for weakly supervised instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3116-3125, 2019. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23549" + } + ] +] \ No newline at end of file diff --git a/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_origin.pdf b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5acec8b17ed452fab8699f8ae7b042c0c895946f --- /dev/null +++ b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/d8a0bc48-f3dc-4cd8-b893-b6a4dd0a47ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b30eadf3e0d400073c63aebb26ef36cd090dc27a3b622ec8fa97f6af6748032f +size 573904 diff --git a/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/full.md b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b571ecebf758ed6f8ee8e8c70ba8a26484e9011c --- /dev/null +++ b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/full.md @@ -0,0 +1,286 @@ +# Mask-free OVIS: Open-Vocabulary Instance Segmentation without Manual Mask Annotations + +Vibashan VS\*, Ning Yu†, Chen Xing†, Can Qin‡, Mingfei Gao†, Juan Carlos Niebles†, Vishal M. Patel\*, Ran Xu† + +*Johns Hopkins University, ‡Northeastern University, †Salesforce Research + +{vvishnu2, vpatel36}@jhu.edu, qin.ca@northeastern.edu, + +{ning.yu, cxing, jniebles, ran.xu}@salesforce.com + +# Abstract + +Existing instance segmentation models learn task-specific information using manual mask annotations from base (training) categories. These mask annotations require tremendous human effort, limiting the scalability to annotate novel (new) categories. To alleviate this problem, Open-Vocabulary (OV) methods leverage large-scale image-caption pairs and vision-language models to learn novel categories. In summary, an OV method learns task-specific information using strong supervision from base annotations and novel category information using weak supervision from image-captions pairs. This difference between strong and weak supervision leads to overfitting on base categories, resulting in poor generalization towards novel categories. In this work, we overcome this issue by learning both base and novel categories from pseudomask annotations generated by the vision-language model in a weakly supervised manner using our proposed Mask-free OVIS pipeline. Our method automatically generates pseudo-mask annotations by leveraging the localization ability of a pre-trained vision-language model for objects present in image-caption pairs. The generated pseudomask annotations are then used to supervise an instance segmentation model, freeing the entire pipeline from any labour-expensive instance-level annotations and overfitting. Our extensive experiments show that our method trained with just pseudo-masks significantly improves the mAP scores on the MS-COCO dataset and OpenImages dataset compared to the recent state-of-the-art methods trained with manual masks. Codes and models are provided in https://vibashan.github.io/ovis-web/. + +# 1. Introduction + +Instance segmentation is a challenging task as it requires models to detect objects in an image while also precisely + +![](images/a49cee1ae68a616cd58806ea83eadc22c8d47a685d2fb4326a991229a3386be0.jpg) + +![](images/919bacb6b2095f72cc721d1ac5f5ca9082c6a24b847e596aab35e6a4e33243ac.jpg) +Figure 1. A) Previous Methods: Learn task-specific information (detection/segmentation) in a fully-supervised manner and novel category information with weak supervision. During training, this difference in strong and weak supervision signals leads to overfitting and requires expensive base annotations. B) Our method: Given image-caption pairs, we generate pseudo-annotations for both base and novel categories under weak supervision, solving the problems of labour-expensive annotation and overfitting. + +segment each object at the pixel-level. Although the rise of deep neural networks has significantly boosted the state-of-the-art instance segmentation performance [7, 15, 42], these methods are still trained for a pre-defined set of object categories and are data-hungry [33]. Particularly, one needs to manually annotate thousands of instance-level masks for each object category, which takes around 78 seconds per instance mask [3]. If we look at this quantitatively on Open Images [21], a large-scale dataset with 2.1M instance-level mask annotations requires around 5 years of human labour. Even after extensive annotation, these training datasets are still limited to a small number of categories and segmenting objects from a novel category requires further annotation. Therefore, it is difficult to scale up existing methods to segment a large number of categories due to intensive labour. + +Recently, Open-Vocabulary (OV) methods have gained much attention due to their success in detecting [11, 14, + +![](images/58b34a10e529132414d0d16d8145a05ece8d697b3e81726db38591b9fcc86259.jpg) +Figure 2. RPN is supervised using bounding box annotations from COCO base and WSPN is supervised using image-labels from COCO base. A) WSPN produces better quality proposals for novel object categories compared to fully-supervised RPN. B) WSPN consistently produces better recall for all COCO novel categories than RPN. + +![](images/cbc981d08377389e6e16d676f69b6fdb3338f429e6c728d6041248ef22bb903b.jpg) + +46, 52] and segmenting [17] novel categories beyond base (training) categories. An OV method learns task-specific information (detection/segmentation) from base categories that have manual instance-level bounding boxes or masks and learns novel category information from the pre-trained Vision-Language Model (VLM) [19, 34] (see Fig. 1). All these methods produce promising results on novel categories by leveraging different forms of weak supervision such as caption pretaining [46, 50], knowledge distillation [14, 52] and pseudo-labelling [17, 28]. However, all above-mentioned OV methods still rely on the manually-annotated base categories to improve their performances on novel categories. Without fine-tuning on base categories, existing OV methods lack task/domain specific knowledge and the performances on novel categories will be affected [14, 46]. + +Although manual instance-level annotations of base categories are critical to open-vocabulary segmentation methods, we find empirically that such fully-supervised information causes OV methods to overfit to base categories, leading to a higher failure rate when evaluated on novel categories. Specifically, OV methods utilize a region proposal network (RPN) [37] supervised with bounding box annotations obtained from the base categories to generate a set of bounding box proposals for all objects categories in a given image [46]. The feature representation from these object proposals is later matched with text embedding to learn the visual-semantic space for base and novel categories [46]. Therefore, the quality of proposals generated for novel object categories plays a key role in determining the performance in later stages. However, from our experiments, we find that many objects of novel categories wouldn't be included in such proposals due to the RPN's overfitting to base categories. Fig. 2 (A) - Top gives some examples where the RPN trained with COCO base categories fails to generate high-quality region proposals for novel categories such as elephant, cat and skateboard. Therefore, a fully + +supervised proposal network is a bottleneck in OV pipeline due to its poor generalization towards novel categories. + +Given the aforementioned observations of poor generalization, we raise the question of whether we can improve the generalization by using weak supervision instead of relying on strong supervision from manual base annotations. If so, we can reduce overfitting to the base categories and the requirement for costly instance-level human annotations can be entirely removed from the pipeline. Our preliminary experiments give us some hope. Our experiments show that if we train a weakly-supervised proposal network (WSPN) with image-level annotations instead of box-level annotations, the region proposals it generates can better generalize to novel objects. As shown in Fig. 2 A), the novel objects that the RPN proposals miss are covered by WSPN. Fig. 2 B) shows WSPN proposals have consistently better average recall than RPN for all COCO novel categories, indicating that WSPN proposals are more likely to cover the ground-truth bounding boxes of novel objects. + +Inspired by these observations, we propose open-vocabulary segmentation without manual mask annotations. We do not use any human-provided box-level or pixel-level annotations during the training of our method. We first train a simple WSPN model with image-level annotations on base categories as a proposal generator to generate proposals for all objects given an image. Then, we adopt pre-trained vision-language models to select proposals as pseudo bounding boxes for novel objects. Given a novel object's text name, we can utilize the name as a text prompt to localize this object in an image with a pre-trained vision-language model. To obtain a more accurate pseudo-mask that covers the entire object, we conduct iterative masking with GradCAM [39] given the vision-language model. Finally, we train a weakly-supervised segmentation (WSS) [20] network with previously generated bounding box and GradCAM activation map to obtain pixel-level annotation. + +Our contributions are summarized as follows: (1) We propose a Mask-free OVIS pipeline where we produce manual-effort-free pseudo-mask annotations for base and novel instance segmentation using open-vocabulary and weakly supervised techniques. (2) We propose a novel pseudo-mask generation pipeline leveraging a pre-trained vision-language model to generate instance-level annotations. (3) Benefiting from pseudo-labels, our method sets up SOTA's for both detection and instance segmentation tasks compared to recent methods trained with manual masks on MS-COCO and OpenImages datasets. + +# 2. Related Work + +Weakly-supervised learning (WSL) methods require less manual annotation effort and only rely on partial information about the task compared to fully-supervised methods. WSL applications include object detection [1, 16, 44, 47], segmentation [22,23,40], action recognition, etc [26,41,48]. The class activation map (CAM) [53] and grad-CAM [39] are popular approaches which exploits pre-trained network to generate saliency or activation map for target classes. Li et al. [26] enables weakly-supervised semantic segmentation via CAM-guided attention. + +Vision-Language Models. With the rise of large-scale text-pretraining with attention-based models [9, 49], vision-language models (VLM) have caught increasing attention due to their strong performance in downstream visual understanding tasks [27, 31, 34]. Early VLM models [27, 31] applied the pre-trained models with constraints of the class set. With the help of contrastive learning, CLIP [34] has enabled VLM in the wild with large-scale multimodal learning upon the 400M noisy data crawled from the web. Several methods have extended CLIP for high efficiency model training and cycle consistency [13, 24, 25]. ALBEF [25] utilizes cross-modal attention between image patches and text, leading to more grounded vision and language representation learning. Compared to CLIP, ALBEF excels in object localization, while CLIP performs better at zero-shot classification due to its larger training dataset. + +Open-Vocabulary. Open-Vocabulary methods [6, 10, 32, 36, 46, 50] scale up their vocabulary size for object detection or instance segmentation tasks by transferring knowledge from pre-trained vision-language models. Zareian et al. [46] proposed an open-vocabulary object detection method where a visual encoder is trained on image-caption pair to model object semantics and then transfer to zero-shot object detection. ViLD [14] and RegionCLIP [52] use pre-trained CLIP [34] to distill knowledge and enhance the vision-text embedding space effectively. Similarly, Huynh et al. [17] proposes a robust pseudo-labelling, for instance segmentation where a teacher model generates pseudo-label for novel categories and is later used to supervise a student model. Therefore, all these methods require manually + +annotated base categories to improve their performance on novel categories. In contrast, our method generates pseudolabels for both base and novel categories by leveraging a pre-trained VLM, freeing the entire pipeline from human-provided instance-level annotations. Similar to our method, Gao et al. [11] generates pseudo-labels using pre-trained VLM, but their approach still uses a proposal generator that relies on human-provided bounding boxes for the base categories. Our method, on the other hand, requires no such annotations, making it completely free from any human-provided box-level or pixel-level annotations. + +# 3. Mask-free OVIS + +Our pipeline consists of two stages: (I) Pseudo-mask generation, and (II) Open-vocabulary instance segmentation. The goal of (I) is to generate box-level and pixel-level annotations by leveraging the region-text alignment property of a pre-trained vision-language model as illustrated in Fig. 3. In (II), we train a Mask-RCNN using the generated pseudo annotations. + +# 3.1. Pseudo-Mask Generation + +Region-Text Attention Scores. The inputs to the vision-language model are image $\mathbf{I}$ and caption $\mathbf{C} = \{c_1, c_2, \dots, c_{N_c}\}$ pair, where $N_c$ is the number of words in the caption (including [CLS] and [SEP]) [9, 25]. In a VLM, a text encoder is utilized to get text representations $\mathbf{T} \in \mathbb{R}^{N_C \times d}$ and an image encoder is utilized to extract region representation $\mathbf{R} \in \mathbb{R}^{N_R \times d}$ , where $N_R$ is the number of regions in the image. To fuse the information from both image and text encoders, a multi-modal encoder with $M$ consecutive cross-attention layers is utilized [19, 25]. These cross-attention layers learn a good region-text alignment [25, 34] and are utilized to obtain regions corresponding to the object of interest $c_t$ from the caption. In particular, for the $m$ -th cross-attention layer, the visual region attention scores $\mathbf{X}_t^m$ for the object of interest $c_t$ is calculated as: + +$$ +\mathbf {X} _ {t} ^ {m} = \operatorname {S o f t m a x} \left(\frac {\mathbf {h} _ {t} ^ {m - 1} \mathbf {R} ^ {T}}{\sqrt {d}}\right), \tag {1} +$$ + +$$ +\mathbf {h} _ {t} ^ {n} = \mathbf {X} _ {t} ^ {m} \cdot \mathbf {R}. \tag {2} +$$ + +where $d$ is a scalar and $\mathbf{h}_t^{m-1}$ is the hidden representation obtained from the previous $(m-1)$ -th cross-attention layer. GradCAM Activation Map. After obtaining attention scores $\mathbf{X}_t^m$ , we employ Grad-CAM [39] to visualize the activated regions. As in [39], we take the image-caption similarity $(S)$ output from the multi-modal encoder's final layer and calculate the gradient with respect to the attention scores. The activation map $\phi_t$ for object $c_t$ is: + +$$ +\phi_ {t} = \mathbf {X} _ {t} ^ {m} \cdot \max \left(\frac {\partial S}{\partial \mathbf {X} _ {t} ^ {m}}, 0\right). \tag {3} +$$ + +Iterative Masking. During VLM training, an object's most discriminative regions easily get aligned towards object text representation [25, 34]. As a result, $\phi_t$ is localized towards + +![](images/e6e0885f1a8316d49654e267400d7cc5c755ea0329656cd230f818899ba46f98.jpg) +Figure 3. Illustrative overview of our pseudo-mask generation pipeline. Given an image-caption pair and pre-trained VLM, we generate an activation map for the object of interest ("umbrella") and enhance it using iterative masking strategy. We generate box-level annotations using an activation map as a guidance function to select the best WSPN proposals covering the object. We crop the image corresponding to the generated pseduo bounding box and perform weakly-supervised segmentation to obtain pixel-level annotations. + +the most discriminative region and fails to cover the object completely [39]. However, when we mask out the most discriminative regions, GradCAM activations are shifted towards other discriminative regions (see Fig. 4). We propose a simple iterative masking strategy to obtain better activation where the most activated part of the object is replaced with image mean and the new activation map is computed following Eq. 1 and 3. The final activation map is: + +$$ +\boldsymbol {\Phi} _ {t} = \bigcup_ {i = 1} ^ {G} \mathcal {I M} \left(\phi_ {t} ^ {i}\right) \tag {4} +$$ + +where $G$ is a hyper-parameter indicating the number of masking iterations and $\mathcal{IM}(\cdot)$ normalize and threshold $\phi_t$ by 0.5. We utilize the activation map $\Phi_t$ as a guidance function to generate box-level and pixel-level annotations. + +Weakly-Supervised Proposal Network. To generate box-level annotations, we require bounding box proposals covering the activated region $\Phi_t$ . As explained in the introduction, the fully-supervised RPN network produces poor proposals for novel categories, making it less preferable. To overcome this, we propose a weakly-supervised proposal network (WSPN), which generates object proposals trained in a weakly-supervised manner. Given an Image I, the WSPN is supervised with image-labels $\mathbf{Y} = \{y_1, y_2, \dots, y_C\}$ , where $y_c = 0$ or 1 indicates the absence or presence of class $c$ in $\mathbf{I}$ and $C$ denotes total number of classes [5]. First, we utilize Selective Search to generate a set of unsupervised proposal $\mathbf{U} = \{u_1, u_2, \dots, u_N\}$ where $N$ is the total number of bounding box proposals. Then, the Image I and proposal $\mathbf{U}$ are fed into a CNN backbone + +![](images/8416f981e3aff51cf2fa537af902c1472958da128a51fa9440c26527c5e37739.jpg) +Figure 4. Comparison between activation map $\Phi_t$ generated at $G$ 0 and 1. For $G = 0$ , the most discriminative parts of an object (bird's head) gets activated (bird's head). After masking, for $G = 1$ we can observe that the activation map has shifted to less discriminative part (bird's body). Thus, by combining activation from both steps, we obtain a better activation map trying to cover entire object $\Phi_t$ . + +to extract features and RoI pooling layer [37] to obtain RoI pooled feature vectors. Following [5], the pooled feature vectors are passed to a classification and detection branch to generate two matrices $\mathbf{W}^{cls}$ , $\mathbf{W}^{det} \in \mathbb{R}^{C \times N}$ . Then, $\mathbf{W}^{cls}$ and $\mathbf{W}^{det}$ matrices are normalized along the category direction (column-wise) and proposal direction (row-wise) by the softmax layers $\sigma(\cdot)$ respectively. From $\mathbf{W}^{cls}$ and $\mathbf{W}^{det}$ , the instance-level classification scores for object proposals are computed by the element-wise product $\mathbf{W}_{\mathbf{C}} = \sigma(\mathbf{W}^{cls}) \odot \sigma(\mathbf{W}^{det})$ . and image-level classification score for the $c_{th}$ class is computed as $p_c = \sum_{i=1}^N w_{i,c}$ . Following [38], we select the high-scoring proposals are used as + +pseudo-regression targets ( $\hat{\mathbf{T}} = \{\hat{t}(u_1), \hat{t}(u_2), \dots, \hat{t}(u_N)\}$ ) for low-scoring proposals to make sure objects can be more tightly captured. To this end, the classification loss and regression loss for WSPN are calculated as: + +$$ +\begin{array}{l} \mathcal {L} _ {w s p n} = - \sum_ {c = 1} ^ {C} y _ {c} \log p _ {c} + (1 - y _ {c}) \log \left(1 - p _ {c}\right) + \tag {5} \\ \frac {1}{N} \sum_ {u = 1} ^ {N} \mathcal {L} _ {\text {s m o o t h L 1}} (\hat {t} (u _ {i}), u _ {i}). \\ \end{array} +$$ + +WSPN is trained to localize and classify objects by minimizing Eq. 5. The trained WSPN model is used to generate object proposals and top $K$ proposal candidates over all the classes $\mathcal{B} = \{\mathbf{b}_1, \mathbf{b}_2, \dots, \mathbf{b}_K\}$ are selected by sorting proposal confidence scores obtained from $\mathbf{W}^{det}$ . Finally, from the top proposal candidates $\mathcal{B}$ , we select the proposal which overlaps the most with $\Phi_t$ as pseudo box-bounding: + +$$ +\mathbf {b} ^ {*} = \underset {\mathbf {b} \in \mathcal {B}} {\arg \max } \frac {\sum_ {\mathbf {b}} \Phi_ {t}}{\sqrt {| \mathbf {b} |}}, \tag {6} +$$ + +where $\mathbf{b}^*$ is the pseudo box-level annotation and $\sum_{\mathbf{b}}\Phi_t$ indicates the summation of the activation map values within a box proposal $\mathbf{b}$ , and $|\mathbf{b}|$ indicates the proposal area. + +Weakly-Supervised Segmentation. Once we obtain the pseudo bounding box $\mathbf{b}^*$ , we crop image $\mathbf{I}$ to obtain the corresponding image patch. The cropped patch is then fed into a simple three-layer CNN network to perform pixel-level segmentation. To supervise the CNN network, we generate pseudo ground-truth $\Theta$ using $\Phi_t$ and $\mathbf{b}^*$ , where we sample $Z$ points as foreground $F_z = \{f_i\}_{i=1,\dots,Z}$ and background $B_z = \{b_i\}_{i=1,\dots,Z}$ and each point is set to 1 or 0, respectively. Specifically, the foreground and background points are sampled from the most and least activated part of $\Phi_t$ inside $\mathbf{b}^*$ . The pseudo ground-truth $\Theta$ is of size $\mathbf{b}^*$ and we supervise the network predictions only at sampled points. Thus, the segmentation loss obtained from these weak points is computed as follows: + +$$ +\mathcal {L} _ {w s s} = \sum_ {i = 1} ^ {G} \mathcal {L} _ {c e} \left(\mathbf {s} ^ {*} \left(f _ {i}\right), \boldsymbol {\Theta} \left(f _ {i}\right)\right) + \sum_ {i = 1} ^ {G} \mathcal {L} _ {c e} \left(\mathbf {s} ^ {*} \left(b _ {i}\right), \boldsymbol {\Theta} \left(b _ {i}\right)\right), \tag {7} +$$ + +where $\mathbf{s}^*$ is the pseudo pixel-level annotation of size $\mathbf{P}$ and $\mathcal{L}_{ce}$ indicates cross-entropy loss. + +To this end, given an image $\mathbf{I}$ and caption $\mathbf{C}$ pair, we generate pseudo box-level $\mathbf{b}^*$ and pixel-level $\mathbf{s}^*$ annotation for the object of interest $c_{t}$ . In practice, the pseduo-mask annotations are generated for a pre-defined set of object categories obtained from training vocabulary. Fig. 5 visualizes the generated pseduo-mask annotations for various categories. Activated regions correspond well with objects of interest and the generated pseduo-mask are of good quality. + +# 3.2. Open-Vocabulary Instance Segmentation + +After generating pseudo-mask annotations, we train an open-vocabulary instance segmentation model. Following + +[46], we employ a Mask-RCNN as the instance segmentation model, where a class-agnostic mask head is utilized to segment objects and the classification head is replaced with embedding head $h_{emb}$ . Given a Image I, an encoder network extracts image features and region embeddings, $\mathbf{R} = \{\mathbf{r}_i\}_{i=1,\dots,N_r}$ , are obtained by RoI align [15] followed by a fully connected layer, where $N_r$ denotes the number of regions. The similarity between the region and text embedding pair is calculated as follows: + +$$ +p \left(\mathbf {r} _ {i}, \mathbf {c} _ {j}\right) = \frac {\exp \left(h _ {e m b} \left(\mathbf {r} _ {i}\right) \cdot \mathbf {c} _ {j}\right)}{\exp \left(h _ {e m b} \left(\mathbf {r} _ {i}\right) \cdot \mathbf {b g}\right) + \sum_ {k} \exp \left(h _ {e m b} \left(\mathbf {r} _ {i}\right) \cdot \mathbf {c} _ {k}\right)}, \tag {8} +$$ + +where, $\mathbf{C} = \{\mathbf{bg},\mathbf{c}_1,\mathbf{c}_2,\dots,\mathbf{c}_{N_C}\}$ , are object vocabulary text representation obtained from pre-trained text encoder, where $N_{C}$ is the training object vocabulary size. To learn the semantic space, negative pairs are pushed away and positive pairs are pulled together using cross entropy loss obtained from $\mathbf{b}^*$ . The class-agnostic mask head is supervised by minimizing standard segmentation loss [15] obtained from $\mathbf{s}^*$ . During inference, the similarity between the region proposals embedding and text embedding from a group of object classes of interest is calculated. The region is then assigned to a class with the highest similarity. + +# 4. Experiments and Results + +Datasets. Following [46], we conduct experiments on MS-COCO [29] with data split of 48 base categories and 17 novel categories. The processed COCO dataset contains 107,761 training images and 4,836 test images. Following [17], we conduct experiments on Open Images [21] to verify the effectiveness of our method on the large-scale dataset. The Open Images dataset consists of 300 categories with a class split of 200 base categories (frequent objects) and 100 novel categories (rare objects). Following [34], we leverage image-labels obtained from MS-COCO and Open Images to learn the novel category information. We also experiment using image-caption datasets to show our method's effectiveness irrespective of training vocabulary. + +Evaluation Metrics and Protocols. Following open-vocabulary methods [17, 52], for both detection and segmentation tasks, we report the mean Average Precision at intersection-over-union (IoU) of $0.5\left(mAP_{50}\right)$ . Following zero-shot settings [46], we report novel category performance for both constrained setting and generalized setting. In constrained setting, the model is evaluated only on novel class test images and in generalized setting, the model is evaluated on both base and novel class test images. + +Implementation Details. In pseudo-mask generation framework, we use pre-trained ALBEF [25] as our vision-language model. We conducted all our pseudo-mask generation experiments using ALBEF due to the good region-text alignment when image and caption pair are present [25]. Following ALBEF, the cross-attention layer $m$ used for + +Table 1. Object Detection (mAP) performances for MS-COCO under constrained and generalized setting. $\mathcal{C}_{\mathrm{B}}$ and $\mathcal{C}_{\mathrm{N}}$ are subset of $\mathcal{C}_{\Omega}$ , where $\mathcal{C}_{\Omega}$ contains training vocabulary larger than COCO categories. + +
MethodProposal GeneratorLanguage SupervisionBase AnnotationConstrained NovelGeneralized Novel
WSDDN [5]-Image-labels in C_B ∪ C_NX-19.7
Cap2Det [45]-Image-labels in C_B ∪ C_NX-20.3
SB [2]RPN COCObase-✓0.700.31
DELO [56]RPN COCObase-✓7.603.41
PL [35]RPN COCObase-✓10.04.12
OV-RCNN [46]RPN COCObaseImage-caption in C_B ∪ C_N✓27.522.8
CLIP-RPN [14]RPN COCObaseCLIP image-text pair C_Ω✓-26.3
ViLD [14]RPN COCObaseCLIP image-text pair C_Ω✓-27.6
Detic [54]RPN COCObaseImage-caption in C_B ∪ C_N✓-27.8
RegionCLIP [52]RPN LVISbaseConceptual caption C_Ω✓30.826.8
PB-OVD [11]RCNN COCObaseImage-caption in C_B ∪ C_N✓32.330.7
XPM [17]RPN COCObaseImage-caption in C_B ∪ C_N✓29.927.0
Mask-free OVIS (Ours)WSPN COCObaseImage-labels in C_B ∪ C_NX31.527.4
Mask-free OVIS (Ours)WSPN COCObaseImage-labels in C_B ∪ C_N✓35.931.5
+ +Table 2. Instance Segmentation (mAP) performances for MS-COCO and Open Images under constrained and generalized setting. + +
MethodProposal Generator (MS-COCO/OpenImages)Base AnnotationMS-COCOOpen Images
Constrained NovelGeneralized NovelConstrained NovelGeneralized Novel
OVR+OMP [4]-✓14.18.324.916.8
SB [2]-✓20.816.024.817.3
BA-RPN [51]-✓20.115.425.316.9
Soft-Teacher [43]RPN COCObase/RPN OpenImgbase✓14.89.625.917.6
Unbiased-Teacher [30]RPN COCObase/RPN OpenImgbase✓15.19.822.214.5
OV-RCNN [46]RPN COCObase/RPN OpenImgbase✓20.917.123.817.5
XPM [17]RPN COCObase/RPN OpenImgbase✓24.021.631.622.7
Mask-free OVIS (Ours)WSPN COCObase/WSPN COCObase✗27.425.035.925.8
+ +Grad-CAM visualization is set to 8. For attention score, we directly employ the original setting of ALBEF and no additional modification is performed. Note that other pretrained vision-language models can also be integrated into our pipeline without major modifications. For the proposal generation pipeline, the WSPN network is trained using COCO base image-labels and the top $K$ proposals candidates is set to 50. The WSPN network is trained for 40k iterations with learn rate 0.001 and weight decay 0.0001. For iterative masking, the hyper-parameter $G$ is set to 3. In the segmentation pipeline, for each patch, the segmentation network is trained for 500 iterations with lr 0.25. + +For fair-comparison [17, 46], we use Mask R-CNN with a ResNet50 backbone as our open-vocabulary instance segmentation model. During pseudo-mask training, we train Mask-RCNN on MS-COCO and OpenImages using batch size 8 on 8 A5000 GPUs for 90k iterations. Following [11, 14], we use text embeddings obtained from a pretrained CLIP text encoder. During pseduo-mask training, the initial learning rate is set 0.01 and the background class weight is set to 0.2 to improve the recall of novel classes [46]. For base fine-tuning, the initial learning rate is set to 0.0005 and the weight decay is set to 0.0001. We run fine-tuning for 90k iterations where the learning rate is updated + +by a decreasing factor of 0.1 at 60k and 80k iterations. + +Object Detection: As shown in Table 1, we compare our method with previous established open-vocabulary detection methods on the MS-COCO dataset. Compared to weakly-supervised methods such as WSDDN [5] and zero-shot methods such as SB [2], DELO [56], our method outperforms them by a large margin. OV-RCNN [46], XPM [17] are OVD methods based on caption pre-training and our method trained with only pseudo-labels improves the novel category performance by $20.2\%$ and $2.4\%$ in generalized setting, respectively. Also, when compared to the method which leverages pre-trained vision-language models such as ViLD [14], RegionCLIP [52], PB-OVD [11], our method with just pseudo-labels produces similar performance. However, with fine-tuning on base annotations, our method outperforms ViLD [14], RegionCLIP [52], PB-OVD [11] by $13.3\%$ , $14.9\%$ and $2.8\%$ in generalized setting, respectively. This is because with fine-tuning, the model learns task/domain specific information from noise-free annotations, boosting the novel category performance. Even without base annotations, our method outperforms most of the existing OVD methods supervised using base annotations. This shows the effectiveness of our method for learning quality representation for novel categories. Specif + +Table 3. Object detection and Instance segmentation ablation analysis for GradCAM, Pseudo-label and Mask-RCNN training. + +
MethodLanguage SupervisionMask-RCNN TrainingBase AnnotationObject DetectionInstance Segmentation
ConstrainedGeneralizedConstrainedGeneralized
Novel/BaseNovel/Base/AllNovel/BaseNovel/Base/All
GradCAMCNXX8.6/0.0-5.2/0.0-
PLCNXX17.3/0.0-14.8/0.0-
PL + Mask-RCNNCN✓X31.1/0.427.1/0.6/7.627.0/0.524.7/0.5/6.9
PL + Mask-RCNNCB ∪ CN✓X32.4/22.429.3/22.8/24.527.4/18.425.0/18.3/20.1
PL + Mask-RCNNCB ∪ CN✓✓35.9/40.740.0/31.5/37.731.2/36.736.0/28.7/34.0
+ +ically, the quality representation is learned due to the quality proposal generated by WSPN compared to fully-supervised RPN and RCNN proposal generators. + +Instance Segmentation: Table 2 compares our method with previous open-vocabulary instance segmentation methods on the MS-COCO and Open Images datasets. SB [2] and BA-RPN [51] are zero-shot methods which utilize different background modelling strategies and caption pre-training to learn novel categories. Compared to these, our method improves the novel category performance by a large margin on both datasets and settings. When compared against conventional pseudo-labelling methods, such as soft-teacher [43] and unbiased teacher [30], our method significantly improves on MS-COCO and Open Images datasets. Finally, when compared to open-vocabulary methods such as OV-RCNN [46], XPM [17], our method outperforms by 4.6 and $3.1\mathrm{mAP}$ in COCO and Open Images in generalized setting, respectively. All these comparisons are performed against our method trained with just pseudo-labels and no base annotation is used during training. This shows the effectiveness of our overall pseudo-mask generation pipeline. + +# 4.1. Ablation Study + +GradCAM vs Pseudo-Labels vs Mask-RCNN Training. In Table 3, we analyze the quality of pseudo-labels generated from the GradCAM activation map and our method (PL). After pseudo-label generation, we show how Mask-RCNN training helps to improve the quality of prediction compared to pseudo-labels. Finally, we show how fine-tuning on base annotation improves our method. In the first row, we evaluate the pseudo-mask for novel samples, where the pseudo-mask is generated by normalizing and threshold the GradCAM activation map. In the second row, we evaluate the pseudo-mask for novel samples, where the pseudomask is generated by our method. From Table 3, we can observe the quality of the pseudo-mask generated by our method for novel samples is much better than the GradCAM activation map as a pseudo-mask. From Table 3 third row, we can observe training a Mask-RCNN on pseudo-labels improves the performance on novel categories by modelling fine-grained information. By including pseudo-labels from base categories, we observe that the performance on novel samples further improves. Finally, when fine-tuning on base annotations, the performance on novel categories signifi + +Table 4. Ablation analysis for object detection and instance segmentation under different language supervision. + +
MethodLanguage SupervisionObject DetectionInstance Segmentatn
Constrd NovelGenrlzd NovelConstrd NovelGenrlzd Novel
OV-RCNN [46]Image-caption COCO27.522.8--
PB-OVD [11]Image-caption COCO-29.1--
PB-OVD [11]Img-Cap COCO, SBU, VG32.330.8--
OursImage-labels COCO35.931.531.028.3
OursImage-caption COCO36.131.831.528.8
+ +Table 5. Ablation analysis for object detection under different proposal generator. All models are fined-tuned on COCO base. + +
MethodProposal GeneratorConstrained NovelGeneralized Novel
OV-RCNN [46]RPN COCO Base27.822.8
PB-OVD [11]RCNN COCO Base32.330.8
OursSelective Search34.531.0
OursWSPN COCO Base35.931.5
+ +cantly improves by learning task/domain specific information from noise-free manual annotations. + +Language supervision. Given an image and caption pair, our method can generate a pseudo-mask leveraging a pretrained vision-language model. Thus to analyze the effect of captions, we conduct experiments between human-provided captions and pseudo-captions generated from image-labels. As show in Table 4, human-provided captions and image-labels based pseudo-caption produce similar performance showing that irrespective of caption type, our method can generate pseudo-mask for the object of interest (see supplementary material for visual comparison). Therefore, our method is more data efficient as it requires cheap image-labels compared to human-provided captions. Table 4 compares our method with other pseudo-label generation methods trained with extra language supervision [11]. Our method, with lesser language supervision, outperforms [11] by a considerable margin. + +Proposal Generator Quality vs Performance. In general, better quality of proposal provides better quality of pseudolabels. Therefore, we generated pseduo-labels using different proposal generator and the results are reported in Table 5. As shown in Table 5, our method trained with WSPN as proposal generator is produces better performance compared to methods which rely on fully-supervised proposal generator such as RPN and RCNN. Also when compared to selective search as proposal generator, WSPN demonstrates better performance for novel categories. This is because WSPN refines selective search proposals and localizes them + +![](images/f5323d98e0678f660fcdb53d6b94c82a907a0cd30f5fc2d2a3041ba058e053bf.jpg) +A woman holding an umbrella. + +![](images/98b404cd3d4b36803dea95a082eb36911ddce17cb4300c931a5f55a2397d3216.jpg) +A photo of Dog. + +![](images/01d5ff920c96207505dd4bb367db67ed6c6e089a0747e95477739ee9f1272a35.jpg) +A) Visualization of activation map and pseudo-bbox $(\mathbf{b}^{*})$ +B) Pseudo-mask annotations + +![](images/e9d24e5a58885e18be4417c6cb919917e1f8c8dd29da202846eee89a23983713.jpg) + +![](images/1da99ed894fdfd008face366861bcdac4ad6e7516fa7463994dddbb9c758d9fa.jpg) +Figure 5. A) Pseduo bounding box selection guided by GradCAM activation. B) Visualization of pseudo-mask generated for Open Images. +(a) Object detection +Figure 6. Ablation study for object detection and instance segmentation by increase number of masking iteration during PL generation. As iteration increases, GradCAM activation becomes noisy resulting in poor guidance and pseudo-mask generation. + +![](images/5b9f6f8885241010f0532ac53d6586c4c0265a6882a6aa591f1f1aec46ea1802.jpg) +(b) Instance segmentation + +towards objects producing better quality proposals. + +Iterative Masking Steps vs Performance. Fig. 6 presents ablation analysis for iterative masking hyper-parametr $G$ . We see that for initial few masking steps the performance increase. Because, without any masking the GradCAM activation map is towards most discriminative parts of an object. As we mask out these regions, the less discriminative regions gets activated and combining activation from previous steps produce a strong guidance function $\Phi_t$ . However, performing too many mask iterations could completely mask out the object of interest and unrelated background regions might get activated resulting in a poor $\Phi_t$ . + +Weakly-supervised Instance Segmentation. Weakly-supervised Instance Segmentation (WSIS) methods performs instance-segmentation using supervision from image-labels. Similar to WSIS setting, our method performs instance-segmentation using just image-labels. Following this, we simply extended our method to WSIS setting by comparing with VOC benchmark. PENet [12] is a recent WSIS method, which relies on multiple downstream task network such as segmentation, detection and multi-class classification to perfrom segmentation. CL [18] and IAM [57] are WSIS methods utilize advanced attention mechanism to generate instance segmentation mask. As shown in Table 6, our method outperforms existing WSIS methods by considerable margin. Thus, the proposed method is simple and can be easily extended to OV setting or WSIS setting producing SOTA performance. Note that, other OV methods cannot be extended to WSIS setting as they rely on base + +Table 6. Instance segmentation(mAP) performance for VOC in weakly-supervised instance segmentation setting. + +
MethodSupervisionmAP
Mask-RCNN [15]F67.9
OCIS [8]I26.8
PRM [55]I28.3
IAM [57]I30.2
CL [18]I+C30.2
PENet [12]I38.1
OursI38.6
+ +annotation and similarly WSIS methods cannot be extended to OV setting due to constrained vocabulary space. + +Qualitative Analysis. Fig. 5 (A) presents a visualization of activation maps generated for the object of interest (woman and dog). As we can see, the generated activation map covers the entire object and it can be used as a guidance function to choose the best bounding box proposal. Note that the activation are square-shaped because the original activation map is $1/16$ th of the image size. We perform nearest interpolation to obtain an activation map of image size. In Fig 5 (B), we visualize the pseudo-mask for Open Images generated from our pipeline. We can observe that the generated pseudo-mask is of good quality; still contains some false positives. However, with Mask-RCNN training, the model learns to filter the noise present in the pseudo-mask producing better-quality predictions. + +# 5. Conclusion + +We propose a novel pipeline called Mask-free OVIS for open-vocabulary instance segmentation, which eliminates the need for costly and time-consuming manual mask annotations. We achieve this by generating pseudo-mask annotation for base and novel categories by leveraging a pretrained vision-language model and weakly supervised techniques. The generated pseudo-mask annotations are then used to supervise the Mask-RCNN model. Extensive experiments are conducted to show the effectiveness of our method on MS-COCO and Open Images datasets. Instead of relying on labour-expensive instance-level annotations, we hope our simple and effective pipeline provides an alternative to detect and segment long-tailed novel categories. + +# References + +[1] Aditya Arun, CV Jawahar, and M Pawan Kumar. Dissimilarity coefficient based weakly supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9432-9441, 2019. 3 +[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 384-400, 2018. 6, 7 +[3] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In European conference on computer vision, pages 549-565. Springer, 2016. 1 +[4] David Biertimpel, Sindi Shkodrani, Anil S Baslamisli, and Nora Baka. Prior to segment: Foreground cues for weakly annotated classes in partially supervised instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2824-2833, 2021. 6 +[5] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2846-2854, 2016. 4, 6 +[6] Maria A Bravo, Sudhanshu Mittal, and Thomas Brox. Localized vision-language matching for open-vocabulary object detection. In Pattern Recognition: 44th DAGM German Conference, DAGM GCR 2022, Konstanz, Germany, September 27–30, 2022, Proceedings, pages 393–408. Springer, 2022. 3 +[7] Liang-Chieh Chen, Alexander Hermans, George Papandreou, Florian Schroff, Peng Wang, and Hartwig Adam. Masklab: Instance segmentation by refining object detection with semantic and direction features. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4013-4022, 2018. 1 +[8] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12405, 2019. 8 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 +[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 3 +[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 266-282. Springer, 2022. 1, 3, 6, 7 +[12] Weifeng Ge, Sheng Guo, Weilin Huang, and Matthew R Scott. Label-penet: Sequential label propagation and enhancement networks for weakly supervised instance segmen- + +tation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3345-3354, 2019. 8 +[13] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. CyCLIP: Cyclic contrastive language-image pretraining. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3 +[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 1, 2, 3, 6 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 5, 8 +[16] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Advances in neural information processing systems, 33:16797-16807, 2020. 3 +[17] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2, 3, 5, 6, 7 +[18] Jaedong Hwang, Seohyun Kim, Jeany Son, and Bohyung Han. Weakly supervised instance segmentation by deep community learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1020-1029, 2021. 8 +[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2, 3 +[20] Wonjik Kim, Asako Kanezaki, and Masayuki Tanaka. Unsupervised learning of image segmentation based on differentiable feature clustering. IEEE Transactions on Image Processing, 29:8055-8068, 2020. 2 +[21] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 1, 5 +[22] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 3 +[23] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 3 +[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for uni + +fied vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3 +[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3, 5 +[26] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell me where to look: Guided attention inference network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 3 +[27] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3 +[28] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5 +[30] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In International Conference on Learning Representations, 2021. 6, 7 +[31] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3 +[32] Zongyang Ma, Guan Luo, Jin Gao, Liang Li, Yuxin Chen, Shaoru Wang, Congxuan Zhang, and Weiming Hu. Open-vocabulary one-stage detection with hierarchical visual-language knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14074-14083, 2022. 3 +[33] Niall O'Mahony, Sean Campbell, Anderson Carvalho, Suman Harapanahalli, Gustavo Velasco Hernandez, Lenka Krpalkova, Daniel Riordan, and Joseph Walsh. Deep learning vs. traditional computer vision. In Science and information conference, pages 128-144. Springer, 2019. 1 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5 +[35] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11932–11939, 2020. 6 + +[36] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In 36th Conference on Neural Information Processing Systems (NIPS), 2022. 3 +[37] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2, 4 +[38] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, MingYu Liu, Yong Jae Lee, Alexander G Schwing, and Jan Kautz. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598-10607, 2020. 4 +[39] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 3, 4 +[40] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 3 +[41] Limin Wang, Yuanjun Xiong, Dahua Lin, and Luc Van Gool. Untrimmednets for weakly supervised action recognition and detection. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4325-4334, 2017. 3 +[42] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In European Conference on Computer Vision, pages 649-665. Springer, 2020. 1 +[43] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 6, 7 +[44] Ke Yang, Dongsheng Li, and Yong Dou. Towards precise end-to-end weakly supervised object detection network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8372-8381, 2019. 3 +[45] Keren Ye, Mingda Zhang, Adriana Kovashka, Wei Li, Danfeng Qin, and Jesse Berent. Cap2det: Learning to amplify weak caption supervision for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9686-9695, 2019. 6 +[46] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 1, 2, 3, 5, 6, 7 +[47] Zhaoyang Zeng, Bei Liu, Jianlong Fu, Hongyang Chao, and Lei Zhang. Wsod2: Learning bottom-up and top-down objectness distillation for weakly-supervised object detection. + +In Proceedings of the IEEE/CVF international conference on computer vision, pages 8292-8300, 2019. 3 +[48] Dingwen Zhang, Junwei Han, Gong Cheng, and Ming-Hsuan Yang. Weakly supervised object localization and detection: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5866-5885, 2021. 3 +[49] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3 +[50] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 159-175. Springer, 2022. 2, 3 +[51] Ye Zheng, Jiahong Wu, Yongqiang Qin, Faen Zhang, and Li Cui. Zero-shot instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2593-2602, 2021. 6, 7 +[52] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 1, 2, 3, 5, 6 +[53] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3 +[54] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 6 +[55] Yanzhao Zhou, Yi Zhu, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Weakly supervised instance segmentation using class peak response. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3791-3800, 2018. 8 +[56] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't even look once: Synthesizing features for zero-shot detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11693-11702, 2020. 6 +[57] Yi Zhu, Yanzhao Zhou, Huijuan Xu, Qixiang Ye, David Doermann, and Jianbin Jiao. Learning instance activation maps for weakly supervised instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3116-3125, 2019. 8 \ No newline at end of file diff --git a/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/images.zip b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c9fb24bb6fe0c5387917fdc9c18cbf1db40f245 --- /dev/null +++ b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33302f5d6e218aa2ccc6580229eb5747b2652d7d067dff6a167e51944b6c98fc +size 686073 diff --git a/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/layout.json b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d43be82d46a7c778132ed9becc3a1b55a1c6590c --- /dev/null +++ b/2023/Mask-Free OVIS_ Open-Vocabulary Instance Segmentation Without Manual Mask Annotations/layout.json @@ -0,0 +1,8516 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 478, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 478, + 138 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 478, + 138 + ], + "type": "text", + "content": "Mask-free OVIS: Open-Vocabulary Instance Segmentation without Manual Mask Annotations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 160, + 451, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 160, + 451, + 189 + ], + "spans": [ + { + "bbox": [ + 142, + 160, + 451, + 189 + ], + "type": "text", + "content": "Vibashan VS\\*, Ning Yu†, Chen Xing†, Can Qin‡, Mingfei Gao†, Juan Carlos Niebles†, Vishal M. Patel\\*, Ran Xu†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 190, + 477, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 190, + 477, + 204 + ], + "spans": [ + { + "bbox": [ + 116, + 190, + 477, + 204 + ], + "type": "text", + "content": "*Johns Hopkins University, ‡Northeastern University, †Salesforce Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 205, + 443, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 205, + 443, + 217 + ], + "spans": [ + { + "bbox": [ + 153, + 205, + 443, + 217 + ], + "type": "text", + "content": "{vvishnu2, vpatel36}@jhu.edu, qin.ca@northeastern.edu," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 166, + 220, + 431, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 220, + 431, + 232 + ], + "spans": [ + { + "bbox": [ + 166, + 220, + 431, + 232 + ], + "type": "text", + "content": "{ning.yu, cxing, jniebles, ran.xu}@salesforce.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 274, + 290, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 290, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 290, + 634 + ], + "type": "text", + "content": "Existing instance segmentation models learn task-specific information using manual mask annotations from base (training) categories. These mask annotations require tremendous human effort, limiting the scalability to annotate novel (new) categories. To alleviate this problem, Open-Vocabulary (OV) methods leverage large-scale image-caption pairs and vision-language models to learn novel categories. In summary, an OV method learns task-specific information using strong supervision from base annotations and novel category information using weak supervision from image-captions pairs. This difference between strong and weak supervision leads to overfitting on base categories, resulting in poor generalization towards novel categories. In this work, we overcome this issue by learning both base and novel categories from pseudomask annotations generated by the vision-language model in a weakly supervised manner using our proposed Mask-free OVIS pipeline. Our method automatically generates pseudo-mask annotations by leveraging the localization ability of a pre-trained vision-language model for objects present in image-caption pairs. The generated pseudomask annotations are then used to supervise an instance segmentation model, freeing the entire pipeline from any labour-expensive instance-level annotations and overfitting. Our extensive experiments show that our method trained with just pseudo-masks significantly improves the mAP scores on the MS-COCO dataset and OpenImages dataset compared to the recent state-of-the-art methods trained with manual masks. Codes and models are provided in https://vibashan.github.io/ovis-web/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 641, + 128, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 128, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 128, + 653 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "type": "text", + "content": "Instance segmentation is a challenging task as it requires models to detect objects in an image while also precisely" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 258, + 545, + 338 + ], + "blocks": [ + { + "bbox": [ + 309, + 258, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 309, + 258, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 309, + 258, + 545, + 338 + ], + "type": "image", + "image_path": "a49cee1ae68a616cd58806ea83eadc22c8d47a685d2fb4326a991229a3386be0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 339, + 545, + 409 + ], + "blocks": [ + { + "bbox": [ + 308, + 339, + 545, + 409 + ], + "lines": [ + { + "bbox": [ + 308, + 339, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 545, + 409 + ], + "type": "image", + "image_path": "919bacb6b2095f72cc721d1ac5f5ca9082c6a24b847e596aab35e6a4e33243ac.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 414, + 547, + 503 + ], + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 503 + ], + "type": "text", + "content": "Figure 1. A) Previous Methods: Learn task-specific information (detection/segmentation) in a fully-supervised manner and novel category information with weak supervision. During training, this difference in strong and weak supervision signals leads to overfitting and requires expensive base annotations. B) Our method: Given image-caption pairs, we generate pseudo-annotations for both base and novel categories under weak supervision, solving the problems of labour-expensive annotation and overfitting." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 507, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 687 + ], + "type": "text", + "content": "segment each object at the pixel-level. Although the rise of deep neural networks has significantly boosted the state-of-the-art instance segmentation performance [7, 15, 42], these methods are still trained for a pre-defined set of object categories and are data-hungry [33]. Particularly, one needs to manually annotate thousands of instance-level masks for each object category, which takes around 78 seconds per instance mask [3]. If we look at this quantitatively on Open Images [21], a large-scale dataset with 2.1M instance-level mask annotations requires around 5 years of human labour. Even after extensive annotation, these training datasets are still limited to a small number of categories and segmenting objects from a novel category requires further annotation. Therefore, it is difficult to scale up existing methods to segment a large number of categories due to intensive labour." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Recently, Open-Vocabulary (OV) methods have gained much attention due to their success in detecting [11, 14," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "*This work was done when Vibashan VS interned at Salesforce Research. Primary contact: vvishnu2@jhu.com" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23539" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 355, + 242 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 355, + 242 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 355, + 242 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 355, + 242 + ], + "type": "image", + "image_path": "58b34a10e529132414d0d16d8145a05ece8d697b3e81726db38591b9fcc86259.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 244, + 546, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 546, + 278 + ], + "type": "text", + "content": "Figure 2. RPN is supervised using bounding box annotations from COCO base and WSPN is supervised using image-labels from COCO base. A) WSPN produces better quality proposals for novel object categories compared to fully-supervised RPN. B) WSPN consistently produces better recall for all COCO novel categories than RPN." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 357, + 76, + 543, + 242 + ], + "blocks": [ + { + "bbox": [ + 357, + 76, + 543, + 242 + ], + "lines": [ + { + "bbox": [ + 357, + 76, + 543, + 242 + ], + "spans": [ + { + "bbox": [ + 357, + 76, + 543, + 242 + ], + "type": "image", + "image_path": "cbc981d08377389e6e16d676f69b6fdb3338f429e6c728d6041248ef22bb903b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 281, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 461 + ], + "type": "text", + "content": "46, 52] and segmenting [17] novel categories beyond base (training) categories. An OV method learns task-specific information (detection/segmentation) from base categories that have manual instance-level bounding boxes or masks and learns novel category information from the pre-trained Vision-Language Model (VLM) [19, 34] (see Fig. 1). All these methods produce promising results on novel categories by leveraging different forms of weak supervision such as caption pretaining [46, 50], knowledge distillation [14, 52] and pseudo-labelling [17, 28]. However, all above-mentioned OV methods still rely on the manually-annotated base categories to improve their performances on novel categories. Without fine-tuning on base categories, existing OV methods lack task/domain specific knowledge and the performances on novel categories will be affected [14, 46]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "content": "Although manual instance-level annotations of base categories are critical to open-vocabulary segmentation methods, we find empirically that such fully-supervised information causes OV methods to overfit to base categories, leading to a higher failure rate when evaluated on novel categories. Specifically, OV methods utilize a region proposal network (RPN) [37] supervised with bounding box annotations obtained from the base categories to generate a set of bounding box proposals for all objects categories in a given image [46]. The feature representation from these object proposals is later matched with text embedding to learn the visual-semantic space for base and novel categories [46]. Therefore, the quality of proposals generated for novel object categories plays a key role in determining the performance in later stages. However, from our experiments, we find that many objects of novel categories wouldn't be included in such proposals due to the RPN's overfitting to base categories. Fig. 2 (A) - Top gives some examples where the RPN trained with COCO base categories fails to generate high-quality region proposals for novel categories such as elephant, cat and skateboard. Therefore, a fully" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 305 + ], + "type": "text", + "content": "supervised proposal network is a bottleneck in OV pipeline due to its poor generalization towards novel categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 306, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 546, + 510 + ], + "type": "text", + "content": "Given the aforementioned observations of poor generalization, we raise the question of whether we can improve the generalization by using weak supervision instead of relying on strong supervision from manual base annotations. If so, we can reduce overfitting to the base categories and the requirement for costly instance-level human annotations can be entirely removed from the pipeline. Our preliminary experiments give us some hope. Our experiments show that if we train a weakly-supervised proposal network (WSPN) with image-level annotations instead of box-level annotations, the region proposals it generates can better generalize to novel objects. As shown in Fig. 2 A), the novel objects that the RPN proposals miss are covered by WSPN. Fig. 2 B) shows WSPN proposals have consistently better average recall than RPN for all COCO novel categories, indicating that WSPN proposals are more likely to cover the ground-truth bounding boxes of novel objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 511, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 547, + 713 + ], + "type": "text", + "content": "Inspired by these observations, we propose open-vocabulary segmentation without manual mask annotations. We do not use any human-provided box-level or pixel-level annotations during the training of our method. We first train a simple WSPN model with image-level annotations on base categories as a proposal generator to generate proposals for all objects given an image. Then, we adopt pre-trained vision-language models to select proposals as pseudo bounding boxes for novel objects. Given a novel object's text name, we can utilize the name as a text prompt to localize this object in an image with a pre-trained vision-language model. To obtain a more accurate pseudo-mask that covers the entire object, we conduct iterative masking with GradCAM [39] given the vision-language model. Finally, we train a weakly-supervised segmentation (WSS) [20] network with previously generated bounding box and GradCAM activation map to obtain pixel-level annotation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23540" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": "Our contributions are summarized as follows: (1) We propose a Mask-free OVIS pipeline where we produce manual-effort-free pseudo-mask annotations for base and novel instance segmentation using open-vocabulary and weakly supervised techniques. (2) We propose a novel pseudo-mask generation pipeline leveraging a pre-trained vision-language model to generate instance-level annotations. (3) Benefiting from pseudo-labels, our method sets up SOTA's for both detection and instance segmentation tasks compared to recent methods trained with manual masks on MS-COCO and OpenImages datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 215, + 134, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 134, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 134, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 235, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 287, + 354 + ], + "type": "text", + "content": "Weakly-supervised learning (WSL) methods require less manual annotation effort and only rely on partial information about the task compared to fully-supervised methods. WSL applications include object detection [1, 16, 44, 47], segmentation [22,23,40], action recognition, etc [26,41,48]. The class activation map (CAM) [53] and grad-CAM [39] are popular approaches which exploits pre-trained network to generate saliency or activation map for target classes. Li et al. [26] enables weakly-supervised semantic segmentation via CAM-guided attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 355, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 289, + 545 + ], + "type": "text", + "content": "Vision-Language Models. With the rise of large-scale text-pretraining with attention-based models [9, 49], vision-language models (VLM) have caught increasing attention due to their strong performance in downstream visual understanding tasks [27, 31, 34]. Early VLM models [27, 31] applied the pre-trained models with constraints of the class set. With the help of contrastive learning, CLIP [34] has enabled VLM in the wild with large-scale multimodal learning upon the 400M noisy data crawled from the web. Several methods have extended CLIP for high efficiency model training and cycle consistency [13, 24, 25]. ALBEF [25] utilizes cross-modal attention between image patches and text, leading to more grounded vision and language representation learning. Compared to CLIP, ALBEF excels in object localization, while CLIP performs better at zero-shot classification due to its larger training dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": "Open-Vocabulary. Open-Vocabulary methods [6, 10, 32, 36, 46, 50] scale up their vocabulary size for object detection or instance segmentation tasks by transferring knowledge from pre-trained vision-language models. Zareian et al. [46] proposed an open-vocabulary object detection method where a visual encoder is trained on image-caption pair to model object semantics and then transfer to zero-shot object detection. ViLD [14] and RegionCLIP [52] use pre-trained CLIP [34] to distill knowledge and enhance the vision-text embedding space effectively. Similarly, Huynh et al. [17] proposes a robust pseudo-labelling, for instance segmentation where a teacher model generates pseudo-label for novel categories and is later used to supervise a student model. Therefore, all these methods require manually" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "annotated base categories to improve their performance on novel categories. In contrast, our method generates pseudolabels for both base and novel categories by leveraging a pre-trained VLM, freeing the entire pipeline from human-provided instance-level annotations. Similar to our method, Gao et al. [11] generates pseudo-labels using pre-trained VLM, but their approach still uses a proposal generator that relies on human-provided bounding boxes for the base categories. Our method, on the other hand, requires no such annotations, making it completely free from any human-provided box-level or pixel-level annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 209, + 406, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 406, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 406, + 221 + ], + "type": "text", + "content": "3. Mask-free OVIS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 224, + 547, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 308 + ], + "type": "text", + "content": "Our pipeline consists of two stages: (I) Pseudo-mask generation, and (II) Open-vocabulary instance segmentation. The goal of (I) is to generate box-level and pixel-level annotations by leveraging the region-text alignment property of a pre-trained vision-language model as illustrated in Fig. 3. In (II), we train a Mask-RCNN using the generated pseudo annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 310, + 448, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 448, + 321 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 448, + 321 + ], + "type": "text", + "content": "3.1. Pseudo-Mask Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": "Region-Text Attention Scores. The inputs to the vision-language model are image " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " and caption " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{c_1, c_2, \\dots, c_{N_c}\\}" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " pair, where " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " is the number of words in the caption (including [CLS] and [SEP]) [9, 25]. In a VLM, a text encoder is utilized to get text representations " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{N_C \\times d}" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " and an image encoder is utilized to extract region representation " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{N_R \\times d}" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "N_R" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " is the number of regions in the image. To fuse the information from both image and text encoders, a multi-modal encoder with " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " consecutive cross-attention layers is utilized [19, 25]. These cross-attention layers learn a good region-text alignment [25, 34] and are utilized to obtain regions corresponding to the object of interest " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " from the caption. In particular, for the " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": "-th cross-attention layer, the visual region attention scores " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^m" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " for the object of interest " + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 304, + 327, + 547, + 494 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 370, + 500, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 500, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 370, + 500, + 545, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {t} ^ {m} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {h} _ {t} ^ {m - 1} \\mathbf {R} ^ {T}}{\\sqrt {d}}\\right), \\tag {1}", + "image_path": "2e10de5a5f4d80d82c5f01a3cceb50af53ffba2355a9f1072c939038088f7df6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 375, + 526, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 526, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 375, + 526, + 545, + 540 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {t} ^ {n} = \\mathbf {X} _ {t} ^ {m} \\cdot \\mathbf {R}. \\tag {2}", + "image_path": "c57151798d2df5a1698fe55d405a503af85b7efce1df780c11f3d0ca6b501de0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "content": " is a scalar and " + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t^{m-1}" + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "content": " is the hidden representation obtained from the previous " + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "inline_equation", + "content": "(m-1)" + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "content": "-th cross-attention layer. GradCAM Activation Map. After obtaining attention scores " + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^m" + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "content": ", we employ Grad-CAM [39] to visualize the activated regions. As in [39], we take the image-caption similarity " + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "content": " output from the multi-modal encoder's final layer and calculate the gradient with respect to the attention scores. The activation map " + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "content": " for object " + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 304, + 548, + 545, + 639 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 646, + 545, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 646, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 368, + 646, + 545, + 671 + ], + "type": "interline_equation", + "content": "\\phi_ {t} = \\mathbf {X} _ {t} ^ {m} \\cdot \\max \\left(\\frac {\\partial S}{\\partial \\mathbf {X} _ {t} ^ {m}}, 0\\right). \\tag {3}", + "image_path": "7cddc4230a205f5204b41e81c7a96afb2c02a63f3b6b810ecb88a78d3fdc86a7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "Iterative Masking. During VLM training, an object's most discriminative regions easily get aligned towards object text representation [25, 34]. As a result, " + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": " is localized towards" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23541" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 512, + 311 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 512, + 311 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 512, + 311 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 512, + 311 + ], + "type": "image", + "image_path": "e6e0885f1a8316d49654e267400d7cc5c755ea0329656cd230f818899ba46f98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 548, + 364 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 548, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 548, + 364 + ], + "type": "text", + "content": "Figure 3. Illustrative overview of our pseudo-mask generation pipeline. Given an image-caption pair and pre-trained VLM, we generate an activation map for the object of interest (\"umbrella\") and enhance it using iterative masking strategy. We generate box-level annotations using an activation map as a guidance function to select the best WSPN proposals covering the object. We crop the image corresponding to the generated pseduo bounding box and perform weakly-supervised segmentation to obtain pixel-level annotations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 289, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 464 + ], + "type": "text", + "content": "the most discriminative region and fails to cover the object completely [39]. However, when we mask out the most discriminative regions, GradCAM activations are shifted towards other discriminative regions (see Fig. 4). We propose a simple iterative masking strategy to obtain better activation where the most activated part of the object is replaced with image mean and the new activation map is computed following Eq. 1 and 3. The final activation map is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 465, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 465, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 127, + 465, + 287, + 496 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Phi} _ {t} = \\bigcup_ {i = 1} ^ {G} \\mathcal {I M} \\left(\\phi_ {t} ^ {i}\\right) \\tag {4}", + "image_path": "ca5b5294045da042f83120b1243799939886c4607c48cd460338748361a6a5eb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "text", + "content": " is a hyper-parameter indicating the number of masking iterations and " + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{IM}(\\cdot)" + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "text", + "content": " normalize and threshold " + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "text", + "content": " by 0.5. We utilize the activation map " + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 46, + 496, + 288, + 544 + ], + "type": "text", + "content": " as a guidance function to generate box-level and pixel-level annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": "Weakly-Supervised Proposal Network. To generate box-level annotations, we require bounding box proposals covering the activated region " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": ". As explained in the introduction, the fully-supervised RPN network produces poor proposals for novel categories, making it less preferable. To overcome this, we propose a weakly-supervised proposal network (WSPN), which generates object proposals trained in a weakly-supervised manner. Given an Image I, the WSPN is supervised with image-labels " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{y_1, y_2, \\dots, y_C\\}" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "y_c = 0" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": " or 1 indicates the absence or presence of class " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": " denotes total number of classes [5]. First, we utilize Selective Search to generate a set of unsupervised proposal " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = \\{u_1, u_2, \\dots, u_N\\}" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": " is the total number of bounding box proposals. Then, the Image I and proposal " + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 46, + 548, + 289, + 714 + ], + "type": "text", + "content": " are fed into a CNN backbone" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 367, + 545, + 496 + ], + "blocks": [ + { + "bbox": [ + 309, + 367, + 545, + 496 + ], + "lines": [ + { + "bbox": [ + 309, + 367, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 309, + 367, + 545, + 496 + ], + "type": "image", + "image_path": "8416f981e3aff51cf2fa537af902c1472958da128a51fa9440c26527c5e37739.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "text", + "content": "Figure 4. Comparison between activation map " + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "text", + "content": " generated at " + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "text", + "content": " 0 and 1. For " + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "inline_equation", + "content": "G = 0" + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "text", + "content": ", the most discriminative parts of an object (bird's head) gets activated (bird's head). After masking, for " + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "inline_equation", + "content": "G = 1" + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "text", + "content": " we can observe that the activation map has shifted to less discriminative part (bird's body). Thus, by combining activation from both steps, we obtain a better activation map trying to cover entire object " + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 304, + 498, + 547, + 564 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": "to extract features and RoI pooling layer [37] to obtain RoI pooled feature vectors. Following [5], the pooled feature vectors are passed to a classification and detection branch to generate two matrices " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{cls}" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{det} \\in \\mathbb{R}^{C \\times N}" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{cls}" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{det}" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": " matrices are normalized along the category direction (column-wise) and proposal direction (row-wise) by the softmax layers " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": " respectively. From " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{cls}" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{det}" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": ", the instance-level classification scores for object proposals are computed by the element-wise product " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{C}} = \\sigma(\\mathbf{W}^{cls}) \\odot \\sigma(\\mathbf{W}^{det})" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": ". and image-level classification score for the " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "c_{th}" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": " class is computed as " + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "inline_equation", + "content": "p_c = \\sum_{i=1}^N w_{i,c}" + }, + { + "bbox": [ + 304, + 578, + 547, + 715 + ], + "type": "text", + "content": ". Following [38], we select the high-scoring proposals are used as" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23542" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 287, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 287, + 118 + ], + "type": "text", + "content": "pseudo-regression targets (" + }, + { + "bbox": [ + 46, + 71, + 287, + 118 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}} = \\{\\hat{t}(u_1), \\hat{t}(u_2), \\dots, \\hat{t}(u_N)\\}" + }, + { + "bbox": [ + 46, + 71, + 287, + 118 + ], + "type": "text", + "content": ") for low-scoring proposals to make sure objects can be more tightly captured. To this end, the classification loss and regression loss for WSPN are calculated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 125, + 287, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 125, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 71, + 125, + 287, + 185 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {w s p n} = - \\sum_ {c = 1} ^ {C} y _ {c} \\log p _ {c} + (1 - y _ {c}) \\log \\left(1 - p _ {c}\\right) + \\tag {5} \\\\ \\frac {1}{N} \\sum_ {u = 1} ^ {N} \\mathcal {L} _ {\\text {s m o o t h L 1}} (\\hat {t} (u _ {i}), u _ {i}). \\\\ \\end{array}", + "image_path": "46ff412a29ef2a844281cc7f3b1ae9aac6e62a766100c2408390bb4a2fc0b7b2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": "WSPN is trained to localize and classify objects by minimizing Eq. 5. The trained WSPN model is used to generate object proposals and top " + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": " proposal candidates over all the classes " + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathbf{b}_1, \\mathbf{b}_2, \\dots, \\mathbf{b}_K\\}" + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": " are selected by sorting proposal confidence scores obtained from " + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{det}" + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": ". Finally, from the top proposal candidates " + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": ", we select the proposal which overlaps the most with " + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": " as pseudo box-bounding:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 275, + 287, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 275, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 117, + 275, + 287, + 301 + ], + "type": "interline_equation", + "content": "\\mathbf {b} ^ {*} = \\underset {\\mathbf {b} \\in \\mathcal {B}} {\\arg \\max } \\frac {\\sum_ {\\mathbf {b}} \\Phi_ {t}}{\\sqrt {| \\mathbf {b} |}}, \\tag {6}", + "image_path": "6c6e22dc4546ccb6b5f6f0f0808f132db82d397cc14b36decf6fcaa9638a5d67.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "text", + "content": " is the pseudo box-level annotation and " + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\sum_{\\mathbf{b}}\\Phi_t" + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "text", + "content": " indicates the summation of the activation map values within a box proposal " + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "inline_equation", + "content": "|\\mathbf{b}|" + }, + { + "bbox": [ + 46, + 301, + 287, + 336 + ], + "type": "text", + "content": " indicates the proposal area." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": "Weakly-Supervised Segmentation. Once we obtain the pseudo bounding box " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": ", we crop image " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " to obtain the corresponding image patch. The cropped patch is then fed into a simple three-layer CNN network to perform pixel-level segmentation. To supervise the CNN network, we generate pseudo ground-truth " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": ", where we sample " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " points as foreground " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "F_z = \\{f_i\\}_{i=1,\\dots,Z}" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " and background " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "B_z = \\{b_i\\}_{i=1,\\dots,Z}" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " and each point is set to 1 or 0, respectively. Specifically, the foreground and background points are sampled from the most and least activated part of " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " inside " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": ". The pseudo ground-truth " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " is of size " + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 46, + 337, + 287, + 503 + ], + "type": "text", + "content": " and we supervise the network predictions only at sampled points. Thus, the segmentation loss obtained from these weak points is computed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 504, + 287, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 545 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w s s} = \\sum_ {i = 1} ^ {G} \\mathcal {L} _ {c e} \\left(\\mathbf {s} ^ {*} \\left(f _ {i}\\right), \\boldsymbol {\\Theta} \\left(f _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {G} \\mathcal {L} _ {c e} \\left(\\mathbf {s} ^ {*} \\left(b _ {i}\\right), \\boldsymbol {\\Theta} \\left(b _ {i}\\right)\\right), \\tag {7}", + "image_path": "c0218aaf223f06c4ebf4b6879158af0e83ba1fe483bdfaa04592f6c92637a748.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^*" + }, + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "type": "text", + "content": " is the pseudo pixel-level annotation of size " + }, + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 47, + 546, + 287, + 570 + ], + "type": "text", + "content": " indicates cross-entropy loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": "To this end, given an image " + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": " and caption " + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": " pair, we generate pseudo box-level " + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": " and pixel-level " + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^*" + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": " annotation for the object of interest " + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": ". In practice, the pseduo-mask annotations are generated for a pre-defined set of object categories obtained from training vocabulary. Fig. 5 visualizes the generated pseduo-mask annotations for various categories. Activated regions correspond well with objects of interest and the generated pseduo-mask are of good quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 673, + 265, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 265, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 265, + 686 + ], + "type": "text", + "content": "3.2. Open-Vocabulary Instance Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 690, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 690, + 287, + 714 + ], + "type": "text", + "content": "After generating pseudo-mask annotations, we train an open-vocabulary instance segmentation model. Following" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": "[46], we employ a Mask-RCNN as the instance segmentation model, where a class-agnostic mask head is utilized to segment objects and the classification head is replaced with embedding head " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "h_{emb}" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": ". Given a Image I, an encoder network extracts image features and region embeddings, " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = \\{\\mathbf{r}_i\\}_{i=1,\\dots,N_r}" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": ", are obtained by RoI align [15] followed by a fully connected layer, where " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": " denotes the number of regions. The similarity between the region and text embedding pair is calculated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 180, + 545, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 180, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 312, + 180, + 545, + 213 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {r} _ {i}, \\mathbf {c} _ {j}\\right) = \\frac {\\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {c} _ {j}\\right)}{\\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {b g}\\right) + \\sum_ {k} \\exp \\left(h _ {e m b} \\left(\\mathbf {r} _ {i}\\right) \\cdot \\mathbf {c} _ {k}\\right)}, \\tag {8}", + "image_path": "e634a1ca42643b639eb5080d654380e38f5adcbd7eafa8c94696015308b0de18.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{\\mathbf{bg},\\mathbf{c}_1,\\mathbf{c}_2,\\dots,\\mathbf{c}_{N_C}\\}" + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "text", + "content": ", are object vocabulary text representation obtained from pre-trained text encoder, where " + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "inline_equation", + "content": "N_{C}" + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "text", + "content": " is the training object vocabulary size. To learn the semantic space, negative pairs are pushed away and positive pairs are pulled together using cross entropy loss obtained from " + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "text", + "content": ". The class-agnostic mask head is supervised by minimizing standard segmentation loss [15] obtained from " + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^*" + }, + { + "bbox": [ + 304, + 213, + 545, + 345 + ], + "type": "text", + "content": ". During inference, the similarity between the region proposals embedding and text embedding from a group of object classes of interest is calculated. The region is then assigned to a class with the highest similarity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 357, + 449, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 449, + 370 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 449, + 370 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 377, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 533 + ], + "type": "text", + "content": "Datasets. Following [46], we conduct experiments on MS-COCO [29] with data split of 48 base categories and 17 novel categories. The processed COCO dataset contains 107,761 training images and 4,836 test images. Following [17], we conduct experiments on Open Images [21] to verify the effectiveness of our method on the large-scale dataset. The Open Images dataset consists of 300 categories with a class split of 200 base categories (frequent objects) and 100 novel categories (rare objects). Following [34], we leverage image-labels obtained from MS-COCO and Open Images to learn the novel category information. We also experiment using image-caption datasets to show our method's effectiveness irrespective of training vocabulary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": "Evaluation Metrics and Protocols. Following open-vocabulary methods [17, 52], for both detection and segmentation tasks, we report the mean Average Precision at intersection-over-union (IoU) of " + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "0.5\\left(mAP_{50}\\right)" + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": ". Following zero-shot settings [46], we report novel category performance for both constrained setting and generalized setting. In constrained setting, the model is evaluated only on novel class test images and in generalized setting, the model is evaluated on both base and novel class test images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. In pseudo-mask generation framework, we use pre-trained ALBEF [25] as our vision-language model. We conducted all our pseudo-mask generation experiments using ALBEF due to the good region-text alignment when image and caption pair are present [25]. Following ALBEF, the cross-attention layer " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " used for" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23543" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 93, + 507, + 268 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Object Detection (mAP) performances for MS-COCO under constrained and generalized setting. " + }, + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{B}}" + }, + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": " are subset of " + }, + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\Omega}" + }, + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\Omega}" + }, + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": " contains training vocabulary larger than COCO categories." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 93, + 507, + 268 + ], + "lines": [ + { + "bbox": [ + 82, + 93, + 507, + 268 + ], + "spans": [ + { + "bbox": [ + 82, + 93, + 507, + 268 + ], + "type": "table", + "html": "
MethodProposal GeneratorLanguage SupervisionBase AnnotationConstrained NovelGeneralized Novel
WSDDN [5]-Image-labels in C_B ∪ C_NX-19.7
Cap2Det [45]-Image-labels in C_B ∪ C_NX-20.3
SB [2]RPN COCObase-✓0.700.31
DELO [56]RPN COCObase-✓7.603.41
PL [35]RPN COCObase-✓10.04.12
OV-RCNN [46]RPN COCObaseImage-caption in C_B ∪ C_N✓27.522.8
CLIP-RPN [14]RPN COCObaseCLIP image-text pair C_Ω✓-26.3
ViLD [14]RPN COCObaseCLIP image-text pair C_Ω✓-27.6
Detic [54]RPN COCObaseImage-caption in C_B ∪ C_N✓-27.8
RegionCLIP [52]RPN LVISbaseConceptual caption C_Ω✓30.826.8
PB-OVD [11]RCNN COCObaseImage-caption in C_B ∪ C_N✓32.330.7
XPM [17]RPN COCObaseImage-caption in C_B ∪ C_N✓29.927.0
Mask-free OVIS (Ours)WSPN COCObaseImage-labels in C_B ∪ C_NX31.527.4
Mask-free OVIS (Ours)WSPN COCObaseImage-labels in C_B ∪ C_N✓35.931.5
", + "image_path": "3141744ff0c2c7c223599490997db2c6c2e9a7c053810e75bd26ef8ae26212d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 285, + 547, + 408 + ], + "blocks": [ + { + "bbox": [ + 62, + 274, + 529, + 285 + ], + "lines": [ + { + "bbox": [ + 62, + 274, + 529, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 529, + 285 + ], + "type": "text", + "content": "Table 2. Instance Segmentation (mAP) performances for MS-COCO and Open Images under constrained and generalized setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 285, + 547, + 408 + ], + "lines": [ + { + "bbox": [ + 48, + 285, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 547, + 408 + ], + "type": "table", + "html": "
MethodProposal Generator (MS-COCO/OpenImages)Base AnnotationMS-COCOOpen Images
Constrained NovelGeneralized NovelConstrained NovelGeneralized Novel
OVR+OMP [4]-✓14.18.324.916.8
SB [2]-✓20.816.024.817.3
BA-RPN [51]-✓20.115.425.316.9
Soft-Teacher [43]RPN COCObase/RPN OpenImgbase✓14.89.625.917.6
Unbiased-Teacher [30]RPN COCObase/RPN OpenImgbase✓15.19.822.214.5
OV-RCNN [46]RPN COCObase/RPN OpenImgbase✓20.917.123.817.5
XPM [17]RPN COCObase/RPN OpenImgbase✓24.021.631.622.7
Mask-free OVIS (Ours)WSPN COCObase/WSPN COCObase✗27.425.035.925.8
", + "image_path": "ff6ab62685df579a60ce77157fe941c4e84bb546b38e110ad240193b4613aac7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 559 + ], + "type": "text", + "content": "Grad-CAM visualization is set to 8. For attention score, we directly employ the original setting of ALBEF and no additional modification is performed. Note that other pretrained vision-language models can also be integrated into our pipeline without major modifications. For the proposal generation pipeline, the WSPN network is trained using COCO base image-labels and the top " + }, + { + "bbox": [ + 46, + 415, + 288, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 415, + 288, + 559 + ], + "type": "text", + "content": " proposals candidates is set to 50. The WSPN network is trained for 40k iterations with learn rate 0.001 and weight decay 0.0001. For iterative masking, the hyper-parameter " + }, + { + "bbox": [ + 46, + 415, + 288, + 559 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 415, + 288, + 559 + ], + "type": "text", + "content": " is set to 3. In the segmentation pipeline, for each patch, the segmentation network is trained for 500 iterations with lr 0.25." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "For fair-comparison [17, 46], we use Mask R-CNN with a ResNet50 backbone as our open-vocabulary instance segmentation model. During pseudo-mask training, we train Mask-RCNN on MS-COCO and OpenImages using batch size 8 on 8 A5000 GPUs for 90k iterations. Following [11, 14], we use text embeddings obtained from a pretrained CLIP text encoder. During pseduo-mask training, the initial learning rate is set 0.01 and the background class weight is set to 0.2 to improve the recall of novel classes [46]. For base fine-tuning, the initial learning rate is set to 0.0005 and the weight decay is set to 0.0001. We run fine-tuning for 90k iterations where the learning rate is updated" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 415, + 526, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 526, + 428 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 526, + 428 + ], + "type": "text", + "content": "by a decreasing factor of 0.1 at 60k and 80k iterations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": "Object Detection: As shown in Table 1, we compare our method with previous established open-vocabulary detection methods on the MS-COCO dataset. Compared to weakly-supervised methods such as WSDDN [5] and zero-shot methods such as SB [2], DELO [56], our method outperforms them by a large margin. OV-RCNN [46], XPM [17] are OVD methods based on caption pre-training and our method trained with only pseudo-labels improves the novel category performance by " + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "inline_equation", + "content": "20.2\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": " in generalized setting, respectively. Also, when compared to the method which leverages pre-trained vision-language models such as ViLD [14], RegionCLIP [52], PB-OVD [11], our method with just pseudo-labels produces similar performance. However, with fine-tuning on base annotations, our method outperforms ViLD [14], RegionCLIP [52], PB-OVD [11] by " + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "inline_equation", + "content": "13.3\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "inline_equation", + "content": "14.9\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": " in generalized setting, respectively. This is because with fine-tuning, the model learns task/domain specific information from noise-free annotations, boosting the novel category performance. Even without base annotations, our method outperforms most of the existing OVD methods supervised using base annotations. This shows the effectiveness of our method for learning quality representation for novel categories. Specif" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23544" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 83, + 520, + 170 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 523, + 82 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 523, + 82 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 523, + 82 + ], + "type": "text", + "content": "Table 3. Object detection and Instance segmentation ablation analysis for GradCAM, Pseudo-label and Mask-RCNN training." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 83, + 520, + 170 + ], + "lines": [ + { + "bbox": [ + 70, + 83, + 520, + 170 + ], + "spans": [ + { + "bbox": [ + 70, + 83, + 520, + 170 + ], + "type": "table", + "html": "
MethodLanguage SupervisionMask-RCNN TrainingBase AnnotationObject DetectionInstance Segmentation
ConstrainedGeneralizedConstrainedGeneralized
Novel/BaseNovel/Base/AllNovel/BaseNovel/Base/All
GradCAMCNXX8.6/0.0-5.2/0.0-
PLCNXX17.3/0.0-14.8/0.0-
PL + Mask-RCNNCN✓X31.1/0.427.1/0.6/7.627.0/0.524.7/0.5/6.9
PL + Mask-RCNNCB ∪ CN✓X32.4/22.429.3/22.8/24.527.4/18.425.0/18.3/20.1
PL + Mask-RCNNCB ∪ CN✓✓35.9/40.740.0/31.5/37.731.2/36.736.0/28.7/34.0
", + "image_path": "0d4b824add802320429ea58b6087469e72b4f7cd85321ef07b335d96b606c4a0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 287, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 212 + ], + "type": "text", + "content": "ically, the quality representation is learned due to the quality proposal generated by WSPN compared to fully-supervised RPN and RCNN proposal generators." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 213, + 289, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 429 + ], + "type": "text", + "content": "Instance Segmentation: Table 2 compares our method with previous open-vocabulary instance segmentation methods on the MS-COCO and Open Images datasets. SB [2] and BA-RPN [51] are zero-shot methods which utilize different background modelling strategies and caption pre-training to learn novel categories. Compared to these, our method improves the novel category performance by a large margin on both datasets and settings. When compared against conventional pseudo-labelling methods, such as soft-teacher [43] and unbiased teacher [30], our method significantly improves on MS-COCO and Open Images datasets. Finally, when compared to open-vocabulary methods such as OV-RCNN [46], XPM [17], our method outperforms by 4.6 and " + }, + { + "bbox": [ + 46, + 213, + 289, + 429 + ], + "type": "inline_equation", + "content": "3.1\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 213, + 289, + 429 + ], + "type": "text", + "content": " in COCO and Open Images in generalized setting, respectively. All these comparisons are performed against our method trained with just pseudo-labels and no base annotation is used during training. This shows the effectiveness of our overall pseudo-mask generation pipeline." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 441, + 141, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 141, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 141, + 455 + ], + "type": "text", + "content": "4.1. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": "GradCAM vs Pseudo-Labels vs Mask-RCNN Training. In Table 3, we analyze the quality of pseudo-labels generated from the GradCAM activation map and our method (PL). After pseudo-label generation, we show how Mask-RCNN training helps to improve the quality of prediction compared to pseudo-labels. Finally, we show how fine-tuning on base annotation improves our method. In the first row, we evaluate the pseudo-mask for novel samples, where the pseudo-mask is generated by normalizing and threshold the GradCAM activation map. In the second row, we evaluate the pseudo-mask for novel samples, where the pseudomask is generated by our method. From Table 3, we can observe the quality of the pseudo-mask generated by our method for novel samples is much better than the GradCAM activation map as a pseudo-mask. From Table 3 third row, we can observe training a Mask-RCNN on pseudo-labels improves the performance on novel categories by modelling fine-grained information. By including pseudo-labels from base categories, we observe that the performance on novel samples further improves. Finally, when fine-tuning on base annotations, the performance on novel categories signifi" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 200, + 545, + 271 + ], + "blocks": [ + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "type": "text", + "content": "Table 4. Ablation analysis for object detection and instance segmentation under different language supervision." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 200, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 306, + 200, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 545, + 271 + ], + "type": "table", + "html": "
MethodLanguage SupervisionObject DetectionInstance Segmentatn
Constrd NovelGenrlzd NovelConstrd NovelGenrlzd Novel
OV-RCNN [46]Image-caption COCO27.522.8--
PB-OVD [11]Image-caption COCO-29.1--
PB-OVD [11]Img-Cap COCO, SBU, VG32.330.8--
OursImage-labels COCO35.931.531.028.3
OursImage-caption COCO36.131.831.528.8
", + "image_path": "5f1a171af983dfd320ecceeee652b93e944d2b73fc8c9e57e5b08c0b1e8057a8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 296, + 531, + 363 + ], + "blocks": [ + { + "bbox": [ + 306, + 273, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 306, + 273, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 545, + 294 + ], + "type": "text", + "content": "Table 5. Ablation analysis for object detection under different proposal generator. All models are fined-tuned on COCO base." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 296, + 531, + 363 + ], + "lines": [ + { + "bbox": [ + 316, + 296, + 531, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 296, + 531, + 363 + ], + "type": "table", + "html": "
MethodProposal GeneratorConstrained NovelGeneralized Novel
OV-RCNN [46]RPN COCO Base27.822.8
PB-OVD [11]RCNN COCO Base32.330.8
OursSelective Search34.531.0
OursWSPN COCO Base35.931.5
", + "image_path": "b4d29443a576e64cdf634251182af680f57c98fff1cf96c15dd456f013c25ed8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 366, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 389 + ], + "type": "text", + "content": "cantly improves by learning task/domain specific information from noise-free manual annotations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 581 + ], + "type": "text", + "content": "Language supervision. Given an image and caption pair, our method can generate a pseudo-mask leveraging a pretrained vision-language model. Thus to analyze the effect of captions, we conduct experiments between human-provided captions and pseudo-captions generated from image-labels. As show in Table 4, human-provided captions and image-labels based pseudo-caption produce similar performance showing that irrespective of caption type, our method can generate pseudo-mask for the object of interest (see supplementary material for visual comparison). Therefore, our method is more data efficient as it requires cheap image-labels compared to human-provided captions. Table 4 compares our method with other pseudo-label generation methods trained with extra language supervision [11]. Our method, with lesser language supervision, outperforms [11] by a considerable margin." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "Proposal Generator Quality vs Performance. In general, better quality of proposal provides better quality of pseudolabels. Therefore, we generated pseduo-labels using different proposal generator and the results are reported in Table 5. As shown in Table 5, our method trained with WSPN as proposal generator is produces better performance compared to methods which rely on fully-supervised proposal generator such as RPN and RCNN. Also when compared to selective search as proposal generator, WSPN demonstrates better performance for novel categories. This is because WSPN refines selective search proposals and localizes them" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23545" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 93, + 144, + 184 + ], + "blocks": [ + { + "bbox": [ + 76, + 72, + 143, + 92 + ], + "lines": [ + { + "bbox": [ + 76, + 72, + 143, + 92 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 143, + 92 + ], + "type": "text", + "content": "A woman holding an umbrella." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 93, + 144, + 184 + ], + "lines": [ + { + "bbox": [ + 75, + 93, + 144, + 184 + ], + "spans": [ + { + "bbox": [ + 75, + 93, + 144, + 184 + ], + "type": "image", + "image_path": "f5323d98e0678f660fcdb53d6b94c82a907a0cd30f5fc2d2a3041ba058e053bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 144, + 92, + 281, + 184 + ], + "blocks": [ + { + "bbox": [ + 182, + 82, + 242, + 91 + ], + "lines": [ + { + "bbox": [ + 182, + 82, + 242, + 91 + ], + "spans": [ + { + "bbox": [ + 182, + 82, + 242, + 91 + ], + "type": "text", + "content": "A photo of Dog." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 144, + 92, + 281, + 184 + ], + "lines": [ + { + "bbox": [ + 144, + 92, + 281, + 184 + ], + "spans": [ + { + "bbox": [ + 144, + 92, + 281, + 184 + ], + "type": "image", + "image_path": "98b404cd3d4b36803dea95a082eb36911ddce17cb4300c931a5f55a2397d3216.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 282, + 93, + 397, + 184 + ], + "blocks": [ + { + "bbox": [ + 282, + 93, + 397, + 184 + ], + "lines": [ + { + "bbox": [ + 282, + 93, + 397, + 184 + ], + "spans": [ + { + "bbox": [ + 282, + 93, + 397, + 184 + ], + "type": "image", + "image_path": "01d5ff920c96207505dd4bb367db67ed6c6e089a0747e95477739ee9f1272a35.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 187, + 282, + 199 + ], + "lines": [ + { + "bbox": [ + 75, + 187, + 282, + 199 + ], + "spans": [ + { + "bbox": [ + 75, + 187, + 282, + 199 + ], + "type": "text", + "content": "A) Visualization of activation map and pseudo-bbox " + }, + { + "bbox": [ + 75, + 187, + 282, + 199 + ], + "type": "inline_equation", + "content": "(\\mathbf{b}^{*})" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 188, + 451, + 198 + ], + "lines": [ + { + "bbox": [ + 347, + 188, + 451, + 198 + ], + "spans": [ + { + "bbox": [ + 347, + 188, + 451, + 198 + ], + "type": "text", + "content": "B) Pseudo-mask annotations" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 93, + 516, + 184 + ], + "blocks": [ + { + "bbox": [ + 399, + 93, + 516, + 184 + ], + "lines": [ + { + "bbox": [ + 399, + 93, + 516, + 184 + ], + "spans": [ + { + "bbox": [ + 399, + 93, + 516, + 184 + ], + "type": "image", + "image_path": "e9d24e5a58885e18be4417c6cb919917e1f8c8dd29da202846eee89a23983713.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 214, + 165, + 303 + ], + "blocks": [ + { + "bbox": [ + 47, + 201, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 201, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 545, + 213 + ], + "type": "text", + "content": "Figure 5. A) Pseduo bounding box selection guided by GradCAM activation. B) Visualization of pseudo-mask generated for Open Images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 214, + 165, + 303 + ], + "lines": [ + { + "bbox": [ + 50, + 214, + 165, + 303 + ], + "spans": [ + { + "bbox": [ + 50, + 214, + 165, + 303 + ], + "type": "image", + "image_path": "1da99ed894fdfd008face366861bcdac4ad6e7516fa7463994dddbb9c758d9fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 306, + 140, + 315 + ], + "lines": [ + { + "bbox": [ + 76, + 306, + 140, + 315 + ], + "spans": [ + { + "bbox": [ + 76, + 306, + 140, + 315 + ], + "type": "text", + "content": "(a) Object detection" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 319, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 365 + ], + "type": "text", + "content": "Figure 6. Ablation study for object detection and instance segmentation by increase number of masking iteration during PL generation. As iteration increases, GradCAM activation becomes noisy resulting in poor guidance and pseudo-mask generation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 168, + 214, + 284, + 303 + ], + "blocks": [ + { + "bbox": [ + 168, + 214, + 284, + 303 + ], + "lines": [ + { + "bbox": [ + 168, + 214, + 284, + 303 + ], + "spans": [ + { + "bbox": [ + 168, + 214, + 284, + 303 + ], + "type": "image", + "image_path": "5b9f6f8885241010f0532ac53d6586c4c0265a6882a6aa591f1f1aec46ea1802.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 306, + 269, + 315 + ], + "lines": [ + { + "bbox": [ + 186, + 306, + 269, + 315 + ], + "spans": [ + { + "bbox": [ + 186, + 306, + 269, + 315 + ], + "type": "text", + "content": "(b) Instance segmentation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 367, + 253, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 253, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 253, + 378 + ], + "type": "text", + "content": "towards objects producing better quality proposals." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "content": "Iterative Masking Steps vs Performance. Fig. 6 presents ablation analysis for iterative masking hyper-parametr " + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "content": ". We see that for initial few masking steps the performance increase. Because, without any masking the GradCAM activation map is towards most discriminative parts of an object. As we mask out these regions, the less discriminative regions gets activated and combining activation from previous steps produce a strong guidance function " + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "content": ". However, performing too many mask iterations could completely mask out the object of interest and unrelated background regions might get activated resulting in a poor " + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\Phi_t" + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "Weakly-supervised Instance Segmentation. Weakly-supervised Instance Segmentation (WSIS) methods performs instance-segmentation using supervision from image-labels. Similar to WSIS setting, our method performs instance-segmentation using just image-labels. Following this, we simply extended our method to WSIS setting by comparing with VOC benchmark. PENet [12] is a recent WSIS method, which relies on multiple downstream task network such as segmentation, detection and multi-class classification to perfrom segmentation. CL [18] and IAM [57] are WSIS methods utilize advanced attention mechanism to generate instance segmentation mask. As shown in Table 6, our method outperforms existing WSIS methods by considerable margin. Thus, the proposed method is simple and can be easily extended to OV setting or WSIS setting producing SOTA performance. Note that, other OV methods cannot be extended to WSIS setting as they rely on base" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 351, + 241, + 496, + 330 + ], + "blocks": [ + { + "bbox": [ + 305, + 218, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 305, + 218, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 545, + 240 + ], + "type": "text", + "content": "Table 6. Instance segmentation(mAP) performance for VOC in weakly-supervised instance segmentation setting." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 351, + 241, + 496, + 330 + ], + "lines": [ + { + "bbox": [ + 351, + 241, + 496, + 330 + ], + "spans": [ + { + "bbox": [ + 351, + 241, + 496, + 330 + ], + "type": "table", + "html": "
MethodSupervisionmAP
Mask-RCNN [15]F67.9
OCIS [8]I26.8
PRM [55]I28.3
IAM [57]I30.2
CL [18]I+C30.2
PENet [12]I38.1
OursI38.6
", + "image_path": "f426c5c6c70448a8cb2cf59e03b9d1d4f950226199f36365fb4a314b4caa580e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 338, + 545, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 362 + ], + "type": "text", + "content": "annotation and similarly WSIS methods cannot be extended to OV setting due to constrained vocabulary space." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 365, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 533 + ], + "type": "text", + "content": "Qualitative Analysis. Fig. 5 (A) presents a visualization of activation maps generated for the object of interest (woman and dog). As we can see, the generated activation map covers the entire object and it can be used as a guidance function to choose the best bounding box proposal. Note that the activation are square-shaped because the original activation map is " + }, + { + "bbox": [ + 304, + 365, + 546, + 533 + ], + "type": "inline_equation", + "content": "1/16" + }, + { + "bbox": [ + 304, + 365, + 546, + 533 + ], + "type": "text", + "content": "th of the image size. We perform nearest interpolation to obtain an activation map of image size. In Fig 5 (B), we visualize the pseudo-mask for Open Images generated from our pipeline. We can observe that the generated pseudo-mask is of good quality; still contains some false positives. However, with Mask-RCNN training, the model learns to filter the noise present in the pseudo-mask producing better-quality predictions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 540, + 378, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 378, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 378, + 552 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "We propose a novel pipeline called Mask-free OVIS for open-vocabulary instance segmentation, which eliminates the need for costly and time-consuming manual mask annotations. We achieve this by generating pseudo-mask annotation for base and novel categories by leveraging a pretrained vision-language model and weakly supervised techniques. The generated pseudo-mask annotations are then used to supervise the Mask-RCNN model. Extensive experiments are conducted to show the effectiveness of our method on MS-COCO and Open Images datasets. Instead of relying on labour-expensive instance-level annotations, we hope our simple and effective pipeline provides an alternative to detect and segment long-tailed novel categories." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23546" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Aditya Arun, CV Jawahar, and M Pawan Kumar. Dissimilarity coefficient based weakly supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9432-9441, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "text", + "content": "[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 384-400, 2018. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 224 + ], + "type": "text", + "content": "[3] Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What's the point: Semantic segmentation with point supervision. In European conference on computer vision, pages 549-565. Springer, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 287, + 280 + ], + "type": "text", + "content": "[4] David Biertimpel, Sindi Shkodrani, Anil S Baslamisli, and Nora Baka. Prior to segment: Foreground cues for weakly annotated classes in partially supervised instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2824-2833, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "type": "text", + "content": "[5] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2846-2854, 2016. 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 392 + ], + "type": "text", + "content": "[6] Maria A Bravo, Sudhanshu Mittal, and Thomas Brox. Localized vision-language matching for open-vocabulary object detection. In Pattern Recognition: 44th DAGM German Conference, DAGM GCR 2022, Konstanz, Germany, September 27–30, 2022, Proceedings, pages 393–408. Springer, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 457 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, Alexander Hermans, George Papandreou, Florian Schroff, Peng Wang, and Hartwig Adam. Masklab: Instance segmentation by refining object detection with semantic and direction features. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4013-4022, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 513 + ], + "type": "text", + "content": "[8] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12405, 2019. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 557 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 613 + ], + "type": "text", + "content": "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 680 + ], + "type": "text", + "content": "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 266-282. Springer, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "text", + "content": "[12] Weifeng Ge, Sheng Guo, Weilin Huang, and Matthew R Scott. Label-penet: Sequential label propagation and enhancement networks for weakly supervised instance segmen-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3345-3354, 2019. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "type": "text", + "content": "[13] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. CyCLIP: Cyclic contrastive language-image pretraining. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 161, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 161, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 161, + 545, + 205 + ], + "type": "text", + "content": "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 5, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "text", + "content": "[16] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Advances in neural information processing systems, 33:16797-16807, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 294, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 350 + ], + "type": "text", + "content": "[17] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[18] Jaedong Hwang, Seohyun Kim, Jeany Son, and Bohyung Han. Weakly supervised instance segmentation by deep community learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1020-1029, 2021. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "text", + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[20] Wonjik Kim, Asako Kanezaki, and Masayuki Tanaka. Unsupervised learning of image segmentation based on differentiable feature clustering. IEEE Transactions on Image Processing, 29:8055-8068, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 514, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 569 + ], + "type": "text", + "content": "[21] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 1, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "type": "text", + "content": "[22] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "type": "text", + "content": "[23] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for uni" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23547" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "type": "text", + "content": "fied vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "text", + "content": "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 208 + ], + "type": "text", + "content": "[26] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell me where to look: Guided attention inference network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "text", + "content": "[27] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 255, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 288, + 319 + ], + "type": "text", + "content": "[28] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 378, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 288, + 432 + ], + "type": "text", + "content": "[30] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In International Conference on Learning Representations, 2021. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 288, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 478 + ], + "type": "text", + "content": "[31] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 479, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 544 + ], + "type": "text", + "content": "[32] Zongyang Ma, Guan Luo, Jin Gao, Liang Li, Yuxin Chen, Shaoru Wang, Congxuan Zhang, and Weiming Hu. Open-vocabulary one-stage detection with hierarchical visual-language knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14074-14083, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 546, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 288, + 601 + ], + "type": "text", + "content": "[33] Niall O'Mahony, Sean Campbell, Anderson Carvalho, Suman Harapanahalli, Gustavo Velasco Hernandez, Lenka Krpalkova, Daniel Riordan, and Joseph Walsh. Deep learning vs. traditional computer vision. In Science and information conference, pages 128-144. Springer, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[35] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11932–11939, 2020. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[36] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In 36th Conference on Neural Information Processing Systems (NIPS), 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "text", + "content": "[37] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 546, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 240 + ], + "type": "text", + "content": "[38] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, MingYu Liu, Yong Jae Lee, Alexander G Schwing, and Jan Kautz. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598-10607, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 241, + 546, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 546, + 306 + ], + "type": "text", + "content": "[39] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 363 + ], + "type": "text", + "content": "[40] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 365, + 546, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 419 + ], + "type": "text", + "content": "[41] Limin Wang, Yuanjun Xiong, Dahua Lin, and Luc Van Gool. Untrimmednets for weakly supervised action recognition and detection. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4325-4334, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 421, + 546, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 546, + 464 + ], + "type": "text", + "content": "[42] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In European Conference on Computer Vision, pages 649-665. Springer, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 466, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 546, + 521 + ], + "type": "text", + "content": "[43] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "type": "text", + "content": "[44] Ke Yang, Dongsheng Li, and Yong Dou. Towards precise end-to-end weakly supervised object detection network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8372-8381, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 568, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 546, + 624 + ], + "type": "text", + "content": "[45] Keren Ye, Mingda Zhang, Adriana Kovashka, Wei Li, Danfeng Qin, and Jesse Berent. Cap2det: Learning to amplify weak caption supervision for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9686-9695, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 624, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 546, + 678 + ], + "type": "text", + "content": "[46] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "text", + "content": "[47] Zhaoyang Zeng, Bei Liu, Jianlong Fu, Hongyang Chao, and Lei Zhang. Wsod2: Learning bottom-up and top-down objectness distillation for weakly-supervised object detection." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23548" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 664 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF international conference on computer vision, pages 8292-8300, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[48] Dingwen Zhang, Junwei Han, Gong Cheng, and Ming-Hsuan Yang. Weakly supervised object localization and detection: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5866-5885, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 194 + ], + "type": "text", + "content": "[49] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 274 + ], + "type": "text", + "content": "[50] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 159-175. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[51] Ye Zheng, Jiahong Wu, Yongqiang Qin, Faen Zhang, and Li Cui. Zero-shot instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2593-2602, 2021. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 385 + ], + "type": "text", + "content": "[52] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 287, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 287, + 439 + ], + "type": "text", + "content": "[53] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 441, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 495 + ], + "type": "text", + "content": "[54] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 498, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 551 + ], + "type": "text", + "content": "[55] Yanzhao Zhou, Yi Zhu, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Weakly supervised instance segmentation using class peak response. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3791-3800, 2018. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 554, + 287, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 607 + ], + "type": "text", + "content": "[56] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't even look once: Synthesizing features for zero-shot detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11693-11702, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 609, + 287, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 287, + 664 + ], + "type": "text", + "content": "[57] Yi Zhu, Yanzhao Zhou, Huijuan Xu, Qixiang Ye, David Doermann, and Jianbin Jiao. Learning instance activation maps for weakly supervised instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3116-3125, 2019. 8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23549" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_content_list.json b/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..86f2b20b2e3c0090bfca7e4b017acb8b3c74a5c1 --- /dev/null +++ b/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "Mask-Free Video Instance Segmentation", + "text_level": 1, + "bbox": [ + 277, + 121, + 691, + 142 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Ke $^{1,2}$ Martin Danelljan $^{1}$ Henghui Ding $^{1}$ Yu-Wing Tai $^{2}$ Chi-Keung Tang $^{2}$ Fisher Yu $^{1}$ ETH Zürich ${}^{2}\\mathrm{HKUST}$", + "bbox": [ + 109, + 170, + 857, + 205 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2df21adc52ff236309c7cb72bd869ca6ddfe29f15f25f57a188924dfbe5cd526.jpg", + "image_caption": [ + "Figure 1. Video instance segmentation (VIS) results of our MaskFreeVIS, trained without using any video or image mask annotation. By achieving a remarkable $42.5\\%$ mask AP on the YouTube-VIS val dataset, with a ResNet-50 backbone, our approach demonstrates that high-performing VIS can be learned even without any mask annotations." + ], + "image_footnote": [], + "bbox": [ + 81, + 234, + 888, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 469, + 313, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent advancement in Video Instance Segmentation (VIS) has largely been driven by the use of deeper and increasingly data-hungry transformer-based models. However, video masks are tedious and expensive to annotate, limiting the scale and diversity of existing VIS datasets. In this work, we aim to remove the mask-annotation requirement. We propose MaskFreeVIS, achieving highly competitive VIS performance, while only using bounding box annotations for the object state. We leverage the rich temporal mask consistency constraints in videos by introducing the Temporal KNN-patch Loss (TK-Loss), providing strong mask supervision without any labels. Our TK-Loss finds one-to-many matches across frames, through an efficient patch-matching step followed by a $K$ -nearest neighbor selection. A consistency loss is then enforced on the found matches. Our mask-free objective is simple to implement, has no trainable parameters, is computationally efficient, yet outperforms baselines employing, e.g., state-of-the-art optical flow to enforce temporal mask consistency. We validate MaskFreeVIS on the YouTube-VIS 2019/2021, OVIS and BDD100K MOTS benchmarks. The results clearly demonstrate the efficacy of our method by drastically narrowing the gap between fully and weakly-supervised VIS performance. Our code and trained models are available at http://vis.xyz/pub/maskfreevis.", + "bbox": [ + 75, + 517, + 472, + 895 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 469, + 630, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Instance Segmentation (VIS) requires jointly detecting, tracking and segmenting all objects in a video from a given set of categories. To perform this challenging task, state-of-the-art VIS models are trained with complete video annotations from VIS datasets [39, 61, 64]. However, video annotation is costly, in particular regarding object mask labels. Even coarse polygon-based mask annotation is multiple times slower than annotating video bounding boxes [7]. Expensive mask annotation makes existing VIS benchmarks difficult to scale, limiting the number of object categories covered. This is particularly a problem for the recent transformer-based VIS models [6, 17, 57], which tend to be exceptionally data-hungry. We therefore revisit the need for complete mask annotation by studying the problem of weakly supervised VIS under the mask-free setting.", + "bbox": [ + 496, + 500, + 893, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While there exist box-supervised instance segmentation models [13, 23, 27, 50], they are designed for images. These weakly-supervised single-image methods do not utilize temporal cues when learning mask prediction, leading to lower accuracy when directly applied to videos. As a source for weakly supervised learning, videos contain much richer information about the scene. In particular, videos adhere to the temporal mask consistency constraint, where the regions corresponding to the same underlying object across different frames should have the same mask label. In this work, we set out to leverage this important constraint for mask-free learning of VIS.", + "bbox": [ + 496, + 729, + 895, + 912 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22857", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad65f7aa1ab4c1a9f2dff2f18954500f56a51c3dda909f4eac4ed3f9e17ef030.jpg", + "image_caption": [ + "Temporal One-to-K Patch Correspondence", + "Figure 2. Our Temporal KNN-patch Loss enforces mask consistency between one-to- $k$ patch correspondences found across frames, which allow us to cover the cases where: (i) A unique one-to-one match exists (blue); (ii) Multiple matches are found due to ambiguities in homogenous regions (orange) or along image edges (white and yellow); (iii) No match is found due to $e.g.$ occlusions (green). This allows us to robustly leverage mask consistency constraints in challenging videos." + ], + "image_footnote": [], + "bbox": [ + 80, + 82, + 467, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose MaskFreeVIS method, for high performance VIS without any mask annotations. To leverage temporal mask consistency, we introduce the Temporal KNN-patch Loss (TK-Loss), as in Figure 2. To find regions corresponding to the same underlying video object, our TK-Loss first builds correspondences across frames by patch-wise matching. For each target patch, only the top $K$ matches in the neighboring frame with high enough matching score are selected. A temporal consistency loss is then applied to all found matches to promote the mask consistency. Specifically, our surrogate objective function not only promotes the one-to- $k$ matched regions to reach the same mask probabilities, but also commits their mask prediction to a confident foreground or background prediction by entropy minimization. Unlike flow-based models [32,45], which assume one-to-one matching, our approach builds robust and flexible one-to- $k$ correspondences to cope with e.g. occlusions and homogeneous regions, without introducing additional model parameters or inference cost.", + "bbox": [ + 75, + 332, + 468, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The TK-Loss is easily integrated into existing VIS methods, with no architecture modifications required. During training, our TK-Loss simply replaces the conventional video mask losses in supervising video mask generation. To further enforce temporal consistency through the video clip, TK-Loss is employed in a cyclic manner instead of using dense frame-wise connections. This greatly reduces memory cost with negligible performance drop.", + "bbox": [ + 75, + 621, + 468, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We extensively evaluate our MaskFreeVIS on four large-scale VIS benchmarks, i.e., YouTube-VIS 2019/2021 [61], OVIS [39], and BDD100K MOTS [64]. MaskFreeVIS achieves competitive VIS performance without using any video masks or even image mask labels on all datasets. Validated on various methods and backbones, MaskFreeVIS achieves $91.25\\%$ performance of its fully supervised counterparts, even outperforming a few recent fully-supervised methods [10, 15, 18, 59] on the popular YTVIS benchmark. Our simple yet effective design greatly narrows the performance gap between weakly-supervised and fully", + "bbox": [ + 75, + 744, + 470, + 912 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Mask annotation requirement for state-of-the-art VIS methods. Results are reported using ResNet-50 as backbone on the YTVIS 2019 [61] benchmark. Video Mask: using YTVIS video mask labels. Image Mask: using COCO [30] image mask labels for image-based pretraining. Pseudo Video: using Pseudo Videos from COCO images for joint training [57]. MaskFreeVIS achieves $91.5\\%$ (42.5 vs. 46.4) of its fully-supervised baseline performance (Mask2Former) without using any masks during training.", + "bbox": [ + 498, + 79, + 890, + 191 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/1ff69b0913aabd0b5af8a7762d33d57189dba1c6801dc2c0cb43fd812c2ac214.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVideo MaskImage MaskPseudo VideoAP
SeqFormer [57]✓✓✓47.4
VMT [17]✓✓✓47.9
Mask2Former [6]✓✓✓47.8
MaskFreeVIS (ours)✗✓✓46.6
Mask2Former [6]✓✓✗46.4
MaskFreeVIS (ours)✗✗✗42.5
", + "bbox": [ + 501, + 193, + 888, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "supervised video instance segmentation. It further demonstrates that expensive video masks, or even image masks, is not necessary for training high-performing VIS models.", + "bbox": [ + 498, + 325, + 890, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: (i) To utilize temporal information, we develop a new parameter-free Temporal KNN-patch Loss, which leverages temporal masks consistency using unsupervised one-to- $k$ patch correspondence. We extensively analyze the TK-Loss through ablative experiments. (ii) Based on the TK-Loss, we develop the MaskFreeVIS method, enabling training existing state-of-the-art VIS models without any mask annotation. (iii) To the best of our knowledge, MaskFreeVIS is the first mask-free VIS method attaining high-performing segmentation results. We provide qualitative results in Figure 1. As in Table 1, when integrated into the Mask2Former [6] baseline with ResNet-50, our MaskFreeVIS achieves $42.5\\%$ mask AP on the challenging YTVIS 2019 benchmark while using no video or image mask annotations. Our approach further scales to larger backbones, achieving $55.3\\%$ mask AP on Swin-L backbone with no video mask annotations.", + "bbox": [ + 496, + 372, + 892, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We hope our approach will facilitate achieving label-efficient video instance segmentation, enabling building even larger-scale VIS benchmarks with diverse categories by lifting the mask annotation restriction.", + "bbox": [ + 496, + 628, + 890, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 704, + 640, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Instance Segmentation (VIS) Existing VIS methods can be summarized into three categories: two-stage, one-stage, and transformer-based. Two-stage approaches [2, 18, 28, 29, 61] extend the Mask R-CNN family [11, 19] by designing an additional tracking branch for object association. One-stage works [4, 26, 31, 62] adopt anchor-free detectors [49], generally using linear masks basis combination [3] or conditional mask prediction generation [48]. For the transformer-based models [6, 12, 46, 57, 63], VisTr [54] firstly adapts the transformer [5] for VIS, and IFC [15] further improves its efficiency via memory tokens. Seqformer [57] proposes frame query decompo", + "bbox": [ + 496, + 729, + 890, + 912 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22858", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sition while Mask2Former [6] includes masked attention. VMT [17] extends Mask Transfiner [16] to video for high-quality VIS, and IDOL [58] focuses on online VIS. State-of-the-art VIS methods with growing capacity put limited emphasis on weak supervision. In contrast, the proposed MaskFreeVIS is the first method targeting mask-free VIS while attaining competitive performance.", + "bbox": [ + 75, + 82, + 467, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiple Object Tracking and Segmentation (MOTS) Most MOTS methods [1,34,35,53,56] follow the tracking-by-detection principle. PCAN [18] improves temporal segmentation by utilizing space-time memory prototypes, while the one-stage method Unicorn [60] focuses on unification of different tracking frameworks. Compared to the aforementioned fully-supervised MOTS methods, Mask-FreeVIS focuses on label efficient training without GT masks by proposing a new surrogate temporal loss which can be easily integrated on them.", + "bbox": [ + 75, + 188, + 467, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mask-Free VIS Most mask-free instance segmentation works [7, 13, 20, 25, 27, 37, 40, 44] are designed for single images and thus neglect temporal information. Earlier works BoxSup [8] and Box2Seg [22] rely on region proposals produced by MCG [38] or GrabCut [42], leading to slow training. BoxInst [50] proposes the surrogate projection and pixel pairwise losses to replace the original mask learning loss of CondInst [48], while DiscoBox [23] focuses on generating pseudo mask labels guided by a teacher model.", + "bbox": [ + 75, + 340, + 467, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Earlier works have investigated the use of videos for weakly-, semi-, or un-supervised segmentation by leveraging motion or temporal consistency [21,51,52]. Most aforementioned approaches do not address the VIS problem, and use optical flow for frame-to-frame matching [24, 32, 43]. In particular, FlowIRN [32] explores VIS using only classification labels and incorporates optical flow to leverage mask consistency. The limited performance makes the class-label only or fully-unsupervised setting difficult to deploy in the real world. SOLO-Track [9] aims to train VIS models without video annotations, and one concurrent work MinVIS [14] performs VIS without video-based model architectures. Unlike the above weakly-supervised training settings, our MaskFreeVIS is designed for eliminating the mask annotation requirement for VIS, as we note that video mask labeling is particularly expensive. MaskFreeVIS enables training VIS models without any video masks, or even image masks. Despite its simplicity, MaskFreeVIS drastically reduces the gap between fully-supervised and weakly-supervised VIS models, making weakly-supervised models more accessible in practice.", + "bbox": [ + 75, + 477, + 467, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 810, + 166, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose MaskFreeVIS to tackle video instance segmentation (VIS) without using any video or even image mask labels. Our approach is generic and can be directly applied to train state-of-the-art VIS methods, such as Mask2Former [6] and SeqFormer [57]. In Sec. 3.1, we first", + "bbox": [ + 75, + 835, + 467, + 911 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "present the core component of MaskFreeVIS: the Temporal KNN-patch Loss (TK-Loss), which leverages temporal consistency to supervise accurate mask prediction, without any human mask annotations. In Sec. 3.2, we then describe how to integrate the TK-Loss with existing spatial weak segmentation losses for transformer-based VIS methods, to achieve mask-free training of VIS approaches. Finally, we introduce image-based pretraining details of our MaskFreeVIS in Sec. 3.2.3.", + "bbox": [ + 496, + 82, + 890, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. MaskFreeVIS", + "text_level": 1, + "bbox": [ + 500, + 227, + 643, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the Temporal KNN-patch Loss, illustrated in Figure 3. It serves as an unsupervised objective for mask prediction that leverages the rich spatiotemporal consistency constraints in unlabelled videos.", + "bbox": [ + 496, + 250, + 890, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Temporal Mask Consistency", + "text_level": 1, + "bbox": [ + 498, + 329, + 751, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While an image constitutes a single snapshot of a scene, a video provides multiple snapshots displaced in time. Thereby, a video depicts continuous change in the scene. Objects and background move, deform, are occluded, experience variations in lighting, motion blur, and noise, leading to a sequence of different images that are closely related through gradual transformations.", + "bbox": [ + 496, + 352, + 890, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a small region in the scene (Fig. 2), belonging either to an object or background. The pixels corresponding to the projection of this region should have the same mask prediction in every frame, as they belong to the same underlying physical object or background region. However, the aforementioned dynamic changes in the video lead to substantial appearance variations, serving as a natural form of data augmentation. The fact that the pixels corresponding to the same underlying object region should have the same mask prediction under temporal change therefore provides a powerful constraint, i.e., temporal mask consistency, which can be used for mask supervision [21, 24, 32, 51, 52].", + "bbox": [ + 496, + 458, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The difficulty in leveraging the temporal mask consistency constraint stems from the problem of establishing reliable correspondences between video frames. Objects often undergo fast motion, deformations, etc., resulting in substantial appearance change. Furthermore, regions in the scene may be occluded or move out of the image from one frame to the other. In such cases, no correspondence exist. Lastly, videos are often dominated by homogenous regions, such as sky and ground, where the establishment of one-to-one correspondences are error-prone or even ill-defined.", + "bbox": [ + 496, + 640, + 890, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The problem of establishing dense one-to-one correspondences between subsequent video frames, known as optical flow, is a long-standing and popular research topic. However, when attempting to enforce temporal mask consistency through optical flow estimation [24,32,43], one encounters two key problems. 1) The one-to-one assumption of optical flow is not suitable in cases of occlusions, homogenous regions, and along single edges, where the corre", + "bbox": [ + 496, + 791, + 890, + 911 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22859", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08adf8978d71ca66c5814d3e778b63990cadabda627e958f8883c6a02009439f.jpg", + "image_caption": [ + "(1) Temporal Candidates Extraction", + "Figure 3. Temporal KNN-patch Loss has four steps: 1) Patch Candidate Extraction: Patch candidates searching across frames with radius $R$ . 2) Temporal KNN-Matching: Match $k$ high-confidence candidates by patch affinities. 3) Consistency loss: Enforce mask consistency objective (Eq. 2) among the matches. 4) Cyclic Tube Connection: Temporal loss aggregation in the 5-frame tube, detailed in Figure 4." + ], + "image_footnote": [], + "bbox": [ + 80, + 61, + 225, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72a30751e06486c0c306acfc9f139e3694ae338d6b3e7a905476a6ab62e523b8.jpg", + "image_caption": [ + "(2) Temporal KNN-Matching" + ], + "image_footnote": [], + "bbox": [ + 230, + 61, + 357, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2e8f94bf5ff32f721a2b28f5d96974c30864adc36a2bddeadea1ef7b5be769c1.jpg", + "image_caption": [ + "Loss Backward" + ], + "image_footnote": [], + "bbox": [ + 359, + 61, + 532, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9dc5c571ae48cdb53660ba47af13fcd4b616a62a40e9d40b586b29b3cddc1957.jpg", + "image_caption": [ + "(3) KNN-Consistency Loss" + ], + "image_footnote": [], + "bbox": [ + 532, + 61, + 694, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/faa4326afa7e0473b380c95f8a1f02eefc5ed465e09785eb251695d39413c4b1.jpg", + "image_caption": [ + "4) Cyclic Tube Connection" + ], + "image_footnote": [], + "bbox": [ + 694, + 61, + 761, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77f665369b9c035734a800aeb31c05f322310ee968e66b657ddce88211558705.jpg", + "image_caption": [ + "MaskFreeVIS Prediction" + ], + "image_footnote": [], + "bbox": [ + 763, + 63, + 888, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "spondence is either nonexistent, undefined, ambiguous, uncertain, or very difficult to determine. 2) State-of-the-art optical flow estimation relies on large and complex deep networks, with large computational and memory requirements.", + "bbox": [ + 75, + 280, + 468, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of using optical flow, we aim to design a simple, efficient, and parameter-free approach that effectively enforces the temporal mask consistency constraint.", + "bbox": [ + 75, + 340, + 468, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Temporal KNN-patch Loss", + "text_level": 1, + "bbox": [ + 76, + 390, + 318, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our Temporal KNN-patch Loss (TK-Loss) is based on a simple and flexible correspondence estimation across frames. In contrast to optical flow, we do not restrict our formulation to one-to-one correspondences. Instead, we establish one-to- $K$ correspondences. This include the conventional one-to-one $(K = 1)$ , where a unique well-defined match exists. However, this allows us to also handle the cases of nonexistent correspondences $(K = 0)$ in case of occlusions, and one-to-many $(K \\geq 2)$ in case of homogenous regions. In cases where multiple matches are found, these most often belong to the same underlying object or background due to their similar appearance, as in Figure 2. This further benefits our mask consistency objective through denser supervision. Lastly, our approach is simple to implement, with negligible computational overhead and no learnable parameters. Our approach is in Figure 3, and contains four main steps, which are detailed next.", + "bbox": [ + 75, + 414, + 468, + 670 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Patch Candidate Extraction: Let $X_p^t$ denote an $N \\times N$ target image patch centered at spatial location $p = (x, y)$ in frame $t$ . Our aim is to find a set of corresponding positions $S_p^{t \\to \\hat{t}} = \\{\\hat{p}_i\\}_i$ in frame number $\\hat{t}$ that represent the same object region. To this end, we first select candidate locations $\\hat{p}$ within a radius $R$ such that $\\| p - \\hat{p} \\| \\leq R$ . Such windowed patch search exploits spatial proximity across neighboring frames in order to avoid an exhaustive global search. For a fast implementation, the windowed search is performed for all target image patches $X_p^t$ in parallel.", + "2) Temporal KNN-Matching: We match patch candidate patches through a simple distance computation," + ], + "bbox": [ + 75, + 670, + 468, + 852 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {p \\rightarrow \\hat {p}} ^ {t \\rightarrow \\hat {t}} = \\left\\| X _ {p} ^ {t} - X _ {\\hat {p}} ^ {\\hat {t}} \\right\\|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 862, + 468, + 887 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our ablative experiments (Sec. 4.3), we found the $L_{2}$", + "bbox": [ + 96, + 896, + 468, + 912 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "norm to be the most effective patch matching metric. We select the top $K$ matches with smallest patch distance $\\mathbf{d}_{p\\rightarrow \\hat{p}}^{t\\rightarrow \\hat{t}}$ . Lastly low-confidence matches are removed by enforcing a maximal patch distance $D$ as $\\mathbf{d}_{p\\rightarrow \\hat{p}}^{t\\rightarrow \\hat{t}} < D$ . The remaining matches form the set $S_{p}^{t\\rightarrow \\hat{t}} = \\{\\hat{p}_{i}\\}_{i}$ for each location $p$ .", + "bbox": [ + 496, + 280, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3) Consistency loss: Let $M_p^t \\in [0,1]$ denote the predicted binary instance mask of an object, evaluated at position $p$ in frame $t$ . To ensure temporal mask consistency constraints, we penalize inconsistent mask predictions between a spatiotemporal point $(p,t)$ and its estimated corresponding points in $S_p^{t \\to \\hat{t}}$ . In particular we use the following objective,", + "bbox": [ + 496, + 359, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} ^ {t \\rightarrow \\hat {t}} = \\frac {1}{H W} \\sum_ {p} \\sum_ {\\hat {p} _ {i} \\in \\mathcal {S} _ {p} ^ {t \\rightarrow \\hat {t}}} L _ {\\text {c o n s}} \\left(M _ {p} ^ {t}, M _ {\\hat {p} _ {i}} ^ {\\hat {t}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 468, + 890, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where mask consistency is measured as", + "bbox": [ + 500, + 523, + 763, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o n s}} \\left(M _ {p} ^ {t}, M _ {\\hat {p}} ^ {\\hat {t}}\\right) = - \\log \\left(M _ {p} ^ {t} M _ {\\hat {p}} ^ {\\hat {t}} + \\left(1 - M _ {p} ^ {t}\\right) \\left(1 - M _ {\\hat {p}} ^ {\\hat {t}}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 551, + 890, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that Eq. (3) only attains its minimum value of zero if both predictions indicate exactly background ( $M_p^t = M_{\\hat{p}}^{\\hat{t}} = 0$ ) or foreground ( $M_p^t = M_{\\hat{p}}^{\\hat{t}} = 1$ ). The objective does thus not only promote the two mask predictions to achieve the same probability value $M_p^t = M_{\\hat{p}}^{\\hat{t}}$ , but also to commit to a certain foreground or background prediction.", + "bbox": [ + 496, + 585, + 890, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4) Cyclic Tube Connection: Suppose the temporal tube consists of $T$ frames. We compute the temporal loss for the whole tube in a cyclic manner, as in Figure 4. The start frame is connected to the end frame, which introduces direct long-term mask consistency across the two temporally most distant frames. The temporal TK-Loss for the whole tube is given by", + "bbox": [ + 496, + 681, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e m p}} = \\sum_ {t = 1} ^ {T} \\left\\{\\begin{array}{l l}\\mathcal {L} _ {f} ^ {t \\rightarrow (t + 1)}&\\text {i f} t < T - 1\\\\\\mathcal {L} _ {f} ^ {t \\rightarrow 0}&\\text {i f} t = T - 1.\\end{array}\\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 801, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to inter-frame dense connections in Figure 4, we find the cyclic loss to achieve similar performance but greatly reduce the memory usage as validated in the experiment section.", + "bbox": [ + 496, + 851, + 890, + 910 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22860", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/025f71d424fdc899da266bf565e2003fb4f9a1955446db6c208414ca14289cce.jpg", + "image_caption": [ + "Figure 4. Illustration of different frame-wise tube connection settings (connection number) in the temporal loss design." + ], + "image_footnote": [], + "bbox": [ + 83, + 74, + 460, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Training MaskFreeVIS", + "text_level": 1, + "bbox": [ + 76, + 265, + 294, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe how to train state-of-the-art VIS methods using our TK-Loss, without any mask annotations. Our MaskFreeVIS approach is jointly supervised with spatial-temporal surrogate losses, and is easily integrated with existing transformer-based methods. We also detail mask-free image-based pre-training for MaskFree-VIS to fully eliminate mask usage during training.", + "bbox": [ + 75, + 287, + 468, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Joint Spatio-temporal Regularization", + "text_level": 1, + "bbox": [ + 76, + 402, + 390, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train MaskFreeVIS, in addition to our proposed Temporal KNN-patch Loss for temporal mask consistency, we leverage existing spatial weak segmentation losses to jointly enforce intra-frame consistency.", + "bbox": [ + 75, + 426, + 468, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Consistency To explore spatial weak supervision signals from image bounding boxes and pixel color, we utilize the representative Box Projection Loss $L_{\\mathrm{proj}}$ and Pairwise Loss $L_{\\mathrm{pair}}$ in [50], to replace the supervised mask learning loss. The Projection Loss $L_{\\mathrm{proj}}$ enforces the projection $P'$ of the object mask onto the $\\vec{x}$ -axis and $\\vec{y}$ -axis of image to be consistent with its ground-truth box mask. For the temporal tube with $T$ frames, we concurrently optimize all predicted frame masks of the tube as,", + "bbox": [ + 75, + 486, + 468, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p r o j}} = \\sum_ {t = 1} ^ {T} \\sum_ {d \\in \\{\\vec {x}, \\vec {y} \\}} D \\left(P _ {d} ^ {\\prime} \\left(M _ {p} ^ {t}\\right), P _ {d} ^ {\\prime} \\left(M _ {b} ^ {t}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 628, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D$ denotes dice loss, $P'$ is the projection function along $\\vec{x} / \\vec{y}$ -axis direction, $M_p^t$ and $M_b^t$ denote predicted instance mask and its GT box mask at frame $t$ respectively. The object instance index is omitted here for clarity.", + "bbox": [ + 75, + 680, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Pairwise Loss $L_{\\mathrm{pair}}$ , on the other hand, constrains spatially neighboring pixels of single frame. For pixel of locations $p_i'$ and $p_j'$ of with color similarity $\\geqslant \\sigma_{\\mathrm{pixel}}$ , we enforce their predicted mask labels to be consistent, following Eq. (3) as,", + "bbox": [ + 75, + 739, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p a i r}} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {p _ {i} ^ {\\prime} \\in H \\times W} L _ {\\text {c o n s}} \\left(M _ {p _ {i} ^ {\\prime}} ^ {t}, M _ {p _ {j} ^ {\\prime}} ^ {t}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 823, + 468, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The spatial losses are combined with a weight factor $\\lambda_{\\mathrm{pair}}$", + "bbox": [ + 76, + 873, + 460, + 890 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s p a t i a l}} = \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {p a i r}} \\mathcal {L} _ {\\text {p a i r}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 896, + 468, + 914 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Consistency We adopt the Temporal KNN-patch Loss in Sec. 3.1.2 as $\\mathcal{L}_{\\mathrm{temp}}$ to leverage temporal mask consistency. The overall spatio-temporal objective $\\mathcal{L}_{\\mathrm{seg}}$ for optimizing video segmentation is summarized as,", + "bbox": [ + 498, + 82, + 890, + 142 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e g}} = \\mathcal {L} _ {\\text {s p a t i a l}} + \\lambda_ {\\text {t e m p}} \\mathcal {L} _ {\\text {t e m p}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 155, + 890, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Integration with Transformer-based Methods", + "text_level": 1, + "bbox": [ + 498, + 189, + 870, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing works [13, 48] on box-supervised segmentation losses are coupled with either one-stage or two-stage detectors, such as Faster R-CNN [41] and CondInst [48], and only address the single image case. However, state-of-the-art VIS methods [6, 57] are based on transformers. These works perform object detection via set prediction, where predicted instance masks need to be matched with mask annotations when evaluating the loss. To integrate mask-free VIS training with transformers, one key modification is in this instance-sequence matching step.", + "bbox": [ + 496, + 213, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since only ground-truth bounding boxes are available for box sequence matching, as an initial attempt, we first produce bounding box predictions from the estimated instance masks. Then, we employ the sequential box matching cost function used in VIS methods [55, 57]. To compute matching cost for whole sequence, $\\mathcal{L}_1$ loss and generalized IoU loss for each individual bounding box is averaged across the frames. However, we observe the matching results of frame-wise averaging can easily be affected by a single outlier frame, especially under weak segmentation setup, leading to instability during training and performance decrease. Spatio-temporal Box Mask Matching Instead of using the aforementioned frame-wise matching, we empirically find spatio-temporal box-to-mask matching to produce substantial improvement under the weak segmentation setting. We first convert each predicted instance mask to a bounding box mask, and convert the ground-truth box to box mask. We then randomly sample a equal number of points from the ground-truth box mask sequence and predicted box mask sequence, respectively. Different from Mask2Former [6], we only adopt the dice IoU loss to compute sequence matching cost. We find that cross-entropy accumulates errors per pixel, leading to imbalanced values between large and small objects. In contrast, the IoU loss in normalized per object, leading to a balanced metric. We study different instance sequence matching strategies under the mask-free VIS setting in the ablation experiments.", + "bbox": [ + 496, + 364, + 892, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Image-based MaskFreeVIS Pre-training", + "text_level": 1, + "bbox": [ + 498, + 782, + 834, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most VIS models [6, 57, 61] are initialized from a model pretrained on the COCO instance segmentation dataset. To completely eliminate mask supervision, we pretrain our MaskFreeVIS on COCO using only box supervision. We adopt the spatial consistency loss described in Sec. 3.2.1 on single frame to replace the original GT mask losses in Mask2Former [6], while following the same image-based", + "bbox": [ + 496, + 806, + 890, + 912 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22861", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training setup on COCO. Thus, we provide two training settings in our experiments, one eliminates both image and video mask during training, while the other adopts weights pretrained with COCO mask annotations. In both cases, no video mask annotations are used.", + "bbox": [ + 78, + 82, + 467, + 156 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 78, + 167, + 207, + 185 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 78, + 193, + 176, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YTVIS 2019/2021 We perform experiments on the largescale YouTube-VIS [61] 2019 and 2021. YTVIS 2019 includes 2,883 videos of 131k annotated object instances belonging to 40 categories. To handle more complex cases, YTVIS 2021 updates YTVIS 2019 with additional 794 videos for training and 129 videos for validation, including more tracklets with confusing motion trajectories.", + "bbox": [ + 78, + 215, + 467, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OVIS We also train and evaluate on OVIS [39], a VIS benchmark on occlusion learning. OVIS consists of instance masks covering 25 categories with 607, 140 and 154 videos for train, valid and test respectively.", + "bbox": [ + 78, + 321, + 467, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BDD100K MOTS We further report results of Mask-FreeVIS on the large-scale self-driving benchmark BDD100K [64] MOTS. The dataset annotates 154 videos (30,817 images) for training, 32 videos (6,475 images) for validation, and 37 videos (7,484 images) for testing.", + "bbox": [ + 78, + 382, + 467, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 78, + 465, + 292, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed approach only requires replacing the original video mask loss in state-of-the-art VIS methods. In particular, we adopt Mask2Former [6] and SeqFormer [57] due to their excellent VIS results. Unless specified, we kept all other training schedules and settings the same as in the original methods. For the Temporal KNN-patch Loss, we set the patch size to $3 \\times 3$ , search radius to 5 and $K = 5$ . We adopt the $L_{2}$ distance as the patch matching metric and set the matching threshold to 0.05. On YTVIS 2019/2021, the Mask2Former based models are trained with AdamW [33] with learning rate $10^{-4}$ and weight decay 0.05. The learning rate decays by 10 times at with a factor of 2/3. We set batch size to 16, and train 6k/8k iterations on YTVIS 2019/2021. For experiments on OVIS and BDD100K, we adopted the COCO mask pretrained models by VITA and Unicorn. For the sampled temporal tube at training, we use 5 frames with shuffling instead of 2 frames for better temporal regularization. The compared baselines are adjusted accordingly. During testing, since there is no architecture modification, the inference of MaskFreeVIS is the same to the baselines. More details are in the Supp. file.", + "bbox": [ + 78, + 489, + 467, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Experiments", + "text_level": 1, + "bbox": [ + 78, + 811, + 279, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform detailed ablation studies for MaskFreeVIS using ResNet-50 as backbone on the YTVIS 2019 val set. We adopt the COCO box-pretrained model as initialization to eliminate all mask annotations from the training. Taking Mask2Former [6] as the base VIS method, we analyze", + "bbox": [ + 78, + 835, + 467, + 911 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7cd22fd597b5a1002c8105b63477b5b58d92f04261a6c206e3163b43489547c9.jpg", + "table_caption": [ + "Table 2. Different temporal matching schemes under the mask-free training setting on YTVIS2019 val. 'Param' indicates whether the matching scheme brings extra model parameters." + ], + "table_footnote": [], + "table_body": "
Temporal Matching SchemeParamAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Baseline (No Matching)✘38.665.938.838.447.7
Flow-based Matching✓40.266.341.940.549.1
Temporal Deformable Matching✓39.665.940.139.948.6
Learnable Pixel Matching✓39.565.740.239.748.4
Learnable Patch Matching✓40.666.542.640.349.2
3D Pairwise Matching✘39.465.041.740.248.0
Temporal KNN-Matching (Ours)✘42.566.845.741.251.2
", + "bbox": [ + 503, + 125, + 888, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the impact of individual proposed components. Moreover, we study several alternative solutions for temporal matching and influence of different hyper-parameters to TK-Loss.", + "bbox": [ + 501, + 244, + 890, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on Temporal Matching Schemes Table 2 compares our Temporal KNN-Matching to four alternative frame-wise matching approaches for enforcing temporal mask consistency. Flow-based Matching employs the pretrained optical flow model RAFT [45] to build pixel correspondence [32]. Temporal Deformable Matching adopts the temporal deformable kernels [47] to predict the pixel offsets between the target and alignment frame. Instead of using raw patches, Learnable Pixel/Patch Matching employs jointly learned deep pixel/patch embeddings via three learnable FC layers, which are then used to compute the affinities in a soft attention-like manner. 3D Pairwise Matching directly extends $\\mathcal{L}_{\\mathrm{pair}}$ designed for spatial images to the temporal dimension, where pairwise affinity loss is computed among pixels not only in within the frame but also across multiple frames.", + "bbox": [ + 501, + 291, + 890, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, compared to Flow-based Matching with one-to-one pixel correspondence, our paramter-free Temporal KNN-Matching with one-to- $K$ improves by about 2.3 AP. The prediction of flow-based models are not reliable in case of occlusions and homogeneous regions, and are also influenced by the gap between the training dataset and realworld video data. For the above mentioned deformable and learnable matching schemes, since there are only weak bounding box labels during training, the temporal matching relation is learnt implicitly. We empirically observe the instability during training with limited improvement under the mask-free training setting. For direct generalization of $\\mathcal{L}_{\\mathrm{pair}}$ , it only leads to 0.8 mask AP performance improvement. Despite the simplicity and efficiency of the TK-loss, it significantly improves VIS performance by 3.9 mask AP. Effect of Temporal KNN-patch Loss MaskFreeVIS is trained with joint spatio-temporal losses. In Table 3, to evaluate the effectiveness of each loss component, we compare the performance of MaskFreeVIS solely under the spatial pairwise loss [50] or our proposed TK-Loss. Compared to the 2.0 mask AP improvement by the spatial pairwise loss, the TK-Loss substantially promotes the mask AP from 36.6 to 41.6, showing the advantage of our flexible one-to- $K$ patch correspondence design in leveraging temporal consistency. We show the VIS results in Figure 5 to visualize", + "bbox": [ + 501, + 534, + 890, + 910 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22862", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/102036d78dfcc4ce146408b380d953b223a742098e964317dffe77343adb668c.jpg", + "table_caption": [ + "Table 3. Effect of the Spatial Pairwise loss and our Temporal KNN-patch Loss on YTVIS2019 val." + ], + "table_footnote": [], + "table_body": "
Box ProjPairwiseTK-LossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
✓36.666.536.237.145.0
✓✓\\( 38.6_{\\uparrow 2.0} \\)65.938.838.447.7
✓✓\\( 41.6_{\\uparrow 5.0} \\)68.443.540.150.5
✓✓✓\\( 42.5_{\\uparrow 5.9} \\)66.845.741.251.2
", + "bbox": [ + 81, + 106, + 465, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1da78a330a6eb6ad40d0ad6f8fd67a0e85a796eb3428f7b409c9e846a25d2f20.jpg", + "table_caption": [ + "Table 4. Effect of using max K matches on YTVIS2019 val." + ], + "table_footnote": [], + "table_body": "
KAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
140.865.844.140.3
341.966.945.141.9
542.566.845.741.2
742.367.144.640.6
", + "bbox": [ + 81, + 204, + 261, + 271 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fa44ef0ed8c9c9c89dd1f364904cbc235755111f2860fcdea09ddef27ad47207.jpg", + "table_caption": [ + "Table 6. Impact of search radius R on YTVIS2019 val." + ], + "table_footnote": [], + "table_body": "
RAPAP50AP75AR1
139.665.740.239.7
341.366.643.040.3
542.566.845.741.2
742.367.144.640.6
", + "bbox": [ + 81, + 309, + 261, + 377 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c1567b240016a425c683e3945588765fd013dfbee9051acbb3160dc11cc7edb.jpg", + "table_caption": [ + "Table 5. Patch Matching metrics comparison. NCC is Norm. Cross-Correlation." + ], + "table_footnote": [], + "table_body": "
MetricAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
NCC41.766.743.441.4
\\( L_1 \\)41.266.243.640.3
\\( L_2 \\)42.566.845.741.2
", + "bbox": [ + 279, + 218, + 460, + 273 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/73a06ea5b38111c0e07eec04eabdca8514c9bdf45f67c14a904f62170e2a3b8a.jpg", + "table_caption": [ + "Table 7. Influence of patch size N on YTVIS2019 val." + ], + "table_footnote": [], + "table_body": "
NAPAP50AP75AR1
140.165.242.240.0
342.566.845.741.2
542.168.744.442.1
741.566.344.841.3
", + "bbox": [ + 279, + 309, + 460, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the effectiveness of each loss component.", + "bbox": [ + 76, + 386, + 351, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Temporal KNN-patch Loss In Table 4, we study the influence of $\\mathbf{K}$ , the maximum number of matches selected in Temporal KNN-patch Loss. The best result is obtained for $K = 5$ , while $K = 1$ only allows for the one-to-one and no-match cases. The improvement from 40.8 mask AP to 42.5 mask AP reveals the benefit brought by the flexible one-to-many correspondence design. We also analyze the matching metric in Table 5, search radius in Table 6, and patch size in Table 7 (see Sec. 3.1.2 for details). When patch size $N$ is increased from 1 to 3 in Table 7, the performance of MaskFreeVIS is improved by 2.4 AP, validating the importance of patch structure in robust matching. Effect of the Cyclic Tube Connection We compare three frame-wise tube connection schemes (Figure 4) for the TK-Loss in Table 8. While dense connection brings forth the best performance, it doubles the training memory with minor improvement compared to Cyclic connection. Comparing to Sequential connection, our Cyclic connection benefits from long-range consistency, improving the performance of 0.6 mask AP with an affordable memory cost growth.", + "bbox": [ + 75, + 401, + 468, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Comparison of the tube connection schemes, illustrated in Fig. 4. The tube consist of 5 frames. 'Mem' denotes the memory consumption per sampled video by the TK-Loss during training.", + "bbox": [ + 75, + 705, + 468, + 747 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b07ac3667d1e0b2bf64c787e778f723e15f04218b7f2fcdd8c14ee055e691d76.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Tube ConnectConnect Num.Mem (MB)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
Dense10152642.768.044.341.5
Sequential463141.966.544.741.2
Cyclic (Ours)577342.566.845.741.2
", + "bbox": [ + 81, + 750, + 465, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on Sequence Matching Functions Besides TK-Loss, we analyze the influence of sequence matching cost functions for transformer-based VIS methods under the mask-free setting. In Table 9, we identify the substantial advantage of Spatio-temporal box-mask matching over frameworkwise cost averaging [55, 57]. As discussed in Sec. 3.2.2, we", + "bbox": [ + 75, + 821, + 468, + 912 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43bdb206943dc1d9df06ea1b12062d697ca0ec1a145beb2c8409267a3f09bfbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 75, + 632, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4fbc8d99e6d880551ba0ed3394635374723593ad0f2d7dd968f15ad2959c850.jpg", + "image_caption": [ + "Mask2Fomer + BoxInst (Baseline)" + ], + "image_footnote": [], + "bbox": [ + 504, + 160, + 632, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7655795ebbce2315e34dabc4215aba1261f7029ae455aefd099e24f42776c1f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 75, + 759, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/225a7b23ffb2341280d298fd5411a9d9e393c666f883ac2ba332283facf3da24.jpg", + "image_caption": [ + "Mask2Fomer + TK-Loss (Ours)" + ], + "image_footnote": [], + "bbox": [ + 632, + 160, + 759, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0076021faf8322db5ee9e859464a1c2998f65a9528ecd1bdf7c9062a7600c84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 75, + 888, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa94a76cb5bf4e5b7e56ac8a3fcf338204a41db130dafd40a61e6958f4089a65.jpg", + "image_caption": [ + "Mask2Former (Oracle)" + ], + "image_footnote": [], + "bbox": [ + 761, + 160, + 888, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78f10297b12cc2e0186191abe0d636b481f6a79e613d2cad1a89296e4b5e03f2.jpg", + "image_caption": [ + "Figure 5. Qualitative results comparison between using Spatial Pairwise loss [50], our TK-Loss, and Mask2Former (oracle) trained with GT video and image masks.", + "Figure 6. Results on YTVIS 2019 val with various percentages of the YTVIS training data. Baseline denotes Mask2Former [6] trained with GT video boxes using BoxInst [50], while Oracle denotes Mask2Former trained with GT video masks." + ], + "image_footnote": [], + "bbox": [ + 503, + 297, + 683, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8f096fdf99645268ff2e1f6c2356be87722b38c12a81eea94d9428d3442e35d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 297, + 885, + 393 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ad7d8d2707745c26ce8dbdbb268b9de35d00b7959b49fa3c8bc75a93b20c681.jpg", + "table_caption": [ + "Table 9. Comparison of Set Matching Cost Functions on YTVIS2019 val. ST-BoxMask denotes our Spatio-temporal Box Mask matching. w/o CE denotes removing cross-entropy cost." + ], + "table_footnote": [], + "table_body": "
Matching Cost FunctionAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Frame-wise Averaging [55, 57]37.664.239.537.545.7
ST-BoxMask40.867.842.240.049.2
ST-BoxMask w/o CE42.566.845.741.251.2
", + "bbox": [ + 511, + 500, + 877, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieve further gain by removing the object size imbalanced cross-entropy cost computation.", + "bbox": [ + 498, + 559, + 890, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training on Various Amounts of Data To further study label-efficient VIS, we validate the effect of MaskFreeVIS under various percentages of the YTVIS 2019 training data. We uniformly sample frames and their labels for each video, and set the minimum sampled number of frames to 1. Figure 6 presents the experimental results, which shows the consistent large improvement (over 3.0 AP) brought by our TK-Loss under various amounts of data. In particular, we note that our approach with $50\\%$ data even outperforms the fully-supervised model with $10\\%$ training data.", + "bbox": [ + 496, + 590, + 890, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 500, + 751, + 862, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare MaskFreeVIS with the state-of-the-art fully/weakly supervised methods on benchmarks YTVIS 2019/2021, OVIS and BDD100K MOTS. We integrate MaskFreeVIS on four representative methods [6,12,57,60], attaining consistent large gains over the strong baselines.", + "bbox": [ + 496, + 775, + 890, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "YTVIS 2019/2021 Table 10 compares the performance on YTVIS 2019. Using R50/R101 as backbone and with the same training setting, MaskFreeVIS achieves 42.5/45.8 AP, improving 3.9/5.0 AP over the strong baseline adopt", + "bbox": [ + 496, + 851, + 890, + 912 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22863", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3b15d68d7330d45d7c9c5fd4c13058bfe96009d26b453a16c3280fffbb5f661c.jpg", + "table_caption": [ + "Table 10. Comparison on YTVIS 2019. I: using COCO mask pretrained model as initialization. V: using YTVIS video masks during training. $\\ast$ : using pseudo mask from COCO images for joint training [57]. M2F: Mask2Former [6], SeqF: SeqFormer [57]." + ], + "table_footnote": [], + "table_body": "
MethodMask ann.Back-boneAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)
Fully-supervised:
PCAN [18]I+VR5036.154.939.436.341.6
EfficientVIS [59]I+VR5037.959.743.040.346.6
InsPro [10]I+VR5040.262.943.137.644.5
IFC [15]I+VR5042.865.846.843.851.2
VMT* [17]I+VR5047.9-52.045.8-
SeqF* [57]I+VR5047.469.851.845.554.8
M2FI+VR5046.468.050.0--
M2F*I+VR5047.869.252.746.256.6
Prev. Weakly-supervised:
FlowIRN [32]-R5010.527.26.212.313.6
SOLO-Track [9]IR5030.650.733.531.637.1
Mask-free:
M2F + Flow Consist [45]-R5040.266.341.940.549.1
M2F + BoxInst [50]-R5038.664.238.538.046.8
M2F + MaskFreeVIS-R5042.5↑3.966.845.741.251.2
M2F + MaskFreeVISIR5043.8↑5.270.746.941.552.3
M2F + MaskFreeVIS*IR5046.6↑8.072.549.744.955.7
Fully-supervised:
M2FVR10145.672.648.944.354.5
M2FI+VR10149.272.854.2--
M2F*I+VR10149.873.655.448.058.0
SeqF*I+VR10149.071.155.746.856.9
Mask-free:
M2F + BoxInst [50]-R10140.867.842.240.049.2
M2F + MaskFreeVIS-R10145.8↑5.070.848.645.355.2
M2F + MaskFreeVISIR10147.3↑6.575.449.944.655.2
M2F + MaskFreeVIS*IR10148.9↑8.174.954.744.957.0
SeqF + MaskFreeVIS*IR10148.674.052.245.957.2
Fully-supervised:
M2FI+VSwinL60.484.467.0--
Mask-free:
M2F + BoxInst [50]-SwinL49.873.255.548.258.1
M2F + MaskFreeVIS-SwinL54.3↑4.582.661.150.261.3
M2F + MaskFreeVIS*ISwinL55.3↑5.582.560.850.762.2
", + "bbox": [ + 76, + 122, + 470, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing BoxInst [50] losses. MaskFreeVIS, without any mask labels, even significantly outperforms some recent fully-supervised methods such as EfficientVIS [59] and In-sPro [10]. On R50/R101/Swin-L, our MaskFreeVIS consistently attains over $91\\%$ of its fully-supervised counterpart trained with both GT image and video masks. We also observe similar larger performance growth over the baseline on YTVIS 2021 in Table 11. The excellent performance substantially narrows the performance gap between fully-supervised and weakly-supervised VIS.", + "bbox": [ + 75, + 604, + 468, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "OVIS We also conduct experiments on OVIS in Table 12 using R50 as backbone. We integrate MaskFreeVIS with VITA [12], promoting the baseline performance from 12.1 to 15.7 under the mask-free training setting.", + "bbox": [ + 75, + 755, + 468, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BDD100K MOTS Table 13 further validates our approach on BDD100K MOTS. Integrated with Unicorn [60], Mask-FreeVIS achieves $23.8\\mathrm{mMOTSA}$ by improving over 4.9 points over the strong baseline and thus surpassing the fully", + "bbox": [ + 75, + 816, + 468, + 878 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a459e4072641e57c49f7968446fa68fc13b74252cf870e151650e6090f2f2b92.jpg", + "table_caption": [ + "Table 11. Comparison on YTVIS 2021. Refer to Table 10 for the symbol abbreviations." + ], + "table_footnote": [], + "table_body": "
MethodMask ann.Back-boneAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Fully-supervised:
MaskTrack [61]I+VR5028.648.929.626.533.8
IFC [15]I+VR5036.657.939.3--
SeqF* [57]I+VR5040.562.443.736.148.1
M2FI+VR5040.660.941.8--
Mask-free:
M2F + BoxInst [50]-R5032.152.834.431.038.1
M2F + MaskFreeVIS-R5036.2↑4.160.839.234.645.6
M2F + MaskFreeVISIR5037.2↑5.161.940.335.346.1
M2F + MaskFreeVIS*IR5040.9↑8.865.843.337.150.5
Fully-supervised:
M2FI+VR10142.465.945.8--
Mask-free:
M2F + BoxInst [50]-R10133.355.232.532.141.9
M2F + MaskFreeVIS-R10137.3↑4.061.639.434.145.6
M2F + MaskFreeVISIR10138.2↑4.962.440.034.946.2
M2F + MaskFreeVIS*IR10141.6↑8.366.244.836.349.2
", + "bbox": [ + 501, + 93, + 893, + 344 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bd13df803b2e22a6609463978b1445a4a69c5ed001247150106484e3d3f5f160.jpg", + "table_caption": [ + "Table 12. State-of-the-art comparison on the OVIS using R50." + ], + "table_footnote": [], + "table_body": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Fully-supervised:
CrossVIS [62]14.932.712.110.319.8
Mask2Former [6]17.337.315.110.523.5
VMT [17]16.936.413.710.422.7
VITA [12]19.641.217.411.726.0
Video Mask-free:
VITA [12] + BoxInst [50]12.128.310.28.817.9
VITA [12] + MaskFreeVIS\\( 15.7_{\\uparrow 3,6} \\)35.113.110.120.4
", + "bbox": [ + 501, + 367, + 885, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b98f1efffece4f0ef723115a55989e1566855e8cf44f302afaf3381a566eee29.jpg", + "table_caption": [ + "Table 13. State-of-the-art comparison on the BDD100K segmentation tracking validation set." + ], + "table_footnote": [], + "table_body": "
MethodmMOTSA↑mMOTSP↑mIDF↑ID sw.↓mAP↑
Fully-supervised:
STEm-Seg [1]12.258.225.4873221.8
QDTrack-mots-fix [36]23.566.344.597325.5
PCAN [18]27.466.745.187626.6
Unicorn [60]29.667.744.2173132.1
Video Mask-free:
Unicorn [60] + BoxInst [50]18.958.736.3329822.1
Unicorn [60] + MaskFreeVIS23.8↑4.966.744.9208624.8
", + "bbox": [ + 501, + 534, + 893, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "supervised QDTrack-mots [36]. The consistent large gains on four benchmarks and four base VIS methods validate the generalizability of our MaskFreeVIS.", + "bbox": [ + 498, + 659, + 890, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 719, + 619, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MaskFreeVIS is the first competitive VIS method that does not need any mask annotations during training. The strong results lead to a remarkable conclusion: mask labels are not a necessity for high-performing VIS. Our key component is the unsupervised Temporal KNN-patch Loss, which replaces the conventional video masks losses by leveraging temporal mask consistency constraints. Our approach greatly reduces the long-standing gap between fully-supervised and weakly-supervised VIS on four large-scale benchmarks. MaskFreeVIS thus opens up many opportunities for researchers and practitioners for label-efficient VIS.", + "bbox": [ + 496, + 744, + 890, + 912 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1This work is supported in part by Research Grant Council of the Hong Kong SAR under grant no. 16201420 and Huawei Technologies.", + "bbox": [ + 76, + 885, + 468, + 912 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22864", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 80, + 174, + 97 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, 2020. 3, 8", + "[2] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, 2020. 2", + "[3] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, 2019. 2", + "[4] Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In ECCV, 2020. 2", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2", + "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 2, 3, 5, 6, 7, 8", + "[7] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In CVPR, 2022. 1, 3", + "[8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In ICCV, 2015. 3", + "[9] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to track instances without video annotations. In CVPR, 2021. 3, 8", + "[10] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. In NeurIPS, 2022. 2, 8", + "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2", + "[12] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 7, 8", + "[13] Cheng-Chun Hsu, Kuang-Jui Hsu, Chung-Chi Tsai, Yen-Yu Lin, and Yung-Yu Chuang. Weakly supervised instance segmentation using the bounding box tightness prior. In NeurIPS, 2019. 1, 3, 5", + "[14] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 3", + "[15] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. In NeurIPS, 2021. 2, 8", + "[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In CVPR, 2022. 3", + "[17] Lei Ke, Henghui Ding, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Video mask transfiner for high-quality video instance segmentation. In ECCV, 2022. 1, 2, 3, 8", + "[18] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks" + ], + "bbox": [ + 78, + 104, + 468, + 911 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for multiple object tracking and segmentation. In NeurIPS, 2021. 2, 3, 8", + "[19] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2", + "[20] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In CVPR, 2017. 3", + "[21] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In ICLR, 2022. 3", + "[22] Viveka Kulharia, Siddhartha Chandra, Amit Agrawal, Philip Torr, and Ambrish Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In ECCV, 2020. 3", + "[23] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In ICCV, 2021. 1, 3", + "[24] Jungbeom Lee, Eunji Kim, Sungmin Lee, Jangho Lee, and Sungroh Yoon. Frame-to-frame aggregation of active regions in web videos for weakly supervised semantic segmentation. In ICCV, 2019. 3", + "[25] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In CVPR, 2021. 3", + "[26] Minghan Li, Shuai Li, Lida Li, and Lei Zhang. Spatial feature calibration and temporal fusion for effective one-stage video instance segmentation. In CVPR, 2021. 2", + "[27] Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, XianSheng Hua, and Lei Zhang. Box-supervised instance segmentation with level set evolution. In ECCV, 2022. 1, 3", + "[28] Chung-Ching Lin, Ying Hung, Rogerio Feris, and Linglin He. Video instance segmentation tracking with a modified vae architecture. In CVPR, 2020. 2", + "[29] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In CVPR, 2021. 2", + "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2", + "[31] Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In CVPR, 2021. 2", + "[32] Qing Liu, Vignesh Ramanathan, Dhruv Mahajan, Alan Yuille, and Zhenheng Yang. Weakly supervised instance segmentation for videos with temporal mask consistency. In CVPR, 2021. 2, 3, 6, 8", + "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6", + "[34] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 3", + "[35] Anton Milan, Laura Leal-Taixe, Konrad Schindler, and Ian Reid. Joint tracking and segmentation of multiple targets. In CVPR, 2015. 3" + ], + "bbox": [ + 501, + 83, + 890, + 910 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22865", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 8", + "[37] George Papandreou, Liang-Chieh Chen, Kevin P Murphy, and Alan L Yuille. Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In ICCV, 2015. 3", + "[38] Jordi Pont-Tuset, Pablo Arbelaez, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping for image segmentation and object proposal generation. IEEE TPAMI, 39(1):128-140, 2016. 3", + "[39] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 2021. 1, 2, 6", + "[40] Martin Rajchl, Matthew CH Lee, Ozan Oktay, Konstantinos Kamnitsas, Jonathan Passerat-Palmbach, Wenjia Bai, Mellisa Damodaram, Mary A Rutherford, Joseph V Hajnal, Bernhard Kainz, et al. Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging, 36(2):674-683, 2016. 3", + "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 5", + "[42] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. Grabcut: interactive foreground extraction using iterated graph cuts. ACM TOG, 23(3):309-314, 2004. 3", + "[43] Fatemeh Sadat Saleh, Mohammad Sadegh Aliakbarian, Mathieu Salzmann, Lars Petersson, and Jose M Alvarez. Bringing background into the foreground: Making all classes equal in weakly-supervised video semantic segmentation. In ICCV, 2017. 3", + "[44] Chunfeng Song, Yan Huang, Wanli Ouyang, and Liang Wang. Box-driven class-wise region masking and filling rate guided loss for weakly supervised semantic segmentation. In CVPR, 2019. 3", + "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 6, 8", + "[46] Omkar Thawakar, Sanath Narayan, Jiale Cao, Hisham Cholakkal, Rao Muhammad Anwer, Muhammad Haris Khan, Salman Khan, Michael Felsberg, and Fahad Shahbaz Khan. Video instance segmentation via multi-scale spatiotemporal split attention transformer. In ECCV, 2022. 2", + "[47] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. 6", + "[48] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, 2020. 2, 3, 5", + "[49] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 2", + "[50] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In CVPR, 2021. 1, 3, 5, 6, 7, 8", + "[51] Pavel Tokmakov, Karteek Alahari, and Cordelia Schmid. Learning semantic segmentation with weakly-annotated videos. In ECCV, 2016. 3" + ], + "bbox": [ + 78, + 83, + 468, + 910 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yi-Hsuan Tsai, Guangyu Zhong, and Ming-Hsuan Yang. Semantic co-segmentation in videos. In ECCV, 2016. 3", + "[53] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 3", + "[54] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2", + "[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 5, 7", + "[56] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In CVPR, 2021. 3", + "[57] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. In ECCV, 2022. 1, 2, 3, 5, 6, 7, 8", + "[58] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 3", + "[59] Jialian Wu, Sudhir Yarram, Hui Liang, Tian Lan, Junsong Yuan, Jayan Eledath, and Gerard Medioni. Efficient video instance segmentation via tracklet query and proposal. In CVPR, 2022. 2, 8", + "[60] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 3, 7, 8", + "[61] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 2, 5, 6, 8", + "[62] Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In ICCV, 2021. 2, 8", + "[63] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In CVPR, 2022. 2", + "[64] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 1, 2, 6" + ], + "bbox": [ + 503, + 83, + 890, + 705 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22866", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_model.json b/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f95a7a1460be5aa567145b4e4cc4363cba130fa --- /dev/null +++ b/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_model.json @@ -0,0 +1,2596 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.279, + 0.122, + 0.692, + 0.143 + ], + "angle": 0, + "content": "Mask-Free Video Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.171, + 0.858, + 0.207 + ], + "angle": 0, + "content": "Lei Ke\\(^{1,2}\\) Martin Danelljan\\(^{1}\\) Henghui Ding\\(^{1}\\) Yu-Wing Tai\\(^{2}\\) Chi-Keung Tang\\(^{2}\\) Fisher Yu\\(^{1}\\) ETH Zürich \\({}^{2}\\mathrm{HKUST}\\)" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.235, + 0.889, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.408, + 0.893, + 0.451 + ], + "angle": 0, + "content": "Figure 1. Video instance segmentation (VIS) results of our MaskFreeVIS, trained without using any video or image mask annotation. By achieving a remarkable \\(42.5\\%\\) mask AP on the YouTube-VIS val dataset, with a ResNet-50 backbone, our approach demonstrates that high-performing VIS can be learned even without any mask annotations." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.47, + 0.314, + 0.485 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.473, + 0.896 + ], + "angle": 0, + "content": "The recent advancement in Video Instance Segmentation (VIS) has largely been driven by the use of deeper and increasingly data-hungry transformer-based models. However, video masks are tedious and expensive to annotate, limiting the scale and diversity of existing VIS datasets. In this work, we aim to remove the mask-annotation requirement. We propose MaskFreeVIS, achieving highly competitive VIS performance, while only using bounding box annotations for the object state. We leverage the rich temporal mask consistency constraints in videos by introducing the Temporal KNN-patch Loss (TK-Loss), providing strong mask supervision without any labels. Our TK-Loss finds one-to-many matches across frames, through an efficient patch-matching step followed by a \\(K\\)-nearest neighbor selection. A consistency loss is then enforced on the found matches. Our mask-free objective is simple to implement, has no trainable parameters, is computationally efficient, yet outperforms baselines employing, e.g., state-of-the-art optical flow to enforce temporal mask consistency. We validate MaskFreeVIS on the YouTube-VIS 2019/2021, OVIS and BDD100K MOTS benchmarks. The results clearly demonstrate the efficacy of our method by drastically narrowing the gap between fully and weakly-supervised VIS performance. Our code and trained models are available at http://vis.xyz/pub/maskfreevis." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.47, + 0.631, + 0.485 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.727 + ], + "angle": 0, + "content": "Video Instance Segmentation (VIS) requires jointly detecting, tracking and segmenting all objects in a video from a given set of categories. To perform this challenging task, state-of-the-art VIS models are trained with complete video annotations from VIS datasets [39, 61, 64]. However, video annotation is costly, in particular regarding object mask labels. Even coarse polygon-based mask annotation is multiple times slower than annotating video bounding boxes [7]. Expensive mask annotation makes existing VIS benchmarks difficult to scale, limiting the number of object categories covered. This is particularly a problem for the recent transformer-based VIS models [6, 17, 57], which tend to be exceptionally data-hungry. We therefore revisit the need for complete mask annotation by studying the problem of weakly supervised VIS under the mask-free setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.896, + 0.913 + ], + "angle": 0, + "content": "While there exist box-supervised instance segmentation models [13, 23, 27, 50], they are designed for images. These weakly-supervised single-image methods do not utilize temporal cues when learning mask prediction, leading to lower accuracy when directly applied to videos. As a source for weakly supervised learning, videos contain much richer information about the scene. In particular, videos adhere to the temporal mask consistency constraint, where the regions corresponding to the same underlying object across different frames should have the same mask label. In this work, we set out to leverage this important constraint for mask-free learning of VIS." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.951, + 0.519, + 0.963 + ], + "angle": 0, + "content": "22857" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.073, + 0.385, + 0.083 + ], + "angle": 0, + "content": "Temporal One-to-K Patch Correspondence" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.083, + 0.468, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Figure 2. Our Temporal KNN-patch Loss enforces mask consistency between one-to-\\(k\\) patch correspondences found across frames, which allow us to cover the cases where: (i) A unique one-to-one match exists (blue); (ii) Multiple matches are found due to ambiguities in homogenous regions (orange) or along image edges (white and yellow); (iii) No match is found due to \\(e.g.\\) occlusions (green). This allows us to robustly leverage mask consistency constraints in challenging videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.47, + 0.62 + ], + "angle": 0, + "content": "We propose MaskFreeVIS method, for high performance VIS without any mask annotations. To leverage temporal mask consistency, we introduce the Temporal KNN-patch Loss (TK-Loss), as in Figure 2. To find regions corresponding to the same underlying video object, our TK-Loss first builds correspondences across frames by patch-wise matching. For each target patch, only the top \\(K\\) matches in the neighboring frame with high enough matching score are selected. A temporal consistency loss is then applied to all found matches to promote the mask consistency. Specifically, our surrogate objective function not only promotes the one-to- \\(k\\) matched regions to reach the same mask probabilities, but also commits their mask prediction to a confident foreground or background prediction by entropy minimization. Unlike flow-based models [32,45], which assume one-to-one matching, our approach builds robust and flexible one-to- \\(k\\) correspondences to cope with e.g. occlusions and homogeneous regions, without introducing additional model parameters or inference cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.744 + ], + "angle": 0, + "content": "The TK-Loss is easily integrated into existing VIS methods, with no architecture modifications required. During training, our TK-Loss simply replaces the conventional video mask losses in supervising video mask generation. To further enforce temporal consistency through the video clip, TK-Loss is employed in a cyclic manner instead of using dense frame-wise connections. This greatly reduces memory cost with negligible performance drop." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.471, + 0.913 + ], + "angle": 0, + "content": "We extensively evaluate our MaskFreeVIS on four large-scale VIS benchmarks, i.e., YouTube-VIS 2019/2021 [61], OVIS [39], and BDD100K MOTS [64]. MaskFreeVIS achieves competitive VIS performance without using any video masks or even image mask labels on all datasets. Validated on various methods and backbones, MaskFreeVIS achieves \\(91.25\\%\\) performance of its fully supervised counterparts, even outperforming a few recent fully-supervised methods [10, 15, 18, 59] on the popular YTVIS benchmark. Our simple yet effective design greatly narrows the performance gap between weakly-supervised and fully" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.08, + 0.892, + 0.193 + ], + "angle": 0, + "content": "Table 1. Mask annotation requirement for state-of-the-art VIS methods. Results are reported using ResNet-50 as backbone on the YTVIS 2019 [61] benchmark. Video Mask: using YTVIS video mask labels. Image Mask: using COCO [30] image mask labels for image-based pretraining. Pseudo Video: using Pseudo Videos from COCO images for joint training [57]. MaskFreeVIS achieves \\(91.5\\%\\) (42.5 vs. 46.4) of its fully-supervised baseline performance (Mask2Former) without using any masks during training." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.194, + 0.89, + 0.32 + ], + "angle": 0, + "content": "
MethodVideo MaskImage MaskPseudo VideoAP
SeqFormer [57]✓✓✓47.4
VMT [17]✓✓✓47.9
Mask2Former [6]✓✓✓47.8
MaskFreeVIS (ours)✗✓✓46.6
Mask2Former [6]✓✓✗46.4
MaskFreeVIS (ours)✗✗✗42.5
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.372 + ], + "angle": 0, + "content": "supervised video instance segmentation. It further demonstrates that expensive video masks, or even image masks, is not necessary for training high-performing VIS models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.893, + 0.628 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: (i) To utilize temporal information, we develop a new parameter-free Temporal KNN-patch Loss, which leverages temporal masks consistency using unsupervised one-to-\\(k\\) patch correspondence. We extensively analyze the TK-Loss through ablative experiments. (ii) Based on the TK-Loss, we develop the MaskFreeVIS method, enabling training existing state-of-the-art VIS models without any mask annotation. (iii) To the best of our knowledge, MaskFreeVIS is the first mask-free VIS method attaining high-performing segmentation results. We provide qualitative results in Figure 1. As in Table 1, when integrated into the Mask2Former [6] baseline with ResNet-50, our MaskFreeVIS achieves \\(42.5\\%\\) mask AP on the challenging YTVIS 2019 benchmark while using no video or image mask annotations. Our approach further scales to larger backbones, achieving \\(55.3\\%\\) mask AP on Swin-L backbone with no video mask annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.69 + ], + "angle": 0, + "content": "We hope our approach will facilitate achieving label-efficient video instance segmentation, enabling building even larger-scale VIS benchmarks with diverse categories by lifting the mask annotation restriction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.642, + 0.721 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.913 + ], + "angle": 0, + "content": "Video Instance Segmentation (VIS) Existing VIS methods can be summarized into three categories: two-stage, one-stage, and transformer-based. Two-stage approaches [2, 18, 28, 29, 61] extend the Mask R-CNN family [11, 19] by designing an additional tracking branch for object association. One-stage works [4, 26, 31, 62] adopt anchor-free detectors [49], generally using linear masks basis combination [3] or conditional mask prediction generation [48]. For the transformer-based models [6, 12, 46, 57, 63], VisTr [54] firstly adapts the transformer [5] for VIS, and IFC [15] further improves its efficiency via memory tokens. Seqformer [57] proposes frame query decompo" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.951, + 0.519, + 0.963 + ], + "angle": 0, + "content": "22858" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.083, + 0.468, + 0.188 + ], + "angle": 0, + "content": "sition while Mask2Former [6] includes masked attention. VMT [17] extends Mask Transfiner [16] to video for high-quality VIS, and IDOL [58] focuses on online VIS. State-of-the-art VIS methods with growing capacity put limited emphasis on weak supervision. In contrast, the proposed MaskFreeVIS is the first method targeting mask-free VIS while attaining competitive performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.468, + 0.339 + ], + "angle": 0, + "content": "Multiple Object Tracking and Segmentation (MOTS) Most MOTS methods [1,34,35,53,56] follow the tracking-by-detection principle. PCAN [18] improves temporal segmentation by utilizing space-time memory prototypes, while the one-stage method Unicorn [60] focuses on unification of different tracking frameworks. Compared to the aforementioned fully-supervised MOTS methods, Mask-FreeVIS focuses on label efficient training without GT masks by proposing a new surrogate temporal loss which can be easily integrated on them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.468, + 0.476 + ], + "angle": 0, + "content": "Mask-Free VIS Most mask-free instance segmentation works [7, 13, 20, 25, 27, 37, 40, 44] are designed for single images and thus neglect temporal information. Earlier works BoxSup [8] and Box2Seg [22] rely on region proposals produced by MCG [38] or GrabCut [42], leading to slow training. BoxInst [50] proposes the surrogate projection and pixel pairwise losses to replace the original mask learning loss of CondInst [48], while DiscoBox [23] focuses on generating pseudo mask labels guided by a teacher model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Earlier works have investigated the use of videos for weakly-, semi-, or un-supervised segmentation by leveraging motion or temporal consistency [21,51,52]. Most aforementioned approaches do not address the VIS problem, and use optical flow for frame-to-frame matching [24, 32, 43]. In particular, FlowIRN [32] explores VIS using only classification labels and incorporates optical flow to leverage mask consistency. The limited performance makes the class-label only or fully-unsupervised setting difficult to deploy in the real world. SOLO-Track [9] aims to train VIS models without video annotations, and one concurrent work MinVIS [14] performs VIS without video-based model architectures. Unlike the above weakly-supervised training settings, our MaskFreeVIS is designed for eliminating the mask annotation requirement for VIS, as we note that video mask labeling is particularly expensive. MaskFreeVIS enables training VIS models without any video masks, or even image masks. Despite its simplicity, MaskFreeVIS drastically reduces the gap between fully-supervised and weakly-supervised VIS models, making weakly-supervised models more accessible in practice." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.811, + 0.168, + 0.825 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.837, + 0.468, + 0.912 + ], + "angle": 0, + "content": "We propose MaskFreeVIS to tackle video instance segmentation (VIS) without using any video or even image mask labels. Our approach is generic and can be directly applied to train state-of-the-art VIS methods, such as Mask2Former [6] and SeqFormer [57]. In Sec. 3.1, we first" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.083, + 0.892, + 0.218 + ], + "angle": 0, + "content": "present the core component of MaskFreeVIS: the Temporal KNN-patch Loss (TK-Loss), which leverages temporal consistency to supervise accurate mask prediction, without any human mask annotations. In Sec. 3.2, we then describe how to integrate the TK-Loss with existing spatial weak segmentation losses for transformer-based VIS methods, to achieve mask-free training of VIS approaches. Finally, we introduce image-based pretraining details of our MaskFreeVIS in Sec. 3.2.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.644, + 0.243 + ], + "angle": 0, + "content": "3.1. MaskFreeVIS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.311 + ], + "angle": 0, + "content": "In this section, we introduce the Temporal KNN-patch Loss, illustrated in Figure 3. It serves as an unsupervised objective for mask prediction that leverages the rich spatiotemporal consistency constraints in unlabelled videos." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.33, + 0.753, + 0.345 + ], + "angle": 0, + "content": "3.1.1 Temporal Mask Consistency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.459 + ], + "angle": 0, + "content": "While an image constitutes a single snapshot of a scene, a video provides multiple snapshots displaced in time. Thereby, a video depicts continuous change in the scene. Objects and background move, deform, are occluded, experience variations in lighting, motion blur, and noise, leading to a sequence of different images that are closely related through gradual transformations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Consider a small region in the scene (Fig. 2), belonging either to an object or background. The pixels corresponding to the projection of this region should have the same mask prediction in every frame, as they belong to the same underlying physical object or background region. However, the aforementioned dynamic changes in the video lead to substantial appearance variations, serving as a natural form of data augmentation. The fact that the pixels corresponding to the same underlying object region should have the same mask prediction under temporal change therefore provides a powerful constraint, i.e., temporal mask consistency, which can be used for mask supervision [21, 24, 32, 51, 52]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.791 + ], + "angle": 0, + "content": "The difficulty in leveraging the temporal mask consistency constraint stems from the problem of establishing reliable correspondences between video frames. Objects often undergo fast motion, deformations, etc., resulting in substantial appearance change. Furthermore, regions in the scene may be occluded or move out of the image from one frame to the other. In such cases, no correspondence exist. Lastly, videos are often dominated by homogenous regions, such as sky and ground, where the establishment of one-to-one correspondences are error-prone or even ill-defined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.892, + 0.912 + ], + "angle": 0, + "content": "The problem of establishing dense one-to-one correspondences between subsequent video frames, known as optical flow, is a long-standing and popular research topic. However, when attempting to enforce temporal mask consistency through optical flow estimation [24,32,43], one encounters two key problems. 1) The one-to-one assumption of optical flow is not suitable in cases of occlusions, homogenous regions, and along single edges, where the corre" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "22859" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.063, + 0.226, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.217, + 0.231, + 0.227 + ], + "angle": 0, + "content": "(1) Temporal Candidates Extraction" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.063, + 0.359, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.217, + 0.353, + 0.227 + ], + "angle": 0, + "content": "(2) Temporal KNN-Matching" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.063, + 0.533, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.217, + 0.446, + 0.225 + ], + "angle": 0, + "content": "Loss Backward" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.062, + 0.695, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.217, + 0.67, + 0.227 + ], + "angle": 0, + "content": "(3) KNN-Consistency Loss" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.063, + 0.762, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.214, + 0.758, + 0.227 + ], + "angle": 0, + "content": "4) Cyclic Tube Connection" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.064, + 0.89, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.217, + 0.88, + 0.226 + ], + "angle": 0, + "content": "MaskFreeVIS Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Figure 3. Temporal KNN-patch Loss has four steps: 1) Patch Candidate Extraction: Patch candidates searching across frames with radius \\( R \\). 2) Temporal KNN-Matching: Match \\( k \\) high-confidence candidates by patch affinities. 3) Consistency loss: Enforce mask consistency objective (Eq. 2) among the matches. 4) Cyclic Tube Connection: Temporal loss aggregation in the 5-frame tube, detailed in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.469, + 0.342 + ], + "angle": 0, + "content": "spondence is either nonexistent, undefined, ambiguous, uncertain, or very difficult to determine. 2) State-of-the-art optical flow estimation relies on large and complex deep networks, with large computational and memory requirements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.469, + 0.387 + ], + "angle": 0, + "content": "Instead of using optical flow, we aim to design a simple, efficient, and parameter-free approach that effectively enforces the temporal mask consistency constraint." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.391, + 0.32, + 0.407 + ], + "angle": 0, + "content": "3.1.2 Temporal KNN-patch Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Our Temporal KNN-patch Loss (TK-Loss) is based on a simple and flexible correspondence estimation across frames. In contrast to optical flow, we do not restrict our formulation to one-to-one correspondences. Instead, we establish one-to- \\(K\\) correspondences. This include the conventional one-to-one \\((K = 1)\\), where a unique well-defined match exists. However, this allows us to also handle the cases of nonexistent correspondences \\((K = 0)\\) in case of occlusions, and one-to-many \\((K \\geq 2)\\) in case of homogenous regions. In cases where multiple matches are found, these most often belong to the same underlying object or background due to their similar appearance, as in Figure 2. This further benefits our mask consistency objective through denser supervision. Lastly, our approach is simple to implement, with negligible computational overhead and no learnable parameters. Our approach is in Figure 3, and contains four main steps, which are detailed next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.822 + ], + "angle": 0, + "content": "1) Patch Candidate Extraction: Let \\( X_p^t \\) denote an \\( N \\times N \\) target image patch centered at spatial location \\( p = (x, y) \\) in frame \\( t \\). Our aim is to find a set of corresponding positions \\( S_p^{t \\to \\hat{t}} = \\{\\hat{p}_i\\}_i \\) in frame number \\( \\hat{t} \\) that represent the same object region. To this end, we first select candidate locations \\( \\hat{p} \\) within a radius \\( R \\) such that \\( \\| p - \\hat{p} \\| \\leq R \\). Such windowed patch search exploits spatial proximity across neighboring frames in order to avoid an exhaustive global search. For a fast implementation, the windowed search is performed for all target image patches \\( X_p^t \\) in parallel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.469, + 0.853 + ], + "angle": 0, + "content": "2) Temporal KNN-Matching: We match patch candidate patches through a simple distance computation," + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.863, + 0.469, + 0.888 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {p \\rightarrow \\hat {p}} ^ {t \\rightarrow \\hat {t}} = \\left\\| X _ {p} ^ {t} - X _ {\\hat {p}} ^ {\\hat {t}} \\right\\|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.897, + 0.469, + 0.913 + ], + "angle": 0, + "content": "In our ablative experiments (Sec. 4.3), we found the \\(L_{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.892, + 0.361 + ], + "angle": 0, + "content": "norm to be the most effective patch matching metric. We select the top \\(K\\) matches with smallest patch distance \\(\\mathbf{d}_{p\\rightarrow \\hat{p}}^{t\\rightarrow \\hat{t}}\\). Lastly low-confidence matches are removed by enforcing a maximal patch distance \\(D\\) as \\(\\mathbf{d}_{p\\rightarrow \\hat{p}}^{t\\rightarrow \\hat{t}} < D\\). The remaining matches form the set \\(S_{p}^{t\\rightarrow \\hat{t}} = \\{\\hat{p}_{i}\\}_{i}\\) for each location \\(p\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.456 + ], + "angle": 0, + "content": "3) Consistency loss: Let \\( M_p^t \\in [0,1] \\) denote the predicted binary instance mask of an object, evaluated at position \\( p \\) in frame \\( t \\). To ensure temporal mask consistency constraints, we penalize inconsistent mask predictions between a spatiotemporal point \\( (p,t) \\) and its estimated corresponding points in \\( S_p^{t \\to \\hat{t}} \\). In particular we use the following objective," + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.469, + 0.891, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} ^ {t \\rightarrow \\hat {t}} = \\frac {1}{H W} \\sum_ {p} \\sum_ {\\hat {p} _ {i} \\in \\mathcal {S} _ {p} ^ {t \\rightarrow \\hat {t}}} L _ {\\text {c o n s}} \\left(M _ {p} ^ {t}, M _ {\\hat {p} _ {i}} ^ {\\hat {t}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.524, + 0.764, + 0.539 + ], + "angle": 0, + "content": "where mask consistency is measured as" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.553, + 0.891, + 0.575 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o n s}} \\left(M _ {p} ^ {t}, M _ {\\hat {p}} ^ {\\hat {t}}\\right) = - \\log \\left(M _ {p} ^ {t} M _ {\\hat {p}} ^ {\\hat {t}} + \\left(1 - M _ {p} ^ {t}\\right) \\left(1 - M _ {\\hat {p}} ^ {\\hat {t}}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.681 + ], + "angle": 0, + "content": "Note that Eq. (3) only attains its minimum value of zero if both predictions indicate exactly background (\\(M_p^t = M_{\\hat{p}}^{\\hat{t}} = 0\\)) or foreground (\\(M_p^t = M_{\\hat{p}}^{\\hat{t}} = 1\\)). The objective does thus not only promote the two mask predictions to achieve the same probability value \\(M_p^t = M_{\\hat{p}}^{\\hat{t}}\\), but also to commit to a certain foreground or background prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.789 + ], + "angle": 0, + "content": "4) Cyclic Tube Connection: Suppose the temporal tube consists of \\( T \\) frames. We compute the temporal loss for the whole tube in a cyclic manner, as in Figure 4. The start frame is connected to the end frame, which introduces direct long-term mask consistency across the two temporally most distant frames. The temporal TK-Loss for the whole tube is given by" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.802, + 0.891, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e m p}} = \\sum_ {t = 1} ^ {T} \\left\\{\\begin{array}{l l}\\mathcal {L} _ {f} ^ {t \\rightarrow (t + 1)}&\\text {i f} t < T - 1\\\\\\mathcal {L} _ {f} ^ {t \\rightarrow 0}&\\text {i f} t = T - 1.\\end{array}\\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.892, + 0.911 + ], + "angle": 0, + "content": "Compared to inter-frame dense connections in Figure 4, we find the cyclic loss to achieve similar performance but greatly reduce the memory usage as validated in the experiment section." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.951, + 0.519, + 0.963 + ], + "angle": 0, + "content": "22860" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.075, + 0.462, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.226, + 0.468, + 0.255 + ], + "angle": 0, + "content": "Figure 4. Illustration of different frame-wise tube connection settings (connection number) in the temporal loss design." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.266, + 0.295, + 0.282 + ], + "angle": 0, + "content": "3.2. Training MaskFreeVIS" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.47, + 0.396 + ], + "angle": 0, + "content": "In this section, we describe how to train state-of-the-art VIS methods using our TK-Loss, without any mask annotations. Our MaskFreeVIS approach is jointly supervised with spatial-temporal surrogate losses, and is easily integrated with existing transformer-based methods. We also detail mask-free image-based pre-training for MaskFree-VIS to fully eliminate mask usage during training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.403, + 0.392, + 0.419 + ], + "angle": 0, + "content": "3.2.1 Joint Spatio-temporal Regularization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.469, + 0.487 + ], + "angle": 0, + "content": "To train MaskFreeVIS, in addition to our proposed Temporal KNN-patch Loss for temporal mask consistency, we leverage existing spatial weak segmentation losses to jointly enforce intra-frame consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.623 + ], + "angle": 0, + "content": "Spatial Consistency To explore spatial weak supervision signals from image bounding boxes and pixel color, we utilize the representative Box Projection Loss \\( L_{\\mathrm{proj}} \\) and Pairwise Loss \\( L_{\\mathrm{pair}} \\) in [50], to replace the supervised mask learning loss. The Projection Loss \\( L_{\\mathrm{proj}} \\) enforces the projection \\( P' \\) of the object mask onto the \\( \\vec{x} \\)-axis and \\( \\vec{y} \\)-axis of image to be consistent with its ground-truth box mask. For the temporal tube with \\( T \\) frames, we concurrently optimize all predicted frame masks of the tube as," + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.63, + 0.469, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p r o j}} = \\sum_ {t = 1} ^ {T} \\sum_ {d \\in \\{\\vec {x}, \\vec {y} \\}} D \\left(P _ {d} ^ {\\prime} \\left(M _ {p} ^ {t}\\right), P _ {d} ^ {\\prime} \\left(M _ {b} ^ {t}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.741 + ], + "angle": 0, + "content": "where \\(D\\) denotes dice loss, \\(P'\\) is the projection function along \\(\\vec{x} / \\vec{y}\\)-axis direction, \\(M_p^t\\) and \\(M_b^t\\) denote predicted instance mask and its GT box mask at frame \\(t\\) respectively. The object instance index is omitted here for clarity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.469, + 0.817 + ], + "angle": 0, + "content": "The Pairwise Loss \\( L_{\\mathrm{pair}} \\), on the other hand, constrains spatially neighboring pixels of single frame. For pixel of locations \\( p_i' \\) and \\( p_j' \\) of with color similarity \\( \\geqslant \\sigma_{\\mathrm{pixel}} \\), we enforce their predicted mask labels to be consistent, following Eq. (3) as," + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.824, + 0.469, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p a i r}} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {p _ {i} ^ {\\prime} \\in H \\times W} L _ {\\text {c o n s}} \\left(M _ {p _ {i} ^ {\\prime}} ^ {t}, M _ {p _ {j} ^ {\\prime}} ^ {t}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.874, + 0.462, + 0.891 + ], + "angle": 0, + "content": "The spatial losses are combined with a weight factor \\(\\lambda_{\\mathrm{pair}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.897, + 0.469, + 0.915 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s p a t i a l}} = \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {p a i r}} \\mathcal {L} _ {\\text {p a i r}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.083, + 0.892, + 0.143 + ], + "angle": 0, + "content": "Temporal Consistency We adopt the Temporal KNN-patch Loss in Sec. 3.1.2 as \\(\\mathcal{L}_{\\mathrm{temp}}\\) to leverage temporal mask consistency. The overall spatio-temporal objective \\(\\mathcal{L}_{\\mathrm{seg}}\\) for optimizing video segmentation is summarized as," + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.156, + 0.892, + 0.173 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e g}} = \\mathcal {L} _ {\\text {s p a t i a l}} + \\lambda_ {\\text {t e m p}} \\mathcal {L} _ {\\text {t e m p}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.19, + 0.871, + 0.205 + ], + "angle": 0, + "content": "3.2.2 Integration with Transformer-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Existing works [13, 48] on box-supervised segmentation losses are coupled with either one-stage or two-stage detectors, such as Faster R-CNN [41] and CondInst [48], and only address the single image case. However, state-of-the-art VIS methods [6, 57] are based on transformers. These works perform object detection via set prediction, where predicted instance masks need to be matched with mask annotations when evaluating the loss. To integrate mask-free VIS training with transformers, one key modification is in this instance-sequence matching step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.893, + 0.773 + ], + "angle": 0, + "content": "Since only ground-truth bounding boxes are available for box sequence matching, as an initial attempt, we first produce bounding box predictions from the estimated instance masks. Then, we employ the sequential box matching cost function used in VIS methods [55, 57]. To compute matching cost for whole sequence, \\(\\mathcal{L}_1\\) loss and generalized IoU loss for each individual bounding box is averaged across the frames. However, we observe the matching results of frame-wise averaging can easily be affected by a single outlier frame, especially under weak segmentation setup, leading to instability during training and performance decrease. Spatio-temporal Box Mask Matching Instead of using the aforementioned frame-wise matching, we empirically find spatio-temporal box-to-mask matching to produce substantial improvement under the weak segmentation setting. We first convert each predicted instance mask to a bounding box mask, and convert the ground-truth box to box mask. We then randomly sample a equal number of points from the ground-truth box mask sequence and predicted box mask sequence, respectively. Different from Mask2Former [6], we only adopt the dice IoU loss to compute sequence matching cost. We find that cross-entropy accumulates errors per pixel, leading to imbalanced values between large and small objects. In contrast, the IoU loss in normalized per object, leading to a balanced metric. We study different instance sequence matching strategies under the mask-free VIS setting in the ablation experiments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.783, + 0.836, + 0.799 + ], + "angle": 0, + "content": "3.2.3 Image-based MaskFreeVIS Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.892, + 0.913 + ], + "angle": 0, + "content": "Most VIS models [6, 57, 61] are initialized from a model pretrained on the COCO instance segmentation dataset. To completely eliminate mask supervision, we pretrain our MaskFreeVIS on COCO using only box supervision. We adopt the spatial consistency loss described in Sec. 3.2.1 on single frame to replace the original GT mask losses in Mask2Former [6], while following the same image-based" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.518, + 0.964 + ], + "angle": 0, + "content": "22861" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.083, + 0.468, + 0.157 + ], + "angle": 0, + "content": "training setup on COCO. Thus, we provide two training settings in our experiments, one eliminates both image and video mask during training, while the other adopts weights pretrained with COCO mask annotations. In both cases, no video mask annotations are used." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.169, + 0.208, + 0.186 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.194, + 0.178, + 0.209 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.217, + 0.468, + 0.322 + ], + "angle": 0, + "content": "YTVIS 2019/2021 We perform experiments on the largescale YouTube-VIS [61] 2019 and 2021. YTVIS 2019 includes 2,883 videos of 131k annotated object instances belonging to 40 categories. To handle more complex cases, YTVIS 2021 updates YTVIS 2019 with additional 794 videos for training and 129 videos for validation, including more tracklets with confusing motion trajectories." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.323, + 0.468, + 0.382 + ], + "angle": 0, + "content": "OVIS We also train and evaluate on OVIS [39], a VIS benchmark on occlusion learning. OVIS consists of instance masks covering 25 categories with 607, 140 and 154 videos for train, valid and test respectively." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.383, + 0.468, + 0.459 + ], + "angle": 0, + "content": "BDD100K MOTS We further report results of Mask-FreeVIS on the large-scale self-driving benchmark BDD100K [64] MOTS. The dataset annotates 154 videos (30,817 images) for training, 32 videos (6,475 images) for validation, and 37 videos (7,484 images) for testing." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.466, + 0.293, + 0.482 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.49, + 0.468, + 0.806 + ], + "angle": 0, + "content": "Our proposed approach only requires replacing the original video mask loss in state-of-the-art VIS methods. In particular, we adopt Mask2Former [6] and SeqFormer [57] due to their excellent VIS results. Unless specified, we kept all other training schedules and settings the same as in the original methods. For the Temporal KNN-patch Loss, we set the patch size to \\(3 \\times 3\\), search radius to 5 and \\(K = 5\\). We adopt the \\(L_{2}\\) distance as the patch matching metric and set the matching threshold to 0.05. On YTVIS 2019/2021, the Mask2Former based models are trained with AdamW [33] with learning rate \\(10^{-4}\\) and weight decay 0.05. The learning rate decays by 10 times at with a factor of 2/3. We set batch size to 16, and train 6k/8k iterations on YTVIS 2019/2021. For experiments on OVIS and BDD100K, we adopted the COCO mask pretrained models by VITA and Unicorn. For the sampled temporal tube at training, we use 5 frames with shuffling instead of 2 frames for better temporal regularization. The compared baselines are adjusted accordingly. During testing, since there is no architecture modification, the inference of MaskFreeVIS is the same to the baselines. More details are in the Supp. file." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.813, + 0.281, + 0.83 + ], + "angle": 0, + "content": "4.3. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.836, + 0.468, + 0.912 + ], + "angle": 0, + "content": "We perform detailed ablation studies for MaskFreeVIS using ResNet-50 as backbone on the YTVIS 2019 val set. We adopt the COCO box-pretrained model as initialization to eliminate all mask annotations from the training. Taking Mask2Former [6] as the base VIS method, we analyze" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.081, + 0.891, + 0.122 + ], + "angle": 0, + "content": "Table 2. Different temporal matching schemes under the mask-free training setting on YTVIS2019 val. 'Param' indicates whether the matching scheme brings extra model parameters." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.125, + 0.89, + 0.24 + ], + "angle": 0, + "content": "
Temporal Matching SchemeParamAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Baseline (No Matching)✘38.665.938.838.447.7
Flow-based Matching✓40.266.341.940.549.1
Temporal Deformable Matching✓39.665.940.139.948.6
Learnable Pixel Matching✓39.565.740.239.748.4
Learnable Patch Matching✓40.666.542.640.349.2
3D Pairwise Matching✘39.465.041.740.248.0
Temporal KNN-Matching (Ours)✘42.566.845.741.251.2
" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.246, + 0.891, + 0.291 + ], + "angle": 0, + "content": "the impact of individual proposed components. Moreover, we study several alternative solutions for temporal matching and influence of different hyper-parameters to TK-Loss." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.292, + 0.891, + 0.533 + ], + "angle": 0, + "content": "Comparison on Temporal Matching Schemes Table 2 compares our Temporal KNN-Matching to four alternative frame-wise matching approaches for enforcing temporal mask consistency. Flow-based Matching employs the pretrained optical flow model RAFT [45] to build pixel correspondence [32]. Temporal Deformable Matching adopts the temporal deformable kernels [47] to predict the pixel offsets between the target and alignment frame. Instead of using raw patches, Learnable Pixel/Patch Matching employs jointly learned deep pixel/patch embeddings via three learnable FC layers, which are then used to compute the affinities in a soft attention-like manner. 3D Pairwise Matching directly extends \\(\\mathcal{L}_{\\mathrm{pair}}\\) designed for spatial images to the temporal dimension, where pairwise affinity loss is computed among pixels not only in within the frame but also across multiple frames." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.535, + 0.891, + 0.911 + ], + "angle": 0, + "content": "In Table 2, compared to Flow-based Matching with one-to-one pixel correspondence, our paramter-free Temporal KNN-Matching with one-to- \\(K\\) improves by about 2.3 AP. The prediction of flow-based models are not reliable in case of occlusions and homogeneous regions, and are also influenced by the gap between the training dataset and realworld video data. For the above mentioned deformable and learnable matching schemes, since there are only weak bounding box labels during training, the temporal matching relation is learnt implicitly. We empirically observe the instability during training with limited improvement under the mask-free training setting. For direct generalization of \\(\\mathcal{L}_{\\mathrm{pair}}\\), it only leads to 0.8 mask AP performance improvement. Despite the simplicity and efficiency of the TK-loss, it significantly improves VIS performance by 3.9 mask AP. Effect of Temporal KNN-patch Loss MaskFreeVIS is trained with joint spatio-temporal losses. In Table 3, to evaluate the effectiveness of each loss component, we compare the performance of MaskFreeVIS solely under the spatial pairwise loss [50] or our proposed TK-Loss. Compared to the 2.0 mask AP improvement by the spatial pairwise loss, the TK-Loss substantially promotes the mask AP from 36.6 to 41.6, showing the advantage of our flexible one-to- \\(K\\) patch correspondence design in leveraging temporal consistency. We show the VIS results in Figure 5 to visualize" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "22862" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.076, + 0.47, + 0.103 + ], + "angle": 0, + "content": "Table 3. Effect of the Spatial Pairwise loss and our Temporal KNN-patch Loss on YTVIS2019 val." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.107, + 0.466, + 0.174 + ], + "angle": 0, + "content": "
Box ProjPairwiseTK-LossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
✓36.666.536.237.145.0
✓✓\\( 38.6_{\\uparrow 2.0} \\)65.938.838.447.7
✓✓\\( 41.6_{\\uparrow 5.0} \\)68.443.540.150.5
✓✓✓\\( 42.5_{\\uparrow 5.9} \\)66.845.741.251.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.176, + 0.269, + 0.203 + ], + "angle": 0, + "content": "Table 4. Effect of using max K matches on YTVIS2019 val." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.205, + 0.263, + 0.272 + ], + "angle": 0, + "content": "
KAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
140.865.844.140.3
341.966.945.141.9
542.566.845.741.2
742.367.144.640.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.282, + 0.268, + 0.308 + ], + "angle": 0, + "content": "Table 6. Impact of search radius R on YTVIS2019 val." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.31, + 0.263, + 0.378 + ], + "angle": 0, + "content": "
RAPAP50AP75AR1
139.665.740.239.7
341.366.643.040.3
542.566.845.741.2
742.367.144.640.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.176, + 0.466, + 0.216 + ], + "angle": 0, + "content": "Table 5. Patch Matching metrics comparison. NCC is Norm. Cross-Correlation." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.219, + 0.462, + 0.274 + ], + "angle": 0, + "content": "
MetricAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
NCC41.766.743.441.4
\\( L_1 \\)41.266.243.640.3
\\( L_2 \\)42.566.845.741.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.281, + 0.466, + 0.308 + ], + "angle": 0, + "content": "Table 7. Influence of patch size N on YTVIS2019 val." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.31, + 0.461, + 0.377 + ], + "angle": 0, + "content": "
NAPAP50AP75AR1
140.165.242.240.0
342.566.845.741.2
542.168.744.442.1
741.566.344.841.3
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.387, + 0.352, + 0.401 + ], + "angle": 0, + "content": "the effectiveness of each loss component." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.705 + ], + "angle": 0, + "content": "Analysis of Temporal KNN-patch Loss In Table 4, we study the influence of \\(\\mathbf{K}\\), the maximum number of matches selected in Temporal KNN-patch Loss. The best result is obtained for \\(K = 5\\), while \\(K = 1\\) only allows for the one-to-one and no-match cases. The improvement from 40.8 mask AP to 42.5 mask AP reveals the benefit brought by the flexible one-to-many correspondence design. We also analyze the matching metric in Table 5, search radius in Table 6, and patch size in Table 7 (see Sec. 3.1.2 for details). When patch size \\(N\\) is increased from 1 to 3 in Table 7, the performance of MaskFreeVIS is improved by 2.4 AP, validating the importance of patch structure in robust matching. Effect of the Cyclic Tube Connection We compare three frame-wise tube connection schemes (Figure 4) for the TK-Loss in Table 8. While dense connection brings forth the best performance, it doubles the training memory with minor improvement compared to Cyclic connection. Comparing to Sequential connection, our Cyclic connection benefits from long-range consistency, improving the performance of 0.6 mask AP with an affordable memory cost growth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.469, + 0.748 + ], + "angle": 0, + "content": "Table 8. Comparison of the tube connection schemes, illustrated in Fig. 4. The tube consist of 5 frames. 'Mem' denotes the memory consumption per sampled video by the TK-Loss during training." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.75, + 0.466, + 0.813 + ], + "angle": 0, + "content": "
Tube ConnectConnect Num.Mem (MB)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
Dense10152642.768.044.341.5
Sequential463141.966.544.741.2
Cyclic (Ours)577342.566.845.741.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.47, + 0.913 + ], + "angle": 0, + "content": "Comparison on Sequence Matching Functions Besides TK-Loss, we analyze the influence of sequence matching cost functions for transformer-based VIS methods under the mask-free setting. In Table 9, we identify the substantial advantage of Spatio-temporal box-mask matching over frameworkwise cost averaging [55, 57]. As discussed in Sec. 3.2.2, we" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.077, + 0.633, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.161, + 0.633, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.236, + 0.635, + 0.245 + ], + "angle": 0, + "content": "Mask2Fomer + BoxInst (Baseline)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.077, + 0.761, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.161, + 0.761, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.237, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Mask2Fomer + TK-Loss (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.077, + 0.89, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.161, + 0.89, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.237, + 0.875, + 0.245 + ], + "angle": 0, + "content": "Mask2Former (Oracle)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.25, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Figure 5. Qualitative results comparison between using Spatial Pairwise loss [50], our TK-Loss, and Mask2Former (oracle) trained with GT video and image masks." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.299, + 0.684, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.299, + 0.887, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.399, + 0.892, + 0.454 + ], + "angle": 0, + "content": "Figure 6. Results on YTVIS 2019 val with various percentages of the YTVIS training data. Baseline denotes Mask2Former [6] trained with GT video boxes using BoxInst [50], while Oracle denotes Mask2Former trained with GT video masks." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.457, + 0.892, + 0.499 + ], + "angle": 0, + "content": "Table 9. Comparison of Set Matching Cost Functions on YTVIS2019 val. ST-BoxMask denotes our Spatio-temporal Box Mask matching. w/o CE denotes removing cross-entropy cost." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.5, + 0.878, + 0.556 + ], + "angle": 0, + "content": "
Matching Cost FunctionAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Frame-wise Averaging [55, 57]37.664.239.537.545.7
ST-BoxMask40.867.842.240.049.2
ST-BoxMask w/o CE42.566.845.741.251.2
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.56, + 0.892, + 0.59 + ], + "angle": 0, + "content": "achieve further gain by removing the object size imbalanced cross-entropy cost computation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.741 + ], + "angle": 0, + "content": "Training on Various Amounts of Data To further study label-efficient VIS, we validate the effect of MaskFreeVIS under various percentages of the YTVIS 2019 training data. We uniformly sample frames and their labels for each video, and set the minimum sampled number of frames to 1. Figure 6 presents the experimental results, which shows the consistent large improvement (over 3.0 AP) brought by our TK-Loss under various amounts of data. In particular, we note that our approach with \\(50\\%\\) data even outperforms the fully-supervised model with \\(10\\%\\) training data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.864, + 0.768 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.892, + 0.852 + ], + "angle": 0, + "content": "We compare MaskFreeVIS with the state-of-the-art fully/weakly supervised methods on benchmarks YTVIS 2019/2021, OVIS and BDD100K MOTS. We integrate MaskFreeVIS on four representative methods [6,12,57,60], attaining consistent large gains over the strong baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.892, + 0.913 + ], + "angle": 0, + "content": "YTVIS 2019/2021 Table 10 compares the performance on YTVIS 2019. Using R50/R101 as backbone and with the same training setting, MaskFreeVIS achieves 42.5/45.8 AP, improving 3.9/5.0 AP over the strong baseline adopt" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.518, + 0.964 + ], + "angle": 0, + "content": "22863" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.066, + 0.47, + 0.122 + ], + "angle": 0, + "content": "Table 10. Comparison on YTVIS 2019. I: using COCO mask pretrained model as initialization. V: using YTVIS video masks during training. \\(\\ast\\): using pseudo mask from COCO images for joint training [57]. M2F: Mask2Former [6], SeqF: SeqFormer [57]." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.123, + 0.472, + 0.603 + ], + "angle": 0, + "content": "
MethodMask ann.Back-boneAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)
Fully-supervised:
PCAN [18]I+VR5036.154.939.436.341.6
EfficientVIS [59]I+VR5037.959.743.040.346.6
InsPro [10]I+VR5040.262.943.137.644.5
IFC [15]I+VR5042.865.846.843.851.2
VMT* [17]I+VR5047.9-52.045.8-
SeqF* [57]I+VR5047.469.851.845.554.8
M2FI+VR5046.468.050.0--
M2F*I+VR5047.869.252.746.256.6
Prev. Weakly-supervised:
FlowIRN [32]-R5010.527.26.212.313.6
SOLO-Track [9]IR5030.650.733.531.637.1
Mask-free:
M2F + Flow Consist [45]-R5040.266.341.940.549.1
M2F + BoxInst [50]-R5038.664.238.538.046.8
M2F + MaskFreeVIS-R5042.5↑3.966.845.741.251.2
M2F + MaskFreeVISIR5043.8↑5.270.746.941.552.3
M2F + MaskFreeVIS*IR5046.6↑8.072.549.744.955.7
Fully-supervised:
M2FVR10145.672.648.944.354.5
M2FI+VR10149.272.854.2--
M2F*I+VR10149.873.655.448.058.0
SeqF*I+VR10149.071.155.746.856.9
Mask-free:
M2F + BoxInst [50]-R10140.867.842.240.049.2
M2F + MaskFreeVIS-R10145.8↑5.070.848.645.355.2
M2F + MaskFreeVISIR10147.3↑6.575.449.944.655.2
M2F + MaskFreeVIS*IR10148.9↑8.174.954.744.957.0
SeqF + MaskFreeVIS*IR10148.674.052.245.957.2
Fully-supervised:
M2FI+VSwinL60.484.467.0--
Mask-free:
M2F + BoxInst [50]-SwinL49.873.255.548.258.1
M2F + MaskFreeVIS-SwinL54.3↑4.582.661.150.261.3
M2F + MaskFreeVIS*ISwinL55.3↑5.582.560.850.762.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.756 + ], + "angle": 0, + "content": "ing BoxInst [50] losses. MaskFreeVIS, without any mask labels, even significantly outperforms some recent fully-supervised methods such as EfficientVIS [59] and In-sPro [10]. On R50/R101/Swin-L, our MaskFreeVIS consistently attains over \\(91\\%\\) of its fully-supervised counterpart trained with both GT image and video masks. We also observe similar larger performance growth over the baseline on YTVIS 2021 in Table 11. The excellent performance substantially narrows the performance gap between fully-supervised and weakly-supervised VIS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.817 + ], + "angle": 0, + "content": "OVIS We also conduct experiments on OVIS in Table 12 using R50 as backbone. We integrate MaskFreeVIS with VITA [12], promoting the baseline performance from 12.1 to 15.7 under the mask-free training setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.817, + 0.47, + 0.879 + ], + "angle": 0, + "content": "BDD100K MOTS Table 13 further validates our approach on BDD100K MOTS. Integrated with Unicorn [60], Mask-FreeVIS achieves \\(23.8\\mathrm{mMOTSA}\\) by improving over 4.9 points over the strong baseline and thus surpassing the fully" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.913 + ], + "angle": 0, + "content": "1This work is supported in part by Research Grant Council of the Hong Kong SAR under grant no. 16201420 and Huawei Technologies." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.066, + 0.892, + 0.093 + ], + "angle": 0, + "content": "Table 11. Comparison on YTVIS 2021. Refer to Table 10 for the symbol abbreviations." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.094, + 0.895, + 0.345 + ], + "angle": 0, + "content": "
MethodMask ann.Back-boneAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Fully-supervised:
MaskTrack [61]I+VR5028.648.929.626.533.8
IFC [15]I+VR5036.657.939.3--
SeqF* [57]I+VR5040.562.443.736.148.1
M2FI+VR5040.660.941.8--
Mask-free:
M2F + BoxInst [50]-R5032.152.834.431.038.1
M2F + MaskFreeVIS-R5036.2↑4.160.839.234.645.6
M2F + MaskFreeVISIR5037.2↑5.161.940.335.346.1
M2F + MaskFreeVIS*IR5040.9↑8.865.843.337.150.5
Fully-supervised:
M2FI+VR10142.465.945.8--
Mask-free:
M2F + BoxInst [50]-R10133.355.232.532.141.9
M2F + MaskFreeVIS-R10137.3↑4.061.639.434.145.6
M2F + MaskFreeVISIR10138.2↑4.962.440.034.946.2
M2F + MaskFreeVIS*IR10141.6↑8.366.244.836.349.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.351, + 0.88, + 0.365 + ], + "angle": 0, + "content": "Table 12. State-of-the-art comparison on the OVIS using R50." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.368, + 0.887, + 0.5 + ], + "angle": 0, + "content": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Fully-supervised:
CrossVIS [62]14.932.712.110.319.8
Mask2Former [6]17.337.315.110.523.5
VMT [17]16.936.413.710.422.7
VITA [12]19.641.217.411.726.0
Video Mask-free:
VITA [12] + BoxInst [50]12.128.310.28.817.9
VITA [12] + MaskFreeVIS\\( 15.7_{\\uparrow 3,6} \\)35.113.110.120.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.505, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Table 13. State-of-the-art comparison on the BDD100K segmentation tracking validation set." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.535, + 0.894, + 0.652 + ], + "angle": 0, + "content": "
MethodmMOTSA↑mMOTSP↑mIDF↑ID sw.↓mAP↑
Fully-supervised:
STEm-Seg [1]12.258.225.4873221.8
QDTrack-mots-fix [36]23.566.344.597325.5
PCAN [18]27.466.745.187626.6
Unicorn [60]29.667.744.2173132.1
Video Mask-free:
Unicorn [60] + BoxInst [50]18.958.736.3329822.1
Unicorn [60] + MaskFreeVIS23.8↑4.966.744.9208624.8
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.66, + 0.892, + 0.706 + ], + "angle": 0, + "content": "supervised QDTrack-mots [36]. The consistent large gains on four benchmarks and four base VIS methods validate the generalizability of our MaskFreeVIS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.72, + 0.62, + 0.736 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.913 + ], + "angle": 0, + "content": "MaskFreeVIS is the first competitive VIS method that does not need any mask annotations during training. The strong results lead to a remarkable conclusion: mask labels are not a necessity for high-performing VIS. Our key component is the unsupervised Temporal KNN-patch Loss, which replaces the conventional video masks losses by leveraging temporal mask consistency constraints. Our approach greatly reduces the long-standing gap between fully-supervised and weakly-supervised VIS on four large-scale benchmarks. MaskFreeVIS thus opens up many opportunities for researchers and practitioners for label-efficient VIS." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.521, + 0.964 + ], + "angle": 0, + "content": "22864" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.081, + 0.176, + 0.098 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.106, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[1] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[2] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.206, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[3] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.234, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[4] Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.468, + 0.331 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.333, + 0.468, + 0.388 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.39, + 0.468, + 0.43 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.468, + 0.472 + ], + "angle": 0, + "content": "[8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.468, + 0.515 + ], + "angle": 0, + "content": "[9] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to track instances without video annotations. In CVPR, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[10] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. In NeurIPS, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.468, + 0.6 + ], + "angle": 0, + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.468, + 0.643 + ], + "angle": 0, + "content": "[12] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.468, + 0.699 + ], + "angle": 0, + "content": "[13] Cheng-Chun Hsu, Kuang-Jui Hsu, Chung-Chi Tsai, Yen-Yu Lin, and Yung-Yu Chuang. Weakly supervised instance segmentation using the bounding box tightness prior. In NeurIPS, 2019. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.468, + 0.741 + ], + "angle": 0, + "content": "[14] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.468, + 0.784 + ], + "angle": 0, + "content": "[15] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. In NeurIPS, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.468, + 0.827 + ], + "angle": 0, + "content": "[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.468, + 0.882 + ], + "angle": 0, + "content": "[17] Lei Ke, Henghui Ding, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Video mask transfiner for high-quality video instance segmentation. In ECCV, 2022. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.885, + 0.468, + 0.912 + ], + "angle": 0, + "content": "[18] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.106, + 0.47, + 0.912 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.084, + 0.892, + 0.11 + ], + "angle": 0, + "content": "for multiple object tracking and segmentation. In NeurIPS, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.113, + 0.892, + 0.152 + ], + "angle": 0, + "content": "[19] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.155, + 0.892, + 0.195 + ], + "angle": 0, + "content": "[20] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.198, + 0.892, + 0.251 + ], + "angle": 0, + "content": "[21] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.254, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[22] Viveka Kulharia, Siddhartha Chandra, Amit Agrawal, Philip Torr, and Ambrish Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[23] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[24] Jungbeom Lee, Eunji Kim, Sungmin Lee, Jangho Lee, and Sungroh Yoon. Frame-to-frame aggregation of active regions in web videos for weakly supervised semantic segmentation. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[25] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[26] Minghan Li, Shuai Li, Lida Li, and Lei Zhang. Spatial feature calibration and temporal fusion for effective one-stage video instance segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[27] Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, XianSheng Hua, and Lei Zhang. Box-supervised instance segmentation with level set evolution. In ECCV, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[28] Chung-Ching Lin, Ying Hung, Rogerio Feris, and Linglin He. Video instance segmentation tracking with a modified vae architecture. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[29] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[31] Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[32] Qing Liu, Vignesh Ramanathan, Dhruv Mahajan, Alan Yuille, and Zhenheng Yang. Weakly supervised instance segmentation for videos with temporal mask consistency. In CVPR, 2021. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.829, + 0.892, + 0.869 + ], + "angle": 0, + "content": "[34] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.871, + 0.892, + 0.911 + ], + "angle": 0, + "content": "[35] Anton Milan, Laura Leal-Taixe, Konrad Schindler, and Ian Reid. Joint tracking and segmentation of multiple targets. In CVPR, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.084, + 0.892, + 0.911 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "22865" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.084, + 0.469, + 0.125 + ], + "angle": 0, + "content": "[36] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.127, + 0.47, + 0.18 + ], + "angle": 0, + "content": "[37] George Papandreou, Liang-Chieh Chen, Kevin P Murphy, and Alan L Yuille. Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.183, + 0.469, + 0.236 + ], + "angle": 0, + "content": "[38] Jordi Pont-Tuset, Pablo Arbelaez, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping for image segmentation and object proposal generation. IEEE TPAMI, 39(1):128-140, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.239, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[39] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.295, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[40] Martin Rajchl, Matthew CH Lee, Ozan Oktay, Konstantinos Kamnitsas, Jonathan Passerat-Palmbach, Wenjia Bai, Mellisa Damodaram, Mary A Rutherford, Joseph V Hajnal, Bernhard Kainz, et al. Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging, 36(2):674-683, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.435, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[42] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. Grabcut: interactive foreground extraction using iterated graph cuts. ACM TOG, 23(3):309-314, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.478, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[43] Fatemeh Sadat Saleh, Mohammad Sadegh Aliakbarian, Mathieu Salzmann, Lars Petersson, and Jose M Alvarez. Bringing background into the foreground: Making all classes equal in weakly-supervised video semantic segmentation. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[44] Chunfeng Song, Yan Huang, Wanli Ouyang, and Liang Wang. Box-driven class-wise region masking and filling rate guided loss for weakly supervised semantic segmentation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[46] Omkar Thawakar, Sanath Narayan, Jiale Cao, Hisham Cholakkal, Rao Muhammad Anwer, Muhammad Haris Khan, Salman Khan, Michael Felsberg, and Fahad Shahbaz Khan. Video instance segmentation via multi-scale spatiotemporal split attention transformer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[47] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.745, + 0.469, + 0.783 + ], + "angle": 0, + "content": "[48] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.787, + 0.469, + 0.825 + ], + "angle": 0, + "content": "[49] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.829, + 0.469, + 0.868 + ], + "angle": 0, + "content": "[50] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In CVPR, 2021. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.872, + 0.469, + 0.911 + ], + "angle": 0, + "content": "[51] Pavel Tokmakov, Karteek Alahari, and Cordelia Schmid. Learning semantic segmentation with weakly-annotated videos. In ECCV, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.084, + 0.47, + 0.911 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.084, + 0.891, + 0.11 + ], + "angle": 0, + "content": "[52] Yi-Hsuan Tsai, Guangyu Zhong, and Ming-Hsuan Yang. Semantic co-segmentation in videos. In ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.113, + 0.892, + 0.166 + ], + "angle": 0, + "content": "[53] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.17, + 0.892, + 0.223 + ], + "angle": 0, + "content": "[54] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.226, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.283, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[56] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[57] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. In ECCV, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[58] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.411, + 0.892, + 0.464 + ], + "angle": 0, + "content": "[59] Jialian Wu, Sudhir Yarram, Hui Liang, Tian Lan, Junsong Yuan, Jayan Eledath, and Gerard Medioni. Efficient video instance segmentation via tracklet query and proposal. In CVPR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.468, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[60] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.536 + ], + "angle": 0, + "content": "[61] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.592 + ], + "angle": 0, + "content": "[62] Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In ICCV, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[63] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.653, + 0.892, + 0.707 + ], + "angle": 0, + "content": "[64] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.084, + 0.892, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "22866" + } + ] +] \ No newline at end of file diff --git a/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_origin.pdf b/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0c513151ed0f30aeb6d36ae3cf930e18ef075eb0 --- /dev/null +++ b/2023/Mask-Free Video Instance Segmentation/91f920d3-e27b-4553-9ceb-9dfd929137de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a77a4578fe121dd848a277e6c217e958d44c62661f9ec44031b380e97b142e5b +size 4471272 diff --git a/2023/Mask-Free Video Instance Segmentation/full.md b/2023/Mask-Free Video Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..56742b86188f32a7380cb400764b86944b300445 --- /dev/null +++ b/2023/Mask-Free Video Instance Segmentation/full.md @@ -0,0 +1,369 @@ +# Mask-Free Video Instance Segmentation + +Lei Ke $^{1,2}$ Martin Danelljan $^{1}$ Henghui Ding $^{1}$ Yu-Wing Tai $^{2}$ Chi-Keung Tang $^{2}$ Fisher Yu $^{1}$ ETH Zürich ${}^{2}\mathrm{HKUST}$ + +![](images/2df21adc52ff236309c7cb72bd869ca6ddfe29f15f25f57a188924dfbe5cd526.jpg) +Figure 1. Video instance segmentation (VIS) results of our MaskFreeVIS, trained without using any video or image mask annotation. By achieving a remarkable $42.5\%$ mask AP on the YouTube-VIS val dataset, with a ResNet-50 backbone, our approach demonstrates that high-performing VIS can be learned even without any mask annotations. + +# Abstract + +The recent advancement in Video Instance Segmentation (VIS) has largely been driven by the use of deeper and increasingly data-hungry transformer-based models. However, video masks are tedious and expensive to annotate, limiting the scale and diversity of existing VIS datasets. In this work, we aim to remove the mask-annotation requirement. We propose MaskFreeVIS, achieving highly competitive VIS performance, while only using bounding box annotations for the object state. We leverage the rich temporal mask consistency constraints in videos by introducing the Temporal KNN-patch Loss (TK-Loss), providing strong mask supervision without any labels. Our TK-Loss finds one-to-many matches across frames, through an efficient patch-matching step followed by a $K$ -nearest neighbor selection. A consistency loss is then enforced on the found matches. Our mask-free objective is simple to implement, has no trainable parameters, is computationally efficient, yet outperforms baselines employing, e.g., state-of-the-art optical flow to enforce temporal mask consistency. We validate MaskFreeVIS on the YouTube-VIS 2019/2021, OVIS and BDD100K MOTS benchmarks. The results clearly demonstrate the efficacy of our method by drastically narrowing the gap between fully and weakly-supervised VIS performance. Our code and trained models are available at http://vis.xyz/pub/maskfreevis. + +# 1. Introduction + +Video Instance Segmentation (VIS) requires jointly detecting, tracking and segmenting all objects in a video from a given set of categories. To perform this challenging task, state-of-the-art VIS models are trained with complete video annotations from VIS datasets [39, 61, 64]. However, video annotation is costly, in particular regarding object mask labels. Even coarse polygon-based mask annotation is multiple times slower than annotating video bounding boxes [7]. Expensive mask annotation makes existing VIS benchmarks difficult to scale, limiting the number of object categories covered. This is particularly a problem for the recent transformer-based VIS models [6, 17, 57], which tend to be exceptionally data-hungry. We therefore revisit the need for complete mask annotation by studying the problem of weakly supervised VIS under the mask-free setting. + +While there exist box-supervised instance segmentation models [13, 23, 27, 50], they are designed for images. These weakly-supervised single-image methods do not utilize temporal cues when learning mask prediction, leading to lower accuracy when directly applied to videos. As a source for weakly supervised learning, videos contain much richer information about the scene. In particular, videos adhere to the temporal mask consistency constraint, where the regions corresponding to the same underlying object across different frames should have the same mask label. In this work, we set out to leverage this important constraint for mask-free learning of VIS. + +![](images/ad65f7aa1ab4c1a9f2dff2f18954500f56a51c3dda909f4eac4ed3f9e17ef030.jpg) +Temporal One-to-K Patch Correspondence +Figure 2. Our Temporal KNN-patch Loss enforces mask consistency between one-to- $k$ patch correspondences found across frames, which allow us to cover the cases where: (i) A unique one-to-one match exists (blue); (ii) Multiple matches are found due to ambiguities in homogenous regions (orange) or along image edges (white and yellow); (iii) No match is found due to $e.g.$ occlusions (green). This allows us to robustly leverage mask consistency constraints in challenging videos. + +We propose MaskFreeVIS method, for high performance VIS without any mask annotations. To leverage temporal mask consistency, we introduce the Temporal KNN-patch Loss (TK-Loss), as in Figure 2. To find regions corresponding to the same underlying video object, our TK-Loss first builds correspondences across frames by patch-wise matching. For each target patch, only the top $K$ matches in the neighboring frame with high enough matching score are selected. A temporal consistency loss is then applied to all found matches to promote the mask consistency. Specifically, our surrogate objective function not only promotes the one-to- $k$ matched regions to reach the same mask probabilities, but also commits their mask prediction to a confident foreground or background prediction by entropy minimization. Unlike flow-based models [32,45], which assume one-to-one matching, our approach builds robust and flexible one-to- $k$ correspondences to cope with e.g. occlusions and homogeneous regions, without introducing additional model parameters or inference cost. + +The TK-Loss is easily integrated into existing VIS methods, with no architecture modifications required. During training, our TK-Loss simply replaces the conventional video mask losses in supervising video mask generation. To further enforce temporal consistency through the video clip, TK-Loss is employed in a cyclic manner instead of using dense frame-wise connections. This greatly reduces memory cost with negligible performance drop. + +We extensively evaluate our MaskFreeVIS on four large-scale VIS benchmarks, i.e., YouTube-VIS 2019/2021 [61], OVIS [39], and BDD100K MOTS [64]. MaskFreeVIS achieves competitive VIS performance without using any video masks or even image mask labels on all datasets. Validated on various methods and backbones, MaskFreeVIS achieves $91.25\%$ performance of its fully supervised counterparts, even outperforming a few recent fully-supervised methods [10, 15, 18, 59] on the popular YTVIS benchmark. Our simple yet effective design greatly narrows the performance gap between weakly-supervised and fully + +Table 1. Mask annotation requirement for state-of-the-art VIS methods. Results are reported using ResNet-50 as backbone on the YTVIS 2019 [61] benchmark. Video Mask: using YTVIS video mask labels. Image Mask: using COCO [30] image mask labels for image-based pretraining. Pseudo Video: using Pseudo Videos from COCO images for joint training [57]. MaskFreeVIS achieves $91.5\%$ (42.5 vs. 46.4) of its fully-supervised baseline performance (Mask2Former) without using any masks during training. + +
MethodVideo MaskImage MaskPseudo VideoAP
SeqFormer [57]✓✓✓47.4
VMT [17]✓✓✓47.9
Mask2Former [6]✓✓✓47.8
MaskFreeVIS (ours)✗✓✓46.6
Mask2Former [6]✓✓✗46.4
MaskFreeVIS (ours)✗✗✗42.5
+ +supervised video instance segmentation. It further demonstrates that expensive video masks, or even image masks, is not necessary for training high-performing VIS models. + +Our contributions are summarized as follows: (i) To utilize temporal information, we develop a new parameter-free Temporal KNN-patch Loss, which leverages temporal masks consistency using unsupervised one-to- $k$ patch correspondence. We extensively analyze the TK-Loss through ablative experiments. (ii) Based on the TK-Loss, we develop the MaskFreeVIS method, enabling training existing state-of-the-art VIS models without any mask annotation. (iii) To the best of our knowledge, MaskFreeVIS is the first mask-free VIS method attaining high-performing segmentation results. We provide qualitative results in Figure 1. As in Table 1, when integrated into the Mask2Former [6] baseline with ResNet-50, our MaskFreeVIS achieves $42.5\%$ mask AP on the challenging YTVIS 2019 benchmark while using no video or image mask annotations. Our approach further scales to larger backbones, achieving $55.3\%$ mask AP on Swin-L backbone with no video mask annotations. + +We hope our approach will facilitate achieving label-efficient video instance segmentation, enabling building even larger-scale VIS benchmarks with diverse categories by lifting the mask annotation restriction. + +# 2. Related Work + +Video Instance Segmentation (VIS) Existing VIS methods can be summarized into three categories: two-stage, one-stage, and transformer-based. Two-stage approaches [2, 18, 28, 29, 61] extend the Mask R-CNN family [11, 19] by designing an additional tracking branch for object association. One-stage works [4, 26, 31, 62] adopt anchor-free detectors [49], generally using linear masks basis combination [3] or conditional mask prediction generation [48]. For the transformer-based models [6, 12, 46, 57, 63], VisTr [54] firstly adapts the transformer [5] for VIS, and IFC [15] further improves its efficiency via memory tokens. Seqformer [57] proposes frame query decompo + +sition while Mask2Former [6] includes masked attention. VMT [17] extends Mask Transfiner [16] to video for high-quality VIS, and IDOL [58] focuses on online VIS. State-of-the-art VIS methods with growing capacity put limited emphasis on weak supervision. In contrast, the proposed MaskFreeVIS is the first method targeting mask-free VIS while attaining competitive performance. + +Multiple Object Tracking and Segmentation (MOTS) Most MOTS methods [1,34,35,53,56] follow the tracking-by-detection principle. PCAN [18] improves temporal segmentation by utilizing space-time memory prototypes, while the one-stage method Unicorn [60] focuses on unification of different tracking frameworks. Compared to the aforementioned fully-supervised MOTS methods, Mask-FreeVIS focuses on label efficient training without GT masks by proposing a new surrogate temporal loss which can be easily integrated on them. + +Mask-Free VIS Most mask-free instance segmentation works [7, 13, 20, 25, 27, 37, 40, 44] are designed for single images and thus neglect temporal information. Earlier works BoxSup [8] and Box2Seg [22] rely on region proposals produced by MCG [38] or GrabCut [42], leading to slow training. BoxInst [50] proposes the surrogate projection and pixel pairwise losses to replace the original mask learning loss of CondInst [48], while DiscoBox [23] focuses on generating pseudo mask labels guided by a teacher model. + +Earlier works have investigated the use of videos for weakly-, semi-, or un-supervised segmentation by leveraging motion or temporal consistency [21,51,52]. Most aforementioned approaches do not address the VIS problem, and use optical flow for frame-to-frame matching [24, 32, 43]. In particular, FlowIRN [32] explores VIS using only classification labels and incorporates optical flow to leverage mask consistency. The limited performance makes the class-label only or fully-unsupervised setting difficult to deploy in the real world. SOLO-Track [9] aims to train VIS models without video annotations, and one concurrent work MinVIS [14] performs VIS without video-based model architectures. Unlike the above weakly-supervised training settings, our MaskFreeVIS is designed for eliminating the mask annotation requirement for VIS, as we note that video mask labeling is particularly expensive. MaskFreeVIS enables training VIS models without any video masks, or even image masks. Despite its simplicity, MaskFreeVIS drastically reduces the gap between fully-supervised and weakly-supervised VIS models, making weakly-supervised models more accessible in practice. + +# 3. Method + +We propose MaskFreeVIS to tackle video instance segmentation (VIS) without using any video or even image mask labels. Our approach is generic and can be directly applied to train state-of-the-art VIS methods, such as Mask2Former [6] and SeqFormer [57]. In Sec. 3.1, we first + +present the core component of MaskFreeVIS: the Temporal KNN-patch Loss (TK-Loss), which leverages temporal consistency to supervise accurate mask prediction, without any human mask annotations. In Sec. 3.2, we then describe how to integrate the TK-Loss with existing spatial weak segmentation losses for transformer-based VIS methods, to achieve mask-free training of VIS approaches. Finally, we introduce image-based pretraining details of our MaskFreeVIS in Sec. 3.2.3. + +# 3.1. MaskFreeVIS + +In this section, we introduce the Temporal KNN-patch Loss, illustrated in Figure 3. It serves as an unsupervised objective for mask prediction that leverages the rich spatiotemporal consistency constraints in unlabelled videos. + +# 3.1.1 Temporal Mask Consistency + +While an image constitutes a single snapshot of a scene, a video provides multiple snapshots displaced in time. Thereby, a video depicts continuous change in the scene. Objects and background move, deform, are occluded, experience variations in lighting, motion blur, and noise, leading to a sequence of different images that are closely related through gradual transformations. + +Consider a small region in the scene (Fig. 2), belonging either to an object or background. The pixels corresponding to the projection of this region should have the same mask prediction in every frame, as they belong to the same underlying physical object or background region. However, the aforementioned dynamic changes in the video lead to substantial appearance variations, serving as a natural form of data augmentation. The fact that the pixels corresponding to the same underlying object region should have the same mask prediction under temporal change therefore provides a powerful constraint, i.e., temporal mask consistency, which can be used for mask supervision [21, 24, 32, 51, 52]. + +The difficulty in leveraging the temporal mask consistency constraint stems from the problem of establishing reliable correspondences between video frames. Objects often undergo fast motion, deformations, etc., resulting in substantial appearance change. Furthermore, regions in the scene may be occluded or move out of the image from one frame to the other. In such cases, no correspondence exist. Lastly, videos are often dominated by homogenous regions, such as sky and ground, where the establishment of one-to-one correspondences are error-prone or even ill-defined. + +The problem of establishing dense one-to-one correspondences between subsequent video frames, known as optical flow, is a long-standing and popular research topic. However, when attempting to enforce temporal mask consistency through optical flow estimation [24,32,43], one encounters two key problems. 1) The one-to-one assumption of optical flow is not suitable in cases of occlusions, homogenous regions, and along single edges, where the corre + +![](images/08adf8978d71ca66c5814d3e778b63990cadabda627e958f8883c6a02009439f.jpg) +(1) Temporal Candidates Extraction +Figure 3. Temporal KNN-patch Loss has four steps: 1) Patch Candidate Extraction: Patch candidates searching across frames with radius $R$ . 2) Temporal KNN-Matching: Match $k$ high-confidence candidates by patch affinities. 3) Consistency loss: Enforce mask consistency objective (Eq. 2) among the matches. 4) Cyclic Tube Connection: Temporal loss aggregation in the 5-frame tube, detailed in Figure 4. + +![](images/72a30751e06486c0c306acfc9f139e3694ae338d6b3e7a905476a6ab62e523b8.jpg) +(2) Temporal KNN-Matching + +![](images/2e8f94bf5ff32f721a2b28f5d96974c30864adc36a2bddeadea1ef7b5be769c1.jpg) +Loss Backward + +![](images/9dc5c571ae48cdb53660ba47af13fcd4b616a62a40e9d40b586b29b3cddc1957.jpg) +(3) KNN-Consistency Loss + +![](images/faa4326afa7e0473b380c95f8a1f02eefc5ed465e09785eb251695d39413c4b1.jpg) +4) Cyclic Tube Connection + +![](images/77f665369b9c035734a800aeb31c05f322310ee968e66b657ddce88211558705.jpg) +MaskFreeVIS Prediction + +spondence is either nonexistent, undefined, ambiguous, uncertain, or very difficult to determine. 2) State-of-the-art optical flow estimation relies on large and complex deep networks, with large computational and memory requirements. + +Instead of using optical flow, we aim to design a simple, efficient, and parameter-free approach that effectively enforces the temporal mask consistency constraint. + +# 3.1.2 Temporal KNN-patch Loss + +Our Temporal KNN-patch Loss (TK-Loss) is based on a simple and flexible correspondence estimation across frames. In contrast to optical flow, we do not restrict our formulation to one-to-one correspondences. Instead, we establish one-to- $K$ correspondences. This include the conventional one-to-one $(K = 1)$ , where a unique well-defined match exists. However, this allows us to also handle the cases of nonexistent correspondences $(K = 0)$ in case of occlusions, and one-to-many $(K \geq 2)$ in case of homogenous regions. In cases where multiple matches are found, these most often belong to the same underlying object or background due to their similar appearance, as in Figure 2. This further benefits our mask consistency objective through denser supervision. Lastly, our approach is simple to implement, with negligible computational overhead and no learnable parameters. Our approach is in Figure 3, and contains four main steps, which are detailed next. + +1) Patch Candidate Extraction: Let $X_p^t$ denote an $N \times N$ target image patch centered at spatial location $p = (x, y)$ in frame $t$ . Our aim is to find a set of corresponding positions $S_p^{t \to \hat{t}} = \{\hat{p}_i\}_i$ in frame number $\hat{t}$ that represent the same object region. To this end, we first select candidate locations $\hat{p}$ within a radius $R$ such that $\| p - \hat{p} \| \leq R$ . Such windowed patch search exploits spatial proximity across neighboring frames in order to avoid an exhaustive global search. For a fast implementation, the windowed search is performed for all target image patches $X_p^t$ in parallel. +2) Temporal KNN-Matching: We match patch candidate patches through a simple distance computation, + +$$ +\mathbf {d} _ {p \rightarrow \hat {p}} ^ {t \rightarrow \hat {t}} = \left\| X _ {p} ^ {t} - X _ {\hat {p}} ^ {\hat {t}} \right\|, \tag {1} +$$ + +In our ablative experiments (Sec. 4.3), we found the $L_{2}$ + +norm to be the most effective patch matching metric. We select the top $K$ matches with smallest patch distance $\mathbf{d}_{p\rightarrow \hat{p}}^{t\rightarrow \hat{t}}$ . Lastly low-confidence matches are removed by enforcing a maximal patch distance $D$ as $\mathbf{d}_{p\rightarrow \hat{p}}^{t\rightarrow \hat{t}} < D$ . The remaining matches form the set $S_{p}^{t\rightarrow \hat{t}} = \{\hat{p}_{i}\}_{i}$ for each location $p$ . + +3) Consistency loss: Let $M_p^t \in [0,1]$ denote the predicted binary instance mask of an object, evaluated at position $p$ in frame $t$ . To ensure temporal mask consistency constraints, we penalize inconsistent mask predictions between a spatiotemporal point $(p,t)$ and its estimated corresponding points in $S_p^{t \to \hat{t}}$ . In particular we use the following objective, + +$$ +\mathcal {L} _ {f} ^ {t \rightarrow \hat {t}} = \frac {1}{H W} \sum_ {p} \sum_ {\hat {p} _ {i} \in \mathcal {S} _ {p} ^ {t \rightarrow \hat {t}}} L _ {\text {c o n s}} \left(M _ {p} ^ {t}, M _ {\hat {p} _ {i}} ^ {\hat {t}}\right), \tag {2} +$$ + +where mask consistency is measured as + +$$ +L _ {\text {c o n s}} \left(M _ {p} ^ {t}, M _ {\hat {p}} ^ {\hat {t}}\right) = - \log \left(M _ {p} ^ {t} M _ {\hat {p}} ^ {\hat {t}} + \left(1 - M _ {p} ^ {t}\right) \left(1 - M _ {\hat {p}} ^ {\hat {t}}\right)\right). \tag {3} +$$ + +Note that Eq. (3) only attains its minimum value of zero if both predictions indicate exactly background ( $M_p^t = M_{\hat{p}}^{\hat{t}} = 0$ ) or foreground ( $M_p^t = M_{\hat{p}}^{\hat{t}} = 1$ ). The objective does thus not only promote the two mask predictions to achieve the same probability value $M_p^t = M_{\hat{p}}^{\hat{t}}$ , but also to commit to a certain foreground or background prediction. + +4) Cyclic Tube Connection: Suppose the temporal tube consists of $T$ frames. We compute the temporal loss for the whole tube in a cyclic manner, as in Figure 4. The start frame is connected to the end frame, which introduces direct long-term mask consistency across the two temporally most distant frames. The temporal TK-Loss for the whole tube is given by + +$$ +\mathcal {L} _ {\text {t e m p}} = \sum_ {t = 1} ^ {T} \left\{\begin{array}{l l}\mathcal {L} _ {f} ^ {t \rightarrow (t + 1)}&\text {i f} t < T - 1\\\mathcal {L} _ {f} ^ {t \rightarrow 0}&\text {i f} t = T - 1.\end{array}\right. \tag {4} +$$ + +Compared to inter-frame dense connections in Figure 4, we find the cyclic loss to achieve similar performance but greatly reduce the memory usage as validated in the experiment section. + +![](images/025f71d424fdc899da266bf565e2003fb4f9a1955446db6c208414ca14289cce.jpg) +Figure 4. Illustration of different frame-wise tube connection settings (connection number) in the temporal loss design. + +# 3.2. Training MaskFreeVIS + +In this section, we describe how to train state-of-the-art VIS methods using our TK-Loss, without any mask annotations. Our MaskFreeVIS approach is jointly supervised with spatial-temporal surrogate losses, and is easily integrated with existing transformer-based methods. We also detail mask-free image-based pre-training for MaskFree-VIS to fully eliminate mask usage during training. + +# 3.2.1 Joint Spatio-temporal Regularization + +To train MaskFreeVIS, in addition to our proposed Temporal KNN-patch Loss for temporal mask consistency, we leverage existing spatial weak segmentation losses to jointly enforce intra-frame consistency. + +Spatial Consistency To explore spatial weak supervision signals from image bounding boxes and pixel color, we utilize the representative Box Projection Loss $L_{\mathrm{proj}}$ and Pairwise Loss $L_{\mathrm{pair}}$ in [50], to replace the supervised mask learning loss. The Projection Loss $L_{\mathrm{proj}}$ enforces the projection $P'$ of the object mask onto the $\vec{x}$ -axis and $\vec{y}$ -axis of image to be consistent with its ground-truth box mask. For the temporal tube with $T$ frames, we concurrently optimize all predicted frame masks of the tube as, + +$$ +\mathcal {L} _ {\mathrm {p r o j}} = \sum_ {t = 1} ^ {T} \sum_ {d \in \{\vec {x}, \vec {y} \}} D \left(P _ {d} ^ {\prime} \left(M _ {p} ^ {t}\right), P _ {d} ^ {\prime} \left(M _ {b} ^ {t}\right)\right), \tag {5} +$$ + +where $D$ denotes dice loss, $P'$ is the projection function along $\vec{x} / \vec{y}$ -axis direction, $M_p^t$ and $M_b^t$ denote predicted instance mask and its GT box mask at frame $t$ respectively. The object instance index is omitted here for clarity. + +The Pairwise Loss $L_{\mathrm{pair}}$ , on the other hand, constrains spatially neighboring pixels of single frame. For pixel of locations $p_i'$ and $p_j'$ of with color similarity $\geqslant \sigma_{\mathrm{pixel}}$ , we enforce their predicted mask labels to be consistent, following Eq. (3) as, + +$$ +\mathcal {L} _ {\text {p a i r}} = \frac {1}{T} \sum_ {t = 1} ^ {T} \sum_ {p _ {i} ^ {\prime} \in H \times W} L _ {\text {c o n s}} \left(M _ {p _ {i} ^ {\prime}} ^ {t}, M _ {p _ {j} ^ {\prime}} ^ {t}\right). \tag {6} +$$ + +The spatial losses are combined with a weight factor $\lambda_{\mathrm{pair}}$ + +$$ +\mathcal {L} _ {\text {s p a t i a l}} = \mathcal {L} _ {\text {p r o j}} + \lambda_ {\text {p a i r}} \mathcal {L} _ {\text {p a i r}}. \tag {7} +$$ + +Temporal Consistency We adopt the Temporal KNN-patch Loss in Sec. 3.1.2 as $\mathcal{L}_{\mathrm{temp}}$ to leverage temporal mask consistency. The overall spatio-temporal objective $\mathcal{L}_{\mathrm{seg}}$ for optimizing video segmentation is summarized as, + +$$ +\mathcal {L} _ {\text {s e g}} = \mathcal {L} _ {\text {s p a t i a l}} + \lambda_ {\text {t e m p}} \mathcal {L} _ {\text {t e m p}}. \tag {8} +$$ + +# 3.2.2 Integration with Transformer-based Methods + +Existing works [13, 48] on box-supervised segmentation losses are coupled with either one-stage or two-stage detectors, such as Faster R-CNN [41] and CondInst [48], and only address the single image case. However, state-of-the-art VIS methods [6, 57] are based on transformers. These works perform object detection via set prediction, where predicted instance masks need to be matched with mask annotations when evaluating the loss. To integrate mask-free VIS training with transformers, one key modification is in this instance-sequence matching step. + +Since only ground-truth bounding boxes are available for box sequence matching, as an initial attempt, we first produce bounding box predictions from the estimated instance masks. Then, we employ the sequential box matching cost function used in VIS methods [55, 57]. To compute matching cost for whole sequence, $\mathcal{L}_1$ loss and generalized IoU loss for each individual bounding box is averaged across the frames. However, we observe the matching results of frame-wise averaging can easily be affected by a single outlier frame, especially under weak segmentation setup, leading to instability during training and performance decrease. Spatio-temporal Box Mask Matching Instead of using the aforementioned frame-wise matching, we empirically find spatio-temporal box-to-mask matching to produce substantial improvement under the weak segmentation setting. We first convert each predicted instance mask to a bounding box mask, and convert the ground-truth box to box mask. We then randomly sample a equal number of points from the ground-truth box mask sequence and predicted box mask sequence, respectively. Different from Mask2Former [6], we only adopt the dice IoU loss to compute sequence matching cost. We find that cross-entropy accumulates errors per pixel, leading to imbalanced values between large and small objects. In contrast, the IoU loss in normalized per object, leading to a balanced metric. We study different instance sequence matching strategies under the mask-free VIS setting in the ablation experiments. + +# 3.2.3 Image-based MaskFreeVIS Pre-training + +Most VIS models [6, 57, 61] are initialized from a model pretrained on the COCO instance segmentation dataset. To completely eliminate mask supervision, we pretrain our MaskFreeVIS on COCO using only box supervision. We adopt the spatial consistency loss described in Sec. 3.2.1 on single frame to replace the original GT mask losses in Mask2Former [6], while following the same image-based + +training setup on COCO. Thus, we provide two training settings in our experiments, one eliminates both image and video mask during training, while the other adopts weights pretrained with COCO mask annotations. In both cases, no video mask annotations are used. + +# 4. Experiments + +# 4.1. Datasets + +YTVIS 2019/2021 We perform experiments on the largescale YouTube-VIS [61] 2019 and 2021. YTVIS 2019 includes 2,883 videos of 131k annotated object instances belonging to 40 categories. To handle more complex cases, YTVIS 2021 updates YTVIS 2019 with additional 794 videos for training and 129 videos for validation, including more tracklets with confusing motion trajectories. + +OVIS We also train and evaluate on OVIS [39], a VIS benchmark on occlusion learning. OVIS consists of instance masks covering 25 categories with 607, 140 and 154 videos for train, valid and test respectively. + +BDD100K MOTS We further report results of Mask-FreeVIS on the large-scale self-driving benchmark BDD100K [64] MOTS. The dataset annotates 154 videos (30,817 images) for training, 32 videos (6,475 images) for validation, and 37 videos (7,484 images) for testing. + +# 4.2. Implementation Details + +Our proposed approach only requires replacing the original video mask loss in state-of-the-art VIS methods. In particular, we adopt Mask2Former [6] and SeqFormer [57] due to their excellent VIS results. Unless specified, we kept all other training schedules and settings the same as in the original methods. For the Temporal KNN-patch Loss, we set the patch size to $3 \times 3$ , search radius to 5 and $K = 5$ . We adopt the $L_{2}$ distance as the patch matching metric and set the matching threshold to 0.05. On YTVIS 2019/2021, the Mask2Former based models are trained with AdamW [33] with learning rate $10^{-4}$ and weight decay 0.05. The learning rate decays by 10 times at with a factor of 2/3. We set batch size to 16, and train 6k/8k iterations on YTVIS 2019/2021. For experiments on OVIS and BDD100K, we adopted the COCO mask pretrained models by VITA and Unicorn. For the sampled temporal tube at training, we use 5 frames with shuffling instead of 2 frames for better temporal regularization. The compared baselines are adjusted accordingly. During testing, since there is no architecture modification, the inference of MaskFreeVIS is the same to the baselines. More details are in the Supp. file. + +# 4.3. Ablation Experiments + +We perform detailed ablation studies for MaskFreeVIS using ResNet-50 as backbone on the YTVIS 2019 val set. We adopt the COCO box-pretrained model as initialization to eliminate all mask annotations from the training. Taking Mask2Former [6] as the base VIS method, we analyze + +Table 2. Different temporal matching schemes under the mask-free training setting on YTVIS2019 val. 'Param' indicates whether the matching scheme brings extra model parameters. + +
Temporal Matching SchemeParamAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
Baseline (No Matching)✘38.665.938.838.447.7
Flow-based Matching✓40.266.341.940.549.1
Temporal Deformable Matching✓39.665.940.139.948.6
Learnable Pixel Matching✓39.565.740.239.748.4
Learnable Patch Matching✓40.666.542.640.349.2
3D Pairwise Matching✘39.465.041.740.248.0
Temporal KNN-Matching (Ours)✘42.566.845.741.251.2
+ +the impact of individual proposed components. Moreover, we study several alternative solutions for temporal matching and influence of different hyper-parameters to TK-Loss. + +Comparison on Temporal Matching Schemes Table 2 compares our Temporal KNN-Matching to four alternative frame-wise matching approaches for enforcing temporal mask consistency. Flow-based Matching employs the pretrained optical flow model RAFT [45] to build pixel correspondence [32]. Temporal Deformable Matching adopts the temporal deformable kernels [47] to predict the pixel offsets between the target and alignment frame. Instead of using raw patches, Learnable Pixel/Patch Matching employs jointly learned deep pixel/patch embeddings via three learnable FC layers, which are then used to compute the affinities in a soft attention-like manner. 3D Pairwise Matching directly extends $\mathcal{L}_{\mathrm{pair}}$ designed for spatial images to the temporal dimension, where pairwise affinity loss is computed among pixels not only in within the frame but also across multiple frames. + +In Table 2, compared to Flow-based Matching with one-to-one pixel correspondence, our paramter-free Temporal KNN-Matching with one-to- $K$ improves by about 2.3 AP. The prediction of flow-based models are not reliable in case of occlusions and homogeneous regions, and are also influenced by the gap between the training dataset and realworld video data. For the above mentioned deformable and learnable matching schemes, since there are only weak bounding box labels during training, the temporal matching relation is learnt implicitly. We empirically observe the instability during training with limited improvement under the mask-free training setting. For direct generalization of $\mathcal{L}_{\mathrm{pair}}$ , it only leads to 0.8 mask AP performance improvement. Despite the simplicity and efficiency of the TK-loss, it significantly improves VIS performance by 3.9 mask AP. Effect of Temporal KNN-patch Loss MaskFreeVIS is trained with joint spatio-temporal losses. In Table 3, to evaluate the effectiveness of each loss component, we compare the performance of MaskFreeVIS solely under the spatial pairwise loss [50] or our proposed TK-Loss. Compared to the 2.0 mask AP improvement by the spatial pairwise loss, the TK-Loss substantially promotes the mask AP from 36.6 to 41.6, showing the advantage of our flexible one-to- $K$ patch correspondence design in leveraging temporal consistency. We show the VIS results in Figure 5 to visualize + +Table 3. Effect of the Spatial Pairwise loss and our Temporal KNN-patch Loss on YTVIS2019 val. + +
Box ProjPairwiseTK-LossAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
✓36.666.536.237.145.0
✓✓\( 38.6_{\uparrow 2.0} \)65.938.838.447.7
✓✓\( 41.6_{\uparrow 5.0} \)68.443.540.150.5
✓✓✓\( 42.5_{\uparrow 5.9} \)66.845.741.251.2
+ +Table 4. Effect of using max K matches on YTVIS2019 val. + +
KAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)
140.865.844.140.3
341.966.945.141.9
542.566.845.741.2
742.367.144.640.6
+ +Table 6. Impact of search radius R on YTVIS2019 val. + +
RAPAP50AP75AR1
139.665.740.239.7
341.366.643.040.3
542.566.845.741.2
742.367.144.640.6
+ +Table 5. Patch Matching metrics comparison. NCC is Norm. Cross-Correlation. + +
MetricAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)
NCC41.766.743.441.4
\( L_1 \)41.266.243.640.3
\( L_2 \)42.566.845.741.2
+ +Table 7. Influence of patch size N on YTVIS2019 val. + +
NAPAP50AP75AR1
140.165.242.240.0
342.566.845.741.2
542.168.744.442.1
741.566.344.841.3
+ +the effectiveness of each loss component. + +Analysis of Temporal KNN-patch Loss In Table 4, we study the influence of $\mathbf{K}$ , the maximum number of matches selected in Temporal KNN-patch Loss. The best result is obtained for $K = 5$ , while $K = 1$ only allows for the one-to-one and no-match cases. The improvement from 40.8 mask AP to 42.5 mask AP reveals the benefit brought by the flexible one-to-many correspondence design. We also analyze the matching metric in Table 5, search radius in Table 6, and patch size in Table 7 (see Sec. 3.1.2 for details). When patch size $N$ is increased from 1 to 3 in Table 7, the performance of MaskFreeVIS is improved by 2.4 AP, validating the importance of patch structure in robust matching. Effect of the Cyclic Tube Connection We compare three frame-wise tube connection schemes (Figure 4) for the TK-Loss in Table 8. While dense connection brings forth the best performance, it doubles the training memory with minor improvement compared to Cyclic connection. Comparing to Sequential connection, our Cyclic connection benefits from long-range consistency, improving the performance of 0.6 mask AP with an affordable memory cost growth. + +Table 8. Comparison of the tube connection schemes, illustrated in Fig. 4. The tube consist of 5 frames. 'Mem' denotes the memory consumption per sampled video by the TK-Loss during training. + +
Tube ConnectConnect Num.Mem (MB)AP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)
Dense10152642.768.044.341.5
Sequential463141.966.544.741.2
Cyclic (Ours)577342.566.845.741.2
+ +Comparison on Sequence Matching Functions Besides TK-Loss, we analyze the influence of sequence matching cost functions for transformer-based VIS methods under the mask-free setting. In Table 9, we identify the substantial advantage of Spatio-temporal box-mask matching over frameworkwise cost averaging [55, 57]. As discussed in Sec. 3.2.2, we + +![](images/43bdb206943dc1d9df06ea1b12062d697ca0ec1a145beb2c8409267a3f09bfbd.jpg) + +![](images/c4fbc8d99e6d880551ba0ed3394635374723593ad0f2d7dd968f15ad2959c850.jpg) +Mask2Fomer + BoxInst (Baseline) + +![](images/7655795ebbce2315e34dabc4215aba1261f7029ae455aefd099e24f42776c1f6.jpg) + +![](images/225a7b23ffb2341280d298fd5411a9d9e393c666f883ac2ba332283facf3da24.jpg) +Mask2Fomer + TK-Loss (Ours) + +![](images/b0076021faf8322db5ee9e859464a1c2998f65a9528ecd1bdf7c9062a7600c84.jpg) + +![](images/fa94a76cb5bf4e5b7e56ac8a3fcf338204a41db130dafd40a61e6958f4089a65.jpg) +Mask2Former (Oracle) + +![](images/78f10297b12cc2e0186191abe0d636b481f6a79e613d2cad1a89296e4b5e03f2.jpg) +Figure 5. Qualitative results comparison between using Spatial Pairwise loss [50], our TK-Loss, and Mask2Former (oracle) trained with GT video and image masks. +Figure 6. Results on YTVIS 2019 val with various percentages of the YTVIS training data. Baseline denotes Mask2Former [6] trained with GT video boxes using BoxInst [50], while Oracle denotes Mask2Former trained with GT video masks. + +![](images/c8f096fdf99645268ff2e1f6c2356be87722b38c12a81eea94d9428d3442e35d.jpg) + +Table 9. Comparison of Set Matching Cost Functions on YTVIS2019 val. ST-BoxMask denotes our Spatio-temporal Box Mask matching. w/o CE denotes removing cross-entropy cost. + +
Matching Cost FunctionAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
Frame-wise Averaging [55, 57]37.664.239.537.545.7
ST-BoxMask40.867.842.240.049.2
ST-BoxMask w/o CE42.566.845.741.251.2
+ +achieve further gain by removing the object size imbalanced cross-entropy cost computation. + +Training on Various Amounts of Data To further study label-efficient VIS, we validate the effect of MaskFreeVIS under various percentages of the YTVIS 2019 training data. We uniformly sample frames and their labels for each video, and set the minimum sampled number of frames to 1. Figure 6 presents the experimental results, which shows the consistent large improvement (over 3.0 AP) brought by our TK-Loss under various amounts of data. In particular, we note that our approach with $50\%$ data even outperforms the fully-supervised model with $10\%$ training data. + +# 4.4. Comparison with State-of-the-art Methods + +We compare MaskFreeVIS with the state-of-the-art fully/weakly supervised methods on benchmarks YTVIS 2019/2021, OVIS and BDD100K MOTS. We integrate MaskFreeVIS on four representative methods [6,12,57,60], attaining consistent large gains over the strong baselines. + +YTVIS 2019/2021 Table 10 compares the performance on YTVIS 2019. Using R50/R101 as backbone and with the same training setting, MaskFreeVIS achieves 42.5/45.8 AP, improving 3.9/5.0 AP over the strong baseline adopt + +Table 10. Comparison on YTVIS 2019. I: using COCO mask pretrained model as initialization. V: using YTVIS video masks during training. $\ast$ : using pseudo mask from COCO images for joint training [57]. M2F: Mask2Former [6], SeqF: SeqFormer [57]. + +
MethodMask ann.Back-boneAP\( AP_{50} \)\( AP_{75} \)\( AR_{1} \)\( AR_{10} \)
Fully-supervised:
PCAN [18]I+VR5036.154.939.436.341.6
EfficientVIS [59]I+VR5037.959.743.040.346.6
InsPro [10]I+VR5040.262.943.137.644.5
IFC [15]I+VR5042.865.846.843.851.2
VMT* [17]I+VR5047.9-52.045.8-
SeqF* [57]I+VR5047.469.851.845.554.8
M2FI+VR5046.468.050.0--
M2F*I+VR5047.869.252.746.256.6
Prev. Weakly-supervised:
FlowIRN [32]-R5010.527.26.212.313.6
SOLO-Track [9]IR5030.650.733.531.637.1
Mask-free:
M2F + Flow Consist [45]-R5040.266.341.940.549.1
M2F + BoxInst [50]-R5038.664.238.538.046.8
M2F + MaskFreeVIS-R5042.5↑3.966.845.741.251.2
M2F + MaskFreeVISIR5043.8↑5.270.746.941.552.3
M2F + MaskFreeVIS*IR5046.6↑8.072.549.744.955.7
Fully-supervised:
M2FVR10145.672.648.944.354.5
M2FI+VR10149.272.854.2--
M2F*I+VR10149.873.655.448.058.0
SeqF*I+VR10149.071.155.746.856.9
Mask-free:
M2F + BoxInst [50]-R10140.867.842.240.049.2
M2F + MaskFreeVIS-R10145.8↑5.070.848.645.355.2
M2F + MaskFreeVISIR10147.3↑6.575.449.944.655.2
M2F + MaskFreeVIS*IR10148.9↑8.174.954.744.957.0
SeqF + MaskFreeVIS*IR10148.674.052.245.957.2
Fully-supervised:
M2FI+VSwinL60.484.467.0--
Mask-free:
M2F + BoxInst [50]-SwinL49.873.255.548.258.1
M2F + MaskFreeVIS-SwinL54.3↑4.582.661.150.261.3
M2F + MaskFreeVIS*ISwinL55.3↑5.582.560.850.762.2
+ +ing BoxInst [50] losses. MaskFreeVIS, without any mask labels, even significantly outperforms some recent fully-supervised methods such as EfficientVIS [59] and In-sPro [10]. On R50/R101/Swin-L, our MaskFreeVIS consistently attains over $91\%$ of its fully-supervised counterpart trained with both GT image and video masks. We also observe similar larger performance growth over the baseline on YTVIS 2021 in Table 11. The excellent performance substantially narrows the performance gap between fully-supervised and weakly-supervised VIS. + +OVIS We also conduct experiments on OVIS in Table 12 using R50 as backbone. We integrate MaskFreeVIS with VITA [12], promoting the baseline performance from 12.1 to 15.7 under the mask-free training setting. + +BDD100K MOTS Table 13 further validates our approach on BDD100K MOTS. Integrated with Unicorn [60], Mask-FreeVIS achieves $23.8\mathrm{mMOTSA}$ by improving over 4.9 points over the strong baseline and thus surpassing the fully + +Table 11. Comparison on YTVIS 2021. Refer to Table 10 for the symbol abbreviations. + +
MethodMask ann.Back-boneAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
Fully-supervised:
MaskTrack [61]I+VR5028.648.929.626.533.8
IFC [15]I+VR5036.657.939.3--
SeqF* [57]I+VR5040.562.443.736.148.1
M2FI+VR5040.660.941.8--
Mask-free:
M2F + BoxInst [50]-R5032.152.834.431.038.1
M2F + MaskFreeVIS-R5036.2↑4.160.839.234.645.6
M2F + MaskFreeVISIR5037.2↑5.161.940.335.346.1
M2F + MaskFreeVIS*IR5040.9↑8.865.843.337.150.5
Fully-supervised:
M2FI+VR10142.465.945.8--
Mask-free:
M2F + BoxInst [50]-R10133.355.232.532.141.9
M2F + MaskFreeVIS-R10137.3↑4.061.639.434.145.6
M2F + MaskFreeVISIR10138.2↑4.962.440.034.946.2
M2F + MaskFreeVIS*IR10141.6↑8.366.244.836.349.2
+ +Table 12. State-of-the-art comparison on the OVIS using R50. + +
MethodAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
Fully-supervised:
CrossVIS [62]14.932.712.110.319.8
Mask2Former [6]17.337.315.110.523.5
VMT [17]16.936.413.710.422.7
VITA [12]19.641.217.411.726.0
Video Mask-free:
VITA [12] + BoxInst [50]12.128.310.28.817.9
VITA [12] + MaskFreeVIS\( 15.7_{\uparrow 3,6} \)35.113.110.120.4
+ +Table 13. State-of-the-art comparison on the BDD100K segmentation tracking validation set. + +
MethodmMOTSA↑mMOTSP↑mIDF↑ID sw.↓mAP↑
Fully-supervised:
STEm-Seg [1]12.258.225.4873221.8
QDTrack-mots-fix [36]23.566.344.597325.5
PCAN [18]27.466.745.187626.6
Unicorn [60]29.667.744.2173132.1
Video Mask-free:
Unicorn [60] + BoxInst [50]18.958.736.3329822.1
Unicorn [60] + MaskFreeVIS23.8↑4.966.744.9208624.8
+ +supervised QDTrack-mots [36]. The consistent large gains on four benchmarks and four base VIS methods validate the generalizability of our MaskFreeVIS. + +# 5. Conclusion + +MaskFreeVIS is the first competitive VIS method that does not need any mask annotations during training. The strong results lead to a remarkable conclusion: mask labels are not a necessity for high-performing VIS. Our key component is the unsupervised Temporal KNN-patch Loss, which replaces the conventional video masks losses by leveraging temporal mask consistency constraints. Our approach greatly reduces the long-standing gap between fully-supervised and weakly-supervised VIS on four large-scale benchmarks. MaskFreeVIS thus opens up many opportunities for researchers and practitioners for label-efficient VIS. + +# References + +[1] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, 2020. 3, 8 +[2] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, 2020. 2 +[3] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, 2019. 2 +[4] Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In ECCV, 2020. 2 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2 +[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 2, 3, 5, 6, 7, 8 +[7] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In CVPR, 2022. 1, 3 +[8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In ICCV, 2015. 3 +[9] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to track instances without video annotations. In CVPR, 2021. 3, 8 +[10] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. In NeurIPS, 2022. 2, 8 +[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2 +[12] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 7, 8 +[13] Cheng-Chun Hsu, Kuang-Jui Hsu, Chung-Chi Tsai, Yen-Yu Lin, and Yung-Yu Chuang. Weakly supervised instance segmentation using the bounding box tightness prior. In NeurIPS, 2019. 1, 3, 5 +[14] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 3 +[15] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. In NeurIPS, 2021. 2, 8 +[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In CVPR, 2022. 3 +[17] Lei Ke, Henghui Ding, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Video mask transfiner for high-quality video instance segmentation. In ECCV, 2022. 1, 2, 3, 8 +[18] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks + +for multiple object tracking and segmentation. In NeurIPS, 2021. 2, 3, 8 +[19] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2 +[20] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In CVPR, 2017. 3 +[21] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In ICLR, 2022. 3 +[22] Viveka Kulharia, Siddhartha Chandra, Amit Agrawal, Philip Torr, and Ambrish Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In ECCV, 2020. 3 +[23] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In ICCV, 2021. 1, 3 +[24] Jungbeom Lee, Eunji Kim, Sungmin Lee, Jangho Lee, and Sungroh Yoon. Frame-to-frame aggregation of active regions in web videos for weakly supervised semantic segmentation. In ICCV, 2019. 3 +[25] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In CVPR, 2021. 3 +[26] Minghan Li, Shuai Li, Lida Li, and Lei Zhang. Spatial feature calibration and temporal fusion for effective one-stage video instance segmentation. In CVPR, 2021. 2 +[27] Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, XianSheng Hua, and Lei Zhang. Box-supervised instance segmentation with level set evolution. In ECCV, 2022. 1, 3 +[28] Chung-Ching Lin, Ying Hung, Rogerio Feris, and Linglin He. Video instance segmentation tracking with a modified vae architecture. In CVPR, 2020. 2 +[29] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In CVPR, 2021. 2 +[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2 +[31] Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In CVPR, 2021. 2 +[32] Qing Liu, Vignesh Ramanathan, Dhruv Mahajan, Alan Yuille, and Zhenheng Yang. Weakly supervised instance segmentation for videos with temporal mask consistency. In CVPR, 2021. 2, 3, 6, 8 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 +[34] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 3 +[35] Anton Milan, Laura Leal-Taixe, Konrad Schindler, and Ian Reid. Joint tracking and segmentation of multiple targets. In CVPR, 2015. 3 + +[36] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 8 +[37] George Papandreou, Liang-Chieh Chen, Kevin P Murphy, and Alan L Yuille. Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In ICCV, 2015. 3 +[38] Jordi Pont-Tuset, Pablo Arbelaez, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping for image segmentation and object proposal generation. IEEE TPAMI, 39(1):128-140, 2016. 3 +[39] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 2021. 1, 2, 6 +[40] Martin Rajchl, Matthew CH Lee, Ozan Oktay, Konstantinos Kamnitsas, Jonathan Passerat-Palmbach, Wenjia Bai, Mellisa Damodaram, Mary A Rutherford, Joseph V Hajnal, Bernhard Kainz, et al. Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging, 36(2):674-683, 2016. 3 +[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 5 +[42] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. Grabcut: interactive foreground extraction using iterated graph cuts. ACM TOG, 23(3):309-314, 2004. 3 +[43] Fatemeh Sadat Saleh, Mohammad Sadegh Aliakbarian, Mathieu Salzmann, Lars Petersson, and Jose M Alvarez. Bringing background into the foreground: Making all classes equal in weakly-supervised video semantic segmentation. In ICCV, 2017. 3 +[44] Chunfeng Song, Yan Huang, Wanli Ouyang, and Liang Wang. Box-driven class-wise region masking and filling rate guided loss for weakly supervised semantic segmentation. In CVPR, 2019. 3 +[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 6, 8 +[46] Omkar Thawakar, Sanath Narayan, Jiale Cao, Hisham Cholakkal, Rao Muhammad Anwer, Muhammad Haris Khan, Salman Khan, Michael Felsberg, and Fahad Shahbaz Khan. Video instance segmentation via multi-scale spatiotemporal split attention transformer. In ECCV, 2022. 2 +[47] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. 6 +[48] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, 2020. 2, 3, 5 +[49] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 2 +[50] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In CVPR, 2021. 1, 3, 5, 6, 7, 8 +[51] Pavel Tokmakov, Karteek Alahari, and Cordelia Schmid. Learning semantic segmentation with weakly-annotated videos. In ECCV, 2016. 3 + +[52] Yi-Hsuan Tsai, Guangyu Zhong, and Ming-Hsuan Yang. Semantic co-segmentation in videos. In ECCV, 2016. 3 +[53] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 3 +[54] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2 +[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 5, 7 +[56] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In CVPR, 2021. 3 +[57] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. In ECCV, 2022. 1, 2, 3, 5, 6, 7, 8 +[58] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 3 +[59] Jialian Wu, Sudhir Yarram, Hui Liang, Tian Lan, Junsong Yuan, Jayan Eledath, and Gerard Medioni. Efficient video instance segmentation via tracklet query and proposal. In CVPR, 2022. 2, 8 +[60] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 3, 7, 8 +[61] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 2, 5, 6, 8 +[62] Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In ICCV, 2021. 2, 8 +[63] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In CVPR, 2022. 2 +[64] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 1, 2, 6 \ No newline at end of file diff --git a/2023/Mask-Free Video Instance Segmentation/images.zip b/2023/Mask-Free Video Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8111991e3afedae72ec2d6d715a47ea1a8e087e --- /dev/null +++ b/2023/Mask-Free Video Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6680aba6872e248d240f00f59adacacc201ea8242fe9c84e503e7e28c2f1dc83 +size 762351 diff --git a/2023/Mask-Free Video Instance Segmentation/layout.json b/2023/Mask-Free Video Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..14b361ee92c2bf5b7b18e68184f1634f015b1314 --- /dev/null +++ b/2023/Mask-Free Video Instance Segmentation/layout.json @@ -0,0 +1,9874 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 96, + 423, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 96, + 423, + 113 + ], + "spans": [ + { + "bbox": [ + 170, + 96, + 423, + 113 + ], + "type": "text", + "content": "Mask-Free Video Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "spans": [ + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "text", + "content": "Lei Ke" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "text", + "content": " Martin Danelljan" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "text", + "content": " Henghui Ding" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "text", + "content": " Yu-Wing Tai" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "text", + "content": " Chi-Keung Tang" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "text", + "content": " Fisher Yu" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "text", + "content": " ETH Zürich " + }, + { + "bbox": [ + 67, + 135, + 525, + 163 + ], + "type": "inline_equation", + "content": "{}^{2}\\mathrm{HKUST}" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 186, + 544, + 319 + ], + "blocks": [ + { + "bbox": [ + 50, + 186, + 544, + 319 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 544, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 544, + 319 + ], + "type": "image", + "image_path": "2df21adc52ff236309c7cb72bd869ca6ddfe29f15f25f57a188924dfbe5cd526.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 323, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 546, + 357 + ], + "type": "text", + "content": "Figure 1. Video instance segmentation (VIS) results of our MaskFreeVIS, trained without using any video or image mask annotation. By achieving a remarkable " + }, + { + "bbox": [ + 46, + 323, + 546, + 357 + ], + "type": "inline_equation", + "content": "42.5\\%" + }, + { + "bbox": [ + 46, + 323, + 546, + 357 + ], + "type": "text", + "content": " mask AP on the YouTube-VIS val dataset, with a ResNet-50 backbone, our approach demonstrates that high-performing VIS can be learned even without any mask annotations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 372, + 192, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 372, + 192, + 384 + ], + "spans": [ + { + "bbox": [ + 143, + 372, + 192, + 384 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 410, + 289, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 289, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 289, + 709 + ], + "type": "text", + "content": "The recent advancement in Video Instance Segmentation (VIS) has largely been driven by the use of deeper and increasingly data-hungry transformer-based models. However, video masks are tedious and expensive to annotate, limiting the scale and diversity of existing VIS datasets. In this work, we aim to remove the mask-annotation requirement. We propose MaskFreeVIS, achieving highly competitive VIS performance, while only using bounding box annotations for the object state. We leverage the rich temporal mask consistency constraints in videos by introducing the Temporal KNN-patch Loss (TK-Loss), providing strong mask supervision without any labels. Our TK-Loss finds one-to-many matches across frames, through an efficient patch-matching step followed by a " + }, + { + "bbox": [ + 46, + 410, + 289, + 709 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 410, + 289, + 709 + ], + "type": "text", + "content": "-nearest neighbor selection. A consistency loss is then enforced on the found matches. Our mask-free objective is simple to implement, has no trainable parameters, is computationally efficient, yet outperforms baselines employing, e.g., state-of-the-art optical flow to enforce temporal mask consistency. We validate MaskFreeVIS on the YouTube-VIS 2019/2021, OVIS and BDD100K MOTS benchmarks. The results clearly demonstrate the efficacy of our method by drastically narrowing the gap between fully and weakly-supervised VIS performance. Our code and trained models are available at http://vis.xyz/pub/maskfreevis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 372, + 386, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 386, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 386, + 384 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 396, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 575 + ], + "type": "text", + "content": "Video Instance Segmentation (VIS) requires jointly detecting, tracking and segmenting all objects in a video from a given set of categories. To perform this challenging task, state-of-the-art VIS models are trained with complete video annotations from VIS datasets [39, 61, 64]. However, video annotation is costly, in particular regarding object mask labels. Even coarse polygon-based mask annotation is multiple times slower than annotating video bounding boxes [7]. Expensive mask annotation makes existing VIS benchmarks difficult to scale, limiting the number of object categories covered. This is particularly a problem for the recent transformer-based VIS models [6, 17, 57], which tend to be exceptionally data-hungry. We therefore revisit the need for complete mask annotation by studying the problem of weakly supervised VIS under the mask-free setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 578, + 548, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 548, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 548, + 723 + ], + "type": "text", + "content": "While there exist box-supervised instance segmentation models [13, 23, 27, 50], they are designed for images. These weakly-supervised single-image methods do not utilize temporal cues when learning mask prediction, leading to lower accuracy when directly applied to videos. As a source for weakly supervised learning, videos contain much richer information about the scene. In particular, videos adhere to the temporal mask consistency constraint, where the regions corresponding to the same underlying object across different frames should have the same mask label. In this work, we set out to leverage this important constraint for mask-free learning of VIS." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "22857" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 65, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 107, + 57, + 235, + 65 + ], + "lines": [ + { + "bbox": [ + 107, + 57, + 235, + 65 + ], + "spans": [ + { + "bbox": [ + 107, + 57, + 235, + 65 + ], + "type": "text", + "content": "Temporal One-to-K Patch Correspondence" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 65, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 49, + 65, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 49, + 65, + 286, + 167 + ], + "type": "image", + "image_path": "ad65f7aa1ab4c1a9f2dff2f18954500f56a51c3dda909f4eac4ed3f9e17ef030.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 257 + ], + "type": "text", + "content": "Figure 2. Our Temporal KNN-patch Loss enforces mask consistency between one-to-" + }, + { + "bbox": [ + 46, + 168, + 287, + 257 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 168, + 287, + 257 + ], + "type": "text", + "content": " patch correspondences found across frames, which allow us to cover the cases where: (i) A unique one-to-one match exists (blue); (ii) Multiple matches are found due to ambiguities in homogenous regions (orange) or along image edges (white and yellow); (iii) No match is found due to " + }, + { + "bbox": [ + 46, + 168, + 287, + 257 + ], + "type": "inline_equation", + "content": "e.g." + }, + { + "bbox": [ + 46, + 168, + 287, + 257 + ], + "type": "text", + "content": " occlusions (green). This allows us to robustly leverage mask consistency constraints in challenging videos." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "type": "text", + "content": "We propose MaskFreeVIS method, for high performance VIS without any mask annotations. To leverage temporal mask consistency, we introduce the Temporal KNN-patch Loss (TK-Loss), as in Figure 2. To find regions corresponding to the same underlying video object, our TK-Loss first builds correspondences across frames by patch-wise matching. For each target patch, only the top " + }, + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "type": "text", + "content": " matches in the neighboring frame with high enough matching score are selected. A temporal consistency loss is then applied to all found matches to promote the mask consistency. Specifically, our surrogate objective function not only promotes the one-to- " + }, + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "type": "text", + "content": " matched regions to reach the same mask probabilities, but also commits their mask prediction to a confident foreground or background prediction by entropy minimization. Unlike flow-based models [32,45], which assume one-to-one matching, our approach builds robust and flexible one-to- " + }, + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 263, + 287, + 491 + ], + "type": "text", + "content": " correspondences to cope with e.g. occlusions and homogeneous regions, without introducing additional model parameters or inference cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 492, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 589 + ], + "type": "text", + "content": "The TK-Loss is easily integrated into existing VIS methods, with no architecture modifications required. During training, our TK-Loss simply replaces the conventional video mask losses in supervising video mask generation. To further enforce temporal consistency through the video clip, TK-Loss is employed in a cyclic manner instead of using dense frame-wise connections. This greatly reduces memory cost with negligible performance drop." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 590, + 288, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 288, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 288, + 723 + ], + "type": "text", + "content": "We extensively evaluate our MaskFreeVIS on four large-scale VIS benchmarks, i.e., YouTube-VIS 2019/2021 [61], OVIS [39], and BDD100K MOTS [64]. MaskFreeVIS achieves competitive VIS performance without using any video masks or even image mask labels on all datasets. Validated on various methods and backbones, MaskFreeVIS achieves " + }, + { + "bbox": [ + 46, + 590, + 288, + 723 + ], + "type": "inline_equation", + "content": "91.25\\%" + }, + { + "bbox": [ + 46, + 590, + 288, + 723 + ], + "type": "text", + "content": " performance of its fully supervised counterparts, even outperforming a few recent fully-supervised methods [10, 15, 18, 59] on the popular YTVIS benchmark. Our simple yet effective design greatly narrows the performance gap between weakly-supervised and fully" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 63, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 63, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 63, + 545, + 152 + ], + "type": "text", + "content": "Table 1. Mask annotation requirement for state-of-the-art VIS methods. Results are reported using ResNet-50 as backbone on the YTVIS 2019 [61] benchmark. Video Mask: using YTVIS video mask labels. Image Mask: using COCO [30] image mask labels for image-based pretraining. Pseudo Video: using Pseudo Videos from COCO images for joint training [57]. MaskFreeVIS achieves " + }, + { + "bbox": [ + 305, + 63, + 545, + 152 + ], + "type": "inline_equation", + "content": "91.5\\%" + }, + { + "bbox": [ + 305, + 63, + 545, + 152 + ], + "type": "text", + "content": " (42.5 vs. 46.4) of its fully-supervised baseline performance (Mask2Former) without using any masks during training." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 153, + 544, + 253 + ], + "blocks": [ + { + "bbox": [ + 307, + 153, + 544, + 253 + ], + "lines": [ + { + "bbox": [ + 307, + 153, + 544, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 544, + 253 + ], + "type": "table", + "html": "
MethodVideo MaskImage MaskPseudo VideoAP
SeqFormer [57]✓✓✓47.4
VMT [17]✓✓✓47.9
Mask2Former [6]✓✓✓47.8
MaskFreeVIS (ours)✗✓✓46.6
Mask2Former [6]✓✓✗46.4
MaskFreeVIS (ours)✗✗✗42.5
", + "image_path": "1ff69b0913aabd0b5af8a7762d33d57189dba1c6801dc2c0cb43fd812c2ac214.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 258, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 294 + ], + "type": "text", + "content": "supervised video instance segmentation. It further demonstrates that expensive video masks, or even image masks, is not necessary for training high-performing VIS models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "type": "text", + "content": "Our contributions are summarized as follows: (i) To utilize temporal information, we develop a new parameter-free Temporal KNN-patch Loss, which leverages temporal masks consistency using unsupervised one-to-" + }, + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "type": "text", + "content": " patch correspondence. We extensively analyze the TK-Loss through ablative experiments. (ii) Based on the TK-Loss, we develop the MaskFreeVIS method, enabling training existing state-of-the-art VIS models without any mask annotation. (iii) To the best of our knowledge, MaskFreeVIS is the first mask-free VIS method attaining high-performing segmentation results. We provide qualitative results in Figure 1. As in Table 1, when integrated into the Mask2Former [6] baseline with ResNet-50, our MaskFreeVIS achieves " + }, + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "type": "inline_equation", + "content": "42.5\\%" + }, + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "type": "text", + "content": " mask AP on the challenging YTVIS 2019 benchmark while using no video or image mask annotations. Our approach further scales to larger backbones, achieving " + }, + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "type": "inline_equation", + "content": "55.3\\%" + }, + { + "bbox": [ + 304, + 295, + 546, + 497 + ], + "type": "text", + "content": " mask AP on Swin-L backbone with no video mask annotations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 546 + ], + "type": "text", + "content": "We hope our approach will facilitate achieving label-efficient video instance segmentation, enabling building even larger-scale VIS benchmarks with diverse categories by lifting the mask annotation restriction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 558, + 392, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 392, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 392, + 571 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 578, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 723 + ], + "type": "text", + "content": "Video Instance Segmentation (VIS) Existing VIS methods can be summarized into three categories: two-stage, one-stage, and transformer-based. Two-stage approaches [2, 18, 28, 29, 61] extend the Mask R-CNN family [11, 19] by designing an additional tracking branch for object association. One-stage works [4, 26, 31, 62] adopt anchor-free detectors [49], generally using linear masks basis combination [3] or conditional mask prediction generation [48]. For the transformer-based models [6, 12, 46, 57, 63], VisTr [54] firstly adapts the transformer [5] for VIS, and IFC [15] further improves its efficiency via memory tokens. Seqformer [57] proposes frame query decompo" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "22858" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 65, + 286, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 65, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 46, + 65, + 286, + 148 + ], + "type": "text", + "content": "sition while Mask2Former [6] includes masked attention. VMT [17] extends Mask Transfiner [16] to video for high-quality VIS, and IDOL [58] focuses on online VIS. State-of-the-art VIS methods with growing capacity put limited emphasis on weak supervision. In contrast, the proposed MaskFreeVIS is the first method targeting mask-free VIS while attaining competitive performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 149, + 286, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 286, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 286, + 268 + ], + "type": "text", + "content": "Multiple Object Tracking and Segmentation (MOTS) Most MOTS methods [1,34,35,53,56] follow the tracking-by-detection principle. PCAN [18] improves temporal segmentation by utilizing space-time memory prototypes, while the one-stage method Unicorn [60] focuses on unification of different tracking frameworks. Compared to the aforementioned fully-supervised MOTS methods, Mask-FreeVIS focuses on label efficient training without GT masks by proposing a new surrogate temporal loss which can be easily integrated on them." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 270, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 286, + 376 + ], + "type": "text", + "content": "Mask-Free VIS Most mask-free instance segmentation works [7, 13, 20, 25, 27, 37, 40, 44] are designed for single images and thus neglect temporal information. Earlier works BoxSup [8] and Box2Seg [22] rely on region proposals produced by MCG [38] or GrabCut [42], leading to slow training. BoxInst [50] proposes the surrogate projection and pixel pairwise losses to replace the original mask learning loss of CondInst [48], while DiscoBox [23] focuses on generating pseudo mask labels guided by a teacher model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 378, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 286, + 628 + ], + "type": "text", + "content": "Earlier works have investigated the use of videos for weakly-, semi-, or un-supervised segmentation by leveraging motion or temporal consistency [21,51,52]. Most aforementioned approaches do not address the VIS problem, and use optical flow for frame-to-frame matching [24, 32, 43]. In particular, FlowIRN [32] explores VIS using only classification labels and incorporates optical flow to leverage mask consistency. The limited performance makes the class-label only or fully-unsupervised setting difficult to deploy in the real world. SOLO-Track [9] aims to train VIS models without video annotations, and one concurrent work MinVIS [14] performs VIS without video-based model architectures. Unlike the above weakly-supervised training settings, our MaskFreeVIS is designed for eliminating the mask annotation requirement for VIS, as we note that video mask labeling is particularly expensive. MaskFreeVIS enables training VIS models without any video masks, or even image masks. Despite its simplicity, MaskFreeVIS drastically reduces the gap between fully-supervised and weakly-supervised VIS models, making weakly-supervised models more accessible in practice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 642, + 102, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 102, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 102, + 653 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 662, + 286, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 662, + 286, + 722 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 286, + 722 + ], + "type": "text", + "content": "We propose MaskFreeVIS to tackle video instance segmentation (VIS) without using any video or even image mask labels. Our approach is generic and can be directly applied to train state-of-the-art VIS methods, such as Mask2Former [6] and SeqFormer [57]. In Sec. 3.1, we first" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 65, + 545, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 65, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 304, + 65, + 545, + 172 + ], + "type": "text", + "content": "present the core component of MaskFreeVIS: the Temporal KNN-patch Loss (TK-Loss), which leverages temporal consistency to supervise accurate mask prediction, without any human mask annotations. In Sec. 3.2, we then describe how to integrate the TK-Loss with existing spatial weak segmentation losses for transformer-based VIS methods, to achieve mask-free training of VIS approaches. Finally, we introduce image-based pretraining details of our MaskFreeVIS in Sec. 3.2.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 180, + 394, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 394, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 394, + 192 + ], + "type": "text", + "content": "3.1. MaskFreeVIS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 198, + 545, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 246 + ], + "type": "text", + "content": "In this section, we introduce the Temporal KNN-patch Loss, illustrated in Figure 3. It serves as an unsupervised objective for mask prediction that leverages the rich spatiotemporal consistency constraints in unlabelled videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 261, + 460, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 261, + 460, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 460, + 273 + ], + "type": "text", + "content": "3.1.1 Temporal Mask Consistency" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "content": "While an image constitutes a single snapshot of a scene, a video provides multiple snapshots displaced in time. Thereby, a video depicts continuous change in the scene. Objects and background move, deform, are occluded, experience variations in lighting, motion blur, and noise, leading to a sequence of different images that are closely related through gradual transformations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 363, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 506 + ], + "type": "text", + "content": "Consider a small region in the scene (Fig. 2), belonging either to an object or background. The pixels corresponding to the projection of this region should have the same mask prediction in every frame, as they belong to the same underlying physical object or background region. However, the aforementioned dynamic changes in the video lead to substantial appearance variations, serving as a natural form of data augmentation. The fact that the pixels corresponding to the same underlying object region should have the same mask prediction under temporal change therefore provides a powerful constraint, i.e., temporal mask consistency, which can be used for mask supervision [21, 24, 32, 51, 52]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 507, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 626 + ], + "type": "text", + "content": "The difficulty in leveraging the temporal mask consistency constraint stems from the problem of establishing reliable correspondences between video frames. Objects often undergo fast motion, deformations, etc., resulting in substantial appearance change. Furthermore, regions in the scene may be occluded or move out of the image from one frame to the other. In such cases, no correspondence exist. Lastly, videos are often dominated by homogenous regions, such as sky and ground, where the establishment of one-to-one correspondences are error-prone or even ill-defined." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 627, + 545, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 627, + 545, + 722 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 545, + 722 + ], + "type": "text", + "content": "The problem of establishing dense one-to-one correspondences between subsequent video frames, known as optical flow, is a long-standing and popular research topic. However, when attempting to enforce temporal mask consistency through optical flow estimation [24,32,43], one encounters two key problems. 1) The one-to-one assumption of optical flow is not suitable in cases of occlusions, homogenous regions, and along single edges, where the corre" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "22859" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 49, + 138, + 171 + ], + "blocks": [ + { + "bbox": [ + 49, + 49, + 138, + 171 + ], + "lines": [ + { + "bbox": [ + 49, + 49, + 138, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 49, + 138, + 171 + ], + "type": "image", + "image_path": "08adf8978d71ca66c5814d3e778b63990cadabda627e958f8883c6a02009439f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 171, + 141, + 179 + ], + "lines": [ + { + "bbox": [ + 48, + 171, + 141, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 141, + 179 + ], + "type": "text", + "content": "(1) Temporal Candidates Extraction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "text", + "content": "Figure 3. Temporal KNN-patch Loss has four steps: 1) Patch Candidate Extraction: Patch candidates searching across frames with radius " + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "text", + "content": ". 2) Temporal KNN-Matching: Match " + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "text", + "content": " high-confidence candidates by patch affinities. 3) Consistency loss: Enforce mask consistency objective (Eq. 2) among the matches. 4) Cyclic Tube Connection: Temporal loss aggregation in the 5-frame tube, detailed in Figure 4." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 49, + 219, + 171 + ], + "blocks": [ + { + "bbox": [ + 141, + 49, + 219, + 171 + ], + "lines": [ + { + "bbox": [ + 141, + 49, + 219, + 171 + ], + "spans": [ + { + "bbox": [ + 141, + 49, + 219, + 171 + ], + "type": "image", + "image_path": "72a30751e06486c0c306acfc9f139e3694ae338d6b3e7a905476a6ab62e523b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 171, + 216, + 179 + ], + "lines": [ + { + "bbox": [ + 141, + 171, + 216, + 179 + ], + "spans": [ + { + "bbox": [ + 141, + 171, + 216, + 179 + ], + "type": "text", + "content": "(2) Temporal KNN-Matching" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 49, + 326, + 171 + ], + "blocks": [ + { + "bbox": [ + 220, + 49, + 326, + 171 + ], + "lines": [ + { + "bbox": [ + 220, + 49, + 326, + 171 + ], + "spans": [ + { + "bbox": [ + 220, + 49, + 326, + 171 + ], + "type": "image", + "image_path": "2e8f94bf5ff32f721a2b28f5d96974c30864adc36a2bddeadea1ef7b5be769c1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 171, + 272, + 178 + ], + "lines": [ + { + "bbox": [ + 234, + 171, + 272, + 178 + ], + "spans": [ + { + "bbox": [ + 234, + 171, + 272, + 178 + ], + "type": "text", + "content": "Loss Backward" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 49, + 425, + 171 + ], + "blocks": [ + { + "bbox": [ + 326, + 49, + 425, + 171 + ], + "lines": [ + { + "bbox": [ + 326, + 49, + 425, + 171 + ], + "spans": [ + { + "bbox": [ + 326, + 49, + 425, + 171 + ], + "type": "image", + "image_path": "9dc5c571ae48cdb53660ba47af13fcd4b616a62a40e9d40b586b29b3cddc1957.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 171, + 410, + 179 + ], + "lines": [ + { + "bbox": [ + 340, + 171, + 410, + 179 + ], + "spans": [ + { + "bbox": [ + 340, + 171, + 410, + 179 + ], + "type": "text", + "content": "(3) KNN-Consistency Loss" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 425, + 49, + 466, + 169 + ], + "blocks": [ + { + "bbox": [ + 425, + 49, + 466, + 169 + ], + "lines": [ + { + "bbox": [ + 425, + 49, + 466, + 169 + ], + "spans": [ + { + "bbox": [ + 425, + 49, + 466, + 169 + ], + "type": "image", + "image_path": "faa4326afa7e0473b380c95f8a1f02eefc5ed465e09785eb251695d39413c4b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 169, + 463, + 179 + ], + "lines": [ + { + "bbox": [ + 429, + 169, + 463, + 179 + ], + "spans": [ + { + "bbox": [ + 429, + 169, + 463, + 179 + ], + "type": "text", + "content": "4) Cyclic Tube Connection" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 467, + 50, + 544, + 171 + ], + "blocks": [ + { + "bbox": [ + 467, + 50, + 544, + 171 + ], + "lines": [ + { + "bbox": [ + 467, + 50, + 544, + 171 + ], + "spans": [ + { + "bbox": [ + 467, + 50, + 544, + 171 + ], + "type": "image", + "image_path": "77f665369b9c035734a800aeb31c05f322310ee968e66b657ddce88211558705.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 171, + 538, + 178 + ], + "lines": [ + { + "bbox": [ + 476, + 171, + 538, + 178 + ], + "spans": [ + { + "bbox": [ + 476, + 171, + 538, + 178 + ], + "type": "text", + "content": "MaskFreeVIS Prediction" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 222, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 287, + 270 + ], + "type": "text", + "content": "spondence is either nonexistent, undefined, ambiguous, uncertain, or very difficult to determine. 2) State-of-the-art optical flow estimation relies on large and complex deep networks, with large computational and memory requirements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 270, + 287, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 306 + ], + "type": "text", + "content": "Instead of using optical flow, we aim to design a simple, efficient, and parameter-free approach that effectively enforces the temporal mask consistency constraint." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 309, + 195, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 195, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 195, + 322 + ], + "type": "text", + "content": "3.1.2 Temporal KNN-patch Loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "text", + "content": "Our Temporal KNN-patch Loss (TK-Loss) is based on a simple and flexible correspondence estimation across frames. In contrast to optical flow, we do not restrict our formulation to one-to-one correspondences. Instead, we establish one-to- " + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "text", + "content": " correspondences. This include the conventional one-to-one " + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "inline_equation", + "content": "(K = 1)" + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "text", + "content": ", where a unique well-defined match exists. However, this allows us to also handle the cases of nonexistent correspondences " + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "inline_equation", + "content": "(K = 0)" + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "text", + "content": " in case of occlusions, and one-to-many " + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "inline_equation", + "content": "(K \\geq 2)" + }, + { + "bbox": [ + 46, + 328, + 287, + 531 + ], + "type": "text", + "content": " in case of homogenous regions. In cases where multiple matches are found, these most often belong to the same underlying object or background due to their similar appearance, as in Figure 2. This further benefits our mask consistency objective through denser supervision. Lastly, our approach is simple to implement, with negligible computational overhead and no learnable parameters. Our approach is in Figure 3, and contains four main steps, which are detailed next." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 531, + 287, + 675 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": "1) Patch Candidate Extraction: Let " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "X_p^t" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": " denote an " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": " target image patch centered at spatial location " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "p = (x, y)" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": ". Our aim is to find a set of corresponding positions " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "S_p^{t \\to \\hat{t}} = \\{\\hat{p}_i\\}_i" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": " in frame number " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": " that represent the same object region. To this end, we first select candidate locations " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": " within a radius " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\| p - \\hat{p} \\| \\leq R" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": ". Such windowed patch search exploits spatial proximity across neighboring frames in order to avoid an exhaustive global search. For a fast implementation, the windowed search is performed for all target image patches " + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "inline_equation", + "content": "X_p^t" + }, + { + "bbox": [ + 46, + 531, + 287, + 651 + ], + "type": "text", + "content": " in parallel." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 651, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 675 + ], + "type": "text", + "content": "2) Temporal KNN-Matching: We match patch candidate patches through a simple distance computation," + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 121, + 683, + 287, + 703 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 683, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 121, + 683, + 287, + 703 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {p \\rightarrow \\hat {p}} ^ {t \\rightarrow \\hat {t}} = \\left\\| X _ {p} ^ {t} - X _ {\\hat {p}} ^ {\\hat {t}} \\right\\|, \\tag {1}", + "image_path": "78ffae296d425ab36a411f0a22038a276a2ad16ef369e857e7614e2f65a9b94e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 710, + 287, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 710, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 59, + 710, + 287, + 723 + ], + "type": "text", + "content": "In our ablative experiments (Sec. 4.3), we found the " + }, + { + "bbox": [ + 59, + 710, + 287, + 723 + ], + "type": "inline_equation", + "content": "L_{2}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "text", + "content": "norm to be the most effective patch matching metric. We select the top " + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "text", + "content": " matches with smallest patch distance " + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{p\\rightarrow \\hat{p}}^{t\\rightarrow \\hat{t}}" + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "text", + "content": ". Lastly low-confidence matches are removed by enforcing a maximal patch distance " + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{p\\rightarrow \\hat{p}}^{t\\rightarrow \\hat{t}} < D" + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "text", + "content": ". The remaining matches form the set " + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "inline_equation", + "content": "S_{p}^{t\\rightarrow \\hat{t}} = \\{\\hat{p}_{i}\\}_{i}" + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "text", + "content": " for each location " + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 222, + 545, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "text", + "content": "3) Consistency loss: Let " + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "inline_equation", + "content": "M_p^t \\in [0,1]" + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "text", + "content": " denote the predicted binary instance mask of an object, evaluated at position " + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "text", + "content": ". To ensure temporal mask consistency constraints, we penalize inconsistent mask predictions between a spatiotemporal point " + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "inline_equation", + "content": "(p,t)" + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "text", + "content": " and its estimated corresponding points in " + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "inline_equation", + "content": "S_p^{t \\to \\hat{t}}" + }, + { + "bbox": [ + 304, + 285, + 545, + 361 + ], + "type": "text", + "content": ". In particular we use the following objective," + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 336, + 371, + 545, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 371, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 336, + 371, + 545, + 404 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} ^ {t \\rightarrow \\hat {t}} = \\frac {1}{H W} \\sum_ {p} \\sum_ {\\hat {p} _ {i} \\in \\mathcal {S} _ {p} ^ {t \\rightarrow \\hat {t}}} L _ {\\text {c o n s}} \\left(M _ {p} ^ {t}, M _ {\\hat {p} _ {i}} ^ {\\hat {t}}\\right), \\tag {2}", + "image_path": "59b78ef85f0a7dc60e799acb514c4aba42884bbbf5d9655a825d08e0fc98ff5a.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 415, + 467, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 467, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 467, + 426 + ], + "type": "text", + "content": "where mask consistency is measured as" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 310, + 437, + 545, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 437, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 310, + 437, + 545, + 455 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o n s}} \\left(M _ {p} ^ {t}, M _ {\\hat {p}} ^ {\\hat {t}}\\right) = - \\log \\left(M _ {p} ^ {t} M _ {\\hat {p}} ^ {\\hat {t}} + \\left(1 - M _ {p} ^ {t}\\right) \\left(1 - M _ {\\hat {p}} ^ {\\hat {t}}\\right)\\right). \\tag {3}", + "image_path": "39e803de391478b22897bfa550074ef973306a520ea1845afadbbfea770f49c9.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "type": "text", + "content": "Note that Eq. (3) only attains its minimum value of zero if both predictions indicate exactly background (" + }, + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "type": "inline_equation", + "content": "M_p^t = M_{\\hat{p}}^{\\hat{t}} = 0" + }, + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "type": "text", + "content": ") or foreground (" + }, + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "type": "inline_equation", + "content": "M_p^t = M_{\\hat{p}}^{\\hat{t}} = 1" + }, + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "type": "text", + "content": "). The objective does thus not only promote the two mask predictions to achieve the same probability value " + }, + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "type": "inline_equation", + "content": "M_p^t = M_{\\hat{p}}^{\\hat{t}}" + }, + { + "bbox": [ + 304, + 464, + 545, + 539 + ], + "type": "text", + "content": ", but also to commit to a certain foreground or background prediction." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": "4) Cyclic Tube Connection: Suppose the temporal tube consists of " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " frames. We compute the temporal loss for the whole tube in a cyclic manner, as in Figure 4. The start frame is connected to the end frame, which introduces direct long-term mask consistency across the two temporally most distant frames. The temporal TK-Loss for the whole tube is given by" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 344, + 635, + 545, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 635, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 344, + 635, + 545, + 664 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e m p}} = \\sum_ {t = 1} ^ {T} \\left\\{\\begin{array}{l l}\\mathcal {L} _ {f} ^ {t \\rightarrow (t + 1)}&\\text {i f} t < T - 1\\\\\\mathcal {L} _ {f} ^ {t \\rightarrow 0}&\\text {i f} t = T - 1.\\end{array}\\right. \\tag {4}", + "image_path": "f66ce6fda3ca01104b394b3ecf4af819c6cf36c741df4d8a1bf3ef2ce6bf7e2e.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 674, + 545, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 545, + 721 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 545, + 721 + ], + "type": "text", + "content": "Compared to inter-frame dense connections in Figure 4, we find the cyclic loss to achieve similar performance but greatly reduce the memory usage as validated in the experiment section." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "22860" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 59, + 282, + 178 + ], + "blocks": [ + { + "bbox": [ + 51, + 59, + 282, + 178 + ], + "lines": [ + { + "bbox": [ + 51, + 59, + 282, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 59, + 282, + 178 + ], + "type": "image", + "image_path": "025f71d424fdc899da266bf565e2003fb4f9a1955446db6c208414ca14289cce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 178, + 286, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 286, + 201 + ], + "type": "text", + "content": "Figure 4. Illustration of different frame-wise tube connection settings (connection number) in the temporal loss design." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 210, + 180, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 180, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 180, + 223 + ], + "type": "text", + "content": "3.2. Training MaskFreeVIS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 228, + 287, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 313 + ], + "type": "text", + "content": "In this section, we describe how to train state-of-the-art VIS methods using our TK-Loss, without any mask annotations. Our MaskFreeVIS approach is jointly supervised with spatial-temporal surrogate losses, and is easily integrated with existing transformer-based methods. We also detail mask-free image-based pre-training for MaskFree-VIS to fully eliminate mask usage during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 319, + 239, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 239, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 239, + 331 + ], + "type": "text", + "content": "3.2.1 Joint Spatio-temporal Regularization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 338, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 385 + ], + "type": "text", + "content": "To train MaskFreeVIS, in addition to our proposed Temporal KNN-patch Loss for temporal mask consistency, we leverage existing spatial weak segmentation losses to jointly enforce intra-frame consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": "Spatial Consistency To explore spatial weak supervision signals from image bounding boxes and pixel color, we utilize the representative Box Projection Loss " + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": " and Pairwise Loss " + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pair}}" + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": " in [50], to replace the supervised mask learning loss. The Projection Loss " + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": " enforces the projection " + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": " of the object mask onto the " + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\vec{x}" + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": "-axis and " + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\vec{y}" + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": "-axis of image to be consistent with its ground-truth box mask. For the temporal tube with " + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 385, + 287, + 493 + ], + "type": "text", + "content": " frames, we concurrently optimize all predicted frame masks of the tube as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 498, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 498, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 81, + 498, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p r o j}} = \\sum_ {t = 1} ^ {T} \\sum_ {d \\in \\{\\vec {x}, \\vec {y} \\}} D \\left(P _ {d} ^ {\\prime} \\left(M _ {p} ^ {t}\\right), P _ {d} ^ {\\prime} \\left(M _ {b} ^ {t}\\right)\\right), \\tag {5}", + "image_path": "cc32e4de01ae5943a5878c684b6d19d751ffd014f35d979b65719b1b0bea88f2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "text", + "content": " denotes dice loss, " + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "text", + "content": " is the projection function along " + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\vec{x} / \\vec{y}" + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "text", + "content": "-axis direction, " + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "inline_equation", + "content": "M_p^t" + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "inline_equation", + "content": "M_b^t" + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "text", + "content": " denote predicted instance mask and its GT box mask at frame " + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 539, + 287, + 586 + ], + "type": "text", + "content": " respectively. The object instance index is omitted here for clarity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "text", + "content": "The Pairwise Loss " + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pair}}" + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "text", + "content": ", on the other hand, constrains spatially neighboring pixels of single frame. For pixel of locations " + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "inline_equation", + "content": "p_i'" + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "inline_equation", + "content": "p_j'" + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "text", + "content": " of with color similarity " + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\geqslant \\sigma_{\\mathrm{pixel}}" + }, + { + "bbox": [ + 46, + 586, + 287, + 647 + ], + "type": "text", + "content": ", we enforce their predicted mask labels to be consistent, following Eq. (3) as," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 652, + 287, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 652, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 82, + 652, + 287, + 687 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p a i r}} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {p _ {i} ^ {\\prime} \\in H \\times W} L _ {\\text {c o n s}} \\left(M _ {p _ {i} ^ {\\prime}} ^ {t}, M _ {p _ {j} ^ {\\prime}} ^ {t}\\right). \\tag {6}", + "image_path": "549bb723aedc820f79b4f962695839f2812713bcaf5d872064354edc9340d2fb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 692, + 282, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 692, + 282, + 705 + ], + "spans": [ + { + "bbox": [ + 47, + 692, + 282, + 705 + ], + "type": "text", + "content": "The spatial losses are combined with a weight factor " + }, + { + "bbox": [ + 47, + 692, + 282, + 705 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{pair}}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 111, + 710, + 287, + 724 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 710, + 287, + 724 + ], + "spans": [ + { + "bbox": [ + 111, + 710, + 287, + 724 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s p a t i a l}} = \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {p a i r}} \\mathcal {L} _ {\\text {p a i r}}. \\tag {7}", + "image_path": "1b12649c9232eac85c4ed95dba14440b9cf8977e8feae8e636c880e3d49b7e32.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 65, + 545, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 65, + 545, + 113 + ], + "spans": [ + { + "bbox": [ + 305, + 65, + 545, + 113 + ], + "type": "text", + "content": "Temporal Consistency We adopt the Temporal KNN-patch Loss in Sec. 3.1.2 as " + }, + { + "bbox": [ + 305, + 65, + 545, + 113 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{temp}}" + }, + { + "bbox": [ + 305, + 65, + 545, + 113 + ], + "type": "text", + "content": " to leverage temporal mask consistency. The overall spatio-temporal objective " + }, + { + "bbox": [ + 305, + 65, + 545, + 113 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 305, + 65, + 545, + 113 + ], + "type": "text", + "content": " for optimizing video segmentation is summarized as," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 123, + 545, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 123, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 367, + 123, + 545, + 137 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e g}} = \\mathcal {L} _ {\\text {s p a t i a l}} + \\lambda_ {\\text {t e m p}} \\mathcal {L} _ {\\text {t e m p}}. \\tag {8}", + "image_path": "0c4d5c3e4d627c21f7a4756baaa746e8b9bec47d17996730635c18d676ca6543.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 150, + 533, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 150, + 533, + 162 + ], + "spans": [ + { + "bbox": [ + 305, + 150, + 533, + 162 + ], + "type": "text", + "content": "3.2.2 Integration with Transformer-based Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 169, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 289 + ], + "type": "text", + "content": "Existing works [13, 48] on box-supervised segmentation losses are coupled with either one-stage or two-stage detectors, such as Faster R-CNN [41] and CondInst [48], and only address the single image case. However, state-of-the-art VIS methods [6, 57] are based on transformers. These works perform object detection via set prediction, where predicted instance masks need to be matched with mask annotations when evaluating the loss. To integrate mask-free VIS training with transformers, one key modification is in this instance-sequence matching step." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 289, + 546, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 612 + ], + "type": "text", + "content": "Since only ground-truth bounding boxes are available for box sequence matching, as an initial attempt, we first produce bounding box predictions from the estimated instance masks. Then, we employ the sequential box matching cost function used in VIS methods [55, 57]. To compute matching cost for whole sequence, " + }, + { + "bbox": [ + 304, + 289, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 304, + 289, + 546, + 612 + ], + "type": "text", + "content": " loss and generalized IoU loss for each individual bounding box is averaged across the frames. However, we observe the matching results of frame-wise averaging can easily be affected by a single outlier frame, especially under weak segmentation setup, leading to instability during training and performance decrease. Spatio-temporal Box Mask Matching Instead of using the aforementioned frame-wise matching, we empirically find spatio-temporal box-to-mask matching to produce substantial improvement under the weak segmentation setting. We first convert each predicted instance mask to a bounding box mask, and convert the ground-truth box to box mask. We then randomly sample a equal number of points from the ground-truth box mask sequence and predicted box mask sequence, respectively. Different from Mask2Former [6], we only adopt the dice IoU loss to compute sequence matching cost. We find that cross-entropy accumulates errors per pixel, leading to imbalanced values between large and small objects. In contrast, the IoU loss in normalized per object, leading to a balanced metric. We study different instance sequence matching strategies under the mask-free VIS setting in the ablation experiments." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 620, + 511, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 620, + 511, + 632 + ], + "spans": [ + { + "bbox": [ + 305, + 620, + 511, + 632 + ], + "type": "text", + "content": "3.2.3 Image-based MaskFreeVIS Pre-training" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 639, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 723 + ], + "type": "text", + "content": "Most VIS models [6, 57, 61] are initialized from a model pretrained on the COCO instance segmentation dataset. To completely eliminate mask supervision, we pretrain our MaskFreeVIS on COCO using only box supervision. We adopt the spatial consistency loss described in Sec. 3.2.1 on single frame to replace the original GT mask losses in Mask2Former [6], while following the same image-based" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "22861" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 65, + 286, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 65, + 286, + 124 + ], + "spans": [ + { + "bbox": [ + 48, + 65, + 286, + 124 + ], + "type": "text", + "content": "training setup on COCO. Thus, we provide two training settings in our experiments, one eliminates both image and video mask during training, while the other adopts weights pretrained with COCO mask annotations. In both cases, no video mask annotations are used." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 133, + 127, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 133, + 127, + 147 + ], + "spans": [ + { + "bbox": [ + 48, + 133, + 127, + 147 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 108, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 108, + 165 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 108, + 165 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 171, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 171, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 286, + 255 + ], + "type": "text", + "content": "YTVIS 2019/2021 We perform experiments on the largescale YouTube-VIS [61] 2019 and 2021. YTVIS 2019 includes 2,883 videos of 131k annotated object instances belonging to 40 categories. To handle more complex cases, YTVIS 2021 updates YTVIS 2019 with additional 794 videos for training and 129 videos for validation, including more tracklets with confusing motion trajectories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 255, + 286, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 286, + 302 + ], + "type": "text", + "content": "OVIS We also train and evaluate on OVIS [39], a VIS benchmark on occlusion learning. OVIS consists of instance masks covering 25 categories with 607, 140 and 154 videos for train, valid and test respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 303, + 286, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 286, + 363 + ], + "type": "text", + "content": "BDD100K MOTS We further report results of Mask-FreeVIS on the large-scale self-driving benchmark BDD100K [64] MOTS. The dataset annotates 154 videos (30,817 images) for training, 32 videos (6,475 images) for validation, and 37 videos (7,484 images) for testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 369, + 179, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 369, + 179, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 369, + 179, + 381 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "text", + "content": "Our proposed approach only requires replacing the original video mask loss in state-of-the-art VIS methods. In particular, we adopt Mask2Former [6] and SeqFormer [57] due to their excellent VIS results. Unless specified, we kept all other training schedules and settings the same as in the original methods. For the Temporal KNN-patch Loss, we set the patch size to " + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "text", + "content": ", search radius to 5 and " + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "text", + "content": ". We adopt the " + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "text", + "content": " distance as the patch matching metric and set the matching threshold to 0.05. On YTVIS 2019/2021, the Mask2Former based models are trained with AdamW [33] with learning rate " + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 48, + 388, + 286, + 638 + ], + "type": "text", + "content": " and weight decay 0.05. The learning rate decays by 10 times at with a factor of 2/3. We set batch size to 16, and train 6k/8k iterations on YTVIS 2019/2021. For experiments on OVIS and BDD100K, we adopted the COCO mask pretrained models by VITA and Unicorn. For the sampled temporal tube at training, we use 5 frames with shuffling instead of 2 frames for better temporal regularization. The compared baselines are adjusted accordingly. During testing, since there is no architecture modification, the inference of MaskFreeVIS is the same to the baselines. More details are in the Supp. file." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 643, + 171, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 171, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 171, + 657 + ], + "type": "text", + "content": "4.3. Ablation Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 662, + 286, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 286, + 722 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 286, + 722 + ], + "type": "text", + "content": "We perform detailed ablation studies for MaskFreeVIS using ResNet-50 as backbone on the YTVIS 2019 val set. We adopt the COCO box-pretrained model as initialization to eliminate all mask annotations from the training. Taking Mask2Former [6] as the base VIS method, we analyze" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 99, + 544, + 190 + ], + "blocks": [ + { + "bbox": [ + 307, + 64, + 545, + 96 + ], + "lines": [ + { + "bbox": [ + 307, + 64, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 307, + 64, + 545, + 96 + ], + "type": "text", + "content": "Table 2. Different temporal matching schemes under the mask-free training setting on YTVIS2019 val. 'Param' indicates whether the matching scheme brings extra model parameters." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 99, + 544, + 190 + ], + "lines": [ + { + "bbox": [ + 308, + 99, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 308, + 99, + 544, + 190 + ], + "type": "table", + "html": "
Temporal Matching SchemeParamAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Baseline (No Matching)✘38.665.938.838.447.7
Flow-based Matching✓40.266.341.940.549.1
Temporal Deformable Matching✓39.665.940.139.948.6
Learnable Pixel Matching✓39.565.740.239.748.4
Learnable Patch Matching✓40.666.542.640.349.2
3D Pairwise Matching✘39.465.041.740.248.0
Temporal KNN-Matching (Ours)✘42.566.845.741.251.2
", + "image_path": "7cd22fd597b5a1002c8105b63477b5b58d92f04261a6c206e3163b43489547c9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 194, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 230 + ], + "type": "text", + "content": "the impact of individual proposed components. Moreover, we study several alternative solutions for temporal matching and influence of different hyper-parameters to TK-Loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 231, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 422 + ], + "type": "text", + "content": "Comparison on Temporal Matching Schemes Table 2 compares our Temporal KNN-Matching to four alternative frame-wise matching approaches for enforcing temporal mask consistency. Flow-based Matching employs the pretrained optical flow model RAFT [45] to build pixel correspondence [32]. Temporal Deformable Matching adopts the temporal deformable kernels [47] to predict the pixel offsets between the target and alignment frame. Instead of using raw patches, Learnable Pixel/Patch Matching employs jointly learned deep pixel/patch embeddings via three learnable FC layers, which are then used to compute the affinities in a soft attention-like manner. 3D Pairwise Matching directly extends " + }, + { + "bbox": [ + 307, + 231, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pair}}" + }, + { + "bbox": [ + 307, + 231, + 545, + 422 + ], + "type": "text", + "content": " designed for spatial images to the temporal dimension, where pairwise affinity loss is computed among pixels not only in within the frame but also across multiple frames." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "type": "text", + "content": "In Table 2, compared to Flow-based Matching with one-to-one pixel correspondence, our paramter-free Temporal KNN-Matching with one-to- " + }, + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "type": "text", + "content": " improves by about 2.3 AP. The prediction of flow-based models are not reliable in case of occlusions and homogeneous regions, and are also influenced by the gap between the training dataset and realworld video data. For the above mentioned deformable and learnable matching schemes, since there are only weak bounding box labels during training, the temporal matching relation is learnt implicitly. We empirically observe the instability during training with limited improvement under the mask-free training setting. For direct generalization of " + }, + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pair}}" + }, + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "type": "text", + "content": ", it only leads to 0.8 mask AP performance improvement. Despite the simplicity and efficiency of the TK-loss, it significantly improves VIS performance by 3.9 mask AP. Effect of Temporal KNN-patch Loss MaskFreeVIS is trained with joint spatio-temporal losses. In Table 3, to evaluate the effectiveness of each loss component, we compare the performance of MaskFreeVIS solely under the spatial pairwise loss [50] or our proposed TK-Loss. Compared to the 2.0 mask AP improvement by the spatial pairwise loss, the TK-Loss substantially promotes the mask AP from 36.6 to 41.6, showing the advantage of our flexible one-to- " + }, + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 307, + 423, + 545, + 721 + ], + "type": "text", + "content": " patch correspondence design in leveraging temporal consistency. We show the VIS results in Figure 5 to visualize" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "22862" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 84, + 285, + 137 + ], + "blocks": [ + { + "bbox": [ + 47, + 60, + 287, + 81 + ], + "lines": [ + { + "bbox": [ + 47, + 60, + 287, + 81 + ], + "spans": [ + { + "bbox": [ + 47, + 60, + 287, + 81 + ], + "type": "text", + "content": "Table 3. Effect of the Spatial Pairwise loss and our Temporal KNN-patch Loss on YTVIS2019 val." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 84, + 285, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 84, + 285, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 84, + 285, + 137 + ], + "type": "table", + "html": "
Box ProjPairwiseTK-LossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
✓36.666.536.237.145.0
✓✓\\( 38.6_{\\uparrow 2.0} \\)65.938.838.447.7
✓✓\\( 41.6_{\\uparrow 5.0} \\)68.443.540.150.5
✓✓✓\\( 42.5_{\\uparrow 5.9} \\)66.845.741.251.2
", + "image_path": "102036d78dfcc4ce146408b380d953b223a742098e964317dffe77343adb668c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 162, + 160, + 215 + ], + "blocks": [ + { + "bbox": [ + 47, + 139, + 164, + 160 + ], + "lines": [ + { + "bbox": [ + 47, + 139, + 164, + 160 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 164, + 160 + ], + "type": "text", + "content": "Table 4. Effect of using max K matches on YTVIS2019 val." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 162, + 160, + 215 + ], + "lines": [ + { + "bbox": [ + 50, + 162, + 160, + 215 + ], + "spans": [ + { + "bbox": [ + 50, + 162, + 160, + 215 + ], + "type": "table", + "html": "
KAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
140.865.844.140.3
341.966.945.141.9
542.566.845.741.2
742.367.144.640.6
", + "image_path": "1da78a330a6eb6ad40d0ad6f8fd67a0e85a796eb3428f7b409c9e846a25d2f20.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 245, + 160, + 299 + ], + "blocks": [ + { + "bbox": [ + 47, + 223, + 164, + 243 + ], + "lines": [ + { + "bbox": [ + 47, + 223, + 164, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 164, + 243 + ], + "type": "text", + "content": "Table 6. Impact of search radius R on YTVIS2019 val." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 245, + 160, + 299 + ], + "lines": [ + { + "bbox": [ + 50, + 245, + 160, + 299 + ], + "spans": [ + { + "bbox": [ + 50, + 245, + 160, + 299 + ], + "type": "table", + "html": "
RAPAP50AP75AR1
139.665.740.239.7
341.366.643.040.3
542.566.845.741.2
742.367.144.640.6
", + "image_path": "fa44ef0ed8c9c9c89dd1f364904cbc235755111f2860fcdea09ddef27ad47207.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 171, + 173, + 282, + 217 + ], + "blocks": [ + { + "bbox": [ + 168, + 139, + 285, + 171 + ], + "lines": [ + { + "bbox": [ + 168, + 139, + 285, + 171 + ], + "spans": [ + { + "bbox": [ + 168, + 139, + 285, + 171 + ], + "type": "text", + "content": "Table 5. Patch Matching metrics comparison. NCC is Norm. Cross-Correlation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 173, + 282, + 217 + ], + "lines": [ + { + "bbox": [ + 171, + 173, + 282, + 217 + ], + "spans": [ + { + "bbox": [ + 171, + 173, + 282, + 217 + ], + "type": "table", + "html": "
MetricAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
NCC41.766.743.441.4
\\( L_1 \\)41.266.243.640.3
\\( L_2 \\)42.566.845.741.2
", + "image_path": "6c1567b240016a425c683e3945588765fd013dfbee9051acbb3160dc11cc7edb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 171, + 245, + 282, + 298 + ], + "blocks": [ + { + "bbox": [ + 168, + 222, + 285, + 243 + ], + "lines": [ + { + "bbox": [ + 168, + 222, + 285, + 243 + ], + "spans": [ + { + "bbox": [ + 168, + 222, + 285, + 243 + ], + "type": "text", + "content": "Table 7. Influence of patch size N on YTVIS2019 val." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 245, + 282, + 298 + ], + "lines": [ + { + "bbox": [ + 171, + 245, + 282, + 298 + ], + "spans": [ + { + "bbox": [ + 171, + 245, + 282, + 298 + ], + "type": "table", + "html": "
NAPAP50AP75AR1
140.165.242.240.0
342.566.845.741.2
542.168.744.442.1
741.566.344.841.3
", + "image_path": "73a06ea5b38111c0e07eec04eabdca8514c9bdf45f67c14a904f62170e2a3b8a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 306, + 215, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 215, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 215, + 317 + ], + "type": "text", + "content": "the effectiveness of each loss component." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "text", + "content": "Analysis of Temporal KNN-patch Loss In Table 4, we study the influence of " + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "text", + "content": ", the maximum number of matches selected in Temporal KNN-patch Loss. The best result is obtained for " + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "text", + "content": " only allows for the one-to-one and no-match cases. The improvement from 40.8 mask AP to 42.5 mask AP reveals the benefit brought by the flexible one-to-many correspondence design. We also analyze the matching metric in Table 5, search radius in Table 6, and patch size in Table 7 (see Sec. 3.1.2 for details). When patch size " + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 318, + 287, + 558 + ], + "type": "text", + "content": " is increased from 1 to 3 in Table 7, the performance of MaskFreeVIS is improved by 2.4 AP, validating the importance of patch structure in robust matching. Effect of the Cyclic Tube Connection We compare three frame-wise tube connection schemes (Figure 4) for the TK-Loss in Table 8. While dense connection brings forth the best performance, it doubles the training memory with minor improvement compared to Cyclic connection. Comparing to Sequential connection, our Cyclic connection benefits from long-range consistency, improving the performance of 0.6 mask AP with an affordable memory cost growth." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 559, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 592 + ], + "type": "text", + "content": "Table 8. Comparison of the tube connection schemes, illustrated in Fig. 4. The tube consist of 5 frames. 'Mem' denotes the memory consumption per sampled video by the TK-Loss during training." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 50, + 594, + 285, + 643 + ], + "blocks": [ + { + "bbox": [ + 50, + 594, + 285, + 643 + ], + "lines": [ + { + "bbox": [ + 50, + 594, + 285, + 643 + ], + "spans": [ + { + "bbox": [ + 50, + 594, + 285, + 643 + ], + "type": "table", + "html": "
Tube ConnectConnect Num.Mem (MB)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)
Dense10152642.768.044.341.5
Sequential463141.966.544.741.2
Cyclic (Ours)577342.566.845.741.2
", + "image_path": "b07ac3667d1e0b2bf64c787e778f723e15f04218b7f2fcdd8c14ee055e691d76.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 651, + 287, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 723 + ], + "type": "text", + "content": "Comparison on Sequence Matching Functions Besides TK-Loss, we analyze the influence of sequence matching cost functions for transformer-based VIS methods under the mask-free setting. In Table 9, we identify the substantial advantage of Spatio-temporal box-mask matching over frameworkwise cost averaging [55, 57]. As discussed in Sec. 3.2.2, we" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 60, + 387, + 127 + ], + "blocks": [ + { + "bbox": [ + 309, + 60, + 387, + 127 + ], + "lines": [ + { + "bbox": [ + 309, + 60, + 387, + 127 + ], + "spans": [ + { + "bbox": [ + 309, + 60, + 387, + 127 + ], + "type": "image", + "image_path": "43bdb206943dc1d9df06ea1b12062d697ca0ec1a145beb2c8409267a3f09bfbd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 309, + 127, + 387, + 186 + ], + "blocks": [ + { + "bbox": [ + 309, + 127, + 387, + 186 + ], + "lines": [ + { + "bbox": [ + 309, + 127, + 387, + 186 + ], + "spans": [ + { + "bbox": [ + 309, + 127, + 387, + 186 + ], + "type": "image", + "image_path": "c4fbc8d99e6d880551ba0ed3394635374723593ad0f2d7dd968f15ad2959c850.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 186, + 388, + 194 + ], + "lines": [ + { + "bbox": [ + 309, + 186, + 388, + 194 + ], + "spans": [ + { + "bbox": [ + 309, + 186, + 388, + 194 + ], + "type": "text", + "content": "Mask2Fomer + BoxInst (Baseline)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 387, + 60, + 465, + 127 + ], + "blocks": [ + { + "bbox": [ + 387, + 60, + 465, + 127 + ], + "lines": [ + { + "bbox": [ + 387, + 60, + 465, + 127 + ], + "spans": [ + { + "bbox": [ + 387, + 60, + 465, + 127 + ], + "type": "image", + "image_path": "7655795ebbce2315e34dabc4215aba1261f7029ae455aefd099e24f42776c1f6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 387, + 127, + 465, + 186 + ], + "blocks": [ + { + "bbox": [ + 387, + 127, + 465, + 186 + ], + "lines": [ + { + "bbox": [ + 387, + 127, + 465, + 186 + ], + "spans": [ + { + "bbox": [ + 387, + 127, + 465, + 186 + ], + "type": "image", + "image_path": "225a7b23ffb2341280d298fd5411a9d9e393c666f883ac2ba332283facf3da24.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 187, + 465, + 194 + ], + "lines": [ + { + "bbox": [ + 391, + 187, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 391, + 187, + 465, + 194 + ], + "type": "text", + "content": "Mask2Fomer + TK-Loss (Ours)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 466, + 60, + 544, + 127 + ], + "blocks": [ + { + "bbox": [ + 466, + 60, + 544, + 127 + ], + "lines": [ + { + "bbox": [ + 466, + 60, + 544, + 127 + ], + "spans": [ + { + "bbox": [ + 466, + 60, + 544, + 127 + ], + "type": "image", + "image_path": "b0076021faf8322db5ee9e859464a1c2998f65a9528ecd1bdf7c9062a7600c84.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 466, + 127, + 544, + 186 + ], + "blocks": [ + { + "bbox": [ + 466, + 127, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 466, + 127, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 466, + 127, + 544, + 186 + ], + "type": "image", + "image_path": "fa94a76cb5bf4e5b7e56ac8a3fcf338204a41db130dafd40a61e6958f4089a65.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 187, + 535, + 194 + ], + "lines": [ + { + "bbox": [ + 481, + 187, + 535, + 194 + ], + "spans": [ + { + "bbox": [ + 481, + 187, + 535, + 194 + ], + "type": "text", + "content": "Mask2Former (Oracle)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 308, + 236, + 418, + 312 + ], + "blocks": [ + { + "bbox": [ + 305, + 198, + 545, + 230 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 230 + ], + "type": "text", + "content": "Figure 5. Qualitative results comparison between using Spatial Pairwise loss [50], our TK-Loss, and Mask2Former (oracle) trained with GT video and image masks." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 236, + 418, + 312 + ], + "lines": [ + { + "bbox": [ + 308, + 236, + 418, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 236, + 418, + 312 + ], + "type": "image", + "image_path": "78f10297b12cc2e0186191abe0d636b481f6a79e613d2cad1a89296e4b5e03f2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 316, + 545, + 359 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 359 + ], + "type": "text", + "content": "Figure 6. Results on YTVIS 2019 val with various percentages of the YTVIS training data. Baseline denotes Mask2Former [6] trained with GT video boxes using BoxInst [50], while Oracle denotes Mask2Former trained with GT video masks." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 423, + 236, + 542, + 312 + ], + "blocks": [ + { + "bbox": [ + 423, + 236, + 542, + 312 + ], + "lines": [ + { + "bbox": [ + 423, + 236, + 542, + 312 + ], + "spans": [ + { + "bbox": [ + 423, + 236, + 542, + 312 + ], + "type": "image", + "image_path": "c8f096fdf99645268ff2e1f6c2356be87722b38c12a81eea94d9428d3442e35d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 313, + 396, + 537, + 440 + ], + "blocks": [ + { + "bbox": [ + 305, + 361, + 545, + 395 + ], + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 395 + ], + "type": "text", + "content": "Table 9. Comparison of Set Matching Cost Functions on YTVIS2019 val. ST-BoxMask denotes our Spatio-temporal Box Mask matching. w/o CE denotes removing cross-entropy cost." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 396, + 537, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 537, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 537, + 440 + ], + "type": "table", + "html": "
Matching Cost FunctionAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Frame-wise Averaging [55, 57]37.664.239.537.545.7
ST-BoxMask40.867.842.240.049.2
ST-BoxMask w/o CE42.566.845.741.251.2
", + "image_path": "8ad7d8d2707745c26ce8dbdbb268b9de35d00b7959b49fa3c8bc75a93b20c681.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 443, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 443, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 443, + 545, + 467 + ], + "type": "text", + "content": "achieve further gain by removing the object size imbalanced cross-entropy cost computation." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 468, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 586 + ], + "type": "text", + "content": "Training on Various Amounts of Data To further study label-efficient VIS, we validate the effect of MaskFreeVIS under various percentages of the YTVIS 2019 training data. We uniformly sample frames and their labels for each video, and set the minimum sampled number of frames to 1. Figure 6 presents the experimental results, which shows the consistent large improvement (over 3.0 AP) brought by our TK-Loss under various amounts of data. In particular, we note that our approach with " + }, + { + "bbox": [ + 304, + 468, + 545, + 586 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 468, + 545, + 586 + ], + "type": "text", + "content": " data even outperforms the fully-supervised model with " + }, + { + "bbox": [ + 304, + 468, + 545, + 586 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 468, + 545, + 586 + ], + "type": "text", + "content": " training data." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 595, + 528, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 528, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 528, + 608 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 614, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 545, + 674 + ], + "type": "text", + "content": "We compare MaskFreeVIS with the state-of-the-art fully/weakly supervised methods on benchmarks YTVIS 2019/2021, OVIS and BDD100K MOTS. We integrate MaskFreeVIS on four representative methods [6,12,57,60], attaining consistent large gains over the strong baselines." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 674, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 545, + 723 + ], + "type": "text", + "content": "YTVIS 2019/2021 Table 10 compares the performance on YTVIS 2019. Using R50/R101 as backbone and with the same training setting, MaskFreeVIS achieves 42.5/45.8 AP, improving 3.9/5.0 AP over the strong baseline adopt" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "22863" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 97, + 288, + 477 + ], + "blocks": [ + { + "bbox": [ + 47, + 52, + 287, + 96 + ], + "lines": [ + { + "bbox": [ + 47, + 52, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 52, + 287, + 96 + ], + "type": "text", + "content": "Table 10. Comparison on YTVIS 2019. I: using COCO mask pretrained model as initialization. V: using YTVIS video masks during training. " + }, + { + "bbox": [ + 47, + 52, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\ast" + }, + { + "bbox": [ + 47, + 52, + 287, + 96 + ], + "type": "text", + "content": ": using pseudo mask from COCO images for joint training [57]. M2F: Mask2Former [6], SeqF: SeqFormer [57]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 97, + 288, + 477 + ], + "lines": [ + { + "bbox": [ + 47, + 97, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 97, + 288, + 477 + ], + "type": "table", + "html": "
MethodMask ann.Back-boneAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_{1} \\)\\( AR_{10} \\)
Fully-supervised:
PCAN [18]I+VR5036.154.939.436.341.6
EfficientVIS [59]I+VR5037.959.743.040.346.6
InsPro [10]I+VR5040.262.943.137.644.5
IFC [15]I+VR5042.865.846.843.851.2
VMT* [17]I+VR5047.9-52.045.8-
SeqF* [57]I+VR5047.469.851.845.554.8
M2FI+VR5046.468.050.0--
M2F*I+VR5047.869.252.746.256.6
Prev. Weakly-supervised:
FlowIRN [32]-R5010.527.26.212.313.6
SOLO-Track [9]IR5030.650.733.531.637.1
Mask-free:
M2F + Flow Consist [45]-R5040.266.341.940.549.1
M2F + BoxInst [50]-R5038.664.238.538.046.8
M2F + MaskFreeVIS-R5042.5↑3.966.845.741.251.2
M2F + MaskFreeVISIR5043.8↑5.270.746.941.552.3
M2F + MaskFreeVIS*IR5046.6↑8.072.549.744.955.7
Fully-supervised:
M2FVR10145.672.648.944.354.5
M2FI+VR10149.272.854.2--
M2F*I+VR10149.873.655.448.058.0
SeqF*I+VR10149.071.155.746.856.9
Mask-free:
M2F + BoxInst [50]-R10140.867.842.240.049.2
M2F + MaskFreeVIS-R10145.8↑5.070.848.645.355.2
M2F + MaskFreeVISIR10147.3↑6.575.449.944.655.2
M2F + MaskFreeVIS*IR10148.9↑8.174.954.744.957.0
SeqF + MaskFreeVIS*IR10148.674.052.245.957.2
Fully-supervised:
M2FI+VSwinL60.484.467.0--
Mask-free:
M2F + BoxInst [50]-SwinL49.873.255.548.258.1
M2F + MaskFreeVIS-SwinL54.3↑4.582.661.150.261.3
M2F + MaskFreeVIS*ISwinL55.3↑5.582.560.850.762.2
", + "image_path": "3b15d68d7330d45d7c9c5fd4c13058bfe96009d26b453a16c3280fffbb5f661c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "type": "text", + "content": "ing BoxInst [50] losses. MaskFreeVIS, without any mask labels, even significantly outperforms some recent fully-supervised methods such as EfficientVIS [59] and In-sPro [10]. On R50/R101/Swin-L, our MaskFreeVIS consistently attains over " + }, + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "type": "text", + "content": " of its fully-supervised counterpart trained with both GT image and video masks. We also observe similar larger performance growth over the baseline on YTVIS 2021 in Table 11. The excellent performance substantially narrows the performance gap between fully-supervised and weakly-supervised VIS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 598, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 287, + 647 + ], + "type": "text", + "content": "OVIS We also conduct experiments on OVIS in Table 12 using R50 as backbone. We integrate MaskFreeVIS with VITA [12], promoting the baseline performance from 12.1 to 15.7 under the mask-free training setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "content": "BDD100K MOTS Table 13 further validates our approach on BDD100K MOTS. Integrated with Unicorn [60], Mask-FreeVIS achieves " + }, + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "inline_equation", + "content": "23.8\\mathrm{mMOTSA}" + }, + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "content": " by improving over 4.9 points over the strong baseline and thus surpassing the fully" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 74, + 547, + 273 + ], + "blocks": [ + { + "bbox": [ + 305, + 52, + 545, + 73 + ], + "lines": [ + { + "bbox": [ + 305, + 52, + 545, + 73 + ], + "spans": [ + { + "bbox": [ + 305, + 52, + 545, + 73 + ], + "type": "text", + "content": "Table 11. Comparison on YTVIS 2021. Refer to Table 10 for the symbol abbreviations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 74, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 307, + 74, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 74, + 547, + 273 + ], + "type": "table", + "html": "
MethodMask ann.Back-boneAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Fully-supervised:
MaskTrack [61]I+VR5028.648.929.626.533.8
IFC [15]I+VR5036.657.939.3--
SeqF* [57]I+VR5040.562.443.736.148.1
M2FI+VR5040.660.941.8--
Mask-free:
M2F + BoxInst [50]-R5032.152.834.431.038.1
M2F + MaskFreeVIS-R5036.2↑4.160.839.234.645.6
M2F + MaskFreeVISIR5037.2↑5.161.940.335.346.1
M2F + MaskFreeVIS*IR5040.9↑8.865.843.337.150.5
Fully-supervised:
M2FI+VR10142.465.945.8--
Mask-free:
M2F + BoxInst [50]-R10133.355.232.532.141.9
M2F + MaskFreeVIS-R10137.3↑4.061.639.434.145.6
M2F + MaskFreeVISIR10138.2↑4.962.440.034.946.2
M2F + MaskFreeVIS*IR10141.6↑8.366.244.836.349.2
", + "image_path": "a459e4072641e57c49f7968446fa68fc13b74252cf870e151650e6090f2f2b92.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 291, + 542, + 396 + ], + "blocks": [ + { + "bbox": [ + 312, + 277, + 538, + 289 + ], + "lines": [ + { + "bbox": [ + 312, + 277, + 538, + 289 + ], + "spans": [ + { + "bbox": [ + 312, + 277, + 538, + 289 + ], + "type": "text", + "content": "Table 12. State-of-the-art comparison on the OVIS using R50." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 291, + 542, + 396 + ], + "lines": [ + { + "bbox": [ + 307, + 291, + 542, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 542, + 396 + ], + "type": "table", + "html": "
MethodAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Fully-supervised:
CrossVIS [62]14.932.712.110.319.8
Mask2Former [6]17.337.315.110.523.5
VMT [17]16.936.413.710.422.7
VITA [12]19.641.217.411.726.0
Video Mask-free:
VITA [12] + BoxInst [50]12.128.310.28.817.9
VITA [12] + MaskFreeVIS\\( 15.7_{\\uparrow 3,6} \\)35.113.110.120.4
", + "image_path": "bd13df803b2e22a6609463978b1445a4a69c5ed001247150106484e3d3f5f160.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 423, + 547, + 516 + ], + "blocks": [ + { + "bbox": [ + 306, + 399, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 422 + ], + "type": "text", + "content": "Table 13. State-of-the-art comparison on the BDD100K segmentation tracking validation set." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 423, + 547, + 516 + ], + "lines": [ + { + "bbox": [ + 307, + 423, + 547, + 516 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 547, + 516 + ], + "type": "table", + "html": "
MethodmMOTSA↑mMOTSP↑mIDF↑ID sw.↓mAP↑
Fully-supervised:
STEm-Seg [1]12.258.225.4873221.8
QDTrack-mots-fix [36]23.566.344.597325.5
PCAN [18]27.466.745.187626.6
Unicorn [60]29.667.744.2173132.1
Video Mask-free:
Unicorn [60] + BoxInst [50]18.958.736.3329822.1
Unicorn [60] + MaskFreeVIS23.8↑4.966.744.9208624.8
", + "image_path": "b98f1efffece4f0ef723115a55989e1566855e8cf44f302afaf3381a566eee29.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 522, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 522, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 545, + 559 + ], + "type": "text", + "content": "supervised QDTrack-mots [36]. The consistent large gains on four benchmarks and four base VIS methods validate the generalizability of our MaskFreeVIS." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 570, + 379, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 379, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 379, + 582 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 590, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 723 + ], + "type": "text", + "content": "MaskFreeVIS is the first competitive VIS method that does not need any mask annotations during training. The strong results lead to a remarkable conclusion: mask labels are not a necessity for high-performing VIS. Our key component is the unsupervised Temporal KNN-patch Loss, which replaces the conventional video masks losses by leveraging temporal mask consistency constraints. Our approach greatly reduces the long-standing gap between fully-supervised and weakly-supervised VIS on four large-scale benchmarks. MaskFreeVIS thus opens up many opportunities for researchers and practitioners for label-efficient VIS." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 701, + 287, + 723 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 723 + ], + "type": "text", + "content": "1This work is supported in part by Research Grant Council of the Hong Kong SAR under grant no. 16201420 and Huawei Technologies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "22864" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 64, + 107, + 77 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 64, + 107, + 77 + ], + "spans": [ + { + "bbox": [ + 48, + 64, + 107, + 77 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 83, + 287, + 722 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 83, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 83, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 83, + 287, + 127 + ], + "type": "text", + "content": "[1] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, 2020. 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 287, + 161 + ], + "type": "text", + "content": "[2] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 163, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 286, + 184 + ], + "type": "text", + "content": "[3] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 185, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 185, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 185, + 286, + 228 + ], + "type": "text", + "content": "[4] Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 286, + 262 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 263, + 286, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 263, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 53, + 263, + 286, + 307 + ], + "type": "text", + "content": "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 308, + 286, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 286, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 286, + 340 + ], + "type": "text", + "content": "[7] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 342, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 286, + 373 + ], + "type": "text", + "content": "[8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In ICCV, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 376, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 286, + 407 + ], + "type": "text", + "content": "[9] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to track instances without video annotations. In CVPR, 2021. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 409, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 286, + 453 + ], + "type": "text", + "content": "[10] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. In NeurIPS, 2022. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 454, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 286, + 475 + ], + "type": "text", + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 476, + 286, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 286, + 509 + ], + "type": "text", + "content": "[12] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 510, + 286, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 286, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 286, + 553 + ], + "type": "text", + "content": "[13] Cheng-Chun Hsu, Kuang-Jui Hsu, Chung-Chi Tsai, Yen-Yu Lin, and Yung-Yu Chuang. Weakly supervised instance segmentation using the bounding box tightness prior. In NeurIPS, 2019. 1, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 555, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 286, + 586 + ], + "type": "text", + "content": "[14] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "type": "text", + "content": "[15] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. In NeurIPS, 2021. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 622, + 286, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 286, + 654 + ], + "type": "text", + "content": "[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 656, + 286, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 286, + 698 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 286, + 698 + ], + "type": "text", + "content": "[17] Lei Ke, Henghui Ding, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Video mask transfiner for high-quality video instance segmentation. In ECCV, 2022. 1, 2, 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 700, + 286, + 722 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 700, + 286, + 722 + ], + "spans": [ + { + "bbox": [ + 48, + 700, + 286, + 722 + ], + "type": "text", + "content": "[18] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 66, + 545, + 721 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 327, + 66, + 545, + 87 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 66, + 545, + 87 + ], + "spans": [ + { + "bbox": [ + 327, + 66, + 545, + 87 + ], + "type": "text", + "content": "for multiple object tracking and segmentation. In NeurIPS, 2021. 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 89, + 545, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 89, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 308, + 89, + 545, + 120 + ], + "type": "text", + "content": "[19] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 122, + 545, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 122, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 307, + 122, + 545, + 154 + ], + "type": "text", + "content": "[20] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In CVPR, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 156, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 156, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 156, + 545, + 198 + ], + "type": "text", + "content": "[21] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In ICLR, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "text", + "content": "[22] Viveka Kulharia, Siddhartha Chandra, Amit Agrawal, Philip Torr, and Ambrish Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In ECCV, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 245, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 298 + ], + "type": "text", + "content": "[23] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "text", + "content": "[24] Jungbeom Lee, Eunji Kim, Sungmin Lee, Jangho Lee, and Sungroh Yoon. Frame-to-frame aggregation of active regions in web videos for weakly supervised semantic segmentation. In ICCV, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 344, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 376 + ], + "type": "text", + "content": "[25] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In CVPR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 378, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 378, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 545, + 409 + ], + "type": "text", + "content": "[26] Minghan Li, Shuai Li, Lida Li, and Lei Zhang. Spatial feature calibration and temporal fusion for effective one-stage video instance segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "text", + "content": "[27] Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, XianSheng Hua, and Lei Zhang. Box-supervised instance segmentation with level set evolution. In ECCV, 2022. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "text", + "content": "[28] Chung-Ching Lin, Ying Hung, Rogerio Feris, and Linglin He. Video instance segmentation tracking with a modified vae architecture. In CVPR, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[29] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In CVPR, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 512, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 554 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 556, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 587 + ], + "type": "text", + "content": "[31] Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "type": "text", + "content": "[32] Qing Liu, Vignesh Ramanathan, Dhruv Mahajan, Alan Yuille, and Zhenheng Yang. Weakly supervised instance segmentation for videos with temporal mask consistency. In CVPR, 2021. 2, 3, 6, 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 634, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 654 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 656, + 545, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 656, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 308, + 656, + 545, + 688 + ], + "type": "text", + "content": "[34] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 3" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 308, + 689, + 545, + 721 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 689, + 545, + 721 + ], + "spans": [ + { + "bbox": [ + 308, + 689, + 545, + 721 + ], + "type": "text", + "content": "[35] Anton Milan, Laura Leal-Taixe, Konrad Schindler, and Ian Reid. Joint tracking and segmentation of multiple targets. In CVPR, 2015. 3" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "22865" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 66, + 287, + 721 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 66, + 287, + 99 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 66, + 287, + 99 + ], + "spans": [ + { + "bbox": [ + 48, + 66, + 287, + 99 + ], + "type": "text", + "content": "[36] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 100, + 287, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 100, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 49, + 100, + 287, + 142 + ], + "type": "text", + "content": "[37] George Papandreou, Liang-Chieh Chen, Kevin P Murphy, and Alan L Yuille. Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In ICCV, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 144, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 144, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 287, + 186 + ], + "type": "text", + "content": "[38] Jordi Pont-Tuset, Pablo Arbelaez, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping for image segmentation and object proposal generation. IEEE TPAMI, 39(1):128-140, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 189, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 189, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 287, + 231 + ], + "type": "text", + "content": "[39] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 2021. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 233, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 233, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 233, + 287, + 308 + ], + "type": "text", + "content": "[40] Martin Rajchl, Matthew CH Lee, Ozan Oktay, Konstantinos Kamnitsas, Jonathan Passerat-Palmbach, Wenjia Bai, Mellisa Damodaram, Mary A Rutherford, Joseph V Hajnal, Bernhard Kainz, et al. Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging, 36(2):674-683, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 310, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 287, + 342 + ], + "type": "text", + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 344, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 344, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 49, + 344, + 287, + 376 + ], + "type": "text", + "content": "[42] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. Grabcut: interactive foreground extraction using iterated graph cuts. ACM TOG, 23(3):309-314, 2004. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 378, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 378, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 378, + 287, + 430 + ], + "type": "text", + "content": "[43] Fatemeh Sadat Saleh, Mohammad Sadegh Aliakbarian, Mathieu Salzmann, Lars Petersson, and Jose M Alvarez. Bringing background into the foreground: Making all classes equal in weakly-supervised video semantic segmentation. In ICCV, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 433, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 287, + 475 + ], + "type": "text", + "content": "[44] Chunfeng Song, Yan Huang, Wanli Ouyang, and Liang Wang. Box-driven class-wise region masking and filling rate guided loss for weakly supervised semantic segmentation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 478, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 287, + 498 + ], + "type": "text", + "content": "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 501, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 553 + ], + "type": "text", + "content": "[46] Omkar Thawakar, Sanath Narayan, Jiale Cao, Hisham Cholakkal, Rao Muhammad Anwer, Muhammad Haris Khan, Salman Khan, Michael Felsberg, and Fahad Shahbaz Khan. Video instance segmentation via multi-scale spatiotemporal split attention transformer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 556, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 287, + 587 + ], + "type": "text", + "content": "[47] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 590, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 620 + ], + "type": "text", + "content": "[48] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, 2020. 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 623, + 287, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 623, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 49, + 623, + 287, + 653 + ], + "type": "text", + "content": "[49] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 656, + 287, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 656, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 49, + 656, + 287, + 687 + ], + "type": "text", + "content": "[50] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In CVPR, 2021. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 690, + 287, + 721 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 690, + 287, + 721 + ], + "spans": [ + { + "bbox": [ + 49, + 690, + 287, + 721 + ], + "type": "text", + "content": "[51] Pavel Tokmakov, Karteek Alahari, and Cordelia Schmid. Learning semantic segmentation with weakly-annotated videos. In ECCV, 2016. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 66, + 545, + 559 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 308, + 66, + 545, + 87 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 66, + 545, + 87 + ], + "spans": [ + { + "bbox": [ + 308, + 66, + 545, + 87 + ], + "type": "text", + "content": "[52] Yi-Hsuan Tsai, Guangyu Zhong, and Ming-Hsuan Yang. Semantic co-segmentation in videos. In ECCV, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 89, + 545, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 89, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 308, + 89, + 545, + 131 + ], + "type": "text", + "content": "[53] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 134, + 545, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 134, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 134, + 545, + 176 + ], + "type": "text", + "content": "[54] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 178, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 178, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 178, + 545, + 220 + ], + "type": "text", + "content": "[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 224, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 224, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 224, + 545, + 255 + ], + "type": "text", + "content": "[56] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In CVPR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 258, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 289 + ], + "type": "text", + "content": "[57] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. In ECCV, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 292, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 545, + 323 + ], + "type": "text", + "content": "[58] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 325, + 545, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 325, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 308, + 325, + 545, + 367 + ], + "type": "text", + "content": "[59] Jialian Wu, Sudhir Yarram, Hui Liang, Tian Lan, Junsong Yuan, Jayan Eledath, and Gerard Medioni. Efficient video instance segmentation via tracklet query and proposal. In CVPR, 2022. 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 370, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 370, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 370, + 545, + 402 + ], + "type": "text", + "content": "[60] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 3, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 404, + 545, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 424 + ], + "type": "text", + "content": "[61] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 427, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 468 + ], + "type": "text", + "content": "[62] Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In ICCV, 2021. 2, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "text", + "content": "[63] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 517, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 517, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 517, + 545, + 559 + ], + "type": "text", + "content": "[64] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 1, 2, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "22866" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_content_list.json b/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..13847f110c53fe87889c29348d0a755a2814a2f6 --- /dev/null +++ b/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_content_list.json @@ -0,0 +1,1668 @@ +[ + { + "type": "text", + "text": "Mask-guided Matting in the Wild", + "text_level": 1, + "bbox": [ + 313, + 130, + 656, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kwanyong Park $^{1}$ Sanghyun Woo $^{1}$ Seoung Wug Oh $^{2}$ In So Kweon $^{1}$ Joon-Young Lee $^{2}$ $^{1}$ KAIST $^{2}$ Adobe Research", + "bbox": [ + 109, + 179, + 859, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mask-guided matting has shown great practicality compared to traditional trimap-based methods. The mask-guided approach takes an easily-obtainable coarse mask as guidance and produces an accurate alpha matte. To extend the success toward practical usage, we tackle mask-guided matting in the wild, which covers a wide range of categories in their complex context robustly. To this end, we propose a simple yet effective learning framework based on two core insights: 1) learning a generalized matting model that can better understand the given mask guidance and 2) leveraging weak supervision datasets (e.g., instance segmentation dataset) to alleviate the limited diversity and scale of existing matting datasets. Extensive experimental results on multiple benchmarks, consisting of a newly proposed synthetic benchmark (Composition-Wild) and existing natural datasets, demonstrate the superiority of the proposed method. Moreover, we provide appealing results on new practical applications (e.g., panoptic matting and mask-guided video matting), showing the great generality and potential of our model.", + "bbox": [ + 75, + 300, + 473, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 632, + 209, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image matting aims to predict the opacity of objects, which enables precise separation from surrounding backgrounds. Due to the ill-posed nature of the task, many works [7, 13, 21, 27, 30, 48] have improved matting performance by relying on the manual guidance of a trimap. However, pixel-level annotation of foreground/background/unknown is extremely burdensome, restricting its usage in many practical applications such as image/video editing and film production. Recently, many efficient alternatives for user guidance have been proposed, including trimap-free [15, 32], additional background images [22, 34], scribble [43], and the user clicks [45].", + "bbox": [ + 75, + 657, + 468, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among them, the mask-guided approach [50] shows a great trade-off between performance and intensity of user interaction. It utilizes a coarse mask as guidance, which is much easier to obtain either manually or from off-the-shelf", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c5b876cfad515532688621335ac06fec5c6ded6040dfb44f7c80a11e3e8d53d.jpg", + "image_caption": [ + "Figure 1. Qualitative Comparisons of MGMatting [50] and Ours in the wild. The mask guidance is overlaid on images with blue color. Best viewed zoomed in." + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 910, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "segmentation models [2, 10]. With only the coarse spatial prior, the mask-guided matting model [50] shows comparable or even better performance than the trimap-based competitors [13, 17, 21, 27, 48] on synthetic Composition-1k [48] and a real-world human matting dataset [50]. However, despite the encouraging results, we see the previous state-of-the-art model [50] struggles to obtain desirable alpha matte in complex real-world scenes (see Fig. 1).", + "bbox": [ + 496, + 578, + 893, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With this observation, we tackle mask-guided matting in the wild. Specifically, we formulate unique setups and emerging challenges of the new task as follows: (1) We aim to handle objects in their complex context, reflecting the characteristics of natural images. The previous method [50] evaluates their model on iconic-object images [1,29] where only a single object is in the center. As the model can easily find the target object in such images, the model's real instance discrimination ability is, in fact, veiled. On the contrary, in an 'in-the-wild' setting, it is crucial to precisely localize the target object from the given coarse/noisy mask guidance (i.e. mask awareness). (2) Our model targets to deal with diverse categories of objects in natural", + "bbox": [ + 496, + 703, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1992", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images. Unlike most previous methods that improve generalization performance at the expense of category-specific regime (e.g., limiting to humans [15] or animals [19]), we aim to understand distinctive matting patterns of vast categories. (3) Limited data problem makes the new setting more complicated. Due to the labeling complexity, annotating alpha matte for objects in common scenes, e.g., the COCO dataset [24], is infeasible. As a sidestep, previous benchmarks [32, 48] extract the alpha matte and foreground colors from images with simple backgrounds. These are composited on various backgrounds [6, 24], and resulting samples are used to train and evaluate matting models. However, due to the inevitable composition artifacts, the models usually show limited generalization performance. In that sense, how to train and evaluate the in-the-wild matting model remains an open question.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Toward this goal, we propose a simple yet effective learning framework for a generalized mask-guided matting model. First, we investigate fundamental reasons for the poor generalization of the previous mask-guided matting model [50] and find that this is mainly from the training data generation process. Specifically, the previous composition process includes instance merging data augmentation, which merges several foreground objects into a single object. While this augmentation is effective in the trimap-based methods [21, 30, 41], it implicitly makes a negative bias for the mask-guided matting model to ignore the guidance. Thus, the model struggles to localize the target objects in complex natural scenes. We alleviate the bias by proposing an instance-wise learning objective, where the model is supervised to segment one of multiple instances according to the guidance. By doing so, the model learns strong semantic representation regarding complex relations and soft transitions between objects. Despite the simplicity of the proposal, this greatly improves performance in the wild.", + "bbox": [ + 75, + 335, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we explore a practical solution to make the mask-guided model handle various categories of objects robustly. Instead of scaling the matting dataset, we leverage a dataset with weak supervision [14, 46] (i.e., instance segmentation dataset [24]), as the coarse instance masks are easier to obtain over the diverse categories of objects. To effectively hallucinate the fine supervision signal with the weak localization guidance, we come up with a self-training framework [36, 37, 47]. Specifically, a pseudo label is generated based on a weakly-augmented input (both image and instance mask annotation as guidance), which supervises the model prediction on a strongly augmented version of them. During self-training, the model is not only adapting to the in-the-wild scenario in a self-evolving manner but also being robust to noise in both image and guidance.", + "bbox": [ + 75, + 626, + 472, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To verify the in-the-wild performance of mask-guided matting, we formally define an evaluation protocol involving multiple sub-benchmarks: Composition-Wild, AIM-", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "500 [20], COCO [24]. We first design an in-the-wild extension of the popular synthetic Composition-1k benchmark [48], namely Composition-Wild. We simulate complex real-world images by compositing multiple foreground objects. To bring valuable insight on the failure cases of the model, we design sub-metrics for Composition-Wild. In addition, we use the AIM-500 dataset to establish quantitative results on natural images (i.e., with no composition artifacts), although most images are iconic-object images with simple backgrounds. Finally, we provide qualitative outputs of our mask-guided matting model on the COCO dataset [24] which is one of the most representative in-the-wild datasets.", + "bbox": [ + 496, + 90, + 893, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we make the following main contributions. 1) To our best knowledge, it is the first work to explore mask-guided matting in the wild. 2) We develop a simple yet effective learning framework leveraging both composited and weak-guidance images. 3) We design an evaluation setup for the new task. 4) We initiate several interesting extensions: video and panoptic matting.", + "bbox": [ + 496, + 287, + 893, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 407, + 650, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Natural Image Matting. Most image matting methods require a trimap as additional input, which conveys pixel-level annotations of foreground, background, and unknown region. Traditional image matting methods can be categorized into two groups. Sampling-based methods [5, 8, 11, 35, 44] estimate the alpha matte of the unknown region based on sampled colors of foreground and background. Propagation-based methods [4, 12, 16-18, 38] propagate the neighboring known alpha values to unknown regions according to the affinity between pixels. Both methods primarily rely on color or low-level features, showing limitations in complicated scenarios.", + "bbox": [ + 496, + 431, + 893, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these issues, many deep-learning based mating methods have been proposed. DIM [48] is a representative work which proposes a convolutional encoder-decoder matting network as well as a large-scale synthetic dataset (Composition-1k) to train the model. Many follow-up works have made tremendous improvements in diverse aspects, such as sophisticated loss design [7, 13, 28] and architectural advances by introducing attention mechanisms [21, 26] or transformer architecture [30].", + "bbox": [ + 496, + 614, + 893, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There is a large volume of trials to relax the heavy user-supplied constraints. Some methods [32, 52] attempt to get rid of trimap. However, they show inferior performance and cannot be generalized to unseen objects in the real world. Sengupta et al. [34] utilize additional background images along with other lightweight priors (e.g., segmentation mask) to perform the matting task. Wei et al. [45] introduce user click interaction to effectively eliminate the ambiguity of target object. Recently, MGMatting [50] proposes a mask-guided matting framework, where only the", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1993", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "easily obtainable coarse mask is needed as guidance.", + "bbox": [ + 76, + 90, + 426, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we further study mask-guided matting in an in-the-wild setting, inspired by the generality and practicality of the framework. Different from MGMatting, we target diverse categories of objects in their complex background, enabling real-world applications.", + "bbox": [ + 75, + 106, + 467, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-specific Matting is a special type of matting that limits target objects to one or few classes such as humans [3] or animals [19]. BSHM [25] leverages the coarse segmentation mask data to build a generalized human matting network. MODNet [15] presents a lightweight network architecture with decomposed multi-scale network designs. GFM [19] first proposes the task of animal matting and a real-world animal matting dataset. In general, known semantics effectively ease the difficulty of matting, and these methods show great generality without the necessity of additional guidance input. However, the specialized methods and trained model cannot be generalized to the in-the-wild setting since limited semantics do not cover diverse patterns of visual worlds.", + "bbox": [ + 75, + 181, + 467, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Apart from previous efforts, we attempt to extend the short category regime of matting networks. For this purpose, we propose an efficient solution by leveraging the instance segmentation dataset [24], which covers diverse categories of objects with weak localization annotations. Given the weak guidance dataset, we study the challenges in developing an in-the-wild mask-guided matting model.", + "bbox": [ + 75, + 393, + 467, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Mask-guided Matting in the Wild", + "text_level": 1, + "bbox": [ + 76, + 512, + 382, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In natural scenes, objects are rarely in isolation: diverse categories of entities interact with each other. We aim to robustly separate a target object from the surroundings in such scenes. Specifically, we formulate the problem in the form of mask-guided matting, namely mask-guided matting in the wild. Our matting model takes an easily-obtainable coarse mask as guidance, either manually drawn or predicted from the off-the-shelf segmentation models. With in-the-wild data, the model needs to understand semantics so that it can precisely localize the target object from the given noisy masks. In the meantime, the model also needs to have strong low-level feature representation so that it can capture fine details and accurate opacity of the diverse target matte.", + "bbox": [ + 75, + 537, + 467, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To concrete the new challenges toward such an in-the-wild setting, we first study how the previous model [50] performs. We derive predictions with different mask guidance and summarize the results on Fig. 2. As shown in the upper example, in the synthetic Composition-1k [48] example, the previous model perfectly separates the foreground object regardless of the given guidance. Surprisingly, it produces an alpha matte of the foreground object when the mask of the background region is given as guidance. On the contrary, the model lacks generalization on", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2b76c8772c71bce62191100fc3b91dd8c868b9db95412ce38c0475a5b25abb0f.jpg", + "image_caption": [ + "Image and Guidance" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 638, + 231 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eacb1b650023f83cdb64dc8a8df0a693bd1c1a506c95be8140057038e68bd50e.jpg", + "image_caption": [ + "With $1^{\\mathrm{st}}$ Mask Guidance" + ], + "image_footnote": [], + "bbox": [ + 638, + 89, + 771, + 232 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/be7855bebc8978d1a1b719e61b8db458d1e916b662cc9bad1872c4c8c3267b1c.jpg", + "image_caption": [ + "With $2^{\\text{nd}}$ Mask Guidance", + "Figure 2. Predictions of MGMatting [50] with different guidance. (upper) synthetic Composition-1k image [48]. (bottom) Complex real world image [49]. Best viewed zoomed in." + ], + "image_footnote": [], + "bbox": [ + 772, + 89, + 908, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "complex real-world scenes (see bottom example). In particular, the model fails to localize the target object and makes wrong predictions on regions of similar color distributions (e.g., shadows) or another nearby object. From the two representative examples, we make several observations: 1) The previous method yields a model overfitted to the (trained) synthetic matting dataset, which memorizes the limited patterns of foreground objects rather than semantically understanding the mask guidance for localizing and segmenting the objects; 2) Such a model fails to precisely localize and segment a target object in the wild; 3) Previously used synthetic benchmark, Composition-1k [48], neither represents the complex nature of the in-the-wild setting nor systematically blocks the shortcut of the overfitting issue. Thus the Composition-1k may not be a suitable benchmark for mask-guided matting in the wild.", + "bbox": [ + 496, + 330, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we explore how to train a generalized mask-guided matting model given the limited scale and diversity of the matting dataset (Sec. 4), as well as revisit the evaluation benchmark to test the model (Sec. 5).", + "bbox": [ + 496, + 575, + 890, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 660, + 671, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of the proposed training framework is illustrated in Fig. 3. Our framework involves three types of datasets: 1) matting dataset that includes accurate alpha matte and corresponding foreground colors; 2) background dataset, where the foreground objects are composited to create training samples; 3) weak supervision dataset that is introduced to extend the matting model to the in-the-wild setting. The main goal of our framework is to learn accurate matting ability from the synthetic composited images and generalize the knowledge to the in-the-wild domain. To this end, we first investigate how the model could learn generalized knowledge from the composited images (Sec. 4.1), then how the model is further improved with weak supervision data (Sec. 4.2).", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1994", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d8690695c1e705c06e4da75d42df126bfe25625e99137ab44b9a87a179b5a0d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 89, + 834, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b37ab59244c182cbd48597bd63b1400e7de8680145ae377aeaaf83ad18bc3498.jpg", + "image_caption": [ + "Figure 3. The Overview of the Proposed Framework. (a) Learning from composited images with instance-wise learning (Sec. 4.1). (b) Learning from weak-supervision images with self-training (Sec. 4.2)." + ], + "image_footnote": [], + "bbox": [ + 140, + 260, + 813, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Learning from Composed Images", + "text_level": 1, + "bbox": [ + 76, + 439, + 382, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in the Sec. 3, the previous model [50] shows poor generalization performance in the challenging real-world setting. Specifically, it often fails to identify the target object in a given guidance and produces a matting result for other objects. We found that the problem, the lack of mask awareness, comes from the previous data pipeline to generate training samples.", + "bbox": [ + 75, + 462, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a clear understanding, we first briefly summarize the previous training data generation step proposed in MGMatting [50]. Here, for clarity, generic data augmentation such as random-crop, resize, and affine-transform are omitted. A foreground object (consisting of foreground color $F$ and alpha matte $\\alpha$ ) and a background image $B$ are randomly sampled from the matting and background datasets, respectively. These are composited to form a training image $I$ following the composition formulation as follows:", + "bbox": [ + 75, + 569, + 468, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI = \\operatorname {C o m p o s i t e} (F, \\alpha , B) = \\alpha F + (1 - \\alpha) B. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 710, + 468, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To simulate the mask guidance $M$ , they first binarize the alpha matte with a threshold randomly sampled from 0 to 1, then apply dilation and/or erosion to it. In addition, they propose a stronger mask augmentation, CutMask [50], originally designed for robustness to noise in mask guidance. Like CutMix [51], a random-sized patch is selected, and the content is pasted on a random position.", + "bbox": [ + 75, + 734, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, MGMatting [50] adopted instance merging augmentation as most state-of-the-art trimap-based matting methods [21, 30, 41] did. Specifically, they select two random foreground objects, merge them into a single object,", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and composite it on a random background image. The instance merging augmentation effectively alleviates the limited scale and diversity of foreground objects in the matting dataset. However, in the context of mask-guided matting, we found that this augmentation hinders instance-wise understanding and makes a negative bias that the model eventually segments all the existing composited objects regardless of the given mask guidance. The different effects of the same augmentation are originated from the uncertainty of mask guidance. While the trimap brings pixel-level accurate localization of foreground/background/unknown regions, the mask guidance naturally conveys noisy information, and it is hard to localize them accurately. Under such uncertainty, the model tends to find the shortcut, segmenting all the composited objects rather than learning to utilize the uncertain mask guidance.", + "bbox": [ + 496, + 439, + 892, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-wise Learning. We systematically block the above shortcut and enforce the model to learn complex relations and soft transitions between objects. The idea is simple: supervise the model to segment one of them according to the guidance (See Fig. 3-(a)). Specifically, given the two foreground objects (colors $F_{1}, F_{2}$ and opacity $\\alpha_{1}, \\alpha_{2}$ ), we sequentially composite these foreground layers (the first object is on top.) on background image $B$ as follows:", + "bbox": [ + 496, + 681, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I = \\text {C o m p o s i t e} \\left(F _ {1}, \\alpha_ {1}, \\text {C o m p o s i t e} \\left(F _ {2}, \\alpha_ {2}, B\\right)\\right) \\tag {2} \\\\ = \\alpha_ {1} F _ {1} + \\alpha_ {2} (1 - \\alpha_ {1}) F _ {2} + (1 - \\alpha_ {1}) (1 - \\alpha_ {2}) B. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 811, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above equation, we re-calibrate the fractional contribution of each foreground object to images at pixel level as follows: $\\overline{\\alpha}_1 = \\alpha_1, \\overline{\\alpha}_2 = \\alpha_2 (1 - \\alpha_1)$ . We randomly", + "bbox": [ + 496, + 854, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1995", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "select an index $i_{rand} \\in \\{1,2\\}$ of a target object. With the guidance $M(\\overline{\\alpha}_i)$ , the model predicts the opacity of the target object $\\widetilde{\\alpha}_i$ , which is supervised by corresponding ground-truth alpha matte $\\overline{\\alpha}_i$ . Such instance-wise training makes the model learn fractional differentiation between objects according to the given coarse mask guidance.", + "bbox": [ + 75, + 90, + 470, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Learning from Images with Weak Supervision", + "text_level": 1, + "bbox": [ + 76, + 193, + 467, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the annotation difficulties, the labels of the alpha matte are practically infeasible to collect for large quantities and diverse categories. By comparison, coarse mask annotations are more abundant and accessible, thus extending the dataset scale and taxonomies is easier to achieve. Motivated by this, we leverage the weak supervision dataset (i.e., instance segmentation [24]), and investigate how these new data benefit the generalization of the mask-guided matting model. The main challenge lies in how to get matting supervision because the instance mask labels are noisy in fine details and do not include opacity annotations.", + "bbox": [ + 75, + 217, + 467, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To effectively tackle this challenge, we design a self-training framework (See Fig. 3-(b)). The framework generates pseudo matting labels $\\widehat{\\alpha}$ under the guidance of instance masks, and the pseudo labels supervise the matting model on natural images. To obtain high-quality pseudo labels, we adopt several design choices as follows.", + "bbox": [ + 75, + 383, + 467, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Teacher-Student Framework. Motivated by the success in self-/semi-supervised learning [9, 42], a teacher network is introduced, which is a slowly advancing version of a student model via the exponential moving average (EMA). We utilize stable predictions of the teacher network as the pseudo labels $\\widehat{\\alpha}$ and guide the student model during training.", + "bbox": [ + 75, + 474, + 467, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Weakly-augmented Input. The teacher network takes weakly-augmented samples to generate the pseudo labels. We only apply standard geometric augmentation, such as affine transform and random crop, for both image and mask guidance. Despite the noise in fine details, the ground-truth instance mask conveys a strong localization cue to the model, resulting in high-quality pseudo labels.", + "bbox": [ + 75, + 566, + 467, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For strongly-augmented inputs, the student model produces predictions $\\widehat{\\alpha}^{\\prime}$ , which are supervised to be the same as the pseudo labels $\\widehat{\\alpha}$ obtained from the teacher model. To form the strongly-augmented version, we design two types of augmentations: image and guidance perturbations.", + "bbox": [ + 75, + 672, + 467, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Perturbations. Previous self-training-based methods [36, 37, 47] often employ region-level augmentation (e.g., Gaussian blur) to generate the strongly-augmented images. However, in the matting task, these augmentations are unsuitable since they interpolate the color value across the pixels and disturb each object's fractional contribution. This results in a mismatch of ground-truth alpha matte between the teacher and student inputs, lowering the effectiveness of self-training. We instead adopt linear pixel-level augmentations such as additive Gaussian noise and linear", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "contrast.", + "bbox": [ + 498, + 92, + 560, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Guidance Perturbations. We also modulate the reliability of mask guidance via dilation and/or erosion of the instance mask. Considering the diverse size of objects in natural scenes, we decide the kernel size based on the size of the object. Specifically, we set the kernel size as $\\beta \\%$ of the shorter side of an object bounding box. The guidance perturbation is crucial to not only the success of self-training but also the robustness against common noise in mask guidance.", + "bbox": [ + 496, + 106, + 890, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Unified Training", + "text_level": 1, + "bbox": [ + 498, + 252, + 663, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, under the proposed framework, the (student) mask-guided matting model can jointly learn from composited images with ground-truth labels and natural images with weak supervision. The final loss function is a combination of losses on both images as: $L_{final} = L_{composed} + L_{natural}$ .", + "bbox": [ + 496, + 275, + 890, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the composited images, we use the same loss function as MGMatting [50], which is the summation of $l_{1}$ regression loss, composition loss [48], and Laplacian loss [13]. Thus, $L_{\\text{composed}} = L_{l1}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i) + L_{\\text{comp}}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i) + L_{\\text{lap}}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i)$ . To calculate the composition loss, we use ground-truth foreground color and alpha matte for non-target objects. For the natural images, we use $l_{1}$ regression loss as follows: $L_{\\text{natural}} = L_{l1}(\\widehat{\\alpha}', \\widehat{\\alpha})$ .", + "bbox": [ + 496, + 367, + 890, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Benchmarks", + "text_level": 1, + "bbox": [ + 498, + 501, + 630, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are several public matting benchmarks [19, 20, 32, 39, 48, 50] which provide high-quality alpha matte. However, none of them reflects the complexity of in-the-wild images and the consisting samples only contain a single object with simple backgrounds. Thus, we carefully design a new evaluation setup to facilitate an extensive evaluation of the mask-guided matting model in in-the-wild scenarios. The new setup consists of multiple benchmarks, and we describe the distinctive role of each benchmark below.", + "bbox": [ + 496, + 527, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Composition-Wild. We first propose an in-the-wild extension of standard synthetic benchmark Composition-1k [48], namely Composition-Wild. We simulate the complex context of in-the-wild images by compositing multiple foreground objects onto a background image, similar to Eq.(2). The same foreground object data and background dataset (i.e., PASCAL VOC [6]) are utilized as Composition-1k. We report the performance over the forefront object from each sample unless otherwise specified. In addition, we design comprehensive metrics to allow in-depth analysis.", + "bbox": [ + 496, + 662, + 890, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- $\\mathrm{SAD}_{FG}$ reports the SAD (Sum of Absolute Difference) over the target object region where has non-zero ground-truth alpha matte. It indicates how well the prediction captures fine matting details of target objects (i.e., Target Detail Quality).", + "bbox": [ + 504, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1996", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\mathrm{SAD}_{BG}$ quantifies the wrong predictions on other objects or background regions by measuring the SAD over the non-target object region. It represents the capability of target object localization from the noisy mask guidance (i.e., Localization).", + "- $\\mathrm{SAD}_{OCC}$ measures the SAD error of occluded objects. It indicates how well the model deals with occluded objects (i.e., Occlusion handling)." + ], + "bbox": [ + 81, + 90, + 467, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AIM-500 [20] is a natural image matting benchmark that contains diverse categories of objects. The performance evaluation on the AIM benchmark denotes whether the model generalizes well to natural images rather than overfitting to the synthetic distribution of composited images. However, AIM is limited at capturing matting performance in complex contexts since they only contain images with simple backgrounds. In that regard, the evaluations on Composition-Wild and AIM complement each other and provide a holistic understanding of the mask-guided matting model in the wild.", + "bbox": [ + 75, + 239, + 467, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "COCO [24] is a representative example of in-the-wild images. While we cannot access ground-truth alpha mattes on COCO, we conduct qualitative comparisons over diverse objects in their natural context.", + "bbox": [ + 75, + 407, + 467, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 492, + 207, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 520, + 292, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architecture details. We build our learning framework with a state-of-the-art mask-guided matting network, PRN from MGMatting [50], which adopts a standard encoder-decoder structure [21,33] and progressively refines the output through the decoding process.", + "bbox": [ + 75, + 546, + 467, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. In practice, we adopt the two-stage training: pre-training on composited images and fine-tuning on both composited and natural images. This ensures the quality of pseudo labels at the beginning of the self-training, resulting in better performance. Following the previous works [21, 27, 30, 48], we use DIM [48] and COCO [24] as matting and background datasets, respectively. COCO dataset [24] is also utilized for a weak guidance dataset. We set the kernel ratio $\\beta$ as 0.2.", + "bbox": [ + 75, + 625, + 467, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation details. We report the widely used four matting metrics, SAD (Sum of Absolute Differences), MSE (Mean Squared Error), Grad (Gradient), and Conn (Connectivity), over the whole image, when the ground-truth alpha matte is available. To simulate the coarse mask guidance, we binarize the alpha matte with threshold 0.5 and then dilate the mask with kernel size $25 \\times 25$ , similar to [53]. We use a single trained model for all the benchmark evaluations and real-world applications.", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85494dbe39f0d64f6f98bdd7e733b9d89712e274bcaf7d2f5e5361e3741c7514.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 89, + 638, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d5f460f9a0447df5882b22c361b98d93a02b47362a37f2540e81c8b45b25518.jpg", + "image_caption": [ + "Image and Guidance" + ], + "image_footnote": [], + "bbox": [ + 503, + 166, + 637, + 291 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/991b689f261faabd25116c9e2e44ce0fa6a62addcc71db2dea685d71754872d8.jpg", + "image_caption": [ + "MGMatting", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 638, + 89, + 900, + 291 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7bdae739ce83e3eb04d28609ee72e38147012b5fa5c015c34fcc4e6a8afe6e39.jpg", + "image_caption": [ + "Figure 4. Qualitative Results on COCO dataset. The guidance are either manually drawn [24] (upper) or a prediction of the off-the-self model [10] (bottom). Best viewed zoomed in.", + "(a) Image", + "Figure 5. Effect of Instance-wise Learning." + ], + "image_footnote": [], + "bbox": [ + 501, + 368, + 602, + 472 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/850a5a2772428642e7211cd0982d543f571e6268834e5a050991c7f31a267384.jpg", + "image_caption": [ + "(b) Guidance" + ], + "image_footnote": [], + "bbox": [ + 602, + 368, + 687, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/330801958f40f89e03e7c8e609fa6d87da55a0556c07559395908fdbe7dfafaf.jpg", + "image_caption": [ + "(c) MGMatting*" + ], + "image_footnote": [], + "bbox": [ + 705, + 368, + 800, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e70c3092367d9616a17feb8b0e86241f38936383df0af1c909b76a7b3fbc9f34.jpg", + "image_caption": [ + "(d) +IWL" + ], + "image_footnote": [], + "bbox": [ + 802, + 368, + 900, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 500, + 527, + 728, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 1 and Fig. 4 illustrate the qualitative comparisons on images in the COCO validation sets. We test the models under challenging scenarios such as occlusions, an unseen object in a matting dataset, complex surrounding context, and large interior errors on mask guidance. Compared with the previous mask-guided matting model [50], our model shows significantly better localization of a target object, as well as, produces much more accurate alpha mattes for challenging data. More qualitative comparisons can be found in the supplementary material.", + "bbox": [ + 496, + 550, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 500, + 710, + 754, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative evaluations on Composition-Wild and AIM-500 datasets are shown in Table 1.", + "bbox": [ + 500, + 734, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of Instance-wise Learning. We first analyze how instance-wise learning contributes to learn generalized matting capability from composited images. We set the two baseline models: (naive) MGMatting and MGMatting*. For the (naive) MGMatting, we reproduce the model with the same training recipes proposed in MGMatting [50]. In addition, we report the stronger baseline score of MGMatting* by excluding a heavy guidance augmentation, CutMask [50]. We empirically found that CutMask", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1997", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ddaaccbfd20b729cb3aa56cf03eeb413e3479ba94b43473fb31fe99306070031.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodComposition-WildAIM-500 [20]
SADMSEGradConn\\( SAD_{FG} \\)\\( SAD_{BG} \\)\\( SAD_{OCC} \\)SADMSEGradConn
Stage 1: Learning with Composited Images
MGMatting [50]583.360.1328139.8452.5465.51517.85842.8871.910.026823.3721.97
MGMatting* [50]389.150.0727141.8044.2155.74333.42524.9926.180.005615.8114.53
+Instance-wise (Ours)67.270.005842.3543.5249.5017.7759.5921.120.003816.8814.72
Stage 2: Learning with Composited and Weak-guidance Images
Segmentation-training69.240.006957.0846.5654.3614.8865.4220.940.004320.4614.20
Self-training (Ours)58.160.004639.0441.3747.3210.8453.0216.720.003014.6812.02
", + "bbox": [ + 81, + 88, + 890, + 232 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "makes substantial random localization errors in guidance and lowers the reliability of mask guidance during training. Training with such guidance makes models ignore the mask guidance and overfit to synthetic samples, resulting in inferior in-the-wild performance. On top of MGMatting*, the \"Instance-wise Learning\" strategy drastically benefits the mask-guided matting model. Unlike the baseline model that segments all the composited objects (Fig. 5-(c)), the proposed strategy allows the model to well discriminate the target object (Fig. 5-(d)). Our model faithfully utilizes the coarse mask guidance and shows a huge gain in localization performance (see $\\mathrm{SAD}_{BG}$ metric). Although the strategy assumes multi-object scenarios and learns the complex relations between them, it also contributes to the better matting details for an object (see $\\mathrm{SAD}_{FG}$ in Composition-Wild and SAD in AIM). By design, the instance-wise learning naturally improves occlusion handling by learning the order of layered foreground objects and their fractional contributions.", + "bbox": [ + 75, + 295, + 472, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Self-training. To extend the mask-guided matting model to the in-the-wild setting, we leverage the weak-supervision dataset under the proposed self-training framework. We compare our self-training framework to a baseline, namely \"segmentation-training\". This baseline is motivated by previous human-matting methods [23, 25]. It attempts to learn generalized matting capability via simultaneous segmentation training on natural images. To be specific, the model takes strongly augmented instance masks as guidance and is supervised to recover ground-truth instance masks through the additional segmentation head. As deteriorated performance indicates, the segmentation-training strategy fails to scale up to the in-the-wild setting. While it enhances the localization of target objects, supervision from binary instance masks hinders learning diverse matting patterns in the wild (see $\\mathrm{SAD}_{FG}$ metric), especially for transparent objects. On the contrary, our self-training framework successfully generates the soft-pseudo label and utilizes it as a direct supervision to matting tasks, resulting in huge improvements in all the metrics.", + "bbox": [ + 75, + 583, + 472, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on Perturbation in Self-training. To demon-", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/23bf6862a4f17113e4fc07af0f26b35eec848c508bab1ca0e17a17b67d6880f9.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on Composition-Wild and AIM-500 [20]. * denotes the improved version of MGMatting [50]. \"Instance-wise\" represents instance-wise learning. A lower score is better in all the metrics." + ], + "table_footnote": [], + "table_body": "
MethodComposition-WildAIM-500
SADMSE\\( SAD_{FG} \\)\\( SAD_{BG} \\)SADMSE
Ours58.160.004647.3210.8416.720.0030
No Image.60.080.004845.1914.8918.570.0039
No Guide.60.140.004947.9512.1918.270.0035
", + "bbox": [ + 501, + 292, + 905, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strate the effectiveness of perturbations in self-training, we ablate the image and guidance perturbations separately. The results are summarized in Table 2. As degraded performance denoted, encouraging robustness against image and guidance perturbations is crucial to the success of the self-training.", + "bbox": [ + 496, + 445, + 890, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further investigate whether the matting-specific design of image perturbation is necessary. To this end, we train three variants with different configurations of image perturbations: (1) Pixel-level linear augmentation (Pixel), (2) Region-level augmentation (Region) and (3) Both of them (Pixel+Region). Below, we report the SAD metric for Composition-Wild and AIM-500, respectively.", + "bbox": [ + 496, + 537, + 892, + 643 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/067a6aaa2edb0d525533ce28df07c5fcc200d89bcd4e05f560d4b36985b026af.jpg", + "table_caption": [ + "Table 2. Ablation Study on Perturbations in Self-training." + ], + "table_footnote": [], + "table_body": "
PixelRegionPixel+Region
58.16 / 16.7269.8 / 25.9661.88 / 19.00
", + "bbox": [ + 527, + 654, + 864, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can observe that the region-level augmentation, which is widely used in other tasks [37, 47], disturbs the effects of self-training objectives. As discussed in Sec. 4.2, this is because the region-level augmentation breaks the underlying opacity of objects and make a mismatch between inputs of the student and the teacher networks. On the contrary, pixel-level linear augmentation brings clear improvements.", + "bbox": [ + 496, + 691, + 890, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Real-world Applications", + "text_level": 1, + "bbox": [ + 500, + 813, + 728, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To show the generality and potential of the in-the-wild mask-guided matting model, we illustrate two new applications: mask-guided video matting and panoptic matting. Fig. 6 and Fig. 7 show examples of each application.", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1998", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df35dffa257ae87bf259cec91d4214e5e7450c408112602afa8a2503d958df7a.jpg", + "image_caption": [ + "(a) Frame" + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 256, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e9de2c76caaac8aa0a266288a8f61d7e1c3ed418a21e73e787fc981f61e7f00.jpg", + "image_caption": [ + "(b) Guidance" + ], + "image_footnote": [], + "bbox": [ + 259, + 88, + 354, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5cbbd3e6a457d0eb16f53a3859b2898334e3d573f026143dc7a6d4f47941bd6d.jpg", + "image_caption": [ + "(c) MGMatting" + ], + "image_footnote": [], + "bbox": [ + 405, + 89, + 558, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6531bcc6bed2208c88443aebe8bc15bcbcf08a5c157b8bf46724e41ae4f25859.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 558, + 90, + 705, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66ce27abec34c50cf96fa1adf7bc7507939e0e06e27c01b5dd92f06e1bb2c74f.jpg", + "image_caption": [ + "(e) Editing Result" + ], + "image_footnote": [], + "bbox": [ + 707, + 89, + 856, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1675affb9f2e296a2dbd2395b28fa95cbf03fb58770c4c717f70d8a27156e4e9.jpg", + "image_caption": [ + "Figure 6. Extension to Mask-guided Video Matting. We use propagated masks [31] as guidance." + ], + "image_footnote": [], + "bbox": [ + 109, + 260, + 406, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/806838b63f2647d6f10ef948ea06202462efe355996cb41a7562946d6e6a2302.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 260, + 557, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/201a1a4f25cb8dad2c51e2b1e60071e2971b5bb8fc13d611f56e404bec6df81e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 260, + 705, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6091fd7c8acb78a6215beb65c9a19aebdde2ddc179c98a746eefcda6ebf432ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 261, + 856, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ab6844626e07b606e6391ffeea22475bfd024fbf199e97a090878a54c07d237.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 109, + 338, + 259, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c11a8b55ec1aebfc600dd74aa08e4d13f17b4e295cd9317a548edfc3085d545.jpg", + "image_caption": [ + "(b) Guidance" + ], + "image_footnote": [], + "bbox": [ + 259, + 338, + 410, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e87bd9bd5a3d84b551c8f626628476d82b338360267358f46aaa5ec7d99919d0.jpg", + "image_caption": [ + "(c) MGMatting" + ], + "image_footnote": [], + "bbox": [ + 410, + 338, + 555, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5e894778455953555281dbb3978bc6e85ff6bad80304c7da9c69fef306fd2ef.jpg", + "image_caption": [ + "(d) Ours", + "Figure 7. Extension to Panoptic Matting. Hand drawn panoptic masks [24] are utilized as guidance. We illustrate independent regions with different colors. The black color represents ignored regions of labels or predictions." + ], + "image_footnote": [], + "bbox": [ + 557, + 338, + 705, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4decbd3b062b64b1a670503cd46657166d80d97315fc538694d288346b7f772.jpg", + "image_caption": [ + "(e) Editing Result" + ], + "image_footnote": [], + "bbox": [ + 705, + 338, + 856, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.1. Mask-guided Video Matting", + "text_level": 1, + "bbox": [ + 76, + 515, + 328, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Current video matting solutions [40, 53] require multiple frames of trimap to produce an alpha matte for a video. Therefore, required user interaction is more burdensome in the video domain. With the mask-guided matting model, we explore a more practical scenario where only a single instance mask is given in the first frame. In this challenging setting, we first propagate the given mask to the rest of the video frames using the video object segmentation model [31]. Then, the mask-guided matting model utilizes the propagated mask as guidance and produces alpha matte in a frame-by-frame manner. Our model not only corrects the noise in propagated masks but also predicts reasonable opacity in challenging cases such as motion blur. With the prediction, we also provide visually appealing video editing results (i.e., background replacement).", + "bbox": [ + 75, + 541, + 468, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. Panoptic Matting", + "text_level": 1, + "bbox": [ + 76, + 784, + 246, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Panoptic segmentation aims to parse an image into nonoverlapping regions of stuff and things. By the definition of the task format, the soft transition between the regions is overlooked. We explore whether our mask-guided matting model could provide such understanding from panoptic segmentation results. Given a panoptic mask, our model con", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "siders the segment of each region separately and produces the corresponding alpha matte. We form panoptic matting results by aggregating the predictions. Unlike the baseline, our model successfully refines the fine details and captures the soft transition between the regions. Interestingly, while the model is trained on objects (i.e., things), our method is well generalized on the stuff regions (See sky region in Fig. 7). Panoptic matting enables us to do region-level image editing and the qualitative results are shown in Fig. 7.", + "bbox": [ + 496, + 516, + 890, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 664, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a simple yet effective learning framework for mask-guided matting in the wild. We first introduce instance-wise learning to learn generalized matting ability from the composited images. Then, the model is further extended to handle diverse objects under the self-training on weak-guidance images. Extensive experiments on the newly proposed evaluation benchmark provide an in-depth understanding of the unique challenges. Using our approach, we instantiate new practical applications: mask-guided video matting and panoptic matting. We hope our proposals inspire future work on mask-guided matting.", + "bbox": [ + 496, + 689, + 890, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was partially supported by the NRF (NRF-2020M3H8A1115028, FY2021) and Samsung Electronics Co., Ltd (G01200447).", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1999", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tamara L Berg and Alexander C Berg. Finding iconic images. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, pages 1-8. IEEE, 2009. 1", + "[2] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 1", + "[3] Quan Chen, Tiezheng Ge, Yanyu Xu, Zhiqiang Zhang, Xinxin Yang, and Kun Gai. Semantic human matting. In Proceedings of the 26th ACM international conference on Multimedia, pages 618-626, 2018. 3", + "[4] Qifeng Chen, Dingzeyu Li, and Chi-Keung Tang. Knn matting. IEEE transactions on pattern analysis and machine intelligence, 35(9):2175-2188, 2013. 2", + "[5] Yung-Yu Chuang, Brian Curless, David H Salesin, and Richard Szeliski. A bayesian approach to digital matting. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 2, pages II-II. IEEE, 2001. 2", + "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 5", + "[7] Marco Forte and François Pitie. $f, b$ , alpha matting. arXiv preprint arXiv:2003.07711, 2020. 1, 2", + "[8] Eduardo SL Gastal and Manuel M Oliveira. Shared sampling for real-time alpha matting. In Computer Graphics Forum, volume 29, pages 575-584. Wiley Online Library, 2010. 2", + "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 5", + "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 6", + "[11] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaouu Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 2", + "[12] Kaiming He, Jian Sun, and Xiaou Tang. Fast matting using large kernel matting laplacian matrices. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2165-2172. IEEE, 2010. 2", + "[13] Qiqi Hou and Feng Liu. Context-aware image matting for simultaneous foreground and alpha estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4130-4139, 2019. 1, 2, 5", + "[14] Ronghang Hu, Piotr Dólar, Kaiming He, Trevor Darrell, and Ross Girshick. Learning to segment every thing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4233-4241, 2018. 2" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson W.H. Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In AAAI, 2022. 1, 2, 3", + "[16] Philip Lee and Ying Wu. Nonlocal matting. In CVPR 2011, pages 2193-2200. IEEE, 2011. 2", + "[17] Anat Levin, Dani Lischinski, and Yair Weiss. A closed-form solution to natural image matting. IEEE transactions on pattern analysis and machine intelligence, 30(2):228-242, 2007. 1, 2", + "[18] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 2", + "[19] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 2, 3, 5", + "[20] Jizhizi Li, Jing Zhang, and Dacheng Tao. Deep automatic natural image matting. arXiv preprint arXiv:2107.07235, 2021. 2, 5, 6, 7", + "[21] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11450-11457, 2020. 1, 2, 4, 6", + "[22] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 1", + "[23] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 7", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 3, 5, 6, 8", + "[25] Jinlin Liu, Yuan Yao, Wendi Hou, Miaomiao Cui, Xuansong Xie, Changshui Zhang, and Xian-sheng Hua. Boosting semantic human matting with coarse annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8572, 2020. 3, 7", + "[26] Yuhao Liu, Jiake Xie, Xiao Shi, Yu Qiao, Yujie Huang, Yong Tang, and Xin Yang. Tripartite information mining and integration for image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7555-7564, 2021. 2", + "[27] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3266-3275, 2019. 1, 6", + "[28] Sebastian Lutz, Konstantinos Amplanitis, and Aljosa Smolic. Alphagan: Generative adversarial networks for natural image matting. arXiv preprint arXiv:1807.10088, 2018. 2", + "[29] Stephen Palmer. Canonical perspective and the perception of objects. Attention and performance, pages 135-151, 1981. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2000", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11696-11706, 2022. 1, 2, 4, 6", + "[31] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Per-clip video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1352–1361, 2022. 8", + "[32] Yu Qiao, Yuhao Liu, Xin Yang, Dongsheng Zhou, Mingliang Xu, Qiang Zhang, and Xiaopeng Wei. Attention-guided hierarchical structure aggregation for image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13676-13685, 2020. 1, 2, 5", + "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6", + "[34] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 1, 2", + "[35] Ehsan Shahrian, Deepu Rajan, Brian Price, and Scott Cohen. Improving image matting using comprehensive sampling sets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 636-643, 2013. 2", + "[36] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 5", + "[37] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2, 5, 7", + "[38] Jian Sun, Jiaya Jia, Chi-Keung Tang, and Heung-Yeung Shum. Poisson matting. In ACM SIGGRAPH 2004 Papers, pages 315–321. 2004. 2", + "[39] Yanan Sun, Chi-Keung Tang, and Yu-Wing Tai. Semantic image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11120–11129, 2021. 5", + "[40] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 8", + "[41] Jingwei Tang, Yagiz Aksoy, Cengiz Oztireli, Markus Gross, and Tunc Ozan Aydin. Learning-based sampling for natural image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3055–3063, 2019. 2, 4", + "[42] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 5", + "[43] Jue Wang and Michael F Cohen. An iterative optimization approach for unified image segmentation and matting. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 936-943. IEEE, 2005. 1", + "[44] Jue Wang and Michael F Cohen. Optimized color sampling for robust matting. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 2", + "[45] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Hanqing Zhao, Weiming Zhang, and Nenghai Yu. Improved image matting via real-time user clicks and uncertainty estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15374-15383, 2021. 1, 2", + "[46] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXV, pages 238–258. Springer, 2022. 2", + "[47] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 2, 5, 7", + "[48] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 3, 5, 6", + "[49] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 3", + "[50] Qihang Yu, Jianming Zhang, He Zhang, Yilin Wang, Zhe Lin, Ning Xu, Yutong Bai, and Alan Yuille. Mask guided matting via progressive refinement network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1154-1163, 2021. 1, 2, 3, 4, 5, 6, 7", + "[51] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 4", + "[52] Yunke Zhang, Lixue Gong, Lubin Fan, Peiran Ren, Qixing Huang, Hujun Bao, and Weiwei Xu. A late fusion cnn for digital matting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7469-7478, 2019. 2", + "[53] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 881 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2001", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_model.json b/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c409c03a5452ff36e31d080915125803569de548 --- /dev/null +++ b/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.315, + 0.131, + 0.658, + 0.154 + ], + "angle": 0, + "content": "Mask-guided Matting in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.18, + 0.861, + 0.234 + ], + "angle": 0, + "content": "Kwanyong Park\\(^{1}\\) Sanghyun Woo\\(^{1}\\) Seoung Wug Oh\\(^{2}\\) In So Kweon\\(^{1}\\) Joon-Young Lee\\(^{2}\\) \n\\(^{1}\\)KAIST \\(^{2}\\)Adobe Research" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.604 + ], + "angle": 0, + "content": "Mask-guided matting has shown great practicality compared to traditional trimap-based methods. The mask-guided approach takes an easily-obtainable coarse mask as guidance and produces an accurate alpha matte. To extend the success toward practical usage, we tackle mask-guided matting in the wild, which covers a wide range of categories in their complex context robustly. To this end, we propose a simple yet effective learning framework based on two core insights: 1) learning a generalized matting model that can better understand the given mask guidance and 2) leveraging weak supervision datasets (e.g., instance segmentation dataset) to alleviate the limited diversity and scale of existing matting datasets. Extensive experimental results on multiple benchmarks, consisting of a newly proposed synthetic benchmark (Composition-Wild) and existing natural datasets, demonstrate the superiority of the proposed method. Moreover, we provide appealing results on new practical applications (e.g., panoptic matting and mask-guided video matting), showing the great generality and potential of our model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.633, + 0.21, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Image matting aims to predict the opacity of objects, which enables precise separation from surrounding backgrounds. Due to the ill-posed nature of the task, many works [7, 13, 21, 27, 30, 48] have improved matting performance by relying on the manual guidance of a trimap. However, pixel-level annotation of foreground/background/unknown is extremely burdensome, restricting its usage in many practical applications such as image/video editing and film production. Recently, many efficient alternatives for user guidance have been proposed, including trimap-free [15, 32], additional background images [22, 34], scribble [43], and the user clicks [45]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Among them, the mask-guided approach [50] shows a great trade-off between performance and intensity of user interaction. It utilizes a coarse mask as guidance, which is much easier to obtain either manually or from off-the-shelf" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.268, + 0.911, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.501, + 0.895, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Qualitative Comparisons of MGMatting [50] and Ours in the wild. The mask guidance is overlaid on images with blue color. Best viewed zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.701 + ], + "angle": 0, + "content": "segmentation models [2, 10]. With only the coarse spatial prior, the mask-guided matting model [50] shows comparable or even better performance than the trimap-based competitors [13, 17, 21, 27, 48] on synthetic Composition-1k [48] and a real-world human matting dataset [50]. However, despite the encouraging results, we see the previous state-of-the-art model [50] struggles to obtain desirable alpha matte in complex real-world scenes (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.896, + 0.903 + ], + "angle": 0, + "content": "With this observation, we tackle mask-guided matting in the wild. Specifically, we formulate unique setups and emerging challenges of the new task as follows: (1) We aim to handle objects in their complex context, reflecting the characteristics of natural images. The previous method [50] evaluates their model on iconic-object images [1,29] where only a single object is in the center. As the model can easily find the target object in such images, the model's real instance discrimination ability is, in fact, veiled. On the contrary, in an 'in-the-wild' setting, it is crucial to precisely localize the target object from the given coarse/noisy mask guidance (i.e. mask awareness). (2) Our model targets to deal with diverse categories of objects in natural" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1992" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "images. Unlike most previous methods that improve generalization performance at the expense of category-specific regime (e.g., limiting to humans [15] or animals [19]), we aim to understand distinctive matting patterns of vast categories. (3) Limited data problem makes the new setting more complicated. Due to the labeling complexity, annotating alpha matte for objects in common scenes, e.g., the COCO dataset [24], is infeasible. As a sidestep, previous benchmarks [32, 48] extract the alpha matte and foreground colors from images with simple backgrounds. These are composited on various backgrounds [6, 24], and resulting samples are used to train and evaluate matting models. However, due to the inevitable composition artifacts, the models usually show limited generalization performance. In that sense, how to train and evaluate the in-the-wild matting model remains an open question." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.473, + 0.625 + ], + "angle": 0, + "content": "Toward this goal, we propose a simple yet effective learning framework for a generalized mask-guided matting model. First, we investigate fundamental reasons for the poor generalization of the previous mask-guided matting model [50] and find that this is mainly from the training data generation process. Specifically, the previous composition process includes instance merging data augmentation, which merges several foreground objects into a single object. While this augmentation is effective in the trimap-based methods [21, 30, 41], it implicitly makes a negative bias for the mask-guided matting model to ignore the guidance. Thus, the model struggles to localize the target objects in complex natural scenes. We alleviate the bias by proposing an instance-wise learning objective, where the model is supervised to segment one of multiple instances according to the guidance. By doing so, the model learns strong semantic representation regarding complex relations and soft transitions between objects. Despite the simplicity of the proposal, this greatly improves performance in the wild." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.473, + 0.853 + ], + "angle": 0, + "content": "Second, we explore a practical solution to make the mask-guided model handle various categories of objects robustly. Instead of scaling the matting dataset, we leverage a dataset with weak supervision [14, 46] (i.e., instance segmentation dataset [24]), as the coarse instance masks are easier to obtain over the diverse categories of objects. To effectively hallucinate the fine supervision signal with the weak localization guidance, we come up with a self-training framework [36, 37, 47]. Specifically, a pseudo label is generated based on a weakly-augmented input (both image and instance mask annotation as guidance), which supervises the model prediction on a strongly augmented version of them. During self-training, the model is not only adapting to the in-the-wild scenario in a self-evolving manner but also being robust to noise in both image and guidance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To verify the in-the-wild performance of mask-guided matting, we formally define an evaluation protocol involving multiple sub-benchmarks: Composition-Wild, AIM-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.287 + ], + "angle": 0, + "content": "500 [20], COCO [24]. We first design an in-the-wild extension of the popular synthetic Composition-1k benchmark [48], namely Composition-Wild. We simulate complex real-world images by compositing multiple foreground objects. To bring valuable insight on the failure cases of the model, we design sub-metrics for Composition-Wild. In addition, we use the AIM-500 dataset to establish quantitative results on natural images (i.e., with no composition artifacts), although most images are iconic-object images with simple backgrounds. Finally, we provide qualitative outputs of our mask-guided matting model on the COCO dataset [24] which is one of the most representative in-the-wild datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.396 + ], + "angle": 0, + "content": "To summarize, we make the following main contributions. 1) To our best knowledge, it is the first work to explore mask-guided matting in the wild. 2) We develop a simple yet effective learning framework leveraging both composited and weak-guidance images. 3) We design an evaluation setup for the new task. 4) We initiate several interesting extensions: video and panoptic matting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.651, + 0.423 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Natural Image Matting. Most image matting methods require a trimap as additional input, which conveys pixel-level annotations of foreground, background, and unknown region. Traditional image matting methods can be categorized into two groups. Sampling-based methods [5, 8, 11, 35, 44] estimate the alpha matte of the unknown region based on sampled colors of foreground and background. Propagation-based methods [4, 12, 16-18, 38] propagate the neighboring known alpha values to unknown regions according to the affinity between pixels. Both methods primarily rely on color or low-level features, showing limitations in complicated scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.895, + 0.75 + ], + "angle": 0, + "content": "To tackle these issues, many deep-learning based mating methods have been proposed. DIM [48] is a representative work which proposes a convolutional encoder-decoder matting network as well as a large-scale synthetic dataset (Composition-1k) to train the model. Many follow-up works have made tremendous improvements in diverse aspects, such as sophisticated loss design [7, 13, 28] and architectural advances by introducing attention mechanisms [21, 26] or transformer architecture [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "There is a large volume of trials to relax the heavy user-supplied constraints. Some methods [32, 52] attempt to get rid of trimap. However, they show inferior performance and cannot be generalized to unseen objects in the real world. Sengupta et al. [34] utilize additional background images along with other lightweight priors (e.g., segmentation mask) to perform the matting task. Wei et al. [45] introduce user click interaction to effectively eliminate the ambiguity of target object. Recently, MGMatting [50] proposes a mask-guided matting framework, where only the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1993" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.427, + 0.106 + ], + "angle": 0, + "content": "easily obtainable coarse mask is needed as guidance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.182 + ], + "angle": 0, + "content": "In this paper, we further study mask-guided matting in an in-the-wild setting, inspired by the generality and practicality of the framework. Different from MGMatting, we target diverse categories of objects in their complex background, enabling real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.392 + ], + "angle": 0, + "content": "Class-specific Matting is a special type of matting that limits target objects to one or few classes such as humans [3] or animals [19]. BSHM [25] leverages the coarse segmentation mask data to build a generalized human matting network. MODNet [15] presents a lightweight network architecture with decomposed multi-scale network designs. GFM [19] first proposes the task of animal matting and a real-world animal matting dataset. In general, known semantics effectively ease the difficulty of matting, and these methods show great generality without the necessity of additional guidance input. However, the specialized methods and trained model cannot be generalized to the in-the-wild setting since limited semantics do not cover diverse patterns of visual worlds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.468, + 0.5 + ], + "angle": 0, + "content": "Apart from previous efforts, we attempt to extend the short category regime of matting networks. For this purpose, we propose an efficient solution by leveraging the instance segmentation dataset [24], which covers diverse categories of objects with weak localization annotations. Given the weak guidance dataset, we study the challenges in developing an in-the-wild mask-guided matting model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.383, + 0.53 + ], + "angle": 0, + "content": "3. Mask-guided Matting in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.468, + 0.748 + ], + "angle": 0, + "content": "In natural scenes, objects are rarely in isolation: diverse categories of entities interact with each other. We aim to robustly separate a target object from the surroundings in such scenes. Specifically, we formulate the problem in the form of mask-guided matting, namely mask-guided matting in the wild. Our matting model takes an easily-obtainable coarse mask as guidance, either manually drawn or predicted from the off-the-shelf segmentation models. With in-the-wild data, the model needs to understand semantics so that it can precisely localize the target object from the given noisy masks. In the meantime, the model also needs to have strong low-level feature representation so that it can capture fine details and accurate opacity of the diverse target matte." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To concrete the new challenges toward such an in-the-wild setting, we first study how the previous model [50] performs. We derive predictions with different mask guidance and summarize the results on Fig. 2. As shown in the upper example, in the synthetic Composition-1k [48] example, the previous model perfectly separates the foreground object regardless of the given guidance. Surprisingly, it produces an alpha matte of the foreground object when the mask of the background region is given as guidance. On the contrary, the model lacks generalization on" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.639, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.234, + 0.621, + 0.244 + ], + "angle": 0, + "content": "Image and Guidance" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.09, + 0.772, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.234, + 0.766, + 0.244 + ], + "angle": 0, + "content": "With \\(1^{\\mathrm{st}}\\) Mask Guidance" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.09, + 0.91, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.234, + 0.9, + 0.244 + ], + "angle": 0, + "content": "With \\(2^{\\text{nd}}\\) Mask Guidance" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.255, + 0.892, + 0.297 + ], + "angle": 0, + "content": "Figure 2. Predictions of MGMatting [50] with different guidance. (upper) synthetic Composition-1k image [48]. (bottom) Complex real world image [49]. Best viewed zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.892, + 0.573 + ], + "angle": 0, + "content": "complex real-world scenes (see bottom example). In particular, the model fails to localize the target object and makes wrong predictions on regions of similar color distributions (e.g., shadows) or another nearby object. From the two representative examples, we make several observations: 1) The previous method yields a model overfitted to the (trained) synthetic matting dataset, which memorizes the limited patterns of foreground objects rather than semantically understanding the mask guidance for localizing and segmenting the objects; 2) Such a model fails to precisely localize and segment a target object in the wild; 3) Previously used synthetic benchmark, Composition-1k [48], neither represents the complex nature of the in-the-wild setting nor systematically blocks the shortcut of the overfitting issue. Thus the Composition-1k may not be a suitable benchmark for mask-guided matting in the wild." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.891, + 0.636 + ], + "angle": 0, + "content": "To this end, we explore how to train a generalized mask-guided matting model given the limited scale and diversity of the matting dataset (Sec. 4), as well as revisit the evaluation benchmark to test the model (Sec. 5)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.661, + 0.673, + 0.679 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The overview of the proposed training framework is illustrated in Fig. 3. Our framework involves three types of datasets: 1) matting dataset that includes accurate alpha matte and corresponding foreground colors; 2) background dataset, where the foreground objects are composited to create training samples; 3) weak supervision dataset that is introduced to extend the matting model to the in-the-wild setting. The main goal of our framework is to learn accurate matting ability from the synthetic composited images and generalize the knowledge to the in-the-wild domain. To this end, we first investigate how the model could learn generalized knowledge from the composited images (Sec. 4.1), then how the model is further improved with weak supervision data (Sec. 4.2)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1994" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.835, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.261, + 0.815, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.386, + 0.893, + 0.415 + ], + "angle": 0, + "content": "Figure 3. The Overview of the Proposed Framework. (a) Learning from composited images with instance-wise learning (Sec. 4.1). (b) Learning from weak-supervision images with self-training (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.44, + 0.383, + 0.458 + ], + "angle": 0, + "content": "4.1. Learning from Composed Images" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.569 + ], + "angle": 0, + "content": "As discussed in the Sec. 3, the previous model [50] shows poor generalization performance in the challenging real-world setting. Specifically, it often fails to identify the target object in a given guidance and produces a matting result for other objects. We found that the problem, the lack of mask awareness, comes from the previous data pipeline to generate training samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.705 + ], + "angle": 0, + "content": "For a clear understanding, we first briefly summarize the previous training data generation step proposed in MGMatting [50]. Here, for clarity, generic data augmentation such as random-crop, resize, and affine-transform are omitted. A foreground object (consisting of foreground color \\( F \\) and alpha matte \\( \\alpha \\)) and a background image \\( B \\) are randomly sampled from the matting and background datasets, respectively. These are composited to form a training image \\( I \\) following the composition formulation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.712, + 0.469, + 0.728 + ], + "angle": 0, + "content": "\\[\nI = \\operatorname {C o m p o s i t e} (F, \\alpha , B) = \\alpha F + (1 - \\alpha) B. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.84 + ], + "angle": 0, + "content": "To simulate the mask guidance \\(M\\), they first binarize the alpha matte with a threshold randomly sampled from 0 to 1, then apply dilation and/or erosion to it. In addition, they propose a stronger mask augmentation, CutMask [50], originally designed for robustness to noise in mask guidance. Like CutMix [51], a random-sized patch is selected, and the content is pasted on a random position." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In addition, MGMatting [50] adopted instance merging augmentation as most state-of-the-art trimap-based matting methods [21, 30, 41] did. Specifically, they select two random foreground objects, merge them into a single object," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.893, + 0.682 + ], + "angle": 0, + "content": "and composite it on a random background image. The instance merging augmentation effectively alleviates the limited scale and diversity of foreground objects in the matting dataset. However, in the context of mask-guided matting, we found that this augmentation hinders instance-wise understanding and makes a negative bias that the model eventually segments all the existing composited objects regardless of the given mask guidance. The different effects of the same augmentation are originated from the uncertainty of mask guidance. While the trimap brings pixel-level accurate localization of foreground/background/unknown regions, the mask guidance naturally conveys noisy information, and it is hard to localize them accurately. Under such uncertainty, the model tends to find the shortcut, segmenting all the composited objects rather than learning to utilize the uncertain mask guidance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Instance-wise Learning. We systematically block the above shortcut and enforce the model to learn complex relations and soft transitions between objects. The idea is simple: supervise the model to segment one of them according to the guidance (See Fig. 3-(a)). Specifically, given the two foreground objects (colors \\( F_{1}, F_{2} \\) and opacity \\( \\alpha_{1}, \\alpha_{2} \\)), we sequentially composite these foreground layers (the first object is on top.) on background image \\( B \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.812, + 0.892, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I = \\text {C o m p o s i t e} \\left(F _ {1}, \\alpha_ {1}, \\text {C o m p o s i t e} \\left(F _ {2}, \\alpha_ {2}, B\\right)\\right) \\tag {2} \\\\ = \\alpha_ {1} F _ {1} + \\alpha_ {2} (1 - \\alpha_ {1}) F _ {2} + (1 - \\alpha_ {1}) (1 - \\alpha_ {2}) B. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Based on the above equation, we re-calibrate the fractional contribution of each foreground object to images at pixel level as follows: \\(\\overline{\\alpha}_1 = \\alpha_1, \\overline{\\alpha}_2 = \\alpha_2 (1 - \\alpha_1)\\). We randomly" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1995" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.184 + ], + "angle": 0, + "content": "select an index \\( i_{rand} \\in \\{1,2\\} \\) of a target object. With the guidance \\( M(\\overline{\\alpha}_i) \\), the model predicts the opacity of the target object \\( \\widetilde{\\alpha}_i \\), which is supervised by corresponding ground-truth alpha matte \\( \\overline{\\alpha}_i \\). Such instance-wise training makes the model learn fractional differentiation between objects according to the given coarse mask guidance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.194, + 0.468, + 0.21 + ], + "angle": 0, + "content": "4.2. Learning from Images with Weak Supervision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.468, + 0.383 + ], + "angle": 0, + "content": "Due to the annotation difficulties, the labels of the alpha matte are practically infeasible to collect for large quantities and diverse categories. By comparison, coarse mask annotations are more abundant and accessible, thus extending the dataset scale and taxonomies is easier to achieve. Motivated by this, we leverage the weak supervision dataset (i.e., instance segmentation [24]), and investigate how these new data benefit the generalization of the mask-guided matting model. The main challenge lies in how to get matting supervision because the instance mask labels are noisy in fine details and do not include opacity annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.468, + 0.476 + ], + "angle": 0, + "content": "To effectively tackle this challenge, we design a self-training framework (See Fig. 3-(b)). The framework generates pseudo matting labels \\(\\widehat{\\alpha}\\) under the guidance of instance masks, and the pseudo labels supervise the matting model on natural images. To obtain high-quality pseudo labels, we adopt several design choices as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.567 + ], + "angle": 0, + "content": "Teacher-Student Framework. Motivated by the success in self-/semi-supervised learning [9, 42], a teacher network is introduced, which is a slowly advancing version of a student model via the exponential moving average (EMA). We utilize stable predictions of the teacher network as the pseudo labels \\(\\widehat{\\alpha}\\) and guide the student model during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.468, + 0.673 + ], + "angle": 0, + "content": "Weakly-augmented Input. The teacher network takes weakly-augmented samples to generate the pseudo labels. We only apply standard geometric augmentation, such as affine transform and random crop, for both image and mask guidance. Despite the noise in fine details, the ground-truth instance mask conveys a strong localization cue to the model, resulting in high-quality pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.749 + ], + "angle": 0, + "content": "For strongly-augmented inputs, the student model produces predictions \\(\\widehat{\\alpha}^{\\prime}\\), which are supervised to be the same as the pseudo labels \\(\\widehat{\\alpha}\\) obtained from the teacher model. To form the strongly-augmented version, we design two types of augmentations: image and guidance perturbations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Image Perturbations. Previous self-training-based methods [36, 37, 47] often employ region-level augmentation (e.g., Gaussian blur) to generate the strongly-augmented images. However, in the matting task, these augmentations are unsuitable since they interpolate the color value across the pixels and disturb each object's fractional contribution. This results in a mismatch of ground-truth alpha matte between the teacher and student inputs, lowering the effectiveness of self-training. We instead adopt linear pixel-level augmentations such as additive Gaussian noise and linear" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.093, + 0.561, + 0.105 + ], + "angle": 0, + "content": "contrast." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.241 + ], + "angle": 0, + "content": "Guidance Perturbations. We also modulate the reliability of mask guidance via dilation and/or erosion of the instance mask. Considering the diverse size of objects in natural scenes, we decide the kernel size based on the size of the object. Specifically, we set the kernel size as \\(\\beta \\%\\) of the shorter side of an object bounding box. The guidance perturbation is crucial to not only the success of self-training but also the robustness against common noise in mask guidance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.253, + 0.665, + 0.269 + ], + "angle": 0, + "content": "4.3. Unified Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.367 + ], + "angle": 0, + "content": "As shown in Fig. 3, under the proposed framework, the (student) mask-guided matting model can jointly learn from composited images with ground-truth labels and natural images with weak supervision. The final loss function is a combination of losses on both images as: \\( L_{final} = L_{composed} + L_{natural} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.488 + ], + "angle": 0, + "content": "For the composited images, we use the same loss function as MGMatting [50], which is the summation of \\( l_{1} \\) regression loss, composition loss [48], and Laplacian loss [13]. Thus, \\( L_{\\text{composed}} = L_{l1}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i) + L_{\\text{comp}}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i) + L_{\\text{lap}}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i) \\). To calculate the composition loss, we use ground-truth foreground color and alpha matte for non-target objects. For the natural images, we use \\( l_{1} \\) regression loss as follows: \\( L_{\\text{natural}} = L_{l1}(\\widehat{\\alpha}', \\widehat{\\alpha}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.502, + 0.631, + 0.518 + ], + "angle": 0, + "content": "5. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.663 + ], + "angle": 0, + "content": "There are several public matting benchmarks [19, 20, 32, 39, 48, 50] which provide high-quality alpha matte. However, none of them reflects the complexity of in-the-wild images and the consisting samples only contain a single object with simple backgrounds. Thus, we carefully design a new evaluation setup to facilitate an extensive evaluation of the mask-guided matting model in in-the-wild scenarios. The new setup consists of multiple benchmarks, and we describe the distinctive role of each benchmark below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.815 + ], + "angle": 0, + "content": "Composition-Wild. We first propose an in-the-wild extension of standard synthetic benchmark Composition-1k [48], namely Composition-Wild. We simulate the complex context of in-the-wild images by compositing multiple foreground objects onto a background image, similar to Eq.(2). The same foreground object data and background dataset (i.e., PASCAL VOC [6]) are utilized as Composition-1k. We report the performance over the forefront object from each sample unless otherwise specified. In addition, we design comprehensive metrics to allow in-depth analysis." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- \\(\\mathrm{SAD}_{FG}\\) reports the SAD (Sum of Absolute Difference) over the target object region where has non-zero ground-truth alpha matte. It indicates how well the prediction captures fine matting details of target objects (i.e., Target Detail Quality)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1996" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "- \\(\\mathrm{SAD}_{BG}\\) quantifies the wrong predictions on other objects or background regions by measuring the SAD over the non-target object region. It represents the capability of target object localization from the noisy mask guidance (i.e., Localization)." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.175, + 0.468, + 0.22 + ], + "angle": 0, + "content": "- \\(\\mathrm{SAD}_{OCC}\\) measures the SAD error of occluded objects. It indicates how well the model deals with occluded objects (i.e., Occlusion handling)." + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.468, + 0.405 + ], + "angle": 0, + "content": "AIM-500 [20] is a natural image matting benchmark that contains diverse categories of objects. The performance evaluation on the AIM benchmark denotes whether the model generalizes well to natural images rather than overfitting to the synthetic distribution of composited images. However, AIM is limited at capturing matting performance in complex contexts since they only contain images with simple backgrounds. In that regard, the evaluations on Composition-Wild and AIM complement each other and provide a holistic understanding of the mask-guided matting model in the wild." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.468, + 0.469 + ], + "angle": 0, + "content": "COCO [24] is a representative example of in-the-wild images. While we cannot access ground-truth alpha mattes on COCO, we conduct qualitative comparisons over diverse objects in their natural context." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.493, + 0.208, + 0.51 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.521, + 0.294, + 0.537 + ], + "angle": 0, + "content": "6.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.468, + 0.623 + ], + "angle": 0, + "content": "Architecture details. We build our learning framework with a state-of-the-art mask-guided matting network, PRN from MGMatting [50], which adopts a standard encoder-decoder structure [21,33] and progressively refines the output through the decoding process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.468, + 0.761 + ], + "angle": 0, + "content": "Training details. In practice, we adopt the two-stage training: pre-training on composited images and fine-tuning on both composited and natural images. This ensures the quality of pseudo labels at the beginning of the self-training, resulting in better performance. Following the previous works [21, 27, 30, 48], we use DIM [48] and COCO [24] as matting and background datasets, respectively. COCO dataset [24] is also utilized for a weak guidance dataset. We set the kernel ratio \\(\\beta\\) as 0.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Evaluation details. We report the widely used four matting metrics, SAD (Sum of Absolute Differences), MSE (Mean Squared Error), Grad (Gradient), and Conn (Connectivity), over the whole image, when the ground-truth alpha matte is available. To simulate the coarse mask guidance, we binarize the alpha matte with threshold 0.5 and then dilate the mask with kernel size \\(25 \\times 25\\), similar to [53]. We use a single trained model for all the benchmark evaluations and real-world applications." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.09, + 0.639, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.167, + 0.638, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.293, + 0.619, + 0.304 + ], + "angle": 0, + "content": "Image and Guidance" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.09, + 0.901, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.293, + 0.732, + 0.304 + ], + "angle": 0, + "content": "MGMatting" + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.293, + 0.846, + 0.303 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results on COCO dataset. The guidance are either manually drawn [24] (upper) or a prediction of the off-the-self model [10] (bottom). Best viewed zoomed in." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.369, + 0.603, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.474, + 0.576, + 0.485 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.369, + 0.688, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.474, + 0.682, + 0.485 + ], + "angle": 0, + "content": "(b) Guidance" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.369, + 0.801, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.474, + 0.788, + 0.485 + ], + "angle": 0, + "content": "(c) MGMatting*" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.369, + 0.901, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.474, + 0.872, + 0.484 + ], + "angle": 0, + "content": "(d) +IWL" + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.494, + 0.824, + 0.508 + ], + "angle": 0, + "content": "Figure 5. Effect of Instance-wise Learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.729, + 0.544 + ], + "angle": 0, + "content": "6.2. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Fig. 1 and Fig. 4 illustrate the qualitative comparisons on images in the COCO validation sets. We test the models under challenging scenarios such as occlusions, an unseen object in a matting dataset, complex surrounding context, and large interior errors on mask guidance. Compared with the previous mask-guided matting model [50], our model shows significantly better localization of a target object, as well as, produces much more accurate alpha mattes for challenging data. More qualitative comparisons can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.756, + 0.728 + ], + "angle": 0, + "content": "6.3. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.735, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Quantitative evaluations on Composition-Wild and AIM-500 datasets are shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Instance-wise Learning. We first analyze how instance-wise learning contributes to learn generalized matting capability from composited images. We set the two baseline models: (naive) MGMatting and MGMatting*. For the (naive) MGMatting, we reproduce the model with the same training recipes proposed in MGMatting [50]. In addition, we report the stronger baseline score of MGMatting* by excluding a heavy guidance augmentation, CutMask [50]. We empirically found that CutMask" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1997" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.233 + ], + "angle": 0, + "content": "
MethodComposition-WildAIM-500 [20]
SADMSEGradConn\\( SAD_{FG} \\)\\( SAD_{BG} \\)\\( SAD_{OCC} \\)SADMSEGradConn
Stage 1: Learning with Composited Images
MGMatting [50]583.360.1328139.8452.5465.51517.85842.8871.910.026823.3721.97
MGMatting* [50]389.150.0727141.8044.2155.74333.42524.9926.180.005615.8114.53
+Instance-wise (Ours)67.270.005842.3543.5249.5017.7759.5921.120.003816.8814.72
Stage 2: Learning with Composited and Weak-guidance Images
Segmentation-training69.240.006957.0846.5654.3614.8865.4220.940.004320.4614.20
Self-training (Ours)58.160.004639.0441.3747.3210.8453.0216.720.003014.6812.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.244, + 0.893, + 0.27 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on Composition-Wild and AIM-500 [20]. * denotes the improved version of MGMatting [50]. \"Instance-wise\" represents instance-wise learning. A lower score is better in all the metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.473, + 0.583 + ], + "angle": 0, + "content": "makes substantial random localization errors in guidance and lowers the reliability of mask guidance during training. Training with such guidance makes models ignore the mask guidance and overfit to synthetic samples, resulting in inferior in-the-wild performance. On top of MGMatting*, the \"Instance-wise Learning\" strategy drastically benefits the mask-guided matting model. Unlike the baseline model that segments all the composited objects (Fig. 5-(c)), the proposed strategy allows the model to well discriminate the target object (Fig. 5-(d)). Our model faithfully utilizes the coarse mask guidance and shows a huge gain in localization performance (see \\(\\mathrm{SAD}_{BG}\\) metric). Although the strategy assumes multi-object scenarios and learns the complex relations between them, it also contributes to the better matting details for an object (see \\(\\mathrm{SAD}_{FG}\\) in Composition-Wild and SAD in AIM). By design, the instance-wise learning naturally improves occlusion handling by learning the order of layered foreground objects and their fractional contributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.885 + ], + "angle": 0, + "content": "Effectiveness of Self-training. To extend the mask-guided matting model to the in-the-wild setting, we leverage the weak-supervision dataset under the proposed self-training framework. We compare our self-training framework to a baseline, namely \"segmentation-training\". This baseline is motivated by previous human-matting methods [23, 25]. It attempts to learn generalized matting capability via simultaneous segmentation training on natural images. To be specific, the model takes strongly augmented instance masks as guidance and is supervised to recover ground-truth instance masks through the additional segmentation head. As deteriorated performance indicates, the segmentation-training strategy fails to scale up to the in-the-wild setting. While it enhances the localization of target objects, supervision from binary instance masks hinders learning diverse matting patterns in the wild (see \\(\\mathrm{SAD}_{FG}\\) metric), especially for transparent objects. On the contrary, our self-training framework successfully generates the soft-pseudo label and utilizes it as a direct supervision to matting tasks, resulting in huge improvements in all the metrics." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Analysis on Perturbation in Self-training. To demon-" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.294, + 0.906, + 0.392 + ], + "angle": 0, + "content": "
MethodComposition-WildAIM-500
SADMSE\\( SAD_{FG} \\)\\( SAD_{BG} \\)SADMSE
Ours58.160.004647.3210.8416.720.0030
No Image.60.080.004845.1914.8918.570.0039
No Guide.60.140.004947.9512.1918.270.0035
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.403, + 0.878, + 0.418 + ], + "angle": 0, + "content": "Table 2. Ablation Study on Perturbations in Self-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.536 + ], + "angle": 0, + "content": "strate the effectiveness of perturbations in self-training, we ablate the image and guidance perturbations separately. The results are summarized in Table 2. As degraded performance denoted, encouraging robustness against image and guidance perturbations is crucial to the success of the self-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.893, + 0.645 + ], + "angle": 0, + "content": "We further investigate whether the matting-specific design of image perturbation is necessary. To this end, we train three variants with different configurations of image perturbations: (1) Pixel-level linear augmentation (Pixel), (2) Region-level augmentation (Region) and (3) Both of them (Pixel+Region). Below, we report the SAD metric for Composition-Wild and AIM-500, respectively." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.655, + 0.866, + 0.682 + ], + "angle": 0, + "content": "
PixelRegionPixel+Region
58.16 / 16.7269.8 / 25.9661.88 / 19.00
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.799 + ], + "angle": 0, + "content": "We can observe that the region-level augmentation, which is widely used in other tasks [37, 47], disturbs the effects of self-training objectives. As discussed in Sec. 4.2, this is because the region-level augmentation breaks the underlying opacity of objects and make a mismatch between inputs of the student and the teacher networks. On the contrary, pixel-level linear augmentation brings clear improvements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.729, + 0.832 + ], + "angle": 0, + "content": "7. Real-world Applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To show the generality and potential of the in-the-wild mask-guided matting model, we illustrate two new applications: mask-guided video matting and panoptic matting. Fig. 6 and Fig. 7 show examples of each application." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1998" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.258, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.222, + 0.21, + 0.232 + ], + "angle": 0, + "content": "(a) Frame" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.089, + 0.355, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.222, + 0.368, + 0.233 + ], + "angle": 0, + "content": "(b) Guidance" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.09, + 0.56, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.222, + 0.52, + 0.233 + ], + "angle": 0, + "content": "(c) MGMatting" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.091, + 0.707, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.222, + 0.654, + 0.233 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.09, + 0.857, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.222, + 0.826, + 0.233 + ], + "angle": 0, + "content": "(e) Editing Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.239, + 0.776, + 0.255 + ], + "angle": 0, + "content": "Figure 6. Extension to Mask-guided Video Matting. We use propagated masks [31] as guidance." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.261, + 0.408, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.261, + 0.558, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.261, + 0.706, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.262, + 0.857, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.339, + 0.26, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.45, + 0.211, + 0.46 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.339, + 0.411, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.45, + 0.368, + 0.46 + ], + "angle": 0, + "content": "(b) Guidance" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.339, + 0.557, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.45, + 0.522, + 0.461 + ], + "angle": 0, + "content": "(c) MGMatting" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.339, + 0.706, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.45, + 0.654, + 0.46 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.339, + 0.857, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.45, + 0.826, + 0.46 + ], + "angle": 0, + "content": "(e) Editing Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.469, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Figure 7. Extension to Panoptic Matting. Hand drawn panoptic masks [24] are utilized as guidance. We illustrate independent regions with different colors. The black color represents ignored regions of labels or predictions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.516, + 0.33, + 0.533 + ], + "angle": 0, + "content": "7.1. Mask-guided Video Matting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.77 + ], + "angle": 0, + "content": "Current video matting solutions [40, 53] require multiple frames of trimap to produce an alpha matte for a video. Therefore, required user interaction is more burdensome in the video domain. With the mask-guided matting model, we explore a more practical scenario where only a single instance mask is given in the first frame. In this challenging setting, we first propagate the given mask to the rest of the video frames using the video object segmentation model [31]. Then, the mask-guided matting model utilizes the propagated mask as guidance and produces alpha matte in a frame-by-frame manner. Our model not only corrects the noise in propagated masks but also predicts reasonable opacity in challenging cases such as motion blur. With the prediction, we also provide visually appealing video editing results (i.e., background replacement)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.248, + 0.801 + ], + "angle": 0, + "content": "7.2. Panoptic Matting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Panoptic segmentation aims to parse an image into nonoverlapping regions of stuff and things. By the definition of the task format, the soft transition between the regions is overlooked. We explore whether our mask-guided matting model could provide such understanding from panoptic segmentation results. Given a panoptic mask, our model con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.654 + ], + "angle": 0, + "content": "siders the segment of each region separately and produces the corresponding alpha matte. We form panoptic matting results by aggregating the predictions. Unlike the baseline, our model successfully refines the fine details and captures the soft transition between the regions. Interestingly, while the model is trained on objects (i.e., things), our method is well generalized on the stuff regions (See sky region in Fig. 7). Panoptic matting enables us to do region-level image editing and the qualitative results are shown in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.618, + 0.68 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.856 + ], + "angle": 0, + "content": "In this paper, we propose a simple yet effective learning framework for mask-guided matting in the wild. We first introduce instance-wise learning to learn generalized matting ability from the composited images. Then, the model is further extended to handle diverse objects under the self-training on weak-guidance images. Extensive experiments on the newly proposed evaluation benchmark provide an in-depth understanding of the unique challenges. Using our approach, we instantiate new practical applications: mask-guided video matting and panoptic matting. We hope our proposals inspire future work on mask-guided matting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work was partially supported by the NRF (NRF-2020M3H8A1115028, FY2021) and Samsung Electronics Co., Ltd (G01200447)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1999" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Tamara L Berg and Alexander C Berg. Finding iconic images. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, pages 1-8. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.175, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[2] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[3] Quan Chen, Tiezheng Ge, Yanyu Xu, Zhiqiang Zhang, Xinxin Yang, and Kun Gai. Semantic human matting. In Proceedings of the 26th ACM international conference on Multimedia, pages 618-626, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[4] Qifeng Chen, Dingzeyu Li, and Chi-Keung Tang. Knn matting. IEEE transactions on pattern analysis and machine intelligence, 35(9):2175-2188, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.471, + 0.419 + ], + "angle": 0, + "content": "[5] Yung-Yu Chuang, Brian Curless, David H Salesin, and Richard Szeliski. A bayesian approach to digital matting. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 2, pages II-II. IEEE, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.421, + 0.471, + 0.477 + ], + "angle": 0, + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.479, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[7] Marco Forte and François Pitie. \\(f, b\\), alpha matting. arXiv preprint arXiv:2003.07711, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[8] Eduardo SL Gastal and Manuel M Oliveira. Shared sampling for real-time alpha matting. In Computer Graphics Forum, volume 29, pages 575-584. Wiley Online Library, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.554, + 0.471, + 0.623 + ], + "angle": 0, + "content": "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.626, + 0.47, + 0.68 + ], + "angle": 0, + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.684, + 0.47, + 0.727 + ], + "angle": 0, + "content": "[11] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaouu Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[12] Kaiming He, Jian Sun, and Xiaou Tang. Fast matting using large kernel matting laplacian matrices. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2165-2172. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[13] Qiqi Hou and Feng Liu. Context-aware image matting for simultaneous foreground and alpha estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4130-4139, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Ronghang Hu, Piotr Dólar, Kaiming He, Trevor Darrell, and Ross Girshick. Learning to segment every thing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4233-4241, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[15] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson W.H. Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In AAAI, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[16] Philip Lee and Ying Wu. Nonlocal matting. In CVPR 2011, pages 2193-2200. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[17] Anat Levin, Dani Lischinski, and Yair Weiss. A closed-form solution to natural image matting. IEEE transactions on pattern analysis and machine intelligence, 30(2):228-242, 2007. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[18] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[19] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[20] Jizhizi Li, Jing Zhang, and Dacheng Tao. Deep automatic natural image matting. arXiv preprint arXiv:2107.07235, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[21] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11450-11457, 2020. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.415, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[22] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[23] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[25] Jinlin Liu, Yuan Yao, Wendi Hou, Miaomiao Cui, Xuansong Xie, Changshui Zhang, and Xian-sheng Hua. Boosting semantic human matting with coarse annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8572, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[26] Yuhao Liu, Jiake Xie, Xiao Shi, Yu Qiao, Yujie Huang, Yong Tang, and Xin Yang. Tripartite information mining and integration for image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7555-7564, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[27] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3266-3275, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[28] Sebastian Lutz, Konstantinos Amplanitis, and Aljosa Smolic. Alphagan: Generative adversarial networks for natural image matting. arXiv preprint arXiv:1807.10088, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Stephen Palmer. Canonical perspective and the perception of objects. Attention and performance, pages 135-151, 1981. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2000" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11696-11706, 2022. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[31] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Per-clip video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1352–1361, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[32] Yu Qiao, Yuhao Liu, Xin Yang, Dongsheng Zhou, Mingliang Xu, Qiang Zhang, and Xiaopeng Wei. Attention-guided hierarchical structure aggregation for image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13676-13685, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.441 + ], + "angle": 0, + "content": "[34] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[35] Ehsan Shahrian, Deepu Rajan, Brian Price, and Scott Cohen. Improving image matting using comprehensive sampling sets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 636-643, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[36] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[37] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.692 + ], + "angle": 0, + "content": "[38] Jian Sun, Jiaya Jia, Chi-Keung Tang, and Heung-Yeung Shum. Poisson matting. In ACM SIGGRAPH 2004 Papers, pages 315–321. 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[39] Yanan Sun, Chi-Keung Tang, and Yu-Wing Tai. Semantic image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11120–11129, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.804 + ], + "angle": 0, + "content": "[40] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[41] Jingwei Tang, Yagiz Aksoy, Cengiz Oztireli, Markus Gross, and Tunc Ozan Aydin. Learning-based sampling for natural image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3055–3063, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[42] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[43] Jue Wang and Michael F Cohen. An iterative optimization approach for unified image segmentation and matting. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 936-943. IEEE, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[44] Jue Wang and Michael F Cohen. Optimized color sampling for robust matting. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[45] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Hanqing Zhao, Weiming Zhang, and Nenghai Yu. Improved image matting via real-time user clicks and uncertainty estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15374-15383, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[46] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXV, pages 238–258. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[47] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.473, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[48] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[49] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[50] Qihang Yu, Jianming Zhang, He Zhang, Yilin Wang, Zhe Lin, Ning Xu, Yutong Bai, and Alan Yuille. Mask guided matting via progressive refinement network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1154-1163, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[51] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.741, + 0.892, + 0.811 + ], + "angle": 0, + "content": "[52] Yunke Zhang, Lixue Gong, Lubin Fan, Peiran Ren, Qixing Huang, Hujun Bao, and Weiwei Xu. A late fusion cnn for digital matting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7469-7478, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.812, + 0.892, + 0.882 + ], + "angle": 0, + "content": "[53] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2001" + } + ] +] \ No newline at end of file diff --git a/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_origin.pdf b/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..058e8764f0cd7266e5601bfe46909cb535394a15 --- /dev/null +++ b/2023/Mask-Guided Matting in the Wild/09ad983f-3200-4b62-8566-4c22c8b72f84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b83e8183b0596bd9e2e8e7243a9d9920dacd0e4027f85a5e2571cf05c19a75f2 +size 3202062 diff --git a/2023/Mask-Guided Matting in the Wild/full.md b/2023/Mask-Guided Matting in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7842169aabf3e0abe18ddf6eb61d4f08078a5df2 --- /dev/null +++ b/2023/Mask-Guided Matting in the Wild/full.md @@ -0,0 +1,327 @@ +# Mask-guided Matting in the Wild + +Kwanyong Park $^{1}$ Sanghyun Woo $^{1}$ Seoung Wug Oh $^{2}$ In So Kweon $^{1}$ Joon-Young Lee $^{2}$ $^{1}$ KAIST $^{2}$ Adobe Research + +# Abstract + +Mask-guided matting has shown great practicality compared to traditional trimap-based methods. The mask-guided approach takes an easily-obtainable coarse mask as guidance and produces an accurate alpha matte. To extend the success toward practical usage, we tackle mask-guided matting in the wild, which covers a wide range of categories in their complex context robustly. To this end, we propose a simple yet effective learning framework based on two core insights: 1) learning a generalized matting model that can better understand the given mask guidance and 2) leveraging weak supervision datasets (e.g., instance segmentation dataset) to alleviate the limited diversity and scale of existing matting datasets. Extensive experimental results on multiple benchmarks, consisting of a newly proposed synthetic benchmark (Composition-Wild) and existing natural datasets, demonstrate the superiority of the proposed method. Moreover, we provide appealing results on new practical applications (e.g., panoptic matting and mask-guided video matting), showing the great generality and potential of our model. + +# 1. Introduction + +Image matting aims to predict the opacity of objects, which enables precise separation from surrounding backgrounds. Due to the ill-posed nature of the task, many works [7, 13, 21, 27, 30, 48] have improved matting performance by relying on the manual guidance of a trimap. However, pixel-level annotation of foreground/background/unknown is extremely burdensome, restricting its usage in many practical applications such as image/video editing and film production. Recently, many efficient alternatives for user guidance have been proposed, including trimap-free [15, 32], additional background images [22, 34], scribble [43], and the user clicks [45]. + +Among them, the mask-guided approach [50] shows a great trade-off between performance and intensity of user interaction. It utilizes a coarse mask as guidance, which is much easier to obtain either manually or from off-the-shelf + +![](images/3c5b876cfad515532688621335ac06fec5c6ded6040dfb44f7c80a11e3e8d53d.jpg) +Figure 1. Qualitative Comparisons of MGMatting [50] and Ours in the wild. The mask guidance is overlaid on images with blue color. Best viewed zoomed in. + +segmentation models [2, 10]. With only the coarse spatial prior, the mask-guided matting model [50] shows comparable or even better performance than the trimap-based competitors [13, 17, 21, 27, 48] on synthetic Composition-1k [48] and a real-world human matting dataset [50]. However, despite the encouraging results, we see the previous state-of-the-art model [50] struggles to obtain desirable alpha matte in complex real-world scenes (see Fig. 1). + +With this observation, we tackle mask-guided matting in the wild. Specifically, we formulate unique setups and emerging challenges of the new task as follows: (1) We aim to handle objects in their complex context, reflecting the characteristics of natural images. The previous method [50] evaluates their model on iconic-object images [1,29] where only a single object is in the center. As the model can easily find the target object in such images, the model's real instance discrimination ability is, in fact, veiled. On the contrary, in an 'in-the-wild' setting, it is crucial to precisely localize the target object from the given coarse/noisy mask guidance (i.e. mask awareness). (2) Our model targets to deal with diverse categories of objects in natural + +images. Unlike most previous methods that improve generalization performance at the expense of category-specific regime (e.g., limiting to humans [15] or animals [19]), we aim to understand distinctive matting patterns of vast categories. (3) Limited data problem makes the new setting more complicated. Due to the labeling complexity, annotating alpha matte for objects in common scenes, e.g., the COCO dataset [24], is infeasible. As a sidestep, previous benchmarks [32, 48] extract the alpha matte and foreground colors from images with simple backgrounds. These are composited on various backgrounds [6, 24], and resulting samples are used to train and evaluate matting models. However, due to the inevitable composition artifacts, the models usually show limited generalization performance. In that sense, how to train and evaluate the in-the-wild matting model remains an open question. + +Toward this goal, we propose a simple yet effective learning framework for a generalized mask-guided matting model. First, we investigate fundamental reasons for the poor generalization of the previous mask-guided matting model [50] and find that this is mainly from the training data generation process. Specifically, the previous composition process includes instance merging data augmentation, which merges several foreground objects into a single object. While this augmentation is effective in the trimap-based methods [21, 30, 41], it implicitly makes a negative bias for the mask-guided matting model to ignore the guidance. Thus, the model struggles to localize the target objects in complex natural scenes. We alleviate the bias by proposing an instance-wise learning objective, where the model is supervised to segment one of multiple instances according to the guidance. By doing so, the model learns strong semantic representation regarding complex relations and soft transitions between objects. Despite the simplicity of the proposal, this greatly improves performance in the wild. + +Second, we explore a practical solution to make the mask-guided model handle various categories of objects robustly. Instead of scaling the matting dataset, we leverage a dataset with weak supervision [14, 46] (i.e., instance segmentation dataset [24]), as the coarse instance masks are easier to obtain over the diverse categories of objects. To effectively hallucinate the fine supervision signal with the weak localization guidance, we come up with a self-training framework [36, 37, 47]. Specifically, a pseudo label is generated based on a weakly-augmented input (both image and instance mask annotation as guidance), which supervises the model prediction on a strongly augmented version of them. During self-training, the model is not only adapting to the in-the-wild scenario in a self-evolving manner but also being robust to noise in both image and guidance. + +To verify the in-the-wild performance of mask-guided matting, we formally define an evaluation protocol involving multiple sub-benchmarks: Composition-Wild, AIM- + +500 [20], COCO [24]. We first design an in-the-wild extension of the popular synthetic Composition-1k benchmark [48], namely Composition-Wild. We simulate complex real-world images by compositing multiple foreground objects. To bring valuable insight on the failure cases of the model, we design sub-metrics for Composition-Wild. In addition, we use the AIM-500 dataset to establish quantitative results on natural images (i.e., with no composition artifacts), although most images are iconic-object images with simple backgrounds. Finally, we provide qualitative outputs of our mask-guided matting model on the COCO dataset [24] which is one of the most representative in-the-wild datasets. + +To summarize, we make the following main contributions. 1) To our best knowledge, it is the first work to explore mask-guided matting in the wild. 2) We develop a simple yet effective learning framework leveraging both composited and weak-guidance images. 3) We design an evaluation setup for the new task. 4) We initiate several interesting extensions: video and panoptic matting. + +# 2. Related Works + +Natural Image Matting. Most image matting methods require a trimap as additional input, which conveys pixel-level annotations of foreground, background, and unknown region. Traditional image matting methods can be categorized into two groups. Sampling-based methods [5, 8, 11, 35, 44] estimate the alpha matte of the unknown region based on sampled colors of foreground and background. Propagation-based methods [4, 12, 16-18, 38] propagate the neighboring known alpha values to unknown regions according to the affinity between pixels. Both methods primarily rely on color or low-level features, showing limitations in complicated scenarios. + +To tackle these issues, many deep-learning based mating methods have been proposed. DIM [48] is a representative work which proposes a convolutional encoder-decoder matting network as well as a large-scale synthetic dataset (Composition-1k) to train the model. Many follow-up works have made tremendous improvements in diverse aspects, such as sophisticated loss design [7, 13, 28] and architectural advances by introducing attention mechanisms [21, 26] or transformer architecture [30]. + +There is a large volume of trials to relax the heavy user-supplied constraints. Some methods [32, 52] attempt to get rid of trimap. However, they show inferior performance and cannot be generalized to unseen objects in the real world. Sengupta et al. [34] utilize additional background images along with other lightweight priors (e.g., segmentation mask) to perform the matting task. Wei et al. [45] introduce user click interaction to effectively eliminate the ambiguity of target object. Recently, MGMatting [50] proposes a mask-guided matting framework, where only the + +easily obtainable coarse mask is needed as guidance. + +In this paper, we further study mask-guided matting in an in-the-wild setting, inspired by the generality and practicality of the framework. Different from MGMatting, we target diverse categories of objects in their complex background, enabling real-world applications. + +Class-specific Matting is a special type of matting that limits target objects to one or few classes such as humans [3] or animals [19]. BSHM [25] leverages the coarse segmentation mask data to build a generalized human matting network. MODNet [15] presents a lightweight network architecture with decomposed multi-scale network designs. GFM [19] first proposes the task of animal matting and a real-world animal matting dataset. In general, known semantics effectively ease the difficulty of matting, and these methods show great generality without the necessity of additional guidance input. However, the specialized methods and trained model cannot be generalized to the in-the-wild setting since limited semantics do not cover diverse patterns of visual worlds. + +Apart from previous efforts, we attempt to extend the short category regime of matting networks. For this purpose, we propose an efficient solution by leveraging the instance segmentation dataset [24], which covers diverse categories of objects with weak localization annotations. Given the weak guidance dataset, we study the challenges in developing an in-the-wild mask-guided matting model. + +# 3. Mask-guided Matting in the Wild + +In natural scenes, objects are rarely in isolation: diverse categories of entities interact with each other. We aim to robustly separate a target object from the surroundings in such scenes. Specifically, we formulate the problem in the form of mask-guided matting, namely mask-guided matting in the wild. Our matting model takes an easily-obtainable coarse mask as guidance, either manually drawn or predicted from the off-the-shelf segmentation models. With in-the-wild data, the model needs to understand semantics so that it can precisely localize the target object from the given noisy masks. In the meantime, the model also needs to have strong low-level feature representation so that it can capture fine details and accurate opacity of the diverse target matte. + +To concrete the new challenges toward such an in-the-wild setting, we first study how the previous model [50] performs. We derive predictions with different mask guidance and summarize the results on Fig. 2. As shown in the upper example, in the synthetic Composition-1k [48] example, the previous model perfectly separates the foreground object regardless of the given guidance. Surprisingly, it produces an alpha matte of the foreground object when the mask of the background region is given as guidance. On the contrary, the model lacks generalization on + +![](images/2b76c8772c71bce62191100fc3b91dd8c868b9db95412ce38c0475a5b25abb0f.jpg) +Image and Guidance + +![](images/eacb1b650023f83cdb64dc8a8df0a693bd1c1a506c95be8140057038e68bd50e.jpg) +With $1^{\mathrm{st}}$ Mask Guidance + +![](images/be7855bebc8978d1a1b719e61b8db458d1e916b662cc9bad1872c4c8c3267b1c.jpg) +With $2^{\text{nd}}$ Mask Guidance +Figure 2. Predictions of MGMatting [50] with different guidance. (upper) synthetic Composition-1k image [48]. (bottom) Complex real world image [49]. Best viewed zoomed in. + +complex real-world scenes (see bottom example). In particular, the model fails to localize the target object and makes wrong predictions on regions of similar color distributions (e.g., shadows) or another nearby object. From the two representative examples, we make several observations: 1) The previous method yields a model overfitted to the (trained) synthetic matting dataset, which memorizes the limited patterns of foreground objects rather than semantically understanding the mask guidance for localizing and segmenting the objects; 2) Such a model fails to precisely localize and segment a target object in the wild; 3) Previously used synthetic benchmark, Composition-1k [48], neither represents the complex nature of the in-the-wild setting nor systematically blocks the shortcut of the overfitting issue. Thus the Composition-1k may not be a suitable benchmark for mask-guided matting in the wild. + +To this end, we explore how to train a generalized mask-guided matting model given the limited scale and diversity of the matting dataset (Sec. 4), as well as revisit the evaluation benchmark to test the model (Sec. 5). + +# 4. Proposed Method + +The overview of the proposed training framework is illustrated in Fig. 3. Our framework involves three types of datasets: 1) matting dataset that includes accurate alpha matte and corresponding foreground colors; 2) background dataset, where the foreground objects are composited to create training samples; 3) weak supervision dataset that is introduced to extend the matting model to the in-the-wild setting. The main goal of our framework is to learn accurate matting ability from the synthetic composited images and generalize the knowledge to the in-the-wild domain. To this end, we first investigate how the model could learn generalized knowledge from the composited images (Sec. 4.1), then how the model is further improved with weak supervision data (Sec. 4.2). + +![](images/d8690695c1e705c06e4da75d42df126bfe25625e99137ab44b9a87a179b5a0d8.jpg) + +![](images/b37ab59244c182cbd48597bd63b1400e7de8680145ae377aeaaf83ad18bc3498.jpg) +Figure 3. The Overview of the Proposed Framework. (a) Learning from composited images with instance-wise learning (Sec. 4.1). (b) Learning from weak-supervision images with self-training (Sec. 4.2). + +# 4.1. Learning from Composed Images + +As discussed in the Sec. 3, the previous model [50] shows poor generalization performance in the challenging real-world setting. Specifically, it often fails to identify the target object in a given guidance and produces a matting result for other objects. We found that the problem, the lack of mask awareness, comes from the previous data pipeline to generate training samples. + +For a clear understanding, we first briefly summarize the previous training data generation step proposed in MGMatting [50]. Here, for clarity, generic data augmentation such as random-crop, resize, and affine-transform are omitted. A foreground object (consisting of foreground color $F$ and alpha matte $\alpha$ ) and a background image $B$ are randomly sampled from the matting and background datasets, respectively. These are composited to form a training image $I$ following the composition formulation as follows: + +$$ +I = \operatorname {C o m p o s i t e} (F, \alpha , B) = \alpha F + (1 - \alpha) B. \tag {1} +$$ + +To simulate the mask guidance $M$ , they first binarize the alpha matte with a threshold randomly sampled from 0 to 1, then apply dilation and/or erosion to it. In addition, they propose a stronger mask augmentation, CutMask [50], originally designed for robustness to noise in mask guidance. Like CutMix [51], a random-sized patch is selected, and the content is pasted on a random position. + +In addition, MGMatting [50] adopted instance merging augmentation as most state-of-the-art trimap-based matting methods [21, 30, 41] did. Specifically, they select two random foreground objects, merge them into a single object, + +and composite it on a random background image. The instance merging augmentation effectively alleviates the limited scale and diversity of foreground objects in the matting dataset. However, in the context of mask-guided matting, we found that this augmentation hinders instance-wise understanding and makes a negative bias that the model eventually segments all the existing composited objects regardless of the given mask guidance. The different effects of the same augmentation are originated from the uncertainty of mask guidance. While the trimap brings pixel-level accurate localization of foreground/background/unknown regions, the mask guidance naturally conveys noisy information, and it is hard to localize them accurately. Under such uncertainty, the model tends to find the shortcut, segmenting all the composited objects rather than learning to utilize the uncertain mask guidance. + +Instance-wise Learning. We systematically block the above shortcut and enforce the model to learn complex relations and soft transitions between objects. The idea is simple: supervise the model to segment one of them according to the guidance (See Fig. 3-(a)). Specifically, given the two foreground objects (colors $F_{1}, F_{2}$ and opacity $\alpha_{1}, \alpha_{2}$ ), we sequentially composite these foreground layers (the first object is on top.) on background image $B$ as follows: + +$$ +\begin{array}{l} I = \text {C o m p o s i t e} \left(F _ {1}, \alpha_ {1}, \text {C o m p o s i t e} \left(F _ {2}, \alpha_ {2}, B\right)\right) \tag {2} \\ = \alpha_ {1} F _ {1} + \alpha_ {2} (1 - \alpha_ {1}) F _ {2} + (1 - \alpha_ {1}) (1 - \alpha_ {2}) B. \\ \end{array} +$$ + +Based on the above equation, we re-calibrate the fractional contribution of each foreground object to images at pixel level as follows: $\overline{\alpha}_1 = \alpha_1, \overline{\alpha}_2 = \alpha_2 (1 - \alpha_1)$ . We randomly + +select an index $i_{rand} \in \{1,2\}$ of a target object. With the guidance $M(\overline{\alpha}_i)$ , the model predicts the opacity of the target object $\widetilde{\alpha}_i$ , which is supervised by corresponding ground-truth alpha matte $\overline{\alpha}_i$ . Such instance-wise training makes the model learn fractional differentiation between objects according to the given coarse mask guidance. + +# 4.2. Learning from Images with Weak Supervision + +Due to the annotation difficulties, the labels of the alpha matte are practically infeasible to collect for large quantities and diverse categories. By comparison, coarse mask annotations are more abundant and accessible, thus extending the dataset scale and taxonomies is easier to achieve. Motivated by this, we leverage the weak supervision dataset (i.e., instance segmentation [24]), and investigate how these new data benefit the generalization of the mask-guided matting model. The main challenge lies in how to get matting supervision because the instance mask labels are noisy in fine details and do not include opacity annotations. + +To effectively tackle this challenge, we design a self-training framework (See Fig. 3-(b)). The framework generates pseudo matting labels $\widehat{\alpha}$ under the guidance of instance masks, and the pseudo labels supervise the matting model on natural images. To obtain high-quality pseudo labels, we adopt several design choices as follows. + +Teacher-Student Framework. Motivated by the success in self-/semi-supervised learning [9, 42], a teacher network is introduced, which is a slowly advancing version of a student model via the exponential moving average (EMA). We utilize stable predictions of the teacher network as the pseudo labels $\widehat{\alpha}$ and guide the student model during training. + +Weakly-augmented Input. The teacher network takes weakly-augmented samples to generate the pseudo labels. We only apply standard geometric augmentation, such as affine transform and random crop, for both image and mask guidance. Despite the noise in fine details, the ground-truth instance mask conveys a strong localization cue to the model, resulting in high-quality pseudo labels. + +For strongly-augmented inputs, the student model produces predictions $\widehat{\alpha}^{\prime}$ , which are supervised to be the same as the pseudo labels $\widehat{\alpha}$ obtained from the teacher model. To form the strongly-augmented version, we design two types of augmentations: image and guidance perturbations. + +Image Perturbations. Previous self-training-based methods [36, 37, 47] often employ region-level augmentation (e.g., Gaussian blur) to generate the strongly-augmented images. However, in the matting task, these augmentations are unsuitable since they interpolate the color value across the pixels and disturb each object's fractional contribution. This results in a mismatch of ground-truth alpha matte between the teacher and student inputs, lowering the effectiveness of self-training. We instead adopt linear pixel-level augmentations such as additive Gaussian noise and linear + +contrast. + +Guidance Perturbations. We also modulate the reliability of mask guidance via dilation and/or erosion of the instance mask. Considering the diverse size of objects in natural scenes, we decide the kernel size based on the size of the object. Specifically, we set the kernel size as $\beta \%$ of the shorter side of an object bounding box. The guidance perturbation is crucial to not only the success of self-training but also the robustness against common noise in mask guidance. + +# 4.3. Unified Training + +As shown in Fig. 3, under the proposed framework, the (student) mask-guided matting model can jointly learn from composited images with ground-truth labels and natural images with weak supervision. The final loss function is a combination of losses on both images as: $L_{final} = L_{composed} + L_{natural}$ . + +For the composited images, we use the same loss function as MGMatting [50], which is the summation of $l_{1}$ regression loss, composition loss [48], and Laplacian loss [13]. Thus, $L_{\text{composed}} = L_{l1}(\widetilde{\alpha}_i, \overline{\alpha}_i) + L_{\text{comp}}(\widetilde{\alpha}_i, \overline{\alpha}_i) + L_{\text{lap}}(\widetilde{\alpha}_i, \overline{\alpha}_i)$ . To calculate the composition loss, we use ground-truth foreground color and alpha matte for non-target objects. For the natural images, we use $l_{1}$ regression loss as follows: $L_{\text{natural}} = L_{l1}(\widehat{\alpha}', \widehat{\alpha})$ . + +# 5. Benchmarks + +There are several public matting benchmarks [19, 20, 32, 39, 48, 50] which provide high-quality alpha matte. However, none of them reflects the complexity of in-the-wild images and the consisting samples only contain a single object with simple backgrounds. Thus, we carefully design a new evaluation setup to facilitate an extensive evaluation of the mask-guided matting model in in-the-wild scenarios. The new setup consists of multiple benchmarks, and we describe the distinctive role of each benchmark below. + +Composition-Wild. We first propose an in-the-wild extension of standard synthetic benchmark Composition-1k [48], namely Composition-Wild. We simulate the complex context of in-the-wild images by compositing multiple foreground objects onto a background image, similar to Eq.(2). The same foreground object data and background dataset (i.e., PASCAL VOC [6]) are utilized as Composition-1k. We report the performance over the forefront object from each sample unless otherwise specified. In addition, we design comprehensive metrics to allow in-depth analysis. + +- $\mathrm{SAD}_{FG}$ reports the SAD (Sum of Absolute Difference) over the target object region where has non-zero ground-truth alpha matte. It indicates how well the prediction captures fine matting details of target objects (i.e., Target Detail Quality). + +- $\mathrm{SAD}_{BG}$ quantifies the wrong predictions on other objects or background regions by measuring the SAD over the non-target object region. It represents the capability of target object localization from the noisy mask guidance (i.e., Localization). +- $\mathrm{SAD}_{OCC}$ measures the SAD error of occluded objects. It indicates how well the model deals with occluded objects (i.e., Occlusion handling). + +AIM-500 [20] is a natural image matting benchmark that contains diverse categories of objects. The performance evaluation on the AIM benchmark denotes whether the model generalizes well to natural images rather than overfitting to the synthetic distribution of composited images. However, AIM is limited at capturing matting performance in complex contexts since they only contain images with simple backgrounds. In that regard, the evaluations on Composition-Wild and AIM complement each other and provide a holistic understanding of the mask-guided matting model in the wild. + +COCO [24] is a representative example of in-the-wild images. While we cannot access ground-truth alpha mattes on COCO, we conduct qualitative comparisons over diverse objects in their natural context. + +# 6. Experiments + +# 6.1. Implementation Details + +Architecture details. We build our learning framework with a state-of-the-art mask-guided matting network, PRN from MGMatting [50], which adopts a standard encoder-decoder structure [21,33] and progressively refines the output through the decoding process. + +Training details. In practice, we adopt the two-stage training: pre-training on composited images and fine-tuning on both composited and natural images. This ensures the quality of pseudo labels at the beginning of the self-training, resulting in better performance. Following the previous works [21, 27, 30, 48], we use DIM [48] and COCO [24] as matting and background datasets, respectively. COCO dataset [24] is also utilized for a weak guidance dataset. We set the kernel ratio $\beta$ as 0.2. + +Evaluation details. We report the widely used four matting metrics, SAD (Sum of Absolute Differences), MSE (Mean Squared Error), Grad (Gradient), and Conn (Connectivity), over the whole image, when the ground-truth alpha matte is available. To simulate the coarse mask guidance, we binarize the alpha matte with threshold 0.5 and then dilate the mask with kernel size $25 \times 25$ , similar to [53]. We use a single trained model for all the benchmark evaluations and real-world applications. + +![](images/85494dbe39f0d64f6f98bdd7e733b9d89712e274bcaf7d2f5e5361e3741c7514.jpg) + +![](images/9d5f460f9a0447df5882b22c361b98d93a02b47362a37f2540e81c8b45b25518.jpg) +Image and Guidance + +![](images/991b689f261faabd25116c9e2e44ce0fa6a62addcc71db2dea685d71754872d8.jpg) +MGMatting +Ours + +![](images/7bdae739ce83e3eb04d28609ee72e38147012b5fa5c015c34fcc4e6a8afe6e39.jpg) +Figure 4. Qualitative Results on COCO dataset. The guidance are either manually drawn [24] (upper) or a prediction of the off-the-self model [10] (bottom). Best viewed zoomed in. +(a) Image +Figure 5. Effect of Instance-wise Learning. + +![](images/850a5a2772428642e7211cd0982d543f571e6268834e5a050991c7f31a267384.jpg) +(b) Guidance + +![](images/330801958f40f89e03e7c8e609fa6d87da55a0556c07559395908fdbe7dfafaf.jpg) +(c) MGMatting* + +![](images/e70c3092367d9616a17feb8b0e86241f38936383df0af1c909b76a7b3fbc9f34.jpg) +(d) +IWL + +# 6.2. Qualitative Comparisons + +Fig. 1 and Fig. 4 illustrate the qualitative comparisons on images in the COCO validation sets. We test the models under challenging scenarios such as occlusions, an unseen object in a matting dataset, complex surrounding context, and large interior errors on mask guidance. Compared with the previous mask-guided matting model [50], our model shows significantly better localization of a target object, as well as, produces much more accurate alpha mattes for challenging data. More qualitative comparisons can be found in the supplementary material. + +# 6.3. Ablation Study and Analysis + +Quantitative evaluations on Composition-Wild and AIM-500 datasets are shown in Table 1. + +Effectiveness of Instance-wise Learning. We first analyze how instance-wise learning contributes to learn generalized matting capability from composited images. We set the two baseline models: (naive) MGMatting and MGMatting*. For the (naive) MGMatting, we reproduce the model with the same training recipes proposed in MGMatting [50]. In addition, we report the stronger baseline score of MGMatting* by excluding a heavy guidance augmentation, CutMask [50]. We empirically found that CutMask + +
MethodComposition-WildAIM-500 [20]
SADMSEGradConn\( SAD_{FG} \)\( SAD_{BG} \)\( SAD_{OCC} \)SADMSEGradConn
Stage 1: Learning with Composited Images
MGMatting [50]583.360.1328139.8452.5465.51517.85842.8871.910.026823.3721.97
MGMatting* [50]389.150.0727141.8044.2155.74333.42524.9926.180.005615.8114.53
+Instance-wise (Ours)67.270.005842.3543.5249.5017.7759.5921.120.003816.8814.72
Stage 2: Learning with Composited and Weak-guidance Images
Segmentation-training69.240.006957.0846.5654.3614.8865.4220.940.004320.4614.20
Self-training (Ours)58.160.004639.0441.3747.3210.8453.0216.720.003014.6812.02
+ +makes substantial random localization errors in guidance and lowers the reliability of mask guidance during training. Training with such guidance makes models ignore the mask guidance and overfit to synthetic samples, resulting in inferior in-the-wild performance. On top of MGMatting*, the "Instance-wise Learning" strategy drastically benefits the mask-guided matting model. Unlike the baseline model that segments all the composited objects (Fig. 5-(c)), the proposed strategy allows the model to well discriminate the target object (Fig. 5-(d)). Our model faithfully utilizes the coarse mask guidance and shows a huge gain in localization performance (see $\mathrm{SAD}_{BG}$ metric). Although the strategy assumes multi-object scenarios and learns the complex relations between them, it also contributes to the better matting details for an object (see $\mathrm{SAD}_{FG}$ in Composition-Wild and SAD in AIM). By design, the instance-wise learning naturally improves occlusion handling by learning the order of layered foreground objects and their fractional contributions. + +Effectiveness of Self-training. To extend the mask-guided matting model to the in-the-wild setting, we leverage the weak-supervision dataset under the proposed self-training framework. We compare our self-training framework to a baseline, namely "segmentation-training". This baseline is motivated by previous human-matting methods [23, 25]. It attempts to learn generalized matting capability via simultaneous segmentation training on natural images. To be specific, the model takes strongly augmented instance masks as guidance and is supervised to recover ground-truth instance masks through the additional segmentation head. As deteriorated performance indicates, the segmentation-training strategy fails to scale up to the in-the-wild setting. While it enhances the localization of target objects, supervision from binary instance masks hinders learning diverse matting patterns in the wild (see $\mathrm{SAD}_{FG}$ metric), especially for transparent objects. On the contrary, our self-training framework successfully generates the soft-pseudo label and utilizes it as a direct supervision to matting tasks, resulting in huge improvements in all the metrics. + +Analysis on Perturbation in Self-training. To demon- + +Table 1. Quantitative evaluation on Composition-Wild and AIM-500 [20]. * denotes the improved version of MGMatting [50]. "Instance-wise" represents instance-wise learning. A lower score is better in all the metrics. + +
MethodComposition-WildAIM-500
SADMSE\( SAD_{FG} \)\( SAD_{BG} \)SADMSE
Ours58.160.004647.3210.8416.720.0030
No Image.60.080.004845.1914.8918.570.0039
No Guide.60.140.004947.9512.1918.270.0035
+ +strate the effectiveness of perturbations in self-training, we ablate the image and guidance perturbations separately. The results are summarized in Table 2. As degraded performance denoted, encouraging robustness against image and guidance perturbations is crucial to the success of the self-training. + +We further investigate whether the matting-specific design of image perturbation is necessary. To this end, we train three variants with different configurations of image perturbations: (1) Pixel-level linear augmentation (Pixel), (2) Region-level augmentation (Region) and (3) Both of them (Pixel+Region). Below, we report the SAD metric for Composition-Wild and AIM-500, respectively. + +Table 2. Ablation Study on Perturbations in Self-training. + +
PixelRegionPixel+Region
58.16 / 16.7269.8 / 25.9661.88 / 19.00
+ +We can observe that the region-level augmentation, which is widely used in other tasks [37, 47], disturbs the effects of self-training objectives. As discussed in Sec. 4.2, this is because the region-level augmentation breaks the underlying opacity of objects and make a mismatch between inputs of the student and the teacher networks. On the contrary, pixel-level linear augmentation brings clear improvements. + +# 7. Real-world Applications + +To show the generality and potential of the in-the-wild mask-guided matting model, we illustrate two new applications: mask-guided video matting and panoptic matting. Fig. 6 and Fig. 7 show examples of each application. + +![](images/df35dffa257ae87bf259cec91d4214e5e7450c408112602afa8a2503d958df7a.jpg) +(a) Frame + +![](images/1e9de2c76caaac8aa0a266288a8f61d7e1c3ed418a21e73e787fc981f61e7f00.jpg) +(b) Guidance + +![](images/5cbbd3e6a457d0eb16f53a3859b2898334e3d573f026143dc7a6d4f47941bd6d.jpg) +(c) MGMatting + +![](images/6531bcc6bed2208c88443aebe8bc15bcbcf08a5c157b8bf46724e41ae4f25859.jpg) +(d) Ours + +![](images/66ce27abec34c50cf96fa1adf7bc7507939e0e06e27c01b5dd92f06e1bb2c74f.jpg) +(e) Editing Result + +![](images/1675affb9f2e296a2dbd2395b28fa95cbf03fb58770c4c717f70d8a27156e4e9.jpg) +Figure 6. Extension to Mask-guided Video Matting. We use propagated masks [31] as guidance. + +![](images/806838b63f2647d6f10ef948ea06202462efe355996cb41a7562946d6e6a2302.jpg) + +![](images/201a1a4f25cb8dad2c51e2b1e60071e2971b5bb8fc13d611f56e404bec6df81e.jpg) + +![](images/6091fd7c8acb78a6215beb65c9a19aebdde2ddc179c98a746eefcda6ebf432ab.jpg) + +![](images/9ab6844626e07b606e6391ffeea22475bfd024fbf199e97a090878a54c07d237.jpg) +(a) Image + +![](images/6c11a8b55ec1aebfc600dd74aa08e4d13f17b4e295cd9317a548edfc3085d545.jpg) +(b) Guidance + +![](images/e87bd9bd5a3d84b551c8f626628476d82b338360267358f46aaa5ec7d99919d0.jpg) +(c) MGMatting + +![](images/f5e894778455953555281dbb3978bc6e85ff6bad80304c7da9c69fef306fd2ef.jpg) +(d) Ours +Figure 7. Extension to Panoptic Matting. Hand drawn panoptic masks [24] are utilized as guidance. We illustrate independent regions with different colors. The black color represents ignored regions of labels or predictions. + +![](images/b4decbd3b062b64b1a670503cd46657166d80d97315fc538694d288346b7f772.jpg) +(e) Editing Result + +# 7.1. Mask-guided Video Matting + +Current video matting solutions [40, 53] require multiple frames of trimap to produce an alpha matte for a video. Therefore, required user interaction is more burdensome in the video domain. With the mask-guided matting model, we explore a more practical scenario where only a single instance mask is given in the first frame. In this challenging setting, we first propagate the given mask to the rest of the video frames using the video object segmentation model [31]. Then, the mask-guided matting model utilizes the propagated mask as guidance and produces alpha matte in a frame-by-frame manner. Our model not only corrects the noise in propagated masks but also predicts reasonable opacity in challenging cases such as motion blur. With the prediction, we also provide visually appealing video editing results (i.e., background replacement). + +# 7.2. Panoptic Matting + +Panoptic segmentation aims to parse an image into nonoverlapping regions of stuff and things. By the definition of the task format, the soft transition between the regions is overlooked. We explore whether our mask-guided matting model could provide such understanding from panoptic segmentation results. Given a panoptic mask, our model con + +siders the segment of each region separately and produces the corresponding alpha matte. We form panoptic matting results by aggregating the predictions. Unlike the baseline, our model successfully refines the fine details and captures the soft transition between the regions. Interestingly, while the model is trained on objects (i.e., things), our method is well generalized on the stuff regions (See sky region in Fig. 7). Panoptic matting enables us to do region-level image editing and the qualitative results are shown in Fig. 7. + +# 8. Conclusion + +In this paper, we propose a simple yet effective learning framework for mask-guided matting in the wild. We first introduce instance-wise learning to learn generalized matting ability from the composited images. Then, the model is further extended to handle diverse objects under the self-training on weak-guidance images. Extensive experiments on the newly proposed evaluation benchmark provide an in-depth understanding of the unique challenges. Using our approach, we instantiate new practical applications: mask-guided video matting and panoptic matting. We hope our proposals inspire future work on mask-guided matting. + +Acknowledgement This work was partially supported by the NRF (NRF-2020M3H8A1115028, FY2021) and Samsung Electronics Co., Ltd (G01200447). + +# References + +[1] Tamara L Berg and Alexander C Berg. Finding iconic images. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, pages 1-8. IEEE, 2009. 1 +[2] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 1 +[3] Quan Chen, Tiezheng Ge, Yanyu Xu, Zhiqiang Zhang, Xinxin Yang, and Kun Gai. Semantic human matting. In Proceedings of the 26th ACM international conference on Multimedia, pages 618-626, 2018. 3 +[4] Qifeng Chen, Dingzeyu Li, and Chi-Keung Tang. Knn matting. IEEE transactions on pattern analysis and machine intelligence, 35(9):2175-2188, 2013. 2 +[5] Yung-Yu Chuang, Brian Curless, David H Salesin, and Richard Szeliski. A bayesian approach to digital matting. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 2, pages II-II. IEEE, 2001. 2 +[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 5 +[7] Marco Forte and François Pitie. $f, b$ , alpha matting. arXiv preprint arXiv:2003.07711, 2020. 1, 2 +[8] Eduardo SL Gastal and Manuel M Oliveira. Shared sampling for real-time alpha matting. In Computer Graphics Forum, volume 29, pages 575-584. Wiley Online Library, 2010. 2 +[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 5 +[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 6 +[11] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaouu Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 2 +[12] Kaiming He, Jian Sun, and Xiaou Tang. Fast matting using large kernel matting laplacian matrices. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2165-2172. IEEE, 2010. 2 +[13] Qiqi Hou and Feng Liu. Context-aware image matting for simultaneous foreground and alpha estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4130-4139, 2019. 1, 2, 5 +[14] Ronghang Hu, Piotr Dólar, Kaiming He, Trevor Darrell, and Ross Girshick. Learning to segment every thing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4233-4241, 2018. 2 + +[15] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson W.H. Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In AAAI, 2022. 1, 2, 3 +[16] Philip Lee and Ying Wu. Nonlocal matting. In CVPR 2011, pages 2193-2200. IEEE, 2011. 2 +[17] Anat Levin, Dani Lischinski, and Yair Weiss. A closed-form solution to natural image matting. IEEE transactions on pattern analysis and machine intelligence, 30(2):228-242, 2007. 1, 2 +[18] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 2 +[19] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 2, 3, 5 +[20] Jizhizi Li, Jing Zhang, and Dacheng Tao. Deep automatic natural image matting. arXiv preprint arXiv:2107.07235, 2021. 2, 5, 6, 7 +[21] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11450-11457, 2020. 1, 2, 4, 6 +[22] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 1 +[23] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 7 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 3, 5, 6, 8 +[25] Jinlin Liu, Yuan Yao, Wendi Hou, Miaomiao Cui, Xuansong Xie, Changshui Zhang, and Xian-sheng Hua. Boosting semantic human matting with coarse annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8572, 2020. 3, 7 +[26] Yuhao Liu, Jiake Xie, Xiao Shi, Yu Qiao, Yujie Huang, Yong Tang, and Xin Yang. Tripartite information mining and integration for image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7555-7564, 2021. 2 +[27] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3266-3275, 2019. 1, 6 +[28] Sebastian Lutz, Konstantinos Amplanitis, and Aljosa Smolic. Alphagan: Generative adversarial networks for natural image matting. arXiv preprint arXiv:1807.10088, 2018. 2 +[29] Stephen Palmer. Canonical perspective and the perception of objects. Attention and performance, pages 135-151, 1981. 1 + +[30] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11696-11706, 2022. 1, 2, 4, 6 +[31] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Per-clip video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1352–1361, 2022. 8 +[32] Yu Qiao, Yuhao Liu, Xin Yang, Dongsheng Zhou, Mingliang Xu, Qiang Zhang, and Xiaopeng Wei. Attention-guided hierarchical structure aggregation for image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13676-13685, 2020. 1, 2, 5 +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6 +[34] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 1, 2 +[35] Ehsan Shahrian, Deepu Rajan, Brian Price, and Scott Cohen. Improving image matting using comprehensive sampling sets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 636-643, 2013. 2 +[36] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 5 +[37] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2, 5, 7 +[38] Jian Sun, Jiaya Jia, Chi-Keung Tang, and Heung-Yeung Shum. Poisson matting. In ACM SIGGRAPH 2004 Papers, pages 315–321. 2004. 2 +[39] Yanan Sun, Chi-Keung Tang, and Yu-Wing Tai. Semantic image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11120–11129, 2021. 5 +[40] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 8 +[41] Jingwei Tang, Yagiz Aksoy, Cengiz Oztireli, Markus Gross, and Tunc Ozan Aydin. Learning-based sampling for natural image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3055–3063, 2019. 2, 4 +[42] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve + +semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 5 +[43] Jue Wang and Michael F Cohen. An iterative optimization approach for unified image segmentation and matting. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 936-943. IEEE, 2005. 1 +[44] Jue Wang and Michael F Cohen. Optimized color sampling for robust matting. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 2 +[45] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Hanqing Zhao, Weiming Zhang, and Nenghai Yu. Improved image matting via real-time user clicks and uncertainty estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15374-15383, 2021. 1, 2 +[46] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXV, pages 238–258. Springer, 2022. 2 +[47] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 2, 5, 7 +[48] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 3, 5, 6 +[49] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 3 +[50] Qihang Yu, Jianming Zhang, He Zhang, Yilin Wang, Zhe Lin, Ning Xu, Yutong Bai, and Alan Yuille. Mask guided matting via progressive refinement network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1154-1163, 2021. 1, 2, 3, 4, 5, 6, 7 +[51] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 4 +[52] Yunke Zhang, Lixue Gong, Lubin Fan, Peiran Ren, Qixing Huang, Hujun Bao, and Weiwei Xu. A late fusion cnn for digital matting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7469-7478, 2019. 2 +[53] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 6, 8 \ No newline at end of file diff --git a/2023/Mask-Guided Matting in the Wild/images.zip b/2023/Mask-Guided Matting in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e5a111c6684a58eb711a067df98f157aeb01462e --- /dev/null +++ b/2023/Mask-Guided Matting in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:476156b55c20316aa99ca31bcd253ac8be23f3937001f19f01d77ca642cfa65a +size 420855 diff --git a/2023/Mask-Guided Matting in the Wild/layout.json b/2023/Mask-Guided Matting in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d65d20ee33d3254c6486969d161f95867243b95 --- /dev/null +++ b/2023/Mask-Guided Matting in the Wild/layout.json @@ -0,0 +1,8390 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "spans": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "type": "text", + "content": "Mask-guided Matting in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "text", + "content": "Kwanyong Park" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "text", + "content": " Sanghyun Woo" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "text", + "content": " Seoung Wug Oh" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "text", + "content": " In So Kweon" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "text", + "content": " Joon-Young Lee" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "text", + "content": "KAIST " + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 526, + 185 + ], + "type": "text", + "content": "Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "content": "Mask-guided matting has shown great practicality compared to traditional trimap-based methods. The mask-guided approach takes an easily-obtainable coarse mask as guidance and produces an accurate alpha matte. To extend the success toward practical usage, we tackle mask-guided matting in the wild, which covers a wide range of categories in their complex context robustly. To this end, we propose a simple yet effective learning framework based on two core insights: 1) learning a generalized matting model that can better understand the given mask guidance and 2) leveraging weak supervision datasets (e.g., instance segmentation dataset) to alleviate the limited diversity and scale of existing matting datasets. Extensive experimental results on multiple benchmarks, consisting of a newly proposed synthetic benchmark (Composition-Wild) and existing natural datasets, demonstrate the superiority of the proposed method. Moreover, we provide appealing results on new practical applications (e.g., panoptic matting and mask-guided video matting), showing the great generality and potential of our model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 501, + 128, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 128, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 128, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": "Image matting aims to predict the opacity of objects, which enables precise separation from surrounding backgrounds. Due to the ill-posed nature of the task, many works [7, 13, 21, 27, 30, 48] have improved matting performance by relying on the manual guidance of a trimap. However, pixel-level annotation of foreground/background/unknown is extremely burdensome, restricting its usage in many practical applications such as image/video editing and film production. Recently, many efficient alternatives for user guidance have been proposed, including trimap-free [15, 32], additional background images [22, 34], scribble [43], and the user clicks [45]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Among them, the mask-guided approach [50] shows a great trade-off between performance and intensity of user interaction. It utilizes a coarse mask as guidance, which is much easier to obtain either manually or from off-the-shelf" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 557, + 386 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 557, + 386 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 557, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 557, + 386 + ], + "type": "image", + "image_path": "3c5b876cfad515532688621335ac06fec5c6ded6040dfb44f7c80a11e3e8d53d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 396, + 547, + 430 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 547, + 430 + ], + "type": "text", + "content": "Figure 1. Qualitative Comparisons of MGMatting [50] and Ours in the wild. The mask guidance is overlaid on images with blue color. Best viewed zoomed in." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 458, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 555 + ], + "type": "text", + "content": "segmentation models [2, 10]. With only the coarse spatial prior, the mask-guided matting model [50] shows comparable or even better performance than the trimap-based competitors [13, 17, 21, 27, 48] on synthetic Composition-1k [48] and a real-world human matting dataset [50]. However, despite the encouraging results, we see the previous state-of-the-art model [50] struggles to obtain desirable alpha matte in complex real-world scenes (see Fig. 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "type": "text", + "content": "With this observation, we tackle mask-guided matting in the wild. Specifically, we formulate unique setups and emerging challenges of the new task as follows: (1) We aim to handle objects in their complex context, reflecting the characteristics of natural images. The previous method [50] evaluates their model on iconic-object images [1,29] where only a single object is in the center. As the model can easily find the target object in such images, the model's real instance discrimination ability is, in fact, veiled. On the contrary, in an 'in-the-wild' setting, it is crucial to precisely localize the target object from the given coarse/noisy mask guidance (i.e. mask awareness). (2) Our model targets to deal with diverse categories of objects in natural" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1992" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "images. Unlike most previous methods that improve generalization performance at the expense of category-specific regime (e.g., limiting to humans [15] or animals [19]), we aim to understand distinctive matting patterns of vast categories. (3) Limited data problem makes the new setting more complicated. Due to the labeling complexity, annotating alpha matte for objects in common scenes, e.g., the COCO dataset [24], is infeasible. As a sidestep, previous benchmarks [32, 48] extract the alpha matte and foreground colors from images with simple backgrounds. These are composited on various backgrounds [6, 24], and resulting samples are used to train and evaluate matting models. However, due to the inevitable composition artifacts, the models usually show limited generalization performance. In that sense, how to train and evaluate the in-the-wild matting model remains an open question." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 495 + ], + "type": "text", + "content": "Toward this goal, we propose a simple yet effective learning framework for a generalized mask-guided matting model. First, we investigate fundamental reasons for the poor generalization of the previous mask-guided matting model [50] and find that this is mainly from the training data generation process. Specifically, the previous composition process includes instance merging data augmentation, which merges several foreground objects into a single object. While this augmentation is effective in the trimap-based methods [21, 30, 41], it implicitly makes a negative bias for the mask-guided matting model to ignore the guidance. Thus, the model struggles to localize the target objects in complex natural scenes. We alleviate the bias by proposing an instance-wise learning objective, where the model is supervised to segment one of multiple instances according to the guidance. By doing so, the model learns strong semantic representation regarding complex relations and soft transitions between objects. Despite the simplicity of the proposal, this greatly improves performance in the wild." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 496, + 289, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 289, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 289, + 675 + ], + "type": "text", + "content": "Second, we explore a practical solution to make the mask-guided model handle various categories of objects robustly. Instead of scaling the matting dataset, we leverage a dataset with weak supervision [14, 46] (i.e., instance segmentation dataset [24]), as the coarse instance masks are easier to obtain over the diverse categories of objects. To effectively hallucinate the fine supervision signal with the weak localization guidance, we come up with a self-training framework [36, 37, 47]. Specifically, a pseudo label is generated based on a weakly-augmented input (both image and instance mask annotation as guidance), which supervises the model prediction on a strongly augmented version of them. During self-training, the model is not only adapting to the in-the-wild scenario in a self-evolving manner but also being robust to noise in both image and guidance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "To verify the in-the-wild performance of mask-guided matting, we formally define an evaluation protocol involving multiple sub-benchmarks: Composition-Wild, AIM-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "type": "text", + "content": "500 [20], COCO [24]. We first design an in-the-wild extension of the popular synthetic Composition-1k benchmark [48], namely Composition-Wild. We simulate complex real-world images by compositing multiple foreground objects. To bring valuable insight on the failure cases of the model, we design sub-metrics for Composition-Wild. In addition, we use the AIM-500 dataset to establish quantitative results on natural images (i.e., with no composition artifacts), although most images are iconic-object images with simple backgrounds. Finally, we provide qualitative outputs of our mask-guided matting model on the COCO dataset [24] which is one of the most representative in-the-wild datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 228, + 547, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 313 + ], + "type": "text", + "content": "To summarize, we make the following main contributions. 1) To our best knowledge, it is the first work to explore mask-guided matting in the wild. 2) We develop a simple yet effective learning framework leveraging both composited and weak-guidance images. 3) We design an evaluation setup for the new task. 4) We initiate several interesting extensions: video and panoptic matting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "type": "text", + "content": "Natural Image Matting. Most image matting methods require a trimap as additional input, which conveys pixel-level annotations of foreground, background, and unknown region. Traditional image matting methods can be categorized into two groups. Sampling-based methods [5, 8, 11, 35, 44] estimate the alpha matte of the unknown region based on sampled colors of foreground and background. Propagation-based methods [4, 12, 16-18, 38] propagate the neighboring known alpha values to unknown regions according to the affinity between pixels. Both methods primarily rely on color or low-level features, showing limitations in complicated scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "type": "text", + "content": "To tackle these issues, many deep-learning based mating methods have been proposed. DIM [48] is a representative work which proposes a convolutional encoder-decoder matting network as well as a large-scale synthetic dataset (Composition-1k) to train the model. Many follow-up works have made tremendous improvements in diverse aspects, such as sophisticated loss design [7, 13, 28] and architectural advances by introducing attention mechanisms [21, 26] or transformer architecture [30]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "There is a large volume of trials to relax the heavy user-supplied constraints. Some methods [32, 52] attempt to get rid of trimap. However, they show inferior performance and cannot be generalized to unseen objects in the real world. Sengupta et al. [34] utilize additional background images along with other lightweight priors (e.g., segmentation mask) to perform the matting task. Wei et al. [45] introduce user click interaction to effectively eliminate the ambiguity of target object. Recently, MGMatting [50] proposes a mask-guided matting framework, where only the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1993" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 261, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 261, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 261, + 83 + ], + "type": "text", + "content": "easily obtainable coarse mask is needed as guidance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 144 + ], + "type": "text", + "content": "In this paper, we further study mask-guided matting in an in-the-wild setting, inspired by the generality and practicality of the framework. Different from MGMatting, we target diverse categories of objects in their complex background, enabling real-world applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 144, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 310 + ], + "type": "text", + "content": "Class-specific Matting is a special type of matting that limits target objects to one or few classes such as humans [3] or animals [19]. BSHM [25] leverages the coarse segmentation mask data to build a generalized human matting network. MODNet [15] presents a lightweight network architecture with decomposed multi-scale network designs. GFM [19] first proposes the task of animal matting and a real-world animal matting dataset. In general, known semantics effectively ease the difficulty of matting, and these methods show great generality without the necessity of additional guidance input. However, the specialized methods and trained model cannot be generalized to the in-the-wild setting since limited semantics do not cover diverse patterns of visual worlds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 286, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 286, + 396 + ], + "type": "text", + "content": "Apart from previous efforts, we attempt to extend the short category regime of matting networks. For this purpose, we propose an efficient solution by leveraging the instance segmentation dataset [24], which covers diverse categories of objects with weak localization annotations. Given the weak guidance dataset, we study the challenges in developing an in-the-wild mask-guided matting model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 406, + 234, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 234, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 234, + 419 + ], + "type": "text", + "content": "3. Mask-guided Matting in the Wild" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 286, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 592 + ], + "type": "text", + "content": "In natural scenes, objects are rarely in isolation: diverse categories of entities interact with each other. We aim to robustly separate a target object from the surroundings in such scenes. Specifically, we formulate the problem in the form of mask-guided matting, namely mask-guided matting in the wild. Our matting model takes an easily-obtainable coarse mask as guidance, either manually drawn or predicted from the off-the-shelf segmentation models. With in-the-wild data, the model needs to understand semantics so that it can precisely localize the target object from the given noisy masks. In the meantime, the model also needs to have strong low-level feature representation so that it can capture fine details and accurate opacity of the diverse target matte." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "To concrete the new challenges toward such an in-the-wild setting, we first study how the previous model [50] performs. We derive predictions with different mask guidance and summarize the results on Fig. 2. As shown in the upper example, in the synthetic Composition-1k [48] example, the previous model perfectly separates the foreground object regardless of the given guidance. Surprisingly, it produces an alpha matte of the foreground object when the mask of the background region is given as guidance. On the contrary, the model lacks generalization on" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 391, + 183 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 391, + 183 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 391, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 391, + 183 + ], + "type": "image", + "image_path": "2b76c8772c71bce62191100fc3b91dd8c868b9db95412ce38c0475a5b25abb0f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 185, + 380, + 193 + ], + "lines": [ + { + "bbox": [ + 318, + 185, + 380, + 193 + ], + "spans": [ + { + "bbox": [ + 318, + 185, + 380, + 193 + ], + "type": "text", + "content": "Image and Guidance" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 391, + 71, + 472, + 184 + ], + "blocks": [ + { + "bbox": [ + 391, + 71, + 472, + 184 + ], + "lines": [ + { + "bbox": [ + 391, + 71, + 472, + 184 + ], + "spans": [ + { + "bbox": [ + 391, + 71, + 472, + 184 + ], + "type": "image", + "image_path": "eacb1b650023f83cdb64dc8a8df0a693bd1c1a506c95be8140057038e68bd50e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 185, + 468, + 193 + ], + "lines": [ + { + "bbox": [ + 395, + 185, + 468, + 193 + ], + "spans": [ + { + "bbox": [ + 395, + 185, + 468, + 193 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 395, + 185, + 468, + 193 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 395, + 185, + 468, + 193 + ], + "type": "text", + "content": " Mask Guidance" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 473, + 71, + 556, + 184 + ], + "blocks": [ + { + "bbox": [ + 473, + 71, + 556, + 184 + ], + "lines": [ + { + "bbox": [ + 473, + 71, + 556, + 184 + ], + "spans": [ + { + "bbox": [ + 473, + 71, + 556, + 184 + ], + "type": "image", + "image_path": "be7855bebc8978d1a1b719e61b8db458d1e916b662cc9bad1872c4c8c3267b1c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 185, + 550, + 193 + ], + "lines": [ + { + "bbox": [ + 478, + 185, + 550, + 193 + ], + "spans": [ + { + "bbox": [ + 478, + 185, + 550, + 193 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 478, + 185, + 550, + 193 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 478, + 185, + 550, + 193 + ], + "type": "text", + "content": " Mask Guidance" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 201, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 235 + ], + "type": "text", + "content": "Figure 2. Predictions of MGMatting [50] with different guidance. (upper) synthetic Composition-1k image [48]. (bottom) Complex real world image [49]. Best viewed zoomed in." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 453 + ], + "type": "text", + "content": "complex real-world scenes (see bottom example). In particular, the model fails to localize the target object and makes wrong predictions on regions of similar color distributions (e.g., shadows) or another nearby object. From the two representative examples, we make several observations: 1) The previous method yields a model overfitted to the (trained) synthetic matting dataset, which memorizes the limited patterns of foreground objects rather than semantically understanding the mask guidance for localizing and segmenting the objects; 2) Such a model fails to precisely localize and segment a target object in the wild; 3) Previously used synthetic benchmark, Composition-1k [48], neither represents the complex nature of the in-the-wild setting nor systematically blocks the shortcut of the overfitting issue. Thus the Composition-1k may not be a suitable benchmark for mask-guided matting in the wild." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "type": "text", + "content": "To this end, we explore how to train a generalized mask-guided matting model given the limited scale and diversity of the matting dataset (Sec. 4), as well as revisit the evaluation benchmark to test the model (Sec. 5)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 523, + 411, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 411, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 411, + 537 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "The overview of the proposed training framework is illustrated in Fig. 3. Our framework involves three types of datasets: 1) matting dataset that includes accurate alpha matte and corresponding foreground colors; 2) background dataset, where the foreground objects are composited to create training samples; 3) weak supervision dataset that is introduced to extend the matting model to the in-the-wild setting. The main goal of our framework is to learn accurate matting ability from the synthetic composited images and generalize the knowledge to the in-the-wild domain. To this end, we first investigate how the model could learn generalized knowledge from the composited images (Sec. 4.1), then how the model is further improved with weak supervision data (Sec. 4.2)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1994" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 511, + 203 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 511, + 203 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 511, + 203 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 511, + 203 + ], + "type": "image", + "image_path": "d8690695c1e705c06e4da75d42df126bfe25625e99137ab44b9a87a179b5a0d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 86, + 206, + 498, + 295 + ], + "blocks": [ + { + "bbox": [ + 86, + 206, + 498, + 295 + ], + "lines": [ + { + "bbox": [ + 86, + 206, + 498, + 295 + ], + "spans": [ + { + "bbox": [ + 86, + 206, + 498, + 295 + ], + "type": "image", + "image_path": "b37ab59244c182cbd48597bd63b1400e7de8680145ae377aeaaf83ad18bc3498.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 305, + 546, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 305, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 546, + 328 + ], + "type": "text", + "content": "Figure 3. The Overview of the Proposed Framework. (a) Learning from composited images with instance-wise learning (Sec. 4.1). (b) Learning from weak-supervision images with self-training (Sec. 4.2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 348, + 234, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 234, + 362 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 234, + 362 + ], + "type": "text", + "content": "4.1. Learning from Composed Images" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "type": "text", + "content": "As discussed in the Sec. 3, the previous model [50] shows poor generalization performance in the challenging real-world setting. Specifically, it often fails to identify the target object in a given guidance and produces a matting result for other objects. We found that the problem, the lack of mask awareness, comes from the previous data pipeline to generate training samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "text", + "content": "For a clear understanding, we first briefly summarize the previous training data generation step proposed in MGMatting [50]. Here, for clarity, generic data augmentation such as random-crop, resize, and affine-transform are omitted. A foreground object (consisting of foreground color " + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "text", + "content": " and alpha matte " + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "text", + "content": ") and a background image " + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "text", + "content": " are randomly sampled from the matting and background datasets, respectively. These are composited to form a training image " + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 451, + 287, + 558 + ], + "type": "text", + "content": " following the composition formulation as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 563, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 563, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 72, + 563, + 287, + 576 + ], + "type": "interline_equation", + "content": "I = \\operatorname {C o m p o s i t e} (F, \\alpha , B) = \\alpha F + (1 - \\alpha) B. \\tag {1}", + "image_path": "ea519f68d4a36610ff207a7fd688fb9c58953e4c05701cf16b652113ece32be2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": "To simulate the mask guidance " + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": ", they first binarize the alpha matte with a threshold randomly sampled from 0 to 1, then apply dilation and/or erosion to it. In addition, they propose a stronger mask augmentation, CutMask [50], originally designed for robustness to noise in mask guidance. Like CutMix [51], a random-sized patch is selected, and the content is pasted on a random position." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "In addition, MGMatting [50] adopted instance merging augmentation as most state-of-the-art trimap-based matting methods [21, 30, 41] did. Specifically, they select two random foreground objects, merge them into a single object," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 348, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 546, + 540 + ], + "type": "text", + "content": "and composite it on a random background image. The instance merging augmentation effectively alleviates the limited scale and diversity of foreground objects in the matting dataset. However, in the context of mask-guided matting, we found that this augmentation hinders instance-wise understanding and makes a negative bias that the model eventually segments all the existing composited objects regardless of the given mask guidance. The different effects of the same augmentation are originated from the uncertainty of mask guidance. While the trimap brings pixel-level accurate localization of foreground/background/unknown regions, the mask guidance naturally conveys noisy information, and it is hard to localize them accurately. Under such uncertainty, the model tends to find the shortcut, segmenting all the composited objects rather than learning to utilize the uncertain mask guidance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "text", + "content": "Instance-wise Learning. We systematically block the above shortcut and enforce the model to learn complex relations and soft transitions between objects. The idea is simple: supervise the model to segment one of them according to the guidance (See Fig. 3-(a)). Specifically, given the two foreground objects (colors " + }, + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "inline_equation", + "content": "F_{1}, F_{2}" + }, + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\alpha_{1}, \\alpha_{2}" + }, + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "text", + "content": "), we sequentially composite these foreground layers (the first object is on top.) on background image " + }, + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 540, + 545, + 635 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 643, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 643, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 318, + 643, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I = \\text {C o m p o s i t e} \\left(F _ {1}, \\alpha_ {1}, \\text {C o m p o s i t e} \\left(F _ {2}, \\alpha_ {2}, B\\right)\\right) \\tag {2} \\\\ = \\alpha_ {1} F _ {1} + \\alpha_ {2} (1 - \\alpha_ {1}) F _ {2} + (1 - \\alpha_ {1}) (1 - \\alpha_ {2}) B. \\\\ \\end{array}", + "image_path": "549e9562d0e96ad362f303b9e047b6d814efa28b05d41999c5636fe373b77cbb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": "Based on the above equation, we re-calibrate the fractional contribution of each foreground object to images at pixel level as follows: " + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_1 = \\alpha_1, \\overline{\\alpha}_2 = \\alpha_2 (1 - \\alpha_1)" + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": ". We randomly" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1995" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": "select an index " + }, + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "inline_equation", + "content": "i_{rand} \\in \\{1,2\\}" + }, + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": " of a target object. With the guidance " + }, + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "inline_equation", + "content": "M(\\overline{\\alpha}_i)" + }, + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": ", the model predicts the opacity of the target object " + }, + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "inline_equation", + "content": "\\widetilde{\\alpha}_i" + }, + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": ", which is supervised by corresponding ground-truth alpha matte " + }, + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_i" + }, + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": ". Such instance-wise training makes the model learn fractional differentiation between objects according to the given coarse mask guidance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 153, + 286, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 153, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 153, + 286, + 166 + ], + "type": "text", + "content": "4.2. Learning from Images with Weak Supervision" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 172, + 286, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 286, + 303 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 286, + 303 + ], + "type": "text", + "content": "Due to the annotation difficulties, the labels of the alpha matte are practically infeasible to collect for large quantities and diverse categories. By comparison, coarse mask annotations are more abundant and accessible, thus extending the dataset scale and taxonomies is easier to achieve. Motivated by this, we leverage the weak supervision dataset (i.e., instance segmentation [24]), and investigate how these new data benefit the generalization of the mask-guided matting model. The main challenge lies in how to get matting supervision because the instance mask labels are noisy in fine details and do not include opacity annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 286, + 376 + ], + "type": "text", + "content": "To effectively tackle this challenge, we design a self-training framework (See Fig. 3-(b)). The framework generates pseudo matting labels " + }, + { + "bbox": [ + 46, + 304, + 286, + 376 + ], + "type": "inline_equation", + "content": "\\widehat{\\alpha}" + }, + { + "bbox": [ + 46, + 304, + 286, + 376 + ], + "type": "text", + "content": " under the guidance of instance masks, and the pseudo labels supervise the matting model on natural images. To obtain high-quality pseudo labels, we adopt several design choices as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 449 + ], + "type": "text", + "content": "Teacher-Student Framework. Motivated by the success in self-/semi-supervised learning [9, 42], a teacher network is introduced, which is a slowly advancing version of a student model via the exponential moving average (EMA). We utilize stable predictions of the teacher network as the pseudo labels " + }, + { + "bbox": [ + 46, + 376, + 286, + 449 + ], + "type": "inline_equation", + "content": "\\widehat{\\alpha}" + }, + { + "bbox": [ + 46, + 376, + 286, + 449 + ], + "type": "text", + "content": " and guide the student model during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "text", + "content": "Weakly-augmented Input. The teacher network takes weakly-augmented samples to generate the pseudo labels. We only apply standard geometric augmentation, such as affine transform and random crop, for both image and mask guidance. Despite the noise in fine details, the ground-truth instance mask conveys a strong localization cue to the model, resulting in high-quality pseudo labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "text", + "content": "For strongly-augmented inputs, the student model produces predictions " + }, + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\widehat{\\alpha}^{\\prime}" + }, + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "text", + "content": ", which are supervised to be the same as the pseudo labels " + }, + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\widehat{\\alpha}" + }, + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "text", + "content": " obtained from the teacher model. To form the strongly-augmented version, we design two types of augmentations: image and guidance perturbations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Image Perturbations. Previous self-training-based methods [36, 37, 47] often employ region-level augmentation (e.g., Gaussian blur) to generate the strongly-augmented images. However, in the matting task, these augmentations are unsuitable since they interpolate the color value across the pixels and disturb each object's fractional contribution. This results in a mismatch of ground-truth alpha matte between the teacher and student inputs, lowering the effectiveness of self-training. We instead adopt linear pixel-level augmentations such as additive Gaussian noise and linear" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 73, + 343, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 73, + 343, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 73, + 343, + 83 + ], + "type": "text", + "content": "contrast." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 84, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 190 + ], + "type": "text", + "content": "Guidance Perturbations. We also modulate the reliability of mask guidance via dilation and/or erosion of the instance mask. Considering the diverse size of objects in natural scenes, we decide the kernel size based on the size of the object. Specifically, we set the kernel size as " + }, + { + "bbox": [ + 304, + 84, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\beta \\%" + }, + { + "bbox": [ + 304, + 84, + 545, + 190 + ], + "type": "text", + "content": " of the shorter side of an object bounding box. The guidance perturbation is crucial to not only the success of self-training but also the robustness against common noise in mask guidance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 200, + 406, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 200, + 406, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 406, + 213 + ], + "type": "text", + "content": "4.3. Unified Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 218, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 290 + ], + "type": "text", + "content": "As shown in Fig. 3, under the proposed framework, the (student) mask-guided matting model can jointly learn from composited images with ground-truth labels and natural images with weak supervision. The final loss function is a combination of losses on both images as: " + }, + { + "bbox": [ + 304, + 218, + 545, + 290 + ], + "type": "inline_equation", + "content": "L_{final} = L_{composed} + L_{natural}" + }, + { + "bbox": [ + 304, + 218, + 545, + 290 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "text", + "content": "For the composited images, we use the same loss function as MGMatting [50], which is the summation of " + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "text", + "content": " regression loss, composition loss [48], and Laplacian loss [13]. Thus, " + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "inline_equation", + "content": "L_{\\text{composed}} = L_{l1}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i) + L_{\\text{comp}}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i) + L_{\\text{lap}}(\\widetilde{\\alpha}_i, \\overline{\\alpha}_i)" + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "text", + "content": ". To calculate the composition loss, we use ground-truth foreground color and alpha matte for non-target objects. For the natural images, we use " + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "text", + "content": " regression loss as follows: " + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "inline_equation", + "content": "L_{\\text{natural}} = L_{l1}(\\widehat{\\alpha}', \\widehat{\\alpha})" + }, + { + "bbox": [ + 304, + 291, + 545, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 397, + 386, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 397, + 386, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 386, + 410 + ], + "type": "text", + "content": "5. Benchmarks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "content": "There are several public matting benchmarks [19, 20, 32, 39, 48, 50] which provide high-quality alpha matte. However, none of them reflects the complexity of in-the-wild images and the consisting samples only contain a single object with simple backgrounds. Thus, we carefully design a new evaluation setup to facilitate an extensive evaluation of the mask-guided matting model in in-the-wild scenarios. The new setup consists of multiple benchmarks, and we describe the distinctive role of each benchmark below." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 525, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 645 + ], + "type": "text", + "content": "Composition-Wild. We first propose an in-the-wild extension of standard synthetic benchmark Composition-1k [48], namely Composition-Wild. We simulate the complex context of in-the-wild images by compositing multiple foreground objects onto a background image, similar to Eq.(2). The same foreground object data and background dataset (i.e., PASCAL VOC [6]) are utilized as Composition-1k. We report the performance over the forefront object from each sample unless otherwise specified. In addition, we design comprehensive metrics to allow in-depth analysis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 309, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 309, + 653, + 545, + 713 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 309, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SAD}_{FG}" + }, + { + "bbox": [ + 309, + 653, + 545, + 713 + ], + "type": "text", + "content": " reports the SAD (Sum of Absolute Difference) over the target object region where has non-zero ground-truth alpha matte. It indicates how well the prediction captures fine matting details of target objects (i.e., Target Detail Quality)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1996" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 174 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 132 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 50, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "\\mathrm{SAD}_{BG}" + }, + { + "bbox": [ + 50, + 72, + 286, + 132 + ], + "type": "text", + "content": " quantifies the wrong predictions on other objects or background regions by measuring the SAD over the non-target object region. It represents the capability of target object localization from the noisy mask guidance (i.e., Localization)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 138, + 286, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 138, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 50, + 138, + 286, + 174 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 50, + 138, + 286, + 174 + ], + "type": "inline_equation", + "content": "\\mathrm{SAD}_{OCC}" + }, + { + "bbox": [ + 50, + 138, + 286, + 174 + ], + "type": "text", + "content": " measures the SAD error of occluded objects. It indicates how well the model deals with occluded objects (i.e., Occlusion handling)." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 190, + 286, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 286, + 320 + ], + "type": "text", + "content": "AIM-500 [20] is a natural image matting benchmark that contains diverse categories of objects. The performance evaluation on the AIM benchmark denotes whether the model generalizes well to natural images rather than overfitting to the synthetic distribution of composited images. However, AIM is limited at capturing matting performance in complex contexts since they only contain images with simple backgrounds. In that regard, the evaluations on Composition-Wild and AIM complement each other and provide a holistic understanding of the mask-guided matting model in the wild." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 323, + 286, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 371 + ], + "type": "text", + "content": "COCO [24] is a representative example of in-the-wild images. While we cannot access ground-truth alpha mattes on COCO, we conduct qualitative comparisons over diverse objects in their natural context." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 390, + 127, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 127, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 127, + 403 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 412, + 179, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 179, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 179, + 425 + ], + "type": "text", + "content": "6.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 433, + 286, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 286, + 493 + ], + "type": "text", + "content": "Architecture details. We build our learning framework with a state-of-the-art mask-guided matting network, PRN from MGMatting [50], which adopts a standard encoder-decoder structure [21,33] and progressively refines the output through the decoding process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 495, + 286, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 286, + 602 + ], + "type": "text", + "content": "Training details. In practice, we adopt the two-stage training: pre-training on composited images and fine-tuning on both composited and natural images. This ensures the quality of pseudo labels at the beginning of the self-training, resulting in better performance. Following the previous works [21, 27, 30, 48], we use DIM [48] and COCO [24] as matting and background datasets, respectively. COCO dataset [24] is also utilized for a weak guidance dataset. We set the kernel ratio " + }, + { + "bbox": [ + 46, + 495, + 286, + 602 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 495, + 286, + 602 + ], + "type": "text", + "content": " as 0.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Evaluation details. We report the widely used four matting metrics, SAD (Sum of Absolute Differences), MSE (Mean Squared Error), Grad (Gradient), and Conn (Connectivity), over the whole image, when the ground-truth alpha matte is available. To simulate the coarse mask guidance, we binarize the alpha matte with threshold 0.5 and then dilate the mask with kernel size " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "25 \\times 25" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": ", similar to [53]. We use a single trained model for all the benchmark evaluations and real-world applications." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 71, + 391, + 133 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 391, + 133 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 391, + 133 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 391, + 133 + ], + "type": "image", + "image_path": "85494dbe39f0d64f6f98bdd7e733b9d89712e274bcaf7d2f5e5361e3741c7514.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 132, + 390, + 231 + ], + "blocks": [ + { + "bbox": [ + 308, + 132, + 390, + 231 + ], + "lines": [ + { + "bbox": [ + 308, + 132, + 390, + 231 + ], + "spans": [ + { + "bbox": [ + 308, + 132, + 390, + 231 + ], + "type": "image", + "image_path": "9d5f460f9a0447df5882b22c361b98d93a02b47362a37f2540e81c8b45b25518.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 232, + 378, + 240 + ], + "lines": [ + { + "bbox": [ + 318, + 232, + 378, + 240 + ], + "spans": [ + { + "bbox": [ + 318, + 232, + 378, + 240 + ], + "type": "text", + "content": "Image and Guidance" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 391, + 71, + 551, + 231 + ], + "blocks": [ + { + "bbox": [ + 391, + 71, + 551, + 231 + ], + "lines": [ + { + "bbox": [ + 391, + 71, + 551, + 231 + ], + "spans": [ + { + "bbox": [ + 391, + 71, + 551, + 231 + ], + "type": "image", + "image_path": "991b689f261faabd25116c9e2e44ce0fa6a62addcc71db2dea685d71754872d8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 232, + 447, + 240 + ], + "lines": [ + { + "bbox": [ + 413, + 232, + 447, + 240 + ], + "spans": [ + { + "bbox": [ + 413, + 232, + 447, + 240 + ], + "type": "text", + "content": "MGMatting" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 502, + 232, + 517, + 239 + ], + "lines": [ + { + "bbox": [ + 502, + 232, + 517, + 239 + ], + "spans": [ + { + "bbox": [ + 502, + 232, + 517, + 239 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 292, + 369, + 374 + ], + "blocks": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "type": "text", + "content": "Figure 4. Qualitative Results on COCO dataset. The guidance are either manually drawn [24] (upper) or a prediction of the off-the-self model [10] (bottom). Best viewed zoomed in." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 292, + 369, + 374 + ], + "lines": [ + { + "bbox": [ + 307, + 292, + 369, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 369, + 374 + ], + "type": "image", + "image_path": "7bdae739ce83e3eb04d28609ee72e38147012b5fa5c015c34fcc4e6a8afe6e39.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 375, + 352, + 384 + ], + "lines": [ + { + "bbox": [ + 325, + 375, + 352, + 384 + ], + "spans": [ + { + "bbox": [ + 325, + 375, + 352, + 384 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 346, + 391, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 346, + 391, + 504, + 402 + ], + "spans": [ + { + "bbox": [ + 346, + 391, + 504, + 402 + ], + "type": "text", + "content": "Figure 5. Effect of Instance-wise Learning." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 369, + 292, + 421, + 373 + ], + "blocks": [ + { + "bbox": [ + 369, + 292, + 421, + 373 + ], + "lines": [ + { + "bbox": [ + 369, + 292, + 421, + 373 + ], + "spans": [ + { + "bbox": [ + 369, + 292, + 421, + 373 + ], + "type": "image", + "image_path": "850a5a2772428642e7211cd0982d543f571e6268834e5a050991c7f31a267384.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 375, + 417, + 384 + ], + "lines": [ + { + "bbox": [ + 381, + 375, + 417, + 384 + ], + "spans": [ + { + "bbox": [ + 381, + 375, + 417, + 384 + ], + "type": "text", + "content": "(b) Guidance" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 432, + 292, + 490, + 373 + ], + "blocks": [ + { + "bbox": [ + 432, + 292, + 490, + 373 + ], + "lines": [ + { + "bbox": [ + 432, + 292, + 490, + 373 + ], + "spans": [ + { + "bbox": [ + 432, + 292, + 490, + 373 + ], + "type": "image", + "image_path": "330801958f40f89e03e7c8e609fa6d87da55a0556c07559395908fdbe7dfafaf.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 375, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 437, + 375, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 437, + 375, + 482, + 384 + ], + "type": "text", + "content": "(c) MGMatting*" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 491, + 292, + 551, + 374 + ], + "blocks": [ + { + "bbox": [ + 491, + 292, + 551, + 374 + ], + "lines": [ + { + "bbox": [ + 491, + 292, + 551, + 374 + ], + "spans": [ + { + "bbox": [ + 491, + 292, + 551, + 374 + ], + "type": "image", + "image_path": "e70c3092367d9616a17feb8b0e86241f38936383df0af1c909b76a7b3fbc9f34.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 375, + 533, + 383 + ], + "lines": [ + { + "bbox": [ + 507, + 375, + 533, + 383 + ], + "spans": [ + { + "bbox": [ + 507, + 375, + 533, + 383 + ], + "type": "text", + "content": "(d) +IWL" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 418, + 446, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 446, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 446, + 430 + ], + "type": "text", + "content": "6.2. Qualitative Comparisons" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 555 + ], + "type": "text", + "content": "Fig. 1 and Fig. 4 illustrate the qualitative comparisons on images in the COCO validation sets. We test the models under challenging scenarios such as occlusions, an unseen object in a matting dataset, complex surrounding context, and large interior errors on mask guidance. Compared with the previous mask-guided matting model [50], our model shows significantly better localization of a target object, as well as, produces much more accurate alpha mattes for challenging data. More qualitative comparisons can be found in the supplementary material." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 563, + 462, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 462, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 462, + 576 + ], + "type": "text", + "content": "6.3. Ablation Study and Analysis" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 582, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 545, + 604 + ], + "type": "text", + "content": "Quantitative evaluations on Composition-Wild and AIM-500 datasets are shown in Table 1." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of Instance-wise Learning. We first analyze how instance-wise learning contributes to learn generalized matting capability from composited images. We set the two baseline models: (naive) MGMatting and MGMatting*. For the (naive) MGMatting, we reproduce the model with the same training recipes proposed in MGMatting [50]. In addition, we report the stronger baseline score of MGMatting* by excluding a heavy guidance augmentation, CutMask [50]. We empirically found that CutMask" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1997" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 184 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 184 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 184 + ], + "type": "table", + "html": "
MethodComposition-WildAIM-500 [20]
SADMSEGradConn\\( SAD_{FG} \\)\\( SAD_{BG} \\)\\( SAD_{OCC} \\)SADMSEGradConn
Stage 1: Learning with Composited Images
MGMatting [50]583.360.1328139.8452.5465.51517.85842.8871.910.026823.3721.97
MGMatting* [50]389.150.0727141.8044.2155.74333.42524.9926.180.005615.8114.53
+Instance-wise (Ours)67.270.005842.3543.5249.5017.7759.5921.120.003816.8814.72
Stage 2: Learning with Composited and Weak-guidance Images
Segmentation-training69.240.006957.0846.5654.3614.8865.4220.940.004320.4614.20
Self-training (Ours)58.160.004639.0441.3747.3210.8453.0216.720.003014.6812.02
", + "image_path": "ddaaccbfd20b729cb3aa56cf03eeb413e3479ba94b43473fb31fe99306070031.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 234, + 289, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 289, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 289, + 461 + ], + "type": "text", + "content": "makes substantial random localization errors in guidance and lowers the reliability of mask guidance during training. Training with such guidance makes models ignore the mask guidance and overfit to synthetic samples, resulting in inferior in-the-wild performance. On top of MGMatting*, the \"Instance-wise Learning\" strategy drastically benefits the mask-guided matting model. Unlike the baseline model that segments all the composited objects (Fig. 5-(c)), the proposed strategy allows the model to well discriminate the target object (Fig. 5-(d)). Our model faithfully utilizes the coarse mask guidance and shows a huge gain in localization performance (see " + }, + { + "bbox": [ + 46, + 234, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathrm{SAD}_{BG}" + }, + { + "bbox": [ + 46, + 234, + 289, + 461 + ], + "type": "text", + "content": " metric). Although the strategy assumes multi-object scenarios and learns the complex relations between them, it also contributes to the better matting details for an object (see " + }, + { + "bbox": [ + 46, + 234, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathrm{SAD}_{FG}" + }, + { + "bbox": [ + 46, + 234, + 289, + 461 + ], + "type": "text", + "content": " in Composition-Wild and SAD in AIM). By design, the instance-wise learning naturally improves occlusion handling by learning the order of layered foreground objects and their fractional contributions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 462, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 700 + ], + "type": "text", + "content": "Effectiveness of Self-training. To extend the mask-guided matting model to the in-the-wild setting, we leverage the weak-supervision dataset under the proposed self-training framework. We compare our self-training framework to a baseline, namely \"segmentation-training\". This baseline is motivated by previous human-matting methods [23, 25]. It attempts to learn generalized matting capability via simultaneous segmentation training on natural images. To be specific, the model takes strongly augmented instance masks as guidance and is supervised to recover ground-truth instance masks through the additional segmentation head. As deteriorated performance indicates, the segmentation-training strategy fails to scale up to the in-the-wild setting. While it enhances the localization of target objects, supervision from binary instance masks hinders learning diverse matting patterns in the wild (see " + }, + { + "bbox": [ + 46, + 462, + 289, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{SAD}_{FG}" + }, + { + "bbox": [ + 46, + 462, + 289, + 700 + ], + "type": "text", + "content": " metric), especially for transparent objects. On the contrary, our self-training framework successfully generates the soft-pseudo label and utilizes it as a direct supervision to matting tasks, resulting in huge improvements in all the metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "Analysis on Perturbation in Self-training. To demon-" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 232, + 554, + 310 + ], + "blocks": [ + { + "bbox": [ + 46, + 193, + 546, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 546, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 546, + 213 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on Composition-Wild and AIM-500 [20]. * denotes the improved version of MGMatting [50]. \"Instance-wise\" represents instance-wise learning. A lower score is better in all the metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 232, + 554, + 310 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 554, + 310 + ], + "type": "table", + "html": "
MethodComposition-WildAIM-500
SADMSE\\( SAD_{FG} \\)\\( SAD_{BG} \\)SADMSE
Ours58.160.004647.3210.8416.720.0030
No Image.60.080.004845.1914.8918.570.0039
No Guide.60.140.004947.9512.1918.270.0035
", + "image_path": "23bf6862a4f17113e4fc07af0f26b35eec848c508bab1ca0e17a17b67d6880f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 353, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 424 + ], + "type": "text", + "content": "strate the effectiveness of perturbations in self-training, we ablate the image and guidance perturbations separately. The results are summarized in Table 2. As degraded performance denoted, encouraging robustness against image and guidance perturbations is crucial to the success of the self-training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 426, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 510 + ], + "type": "text", + "content": "We further investigate whether the matting-specific design of image perturbation is necessary. To this end, we train three variants with different configurations of image perturbations: (1) Pixel-level linear augmentation (Pixel), (2) Region-level augmentation (Region) and (3) Both of them (Pixel+Region). Below, we report the SAD metric for Composition-Wild and AIM-500, respectively." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 323, + 518, + 529, + 540 + ], + "blocks": [ + { + "bbox": [ + 314, + 319, + 537, + 331 + ], + "lines": [ + { + "bbox": [ + 314, + 319, + 537, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 537, + 331 + ], + "type": "text", + "content": "Table 2. Ablation Study on Perturbations in Self-training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 518, + 529, + 540 + ], + "lines": [ + { + "bbox": [ + 323, + 518, + 529, + 540 + ], + "spans": [ + { + "bbox": [ + 323, + 518, + 529, + 540 + ], + "type": "table", + "html": "
PixelRegionPixel+Region
58.16 / 16.7269.8 / 25.9661.88 / 19.00
", + "image_path": "067a6aaa2edb0d525533ce28df07c5fcc200d89bcd4e05f560d4b36985b026af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "type": "text", + "content": "We can observe that the region-level augmentation, which is widely used in other tasks [37, 47], disturbs the effects of self-training objectives. As discussed in Sec. 4.2, this is because the region-level augmentation breaks the underlying opacity of objects and make a mismatch between inputs of the student and the teacher networks. On the contrary, pixel-level linear augmentation brings clear improvements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 644, + 446, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 446, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 446, + 658 + ], + "type": "text", + "content": "7. Real-world Applications" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "To show the generality and potential of the in-the-wild mask-guided matting model, we illustrate two new applications: mask-guided video matting and panoptic matting. Fig. 6 and Fig. 7 show examples of each application." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1998" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 157, + 174 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 157, + 174 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 157, + 174 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 157, + 174 + ], + "type": "image", + "image_path": "df35dffa257ae87bf259cec91d4214e5e7450c408112602afa8a2503d958df7a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 175, + 128, + 183 + ], + "lines": [ + { + "bbox": [ + 97, + 175, + 128, + 183 + ], + "spans": [ + { + "bbox": [ + 97, + 175, + 128, + 183 + ], + "type": "text", + "content": "(a) Frame" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 159, + 70, + 217, + 174 + ], + "blocks": [ + { + "bbox": [ + 159, + 70, + 217, + 174 + ], + "lines": [ + { + "bbox": [ + 159, + 70, + 217, + 174 + ], + "spans": [ + { + "bbox": [ + 159, + 70, + 217, + 174 + ], + "type": "image", + "image_path": "1e9de2c76caaac8aa0a266288a8f61d7e1c3ed418a21e73e787fc981f61e7f00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 175, + 225, + 184 + ], + "lines": [ + { + "bbox": [ + 184, + 175, + 225, + 184 + ], + "spans": [ + { + "bbox": [ + 184, + 175, + 225, + 184 + ], + "type": "text", + "content": "(b) Guidance" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 71, + 342, + 174 + ], + "blocks": [ + { + "bbox": [ + 248, + 71, + 342, + 174 + ], + "lines": [ + { + "bbox": [ + 248, + 71, + 342, + 174 + ], + "spans": [ + { + "bbox": [ + 248, + 71, + 342, + 174 + ], + "type": "image", + "image_path": "5cbbd3e6a457d0eb16f53a3859b2898334e3d573f026143dc7a6d4f47941bd6d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 175, + 318, + 184 + ], + "lines": [ + { + "bbox": [ + 270, + 175, + 318, + 184 + ], + "spans": [ + { + "bbox": [ + 270, + 175, + 318, + 184 + ], + "type": "text", + "content": "(c) MGMatting" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 72, + 432, + 174 + ], + "blocks": [ + { + "bbox": [ + 342, + 72, + 432, + 174 + ], + "lines": [ + { + "bbox": [ + 342, + 72, + 432, + 174 + ], + "spans": [ + { + "bbox": [ + 342, + 72, + 432, + 174 + ], + "type": "image", + "image_path": "6531bcc6bed2208c88443aebe8bc15bcbcf08a5c157b8bf46724e41ae4f25859.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 175, + 400, + 184 + ], + "lines": [ + { + "bbox": [ + 373, + 175, + 400, + 184 + ], + "spans": [ + { + "bbox": [ + 373, + 175, + 400, + 184 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 433, + 71, + 524, + 174 + ], + "blocks": [ + { + "bbox": [ + 433, + 71, + 524, + 174 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 524, + 174 + ], + "type": "image", + "image_path": "66ce27abec34c50cf96fa1adf7bc7507939e0e06e27c01b5dd92f06e1bb2c74f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 175, + 505, + 184 + ], + "lines": [ + { + "bbox": [ + 451, + 175, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 451, + 175, + 505, + 184 + ], + "type": "text", + "content": "(e) Editing Result" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 67, + 206, + 249, + 264 + ], + "blocks": [ + { + "bbox": [ + 117, + 189, + 474, + 201 + ], + "lines": [ + { + "bbox": [ + 117, + 189, + 474, + 201 + ], + "spans": [ + { + "bbox": [ + 117, + 189, + 474, + 201 + ], + "type": "text", + "content": "Figure 6. Extension to Mask-guided Video Matting. We use propagated masks [31] as guidance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 206, + 249, + 264 + ], + "lines": [ + { + "bbox": [ + 67, + 206, + 249, + 264 + ], + "spans": [ + { + "bbox": [ + 67, + 206, + 249, + 264 + ], + "type": "image", + "image_path": "1675affb9f2e296a2dbd2395b28fa95cbf03fb58770c4c717f70d8a27156e4e9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 250, + 206, + 341, + 264 + ], + "blocks": [ + { + "bbox": [ + 250, + 206, + 341, + 264 + ], + "lines": [ + { + "bbox": [ + 250, + 206, + 341, + 264 + ], + "spans": [ + { + "bbox": [ + 250, + 206, + 341, + 264 + ], + "type": "image", + "image_path": "806838b63f2647d6f10ef948ea06202462efe355996cb41a7562946d6e6a2302.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 341, + 206, + 432, + 264 + ], + "blocks": [ + { + "bbox": [ + 341, + 206, + 432, + 264 + ], + "lines": [ + { + "bbox": [ + 341, + 206, + 432, + 264 + ], + "spans": [ + { + "bbox": [ + 341, + 206, + 432, + 264 + ], + "type": "image", + "image_path": "201a1a4f25cb8dad2c51e2b1e60071e2971b5bb8fc13d611f56e404bec6df81e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 207, + 524, + 264 + ], + "blocks": [ + { + "bbox": [ + 432, + 207, + 524, + 264 + ], + "lines": [ + { + "bbox": [ + 432, + 207, + 524, + 264 + ], + "spans": [ + { + "bbox": [ + 432, + 207, + 524, + 264 + ], + "type": "image", + "image_path": "6091fd7c8acb78a6215beb65c9a19aebdde2ddc179c98a746eefcda6ebf432ab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 67, + 268, + 159, + 354 + ], + "blocks": [ + { + "bbox": [ + 67, + 268, + 159, + 354 + ], + "lines": [ + { + "bbox": [ + 67, + 268, + 159, + 354 + ], + "spans": [ + { + "bbox": [ + 67, + 268, + 159, + 354 + ], + "type": "image", + "image_path": "9ab6844626e07b606e6391ffeea22475bfd024fbf199e97a090878a54c07d237.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 98, + 356, + 129, + 364 + ], + "lines": [ + { + "bbox": [ + 98, + 356, + 129, + 364 + ], + "spans": [ + { + "bbox": [ + 98, + 356, + 129, + 364 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 159, + 268, + 251, + 354 + ], + "blocks": [ + { + "bbox": [ + 159, + 268, + 251, + 354 + ], + "lines": [ + { + "bbox": [ + 159, + 268, + 251, + 354 + ], + "spans": [ + { + "bbox": [ + 159, + 268, + 251, + 354 + ], + "type": "image", + "image_path": "6c11a8b55ec1aebfc600dd74aa08e4d13f17b4e295cd9317a548edfc3085d545.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 356, + 225, + 364 + ], + "lines": [ + { + "bbox": [ + 184, + 356, + 225, + 364 + ], + "spans": [ + { + "bbox": [ + 184, + 356, + 225, + 364 + ], + "type": "text", + "content": "(b) Guidance" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 251, + 268, + 340, + 354 + ], + "blocks": [ + { + "bbox": [ + 251, + 268, + 340, + 354 + ], + "lines": [ + { + "bbox": [ + 251, + 268, + 340, + 354 + ], + "spans": [ + { + "bbox": [ + 251, + 268, + 340, + 354 + ], + "type": "image", + "image_path": "e87bd9bd5a3d84b551c8f626628476d82b338360267358f46aaa5ec7d99919d0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 356, + 319, + 365 + ], + "lines": [ + { + "bbox": [ + 272, + 356, + 319, + 365 + ], + "spans": [ + { + "bbox": [ + 272, + 356, + 319, + 365 + ], + "type": "text", + "content": "(c) MGMatting" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 341, + 268, + 432, + 354 + ], + "blocks": [ + { + "bbox": [ + 341, + 268, + 432, + 354 + ], + "lines": [ + { + "bbox": [ + 341, + 268, + 432, + 354 + ], + "spans": [ + { + "bbox": [ + 341, + 268, + 432, + 354 + ], + "type": "image", + "image_path": "f5e894778455953555281dbb3978bc6e85ff6bad80304c7da9c69fef306fd2ef.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 356, + 400, + 364 + ], + "lines": [ + { + "bbox": [ + 373, + 356, + 400, + 364 + ], + "spans": [ + { + "bbox": [ + 373, + 356, + 400, + 364 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 371, + 545, + 394 + ], + "lines": [ + { + "bbox": [ + 46, + 371, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 545, + 394 + ], + "type": "text", + "content": "Figure 7. Extension to Panoptic Matting. Hand drawn panoptic masks [24] are utilized as guidance. We illustrate independent regions with different colors. The black color represents ignored regions of labels or predictions." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 432, + 268, + 524, + 354 + ], + "blocks": [ + { + "bbox": [ + 432, + 268, + 524, + 354 + ], + "lines": [ + { + "bbox": [ + 432, + 268, + 524, + 354 + ], + "spans": [ + { + "bbox": [ + 432, + 268, + 524, + 354 + ], + "type": "image", + "image_path": "b4decbd3b062b64b1a670503cd46657166d80d97315fc538694d288346b7f772.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 356, + 505, + 364 + ], + "lines": [ + { + "bbox": [ + 451, + 356, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 451, + 356, + 505, + 364 + ], + "type": "text", + "content": "(e) Editing Result" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 408, + 201, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 201, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 201, + 422 + ], + "type": "text", + "content": "7.1. Mask-guided Video Matting" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 429, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 609 + ], + "type": "text", + "content": "Current video matting solutions [40, 53] require multiple frames of trimap to produce an alpha matte for a video. Therefore, required user interaction is more burdensome in the video domain. With the mask-guided matting model, we explore a more practical scenario where only a single instance mask is given in the first frame. In this challenging setting, we first propagate the given mask to the rest of the video frames using the video object segmentation model [31]. Then, the mask-guided matting model utilizes the propagated mask as guidance and produces alpha matte in a frame-by-frame manner. Our model not only corrects the noise in propagated masks but also predicts reasonable opacity in challenging cases such as motion blur. With the prediction, we also provide visually appealing video editing results (i.e., background replacement)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 621, + 151, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 151, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 151, + 634 + ], + "type": "text", + "content": "7.2. Panoptic Matting" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Panoptic segmentation aims to parse an image into nonoverlapping regions of stuff and things. By the definition of the task format, the soft transition between the regions is overlooked. We explore whether our mask-guided matting model could provide such understanding from panoptic segmentation results. Given a panoptic mask, our model con" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 409, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 517 + ], + "type": "text", + "content": "siders the segment of each region separately and produces the corresponding alpha matte. We form panoptic matting results by aggregating the predictions. Unlike the baseline, our model successfully refines the fine details and captures the soft transition between the regions. Interestingly, while the model is trained on objects (i.e., things), our method is well generalized on the stuff regions (See sky region in Fig. 7). Panoptic matting enables us to do region-level image editing and the qualitative results are shown in Fig. 7." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 546, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 677 + ], + "type": "text", + "content": "In this paper, we propose a simple yet effective learning framework for mask-guided matting in the wild. We first introduce instance-wise learning to learn generalized matting ability from the composited images. Then, the model is further extended to handle diverse objects under the self-training on weak-guidance images. Extensive experiments on the newly proposed evaluation benchmark provide an in-depth understanding of the unique challenges. Using our approach, we instantiate new practical applications: mask-guided video matting and panoptic matting. We hope our proposals inspire future work on mask-guided matting." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement This work was partially supported by the NRF (NRF-2020M3H8A1115028, FY2021) and Samsung Electronics Co., Ltd (G01200447)." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1999" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "text", + "content": "[1] Tamara L Berg and Alexander C Berg. Finding iconic images. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, pages 1-8. IEEE, 2009. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 138, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 138, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 138, + 288, + 194 + ], + "type": "text", + "content": "[2] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 288, + 239 + ], + "type": "text", + "content": "[3] Quan Chen, Tiezheng Ge, Yanyu Xu, Zhiqiang Zhang, Xinxin Yang, and Kun Gai. Semantic human matting. In Proceedings of the 26th ACM international conference on Multimedia, pages 618-626, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 241, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 274 + ], + "type": "text", + "content": "[4] Qifeng Chen, Dingzeyu Li, and Chi-Keung Tang. Knn matting. IEEE transactions on pattern analysis and machine intelligence, 35(9):2175-2188, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 331 + ], + "type": "text", + "content": "[5] Yung-Yu Chuang, Brian Curless, David H Salesin, and Richard Szeliski. A bayesian approach to digital matting. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 2, pages II-II. IEEE, 2001. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 333, + 288, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 333, + 288, + 377 + ], + "spans": [ + { + "bbox": [ + 53, + 333, + 288, + 377 + ], + "type": "text", + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 379, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 379, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 379, + 288, + 402 + ], + "type": "text", + "content": "[7] Marco Forte and François Pitie. " + }, + { + "bbox": [ + 53, + 379, + 288, + 402 + ], + "type": "inline_equation", + "content": "f, b" + }, + { + "bbox": [ + 53, + 379, + 288, + 402 + ], + "type": "text", + "content": ", alpha matting. arXiv preprint arXiv:2003.07711, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 437 + ], + "type": "text", + "content": "[8] Eduardo SL Gastal and Manuel M Oliveira. Shared sampling for real-time alpha matting. In Computer Graphics Forum, volume 29, pages 575-584. Wiley Online Library, 2010. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 438, + 288, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 438, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 288, + 493 + ], + "type": "text", + "content": "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 495, + 287, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 287, + 538 + ], + "type": "text", + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 541, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 541, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 541, + 287, + 575 + ], + "type": "text", + "content": "[11] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaouu Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "text", + "content": "[12] Kaiming He, Jian Sun, and Xiaou Tang. Fast matting using large kernel matting laplacian matrices. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2165-2172. IEEE, 2010. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "type": "text", + "content": "[13] Qiqi Hou and Feng Liu. Context-aware image matting for simultaneous foreground and alpha estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4130-4139, 2019. 1, 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[14] Ronghang Hu, Piotr Dólar, Kaiming He, Trevor Darrell, and Ross Girshick. Learning to segment every thing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4233-4241, 2018. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson W.H. Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In AAAI, 2022. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "text", + "content": "[16] Philip Lee and Ying Wu. Nonlocal matting. In CVPR 2011, pages 2193-2200. IEEE, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "type": "text", + "content": "[17] Anat Levin, Dani Lischinski, and Yair Weiss. A closed-form solution to natural image matting. IEEE transactions on pattern analysis and machine intelligence, 30(2):228-242, 2007. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "text", + "content": "[18] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[19] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "type": "text", + "content": "[20] Jizhizi Li, Jing Zhang, and Dacheng Tao. Deep automatic natural image matting. arXiv preprint arXiv:2107.07235, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "text", + "content": "[21] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11450-11457, 2020. 1, 2, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 328, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 328, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 328, + 545, + 382 + ], + "type": "text", + "content": "[22] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "text", + "content": "[23] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "text", + "content": "[25] Jinlin Liu, Yuan Yao, Wendi Hou, Miaomiao Cui, Xuansong Xie, Changshui Zhang, and Xian-sheng Hua. Boosting semantic human matting with coarse annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8572, 2020. 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "text", + "content": "[26] Yuhao Liu, Jiake Xie, Xiao Shi, Yu Qiao, Yujie Huang, Yong Tang, and Xin Yang. Tripartite information mining and integration for image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7555-7564, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "type": "text", + "content": "[27] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3266-3275, 2019. 1, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[28] Sebastian Lutz, Konstantinos Amplanitis, and Aljosa Smolic. Alphagan: Generative adversarial networks for natural image matting. arXiv preprint arXiv:1807.10088, 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Stephen Palmer. Canonical perspective and the perception of objects. Attention and performance, pages 135-151, 1981. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2000" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11696-11706, 2022. 1, 2, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[31] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Per-clip video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1352–1361, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 239 + ], + "type": "text", + "content": "[32] Yu Qiao, Yuhao Liu, Xin Yang, Dongsheng Zhou, Mingliang Xu, Qiang Zhang, and Xiaopeng Wei. Attention-guided hierarchical structure aggregation for image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13676-13685, 2020. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "type": "text", + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 294, + 288, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 349 + ], + "type": "text", + "content": "[34] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "text", + "content": "[35] Ehsan Shahrian, Deepu Rajan, Brian Price, and Scott Cohen. Improving image matting using comprehensive sampling sets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 636-643, 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 470 + ], + "type": "text", + "content": "[36] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[37] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 514, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 548 + ], + "type": "text", + "content": "[38] Jian Sun, Jiaya Jia, Chi-Keung Tang, and Heung-Yeung Shum. Poisson matting. In ACM SIGGRAPH 2004 Papers, pages 315–321. 2004. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "text", + "content": "[39] Yanan Sun, Chi-Keung Tang, and Yu-Wing Tai. Semantic image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11120–11129, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 636 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 636 + ], + "type": "text", + "content": "[40] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 637, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 288, + 691 + ], + "type": "text", + "content": "[41] Jingwei Tang, Yagiz Aksoy, Cengiz Oztireli, Markus Gross, and Tunc Ozan Aydin. Learning-based sampling for natural image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3055–3063, 2019. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[42] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 698 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[43] Jue Wang and Michael F Cohen. An iterative optimization approach for unified image segmentation and matting. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 936-943. IEEE, 2005. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[44] Jue Wang and Michael F Cohen. Optimized color sampling for robust matting. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 251 + ], + "type": "text", + "content": "[45] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Hanqing Zhao, Weiming Zhang, and Nenghai Yu. Improved image matting via real-time user clicks and uncertainty estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15374-15383, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 319 + ], + "type": "text", + "content": "[46] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXV, pages 238–258. Springer, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 319, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 374 + ], + "type": "text", + "content": "[47] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060-3069, 2021. 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 374, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 374, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 374, + 545, + 419 + ], + "type": "text", + "content": "[48] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "type": "text", + "content": "[49] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "text", + "content": "[50] Qihang Yu, Jianming Zhang, He Zhang, Yilin Wang, Zhe Lin, Ning Xu, Yutong Bai, and Alan Yuille. Mask guided matting via progressive refinement network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1154-1163, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "type": "text", + "content": "[51] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 586, + 545, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 545, + 642 + ], + "type": "text", + "content": "[52] Yunke Zhang, Lixue Gong, Lubin Fan, Peiran Ren, Qixing Huang, Hujun Bao, and Weiwei Xu. A late fusion cnn for digital matting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7469-7478, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 643, + 545, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 643, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 307, + 643, + 545, + 698 + ], + "type": "text", + "content": "[53] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 6, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2001" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_content_list.json b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c91c2dec2611fb4482b73e79555ff913a7c54c7 --- /dev/null +++ b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_content_list.json @@ -0,0 +1,1506 @@ +[ + { + "type": "text", + "text": "Mask3D: Pre-training 2D Vision Transformers by Learning Masked 3D Priors", + "text_level": 1, + "bbox": [ + 89, + 130, + 879, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ji Hou $^{1}$ Xiaoliang Dai $^{1}$ Zijian He $^{1}$ Angela Dai $^{2}$ Matthias Nießner $^{2}$", + "bbox": [ + 179, + 179, + 784, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Meta Reality Labs $^{2}$ Technical University of Munich", + "bbox": [ + 263, + 204, + 700, + 223 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/294f8ce22f3c51be1180ed92dff051729bcb6668902d19a81160b7e7e1602c6c.jpg", + "image_caption": [ + "Pre-text Task" + ], + "image_footnote": [], + "bbox": [ + 84, + 266, + 323, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/babdccce19462cd31a941615100d97b2a4e192b7d2874d2f2a3b4893870d21dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 267, + 514, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30246612520851c198dfb1d75ca6882d1aa6aaac19c3384f02eceef7a70f8597.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 252, + 692, + 382 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0e998b87a3f7e0889f2e41d6b518702c5be306ded28fa007fcf31779d30086a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 266, + 885, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14e45ed55fdb4020a244be26d6ba9c967f0d58df86546eccae2801124c2016a6.jpg", + "image_caption": [ + "Scene Understanding Tasks" + ], + "image_footnote": [], + "bbox": [ + 84, + 391, + 276, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b166da63fcf88b310c38ae24a402d3fc0b067e17ac3e5890ae6d1a5bab0afd5.jpg", + "image_caption": [ + "Figure 1. We present Mask3D, which learns to embed 3D priors to 2D representations for image understanding tasks, based on a self-supervised pre-training formulation from single RGB-D views, without requiring any camera pose or multi-view correspondence information. Our pre-training takes masked RGB and depth patches as input to reconstruct the dense depth map, and the pre-trained color backbone is used to fine-tune various downstream image understanding tasks. This results in effective ViT pre-training for a variety of downstream tasks and datasets." + ], + "image_footnote": [], + "bbox": [ + 282, + 391, + 313, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/756f1963bd8c1c790262dc066e3f5adf68dee1fd1a180c4212b19baa97cb4729.jpg", + "image_caption": [ + "Instance Segmentation" + ], + "image_footnote": [], + "bbox": [ + 325, + 392, + 509, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21fa2419c37ca942862ff3b050beba1500a6d796624183411b868b977efa95d9.jpg", + "image_caption": [ + "Semantic Segmentation" + ], + "image_footnote": [], + "bbox": [ + 511, + 392, + 694, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c8a172a3aece2f73bbd956519ad499ab652a31138813cec0336ef3f6cce6eae7.jpg", + "image_caption": [ + "Object Detection" + ], + "image_footnote": [], + "bbox": [ + 696, + 392, + 885, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 619, + 312, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current popular backbones in computer vision, such as Vision Transformers (ViT) and ResNets are trained to perceive the world from 2D images. However, to more effectively understand 3D structural priors in 2D backbones, we propose Mask3D to leverage existing large-scale RGB-D data in a self-supervised pre-training to embed these 3D priors into 2D learned feature representations. In contrast to traditional 3D contrastive learning paradigms requiring 3D reconstructions or multi-view correspondences, our approach is simple: we formulate a pre-text reconstruction task by masking RGB and depth patches in individual RGB-D frames. We demonstrate the Mask3D is particularly effective in embedding 3D priors into the powerful 2D ViT backbone, enabling improved representation learning for various scene understanding tasks, such as semantic segmentation, instance segmentation and object detection.", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Experiments show that Mask3D notably outperforms existing self-supervised 3D pre-training approaches on ScanNet, NYUv2, and Cityscapes image understanding tasks, with an improvement of $+6.5\\%$ mIoU against the state-of-the-art Pri3D on ScanNet image semantic segmentation.", + "bbox": [ + 500, + 622, + 890, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 724, + 630, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have seen remarkable advances in 2D image understanding as well as 3D scene understanding, although their representation learning has generally been treated separately. Powerful 2D architectures such as ResNets [21] and Vision Transformers (ViT) [14] have achieved notable success in various 2D recognition and segmentation tasks, but focus on learning from 2D image data. Current large-scale RGB-D datasets [1,4,10,34,35] provide an opportunity to learn key geometric and structural priors to provide more informed reasoning about the scale and cir", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13510", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cumvent view-dependent effects, which can provide more efficient representation learning. In 3D, various successful methods have been leveraging the RGB-D datasets for constructive point discrimination [6,24,39,44] for downstream 3D tasks, including high-level scene understanding tasks as well as low-level point matching tasks [15,42,43]. However, the other direction from 3D to 2D is less explored.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We thus aim to embed such 3D priors into 2D backbones to effectively learn the structural and geometric priors underlying the 3D scenes captured in 2D image projections. Recently, Pri3D [25] adopted similar multi-view and reconstruction-based constraints to induce 3D priors in learned 2D representations. However, this relies on not only acquiring RGB-D frame data but also the robust registration of multiple views to obtain camera pose information for each frame. Instead, we consider how to effectively learn such geometric priors from only single-view RGB-D data in a more broadly applicable setting for 3D-based pre-training.", + "bbox": [ + 75, + 198, + 467, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We thus propose Mask3D, which learns effective 3D priors for 2D backbones in a self-supervised fashion by pretraining with single-view RGB-D frame data. We propose a pre-text reconstruction task to reconstruct the depth map by masking different random RGB and depth patches of an input frame. These masked input RGB and depth are encoded simultaneously in separate encoding branches and decoded to reconstruct the dense depth map. This imbues 3D priors into the RGB backbone which can then be used for fine-tuning downstream image based scene understanding tasks.", + "bbox": [ + 75, + 363, + 467, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, our self-supervised approach to embedding 3D priors from single-view RGB-D data to 2D learned features is not only more generally applicable, but we also demonstrate that it is particularly effective for pre-training vision transformers. Our experiments demonstrate the effectiveness of Mask3D on a variety of datasets and image understanding tasks. We pre-train on ScanNet [10] with our masked 3D pre-training paradigm and fine-tune for 2D semantic segmentation, instance segmentation, and object detection. This enables notable improvements not only on ScanNet data but also generalizes to NYUv2 [34] and even Cityscapes [8] data. We believe that Mask3D makes an important step to shed light on the paradigm of incorporating 3D representation learning to powerful 2D backbones.", + "bbox": [ + 75, + 513, + 467, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 96, + 724, + 326, + 739 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a self-supervised pre-training approach to learn masked 3D priors for 2D image understanding tasks based on learning from only single-view RGB-D data, without requiring any camera pose or 3D reconstruction information, and thus enabling more general applicability.", + "- We demonstrate that our masked depth reconstruction pre-training is particularly effective for the modern, powerful ViT architecture, across a variety of datasets and image understanding tasks." + ], + "bbox": [ + 94, + 741, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-training in Visual Transformers. Recently, visual transformers have revolutionized computer vision and attracted wide attention. In contrast to popular CNNs that operate in a sliding window fashion, Vision Transformers (ViT) describe the image as patches of 16x16 pixels. The Swin Trasnformer [28] has set new records with its hierarchical transformer formulation on major vision benchmarks. The dominance of visual transformers in many vision tasks has inspired study into how to pre-training such backbones. MoCoV3 [5] first investigated the effects of several fundamental components for self-supervised ViT training. MAE [19] then proposed an approach inspired by BERT [13], which randomly masks words in sentences and leveraged masked image reconstruction for self-supervised pre-training that achieved state-of-the-art results in ViT. A similar self-supervision has also been proposed by MaskFeat [37] for self-supervised video pre-training. MaskFeat randomly masks out pixels of the input sequence and then predicts the Oriented Gradients (HOG) of the masked regions. However, such ViT pre-training methods focus on image or video data, without exploring how 3D priors can potentially be exploited. MultiMAE [2] on the other hand introduces depth priors. However, it requires depth as input not only in pre-training but also in downstream tasks. In addition to depth, human annotations (e.g., semantics) are also leveraged in the pre-training. To achieve a self-supervised pre-training, we do not use semantics in the pre-training and only use RGB images as input in downstream tasks.", + "bbox": [ + 496, + 114, + 890, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RGB-D Scene Understanding. Research in 3D scene understanding have been spurred forward with the introduction of larger-scale, annotated real-world RGB-D datasets [1, 4, 10]. This has enabled data-driven semantic understanding of 3D reconstructed environments, where we have now seen notable progress, such as for 3D semantic segmentation [7, 11, 17, 31, 32, 36], object detection [29, 30, 45], instance segmentation [16, 18, 22, 23, 26, 27, 40, 41], and recently panoptic segmentation [9]. Such 3D scene understanding tasks have been analogously defined to 2D image understanding, which considers RGB-only input without depth information. However, learning from 3D enables geometric reasoning without requiring learning view-dependent effects or resolving depth/scale ambiguity that must be learned when considering solely 2D data. We thus take advantage of existing large-scale RGB-D data to explore how to effectively embed 3D priors for better representation learning for 2D scene understanding tasks.", + "bbox": [ + 496, + 545, + 890, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Embedding 3D Priors in 2D Backbones. Learning cross-modality features has been seen in extensive studies of the ties between languages and images. In particular, CLIP [33] learns visual features from language supervision during pre-training, showing promising results in zero-", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13511", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shot learning for image classification. Pri3D [25] explores 3D-based pre-training for image-based tasks by leveraging multi-view consistency and 2D-3D correspondence with contrastive learning to embed 3D priors into ResNet backbones. This results in enhanced features over ImageNet pretraining on 2D scene understanding tasks. However, Pri3D requires camera pose registration across RGB-D video sequences and is specifically designed for CNNs-based architectures. In contrast, we formulate a self-supervised pretraining that operates on only single-view RGB-D frames and leverages masked 3D priors that can effectively pretrain powerful ViT backbones.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 284, + 169, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce Mask3D to embed 3D priors into learned 2D representations by self-supervised pre-training from only single-view RGB-D frames. To effectively learn 3D structural priors without requiring any camera pose information or multi-view constraints, we formulate a pre-text depth reconstruction task to inform the RGB feature extraction to be geometrically aware. Randomly masked color and depth images are used as input to reconstruct the dense depth map, and the RGB backbone can then be used to fine-tune downstream image understanding tasks. In particular, we show in Sec. 4 that this single-frame self-supervision is particularly well-suited for powerful vision transformer (ViT) backbones, even without any multi-view information.", + "bbox": [ + 75, + 310, + 472, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Learning Masked 3D Priors", + "text_level": 1, + "bbox": [ + 76, + 513, + 328, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose to learn masked 3D priors to embed to learned 2D backbones by pre-training to reconstruct dense depth from RGB images with the guidance of sparse depth. That is, for an RGB-D frame $F = (C, D)$ with RGB image $C$ and depth map $D$ , we train to reconstruct $D$ from masked patches of $C$ guided with sparse masked patches of $D$ . An overview of our approach is shown in Fig. 2.", + "bbox": [ + 75, + 537, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To create masked color and depth $M_{c}$ and $M_{d}$ from $C$ and $D$ as input for reconstruction, a 240x320 RGB image $C$ is uniformly divided into 300 16x16 patches, from which we randomly keep a percentage $p_{c}$ of patches, masking out the others, to obtain $M_{c}$ . $M_{d}$ is created similarly by keeping only a percentage $p_{d}$ of patches, such that the resulting depth patches do not coincide with the RGB patches in $M_{c}$ .", + "bbox": [ + 75, + 643, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then train color and depth encoders $\\Psi_{c}$ and $\\Psi_{d}$ to separately encode RGB and depth signals. RGB patches are fed into $\\Psi_{c}$ and concatenated with a positional embedding, following the ViT architecture, and similarly for depth. The positional embedding used encodes the patch location by a cosine function. Patches and their positional embeddings are then mapped into higher dimensional feature vectors via $\\Psi_{c}$ and $\\Psi_{d}$ . The encoders $\\Psi_{c}$ and $\\Psi_{d}$ are built by blocks composed of linear and norm layers. The features from $\\Psi_{c}$ and $\\Psi_{d}$ are then fused in the bottleneck; since depth patches", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "were selected in regions where no RGB patches were selected, there are no duplicate patches representing the same patch location.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For those regions which do not have any associated RGB or depth patch, we use patches of constant values as mask tokens to create a placeholder in the bottleneck to enable reconstructing dense depth at the original image resolution. In the bottleneck, the RGB and depth patch feature vectors, along with the mask tokens, form the input to the decoder. This formulates a reconstruction task from sparse RGB and depth; the joint RGB-D pre-training enables reconstruction from very sparse input, as shown by our ablation on the masked input rations in Sec. 4.5. Note that the depth encoder is trained only during pre-training, and only the color ViT encoder (and decoder, if applicable) are used for downstream fine-tuning.", + "bbox": [ + 496, + 136, + 892, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the pre-training task, we demonstrate the depth completion results from the pretraining phase in Fig. 6. A detailed analysis of masking different ratios of color and depth signals is shown in Sec. 4.5.", + "bbox": [ + 496, + 333, + 893, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-training Loss In contrast to the widely used contrastive loss in 3D representation learning, we train for dense depth reconstruction with an $\\ell_2$ reconstruction loss. Similar to MAE [19], we normalize the output patches as well as the target patches prior to computing the loss, which we found to empirically improve the performance.", + "bbox": [ + 496, + 396, + 893, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 501, + 586, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of Mask3D pretraining for ViT [14] backbones on semantic segmentation, instance segmentation, and object detection tasks. We pre-train on ScanNet [10] data and demonstrate the effectiveness of learned masked 3D priors for not only ScanNet downstream tasks but also their transferability to NYUv2 [34] and even across the indoor/outdoor domain gap to Cityscapes [8] data.", + "bbox": [ + 496, + 526, + 890, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 657, + 689, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the pre-training and finetuning procedures. Our method uses a two-stage pretraining design introduced in the following.", + "bbox": [ + 496, + 681, + 890, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage-I: Mask3D Encoder Initialization. We initialize the RGB encoder with network weights trained on ImageNet [12] (as pre-training for our pre-training). To maintain a fully self-supervised pre-training paradigm, we initialize with weights obtained by self-supervised ImageNet pre-training [19].", + "bbox": [ + 496, + 737, + 890, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage-II: Mask3D Pre-training on ScanNet. Mask3D pre-training leverages 3D priors in RGB-D frame data, for which we use the color and depth maps of ScanNet [10]. Note that this does not use any semantic or reconstruction", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13512", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dc8cb09a518a2067cae2580897a845ac376e471c53f11e00667b29b7af8f27c8.jpg", + "image_caption": [ + "Figure 2. Overview of Mask3D Pre-training. As a pretext task, we predict dense depth from color and sparse depth signals. We use masked input by randomly selecting a set of patches from the input color image, which are then mapped to higher dimensional feature vectors; input depth is similarly masked and encoded. The color and depth features are then fused into a bottleneck from which dense depth is reconstructed as a self-supervised loss." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 890, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "information during pre-training. ScanNet contains 2.5M RGB-D frames from 1513 ScanNet train video sequences. We regularly sample every $25^{\\text{th}}$ frame without any other filtering (e.g., no control on viewpoint variation).", + "bbox": [ + 75, + 407, + 470, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Downstream Fine-tuning. We evaluate our Mask3D pretraining by fine-tuning a variety of downstream image understanding tasks (semantic segmentation, instance segmentation, object detection). We consider in-domain transfers on ScanNet image understanding, and further evaluate the out-of-domain transfer on datasets with different statistical characteristics: the indoor image data of NYUv2 [34], as well as across a strong domain gap to the outdoor image data of Cityscapes [8]. For semantic segmentation tasks, we use both encoder and decoder pre-trained with Mask3D, and for instance segmentation and detection, only the backbone encoder is pre-trained.", + "bbox": [ + 75, + 481, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines. To evaluate the effectiveness of our learned masked 3D priors for 2D representations, we benchmark our method against relevant baselines:", + "bbox": [ + 75, + 672, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervised ImageNet Pre-training (supIN). This uses the pre-trained weights from ImageNet, provided by torchvision, as is commonly used for image understanding tasks. Here, only ImageNet data is used, and no ScanNet data is involved in the pre-training phase.", + "bbox": [ + 75, + 718, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2-Stage MoCoV2 (MoCoV2-supIN $\\rightarrow$ SN). Supervised ImageNet pre-trained (supIN) weights are used as network initialization for pre-training. MoCoV2 [20] is used for pre-training with randomly shuffled ScanNet images. Here, both ImageNet and ScanNet image data are used without any geometric priors.", + "bbox": [ + 75, + 794, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2-Stage MAE (MAE-unsupIN $\\rightarrow$ SN). Self-supervised Im", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ageNet pre-trained weights are used as network initialization for pre-training. MAE [19] is used for pre-training with randomly shuffled ScanNet images. Here, both ImageNet and ScanNet image data are used without any geometric priors.", + "bbox": [ + 496, + 409, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pri3D [25]. Supervised ImageNet pre-trained are used to initialize Pri3D pre-training, which leverages multi-view and reconstruction constraints from ScanNet data under a contrastive loss. Here, both ImageNet and ScanNet data are used, incorporating 3D priors from reconstructed RGB-D video sequences for pre-training.", + "bbox": [ + 496, + 484, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation Details. We use a ViT-B backbone to train our approach. For pre-training, we use an SGD optimizer with a learning rate of 0.1 and an effective batch size of 128 (accumulated gradients from an actual batch size of 64). The learning rate is decreased by a factor of 0.99 every 1000 steps, and our method is trained for 100 epochs. Fine-tuning on semantic segmentation is trained with a batch size of 8 for 80 epochs. The initial learning rate is 0.01, with polynomial decay with a power of 0.9. Finetuning on detection and instance segmentation is trained using Detector2 [38] with the 1x schedule. Pre-training experiments are conducted on a single NVIDIA A6000 GPU, or 2 NVIDIA RTX3090 GPUs, or 4 NVIDIA RTX2080Ti GPUs; semantic segmentation experiments are conducted on a single NVIDIA A6000 GPU; instance segmentation and detection experiments are conducted on 8 V100 GPUs.", + "bbox": [ + 496, + 580, + 892, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. ScanNet Downstream Tasks", + "text_level": 1, + "bbox": [ + 500, + 832, + 750, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of representation learning with 3D priors via downstream tasks on ScanNet [10] images. For fine-tuning, we follow the standard", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/62e7cf64421ce4c3177570038e0d429b7c461672494753fe6e0b7150375b77b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-training MethodBackbonePre-training DatamIoU
ScratchResNet-50None39.1
ImageNet Pre-training (supIN)ResNet-50ImageNet55.7
Supervised Pre-trainingViTImageNet+ScanNet65.9 (+10.2)
MoCoV2-supIN→SN [20]ResNet-50ImageNet+ScanNet56.6 (+0.9)
Pri3D [25]ResNet-50ImageNet+ScanNet60.2 (+4.5)
Pri3D [25]ViTImageNet+ScanNet59.3 (+3.6)
DINO [3]ViTImageNet+ScanNet58.1 (+3.6)
MAE-unsupIN→SN [19]ViTImageNet+ScanNet63.3 (+7.6)
Ours - Mask3D (DINO)ViTImageNet+ScanNet60.5 (+4.8)
Ours - Mask3D (MAE)ViTImageNet+ScanNet66.7 (+11.0)
", + "bbox": [ + 165, + 88, + 802, + 244 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7c0da3f62e32de6c9dd1d5054ca626bce7090efdb676a58f74bfa2f742a816dc.jpg", + "table_caption": [ + "Table 1. ScanNet 2D Semantic Segmentation. Mask3D significantly outperforms Pri3D as well as other state-of-the-art pre-training approaches that leverage both ImageNet and ScanNet data." + ], + "table_footnote": [], + "table_body": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch32.717.716.9
ImageNet Pretrain (supIN)41.725.925.1
MoCoV2-supIN→SN [20]43.5 (+1.8)26.8 (+0.9)25.8 (+0.7)
Pri3D [25]43.7 (+2.0)27.0 (+1.1)26.3 (+1.2)
MAE-unsupIN→SN [19]46.1 (+4.4)32.7 (+6.8)30.5 (+5.4)
Mask3D (Ours)50.4 (+8.7)35.3 (+9.4)32.7 (+7.6)
", + "bbox": [ + 78, + 295, + 478, + 398 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/16f72ee8d05e74abdc1c64fd06ea2c78c10565350809664a76b78cfa449968a0.jpg", + "table_caption": [ + "Table 2. ScanNet 2D Object Detection. Fine-tuning with Mask3D pre-trained models leads to improved object detection results across different metrics, in comparison to ImageNet pretraining, MoCo-style pre-training, and a strong MAE-style pre-training method." + ], + "table_footnote": [], + "table_body": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch25.813.112.2
ImageNet Pretrain (supIN)32.617.817.6
MoCoV2-supIN→SN [20]33.9 (+1.3)18.1 (+0.3)18.3 (+0.7)
Pri3D [25]34.3 (+1.7)18.7 (+0.9)18.3 (+0.7)
MAE-unsupIN→SN [19]37.4 (+4.8)20.3 (+2.5)20.7 (+3.1)
Mask3D (Ours)41.2 (+8.6)22.7 (+4.9)22.8 (+5.2)
", + "bbox": [ + 78, + 487, + 480, + 590 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ae074d9b53aa3fd19819ba96a4c9dd4f58580f8816a7f819c600008fea7b19bc.jpg", + "table_caption": [ + "Table 3. ScanNet 2D Instance Segmentation. Fine-tuning with Mask3D pre-trained models leads to improved instance segmentation results across different metrics compared to ImageNet pre-training, MoCo-style pre-training, and a strong MAE-style pre-training method." + ], + "table_footnote": [], + "table_body": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch17.29.28.8
ImageNet Pretrain (supIN)25.113.913.4
MoCoV2-supIN→SN [20]27.2 (+2.1)14.7 (+0.2)14.8 (+1.4)
Pri3D [25]28.1 (+3.0)15.7 (+1.8)15.7 (+2.3)
MAE unsupIN→SN [19]33.6 (+8.5)19.0 (+5.1)19.0 (+5.6)
Mask3D (Ours)37.0 (+11.9)21.6 (+7.7)21.3 (+7.9)
", + "bbox": [ + 78, + 679, + 480, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4. NYUv2 2D Instance Segmentation. Fine-tuning with Mask3D pre-trained models leads to improved instance segmentation results across different metrics compared to previous methods, demonstrating the cross-dataset transfer ability of Mask3D.", + "bbox": [ + 75, + 787, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "protocol of the ScanNet benchmark [10] and sample every $100^{\\mathrm{th}}$ frame, resulting in 20,000 train images and 5,000 validation images.", + "bbox": [ + 75, + 864, + 468, + 912 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2D Semantic Segmentation. Tab. 1 shows the fine-tuning for semantic segmentation, in comparison with baseline pre-training approaches. All pre-training methods significantly improve performance over training the semantic segmentation model from scratch. In particular, Mask3D provides substantially better representation quality leading to a much stronger improvement over supervised ImageNet pretraining $(+11\\mathrm{mIoU})$ , and notably improving over MAE unsupIN $\\rightarrow$ SN with ImageNet and ScanNet $(+3.4\\mathrm{mIoU})$ and the 3D-based pre-training of Pri3D $(+6.5\\mathrm{mIoU})$ . We note that the multi-view 3D pre-training of Pri3D does not effectively embed informative 3D priors to ViT backbones, rather suffering from performance degradation from a ResNet-50 backbone. In contrast, our Mask3D pretraining can notably improve performance with a ViT backbone, indicating the effectiveness of our learned 3D priors.", + "bbox": [ + 496, + 297, + 893, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2D Object Detection and Instance Segmentation We show that Mask3D provides effective general 3D priors for a variety of image-based tasks, by evaluating downstream object detection and instance segmentation in Tab. 2 and Tab. 3, respectively. Across all tasks, various pre-training approaches yield substantial improvement over training from scratch. Our masked 3D prior learning transfers effectively learned representations for object detection and instance segmentation, notably improving over the best-performing MAE-unsupIN $\\rightarrow$ SN (+4.3 AP@0.5 and +3.8 AP@0.5, respectively).", + "bbox": [ + 496, + 545, + 893, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data-Efficient Scenarios. We additionally show that our single-view RGB-D pre-training to embed 3D priors in limited data scenarios for downstream ScanNet semantic segmentation in Fig. 5. Mask3D shows consistent improvements across a range of limited data; even with only $20\\%$ of the training data, we recover $80\\%$ performance with $100\\%$ training data available and improving $+15.2$ mIoU over Pri3D pre-training on a ViT backbone.", + "bbox": [ + 496, + 717, + 893, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. NYUv2 Downstream Tasks", + "text_level": 1, + "bbox": [ + 500, + 847, + 743, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate the generalizability of our 3D-imbued learned feature representations across datasets, us", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13514", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3fe270dc8c44a7b29c5a15ad45e7c0f0efbb3a2e532d13043902dff2d5200c37.jpg", + "image_caption": [ + "Figure 3. Qualitative Results on Various Tasks across Different Benchmarks. We visualize predictions on different tasks across various scene understanding benchmarks. From top to bottom rows: instance segmentation on ScanNet, instance segmentation on NYUv2, semantic segmentation on NYUv2, and semantic segmentation results in ScanNet." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 869, + 444 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab080659c103eb936871270160b0774b951d7aeab8fa96ac1d99dc8dbecb50b1.jpg", + "image_caption": [ + "Figure 4. More Qualitative Results on Semantic Segmentation. We visualize semantic segmentation predictions on various scene understanding benchmarks including ScanNet and NYUv2." + ], + "image_footnote": [], + "bbox": [ + 84, + 500, + 890, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing Mask3D pre-trained on ScanNet and fine-tuned on NYUv2 [34] following the same fine-tuning setup as before. NYUv2 contains Microsoft Kinect RGB-D video", + "bbox": [ + 75, + 840, + 472, + 887 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sequences of indoor scenes, comprising 1449 densely labeled RGB images. We use the official 795/654 train/val split. Tables 5, 6, and 4 evaluate the downstream tasks", + "bbox": [ + 498, + 840, + 893, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13515", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ce03959531e20140c76015beae5396c6c611504af19958f91c251e8235a0c059.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-training MethodBackbonePre-training DatamIoU
ScratchResNet-50None24.8
ImageNet Pre-training (supIN)ResNet-50ImageNet50.0
Supervised Pre-trainingViTImageNet+ScanNet55.5 (+5.5)
MoCoV2-supIN→SN [20]ResNet-50ImageNet+ScanNet47.6 (-2.4)
Pri3D [25]ResNet-50ImageNet+ScanNet54.2 (+4.2)
Pri3D [25]ViTImageNet+ScanNet53.2 (+3.2)
MAE-unsupIN→SN [19]ViTImageNet+ScanNet54.9 (+4.9)
Mask3D (Ours)ViTImageNet+ScanNet56.9 (+6.9)
", + "bbox": [ + 200, + 88, + 767, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b2a353062552324ebf8eb7a23902eb9505ec57870a964a21edd760623e9241a.jpg", + "image_caption": [ + "Figure 5. Data-Efficient Results. Compared to previous methods, Mask3D demonstrates consistent improvements on ScanNet 2D semantic segmentation across a range of limited data scenarios. Mask3D is particularly effective for ViT pre-training, improving $+15.2\\%$ mIoU over state-of-the-art Pri3D [25] on a ViT backbone at $20\\%$ of the training data." + ], + "image_footnote": [], + "bbox": [ + 86, + 267, + 460, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d6b653bfa166aabb43e0e7a690999f9ac5a46445a705f245687dd1af260eca6a.jpg", + "table_caption": [ + "Table 5. NYUv2 2D Semantic Segmentation. Mask3D significantly outperforms state-of-the-art pre-training approaches, demonstrating its effectiveness in transferring to different dataset characteristics." + ], + "table_footnote": [], + "table_body": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch21.310.39.0
ImageNet Pretrain (supIN)29.917.316.8
MoCoV2-supIN→SN [20]30.1 (+0.20)18.1 (+0.80)17.3 (+0.50)
Pri3D [25]33.0 (+2.10)19.8 (+2.60)18.9 (+2.10)
MAE-unsupIN→SN [19]40.3 (+10.4)24.5 (+7.20)23.2 (+6.40)
Mask3D (Ours)44.0 (+14.1)28.3 (+6.40)25.9 (+9.10)
", + "bbox": [ + 78, + 595, + 485, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of 2D semantic segmentation, object detection, and instance segmentation, respectively. Across all three tasks on NYUv2 data, our Mask3D pre-training achieves notably improved performance than training from scratch as well as the various baseline pre-training methods. In particular, we achieve an improvement of $+6.9$ mIoU, $+14.1$ AP@0.5, and $+11.9$ AP@0.5 over the common supervised ImageNet pretraining on semantic segmentation, object detection, and instance segmentation.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dc793ad43c6d5e266e38f040d39df204001cffa534cd4776a4ae5bb8737ef93e.jpg", + "table_caption": [ + "Table 6. NYUv2 2D Object Detection. Fine-tuning with Mask3D pre-trained models leads to improved object detection results across different metrics, showing an effective transfer across dataset characteristics." + ], + "table_footnote": [], + "table_body": "
Pre-training MethodBackbonemIoU
ImageNet Pre-training (supIN)ResNet-5054.1
Pri3D [25]ResNet-5055.1 (+1.00)
MAE-unsupIN→SN [19]ViT64.7 (+10.6)
Mask3D (Ours)ViT66.4 (+12.3)
", + "bbox": [ + 501, + 262, + 890, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Cityscapes 2D Semantic Segmentation. Mask3D significantly outperforms state-of-the-art Pri3D as well as a strong MAE-style pre-training. This demonstrates the effectiveness of the transferability of Mask3D, even under a significant domain gap.", + "bbox": [ + 498, + 339, + 893, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Out-of-domain Transfer", + "text_level": 1, + "bbox": [ + 500, + 401, + 723, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While Mask3D concentrates on pre-training for improving indoor scene understanding, we further demonstrate the effectiveness of our Mask3D pre-training for the out-of-domain transfer on outdoor data, such as Cityscapes [8]. We use the official data split of 3000 images for training and 500 images for the test. To evaluate the transferability in such a large domain gap scenario, we fine-tune the pretrained models for the 2D semantic segmentation task in Tab. 7. Our approach maintains performance improvement over baseline pre-training methods such as Pri3D (+11.3 mIoU) and MAE-unsupIN→SN (+1.7 mIoU). This indicates an encouraging transferability of our learned representations and their applicability to a variety of scenarios. Please refer to the supplemental material for more out-of-domain transfer results on more generally distributed data, such as ADE20K [46].", + "bbox": [ + 496, + 425, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 676, + 665, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Does the pre-training masking ratio matter? We show how different masking ratios influence downstream task results in Tab. 8 on ScanNet semantic segmentation. We found a performance gain when masking more RGB values (keeping $20\\%$ ), which in combination with the heavy depth masking (keeping $20\\%$ ) leads to the best performance.", + "bbox": [ + 496, + 699, + 890, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What about other ViT variants? In our experiments, we use ViT-B as the meta-architecture. We show Mask3D also works in other ViT variants, such as ViT-L (see Tab. 12), which exhibits a similar trend of improvements.", + "bbox": [ + 496, + 800, + 893, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Does the normalization in the reconstruction loss help? We normalize the features when computing the reconstruc", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13516", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d9040d351335d5e679882a07035870c0c2b212ef683dc01166881cd5bed94cdd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RGB RatioDepth RatiomIoU
20.0%0.0%65.2
20.0%20.0%66.7
20.0%80.0%65.5
50.0%20.0%65.9
50.0%50.0%64.7
80.0%20.0%64.8
80.0%50.0%64.8
100.0%0.0%64.6
100.0%20.0%64.8
100.0%100.0%64.5
", + "bbox": [ + 147, + 88, + 395, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion loss and observe an improvement of $+0.8\\%$ mIoU in the semantic segmentation task on ScanNet.", + "bbox": [ + 76, + 313, + 468, + 343 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/41d83f121278ce92cb4dbf81148bb934406239ae220c112c653718b9a2a0dec4.jpg", + "table_caption": [ + "Table 8. Ablation Study of Masking Ratios. on ScanNet 2D semantic segmentation. We mask out different ratios of RGB and depth patches, where the ratio indicates the percentage of kept patches. Refer to supplemental material for a full list." + ], + "table_footnote": [], + "table_body": "
MethodBackbonemIoU
Train from ScratchViT32.6
MAEViT37.1
Mask3DViT42.2
", + "bbox": [ + 137, + 356, + 405, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared to a pure depth prediction baseline. In Tab. 8, we demonstrate a superior performance with a $20\\%$ kept patches of RGB and depth, compared to a pure depth prediction method (66.7 vs. 64.6). Note in the table, pretraining with $100\\%$ RGB ratio and $0\\%$ depth ratio is equivalent to a pure depth prediction from a RGB image.", + "bbox": [ + 76, + 467, + 468, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Color + depth reconstruction? We found that having joint losses on color and depth during pre-training does not benefit performance (see the following Tab. 10). The RGB reconstruction loss potentially makes pre-training easier, as we already have additional depth priors as guidance.", + "bbox": [ + 75, + 569, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a8a60743a9ffa7c7ee8100ee633578b93eef671de55232ea802aec0d0790cd50.jpg", + "table_caption": [ + "Table 9. Results on ScanNet Semantic Segmentation without ImageNet pre-training. Similar trend is seen as ImageNet pretraining. The Gap gets larger compared to ImageNet pre-training." + ], + "table_footnote": [], + "table_body": "
MethodReconstructionmIoU
Mask3DRGB+Depth65.6
Mask3DDepth66.7
", + "bbox": [ + 153, + 657, + 390, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "No Stage-I pre-training. We observe a performance drop without ImageNet pre-training model as initialization for our pre-training. Since ImageNet pre-training is readily available and ScanNet has a relatively small amount of indoor data, we make ImageNet pre-training initialization as default, similar to Pri3D. Meanwhile, we conduct experiments without ImageNet pre-training in Tab. 9, and observe similar trends as when using ImageNet pre-training.", + "bbox": [ + 76, + 739, + 468, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RGB + semantic segmentation as pre-training. Using RGB and semantic segmentation for pre-training rather than", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ed3ac8b6e1010709ba434d44cac4590c75ad4ae1237e94337e08e0744f1c98dd.jpg", + "table_caption": [ + "Table 10. ScanNet Semantic Segmentation. RGB as an additional signal does not bring a significant improvement." + ], + "table_footnote": [], + "table_body": "
DatasetsMask3D - SemanticsMask3D
ScanNet65.966.7
NYUv255.556.9
CityScapes63.066.4
", + "bbox": [ + 542, + 88, + 846, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae5dee2d82c18abc68501d66a71e68178ba1d44ff3f760b832606fc81b5e0c10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 186, + 630, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb7469c911f93ad67b9482aeaaf11ff8ebb6d53280d622c65f786515b1518ce1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 186, + 759, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3ec038a4326ee242fd96bb273d9b5b6a362828e63f27906cb952b7ff3142598.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 186, + 888, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2da07e039d5b83d4b0ecb61be3cec12a41e2cc3b5eb91514a06210e793cac89.jpg", + "image_caption": [ + "Figure 6. Pre-trained ViT learns 3D structural priors. Our proposed pre-training method learns spatial structures from heavily masked RGB images." + ], + "image_footnote": [], + "bbox": [ + 501, + 263, + 630, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93f69c2a947fb008725e3a0a2582e6f1b68bd82dd9a8268c3f89880376d7d6f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 263, + 759, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4364ded99b9cc21523aa9333edc5219121b5264aae70e3f4094f0f75d53400bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 263, + 888, + 338 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5b4481c68c8016e448be80e8fa43f17edc68735078888f33a28a3d3ed2654092.jpg", + "table_caption": [ + "Table 11. Semantic segmentation results (mIoU). \"Mask3D - Semantics\" denotes pre-training using RGB+Semantics." + ], + "table_footnote": [], + "table_body": "
Pre-training MethodBackbonemIoU
Pri3D [25]ResNet-5060.2
Pri3D [25]ViT-B59.3 (-0.9)
MAE-unsupIN→SN [19]ViT-B63.3 (+3.1)
Mask3D (Ours)ViT-B66.7 (+6.5)
Pri3D [25]ViT-L64.3 (+4.1)
MAE-unsupIN→SN [19]ViT-L68.2 (+8.0)
Mask3D (Ours)ViT-L70.7 (+10.5)
", + "bbox": [ + 529, + 396, + 859, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 12. ViT Variants on ScanNet 2D Semantic Segmentation. Mask3D yields consistent improvements for both ViT-B and ViT-L backbone architectures.", + "bbox": [ + 500, + 518, + 890, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "depth completion achieved competitive results on ScanNet semantic segmentation, although this requires the use of semantic labels for the pre-training dataset, and is likely to be less transferable across domains than using depth completion. As shown in the following Tab. 11, the gap becomes larger when transferring to both NYUv2 and Cityscapes.", + "bbox": [ + 500, + 569, + 890, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 676, + 617, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present Mask3D, a new self-supervised approach to embed 3D priors into learned 2D representations for image scene understanding. We leverage existing large-scale RGB-D data to learn 3D priors without requiring any camera pose or multi-view correspondence information, instead learning geometric and structural cues through a pre-text reconstruction task from masked color and depth. We show that Mask3D is particularly effective in pre-training the modern, powerful ViT backbones, with notable improvements across a variety of image-based tasks and datasets. We believe this shows the strong potential in effectively learning 3D priors and provides new avenues for such 3D-grounded representation learning.", + "bbox": [ + 500, + 704, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13517", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In ICCV, 2016. 1, 2", + "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 348-367. Springer, 2022. 2", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5", + "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 1, 2", + "[5] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2", + "[6] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. arXiv preprint arXiv:2112.02990, 2021. 2", + "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2", + "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3, 4, 7", + "[9] Manuel Dahnert, Ji Hou, Matthias Nießner, and Angela Dai. Panoptic 3d scene reconstruction from a single rgb image. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 1, 2, 3, 4, 5", + "[11] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-468, 2018. 2", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.1,3", + "[15] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6433-6442, 2021. 2", + "[16] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation. In CVPR, 2020. 2", + "[17] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2", + "[18] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-aware 3D instance segmentation. In CVPR, 2020. 2", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4, 5, 7, 8", + "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 4, 5, 7", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1", + "[22] Ji Hou, Angela Dai, and Matthias Nießner. 3D-SIS: 3D Semantic Instance Segmentation of RGB-D Scans. In CVPR, 2019. 2", + "[23] Ji Hou, Angela Dai, and Matthias Nießner. **RevealNet: Seeing Behind Objects in RGB-D Scans.** In CVPR, 2020. 2", + "[24] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2", + "[25] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In ICCV, 2021. 2, 3, 4, 5, 7, 8", + "[26] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation. In CVPR, 2020. 2", + "[27] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In ICCV, 2019. 2", + "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2", + "[29] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In CVPR, 2021. 2", + "[30] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3D object detection in point clouds. ICCV, 2019. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13518", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. CVPR, 2017. 2", + "[32] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[34] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGB-D images. ECCV, 2012. 1, 2, 3, 4, 6", + "[35] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite. In CVPR, 2015. 1", + "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. KPConv: Flexible and deformable convolution for point clouds. In CVPR, 2019. 2", + "[37] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021. 2", + "[38] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.4", + "[39] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas J Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3D point cloud understanding. ECCV, 2020. 2", + "[40] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3D instance segmentation on point clouds. In NeurIPS, 2019. 2", + "[41] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas Guibas. GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In CVPR, 2019. 2", + "[42] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. arXiv preprint arXiv:2303.08231, 2023. 2", + "[43] Yu Zhang, Junle Yu, Xiaolin Huang, Wenhui Zhou, and Ji Hou. Pcr-cg: Point cloud registration via deep explicit color and geometry. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 443-459. Springer, 2022. 2", + "[44] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2", + "[45] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In European Conference on Computer Vision, pages 311-329. Springer, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7", + "bbox": [ + 501, + 90, + 892, + 159 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13519", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_model.json b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb6779c1062577f43f2c08d5aefffaa500baf8eb --- /dev/null +++ b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.131, + 0.88, + 0.152 + ], + "angle": 0, + "content": "Mask3D: Pre-training 2D Vision Transformers by Learning Masked 3D Priors" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.18, + 0.785, + 0.201 + ], + "angle": 0, + "content": "Ji Hou\\(^{1}\\) Xiaoliang Dai\\(^{1}\\) Zijian He\\(^{1}\\) Angela Dai\\(^{2}\\) Matthias Nießner\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.205, + 0.701, + 0.224 + ], + "angle": 0, + "content": "\\(^{1}\\)Meta Reality Labs \\(^{2}\\)Technical University of Munich" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.253, + 0.16, + 0.264 + ], + "angle": 0, + "content": "Pre-text Task" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.267, + 0.325, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.268, + 0.516, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.253, + 0.694, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.267, + 0.886, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.392, + 0.277, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.507, + 0.231, + 0.518 + ], + "angle": 0, + "content": "Scene Understanding Tasks" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.392, + 0.315, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.393, + 0.51, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.507, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Instance Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.393, + 0.695, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.507, + 0.673, + 0.518 + ], + "angle": 0, + "content": "Semantic Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.393, + 0.887, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.507, + 0.844, + 0.518 + ], + "angle": 0, + "content": "Object Detection" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.541, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Figure 1. We present Mask3D, which learns to embed 3D priors to 2D representations for image understanding tasks, based on a self-supervised pre-training formulation from single RGB-D views, without requiring any camera pose or multi-view correspondence information. Our pre-training takes masked RGB and depth patches as input to reconstruct the dense depth map, and the pre-trained color backbone is used to fine-tune various downstream image understanding tasks. This results in effective ViT pre-training for a variety of downstream tasks and datasets." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.621, + 0.313, + 0.637 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Current popular backbones in computer vision, such as Vision Transformers (ViT) and ResNets are trained to perceive the world from 2D images. However, to more effectively understand 3D structural priors in 2D backbones, we propose Mask3D to leverage existing large-scale RGB-D data in a self-supervised pre-training to embed these 3D priors into 2D learned feature representations. In contrast to traditional 3D contrastive learning paradigms requiring 3D reconstructions or multi-view correspondences, our approach is simple: we formulate a pre-text reconstruction task by masking RGB and depth patches in individual RGB-D frames. We demonstrate the Mask3D is particularly effective in embedding 3D priors into the powerful 2D ViT backbone, enabling improved representation learning for various scene understanding tasks, such as semantic segmentation, instance segmentation and object detection." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.623, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Experiments show that Mask3D notably outperforms existing self-supervised 3D pre-training approaches on ScanNet, NYUv2, and Cityscapes image understanding tasks, with an improvement of \\(+6.5\\%\\) mIoU against the state-of-the-art Pri3D on ScanNet image semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.725, + 0.631, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recent years have seen remarkable advances in 2D image understanding as well as 3D scene understanding, although their representation learning has generally been treated separately. Powerful 2D architectures such as ResNets [21] and Vision Transformers (ViT) [14] have achieved notable success in various 2D recognition and segmentation tasks, but focus on learning from 2D image data. Current large-scale RGB-D datasets [1,4,10,34,35] provide an opportunity to learn key geometric and structural priors to provide more informed reasoning about the scale and cir" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13510" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "cumvent view-dependent effects, which can provide more efficient representation learning. In 3D, various successful methods have been leveraging the RGB-D datasets for constructive point discrimination [6,24,39,44] for downstream 3D tasks, including high-level scene understanding tasks as well as low-level point matching tasks [15,42,43]. However, the other direction from 3D to 2D is less explored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.468, + 0.363 + ], + "angle": 0, + "content": "We thus aim to embed such 3D priors into 2D backbones to effectively learn the structural and geometric priors underlying the 3D scenes captured in 2D image projections. Recently, Pri3D [25] adopted similar multi-view and reconstruction-based constraints to induce 3D priors in learned 2D representations. However, this relies on not only acquiring RGB-D frame data but also the robust registration of multiple views to obtain camera pose information for each frame. Instead, we consider how to effectively learn such geometric priors from only single-view RGB-D data in a more broadly applicable setting for 3D-based pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.468, + 0.514 + ], + "angle": 0, + "content": "We thus propose Mask3D, which learns effective 3D priors for 2D backbones in a self-supervised fashion by pretraining with single-view RGB-D frame data. We propose a pre-text reconstruction task to reconstruct the depth map by masking different random RGB and depth patches of an input frame. These masked input RGB and depth are encoded simultaneously in separate encoding branches and decoded to reconstruct the dense depth map. This imbues 3D priors into the RGB backbone which can then be used for fine-tuning downstream image based scene understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.468, + 0.725 + ], + "angle": 0, + "content": "In particular, our self-supervised approach to embedding 3D priors from single-view RGB-D data to 2D learned features is not only more generally applicable, but we also demonstrate that it is particularly effective for pre-training vision transformers. Our experiments demonstrate the effectiveness of Mask3D on a variety of datasets and image understanding tasks. We pre-train on ScanNet [10] with our masked 3D pre-training paradigm and fine-tune for 2D semantic segmentation, instance segmentation, and object detection. This enables notable improvements not only on ScanNet data but also generalizes to NYUv2 [34] and even Cityscapes [8] data. We believe that Mask3D makes an important step to shed light on the paradigm of incorporating 3D representation learning to powerful 2D backbones." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.726, + 0.327, + 0.74 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.742, + 0.468, + 0.833 + ], + "angle": 0, + "content": "- We introduce a self-supervised pre-training approach to learn masked 3D priors for 2D image understanding tasks based on learning from only single-view RGB-D data, without requiring any camera pose or 3D reconstruction information, and thus enabling more general applicability." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.841, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We demonstrate that our masked depth reconstruction pre-training is particularly effective for the modern, powerful ViT architecture, across a variety of datasets and image understanding tasks." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.742, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Pre-training in Visual Transformers. Recently, visual transformers have revolutionized computer vision and attracted wide attention. In contrast to popular CNNs that operate in a sliding window fashion, Vision Transformers (ViT) describe the image as patches of 16x16 pixels. The Swin Trasnformer [28] has set new records with its hierarchical transformer formulation on major vision benchmarks. The dominance of visual transformers in many vision tasks has inspired study into how to pre-training such backbones. MoCoV3 [5] first investigated the effects of several fundamental components for self-supervised ViT training. MAE [19] then proposed an approach inspired by BERT [13], which randomly masks words in sentences and leveraged masked image reconstruction for self-supervised pre-training that achieved state-of-the-art results in ViT. A similar self-supervision has also been proposed by MaskFeat [37] for self-supervised video pre-training. MaskFeat randomly masks out pixels of the input sequence and then predicts the Oriented Gradients (HOG) of the masked regions. However, such ViT pre-training methods focus on image or video data, without exploring how 3D priors can potentially be exploited. MultiMAE [2] on the other hand introduces depth priors. However, it requires depth as input not only in pre-training but also in downstream tasks. In addition to depth, human annotations (e.g., semantics) are also leveraged in the pre-training. To achieve a self-supervised pre-training, we do not use semantics in the pre-training and only use RGB images as input in downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.818 + ], + "angle": 0, + "content": "RGB-D Scene Understanding. Research in 3D scene understanding have been spurred forward with the introduction of larger-scale, annotated real-world RGB-D datasets [1, 4, 10]. This has enabled data-driven semantic understanding of 3D reconstructed environments, where we have now seen notable progress, such as for 3D semantic segmentation [7, 11, 17, 31, 32, 36], object detection [29, 30, 45], instance segmentation [16, 18, 22, 23, 26, 27, 40, 41], and recently panoptic segmentation [9]. Such 3D scene understanding tasks have been analogously defined to 2D image understanding, which considers RGB-only input without depth information. However, learning from 3D enables geometric reasoning without requiring learning view-dependent effects or resolving depth/scale ambiguity that must be learned when considering solely 2D data. We thus take advantage of existing large-scale RGB-D data to explore how to effectively embed 3D priors for better representation learning for 2D scene understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Embedding 3D Priors in 2D Backbones. Learning cross-modality features has been seen in extensive studies of the ties between languages and images. In particular, CLIP [33] learns visual features from language supervision during pre-training, showing promising results in zero-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13511" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "shot learning for image classification. Pri3D [25] explores 3D-based pre-training for image-based tasks by leveraging multi-view consistency and 2D-3D correspondence with contrastive learning to embed 3D priors into ResNet backbones. This results in enhanced features over ImageNet pretraining on 2D scene understanding tasks. However, Pri3D requires camera pose registration across RGB-D video sequences and is specifically designed for CNNs-based architectures. In contrast, we formulate a self-supervised pretraining that operates on only single-view RGB-D frames and leverages masked 3D priors that can effectively pretrain powerful ViT backbones." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.285, + 0.17, + 0.301 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.473, + 0.508 + ], + "angle": 0, + "content": "We introduce Mask3D to embed 3D priors into learned 2D representations by self-supervised pre-training from only single-view RGB-D frames. To effectively learn 3D structural priors without requiring any camera pose information or multi-view constraints, we formulate a pre-text depth reconstruction task to inform the RGB feature extraction to be geometrically aware. Randomly masked color and depth images are used as input to reconstruct the dense depth map, and the RGB backbone can then be used to fine-tune downstream image understanding tasks. In particular, we show in Sec. 4 that this single-frame self-supervision is particularly well-suited for powerful vision transformer (ViT) backbones, even without any multi-view information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.33, + 0.531 + ], + "angle": 0, + "content": "3.1. Learning Masked 3D Priors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.645 + ], + "angle": 0, + "content": "We propose to learn masked 3D priors to embed to learned 2D backbones by pre-training to reconstruct dense depth from RGB images with the guidance of sparse depth. That is, for an RGB-D frame \\( F = (C, D) \\) with RGB image \\( C \\) and depth map \\( D \\), we train to reconstruct \\( D \\) from masked patches of \\( C \\) guided with sparse masked patches of \\( D \\). An overview of our approach is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.75 + ], + "angle": 0, + "content": "To create masked color and depth \\( M_{c} \\) and \\( M_{d} \\) from \\( C \\) and \\( D \\) as input for reconstruction, a 240x320 RGB image \\( C \\) is uniformly divided into 300 16x16 patches, from which we randomly keep a percentage \\( p_{c} \\) of patches, masking out the others, to obtain \\( M_{c} \\). \\( M_{d} \\) is created similarly by keeping only a percentage \\( p_{d} \\) of patches, such that the resulting depth patches do not coincide with the RGB patches in \\( M_{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We then train color and depth encoders \\(\\Psi_{c}\\) and \\(\\Psi_{d}\\) to separately encode RGB and depth signals. RGB patches are fed into \\(\\Psi_{c}\\) and concatenated with a positional embedding, following the ViT architecture, and similarly for depth. The positional embedding used encodes the patch location by a cosine function. Patches and their positional embeddings are then mapped into higher dimensional feature vectors via \\(\\Psi_{c}\\) and \\(\\Psi_{d}\\). The encoders \\(\\Psi_{c}\\) and \\(\\Psi_{d}\\) are built by blocks composed of linear and norm layers. The features from \\(\\Psi_{c}\\) and \\(\\Psi_{d}\\) are then fused in the bottleneck; since depth patches" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "were selected in regions where no RGB patches were selected, there are no duplicate patches representing the same patch location." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.333 + ], + "angle": 0, + "content": "For those regions which do not have any associated RGB or depth patch, we use patches of constant values as mask tokens to create a placeholder in the bottleneck to enable reconstructing dense depth at the original image resolution. In the bottleneck, the RGB and depth patch feature vectors, along with the mask tokens, form the input to the decoder. This formulates a reconstruction task from sparse RGB and depth; the joint RGB-D pre-training enables reconstruction from very sparse input, as shown by our ablation on the masked input rations in Sec. 4.5. Note that the depth encoder is trained only during pre-training, and only the color ViT encoder (and decoder, if applicable) are used for downstream fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.894, + 0.395 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the pre-training task, we demonstrate the depth completion results from the pretraining phase in Fig. 6. A detailed analysis of masking different ratios of color and depth signals is shown in Sec. 4.5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.894, + 0.488 + ], + "angle": 0, + "content": "Pre-training Loss In contrast to the widely used contrastive loss in 3D representation learning, we train for dense depth reconstruction with an \\(\\ell_2\\) reconstruction loss. Similar to MAE [19], we normalize the output patches as well as the target patches prior to computing the loss, which we found to empirically improve the performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.502, + 0.587, + 0.517 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.65 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of Mask3D pretraining for ViT [14] backbones on semantic segmentation, instance segmentation, and object detection tasks. We pre-train on ScanNet [10] data and demonstrate the effectiveness of learned masked 3D priors for not only ScanNet downstream tasks but also their transferability to NYUv2 [34] and even across the indoor/outdoor domain gap to Cityscapes [8] data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.659, + 0.691, + 0.675 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.728 + ], + "angle": 0, + "content": "In this section, we introduce the pre-training and finetuning procedures. Our method uses a two-stage pretraining design introduced in the following." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Stage-I: Mask3D Encoder Initialization. We initialize the RGB encoder with network weights trained on ImageNet [12] (as pre-training for our pre-training). To maintain a fully self-supervised pre-training paradigm, we initialize with weights obtained by self-supervised ImageNet pre-training [19]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Stage-II: Mask3D Pre-training on ScanNet. Mask3D pre-training leverages 3D priors in RGB-D frame data, for which we use the color and depth maps of ScanNet [10]. Note that this does not use any semantic or reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13512" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.893, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Overview of Mask3D Pre-training. As a pretext task, we predict dense depth from color and sparse depth signals. We use masked input by randomly selecting a set of patches from the input color image, which are then mapped to higher dimensional feature vectors; input depth is similarly masked and encoded. The color and depth features are then fused into a bottleneck from which dense depth is reconstructed as a self-supervised loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.471, + 0.471 + ], + "angle": 0, + "content": "information during pre-training. ScanNet contains 2.5M RGB-D frames from 1513 ScanNet train video sequences. We regularly sample every \\(25^{\\text{th}}\\) frame without any other filtering (e.g., no control on viewpoint variation)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.663 + ], + "angle": 0, + "content": "Downstream Fine-tuning. We evaluate our Mask3D pretraining by fine-tuning a variety of downstream image understanding tasks (semantic segmentation, instance segmentation, object detection). We consider in-domain transfers on ScanNet image understanding, and further evaluate the out-of-domain transfer on datasets with different statistical characteristics: the indoor image data of NYUv2 [34], as well as across a strong domain gap to the outdoor image data of Cityscapes [8]. For semantic segmentation tasks, we use both encoder and decoder pre-trained with Mask3D, and for instance segmentation and detection, only the backbone encoder is pre-trained." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Baselines. To evaluate the effectiveness of our learned masked 3D priors for 2D representations, we benchmark our method against relevant baselines:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Supervised ImageNet Pre-training (supIN). This uses the pre-trained weights from ImageNet, provided by torchvision, as is commonly used for image understanding tasks. Here, only ImageNet data is used, and no ScanNet data is involved in the pre-training phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.884 + ], + "angle": 0, + "content": "2-Stage MoCoV2 (MoCoV2-supIN \\(\\rightarrow\\) SN). Supervised ImageNet pre-trained (supIN) weights are used as network initialization for pre-training. MoCoV2 [20] is used for pre-training with randomly shuffled ScanNet images. Here, both ImageNet and ScanNet image data are used without any geometric priors." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "2-Stage MAE (MAE-unsupIN \\(\\rightarrow\\) SN). Self-supervised Im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.485 + ], + "angle": 0, + "content": "ageNet pre-trained weights are used as network initialization for pre-training. MAE [19] is used for pre-training with randomly shuffled ScanNet images. Here, both ImageNet and ScanNet image data are used without any geometric priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Pri3D [25]. Supervised ImageNet pre-trained are used to initialize Pri3D pre-training, which leverages multi-view and reconstruction constraints from ScanNet data under a contrastive loss. Here, both ImageNet and ScanNet data are used, incorporating 3D priors from reconstructed RGB-D video sequences for pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.893, + 0.823 + ], + "angle": 0, + "content": "Implementation Details. We use a ViT-B backbone to train our approach. For pre-training, we use an SGD optimizer with a learning rate of 0.1 and an effective batch size of 128 (accumulated gradients from an actual batch size of 64). The learning rate is decreased by a factor of 0.99 every 1000 steps, and our method is trained for 100 epochs. Fine-tuning on semantic segmentation is trained with a batch size of 8 for 80 epochs. The initial learning rate is 0.01, with polynomial decay with a power of 0.9. Finetuning on detection and instance segmentation is trained using Detector2 [38] with the 1x schedule. Pre-training experiments are conducted on a single NVIDIA A6000 GPU, or 2 NVIDIA RTX3090 GPUs, or 4 NVIDIA RTX2080Ti GPUs; semantic segmentation experiments are conducted on a single NVIDIA A6000 GPU; instance segmentation and detection experiments are conducted on 8 V100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.75, + 0.848 + ], + "angle": 0, + "content": "4.2. ScanNet Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of representation learning with 3D priors via downstream tasks on ScanNet [10] images. For fine-tuning, we follow the standard" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13513" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.166, + 0.089, + 0.803, + 0.246 + ], + "angle": 0, + "content": "
Pre-training MethodBackbonePre-training DatamIoU
ScratchResNet-50None39.1
ImageNet Pre-training (supIN)ResNet-50ImageNet55.7
Supervised Pre-trainingViTImageNet+ScanNet65.9 (+10.2)
MoCoV2-supIN→SN [20]ResNet-50ImageNet+ScanNet56.6 (+0.9)
Pri3D [25]ResNet-50ImageNet+ScanNet60.2 (+4.5)
Pri3D [25]ViTImageNet+ScanNet59.3 (+3.6)
DINO [3]ViTImageNet+ScanNet58.1 (+3.6)
MAE-unsupIN→SN [19]ViTImageNet+ScanNet63.3 (+7.6)
Ours - Mask3D (DINO)ViTImageNet+ScanNet60.5 (+4.8)
Ours - Mask3D (MAE)ViTImageNet+ScanNet66.7 (+11.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.253, + 0.895, + 0.283 + ], + "angle": 0, + "content": "Table 1. ScanNet 2D Semantic Segmentation. Mask3D significantly outperforms Pri3D as well as other state-of-the-art pre-training approaches that leverage both ImageNet and ScanNet data." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.296, + 0.48, + 0.399 + ], + "angle": 0, + "content": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch32.717.716.9
ImageNet Pretrain (supIN)41.725.925.1
MoCoV2-supIN→SN [20]43.5 (+1.8)26.8 (+0.9)25.8 (+0.7)
Pri3D [25]43.7 (+2.0)27.0 (+1.1)26.3 (+1.2)
MAE-unsupIN→SN [19]46.1 (+4.4)32.7 (+6.8)30.5 (+5.4)
Mask3D (Ours)50.4 (+8.7)35.3 (+9.4)32.7 (+7.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.405, + 0.472, + 0.475 + ], + "angle": 0, + "content": "Table 2. ScanNet 2D Object Detection. Fine-tuning with Mask3D pre-trained models leads to improved object detection results across different metrics, in comparison to ImageNet pretraining, MoCo-style pre-training, and a strong MAE-style pre-training method." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.488, + 0.481, + 0.591 + ], + "angle": 0, + "content": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch25.813.112.2
ImageNet Pretrain (supIN)32.617.817.6
MoCoV2-supIN→SN [20]33.9 (+1.3)18.1 (+0.3)18.3 (+0.7)
Pri3D [25]34.3 (+1.7)18.7 (+0.9)18.3 (+0.7)
MAE-unsupIN→SN [19]37.4 (+4.8)20.3 (+2.5)20.7 (+3.1)
Mask3D (Ours)41.2 (+8.6)22.7 (+4.9)22.8 (+5.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.597, + 0.472, + 0.667 + ], + "angle": 0, + "content": "Table 3. ScanNet 2D Instance Segmentation. Fine-tuning with Mask3D pre-trained models leads to improved instance segmentation results across different metrics compared to ImageNet pre-training, MoCo-style pre-training, and a strong MAE-style pre-training method." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.68, + 0.481, + 0.782 + ], + "angle": 0, + "content": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch17.29.28.8
ImageNet Pretrain (supIN)25.113.913.4
MoCoV2-supIN→SN [20]27.2 (+2.1)14.7 (+0.2)14.8 (+1.4)
Pri3D [25]28.1 (+3.0)15.7 (+1.8)15.7 (+2.3)
MAE unsupIN→SN [19]33.6 (+8.5)19.0 (+5.1)19.0 (+5.6)
Mask3D (Ours)37.0 (+11.9)21.6 (+7.7)21.3 (+7.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.788, + 0.47, + 0.845 + ], + "angle": 0, + "content": "Table 4. NYUv2 2D Instance Segmentation. Fine-tuning with Mask3D pre-trained models leads to improved instance segmentation results across different metrics compared to previous methods, demonstrating the cross-dataset transfer ability of Mask3D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.865, + 0.47, + 0.913 + ], + "angle": 0, + "content": "protocol of the ScanNet benchmark [10] and sample every \\(100^{\\mathrm{th}}\\) frame, resulting in 20,000 train images and 5,000 validation images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.895, + 0.542 + ], + "angle": 0, + "content": "2D Semantic Segmentation. Tab. 1 shows the fine-tuning for semantic segmentation, in comparison with baseline pre-training approaches. All pre-training methods significantly improve performance over training the semantic segmentation model from scratch. In particular, Mask3D provides substantially better representation quality leading to a much stronger improvement over supervised ImageNet pretraining \\((+11\\mathrm{mIoU})\\), and notably improving over MAE unsupIN \\(\\rightarrow\\) SN with ImageNet and ScanNet \\((+3.4\\mathrm{mIoU})\\) and the 3D-based pre-training of Pri3D \\((+6.5\\mathrm{mIoU})\\). We note that the multi-view 3D pre-training of Pri3D does not effectively embed informative 3D priors to ViT backbones, rather suffering from performance degradation from a ResNet-50 backbone. In contrast, our Mask3D pretraining can notably improve performance with a ViT backbone, indicating the effectiveness of our learned 3D priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.895, + 0.713 + ], + "angle": 0, + "content": "2D Object Detection and Instance Segmentation We show that Mask3D provides effective general 3D priors for a variety of image-based tasks, by evaluating downstream object detection and instance segmentation in Tab. 2 and Tab. 3, respectively. Across all tasks, various pre-training approaches yield substantial improvement over training from scratch. Our masked 3D prior learning transfers effectively learned representations for object detection and instance segmentation, notably improving over the best-performing MAE-unsupIN \\(\\rightarrow\\) SN (+4.3 AP@0.5 and +3.8 AP@0.5, respectively)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Data-Efficient Scenarios. We additionally show that our single-view RGB-D pre-training to embed 3D priors in limited data scenarios for downstream ScanNet semantic segmentation in Fig. 5. Mask3D shows consistent improvements across a range of limited data; even with only \\(20\\%\\) of the training data, we recover \\(80\\%\\) performance with \\(100\\%\\) training data available and improving \\(+15.2\\) mIoU over Pri3D pre-training on a ViT backbone." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.744, + 0.862 + ], + "angle": 0, + "content": "4.3. NYUv2 Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We demonstrate the generalizability of our 3D-imbued learned feature representations across datasets, us" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13514" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.87, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.449, + 0.896, + 0.493 + ], + "angle": 0, + "content": "Figure 3. Qualitative Results on Various Tasks across Different Benchmarks. We visualize predictions on different tasks across various scene understanding benchmarks. From top to bottom rows: instance segmentation on ScanNet, instance segmentation on NYUv2, semantic segmentation on NYUv2, and semantic segmentation results in ScanNet." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.501, + 0.891, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.8, + 0.895, + 0.829 + ], + "angle": 0, + "content": "Figure 4. More Qualitative Results on Semantic Segmentation. We visualize semantic segmentation predictions on various scene understanding benchmarks including ScanNet and NYUv2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.842, + 0.473, + 0.888 + ], + "angle": 0, + "content": "ing Mask3D pre-trained on ScanNet and fine-tuned on NYUv2 [34] following the same fine-tuning setup as before. NYUv2 contains Microsoft Kinect RGB-D video" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.842, + 0.895, + 0.889 + ], + "angle": 0, + "content": "sequences of indoor scenes, comprising 1449 densely labeled RGB images. We use the official 795/654 train/val split. Tables 5, 6, and 4 evaluate the downstream tasks" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13515" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.202, + 0.089, + 0.768, + 0.219 + ], + "angle": 0, + "content": "
Pre-training MethodBackbonePre-training DatamIoU
ScratchResNet-50None24.8
ImageNet Pre-training (supIN)ResNet-50ImageNet50.0
Supervised Pre-trainingViTImageNet+ScanNet55.5 (+5.5)
MoCoV2-supIN→SN [20]ResNet-50ImageNet+ScanNet47.6 (-2.4)
Pri3D [25]ResNet-50ImageNet+ScanNet54.2 (+4.2)
Pri3D [25]ViTImageNet+ScanNet53.2 (+3.2)
MAE-unsupIN→SN [19]ViTImageNet+ScanNet54.9 (+4.9)
Mask3D (Ours)ViTImageNet+ScanNet56.9 (+6.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.224, + 0.895, + 0.253 + ], + "angle": 0, + "content": "Table 5. NYUv2 2D Semantic Segmentation. Mask3D significantly outperforms state-of-the-art pre-training approaches, demonstrating its effectiveness in transferring to different dataset characteristics." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.268, + 0.462, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.498, + 0.472, + 0.584 + ], + "angle": 0, + "content": "Figure 5. Data-Efficient Results. Compared to previous methods, Mask3D demonstrates consistent improvements on ScanNet 2D semantic segmentation across a range of limited data scenarios. Mask3D is particularly effective for ViT pre-training, improving \\(+15.2\\%\\) mIoU over state-of-the-art Pri3D [25] on a ViT backbone at \\(20\\%\\) of the training data." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.596, + 0.486, + 0.699 + ], + "angle": 0, + "content": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch21.310.39.0
ImageNet Pretrain (supIN)29.917.316.8
MoCoV2-supIN→SN [20]30.1 (+0.20)18.1 (+0.80)17.3 (+0.50)
Pri3D [25]33.0 (+2.10)19.8 (+2.60)18.9 (+2.10)
MAE-unsupIN→SN [19]40.3 (+10.4)24.5 (+7.20)23.2 (+6.40)
Mask3D (Ours)44.0 (+14.1)28.3 (+6.40)25.9 (+9.10)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.757 + ], + "angle": 0, + "content": "Table 6. NYUv2 2D Object Detection. Fine-tuning with Mask3D pre-trained models leads to improved object detection results across different metrics, showing an effective transfer across dataset characteristics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "of 2D semantic segmentation, object detection, and instance segmentation, respectively. Across all three tasks on NYUv2 data, our Mask3D pre-training achieves notably improved performance than training from scratch as well as the various baseline pre-training methods. In particular, we achieve an improvement of \\(+6.9\\) mIoU, \\(+14.1\\) AP@0.5, and \\(+11.9\\) AP@0.5 over the common supervised ImageNet pretraining on semantic segmentation, object detection, and instance segmentation." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.338 + ], + "angle": 0, + "content": "
Pre-training MethodBackbonemIoU
ImageNet Pre-training (supIN)ResNet-5054.1
Pri3D [25]ResNet-5055.1 (+1.00)
MAE-unsupIN→SN [19]ViT64.7 (+10.6)
Mask3D (Ours)ViT66.4 (+12.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.34, + 0.894, + 0.397 + ], + "angle": 0, + "content": "Table 7. Cityscapes 2D Semantic Segmentation. Mask3D significantly outperforms state-of-the-art Pri3D as well as a strong MAE-style pre-training. This demonstrates the effectiveness of the transferability of Mask3D, even under a significant domain gap." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.402, + 0.724, + 0.418 + ], + "angle": 0, + "content": "4.4. Out-of-domain Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.669 + ], + "angle": 0, + "content": "While Mask3D concentrates on pre-training for improving indoor scene understanding, we further demonstrate the effectiveness of our Mask3D pre-training for the out-of-domain transfer on outdoor data, such as Cityscapes [8]. We use the official data split of 3000 images for training and 500 images for the test. To evaluate the transferability in such a large domain gap scenario, we fine-tune the pretrained models for the 2D semantic segmentation task in Tab. 7. Our approach maintains performance improvement over baseline pre-training methods such as Pri3D (+11.3 mIoU) and MAE-unsupIN→SN (+1.7 mIoU). This indicates an encouraging transferability of our learned representations and their applicability to a variety of scenarios. Please refer to the supplemental material for more out-of-domain transfer results on more generally distributed data, such as ADE20K [46]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.666, + 0.693 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.792 + ], + "angle": 0, + "content": "Does the pre-training masking ratio matter? We show how different masking ratios influence downstream task results in Tab. 8 on ScanNet semantic segmentation. We found a performance gain when masking more RGB values (keeping \\(20\\%\\)), which in combination with the heavy depth masking (keeping \\(20\\%\\)) leads to the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.801, + 0.894, + 0.862 + ], + "angle": 0, + "content": "What about other ViT variants? In our experiments, we use ViT-B as the meta-architecture. We show Mask3D also works in other ViT variants, such as ViT-L (see Tab. 12), which exhibits a similar trend of improvements." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Does the normalization in the reconstruction loss help? We normalize the features when computing the reconstruc" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13516" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.089, + 0.396, + 0.244 + ], + "angle": 0, + "content": "
RGB RatioDepth RatiomIoU
20.0%0.0%65.2
20.0%20.0%66.7
20.0%80.0%65.5
50.0%20.0%65.9
50.0%50.0%64.7
80.0%20.0%64.8
80.0%50.0%64.8
100.0%0.0%64.6
100.0%20.0%64.8
100.0%100.0%64.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.249, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Table 8. Ablation Study of Masking Ratios. on ScanNet 2D semantic segmentation. We mask out different ratios of RGB and depth patches, where the ratio indicates the percentage of kept patches. Refer to supplemental material for a full list." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.314, + 0.469, + 0.344 + ], + "angle": 0, + "content": "tion loss and observe an improvement of \\(+0.8\\%\\) mIoU in the semantic segmentation task on ScanNet." + }, + { + "type": "table", + "bbox": [ + 0.138, + 0.357, + 0.406, + 0.416 + ], + "angle": 0, + "content": "
MethodBackbonemIoU
Train from ScratchViT32.6
MAEViT37.1
Mask3DViT42.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.419, + 0.47, + 0.461 + ], + "angle": 0, + "content": "Table 9. Results on ScanNet Semantic Segmentation without ImageNet pre-training. Similar trend is seen as ImageNet pretraining. The Gap gets larger compared to ImageNet pre-training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.468, + 0.469, + 0.559 + ], + "angle": 0, + "content": "Compared to a pure depth prediction baseline. In Tab. 8, we demonstrate a superior performance with a \\(20\\%\\) kept patches of RGB and depth, compared to a pure depth prediction method (66.7 vs. 64.6). Note in the table, pretraining with \\(100\\%\\) RGB ratio and \\(0\\%\\) depth ratio is equivalent to a pure depth prediction from a RGB image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.469, + 0.645 + ], + "angle": 0, + "content": "Color + depth reconstruction? We found that having joint losses on color and depth during pre-training does not benefit performance (see the following Tab. 10). The RGB reconstruction loss potentially makes pre-training easier, as we already have additional depth priors as guidance." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.658, + 0.391, + 0.702 + ], + "angle": 0, + "content": "
MethodReconstructionmIoU
Mask3DRGB+Depth65.6
Mask3DDepth66.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.706, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Table 10. ScanNet Semantic Segmentation. RGB as an additional signal does not bring a significant improvement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.741, + 0.469, + 0.862 + ], + "angle": 0, + "content": "No Stage-I pre-training. We observe a performance drop without ImageNet pre-training model as initialization for our pre-training. Since ImageNet pre-training is readily available and ScanNet has a relatively small amount of indoor data, we make ImageNet pre-training initialization as default, similar to Pri3D. Meanwhile, we conduct experiments without ImageNet pre-training in Tab. 9, and observe similar trends as when using ImageNet pre-training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "RGB + semantic segmentation as pre-training. Using RGB and semantic segmentation for pre-training rather than" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.089, + 0.847, + 0.148 + ], + "angle": 0, + "content": "
DatasetsMask3D - SemanticsMask3D
ScanNet65.966.7
NYUv255.556.9
CityScapes63.066.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.15, + 0.892, + 0.178 + ], + "angle": 0, + "content": "Table 11. Semantic segmentation results (mIoU). \"Mask3D - Semantics\" denotes pre-training using RGB+Semantics." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.188, + 0.631, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.188, + 0.761, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.188, + 0.89, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.264, + 0.631, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.264, + 0.761, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.264, + 0.89, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.346, + 0.892, + 0.388 + ], + "angle": 0, + "content": "Figure 6. Pre-trained ViT learns 3D structural priors. Our proposed pre-training method learns spatial structures from heavily masked RGB images." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.397, + 0.86, + 0.513 + ], + "angle": 0, + "content": "
Pre-training MethodBackbonemIoU
Pri3D [25]ResNet-5060.2
Pri3D [25]ViT-B59.3 (-0.9)
MAE-unsupIN→SN [19]ViT-B63.3 (+3.1)
Mask3D (Ours)ViT-B66.7 (+6.5)
Pri3D [25]ViT-L64.3 (+4.1)
MAE-unsupIN→SN [19]ViT-L68.2 (+8.0)
Mask3D (Ours)ViT-L70.7 (+10.5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Table 12. ViT Variants on ScanNet 2D Semantic Segmentation. Mask3D yields consistent improvements for both ViT-B and ViT-L backbone architectures." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.57, + 0.891, + 0.661 + ], + "angle": 0, + "content": "depth completion achieved competitive results on ScanNet semantic segmentation, although this requires the use of semantic labels for the pre-training dataset, and is likely to be less transferable across domains than using depth completion. As shown in the following Tab. 11, the gap becomes larger when transferring to both NYUv2 and Cityscapes." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.678, + 0.618, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present Mask3D, a new self-supervised approach to embed 3D priors into learned 2D representations for image scene understanding. We leverage existing large-scale RGB-D data to learn 3D priors without requiring any camera pose or multi-view correspondence information, instead learning geometric and structural cues through a pre-text reconstruction task from masked color and depth. We show that Mask3D is particularly effective in pre-training the modern, powerful ViT backbones, with notable improvements across a variety of image-based tasks and datasets. We believe this shows the strong potential in effectively learning 3D priors and provides new avenues for such 3D-grounded representation learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13517" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In ICCV, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 348-367. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.313 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.386, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.497 + ], + "angle": 0, + "content": "[6] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. arXiv preprint arXiv:2112.02990, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.542 + ], + "angle": 0, + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.47, + 0.669 + ], + "angle": 0, + "content": "[9] Manuel Dahnert, Ji Hou, Matthias Nießner, and Angela Dai. Panoptic 3d scene reconstruction from a single rgb image. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[11] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-468, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[15] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6433-6442, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[16] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[17] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.309, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[18] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-aware 3D instance segmentation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.352, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.41, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.453, + 0.892, + 0.493 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[22] Ji Hou, Angela Dai, and Matthias Nießner. 3D-SIS: 3D Semantic Instance Segmentation of RGB-D Scans. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[23] Ji Hou, Angela Dai, and Matthias Nießner. **RevealNet: Seeing Behind Objects in RGB-D Scans.** In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[24] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.614, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[25] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In ICCV, 2021. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[26] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[27] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[29] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3D object detection in point clouds. ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13518" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[31] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[32] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[34] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGB-D images. ECCV, 2012. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[35] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite. In CVPR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. KPConv: Flexible and deformable convolution for point clouds. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[37] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[38] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[39] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas J Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3D point cloud understanding. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[40] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3D instance segmentation on point clouds. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[41] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas Guibas. GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[42] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. arXiv preprint arXiv:2303.08231, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[43] Yu Zhang, Junle Yu, Xiaolin Huang, Wenhui Zhou, and Ji Hou. Pcr-cg: Point cloud registration via deep explicit color and geometry. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 443-459. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[44] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In European Conference on Computer Vision, pages 311-329. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13519" + } + ] +] \ No newline at end of file diff --git a/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_origin.pdf b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e10a241fb5eb8c65947776d38e590a8595fddf5e --- /dev/null +++ b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/2e3186da-df36-4393-adc4-0fa05e22b0e3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1478bef74d836c60d5b319ba3e38e6aa1f51240a5cf5b992c01fad7c53b4e53b +size 4366809 diff --git a/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/full.md b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0921acad2122c97d6e1d93000d980f013ff4ffeb --- /dev/null +++ b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/full.md @@ -0,0 +1,288 @@ +# Mask3D: Pre-training 2D Vision Transformers by Learning Masked 3D Priors + +Ji Hou $^{1}$ Xiaoliang Dai $^{1}$ Zijian He $^{1}$ Angela Dai $^{2}$ Matthias Nießner $^{2}$ + +$^{1}$ Meta Reality Labs $^{2}$ Technical University of Munich + +![](images/294f8ce22f3c51be1180ed92dff051729bcb6668902d19a81160b7e7e1602c6c.jpg) +Pre-text Task + +![](images/babdccce19462cd31a941615100d97b2a4e192b7d2874d2f2a3b4893870d21dd.jpg) + +![](images/30246612520851c198dfb1d75ca6882d1aa6aaac19c3384f02eceef7a70f8597.jpg) + +![](images/c0e998b87a3f7e0889f2e41d6b518702c5be306ded28fa007fcf31779d30086a.jpg) + +![](images/14e45ed55fdb4020a244be26d6ba9c967f0d58df86546eccae2801124c2016a6.jpg) +Scene Understanding Tasks + +![](images/8b166da63fcf88b310c38ae24a402d3fc0b067e17ac3e5890ae6d1a5bab0afd5.jpg) +Figure 1. We present Mask3D, which learns to embed 3D priors to 2D representations for image understanding tasks, based on a self-supervised pre-training formulation from single RGB-D views, without requiring any camera pose or multi-view correspondence information. Our pre-training takes masked RGB and depth patches as input to reconstruct the dense depth map, and the pre-trained color backbone is used to fine-tune various downstream image understanding tasks. This results in effective ViT pre-training for a variety of downstream tasks and datasets. + +![](images/756f1963bd8c1c790262dc066e3f5adf68dee1fd1a180c4212b19baa97cb4729.jpg) +Instance Segmentation + +![](images/21fa2419c37ca942862ff3b050beba1500a6d796624183411b868b977efa95d9.jpg) +Semantic Segmentation + +![](images/c8a172a3aece2f73bbd956519ad499ab652a31138813cec0336ef3f6cce6eae7.jpg) +Object Detection + +# Abstract + +Current popular backbones in computer vision, such as Vision Transformers (ViT) and ResNets are trained to perceive the world from 2D images. However, to more effectively understand 3D structural priors in 2D backbones, we propose Mask3D to leverage existing large-scale RGB-D data in a self-supervised pre-training to embed these 3D priors into 2D learned feature representations. In contrast to traditional 3D contrastive learning paradigms requiring 3D reconstructions or multi-view correspondences, our approach is simple: we formulate a pre-text reconstruction task by masking RGB and depth patches in individual RGB-D frames. We demonstrate the Mask3D is particularly effective in embedding 3D priors into the powerful 2D ViT backbone, enabling improved representation learning for various scene understanding tasks, such as semantic segmentation, instance segmentation and object detection. + +Experiments show that Mask3D notably outperforms existing self-supervised 3D pre-training approaches on ScanNet, NYUv2, and Cityscapes image understanding tasks, with an improvement of $+6.5\%$ mIoU against the state-of-the-art Pri3D on ScanNet image semantic segmentation. + +# 1. Introduction + +Recent years have seen remarkable advances in 2D image understanding as well as 3D scene understanding, although their representation learning has generally been treated separately. Powerful 2D architectures such as ResNets [21] and Vision Transformers (ViT) [14] have achieved notable success in various 2D recognition and segmentation tasks, but focus on learning from 2D image data. Current large-scale RGB-D datasets [1,4,10,34,35] provide an opportunity to learn key geometric and structural priors to provide more informed reasoning about the scale and cir + +cumvent view-dependent effects, which can provide more efficient representation learning. In 3D, various successful methods have been leveraging the RGB-D datasets for constructive point discrimination [6,24,39,44] for downstream 3D tasks, including high-level scene understanding tasks as well as low-level point matching tasks [15,42,43]. However, the other direction from 3D to 2D is less explored. + +We thus aim to embed such 3D priors into 2D backbones to effectively learn the structural and geometric priors underlying the 3D scenes captured in 2D image projections. Recently, Pri3D [25] adopted similar multi-view and reconstruction-based constraints to induce 3D priors in learned 2D representations. However, this relies on not only acquiring RGB-D frame data but also the robust registration of multiple views to obtain camera pose information for each frame. Instead, we consider how to effectively learn such geometric priors from only single-view RGB-D data in a more broadly applicable setting for 3D-based pre-training. + +We thus propose Mask3D, which learns effective 3D priors for 2D backbones in a self-supervised fashion by pretraining with single-view RGB-D frame data. We propose a pre-text reconstruction task to reconstruct the depth map by masking different random RGB and depth patches of an input frame. These masked input RGB and depth are encoded simultaneously in separate encoding branches and decoded to reconstruct the dense depth map. This imbues 3D priors into the RGB backbone which can then be used for fine-tuning downstream image based scene understanding tasks. + +In particular, our self-supervised approach to embedding 3D priors from single-view RGB-D data to 2D learned features is not only more generally applicable, but we also demonstrate that it is particularly effective for pre-training vision transformers. Our experiments demonstrate the effectiveness of Mask3D on a variety of datasets and image understanding tasks. We pre-train on ScanNet [10] with our masked 3D pre-training paradigm and fine-tune for 2D semantic segmentation, instance segmentation, and object detection. This enables notable improvements not only on ScanNet data but also generalizes to NYUv2 [34] and even Cityscapes [8] data. We believe that Mask3D makes an important step to shed light on the paradigm of incorporating 3D representation learning to powerful 2D backbones. + +In summary, our contributions are: + +- We introduce a self-supervised pre-training approach to learn masked 3D priors for 2D image understanding tasks based on learning from only single-view RGB-D data, without requiring any camera pose or 3D reconstruction information, and thus enabling more general applicability. +- We demonstrate that our masked depth reconstruction pre-training is particularly effective for the modern, powerful ViT architecture, across a variety of datasets and image understanding tasks. + +# 2. Related Work + +Pre-training in Visual Transformers. Recently, visual transformers have revolutionized computer vision and attracted wide attention. In contrast to popular CNNs that operate in a sliding window fashion, Vision Transformers (ViT) describe the image as patches of 16x16 pixels. The Swin Trasnformer [28] has set new records with its hierarchical transformer formulation on major vision benchmarks. The dominance of visual transformers in many vision tasks has inspired study into how to pre-training such backbones. MoCoV3 [5] first investigated the effects of several fundamental components for self-supervised ViT training. MAE [19] then proposed an approach inspired by BERT [13], which randomly masks words in sentences and leveraged masked image reconstruction for self-supervised pre-training that achieved state-of-the-art results in ViT. A similar self-supervision has also been proposed by MaskFeat [37] for self-supervised video pre-training. MaskFeat randomly masks out pixels of the input sequence and then predicts the Oriented Gradients (HOG) of the masked regions. However, such ViT pre-training methods focus on image or video data, without exploring how 3D priors can potentially be exploited. MultiMAE [2] on the other hand introduces depth priors. However, it requires depth as input not only in pre-training but also in downstream tasks. In addition to depth, human annotations (e.g., semantics) are also leveraged in the pre-training. To achieve a self-supervised pre-training, we do not use semantics in the pre-training and only use RGB images as input in downstream tasks. + +RGB-D Scene Understanding. Research in 3D scene understanding have been spurred forward with the introduction of larger-scale, annotated real-world RGB-D datasets [1, 4, 10]. This has enabled data-driven semantic understanding of 3D reconstructed environments, where we have now seen notable progress, such as for 3D semantic segmentation [7, 11, 17, 31, 32, 36], object detection [29, 30, 45], instance segmentation [16, 18, 22, 23, 26, 27, 40, 41], and recently panoptic segmentation [9]. Such 3D scene understanding tasks have been analogously defined to 2D image understanding, which considers RGB-only input without depth information. However, learning from 3D enables geometric reasoning without requiring learning view-dependent effects or resolving depth/scale ambiguity that must be learned when considering solely 2D data. We thus take advantage of existing large-scale RGB-D data to explore how to effectively embed 3D priors for better representation learning for 2D scene understanding tasks. + +Embedding 3D Priors in 2D Backbones. Learning cross-modality features has been seen in extensive studies of the ties between languages and images. In particular, CLIP [33] learns visual features from language supervision during pre-training, showing promising results in zero- + +shot learning for image classification. Pri3D [25] explores 3D-based pre-training for image-based tasks by leveraging multi-view consistency and 2D-3D correspondence with contrastive learning to embed 3D priors into ResNet backbones. This results in enhanced features over ImageNet pretraining on 2D scene understanding tasks. However, Pri3D requires camera pose registration across RGB-D video sequences and is specifically designed for CNNs-based architectures. In contrast, we formulate a self-supervised pretraining that operates on only single-view RGB-D frames and leverages masked 3D priors that can effectively pretrain powerful ViT backbones. + +# 3. Method + +We introduce Mask3D to embed 3D priors into learned 2D representations by self-supervised pre-training from only single-view RGB-D frames. To effectively learn 3D structural priors without requiring any camera pose information or multi-view constraints, we formulate a pre-text depth reconstruction task to inform the RGB feature extraction to be geometrically aware. Randomly masked color and depth images are used as input to reconstruct the dense depth map, and the RGB backbone can then be used to fine-tune downstream image understanding tasks. In particular, we show in Sec. 4 that this single-frame self-supervision is particularly well-suited for powerful vision transformer (ViT) backbones, even without any multi-view information. + +# 3.1. Learning Masked 3D Priors + +We propose to learn masked 3D priors to embed to learned 2D backbones by pre-training to reconstruct dense depth from RGB images with the guidance of sparse depth. That is, for an RGB-D frame $F = (C, D)$ with RGB image $C$ and depth map $D$ , we train to reconstruct $D$ from masked patches of $C$ guided with sparse masked patches of $D$ . An overview of our approach is shown in Fig. 2. + +To create masked color and depth $M_{c}$ and $M_{d}$ from $C$ and $D$ as input for reconstruction, a 240x320 RGB image $C$ is uniformly divided into 300 16x16 patches, from which we randomly keep a percentage $p_{c}$ of patches, masking out the others, to obtain $M_{c}$ . $M_{d}$ is created similarly by keeping only a percentage $p_{d}$ of patches, such that the resulting depth patches do not coincide with the RGB patches in $M_{c}$ . + +We then train color and depth encoders $\Psi_{c}$ and $\Psi_{d}$ to separately encode RGB and depth signals. RGB patches are fed into $\Psi_{c}$ and concatenated with a positional embedding, following the ViT architecture, and similarly for depth. The positional embedding used encodes the patch location by a cosine function. Patches and their positional embeddings are then mapped into higher dimensional feature vectors via $\Psi_{c}$ and $\Psi_{d}$ . The encoders $\Psi_{c}$ and $\Psi_{d}$ are built by blocks composed of linear and norm layers. The features from $\Psi_{c}$ and $\Psi_{d}$ are then fused in the bottleneck; since depth patches + +were selected in regions where no RGB patches were selected, there are no duplicate patches representing the same patch location. + +For those regions which do not have any associated RGB or depth patch, we use patches of constant values as mask tokens to create a placeholder in the bottleneck to enable reconstructing dense depth at the original image resolution. In the bottleneck, the RGB and depth patch feature vectors, along with the mask tokens, form the input to the decoder. This formulates a reconstruction task from sparse RGB and depth; the joint RGB-D pre-training enables reconstruction from very sparse input, as shown by our ablation on the masked input rations in Sec. 4.5. Note that the depth encoder is trained only during pre-training, and only the color ViT encoder (and decoder, if applicable) are used for downstream fine-tuning. + +To demonstrate the effectiveness of the pre-training task, we demonstrate the depth completion results from the pretraining phase in Fig. 6. A detailed analysis of masking different ratios of color and depth signals is shown in Sec. 4.5. + +Pre-training Loss In contrast to the widely used contrastive loss in 3D representation learning, we train for dense depth reconstruction with an $\ell_2$ reconstruction loss. Similar to MAE [19], we normalize the output patches as well as the target patches prior to computing the loss, which we found to empirically improve the performance. + +# 4. Results + +We demonstrate the effectiveness of Mask3D pretraining for ViT [14] backbones on semantic segmentation, instance segmentation, and object detection tasks. We pre-train on ScanNet [10] data and demonstrate the effectiveness of learned masked 3D priors for not only ScanNet downstream tasks but also their transferability to NYUv2 [34] and even across the indoor/outdoor domain gap to Cityscapes [8] data. + +# 4.1. Experimental Setup + +In this section, we introduce the pre-training and finetuning procedures. Our method uses a two-stage pretraining design introduced in the following. + +Stage-I: Mask3D Encoder Initialization. We initialize the RGB encoder with network weights trained on ImageNet [12] (as pre-training for our pre-training). To maintain a fully self-supervised pre-training paradigm, we initialize with weights obtained by self-supervised ImageNet pre-training [19]. + +Stage-II: Mask3D Pre-training on ScanNet. Mask3D pre-training leverages 3D priors in RGB-D frame data, for which we use the color and depth maps of ScanNet [10]. Note that this does not use any semantic or reconstruction + +![](images/dc8cb09a518a2067cae2580897a845ac376e471c53f11e00667b29b7af8f27c8.jpg) +Figure 2. Overview of Mask3D Pre-training. As a pretext task, we predict dense depth from color and sparse depth signals. We use masked input by randomly selecting a set of patches from the input color image, which are then mapped to higher dimensional feature vectors; input depth is similarly masked and encoded. The color and depth features are then fused into a bottleneck from which dense depth is reconstructed as a self-supervised loss. + +information during pre-training. ScanNet contains 2.5M RGB-D frames from 1513 ScanNet train video sequences. We regularly sample every $25^{\text{th}}$ frame without any other filtering (e.g., no control on viewpoint variation). + +Downstream Fine-tuning. We evaluate our Mask3D pretraining by fine-tuning a variety of downstream image understanding tasks (semantic segmentation, instance segmentation, object detection). We consider in-domain transfers on ScanNet image understanding, and further evaluate the out-of-domain transfer on datasets with different statistical characteristics: the indoor image data of NYUv2 [34], as well as across a strong domain gap to the outdoor image data of Cityscapes [8]. For semantic segmentation tasks, we use both encoder and decoder pre-trained with Mask3D, and for instance segmentation and detection, only the backbone encoder is pre-trained. + +Baselines. To evaluate the effectiveness of our learned masked 3D priors for 2D representations, we benchmark our method against relevant baselines: + +Supervised ImageNet Pre-training (supIN). This uses the pre-trained weights from ImageNet, provided by torchvision, as is commonly used for image understanding tasks. Here, only ImageNet data is used, and no ScanNet data is involved in the pre-training phase. + +2-Stage MoCoV2 (MoCoV2-supIN $\rightarrow$ SN). Supervised ImageNet pre-trained (supIN) weights are used as network initialization for pre-training. MoCoV2 [20] is used for pre-training with randomly shuffled ScanNet images. Here, both ImageNet and ScanNet image data are used without any geometric priors. + +2-Stage MAE (MAE-unsupIN $\rightarrow$ SN). Self-supervised Im + +ageNet pre-trained weights are used as network initialization for pre-training. MAE [19] is used for pre-training with randomly shuffled ScanNet images. Here, both ImageNet and ScanNet image data are used without any geometric priors. + +Pri3D [25]. Supervised ImageNet pre-trained are used to initialize Pri3D pre-training, which leverages multi-view and reconstruction constraints from ScanNet data under a contrastive loss. Here, both ImageNet and ScanNet data are used, incorporating 3D priors from reconstructed RGB-D video sequences for pre-training. + +Implementation Details. We use a ViT-B backbone to train our approach. For pre-training, we use an SGD optimizer with a learning rate of 0.1 and an effective batch size of 128 (accumulated gradients from an actual batch size of 64). The learning rate is decreased by a factor of 0.99 every 1000 steps, and our method is trained for 100 epochs. Fine-tuning on semantic segmentation is trained with a batch size of 8 for 80 epochs. The initial learning rate is 0.01, with polynomial decay with a power of 0.9. Finetuning on detection and instance segmentation is trained using Detector2 [38] with the 1x schedule. Pre-training experiments are conducted on a single NVIDIA A6000 GPU, or 2 NVIDIA RTX3090 GPUs, or 4 NVIDIA RTX2080Ti GPUs; semantic segmentation experiments are conducted on a single NVIDIA A6000 GPU; instance segmentation and detection experiments are conducted on 8 V100 GPUs. + +# 4.2. ScanNet Downstream Tasks + +We demonstrate the effectiveness of representation learning with 3D priors via downstream tasks on ScanNet [10] images. For fine-tuning, we follow the standard + +
Pre-training MethodBackbonePre-training DatamIoU
ScratchResNet-50None39.1
ImageNet Pre-training (supIN)ResNet-50ImageNet55.7
Supervised Pre-trainingViTImageNet+ScanNet65.9 (+10.2)
MoCoV2-supIN→SN [20]ResNet-50ImageNet+ScanNet56.6 (+0.9)
Pri3D [25]ResNet-50ImageNet+ScanNet60.2 (+4.5)
Pri3D [25]ViTImageNet+ScanNet59.3 (+3.6)
DINO [3]ViTImageNet+ScanNet58.1 (+3.6)
MAE-unsupIN→SN [19]ViTImageNet+ScanNet63.3 (+7.6)
Ours - Mask3D (DINO)ViTImageNet+ScanNet60.5 (+4.8)
Ours - Mask3D (MAE)ViTImageNet+ScanNet66.7 (+11.0)
+ +Table 1. ScanNet 2D Semantic Segmentation. Mask3D significantly outperforms Pri3D as well as other state-of-the-art pre-training approaches that leverage both ImageNet and ScanNet data. + +
Pre-training MethodAP@0.5AP@0.75AP
Scratch32.717.716.9
ImageNet Pretrain (supIN)41.725.925.1
MoCoV2-supIN→SN [20]43.5 (+1.8)26.8 (+0.9)25.8 (+0.7)
Pri3D [25]43.7 (+2.0)27.0 (+1.1)26.3 (+1.2)
MAE-unsupIN→SN [19]46.1 (+4.4)32.7 (+6.8)30.5 (+5.4)
Mask3D (Ours)50.4 (+8.7)35.3 (+9.4)32.7 (+7.6)
+ +Table 2. ScanNet 2D Object Detection. Fine-tuning with Mask3D pre-trained models leads to improved object detection results across different metrics, in comparison to ImageNet pretraining, MoCo-style pre-training, and a strong MAE-style pre-training method. + +
Pre-training MethodAP@0.5AP@0.75AP
Scratch25.813.112.2
ImageNet Pretrain (supIN)32.617.817.6
MoCoV2-supIN→SN [20]33.9 (+1.3)18.1 (+0.3)18.3 (+0.7)
Pri3D [25]34.3 (+1.7)18.7 (+0.9)18.3 (+0.7)
MAE-unsupIN→SN [19]37.4 (+4.8)20.3 (+2.5)20.7 (+3.1)
Mask3D (Ours)41.2 (+8.6)22.7 (+4.9)22.8 (+5.2)
+ +Table 3. ScanNet 2D Instance Segmentation. Fine-tuning with Mask3D pre-trained models leads to improved instance segmentation results across different metrics compared to ImageNet pre-training, MoCo-style pre-training, and a strong MAE-style pre-training method. + +
Pre-training MethodAP@0.5AP@0.75AP
Scratch17.29.28.8
ImageNet Pretrain (supIN)25.113.913.4
MoCoV2-supIN→SN [20]27.2 (+2.1)14.7 (+0.2)14.8 (+1.4)
Pri3D [25]28.1 (+3.0)15.7 (+1.8)15.7 (+2.3)
MAE unsupIN→SN [19]33.6 (+8.5)19.0 (+5.1)19.0 (+5.6)
Mask3D (Ours)37.0 (+11.9)21.6 (+7.7)21.3 (+7.9)
+ +Table 4. NYUv2 2D Instance Segmentation. Fine-tuning with Mask3D pre-trained models leads to improved instance segmentation results across different metrics compared to previous methods, demonstrating the cross-dataset transfer ability of Mask3D. + +protocol of the ScanNet benchmark [10] and sample every $100^{\mathrm{th}}$ frame, resulting in 20,000 train images and 5,000 validation images. + +2D Semantic Segmentation. Tab. 1 shows the fine-tuning for semantic segmentation, in comparison with baseline pre-training approaches. All pre-training methods significantly improve performance over training the semantic segmentation model from scratch. In particular, Mask3D provides substantially better representation quality leading to a much stronger improvement over supervised ImageNet pretraining $(+11\mathrm{mIoU})$ , and notably improving over MAE unsupIN $\rightarrow$ SN with ImageNet and ScanNet $(+3.4\mathrm{mIoU})$ and the 3D-based pre-training of Pri3D $(+6.5\mathrm{mIoU})$ . We note that the multi-view 3D pre-training of Pri3D does not effectively embed informative 3D priors to ViT backbones, rather suffering from performance degradation from a ResNet-50 backbone. In contrast, our Mask3D pretraining can notably improve performance with a ViT backbone, indicating the effectiveness of our learned 3D priors. + +2D Object Detection and Instance Segmentation We show that Mask3D provides effective general 3D priors for a variety of image-based tasks, by evaluating downstream object detection and instance segmentation in Tab. 2 and Tab. 3, respectively. Across all tasks, various pre-training approaches yield substantial improvement over training from scratch. Our masked 3D prior learning transfers effectively learned representations for object detection and instance segmentation, notably improving over the best-performing MAE-unsupIN $\rightarrow$ SN (+4.3 AP@0.5 and +3.8 AP@0.5, respectively). + +Data-Efficient Scenarios. We additionally show that our single-view RGB-D pre-training to embed 3D priors in limited data scenarios for downstream ScanNet semantic segmentation in Fig. 5. Mask3D shows consistent improvements across a range of limited data; even with only $20\%$ of the training data, we recover $80\%$ performance with $100\%$ training data available and improving $+15.2$ mIoU over Pri3D pre-training on a ViT backbone. + +# 4.3. NYUv2 Downstream Tasks + +We demonstrate the generalizability of our 3D-imbued learned feature representations across datasets, us + +![](images/3fe270dc8c44a7b29c5a15ad45e7c0f0efbb3a2e532d13043902dff2d5200c37.jpg) +Figure 3. Qualitative Results on Various Tasks across Different Benchmarks. We visualize predictions on different tasks across various scene understanding benchmarks. From top to bottom rows: instance segmentation on ScanNet, instance segmentation on NYUv2, semantic segmentation on NYUv2, and semantic segmentation results in ScanNet. + +![](images/ab080659c103eb936871270160b0774b951d7aeab8fa96ac1d99dc8dbecb50b1.jpg) +Figure 4. More Qualitative Results on Semantic Segmentation. We visualize semantic segmentation predictions on various scene understanding benchmarks including ScanNet and NYUv2. + +ing Mask3D pre-trained on ScanNet and fine-tuned on NYUv2 [34] following the same fine-tuning setup as before. NYUv2 contains Microsoft Kinect RGB-D video + +sequences of indoor scenes, comprising 1449 densely labeled RGB images. We use the official 795/654 train/val split. Tables 5, 6, and 4 evaluate the downstream tasks + +
Pre-training MethodBackbonePre-training DatamIoU
ScratchResNet-50None24.8
ImageNet Pre-training (supIN)ResNet-50ImageNet50.0
Supervised Pre-trainingViTImageNet+ScanNet55.5 (+5.5)
MoCoV2-supIN→SN [20]ResNet-50ImageNet+ScanNet47.6 (-2.4)
Pri3D [25]ResNet-50ImageNet+ScanNet54.2 (+4.2)
Pri3D [25]ViTImageNet+ScanNet53.2 (+3.2)
MAE-unsupIN→SN [19]ViTImageNet+ScanNet54.9 (+4.9)
Mask3D (Ours)ViTImageNet+ScanNet56.9 (+6.9)
+ +![](images/7b2a353062552324ebf8eb7a23902eb9505ec57870a964a21edd760623e9241a.jpg) +Figure 5. Data-Efficient Results. Compared to previous methods, Mask3D demonstrates consistent improvements on ScanNet 2D semantic segmentation across a range of limited data scenarios. Mask3D is particularly effective for ViT pre-training, improving $+15.2\%$ mIoU over state-of-the-art Pri3D [25] on a ViT backbone at $20\%$ of the training data. + +Table 5. NYUv2 2D Semantic Segmentation. Mask3D significantly outperforms state-of-the-art pre-training approaches, demonstrating its effectiveness in transferring to different dataset characteristics. + +
Pre-training MethodAP@0.5AP@0.75AP
Scratch21.310.39.0
ImageNet Pretrain (supIN)29.917.316.8
MoCoV2-supIN→SN [20]30.1 (+0.20)18.1 (+0.80)17.3 (+0.50)
Pri3D [25]33.0 (+2.10)19.8 (+2.60)18.9 (+2.10)
MAE-unsupIN→SN [19]40.3 (+10.4)24.5 (+7.20)23.2 (+6.40)
Mask3D (Ours)44.0 (+14.1)28.3 (+6.40)25.9 (+9.10)
+ +of 2D semantic segmentation, object detection, and instance segmentation, respectively. Across all three tasks on NYUv2 data, our Mask3D pre-training achieves notably improved performance than training from scratch as well as the various baseline pre-training methods. In particular, we achieve an improvement of $+6.9$ mIoU, $+14.1$ AP@0.5, and $+11.9$ AP@0.5 over the common supervised ImageNet pretraining on semantic segmentation, object detection, and instance segmentation. + +Table 6. NYUv2 2D Object Detection. Fine-tuning with Mask3D pre-trained models leads to improved object detection results across different metrics, showing an effective transfer across dataset characteristics. + +
Pre-training MethodBackbonemIoU
ImageNet Pre-training (supIN)ResNet-5054.1
Pri3D [25]ResNet-5055.1 (+1.00)
MAE-unsupIN→SN [19]ViT64.7 (+10.6)
Mask3D (Ours)ViT66.4 (+12.3)
+ +Table 7. Cityscapes 2D Semantic Segmentation. Mask3D significantly outperforms state-of-the-art Pri3D as well as a strong MAE-style pre-training. This demonstrates the effectiveness of the transferability of Mask3D, even under a significant domain gap. + +# 4.4. Out-of-domain Transfer + +While Mask3D concentrates on pre-training for improving indoor scene understanding, we further demonstrate the effectiveness of our Mask3D pre-training for the out-of-domain transfer on outdoor data, such as Cityscapes [8]. We use the official data split of 3000 images for training and 500 images for the test. To evaluate the transferability in such a large domain gap scenario, we fine-tune the pretrained models for the 2D semantic segmentation task in Tab. 7. Our approach maintains performance improvement over baseline pre-training methods such as Pri3D (+11.3 mIoU) and MAE-unsupIN→SN (+1.7 mIoU). This indicates an encouraging transferability of our learned representations and their applicability to a variety of scenarios. Please refer to the supplemental material for more out-of-domain transfer results on more generally distributed data, such as ADE20K [46]. + +# 4.5. Ablation Studies + +Does the pre-training masking ratio matter? We show how different masking ratios influence downstream task results in Tab. 8 on ScanNet semantic segmentation. We found a performance gain when masking more RGB values (keeping $20\%$ ), which in combination with the heavy depth masking (keeping $20\%$ ) leads to the best performance. + +What about other ViT variants? In our experiments, we use ViT-B as the meta-architecture. We show Mask3D also works in other ViT variants, such as ViT-L (see Tab. 12), which exhibits a similar trend of improvements. + +Does the normalization in the reconstruction loss help? We normalize the features when computing the reconstruc + +
RGB RatioDepth RatiomIoU
20.0%0.0%65.2
20.0%20.0%66.7
20.0%80.0%65.5
50.0%20.0%65.9
50.0%50.0%64.7
80.0%20.0%64.8
80.0%50.0%64.8
100.0%0.0%64.6
100.0%20.0%64.8
100.0%100.0%64.5
+ +tion loss and observe an improvement of $+0.8\%$ mIoU in the semantic segmentation task on ScanNet. + +Table 8. Ablation Study of Masking Ratios. on ScanNet 2D semantic segmentation. We mask out different ratios of RGB and depth patches, where the ratio indicates the percentage of kept patches. Refer to supplemental material for a full list. + +
MethodBackbonemIoU
Train from ScratchViT32.6
MAEViT37.1
Mask3DViT42.2
+ +Compared to a pure depth prediction baseline. In Tab. 8, we demonstrate a superior performance with a $20\%$ kept patches of RGB and depth, compared to a pure depth prediction method (66.7 vs. 64.6). Note in the table, pretraining with $100\%$ RGB ratio and $0\%$ depth ratio is equivalent to a pure depth prediction from a RGB image. + +Color + depth reconstruction? We found that having joint losses on color and depth during pre-training does not benefit performance (see the following Tab. 10). The RGB reconstruction loss potentially makes pre-training easier, as we already have additional depth priors as guidance. + +Table 9. Results on ScanNet Semantic Segmentation without ImageNet pre-training. Similar trend is seen as ImageNet pretraining. The Gap gets larger compared to ImageNet pre-training. + +
MethodReconstructionmIoU
Mask3DRGB+Depth65.6
Mask3DDepth66.7
+ +No Stage-I pre-training. We observe a performance drop without ImageNet pre-training model as initialization for our pre-training. Since ImageNet pre-training is readily available and ScanNet has a relatively small amount of indoor data, we make ImageNet pre-training initialization as default, similar to Pri3D. Meanwhile, we conduct experiments without ImageNet pre-training in Tab. 9, and observe similar trends as when using ImageNet pre-training. + +RGB + semantic segmentation as pre-training. Using RGB and semantic segmentation for pre-training rather than + +Table 10. ScanNet Semantic Segmentation. RGB as an additional signal does not bring a significant improvement. + +
DatasetsMask3D - SemanticsMask3D
ScanNet65.966.7
NYUv255.556.9
CityScapes63.066.4
+ +![](images/ae5dee2d82c18abc68501d66a71e68178ba1d44ff3f760b832606fc81b5e0c10.jpg) + +![](images/eb7469c911f93ad67b9482aeaaf11ff8ebb6d53280d622c65f786515b1518ce1.jpg) + +![](images/f3ec038a4326ee242fd96bb273d9b5b6a362828e63f27906cb952b7ff3142598.jpg) + +![](images/b2da07e039d5b83d4b0ecb61be3cec12a41e2cc3b5eb91514a06210e793cac89.jpg) +Figure 6. Pre-trained ViT learns 3D structural priors. Our proposed pre-training method learns spatial structures from heavily masked RGB images. + +![](images/93f69c2a947fb008725e3a0a2582e6f1b68bd82dd9a8268c3f89880376d7d6f6.jpg) + +![](images/4364ded99b9cc21523aa9333edc5219121b5264aae70e3f4094f0f75d53400bf.jpg) + +Table 11. Semantic segmentation results (mIoU). "Mask3D - Semantics" denotes pre-training using RGB+Semantics. + +
Pre-training MethodBackbonemIoU
Pri3D [25]ResNet-5060.2
Pri3D [25]ViT-B59.3 (-0.9)
MAE-unsupIN→SN [19]ViT-B63.3 (+3.1)
Mask3D (Ours)ViT-B66.7 (+6.5)
Pri3D [25]ViT-L64.3 (+4.1)
MAE-unsupIN→SN [19]ViT-L68.2 (+8.0)
Mask3D (Ours)ViT-L70.7 (+10.5)
+ +Table 12. ViT Variants on ScanNet 2D Semantic Segmentation. Mask3D yields consistent improvements for both ViT-B and ViT-L backbone architectures. + +depth completion achieved competitive results on ScanNet semantic segmentation, although this requires the use of semantic labels for the pre-training dataset, and is likely to be less transferable across domains than using depth completion. As shown in the following Tab. 11, the gap becomes larger when transferring to both NYUv2 and Cityscapes. + +# 5. Conclusion + +In this paper, we present Mask3D, a new self-supervised approach to embed 3D priors into learned 2D representations for image scene understanding. We leverage existing large-scale RGB-D data to learn 3D priors without requiring any camera pose or multi-view correspondence information, instead learning geometric and structural cues through a pre-text reconstruction task from masked color and depth. We show that Mask3D is particularly effective in pre-training the modern, powerful ViT backbones, with notable improvements across a variety of image-based tasks and datasets. We believe this shows the strong potential in effectively learning 3D priors and provides new avenues for such 3D-grounded representation learning. + +# References + +[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In ICCV, 2016. 1, 2 +[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 348-367. Springer, 2022. 2 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5 +[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 1, 2 +[5] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2 +[6] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. arXiv preprint arXiv:2112.02990, 2021. 2 +[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2 +[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3, 4, 7 +[9] Manuel Dahnert, Ji Hou, Matthias Nießner, and Angela Dai. Panoptic 3d scene reconstruction from a single rgb image. Advances in Neural Information Processing Systems, 34, 2021. 2 +[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 1, 2, 3, 4, 5 +[11] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-468, 2018. 2 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.1,3 +[15] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6433-6442, 2021. 2 +[16] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation. In CVPR, 2020. 2 +[17] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2 +[18] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-aware 3D instance segmentation. In CVPR, 2020. 2 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4, 5, 7, 8 +[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 4, 5, 7 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1 +[22] Ji Hou, Angela Dai, and Matthias Nießner. 3D-SIS: 3D Semantic Instance Segmentation of RGB-D Scans. In CVPR, 2019. 2 +[23] Ji Hou, Angela Dai, and Matthias Nießner. **RevealNet: Seeing Behind Objects in RGB-D Scans.** In CVPR, 2020. 2 +[24] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2 +[25] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In ICCV, 2021. 2, 3, 4, 5, 7, 8 +[26] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation. In CVPR, 2020. 2 +[27] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In ICCV, 2019. 2 +[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2 +[29] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In CVPR, 2021. 2 +[30] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3D object detection in point clouds. ICCV, 2019. 2 + +[31] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. CVPR, 2017. 2 +[32] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[34] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGB-D images. ECCV, 2012. 1, 2, 3, 4, 6 +[35] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite. In CVPR, 2015. 1 +[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. KPConv: Flexible and deformable convolution for point clouds. In CVPR, 2019. 2 +[37] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021. 2 +[38] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.4 +[39] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas J Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3D point cloud understanding. ECCV, 2020. 2 +[40] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3D instance segmentation on point clouds. In NeurIPS, 2019. 2 +[41] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas Guibas. GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In CVPR, 2019. 2 +[42] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. arXiv preprint arXiv:2303.08231, 2023. 2 +[43] Yu Zhang, Junle Yu, Xiaolin Huang, Wenhui Zhou, and Ji Hou. Pcr-cg: Point cloud registration via deep explicit color and geometry. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 443-459. Springer, 2022. 2 +[44] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2 +[45] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In European Conference on Computer Vision, pages 311-329. Springer, 2020. 2 + +[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7 \ No newline at end of file diff --git a/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/images.zip b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed9988f0bc2ee16cefa38dc3fa3cf12a59f87d7c --- /dev/null +++ b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27e035fb07946b27f22cc7c4e315d196fed793e8b7fd84063b85176877cf72dc +size 858297 diff --git a/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/layout.json b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..932c102029d918eb9664922523731b8ca1165b3f --- /dev/null +++ b/2023/Mask3D_ Pre-Training 2D Vision Transformers by Learning Masked 3D Priors/layout.json @@ -0,0 +1,7768 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 538, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 538, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 538, + 120 + ], + "type": "text", + "content": "Mask3D: Pre-training 2D Vision Transformers by Learning Masked 3D Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "text", + "content": "Ji Hou" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "text", + "content": " Xiaoliang Dai" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "text", + "content": " Zijian He" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "text", + "content": " Angela Dai" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "text", + "content": " Matthias Nießner" + }, + { + "bbox": [ + 110, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 162, + 429, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 162, + 429, + 177 + ], + "spans": [ + { + "bbox": [ + 161, + 162, + 429, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 161, + 162, + 429, + 177 + ], + "type": "text", + "content": "Meta Reality Labs " + }, + { + "bbox": [ + 161, + 162, + 429, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 162, + 429, + 177 + ], + "type": "text", + "content": "Technical University of Munich" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 211, + 198, + 300 + ], + "blocks": [ + { + "bbox": [ + 51, + 200, + 97, + 209 + ], + "lines": [ + { + "bbox": [ + 51, + 200, + 97, + 209 + ], + "spans": [ + { + "bbox": [ + 51, + 200, + 97, + 209 + ], + "type": "text", + "content": "Pre-text Task" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 211, + 198, + 300 + ], + "lines": [ + { + "bbox": [ + 52, + 211, + 198, + 300 + ], + "spans": [ + { + "bbox": [ + 52, + 211, + 198, + 300 + ], + "type": "image", + "image_path": "294f8ce22f3c51be1180ed92dff051729bcb6668902d19a81160b7e7e1602c6c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 200, + 212, + 315, + 300 + ], + "blocks": [ + { + "bbox": [ + 200, + 212, + 315, + 300 + ], + "lines": [ + { + "bbox": [ + 200, + 212, + 315, + 300 + ], + "spans": [ + { + "bbox": [ + 200, + 212, + 315, + 300 + ], + "type": "image", + "image_path": "babdccce19462cd31a941615100d97b2a4e192b7d2874d2f2a3b4893870d21dd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 200, + 424, + 303 + ], + "blocks": [ + { + "bbox": [ + 318, + 200, + 424, + 303 + ], + "lines": [ + { + "bbox": [ + 318, + 200, + 424, + 303 + ], + "spans": [ + { + "bbox": [ + 318, + 200, + 424, + 303 + ], + "type": "image", + "image_path": "30246612520851c198dfb1d75ca6882d1aa6aaac19c3384f02eceef7a70f8597.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 211, + 542, + 299 + ], + "blocks": [ + { + "bbox": [ + 425, + 211, + 542, + 299 + ], + "lines": [ + { + "bbox": [ + 425, + 211, + 542, + 299 + ], + "spans": [ + { + "bbox": [ + 425, + 211, + 542, + 299 + ], + "type": "image", + "image_path": "c0e998b87a3f7e0889f2e41d6b518702c5be306ded28fa007fcf31779d30086a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 52, + 310, + 169, + 398 + ], + "blocks": [ + { + "bbox": [ + 52, + 310, + 169, + 398 + ], + "lines": [ + { + "bbox": [ + 52, + 310, + 169, + 398 + ], + "spans": [ + { + "bbox": [ + 52, + 310, + 169, + 398 + ], + "type": "image", + "image_path": "14e45ed55fdb4020a244be26d6ba9c967f0d58df86546eccae2801124c2016a6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 401, + 141, + 410 + ], + "lines": [ + { + "bbox": [ + 54, + 401, + 141, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 141, + 410 + ], + "type": "text", + "content": "Scene Understanding Tasks" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 173, + 310, + 192, + 403 + ], + "blocks": [ + { + "bbox": [ + 173, + 310, + 192, + 403 + ], + "lines": [ + { + "bbox": [ + 173, + 310, + 192, + 403 + ], + "spans": [ + { + "bbox": [ + 173, + 310, + 192, + 403 + ], + "type": "image", + "image_path": "8b166da63fcf88b310c38ae24a402d3fc0b067e17ac3e5890ae6d1a5bab0afd5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 428, + 545, + 481 + ], + "lines": [ + { + "bbox": [ + 46, + 428, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 428, + 545, + 481 + ], + "type": "text", + "content": "Figure 1. We present Mask3D, which learns to embed 3D priors to 2D representations for image understanding tasks, based on a self-supervised pre-training formulation from single RGB-D views, without requiring any camera pose or multi-view correspondence information. Our pre-training takes masked RGB and depth patches as input to reconstruct the dense depth map, and the pre-trained color backbone is used to fine-tune various downstream image understanding tasks. This results in effective ViT pre-training for a variety of downstream tasks and datasets." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 199, + 311, + 312, + 398 + ], + "blocks": [ + { + "bbox": [ + 199, + 311, + 312, + 398 + ], + "lines": [ + { + "bbox": [ + 199, + 311, + 312, + 398 + ], + "spans": [ + { + "bbox": [ + 199, + 311, + 312, + 398 + ], + "type": "image", + "image_path": "756f1963bd8c1c790262dc066e3f5adf68dee1fd1a180c4212b19baa97cb4729.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 401, + 295, + 410 + ], + "lines": [ + { + "bbox": [ + 220, + 401, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 220, + 401, + 295, + 410 + ], + "type": "text", + "content": "Instance Segmentation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 311, + 425, + 398 + ], + "blocks": [ + { + "bbox": [ + 313, + 311, + 425, + 398 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 425, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 425, + 398 + ], + "type": "image", + "image_path": "21fa2419c37ca942862ff3b050beba1500a6d796624183411b868b977efa95d9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 401, + 411, + 410 + ], + "lines": [ + { + "bbox": [ + 334, + 401, + 411, + 410 + ], + "spans": [ + { + "bbox": [ + 334, + 401, + 411, + 410 + ], + "type": "text", + "content": "Semantic Segmentation" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 426, + 311, + 542, + 398 + ], + "blocks": [ + { + "bbox": [ + 426, + 311, + 542, + 398 + ], + "lines": [ + { + "bbox": [ + 426, + 311, + 542, + 398 + ], + "spans": [ + { + "bbox": [ + 426, + 311, + 542, + 398 + ], + "type": "image", + "image_path": "c8a172a3aece2f73bbd956519ad499ab652a31138813cec0336ef3f6cce6eae7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 401, + 516, + 410 + ], + "lines": [ + { + "bbox": [ + 462, + 401, + 516, + 410 + ], + "spans": [ + { + "bbox": [ + 462, + 401, + 516, + 410 + ], + "type": "text", + "content": "Object Detection" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 491, + 191, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 491, + 191, + 504 + ], + "spans": [ + { + "bbox": [ + 143, + 491, + 191, + 504 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "Current popular backbones in computer vision, such as Vision Transformers (ViT) and ResNets are trained to perceive the world from 2D images. However, to more effectively understand 3D structural priors in 2D backbones, we propose Mask3D to leverage existing large-scale RGB-D data in a self-supervised pre-training to embed these 3D priors into 2D learned feature representations. In contrast to traditional 3D contrastive learning paradigms requiring 3D reconstructions or multi-view correspondences, our approach is simple: we formulate a pre-text reconstruction task by masking RGB and depth patches in individual RGB-D frames. We demonstrate the Mask3D is particularly effective in embedding 3D priors into the powerful 2D ViT backbone, enabling improved representation learning for various scene understanding tasks, such as semantic segmentation, instance segmentation and object detection." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 493, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 493, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 306, + 493, + 545, + 554 + ], + "type": "text", + "content": "Experiments show that Mask3D notably outperforms existing self-supervised 3D pre-training approaches on ScanNet, NYUv2, and Cityscapes image understanding tasks, with an improvement of " + }, + { + "bbox": [ + 306, + 493, + 545, + 554 + ], + "type": "inline_equation", + "content": "+6.5\\%" + }, + { + "bbox": [ + 306, + 493, + 545, + 554 + ], + "type": "text", + "content": " mIoU against the state-of-the-art Pri3D on ScanNet image semantic segmentation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 574, + 386, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 386, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 386, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Recent years have seen remarkable advances in 2D image understanding as well as 3D scene understanding, although their representation learning has generally been treated separately. Powerful 2D architectures such as ResNets [21] and Vision Transformers (ViT) [14] have achieved notable success in various 2D recognition and segmentation tasks, but focus on learning from 2D image data. Current large-scale RGB-D datasets [1,4,10,34,35] provide an opportunity to learn key geometric and structural priors to provide more informed reasoning about the scale and cir" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13510" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "cumvent view-dependent effects, which can provide more efficient representation learning. In 3D, various successful methods have been leveraging the RGB-D datasets for constructive point discrimination [6,24,39,44] for downstream 3D tasks, including high-level scene understanding tasks as well as low-level point matching tasks [15,42,43]. However, the other direction from 3D to 2D is less explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 286, + 287 + ], + "type": "text", + "content": "We thus aim to embed such 3D priors into 2D backbones to effectively learn the structural and geometric priors underlying the 3D scenes captured in 2D image projections. Recently, Pri3D [25] adopted similar multi-view and reconstruction-based constraints to induce 3D priors in learned 2D representations. However, this relies on not only acquiring RGB-D frame data but also the robust registration of multiple views to obtain camera pose information for each frame. Instead, we consider how to effectively learn such geometric priors from only single-view RGB-D data in a more broadly applicable setting for 3D-based pre-training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 286, + 407 + ], + "type": "text", + "content": "We thus propose Mask3D, which learns effective 3D priors for 2D backbones in a self-supervised fashion by pretraining with single-view RGB-D frame data. We propose a pre-text reconstruction task to reconstruct the depth map by masking different random RGB and depth patches of an input frame. These masked input RGB and depth are encoded simultaneously in separate encoding branches and decoded to reconstruct the dense depth map. This imbues 3D priors into the RGB backbone which can then be used for fine-tuning downstream image based scene understanding tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 286, + 574 + ], + "type": "text", + "content": "In particular, our self-supervised approach to embedding 3D priors from single-view RGB-D data to 2D learned features is not only more generally applicable, but we also demonstrate that it is particularly effective for pre-training vision transformers. Our experiments demonstrate the effectiveness of Mask3D on a variety of datasets and image understanding tasks. We pre-train on ScanNet [10] with our masked 3D pre-training paradigm and fine-tune for 2D semantic segmentation, instance segmentation, and object detection. This enables notable improvements not only on ScanNet data but also generalizes to NYUv2 [34] and even Cityscapes [8] data. We believe that Mask3D makes an important step to shed light on the paradigm of incorporating 3D representation learning to powerful 2D backbones." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 574, + 200, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 574, + 200, + 586 + ], + "spans": [ + { + "bbox": [ + 59, + 574, + 200, + 586 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 587, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 587, + 286, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 587, + 286, + 659 + ], + "spans": [ + { + "bbox": [ + 58, + 587, + 286, + 659 + ], + "type": "text", + "content": "- We introduce a self-supervised pre-training approach to learn masked 3D priors for 2D image understanding tasks based on learning from only single-view RGB-D data, without requiring any camera pose or 3D reconstruction information, and thus enabling more general applicability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 666, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 666, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 666, + 286, + 713 + ], + "type": "text", + "content": "- We demonstrate that our masked depth reconstruction pre-training is particularly effective for the modern, powerful ViT architecture, across a variety of datasets and image understanding tasks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 426 + ], + "type": "text", + "content": "Pre-training in Visual Transformers. Recently, visual transformers have revolutionized computer vision and attracted wide attention. In contrast to popular CNNs that operate in a sliding window fashion, Vision Transformers (ViT) describe the image as patches of 16x16 pixels. The Swin Trasnformer [28] has set new records with its hierarchical transformer formulation on major vision benchmarks. The dominance of visual transformers in many vision tasks has inspired study into how to pre-training such backbones. MoCoV3 [5] first investigated the effects of several fundamental components for self-supervised ViT training. MAE [19] then proposed an approach inspired by BERT [13], which randomly masks words in sentences and leveraged masked image reconstruction for self-supervised pre-training that achieved state-of-the-art results in ViT. A similar self-supervision has also been proposed by MaskFeat [37] for self-supervised video pre-training. MaskFeat randomly masks out pixels of the input sequence and then predicts the Oriented Gradients (HOG) of the masked regions. However, such ViT pre-training methods focus on image or video data, without exploring how 3D priors can potentially be exploited. MultiMAE [2] on the other hand introduces depth priors. However, it requires depth as input not only in pre-training but also in downstream tasks. In addition to depth, human annotations (e.g., semantics) are also leveraged in the pre-training. To achieve a self-supervised pre-training, we do not use semantics in the pre-training and only use RGB images as input in downstream tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 432, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 647 + ], + "type": "text", + "content": "RGB-D Scene Understanding. Research in 3D scene understanding have been spurred forward with the introduction of larger-scale, annotated real-world RGB-D datasets [1, 4, 10]. This has enabled data-driven semantic understanding of 3D reconstructed environments, where we have now seen notable progress, such as for 3D semantic segmentation [7, 11, 17, 31, 32, 36], object detection [29, 30, 45], instance segmentation [16, 18, 22, 23, 26, 27, 40, 41], and recently panoptic segmentation [9]. Such 3D scene understanding tasks have been analogously defined to 2D image understanding, which considers RGB-only input without depth information. However, learning from 3D enables geometric reasoning without requiring learning view-dependent effects or resolving depth/scale ambiguity that must be learned when considering solely 2D data. We thus take advantage of existing large-scale RGB-D data to explore how to effectively embed 3D priors for better representation learning for 2D scene understanding tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Embedding 3D Priors in 2D Backbones. Learning cross-modality features has been seen in extensive studies of the ties between languages and images. In particular, CLIP [33] learns visual features from language supervision during pre-training, showing promising results in zero-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13511" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "shot learning for image classification. Pri3D [25] explores 3D-based pre-training for image-based tasks by leveraging multi-view consistency and 2D-3D correspondence with contrastive learning to embed 3D priors into ResNet backbones. This results in enhanced features over ImageNet pretraining on 2D scene understanding tasks. However, Pri3D requires camera pose registration across RGB-D video sequences and is specifically designed for CNNs-based architectures. In contrast, we formulate a self-supervised pretraining that operates on only single-view RGB-D frames and leverages masked 3D priors that can effectively pretrain powerful ViT backbones." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 225, + 104, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 104, + 238 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 104, + 238 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 246, + 289, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 289, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 289, + 402 + ], + "type": "text", + "content": "We introduce Mask3D to embed 3D priors into learned 2D representations by self-supervised pre-training from only single-view RGB-D frames. To effectively learn 3D structural priors without requiring any camera pose information or multi-view constraints, we formulate a pre-text depth reconstruction task to inform the RGB feature extraction to be geometrically aware. Randomly masked color and depth images are used as input to reconstruct the dense depth map, and the RGB backbone can then be used to fine-tune downstream image understanding tasks. In particular, we show in Sec. 4 that this single-frame self-supervision is particularly well-suited for powerful vision transformer (ViT) backbones, even without any multi-view information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 201, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 201, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 201, + 420 + ], + "type": "text", + "content": "3.1. Learning Masked 3D Priors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": "We propose to learn masked 3D priors to embed to learned 2D backbones by pre-training to reconstruct dense depth from RGB images with the guidance of sparse depth. That is, for an RGB-D frame " + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "inline_equation", + "content": "F = (C, D)" + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": " with RGB image " + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": " and depth map " + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": ", we train to reconstruct " + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": " from masked patches of " + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": " guided with sparse masked patches of " + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": ". An overview of our approach is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": "To create masked color and depth " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "M_{d}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " as input for reconstruction, a 240x320 RGB image " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " is uniformly divided into 300 16x16 patches, from which we randomly keep a percentage " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "p_{c}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " of patches, masking out the others, to obtain " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "M_{d}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " is created similarly by keeping only a percentage " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "p_{d}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " of patches, such that the resulting depth patches do not coincide with the RGB patches in " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "We then train color and depth encoders " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{c}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{d}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " to separately encode RGB and depth signals. RGB patches are fed into " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{c}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and concatenated with a positional embedding, following the ViT architecture, and similarly for depth. The positional embedding used encodes the patch location by a cosine function. Patches and their positional embeddings are then mapped into higher dimensional feature vectors via " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{c}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{d}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". The encoders " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{c}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{d}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " are built by blocks composed of linear and norm layers. The features from " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{c}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Psi_{d}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " are then fused in the bottleneck; since depth patches" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "were selected in regions where no RGB patches were selected, there are no duplicate patches representing the same patch location." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 546, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 263 + ], + "type": "text", + "content": "For those regions which do not have any associated RGB or depth patch, we use patches of constant values as mask tokens to create a placeholder in the bottleneck to enable reconstructing dense depth at the original image resolution. In the bottleneck, the RGB and depth patch feature vectors, along with the mask tokens, form the input to the decoder. This formulates a reconstruction task from sparse RGB and depth; the joint RGB-D pre-training enables reconstruction from very sparse input, as shown by our ablation on the masked input rations in Sec. 4.5. Note that the depth encoder is trained only during pre-training, and only the color ViT encoder (and decoder, if applicable) are used for downstream fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 264, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 547, + 312 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the pre-training task, we demonstrate the depth completion results from the pretraining phase in Fig. 6. A detailed analysis of masking different ratios of color and depth signals is shown in Sec. 4.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 314, + 547, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 386 + ], + "type": "text", + "content": "Pre-training Loss In contrast to the widely used contrastive loss in 3D representation learning, we train for dense depth reconstruction with an " + }, + { + "bbox": [ + 304, + 314, + 547, + 386 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 314, + 547, + 386 + ], + "type": "text", + "content": " reconstruction loss. Similar to MAE [19], we normalize the output patches as well as the target patches prior to computing the loss, which we found to empirically improve the performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 397, + 359, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 359, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 359, + 409 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 417, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 514 + ], + "type": "text", + "content": "We demonstrate the effectiveness of Mask3D pretraining for ViT [14] backbones on semantic segmentation, instance segmentation, and object detection tasks. We pre-train on ScanNet [10] data and demonstrate the effectiveness of learned masked 3D priors for not only ScanNet downstream tasks but also their transferability to NYUv2 [34] and even across the indoor/outdoor domain gap to Cityscapes [8] data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 521, + 422, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 422, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 422, + 534 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 576 + ], + "type": "text", + "content": "In this section, we introduce the pre-training and finetuning procedures. Our method uses a two-stage pretraining design introduced in the following." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 584, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 545, + 657 + ], + "type": "text", + "content": "Stage-I: Mask3D Encoder Initialization. We initialize the RGB encoder with network weights trained on ImageNet [12] (as pre-training for our pre-training). To maintain a fully self-supervised pre-training paradigm, we initialize with weights obtained by self-supervised ImageNet pre-training [19]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Stage-II: Mask3D Pre-training on ScanNet. Mask3D pre-training leverages 3D priors in RGB-D frame data, for which we use the color and depth maps of ScanNet [10]. Note that this does not use any semantic or reconstruction" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13512" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 545, + 261 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 261 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 261 + ], + "type": "image", + "image_path": "dc8cb09a518a2067cae2580897a845ac376e471c53f11e00667b29b7af8f27c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 546, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 546, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 546, + 313 + ], + "type": "text", + "content": "Figure 2. Overview of Mask3D Pre-training. As a pretext task, we predict dense depth from color and sparse depth signals. We use masked input by randomly selecting a set of patches from the input color image, which are then mapped to higher dimensional feature vectors; input depth is similarly masked and encoded. The color and depth features are then fused into a bottleneck from which dense depth is reconstructed as a self-supervised loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 323, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 373 + ], + "type": "text", + "content": "information during pre-training. ScanNet contains 2.5M RGB-D frames from 1513 ScanNet train video sequences. We regularly sample every " + }, + { + "bbox": [ + 46, + 323, + 288, + 373 + ], + "type": "inline_equation", + "content": "25^{\\text{th}}" + }, + { + "bbox": [ + 46, + 323, + 288, + 373 + ], + "type": "text", + "content": " frame without any other filtering (e.g., no control on viewpoint variation)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 381, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 525 + ], + "type": "text", + "content": "Downstream Fine-tuning. We evaluate our Mask3D pretraining by fine-tuning a variety of downstream image understanding tasks (semantic segmentation, instance segmentation, object detection). We consider in-domain transfers on ScanNet image understanding, and further evaluate the out-of-domain transfer on datasets with different statistical characteristics: the indoor image data of NYUv2 [34], as well as across a strong domain gap to the outdoor image data of Cityscapes [8]. For semantic segmentation tasks, we use both encoder and decoder pre-trained with Mask3D, and for instance segmentation and detection, only the backbone encoder is pre-trained." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 568 + ], + "type": "text", + "content": "Baselines. To evaluate the effectiveness of our learned masked 3D priors for 2D representations, we benchmark our method against relevant baselines:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 569, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 629 + ], + "type": "text", + "content": "Supervised ImageNet Pre-training (supIN). This uses the pre-trained weights from ImageNet, provided by torchvision, as is commonly used for image understanding tasks. Here, only ImageNet data is used, and no ScanNet data is involved in the pre-training phase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "text", + "content": "2-Stage MoCoV2 (MoCoV2-supIN " + }, + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "text", + "content": " SN). Supervised ImageNet pre-trained (supIN) weights are used as network initialization for pre-training. MoCoV2 [20] is used for pre-training with randomly shuffled ScanNet images. Here, both ImageNet and ScanNet image data are used without any geometric priors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "2-Stage MAE (MAE-unsupIN " + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": " SN). Self-supervised Im" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 324, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 384 + ], + "type": "text", + "content": "ageNet pre-trained weights are used as network initialization for pre-training. MAE [19] is used for pre-training with randomly shuffled ScanNet images. Here, both ImageNet and ScanNet image data are used without any geometric priors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "type": "text", + "content": "Pri3D [25]. Supervised ImageNet pre-trained are used to initialize Pri3D pre-training, which leverages multi-view and reconstruction constraints from ScanNet data under a contrastive loss. Here, both ImageNet and ScanNet data are used, incorporating 3D priors from reconstructed RGB-D video sequences for pre-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 460, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 651 + ], + "type": "text", + "content": "Implementation Details. We use a ViT-B backbone to train our approach. For pre-training, we use an SGD optimizer with a learning rate of 0.1 and an effective batch size of 128 (accumulated gradients from an actual batch size of 64). The learning rate is decreased by a factor of 0.99 every 1000 steps, and our method is trained for 100 epochs. Fine-tuning on semantic segmentation is trained with a batch size of 8 for 80 epochs. The initial learning rate is 0.01, with polynomial decay with a power of 0.9. Finetuning on detection and instance segmentation is trained using Detector2 [38] with the 1x schedule. Pre-training experiments are conducted on a single NVIDIA A6000 GPU, or 2 NVIDIA RTX3090 GPUs, or 4 NVIDIA RTX2080Ti GPUs; semantic segmentation experiments are conducted on a single NVIDIA A6000 GPU; instance segmentation and detection experiments are conducted on 8 V100 GPUs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 659, + 459, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 459, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 459, + 671 + ], + "type": "text", + "content": "4.2. ScanNet Downstream Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "We demonstrate the effectiveness of representation learning with 3D priors via downstream tasks on ScanNet [10] images. For fine-tuning, we follow the standard" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13513" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 101, + 70, + 491, + 194 + ], + "blocks": [ + { + "bbox": [ + 101, + 70, + 491, + 194 + ], + "lines": [ + { + "bbox": [ + 101, + 70, + 491, + 194 + ], + "spans": [ + { + "bbox": [ + 101, + 70, + 491, + 194 + ], + "type": "table", + "html": "
Pre-training MethodBackbonePre-training DatamIoU
ScratchResNet-50None39.1
ImageNet Pre-training (supIN)ResNet-50ImageNet55.7
Supervised Pre-trainingViTImageNet+ScanNet65.9 (+10.2)
MoCoV2-supIN→SN [20]ResNet-50ImageNet+ScanNet56.6 (+0.9)
Pri3D [25]ResNet-50ImageNet+ScanNet60.2 (+4.5)
Pri3D [25]ViTImageNet+ScanNet59.3 (+3.6)
DINO [3]ViTImageNet+ScanNet58.1 (+3.6)
MAE-unsupIN→SN [19]ViTImageNet+ScanNet63.3 (+7.6)
Ours - Mask3D (DINO)ViTImageNet+ScanNet60.5 (+4.8)
Ours - Mask3D (MAE)ViTImageNet+ScanNet66.7 (+11.0)
", + "image_path": "62e7cf64421ce4c3177570038e0d429b7c461672494753fe6e0b7150375b77b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 234, + 293, + 316 + ], + "blocks": [ + { + "bbox": [ + 46, + 200, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 547, + 224 + ], + "type": "text", + "content": "Table 1. ScanNet 2D Semantic Segmentation. Mask3D significantly outperforms Pri3D as well as other state-of-the-art pre-training approaches that leverage both ImageNet and ScanNet data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 234, + 293, + 316 + ], + "lines": [ + { + "bbox": [ + 48, + 234, + 293, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 293, + 316 + ], + "type": "table", + "html": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch32.717.716.9
ImageNet Pretrain (supIN)41.725.925.1
MoCoV2-supIN→SN [20]43.5 (+1.8)26.8 (+0.9)25.8 (+0.7)
Pri3D [25]43.7 (+2.0)27.0 (+1.1)26.3 (+1.2)
MAE-unsupIN→SN [19]46.1 (+4.4)32.7 (+6.8)30.5 (+5.4)
Mask3D (Ours)50.4 (+8.7)35.3 (+9.4)32.7 (+7.6)
", + "image_path": "7c0da3f62e32de6c9dd1d5054ca626bce7090efdb676a58f74bfa2f742a816dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 386, + 294, + 468 + ], + "blocks": [ + { + "bbox": [ + 46, + 320, + 288, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 288, + 376 + ], + "type": "text", + "content": "Table 2. ScanNet 2D Object Detection. Fine-tuning with Mask3D pre-trained models leads to improved object detection results across different metrics, in comparison to ImageNet pretraining, MoCo-style pre-training, and a strong MAE-style pre-training method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 386, + 294, + 468 + ], + "lines": [ + { + "bbox": [ + 48, + 386, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 294, + 468 + ], + "type": "table", + "html": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch25.813.112.2
ImageNet Pretrain (supIN)32.617.817.6
MoCoV2-supIN→SN [20]33.9 (+1.3)18.1 (+0.3)18.3 (+0.7)
Pri3D [25]34.3 (+1.7)18.7 (+0.9)18.3 (+0.7)
MAE-unsupIN→SN [19]37.4 (+4.8)20.3 (+2.5)20.7 (+3.1)
Mask3D (Ours)41.2 (+8.6)22.7 (+4.9)22.8 (+5.2)
", + "image_path": "16f72ee8d05e74abdc1c64fd06ea2c78c10565350809664a76b78cfa449968a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 538, + 294, + 619 + ], + "blocks": [ + { + "bbox": [ + 46, + 472, + 288, + 528 + ], + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 528 + ], + "type": "text", + "content": "Table 3. ScanNet 2D Instance Segmentation. Fine-tuning with Mask3D pre-trained models leads to improved instance segmentation results across different metrics compared to ImageNet pre-training, MoCo-style pre-training, and a strong MAE-style pre-training method." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 538, + 294, + 619 + ], + "lines": [ + { + "bbox": [ + 48, + 538, + 294, + 619 + ], + "spans": [ + { + "bbox": [ + 48, + 538, + 294, + 619 + ], + "type": "table", + "html": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch17.29.28.8
ImageNet Pretrain (supIN)25.113.913.4
MoCoV2-supIN→SN [20]27.2 (+2.1)14.7 (+0.2)14.8 (+1.4)
Pri3D [25]28.1 (+3.0)15.7 (+1.8)15.7 (+2.3)
MAE unsupIN→SN [19]33.6 (+8.5)19.0 (+5.1)19.0 (+5.6)
Mask3D (Ours)37.0 (+11.9)21.6 (+7.7)21.3 (+7.9)
", + "image_path": "ae074d9b53aa3fd19819ba96a4c9dd4f58580f8816a7f819c600008fea7b19bc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 624, + 287, + 669 + ], + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 669 + ], + "type": "text", + "content": "Table 4. NYUv2 2D Instance Segmentation. Fine-tuning with Mask3D pre-trained models leads to improved instance segmentation results across different metrics compared to previous methods, demonstrating the cross-dataset transfer ability of Mask3D." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 685, + 287, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 685, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 685, + 287, + 723 + ], + "type": "text", + "content": "protocol of the ScanNet benchmark [10] and sample every " + }, + { + "bbox": [ + 46, + 685, + 287, + 723 + ], + "type": "inline_equation", + "content": "100^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 685, + 287, + 723 + ], + "type": "text", + "content": " frame, resulting in 20,000 train images and 5,000 validation images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "text", + "content": "2D Semantic Segmentation. Tab. 1 shows the fine-tuning for semantic segmentation, in comparison with baseline pre-training approaches. All pre-training methods significantly improve performance over training the semantic segmentation model from scratch. In particular, Mask3D provides substantially better representation quality leading to a much stronger improvement over supervised ImageNet pretraining " + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "inline_equation", + "content": "(+11\\mathrm{mIoU})" + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "text", + "content": ", and notably improving over MAE unsupIN " + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "text", + "content": " SN with ImageNet and ScanNet " + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "inline_equation", + "content": "(+3.4\\mathrm{mIoU})" + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "text", + "content": " and the 3D-based pre-training of Pri3D " + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "inline_equation", + "content": "(+6.5\\mathrm{mIoU})" + }, + { + "bbox": [ + 304, + 236, + 547, + 429 + ], + "type": "text", + "content": ". We note that the multi-view 3D pre-training of Pri3D does not effectively embed informative 3D priors to ViT backbones, rather suffering from performance degradation from a ResNet-50 backbone. In contrast, our Mask3D pretraining can notably improve performance with a ViT backbone, indicating the effectiveness of our learned 3D priors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 432, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 547, + 564 + ], + "type": "text", + "content": "2D Object Detection and Instance Segmentation We show that Mask3D provides effective general 3D priors for a variety of image-based tasks, by evaluating downstream object detection and instance segmentation in Tab. 2 and Tab. 3, respectively. Across all tasks, various pre-training approaches yield substantial improvement over training from scratch. Our masked 3D prior learning transfers effectively learned representations for object detection and instance segmentation, notably improving over the best-performing MAE-unsupIN " + }, + { + "bbox": [ + 304, + 432, + 547, + 564 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 432, + 547, + 564 + ], + "type": "text", + "content": " SN (+4.3 AP@0.5 and +3.8 AP@0.5, respectively)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "text", + "content": "Data-Efficient Scenarios. We additionally show that our single-view RGB-D pre-training to embed 3D priors in limited data scenarios for downstream ScanNet semantic segmentation in Fig. 5. Mask3D shows consistent improvements across a range of limited data; even with only " + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "text", + "content": " of the training data, we recover " + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "text", + "content": " performance with " + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "text", + "content": " training data available and improving " + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "inline_equation", + "content": "+15.2" + }, + { + "bbox": [ + 304, + 568, + 547, + 665 + ], + "type": "text", + "content": " mIoU over Pri3D pre-training on a ViT backbone." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 671, + 455, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 455, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 455, + 682 + ], + "type": "text", + "content": "4.3. NYUv2 Downstream Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "We demonstrate the generalizability of our 3D-imbued learned feature representations across datasets, us" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13514" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 532, + 352 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 532, + 352 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 532, + 352 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 532, + 352 + ], + "type": "image", + "image_path": "3fe270dc8c44a7b29c5a15ad45e7c0f0efbb3a2e532d13043902dff2d5200c37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 355, + 548, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 355, + 548, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 548, + 390 + ], + "type": "text", + "content": "Figure 3. Qualitative Results on Various Tasks across Different Benchmarks. We visualize predictions on different tasks across various scene understanding benchmarks. From top to bottom rows: instance segmentation on ScanNet, instance segmentation on NYUv2, semantic segmentation on NYUv2, and semantic segmentation results in ScanNet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 396, + 545, + 630 + ], + "blocks": [ + { + "bbox": [ + 52, + 396, + 545, + 630 + ], + "lines": [ + { + "bbox": [ + 52, + 396, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 52, + 396, + 545, + 630 + ], + "type": "image", + "image_path": "ab080659c103eb936871270160b0774b951d7aeab8fa96ac1d99dc8dbecb50b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 633, + 547, + 656 + ], + "lines": [ + { + "bbox": [ + 46, + 633, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 633, + 547, + 656 + ], + "type": "text", + "content": "Figure 4. More Qualitative Results on Semantic Segmentation. We visualize semantic segmentation predictions on various scene understanding benchmarks including ScanNet and NYUv2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 666, + 289, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 289, + 703 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 289, + 703 + ], + "type": "text", + "content": "ing Mask3D pre-trained on ScanNet and fine-tuned on NYUv2 [34] following the same fine-tuning setup as before. NYUv2 contains Microsoft Kinect RGB-D video" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 666, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 547, + 704 + ], + "type": "text", + "content": "sequences of indoor scenes, comprising 1449 densely labeled RGB images. We use the official 795/654 train/val split. Tables 5, 6, and 4 evaluate the downstream tasks" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13515" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 70, + 470, + 173 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 470, + 173 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 470, + 173 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 470, + 173 + ], + "type": "table", + "html": "
Pre-training MethodBackbonePre-training DatamIoU
ScratchResNet-50None24.8
ImageNet Pre-training (supIN)ResNet-50ImageNet50.0
Supervised Pre-trainingViTImageNet+ScanNet55.5 (+5.5)
MoCoV2-supIN→SN [20]ResNet-50ImageNet+ScanNet47.6 (-2.4)
Pri3D [25]ResNet-50ImageNet+ScanNet54.2 (+4.2)
Pri3D [25]ViTImageNet+ScanNet53.2 (+3.2)
MAE-unsupIN→SN [19]ViTImageNet+ScanNet54.9 (+4.9)
Mask3D (Ours)ViTImageNet+ScanNet56.9 (+6.9)
", + "image_path": "ce03959531e20140c76015beae5396c6c611504af19958f91c251e8235a0c059.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 212, + 282, + 392 + ], + "blocks": [ + { + "bbox": [ + 53, + 212, + 282, + 392 + ], + "lines": [ + { + "bbox": [ + 53, + 212, + 282, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 212, + 282, + 392 + ], + "type": "image", + "image_path": "7b2a353062552324ebf8eb7a23902eb9505ec57870a964a21edd760623e9241a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 394, + 288, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 394, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 288, + 462 + ], + "type": "text", + "content": "Figure 5. Data-Efficient Results. Compared to previous methods, Mask3D demonstrates consistent improvements on ScanNet 2D semantic segmentation across a range of limited data scenarios. Mask3D is particularly effective for ViT pre-training, improving " + }, + { + "bbox": [ + 46, + 394, + 288, + 462 + ], + "type": "inline_equation", + "content": "+15.2\\%" + }, + { + "bbox": [ + 46, + 394, + 288, + 462 + ], + "type": "text", + "content": " mIoU over state-of-the-art Pri3D [25] on a ViT backbone at " + }, + { + "bbox": [ + 46, + 394, + 288, + 462 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 394, + 288, + 462 + ], + "type": "text", + "content": " of the training data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 472, + 297, + 553 + ], + "blocks": [ + { + "bbox": [ + 46, + 177, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 547, + 200 + ], + "type": "text", + "content": "Table 5. NYUv2 2D Semantic Segmentation. Mask3D significantly outperforms state-of-the-art pre-training approaches, demonstrating its effectiveness in transferring to different dataset characteristics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 472, + 297, + 553 + ], + "lines": [ + { + "bbox": [ + 48, + 472, + 297, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 297, + 553 + ], + "type": "table", + "html": "
Pre-training MethodAP@0.5AP@0.75AP
Scratch21.310.39.0
ImageNet Pretrain (supIN)29.917.316.8
MoCoV2-supIN→SN [20]30.1 (+0.20)18.1 (+0.80)17.3 (+0.50)
Pri3D [25]33.0 (+2.10)19.8 (+2.60)18.9 (+2.10)
MAE-unsupIN→SN [19]40.3 (+10.4)24.5 (+7.20)23.2 (+6.40)
Mask3D (Ours)44.0 (+14.1)28.3 (+6.40)25.9 (+9.10)
", + "image_path": "d6b653bfa166aabb43e0e7a690999f9ac5a46445a705f245687dd1af260eca6a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "of 2D semantic segmentation, object detection, and instance segmentation, respectively. Across all three tasks on NYUv2 data, our Mask3D pre-training achieves notably improved performance than training from scratch as well as the various baseline pre-training methods. In particular, we achieve an improvement of " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "+6.9" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " mIoU, " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "+14.1" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " AP@0.5, and " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "+11.9" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " AP@0.5 over the common supervised ImageNet pretraining on semantic segmentation, object detection, and instance segmentation." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 208, + 545, + 267 + ], + "blocks": [ + { + "bbox": [ + 46, + 555, + 287, + 599 + ], + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 599 + ], + "type": "text", + "content": "Table 6. NYUv2 2D Object Detection. Fine-tuning with Mask3D pre-trained models leads to improved object detection results across different metrics, showing an effective transfer across dataset characteristics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 208, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 267 + ], + "type": "table", + "html": "
Pre-training MethodBackbonemIoU
ImageNet Pre-training (supIN)ResNet-5054.1
Pri3D [25]ResNet-5055.1 (+1.00)
MAE-unsupIN→SN [19]ViT64.7 (+10.6)
Mask3D (Ours)ViT66.4 (+12.3)
", + "image_path": "dc793ad43c6d5e266e38f040d39df204001cffa534cd4776a4ae5bb8737ef93e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 269, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 305, + 269, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 547, + 314 + ], + "type": "text", + "content": "Table 7. Cityscapes 2D Semantic Segmentation. Mask3D significantly outperforms state-of-the-art Pri3D as well as a strong MAE-style pre-training. This demonstrates the effectiveness of the transferability of Mask3D, even under a significant domain gap." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 318, + 443, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 318, + 443, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 318, + 443, + 331 + ], + "type": "text", + "content": "4.4. Out-of-domain Transfer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 337, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 529 + ], + "type": "text", + "content": "While Mask3D concentrates on pre-training for improving indoor scene understanding, we further demonstrate the effectiveness of our Mask3D pre-training for the out-of-domain transfer on outdoor data, such as Cityscapes [8]. We use the official data split of 3000 images for training and 500 images for the test. To evaluate the transferability in such a large domain gap scenario, we fine-tune the pretrained models for the 2D semantic segmentation task in Tab. 7. Our approach maintains performance improvement over baseline pre-training methods such as Pri3D (+11.3 mIoU) and MAE-unsupIN→SN (+1.7 mIoU). This indicates an encouraging transferability of our learned representations and their applicability to a variety of scenarios. Please refer to the supplemental material for more out-of-domain transfer results on more generally distributed data, such as ADE20K [46]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 536, + 407, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 407, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 407, + 548 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 554, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 627 + ], + "type": "text", + "content": "Does the pre-training masking ratio matter? We show how different masking ratios influence downstream task results in Tab. 8 on ScanNet semantic segmentation. We found a performance gain when masking more RGB values (keeping " + }, + { + "bbox": [ + 304, + 554, + 545, + 627 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 554, + 545, + 627 + ], + "type": "text", + "content": "), which in combination with the heavy depth masking (keeping " + }, + { + "bbox": [ + 304, + 554, + 545, + 627 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 554, + 545, + 627 + ], + "type": "text", + "content": ") leads to the best performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "content": "What about other ViT variants? In our experiments, we use ViT-B as the meta-architecture. We show Mask3D also works in other ViT variants, such as ViT-L (see Tab. 12), which exhibits a similar trend of improvements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Does the normalization in the reconstruction loss help? We normalize the features when computing the reconstruc" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13516" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 70, + 242, + 193 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 242, + 193 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 242, + 193 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 242, + 193 + ], + "type": "table", + "html": "
RGB RatioDepth RatiomIoU
20.0%0.0%65.2
20.0%20.0%66.7
20.0%80.0%65.5
50.0%20.0%65.9
50.0%50.0%64.7
80.0%20.0%64.8
80.0%50.0%64.8
100.0%0.0%64.6
100.0%20.0%64.8
100.0%100.0%64.5
", + "image_path": "d9040d351335d5e679882a07035870c0c2b212ef683dc01166881cd5bed94cdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "text", + "content": "tion loss and observe an improvement of " + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "inline_equation", + "content": "+0.8\\%" + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "text", + "content": " mIoU in the semantic segmentation task on ScanNet." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 84, + 282, + 248, + 329 + ], + "blocks": [ + { + "bbox": [ + 47, + 197, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 47, + 197, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 287, + 242 + ], + "type": "text", + "content": "Table 8. Ablation Study of Masking Ratios. on ScanNet 2D semantic segmentation. We mask out different ratios of RGB and depth patches, where the ratio indicates the percentage of kept patches. Refer to supplemental material for a full list." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 84, + 282, + 248, + 329 + ], + "lines": [ + { + "bbox": [ + 84, + 282, + 248, + 329 + ], + "spans": [ + { + "bbox": [ + 84, + 282, + 248, + 329 + ], + "type": "table", + "html": "
MethodBackbonemIoU
Train from ScratchViT32.6
MAEViT37.1
Mask3DViT42.2
", + "image_path": "41d83f121278ce92cb4dbf81148bb934406239ae220c112c653718b9a2a0dec4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "type": "text", + "content": "Compared to a pure depth prediction baseline. In Tab. 8, we demonstrate a superior performance with a " + }, + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "type": "text", + "content": " kept patches of RGB and depth, compared to a pure depth prediction method (66.7 vs. 64.6). Note in the table, pretraining with " + }, + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "type": "text", + "content": " RGB ratio and " + }, + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 47, + 370, + 287, + 442 + ], + "type": "text", + "content": " depth ratio is equivalent to a pure depth prediction from a RGB image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 451, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 510 + ], + "type": "text", + "content": "Color + depth reconstruction? We found that having joint losses on color and depth during pre-training does not benefit performance (see the following Tab. 10). The RGB reconstruction loss potentially makes pre-training easier, as we already have additional depth priors as guidance." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 94, + 521, + 239, + 555 + ], + "blocks": [ + { + "bbox": [ + 47, + 331, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 47, + 331, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 287, + 365 + ], + "type": "text", + "content": "Table 9. Results on ScanNet Semantic Segmentation without ImageNet pre-training. Similar trend is seen as ImageNet pretraining. The Gap gets larger compared to ImageNet pre-training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 521, + 239, + 555 + ], + "lines": [ + { + "bbox": [ + 94, + 521, + 239, + 555 + ], + "spans": [ + { + "bbox": [ + 94, + 521, + 239, + 555 + ], + "type": "table", + "html": "
MethodReconstructionmIoU
Mask3DRGB+Depth65.6
Mask3DDepth66.7
", + "image_path": "a8a60743a9ffa7c7ee8100ee633578b93eef671de55232ea802aec0d0790cd50.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 682 + ], + "type": "text", + "content": "No Stage-I pre-training. We observe a performance drop without ImageNet pre-training model as initialization for our pre-training. Since ImageNet pre-training is readily available and ScanNet has a relatively small amount of indoor data, we make ImageNet pre-training initialization as default, similar to Pri3D. Meanwhile, we conduct experiments without ImageNet pre-training in Tab. 9, and observe similar trends as when using ImageNet pre-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "RGB + semantic segmentation as pre-training. Using RGB and semantic segmentation for pre-training rather than" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 332, + 70, + 518, + 117 + ], + "blocks": [ + { + "bbox": [ + 47, + 559, + 286, + 581 + ], + "lines": [ + { + "bbox": [ + 47, + 559, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 286, + 581 + ], + "type": "text", + "content": "Table 10. ScanNet Semantic Segmentation. RGB as an additional signal does not bring a significant improvement." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 70, + 518, + 117 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 518, + 117 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 518, + 117 + ], + "type": "table", + "html": "
DatasetsMask3D - SemanticsMask3D
ScanNet65.966.7
NYUv255.556.9
CityScapes63.066.4
", + "image_path": "ed3ac8b6e1010709ba434d44cac4590c75ad4ae1237e94337e08e0744f1c98dd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 148, + 386, + 208 + ], + "blocks": [ + { + "bbox": [ + 307, + 148, + 386, + 208 + ], + "lines": [ + { + "bbox": [ + 307, + 148, + 386, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 386, + 208 + ], + "type": "image", + "image_path": "ae5dee2d82c18abc68501d66a71e68178ba1d44ff3f760b832606fc81b5e0c10.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 386, + 148, + 465, + 208 + ], + "blocks": [ + { + "bbox": [ + 386, + 148, + 465, + 208 + ], + "lines": [ + { + "bbox": [ + 386, + 148, + 465, + 208 + ], + "spans": [ + { + "bbox": [ + 386, + 148, + 465, + 208 + ], + "type": "image", + "image_path": "eb7469c911f93ad67b9482aeaaf11ff8ebb6d53280d622c65f786515b1518ce1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 465, + 148, + 544, + 208 + ], + "blocks": [ + { + "bbox": [ + 465, + 148, + 544, + 208 + ], + "lines": [ + { + "bbox": [ + 465, + 148, + 544, + 208 + ], + "spans": [ + { + "bbox": [ + 465, + 148, + 544, + 208 + ], + "type": "image", + "image_path": "f3ec038a4326ee242fd96bb273d9b5b6a362828e63f27906cb952b7ff3142598.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 209, + 386, + 268 + ], + "blocks": [ + { + "bbox": [ + 307, + 209, + 386, + 268 + ], + "lines": [ + { + "bbox": [ + 307, + 209, + 386, + 268 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 386, + 268 + ], + "type": "image", + "image_path": "b2da07e039d5b83d4b0ecb61be3cec12a41e2cc3b5eb91514a06210e793cac89.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 274, + 545, + 307 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 545, + 307 + ], + "type": "text", + "content": "Figure 6. Pre-trained ViT learns 3D structural priors. Our proposed pre-training method learns spatial structures from heavily masked RGB images." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 386, + 209, + 465, + 268 + ], + "blocks": [ + { + "bbox": [ + 386, + 209, + 465, + 268 + ], + "lines": [ + { + "bbox": [ + 386, + 209, + 465, + 268 + ], + "spans": [ + { + "bbox": [ + 386, + 209, + 465, + 268 + ], + "type": "image", + "image_path": "93f69c2a947fb008725e3a0a2582e6f1b68bd82dd9a8268c3f89880376d7d6f6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 465, + 209, + 544, + 268 + ], + "blocks": [ + { + "bbox": [ + 465, + 209, + 544, + 268 + ], + "lines": [ + { + "bbox": [ + 465, + 209, + 544, + 268 + ], + "spans": [ + { + "bbox": [ + 465, + 209, + 544, + 268 + ], + "type": "image", + "image_path": "4364ded99b9cc21523aa9333edc5219121b5264aae70e3f4094f0f75d53400bf.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 324, + 314, + 526, + 406 + ], + "blocks": [ + { + "bbox": [ + 306, + 118, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 306, + 118, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 118, + 545, + 140 + ], + "type": "text", + "content": "Table 11. Semantic segmentation results (mIoU). \"Mask3D - Semantics\" denotes pre-training using RGB+Semantics." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 314, + 526, + 406 + ], + "lines": [ + { + "bbox": [ + 324, + 314, + 526, + 406 + ], + "spans": [ + { + "bbox": [ + 324, + 314, + 526, + 406 + ], + "type": "table", + "html": "
Pre-training MethodBackbonemIoU
Pri3D [25]ResNet-5060.2
Pri3D [25]ViT-B59.3 (-0.9)
MAE-unsupIN→SN [19]ViT-B63.3 (+3.1)
Mask3D (Ours)ViT-B66.7 (+6.5)
Pri3D [25]ViT-L64.3 (+4.1)
MAE-unsupIN→SN [19]ViT-L68.2 (+8.0)
Mask3D (Ours)ViT-L70.7 (+10.5)
", + "image_path": "5b4481c68c8016e448be80e8fa43f17edc68735078888f33a28a3d3ed2654092.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 411, + 545, + 443 + ], + "lines": [ + { + "bbox": [ + 306, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 545, + 443 + ], + "type": "text", + "content": "Table 12. ViT Variants on ScanNet 2D Semantic Segmentation. Mask3D yields consistent improvements for both ViT-B and ViT-L backbone architectures." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 451, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 451, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 451, + 545, + 523 + ], + "type": "text", + "content": "depth completion achieved competitive results on ScanNet semantic segmentation, although this requires the use of semantic labels for the pre-training dataset, and is likely to be less transferable across domains than using depth completion. As shown in the following Tab. 11, the gap becomes larger when transferring to both NYUv2 and Cityscapes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we present Mask3D, a new self-supervised approach to embed 3D priors into learned 2D representations for image scene understanding. We leverage existing large-scale RGB-D data to learn 3D priors without requiring any camera pose or multi-view correspondence information, instead learning geometric and structural cues through a pre-text reconstruction task from masked color and depth. We show that Mask3D is particularly effective in pre-training the modern, powerful ViT backbones, with notable improvements across a variety of image-based tasks and datasets. We believe this shows the strong potential in effectively learning 3D priors and provides new avenues for such 3D-grounded representation learning." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13517" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In ICCV, 2016. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 348-367. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 287, + 304 + ], + "type": "text", + "content": "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 305, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 287, + 350 + ], + "type": "text", + "content": "[5] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 393 + ], + "type": "text", + "content": "[6] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. arXiv preprint arXiv:2112.02990, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "text", + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 485 + ], + "type": "text", + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 529 + ], + "type": "text", + "content": "[9] Manuel Dahnert, Ji Hou, Matthias Nießner, and Angela Dai. Panoptic 3d scene reconstruction from a single rgb image. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "type": "text", + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "text", + "content": "[11] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-468, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 622, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 655 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.1,3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 545, + 162 + ], + "type": "text", + "content": "[15] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6433-6442, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "type": "text", + "content": "[16] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3D-MPA: Multi-Proposal Aggregation for 3D Semantic Instance Segmentation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "type": "text", + "content": "[17] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "type": "text", + "content": "[18] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-aware 3D instance segmentation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 278, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 278, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 278, + 545, + 322 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 324, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 324, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 308, + 324, + 545, + 357 + ], + "type": "text", + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 4, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 358, + 545, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 358, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 308, + 358, + 545, + 390 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 393, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 425 + ], + "type": "text", + "content": "[22] Ji Hou, Angela Dai, and Matthias Nießner. 3D-SIS: 3D Semantic Instance Segmentation of RGB-D Scans. In CVPR, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 428, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 545, + 449 + ], + "type": "text", + "content": "[23] Ji Hou, Angela Dai, and Matthias Nießner. **RevealNet: Seeing Behind Objects in RGB-D Scans.** In CVPR, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "type": "text", + "content": "[24] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 486, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 486, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 486, + 545, + 518 + ], + "type": "text", + "content": "[25] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In ICCV, 2021. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 520, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 545, + 553 + ], + "type": "text", + "content": "[26] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 555, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 587 + ], + "type": "text", + "content": "[27] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In ICCV, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "text", + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "text", + "content": "[29] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In CVPR, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[30] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3D object detection in point clouds. ICCV, 2019. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13518" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[31] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. CVPR, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[32] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "text", + "content": "[34] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGB-D images. ECCV, 2012. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[35] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite. In CVPR, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. KPConv: Flexible and deformable convolution for point clouds. In CVPR, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "type": "text", + "content": "[37] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "type": "text", + "content": "[38] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "type": "text", + "content": "[39] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas J Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3D point cloud understanding. ECCV, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "text", + "content": "[40] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3D instance segmentation on point clouds. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 479, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 511 + ], + "type": "text", + "content": "[41] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas Guibas. GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In CVPR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "type": "text", + "content": "[42] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. arXiv preprint arXiv:2303.08231, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "text", + "content": "[43] Yu Zhang, Junle Yu, Xiaolin Huang, Wenhui Zhou, and Ji Hou. Pcr-cg: Point cloud registration via deep explicit color and geometry. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 443-459. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[44] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[45] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In European Conference on Computer Vision, pages 311-329. Springer, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "type": "text", + "content": "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13519" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_content_list.json b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..858179ea2c5c8a69ab7d6b43a7c5f9bbeb552a26 --- /dev/null +++ b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_content_list.json @@ -0,0 +1,1700 @@ +[ + { + "type": "text", + "text": "MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image Pretraining", + "text_level": 1, + "bbox": [ + 186, + 161, + 783, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyi Dong $^{1*}$ , Jianmin Bao $^{2*}$ , Yinglin Zhang $^{3}$ , Ting Zhang $^{2}$ , Dongdong Chen $^{4,\\dagger}$ , Hao Yang $^{2}$ , Ming Zeng $^{3}$ , Weiming Zhang $^{1}$ , Lu Yuan $^{4}$ , Dong Chen $^{2}$ , Fang Wen $^{2}$ , Nenghai Yu $^{1}$ $^{1}$ University of Science and Technology of China $^{2}$ Microsoft Research Asia \n $^{3}$ Xiamen University $^{4}$ Microsoft Cloud + AI", + "bbox": [ + 115, + 226, + 855, + 293 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dlight@mail., zhangwm@, ynh@}.ustc.edu.cn cddlyf@gmail.com", + "bbox": [ + 225, + 296, + 746, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jianbao, ting.zhang, luyuan, doch, fangwen}@microsoft.com", + "bbox": [ + 225, + 312, + 739, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhengyinglin@stu，zengming@}xmu.edu.cn yanghao.alexis@foxmail.com", + "bbox": [ + 189, + 328, + 776, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 372, + 312, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents a simple yet effective framework MaskCLIP, which incorporates a newly proposed masked self-distillation into contrastive language-image pretraining. The core idea of masked self-distillation is to distill representation from a full image to the representation predicted from a masked image. Such incorporation enjoys two vital benefits. First, masked self-distillation targets local patch representation learning, which is complementary to vision-language contrastive focusing on text-related representation. Second, masked self-distillation is also consistent with vision-language contrastive from the perspective of training objective as both utilize the visual encoder for feature aligning, and thus is able to learn local semantics getting indirect supervision from the language. We provide specially designed experiments with a comprehensive analysis to validate the two benefits. Symmetrically, we also introduce the local semantic supervision into the text branch, which further improves the pretraining performance. With extensive experiments, we show that MaskCLIP, when applied to various challenging downstream tasks, achieves superior results in linear probing, finetuning, and zero-shot performance with the guidance of the language encoder. Code will be release at https://github.com/LightDXY/MaskCLIP.", + "bbox": [ + 75, + 403, + 472, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 750, + 208, + 764 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language (VL) contrastive learning [31, 51] has shown remarkable success in pretraining for various tasks. With large-scale image-text pairs available on the Internet, the model composed of a simple dual encoder design learns", + "bbox": [ + 75, + 774, + 470, + 831 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strong semantic prior by aligning between image and text. The resulting visual encoder not only exhibits excellent linear probing and finetuning performance, but also enables impressive zero-shot performance with the guidance of the language encoder, showing the generality of natural language and its ability to supervise a wide range of visual concepts.", + "bbox": [ + 499, + 374, + 894, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nonetheless, the associated language description, though providing richer information than mere class labels, still can hardly describe all the information in the corresponding image, as images are continuous signals with fine-grained details and complex semantics. As a result, the VL contrastive by aligning global representations may only focus on the text-described objects and ignore the rest which might be useful for downstream tasks.", + "bbox": [ + 497, + 461, + 894, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we are interested in how to fully leverage the image itself to facilitate the VL contrastive to further improve the transfer capability. (1) Firstly, the learned feature representation shall characterize local patches, serving as a complementary for global representation in VL contrastive. Inspired by the recent success of masked image modeling [4, 19, 26, 51, 60, 61] in learning patch representations, we also randomly mask the input image with a large portion to force the visual encoder to focus on the remaining visible patches. (2) Secondly, the learned representation for local patches shall possess semantic meanings, being consistent with the global representation receiving semantic text supervision. We bring mean teacher self-distillation [25, 57] to supervise the learned patch representations with the visual feature representations, enabling implicit supervision from natural language. The resulting objective is denoted as masked self-distillation where the student model and the teacher model come from the same neural networks and the knowledge is distilled from the full image (fed to the teacher model) to the masked image (fed to student model). To this end, we introduce MaskCLIP by incorporating masked self", + "bbox": [ + 497, + 574, + 895, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, † Corresponding Author", + "bbox": [ + 92, + 842, + 337, + 853 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Work done during an internship at Microsoft Research Asia", + "bbox": [ + 92, + 853, + 416, + 865 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10995", + "bbox": [ + 465, + 885, + 505, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "distillation into VL contrastive to advance the transferable visual encoder.", + "bbox": [ + 75, + 125, + 468, + 153 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are several recent attempts [49, 68] also exploring the capability of the visual encoder under natural language supervision. The common approach is to introduce contrastive learning or masked image modeling on the vision side together with contrastive language-image pretraining. However, the performance indeed improves based on CLIP but does not as well as our masked self-distillation. We argue that (1) the contrastive learning objective based on central crop augmentation actually learns global representations for salient objects while lack of attention on the surrounding backgrounds [11]; and (2) masked image modeling usually needs to remap the learned representation to pixels [26] or discrete tokens [4]. Such low-level prediction target is inefficient for semantic feature learning and thus also conflicts with high-level language supervision in VL contrastive. A brief illustration is presented in Figure 1. In the experiments, we conduct comprehensive ablations to analyze the difference and provide numerical and visual evidence for better understanding.", + "bbox": [ + 75, + 154, + 472, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Symmetrically, we argue that local semantic supervision on the text branch is also helpful for the text encoder and eventually beneficial for zero-shot performance. So we introduce the same mask-data-modeling format supervision into the text branch as well. Different from images where the pixel is low-level signal, the words crafted by human beings are already highly semantic, so we use the tokenized word piece as the prediction target directly, following the well-studied mask language modeling method BERT. Meanwhile, to reduce the output conflicts between contrastive learning and mask language modeling, we introduce a small decoder for the mask language modeling branch.", + "bbox": [ + 75, + 416, + 472, + 581 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train our MaskCLIP on a subset of a publicly available image-text pairs dataset, YFCC [58], and thoroughly evaluate the transfer ability of visual representations on several vision benchmarks: ImageNet-1K [17] for classification, ADE20K [69] for semantic segmentation, MS-COCO [40] for detection and segmentation, as well as a batch of other classification benchmarks. When it comes to ImageNet-1K [17] classification, MaskCLIP achieves $+6.9\\%$ , $+7.2\\%$ , $+1.3\\%$ higher than CLIP for zero-shot transfer, linear probing, and finetuning respectively. For vision downstream tasks, we reach $+2.7$ mIoU on ADE20K [69] and $+1.8$ APb, $+1.4$ APm on MS-COCO [40]. For vision-language tasks, MaskCLIP achieves $+6.1\\%$ average zero-shot accuracy on 20 datasets, and $+17.2\\%$ , $+12.8\\%$ rank@1 improvement on the Flickr30K [67] image-test retrieval. In the recent Image Classification in the Wild challenge academic track, our MaskCLIP gets the $1_{st}$ result with $48.9\\%$ TOP-1 average accuracy, surpassing the second team with $3.4\\%$ .", + "bbox": [ + 75, + 581, + 472, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the major contributions of this work are:", + "bbox": [ + 95, + 830, + 450, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We present a novel vision-language pretraining", + "bbox": [ + 90, + 852, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "framework MaskCLIP, by introducing masked self-distillation objective to facilitate VL contrastive for better transferable visual models.", + "bbox": [ + 529, + 125, + 894, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. We present extensive ablation studies on MaskCLIP variants and provide in-depth analysis numerically and visually to help understand how the proposed masked self-distillation assists VL contrastive.", + "3. We demonstrate our MaskCLIP on tens of benchmarks, showing the superiority under all three settings: zero-shot, linear probing, and finetuning." + ], + "bbox": [ + 510, + 174, + 894, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 290, + 640, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language pretraining. Recent years have seen rapid progress made in vision-language pretraining [15, 18, 33, 35-39, 44-46, 50, 55, 56, 72]. Several multiple cross-modality loss functions have been proposed for the training objective, such as image-text matching [15, 37, 44, 56, 64], masked language modeling [15, 37, 44, 55, 56], masked image modeling [15, 44, 55, 56], contrastive loss [35, 38, 39]. These objects are often mixed with each other to form a compound objective. While a variety of approaches have been proposed, few works investigate the performance on visual representation learning for image classification. Recently, CLIP [51] and ALIGN [31] show that the image-text contrastive learning objective achieves promising performance for visual representation learning. There are many following works proposed to further improve the pretraining performance, DeCLIP [70], SLIP [49], COTS [43], ViCHA [54], CYCLIP [24] use additional uni/multi-modality supervision to improve the model capability, and PyramidCLIP [23], KLITE [53], IDEA [30] seek to external knowledge from pre-trained models or datasets as the additional guidance. FILIP [66] and LOUPE [34] introduce fine-grained alignment to the model. Focusing on this research direction, we analyze the desired properties of supervision which could be complementary to CLIP, and propose the masked self-distillation objective incorporated with the image-text contrastive loss to further improve pretraining performance for various visual understanding tasks.", + "bbox": [ + 497, + 313, + 895, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning Self-supervised visual representation learning has attracted increasing attention over the past few years. The objective of the self-supervised learning is mainly divided into two categories: contrastive and generative [41]. The contrastive methods, such as MOCO [12, 27], SimCLR [9, 10], BYOL [25], SimSiam [13], and DINO [6] measure the similar and dissimilar samples by contrastive loss. Their success heavily depends on the strong data augmentation. The generative methods, such as BEiT [4], MAE [26], PeCo [19], BEVT [60], BootMAE [20] and MaskFeat [61] leverage masked image modeling to reconstruct the remaining masked part of its original input from the given visible parts. The generative methods show more", + "bbox": [ + 497, + 687, + 895, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10996", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/572575a2acca74311944aa1ecbe913af6b324e8b33d2284f1a8940f199fd4db6.jpg", + "image_caption": [ + "(a) CLIP" + ], + "image_footnote": [], + "bbox": [ + 77, + 122, + 200, + 309 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ccea09b0b77ceeb4b93c215b19706025b070148e41d5239e61a643d403af67e3.jpg", + "image_caption": [ + "(b) + Contrastive Learning" + ], + "image_footnote": [], + "bbox": [ + 206, + 125, + 396, + 309 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aff0d5e3397fa30243d23918cc5d12debb9822f4031e7da9ce75fa8982c3c5de.jpg", + "image_caption": [ + "(c) + Mask Image Modeling" + ], + "image_footnote": [], + "bbox": [ + 400, + 125, + 576, + 309 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8c1639d6fc1293107cec6749c9c6f287453f4383955d519a03bc0bc447e41862.jpg", + "image_caption": [ + "(d) MaskCLIP", + "Figure 1. Pipeline comparison between combination CLIP with different vision self-supervised learning methods. (a) Vanilla CLIP. (b) CLIP + contrastive learning. (c) CLIP + pixel prediction mask image modeling. (d) CLIP + mask self-distillation, i.e. MaskCLIP. The $E_{T}$ , $E_{I}$ is the text encoder and image encoder respectively, and all the $E_{I}$ , $E_{T}$ within each pipeline share the weight. $\\bar{E}_{I}$ is the mean-teacher model, whose weight is updated by the exponential moving average of $E_{I}$ and does not require gradient." + ], + "image_footnote": [], + "bbox": [ + 581, + 122, + 894, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "promising transfer performance than the contrastive methods, as generative objective learns patch representations while contrastive objective focuses on learning centric global representations [11].", + "bbox": [ + 75, + 396, + 470, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-knowledge distillation Self-knowledge distillation [32] aims to distill the knowledge in a model itself and uses it for training the model. Instead of distilling knowledge from a pretrained teacher model [29], self-knowledge distillation regards a temporal ensemble of the student model as the teacher. It means that a student model becomes a teacher model itself, which gradually utilizes its own knowledge for softening the hard targets to be more informative during training. Self-knowledge distillation has been explored in semi-supervised learning [57], contrastive learning [16, 35], self-supervised learning [3, 7]. In this paper, we use visual features supervised by natural language for guidance in masked self-distillation which naturally fit VL contrastive to learn more transferable visual representations.", + "bbox": [ + 75, + 451, + 470, + 644 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. MaskCLIP", + "text_level": 1, + "bbox": [ + 75, + 656, + 194, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce MaskCLIP, a novel framework that learns visual representations. The core part of MaskCLIP is its backbone image encoder, denoted by $E_{I}$ as shown in Figure 1. It obtains the transferable capability during pretraining that could benefit downstream vision tasks. Following recent self-supervised approaches [4, 14, 26, 49], we implement the backbone $E_{I}$ as a Vision Transformer (ViT) [22]. The prediction results from $E_{I}$ given an input image $I$ then should be a collection of visual feature tokens, represented as", + "bbox": [ + 75, + 679, + 468, + 804 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {I} (I) = \\left\\{f _ {\\text {c l s}}, f _ {1}, f _ {2}, \\dots , f _ {N} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 814, + 468, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $cls$ is short for class token. $1,\\ldots ,N$ are the indexes of the non-class tokens.", + "bbox": [ + 75, + 838, + 468, + 865 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The rest of this section starts with the utilization of language supervision. More shall be emphasized on the masked self-distillation, which we deem crucial for visual pretraining.", + "bbox": [ + 497, + 396, + 892, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Vision-language Contrastive", + "text_level": 1, + "bbox": [ + 499, + 472, + 752, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [31, 51], we introduce a Transformer-based text encoder $E_{T}$ to leverage language knowledge. It aims to align the global feature representations of an image and a text with respect to some forms of similarity. Precisely, consider a given image-text pair $\\{I, T\\}$ , besides extracting the visual feature representation $E_{I}(I)$ using the vision backbone as shown by Equation 1, we additionally use the text encoder $E_{T}$ to extract linguistic features from the text $T$ .", + "bbox": [ + 497, + 498, + 892, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mean feature of the two branches are regarded as the global representations and are fed into a projection head (implemented as a fully-connected layer) respectively to obtain the metric embeddings $e^T$ and $e^I$ . Image-text contrastive loss is employed to align them during pretraining. The loss can be formulated as $\\mathcal{L}_T + \\mathcal{L}_I$ , with", + "bbox": [ + 497, + 612, + 892, + 695 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp (e _ {i} ^ {I} e _ {i} ^ {T} / \\sigma)}{\\sum_ {j = 1} ^ {B} \\exp (e _ {i} ^ {I} e _ {j} ^ {T} / \\sigma)}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 714, + 830, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(e _ {i} ^ {T} e _ {i} ^ {I} / \\sigma\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(e _ {i} ^ {T} e _ {j} ^ {I} / \\sigma\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 559, + 755, + 890, + 793 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B$ stands for the number of image-text pairs within a training mini-batch, $i,j$ are indexes within the batch; $\\sigma$ stands for the temperature for the loss functions, which is learned together with all other parameters during training.", + "bbox": [ + 499, + 809, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10997", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Masked Self-distillation for Visual Encoder", + "text_level": 1, + "bbox": [ + 77, + 125, + 443, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowledge distillation is a learning paradigm where a student model is trained to match the output of a given teacher model, so that the student model can be improved by the teacher. Instead of bringing in an external teacher, self-distillation methods such as [7, 25, 57] proposes using a mean teacher model that is derived from the student itself. In specific, the teacher shares the same structure with the student, while the parameters of the teacher are exponential moving averages (EMA) of the parameters from the student. In the following, we would use the term \"EMA model\" to represent such mean teacher model constructed from the student.", + "bbox": [ + 75, + 146, + 472, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MaskCLIP leverages the mean teacher self-distillation to enhance its vision representations. Let $\\bar{E}_I$ be the EMA model of the backbone encoder $E_I$ . $\\theta_t$ and $\\bar{\\theta}_t$ are the parameters of $E_I$ and $\\bar{E}_I$ at training step $t$ . $\\bar{\\theta}_t$ is updated with", + "bbox": [ + 75, + 312, + 472, + 368 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\theta} _ {t} = \\alpha \\bar {\\theta} _ {t - 1} + (1 - \\alpha) \\theta_ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 375, + 468, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyper-parameter for smoothing updates. We propose to incorporate masked image modeling into self-distillation, resulting in masked self-distillation with asymmetric input for student model and teacher model.", + "bbox": [ + 75, + 397, + 472, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In specific, considering a given input image $I$ , we first feed it to the EMA model $\\bar{E}_I$ (teacher model) to obtain the distillation targets. These target features can be represented as", + "bbox": [ + 75, + 453, + 472, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {E} _ {I} (I) = \\{\\bar {f} _ {c l s}, \\bar {f} _ {1}, \\bar {f} _ {2}, \\dots , \\bar {f} _ {N} \\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 507, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the meantime, we randomly mask a large portion of the input image patches and then feed it into the original backbone $E_I$ (student model). Following [26], we only feed the visible (unmasked) patches, denoted by $I'$ , into the original backbone $E_I$ to speed up computation and save memory. Let $\\mathcal{M}$ be the indexes of all the masked tokens. These encoded features corresponding to visible tokens can then be denoted as $E_I(I') = \\{f_{cls}'\\} \\cup \\{f_{k \\notin \\mathcal{M}}'\\}$ . They are then joined with a shared and learnable feature vector, denoted as $m$ , that represents mask tokens, to form a complete set of features $\\{f_{cls}', f_1', f_2', \\ldots, f_N'\\}$ , with $f_{i \\in \\mathcal{M}}' = m$ . We attach positional embeddings onto all these tokens, and append a small Transformer $D$ as a decoder to predict features of the masked region from the visible tokens, which could be formulated as", + "bbox": [ + 75, + 527, + 472, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(D \\circ E _ {I}\\right) \\left(I ^ {\\prime}\\right) = D \\left(\\left\\{f _ {c l s} ^ {\\prime}, f _ {1} ^ {\\prime}, f _ {2} ^ {\\prime}, \\dots , f _ {N} ^ {\\prime} \\right\\}\\right) \\\\ = \\left\\{f _ {c l s} ^ {\\prime \\prime}, f _ {1} ^ {\\prime \\prime}, f _ {2} ^ {\\prime \\prime}, \\dots , f _ {N} ^ {\\prime \\prime} \\right\\}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 733, + 468, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [71], we use an online quantizer $h(\\cdot)$ to transform the output features into a soft codewords distribution, and minimize the cross-entropy between the target features and the predicted features. Formally,", + "bbox": [ + 75, + 773, + 468, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {D i s t}} = \\frac {1}{| \\mathcal {M} |} \\sum_ {k \\in \\mathcal {M}} - \\bar {h} \\left(\\bar {f} _ {k}\\right) ^ {T} \\log h \\left(f _ {k} ^ {\\prime \\prime}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 144, + 834, + 468, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here the parameter of the teacher quantizer $\\bar{h}()$ is also EMA updated by the online quantizer, similar to the teacher model.", + "bbox": [ + 497, + 125, + 892, + 153 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Local Semantic Learning for Text Encoder", + "text_level": 1, + "bbox": [ + 497, + 161, + 863, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides the local semantic supervision for the visual encoder, we argue it is also helpful for the text encoder. So we introduce the BERT pretraining into the text branch. For the text $T = \\{t_{\\text{sos}}, t_1, t_2, \\dots, t_M, t_{\\text{eos}}\\}$ , we denote the masked input as $T' = \\{t_{\\text{sos}}', t_1', t_2', \\dots, t_M', t_{\\text{eos}}'\\}$ , where $t_{i \\in \\mathcal{M}_T}' = m_t$ and $t_{i \\notin \\mathcal{M}_T}' = t_i$ , and $\\mathcal{M}_T$ be the indexes of all the masked text tokens. The output feature of the encoder is $E_T(T')$ .", + "bbox": [ + 497, + 182, + 892, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reduce the output conflict between the global image-text contrastive learning and the local mask language modeling, we further introduce a small text decoder, which shares the same architecture as the encoder but with only a few layers. So that the global prediction and local prediction are conducted at different layers. We denote the output feature as: $(D_T\\circ E_T)(T') = \\{t_{sos}''$ , $t_1''$ , $t_2''$ , ..., $t_M''$ , $t_{eos}''\\}$ and the loss could be formulated as:", + "bbox": [ + 497, + 293, + 892, + 403 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M L M} = \\frac {1}{| \\mathcal {M} _ {T} |} \\sum_ {k \\in \\mathcal {M} _ {T}} - t _ {k} ^ {T} \\log t _ {k} ^ {\\prime \\prime}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 409, + 890, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Overall Loss Functions", + "text_level": 1, + "bbox": [ + 497, + 451, + 714, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we pretrain MaskCLIP with all these losses combined:", + "bbox": [ + 497, + 473, + 892, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I} + \\mathcal {L} _ {T} + \\lambda \\mathcal {L} _ {\\text {D i s t}} + \\beta \\mathcal {L} _ {\\text {M L M}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 510, + 890, + 524 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\lambda, \\beta$ being the hyper-parameter weighting between VL contrastive loss and self-supervised learning loss. All the components of MaskCLIP are trained from scratch, including the visual backbone $E_I$ , the visual decoder $D$ , the text encoder $E_T$ , as well as the text decoder $D_T$ .", + "bbox": [ + 497, + 533, + 892, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 497, + 614, + 633, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 497, + 638, + 581, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model architecture. Our framework consists of the visual encoder $E_{I}$ , the text encoder $E_{T}$ , the visual decoder $D$ , and the text decoder $D_{T}$ . We adopt the widely used Transformer ViT-B/16 [22] for a fair comparison. It is composed of 12 layers, 768 width, and 12 head. The input image is $224 \\times 224$ resolution and is further split into $14 \\times 14$ patches with size $16 \\times 16$ . A learnablecls token is pretended to the 196 embeddings. For the text encoder, we adopt a 12-layer, 512-width, and 8-head Transformer following CLIP [51], and the text decoder has 4 layers. The number of text tokens is fixed to 77 with necessary truncations or paddings. For the image decoder, we directly use a one-layer Vision Transformer.", + "bbox": [ + 497, + 657, + 892, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining details. We train our proposed MaskCLIP from scratch for 25 epochs, the batch size is fixed to 4096 for all the experiments. The masks used in the mask self-distillation", + "bbox": [ + 497, + 824, + 892, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10998", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/14487a7fcb5cc6382e984dc0b94f915e3aa1a7bab539289409b1dceb7dacd090.jpg", + "image_caption": [ + "Figure 2. Visualization of the similarity between text and image features. The images and captions are from the MS-COCO val set. Here we show the image feature similarity with both full caption and different objects in it. The caption is \"Three teddy bears sit in a sled in snow\". More results could be found in the supplemental materials." + ], + "image_footnote": [], + "bbox": [ + 82, + 119, + 873, + 429 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cb825e38a92ba65ededb5551252a33c6fd760bd5564fab181f6830a28b7a4b13.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainingIN-1KFlicker30K
MemoryTime0-shotLinearFinetuneI2TT2I
CLIP14G1.00×37.666.582.352.932.8
CLIP+SimCLR30G2.67×42.872.182.658.641.3
CLIP+MAE16G1.30×42.168.583.257.341.1
MaskCLIP19G1.75×44.573.783.670.145.6
", + "bbox": [ + 78, + 480, + 480, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Results of boosting CLIP with different kinds of vision self-supervised learning methods.", + "bbox": [ + 75, + 573, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "branch and mask language modeling branch are random mask with a mask ratio of $75\\%$ and $20\\%$ . We pretrain all the models with the commonly used YFCC15M dataset, which is flited from the YFCC100M [58] dataset by [51].", + "bbox": [ + 75, + 609, + 468, + 663 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream details. We evaluated MaskCLIP on several downstream datasets, including ImageNet-1K [17], ADE20K [69], MS-COCO [40], Flicr30K [67] et al. For ImageNet-1K, we report zero-shot, linear probing, and finetuning performance. The zero-shot is conducted following the label prompt setting in SLIP [49]. For linear probing, we fix the backbone and train a new linear classifier for 90 epochs. For finetuning, we follow the setting in BEiT [4] and finetune the model for 100 epochs with a layer-decayed learning rate. See supplemental materials for more details.", + "bbox": [ + 75, + 665, + 468, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Analysis", + "text_level": 1, + "bbox": [ + 75, + 815, + 179, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first present our analysis by studying different ways of boosting CLIP. The baseline is CLIP [51] trained on the", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "YFCC-15M. Besides the introduced masked self-distillation, we consider two other popular methods: (1) SimCLR [9], a representative contrastive method; and (2) MAE [26] the state-of-the-art masked image modeling approaches. All the compared methods are trained on the YFCC-15M for a fair comparison. We have the following observations.", + "bbox": [ + 497, + 483, + 892, + 567 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision self-supervision helps VL contrastive. We evaluate the models on both vision task ImageNet-1K [17] classification and vision-language task image-text retrieval on Flicker30K [67] and present the comparison in Table 1. All the added vision self-supervision, regardless of contrastive or generative, improves the baseline CLIP. Among them, our proposed MaskCLIP achieves the best results in terms of all the evaluation metrics, outperforming CLIP with $+6.9\\%$ , $+7.2\\%$ , $+1.3\\%$ on ImageNet-1K classification for zero-shot, linear probing, and finetuning respectively, and $+17.2\\%$ , $+12.8\\%$ on Flicker30K for image-to-text retrieval and text-to-image retrieval. We also report the training GPU memory usage and time-consuming cost in Table 1. It is worth noting that the contrastive model (CLIP + SimCLR) compares two additional views of the input image, resulting in larger GPU memory usage and longer training time.", + "bbox": [ + 497, + 571, + 892, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Masked image modeling is able to learn representations for local patches. We argue that the image encoder only pays attention to the text-described objects under VL contrastive due to sparse text description and to the centric objects under image contrastive due to central-crop augmenta", + "bbox": [ + 497, + 796, + 892, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10999", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5a13b4fd09b7a76e4cc06d7a4f4635c2f678e0d747b1b681c22a0f49fdbe2a87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodObjectiveADE20K mIoUPascal mIoU
CLIPGlobal7.213.5
CLIP + SimCLRGlobal + Global6.311.9
CLIP + MAEGlobal + Pixel-wise Local8.316.4
MaskCLIPGlobal + Token-wise Local10.217.2
", + "bbox": [ + 84, + 124, + 470, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion. In contrast, masked image modeling forces the image encoder to focus on local patches using token-wise objectives by mandatorily masking a large portion of patches. Here, we provide numerical comparisons for evidence. We conduct an \"Annotation-free zero-shot segmentation\" experiment to test the zero-shot segmentation. The results on such a dense prediction task would better reveal the ability of local patch representations than global classification. Following the design in DenseCLIP [70], we use the prompted label feature as the linear classification weight to realize segmentation, without any training procedure. We evaluate the performance on two widely used datasets: ADE20K [69] and Pascal Context [48]. The results are shown in Table.2. We can see that equipped with masked image modeling, our MaskCLIP as well as CLIP+MAE achieves better results than CLIP and CLIP+SimCLR, validating our hypothesis.", + "bbox": [ + 75, + 248, + 470, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Masked self-distillation learns semantic representations for local patches. Our masked self-distillation predicts visual features dynamically outputted by the visual encoder and thus implicitly gets supervision from the text side via VL contrastive. While MAE predicts fixed low-level pixels, making it inefficient to learn semantic representations (as the objective may force the representation to memorize low-level details) and thus causing conflict with VL contrastive. To show this, we select images from MS-COCO [40] and calculate the feature similarity between image features and their corresponding caption features. We also select objects in the caption, prompt it to a new caption, such as \"a photo of teddy bears\", and calculate the similarities. An example is shown in Figure 2 (More can be found in the supplementary material). Comparing MaskCLIP with CLIP+MAE in the fourth column, we can see that CLIP+MAE uses color as evidence and fails to distinguish the white teddy bear from the white snow. While our MaskCLIP successfully differentiates the two objects, suggesting ours learn more semantic features. On the other hand, the superior results of MaskCLIP shown in Table 1 and Table 2 also validate this. It is worth mentioning that CLIP and CLIP+SimCLR fail to have a correct response partition for different single objects like MaskCLIP, further justifying our second observation.", + "bbox": [ + 75, + 469, + 470, + 802 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with Previous Methods", + "text_level": 1, + "bbox": [ + 75, + 806, + 389, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To show the effectiveness of MaskCLIP as a general vision-language pretrain method, we conduct experiments on both vision tasks and vision-language tasks. For vision", + "bbox": [ + 75, + 824, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/00194b401d3638d249f4357215982875f5189df5992229fb375e856017248cac.jpg", + "table_caption": [ + "Table 2. Annotation-free zero-shot segmentation results on ADE20K and Pascal Context." + ], + "table_footnote": [], + "table_body": "
MethodEpochIN-1KADE20KMS-COCO
0-ShotLinFTmIoUAPbAPm
DeiT [59]300*--81.847.444.139.8
SimCLR [9]25-64.082.548.044.640.2
MAE [26]25-56.282.546.543.239.1
CLIP [51]2537.666.582.347.843.639.5
SLIP [49]2542.872.182.648.544.040.3
MaskCLIP2544.573.783.650.545.440.9
", + "bbox": [ + 507, + 124, + 894, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with previous methods, including supervised baselines, self-supervised pretraining methods, and vision-language pretraining methods. * is the epoch of the supervised baseline on ImageNet-1K.", + "bbox": [ + 497, + 245, + 892, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tasks, we report results on ImageNet-1K [17] classification, MS-COCO [40] object detection, and ADE20K [69] semantic segmentation. For vision-language tasks, we report zero-shot results on recent challenging ICinW 20 datasets benchmark and image-text retrieval results on Flickr30K [67] and MS-COCO [40]. In the following, we compare with the supervised baseline DeiT [59], self-supervised methods SimCLR [9] and MAE [26], and vision-language methods CLIP [51] and SLIP [49]. For a fair comparison, we train SimCLR and MAE on YFCC-15M [58] with the same epochs.", + "bbox": [ + 497, + 309, + 894, + 461 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classification on ImageNet-1K. As shown in Table 3, MaskCLIP benefits from the advantages of both VL pretraining and image mask self-distillation that shows strong performance on all the metrics. For zero-shot tasks, MaskCLIP outperforms CLIP by $+6.9\\%$ with 25 epoch training and achieves $+1.7\\%$ higher than the recent work SLIP. When it comes to finetune, MaskCLIP reaches $83.6\\%$ top-1 accuracy, and outperforms CLIP by $+1.3\\%$ .", + "bbox": [ + 497, + 461, + 894, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic segmentation on ADE20K. Then we apply our MaskCLIP to the semantic segmentation task. Here we use the UperNet [63] framework with $512 \\times 512$ input and end-to-end training for 160K iterations. The evaluation metric is the mean Intersection of Union (mIoU) and we report single-scale evaluation results here. The results are given in Table 3. Our method achieves $50.5\\mathrm{mIoU}$ , $+2.7\\mathrm{mIoU}$ than our baseline method CLIP, and $+2.0\\mathrm{mIoU}$ than SLIP. This verifies the effectiveness of our introduced incorporation.", + "bbox": [ + 497, + 574, + 892, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object detection and instance segmentation on MS-COCO. We further investigate our transfer performance on object detection and instance segmentation in Table.3. Here we use Mask-RCNN [28] framework with single-scale input and $1 \\times$ schedule (12 epochs). Our method achieves 45.4 box AP and 40.9 mask AP, $+1.8 / 1.4$ better than CLIP, and $+1.4 / 0.6$ better than SLIP.", + "bbox": [ + 497, + 699, + 894, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot on small datasets. We also report zero-shot performance on 20 small datasets under the ICinW setting (see the introduction below) in Table 4. We find that all the methods perform poorly on some datasets such as Aircraft(1% acc for random guessing, we omit the description in the", + "bbox": [ + 497, + 796, + 894, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11000", + "bbox": [ + 465, + 885, + 505, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f16ac5741607f4ef3518057800ca0888c8d8c789218e3708f1bd59641c53d93.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AverageCaltech-101CIFAR-10CIFAR-100Country211DTDEuroSATFER-2013AircraftFood-101GTSRBMemesKittiDisMNISTFlowersPetsPatchCamSST2RESISC45CarsVoc2007
Pretraining on YFCC-15M
CLIP34.058.668.536.910.821.430.516.95.151.66.551.125.95.052.728.651.752.522.44.579.1
SLIP37.870.982.648.611.826.619.818.15.659.912.651.829.49.856.331.455.351.528.55.480.5
MaskCLIP40.172.080.257.512.627.944.020.36.164.98.552.034.34.957.034.350.149.935.76.782.1
Pretraining on ICinW Academic Track Siting: YFCC-15M, GCC3M+12M, ImageNet-21K(ImageNet-1K is removed)
1st MaskCLIP48.986.495.378.311.633.057.718.88.078.917.352.816.07.374.274.452.146.254.326.582.3
2nd KLITE*45.587.492.768.88.232.227.917.44.372.411.448.431.112.875.665.950.652.944.410.282.3
3rd YT-CLIP44.577.883.558.411.931.940.727.16.968.718.852.39.118.853.169.351.550.352.719.779.3
4th UniCL†44.084.890.267.86.725.435.330.83.568.311.151.017.911.371.744.952.149.541.424.281.3
5th Gramer*43.283.992.969.57.325.524.430.42.771.09.052.612.410.170.452.450.650.144.813.881.3
", + "bbox": [ + 80, + 117, + 890, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ae827564b38c8c1c0547513255bd72f70ecfe97307c5dbb8a7cdb347e00b939.jpg", + "table_caption": [ + "Table 4. Zero-shot evaluation on ICinW classification benchmarks. Best results in bold. * indicates using Swin-B as the backbone, † indicates using Focal-B as the backbone." + ], + "table_footnote": [], + "table_body": "
Training EpochFlickr30KMS-COCO
Image-to-textText-to-imageImage-to-textText-to-image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP [51]2552.979.687.232.860.871.227.553.565.017.738.850.5
SLIP [49]2558.685.191.741.368.778.633.459.870.621.544.456.3
MaskCLIP2570.190.395.345.673.482.141.467.977.525.549.761.3
", + "bbox": [ + 85, + 339, + 890, + 427 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results of zero-shot image-text retrieval on Flickr30K and MS-COCO datasets. Best results in bold.", + "bbox": [ + 159, + 429, + 806, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "following), $\\mathrm{Fer}(24.7\\%)$ , Country211 $(0.5\\%)$ , GTSRB $(5.9\\%)$ , Cars $(0.8\\%)$ . This might be caused by the data domain gap that the YFCC-15M contains few related images and descriptions. For the rest of the datasets, all the methods get reasonable performance and our MaskCLIP gets the best performance on most datasets.", + "bbox": [ + 75, + 456, + 470, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Classification in the Wild (ICinW) Challenge The ICinW challenge [1] is a newly proposed visual pretraining benchmark, which contains 20 diverse downstream classification datasets, measuring the ability of pre-training models on both the prediction accuracy and their transfer efficiency in a new task. The pretraining is limited to three datasets: YFCC-15M [58], GCC3M [52] $+12\\mathrm{M}$ [8] and ImageNet-21K [17] (ImageNet-1K data is excluded). We pretrain our MaskCLIP on it and get the $\\mathbf{1}_{st}$ result in the zero-shot track [2] (we submit the results anonymously). As shown in Table 4, the $2_{nd}$ team KLITE uses a strong Swin-B [42] as the backbone and additional knowledge from GPT-3 [5] and Wiktionary [47], and the $4_{th}$ use the strong Focal-B [65] as the backbone, while our MaskCLIP greatly outperforms these methods with a simple ViT-B backbone and no additional knowledge.", + "bbox": [ + 75, + 543, + 473, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot on text-image retrieval. We further report the zero-shot text-image retrieval results on two benchmark datasets, Flickr30K [67] and MS-COCO [40]. We find that the text without any prefixes or suffixes works well for all the models. Table 5 shows the results. We can see that MaskCLIP exhibits a strong zero-shot performance. For example, with 25 epochs training, MaskCLIP reaches $41.4\\%$ Rank@1 image-to-text accuracy on MS-COCO, outperform-", + "bbox": [ + 75, + 755, + 470, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing CLIP with $13.9\\%$ and $25.5\\%$ Rank@1 text-to-image accuracy, $+7.8\\%$ higher than CLIP.", + "bbox": [ + 497, + 456, + 890, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 499, + 489, + 610, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our default settings with other alternatives to justify the efficacy of our model designs.", + "bbox": [ + 497, + 507, + 890, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training objectives ablation. As shown in Table.6a, when we remove the mask language modeling loss $\\mathcal{L}_{\\mathrm{MLM}}$ , the performance of the image-text task drops, including the zero-shot accuracy and retrial performance. While benefiting from the distillation loss, the finetuning performance on ImageNet-1K is not influenced. When we remove the distillation loss $\\mathcal{L}_{\\mathrm{Dis}}$ , we observe a performance drop on all tasks, especially the finetuning results.", + "bbox": [ + 497, + 534, + 892, + 644 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distillation loss format. Different from previous methods [3,21,26] that calculate the per-element distance as the loss function, we use an online tokenizer to map the feature to soft codewords and use the cross-entropy loss as the supervision. Here we study their difference in Table.6b. We find that although they get similar fine-tuning performance, the CE loss gets better zero-shot and linear probing performance. The reason may be that the per-element MSE loss leads the model to fit some unnecessary details of the target feature, while the CE loss with soft tokenizer helps the model to focus more on the important feature.", + "bbox": [ + 497, + 644, + 894, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distillation & MLM loss weight. Here we set the loss weight of the CLIP branch as 1 and study the loss weight of the two additional branches. As shown in Table.6e and Table.6f, setting $\\lambda = 1$ or $\\beta = 1$ emphasize too much on new tasks, which mislead the model to a wrong converge", + "bbox": [ + 497, + 796, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11001", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6f9990abb2bf4a59ed8e24b84f02e432963119a929b5171ef95416d1f84fcc12.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model0-ShotFTI2T/T2I
MaskCLIP44.583.670.1/45.6
w/o LMLM42.883.665.0/41.6
w/o LDis42.082.465.4/40.5
", + "bbox": [ + 87, + 122, + 339, + 182 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e67951f014a0995a170122980409bbb78cd1ba8e603324e7595e43e9664ae6cd.jpg", + "table_caption": [ + "(a) Training Objectives ablation. Both is necessary for MaskCLIP." + ], + "table_footnote": [], + "table_body": "
Depth0-ShotI2T/T2I
043.565.2/44.1
144.370.4/45.3
244.370.2/45.4
444.570.1/45.6
844.267.5/44.7
", + "bbox": [ + 104, + 220, + 310, + 305 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/598e85f86ffa043ccde524e12058f06e3a1d368c20a33e0034327d3e57e24a88.jpg", + "table_caption": [ + "(d) Text decoder depth. The decoder is necessary and a shallow one works better." + ], + "table_footnote": [], + "table_body": "
Loss0-ShotLinFT
MSE43.873.283.6
CE44.573.783.6
", + "bbox": [ + 384, + 122, + 588, + 182 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/108f3f8d0001d1385786867ad3e246c537136b405ad77c65d3691e17c53b4b6b.jpg", + "table_caption": [ + "(b) Distillation loss format. The online tokenizer with cross-entropy loss works slightly better than MSE loss." + ], + "table_footnote": [], + "table_body": "
Weight0-ShotLinFT
138.568.282.5
0.144.473.583.5
0.0544.573.783.6
0.0143.673.083.4
", + "bbox": [ + 383, + 220, + 589, + 305 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d61ad7835f76f4711d36b8203f24957d5377a022edf767e19cbf8e147b1c5f93.jpg", + "table_caption": [ + "(e) Distillation loss weight. A small loss weight works well for MaskCLIP." + ], + "table_footnote": [], + "table_body": "
Depth0-ShotLinFT
144.573.783.6
243.772.983.4
443.572.583.3
", + "bbox": [ + 675, + 122, + 855, + 182 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/098ca3f538339954e7552328387701cdac7e79c57b43e8021e9c014a3608c1b9.jpg", + "table_caption": [ + "(c) Visual decoder Depth. A shallow decoder gets better performance." + ], + "table_footnote": [], + "table_body": "
Weight0-ShotI2T/T2I
136.551.7/32.1
0.144.369.2/45.9
0.0544.570.1/45.6
0.0143.270.6/45.6
", + "bbox": [ + 673, + 220, + 857, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. MaskCLIP ablation experiments with YFCC-15M dataset. We report zero-shot(0-Shot), fine-tuning (FT), and linear probing (Lin) accuracy (%) for image-encoder-related ablation. And zero shot image-to-text, text-to-image retrieval (I2T/T2I) for text encoder-related ablations. Default settings are marked in gray .", + "bbox": [ + 75, + 331, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "direction, resulting in poor performance. When we reduce the loss weight by $10 \\times$ , the two additional tasks are helpful for the model and show a consistent gain on all the metrics. We suspect this is because the CLIP loss requires two different capabilities: understanding the input content and aligning them into a shared feature space. And the goal of the two additional self-supervised learning tasks is to facilitate understanding.", + "bbox": [ + 75, + 376, + 468, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image & Text decoder depth. Then we study the influence of the decoder depth for both image and text decoders. As shown in Table.6c, we find the image decoder with only one layer works well, increasing the decoder depth leads to worse performance on all metrics. Similarly, Table.6d shows that the text branch benefits from a shallow decoder design. We argue that a too-deep decoder would make the encoder lazy, relying on the strong decoder to resolve the challenging mask feature/language modeling tasks. And the different depth choice between the image and text branches is caused by the framework difference: the image branch sees the mask tokens at the decoder, while the text branch takes the mask tokens as the encoder input. Note that if we remove the text decoder, the performance gets worse. We think this is largely caused by the output conflict that the global recognition feature aggregation and local word prediction are conducted at the same layer.", + "bbox": [ + 75, + 490, + 468, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single-Stage v.s. two-Stage. Our MaskCLIP learns the VL contrastive and masked self-distillation simultaneously and jointly in a single stage. One possible variant is to first train CLIP and then use CLIP feature from the first stage to train masked image modeling as in [61, 62]. We report results on three datasets in Table 7. We can see that the second stage achieves better finetuning results compared with results from the stage one, showing the effectiveness of masked image modeling. Nonetheless, such two-stage training requires longer training time and loses the transfer capability in a zero", + "bbox": [ + 75, + 726, + 468, + 865 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fc667e51881d97bc34b72dc1747a68706fc2333a0537b5221c6a87a0d2899df2.jpg", + "table_caption": [ + "(f) MLM loss weight. A small loss weight works better." + ], + "table_footnote": [], + "table_body": "
MethodEpochIN-1KFlicker30KADE20K
0-shotFTI2TT2I0-shotFT
Two-StageStage12537.682.352.932.87.247.8
Stage225—83.4———48.2
MaskCLIP2544.583.670.145.610.250.5
", + "bbox": [ + 507, + 374, + 892, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Comparison between two-stage method and our single-stage MaskCLIP.", + "bbox": [ + 499, + 467, + 892, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shot setting. In contrast, our MaskCLIP achieves superior results under all settings with fewer epochs.", + "bbox": [ + 499, + 502, + 890, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 535, + 616, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present MaskCLIP, a new VL pretraining framework that incorporates masked self-distillation into VL contrastive. We point out that masked self-distillation learns local semantics, fitting nicely to the VL contrastive that aims to learn global semantics, and this is supported with comprehensively designed experiments. We also utilize mask language modeling to enhance the text encoder which is critical for zero-shot performance. The resulting visual encoder shows strong transfer capability across widely adopted benchmarks for linear probing, fine-tuning, and also zero-shot evaluation.", + "bbox": [ + 497, + 555, + 892, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 499, + 704, + 665, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Natural Science Foundation of China under Grant U20B2047, 62072421, 62002334, 62102386, and 62121002. Key Research and Development program of Anhui Province under Grant 2022k07020008, Ant Group through CCF-Ant Innovative Research Program CCF-AFSG RF20210025, Alibaba Group through Alibaba Innovative Research Program. This work was also partly supported by Shenzhen Key Laboratory of Media Security, Shenzhen University, Shenzhen 518060, China.", + "bbox": [ + 497, + 726, + 892, + 864 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11002", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 173, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https : / / eval . ai / web / challenges / challenge-page/1832/overview.", + "[2] https : / / eval . ai / web / challenges / challenge-page/1832/leaderboard/4298.", + "[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022.", + "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020.", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021.", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021.", + "[8] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021.", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[10] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020.", + "[11] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.", + "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020.", + "[13] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning, 2020.", + "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021.", + "[15] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter:" + ], + "bbox": [ + 78, + 148, + 470, + 865 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020.", + "[16] Ruizhe Cheng, Bichen Wu, Peizhao Zhang, Peter Vajda, and Joseph E Gonzalez. Data-efficient language-supervised zero-shot learning with self-distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3119-3124, 2021.", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[18] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11162-11173, 2021.", + "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021.", + "[20] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022.", + "[21] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022.", + "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[23] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022.", + "[24] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022.", + "[25] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020.", + "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021.", + "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019.", + "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + ], + "bbox": [ + 502, + 127, + 892, + 865 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11003", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015.", + "[30] Xinyu Huang, Youcai Zhang, Ying Cheng, Weiwei Tian, Ruiwei Zhao, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Xiaobo Zhang. Idea: Increasing text diversity via online multi-label recognition for vision-language pre-training. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4573-4583.", + "[31] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021.", + "[32] Kyungyul Kim, ByeongMoon Ji, Doyoung Yoon, and Sangheum Hwang. Self-knowledge distillation: A simple way for better generalization. arXiv preprint arXiv:2006.12000, 3, 2020.", + "[33] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11336-11344, 2020.", + "[34] Juncheng Li, Xin He, Longhui Wei, Long Qian, Linchao Zhu, Lingxi Xie, Yueting Zhuang, Qi Tian, and Siliang Tang. Fine-grained semantically aligned vision-language pre-training. arXiv preprint arXiv:2208.02515, 2022.", + "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34, 2021.", + "[36] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020.", + "[37] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019.", + "[38] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409, 2020.", + "[39] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020.", + "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.", + "[41] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Gener-" + ], + "bbox": [ + 78, + 125, + 470, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 2021.", + "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021.", + "[43] Haoyu Lu, Nanyi Fei, Yuqi Huo, Yizhao Gao, Zhiwu Lu, and Ji-Rong Wen. Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15692-15701, 2022.", + "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019.", + "[45] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10437-10446, 2020.", + "[46] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020.", + "[47] Christian M Meyer and Iryna Gurevych. Wiktionary: A new rival for expert-built lexicons? Exploring the possibilities of collaborative lexicography. na, 2012.", + "[48] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014.", + "[49] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. arXiv preprint arXiv:2112.12750, 2021.", + "[50] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020.", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.", + "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018.", + "[53] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Anna Rohrbach, Zhe Gan, Lijuan Wang, Lu Yuan, et al. K-lite: Learning transfer" + ], + "bbox": [ + 502, + 127, + 892, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11004", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "able visual models with external knowledge. arXiv preprint arXiv:2204.09222, 2022.", + "[54] Mustafa Shukor, Guillaume Couairon, and Matthieu Cord. Efficient vision-language pretraining with visual concepts and hierarchical alignment. arXiv preprint arXiv:2208.13628, 2022.", + "[55] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019.", + "[56] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019.", + "[57] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017.", + "[58] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016.", + "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021.", + "[60] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022.", + "[61] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021.", + "[62] Longhui Wei, Lingxi Xie, Wengang Zhou, Houqiang Li, and Qi Tian. Mvp: Multimodality-guided visual pre-training. arXiv preprint arXiv:2203.05175, 2022.", + "[63] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018.", + "[64] Haiyang Xu, Ming Yan, Chenliang Li, Bin Bi, Songfang Huang, Wenming Xiao, and Fei Huang. E2e-vlp: End-to-end vision-language pre-training enhanced by visual learning. arXiv preprint arXiv:2106.01804, 2021.", + "[65] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021.", + "[66] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021.", + "[67] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New" + ], + "bbox": [ + 78, + 125, + 470, + 866 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014.", + "[68] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021.", + "[69] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017.", + "[70] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021.", + "[71] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2021.", + "[72] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020." + ], + "bbox": [ + 502, + 127, + 892, + 435 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11005", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_model.json b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..836580d029c14774293ca68f077d816f91f0dbce --- /dev/null +++ b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.793, + 0.088 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.162, + 0.784, + 0.205 + ], + "angle": 0, + "content": "MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.228, + 0.856, + 0.294 + ], + "angle": 0, + "content": "Xiaoyi Dong\\(^{1*}\\), Jianmin Bao\\(^{2*}\\), Yinglin Zhang\\(^{3}\\), Ting Zhang\\(^{2}\\), Dongdong Chen\\(^{4,\\dagger}\\), Hao Yang\\(^{2}\\), Ming Zeng\\(^{3}\\), Weiming Zhang\\(^{1}\\), Lu Yuan\\(^{4}\\), Dong Chen\\(^{2}\\), Fang Wen\\(^{2}\\), Nenghai Yu\\(^{1}\\) \n\\(^{1}\\)University of Science and Technology of China \\(^{2}\\)Microsoft Research Asia \n\\(^{3}\\)Xiamen University \\(^{4}\\)Microsoft Cloud + AI" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.297, + 0.747, + 0.31 + ], + "angle": 0, + "content": "{dlight@mail., zhangwm@, ynh@}.ustc.edu.cn cddlyf@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.313, + 0.74, + 0.326 + ], + "angle": 0, + "content": "{jianbao, ting.zhang, luyuan, doch, fangwen}@microsoft.com" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.329, + 0.778, + 0.342 + ], + "angle": 0, + "content": "{zhengyinglin@stu，zengming@}xmu.edu.cn yanghao.alexis@foxmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.374, + 0.314, + 0.388 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.473, + 0.722 + ], + "angle": 0, + "content": "This paper presents a simple yet effective framework MaskCLIP, which incorporates a newly proposed masked self-distillation into contrastive language-image pretraining. The core idea of masked self-distillation is to distill representation from a full image to the representation predicted from a masked image. Such incorporation enjoys two vital benefits. First, masked self-distillation targets local patch representation learning, which is complementary to vision-language contrastive focusing on text-related representation. Second, masked self-distillation is also consistent with vision-language contrastive from the perspective of training objective as both utilize the visual encoder for feature aligning, and thus is able to learn local semantics getting indirect supervision from the language. We provide specially designed experiments with a comprehensive analysis to validate the two benefits. Symmetrically, we also introduce the local semantic supervision into the text branch, which further improves the pretraining performance. With extensive experiments, we show that MaskCLIP, when applied to various challenging downstream tasks, achieves superior results in linear probing, finetuning, and zero-shot performance with the guidance of the language encoder. Code will be release at https://github.com/LightDXY/MaskCLIP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.751, + 0.21, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.832 + ], + "angle": 0, + "content": "Vision-language (VL) contrastive learning [31, 51] has shown remarkable success in pretraining for various tasks. With large-scale image-text pairs available on the Internet, the model composed of a simple dual encoder design learns" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.375, + 0.895, + 0.459 + ], + "angle": 0, + "content": "strong semantic prior by aligning between image and text. The resulting visual encoder not only exhibits excellent linear probing and finetuning performance, but also enables impressive zero-shot performance with the guidance of the language encoder, showing the generality of natural language and its ability to supervise a wide range of visual concepts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.572 + ], + "angle": 0, + "content": "Nonetheless, the associated language description, though providing richer information than mere class labels, still can hardly describe all the information in the corresponding image, as images are continuous signals with fine-grained details and complex semantics. As a result, the VL contrastive by aligning global representations may only focus on the text-described objects and ignore the rest which might be useful for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.897, + 0.868 + ], + "angle": 0, + "content": "In this paper, we are interested in how to fully leverage the image itself to facilitate the VL contrastive to further improve the transfer capability. (1) Firstly, the learned feature representation shall characterize local patches, serving as a complementary for global representation in VL contrastive. Inspired by the recent success of masked image modeling [4, 19, 26, 51, 60, 61] in learning patch representations, we also randomly mask the input image with a large portion to force the visual encoder to focus on the remaining visible patches. (2) Secondly, the learned representation for local patches shall possess semantic meanings, being consistent with the global representation receiving semantic text supervision. We bring mean teacher self-distillation [25, 57] to supervise the learned patch representations with the visual feature representations, enabling implicit supervision from natural language. The resulting objective is denoted as masked self-distillation where the student model and the teacher model come from the same neural networks and the knowledge is distilled from the full image (fed to the teacher model) to the masked image (fed to student model). To this end, we introduce MaskCLIP by incorporating masked self" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.843, + 0.339, + 0.855 + ], + "angle": 0, + "content": "*Equal contribution, † Corresponding Author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.855, + 0.418, + 0.866 + ], + "angle": 0, + "content": "†Work done during an internship at Microsoft Research Asia" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.843, + 0.418, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "10995" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.47, + 0.154 + ], + "angle": 0, + "content": "distillation into VL contrastive to advance the transferable visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.155, + 0.473, + 0.417 + ], + "angle": 0, + "content": "There are several recent attempts [49, 68] also exploring the capability of the visual encoder under natural language supervision. The common approach is to introduce contrastive learning or masked image modeling on the vision side together with contrastive language-image pretraining. However, the performance indeed improves based on CLIP but does not as well as our masked self-distillation. We argue that (1) the contrastive learning objective based on central crop augmentation actually learns global representations for salient objects while lack of attention on the surrounding backgrounds [11]; and (2) masked image modeling usually needs to remap the learned representation to pixels [26] or discrete tokens [4]. Such low-level prediction target is inefficient for semantic feature learning and thus also conflicts with high-level language supervision in VL contrastive. A brief illustration is presented in Figure 1. In the experiments, we conduct comprehensive ablations to analyze the difference and provide numerical and visual evidence for better understanding." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.473, + 0.582 + ], + "angle": 0, + "content": "Symmetrically, we argue that local semantic supervision on the text branch is also helpful for the text encoder and eventually beneficial for zero-shot performance. So we introduce the same mask-data-modeling format supervision into the text branch as well. Different from images where the pixel is low-level signal, the words crafted by human beings are already highly semantic, so we use the tokenized word piece as the prediction target directly, following the well-studied mask language modeling method BERT. Meanwhile, to reduce the output conflicts between contrastive learning and mask language modeling, we introduce a small decoder for the mask language modeling branch." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.582, + 0.473, + 0.831 + ], + "angle": 0, + "content": "We train our MaskCLIP on a subset of a publicly available image-text pairs dataset, YFCC [58], and thoroughly evaluate the transfer ability of visual representations on several vision benchmarks: ImageNet-1K [17] for classification, ADE20K [69] for semantic segmentation, MS-COCO [40] for detection and segmentation, as well as a batch of other classification benchmarks. When it comes to ImageNet-1K [17] classification, MaskCLIP achieves \\(+6.9\\%\\), \\(+7.2\\%\\), \\(+1.3\\%\\) higher than CLIP for zero-shot transfer, linear probing, and finetuning respectively. For vision downstream tasks, we reach \\(+2.7\\) mIoU on ADE20K [69] and \\(+1.8\\) APb, \\(+1.4\\) APm on MS-COCO [40]. For vision-language tasks, MaskCLIP achieves \\(+6.1\\%\\) average zero-shot accuracy on 20 datasets, and \\(+17.2\\%\\), \\(+12.8\\%\\) rank@1 improvement on the Flickr30K [67] image-test retrieval. In the recent Image Classification in the Wild challenge academic track, our MaskCLIP gets the \\(1_{st}\\) result with \\(48.9\\%\\) TOP-1 average accuracy, surpassing the second team with \\(3.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.831, + 0.451, + 0.845 + ], + "angle": 0, + "content": "In summary, the major contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.853, + 0.47, + 0.868 + ], + "angle": 0, + "content": "1. We present a novel vision-language pretraining" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.127, + 0.895, + 0.168 + ], + "angle": 0, + "content": "framework MaskCLIP, by introducing masked self-distillation objective to facilitate VL contrastive for better transferable visual models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.175, + 0.895, + 0.232 + ], + "angle": 0, + "content": "2. We present extensive ablation studies on MaskCLIP variants and provide in-depth analysis numerically and visually to help understand how the proposed masked self-distillation assists VL contrastive." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.895, + 0.281 + ], + "angle": 0, + "content": "3. We demonstrate our MaskCLIP on tens of benchmarks, showing the superiority under all three settings: zero-shot, linear probing, and finetuning." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.175, + 0.895, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.291, + 0.642, + 0.306 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.896, + 0.687 + ], + "angle": 0, + "content": "Vision-language pretraining. Recent years have seen rapid progress made in vision-language pretraining [15, 18, 33, 35-39, 44-46, 50, 55, 56, 72]. Several multiple cross-modality loss functions have been proposed for the training objective, such as image-text matching [15, 37, 44, 56, 64], masked language modeling [15, 37, 44, 55, 56], masked image modeling [15, 44, 55, 56], contrastive loss [35, 38, 39]. These objects are often mixed with each other to form a compound objective. While a variety of approaches have been proposed, few works investigate the performance on visual representation learning for image classification. Recently, CLIP [51] and ALIGN [31] show that the image-text contrastive learning objective achieves promising performance for visual representation learning. There are many following works proposed to further improve the pretraining performance, DeCLIP [70], SLIP [49], COTS [43], ViCHA [54], CYCLIP [24] use additional uni/multi-modality supervision to improve the model capability, and PyramidCLIP [23], KLITE [53], IDEA [30] seek to external knowledge from pre-trained models or datasets as the additional guidance. FILIP [66] and LOUPE [34] introduce fine-grained alignment to the model. Focusing on this research direction, we analyze the desired properties of supervision which could be complementary to CLIP, and propose the masked self-distillation objective incorporated with the image-text contrastive loss to further improve pretraining performance for various visual understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.897, + 0.868 + ], + "angle": 0, + "content": "Self-supervised learning Self-supervised visual representation learning has attracted increasing attention over the past few years. The objective of the self-supervised learning is mainly divided into two categories: contrastive and generative [41]. The contrastive methods, such as MOCO [12, 27], SimCLR [9, 10], BYOL [25], SimSiam [13], and DINO [6] measure the similar and dissimilar samples by contrastive loss. Their success heavily depends on the strong data augmentation. The generative methods, such as BEiT [4], MAE [26], PeCo [19], BEVT [60], BootMAE [20] and MaskFeat [61] leverage masked image modeling to reconstruct the remaining masked part of its original input from the given visible parts. The generative methods show more" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "10996" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.123, + 0.201, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.314, + 0.171, + 0.326 + ], + "angle": 0, + "content": "(a) CLIP" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.127, + 0.398, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.314, + 0.38, + 0.326 + ], + "angle": 0, + "content": "(b) + Contrastive Learning" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.127, + 0.577, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.314, + 0.565, + 0.326 + ], + "angle": 0, + "content": "(c) + Mask Image Modeling" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.123, + 0.895, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.314, + 0.779, + 0.326 + ], + "angle": 0, + "content": "(d) MaskCLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.334, + 0.894, + 0.386 + ], + "angle": 0, + "content": "Figure 1. Pipeline comparison between combination CLIP with different vision self-supervised learning methods. (a) Vanilla CLIP. (b) CLIP + contrastive learning. (c) CLIP + pixel prediction mask image modeling. (d) CLIP + mask self-distillation, i.e. MaskCLIP. The \\( E_{T} \\), \\( E_{I} \\) is the text encoder and image encoder respectively, and all the \\( E_{I} \\), \\( E_{T} \\) within each pipeline share the weight. \\( \\bar{E}_{I} \\) is the mean-teacher model, whose weight is updated by the exponential moving average of \\( E_{I} \\) and does not require gradient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.471, + 0.451 + ], + "angle": 0, + "content": "promising transfer performance than the contrastive methods, as generative objective learns patch representations while contrastive objective focuses on learning centric global representations [11]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.471, + 0.646 + ], + "angle": 0, + "content": "Self-knowledge distillation Self-knowledge distillation [32] aims to distill the knowledge in a model itself and uses it for training the model. Instead of distilling knowledge from a pretrained teacher model [29], self-knowledge distillation regards a temporal ensemble of the student model as the teacher. It means that a student model becomes a teacher model itself, which gradually utilizes its own knowledge for softening the hard targets to be more informative during training. Self-knowledge distillation has been explored in semi-supervised learning [57], contrastive learning [16, 35], self-supervised learning [3, 7]. In this paper, we use visual features supervised by natural language for guidance in masked self-distillation which naturally fit VL contrastive to learn more transferable visual representations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.195, + 0.671 + ], + "angle": 0, + "content": "3. MaskCLIP" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.805 + ], + "angle": 0, + "content": "We introduce MaskCLIP, a novel framework that learns visual representations. The core part of MaskCLIP is its backbone image encoder, denoted by \\( E_{I} \\) as shown in Figure 1. It obtains the transferable capability during pretraining that could benefit downstream vision tasks. Following recent self-supervised approaches [4, 14, 26, 49], we implement the backbone \\( E_{I} \\) as a Vision Transformer (ViT) [22]. The prediction results from \\( E_{I} \\) given an input image \\( I \\) then should be a collection of visual feature tokens, represented as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.815, + 0.47, + 0.83 + ], + "angle": 0, + "content": "\\[\nE _ {I} (I) = \\left\\{f _ {\\text {c l s}}, f _ {1}, f _ {2}, \\dots , f _ {N} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Here \\(cls\\) is short for class token. \\(1,\\ldots ,N\\) are the indexes of the non-class tokens." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.894, + 0.453 + ], + "angle": 0, + "content": "The rest of this section starts with the utilization of language supervision. More shall be emphasized on the masked self-distillation, which we deem crucial for visual pretraining." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.473, + 0.753, + 0.488 + ], + "angle": 0, + "content": "3.1. Vision-language Contrastive" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.499, + 0.893, + 0.61 + ], + "angle": 0, + "content": "Following [31, 51], we introduce a Transformer-based text encoder \\( E_{T} \\) to leverage language knowledge. It aims to align the global feature representations of an image and a text with respect to some forms of similarity. Precisely, consider a given image-text pair \\( \\{I, T\\} \\), besides extracting the visual feature representation \\( E_{I}(I) \\) using the vision backbone as shown by Equation 1, we additionally use the text encoder \\( E_{T} \\) to extract linguistic features from the text \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.614, + 0.894, + 0.697 + ], + "angle": 0, + "content": "The mean feature of the two branches are regarded as the global representations and are fed into a projection head (implemented as a fully-connected layer) respectively to obtain the metric embeddings \\(e^T\\) and \\(e^I\\). Image-text contrastive loss is employed to align them during pretraining. The loss can be formulated as \\(\\mathcal{L}_T + \\mathcal{L}_I\\), with" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.715, + 0.831, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp (e _ {i} ^ {I} e _ {i} ^ {T} / \\sigma)}{\\sum_ {j = 1} ^ {B} \\exp (e _ {i} ^ {I} e _ {j} ^ {T} / \\sigma)}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.756, + 0.892, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(e _ {i} ^ {T} e _ {i} ^ {I} / \\sigma\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(e _ {i} ^ {T} e _ {j} ^ {I} / \\sigma\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.811, + 0.892, + 0.868 + ], + "angle": 0, + "content": "where \\(B\\) stands for the number of image-text pairs within a training mini-batch, \\(i,j\\) are indexes within the batch; \\(\\sigma\\) stands for the temperature for the loss functions, which is learned together with all other parameters during training." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "10997" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.126, + 0.444, + 0.14 + ], + "angle": 0, + "content": "3.2. Masked Self-distillation for Visual Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.147, + 0.473, + 0.312 + ], + "angle": 0, + "content": "Knowledge distillation is a learning paradigm where a student model is trained to match the output of a given teacher model, so that the student model can be improved by the teacher. Instead of bringing in an external teacher, self-distillation methods such as [7, 25, 57] proposes using a mean teacher model that is derived from the student itself. In specific, the teacher shares the same structure with the student, while the parameters of the teacher are exponential moving averages (EMA) of the parameters from the student. In the following, we would use the term \"EMA model\" to represent such mean teacher model constructed from the student." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.313, + 0.473, + 0.369 + ], + "angle": 0, + "content": "MaskCLIP leverages the mean teacher self-distillation to enhance its vision representations. Let \\(\\bar{E}_I\\) be the EMA model of the backbone encoder \\(E_I\\). \\(\\theta_t\\) and \\(\\bar{\\theta}_t\\) are the parameters of \\(E_I\\) and \\(\\bar{E}_I\\) at training step \\(t\\). \\(\\bar{\\theta}_t\\) is updated with" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.376, + 0.47, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\bar {\\theta} _ {t} = \\alpha \\bar {\\theta} _ {t - 1} + (1 - \\alpha) \\theta_ {t}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.453 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyper-parameter for smoothing updates. We propose to incorporate masked image modeling into self-distillation, resulting in masked self-distillation with asymmetric input for student model and teacher model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.473, + 0.507 + ], + "angle": 0, + "content": "In specific, considering a given input image \\(I\\), we first feed it to the EMA model \\(\\bar{E}_I\\) (teacher model) to obtain the distillation targets. These target features can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.508, + 0.47, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\bar {E} _ {I} (I) = \\{\\bar {f} _ {c l s}, \\bar {f} _ {1}, \\bar {f} _ {2}, \\dots , \\bar {f} _ {N} \\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.473, + 0.728 + ], + "angle": 0, + "content": "In the meantime, we randomly mask a large portion of the input image patches and then feed it into the original backbone \\(E_I\\) (student model). Following [26], we only feed the visible (unmasked) patches, denoted by \\(I'\\), into the original backbone \\(E_I\\) to speed up computation and save memory. Let \\(\\mathcal{M}\\) be the indexes of all the masked tokens. These encoded features corresponding to visible tokens can then be denoted as \\(E_I(I') = \\{f_{cls}'\\} \\cup \\{f_{k \\notin \\mathcal{M}}'\\}\\). They are then joined with a shared and learnable feature vector, denoted as \\(m\\), that represents mask tokens, to form a complete set of features \\(\\{f_{cls}', f_1', f_2', \\ldots, f_N'\\}\\), with \\(f_{i \\in \\mathcal{M}}' = m\\). We attach positional embeddings onto all these tokens, and append a small Transformer \\(D\\) as a decoder to predict features of the masked region from the visible tokens, which could be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.734, + 0.47, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(D \\circ E _ {I}\\right) \\left(I ^ {\\prime}\\right) = D \\left(\\left\\{f _ {c l s} ^ {\\prime}, f _ {1} ^ {\\prime}, f _ {2} ^ {\\prime}, \\dots , f _ {N} ^ {\\prime} \\right\\}\\right) \\\\ = \\left\\{f _ {c l s} ^ {\\prime \\prime}, f _ {1} ^ {\\prime \\prime}, f _ {2} ^ {\\prime \\prime}, \\dots , f _ {N} ^ {\\prime \\prime} \\right\\}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.47, + 0.83 + ], + "angle": 0, + "content": "Inspired by [71], we use an online quantizer \\( h(\\cdot) \\) to transform the output features into a soft codewords distribution, and minimize the cross-entropy between the target features and the predicted features. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.836, + 0.47, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {D i s t}} = \\frac {1}{| \\mathcal {M} |} \\sum_ {k \\in \\mathcal {M}} - \\bar {h} \\left(\\bar {f} _ {k}\\right) ^ {T} \\log h \\left(f _ {k} ^ {\\prime \\prime}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.126, + 0.894, + 0.154 + ], + "angle": 0, + "content": "here the parameter of the teacher quantizer \\(\\bar{h}()\\) is also EMA updated by the online quantizer, similar to the teacher model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.162, + 0.864, + 0.177 + ], + "angle": 0, + "content": "3.3. Local Semantic Learning for Text Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.894, + 0.293 + ], + "angle": 0, + "content": "Besides the local semantic supervision for the visual encoder, we argue it is also helpful for the text encoder. So we introduce the BERT pretraining into the text branch. For the text \\( T = \\{t_{\\text{sos}}, t_1, t_2, \\dots, t_M, t_{\\text{eos}}\\} \\), we denote the masked input as \\( T' = \\{t_{\\text{sos}}', t_1', t_2', \\dots, t_M', t_{\\text{eos}}'\\} \\), where \\( t_{i \\in \\mathcal{M}_T}' = m_t \\) and \\( t_{i \\notin \\mathcal{M}_T}' = t_i \\), and \\( \\mathcal{M}_T \\) be the indexes of all the masked text tokens. The output feature of the encoder is \\( E_T(T') \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.894, + 0.404 + ], + "angle": 0, + "content": "To reduce the output conflict between the global image-text contrastive learning and the local mask language modeling, we further introduce a small text decoder, which shares the same architecture as the encoder but with only a few layers. So that the global prediction and local prediction are conducted at different layers. We denote the output feature as: \\((D_T\\circ E_T)(T') = \\{t_{sos}''\\), \\(t_1''\\), \\(t_2''\\), ..., \\(t_M''\\), \\(t_{eos}''\\}\\) and the loss could be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.41, + 0.892, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M L M} = \\frac {1}{| \\mathcal {M} _ {T} |} \\sum_ {k \\in \\mathcal {M} _ {T}} - t _ {k} ^ {T} \\log t _ {k} ^ {\\prime \\prime}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.452, + 0.715, + 0.466 + ], + "angle": 0, + "content": "3.4. Overall Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.894, + 0.5 + ], + "angle": 0, + "content": "Finally, we pretrain MaskCLIP with all these losses combined:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.511, + 0.892, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I} + \\mathcal {L} _ {T} + \\lambda \\mathcal {L} _ {\\text {D i s t}} + \\beta \\mathcal {L} _ {\\text {M L M}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.894, + 0.604 + ], + "angle": 0, + "content": "with \\(\\lambda, \\beta\\) being the hyper-parameter weighting between VL contrastive loss and self-supervised learning loss. All the components of MaskCLIP are trained from scratch, including the visual backbone \\(E_I\\), the visual decoder \\(D\\), the text encoder \\(E_T\\), as well as the text decoder \\(D_T\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.615, + 0.634, + 0.631 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.639, + 0.583, + 0.653 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.894, + 0.825 + ], + "angle": 0, + "content": "Model architecture. Our framework consists of the visual encoder \\( E_{I} \\), the text encoder \\( E_{T} \\), the visual decoder \\( D \\), and the text decoder \\( D_{T} \\). We adopt the widely used Transformer ViT-B/16 [22] for a fair comparison. It is composed of 12 layers, 768 width, and 12 head. The input image is \\( 224 \\times 224 \\) resolution and is further split into \\( 14 \\times 14 \\) patches with size \\( 16 \\times 16 \\). A learnablecls token is pretended to the 196 embeddings. For the text encoder, we adopt a 12-layer, 512-width, and 8-head Transformer following CLIP [51], and the text decoder has 4 layers. The number of text tokens is fixed to 77 with necessary truncations or paddings. For the image decoder, we directly use a one-layer Vision Transformer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.867 + ], + "angle": 0, + "content": "Pretraining details. We train our proposed MaskCLIP from scratch for 25 epochs, the batch size is fixed to 4096 for all the experiments. The masks used in the mask self-distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "10998" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.12, + 0.875, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.432, + 0.893, + 0.47 + ], + "angle": 0, + "content": "Figure 2. Visualization of the similarity between text and image features. The images and captions are from the MS-COCO val set. Here we show the image feature similarity with both full caption and different objects in it. The caption is \"Three teddy bears sit in a sled in snow\". More results could be found in the supplemental materials." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.482, + 0.482, + 0.572 + ], + "angle": 0, + "content": "
TrainingIN-1KFlicker30K
MemoryTime0-shotLinearFinetuneI2TT2I
CLIP14G1.00×37.666.582.352.932.8
CLIP+SimCLR30G2.67×42.872.182.658.641.3
CLIP+MAE16G1.30×42.168.583.257.341.1
MaskCLIP19G1.75×44.573.783.670.145.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.574, + 0.47, + 0.6 + ], + "angle": 0, + "content": "Table 1. Results of boosting CLIP with different kinds of vision self-supervised learning methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.469, + 0.665 + ], + "angle": 0, + "content": "branch and mask language modeling branch are random mask with a mask ratio of \\(75\\%\\) and \\(20\\%\\). We pretrain all the models with the commonly used YFCC15M dataset, which is flited from the YFCC100M [58] dataset by [51]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.805 + ], + "angle": 0, + "content": "Downstream details. We evaluated MaskCLIP on several downstream datasets, including ImageNet-1K [17], ADE20K [69], MS-COCO [40], Flicr30K [67] et al. For ImageNet-1K, we report zero-shot, linear probing, and finetuning performance. The zero-shot is conducted following the label prompt setting in SLIP [49]. For linear probing, we fix the backbone and train a new linear classifier for 90 epochs. For finetuning, we follow the setting in BEiT [4] and finetune the model for 100 epochs with a layer-decayed learning rate. See supplemental materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.18, + 0.831 + ], + "angle": 0, + "content": "4.2. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.867 + ], + "angle": 0, + "content": "We first present our analysis by studying different ways of boosting CLIP. The baseline is CLIP [51] trained on the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.893, + 0.568 + ], + "angle": 0, + "content": "YFCC-15M. Besides the introduced masked self-distillation, we consider two other popular methods: (1) SimCLR [9], a representative contrastive method; and (2) MAE [26] the state-of-the-art masked image modeling approaches. All the compared methods are trained on the YFCC-15M for a fair comparison. We have the following observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.894, + 0.793 + ], + "angle": 0, + "content": "Vision self-supervision helps VL contrastive. We evaluate the models on both vision task ImageNet-1K [17] classification and vision-language task image-text retrieval on Flicker30K [67] and present the comparison in Table 1. All the added vision self-supervision, regardless of contrastive or generative, improves the baseline CLIP. Among them, our proposed MaskCLIP achieves the best results in terms of all the evaluation metrics, outperforming CLIP with \\(+6.9\\%\\), \\(+7.2\\%\\), \\(+1.3\\%\\) on ImageNet-1K classification for zero-shot, linear probing, and finetuning respectively, and \\(+17.2\\%\\), \\(+12.8\\%\\) on Flicker30K for image-to-text retrieval and text-to-image retrieval. We also report the training GPU memory usage and time-consuming cost in Table 1. It is worth noting that the contrastive model (CLIP + SimCLR) compares two additional views of the input image, resulting in larger GPU memory usage and longer training time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.894, + 0.868 + ], + "angle": 0, + "content": "Masked image modeling is able to learn representations for local patches. We argue that the image encoder only pays attention to the text-described objects under VL contrastive due to sparse text description and to the centric objects under image contrastive due to central-crop augmenta" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "10999" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.085, + 0.125, + 0.472, + 0.214 + ], + "angle": 0, + "content": "
MethodObjectiveADE20K mIoUPascal mIoU
CLIPGlobal7.213.5
CLIP + SimCLRGlobal + Global6.311.9
CLIP + MAEGlobal + Pixel-wise Local8.316.4
MaskCLIPGlobal + Token-wise Local10.217.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.217, + 0.47, + 0.241 + ], + "angle": 0, + "content": "Table 2. Annotation-free zero-shot segmentation results on ADE20K and Pascal Context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.249, + 0.471, + 0.47 + ], + "angle": 0, + "content": "tion. In contrast, masked image modeling forces the image encoder to focus on local patches using token-wise objectives by mandatorily masking a large portion of patches. Here, we provide numerical comparisons for evidence. We conduct an \"Annotation-free zero-shot segmentation\" experiment to test the zero-shot segmentation. The results on such a dense prediction task would better reveal the ability of local patch representations than global classification. Following the design in DenseCLIP [70], we use the prompted label feature as the linear classification weight to realize segmentation, without any training procedure. We evaluate the performance on two widely used datasets: ADE20K [69] and Pascal Context [48]. The results are shown in Table.2. We can see that equipped with masked image modeling, our MaskCLIP as well as CLIP+MAE achieves better results than CLIP and CLIP+SimCLR, validating our hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.471, + 0.803 + ], + "angle": 0, + "content": "Masked self-distillation learns semantic representations for local patches. Our masked self-distillation predicts visual features dynamically outputted by the visual encoder and thus implicitly gets supervision from the text side via VL contrastive. While MAE predicts fixed low-level pixels, making it inefficient to learn semantic representations (as the objective may force the representation to memorize low-level details) and thus causing conflict with VL contrastive. To show this, we select images from MS-COCO [40] and calculate the feature similarity between image features and their corresponding caption features. We also select objects in the caption, prompt it to a new caption, such as \"a photo of teddy bears\", and calculate the similarities. An example is shown in Figure 2 (More can be found in the supplementary material). Comparing MaskCLIP with CLIP+MAE in the fourth column, we can see that CLIP+MAE uses color as evidence and fails to distinguish the white teddy bear from the white snow. While our MaskCLIP successfully differentiates the two objects, suggesting ours learn more semantic features. On the other hand, the superior results of MaskCLIP shown in Table 1 and Table 2 also validate this. It is worth mentioning that CLIP and CLIP+SimCLR fail to have a correct response partition for different single objects like MaskCLIP, further justifying our second observation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.391, + 0.822 + ], + "angle": 0, + "content": "4.3. Comparison with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.867 + ], + "angle": 0, + "content": "To show the effectiveness of MaskCLIP as a general vision-language pretrain method, we conduct experiments on both vision tasks and vision-language tasks. For vision" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.125, + 0.895, + 0.243 + ], + "angle": 0, + "content": "
MethodEpochIN-1KADE20KMS-COCO
0-ShotLinFTmIoUAPbAPm
DeiT [59]300*--81.847.444.139.8
SimCLR [9]25-64.082.548.044.640.2
MAE [26]25-56.282.546.543.239.1
CLIP [51]2537.666.582.347.843.639.5
SLIP [49]2542.872.182.648.544.040.3
MaskCLIP2544.573.783.650.545.440.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.246, + 0.893, + 0.298 + ], + "angle": 0, + "content": "Table 3. Comparison with previous methods, including supervised baselines, self-supervised pretraining methods, and vision-language pretraining methods. * is the epoch of the supervised baseline on ImageNet-1K." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.895, + 0.463 + ], + "angle": 0, + "content": "tasks, we report results on ImageNet-1K [17] classification, MS-COCO [40] object detection, and ADE20K [69] semantic segmentation. For vision-language tasks, we report zero-shot results on recent challenging ICinW 20 datasets benchmark and image-text retrieval results on Flickr30K [67] and MS-COCO [40]. In the following, we compare with the supervised baseline DeiT [59], self-supervised methods SimCLR [9] and MAE [26], and vision-language methods CLIP [51] and SLIP [49]. For a fair comparison, we train SimCLR and MAE on YFCC-15M [58] with the same epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.574 + ], + "angle": 0, + "content": "Classification on ImageNet-1K. As shown in Table 3, MaskCLIP benefits from the advantages of both VL pretraining and image mask self-distillation that shows strong performance on all the metrics. For zero-shot tasks, MaskCLIP outperforms CLIP by \\(+6.9\\%\\) with 25 epoch training and achieves \\(+1.7\\%\\) higher than the recent work SLIP. When it comes to finetune, MaskCLIP reaches \\(83.6\\%\\) top-1 accuracy, and outperforms CLIP by \\(+1.3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.894, + 0.699 + ], + "angle": 0, + "content": "Semantic segmentation on ADE20K. Then we apply our MaskCLIP to the semantic segmentation task. Here we use the UperNet [63] framework with \\(512 \\times 512\\) input and end-to-end training for 160K iterations. The evaluation metric is the mean Intersection of Union (mIoU) and we report single-scale evaluation results here. The results are given in Table 3. Our method achieves \\(50.5\\mathrm{mIoU}\\), \\(+2.7\\mathrm{mIoU}\\) than our baseline method CLIP, and \\(+2.0\\mathrm{mIoU}\\) than SLIP. This verifies the effectiveness of our introduced incorporation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.895, + 0.796 + ], + "angle": 0, + "content": "Object detection and instance segmentation on MS-COCO. We further investigate our transfer performance on object detection and instance segmentation in Table.3. Here we use Mask-RCNN [28] framework with single-scale input and \\(1 \\times\\) schedule (12 epochs). Our method achieves 45.4 box AP and 40.9 mask AP, \\(+1.8 / 1.4\\) better than CLIP, and \\(+1.4 / 0.6\\) better than SLIP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Zero-shot on small datasets. We also report zero-shot performance on 20 small datasets under the ICinW setting (see the introduction below) in Table 4. We find that all the methods perform poorly on some datasets such as Aircraft(1% acc for random guessing, we omit the description in the" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "11000" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.118, + 0.892, + 0.309 + ], + "angle": 0, + "content": "
AverageCaltech-101CIFAR-10CIFAR-100Country211DTDEuroSATFER-2013AircraftFood-101GTSRBMemesKittiDisMNISTFlowersPetsPatchCamSST2RESISC45CarsVoc2007
Pretraining on YFCC-15M
CLIP34.058.668.536.910.821.430.516.95.151.66.551.125.95.052.728.651.752.522.44.579.1
SLIP37.870.982.648.611.826.619.818.15.659.912.651.829.49.856.331.455.351.528.55.480.5
MaskCLIP40.172.080.257.512.627.944.020.36.164.98.552.034.34.957.034.350.149.935.76.782.1
Pretraining on ICinW Academic Track Siting: YFCC-15M, GCC3M+12M, ImageNet-21K(ImageNet-1K is removed)
1st MaskCLIP48.986.495.378.311.633.057.718.88.078.917.352.816.07.374.274.452.146.254.326.582.3
2nd KLITE*45.587.492.768.88.232.227.917.44.372.411.448.431.112.875.665.950.652.944.410.282.3
3rd YT-CLIP44.577.883.558.411.931.940.727.16.968.718.852.39.118.853.169.351.550.352.719.779.3
4th UniCL†44.084.890.267.86.725.435.330.83.568.311.151.017.911.371.744.952.149.541.424.281.3
5th Gramer*43.283.992.969.57.325.524.430.42.771.09.052.612.410.170.452.450.650.144.813.881.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.312, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Table 4. Zero-shot evaluation on ICinW classification benchmarks. Best results in bold. * indicates using Swin-B as the backbone, † indicates using Focal-B as the backbone." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.34, + 0.891, + 0.428 + ], + "angle": 0, + "content": "
Training EpochFlickr30KMS-COCO
Image-to-textText-to-imageImage-to-textText-to-image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP [51]2552.979.687.232.860.871.227.553.565.017.738.850.5
SLIP [49]2558.685.191.741.368.778.633.459.870.621.544.456.3
MaskCLIP2570.190.395.345.673.482.141.467.977.525.549.761.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.43, + 0.807, + 0.442 + ], + "angle": 0, + "content": "Table 5. Results of zero-shot image-text retrieval on Flickr30K and MS-COCO datasets. Best results in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.471, + 0.54 + ], + "angle": 0, + "content": "following), \\(\\mathrm{Fer}(24.7\\%)\\), Country211 \\((0.5\\%)\\), GTSRB \\((5.9\\%)\\), Cars \\((0.8\\%)\\). This might be caused by the data domain gap that the YFCC-15M contains few related images and descriptions. For the rest of the datasets, all the methods get reasonable performance and our MaskCLIP gets the best performance on most datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.474, + 0.752 + ], + "angle": 0, + "content": "Image Classification in the Wild (ICinW) Challenge The ICinW challenge [1] is a newly proposed visual pretraining benchmark, which contains 20 diverse downstream classification datasets, measuring the ability of pre-training models on both the prediction accuracy and their transfer efficiency in a new task. The pretraining is limited to three datasets: YFCC-15M [58], GCC3M [52] \\(+12\\mathrm{M}\\) [8] and ImageNet-21K [17] (ImageNet-1K data is excluded). We pretrain our MaskCLIP on it and get the \\(\\mathbf{1}_{st}\\) result in the zero-shot track [2] (we submit the results anonymously). As shown in Table 4, the \\(2_{nd}\\) team KLITE uses a strong Swin-B [42] as the backbone and additional knowledge from GPT-3 [5] and Wiktionary [47], and the \\(4_{th}\\) use the strong Focal-B [65] as the backbone, while our MaskCLIP greatly outperforms these methods with a simple ViT-B backbone and no additional knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.472, + 0.867 + ], + "angle": 0, + "content": "Zero-shot on text-image retrieval. We further report the zero-shot text-image retrieval results on two benchmark datasets, Flickr30K [67] and MS-COCO [40]. We find that the text without any prefixes or suffixes works well for all the models. Table 5 shows the results. We can see that MaskCLIP exhibits a strong zero-shot performance. For example, with 25 epochs training, MaskCLIP reaches \\(41.4\\%\\) Rank@1 image-to-text accuracy on MS-COCO, outperform-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.457, + 0.892, + 0.485 + ], + "angle": 0, + "content": "ing CLIP with \\(13.9\\%\\) and \\(25.5\\%\\) Rank@1 text-to-image accuracy, \\(+7.8\\%\\) higher than CLIP." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.49, + 0.611, + 0.503 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.891, + 0.535 + ], + "angle": 0, + "content": "We compare our default settings with other alternatives to justify the efficacy of our model designs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.894, + 0.645 + ], + "angle": 0, + "content": "Training objectives ablation. As shown in Table.6a, when we remove the mask language modeling loss \\(\\mathcal{L}_{\\mathrm{MLM}}\\), the performance of the image-text task drops, including the zero-shot accuracy and retrial performance. While benefiting from the distillation loss, the finetuning performance on ImageNet-1K is not influenced. When we remove the distillation loss \\(\\mathcal{L}_{\\mathrm{Dis}}\\), we observe a performance drop on all tasks, especially the finetuning results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.895, + 0.797 + ], + "angle": 0, + "content": "Distillation loss format. Different from previous methods [3,21,26] that calculate the per-element distance as the loss function, we use an online tokenizer to map the feature to soft codewords and use the cross-entropy loss as the supervision. Here we study their difference in Table.6b. We find that although they get similar fine-tuning performance, the CE loss gets better zero-shot and linear probing performance. The reason may be that the per-element MSE loss leads the model to fit some unnecessary details of the target feature, while the CE loss with soft tokenizer helps the model to focus more on the important feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Distillation & MLM loss weight. Here we set the loss weight of the CLIP branch as 1 and study the loss weight of the two additional branches. As shown in Table.6e and Table.6f, setting \\(\\lambda = 1\\) or \\(\\beta = 1\\) emphasize too much on new tasks, which mislead the model to a wrong converge" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "11001" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.123, + 0.34, + 0.184 + ], + "angle": 0, + "content": "
Model0-ShotFTI2T/T2I
MaskCLIP44.583.670.1/45.6
w/o LMLM42.883.665.0/41.6
w/o LDis42.082.465.4/40.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.186, + 0.331, + 0.207 + ], + "angle": 0, + "content": "(a) Training Objectives ablation. Both is necessary for MaskCLIP." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.222, + 0.311, + 0.306 + ], + "angle": 0, + "content": "
Depth0-ShotI2T/T2I
043.565.2/44.1
144.370.4/45.3
244.370.2/45.4
444.570.1/45.6
844.267.5/44.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.308, + 0.33, + 0.33 + ], + "angle": 0, + "content": "(d) Text decoder depth. The decoder is necessary and a shallow one works better." + }, + { + "type": "table", + "bbox": [ + 0.386, + 0.123, + 0.589, + 0.184 + ], + "angle": 0, + "content": "
Loss0-ShotLinFT
MSE43.873.283.6
CE44.573.783.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.186, + 0.61, + 0.218 + ], + "angle": 0, + "content": "(b) Distillation loss format. The online tokenizer with cross-entropy loss works slightly better than MSE loss." + }, + { + "type": "table", + "bbox": [ + 0.384, + 0.222, + 0.59, + 0.306 + ], + "angle": 0, + "content": "
Weight0-ShotLinFT
138.568.282.5
0.144.473.583.5
0.0544.573.783.6
0.0143.673.083.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.308, + 0.608, + 0.33 + ], + "angle": 0, + "content": "(e) Distillation loss weight. A small loss weight works well for MaskCLIP." + }, + { + "type": "table", + "bbox": [ + 0.677, + 0.123, + 0.857, + 0.184 + ], + "angle": 0, + "content": "
Depth0-ShotLinFT
144.573.783.6
243.772.983.4
443.572.583.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.641, + 0.186, + 0.888, + 0.208 + ], + "angle": 0, + "content": "(c) Visual decoder Depth. A shallow decoder gets better performance." + }, + { + "type": "table", + "bbox": [ + 0.675, + 0.222, + 0.858, + 0.306 + ], + "angle": 0, + "content": "
Weight0-ShotI2T/T2I
136.551.7/32.1
0.144.369.2/45.9
0.0544.570.1/45.6
0.0143.270.6/45.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.641, + 0.308, + 0.887, + 0.329 + ], + "angle": 0, + "content": "(f) MLM loss weight. A small loss weight works better." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Table 6. MaskCLIP ablation experiments with YFCC-15M dataset. We report zero-shot(0-Shot), fine-tuning (FT), and linear probing (Lin) accuracy (%) for image-encoder-related ablation. And zero shot image-to-text, text-to-image retrieval (I2T/T2I) for text encoder-related ablations. Default settings are marked in gray ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.469, + 0.488 + ], + "angle": 0, + "content": "direction, resulting in poor performance. When we reduce the loss weight by \\(10 \\times\\), the two additional tasks are helpful for the model and show a consistent gain on all the metrics. We suspect this is because the CLIP loss requires two different capabilities: understanding the input content and aligning them into a shared feature space. And the goal of the two additional self-supervised learning tasks is to facilitate understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.469, + 0.725 + ], + "angle": 0, + "content": "Image & Text decoder depth. Then we study the influence of the decoder depth for both image and text decoders. As shown in Table.6c, we find the image decoder with only one layer works well, increasing the decoder depth leads to worse performance on all metrics. Similarly, Table.6d shows that the text branch benefits from a shallow decoder design. We argue that a too-deep decoder would make the encoder lazy, relying on the strong decoder to resolve the challenging mask feature/language modeling tasks. And the different depth choice between the image and text branches is caused by the framework difference: the image branch sees the mask tokens at the decoder, while the text branch takes the mask tokens as the encoder input. Note that if we remove the text decoder, the performance gets worse. We think this is largely caused by the output conflict that the global recognition feature aggregation and local word prediction are conducted at the same layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Single-Stage v.s. two-Stage. Our MaskCLIP learns the VL contrastive and masked self-distillation simultaneously and jointly in a single stage. One possible variant is to first train CLIP and then use CLIP feature from the first stage to train masked image modeling as in [61, 62]. We report results on three datasets in Table 7. We can see that the second stage achieves better finetuning results compared with results from the stage one, showing the effectiveness of masked image modeling. Nonetheless, such two-stage training requires longer training time and loses the transfer capability in a zero" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.375, + 0.893, + 0.466 + ], + "angle": 0, + "content": "
MethodEpochIN-1KFlicker30KADE20K
0-shotFTI2TT2I0-shotFT
Two-StageStage12537.682.352.932.87.247.8
Stage225—83.4———48.2
MaskCLIP2544.583.670.145.610.250.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.469, + 0.893, + 0.494 + ], + "angle": 0, + "content": "Table 7. Comparison between two-stage method and our single-stage MaskCLIP." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.503, + 0.891, + 0.531 + ], + "angle": 0, + "content": "shot setting. In contrast, our MaskCLIP achieves superior results under all settings with fewer epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.536, + 0.618, + 0.55 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.894, + 0.694 + ], + "angle": 0, + "content": "We present MaskCLIP, a new VL pretraining framework that incorporates masked self-distillation into VL contrastive. We point out that masked self-distillation learns local semantics, fitting nicely to the VL contrastive that aims to learn global semantics, and this is supported with comprehensively designed experiments. We also utilize mask language modeling to enhance the text encoder which is critical for zero-shot performance. The resulting visual encoder shows strong transfer capability across widely adopted benchmarks for linear probing, fine-tuning, and also zero-shot evaluation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.667, + 0.721 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.893, + 0.865 + ], + "angle": 0, + "content": "This work was supported in part by the Natural Science Foundation of China under Grant U20B2047, 62072421, 62002334, 62102386, and 62121002. Key Research and Development program of Anhui Province under Grant 2022k07020008, Ant Group through CCF-Ant Innovative Research Program CCF-AFSG RF20210025, Alibaba Group through Alibaba Innovative Research Program. This work was also partly supported by Shenzhen Key Laboratory of Media Security, Shenzhen University, Shenzhen 518060, China." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "11002" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.14 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[1] https : / / eval . ai / web / challenges / challenge-page/1832/overview." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.176, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[2] https : / / eval . ai / web / challenges / challenge-page/1832/leaderboard/4298." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.472, + 0.252 + ], + "angle": 0, + "content": "[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.471, + 0.291 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.295, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.487 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.471, + 0.551 + ], + "angle": 0, + "content": "[8] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.554, + 0.471, + 0.604 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.471, + 0.655 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.471, + 0.719 + ], + "angle": 0, + "content": "[11] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.723, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[13] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.839 + ], + "angle": 0, + "content": "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.841, + 0.471, + 0.866 + ], + "angle": 0, + "content": "[15] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.128, + 0.893, + 0.164 + ], + "angle": 0, + "content": "Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.894, + 0.23 + ], + "angle": 0, + "content": "[16] Ruizhe Cheng, Bichen Wu, Peizhao Zhang, Peter Vajda, and Joseph E Gonzalez. Data-efficient language-supervised zero-shot learning with self-distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3119-3124, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.893, + 0.281 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.284, + 0.894, + 0.333 + ], + "angle": 0, + "content": "[18] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11162-11173, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.388, + 0.893, + 0.437 + ], + "angle": 0, + "content": "[20] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.893, + 0.489 + ], + "angle": 0, + "content": "[21] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.893, + 0.566 + ], + "angle": 0, + "content": "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.893, + 0.618 + ], + "angle": 0, + "content": "[23] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.621, + 0.893, + 0.67 + ], + "angle": 0, + "content": "[24] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.893, + 0.748 + ], + "angle": 0, + "content": "[25] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.75, + 0.893, + 0.787 + ], + "angle": 0, + "content": "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.893, + 0.827 + ], + "angle": 0, + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.893, + 0.866 + ], + "angle": 0, + "content": "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.894, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "11003" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.165 + ], + "angle": 0, + "content": "[29] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[30] Xinyu Huang, Youcai Zhang, Ying Cheng, Weiwei Tian, Ruiwei Zhao, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Xiaobo Zhang. Idea: Increasing text diversity via online multi-label recognition for vision-language pre-training. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4573-4583." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.245, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[31] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[32] Kyungyul Kim, ByeongMoon Ji, Doyoung Yoon, and Sangheum Hwang. Self-knowledge distillation: A simple way for better generalization. arXiv preprint arXiv:2006.12000, 3, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.374, + 0.472, + 0.436 + ], + "angle": 0, + "content": "[33] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11336-11344, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.472, + 0.49 + ], + "angle": 0, + "content": "[34] Juncheng Li, Xin He, Longhui Wei, Long Qian, Linchao Zhu, Lingxi Xie, Yueting Zhuang, Qi Tian, and Siliang Tang. Fine-grained semantically aligned vision-language pre-training. arXiv preprint arXiv:2208.02515, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.472, + 0.554 + ], + "angle": 0, + "content": "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.472, + 0.606 + ], + "angle": 0, + "content": "[36] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.472, + 0.658 + ], + "angle": 0, + "content": "[37] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.711 + ], + "angle": 0, + "content": "[38] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.712, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[39] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.472, + 0.84 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.841, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[41] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Gener-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.472, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.128, + 0.893, + 0.154 + ], + "angle": 0, + "content": "ative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.155, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.283 + ], + "angle": 0, + "content": "[43] Haoyu Lu, Nanyi Fei, Yuqi Huo, Yizhao Gao, Zhiwu Lu, and Ji-Rong Wen. Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15692-15701, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.284, + 0.893, + 0.334 + ], + "angle": 0, + "content": "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[45] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10437-10446, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.401, + 0.893, + 0.463 + ], + "angle": 0, + "content": "[46] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.465, + 0.894, + 0.504 + ], + "angle": 0, + "content": "[47] Christian M Meyer and Iryna Gurevych. Wiktionary: A new rival for expert-built lexicons? Exploring the possibilities of collaborative lexicography. na, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.893, + 0.581 + ], + "angle": 0, + "content": "[48] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.582, + 0.893, + 0.62 + ], + "angle": 0, + "content": "[49] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. arXiv preprint arXiv:2112.12750, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[50] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.893, + 0.749 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.751, + 0.893, + 0.826 + ], + "angle": 0, + "content": "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.828, + 0.893, + 0.867 + ], + "angle": 0, + "content": "[53] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Anna Rohrbach, Zhe Gan, Lijuan Wang, Lu Yuan, et al. K-lite: Learning transfer" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.894, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "11004" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.127, + 0.47, + 0.152 + ], + "angle": 0, + "content": "able visual models with external knowledge. arXiv preprint arXiv:2204.09222, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.155, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[54] Mustafa Shukor, Guillaume Couairon, and Matthieu Cord. Efficient vision-language pretraining with visual concepts and hierarchical alignment. arXiv preprint arXiv:2208.13628, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[55] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[56] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.3, + 0.471, + 0.351 + ], + "angle": 0, + "content": "[57] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.471, + 0.404 + ], + "angle": 0, + "content": "[58] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.471, + 0.469 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.472, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[60] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.471, + 0.587 + ], + "angle": 0, + "content": "[61] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.628 + ], + "angle": 0, + "content": "[62] Longhui Wei, Lingxi Xie, Wengang Zhou, Houqiang Li, and Qi Tian. Mvp: Multimodality-guided visual pre-training. arXiv preprint arXiv:2203.05175, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.471, + 0.68 + ], + "angle": 0, + "content": "[63] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.683, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[64] Haiyang Xu, Ming Yan, Chenliang Li, Bin Bi, Songfang Huang, Wenming Xiao, and Fei Huang. E2e-vlp: End-to-end vision-language pre-training enhanced by visual learning. arXiv preprint arXiv:2106.01804, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.736, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[65] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.839 + ], + "angle": 0, + "content": "[66] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.841, + 0.471, + 0.867 + ], + "angle": 0, + "content": "[67] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.472, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.128, + 0.893, + 0.165 + ], + "angle": 0, + "content": "similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.894, + 0.229 + ], + "angle": 0, + "content": "[68] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.231, + 0.894, + 0.281 + ], + "angle": 0, + "content": "[69] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.894, + 0.32 + ], + "angle": 0, + "content": "[70] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.893, + 0.372 + ], + "angle": 0, + "content": "[71] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.894, + 0.436 + ], + "angle": 0, + "content": "[72] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.894, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "11005" + } + ] +] \ No newline at end of file diff --git a/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_origin.pdf b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4d94cd3a60501d20147a971b5e87471d515c2f49 --- /dev/null +++ b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/903de9b2-1bcc-4b06-bff8-9b57aa0823cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29295dcc59ecc6cd2431972851700d431e13b75c30bcbcbedf175182a1edfc21 +size 1174457 diff --git a/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/full.md b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ffb2f68dc229bdee7414e986f761f77cfeda71b --- /dev/null +++ b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/full.md @@ -0,0 +1,355 @@ +# MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image Pretraining + +Xiaoyi Dong $^{1*}$ , Jianmin Bao $^{2*}$ , Yinglin Zhang $^{3}$ , Ting Zhang $^{2}$ , Dongdong Chen $^{4,\dagger}$ , Hao Yang $^{2}$ , Ming Zeng $^{3}$ , Weiming Zhang $^{1}$ , Lu Yuan $^{4}$ , Dong Chen $^{2}$ , Fang Wen $^{2}$ , Nenghai Yu $^{1}$ $^{1}$ University of Science and Technology of China $^{2}$ Microsoft Research Asia + $^{3}$ Xiamen University $^{4}$ Microsoft Cloud + AI + +{dlight@mail., zhangwm@, ynh@}.ustc.edu.cn cddlyf@gmail.com + +{jianbao, ting.zhang, luyuan, doch, fangwen}@microsoft.com + +{zhengyinglin@stu，zengming@}xmu.edu.cn yanghao.alexis@foxmail.com + +# Abstract + +This paper presents a simple yet effective framework MaskCLIP, which incorporates a newly proposed masked self-distillation into contrastive language-image pretraining. The core idea of masked self-distillation is to distill representation from a full image to the representation predicted from a masked image. Such incorporation enjoys two vital benefits. First, masked self-distillation targets local patch representation learning, which is complementary to vision-language contrastive focusing on text-related representation. Second, masked self-distillation is also consistent with vision-language contrastive from the perspective of training objective as both utilize the visual encoder for feature aligning, and thus is able to learn local semantics getting indirect supervision from the language. We provide specially designed experiments with a comprehensive analysis to validate the two benefits. Symmetrically, we also introduce the local semantic supervision into the text branch, which further improves the pretraining performance. With extensive experiments, we show that MaskCLIP, when applied to various challenging downstream tasks, achieves superior results in linear probing, finetuning, and zero-shot performance with the guidance of the language encoder. Code will be release at https://github.com/LightDXY/MaskCLIP. + +# 1. Introduction + +Vision-language (VL) contrastive learning [31, 51] has shown remarkable success in pretraining for various tasks. With large-scale image-text pairs available on the Internet, the model composed of a simple dual encoder design learns + +strong semantic prior by aligning between image and text. The resulting visual encoder not only exhibits excellent linear probing and finetuning performance, but also enables impressive zero-shot performance with the guidance of the language encoder, showing the generality of natural language and its ability to supervise a wide range of visual concepts. + +Nonetheless, the associated language description, though providing richer information than mere class labels, still can hardly describe all the information in the corresponding image, as images are continuous signals with fine-grained details and complex semantics. As a result, the VL contrastive by aligning global representations may only focus on the text-described objects and ignore the rest which might be useful for downstream tasks. + +In this paper, we are interested in how to fully leverage the image itself to facilitate the VL contrastive to further improve the transfer capability. (1) Firstly, the learned feature representation shall characterize local patches, serving as a complementary for global representation in VL contrastive. Inspired by the recent success of masked image modeling [4, 19, 26, 51, 60, 61] in learning patch representations, we also randomly mask the input image with a large portion to force the visual encoder to focus on the remaining visible patches. (2) Secondly, the learned representation for local patches shall possess semantic meanings, being consistent with the global representation receiving semantic text supervision. We bring mean teacher self-distillation [25, 57] to supervise the learned patch representations with the visual feature representations, enabling implicit supervision from natural language. The resulting objective is denoted as masked self-distillation where the student model and the teacher model come from the same neural networks and the knowledge is distilled from the full image (fed to the teacher model) to the masked image (fed to student model). To this end, we introduce MaskCLIP by incorporating masked self + +distillation into VL contrastive to advance the transferable visual encoder. + +There are several recent attempts [49, 68] also exploring the capability of the visual encoder under natural language supervision. The common approach is to introduce contrastive learning or masked image modeling on the vision side together with contrastive language-image pretraining. However, the performance indeed improves based on CLIP but does not as well as our masked self-distillation. We argue that (1) the contrastive learning objective based on central crop augmentation actually learns global representations for salient objects while lack of attention on the surrounding backgrounds [11]; and (2) masked image modeling usually needs to remap the learned representation to pixels [26] or discrete tokens [4]. Such low-level prediction target is inefficient for semantic feature learning and thus also conflicts with high-level language supervision in VL contrastive. A brief illustration is presented in Figure 1. In the experiments, we conduct comprehensive ablations to analyze the difference and provide numerical and visual evidence for better understanding. + +Symmetrically, we argue that local semantic supervision on the text branch is also helpful for the text encoder and eventually beneficial for zero-shot performance. So we introduce the same mask-data-modeling format supervision into the text branch as well. Different from images where the pixel is low-level signal, the words crafted by human beings are already highly semantic, so we use the tokenized word piece as the prediction target directly, following the well-studied mask language modeling method BERT. Meanwhile, to reduce the output conflicts between contrastive learning and mask language modeling, we introduce a small decoder for the mask language modeling branch. + +We train our MaskCLIP on a subset of a publicly available image-text pairs dataset, YFCC [58], and thoroughly evaluate the transfer ability of visual representations on several vision benchmarks: ImageNet-1K [17] for classification, ADE20K [69] for semantic segmentation, MS-COCO [40] for detection and segmentation, as well as a batch of other classification benchmarks. When it comes to ImageNet-1K [17] classification, MaskCLIP achieves $+6.9\%$ , $+7.2\%$ , $+1.3\%$ higher than CLIP for zero-shot transfer, linear probing, and finetuning respectively. For vision downstream tasks, we reach $+2.7$ mIoU on ADE20K [69] and $+1.8$ APb, $+1.4$ APm on MS-COCO [40]. For vision-language tasks, MaskCLIP achieves $+6.1\%$ average zero-shot accuracy on 20 datasets, and $+17.2\%$ , $+12.8\%$ rank@1 improvement on the Flickr30K [67] image-test retrieval. In the recent Image Classification in the Wild challenge academic track, our MaskCLIP gets the $1_{st}$ result with $48.9\%$ TOP-1 average accuracy, surpassing the second team with $3.4\%$ . + +In summary, the major contributions of this work are: + +1. We present a novel vision-language pretraining + +framework MaskCLIP, by introducing masked self-distillation objective to facilitate VL contrastive for better transferable visual models. + +2. We present extensive ablation studies on MaskCLIP variants and provide in-depth analysis numerically and visually to help understand how the proposed masked self-distillation assists VL contrastive. +3. We demonstrate our MaskCLIP on tens of benchmarks, showing the superiority under all three settings: zero-shot, linear probing, and finetuning. + +# 2. Related Work + +Vision-language pretraining. Recent years have seen rapid progress made in vision-language pretraining [15, 18, 33, 35-39, 44-46, 50, 55, 56, 72]. Several multiple cross-modality loss functions have been proposed for the training objective, such as image-text matching [15, 37, 44, 56, 64], masked language modeling [15, 37, 44, 55, 56], masked image modeling [15, 44, 55, 56], contrastive loss [35, 38, 39]. These objects are often mixed with each other to form a compound objective. While a variety of approaches have been proposed, few works investigate the performance on visual representation learning for image classification. Recently, CLIP [51] and ALIGN [31] show that the image-text contrastive learning objective achieves promising performance for visual representation learning. There are many following works proposed to further improve the pretraining performance, DeCLIP [70], SLIP [49], COTS [43], ViCHA [54], CYCLIP [24] use additional uni/multi-modality supervision to improve the model capability, and PyramidCLIP [23], KLITE [53], IDEA [30] seek to external knowledge from pre-trained models or datasets as the additional guidance. FILIP [66] and LOUPE [34] introduce fine-grained alignment to the model. Focusing on this research direction, we analyze the desired properties of supervision which could be complementary to CLIP, and propose the masked self-distillation objective incorporated with the image-text contrastive loss to further improve pretraining performance for various visual understanding tasks. + +Self-supervised learning Self-supervised visual representation learning has attracted increasing attention over the past few years. The objective of the self-supervised learning is mainly divided into two categories: contrastive and generative [41]. The contrastive methods, such as MOCO [12, 27], SimCLR [9, 10], BYOL [25], SimSiam [13], and DINO [6] measure the similar and dissimilar samples by contrastive loss. Their success heavily depends on the strong data augmentation. The generative methods, such as BEiT [4], MAE [26], PeCo [19], BEVT [60], BootMAE [20] and MaskFeat [61] leverage masked image modeling to reconstruct the remaining masked part of its original input from the given visible parts. The generative methods show more + +![](images/572575a2acca74311944aa1ecbe913af6b324e8b33d2284f1a8940f199fd4db6.jpg) +(a) CLIP + +![](images/ccea09b0b77ceeb4b93c215b19706025b070148e41d5239e61a643d403af67e3.jpg) +(b) + Contrastive Learning + +![](images/aff0d5e3397fa30243d23918cc5d12debb9822f4031e7da9ce75fa8982c3c5de.jpg) +(c) + Mask Image Modeling + +![](images/8c1639d6fc1293107cec6749c9c6f287453f4383955d519a03bc0bc447e41862.jpg) +(d) MaskCLIP +Figure 1. Pipeline comparison between combination CLIP with different vision self-supervised learning methods. (a) Vanilla CLIP. (b) CLIP + contrastive learning. (c) CLIP + pixel prediction mask image modeling. (d) CLIP + mask self-distillation, i.e. MaskCLIP. The $E_{T}$ , $E_{I}$ is the text encoder and image encoder respectively, and all the $E_{I}$ , $E_{T}$ within each pipeline share the weight. $\bar{E}_{I}$ is the mean-teacher model, whose weight is updated by the exponential moving average of $E_{I}$ and does not require gradient. + +promising transfer performance than the contrastive methods, as generative objective learns patch representations while contrastive objective focuses on learning centric global representations [11]. + +Self-knowledge distillation Self-knowledge distillation [32] aims to distill the knowledge in a model itself and uses it for training the model. Instead of distilling knowledge from a pretrained teacher model [29], self-knowledge distillation regards a temporal ensemble of the student model as the teacher. It means that a student model becomes a teacher model itself, which gradually utilizes its own knowledge for softening the hard targets to be more informative during training. Self-knowledge distillation has been explored in semi-supervised learning [57], contrastive learning [16, 35], self-supervised learning [3, 7]. In this paper, we use visual features supervised by natural language for guidance in masked self-distillation which naturally fit VL contrastive to learn more transferable visual representations. + +# 3. MaskCLIP + +We introduce MaskCLIP, a novel framework that learns visual representations. The core part of MaskCLIP is its backbone image encoder, denoted by $E_{I}$ as shown in Figure 1. It obtains the transferable capability during pretraining that could benefit downstream vision tasks. Following recent self-supervised approaches [4, 14, 26, 49], we implement the backbone $E_{I}$ as a Vision Transformer (ViT) [22]. The prediction results from $E_{I}$ given an input image $I$ then should be a collection of visual feature tokens, represented as + +$$ +E _ {I} (I) = \left\{f _ {\text {c l s}}, f _ {1}, f _ {2}, \dots , f _ {N} \right\}. \tag {1} +$$ + +Here $cls$ is short for class token. $1,\ldots ,N$ are the indexes of the non-class tokens. + +The rest of this section starts with the utilization of language supervision. More shall be emphasized on the masked self-distillation, which we deem crucial for visual pretraining. + +# 3.1. Vision-language Contrastive + +Following [31, 51], we introduce a Transformer-based text encoder $E_{T}$ to leverage language knowledge. It aims to align the global feature representations of an image and a text with respect to some forms of similarity. Precisely, consider a given image-text pair $\{I, T\}$ , besides extracting the visual feature representation $E_{I}(I)$ using the vision backbone as shown by Equation 1, we additionally use the text encoder $E_{T}$ to extract linguistic features from the text $T$ . + +The mean feature of the two branches are regarded as the global representations and are fed into a projection head (implemented as a fully-connected layer) respectively to obtain the metric embeddings $e^T$ and $e^I$ . Image-text contrastive loss is employed to align them during pretraining. The loss can be formulated as $\mathcal{L}_T + \mathcal{L}_I$ , with + +$$ +\mathcal {L} _ {I} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp (e _ {i} ^ {I} e _ {i} ^ {T} / \sigma)}{\sum_ {j = 1} ^ {B} \exp (e _ {i} ^ {I} e _ {j} ^ {T} / \sigma)} +$$ + +$$ +\mathcal {L} _ {T} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp \left(e _ {i} ^ {T} e _ {i} ^ {I} / \sigma\right)}{\sum_ {j = 1} ^ {B} \exp \left(e _ {i} ^ {T} e _ {j} ^ {I} / \sigma\right)}, \tag {2} +$$ + +where $B$ stands for the number of image-text pairs within a training mini-batch, $i,j$ are indexes within the batch; $\sigma$ stands for the temperature for the loss functions, which is learned together with all other parameters during training. + +# 3.2. Masked Self-distillation for Visual Encoder + +Knowledge distillation is a learning paradigm where a student model is trained to match the output of a given teacher model, so that the student model can be improved by the teacher. Instead of bringing in an external teacher, self-distillation methods such as [7, 25, 57] proposes using a mean teacher model that is derived from the student itself. In specific, the teacher shares the same structure with the student, while the parameters of the teacher are exponential moving averages (EMA) of the parameters from the student. In the following, we would use the term "EMA model" to represent such mean teacher model constructed from the student. + +MaskCLIP leverages the mean teacher self-distillation to enhance its vision representations. Let $\bar{E}_I$ be the EMA model of the backbone encoder $E_I$ . $\theta_t$ and $\bar{\theta}_t$ are the parameters of $E_I$ and $\bar{E}_I$ at training step $t$ . $\bar{\theta}_t$ is updated with + +$$ +\bar {\theta} _ {t} = \alpha \bar {\theta} _ {t - 1} + (1 - \alpha) \theta_ {t}, \tag {3} +$$ + +where $\alpha$ is a hyper-parameter for smoothing updates. We propose to incorporate masked image modeling into self-distillation, resulting in masked self-distillation with asymmetric input for student model and teacher model. + +In specific, considering a given input image $I$ , we first feed it to the EMA model $\bar{E}_I$ (teacher model) to obtain the distillation targets. These target features can be represented as + +$$ +\bar {E} _ {I} (I) = \{\bar {f} _ {c l s}, \bar {f} _ {1}, \bar {f} _ {2}, \dots , \bar {f} _ {N} \}. \tag {4} +$$ + +In the meantime, we randomly mask a large portion of the input image patches and then feed it into the original backbone $E_I$ (student model). Following [26], we only feed the visible (unmasked) patches, denoted by $I'$ , into the original backbone $E_I$ to speed up computation and save memory. Let $\mathcal{M}$ be the indexes of all the masked tokens. These encoded features corresponding to visible tokens can then be denoted as $E_I(I') = \{f_{cls}'\} \cup \{f_{k \notin \mathcal{M}}'\}$ . They are then joined with a shared and learnable feature vector, denoted as $m$ , that represents mask tokens, to form a complete set of features $\{f_{cls}', f_1', f_2', \ldots, f_N'\}$ , with $f_{i \in \mathcal{M}}' = m$ . We attach positional embeddings onto all these tokens, and append a small Transformer $D$ as a decoder to predict features of the masked region from the visible tokens, which could be formulated as + +$$ +\begin{array}{l} \left(D \circ E _ {I}\right) \left(I ^ {\prime}\right) = D \left(\left\{f _ {c l s} ^ {\prime}, f _ {1} ^ {\prime}, f _ {2} ^ {\prime}, \dots , f _ {N} ^ {\prime} \right\}\right) \\ = \left\{f _ {c l s} ^ {\prime \prime}, f _ {1} ^ {\prime \prime}, f _ {2} ^ {\prime \prime}, \dots , f _ {N} ^ {\prime \prime} \right\}. \tag {5} \\ \end{array} +$$ + +Inspired by [71], we use an online quantizer $h(\cdot)$ to transform the output features into a soft codewords distribution, and minimize the cross-entropy between the target features and the predicted features. Formally, + +$$ +\mathcal {L} _ {\text {D i s t}} = \frac {1}{| \mathcal {M} |} \sum_ {k \in \mathcal {M}} - \bar {h} \left(\bar {f} _ {k}\right) ^ {T} \log h \left(f _ {k} ^ {\prime \prime}\right). \tag {6} +$$ + +here the parameter of the teacher quantizer $\bar{h}()$ is also EMA updated by the online quantizer, similar to the teacher model. + +# 3.3. Local Semantic Learning for Text Encoder + +Besides the local semantic supervision for the visual encoder, we argue it is also helpful for the text encoder. So we introduce the BERT pretraining into the text branch. For the text $T = \{t_{\text{sos}}, t_1, t_2, \dots, t_M, t_{\text{eos}}\}$ , we denote the masked input as $T' = \{t_{\text{sos}}', t_1', t_2', \dots, t_M', t_{\text{eos}}'\}$ , where $t_{i \in \mathcal{M}_T}' = m_t$ and $t_{i \notin \mathcal{M}_T}' = t_i$ , and $\mathcal{M}_T$ be the indexes of all the masked text tokens. The output feature of the encoder is $E_T(T')$ . + +To reduce the output conflict between the global image-text contrastive learning and the local mask language modeling, we further introduce a small text decoder, which shares the same architecture as the encoder but with only a few layers. So that the global prediction and local prediction are conducted at different layers. We denote the output feature as: $(D_T\circ E_T)(T') = \{t_{sos}''$ , $t_1''$ , $t_2''$ , ..., $t_M''$ , $t_{eos}''\}$ and the loss could be formulated as: + +$$ +\mathcal {L} _ {M L M} = \frac {1}{| \mathcal {M} _ {T} |} \sum_ {k \in \mathcal {M} _ {T}} - t _ {k} ^ {T} \log t _ {k} ^ {\prime \prime}. \tag {7} +$$ + +# 3.4. Overall Loss Functions + +Finally, we pretrain MaskCLIP with all these losses combined: + +$$ +\mathcal {L} _ {I} + \mathcal {L} _ {T} + \lambda \mathcal {L} _ {\text {D i s t}} + \beta \mathcal {L} _ {\text {M L M}}, \tag {8} +$$ + +with $\lambda, \beta$ being the hyper-parameter weighting between VL contrastive loss and self-supervised learning loss. All the components of MaskCLIP are trained from scratch, including the visual backbone $E_I$ , the visual decoder $D$ , the text encoder $E_T$ , as well as the text decoder $D_T$ . + +# 4. Experiments + +# 4.1. Setup + +Model architecture. Our framework consists of the visual encoder $E_{I}$ , the text encoder $E_{T}$ , the visual decoder $D$ , and the text decoder $D_{T}$ . We adopt the widely used Transformer ViT-B/16 [22] for a fair comparison. It is composed of 12 layers, 768 width, and 12 head. The input image is $224 \times 224$ resolution and is further split into $14 \times 14$ patches with size $16 \times 16$ . A learnablecls token is pretended to the 196 embeddings. For the text encoder, we adopt a 12-layer, 512-width, and 8-head Transformer following CLIP [51], and the text decoder has 4 layers. The number of text tokens is fixed to 77 with necessary truncations or paddings. For the image decoder, we directly use a one-layer Vision Transformer. + +Pretraining details. We train our proposed MaskCLIP from scratch for 25 epochs, the batch size is fixed to 4096 for all the experiments. The masks used in the mask self-distillation + +![](images/14487a7fcb5cc6382e984dc0b94f915e3aa1a7bab539289409b1dceb7dacd090.jpg) +Figure 2. Visualization of the similarity between text and image features. The images and captions are from the MS-COCO val set. Here we show the image feature similarity with both full caption and different objects in it. The caption is "Three teddy bears sit in a sled in snow". More results could be found in the supplemental materials. + +
TrainingIN-1KFlicker30K
MemoryTime0-shotLinearFinetuneI2TT2I
CLIP14G1.00×37.666.582.352.932.8
CLIP+SimCLR30G2.67×42.872.182.658.641.3
CLIP+MAE16G1.30×42.168.583.257.341.1
MaskCLIP19G1.75×44.573.783.670.145.6
+ +Table 1. Results of boosting CLIP with different kinds of vision self-supervised learning methods. + +branch and mask language modeling branch are random mask with a mask ratio of $75\%$ and $20\%$ . We pretrain all the models with the commonly used YFCC15M dataset, which is flited from the YFCC100M [58] dataset by [51]. + +Downstream details. We evaluated MaskCLIP on several downstream datasets, including ImageNet-1K [17], ADE20K [69], MS-COCO [40], Flicr30K [67] et al. For ImageNet-1K, we report zero-shot, linear probing, and finetuning performance. The zero-shot is conducted following the label prompt setting in SLIP [49]. For linear probing, we fix the backbone and train a new linear classifier for 90 epochs. For finetuning, we follow the setting in BEiT [4] and finetune the model for 100 epochs with a layer-decayed learning rate. See supplemental materials for more details. + +# 4.2. Analysis + +We first present our analysis by studying different ways of boosting CLIP. The baseline is CLIP [51] trained on the + +YFCC-15M. Besides the introduced masked self-distillation, we consider two other popular methods: (1) SimCLR [9], a representative contrastive method; and (2) MAE [26] the state-of-the-art masked image modeling approaches. All the compared methods are trained on the YFCC-15M for a fair comparison. We have the following observations. + +Vision self-supervision helps VL contrastive. We evaluate the models on both vision task ImageNet-1K [17] classification and vision-language task image-text retrieval on Flicker30K [67] and present the comparison in Table 1. All the added vision self-supervision, regardless of contrastive or generative, improves the baseline CLIP. Among them, our proposed MaskCLIP achieves the best results in terms of all the evaluation metrics, outperforming CLIP with $+6.9\%$ , $+7.2\%$ , $+1.3\%$ on ImageNet-1K classification for zero-shot, linear probing, and finetuning respectively, and $+17.2\%$ , $+12.8\%$ on Flicker30K for image-to-text retrieval and text-to-image retrieval. We also report the training GPU memory usage and time-consuming cost in Table 1. It is worth noting that the contrastive model (CLIP + SimCLR) compares two additional views of the input image, resulting in larger GPU memory usage and longer training time. + +Masked image modeling is able to learn representations for local patches. We argue that the image encoder only pays attention to the text-described objects under VL contrastive due to sparse text description and to the centric objects under image contrastive due to central-crop augmenta + +
MethodObjectiveADE20K mIoUPascal mIoU
CLIPGlobal7.213.5
CLIP + SimCLRGlobal + Global6.311.9
CLIP + MAEGlobal + Pixel-wise Local8.316.4
MaskCLIPGlobal + Token-wise Local10.217.2
+ +tion. In contrast, masked image modeling forces the image encoder to focus on local patches using token-wise objectives by mandatorily masking a large portion of patches. Here, we provide numerical comparisons for evidence. We conduct an "Annotation-free zero-shot segmentation" experiment to test the zero-shot segmentation. The results on such a dense prediction task would better reveal the ability of local patch representations than global classification. Following the design in DenseCLIP [70], we use the prompted label feature as the linear classification weight to realize segmentation, without any training procedure. We evaluate the performance on two widely used datasets: ADE20K [69] and Pascal Context [48]. The results are shown in Table.2. We can see that equipped with masked image modeling, our MaskCLIP as well as CLIP+MAE achieves better results than CLIP and CLIP+SimCLR, validating our hypothesis. + +Masked self-distillation learns semantic representations for local patches. Our masked self-distillation predicts visual features dynamically outputted by the visual encoder and thus implicitly gets supervision from the text side via VL contrastive. While MAE predicts fixed low-level pixels, making it inefficient to learn semantic representations (as the objective may force the representation to memorize low-level details) and thus causing conflict with VL contrastive. To show this, we select images from MS-COCO [40] and calculate the feature similarity between image features and their corresponding caption features. We also select objects in the caption, prompt it to a new caption, such as "a photo of teddy bears", and calculate the similarities. An example is shown in Figure 2 (More can be found in the supplementary material). Comparing MaskCLIP with CLIP+MAE in the fourth column, we can see that CLIP+MAE uses color as evidence and fails to distinguish the white teddy bear from the white snow. While our MaskCLIP successfully differentiates the two objects, suggesting ours learn more semantic features. On the other hand, the superior results of MaskCLIP shown in Table 1 and Table 2 also validate this. It is worth mentioning that CLIP and CLIP+SimCLR fail to have a correct response partition for different single objects like MaskCLIP, further justifying our second observation. + +# 4.3. Comparison with Previous Methods + +To show the effectiveness of MaskCLIP as a general vision-language pretrain method, we conduct experiments on both vision tasks and vision-language tasks. For vision + +Table 2. Annotation-free zero-shot segmentation results on ADE20K and Pascal Context. + +
MethodEpochIN-1KADE20KMS-COCO
0-ShotLinFTmIoUAPbAPm
DeiT [59]300*--81.847.444.139.8
SimCLR [9]25-64.082.548.044.640.2
MAE [26]25-56.282.546.543.239.1
CLIP [51]2537.666.582.347.843.639.5
SLIP [49]2542.872.182.648.544.040.3
MaskCLIP2544.573.783.650.545.440.9
+ +Table 3. Comparison with previous methods, including supervised baselines, self-supervised pretraining methods, and vision-language pretraining methods. * is the epoch of the supervised baseline on ImageNet-1K. + +tasks, we report results on ImageNet-1K [17] classification, MS-COCO [40] object detection, and ADE20K [69] semantic segmentation. For vision-language tasks, we report zero-shot results on recent challenging ICinW 20 datasets benchmark and image-text retrieval results on Flickr30K [67] and MS-COCO [40]. In the following, we compare with the supervised baseline DeiT [59], self-supervised methods SimCLR [9] and MAE [26], and vision-language methods CLIP [51] and SLIP [49]. For a fair comparison, we train SimCLR and MAE on YFCC-15M [58] with the same epochs. + +Classification on ImageNet-1K. As shown in Table 3, MaskCLIP benefits from the advantages of both VL pretraining and image mask self-distillation that shows strong performance on all the metrics. For zero-shot tasks, MaskCLIP outperforms CLIP by $+6.9\%$ with 25 epoch training and achieves $+1.7\%$ higher than the recent work SLIP. When it comes to finetune, MaskCLIP reaches $83.6\%$ top-1 accuracy, and outperforms CLIP by $+1.3\%$ . + +Semantic segmentation on ADE20K. Then we apply our MaskCLIP to the semantic segmentation task. Here we use the UperNet [63] framework with $512 \times 512$ input and end-to-end training for 160K iterations. The evaluation metric is the mean Intersection of Union (mIoU) and we report single-scale evaluation results here. The results are given in Table 3. Our method achieves $50.5\mathrm{mIoU}$ , $+2.7\mathrm{mIoU}$ than our baseline method CLIP, and $+2.0\mathrm{mIoU}$ than SLIP. This verifies the effectiveness of our introduced incorporation. + +Object detection and instance segmentation on MS-COCO. We further investigate our transfer performance on object detection and instance segmentation in Table.3. Here we use Mask-RCNN [28] framework with single-scale input and $1 \times$ schedule (12 epochs). Our method achieves 45.4 box AP and 40.9 mask AP, $+1.8 / 1.4$ better than CLIP, and $+1.4 / 0.6$ better than SLIP. + +Zero-shot on small datasets. We also report zero-shot performance on 20 small datasets under the ICinW setting (see the introduction below) in Table 4. We find that all the methods perform poorly on some datasets such as Aircraft(1% acc for random guessing, we omit the description in the + +
AverageCaltech-101CIFAR-10CIFAR-100Country211DTDEuroSATFER-2013AircraftFood-101GTSRBMemesKittiDisMNISTFlowersPetsPatchCamSST2RESISC45CarsVoc2007
Pretraining on YFCC-15M
CLIP34.058.668.536.910.821.430.516.95.151.66.551.125.95.052.728.651.752.522.44.579.1
SLIP37.870.982.648.611.826.619.818.15.659.912.651.829.49.856.331.455.351.528.55.480.5
MaskCLIP40.172.080.257.512.627.944.020.36.164.98.552.034.34.957.034.350.149.935.76.782.1
Pretraining on ICinW Academic Track Siting: YFCC-15M, GCC3M+12M, ImageNet-21K(ImageNet-1K is removed)
1st MaskCLIP48.986.495.378.311.633.057.718.88.078.917.352.816.07.374.274.452.146.254.326.582.3
2nd KLITE*45.587.492.768.88.232.227.917.44.372.411.448.431.112.875.665.950.652.944.410.282.3
3rd YT-CLIP44.577.883.558.411.931.940.727.16.968.718.852.39.118.853.169.351.550.352.719.779.3
4th UniCL†44.084.890.267.86.725.435.330.83.568.311.151.017.911.371.744.952.149.541.424.281.3
5th Gramer*43.283.992.969.57.325.524.430.42.771.09.052.612.410.170.452.450.650.144.813.881.3
+ +Table 4. Zero-shot evaluation on ICinW classification benchmarks. Best results in bold. * indicates using Swin-B as the backbone, † indicates using Focal-B as the backbone. + +
Training EpochFlickr30KMS-COCO
Image-to-textText-to-imageImage-to-textText-to-image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP [51]2552.979.687.232.860.871.227.553.565.017.738.850.5
SLIP [49]2558.685.191.741.368.778.633.459.870.621.544.456.3
MaskCLIP2570.190.395.345.673.482.141.467.977.525.549.761.3
+ +Table 5. Results of zero-shot image-text retrieval on Flickr30K and MS-COCO datasets. Best results in bold. + +following), $\mathrm{Fer}(24.7\%)$ , Country211 $(0.5\%)$ , GTSRB $(5.9\%)$ , Cars $(0.8\%)$ . This might be caused by the data domain gap that the YFCC-15M contains few related images and descriptions. For the rest of the datasets, all the methods get reasonable performance and our MaskCLIP gets the best performance on most datasets. + +Image Classification in the Wild (ICinW) Challenge The ICinW challenge [1] is a newly proposed visual pretraining benchmark, which contains 20 diverse downstream classification datasets, measuring the ability of pre-training models on both the prediction accuracy and their transfer efficiency in a new task. The pretraining is limited to three datasets: YFCC-15M [58], GCC3M [52] $+12\mathrm{M}$ [8] and ImageNet-21K [17] (ImageNet-1K data is excluded). We pretrain our MaskCLIP on it and get the $\mathbf{1}_{st}$ result in the zero-shot track [2] (we submit the results anonymously). As shown in Table 4, the $2_{nd}$ team KLITE uses a strong Swin-B [42] as the backbone and additional knowledge from GPT-3 [5] and Wiktionary [47], and the $4_{th}$ use the strong Focal-B [65] as the backbone, while our MaskCLIP greatly outperforms these methods with a simple ViT-B backbone and no additional knowledge. + +Zero-shot on text-image retrieval. We further report the zero-shot text-image retrieval results on two benchmark datasets, Flickr30K [67] and MS-COCO [40]. We find that the text without any prefixes or suffixes works well for all the models. Table 5 shows the results. We can see that MaskCLIP exhibits a strong zero-shot performance. For example, with 25 epochs training, MaskCLIP reaches $41.4\%$ Rank@1 image-to-text accuracy on MS-COCO, outperform- + +ing CLIP with $13.9\%$ and $25.5\%$ Rank@1 text-to-image accuracy, $+7.8\%$ higher than CLIP. + +# 4.4. Ablations + +We compare our default settings with other alternatives to justify the efficacy of our model designs. + +Training objectives ablation. As shown in Table.6a, when we remove the mask language modeling loss $\mathcal{L}_{\mathrm{MLM}}$ , the performance of the image-text task drops, including the zero-shot accuracy and retrial performance. While benefiting from the distillation loss, the finetuning performance on ImageNet-1K is not influenced. When we remove the distillation loss $\mathcal{L}_{\mathrm{Dis}}$ , we observe a performance drop on all tasks, especially the finetuning results. + +Distillation loss format. Different from previous methods [3,21,26] that calculate the per-element distance as the loss function, we use an online tokenizer to map the feature to soft codewords and use the cross-entropy loss as the supervision. Here we study their difference in Table.6b. We find that although they get similar fine-tuning performance, the CE loss gets better zero-shot and linear probing performance. The reason may be that the per-element MSE loss leads the model to fit some unnecessary details of the target feature, while the CE loss with soft tokenizer helps the model to focus more on the important feature. + +Distillation & MLM loss weight. Here we set the loss weight of the CLIP branch as 1 and study the loss weight of the two additional branches. As shown in Table.6e and Table.6f, setting $\lambda = 1$ or $\beta = 1$ emphasize too much on new tasks, which mislead the model to a wrong converge + +
Model0-ShotFTI2T/T2I
MaskCLIP44.583.670.1/45.6
w/o LMLM42.883.665.0/41.6
w/o LDis42.082.465.4/40.5
+ +(a) Training Objectives ablation. Both is necessary for MaskCLIP. + +
Depth0-ShotI2T/T2I
043.565.2/44.1
144.370.4/45.3
244.370.2/45.4
444.570.1/45.6
844.267.5/44.7
+ +(d) Text decoder depth. The decoder is necessary and a shallow one works better. + +
Loss0-ShotLinFT
MSE43.873.283.6
CE44.573.783.6
+ +(b) Distillation loss format. The online tokenizer with cross-entropy loss works slightly better than MSE loss. + +
Weight0-ShotLinFT
138.568.282.5
0.144.473.583.5
0.0544.573.783.6
0.0143.673.083.4
+ +(e) Distillation loss weight. A small loss weight works well for MaskCLIP. + +
Depth0-ShotLinFT
144.573.783.6
243.772.983.4
443.572.583.3
+ +(c) Visual decoder Depth. A shallow decoder gets better performance. + +
Weight0-ShotI2T/T2I
136.551.7/32.1
0.144.369.2/45.9
0.0544.570.1/45.6
0.0143.270.6/45.6
+ +Table 6. MaskCLIP ablation experiments with YFCC-15M dataset. We report zero-shot(0-Shot), fine-tuning (FT), and linear probing (Lin) accuracy (%) for image-encoder-related ablation. And zero shot image-to-text, text-to-image retrieval (I2T/T2I) for text encoder-related ablations. Default settings are marked in gray . + +direction, resulting in poor performance. When we reduce the loss weight by $10 \times$ , the two additional tasks are helpful for the model and show a consistent gain on all the metrics. We suspect this is because the CLIP loss requires two different capabilities: understanding the input content and aligning them into a shared feature space. And the goal of the two additional self-supervised learning tasks is to facilitate understanding. + +Image & Text decoder depth. Then we study the influence of the decoder depth for both image and text decoders. As shown in Table.6c, we find the image decoder with only one layer works well, increasing the decoder depth leads to worse performance on all metrics. Similarly, Table.6d shows that the text branch benefits from a shallow decoder design. We argue that a too-deep decoder would make the encoder lazy, relying on the strong decoder to resolve the challenging mask feature/language modeling tasks. And the different depth choice between the image and text branches is caused by the framework difference: the image branch sees the mask tokens at the decoder, while the text branch takes the mask tokens as the encoder input. Note that if we remove the text decoder, the performance gets worse. We think this is largely caused by the output conflict that the global recognition feature aggregation and local word prediction are conducted at the same layer. + +Single-Stage v.s. two-Stage. Our MaskCLIP learns the VL contrastive and masked self-distillation simultaneously and jointly in a single stage. One possible variant is to first train CLIP and then use CLIP feature from the first stage to train masked image modeling as in [61, 62]. We report results on three datasets in Table 7. We can see that the second stage achieves better finetuning results compared with results from the stage one, showing the effectiveness of masked image modeling. Nonetheless, such two-stage training requires longer training time and loses the transfer capability in a zero + +(f) MLM loss weight. A small loss weight works better. + +
MethodEpochIN-1KFlicker30KADE20K
0-shotFTI2TT2I0-shotFT
Two-StageStage12537.682.352.932.87.247.8
Stage225—83.4———48.2
MaskCLIP2544.583.670.145.610.250.5
+ +Table 7. Comparison between two-stage method and our single-stage MaskCLIP. + +shot setting. In contrast, our MaskCLIP achieves superior results under all settings with fewer epochs. + +# 5. Conclusion + +We present MaskCLIP, a new VL pretraining framework that incorporates masked self-distillation into VL contrastive. We point out that masked self-distillation learns local semantics, fitting nicely to the VL contrastive that aims to learn global semantics, and this is supported with comprehensively designed experiments. We also utilize mask language modeling to enhance the text encoder which is critical for zero-shot performance. The resulting visual encoder shows strong transfer capability across widely adopted benchmarks for linear probing, fine-tuning, and also zero-shot evaluation. + +# Acknowledgements + +This work was supported in part by the Natural Science Foundation of China under Grant U20B2047, 62072421, 62002334, 62102386, and 62121002. Key Research and Development program of Anhui Province under Grant 2022k07020008, Ant Group through CCF-Ant Innovative Research Program CCF-AFSG RF20210025, Alibaba Group through Alibaba Innovative Research Program. This work was also partly supported by Shenzhen Key Laboratory of Media Security, Shenzhen University, Shenzhen 518060, China. + +# References + +[1] https : / / eval . ai / web / challenges / challenge-page/1832/overview. +[2] https : / / eval . ai / web / challenges / challenge-page/1832/leaderboard/4298. +[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022. +[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. +[8] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[10] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020. +[11] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. +[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. +[13] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning, 2020. +[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. +[15] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: + +Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. +[16] Ruizhe Cheng, Bichen Wu, Peizhao Zhang, Peter Vajda, and Joseph E Gonzalez. Data-efficient language-supervised zero-shot learning with self-distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3119-3124, 2021. +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[18] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11162-11173, 2021. +[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. +[20] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022. +[21] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022. +[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[23] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022. +[24] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022. +[25] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. +[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. +[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. +[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. + +[29] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. +[30] Xinyu Huang, Youcai Zhang, Ying Cheng, Weiwei Tian, Ruiwei Zhao, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Xiaobo Zhang. Idea: Increasing text diversity via online multi-label recognition for vision-language pre-training. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4573-4583. +[31] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. +[32] Kyungyul Kim, ByeongMoon Ji, Doyoung Yoon, and Sangheum Hwang. Self-knowledge distillation: A simple way for better generalization. arXiv preprint arXiv:2006.12000, 3, 2020. +[33] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11336-11344, 2020. +[34] Juncheng Li, Xin He, Longhui Wei, Long Qian, Linchao Zhu, Lingxi Xie, Yueting Zhuang, Qi Tian, and Siliang Tang. Fine-grained semantically aligned vision-language pre-training. arXiv preprint arXiv:2208.02515, 2022. +[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34, 2021. +[36] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. +[37] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. +[38] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409, 2020. +[39] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. +[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. +[41] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Gener- + +ative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 2021. +[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. +[43] Haoyu Lu, Nanyi Fei, Yuqi Huo, Yizhao Gao, Zhiwu Lu, and Ji-Rong Wen. Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15692-15701, 2022. +[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. +[45] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10437-10446, 2020. +[46] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. +[47] Christian M Meyer and Iryna Gurevych. Wiktionary: A new rival for expert-built lexicons? Exploring the possibilities of collaborative lexicography. na, 2012. +[48] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. +[49] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. arXiv preprint arXiv:2112.12750, 2021. +[50] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. +[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. +[53] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Anna Rohrbach, Zhe Gan, Lijuan Wang, Lu Yuan, et al. K-lite: Learning transfer + +able visual models with external knowledge. arXiv preprint arXiv:2204.09222, 2022. +[54] Mustafa Shukor, Guillaume Couairon, and Matthieu Cord. Efficient vision-language pretraining with visual concepts and hierarchical alignment. arXiv preprint arXiv:2208.13628, 2022. +[55] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. +[56] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. +[57] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. +[58] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. +[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. +[60] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. +[61] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021. +[62] Longhui Wei, Lingxi Xie, Wengang Zhou, Houqiang Li, and Qi Tian. Mvp: Multimodality-guided visual pre-training. arXiv preprint arXiv:2203.05175, 2022. +[63] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018. +[64] Haiyang Xu, Ming Yan, Chenliang Li, Bin Bi, Songfang Huang, Wenming Xiao, and Fei Huang. E2e-vlp: End-to-end vision-language pre-training enhanced by visual learning. arXiv preprint arXiv:2106.01804, 2021. +[65] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021. +[66] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. +[67] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New + +similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. +[68] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021. +[69] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. +[70] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. +[71] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2021. +[72] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020. \ No newline at end of file diff --git a/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/images.zip b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a4d5769bd2b3d477eb5e0970f3f157fed82f2bd7 --- /dev/null +++ b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18678fae22b87643b9e0facf891da3088837144553a89af8ca7624fb2ecae399 +size 557379 diff --git a/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/layout.json b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..538fc5e2036ecdfef85f18ea24d38cec9d2a4006 --- /dev/null +++ b/2023/MaskCLIP_ Masked Self-Distillation Advances Contrastive Language-Image Pretraining/layout.json @@ -0,0 +1,10386 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 136, + 466, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 136, + 466, + 172 + ], + "spans": [ + { + "bbox": [ + 111, + 136, + 466, + 172 + ], + "type": "text", + "content": "MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image Pretraining" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "spans": [ + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": "Xiaoyi Dong" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Jianmin Bao" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Yinglin Zhang" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Ting Zhang" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Dongdong Chen" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{4,\\dagger}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Hao Yang" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Ming Zeng" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Weiming Zhang" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Lu Yuan" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Dong Chen" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Fang Wen" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": ", Nenghai Yu" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": "Microsoft Research Asia \n" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": "Xiamen University " + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 69, + 191, + 509, + 247 + ], + "type": "text", + "content": "Microsoft Cloud + AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 250, + 444, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 250, + 444, + 261 + ], + "spans": [ + { + "bbox": [ + 134, + 250, + 444, + 261 + ], + "type": "text", + "content": "{dlight@mail., zhangwm@, ynh@}.ustc.edu.cn cddlyf@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 263, + 440, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 263, + 440, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 263, + 440, + 274 + ], + "type": "text", + "content": "{jianbao, ting.zhang, luyuan, doch, fangwen}@microsoft.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 277, + 462, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 277, + 462, + 287 + ], + "spans": [ + { + "bbox": [ + 113, + 277, + 462, + 287 + ], + "type": "text", + "content": "{zhengyinglin@stu，zengming@}xmu.edu.cn yanghao.alexis@foxmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "spans": [ + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 340, + 281, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 340, + 281, + 607 + ], + "spans": [ + { + "bbox": [ + 45, + 340, + 281, + 607 + ], + "type": "text", + "content": "This paper presents a simple yet effective framework MaskCLIP, which incorporates a newly proposed masked self-distillation into contrastive language-image pretraining. The core idea of masked self-distillation is to distill representation from a full image to the representation predicted from a masked image. Such incorporation enjoys two vital benefits. First, masked self-distillation targets local patch representation learning, which is complementary to vision-language contrastive focusing on text-related representation. Second, masked self-distillation is also consistent with vision-language contrastive from the perspective of training objective as both utilize the visual encoder for feature aligning, and thus is able to learn local semantics getting indirect supervision from the language. We provide specially designed experiments with a comprehensive analysis to validate the two benefits. Symmetrically, we also introduce the local semantic supervision into the text branch, which further improves the pretraining performance. With extensive experiments, we show that MaskCLIP, when applied to various challenging downstream tasks, achieves superior results in linear probing, finetuning, and zero-shot performance with the guidance of the language encoder. Code will be release at https://github.com/LightDXY/MaskCLIP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 632, + 124, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 124, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 124, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 652, + 280, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 652, + 280, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 652, + 280, + 700 + ], + "type": "text", + "content": "Vision-language (VL) contrastive learning [31, 51] has shown remarkable success in pretraining for various tasks. With large-scale image-text pairs available on the Internet, the model composed of a simple dual encoder design learns" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 315, + 532, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 315, + 532, + 386 + ], + "spans": [ + { + "bbox": [ + 297, + 315, + 532, + 386 + ], + "type": "text", + "content": "strong semantic prior by aligning between image and text. The resulting visual encoder not only exhibits excellent linear probing and finetuning performance, but also enables impressive zero-shot performance with the guidance of the language encoder, showing the generality of natural language and its ability to supervise a wide range of visual concepts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 389, + 532, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 389, + 532, + 481 + ], + "spans": [ + { + "bbox": [ + 296, + 389, + 532, + 481 + ], + "type": "text", + "content": "Nonetheless, the associated language description, though providing richer information than mere class labels, still can hardly describe all the information in the corresponding image, as images are continuous signals with fine-grained details and complex semantics. As a result, the VL contrastive by aligning global representations may only focus on the text-described objects and ignore the rest which might be useful for downstream tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 484, + 533, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 484, + 533, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 484, + 533, + 730 + ], + "type": "text", + "content": "In this paper, we are interested in how to fully leverage the image itself to facilitate the VL contrastive to further improve the transfer capability. (1) Firstly, the learned feature representation shall characterize local patches, serving as a complementary for global representation in VL contrastive. Inspired by the recent success of masked image modeling [4, 19, 26, 51, 60, 61] in learning patch representations, we also randomly mask the input image with a large portion to force the visual encoder to focus on the remaining visible patches. (2) Secondly, the learned representation for local patches shall possess semantic meanings, being consistent with the global representation receiving semantic text supervision. We bring mean teacher self-distillation [25, 57] to supervise the learned patch representations with the visual feature representations, enabling implicit supervision from natural language. The resulting objective is denoted as masked self-distillation where the student model and the teacher model come from the same neural networks and the knowledge is distilled from the full image (fed to the teacher model) to the masked image (fed to student model). To this end, we introduce MaskCLIP by incorporating masked self" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 709, + 201, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 709, + 201, + 719 + ], + "spans": [ + { + "bbox": [ + 55, + 709, + 201, + 719 + ], + "type": "text", + "content": "*Equal contribution, † Corresponding Author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 719, + 248, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 719, + 248, + 729 + ], + "spans": [ + { + "bbox": [ + 55, + 719, + 248, + 729 + ], + "type": "text", + "content": "†Work done during an internship at Microsoft Research Asia" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "text", + "content": "10995" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "content": "distillation into VL contrastive to advance the transferable visual encoder." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 130, + 281, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 130, + 281, + 351 + ], + "spans": [ + { + "bbox": [ + 45, + 130, + 281, + 351 + ], + "type": "text", + "content": "There are several recent attempts [49, 68] also exploring the capability of the visual encoder under natural language supervision. The common approach is to introduce contrastive learning or masked image modeling on the vision side together with contrastive language-image pretraining. However, the performance indeed improves based on CLIP but does not as well as our masked self-distillation. We argue that (1) the contrastive learning objective based on central crop augmentation actually learns global representations for salient objects while lack of attention on the surrounding backgrounds [11]; and (2) masked image modeling usually needs to remap the learned representation to pixels [26] or discrete tokens [4]. Such low-level prediction target is inefficient for semantic feature learning and thus also conflicts with high-level language supervision in VL contrastive. A brief illustration is presented in Figure 1. In the experiments, we conduct comprehensive ablations to analyze the difference and provide numerical and visual evidence for better understanding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 351, + 281, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 351, + 281, + 490 + ], + "spans": [ + { + "bbox": [ + 45, + 351, + 281, + 490 + ], + "type": "text", + "content": "Symmetrically, we argue that local semantic supervision on the text branch is also helpful for the text encoder and eventually beneficial for zero-shot performance. So we introduce the same mask-data-modeling format supervision into the text branch as well. Different from images where the pixel is low-level signal, the words crafted by human beings are already highly semantic, so we use the tokenized word piece as the prediction target directly, following the well-studied mask language modeling method BERT. Meanwhile, to reduce the output conflicts between contrastive learning and mask language modeling, we introduce a small decoder for the mask language modeling branch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "spans": [ + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": "We train our MaskCLIP on a subset of a publicly available image-text pairs dataset, YFCC [58], and thoroughly evaluate the transfer ability of visual representations on several vision benchmarks: ImageNet-1K [17] for classification, ADE20K [69] for semantic segmentation, MS-COCO [40] for detection and segmentation, as well as a batch of other classification benchmarks. When it comes to ImageNet-1K [17] classification, MaskCLIP achieves " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+6.9\\%" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+7.2\\%" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+1.3\\%" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": " higher than CLIP for zero-shot transfer, linear probing, and finetuning respectively. For vision downstream tasks, we reach " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+2.7" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": " mIoU on ADE20K [69] and " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+1.8" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": " APb, " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+1.4" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": " APm on MS-COCO [40]. For vision-language tasks, MaskCLIP achieves " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+6.1\\%" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": " average zero-shot accuracy on 20 datasets, and " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+17.2\\%" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "+12.8\\%" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": " rank@1 improvement on the Flickr30K [67] image-test retrieval. In the recent Image Classification in the Wild challenge academic track, our MaskCLIP gets the " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "1_{st}" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": " result with " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "48.9\\%" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": " TOP-1 average accuracy, surpassing the second team with " + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 45, + 490, + 281, + 699 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 699, + 268, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 699, + 268, + 711 + ], + "spans": [ + { + "bbox": [ + 57, + 699, + 268, + 711 + ], + "type": "text", + "content": "In summary, the major contributions of this work are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 718, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 718, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 54, + 718, + 279, + 730 + ], + "type": "text", + "content": "1. We present a novel vision-language pretraining" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 106, + 532, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 106, + 532, + 141 + ], + "spans": [ + { + "bbox": [ + 315, + 106, + 532, + 141 + ], + "type": "text", + "content": "framework MaskCLIP, by introducing masked self-distillation objective to facilitate VL contrastive for better transferable visual models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 147, + 532, + 236 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 304, + 147, + 532, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 532, + 195 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 532, + 195 + ], + "type": "text", + "content": "2. We present extensive ablation studies on MaskCLIP variants and provide in-depth analysis numerically and visually to help understand how the proposed masked self-distillation assists VL contrastive." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 201, + 532, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 532, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 532, + 236 + ], + "type": "text", + "content": "3. We demonstrate our MaskCLIP on tens of benchmarks, showing the superiority under all three settings: zero-shot, linear probing, and finetuning." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 245, + 381, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 245, + 381, + 257 + ], + "spans": [ + { + "bbox": [ + 297, + 245, + 381, + 257 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 264, + 533, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 264, + 533, + 578 + ], + "spans": [ + { + "bbox": [ + 296, + 264, + 533, + 578 + ], + "type": "text", + "content": "Vision-language pretraining. Recent years have seen rapid progress made in vision-language pretraining [15, 18, 33, 35-39, 44-46, 50, 55, 56, 72]. Several multiple cross-modality loss functions have been proposed for the training objective, such as image-text matching [15, 37, 44, 56, 64], masked language modeling [15, 37, 44, 55, 56], masked image modeling [15, 44, 55, 56], contrastive loss [35, 38, 39]. These objects are often mixed with each other to form a compound objective. While a variety of approaches have been proposed, few works investigate the performance on visual representation learning for image classification. Recently, CLIP [51] and ALIGN [31] show that the image-text contrastive learning objective achieves promising performance for visual representation learning. There are many following works proposed to further improve the pretraining performance, DeCLIP [70], SLIP [49], COTS [43], ViCHA [54], CYCLIP [24] use additional uni/multi-modality supervision to improve the model capability, and PyramidCLIP [23], KLITE [53], IDEA [30] seek to external knowledge from pre-trained models or datasets as the additional guidance. FILIP [66] and LOUPE [34] introduce fine-grained alignment to the model. Focusing on this research direction, we analyze the desired properties of supervision which could be complementary to CLIP, and propose the masked self-distillation objective incorporated with the image-text contrastive loss to further improve pretraining performance for various visual understanding tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 579, + 533, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 579, + 533, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 579, + 533, + 730 + ], + "type": "text", + "content": "Self-supervised learning Self-supervised visual representation learning has attracted increasing attention over the past few years. The objective of the self-supervised learning is mainly divided into two categories: contrastive and generative [41]. The contrastive methods, such as MOCO [12, 27], SimCLR [9, 10], BYOL [25], SimSiam [13], and DINO [6] measure the similar and dissimilar samples by contrastive loss. Their success heavily depends on the strong data augmentation. The generative methods, such as BEiT [4], MAE [26], PeCo [19], BEVT [60], BootMAE [20] and MaskFeat [61] leverage masked image modeling to reconstruct the remaining masked part of its original input from the given visible parts. The generative methods show more" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "10996" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 46, + 103, + 119, + 261 + ], + "blocks": [ + { + "bbox": [ + 46, + 103, + 119, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 103, + 119, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 103, + 119, + 261 + ], + "type": "image", + "image_path": "572575a2acca74311944aa1ecbe913af6b324e8b33d2284f1a8940f199fd4db6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 264, + 101, + 274 + ], + "lines": [ + { + "bbox": [ + 69, + 264, + 101, + 274 + ], + "spans": [ + { + "bbox": [ + 69, + 264, + 101, + 274 + ], + "type": "text", + "content": "(a) CLIP" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 106, + 236, + 261 + ], + "blocks": [ + { + "bbox": [ + 123, + 106, + 236, + 261 + ], + "lines": [ + { + "bbox": [ + 123, + 106, + 236, + 261 + ], + "spans": [ + { + "bbox": [ + 123, + 106, + 236, + 261 + ], + "type": "image", + "image_path": "ccea09b0b77ceeb4b93c215b19706025b070148e41d5239e61a643d403af67e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 264, + 226, + 274 + ], + "lines": [ + { + "bbox": [ + 134, + 264, + 226, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 264, + 226, + 274 + ], + "type": "text", + "content": "(b) + Contrastive Learning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 106, + 343, + 261 + ], + "blocks": [ + { + "bbox": [ + 238, + 106, + 343, + 261 + ], + "lines": [ + { + "bbox": [ + 238, + 106, + 343, + 261 + ], + "spans": [ + { + "bbox": [ + 238, + 106, + 343, + 261 + ], + "type": "image", + "image_path": "aff0d5e3397fa30243d23918cc5d12debb9822f4031e7da9ce75fa8982c3c5de.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 264, + 336, + 274 + ], + "lines": [ + { + "bbox": [ + 240, + 264, + 336, + 274 + ], + "spans": [ + { + "bbox": [ + 240, + 264, + 336, + 274 + ], + "type": "text", + "content": "(c) + Mask Image Modeling" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 346, + 103, + 532, + 262 + ], + "blocks": [ + { + "bbox": [ + 346, + 103, + 532, + 262 + ], + "lines": [ + { + "bbox": [ + 346, + 103, + 532, + 262 + ], + "spans": [ + { + "bbox": [ + 346, + 103, + 532, + 262 + ], + "type": "image", + "image_path": "8c1639d6fc1293107cec6749c9c6f287453f4383955d519a03bc0bc447e41862.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 264, + 463, + 274 + ], + "lines": [ + { + "bbox": [ + 413, + 264, + 463, + 274 + ], + "spans": [ + { + "bbox": [ + 413, + 264, + 463, + 274 + ], + "type": "text", + "content": "(d) MaskCLIP" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "lines": [ + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "spans": [ + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "text", + "content": "Figure 1. Pipeline comparison between combination CLIP with different vision self-supervised learning methods. (a) Vanilla CLIP. (b) CLIP + contrastive learning. (c) CLIP + pixel prediction mask image modeling. (d) CLIP + mask self-distillation, i.e. MaskCLIP. The " + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "text", + "content": " is the text encoder and image encoder respectively, and all the " + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "text", + "content": " within each pipeline share the weight. " + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "inline_equation", + "content": "\\bar{E}_{I}" + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "text", + "content": " is the mean-teacher model, whose weight is updated by the exponential moving average of " + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 44, + 281, + 531, + 325 + ], + "type": "text", + "content": " and does not require gradient." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 334, + 280, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 334, + 280, + 379 + ], + "spans": [ + { + "bbox": [ + 45, + 334, + 280, + 379 + ], + "type": "text", + "content": "promising transfer performance than the contrastive methods, as generative objective learns patch representations while contrastive objective focuses on learning centric global representations [11]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 380, + 280, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 380, + 280, + 543 + ], + "spans": [ + { + "bbox": [ + 45, + 380, + 280, + 543 + ], + "type": "text", + "content": "Self-knowledge distillation Self-knowledge distillation [32] aims to distill the knowledge in a model itself and uses it for training the model. Instead of distilling knowledge from a pretrained teacher model [29], self-knowledge distillation regards a temporal ensemble of the student model as the teacher. It means that a student model becomes a teacher model itself, which gradually utilizes its own knowledge for softening the hard targets to be more informative during training. Self-knowledge distillation has been explored in semi-supervised learning [57], contrastive learning [16, 35], self-supervised learning [3, 7]. In this paper, we use visual features supervised by natural language for guidance in masked self-distillation which naturally fit VL contrastive to learn more transferable visual representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 553, + 116, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 553, + 116, + 564 + ], + "spans": [ + { + "bbox": [ + 45, + 553, + 116, + 564 + ], + "type": "text", + "content": "3. MaskCLIP" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "spans": [ + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "text", + "content": "We introduce MaskCLIP, a novel framework that learns visual representations. The core part of MaskCLIP is its backbone image encoder, denoted by " + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "text", + "content": " as shown in Figure 1. It obtains the transferable capability during pretraining that could benefit downstream vision tasks. Following recent self-supervised approaches [4, 14, 26, 49], we implement the backbone " + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "text", + "content": " as a Vision Transformer (ViT) [22]. The prediction results from " + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "text", + "content": " given an input image " + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 45, + 572, + 279, + 677 + ], + "type": "text", + "content": " then should be a collection of visual feature tokens, represented as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 99, + 686, + 279, + 698 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 686, + 279, + 698 + ], + "spans": [ + { + "bbox": [ + 99, + 686, + 279, + 698 + ], + "type": "interline_equation", + "content": "E _ {I} (I) = \\left\\{f _ {\\text {c l s}}, f _ {1}, f _ {2}, \\dots , f _ {N} \\right\\}. \\tag {1}", + "image_path": "a99c648e2b2aab8cc6fa245659a82dc1db6b32d4c845e5d01eba1f9f867d64e0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 706, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 729 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 45, + 706, + 279, + 729 + ], + "type": "inline_equation", + "content": "cls" + }, + { + "bbox": [ + 45, + 706, + 279, + 729 + ], + "type": "text", + "content": " is short for class token. " + }, + { + "bbox": [ + 45, + 706, + 279, + 729 + ], + "type": "inline_equation", + "content": "1,\\ldots ,N" + }, + { + "bbox": [ + 45, + 706, + 279, + 729 + ], + "type": "text", + "content": " are the indexes of the non-class tokens." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 334, + 531, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 334, + 531, + 381 + ], + "spans": [ + { + "bbox": [ + 296, + 334, + 531, + 381 + ], + "type": "text", + "content": "The rest of this section starts with the utilization of language supervision. More shall be emphasized on the masked self-distillation, which we deem crucial for visual pretraining." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 398, + 448, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 398, + 448, + 410 + ], + "spans": [ + { + "bbox": [ + 297, + 398, + 448, + 410 + ], + "type": "text", + "content": "3.1. Vision-language Contrastive" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "spans": [ + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "text", + "content": "Following [31, 51], we introduce a Transformer-based text encoder " + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "text", + "content": " to leverage language knowledge. It aims to align the global feature representations of an image and a text with respect to some forms of similarity. Precisely, consider a given image-text pair " + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "inline_equation", + "content": "\\{I, T\\}" + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "text", + "content": ", besides extracting the visual feature representation " + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "inline_equation", + "content": "E_{I}(I)" + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "text", + "content": " using the vision backbone as shown by Equation 1, we additionally use the text encoder " + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "text", + "content": " to extract linguistic features from the text " + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 296, + 420, + 531, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "spans": [ + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "type": "text", + "content": "The mean feature of the two branches are regarded as the global representations and are fed into a projection head (implemented as a fully-connected layer) respectively to obtain the metric embeddings " + }, + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "type": "inline_equation", + "content": "e^T" + }, + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "type": "inline_equation", + "content": "e^I" + }, + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "type": "text", + "content": ". Image-text contrastive loss is employed to align them during pretraining. The loss can be formulated as " + }, + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_T + \\mathcal{L}_I" + }, + { + "bbox": [ + 296, + 516, + 531, + 586 + ], + "type": "text", + "content": ", with" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 337, + 602, + 494, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 602, + 494, + 634 + ], + "spans": [ + { + "bbox": [ + 337, + 602, + 494, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp (e _ {i} ^ {I} e _ {i} ^ {T} / \\sigma)}{\\sum_ {j = 1} ^ {B} \\exp (e _ {i} ^ {I} e _ {j} ^ {T} / \\sigma)}", + "image_path": "e4a776d04bba0c64abf414ccf140910239fbf9cb1e4ba04cbd55cef799d26636.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 333, + 636, + 530, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 636, + 530, + 668 + ], + "spans": [ + { + "bbox": [ + 333, + 636, + 530, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(e _ {i} ^ {T} e _ {i} ^ {I} / \\sigma\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(e _ {i} ^ {T} e _ {j} ^ {I} / \\sigma\\right)}, \\tag {2}", + "image_path": "7b3308d243525864dc9ec81efc791a1c4e81115fc9bbf4485d162e68b0747eb8.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "type": "text", + "content": " stands for the number of image-text pairs within a training mini-batch, " + }, + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "type": "text", + "content": " are indexes within the batch; " + }, + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 297, + 682, + 530, + 730 + ], + "type": "text", + "content": " stands for the temperature for the loss functions, which is learned together with all other parameters during training." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "10997" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 106, + 264, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 106, + 264, + 117 + ], + "spans": [ + { + "bbox": [ + 46, + 106, + 264, + 117 + ], + "type": "text", + "content": "3.2. Masked Self-distillation for Visual Encoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 123, + 281, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 123, + 281, + 262 + ], + "spans": [ + { + "bbox": [ + 45, + 123, + 281, + 262 + ], + "type": "text", + "content": "Knowledge distillation is a learning paradigm where a student model is trained to match the output of a given teacher model, so that the student model can be improved by the teacher. Instead of bringing in an external teacher, self-distillation methods such as [7, 25, 57] proposes using a mean teacher model that is derived from the student itself. In specific, the teacher shares the same structure with the student, while the parameters of the teacher are exponential moving averages (EMA) of the parameters from the student. In the following, we would use the term \"EMA model\" to represent such mean teacher model constructed from the student." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "spans": [ + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": "MaskCLIP leverages the mean teacher self-distillation to enhance its vision representations. Let " + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "inline_equation", + "content": "\\bar{E}_I" + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": " be the EMA model of the backbone encoder " + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "inline_equation", + "content": "\\bar{\\theta}_t" + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": " are the parameters of " + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "inline_equation", + "content": "\\bar{E}_I" + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": " at training step " + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "inline_equation", + "content": "\\bar{\\theta}_t" + }, + { + "bbox": [ + 45, + 263, + 281, + 310 + ], + "type": "text", + "content": " is updated with" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 316, + 279, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 316, + 279, + 330 + ], + "spans": [ + { + "bbox": [ + 111, + 316, + 279, + 330 + ], + "type": "interline_equation", + "content": "\\bar {\\theta} _ {t} = \\alpha \\bar {\\theta} _ {t - 1} + (1 - \\alpha) \\theta_ {t}, \\tag {3}", + "image_path": "874f5cafa48702afe17dcec52e0d141aa85eed746124be0ad12eaf82ff90bd64.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 335, + 281, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 335, + 281, + 381 + ], + "spans": [ + { + "bbox": [ + 45, + 335, + 281, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 335, + 281, + 381 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 335, + 281, + 381 + ], + "type": "text", + "content": " is a hyper-parameter for smoothing updates. We propose to incorporate masked image modeling into self-distillation, resulting in masked self-distillation with asymmetric input for student model and teacher model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 382, + 281, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 382, + 281, + 426 + ], + "spans": [ + { + "bbox": [ + 45, + 382, + 281, + 426 + ], + "type": "text", + "content": "In specific, considering a given input image " + }, + { + "bbox": [ + 45, + 382, + 281, + 426 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 45, + 382, + 281, + 426 + ], + "type": "text", + "content": ", we first feed it to the EMA model " + }, + { + "bbox": [ + 45, + 382, + 281, + 426 + ], + "type": "inline_equation", + "content": "\\bar{E}_I" + }, + { + "bbox": [ + 45, + 382, + 281, + 426 + ], + "type": "text", + "content": " (teacher model) to obtain the distillation targets. These target features can be represented as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 427, + 279, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 427, + 279, + 441 + ], + "spans": [ + { + "bbox": [ + 99, + 427, + 279, + 441 + ], + "type": "interline_equation", + "content": "\\bar {E} _ {I} (I) = \\{\\bar {f} _ {c l s}, \\bar {f} _ {1}, \\bar {f} _ {2}, \\dots , \\bar {f} _ {N} \\}. \\tag {4}", + "image_path": "cb9268c7f5ec06ca49c6b1f6fea6571028ad921e470d87bf256bf2c87ec55bbb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "spans": [ + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": "In the meantime, we randomly mask a large portion of the input image patches and then feed it into the original backbone " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": " (student model). Following [26], we only feed the visible (unmasked) patches, denoted by " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": ", into the original backbone " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": " to speed up computation and save memory. Let " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": " be the indexes of all the masked tokens. These encoded features corresponding to visible tokens can then be denoted as " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "E_I(I') = \\{f_{cls}'\\} \\cup \\{f_{k \\notin \\mathcal{M}}'\\}" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": ". They are then joined with a shared and learnable feature vector, denoted as " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": ", that represents mask tokens, to form a complete set of features " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "\\{f_{cls}', f_1', f_2', \\ldots, f_N'\\}" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "f_{i \\in \\mathcal{M}}' = m" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": ". We attach positional embeddings onto all these tokens, and append a small Transformer " + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 45, + 444, + 281, + 612 + ], + "type": "text", + "content": " as a decoder to predict features of the masked region from the visible tokens, which could be formulated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 618, + 279, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 618, + 279, + 646 + ], + "spans": [ + { + "bbox": [ + 77, + 618, + 279, + 646 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(D \\circ E _ {I}\\right) \\left(I ^ {\\prime}\\right) = D \\left(\\left\\{f _ {c l s} ^ {\\prime}, f _ {1} ^ {\\prime}, f _ {2} ^ {\\prime}, \\dots , f _ {N} ^ {\\prime} \\right\\}\\right) \\\\ = \\left\\{f _ {c l s} ^ {\\prime \\prime}, f _ {1} ^ {\\prime \\prime}, f _ {2} ^ {\\prime \\prime}, \\dots , f _ {N} ^ {\\prime \\prime} \\right\\}. \\tag {5} \\\\ \\end{array}", + "image_path": "1f0401fa61824dcd22da61bcbb7790e73e8548546179d0b3abf3880c37687ef4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 651, + 279, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 651, + 279, + 698 + ], + "spans": [ + { + "bbox": [ + 45, + 651, + 279, + 698 + ], + "type": "text", + "content": "Inspired by [71], we use an online quantizer " + }, + { + "bbox": [ + 45, + 651, + 279, + 698 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 45, + 651, + 279, + 698 + ], + "type": "text", + "content": " to transform the output features into a soft codewords distribution, and minimize the cross-entropy between the target features and the predicted features. Formally," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 703, + 279, + 731 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 703, + 279, + 731 + ], + "spans": [ + { + "bbox": [ + 86, + 703, + 279, + 731 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {D i s t}} = \\frac {1}{| \\mathcal {M} |} \\sum_ {k \\in \\mathcal {M}} - \\bar {h} \\left(\\bar {f} _ {k}\\right) ^ {T} \\log h \\left(f _ {k} ^ {\\prime \\prime}\\right). \\tag {6}", + "image_path": "3e7d0a51377f0035487d1c7079a90fee217d56b90938bdc22633b4bedb454c9f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 106, + 531, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 531, + 129 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 531, + 129 + ], + "type": "text", + "content": "here the parameter of the teacher quantizer " + }, + { + "bbox": [ + 296, + 106, + 531, + 129 + ], + "type": "inline_equation", + "content": "\\bar{h}()" + }, + { + "bbox": [ + 296, + 106, + 531, + 129 + ], + "type": "text", + "content": " is also EMA updated by the online quantizer, similar to the teacher model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 136, + 514, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 136, + 514, + 149 + ], + "spans": [ + { + "bbox": [ + 296, + 136, + 514, + 149 + ], + "type": "text", + "content": "3.3. Local Semantic Learning for Text Encoder" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "spans": [ + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "text", + "content": "Besides the local semantic supervision for the visual encoder, we argue it is also helpful for the text encoder. So we introduce the BERT pretraining into the text branch. For the text " + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "inline_equation", + "content": "T = \\{t_{\\text{sos}}, t_1, t_2, \\dots, t_M, t_{\\text{eos}}\\}" + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "text", + "content": ", we denote the masked input as " + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "inline_equation", + "content": "T' = \\{t_{\\text{sos}}', t_1', t_2', \\dots, t_M', t_{\\text{eos}}'\\}" + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "inline_equation", + "content": "t_{i \\in \\mathcal{M}_T}' = m_t" + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "inline_equation", + "content": "t_{i \\notin \\mathcal{M}_T}' = t_i" + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_T" + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "text", + "content": " be the indexes of all the masked text tokens. The output feature of the encoder is " + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "inline_equation", + "content": "E_T(T')" + }, + { + "bbox": [ + 296, + 154, + 531, + 246 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "spans": [ + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "text", + "content": "To reduce the output conflict between the global image-text contrastive learning and the local mask language modeling, we further introduce a small text decoder, which shares the same architecture as the encoder but with only a few layers. So that the global prediction and local prediction are conducted at different layers. We denote the output feature as: " + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "inline_equation", + "content": "(D_T\\circ E_T)(T') = \\{t_{sos}''" + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "inline_equation", + "content": "t_1''" + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "inline_equation", + "content": "t_2''" + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "text", + "content": ", ..., " + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "inline_equation", + "content": "t_M''" + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "inline_equation", + "content": "t_{eos}''\\}" + }, + { + "bbox": [ + 296, + 247, + 531, + 340 + ], + "type": "text", + "content": " and the loss could be formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 346, + 345, + 530, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 345, + 530, + 374 + ], + "spans": [ + { + "bbox": [ + 346, + 345, + 530, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M L M} = \\frac {1}{| \\mathcal {M} _ {T} |} \\sum_ {k \\in \\mathcal {M} _ {T}} - t _ {k} ^ {T} \\log t _ {k} ^ {\\prime \\prime}. \\tag {7}", + "image_path": "ba0ead08196a1d6ee4aca41f89435d6e8ab640e886ffd3d5630291e38dbe0f2b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 380, + 425, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 380, + 425, + 392 + ], + "spans": [ + { + "bbox": [ + 296, + 380, + 425, + 392 + ], + "type": "text", + "content": "3.4. Overall Loss Functions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 399, + 531, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 399, + 531, + 421 + ], + "spans": [ + { + "bbox": [ + 296, + 399, + 531, + 421 + ], + "type": "text", + "content": "Finally, we pretrain MaskCLIP with all these losses combined:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 355, + 430, + 530, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 430, + 530, + 442 + ], + "spans": [ + { + "bbox": [ + 355, + 430, + 530, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I} + \\mathcal {L} _ {T} + \\lambda \\mathcal {L} _ {\\text {D i s t}} + \\beta \\mathcal {L} _ {\\text {M L M}}, \\tag {8}", + "image_path": "8f766e8c28fdc0972d94ebceab9f0090391f280dec7ad9a08ee89d0173d1b7d0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "spans": [ + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "inline_equation", + "content": "\\lambda, \\beta" + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "text", + "content": " being the hyper-parameter weighting between VL contrastive loss and self-supervised learning loss. All the components of MaskCLIP are trained from scratch, including the visual backbone " + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "text", + "content": ", the visual decoder " + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "text", + "content": ", the text encoder " + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "inline_equation", + "content": "E_T" + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "text", + "content": ", as well as the text decoder " + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 296, + 449, + 531, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 517, + 377, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 517, + 377, + 531 + ], + "spans": [ + { + "bbox": [ + 296, + 517, + 377, + 531 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 538, + 346, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 538, + 346, + 549 + ], + "spans": [ + { + "bbox": [ + 296, + 538, + 346, + 549 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "spans": [ + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "content": "Model architecture. Our framework consists of the visual encoder " + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "content": ", the text encoder " + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "content": ", the visual decoder " + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "content": ", and the text decoder " + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "content": ". We adopt the widely used Transformer ViT-B/16 [22] for a fair comparison. It is composed of 12 layers, 768 width, and 12 head. The input image is " + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "content": " resolution and is further split into " + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "content": " patches with size " + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 296, + 554, + 531, + 694 + ], + "type": "text", + "content": ". A learnablecls token is pretended to the 196 embeddings. For the text encoder, we adopt a 12-layer, 512-width, and 8-head Transformer following CLIP [51], and the text decoder has 4 layers. The number of text tokens is fixed to 77 with necessary truncations or paddings. For the image decoder, we directly use a one-layer Vision Transformer." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 296, + 694, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 694, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 694, + 531, + 730 + ], + "type": "text", + "content": "Pretraining details. We train our proposed MaskCLIP from scratch for 25 epochs, the batch size is fixed to 4096 for all the experiments. The masks used in the mask self-distillation" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "10998" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 101, + 520, + 362 + ], + "blocks": [ + { + "bbox": [ + 49, + 101, + 520, + 362 + ], + "lines": [ + { + "bbox": [ + 49, + 101, + 520, + 362 + ], + "spans": [ + { + "bbox": [ + 49, + 101, + 520, + 362 + ], + "type": "image", + "image_path": "14487a7fcb5cc6382e984dc0b94f915e3aa1a7bab539289409b1dceb7dacd090.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 363, + 531, + 395 + ], + "lines": [ + { + "bbox": [ + 45, + 363, + 531, + 395 + ], + "spans": [ + { + "bbox": [ + 45, + 363, + 531, + 395 + ], + "type": "text", + "content": "Figure 2. Visualization of the similarity between text and image features. The images and captions are from the MS-COCO val set. Here we show the image feature similarity with both full caption and different objects in it. The caption is \"Three teddy bears sit in a sled in snow\". More results could be found in the supplemental materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 405, + 286, + 481 + ], + "blocks": [ + { + "bbox": [ + 47, + 405, + 286, + 481 + ], + "lines": [ + { + "bbox": [ + 47, + 405, + 286, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 286, + 481 + ], + "type": "table", + "html": "
TrainingIN-1KFlicker30K
MemoryTime0-shotLinearFinetuneI2TT2I
CLIP14G1.00×37.666.582.352.932.8
CLIP+SimCLR30G2.67×42.872.182.658.641.3
CLIP+MAE16G1.30×42.168.583.257.341.1
MaskCLIP19G1.75×44.573.783.670.145.6
", + "image_path": "cb825e38a92ba65ededb5551252a33c6fd760bd5564fab181f6830a28b7a4b13.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 483, + 279, + 505 + ], + "lines": [ + { + "bbox": [ + 45, + 483, + 279, + 505 + ], + "spans": [ + { + "bbox": [ + 45, + 483, + 279, + 505 + ], + "type": "text", + "content": "Table 1. Results of boosting CLIP with different kinds of vision self-supervised learning methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 513, + 279, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 513, + 279, + 559 + ], + "spans": [ + { + "bbox": [ + 45, + 513, + 279, + 559 + ], + "type": "text", + "content": "branch and mask language modeling branch are random mask with a mask ratio of " + }, + { + "bbox": [ + 45, + 513, + 279, + 559 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 45, + 513, + 279, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 513, + 279, + 559 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 45, + 513, + 279, + 559 + ], + "type": "text", + "content": ". We pretrain all the models with the commonly used YFCC15M dataset, which is flited from the YFCC100M [58] dataset by [51]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 560, + 279, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 560, + 279, + 677 + ], + "spans": [ + { + "bbox": [ + 45, + 560, + 279, + 677 + ], + "type": "text", + "content": "Downstream details. We evaluated MaskCLIP on several downstream datasets, including ImageNet-1K [17], ADE20K [69], MS-COCO [40], Flicr30K [67] et al. For ImageNet-1K, we report zero-shot, linear probing, and finetuning performance. The zero-shot is conducted following the label prompt setting in SLIP [49]. For linear probing, we fix the backbone and train a new linear classifier for 90 epochs. For finetuning, we follow the setting in BEiT [4] and finetune the model for 100 epochs with a layer-decayed learning rate. See supplemental materials for more details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 687, + 107, + 699 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 687, + 107, + 699 + ], + "spans": [ + { + "bbox": [ + 45, + 687, + 107, + 699 + ], + "type": "text", + "content": "4.2. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "We first present our analysis by studying different ways of boosting CLIP. The baseline is CLIP [51] trained on the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 407, + 531, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 407, + 531, + 478 + ], + "spans": [ + { + "bbox": [ + 296, + 407, + 531, + 478 + ], + "type": "text", + "content": "YFCC-15M. Besides the introduced masked self-distillation, we consider two other popular methods: (1) SimCLR [9], a representative contrastive method; and (2) MAE [26] the state-of-the-art masked image modeling approaches. All the compared methods are trained on the YFCC-15M for a fair comparison. We have the following observations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "spans": [ + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "text", + "content": "Vision self-supervision helps VL contrastive. We evaluate the models on both vision task ImageNet-1K [17] classification and vision-language task image-text retrieval on Flicker30K [67] and present the comparison in Table 1. All the added vision self-supervision, regardless of contrastive or generative, improves the baseline CLIP. Among them, our proposed MaskCLIP achieves the best results in terms of all the evaluation metrics, outperforming CLIP with " + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "inline_equation", + "content": "+6.9\\%" + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "inline_equation", + "content": "+7.2\\%" + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "inline_equation", + "content": "+1.3\\%" + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "text", + "content": " on ImageNet-1K classification for zero-shot, linear probing, and finetuning respectively, and " + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "inline_equation", + "content": "+17.2\\%" + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "inline_equation", + "content": "+12.8\\%" + }, + { + "bbox": [ + 296, + 481, + 531, + 667 + ], + "type": "text", + "content": " on Flicker30K for image-to-text retrieval and text-to-image retrieval. We also report the training GPU memory usage and time-consuming cost in Table 1. It is worth noting that the contrastive model (CLIP + SimCLR) compares two additional views of the input image, resulting in larger GPU memory usage and longer training time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 671, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 671, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 671, + 531, + 730 + ], + "type": "text", + "content": "Masked image modeling is able to learn representations for local patches. We argue that the image encoder only pays attention to the text-described objects under VL contrastive due to sparse text description and to the centric objects under image contrastive due to central-crop augmenta" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "10999" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 105, + 280, + 180 + ], + "blocks": [ + { + "bbox": [ + 50, + 105, + 280, + 180 + ], + "lines": [ + { + "bbox": [ + 50, + 105, + 280, + 180 + ], + "spans": [ + { + "bbox": [ + 50, + 105, + 280, + 180 + ], + "type": "table", + "html": "
MethodObjectiveADE20K mIoUPascal mIoU
CLIPGlobal7.213.5
CLIP + SimCLRGlobal + Global6.311.9
CLIP + MAEGlobal + Pixel-wise Local8.316.4
MaskCLIPGlobal + Token-wise Local10.217.2
", + "image_path": "5a13b4fd09b7a76e4cc06d7a4f4635c2f678e0d747b1b681c22a0f49fdbe2a87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 209, + 280, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 209, + 280, + 395 + ], + "spans": [ + { + "bbox": [ + 45, + 209, + 280, + 395 + ], + "type": "text", + "content": "tion. In contrast, masked image modeling forces the image encoder to focus on local patches using token-wise objectives by mandatorily masking a large portion of patches. Here, we provide numerical comparisons for evidence. We conduct an \"Annotation-free zero-shot segmentation\" experiment to test the zero-shot segmentation. The results on such a dense prediction task would better reveal the ability of local patch representations than global classification. Following the design in DenseCLIP [70], we use the prompted label feature as the linear classification weight to realize segmentation, without any training procedure. We evaluate the performance on two widely used datasets: ADE20K [69] and Pascal Context [48]. The results are shown in Table.2. We can see that equipped with masked image modeling, our MaskCLIP as well as CLIP+MAE achieves better results than CLIP and CLIP+SimCLR, validating our hypothesis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 395, + 280, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 395, + 280, + 676 + ], + "spans": [ + { + "bbox": [ + 45, + 395, + 280, + 676 + ], + "type": "text", + "content": "Masked self-distillation learns semantic representations for local patches. Our masked self-distillation predicts visual features dynamically outputted by the visual encoder and thus implicitly gets supervision from the text side via VL contrastive. While MAE predicts fixed low-level pixels, making it inefficient to learn semantic representations (as the objective may force the representation to memorize low-level details) and thus causing conflict with VL contrastive. To show this, we select images from MS-COCO [40] and calculate the feature similarity between image features and their corresponding caption features. We also select objects in the caption, prompt it to a new caption, such as \"a photo of teddy bears\", and calculate the similarities. An example is shown in Figure 2 (More can be found in the supplementary material). Comparing MaskCLIP with CLIP+MAE in the fourth column, we can see that CLIP+MAE uses color as evidence and fails to distinguish the white teddy bear from the white snow. While our MaskCLIP successfully differentiates the two objects, suggesting ours learn more semantic features. On the other hand, the superior results of MaskCLIP shown in Table 1 and Table 2 also validate this. It is worth mentioning that CLIP and CLIP+SimCLR fail to have a correct response partition for different single objects like MaskCLIP, further justifying our second observation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 679, + 232, + 692 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 679, + 232, + 692 + ], + "spans": [ + { + "bbox": [ + 45, + 679, + 232, + 692 + ], + "type": "text", + "content": "4.3. Comparison with Previous Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": "To show the effectiveness of MaskCLIP as a general vision-language pretrain method, we conduct experiments on both vision tasks and vision-language tasks. For vision" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 302, + 105, + 532, + 204 + ], + "blocks": [ + { + "bbox": [ + 45, + 182, + 279, + 202 + ], + "lines": [ + { + "bbox": [ + 45, + 182, + 279, + 202 + ], + "spans": [ + { + "bbox": [ + 45, + 182, + 279, + 202 + ], + "type": "text", + "content": "Table 2. Annotation-free zero-shot segmentation results on ADE20K and Pascal Context." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 302, + 105, + 532, + 204 + ], + "lines": [ + { + "bbox": [ + 302, + 105, + 532, + 204 + ], + "spans": [ + { + "bbox": [ + 302, + 105, + 532, + 204 + ], + "type": "table", + "html": "
MethodEpochIN-1KADE20KMS-COCO
0-ShotLinFTmIoUAPbAPm
DeiT [59]300*--81.847.444.139.8
SimCLR [9]25-64.082.548.044.640.2
MAE [26]25-56.282.546.543.239.1
CLIP [51]2537.666.582.347.843.639.5
SLIP [49]2542.872.182.648.544.040.3
MaskCLIP2544.573.783.650.545.440.9
", + "image_path": "00194b401d3638d249f4357215982875f5189df5992229fb375e856017248cac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 207, + 531, + 250 + ], + "lines": [ + { + "bbox": [ + 296, + 207, + 531, + 250 + ], + "spans": [ + { + "bbox": [ + 296, + 207, + 531, + 250 + ], + "type": "text", + "content": "Table 3. Comparison with previous methods, including supervised baselines, self-supervised pretraining methods, and vision-language pretraining methods. * is the epoch of the supervised baseline on ImageNet-1K." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 261, + 532, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 261, + 532, + 389 + ], + "spans": [ + { + "bbox": [ + 296, + 261, + 532, + 389 + ], + "type": "text", + "content": "tasks, we report results on ImageNet-1K [17] classification, MS-COCO [40] object detection, and ADE20K [69] semantic segmentation. For vision-language tasks, we report zero-shot results on recent challenging ICinW 20 datasets benchmark and image-text retrieval results on Flickr30K [67] and MS-COCO [40]. In the following, we compare with the supervised baseline DeiT [59], self-supervised methods SimCLR [9] and MAE [26], and vision-language methods CLIP [51] and SLIP [49]. For a fair comparison, we train SimCLR and MAE on YFCC-15M [58] with the same epochs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "spans": [ + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "text", + "content": "Classification on ImageNet-1K. As shown in Table 3, MaskCLIP benefits from the advantages of both VL pretraining and image mask self-distillation that shows strong performance on all the metrics. For zero-shot tasks, MaskCLIP outperforms CLIP by " + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "inline_equation", + "content": "+6.9\\%" + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "text", + "content": " with 25 epoch training and achieves " + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "inline_equation", + "content": "+1.7\\%" + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "text", + "content": " higher than the recent work SLIP. When it comes to finetune, MaskCLIP reaches " + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "inline_equation", + "content": "83.6\\%" + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "text", + "content": " top-1 accuracy, and outperforms CLIP by " + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "inline_equation", + "content": "+1.3\\%" + }, + { + "bbox": [ + 296, + 389, + 532, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "spans": [ + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "text", + "content": "Semantic segmentation on ADE20K. Then we apply our MaskCLIP to the semantic segmentation task. Here we use the UperNet [63] framework with " + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "text", + "content": " input and end-to-end training for 160K iterations. The evaluation metric is the mean Intersection of Union (mIoU) and we report single-scale evaluation results here. The results are given in Table 3. Our method achieves " + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "inline_equation", + "content": "50.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "inline_equation", + "content": "+2.7\\mathrm{mIoU}" + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "text", + "content": " than our baseline method CLIP, and " + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "inline_equation", + "content": "+2.0\\mathrm{mIoU}" + }, + { + "bbox": [ + 296, + 484, + 531, + 588 + ], + "type": "text", + "content": " than SLIP. This verifies the effectiveness of our introduced incorporation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "spans": [ + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "type": "text", + "content": "Object detection and instance segmentation on MS-COCO. We further investigate our transfer performance on object detection and instance segmentation in Table.3. Here we use Mask-RCNN [28] framework with single-scale input and " + }, + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "type": "text", + "content": " schedule (12 epochs). Our method achieves 45.4 box AP and 40.9 mask AP, " + }, + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "type": "inline_equation", + "content": "+1.8 / 1.4" + }, + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "type": "text", + "content": " better than CLIP, and " + }, + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "type": "inline_equation", + "content": "+1.4 / 0.6" + }, + { + "bbox": [ + 296, + 589, + 532, + 670 + ], + "type": "text", + "content": " better than SLIP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 671, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 671, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 671, + 532, + 730 + ], + "type": "text", + "content": "Zero-shot on small datasets. We also report zero-shot performance on 20 small datasets under the ICinW setting (see the introduction below) in Table 4. We find that all the methods perform poorly on some datasets such as Aircraft(1% acc for random guessing, we omit the description in the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "text", + "content": "11000" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 99, + 530, + 260 + ], + "blocks": [ + { + "bbox": [ + 48, + 99, + 530, + 260 + ], + "lines": [ + { + "bbox": [ + 48, + 99, + 530, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 99, + 530, + 260 + ], + "type": "table", + "html": "
AverageCaltech-101CIFAR-10CIFAR-100Country211DTDEuroSATFER-2013AircraftFood-101GTSRBMemesKittiDisMNISTFlowersPetsPatchCamSST2RESISC45CarsVoc2007
Pretraining on YFCC-15M
CLIP34.058.668.536.910.821.430.516.95.151.66.551.125.95.052.728.651.752.522.44.579.1
SLIP37.870.982.648.611.826.619.818.15.659.912.651.829.49.856.331.455.351.528.55.480.5
MaskCLIP40.172.080.257.512.627.944.020.36.164.98.552.034.34.957.034.350.149.935.76.782.1
Pretraining on ICinW Academic Track Siting: YFCC-15M, GCC3M+12M, ImageNet-21K(ImageNet-1K is removed)
1st MaskCLIP48.986.495.378.311.633.057.718.88.078.917.352.816.07.374.274.452.146.254.326.582.3
2nd KLITE*45.587.492.768.88.232.227.917.44.372.411.448.431.112.875.665.950.652.944.410.282.3
3rd YT-CLIP44.577.883.558.411.931.940.727.16.968.718.852.39.118.853.169.351.550.352.719.779.3
4th UniCL†44.084.890.267.86.725.435.330.83.568.311.151.017.911.371.744.952.149.541.424.281.3
5th Gramer*43.283.992.969.57.325.524.430.42.771.09.052.612.410.170.452.450.650.144.813.881.3
", + "image_path": "8f16ac5741607f4ef3518057800ca0888c8d8c789218e3708f1bd59641c53d93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 286, + 530, + 360 + ], + "blocks": [ + { + "bbox": [ + 46, + 262, + 531, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 531, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 531, + 283 + ], + "type": "text", + "content": "Table 4. Zero-shot evaluation on ICinW classification benchmarks. Best results in bold. * indicates using Swin-B as the backbone, † indicates using Focal-B as the backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 286, + 530, + 360 + ], + "lines": [ + { + "bbox": [ + 51, + 286, + 530, + 360 + ], + "spans": [ + { + "bbox": [ + 51, + 286, + 530, + 360 + ], + "type": "table", + "html": "
Training EpochFlickr30KMS-COCO
Image-to-textText-to-imageImage-to-textText-to-image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP [51]2552.979.687.232.860.871.227.553.565.017.738.850.5
SLIP [49]2558.685.191.741.368.778.633.459.870.621.544.456.3
MaskCLIP2570.190.395.345.673.482.141.467.977.525.549.761.3
", + "image_path": "6ae827564b38c8c1c0547513255bd72f70ecfe97307c5dbb8a7cdb347e00b939.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 362, + 480, + 372 + ], + "lines": [ + { + "bbox": [ + 95, + 362, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 95, + 362, + 480, + 372 + ], + "type": "text", + "content": "Table 5. Results of zero-shot image-text retrieval on Flickr30K and MS-COCO datasets. Best results in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "spans": [ + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "text", + "content": "following), " + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "inline_equation", + "content": "\\mathrm{Fer}(24.7\\%)" + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "text", + "content": ", Country211 " + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "inline_equation", + "content": "(0.5\\%)" + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "text", + "content": ", GTSRB " + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "inline_equation", + "content": "(5.9\\%)" + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "text", + "content": ", Cars " + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "inline_equation", + "content": "(0.8\\%)" + }, + { + "bbox": [ + 45, + 384, + 280, + 454 + ], + "type": "text", + "content": ". This might be caused by the data domain gap that the YFCC-15M contains few related images and descriptions. For the rest of the datasets, all the methods get reasonable performance and our MaskCLIP gets the best performance on most datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "spans": [ + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "text", + "content": "Image Classification in the Wild (ICinW) Challenge The ICinW challenge [1] is a newly proposed visual pretraining benchmark, which contains 20 diverse downstream classification datasets, measuring the ability of pre-training models on both the prediction accuracy and their transfer efficiency in a new task. The pretraining is limited to three datasets: YFCC-15M [58], GCC3M [52] " + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "inline_equation", + "content": "+12\\mathrm{M}" + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "text", + "content": " [8] and ImageNet-21K [17] (ImageNet-1K data is excluded). We pretrain our MaskCLIP on it and get the " + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{st}" + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "text", + "content": " result in the zero-shot track [2] (we submit the results anonymously). As shown in Table 4, the " + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "inline_equation", + "content": "2_{nd}" + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "text", + "content": " team KLITE uses a strong Swin-B [42] as the backbone and additional knowledge from GPT-3 [5] and Wiktionary [47], and the " + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "inline_equation", + "content": "4_{th}" + }, + { + "bbox": [ + 45, + 458, + 282, + 633 + ], + "type": "text", + "content": " use the strong Focal-B [65] as the backbone, while our MaskCLIP greatly outperforms these methods with a simple ViT-B backbone and no additional knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 636, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 636, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 636, + 280, + 730 + ], + "type": "text", + "content": "Zero-shot on text-image retrieval. We further report the zero-shot text-image retrieval results on two benchmark datasets, Flickr30K [67] and MS-COCO [40]. We find that the text without any prefixes or suffixes works well for all the models. Table 5 shows the results. We can see that MaskCLIP exhibits a strong zero-shot performance. For example, with 25 epochs training, MaskCLIP reaches " + }, + { + "bbox": [ + 45, + 636, + 280, + 730 + ], + "type": "inline_equation", + "content": "41.4\\%" + }, + { + "bbox": [ + 45, + 636, + 280, + 730 + ], + "type": "text", + "content": " Rank@1 image-to-text accuracy on MS-COCO, outperform-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "spans": [ + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "type": "text", + "content": "ing CLIP with " + }, + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "type": "inline_equation", + "content": "13.9\\%" + }, + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "type": "inline_equation", + "content": "25.5\\%" + }, + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "type": "text", + "content": " Rank@1 text-to-image accuracy, " + }, + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "type": "inline_equation", + "content": "+7.8\\%" + }, + { + "bbox": [ + 296, + 384, + 530, + 408 + ], + "type": "text", + "content": " higher than CLIP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 412, + 363, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 412, + 363, + 423 + ], + "spans": [ + { + "bbox": [ + 297, + 412, + 363, + 423 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 427, + 530, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 427, + 530, + 450 + ], + "spans": [ + { + "bbox": [ + 296, + 427, + 530, + 450 + ], + "type": "text", + "content": "We compare our default settings with other alternatives to justify the efficacy of our model designs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 450, + 531, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 450, + 531, + 543 + ], + "spans": [ + { + "bbox": [ + 296, + 450, + 531, + 543 + ], + "type": "text", + "content": "Training objectives ablation. As shown in Table.6a, when we remove the mask language modeling loss " + }, + { + "bbox": [ + 296, + 450, + 531, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MLM}}" + }, + { + "bbox": [ + 296, + 450, + 531, + 543 + ], + "type": "text", + "content": ", the performance of the image-text task drops, including the zero-shot accuracy and retrial performance. While benefiting from the distillation loss, the finetuning performance on ImageNet-1K is not influenced. When we remove the distillation loss " + }, + { + "bbox": [ + 296, + 450, + 531, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Dis}}" + }, + { + "bbox": [ + 296, + 450, + 531, + 543 + ], + "type": "text", + "content": ", we observe a performance drop on all tasks, especially the finetuning results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 543, + 532, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 543, + 532, + 671 + ], + "spans": [ + { + "bbox": [ + 296, + 543, + 532, + 671 + ], + "type": "text", + "content": "Distillation loss format. Different from previous methods [3,21,26] that calculate the per-element distance as the loss function, we use an online tokenizer to map the feature to soft codewords and use the cross-entropy loss as the supervision. Here we study their difference in Table.6b. We find that although they get similar fine-tuning performance, the CE loss gets better zero-shot and linear probing performance. The reason may be that the per-element MSE loss leads the model to fit some unnecessary details of the target feature, while the CE loss with soft tokenizer helps the model to focus more on the important feature." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "type": "text", + "content": "Distillation & MLM loss weight. Here we set the loss weight of the CLIP branch as 1 and study the loss weight of the two additional branches. As shown in Table.6e and Table.6f, setting " + }, + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 296, + 671, + 530, + 730 + ], + "type": "text", + "content": " emphasize too much on new tasks, which mislead the model to a wrong converge" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "11001" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 103, + 202, + 154 + ], + "blocks": [ + { + "bbox": [ + 52, + 103, + 202, + 154 + ], + "lines": [ + { + "bbox": [ + 52, + 103, + 202, + 154 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 202, + 154 + ], + "type": "table", + "html": "
Model0-ShotFTI2T/T2I
MaskCLIP44.583.670.1/45.6
w/o LMLM42.883.665.0/41.6
w/o LDis42.082.465.4/40.5
", + "image_path": "6f9990abb2bf4a59ed8e24b84f02e432963119a929b5171ef95416d1f84fcc12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 186, + 185, + 257 + ], + "blocks": [ + { + "bbox": [ + 48, + 156, + 196, + 174 + ], + "lines": [ + { + "bbox": [ + 48, + 156, + 196, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 156, + 196, + 174 + ], + "type": "text", + "content": "(a) Training Objectives ablation. Both is necessary for MaskCLIP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 186, + 185, + 257 + ], + "lines": [ + { + "bbox": [ + 62, + 186, + 185, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 185, + 257 + ], + "type": "table", + "html": "
Depth0-ShotI2T/T2I
043.565.2/44.1
144.370.4/45.3
244.370.2/45.4
444.570.1/45.6
844.267.5/44.7
", + "image_path": "e67951f014a0995a170122980409bbb78cd1ba8e603324e7595e43e9664ae6cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 229, + 103, + 350, + 154 + ], + "blocks": [ + { + "bbox": [ + 48, + 259, + 196, + 277 + ], + "lines": [ + { + "bbox": [ + 48, + 259, + 196, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 196, + 277 + ], + "type": "text", + "content": "(d) Text decoder depth. The decoder is necessary and a shallow one works better." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 229, + 103, + 350, + 154 + ], + "lines": [ + { + "bbox": [ + 229, + 103, + 350, + 154 + ], + "spans": [ + { + "bbox": [ + 229, + 103, + 350, + 154 + ], + "type": "table", + "html": "
Loss0-ShotLinFT
MSE43.873.283.6
CE44.573.783.6
", + "image_path": "598e85f86ffa043ccde524e12058f06e3a1d368c20a33e0034327d3e57e24a88.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 228, + 186, + 351, + 257 + ], + "blocks": [ + { + "bbox": [ + 214, + 156, + 362, + 183 + ], + "lines": [ + { + "bbox": [ + 214, + 156, + 362, + 183 + ], + "spans": [ + { + "bbox": [ + 214, + 156, + 362, + 183 + ], + "type": "text", + "content": "(b) Distillation loss format. The online tokenizer with cross-entropy loss works slightly better than MSE loss." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 228, + 186, + 351, + 257 + ], + "lines": [ + { + "bbox": [ + 228, + 186, + 351, + 257 + ], + "spans": [ + { + "bbox": [ + 228, + 186, + 351, + 257 + ], + "type": "table", + "html": "
Weight0-ShotLinFT
138.568.282.5
0.144.473.583.5
0.0544.573.783.6
0.0143.673.083.4
", + "image_path": "108f3f8d0001d1385786867ad3e246c537136b405ad77c65d3691e17c53b4b6b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 402, + 103, + 509, + 154 + ], + "blocks": [ + { + "bbox": [ + 214, + 259, + 361, + 277 + ], + "lines": [ + { + "bbox": [ + 214, + 259, + 361, + 277 + ], + "spans": [ + { + "bbox": [ + 214, + 259, + 361, + 277 + ], + "type": "text", + "content": "(e) Distillation loss weight. A small loss weight works well for MaskCLIP." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 402, + 103, + 509, + 154 + ], + "lines": [ + { + "bbox": [ + 402, + 103, + 509, + 154 + ], + "spans": [ + { + "bbox": [ + 402, + 103, + 509, + 154 + ], + "type": "table", + "html": "
Depth0-ShotLinFT
144.573.783.6
243.772.983.4
443.572.583.3
", + "image_path": "d61ad7835f76f4711d36b8203f24957d5377a022edf767e19cbf8e147b1c5f93.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 401, + 186, + 510, + 257 + ], + "blocks": [ + { + "bbox": [ + 381, + 156, + 528, + 175 + ], + "lines": [ + { + "bbox": [ + 381, + 156, + 528, + 175 + ], + "spans": [ + { + "bbox": [ + 381, + 156, + 528, + 175 + ], + "type": "text", + "content": "(c) Visual decoder Depth. A shallow decoder gets better performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 401, + 186, + 510, + 257 + ], + "lines": [ + { + "bbox": [ + 401, + 186, + 510, + 257 + ], + "spans": [ + { + "bbox": [ + 401, + 186, + 510, + 257 + ], + "type": "table", + "html": "
Weight0-ShotI2T/T2I
136.551.7/32.1
0.144.369.2/45.9
0.0544.570.1/45.6
0.0143.270.6/45.6
", + "image_path": "098ca3f538339954e7552328387701cdac7e79c57b43e8021e9c014a3608c1b9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 279, + 530, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 279, + 530, + 311 + ], + "spans": [ + { + "bbox": [ + 45, + 279, + 530, + 311 + ], + "type": "text", + "content": "Table 6. MaskCLIP ablation experiments with YFCC-15M dataset. We report zero-shot(0-Shot), fine-tuning (FT), and linear probing (Lin) accuracy (%) for image-encoder-related ablation. And zero shot image-to-text, text-to-image retrieval (I2T/T2I) for text encoder-related ablations. Default settings are marked in gray ." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 317, + 279, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 317, + 279, + 410 + ], + "spans": [ + { + "bbox": [ + 45, + 317, + 279, + 410 + ], + "type": "text", + "content": "direction, resulting in poor performance. When we reduce the loss weight by " + }, + { + "bbox": [ + 45, + 317, + 279, + 410 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 45, + 317, + 279, + 410 + ], + "type": "text", + "content": ", the two additional tasks are helpful for the model and show a consistent gain on all the metrics. We suspect this is because the CLIP loss requires two different capabilities: understanding the input content and aligning them into a shared feature space. And the goal of the two additional self-supervised learning tasks is to facilitate understanding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 413, + 279, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 413, + 279, + 610 + ], + "spans": [ + { + "bbox": [ + 45, + 413, + 279, + 610 + ], + "type": "text", + "content": "Image & Text decoder depth. Then we study the influence of the decoder depth for both image and text decoders. As shown in Table.6c, we find the image decoder with only one layer works well, increasing the decoder depth leads to worse performance on all metrics. Similarly, Table.6d shows that the text branch benefits from a shallow decoder design. We argue that a too-deep decoder would make the encoder lazy, relying on the strong decoder to resolve the challenging mask feature/language modeling tasks. And the different depth choice between the image and text branches is caused by the framework difference: the image branch sees the mask tokens at the decoder, while the text branch takes the mask tokens as the encoder input. Note that if we remove the text decoder, the performance gets worse. We think this is largely caused by the output conflict that the global recognition feature aggregation and local word prediction are conducted at the same layer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 612, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 612, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 612, + 279, + 729 + ], + "type": "text", + "content": "Single-Stage v.s. two-Stage. Our MaskCLIP learns the VL contrastive and masked self-distillation simultaneously and jointly in a single stage. One possible variant is to first train CLIP and then use CLIP feature from the first stage to train masked image modeling as in [61, 62]. We report results on three datasets in Table 7. We can see that the second stage achieves better finetuning results compared with results from the stage one, showing the effectiveness of masked image modeling. Nonetheless, such two-stage training requires longer training time and loses the transfer capability in a zero" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 302, + 315, + 531, + 392 + ], + "blocks": [ + { + "bbox": [ + 381, + 259, + 527, + 277 + ], + "lines": [ + { + "bbox": [ + 381, + 259, + 527, + 277 + ], + "spans": [ + { + "bbox": [ + 381, + 259, + 527, + 277 + ], + "type": "text", + "content": "(f) MLM loss weight. A small loss weight works better." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 302, + 315, + 531, + 392 + ], + "lines": [ + { + "bbox": [ + 302, + 315, + 531, + 392 + ], + "spans": [ + { + "bbox": [ + 302, + 315, + 531, + 392 + ], + "type": "table", + "html": "
MethodEpochIN-1KFlicker30KADE20K
0-shotFTI2TT2I0-shotFT
Two-StageStage12537.682.352.932.87.247.8
Stage225—83.4———48.2
MaskCLIP2544.583.670.145.610.250.5
", + "image_path": "fc667e51881d97bc34b72dc1747a68706fc2333a0537b5221c6a87a0d2899df2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 394, + 531, + 415 + ], + "lines": [ + { + "bbox": [ + 297, + 394, + 531, + 415 + ], + "spans": [ + { + "bbox": [ + 297, + 394, + 531, + 415 + ], + "type": "text", + "content": "Table 7. Comparison between two-stage method and our single-stage MaskCLIP." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 297, + 423, + 530, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 423, + 530, + 447 + ], + "spans": [ + { + "bbox": [ + 297, + 423, + 530, + 447 + ], + "type": "text", + "content": "shot setting. In contrast, our MaskCLIP achieves superior results under all settings with fewer epochs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 297, + 451, + 367, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 451, + 367, + 463 + ], + "spans": [ + { + "bbox": [ + 297, + 451, + 367, + 463 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 468, + 531, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 468, + 531, + 584 + ], + "spans": [ + { + "bbox": [ + 296, + 468, + 531, + 584 + ], + "type": "text", + "content": "We present MaskCLIP, a new VL pretraining framework that incorporates masked self-distillation into VL contrastive. We point out that masked self-distillation learns local semantics, fitting nicely to the VL contrastive that aims to learn global semantics, and this is supported with comprehensively designed experiments. We also utilize mask language modeling to enhance the text encoder which is critical for zero-shot performance. The resulting visual encoder shows strong transfer capability across widely adopted benchmarks for linear probing, fine-tuning, and also zero-shot evaluation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 297, + 593, + 396, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 593, + 396, + 607 + ], + "spans": [ + { + "bbox": [ + 297, + 593, + 396, + 607 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 296, + 612, + 531, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 612, + 531, + 728 + ], + "spans": [ + { + "bbox": [ + 296, + 612, + 531, + 728 + ], + "type": "text", + "content": "This work was supported in part by the Natural Science Foundation of China under Grant U20B2047, 62072421, 62002334, 62102386, and 62121002. Key Research and Development program of Anhui Province under Grant 2022k07020008, Ant Group through CCF-Ant Innovative Research Program CCF-AFSG RF20210025, Alibaba Group through Alibaba Innovative Research Program. This work was also partly supported by Shenzhen Key Laboratory of Media Security, Shenzhen University, Shenzhen 518060, China." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "11002" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 125, + 280, + 729 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 51, + 125, + 279, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 125, + 279, + 145 + ], + "spans": [ + { + "bbox": [ + 51, + 125, + 279, + 145 + ], + "type": "text", + "content": "[1] https : / / eval . ai / web / challenges / challenge-page/1832/overview." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 148, + 280, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 148, + 280, + 168 + ], + "spans": [ + { + "bbox": [ + 51, + 148, + 280, + 168 + ], + "type": "text", + "content": "[2] https : / / eval . ai / web / challenges / challenge-page/1832/leaderboard/4298." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 170, + 280, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 170, + 280, + 212 + ], + "spans": [ + { + "bbox": [ + 51, + 170, + 280, + 212 + ], + "type": "text", + "content": "[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 214, + 280, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 214, + 280, + 245 + ], + "spans": [ + { + "bbox": [ + 51, + 214, + 280, + 245 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 248, + 280, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 248, + 280, + 300 + ], + "spans": [ + { + "bbox": [ + 51, + 248, + 280, + 300 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 303, + 280, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 303, + 280, + 355 + ], + "spans": [ + { + "bbox": [ + 51, + 303, + 280, + 355 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 357, + 280, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 357, + 280, + 410 + ], + "spans": [ + { + "bbox": [ + 51, + 357, + 280, + 410 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 411, + 280, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 411, + 280, + 463 + ], + "spans": [ + { + "bbox": [ + 51, + 411, + 280, + 463 + ], + "type": "text", + "content": "[8] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 466, + 280, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 466, + 280, + 508 + ], + "spans": [ + { + "bbox": [ + 51, + 466, + 280, + 508 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 510, + 280, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 280, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 280, + 551 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 554, + 280, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 280, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 280, + 605 + ], + "type": "text", + "content": "[11] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 608, + 280, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 280, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 280, + 639 + ], + "type": "text", + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 642, + 280, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 280, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 280, + 662 + ], + "type": "text", + "content": "[13] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 664, + 280, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 280, + 706 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 280, + 706 + ], + "type": "text", + "content": "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 708, + 280, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 708, + 280, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 708, + 280, + 729 + ], + "type": "text", + "content": "[15] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter:" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 531, + 729 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 317, + 107, + 531, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 531, + 138 + ], + "type": "text", + "content": "Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 140, + 531, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 531, + 193 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 531, + 193 + ], + "type": "text", + "content": "[16] Ruizhe Cheng, Bichen Wu, Peizhao Zhang, Peter Vajda, and Joseph E Gonzalez. Data-efficient language-supervised zero-shot learning with self-distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3119-3124, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 195, + 531, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 195, + 531, + 236 + ], + "spans": [ + { + "bbox": [ + 299, + 195, + 531, + 236 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 239, + 531, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 239, + 531, + 280 + ], + "spans": [ + { + "bbox": [ + 299, + 239, + 531, + 280 + ], + "type": "text", + "content": "[18] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11162-11173, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 282, + 531, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 282, + 531, + 324 + ], + "spans": [ + { + "bbox": [ + 299, + 282, + 531, + 324 + ], + "type": "text", + "content": "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 326, + 531, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 326, + 531, + 367 + ], + "spans": [ + { + "bbox": [ + 299, + 326, + 531, + 367 + ], + "type": "text", + "content": "[20] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 370, + 531, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 370, + 531, + 411 + ], + "spans": [ + { + "bbox": [ + 299, + 370, + 531, + 411 + ], + "type": "text", + "content": "[21] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 413, + 531, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 413, + 531, + 476 + ], + "spans": [ + { + "bbox": [ + 299, + 413, + 531, + 476 + ], + "type": "text", + "content": "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 479, + 531, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 479, + 531, + 520 + ], + "spans": [ + { + "bbox": [ + 299, + 479, + 531, + 520 + ], + "type": "text", + "content": "[23] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 522, + 531, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 522, + 531, + 564 + ], + "spans": [ + { + "bbox": [ + 299, + 522, + 531, + 564 + ], + "type": "text", + "content": "[24] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 566, + 531, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 566, + 531, + 629 + ], + "spans": [ + { + "bbox": [ + 299, + 566, + 531, + 629 + ], + "type": "text", + "content": "[25] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 631, + 531, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 631, + 531, + 662 + ], + "spans": [ + { + "bbox": [ + 299, + 631, + 531, + 662 + ], + "type": "text", + "content": "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 664, + 531, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 664, + 531, + 696 + ], + "spans": [ + { + "bbox": [ + 299, + 664, + 531, + 696 + ], + "type": "text", + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 698, + 531, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 698, + 531, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 698, + 531, + 729 + ], + "type": "text", + "content": "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "11003" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 280, + 730 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[29] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 280, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 280, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 280, + 204 + ], + "type": "text", + "content": "[30] Xinyu Huang, Youcai Zhang, Ying Cheng, Weiwei Tian, Ruiwei Zhao, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Xiaobo Zhang. Idea: Increasing text diversity via online multi-label recognition for vision-language pre-training. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4573-4583." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 206, + 280, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 280, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 280, + 268 + ], + "type": "text", + "content": "[31] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 271, + 280, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 280, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 280, + 313 + ], + "type": "text", + "content": "[32] Kyungyul Kim, ByeongMoon Ji, Doyoung Yoon, and Sangheum Hwang. Self-knowledge distillation: A simple way for better generalization. arXiv preprint arXiv:2006.12000, 3, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 314, + 280, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 280, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 280, + 367 + ], + "type": "text", + "content": "[33] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11336-11344, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 280, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 280, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 280, + 412 + ], + "type": "text", + "content": "[34] Juncheng Li, Xin He, Longhui Wei, Long Qian, Linchao Zhu, Lingxi Xie, Yueting Zhuang, Qi Tian, and Siliang Tang. Fine-grained semantically aligned vision-language pre-training. arXiv preprint arXiv:2208.02515, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 413, + 280, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 280, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 280, + 466 + ], + "type": "text", + "content": "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 467, + 280, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 280, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 280, + 510 + ], + "type": "text", + "content": "[36] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 511, + 280, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 280, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 280, + 554 + ], + "type": "text", + "content": "[37] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 555, + 280, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 280, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 280, + 598 + ], + "type": "text", + "content": "[38] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 599, + 280, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 280, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 280, + 652 + ], + "type": "text", + "content": "[39] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 654, + 280, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 280, + 707 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 280, + 707 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "type": "text", + "content": "[41] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Gener-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 531, + 730 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 317, + 107, + 531, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 531, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 531, + 129 + ], + "type": "text", + "content": "ative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 130, + 531, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 130, + 531, + 183 + ], + "spans": [ + { + "bbox": [ + 299, + 130, + 531, + 183 + ], + "type": "text", + "content": "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 184, + 531, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 184, + 531, + 238 + ], + "spans": [ + { + "bbox": [ + 299, + 184, + 531, + 238 + ], + "type": "text", + "content": "[43] Haoyu Lu, Nanyi Fei, Yuqi Huo, Yizhao Gao, Zhiwu Lu, and Ji-Rong Wen. Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15692-15701, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 239, + 531, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 239, + 531, + 281 + ], + "spans": [ + { + "bbox": [ + 299, + 239, + 531, + 281 + ], + "type": "text", + "content": "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 282, + 531, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 282, + 531, + 335 + ], + "spans": [ + { + "bbox": [ + 299, + 282, + 531, + 335 + ], + "type": "text", + "content": "[45] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10437-10446, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 337, + 531, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 337, + 531, + 389 + ], + "spans": [ + { + "bbox": [ + 299, + 337, + 531, + 389 + ], + "type": "text", + "content": "[46] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 391, + 531, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 391, + 531, + 424 + ], + "spans": [ + { + "bbox": [ + 299, + 391, + 531, + 424 + ], + "type": "text", + "content": "[47] Christian M Meyer and Iryna Gurevych. Wiktionary: A new rival for expert-built lexicons? Exploring the possibilities of collaborative lexicography. na, 2012." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 425, + 531, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 425, + 531, + 489 + ], + "spans": [ + { + "bbox": [ + 299, + 425, + 531, + 489 + ], + "type": "text", + "content": "[48] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 490, + 531, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 490, + 531, + 522 + ], + "spans": [ + { + "bbox": [ + 299, + 490, + 531, + 522 + ], + "type": "text", + "content": "[49] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. arXiv preprint arXiv:2112.12750, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 523, + 531, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 523, + 531, + 565 + ], + "spans": [ + { + "bbox": [ + 299, + 523, + 531, + 565 + ], + "type": "text", + "content": "[50] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 566, + 531, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 566, + 531, + 630 + ], + "spans": [ + { + "bbox": [ + 299, + 566, + 531, + 630 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 632, + 531, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 632, + 531, + 695 + ], + "spans": [ + { + "bbox": [ + 299, + 632, + 531, + 695 + ], + "type": "text", + "content": "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 697, + 531, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 697, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 697, + 531, + 730 + ], + "type": "text", + "content": "[53] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Anna Rohrbach, Zhe Gan, Lijuan Wang, Lu Yuan, et al. K-lite: Learning transfer" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "11004" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 280, + 730 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 65, + 106, + 279, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 106, + 279, + 127 + ], + "spans": [ + { + "bbox": [ + 65, + 106, + 279, + 127 + ], + "type": "text", + "content": "able visual models with external knowledge. arXiv preprint arXiv:2204.09222, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 130, + 280, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 130, + 280, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 280, + 172 + ], + "type": "text", + "content": "[54] Mustafa Shukor, Guillaume Couairon, and Matthieu Cord. Efficient vision-language pretraining with visual concepts and hierarchical alignment. arXiv preprint arXiv:2208.13628, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 175, + 280, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 280, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 280, + 216 + ], + "type": "text", + "content": "[55] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 218, + 280, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 280, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 280, + 250 + ], + "type": "text", + "content": "[56] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 252, + 280, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 280, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 280, + 295 + ], + "type": "text", + "content": "[57] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 298, + 280, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 280, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 280, + 340 + ], + "type": "text", + "content": "[58] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 341, + 280, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 280, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 280, + 394 + ], + "type": "text", + "content": "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 397, + 280, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 280, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 280, + 450 + ], + "type": "text", + "content": "[60] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 452, + 280, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 280, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 280, + 494 + ], + "type": "text", + "content": "[61] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 496, + 280, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 280, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 280, + 528 + ], + "type": "text", + "content": "[62] Longhui Wei, Lingxi Xie, Wengang Zhou, Houqiang Li, and Qi Tian. Mvp: Multimodality-guided visual pre-training. arXiv preprint arXiv:2203.05175, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 530, + 280, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 280, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 280, + 572 + ], + "type": "text", + "content": "[63] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 575, + 280, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 280, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 280, + 618 + ], + "type": "text", + "content": "[64] Haiyang Xu, Ming Yan, Chenliang Li, Bin Bi, Songfang Huang, Wenming Xiao, and Fei Huang. E2e-vlp: End-to-end vision-language pre-training enhanced by visual learning. arXiv preprint arXiv:2106.01804, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 619, + 280, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 280, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 280, + 661 + ], + "type": "text", + "content": "[65] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 664, + 280, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 280, + 706 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 280, + 706 + ], + "type": "text", + "content": "[66] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "type": "text", + "content": "[67] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 531, + 367 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 317, + 107, + 531, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 531, + 138 + ], + "type": "text", + "content": "similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 140, + 531, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 531, + 192 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 531, + 192 + ], + "type": "text", + "content": "[68] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 194, + 531, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 194, + 531, + 236 + ], + "spans": [ + { + "bbox": [ + 299, + 194, + 531, + 236 + ], + "type": "text", + "content": "[69] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 238, + 531, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 238, + 531, + 269 + ], + "spans": [ + { + "bbox": [ + 299, + 238, + 531, + 269 + ], + "type": "text", + "content": "[70] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 271, + 531, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 271, + 531, + 313 + ], + "spans": [ + { + "bbox": [ + 299, + 271, + 531, + 313 + ], + "type": "text", + "content": "[71] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 314, + 531, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 314, + 531, + 367 + ], + "spans": [ + { + "bbox": [ + 299, + 314, + 531, + 367 + ], + "type": "text", + "content": "[72] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "11005" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_content_list.json b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..68b33b2883f9c9b5e2515f7955534ddd1eaafe28 --- /dev/null +++ b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_content_list.json @@ -0,0 +1,1799 @@ +[ + { + "type": "text", + "text": "MaskCon: Masked Contrastive Learning for Coarse-Labelled Dataset", + "text_level": 1, + "bbox": [ + 130, + 130, + 836, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Feng, Ioannis Patras \nQueen Mary University of London, UK", + "bbox": [ + 326, + 181, + 640, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chen.feng, i.patras}@qmul.ac.uk", + "bbox": [ + 341, + 219, + 622, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has achieved great success in recent years with the aid of advanced neural network structures and large-scale human-annotated datasets. However, it is often costly and difficult to accurately and efficiently annotate large-scale datasets, especially for some specialized domains where fine-grained labels are required. In this setting, coarse labels are much easier to acquire as they do not require expert knowledge. In this work, we propose a contrastive learning method, called masked contrastive learning (MaskCon) to address the under-explored problem setting, where we learn with a coarse-labelled dataset in order to address a finer labelling problem. More specifically, within the contrastive learning framework, for each sample our method generates soft-labels with the aid of coarse labels against other samples and another augmented view of the sample in question. By contrast to self-supervised contrastive learning where only the sample's augmentations are considered hard positives, and in supervised contrastive learning where only samples with the same coarse labels are considered hard positives, we propose soft labels based on sample distances, that are masked by the coarse labels. This allows us to utilize both inter-sample relations and coarse labels. We demonstrate that our method can obtain as special cases many existing state-of-the-art works and that it provides tighter bounds on the generalization error. Experimentally, our method achieves significant improvement over the current state-of-the-art in various datasets, including CIFAR10, CIFAR100, ImageNet-1K, Stanford Online Products and Stanford Cars196 datasets. Code and annotations are available at https://github.com/MrChenFeng/MaskCon_CVPR2023.", + "bbox": [ + 76, + 301, + 473, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 799, + 209, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Supervised learning with deep neural networks has achieved great success in various computer vision tasks such as image classification, action detection and object localization. However, the success of supervised learning relies on large-scale and high-quality human-", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e24bf4d31d76eadea5bd76eebbc512b3bfd0c539e06bf653b132248009e9aaa.jpg", + "image_caption": [ + "Figure 1. Contrastive learning sample relations using MaskCon (ours) and other learning paradigms when only coarse labels are available. MaskCon are closer to the fine ones." + ], + "image_footnote": [], + "bbox": [ + 501, + 271, + 890, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "annotated datasets, whose annotations are time-consuming and labour-intensive to produce. To avoid such reliance, various learning frameworks have been proposed and investigated: Self-supervised learning aims to learn meaningful representations with heuristic proxy visual tasks, such as rotation prediction [16] and the more prevalent instance discrimination task, the latter, being widely applied in self-supervised contrastive learning framework; semi-supervised learning usually considers a dataset for which only a small part is annotated - within this setting, pseudo labelling methods [24] and consistency regularization techniques [1, 29] are typically used; Moreover, learning using more accessible but noisy data, such as web-crawled data, has also received increasing attention [13, 25].", + "bbox": [ + 496, + 507, + 892, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we consider an under-explored problem setting aiming at reducing the annotation effort - learning fine-grained representations with a coarsely-labelled dataset. Specifically, we learn with a dataset that is fully labeled, albeit at a coarser granularity than we are interested in (i.e., that of the test set). Compared to fine-grained labels, coarse labels are often significantly easier to obtain, especially in some of the more specialized domains, such as the recognition and classification of medical pathology images. As a simple example, for the task of differentiation between different pets, we need a knowledgeable cat lover to distinguish between 'British short' and 'Siamese',", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19913", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "but even a child annotator may help to discriminate between 'cat' and 'non-cat' (Fig. 1). Unfortunately, learning with a coarse labelled dataset has been less investigated compared to other weakly supervised learning paradigms. Recently, Bukchi et al. [2] investigate on learning with coarse labels in the few-shot setting. More closely related to us, Grafit [31] proposes a multi-task framework by a weighted combination of self-supervised contrastive learning and supervised contrastive learning cost; Similarly, CoIns [37] uses both a self-supervised contrastive learning cost and a supervised learning cross-entropy loss. Both works combine a fully supervised learning cost (cross entropy or contrastive) with a self-supervised contrastive loss – these works are the main ones with which we compare.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Differently than them, instead of using self-supervised contrastive learning as an auxiliary task, we propose a novel learning scheme, namely Masked Contrastive Learning (MaskCon). Our method aims to learn by considering inter-sample relations of each sample with other samples in the dataset. Specifically, we always consider the relation to oneself as confidently positive. To estimate the relations to other samples, we derive soft labels by contrasting an augmented view of the sample in question with other samples, and further improve it by utilizing the mask generated based on the coarse labels. Our approach generates soft inter-sample relations that can more accurately estimate fine inter-sample relations compared to the baseline methods (Fig. 1). Efficiently and effectively, our method achieves significant improvements over the state-of-the-art in various datasets, including CIFARtoy, CIFAR100, ImageNet-1K and more challenging fine-grained datasets Stanford Online Products and Stanford Cars196.", + "bbox": [ + 75, + 301, + 472, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 76, + 585, + 222, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first briefly review the contrastive learning works which are the basis of our method and then, we briefly introduce the state-of-the-art works in related problem settings.", + "bbox": [ + 75, + 611, + 468, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised contrastive learning Contrastive learning has recently emerged as a powerful method for representation learning without labels. In the context of self-supervision, contrastive learning has been used for instance discrimination, producing models whose performance rivals that of fully supervised training. The objective of contrastive instance discrimination is to minimize the distance between transformed views of the same sample in the feature space, while maximizing their distance from views of other samples. Notable methods in this area are SimCLR [5] and MoCo [18], as well as other variants including SwAV [4], which leverages clustering to identify prototypes for the contrastive learning objective and [33], that reformulates the instance discrimination contrastive objective in the", + "bbox": [ + 75, + 689, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "form of a triplet loss. These methods have been found to produce models that, requiring only limited amounts of annotated data for fine-tuning on a particular task or domain, achieve performance that reaches or even surpasses that of supervised models [6, 11]. Furthermore, self-supervised training has been found to compare positively with supervised training in other respects, such as producing models that perform better in the context of continual learning [14].", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, self-supervised contrastive learning based on instance discrimination task usually suffers from the 'false negative' problem – that samples from the same class should not be considered as negative. For example, pictures of different cats should not be considered completely negative to each other. To this end, heuristically modifying the inter-sample relations have been widely applied in self-supervised learning [10, 12, 22, 34], intuitively similar to us. Many of the state-of-the-art methods can be considered as special cases of our method by the adjustment of hyperparameters (Sec. 3.2).", + "bbox": [ + 496, + 212, + 892, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The success of self-supervised contrastive learning has led to increasing attention on how to integrate contrastive learning into existing supervised learning paradigms. One approach is supervised contrastive learning, which adapts contrastive learning to the fully supervised setting. Supervised contrastive learning was first introduced in [17, 28] under the name of neighbourhood components analysis (NCA). Recently, supervised contrastive learning has regained interest due to the significant progress made in [20] by employing more advanced network structures and image augmentations. Compared to standard supervised learning using cross-entropy loss, supervised contrastive learning exhibits superior performance in hard sample mining [20].", + "bbox": [ + 496, + 397, + 892, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other works have attempted to introduce self-supervised contrastive learning as an auxiliary task, especially for weakly supervised learning problems. Gidaris et al. [15] jointly train a supervised classifier and a rotation-prediction objective. Komodakis and Gidaris [21] seeking to improve model performance on few-shot learning. Wei et al. [35] propose a label-filtered instance discrimination objective for the pre-training of models to be used for transfer learning in other datasets. S4 [38] uses rotation and exemplar [9] prediction to improve performance in semi-supervised learning tasks. Feng et al. [13] apply a self-supervised contrastive loss to safely learn meaningful representations in the presence of noisy labels.", + "bbox": [ + 496, + 609, + 893, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hierarchical image classification & Unsupervised clustering Hierarchical image classification methods aim to learn better representations with datasets having hierarchically structured labels. To utilize information from different levels of labels, Zhang et al. [39] propose a multi-losses", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19914", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "contrastive learning framework with each loss considering a specific level of labels. Unsupervised clustering methods often have the same problem setting but different goals than self-supervised learning, that is, to recover and identify the ground-truth semantic labels of each sample [3, 32]. Deep-Cluster [3] simultaneously learns the clustering assignments and a parametric classifier, by feeding the clustering assignments as pseudo labels to the classifier. SCAN [32] inherits the idea of DeepClustering while augmenting extra self-labelling and neighbor mining techniques.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 255, + 210, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider the problem of learning when the labels of the training data and the labels of the test data are inconsistent in granularity, and in particular when the labels of the training set are coarser. More specifically, let us denote with $\\mathbf{X} = \\{\\pmb{x}_i\\}_{i=1}^N$ an i.i.d sampled train dataset with annotated coarse labels $\\mathbf{Y} = \\{\\pmb{y}_i \\in \\{0,1\\}^M\\}_{i=1}^N$ subject to $\\sum_{m=1}^{M} \\pmb{y}_{im} = 1$ . Here, $N$ denotes the number of samples and $M$ denotes the number of coarse classes. Let us also define a finely labelled set $\\mathbf{Y}' = \\{\\pmb{y}_i' \\in \\{0,1\\}^{M'}\\}_{i=1}^N$ subject to $\\sum_{k=1}^{M'} \\pmb{y}_i^k = 1$ , where $M'$ denotes the number of fine classes. Typically, $M' > M$ .", + "bbox": [ + 75, + 280, + 472, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 460, + 217, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first quickly review the common learning paradigms utilized in the baseline methods and our method.", + "bbox": [ + 76, + 484, + 468, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Supervised learning", + "text_level": 1, + "bbox": [ + 76, + 532, + 271, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In most supervised learning frameworks, the objective is the minimization of the below empirical risk:", + "bbox": [ + 76, + 555, + 468, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR (f, g) = \\sum_ {i = 1} ^ {N} L _ {c e} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {y} _ {i}; f, g\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 595, + 468, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ denotes the feature encoder, and $g$ denotes the classifier head (usually a single fully-connected layer).", + "bbox": [ + 76, + 647, + 468, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For brevity, we denote $\\pmb{f}_i \\triangleq f(\\pmb{x}_i)$ , $\\pmb{g}_i \\triangleq g(f(\\pmb{x}_i))$ and $\\pmb{p}_i \\triangleq \\mathrm{softmax}(\\pmb{g}_i)$ as the feature, logit and prediction of a sample view $\\pmb{x}_i$ respectively. Here,", + "bbox": [ + 76, + 676, + 468, + 722 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c e} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {y} _ {i}; f, g\\right) = - \\sum_ {m = 1} ^ {M} \\boldsymbol {y} _ {i} ^ {m} \\log \\boldsymbol {p} _ {i} ^ {m} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 733, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is the cross-entropy loss function - the default loss function for classification problems.", + "bbox": [ + 76, + 782, + 468, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Contrastive learning", + "text_level": 1, + "bbox": [ + 76, + 830, + 274, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the common supervised learning model above where a parametric classifier $g$ is learned based on the semantic labels $Y$ of the samples $X$ , we can also perform contrastive", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learning based on the inter-sample relations $Z = \\{z_{i} \\in (0,1)^{N}\\}_{i=1}^{N}$ , with each entry $z_{ij}$ depicting the inter-sample relation between $x_{i}$ and $x_{j}$ . Intuitively, $z_{ij} = 1$ means that sample $x_{i}$ and $x_{j}$ generate a strong positive pair. Since each sample may form multiple positive sample pairs, for brevity, we abuse the notation here with $Z$ denoting also the sample-wise normalized inter-sample relations. To learn such intersample relations, instead of a parametric classifier $g$ , the $f$ encoder is usually followed by a projector $h$ , which is often implemented as an MLP and learned by regularizing the inter-sample relations $Z$ (Eq. (4)).", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, let us denote by $\\pmb{h}_i \\triangleq h(f(\\pmb{x}_i))$ the projection. We first calculate the cosine similarity $\\pmb{d}_i$ between a sample $\\pmb{x}_i$ and the dataset $\\pmb{H} = \\{\\pmb{h}_n\\}_{n=1}^N$ :", + "bbox": [ + 496, + 256, + 890, + 304 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {d} _ {i} = \\left[ \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {1}\\right), \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {2}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {N}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 310, + 890, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us further define $\\pmb{q}_i \\triangleq \\mathrm{softmax}(\\pmb{d}_i / \\tau_0)$ , where $\\tau_0$ is the temperature hyperparameter. Then the following empirical risk will be optimized:", + "bbox": [ + 496, + 334, + 890, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR (f, h) = \\sum_ {i = 1} ^ {N} L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {z} _ {i}; f, h\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 386, + 890, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the contrastive loss $L_{con}$ is defined as follows:", + "bbox": [ + 500, + 434, + 849, + 449 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o n} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {z} _ {i}; f, h\\right) = - \\sum_ {n = 1} ^ {N} \\boldsymbol {z} _ {i} ^ {n} \\log \\boldsymbol {q} _ {i} ^ {n}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 455, + 890, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised contrastive learning We first introduce the most prevalent form of contrastive learning currently — self-supervised contrastive learning for learning without labels. Since there are no annotated labels, we usually set $Z^{self}$ by considering each sample having only itself as positive:", + "bbox": [ + 496, + 502, + 890, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i j} ^ {\\text {s e l f}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} i = j \\\\ 0, & \\text {i f} i \\neq j \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 590, + 890, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typically, we aim at maximizing the relations between different augmented views of the same sample while minimizing the relations between augmented views of different samples. This is also widely known as instance discrimination task. We denote such self-supervised contrastive loss as $L_{selfcon}$ .", + "bbox": [ + 496, + 625, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised contrastive learning We can easily transition from self-supervised contrastive learning to supervised contrastive learning, by simply changing the inter-sample relations $Z^{sup}$ for Eq. (4) and Eq. (5). More specifically, as in [17, 20, 28], samples from the same semantic classes are considered positive pairs,", + "bbox": [ + 496, + 715, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i j} ^ {\\text {s u p}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\\\ 0, & \\text {i f} \\boldsymbol {y} _ {i} \\neq \\boldsymbol {y} _ {j} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 811, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in addition to the sample itself (and its augmented views) as in self-supervised contrastive learning. We denote the supervised contrastive loss as $L_{supcon}$ .", + "bbox": [ + 496, + 851, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19915", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Memory bank For consistency with the supervised cross-entropy loss and for clarity, we present all the essential formulas above in terms of the entire dataset $\\mathbf{X}$ . However, due to computation and memory constraints, it is often unrealistic to contrast a specific sample $\\mathbf{x}_i$ with the whole dataset. In the actual implementation in this work, following MoCo [18], we use a dynamic FIFO memory bank $\\mathbf{H} = \\{\\mathbf{h}_p\\}_{p=1}^P$ consisting of cached projections of $P$ samples. For each sample $\\mathbf{x}_i$ , we generate two random augmented views as $\\mathbf{x}_q$ (query) and $\\mathbf{x}_k$ (key) with their corresponding projections denoted by $\\mathbf{h}_q$ and $\\mathbf{h}_k$ , respectively. Then, we calculate the cosine similarity between $\\mathbf{h}_q$ on the one hand, and $\\mathbf{h}_k$ and each projection in the memory bank on the other:", + "bbox": [ + 78, + 89, + 467, + 300 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {d} _ {i} = \\overbrace {\\left[ \\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {k}\\right) , \\overbrace {\\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {1}\\right) , \\dots , \\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {P}\\right)} ^ {\\text {s e l f}} \\right]} ^ {\\text {m e m o r y b a n k}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 309, + 465, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For more details, please refer to the MoCo paper [18].", + "bbox": [ + 80, + 357, + 442, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 Baseline methods", + "text_level": 1, + "bbox": [ + 76, + 398, + 253, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clearly, with respect to the finer target question labels, each sample $\\boldsymbol{x}_i$ , is a trustworthy positive sample in relation to itself (i.e., they have the same fine-grained label). However, considering the samples with the same coarse label as positives may lead to under-clustering issues (Fig. 3) (they may have different fine-grained labels). Recent works attempt to address such under-clustering by combining an additional self-supervised contrastive loss with a specific supervised loss that utilizes coarse labels. Grafit [31] combine it with the supervised contrastive loss and CoIns [37] with the supervised cross-entropy loss. In both cases, as a result, the self-supervised contrastive loss that considers each sample as a class by itself, aims to mitigate the tendency for under-clustering. Formally, given our definitions of the previously mentioned loss functions in Sec. 3.1 ( $L_{ce}$ , $L_{selfcon}$ and $L_{supcon}$ ) it turns out that the losses used by Grafit [31] and CoIns [37] can be expressed as follows:", + "bbox": [ + 76, + 422, + 468, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {G r a f i t} = w L _ {s u p c o n} + (1 - w) L _ {s e l f c o n} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 694, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C o I n s} = w L _ {c e} + (1 - w) L _ {s e l f c o n} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 714, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w$ controls the relative weight of each loss.", + "bbox": [ + 76, + 743, + 403, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. MaskCon: Masked Contrastive learning", + "text_level": 1, + "bbox": [ + 76, + 770, + 423, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of equally weighing all the samples within the same coarse classes, we aim to emphasize the samples with the same fine labels and reduce the importance of the others. To achieve this, we introduce a novel contrastive learning method, namely Masked Contrastive learning (MaskCon), within the framework of contrastive learning that utilizes inter-sample relations directly.", + "bbox": [ + 76, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, for sample $\\boldsymbol{x}_i$ , we estimate its intersample relations $\\boldsymbol{z}_i'$ to other samples utilizing the key view projection $\\boldsymbol{h}_k$ and the whole dataset $\\{\\boldsymbol{h}_1, \\dots, \\boldsymbol{h}_N\\}$ excluding itself (since it will always be considered as a trustworthy positive), as below:", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i j} ^ {\\prime} = \\frac {\\mathbb {1} \\left(\\boldsymbol {y} _ {j} = \\boldsymbol {y} _ {i}\\right) \\cdot \\exp \\left(d _ {i j} ^ {\\prime} / \\tau\\right)}{\\sum_ {n = 1 , n \\neq i} ^ {N} \\mathbb {1} \\left(\\boldsymbol {y} _ {n} = \\boldsymbol {y} _ {i}\\right) \\cdot \\exp \\left(d _ {i n} ^ {\\prime} / \\tau\\right)}, i \\neq j, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 176, + 890, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the similarity $d_i^{\\prime}$ is given by", + "bbox": [ + 500, + 228, + 730, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {d} _ {i} ^ {\\prime} = \\left[ \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {1}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {i - 1}\\right), \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 253, + 890, + 280 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {i + 1}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {N}\\right) \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 277, + 823, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Please note the use of the mask $(\\mathbb{1}(\\pmb {y}_j = \\pmb {y}_i))$ that excludes from the softmax that estimates inter-sample relationships, the samples $j$ that have a different coarse label with the sample $i$ (and sets their $z_{ij}^{\\prime}$ to 0). While it is risky to consider all samples from the same coarse class as positive, we can confidently identify those samples that do not have the same coarse class as negative. This reduces the noise in $z_i^\\prime$ . Finally, we re-scale the $z_{i}^{\\prime}$ with its maximum", + "bbox": [ + 498, + 305, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i j} ^ {\\prime} = z _ {i j} ^ {\\prime} / \\max \\left(\\boldsymbol {z} _ {i} ^ {\\prime}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 438, + 890, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to make the closest neighbour as positive as the sample itself and arrive at:", + "bbox": [ + 498, + 467, + 890, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i j} ^ {\\text {m a s k}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} i = j \\\\ z _ {i j} ^ {\\prime}, & \\text {i f} i \\neq j \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 507, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to $Z^{supcon}$ , we thus reweight the samples of the same coarse label according to the similarities in the feature space.", + "bbox": [ + 498, + 550, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the masked contrastive loss as $L_{maskcon}$ and, similarly to Grafit and CoIns, we also consider a weighted loss as the final objective:", + "bbox": [ + 498, + 595, + 890, + 641 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = w L _ {\\text {m a s k c o n}} + (1 - w) L _ {\\text {s e l f c o n}} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 654, + 890, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relations to SOTA works By adjusting $w$ and $\\tau$ , our method can obtain various existing SOTA methods as special cases. More specifically, by setting $\\tau$ to $\\infty$ , our method degenerates to Grafit [31]. Ignoring the mask, i.e., treating all samples as having the same coarse label, our method can also generalize to existing SOTAs in self-supervised learning. For example, NNCLR [10] instead of considering only each sample itself as positive it considers as positive the nearest neighbors as well. Formally, by setting $w = 1$ and $\\tau = 0$ , MaskCon degenerates to NNCLR; ASCL [12] proposes to introduce extra positive samples by calculating soft inter-sample relations, adaptively weighted by its own normalized entropy. By setting $w = \\text{norm\\_entropy}(z_i')$ , MaskCon degenerates to ASCL.", + "bbox": [ + 498, + 689, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19916", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Theoretical justification of MaskCon", + "text_level": 1, + "bbox": [ + 76, + 90, + 395, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The fundamental goal of contrastive learning is to identify $Z$ that suits the problem granularity and accurately represents the relationship between samples. Assuming that there exists an optimal hidden $\\hat{Z}$ , we can define the expected population risk in terms of $\\hat{Z}$ as shown in Eq. (16).", + "bbox": [ + 76, + 113, + 468, + 190 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {E} (f, h) = E _ {\\boldsymbol {x} _ {i}} \\left[ L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {z}} _ {i}; f, h\\right) \\right] \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 204, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, let us define the empirical risk with respect to $\\hat{Z}$ as shown in Eq. (17).", + "bbox": [ + 76, + 226, + 468, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {R} (f, h) = \\sum_ {i = 1} ^ {N} L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {z}} _ {i}; f, h\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 266, + 468, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "THEOREM 1. The generalization error bounds of contrastive learning depends on the difference between the inter-sample relations $\\mathbf{Z}$ and the optimal hidden intersample relations $\\hat{\\mathbf{Z}}$ , i.e.,", + "bbox": [ + 76, + 316, + 468, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {z} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| \\boldsymbol {z} _ {i} - \\hat {\\boldsymbol {z}} _ {i} \\| _ {2} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 387, + 468, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. Let us assume $L_{con}(\\pmb{x}_i, \\pmb{z}_i; f, h) \\in [a, b]$ and is $\\lambda$ -Lipschitz continuous w.r.t $\\pmb{z}_i$ . Let $|\\mathcal{F}|$ and $|\\mathcal{H}|$ be the covering numbers that correspond to the finite hypothesis space of $f$ and $g$ . We have:", + "bbox": [ + 76, + 434, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| R _ {E} (f, h) - R (f, h) \\right| \\leq \\overbrace {\\left| R _ {E} (f , h) - \\hat {R} (f , h) \\right|} ^ {\\text {H o e f f d i n g ' s i n e q u a l i t y}} \\\\ L i p s c h i t z \\quad c o n t i n u o u s \\\\ + \\overbrace {| \\hat {R} (f , h) - R (f , h) |} \\\\ \\leq | a - b | \\sqrt {\\frac {\\log (2 | \\mathcal {F} | \\cdot | \\mathcal {H} | / \\delta)}{2 n}} \\\\ + \\lambda \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| z _ {i} - \\hat {z} _ {i} \\| _ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 503, + 444, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with probability at least $1 - \\delta$", + "bbox": [ + 76, + 671, + 279, + 686 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/08164868a0cfd8e4ddc80431a588496722dbb4323ac57bb5138258b55e39bdf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 672, + 468, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on Theorem 1, it can be concluded that our method performs better than Grafit with high probability, because an appropriate temperature $\\tau$ allows $Z^{mask}$ to more accurately estimate the optimal $\\hat{Z}$ (Fig. 1) in comparison to assigning equal weights to all coarse neighbours.", + "bbox": [ + 76, + 696, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 784, + 207, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, extensive experiments are conducted to validate the findings and effectiveness of the proposed Masked Contrastive Learning (MaskCon) framework. The impact of hyperparameters is thoroughly investigated in Sec. 4.1. In Sec. 4.2, Sec. 4.3 and Sec. 4.4, results are provided for experiments on CIFAR datasets and ImageNet-1K", + "bbox": [ + 76, + 810, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset with coarse labels. In Sec. 4.5, we apply our method on more challenging fine-grained datasets, including Stanford Online Products (SOP) dataset [30] and Stanford Cars 196 dataset [23].", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure the robustness of the proposed method and exclude the impact of model capacity, the model settings are kept consistent throughout all the experiments, except for the two hyperparameters, $w$ and $\\tau$ , which are ablated in Sec. 4.1. ResNet18 [19] is employed in all experiments, with the initial convolutional layer modified to have a kernel size of $3 \\times 3$ and a stride of 1, and the initial max-pooling operator removed for CIFAR experiments, to account for the smaller image size $(32 \\times 32)$ as suggested in prior works [18]. More implementation and dataset details can be found in APPENDIX A.", + "bbox": [ + 496, + 152, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with two competing methods: Grafit and CoIns. For a fair comparison, we exhaust the weight $w$ choices for both methods and report the best achievable results in all experiments. Note that when $w = 0$ , Grafit and CoIns degenerate to self-supervised contrastive learning denoted as SelfCon; Conversely, when $w = 1$ , Grafit degenerates to supervised contrastive learning [20] denoted as SupCon, while CoIns degenerates to conventional supervised cross-entropy learning denoted as SupCE. For reference, we also show the results when training with fine labels – this is denoted as SupFINE.", + "bbox": [ + 496, + 316, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation protocol To evaluate the different methods on the test set with fine labels, we use the recall@K [27] metric widely used in the image retrieval task. Each test image first retrieves top-K nearest neighbours from the test set and receives 1 if there exists at least one image from the same fine class among the top-K nearest neighbours, otherwise 0. Recall@K averages this score over all the test images.", + "bbox": [ + 498, + 501, + 890, + 607 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35584c21fd18971c3edb37faf03fd9f1e8bcf6ac7342541c0ea33460bf31a10e.jpg", + "image_caption": [ + "Figure 2. Recall@1 w.r.t $w$ and $\\tau$ on CIFAR100 dataset" + ], + "image_footnote": [], + "bbox": [ + 531, + 635, + 836, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Effect of $w$ and $\\tau$", + "text_level": 1, + "bbox": [ + 500, + 847, + 665, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we extensively ablate the effect of hyperparameters of MaskCon - weight $w$ and temperature $\\tau$ .", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19917", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0afe7e2d18515a77a3448666a4a616b93b2d5fcce67f0e9c021a9e7ee1a218a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFARtoy-goodsplitCIFARtoy-badsplit
Recall@1Recall@2Recall@5Recall@10Recall@1Recall@2Recall@5Recall@10
SelfCon84.8391.5596.3598.1684.8391.5596.3598.16
Graft86.6192.3397.0198.3889.9694.3697.6198.10
SupCon73.8484.2592.1495.4684.6690.9395.1596.71
CoIns86.1592.7697.2198.4690.5594.9497.7398.71
SupCE76.3085.2694.6597.4687.1592.8596.7898.34
SupFINE94.1196.5398.2598.9694.1196.5398.2598.96
MaskCon (Ours)90.28 (13.98↑)94.0497.3398.5391.56 (4.41↑)95.2397.7098.70
", + "bbox": [ + 153, + 88, + 818, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results on CIFARtoy dataset.", + "bbox": [ + 369, + 234, + 598, + 248 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/178b97c5e77a93fea5399a99b0608387bcab4cb3c812a2245fb34c12f6b6f53a.jpg", + "image_caption": [ + "Figure 3. t-SNE visualization of learned representation on CIFARtoy dataset." + ], + "image_footnote": [], + "bbox": [ + 174, + 263, + 310, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1567f0ddfea23c34bd0014d92139a0f7445dc4f315629ce625b5ea2e0366b730.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 263, + 473, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d066dcae930ba7b28f5ef8fc1f1e1ed82805a53a770e66fd9ea74eb5f5a851d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 263, + 633, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/49148ea15d2793ba7d457ca504516144ea116adf981bdf6e85964ed110c5d239.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 263, + 795, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we show the recall@1 results on CIFAR100 dataset in Fig. 2, with $w = \\{0, 0.2, 0.5, 0.8, 1.0\\}$ and $\\tau = \\{0, 0.01, 0.05, 0.1, 0.5, \\infty\\}$ .", + "bbox": [ + 76, + 421, + 468, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results demonstrate a significant improvement in performance upon the inclusion of additional inter-sample relationships ( $w \\neq 0$ ) in comparison to self-supervised learning alone ( $w = 0$ ) on the CIFAR dataset. Furthermore, a suitable temperature $\\tau$ , such as 0.05 or 0.1, consistently yields superior results compared to the simple weighted combination in Grafit ( $\\tau = \\infty$ ). In general, we recommend initiating the hyperparameter search with $w = 1$ and $\\tau = \\tau_0$ , where $\\tau_0$ denotes the temperature for the projection head. Additional results concerning various hyperparameters on other datasets can be found in APPENDIX B.", + "bbox": [ + 75, + 468, + 468, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experiments on CIFARtoy dataset", + "text_level": 1, + "bbox": [ + 76, + 647, + 377, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To simulate the coarse labelling process, we manually generate toy datasets based on the CIFAR10 dataset. Following the problem motivation, a subset of 8 classes is selected from the original 10 classes. Specifically, the classes 'airplane', 'truck', 'automobile', and 'ship' are designated as Class A, while 'horse', 'dog', 'bird', and 'cat' are assigned to Class B. Note that Class A comprises non-organic objects, while Class B comprises animals. Moreover, a bad split of the aforementioned dataset is defined, wherein 'airplane', 'automobile', 'bird', and 'cat' are Class A, and 'horse', 'dog', 'ship', and 'truck' are Class B.", + "bbox": [ + 75, + 672, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1, our method achieves the best performance in both splits. Moreover, the improvement over the supervised learning with coarse labels (SupCE) on good split $(13.98\\%)$ is much higher than the bad split $(4.41\\%)$ ,", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which shows the substantial potential of MaskCon in dealing with more realistic coarse-labelled datasets1.", + "bbox": [ + 496, + 421, + 888, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 3 we visualize the learned features of all test samples in the good split. Clearly, the learned representations with supervised contrastive learning (SupCon) and self-supervised contrastive learning (SelfCon) tend to under-cluster and over-cluster the samples, respectively. By contrast, our method gets more compact and clear clusters, in line with the results when trained with fine labels (SupFINE).", + "bbox": [ + 496, + 450, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experiments on CIFAR100 dataset", + "text_level": 1, + "bbox": [ + 498, + 580, + 803, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The common CIFAR100 dataset has 20 classes of coarse labels in addition to the 100 classes of fine labels, with each coarse class containing five fine-grained classes (500 samples). The results in Tab. 2 show that our method achieves significant improvements over the SOTAs. In particular, it improves the top-1 retrieval precision from $47.25\\%$ to $65.52\\%$ , approaching the results by the model learned with fine labels $(71.13\\%)$ .", + "bbox": [ + 496, + 603, + 890, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Experiments on ImageNet-1K dataset", + "text_level": 1, + "bbox": [ + 498, + 732, + 823, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our method on the largescale ImageNet-1K dataset. For efficiency, we experiment with the downsampled version of ImageNet-1K dataset [7], where each sample was resized to $32 \\times 32$ . Since no official coarse labels exist for ImageNet, we introduce coarse", + "bbox": [ + 496, + 755, + 890, + 830 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1It is worth noting that the better performance on the bad split may seem counter-intuitive at first glance. However, note that a bad split actually makes the labels more informative, as it explicitly helps to discriminate visually similar classes, e.g., 'cat' and 'dog' in different coarse classes in our setting.", + "bbox": [ + 496, + 838, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19918", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/69cd1925bd0f42b498b6753bb87898584904bb79211376c727ddbc32204c22bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon40.5051.8366.2376.66
Graft60.5771.1382.3289.21
SupCon58.6570.0482.1889.09
CoIns60.1070.8983.1489.52
SupCE47.2561.2477.7887.01
SupFINE71.1380.0387.6191.59
MaskCon (Ours)65.52 (18.17†)74.4683.6489.25
", + "bbox": [ + 89, + 88, + 457, + 203 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "labels based on the WordNet [26] hierarchy so as to artificially group the whole dataset into 12 coarse classes: '0: Invertebrate', '1: Domestic animal', '2: Bird', '3: Mammal', '4: Reptile/Aquatic vertebrate', '5: Device', '6: Vehicle', '7: Container', '8: Instrument', '9: Artifact', '10: Clothing' and '11: Others'. In Tab. 3, we show the results on down-sampled ImageNet-1K datasets.", + "bbox": [ + 75, + 251, + 468, + 358 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f82298af1f6504929ab3f92730623ae46739d512d9e2509ae16344d2c9f1fa03.jpg", + "table_caption": [ + "Table 2. Results on CIFAR100 dataset." + ], + "table_footnote": [], + "table_body": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon10.2814.1522.3630.34
Graft18.1325.4637.1946.64
SupCon13.3619.4029.7739.38
CoIns18.3625.5437.0946.89
SupCE12.2314.1527.7637.03
SupFINE33.9744.5557.2365.77
MaskCon (Ours)19.08 (6.86↑)26.2138.1747.96
", + "bbox": [ + 89, + 369, + 457, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can again validate that with temperature-controlled soft relations, our method surpasses SupCE, Grafit and CoIns consistently, especially for top-10 retrieval results.", + "bbox": [ + 75, + 525, + 468, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Experiments on fine-grained datasets", + "text_level": 1, + "bbox": [ + 76, + 579, + 398, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct experiments in a more challenging scenario - fine-grained datasets with only coarse labels. Please note, that in this work, we do not aim to compare with the state-of-the-art works in fine-grained classification, which usually involve more specialized techniques, such as object localization and local feature extraction.", + "bbox": [ + 75, + 602, + 468, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.1 Stanford Online Products (SOP)", + "text_level": 1, + "bbox": [ + 76, + 710, + 356, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Stanford Online Products (SOP) dataset [30] consists of 22,634 products, with each product having between two and twelve photos from different perspectives, for a total of 120,053 images. In addition, there are 12 coarse classes based on the semantic categories of the products, such as 'bicycle' and 'kettle'. For the common image retrieval task, the category of the test set is possibly unknown, so we firstly extract a subset with an unknown test category based on the SOP dataset. We then selected the categories with eight or more images. We then split almost equally this subset to obtain the training set and the test set with 25,368 and 25,278", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images, consisting of 2,517 and 2,518 classes respectively, denoted as SOP-split1. Moreover, we select all products with twelve images, and then randomly select ten train images and two test images. Thus, we have 1,498 classes with a total of 17,976 images (2,996 test images and 14,980 train images), denoted as SOP-split2.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cf01f81c30bb82b660213804d0f3cef27cdc33a87d9ff62699ced4658c2a2ed7.jpg", + "table_caption": [ + "Table 3. Results on ImageNet-1K dataset." + ], + "table_footnote": [], + "table_body": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon70.3675.5781.5385.13
Graft74.0278.8284.1387.91
SupCon53.6959.5567.1272.78
CoIns70.8476.0182.286.08
SupCE36.3542.3950.3056.52
SupFINE83.9488.0491.9594.00
MaskCon (Ours)74.05 (37.7↑)78.9784.4887.96
", + "bbox": [ + 511, + 194, + 879, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d4d784ab57b83358ac627e5008d0a5d99a5587b9b9be01813bea92bbd18db312.jpg", + "table_caption": [ + "Table 4. Results on SOP-split1 dataset." + ], + "table_footnote": [], + "table_body": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon35.8541.4649.7756.11
Graft39.1244.6653.1059.65
SupCon25.0729.2435.8541.59
CoIns38.2245.1954.3761.18
SupCE22.5626.3433.2838.95
SupFINE69.5675.7083.2487.88
MaskCon (Ours)45.36 (22.8†)51.0758.9165.52
", + "bbox": [ + 511, + 364, + 879, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results on SOP-split2 dataset.", + "bbox": [ + 576, + 492, + 812, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 4 and Tab. 5 we show results on SOP-split1 and SOP-split2, respectively. Similarly, regardless of whether the test class is known or not, our method still achieves significant improvements over state-of-the-art methods. An interesting phenomenon here is, that on the SOP dataset, SelfCon alone is significantly better than its SupCE/SupCon counterpart. We conjecture that this is due to the fact that each fine class of the SOP dataset in fact consists of different views of the same product, and the number of images per fine class is quite small (2 to 12). For such a dataset, the classification of fine classes is closer to instance discrimination rather than the coarse classes classification.", + "bbox": [ + 496, + 523, + 890, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.2 Stanford Cars196", + "text_level": 1, + "bbox": [ + 500, + 724, + 678, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stanford Cars196 is another widely used fine-grained image classification benchmark. As there are no official annotated coarse labels, we manually group 196 car models into 8 coarse classes based on the common types of cars: '0: Cab', '1: Sedan', '2: SUV', '3: Convertible', '4: Coupe', '5: Hatchback', '6: Wagon' and '7: Van'.", + "bbox": [ + 496, + 750, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 6, our method again has the best performance. In addition, unlike in the SOP datasets, SupCon and SupCE have an overwhelming advantage over SelfCon. In this dataset, this is not surprising because for each fine class,", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19919", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58813795238c845a958de7eb7661293ea55d43e8983f551fd6346fa83bd9390b.jpg", + "image_caption": [ + "Figure 4. Examples of Top-10 image retrieval results on Cars196 dataset." + ], + "image_footnote": [], + "bbox": [ + 140, + 92, + 825, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "there are more images (ranging from 24 images to 68 images) with different colours/backgrounds in the cars dataset, leading to a much higher intra-class variance. Some image retrieval examples can be found in Fig. 4.", + "bbox": [ + 75, + 500, + 468, + 560 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f33cc3c655b9d9fc392d6519d64c9dc8274d25ee28df31471457defb30d0773e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon20.9728.7541.4452.61
Graft42.3054.7971.181.74
SupCon42.3054.7971.181.74
CoIns42.7755.6072.2982.53
SupCE42.7755.6072.2982.53
SupFINE78.0982.9786.5188.04
MaskCon (Ours)45.53 (2.76†)58.5674.3684.36
", + "bbox": [ + 89, + 574, + 457, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Results on Stanford Cars196 dataset.", + "bbox": [ + 135, + 700, + 408, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.3 Additional discussion", + "text_level": 1, + "bbox": [ + 76, + 739, + 279, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared to the state-of-the-art, our method achieves the best results in all experiments. We would like to note here some of our insights from the results on the CIFAR and the fine-grained ones (including ImageNet-1K). More specifically, we note that on the CIFAR datasets the performance of our method approaches that of the supervised one with fine labels (SupFINE), while on the fine-grained datasets there is a larger gap. We note that the critical question is to consider whether the instance discrimination task is", + "bbox": [ + 75, + 763, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "useful for fine-grained classification in different scenarios, since coarse labelling usually does not lead to worse results. Towards that, recent works have shown that the core idea of self-supervised contrastive learning — augmentation invariance may be destructive for fine-grained tasks [8, 36]. For example, the commonly applied colour distortion augmentation will promote the model to be non-sensitive to colour information. However, colour may be the key to discriminate between different breeds of birds. How to adapt the instance discrimination task to the fine-grained coarse-labelled dataset is a future direction we wish to pursue.", + "bbox": [ + 496, + 500, + 890, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 679, + 617, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a Masked Contrastive learning framework (MaskCon) for learning fine-grained information with coarse-labelled datasets. On the basis of two baseline methods, we utilize coarse labels and the instance discrimination task to better estimate inter-sample relations. We show theoretically that our method can reduce the optimization error bound. Extensive experiments with various hyperparameter settings on multiple benchmarks, including the CIFAR datasets and the more challenging fine-grained classification datasets show that our method achieves consistent and large improvement over the baselines.", + "bbox": [ + 496, + 704, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This work was supported by the EU H2020 AI4Media No. 951911 project.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19920", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1", + "[2] Guy Bukchin, Eli Schwartz, Kate Saenko, Ori Shahar, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. Fine-grained angular contrastive learning with coarse labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8730-8740, June 2021. 2", + "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3", + "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2", + "[7] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. arXiv preprint arXiv:1707.08819, 2017. 6", + "[8] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 8", + "[9] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Ried-miller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. Advances in neural information processing systems, 27, 2014. 2", + "[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. 2, 4", + "[11] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5423, 2021. 2", + "[12] Chen Feng and Ioannis Patras. Adaptive soft contrastive learning. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 2721-2727, 2022. 2, 4", + "[13] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 1, 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jhair Gallardo, Tyler L Hayes, and Christopher Kanan. Self-supervised training enhances online continual learning. arXiv preprint arXiv:2103.14010, 2021. 2", + "[15] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8059-8068, 2019. 2", + "[16] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 1", + "[17] Jacob Goldberger, Geoffrey E Hinton, Sam Roweis, and Russ R Salakhutdinov. Neighbourhood components analysis. Advances in neural information processing systems, 17, 2004. 2, 3", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2, 4, 5", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[20] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3, 5", + "[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. 2", + "[22] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsivash. Mean shift for self-supervised learning. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10306-10315, 2021. 2", + "[23] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 5", + "[24] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1", + "[25] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1", + "[26] George A Miller. WordNet: An electronic lexical database. MIT press, 1998. 7", + "[27] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. 5", + "[28] Ruslan Salakhutdinov and Geoff Hinton. Learning a nonlinear embedding by preserving class neighbourhood struc" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19921", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ture. In Artificial Intelligence and Statistics, pages 412-419. PMLR, 2007. 2, 3", + "[29] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020. 1", + "[30] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 7", + "[31] Hugo Touvron, Alexandre Sablayrolles, Matthijs Douze, Matthieu Cord, and Hervé Jégou. Graft: Learning fine-grained image representations with coarse labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 874-884, 2021. 2, 4", + "[32] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 3", + "[33] Guangrun Wang, Keze Wang, Guangcong Wang, Philip HS Torr, and Liang Lin. Solving inefficiency of self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9505-9515, 2021. 2", + "[34] Xudong Wang, Ziwei Liu, and Stella X. Yu. Unsupervised feature learning by cross-level instance-group discrimination. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12581-12590, 2021. 2", + "[35] Longhui Wei, Lingxi Xie, Jianzhong He, Jianlong Chang, Xiaopeng Zhang, Wengang Zhou, Houqiang Li, and Qi Tian. Can semantic labels assist self-supervised visual representation learning? arXiv preprint arXiv:2011.08621, 2020. 2", + "[36] Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darell. What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659, 2020. 8", + "[37] Yuanhong Xu, Qi Qian, Hao Li, Rong Jin, and Juhua Hu. Weakly supervised representation learning with coarse labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10593-10601, 2021. 2, 4", + "[38] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1476-1485, 2019. 2", + "[39] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16660-16669, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19922", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_model.json b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bf88cbae0c528050d6b0c643916e8b5df1f770e5 --- /dev/null +++ b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.131, + 0.838, + 0.152 + ], + "angle": 0, + "content": "MaskCon: Masked Contrastive Learning for Coarse-Labelled Dataset" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.182, + 0.642, + 0.217 + ], + "angle": 0, + "content": "Chen Feng, Ioannis Patras \nQueen Mary University of London, UK" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.22, + 0.624, + 0.234 + ], + "angle": 0, + "content": "{chen.feng, i.patras}@qmul.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.771 + ], + "angle": 0, + "content": "Deep learning has achieved great success in recent years with the aid of advanced neural network structures and large-scale human-annotated datasets. However, it is often costly and difficult to accurately and efficiently annotate large-scale datasets, especially for some specialized domains where fine-grained labels are required. In this setting, coarse labels are much easier to acquire as they do not require expert knowledge. In this work, we propose a contrastive learning method, called masked contrastive learning (MaskCon) to address the under-explored problem setting, where we learn with a coarse-labelled dataset in order to address a finer labelling problem. More specifically, within the contrastive learning framework, for each sample our method generates soft-labels with the aid of coarse labels against other samples and another augmented view of the sample in question. By contrast to self-supervised contrastive learning where only the sample's augmentations are considered hard positives, and in supervised contrastive learning where only samples with the same coarse labels are considered hard positives, we propose soft labels based on sample distances, that are masked by the coarse labels. This allows us to utilize both inter-sample relations and coarse labels. We demonstrate that our method can obtain as special cases many existing state-of-the-art works and that it provides tighter bounds on the generalization error. Experimentally, our method achieves significant improvement over the current state-of-the-art in various datasets, including CIFAR10, CIFAR100, ImageNet-1K, Stanford Online Products and Stanford Cars196 datasets. Code and annotations are available at https://github.com/MrChenFeng/MaskCon_CVPR2023." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.8, + 0.21, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Supervised learning with deep neural networks has achieved great success in various computer vision tasks such as image classification, action detection and object localization. However, the success of supervised learning relies on large-scale and high-quality human-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.272, + 0.891, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.437, + 0.893, + 0.478 + ], + "angle": 0, + "content": "Figure 1. Contrastive learning sample relations using MaskCon (ours) and other learning paradigms when only coarse labels are available. MaskCon are closer to the fine ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.719 + ], + "angle": 0, + "content": "annotated datasets, whose annotations are time-consuming and labour-intensive to produce. To avoid such reliance, various learning frameworks have been proposed and investigated: Self-supervised learning aims to learn meaningful representations with heuristic proxy visual tasks, such as rotation prediction [16] and the more prevalent instance discrimination task, the latter, being widely applied in self-supervised contrastive learning framework; semi-supervised learning usually considers a dataset for which only a small part is annotated - within this setting, pseudo labelling methods [24] and consistency regularization techniques [1, 29] are typically used; Moreover, learning using more accessible but noisy data, such as web-crawled data, has also received increasing attention [13, 25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this work, we consider an under-explored problem setting aiming at reducing the annotation effort - learning fine-grained representations with a coarsely-labelled dataset. Specifically, we learn with a dataset that is fully labeled, albeit at a coarser granularity than we are interested in (i.e., that of the test set). Compared to fine-grained labels, coarse labels are often significantly easier to obtain, especially in some of the more specialized domains, such as the recognition and classification of medical pathology images. As a simple example, for the task of differentiation between different pets, we need a knowledgeable cat lover to distinguish between 'British short' and 'Siamese'," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19913" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "but even a child annotator may help to discriminate between 'cat' and 'non-cat' (Fig. 1). Unfortunately, learning with a coarse labelled dataset has been less investigated compared to other weakly supervised learning paradigms. Recently, Bukchi et al. [2] investigate on learning with coarse labels in the few-shot setting. More closely related to us, Grafit [31] proposes a multi-task framework by a weighted combination of self-supervised contrastive learning and supervised contrastive learning cost; Similarly, CoIns [37] uses both a self-supervised contrastive learning cost and a supervised learning cross-entropy loss. Both works combine a fully supervised learning cost (cross entropy or contrastive) with a self-supervised contrastive loss – these works are the main ones with which we compare." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.575 + ], + "angle": 0, + "content": "Differently than them, instead of using self-supervised contrastive learning as an auxiliary task, we propose a novel learning scheme, namely Masked Contrastive Learning (MaskCon). Our method aims to learn by considering inter-sample relations of each sample with other samples in the dataset. Specifically, we always consider the relation to oneself as confidently positive. To estimate the relations to other samples, we derive soft labels by contrasting an augmented view of the sample in question with other samples, and further improve it by utilizing the mask generated based on the coarse labels. Our approach generates soft inter-sample relations that can more accurately estimate fine inter-sample relations compared to the baseline methods (Fig. 1). Efficiently and effectively, our method achieves significant improvements over the state-of-the-art in various datasets, including CIFARtoy, CIFAR100, ImageNet-1K and more challenging fine-grained datasets Stanford Online Products and Stanford Cars196." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.223, + 0.603 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.673 + ], + "angle": 0, + "content": "In this section, we first briefly review the contrastive learning works which are the basis of our method and then, we briefly introduce the state-of-the-art works in related problem settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Self-supervised contrastive learning Contrastive learning has recently emerged as a powerful method for representation learning without labels. In the context of self-supervision, contrastive learning has been used for instance discrimination, producing models whose performance rivals that of fully supervised training. The objective of contrastive instance discrimination is to minimize the distance between transformed views of the same sample in the feature space, while maximizing their distance from views of other samples. Notable methods in this area are SimCLR [5] and MoCo [18], as well as other variants including SwAV [4], which leverages clustering to identify prototypes for the contrastive learning objective and [33], that reformulates the instance discrimination contrastive objective in the" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "form of a triplet loss. These methods have been found to produce models that, requiring only limited amounts of annotated data for fine-tuning on a particular task or domain, achieve performance that reaches or even surpasses that of supervised models [6, 11]. Furthermore, self-supervised training has been found to compare positively with supervised training in other respects, such as producing models that perform better in the context of continual learning [14]." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.213, + 0.893, + 0.379 + ], + "angle": 0, + "content": "However, self-supervised contrastive learning based on instance discrimination task usually suffers from the 'false negative' problem – that samples from the same class should not be considered as negative. For example, pictures of different cats should not be considered completely negative to each other. To this end, heuristically modifying the inter-sample relations have been widely applied in self-supervised learning [10, 12, 22, 34], intuitively similar to us. Many of the state-of-the-art methods can be considered as special cases of our method by the adjustment of hyperparameters (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.398, + 0.893, + 0.61 + ], + "angle": 0, + "content": "The success of self-supervised contrastive learning has led to increasing attention on how to integrate contrastive learning into existing supervised learning paradigms. One approach is supervised contrastive learning, which adapts contrastive learning to the fully supervised setting. Supervised contrastive learning was first introduced in [17, 28] under the name of neighbourhood components analysis (NCA). Recently, supervised contrastive learning has regained interest due to the significant progress made in [20] by employing more advanced network structures and image augmentations. Compared to standard supervised learning using cross-entropy loss, supervised contrastive learning exhibits superior performance in hard sample mining [20]." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.61, + 0.894, + 0.806 + ], + "angle": 0, + "content": "Other works have attempted to introduce self-supervised contrastive learning as an auxiliary task, especially for weakly supervised learning problems. Gidaris et al. [15] jointly train a supervised classifier and a rotation-prediction objective. Komodakis and Gidaris [21] seeking to improve model performance on few-shot learning. Wei et al. [35] propose a label-filtered instance discrimination objective for the pre-training of models to be used for transfer learning in other datasets. S4 [38] uses rotation and exemplar [9] prediction to improve performance in semi-supervised learning tasks. Feng et al. [13] apply a self-supervised contrastive loss to safely learn meaningful representations in the presence of noisy labels." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Hierarchical image classification & Unsupervised clustering Hierarchical image classification methods aim to learn better representations with datasets having hierarchically structured labels. To utilize information from different levels of labels, Zhang et al. [39] propose a multi-losses" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19914" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "contrastive learning framework with each loss considering a specific level of labels. Unsupervised clustering methods often have the same problem setting but different goals than self-supervised learning, that is, to recover and identify the ground-truth semantic labels of each sample [3, 32]. Deep-Cluster [3] simultaneously learns the clustering assignments and a parametric classifier, by feeding the clustering assignments as pseudo labels to the classifier. SCAN [32] inherits the idea of DeepClustering while augmenting extra self-labelling and neighbor mining techniques." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.256, + 0.212, + 0.273 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.473, + 0.452 + ], + "angle": 0, + "content": "We consider the problem of learning when the labels of the training data and the labels of the test data are inconsistent in granularity, and in particular when the labels of the training set are coarser. More specifically, let us denote with \\( \\mathbf{X} = \\{\\pmb{x}_i\\}_{i=1}^N \\) an i.i.d sampled train dataset with annotated coarse labels \\( \\mathbf{Y} = \\{\\pmb{y}_i \\in \\{0,1\\}^M\\}_{i=1}^N \\) subject to \\( \\sum_{m=1}^{M} \\pmb{y}_{im} = 1 \\). Here, \\( N \\) denotes the number of samples and \\( M \\) denotes the number of coarse classes. Let us also define a finely labelled set \\( \\mathbf{Y}' = \\{\\pmb{y}_i' \\in \\{0,1\\}^{M'}\\}_{i=1}^N \\) subject to \\( \\sum_{k=1}^{M'} \\pmb{y}_i^k = 1 \\), where \\( M' \\) denotes the number of fine classes. Typically, \\( M' > M \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.461, + 0.218, + 0.475 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.485, + 0.469, + 0.514 + ], + "angle": 0, + "content": "We first quickly review the common learning paradigms utilized in the baseline methods and our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.533, + 0.272, + 0.549 + ], + "angle": 0, + "content": "3.1.1 Supervised learning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.556, + 0.47, + 0.587 + ], + "angle": 0, + "content": "In most supervised learning frameworks, the objective is the minimization of the below empirical risk:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.597, + 0.47, + 0.639 + ], + "angle": 0, + "content": "\\[\nR (f, g) = \\sum_ {i = 1} ^ {N} L _ {c e} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {y} _ {i}; f, g\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.648, + 0.469, + 0.678 + ], + "angle": 0, + "content": "where \\( f \\) denotes the feature encoder, and \\( g \\) denotes the classifier head (usually a single fully-connected layer)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.678, + 0.47, + 0.723 + ], + "angle": 0, + "content": "For brevity, we denote \\( \\pmb{f}_i \\triangleq f(\\pmb{x}_i) \\), \\( \\pmb{g}_i \\triangleq g(f(\\pmb{x}_i)) \\) and \\( \\pmb{p}_i \\triangleq \\mathrm{softmax}(\\pmb{g}_i) \\) as the feature, logit and prediction of a sample view \\( \\pmb{x}_i \\) respectively. Here," + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.734, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\[\nL _ {c e} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {y} _ {i}; f, g\\right) = - \\sum_ {m = 1} ^ {M} \\boldsymbol {y} _ {i} ^ {m} \\log \\boldsymbol {p} _ {i} ^ {m} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.469, + 0.814 + ], + "angle": 0, + "content": "is the cross-entropy loss function - the default loss function for classification problems." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.275, + 0.848 + ], + "angle": 0, + "content": "3.1.2 Contrastive learning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Unlike the common supervised learning model above where a parametric classifier \\( g \\) is learned based on the semantic labels \\( Y \\) of the samples \\( X \\), we can also perform contrastive" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "learning based on the inter-sample relations \\( Z = \\{z_{i} \\in (0,1)^{N}\\}_{i=1}^{N} \\), with each entry \\( z_{ij} \\) depicting the inter-sample relation between \\( x_{i} \\) and \\( x_{j} \\). Intuitively, \\( z_{ij} = 1 \\) means that sample \\( x_{i} \\) and \\( x_{j} \\) generate a strong positive pair. Since each sample may form multiple positive sample pairs, for brevity, we abuse the notation here with \\( Z \\) denoting also the sample-wise normalized inter-sample relations. To learn such intersample relations, instead of a parametric classifier \\( g \\), the \\( f \\) encoder is usually followed by a projector \\( h \\), which is often implemented as an MLP and learned by regularizing the inter-sample relations \\( Z \\) (Eq. (4))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.305 + ], + "angle": 0, + "content": "More specifically, let us denote by \\( \\pmb{h}_i \\triangleq h(f(\\pmb{x}_i)) \\) the projection. We first calculate the cosine similarity \\( \\pmb{d}_i \\) between a sample \\( \\pmb{x}_i \\) and the dataset \\( \\pmb{H} = \\{\\pmb{h}_n\\}_{n=1}^N \\):" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.311, + 0.892, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {d} _ {i} = \\left[ \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {1}\\right), \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {2}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {N}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Let us further define \\( \\pmb{q}_i \\triangleq \\mathrm{softmax}(\\pmb{d}_i / \\tau_0) \\), where \\( \\tau_0 \\) is the temperature hyperparameter. Then the following empirical risk will be optimized:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.387, + 0.892, + 0.428 + ], + "angle": 0, + "content": "\\[\nR (f, h) = \\sum_ {i = 1} ^ {N} L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {z} _ {i}; f, h\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.435, + 0.851, + 0.45 + ], + "angle": 0, + "content": "where the contrastive loss \\( L_{con} \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.456, + 0.892, + 0.497 + ], + "angle": 0, + "content": "\\[\nL _ {c o n} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {z} _ {i}; f, h\\right) = - \\sum_ {n = 1} ^ {N} \\boldsymbol {z} _ {i} ^ {n} \\log \\boldsymbol {q} _ {i} ^ {n}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.591 + ], + "angle": 0, + "content": "Self-supervised contrastive learning We first introduce the most prevalent form of contrastive learning currently — self-supervised contrastive learning for learning without labels. Since there are no annotated labels, we usually set \\( Z^{self} \\) by considering each sample having only itself as positive:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.591, + 0.891, + 0.623 + ], + "angle": 0, + "content": "\\[\nz _ {i j} ^ {\\text {s e l f}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} i = j \\\\ 0, & \\text {i f} i \\neq j \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Typically, we aim at maximizing the relations between different augmented views of the same sample while minimizing the relations between augmented views of different samples. This is also widely known as instance discrimination task. We denote such self-supervised contrastive loss as \\( L_{selfcon} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Supervised contrastive learning We can easily transition from self-supervised contrastive learning to supervised contrastive learning, by simply changing the inter-sample relations \\( Z^{sup} \\) for Eq. (4) and Eq. (5). More specifically, as in [17, 20, 28], samples from the same semantic classes are considered positive pairs," + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.813, + 0.892, + 0.846 + ], + "angle": 0, + "content": "\\[\nz _ {i j} ^ {\\text {s u p}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\\\ 0, & \\text {i f} \\boldsymbol {y} _ {i} \\neq \\boldsymbol {y} _ {j} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.892, + 0.899 + ], + "angle": 0, + "content": "in addition to the sample itself (and its augmented views) as in self-supervised contrastive learning. We denote the supervised contrastive loss as \\( L_{supcon} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19915" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.09, + 0.468, + 0.301 + ], + "angle": 0, + "content": "Memory bank For consistency with the supervised cross-entropy loss and for clarity, we present all the essential formulas above in terms of the entire dataset \\( \\mathbf{X} \\). However, due to computation and memory constraints, it is often unrealistic to contrast a specific sample \\( \\mathbf{x}_i \\) with the whole dataset. In the actual implementation in this work, following MoCo [18], we use a dynamic FIFO memory bank \\( \\mathbf{H} = \\{\\mathbf{h}_p\\}_{p=1}^P \\) consisting of cached projections of \\( P \\) samples. For each sample \\( \\mathbf{x}_i \\), we generate two random augmented views as \\( \\mathbf{x}_q \\) (query) and \\( \\mathbf{x}_k \\) (key) with their corresponding projections denoted by \\( \\mathbf{h}_q \\) and \\( \\mathbf{h}_k \\), respectively. Then, we calculate the cosine similarity between \\( \\mathbf{h}_q \\) on the one hand, and \\( \\mathbf{h}_k \\) and each projection in the memory bank on the other:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.31, + 0.466, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {d} _ {i} = \\overbrace {\\left[ \\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {k}\\right) , \\overbrace {\\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {1}\\right) , \\dots , \\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {P}\\right)} ^ {\\text {s e l f}} \\right]} ^ {\\text {m e m o r y b a n k}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.358, + 0.443, + 0.373 + ], + "angle": 0, + "content": "For more details, please refer to the MoCo paper [18]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.399, + 0.254, + 0.413 + ], + "angle": 0, + "content": "3.1.3 Baseline methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.424, + 0.47, + 0.681 + ], + "angle": 0, + "content": "Clearly, with respect to the finer target question labels, each sample \\( \\boldsymbol{x}_i \\), is a trustworthy positive sample in relation to itself (i.e., they have the same fine-grained label). However, considering the samples with the same coarse label as positives may lead to under-clustering issues (Fig. 3) (they may have different fine-grained labels). Recent works attempt to address such under-clustering by combining an additional self-supervised contrastive loss with a specific supervised loss that utilizes coarse labels. Grafit [31] combine it with the supervised contrastive loss and CoIns [37] with the supervised cross-entropy loss. In both cases, as a result, the self-supervised contrastive loss that considers each sample as a class by itself, aims to mitigate the tendency for under-clustering. Formally, given our definitions of the previously mentioned loss functions in Sec. 3.1 (\\( L_{ce} \\), \\( L_{selfcon} \\) and \\( L_{supcon} \\)) it turns out that the losses used by Grafit [31] and CoIns [37] can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.695, + 0.469, + 0.712 + ], + "angle": 0, + "content": "\\[\nL _ {G r a f i t} = w L _ {s u p c o n} + (1 - w) L _ {s e l f c o n} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.715, + 0.469, + 0.731 + ], + "angle": 0, + "content": "\\[\nL _ {C o I n s} = w L _ {c e} + (1 - w) L _ {s e l f c o n} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.744, + 0.405, + 0.759 + ], + "angle": 0, + "content": "where \\(w\\) controls the relative weight of each loss." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.424, + 0.788 + ], + "angle": 0, + "content": "3.2. MaskCon: Masked Contrastive learning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Instead of equally weighing all the samples within the same coarse classes, we aim to emphasize the samples with the same fine labels and reduce the importance of the others. To achieve this, we introduce a novel contrastive learning method, namely Masked Contrastive learning (MaskCon), within the framework of contrastive learning that utilizes inter-sample relations directly." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "More specifically, for sample \\( \\boldsymbol{x}_i \\), we estimate its intersample relations \\( \\boldsymbol{z}_i' \\) to other samples utilizing the key view projection \\( \\boldsymbol{h}_k \\) and the whole dataset \\( \\{\\boldsymbol{h}_1, \\dots, \\boldsymbol{h}_N\\} \\) excluding itself (since it will always be considered as a trustworthy positive), as below:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.178, + 0.892, + 0.218 + ], + "angle": 0, + "content": "\\[\nz _ {i j} ^ {\\prime} = \\frac {\\mathbb {1} \\left(\\boldsymbol {y} _ {j} = \\boldsymbol {y} _ {i}\\right) \\cdot \\exp \\left(d _ {i j} ^ {\\prime} / \\tau\\right)}{\\sum_ {n = 1 , n \\neq i} ^ {N} \\mathbb {1} \\left(\\boldsymbol {y} _ {n} = \\boldsymbol {y} _ {i}\\right) \\cdot \\exp \\left(d _ {i n} ^ {\\prime} / \\tau\\right)}, i \\neq j, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.229, + 0.731, + 0.244 + ], + "angle": 0, + "content": "where the similarity \\(d_i^{\\prime}\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.255, + 0.892, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {d} _ {i} ^ {\\prime} = \\left[ \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {1}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {i - 1}\\right), \\right. \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.278, + 0.824, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\left. \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {i + 1}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {N}\\right) \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.306, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Please note the use of the mask \\((\\mathbb{1}(\\pmb {y}_j = \\pmb {y}_i))\\) that excludes from the softmax that estimates inter-sample relationships, the samples \\(j\\) that have a different coarse label with the sample \\(i\\) (and sets their \\(z_{ij}^{\\prime}\\) to 0). While it is risky to consider all samples from the same coarse class as positive, we can confidently identify those samples that do not have the same coarse class as negative. This reduces the noise in \\(z_i^\\prime\\). Finally, we re-scale the \\(z_{i}^{\\prime}\\) with its maximum" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.439, + 0.892, + 0.459 + ], + "angle": 0, + "content": "\\[\nz _ {i j} ^ {\\prime} = z _ {i j} ^ {\\prime} / \\max \\left(\\boldsymbol {z} _ {i} ^ {\\prime}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.892, + 0.497 + ], + "angle": 0, + "content": "to make the closest neighbour as positive as the sample itself and arrive at:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.508, + 0.892, + 0.542 + ], + "angle": 0, + "content": "\\[\nz _ {i j} ^ {\\text {m a s k}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} i = j \\\\ z _ {i j} ^ {\\prime}, & \\text {i f} i \\neq j \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Compared to \\( Z^{supcon} \\), we thus reweight the samples of the same coarse label according to the similarities in the feature space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.597, + 0.892, + 0.642 + ], + "angle": 0, + "content": "We denote the masked contrastive loss as \\( L_{maskcon} \\) and, similarly to Grafit and CoIns, we also consider a weighted loss as the final objective:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.655, + 0.892, + 0.672 + ], + "angle": 0, + "content": "\\[\nL = w L _ {\\text {m a s k c o n}} + (1 - w) L _ {\\text {s e l f c o n}} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Relations to SOTA works By adjusting \\( w \\) and \\( \\tau \\), our method can obtain various existing SOTA methods as special cases. More specifically, by setting \\( \\tau \\) to \\( \\infty \\), our method degenerates to Grafit [31]. Ignoring the mask, i.e., treating all samples as having the same coarse label, our method can also generalize to existing SOTAs in self-supervised learning. For example, NNCLR [10] instead of considering only each sample itself as positive it considers as positive the nearest neighbors as well. Formally, by setting \\( w = 1 \\) and \\( \\tau = 0 \\), MaskCon degenerates to NNCLR; ASCL [12] proposes to introduce extra positive samples by calculating soft inter-sample relations, adaptively weighted by its own normalized entropy. By setting \\( w = \\text{norm\\_entropy}(z_i') \\), MaskCon degenerates to ASCL." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19916" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.396, + 0.108 + ], + "angle": 0, + "content": "3.3. Theoretical justification of MaskCon" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.114, + 0.47, + 0.191 + ], + "angle": 0, + "content": "The fundamental goal of contrastive learning is to identify \\( Z \\) that suits the problem granularity and accurately represents the relationship between samples. Assuming that there exists an optimal hidden \\( \\hat{Z} \\), we can define the expected population risk in terms of \\( \\hat{Z} \\) as shown in Eq. (16)." + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.205, + 0.469, + 0.222 + ], + "angle": 0, + "content": "\\[\nR _ {E} (f, h) = E _ {\\boldsymbol {x} _ {i}} \\left[ L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {z}} _ {i}; f, h\\right) \\right] \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.227, + 0.469, + 0.259 + ], + "angle": 0, + "content": "Moreover, let us define the empirical risk with respect to \\(\\hat{Z}\\) as shown in Eq. (17)." + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.267, + 0.469, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\hat {R} (f, h) = \\sum_ {i = 1} ^ {N} L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {z}} _ {i}; f, h\\right) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.47, + 0.378 + ], + "angle": 0, + "content": "THEOREM 1. The generalization error bounds of contrastive learning depends on the difference between the inter-sample relations \\(\\mathbf{Z}\\) and the optimal hidden intersample relations \\(\\hat{\\mathbf{Z}}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.388, + 0.469, + 0.426 + ], + "angle": 0, + "content": "\\[\nd _ {z} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| \\boldsymbol {z} _ {i} - \\hat {\\boldsymbol {z}} _ {i} \\| _ {2} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.435, + 0.469, + 0.496 + ], + "angle": 0, + "content": "Proof. Let us assume \\( L_{con}(\\pmb{x}_i, \\pmb{z}_i; f, h) \\in [a, b] \\) and is \\( \\lambda \\)-Lipschitz continuous w.r.t \\( \\pmb{z}_i \\). Let \\( |\\mathcal{F}| \\) and \\( |\\mathcal{H}| \\) be the covering numbers that correspond to the finite hypothesis space of \\( f \\) and \\( g \\). We have:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.505, + 0.446, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| R _ {E} (f, h) - R (f, h) \\right| \\leq \\overbrace {\\left| R _ {E} (f , h) - \\hat {R} (f , h) \\right|} ^ {\\text {H o e f f d i n g ' s i n e q u a l i t y}} \\\\ L i p s c h i t z \\quad c o n t i n u o u s \\\\ + \\overbrace {| \\hat {R} (f , h) - R (f , h) |} \\\\ \\leq | a - b | \\sqrt {\\frac {\\log (2 | \\mathcal {F} | \\cdot | \\mathcal {H} | / \\delta)}{2 n}} \\\\ + \\lambda \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| z _ {i} - \\hat {z} _ {i} \\| _ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.672, + 0.28, + 0.687 + ], + "angle": 0, + "content": "with probability at least \\(1 - \\delta\\)" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.673, + 0.469, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.697, + 0.469, + 0.773 + ], + "angle": 0, + "content": "Based on Theorem 1, it can be concluded that our method performs better than Grafit with high probability, because an appropriate temperature \\(\\tau\\) allows \\(Z^{mask}\\) to more accurately estimate the optimal \\(\\hat{Z}\\) (Fig. 1) in comparison to assigning equal weights to all coarse neighbours." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.209, + 0.802 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, extensive experiments are conducted to validate the findings and effectiveness of the proposed Masked Contrastive Learning (MaskCon) framework. The impact of hyperparameters is thoroughly investigated in Sec. 4.1. In Sec. 4.2, Sec. 4.3 and Sec. 4.4, results are provided for experiments on CIFAR datasets and ImageNet-1K" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "dataset with coarse labels. In Sec. 4.5, we apply our method on more challenging fine-grained datasets, including Stanford Online Products (SOP) dataset [30] and Stanford Cars 196 dataset [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.317 + ], + "angle": 0, + "content": "To ensure the robustness of the proposed method and exclude the impact of model capacity, the model settings are kept consistent throughout all the experiments, except for the two hyperparameters, \\( w \\) and \\( \\tau \\), which are ablated in Sec. 4.1. ResNet18 [19] is employed in all experiments, with the initial convolutional layer modified to have a kernel size of \\( 3 \\times 3 \\) and a stride of 1, and the initial max-pooling operator removed for CIFAR experiments, to account for the smaller image size \\( (32 \\times 32) \\) as suggested in prior works [18]. More implementation and dataset details can be found in APPENDIX A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.485 + ], + "angle": 0, + "content": "We compare our method with two competing methods: Grafit and CoIns. For a fair comparison, we exhaust the weight \\( w \\) choices for both methods and report the best achievable results in all experiments. Note that when \\( w = 0 \\), Grafit and CoIns degenerate to self-supervised contrastive learning denoted as SelfCon; Conversely, when \\( w = 1 \\), Grafit degenerates to supervised contrastive learning [20] denoted as SupCon, while CoIns degenerates to conventional supervised cross-entropy learning denoted as SupCE. For reference, we also show the results when training with fine labels – this is denoted as SupFINE." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.502, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Evaluation protocol To evaluate the different methods on the test set with fine labels, we use the recall@K [27] metric widely used in the image retrieval task. Each test image first retrieves top-K nearest neighbours from the test set and receives 1 if there exists at least one image from the same fine class among the top-K nearest neighbours, otherwise 0. Recall@K averages this score over all the test images." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.636, + 0.838, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.811, + 0.862, + 0.825 + ], + "angle": 0, + "content": "Figure 2. Recall@1 w.r.t \\(w\\) and \\(\\tau\\) on CIFAR100 dataset" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.666, + 0.862 + ], + "angle": 0, + "content": "4.1. Effect of \\( w \\) and \\( \\tau \\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "In this section, we extensively ablate the effect of hyperparameters of MaskCon - weight \\( w \\) and temperature \\( \\tau \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19917" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.154, + 0.089, + 0.819, + 0.226 + ], + "angle": 0, + "content": "
MethodCIFARtoy-goodsplitCIFARtoy-badsplit
Recall@1Recall@2Recall@5Recall@10Recall@1Recall@2Recall@5Recall@10
SelfCon84.8391.5596.3598.1684.8391.5596.3598.16
Graft86.6192.3397.0198.3889.9694.3697.6198.10
SupCon73.8484.2592.1495.4684.6690.9395.1596.71
CoIns86.1592.7697.2198.4690.5594.9497.7398.71
SupCE76.3085.2694.6597.4687.1592.8596.7898.34
SupFINE94.1196.5398.2598.9694.1196.5398.2598.96
MaskCon (Ours)90.28 (13.98↑)94.0497.3398.5391.56 (4.41↑)95.2397.7098.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.37, + 0.235, + 0.599, + 0.249 + ], + "angle": 0, + "content": "Table 1. Results on CIFARtoy dataset." + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.264, + 0.312, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.264, + 0.474, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.264, + 0.635, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.264, + 0.796, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.381, + 0.713, + 0.396 + ], + "angle": 0, + "content": "Figure 3. t-SNE visualization of learned representation on CIFARtoy dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.422, + 0.47, + 0.468 + ], + "angle": 0, + "content": "Specifically, we show the recall@1 results on CIFAR100 dataset in Fig. 2, with \\( w = \\{0, 0.2, 0.5, 0.8, 1.0\\} \\) and \\( \\tau = \\{0, 0.01, 0.05, 0.1, 0.5, \\infty\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.635 + ], + "angle": 0, + "content": "The results demonstrate a significant improvement in performance upon the inclusion of additional inter-sample relationships (\\(w \\neq 0\\)) in comparison to self-supervised learning alone (\\(w = 0\\)) on the CIFAR dataset. Furthermore, a suitable temperature \\(\\tau\\), such as 0.05 or 0.1, consistently yields superior results compared to the simple weighted combination in Grafit (\\(\\tau = \\infty\\)). In general, we recommend initiating the hyperparameter search with \\(w = 1\\) and \\(\\tau = \\tau_0\\), where \\(\\tau_0\\) denotes the temperature for the projection head. Additional results concerning various hyperparameters on other datasets can be found in APPENDIX B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.379, + 0.665 + ], + "angle": 0, + "content": "4.2. Experiments on CIFARtoy dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.839 + ], + "angle": 0, + "content": "To simulate the coarse labelling process, we manually generate toy datasets based on the CIFAR10 dataset. Following the problem motivation, a subset of 8 classes is selected from the original 10 classes. Specifically, the classes 'airplane', 'truck', 'automobile', and 'ship' are designated as Class A, while 'horse', 'dog', 'bird', and 'cat' are assigned to Class B. Note that Class A comprises non-organic objects, while Class B comprises animals. Moreover, a bad split of the aforementioned dataset is defined, wherein 'airplane', 'automobile', 'bird', and 'cat' are Class A, and 'horse', 'dog', 'ship', and 'truck' are Class B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As shown in Tab. 1, our method achieves the best performance in both splits. Moreover, the improvement over the supervised learning with coarse labels (SupCE) on good split \\((13.98\\%)\\) is much higher than the bad split \\((4.41\\%)\\)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.89, + 0.451 + ], + "angle": 0, + "content": "which shows the substantial potential of MaskCon in dealing with more realistic coarse-labelled datasets1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.573 + ], + "angle": 0, + "content": "In Fig. 3 we visualize the learned features of all test samples in the good split. Clearly, the learned representations with supervised contrastive learning (SupCon) and self-supervised contrastive learning (SelfCon) tend to under-cluster and over-cluster the samples, respectively. By contrast, our method gets more compact and clear clusters, in line with the results when trained with fine labels (SupFINE)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.581, + 0.805, + 0.597 + ], + "angle": 0, + "content": "4.3. Experiments on CIFAR100 dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.725 + ], + "angle": 0, + "content": "The common CIFAR100 dataset has 20 classes of coarse labels in addition to the 100 classes of fine labels, with each coarse class containing five fine-grained classes (500 samples). The results in Tab. 2 show that our method achieves significant improvements over the SOTAs. In particular, it improves the top-1 retrieval precision from \\(47.25\\%\\) to \\(65.52\\%\\), approaching the results by the model learned with fine labels \\((71.13\\%)\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.733, + 0.825, + 0.75 + ], + "angle": 0, + "content": "4.4. Experiments on ImageNet-1K dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.832 + ], + "angle": 0, + "content": "In this section, we evaluate our method on the largescale ImageNet-1K dataset. For efficiency, we experiment with the downsampled version of ImageNet-1K dataset [7], where each sample was resized to \\(32 \\times 32\\). Since no official coarse labels exist for ImageNet, we introduce coarse" + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.839, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1It is worth noting that the better performance on the bad split may seem counter-intuitive at first glance. However, note that a bad split actually makes the labels more informative, as it explicitly helps to discriminate visually similar classes, e.g., 'cat' and 'dog' in different coarse classes in our setting." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19918" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.09, + 0.089, + 0.458, + 0.204 + ], + "angle": 0, + "content": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon40.5051.8366.2376.66
Graft60.5771.1382.3289.21
SupCon58.6570.0482.1889.09
CoIns60.1070.8983.1489.52
SupCE47.2561.2477.7887.01
SupFINE71.1380.0387.6191.59
MaskCon (Ours)65.52 (18.17†)74.4683.6489.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.214, + 0.393, + 0.227 + ], + "angle": 0, + "content": "Table 2. Results on CIFAR100 dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.47, + 0.359 + ], + "angle": 0, + "content": "labels based on the WordNet [26] hierarchy so as to artificially group the whole dataset into 12 coarse classes: '0: Invertebrate', '1: Domestic animal', '2: Bird', '3: Mammal', '4: Reptile/Aquatic vertebrate', '5: Device', '6: Vehicle', '7: Container', '8: Instrument', '9: Artifact', '10: Clothing' and '11: Others'. In Tab. 3, we show the results on down-sampled ImageNet-1K datasets." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.37, + 0.458, + 0.487 + ], + "angle": 0, + "content": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon10.2814.1522.3630.34
Graft18.1325.4637.1946.64
SupCon13.3619.4029.7739.38
CoIns18.3625.5437.0946.89
SupCE12.2314.1527.7637.03
SupFINE33.9744.5557.2365.77
MaskCon (Ours)19.08 (6.86↑)26.2138.1747.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.497, + 0.399, + 0.511 + ], + "angle": 0, + "content": "Table 3. Results on ImageNet-1K dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.573 + ], + "angle": 0, + "content": "We can again validate that with temperature-controlled soft relations, our method surpasses SupCE, Grafit and CoIns consistently, especially for top-10 retrieval results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.58, + 0.4, + 0.596 + ], + "angle": 0, + "content": "4.5. Experiments on fine-grained datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.469, + 0.693 + ], + "angle": 0, + "content": "In this section, we conduct experiments in a more challenging scenario - fine-grained datasets with only coarse labels. Please note, that in this work, we do not aim to compare with the state-of-the-art works in fine-grained classification, which usually involve more specialized techniques, such as object localization and local feature extraction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.357, + 0.726 + ], + "angle": 0, + "content": "4.5.1 Stanford Online Products (SOP)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The Stanford Online Products (SOP) dataset [30] consists of 22,634 products, with each product having between two and twelve photos from different perspectives, for a total of 120,053 images. In addition, there are 12 coarse classes based on the semantic categories of the products, such as 'bicycle' and 'kettle'. For the common image retrieval task, the category of the test set is possibly unknown, so we firstly extract a subset with an unknown test category based on the SOP dataset. We then selected the categories with eight or more images. We then split almost equally this subset to obtain the training set and the test set with 25,368 and 25,278" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "images, consisting of 2,517 and 2,518 classes respectively, denoted as SOP-split1. Moreover, we select all products with twelve images, and then randomly select ten train images and two test images. Thus, we have 1,498 classes with a total of 17,976 images (2,996 test images and 14,980 train images), denoted as SOP-split2." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.195, + 0.88, + 0.312 + ], + "angle": 0, + "content": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon70.3675.5781.5385.13
Graft74.0278.8284.1387.91
SupCon53.6959.5567.1272.78
CoIns70.8476.0182.286.08
SupCE36.3542.3950.3056.52
SupFINE83.9488.0491.9594.00
MaskCon (Ours)74.05 (37.7↑)78.9784.4887.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.322, + 0.813, + 0.336 + ], + "angle": 0, + "content": "Table 4. Results on SOP-split1 dataset." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.366, + 0.88, + 0.483 + ], + "angle": 0, + "content": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon35.8541.4649.7756.11
Graft39.1244.6653.1059.65
SupCon25.0729.2435.8541.59
CoIns38.2245.1954.3761.18
SupCE22.5626.3433.2838.95
SupFINE69.5675.7083.2487.88
MaskCon (Ours)45.36 (22.8†)51.0758.9165.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.493, + 0.813, + 0.507 + ], + "angle": 0, + "content": "Table 5. Results on SOP-split2 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.705 + ], + "angle": 0, + "content": "In Tab. 4 and Tab. 5 we show results on SOP-split1 and SOP-split2, respectively. Similarly, regardless of whether the test class is known or not, our method still achieves significant improvements over state-of-the-art methods. An interesting phenomenon here is, that on the SOP dataset, SelfCon alone is significantly better than its SupCE/SupCon counterpart. We conjecture that this is due to the fact that each fine class of the SOP dataset in fact consists of different views of the same product, and the number of images per fine class is quite small (2 to 12). For such a dataset, the classification of fine classes is closer to instance discrimination rather than the coarse classes classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.679, + 0.739 + ], + "angle": 0, + "content": "4.5.2 Stanford Cars196" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Stanford Cars196 is another widely used fine-grained image classification benchmark. As there are no official annotated coarse labels, we manually group 196 car models into 8 coarse classes based on the common types of cars: '0: Cab', '1: Sedan', '2: SUV', '3: Convertible', '4: Coupe', '5: Hatchback', '6: Wagon' and '7: Van'." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Tab. 6, our method again has the best performance. In addition, unlike in the SOP datasets, SupCon and SupCE have an overwhelming advantage over SelfCon. In this dataset, this is not surprising because for each fine class," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19919" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.093, + 0.826, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.459, + 0.7, + 0.473 + ], + "angle": 0, + "content": "Figure 4. Examples of Top-10 image retrieval results on Cars196 dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.561 + ], + "angle": 0, + "content": "there are more images (ranging from 24 images to 68 images) with different colours/backgrounds in the cars dataset, leading to a much higher intra-class variance. Some image retrieval examples can be found in Fig. 4." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.575, + 0.458, + 0.692 + ], + "angle": 0, + "content": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon20.9728.7541.4452.61
Graft42.3054.7971.181.74
SupCon42.3054.7971.181.74
CoIns42.7755.6072.2982.53
SupCE42.7755.6072.2982.53
SupFINE78.0982.9786.5188.04
MaskCon (Ours)45.53 (2.76†)58.5674.3684.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.702, + 0.409, + 0.715 + ], + "angle": 0, + "content": "Table 6. Results on Stanford Cars196 dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.281, + 0.754 + ], + "angle": 0, + "content": "4.5.3 Additional discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Compared to the state-of-the-art, our method achieves the best results in all experiments. We would like to note here some of our insights from the results on the CIFAR and the fine-grained ones (including ImageNet-1K). More specifically, we note that on the CIFAR datasets the performance of our method approaches that of the supervised one with fine labels (SupFINE), while on the fine-grained datasets there is a larger gap. We note that the critical question is to consider whether the instance discrimination task is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.668 + ], + "angle": 0, + "content": "useful for fine-grained classification in different scenarios, since coarse labelling usually does not lead to worse results. Towards that, recent works have shown that the core idea of self-supervised contrastive learning — augmentation invariance may be destructive for fine-grained tasks [8, 36]. For example, the commonly applied colour distortion augmentation will promote the model to be non-sensitive to colour information. However, colour may be the key to discriminate between different breeds of birds. How to adapt the instance discrimination task to the fine-grained coarse-labelled dataset is a future direction we wish to pursue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.619, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In this work, we propose a Masked Contrastive learning framework (MaskCon) for learning fine-grained information with coarse-labelled datasets. On the basis of two baseline methods, we utilize coarse labels and the instance discrimination task to better estimate inter-sample relations. We show theoretically that our method can reduce the optimization error bound. Extensive experiments with various hyperparameter settings on multiple benchmarks, including the CIFAR datasets and the more challenging fine-grained classification datasets show that our method achieves consistent and large improvement over the baselines." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: This work was supported by the EU H2020 AI4Media No. 951911 project." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19920" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Guy Bukchin, Eli Schwartz, Kate Saenko, Ori Shahar, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. Fine-grained angular contrastive learning with coarse labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8730-8740, June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. arXiv preprint arXiv:1707.08819, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.472, + 0.592 + ], + "angle": 0, + "content": "[8] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Ried-miller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[11] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5423, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[12] Chen Feng and Ioannis Patras. Adaptive soft contrastive learning. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 2721-2727, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Jhair Gallardo, Tyler L Hayes, and Christopher Kanan. Self-supervised training enhances online continual learning. arXiv preprint arXiv:2103.14010, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[15] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8059-8068, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[16] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[17] Jacob Goldberger, Geoffrey E Hinton, Sam Roweis, and Russ R Salakhutdinov. Neighbourhood components analysis. Advances in neural information processing systems, 17, 2004. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[20] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[22] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsivash. Mean shift for self-supervised learning. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10306-10315, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[23] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[24] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[25] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[26] George A Miller. WordNet: An electronic lexical database. MIT press, 1998. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[27] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[28] Ruslan Salakhutdinov and Geoff Hinton. Learning a nonlinear embedding by preserving class neighbourhood struc" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19921" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "ture. In Artificial Intelligence and Statistics, pages 412-419. PMLR, 2007. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[29] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[30] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[31] Hugo Touvron, Alexandre Sablayrolles, Matthijs Douze, Matthieu Cord, and Hervé Jégou. Graft: Learning fine-grained image representations with coarse labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 874-884, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.386 + ], + "angle": 0, + "content": "[32] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.389, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[33] Guangrun Wang, Keze Wang, Guangcong Wang, Philip HS Torr, and Liang Lin. Solving inefficiency of self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9505-9515, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[34] Xudong Wang, Ziwei Liu, and Stella X. Yu. Unsupervised feature learning by cross-level instance-group discrimination. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12581-12590, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[35] Longhui Wei, Lingxi Xie, Jianzhong He, Jianlong Chang, Xiaopeng Zhang, Wengang Zhou, Houqiang Li, and Qi Tian. Can semantic labels assist self-supervised visual representation learning? arXiv preprint arXiv:2011.08621, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[36] Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darell. What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[37] Yuanhong Xu, Qi Qian, Hao Li, Rong Jin, and Juhua Hu. Weakly supervised representation learning with coarse labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10593-10601, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[38] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1476-1485, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.811 + ], + "angle": 0, + "content": "[39] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16660-16669, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19922" + } + ] +] \ No newline at end of file diff --git a/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_origin.pdf b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..161fc0f3ec5bb63684ad19848fa375bb9b4a5e67 --- /dev/null +++ b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/4e73c1aa-93be-4570-b85f-4e3d7202a290_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3143d69a69e1e2dd9217c6d89bffc19354653133a8683f1b2321cf28f927943 +size 2465094 diff --git a/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/full.md b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..20ff1343e62287c6bb9a61f4d827e6bbddd3b8bf --- /dev/null +++ b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/full.md @@ -0,0 +1,363 @@ +# MaskCon: Masked Contrastive Learning for Coarse-Labelled Dataset + +Chen Feng, Ioannis Patras +Queen Mary University of London, UK + +{chen.feng, i.patras}@qmul.ac.uk + +# Abstract + +Deep learning has achieved great success in recent years with the aid of advanced neural network structures and large-scale human-annotated datasets. However, it is often costly and difficult to accurately and efficiently annotate large-scale datasets, especially for some specialized domains where fine-grained labels are required. In this setting, coarse labels are much easier to acquire as they do not require expert knowledge. In this work, we propose a contrastive learning method, called masked contrastive learning (MaskCon) to address the under-explored problem setting, where we learn with a coarse-labelled dataset in order to address a finer labelling problem. More specifically, within the contrastive learning framework, for each sample our method generates soft-labels with the aid of coarse labels against other samples and another augmented view of the sample in question. By contrast to self-supervised contrastive learning where only the sample's augmentations are considered hard positives, and in supervised contrastive learning where only samples with the same coarse labels are considered hard positives, we propose soft labels based on sample distances, that are masked by the coarse labels. This allows us to utilize both inter-sample relations and coarse labels. We demonstrate that our method can obtain as special cases many existing state-of-the-art works and that it provides tighter bounds on the generalization error. Experimentally, our method achieves significant improvement over the current state-of-the-art in various datasets, including CIFAR10, CIFAR100, ImageNet-1K, Stanford Online Products and Stanford Cars196 datasets. Code and annotations are available at https://github.com/MrChenFeng/MaskCon_CVPR2023. + +# 1. Introduction + +Supervised learning with deep neural networks has achieved great success in various computer vision tasks such as image classification, action detection and object localization. However, the success of supervised learning relies on large-scale and high-quality human- + +![](images/2e24bf4d31d76eadea5bd76eebbc512b3bfd0c539e06bf653b132248009e9aaa.jpg) +Figure 1. Contrastive learning sample relations using MaskCon (ours) and other learning paradigms when only coarse labels are available. MaskCon are closer to the fine ones. + +annotated datasets, whose annotations are time-consuming and labour-intensive to produce. To avoid such reliance, various learning frameworks have been proposed and investigated: Self-supervised learning aims to learn meaningful representations with heuristic proxy visual tasks, such as rotation prediction [16] and the more prevalent instance discrimination task, the latter, being widely applied in self-supervised contrastive learning framework; semi-supervised learning usually considers a dataset for which only a small part is annotated - within this setting, pseudo labelling methods [24] and consistency regularization techniques [1, 29] are typically used; Moreover, learning using more accessible but noisy data, such as web-crawled data, has also received increasing attention [13, 25]. + +In this work, we consider an under-explored problem setting aiming at reducing the annotation effort - learning fine-grained representations with a coarsely-labelled dataset. Specifically, we learn with a dataset that is fully labeled, albeit at a coarser granularity than we are interested in (i.e., that of the test set). Compared to fine-grained labels, coarse labels are often significantly easier to obtain, especially in some of the more specialized domains, such as the recognition and classification of medical pathology images. As a simple example, for the task of differentiation between different pets, we need a knowledgeable cat lover to distinguish between 'British short' and 'Siamese', + +but even a child annotator may help to discriminate between 'cat' and 'non-cat' (Fig. 1). Unfortunately, learning with a coarse labelled dataset has been less investigated compared to other weakly supervised learning paradigms. Recently, Bukchi et al. [2] investigate on learning with coarse labels in the few-shot setting. More closely related to us, Grafit [31] proposes a multi-task framework by a weighted combination of self-supervised contrastive learning and supervised contrastive learning cost; Similarly, CoIns [37] uses both a self-supervised contrastive learning cost and a supervised learning cross-entropy loss. Both works combine a fully supervised learning cost (cross entropy or contrastive) with a self-supervised contrastive loss – these works are the main ones with which we compare. + +Differently than them, instead of using self-supervised contrastive learning as an auxiliary task, we propose a novel learning scheme, namely Masked Contrastive Learning (MaskCon). Our method aims to learn by considering inter-sample relations of each sample with other samples in the dataset. Specifically, we always consider the relation to oneself as confidently positive. To estimate the relations to other samples, we derive soft labels by contrasting an augmented view of the sample in question with other samples, and further improve it by utilizing the mask generated based on the coarse labels. Our approach generates soft inter-sample relations that can more accurately estimate fine inter-sample relations compared to the baseline methods (Fig. 1). Efficiently and effectively, our method achieves significant improvements over the state-of-the-art in various datasets, including CIFARtoy, CIFAR100, ImageNet-1K and more challenging fine-grained datasets Stanford Online Products and Stanford Cars196. + +# 2. Related works + +In this section, we first briefly review the contrastive learning works which are the basis of our method and then, we briefly introduce the state-of-the-art works in related problem settings. + +Self-supervised contrastive learning Contrastive learning has recently emerged as a powerful method for representation learning without labels. In the context of self-supervision, contrastive learning has been used for instance discrimination, producing models whose performance rivals that of fully supervised training. The objective of contrastive instance discrimination is to minimize the distance between transformed views of the same sample in the feature space, while maximizing their distance from views of other samples. Notable methods in this area are SimCLR [5] and MoCo [18], as well as other variants including SwAV [4], which leverages clustering to identify prototypes for the contrastive learning objective and [33], that reformulates the instance discrimination contrastive objective in the + +form of a triplet loss. These methods have been found to produce models that, requiring only limited amounts of annotated data for fine-tuning on a particular task or domain, achieve performance that reaches or even surpasses that of supervised models [6, 11]. Furthermore, self-supervised training has been found to compare positively with supervised training in other respects, such as producing models that perform better in the context of continual learning [14]. + +However, self-supervised contrastive learning based on instance discrimination task usually suffers from the 'false negative' problem – that samples from the same class should not be considered as negative. For example, pictures of different cats should not be considered completely negative to each other. To this end, heuristically modifying the inter-sample relations have been widely applied in self-supervised learning [10, 12, 22, 34], intuitively similar to us. Many of the state-of-the-art methods can be considered as special cases of our method by the adjustment of hyperparameters (Sec. 3.2). + +The success of self-supervised contrastive learning has led to increasing attention on how to integrate contrastive learning into existing supervised learning paradigms. One approach is supervised contrastive learning, which adapts contrastive learning to the fully supervised setting. Supervised contrastive learning was first introduced in [17, 28] under the name of neighbourhood components analysis (NCA). Recently, supervised contrastive learning has regained interest due to the significant progress made in [20] by employing more advanced network structures and image augmentations. Compared to standard supervised learning using cross-entropy loss, supervised contrastive learning exhibits superior performance in hard sample mining [20]. + +Other works have attempted to introduce self-supervised contrastive learning as an auxiliary task, especially for weakly supervised learning problems. Gidaris et al. [15] jointly train a supervised classifier and a rotation-prediction objective. Komodakis and Gidaris [21] seeking to improve model performance on few-shot learning. Wei et al. [35] propose a label-filtered instance discrimination objective for the pre-training of models to be used for transfer learning in other datasets. S4 [38] uses rotation and exemplar [9] prediction to improve performance in semi-supervised learning tasks. Feng et al. [13] apply a self-supervised contrastive loss to safely learn meaningful representations in the presence of noisy labels. + +Hierarchical image classification & Unsupervised clustering Hierarchical image classification methods aim to learn better representations with datasets having hierarchically structured labels. To utilize information from different levels of labels, Zhang et al. [39] propose a multi-losses + +contrastive learning framework with each loss considering a specific level of labels. Unsupervised clustering methods often have the same problem setting but different goals than self-supervised learning, that is, to recover and identify the ground-truth semantic labels of each sample [3, 32]. Deep-Cluster [3] simultaneously learns the clustering assignments and a parametric classifier, by feeding the clustering assignments as pseudo labels to the classifier. SCAN [32] inherits the idea of DeepClustering while augmenting extra self-labelling and neighbor mining techniques. + +# 3. Methodology + +We consider the problem of learning when the labels of the training data and the labels of the test data are inconsistent in granularity, and in particular when the labels of the training set are coarser. More specifically, let us denote with $\mathbf{X} = \{\pmb{x}_i\}_{i=1}^N$ an i.i.d sampled train dataset with annotated coarse labels $\mathbf{Y} = \{\pmb{y}_i \in \{0,1\}^M\}_{i=1}^N$ subject to $\sum_{m=1}^{M} \pmb{y}_{im} = 1$ . Here, $N$ denotes the number of samples and $M$ denotes the number of coarse classes. Let us also define a finely labelled set $\mathbf{Y}' = \{\pmb{y}_i' \in \{0,1\}^{M'}\}_{i=1}^N$ subject to $\sum_{k=1}^{M'} \pmb{y}_i^k = 1$ , where $M'$ denotes the number of fine classes. Typically, $M' > M$ . + +# 3.1. Preliminaries + +We first quickly review the common learning paradigms utilized in the baseline methods and our method. + +# 3.1.1 Supervised learning + +In most supervised learning frameworks, the objective is the minimization of the below empirical risk: + +$$ +R (f, g) = \sum_ {i = 1} ^ {N} L _ {c e} \left(\boldsymbol {x} _ {i}, \boldsymbol {y} _ {i}; f, g\right), \tag {1} +$$ + +where $f$ denotes the feature encoder, and $g$ denotes the classifier head (usually a single fully-connected layer). + +For brevity, we denote $\pmb{f}_i \triangleq f(\pmb{x}_i)$ , $\pmb{g}_i \triangleq g(f(\pmb{x}_i))$ and $\pmb{p}_i \triangleq \mathrm{softmax}(\pmb{g}_i)$ as the feature, logit and prediction of a sample view $\pmb{x}_i$ respectively. Here, + +$$ +L _ {c e} \left(\boldsymbol {x} _ {i}, \boldsymbol {y} _ {i}; f, g\right) = - \sum_ {m = 1} ^ {M} \boldsymbol {y} _ {i} ^ {m} \log \boldsymbol {p} _ {i} ^ {m} \tag {2} +$$ + +is the cross-entropy loss function - the default loss function for classification problems. + +# 3.1.2 Contrastive learning + +Unlike the common supervised learning model above where a parametric classifier $g$ is learned based on the semantic labels $Y$ of the samples $X$ , we can also perform contrastive + +learning based on the inter-sample relations $Z = \{z_{i} \in (0,1)^{N}\}_{i=1}^{N}$ , with each entry $z_{ij}$ depicting the inter-sample relation between $x_{i}$ and $x_{j}$ . Intuitively, $z_{ij} = 1$ means that sample $x_{i}$ and $x_{j}$ generate a strong positive pair. Since each sample may form multiple positive sample pairs, for brevity, we abuse the notation here with $Z$ denoting also the sample-wise normalized inter-sample relations. To learn such intersample relations, instead of a parametric classifier $g$ , the $f$ encoder is usually followed by a projector $h$ , which is often implemented as an MLP and learned by regularizing the inter-sample relations $Z$ (Eq. (4)). + +More specifically, let us denote by $\pmb{h}_i \triangleq h(f(\pmb{x}_i))$ the projection. We first calculate the cosine similarity $\pmb{d}_i$ between a sample $\pmb{x}_i$ and the dataset $\pmb{H} = \{\pmb{h}_n\}_{n=1}^N$ : + +$$ +\boldsymbol {d} _ {i} = \left[ \cos \left(\boldsymbol {h} _ {i}, \boldsymbol {h} _ {1}\right), \cos \left(\boldsymbol {h} _ {i}, \boldsymbol {h} _ {2}\right), \dots , \cos \left(\boldsymbol {h} _ {i}, \boldsymbol {h} _ {N}\right) \right], \tag {3} +$$ + +Let us further define $\pmb{q}_i \triangleq \mathrm{softmax}(\pmb{d}_i / \tau_0)$ , where $\tau_0$ is the temperature hyperparameter. Then the following empirical risk will be optimized: + +$$ +R (f, h) = \sum_ {i = 1} ^ {N} L _ {\text {c o n}} \left(\boldsymbol {x} _ {i}, \boldsymbol {z} _ {i}; f, h\right), \tag {4} +$$ + +where the contrastive loss $L_{con}$ is defined as follows: + +$$ +L _ {c o n} \left(\boldsymbol {x} _ {i}, \boldsymbol {z} _ {i}; f, h\right) = - \sum_ {n = 1} ^ {N} \boldsymbol {z} _ {i} ^ {n} \log \boldsymbol {q} _ {i} ^ {n}. \tag {5} +$$ + +Self-supervised contrastive learning We first introduce the most prevalent form of contrastive learning currently — self-supervised contrastive learning for learning without labels. Since there are no annotated labels, we usually set $Z^{self}$ by considering each sample having only itself as positive: + +$$ +z _ {i j} ^ {\text {s e l f}} = \left\{ \begin{array}{l l} 1, & \text {i f} i = j \\ 0, & \text {i f} i \neq j \end{array} \right. \tag {6} +$$ + +Typically, we aim at maximizing the relations between different augmented views of the same sample while minimizing the relations between augmented views of different samples. This is also widely known as instance discrimination task. We denote such self-supervised contrastive loss as $L_{selfcon}$ . + +Supervised contrastive learning We can easily transition from self-supervised contrastive learning to supervised contrastive learning, by simply changing the inter-sample relations $Z^{sup}$ for Eq. (4) and Eq. (5). More specifically, as in [17, 20, 28], samples from the same semantic classes are considered positive pairs, + +$$ +z _ {i j} ^ {\text {s u p}} = \left\{ \begin{array}{l l} 1, & \text {i f} \boldsymbol {y} _ {i} = \boldsymbol {y} _ {j} \\ 0, & \text {i f} \boldsymbol {y} _ {i} \neq \boldsymbol {y} _ {j} \end{array} \right. \tag {7} +$$ + +in addition to the sample itself (and its augmented views) as in self-supervised contrastive learning. We denote the supervised contrastive loss as $L_{supcon}$ . + +Memory bank For consistency with the supervised cross-entropy loss and for clarity, we present all the essential formulas above in terms of the entire dataset $\mathbf{X}$ . However, due to computation and memory constraints, it is often unrealistic to contrast a specific sample $\mathbf{x}_i$ with the whole dataset. In the actual implementation in this work, following MoCo [18], we use a dynamic FIFO memory bank $\mathbf{H} = \{\mathbf{h}_p\}_{p=1}^P$ consisting of cached projections of $P$ samples. For each sample $\mathbf{x}_i$ , we generate two random augmented views as $\mathbf{x}_q$ (query) and $\mathbf{x}_k$ (key) with their corresponding projections denoted by $\mathbf{h}_q$ and $\mathbf{h}_k$ , respectively. Then, we calculate the cosine similarity between $\mathbf{h}_q$ on the one hand, and $\mathbf{h}_k$ and each projection in the memory bank on the other: + +$$ +\boldsymbol {d} _ {i} = \overbrace {\left[ \cos \left(\boldsymbol {h} _ {q} , \boldsymbol {h} _ {k}\right) , \overbrace {\cos \left(\boldsymbol {h} _ {q} , \boldsymbol {h} _ {1}\right) , \dots , \cos \left(\boldsymbol {h} _ {q} , \boldsymbol {h} _ {P}\right)} ^ {\text {s e l f}} \right]} ^ {\text {m e m o r y b a n k}}, \tag {8} +$$ + +For more details, please refer to the MoCo paper [18]. + +# 3.1.3 Baseline methods + +Clearly, with respect to the finer target question labels, each sample $\boldsymbol{x}_i$ , is a trustworthy positive sample in relation to itself (i.e., they have the same fine-grained label). However, considering the samples with the same coarse label as positives may lead to under-clustering issues (Fig. 3) (they may have different fine-grained labels). Recent works attempt to address such under-clustering by combining an additional self-supervised contrastive loss with a specific supervised loss that utilizes coarse labels. Grafit [31] combine it with the supervised contrastive loss and CoIns [37] with the supervised cross-entropy loss. In both cases, as a result, the self-supervised contrastive loss that considers each sample as a class by itself, aims to mitigate the tendency for under-clustering. Formally, given our definitions of the previously mentioned loss functions in Sec. 3.1 ( $L_{ce}$ , $L_{selfcon}$ and $L_{supcon}$ ) it turns out that the losses used by Grafit [31] and CoIns [37] can be expressed as follows: + +$$ +L _ {G r a f i t} = w L _ {s u p c o n} + (1 - w) L _ {s e l f c o n} \tag {9} +$$ + +$$ +L _ {C o I n s} = w L _ {c e} + (1 - w) L _ {s e l f c o n} \tag {10} +$$ + +where $w$ controls the relative weight of each loss. + +# 3.2. MaskCon: Masked Contrastive learning + +Instead of equally weighing all the samples within the same coarse classes, we aim to emphasize the samples with the same fine labels and reduce the importance of the others. To achieve this, we introduce a novel contrastive learning method, namely Masked Contrastive learning (MaskCon), within the framework of contrastive learning that utilizes inter-sample relations directly. + +More specifically, for sample $\boldsymbol{x}_i$ , we estimate its intersample relations $\boldsymbol{z}_i'$ to other samples utilizing the key view projection $\boldsymbol{h}_k$ and the whole dataset $\{\boldsymbol{h}_1, \dots, \boldsymbol{h}_N\}$ excluding itself (since it will always be considered as a trustworthy positive), as below: + +$$ +z _ {i j} ^ {\prime} = \frac {\mathbb {1} \left(\boldsymbol {y} _ {j} = \boldsymbol {y} _ {i}\right) \cdot \exp \left(d _ {i j} ^ {\prime} / \tau\right)}{\sum_ {n = 1 , n \neq i} ^ {N} \mathbb {1} \left(\boldsymbol {y} _ {n} = \boldsymbol {y} _ {i}\right) \cdot \exp \left(d _ {i n} ^ {\prime} / \tau\right)}, i \neq j, \tag {11} +$$ + +where the similarity $d_i^{\prime}$ is given by + +$$ +\boldsymbol {d} _ {i} ^ {\prime} = \left[ \cos \left(\boldsymbol {h} _ {i} ^ {k}, \boldsymbol {h} _ {1}\right), \dots , \cos \left(\boldsymbol {h} _ {i} ^ {k}, \boldsymbol {h} _ {i - 1}\right), \right. \tag {12} +$$ + +$$ +\left. \cos \left(\boldsymbol {h} _ {i} ^ {k}, \boldsymbol {h} _ {i + 1}\right), \dots , \cos \left(\boldsymbol {h} _ {i} ^ {k}, \boldsymbol {h} _ {N}\right) \right]. +$$ + +Please note the use of the mask $(\mathbb{1}(\pmb {y}_j = \pmb {y}_i))$ that excludes from the softmax that estimates inter-sample relationships, the samples $j$ that have a different coarse label with the sample $i$ (and sets their $z_{ij}^{\prime}$ to 0). While it is risky to consider all samples from the same coarse class as positive, we can confidently identify those samples that do not have the same coarse class as negative. This reduces the noise in $z_i^\prime$ . Finally, we re-scale the $z_{i}^{\prime}$ with its maximum + +$$ +z _ {i j} ^ {\prime} = z _ {i j} ^ {\prime} / \max \left(\boldsymbol {z} _ {i} ^ {\prime}\right), \tag {13} +$$ + +to make the closest neighbour as positive as the sample itself and arrive at: + +$$ +z _ {i j} ^ {\text {m a s k}} = \left\{ \begin{array}{l l} 1, & \text {i f} i = j \\ z _ {i j} ^ {\prime}, & \text {i f} i \neq j \end{array} \right. \tag {14} +$$ + +Compared to $Z^{supcon}$ , we thus reweight the samples of the same coarse label according to the similarities in the feature space. + +We denote the masked contrastive loss as $L_{maskcon}$ and, similarly to Grafit and CoIns, we also consider a weighted loss as the final objective: + +$$ +L = w L _ {\text {m a s k c o n}} + (1 - w) L _ {\text {s e l f c o n}} \tag {15} +$$ + +Relations to SOTA works By adjusting $w$ and $\tau$ , our method can obtain various existing SOTA methods as special cases. More specifically, by setting $\tau$ to $\infty$ , our method degenerates to Grafit [31]. Ignoring the mask, i.e., treating all samples as having the same coarse label, our method can also generalize to existing SOTAs in self-supervised learning. For example, NNCLR [10] instead of considering only each sample itself as positive it considers as positive the nearest neighbors as well. Formally, by setting $w = 1$ and $\tau = 0$ , MaskCon degenerates to NNCLR; ASCL [12] proposes to introduce extra positive samples by calculating soft inter-sample relations, adaptively weighted by its own normalized entropy. By setting $w = \text{norm\_entropy}(z_i')$ , MaskCon degenerates to ASCL. + +# 3.3. Theoretical justification of MaskCon + +The fundamental goal of contrastive learning is to identify $Z$ that suits the problem granularity and accurately represents the relationship between samples. Assuming that there exists an optimal hidden $\hat{Z}$ , we can define the expected population risk in terms of $\hat{Z}$ as shown in Eq. (16). + +$$ +R _ {E} (f, h) = E _ {\boldsymbol {x} _ {i}} \left[ L _ {\text {c o n}} \left(\boldsymbol {x} _ {i}, \hat {\boldsymbol {z}} _ {i}; f, h\right) \right] \tag {16} +$$ + +Moreover, let us define the empirical risk with respect to $\hat{Z}$ as shown in Eq. (17). + +$$ +\hat {R} (f, h) = \sum_ {i = 1} ^ {N} L _ {\text {c o n}} \left(\boldsymbol {x} _ {i}, \hat {\boldsymbol {z}} _ {i}; f, h\right) \tag {17} +$$ + +THEOREM 1. The generalization error bounds of contrastive learning depends on the difference between the inter-sample relations $\mathbf{Z}$ and the optimal hidden intersample relations $\hat{\mathbf{Z}}$ , i.e., + +$$ +d _ {z} = \frac {1}{n} \sum_ {i = 1} ^ {n} \| \boldsymbol {z} _ {i} - \hat {\boldsymbol {z}} _ {i} \| _ {2} \tag {18} +$$ + +Proof. Let us assume $L_{con}(\pmb{x}_i, \pmb{z}_i; f, h) \in [a, b]$ and is $\lambda$ -Lipschitz continuous w.r.t $\pmb{z}_i$ . Let $|\mathcal{F}|$ and $|\mathcal{H}|$ be the covering numbers that correspond to the finite hypothesis space of $f$ and $g$ . We have: + +$$ +\begin{array}{l} \left| R _ {E} (f, h) - R (f, h) \right| \leq \overbrace {\left| R _ {E} (f , h) - \hat {R} (f , h) \right|} ^ {\text {H o e f f d i n g ' s i n e q u a l i t y}} \\ L i p s c h i t z \quad c o n t i n u o u s \\ + \overbrace {| \hat {R} (f , h) - R (f , h) |} \\ \leq | a - b | \sqrt {\frac {\log (2 | \mathcal {F} | \cdot | \mathcal {H} | / \delta)}{2 n}} \\ + \lambda \frac {1}{n} \sum_ {i = 1} ^ {n} \| z _ {i} - \hat {z} _ {i} \| _ {2} \\ \end{array} +$$ + +with probability at least $1 - \delta$ + +![](images/08164868a0cfd8e4ddc80431a588496722dbb4323ac57bb5138258b55e39bdf2.jpg) + +Based on Theorem 1, it can be concluded that our method performs better than Grafit with high probability, because an appropriate temperature $\tau$ allows $Z^{mask}$ to more accurately estimate the optimal $\hat{Z}$ (Fig. 1) in comparison to assigning equal weights to all coarse neighbours. + +# 4. Experiments + +In this section, extensive experiments are conducted to validate the findings and effectiveness of the proposed Masked Contrastive Learning (MaskCon) framework. The impact of hyperparameters is thoroughly investigated in Sec. 4.1. In Sec. 4.2, Sec. 4.3 and Sec. 4.4, results are provided for experiments on CIFAR datasets and ImageNet-1K + +dataset with coarse labels. In Sec. 4.5, we apply our method on more challenging fine-grained datasets, including Stanford Online Products (SOP) dataset [30] and Stanford Cars 196 dataset [23]. + +To ensure the robustness of the proposed method and exclude the impact of model capacity, the model settings are kept consistent throughout all the experiments, except for the two hyperparameters, $w$ and $\tau$ , which are ablated in Sec. 4.1. ResNet18 [19] is employed in all experiments, with the initial convolutional layer modified to have a kernel size of $3 \times 3$ and a stride of 1, and the initial max-pooling operator removed for CIFAR experiments, to account for the smaller image size $(32 \times 32)$ as suggested in prior works [18]. More implementation and dataset details can be found in APPENDIX A. + +We compare our method with two competing methods: Grafit and CoIns. For a fair comparison, we exhaust the weight $w$ choices for both methods and report the best achievable results in all experiments. Note that when $w = 0$ , Grafit and CoIns degenerate to self-supervised contrastive learning denoted as SelfCon; Conversely, when $w = 1$ , Grafit degenerates to supervised contrastive learning [20] denoted as SupCon, while CoIns degenerates to conventional supervised cross-entropy learning denoted as SupCE. For reference, we also show the results when training with fine labels – this is denoted as SupFINE. + +Evaluation protocol To evaluate the different methods on the test set with fine labels, we use the recall@K [27] metric widely used in the image retrieval task. Each test image first retrieves top-K nearest neighbours from the test set and receives 1 if there exists at least one image from the same fine class among the top-K nearest neighbours, otherwise 0. Recall@K averages this score over all the test images. + +![](images/35584c21fd18971c3edb37faf03fd9f1e8bcf6ac7342541c0ea33460bf31a10e.jpg) +Figure 2. Recall@1 w.r.t $w$ and $\tau$ on CIFAR100 dataset + +# 4.1. Effect of $w$ and $\tau$ + +In this section, we extensively ablate the effect of hyperparameters of MaskCon - weight $w$ and temperature $\tau$ . + +
MethodCIFARtoy-goodsplitCIFARtoy-badsplit
Recall@1Recall@2Recall@5Recall@10Recall@1Recall@2Recall@5Recall@10
SelfCon84.8391.5596.3598.1684.8391.5596.3598.16
Graft86.6192.3397.0198.3889.9694.3697.6198.10
SupCon73.8484.2592.1495.4684.6690.9395.1596.71
CoIns86.1592.7697.2198.4690.5594.9497.7398.71
SupCE76.3085.2694.6597.4687.1592.8596.7898.34
SupFINE94.1196.5398.2598.9694.1196.5398.2598.96
MaskCon (Ours)90.28 (13.98↑)94.0497.3398.5391.56 (4.41↑)95.2397.7098.70
+ +Table 1. Results on CIFARtoy dataset. + +![](images/178b97c5e77a93fea5399a99b0608387bcab4cb3c812a2245fb34c12f6b6f53a.jpg) +Figure 3. t-SNE visualization of learned representation on CIFARtoy dataset. + +![](images/1567f0ddfea23c34bd0014d92139a0f7445dc4f315629ce625b5ea2e0366b730.jpg) + +![](images/d066dcae930ba7b28f5ef8fc1f1e1ed82805a53a770e66fd9ea74eb5f5a851d1.jpg) + +![](images/49148ea15d2793ba7d457ca504516144ea116adf981bdf6e85964ed110c5d239.jpg) + +Specifically, we show the recall@1 results on CIFAR100 dataset in Fig. 2, with $w = \{0, 0.2, 0.5, 0.8, 1.0\}$ and $\tau = \{0, 0.01, 0.05, 0.1, 0.5, \infty\}$ . + +The results demonstrate a significant improvement in performance upon the inclusion of additional inter-sample relationships ( $w \neq 0$ ) in comparison to self-supervised learning alone ( $w = 0$ ) on the CIFAR dataset. Furthermore, a suitable temperature $\tau$ , such as 0.05 or 0.1, consistently yields superior results compared to the simple weighted combination in Grafit ( $\tau = \infty$ ). In general, we recommend initiating the hyperparameter search with $w = 1$ and $\tau = \tau_0$ , where $\tau_0$ denotes the temperature for the projection head. Additional results concerning various hyperparameters on other datasets can be found in APPENDIX B. + +# 4.2. Experiments on CIFARtoy dataset + +To simulate the coarse labelling process, we manually generate toy datasets based on the CIFAR10 dataset. Following the problem motivation, a subset of 8 classes is selected from the original 10 classes. Specifically, the classes 'airplane', 'truck', 'automobile', and 'ship' are designated as Class A, while 'horse', 'dog', 'bird', and 'cat' are assigned to Class B. Note that Class A comprises non-organic objects, while Class B comprises animals. Moreover, a bad split of the aforementioned dataset is defined, wherein 'airplane', 'automobile', 'bird', and 'cat' are Class A, and 'horse', 'dog', 'ship', and 'truck' are Class B. + +As shown in Tab. 1, our method achieves the best performance in both splits. Moreover, the improvement over the supervised learning with coarse labels (SupCE) on good split $(13.98\%)$ is much higher than the bad split $(4.41\%)$ , + +which shows the substantial potential of MaskCon in dealing with more realistic coarse-labelled datasets1. + +In Fig. 3 we visualize the learned features of all test samples in the good split. Clearly, the learned representations with supervised contrastive learning (SupCon) and self-supervised contrastive learning (SelfCon) tend to under-cluster and over-cluster the samples, respectively. By contrast, our method gets more compact and clear clusters, in line with the results when trained with fine labels (SupFINE). + +# 4.3. Experiments on CIFAR100 dataset + +The common CIFAR100 dataset has 20 classes of coarse labels in addition to the 100 classes of fine labels, with each coarse class containing five fine-grained classes (500 samples). The results in Tab. 2 show that our method achieves significant improvements over the SOTAs. In particular, it improves the top-1 retrieval precision from $47.25\%$ to $65.52\%$ , approaching the results by the model learned with fine labels $(71.13\%)$ . + +# 4.4. Experiments on ImageNet-1K dataset + +In this section, we evaluate our method on the largescale ImageNet-1K dataset. For efficiency, we experiment with the downsampled version of ImageNet-1K dataset [7], where each sample was resized to $32 \times 32$ . Since no official coarse labels exist for ImageNet, we introduce coarse + +
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon40.5051.8366.2376.66
Graft60.5771.1382.3289.21
SupCon58.6570.0482.1889.09
CoIns60.1070.8983.1489.52
SupCE47.2561.2477.7887.01
SupFINE71.1380.0387.6191.59
MaskCon (Ours)65.52 (18.17†)74.4683.6489.25
+ +labels based on the WordNet [26] hierarchy so as to artificially group the whole dataset into 12 coarse classes: '0: Invertebrate', '1: Domestic animal', '2: Bird', '3: Mammal', '4: Reptile/Aquatic vertebrate', '5: Device', '6: Vehicle', '7: Container', '8: Instrument', '9: Artifact', '10: Clothing' and '11: Others'. In Tab. 3, we show the results on down-sampled ImageNet-1K datasets. + +Table 2. Results on CIFAR100 dataset. + +
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon10.2814.1522.3630.34
Graft18.1325.4637.1946.64
SupCon13.3619.4029.7739.38
CoIns18.3625.5437.0946.89
SupCE12.2314.1527.7637.03
SupFINE33.9744.5557.2365.77
MaskCon (Ours)19.08 (6.86↑)26.2138.1747.96
+ +We can again validate that with temperature-controlled soft relations, our method surpasses SupCE, Grafit and CoIns consistently, especially for top-10 retrieval results. + +# 4.5. Experiments on fine-grained datasets + +In this section, we conduct experiments in a more challenging scenario - fine-grained datasets with only coarse labels. Please note, that in this work, we do not aim to compare with the state-of-the-art works in fine-grained classification, which usually involve more specialized techniques, such as object localization and local feature extraction. + +# 4.5.1 Stanford Online Products (SOP) + +The Stanford Online Products (SOP) dataset [30] consists of 22,634 products, with each product having between two and twelve photos from different perspectives, for a total of 120,053 images. In addition, there are 12 coarse classes based on the semantic categories of the products, such as 'bicycle' and 'kettle'. For the common image retrieval task, the category of the test set is possibly unknown, so we firstly extract a subset with an unknown test category based on the SOP dataset. We then selected the categories with eight or more images. We then split almost equally this subset to obtain the training set and the test set with 25,368 and 25,278 + +images, consisting of 2,517 and 2,518 classes respectively, denoted as SOP-split1. Moreover, we select all products with twelve images, and then randomly select ten train images and two test images. Thus, we have 1,498 classes with a total of 17,976 images (2,996 test images and 14,980 train images), denoted as SOP-split2. + +Table 3. Results on ImageNet-1K dataset. + +
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon70.3675.5781.5385.13
Graft74.0278.8284.1387.91
SupCon53.6959.5567.1272.78
CoIns70.8476.0182.286.08
SupCE36.3542.3950.3056.52
SupFINE83.9488.0491.9594.00
MaskCon (Ours)74.05 (37.7↑)78.9784.4887.96
+ +Table 4. Results on SOP-split1 dataset. + +
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon35.8541.4649.7756.11
Graft39.1244.6653.1059.65
SupCon25.0729.2435.8541.59
CoIns38.2245.1954.3761.18
SupCE22.5626.3433.2838.95
SupFINE69.5675.7083.2487.88
MaskCon (Ours)45.36 (22.8†)51.0758.9165.52
+ +Table 5. Results on SOP-split2 dataset. + +In Tab. 4 and Tab. 5 we show results on SOP-split1 and SOP-split2, respectively. Similarly, regardless of whether the test class is known or not, our method still achieves significant improvements over state-of-the-art methods. An interesting phenomenon here is, that on the SOP dataset, SelfCon alone is significantly better than its SupCE/SupCon counterpart. We conjecture that this is due to the fact that each fine class of the SOP dataset in fact consists of different views of the same product, and the number of images per fine class is quite small (2 to 12). For such a dataset, the classification of fine classes is closer to instance discrimination rather than the coarse classes classification. + +# 4.5.2 Stanford Cars196 + +Stanford Cars196 is another widely used fine-grained image classification benchmark. As there are no official annotated coarse labels, we manually group 196 car models into 8 coarse classes based on the common types of cars: '0: Cab', '1: Sedan', '2: SUV', '3: Convertible', '4: Coupe', '5: Hatchback', '6: Wagon' and '7: Van'. + +In Tab. 6, our method again has the best performance. In addition, unlike in the SOP datasets, SupCon and SupCE have an overwhelming advantage over SelfCon. In this dataset, this is not surprising because for each fine class, + +![](images/58813795238c845a958de7eb7661293ea55d43e8983f551fd6346fa83bd9390b.jpg) +Figure 4. Examples of Top-10 image retrieval results on Cars196 dataset. + +there are more images (ranging from 24 images to 68 images) with different colours/backgrounds in the cars dataset, leading to a much higher intra-class variance. Some image retrieval examples can be found in Fig. 4. + +
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon20.9728.7541.4452.61
Graft42.3054.7971.181.74
SupCon42.3054.7971.181.74
CoIns42.7755.6072.2982.53
SupCE42.7755.6072.2982.53
SupFINE78.0982.9786.5188.04
MaskCon (Ours)45.53 (2.76†)58.5674.3684.36
+ +Table 6. Results on Stanford Cars196 dataset. + +# 4.5.3 Additional discussion + +Compared to the state-of-the-art, our method achieves the best results in all experiments. We would like to note here some of our insights from the results on the CIFAR and the fine-grained ones (including ImageNet-1K). More specifically, we note that on the CIFAR datasets the performance of our method approaches that of the supervised one with fine labels (SupFINE), while on the fine-grained datasets there is a larger gap. We note that the critical question is to consider whether the instance discrimination task is + +useful for fine-grained classification in different scenarios, since coarse labelling usually does not lead to worse results. Towards that, recent works have shown that the core idea of self-supervised contrastive learning — augmentation invariance may be destructive for fine-grained tasks [8, 36]. For example, the commonly applied colour distortion augmentation will promote the model to be non-sensitive to colour information. However, colour may be the key to discriminate between different breeds of birds. How to adapt the instance discrimination task to the fine-grained coarse-labelled dataset is a future direction we wish to pursue. + +# 5. Conclusion + +In this work, we propose a Masked Contrastive learning framework (MaskCon) for learning fine-grained information with coarse-labelled datasets. On the basis of two baseline methods, we utilize coarse labels and the instance discrimination task to better estimate inter-sample relations. We show theoretically that our method can reduce the optimization error bound. Extensive experiments with various hyperparameter settings on multiple benchmarks, including the CIFAR datasets and the more challenging fine-grained classification datasets show that our method achieves consistent and large improvement over the baselines. + +Acknowledgments: This work was supported by the EU H2020 AI4Media No. 951911 project. + +# References + +[1] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1 +[2] Guy Bukchin, Eli Schwartz, Kate Saenko, Ori Shahar, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. Fine-grained angular contrastive learning with coarse labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8730-8740, June 2021. 2 +[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3 +[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2 +[7] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. arXiv preprint arXiv:1707.08819, 2017. 6 +[8] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 8 +[9] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Ried-miller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. Advances in neural information processing systems, 27, 2014. 2 +[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. 2, 4 +[11] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5423, 2021. 2 +[12] Chen Feng and Ioannis Patras. Adaptive soft contrastive learning. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 2721-2727, 2022. 2, 4 +[13] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 1, 2 + +[14] Jhair Gallardo, Tyler L Hayes, and Christopher Kanan. Self-supervised training enhances online continual learning. arXiv preprint arXiv:2103.14010, 2021. 2 +[15] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8059-8068, 2019. 2 +[16] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 1 +[17] Jacob Goldberger, Geoffrey E Hinton, Sam Roweis, and Russ R Salakhutdinov. Neighbourhood components analysis. Advances in neural information processing systems, 17, 2004. 2, 3 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2, 4, 5 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[20] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3, 5 +[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. 2 +[22] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsivash. Mean shift for self-supervised learning. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10306-10315, 2021. 2 +[23] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 5 +[24] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1 +[25] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1 +[26] George A Miller. WordNet: An electronic lexical database. MIT press, 1998. 7 +[27] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. 5 +[28] Ruslan Salakhutdinov and Geoff Hinton. Learning a nonlinear embedding by preserving class neighbourhood struc + +ture. In Artificial Intelligence and Statistics, pages 412-419. PMLR, 2007. 2, 3 +[29] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020. 1 +[30] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 7 +[31] Hugo Touvron, Alexandre Sablayrolles, Matthijs Douze, Matthieu Cord, and Hervé Jégou. Graft: Learning fine-grained image representations with coarse labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 874-884, 2021. 2, 4 +[32] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 3 +[33] Guangrun Wang, Keze Wang, Guangcong Wang, Philip HS Torr, and Liang Lin. Solving inefficiency of self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9505-9515, 2021. 2 +[34] Xudong Wang, Ziwei Liu, and Stella X. Yu. Unsupervised feature learning by cross-level instance-group discrimination. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12581-12590, 2021. 2 +[35] Longhui Wei, Lingxi Xie, Jianzhong He, Jianlong Chang, Xiaopeng Zhang, Wengang Zhou, Houqiang Li, and Qi Tian. Can semantic labels assist self-supervised visual representation learning? arXiv preprint arXiv:2011.08621, 2020. 2 +[36] Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darell. What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659, 2020. 8 +[37] Yuanhong Xu, Qi Qian, Hao Li, Rong Jin, and Juhua Hu. Weakly supervised representation learning with coarse labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10593-10601, 2021. 2, 4 +[38] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1476-1485, 2019. 2 +[39] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16660-16669, 2022. 2 \ No newline at end of file diff --git a/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/images.zip b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..23679a98588646bb8d02e29c3c35ac8c3484aa4f --- /dev/null +++ b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:663f264c3b2dc027e96467804c6658435d0c572c313cef8740d0b5dcfc07b99c +size 584467 diff --git a/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/layout.json b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ff3e68f93ffc0a0a6faa852542857ceeea13d06 --- /dev/null +++ b/2023/MaskCon_ Masked Contrastive Learning for Coarse-Labelled Dataset/layout.json @@ -0,0 +1,9207 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 512, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 512, + 120 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 512, + 120 + ], + "type": "text", + "content": "MaskCon: Masked Contrastive Learning for Coarse-Labelled Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 200, + 144, + 392, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 144, + 392, + 171 + ], + "spans": [ + { + "bbox": [ + 200, + 144, + 392, + 171 + ], + "type": "text", + "content": "Chen Feng, Ioannis Patras \nQueen Mary University of London, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 174, + 381, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 174, + 381, + 185 + ], + "spans": [ + { + "bbox": [ + 209, + 174, + 381, + 185 + ], + "type": "text", + "content": "{chen.feng, i.patras}@qmul.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 239, + 290, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 610 + ], + "type": "text", + "content": "Deep learning has achieved great success in recent years with the aid of advanced neural network structures and large-scale human-annotated datasets. However, it is often costly and difficult to accurately and efficiently annotate large-scale datasets, especially for some specialized domains where fine-grained labels are required. In this setting, coarse labels are much easier to acquire as they do not require expert knowledge. In this work, we propose a contrastive learning method, called masked contrastive learning (MaskCon) to address the under-explored problem setting, where we learn with a coarse-labelled dataset in order to address a finer labelling problem. More specifically, within the contrastive learning framework, for each sample our method generates soft-labels with the aid of coarse labels against other samples and another augmented view of the sample in question. By contrast to self-supervised contrastive learning where only the sample's augmentations are considered hard positives, and in supervised contrastive learning where only samples with the same coarse labels are considered hard positives, we propose soft labels based on sample distances, that are masked by the coarse labels. This allows us to utilize both inter-sample relations and coarse labels. We demonstrate that our method can obtain as special cases many existing state-of-the-art works and that it provides tighter bounds on the generalization error. Experimentally, our method achieves significant improvement over the current state-of-the-art in various datasets, including CIFAR10, CIFAR100, ImageNet-1K, Stanford Online Products and Stanford Cars196 datasets. Code and annotations are available at https://github.com/MrChenFeng/MaskCon_CVPR2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 633, + 128, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 633, + 128, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 633, + 128, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Supervised learning with deep neural networks has achieved great success in various computer vision tasks such as image classification, action detection and object localization. However, the success of supervised learning relies on large-scale and high-quality human-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 215, + 545, + 336 + ], + "blocks": [ + { + "bbox": [ + 307, + 215, + 545, + 336 + ], + "lines": [ + { + "bbox": [ + 307, + 215, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 215, + 545, + 336 + ], + "type": "image", + "image_path": "2e24bf4d31d76eadea5bd76eebbc512b3bfd0c539e06bf653b132248009e9aaa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 346, + 546, + 378 + ], + "lines": [ + { + "bbox": [ + 305, + 346, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 546, + 378 + ], + "type": "text", + "content": "Figure 1. Contrastive learning sample relations using MaskCon (ours) and other learning paradigms when only coarse labels are available. MaskCon are closer to the fine ones." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 402, + 546, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 569 + ], + "type": "text", + "content": "annotated datasets, whose annotations are time-consuming and labour-intensive to produce. To avoid such reliance, various learning frameworks have been proposed and investigated: Self-supervised learning aims to learn meaningful representations with heuristic proxy visual tasks, such as rotation prediction [16] and the more prevalent instance discrimination task, the latter, being widely applied in self-supervised contrastive learning framework; semi-supervised learning usually considers a dataset for which only a small part is annotated - within this setting, pseudo labelling methods [24] and consistency regularization techniques [1, 29] are typically used; Moreover, learning using more accessible but noisy data, such as web-crawled data, has also received increasing attention [13, 25]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "In this work, we consider an under-explored problem setting aiming at reducing the annotation effort - learning fine-grained representations with a coarsely-labelled dataset. Specifically, we learn with a dataset that is fully labeled, albeit at a coarser granularity than we are interested in (i.e., that of the test set). Compared to fine-grained labels, coarse labels are often significantly easier to obtain, especially in some of the more specialized domains, such as the recognition and classification of medical pathology images. As a simple example, for the task of differentiation between different pets, we need a knowledgeable cat lover to distinguish between 'British short' and 'Siamese'," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19913" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "but even a child annotator may help to discriminate between 'cat' and 'non-cat' (Fig. 1). Unfortunately, learning with a coarse labelled dataset has been less investigated compared to other weakly supervised learning paradigms. Recently, Bukchi et al. [2] investigate on learning with coarse labels in the few-shot setting. More closely related to us, Grafit [31] proposes a multi-task framework by a weighted combination of self-supervised contrastive learning and supervised contrastive learning cost; Similarly, CoIns [37] uses both a self-supervised contrastive learning cost and a supervised learning cross-entropy loss. Both works combine a fully supervised learning cost (cross entropy or contrastive) with a self-supervised contrastive loss – these works are the main ones with which we compare." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 455 + ], + "type": "text", + "content": "Differently than them, instead of using self-supervised contrastive learning as an auxiliary task, we propose a novel learning scheme, namely Masked Contrastive Learning (MaskCon). Our method aims to learn by considering inter-sample relations of each sample with other samples in the dataset. Specifically, we always consider the relation to oneself as confidently positive. To estimate the relations to other samples, we derive soft labels by contrasting an augmented view of the sample in question with other samples, and further improve it by utilizing the mask generated based on the coarse labels. Our approach generates soft inter-sample relations that can more accurately estimate fine inter-sample relations compared to the baseline methods (Fig. 1). Efficiently and effectively, our method achieves significant improvements over the state-of-the-art in various datasets, including CIFARtoy, CIFAR100, ImageNet-1K and more challenging fine-grained datasets Stanford Online Products and Stanford Cars196." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 464, + 136, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 136, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 136, + 477 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 533 + ], + "type": "text", + "content": "In this section, we first briefly review the contrastive learning works which are the basis of our method and then, we briefly introduce the state-of-the-art works in related problem settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 546, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 289, + 713 + ], + "type": "text", + "content": "Self-supervised contrastive learning Contrastive learning has recently emerged as a powerful method for representation learning without labels. In the context of self-supervision, contrastive learning has been used for instance discrimination, producing models whose performance rivals that of fully supervised training. The objective of contrastive instance discrimination is to minimize the distance between transformed views of the same sample in the feature space, while maximizing their distance from views of other samples. Notable methods in this area are SimCLR [5] and MoCo [18], as well as other variants including SwAV [4], which leverages clustering to identify prototypes for the contrastive learning objective and [33], that reformulates the instance discrimination contrastive objective in the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "form of a triplet loss. These methods have been found to produce models that, requiring only limited amounts of annotated data for fine-tuning on a particular task or domain, achieve performance that reaches or even surpasses that of supervised models [6, 11]. Furthermore, self-supervised training has been found to compare positively with supervised training in other respects, such as producing models that perform better in the context of continual learning [14]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 168, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 546, + 300 + ], + "type": "text", + "content": "However, self-supervised contrastive learning based on instance discrimination task usually suffers from the 'false negative' problem – that samples from the same class should not be considered as negative. For example, pictures of different cats should not be considered completely negative to each other. To this end, heuristically modifying the inter-sample relations have been widely applied in self-supervised learning [10, 12, 22, 34], intuitively similar to us. Many of the state-of-the-art methods can be considered as special cases of our method by the adjustment of hyperparameters (Sec. 3.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 315, + 546, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 546, + 483 + ], + "type": "text", + "content": "The success of self-supervised contrastive learning has led to increasing attention on how to integrate contrastive learning into existing supervised learning paradigms. One approach is supervised contrastive learning, which adapts contrastive learning to the fully supervised setting. Supervised contrastive learning was first introduced in [17, 28] under the name of neighbourhood components analysis (NCA). Recently, supervised contrastive learning has regained interest due to the significant progress made in [20] by employing more advanced network structures and image augmentations. Compared to standard supervised learning using cross-entropy loss, supervised contrastive learning exhibits superior performance in hard sample mining [20]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 483, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 638 + ], + "type": "text", + "content": "Other works have attempted to introduce self-supervised contrastive learning as an auxiliary task, especially for weakly supervised learning problems. Gidaris et al. [15] jointly train a supervised classifier and a rotation-prediction objective. Komodakis and Gidaris [21] seeking to improve model performance on few-shot learning. Wei et al. [35] propose a label-filtered instance discrimination objective for the pre-training of models to be used for transfer learning in other datasets. S4 [38] uses rotation and exemplar [9] prediction to improve performance in semi-supervised learning tasks. Feng et al. [13] apply a self-supervised contrastive loss to safely learn meaningful representations in the presence of noisy labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Hierarchical image classification & Unsupervised clustering Hierarchical image classification methods aim to learn better representations with datasets having hierarchically structured labels. To utilize information from different levels of labels, Zhang et al. [39] propose a multi-losses" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19914" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "contrastive learning framework with each loss considering a specific level of labels. Unsupervised clustering methods often have the same problem setting but different goals than self-supervised learning, that is, to recover and identify the ground-truth semantic labels of each sample [3, 32]. Deep-Cluster [3] simultaneously learns the clustering assignments and a parametric classifier, by feeding the clustering assignments as pseudo labels to the classifier. SCAN [32] inherits the idea of DeepClustering while augmenting extra self-labelling and neighbor mining techniques." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 202, + 129, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 202, + 129, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 202, + 129, + 216 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": "We consider the problem of learning when the labels of the training data and the labels of the test data are inconsistent in granularity, and in particular when the labels of the training set are coarser. More specifically, let us denote with " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{\\pmb{x}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": " an i.i.d sampled train dataset with annotated coarse labels " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{\\pmb{y}_i \\in \\{0,1\\}^M\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": " subject to " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\sum_{m=1}^{M} \\pmb{y}_{im} = 1" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": " denotes the number of samples and " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": " denotes the number of coarse classes. Let us also define a finely labelled set " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}' = \\{\\pmb{y}_i' \\in \\{0,1\\}^{M'}\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": " subject to " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{M'} \\pmb{y}_i^k = 1" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": " denotes the number of fine classes. Typically, " + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "inline_equation", + "content": "M' > M" + }, + { + "bbox": [ + 46, + 222, + 289, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 365, + 133, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 133, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 133, + 376 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "type": "text", + "content": "We first quickly review the common learning paradigms utilized in the baseline methods and our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 422, + 166, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 166, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 166, + 434 + ], + "type": "text", + "content": "3.1.1 Supervised learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 440, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 287, + 464 + ], + "type": "text", + "content": "In most supervised learning frameworks, the objective is the minimization of the below empirical risk:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 472, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 472, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 101, + 472, + 287, + 506 + ], + "type": "interline_equation", + "content": "R (f, g) = \\sum_ {i = 1} ^ {N} L _ {c e} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {y} _ {i}; f, g\\right), \\tag {1}", + "image_path": "f7012f10357d70992071011e5ad951f59dde1ecc46717e71103d560b782b307d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "content": " denotes the feature encoder, and " + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "content": " denotes the classifier head (usually a single fully-connected layer)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "text", + "content": "For brevity, we denote " + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i \\triangleq f(\\pmb{x}_i)" + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i \\triangleq g(f(\\pmb{x}_i))" + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i \\triangleq \\mathrm{softmax}(\\pmb{g}_i)" + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "text", + "content": " as the feature, logit and prediction of a sample view " + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 47, + 536, + 287, + 572 + ], + "type": "text", + "content": " respectively. Here," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 581, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 90, + 581, + 287, + 613 + ], + "type": "interline_equation", + "content": "L _ {c e} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {y} _ {i}; f, g\\right) = - \\sum_ {m = 1} ^ {M} \\boldsymbol {y} _ {i} ^ {m} \\log \\boldsymbol {p} _ {i} ^ {m} \\tag {2}", + "image_path": "2720b802a2e0b0ecd8f9f72027723c4e0bdb56d439a5107b71b879d9708b734d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 620, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 287, + 644 + ], + "type": "text", + "content": "is the cross-entropy loss function - the default loss function for classification problems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 658, + 168, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 168, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 168, + 671 + ], + "type": "text", + "content": "3.1.2 Contrastive learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Unlike the common supervised learning model above where a parametric classifier " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " is learned based on the semantic labels " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " of the samples " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": ", we can also perform contrastive" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "learning based on the inter-sample relations " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "Z = \\{z_{i} \\in (0,1)^{N}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": ", with each entry " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "z_{ij}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " depicting the inter-sample relation between " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "z_{ij} = 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " means that sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " generate a strong positive pair. Since each sample may form multiple positive sample pairs, for brevity, we abuse the notation here with " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " denoting also the sample-wise normalized inter-sample relations. To learn such intersample relations, instead of a parametric classifier " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " encoder is usually followed by a projector " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": ", which is often implemented as an MLP and learned by regularizing the inter-sample relations " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " (Eq. (4))." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "text", + "content": "More specifically, let us denote by " + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\pmb{h}_i \\triangleq h(f(\\pmb{x}_i))" + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "text", + "content": " the projection. We first calculate the cosine similarity " + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\pmb{d}_i" + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "text", + "content": " between a sample " + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "text", + "content": " and the dataset " + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\pmb{H} = \\{\\pmb{h}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 203, + 545, + 241 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 246, + 545, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 246, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 320, + 246, + 545, + 259 + ], + "type": "interline_equation", + "content": "\\boldsymbol {d} _ {i} = \\left[ \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {1}\\right), \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {2}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i}, \\boldsymbol {h} _ {N}\\right) \\right], \\tag {3}", + "image_path": "7e36e11ee118d0882a6ec000b4089b354555b6efdcdb9823a991eb4c99bc380d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "text", + "content": "Let us further define " + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{q}_i \\triangleq \\mathrm{softmax}(\\pmb{d}_i / \\tau_0)" + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 304, + 265, + 545, + 300 + ], + "type": "text", + "content": " is the temperature hyperparameter. Then the following empirical risk will be optimized:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 357, + 306, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 306, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 357, + 306, + 545, + 338 + ], + "type": "interline_equation", + "content": "R (f, h) = \\sum_ {i = 1} ^ {N} L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {z} _ {i}; f, h\\right), \\tag {4}", + "image_path": "4299d3cf950a23509e786cafb5b1604d40ed34a48335b7ce0e5ba722cbe5c840.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 344, + 520, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 520, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 520, + 356 + ], + "type": "text", + "content": "where the contrastive loss " + }, + { + "bbox": [ + 306, + 344, + 520, + 356 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 306, + 344, + 520, + 356 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 348, + 361, + 545, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 361, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 348, + 361, + 545, + 393 + ], + "type": "interline_equation", + "content": "L _ {c o n} \\left(\\boldsymbol {x} _ {i}, \\boldsymbol {z} _ {i}; f, h\\right) = - \\sum_ {n = 1} ^ {N} \\boldsymbol {z} _ {i} ^ {n} \\log \\boldsymbol {q} _ {i} ^ {n}. \\tag {5}", + "image_path": "1f66a50a6635af3b48c7681e693a6d9ea89d8bde1f7ffc28d5e2702c8f78772c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 398, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 468 + ], + "type": "text", + "content": "Self-supervised contrastive learning We first introduce the most prevalent form of contrastive learning currently — self-supervised contrastive learning for learning without labels. Since there are no annotated labels, we usually set " + }, + { + "bbox": [ + 304, + 398, + 545, + 468 + ], + "type": "inline_equation", + "content": "Z^{self}" + }, + { + "bbox": [ + 304, + 398, + 545, + 468 + ], + "type": "text", + "content": " by considering each sample having only itself as positive:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 373, + 468, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 468, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 373, + 468, + 545, + 493 + ], + "type": "interline_equation", + "content": "z _ {i j} ^ {\\text {s e l f}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} i = j \\\\ 0, & \\text {i f} i \\neq j \\end{array} \\right. \\tag {6}", + "image_path": "ce5d06a7fc24b694bc2536fb426c67c7c3a4cbfe009918a7908657f5bdc40e72.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": "Typically, we aim at maximizing the relations between different augmented views of the same sample while minimizing the relations between augmented views of different samples. This is also widely known as instance discrimination task. We denote such self-supervised contrastive loss as " + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "inline_equation", + "content": "L_{selfcon}" + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": "Supervised contrastive learning We can easily transition from self-supervised contrastive learning to supervised contrastive learning, by simply changing the inter-sample relations " + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "inline_equation", + "content": "Z^{sup}" + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": " for Eq. (4) and Eq. (5). More specifically, as in [17, 20, 28], samples from the same semantic classes are considered positive pairs," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 369, + 643, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 643, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 369, + 643, + 545, + 670 + ], + "type": "interline_equation", + "content": "z _ {i j} ^ {\\text {s u p}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\\\ 0, & \\text {i f} \\boldsymbol {y} _ {i} \\neq \\boldsymbol {y} _ {j} \\end{array} \\right. \\tag {7}", + "image_path": "e3fc9474e8d70ee09e65e2f172a1bef32145e8f71780b63a3c7c0614a87d64a0.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 674, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 545, + 712 + ], + "type": "text", + "content": "in addition to the sample itself (and its augmented views) as in self-supervised contrastive learning. We denote the supervised contrastive loss as " + }, + { + "bbox": [ + 304, + 674, + 545, + 712 + ], + "type": "inline_equation", + "content": "L_{supcon}" + }, + { + "bbox": [ + 304, + 674, + 545, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19915" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": "Memory bank For consistency with the supervised cross-entropy loss and for clarity, we present all the essential formulas above in terms of the entire dataset " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": ". However, due to computation and memory constraints, it is often unrealistic to contrast a specific sample " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": " with the whole dataset. In the actual implementation in this work, following MoCo [18], we use a dynamic FIFO memory bank " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = \\{\\mathbf{h}_p\\}_{p=1}^P" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": " consisting of cached projections of " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": " samples. For each sample " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": ", we generate two random augmented views as " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_q" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": " (query) and " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": " (key) with their corresponding projections denoted by " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_q" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_k" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": ", respectively. Then, we calculate the cosine similarity between " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_q" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": " on the one hand, and " + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_k" + }, + { + "bbox": [ + 48, + 71, + 286, + 238 + ], + "type": "text", + "content": " and each projection in the memory bank on the other:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 60, + 245, + 285, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 245, + 285, + 275 + ], + "spans": [ + { + "bbox": [ + 60, + 245, + 285, + 275 + ], + "type": "interline_equation", + "content": "\\boldsymbol {d} _ {i} = \\overbrace {\\left[ \\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {k}\\right) , \\overbrace {\\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {1}\\right) , \\dots , \\cos \\left(\\boldsymbol {h} _ {q} , \\boldsymbol {h} _ {P}\\right)} ^ {\\text {s e l f}} \\right]} ^ {\\text {m e m o r y b a n k}}, \\tag {8}", + "image_path": "8fcfb1481617f5528cbdf6acba5085637353357c1809c6498d344ed92b29359a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 283, + 271, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 283, + 271, + 295 + ], + "spans": [ + { + "bbox": [ + 49, + 283, + 271, + 295 + ], + "type": "text", + "content": "For more details, please refer to the MoCo paper [18]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 316, + 155, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 155, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 155, + 327 + ], + "type": "text", + "content": "3.1.3 Baseline methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "text", + "content": "Clearly, with respect to the finer target question labels, each sample " + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i" + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "text", + "content": ", is a trustworthy positive sample in relation to itself (i.e., they have the same fine-grained label). However, considering the samples with the same coarse label as positives may lead to under-clustering issues (Fig. 3) (they may have different fine-grained labels). Recent works attempt to address such under-clustering by combining an additional self-supervised contrastive loss with a specific supervised loss that utilizes coarse labels. Grafit [31] combine it with the supervised contrastive loss and CoIns [37] with the supervised cross-entropy loss. In both cases, as a result, the self-supervised contrastive loss that considers each sample as a class by itself, aims to mitigate the tendency for under-clustering. Formally, given our definitions of the previously mentioned loss functions in Sec. 3.1 (" + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "inline_equation", + "content": "L_{selfcon}" + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "inline_equation", + "content": "L_{supcon}" + }, + { + "bbox": [ + 47, + 335, + 287, + 539 + ], + "type": "text", + "content": ") it turns out that the losses used by Grafit [31] and CoIns [37] can be expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 550, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 550, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 83, + 550, + 287, + 563 + ], + "type": "interline_equation", + "content": "L _ {G r a f i t} = w L _ {s u p c o n} + (1 - w) L _ {s e l f c o n} \\tag {9}", + "image_path": "c5ae9d7361992b097819e19ffccba57ea3669349ddb0c28a4a00fae71162a2a6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 566, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 566, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 86, + 566, + 287, + 578 + ], + "type": "interline_equation", + "content": "L _ {C o I n s} = w L _ {c e} + (1 - w) L _ {s e l f c o n} \\tag {10}", + "image_path": "4a664db60319bb287c273d0dc46e9b5fb9b3f38d28b5f23cfad1af136c16d0ba.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 589, + 247, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 247, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 247, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 589, + 247, + 601 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 589, + 247, + 601 + ], + "type": "text", + "content": " controls the relative weight of each loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 610, + 259, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 259, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 259, + 624 + ], + "type": "text", + "content": "3.2. MaskCon: Masked Contrastive learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": "Instead of equally weighing all the samples within the same coarse classes, we aim to emphasize the samples with the same fine labels and reduce the importance of the others. To achieve this, we introduce a novel contrastive learning method, namely Masked Contrastive learning (MaskCon), within the framework of contrastive learning that utilizes inter-sample relations directly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "More specifically, for sample " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ", we estimate its intersample relations " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_i'" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " to other samples utilizing the key view projection " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\boldsymbol{h}_k" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and the whole dataset " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{h}_1, \\dots, \\boldsymbol{h}_N\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " excluding itself (since it will always be considered as a trustworthy positive), as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 140, + 545, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 140, + 545, + 172 + ], + "type": "interline_equation", + "content": "z _ {i j} ^ {\\prime} = \\frac {\\mathbb {1} \\left(\\boldsymbol {y} _ {j} = \\boldsymbol {y} _ {i}\\right) \\cdot \\exp \\left(d _ {i j} ^ {\\prime} / \\tau\\right)}{\\sum_ {n = 1 , n \\neq i} ^ {N} \\mathbb {1} \\left(\\boldsymbol {y} _ {n} = \\boldsymbol {y} _ {i}\\right) \\cdot \\exp \\left(d _ {i n} ^ {\\prime} / \\tau\\right)}, i \\neq j, \\tag {11}", + "image_path": "c4f954f0be13ddb021bb5055c880b15d754c6f100f51dda4b65f046a40738336.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 181, + 447, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 181, + 447, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 181, + 447, + 193 + ], + "type": "text", + "content": "where the similarity " + }, + { + "bbox": [ + 306, + 181, + 447, + 193 + ], + "type": "inline_equation", + "content": "d_i^{\\prime}" + }, + { + "bbox": [ + 306, + 181, + 447, + 193 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 201, + 545, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 201, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 347, + 201, + 545, + 222 + ], + "type": "interline_equation", + "content": "\\boldsymbol {d} _ {i} ^ {\\prime} = \\left[ \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {1}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {i - 1}\\right), \\right. \\tag {12}", + "image_path": "374bf25efd49bab95057d492b614234fb518a7cfca9cbb6d260845b9a176b816.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 220, + 504, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 220, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 370, + 220, + 504, + 233 + ], + "type": "interline_equation", + "content": "\\left. \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {i + 1}\\right), \\dots , \\cos \\left(\\boldsymbol {h} _ {i} ^ {k}, \\boldsymbol {h} _ {N}\\right) \\right].", + "image_path": "9e39d384827ef45bf83a6574eff7b65b9c41f99e30796d455cd095251549a989.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "text", + "content": "Please note the use of the mask " + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "inline_equation", + "content": "(\\mathbb{1}(\\pmb {y}_j = \\pmb {y}_i))" + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "text", + "content": " that excludes from the softmax that estimates inter-sample relationships, the samples " + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "text", + "content": " that have a different coarse label with the sample " + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "text", + "content": " (and sets their " + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "inline_equation", + "content": "z_{ij}^{\\prime}" + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "text", + "content": " to 0). While it is risky to consider all samples from the same coarse class as positive, we can confidently identify those samples that do not have the same coarse class as negative. This reduces the noise in " + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "inline_equation", + "content": "z_i^\\prime" + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "text", + "content": ". Finally, we re-scale the " + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "inline_equation", + "content": "z_{i}^{\\prime}" + }, + { + "bbox": [ + 305, + 242, + 545, + 338 + ], + "type": "text", + "content": " with its maximum" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 347, + 545, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 347, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 383, + 347, + 545, + 363 + ], + "type": "interline_equation", + "content": "z _ {i j} ^ {\\prime} = z _ {i j} ^ {\\prime} / \\max \\left(\\boldsymbol {z} _ {i} ^ {\\prime}\\right), \\tag {13}", + "image_path": "f8573bbaca7ed9864fe61bba0cd35c904309d2371c57dbf09d381d659b059495.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 370, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 545, + 393 + ], + "type": "text", + "content": "to make the closest neighbour as positive as the sample itself and arrive at:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 368, + 402, + 545, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 402, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 368, + 402, + 545, + 429 + ], + "type": "interline_equation", + "content": "z _ {i j} ^ {\\text {m a s k}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} i = j \\\\ z _ {i j} ^ {\\prime}, & \\text {i f} i \\neq j \\end{array} \\right. \\tag {14}", + "image_path": "26199b772ca32afdfb2318bec9bdc64379147e99cbe782cc817eb08e9bb454bc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 436, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 545, + 471 + ], + "type": "text", + "content": "Compared to " + }, + { + "bbox": [ + 305, + 436, + 545, + 471 + ], + "type": "inline_equation", + "content": "Z^{supcon}" + }, + { + "bbox": [ + 305, + 436, + 545, + 471 + ], + "type": "text", + "content": ", we thus reweight the samples of the same coarse label according to the similarities in the feature space." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 472, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 508 + ], + "type": "text", + "content": "We denote the masked contrastive loss as " + }, + { + "bbox": [ + 305, + 472, + 545, + 508 + ], + "type": "inline_equation", + "content": "L_{maskcon}" + }, + { + "bbox": [ + 305, + 472, + 545, + 508 + ], + "type": "text", + "content": " and, similarly to Grafit and CoIns, we also consider a weighted loss as the final objective:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 351, + 518, + 545, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 518, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 351, + 518, + 545, + 532 + ], + "type": "interline_equation", + "content": "L = w L _ {\\text {m a s k c o n}} + (1 - w) L _ {\\text {s e l f c o n}} \\tag {15}", + "image_path": "3389061b164f643f2e5f95ca2c1a56b2e1e9af4dfe8f111c9f2f3bbcc2ff6495.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": "Relations to SOTA works By adjusting " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ", our method can obtain various existing SOTA methods as special cases. More specifically, by setting " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ", our method degenerates to Grafit [31]. Ignoring the mask, i.e., treating all samples as having the same coarse label, our method can also generalize to existing SOTAs in self-supervised learning. For example, NNCLR [10] instead of considering only each sample itself as positive it considers as positive the nearest neighbors as well. Formally, by setting " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ", MaskCon degenerates to NNCLR; ASCL [12] proposes to introduce extra positive samples by calculating soft inter-sample relations, adaptively weighted by its own normalized entropy. By setting " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "w = \\text{norm\\_entropy}(z_i')" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ", MaskCon degenerates to ASCL." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19916" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "type": "text", + "content": "3.3. Theoretical justification of MaskCon" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "type": "text", + "content": "The fundamental goal of contrastive learning is to identify " + }, + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "type": "text", + "content": " that suits the problem granularity and accurately represents the relationship between samples. Assuming that there exists an optimal hidden " + }, + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "type": "text", + "content": ", we can define the expected population risk in terms of " + }, + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 47, + 90, + 287, + 151 + ], + "type": "text", + "content": " as shown in Eq. (16)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 162, + 287, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 162, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 94, + 162, + 287, + 175 + ], + "type": "interline_equation", + "content": "R _ {E} (f, h) = E _ {\\boldsymbol {x} _ {i}} \\left[ L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {z}} _ {i}; f, h\\right) \\right] \\tag {16}", + "image_path": "c4a8164e99baf7d8d20e36464c2d966da28af416fa34f3141a490b2212d46c46.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 179, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 179, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 179, + 287, + 205 + ], + "type": "text", + "content": "Moreover, let us define the empirical risk with respect to " + }, + { + "bbox": [ + 47, + 179, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 47, + 179, + 287, + 205 + ], + "type": "text", + "content": " as shown in Eq. (17)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 211, + 287, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 211, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 100, + 211, + 287, + 244 + ], + "type": "interline_equation", + "content": "\\hat {R} (f, h) = \\sum_ {i = 1} ^ {N} L _ {\\text {c o n}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {z}} _ {i}; f, h\\right) \\tag {17}", + "image_path": "8f401f5585fe0c36f7bc9b7005380c33708d0bd1180d407ef4bf9e452e61f377.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 251, + 287, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 287, + 299 + ], + "type": "text", + "content": "THEOREM 1. The generalization error bounds of contrastive learning depends on the difference between the inter-sample relations " + }, + { + "bbox": [ + 47, + 251, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 47, + 251, + 287, + 299 + ], + "type": "text", + "content": " and the optimal hidden intersample relations " + }, + { + "bbox": [ + 47, + 251, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 47, + 251, + 287, + 299 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 307, + 287, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 307, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 119, + 307, + 287, + 337 + ], + "type": "interline_equation", + "content": "d _ {z} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| \\boldsymbol {z} _ {i} - \\hat {\\boldsymbol {z}} _ {i} \\| _ {2} \\tag {18}", + "image_path": "d09a20b7c48c38b5e498d53d56464b104a1d98fe3d2e4a0f32e5510a61022eef.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "content": "Proof. Let us assume " + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "inline_equation", + "content": "L_{con}(\\pmb{x}_i, \\pmb{z}_i; f, h) \\in [a, b]" + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "content": " and is " + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "content": "-Lipschitz continuous w.r.t " + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{z}_i" + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "inline_equation", + "content": "|\\mathcal{F}|" + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "inline_equation", + "content": "|\\mathcal{H}|" + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "content": " be the covering numbers that correspond to the finite hypothesis space of " + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 344, + 287, + 392 + ], + "type": "text", + "content": ". We have:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 399, + 272, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 399, + 272, + 525 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 272, + 525 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| R _ {E} (f, h) - R (f, h) \\right| \\leq \\overbrace {\\left| R _ {E} (f , h) - \\hat {R} (f , h) \\right|} ^ {\\text {H o e f f d i n g ' s i n e q u a l i t y}} \\\\ L i p s c h i t z \\quad c o n t i n u o u s \\\\ + \\overbrace {| \\hat {R} (f , h) - R (f , h) |} \\\\ \\leq | a - b | \\sqrt {\\frac {\\log (2 | \\mathcal {F} | \\cdot | \\mathcal {H} | / \\delta)}{2 n}} \\\\ + \\lambda \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| z _ {i} - \\hat {z} _ {i} \\| _ {2} \\\\ \\end{array}", + "image_path": "571b0cc7b86daa1d87b7238289406a8a93c496cf7a5bfeaf59b2cafa5c85034d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 532, + 171, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 171, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 171, + 544 + ], + "type": "text", + "content": "with probability at least " + }, + { + "bbox": [ + 47, + 532, + 171, + 544 + ], + "type": "inline_equation", + "content": "1 - \\delta" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 277, + 533, + 287, + 541 + ], + "blocks": [ + { + "bbox": [ + 277, + 533, + 287, + 541 + ], + "lines": [ + { + "bbox": [ + 277, + 533, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 277, + 533, + 287, + 541 + ], + "type": "image", + "image_path": "08164868a0cfd8e4ddc80431a588496722dbb4323ac57bb5138258b55e39bdf2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "content": "Based on Theorem 1, it can be concluded that our method performs better than Grafit with high probability, because an appropriate temperature " + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "content": " allows " + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "Z^{mask}" + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "content": " to more accurately estimate the optimal " + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "content": " (Fig. 1) in comparison to assigning equal weights to all coarse neighbours." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 621, + 127, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 127, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 127, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "content": "In this section, extensive experiments are conducted to validate the findings and effectiveness of the proposed Masked Contrastive Learning (MaskCon) framework. The impact of hyperparameters is thoroughly investigated in Sec. 4.1. In Sec. 4.2, Sec. 4.3 and Sec. 4.4, results are provided for experiments on CIFAR datasets and ImageNet-1K" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "dataset with coarse labels. In Sec. 4.5, we apply our method on more challenging fine-grained datasets, including Stanford Online Products (SOP) dataset [30] and Stanford Cars 196 dataset [23]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "text", + "content": "To ensure the robustness of the proposed method and exclude the impact of model capacity, the model settings are kept consistent throughout all the experiments, except for the two hyperparameters, " + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "text", + "content": ", which are ablated in Sec. 4.1. ResNet18 [19] is employed in all experiments, with the initial convolutional layer modified to have a kernel size of " + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "text", + "content": " and a stride of 1, and the initial max-pooling operator removed for CIFAR experiments, to account for the smaller image size " + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 304, + 121, + 545, + 251 + ], + "type": "text", + "content": " as suggested in prior works [18]. More implementation and dataset details can be found in APPENDIX A." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "text", + "content": "We compare our method with two competing methods: Grafit and CoIns. For a fair comparison, we exhaust the weight " + }, + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "text", + "content": " choices for both methods and report the best achievable results in all experiments. Note that when " + }, + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "inline_equation", + "content": "w = 0" + }, + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "text", + "content": ", Grafit and CoIns degenerate to self-supervised contrastive learning denoted as SelfCon; Conversely, when " + }, + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "text", + "content": ", Grafit degenerates to supervised contrastive learning [20] denoted as SupCon, while CoIns degenerates to conventional supervised cross-entropy learning denoted as SupCE. For reference, we also show the results when training with fine labels – this is denoted as SupFINE." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 397, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 397, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 545, + 481 + ], + "type": "text", + "content": "Evaluation protocol To evaluate the different methods on the test set with fine labels, we use the recall@K [27] metric widely used in the image retrieval task. Each test image first retrieves top-K nearest neighbours from the test set and receives 1 if there exists at least one image from the same fine class among the top-K nearest neighbours, otherwise 0. Recall@K averages this score over all the test images." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 325, + 503, + 512, + 635 + ], + "blocks": [ + { + "bbox": [ + 325, + 503, + 512, + 635 + ], + "lines": [ + { + "bbox": [ + 325, + 503, + 512, + 635 + ], + "spans": [ + { + "bbox": [ + 325, + 503, + 512, + 635 + ], + "type": "image", + "image_path": "35584c21fd18971c3edb37faf03fd9f1e8bcf6ac7342541c0ea33460bf31a10e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 642, + 527, + 653 + ], + "lines": [ + { + "bbox": [ + 323, + 642, + 527, + 653 + ], + "spans": [ + { + "bbox": [ + 323, + 642, + 527, + 653 + ], + "type": "text", + "content": "Figure 2. Recall@1 w.r.t " + }, + { + "bbox": [ + 323, + 642, + 527, + 653 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 323, + 642, + 527, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 323, + 642, + 527, + 653 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 323, + 642, + 527, + 653 + ], + "type": "text", + "content": " on CIFAR100 dataset" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "type": "text", + "content": "4.1. Effect of " + }, + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "type": "inline_equation", + "content": "\\tau" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In this section, we extensively ablate the effect of hyperparameters of MaskCon - weight " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " and temperature " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19917" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 70, + 501, + 178 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 501, + 178 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 501, + 178 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 501, + 178 + ], + "type": "table", + "html": "
MethodCIFARtoy-goodsplitCIFARtoy-badsplit
Recall@1Recall@2Recall@5Recall@10Recall@1Recall@2Recall@5Recall@10
SelfCon84.8391.5596.3598.1684.8391.5596.3598.16
Graft86.6192.3397.0198.3889.9694.3697.6198.10
SupCon73.8484.2592.1495.4684.6690.9395.1596.71
CoIns86.1592.7697.2198.4690.5594.9497.7398.71
SupCE76.3085.2694.6597.4687.1592.8596.7898.34
SupFINE94.1196.5398.2598.9694.1196.5398.2598.96
MaskCon (Ours)90.28 (13.98↑)94.0497.3398.5391.56 (4.41↑)95.2397.7098.70
", + "image_path": "0afe7e2d18515a77a3448666a4a616b93b2d5fcce67f0e9c021a9e7ee1a218a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 226, + 186, + 366, + 197 + ], + "lines": [ + { + "bbox": [ + 226, + 186, + 366, + 197 + ], + "spans": [ + { + "bbox": [ + 226, + 186, + 366, + 197 + ], + "type": "text", + "content": "Table 1. Results on CIFARtoy dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 107, + 209, + 190, + 286 + ], + "blocks": [ + { + "bbox": [ + 107, + 209, + 190, + 286 + ], + "lines": [ + { + "bbox": [ + 107, + 209, + 190, + 286 + ], + "spans": [ + { + "bbox": [ + 107, + 209, + 190, + 286 + ], + "type": "image", + "image_path": "178b97c5e77a93fea5399a99b0608387bcab4cb3c812a2245fb34c12f6b6f53a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 301, + 436, + 313 + ], + "lines": [ + { + "bbox": [ + 156, + 301, + 436, + 313 + ], + "spans": [ + { + "bbox": [ + 156, + 301, + 436, + 313 + ], + "type": "text", + "content": "Figure 3. t-SNE visualization of learned representation on CIFARtoy dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 209, + 290, + 285 + ], + "blocks": [ + { + "bbox": [ + 205, + 209, + 290, + 285 + ], + "lines": [ + { + "bbox": [ + 205, + 209, + 290, + 285 + ], + "spans": [ + { + "bbox": [ + 205, + 209, + 290, + 285 + ], + "type": "image", + "image_path": "1567f0ddfea23c34bd0014d92139a0f7445dc4f315629ce625b5ea2e0366b730.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 304, + 209, + 388, + 285 + ], + "blocks": [ + { + "bbox": [ + 304, + 209, + 388, + 285 + ], + "lines": [ + { + "bbox": [ + 304, + 209, + 388, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 388, + 285 + ], + "type": "image", + "image_path": "d066dcae930ba7b28f5ef8fc1f1e1ed82805a53a770e66fd9ea74eb5f5a851d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 403, + 209, + 487, + 285 + ], + "blocks": [ + { + "bbox": [ + 403, + 209, + 487, + 285 + ], + "lines": [ + { + "bbox": [ + 403, + 209, + 487, + 285 + ], + "spans": [ + { + "bbox": [ + 403, + 209, + 487, + 285 + ], + "type": "image", + "image_path": "49148ea15d2793ba7d457ca504516144ea116adf981bdf6e85964ed110c5d239.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 334, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 370 + ], + "type": "text", + "content": "Specifically, we show the recall@1 results on CIFAR100 dataset in Fig. 2, with " + }, + { + "bbox": [ + 47, + 334, + 287, + 370 + ], + "type": "inline_equation", + "content": "w = \\{0, 0.2, 0.5, 0.8, 1.0\\}" + }, + { + "bbox": [ + 47, + 334, + 287, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 334, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\tau = \\{0, 0.01, 0.05, 0.1, 0.5, \\infty\\}" + }, + { + "bbox": [ + 47, + 334, + 287, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "content": "The results demonstrate a significant improvement in performance upon the inclusion of additional inter-sample relationships (" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "inline_equation", + "content": "w \\neq 0" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "content": ") in comparison to self-supervised learning alone (" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "inline_equation", + "content": "w = 0" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "content": ") on the CIFAR dataset. Furthermore, a suitable temperature " + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "content": ", such as 0.05 or 0.1, consistently yields superior results compared to the simple weighted combination in Grafit (" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\tau = \\infty" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "content": "). In general, we recommend initiating the hyperparameter search with " + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\tau = \\tau_0" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 46, + 371, + 287, + 502 + ], + "type": "text", + "content": " denotes the temperature for the projection head. Additional results concerning various hyperparameters on other datasets can be found in APPENDIX B." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 513, + 231, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 231, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 231, + 526 + ], + "type": "text", + "content": "4.2. Experiments on CIFARtoy dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": "To simulate the coarse labelling process, we manually generate toy datasets based on the CIFAR10 dataset. Following the problem motivation, a subset of 8 classes is selected from the original 10 classes. Specifically, the classes 'airplane', 'truck', 'automobile', and 'ship' are designated as Class A, while 'horse', 'dog', 'bird', and 'cat' are assigned to Class B. Note that Class A comprises non-organic objects, while Class B comprises animals. Moreover, a bad split of the aforementioned dataset is defined, wherein 'airplane', 'automobile', 'bird', and 'cat' are Class A, and 'horse', 'dog', 'ship', and 'truck' are Class B." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "As shown in Tab. 1, our method achieves the best performance in both splits. Moreover, the improvement over the supervised learning with coarse labels (SupCE) on good split " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "(13.98\\%)" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " is much higher than the bad split " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "(4.41\\%)" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 334, + 544, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 544, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 544, + 357 + ], + "type": "text", + "content": "which shows the substantial potential of MaskCon in dealing with more realistic coarse-labelled datasets1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 357, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 453 + ], + "type": "text", + "content": "In Fig. 3 we visualize the learned features of all test samples in the good split. Clearly, the learned representations with supervised contrastive learning (SupCon) and self-supervised contrastive learning (SelfCon) tend to under-cluster and over-cluster the samples, respectively. By contrast, our method gets more compact and clear clusters, in line with the results when trained with fine labels (SupFINE)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 460, + 492, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 492, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 492, + 472 + ], + "type": "text", + "content": "4.3. Experiments on CIFAR100 dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "type": "text", + "content": "The common CIFAR100 dataset has 20 classes of coarse labels in addition to the 100 classes of fine labels, with each coarse class containing five fine-grained classes (500 samples). The results in Tab. 2 show that our method achieves significant improvements over the SOTAs. In particular, it improves the top-1 retrieval precision from " + }, + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "type": "inline_equation", + "content": "47.25\\%" + }, + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "type": "inline_equation", + "content": "65.52\\%" + }, + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "type": "text", + "content": ", approaching the results by the model learned with fine labels " + }, + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "type": "inline_equation", + "content": "(71.13\\%)" + }, + { + "bbox": [ + 304, + 478, + 545, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 580, + 504, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 580, + 504, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 580, + 504, + 594 + ], + "type": "text", + "content": "4.4. Experiments on ImageNet-1K dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 598, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 545, + 658 + ], + "type": "text", + "content": "In this section, we evaluate our method on the largescale ImageNet-1K dataset. For efficiency, we experiment with the downsampled version of ImageNet-1K dataset [7], where each sample was resized to " + }, + { + "bbox": [ + 304, + 598, + 545, + 658 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 598, + 545, + 658 + ], + "type": "text", + "content": ". Since no official coarse labels exist for ImageNet, we introduce coarse" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "text", + "content": "1It is worth noting that the better performance on the bad split may seem counter-intuitive at first glance. However, note that a bad split actually makes the labels more informative, as it explicitly helps to discriminate visually similar classes, e.g., 'cat' and 'dog' in different coarse classes in our setting." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19918" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 280, + 161 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 280, + 161 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 280, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 280, + 161 + ], + "type": "table", + "html": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon40.5051.8366.2376.66
Graft60.5771.1382.3289.21
SupCon58.6570.0482.1889.09
CoIns60.1070.8983.1489.52
SupCE47.2561.2477.7887.01
SupFINE71.1380.0387.6191.59
MaskCon (Ours)65.52 (18.17†)74.4683.6489.25
", + "image_path": "69cd1925bd0f42b498b6753bb87898584904bb79211376c727ddbc32204c22bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 199, + 287, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 284 + ], + "type": "text", + "content": "labels based on the WordNet [26] hierarchy so as to artificially group the whole dataset into 12 coarse classes: '0: Invertebrate', '1: Domestic animal', '2: Bird', '3: Mammal', '4: Reptile/Aquatic vertebrate', '5: Device', '6: Vehicle', '7: Container', '8: Instrument', '9: Artifact', '10: Clothing' and '11: Others'. In Tab. 3, we show the results on down-sampled ImageNet-1K datasets." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 293, + 280, + 385 + ], + "blocks": [ + { + "bbox": [ + 95, + 169, + 240, + 179 + ], + "lines": [ + { + "bbox": [ + 95, + 169, + 240, + 179 + ], + "spans": [ + { + "bbox": [ + 95, + 169, + 240, + 179 + ], + "type": "text", + "content": "Table 2. Results on CIFAR100 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 293, + 280, + 385 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 280, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 280, + 385 + ], + "type": "table", + "html": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon10.2814.1522.3630.34
Graft18.1325.4637.1946.64
SupCon13.3619.4029.7739.38
CoIns18.3625.5437.0946.89
SupCE12.2314.1527.7637.03
SupFINE33.9744.5557.2365.77
MaskCon (Ours)19.08 (6.86↑)26.2138.1747.96
", + "image_path": "f82298af1f6504929ab3f92730623ae46739d512d9e2509ae16344d2c9f1fa03.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 416, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 453 + ], + "type": "text", + "content": "We can again validate that with temperature-controlled soft relations, our method surpasses SupCE, Grafit and CoIns consistently, especially for top-10 retrieval results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 459, + 244, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 244, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 244, + 472 + ], + "type": "text", + "content": "4.5. Experiments on fine-grained datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 477, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 477, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 287, + 548 + ], + "type": "text", + "content": "In this section, we conduct experiments in a more challenging scenario - fine-grained datasets with only coarse labels. Please note, that in this work, we do not aim to compare with the state-of-the-art works in fine-grained classification, which usually involve more specialized techniques, such as object localization and local feature extraction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 218, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 218, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 218, + 574 + ], + "type": "text", + "content": "4.5.1 Stanford Online Products (SOP)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "The Stanford Online Products (SOP) dataset [30] consists of 22,634 products, with each product having between two and twelve photos from different perspectives, for a total of 120,053 images. In addition, there are 12 coarse classes based on the semantic categories of the products, such as 'bicycle' and 'kettle'. For the common image retrieval task, the category of the test set is possibly unknown, so we firstly extract a subset with an unknown test category based on the SOP dataset. We then selected the categories with eight or more images. We then split almost equally this subset to obtain the training set and the test set with 25,368 and 25,278" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "images, consisting of 2,517 and 2,518 classes respectively, denoted as SOP-split1. Moreover, we select all products with twelve images, and then randomly select ten train images and two test images. Thus, we have 1,498 classes with a total of 17,976 images (2,996 test images and 14,980 train images), denoted as SOP-split2." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 313, + 154, + 538, + 247 + ], + "blocks": [ + { + "bbox": [ + 90, + 393, + 244, + 404 + ], + "lines": [ + { + "bbox": [ + 90, + 393, + 244, + 404 + ], + "spans": [ + { + "bbox": [ + 90, + 393, + 244, + 404 + ], + "type": "text", + "content": "Table 3. Results on ImageNet-1K dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 154, + 538, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 538, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 538, + 247 + ], + "type": "table", + "html": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon70.3675.5781.5385.13
Graft74.0278.8284.1387.91
SupCon53.6959.5567.1272.78
CoIns70.8476.0182.286.08
SupCE36.3542.3950.3056.52
SupFINE83.9488.0491.9594.00
MaskCon (Ours)74.05 (37.7↑)78.9784.4887.96
", + "image_path": "cf01f81c30bb82b660213804d0f3cef27cdc33a87d9ff62699ced4658c2a2ed7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 313, + 289, + 538, + 382 + ], + "blocks": [ + { + "bbox": [ + 353, + 255, + 497, + 266 + ], + "lines": [ + { + "bbox": [ + 353, + 255, + 497, + 266 + ], + "spans": [ + { + "bbox": [ + 353, + 255, + 497, + 266 + ], + "type": "text", + "content": "Table 4. Results on SOP-split1 dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 289, + 538, + 382 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 538, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 538, + 382 + ], + "type": "table", + "html": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon35.8541.4649.7756.11
Graft39.1244.6653.1059.65
SupCon25.0729.2435.8541.59
CoIns38.2245.1954.3761.18
SupCE22.5626.3433.2838.95
SupFINE69.5675.7083.2487.88
MaskCon (Ours)45.36 (22.8†)51.0758.9165.52
", + "image_path": "d4d784ab57b83358ac627e5008d0a5d99a5587b9b9be01813bea92bbd18db312.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 390, + 497, + 401 + ], + "lines": [ + { + "bbox": [ + 353, + 390, + 497, + 401 + ], + "spans": [ + { + "bbox": [ + 353, + 390, + 497, + 401 + ], + "type": "text", + "content": "Table 5. Results on SOP-split2 dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 415, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 558 + ], + "type": "text", + "content": "In Tab. 4 and Tab. 5 we show results on SOP-split1 and SOP-split2, respectively. Similarly, regardless of whether the test class is known or not, our method still achieves significant improvements over state-of-the-art methods. An interesting phenomenon here is, that on the SOP dataset, SelfCon alone is significantly better than its SupCE/SupCon counterpart. We conjecture that this is due to the fact that each fine class of the SOP dataset in fact consists of different views of the same product, and the number of images per fine class is quite small (2 to 12). For such a dataset, the classification of fine classes is closer to instance discrimination rather than the coarse classes classification." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 574, + 415, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 415, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 415, + 585 + ], + "type": "text", + "content": "4.5.2 Stanford Cars196" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": "Stanford Cars196 is another widely used fine-grained image classification benchmark. As there are no official annotated coarse labels, we manually group 196 car models into 8 coarse classes based on the common types of cars: '0: Cab', '1: Sedan', '2: SUV', '3: Convertible', '4: Coupe', '5: Hatchback', '6: Wagon' and '7: Van'." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "In Tab. 6, our method again has the best performance. In addition, unlike in the SOP datasets, SupCon and SupCE have an overwhelming advantage over SelfCon. In this dataset, this is not surprising because for each fine class," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19919" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 73, + 505, + 352 + ], + "blocks": [ + { + "bbox": [ + 86, + 73, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 86, + 73, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 86, + 73, + 505, + 352 + ], + "type": "image", + "image_path": "58813795238c845a958de7eb7661293ea55d43e8983f551fd6346fa83bd9390b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 363, + 428, + 374 + ], + "lines": [ + { + "bbox": [ + 163, + 363, + 428, + 374 + ], + "spans": [ + { + "bbox": [ + 163, + 363, + 428, + 374 + ], + "type": "text", + "content": "Figure 4. Examples of Top-10 image retrieval results on Cars196 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 396, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 444 + ], + "type": "text", + "content": "there are more images (ranging from 24 images to 68 images) with different colours/backgrounds in the cars dataset, leading to a much higher intra-class variance. Some image retrieval examples can be found in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 455, + 280, + 548 + ], + "blocks": [ + { + "bbox": [ + 55, + 455, + 280, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 455, + 280, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 280, + 548 + ], + "type": "table", + "html": "
MethodRecall@1Recall@2Recall@5Recall@10
SelfCon20.9728.7541.4452.61
Graft42.3054.7971.181.74
SupCon42.3054.7971.181.74
CoIns42.7755.6072.2982.53
SupCE42.7755.6072.2982.53
SupFINE78.0982.9786.5188.04
MaskCon (Ours)45.53 (2.76†)58.5674.3684.36
", + "image_path": "f33cc3c655b9d9fc392d6519d64c9dc8274d25ee28df31471457defb30d0773e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 555, + 250, + 566 + ], + "lines": [ + { + "bbox": [ + 83, + 555, + 250, + 566 + ], + "spans": [ + { + "bbox": [ + 83, + 555, + 250, + 566 + ], + "type": "text", + "content": "Table 6. Results on Stanford Cars196 dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 586, + 171, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 171, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 171, + 597 + ], + "type": "text", + "content": "4.5.3 Additional discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 712 + ], + "type": "text", + "content": "Compared to the state-of-the-art, our method achieves the best results in all experiments. We would like to note here some of our insights from the results on the CIFAR and the fine-grained ones (including ImageNet-1K). More specifically, we note that on the CIFAR datasets the performance of our method approaches that of the supervised one with fine labels (SupFINE), while on the fine-grained datasets there is a larger gap. We note that the critical question is to consider whether the instance discrimination task is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 396, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 529 + ], + "type": "text", + "content": "useful for fine-grained classification in different scenarios, since coarse labelling usually does not lead to worse results. Towards that, recent works have shown that the core idea of self-supervised contrastive learning — augmentation invariance may be destructive for fine-grained tasks [8, 36]. For example, the commonly applied colour distortion augmentation will promote the model to be non-sensitive to colour information. However, colour may be the key to discriminate between different breeds of birds. How to adapt the instance discrimination task to the fine-grained coarse-labelled dataset is a future direction we wish to pursue." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 538, + 378, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 378, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 378, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": "In this work, we propose a Masked Contrastive learning framework (MaskCon) for learning fine-grained information with coarse-labelled datasets. On the basis of two baseline methods, we utilize coarse labels and the instance discrimination task to better estimate inter-sample relations. We show theoretically that our method can reduce the optimization error bound. Extensive experiments with various hyperparameter settings on multiple benchmarks, including the CIFAR datasets and the more challenging fine-grained classification datasets show that our method achieves consistent and large improvement over the baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments: This work was supported by the EU H2020 AI4Media No. 951911 project." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19920" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 200 + ], + "type": "text", + "content": "[2] Guy Bukchin, Eli Schwartz, Kate Saenko, Ori Shahar, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. Fine-grained angular contrastive learning with coarse labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8730-8740, June 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "text", + "content": "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 413 + ], + "type": "text", + "content": "[7] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. arXiv preprint arXiv:1707.08819, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[8] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Ried-miller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 578 + ], + "type": "text", + "content": "[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[11] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5423, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "text", + "content": "[12] Chen Feng and Ioannis Patras. Adaptive soft contrastive learning. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 2721-2727, 2022. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[13] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Jhair Gallardo, Tyler L Hayes, and Christopher Kanan. Self-supervised training enhances online continual learning. arXiv preprint arXiv:2103.14010, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "text", + "content": "[15] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8059-8068, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[16] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[17] Jacob Goldberger, Geoffrey E Hinton, Sam Roweis, and Russ R Salakhutdinov. Neighbourhood components analysis. Advances in neural information processing systems, 17, 2004. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "type": "text", + "content": "[20] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 397, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 545, + 441 + ], + "type": "text", + "content": "[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "type": "text", + "content": "[22] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsivash. Mean shift for self-supervised learning. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10306-10315, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 488, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 533 + ], + "type": "text", + "content": "[23] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "text", + "content": "[24] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "text", + "content": "[25] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 612, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 635 + ], + "type": "text", + "content": "[26] George A Miller. WordNet: An electronic lexical database. MIT press, 1998. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[27] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[28] Ruslan Salakhutdinov and Geoff Hinton. Learning a nonlinear embedding by preserving class neighbourhood struc" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19921" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 642 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "ture. In Artificial Intelligence and Statistics, pages 412-419. PMLR, 2007. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[29] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[30] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "text", + "content": "[31] Hugo Touvron, Alexandre Sablayrolles, Matthijs Douze, Matthieu Cord, and Hervé Jégou. Graft: Learning fine-grained image representations with coarse labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 874-884, 2021. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 305 + ], + "type": "text", + "content": "[32] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "text", + "content": "[33] Guangrun Wang, Keze Wang, Guangcong Wang, Philip HS Torr, and Liang Lin. Solving inefficiency of self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9505-9515, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "text", + "content": "[34] Xudong Wang, Ziwei Liu, and Stella X. Yu. Unsupervised feature learning by cross-level instance-group discrimination. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12581-12590, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "text", + "content": "[35] Longhui Wei, Lingxi Xie, Jianzhong He, Jianlong Chang, Xiaopeng Zhang, Wengang Zhou, Houqiang Li, and Qi Tian. Can semantic labels assist self-supervised visual representation learning? arXiv preprint arXiv:2011.08621, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "type": "text", + "content": "[36] Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darell. What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659, 2020. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "type": "text", + "content": "[37] Yuanhong Xu, Qi Qian, Hao Li, Rong Jin, and Juhua Hu. Weakly supervised representation learning with coarse labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10593-10601, 2021. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[38] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1476-1485, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 588, + 287, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 642 + ], + "type": "text", + "content": "[39] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16660-16669, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19922" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_content_list.json b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..93eac4c0ed9b2fdaeb478fae43c231626e9ef889 --- /dev/null +++ b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_content_list.json @@ -0,0 +1,1335 @@ +[ + { + "type": "text", + "text": "MaskSketch: Unpaired Structure-guided Masked Image Generation", + "text_level": 1, + "bbox": [ + 140, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dina Bashkirova1*, José Lezama2, Kihyuk Sohn2, Kate Saenko1,3, Irfan Essa2,4", + "bbox": [ + 163, + 179, + 803, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Boston University, 2Google Research, 3MIT-IBM Watson AI Lab, 4Georgia Institute of Technology", + "bbox": [ + 89, + 198, + 879, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent conditional image generation methods produce images of remarkable diversity, fidelity and realism. However, the majority of these methods allow conditioning only on labels or text prompts, which limits their level of control over the generation result. In this paper, we introduce MaskSketch, an image generation method that allows spatial conditioning of the generation result using a guiding sketch as an extra conditioning signal during sampling. MaskSketch utilizes a pre-trained masked generative transformer, requiring no model training or paired supervision, and works with input sketches of different levels of abstraction. We show that intermediate self-attention maps of a masked generative transformer encode important structural information of the input image, such as scene layout and object shape, and we propose a novel sampling method based on this observation to enable structure-guided generation. Our results show that MaskSketch achieves high image realism and fidelity to the guiding structure. Evaluated on standard benchmark datasets, MaskSketch outperforms state-of-the-art methods for sketch-to-image translation, as well as unpaired image-to-image translation approaches. The code can be found on our project website: https://masksketch.github.io/", + "bbox": [ + 75, + 282, + 473, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 657, + 209, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent Image generation methods achieved remarkable success, allowing diverse and photorealistic image synthesis [4, 11, 44, 46]. The majority of state-of-the-art generative models allow conditioning with class labels [2, 4, 11, 13] or text prompts [40, 41, 44, 46]. However, some applications require a more fine-grained control over the spatial composition of the generation result. While methods conditioned with segmentation maps [14] or strokes [34] achieve some spatial control over the generated image, sketching allows a more fine-grained specification of the target spatial layout, which makes it desirable for many creative applications.", + "bbox": [ + 75, + 683, + 468, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose MaskSketch, a method for conditional image synthesis that uses sketch guidance to de", + "bbox": [ + 76, + 849, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91011f0629c36a9af325754519411ba63b01eb471e32025c05cc51181d096fd1.jpg", + "image_caption": [ + "Figure 1. Given an input sketch and class label, MaskSketch samples realistic images that follow the given structure. MaskSketch works on sketches of various degree of abstraction by leveraging a pre-trained masked image generator [4], while not requiring model finetuning or pairwise supervision." + ], + "image_footnote": [], + "bbox": [ + 501, + 248, + 893, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fine the desired structure, and a pre-trained state-of-the-art masked generative transformer, MaskGIT [4], to leverage a strong generative prior. We demonstrate the capability of MaskSketch to generate realistic images of a given structure for sketch-to-photo image translation. Sketch-to-photo [5, 20, 32] is one of the most challenging applications of structure-conditional generation due to the large domain gap between sketches and natural images. MaskSketch achieves a balance between realism and structure fidelity. Our experiments show that MaskSketch outperforms state-of-the-art sketch-to-photo [20] and unpaired image translation methods [6, 25, 37], according to standard metrics for image generation [23] and user preference studies.", + "bbox": [ + 496, + 642, + 893, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In MaskSketch, we formulate a structure similarity constraint based on the observation that the intermediate self-attention maps of a masked generative transformer [4] encode rich structural information (see Fig. 2). We use this", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Google.", + "bbox": [ + 94, + 886, + 328, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1879", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5901986f3e6127051d29fc7885def284cca4e7d07e8fb88cdf39b7cb61810b16.jpg", + "image_caption": [ + "Figure 2. Self-attention maps (PCA) of the intermediate layers of a pre-trained masked generative transformer [4] encode information about the spatial layout of the input. Notably, they are robust to the domain shift between natural images (left) and sketches (right)." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structure similarity constraint to guide the generated image towards the desired spatial layout [22, 48]. Our study shows that the proposed attention-based structure similarity objective is robust to the domain shift occurring in sketch-to-photo translation. The proposed structure-based sampling method leverages a pre-trained image generator, and does not require model finetuning or sketch-photo paired data. Moreover, it is significantly faster than other methods that exploit self-attention maps for guided image generation [48]. Figure 1 shows the translation results produced by our method on sketches of various levels of abstraction.", + "bbox": [ + 75, + 332, + 468, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The limitations of existing sketch-to-photo translation methods [5, 20, 32] come from having to learn both an implicit natural domain prior and the mapping that aligns sketches to natural images, for which the domain gap is severe. MaskSketch, on the other hand, uses the strong generative prior of a pre-trained generative transformer, which allows highly realistic generation. In addition, MaskSketch uses the domain-invariant self-attention maps for structure conditioning, allowing its use on sketches of a wide range of abstraction levels.", + "bbox": [ + 75, + 505, + 468, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 662, + 421, + 676 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that the self-attention maps of a masked generative transformer encode important structural information and are robust to the domain shift between images and sketches.", + "- We propose a sampling method based on self-attention similarity, balancing the structural guidance of an input sketch and the natural image prior.", + "- We demonstrate that the proposed sampling approach, MaskSketch, outperforms state-of-the-art methods in unpaired sketch-to-photo translation.", + "- To the best of our knowledge, MaskSketch is the first method for sketch-to-photo translation in the existing literature that produces photorealistic results requiring only class label supervision." + ], + "bbox": [ + 93, + 685, + 467, + 895 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While there is a vast volume of literature on image generative models thanks to recent progress ranging from generative adversarial networks [2, 18, 27] generative transformers [4, 13, 54] and diffusion models [11, 35, 40, 46], in this section, we focus on reviewing image-conditioned image generation, also known as image translation.", + "bbox": [ + 496, + 114, + 890, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised image conditional generation Sketch-to-photo image translation is a special case of image-conditional image generation. Early conditional image generation methods were based on generative adversarial networks. For example, pix2pix [26] conditioned the generation result by minimizing the patchwise distance between the ground truth and the generated image; SPADE [38] and OASIS [47] used spatially-adaptive instance normalization to condition generation on a segmentation map; CoCosNet [55], CoCosNet V2 [57] warped the reference image using a correlation matrix between the image and the given segmentation map. Similarly to MaskSketch, Makea-Scene and NUWA [14, 52] are designed to condition generation on semantic segmentation and text prompts with a VQ-based transformer. While these methods allow spatial conditioning, they are inapplicable for sketch-to-photo due to the lack of ground truth paired data, domain gap between sketches and segmentation maps and lack of efficient methods that extract semantic segmentation from sketches.", + "bbox": [ + 496, + 212, + 892, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised image-conditional generation In unsupervised image-conditioned translation, the ground truth input and translation pairs are not available for training. For example, CycleGAN [58] used a cycle reconstruction loss to ensure a semantically consistent translation, UNIT [31], MUNIT [25], and StarGANv2 [8] disentangled domain-specific and shared information between the source and target image domains by mapping them to a shared latent embedding space. PSP [42] used StyleGAN [46] inversion along with style mixing for segmentation- and edge-guided translation. SDEdit [33] uses a diffusion model to translate the input strokes or segmentation maps to natural images.", + "bbox": [ + 496, + 507, + 892, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The closest work to ours in this line may be SpliceVIT [48], which uses self-attention key self-similarity extracted from the discriminative ViT (Dino [3]) to represent the structure of an input image. As pointed in [48], SpliceVIT works only in case when both the input and expected output images come from the same domain, which makes it inapplicable for sketch-to-photo translation. VQ-I2I [6] is another work on unsupervised image translation that leverages the generative power of a VQ-GAN [13]-based generative transformer. Unlike MaskSketch, VQ-I2I uses the embedding reconstruction loss for controllable generation.", + "bbox": [ + 496, + 689, + 892, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, [15, 28, 45] also demonstrated how to leverage pre-trained image generators for image synthesis based on novel conditioning inputs. These methods allow to replicate", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1880", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9b19bd7055960bb16ca14f85251f7484f28dbecc1da231c76acd4eddfaaafe8.jpg", + "image_caption": [ + "Figure 3. Left: confidence-based token rejection (masking) in MaskGIT. Right: structure-based token rejection. Confidence-based rejection masks out the least 'likely' tokens, while structure-based rejection masks the tokens with the highest structural distance (Eq. 1) w.r.t. the input sketch." + ], + "image_footnote": [], + "bbox": [ + 81, + 63, + 470, + 232 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9821e08a1aeb7ae80e6b3ea7219402b4e1790d5ab33b43137f8166dd54580ce6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 63, + 890, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a given object or subject in the generated image. However, they do not allow to finely specify the spatial layout of the generated image as MaskSketch and typically require some degree of fine-tuning. Prompt-to-prompt tuning [22] uses the attention features to perform spatially aligned prompt-conditional generation. General image-conditional methods show remarkable results when the source and target domains are visually similar, e.g., translating horses to zebras, performing artistic style transfer, etc, however, they tend to struggle on the more challenging sketch-to-photo translation task. MaskSketch shows a promising alternative for transferring the spatial composition from sketches since it requires no paired data or model training, thanks to leveraging a powerful pre-trained generator.", + "bbox": [ + 75, + 300, + 473, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch-to-photo translation The sketch-to-photo application received attention in recent years thanks to the advancement in the field of image generation. For example, SketchyGAN [5] proposes an a GAN-based approach based on edge-preserving image augmentations, ContextualGAN [32] leverages conditional GAN along with joint image-sketch representation, iSketchNFill [17] uses a gating mechanism to condition output images on the class label and an MUNIT-based generator to synthesize images with diverse appearance. Photosketcher [12] uses sketch-based image retrieval to compose a real image. PITI [51] is pretrained with ground truth edge maps and semantic segmentation maps to learn a domain invariant semantic representation. The state-of-the-art supervised method CoGS [20] learns the structure embeddings using the ground truth image-sketch pairs in the vector-quantized space of a VQGAN [13]. In contrast, MaskSketch does not rely on paired data for training, which allows it to use sketches of different abstraction levels, as well as real photos.", + "bbox": [ + 75, + 523, + 470, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 828, + 168, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the main components in MaskSketch that introduce sketch-guided spatial control to a conditional masked image generator. We first review", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "masked image generation in Section 3.1. Then we introduce the two main components of MaskSketch, a structure similarity distance in Section 3.2, and structure-guided parallel sampling in Section 3.3. Finally, we discuss how to balance the trade-off between structure fidelity and generation realism in Section 3.4.", + "bbox": [ + 496, + 300, + 892, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: Masked Image Generation", + "text_level": 1, + "bbox": [ + 498, + 402, + 846, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked image generation is a state-of-the-art approach for efficient generation [4, 19, 29], combining the strengths of masked token modeling [10] and non-autoregressive sampling [16]. It encodes images as discrete sequences of visual tokens using a VQ-GAN encoder [13], and then trains a bi-directional transformer (BERT [10]) to model natural image distribution in the discrete token sequence space. Generation is performed iteratively, where significant gains in efficiency are obtained by using parallel sampling instead of auto-regressive sampling. MaskGIT [4] starts from a blank canvas where all visual tokens are masked. At each sampling iteration, all the missing tokens are sampled in parallel, and a rejection criteria is used, where the tokens with low model likelihood are masked and will be re-predicted in the next refinement iteration. See Figure 3 (left) for an illustration of a single MaskGIT decoding step. MaskSketch extends the parallel sampling of MaskGIT to sample images that follow the structure determined by an input image (Fig. 3, right), as described in the following sections.", + "bbox": [ + 496, + 426, + 892, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Structure Similarity via Attention Maps", + "text_level": 1, + "bbox": [ + 500, + 739, + 843, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider two images to be structurally similar when their self-similarity maps are close to each other. MaskSketch leverages the self-similarity encoded in the self-attention maps of a masked generative transformer (Section 3.1) to define structural distance. One key observation in our work is that a class-conditional MaskGIT trained on ImageNet shows a high degree of domain invariance in its attention maps and is able to capture the self-similarity in out-of-distribution domains such as sketches (Fig. 2).", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1881", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bf9299ba1b57de4fc69ded82acaed74c811198ef8f2fe2fe5462aa14b618fecc.jpg", + "image_caption": [ + "Figure 4. Structure-guided sampling using attention maps from the first three layers of a masked generative transformer results in a nearly perfect reconstruction of the input, whereas using the final layers (16, 18, 20), out of 24, yields realistic images with a similar structure but highly diverse appearance. (Best viewed in color.)" + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 467, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, we define a structural distance based on a comparison of self-attention maps. Let $\\mathbb{Z}$ be the indices representing the VQ-GAN [13] dictionary of vector-quantized image tokens. Let $\\pmb{x} \\in \\mathbb{Z}^N$ be sequence of $N$ discrete tokens obtained using a VQ-GAN encoding an input image in the vector-quantized space. Given an input image $\\pmb{x}$ and a generated image $\\pmb{y}$ , let $A^\\ell(\\pmb{x}) \\in [0,1]^{N \\times N}$ be the transformer self-attention map at layer $\\ell$ . Each row in $A^\\ell(\\pmb{x})$ represents the attention weights of each token with respect to all tokens, normalized with a softmax function. We define the structural distance between the $i^{\\text{th}}$ tokens of images $\\pmb{x}$ and $\\pmb{y}$ across layers $\\mathcal{L}$ as:", + "bbox": [ + 75, + 319, + 472, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {S} ^ {i} (\\boldsymbol {x}, \\boldsymbol {y}) = \\sum_ {\\ell \\in \\mathcal {L}} d _ {J} \\left(A _ {i} ^ {\\ell} (\\boldsymbol {x}), A _ {i} ^ {\\ell} (\\boldsymbol {y})\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 512, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_J$ is the Jeffrey's divergence:", + "bbox": [ + 76, + 556, + 323, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {J} (\\boldsymbol {u}, \\boldsymbol {v}) = \\frac {K L (\\boldsymbol {u} \\| \\boldsymbol {v}) + K L (\\boldsymbol {v} \\| \\boldsymbol {u})}{2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 580, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, the image regions represented by the $i^{\\mathrm{th}}$ tokens of $\\pmb{x}$ and $\\pmb{y}$ are structurally similar if their distributions of attention self-similarities are close to each other.", + "bbox": [ + 75, + 625, + 468, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Structure-guided Parallel Decoding", + "text_level": 1, + "bbox": [ + 76, + 680, + 387, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MaskSketch adapts the parallel sampling of MaskGIT to take into account the structural similarity between the output and the reference input sketch. More precisely, the token rejection criteria in each decoding iteration is modified to also reject the sampled tokens that have low self-similarity score (1). The proposed structure-guided decoding strategy can also be seen as a greedy optimization technique that balances minimizing the structural distance and following the model's image prior.", + "bbox": [ + 75, + 703, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While MaskGIT sampling rejects token candidates with the lowest likelihood by masking them at the end of each decoding iteration, MaskSketch creates an additional mask that rejects tokens based on the structural similarity to the", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72cccc7200eda24c43d656909af9a14ae96ca5b9ed6f322a518f2f09205d485c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 90, + 901, + 287 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/142ce0aaf7d496ede3ead55be73944a7d69ed632945ee0f453305d2fdf2bc6cd.jpg", + "image_caption": [ + "Figure 5. Top: Traversal of the realism-fidelity trade-off by varying the classifier-free guidance scale. Bottom: Increasing the guidance scale leads to higher realism and a high CLIP similarity to the prompt \"photo of a $c$ \", where $c$ is the class name, at the cost of lower structure fidelity and lower CLIP feature similarity." + ], + "image_footnote": [], + "bbox": [ + 519, + 297, + 898, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input sketch (see Fig. 3). At the end of one decoding iteration, we compute the logical OR between the confidence-based and the structure-based masks to optimize both realism and structure similarity (Sec. 3.4). The pseudocode of our algorithm is described in Algorithm 1. It relies on the function sample_mask, which takes as input a vector of structure similarity scores $s^s$ and the number of masked elements $k$ , and samples a mask by Gumbel top- $k$ using $s^s$ to mask the tokens with the highest structure distance.", + "bbox": [ + 496, + 564, + 892, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The choice of layers $\\mathcal{L}$ selected for computing the structure similarity significantly impacts the generation results. Our experiments show that sampling with the attention maps extracted from the first layers results in nearly identical reconstruction of the given input image, whereas minimizing the structure distance based on the last layers results in images of diverse appearance that are spatially aligned with the input image at a high level, as shown in Fig. 4.", + "bbox": [ + 496, + 700, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Structure fidelity vs realism trade-off", + "text_level": 1, + "bbox": [ + 500, + 830, + 821, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One of the biggest challenges in sketch-to-real translation is the immense domain gap between the source and target domains. The input sketches and natural images differ", + "bbox": [ + 496, + 854, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1882", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "significantly not only in appearance, but also in the distribution of shapes and spatial composition. Due to the domain gap, optimization based solely on the structure distance often results in structurally similar but unrealistic images. To overcome this issue, we propose a combined masking approach that optimizes both structure fidelity and realism.", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To navigate this trade-off, we use a parameter $\\lambda_s \\in [0,1]$ to determine the proportion of tokens masked according to the structure similarity scores and those masked according to the model confidence or likelihood scores. Given an overall masking rate schedule function $\\gamma(t)$ at step $t$ , the structure-based mask rate is computed as $\\lambda_s \\gamma(t)$ , whereas the confidence-based mask rate is $(1 - \\lambda_s) \\gamma(t)$ . Two independent masks, $m_t^s$ and $m_t^c$ , are computed for the structure-based and confidence-based scores, respectively. The final mask at iteration $t$ is then computed as the logical OR between $m_t^s$ and $m_t^c$ . Please refer to Fig. 9 in the Appendix for an ablation study on the parameter $\\lambda_s$ .", + "bbox": [ + 75, + 181, + 467, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classifier-free Guidance To further improve the level of realism in the translation result, we use classifier-free guidance [14, 24, 35] when computing the model likelihood scores. Specifically, for a given sequence of sampled tokens $\\bar{\\pmb{y}}$ and input image $\\pmb{x}$ , we use the pre-trained generator $G$ to compute the per-token logits $\\log p(\\bar{\\pmb{y}}(i)|\\pmb{x}, c)$ conditioned on the correct class $c$ and logits conditioned on a random class $r$ : $\\log p(\\bar{\\pmb{y}}(i)|\\pmb{x}, r)$ , and calibrate the final confidence-based scores as follows:", + "bbox": [ + 75, + 368, + 467, + 502 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {s} ^ {c} (i) = \\log p (\\bar {\\boldsymbol {y}} (i) | \\boldsymbol {x}, c) \\\\ - \\beta (\\log p (\\bar {\\mathbf {y}} (i) | \\boldsymbol {x}, c) - \\log p (\\bar {\\mathbf {y}} (i) | \\boldsymbol {x}, r)) \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 513, + 467, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is the classifier-free guidance scale. Figure 5 shows how varying $\\beta$ affects the fidelity-realism trade-off.", + "bbox": [ + 75, + 559, + 467, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global CLIP-based rejection sampling Minimization of the structure similarity distance in the space of visual", + "bbox": [ + 75, + 595, + 467, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 MaskSketch sampling", + "text_level": 1, + "bbox": [ + 80, + 648, + 315, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Pre-trained BERT generator $G$ , structure and confidence masking schedule function $\\gamma(t)$ , structure-based sampling ratio $\\lambda_s$ , input sketch $x$ , layer(s) $\\ell$ .", + "bbox": [ + 75, + 666, + 467, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Generated image encoding $y_0$", + "bbox": [ + 76, + 712, + 336, + 726 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $A^{\\ell}(\\pmb {x})\\gets \\mathrm{attn\\_map}(G,\\pmb {x},\\ell)$", + "2: Initialize $\\mathbf{y}_T$", + "3: for $t = T - 1 \\dots 0$ do", + "4: $\\bar{\\pmb{y}}_t, \\pmb{s}_t^c = G(\\pmb{y}_{t + 1})$", + "5: $A^{\\ell}(\\bar{\\mathbf{y}}_t) = \\mathrm{attn\\_map}(G,\\bar{\\mathbf{y}}_t,\\ell)$", + "6: $\\pmb{s}_t^s\\gets \\left\\{d_J\\left(A_i^\\ell (\\bar{\\pmb{y}}_t),A_i^\\ell (\\pmb {x})\\right)\\right\\}_{i = 1\\dots N}$", + "7: $\\pmb{m}_{t}^{s} = \\text{sample\\_mask}(\\pmb{s}_{t}^{s}, \\lfloor \\lambda_{s}\\gamma(\\overline{t}) \\cdot N \\rfloor)$", + "8: $\\pmb{m}_{t}^{c} = \\text{sample\\_mask}(\\pmb{s}_{t}^{c},\\lfloor (1 - \\lambda_{s})\\gamma (t)\\cdot N\\rfloor)$", + "9: $\\pmb{m}_t = \\pmb{m}_t^s \\vee \\pmb{m}_t^c$", + "10: $\\pmb{y}_t = \\bar{\\pmb{y}}_t \\odot \\pmb{m}_t$", + "11: end for" + ], + "bbox": [ + 83, + 727, + 447, + 893 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tokens is a discrete optimization problem that cannot be efficiently tackled via continuous optimization methods such as gradient descent. Moreover, since MaskGIT was trained to minimize a different objective, such a greedy optimization process requires more iterations than regular sampling. To increase the stability of the proposed method, we improve the overall fidelity by producing multiple translation samples for a given sketch with different random seeds and guidance scales $\\beta$ , and selecting the image that yields the highest structure fidelity and realism according to a CLIP-based score. Inspired by the recent success in photo-to-sketch mapping with CLIP [39] domain-invariant representations [49], we use the $L_{1}$ distance between features of a CLIP encoder $\\mathrm{CLIP}_s(\\pmb{x},\\pmb{y})$ of the input image $\\pmb{x}$ of class $c$ and generation result $\\pmb{y}$ to estimate the structure similarity. We also use the CLIP similarity score $\\mathrm{CLIP}_r(c,\\pmb{y})$ between the translated image and the corresponding prompt prompt $(c) = \\text{'photo of a } c'$ to assess the realism for each generated example, more details can be found in Appendix F. We normalize the scores across $R$ trials and keep the result with the highest overall quality score:", + "bbox": [ + 496, + 90, + 890, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {\\text {f i n a l}} = \\underset {\\boldsymbol {y} \\in \\left\\{\\boldsymbol {y} _ {1} \\dots \\boldsymbol {y} _ {R} \\right\\}} {\\operatorname {a r g m a x}} (1 - \\operatorname {C L I P} _ {s} (\\boldsymbol {x}, \\boldsymbol {y})) ^ {2} \\operatorname {C L I P} _ {r} (c, \\boldsymbol {y}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 416, + 890, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 5 in the Appendix shows how the proposed CLIP-based selection approach improves the overall generation result as the number of sampling trials increases.", + "bbox": [ + 496, + 452, + 890, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 508, + 630, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setup In all experiments, we used a class-conditional MaskGIT model pretrained on the ImageNet 2012 [9] dataset with the output resolution $256 \\times 256$ . We used layers 1, 3, 16, 20, 21 and 22 to formulate the structure preservation objective. We validated this choice on 100 random sketches considering structure preservation and realism. In our experiments, for each input sketch, we sample the images four times ( $R = 4$ ) with different classifier-free guidance scales (i.e. $\\beta \\in \\{0.0, 0.05, 0.1, 0.25\\}$ ), and select the one that maximizes the CLIP-based objective in Eq. (4). We use a linear decay mask rate schedule in all experiments, starting from $\\gamma(T) = 0.95$ and stopping sampling at the mask rate $\\gamma(0) = 0.25$ , which results in higher realism and reduces artifacts associated with structure-based sampling. To further increase realism, we postprocess the generated samples with Token-Critic refinement [29], which adds 32 sampling iterations. See Appendix D for more details. We generate each image using $T = 500$ sampling iterations, and the overall sampling time for a batch of 8 images is on average 750 seconds on a single TPUv4, including four trials and the CLIP-based evaluation.", + "bbox": [ + 496, + 535, + 890, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines We consider well-established unpaired image-to-image translation methods as baselines. Specifically, we", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1883", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "used CUT [37], MUNIT [25] and VQI2I [6] in our comparisons. We note that for sketch-to-photo translation, methods that use ground truth attribute information to translate from one attribute to another, e.g. StarGAN [7, 8], fail to minimize the gap between sketch and real domains [20]. CUT [37] uses a contrastive objective to ensure structural similarity between the corresponding patches of the input image and the translation result. MUNIT [25] is a GAN-based model that uses latent embedding reconstruction losses to disentangle appearance from structure. VQI2I [6] uses a vector-quantized GAN to encode images into sequences of tokens representing the structure and appearance of the input images, and uses embedding reconstruction losses to enforce the disentanglement of the structure. Since these methods are not class-conditional, we trained them on each class separately. We report the average result across the examples of all classes as well as the results of training on the entire datasets.", + "bbox": [ + 75, + 90, + 472, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although MaskSketch does not utilize paired data, we also consider as baseline the state-of-the-art paired sketch-to-photo method CoGS [20]. We note that VQI2I, CoGS and MUNIT allow diverse sampling with an additional appearance image or vector as input, whereas MaskSketch samples diverse results by varying the random seed.", + "bbox": [ + 75, + 362, + 470, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets For qualitative evaluation of MaskSketch, we propose OpenSketches, a novel dataset made of 200 openly licensed sketches. OpenSketches contains real sketches drawn with pencil and paper, as well as digital sketches. Furthermore, to mimic realistic, highly detailed sketches, we utilized the open source implementation of Stable Diffusion [44] to generate input examples. All the sketches shown in this manuscript are from OpenSketches.", + "bbox": [ + 75, + 470, + 468, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For quantitative evaluation, we considered two datasets: ImageNet-Sketch [50], a dataset of 50 real sketches of 1000 classes of ImageNet-2012 [9] and the Pseudosketches dataset [20], consisting of pairs of ground truth real images and their corresponding automatically extracted edge maps from 125 classes from the ImageNet21K [43] dataset. We present qualitative results for these datasets in Appendix C1. In our quantitative experiments, we report the results on two versions of the datasets: 1) full: using all examples from each of the datasets, and 2) 10-class: using the 10 classes that are reported to result in the highest-quality translation results in CoGS [20]: \"songbird\", \"pizza\", \"volcano\", \"zebra\", \"castle\", \"door\", \"shark\", \"mushroom\", \"cup\", \"lion\". The 10-class subsets of Pseudosketches and ImageNet-Sketch consist of 1,749 and 508 examples respectively, whereas the full datasets consist of 113,370 examples and 52,888 examples, respectively. For the 10-class subsets, we trained unpaired image translation baselines that are not class-conditional on each class separately and", + "bbox": [ + 75, + 592, + 472, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "reported the aggregated results over all 10 classes for a fair comparison with the class-conditional MaskSketch. For the full version of the datasets, we train the baseline methods on all classes without class conditioning. Since ImageNet-Sketch does not provide ground truth paired data, it is impossible to train CoGS [20] on this dataset, therefore we use the model trained on Pseudosketches for both datasets.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Quantitative evaluation of sketch-to-photo translation consists of two aspects: evaluation of realism of the generation results, and evaluation of structure fidelity with respect to the input sketch. To estimate realism, we use the FID score [23]. To assess generation diversity, we used the LPIPS-based diversity score [36], which computes the average LPIPS [56] distance between the generated examples. For a fair comparison and due to the limited number of samples in the 10-class subsets, we report the FID and LPIPS results over 10,000 examples generated with different 'appearance' inputs with the baseline methods CoGS, MUNIT and VQI2I, and with different seeds for MaskS-ketch. For CUT, we diversify the generated set with augmentations. The FID score is computed with respect to the images from ImageNet [9] for the ImageNet-Sketch experiments, and with respect to the ground truth Pseudosketches images for the Pseudosketches experiments.", + "bbox": [ + 496, + 222, + 893, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To provide additional quantitative evaluation of structure preservation quality and realism, we also report the two CLIP-based metrics defined in Sec. 3.4: image feature distance and prompt similarity score. The CLIP feature distance metric is more appropriate for the evaluation of structure preservation quality than the edge-based metrics [20] since the CLIP features are more invariant to the domain gap as shown in the recent works on image-to-sketch translation [49]. We note that these metrics are identical to the CLIP-based rejection sampling in Sec. 3, and we include the quantitative results without CLIP-based sampling in Appendix C.1.", + "bbox": [ + 496, + 482, + 893, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Preference Studies Quantitative evaluation of structure fidelity is challenging due to the distribution shift between the shapes of real objects and abstract sketches and outlines. To complement the quantitative results, we performed user preference studies. We asked users the question: \"Given the task of converting the sketch shown on the left into a realistic photo, which result do you prefer?\" Users were asked to pick one result among the five compared methods (CoGS, MUNIT, VQI2I, CUT and MaskSketch) according to their preference. We collected three preference evaluations for each example in the 10-class ImageNet-Sketch and Pseudosketches datasets. Finally, we counted only unanimous votes to guarantee statistical significance. Please see Appendix E for more details.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 Not shown in the main manuscript due to copyright concerns.", + "bbox": [ + 76, + 886, + 415, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1884", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/60c68998a43d5b1e58cf3e725656fcd73ca8e54fe9bb88c7e7c9b7695a45b280.jpg", + "image_caption": [ + "Figure 6. Example translations by MaskSketch on the OpenSketches dataset. The model takes as input a sketch and a class label." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 872, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 76, + 527, + 163, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results In Tab. 1, we report the quantitative evaluation results on the 10-class subsets of ImageNet-Sketch and Pseudosketch. Additionally, we report the FID and LPIPS diversity scores over the full ImageNet-Sketch and Pseudosketch datasets in Tab. 2. The results on the 10-class subsets indicate an advantage of MaskSketch in terms of realism and diversity, with a two-fold decrease in the FID score compared to the baseline MUNIT on both ImageNet-Sketch and Pseudosketch datasets. In our experiments, MaskSketch outperformed the baselines on the entire ImageNet-Sketch dataset of real sketches, including the fully-supervised CoGS on the Pseudosketches. Notably, general image translation methods, such as MUNIT and CUT, outperform the fully-supervised CoGS in a class-supervised setup. As seen from the FID and LPIPS results, VQI2I struggles to generalize on the relatively small ImageNet-Sketch dataset that contains only 50 examples in each class, mainly due to mode collapse.", + "bbox": [ + 75, + 554, + 472, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results Qualitative comparison shows that the baseline image translation methods, including the supervised CoGS, capture the overall layout and outlines of", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the input sketch but sometimes fail to produce realistic results. For instance, the GAN-based architectures, namely CUT and MUNIT, produce structurally similar results by practically recoloring the input sketch, which results in a sub-par realism, especially on the more abstract sketches. In our experiments, the VQ-GAN-based VQI2I model failed to learn the correspondences between hand-drawn sketches from ImageNet-Sketch and images from the real photo domain due to a limited number of examples in ImageNet-Sketch, therefore we observe a severe mode collapse on most classes. The fully-supervised CoGS sometimes failed to produce realistic and semantically meaningful results, especially on the hand-drawn sketches. MaskSketch achieved a good balance between realism and structure fidelity on the majority of sketches from ImageNet-Sketch. However, MaskSketch struggled to preserve structure on some examples from Pseudosketches due to the extreme complexity of the extracted edge maps.", + "bbox": [ + 496, + 527, + 893, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further refine the resulting token sequence to produce a more realistic image, we use token-critic refinement as proposed by Lezama et al. [29]. Token-critic approach uses a discriminator to detect and refine the tokens that decrease the probability of the overall sequence.", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1885", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c26c89e33d03877753ea2a2045c16208005d2593d337afe1afc054bcf90ef5be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
supervisionFID ↓LPIPS ↑CLIP prompt ↑CLIP feat. ↓User preference ↑
ImageNet-Sketch
MUNITclass68.650.5852.1030.2710.70%
CUTclass77.740.6865.5928.0519.78%
VQ-I2Iclass181.770.3253.7631.050%
CoGSclass + pairs97.310.6456.6229.528.55%
MaskSketch(ours)class33.240.7867.1026.6359.35%
Pseudosketches
MUNITclass93.230.6941.9127.6523.08%
CUTclass112.110.4245.6727.6225.0%
VQ-I2Iclass169.10.7734.528.470.64%
CoGSclass + pairs102.660.6834.7927.5214.10%
MaskSketch (ours)class56.550.7859.4825.6035.25%
", + "bbox": [ + 114, + 88, + 854, + 320 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a9836d9af61dc45861321ad2dd2760280e045f448369fc7f2f9e1fc321f74e45.jpg", + "table_caption": [ + "Table 1. Sketch-to-photo translation performance on ImageNet-Sketch (top) and Pseudosketches (bottom) 10-classes subsets." + ], + "table_footnote": [], + "table_body": "
ImageNet-SketchPseudosketches
FID ↓LPIPS ↑FID ↓LPIPS ↑
MUNIT113.450.74121.690.71
CUT161.330.74163.820.70
VQI2I131.700.72135.470.71
CoGS (sup.)85.090.7249.310.71
MaskSketch23.890.7746.440.78
", + "bbox": [ + 81, + 369, + 465, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Comparison on the full ImageNet-Sketch (left) and Pseudosketches-validation (right) datasets.", + "bbox": [ + 76, + 507, + 470, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations The main limitation of MaskSketch is computational efficiency. To achieve a successful optimization of the structural constraint, MaskSketch requires significantly more sampling iterations than the regular MaskGIT. Furthermore, to improve the stability of results it was necessary to apply a multiple trials rejection scheme. Two other important limitations for MaskSketch are the coarse granularity of the attention maps in a transformer, and the flexibility of the prior model, in our case an ImageNet-pretrained MaskGIT. Figure 7 illustrates the common failure cases of our method: out-of-distribution scene composition scarcely or not represented in the training set of MaskGIT, multiple objects forming an unrealistic scene, as well as the complex scenes with multiple foreground and background objects.", + "bbox": [ + 75, + 566, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 797, + 194, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed MaskSketch, a sketch-guided image generation method that allows control over the spatial layout of the generation result. MaskSketch achieves high realism and structure preservation without pairwise supervision, does not require model finetuning and works on sketches of", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07a3e9d0399af67b3e02c8598853dd84df257780625aae71cb3c8827c87c454f.jpg", + "image_caption": [ + "Figure 7. Failure cases of MaskSketch on the hand-drawn sketch examples: out-of-distribution composition and shapes, complex scenes containing multiple objects." + ], + "image_footnote": [], + "bbox": [ + 506, + 369, + 888, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "various levels of abstraction. We show that the self-attention maps of the intermediate layers of a masked generative transformer encode important structural information of the input image and are sufficiently domain-invariant, which allows their use in a structure similarity constraint. Our experimental results show that the proposed attention-based sampling approach outperforms state-of-the-art sketch-tophoto and general image translation methods in terms of both realism and structure fidelity.", + "bbox": [ + 496, + 680, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We thank Tali Dekel, Huiwen Chang, Lu Jiang, and David Salesin for their insightful advice and guidance. This work was done during an internship at Google Research.", + "bbox": [ + 496, + 834, + 892, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1886", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 12", + "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1, 2", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2", + "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1, 2, 3", + "[5] Wengling Chen and James Hays. Sketchygan: Towards diverse and realistic sketch to image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9416-9425, 2018. 1, 2, 3", + "[6] Yu-Jie Chen, Shin-I Cheng, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Vector quantized image-to-image translation. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6", + "[7] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 6", + "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 6", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 6", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3", + "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2", + "[12] Mathias Eitz, Ronald Richter, Kristian Hildebrand, Tamy Boubekeur, and Marc Alexa. Photosketcher: interactive sketch-based image synthesis. IEEE Computer Graphics and Applications, 31(6):56-66, 2011. 3", + "[13] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020. 1, 2, 3, 4", + "[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1, 2, 5" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[16] Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324, 2019. 3", + "[17] Arnab Ghosh, Richard Zhang, Puneet K Dokania, Oliver Wang, Alexei A Efros, Philip HS Torr, and Eli Shechtman. Interactive sketch & fill: Multiclass sketch-to-image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1171-1180, 2019. 3", + "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[19] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. arXiv preprint arXiv:2111.14822, 2021. 3", + "[20] Cusuh Ham, Gemma Canet Tarres, Tu Bui, James Hays, Zhe Lin, and John Collomosse. Cogs: Controllable generation and search from sketch and style. European Conference on Computer Vision, 2022. 1, 2, 3, 6", + "[21] Jonathan Heek, Anselm Levskaya, Avital Oliver, Marvin Ritter, Bertrand Rondepierre, Andreas Steiner, and Marc van Zee. Flax: A neural network library and ecosystem for JAX, 2020. 12", + "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3", + "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 1, 6", + "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5", + "[25] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In ECCV, 2018. 1, 2, 6", + "[26] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[28] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1887", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] José Lezama, Huiwen Chang, Lu Jiang, , and Irfan Essa. Improved masked image generation with token-critic. European Conference on Computer Vision, 2022. 3, 5, 7, 13", + "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 14", + "[31] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in neural information processing systems, 30, 2017. 2", + "[32] Yongyi Lu, Shangzhe Wu, Yu-Wing Tai, and Chi-Keung Tang. Image generation from sketch constraint using contextual gan. In Proceedings of the European conference on computer vision (ECCV), pages 205–220, 2018. 1, 2, 3", + "[33] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 2", + "[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 1", + "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 5", + "[36] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 6", + "[37] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In European conference on computer vision, pages 319-345. Springer, 2020. 1, 6", + "[38] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5", + "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2", + "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2", + "[43] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 6", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 6", + "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2", + "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2", + "[47] Vadim Sushko, Edgar Schonfeld, Dan Zhang, Juergen Gall, Bernt Schiele, and Anna Khoreva. You only need adversarial supervision for semantic image synthesis. arXiv preprint arXiv:2012.04781, 2020. 2", + "[48] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. arXiv preprint arXiv:2201.00424, 2022. 2", + "[49] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching, 2022. 5, 6", + "[50] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 6", + "[51] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. In arXiv, 2022. 3, 14", + "[52] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. N\\`uwa: Visual synthesis pre-training for neural visual world creation. arXiv preprint arXiv:2111.12417, 2021. 2", + "[53] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 14", + "[54] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2", + "[55] Pan Zhang, Bo Zhang, Dong Chen, Lu Yuan, and Fang Wen. Cross-domain correspondence learning for exemplar-based image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5143-5153, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1888", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[57] Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocosnet v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11465–11475, 2021. 2", + "[58] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + ], + "bbox": [ + 78, + 90, + 467, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1889", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_model.json b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0cabf1a53b05e18560e3ffe401882f983dc21247 --- /dev/null +++ b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_model.json @@ -0,0 +1,2103 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "MaskSketch: Unpaired Structure-guided Masked Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.18, + 0.804, + 0.199 + ], + "angle": 0, + "content": "Dina Bashkirova1*, José Lezama2, Kihyuk Sohn2, Kate Saenko1,3, Irfan Essa2,4" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.88, + 0.218 + ], + "angle": 0, + "content": "1Boston University, 2Google Research, 3MIT-IBM Watson AI Lab, 4Georgia Institute of Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.474, + 0.631 + ], + "angle": 0, + "content": "Recent conditional image generation methods produce images of remarkable diversity, fidelity and realism. However, the majority of these methods allow conditioning only on labels or text prompts, which limits their level of control over the generation result. In this paper, we introduce MaskSketch, an image generation method that allows spatial conditioning of the generation result using a guiding sketch as an extra conditioning signal during sampling. MaskSketch utilizes a pre-trained masked generative transformer, requiring no model training or paired supervision, and works with input sketches of different levels of abstraction. We show that intermediate self-attention maps of a masked generative transformer encode important structural information of the input image, such as scene layout and object shape, and we propose a novel sampling method based on this observation to enable structure-guided generation. Our results show that MaskSketch achieves high image realism and fidelity to the guiding structure. Evaluated on standard benchmark datasets, MaskSketch outperforms state-of-the-art methods for sketch-to-image translation, as well as unpaired image-to-image translation approaches. The code can be found on our project website: https://masksketch.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.658, + 0.21, + 0.673 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.849 + ], + "angle": 0, + "content": "Recent Image generation methods achieved remarkable success, allowing diverse and photorealistic image synthesis [4, 11, 44, 46]. The majority of state-of-the-art generative models allow conditioning with class labels [2, 4, 11, 13] or text prompts [40, 41, 44, 46]. However, some applications require a more fine-grained control over the spatial composition of the generation result. While methods conditioned with segmentation maps [14] or strokes [34] achieve some spatial control over the generated image, sketching allows a more fine-grained specification of the target spatial layout, which makes it desirable for many creative applications." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.85, + 0.47, + 0.88 + ], + "angle": 0, + "content": "In this paper, we propose MaskSketch, a method for conditional image synthesis that uses sketch guidance to de" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.249, + 0.895, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.55, + 0.894, + 0.62 + ], + "angle": 0, + "content": "Figure 1. Given an input sketch and class label, MaskSketch samples realistic images that follow the given structure. MaskSketch works on sketches of various degree of abstraction by leveraging a pre-trained masked image generator [4], while not requiring model finetuning or pairwise supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.839 + ], + "angle": 0, + "content": "fine the desired structure, and a pre-trained state-of-the-art masked generative transformer, MaskGIT [4], to leverage a strong generative prior. We demonstrate the capability of MaskSketch to generate realistic images of a given structure for sketch-to-photo image translation. Sketch-to-photo [5, 20, 32] is one of the most challenging applications of structure-conditional generation due to the large domain gap between sketches and natural images. MaskSketch achieves a balance between realism and structure fidelity. Our experiments show that MaskSketch outperforms state-of-the-art sketch-to-photo [20] and unpaired image translation methods [6, 25, 37], according to standard metrics for image generation [23] and user preference studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In MaskSketch, we formulate a structure similarity constraint based on the observation that the intermediate self-attention maps of a masked generative transformer [4] encode rich structural information (see Fig. 2). We use this" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.329, + 0.901 + ], + "angle": 0, + "content": "*Work done during an internship at Google." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1879" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.472, + 0.306 + ], + "angle": 0, + "content": "Figure 2. Self-attention maps (PCA) of the intermediate layers of a pre-trained masked generative transformer [4] encode information about the spatial layout of the input. Notably, they are robust to the domain shift between natural images (left) and sketches (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.47, + 0.499 + ], + "angle": 0, + "content": "structure similarity constraint to guide the generated image towards the desired spatial layout [22, 48]. Our study shows that the proposed attention-based structure similarity objective is robust to the domain shift occurring in sketch-to-photo translation. The proposed structure-based sampling method leverages a pre-trained image generator, and does not require model finetuning or sketch-photo paired data. Moreover, it is significantly faster than other methods that exploit self-attention maps for guided image generation [48]. Figure 1 shows the translation results produced by our method on sketches of various levels of abstraction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.656 + ], + "angle": 0, + "content": "The limitations of existing sketch-to-photo translation methods [5, 20, 32] come from having to learn both an implicit natural domain prior and the mapping that aligns sketches to natural images, for which the domain gap is severe. MaskSketch, on the other hand, uses the strong generative prior of a pre-trained generative transformer, which allows highly realistic generation. In addition, MaskSketch uses the domain-invariant self-attention maps for structure conditioning, allowing its use on sketches of a wide range of abstraction levels." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.664, + 0.422, + 0.678 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.686, + 0.468, + 0.745 + ], + "angle": 0, + "content": "- We show that the self-attention maps of a masked generative transformer encode important structural information and are robust to the domain shift between images and sketches." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.747, + 0.468, + 0.79 + ], + "angle": 0, + "content": "- We propose a sampling method based on self-attention similarity, balancing the structural guidance of an input sketch and the natural image prior." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.792, + 0.468, + 0.835 + ], + "angle": 0, + "content": "- We demonstrate that the proposed sampling approach, MaskSketch, outperforms state-of-the-art methods in unpaired sketch-to-photo translation." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.838, + 0.468, + 0.896 + ], + "angle": 0, + "content": "- To the best of our knowledge, MaskSketch is the first method for sketch-to-photo translation in the existing literature that produces photorealistic results requiring only class label supervision." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.686, + 0.468, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.207 + ], + "angle": 0, + "content": "While there is a vast volume of literature on image generative models thanks to recent progress ranging from generative adversarial networks [2, 18, 27] generative transformers [4, 13, 54] and diffusion models [11, 35, 40, 46], in this section, we focus on reviewing image-conditioned image generation, also known as image translation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.893, + 0.501 + ], + "angle": 0, + "content": "Supervised image conditional generation Sketch-to-photo image translation is a special case of image-conditional image generation. Early conditional image generation methods were based on generative adversarial networks. For example, pix2pix [26] conditioned the generation result by minimizing the patchwise distance between the ground truth and the generated image; SPADE [38] and OASIS [47] used spatially-adaptive instance normalization to condition generation on a segmentation map; CoCosNet [55], CoCosNet V2 [57] warped the reference image using a correlation matrix between the image and the given segmentation map. Similarly to MaskSketch, Makea-Scene and NUWA [14, 52] are designed to condition generation on semantic segmentation and text prompts with a VQ-based transformer. While these methods allow spatial conditioning, they are inapplicable for sketch-to-photo due to the lack of ground truth paired data, domain gap between sketches and segmentation maps and lack of efficient methods that extract semantic segmentation from sketches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.689 + ], + "angle": 0, + "content": "Unsupervised image-conditional generation In unsupervised image-conditioned translation, the ground truth input and translation pairs are not available for training. For example, CycleGAN [58] used a cycle reconstruction loss to ensure a semantically consistent translation, UNIT [31], MUNIT [25], and StarGANv2 [8] disentangled domain-specific and shared information between the source and target image domains by mapping them to a shared latent embedding space. PSP [42] used StyleGAN [46] inversion along with style mixing for segmentation- and edge-guided translation. SDEdit [33] uses a diffusion model to translate the input strokes or segmentation maps to natural images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.855 + ], + "angle": 0, + "content": "The closest work to ours in this line may be SpliceVIT [48], which uses self-attention key self-similarity extracted from the discriminative ViT (Dino [3]) to represent the structure of an input image. As pointed in [48], SpliceVIT works only in case when both the input and expected output images come from the same domain, which makes it inapplicable for sketch-to-photo translation. VQ-I2I [6] is another work on unsupervised image translation that leverages the generative power of a VQ-GAN [13]-based generative transformer. Unlike MaskSketch, VQ-I2I uses the embedding reconstruction loss for controllable generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, [15, 28, 45] also demonstrated how to leverage pre-trained image generators for image synthesis based on novel conditioning inputs. These methods allow to replicate" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1880" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.064, + 0.472, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.064, + 0.891, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.894, + 0.288 + ], + "angle": 0, + "content": "Figure 3. Left: confidence-based token rejection (masking) in MaskGIT. Right: structure-based token rejection. Confidence-based rejection masks out the least 'likely' tokens, while structure-based rejection masks the tokens with the highest structural distance (Eq. 1) w.r.t. the input sketch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.513 + ], + "angle": 0, + "content": "a given object or subject in the generated image. However, they do not allow to finely specify the spatial layout of the generated image as MaskSketch and typically require some degree of fine-tuning. Prompt-to-prompt tuning [22] uses the attention features to perform spatially aligned prompt-conditional generation. General image-conditional methods show remarkable results when the source and target domains are visually similar, e.g., translating horses to zebras, performing artistic style transfer, etc, however, they tend to struggle on the more challenging sketch-to-photo translation task. MaskSketch shows a promising alternative for transferring the spatial composition from sketches since it requires no paired data or model training, thanks to leveraging a powerful pre-trained generator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.472, + 0.812 + ], + "angle": 0, + "content": "Sketch-to-photo translation The sketch-to-photo application received attention in recent years thanks to the advancement in the field of image generation. For example, SketchyGAN [5] proposes an a GAN-based approach based on edge-preserving image augmentations, ContextualGAN [32] leverages conditional GAN along with joint image-sketch representation, iSketchNFill [17] uses a gating mechanism to condition output images on the class label and an MUNIT-based generator to synthesize images with diverse appearance. Photosketcher [12] uses sketch-based image retrieval to compose a real image. PITI [51] is pretrained with ground truth edge maps and semantic segmentation maps to learn a domain invariant semantic representation. The state-of-the-art supervised method CoGS [20] learns the structure embeddings using the ground truth image-sketch pairs in the vector-quantized space of a VQGAN [13]. In contrast, MaskSketch does not rely on paired data for training, which allows it to use sketches of different abstraction levels, as well as real photos." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.829, + 0.169, + 0.844 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this section, we describe the main components in MaskSketch that introduce sketch-guided spatial control to a conditional masked image generator. We first review" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.893, + 0.39 + ], + "angle": 0, + "content": "masked image generation in Section 3.1. Then we introduce the two main components of MaskSketch, a structure similarity distance in Section 3.2, and structure-guided parallel sampling in Section 3.3. Finally, we discuss how to balance the trade-off between structure fidelity and generation realism in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.403, + 0.848, + 0.42 + ], + "angle": 0, + "content": "3.1. Background: Masked Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.893, + 0.73 + ], + "angle": 0, + "content": "Masked image generation is a state-of-the-art approach for efficient generation [4, 19, 29], combining the strengths of masked token modeling [10] and non-autoregressive sampling [16]. It encodes images as discrete sequences of visual tokens using a VQ-GAN encoder [13], and then trains a bi-directional transformer (BERT [10]) to model natural image distribution in the discrete token sequence space. Generation is performed iteratively, where significant gains in efficiency are obtained by using parallel sampling instead of auto-regressive sampling. MaskGIT [4] starts from a blank canvas where all visual tokens are masked. At each sampling iteration, all the missing tokens are sampled in parallel, and a rejection criteria is used, where the tokens with low model likelihood are masked and will be re-predicted in the next refinement iteration. See Figure 3 (left) for an illustration of a single MaskGIT decoding step. MaskSketch extends the parallel sampling of MaskGIT to sample images that follow the structure determined by an input image (Fig. 3, right), as described in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.844, + 0.758 + ], + "angle": 0, + "content": "3.2. Structure Similarity via Attention Maps" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We consider two images to be structurally similar when their self-similarity maps are close to each other. MaskSketch leverages the self-similarity encoded in the self-attention maps of a masked generative transformer (Section 3.1) to define structural distance. One key observation in our work is that a class-conditional MaskGIT trained on ImageNet shows a high degree of domain invariance in its attention maps and is able to capture the self-similarity in out-of-distribution domains such as sketches (Fig. 2)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1881" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.468, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.229, + 0.471, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Structure-guided sampling using attention maps from the first three layers of a masked generative transformer results in a nearly perfect reconstruction of the input, whereas using the final layers (16, 18, 20), out of 24, yields realistic images with a similar structure but highly diverse appearance. (Best viewed in color.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.473, + 0.502 + ], + "angle": 0, + "content": "Formally, we define a structural distance based on a comparison of self-attention maps. Let \\(\\mathbb{Z}\\) be the indices representing the VQ-GAN [13] dictionary of vector-quantized image tokens. Let \\(\\pmb{x} \\in \\mathbb{Z}^N\\) be sequence of \\(N\\) discrete tokens obtained using a VQ-GAN encoding an input image in the vector-quantized space. Given an input image \\(\\pmb{x}\\) and a generated image \\(\\pmb{y}\\), let \\(A^\\ell(\\pmb{x}) \\in [0,1]^{N \\times N}\\) be the transformer self-attention map at layer \\(\\ell\\). Each row in \\(A^\\ell(\\pmb{x})\\) represents the attention weights of each token with respect to all tokens, normalized with a softmax function. We define the structural distance between the \\(i^{\\text{th}}\\) tokens of images \\(\\pmb{x}\\) and \\(\\pmb{y}\\) across layers \\(\\mathcal{L}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.513, + 0.47, + 0.546 + ], + "angle": 0, + "content": "\\[\nd _ {S} ^ {i} (\\boldsymbol {x}, \\boldsymbol {y}) = \\sum_ {\\ell \\in \\mathcal {L}} d _ {J} \\left(A _ {i} ^ {\\ell} (\\boldsymbol {x}), A _ {i} ^ {\\ell} (\\boldsymbol {y})\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.557, + 0.324, + 0.572 + ], + "angle": 0, + "content": "where \\(d_J\\) is the Jeffrey's divergence:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.582, + 0.47, + 0.613 + ], + "angle": 0, + "content": "\\[\nd _ {J} (\\boldsymbol {u}, \\boldsymbol {v}) = \\frac {K L (\\boldsymbol {u} \\| \\boldsymbol {v}) + K L (\\boldsymbol {v} \\| \\boldsymbol {u})}{2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Intuitively, the image regions represented by the \\(i^{\\mathrm{th}}\\) tokens of \\(\\pmb{x}\\) and \\(\\pmb{y}\\) are structurally similar if their distributions of attention self-similarities are close to each other." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.388, + 0.698 + ], + "angle": 0, + "content": "3.3. Structure-guided Parallel Decoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.84 + ], + "angle": 0, + "content": "MaskSketch adapts the parallel sampling of MaskGIT to take into account the structural similarity between the output and the reference input sketch. More precisely, the token rejection criteria in each decoding iteration is modified to also reject the sampled tokens that have low self-similarity score (1). The proposed structure-guided decoding strategy can also be seen as a greedy optimization technique that balances minimizing the structural distance and following the model's image prior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "While MaskGIT sampling rejects token candidates with the lowest likelihood by masking them at the end of each decoding iteration, MaskSketch creates an additional mask that rejects tokens based on the structural similarity to the" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.902, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.299, + 0.899, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.487, + 0.893, + 0.557 + ], + "angle": 0, + "content": "Figure 5. Top: Traversal of the realism-fidelity trade-off by varying the classifier-free guidance scale. Bottom: Increasing the guidance scale leads to higher realism and a high CLIP similarity to the prompt \"photo of a \\( c \\)\", where \\( c \\) is the class name, at the cost of lower structure fidelity and lower CLIP feature similarity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.893, + 0.701 + ], + "angle": 0, + "content": "input sketch (see Fig. 3). At the end of one decoding iteration, we compute the logical OR between the confidence-based and the structure-based masks to optimize both realism and structure similarity (Sec. 3.4). The pseudocode of our algorithm is described in Algorithm 1. It relies on the function sample_mask, which takes as input a vector of structure similarity scores \\( s^s \\) and the number of masked elements \\( k \\), and samples a mask by Gumbel top- \\( k \\) using \\( s^s \\) to mask the tokens with the highest structure distance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.822 + ], + "angle": 0, + "content": "The choice of layers \\(\\mathcal{L}\\) selected for computing the structure similarity significantly impacts the generation results. Our experiments show that sampling with the attention maps extracted from the first layers results in nearly identical reconstruction of the given input image, whereas minimizing the structure distance based on the last layers results in images of diverse appearance that are spatially aligned with the input image at a high level, as shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.822, + 0.848 + ], + "angle": 0, + "content": "3.4. Structure fidelity vs realism trade-off" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "One of the biggest challenges in sketch-to-real translation is the immense domain gap between the source and target domains. The input sketches and natural images differ" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1882" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "significantly not only in appearance, but also in the distribution of shapes and spatial composition. Due to the domain gap, optimization based solely on the structure distance often results in structurally similar but unrealistic images. To overcome this issue, we propose a combined masking approach that optimizes both structure fidelity and realism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.468, + 0.363 + ], + "angle": 0, + "content": "To navigate this trade-off, we use a parameter \\(\\lambda_s \\in [0,1]\\) to determine the proportion of tokens masked according to the structure similarity scores and those masked according to the model confidence or likelihood scores. Given an overall masking rate schedule function \\(\\gamma(t)\\) at step \\(t\\), the structure-based mask rate is computed as \\(\\lambda_s \\gamma(t)\\), whereas the confidence-based mask rate is \\((1 - \\lambda_s) \\gamma(t)\\). Two independent masks, \\(m_t^s\\) and \\(m_t^c\\), are computed for the structure-based and confidence-based scores, respectively. The final mask at iteration \\(t\\) is then computed as the logical OR between \\(m_t^s\\) and \\(m_t^c\\). Please refer to Fig. 9 in the Appendix for an ablation study on the parameter \\(\\lambda_s\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.468, + 0.503 + ], + "angle": 0, + "content": "Classifier-free Guidance To further improve the level of realism in the translation result, we use classifier-free guidance [14, 24, 35] when computing the model likelihood scores. Specifically, for a given sequence of sampled tokens \\(\\bar{\\pmb{y}}\\) and input image \\(\\pmb{x}\\), we use the pre-trained generator \\(G\\) to compute the per-token logits \\(\\log p(\\bar{\\pmb{y}}(i)|\\pmb{x}, c)\\) conditioned on the correct class \\(c\\) and logits conditioned on a random class \\(r\\): \\(\\log p(\\bar{\\pmb{y}}(i)|\\pmb{x}, r)\\), and calibrate the final confidence-based scores as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.515, + 0.468, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {s} ^ {c} (i) = \\log p (\\bar {\\boldsymbol {y}} (i) | \\boldsymbol {x}, c) \\\\ - \\beta (\\log p (\\bar {\\mathbf {y}} (i) | \\boldsymbol {x}, c) - \\log p (\\bar {\\mathbf {y}} (i) | \\boldsymbol {x}, r)) \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.468, + 0.59 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the classifier-free guidance scale. Figure 5 shows how varying \\(\\beta\\) affects the fidelity-realism trade-off." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.468, + 0.626 + ], + "angle": 0, + "content": "Global CLIP-based rejection sampling Minimization of the structure similarity distance in the space of visual" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.65, + 0.316, + 0.665 + ], + "angle": 0, + "content": "Algorithm 1 MaskSketch sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.468, + 0.711 + ], + "angle": 0, + "content": "Input: Pre-trained BERT generator \\(G\\), structure and confidence masking schedule function \\(\\gamma(t)\\), structure-based sampling ratio \\(\\lambda_s\\), input sketch \\(x\\), layer(s) \\(\\ell\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.713, + 0.338, + 0.727 + ], + "angle": 0, + "content": "Output: Generated image encoding \\( y_0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.728, + 0.309, + 0.745 + ], + "angle": 0, + "content": "1: \\(A^{\\ell}(\\pmb {x})\\gets \\mathrm{attn\\_map}(G,\\pmb {x},\\ell)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.746, + 0.192, + 0.76 + ], + "angle": 0, + "content": "2: Initialize \\( \\mathbf{y}_T \\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.761, + 0.258, + 0.774 + ], + "angle": 0, + "content": "3: for \\( t = T - 1 \\dots 0 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.776, + 0.259, + 0.79 + ], + "angle": 0, + "content": "4: \\(\\bar{\\pmb{y}}_t, \\pmb{s}_t^c = G(\\pmb{y}_{t + 1})\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.79, + 0.343, + 0.805 + ], + "angle": 0, + "content": "5: \\(A^{\\ell}(\\bar{\\mathbf{y}}_t) = \\mathrm{attn\\_map}(G,\\bar{\\mathbf{y}}_t,\\ell)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.806, + 0.377, + 0.821 + ], + "angle": 0, + "content": "6: \\(\\pmb{s}_t^s\\gets \\left\\{d_J\\left(A_i^\\ell (\\bar{\\pmb{y}}_t),A_i^\\ell (\\pmb {x})\\right)\\right\\}_{i = 1\\dots N}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.821, + 0.408, + 0.836 + ], + "angle": 0, + "content": "7: \\(\\pmb{m}_{t}^{s} = \\text{sample\\_mask}(\\pmb{s}_{t}^{s}, \\lfloor \\lambda_{s}\\gamma(\\overline{t}) \\cdot N \\rfloor)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.836, + 0.448, + 0.851 + ], + "angle": 0, + "content": "8: \\(\\pmb{m}_{t}^{c} = \\text{sample\\_mask}(\\pmb{s}_{t}^{c},\\lfloor (1 - \\lambda_{s})\\gamma (t)\\cdot N\\rfloor)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.852, + 0.256, + 0.866 + ], + "angle": 0, + "content": "9: \\(\\pmb{m}_t = \\pmb{m}_t^s \\vee \\pmb{m}_t^c\\)" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.867, + 0.233, + 0.88 + ], + "angle": 0, + "content": "10: \\( \\pmb{y}_t = \\bar{\\pmb{y}}_t \\odot \\pmb{m}_t \\)" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.882, + 0.16, + 0.894 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.728, + 0.448, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.409 + ], + "angle": 0, + "content": "tokens is a discrete optimization problem that cannot be efficiently tackled via continuous optimization methods such as gradient descent. Moreover, since MaskGIT was trained to minimize a different objective, such a greedy optimization process requires more iterations than regular sampling. To increase the stability of the proposed method, we improve the overall fidelity by producing multiple translation samples for a given sketch with different random seeds and guidance scales \\(\\beta\\), and selecting the image that yields the highest structure fidelity and realism according to a CLIP-based score. Inspired by the recent success in photo-to-sketch mapping with CLIP [39] domain-invariant representations [49], we use the \\(L_{1}\\) distance between features of a CLIP encoder \\(\\mathrm{CLIP}_s(\\pmb{x},\\pmb{y})\\) of the input image \\(\\pmb{x}\\) of class \\(c\\) and generation result \\(\\pmb{y}\\) to estimate the structure similarity. We also use the CLIP similarity score \\(\\mathrm{CLIP}_r(c,\\pmb{y})\\) between the translated image and the corresponding prompt prompt \\((c) = \\text{'photo of a } c'\\) to assess the realism for each generated example, more details can be found in Appendix F. We normalize the scores across \\(R\\) trials and keep the result with the highest overall quality score:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.417, + 0.892, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {\\text {f i n a l}} = \\underset {\\boldsymbol {y} \\in \\left\\{\\boldsymbol {y} _ {1} \\dots \\boldsymbol {y} _ {R} \\right\\}} {\\operatorname {a r g m a x}} (1 - \\operatorname {C L I P} _ {s} (\\boldsymbol {x}, \\boldsymbol {y})) ^ {2} \\operatorname {C L I P} _ {r} (c, \\boldsymbol {y}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Tab. 5 in the Appendix shows how the proposed CLIP-based selection approach improves the overall generation result as the number of sampling trials increases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.632, + 0.527 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Experimental Setup In all experiments, we used a class-conditional MaskGIT model pretrained on the ImageNet 2012 [9] dataset with the output resolution \\(256 \\times 256\\). We used layers 1, 3, 16, 20, 21 and 22 to formulate the structure preservation objective. We validated this choice on 100 random sketches considering structure preservation and realism. In our experiments, for each input sketch, we sample the images four times (\\(R = 4\\)) with different classifier-free guidance scales (i.e. \\(\\beta \\in \\{0.0, 0.05, 0.1, 0.25\\}\\)), and select the one that maximizes the CLIP-based objective in Eq. (4). We use a linear decay mask rate schedule in all experiments, starting from \\(\\gamma(T) = 0.95\\) and stopping sampling at the mask rate \\(\\gamma(0) = 0.25\\), which results in higher realism and reduces artifacts associated with structure-based sampling. To further increase realism, we postprocess the generated samples with Token-Critic refinement [29], which adds 32 sampling iterations. See Appendix D for more details. We generate each image using \\(T = 500\\) sampling iterations, and the overall sampling time for a batch of 8 images is on average 750 seconds on a single TPUv4, including four trials and the CLIP-based evaluation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baselines We consider well-established unpaired image-to-image translation methods as baselines. Specifically, we" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1883" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.362 + ], + "angle": 0, + "content": "used CUT [37], MUNIT [25] and VQI2I [6] in our comparisons. We note that for sketch-to-photo translation, methods that use ground truth attribute information to translate from one attribute to another, e.g. StarGAN [7, 8], fail to minimize the gap between sketch and real domains [20]. CUT [37] uses a contrastive objective to ensure structural similarity between the corresponding patches of the input image and the translation result. MUNIT [25] is a GAN-based model that uses latent embedding reconstruction losses to disentangle appearance from structure. VQI2I [6] uses a vector-quantized GAN to encode images into sequences of tokens representing the structure and appearance of the input images, and uses embedding reconstruction losses to enforce the disentanglement of the structure. Since these methods are not class-conditional, we trained them on each class separately. We report the average result across the examples of all classes as well as the results of training on the entire datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.472, + 0.454 + ], + "angle": 0, + "content": "Although MaskSketch does not utilize paired data, we also consider as baseline the state-of-the-art paired sketch-to-photo method CoGS [20]. We note that VQI2I, CoGS and MUNIT allow diverse sampling with an additional appearance image or vector as input, whereas MaskSketch samples diverse results by varying the random seed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Datasets For qualitative evaluation of MaskSketch, we propose OpenSketches, a novel dataset made of 200 openly licensed sketches. OpenSketches contains real sketches drawn with pencil and paper, as well as digital sketches. Furthermore, to mimic realistic, highly detailed sketches, we utilized the open source implementation of Stable Diffusion [44] to generate input examples. All the sketches shown in this manuscript are from OpenSketches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.473, + 0.88 + ], + "angle": 0, + "content": "For quantitative evaluation, we considered two datasets: ImageNet-Sketch [50], a dataset of 50 real sketches of 1000 classes of ImageNet-2012 [9] and the Pseudosketches dataset [20], consisting of pairs of ground truth real images and their corresponding automatically extracted edge maps from 125 classes from the ImageNet21K [43] dataset. We present qualitative results for these datasets in Appendix C1. In our quantitative experiments, we report the results on two versions of the datasets: 1) full: using all examples from each of the datasets, and 2) 10-class: using the 10 classes that are reported to result in the highest-quality translation results in CoGS [20]: \"songbird\", \"pizza\", \"volcano\", \"zebra\", \"castle\", \"door\", \"shark\", \"mushroom\", \"cup\", \"lion\". The 10-class subsets of Pseudosketches and ImageNet-Sketch consist of 1,749 and 508 examples respectively, whereas the full datasets consist of 113,370 examples and 52,888 examples, respectively. For the 10-class subsets, we trained unpaired image translation baselines that are not class-conditional on each class separately and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "reported the aggregated results over all 10 classes for a fair comparison with the class-conditional MaskSketch. For the full version of the datasets, we train the baseline methods on all classes without class conditioning. Since ImageNet-Sketch does not provide ground truth paired data, it is impossible to train CoGS [20] on this dataset, therefore we use the model trained on Pseudosketches for both datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.895, + 0.48 + ], + "angle": 0, + "content": "Metrics Quantitative evaluation of sketch-to-photo translation consists of two aspects: evaluation of realism of the generation results, and evaluation of structure fidelity with respect to the input sketch. To estimate realism, we use the FID score [23]. To assess generation diversity, we used the LPIPS-based diversity score [36], which computes the average LPIPS [56] distance between the generated examples. For a fair comparison and due to the limited number of samples in the 10-class subsets, we report the FID and LPIPS results over 10,000 examples generated with different 'appearance' inputs with the baseline methods CoGS, MUNIT and VQI2I, and with different seeds for MaskS-ketch. For CUT, we diversify the generated set with augmentations. The FID score is computed with respect to the images from ImageNet [9] for the ImageNet-Sketch experiments, and with respect to the ground truth Pseudosketches images for the Pseudosketches experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.895, + 0.664 + ], + "angle": 0, + "content": "To provide additional quantitative evaluation of structure preservation quality and realism, we also report the two CLIP-based metrics defined in Sec. 3.4: image feature distance and prompt similarity score. The CLIP feature distance metric is more appropriate for the evaluation of structure preservation quality than the edge-based metrics [20] since the CLIP features are more invariant to the domain gap as shown in the recent works on image-to-sketch translation [49]. We note that these metrics are identical to the CLIP-based rejection sampling in Sec. 3, and we include the quantitative results without CLIP-based sampling in Appendix C.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "User Preference Studies Quantitative evaluation of structure fidelity is challenging due to the distribution shift between the shapes of real objects and abstract sketches and outlines. To complement the quantitative results, we performed user preference studies. We asked users the question: \"Given the task of converting the sketch shown on the left into a realistic photo, which result do you prefer?\" Users were asked to pick one result among the five compared methods (CoGS, MUNIT, VQI2I, CUT and MaskSketch) according to their preference. We collected three preference evaluations for each example in the 10-class ImageNet-Sketch and Pseudosketches datasets. Finally, we counted only unanimous votes to guarantee statistical significance. Please see Appendix E for more details." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.416, + 0.901 + ], + "angle": 0, + "content": "1 Not shown in the main manuscript due to copyright concerns." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1884" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.873, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.489, + 0.867, + 0.504 + ], + "angle": 0, + "content": "Figure 6. Example translations by MaskSketch on the OpenSketches dataset. The model takes as input a sketch and a class label." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.164, + 0.543 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.473, + 0.829 + ], + "angle": 0, + "content": "Quantitative Results In Tab. 1, we report the quantitative evaluation results on the 10-class subsets of ImageNet-Sketch and Pseudosketch. Additionally, we report the FID and LPIPS diversity scores over the full ImageNet-Sketch and Pseudosketch datasets in Tab. 2. The results on the 10-class subsets indicate an advantage of MaskSketch in terms of realism and diversity, with a two-fold decrease in the FID score compared to the baseline MUNIT on both ImageNet-Sketch and Pseudosketch datasets. In our experiments, MaskSketch outperformed the baselines on the entire ImageNet-Sketch dataset of real sketches, including the fully-supervised CoGS on the Pseudosketches. Notably, general image translation methods, such as MUNIT and CUT, outperform the fully-supervised CoGS in a class-supervised setup. As seen from the FID and LPIPS results, VQI2I struggles to generalize on the relatively small ImageNet-Sketch dataset that contains only 50 examples in each class, mainly due to mode collapse." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results Qualitative comparison shows that the baseline image translation methods, including the supervised CoGS, capture the overall layout and outlines of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.895, + 0.803 + ], + "angle": 0, + "content": "the input sketch but sometimes fail to produce realistic results. For instance, the GAN-based architectures, namely CUT and MUNIT, produce structurally similar results by practically recoloring the input sketch, which results in a sub-par realism, especially on the more abstract sketches. In our experiments, the VQ-GAN-based VQI2I model failed to learn the correspondences between hand-drawn sketches from ImageNet-Sketch and images from the real photo domain due to a limited number of examples in ImageNet-Sketch, therefore we observe a severe mode collapse on most classes. The fully-supervised CoGS sometimes failed to produce realistic and semantically meaningful results, especially on the hand-drawn sketches. MaskSketch achieved a good balance between realism and structure fidelity on the majority of sketches from ImageNet-Sketch. However, MaskSketch struggled to preserve structure on some examples from Pseudosketches due to the extreme complexity of the extracted edge maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To further refine the resulting token sequence to produce a more realistic image, we use token-critic refinement as proposed by Lezama et al. [29]. Token-critic approach uses a discriminator to detect and refine the tokens that decrease the probability of the overall sequence." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1885" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.855, + 0.321 + ], + "angle": 0, + "content": "
supervisionFID ↓LPIPS ↑CLIP prompt ↑CLIP feat. ↓User preference ↑
ImageNet-Sketch
MUNITclass68.650.5852.1030.2710.70%
CUTclass77.740.6865.5928.0519.78%
VQ-I2Iclass181.770.3253.7631.050%
CoGSclass + pairs97.310.6456.6229.528.55%
MaskSketch(ours)class33.240.7867.1026.6359.35%
Pseudosketches
MUNITclass93.230.6941.9127.6523.08%
CUTclass112.110.4245.6727.6225.0%
VQ-I2Iclass169.10.7734.528.470.64%
CoGSclass + pairs102.660.6834.7927.5214.10%
MaskSketch (ours)class56.550.7859.4825.6035.25%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.331, + 0.856, + 0.346 + ], + "angle": 0, + "content": "Table 1. Sketch-to-photo translation performance on ImageNet-Sketch (top) and Pseudosketches (bottom) 10-classes subsets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.37, + 0.466, + 0.498 + ], + "angle": 0, + "content": "
ImageNet-SketchPseudosketches
FID ↓LPIPS ↑FID ↓LPIPS ↑
MUNIT113.450.74121.690.71
CUT161.330.74163.820.70
VQI2I131.700.72135.470.71
CoGS (sup.)85.090.7249.310.71
MaskSketch23.890.7746.440.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.508, + 0.471, + 0.537 + ], + "angle": 0, + "content": "Table 2. Comparison on the full ImageNet-Sketch (left) and Pseudosketches-validation (right) datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Limitations The main limitation of MaskSketch is computational efficiency. To achieve a successful optimization of the structural constraint, MaskSketch requires significantly more sampling iterations than the regular MaskGIT. Furthermore, to improve the stability of results it was necessary to apply a multiple trials rejection scheme. Two other important limitations for MaskSketch are the coarse granularity of the attention maps in a transformer, and the flexibility of the prior model, in our case an ImageNet-pretrained MaskGIT. Figure 7 illustrates the common failure cases of our method: out-of-distribution scene composition scarcely or not represented in the training set of MaskGIT, multiple objects forming an unrealistic scene, as well as the complex scenes with multiple foreground and background objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.798, + 0.196, + 0.814 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We proposed MaskSketch, a sketch-guided image generation method that allows control over the spatial layout of the generation result. MaskSketch achieves high realism and structure preservation without pairwise supervision, does not require model finetuning and works on sketches of" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.371, + 0.889, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.612, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Figure 7. Failure cases of MaskSketch on the hand-drawn sketch examples: out-of-distribution composition and shapes, complex scenes containing multiple objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.817 + ], + "angle": 0, + "content": "various levels of abstraction. We show that the self-attention maps of the intermediate layers of a masked generative transformer encode important structural information of the input image and are sufficiently domain-invariant, which allows their use in a structure similarity constraint. Our experimental results show that the proposed attention-based sampling approach outperforms state-of-the-art sketch-tophoto and general image translation methods in terms of both realism and structure fidelity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.835, + 0.893, + 0.896 + ], + "angle": 0, + "content": "Acknowledgements We thank Tali Dekel, Huiwen Chang, Lu Jiang, and David Salesin for their insightful advice and guidance. This work was done during an internship at Google Research." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1886" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.469, + 0.296 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.469, + 0.353 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.409 + ], + "angle": 0, + "content": "[5] Wengling Chen and James Hays. Sketchygan: Towards diverse and realistic sketch to image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9416-9425, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[6] Yu-Jie Chen, Shin-I Cheng, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Vector quantized image-to-image translation. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[7] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.469, + 0.591 + ], + "angle": 0, + "content": "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[12] Mathias Eitz, Ronald Richter, Kristian Hildebrand, Tamy Boubekeur, and Marc Alexa. Photosketcher: interactive sketch-based image synthesis. IEEE Computer Graphics and Applications, 31(6):56-66, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[13] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[16] Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Arnab Ghosh, Richard Zhang, Puneet K Dokania, Oliver Wang, Alexei A Efros, Philip HS Torr, and Eli Shechtman. Interactive sketch & fill: Multiclass sketch-to-image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1171-1180, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[19] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. arXiv preprint arXiv:2111.14822, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[20] Cusuh Ham, Gemma Canet Tarres, Tu Bui, James Hays, Zhe Lin, and John Collomosse. Cogs: Controllable generation and search from sketch and style. European Conference on Computer Vision, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[21] Jonathan Heek, Anselm Levskaya, Avital Oliver, Marvin Ritter, Bertrand Rondepierre, Andreas Steiner, and Marc van Zee. Flax: A neural network library and ecosystem for JAX, 2020. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[25] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In ECCV, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[26] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1887" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[29] José Lezama, Huiwen Chang, Lu Jiang, , and Irfan Essa. Improved masked image generation with token-critic. European Conference on Computer Vision, 2022. 3, 5, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[31] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[32] Yongyi Lu, Shangzhe Wu, Yu-Wing Tai, and Chi-Keung Tang. Image generation from sketch constraint using contextual gan. In Proceedings of the European conference on computer vision (ECCV), pages 205–220, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[33] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[36] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.469, + 0.637 + ], + "angle": 0, + "content": "[37] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In European conference on computer vision, pages 319-345. Springer, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[38] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[42] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[43] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[47] Vadim Sushko, Edgar Schonfeld, Dan Zhang, Juergen Gall, Bernt Schiele, and Anna Khoreva. You only need adversarial supervision for semantic image synthesis. arXiv preprint arXiv:2012.04781, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[48] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. arXiv preprint arXiv:2201.00424, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[49] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[50] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[51] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. In arXiv, 2022. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[52] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. N\\`uwa: Visual synthesis pre-training for neural visual world creation. arXiv preprint arXiv:2111.12417, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[53] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[54] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[55] Pan Zhang, Bo Zhang, Dong Chen, Lu Yuan, and Fang Wen. Cross-domain correspondence learning for exemplar-based image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5143-5153, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1888" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.218 + ], + "angle": 0, + "content": "[57] Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocosnet v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11465–11475, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.468, + 0.288 + ], + "angle": 0, + "content": "[58] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1889" + } + ] +] \ No newline at end of file diff --git a/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_origin.pdf b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3634535ce700519d89a905bfdb8ee891c2edba5e --- /dev/null +++ b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/73d92261-ba42-4e3c-9334-1f86ad52f64d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b01dcbdb925d6b8254c443943d4ba761f2848df6321c6e564907ee8ff498720e +size 4658500 diff --git a/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/full.md b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ca990efd551def5836bfa2396f00c69fe119a35b --- /dev/null +++ b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/full.md @@ -0,0 +1,268 @@ +# MaskSketch: Unpaired Structure-guided Masked Image Generation + +Dina Bashkirova1*, José Lezama2, Kihyuk Sohn2, Kate Saenko1,3, Irfan Essa2,4 + +1Boston University, 2Google Research, 3MIT-IBM Watson AI Lab, 4Georgia Institute of Technology + +# Abstract + +Recent conditional image generation methods produce images of remarkable diversity, fidelity and realism. However, the majority of these methods allow conditioning only on labels or text prompts, which limits their level of control over the generation result. In this paper, we introduce MaskSketch, an image generation method that allows spatial conditioning of the generation result using a guiding sketch as an extra conditioning signal during sampling. MaskSketch utilizes a pre-trained masked generative transformer, requiring no model training or paired supervision, and works with input sketches of different levels of abstraction. We show that intermediate self-attention maps of a masked generative transformer encode important structural information of the input image, such as scene layout and object shape, and we propose a novel sampling method based on this observation to enable structure-guided generation. Our results show that MaskSketch achieves high image realism and fidelity to the guiding structure. Evaluated on standard benchmark datasets, MaskSketch outperforms state-of-the-art methods for sketch-to-image translation, as well as unpaired image-to-image translation approaches. The code can be found on our project website: https://masksketch.github.io/ + +# 1. Introduction + +Recent Image generation methods achieved remarkable success, allowing diverse and photorealistic image synthesis [4, 11, 44, 46]. The majority of state-of-the-art generative models allow conditioning with class labels [2, 4, 11, 13] or text prompts [40, 41, 44, 46]. However, some applications require a more fine-grained control over the spatial composition of the generation result. While methods conditioned with segmentation maps [14] or strokes [34] achieve some spatial control over the generated image, sketching allows a more fine-grained specification of the target spatial layout, which makes it desirable for many creative applications. + +In this paper, we propose MaskSketch, a method for conditional image synthesis that uses sketch guidance to de + +![](images/91011f0629c36a9af325754519411ba63b01eb471e32025c05cc51181d096fd1.jpg) +Figure 1. Given an input sketch and class label, MaskSketch samples realistic images that follow the given structure. MaskSketch works on sketches of various degree of abstraction by leveraging a pre-trained masked image generator [4], while not requiring model finetuning or pairwise supervision. + +fine the desired structure, and a pre-trained state-of-the-art masked generative transformer, MaskGIT [4], to leverage a strong generative prior. We demonstrate the capability of MaskSketch to generate realistic images of a given structure for sketch-to-photo image translation. Sketch-to-photo [5, 20, 32] is one of the most challenging applications of structure-conditional generation due to the large domain gap between sketches and natural images. MaskSketch achieves a balance between realism and structure fidelity. Our experiments show that MaskSketch outperforms state-of-the-art sketch-to-photo [20] and unpaired image translation methods [6, 25, 37], according to standard metrics for image generation [23] and user preference studies. + +In MaskSketch, we formulate a structure similarity constraint based on the observation that the intermediate self-attention maps of a masked generative transformer [4] encode rich structural information (see Fig. 2). We use this + +![](images/5901986f3e6127051d29fc7885def284cca4e7d07e8fb88cdf39b7cb61810b16.jpg) +Figure 2. Self-attention maps (PCA) of the intermediate layers of a pre-trained masked generative transformer [4] encode information about the spatial layout of the input. Notably, they are robust to the domain shift between natural images (left) and sketches (right). + +structure similarity constraint to guide the generated image towards the desired spatial layout [22, 48]. Our study shows that the proposed attention-based structure similarity objective is robust to the domain shift occurring in sketch-to-photo translation. The proposed structure-based sampling method leverages a pre-trained image generator, and does not require model finetuning or sketch-photo paired data. Moreover, it is significantly faster than other methods that exploit self-attention maps for guided image generation [48]. Figure 1 shows the translation results produced by our method on sketches of various levels of abstraction. + +The limitations of existing sketch-to-photo translation methods [5, 20, 32] come from having to learn both an implicit natural domain prior and the mapping that aligns sketches to natural images, for which the domain gap is severe. MaskSketch, on the other hand, uses the strong generative prior of a pre-trained generative transformer, which allows highly realistic generation. In addition, MaskSketch uses the domain-invariant self-attention maps for structure conditioning, allowing its use on sketches of a wide range of abstraction levels. + +Our contributions can be summarized as follows: + +- We show that the self-attention maps of a masked generative transformer encode important structural information and are robust to the domain shift between images and sketches. +- We propose a sampling method based on self-attention similarity, balancing the structural guidance of an input sketch and the natural image prior. +- We demonstrate that the proposed sampling approach, MaskSketch, outperforms state-of-the-art methods in unpaired sketch-to-photo translation. +- To the best of our knowledge, MaskSketch is the first method for sketch-to-photo translation in the existing literature that produces photorealistic results requiring only class label supervision. + +# 2. Related Work + +While there is a vast volume of literature on image generative models thanks to recent progress ranging from generative adversarial networks [2, 18, 27] generative transformers [4, 13, 54] and diffusion models [11, 35, 40, 46], in this section, we focus on reviewing image-conditioned image generation, also known as image translation. + +Supervised image conditional generation Sketch-to-photo image translation is a special case of image-conditional image generation. Early conditional image generation methods were based on generative adversarial networks. For example, pix2pix [26] conditioned the generation result by minimizing the patchwise distance between the ground truth and the generated image; SPADE [38] and OASIS [47] used spatially-adaptive instance normalization to condition generation on a segmentation map; CoCosNet [55], CoCosNet V2 [57] warped the reference image using a correlation matrix between the image and the given segmentation map. Similarly to MaskSketch, Makea-Scene and NUWA [14, 52] are designed to condition generation on semantic segmentation and text prompts with a VQ-based transformer. While these methods allow spatial conditioning, they are inapplicable for sketch-to-photo due to the lack of ground truth paired data, domain gap between sketches and segmentation maps and lack of efficient methods that extract semantic segmentation from sketches. + +Unsupervised image-conditional generation In unsupervised image-conditioned translation, the ground truth input and translation pairs are not available for training. For example, CycleGAN [58] used a cycle reconstruction loss to ensure a semantically consistent translation, UNIT [31], MUNIT [25], and StarGANv2 [8] disentangled domain-specific and shared information between the source and target image domains by mapping them to a shared latent embedding space. PSP [42] used StyleGAN [46] inversion along with style mixing for segmentation- and edge-guided translation. SDEdit [33] uses a diffusion model to translate the input strokes or segmentation maps to natural images. + +The closest work to ours in this line may be SpliceVIT [48], which uses self-attention key self-similarity extracted from the discriminative ViT (Dino [3]) to represent the structure of an input image. As pointed in [48], SpliceVIT works only in case when both the input and expected output images come from the same domain, which makes it inapplicable for sketch-to-photo translation. VQ-I2I [6] is another work on unsupervised image translation that leverages the generative power of a VQ-GAN [13]-based generative transformer. Unlike MaskSketch, VQ-I2I uses the embedding reconstruction loss for controllable generation. + +Recently, [15, 28, 45] also demonstrated how to leverage pre-trained image generators for image synthesis based on novel conditioning inputs. These methods allow to replicate + +![](images/e9b19bd7055960bb16ca14f85251f7484f28dbecc1da231c76acd4eddfaaafe8.jpg) +Figure 3. Left: confidence-based token rejection (masking) in MaskGIT. Right: structure-based token rejection. Confidence-based rejection masks out the least 'likely' tokens, while structure-based rejection masks the tokens with the highest structural distance (Eq. 1) w.r.t. the input sketch. + +![](images/9821e08a1aeb7ae80e6b3ea7219402b4e1790d5ab33b43137f8166dd54580ce6.jpg) + +a given object or subject in the generated image. However, they do not allow to finely specify the spatial layout of the generated image as MaskSketch and typically require some degree of fine-tuning. Prompt-to-prompt tuning [22] uses the attention features to perform spatially aligned prompt-conditional generation. General image-conditional methods show remarkable results when the source and target domains are visually similar, e.g., translating horses to zebras, performing artistic style transfer, etc, however, they tend to struggle on the more challenging sketch-to-photo translation task. MaskSketch shows a promising alternative for transferring the spatial composition from sketches since it requires no paired data or model training, thanks to leveraging a powerful pre-trained generator. + +Sketch-to-photo translation The sketch-to-photo application received attention in recent years thanks to the advancement in the field of image generation. For example, SketchyGAN [5] proposes an a GAN-based approach based on edge-preserving image augmentations, ContextualGAN [32] leverages conditional GAN along with joint image-sketch representation, iSketchNFill [17] uses a gating mechanism to condition output images on the class label and an MUNIT-based generator to synthesize images with diverse appearance. Photosketcher [12] uses sketch-based image retrieval to compose a real image. PITI [51] is pretrained with ground truth edge maps and semantic segmentation maps to learn a domain invariant semantic representation. The state-of-the-art supervised method CoGS [20] learns the structure embeddings using the ground truth image-sketch pairs in the vector-quantized space of a VQGAN [13]. In contrast, MaskSketch does not rely on paired data for training, which allows it to use sketches of different abstraction levels, as well as real photos. + +# 3. Method + +In this section, we describe the main components in MaskSketch that introduce sketch-guided spatial control to a conditional masked image generator. We first review + +masked image generation in Section 3.1. Then we introduce the two main components of MaskSketch, a structure similarity distance in Section 3.2, and structure-guided parallel sampling in Section 3.3. Finally, we discuss how to balance the trade-off between structure fidelity and generation realism in Section 3.4. + +# 3.1. Background: Masked Image Generation + +Masked image generation is a state-of-the-art approach for efficient generation [4, 19, 29], combining the strengths of masked token modeling [10] and non-autoregressive sampling [16]. It encodes images as discrete sequences of visual tokens using a VQ-GAN encoder [13], and then trains a bi-directional transformer (BERT [10]) to model natural image distribution in the discrete token sequence space. Generation is performed iteratively, where significant gains in efficiency are obtained by using parallel sampling instead of auto-regressive sampling. MaskGIT [4] starts from a blank canvas where all visual tokens are masked. At each sampling iteration, all the missing tokens are sampled in parallel, and a rejection criteria is used, where the tokens with low model likelihood are masked and will be re-predicted in the next refinement iteration. See Figure 3 (left) for an illustration of a single MaskGIT decoding step. MaskSketch extends the parallel sampling of MaskGIT to sample images that follow the structure determined by an input image (Fig. 3, right), as described in the following sections. + +# 3.2. Structure Similarity via Attention Maps + +We consider two images to be structurally similar when their self-similarity maps are close to each other. MaskSketch leverages the self-similarity encoded in the self-attention maps of a masked generative transformer (Section 3.1) to define structural distance. One key observation in our work is that a class-conditional MaskGIT trained on ImageNet shows a high degree of domain invariance in its attention maps and is able to capture the self-similarity in out-of-distribution domains such as sketches (Fig. 2). + +![](images/bf9299ba1b57de4fc69ded82acaed74c811198ef8f2fe2fe5462aa14b618fecc.jpg) +Figure 4. Structure-guided sampling using attention maps from the first three layers of a masked generative transformer results in a nearly perfect reconstruction of the input, whereas using the final layers (16, 18, 20), out of 24, yields realistic images with a similar structure but highly diverse appearance. (Best viewed in color.) + +Formally, we define a structural distance based on a comparison of self-attention maps. Let $\mathbb{Z}$ be the indices representing the VQ-GAN [13] dictionary of vector-quantized image tokens. Let $\pmb{x} \in \mathbb{Z}^N$ be sequence of $N$ discrete tokens obtained using a VQ-GAN encoding an input image in the vector-quantized space. Given an input image $\pmb{x}$ and a generated image $\pmb{y}$ , let $A^\ell(\pmb{x}) \in [0,1]^{N \times N}$ be the transformer self-attention map at layer $\ell$ . Each row in $A^\ell(\pmb{x})$ represents the attention weights of each token with respect to all tokens, normalized with a softmax function. We define the structural distance between the $i^{\text{th}}$ tokens of images $\pmb{x}$ and $\pmb{y}$ across layers $\mathcal{L}$ as: + +$$ +d _ {S} ^ {i} (\boldsymbol {x}, \boldsymbol {y}) = \sum_ {\ell \in \mathcal {L}} d _ {J} \left(A _ {i} ^ {\ell} (\boldsymbol {x}), A _ {i} ^ {\ell} (\boldsymbol {y})\right), \tag {1} +$$ + +where $d_J$ is the Jeffrey's divergence: + +$$ +d _ {J} (\boldsymbol {u}, \boldsymbol {v}) = \frac {K L (\boldsymbol {u} \| \boldsymbol {v}) + K L (\boldsymbol {v} \| \boldsymbol {u})}{2}. \tag {2} +$$ + +Intuitively, the image regions represented by the $i^{\mathrm{th}}$ tokens of $\pmb{x}$ and $\pmb{y}$ are structurally similar if their distributions of attention self-similarities are close to each other. + +# 3.3. Structure-guided Parallel Decoding + +MaskSketch adapts the parallel sampling of MaskGIT to take into account the structural similarity between the output and the reference input sketch. More precisely, the token rejection criteria in each decoding iteration is modified to also reject the sampled tokens that have low self-similarity score (1). The proposed structure-guided decoding strategy can also be seen as a greedy optimization technique that balances minimizing the structural distance and following the model's image prior. + +While MaskGIT sampling rejects token candidates with the lowest likelihood by masking them at the end of each decoding iteration, MaskSketch creates an additional mask that rejects tokens based on the structural similarity to the + +![](images/72cccc7200eda24c43d656909af9a14ae96ca5b9ed6f322a518f2f09205d485c.jpg) + +![](images/142ce0aaf7d496ede3ead55be73944a7d69ed632945ee0f453305d2fdf2bc6cd.jpg) +Figure 5. Top: Traversal of the realism-fidelity trade-off by varying the classifier-free guidance scale. Bottom: Increasing the guidance scale leads to higher realism and a high CLIP similarity to the prompt "photo of a $c$ ", where $c$ is the class name, at the cost of lower structure fidelity and lower CLIP feature similarity. + +input sketch (see Fig. 3). At the end of one decoding iteration, we compute the logical OR between the confidence-based and the structure-based masks to optimize both realism and structure similarity (Sec. 3.4). The pseudocode of our algorithm is described in Algorithm 1. It relies on the function sample_mask, which takes as input a vector of structure similarity scores $s^s$ and the number of masked elements $k$ , and samples a mask by Gumbel top- $k$ using $s^s$ to mask the tokens with the highest structure distance. + +The choice of layers $\mathcal{L}$ selected for computing the structure similarity significantly impacts the generation results. Our experiments show that sampling with the attention maps extracted from the first layers results in nearly identical reconstruction of the given input image, whereas minimizing the structure distance based on the last layers results in images of diverse appearance that are spatially aligned with the input image at a high level, as shown in Fig. 4. + +# 3.4. Structure fidelity vs realism trade-off + +One of the biggest challenges in sketch-to-real translation is the immense domain gap between the source and target domains. The input sketches and natural images differ + +significantly not only in appearance, but also in the distribution of shapes and spatial composition. Due to the domain gap, optimization based solely on the structure distance often results in structurally similar but unrealistic images. To overcome this issue, we propose a combined masking approach that optimizes both structure fidelity and realism. + +To navigate this trade-off, we use a parameter $\lambda_s \in [0,1]$ to determine the proportion of tokens masked according to the structure similarity scores and those masked according to the model confidence or likelihood scores. Given an overall masking rate schedule function $\gamma(t)$ at step $t$ , the structure-based mask rate is computed as $\lambda_s \gamma(t)$ , whereas the confidence-based mask rate is $(1 - \lambda_s) \gamma(t)$ . Two independent masks, $m_t^s$ and $m_t^c$ , are computed for the structure-based and confidence-based scores, respectively. The final mask at iteration $t$ is then computed as the logical OR between $m_t^s$ and $m_t^c$ . Please refer to Fig. 9 in the Appendix for an ablation study on the parameter $\lambda_s$ . + +Classifier-free Guidance To further improve the level of realism in the translation result, we use classifier-free guidance [14, 24, 35] when computing the model likelihood scores. Specifically, for a given sequence of sampled tokens $\bar{\pmb{y}}$ and input image $\pmb{x}$ , we use the pre-trained generator $G$ to compute the per-token logits $\log p(\bar{\pmb{y}}(i)|\pmb{x}, c)$ conditioned on the correct class $c$ and logits conditioned on a random class $r$ : $\log p(\bar{\pmb{y}}(i)|\pmb{x}, r)$ , and calibrate the final confidence-based scores as follows: + +$$ +\begin{array}{l} \boldsymbol {s} ^ {c} (i) = \log p (\bar {\boldsymbol {y}} (i) | \boldsymbol {x}, c) \\ - \beta (\log p (\bar {\mathbf {y}} (i) | \boldsymbol {x}, c) - \log p (\bar {\mathbf {y}} (i) | \boldsymbol {x}, r)) \tag {3} \\ \end{array} +$$ + +where $\beta$ is the classifier-free guidance scale. Figure 5 shows how varying $\beta$ affects the fidelity-realism trade-off. + +Global CLIP-based rejection sampling Minimization of the structure similarity distance in the space of visual + +# Algorithm 1 MaskSketch sampling + +Input: Pre-trained BERT generator $G$ , structure and confidence masking schedule function $\gamma(t)$ , structure-based sampling ratio $\lambda_s$ , input sketch $x$ , layer(s) $\ell$ . + +Output: Generated image encoding $y_0$ + +1: $A^{\ell}(\pmb {x})\gets \mathrm{attn\_map}(G,\pmb {x},\ell)$ +2: Initialize $\mathbf{y}_T$ +3: for $t = T - 1 \dots 0$ do +4: $\bar{\pmb{y}}_t, \pmb{s}_t^c = G(\pmb{y}_{t + 1})$ +5: $A^{\ell}(\bar{\mathbf{y}}_t) = \mathrm{attn\_map}(G,\bar{\mathbf{y}}_t,\ell)$ +6: $\pmb{s}_t^s\gets \left\{d_J\left(A_i^\ell (\bar{\pmb{y}}_t),A_i^\ell (\pmb {x})\right)\right\}_{i = 1\dots N}$ +7: $\pmb{m}_{t}^{s} = \text{sample\_mask}(\pmb{s}_{t}^{s}, \lfloor \lambda_{s}\gamma(\overline{t}) \cdot N \rfloor)$ +8: $\pmb{m}_{t}^{c} = \text{sample\_mask}(\pmb{s}_{t}^{c},\lfloor (1 - \lambda_{s})\gamma (t)\cdot N\rfloor)$ +9: $\pmb{m}_t = \pmb{m}_t^s \vee \pmb{m}_t^c$ +10: $\pmb{y}_t = \bar{\pmb{y}}_t \odot \pmb{m}_t$ +11: end for + +tokens is a discrete optimization problem that cannot be efficiently tackled via continuous optimization methods such as gradient descent. Moreover, since MaskGIT was trained to minimize a different objective, such a greedy optimization process requires more iterations than regular sampling. To increase the stability of the proposed method, we improve the overall fidelity by producing multiple translation samples for a given sketch with different random seeds and guidance scales $\beta$ , and selecting the image that yields the highest structure fidelity and realism according to a CLIP-based score. Inspired by the recent success in photo-to-sketch mapping with CLIP [39] domain-invariant representations [49], we use the $L_{1}$ distance between features of a CLIP encoder $\mathrm{CLIP}_s(\pmb{x},\pmb{y})$ of the input image $\pmb{x}$ of class $c$ and generation result $\pmb{y}$ to estimate the structure similarity. We also use the CLIP similarity score $\mathrm{CLIP}_r(c,\pmb{y})$ between the translated image and the corresponding prompt prompt $(c) = \text{'photo of a } c'$ to assess the realism for each generated example, more details can be found in Appendix F. We normalize the scores across $R$ trials and keep the result with the highest overall quality score: + +$$ +\boldsymbol {y} _ {\text {f i n a l}} = \underset {\boldsymbol {y} \in \left\{\boldsymbol {y} _ {1} \dots \boldsymbol {y} _ {R} \right\}} {\operatorname {a r g m a x}} (1 - \operatorname {C L I P} _ {s} (\boldsymbol {x}, \boldsymbol {y})) ^ {2} \operatorname {C L I P} _ {r} (c, \boldsymbol {y}) \tag {4} +$$ + +Tab. 5 in the Appendix shows how the proposed CLIP-based selection approach improves the overall generation result as the number of sampling trials increases. + +# 4. Experiments + +Experimental Setup In all experiments, we used a class-conditional MaskGIT model pretrained on the ImageNet 2012 [9] dataset with the output resolution $256 \times 256$ . We used layers 1, 3, 16, 20, 21 and 22 to formulate the structure preservation objective. We validated this choice on 100 random sketches considering structure preservation and realism. In our experiments, for each input sketch, we sample the images four times ( $R = 4$ ) with different classifier-free guidance scales (i.e. $\beta \in \{0.0, 0.05, 0.1, 0.25\}$ ), and select the one that maximizes the CLIP-based objective in Eq. (4). We use a linear decay mask rate schedule in all experiments, starting from $\gamma(T) = 0.95$ and stopping sampling at the mask rate $\gamma(0) = 0.25$ , which results in higher realism and reduces artifacts associated with structure-based sampling. To further increase realism, we postprocess the generated samples with Token-Critic refinement [29], which adds 32 sampling iterations. See Appendix D for more details. We generate each image using $T = 500$ sampling iterations, and the overall sampling time for a batch of 8 images is on average 750 seconds on a single TPUv4, including four trials and the CLIP-based evaluation. + +Baselines We consider well-established unpaired image-to-image translation methods as baselines. Specifically, we + +used CUT [37], MUNIT [25] and VQI2I [6] in our comparisons. We note that for sketch-to-photo translation, methods that use ground truth attribute information to translate from one attribute to another, e.g. StarGAN [7, 8], fail to minimize the gap between sketch and real domains [20]. CUT [37] uses a contrastive objective to ensure structural similarity between the corresponding patches of the input image and the translation result. MUNIT [25] is a GAN-based model that uses latent embedding reconstruction losses to disentangle appearance from structure. VQI2I [6] uses a vector-quantized GAN to encode images into sequences of tokens representing the structure and appearance of the input images, and uses embedding reconstruction losses to enforce the disentanglement of the structure. Since these methods are not class-conditional, we trained them on each class separately. We report the average result across the examples of all classes as well as the results of training on the entire datasets. + +Although MaskSketch does not utilize paired data, we also consider as baseline the state-of-the-art paired sketch-to-photo method CoGS [20]. We note that VQI2I, CoGS and MUNIT allow diverse sampling with an additional appearance image or vector as input, whereas MaskSketch samples diverse results by varying the random seed. + +Datasets For qualitative evaluation of MaskSketch, we propose OpenSketches, a novel dataset made of 200 openly licensed sketches. OpenSketches contains real sketches drawn with pencil and paper, as well as digital sketches. Furthermore, to mimic realistic, highly detailed sketches, we utilized the open source implementation of Stable Diffusion [44] to generate input examples. All the sketches shown in this manuscript are from OpenSketches. + +For quantitative evaluation, we considered two datasets: ImageNet-Sketch [50], a dataset of 50 real sketches of 1000 classes of ImageNet-2012 [9] and the Pseudosketches dataset [20], consisting of pairs of ground truth real images and their corresponding automatically extracted edge maps from 125 classes from the ImageNet21K [43] dataset. We present qualitative results for these datasets in Appendix C1. In our quantitative experiments, we report the results on two versions of the datasets: 1) full: using all examples from each of the datasets, and 2) 10-class: using the 10 classes that are reported to result in the highest-quality translation results in CoGS [20]: "songbird", "pizza", "volcano", "zebra", "castle", "door", "shark", "mushroom", "cup", "lion". The 10-class subsets of Pseudosketches and ImageNet-Sketch consist of 1,749 and 508 examples respectively, whereas the full datasets consist of 113,370 examples and 52,888 examples, respectively. For the 10-class subsets, we trained unpaired image translation baselines that are not class-conditional on each class separately and + +reported the aggregated results over all 10 classes for a fair comparison with the class-conditional MaskSketch. For the full version of the datasets, we train the baseline methods on all classes without class conditioning. Since ImageNet-Sketch does not provide ground truth paired data, it is impossible to train CoGS [20] on this dataset, therefore we use the model trained on Pseudosketches for both datasets. + +Metrics Quantitative evaluation of sketch-to-photo translation consists of two aspects: evaluation of realism of the generation results, and evaluation of structure fidelity with respect to the input sketch. To estimate realism, we use the FID score [23]. To assess generation diversity, we used the LPIPS-based diversity score [36], which computes the average LPIPS [56] distance between the generated examples. For a fair comparison and due to the limited number of samples in the 10-class subsets, we report the FID and LPIPS results over 10,000 examples generated with different 'appearance' inputs with the baseline methods CoGS, MUNIT and VQI2I, and with different seeds for MaskS-ketch. For CUT, we diversify the generated set with augmentations. The FID score is computed with respect to the images from ImageNet [9] for the ImageNet-Sketch experiments, and with respect to the ground truth Pseudosketches images for the Pseudosketches experiments. + +To provide additional quantitative evaluation of structure preservation quality and realism, we also report the two CLIP-based metrics defined in Sec. 3.4: image feature distance and prompt similarity score. The CLIP feature distance metric is more appropriate for the evaluation of structure preservation quality than the edge-based metrics [20] since the CLIP features are more invariant to the domain gap as shown in the recent works on image-to-sketch translation [49]. We note that these metrics are identical to the CLIP-based rejection sampling in Sec. 3, and we include the quantitative results without CLIP-based sampling in Appendix C.1. + +User Preference Studies Quantitative evaluation of structure fidelity is challenging due to the distribution shift between the shapes of real objects and abstract sketches and outlines. To complement the quantitative results, we performed user preference studies. We asked users the question: "Given the task of converting the sketch shown on the left into a realistic photo, which result do you prefer?" Users were asked to pick one result among the five compared methods (CoGS, MUNIT, VQI2I, CUT and MaskSketch) according to their preference. We collected three preference evaluations for each example in the 10-class ImageNet-Sketch and Pseudosketches datasets. Finally, we counted only unanimous votes to guarantee statistical significance. Please see Appendix E for more details. + +![](images/60c68998a43d5b1e58cf3e725656fcd73ca8e54fe9bb88c7e7c9b7695a45b280.jpg) +Figure 6. Example translations by MaskSketch on the OpenSketches dataset. The model takes as input a sketch and a class label. + +# 5. Results + +Quantitative Results In Tab. 1, we report the quantitative evaluation results on the 10-class subsets of ImageNet-Sketch and Pseudosketch. Additionally, we report the FID and LPIPS diversity scores over the full ImageNet-Sketch and Pseudosketch datasets in Tab. 2. The results on the 10-class subsets indicate an advantage of MaskSketch in terms of realism and diversity, with a two-fold decrease in the FID score compared to the baseline MUNIT on both ImageNet-Sketch and Pseudosketch datasets. In our experiments, MaskSketch outperformed the baselines on the entire ImageNet-Sketch dataset of real sketches, including the fully-supervised CoGS on the Pseudosketches. Notably, general image translation methods, such as MUNIT and CUT, outperform the fully-supervised CoGS in a class-supervised setup. As seen from the FID and LPIPS results, VQI2I struggles to generalize on the relatively small ImageNet-Sketch dataset that contains only 50 examples in each class, mainly due to mode collapse. + +Qualitative Results Qualitative comparison shows that the baseline image translation methods, including the supervised CoGS, capture the overall layout and outlines of + +the input sketch but sometimes fail to produce realistic results. For instance, the GAN-based architectures, namely CUT and MUNIT, produce structurally similar results by practically recoloring the input sketch, which results in a sub-par realism, especially on the more abstract sketches. In our experiments, the VQ-GAN-based VQI2I model failed to learn the correspondences between hand-drawn sketches from ImageNet-Sketch and images from the real photo domain due to a limited number of examples in ImageNet-Sketch, therefore we observe a severe mode collapse on most classes. The fully-supervised CoGS sometimes failed to produce realistic and semantically meaningful results, especially on the hand-drawn sketches. MaskSketch achieved a good balance between realism and structure fidelity on the majority of sketches from ImageNet-Sketch. However, MaskSketch struggled to preserve structure on some examples from Pseudosketches due to the extreme complexity of the extracted edge maps. + +To further refine the resulting token sequence to produce a more realistic image, we use token-critic refinement as proposed by Lezama et al. [29]. Token-critic approach uses a discriminator to detect and refine the tokens that decrease the probability of the overall sequence. + +
supervisionFID ↓LPIPS ↑CLIP prompt ↑CLIP feat. ↓User preference ↑
ImageNet-Sketch
MUNITclass68.650.5852.1030.2710.70%
CUTclass77.740.6865.5928.0519.78%
VQ-I2Iclass181.770.3253.7631.050%
CoGSclass + pairs97.310.6456.6229.528.55%
MaskSketch(ours)class33.240.7867.1026.6359.35%
Pseudosketches
MUNITclass93.230.6941.9127.6523.08%
CUTclass112.110.4245.6727.6225.0%
VQ-I2Iclass169.10.7734.528.470.64%
CoGSclass + pairs102.660.6834.7927.5214.10%
MaskSketch (ours)class56.550.7859.4825.6035.25%
+ +Table 1. Sketch-to-photo translation performance on ImageNet-Sketch (top) and Pseudosketches (bottom) 10-classes subsets. + +
ImageNet-SketchPseudosketches
FID ↓LPIPS ↑FID ↓LPIPS ↑
MUNIT113.450.74121.690.71
CUT161.330.74163.820.70
VQI2I131.700.72135.470.71
CoGS (sup.)85.090.7249.310.71
MaskSketch23.890.7746.440.78
+ +Table 2. Comparison on the full ImageNet-Sketch (left) and Pseudosketches-validation (right) datasets. + +Limitations The main limitation of MaskSketch is computational efficiency. To achieve a successful optimization of the structural constraint, MaskSketch requires significantly more sampling iterations than the regular MaskGIT. Furthermore, to improve the stability of results it was necessary to apply a multiple trials rejection scheme. Two other important limitations for MaskSketch are the coarse granularity of the attention maps in a transformer, and the flexibility of the prior model, in our case an ImageNet-pretrained MaskGIT. Figure 7 illustrates the common failure cases of our method: out-of-distribution scene composition scarcely or not represented in the training set of MaskGIT, multiple objects forming an unrealistic scene, as well as the complex scenes with multiple foreground and background objects. + +# 6. Conclusion + +We proposed MaskSketch, a sketch-guided image generation method that allows control over the spatial layout of the generation result. MaskSketch achieves high realism and structure preservation without pairwise supervision, does not require model finetuning and works on sketches of + +![](images/07a3e9d0399af67b3e02c8598853dd84df257780625aae71cb3c8827c87c454f.jpg) +Figure 7. Failure cases of MaskSketch on the hand-drawn sketch examples: out-of-distribution composition and shapes, complex scenes containing multiple objects. + +various levels of abstraction. We show that the self-attention maps of the intermediate layers of a masked generative transformer encode important structural information of the input image and are sufficiently domain-invariant, which allows their use in a structure similarity constraint. Our experimental results show that the proposed attention-based sampling approach outperforms state-of-the-art sketch-tophoto and general image translation methods in terms of both realism and structure fidelity. + +Acknowledgements We thank Tali Dekel, Huiwen Chang, Lu Jiang, and David Salesin for their insightful advice and guidance. This work was done during an internship at Google Research. + +# References + +[1] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 12 +[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1, 2 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2 +[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1, 2, 3 +[5] Wengling Chen and James Hays. Sketchygan: Towards diverse and realistic sketch to image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9416-9425, 2018. 1, 2, 3 +[6] Yu-Jie Chen, Shin-I Cheng, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Vector quantized image-to-image translation. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6 +[7] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 6 +[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 6 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 6 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 +[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2 +[12] Mathias Eitz, Ronald Richter, Kristian Hildebrand, Tamy Boubekeur, and Marc Alexa. Photosketcher: interactive sketch-based image synthesis. IEEE Computer Graphics and Applications, 31(6):56-66, 2011. 3 +[13] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020. 1, 2, 3, 4 +[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1, 2, 5 + +[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[16] Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324, 2019. 3 +[17] Arnab Ghosh, Richard Zhang, Puneet K Dokania, Oliver Wang, Alexei A Efros, Philip HS Torr, and Eli Shechtman. Interactive sketch & fill: Multiclass sketch-to-image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1171-1180, 2019. 3 +[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[19] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. arXiv preprint arXiv:2111.14822, 2021. 3 +[20] Cusuh Ham, Gemma Canet Tarres, Tu Bui, James Hays, Zhe Lin, and John Collomosse. Cogs: Controllable generation and search from sketch and style. European Conference on Computer Vision, 2022. 1, 2, 3, 6 +[21] Jonathan Heek, Anselm Levskaya, Avital Oliver, Marvin Ritter, Bertrand Rondepierre, Andreas Steiner, and Marc van Zee. Flax: A neural network library and ecosystem for JAX, 2020. 12 +[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3 +[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 1, 6 +[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5 +[25] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In ECCV, 2018. 1, 2, 6 +[26] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[28] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2 + +[29] José Lezama, Huiwen Chang, Lu Jiang, , and Irfan Essa. Improved masked image generation with token-critic. European Conference on Computer Vision, 2022. 3, 5, 7, 13 +[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 14 +[31] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in neural information processing systems, 30, 2017. 2 +[32] Yongyi Lu, Shangzhe Wu, Yu-Wing Tai, and Chi-Keung Tang. Image generation from sketch constraint using contextual gan. In Proceedings of the European conference on computer vision (ECCV), pages 205–220, 2018. 1, 2, 3 +[33] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 2 +[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 1 +[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 5 +[36] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 6 +[37] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In European conference on computer vision, pages 319-345. Springer, 2020. 1, 6 +[38] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5 +[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2 +[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1 + +[42] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2 +[43] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 6 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 6 +[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2 +[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2 +[47] Vadim Sushko, Edgar Schonfeld, Dan Zhang, Juergen Gall, Bernt Schiele, and Anna Khoreva. You only need adversarial supervision for semantic image synthesis. arXiv preprint arXiv:2012.04781, 2020. 2 +[48] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. arXiv preprint arXiv:2201.00424, 2022. 2 +[49] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching, 2022. 5, 6 +[50] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 6 +[51] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. In arXiv, 2022. 3, 14 +[52] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. N\`uwa: Visual synthesis pre-training for neural visual world creation. arXiv preprint arXiv:2111.12417, 2021. 2 +[53] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 14 +[54] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2 +[55] Pan Zhang, Bo Zhang, Dong Chen, Lu Yuan, and Fang Wen. Cross-domain correspondence learning for exemplar-based image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5143-5153, 2020. 2 + +[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[57] Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocosnet v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11465–11475, 2021. 2 +[58] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2 \ No newline at end of file diff --git a/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/images.zip b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5858bf3d415ae987a7367b0d85a83904961fe2ec --- /dev/null +++ b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f7de7158b7037e0a3927d6f087d4388fe078bc3feb8c5f45cc0bf8db92f0b70 +size 646326 diff --git a/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/layout.json b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..35eaed9125ba54685552e668b5563ed70d2cbb6b --- /dev/null +++ b/2023/MaskSketch_ Unpaired Structure-Guided Masked Image Generation/layout.json @@ -0,0 +1,7749 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "type": "text", + "content": "MaskSketch: Unpaired Structure-guided Masked Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 142, + 492, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 142, + 492, + 157 + ], + "spans": [ + { + "bbox": [ + 100, + 142, + 492, + 157 + ], + "type": "text", + "content": "Dina Bashkirova1*, José Lezama2, Kihyuk Sohn2, Kate Saenko1,3, Irfan Essa2,4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 157, + 538, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 538, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 538, + 172 + ], + "type": "text", + "content": "1Boston University, 2Google Research, 3MIT-IBM Watson AI Lab, 4Georgia Institute of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 224, + 290, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 290, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 290, + 499 + ], + "type": "text", + "content": "Recent conditional image generation methods produce images of remarkable diversity, fidelity and realism. However, the majority of these methods allow conditioning only on labels or text prompts, which limits their level of control over the generation result. In this paper, we introduce MaskSketch, an image generation method that allows spatial conditioning of the generation result using a guiding sketch as an extra conditioning signal during sampling. MaskSketch utilizes a pre-trained masked generative transformer, requiring no model training or paired supervision, and works with input sketches of different levels of abstraction. We show that intermediate self-attention maps of a masked generative transformer encode important structural information of the input image, such as scene layout and object shape, and we propose a novel sampling method based on this observation to enable structure-guided generation. Our results show that MaskSketch achieves high image realism and fidelity to the guiding structure. Evaluated on standard benchmark datasets, MaskSketch outperforms state-of-the-art methods for sketch-to-image translation, as well as unpaired image-to-image translation approaches. The code can be found on our project website: https://masksketch.github.io/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 541, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 672 + ], + "type": "text", + "content": "Recent Image generation methods achieved remarkable success, allowing diverse and photorealistic image synthesis [4, 11, 44, 46]. The majority of state-of-the-art generative models allow conditioning with class labels [2, 4, 11, 13] or text prompts [40, 41, 44, 46]. However, some applications require a more fine-grained control over the spatial composition of the generation result. While methods conditioned with segmentation maps [14] or strokes [34] achieve some spatial control over the generated image, sketching allows a more fine-grained specification of the target spatial layout, which makes it desirable for many creative applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "type": "text", + "content": "In this paper, we propose MaskSketch, a method for conditional image synthesis that uses sketch guidance to de" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 197, + 547, + 427 + ], + "blocks": [ + { + "bbox": [ + 307, + 197, + 547, + 427 + ], + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 427 + ], + "type": "image", + "image_path": "91011f0629c36a9af325754519411ba63b01eb471e32025c05cc51181d096fd1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 435, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 305, + 435, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 547, + 491 + ], + "type": "text", + "content": "Figure 1. Given an input sketch and class label, MaskSketch samples realistic images that follow the given structure. MaskSketch works on sketches of various degree of abstraction by leveraging a pre-trained masked image generator [4], while not requiring model finetuning or pairwise supervision." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 509, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 664 + ], + "type": "text", + "content": "fine the desired structure, and a pre-trained state-of-the-art masked generative transformer, MaskGIT [4], to leverage a strong generative prior. We demonstrate the capability of MaskSketch to generate realistic images of a given structure for sketch-to-photo image translation. Sketch-to-photo [5, 20, 32] is one of the most challenging applications of structure-conditional generation due to the large domain gap between sketches and natural images. MaskSketch achieves a balance between realism and structure fidelity. Our experiments show that MaskSketch outperforms state-of-the-art sketch-to-photo [20] and unpaired image translation methods [6, 25, 37], according to standard metrics for image generation [23] and user preference studies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "In MaskSketch, we formulate a structure similarity constraint based on the observation that the intermediate self-attention maps of a masked generative transformer [4] encode rich structural information (see Fig. 2). We use this" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 201, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 201, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 201, + 713 + ], + "type": "text", + "content": "*Work done during an internship at Google." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1879" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 187 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 187 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 187 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 187 + ], + "type": "image", + "image_path": "5901986f3e6127051d29fc7885def284cca4e7d07e8fb88cdf39b7cb61810b16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 197, + 288, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 288, + 242 + ], + "type": "text", + "content": "Figure 2. Self-attention maps (PCA) of the intermediate layers of a pre-trained masked generative transformer [4] encode information about the spatial layout of the input. Notably, they are robust to the domain shift between natural images (left) and sketches (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 395 + ], + "type": "text", + "content": "structure similarity constraint to guide the generated image towards the desired spatial layout [22, 48]. Our study shows that the proposed attention-based structure similarity objective is robust to the domain shift occurring in sketch-to-photo translation. The proposed structure-based sampling method leverages a pre-trained image generator, and does not require model finetuning or sketch-photo paired data. Moreover, it is significantly faster than other methods that exploit self-attention maps for guided image generation [48]. Figure 1 shows the translation results produced by our method on sketches of various levels of abstraction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 400, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 519 + ], + "type": "text", + "content": "The limitations of existing sketch-to-photo translation methods [5, 20, 32] come from having to learn both an implicit natural domain prior and the mapping that aligns sketches to natural images, for which the domain gap is severe. MaskSketch, on the other hand, uses the strong generative prior of a pre-trained generative transformer, which allows highly realistic generation. In addition, MaskSketch uses the domain-invariant self-attention maps for structure conditioning, allowing its use on sketches of a wide range of abstraction levels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 525, + 258, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 525, + 258, + 536 + ], + "spans": [ + { + "bbox": [ + 59, + 525, + 258, + 536 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 543, + 286, + 709 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 57, + 543, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 543, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 543, + 286, + 590 + ], + "type": "text", + "content": "- We show that the self-attention maps of a masked generative transformer encode important structural information and are robust to the domain shift between images and sketches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 591, + 286, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 286, + 625 + ], + "type": "text", + "content": "- We propose a sampling method based on self-attention similarity, balancing the structural guidance of an input sketch and the natural image prior." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 627, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 627, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 57, + 627, + 286, + 661 + ], + "type": "text", + "content": "- We demonstrate that the proposed sampling approach, MaskSketch, outperforms state-of-the-art methods in unpaired sketch-to-photo translation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 663, + 286, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 663, + 286, + 709 + ], + "spans": [ + { + "bbox": [ + 57, + 663, + 286, + 709 + ], + "type": "text", + "content": "- To the best of our knowledge, MaskSketch is the first method for sketch-to-photo translation in the existing literature that produces photorealistic results requiring only class label supervision." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": "While there is a vast volume of literature on image generative models thanks to recent progress ranging from generative adversarial networks [2, 18, 27] generative transformers [4, 13, 54] and diffusion models [11, 35, 40, 46], in this section, we focus on reviewing image-conditioned image generation, also known as image translation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 168, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 546, + 396 + ], + "type": "text", + "content": "Supervised image conditional generation Sketch-to-photo image translation is a special case of image-conditional image generation. Early conditional image generation methods were based on generative adversarial networks. For example, pix2pix [26] conditioned the generation result by minimizing the patchwise distance between the ground truth and the generated image; SPADE [38] and OASIS [47] used spatially-adaptive instance normalization to condition generation on a segmentation map; CoCosNet [55], CoCosNet V2 [57] warped the reference image using a correlation matrix between the image and the given segmentation map. Similarly to MaskSketch, Makea-Scene and NUWA [14, 52] are designed to condition generation on semantic segmentation and text prompts with a VQ-based transformer. While these methods allow spatial conditioning, they are inapplicable for sketch-to-photo due to the lack of ground truth paired data, domain gap between sketches and segmentation maps and lack of efficient methods that extract semantic segmentation from sketches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 546, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 545 + ], + "type": "text", + "content": "Unsupervised image-conditional generation In unsupervised image-conditioned translation, the ground truth input and translation pairs are not available for training. For example, CycleGAN [58] used a cycle reconstruction loss to ensure a semantically consistent translation, UNIT [31], MUNIT [25], and StarGANv2 [8] disentangled domain-specific and shared information between the source and target image domains by mapping them to a shared latent embedding space. PSP [42] used StyleGAN [46] inversion along with style mixing for segmentation- and edge-guided translation. SDEdit [33] uses a diffusion model to translate the input strokes or segmentation maps to natural images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 546, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 677 + ], + "type": "text", + "content": "The closest work to ours in this line may be SpliceVIT [48], which uses self-attention key self-similarity extracted from the discriminative ViT (Dino [3]) to represent the structure of an input image. As pointed in [48], SpliceVIT works only in case when both the input and expected output images come from the same domain, which makes it inapplicable for sketch-to-photo translation. VQ-I2I [6] is another work on unsupervised image translation that leverages the generative power of a VQ-GAN [13]-based generative transformer. Unlike MaskSketch, VQ-I2I uses the embedding reconstruction loss for controllable generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Recently, [15, 28, 45] also demonstrated how to leverage pre-trained image generators for image synthesis based on novel conditioning inputs. These methods allow to replicate" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1880" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 50, + 288, + 184 + ], + "blocks": [ + { + "bbox": [ + 50, + 50, + 288, + 184 + ], + "lines": [ + { + "bbox": [ + 50, + 50, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 50, + 50, + 288, + 184 + ], + "type": "image", + "image_path": "e9b19bd7055960bb16ca14f85251f7484f28dbecc1da231c76acd4eddfaaafe8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "type": "text", + "content": "Figure 3. Left: confidence-based token rejection (masking) in MaskGIT. Right: structure-based token rejection. Confidence-based rejection masks out the least 'likely' tokens, while structure-based rejection masks the tokens with the highest structural distance (Eq. 1) w.r.t. the input sketch." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 50, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 294, + 50, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 294, + 50, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 294, + 50, + 545, + 185 + ], + "type": "image", + "image_path": "9821e08a1aeb7ae80e6b3ea7219402b4e1790d5ab33b43137f8166dd54580ce6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 238, + 290, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 406 + ], + "type": "text", + "content": "a given object or subject in the generated image. However, they do not allow to finely specify the spatial layout of the generated image as MaskSketch and typically require some degree of fine-tuning. Prompt-to-prompt tuning [22] uses the attention features to perform spatially aligned prompt-conditional generation. General image-conditional methods show remarkable results when the source and target domains are visually similar, e.g., translating horses to zebras, performing artistic style transfer, etc, however, they tend to struggle on the more challenging sketch-to-photo translation task. MaskSketch shows a promising alternative for transferring the spatial composition from sketches since it requires no paired data or model training, thanks to leveraging a powerful pre-trained generator." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 288, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 643 + ], + "type": "text", + "content": "Sketch-to-photo translation The sketch-to-photo application received attention in recent years thanks to the advancement in the field of image generation. For example, SketchyGAN [5] proposes an a GAN-based approach based on edge-preserving image augmentations, ContextualGAN [32] leverages conditional GAN along with joint image-sketch representation, iSketchNFill [17] uses a gating mechanism to condition output images on the class label and an MUNIT-based generator to synthesize images with diverse appearance. Photosketcher [12] uses sketch-based image retrieval to compose a real image. PITI [51] is pretrained with ground truth edge maps and semantic segmentation maps to learn a domain invariant semantic representation. The state-of-the-art supervised method CoGS [20] learns the structure embeddings using the ground truth image-sketch pairs in the vector-quantized space of a VQGAN [13]. In contrast, MaskSketch does not rely on paired data for training, which allows it to use sketches of different abstraction levels, as well as real photos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 656, + 103, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 103, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 103, + 668 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "In this section, we describe the main components in MaskSketch that introduce sketch-guided spatial control to a conditional masked image generator. We first review" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 238, + 546, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 546, + 308 + ], + "type": "text", + "content": "masked image generation in Section 3.1. Then we introduce the two main components of MaskSketch, a structure similarity distance in Section 3.2, and structure-guided parallel sampling in Section 3.3. Finally, we discuss how to balance the trade-off between structure fidelity and generation realism in Section 3.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 319, + 518, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 518, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 518, + 332 + ], + "type": "text", + "content": "3.1. Background: Masked Image Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 338, + 546, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 578 + ], + "type": "text", + "content": "Masked image generation is a state-of-the-art approach for efficient generation [4, 19, 29], combining the strengths of masked token modeling [10] and non-autoregressive sampling [16]. It encodes images as discrete sequences of visual tokens using a VQ-GAN encoder [13], and then trains a bi-directional transformer (BERT [10]) to model natural image distribution in the discrete token sequence space. Generation is performed iteratively, where significant gains in efficiency are obtained by using parallel sampling instead of auto-regressive sampling. MaskGIT [4] starts from a blank canvas where all visual tokens are masked. At each sampling iteration, all the missing tokens are sampled in parallel, and a rejection criteria is used, where the tokens with low model likelihood are masked and will be re-predicted in the next refinement iteration. See Figure 3 (left) for an illustration of a single MaskGIT decoding step. MaskSketch extends the parallel sampling of MaskGIT to sample images that follow the structure determined by an input image (Fig. 3, right), as described in the following sections." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 586, + 516, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 516, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 516, + 600 + ], + "type": "text", + "content": "3.2. Structure Similarity via Attention Maps" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We consider two images to be structurally similar when their self-similarity maps are close to each other. MaskSketch leverages the self-similarity encoded in the self-attention maps of a masked generative transformer (Section 3.1) to define structural distance. One key observation in our work is that a class-conditional MaskGIT trained on ImageNet shows a high degree of domain invariance in its attention maps and is able to capture the self-similarity in out-of-distribution domains such as sketches (Fig. 2)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1881" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 286, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 286, + 171 + ], + "type": "image", + "image_path": "bf9299ba1b57de4fc69ded82acaed74c811198ef8f2fe2fe5462aa14b618fecc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 181, + 288, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 181, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 288, + 237 + ], + "type": "text", + "content": "Figure 4. Structure-guided sampling using attention maps from the first three layers of a masked generative transformer results in a nearly perfect reconstruction of the input, whereas using the final layers (16, 18, 20), out of 24, yields realistic images with a similar structure but highly diverse appearance. (Best viewed in color.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": "Formally, we define a structural distance based on a comparison of self-attention maps. Let " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " be the indices representing the VQ-GAN [13] dictionary of vector-quantized image tokens. Let " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{Z}^N" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " be sequence of " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " discrete tokens obtained using a VQ-GAN encoding an input image in the vector-quantized space. Given an input image " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " and a generated image " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "A^\\ell(\\pmb{x}) \\in [0,1]^{N \\times N}" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " be the transformer self-attention map at layer " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": ". Each row in " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "A^\\ell(\\pmb{x})" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " represents the attention weights of each token with respect to all tokens, normalized with a softmax function. We define the structural distance between the " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " tokens of images " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " across layers " + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 253, + 289, + 397 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 406, + 287, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 406, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 91, + 406, + 287, + 432 + ], + "type": "interline_equation", + "content": "d _ {S} ^ {i} (\\boldsymbol {x}, \\boldsymbol {y}) = \\sum_ {\\ell \\in \\mathcal {L}} d _ {J} \\left(A _ {i} ^ {\\ell} (\\boldsymbol {x}), A _ {i} ^ {\\ell} (\\boldsymbol {y})\\right), \\tag {1}", + "image_path": "f138f5d77b3a60504e24522743eab37c33009f607d38e09bed91cce007e23b02.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 441, + 198, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 198, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 198, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 441, + 198, + 453 + ], + "type": "inline_equation", + "content": "d_J" + }, + { + "bbox": [ + 47, + 441, + 198, + 453 + ], + "type": "text", + "content": " is the Jeffrey's divergence:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 460, + 287, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 460, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 88, + 460, + 287, + 485 + ], + "type": "interline_equation", + "content": "d _ {J} (\\boldsymbol {u}, \\boldsymbol {v}) = \\frac {K L (\\boldsymbol {u} \\| \\boldsymbol {v}) + K L (\\boldsymbol {v} \\| \\boldsymbol {u})}{2}. \\tag {2}", + "image_path": "839bada2c12bee52274fe6c5e1c8e06bd78abf635e79aac9d7e6e1e7854fea49.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "type": "text", + "content": "Intuitively, the image regions represented by the " + }, + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "type": "text", + "content": " tokens of " + }, + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 495, + 287, + 531 + ], + "type": "text", + "content": " are structurally similar if their distributions of attention self-similarities are close to each other." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 539, + 237, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 237, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 237, + 552 + ], + "type": "text", + "content": "3.3. Structure-guided Parallel Decoding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "content": "MaskSketch adapts the parallel sampling of MaskGIT to take into account the structural similarity between the output and the reference input sketch. More precisely, the token rejection criteria in each decoding iteration is modified to also reject the sampled tokens that have low self-similarity score (1). The proposed structure-guided decoding strategy can also be seen as a greedy optimization technique that balances minimizing the structural distance and following the model's image prior." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "While MaskGIT sampling rejects token candidates with the lowest likelihood by masking them at the end of each decoding iteration, MaskSketch creates an additional mask that rejects tokens based on the structural similarity to the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 552, + 228 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 552, + 228 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 552, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 552, + 228 + ], + "type": "image", + "image_path": "72cccc7200eda24c43d656909af9a14ae96ca5b9ed6f322a518f2f09205d485c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 236, + 550, + 373 + ], + "blocks": [ + { + "bbox": [ + 318, + 236, + 550, + 373 + ], + "lines": [ + { + "bbox": [ + 318, + 236, + 550, + 373 + ], + "spans": [ + { + "bbox": [ + 318, + 236, + 550, + 373 + ], + "type": "image", + "image_path": "142ce0aaf7d496ede3ead55be73944a7d69ed632945ee0f453305d2fdf2bc6cd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 385, + 546, + 441 + ], + "lines": [ + { + "bbox": [ + 304, + 385, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 546, + 441 + ], + "type": "text", + "content": "Figure 5. Top: Traversal of the realism-fidelity trade-off by varying the classifier-free guidance scale. Bottom: Increasing the guidance scale leads to higher realism and a high CLIP similarity to the prompt \"photo of a " + }, + { + "bbox": [ + 304, + 385, + 546, + 441 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 385, + 546, + 441 + ], + "type": "text", + "content": "\", where " + }, + { + "bbox": [ + 304, + 385, + 546, + 441 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 385, + 546, + 441 + ], + "type": "text", + "content": " is the class name, at the cost of lower structure fidelity and lower CLIP feature similarity." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "text", + "content": "input sketch (see Fig. 3). At the end of one decoding iteration, we compute the logical OR between the confidence-based and the structure-based masks to optimize both realism and structure similarity (Sec. 3.4). The pseudocode of our algorithm is described in Algorithm 1. It relies on the function sample_mask, which takes as input a vector of structure similarity scores " + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "inline_equation", + "content": "s^s" + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "text", + "content": " and the number of masked elements " + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "text", + "content": ", and samples a mask by Gumbel top- " + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "inline_equation", + "content": "s^s" + }, + { + "bbox": [ + 304, + 447, + 546, + 555 + ], + "type": "text", + "content": " to mask the tokens with the highest structure distance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 555, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 651 + ], + "type": "text", + "content": "The choice of layers " + }, + { + "bbox": [ + 304, + 555, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 555, + 545, + 651 + ], + "type": "text", + "content": " selected for computing the structure similarity significantly impacts the generation results. Our experiments show that sampling with the attention maps extracted from the first layers results in nearly identical reconstruction of the given input image, whereas minimizing the structure distance based on the last layers results in images of diverse appearance that are spatially aligned with the input image at a high level, as shown in Fig. 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 658, + 503, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 503, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 503, + 671 + ], + "type": "text", + "content": "3.4. Structure fidelity vs realism trade-off" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": "One of the biggest challenges in sketch-to-real translation is the immense domain gap between the source and target domains. The input sketches and natural images differ" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1882" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "significantly not only in appearance, but also in the distribution of shapes and spatial composition. Due to the domain gap, optimization based solely on the structure distance often results in structurally similar but unrealistic images. To overcome this issue, we propose a combined masking approach that optimizes both structure fidelity and realism." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": "To navigate this trade-off, we use a parameter " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "\\lambda_s \\in [0,1]" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": " to determine the proportion of tokens masked according to the structure similarity scores and those masked according to the model confidence or likelihood scores. Given an overall masking rate schedule function " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "\\gamma(t)" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": ", the structure-based mask rate is computed as " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "\\lambda_s \\gamma(t)" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": ", whereas the confidence-based mask rate is " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "(1 - \\lambda_s) \\gamma(t)" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": ". Two independent masks, " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "m_t^s" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "m_t^c" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": ", are computed for the structure-based and confidence-based scores, respectively. The final mask at iteration " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": " is then computed as the logical OR between " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "m_t^s" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "m_t^c" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": ". Please refer to Fig. 9 in the Appendix for an ablation study on the parameter " + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "inline_equation", + "content": "\\lambda_s" + }, + { + "bbox": [ + 46, + 144, + 286, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "content": "Classifier-free Guidance To further improve the level of realism in the translation result, we use classifier-free guidance [14, 24, 35] when computing the model likelihood scores. Specifically, for a given sequence of sampled tokens " + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{y}}" + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "content": " and input image " + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "content": ", we use the pre-trained generator " + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "content": " to compute the per-token logits " + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "inline_equation", + "content": "\\log p(\\bar{\\pmb{y}}(i)|\\pmb{x}, c)" + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "content": " conditioned on the correct class " + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "content": " and logits conditioned on a random class " + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "inline_equation", + "content": "\\log p(\\bar{\\pmb{y}}(i)|\\pmb{x}, r)" + }, + { + "bbox": [ + 46, + 292, + 286, + 398 + ], + "type": "text", + "content": ", and calibrate the final confidence-based scores as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 407, + 286, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 407, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 61, + 407, + 286, + 436 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {s} ^ {c} (i) = \\log p (\\bar {\\boldsymbol {y}} (i) | \\boldsymbol {x}, c) \\\\ - \\beta (\\log p (\\bar {\\mathbf {y}} (i) | \\boldsymbol {x}, c) - \\log p (\\bar {\\mathbf {y}} (i) | \\boldsymbol {x}, r)) \\tag {3} \\\\ \\end{array}", + "image_path": "441fd1fc7c3539f23552bdf7f6a5a6f6e733f29f3ec5f8a88a6d6d4eda44c955.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 443, + 286, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 286, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 443, + 286, + 467 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 443, + 286, + 467 + ], + "type": "text", + "content": " is the classifier-free guidance scale. Figure 5 shows how varying " + }, + { + "bbox": [ + 46, + 443, + 286, + 467 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 443, + 286, + 467 + ], + "type": "text", + "content": " affects the fidelity-realism trade-off." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 286, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 286, + 495 + ], + "type": "text", + "content": "Global CLIP-based rejection sampling Minimization of the structure similarity distance in the space of visual" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 514, + 193, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 514, + 193, + 526 + ], + "spans": [ + { + "bbox": [ + 49, + 514, + 193, + 526 + ], + "type": "text", + "content": "Algorithm 1 MaskSketch sampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "text", + "content": "Input: Pre-trained BERT generator " + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "text", + "content": ", structure and confidence masking schedule function " + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\gamma(t)" + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "text", + "content": ", structure-based sampling ratio " + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\lambda_s" + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "text", + "content": ", input sketch " + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "text", + "content": ", layer(s) " + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 528, + 286, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 564, + 206, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 206, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 206, + 575 + ], + "type": "text", + "content": "Output: Generated image encoding " + }, + { + "bbox": [ + 47, + 564, + 206, + 575 + ], + "type": "inline_equation", + "content": "y_0" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 576, + 274, + 708 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 53, + 576, + 189, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 576, + 189, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 576, + 189, + 590 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 53, + 576, + 189, + 590 + ], + "type": "inline_equation", + "content": "A^{\\ell}(\\pmb {x})\\gets \\mathrm{attn\\_map}(G,\\pmb {x},\\ell)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 590, + 117, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 117, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 117, + 601 + ], + "type": "text", + "content": "2: Initialize " + }, + { + "bbox": [ + 53, + 590, + 117, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_T" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 602, + 157, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 157, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 157, + 613 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 53, + 602, + 157, + 613 + ], + "type": "inline_equation", + "content": "t = T - 1 \\dots 0" + }, + { + "bbox": [ + 53, + 602, + 157, + 613 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 614, + 158, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 158, + 625 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 158, + 625 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 53, + 614, + 158, + 625 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{y}}_t, \\pmb{s}_t^c = G(\\pmb{y}_{t + 1})" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 625, + 209, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 209, + 637 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 209, + 637 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 53, + 625, + 209, + 637 + ], + "type": "inline_equation", + "content": "A^{\\ell}(\\bar{\\mathbf{y}}_t) = \\mathrm{attn\\_map}(G,\\bar{\\mathbf{y}}_t,\\ell)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 638, + 230, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 638, + 230, + 650 + ], + "spans": [ + { + "bbox": [ + 53, + 638, + 230, + 650 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 53, + 638, + 230, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{s}_t^s\\gets \\left\\{d_J\\left(A_i^\\ell (\\bar{\\pmb{y}}_t),A_i^\\ell (\\pmb {x})\\right)\\right\\}_{i = 1\\dots N}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 650, + 249, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 650, + 249, + 662 + ], + "spans": [ + { + "bbox": [ + 53, + 650, + 249, + 662 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 53, + 650, + 249, + 662 + ], + "type": "inline_equation", + "content": "\\pmb{m}_{t}^{s} = \\text{sample\\_mask}(\\pmb{s}_{t}^{s}, \\lfloor \\lambda_{s}\\gamma(\\overline{t}) \\cdot N \\rfloor)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 662, + 274, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 662, + 274, + 673 + ], + "spans": [ + { + "bbox": [ + 53, + 662, + 274, + 673 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 53, + 662, + 274, + 673 + ], + "type": "inline_equation", + "content": "\\pmb{m}_{t}^{c} = \\text{sample\\_mask}(\\pmb{s}_{t}^{c},\\lfloor (1 - \\lambda_{s})\\gamma (t)\\cdot N\\rfloor)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 674, + 156, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 674, + 156, + 685 + ], + "spans": [ + { + "bbox": [ + 53, + 674, + 156, + 685 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 53, + 674, + 156, + 685 + ], + "type": "inline_equation", + "content": "\\pmb{m}_t = \\pmb{m}_t^s \\vee \\pmb{m}_t^c" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 51, + 686, + 142, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 686, + 142, + 696 + ], + "spans": [ + { + "bbox": [ + 51, + 686, + 142, + 696 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 51, + 686, + 142, + 696 + ], + "type": "inline_equation", + "content": "\\pmb{y}_t = \\bar{\\pmb{y}}_t \\odot \\pmb{m}_t" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 51, + 698, + 97, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 698, + 97, + 708 + ], + "spans": [ + { + "bbox": [ + 51, + 698, + 97, + 708 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": "tokens is a discrete optimization problem that cannot be efficiently tackled via continuous optimization methods such as gradient descent. Moreover, since MaskGIT was trained to minimize a different objective, such a greedy optimization process requires more iterations than regular sampling. To increase the stability of the proposed method, we improve the overall fidelity by producing multiple translation samples for a given sketch with different random seeds and guidance scales " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": ", and selecting the image that yields the highest structure fidelity and realism according to a CLIP-based score. Inspired by the recent success in photo-to-sketch mapping with CLIP [39] domain-invariant representations [49], we use the " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": " distance between features of a CLIP encoder " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_s(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": " of the input image " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": " and generation result " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": " to estimate the structure similarity. We also use the CLIP similarity score " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_r(c,\\pmb{y})" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": " between the translated image and the corresponding prompt prompt " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "(c) = \\text{'photo of a } c'" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": " to assess the realism for each generated example, more details can be found in Appendix F. We normalize the scores across " + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": " trials and keep the result with the highest overall quality score:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 330, + 545, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 330, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 310, + 330, + 545, + 352 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {\\text {f i n a l}} = \\underset {\\boldsymbol {y} \\in \\left\\{\\boldsymbol {y} _ {1} \\dots \\boldsymbol {y} _ {R} \\right\\}} {\\operatorname {a r g m a x}} (1 - \\operatorname {C L I P} _ {s} (\\boldsymbol {x}, \\boldsymbol {y})) ^ {2} \\operatorname {C L I P} _ {r} (c, \\boldsymbol {y}) \\tag {4}", + "image_path": "1702ce90b8c25620049a6a57a73a1f826a3a4024e197221b8808e0ba482082ad.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 358, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 394 + ], + "type": "text", + "content": "Tab. 5 in the Appendix shows how the proposed CLIP-based selection approach improves the overall generation result as the number of sampling trials increases." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 403, + 386, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 386, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 386, + 417 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "text", + "content": "Experimental Setup In all experiments, we used a class-conditional MaskGIT model pretrained on the ImageNet 2012 [9] dataset with the output resolution " + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "text", + "content": ". We used layers 1, 3, 16, 20, 21 and 22 to formulate the structure preservation objective. We validated this choice on 100 random sketches considering structure preservation and realism. In our experiments, for each input sketch, we sample the images four times (" + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "inline_equation", + "content": "R = 4" + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "text", + "content": ") with different classifier-free guidance scales (i.e. " + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\beta \\in \\{0.0, 0.05, 0.1, 0.25\\}" + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "text", + "content": "), and select the one that maximizes the CLIP-based objective in Eq. (4). We use a linear decay mask rate schedule in all experiments, starting from " + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\gamma(T) = 0.95" + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "text", + "content": " and stopping sampling at the mask rate " + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\gamma(0) = 0.25" + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "text", + "content": ", which results in higher realism and reduces artifacts associated with structure-based sampling. To further increase realism, we postprocess the generated samples with Token-Critic refinement [29], which adds 32 sampling iterations. See Appendix D for more details. We generate each image using " + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "inline_equation", + "content": "T = 500" + }, + { + "bbox": [ + 304, + 424, + 545, + 674 + ], + "type": "text", + "content": " sampling iterations, and the overall sampling time for a batch of 8 images is on average 750 seconds on a single TPUv4, including four trials and the CLIP-based evaluation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Baselines We consider well-established unpaired image-to-image translation methods as baselines. Specifically, we" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1883" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "content": "used CUT [37], MUNIT [25] and VQI2I [6] in our comparisons. We note that for sketch-to-photo translation, methods that use ground truth attribute information to translate from one attribute to another, e.g. StarGAN [7, 8], fail to minimize the gap between sketch and real domains [20]. CUT [37] uses a contrastive objective to ensure structural similarity between the corresponding patches of the input image and the translation result. MUNIT [25] is a GAN-based model that uses latent embedding reconstruction losses to disentangle appearance from structure. VQI2I [6] uses a vector-quantized GAN to encode images into sequences of tokens representing the structure and appearance of the input images, and uses embedding reconstruction losses to enforce the disentanglement of the structure. Since these methods are not class-conditional, we trained them on each class separately. We report the average result across the examples of all classes as well as the results of training on the entire datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 287, + 288, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 288, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 288, + 359 + ], + "type": "text", + "content": "Although MaskSketch does not utilize paired data, we also consider as baseline the state-of-the-art paired sketch-to-photo method CoGS [20]. We note that VQI2I, CoGS and MUNIT allow diverse sampling with an additional appearance image or vector as input, whereas MaskSketch samples diverse results by varying the random seed." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 373, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 468 + ], + "type": "text", + "content": "Datasets For qualitative evaluation of MaskSketch, we propose OpenSketches, a novel dataset made of 200 openly licensed sketches. OpenSketches contains real sketches drawn with pencil and paper, as well as digital sketches. Furthermore, to mimic realistic, highly detailed sketches, we utilized the open source implementation of Stable Diffusion [44] to generate input examples. All the sketches shown in this manuscript are from OpenSketches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 469, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 289, + 696 + ], + "type": "text", + "content": "For quantitative evaluation, we considered two datasets: ImageNet-Sketch [50], a dataset of 50 real sketches of 1000 classes of ImageNet-2012 [9] and the Pseudosketches dataset [20], consisting of pairs of ground truth real images and their corresponding automatically extracted edge maps from 125 classes from the ImageNet21K [43] dataset. We present qualitative results for these datasets in Appendix C1. In our quantitative experiments, we report the results on two versions of the datasets: 1) full: using all examples from each of the datasets, and 2) 10-class: using the 10 classes that are reported to result in the highest-quality translation results in CoGS [20]: \"songbird\", \"pizza\", \"volcano\", \"zebra\", \"castle\", \"door\", \"shark\", \"mushroom\", \"cup\", \"lion\". The 10-class subsets of Pseudosketches and ImageNet-Sketch consist of 1,749 and 508 examples respectively, whereas the full datasets consist of 113,370 examples and 52,888 examples, respectively. For the 10-class subsets, we trained unpaired image translation baselines that are not class-conditional on each class separately and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "reported the aggregated results over all 10 classes for a fair comparison with the class-conditional MaskSketch. For the full version of the datasets, we train the baseline methods on all classes without class conditioning. Since ImageNet-Sketch does not provide ground truth paired data, it is impossible to train CoGS [20] on this dataset, therefore we use the model trained on Pseudosketches for both datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 176, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 547, + 380 + ], + "type": "text", + "content": "Metrics Quantitative evaluation of sketch-to-photo translation consists of two aspects: evaluation of realism of the generation results, and evaluation of structure fidelity with respect to the input sketch. To estimate realism, we use the FID score [23]. To assess generation diversity, we used the LPIPS-based diversity score [36], which computes the average LPIPS [56] distance between the generated examples. For a fair comparison and due to the limited number of samples in the 10-class subsets, we report the FID and LPIPS results over 10,000 examples generated with different 'appearance' inputs with the baseline methods CoGS, MUNIT and VQI2I, and with different seeds for MaskS-ketch. For CUT, we diversify the generated set with augmentations. The FID score is computed with respect to the images from ImageNet [9] for the ImageNet-Sketch experiments, and with respect to the ground truth Pseudosketches images for the Pseudosketches experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 382, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 547, + 525 + ], + "type": "text", + "content": "To provide additional quantitative evaluation of structure preservation quality and realism, we also report the two CLIP-based metrics defined in Sec. 3.4: image feature distance and prompt similarity score. The CLIP feature distance metric is more appropriate for the evaluation of structure preservation quality than the edge-based metrics [20] since the CLIP features are more invariant to the domain gap as shown in the recent works on image-to-sketch translation [49]. We note that these metrics are identical to the CLIP-based rejection sampling in Sec. 3, and we include the quantitative results without CLIP-based sampling in Appendix C.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "User Preference Studies Quantitative evaluation of structure fidelity is challenging due to the distribution shift between the shapes of real objects and abstract sketches and outlines. To complement the quantitative results, we performed user preference studies. We asked users the question: \"Given the task of converting the sketch shown on the left into a realistic photo, which result do you prefer?\" Users were asked to pick one result among the five compared methods (CoGS, MUNIT, VQI2I, CUT and MaskSketch) according to their preference. We collected three preference evaluations for each example in the 10-class ImageNet-Sketch and Pseudosketches datasets. Finally, we counted only unanimous votes to guarantee statistical significance. Please see Appendix E for more details." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 702, + 254, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 254, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 254, + 713 + ], + "type": "text", + "content": "1 Not shown in the main manuscript due to copyright concerns." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1884" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 534, + 376 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 534, + 376 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 534, + 376 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 534, + 376 + ], + "type": "image", + "image_path": "60c68998a43d5b1e58cf3e725656fcd73ca8e54fe9bb88c7e7c9b7695a45b280.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 387, + 530, + 399 + ], + "lines": [ + { + "bbox": [ + 62, + 387, + 530, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 530, + 399 + ], + "type": "text", + "content": "Figure 6. Example translations by MaskSketch on the OpenSketches dataset. The model takes as input a sketch and a class label." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 418, + 100, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 100, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 100, + 430 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 439, + 289, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 289, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 289, + 656 + ], + "type": "text", + "content": "Quantitative Results In Tab. 1, we report the quantitative evaluation results on the 10-class subsets of ImageNet-Sketch and Pseudosketch. Additionally, we report the FID and LPIPS diversity scores over the full ImageNet-Sketch and Pseudosketch datasets in Tab. 2. The results on the 10-class subsets indicate an advantage of MaskSketch in terms of realism and diversity, with a two-fold decrease in the FID score compared to the baseline MUNIT on both ImageNet-Sketch and Pseudosketch datasets. In our experiments, MaskSketch outperformed the baselines on the entire ImageNet-Sketch dataset of real sketches, including the fully-supervised CoGS on the Pseudosketches. Notably, general image translation methods, such as MUNIT and CUT, outperform the fully-supervised CoGS in a class-supervised setup. As seen from the FID and LPIPS results, VQI2I struggles to generalize on the relatively small ImageNet-Sketch dataset that contains only 50 examples in each class, mainly due to mode collapse." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "Qualitative Results Qualitative comparison shows that the baseline image translation methods, including the supervised CoGS, capture the overall layout and outlines of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 418, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 635 + ], + "type": "text", + "content": "the input sketch but sometimes fail to produce realistic results. For instance, the GAN-based architectures, namely CUT and MUNIT, produce structurally similar results by practically recoloring the input sketch, which results in a sub-par realism, especially on the more abstract sketches. In our experiments, the VQ-GAN-based VQI2I model failed to learn the correspondences between hand-drawn sketches from ImageNet-Sketch and images from the real photo domain due to a limited number of examples in ImageNet-Sketch, therefore we observe a severe mode collapse on most classes. The fully-supervised CoGS sometimes failed to produce realistic and semantically meaningful results, especially on the hand-drawn sketches. MaskSketch achieved a good balance between realism and structure fidelity on the majority of sketches from ImageNet-Sketch. However, MaskSketch struggled to preserve structure on some examples from Pseudosketches due to the extreme complexity of the extracted edge maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "To further refine the resulting token sequence to produce a more realistic image, we use token-critic refinement as proposed by Lezama et al. [29]. Token-critic approach uses a discriminator to detect and refine the tokens that decrease the probability of the overall sequence." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1885" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 523, + 254 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 523, + 254 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 523, + 254 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 523, + 254 + ], + "type": "table", + "html": "
supervisionFID ↓LPIPS ↑CLIP prompt ↑CLIP feat. ↓User preference ↑
ImageNet-Sketch
MUNITclass68.650.5852.1030.2710.70%
CUTclass77.740.6865.5928.0519.78%
VQ-I2Iclass181.770.3253.7631.050%
CoGSclass + pairs97.310.6456.6229.528.55%
MaskSketch(ours)class33.240.7867.1026.6359.35%
Pseudosketches
MUNITclass93.230.6941.9127.6523.08%
CUTclass112.110.4245.6727.6225.0%
VQ-I2Iclass169.10.7734.528.470.64%
CoGSclass + pairs102.660.6834.7927.5214.10%
MaskSketch (ours)class56.550.7859.4825.6035.25%
", + "image_path": "c26c89e33d03877753ea2a2045c16208005d2593d337afe1afc054bcf90ef5be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 293, + 285, + 394 + ], + "blocks": [ + { + "bbox": [ + 68, + 262, + 523, + 274 + ], + "lines": [ + { + "bbox": [ + 68, + 262, + 523, + 274 + ], + "spans": [ + { + "bbox": [ + 68, + 262, + 523, + 274 + ], + "type": "text", + "content": "Table 1. Sketch-to-photo translation performance on ImageNet-Sketch (top) and Pseudosketches (bottom) 10-classes subsets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 293, + 285, + 394 + ], + "lines": [ + { + "bbox": [ + 50, + 293, + 285, + 394 + ], + "spans": [ + { + "bbox": [ + 50, + 293, + 285, + 394 + ], + "type": "table", + "html": "
ImageNet-SketchPseudosketches
FID ↓LPIPS ↑FID ↓LPIPS ↑
MUNIT113.450.74121.690.71
CUT161.330.74163.820.70
VQI2I131.700.72135.470.71
CoGS (sup.)85.090.7249.310.71
MaskSketch23.890.7746.440.78
", + "image_path": "a9836d9af61dc45861321ad2dd2760280e045f448369fc7f2f9e1fc321f74e45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 402, + 288, + 425 + ], + "lines": [ + { + "bbox": [ + 47, + 402, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 288, + 425 + ], + "type": "text", + "content": "Table 2. Comparison on the full ImageNet-Sketch (left) and Pseudosketches-validation (right) datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 449, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 617 + ], + "type": "text", + "content": "Limitations The main limitation of MaskSketch is computational efficiency. To achieve a successful optimization of the structural constraint, MaskSketch requires significantly more sampling iterations than the regular MaskGIT. Furthermore, to improve the stability of results it was necessary to apply a multiple trials rejection scheme. Two other important limitations for MaskSketch are the coarse granularity of the attention maps in a transformer, and the flexibility of the prior model, in our case an ImageNet-pretrained MaskGIT. Figure 7 illustrates the common failure cases of our method: out-of-distribution scene composition scarcely or not represented in the training set of MaskGIT, multiple objects forming an unrealistic scene, as well as the complex scenes with multiple foreground and background objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 632, + 119, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 119, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 119, + 644 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "We proposed MaskSketch, a sketch-guided image generation method that allows control over the spatial layout of the generation result. MaskSketch achieves high realism and structure preservation without pairwise supervision, does not require model finetuning and works on sketches of" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 293, + 544, + 476 + ], + "blocks": [ + { + "bbox": [ + 310, + 293, + 544, + 476 + ], + "lines": [ + { + "bbox": [ + 310, + 293, + 544, + 476 + ], + "spans": [ + { + "bbox": [ + 310, + 293, + 544, + 476 + ], + "type": "image", + "image_path": "07a3e9d0399af67b3e02c8598853dd84df257780625aae71cb3c8827c87c454f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 484, + 545, + 518 + ], + "lines": [ + { + "bbox": [ + 305, + 484, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 484, + 545, + 518 + ], + "type": "text", + "content": "Figure 7. Failure cases of MaskSketch on the hand-drawn sketch examples: out-of-distribution composition and shapes, complex scenes containing multiple objects." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 539, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 647 + ], + "type": "text", + "content": "various levels of abstraction. We show that the self-attention maps of the intermediate layers of a masked generative transformer encode important structural information of the input image and are sufficiently domain-invariant, which allows their use in a structure similarity constraint. Our experimental results show that the proposed attention-based sampling approach outperforms state-of-the-art sketch-tophoto and general image translation methods in terms of both realism and structure fidelity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 661, + 546, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 661, + 546, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 661, + 546, + 709 + ], + "type": "text", + "content": "Acknowledgements We thank Tali Dekel, Huiwen Chang, Lu Jiang, and David Salesin for their insightful advice and guidance. This work was done during an internship at Google Research." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1886" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 287, + 178 + ], + "type": "text", + "content": "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 279 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "type": "text", + "content": "[5] Wengling Chen and James Hays. Sketchygan: Towards diverse and realistic sketch to image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9416-9425, 2018. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "text", + "content": "[6] Yu-Jie Chen, Shin-I Cheng, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Vector quantized image-to-image translation. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "text", + "content": "[7] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "text", + "content": "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 513 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[12] Mathias Eitz, Ronald Richter, Kristian Hildebrand, Tamy Boubekeur, and Marc Alexa. Photosketcher: interactive sketch-based image synthesis. IEEE Computer Graphics and Applications, 31(6):56-66, 2011. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "type": "text", + "content": "[13] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1, 2, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "text", + "content": "[16] Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. Mask-predict: Parallel decoding of conditional masked language models. arXiv preprint arXiv:1904.09324, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 172, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 227 + ], + "type": "text", + "content": "[17] Arnab Ghosh, Richard Zhang, Puneet K Dokania, Oliver Wang, Alexei A Efros, Philip HS Torr, and Eli Shechtman. Interactive sketch & fill: Multiclass sketch-to-image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1171-1180, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "text", + "content": "[19] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. arXiv preprint arXiv:2111.14822, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "text", + "content": "[20] Cusuh Ham, Gemma Canet Tarres, Tu Bui, James Hays, Zhe Lin, and John Collomosse. Cogs: Controllable generation and search from sketch and style. European Conference on Computer Vision, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "type": "text", + "content": "[21] Jonathan Heek, Anselm Levskaya, Avital Oliver, Marvin Ritter, Bertrand Rondepierre, Andreas Steiner, and Marc van Zee. Flax: A neural network library and ecosystem for JAX, 2020. 12" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "type": "text", + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "text", + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 1, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "text", + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "type": "text", + "content": "[25] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In ECCV, 2018. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "text", + "content": "[26] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[28] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1887" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[29] José Lezama, Huiwen Chang, Lu Jiang, , and Irfan Essa. Improved masked image generation with token-critic. European Conference on Computer Vision, 2022. 3, 5, 7, 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "type": "text", + "content": "[31] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "text", + "content": "[32] Yongyi Lu, Shangzhe Wu, Yu-Wing Tai, and Chi-Keung Tang. Image generation from sketch constraint using contextual gan. In Proceedings of the European conference on computer vision (ECCV), pages 205–220, 2018. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[33] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "type": "text", + "content": "[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "text", + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[36] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 504 + ], + "type": "text", + "content": "[37] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In European conference on computer vision, pages 319-345. Springer, 2020. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "type": "text", + "content": "[38] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[42] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[43] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "text", + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "type": "text", + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": "[47] Vadim Sushko, Edgar Schonfeld, Dan Zhang, Juergen Gall, Bernt Schiele, and Anna Khoreva. You only need adversarial supervision for semantic image synthesis. arXiv preprint arXiv:2012.04781, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "type": "text", + "content": "[48] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. arXiv preprint arXiv:2201.00424, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "type": "text", + "content": "[49] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching, 2022. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[50] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "text", + "content": "[51] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. In arXiv, 2022. 3, 14" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "text", + "content": "[52] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. N\\`uwa: Visual synthesis pre-training for neural visual world creation. arXiv preprint arXiv:2111.12417, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 569, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 603 + ], + "type": "text", + "content": "[53] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 14" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "text", + "content": "[54] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[55] Pan Zhang, Bo Zhang, Dong Chen, Lu Yuan, and Fang Wen. Cross-domain correspondence learning for exemplar-based image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5143-5153, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1888" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 172 + ], + "type": "text", + "content": "[57] Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocosnet v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11465–11475, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 286, + 228 + ], + "type": "text", + "content": "[58] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1889" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_content_list.json b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4d301c344dbe4774d9584e84fa094305048b614b --- /dev/null +++ b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_content_list.json @@ -0,0 +1,1475 @@ +[ + { + "type": "text", + "text": "Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond", + "text_level": 1, + "bbox": [ + 102, + 128, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang* Xiaoming Wei, Xiaolin Wei Meituan {feizhengcong, fanmingyuan, zhuli09, huangjunshi}@meituan.com {weixiaoming, weixiaolin02}@meituan.com", + "bbox": [ + 215, + 180, + 754, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 303, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masked Auto-Encoder (MAE) pretraining methods randomly mask image patches and then train a vision Transformer to reconstruct the original pixels based on the unmasked patches. While they demonstrate impressive performance for downstream vision tasks, it generally requires a large amount of training resource. In this paper, we introduce a novel Generative Adversarial Networks alike framework, referred to as GAN-MAE, where a generator is used to generate the masked patches according to the remaining visible patches, and a discriminator is employed to predict whether the patch is synthesized by the generator. We believe this capacity of distinguishing whether the image patch is predicted or original is benefit to representation learning. Another key point lies in that the parameters of the vision Transformer backbone in the generator and discriminator are shared. Extensive experiments demonstrate that adversarial training of GAN-MAE framework is more efficient and accordingly outperforms the standard MAE given the same model size, training data, and computation resource. The gains are substantially robust for different model sizes and datasets, in particular, a ViT-B model trained with GAN-MAE for 200 epochs outperforms the MAE with 1600 epochs on fine-tuning top-1 accuracy of ImageNet-1k with much less FLOPs. Besides, our approach also works well at transferring downstream tasks.", + "bbox": [ + 75, + 335, + 472, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 743, + 209, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, Transformer [62] has become the de facto standard architecture in computer vision, and has surpassed state-of-the-art Convolutional Neural Network (CNN) [31, 58] feature extractors in vision tasks through models such as the Vision Transformer [21]. Meanwhile, self-supervised learning (SSL) algorithms [12, 14, 27, 29] aims to learn transferable representation from unlabeled", + "bbox": [ + 73, + 768, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/987b48bb5a14bd972540c7a2439719338dbd2790d48f3b7bfbd7809f1f0e08bb.jpg", + "image_caption": [ + "Figure 1. Performance comparison in different pre-training epochs for ImageNet-1K Fine-tuning top-1 accuracy. Compared to MAE trained for 1600 epochs, GAN-MAE achieves comparable accuracy with much less training time at 200 epochs." + ], + "image_footnote": [], + "bbox": [ + 503, + 325, + 857, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data by performing instance-level pretext tasks, and has been a long-standing target in the vision community. Particularly, masked image modeling (MIM) in SSL for vision transformers has shown remarkably impressive downstream performance in a wide variety of computer vision tasks [3, 28], attracting increasing attention.", + "bbox": [ + 496, + 625, + 893, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MIM is a simple pretext task that first randomly masks some patches of an image, and then predicts the contents of the masked patches according to the remaining, using various reconstruction targets, e.g., visual tokens [3, 19], semantic features [1, 77] and raw pixels [28, 70]. Essentially, it learns the transferable representation by modeling the image structure itself as content prediction. While more effective than conventional pre-training, masked autoencoder modeling approaches still exist some issues: (i) reconstruction optimization with MSE loss leads to blurrier output images than the raw input, it would be better to use a more perceptual loss over pixels to guide the fine-grained seman", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The corresponding author.", + "bbox": [ + 94, + 887, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24449", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tic understanding and representation learning, leading to more plausible synthesized patches; (ii) inner dependency between masked patches is lacked [71], i.e., generation of masked image patches may lack the surrounding information. This situation becomes more serious when the image patch masking ratio is large. We alleviate this problem by introducing confident synthesized patches as complementary information during training; (iii) mask-reconstruction methods incur a substantial computation cost because the network only learns from part of the visible patches and misses the information of masked patches.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a Generative Adversarial Networks-based pre-training framework, referred to as GAN-MAE, which contains two components: a generator model learns to reconstruct the masked patches according to visible patches in the encoder-decoder architecture and a discriminator model learns to distinguish real image patches from plausible but synthesized remains. Generally, given an image from training dataset, our method first randomly masks parts of patches and reconstructs them using the rest visible patches with a generator, which serves as a standard MAE model. Then we build the corrupt image as the combination of visible and synthesis patches, which is then fed into the discriminator to predict whether each patch is from raw image or synthesized results. In this manner, the discriminator provides a valid guiding for more delicate image patch modeling. Then, with the development of generator capacity, a key advantage of discriminative task is that it integrates the synthesized patches into corrupt images as complementary information, which fills the missing inner relationship between patches during pre-training. Moreover, we shared the parameters of vision transformer backbone in the generator and discriminator to promote memory reduction, training efficiency, as well as performance enhancement.", + "bbox": [ + 75, + 261, + 472, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments follow the same architecture, settings, and pre-training recipe as MAE [28], and we find that the simple incorporation of a discriminator consistently outperforms MAE in variant models, e.g., ViT-S, ViT-B, and ViT-L, when fine-tuning for top-1 accuracy of ImageNet classification. We also conduct extensive ablation studies to validate the effectiveness of our core designs in backbone parameter sharing and advarial training. As pre-training with more epochs usually results in a better downstream performance, we argue that an important consideration for pre-training methods should be computation efficiency as well as absolute downstream performance. From this viewpoint, we also demonstrate that discrimination of pseudo-image patches forces GAN-MAE to train more efficiently than standard MAE. We further provide a comprehensive comparison with MAE in various epochs and various models and show our framework achieves consistently better performance. In particular, as presented in Figure 1, for the ViT-B model structure, our GAN-MAE achieves comparable", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ble classification performance with only 200 pre-training epochs vs. standard MAE 1600 pre-training epochs. Furthermore, the GAN-MAE achieves 0.7 points improvement when pre-training 1600 epochs. Finally, we summarize our contribution as follows:", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new and effective GAN-alike framework for visual representation self-supervised learning, which to our best knowledge is the first trial of integrating GAN idea into MAE framework. As a generic approach, we suggest that this framework can be easily applied on many other MIM-based tasks.", + "- We introduce two core designs: shared weight for the main backbones of generator and discriminator, and an adversarial training process, both of which cost fewer amounts of computing resources while obtaining appreciable performance improvements.", + "- Extensive experiments demonstrate that compared with the original MAE, our method is more compute-efficient and results in better transfer representation learning on downstream tasks." + ], + "bbox": [ + 517, + 178, + 890, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 459, + 650, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoencoding. Autoencoder [6, 7, 35] is an unsupervised learning technique for neural networks that learns efficient representations by training the network to ignore signal noise. It includes an encoder that maps the original data to a low-dimensional latent embedding and a decoder that recovers the data from the latent embedding, with the goal of learning a compressed knowledge representation. The denoising autoencoder [63, 64] learns to reconstruct clean data points from a noisy version. Numerous efforts have been devoted for image denoising, such as masking pixels [11, 55, 64], inpainting [69], removing color channels [39, 72], and shuffling image patches [22, 54]. For a broader overview of denoising autoencoder, we refer to [5, 61].", + "bbox": [ + 496, + 484, + 893, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Image Modeling. Masked language modeling [37,49,57], which generalizes well on language understanding and generation tasks, is the domain self-supervised approach in the field of NLP. Similarly, vision transformer [21, 41, 50] based masked image modeling (MIM) approaches [1,3,28,70,77] for computer vision tasks have also been developed. Generally, these MIM approaches first apply a mask to patches of an image, and then the masked patches are predicted given the visible patches. Feature representation learned through such within-image context prediction demonstrate strong transfer performance in downstream tasks. Recently, lots of works exploring MIM have been concurrently developed from different perspectives.", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24450", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b1b341053e9586bee822eb3f148d6d85160fd4bd229637788059f0a4d011d6a.jpg", + "image_caption": [ + "Figure 2. Overview of GAN-MAE framework, where a generator is used to predict the masked patches and a discriminator is employed to classify whether the patches are selected from the raw image or synthesized results. In particular, both of MAE encoder and discriminator are based on the vision Transformer backbone and share parameters for memory reduction and training efficiency." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 893, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The works include framework design [13, 23, 28, 70], prediction targets [2, 19, 66, 77], and integration with vision-language representation learning [42, 43, 75]. Our work belongs to the first group and introduces a novel GAN framework that discriminates the reconstructed image patches in a standard MAE to learn deeper semantics.", + "bbox": [ + 75, + 335, + 470, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Adversarial Networks. GANs [26] are effective at generating high-quality synthetic data. Usually, the generator generates an image, and the discriminator determines whether the input image is a real image or a generated image. Subsequently, many improvements based on the original GAN focused on speeding up the training of the network and improving the quality of the generated images [8, 52, 74]. These improvements also help GAN achieving a wider range of applications [45, 48, 67]. Methods based on GAN are also widely used in image-to-image translation [36], super-resolution [4, 40], style transfer [15], text generation [10, 24, 73], and representation learning [17, 25], to name a few. Particularly, [56] brings some designs of CNN architecture to stabilize the training of GAN framework, after that the discriminator can be directly used as feature extractor in downstream tasks. In contrast, our method proposes to integrate the GAN as assistant for the MIM task, which focuses on the study of better pattern for masking based self-supervised learning. Besides, our strategy of shared parameters provides a unified backbone for better vision representation.", + "bbox": [ + 75, + 449, + 472, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 782, + 187, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the GAN-MAE framework in details. At first, we briefly review the conventional masked autoencoder model in Sec. 3.1, and then describe the proposed generator-discriminator pre-training framework in Sec. 3.2. The architecture of our framework is presented in Figure 2. Finally, we suggest an adversarial", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "training processes under the proposed method and discuss our framework in Sec. 3.3 and Sec. 3.4, respectively.", + "bbox": [ + 498, + 335, + 890, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 376, + 640, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked Autoencoder (MAE) [28] is a self-supervised approach with a vision transformer encoder and a small transformer decoder, which randomly masks a large portion of input patches, and then reconstructs the masked patches according to the visible patches. Specifically, provided with an image $X \\in \\mathbb{R}^{C \\times H \\times W}$ , where $C$ , $H$ and $W$ are the channel number, image height and image width respectively, MAE partitions $X$ into $N = \\frac{H \\times W}{P^2}$ non-overlapping patches with patch size $P$ . In this way, the image is transformed into a sequence of patches $X = \\{x^1, \\ldots, x^N\\}$ with each element $x^k \\in \\mathbb{R}^{P^2 \\times C}$ . Then, we sample a random set of patch index $M$ in uniform distribution, and split the image patches $X$ into masked patch set $X_m = \\{x^k | k \\in M\\}$ and visible patch set $X_v = \\{x^k | k \\notin M\\}$ . During training, the MAE encoder inputs $X_v$ to achieve the latent representations $H_v$ . Then, the MAE decoder attempts to reconstruct $X_m$ with the input of interpolating [mask] token embedding into the sequence of latent representations $H_v$ according to the index set $M$ , and outputs the reconstructed patches $\\tilde{X}_m$ . Finally, MAE optimizes the mean-squared error reconstruction loss on the masked patches as:", + "bbox": [ + 496, + 398, + 893, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m a e} (X, M, \\theta_ {m a e}) = \\sum_ {k \\in M} | | \\tilde {x} ^ {k} - x ^ {k} | | _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 729, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{mae}$ represents the parameters of MAE model.", + "bbox": [ + 500, + 775, + 857, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. When MAE meet GAN", + "text_level": 1, + "bbox": [ + 500, + 800, + 714, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the proposed GAN-MAE framework, as presented in Figure 2. Generally, our framework consists of two parts, a generator $G$ to reconstruct the masked image patches and a discriminator $D$ to predict the realness of image patches.", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24451", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Patch Generator. Identical to the standard MAE, the generator follows the encoder-decoder paradigm and is trained to perform masked-image reconstruction task. Given the partitioned image patches $X$ , the generator randomly masks some image patches and encodes the remaining visible patches $X_{v}$ into a sequence of contextualized vector representations $H_{v}$ , based on which the masked patches are reconstructed as $\\hat{X}_m$ . Please refer to Sec. 3.1 for more details. In general, the generation process can be formulated as:", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM \\sim \\operatorname {U n i f o r m} (1, N), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 253, + 468, + 268 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {v} = f _ {e} \\left(X _ {v}, M\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 272, + 468, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {X} _ {m} = f _ {d} \\left(H _ {v}, M\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 292, + 468, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the number of image patches. $f_{e}(\\cdot)$ and $f_{d}(\\cdot)$ denote the encoder and decoder in a conventional MAE.", + "bbox": [ + 76, + 321, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Patch Discriminator. For a patch index $k$ and corrupted image sequence $\\overline{X} = \\{X_v, \\tilde{X}_m\\}$ , the discriminator predicts whether the patch token $x^k$ is real or synthesized as binary classification task. Specifically, we create the corrupted image $\\overline{X}$ by maintaining the visible patches $X_v$ in raw image $X$ and replacing the masked patches with generator predicted result $\\tilde{X}_m$ . Note that $X_v$ and $\\tilde{X}_m$ are absolute complement of set to each other, i.e., $X_v \\cup \\tilde{X}_m = \\overline{X}$ and $X_v \\cap \\tilde{X}_m = \\emptyset$ . Formally, the task of discriminator model can be formulated as:", + "bbox": [ + 75, + 369, + 472, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD (\\bar {X}, k) = p _ {\\text {d i s c}} \\left(y ^ {k} | \\bar {X}, k\\right), \\text {f o r} k \\in [ 1, N ]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 531, + 468, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let the ground-truth classification label sequence $Y = \\{y^{1},\\ldots ,y^{N}\\}$ with each element $y_{k}\\in \\{0,1\\}$ , where 0 and 1 denote the corresponding image patch is reconstructed from generator or comes from the original image, respectively. The training objective of the discriminator can be formulated as:", + "bbox": [ + 76, + 560, + 472, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {d i s c} (\\bar {X}, \\theta_ {d i s c}) = \\sum_ {k = 1} ^ {N} - y ^ {k} \\log D (\\bar {X}, k) \\tag {6} \\\\ - \\left(1 - y ^ {k}\\right) \\log \\left(1 - D (\\overline {{X}}, k)\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 659, + 468, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Particularly, though the corrupted image is partially unreal, we suggest that this discriminative task can still be benefit to the learning of feature representation with the plausible corrupted image from a well-trained generator, as presented in experiments. Hitherto, due to the GAN-alike strategy for MAE task, we name our framework as GAN-MAE.", + "bbox": [ + 75, + 731, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training Scheme", + "text_level": 1, + "bbox": [ + 76, + 830, + 243, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We explore the training strategy of our proposed GANMAE in this section. Particularly, to improve synthesis result, we augment the L-2 reconstruction loss with a percep", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Adversarial training for GAN-MAE" + ], + "code_body": "Data: Training data $\\mathcal{D}_{train}$ total epoch number $N_{e}$ GAN-MAE model with generator parameters $\\theta_{mae}$ and discriminator parameters $\\theta_{disc}$ \n1 share weights between generator and discriminator backbones; \n2 while $n_e < N_e$ do \n3 for $x^{i}\\in \\mathcal{D}_{train}$ do \n4 $\\triangleright$ generator training; \n5 sample masking set $M^i$ and mask image $x^i$ . \n6 predict masked image patches $\\tilde{x}_m^i$ . \n7 compute loss $L_{gen}$ . \n8 loss backward for updating $\\theta_{mae}$ . \n9 dicriminator training; \n10 construct $\\overline{x}^i$ based on $x^i$ and $\\tilde{x}_m^i$ . \n11 comput loss $L_{disc}$ . \n12 loss backward for updating $\\theta_{disc}$ . \n13 end \n14 $n_e + = 1$ . \nend", + "bbox": [ + 506, + 111, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tual loss that aims to differentiate the real patches and reconstructed patches as:", + "bbox": [ + 498, + 434, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a d v} \\left(X, \\theta_ {m a e}\\right) = \\log D \\left(X _ {v}\\right) + \\log \\left(1 - D \\left(\\tilde {X} _ {m}\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 473, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Please note that $\\tilde{X}_m$ is the reconstructed patches by MAE. Therefore, the final objective for the generator comes to minimize the combined loss as:", + "bbox": [ + 498, + 500, + 890, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {g e n}} (X, \\theta_ {\\text {m a e}}) = L _ {\\text {m a e}} (X, \\theta_ {\\text {m a e}}) + \\gamma L _ {\\text {a d v}} (X, \\theta_ {\\text {m a e}}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 553, + 890, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we compute the adaptive weight $\\gamma$ according to:", + "bbox": [ + 500, + 578, + 861, + 593 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\frac {\\nabla \\left[ L _ {m a e} \\right]}{\\nabla \\left[ L _ {a d v} \\right] + \\delta}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 599, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\nabla[\\cdot]$ denotes the gradients of different loss function w.r.t. the parameters of last layer in network, and $\\delta = 1e - 6$ is used for numerical stability. Intuitively, the integration of $\\gamma$ adaptively balances the contributions of two loss functions to the gradients of parameters.", + "bbox": [ + 498, + 641, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the aforementioned strategy, the final training algorithm can be formulated in Algorithm 1. Specifically, at each epoch, we conduct the iteration in two steps: (i) train only the generator with $L_{gen}$ ; (ii) Train the discriminator with $L_{disc}$ . During training, we observe that shared weights of generator and discriminator backbones can usually stabilize the training procedure and bring extra-gain in down-stream tasks.", + "bbox": [ + 496, + 717, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Discussion", + "text_level": 1, + "bbox": [ + 500, + 847, + 617, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretically, the integration of discriminator can be considered as a high-level perceptual loss, which forces the", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24452", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generator to learn better feature representation for the plausible synthesis of masked patches. As the quality of synthesized patches improved, we claim that the introduction of corrupted images, serving as complementary information, provides plausible inner dependency between full image patches for representation learning. Furthermore, with the shared parameters of backbones in generator and discriminator, the GAN-alike framework can be considered as a type of multi-task learning, leading to even better result as presented in experiments.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 256, + 209, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Settings", + "text_level": 1, + "bbox": [ + 76, + 281, + 277, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the experiments, we pre-train our GAN-MAE model on the ImageNet-1k [16] and evaluate the performance in end-to-end fine-tuning (FT) pattern for the task of classification, semantic segmentation, object detection and instance segmentation. The evaluation metric of classification is the top-1 validation accuracy on $224 \\times 224$ cropped input images. The input image is partitioned into $14 \\times 14$ patches and each patch is of size $16 \\times 16$ . Following the setting of MAE, we only use the standard random cropping and horizontal flipping for data augmentation. To validate the effectiveness of GAN-MAE framework, the used ViT architecture and most hyper-parameters are exactly the same to [28, 60], i.e., ViT-S (12 transformer blocks with dimension 384), ViT-B (12 transformer blocks with dimension 768), and ViT-L (24 transformer blocks with dimension 1024). All version of ViT models are trained with 4096 batch size on 8 V-100 32GB GPUs. We adopt dynamic token masking with the masked positions decided on-the-fly. We use AdamW [38] optimizer and cosine schedule [51] with warm up for model training. The learning rate is annealed according to the cosine schedule. Unless stated otherwise, results are evaluated on the dev set. Please refer to appendix A for more training implementation and hyperparameter values for different backbone in details.", + "bbox": [ + 75, + 303, + 472, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Analysis of GAN-MAE", + "text_level": 1, + "bbox": [ + 76, + 676, + 290, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We analyze our GAN-MAE framework by proposing and evaluating several extensions to the model. Unless stated otherwise, all these experiments use the same model size as ViT-B and training dataset ImageNet-1K.", + "bbox": [ + 75, + 700, + 468, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameter Sharing of Backbone. In this work, we propose to improve the efficiency of the pre-training by sharing the parameters of the backbone vision transformer between the generator encoder and discriminator. One cause is that the generator and discriminator utilize the same network architecture, and all of the transformer weights can be tied. However, we can release the weight sharing and train the generator and discriminator independently. In between, we", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6a5a4aa56ca93c7647ea65ac3ef1fc57de67b2c84beaf94e1363ca59a667acc3.jpg", + "table_caption": [ + "Table 1. Effect of parameter sharing in GAN-MAE framework. Results demonstrate that shared parameters for backbone benefits both memory cost and performance improvements." + ], + "table_footnote": [], + "table_body": "
ModelsEpochMask ratioFT
Generator80075%83.9
Discriminator80075%84.2
Shared80075%84.3
Generator160075%84.4
Discriminator160075%84.4
Shared160075%84.6
", + "bbox": [ + 550, + 141, + 844, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can adopt both the parameters of generator and discriminator for downstream learning. The experimental results are shown in Table 1, where we employ the adversarial training with the same training epochs. We can see that, generally, the fine-tuning top-1 accuracy of shared weight outperforms the independent generator and discriminator conspicuously, in particular 0.4 point improvements for the generator when pre-training 800 epochs. We hypothesize that GAN-MAE framework benefits from both mask-then-reconstruct and pseudo-patch classification tasks, which can incorporate the visual semantic and consistency understanding.", + "bbox": [ + 496, + 294, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is also surprising that under the GAN process without weight sharing, the discriminator with learning of patch classification can lead to better performance than generator. We suspect that the fine-grained patch classification is more delicate while harder than image-level classification. Thus, we further tried to add an image-level contrastive objective. For this task, we input $50\\%$ of the input image unchanged rather than noising them with the generator. We then added a prediction head to the model that predicted if the entire input image was corrupted or not. However, the result didn't improve the final accuracy. In conclusion, we believe that the design of discriminative task, e.g., getting closer to downstream and be more difficult, is an important exploration direction in the future. Another interesting finding is that with the pre-training epochs increase, the independent generator performance boosts, we believe that it is caused by the adversarial training, where the discriminator becomes an effective guiding for generation.", + "bbox": [ + 496, + 460, + 892, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Schemes. We analyze the effect of proposed adversarial training scheme on GAN-MAE with shared backbones. Two training variants are considered as following:", + "bbox": [ + 498, + 753, + 890, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Two-stage Training. It is natural to consider to remove the discriminator guiding signal during generator training, which leads to the disentangled optimizing process. Specifically, at each epoch, we do the following steps: train the generator only with $L_{mae}$ and train the discriminator only with $L_{disc}$ . The difference with ad", + "bbox": [ + 517, + 810, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24453", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a940b1c4efb60eed18677476545a3f3aeb632b977b72ae2a6d6522f912157a6a.jpg", + "table_caption": [ + "Table 2. Effect of different training schemes." + ], + "table_footnote": [], + "table_body": "
ModelsEpochMask ratioFTGPU Time
Two-stage30075%82.094.3h
Combined30075%82.290.9h
Adversarial30075%82.2118.8h
Two-stage80075%84.0252.2h
Combined80075%84.1240.5h
Adversarial80075%84.3317.5h
", + "bbox": [ + 93, + 113, + 454, + 234 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb35fbef553714a39eedd9041de61f651a1d84275d9ffb74fc29acb8544674b4.jpg", + "table_caption": [ + "Table 3. Effect of different masking ratio in GAN-MAE framework with different pre-training epochs." + ], + "table_footnote": [], + "table_body": "
ModelsEpochMask ratioFT
MAE80075%83.4
GAN-MAE80070%84.3
GAN-MAE80075%84.3
GAN-MAE80080%84.0
MAE160075%83.6
GAN-MAE160070%84.5
GAN-MAE160075%84.6
GAN-MAE160080%84.4
", + "bbox": [ + 135, + 299, + 411, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "versarial training lies in that the training loss of the generator in the first step is changed.", + "bbox": [ + 107, + 481, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Combined Training. Instead of iterative in total dataset, we can jointly trains the generator and discriminator at each step. That is, for each image $X$ in the training dataset $\\mathcal{D}_{train}$ , we can directly minimize the combined loss as:", + "bbox": [ + 96, + 523, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g e n} \\left(\\theta_ {m a e}\\right) + \\lambda L _ {d i s c} \\left(\\theta_ {d i s c}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 612, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set $\\lambda$ , the weight for the discriminator objective in the loss to 2.0, as we searched for $\\lambda$ out of [1,2,5,10] in early experiments.", + "bbox": [ + 107, + 640, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that regardless of the form, the nature of sequential training, i.e., the corrupt image built from generator will be fed into discriminator, is not changed.", + "bbox": [ + 75, + 696, + 468, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The evaluation results for classification are listed in Table 2. GPU time means pre-training time (hours) on 8 V100 32GB GPUs environment. As we can see, the combined training shows a superior training time while adversarial training slows down as the training procedures of generator and discriminator are isolated. On the other hand, the performance of adversarial training is not better than combined training in the early stages; when pre-training epochs come to 800, the benefits of adversarial training appear.", + "bbox": [ + 75, + 742, + 470, + 878 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7cc48d3c06a3cf4be7e604e17c100f42a4b013a3bda596f313ce35ee27177317.jpg", + "table_caption": [ + "Table 4. Comparison of computation resource usage during self-supervised pre-training." + ], + "table_footnote": [], + "table_body": "
BackboneModelsFLOPsParams
ViT-BMAE9.4e9111.654M
ViT-BGAN-MAE2.6e10111.656M
ViT-LMAE2.0e10329.238M
ViT-LGAN-MAE8.0e10329.240M
", + "bbox": [ + 542, + 128, + 849, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Masking Ratio. Table 3 shows the influence of masking ratio in MAE-GAN under different pre-training epochs. The optimal ratio for MAE-GAN is identical to $75\\%$ , showing a obviously better classification performance compared with same masking ratio in previous work [28]. Meanwhile, we present several reconstructed images from MAE and GAN-MAE. Although the MIM model can infer missing patches as different yet plausible outputs when mask ratio comes to large, the generator of GAN-MAE can predict the images with more realistic and fine-grained details, e.g., the outline of a mountain peak in first case, which we believe is relative to the learning of useful representation.", + "bbox": [ + 496, + 242, + 890, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Computation Resource. In terms of pre-training cost, we conduct a computing resource comparison with the baseline MAE on different vision transformer backbones. The results are shown in Table 4. We first point out that GAN-MAE slows down the training process as the discriminator integration, e.g. 317.5h vs. 127.7h for ViT-B model in 800 epochs. Moreover, we also choose to measure computation usage in terms of floating point operations (FLOPs) as it is a metric agnostic to the particular hardware, low-level optimizations, etc. Note that an \"operation\" is a mathematical operation, not a machine instruction and thop package is used to compute FLOPs in practice. As expected, GAN-MAE employs $\\sim 3$ extra theoretical computation cost. Besides, benefits from the weight sharing of backbone, the GAN-MAE incorporates less than $1\\%$ addition parameter, which helps to reduce the running memory usage, and provides a convenience to set large batch size. Last but not least, we want to state that although our GAN-MAE method incorporates additional computation resources during training at each epoch, considering the superiority of reduction of epoch number and classification performance improvements, it is fully acceptable and exploration valuable.", + "bbox": [ + 496, + 444, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. ImageNet Classification Comparison", + "text_level": 1, + "bbox": [ + 498, + 786, + 818, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our GAN-MAE methods with previous state-of-the-art works on the ImageNet-1K classification task. Table 5 reports the top-1 validation accuracy for finetuning results. We can find that compared to the supervised models, trained from scratch, all of the self-supervised pre-training methods achieve significant improvement, sug", + "bbox": [ + 496, + 810, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24454", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4beaed6f5b05eb57a4cf78ad30018dfcf433231e01ff5836ca4a2607b588199c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 92, + 477, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ddcfbc8bfd3b604331fddb3c4f1a123e6120b15a9d44aa1479eaf1c844a3291.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 172, + 475, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0baaf630ac8e81e514ec56c7064ca6ba5cc70acc6661a703f7772ad9089199d5.jpg", + "image_caption": [ + "Figure 3. Qualitative analysis for patch reconstruction. Example results are from ImageNet validation set. For each tuple, we show the raw image, masked image, MAE reconstructed image, and our proposed GAN-MAE reconstructed image from left to right. We can see that the reconstructed images from GAN-MAE are significantly clearer than MAE, which we believe benefits the fine-grained visual semantic understanding." + ], + "image_footnote": [], + "bbox": [ + 88, + 256, + 477, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c9af92e7918f7eb148ab1ff90d6cf8cd8b5e0e4605d30763e0806b85e0505a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 93, + 879, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f912ca23a28c07aa3115c05807a0fde46ca7000513169496914f65f13d17c0ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 172, + 879, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f31ed1cbd7180b371977d634615cb036bf4fe7423bdba14aadc5bf2ae492488c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 256, + 879, + 335 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ebd504e45770b877482d4bc6dea87ca4db3559e882231c18a96d5913873d6219.jpg", + "table_caption": [ + "Table 5. End-to-end fine-tuning on ImageNet-1K. We report the fine-tuning top-1 accuracy for classification in different vision transformer architectures and results show that GAN-MAE outperforms previous self-supervised methods." + ], + "table_footnote": [], + "table_body": "
ModelPre-train dataPre-train epochsViT-SViT-BViT-L
Supervised [59]IN1K w/ labels30079.781.882.6
DINO [9]IN1K80081.582.8-
MoCo v3 [14]IN1K30081.483.284.1
BEiT [3]IN1K+DALLE80081.783.285.2
MSN [1]IN1K600-83.4-
iBOT [77]IN1K80082.384.084.8
BootMAE [20]IN1K800-84.285.9
MAE [28]IN1K800-83.485.4
MAE [28]IN1K1600-83.685.9
GAN-MAEIN1K30082.284.085.6
GAN-MAEIN1K80082.484.386.1
", + "bbox": [ + 199, + 460, + 772, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gesting the effectiveness of pre-training. We further compare our GAN-MAE framework with prior popular self-supervised pre-training models. We can see that the proposed GAN-MAE achieves the best fine-tuning performance either based on the ViT-B or based on the ViT-L architectures. For example, compared with the recent work MAE [28], our GAN-MAE in ViT-L network achieves $86.1\\%$ with 0.7 point improvement when pre-trained on 800 epochs. More encouragingly, for all ViT backbone sizes, GAN-MAE mostly outperforms the previous self-supervised methods. These results suggest that the incorporation of a discriminator scheme could have consistently benefits for various scale ViT models.", + "bbox": [ + 75, + 688, + 470, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we present a comprehensive comparison", + "bbox": [ + 94, + 885, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with MAE under different pre-training epochs for the ViT-B model. We plot the results in Figure 1. We can see that our GAN-MAE approach consistently performs better than MAE. It is worth mentioning that the proposed GAN-MAE at 300 epochs achieves $84.0\\%$ accuracy, which is already better than MAE pre-trained at 1600 epochs. This demonstrates that our approach is more efficient to achieve comparable performance. What's more, no additional speed and parameter cost during inference. Besides, we can confirm that, similar to other MIM-based self-supervised training, the accuracy of GAN-MAE also improves steadily as the pre-training steps increase. Particularly, our GAN-MAE on ViT-B achieves $84.6\\%$ top-1 accuracy at 1600 epochs, which is almost $1\\%$ higher than that of MAE.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24455", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c401f223325a792518d8afe63a7c24ce46de0efb43e57e1281c0b6770cf20856.jpg", + "table_caption": [ + "Table 6. Robustness Evaluation on the four ImageNet-variants: ImageNet-C, ImageNet-A, ImageNet-R, and ImageNet-Sketch. Except for ImageNet-C which is measured in terms of mean Corruption Error (mCE), top-1 accuracy is used as the remaining evaluation metric. For simplicity, we denoted IN-C, IN-A, IN-R, In-Sketch correspondingly." + ], + "table_footnote": [], + "table_body": "
ModelIN-C (mCE ↓)IN-A (top-1 ↑)IN-R (top-1 ↑)IN-Sketch (top-1 ↑)
Supervised [53]42.535.848.736.0
MAE [28]51.735.948.334.5
GAN-MAE49.536.849.635.9
", + "bbox": [ + 166, + 141, + 805, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1cde421a53494c8b04f017b2598223e3e883427b75bb77b74d864bc7c614a02b.jpg", + "table_caption": [ + "Table 7. Semantic segmentation comparison on the ADE20K dataset for mIoU (%) metric with the ViT-B backbone." + ], + "table_footnote": [], + "table_body": "
ModelsPre-train dataEpochsmIoU
Supervised [28]IN1K w/ labels30047.4
MoCo v3 [14]IN1K30047.3
BEiT [3]IN1K+DALLE80047.1
MAE [28]IN1K80047.6
MAE [28]IN1K160048.1
BootMAE [20]IN1K80049.1
GAN-MAEIN1K80049.5
", + "bbox": [ + 99, + 284, + 447, + 414 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4d7aef036b2d3846b58fabad473fe007dcf92223c62756a3823f047577c3530d.jpg", + "table_caption": [ + "Table 8. COCO object detection and segmentation using Mask R-CNN framework with ViT-B backbone." + ], + "table_footnote": [], + "table_body": "
ModelsPre-train dataAP-boxAP-mask
Supervised [28]IN1K w/ labels44.139.8
MoCo v3 [14]IN1K44.940.4
BEiT [3]IN1K+DALLE46.341.1
MSN [1]IN1K46.641.5
iBOT [77]IN1K47.342.2
MAE [28]IN1K47.242.0
BootMAE [20]IN1K48.543.4
GAN-MAEIN1K49.043.8
", + "bbox": [ + 86, + 486, + 460, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Downstream Tasks", + "text_level": 1, + "bbox": [ + 76, + 676, + 259, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic Segmentation. We compare our GAN-MAE with supervised as well as state-of-the-art self-supervised models on the widely used dataset ADE20K [76] for semantic segmentation. Specifically, we use the UperNet framework [68] in the experiments. We train Upernet for 160K iterations with batch size set as 64 and report the results in Table 7. The evaluation metric is mean Intersection of Union (mIoU) averaged over all semantic categories and the single-scale test results are reported. Importantly, we can see that the proposed GAN-MAE gets superior performance than all the other baselines in the same configuration, further validating the effectiveness of adversarial training with a reconstructed patch discriminator.", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Detection and Segmentation. We also perform object detection and instance segmentation, compared with other popular self-supervised methods and the supervised model, on the COCO dataset [47]. In practice, we choose the Mask R-CNN [30] framework and adopt FPNs [46] to scale the feature map into different sizes as introduced in [44]. The performance is tested on the COCO validation set, following the previous work [18]. The results are listed in Table 8 in terms of box AP metric for object detection and mask AP metric for instance segmentation. Importantly, we can observe that our GAN-MAE model achieves $49.0\\%$ for object detection and $43.8\\%$ for segmentation, surpassing the previous state-of-the-art BootMAE by $0.5\\%$ and $0.4\\%$ point, respectively.", + "bbox": [ + 496, + 247, + 893, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Classification Robutness. Similar to [28], we further evaluate the robustness of classification performance on the four ImageNet variants, i.e., ImageNet-C [33], ImageNet-A [34], ImageNet-R [32], and ImageNet-Sketch [65], which are common benchmarks to evaluate robustness for perturbations. Table 6 demonstrates the robustness comparison with GAN-MAE and MAE using the ViT-B backbone, as well as previous supervised SoTA models. The results illustrate that GAN-MAE outperforms the MAE baseline consistently on all robustness datasets, indicating that the promising of adversarial training in representation learning.", + "bbox": [ + 496, + 481, + 893, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 662, + 619, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a new self-supervised GAN-alike framework for visual representation learning, where a generator is used to predict masked image patches according to the visible patches and a discriminator is employed to predict whether the patch is from raw image or generated by generator. The key idea is adversarial training a shared vision Transformer to distinguish the input patches from high-quality negative samples, which we believe is beneficial for the understanding of visual conception. It works well while incorporating no much addition parameter. More encouragingly, compared to standard masked image modeling, our GAN-MAE is more compute-efficient, as fewer pre-training epochs result in a better performance on downstream tasks.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24456", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. arXiv preprint arXiv:2204.07141, 2022. 1, 2, 7, 8", + "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In Proc. ICML, pages 1298-1312. PMLR, 2022. 3", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In Proc. ICLR, 2021. 1, 2, 7, 8", + "[4] Sefi Bell-Kligler, Assaf Shocher, and Michal Irani. Blind super-resolution kernel estimation using an internal-gan. Proc. NIPS, 32, 2019. 3", + "[5] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 2", + "[6] Yoshua Bengio et al. Learning deep architectures for ai. Foundations and trends® in Machine Learning, 2(1):1-127, 2009. 2", + "[7] Yoshua Bengio, Pascal Lamblin, Dan Popovici, and Hugo Larochelle. Greedy layer-wise training of deep networks. Proc. NIPS, 19, 2006. 2", + "[8] David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 3", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proc. IEEE CVPR, pages 9650-9660, 2021. 7", + "[10] Liquun Chen, Shuyang Dai, Chenyang Tao, Haichao Zhang, Zhe Gan, Dinghan Shen, Yizhe Zhang, Guoyin Wang, Ruiyi Zhang, and Lawrence Carin. Adversarial text generation via feature-mover's distance. Proc. NIPS, 31, 2018. 3", + "[11] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proc. ICML, pages 1691-1703. PMLR, 2020. 2", + "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proc. ICML, pages 1597-1607. PMLR, 2020. 1", + "[13] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 3", + "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proc. IEEE ICCV, pages 9640-9649, 2021. 1, 7, 8", + "[15] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE CVPR, pages 8789-8797, 2018. 3" + ], + "bbox": [ + 76, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE CVPR, pages 248–255, 2009. 5", + "[17] Jeff Donahue, Philipp Krahenbuhl, and Trevor Darrell. Adversarial feature learning. arXiv preprint arXiv:1605.09782, 2016.3", + "[18] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proc. IEEE CVPR, pages 12124-12134, 2022. 8", + "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. 1, 3", + "[20] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022. 7, 8, 12", + "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. ICLR, 2020. 1, 2", + "[22] Bo Du, Wei Xiong, Jia Wu, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Stacked convolutional denoising auto-encoders for feature representation. IEEE transactions on cybernetics, 47(4):1017-1027, 2016. 2", + "[23] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 3", + "[24] William Fedus, Ian Goodfellow, and Andrew M Dai. Maskgan: Better text generation via filling in the .. In Proc. ICLR, 2018. 3", + "[25] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 3", + "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3", + "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Proc. NIPS, 33:21271-21284, 2020. 1", + "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE CVPR, pages 16000-16009, 2022. 1, 2, 3, 5, 6, 7, 8, 12" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24457", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE CVPR, pages 9729-9738, 2020. 1", + "[30] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. IEEE ICCV, pages 2961-2969, 2017. 8", + "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE CVPR, pages 770-778, 2016. 1", + "[32] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proc. IEEE CVPR, pages 8340-8349, 2021. 8", + "[33] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 8", + "[34] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proc. IEEE CVPR, pages 15262-15271, 2021. 8", + "[35] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 2", + "[36] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE CVPR, pages 1125-1134, 2017. 3", + "[37] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL, pages 4171–4186, 2019. 2", + "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[39] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In Proc. ECCV, pages 577-593. Springer, 2016. 2", + "[40] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proc. IEEE CVPR, pages 4681-4690, 2017. 3", + "[41] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In Proc. IEEE CVPR, pages 7287-7296, 2022. 2", + "[42] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proc. AAAI, volume 34, pages 11336-11344, 2020. 3", + "[43] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 8", + "[45] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In Proc. IEEE CVPR, pages 9392-9400, 2021. 3", + "[46] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proc. IEEE CVPR, pages 2117-2125, 2017. 8", + "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755. Springer, 2014. 8", + "[48] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, and Jing Liao. Pd-gan: Probabilistic diverse gan for image inpainting. In Proc. IEEE CVPR, pages 9371-9381, 2021. 3", + "[49] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2", + "[50] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. IEEE ICCV, pages 10012-10022, 2021. 2", + "[51] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5", + "[52] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In Proc. IEEE CVPR, pages 2794-2802, 2017. 3", + "[53] Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. In Proc. IEEE CVPR, pages 12042-12051, 2022. 8", + "[54] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proc. ECCV, pages 69-84. Springer, 2016. 2", + "[55] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proc. IEEE CVPR, pages 2536-2544, 2016. 2", + "[56] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 3", + "[57] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2", + "[58] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, pages 6105-6114. PMLR, 2019. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24458", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proc. ICML, pages 10347-10357. PMLR, 2021. 7", + "[60] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 5, 12", + "[61] Michael Tschannen, Olivier Bachem, and Mario Lucic. Recent advances in autoencoder-based representation learning. arXiv preprint arXiv:1812.05069, 2018. 2", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NIPS, pages 5998-6008, 2017. 1", + "[63] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proc. ICML, pages 1096-1103, 2008. 2", + "[64] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 2", + "[65] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Proc. NIPS, 32, 2019. 8", + "[66] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proc. IEEE CVPR, pages 14668-14678, 2022. 3", + "[67] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proc. IEEE CVPR, pages 2256-2265, 2021. 3", + "[68] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proc. ECCV, pages 418-434, 2018. 8", + "[69] Junyuan Xie, Linli Xu, and Enhong Chen. Image denoising and inpainting with deep neural networks. Proc. NIPS, 25, 2012. 2", + "[70] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proc. IEEE CVPR, pages 9653-9663, 2022. 1, 2, 3", + "[71] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Proc. NIPS, 32, 2019. 2", + "[72] Richard Zhang, Phillip Isola, and Alexei A Efros. **Colorful image colorization. In Proc. ECCV, pages 649-666. Springer, 2016. 2**", + "[73] Yizhe Zhang, Zhe Gan, Kai Fan, Zhi Chen, Ricardo Henao, Dinghan Shen, and Lawrence Carin. Adversarial feature matching for text generation. In Proc. ICML, pages 4006-4015. PMLR, 2017. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial networks. In Proc. ICLR, 2017. 3", + "[75] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proc. IEEE CVPR, pages 18697–18709, 2022. 3", + "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proc. IEEE CVPR, pages 633-641, 2017. 8", + "[77] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24459", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_model.json b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e47ca29c3076f5ab5b096ceb36f6a788ac398d89 --- /dev/null +++ b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.181, + 0.756, + 0.272 + ], + "angle": 0, + "content": "Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang* Xiaoming Wei, Xiaolin Wei Meituan {feizhengcong, fanmingyuan, zhuli09, huangjunshi}@meituan.com {weixiaoming, weixiaolin02}@meituan.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.304, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.715 + ], + "angle": 0, + "content": "Masked Auto-Encoder (MAE) pretraining methods randomly mask image patches and then train a vision Transformer to reconstruct the original pixels based on the unmasked patches. While they demonstrate impressive performance for downstream vision tasks, it generally requires a large amount of training resource. In this paper, we introduce a novel Generative Adversarial Networks alike framework, referred to as GAN-MAE, where a generator is used to generate the masked patches according to the remaining visible patches, and a discriminator is employed to predict whether the patch is synthesized by the generator. We believe this capacity of distinguishing whether the image patch is predicted or original is benefit to representation learning. Another key point lies in that the parameters of the vision Transformer backbone in the generator and discriminator are shared. Extensive experiments demonstrate that adversarial training of GAN-MAE framework is more efficient and accordingly outperforms the standard MAE given the same model size, training data, and computation resource. The gains are substantially robust for different model sizes and datasets, in particular, a ViT-B model trained with GAN-MAE for 200 epochs outperforms the MAE with 1600 epochs on fine-tuning top-1 accuracy of ImageNet-1k with much less FLOPs. Besides, our approach also works well at transferring downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.744, + 0.21, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.77, + 0.471, + 0.877 + ], + "angle": 0, + "content": "In recent years, Transformer [62] has become the de facto standard architecture in computer vision, and has surpassed state-of-the-art Convolutional Neural Network (CNN) [31, 58] feature extractors in vision tasks through models such as the Vision Transformer [21]. Meanwhile, self-supervised learning (SSL) algorithms [12, 14, 27, 29] aims to learn transferable representation from unlabeled" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.326, + 0.858, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.536, + 0.895, + 0.594 + ], + "angle": 0, + "content": "Figure 1. Performance comparison in different pre-training epochs for ImageNet-1K Fine-tuning top-1 accuracy. Compared to MAE trained for 1600 epochs, GAN-MAE achieves comparable accuracy with much less training time at 200 epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.718 + ], + "angle": 0, + "content": "data by performing instance-level pretext tasks, and has been a long-standing target in the vision community. Particularly, masked image modeling (MIM) in SSL for vision transformers has shown remarkably impressive downstream performance in a wide variety of computer vision tasks [3, 28], attracting increasing attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "MIM is a simple pretext task that first randomly masks some patches of an image, and then predicts the contents of the masked patches according to the remaining, using various reconstruction targets, e.g., visual tokens [3, 19], semantic features [1, 77] and raw pixels [28, 70]. Essentially, it learns the transferable representation by modeling the image structure itself as content prediction. While more effective than conventional pre-training, masked autoencoder modeling approaches still exist some issues: (i) reconstruction optimization with MSE loss leads to blurrier output images than the raw input, it would be better to use a more perceptual loss over pixels to guide the fine-grained seman" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.245, + 0.901 + ], + "angle": 0, + "content": "*The corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24449" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "tic understanding and representation learning, leading to more plausible synthesized patches; (ii) inner dependency between masked patches is lacked [71], i.e., generation of masked image patches may lack the surrounding information. This situation becomes more serious when the image patch masking ratio is large. We alleviate this problem by introducing confident synthesized patches as complementary information during training; (iii) mask-reconstruction methods incur a substantial computation cost because the network only learns from part of the visible patches and misses the information of masked patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.61 + ], + "angle": 0, + "content": "In this paper, we propose a Generative Adversarial Networks-based pre-training framework, referred to as GAN-MAE, which contains two components: a generator model learns to reconstruct the masked patches according to visible patches in the encoder-decoder architecture and a discriminator model learns to distinguish real image patches from plausible but synthesized remains. Generally, given an image from training dataset, our method first randomly masks parts of patches and reconstructs them using the rest visible patches with a generator, which serves as a standard MAE model. Then we build the corrupt image as the combination of visible and synthesis patches, which is then fed into the discriminator to predict whether each patch is from raw image or synthesized results. In this manner, the discriminator provides a valid guiding for more delicate image patch modeling. Then, with the development of generator capacity, a key advantage of discriminative task is that it integrates the synthesized patches into corrupt images as complementary information, which fills the missing inner relationship between patches during pre-training. Moreover, we shared the parameters of vision transformer backbone in the generator and discriminator to promote memory reduction, training efficiency, as well as performance enhancement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our experiments follow the same architecture, settings, and pre-training recipe as MAE [28], and we find that the simple incorporation of a discriminator consistently outperforms MAE in variant models, e.g., ViT-S, ViT-B, and ViT-L, when fine-tuning for top-1 accuracy of ImageNet classification. We also conduct extensive ablation studies to validate the effectiveness of our core designs in backbone parameter sharing and advarial training. As pre-training with more epochs usually results in a better downstream performance, we argue that an important consideration for pre-training methods should be computation efficiency as well as absolute downstream performance. From this viewpoint, we also demonstrate that discrimination of pseudo-image patches forces GAN-MAE to train more efficiently than standard MAE. We further provide a comprehensive comparison with MAE in various epochs and various models and show our framework achieves consistently better performance. In particular, as presented in Figure 1, for the ViT-B model structure, our GAN-MAE achieves comparable" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "ble classification performance with only 200 pre-training epochs vs. standard MAE 1600 pre-training epochs. Furthermore, the GAN-MAE achieves 0.7 points improvement when pre-training 1600 epochs. Finally, we summarize our contribution as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.179, + 0.892, + 0.271 + ], + "angle": 0, + "content": "- We propose a new and effective GAN-alike framework for visual representation self-supervised learning, which to our best knowledge is the first trial of integrating GAN idea into MAE framework. As a generic approach, we suggest that this framework can be easily applied on many other MIM-based tasks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.283, + 0.892, + 0.358 + ], + "angle": 0, + "content": "- We introduce two core designs: shared weight for the main backbones of generator and discriminator, and an adversarial training process, both of which cost fewer amounts of computing resources while obtaining appreciable performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.372, + 0.892, + 0.432 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that compared with the original MAE, our method is more compute-efficient and results in better transfer representation learning on downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.179, + 0.892, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.651, + 0.476 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.895, + 0.683 + ], + "angle": 0, + "content": "Autoencoding. Autoencoder [6, 7, 35] is an unsupervised learning technique for neural networks that learns efficient representations by training the network to ignore signal noise. It includes an encoder that maps the original data to a low-dimensional latent embedding and a decoder that recovers the data from the latent embedding, with the goal of learning a compressed knowledge representation. The denoising autoencoder [63, 64] learns to reconstruct clean data points from a noisy version. Numerous efforts have been devoted for image denoising, such as masking pixels [11, 55, 64], inpainting [69], removing color channels [39, 72], and shuffling image patches [22, 54]. For a broader overview of denoising autoencoder, we refer to [5, 61]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Masked Image Modeling. Masked language modeling [37,49,57], which generalizes well on language understanding and generation tasks, is the domain self-supervised approach in the field of NLP. Similarly, vision transformer [21, 41, 50] based masked image modeling (MIM) approaches [1,3,28,70,77] for computer vision tasks have also been developed. Generally, these MIM approaches first apply a mask to patches of an image, and then the masked patches are predicted given the visible patches. Feature representation learned through such within-image context prediction demonstrate strong transfer performance in downstream tasks. Recently, lots of works exploring MIM have been concurrently developed from different perspectives." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24450" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.895, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Figure 2. Overview of GAN-MAE framework, where a generator is used to predict the masked patches and a discriminator is employed to classify whether the patches are selected from the raw image or synthesized results. In particular, both of MAE encoder and discriminator are based on the vision Transformer backbone and share parameters for memory reduction and training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.471, + 0.429 + ], + "angle": 0, + "content": "The works include framework design [13, 23, 28, 70], prediction targets [2, 19, 66, 77], and integration with vision-language representation learning [42, 43, 75]. Our work belongs to the first group and introduces a novel GAN framework that discriminates the reconstructed image patches in a standard MAE to learn deeper semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.473, + 0.768 + ], + "angle": 0, + "content": "Generative Adversarial Networks. GANs [26] are effective at generating high-quality synthetic data. Usually, the generator generates an image, and the discriminator determines whether the input image is a real image or a generated image. Subsequently, many improvements based on the original GAN focused on speeding up the training of the network and improving the quality of the generated images [8, 52, 74]. These improvements also help GAN achieving a wider range of applications [45, 48, 67]. Methods based on GAN are also widely used in image-to-image translation [36], super-resolution [4, 40], style transfer [15], text generation [10, 24, 73], and representation learning [17, 25], to name a few. Particularly, [56] brings some designs of CNN architecture to stabilize the training of GAN framework, after that the discriminator can be directly used as feature extractor in downstream tasks. In contrast, our method proposes to integrate the GAN as assistant for the MIM task, which focuses on the study of better pattern for masking based self-supervised learning. Besides, our strategy of shared parameters provides a unified backbone for better vision representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.188, + 0.801 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this section, we introduce the GAN-MAE framework in details. At first, we briefly review the conventional masked autoencoder model in Sec. 3.1, and then describe the proposed generator-discriminator pre-training framework in Sec. 3.2. The architecture of our framework is presented in Figure 2. Finally, we suggest an adversarial" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.336, + 0.892, + 0.368 + ], + "angle": 0, + "content": "training processes under the proposed method and discuss our framework in Sec. 3.3 and Sec. 3.4, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.377, + 0.642, + 0.392 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.72 + ], + "angle": 0, + "content": "Masked Autoencoder (MAE) [28] is a self-supervised approach with a vision transformer encoder and a small transformer decoder, which randomly masks a large portion of input patches, and then reconstructs the masked patches according to the visible patches. Specifically, provided with an image \\(X \\in \\mathbb{R}^{C \\times H \\times W}\\), where \\(C\\), \\(H\\) and \\(W\\) are the channel number, image height and image width respectively, MAE partitions \\(X\\) into \\(N = \\frac{H \\times W}{P^2}\\) non-overlapping patches with patch size \\(P\\). In this way, the image is transformed into a sequence of patches \\(X = \\{x^1, \\ldots, x^N\\}\\) with each element \\(x^k \\in \\mathbb{R}^{P^2 \\times C}\\). Then, we sample a random set of patch index \\(M\\) in uniform distribution, and split the image patches \\(X\\) into masked patch set \\(X_m = \\{x^k | k \\in M\\}\\) and visible patch set \\(X_v = \\{x^k | k \\notin M\\}\\). During training, the MAE encoder inputs \\(X_v\\) to achieve the latent representations \\(H_v\\). Then, the MAE decoder attempts to reconstruct \\(X_m\\) with the input of interpolating [mask] token embedding into the sequence of latent representations \\(H_v\\) according to the index set \\(M\\), and outputs the reconstructed patches \\(\\tilde{X}_m\\). Finally, MAE optimizes the mean-squared error reconstruction loss on the masked patches as:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.731, + 0.892, + 0.765 + ], + "angle": 0, + "content": "\\[\nL _ {m a e} (X, M, \\theta_ {m a e}) = \\sum_ {k \\in M} | | \\tilde {x} ^ {k} - x ^ {k} | | _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.776, + 0.859, + 0.792 + ], + "angle": 0, + "content": "where \\(\\theta_{mae}\\) represents the parameters of MAE model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.715, + 0.816 + ], + "angle": 0, + "content": "3.2. When MAE meet GAN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In this section, we describe the proposed GAN-MAE framework, as presented in Figure 2. Generally, our framework consists of two parts, a generator \\(G\\) to reconstruct the masked image patches and a discriminator \\(D\\) to predict the realness of image patches." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "24451" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.242 + ], + "angle": 0, + "content": "Image Patch Generator. Identical to the standard MAE, the generator follows the encoder-decoder paradigm and is trained to perform masked-image reconstruction task. Given the partitioned image patches \\( X \\), the generator randomly masks some image patches and encodes the remaining visible patches \\( X_{v} \\) into a sequence of contextualized vector representations \\( H_{v} \\), based on which the masked patches are reconstructed as \\( \\hat{X}_m \\). Please refer to Sec. 3.1 for more details. In general, the generation process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.255, + 0.47, + 0.27 + ], + "angle": 0, + "content": "\\[\nM \\sim \\operatorname {U n i f o r m} (1, N), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.273, + 0.469, + 0.289 + ], + "angle": 0, + "content": "\\[\nH _ {v} = f _ {e} \\left(X _ {v}, M\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.293, + 0.469, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\tilde {X} _ {m} = f _ {d} \\left(H _ {v}, M\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.322, + 0.469, + 0.352 + ], + "angle": 0, + "content": "where \\(N\\) is the number of image patches. \\(f_{e}(\\cdot)\\) and \\(f_{d}(\\cdot)\\) denote the encoder and decoder in a conventional MAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.473, + 0.522 + ], + "angle": 0, + "content": "Image Patch Discriminator. For a patch index \\( k \\) and corrupted image sequence \\( \\overline{X} = \\{X_v, \\tilde{X}_m\\} \\), the discriminator predicts whether the patch token \\( x^k \\) is real or synthesized as binary classification task. Specifically, we create the corrupted image \\( \\overline{X} \\) by maintaining the visible patches \\( X_v \\) in raw image \\( X \\) and replacing the masked patches with generator predicted result \\( \\tilde{X}_m \\). Note that \\( X_v \\) and \\( \\tilde{X}_m \\) are absolute complement of set to each other, i.e., \\( X_v \\cup \\tilde{X}_m = \\overline{X} \\) and \\( X_v \\cap \\tilde{X}_m = \\emptyset \\). Formally, the task of discriminator model can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.532, + 0.47, + 0.55 + ], + "angle": 0, + "content": "\\[\nD (\\bar {X}, k) = p _ {\\text {d i s c}} \\left(y ^ {k} | \\bar {X}, k\\right), \\text {f o r} k \\in [ 1, N ]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.561, + 0.473, + 0.651 + ], + "angle": 0, + "content": "Let the ground-truth classification label sequence \\( Y = \\{y^{1},\\ldots ,y^{N}\\} \\) with each element \\( y_{k}\\in \\{0,1\\} \\), where 0 and 1 denote the corresponding image patch is reconstructed from generator or comes from the original image, respectively. The training objective of the discriminator can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.66, + 0.469, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {d i s c} (\\bar {X}, \\theta_ {d i s c}) = \\sum_ {k = 1} ^ {N} - y ^ {k} \\log D (\\bar {X}, k) \\tag {6} \\\\ - \\left(1 - y ^ {k}\\right) \\log \\left(1 - D (\\overline {{X}}, k)\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Particularly, though the corrupted image is partially unreal, we suggest that this discriminative task can still be benefit to the learning of feature representation with the plausible corrupted image from a well-trained generator, as presented in experiments. Hitherto, due to the GAN-alike strategy for MAE task, we name our framework as GAN-MAE." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.245, + 0.849 + ], + "angle": 0, + "content": "3.3. Training Scheme" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We explore the training strategy of our proposed GANMAE in this section. Particularly, to improve synthesis result, we augment the L-2 reconstruction loss with a percep" + }, + { + "type": "code_caption", + "bbox": [ + 0.511, + 0.096, + 0.822, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Adversarial training for GAN-MAE" + }, + { + "type": "algorithm", + "bbox": [ + 0.508, + 0.112, + 0.892, + 0.405 + ], + "angle": 0, + "content": "Data: Training data \\(\\mathcal{D}_{train}\\) total epoch number \\(N_{e}\\) GAN-MAE model with generator parameters \\(\\theta_{mae}\\) and discriminator parameters \\(\\theta_{disc}\\) \n1 share weights between generator and discriminator backbones; \n2 while \\(n_e < N_e\\) do \n3 for \\(x^{i}\\in \\mathcal{D}_{train}\\) do \n4 \\(\\triangleright\\) generator training; \n5 sample masking set \\(M^i\\) and mask image \\(x^i\\) . \n6 predict masked image patches \\(\\tilde{x}_m^i\\) . \n7 compute loss \\(L_{gen}\\) . \n8 loss backward for updating \\(\\theta_{mae}\\) . \n9 dicriminator training; \n10 construct \\(\\overline{x}^i\\) based on \\(x^i\\) and \\(\\tilde{x}_m^i\\) . \n11 comput loss \\(L_{disc}\\) . \n12 loss backward for updating \\(\\theta_{disc}\\) . \n13 end \n14 \\(n_e + = 1\\) . \nend" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.435, + 0.892, + 0.466 + ], + "angle": 0, + "content": "tual loss that aims to differentiate the real patches and reconstructed patches as:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.474, + 0.892, + 0.492 + ], + "angle": 0, + "content": "\\[\nL _ {a d v} \\left(X, \\theta_ {m a e}\\right) = \\log D \\left(X _ {v}\\right) + \\log \\left(1 - D \\left(\\tilde {X} _ {m}\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.501, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Please note that \\(\\tilde{X}_m\\) is the reconstructed patches by MAE. Therefore, the final objective for the generator comes to minimize the combined loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.554, + 0.892, + 0.571 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {g e n}} (X, \\theta_ {\\text {m a e}}) = L _ {\\text {m a e}} (X, \\theta_ {\\text {m a e}}) + \\gamma L _ {\\text {a d v}} (X, \\theta_ {\\text {m a e}}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.579, + 0.862, + 0.594 + ], + "angle": 0, + "content": "where we compute the adaptive weight \\(\\gamma\\) according to:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.601, + 0.891, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\frac {\\nabla \\left[ L _ {m a e} \\right]}{\\nabla \\left[ L _ {a d v} \\right] + \\delta}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.892, + 0.718 + ], + "angle": 0, + "content": "\\(\\nabla[\\cdot]\\) denotes the gradients of different loss function w.r.t. the parameters of last layer in network, and \\(\\delta = 1e - 6\\) is used for numerical stability. Intuitively, the integration of \\(\\gamma\\) adaptively balances the contributions of two loss functions to the gradients of parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Based on the aforementioned strategy, the final training algorithm can be formulated in Algorithm 1. Specifically, at each epoch, we conduct the iteration in two steps: (i) train only the generator with \\( L_{gen} \\); (ii) Train the discriminator with \\( L_{disc} \\). During training, we observe that shared weights of generator and discriminator backbones can usually stabilize the training procedure and bring extra-gain in down-stream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.618, + 0.862 + ], + "angle": 0, + "content": "3.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Theoretically, the integration of discriminator can be considered as a high-level perceptual loss, which forces the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24452" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "generator to learn better feature representation for the plausible synthesis of masked patches. As the quality of synthesized patches improved, we claim that the introduction of corrupted images, serving as complementary information, provides plausible inner dependency between full image patches for representation learning. Furthermore, with the shared parameters of backbones in generator and discriminator, the GAN-alike framework can be considered as a type of multi-task learning, leading to even better result as presented in experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.257, + 0.21, + 0.274 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.282, + 0.279, + 0.298 + ], + "angle": 0, + "content": "4.1. Datasets and Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.473, + 0.668 + ], + "angle": 0, + "content": "In the experiments, we pre-train our GAN-MAE model on the ImageNet-1k [16] and evaluate the performance in end-to-end fine-tuning (FT) pattern for the task of classification, semantic segmentation, object detection and instance segmentation. The evaluation metric of classification is the top-1 validation accuracy on \\(224 \\times 224\\) cropped input images. The input image is partitioned into \\(14 \\times 14\\) patches and each patch is of size \\(16 \\times 16\\). Following the setting of MAE, we only use the standard random cropping and horizontal flipping for data augmentation. To validate the effectiveness of GAN-MAE framework, the used ViT architecture and most hyper-parameters are exactly the same to [28, 60], i.e., ViT-S (12 transformer blocks with dimension 384), ViT-B (12 transformer blocks with dimension 768), and ViT-L (24 transformer blocks with dimension 1024). All version of ViT models are trained with 4096 batch size on 8 V-100 32GB GPUs. We adopt dynamic token masking with the masked positions decided on-the-fly. We use AdamW [38] optimizer and cosine schedule [51] with warm up for model training. The learning rate is annealed according to the cosine schedule. Unless stated otherwise, results are evaluated on the dev set. Please refer to appendix A for more training implementation and hyperparameter values for different backbone in details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.291, + 0.693 + ], + "angle": 0, + "content": "4.2. Analysis of GAN-MAE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.762 + ], + "angle": 0, + "content": "We analyze our GAN-MAE framework by proposing and evaluating several extensions to the model. Unless stated otherwise, all these experiments use the same model size as ViT-B and training dataset ImageNet-1K." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Parameter Sharing of Backbone. In this work, we propose to improve the efficiency of the pre-training by sharing the parameters of the backbone vision transformer between the generator encoder and discriminator. One cause is that the generator and discriminator utilize the same network architecture, and all of the transformer weights can be tied. However, we can release the weight sharing and train the generator and discriminator independently. In between, we" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Table 1. Effect of parameter sharing in GAN-MAE framework. Results demonstrate that shared parameters for backbone benefits both memory cost and performance improvements." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.142, + 0.846, + 0.257 + ], + "angle": 0, + "content": "
ModelsEpochMask ratioFT
Generator80075%83.9
Discriminator80075%84.2
Shared80075%84.3
Generator160075%84.4
Discriminator160075%84.4
Shared160075%84.6
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.461 + ], + "angle": 0, + "content": "can adopt both the parameters of generator and discriminator for downstream learning. The experimental results are shown in Table 1, where we employ the adversarial training with the same training epochs. We can see that, generally, the fine-tuning top-1 accuracy of shared weight outperforms the independent generator and discriminator conspicuously, in particular 0.4 point improvements for the generator when pre-training 800 epochs. We hypothesize that GAN-MAE framework benefits from both mask-then-reconstruct and pseudo-patch classification tasks, which can incorporate the visual semantic and consistency understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.893, + 0.735 + ], + "angle": 0, + "content": "It is also surprising that under the GAN process without weight sharing, the discriminator with learning of patch classification can lead to better performance than generator. We suspect that the fine-grained patch classification is more delicate while harder than image-level classification. Thus, we further tried to add an image-level contrastive objective. For this task, we input \\(50\\%\\) of the input image unchanged rather than noising them with the generator. We then added a prediction head to the model that predicted if the entire input image was corrupted or not. However, the result didn't improve the final accuracy. In conclusion, we believe that the design of discriminative task, e.g., getting closer to downstream and be more difficult, is an important exploration direction in the future. Another interesting finding is that with the pre-training epochs increase, the independent generator performance boosts, we believe that it is caused by the adversarial training, where the discriminator becomes an effective guiding for generation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.754, + 0.892, + 0.8 + ], + "angle": 0, + "content": "Training Schemes. We analyze the effect of proposed adversarial training scheme on GAN-MAE with shared backbones. Two training variants are considered as following:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.811, + 0.894, + 0.902 + ], + "angle": 0, + "content": "- Two-stage Training. It is natural to consider to remove the discriminator guiding signal during generator training, which leads to the disentangled optimizing process. Specifically, at each epoch, we do the following steps: train the generator only with \\( L_{mae} \\) and train the discriminator only with \\( L_{disc} \\). The difference with ad" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24453" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.09, + 0.411, + 0.104 + ], + "angle": 0, + "content": "Table 2. Effect of different training schemes." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.114, + 0.455, + 0.235 + ], + "angle": 0, + "content": "
ModelsEpochMask ratioFTGPU Time
Two-stage30075%82.094.3h
Combined30075%82.290.9h
Adversarial30075%82.2118.8h
Two-stage80075%84.0252.2h
Combined80075%84.1240.5h
Adversarial80075%84.3317.5h
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.26, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Table 3. Effect of different masking ratio in GAN-MAE framework with different pre-training epochs." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.3, + 0.413, + 0.445 + ], + "angle": 0, + "content": "
ModelsEpochMask ratioFT
MAE80075%83.4
GAN-MAE80070%84.3
GAN-MAE80075%84.3
GAN-MAE80080%84.0
MAE160075%83.6
GAN-MAE160070%84.5
GAN-MAE160075%84.6
GAN-MAE160080%84.4
" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.482, + 0.47, + 0.514 + ], + "angle": 0, + "content": "versarial training lies in that the training loss of the generator in the first step is changed." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.524, + 0.47, + 0.6 + ], + "angle": 0, + "content": "- Combined Training. Instead of iterative in total dataset, we can jointly trains the generator and discriminator at each step. That is, for each image \\( X \\) in the training dataset \\( \\mathcal{D}_{train} \\), we can directly minimize the combined loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.613, + 0.47, + 0.63 + ], + "angle": 0, + "content": "\\[\nL _ {g e n} \\left(\\theta_ {m a e}\\right) + \\lambda L _ {d i s c} \\left(\\theta_ {d i s c}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.641, + 0.47, + 0.687 + ], + "angle": 0, + "content": "We set \\(\\lambda\\), the weight for the discriminator objective in the loss to 2.0, as we searched for \\(\\lambda\\) out of [1,2,5,10] in early experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Note that regardless of the form, the nature of sequential training, i.e., the corrupt image built from generator will be fed into discriminator, is not changed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.471, + 0.88 + ], + "angle": 0, + "content": "The evaluation results for classification are listed in Table 2. GPU time means pre-training time (hours) on 8 V100 32GB GPUs environment. As we can see, the combined training shows a superior training time while adversarial training slows down as the training procedures of generator and discriminator are isolated. On the other hand, the performance of adversarial training is not better than combined training in the early stages; when pre-training epochs come to 800, the benefits of adversarial training appear." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 4. Comparison of computation resource usage during self-supervised pre-training." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.129, + 0.851, + 0.206 + ], + "angle": 0, + "content": "
BackboneModelsFLOPsParams
ViT-BMAE9.4e9111.654M
ViT-BGAN-MAE2.6e10111.656M
ViT-LMAE2.0e10329.238M
ViT-LGAN-MAE8.0e10329.240M
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Masking Ratio. Table 3 shows the influence of masking ratio in MAE-GAN under different pre-training epochs. The optimal ratio for MAE-GAN is identical to \\(75\\%\\), showing a obviously better classification performance compared with same masking ratio in previous work [28]. Meanwhile, we present several reconstructed images from MAE and GAN-MAE. Although the MIM model can infer missing patches as different yet plausible outputs when mask ratio comes to large, the generator of GAN-MAE can predict the images with more realistic and fine-grained details, e.g., the outline of a mountain peak in first case, which we believe is relative to the learning of useful representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Computation Resource. In terms of pre-training cost, we conduct a computing resource comparison with the baseline MAE on different vision transformer backbones. The results are shown in Table 4. We first point out that GAN-MAE slows down the training process as the discriminator integration, e.g. 317.5h vs. 127.7h for ViT-B model in 800 epochs. Moreover, we also choose to measure computation usage in terms of floating point operations (FLOPs) as it is a metric agnostic to the particular hardware, low-level optimizations, etc. Note that an \"operation\" is a mathematical operation, not a machine instruction and thop package is used to compute FLOPs in practice. As expected, GAN-MAE employs \\(\\sim 3\\) extra theoretical computation cost. Besides, benefits from the weight sharing of backbone, the GAN-MAE incorporates less than \\(1\\%\\) addition parameter, which helps to reduce the running memory usage, and provides a convenience to set large batch size. Last but not least, we want to state that although our GAN-MAE method incorporates additional computation resources during training at each epoch, considering the superiority of reduction of epoch number and classification performance improvements, it is fully acceptable and exploration valuable." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.819, + 0.804 + ], + "angle": 0, + "content": "4.3. ImageNet Classification Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We compare our GAN-MAE methods with previous state-of-the-art works on the ImageNet-1K classification task. Table 5 reports the top-1 validation accuracy for finetuning results. We can find that compared to the supervised models, trained from scratch, all of the self-supervised pre-training methods achieve significant improvement, sug" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24454" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.093, + 0.478, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.174, + 0.476, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.257, + 0.478, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.094, + 0.88, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.174, + 0.88, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.257, + 0.88, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.895, + 0.409 + ], + "angle": 0, + "content": "Figure 3. Qualitative analysis for patch reconstruction. Example results are from ImageNet validation set. For each tuple, we show the raw image, masked image, MAE reconstructed image, and our proposed GAN-MAE reconstructed image from left to right. We can see that the reconstructed images from GAN-MAE are significantly clearer than MAE, which we believe benefits the fine-grained visual semantic understanding." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.422, + 0.893, + 0.452 + ], + "angle": 0, + "content": "Table 5. End-to-end fine-tuning on ImageNet-1K. We report the fine-tuning top-1 accuracy for classification in different vision transformer architectures and results show that GAN-MAE outperforms previous self-supervised methods." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.461, + 0.773, + 0.653 + ], + "angle": 0, + "content": "
ModelPre-train dataPre-train epochsViT-SViT-BViT-L
Supervised [59]IN1K w/ labels30079.781.882.6
DINO [9]IN1K80081.582.8-
MoCo v3 [14]IN1K30081.483.284.1
BEiT [3]IN1K+DALLE80081.783.285.2
MSN [1]IN1K600-83.4-
iBOT [77]IN1K80082.384.084.8
BootMAE [20]IN1K800-84.285.9
MAE [28]IN1K800-83.485.4
MAE [28]IN1K1600-83.685.9
GAN-MAEIN1K30082.284.085.6
GAN-MAEIN1K80082.484.386.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.884 + ], + "angle": 0, + "content": "gesting the effectiveness of pre-training. We further compare our GAN-MAE framework with prior popular self-supervised pre-training models. We can see that the proposed GAN-MAE achieves the best fine-tuning performance either based on the ViT-B or based on the ViT-L architectures. For example, compared with the recent work MAE [28], our GAN-MAE in ViT-L network achieves \\(86.1\\%\\) with 0.7 point improvement when pre-trained on 800 epochs. More encouragingly, for all ViT backbone sizes, GAN-MAE mostly outperforms the previous self-supervised methods. These results suggest that the incorporation of a discriminator scheme could have consistently benefits for various scale ViT models." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In addition, we present a comprehensive comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "with MAE under different pre-training epochs for the ViT-B model. We plot the results in Figure 1. We can see that our GAN-MAE approach consistently performs better than MAE. It is worth mentioning that the proposed GAN-MAE at 300 epochs achieves \\(84.0\\%\\) accuracy, which is already better than MAE pre-trained at 1600 epochs. This demonstrates that our approach is more efficient to achieve comparable performance. What's more, no additional speed and parameter cost during inference. Besides, we can confirm that, similar to other MIM-based self-supervised training, the accuracy of GAN-MAE also improves steadily as the pre-training steps increase. Particularly, our GAN-MAE on ViT-B achieves \\(84.6\\%\\) top-1 accuracy at 1600 epochs, which is almost \\(1\\%\\) higher than that of MAE." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24455" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 6. Robustness Evaluation on the four ImageNet-variants: ImageNet-C, ImageNet-A, ImageNet-R, and ImageNet-Sketch. Except for ImageNet-C which is measured in terms of mean Corruption Error (mCE), top-1 accuracy is used as the remaining evaluation metric. For simplicity, we denoted IN-C, IN-A, IN-R, In-Sketch correspondingly." + }, + { + "type": "table", + "bbox": [ + 0.167, + 0.142, + 0.807, + 0.212 + ], + "angle": 0, + "content": "
ModelIN-C (mCE ↓)IN-A (top-1 ↑)IN-R (top-1 ↑)IN-Sketch (top-1 ↑)
Supervised [53]42.535.848.736.0
MAE [28]51.735.948.334.5
GAN-MAE49.536.849.635.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.246, + 0.47, + 0.274 + ], + "angle": 0, + "content": "Table 7. Semantic segmentation comparison on the ADE20K dataset for mIoU (%) metric with the ViT-B backbone." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.285, + 0.449, + 0.415 + ], + "angle": 0, + "content": "
ModelsPre-train dataEpochsmIoU
Supervised [28]IN1K w/ labels30047.4
MoCo v3 [14]IN1K30047.3
BEiT [3]IN1K+DALLE80047.1
MAE [28]IN1K80047.6
MAE [28]IN1K160048.1
BootMAE [20]IN1K80049.1
GAN-MAEIN1K80049.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.448, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Table 8. COCO object detection and segmentation using Mask R-CNN framework with ViT-B backbone." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.487, + 0.462, + 0.632 + ], + "angle": 0, + "content": "
ModelsPre-train dataAP-boxAP-mask
Supervised [28]IN1K w/ labels44.139.8
MoCo v3 [14]IN1K44.940.4
BEiT [3]IN1K+DALLE46.341.1
MSN [1]IN1K46.641.5
iBOT [77]IN1K47.342.2
MAE [28]IN1K47.242.0
BootMAE [20]IN1K48.543.4
GAN-MAEIN1K49.043.8
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.26, + 0.692 + ], + "angle": 0, + "content": "4.4. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Semantic Segmentation. We compare our GAN-MAE with supervised as well as state-of-the-art self-supervised models on the widely used dataset ADE20K [76] for semantic segmentation. Specifically, we use the UperNet framework [68] in the experiments. We train Upernet for 160K iterations with batch size set as 64 and report the results in Table 7. The evaluation metric is mean Intersection of Union (mIoU) averaged over all semantic categories and the single-scale test results are reported. Importantly, we can see that the proposed GAN-MAE gets superior performance than all the other baselines in the same configuration, further validating the effectiveness of adversarial training with a reconstructed patch discriminator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.895, + 0.461 + ], + "angle": 0, + "content": "Object Detection and Segmentation. We also perform object detection and instance segmentation, compared with other popular self-supervised methods and the supervised model, on the COCO dataset [47]. In practice, we choose the Mask R-CNN [30] framework and adopt FPNs [46] to scale the feature map into different sizes as introduced in [44]. The performance is tested on the COCO validation set, following the previous work [18]. The results are listed in Table 8 in terms of box AP metric for object detection and mask AP metric for instance segmentation. Importantly, we can observe that our GAN-MAE model achieves \\(49.0\\%\\) for object detection and \\(43.8\\%\\) for segmentation, surpassing the previous state-of-the-art BootMAE by \\(0.5\\%\\) and \\(0.4\\%\\) point, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.895, + 0.649 + ], + "angle": 0, + "content": "Classification Robutness. Similar to [28], we further evaluate the robustness of classification performance on the four ImageNet variants, i.e., ImageNet-C [33], ImageNet-A [34], ImageNet-R [32], and ImageNet-Sketch [65], which are common benchmarks to evaluate robustness for perturbations. Table 6 demonstrates the robustness comparison with GAN-MAE and MAE using the ViT-B backbone, as well as previous supervised SoTA models. The results illustrate that GAN-MAE outperforms the MAE baseline consistently on all robustness datasets, indicating that the promising of adversarial training in representation learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.62, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have proposed a new self-supervised GAN-alike framework for visual representation learning, where a generator is used to predict masked image patches according to the visible patches and a discriminator is employed to predict whether the patch is from raw image or generated by generator. The key idea is adversarial training a shared vision Transformer to distinguish the input patches from high-quality negative samples, which we believe is beneficial for the understanding of visual conception. It works well while incorporating no much addition parameter. More encouragingly, compared to standard masked image modeling, our GAN-MAE is more compute-efficient, as fewer pre-training epochs result in a better performance on downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24456" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. arXiv preprint arXiv:2204.07141, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In Proc. ICML, pages 1298-1312. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In Proc. ICLR, 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[4] Sefi Bell-Kligler, Assaf Shocher, and Michal Irani. Blind super-resolution kernel estimation using an internal-gan. Proc. NIPS, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[5] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[6] Yoshua Bengio et al. Learning deep architectures for ai. Foundations and trends® in Machine Learning, 2(1):1-127, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.469, + 0.465 + ], + "angle": 0, + "content": "[7] Yoshua Bengio, Pascal Lamblin, Dan Popovici, and Hugo Larochelle. Greedy layer-wise training of deep networks. Proc. NIPS, 19, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[8] David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proc. IEEE CVPR, pages 9650-9660, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.565, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[10] Liquun Chen, Shuyang Dai, Chenyang Tao, Haichao Zhang, Zhe Gan, Dinghan Shen, Yizhe Zhang, Guoyin Wang, Ruiyi Zhang, and Lawrence Carin. Adversarial text generation via feature-mover's distance. Proc. NIPS, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.621, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[11] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proc. ICML, pages 1691-1703. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.677, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proc. ICML, pages 1597-1607. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.733, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[13] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.803, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proc. IEEE ICCV, pages 9640-9649, 2021. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE CVPR, pages 8789-8797, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE CVPR, pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[17] Jeff Donahue, Philipp Krahenbuhl, and Trevor Darrell. Adversarial feature learning. arXiv preprint arXiv:1605.09782, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.249 + ], + "angle": 0, + "content": "[18] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proc. IEEE CVPR, pages 12124-12134, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.252, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[20] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022. 7, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.397, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. ICLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[22] Bo Du, Wei Xiong, Jia Wu, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Stacked convolutional denoising auto-encoders for feature representation. IEEE transactions on cybernetics, 47(4):1017-1027, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[23] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[24] William Fedus, Ian Goodfellow, and Andrew M Dai. Maskgan: Better text generation via filling in the .. In Proc. ICLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[25] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Proc. NIPS, 33:21271-21284, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE CVPR, pages 16000-16009, 2022. 1, 2, 3, 5, 6, 7, 8, 12" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24457" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE CVPR, pages 9729-9738, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.151, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[30] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. IEEE ICCV, pages 2961-2969, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.471, + 0.236 + ], + "angle": 0, + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE CVPR, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.239, + 0.47, + 0.307 + ], + "angle": 0, + "content": "[32] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proc. IEEE CVPR, pages 8340-8349, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.47, + 0.352 + ], + "angle": 0, + "content": "[33] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[34] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proc. IEEE CVPR, pages 15262-15271, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.398, + 0.47, + 0.438 + ], + "angle": 0, + "content": "[35] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.47, + 0.495 + ], + "angle": 0, + "content": "[36] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE CVPR, pages 1125-1134, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.499, + 0.47, + 0.553 + ], + "angle": 0, + "content": "[37] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL, pages 4171–4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.557, + 0.47, + 0.598 + ], + "angle": 0, + "content": "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[39] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In Proc. ECCV, pages 577-593. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.47, + 0.74 + ], + "angle": 0, + "content": "[40] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proc. IEEE CVPR, pages 4681-4690, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[41] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In Proc. IEEE CVPR, pages 7287-7296, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[42] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proc. AAAI, volume 34, pages 11336-11344, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[43] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[44] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[45] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In Proc. IEEE CVPR, pages 9392-9400, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[46] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proc. IEEE CVPR, pages 2117-2125, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755. Springer, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[48] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, and Jing Liao. Pd-gan: Probabilistic diverse gan for image inpainting. In Proc. IEEE CVPR, pages 9371-9381, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[49] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[50] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. IEEE ICCV, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[51] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[52] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In Proc. IEEE CVPR, pages 2794-2802, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[53] Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. In Proc. IEEE CVPR, pages 12042-12051, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[54] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proc. ECCV, pages 69-84. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[55] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proc. IEEE CVPR, pages 2536-2544, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[56] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[57] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[58] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, pages 6105-6114. PMLR, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24458" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proc. ICML, pages 10347-10357. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[60] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[61] Michael Tschannen, Olivier Bachem, and Mario Lucic. Recent advances in autoencoder-based representation learning. arXiv preprint arXiv:1812.05069, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NIPS, pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[63] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proc. ICML, pages 1096-1103, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[64] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[65] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Proc. NIPS, 32, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[66] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proc. IEEE CVPR, pages 14668-14678, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[67] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proc. IEEE CVPR, pages 2256-2265, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[68] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proc. ECCV, pages 418-434, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[69] Junyuan Xie, Linli Xu, and Enhong Chen. Image denoising and inpainting with deep neural networks. Proc. NIPS, 25, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[70] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proc. IEEE CVPR, pages 9653-9663, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[71] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Proc. NIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[72] Richard Zhang, Phillip Isola, and Alexei A Efros. **Colorful image colorization. In Proc. ECCV, pages 649-666. Springer, 2016. 2**" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[73] Yizhe Zhang, Zhe Gan, Kai Fan, Zhi Chen, Ricardo Henao, Dinghan Shen, and Lawrence Carin. Adversarial feature matching for text generation. In Proc. ICML, pages 4006-4015. PMLR, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[74] Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial networks. In Proc. ICLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[75] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proc. IEEE CVPR, pages 18697–18709, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proc. IEEE CVPR, pages 633-641, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[77] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24459" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_origin.pdf b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..715c9ca156894486650a2bcc53008a1cfa5ca739 --- /dev/null +++ b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/ba6bf709-c33b-4412-be5d-32997f46a850_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68baa4c816e73d02543a7615083297392ebfe94490f3f16c3eeb940c919d2a9e +size 5141847 diff --git a/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/full.md b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/full.md new file mode 100644 index 0000000000000000000000000000000000000000..94202a77d336ad8e72bccc40597d62cff07b31c8 --- /dev/null +++ b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/full.md @@ -0,0 +1,333 @@ +# Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond + +Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang* Xiaoming Wei, Xiaolin Wei Meituan {feizhengcong, fanmingyuan, zhuli09, huangjunshi}@meituan.com {weixiaoming, weixiaolin02}@meituan.com + +# Abstract + +Masked Auto-Encoder (MAE) pretraining methods randomly mask image patches and then train a vision Transformer to reconstruct the original pixels based on the unmasked patches. While they demonstrate impressive performance for downstream vision tasks, it generally requires a large amount of training resource. In this paper, we introduce a novel Generative Adversarial Networks alike framework, referred to as GAN-MAE, where a generator is used to generate the masked patches according to the remaining visible patches, and a discriminator is employed to predict whether the patch is synthesized by the generator. We believe this capacity of distinguishing whether the image patch is predicted or original is benefit to representation learning. Another key point lies in that the parameters of the vision Transformer backbone in the generator and discriminator are shared. Extensive experiments demonstrate that adversarial training of GAN-MAE framework is more efficient and accordingly outperforms the standard MAE given the same model size, training data, and computation resource. The gains are substantially robust for different model sizes and datasets, in particular, a ViT-B model trained with GAN-MAE for 200 epochs outperforms the MAE with 1600 epochs on fine-tuning top-1 accuracy of ImageNet-1k with much less FLOPs. Besides, our approach also works well at transferring downstream tasks. + +# 1. Introduction + +In recent years, Transformer [62] has become the de facto standard architecture in computer vision, and has surpassed state-of-the-art Convolutional Neural Network (CNN) [31, 58] feature extractors in vision tasks through models such as the Vision Transformer [21]. Meanwhile, self-supervised learning (SSL) algorithms [12, 14, 27, 29] aims to learn transferable representation from unlabeled + +![](images/987b48bb5a14bd972540c7a2439719338dbd2790d48f3b7bfbd7809f1f0e08bb.jpg) +Figure 1. Performance comparison in different pre-training epochs for ImageNet-1K Fine-tuning top-1 accuracy. Compared to MAE trained for 1600 epochs, GAN-MAE achieves comparable accuracy with much less training time at 200 epochs. + +data by performing instance-level pretext tasks, and has been a long-standing target in the vision community. Particularly, masked image modeling (MIM) in SSL for vision transformers has shown remarkably impressive downstream performance in a wide variety of computer vision tasks [3, 28], attracting increasing attention. + +MIM is a simple pretext task that first randomly masks some patches of an image, and then predicts the contents of the masked patches according to the remaining, using various reconstruction targets, e.g., visual tokens [3, 19], semantic features [1, 77] and raw pixels [28, 70]. Essentially, it learns the transferable representation by modeling the image structure itself as content prediction. While more effective than conventional pre-training, masked autoencoder modeling approaches still exist some issues: (i) reconstruction optimization with MSE loss leads to blurrier output images than the raw input, it would be better to use a more perceptual loss over pixels to guide the fine-grained seman + +tic understanding and representation learning, leading to more plausible synthesized patches; (ii) inner dependency between masked patches is lacked [71], i.e., generation of masked image patches may lack the surrounding information. This situation becomes more serious when the image patch masking ratio is large. We alleviate this problem by introducing confident synthesized patches as complementary information during training; (iii) mask-reconstruction methods incur a substantial computation cost because the network only learns from part of the visible patches and misses the information of masked patches. + +In this paper, we propose a Generative Adversarial Networks-based pre-training framework, referred to as GAN-MAE, which contains two components: a generator model learns to reconstruct the masked patches according to visible patches in the encoder-decoder architecture and a discriminator model learns to distinguish real image patches from plausible but synthesized remains. Generally, given an image from training dataset, our method first randomly masks parts of patches and reconstructs them using the rest visible patches with a generator, which serves as a standard MAE model. Then we build the corrupt image as the combination of visible and synthesis patches, which is then fed into the discriminator to predict whether each patch is from raw image or synthesized results. In this manner, the discriminator provides a valid guiding for more delicate image patch modeling. Then, with the development of generator capacity, a key advantage of discriminative task is that it integrates the synthesized patches into corrupt images as complementary information, which fills the missing inner relationship between patches during pre-training. Moreover, we shared the parameters of vision transformer backbone in the generator and discriminator to promote memory reduction, training efficiency, as well as performance enhancement. + +Our experiments follow the same architecture, settings, and pre-training recipe as MAE [28], and we find that the simple incorporation of a discriminator consistently outperforms MAE in variant models, e.g., ViT-S, ViT-B, and ViT-L, when fine-tuning for top-1 accuracy of ImageNet classification. We also conduct extensive ablation studies to validate the effectiveness of our core designs in backbone parameter sharing and advarial training. As pre-training with more epochs usually results in a better downstream performance, we argue that an important consideration for pre-training methods should be computation efficiency as well as absolute downstream performance. From this viewpoint, we also demonstrate that discrimination of pseudo-image patches forces GAN-MAE to train more efficiently than standard MAE. We further provide a comprehensive comparison with MAE in various epochs and various models and show our framework achieves consistently better performance. In particular, as presented in Figure 1, for the ViT-B model structure, our GAN-MAE achieves comparable + +ble classification performance with only 200 pre-training epochs vs. standard MAE 1600 pre-training epochs. Furthermore, the GAN-MAE achieves 0.7 points improvement when pre-training 1600 epochs. Finally, we summarize our contribution as follows: + +- We propose a new and effective GAN-alike framework for visual representation self-supervised learning, which to our best knowledge is the first trial of integrating GAN idea into MAE framework. As a generic approach, we suggest that this framework can be easily applied on many other MIM-based tasks. +- We introduce two core designs: shared weight for the main backbones of generator and discriminator, and an adversarial training process, both of which cost fewer amounts of computing resources while obtaining appreciable performance improvements. +- Extensive experiments demonstrate that compared with the original MAE, our method is more compute-efficient and results in better transfer representation learning on downstream tasks. + +# 2. Related Works + +Autoencoding. Autoencoder [6, 7, 35] is an unsupervised learning technique for neural networks that learns efficient representations by training the network to ignore signal noise. It includes an encoder that maps the original data to a low-dimensional latent embedding and a decoder that recovers the data from the latent embedding, with the goal of learning a compressed knowledge representation. The denoising autoencoder [63, 64] learns to reconstruct clean data points from a noisy version. Numerous efforts have been devoted for image denoising, such as masking pixels [11, 55, 64], inpainting [69], removing color channels [39, 72], and shuffling image patches [22, 54]. For a broader overview of denoising autoencoder, we refer to [5, 61]. + +Masked Image Modeling. Masked language modeling [37,49,57], which generalizes well on language understanding and generation tasks, is the domain self-supervised approach in the field of NLP. Similarly, vision transformer [21, 41, 50] based masked image modeling (MIM) approaches [1,3,28,70,77] for computer vision tasks have also been developed. Generally, these MIM approaches first apply a mask to patches of an image, and then the masked patches are predicted given the visible patches. Feature representation learned through such within-image context prediction demonstrate strong transfer performance in downstream tasks. Recently, lots of works exploring MIM have been concurrently developed from different perspectives. + +![](images/5b1b341053e9586bee822eb3f148d6d85160fd4bd229637788059f0a4d011d6a.jpg) +Figure 2. Overview of GAN-MAE framework, where a generator is used to predict the masked patches and a discriminator is employed to classify whether the patches are selected from the raw image or synthesized results. In particular, both of MAE encoder and discriminator are based on the vision Transformer backbone and share parameters for memory reduction and training efficiency. + +The works include framework design [13, 23, 28, 70], prediction targets [2, 19, 66, 77], and integration with vision-language representation learning [42, 43, 75]. Our work belongs to the first group and introduces a novel GAN framework that discriminates the reconstructed image patches in a standard MAE to learn deeper semantics. + +Generative Adversarial Networks. GANs [26] are effective at generating high-quality synthetic data. Usually, the generator generates an image, and the discriminator determines whether the input image is a real image or a generated image. Subsequently, many improvements based on the original GAN focused on speeding up the training of the network and improving the quality of the generated images [8, 52, 74]. These improvements also help GAN achieving a wider range of applications [45, 48, 67]. Methods based on GAN are also widely used in image-to-image translation [36], super-resolution [4, 40], style transfer [15], text generation [10, 24, 73], and representation learning [17, 25], to name a few. Particularly, [56] brings some designs of CNN architecture to stabilize the training of GAN framework, after that the discriminator can be directly used as feature extractor in downstream tasks. In contrast, our method proposes to integrate the GAN as assistant for the MIM task, which focuses on the study of better pattern for masking based self-supervised learning. Besides, our strategy of shared parameters provides a unified backbone for better vision representation. + +# 3. Approach + +In this section, we introduce the GAN-MAE framework in details. At first, we briefly review the conventional masked autoencoder model in Sec. 3.1, and then describe the proposed generator-discriminator pre-training framework in Sec. 3.2. The architecture of our framework is presented in Figure 2. Finally, we suggest an adversarial + +training processes under the proposed method and discuss our framework in Sec. 3.3 and Sec. 3.4, respectively. + +# 3.1. Preliminaries + +Masked Autoencoder (MAE) [28] is a self-supervised approach with a vision transformer encoder and a small transformer decoder, which randomly masks a large portion of input patches, and then reconstructs the masked patches according to the visible patches. Specifically, provided with an image $X \in \mathbb{R}^{C \times H \times W}$ , where $C$ , $H$ and $W$ are the channel number, image height and image width respectively, MAE partitions $X$ into $N = \frac{H \times W}{P^2}$ non-overlapping patches with patch size $P$ . In this way, the image is transformed into a sequence of patches $X = \{x^1, \ldots, x^N\}$ with each element $x^k \in \mathbb{R}^{P^2 \times C}$ . Then, we sample a random set of patch index $M$ in uniform distribution, and split the image patches $X$ into masked patch set $X_m = \{x^k | k \in M\}$ and visible patch set $X_v = \{x^k | k \notin M\}$ . During training, the MAE encoder inputs $X_v$ to achieve the latent representations $H_v$ . Then, the MAE decoder attempts to reconstruct $X_m$ with the input of interpolating [mask] token embedding into the sequence of latent representations $H_v$ according to the index set $M$ , and outputs the reconstructed patches $\tilde{X}_m$ . Finally, MAE optimizes the mean-squared error reconstruction loss on the masked patches as: + +$$ +L _ {m a e} (X, M, \theta_ {m a e}) = \sum_ {k \in M} | | \tilde {x} ^ {k} - x ^ {k} | | _ {2} ^ {2}, \tag {1} +$$ + +where $\theta_{mae}$ represents the parameters of MAE model. + +# 3.2. When MAE meet GAN + +In this section, we describe the proposed GAN-MAE framework, as presented in Figure 2. Generally, our framework consists of two parts, a generator $G$ to reconstruct the masked image patches and a discriminator $D$ to predict the realness of image patches. + +Image Patch Generator. Identical to the standard MAE, the generator follows the encoder-decoder paradigm and is trained to perform masked-image reconstruction task. Given the partitioned image patches $X$ , the generator randomly masks some image patches and encodes the remaining visible patches $X_{v}$ into a sequence of contextualized vector representations $H_{v}$ , based on which the masked patches are reconstructed as $\hat{X}_m$ . Please refer to Sec. 3.1 for more details. In general, the generation process can be formulated as: + +$$ +M \sim \operatorname {U n i f o r m} (1, N), \tag {2} +$$ + +$$ +H _ {v} = f _ {e} \left(X _ {v}, M\right), \tag {3} +$$ + +$$ +\tilde {X} _ {m} = f _ {d} \left(H _ {v}, M\right), \tag {4} +$$ + +where $N$ is the number of image patches. $f_{e}(\cdot)$ and $f_{d}(\cdot)$ denote the encoder and decoder in a conventional MAE. + +Image Patch Discriminator. For a patch index $k$ and corrupted image sequence $\overline{X} = \{X_v, \tilde{X}_m\}$ , the discriminator predicts whether the patch token $x^k$ is real or synthesized as binary classification task. Specifically, we create the corrupted image $\overline{X}$ by maintaining the visible patches $X_v$ in raw image $X$ and replacing the masked patches with generator predicted result $\tilde{X}_m$ . Note that $X_v$ and $\tilde{X}_m$ are absolute complement of set to each other, i.e., $X_v \cup \tilde{X}_m = \overline{X}$ and $X_v \cap \tilde{X}_m = \emptyset$ . Formally, the task of discriminator model can be formulated as: + +$$ +D (\bar {X}, k) = p _ {\text {d i s c}} \left(y ^ {k} | \bar {X}, k\right), \text {f o r} k \in [ 1, N ]. \tag {5} +$$ + +Let the ground-truth classification label sequence $Y = \{y^{1},\ldots ,y^{N}\}$ with each element $y_{k}\in \{0,1\}$ , where 0 and 1 denote the corresponding image patch is reconstructed from generator or comes from the original image, respectively. The training objective of the discriminator can be formulated as: + +$$ +\begin{array}{l} L _ {d i s c} (\bar {X}, \theta_ {d i s c}) = \sum_ {k = 1} ^ {N} - y ^ {k} \log D (\bar {X}, k) \tag {6} \\ - \left(1 - y ^ {k}\right) \log \left(1 - D (\overline {{X}}, k)\right). \\ \end{array} +$$ + +Particularly, though the corrupted image is partially unreal, we suggest that this discriminative task can still be benefit to the learning of feature representation with the plausible corrupted image from a well-trained generator, as presented in experiments. Hitherto, due to the GAN-alike strategy for MAE task, we name our framework as GAN-MAE. + +# 3.3. Training Scheme + +We explore the training strategy of our proposed GANMAE in this section. Particularly, to improve synthesis result, we augment the L-2 reconstruction loss with a percep + +Algorithm 1: Adversarial training for GAN-MAE +Data: Training data $\mathcal{D}_{train}$ total epoch number $N_{e}$ GAN-MAE model with generator parameters $\theta_{mae}$ and discriminator parameters $\theta_{disc}$ +1 share weights between generator and discriminator backbones; +2 while $n_e < N_e$ do +3 for $x^{i}\in \mathcal{D}_{train}$ do +4 $\triangleright$ generator training; +5 sample masking set $M^i$ and mask image $x^i$ . +6 predict masked image patches $\tilde{x}_m^i$ . +7 compute loss $L_{gen}$ . +8 loss backward for updating $\theta_{mae}$ . +9 dicriminator training; +10 construct $\overline{x}^i$ based on $x^i$ and $\tilde{x}_m^i$ . +11 comput loss $L_{disc}$ . +12 loss backward for updating $\theta_{disc}$ . +13 end +14 $n_e + = 1$ . +end + +tual loss that aims to differentiate the real patches and reconstructed patches as: + +$$ +L _ {a d v} \left(X, \theta_ {m a e}\right) = \log D \left(X _ {v}\right) + \log \left(1 - D \left(\tilde {X} _ {m}\right)\right). \tag {7} +$$ + +Please note that $\tilde{X}_m$ is the reconstructed patches by MAE. Therefore, the final objective for the generator comes to minimize the combined loss as: + +$$ +L _ {\text {g e n}} (X, \theta_ {\text {m a e}}) = L _ {\text {m a e}} (X, \theta_ {\text {m a e}}) + \gamma L _ {\text {a d v}} (X, \theta_ {\text {m a e}}), \tag {8} +$$ + +where we compute the adaptive weight $\gamma$ according to: + +$$ +\gamma = \frac {\nabla \left[ L _ {m a e} \right]}{\nabla \left[ L _ {a d v} \right] + \delta}, \tag {9} +$$ + +$\nabla[\cdot]$ denotes the gradients of different loss function w.r.t. the parameters of last layer in network, and $\delta = 1e - 6$ is used for numerical stability. Intuitively, the integration of $\gamma$ adaptively balances the contributions of two loss functions to the gradients of parameters. + +Based on the aforementioned strategy, the final training algorithm can be formulated in Algorithm 1. Specifically, at each epoch, we conduct the iteration in two steps: (i) train only the generator with $L_{gen}$ ; (ii) Train the discriminator with $L_{disc}$ . During training, we observe that shared weights of generator and discriminator backbones can usually stabilize the training procedure and bring extra-gain in down-stream tasks. + +# 3.4. Discussion + +Theoretically, the integration of discriminator can be considered as a high-level perceptual loss, which forces the + +generator to learn better feature representation for the plausible synthesis of masked patches. As the quality of synthesized patches improved, we claim that the introduction of corrupted images, serving as complementary information, provides plausible inner dependency between full image patches for representation learning. Furthermore, with the shared parameters of backbones in generator and discriminator, the GAN-alike framework can be considered as a type of multi-task learning, leading to even better result as presented in experiments. + +# 4. Experiments + +# 4.1. Datasets and Settings + +In the experiments, we pre-train our GAN-MAE model on the ImageNet-1k [16] and evaluate the performance in end-to-end fine-tuning (FT) pattern for the task of classification, semantic segmentation, object detection and instance segmentation. The evaluation metric of classification is the top-1 validation accuracy on $224 \times 224$ cropped input images. The input image is partitioned into $14 \times 14$ patches and each patch is of size $16 \times 16$ . Following the setting of MAE, we only use the standard random cropping and horizontal flipping for data augmentation. To validate the effectiveness of GAN-MAE framework, the used ViT architecture and most hyper-parameters are exactly the same to [28, 60], i.e., ViT-S (12 transformer blocks with dimension 384), ViT-B (12 transformer blocks with dimension 768), and ViT-L (24 transformer blocks with dimension 1024). All version of ViT models are trained with 4096 batch size on 8 V-100 32GB GPUs. We adopt dynamic token masking with the masked positions decided on-the-fly. We use AdamW [38] optimizer and cosine schedule [51] with warm up for model training. The learning rate is annealed according to the cosine schedule. Unless stated otherwise, results are evaluated on the dev set. Please refer to appendix A for more training implementation and hyperparameter values for different backbone in details. + +# 4.2. Analysis of GAN-MAE + +We analyze our GAN-MAE framework by proposing and evaluating several extensions to the model. Unless stated otherwise, all these experiments use the same model size as ViT-B and training dataset ImageNet-1K. + +Parameter Sharing of Backbone. In this work, we propose to improve the efficiency of the pre-training by sharing the parameters of the backbone vision transformer between the generator encoder and discriminator. One cause is that the generator and discriminator utilize the same network architecture, and all of the transformer weights can be tied. However, we can release the weight sharing and train the generator and discriminator independently. In between, we + +Table 1. Effect of parameter sharing in GAN-MAE framework. Results demonstrate that shared parameters for backbone benefits both memory cost and performance improvements. + +
ModelsEpochMask ratioFT
Generator80075%83.9
Discriminator80075%84.2
Shared80075%84.3
Generator160075%84.4
Discriminator160075%84.4
Shared160075%84.6
+ +can adopt both the parameters of generator and discriminator for downstream learning. The experimental results are shown in Table 1, where we employ the adversarial training with the same training epochs. We can see that, generally, the fine-tuning top-1 accuracy of shared weight outperforms the independent generator and discriminator conspicuously, in particular 0.4 point improvements for the generator when pre-training 800 epochs. We hypothesize that GAN-MAE framework benefits from both mask-then-reconstruct and pseudo-patch classification tasks, which can incorporate the visual semantic and consistency understanding. + +It is also surprising that under the GAN process without weight sharing, the discriminator with learning of patch classification can lead to better performance than generator. We suspect that the fine-grained patch classification is more delicate while harder than image-level classification. Thus, we further tried to add an image-level contrastive objective. For this task, we input $50\%$ of the input image unchanged rather than noising them with the generator. We then added a prediction head to the model that predicted if the entire input image was corrupted or not. However, the result didn't improve the final accuracy. In conclusion, we believe that the design of discriminative task, e.g., getting closer to downstream and be more difficult, is an important exploration direction in the future. Another interesting finding is that with the pre-training epochs increase, the independent generator performance boosts, we believe that it is caused by the adversarial training, where the discriminator becomes an effective guiding for generation. + +Training Schemes. We analyze the effect of proposed adversarial training scheme on GAN-MAE with shared backbones. Two training variants are considered as following: + +- Two-stage Training. It is natural to consider to remove the discriminator guiding signal during generator training, which leads to the disentangled optimizing process. Specifically, at each epoch, we do the following steps: train the generator only with $L_{mae}$ and train the discriminator only with $L_{disc}$ . The difference with ad + +Table 2. Effect of different training schemes. + +
ModelsEpochMask ratioFTGPU Time
Two-stage30075%82.094.3h
Combined30075%82.290.9h
Adversarial30075%82.2118.8h
Two-stage80075%84.0252.2h
Combined80075%84.1240.5h
Adversarial80075%84.3317.5h
+ +Table 3. Effect of different masking ratio in GAN-MAE framework with different pre-training epochs. + +
ModelsEpochMask ratioFT
MAE80075%83.4
GAN-MAE80070%84.3
GAN-MAE80075%84.3
GAN-MAE80080%84.0
MAE160075%83.6
GAN-MAE160070%84.5
GAN-MAE160075%84.6
GAN-MAE160080%84.4
+ +versarial training lies in that the training loss of the generator in the first step is changed. + +- Combined Training. Instead of iterative in total dataset, we can jointly trains the generator and discriminator at each step. That is, for each image $X$ in the training dataset $\mathcal{D}_{train}$ , we can directly minimize the combined loss as: + +$$ +L _ {g e n} \left(\theta_ {m a e}\right) + \lambda L _ {d i s c} \left(\theta_ {d i s c}\right). \tag {10} +$$ + +We set $\lambda$ , the weight for the discriminator objective in the loss to 2.0, as we searched for $\lambda$ out of [1,2,5,10] in early experiments. + +Note that regardless of the form, the nature of sequential training, i.e., the corrupt image built from generator will be fed into discriminator, is not changed. + +The evaluation results for classification are listed in Table 2. GPU time means pre-training time (hours) on 8 V100 32GB GPUs environment. As we can see, the combined training shows a superior training time while adversarial training slows down as the training procedures of generator and discriminator are isolated. On the other hand, the performance of adversarial training is not better than combined training in the early stages; when pre-training epochs come to 800, the benefits of adversarial training appear. + +Table 4. Comparison of computation resource usage during self-supervised pre-training. + +
BackboneModelsFLOPsParams
ViT-BMAE9.4e9111.654M
ViT-BGAN-MAE2.6e10111.656M
ViT-LMAE2.0e10329.238M
ViT-LGAN-MAE8.0e10329.240M
+ +Masking Ratio. Table 3 shows the influence of masking ratio in MAE-GAN under different pre-training epochs. The optimal ratio for MAE-GAN is identical to $75\%$ , showing a obviously better classification performance compared with same masking ratio in previous work [28]. Meanwhile, we present several reconstructed images from MAE and GAN-MAE. Although the MIM model can infer missing patches as different yet plausible outputs when mask ratio comes to large, the generator of GAN-MAE can predict the images with more realistic and fine-grained details, e.g., the outline of a mountain peak in first case, which we believe is relative to the learning of useful representation. + +Computation Resource. In terms of pre-training cost, we conduct a computing resource comparison with the baseline MAE on different vision transformer backbones. The results are shown in Table 4. We first point out that GAN-MAE slows down the training process as the discriminator integration, e.g. 317.5h vs. 127.7h for ViT-B model in 800 epochs. Moreover, we also choose to measure computation usage in terms of floating point operations (FLOPs) as it is a metric agnostic to the particular hardware, low-level optimizations, etc. Note that an "operation" is a mathematical operation, not a machine instruction and thop package is used to compute FLOPs in practice. As expected, GAN-MAE employs $\sim 3$ extra theoretical computation cost. Besides, benefits from the weight sharing of backbone, the GAN-MAE incorporates less than $1\%$ addition parameter, which helps to reduce the running memory usage, and provides a convenience to set large batch size. Last but not least, we want to state that although our GAN-MAE method incorporates additional computation resources during training at each epoch, considering the superiority of reduction of epoch number and classification performance improvements, it is fully acceptable and exploration valuable. + +# 4.3. ImageNet Classification Comparison + +We compare our GAN-MAE methods with previous state-of-the-art works on the ImageNet-1K classification task. Table 5 reports the top-1 validation accuracy for finetuning results. We can find that compared to the supervised models, trained from scratch, all of the self-supervised pre-training methods achieve significant improvement, sug + +![](images/4beaed6f5b05eb57a4cf78ad30018dfcf433231e01ff5836ca4a2607b588199c.jpg) + +![](images/2ddcfbc8bfd3b604331fddb3c4f1a123e6120b15a9d44aa1479eaf1c844a3291.jpg) + +![](images/0baaf630ac8e81e514ec56c7064ca6ba5cc70acc6661a703f7772ad9089199d5.jpg) +Figure 3. Qualitative analysis for patch reconstruction. Example results are from ImageNet validation set. For each tuple, we show the raw image, masked image, MAE reconstructed image, and our proposed GAN-MAE reconstructed image from left to right. We can see that the reconstructed images from GAN-MAE are significantly clearer than MAE, which we believe benefits the fine-grained visual semantic understanding. + +![](images/4c9af92e7918f7eb148ab1ff90d6cf8cd8b5e0e4605d30763e0806b85e0505a1.jpg) + +![](images/f912ca23a28c07aa3115c05807a0fde46ca7000513169496914f65f13d17c0ac.jpg) + +![](images/f31ed1cbd7180b371977d634615cb036bf4fe7423bdba14aadc5bf2ae492488c.jpg) + +Table 5. End-to-end fine-tuning on ImageNet-1K. We report the fine-tuning top-1 accuracy for classification in different vision transformer architectures and results show that GAN-MAE outperforms previous self-supervised methods. + +
ModelPre-train dataPre-train epochsViT-SViT-BViT-L
Supervised [59]IN1K w/ labels30079.781.882.6
DINO [9]IN1K80081.582.8-
MoCo v3 [14]IN1K30081.483.284.1
BEiT [3]IN1K+DALLE80081.783.285.2
MSN [1]IN1K600-83.4-
iBOT [77]IN1K80082.384.084.8
BootMAE [20]IN1K800-84.285.9
MAE [28]IN1K800-83.485.4
MAE [28]IN1K1600-83.685.9
GAN-MAEIN1K30082.284.085.6
GAN-MAEIN1K80082.484.386.1
+ +gesting the effectiveness of pre-training. We further compare our GAN-MAE framework with prior popular self-supervised pre-training models. We can see that the proposed GAN-MAE achieves the best fine-tuning performance either based on the ViT-B or based on the ViT-L architectures. For example, compared with the recent work MAE [28], our GAN-MAE in ViT-L network achieves $86.1\%$ with 0.7 point improvement when pre-trained on 800 epochs. More encouragingly, for all ViT backbone sizes, GAN-MAE mostly outperforms the previous self-supervised methods. These results suggest that the incorporation of a discriminator scheme could have consistently benefits for various scale ViT models. + +In addition, we present a comprehensive comparison + +with MAE under different pre-training epochs for the ViT-B model. We plot the results in Figure 1. We can see that our GAN-MAE approach consistently performs better than MAE. It is worth mentioning that the proposed GAN-MAE at 300 epochs achieves $84.0\%$ accuracy, which is already better than MAE pre-trained at 1600 epochs. This demonstrates that our approach is more efficient to achieve comparable performance. What's more, no additional speed and parameter cost during inference. Besides, we can confirm that, similar to other MIM-based self-supervised training, the accuracy of GAN-MAE also improves steadily as the pre-training steps increase. Particularly, our GAN-MAE on ViT-B achieves $84.6\%$ top-1 accuracy at 1600 epochs, which is almost $1\%$ higher than that of MAE. + +Table 6. Robustness Evaluation on the four ImageNet-variants: ImageNet-C, ImageNet-A, ImageNet-R, and ImageNet-Sketch. Except for ImageNet-C which is measured in terms of mean Corruption Error (mCE), top-1 accuracy is used as the remaining evaluation metric. For simplicity, we denoted IN-C, IN-A, IN-R, In-Sketch correspondingly. + +
ModelIN-C (mCE ↓)IN-A (top-1 ↑)IN-R (top-1 ↑)IN-Sketch (top-1 ↑)
Supervised [53]42.535.848.736.0
MAE [28]51.735.948.334.5
GAN-MAE49.536.849.635.9
+ +Table 7. Semantic segmentation comparison on the ADE20K dataset for mIoU (%) metric with the ViT-B backbone. + +
ModelsPre-train dataEpochsmIoU
Supervised [28]IN1K w/ labels30047.4
MoCo v3 [14]IN1K30047.3
BEiT [3]IN1K+DALLE80047.1
MAE [28]IN1K80047.6
MAE [28]IN1K160048.1
BootMAE [20]IN1K80049.1
GAN-MAEIN1K80049.5
+ +Table 8. COCO object detection and segmentation using Mask R-CNN framework with ViT-B backbone. + +
ModelsPre-train dataAP-boxAP-mask
Supervised [28]IN1K w/ labels44.139.8
MoCo v3 [14]IN1K44.940.4
BEiT [3]IN1K+DALLE46.341.1
MSN [1]IN1K46.641.5
iBOT [77]IN1K47.342.2
MAE [28]IN1K47.242.0
BootMAE [20]IN1K48.543.4
GAN-MAEIN1K49.043.8
+ +# 4.4. Downstream Tasks + +Semantic Segmentation. We compare our GAN-MAE with supervised as well as state-of-the-art self-supervised models on the widely used dataset ADE20K [76] for semantic segmentation. Specifically, we use the UperNet framework [68] in the experiments. We train Upernet for 160K iterations with batch size set as 64 and report the results in Table 7. The evaluation metric is mean Intersection of Union (mIoU) averaged over all semantic categories and the single-scale test results are reported. Importantly, we can see that the proposed GAN-MAE gets superior performance than all the other baselines in the same configuration, further validating the effectiveness of adversarial training with a reconstructed patch discriminator. + +Object Detection and Segmentation. We also perform object detection and instance segmentation, compared with other popular self-supervised methods and the supervised model, on the COCO dataset [47]. In practice, we choose the Mask R-CNN [30] framework and adopt FPNs [46] to scale the feature map into different sizes as introduced in [44]. The performance is tested on the COCO validation set, following the previous work [18]. The results are listed in Table 8 in terms of box AP metric for object detection and mask AP metric for instance segmentation. Importantly, we can observe that our GAN-MAE model achieves $49.0\%$ for object detection and $43.8\%$ for segmentation, surpassing the previous state-of-the-art BootMAE by $0.5\%$ and $0.4\%$ point, respectively. + +Classification Robutness. Similar to [28], we further evaluate the robustness of classification performance on the four ImageNet variants, i.e., ImageNet-C [33], ImageNet-A [34], ImageNet-R [32], and ImageNet-Sketch [65], which are common benchmarks to evaluate robustness for perturbations. Table 6 demonstrates the robustness comparison with GAN-MAE and MAE using the ViT-B backbone, as well as previous supervised SoTA models. The results illustrate that GAN-MAE outperforms the MAE baseline consistently on all robustness datasets, indicating that the promising of adversarial training in representation learning. + +# 5. Conclusion + +In this paper, we have proposed a new self-supervised GAN-alike framework for visual representation learning, where a generator is used to predict masked image patches according to the visible patches and a discriminator is employed to predict whether the patch is from raw image or generated by generator. The key idea is adversarial training a shared vision Transformer to distinguish the input patches from high-quality negative samples, which we believe is beneficial for the understanding of visual conception. It works well while incorporating no much addition parameter. More encouragingly, compared to standard masked image modeling, our GAN-MAE is more compute-efficient, as fewer pre-training epochs result in a better performance on downstream tasks. + +# References + +[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. arXiv preprint arXiv:2204.07141, 2022. 1, 2, 7, 8 +[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In Proc. ICML, pages 1298-1312. PMLR, 2022. 3 +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In Proc. ICLR, 2021. 1, 2, 7, 8 +[4] Sefi Bell-Kligler, Assaf Shocher, and Michal Irani. Blind super-resolution kernel estimation using an internal-gan. Proc. NIPS, 32, 2019. 3 +[5] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 2 +[6] Yoshua Bengio et al. Learning deep architectures for ai. Foundations and trends® in Machine Learning, 2(1):1-127, 2009. 2 +[7] Yoshua Bengio, Pascal Lamblin, Dan Popovici, and Hugo Larochelle. Greedy layer-wise training of deep networks. Proc. NIPS, 19, 2006. 2 +[8] David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 3 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proc. IEEE CVPR, pages 9650-9660, 2021. 7 +[10] Liquun Chen, Shuyang Dai, Chenyang Tao, Haichao Zhang, Zhe Gan, Dinghan Shen, Yizhe Zhang, Guoyin Wang, Ruiyi Zhang, and Lawrence Carin. Adversarial text generation via feature-mover's distance. Proc. NIPS, 31, 2018. 3 +[11] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proc. ICML, pages 1691-1703. PMLR, 2020. 2 +[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proc. ICML, pages 1597-1607. PMLR, 2020. 1 +[13] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 3 +[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proc. IEEE ICCV, pages 9640-9649, 2021. 1, 7, 8 +[15] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE CVPR, pages 8789-8797, 2018. 3 + +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE CVPR, pages 248–255, 2009. 5 +[17] Jeff Donahue, Philipp Krahenbuhl, and Trevor Darrell. Adversarial feature learning. arXiv preprint arXiv:1605.09782, 2016.3 +[18] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proc. IEEE CVPR, pages 12124-12134, 2022. 8 +[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. 1, 3 +[20] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022. 7, 8, 12 +[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. ICLR, 2020. 1, 2 +[22] Bo Du, Wei Xiong, Jia Wu, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Stacked convolutional denoising auto-encoders for feature representation. IEEE transactions on cybernetics, 47(4):1017-1027, 2016. 2 +[23] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 3 +[24] William Fedus, Ian Goodfellow, and Andrew M Dai. Maskgan: Better text generation via filling in the .. In Proc. ICLR, 2018. 3 +[25] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 3 +[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3 +[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Proc. NIPS, 33:21271-21284, 2020. 1 +[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE CVPR, pages 16000-16009, 2022. 1, 2, 3, 5, 6, 7, 8, 12 + +[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE CVPR, pages 9729-9738, 2020. 1 +[30] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. IEEE ICCV, pages 2961-2969, 2017. 8 +[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE CVPR, pages 770-778, 2016. 1 +[32] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proc. IEEE CVPR, pages 8340-8349, 2021. 8 +[33] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 8 +[34] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proc. IEEE CVPR, pages 15262-15271, 2021. 8 +[35] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 2 +[36] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE CVPR, pages 1125-1134, 2017. 3 +[37] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL, pages 4171–4186, 2019. 2 +[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[39] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In Proc. ECCV, pages 577-593. Springer, 2016. 2 +[40] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proc. IEEE CVPR, pages 4681-4690, 2017. 3 +[41] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In Proc. IEEE CVPR, pages 7287-7296, 2022. 2 +[42] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proc. AAAI, volume 34, pages 11336-11344, 2020. 3 +[43] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3 + +[44] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 8 +[45] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In Proc. IEEE CVPR, pages 9392-9400, 2021. 3 +[46] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proc. IEEE CVPR, pages 2117-2125, 2017. 8 +[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755. Springer, 2014. 8 +[48] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, and Jing Liao. Pd-gan: Probabilistic diverse gan for image inpainting. In Proc. IEEE CVPR, pages 9371-9381, 2021. 3 +[49] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2 +[50] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. IEEE ICCV, pages 10012-10022, 2021. 2 +[51] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5 +[52] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In Proc. IEEE CVPR, pages 2794-2802, 2017. 3 +[53] Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. In Proc. IEEE CVPR, pages 12042-12051, 2022. 8 +[54] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proc. ECCV, pages 69-84. Springer, 2016. 2 +[55] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proc. IEEE CVPR, pages 2536-2544, 2016. 2 +[56] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 3 +[57] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2 +[58] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, pages 6105-6114. PMLR, 2019. 1 + +[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proc. ICML, pages 10347-10357. PMLR, 2021. 7 +[60] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 5, 12 +[61] Michael Tschannen, Olivier Bachem, and Mario Lucic. Recent advances in autoencoder-based representation learning. arXiv preprint arXiv:1812.05069, 2018. 2 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NIPS, pages 5998-6008, 2017. 1 +[63] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proc. ICML, pages 1096-1103, 2008. 2 +[64] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 2 +[65] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Proc. NIPS, 32, 2019. 8 +[66] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proc. IEEE CVPR, pages 14668-14678, 2022. 3 +[67] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proc. IEEE CVPR, pages 2256-2265, 2021. 3 +[68] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proc. ECCV, pages 418-434, 2018. 8 +[69] Junyuan Xie, Linli Xu, and Enhong Chen. Image denoising and inpainting with deep neural networks. Proc. NIPS, 25, 2012. 2 +[70] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proc. IEEE CVPR, pages 9653-9663, 2022. 1, 2, 3 +[71] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Proc. NIPS, 32, 2019. 2 +[72] Richard Zhang, Phillip Isola, and Alexei A Efros. **Colorful image colorization. In Proc. ECCV, pages 649-666. Springer, 2016. 2** +[73] Yizhe Zhang, Zhe Gan, Kai Fan, Zhi Chen, Ricardo Henao, Dinghan Shen, and Lawrence Carin. Adversarial feature matching for text generation. In Proc. ICML, pages 4006-4015. PMLR, 2017. 3 + +[74] Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial networks. In Proc. ICLR, 2017. 3 +[75] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proc. IEEE CVPR, pages 18697–18709, 2022. 3 +[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proc. IEEE CVPR, pages 633-641, 2017. 8 +[77] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 7, 8 \ No newline at end of file diff --git a/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/images.zip b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..87644f8e0f5fd7181d52e7793cf50f0d1237ecbd --- /dev/null +++ b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8a2973910eab661aed9ce92d5d06291df10ba21d4e22636444800be39949a0e +size 489421 diff --git a/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/layout.json b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c4f1cc3fecffd9d38234c2fb9f2bb000b29b8349 --- /dev/null +++ b/2023/Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond/layout.json @@ -0,0 +1,8864 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 531, + 121 + ], + "type": "text", + "content": "Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 462, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 462, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 462, + 215 + ], + "type": "text", + "content": "Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang* Xiaoming Wei, Xiaolin Wei Meituan {feizhengcong, fanmingyuan, zhuli09, huangjunshi}@meituan.com {weixiaoming, weixiaolin02}@meituan.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 240, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 240, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 240, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 566 + ], + "type": "text", + "content": "Masked Auto-Encoder (MAE) pretraining methods randomly mask image patches and then train a vision Transformer to reconstruct the original pixels based on the unmasked patches. While they demonstrate impressive performance for downstream vision tasks, it generally requires a large amount of training resource. In this paper, we introduce a novel Generative Adversarial Networks alike framework, referred to as GAN-MAE, where a generator is used to generate the masked patches according to the remaining visible patches, and a discriminator is employed to predict whether the patch is synthesized by the generator. We believe this capacity of distinguishing whether the image patch is predicted or original is benefit to representation learning. Another key point lies in that the parameters of the vision Transformer backbone in the generator and discriminator are shared. Extensive experiments demonstrate that adversarial training of GAN-MAE framework is more efficient and accordingly outperforms the standard MAE given the same model size, training data, and computation resource. The gains are substantially robust for different model sizes and datasets, in particular, a ViT-B model trained with GAN-MAE for 200 epochs outperforms the MAE with 1600 epochs on fine-tuning top-1 accuracy of ImageNet-1k with much less FLOPs. Besides, our approach also works well at transferring downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 589, + 128, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 128, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 128, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 609, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 609, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 45, + 609, + 288, + 694 + ], + "type": "text", + "content": "In recent years, Transformer [62] has become the de facto standard architecture in computer vision, and has surpassed state-of-the-art Convolutional Neural Network (CNN) [31, 58] feature extractors in vision tasks through models such as the Vision Transformer [21]. Meanwhile, self-supervised learning (SSL) algorithms [12, 14, 27, 29] aims to learn transferable representation from unlabeled" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 258, + 525, + 417 + ], + "blocks": [ + { + "bbox": [ + 308, + 258, + 525, + 417 + ], + "lines": [ + { + "bbox": [ + 308, + 258, + 525, + 417 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 525, + 417 + ], + "type": "image", + "image_path": "987b48bb5a14bd972540c7a2439719338dbd2790d48f3b7bfbd7809f1f0e08bb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 424, + 547, + 470 + ], + "lines": [ + { + "bbox": [ + 305, + 424, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 547, + 470 + ], + "type": "text", + "content": "Figure 1. Performance comparison in different pre-training epochs for ImageNet-1K Fine-tuning top-1 accuracy. Compared to MAE trained for 1600 epochs, GAN-MAE achieves comparable accuracy with much less training time at 200 epochs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": "data by performing instance-level pretext tasks, and has been a long-standing target in the vision community. Particularly, masked image modeling (MIM) in SSL for vision transformers has shown remarkably impressive downstream performance in a wide variety of computer vision tasks [3, 28], attracting increasing attention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "MIM is a simple pretext task that first randomly masks some patches of an image, and then predicts the contents of the masked patches according to the remaining, using various reconstruction targets, e.g., visual tokens [3, 19], semantic features [1, 77] and raw pixels [28, 70]. Essentially, it learns the transferable representation by modeling the image structure itself as content prediction. While more effective than conventional pre-training, masked autoencoder modeling approaches still exist some issues: (i) reconstruction optimization with MSE loss leads to blurrier output images than the raw input, it would be better to use a more perceptual loss over pixels to guide the fine-grained seman" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 149, + 713 + ], + "type": "text", + "content": "*The corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24449" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "tic understanding and representation learning, leading to more plausible synthesized patches; (ii) inner dependency between masked patches is lacked [71], i.e., generation of masked image patches may lack the surrounding information. This situation becomes more serious when the image patch masking ratio is large. We alleviate this problem by introducing confident synthesized patches as complementary information during training; (iii) mask-reconstruction methods incur a substantial computation cost because the network only learns from part of the visible patches and misses the information of masked patches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 483 + ], + "type": "text", + "content": "In this paper, we propose a Generative Adversarial Networks-based pre-training framework, referred to as GAN-MAE, which contains two components: a generator model learns to reconstruct the masked patches according to visible patches in the encoder-decoder architecture and a discriminator model learns to distinguish real image patches from plausible but synthesized remains. Generally, given an image from training dataset, our method first randomly masks parts of patches and reconstructs them using the rest visible patches with a generator, which serves as a standard MAE model. Then we build the corrupt image as the combination of visible and synthesis patches, which is then fed into the discriminator to predict whether each patch is from raw image or synthesized results. In this manner, the discriminator provides a valid guiding for more delicate image patch modeling. Then, with the development of generator capacity, a key advantage of discriminative task is that it integrates the synthesized patches into corrupt images as complementary information, which fills the missing inner relationship between patches during pre-training. Moreover, we shared the parameters of vision transformer backbone in the generator and discriminator to promote memory reduction, training efficiency, as well as performance enhancement." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Our experiments follow the same architecture, settings, and pre-training recipe as MAE [28], and we find that the simple incorporation of a discriminator consistently outperforms MAE in variant models, e.g., ViT-S, ViT-B, and ViT-L, when fine-tuning for top-1 accuracy of ImageNet classification. We also conduct extensive ablation studies to validate the effectiveness of our core designs in backbone parameter sharing and advarial training. As pre-training with more epochs usually results in a better downstream performance, we argue that an important consideration for pre-training methods should be computation efficiency as well as absolute downstream performance. From this viewpoint, we also demonstrate that discrimination of pseudo-image patches forces GAN-MAE to train more efficiently than standard MAE. We further provide a comprehensive comparison with MAE in various epochs and various models and show our framework achieves consistently better performance. In particular, as presented in Figure 1, for the ViT-B model structure, our GAN-MAE achieves comparable" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "ble classification performance with only 200 pre-training epochs vs. standard MAE 1600 pre-training epochs. Furthermore, the GAN-MAE achieves 0.7 points improvement when pre-training 1600 epochs. Finally, we summarize our contribution as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 141, + 545, + 342 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 317, + 141, + 545, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 545, + 214 + ], + "type": "text", + "content": "- We propose a new and effective GAN-alike framework for visual representation self-supervised learning, which to our best knowledge is the first trial of integrating GAN idea into MAE framework. As a generic approach, we suggest that this framework can be easily applied on many other MIM-based tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 224, + 545, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 224, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 317, + 224, + 545, + 283 + ], + "type": "text", + "content": "- We introduce two core designs: shared weight for the main backbones of generator and discriminator, and an adversarial training process, both of which cost fewer amounts of computing resources while obtaining appreciable performance improvements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 294, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 545, + 342 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that compared with the original MAE, our method is more compute-efficient and results in better transfer representation learning on downstream tasks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 364, + 398, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 398, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 398, + 376 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 384, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 540 + ], + "type": "text", + "content": "Autoencoding. Autoencoder [6, 7, 35] is an unsupervised learning technique for neural networks that learns efficient representations by training the network to ignore signal noise. It includes an encoder that maps the original data to a low-dimensional latent embedding and a decoder that recovers the data from the latent embedding, with the goal of learning a compressed knowledge representation. The denoising autoencoder [63, 64] learns to reconstruct clean data points from a noisy version. Numerous efforts have been devoted for image denoising, such as masking pixels [11, 55, 64], inpainting [69], removing color channels [39, 72], and shuffling image patches [22, 54]. For a broader overview of denoising autoencoder, we refer to [5, 61]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "Masked Image Modeling. Masked language modeling [37,49,57], which generalizes well on language understanding and generation tasks, is the domain self-supervised approach in the field of NLP. Similarly, vision transformer [21, 41, 50] based masked image modeling (MIM) approaches [1,3,28,70,77] for computer vision tasks have also been developed. Generally, these MIM approaches first apply a mask to patches of an image, and then the masked patches are predicted given the visible patches. Feature representation learned through such within-image context prediction demonstrate strong transfer performance in downstream tasks. Recently, lots of works exploring MIM have been concurrently developed from different perspectives." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24450" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 547, + 201 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 547, + 201 + ], + "type": "image", + "image_path": "5b1b341053e9586bee822eb3f148d6d85160fd4bd229637788059f0a4d011d6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 547, + 246 + ], + "type": "text", + "content": "Figure 2. Overview of GAN-MAE framework, where a generator is used to predict the masked patches and a discriminator is employed to classify whether the patches are selected from the raw image or synthesized results. In particular, both of MAE encoder and discriminator are based on the vision Transformer backbone and share parameters for memory reduction and training efficiency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 339 + ], + "type": "text", + "content": "The works include framework design [13, 23, 28, 70], prediction targets [2, 19, 66, 77], and integration with vision-language representation learning [42, 43, 75]. Our work belongs to the first group and introduces a novel GAN framework that discriminates the reconstructed image patches in a standard MAE to learn deeper semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 356, + 289, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 289, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 289, + 608 + ], + "type": "text", + "content": "Generative Adversarial Networks. GANs [26] are effective at generating high-quality synthetic data. Usually, the generator generates an image, and the discriminator determines whether the input image is a real image or a generated image. Subsequently, many improvements based on the original GAN focused on speeding up the training of the network and improving the quality of the generated images [8, 52, 74]. These improvements also help GAN achieving a wider range of applications [45, 48, 67]. Methods based on GAN are also widely used in image-to-image translation [36], super-resolution [4, 40], style transfer [15], text generation [10, 24, 73], and representation learning [17, 25], to name a few. Particularly, [56] brings some designs of CNN architecture to stabilize the training of GAN framework, after that the discriminator can be directly used as feature extractor in downstream tasks. In contrast, our method proposes to integrate the GAN as assistant for the MIM task, which focuses on the study of better pattern for masking based self-supervised learning. Besides, our strategy of shared parameters provides a unified backbone for better vision representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 620, + 115, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 115, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 115, + 634 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "In this section, we introduce the GAN-MAE framework in details. At first, we briefly review the conventional masked autoencoder model in Sec. 3.1, and then describe the proposed generator-discriminator pre-training framework in Sec. 3.2. The architecture of our framework is presented in Figure 2. Finally, we suggest an adversarial" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 266, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 291 + ], + "type": "text", + "content": "training processes under the proposed method and discuss our framework in Sec. 3.3 and Sec. 3.4, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 298, + 392, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 392, + 310 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 392, + 310 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": "Masked Autoencoder (MAE) [28] is a self-supervised approach with a vision transformer encoder and a small transformer decoder, which randomly masks a large portion of input patches, and then reconstructs the masked patches according to the visible patches. Specifically, provided with an image " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " are the channel number, image height and image width respectively, MAE partitions " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "N = \\frac{H \\times W}{P^2}" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " non-overlapping patches with patch size " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": ". In this way, the image is transformed into a sequence of patches " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "X = \\{x^1, \\ldots, x^N\\}" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " with each element " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "x^k \\in \\mathbb{R}^{P^2 \\times C}" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": ". Then, we sample a random set of patch index " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " in uniform distribution, and split the image patches " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " into masked patch set " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "X_m = \\{x^k | k \\in M\\}" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " and visible patch set " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "X_v = \\{x^k | k \\notin M\\}" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": ". During training, the MAE encoder inputs " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "X_v" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " to achieve the latent representations " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "H_v" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": ". Then, the MAE decoder attempts to reconstruct " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "X_m" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " with the input of interpolating [mask] token embedding into the sequence of latent representations " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "H_v" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": " according to the index set " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": ", and outputs the reconstructed patches " + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\tilde{X}_m" + }, + { + "bbox": [ + 304, + 316, + 547, + 570 + ], + "type": "text", + "content": ". Finally, MAE optimizes the mean-squared error reconstruction loss on the masked patches as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 578, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 578, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 342, + 578, + 545, + 605 + ], + "type": "interline_equation", + "content": "L _ {m a e} (X, M, \\theta_ {m a e}) = \\sum_ {k \\in M} | | \\tilde {x} ^ {k} - x ^ {k} | | _ {2} ^ {2}, \\tag {1}", + "image_path": "efe5849c8e8af44bd37ca54341da0604d165c9882b2d0fa2ab172b2009730bf8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 614, + 525, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 614, + 525, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 614, + 525, + 627 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 614, + 525, + 627 + ], + "type": "inline_equation", + "content": "\\theta_{mae}" + }, + { + "bbox": [ + 306, + 614, + 525, + 627 + ], + "type": "text", + "content": " represents the parameters of MAE model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 437, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 437, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 437, + 646 + ], + "type": "text", + "content": "3.2. When MAE meet GAN" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "In this section, we describe the proposed GAN-MAE framework, as presented in Figure 2. Generally, our framework consists of two parts, a generator " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " to reconstruct the masked image patches and a discriminator " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " to predict the realness of image patches." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "24451" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "Image Patch Generator. Identical to the standard MAE, the generator follows the encoder-decoder paradigm and is trained to perform masked-image reconstruction task. Given the partitioned image patches " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", the generator randomly masks some image patches and encodes the remaining visible patches " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " into a sequence of contextualized vector representations " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", based on which the masked patches are reconstructed as " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\hat{X}_m" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ". Please refer to Sec. 3.1 for more details. In general, the generation process can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 201, + 287, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 201, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 122, + 201, + 287, + 213 + ], + "type": "interline_equation", + "content": "M \\sim \\operatorname {U n i f o r m} (1, N), \\tag {2}", + "image_path": "858b12a2543321132718f7fcb776863aac8610c0da92af393f2181b987bf4bb6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 216, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 216, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 136, + 216, + 287, + 228 + ], + "type": "interline_equation", + "content": "H _ {v} = f _ {e} \\left(X _ {v}, M\\right), \\tag {3}", + "image_path": "ff85282ad7195323b80be4e627dbdc2f404352423463635d9b0e193e9eb959dd.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 232, + 287, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 232, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 232, + 287, + 246 + ], + "type": "interline_equation", + "content": "\\tilde {X} _ {m} = f _ {d} \\left(H _ {v}, M\\right), \\tag {4}", + "image_path": "8e8fafde1ed225272c8021d38df573a5c41313fc45cc0895cf792fa86f0f3c57.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "type": "text", + "content": " is the number of image patches. " + }, + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "type": "inline_equation", + "content": "f_{e}(\\cdot)" + }, + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "type": "inline_equation", + "content": "f_{d}(\\cdot)" + }, + { + "bbox": [ + 47, + 255, + 287, + 278 + ], + "type": "text", + "content": " denote the encoder and decoder in a conventional MAE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": "Image Patch Discriminator. For a patch index " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": " and corrupted image sequence " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "\\overline{X} = \\{X_v, \\tilde{X}_m\\}" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": ", the discriminator predicts whether the patch token " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "x^k" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": " is real or synthesized as binary classification task. Specifically, we create the corrupted image " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "\\overline{X}" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": " by maintaining the visible patches " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "X_v" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": " in raw image " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": " and replacing the masked patches with generator predicted result " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "\\tilde{X}_m" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "X_v" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "\\tilde{X}_m" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": " are absolute complement of set to each other, i.e., " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "X_v \\cup \\tilde{X}_m = \\overline{X}" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "inline_equation", + "content": "X_v \\cap \\tilde{X}_m = \\emptyset" + }, + { + "bbox": [ + 46, + 293, + 289, + 413 + ], + "type": "text", + "content": ". Formally, the task of discriminator model can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 421, + 287, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 421, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 79, + 421, + 287, + 435 + ], + "type": "interline_equation", + "content": "D (\\bar {X}, k) = p _ {\\text {d i s c}} \\left(y ^ {k} | \\bar {X}, k\\right), \\text {f o r} k \\in [ 1, N ]. \\tag {5}", + "image_path": "d113c13a8e923153f9d68b81d3b3c8b1635dbbf21b433aa23e6f5cd0b9bece5c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 444, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 289, + 515 + ], + "type": "text", + "content": "Let the ground-truth classification label sequence " + }, + { + "bbox": [ + 47, + 444, + 289, + 515 + ], + "type": "inline_equation", + "content": "Y = \\{y^{1},\\ldots ,y^{N}\\}" + }, + { + "bbox": [ + 47, + 444, + 289, + 515 + ], + "type": "text", + "content": " with each element " + }, + { + "bbox": [ + 47, + 444, + 289, + 515 + ], + "type": "inline_equation", + "content": "y_{k}\\in \\{0,1\\}" + }, + { + "bbox": [ + 47, + 444, + 289, + 515 + ], + "type": "text", + "content": ", where 0 and 1 denote the corresponding image patch is reconstructed from generator or comes from the original image, respectively. The training objective of the discriminator can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 522, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 522, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 61, + 522, + 287, + 572 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {d i s c} (\\bar {X}, \\theta_ {d i s c}) = \\sum_ {k = 1} ^ {N} - y ^ {k} \\log D (\\bar {X}, k) \\tag {6} \\\\ - \\left(1 - y ^ {k}\\right) \\log \\left(1 - D (\\overline {{X}}, k)\\right). \\\\ \\end{array}", + "image_path": "0bb7bfbae6e54f669723455b6de60296c4e8fe81b64c0853e1f354beeb22b97d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "type": "text", + "content": "Particularly, though the corrupted image is partially unreal, we suggest that this discriminative task can still be benefit to the learning of feature representation with the plausible corrupted image from a well-trained generator, as presented in experiments. Hitherto, due to the GAN-alike strategy for MAE task, we name our framework as GAN-MAE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 658, + 149, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 149, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 149, + 672 + ], + "type": "text", + "content": "3.3. Training Scheme" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "We explore the training strategy of our proposed GANMAE in this section. Particularly, to improve synthesis result, we augment the L-2 reconstruction loss with a percep" + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 310, + 88, + 545, + 320 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 503, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 503, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 503, + 87 + ], + "type": "text", + "content": "Algorithm 1: Adversarial training for GAN-MAE" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": "Data: Training data " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " total epoch number " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " GAN-MAE model with generator parameters " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{mae}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " and discriminator parameters " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{disc}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " \n1 share weights between generator and discriminator backbones; \n2 while " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "n_e < N_e" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " do \n3 for " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "x^{i}\\in \\mathcal{D}_{train}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " do \n4 " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " generator training; \n5 sample masking set " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "M^i" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " and mask image " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " . \n6 predict masked image patches " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\tilde{x}_m^i" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " . \n7 compute loss " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "L_{gen}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " . \n8 loss backward for updating " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{mae}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " . \n9 dicriminator training; \n10 construct " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\overline{x}^i" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\tilde{x}_m^i" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " . \n11 comput loss " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "L_{disc}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " . \n12 loss backward for updating " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{disc}" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " . \n13 end \n14 " + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "inline_equation", + "content": "n_e + = 1" + }, + { + "bbox": [ + 310, + 88, + 545, + 320 + ], + "type": "text", + "content": " . \nend" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "algorithm" + }, + { + "bbox": [ + 305, + 344, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 344, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 545, + 369 + ], + "type": "text", + "content": "tual loss that aims to differentiate the real patches and reconstructed patches as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 375, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 375, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 317, + 375, + 545, + 389 + ], + "type": "interline_equation", + "content": "L _ {a d v} \\left(X, \\theta_ {m a e}\\right) = \\log D \\left(X _ {v}\\right) + \\log \\left(1 - D \\left(\\tilde {X} _ {m}\\right)\\right). \\tag {7}", + "image_path": "1ff1905729dd41ed286dfae0afe9efddad355ce5cf411c8de83dafd1284b9a12.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 396, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 431 + ], + "type": "text", + "content": "Please note that " + }, + { + "bbox": [ + 305, + 396, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\tilde{X}_m" + }, + { + "bbox": [ + 305, + 396, + 545, + 431 + ], + "type": "text", + "content": " is the reconstructed patches by MAE. Therefore, the final objective for the generator comes to minimize the combined loss as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 438, + 545, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 545, + 452 + ], + "type": "interline_equation", + "content": "L _ {\\text {g e n}} (X, \\theta_ {\\text {m a e}}) = L _ {\\text {m a e}} (X, \\theta_ {\\text {m a e}}) + \\gamma L _ {\\text {a d v}} (X, \\theta_ {\\text {m a e}}), \\tag {8}", + "image_path": "3218a2fc0cdc0eb107b02a32a0de81c30396a57af6030bd690d263d303c294d6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 458, + 527, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 458, + 527, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 458, + 527, + 470 + ], + "type": "text", + "content": "where we compute the adaptive weight " + }, + { + "bbox": [ + 306, + 458, + 527, + 470 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 458, + 527, + 470 + ], + "type": "text", + "content": " according to:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 386, + 475, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 475, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 386, + 475, + 545, + 502 + ], + "type": "interline_equation", + "content": "\\gamma = \\frac {\\nabla \\left[ L _ {m a e} \\right]}{\\nabla \\left[ L _ {a d v} \\right] + \\delta}, \\tag {9}", + "image_path": "42eccb250a1f10c175320108b76454f1d884f39325057e585a12bb533a9c3dec.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 508, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\nabla[\\cdot]" + }, + { + "bbox": [ + 305, + 508, + 545, + 568 + ], + "type": "text", + "content": " denotes the gradients of different loss function w.r.t. the parameters of last layer in network, and " + }, + { + "bbox": [ + 305, + 508, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\delta = 1e - 6" + }, + { + "bbox": [ + 305, + 508, + 545, + 568 + ], + "type": "text", + "content": " is used for numerical stability. Intuitively, the integration of " + }, + { + "bbox": [ + 305, + 508, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 508, + 545, + 568 + ], + "type": "text", + "content": " adaptively balances the contributions of two loss functions to the gradients of parameters." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "type": "text", + "content": "Based on the aforementioned strategy, the final training algorithm can be formulated in Algorithm 1. Specifically, at each epoch, we conduct the iteration in two steps: (i) train only the generator with " + }, + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "type": "inline_equation", + "content": "L_{gen}" + }, + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "type": "text", + "content": "; (ii) Train the discriminator with " + }, + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "type": "inline_equation", + "content": "L_{disc}" + }, + { + "bbox": [ + 304, + 568, + 545, + 663 + ], + "type": "text", + "content": ". During training, we observe that shared weights of generator and discriminator backbones can usually stabilize the training procedure and bring extra-gain in down-stream tasks." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 671, + 378, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 378, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 378, + 682 + ], + "type": "text", + "content": "3.4. Discussion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Theoretically, the integration of discriminator can be considered as a high-level perceptual loss, which forces the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24452" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "generator to learn better feature representation for the plausible synthesis of masked patches. As the quality of synthesized patches improved, we claim that the introduction of corrupted images, serving as complementary information, provides plausible inner dependency between full image patches for representation learning. Furthermore, with the shared parameters of backbones in generator and discriminator, the GAN-alike framework can be considered as a type of multi-task learning, leading to even better result as presented in experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 203, + 128, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 128, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 128, + 217 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 223, + 170, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 170, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 170, + 236 + ], + "type": "text", + "content": "4.1. Datasets and Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "type": "text", + "content": "In the experiments, we pre-train our GAN-MAE model on the ImageNet-1k [16] and evaluate the performance in end-to-end fine-tuning (FT) pattern for the task of classification, semantic segmentation, object detection and instance segmentation. The evaluation metric of classification is the top-1 validation accuracy on " + }, + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "type": "text", + "content": " cropped input images. The input image is partitioned into " + }, + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "type": "text", + "content": " patches and each patch is of size " + }, + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 240, + 289, + 529 + ], + "type": "text", + "content": ". Following the setting of MAE, we only use the standard random cropping and horizontal flipping for data augmentation. To validate the effectiveness of GAN-MAE framework, the used ViT architecture and most hyper-parameters are exactly the same to [28, 60], i.e., ViT-S (12 transformer blocks with dimension 384), ViT-B (12 transformer blocks with dimension 768), and ViT-L (24 transformer blocks with dimension 1024). All version of ViT models are trained with 4096 batch size on 8 V-100 32GB GPUs. We adopt dynamic token masking with the masked positions decided on-the-fly. We use AdamW [38] optimizer and cosine schedule [51] with warm up for model training. The learning rate is annealed according to the cosine schedule. Unless stated otherwise, results are evaluated on the dev set. Please refer to appendix A for more training implementation and hyperparameter values for different backbone in details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 536, + 178, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 178, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 178, + 548 + ], + "type": "text", + "content": "4.2. Analysis of GAN-MAE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "type": "text", + "content": "We analyze our GAN-MAE framework by proposing and evaluating several extensions to the model. Unless stated otherwise, all these experiments use the same model size as ViT-B and training dataset ImageNet-1K." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "Parameter Sharing of Backbone. In this work, we propose to improve the efficiency of the pre-training by sharing the parameters of the backbone vision transformer between the generator encoder and discriminator. One cause is that the generator and discriminator utilize the same network architecture, and all of the transformer weights can be tied. However, we can release the weight sharing and train the generator and discriminator independently. In between, we" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 337, + 112, + 517, + 203 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 105 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 105 + ], + "type": "text", + "content": "Table 1. Effect of parameter sharing in GAN-MAE framework. Results demonstrate that shared parameters for backbone benefits both memory cost and performance improvements." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 112, + 517, + 203 + ], + "lines": [ + { + "bbox": [ + 337, + 112, + 517, + 203 + ], + "spans": [ + { + "bbox": [ + 337, + 112, + 517, + 203 + ], + "type": "table", + "html": "
ModelsEpochMask ratioFT
Generator80075%83.9
Discriminator80075%84.2
Shared80075%84.3
Generator160075%84.4
Discriminator160075%84.4
Shared160075%84.6
", + "image_path": "6a5a4aa56ca93c7647ea65ac3ef1fc57de67b2c84beaf94e1363ca59a667acc3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 233, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 365 + ], + "type": "text", + "content": "can adopt both the parameters of generator and discriminator for downstream learning. The experimental results are shown in Table 1, where we employ the adversarial training with the same training epochs. We can see that, generally, the fine-tuning top-1 accuracy of shared weight outperforms the independent generator and discriminator conspicuously, in particular 0.4 point improvements for the generator when pre-training 800 epochs. We hypothesize that GAN-MAE framework benefits from both mask-then-reconstruct and pseudo-patch classification tasks, which can incorporate the visual semantic and consistency understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 546, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 582 + ], + "type": "text", + "content": "It is also surprising that under the GAN process without weight sharing, the discriminator with learning of patch classification can lead to better performance than generator. We suspect that the fine-grained patch classification is more delicate while harder than image-level classification. Thus, we further tried to add an image-level contrastive objective. For this task, we input " + }, + { + "bbox": [ + 304, + 365, + 546, + 582 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 365, + 546, + 582 + ], + "type": "text", + "content": " of the input image unchanged rather than noising them with the generator. We then added a prediction head to the model that predicted if the entire input image was corrupted or not. However, the result didn't improve the final accuracy. In conclusion, we believe that the design of discriminative task, e.g., getting closer to downstream and be more difficult, is an important exploration direction in the future. Another interesting finding is that with the pre-training epochs increase, the independent generator performance boosts, we believe that it is caused by the adversarial training, where the discriminator becomes an effective guiding for generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 597, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 545, + 633 + ], + "type": "text", + "content": "Training Schemes. We analyze the effect of proposed adversarial training scheme on GAN-MAE with shared backbones. Two training variants are considered as following:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 642, + 547, + 714 + ], + "type": "text", + "content": "- Two-stage Training. It is natural to consider to remove the discriminator guiding signal during generator training, which leads to the disentangled optimizing process. Specifically, at each epoch, we do the following steps: train the generator only with " + }, + { + "bbox": [ + 317, + 642, + 547, + 714 + ], + "type": "inline_equation", + "content": "L_{mae}" + }, + { + "bbox": [ + 317, + 642, + 547, + 714 + ], + "type": "text", + "content": " and train the discriminator only with " + }, + { + "bbox": [ + 317, + 642, + 547, + 714 + ], + "type": "inline_equation", + "content": "L_{disc}" + }, + { + "bbox": [ + 317, + 642, + 547, + 714 + ], + "type": "text", + "content": ". The difference with ad" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24453" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 90, + 278, + 186 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 251, + 82 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 251, + 82 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 251, + 82 + ], + "type": "text", + "content": "Table 2. Effect of different training schemes." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 90, + 278, + 186 + ], + "lines": [ + { + "bbox": [ + 57, + 90, + 278, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 90, + 278, + 186 + ], + "type": "table", + "html": "
ModelsEpochMask ratioFTGPU Time
Two-stage30075%82.094.3h
Combined30075%82.290.9h
Adversarial30075%82.2118.8h
Two-stage80075%84.0252.2h
Combined80075%84.1240.5h
Adversarial80075%84.3317.5h
", + "image_path": "a940b1c4efb60eed18677476545a3f3aeb632b977b72ae2a6d6522f912157a6a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 83, + 237, + 252, + 352 + ], + "blocks": [ + { + "bbox": [ + 47, + 205, + 287, + 229 + ], + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 229 + ], + "type": "text", + "content": "Table 3. Effect of different masking ratio in GAN-MAE framework with different pre-training epochs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 237, + 252, + 352 + ], + "lines": [ + { + "bbox": [ + 83, + 237, + 252, + 352 + ], + "spans": [ + { + "bbox": [ + 83, + 237, + 252, + 352 + ], + "type": "table", + "html": "
ModelsEpochMask ratioFT
MAE80075%83.4
GAN-MAE80070%84.3
GAN-MAE80075%84.3
GAN-MAE80080%84.0
MAE160075%83.6
GAN-MAE160070%84.5
GAN-MAE160075%84.6
GAN-MAE160080%84.4
", + "image_path": "cb35fbef553714a39eedd9041de61f651a1d84275d9ffb74fc29acb8544674b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 381, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 381, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 66, + 381, + 287, + 407 + ], + "type": "text", + "content": "versarial training lies in that the training loss of the generator in the first step is changed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 415, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 415, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 59, + 415, + 287, + 475 + ], + "type": "text", + "content": "- Combined Training. Instead of iterative in total dataset, we can jointly trains the generator and discriminator at each step. That is, for each image " + }, + { + "bbox": [ + 59, + 415, + 287, + 475 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 59, + 415, + 287, + 475 + ], + "type": "text", + "content": " in the training dataset " + }, + { + "bbox": [ + 59, + 415, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 59, + 415, + 287, + 475 + ], + "type": "text", + "content": ", we can directly minimize the combined loss as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 485, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 485, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 116, + 485, + 287, + 498 + ], + "type": "interline_equation", + "content": "L _ {g e n} \\left(\\theta_ {m a e}\\right) + \\lambda L _ {d i s c} \\left(\\theta_ {d i s c}\\right). \\tag {10}", + "image_path": "2bfea347baf0cdfa122345b5d5686eac7f1353ba5fd820efeceaa258233dceb5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 507, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 507, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 66, + 507, + 287, + 544 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 66, + 507, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 66, + 507, + 287, + 544 + ], + "type": "text", + "content": ", the weight for the discriminator objective in the loss to 2.0, as we searched for " + }, + { + "bbox": [ + 66, + 507, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 66, + 507, + 287, + 544 + ], + "type": "text", + "content": " out of [1,2,5,10] in early experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 552, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 588 + ], + "type": "text", + "content": "Note that regardless of the form, the nature of sequential training, i.e., the corrupt image built from generator will be fed into discriminator, is not changed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 588, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 288, + 696 + ], + "type": "text", + "content": "The evaluation results for classification are listed in Table 2. GPU time means pre-training time (hours) on 8 V100 32GB GPUs environment. As we can see, the combined training shows a superior training time while adversarial training slows down as the training procedures of generator and discriminator are isolated. On the other hand, the performance of adversarial training is not better than combined training in the early stages; when pre-training epochs come to 800, the benefits of adversarial training appear." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 332, + 102, + 520, + 163 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 94 + ], + "type": "text", + "content": "Table 4. Comparison of computation resource usage during self-supervised pre-training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 102, + 520, + 163 + ], + "lines": [ + { + "bbox": [ + 332, + 102, + 520, + 163 + ], + "spans": [ + { + "bbox": [ + 332, + 102, + 520, + 163 + ], + "type": "table", + "html": "
BackboneModelsFLOPsParams
ViT-BMAE9.4e9111.654M
ViT-BGAN-MAE2.6e10111.656M
ViT-LMAE2.0e10329.238M
ViT-LGAN-MAE8.0e10329.240M
", + "image_path": "7cc48d3c06a3cf4be7e604e17c100f42a4b013a3bda596f313ce35ee27177317.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "text", + "content": "Masking Ratio. Table 3 shows the influence of masking ratio in MAE-GAN under different pre-training epochs. The optimal ratio for MAE-GAN is identical to " + }, + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "text", + "content": ", showing a obviously better classification performance compared with same masking ratio in previous work [28]. Meanwhile, we present several reconstructed images from MAE and GAN-MAE. Although the MIM model can infer missing patches as different yet plausible outputs when mask ratio comes to large, the generator of GAN-MAE can predict the images with more realistic and fine-grained details, e.g., the outline of a mountain peak in first case, which we believe is relative to the learning of useful representation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 352, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 616 + ], + "type": "text", + "content": "Computation Resource. In terms of pre-training cost, we conduct a computing resource comparison with the baseline MAE on different vision transformer backbones. The results are shown in Table 4. We first point out that GAN-MAE slows down the training process as the discriminator integration, e.g. 317.5h vs. 127.7h for ViT-B model in 800 epochs. Moreover, we also choose to measure computation usage in terms of floating point operations (FLOPs) as it is a metric agnostic to the particular hardware, low-level optimizations, etc. Note that an \"operation\" is a mathematical operation, not a machine instruction and thop package is used to compute FLOPs in practice. As expected, GAN-MAE employs " + }, + { + "bbox": [ + 304, + 352, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\sim 3" + }, + { + "bbox": [ + 304, + 352, + 545, + 616 + ], + "type": "text", + "content": " extra theoretical computation cost. Besides, benefits from the weight sharing of backbone, the GAN-MAE incorporates less than " + }, + { + "bbox": [ + 304, + 352, + 545, + 616 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 352, + 545, + 616 + ], + "type": "text", + "content": " addition parameter, which helps to reduce the running memory usage, and provides a convenience to set large batch size. Last but not least, we want to state that although our GAN-MAE method incorporates additional computation resources during training at each epoch, considering the superiority of reduction of epoch number and classification performance improvements, it is fully acceptable and exploration valuable." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 623, + 501, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 501, + 636 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 501, + 636 + ], + "type": "text", + "content": "4.3. ImageNet Classification Comparison" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "content": "We compare our GAN-MAE methods with previous state-of-the-art works on the ImageNet-1K classification task. Table 5 reports the top-1 validation accuracy for finetuning results. We can find that compared to the supervised models, trained from scratch, all of the self-supervised pre-training methods achieve significant improvement, sug" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24454" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 73, + 292, + 135 + ], + "blocks": [ + { + "bbox": [ + 53, + 73, + 292, + 135 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 292, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 292, + 135 + ], + "type": "image", + "image_path": "4beaed6f5b05eb57a4cf78ad30018dfcf433231e01ff5836ca4a2607b588199c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 54, + 137, + 291, + 201 + ], + "blocks": [ + { + "bbox": [ + 54, + 137, + 291, + 201 + ], + "lines": [ + { + "bbox": [ + 54, + 137, + 291, + 201 + ], + "spans": [ + { + "bbox": [ + 54, + 137, + 291, + 201 + ], + "type": "image", + "image_path": "2ddcfbc8bfd3b604331fddb3c4f1a123e6120b15a9d44aa1479eaf1c844a3291.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 54, + 203, + 292, + 266 + ], + "blocks": [ + { + "bbox": [ + 54, + 203, + 292, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 203, + 292, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 203, + 292, + 266 + ], + "type": "image", + "image_path": "0baaf630ac8e81e514ec56c7064ca6ba5cc70acc6661a703f7772ad9089199d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 278, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 323 + ], + "type": "text", + "content": "Figure 3. Qualitative analysis for patch reconstruction. Example results are from ImageNet validation set. For each tuple, we show the raw image, masked image, MAE reconstructed image, and our proposed GAN-MAE reconstructed image from left to right. We can see that the reconstructed images from GAN-MAE are significantly clearer than MAE, which we believe benefits the fine-grained visual semantic understanding." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 301, + 74, + 538, + 135 + ], + "blocks": [ + { + "bbox": [ + 301, + 74, + 538, + 135 + ], + "lines": [ + { + "bbox": [ + 301, + 74, + 538, + 135 + ], + "spans": [ + { + "bbox": [ + 301, + 74, + 538, + 135 + ], + "type": "image", + "image_path": "4c9af92e7918f7eb148ab1ff90d6cf8cd8b5e0e4605d30763e0806b85e0505a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 301, + 137, + 538, + 201 + ], + "blocks": [ + { + "bbox": [ + 301, + 137, + 538, + 201 + ], + "lines": [ + { + "bbox": [ + 301, + 137, + 538, + 201 + ], + "spans": [ + { + "bbox": [ + 301, + 137, + 538, + 201 + ], + "type": "image", + "image_path": "f912ca23a28c07aa3115c05807a0fde46ca7000513169496914f65f13d17c0ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 203, + 538, + 266 + ], + "blocks": [ + { + "bbox": [ + 301, + 203, + 538, + 266 + ], + "lines": [ + { + "bbox": [ + 301, + 203, + 538, + 266 + ], + "spans": [ + { + "bbox": [ + 301, + 203, + 538, + 266 + ], + "type": "image", + "image_path": "f31ed1cbd7180b371977d634615cb036bf4fe7423bdba14aadc5bf2ae492488c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 122, + 365, + 473, + 517 + ], + "blocks": [ + { + "bbox": [ + 46, + 334, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 546, + 357 + ], + "type": "text", + "content": "Table 5. End-to-end fine-tuning on ImageNet-1K. We report the fine-tuning top-1 accuracy for classification in different vision transformer architectures and results show that GAN-MAE outperforms previous self-supervised methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 122, + 365, + 473, + 517 + ], + "lines": [ + { + "bbox": [ + 122, + 365, + 473, + 517 + ], + "spans": [ + { + "bbox": [ + 122, + 365, + 473, + 517 + ], + "type": "table", + "html": "
ModelPre-train dataPre-train epochsViT-SViT-BViT-L
Supervised [59]IN1K w/ labels30079.781.882.6
DINO [9]IN1K80081.582.8-
MoCo v3 [14]IN1K30081.483.284.1
BEiT [3]IN1K+DALLE80081.783.285.2
MSN [1]IN1K600-83.4-
iBOT [77]IN1K80082.384.084.8
BootMAE [20]IN1K800-84.285.9
MAE [28]IN1K800-83.485.4
MAE [28]IN1K1600-83.685.9
GAN-MAEIN1K30082.284.085.6
GAN-MAEIN1K80082.484.386.1
", + "image_path": "ebd504e45770b877482d4bc6dea87ca4db3559e882231c18a96d5913873d6219.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "type": "text", + "content": "gesting the effectiveness of pre-training. We further compare our GAN-MAE framework with prior popular self-supervised pre-training models. We can see that the proposed GAN-MAE achieves the best fine-tuning performance either based on the ViT-B or based on the ViT-L architectures. For example, compared with the recent work MAE [28], our GAN-MAE in ViT-L network achieves " + }, + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "type": "inline_equation", + "content": "86.1\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "type": "text", + "content": " with 0.7 point improvement when pre-trained on 800 epochs. More encouragingly, for all ViT backbone sizes, GAN-MAE mostly outperforms the previous self-supervised methods. These results suggest that the incorporation of a discriminator scheme could have consistently benefits for various scale ViT models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "content": "In addition, we present a comprehensive comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "with MAE under different pre-training epochs for the ViT-B model. We plot the results in Figure 1. We can see that our GAN-MAE approach consistently performs better than MAE. It is worth mentioning that the proposed GAN-MAE at 300 epochs achieves " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " accuracy, which is already better than MAE pre-trained at 1600 epochs. This demonstrates that our approach is more efficient to achieve comparable performance. What's more, no additional speed and parameter cost during inference. Besides, we can confirm that, similar to other MIM-based self-supervised training, the accuracy of GAN-MAE also improves steadily as the pre-training steps increase. Particularly, our GAN-MAE on ViT-B achieves " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "84.6\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " top-1 accuracy at 1600 epochs, which is almost " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " higher than that of MAE." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24455" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 112, + 493, + 167 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 6. Robustness Evaluation on the four ImageNet-variants: ImageNet-C, ImageNet-A, ImageNet-R, and ImageNet-Sketch. Except for ImageNet-C which is measured in terms of mean Corruption Error (mCE), top-1 accuracy is used as the remaining evaluation metric. For simplicity, we denoted IN-C, IN-A, IN-R, In-Sketch correspondingly." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 112, + 493, + 167 + ], + "lines": [ + { + "bbox": [ + 102, + 112, + 493, + 167 + ], + "spans": [ + { + "bbox": [ + 102, + 112, + 493, + 167 + ], + "type": "table", + "html": "
ModelIN-C (mCE ↓)IN-A (top-1 ↑)IN-R (top-1 ↑)IN-Sketch (top-1 ↑)
Supervised [53]42.535.848.736.0
MAE [28]51.735.948.334.5
GAN-MAE49.536.849.635.9
", + "image_path": "c401f223325a792518d8afe63a7c24ce46de0efb43e57e1281c0b6770cf20856.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 225, + 274, + 328 + ], + "blocks": [ + { + "bbox": [ + 47, + 194, + 287, + 217 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 287, + 217 + ], + "type": "text", + "content": "Table 7. Semantic segmentation comparison on the ADE20K dataset for mIoU (%) metric with the ViT-B backbone." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 225, + 274, + 328 + ], + "lines": [ + { + "bbox": [ + 61, + 225, + 274, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 274, + 328 + ], + "type": "table", + "html": "
ModelsPre-train dataEpochsmIoU
Supervised [28]IN1K w/ labels30047.4
MoCo v3 [14]IN1K30047.3
BEiT [3]IN1K+DALLE80047.1
MAE [28]IN1K80047.6
MAE [28]IN1K160048.1
BootMAE [20]IN1K80049.1
GAN-MAEIN1K80049.5
", + "image_path": "1cde421a53494c8b04f017b2598223e3e883427b75bb77b74d864bc7c614a02b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 53, + 385, + 282, + 500 + ], + "blocks": [ + { + "bbox": [ + 47, + 354, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 376 + ], + "type": "text", + "content": "Table 8. COCO object detection and segmentation using Mask R-CNN framework with ViT-B backbone." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 385, + 282, + 500 + ], + "lines": [ + { + "bbox": [ + 53, + 385, + 282, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 282, + 500 + ], + "type": "table", + "html": "
ModelsPre-train dataAP-boxAP-mask
Supervised [28]IN1K w/ labels44.139.8
MoCo v3 [14]IN1K44.940.4
BEiT [3]IN1K+DALLE46.341.1
MSN [1]IN1K46.641.5
iBOT [77]IN1K47.342.2
MAE [28]IN1K47.242.0
BootMAE [20]IN1K48.543.4
GAN-MAEIN1K49.043.8
", + "image_path": "4d7aef036b2d3846b58fabad473fe007dcf92223c62756a3823f047577c3530d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 536, + 159, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 159, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 159, + 548 + ], + "type": "text", + "content": "4.4. Downstream Tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "Semantic Segmentation. We compare our GAN-MAE with supervised as well as state-of-the-art self-supervised models on the widely used dataset ADE20K [76] for semantic segmentation. Specifically, we use the UperNet framework [68] in the experiments. We train Upernet for 160K iterations with batch size set as 64 and report the results in Table 7. The evaluation metric is mean Intersection of Union (mIoU) averaged over all semantic categories and the single-scale test results are reported. Importantly, we can see that the proposed GAN-MAE gets superior performance than all the other baselines in the same configuration, further validating the effectiveness of adversarial training with a reconstructed patch discriminator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "text", + "content": "Object Detection and Segmentation. We also perform object detection and instance segmentation, compared with other popular self-supervised methods and the supervised model, on the COCO dataset [47]. In practice, we choose the Mask R-CNN [30] framework and adopt FPNs [46] to scale the feature map into different sizes as introduced in [44]. The performance is tested on the COCO validation set, following the previous work [18]. The results are listed in Table 8 in terms of box AP metric for object detection and mask AP metric for instance segmentation. Importantly, we can observe that our GAN-MAE model achieves " + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "inline_equation", + "content": "49.0\\%" + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "text", + "content": " for object detection and " + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "inline_equation", + "content": "43.8\\%" + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "text", + "content": " for segmentation, surpassing the previous state-of-the-art BootMAE by " + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 196, + 547, + 365 + ], + "type": "text", + "content": " point, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 381, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 547, + 514 + ], + "type": "text", + "content": "Classification Robutness. Similar to [28], we further evaluate the robustness of classification performance on the four ImageNet variants, i.e., ImageNet-C [33], ImageNet-A [34], ImageNet-R [32], and ImageNet-Sketch [65], which are common benchmarks to evaluate robustness for perturbations. Table 6 demonstrates the robustness comparison with GAN-MAE and MAE using the ViT-B backbone, as well as previous supervised SoTA models. The results illustrate that GAN-MAE outperforms the MAE baseline consistently on all robustness datasets, indicating that the promising of adversarial training in representation learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 525, + 379, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 379, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 379, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we have proposed a new self-supervised GAN-alike framework for visual representation learning, where a generator is used to predict masked image patches according to the visible patches and a discriminator is employed to predict whether the patch is from raw image or generated by generator. The key idea is adversarial training a shared vision Transformer to distinguish the input patches from high-quality negative samples, which we believe is beneficial for the understanding of visual conception. It works well while incorporating no much addition parameter. More encouragingly, compared to standard masked image modeling, our GAN-MAE is more compute-efficient, as fewer pre-training epochs result in a better performance on downstream tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24456" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. arXiv preprint arXiv:2204.07141, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In Proc. ICML, pages 1298-1312. PMLR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 223 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In Proc. ICLR, 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "text", + "content": "[4] Sefi Bell-Kligler, Assaf Shocher, and Michal Irani. Blind super-resolution kernel estimation using an internal-gan. Proc. NIPS, 32, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "text", + "content": "[5] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 301, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 287, + 334 + ], + "type": "text", + "content": "[6] Yoshua Bengio et al. Learning deep architectures for ai. Foundations and trends® in Machine Learning, 2(1):1-127, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "type": "text", + "content": "[7] Yoshua Bengio, Pascal Lamblin, Dan Popovici, and Hugo Larochelle. Greedy layer-wise training of deep networks. Proc. NIPS, 19, 2006. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "text", + "content": "[8] David Berthelot, Thomas Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proc. IEEE CVPR, pages 9650-9660, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 287, + 491 + ], + "type": "text", + "content": "[10] Liquun Chen, Shuyang Dai, Chenyang Tao, Haichao Zhang, Zhe Gan, Dinghan Shen, Yizhe Zhang, Guoyin Wang, Ruiyi Zhang, and Lawrence Carin. Adversarial text generation via feature-mover's distance. Proc. NIPS, 31, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 287, + 535 + ], + "type": "text", + "content": "[11] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proc. ICML, pages 1691-1703. PMLR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 579 + ], + "type": "text", + "content": "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proc. ICML, pages 1597-1607. PMLR, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 580, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 287, + 635 + ], + "type": "text", + "content": "[13] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 635, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 669 + ], + "type": "text", + "content": "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proc. IEEE ICCV, pages 9640-9649, 2021. 1, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 713 + ], + "type": "text", + "content": "[15] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE CVPR, pages 8789-8797, 2018. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE CVPR, pages 248–255, 2009. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[17] Jeff Donahue, Philipp Krahenbuhl, and Trevor Darrell. Adversarial feature learning. arXiv preprint arXiv:1605.09782, 2016.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 143, + 545, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 197 + ], + "type": "text", + "content": "[18] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proc. IEEE CVPR, pages 12124-12134, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 199, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 199, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 545, + 255 + ], + "type": "text", + "content": "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "type": "text", + "content": "[20] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. arXiv preprint arXiv:2207.07116, 2022. 7, 8, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 314, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 314, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 314, + 545, + 380 + ], + "type": "text", + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. ICLR, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 381, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 426 + ], + "type": "text", + "content": "[22] Bo Du, Wei Xiong, Jia Wu, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Stacked convolutional denoising auto-encoders for feature representation. IEEE transactions on cybernetics, 47(4):1017-1027, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 428, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 545, + 472 + ], + "type": "text", + "content": "[23] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "text", + "content": "[24] William Fedus, Ian Goodfellow, and Andrew M Dai. Maskgan: Better text generation via filling in the .. In Proc. ICLR, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 508, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 553 + ], + "type": "text", + "content": "[25] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "type": "text", + "content": "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 600, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 666 + ], + "type": "text", + "content": "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Proc. NIPS, 33:21271-21284, 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE CVPR, pages 16000-16009, 2022. 1, 2, 3, 5, 6, 7, 8, 12" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24457" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE CVPR, pages 9729-9738, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 119, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 119, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 119, + 287, + 151 + ], + "type": "text", + "content": "[30] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. IEEE ICCV, pages 2961-2969, 2017. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 288, + 186 + ], + "type": "text", + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE CVPR, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 189, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 189, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 189, + 287, + 243 + ], + "type": "text", + "content": "[32] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proc. IEEE CVPR, pages 8340-8349, 2021. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 245, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 278 + ], + "type": "text", + "content": "[33] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 280, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 312 + ], + "type": "text", + "content": "[34] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proc. IEEE CVPR, pages 15262-15271, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 315, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 315, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 315, + 287, + 346 + ], + "type": "text", + "content": "[35] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[36] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE CVPR, pages 1125-1134, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 395, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 395, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 395, + 287, + 437 + ], + "type": "text", + "content": "[37] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL, pages 4171–4186, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 441, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 473 + ], + "type": "text", + "content": "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "type": "text", + "content": "[39] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In Proc. ECCV, pages 577-593. Springer, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 586 + ], + "type": "text", + "content": "[40] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proc. IEEE CVPR, pages 4681-4690, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[41] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In Proc. IEEE CVPR, pages 7287-7296, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[42] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proc. AAAI, volume 34, pages 11336-11344, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[43] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[44] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[45] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In Proc. IEEE CVPR, pages 9392-9400, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[46] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proc. IEEE CVPR, pages 2117-2125, 2017. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755. Springer, 2014. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "text", + "content": "[48] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, and Jing Liao. Pd-gan: Probabilistic diverse gan for image inpainting. In Proc. IEEE CVPR, pages 9371-9381, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 351 + ], + "type": "text", + "content": "[49] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[50] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. IEEE ICCV, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "text", + "content": "[51] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 433, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 475 + ], + "type": "text", + "content": "[52] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In Proc. IEEE CVPR, pages 2794-2802, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 478, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 520 + ], + "type": "text", + "content": "[53] Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. In Proc. IEEE CVPR, pages 12042-12051, 2022. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[54] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proc. ECCV, pages 69-84. Springer, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "type": "text", + "content": "[55] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proc. IEEE CVPR, pages 2536-2544, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[56] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[57] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[58] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, pages 6105-6114. PMLR, 2019. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24458" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proc. ICML, pages 10347-10357. PMLR, 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[60] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 5, 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[61] Michael Tschannen, Olivier Bachem, and Mario Lucic. Recent advances in autoencoder-based representation learning. arXiv preprint arXiv:1812.05069, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NIPS, pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "text", + "content": "[63] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proc. ICML, pages 1096-1103, 2008. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "text", + "content": "[64] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "text", + "content": "[65] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Proc. NIPS, 32, 2019. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "text", + "content": "[66] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proc. IEEE CVPR, pages 14668-14678, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "type": "text", + "content": "[67] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proc. IEEE CVPR, pages 2256-2265, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "text", + "content": "[68] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proc. ECCV, pages 418-434, 2018. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "text", + "content": "[69] Junyuan Xie, Linli Xu, and Enhong Chen. Image denoising and inpainting with deep neural networks. Proc. NIPS, 25, 2012. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[70] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proc. IEEE CVPR, pages 9653-9663, 2022. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[71] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Proc. NIPS, 32, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "type": "text", + "content": "[72] Richard Zhang, Phillip Isola, and Alexei A Efros. **Colorful image colorization. In Proc. ECCV, pages 649-666. Springer, 2016. 2**" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[73] Yizhe Zhang, Zhe Gan, Kai Fan, Zhi Chen, Ricardo Henao, Dinghan Shen, and Lawrence Carin. Adversarial feature matching for text generation. In Proc. ICML, pages 4006-4015. PMLR, 2017. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 251 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[74] Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial networks. In Proc. ICLR, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[75] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proc. IEEE CVPR, pages 18697–18709, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proc. IEEE CVPR, pages 633-641, 2017. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[77] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24459" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_content_list.json b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..199a862f6641060b4ef60e662d321278e46d9a21 --- /dev/null +++ b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_content_list.json @@ -0,0 +1,1544 @@ +[ + { + "type": "text", + "text": "Masked Autoencoders Enable Efficient Knowledge Distillers", + "text_level": 1, + "bbox": [ + 181, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yutong Bai1 Zeyu Wang2 Junfei Xiao1 Chen Wei1 Huiyu Wang1 Alan Yuille1 Yuyin Zhou2 Cihang Xie2 1Johns Hopkins University 2 University of California, Santa Cruz ytongbai@gmail.com", + "bbox": [ + 222, + 179, + 746, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, i.e., forward propagate inputs through the first few layers, for obtaining intermediate feature maps.", + "bbox": [ + 73, + 316, + 468, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves $84.0\\%$ ImageNet top-1 accuracy, outperforming the baseline of directly distilling a fine-tuned ViT-L by $1.2\\%$ . More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with $95\\%$ masking ratio where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of $83.6\\%$ ; surprisingly, it can still secure $82.4\\%$ top-1 ImageNet accuracy by aggressively training with just FOUR visible patches (98% masking ratio). The code and models are publicly available at https://github.com/UCSC-VLAA/DMAE.", + "bbox": [ + 75, + 484, + 472, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following the success in the natural language processing [10, 27], the Transformer architecture is showing tremendous potentials in computer vision [2, 11, 25, 26, 33], especially when they are pre-trained with a huge amount of unlabelled data [3] with self-supervised learning techniques [1, 14]. Masked image modeling, which trains models to predict the masked signals (either as raw pixels or as semantic tokens) of the input image, stands as one of the most powerful ways for feature pre-training. With the most re", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cent representative work in this direction, masked autoencoder (MAE) [13], we are now able to efficiently and effectively pre-train high-capacity Vision Transformers (ViTs) with strong feature representations, leading to state-of-the-art solutions for a wide range of downstream visual tasks.", + "bbox": [ + 496, + 287, + 890, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we are interested in applying knowledge distillation [17], which is one of the most popular model compression techniques, to transfer the knowledge from these strong but cumbersome ViTs into smaller ones. In contrast to prior knowledge distillation works [17, 21, 34], the teacher considered here is a pre-trained model whose predictions do not necessarily reveal the fine-grained relationship between categories; therefore, typical solutions like aligning the soft/hard logits between the teacher model and the student model may no longer remain effective. Moreover, after distilling the pre-trained teacher model, these student models need an extra round of fine-tuning to adapt to downstream tasks. These factors turn distilling pretrained models seemingly a less favorable design choice in terms of both performance and computational cost.", + "bbox": [ + 496, + 366, + 892, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nonetheless, surprisingly, we find by building upon MAE, the whole distillation framework can efficiently yield high-performance student models. There are two key designs. Firstly, we follow MAE to let the encoder exclusively operate on a small visible subset of patches and to employ a lightweight decoder for pixel reconstruction. Whereas rather than using the \"luxury\" setups in MAE, we show aggressively simplifying pre-training from 1600 epochs to 100 epochs and pushing masking ratio from $75\\%$ to $95\\%$ suffice to distill strong student models. Secondly, instead of aligning logits, we alternatively seek to match the intermediate feature representation; this enables the cumbersome teacher model to only forward propagate inputs through the first few layers, therefore, reducing computations. We note applying L1 norm for distance measure is essential recipe for ensuring a successful intermediate feature alignment.", + "bbox": [ + 496, + 595, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We name this distilling MAE framework as DMAE. Compared to the traditional knowledge distillation framework where the teacher is a fine-tuned model, DMAE is more efficient and can train much stronger student models", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24256", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/996c5217375d9066e1d46cbc84c713e221d9ebe32bc23f3cfa725364e2451fd1.jpg", + "image_caption": [ + "Figure 1. Illustration of the distillation process in DMAE. There are two key designs. Firstly, following MAE, we hereby only take visible patches as inputs and aims to reconstruct the masked ones. Secondly, knowledge distillation is achieved by aligning the intermediate features between the teacher model and the student model. Note the gray blocks denote the dropped high-level layers of the teacher model during distillation." + ], + "image_footnote": [], + "bbox": [ + 107, + 97, + 861, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "at different capacities. For example, by setting ViT-B as the student model, while the baseline of distilling a fine-tuned ViT-L achieves $82.8\\%$ top-1 ImageNet accuracy, DMAE substantially boosts the performance to $84.0\\%$ $(+1.2\\%)$ top-1 ImageNet accuracy, at a even lower training cost (i.e., 195 GPU hours vs. 208 GPU hours, see Table 9). More intriguingly, we found that DMAE allows for robust training with extremely highly masked images—even with TEN visible patches (i.e., $95\\%$ masking ratio), ViT-B can competitively attain a top-1 ImageNet accuracy of $83.6\\%$ ; this masking ratio can further be aggressively pushed to $98\\%$ (FOUR visible patches) where DMAE still help ViT-B secure $82.4\\%$ top-1 ImageNet accuracy. We hope this work can benefit future research on efficiently unleashing the power of pretrain models.", + "bbox": [ + 75, + 377, + 472, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 618, + 218, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge distillation (KD) is a popular model compression technique that allows models to achieve both strong performances of large models and fast inference speed of small models. The first and seminal KD approach, proposed in [17], transfers the \"dark knowledge\" via minimizing the KL divergence between the soft logits of the teacher model and that of the student model. From then on, many advanced KD methods have been developed, which can be categorized into two branches: logits distillation [8, 12, 21, 30, 34, 35] and intermediate representation distillation [15, 16, 18, 23, 24]. Our DMAE belongs to the second branch, as it minimizes the distance between latent features of the teacher model and those of the student model.", + "bbox": [ + 75, + 643, + 468, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first feature-based distillation method is Fit-Nets [23]. In addition to aligning logits, FitNets requires the student model to learn an intermediate representation that is", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "predictive of the intermediate representations of the teacher network. Heo et al. [15] re-investigates the design of feature distillation and develops a novel KD method to create a synergy among various aspects, including teacher transform, student transform, distillation feature position, and distance function. CRD [24] incorporates contrastive learning into KD to capture correlations and higher-order output dependencies. Unlike these existing works, our DMAE is the first to consider applying KD to extra information from self-supervised pre-trained models.", + "bbox": [ + 496, + 377, + 893, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked image modeling (MIM) helps models to acquire meaningful representations by reconstructing masked images. The pioneering works are built on denoising autoencoders [28] and context encoders [22]. Following the success of BERT in natural language [10], and also with the recent trend of adopting Transformer [27] to computer vision [11], there have emerged a set of promising works on applying MIM for self-supervised visual pre-training. BEiT [3] first successfully adopts MIM to ViT pre-training by learning to predict visual tokens. MaskFeat [29] finds that learning to reconstruct HOG features enables effective visual representation learning. SimMIM [31] and MAE [13] both propose to directly reconstruct the pixel values of the masked image patches. Our work is built on MAE and finds that MAE enables the whole distillation framework to be efficient and effective.", + "bbox": [ + 496, + 534, + 895, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 791, + 609, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Masked Autoencoders", + "text_level": 1, + "bbox": [ + 500, + 816, + 705, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is built upon MAE, a powerful autoencoder-based MIM approach. Specifically, the MAE encoder first projects unmasked patches to a latent space, which are then fed into the MAE decoder to help predict pixel values of", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24257", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "masked patches. The core elements in MAE include:", + "bbox": [ + 76, + 90, + 428, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masking. MAE operates on image tokens, i.e., the image needs to be divided into non-overlapping patches. A random small subset of those patches will be kept for the MAE encoder, and the rest will be set as the predicting target of the MAE decoder. Typically, a high masking ratio (e.g., $75\\%$ ) is applied, preventing models from taking shortcuts (e.g., simply extrapolating missing pixels based on neighbors) in representation learning.", + "bbox": [ + 75, + 111, + 468, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAE encoder. The MAE encoder is a standard ViT architecture except that it only operates on those unmasked patches. This design largely reduces the computation cost of encoders.", + "bbox": [ + 75, + 236, + 468, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAE decoder. In addition to the encoded features of un-masked patches (from MAE encoder), the MAE decoder receives mask tokens as input, a learned vector shared across all missing positions. The mask token is only used during pre-training, allowing independent decoder design. Particularly, MAE adopts a lightweight decoder for saving computations.", + "bbox": [ + 75, + 301, + 468, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstruction. Different from BEiT [3] or MaskFeat [29], MAE directly reconstructs image pixel values. The simple mean squared error is applied to masked tokens for calculating loss.", + "bbox": [ + 76, + 412, + 468, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that, other than distillation-related operations, the whole pre-training and fine-tuning process in this paper exactly follow the default setup in MAE, unless specifically mentioned. Interestingly, compared to MAE, our DMAE robustly enables a much more efficient pre-training setup, e.g., 100 (vs. 1600) training epochs and $95\\%$ (vs. $75\\%$ ) masking ratio.", + "bbox": [ + 75, + 473, + 468, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Knowledge Distillation", + "text_level": 1, + "bbox": [ + 76, + 589, + 287, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAE demonstrates extraordinary capabilities in learning high-capacity models efficiently and effectively. In this work, we seek to combine knowledge distillation with the MAE framework, to efficiently acquire small and fast models with similar performance as those powerful yet cumbersome models. The most straightforward approach is directly applying existing knowledge distillation methods, like the one proposed in DeiT [25], to a fine-tuned MAE model. However, we empirically find that this approach hardly brings in improvements. In addition, this approach fails to leverage the special designs in MAE for reducing computations, e.g., only feeding a small portion of the input image to the encoder. To this end, we hereby study an alternative solution: directly applying knowledge distillation at the pre-training stage.", + "bbox": [ + 75, + 613, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since there are no categorical labels in MAE pretraining, distilling logits can hardly help learn semantically meaningful representations. We, therefore, resort to distilling the intermediate features. This idea is first developed in", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FitNets [23], and inspired a set of followups for advancing knowledge distillation [15, 16, 18, 24]. Concretely, we first extract the features from the specific layers of the student model; after feeding such features into a small project head, the outputs will be asked to mimic the features from the corresponding layers of the teacher model. In practice, the projection head is implemented by simple fully connected layer, which addresses the possible feature dimension mismatch between teacher models and student models.", + "bbox": [ + 496, + 90, + 890, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, let $\\mathbf{x} \\in \\mathbb{R}^{3HW \\times 1}$ be the input pixel RGB values and $\\mathbf{y} \\in \\mathbb{R}^{3HW \\times 1}$ be the predicted pixel values, where $H$ denotes image height and $W$ denotes image width. The MAE reconstruction loss $L_{MAE}$ can be written as", + "bbox": [ + 498, + 226, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M A E} = \\frac {1}{\\Omega (\\mathbf {x} _ {M})} \\sum_ {i \\in M} \\left(\\mathbf {y} _ {i} - \\mathbf {x} _ {i}\\right) ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 292, + 890, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M$ denotes the set of masked pixels, $\\Omega(.)$ is the number of elements, and $i$ is the pixel index.", + "bbox": [ + 496, + 337, + 890, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{z}_l^S, \\mathbf{z}_l^T \\in \\mathbb{R}^{LC \\times 1}$ be the features extracted from the $l$ th layer of the student model and the teacher model, respectively, where $L$ denotes the patch numbers, and $C$ denotes the channel dimension. We use $\\sigma()$ to denote the projection network function. Our feature alignment distillation loss $L_{Dist}$ can be written as", + "bbox": [ + 496, + 366, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D i s t} = \\sum_ {l} \\frac {1}{\\Omega (\\mathbf {z} _ {l} ^ {T})} \\sum_ {i} \\left\\| \\sigma (\\mathbf {z} _ {l} ^ {S}) _ {i} - \\mathbf {z} _ {l, i} ^ {T} \\right\\| _ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 463, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The final loss used in pre-training is a weighted summation of MAE reconstruction loss $L_{MAE}$ and the feature alignment distillation loss $L_{Dist}$ , controlled by the hyperparameter $\\alpha$ :", + "bbox": [ + 496, + 505, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {M A E} + \\alpha \\times L _ {D i s t}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 574, + 890, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The framework of DMAE is summarized in Figure 1. Following MAE, DMAE also takes masked inputs and performs the pretext task of masked image modeling. Besides, the corresponding features are aligned between the teacher model and the student model. It is worthy of highlighting that DMAE is an efficient knowledge distiller: 1) it only operates on a tiny subset of visible patches i.e., a high masking ratio is applied; and 2) aligning intermediate layer features reduce the computation cost of (cumbersome) teacher model. In the next section, we extensively compare our method with three baselines: the original MAE without any distillation, DeiT-style distillation, and feature alignment distillation in the supervised setting.", + "bbox": [ + 496, + 599, + 890, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 806, + 633, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 832, + 718, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following MAE [13], we first perform self-supervised pre-training on ImageNet-1k [9]. Unless otherwise mentioned, the teacher models are public checkpoints released", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24258", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/43ea7bc82a4a166b501a40331fab9ba8e57b11cc47599155ea66b017c132e324.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# of LayersLayer LocationStudent Aligned Layer IndexImageNet Top-1 Acc (%)
SingleBottom382.6
Middle683.6
984.0
Top1283.4
MultipleMiddle+Top6+1284.2
", + "bbox": [ + 202, + 88, + 764, + 178 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the official MAE implementations1. For pre-training, we train all models using AdamW optimizer [20], with a base learning rate of 1.5e-4, weight decay of 0.05, and optimizer momentum $\\beta_{1},\\beta_{2} = 0.9,0.95$ . We use a total batch size of 4096, and pre-train models for 100 epochs with a warmup epoch of 20 and a cosine learning rate decay schedule. We by default use the masking ratio of $75\\%$ ; while the ablation study shows that our method can robustly tackle extremely high masking ratios. After pre-training, the teacher model is dropped, and we exactly follow the default setups in MAE to fine-tune the student model on ImageNet.", + "bbox": [ + 75, + 250, + 470, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For feature alignments, we choose to align the features from the $\\frac{3}{4}$ depth of both the student model and the teacher model, which we find delivers decent results for all model sizes tested. For example, with a 24-layer ViT-L as the teacher model and a 12-layer ViT-B as the student model, features from the 9th layer of ViT-B are aligned with the features from the 18th layer of ViT-L. We set $\\alpha = 1$ in Eq. 3 to balance the tradeoff between MAE reconstruction loss $L_{MAE}$ and the feature alignment distillation loss $L_{Dist}$ in pre-training.", + "bbox": [ + 75, + 417, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Analysis", + "text_level": 1, + "bbox": [ + 76, + 578, + 181, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first provide a detailed analysis of how to set distillation-related parameters in DMAE. Specifically, we set the teacher model as an MAE pre-trained ViT-L (from MAE official GitHub repository, attaining $85.9\\%$ top-1 ImageNet accuracy after fine-tuning), and set the student model as a randomly initialized ViT-B. We analyze the following six factors:", + "bbox": [ + 75, + 601, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where to align. In Table 1 we first check the effect of feature alignment location on model performance. We observe that shallower features are less favored: e.g., the 3rd layer alignment under-performs all other settings. We speculate this is due to the learning process of ViTs—images are much noisier and less semantic than texts, ViTs will first group the raw pixels in the bottom layers (closer to the input), which is harder to transfer. While features from $\\frac{3}{4}$ depth (i.e., the 9th layer in ViT-B) achieve the best performance, a simple rule-of-thumb which we find fits all model scales in our experiments. We adopt this design choice in all", + "bbox": [ + 75, + 710, + 470, + 878 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/db6481f8f95fa365381e9ebb9f41763c5f99d555b24ed77b5a1fa4dab2dd5ec3.jpg", + "table_caption": [ + "Table 1. The effects of feature alignment location. We hereby test with 5 different layer locations, where top layers refer to those closer to network outputs. For single layer feature alignment, features from the $\\frac{3}{4}$ depth of the model leads to the best ImageNet top-1 accuracy. This performance is even comparable to the setup of aligning multiple layers." + ], + "table_footnote": [], + "table_body": "
Teacher Layer (relative position)ImageNet Top-1 Acc (%)
Middle84.0
Top83.3
Bottom82.1
", + "bbox": [ + 501, + 248, + 887, + 310 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a87dd2a6b89c1ec1c741c81138ab5c542731cf7efe968355b92cb2b960f4a05c.jpg", + "table_caption": [ + "Table 2. The analysis of aligning order on ImageNet classification top-1 accuracy $(\\%)$" + ], + "table_footnote": [], + "table_body": "
Projection HeadImageNet Top-1 Acc (%)
Linear84.0
2-layer MLP84.0
3-layer MLP83.8
", + "bbox": [ + 550, + 357, + 839, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3. Projection Head. A simple fully-connected layer works the best. We choose this as the default setting.", + "bbox": [ + 500, + 422, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "other experiments. It is also worth mentioning that simply aligning multiple layers has no clear advantage over aligning features from $\\frac{3}{4}$ depth (84.2% vs. 84.0%); we, therefore, stick to the $\\frac{3}{4}$ depth setting, which is more efficient.", + "bbox": [ + 498, + 479, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aligning order. We next test the importance of alignment ordering on model performance. Specifically, by fixing the layer location in the student model (i.e., the middle layer in our experiment), we then align it to different layers of the teacher model. As shown in Table 2, we observe that when the aligned layers are in the same relative position (e.g., middle to middle), the student model can achieve the best performance.", + "bbox": [ + 498, + 546, + 892, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masking ratio. MAE reveals that the masking ratio in masked image modeling could be surprisingly high (75%). The hypothesis is that by learning to reason about the gestalt of the missing objects and scenes, which cannot be done by extending lines or textures because of the high masking ratio, the model is also learning useful representations. Interestingly, we find that when combining MAE and knowledge distillation, an even much higher masking ratio is possible, as shown in Figure 2.", + "bbox": [ + 496, + 672, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, it is interesting to note that, compared to the typical $75\\%$ masking ratio setting, further raising the masking ratio to $90\\%$ comes at no performance drop, i.e., both attain $84.0\\%$ top-1 ImageNet accuracy. Next, even with an extremely large masking ratio like $98\\%$ (only FOUR visible patches), DMAE still beats the 100-epoch MAE baseline", + "bbox": [ + 496, + 810, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/facebookresearch/mae", + "bbox": [ + 94, + 886, + 321, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24259", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/339f5a89158840705c7d8d3a45b2f9dc22625001b6cb03b561f002fbfacc943d.jpg", + "image_caption": [ + "Figure 2. DMAE allows an extremely high masking ratio. From left to right, we increase the masking ratio from the basic $75\\%$ to the extreme $99\\%$ . We note that our DMAE competitively attains $83.6\\%$ top-1 ImageNet accuracy with $95\\%$ masking ratio (TEN visible patches), and still secures $82.4\\%$ top-1 ImageNet accuracy even by learning with FOUR visible patches ( $98\\%$ masking ratio)." + ], + "image_footnote": [], + "bbox": [ + 197, + 88, + 769, + 292 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/77e982003bc1cc19a8238aec6c8a734dfa933cc67a72043ae82d09969b6ec5e5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Decoder DepthImageNet Top-1 Acc (%)
283.7
483.8
884.0
", + "bbox": [ + 130, + 363, + 411, + 425 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/123de3ee749463ce539e6d2c9d8fa426b79223ac4c9f750f79f75c7fcabd9529.jpg", + "table_caption": [ + "Table 4. Decoder Depth. A deeper decoder (slightly) improves the pre-trained representation quality." + ], + "table_footnote": [], + "table_body": "
Loss DesignImageNet Top-1 Acc (%)
Loss ChoiceL1 with α = 184.0
L2 with α = 183.3
Loss RatioL1 with α = 0.583.9
L1 with α = 184.0
L1 with α = 284.0
L1 with α = 484.0
", + "bbox": [ + 78, + 473, + 464, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 5. Loss Function. From the first block, we can observe that L1 distance yields significantly higher performance than L2 distance. From the second block, we can observe that the hyperparameter $\\alpha$ has little influence on the representation quality of DMAE.", + "bbox": [ + 75, + 580, + 468, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that uses a masking ratio of $75\\%$ (second row in Table 6), by a non-trivial-margin $(82.4\\% \\text{ vs. } 81.6\\%)$ . These results suggest that, with the assistance of distilled knowledge from the teacher model, the student model can make better use of visible patches, even at a very limited amount, for representation learning.", + "bbox": [ + 75, + 680, + 468, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Projection head. The goal of the proposed feature alignment distillation is to encourage the student model to learn features that are predictive of features from a stronger teacher model. To that end, a small projection head is employed on features from the student model, to 1) project them onto a space of the same dimension as the hidden dimension of the teacher model, and 2) provide extra flexibility for feature alignment. We ablate the choice of this pro", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "jection network, as shown in Table 3. We can observe that applying a simple fully connected layer already performs the best among other choices.", + "bbox": [ + 496, + 366, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoder depth. In Table 4 we analyze the effect of decoder depth. Similar to MAE [13], the final performance gets (slightly) increased with a deeper decoder. We choose a decoder depth of 8 as the default setting as in [13]. Note that a decoder depth of 2 is also a competitive choice—compared to an 8-depth decoder, it significantly reduces the computation cost while only marginally sacrificing the accuracy by $0.3\\%$ .", + "bbox": [ + 496, + 415, + 892, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss designs. Table 5 ablates the loss design. While SimMIM [31] shows that L1 distance and L2 distance lead to similar performance, ours suggests that L1 distance exhibits a clear advantage over L2 distance, i.e., $+0.7\\%$ improvement. Furthermore, we note DMAE is quite robust to the specific value of the hyperparameter $\\alpha$ , which controls the relative importance of the distillation loss over the reconstruction loss. Based on these results, we choose L1 in Eq. 2 for distance measure and set $\\alpha = 1$ in Eq. 3 for the rest experiments.", + "bbox": [ + 496, + 541, + 890, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with Baselines.", + "text_level": 1, + "bbox": [ + 500, + 702, + 750, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 6, we compare the performance of our DMAE with various baselines:", + "bbox": [ + 500, + 726, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MAE. The MAE baselines are presented in the first block of Table 6. If MAE is also asked to pre-train for only 100 epochs, DMAE can substantially outperform this baseline by $2.4\\%$ (from $81.6\\%$ to $84.0\\%$ ). When comparing to a much stronger but more computationally expensive MAE baseline with 1600 pre-training epoch, we note DMAE still beats it by $0.4\\%$ .", + "bbox": [ + 496, + 760, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervised model. The second block in Table 6 demonstrates the effectiveness of DMAE compared with models", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24260", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/89df4888d473e4be63d904611b3032a0a40463dc1d3111c830f1a640119e6b9d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre-training epochsSupervised training / fine-tuning epochsImageNet Top-1 Acc (%)
MAE-B10010081.6
MAE-B160010083.6
DeiT-B-10076.8
DeiT-B-30081.8
DeiT-B-Soft Distillation-10077.5
DeiT-B-Hard Distillation-10078.3
CRD [24]-10081.9
SRRL [32]-10082.2
Dear-KD [6]-10082.4
Supervised Feature Alignment-10082.8
DMAE-B10010084.0
", + "bbox": [ + 210, + 88, + 754, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "trained under the supervision of categorical labels, which requires a much longer training time, i.e., $+2.2\\%$ compared to the DeiT 300 epochs supervised training.", + "bbox": [ + 75, + 335, + 468, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Other distillation strategies. We next compare DMAE with other distillation methods. We consider DeiT-style logit-based distillation [25], CRD [24], SRRL [32], DearKD [6], and feature alignment distillation. Note that for these baselines, one significant difference from DMAE is that the student model here directly distill knowledge from a supervisely fine-tuned teacher model. Moreover, to make these baselines more competitive, the teacher models will first be MAE pre-trained and then fine-tuned on ImageNet-1k.", + "bbox": [ + 75, + 382, + 468, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in the third block of Table 6. Firstly, we can observe that the DeiT-style logit-based distillation, either soft or hard, even hurts the student models' performance. This phenomenon potentially suggests that such a distillation strategy may not fit teacher models of ViT architectures. For other baselines, we note that feature alignment distillation performs the best; but this is still worse than DMAE (82.8% vs. 84.0%), indicating the importance and effectiveness of distilling knowledge from a pre-trained teacher model.", + "bbox": [ + 75, + 534, + 470, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Scaling to Different Model Sizes", + "text_level": 1, + "bbox": [ + 76, + 693, + 359, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test DMAE with different model sizes, listed in Table 7. For a fair comparison, both methods only pre-train models for 100 epochs. DMAE shows consistent improvement compared to MAE across different model sizes. With only one middle layer feature alignment, DMAE brings an additional improvement of $+2.4\\%$ with ViT-B, $+2.7\\%$ with ViT-Small, and $+3.4\\%$ with ViT-Tiny. In addition, we are interested in the following two cases:", + "bbox": [ + 75, + 715, + 468, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Same teacher model, different student model. As shown in the first two lines in Table 7, we find that when using ViT-L as the teacher model, both ViT-B and ViT-S benefit from the distillation, demonstrating a clear advantage over", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/edcbe723c60497a602e2a30e5d2fdf886b20c1b98dc4329fbc4f873e32d9c914.jpg", + "table_caption": [ + "Table 6. DMAE shows stronger performance than all three kinds of baselines: MAE, supervised model, and other existing advanced distillation strategies." + ], + "table_footnote": [], + "table_body": "
Student ModelTeacher ModelImageNet Top-1 Acc (%)
MAEDMAE
BaseLarge81.684.0 (+2.4)
SmallLarge77.480.1 (+2.7)
Base77.479.3 (+1.9)
TinyBase66.670.0 (+3.4)
", + "bbox": [ + 501, + 332, + 890, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 7. Across different model sizes, DMAE shows consistent improvements compared with MAE.", + "bbox": [ + 500, + 428, + 890, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the MAE baseline. We argue that the ability to effectively generalize to cases where an even smaller student model is desirable, especially for those computation-constrained real-world applications.", + "bbox": [ + 498, + 483, + 890, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Same student model, different teacher model. As shown in Table 7, with a ViT-S as the student model, enlarging the teacher model from ViT-B to ViT-L further boosts the accuracy by $0.8\\%$ (from $79.3\\%$ to $80.1\\%$ ). This result suggests that DMAE can effectively distill knowledge from the teacher models at different scales.", + "bbox": [ + 498, + 549, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Limited Training Data", + "text_level": 1, + "bbox": [ + 500, + 650, + 710, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In certain real-world applications, data could be hard to acquire because of high data collection and labeling costs or due to privacy concerns. Leveraging models pre-trained on large-scale unlabeled datasets for fine-tuning when only a small dataset of downstream tasks is available becomes a promising solution. Here we seek to test the potential of our DMAE in this data-scarce scenario. We strictly follow [5] to sample $1\\%$ or $10\\%$ of the labeled ILSVRC-12 training datasets in a class-balanced way. We set MAE pre-trained ViT-L as the teacher model and a randomly initialized ViT-B as the student model. In addition, we compare DMAE with three kinds of baselines described in Section 4.3: MAE, supervised model, and other distillation strategies, and similarly, set the teacher model to be a ViT-L that first pre-trained with MAE and then fine", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24261", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0fd9267419bad2f9262a939e2f2796a892e1a0f32c5ca6bf59929b8da88798a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre-training epochsSupervised training / fine-tuning epochsIN-1% Top-1 Acc(%)IN-10% Top-1 Acc(%)
MAE-B10010033.965.0
MAE-B160010049.672.8
DeiT-B-100--
DeiT-B-Soft Distillation160010036.066.4
DeiT-B-Hard Distillation160010037.367.3
Supervised Feature Alignment160010034.267.6
DMAE-B10010050.373.4
", + "bbox": [ + 161, + 88, + 803, + 219 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b0cdbeb78b69908832427830b0243c2f7740137b9be5c2054fda07d4054bf5a9.jpg", + "table_caption": [ + "Table 8. DMAE demonstrates much stronger performance than all other baselines when training data is limited. Note that the DeiT-B baseline is unable to converge because of data insufficiency." + ], + "table_footnote": [], + "table_body": "
ModelTraining Cost (GPU Hours)ImageNet Top-1 Acc(%)
Pre-trainingFine-tuningOverall
MAE-B-100-epoch78h112h190h81.6
MAE-B-1600-epoch1248h112h1360h83.6
DeiT-B-Soft Distillation-213h213h77.5
DeiT-B-Hard Distillation-213h213h78.3
Supervised Feature Alignment-208h208h82.8
DMAE-B83h112h195h84.0
", + "bbox": [ + 207, + 266, + 759, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tuned on ImageNet-1k. Note that since DMAE has full access to the $100\\%$ ImageNet dataset (without labels) during pre-training, to ensure a fair and competitive comparison, we initialize all the baselines as the 1600-epoch MAE pretrained model on ImageNet.", + "bbox": [ + 75, + 457, + 468, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8 shows that DMAE largely surpasses all other baselines. For example, when only $10\\%$ ImageNet data is available for supervised training or fine-tuning, DMAE outperforms the MAE pre-trained baseline by $8.4\\%$ (i.e., $73.4\\%$ vs. $65.0\\%$ ). DMAE also significantly outperforms other distillation strategies, with an improvement ranging from $5.8\\%$ to $7.0\\%$ . We note this accuracy gap is even larger when only $1\\%$ ImageNet is available, demonstrating the data efficiency of DMAE.", + "bbox": [ + 75, + 539, + 470, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Generalization to Other Methods", + "text_level": 1, + "bbox": [ + 76, + 704, + 369, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, we provide preliminary results of integrating DMAE into other self-supervised training frameworks, including DINO [4] and MoCo-V3 [7]. Unlike MAE, which belongs to masked image modeling, DINO and MoCo-V3 are contrastive learning-based methods. Still, as shown in Table 10, without further hyperparameter tuning, DMAE effectively shows non-trivial improvements on top of both DINO (+1.3%) and MoCo-v3 (+1.4%), demonstrating the potential of distilling pre-trained models (rather than fine-tuned models as in most existing knowledge distillation frameworks).", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3e6ce96032e311ef624f179ff3b9b4c71b4f1d406fb8a17a6875fedb026d833b.jpg", + "table_caption": [ + "Table 9. Computational cost comparisons among DMAE and other baselines. The training cost is measured by A5000 GPU hours. We note the proposed DMAE maintains a similar (or even cheaper) training cost than others, while achieving much higher top-1 ImageNet accuracy." + ], + "table_footnote": [], + "table_body": "
w/o DMAEw/ DMAE
DINO80.982.2 (+1.3)
MoCo-V381.182.5 (+1.4)
", + "bbox": [ + 565, + 454, + 823, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Generalization to Downstream tasks", + "text_level": 1, + "bbox": [ + 500, + 592, + 815, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following ViT-Det [19], we conduct downstream fin-tuning on the COCO dataset, where MAE/DMAE pretrained ViT-Base model is adopted as the plain backbone of Mask-RCNN. We train the model on the train2017 split and evaluate it on the val2017 split. We report results on bounding box object detection $(\\mathrm{AP}^{\\mathrm{box}})$ and instance segmentation $(\\mathrm{AP}^{\\mathrm{mask}})$ , shown in Table 11. Our observations indicate that DMAE also demonstrates strong potential in downstream tasks like detection and segmentation.", + "bbox": [ + 496, + 625, + 893, + 762 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8c4577dee5ae6b3dca99b54656da26d9bd7a3e263035e18cd2b5f9b2c85bc546.jpg", + "table_caption": [ + "Table 10. DMAE effectively improves other self-supervised pre-training frameworks (including DINO and MoCo-v3) on ImageNet classification top-1 accuracy $(\\%)$ ." + ], + "table_footnote": [], + "table_body": "
method\\(AP^{box}\\)\\(AP^{mask}\\)
MAE51.245.5
DMAE53.446.9
", + "bbox": [ + 553, + 791, + 836, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 11. DMAE effectively improves the downstream tasks as well, including object detection and segmentation.", + "bbox": [ + 500, + 844, + 890, + 872 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24262", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 89, + 189, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present a quantitative evaluation of the computational cost of DMAE compared to baseline methods. Additionally, we conduct a standard deviation analysis to assess the stability of our DMAE approach. Finally, we propose an advanced fine-tuning recipe for applying DMAE to smaller ViT models.", + "bbox": [ + 75, + 114, + 470, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Computational Costs", + "text_level": 1, + "bbox": [ + 76, + 214, + 276, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 9, we provide a quantitative evaluation on the computational cost, which is tested on a single NVIDIA A5000 GPU. We can observe that, without a significantly increasing of training hours, DMAE substantially outperforms the MAE-100 baseline by $+2.4\\%$ on ImageNet; this result even exceeds the MAE-1600 baseline, which causes $\\sim 7x$ GPU hours than MAE-100 baseline. We additionally provide the actual GPU hours of other baselines, and find that the proposed DMAE stands as the most efficient one, meanwhile achieving the best top-1 ImageNet accuracy.", + "bbox": [ + 75, + 237, + 468, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Standard Deviation Analysis", + "text_level": 1, + "bbox": [ + 76, + 397, + 333, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the above experiments, we kept the same random seed. Following MAE, we perform the statistical analysis for DMAE by changing the random seeds. In Table 12, from top to down, we show three aligning settings; and from left to right, we show the results with the default seed, the average accuracy with three randomly sampled seeds, and their standard deviation, respectively. From these results, we could conclude that our DMAE can bring in statistically stable improvements.", + "bbox": [ + 75, + 420, + 468, + 556 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e91c0a67935d187adea730256c72a4627b4950484f05378cdd5c915aa546dbeb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PosAcc(%)AvgStandard Deviation
Bottom (3th)82.682.500.20
Mid (6th)83.683.670.08
Top (9th)84.084.030.10
", + "bbox": [ + 94, + 568, + 447, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Smaller ViT Models.", + "text_level": 1, + "bbox": [ + 76, + 756, + 272, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the original MAE paper does not offer specialized recipes for ViT-Small and Tiny, we by default use the recipe for ViT-Base to fine-tune these smaller ViTs. However, this recipe includes strong regularization and augmentation techniques that might lead to over-regularization for the smaller ViTs. To address this issue, we experiment with a modified recipe with weaker augmentation and regularization by removing MixUp, CutMix, and Stochastic Depth.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f5149e90040822103b0b1a1128199914e96c7ae78b4a5ff007ed6372b5d5de2f.jpg", + "table_caption": [ + "Table 12. Standard deviation analysis for DMAE. From top to bottom, we show three aligning settings, the model performance with the default seed, the average performance with three randomly sampled seeds, and their standard deviations, respectively. 'Pos' denotes the student distillation position. Acc denotes ImageNet Top-1 accuracy. 'Avg' denotes the average value over three times." + ], + "table_footnote": [], + "table_body": "
ViT-TinyTop-1 Acc (%)
MAE70.1
DeiT74.5
DMAE74.9
", + "bbox": [ + 602, + 88, + 794, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/92eb2b5cc66e6f4dbe7d5c526c679713f560f5cc190ae9957de73ec2cb35bd29.jpg", + "table_caption": [ + "(a) ViT-Tiny" + ], + "table_footnote": [], + "table_body": "
ViT-SmallTop-1 Acc (%)
MAE80.0
DeiT81.2
DMAE82.2
", + "bbox": [ + 602, + 174, + 794, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) ViT-Small", + "bbox": [ + 658, + 238, + 733, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 13. Weaker augmentation and regularization helps smaller ViT models, during finetuning for both MAE and DMAE on ImageNet classification top-1 accuracy $(\\%)$", + "bbox": [ + 498, + 256, + 893, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on ImageNet in Table 13 show that DMAE not only maintains its advantage over MAE, but also outperforms DeiT with this new recipe, highlighting its effectiveness.", + "bbox": [ + 498, + 325, + 890, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 387, + 617, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-supervised pre-training has demonstrated great success for those exponentially growing models in the natural language domain. Recently, the rise of MAE shows that a similar paradigm also works for the computer vision domain, and now the development of vision models may embark on a similar trajectory as in the language domain. Yet, it is often desirable to have a well-balanced model between performance and speed in real-world applications. This work is a small step towards unleashing the potential of knowledge distillation, a popular model compression technique, within the MAE framework. Our DMAE is a simple, efficient, and effective knowledge distillation method: feature alignment during MAE pre-training. Extensive experiments on multiple model scales demonstrate the effectiveness of our approach. Moreover, an intriguing finding is that it allows for a masking ratio even higher than the already large one used in MAE (i.e., $75\\%$ ). We have also validated the effectiveness of our DMAE when in the small-data regime. We hope this work can benefit future research in knowledge distillation with pre-trained models.", + "bbox": [ + 496, + 412, + 893, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 731, + 660, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by a gift from Open Philanthropy, TPU Research Cloud (TRC) program, and Google Cloud Research Credits program. We also want to acknowledge ONR with award N00014-21-1-2812.", + "bbox": [ + 498, + 757, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 832, + 596, + 848 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Yutong Bai, Xinlei Chen, Alexander Kirillov, Alan Yuille, and Alexander C Berg. Point-level region contrast for object detection pre-training. In Proceedings of the IEEE/CVF Con-", + "bbox": [ + 506, + 858, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24263", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 16061-16070, 2022. 1", + "[2] Yutong Bai, Haoqi Fan, Ishan Misra, Ganesh Venkatesh, Yongyi Lu, Yuyin Zhou, Qihang Yu, Vikas Chandra, and Alan Yuille. Can temporal information help with contrastive self-supervised learning? arXiv preprint arXiv:2011.13046, 2020. 1", + "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2021. 1, 2, 3", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, pages 9650-9660, 2021. 7", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 6", + "[6] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. In CVPR, pages 12052-12062, 2022. 6", + "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 7", + "[8] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In CVPR, pages 4794-4802, 2019. 2", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 3", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2", + "[12] Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. In ICML, pages 1607-1616. PMLR, 2018. 2", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 2, 3, 5", + "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 1", + "[15] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In CVPR, pages 1921-1930, 2019. 2, 3", + "[16] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In AAAI, volume 33, pages 3779-3787, 2019. 2, 3" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 1, 2", + "[18] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. NeurIPS, 31, 2018. 2, 3", + "[19] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 280-296. Springer, 2022. 7", + "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 4", + "[21] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, volume 34, pages 5191-5198, 2020. 1, 2", + "[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544, 2016. 2", + "[23] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2, 3", + "[24] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 2, 3, 6", + "[25] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 1, 3, 6", + "[26] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In CVPR, pages 32-42, 2021. 1", + "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2", + "[28] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 11(12), 2010. 2", + "[29] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2, 3", + "[30] Junfei Xiao, Longlong Jing, Lin Zhang, Ju He, Qi She, Zongwei Zhou, Alan Yuille, and Yingwei Li. Learning from temporal gradient for semi-supervised action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3252-3262, 2022. 2", + "[31] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24264", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 6", + "[33] Qihang Yu, Yingda Xia, Yutong Bai, Yongyi Lu, Alan L Yuille, and Wei Shen. Glance-and-gaze vision transformer. Advances in Neural Information Processing Systems, 34:12992-13003, 2021. 1", + "[34] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, pages 4320-4328, 2018. 1, 2", + "[35] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. arXiv preprint arXiv:2203.08679, 2022. 2" + ], + "bbox": [ + 78, + 90, + 467, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24265", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_model.json b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e54021665dd1ebf9e8884e25780eee1d5f212a46 --- /dev/null +++ b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_model.json @@ -0,0 +1,1936 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Masked Autoencoders Enable Efficient Knowledge Distillers" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.18, + 0.748, + 0.252 + ], + "angle": 0, + "content": "Yutong Bai1 Zeyu Wang2 Junfei Xiao1 Chen Wei1 Huiyu Wang1 Alan Yuille1 Yuyin Zhou2 Cihang Xie2 1Johns Hopkins University 2 University of California, Santa Cruz ytongbai@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.318, + 0.47, + 0.485 + ], + "angle": 0, + "content": "This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, i.e., forward propagate inputs through the first few layers, for obtaining intermediate feature maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.473, + 0.712 + ], + "angle": 0, + "content": "Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves \\(84.0\\%\\) ImageNet top-1 accuracy, outperforming the baseline of directly distilling a fine-tuned ViT-L by \\(1.2\\%\\). More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with \\(95\\%\\) masking ratio where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of \\(83.6\\%\\); surprisingly, it can still secure \\(82.4\\%\\) top-1 ImageNet accuracy by aggressively training with just FOUR visible patches (98% masking ratio). The code and models are publicly available at https://github.com/UCSC-VLAA/DMAE." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Following the success in the natural language processing [10, 27], the Transformer architecture is showing tremendous potentials in computer vision [2, 11, 25, 26, 33], especially when they are pre-trained with a huge amount of unlabelled data [3] with self-supervised learning techniques [1, 14]. Masked image modeling, which trains models to predict the masked signals (either as raw pixels or as semantic tokens) of the input image, stands as one of the most powerful ways for feature pre-training. With the most re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.364 + ], + "angle": 0, + "content": "cent representative work in this direction, masked autoencoder (MAE) [13], we are now able to efficiently and effectively pre-train high-capacity Vision Transformers (ViTs) with strong feature representations, leading to state-of-the-art solutions for a wide range of downstream visual tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.893, + 0.593 + ], + "angle": 0, + "content": "In this paper, we are interested in applying knowledge distillation [17], which is one of the most popular model compression techniques, to transfer the knowledge from these strong but cumbersome ViTs into smaller ones. In contrast to prior knowledge distillation works [17, 21, 34], the teacher considered here is a pre-trained model whose predictions do not necessarily reveal the fine-grained relationship between categories; therefore, typical solutions like aligning the soft/hard logits between the teacher model and the student model may no longer remain effective. Moreover, after distilling the pre-trained teacher model, these student models need an extra round of fine-tuning to adapt to downstream tasks. These factors turn distilling pretrained models seemingly a less favorable design choice in terms of both performance and computational cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Nonetheless, surprisingly, we find by building upon MAE, the whole distillation framework can efficiently yield high-performance student models. There are two key designs. Firstly, we follow MAE to let the encoder exclusively operate on a small visible subset of patches and to employ a lightweight decoder for pixel reconstruction. Whereas rather than using the \"luxury\" setups in MAE, we show aggressively simplifying pre-training from 1600 epochs to 100 epochs and pushing masking ratio from \\(75\\%\\) to \\(95\\%\\) suffice to distill strong student models. Secondly, instead of aligning logits, we alternatively seek to match the intermediate feature representation; this enables the cumbersome teacher model to only forward propagate inputs through the first few layers, therefore, reducing computations. We note applying L1 norm for distance measure is essential recipe for ensuring a successful intermediate feature alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We name this distilling MAE framework as DMAE. Compared to the traditional knowledge distillation framework where the teacher is a fine-tuned model, DMAE is more efficient and can train much stronger student models" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24256" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.098, + 0.862, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Figure 1. Illustration of the distillation process in DMAE. There are two key designs. Firstly, following MAE, we hereby only take visible patches as inputs and aims to reconstruct the masked ones. Secondly, knowledge distillation is achieved by aligning the intermediate features between the teacher model and the student model. Note the gray blocks denote the dropped high-level layers of the teacher model during distillation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.473, + 0.604 + ], + "angle": 0, + "content": "at different capacities. For example, by setting ViT-B as the student model, while the baseline of distilling a fine-tuned ViT-L achieves \\(82.8\\%\\) top-1 ImageNet accuracy, DMAE substantially boosts the performance to \\(84.0\\%\\) \\((+1.2\\%)\\) top-1 ImageNet accuracy, at a even lower training cost (i.e., 195 GPU hours vs. 208 GPU hours, see Table 9). More intriguingly, we found that DMAE allows for robust training with extremely highly masked images—even with TEN visible patches (i.e., \\(95\\%\\) masking ratio), ViT-B can competitively attain a top-1 ImageNet accuracy of \\(83.6\\%\\); this masking ratio can further be aggressively pushed to \\(98\\%\\) (FOUR visible patches) where DMAE still help ViT-B secure \\(82.4\\%\\) top-1 ImageNet accuracy. We hope this work can benefit future research on efficiently unleashing the power of pretrain models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.22, + 0.634 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.853 + ], + "angle": 0, + "content": "Knowledge distillation (KD) is a popular model compression technique that allows models to achieve both strong performances of large models and fast inference speed of small models. The first and seminal KD approach, proposed in [17], transfers the \"dark knowledge\" via minimizing the KL divergence between the soft logits of the teacher model and that of the student model. From then on, many advanced KD methods have been developed, which can be categorized into two branches: logits distillation [8, 12, 21, 30, 34, 35] and intermediate representation distillation [15, 16, 18, 23, 24]. Our DMAE belongs to the second branch, as it minimizes the distance between latent features of the teacher model and those of the student model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The first feature-based distillation method is Fit-Nets [23]. In addition to aligning logits, FitNets requires the student model to learn an intermediate representation that is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.895, + 0.529 + ], + "angle": 0, + "content": "predictive of the intermediate representations of the teacher network. Heo et al. [15] re-investigates the design of feature distillation and develops a novel KD method to create a synergy among various aspects, including teacher transform, student transform, distillation feature position, and distance function. CRD [24] incorporates contrastive learning into KD to capture correlations and higher-order output dependencies. Unlike these existing works, our DMAE is the first to consider applying KD to extra information from self-supervised pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.896, + 0.777 + ], + "angle": 0, + "content": "Masked image modeling (MIM) helps models to acquire meaningful representations by reconstructing masked images. The pioneering works are built on denoising autoencoders [28] and context encoders [22]. Following the success of BERT in natural language [10], and also with the recent trend of adopting Transformer [27] to computer vision [11], there have emerged a set of promising works on applying MIM for self-supervised visual pre-training. BEiT [3] first successfully adopts MIM to ViT pre-training by learning to predict visual tokens. MaskFeat [29] finds that learning to reconstruct HOG features enables effective visual representation learning. SimMIM [31] and MAE [13] both propose to directly reconstruct the pixel values of the masked image patches. Our work is built on MAE and finds that MAE enables the whole distillation framework to be efficient and effective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.792, + 0.61, + 0.809 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.707, + 0.832 + ], + "angle": 0, + "content": "3.1. Masked Autoencoders" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our method is built upon MAE, a powerful autoencoder-based MIM approach. Specifically, the MAE encoder first projects unmasked patches to a latent space, which are then fed into the MAE decoder to help predict pixel values of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24257" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.429, + 0.106 + ], + "angle": 0, + "content": "masked patches. The core elements in MAE include:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.112, + 0.47, + 0.233 + ], + "angle": 0, + "content": "Masking. MAE operates on image tokens, i.e., the image needs to be divided into non-overlapping patches. A random small subset of those patches will be kept for the MAE encoder, and the rest will be set as the predicting target of the MAE decoder. Typically, a high masking ratio (e.g., \\(75\\%\\)) is applied, preventing models from taking shortcuts (e.g., simply extrapolating missing pixels based on neighbors) in representation learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.47, + 0.297 + ], + "angle": 0, + "content": "MAE encoder. The MAE encoder is a standard ViT architecture except that it only operates on those unmasked patches. This design largely reduces the computation cost of encoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.469, + 0.408 + ], + "angle": 0, + "content": "MAE decoder. In addition to the encoded features of un-masked patches (from MAE encoder), the MAE decoder receives mask tokens as input, a learned vector shared across all missing positions. The mask token is only used during pre-training, allowing independent decoder design. Particularly, MAE adopts a lightweight decoder for saving computations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.469, + 0.473 + ], + "angle": 0, + "content": "Reconstruction. Different from BEiT [3] or MaskFeat [29], MAE directly reconstructs image pixel values. The simple mean squared error is applied to masked tokens for calculating loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.469, + 0.58 + ], + "angle": 0, + "content": "Note that, other than distillation-related operations, the whole pre-training and fine-tuning process in this paper exactly follow the default setup in MAE, unless specifically mentioned. Interestingly, compared to MAE, our DMAE robustly enables a much more efficient pre-training setup, e.g., 100 (vs. 1600) training epochs and \\(95\\%\\) (vs. \\(75\\%\\)) masking ratio." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.289, + 0.607 + ], + "angle": 0, + "content": "3.2. Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.84 + ], + "angle": 0, + "content": "MAE demonstrates extraordinary capabilities in learning high-capacity models efficiently and effectively. In this work, we seek to combine knowledge distillation with the MAE framework, to efficiently acquire small and fast models with similar performance as those powerful yet cumbersome models. The most straightforward approach is directly applying existing knowledge distillation methods, like the one proposed in DeiT [25], to a fine-tuned MAE model. However, we empirically find that this approach hardly brings in improvements. In addition, this approach fails to leverage the special designs in MAE for reducing computations, e.g., only feeding a small portion of the input image to the encoder. To this end, we hereby study an alternative solution: directly applying knowledge distillation at the pre-training stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Since there are no categorical labels in MAE pretraining, distilling logits can hardly help learn semantically meaningful representations. We, therefore, resort to distilling the intermediate features. This idea is first developed in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.226 + ], + "angle": 0, + "content": "FitNets [23], and inspired a set of followups for advancing knowledge distillation [15, 16, 18, 24]. Concretely, we first extract the features from the specific layers of the student model; after feeding such features into a small project head, the outputs will be asked to mimic the features from the corresponding layers of the teacher model. In practice, the projection head is implemented by simple fully connected layer, which addresses the possible feature dimension mismatch between teacher models and student models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.227, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Formally, let \\(\\mathbf{x} \\in \\mathbb{R}^{3HW \\times 1}\\) be the input pixel RGB values and \\(\\mathbf{y} \\in \\mathbb{R}^{3HW \\times 1}\\) be the predicted pixel values, where \\(H\\) denotes image height and \\(W\\) denotes image width. The MAE reconstruction loss \\(L_{MAE}\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.293, + 0.892, + 0.33 + ], + "angle": 0, + "content": "\\[\nL _ {M A E} = \\frac {1}{\\Omega (\\mathbf {x} _ {M})} \\sum_ {i \\in M} \\left(\\mathbf {y} _ {i} - \\mathbf {x} _ {i}\\right) ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.367 + ], + "angle": 0, + "content": "where \\(M\\) denotes the set of masked pixels, \\(\\Omega(.)\\) is the number of elements, and \\(i\\) is the pixel index." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Let \\(\\mathbf{z}_l^S, \\mathbf{z}_l^T \\in \\mathbb{R}^{LC \\times 1}\\) be the features extracted from the \\(l\\)th layer of the student model and the teacher model, respectively, where \\(L\\) denotes the patch numbers, and \\(C\\) denotes the channel dimension. We use \\(\\sigma()\\) to denote the projection network function. Our feature alignment distillation loss \\(L_{Dist}\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.464, + 0.892, + 0.5 + ], + "angle": 0, + "content": "\\[\nL _ {D i s t} = \\sum_ {l} \\frac {1}{\\Omega (\\mathbf {z} _ {l} ^ {T})} \\sum_ {i} \\left\\| \\sigma (\\mathbf {z} _ {l} ^ {S}) _ {i} - \\mathbf {z} _ {l, i} ^ {T} \\right\\| _ {1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.567 + ], + "angle": 0, + "content": "The final loss used in pre-training is a weighted summation of MAE reconstruction loss \\( L_{MAE} \\) and the feature alignment distillation loss \\( L_{Dist} \\), controlled by the hyperparameter \\( \\alpha \\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.575, + 0.892, + 0.591 + ], + "angle": 0, + "content": "\\[\nL = L _ {M A E} + \\alpha \\times L _ {D i s t}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.796 + ], + "angle": 0, + "content": "The framework of DMAE is summarized in Figure 1. Following MAE, DMAE also takes masked inputs and performs the pretext task of masked image modeling. Besides, the corresponding features are aligned between the teacher model and the student model. It is worthy of highlighting that DMAE is an efficient knowledge distiller: 1) it only operates on a tiny subset of visible patches i.e., a high masking ratio is applied; and 2) aligning intermediate layer features reduce the computation cost of (cumbersome) teacher model. In the next section, we extensively compare our method with three baselines: the original MAE without any distillation, DeiT-style distillation, and feature alignment distillation in the supervised setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.634, + 0.825 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.719, + 0.849 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Following MAE [13], we first perform self-supervised pre-training on ImageNet-1k [9]. Unless otherwise mentioned, the teacher models are public checkpoints released" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24258" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.204, + 0.089, + 0.765, + 0.179 + ], + "angle": 0, + "content": "
# of LayersLayer LocationStudent Aligned Layer IndexImageNet Top-1 Acc (%)
SingleBottom382.6
Middle683.6
984.0
Top1283.4
MultipleMiddle+Top6+1284.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.183, + 0.895, + 0.227 + ], + "angle": 0, + "content": "Table 1. The effects of feature alignment location. We hereby test with 5 different layer locations, where top layers refer to those closer to network outputs. For single layer feature alignment, features from the \\( \\frac{3}{4} \\) depth of the model leads to the best ImageNet top-1 accuracy. This performance is even comparable to the setup of aligning multiple layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.471, + 0.418 + ], + "angle": 0, + "content": "from the official MAE implementations1. For pre-training, we train all models using AdamW optimizer [20], with a base learning rate of 1.5e-4, weight decay of 0.05, and optimizer momentum \\(\\beta_{1},\\beta_{2} = 0.9,0.95\\). We use a total batch size of 4096, and pre-train models for 100 epochs with a warmup epoch of 20 and a cosine learning rate decay schedule. We by default use the masking ratio of \\(75\\%\\); while the ablation study shows that our method can robustly tackle extremely high masking ratios. After pre-training, the teacher model is dropped, and we exactly follow the default setups in MAE to fine-tune the student model on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.569 + ], + "angle": 0, + "content": "For feature alignments, we choose to align the features from the \\(\\frac{3}{4}\\) depth of both the student model and the teacher model, which we find delivers decent results for all model sizes tested. For example, with a 24-layer ViT-L as the teacher model and a 12-layer ViT-B as the student model, features from the 9th layer of ViT-B are aligned with the features from the 18th layer of ViT-L. We set \\(\\alpha = 1\\) in Eq. 3 to balance the tradeoff between MAE reconstruction loss \\(L_{MAE}\\) and the feature alignment distillation loss \\(L_{Dist}\\) in pre-training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.579, + 0.182, + 0.595 + ], + "angle": 0, + "content": "4.2. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.707 + ], + "angle": 0, + "content": "We first provide a detailed analysis of how to set distillation-related parameters in DMAE. Specifically, we set the teacher model as an MAE pre-trained ViT-L (from MAE official GitHub repository, attaining \\(85.9\\%\\) top-1 ImageNet accuracy after fine-tuning), and set the student model as a randomly initialized ViT-B. We analyze the following six factors:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Where to align. In Table 1 we first check the effect of feature alignment location on model performance. We observe that shallower features are less favored: e.g., the 3rd layer alignment under-performs all other settings. We speculate this is due to the learning process of ViTs—images are much noisier and less semantic than texts, ViTs will first group the raw pixels in the bottom layers (closer to the input), which is harder to transfer. While features from \\(\\frac{3}{4}\\) depth (i.e., the 9th layer in ViT-B) achieve the best performance, a simple rule-of-thumb which we find fits all model scales in our experiments. We adopt this design choice in all" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.249, + 0.888, + 0.311 + ], + "angle": 0, + "content": "
Teacher Layer (relative position)ImageNet Top-1 Acc (%)
Middle84.0
Top83.3
Bottom82.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.315, + 0.892, + 0.344 + ], + "angle": 0, + "content": "Table 2. The analysis of aligning order on ImageNet classification top-1 accuracy \\((\\%)\\)" + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.358, + 0.84, + 0.419 + ], + "angle": 0, + "content": "
Projection HeadImageNet Top-1 Acc (%)
Linear84.0
2-layer MLP84.0
3-layer MLP83.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.424, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Table 3. Projection Head. A simple fully-connected layer works the best. We choose this as the default setting." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.481, + 0.892, + 0.543 + ], + "angle": 0, + "content": "other experiments. It is also worth mentioning that simply aligning multiple layers has no clear advantage over aligning features from \\(\\frac{3}{4}\\) depth (84.2% vs. 84.0%); we, therefore, stick to the \\(\\frac{3}{4}\\) depth setting, which is more efficient." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.547, + 0.893, + 0.669 + ], + "angle": 0, + "content": "Aligning order. We next test the importance of alignment ordering on model performance. Specifically, by fixing the layer location in the student model (i.e., the middle layer in our experiment), we then align it to different layers of the teacher model. As shown in Table 2, we observe that when the aligned layers are in the same relative position (e.g., middle to middle), the student model can achieve the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Masking ratio. MAE reveals that the masking ratio in masked image modeling could be surprisingly high (75%). The hypothesis is that by learning to reason about the gestalt of the missing objects and scenes, which cannot be done by extending lines or textures because of the high masking ratio, the model is also learning useful representations. Interestingly, we find that when combining MAE and knowledge distillation, an even much higher masking ratio is possible, as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Firstly, it is interesting to note that, compared to the typical \\(75\\%\\) masking ratio setting, further raising the masking ratio to \\(90\\%\\) comes at no performance drop, i.e., both attain \\(84.0\\%\\) top-1 ImageNet accuracy. Next, even with an extremely large masking ratio like \\(98\\%\\) (only FOUR visible patches), DMAE still beats the 100-epoch MAE baseline" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.322, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/facebookresearch/mae" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24259" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.199, + 0.089, + 0.77, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.895, + 0.342 + ], + "angle": 0, + "content": "Figure 2. DMAE allows an extremely high masking ratio. From left to right, we increase the masking ratio from the basic \\(75\\%\\) to the extreme \\(99\\%\\). We note that our DMAE competitively attains \\(83.6\\%\\) top-1 ImageNet accuracy with \\(95\\%\\) masking ratio (TEN visible patches), and still secures \\(82.4\\%\\) top-1 ImageNet accuracy even by learning with FOUR visible patches (\\(98\\%\\) masking ratio)." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.364, + 0.413, + 0.426 + ], + "angle": 0, + "content": "
Decoder DepthImageNet Top-1 Acc (%)
283.7
483.8
884.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.43, + 0.47, + 0.459 + ], + "angle": 0, + "content": "Table 4. Decoder Depth. A deeper decoder (slightly) improves the pre-trained representation quality." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.474, + 0.465, + 0.577 + ], + "angle": 0, + "content": "
Loss DesignImageNet Top-1 Acc (%)
Loss ChoiceL1 with α = 184.0
L2 with α = 183.3
Loss RatioL1 with α = 0.583.9
L1 with α = 184.0
L1 with α = 284.0
L1 with α = 484.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Table 5. Loss Function. From the first block, we can observe that L1 distance yields significantly higher performance than L2 distance. From the second block, we can observe that the hyperparameter \\(\\alpha\\) has little influence on the representation quality of DMAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.772 + ], + "angle": 0, + "content": "that uses a masking ratio of \\(75\\%\\) (second row in Table 6), by a non-trivial-margin \\((82.4\\% \\text{ vs. } 81.6\\%)\\). These results suggest that, with the assistance of distilled knowledge from the teacher model, the student model can make better use of visible patches, even at a very limited amount, for representation learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Projection head. The goal of the proposed feature alignment distillation is to encourage the student model to learn features that are predictive of features from a stronger teacher model. To that end, a small projection head is employed on features from the student model, to 1) project them onto a space of the same dimension as the hidden dimension of the teacher model, and 2) provide extra flexibility for feature alignment. We ablate the choice of this pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.413 + ], + "angle": 0, + "content": "jection network, as shown in Table 3. We can observe that applying a simple fully connected layer already performs the best among other choices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.537 + ], + "angle": 0, + "content": "Decoder depth. In Table 4 we analyze the effect of decoder depth. Similar to MAE [13], the final performance gets (slightly) increased with a deeper decoder. We choose a decoder depth of 8 as the default setting as in [13]. Note that a decoder depth of 2 is also a competitive choice—compared to an 8-depth decoder, it significantly reduces the computation cost while only marginally sacrificing the accuracy by \\(0.3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.693 + ], + "angle": 0, + "content": "Loss designs. Table 5 ablates the loss design. While SimMIM [31] shows that L1 distance and L2 distance lead to similar performance, ours suggests that L1 distance exhibits a clear advantage over L2 distance, i.e., \\(+0.7\\%\\) improvement. Furthermore, we note DMAE is quite robust to the specific value of the hyperparameter \\(\\alpha\\), which controls the relative importance of the distillation loss over the reconstruction loss. Based on these results, we choose L1 in Eq. 2 for distance measure and set \\(\\alpha = 1\\) in Eq. 3 for the rest experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.703, + 0.75, + 0.72 + ], + "angle": 0, + "content": "4.3. Comparison with Baselines." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.727, + 0.892, + 0.756 + ], + "angle": 0, + "content": "In Table 6, we compare the performance of our DMAE with various baselines:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.867 + ], + "angle": 0, + "content": "MAE. The MAE baselines are presented in the first block of Table 6. If MAE is also asked to pre-train for only 100 epochs, DMAE can substantially outperform this baseline by \\(2.4\\%\\) (from \\(81.6\\%\\) to \\(84.0\\%\\)). When comparing to a much stronger but more computationally expensive MAE baseline with 1600 pre-training epoch, we note DMAE still beats it by \\(0.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Supervised model. The second block in Table 6 demonstrates the effectiveness of DMAE compared with models" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24260" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.212, + 0.089, + 0.756, + 0.276 + ], + "angle": 0, + "content": "
MethodPre-training epochsSupervised training / fine-tuning epochsImageNet Top-1 Acc (%)
MAE-B10010081.6
MAE-B160010083.6
DeiT-B-10076.8
DeiT-B-30081.8
DeiT-B-Soft Distillation-10077.5
DeiT-B-Hard Distillation-10078.3
CRD [24]-10081.9
SRRL [32]-10082.2
Dear-KD [6]-10082.4
Supervised Feature Alignment-10082.8
DMAE-B10010084.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.28, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Table 6. DMAE shows stronger performance than all three kinds of baselines: MAE, supervised model, and other existing advanced distillation strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.381 + ], + "angle": 0, + "content": "trained under the supervision of categorical labels, which requires a much longer training time, i.e., \\(+2.2\\%\\) compared to the DeiT 300 epochs supervised training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Other distillation strategies. We next compare DMAE with other distillation methods. We consider DeiT-style logit-based distillation [25], CRD [24], SRRL [32], DearKD [6], and feature alignment distillation. Note that for these baselines, one significant difference from DMAE is that the student model here directly distill knowledge from a supervisely fine-tuned teacher model. Moreover, to make these baselines more competitive, the teacher models will first be MAE pre-trained and then fine-tuned on ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.471, + 0.686 + ], + "angle": 0, + "content": "The results are shown in the third block of Table 6. Firstly, we can observe that the DeiT-style logit-based distillation, either soft or hard, even hurts the student models' performance. This phenomenon potentially suggests that such a distillation strategy may not fit teacher models of ViT architectures. For other baselines, we note that feature alignment distillation performs the best; but this is still worse than DMAE (82.8% vs. 84.0%), indicating the importance and effectiveness of distilling knowledge from a pre-trained teacher model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.36, + 0.71 + ], + "angle": 0, + "content": "4.4. Scaling to Different Model Sizes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.838 + ], + "angle": 0, + "content": "We test DMAE with different model sizes, listed in Table 7. For a fair comparison, both methods only pre-train models for 100 epochs. DMAE shows consistent improvement compared to MAE across different model sizes. With only one middle layer feature alignment, DMAE brings an additional improvement of \\(+2.4\\%\\) with ViT-B, \\(+2.7\\%\\) with ViT-Small, and \\(+3.4\\%\\) with ViT-Tiny. In addition, we are interested in the following two cases:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Same teacher model, different student model. As shown in the first two lines in Table 7, we find that when using ViT-L as the teacher model, both ViT-B and ViT-S benefit from the distillation, demonstrating a clear advantage over" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.424 + ], + "angle": 0, + "content": "
Student ModelTeacher ModelImageNet Top-1 Acc (%)
MAEDMAE
BaseLarge81.684.0 (+2.4)
SmallLarge77.480.1 (+2.7)
Base77.479.3 (+1.9)
TinyBase66.670.0 (+3.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.429, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Table 7. Across different model sizes, DMAE shows consistent improvements compared with MAE." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.484, + 0.892, + 0.546 + ], + "angle": 0, + "content": "the MAE baseline. We argue that the ability to effectively generalize to cases where an even smaller student model is desirable, especially for those computation-constrained real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.55, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Same student model, different teacher model. As shown in Table 7, with a ViT-S as the student model, enlarging the teacher model from ViT-B to ViT-L further boosts the accuracy by \\(0.8\\%\\) (from \\(79.3\\%\\) to \\(80.1\\%\\)). This result suggests that DMAE can effectively distill knowledge from the teacher models at different scales." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.712, + 0.668 + ], + "angle": 0, + "content": "4.5. Limited Training Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In certain real-world applications, data could be hard to acquire because of high data collection and labeling costs or due to privacy concerns. Leveraging models pre-trained on large-scale unlabeled datasets for fine-tuning when only a small dataset of downstream tasks is available becomes a promising solution. Here we seek to test the potential of our DMAE in this data-scarce scenario. We strictly follow [5] to sample \\(1\\%\\) or \\(10\\%\\) of the labeled ILSVRC-12 training datasets in a class-balanced way. We set MAE pre-trained ViT-L as the teacher model and a randomly initialized ViT-B as the student model. In addition, we compare DMAE with three kinds of baselines described in Section 4.3: MAE, supervised model, and other distillation strategies, and similarly, set the teacher model to be a ViT-L that first pre-trained with MAE and then fine" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24261" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.805, + 0.22 + ], + "angle": 0, + "content": "
MethodPre-training epochsSupervised training / fine-tuning epochsIN-1% Top-1 Acc(%)IN-10% Top-1 Acc(%)
MAE-B10010033.965.0
MAE-B160010049.672.8
DeiT-B-100--
DeiT-B-Soft Distillation160010036.066.4
DeiT-B-Hard Distillation160010037.367.3
Supervised Feature Alignment160010034.267.6
DMAE-B10010050.373.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.225, + 0.894, + 0.255 + ], + "angle": 0, + "content": "Table 8. DMAE demonstrates much stronger performance than all other baselines when training data is limited. Note that the DeiT-B baseline is unable to converge because of data insufficiency." + }, + { + "type": "table", + "bbox": [ + 0.209, + 0.267, + 0.761, + 0.385 + ], + "angle": 0, + "content": "
ModelTraining Cost (GPU Hours)ImageNet Top-1 Acc(%)
Pre-trainingFine-tuningOverall
MAE-B-100-epoch78h112h190h81.6
MAE-B-1600-epoch1248h112h1360h83.6
DeiT-B-Soft Distillation-213h213h77.5
DeiT-B-Hard Distillation-213h213h78.3
Supervised Feature Alignment-208h208h82.8
DMAE-B83h112h195h84.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.389, + 0.893, + 0.434 + ], + "angle": 0, + "content": "Table 9. Computational cost comparisons among DMAE and other baselines. The training cost is measured by A5000 GPU hours. We note the proposed DMAE maintains a similar (or even cheaper) training cost than others, while achieving much higher top-1 ImageNet accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.534 + ], + "angle": 0, + "content": "tuned on ImageNet-1k. Note that since DMAE has full access to the \\(100\\%\\) ImageNet dataset (without labels) during pre-training, to ensure a fair and competitive comparison, we initialize all the baselines as the 1600-epoch MAE pretrained model on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.471, + 0.678 + ], + "angle": 0, + "content": "Table 8 shows that DMAE largely surpasses all other baselines. For example, when only \\(10\\%\\) ImageNet data is available for supervised training or fine-tuning, DMAE outperforms the MAE pre-trained baseline by \\(8.4\\%\\) (i.e., \\(73.4\\%\\) vs. \\(65.0\\%\\)). DMAE also significantly outperforms other distillation strategies, with an improvement ranging from \\(5.8\\%\\) to \\(7.0\\%\\). We note this accuracy gap is even larger when only \\(1\\%\\) ImageNet is available, demonstrating the data efficiency of DMAE." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.705, + 0.37, + 0.721 + ], + "angle": 0, + "content": "4.6. Generalization to Other Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Lastly, we provide preliminary results of integrating DMAE into other self-supervised training frameworks, including DINO [4] and MoCo-V3 [7]. Unlike MAE, which belongs to masked image modeling, DINO and MoCo-V3 are contrastive learning-based methods. Still, as shown in Table 10, without further hyperparameter tuning, DMAE effectively shows non-trivial improvements on top of both DINO (+1.3%) and MoCo-v3 (+1.4%), demonstrating the potential of distilling pre-trained models (rather than fine-tuned models as in most existing knowledge distillation frameworks)." + }, + { + "type": "table", + "bbox": [ + 0.566, + 0.455, + 0.825, + 0.503 + ], + "angle": 0, + "content": "
w/o DMAEw/ DMAE
DINO80.982.2 (+1.3)
MoCo-V381.182.5 (+1.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.507, + 0.894, + 0.551 + ], + "angle": 0, + "content": "Table 10. DMAE effectively improves other self-supervised pre-training frameworks (including DINO and MoCo-v3) on ImageNet classification top-1 accuracy \\((\\%)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.593, + 0.816, + 0.609 + ], + "angle": 0, + "content": "4.7. Generalization to Downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.894, + 0.763 + ], + "angle": 0, + "content": "Following ViT-Det [19], we conduct downstream fin-tuning on the COCO dataset, where MAE/DMAE pretrained ViT-Base model is adopted as the plain backbone of Mask-RCNN. We train the model on the train2017 split and evaluate it on the val2017 split. We report results on bounding box object detection \\((\\mathrm{AP}^{\\mathrm{box}})\\) and instance segmentation \\((\\mathrm{AP}^{\\mathrm{mask}})\\), shown in Table 11. Our observations indicate that DMAE also demonstrates strong potential in downstream tasks like detection and segmentation." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.792, + 0.837, + 0.841 + ], + "angle": 0, + "content": "
method\\(AP^{box}\\)\\(AP^{mask}\\)
MAE51.245.5
DMAE53.446.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.845, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Table 11. DMAE effectively improves the downstream tasks as well, including object detection and segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24262" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.191, + 0.106 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.471, + 0.206 + ], + "angle": 0, + "content": "In this section, we present a quantitative evaluation of the computational cost of DMAE compared to baseline methods. Additionally, we conduct a standard deviation analysis to assess the stability of our DMAE approach. Finally, we propose an advanced fine-tuning recipe for applying DMAE to smaller ViT models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.215, + 0.277, + 0.231 + ], + "angle": 0, + "content": "5.1. Computational Costs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.47, + 0.39 + ], + "angle": 0, + "content": "In Table 9, we provide a quantitative evaluation on the computational cost, which is tested on a single NVIDIA A5000 GPU. We can observe that, without a significantly increasing of training hours, DMAE substantially outperforms the MAE-100 baseline by \\(+2.4\\%\\) on ImageNet; this result even exceeds the MAE-1600 baseline, which causes \\(\\sim 7x\\) GPU hours than MAE-100 baseline. We additionally provide the actual GPU hours of other baselines, and find that the proposed DMAE stands as the most efficient one, meanwhile achieving the best top-1 ImageNet accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.398, + 0.334, + 0.415 + ], + "angle": 0, + "content": "5.2. Standard Deviation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.557 + ], + "angle": 0, + "content": "In the above experiments, we kept the same random seed. Following MAE, we perform the statistical analysis for DMAE by changing the random seeds. In Table 12, from top to down, we show three aligning settings; and from left to right, we show the results with the default seed, the average accuracy with three randomly sampled seeds, and their standard deviation, respectively. From these results, we could conclude that our DMAE can bring in statistically stable improvements." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.569, + 0.449, + 0.631 + ], + "angle": 0, + "content": "
PosAcc(%)AvgStandard Deviation
Bottom (3th)82.682.500.20
Mid (6th)83.683.670.08
Top (9th)84.084.030.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Table 12. Standard deviation analysis for DMAE. From top to bottom, we show three aligning settings, the model performance with the default seed, the average performance with three randomly sampled seeds, and their standard deviations, respectively. 'Pos' denotes the student distillation position. Acc denotes ImageNet Top-1 accuracy. 'Avg' denotes the average value over three times." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.273, + 0.772 + ], + "angle": 0, + "content": "5.3. Smaller ViT Models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Since the original MAE paper does not offer specialized recipes for ViT-Small and Tiny, we by default use the recipe for ViT-Base to fine-tune these smaller ViTs. However, this recipe includes strong regularization and augmentation techniques that might lead to over-regularization for the smaller ViTs. To address this issue, we experiment with a modified recipe with weaker augmentation and regularization by removing MixUp, CutMix, and Stochastic Depth." + }, + { + "type": "table", + "bbox": [ + 0.604, + 0.089, + 0.795, + 0.15 + ], + "angle": 0, + "content": "
ViT-TinyTop-1 Acc (%)
MAE70.1
DeiT74.5
DMAE74.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.663, + 0.152, + 0.731, + 0.164 + ], + "angle": 0, + "content": "(a) ViT-Tiny" + }, + { + "type": "table", + "bbox": [ + 0.603, + 0.175, + 0.795, + 0.237 + ], + "angle": 0, + "content": "
ViT-SmallTop-1 Acc (%)
MAE80.0
DeiT81.2
DMAE82.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.659, + 0.239, + 0.735, + 0.251 + ], + "angle": 0, + "content": "(b) ViT-Small" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.257, + 0.894, + 0.3 + ], + "angle": 0, + "content": "Table 13. Weaker augmentation and regularization helps smaller ViT models, during finetuning for both MAE and DMAE on ImageNet classification top-1 accuracy \\((\\%)\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.373 + ], + "angle": 0, + "content": "Results on ImageNet in Table 13 show that DMAE not only maintains its advantage over MAE, but also outperforms DeiT with this new recipe, highlighting its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.619, + 0.404 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.718 + ], + "angle": 0, + "content": "Self-supervised pre-training has demonstrated great success for those exponentially growing models in the natural language domain. Recently, the rise of MAE shows that a similar paradigm also works for the computer vision domain, and now the development of vision models may embark on a similar trajectory as in the language domain. Yet, it is often desirable to have a well-balanced model between performance and speed in real-world applications. This work is a small step towards unleashing the potential of knowledge distillation, a popular model compression technique, within the MAE framework. Our DMAE is a simple, efficient, and effective knowledge distillation method: feature alignment during MAE pre-training. Extensive experiments on multiple model scales demonstrate the effectiveness of our approach. Moreover, an intriguing finding is that it allows for a masking ratio even higher than the already large one used in MAE (i.e., \\(75\\%\\)). We have also validated the effectiveness of our DMAE when in the small-data regime. We hope this work can benefit future research in knowledge distillation with pre-trained models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.732, + 0.661, + 0.75 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.758, + 0.892, + 0.817 + ], + "angle": 0, + "content": "This work is supported by a gift from Open Philanthropy, TPU Research Cloud (TRC) program, and Google Cloud Research Credits program. We also want to acknowledge ONR with award N00014-21-1-2812." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.833, + 0.597, + 0.849 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.859, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[1] Yutong Bai, Xinlei Chen, Alexander Kirillov, Alan Yuille, and Alexander C Berg. Point-level region contrast for object detection pre-training. In Proceedings of the IEEE/CVF Con-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24263" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 16061-16070, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[2] Yutong Bai, Haoqi Fan, Ishan Misra, Ganesh Venkatesh, Yongyi Lu, Yuyin Zhou, Qihang Yu, Vikas Chandra, and Alan Yuille. Can temporal information help with contrastive self-supervised learning? arXiv preprint arXiv:2011.13046, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.192, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, pages 9650-9660, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[6] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. In CVPR, pages 12052-12062, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.391, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.435, + 0.469, + 0.462 + ], + "angle": 0, + "content": "[8] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In CVPR, pages 4794-4802, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.464, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[12] Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. In ICML, pages 1607-1616. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[15] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In CVPR, pages 1921-1930, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In AAAI, volume 33, pages 3779-3787, 2019. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[18] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. NeurIPS, 31, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[19] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 280-296. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[21] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, volume 34, pages 5191-5198, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[23] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[24] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[25] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[26] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In CVPR, pages 32-42, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[28] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 11(12), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[29] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[30] Junfei Xiao, Longlong Jing, Lin Zhang, Ju He, Qi She, Zongwei Zhou, Alan Yuille, and Yingwei Li. Learning from temporal gradient for semi-supervised action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3252-3262, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[31] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24264" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[32] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[33] Qihang Yu, Yingda Xia, Yutong Bai, Yongyi Lu, Alan L Yuille, and Wei Shen. Glance-and-gaze vision transformer. Advances in Neural Information Processing Systems, 34:12992-13003, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.232 + ], + "angle": 0, + "content": "[34] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, pages 4320-4328, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[35] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. arXiv preprint arXiv:2203.08679, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "24265" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_origin.pdf b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f8cd15e83f481aaee24138275a9eca31f3722b28 --- /dev/null +++ b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/37d75dfa-064e-4dc6-9b5a-a491cac2165f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a1f5dd0993e04e1a15bdbcfd8480fa1883102bac57e6ab65679f29bb69628b3 +size 477361 diff --git a/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/full.md b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..506e701db3242c525cd4451de733bf6b2577fe8a --- /dev/null +++ b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/full.md @@ -0,0 +1,287 @@ +# Masked Autoencoders Enable Efficient Knowledge Distillers + +Yutong Bai1 Zeyu Wang2 Junfei Xiao1 Chen Wei1 Huiyu Wang1 Alan Yuille1 Yuyin Zhou2 Cihang Xie2 1Johns Hopkins University 2 University of California, Santa Cruz ytongbai@gmail.com + +# Abstract + +This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, i.e., forward propagate inputs through the first few layers, for obtaining intermediate feature maps. + +Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves $84.0\%$ ImageNet top-1 accuracy, outperforming the baseline of directly distilling a fine-tuned ViT-L by $1.2\%$ . More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with $95\%$ masking ratio where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of $83.6\%$ ; surprisingly, it can still secure $82.4\%$ top-1 ImageNet accuracy by aggressively training with just FOUR visible patches (98% masking ratio). The code and models are publicly available at https://github.com/UCSC-VLAA/DMAE. + +# 1. Introduction + +Following the success in the natural language processing [10, 27], the Transformer architecture is showing tremendous potentials in computer vision [2, 11, 25, 26, 33], especially when they are pre-trained with a huge amount of unlabelled data [3] with self-supervised learning techniques [1, 14]. Masked image modeling, which trains models to predict the masked signals (either as raw pixels or as semantic tokens) of the input image, stands as one of the most powerful ways for feature pre-training. With the most re + +cent representative work in this direction, masked autoencoder (MAE) [13], we are now able to efficiently and effectively pre-train high-capacity Vision Transformers (ViTs) with strong feature representations, leading to state-of-the-art solutions for a wide range of downstream visual tasks. + +In this paper, we are interested in applying knowledge distillation [17], which is one of the most popular model compression techniques, to transfer the knowledge from these strong but cumbersome ViTs into smaller ones. In contrast to prior knowledge distillation works [17, 21, 34], the teacher considered here is a pre-trained model whose predictions do not necessarily reveal the fine-grained relationship between categories; therefore, typical solutions like aligning the soft/hard logits between the teacher model and the student model may no longer remain effective. Moreover, after distilling the pre-trained teacher model, these student models need an extra round of fine-tuning to adapt to downstream tasks. These factors turn distilling pretrained models seemingly a less favorable design choice in terms of both performance and computational cost. + +Nonetheless, surprisingly, we find by building upon MAE, the whole distillation framework can efficiently yield high-performance student models. There are two key designs. Firstly, we follow MAE to let the encoder exclusively operate on a small visible subset of patches and to employ a lightweight decoder for pixel reconstruction. Whereas rather than using the "luxury" setups in MAE, we show aggressively simplifying pre-training from 1600 epochs to 100 epochs and pushing masking ratio from $75\%$ to $95\%$ suffice to distill strong student models. Secondly, instead of aligning logits, we alternatively seek to match the intermediate feature representation; this enables the cumbersome teacher model to only forward propagate inputs through the first few layers, therefore, reducing computations. We note applying L1 norm for distance measure is essential recipe for ensuring a successful intermediate feature alignment. + +We name this distilling MAE framework as DMAE. Compared to the traditional knowledge distillation framework where the teacher is a fine-tuned model, DMAE is more efficient and can train much stronger student models + +![](images/996c5217375d9066e1d46cbc84c713e221d9ebe32bc23f3cfa725364e2451fd1.jpg) +Figure 1. Illustration of the distillation process in DMAE. There are two key designs. Firstly, following MAE, we hereby only take visible patches as inputs and aims to reconstruct the masked ones. Secondly, knowledge distillation is achieved by aligning the intermediate features between the teacher model and the student model. Note the gray blocks denote the dropped high-level layers of the teacher model during distillation. + +at different capacities. For example, by setting ViT-B as the student model, while the baseline of distilling a fine-tuned ViT-L achieves $82.8\%$ top-1 ImageNet accuracy, DMAE substantially boosts the performance to $84.0\%$ $(+1.2\%)$ top-1 ImageNet accuracy, at a even lower training cost (i.e., 195 GPU hours vs. 208 GPU hours, see Table 9). More intriguingly, we found that DMAE allows for robust training with extremely highly masked images—even with TEN visible patches (i.e., $95\%$ masking ratio), ViT-B can competitively attain a top-1 ImageNet accuracy of $83.6\%$ ; this masking ratio can further be aggressively pushed to $98\%$ (FOUR visible patches) where DMAE still help ViT-B secure $82.4\%$ top-1 ImageNet accuracy. We hope this work can benefit future research on efficiently unleashing the power of pretrain models. + +# 2. Related Work + +Knowledge distillation (KD) is a popular model compression technique that allows models to achieve both strong performances of large models and fast inference speed of small models. The first and seminal KD approach, proposed in [17], transfers the "dark knowledge" via minimizing the KL divergence between the soft logits of the teacher model and that of the student model. From then on, many advanced KD methods have been developed, which can be categorized into two branches: logits distillation [8, 12, 21, 30, 34, 35] and intermediate representation distillation [15, 16, 18, 23, 24]. Our DMAE belongs to the second branch, as it minimizes the distance between latent features of the teacher model and those of the student model. + +The first feature-based distillation method is Fit-Nets [23]. In addition to aligning logits, FitNets requires the student model to learn an intermediate representation that is + +predictive of the intermediate representations of the teacher network. Heo et al. [15] re-investigates the design of feature distillation and develops a novel KD method to create a synergy among various aspects, including teacher transform, student transform, distillation feature position, and distance function. CRD [24] incorporates contrastive learning into KD to capture correlations and higher-order output dependencies. Unlike these existing works, our DMAE is the first to consider applying KD to extra information from self-supervised pre-trained models. + +Masked image modeling (MIM) helps models to acquire meaningful representations by reconstructing masked images. The pioneering works are built on denoising autoencoders [28] and context encoders [22]. Following the success of BERT in natural language [10], and also with the recent trend of adopting Transformer [27] to computer vision [11], there have emerged a set of promising works on applying MIM for self-supervised visual pre-training. BEiT [3] first successfully adopts MIM to ViT pre-training by learning to predict visual tokens. MaskFeat [29] finds that learning to reconstruct HOG features enables effective visual representation learning. SimMIM [31] and MAE [13] both propose to directly reconstruct the pixel values of the masked image patches. Our work is built on MAE and finds that MAE enables the whole distillation framework to be efficient and effective. + +# 3. Approach + +# 3.1. Masked Autoencoders + +Our method is built upon MAE, a powerful autoencoder-based MIM approach. Specifically, the MAE encoder first projects unmasked patches to a latent space, which are then fed into the MAE decoder to help predict pixel values of + +masked patches. The core elements in MAE include: + +Masking. MAE operates on image tokens, i.e., the image needs to be divided into non-overlapping patches. A random small subset of those patches will be kept for the MAE encoder, and the rest will be set as the predicting target of the MAE decoder. Typically, a high masking ratio (e.g., $75\%$ ) is applied, preventing models from taking shortcuts (e.g., simply extrapolating missing pixels based on neighbors) in representation learning. + +MAE encoder. The MAE encoder is a standard ViT architecture except that it only operates on those unmasked patches. This design largely reduces the computation cost of encoders. + +MAE decoder. In addition to the encoded features of un-masked patches (from MAE encoder), the MAE decoder receives mask tokens as input, a learned vector shared across all missing positions. The mask token is only used during pre-training, allowing independent decoder design. Particularly, MAE adopts a lightweight decoder for saving computations. + +Reconstruction. Different from BEiT [3] or MaskFeat [29], MAE directly reconstructs image pixel values. The simple mean squared error is applied to masked tokens for calculating loss. + +Note that, other than distillation-related operations, the whole pre-training and fine-tuning process in this paper exactly follow the default setup in MAE, unless specifically mentioned. Interestingly, compared to MAE, our DMAE robustly enables a much more efficient pre-training setup, e.g., 100 (vs. 1600) training epochs and $95\%$ (vs. $75\%$ ) masking ratio. + +# 3.2. Knowledge Distillation + +MAE demonstrates extraordinary capabilities in learning high-capacity models efficiently and effectively. In this work, we seek to combine knowledge distillation with the MAE framework, to efficiently acquire small and fast models with similar performance as those powerful yet cumbersome models. The most straightforward approach is directly applying existing knowledge distillation methods, like the one proposed in DeiT [25], to a fine-tuned MAE model. However, we empirically find that this approach hardly brings in improvements. In addition, this approach fails to leverage the special designs in MAE for reducing computations, e.g., only feeding a small portion of the input image to the encoder. To this end, we hereby study an alternative solution: directly applying knowledge distillation at the pre-training stage. + +Since there are no categorical labels in MAE pretraining, distilling logits can hardly help learn semantically meaningful representations. We, therefore, resort to distilling the intermediate features. This idea is first developed in + +FitNets [23], and inspired a set of followups for advancing knowledge distillation [15, 16, 18, 24]. Concretely, we first extract the features from the specific layers of the student model; after feeding such features into a small project head, the outputs will be asked to mimic the features from the corresponding layers of the teacher model. In practice, the projection head is implemented by simple fully connected layer, which addresses the possible feature dimension mismatch between teacher models and student models. + +Formally, let $\mathbf{x} \in \mathbb{R}^{3HW \times 1}$ be the input pixel RGB values and $\mathbf{y} \in \mathbb{R}^{3HW \times 1}$ be the predicted pixel values, where $H$ denotes image height and $W$ denotes image width. The MAE reconstruction loss $L_{MAE}$ can be written as + +$$ +L _ {M A E} = \frac {1}{\Omega (\mathbf {x} _ {M})} \sum_ {i \in M} \left(\mathbf {y} _ {i} - \mathbf {x} _ {i}\right) ^ {2}. \tag {1} +$$ + +where $M$ denotes the set of masked pixels, $\Omega(.)$ is the number of elements, and $i$ is the pixel index. + +Let $\mathbf{z}_l^S, \mathbf{z}_l^T \in \mathbb{R}^{LC \times 1}$ be the features extracted from the $l$ th layer of the student model and the teacher model, respectively, where $L$ denotes the patch numbers, and $C$ denotes the channel dimension. We use $\sigma()$ to denote the projection network function. Our feature alignment distillation loss $L_{Dist}$ can be written as + +$$ +L _ {D i s t} = \sum_ {l} \frac {1}{\Omega (\mathbf {z} _ {l} ^ {T})} \sum_ {i} \left\| \sigma (\mathbf {z} _ {l} ^ {S}) _ {i} - \mathbf {z} _ {l, i} ^ {T} \right\| _ {1}. \tag {2} +$$ + +The final loss used in pre-training is a weighted summation of MAE reconstruction loss $L_{MAE}$ and the feature alignment distillation loss $L_{Dist}$ , controlled by the hyperparameter $\alpha$ : + +$$ +L = L _ {M A E} + \alpha \times L _ {D i s t}. \tag {3} +$$ + +The framework of DMAE is summarized in Figure 1. Following MAE, DMAE also takes masked inputs and performs the pretext task of masked image modeling. Besides, the corresponding features are aligned between the teacher model and the student model. It is worthy of highlighting that DMAE is an efficient knowledge distiller: 1) it only operates on a tiny subset of visible patches i.e., a high masking ratio is applied; and 2) aligning intermediate layer features reduce the computation cost of (cumbersome) teacher model. In the next section, we extensively compare our method with three baselines: the original MAE without any distillation, DeiT-style distillation, and feature alignment distillation in the supervised setting. + +# 4. Experiments + +# 4.1. Implementation Details + +Following MAE [13], we first perform self-supervised pre-training on ImageNet-1k [9]. Unless otherwise mentioned, the teacher models are public checkpoints released + +
# of LayersLayer LocationStudent Aligned Layer IndexImageNet Top-1 Acc (%)
SingleBottom382.6
Middle683.6
984.0
Top1283.4
MultipleMiddle+Top6+1284.2
+ +from the official MAE implementations1. For pre-training, we train all models using AdamW optimizer [20], with a base learning rate of 1.5e-4, weight decay of 0.05, and optimizer momentum $\beta_{1},\beta_{2} = 0.9,0.95$ . We use a total batch size of 4096, and pre-train models for 100 epochs with a warmup epoch of 20 and a cosine learning rate decay schedule. We by default use the masking ratio of $75\%$ ; while the ablation study shows that our method can robustly tackle extremely high masking ratios. After pre-training, the teacher model is dropped, and we exactly follow the default setups in MAE to fine-tune the student model on ImageNet. + +For feature alignments, we choose to align the features from the $\frac{3}{4}$ depth of both the student model and the teacher model, which we find delivers decent results for all model sizes tested. For example, with a 24-layer ViT-L as the teacher model and a 12-layer ViT-B as the student model, features from the 9th layer of ViT-B are aligned with the features from the 18th layer of ViT-L. We set $\alpha = 1$ in Eq. 3 to balance the tradeoff between MAE reconstruction loss $L_{MAE}$ and the feature alignment distillation loss $L_{Dist}$ in pre-training. + +# 4.2. Analysis + +We first provide a detailed analysis of how to set distillation-related parameters in DMAE. Specifically, we set the teacher model as an MAE pre-trained ViT-L (from MAE official GitHub repository, attaining $85.9\%$ top-1 ImageNet accuracy after fine-tuning), and set the student model as a randomly initialized ViT-B. We analyze the following six factors: + +Where to align. In Table 1 we first check the effect of feature alignment location on model performance. We observe that shallower features are less favored: e.g., the 3rd layer alignment under-performs all other settings. We speculate this is due to the learning process of ViTs—images are much noisier and less semantic than texts, ViTs will first group the raw pixels in the bottom layers (closer to the input), which is harder to transfer. While features from $\frac{3}{4}$ depth (i.e., the 9th layer in ViT-B) achieve the best performance, a simple rule-of-thumb which we find fits all model scales in our experiments. We adopt this design choice in all + +Table 1. The effects of feature alignment location. We hereby test with 5 different layer locations, where top layers refer to those closer to network outputs. For single layer feature alignment, features from the $\frac{3}{4}$ depth of the model leads to the best ImageNet top-1 accuracy. This performance is even comparable to the setup of aligning multiple layers. + +
Teacher Layer (relative position)ImageNet Top-1 Acc (%)
Middle84.0
Top83.3
Bottom82.1
+ +Table 2. The analysis of aligning order on ImageNet classification top-1 accuracy $(\%)$ + +
Projection HeadImageNet Top-1 Acc (%)
Linear84.0
2-layer MLP84.0
3-layer MLP83.8
+ +Table 3. Projection Head. A simple fully-connected layer works the best. We choose this as the default setting. + +other experiments. It is also worth mentioning that simply aligning multiple layers has no clear advantage over aligning features from $\frac{3}{4}$ depth (84.2% vs. 84.0%); we, therefore, stick to the $\frac{3}{4}$ depth setting, which is more efficient. + +Aligning order. We next test the importance of alignment ordering on model performance. Specifically, by fixing the layer location in the student model (i.e., the middle layer in our experiment), we then align it to different layers of the teacher model. As shown in Table 2, we observe that when the aligned layers are in the same relative position (e.g., middle to middle), the student model can achieve the best performance. + +Masking ratio. MAE reveals that the masking ratio in masked image modeling could be surprisingly high (75%). The hypothesis is that by learning to reason about the gestalt of the missing objects and scenes, which cannot be done by extending lines or textures because of the high masking ratio, the model is also learning useful representations. Interestingly, we find that when combining MAE and knowledge distillation, an even much higher masking ratio is possible, as shown in Figure 2. + +Firstly, it is interesting to note that, compared to the typical $75\%$ masking ratio setting, further raising the masking ratio to $90\%$ comes at no performance drop, i.e., both attain $84.0\%$ top-1 ImageNet accuracy. Next, even with an extremely large masking ratio like $98\%$ (only FOUR visible patches), DMAE still beats the 100-epoch MAE baseline + +![](images/339f5a89158840705c7d8d3a45b2f9dc22625001b6cb03b561f002fbfacc943d.jpg) +Figure 2. DMAE allows an extremely high masking ratio. From left to right, we increase the masking ratio from the basic $75\%$ to the extreme $99\%$ . We note that our DMAE competitively attains $83.6\%$ top-1 ImageNet accuracy with $95\%$ masking ratio (TEN visible patches), and still secures $82.4\%$ top-1 ImageNet accuracy even by learning with FOUR visible patches ( $98\%$ masking ratio). + +
Decoder DepthImageNet Top-1 Acc (%)
283.7
483.8
884.0
+ +Table 4. Decoder Depth. A deeper decoder (slightly) improves the pre-trained representation quality. + +
Loss DesignImageNet Top-1 Acc (%)
Loss ChoiceL1 with α = 184.0
L2 with α = 183.3
Loss RatioL1 with α = 0.583.9
L1 with α = 184.0
L1 with α = 284.0
L1 with α = 484.0
+ +Table 5. Loss Function. From the first block, we can observe that L1 distance yields significantly higher performance than L2 distance. From the second block, we can observe that the hyperparameter $\alpha$ has little influence on the representation quality of DMAE. + +that uses a masking ratio of $75\%$ (second row in Table 6), by a non-trivial-margin $(82.4\% \text{ vs. } 81.6\%)$ . These results suggest that, with the assistance of distilled knowledge from the teacher model, the student model can make better use of visible patches, even at a very limited amount, for representation learning. + +Projection head. The goal of the proposed feature alignment distillation is to encourage the student model to learn features that are predictive of features from a stronger teacher model. To that end, a small projection head is employed on features from the student model, to 1) project them onto a space of the same dimension as the hidden dimension of the teacher model, and 2) provide extra flexibility for feature alignment. We ablate the choice of this pro + +jection network, as shown in Table 3. We can observe that applying a simple fully connected layer already performs the best among other choices. + +Decoder depth. In Table 4 we analyze the effect of decoder depth. Similar to MAE [13], the final performance gets (slightly) increased with a deeper decoder. We choose a decoder depth of 8 as the default setting as in [13]. Note that a decoder depth of 2 is also a competitive choice—compared to an 8-depth decoder, it significantly reduces the computation cost while only marginally sacrificing the accuracy by $0.3\%$ . + +Loss designs. Table 5 ablates the loss design. While SimMIM [31] shows that L1 distance and L2 distance lead to similar performance, ours suggests that L1 distance exhibits a clear advantage over L2 distance, i.e., $+0.7\%$ improvement. Furthermore, we note DMAE is quite robust to the specific value of the hyperparameter $\alpha$ , which controls the relative importance of the distillation loss over the reconstruction loss. Based on these results, we choose L1 in Eq. 2 for distance measure and set $\alpha = 1$ in Eq. 3 for the rest experiments. + +# 4.3. Comparison with Baselines. + +In Table 6, we compare the performance of our DMAE with various baselines: + +MAE. The MAE baselines are presented in the first block of Table 6. If MAE is also asked to pre-train for only 100 epochs, DMAE can substantially outperform this baseline by $2.4\%$ (from $81.6\%$ to $84.0\%$ ). When comparing to a much stronger but more computationally expensive MAE baseline with 1600 pre-training epoch, we note DMAE still beats it by $0.4\%$ . + +Supervised model. The second block in Table 6 demonstrates the effectiveness of DMAE compared with models + +
MethodPre-training epochsSupervised training / fine-tuning epochsImageNet Top-1 Acc (%)
MAE-B10010081.6
MAE-B160010083.6
DeiT-B-10076.8
DeiT-B-30081.8
DeiT-B-Soft Distillation-10077.5
DeiT-B-Hard Distillation-10078.3
CRD [24]-10081.9
SRRL [32]-10082.2
Dear-KD [6]-10082.4
Supervised Feature Alignment-10082.8
DMAE-B10010084.0
+ +trained under the supervision of categorical labels, which requires a much longer training time, i.e., $+2.2\%$ compared to the DeiT 300 epochs supervised training. + +Other distillation strategies. We next compare DMAE with other distillation methods. We consider DeiT-style logit-based distillation [25], CRD [24], SRRL [32], DearKD [6], and feature alignment distillation. Note that for these baselines, one significant difference from DMAE is that the student model here directly distill knowledge from a supervisely fine-tuned teacher model. Moreover, to make these baselines more competitive, the teacher models will first be MAE pre-trained and then fine-tuned on ImageNet-1k. + +The results are shown in the third block of Table 6. Firstly, we can observe that the DeiT-style logit-based distillation, either soft or hard, even hurts the student models' performance. This phenomenon potentially suggests that such a distillation strategy may not fit teacher models of ViT architectures. For other baselines, we note that feature alignment distillation performs the best; but this is still worse than DMAE (82.8% vs. 84.0%), indicating the importance and effectiveness of distilling knowledge from a pre-trained teacher model. + +# 4.4. Scaling to Different Model Sizes + +We test DMAE with different model sizes, listed in Table 7. For a fair comparison, both methods only pre-train models for 100 epochs. DMAE shows consistent improvement compared to MAE across different model sizes. With only one middle layer feature alignment, DMAE brings an additional improvement of $+2.4\%$ with ViT-B, $+2.7\%$ with ViT-Small, and $+3.4\%$ with ViT-Tiny. In addition, we are interested in the following two cases: + +Same teacher model, different student model. As shown in the first two lines in Table 7, we find that when using ViT-L as the teacher model, both ViT-B and ViT-S benefit from the distillation, demonstrating a clear advantage over + +Table 6. DMAE shows stronger performance than all three kinds of baselines: MAE, supervised model, and other existing advanced distillation strategies. + +
Student ModelTeacher ModelImageNet Top-1 Acc (%)
MAEDMAE
BaseLarge81.684.0 (+2.4)
SmallLarge77.480.1 (+2.7)
Base77.479.3 (+1.9)
TinyBase66.670.0 (+3.4)
+ +Table 7. Across different model sizes, DMAE shows consistent improvements compared with MAE. + +the MAE baseline. We argue that the ability to effectively generalize to cases where an even smaller student model is desirable, especially for those computation-constrained real-world applications. + +Same student model, different teacher model. As shown in Table 7, with a ViT-S as the student model, enlarging the teacher model from ViT-B to ViT-L further boosts the accuracy by $0.8\%$ (from $79.3\%$ to $80.1\%$ ). This result suggests that DMAE can effectively distill knowledge from the teacher models at different scales. + +# 4.5. Limited Training Data + +In certain real-world applications, data could be hard to acquire because of high data collection and labeling costs or due to privacy concerns. Leveraging models pre-trained on large-scale unlabeled datasets for fine-tuning when only a small dataset of downstream tasks is available becomes a promising solution. Here we seek to test the potential of our DMAE in this data-scarce scenario. We strictly follow [5] to sample $1\%$ or $10\%$ of the labeled ILSVRC-12 training datasets in a class-balanced way. We set MAE pre-trained ViT-L as the teacher model and a randomly initialized ViT-B as the student model. In addition, we compare DMAE with three kinds of baselines described in Section 4.3: MAE, supervised model, and other distillation strategies, and similarly, set the teacher model to be a ViT-L that first pre-trained with MAE and then fine + +
MethodPre-training epochsSupervised training / fine-tuning epochsIN-1% Top-1 Acc(%)IN-10% Top-1 Acc(%)
MAE-B10010033.965.0
MAE-B160010049.672.8
DeiT-B-100--
DeiT-B-Soft Distillation160010036.066.4
DeiT-B-Hard Distillation160010037.367.3
Supervised Feature Alignment160010034.267.6
DMAE-B10010050.373.4
+ +Table 8. DMAE demonstrates much stronger performance than all other baselines when training data is limited. Note that the DeiT-B baseline is unable to converge because of data insufficiency. + +
ModelTraining Cost (GPU Hours)ImageNet Top-1 Acc(%)
Pre-trainingFine-tuningOverall
MAE-B-100-epoch78h112h190h81.6
MAE-B-1600-epoch1248h112h1360h83.6
DeiT-B-Soft Distillation-213h213h77.5
DeiT-B-Hard Distillation-213h213h78.3
Supervised Feature Alignment-208h208h82.8
DMAE-B83h112h195h84.0
+ +tuned on ImageNet-1k. Note that since DMAE has full access to the $100\%$ ImageNet dataset (without labels) during pre-training, to ensure a fair and competitive comparison, we initialize all the baselines as the 1600-epoch MAE pretrained model on ImageNet. + +Table 8 shows that DMAE largely surpasses all other baselines. For example, when only $10\%$ ImageNet data is available for supervised training or fine-tuning, DMAE outperforms the MAE pre-trained baseline by $8.4\%$ (i.e., $73.4\%$ vs. $65.0\%$ ). DMAE also significantly outperforms other distillation strategies, with an improvement ranging from $5.8\%$ to $7.0\%$ . We note this accuracy gap is even larger when only $1\%$ ImageNet is available, demonstrating the data efficiency of DMAE. + +# 4.6. Generalization to Other Methods + +Lastly, we provide preliminary results of integrating DMAE into other self-supervised training frameworks, including DINO [4] and MoCo-V3 [7]. Unlike MAE, which belongs to masked image modeling, DINO and MoCo-V3 are contrastive learning-based methods. Still, as shown in Table 10, without further hyperparameter tuning, DMAE effectively shows non-trivial improvements on top of both DINO (+1.3%) and MoCo-v3 (+1.4%), demonstrating the potential of distilling pre-trained models (rather than fine-tuned models as in most existing knowledge distillation frameworks). + +Table 9. Computational cost comparisons among DMAE and other baselines. The training cost is measured by A5000 GPU hours. We note the proposed DMAE maintains a similar (or even cheaper) training cost than others, while achieving much higher top-1 ImageNet accuracy. + +
w/o DMAEw/ DMAE
DINO80.982.2 (+1.3)
MoCo-V381.182.5 (+1.4)
+ +# 4.7. Generalization to Downstream tasks + +Following ViT-Det [19], we conduct downstream fin-tuning on the COCO dataset, where MAE/DMAE pretrained ViT-Base model is adopted as the plain backbone of Mask-RCNN. We train the model on the train2017 split and evaluate it on the val2017 split. We report results on bounding box object detection $(\mathrm{AP}^{\mathrm{box}})$ and instance segmentation $(\mathrm{AP}^{\mathrm{mask}})$ , shown in Table 11. Our observations indicate that DMAE also demonstrates strong potential in downstream tasks like detection and segmentation. + +Table 10. DMAE effectively improves other self-supervised pre-training frameworks (including DINO and MoCo-v3) on ImageNet classification top-1 accuracy $(\%)$ . + +
method\(AP^{box}\)\(AP^{mask}\)
MAE51.245.5
DMAE53.446.9
+ +Table 11. DMAE effectively improves the downstream tasks as well, including object detection and segmentation. + +# 5. Discussion + +In this section, we present a quantitative evaluation of the computational cost of DMAE compared to baseline methods. Additionally, we conduct a standard deviation analysis to assess the stability of our DMAE approach. Finally, we propose an advanced fine-tuning recipe for applying DMAE to smaller ViT models. + +# 5.1. Computational Costs + +In Table 9, we provide a quantitative evaluation on the computational cost, which is tested on a single NVIDIA A5000 GPU. We can observe that, without a significantly increasing of training hours, DMAE substantially outperforms the MAE-100 baseline by $+2.4\%$ on ImageNet; this result even exceeds the MAE-1600 baseline, which causes $\sim 7x$ GPU hours than MAE-100 baseline. We additionally provide the actual GPU hours of other baselines, and find that the proposed DMAE stands as the most efficient one, meanwhile achieving the best top-1 ImageNet accuracy. + +# 5.2. Standard Deviation Analysis + +In the above experiments, we kept the same random seed. Following MAE, we perform the statistical analysis for DMAE by changing the random seeds. In Table 12, from top to down, we show three aligning settings; and from left to right, we show the results with the default seed, the average accuracy with three randomly sampled seeds, and their standard deviation, respectively. From these results, we could conclude that our DMAE can bring in statistically stable improvements. + +
PosAcc(%)AvgStandard Deviation
Bottom (3th)82.682.500.20
Mid (6th)83.683.670.08
Top (9th)84.084.030.10
+ +# 5.3. Smaller ViT Models. + +Since the original MAE paper does not offer specialized recipes for ViT-Small and Tiny, we by default use the recipe for ViT-Base to fine-tune these smaller ViTs. However, this recipe includes strong regularization and augmentation techniques that might lead to over-regularization for the smaller ViTs. To address this issue, we experiment with a modified recipe with weaker augmentation and regularization by removing MixUp, CutMix, and Stochastic Depth. + +Table 12. Standard deviation analysis for DMAE. From top to bottom, we show three aligning settings, the model performance with the default seed, the average performance with three randomly sampled seeds, and their standard deviations, respectively. 'Pos' denotes the student distillation position. Acc denotes ImageNet Top-1 accuracy. 'Avg' denotes the average value over three times. + +
ViT-TinyTop-1 Acc (%)
MAE70.1
DeiT74.5
DMAE74.9
+ +(a) ViT-Tiny + +
ViT-SmallTop-1 Acc (%)
MAE80.0
DeiT81.2
DMAE82.2
+ +(b) ViT-Small + +Table 13. Weaker augmentation and regularization helps smaller ViT models, during finetuning for both MAE and DMAE on ImageNet classification top-1 accuracy $(\%)$ + +Results on ImageNet in Table 13 show that DMAE not only maintains its advantage over MAE, but also outperforms DeiT with this new recipe, highlighting its effectiveness. + +# 6. Conclusion + +Self-supervised pre-training has demonstrated great success for those exponentially growing models in the natural language domain. Recently, the rise of MAE shows that a similar paradigm also works for the computer vision domain, and now the development of vision models may embark on a similar trajectory as in the language domain. Yet, it is often desirable to have a well-balanced model between performance and speed in real-world applications. This work is a small step towards unleashing the potential of knowledge distillation, a popular model compression technique, within the MAE framework. Our DMAE is a simple, efficient, and effective knowledge distillation method: feature alignment during MAE pre-training. Extensive experiments on multiple model scales demonstrate the effectiveness of our approach. Moreover, an intriguing finding is that it allows for a masking ratio even higher than the already large one used in MAE (i.e., $75\%$ ). We have also validated the effectiveness of our DMAE when in the small-data regime. We hope this work can benefit future research in knowledge distillation with pre-trained models. + +# Acknowledgement + +This work is supported by a gift from Open Philanthropy, TPU Research Cloud (TRC) program, and Google Cloud Research Credits program. We also want to acknowledge ONR with award N00014-21-1-2812. + +# References + +[1] Yutong Bai, Xinlei Chen, Alexander Kirillov, Alan Yuille, and Alexander C Berg. Point-level region contrast for object detection pre-training. In Proceedings of the IEEE/CVF Con- + +ference on Computer Vision and Pattern Recognition, pages 16061-16070, 2022. 1 +[2] Yutong Bai, Haoqi Fan, Ishan Misra, Ganesh Venkatesh, Yongyi Lu, Yuyin Zhou, Qihang Yu, Vikas Chandra, and Alan Yuille. Can temporal information help with contrastive self-supervised learning? arXiv preprint arXiv:2011.13046, 2020. 1 +[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2021. 1, 2, 3 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, pages 9650-9660, 2021. 7 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 6 +[6] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. In CVPR, pages 12052-12062, 2022. 6 +[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 7 +[8] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In CVPR, pages 4794-4802, 2019. 2 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 3 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2 +[12] Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. In ICML, pages 1607-1616. PMLR, 2018. 2 +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 2, 3, 5 +[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 1 +[15] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In CVPR, pages 1921-1930, 2019. 2, 3 +[16] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In AAAI, volume 33, pages 3779-3787, 2019. 2, 3 + +[17] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 1, 2 +[18] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. NeurIPS, 31, 2018. 2, 3 +[19] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 280-296. Springer, 2022. 7 +[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 4 +[21] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, volume 34, pages 5191-5198, 2020. 1, 2 +[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544, 2016. 2 +[23] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2, 3 +[24] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 2, 3, 6 +[25] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 1, 3, 6 +[26] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In CVPR, pages 32-42, 2021. 1 +[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2 +[28] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 11(12), 2010. 2 +[29] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2, 3 +[30] Junfei Xiao, Longlong Jing, Lin Zhang, Ju He, Qi She, Zongwei Zhou, Alan Yuille, and Yingwei Li. Learning from temporal gradient for semi-supervised action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3252-3262, 2022. 2 +[31] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2, 5 + +[32] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 6 +[33] Qihang Yu, Yingda Xia, Yutong Bai, Yongyi Lu, Alan L Yuille, and Wei Shen. Glance-and-gaze vision transformer. Advances in Neural Information Processing Systems, 34:12992-13003, 2021. 1 +[34] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, pages 4320-4328, 2018. 1, 2 +[35] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. arXiv preprint arXiv:2203.08679, 2022. 2 \ No newline at end of file diff --git a/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/images.zip b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2bb7fbcd3184dcf78f1679329799924649acc6d3 --- /dev/null +++ b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57ec1550a709f3a531e0b596cbccd54ed853da6d7a357fe4318e15557f371266 +size 345470 diff --git a/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/layout.json b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d30c5f176c1593ff865fc1ba717f6117709fd2f2 --- /dev/null +++ b/2023/Masked Autoencoders Enable Efficient Knowledge Distillers/layout.json @@ -0,0 +1,7810 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 483, + 121 + ], + "type": "text", + "content": "Masked Autoencoders Enable Efficient Knowledge Distillers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 142, + 457, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 142, + 457, + 199 + ], + "spans": [ + { + "bbox": [ + 136, + 142, + 457, + 199 + ], + "type": "text", + "content": "Yutong Bai1 Zeyu Wang2 Junfei Xiao1 Chen Wei1 Huiyu Wang1 Alan Yuille1 Yuyin Zhou2 Cihang Xie2 1Johns Hopkins University 2 University of California, Santa Cruz ytongbai@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 251, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 251, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 45, + 251, + 287, + 384 + ], + "type": "text", + "content": "This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, i.e., forward propagate inputs through the first few layers, for obtaining intermediate feature maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "text", + "content": "Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves " + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "text", + "content": " ImageNet top-1 accuracy, outperforming the baseline of directly distilling a fine-tuned ViT-L by " + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "text", + "content": ". More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with " + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "text", + "content": " masking ratio where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of " + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "inline_equation", + "content": "83.6\\%" + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "text", + "content": "; surprisingly, it can still secure " + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "inline_equation", + "content": "82.4\\%" + }, + { + "bbox": [ + 46, + 384, + 289, + 563 + ], + "type": "text", + "content": " top-1 ImageNet accuracy by aggressively training with just FOUR visible patches (98% masking ratio). The code and models are publicly available at https://github.com/UCSC-VLAA/DMAE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Following the success in the natural language processing [10, 27], the Transformer architecture is showing tremendous potentials in computer vision [2, 11, 25, 26, 33], especially when they are pre-trained with a huge amount of unlabelled data [3] with self-supervised learning techniques [1, 14]. Masked image modeling, which trains models to predict the masked signals (either as raw pixels or as semantic tokens) of the input image, stands as one of the most powerful ways for feature pre-training. With the most re" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "text", + "content": "cent representative work in this direction, masked autoencoder (MAE) [13], we are now able to efficiently and effectively pre-train high-capacity Vision Transformers (ViTs) with strong feature representations, leading to state-of-the-art solutions for a wide range of downstream visual tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": "In this paper, we are interested in applying knowledge distillation [17], which is one of the most popular model compression techniques, to transfer the knowledge from these strong but cumbersome ViTs into smaller ones. In contrast to prior knowledge distillation works [17, 21, 34], the teacher considered here is a pre-trained model whose predictions do not necessarily reveal the fine-grained relationship between categories; therefore, typical solutions like aligning the soft/hard logits between the teacher model and the student model may no longer remain effective. Moreover, after distilling the pre-trained teacher model, these student models need an extra round of fine-tuning to adapt to downstream tasks. These factors turn distilling pretrained models seemingly a less favorable design choice in terms of both performance and computational cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 663 + ], + "type": "text", + "content": "Nonetheless, surprisingly, we find by building upon MAE, the whole distillation framework can efficiently yield high-performance student models. There are two key designs. Firstly, we follow MAE to let the encoder exclusively operate on a small visible subset of patches and to employ a lightweight decoder for pixel reconstruction. Whereas rather than using the \"luxury\" setups in MAE, we show aggressively simplifying pre-training from 1600 epochs to 100 epochs and pushing masking ratio from " + }, + { + "bbox": [ + 304, + 472, + 547, + 663 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 472, + 547, + 663 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 472, + 547, + 663 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 472, + 547, + 663 + ], + "type": "text", + "content": " suffice to distill strong student models. Secondly, instead of aligning logits, we alternatively seek to match the intermediate feature representation; this enables the cumbersome teacher model to only forward propagate inputs through the first few layers, therefore, reducing computations. We note applying L1 norm for distance measure is essential recipe for ensuring a successful intermediate feature alignment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "We name this distilling MAE framework as DMAE. Compared to the traditional knowledge distillation framework where the teacher is a fine-tuned model, DMAE is more efficient and can train much stronger student models" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24256" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 77, + 527, + 225 + ], + "blocks": [ + { + "bbox": [ + 66, + 77, + 527, + 225 + ], + "lines": [ + { + "bbox": [ + 66, + 77, + 527, + 225 + ], + "spans": [ + { + "bbox": [ + 66, + 77, + 527, + 225 + ], + "type": "image", + "image_path": "996c5217375d9066e1d46cbc84c713e221d9ebe32bc23f3cfa725364e2451fd1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 547, + 278 + ], + "type": "text", + "content": "Figure 1. Illustration of the distillation process in DMAE. There are two key designs. Firstly, following MAE, we hereby only take visible patches as inputs and aims to reconstruct the masked ones. Secondly, knowledge distillation is achieved by aligning the intermediate features between the teacher model and the student model. Note the gray blocks denote the dropped high-level layers of the teacher model during distillation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "text", + "content": "at different capacities. For example, by setting ViT-B as the student model, while the baseline of distilling a fine-tuned ViT-L achieves " + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "inline_equation", + "content": "82.8\\%" + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "text", + "content": " top-1 ImageNet accuracy, DMAE substantially boosts the performance to " + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "inline_equation", + "content": "(+1.2\\%)" + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "text", + "content": " top-1 ImageNet accuracy, at a even lower training cost (i.e., 195 GPU hours vs. 208 GPU hours, see Table 9). More intriguingly, we found that DMAE allows for robust training with extremely highly masked images—even with TEN visible patches (i.e., " + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "text", + "content": " masking ratio), ViT-B can competitively attain a top-1 ImageNet accuracy of " + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "inline_equation", + "content": "83.6\\%" + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "text", + "content": "; this masking ratio can further be aggressively pushed to " + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "text", + "content": " (FOUR visible patches) where DMAE still help ViT-B secure " + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "inline_equation", + "content": "82.4\\%" + }, + { + "bbox": [ + 46, + 299, + 289, + 478 + ], + "type": "text", + "content": " top-1 ImageNet accuracy. We hope this work can benefit future research on efficiently unleashing the power of pretrain models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 675 + ], + "type": "text", + "content": "Knowledge distillation (KD) is a popular model compression technique that allows models to achieve both strong performances of large models and fast inference speed of small models. The first and seminal KD approach, proposed in [17], transfers the \"dark knowledge\" via minimizing the KL divergence between the soft logits of the teacher model and that of the student model. From then on, many advanced KD methods have been developed, which can be categorized into two branches: logits distillation [8, 12, 21, 30, 34, 35] and intermediate representation distillation [15, 16, 18, 23, 24]. Our DMAE belongs to the second branch, as it minimizes the distance between latent features of the teacher model and those of the student model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "The first feature-based distillation method is Fit-Nets [23]. In addition to aligning logits, FitNets requires the student model to learn an intermediate representation that is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 299, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 547, + 418 + ], + "type": "text", + "content": "predictive of the intermediate representations of the teacher network. Heo et al. [15] re-investigates the design of feature distillation and develops a novel KD method to create a synergy among various aspects, including teacher transform, student transform, distillation feature position, and distance function. CRD [24] incorporates contrastive learning into KD to capture correlations and higher-order output dependencies. Unlike these existing works, our DMAE is the first to consider applying KD to extra information from self-supervised pre-trained models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "type": "text", + "content": "Masked image modeling (MIM) helps models to acquire meaningful representations by reconstructing masked images. The pioneering works are built on denoising autoencoders [28] and context encoders [22]. Following the success of BERT in natural language [10], and also with the recent trend of adopting Transformer [27] to computer vision [11], there have emerged a set of promising works on applying MIM for self-supervised visual pre-training. BEiT [3] first successfully adopts MIM to ViT pre-training by learning to predict visual tokens. MaskFeat [29] finds that learning to reconstruct HOG features enables effective visual representation learning. SimMIM [31] and MAE [13] both propose to directly reconstruct the pixel values of the masked image patches. Our work is built on MAE and finds that MAE enables the whole distillation framework to be efficient and effective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 627, + 373, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 627, + 373, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 627, + 373, + 640 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "type": "text", + "content": "3.1. Masked Autoencoders" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "Our method is built upon MAE, a powerful autoencoder-based MIM approach. Specifically, the MAE encoder first projects unmasked patches to a latent space, which are then fed into the MAE decoder to help predict pixel values of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24257" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 262, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 262, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 262, + 83 + ], + "type": "text", + "content": "masked patches. The core elements in MAE include:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 88, + 287, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 88, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 46, + 88, + 287, + 184 + ], + "type": "text", + "content": "Masking. MAE operates on image tokens, i.e., the image needs to be divided into non-overlapping patches. A random small subset of those patches will be kept for the MAE encoder, and the rest will be set as the predicting target of the MAE decoder. Typically, a high masking ratio (e.g., " + }, + { + "bbox": [ + 46, + 88, + 287, + 184 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 88, + 287, + 184 + ], + "type": "text", + "content": ") is applied, preventing models from taking shortcuts (e.g., simply extrapolating missing pixels based on neighbors) in representation learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 187, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 287, + 235 + ], + "type": "text", + "content": "MAE encoder. The MAE encoder is a standard ViT architecture except that it only operates on those unmasked patches. This design largely reduces the computation cost of encoders." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 323 + ], + "type": "text", + "content": "MAE decoder. In addition to the encoded features of un-masked patches (from MAE encoder), the MAE decoder receives mask tokens as input, a learned vector shared across all missing positions. The mask token is only used during pre-training, allowing independent decoder design. Particularly, MAE adopts a lightweight decoder for saving computations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "type": "text", + "content": "Reconstruction. Different from BEiT [3] or MaskFeat [29], MAE directly reconstructs image pixel values. The simple mean squared error is applied to masked tokens for calculating loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "content": "Note that, other than distillation-related operations, the whole pre-training and fine-tuning process in this paper exactly follow the default setup in MAE, unless specifically mentioned. Interestingly, compared to MAE, our DMAE robustly enables a much more efficient pre-training setup, e.g., 100 (vs. 1600) training epochs and " + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "content": " (vs. " + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "content": ") masking ratio." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 467, + 176, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 176, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 176, + 480 + ], + "type": "text", + "content": "3.2. Knowledge Distillation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 665 + ], + "type": "text", + "content": "MAE demonstrates extraordinary capabilities in learning high-capacity models efficiently and effectively. In this work, we seek to combine knowledge distillation with the MAE framework, to efficiently acquire small and fast models with similar performance as those powerful yet cumbersome models. The most straightforward approach is directly applying existing knowledge distillation methods, like the one proposed in DeiT [25], to a fine-tuned MAE model. However, we empirically find that this approach hardly brings in improvements. In addition, this approach fails to leverage the special designs in MAE for reducing computations, e.g., only feeding a small portion of the input image to the encoder. To this end, we hereby study an alternative solution: directly applying knowledge distillation at the pre-training stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Since there are no categorical labels in MAE pretraining, distilling logits can hardly help learn semantically meaningful representations. We, therefore, resort to distilling the intermediate features. This idea is first developed in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "type": "text", + "content": "FitNets [23], and inspired a set of followups for advancing knowledge distillation [15, 16, 18, 24]. Concretely, we first extract the features from the specific layers of the student model; after feeding such features into a small project head, the outputs will be asked to mimic the features from the corresponding layers of the teacher model. In practice, the projection head is implemented by simple fully connected layer, which addresses the possible feature dimension mismatch between teacher models and student models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "content": "Formally, let " + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{3HW \\times 1}" + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "content": " be the input pixel RGB values and " + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{3HW \\times 1}" + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "content": " be the predicted pixel values, where " + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "content": " denotes image height and " + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "content": " denotes image width. The MAE reconstruction loss " + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "inline_equation", + "content": "L_{MAE}" + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 232, + 545, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 232, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 353, + 232, + 545, + 261 + ], + "type": "interline_equation", + "content": "L _ {M A E} = \\frac {1}{\\Omega (\\mathbf {x} _ {M})} \\sum_ {i \\in M} \\left(\\mathbf {y} _ {i} - \\mathbf {x} _ {i}\\right) ^ {2}. \\tag {1}", + "image_path": "6e192ef0fb36a4f5f0ce6b26181a3928fd55d281539f3b8c7ebbe9d3c0054c6d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "type": "text", + "content": " denotes the set of masked pixels, " + }, + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "type": "inline_equation", + "content": "\\Omega(.)" + }, + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "type": "text", + "content": " is the number of elements, and " + }, + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 267, + 545, + 290 + ], + "type": "text", + "content": " is the pixel index." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_l^S, \\mathbf{z}_l^T \\in \\mathbb{R}^{LC \\times 1}" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " be the features extracted from the " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": "th layer of the student model and the teacher model, respectively, where " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " denotes the patch numbers, and " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " denotes the channel dimension. We use " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\sigma()" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " to denote the projection network function. Our feature alignment distillation loss " + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "L_{Dist}" + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 334, + 367, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 367, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 334, + 367, + 545, + 396 + ], + "type": "interline_equation", + "content": "L _ {D i s t} = \\sum_ {l} \\frac {1}{\\Omega (\\mathbf {z} _ {l} ^ {T})} \\sum_ {i} \\left\\| \\sigma (\\mathbf {z} _ {l} ^ {S}) _ {i} - \\mathbf {z} _ {l, i} ^ {T} \\right\\| _ {1}. \\tag {2}", + "image_path": "d2fe661d33de3a9560c34eead2b5ec72529471dfd2a5a46b2b2e28520de58a13.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "type": "text", + "content": "The final loss used in pre-training is a weighted summation of MAE reconstruction loss " + }, + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "type": "inline_equation", + "content": "L_{MAE}" + }, + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "type": "text", + "content": " and the feature alignment distillation loss " + }, + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "type": "inline_equation", + "content": "L_{Dist}" + }, + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "type": "text", + "content": ", controlled by the hyperparameter " + }, + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 400, + 545, + 449 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 455, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 455, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 372, + 455, + 545, + 468 + ], + "type": "interline_equation", + "content": "L = L _ {M A E} + \\alpha \\times L _ {D i s t}. \\tag {3}", + "image_path": "4b70ec088eb5ded432824e35b0d07434ef9d8e90b25661d6368fa480b4bcd137.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 475, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 630 + ], + "type": "text", + "content": "The framework of DMAE is summarized in Figure 1. Following MAE, DMAE also takes masked inputs and performs the pretext task of masked image modeling. Besides, the corresponding features are aligned between the teacher model and the student model. It is worthy of highlighting that DMAE is an efficient knowledge distiller: 1) it only operates on a tiny subset of visible patches i.e., a high masking ratio is applied; and 2) aligning intermediate layer features reduce the computation cost of (cumbersome) teacher model. In the next section, we extensively compare our method with three baselines: the original MAE without any distillation, DeiT-style distillation, and feature alignment distillation in the supervised setting." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 659, + 440, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 440, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 440, + 672 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Following MAE [13], we first perform self-supervised pre-training on ImageNet-1k [9]. Unless otherwise mentioned, the teacher models are public checkpoints released" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24258" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 70, + 468, + 141 + ], + "blocks": [ + { + "bbox": [ + 124, + 70, + 468, + 141 + ], + "lines": [ + { + "bbox": [ + 124, + 70, + 468, + 141 + ], + "spans": [ + { + "bbox": [ + 124, + 70, + 468, + 141 + ], + "type": "table", + "html": "
# of LayersLayer LocationStudent Aligned Layer IndexImageNet Top-1 Acc (%)
SingleBottom382.6
Middle683.6
984.0
Top1283.4
MultipleMiddle+Top6+1284.2
", + "image_path": "43ea7bc82a4a166b501a40331fab9ba8e57b11cc47599155ea66b017c132e324.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 198, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 288, + 331 + ], + "type": "text", + "content": "from the official MAE implementations1. For pre-training, we train all models using AdamW optimizer [20], with a base learning rate of 1.5e-4, weight decay of 0.05, and optimizer momentum " + }, + { + "bbox": [ + 46, + 198, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\beta_{2} = 0.9,0.95" + }, + { + "bbox": [ + 46, + 198, + 288, + 331 + ], + "type": "text", + "content": ". We use a total batch size of 4096, and pre-train models for 100 epochs with a warmup epoch of 20 and a cosine learning rate decay schedule. We by default use the masking ratio of " + }, + { + "bbox": [ + 46, + 198, + 288, + 331 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 198, + 288, + 331 + ], + "type": "text", + "content": "; while the ablation study shows that our method can robustly tackle extremely high masking ratios. After pre-training, the teacher model is dropped, and we exactly follow the default setups in MAE to fine-tune the student model on ImageNet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": "For feature alignments, we choose to align the features from the " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\frac{3}{4}" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": " depth of both the student model and the teacher model, which we find delivers decent results for all model sizes tested. For example, with a 24-layer ViT-L as the teacher model and a 12-layer ViT-B as the student model, features from the 9th layer of ViT-B are aligned with the features from the 18th layer of ViT-L. We set " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": " in Eq. 3 to balance the tradeoff between MAE reconstruction loss " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "L_{MAE}" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": " and the feature alignment distillation loss " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "L_{Dist}" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": " in pre-training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 458, + 111, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 111, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 111, + 471 + ], + "type": "text", + "content": "4.2. Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 476, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 559 + ], + "type": "text", + "content": "We first provide a detailed analysis of how to set distillation-related parameters in DMAE. Specifically, we set the teacher model as an MAE pre-trained ViT-L (from MAE official GitHub repository, attaining " + }, + { + "bbox": [ + 46, + 476, + 287, + 559 + ], + "type": "inline_equation", + "content": "85.9\\%" + }, + { + "bbox": [ + 46, + 476, + 287, + 559 + ], + "type": "text", + "content": " top-1 ImageNet accuracy after fine-tuning), and set the student model as a randomly initialized ViT-B. We analyze the following six factors:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "content": "Where to align. In Table 1 we first check the effect of feature alignment location on model performance. We observe that shallower features are less favored: e.g., the 3rd layer alignment under-performs all other settings. We speculate this is due to the learning process of ViTs—images are much noisier and less semantic than texts, ViTs will first group the raw pixels in the bottom layers (closer to the input), which is harder to transfer. While features from " + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "inline_equation", + "content": "\\frac{3}{4}" + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "content": " depth (i.e., the 9th layer in ViT-B) achieve the best performance, a simple rule-of-thumb which we find fits all model scales in our experiments. We adopt this design choice in all" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 197, + 543, + 246 + ], + "blocks": [ + { + "bbox": [ + 46, + 144, + 547, + 179 + ], + "lines": [ + { + "bbox": [ + 46, + 144, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 547, + 179 + ], + "type": "text", + "content": "Table 1. The effects of feature alignment location. We hereby test with 5 different layer locations, where top layers refer to those closer to network outputs. For single layer feature alignment, features from the " + }, + { + "bbox": [ + 46, + 144, + 547, + 179 + ], + "type": "inline_equation", + "content": "\\frac{3}{4}" + }, + { + "bbox": [ + 46, + 144, + 547, + 179 + ], + "type": "text", + "content": " depth of the model leads to the best ImageNet top-1 accuracy. This performance is even comparable to the setup of aligning multiple layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 197, + 543, + 246 + ], + "lines": [ + { + "bbox": [ + 307, + 197, + 543, + 246 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 543, + 246 + ], + "type": "table", + "html": "
Teacher Layer (relative position)ImageNet Top-1 Acc (%)
Middle84.0
Top83.3
Bottom82.1
", + "image_path": "db6481f8f95fa365381e9ebb9f41763c5f99d555b24ed77b5a1fa4dab2dd5ec3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 337, + 283, + 514, + 331 + ], + "blocks": [ + { + "bbox": [ + 306, + 249, + 545, + 272 + ], + "lines": [ + { + "bbox": [ + 306, + 249, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 545, + 272 + ], + "type": "text", + "content": "Table 2. The analysis of aligning order on ImageNet classification top-1 accuracy " + }, + { + "bbox": [ + 306, + 249, + 545, + 272 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 283, + 514, + 331 + ], + "lines": [ + { + "bbox": [ + 337, + 283, + 514, + 331 + ], + "spans": [ + { + "bbox": [ + 337, + 283, + 514, + 331 + ], + "type": "table", + "html": "
Projection HeadImageNet Top-1 Acc (%)
Linear84.0
2-layer MLP84.0
3-layer MLP83.8
", + "image_path": "a87dd2a6b89c1ec1c741c81138ab5c542731cf7efe968355b92cb2b960f4a05c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 335, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 306, + 335, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 545, + 358 + ], + "type": "text", + "content": "Table 3. Projection Head. A simple fully-connected layer works the best. We choose this as the default setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": "other experiments. It is also worth mentioning that simply aligning multiple layers has no clear advantage over aligning features from " + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\frac{3}{4}" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": " depth (84.2% vs. 84.0%); we, therefore, stick to the " + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\frac{3}{4}" + }, + { + "bbox": [ + 305, + 380, + 545, + 430 + ], + "type": "text", + "content": " depth setting, which is more efficient." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 433, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 433, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 433, + 546, + 529 + ], + "type": "text", + "content": "Aligning order. We next test the importance of alignment ordering on model performance. Specifically, by fixing the layer location in the student model (i.e., the middle layer in our experiment), we then align it to different layers of the teacher model. As shown in Table 2, we observe that when the aligned layers are in the same relative position (e.g., middle to middle), the student model can achieve the best performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": "Masking ratio. MAE reveals that the masking ratio in masked image modeling could be surprisingly high (75%). The hypothesis is that by learning to reason about the gestalt of the missing objects and scenes, which cannot be done by extending lines or textures because of the high masking ratio, the model is also learning useful representations. Interestingly, we find that when combining MAE and knowledge distillation, an even much higher masking ratio is possible, as shown in Figure 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": "Firstly, it is interesting to note that, compared to the typical " + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": " masking ratio setting, further raising the masking ratio to " + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": " comes at no performance drop, i.e., both attain " + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": " top-1 ImageNet accuracy. Next, even with an extremely large masking ratio like " + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": " (only FOUR visible patches), DMAE still beats the 100-epoch MAE baseline" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 197, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 197, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 197, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 702, + 197, + 713 + ], + "type": "text", + "content": "https://github.com/facebookresearch/mae" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24259" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 121, + 70, + 471, + 232 + ], + "blocks": [ + { + "bbox": [ + 121, + 70, + 471, + 232 + ], + "lines": [ + { + "bbox": [ + 121, + 70, + 471, + 232 + ], + "spans": [ + { + "bbox": [ + 121, + 70, + 471, + 232 + ], + "type": "image", + "image_path": "339f5a89158840705c7d8d3a45b2f9dc22625001b6cb03b561f002fbfacc943d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": "Figure 2. DMAE allows an extremely high masking ratio. From left to right, we increase the masking ratio from the basic " + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": " to the extreme " + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": ". We note that our DMAE competitively attains " + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "inline_equation", + "content": "83.6\\%" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": " top-1 ImageNet accuracy with " + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": " masking ratio (TEN visible patches), and still secures " + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "inline_equation", + "content": "82.4\\%" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": " top-1 ImageNet accuracy even by learning with FOUR visible patches (" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 46, + 236, + 547, + 270 + ], + "type": "text", + "content": " masking ratio)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 288, + 252, + 337 + ], + "blocks": [ + { + "bbox": [ + 80, + 288, + 252, + 337 + ], + "lines": [ + { + "bbox": [ + 80, + 288, + 252, + 337 + ], + "spans": [ + { + "bbox": [ + 80, + 288, + 252, + 337 + ], + "type": "table", + "html": "
Decoder DepthImageNet Top-1 Acc (%)
283.7
483.8
884.0
", + "image_path": "77e982003bc1cc19a8238aec6c8a734dfa933cc67a72043ae82d09969b6ec5e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 375, + 284, + 456 + ], + "blocks": [ + { + "bbox": [ + 47, + 340, + 287, + 363 + ], + "lines": [ + { + "bbox": [ + 47, + 340, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 287, + 363 + ], + "type": "text", + "content": "Table 4. Decoder Depth. A deeper decoder (slightly) improves the pre-trained representation quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 375, + 284, + 456 + ], + "lines": [ + { + "bbox": [ + 48, + 375, + 284, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 284, + 456 + ], + "type": "table", + "html": "
Loss DesignImageNet Top-1 Acc (%)
Loss ChoiceL1 with α = 184.0
L2 with α = 183.3
Loss RatioL1 with α = 0.583.9
L1 with α = 184.0
L1 with α = 284.0
L1 with α = 484.0
", + "image_path": "123de3ee749463ce539e6d2c9d8fa426b79223ac4c9f750f79f75c7fcabd9529.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 515 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 515 + ], + "type": "text", + "content": "Table 5. Loss Function. From the first block, we can observe that L1 distance yields significantly higher performance than L2 distance. From the second block, we can observe that the hyperparameter " + }, + { + "bbox": [ + 46, + 460, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 460, + 287, + 515 + ], + "type": "text", + "content": " has little influence on the representation quality of DMAE." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": "that uses a masking ratio of " + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": " (second row in Table 6), by a non-trivial-margin " + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "inline_equation", + "content": "(82.4\\% \\text{ vs. } 81.6\\%)" + }, + { + "bbox": [ + 46, + 539, + 287, + 611 + ], + "type": "text", + "content": ". These results suggest that, with the assistance of distilled knowledge from the teacher model, the student model can make better use of visible patches, even at a very limited amount, for representation learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "Projection head. The goal of the proposed feature alignment distillation is to encourage the student model to learn features that are predictive of features from a stronger teacher model. To that end, a small projection head is employed on features from the student model, to 1) project them onto a space of the same dimension as the hidden dimension of the teacher model, and 2) provide extra flexibility for feature alignment. We ablate the choice of this pro" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 290, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 327 + ], + "type": "text", + "content": "jection network, as shown in Table 3. We can observe that applying a simple fully connected layer already performs the best among other choices." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 329, + 546, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 425 + ], + "type": "text", + "content": "Decoder depth. In Table 4 we analyze the effect of decoder depth. Similar to MAE [13], the final performance gets (slightly) increased with a deeper decoder. We choose a decoder depth of 8 as the default setting as in [13]. Note that a decoder depth of 2 is also a competitive choice—compared to an 8-depth decoder, it significantly reduces the computation cost while only marginally sacrificing the accuracy by " + }, + { + "bbox": [ + 304, + 329, + 546, + 425 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 304, + 329, + 546, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "content": "Loss designs. Table 5 ablates the loss design. While SimMIM [31] shows that L1 distance and L2 distance lead to similar performance, ours suggests that L1 distance exhibits a clear advantage over L2 distance, i.e., " + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "inline_equation", + "content": "+0.7\\%" + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "content": " improvement. Furthermore, we note DMAE is quite robust to the specific value of the hyperparameter " + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "content": ", which controls the relative importance of the distillation loss over the reconstruction loss. Based on these results, we choose L1 in Eq. 2 for distance measure and set " + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "content": " in Eq. 3 for the rest experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 556, + 459, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 556, + 459, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 556, + 459, + 570 + ], + "type": "text", + "content": "4.3. Comparison with Baselines." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 575, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 545, + 598 + ], + "type": "text", + "content": "In Table 6, we compare the performance of our DMAE with various baselines:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "text", + "content": "MAE. The MAE baselines are presented in the first block of Table 6. If MAE is also asked to pre-train for only 100 epochs, DMAE can substantially outperform this baseline by " + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "inline_equation", + "content": "81.6\\%" + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "text", + "content": "). When comparing to a much stronger but more computationally expensive MAE baseline with 1600 pre-training epoch, we note DMAE still beats it by " + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 602, + 545, + 686 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Supervised model. The second block in Table 6 demonstrates the effectiveness of DMAE compared with models" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24260" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 129, + 70, + 462, + 218 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 462, + 218 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 462, + 218 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 462, + 218 + ], + "type": "table", + "html": "
MethodPre-training epochsSupervised training / fine-tuning epochsImageNet Top-1 Acc (%)
MAE-B10010081.6
MAE-B160010083.6
DeiT-B-10076.8
DeiT-B-30081.8
DeiT-B-Soft Distillation-10077.5
DeiT-B-Hard Distillation-10078.3
CRD [24]-10081.9
SRRL [32]-10082.2
Dear-KD [6]-10082.4
Supervised Feature Alignment-10082.8
DMAE-B10010084.0
", + "image_path": "89df4888d473e4be63d904611b3032a0a40463dc1d3111c830f1a640119e6b9d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "type": "text", + "content": "trained under the supervision of categorical labels, which requires a much longer training time, i.e., " + }, + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "type": "inline_equation", + "content": "+2.2\\%" + }, + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "type": "text", + "content": " compared to the DeiT 300 epochs supervised training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": "Other distillation strategies. We next compare DMAE with other distillation methods. We consider DeiT-style logit-based distillation [25], CRD [24], SRRL [32], DearKD [6], and feature alignment distillation. Note that for these baselines, one significant difference from DMAE is that the student model here directly distill knowledge from a supervisely fine-tuned teacher model. Moreover, to make these baselines more competitive, the teacher models will first be MAE pre-trained and then fine-tuned on ImageNet-1k." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 423, + 288, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 543 + ], + "type": "text", + "content": "The results are shown in the third block of Table 6. Firstly, we can observe that the DeiT-style logit-based distillation, either soft or hard, even hurts the student models' performance. This phenomenon potentially suggests that such a distillation strategy may not fit teacher models of ViT architectures. For other baselines, we note that feature alignment distillation performs the best; but this is still worse than DMAE (82.8% vs. 84.0%), indicating the importance and effectiveness of distilling knowledge from a pre-trained teacher model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 549, + 220, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 220, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 220, + 562 + ], + "type": "text", + "content": "4.4. Scaling to Different Model Sizes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": "We test DMAE with different model sizes, listed in Table 7. For a fair comparison, both methods only pre-train models for 100 epochs. DMAE shows consistent improvement compared to MAE across different model sizes. With only one middle layer feature alignment, DMAE brings an additional improvement of " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "+2.4\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": " with ViT-B, " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "+2.7\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": " with ViT-Small, and " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "+3.4\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": " with ViT-Tiny. In addition, we are interested in the following two cases:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Same teacher model, different student model. As shown in the first two lines in Table 7, we find that when using ViT-L as the teacher model, both ViT-B and ViT-S benefit from the distillation, demonstrating a clear advantage over" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 263, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 46, + 221, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 547, + 246 + ], + "type": "text", + "content": "Table 6. DMAE shows stronger performance than all three kinds of baselines: MAE, supervised model, and other existing advanced distillation strategies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 263, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 335 + ], + "type": "table", + "html": "
Student ModelTeacher ModelImageNet Top-1 Acc (%)
MAEDMAE
BaseLarge81.684.0 (+2.4)
SmallLarge77.480.1 (+2.7)
Base77.479.3 (+1.9)
TinyBase66.670.0 (+3.4)
", + "image_path": "edcbe723c60497a602e2a30e5d2fdf886b20c1b98dc4329fbc4f873e32d9c914.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 339, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 306, + 339, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 545, + 361 + ], + "type": "text", + "content": "Table 7. Across different model sizes, DMAE shows consistent improvements compared with MAE." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 383, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 383, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 545, + 432 + ], + "type": "text", + "content": "the MAE baseline. We argue that the ability to effectively generalize to cases where an even smaller student model is desirable, especially for those computation-constrained real-world applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "type": "text", + "content": "Same student model, different teacher model. As shown in Table 7, with a ViT-S as the student model, enlarging the teacher model from ViT-B to ViT-L further boosts the accuracy by " + }, + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "type": "inline_equation", + "content": "79.3\\%" + }, + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "type": "inline_equation", + "content": "80.1\\%" + }, + { + "bbox": [ + 305, + 435, + 545, + 506 + ], + "type": "text", + "content": "). This result suggests that DMAE can effectively distill knowledge from the teacher models at different scales." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 515, + 435, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 435, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 435, + 529 + ], + "type": "text", + "content": "4.5. Limited Training Data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "In certain real-world applications, data could be hard to acquire because of high data collection and labeling costs or due to privacy concerns. Leveraging models pre-trained on large-scale unlabeled datasets for fine-tuning when only a small dataset of downstream tasks is available becomes a promising solution. Here we seek to test the potential of our DMAE in this data-scarce scenario. We strictly follow [5] to sample " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " of the labeled ILSVRC-12 training datasets in a class-balanced way. We set MAE pre-trained ViT-L as the teacher model and a randomly initialized ViT-B as the student model. In addition, we compare DMAE with three kinds of baselines described in Section 4.3: MAE, supervised model, and other distillation strategies, and similarly, set the teacher model to be a ViT-L that first pre-trained with MAE and then fine" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24261" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 492, + 174 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 492, + 174 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 492, + 174 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 492, + 174 + ], + "type": "table", + "html": "
MethodPre-training epochsSupervised training / fine-tuning epochsIN-1% Top-1 Acc(%)IN-10% Top-1 Acc(%)
MAE-B10010033.965.0
MAE-B160010049.672.8
DeiT-B-100--
DeiT-B-Soft Distillation160010036.066.4
DeiT-B-Hard Distillation160010037.367.3
Supervised Feature Alignment160010034.267.6
DMAE-B10010050.373.4
", + "image_path": "f0fd9267419bad2f9262a939e2f2796a892e1a0f32c5ca6bf59929b8da88798a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 127, + 211, + 465, + 304 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "text", + "content": "Table 8. DMAE demonstrates much stronger performance than all other baselines when training data is limited. Note that the DeiT-B baseline is unable to converge because of data insufficiency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 127, + 211, + 465, + 304 + ], + "lines": [ + { + "bbox": [ + 127, + 211, + 465, + 304 + ], + "spans": [ + { + "bbox": [ + 127, + 211, + 465, + 304 + ], + "type": "table", + "html": "
ModelTraining Cost (GPU Hours)ImageNet Top-1 Acc(%)
Pre-trainingFine-tuningOverall
MAE-B-100-epoch78h112h190h81.6
MAE-B-1600-epoch1248h112h1360h83.6
DeiT-B-Soft Distillation-213h213h77.5
DeiT-B-Hard Distillation-213h213h78.3
Supervised Feature Alignment-208h208h82.8
DMAE-B83h112h195h84.0
", + "image_path": "b0cdbeb78b69908832427830b0243c2f7740137b9be5c2054fda07d4054bf5a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 362, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 422 + ], + "type": "text", + "content": "tuned on ImageNet-1k. Note that since DMAE has full access to the " + }, + { + "bbox": [ + 46, + 362, + 287, + 422 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 422 + ], + "type": "text", + "content": " ImageNet dataset (without labels) during pre-training, to ensure a fair and competitive comparison, we initialize all the baselines as the 1600-epoch MAE pretrained model on ImageNet." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "content": "Table 8 shows that DMAE largely surpasses all other baselines. For example, when only " + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "content": " ImageNet data is available for supervised training or fine-tuning, DMAE outperforms the MAE pre-trained baseline by " + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "inline_equation", + "content": "73.4\\%" + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "inline_equation", + "content": "65.0\\%" + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "content": "). DMAE also significantly outperforms other distillation strategies, with an improvement ranging from " + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "content": ". We note this accuracy gap is even larger when only " + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 427, + 288, + 536 + ], + "type": "text", + "content": " ImageNet is available, demonstrating the data efficiency of DMAE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 558, + 226, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 226, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 226, + 571 + ], + "type": "text", + "content": "4.6. Generalization to Other Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Lastly, we provide preliminary results of integrating DMAE into other self-supervised training frameworks, including DINO [4] and MoCo-V3 [7]. Unlike MAE, which belongs to masked image modeling, DINO and MoCo-V3 are contrastive learning-based methods. Still, as shown in Table 10, without further hyperparameter tuning, DMAE effectively shows non-trivial improvements on top of both DINO (+1.3%) and MoCo-v3 (+1.4%), demonstrating the potential of distilling pre-trained models (rather than fine-tuned models as in most existing knowledge distillation frameworks)." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 346, + 360, + 504, + 398 + ], + "blocks": [ + { + "bbox": [ + 46, + 308, + 546, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 546, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 546, + 343 + ], + "type": "text", + "content": "Table 9. Computational cost comparisons among DMAE and other baselines. The training cost is measured by A5000 GPU hours. We note the proposed DMAE maintains a similar (or even cheaper) training cost than others, while achieving much higher top-1 ImageNet accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 346, + 360, + 504, + 398 + ], + "lines": [ + { + "bbox": [ + 346, + 360, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 346, + 360, + 504, + 398 + ], + "type": "table", + "html": "
w/o DMAEw/ DMAE
DINO80.982.2 (+1.3)
MoCo-V381.182.5 (+1.4)
", + "image_path": "3e6ce96032e311ef624f179ff3b9b4c71b4f1d406fb8a17a6875fedb026d833b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 469, + 499, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 499, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 499, + 482 + ], + "type": "text", + "content": "4.7. Generalization to Downstream tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 604 + ], + "type": "text", + "content": "Following ViT-Det [19], we conduct downstream fin-tuning on the COCO dataset, where MAE/DMAE pretrained ViT-Base model is adopted as the plain backbone of Mask-RCNN. We train the model on the train2017 split and evaluate it on the val2017 split. We report results on bounding box object detection " + }, + { + "bbox": [ + 304, + 495, + 547, + 604 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^{\\mathrm{box}})" + }, + { + "bbox": [ + 304, + 495, + 547, + 604 + ], + "type": "text", + "content": " and instance segmentation " + }, + { + "bbox": [ + 304, + 495, + 547, + 604 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^{\\mathrm{mask}})" + }, + { + "bbox": [ + 304, + 495, + 547, + 604 + ], + "type": "text", + "content": ", shown in Table 11. Our observations indicate that DMAE also demonstrates strong potential in downstream tasks like detection and segmentation." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 339, + 627, + 512, + 666 + ], + "blocks": [ + { + "bbox": [ + 305, + 401, + 547, + 436 + ], + "lines": [ + { + "bbox": [ + 305, + 401, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 547, + 436 + ], + "type": "text", + "content": "Table 10. DMAE effectively improves other self-supervised pre-training frameworks (including DINO and MoCo-v3) on ImageNet classification top-1 accuracy " + }, + { + "bbox": [ + 305, + 401, + 547, + 436 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 401, + 547, + 436 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 627, + 512, + 666 + ], + "lines": [ + { + "bbox": [ + 339, + 627, + 512, + 666 + ], + "spans": [ + { + "bbox": [ + 339, + 627, + 512, + 666 + ], + "type": "table", + "html": "
method\\(AP^{box}\\)\\(AP^{mask}\\)
MAE51.245.5
DMAE53.446.9
", + "image_path": "8c4577dee5ae6b3dca99b54656da26d9bd7a3e263035e18cd2b5f9b2c85bc546.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 669, + 545, + 691 + ], + "lines": [ + { + "bbox": [ + 306, + 669, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 545, + 691 + ], + "type": "text", + "content": "Table 11. DMAE effectively improves the downstream tasks as well, including object detection and segmentation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24262" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 116, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 116, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 116, + 83 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 163 + ], + "type": "text", + "content": "In this section, we present a quantitative evaluation of the computational cost of DMAE compared to baseline methods. Additionally, we conduct a standard deviation analysis to assess the stability of our DMAE approach. Finally, we propose an advanced fine-tuning recipe for applying DMAE to smaller ViT models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 170, + 169, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 170, + 169, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 169, + 182 + ], + "type": "text", + "content": "5.1. Computational Costs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 188, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 287, + 308 + ], + "type": "text", + "content": "In Table 9, we provide a quantitative evaluation on the computational cost, which is tested on a single NVIDIA A5000 GPU. We can observe that, without a significantly increasing of training hours, DMAE substantially outperforms the MAE-100 baseline by " + }, + { + "bbox": [ + 46, + 188, + 287, + 308 + ], + "type": "inline_equation", + "content": "+2.4\\%" + }, + { + "bbox": [ + 46, + 188, + 287, + 308 + ], + "type": "text", + "content": " on ImageNet; this result even exceeds the MAE-1600 baseline, which causes " + }, + { + "bbox": [ + 46, + 188, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\sim 7x" + }, + { + "bbox": [ + 46, + 188, + 287, + 308 + ], + "type": "text", + "content": " GPU hours than MAE-100 baseline. We additionally provide the actual GPU hours of other baselines, and find that the proposed DMAE stands as the most efficient one, meanwhile achieving the best top-1 ImageNet accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 315, + 204, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 204, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 204, + 328 + ], + "type": "text", + "content": "5.2. Standard Deviation Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 333, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 441 + ], + "type": "text", + "content": "In the above experiments, we kept the same random seed. Following MAE, we perform the statistical analysis for DMAE by changing the random seeds. In Table 12, from top to down, we show three aligning settings; and from left to right, we show the results with the default seed, the average accuracy with three randomly sampled seeds, and their standard deviation, respectively. From these results, we could conclude that our DMAE can bring in statistically stable improvements." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 450, + 274, + 499 + ], + "blocks": [ + { + "bbox": [ + 58, + 450, + 274, + 499 + ], + "lines": [ + { + "bbox": [ + 58, + 450, + 274, + 499 + ], + "spans": [ + { + "bbox": [ + 58, + 450, + 274, + 499 + ], + "type": "table", + "html": "
PosAcc(%)AvgStandard Deviation
Bottom (3th)82.682.500.20
Mid (6th)83.683.670.08
Top (9th)84.084.030.10
", + "image_path": "e91c0a67935d187adea730256c72a4627b4950484f05378cdd5c915aa546dbeb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 167, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 167, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 167, + 611 + ], + "type": "text", + "content": "5.3. Smaller ViT Models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Since the original MAE paper does not offer specialized recipes for ViT-Small and Tiny, we by default use the recipe for ViT-Base to fine-tune these smaller ViTs. However, this recipe includes strong regularization and augmentation techniques that might lead to over-regularization for the smaller ViTs. To address this issue, we experiment with a modified recipe with weaker augmentation and regularization by removing MixUp, CutMix, and Stochastic Depth." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 369, + 70, + 486, + 118 + ], + "blocks": [ + { + "bbox": [ + 46, + 502, + 287, + 579 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 579 + ], + "type": "text", + "content": "Table 12. Standard deviation analysis for DMAE. From top to bottom, we show three aligning settings, the model performance with the default seed, the average performance with three randomly sampled seeds, and their standard deviations, respectively. 'Pos' denotes the student distillation position. Acc denotes ImageNet Top-1 accuracy. 'Avg' denotes the average value over three times." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 369, + 70, + 486, + 118 + ], + "lines": [ + { + "bbox": [ + 369, + 70, + 486, + 118 + ], + "spans": [ + { + "bbox": [ + 369, + 70, + 486, + 118 + ], + "type": "table", + "html": "
ViT-TinyTop-1 Acc (%)
MAE70.1
DeiT74.5
DMAE74.9
", + "image_path": "f5149e90040822103b0b1a1128199914e96c7ae78b4a5ff007ed6372b5d5de2f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 369, + 138, + 486, + 187 + ], + "blocks": [ + { + "bbox": [ + 405, + 120, + 447, + 129 + ], + "lines": [ + { + "bbox": [ + 405, + 120, + 447, + 129 + ], + "spans": [ + { + "bbox": [ + 405, + 120, + 447, + 129 + ], + "type": "text", + "content": "(a) ViT-Tiny" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 369, + 138, + 486, + 187 + ], + "lines": [ + { + "bbox": [ + 369, + 138, + 486, + 187 + ], + "spans": [ + { + "bbox": [ + 369, + 138, + 486, + 187 + ], + "type": "table", + "html": "
ViT-SmallTop-1 Acc (%)
MAE80.0
DeiT81.2
DMAE82.2
", + "image_path": "92eb2b5cc66e6f4dbe7d5c526c679713f560f5cc190ae9957de73ec2cb35bd29.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 403, + 189, + 449, + 198 + ], + "lines": [ + { + "bbox": [ + 403, + 189, + 449, + 198 + ], + "spans": [ + { + "bbox": [ + 403, + 189, + 449, + 198 + ], + "type": "text", + "content": "(b) ViT-Small" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 203, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 203, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 547, + 237 + ], + "type": "text", + "content": "Table 13. Weaker augmentation and regularization helps smaller ViT models, during finetuning for both MAE and DMAE on ImageNet classification top-1 accuracy " + }, + { + "bbox": [ + 305, + 203, + 547, + 237 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 258, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 295 + ], + "type": "text", + "content": "Results on ImageNet in Table 13 show that DMAE not only maintains its advantage over MAE, but also outperforms DeiT with this new recipe, highlighting its effectiveness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 307, + 378, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 378, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 378, + 319 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 327, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 568 + ], + "type": "text", + "content": "Self-supervised pre-training has demonstrated great success for those exponentially growing models in the natural language domain. Recently, the rise of MAE shows that a similar paradigm also works for the computer vision domain, and now the development of vision models may embark on a similar trajectory as in the language domain. Yet, it is often desirable to have a well-balanced model between performance and speed in real-world applications. This work is a small step towards unleashing the potential of knowledge distillation, a popular model compression technique, within the MAE framework. Our DMAE is a simple, efficient, and effective knowledge distillation method: feature alignment during MAE pre-training. Extensive experiments on multiple model scales demonstrate the effectiveness of our approach. Moreover, an intriguing finding is that it allows for a masking ratio even higher than the already large one used in MAE (i.e., " + }, + { + "bbox": [ + 304, + 327, + 547, + 568 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 327, + 547, + 568 + ], + "type": "text", + "content": "). We have also validated the effectiveness of our DMAE when in the small-data regime. We hope this work can benefit future research in knowledge distillation with pre-trained models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 579, + 404, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 579, + 404, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 579, + 404, + 594 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 600, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 600, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 600, + 545, + 647 + ], + "type": "text", + "content": "This work is supported by a gift from Open Philanthropy, TPU Research Cloud (TRC) program, and Google Cloud Research Credits program. We also want to acknowledge ONR with award N00014-21-1-2812." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 659, + 365, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 365, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 365, + 672 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 310, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 680, + 547, + 713 + ], + "type": "text", + "content": "[1] Yutong Bai, Xinlei Chen, Alexander Kirillov, Alan Yuille, and Alexander C Berg. Point-level region contrast for object detection pre-training. In Proceedings of the IEEE/CVF Con-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24263" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 16061-16070, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 149 + ], + "type": "text", + "content": "[2] Yutong Bai, Haoqi Fan, Ishan Misra, Ganesh Venkatesh, Yongyi Lu, Yuyin Zhou, Qihang Yu, Vikas Chandra, and Alan Yuille. Can temporal information help with contrastive self-supervised learning? arXiv preprint arXiv:2011.13046, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 287, + 174 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2021. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, pages 9650-9660, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "text", + "content": "[6] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. In CVPR, pages 12052-12062, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 309, + 287, + 342 + ], + "type": "text", + "content": "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 344, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 344, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 344, + 287, + 365 + ], + "type": "text", + "content": "[8] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In CVPR, pages 4794-4802, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 367, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 367, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 287, + 399 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 400, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 443 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "text", + "content": "[12] Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. In ICML, pages 1607-1616. PMLR, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "type": "text", + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "text", + "content": "[15] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In CVPR, pages 1921-1930, 2019. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[16] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In AAAI, volume 33, pages 3779-3787, 2019. 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[18] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. NeurIPS, 31, 2018. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[19] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 280-296. Springer, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "type": "text", + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "text", + "content": "[21] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, volume 34, pages 5191-5198, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "text", + "content": "[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[23] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "type": "text", + "content": "[24] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 378, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 545, + 431 + ], + "type": "text", + "content": "[25] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 1, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "text", + "content": "[26] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In CVPR, pages 32-42, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "text", + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "type": "text", + "content": "[28] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 11(12), 2010. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "type": "text", + "content": "[29] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "type": "text", + "content": "[30] Junfei Xiao, Longlong Jing, Lin Zhang, Ju He, Qi She, Zongwei Zhou, Alan Yuille, and Yingwei Li. Learning from temporal gradient for semi-supervised action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3252-3262, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[31] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24264" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 217 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[32] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "text", + "content": "[33] Qihang Yu, Yingda Xia, Yutong Bai, Yongyi Lu, Alan L Yuille, and Wei Shen. Glance-and-gaze vision transformer. Advances in Neural Information Processing Systems, 34:12992-13003, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 183 + ], + "type": "text", + "content": "[34] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, pages 4320-4328, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "type": "text", + "content": "[35] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. arXiv preprint arXiv:2203.08679, 2022. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24265" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_content_list.json b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..de90b656fa6ef4b3cf92eb1f7b90b3810c8f2ef8 --- /dev/null +++ b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_content_list.json @@ -0,0 +1,1787 @@ +[ + { + "type": "text", + "text": "Masked Autoencoding Does Not Help Natural Language Supervision at Scale", + "text_level": 1, + "bbox": [ + 96, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Floris Weers Vaishaal Shankar Angelos Katharopoulos Yinfei Yang Tom Gunter", + "bbox": [ + 96, + 180, + 872, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fweers,vaishaal_shankar,a_katharopoulos,yinfeiy,tom_gunter}@apple.com", + "bbox": [ + 178, + 200, + 785, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 261, + 313, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self supervision and natural language supervision have emerged as two exciting ways to train general purpose image encoders which excel at a variety of downstream tasks. Recent works such as M3AE [31] and SLIP [63] have suggested that these approaches can be effectively combined, but most notably their results use small (<20M examples) pre-training datasets and don't effectively reflect the large-scale regime (>100M samples) that is commonly used for these approaches. Here we investigate whether a similar approach can be effective when trained with a much larger amount of data. We find that a combination of two state of the art approaches: masked auto-encoders, MAE [37] and contrastive language image pre-training, CLIP [68] provides a benefit over CLIP when trained on a corpus of 11.3M image-text pairs, but little to no benefit (as evaluated on a suite of common vision tasks) over CLIP when trained on a large corpus of 1.4B images. Our work provides some much needed clarity into the effectiveness (or lack thereof) of self supervision for large-scale image-text training.", + "bbox": [ + 75, + 292, + 473, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 611, + 209, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large scale pretraining has become a powerful tool in the arsenal of computer vision researchers to produce state of the art results across a wider variety of tasks [39,88,95,98]. However, when pre-training on tens of millions to billions of images it is difficult to rely on standard supervised methods to train models, as datasets of this size often lack reliable labels. In the presence of these massive but largely under-curated datasets, two general classes of methods to train general purpose image encoders have emerged:", + "bbox": [ + 75, + 637, + 468, + 773 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Self Supervised techniques that learn visual representations from the image data alone [11,36]", + "2. Natural Language Supervised methods that utilize paired free-form text data to learn visual representations [43, 69]" + ], + "bbox": [ + 88, + 785, + 468, + 873 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the unique strengths and weaknesses of each ap-", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proach $^1$ , a recent flurry of work has introduced methods that combine both forms of supervision [31, 56, 64, 78] to varying degrees of success. While each of these methods establishes some regime where the additional supervision helps, none of these \"joint-supervision\" methods advance state of the art in any meaningful way. Additionally, to our knowledge none of these methods have shown comparative results at the scale many large scale vision models are currently trained at ( $>100\\mathrm{M}$ examples) [43, 66, 69, 73, 80, 82, 98]. Furthermore, methods that use both forms of supervision start with the presumption that the additional supervision is helpful and either often lack clean ablations or lack evaluations in a \"high accuracy\" regime—leading to further confusion regarding whether a combination of these methods can actually improve the state of the art. To clarify this issue, in this work, we investigate a simple question:", + "bbox": [ + 496, + 262, + 893, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Does a combination of self supervision and natural language supervision actually lead to higher quality visual representations?", + "bbox": [ + 531, + 513, + 859, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to answer this, we first introduce a straightforward baseline approach that combines standard self-supervision and language supervision techniques. We combine masked auto-encoders (MAE) and contrastive language image-pretraining (CLIP) to make MAE-CLIP. We then present a careful study of the performance of MAE, M3AE, CLIP and MAE-CLIP across a wide variety of tasks in two distinct regimes: a \"low-sample\" $^{2}$ 11.3 million example regime and a \"high-sample\" 1.4 billion example regime. We train self-supervised and language-supervised methods using the same pre-training datasets under the assumption that we have no knowledge about downstream tasks. Our experiments show:", + "bbox": [ + 496, + 566, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. In the low sample size regime, without changing the final pooling operation in the network, we observe a large performance improvement, namely $6\\%$ on ImageNet [18] and $4\\%$ on VTAB [105]. However, when", + "bbox": [ + 511, + 770, + 893, + 830 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Self supervised methods can learn representations without labels, but natural language supervision learns better representations. Natural language supervised methods rely on quality of captions", + "2We note that what low sample means has changed substantially over the last few years" + ], + "bbox": [ + 500, + 838, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23432", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "we modify the pooling operation, the improvement substantially decreases to around $1\\%$ on both ImageNet and VTAB.", + "bbox": [ + 107, + 90, + 467, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. In the high sample size regime, there is virtually no difference in performance between MAE-CLIP and CLIP across ImageNet, VTAB, and VQA tasks.", + "bbox": [ + 89, + 148, + 467, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We believe our work is the first careful study of this form and contextualizes recent progress in both self-supervision and natural language supervision.", + "bbox": [ + 76, + 207, + 467, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rest of the paper is organized as follows: In Section 2, we cover related work in the areas of self supervision and natural language supervision. In Section 3, we give an overview of the baseline methods we study, MAE, M3AE, CLIP and our new baseline MAE-CLIP. Then we present and analyse our small scale and large scale experimental findings in Sections 4 and 5. Finally, we discuss potential explanations for our findings and some future work in 6.", + "bbox": [ + 75, + 253, + 467, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 388, + 217, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work combines natural language supervision and self-supervision in a multi-task approach to visual encoding, and so research from these three areas is relevant.", + "bbox": [ + 76, + 416, + 467, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Natural language supervision for visual encoding covers a variety of approaches that assume access to datasets of images or videos associated with text. Some of the most successful use a contrastive pairwise alignment signal applied to very large batch and dataset sizes [68]. This large batch size means hard negative pairs can be produced by random sampling, avoiding the need for a memory bank [12, 13, 38] or momentum distillation [15]. FILIP [96] further improves zero-shot performance for some tasks by using a more fine-grained elementwise contrastive loss. Image captioning has also shown promise as a pre-training task [20, 90], with [20] in particular demonstrating strong data efficiency. UNITER [16] pursues a similar idea, applying a generative loss to both image and text modalities. [2, 16, 23, 41, 47, 51, 61, 81, 84, 97] apply the masked patch prediction problem from [21, 37, 50] to a joint image-text data space. These approaches either use a pre-trained convolutional neural network (CNN) to generate region of interest (RoI) proposal encodings, or predict the labelled class of the masked patch instead of the raw pixel values (or the quantized patch ID). UNIMO-2 [53] attempts to ground image and text patches for a single data example into a top-k quantized set of embeddings, before passing the grounded embeddings along with the raw non-grounded embeddings into a decoder, targeting a loss based on masked language modeling (MLM) and image-text-matching (ITM).", + "bbox": [ + 75, + 462, + 467, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised approaches for visual encoding may be loosely categorized into those that target consistency constraints between multiple views of the same scene and those", + "bbox": [ + 76, + 854, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that attempt visual reconstruction on corrupted or conditioning representations of images. Consistency constraints are often derived through data-augmentation applied to a single real image [6, 10, 14, 35, 102], but these may also take the form of e.g. a temporal constraint if video data is available [89], or be applied at an image-patch level by making a smoothness assumption and selecting neighboring patches as 'consistent' whilst treating distant patches (or those sourced from another image) as negative examples [101]. Forms of denoising autoencoder (DAE) [87] are popular as a means of self-supervision, and have been investigated at a variety of scales; iGPT [9] learns to generate images as a row-major flattened sequence of pixels, whilst [65], [72], and [26] compress images into a short sequence of discrete codes before then regenerating and scoring at a pixel level. BEiT [3] first encodes an image into a sequence of discrete codes before masking and then predicting a subset of the codes. MAE [37] drops the discretization, instead predicting raw pixel loss for a subset of the encoded patches in a manner strongly reminiscent of BART [50]. SplitMask [25] applies both a masked image patch prediction loss and a pairwise contrastive loss, using the masked example as well as the inverted masked example to form each positive pair. Meanwhile, M3AE [24] and VL-BEiT [4] both propose a masked patch prediction problem applied to image and text modalities jointly. Recent concurrent work shows that while masking does not help generalization, the masking can be used to speed up training by dropping the masked tokens during the forward and backward pass [55]. EVA [29] shows that doing masked image modeling to predict the output embeddings (rather than pixels) of the masked image can be an effective pretraining task when combined with natural language supervision. We note for EVA the masked image modeling occurs only on smaller well curated datasets such as ImageNet, Conceptual Captions and COCO.", + "bbox": [ + 496, + 90, + 890, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-task methods for pre-training visual encoders are a highly active area of research. CoCA [99] combines the cross-modality contrastive task from [68] with image captioning in the style of [20], and by pre-training at very large scale show that the resulting model is more performant than prior art across a very broad array of downstream visual understanding tasks. SLIP [64] combines a SimCLR style self supervision loss with the standard CLIP contrastive loss to train a jointy self-supervised and natural language supervised model, however while they show large performance gains they are all in a low accuracy regime (below $40\\%$ Top-1 accuracy on ImageNet). Florence [100] employs a diverse array of tasks and datasets at pre-training time (including object detection, CLIP-style contrastive modeling, Visual Question Answering (VQA), and supervised classification). ALBEF [52] builds on a CLIP-style architecture by incorporating a single-stream encoder that consumes both", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23433", + "bbox": [ + 478, + 924, + 517, + 935 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8d0073fe48fb99b79de939d95b11dfd5570288c17649fbc5ff4b69761c64648.jpg", + "image_caption": [ + "Figure 1. MAE-CLIP: CLIP [68] augmented with a generative decoder in the style of MAE [31]." + ], + "image_footnote": [], + "bbox": [ + 158, + 77, + 816, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "modalities. It uses a masked language modeling loss for the text modality and an image-text matching loss for within-batch hard-negatives, mined according to contrastive similarity. FLAVA [79] also uses masked-image modeling, and uses both single and joint modality encoders, with additional single-modality datasets (i.e. not just image-text pairs) that allow it to generalize to longer text inputs and visual data which is unlikely to be captioned. X-VLM [103] utilizes image data which includes object-labelled bounding boxes to extend the image-text matching and contrastive losses to cover within-image regions as well as whole-image captions. They also include a masked language modelling objective, achieving impressive performance on many zero-shot downstream tasks—although the requirement for object bounding box data may not scale well for large pre-training datasets. SupMAE [57] adds a supervised labeling task to the MAE architecture, demonstrating that this results in better data efficiency. MVP [91], meanwhile, finetunes a pre-trained CLIP backbone using MIM, and demonstrates that starting with a pre-trained model improves downstream visual recognition performance. MAE-CLIP is reminiscent of M3AE (with the addition of a contrastive task), ALBEF (but applying the masked prediction task to both modalities, dropping the momentum distillation), and most of all FLAVA (without the single-modality masked-patch-prediction tasks, and with the use of high ratio random masking following MAE and M3AE).", + "bbox": [ + 76, + 276, + 472, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 76, + 696, + 205, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Contrastive language-image pre-training", + "text_level": 1, + "bbox": [ + 76, + 724, + 426, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP [68] and ALIGN [44] both demonstrated that contrastive image-language supervision applied at scale was capable of producing an image encoder that excels at a range of downstream tasks—often in a zero-shot setting. In our implementation, we begin by following the design outlined in [68]. Specifically, we use per-modality encoders to produce a single $\\ell 2$ -normalized dense embedding vector for each image or text input, before applying a pairwise InfoNCE [65] loss to a large global batch of paired image-text embeddings, using all non-paired examples as hard nega", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tives. We define the image-to-text loss $L_{i2t}$ and the text-to-image loss $L_{t2i}$ as:", + "bbox": [ + 498, + 276, + 890, + 306 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i 2 t} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(x _ {j} ^ {T} y _ {j} / \\sigma\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(x _ {j} ^ {T} y _ {k} / \\sigma\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 316, + 890, + 359 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t 2 i} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(y _ {j} ^ {T} x _ {j} / \\sigma\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(y _ {j} ^ {T} x _ {k} / \\sigma\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 371, + 890, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ is the global batch size, $x_{j}$ is the normalized embedding of the image for the $j$ -th pair and $y_{k}$ is the normalized embedding of the text in the $k$ -th pair in the batch. $\\sigma$ is a learnable temperature parameter. The total contrastive loss $L_{c} = \\frac{1}{2} (L_{t2i} + L_{i2t})$ is the average of these two losses.", + "bbox": [ + 496, + 417, + 890, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use transformer encoders for both modalities, however, differently to [68] we explore several strategies for aggregating the image-patch or text-token outputs into a single embedding (see Section 4.4). We also eschew the autoregressive masking strategy for the text encoder, instead allowing full bi-directional self attention following [44].", + "bbox": [ + 496, + 494, + 890, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Pooling", + "text_level": 1, + "bbox": [ + 500, + 603, + 607, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In CLIP [68] the authors use a separate 'CLS' token projected through the network as the overall image representation. On the other hand, ALIGN [44] use global-average pooling over the encoded visual features. Whilst both approaches produced good downstream evaluation results, recent work [8, 71, 104], suggested that the choice of pooling strategy can strongly influence the quality of visual semantic embeddings. In particular, [71] demonstrated that this effect is present for CLIP-like visual encoders. Noting this, we opt to investigate three pooling strategies: 1. the default multihead-attention pooling (MAP), 2. global average pooling (GAP)—both described in [104]—as well as 3. non-maximal suppression pooling (MAX) as in [8, 71].", + "bbox": [ + 496, + 626, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Masked Autoencoders", + "text_level": 1, + "bbox": [ + 500, + 832, + 705, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In MAE [37] the authors demonstrate a simple technique for self-supervised image-encoder pre-training that—to our knowledge—is still considered state-of-the-art. They use a", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23434", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ViT [22] encoder-decoder architecture and apply it to heavily masked input images. The input to the encoder consists of the visible, unmasked image patches, first embedded via a linear projection before additive positional embeddings are applied, and then the result is fed through the encoder's transformer layers. They demonstrate that a very high patch masking ratio is critical to achieving good performance, and usually retain only $25\\%$ of the image. The decoder consumes the output of the encoder, as well as a learned 'masked-patch' embedding, which is included to represent each masked token. Positional embeddings are also reapplied, to ensure that spatial information is passed through the decoder for the masked patches. The output of the decoder at the masked positions is then measured against ground truth using a simple mean-squared-error loss. As in [37] we experiment with predicting both normalized and un-normalized patch values, finding that predicting the normalized patch value slightly improves the performance of our MAE implementation.", + "bbox": [ + 75, + 90, + 472, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masking across modalities: BART [50] applies a similar strategy to text data, but uses a cross-entropy loss over the masked-token output distribution and a far lower masking ratio (typically around $15\\%$ ). M3AE [31] extends MAE and BART to incorporate inputs from both text and image modalities, using per-modality input and output projections. Otherwise the encoder-decoder architecture resembles MAE, with the addition of learned modality-indicating embeddings to each transformer input.", + "bbox": [ + 75, + 378, + 470, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The M3AE loss relies only on the contents of the image-text pairs, and comprises mean-squared error for the masked image patches $(L_{gen.i})$ and cross-entropy over the vocabulary for the masked text tokens $(L_{gen.t})$ :", + "bbox": [ + 76, + 517, + 470, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {g e n . i}} = \\frac {1}{N} \\sum_ {j} ^ {N} \\left(p _ {j} - P _ {j}\\right) ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 606, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {g e n . t}} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(t _ {n , T _ {n}}\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(t _ {n , c}\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 652, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_j$ and $P_j$ refer to the predicted and ground truth pixel value respectively and $t_n$ and $T_n$ refer to the predicted token distribution and the ground truth token, with $C$ as the number of unique tokens. For the image reconstruction loss, we normalize the ground-truth per-patch, following [37].", + "bbox": [ + 76, + 710, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. MAE-CLIP", + "text_level": 1, + "bbox": [ + 76, + 800, + 205, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the name suggests, MAE-CLIP attempts to incorporate aspects of MAE/M3AE into CLIP through the addition of a single dual-modality transformer decoder (as in M3AE) which consumes the output of the CLIP encoders. We show this in Figure 1.", + "bbox": [ + 76, + 824, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Components", + "text_level": 1, + "bbox": [ + 500, + 90, + 643, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The model architecture is similar to ALBEF [52] and FLAVA [78], consisting of three components.", + "bbox": [ + 498, + 117, + 890, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An image encoder: Following ViT [22], we divide the input image into equally-sized, non-overlapping patches. After applying a linear projection and adding a 2-D position encoding, we feed the per-patch representations through a number of transformer layers [86].", + "bbox": [ + 498, + 148, + 893, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A text encoder: Following ALIGN [43], our text encoder is based on BERT [21], in which text is first tokenized and embedded, and a 1-D trainable position encoding is added. Differently to BERT, we use pre-layernorm [94] and initialize the parameters according to the simplified scheme outlined in [70].", + "bbox": [ + 498, + 227, + 893, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A cross-modality decoder: The decoder receives per-element encoded image and text representations from the encoders for both masked and un-masked elements. For the image modality, masked patches are added by replacing their values with a shared trainable mask token. Positional encodings are once again added to all elements, following [37]. We also add a per-modality trainable encoding to allow the decoder to easily distinguish between the two modalities. The decoder uses the same transformer implementation as the encoders. The output of the final decoder layer is then projected into a per-patch-modality output space and the overall loss is computed.", + "bbox": [ + 498, + 319, + 892, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Losses", + "text_level": 1, + "bbox": [ + 500, + 523, + 599, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MAE-CLIP employs the losses of both CLIP (Eq. 1, 2) and M3AE (Eq. 3, 4). Our final loss is a weighted sum of the losses from each task, as follows:", + "bbox": [ + 498, + 551, + 890, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = \\frac {1}{2} \\left(L _ {t 2 i} + L _ {i 2 t}\\right) + w _ {i} \\cdot L _ {\\text {g e n . i}} + w _ {t} \\cdot L _ {\\text {g e n . t}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 609, + 890, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}$ and $w_{t}$ are scalars used to control the relative weight of the generative losses. We always provide paired image-text inputs to the model, leaving it to future work to explore the benefits of also incorporating single modality inputs and reconstruction losses, at scale.", + "bbox": [ + 498, + 652, + 890, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To avoid severely impacting the CLIP loss via computing it with masked inputs, we run the encoders twice: once with full unmasked input in order to compute the CLIP loss, before then making a second pass using only the unmasked input to compute the M3AE loss. Finally, we compute the weighted sum of the two losses, and use this to calculate the overall update.", + "bbox": [ + 498, + 729, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3 Alternative, more memory efficient, strategies exist for computing the overall updates (e.g. calculate per-loss gradients, or perform round-robin updates), but prior work [1] has demonstrated that some of these techniques may negatively impact the overall training result.", + "bbox": [ + 498, + 849, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23435", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present our experimental results to study how the addition of self-supervision affects the representation quality of natural language supervised models. We study standard masked auto-encoders (MAE), multimoddal masked auto-encoders (M3AE), contrastive language image pretraining (CLIP) and our newly introduced baseline method (MAE-CLIP), which combines mask based self-supervision with contrastive language-image learning.", + "bbox": [ + 75, + 114, + 468, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 261, + 266, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture: Experiments are performed using an image encoder based on the architecture described as ViT-B-16 in [93], in combination with a 12-layer, 512-wide text encoder with 8 attention heads. For the decoder, we use another 12-layer, 512-wide transformer with 8 attention heads, identical to the text encoder. We also use a byte-pair encoding tokenizer, fit to the OpenWebText corpus [32]. When training MAE-CLIP, we replace the random masking by a similarity masking strategy, which makes use of the element-wise CLIP similarity scores to select the masked element for the reconstruction task.", + "bbox": [ + 75, + 284, + 468, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretraining Datasets: Our analysis is divided into two sections. First, in sections 4.2 to 4.4, we present our study on the \"low-sample\" regime using the combination of CC12M [7] and CC3M [77] excluding all images whose text contains a [PERSON] tag5. The final dataset contains 11.3 million comparatively high quality image-text pairs. We refer to this dataset as CC or \"small\" scale.", + "bbox": [ + 75, + 450, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretraining configuration: We use a minibatch size of 16,384 samples and train using the AdamW optimizer [60] with decoupled weight decay. We use a base learning rate of $5 \\times 10^{-4}$ and train using cosine learning rate decay [59] with an initial linear warmup of 200 steps. We reduce the $\\beta_{2}$ decay factor to 0.98, and increase $\\epsilon$ to $10^{-6}$ , finding that these changes improved rate of convergence for all models trained. To increase training convergence speed, we switch from a local (per-GPU) to a global contrastive loss after 500 steps. During the global contrastive loss phase, we set the generative text and image loss weights to $w_{t} = 0.05$ and $w_{i} = 0.1$ respectively. For all models which rely on masking, we use a $75\\%$ masking ratio, consistent with [31,37], as we did not find an alternative that improved downstream results. In total, we train our models on CC for 32 epochs, which corresponds to $\\sim 22,000$ steps. For our M3AE training runs, we use per-modality encoders so as to produce a baseline that is flop and parameter matched to our MAE-CLIP runs.", + "bbox": [ + 75, + 556, + 470, + 843 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/661f772e83024212c326dd63c6f2dc2a9a0f70d4b96ed4e45c8ac31c93dbaf4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsZero-shotLinear Probing
MAE-33.9
M3AE-52.5
CLIP29.752.6
MAE-CLIP33.858.9
", + "bbox": [ + 550, + 88, + 841, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. ImageNet classification with zero-shot transfer or linear probing after pretraining on the $CC$ dataset (11.3M images). MAE-CLIP significantly improves the classification performance of CLIP in the small scale regime.", + "bbox": [ + 496, + 191, + 890, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Image Classification", + "text_level": 1, + "bbox": [ + 500, + 255, + 691, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the quality of the representations learned from pre-training on $CC$ by measuring the image classification performance, either by zero-shot transfer, or by training a linear classifier using the predicted visual features. For zero-shot classification, we use the average of 80 prompts and follow the routine prescribed in [69]. To train the linear classifier, we pre-compute features using the visual encoder6 and run AdamW for 20 to 80 epochs with a learning rate of 0.01.", + "bbox": [ + 496, + 279, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows the zero-shot and linear probing classification results on the ImageNet [19] dataset while Table 2 shows our linear probing results on most (17) of the VTAB [106] benchmark datasets7. Initially, we observe that the combination of self-supervision and natural language supervision provides a consistent and substantial improvement over either form of supervision by itself. This result concurs with previous works such as [31] and [64], which show that self-supervision aids natural language supervision.", + "bbox": [ + 496, + 415, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.VQA", + "text_level": 1, + "bbox": [ + 500, + 560, + 576, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we compare the performance of MAE-CLIP to the rest of the baselines on the visual question answering task using three datasets. CLEVR [45], VQAv2 [34] and GQA [42]. VQA assesses the model's multimodal reasoning capabilities as well as its visual grounding. In order to finetune our models for VQA, we freeze the image and text encoders and either randomly reinitialize the decoder in models such as MAE-CLIP or M3AE or add a new identical decoder for CLIP. During finetuning, we use a layer-wise learning rate decay, following [3, 17, 37]. For CLEVR and VQAv2 we finetune the decoder for 50 epochs, while for GQA we finetune for 5 epochs on the \"train-all\" split and 2 epochs on the \"train-balanced\" split following the protocol of [54]. In all cases, we treat the problem as a classification problem of selecting one answer out of the set of possible answers for each dataset.", + "bbox": [ + 496, + 584, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 3 we observe that self-supervision combined", + "bbox": [ + 517, + 825, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4We compare the similarity masking strategy with the original random masking strategy on various tasks, the results show they perform similarly.", + "bbox": [ + 75, + 849, + 467, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "5 Preliminary experiments on CLIP suggest removing such data leads to better performance.", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "6We use no data augmentation here", + "bbox": [ + 517, + 849, + 707, + 863 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "7We do not include results for either the diabetic retinopathy or the Sun397 tasks, due to licensing issues. More details in the supplementary material", + "bbox": [ + 500, + 863, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23436", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c1d7ee7eb17a7ec93fe7519a1da9a1b48a23b0f187043ceb61de04ee95cf43b0.jpg", + "table_caption": [ + "Table 2. Linear probing accuracy (%) on classification tasks. All models are trained on the CC Dataset (11.3M images). (● VTAB/natural, ● VTAB/specialized and ● VTAB/structured.)" + ], + "table_footnote": [], + "table_body": "
Caltech101CIFAR-100DTDFlowers102PetsSVHNEuroSATCamelyonResisc45Clevr/CloestClevr/CountDMLabdSprites/OridSprites/LocKITTI/DistsNORB/AzimsNORB/ElevAverage
MAE75.356.258.570.237.471.996.282.884.065.160.642.837.979.936.737.464.562.2
M3AE85.666.669.486.558.469.297.482.990.769.471.248.249.883.342.634.373.469.3
CLIP84.262.857.781.669.752.595.582.686.753.752.944.845.961.345.731.042.661.8
MAE-CLIP89.866.264.886.574.556.795.481.686.953.659.645.944.668.648.036.151.065.3
", + "bbox": [ + 122, + 103, + 849, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e0d22ff91b7f6efdf204110fecec6e56a63c89dd5770da3f5c5388aae68c3fbd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCLEVRVQAv2GQA
MAE93.545.746.1
M3AE97.555.850.9
CLIP87.555.650.0
MAE-CLIP96.058.552.2
", + "bbox": [ + 132, + 252, + 415, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with natural language supervision (MAE-CLIP) performs consistently better than either MAE or CLIP by themselves. Combining information from both modalities is evidently critical for effectively answering visual questions as the performance of MAE on VQAv2 and GQA is several percentage points lower than the rest of the methods. Interestingly, methods that employ a form of self-supervision perform significantly better on CLEVR. We argue that this discrepancy is due to CLEVR requiring object localization and spatial reasoning which benefits from this form of self-supervision.", + "bbox": [ + 75, + 428, + 468, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Pooling Analysis", + "text_level": 1, + "bbox": [ + 76, + 601, + 240, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To better understand whether the gap between MAE-CLIP and CLIP we observed in the previous section is fundamental, we investigate the effect of the pooling operator on different downstream tasks. Table 4 compares the performance of CLIP and MAE-CLIP with the standard multi-head attention pooling (MAP), global average pooling (GAP) and max pooling (MAX) on all image classification tasks and visual question answering. We note that GAP and MAX pooling perform substantially better than the standard MAP pooling across ImageNet, VTAB and VQA tasks. A hypothesis posed by [71] is that the alternative pooling operators lead to better perceptual grouping in the visual representation. We study this with a qualitative analysis in Section 6.1.", + "bbox": [ + 75, + 625, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments at Scale", + "text_level": 1, + "bbox": [ + 76, + 844, + 277, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At small-scale, we showed that self-supervision combined with natural language supervision marginally im", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/41c8ec98937388f3fd0fb5f0fe93ca15a6ae3bfd16c3dbb0aed025d4a1d9c538.jpg", + "table_caption": [ + "Table 3. VQA finetuning results for models pre-trained on CC Dataset (11.3M images). We train a new decoder for all methods while keeping the encoders frozen. MAE-CLIP performs significantly better than either only self-supervised or language supervised methods by themselves." + ], + "table_footnote": [], + "table_body": "
ModelPoolingImageNetVTABVQA
ZSLPLPFT
CLIPMAP29.752.661.864.4
MAE-CLIPMAP33.858.965.368.9
CLIPGAP29.359.870.665.1
MAE-CLIPGAP33.562.571.769.2
CLIPMAX33.162.368.763.5
MAE-CLIPMAX35.263.269.168.5
", + "bbox": [ + 514, + 252, + 877, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Results of CLIP and MAE-CLIP with different pooling options trained on CC dataset (11.3M images). We compare the models on ImageNet Zero-Shot (ZS), ImageNet Linear-Probing (LP), VTAB 17-task Linear-Probing average, and VQA 3 tasks Fine-Tuning (FT) average.", + "bbox": [ + 498, + 402, + 890, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proves the quality of the learned visual representations, and that the choice of image-encoder pooling strategy can heavily influence results. We now investigate whether these conclusions still hold when training on a much larger, 1.4B example dataset.", + "bbox": [ + 498, + 483, + 890, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 571, + 689, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the configuration described in Section 4.1, with a few key differences, noted below.", + "bbox": [ + 498, + 595, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining Dataset: We combine a 2.2B example web-crawled image-text pair dataset, termed the English-WebImage-Text dataset (EWIT-2.2B), with LAION-400M [74], CC3M [77], CC12M [7] and an internal high-quality image-text pair dataset containing approximately 134M image text pairs which we term the High Quality Image Text Pairs Dataset (HQITP-134M). We globally deduplicate this—keyed by image bytes—reducing the number of image-text pairs significantly, to yield a final 1.4B examples. Details provided in the supplementary material. In the rest of the paper we refer to this as the web-crawled or large-scale dataset.", + "bbox": [ + 496, + 627, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining configuration: We increase the learning rate warm-up to 1,000 steps, and compute a local contrastive loss for the first 10,000 steps, as during early training this improves the models rate of convergence. We train for a total of 480,000 steps, which corresponds to 6 full passes through the dataset. We do not increase the batch", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23437", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f913c492a463eb4ffd78eb5c63f193421a441055e531c55213cdf3bb403c19b3.jpg", + "table_caption": [ + "Table 5. Linear probing accuracy (%) on classification tasks. Models are all trained on our web-crawled dataset (1.4B images). (VTAB/natural, VTAB/specialized and VTAB/structured.) In the large scale pretraining regime, the difference between MAE-CLIP and CLIP is reduced to $< 1\\%$ . * At evaluation time, our M3AE model was $50\\%$ trained, so that performance may improve further." + ], + "table_footnote": [], + "table_body": "
Caltech101CIFAR-100DTDFlowers102PetsSVHNEuroSATCarnelyonResisc45Clevi/CloestClevi/CountDMLabdSprites/OridSprites/LocKITTI/DistsNORB/AzimsNORB/ElevAverage
M3AE*93.074.878.295.481.269.697.384.892.765.774.751.052.780.747.536.268.873.2
CLIP94.978.480.097.386.959.094.182.392.745.662.146.046.153.350.920.335.866.2
CLIPMAX96.181.080.997.389.965.796.083.294.152.867.849.959.567.641.223.445.870.1
MAE-CLIPMAX95.879.281.596.888.262.195.881.893.052.066.949.653.772.553.032.345.470.6
CLIPGAP95.880.581.697.688.766.097.084.493.356.771.453.358.070.150.638.355.172.9
MAE-CLIPGAP95.479.382.297.488.672.896.684.593.557.573.652.757.571.251.645.655.273.8
", + "bbox": [ + 107, + 132, + 859, + 303 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/789ccc830566b8b28285a755baa8e461510f24c89b255279abfbd270a313ffd9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsZero-shotLinear Probing
M3AE*-69.3
CLIPGAP61.875.9
CLIPMAX63.777.5
MAE-CLIPGAP57.475.7
MAE-CLIPMAX60.976.6
", + "bbox": [ + 117, + 315, + 431, + 416 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dbf42315a86d4c3464c93980804567fc537880018ab6aac4cfc45f4358c09232.jpg", + "table_caption": [ + "Table 6. ImageNet classification after pretraining on web-crawled dataset (1.4B images). In the large scale regime, self-supervision does not complement natural language supervision." + ], + "table_footnote": [], + "table_body": "
ModelCLEVRVQAv2GQA
M3AE96.959.953.3
CLIPGAP87.861.855.0
CLIPMAX89.560.653.6
MAE-CLIPGAP92.861.955.3
MAE-CLIPMAX93.961.553.7
", + "bbox": [ + 542, + 315, + 849, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. VQA finetuning results on web-crawled dataset (1.4B images). Similar to Table 15a, for large pretraining datasets, self-supervision is not improving the downstream VQA performance.", + "bbox": [ + 498, + 425, + 890, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "size, and use the same learning rate schedule as for $CC$ .", + "bbox": [ + 76, + 487, + 444, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model variants: We train M3AE as well as both GAP and MAX variants of CLIP and MAE-CLIP, having noted in the small-scale regime that these are the highest performing architectures and pre-training strategies. We do not train MAE at scale due to resource limitations, and because its performance was significantly lower than the models that utilize natural language supervision—a finding that is consistent with prior literature [31].", + "bbox": [ + 75, + 502, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Image Classification", + "text_level": 1, + "bbox": [ + 76, + 635, + 267, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 15a provides encoder performance for both zero-shot and linear-probing on ImageNet [19]. We note that at scale, MAE-CLIP typically shows worse performance than CLIP across both pooling strategies and tasks. We hypothesize that this may be due to the fact that masked-patch-prediction results in model capacity being lost to the task of predicting semantically irrelevant patches, meaning that we lose some of the benefits provided by the larger scale dataset (despite tuning the relative loss weights). If we are correct in this hypothesis then it is possible that a larger model capacity would help to alleviate this issue. We leave this to a future investigation. Meanwhile, in Table 14 we once again explore the encoder performance through linear-probing for VTAB [106]. We note that GAP provides a consistent improvement over MAX across both CLIP and MAE-CLIP models, and that M3AE is also a strong performer—as was", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the case in the small-scale regime. M3AE provides particular benefit to the \"structured\" VTAB tasks, however underperforms (sometimes by a large margin) on the \"natural\" tasks. MAE-CLIP ends up in between the two—doing better than CLIP but worse than M3AE on the structured tasks, and worse than CLIP but better than M3AE on the natural tasks.", + "bbox": [ + 496, + 486, + 890, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.VQA", + "text_level": 1, + "bbox": [ + 500, + 611, + 576, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow the same finetuning strategy as before, with results in Table 15b. Notably, we do not see a consistent difference across tasks for CLIP vs MAE-CLIP, with the possible exception of CLEVR [45], where it is clear that M3AE is by far the best performing approach—and its benefits improve the MAE-CLIP results on that task. GAP is perhaps a marginally stronger pooling strategy as measured against MAX, however the difference is smaller than it was in the small-scale regime.", + "bbox": [ + 496, + 637, + 890, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 500, + 796, + 723, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While our results present a pessimistic view on combining self-supervision with natural language supervision, we don't have a clear reason as to why self-supervision fails at scale. Here we present two hypotheses that may be explored in future work.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23438", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a1803115410c8dae5d6e5cc382ef40d248a06721407883a4b035ecb0bd303ff.jpg", + "image_caption": [ + "(a) CLIP" + ], + "image_footnote": [], + "bbox": [ + 133, + 102, + 250, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e77dd4128ede60c24b9a7b39c084248839989742632b53f6c4c84779bd1e161.jpg", + "image_caption": [ + "(b) MAE-CLIP" + ], + "image_footnote": [], + "bbox": [ + 251, + 102, + 367, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc7d15102741e24869e262039d947f9be47c83657f0c2e2e6375acf8fc59b7ba.jpg", + "image_caption": [ + "(c) $\\mathrm{CLIP}_{\\mathrm{GAP}}$", + "Figure 2. Zero-shot segmentation masks (top row) and GradCAM [75] visualizations (bottom row) for the prompt \"a photo of a dog\" for CLIP and MAE-CLIP. Self-supervision qualitatively improves the visual grounding of the model, however, the choice of pooling operator has the largest effect, with max pooling producing significantly better results [71]." + ], + "image_footnote": [], + "bbox": [ + 367, + 102, + 483, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/240f6ae5dfd7ebbdc460c056fe14a4539dd23b1547e82e3e95d74d34f64ab11b.jpg", + "image_caption": [ + "(d) MAE-CLIPGAP" + ], + "image_footnote": [], + "bbox": [ + 486, + 102, + 602, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5ec95de2935b4a5d1c5dfacb2cc60c1a3231cc7a3b9121a07a911cf4a9e887b.jpg", + "image_caption": [ + "(e) $\\mathrm{CLIP}_{\\mathrm{MAX}}$" + ], + "image_footnote": [], + "bbox": [ + 604, + 102, + 718, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa409b56104c878a95724ad919adaa7268c1203d1b6b0ca1179052a8fe931aa2.jpg", + "image_caption": [ + "(f) MAE-CLIPMAX" + ], + "image_footnote": [], + "bbox": [ + 720, + 102, + 834, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Visual Grounding", + "text_level": 1, + "bbox": [ + 76, + 349, + 250, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual grounding measures how well a particular representation or network can localize objects within an image. It is thought that representations with better visual grounding will excel at more general purpose tasks. In section 4, we found that pooling operator had a larger effect on performance than the addition of MAE. We qualitatively study the effect of both pooling and MAE on the visual grounding of the CLIP image and text encoders. We extract zero-shot segmentation masks and GradCAM [75] visualizations for the prompt \"a photo of a dog\", shown in Figure 2. Details are provided in the supplementary material. We observe that, with respect to visual grounding, the pooling operator plays a much more significant role than self-supervision [71]. In particular, $\\mathrm{CLIP}_{\\mathrm{MAX}}$ correctly segments the dog in the picture while CLIP focuses only in the background. However, both for CLIP and $\\mathrm{CLIP}_{\\mathrm{MAX}}$ we observe only an incremental improvement of the localization of the dog in the image when self-supervision is added. Unfortunately, even though the GAP variants of MAE-CLIP and CLIP achieve the best performance on VTAB, they do not also exhibit the most visual grounding. We hope that a more thorough future analysis can elucidate the relationship between self supervision and visual grounding.", + "bbox": [ + 75, + 371, + 472, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Dataset Diversity", + "text_level": 1, + "bbox": [ + 76, + 726, + 243, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternatively, it may be that self-supervision and natural language supervision excel for entirely different parts of the dataset diversity-size spectrum. Historically, the strongest self-supervised visual-encoder baselines (when compared to supervised methods) include methods such as: SimCLR [11], SimCLRV2 [12] and masked auto encoders [36], which all excel on ImageNet—a somewhat diverse but a relatively \"clean\" and object-centric dataset with 1000 disjoint classes. On less diverse datasets such as Cifar-10 [49], specialized self-supervised methods such as ReMixMatch [5]", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "have shown excellent performance in the extremely low data regime. Outside of vision, MAEs have shown near state-of-the-art performance on datasets such as Audioset [30,40], a clean 2M example, 632 class audio/video dataset. Recent works which perform controlled experiments scaling self supervised methods to massive datasets [62] show that self-supervised methods achieve numbers roughly $10\\%$ worse (as measured by ImageNet linear probe) as compared to natural language supervised methods at similar data scale.", + "bbox": [ + 496, + 349, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the other hand, natural language supervised models seem to only show competitive performance on massive and diverse datasets, CLIP [69] was trained on 400M examples, ALIGN [43] was trained on 1.8B examples and BASIC [66] was trained on 6.6B examples. Recently, [28] performs controlled natural language supervision at a smaller scale, showing that natural language supervised models are $5 - 10\\%$ worse (on ImageNet) than their supervised counterparts.", + "bbox": [ + 496, + 500, + 890, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While all of these past results only pose circumstantial and not rigorous experimental evidence, we believe it poses a fruitful line of future work on studying the scaling trends of self supervised methods and contrasting them with other forms of supervision.", + "bbox": [ + 496, + 621, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 709, + 679, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We are grateful to many colleagues for helpful advice and guidance throughout this project. We especially thank Jonathan Shlens, Brandon McKinzie, and Jason Ramuparam for feedback on paper drafts, Ruoming Pang for useful comments and suggestions early in the project, Tom Nickson for his training infrastructure support and Albin Madappally Jose for his assistance in building the training dataset. We also extend thanks to Dr. Nicola Strisciuglio for his clarifying input on earlier incarnations of this work, and Dr. Samuel Albanie for several engaging discussions and helpful pointers to related literature.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23439", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a Visual Language Model for Few-Shot Learning. Technical Report arXiv:2204.14198, arXiv, Apr. 2022. arXiv:2204.14198 [cs] type: article. 4", + "[2] Tarik Arici, Mehmet Saygin Seyfioglu, Tal Neiman, Yi Xu, Son Train, Trishul Chilimbi, Belinda Zeng, and Ismail Tutar. MLIM: Vision-and-Language Model Pre-training with Masked Language and Image Modeling. arXiv:2109.12178 [cs], Sept. 2021. arXiv: 2109.12178. 2", + "[3] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT Pre-Training of Image Transformers. arXiv:2106.08254 [cs], June 2021. arXiv: 2106.08254. 2, 5", + "[4] Hangbo Bao, Wenhui Wang, Li Dong, and Furu Wei. VL-BEiT: Generative Vision-Language Pretraining. Technical Report arXiv:2206.01127, arXiv, June 2022. arXiv:2206.01127 [cs] type: article. 2", + "[5] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. CoRR, abs/1911.09785, 2019. 8", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. DINO: Emerging Properties in Self-Supervised Vision Transformers. arXiv:2104.14294 [cs], May 2021. arXiv: 2104.14294. 2, 16", + "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts. Technical Report arXiv:2102.08981, arXiv, Mar. 2021. arXiv:2102.08981 [cs] type: article. 5, 6, 14", + "[8] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. iGPT: Generative Pretraining From Pixels. In Proceedings of the 37th International Conference on Machine Learning, pages 1691-1703. PMLR, Nov. 2020. ISSN: 2640-3498. 2", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. SimCLR: A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597-1607. PMLR, Nov. 2020. ISSN: 2640-3498. 2", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. CoRR, abs/2002.05709, 2020. 1, 8" + ], + "bbox": [ + 86, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. SimCLRv2: Big Self-Supervised Models are Strong Semi-Supervised Learners. arXiv:2006.10029 [cs, stat], Oct. 2020. arXiv: 2006.10029. 2, 8", + "[13] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. MoCov2: Improved Baselines with Momentum Contrastive Learning. arXiv:2003.04297 [cs], Mar. 2020. arXiv: 2003.04297. 2", + "[14] Xinlei Chen and Kaiming He. SimSiam: Exploring Simple Siamese Representation Learning. arXiv:2011.10566 [cs], Nov. 2020. arXiv: 2011.10566. 2", + "[15] Xinlei Chen, Saining Xie, and Kaiming He. MoCov3: An Empirical Study of Training Self-Supervised Vision Transformers. arXiv:2104.02057 [cs], Aug. 2021. arXiv: 2104.02057. 2", + "[16] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: UNiversal Image-TExt Representation Learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, Lecture Notes in Computer Science, pages 104-120, Cham, Sept. 2020. Springer International Publishing. 2", + "[17] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. Technical Report arXiv:2003.10555, arXiv, Mar. 2020. arXiv:2003.10555 [cs] type: article. 5", + "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1", + "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, June 2009. ISSN: 1063-6919. 5, 7", + "[20] Karan Desai and Justin Johnson. VirTex: Learning Visual Representations from Textual Annotations. arXiv:2006.06666 [cs], Sept. 2021. arXiv:2006.06666. 2", + "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 [cs], May 2019. arXiv: 1810.04805. 2, 4", + "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. ViT: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv:2010.11929 [cs], June 2021. arXiv: 2010.11929.4", + "[23] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuo-hang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. METER: An Empirical Study of Training End-to-End Vision-and-Language Transformers. arXiv:2111.02387 [cs], Nov. 2021. arXiv: 2111.02387. 2" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23440", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jiali Duan, Liquin Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal Alignment using Representation Codebook. arXiv:2203.00048 [cs], Mar. 2022. arXiv: 2203.00048. 2", + "[25] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herve Jegou, and Edouard Grave. SplitMask: Are Large-scale Datasets Necessary for Self-Supervised Pretraining? arXiv:2112.10740 [cs], Dec. 2021. arXiv: 2112.10740. 2", + "[26] Patrick Esser, Robin Rombach, and Björn Ommer. VQGAN: Taming Transformers for High-Resolution Image Synthesis. arXiv:2012.09841 [cs], June 2021. arXiv: 2012.09841.2", + "[27] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2):303-338, June 2010. 19", + "[28] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip), 2022. 8", + "[29] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale, 2022. 2", + "[30] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 8", + "[31] Xinyang Geng, Hao Liu, Lisa Lee, Dale Schuurmans, Sergey Levine, and Pieter Abbeel. M3AE: Multimodal Masked Autoencoders Learn Transferable Representations. Technical Report arXiv:2205.14204, arXiv, May 2022. arXiv:2205.14204 [cs] type: article. 1, 3, 4, 5, 7, 14, 21", + "[32] Aaron Gokaslan and Vanya Cohen. Openwebtext corpus. http://Skylion007.github.io/OpenWebTextCorpus, 2019.5", + "[33] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. arXiv:1706.02677 [cs], Apr. 2018. arXiv: 1706.02677. 15", + "[34] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. VQA2: Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. pages 6904-6913, 2017. 5, 17", + "[35] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. BYOL: Bootstrap your own latent: A new approach to self-supervised Learning. arXiv:2006.07733 [cs, stat], Sept. 2020. arXiv: 2006.07733. 2" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. 1, 8", + "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. MAE: Masked Autoencoders Are Scalable Vision Learners. arXiv:2111.06377 [cs], Nov. 2021. arXiv: 2111.06377. 1, 2, 3, 4, 5", + "[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. MoCov1: Momentum Contrast for Unsupervised Visual Representation Learning. pages 9729-9738, 2020. 2", + "[39] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9068-9077, June 2022. 1", + "[40] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen, 2022. 8", + "[41] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing Out of the Box: End-to-End Pre-Training for Vision-Language Representation Learning. pages 12976-12985, Apr. 2021. 2", + "[42] Drew A. Hudson and Christopher D. Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. pages 6700-6709, 2019. 5, 17", + "[43] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. CoRR, abs/2102.05918, 2021. 1, 4, 8", + "[44] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. ALIGN: Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. arXiv:2102.05918 [cs], June 2021. arXiv: 2102.05918.3, 14", + "[45] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1988–1997, 2017. 5, 7, 17", + "[46] Kaggle and EyePacs. Kaggle diabetic retinopathy detection. https://www.kaggle.com/c/diabetic-retinopathy-detection/data, July 2015. 14", + "[47] Wonjae Kim, Bokyung Son, and Ildoo Kim. ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision. arXiv:2102.03334 [cs, stat], June 2021. arXiv: 2102.03334. 2", + "[48] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. Technical Report arXiv:1412.6980, arXiv, Jan. 2017. arXiv:1412.6980 [cs] type: article. 15" + ], + "bbox": [ + 509, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23441", + "bbox": [ + 478, + 926, + 517, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 8", + "[50] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. arXiv:1910.13461 [cs, stat], Oct. 2019. arXiv: 1910.13461.2, 4", + "[51] Chenliang Li, Ming Yan, Haiyang Xu, Fuli Luo, Wei Wang, Bin Bi, and Songfang Huang. SemVLP: Vision-Language Pre-training by Aligning Semantics at Multiple Levels. arXiv:2103.07829 [cs], Mar. 2021. arXiv: 2103.07829. 2", + "[52] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. arXiv:2107.07651 [cs], Oct. 2021. arXiv: 2107.07651. 2, 4", + "[53] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO-2: End-to-End Unified Vision-Language Grounded Learning. arXiv:2203.09067 [cs], Mar. 2022. arXiv: 2203.09067. 2", + "[54] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, Lecture Notes in Computer Science, pages 121–137, Cham, July 2020. Springer International Publishing. 5", + "[55] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pretraining via masking, 2022. 2", + "[56] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. CoRR, abs/2110.05208, 2021. 1", + "[57] Feng Liang, Yangguang Li, and Diana Marculescu. SupMAE: Supervised Masked Autoencoders Are Efficient Vision Learners. Technical Report arXiv:2205.14540, arXiv, May 2022. arXiv:2205.14540 [cs] type: article. 3", + "[58] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 16, 19", + "[59] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic Gradient Descent with Warm Restarts. arXiv:1608.03983 [cs, math], May 2017. arXiv: 1608.03983. 5, 15", + "[60] Ilya Loshchilov and Frank Hutter. AdamW: Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs, math], Jan. 2019. arXiv: 1711.05101. 5, 15", + "[61] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. arXiv:1908.02265 [cs], Aug. 2019. arXiv: 1908.02265. 2" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Shlok Mishra, Joshua Robinson, Huiwen Chang, David Jacobs, Aaron Sarna, Aaron Maschinot, and Dilip Krishnan. A simple, efficient and scalable contrastive masked autoencoder for learning visual representations, 2022. 8", + "[63] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. SLIP: Self-supervision meets Language-Image Pre-training. arXiv:2112.12750 [cs], Dec. 2021. arXiv: 2112.12750.1", + "[64] Norman Mu, Alexander Kirillov, David A. Wagner, and Saining Xie. SLIP: self-supervision meets language-image pre-training. CoRR, abs/2112.12750, 2021. 1, 2, 5", + "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748 [cs, stat], Jan. 2019. arXiv: 1807.03748. 2, 3", + "[66] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. CoRR, abs/2111.10050, 2021. 1, 8", + "[67] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting Region-to-Phrase Correspondences for Richer Image-to-Sentence Models. pages 2641–2649, 2015. 16", + "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. CLIP: Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020 [cs], Feb. 2021. arXiv: 2103.00020. 1, 2, 3, 14, 15", + "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1, 5, 8", + "[70] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. GPT-2: Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4, 15", + "[71] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander T Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models, 2023. 3, 6, 8", + "[72] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating Diverse High-Fidelity Images with VQ-VAE-2. Technical Report arXiv:1906.00446, arXiv, June 2019. arXiv:1906.00446 [cs, stat] type: article. 2", + "[73] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenna Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 1", + "[74] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23442", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION400M: Open Dataset of CLIP-Filtered 400 Million ImageText Pairs. arXiv:2111.02114 [cs], Nov. 2021. arXiv:2111.02114.6, 14", + "[75] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision, 128(2):336-359, Feb. 2020. arXiv:1610.02391 [cs]. 8", + "[76] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. 14", + "[77] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. GCC: Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 5, 6", + "[78] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. CoRR, abs/2112.04482, 2021. 1, 4", + "[79] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A Foundational Language And Vision Alignment Model. arXiv:2112.04482 [cs], Feb. 2022. arXiv: 2112.04482. 3", + "[80] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross B. Girshick, Piotr Dólar, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. CoRR, abs/2201.08371, 2022. 1", + "[81] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of Generic Visual-Linguistic Representations. arXiv:1908.08530 [cs], Feb. 2020. arXiv: 1908.08530. 2", + "[82] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1", + "[83] Yihong Sun, Adam Kortylewski, and Alan Yuille. Amodal Segmentation through Out-of-Task and Out-of-Distribution Generalization with a Bayesian Model. Technical Report arXiv:2010.13175, arXiv, July 2022. arXiv:2010.13175 [cs] type: article. 16", + "[84] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. arXiv:1908.07490 [cs], Dec. 2019. arXiv: 1908.07490. 2", + "[85] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine learning research, 9(11), 2008. 17", + "[86] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, $\\backslash$ Lukasz Kaiser," + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 4", + "[87] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning - ICML '08, pages 1096-1103, Helsinki, Finland, 2008. ACM Press. 2", + "[88] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for all vision and vision-language tasks, 2022. 1", + "[89] Xiaolong Wang and Abhinav Gupta. Unsupervised Learning of Visual Representations using Videos. Technical Report arXiv:1505.00687, arXiv, Oct. 2015. arXiv:1505.00687 [cs] type: article. 2", + "[90] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple Visual Language Model Pretraining with Weak Supervision. arXiv:2108.10904 [cs], Aug. 2021. arXiv: 2108.10904. 2", + "[91] Longhui Wei, Lingxi Xie, Wengang Zhou, Houqiang Li, and Qi Tian. MVP: Multimodality-guided Visual Pretraining. arXiv:2203.05175 [cs], Mar. 2022. arXiv: 2203.05175. 3", + "[92] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 14", + "[93] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dólar, and Ross Girshick. Early Convolutions Help Transformers See Better. Technical Report arXiv:2106.14881, arXiv, Oct. 2021. arXiv:2106.14881 [cs] type: article. 5", + "[94] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. Pre-layernorm: On Layer Normalization in the Transformer Architecture. In Proceedings of the 37th International Conference on Machine Learning, pages 10524-10533. PMLR, Nov. 2020. ISSN: 2640-3498. 4, 15", + "[95] Jianwei Yang, Chunyuan Li, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Focal modulation networks, 2022. 1", + "[96] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained Interactive Language-Image Pre-Training. arXiv:2111.07783 [cs], Nov. 2021. arXiv: 2111.07783. 2", + "[97] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. ERNIE-ViL: Knowledge Enhanced Vision-Language Representations Through Scene Graph. June 2020. 2", + "[98] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models, 2022. 1" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23443", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[99] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive Captioners are Image-Text Foundation Models. arXiv:2205.01917 [cs], May 2022. arXiv: 2205.01917. 2", + "[100] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A New Foundation Model for Computer Vision. arXiv:2111.11432 [cs], Nov. 2021. arXiv: 2111.11432. 2", + "[101] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-level Representation Learning for Self-supervised Vision Transformers. Technical Report arXiv:2206.07990, arXiv, June 2022. arXiv:2206.07990 [cs] type: article. 2", + "[102] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. arXiv:2103.03230 [cs, q-bio], June 2021. arXiv: 2103.03230. 2", + "[103] Yan Zeng, Xinsong Zhang, and Hang Li. X:VLM: MultiGrained Vision Language Pre-Training: Aligning Texts with Visual Concepts. arXiv:2111.08276 [cs], Feb. 2022. arXiv: 2111.08276. 3", + "[104] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12104-12113, June 2022. 3", + "[105] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, André Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, Lucas Beyer, Olivier Bachem, Michael Tschannen, Marcin Michalski, Olivier Bousquet, Sylvain Gelly, and Neil Houlsby. The visual task adaptation benchmark. CoRR, abs/1910.04867, 2019. 1, 14", + "[106] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djo-longa, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, Lucas Beyer, Olivier Bachem, Michael Tschannen, Marcin Michalski, Olivier Bousquet, Sylvain Gelly, and Neil Houlsby. VTAB: A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark. arXiv:1910.04867 [cs, stat], Feb. 2020. arXiv: 1910.04867.5,7,14", + "[107] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. ADE20K: Semantic Understanding of Scenes through the ADE20K Dataset. arXiv:1608.05442 [cs], Oct. 2018. arXiv: 1608.05442. 19" + ], + "bbox": [ + 80, + 92, + 467, + 805 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "23444", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_model.json b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..11a6b7e5e031ff4cba3a96c7af34b2e48c11ae8c --- /dev/null +++ b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_model.json @@ -0,0 +1,2987 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.873, + 0.153 + ], + "angle": 0, + "content": "Masked Autoencoding Does Not Help Natural Language Supervision at Scale" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.181, + 0.873, + 0.199 + ], + "angle": 0, + "content": "Floris Weers Vaishaal Shankar Angelos Katharopoulos Yinfei Yang Tom Gunter" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.202, + 0.787, + 0.217 + ], + "angle": 0, + "content": "{fweers,vaishaal_shankar,a_katharopoulos,yinfeiy,tom_gunter}@apple.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.262, + 0.314, + 0.279 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.474, + 0.584 + ], + "angle": 0, + "content": "Self supervision and natural language supervision have emerged as two exciting ways to train general purpose image encoders which excel at a variety of downstream tasks. Recent works such as M3AE [31] and SLIP [63] have suggested that these approaches can be effectively combined, but most notably their results use small (<20M examples) pre-training datasets and don't effectively reflect the large-scale regime (>100M samples) that is commonly used for these approaches. Here we investigate whether a similar approach can be effective when trained with a much larger amount of data. We find that a combination of two state of the art approaches: masked auto-encoders, MAE [37] and contrastive language image pre-training, CLIP [68] provides a benefit over CLIP when trained on a corpus of 11.3M image-text pairs, but little to no benefit (as evaluated on a suite of common vision tasks) over CLIP when trained on a large corpus of 1.4B images. Our work provides some much needed clarity into the effectiveness (or lack thereof) of self supervision for large-scale image-text training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.612, + 0.21, + 0.629 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Large scale pretraining has become a powerful tool in the arsenal of computer vision researchers to produce state of the art results across a wider variety of tasks [39,88,95,98]. However, when pre-training on tens of millions to billions of images it is difficult to rely on standard supervised methods to train models, as datasets of this size often lack reliable labels. In the presence of these massive but largely under-curated datasets, two general classes of methods to train general purpose image encoders have emerged:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.47, + 0.817 + ], + "angle": 0, + "content": "1. Self Supervised techniques that learn visual representations from the image data alone [11,36]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.828, + 0.47, + 0.874 + ], + "angle": 0, + "content": "2. Natural Language Supervised methods that utilize paired free-form text data to learn visual representations [43, 69]" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.786, + 0.47, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Due to the unique strengths and weaknesses of each ap-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.895, + 0.506 + ], + "angle": 0, + "content": "proach\\(^1\\), a recent flurry of work has introduced methods that combine both forms of supervision [31, 56, 64, 78] to varying degrees of success. While each of these methods establishes some regime where the additional supervision helps, none of these \"joint-supervision\" methods advance state of the art in any meaningful way. Additionally, to our knowledge none of these methods have shown comparative results at the scale many large scale vision models are currently trained at (\\(>100\\mathrm{M}\\) examples) [43, 66, 69, 73, 80, 82, 98]. Furthermore, methods that use both forms of supervision start with the presumption that the additional supervision is helpful and either often lack clean ablations or lack evaluations in a \"high accuracy\" regime—leading to further confusion regarding whether a combination of these methods can actually improve the state of the art. To clarify this issue, in this work, we investigate a simple question:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.514, + 0.861, + 0.56 + ], + "angle": 0, + "content": "Does a combination of self supervision and natural language supervision actually lead to higher quality visual representations?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.765 + ], + "angle": 0, + "content": "In order to answer this, we first introduce a straightforward baseline approach that combines standard self-supervision and language supervision techniques. We combine masked auto-encoders (MAE) and contrastive language image-pretraining (CLIP) to make MAE-CLIP. We then present a careful study of the performance of MAE, M3AE, CLIP and MAE-CLIP across a wide variety of tasks in two distinct regimes: a \"low-sample\" \\(^{2}\\) 11.3 million example regime and a \"high-sample\" 1.4 billion example regime. We train self-supervised and language-supervised methods using the same pre-training datasets under the assumption that we have no knowledge about downstream tasks. Our experiments show:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.771, + 0.894, + 0.832 + ], + "angle": 0, + "content": "1. In the low sample size regime, without changing the final pooling operation in the network, we observe a large performance improvement, namely \\(6\\%\\) on ImageNet [18] and \\(4\\%\\) on VTAB [105]. However, when" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Self supervised methods can learn representations without labels, but natural language supervision learns better representations. Natural language supervised methods rely on quality of captions" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.876, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2We note that what low sample means has changed substantially over the last few years" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23432" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.136 + ], + "angle": 0, + "content": "we modify the pooling operation, the improvement substantially decreases to around \\(1\\%\\) on both ImageNet and VTAB." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.15, + 0.468, + 0.194 + ], + "angle": 0, + "content": "2. In the high sample size regime, there is virtually no difference in performance between MAE-CLIP and CLIP across ImageNet, VTAB, and VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.208, + 0.468, + 0.252 + ], + "angle": 0, + "content": "We believe our work is the first careful study of this form and contextualizes recent progress in both self-supervision and natural language supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.468, + 0.373 + ], + "angle": 0, + "content": "The rest of the paper is organized as follows: In Section 2, we cover related work in the areas of self supervision and natural language supervision. In Section 3, we give an overview of the baseline methods we study, MAE, M3AE, CLIP and our new baseline MAE-CLIP. Then we present and analyse our small scale and large scale experimental findings in Sections 4 and 5. Finally, we discuss potential explanations for our findings and some future work in 6." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.39, + 0.218, + 0.405 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.468, + 0.46 + ], + "angle": 0, + "content": "Our work combines natural language supervision and self-supervision in a multi-task approach to visual encoding, and so research from these three areas is relevant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.468, + 0.853 + ], + "angle": 0, + "content": "Natural language supervision for visual encoding covers a variety of approaches that assume access to datasets of images or videos associated with text. Some of the most successful use a contrastive pairwise alignment signal applied to very large batch and dataset sizes [68]. This large batch size means hard negative pairs can be produced by random sampling, avoiding the need for a memory bank [12, 13, 38] or momentum distillation [15]. FILIP [96] further improves zero-shot performance for some tasks by using a more fine-grained elementwise contrastive loss. Image captioning has also shown promise as a pre-training task [20, 90], with [20] in particular demonstrating strong data efficiency. UNITER [16] pursues a similar idea, applying a generative loss to both image and text modalities. [2, 16, 23, 41, 47, 51, 61, 81, 84, 97] apply the masked patch prediction problem from [21, 37, 50] to a joint image-text data space. These approaches either use a pre-trained convolutional neural network (CNN) to generate region of interest (RoI) proposal encodings, or predict the labelled class of the masked patch instead of the raw pixel values (or the quantized patch ID). UNIMO-2 [53] attempts to ground image and text patches for a single data example into a top-k quantized set of embeddings, before passing the grounded embeddings along with the raw non-grounded embeddings into a decoder, targeting a loss based on masked language modeling (MLM) and image-text-matching (ITM)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Self-supervised approaches for visual encoding may be loosely categorized into those that target consistency constraints between multiple views of the same scene and those" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.636 + ], + "angle": 0, + "content": "that attempt visual reconstruction on corrupted or conditioning representations of images. Consistency constraints are often derived through data-augmentation applied to a single real image [6, 10, 14, 35, 102], but these may also take the form of e.g. a temporal constraint if video data is available [89], or be applied at an image-patch level by making a smoothness assumption and selecting neighboring patches as 'consistent' whilst treating distant patches (or those sourced from another image) as negative examples [101]. Forms of denoising autoencoder (DAE) [87] are popular as a means of self-supervision, and have been investigated at a variety of scales; iGPT [9] learns to generate images as a row-major flattened sequence of pixels, whilst [65], [72], and [26] compress images into a short sequence of discrete codes before then regenerating and scoring at a pixel level. BEiT [3] first encodes an image into a sequence of discrete codes before masking and then predicting a subset of the codes. MAE [37] drops the discretization, instead predicting raw pixel loss for a subset of the encoded patches in a manner strongly reminiscent of BART [50]. SplitMask [25] applies both a masked image patch prediction loss and a pairwise contrastive loss, using the masked example as well as the inverted masked example to form each positive pair. Meanwhile, M3AE [24] and VL-BEiT [4] both propose a masked patch prediction problem applied to image and text modalities jointly. Recent concurrent work shows that while masking does not help generalization, the masking can be used to speed up training by dropping the masked tokens during the forward and backward pass [55]. EVA [29] shows that doing masked image modeling to predict the output embeddings (rather than pixels) of the masked image can be an effective pretraining task when combined with natural language supervision. We note for EVA the masked image modeling occurs only on smaller well curated datasets such as ImageNet, Conceptual Captions and COCO." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-task methods for pre-training visual encoders are a highly active area of research. CoCA [99] combines the cross-modality contrastive task from [68] with image captioning in the style of [20], and by pre-training at very large scale show that the resulting model is more performant than prior art across a very broad array of downstream visual understanding tasks. SLIP [64] combines a SimCLR style self supervision loss with the standard CLIP contrastive loss to train a jointy self-supervised and natural language supervised model, however while they show large performance gains they are all in a low accuracy regime (below \\(40\\%\\) Top-1 accuracy on ImageNet). Florence [100] employs a diverse array of tasks and datasets at pre-training time (including object detection, CLIP-style contrastive modeling, Visual Question Answering (VQA), and supervised classification). ALBEF [52] builds on a CLIP-style architecture by incorporating a single-stream encoder that consumes both" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.936 + ], + "angle": 0, + "content": "23433" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.078, + 0.817, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.254, + 0.775, + 0.269 + ], + "angle": 0, + "content": "Figure 1. MAE-CLIP: CLIP [68] augmented with a generative decoder in the style of MAE [31]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.277, + 0.473, + 0.685 + ], + "angle": 0, + "content": "modalities. It uses a masked language modeling loss for the text modality and an image-text matching loss for within-batch hard-negatives, mined according to contrastive similarity. FLAVA [79] also uses masked-image modeling, and uses both single and joint modality encoders, with additional single-modality datasets (i.e. not just image-text pairs) that allow it to generalize to longer text inputs and visual data which is unlikely to be captioned. X-VLM [103] utilizes image data which includes object-labelled bounding boxes to extend the image-text matching and contrastive losses to cover within-image regions as well as whole-image captions. They also include a masked language modelling objective, achieving impressive performance on many zero-shot downstream tasks—although the requirement for object bounding box data may not scale well for large pre-training datasets. SupMAE [57] adds a supervised labeling task to the MAE architecture, demonstrating that this results in better data efficiency. MVP [91], meanwhile, finetunes a pre-trained CLIP backbone using MIM, and demonstrates that starting with a pre-trained model improves downstream visual recognition performance. MAE-CLIP is reminiscent of M3AE (with the addition of a contrastive task), ALBEF (but applying the masked prediction task to both modalities, dropping the momentum distillation), and most of all FLAVA (without the single-modality masked-patch-prediction tasks, and with the use of high ratio random masking following MAE and M3AE)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.698, + 0.206, + 0.714 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.427, + 0.741 + ], + "angle": 0, + "content": "3.1. Contrastive language-image pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "CLIP [68] and ALIGN [44] both demonstrated that contrastive image-language supervision applied at scale was capable of producing an image encoder that excels at a range of downstream tasks—often in a zero-shot setting. In our implementation, we begin by following the design outlined in [68]. Specifically, we use per-modality encoders to produce a single \\(\\ell 2\\)-normalized dense embedding vector for each image or text input, before applying a pairwise InfoNCE [65] loss to a large global batch of paired image-text embeddings, using all non-paired examples as hard nega" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.892, + 0.308 + ], + "angle": 0, + "content": "tives. We define the image-to-text loss \\( L_{i2t} \\) and the text-to-image loss \\( L_{t2i} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.318, + 0.892, + 0.36 + ], + "angle": 0, + "content": "\\[\nL _ {i 2 t} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(x _ {j} ^ {T} y _ {j} / \\sigma\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(x _ {j} ^ {T} y _ {k} / \\sigma\\right)} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.372, + 0.892, + 0.414 + ], + "angle": 0, + "content": "\\[\nL _ {t 2 i} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(y _ {j} ^ {T} x _ {j} / \\sigma\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(y _ {j} ^ {T} x _ {k} / \\sigma\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.892, + 0.495 + ], + "angle": 0, + "content": "where \\(N\\) is the global batch size, \\(x_{j}\\) is the normalized embedding of the image for the \\(j\\)-th pair and \\(y_{k}\\) is the normalized embedding of the text in the \\(k\\)-th pair in the batch. \\(\\sigma\\) is a learnable temperature parameter. The total contrastive loss \\(L_{c} = \\frac{1}{2} (L_{t2i} + L_{i2t})\\) is the average of these two losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.585 + ], + "angle": 0, + "content": "We use transformer encoders for both modalities, however, differently to [68] we explore several strategies for aggregating the image-patch or text-token outputs into a single embedding (see Section 4.4). We also eschew the autoregressive masking strategy for the text encoder, instead allowing full bi-directional self attention following [44]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.608, + 0.619 + ], + "angle": 0, + "content": "3.1.1 Pooling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.824 + ], + "angle": 0, + "content": "In CLIP [68] the authors use a separate 'CLS' token projected through the network as the overall image representation. On the other hand, ALIGN [44] use global-average pooling over the encoded visual features. Whilst both approaches produced good downstream evaluation results, recent work [8, 71, 104], suggested that the choice of pooling strategy can strongly influence the quality of visual semantic embeddings. In particular, [71] demonstrated that this effect is present for CLIP-like visual encoders. Noting this, we opt to investigate three pooling strategies: 1. the default multihead-attention pooling (MAP), 2. global average pooling (GAP)—both described in [104]—as well as 3. non-maximal suppression pooling (MAX) as in [8, 71]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.707, + 0.847 + ], + "angle": 0, + "content": "3.2. Masked Autoencoders" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In MAE [37] the authors demonstrate a simple technique for self-supervised image-encoder pre-training that—to our knowledge—is still considered state-of-the-art. They use a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "23434" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.378 + ], + "angle": 0, + "content": "ViT [22] encoder-decoder architecture and apply it to heavily masked input images. The input to the encoder consists of the visible, unmasked image patches, first embedded via a linear projection before additive positional embeddings are applied, and then the result is fed through the encoder's transformer layers. They demonstrate that a very high patch masking ratio is critical to achieving good performance, and usually retain only \\(25\\%\\) of the image. The decoder consumes the output of the encoder, as well as a learned 'masked-patch' embedding, which is included to represent each masked token. Positional embeddings are also reapplied, to ensure that spatial information is passed through the decoder for the masked patches. The output of the decoder at the masked positions is then measured against ground truth using a simple mean-squared-error loss. As in [37] we experiment with predicting both normalized and un-normalized patch values, finding that predicting the normalized patch value slightly improves the performance of our MAE implementation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.472, + 0.517 + ], + "angle": 0, + "content": "Masking across modalities: BART [50] applies a similar strategy to text data, but uses a cross-entropy loss over the masked-token output distribution and a far lower masking ratio (typically around \\(15\\%\\)). M3AE [31] extends MAE and BART to incorporate inputs from both text and image modalities, using per-modality input and output projections. Otherwise the encoder-decoder architecture resembles MAE, with the addition of learned modality-indicating embeddings to each transformer input." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.518, + 0.471, + 0.581 + ], + "angle": 0, + "content": "The M3AE loss relies only on the contents of the image-text pairs, and comprises mean-squared error for the masked image patches \\((L_{gen.i})\\) and cross-entropy over the vocabulary for the masked text tokens \\((L_{gen.t})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.607, + 0.47, + 0.65 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {g e n . i}} = \\frac {1}{N} \\sum_ {j} ^ {N} \\left(p _ {j} - P _ {j}\\right) ^ {2} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.653, + 0.47, + 0.697 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {g e n . t}} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(t _ {n , T _ {n}}\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(t _ {n , c}\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.711, + 0.47, + 0.788 + ], + "angle": 0, + "content": "where \\( p_j \\) and \\( P_j \\) refer to the predicted and ground truth pixel value respectively and \\( t_n \\) and \\( T_n \\) refer to the predicted token distribution and the ground truth token, with \\( C \\) as the number of unique tokens. For the image reconstruction loss, we normalize the ground-truth per-patch, following [37]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.206, + 0.815 + ], + "angle": 0, + "content": "3.3. MAE-CLIP" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As the name suggests, MAE-CLIP attempts to incorporate aspects of MAE/M3AE into CLIP through the addition of a single dual-modality transformer decoder (as in M3AE) which consumes the output of the CLIP encoders. We show this in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.644, + 0.108 + ], + "angle": 0, + "content": "3.3.1 Components" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.118, + 0.892, + 0.149 + ], + "angle": 0, + "content": "The model architecture is similar to ALBEF [52] and FLAVA [78], consisting of three components." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.15, + 0.894, + 0.227 + ], + "angle": 0, + "content": "An image encoder: Following ViT [22], we divide the input image into equally-sized, non-overlapping patches. After applying a linear projection and adding a 2-D position encoding, we feed the per-patch representations through a number of transformer layers [86]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.894, + 0.319 + ], + "angle": 0, + "content": "A text encoder: Following ALIGN [43], our text encoder is based on BERT [21], in which text is first tokenized and embedded, and a 1-D trainable position encoding is added. Differently to BERT, we use pre-layernorm [94] and initialize the parameters according to the simplified scheme outlined in [70]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.32, + 0.893, + 0.502 + ], + "angle": 0, + "content": "A cross-modality decoder: The decoder receives per-element encoded image and text representations from the encoders for both masked and un-masked elements. For the image modality, masked patches are added by replacing their values with a shared trainable mask token. Positional encodings are once again added to all elements, following [37]. We also add a per-modality trainable encoding to allow the decoder to easily distinguish between the two modalities. The decoder uses the same transformer implementation as the encoders. The output of the final decoder layer is then projected into a per-patch-modality output space and the overall loss is computed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.6, + 0.539 + ], + "angle": 0, + "content": "3.3.2 Losses" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.892, + 0.598 + ], + "angle": 0, + "content": "MAE-CLIP employs the losses of both CLIP (Eq. 1, 2) and M3AE (Eq. 3, 4). Our final loss is a weighted sum of the losses from each task, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.611, + 0.892, + 0.641 + ], + "angle": 0, + "content": "\\[\nL = \\frac {1}{2} \\left(L _ {t 2 i} + L _ {i 2 t}\\right) + w _ {i} \\cdot L _ {\\text {g e n . i}} + w _ {t} \\cdot L _ {\\text {g e n . t}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.653, + 0.892, + 0.728 + ], + "angle": 0, + "content": "where \\( w_{i} \\) and \\( w_{t} \\) are scalars used to control the relative weight of the generative losses. We always provide paired image-text inputs to the model, leaving it to future work to explore the benefits of also incorporating single modality inputs and reconstruction losses, at scale." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.73, + 0.892, + 0.836 + ], + "angle": 0, + "content": "To avoid severely impacting the CLIP loss via computing it with masked inputs, we run the encoders twice: once with full unmasked input in order to compute the CLIP loss, before then making a second pass using only the unmasked input to compute the M3AE loss. Finally, we compute the weighted sum of the two losses, and use this to calculate the overall update." + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.851, + 0.894, + 0.901 + ], + "angle": 0, + "content": "3 Alternative, more memory efficient, strategies exist for computing the overall updates (e.g. calculate per-loss gradients, or perform round-robin updates), but prior work [1] has demonstrated that some of these techniques may negatively impact the overall training result." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23435" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.47, + 0.252 + ], + "angle": 0, + "content": "In this section, we present our experimental results to study how the addition of self-supervision affects the representation quality of natural language supervised models. We study standard masked auto-encoders (MAE), multimoddal masked auto-encoders (M3AE), contrastive language image pretraining (CLIP) and our newly introduced baseline method (MAE-CLIP), which combines mask based self-supervision with contrastive language-image learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.262, + 0.267, + 0.279 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.47, + 0.45 + ], + "angle": 0, + "content": "Architecture: Experiments are performed using an image encoder based on the architecture described as ViT-B-16 in [93], in combination with a 12-layer, 512-wide text encoder with 8 attention heads. For the decoder, we use another 12-layer, 512-wide transformer with 8 attention heads, identical to the text encoder. We also use a byte-pair encoding tokenizer, fit to the OpenWebText corpus [32]. When training MAE-CLIP, we replace the random masking by a similarity masking strategy, which makes use of the element-wise CLIP similarity scores to select the masked element for the reconstruction task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Pretraining Datasets: Our analysis is divided into two sections. First, in sections 4.2 to 4.4, we present our study on the \"low-sample\" regime using the combination of CC12M [7] and CC3M [77] excluding all images whose text contains a [PERSON] tag5. The final dataset contains 11.3 million comparatively high quality image-text pairs. We refer to this dataset as CC or \"small\" scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.471, + 0.844 + ], + "angle": 0, + "content": "Pretraining configuration: We use a minibatch size of 16,384 samples and train using the AdamW optimizer [60] with decoupled weight decay. We use a base learning rate of \\(5 \\times 10^{-4}\\) and train using cosine learning rate decay [59] with an initial linear warmup of 200 steps. We reduce the \\(\\beta_{2}\\) decay factor to 0.98, and increase \\(\\epsilon\\) to \\(10^{-6}\\), finding that these changes improved rate of convergence for all models trained. To increase training convergence speed, we switch from a local (per-GPU) to a global contrastive loss after 500 steps. During the global contrastive loss phase, we set the generative text and image loss weights to \\(w_{t} = 0.05\\) and \\(w_{i} = 0.1\\) respectively. For all models which rely on masking, we use a \\(75\\%\\) masking ratio, consistent with [31,37], as we did not find an alternative that improved downstream results. In total, we train our models on CC for 32 epochs, which corresponds to \\(\\sim 22,000\\) steps. For our M3AE training runs, we use per-modality encoders so as to produce a baseline that is flop and parameter matched to our MAE-CLIP runs." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.089, + 0.842, + 0.182 + ], + "angle": 0, + "content": "
ModelsZero-shotLinear Probing
MAE-33.9
M3AE-52.5
CLIP29.752.6
MAE-CLIP33.858.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 1. ImageNet classification with zero-shot transfer or linear probing after pretraining on the \\( CC \\) dataset (11.3M images). MAE-CLIP significantly improves the classification performance of CLIP in the small scale regime." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.256, + 0.692, + 0.273 + ], + "angle": 0, + "content": "4.2. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.414 + ], + "angle": 0, + "content": "We evaluate the quality of the representations learned from pre-training on \\( CC \\) by measuring the image classification performance, either by zero-shot transfer, or by training a linear classifier using the predicted visual features. For zero-shot classification, we use the average of 80 prompts and follow the routine prescribed in [69]. To train the linear classifier, we pre-compute features using the visual encoder6 and run AdamW for 20 to 80 epochs with a learning rate of 0.01." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Table 1 shows the zero-shot and linear probing classification results on the ImageNet [19] dataset while Table 2 shows our linear probing results on most (17) of the VTAB [106] benchmark datasets7. Initially, we observe that the combination of self-supervision and natural language supervision provides a consistent and substantial improvement over either form of supervision by itself. This result concurs with previous works such as [31] and [64], which show that self-supervision aids natural language supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.577, + 0.577 + ], + "angle": 0, + "content": "4.3.VQA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Subsequently, we compare the performance of MAE-CLIP to the rest of the baselines on the visual question answering task using three datasets. CLEVR [45], VQAv2 [34] and GQA [42]. VQA assesses the model's multimodal reasoning capabilities as well as its visual grounding. In order to finetune our models for VQA, we freeze the image and text encoders and either randomly reinitialize the decoder in models such as MAE-CLIP or M3AE or add a new identical decoder for CLIP. During finetuning, we use a layer-wise learning rate decay, following [3, 17, 37]. For CLEVR and VQAv2 we finetune the decoder for 50 epochs, while for GQA we finetune for 5 epochs on the \"train-all\" split and 2 epochs on the \"train-balanced\" split following the protocol of [54]. In all cases, we treat the problem as a classification problem of selecting one answer out of the set of possible answers for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.827, + 0.892, + 0.842 + ], + "angle": 0, + "content": "In Table 3 we observe that self-supervision combined" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.468, + 0.875 + ], + "angle": 0, + "content": "4We compare the similarity masking strategy with the original random masking strategy on various tasks, the results show they perform similarly." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.876, + 0.468, + 0.9 + ], + "angle": 0, + "content": "5 Preliminary experiments on CLIP suggest removing such data leads to better performance." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.851, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.851, + 0.708, + 0.864 + ], + "angle": 0, + "content": "6We use no data augmentation here" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.864, + 0.892, + 0.899 + ], + "angle": 0, + "content": "7We do not include results for either the diabetic retinopathy or the Sun397 tasks, due to licensing issues. More details in the supplementary material" + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.851, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "23436" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.075, + 0.892, + 0.102 + ], + "angle": 0, + "content": "Table 2. Linear probing accuracy (%) on classification tasks. All models are trained on the CC Dataset (11.3M images). (● VTAB/natural, ● VTAB/specialized and ● VTAB/structured.)" + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.104, + 0.85, + 0.248 + ], + "angle": 0, + "content": "
Caltech101CIFAR-100DTDFlowers102PetsSVHNEuroSATCamelyonResisc45Clevr/CloestClevr/CountDMLabdSprites/OridSprites/LocKITTI/DistsNORB/AzimsNORB/ElevAverage
MAE75.356.258.570.237.471.996.282.884.065.160.642.837.979.936.737.464.562.2
M3AE85.666.669.486.558.469.297.482.990.769.471.248.249.883.342.634.373.469.3
CLIP84.262.857.781.669.752.595.582.686.753.752.944.845.961.345.731.042.661.8
MAE-CLIP89.866.264.886.574.556.795.481.686.953.659.645.944.668.648.036.151.065.3
" + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.253, + 0.416, + 0.339 + ], + "angle": 0, + "content": "
ModelCLEVRVQAv2GQA
MAE93.545.746.1
M3AE97.555.850.9
CLIP87.555.650.0
MAE-CLIP96.058.552.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.419 + ], + "angle": 0, + "content": "Table 3. VQA finetuning results for models pre-trained on CC Dataset (11.3M images). We train a new decoder for all methods while keeping the encoders frozen. MAE-CLIP performs significantly better than either only self-supervised or language supervised methods by themselves." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.595 + ], + "angle": 0, + "content": "with natural language supervision (MAE-CLIP) performs consistently better than either MAE or CLIP by themselves. Combining information from both modalities is evidently critical for effectively answering visual questions as the performance of MAE on VQAv2 and GQA is several percentage points lower than the rest of the methods. Interestingly, methods that employ a form of self-supervision perform significantly better on CLEVR. We argue that this discrepancy is due to CLEVR requiring object localization and spatial reasoning which benefits from this form of self-supervision." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.602, + 0.241, + 0.618 + ], + "angle": 0, + "content": "4.4. Pooling Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.837 + ], + "angle": 0, + "content": "To better understand whether the gap between MAE-CLIP and CLIP we observed in the previous section is fundamental, we investigate the effect of the pooling operator on different downstream tasks. Table 4 compares the performance of CLIP and MAE-CLIP with the standard multi-head attention pooling (MAP), global average pooling (GAP) and max pooling (MAX) on all image classification tasks and visual question answering. We note that GAP and MAX pooling perform substantially better than the standard MAP pooling across ImageNet, VTAB and VQA tasks. A hypothesis posed by [71] is that the alternative pooling operators lead to better perceptual grouping in the visual representation. We study this with a qualitative analysis in Section 6.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.279, + 0.862 + ], + "angle": 0, + "content": "5. Experiments at Scale" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "At small-scale, we showed that self-supervision combined with natural language supervision marginally im" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.253, + 0.879, + 0.394 + ], + "angle": 0, + "content": "
ModelPoolingImageNetVTABVQA
ZSLPLPFT
CLIPMAP29.752.661.864.4
MAE-CLIPMAP33.858.965.368.9
CLIPGAP29.359.870.665.1
MAE-CLIPGAP33.562.571.769.2
CLIPMAX33.162.368.763.5
MAE-CLIPMAX35.263.269.168.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Table 4. Results of CLIP and MAE-CLIP with different pooling options trained on CC dataset (11.3M images). We compare the models on ImageNet Zero-Shot (ZS), ImageNet Linear-Probing (LP), VTAB 17-task Linear-Probing average, and VQA 3 tasks Fine-Tuning (FT) average." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.484, + 0.892, + 0.56 + ], + "angle": 0, + "content": "proves the quality of the learned visual representations, and that the choice of image-encoder pooling strategy can heavily influence results. We now investigate whether these conclusions still hold when training on a much larger, 1.4B example dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.691, + 0.589 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.597, + 0.891, + 0.626 + ], + "angle": 0, + "content": "We follow the configuration described in Section 4.1, with a few key differences, noted below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.808 + ], + "angle": 0, + "content": "Pretraining Dataset: We combine a 2.2B example web-crawled image-text pair dataset, termed the English-WebImage-Text dataset (EWIT-2.2B), with LAION-400M [74], CC3M [77], CC12M [7] and an internal high-quality image-text pair dataset containing approximately 134M image text pairs which we term the High Quality Image Text Pairs Dataset (HQITP-134M). We globally deduplicate this—keyed by image bytes—reducing the number of image-text pairs significantly, to yield a final 1.4B examples. Details provided in the supplementary material. In the rest of the paper we refer to this as the web-crawled or large-scale dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pretraining configuration: We increase the learning rate warm-up to 1,000 steps, and compute a local contrastive loss for the first 10,000 steps, as during early training this improves the models rate of convergence. We train for a total of 480,000 steps, which corresponds to 6 full passes through the dataset. We do not increase the batch" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "23437" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 5. Linear probing accuracy (%) on classification tasks. Models are all trained on our web-crawled dataset (1.4B images). (VTAB/natural, VTAB/specialized and VTAB/structured.) In the large scale pretraining regime, the difference between MAE-CLIP and CLIP is reduced to \\( < 1\\% \\). * At evaluation time, our M3AE model was \\( 50\\% \\) trained, so that performance may improve further." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.133, + 0.86, + 0.304 + ], + "angle": 0, + "content": "
Caltech101CIFAR-100DTDFlowers102PetsSVHNEuroSATCarnelyonResisc45Clevi/CloestClevi/CountDMLabdSprites/OridSprites/LocKITTI/DistsNORB/AzimsNORB/ElevAverage
M3AE*93.074.878.295.481.269.697.384.892.765.774.751.052.780.747.536.268.873.2
CLIP94.978.480.097.386.959.094.182.392.745.662.146.046.153.350.920.335.866.2
CLIPMAX96.181.080.997.389.965.796.083.294.152.867.849.959.567.641.223.445.870.1
MAE-CLIPMAX95.879.281.596.888.262.195.881.893.052.066.949.653.772.553.032.345.470.6
CLIPGAP95.880.581.697.688.766.097.084.493.356.771.453.358.070.150.638.355.172.9
MAE-CLIPGAP95.479.382.297.488.672.896.684.593.557.573.652.757.571.251.645.655.273.8
" + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.316, + 0.432, + 0.417 + ], + "angle": 0, + "content": "
ModelsZero-shotLinear Probing
M3AE*-69.3
CLIPGAP61.875.9
CLIPMAX63.777.5
MAE-CLIPGAP57.475.7
MAE-CLIPMAX60.976.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.426, + 0.471, + 0.469 + ], + "angle": 0, + "content": "Table 6. ImageNet classification after pretraining on web-crawled dataset (1.4B images). In the large scale regime, self-supervision does not complement natural language supervision." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.316, + 0.851, + 0.416 + ], + "angle": 0, + "content": "
ModelCLEVRVQAv2GQA
M3AE96.959.953.3
CLIPGAP87.861.855.0
CLIPMAX89.560.653.6
MAE-CLIPGAP92.861.955.3
MAE-CLIPMAX93.961.553.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.426, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Table 7. VQA finetuning results on web-crawled dataset (1.4B images). Similar to Table 15a, for large pretraining datasets, self-supervision is not improving the downstream VQA performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.445, + 0.502 + ], + "angle": 0, + "content": "size, and use the same learning rate schedule as for \\( CC \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Model variants: We train M3AE as well as both GAP and MAX variants of CLIP and MAE-CLIP, having noted in the small-scale regime that these are the highest performing architectures and pre-training strategies. We do not train MAE at scale due to resource limitations, and because its performance was significantly lower than the models that utilize natural language supervision—a finding that is consistent with prior literature [31]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.269, + 0.652 + ], + "angle": 0, + "content": "5.2. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Table 15a provides encoder performance for both zero-shot and linear-probing on ImageNet [19]. We note that at scale, MAE-CLIP typically shows worse performance than CLIP across both pooling strategies and tasks. We hypothesize that this may be due to the fact that masked-patch-prediction results in model capacity being lost to the task of predicting semantically irrelevant patches, meaning that we lose some of the benefits provided by the larger scale dataset (despite tuning the relative loss weights). If we are correct in this hypothesis then it is possible that a larger model capacity would help to alleviate this issue. We leave this to a future investigation. Meanwhile, in Table 14 we once again explore the encoder performance through linear-probing for VTAB [106]. We note that GAP provides a consistent improvement over MAX across both CLIP and MAE-CLIP models, and that M3AE is also a strong performer—as was" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.591 + ], + "angle": 0, + "content": "the case in the small-scale regime. M3AE provides particular benefit to the \"structured\" VTAB tasks, however underperforms (sometimes by a large margin) on the \"natural\" tasks. MAE-CLIP ends up in between the two—doing better than CLIP but worse than M3AE on the structured tasks, and worse than CLIP but better than M3AE on the natural tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.612, + 0.577, + 0.627 + ], + "angle": 0, + "content": "5.3.VQA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.774 + ], + "angle": 0, + "content": "We follow the same finetuning strategy as before, with results in Table 15b. Notably, we do not see a consistent difference across tasks for CLIP vs MAE-CLIP, with the possible exception of CLEVR [45], where it is clear that M3AE is by far the best performing approach—and its benefits improve the MAE-CLIP results on that task. GAP is perhaps a marginally stronger pooling strategy as measured against MAX, however the difference is smaller than it was in the small-scale regime." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.797, + 0.725, + 0.814 + ], + "angle": 0, + "content": "6. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "While our results present a pessimistic view on combining self-supervision with natural language supervision, we don't have a clear reason as to why self-supervision fails at scale. Here we present two hypotheses that may be explored in future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "23438" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.09, + 0.217, + 0.101 + ], + "angle": 0, + "content": "(a) CLIP" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.103, + 0.251, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.09, + 0.351, + 0.1 + ], + "angle": 0, + "content": "(b) MAE-CLIP" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.103, + 0.368, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.091, + 0.461, + 0.102 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{CLIP}_{\\mathrm{GAP}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.103, + 0.484, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.09, + 0.596, + 0.102 + ], + "angle": 0, + "content": "(d) MAE-CLIPGAP" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.103, + 0.603, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.091, + 0.697, + 0.101 + ], + "angle": 0, + "content": "(e) \\(\\mathrm{CLIP}_{\\mathrm{MAX}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.103, + 0.72, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.091, + 0.829, + 0.101 + ], + "angle": 0, + "content": "(f) MAE-CLIPMAX" + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.103, + 0.836, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.895, + 0.335 + ], + "angle": 0, + "content": "Figure 2. Zero-shot segmentation masks (top row) and GradCAM [75] visualizations (bottom row) for the prompt \"a photo of a dog\" for CLIP and MAE-CLIP. Self-supervision qualitatively improves the visual grounding of the model, however, the choice of pooling operator has the largest effect, with max pooling producing significantly better results [71]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.35, + 0.251, + 0.365 + ], + "angle": 0, + "content": "6.1. Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.473, + 0.72 + ], + "angle": 0, + "content": "Visual grounding measures how well a particular representation or network can localize objects within an image. It is thought that representations with better visual grounding will excel at more general purpose tasks. In section 4, we found that pooling operator had a larger effect on performance than the addition of MAE. We qualitatively study the effect of both pooling and MAE on the visual grounding of the CLIP image and text encoders. We extract zero-shot segmentation masks and GradCAM [75] visualizations for the prompt \"a photo of a dog\", shown in Figure 2. Details are provided in the supplementary material. We observe that, with respect to visual grounding, the pooling operator plays a much more significant role than self-supervision [71]. In particular, \\(\\mathrm{CLIP}_{\\mathrm{MAX}}\\) correctly segments the dog in the picture while CLIP focuses only in the background. However, both for CLIP and \\(\\mathrm{CLIP}_{\\mathrm{MAX}}\\) we observe only an incremental improvement of the localization of the dog in the image when self-supervision is added. Unfortunately, even though the GAP variants of MAE-CLIP and CLIP achieve the best performance on VTAB, they do not also exhibit the most visual grounding. We hope that a more thorough future analysis can elucidate the relationship between self supervision and visual grounding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.245, + 0.743 + ], + "angle": 0, + "content": "6.2. Dataset Diversity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Alternatively, it may be that self-supervision and natural language supervision excel for entirely different parts of the dataset diversity-size spectrum. Historically, the strongest self-supervised visual-encoder baselines (when compared to supervised methods) include methods such as: SimCLR [11], SimCLRV2 [12] and masked auto encoders [36], which all excel on ImageNet—a somewhat diverse but a relatively \"clean\" and object-centric dataset with 1000 disjoint classes. On less diverse datasets such as Cifar-10 [49], specialized self-supervised methods such as ReMixMatch [5]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.501 + ], + "angle": 0, + "content": "have shown excellent performance in the extremely low data regime. Outside of vision, MAEs have shown near state-of-the-art performance on datasets such as Audioset [30,40], a clean 2M example, 632 class audio/video dataset. Recent works which perform controlled experiments scaling self supervised methods to massive datasets [62] show that self-supervised methods achieve numbers roughly \\(10\\%\\) worse (as measured by ImageNet linear probe) as compared to natural language supervised methods at similar data scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.622 + ], + "angle": 0, + "content": "On the other hand, natural language supervised models seem to only show competitive performance on massive and diverse datasets, CLIP [69] was trained on 400M examples, ALIGN [43] was trained on 1.8B examples and BASIC [66] was trained on 6.6B examples. Recently, [28] performs controlled natural language supervision at a smaller scale, showing that natural language supervised models are \\(5 - 10\\%\\) worse (on ImageNet) than their supervised counterparts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.698 + ], + "angle": 0, + "content": "While all of these past results only pose circumstantial and not rigorous experimental evidence, we believe it poses a fruitful line of future work on studying the scaling trends of self supervised methods and contrasting them with other forms of supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.68, + 0.726 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We are grateful to many colleagues for helpful advice and guidance throughout this project. We especially thank Jonathan Shlens, Brandon McKinzie, and Jason Ramuparam for feedback on paper drafts, Ruoming Pang for useful comments and suggestions early in the project, Tom Nickson for his training infrastructure support and Albin Madappally Jose for his assistance in building the training dataset. We also extend thanks to Dr. Nicola Strisciuglio for his clarifying input on earlier incarnations of this work, and Dr. Samuel Albanie for several engaging discussions and helpful pointers to related literature." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "23439" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.115, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a Visual Language Model for Few-Shot Learning. Technical Report arXiv:2204.14198, arXiv, Apr. 2022. arXiv:2204.14198 [cs] type: article. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.269, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[2] Tarik Arici, Mehmet Saygin Seyfioglu, Tal Neiman, Yi Xu, Son Train, Trishul Chilimbi, Belinda Zeng, and Ismail Tutar. MLIM: Vision-and-Language Model Pre-training with Masked Language and Image Modeling. arXiv:2109.12178 [cs], Sept. 2021. arXiv: 2109.12178. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.339, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT Pre-Training of Image Transformers. arXiv:2106.08254 [cs], June 2021. arXiv: 2106.08254. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.381, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Wenhui Wang, Li Dong, and Furu Wei. VL-BEiT: Generative Vision-Language Pretraining. Technical Report arXiv:2206.01127, arXiv, June 2022. arXiv:2206.01127 [cs] type: article. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.438, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[5] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. CoRR, abs/1911.09785, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.508, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. DINO: Emerging Properties in Self-Supervised Vision Transformers. arXiv:2104.14294 [cs], May 2021. arXiv: 2104.14294. 2, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.579, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts. Technical Report arXiv:2102.08981, arXiv, Mar. 2021. arXiv:2102.08981 [cs] type: article. 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[8] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. iGPT: Generative Pretraining From Pixels. In Proceedings of the 37th International Conference on Machine Learning, pages 1691-1703. PMLR, Nov. 2020. ISSN: 2640-3498. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.775, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. SimCLR: A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597-1607. PMLR, Nov. 2020. ISSN: 2640-3498. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. CoRR, abs/2002.05709, 2020. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[12] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. SimCLRv2: Big Self-Supervised Models are Strong Semi-Supervised Learners. arXiv:2006.10029 [cs, stat], Oct. 2020. arXiv: 2006.10029. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.162, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[13] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. MoCov2: Improved Baselines with Momentum Contrastive Learning. arXiv:2003.04297 [cs], Mar. 2020. arXiv: 2003.04297. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.218, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[14] Xinlei Chen and Kaiming He. SimSiam: Exploring Simple Siamese Representation Learning. arXiv:2011.10566 [cs], Nov. 2020. arXiv: 2011.10566. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.26, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[15] Xinlei Chen, Saining Xie, and Kaiming He. MoCov3: An Empirical Study of Training Self-Supervised Vision Transformers. arXiv:2104.02057 [cs], Aug. 2021. arXiv: 2104.02057. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.316, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[16] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: UNiversal Image-TExt Representation Learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, Lecture Notes in Computer Science, pages 104-120, Cham, Sept. 2020. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.414, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[17] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. Technical Report arXiv:2003.10555, arXiv, Mar. 2020. arXiv:2003.10555 [cs] type: article. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.484, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.54, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, June 2009. ISSN: 1063-6919. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[20] Karan Desai and Justin Johnson. VirTex: Learning Visual Representations from Textual Annotations. arXiv:2006.06666 [cs], Sept. 2021. arXiv:2006.06666. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 [cs], May 2019. arXiv: 1810.04805. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.72, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. ViT: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv:2010.11929 [cs], June 2021. arXiv: 2010.11929.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[23] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuo-hang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. METER: An Empirical Study of Training End-to-End Vision-and-Language Transformers. arXiv:2111.02387 [cs], Nov. 2021. arXiv: 2111.02387. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "23440" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[24] Jiali Duan, Liquin Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal Alignment using Representation Codebook. arXiv:2203.00048 [cs], Mar. 2022. arXiv: 2203.00048. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.151, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[25] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herve Jegou, and Edouard Grave. SplitMask: Are Large-scale Datasets Necessary for Self-Supervised Pretraining? arXiv:2112.10740 [cs], Dec. 2021. arXiv: 2112.10740. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.223, + 0.472, + 0.277 + ], + "angle": 0, + "content": "[26] Patrick Esser, Robin Rombach, and Björn Ommer. VQGAN: Taming Transformers for High-Resolution Image Synthesis. arXiv:2012.09841 [cs], June 2021. arXiv: 2012.09841.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.281, + 0.472, + 0.337 + ], + "angle": 0, + "content": "[27] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2):303-338, June 2010. 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.472, + 0.396 + ], + "angle": 0, + "content": "[28] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.399, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[29] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.472, + 0.539 + ], + "angle": 0, + "content": "[30] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.543, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[31] Xinyang Geng, Hao Liu, Lisa Lee, Dale Schuurmans, Sergey Levine, and Pieter Abbeel. M3AE: Multimodal Masked Autoencoders Learn Transferable Representations. Technical Report arXiv:2205.14204, arXiv, May 2022. arXiv:2205.14204 [cs] type: article. 1, 3, 4, 5, 7, 14, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[32] Aaron Gokaslan and Vanya Cohen. Openwebtext corpus. http://Skylion007.github.io/OpenWebTextCorpus, 2019.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.66, + 0.472, + 0.728 + ], + "angle": 0, + "content": "[33] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. arXiv:1706.02677 [cs], Apr. 2018. arXiv: 1706.02677. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.732, + 0.472, + 0.788 + ], + "angle": 0, + "content": "[34] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. VQA2: Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. pages 6904-6913, 2017. 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.79, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[35] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. BYOL: Bootstrap your own latent: A new approach to self-supervised Learning. arXiv:2006.07733 [cs, stat], Sept. 2020. arXiv: 2006.07733. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[36] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. MAE: Masked Autoencoders Are Scalable Vision Learners. arXiv:2111.06377 [cs], Nov. 2021. arXiv: 2111.06377. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.193, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. MoCov1: Momentum Contrast for Unsupervised Visual Representation Learning. pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.25, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[39] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9068-9077, June 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[40] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[41] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing Out of the Box: End-to-End Pre-Training for Vision-Language Representation Learning. pages 12976-12985, Apr. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.449, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[42] Drew A. Hudson and Christopher D. Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. pages 6700-6709, 2019. 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.505, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[43] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. CoRR, abs/2102.05918, 2021. 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.577, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[44] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. ALIGN: Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. arXiv:2102.05918 [cs], June 2021. arXiv: 2102.05918.3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.661, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[45] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1988–1997, 2017. 5, 7, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[46] Kaggle and EyePacs. Kaggle diabetic retinopathy detection. https://www.kaggle.com/c/diabetic-retinopathy-detection/data, July 2015. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[47] Wonjae Kim, Bokyung Son, and Ildoo Kim. ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision. arXiv:2102.03334 [cs, stat], June 2021. arXiv: 2102.03334. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[48] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. Technical Report arXiv:1412.6980, arXiv, Jan. 2017. arXiv:1412.6980 [cs] type: article. 15" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.518, + 0.938 + ], + "angle": 0, + "content": "23441" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "[49] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.122, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[50] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. arXiv:1910.13461 [cs, stat], Oct. 2019. arXiv: 1910.13461.2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.221, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[51] Chenliang Li, Ming Yan, Haiyang Xu, Fuli Luo, Wei Wang, Bin Bi, and Songfang Huang. SemVLP: Vision-Language Pre-training by Aligning Semantics at Multiple Levels. arXiv:2103.07829 [cs], Mar. 2021. arXiv: 2103.07829. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[52] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. arXiv:2107.07651 [cs], Oct. 2021. arXiv: 2107.07651. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[53] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO-2: End-to-End Unified Vision-Language Grounded Learning. arXiv:2203.09067 [cs], Mar. 2022. arXiv: 2203.09067. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.405, + 0.471, + 0.516 + ], + "angle": 0, + "content": "[54] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, Lecture Notes in Computer Science, pages 121–137, Cham, July 2020. Springer International Publishing. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[55] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pretraining via masking, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.47, + 0.629 + ], + "angle": 0, + "content": "[56] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. CoRR, abs/2110.05208, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.632, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[57] Feng Liang, Yangguang Li, and Diana Marculescu. SupMAE: Supervised Masked Autoencoders Are Efficient Vision Learners. Technical Report arXiv:2205.14540, arXiv, May 2022. arXiv:2205.14540 [cs] type: article. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.689, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[58] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 16, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.76, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[59] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic Gradient Descent with Warm Restarts. arXiv:1608.03983 [cs, math], May 2017. arXiv: 1608.03983. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.803, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[60] Ilya Loshchilov and Frank Hutter. AdamW: Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs, math], Jan. 2019. arXiv: 1711.05101. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[61] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. arXiv:1908.02265 [cs], Aug. 2019. arXiv: 1908.02265. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[62] Shlok Mishra, Joshua Robinson, Huiwen Chang, David Jacobs, Aaron Sarna, Aaron Maschinot, and Dilip Krishnan. A simple, efficient and scalable contrastive masked autoencoder for learning visual representations, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[63] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. SLIP: Self-supervision meets Language-Image Pre-training. arXiv:2112.12750 [cs], Dec. 2021. arXiv: 2112.12750.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[64] Norman Mu, Alexander Kirillov, David A. Wagner, and Saining Xie. SLIP: self-supervision meets language-image pre-training. CoRR, abs/2112.12750, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748 [cs, stat], Jan. 2019. arXiv: 1807.03748. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[66] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. CoRR, abs/2111.10050, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.359, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[67] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting Region-to-Phrase Correspondences for Richer Image-to-Sentence Models. pages 2641–2649, 2015. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.428, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. CLIP: Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020 [cs], Feb. 2021. arXiv: 2103.00020. 1, 2, 3, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.525, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[70] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. GPT-2: Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[71] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander T Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.72, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[72] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating Diverse High-Fidelity Images with VQ-VAE-2. Technical Report arXiv:1906.00446, arXiv, June 2019. arXiv:1906.00446 [cs, stat] type: article. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.777, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[73] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenna Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[74] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "23442" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.147 + ], + "angle": 0, + "content": "Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION400M: Open Dataset of CLIP-Filtered 400 Million ImageText Pairs. arXiv:2111.02114 [cs], Nov. 2021. arXiv:2111.02114.6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.15, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[75] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision, 128(2):336-359, Feb. 2020. arXiv:1610.02391 [cs]. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.235, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[76] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.292, + 0.47, + 0.388 + ], + "angle": 0, + "content": "[77] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. GCC: Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.39, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[78] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. CoRR, abs/2112.04482, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.447, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[79] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A Foundational Language And Vision Alignment Model. arXiv:2112.04482 [cs], Feb. 2022. arXiv: 2112.04482. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.518, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[80] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross B. Girshick, Piotr Dólar, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. CoRR, abs/2201.08371, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.602, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[81] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of Generic Visual-Linguistic Representations. arXiv:1908.08530 [cs], Feb. 2020. arXiv: 1908.08530. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.66, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[82] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.717, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[83] Yihong Sun, Adam Kortylewski, and Alan Yuille. Amodal Segmentation through Out-of-Task and Out-of-Distribution Generalization with a Bayesian Model. Technical Report arXiv:2010.13175, arXiv, July 2022. arXiv:2010.13175 [cs] type: article. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.788, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[84] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. arXiv:1908.07490 [cs], Dec. 2019. arXiv: 1908.07490. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[85] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine learning research, 9(11), 2008. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[86] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, \\(\\backslash\\) Lukasz Kaiser," + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[87] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning - ICML '08, pages 1096-1103, Helsinki, Finland, 2008. ACM Press. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[88] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for all vision and vision-language tasks, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[89] Xiaolong Wang and Abhinav Gupta. Unsupervised Learning of Visual Representations using Videos. Technical Report arXiv:1505.00687, arXiv, Oct. 2015. arXiv:1505.00687 [cs] type: article. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[90] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple Visual Language Model Pretraining with Weak Supervision. arXiv:2108.10904 [cs], Aug. 2021. arXiv: 2108.10904. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.406, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[91] Longhui Wei, Lingxi Xie, Wengang Zhou, Houqiang Li, and Qi Tian. MVP: Multimodality-guided Visual Pretraining. arXiv:2203.05175 [cs], Mar. 2022. arXiv: 2203.05175. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[92] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[93] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dólar, and Ross Girshick. Early Convolutions Help Transformers See Better. Technical Report arXiv:2106.14881, arXiv, Oct. 2021. arXiv:2106.14881 [cs] type: article. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.604, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[94] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. Pre-layernorm: On Layer Normalization in the Transformer Architecture. In Proceedings of the 37th International Conference on Machine Learning, pages 10524-10533. PMLR, Nov. 2020. ISSN: 2640-3498. 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.703, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[95] Jianwei Yang, Chunyuan Li, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Focal modulation networks, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[96] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained Interactive Language-Image Pre-Training. arXiv:2111.07783 [cs], Nov. 2021. arXiv: 2111.07783. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[97] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. ERNIE-ViL: Knowledge Enhanced Vision-Language Representations Through Scene Graph. June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[98] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23443" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.093, + 0.468, + 0.147 + ], + "angle": 0, + "content": "[99] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive Captioners are Image-Text Foundation Models. arXiv:2205.01917 [cs], May 2022. arXiv: 2205.01917. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.468, + 0.258 + ], + "angle": 0, + "content": "[100] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A New Foundation Model for Computer Vision. arXiv:2111.11432 [cs], Nov. 2021. arXiv: 2111.11432. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.262, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[101] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-level Representation Learning for Self-supervised Vision Transformers. Technical Report arXiv:2206.07990, arXiv, June 2022. arXiv:2206.07990 [cs] type: article. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.332, + 0.468, + 0.385 + ], + "angle": 0, + "content": "[102] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. arXiv:2103.03230 [cs, q-bio], June 2021. arXiv: 2103.03230. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.389, + 0.468, + 0.441 + ], + "angle": 0, + "content": "[103] Yan Zeng, Xinsong Zhang, and Hang Li. X:VLM: MultiGrained Vision Language Pre-Training: Aligning Texts with Visual Concepts. arXiv:2111.08276 [cs], Feb. 2022. arXiv: 2111.08276. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.446, + 0.468, + 0.511 + ], + "angle": 0, + "content": "[104] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12104-12113, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.516, + 0.468, + 0.61 + ], + "angle": 0, + "content": "[105] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, André Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, Lucas Beyer, Olivier Bachem, Michael Tschannen, Marcin Michalski, Olivier Bousquet, Sylvain Gelly, and Neil Houlsby. The visual task adaptation benchmark. CoRR, abs/1910.04867, 2019. 1, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.614, + 0.468, + 0.736 + ], + "angle": 0, + "content": "[106] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djo-longa, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, Lucas Beyer, Olivier Bachem, Michael Tschannen, Marcin Michalski, Olivier Bousquet, Sylvain Gelly, and Neil Houlsby. VTAB: A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark. arXiv:1910.04867 [cs, stat], Feb. 2020. arXiv: 1910.04867.5,7,14" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.74, + 0.468, + 0.806 + ], + "angle": 0, + "content": "[107] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. ADE20K: Semantic Understanding of Scenes through the ADE20K Dataset. arXiv:1608.05442 [cs], Oct. 2018. arXiv: 1608.05442. 19" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.093, + 0.468, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23444" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_origin.pdf b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..99a8784f0fa87ff2d35b6adef6d1ef2c749a12f5 --- /dev/null +++ b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/3ac61577-bf5b-4f86-80cd-9aa136bbba7b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b835cb8cb63c9a22b6dac27ca1051c2ef2f861107e18d0b8836bd6e85e1ba3a9 +size 4420124 diff --git a/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/full.md b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b46cae7fe54491f311a31e4e8bd3157e29b1f342 --- /dev/null +++ b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/full.md @@ -0,0 +1,371 @@ +# Masked Autoencoding Does Not Help Natural Language Supervision at Scale + +Floris Weers Vaishaal Shankar Angelos Katharopoulos Yinfei Yang Tom Gunter + +{fweers,vaishaal_shankar,a_katharopoulos,yinfeiy,tom_gunter}@apple.com + +# Abstract + +Self supervision and natural language supervision have emerged as two exciting ways to train general purpose image encoders which excel at a variety of downstream tasks. Recent works such as M3AE [31] and SLIP [63] have suggested that these approaches can be effectively combined, but most notably their results use small (<20M examples) pre-training datasets and don't effectively reflect the large-scale regime (>100M samples) that is commonly used for these approaches. Here we investigate whether a similar approach can be effective when trained with a much larger amount of data. We find that a combination of two state of the art approaches: masked auto-encoders, MAE [37] and contrastive language image pre-training, CLIP [68] provides a benefit over CLIP when trained on a corpus of 11.3M image-text pairs, but little to no benefit (as evaluated on a suite of common vision tasks) over CLIP when trained on a large corpus of 1.4B images. Our work provides some much needed clarity into the effectiveness (or lack thereof) of self supervision for large-scale image-text training. + +# 1. Introduction + +Large scale pretraining has become a powerful tool in the arsenal of computer vision researchers to produce state of the art results across a wider variety of tasks [39,88,95,98]. However, when pre-training on tens of millions to billions of images it is difficult to rely on standard supervised methods to train models, as datasets of this size often lack reliable labels. In the presence of these massive but largely under-curated datasets, two general classes of methods to train general purpose image encoders have emerged: + +1. Self Supervised techniques that learn visual representations from the image data alone [11,36] +2. Natural Language Supervised methods that utilize paired free-form text data to learn visual representations [43, 69] + +Due to the unique strengths and weaknesses of each ap- + +proach $^1$ , a recent flurry of work has introduced methods that combine both forms of supervision [31, 56, 64, 78] to varying degrees of success. While each of these methods establishes some regime where the additional supervision helps, none of these "joint-supervision" methods advance state of the art in any meaningful way. Additionally, to our knowledge none of these methods have shown comparative results at the scale many large scale vision models are currently trained at ( $>100\mathrm{M}$ examples) [43, 66, 69, 73, 80, 82, 98]. Furthermore, methods that use both forms of supervision start with the presumption that the additional supervision is helpful and either often lack clean ablations or lack evaluations in a "high accuracy" regime—leading to further confusion regarding whether a combination of these methods can actually improve the state of the art. To clarify this issue, in this work, we investigate a simple question: + +Does a combination of self supervision and natural language supervision actually lead to higher quality visual representations? + +In order to answer this, we first introduce a straightforward baseline approach that combines standard self-supervision and language supervision techniques. We combine masked auto-encoders (MAE) and contrastive language image-pretraining (CLIP) to make MAE-CLIP. We then present a careful study of the performance of MAE, M3AE, CLIP and MAE-CLIP across a wide variety of tasks in two distinct regimes: a "low-sample" $^{2}$ 11.3 million example regime and a "high-sample" 1.4 billion example regime. We train self-supervised and language-supervised methods using the same pre-training datasets under the assumption that we have no knowledge about downstream tasks. Our experiments show: + +1. In the low sample size regime, without changing the final pooling operation in the network, we observe a large performance improvement, namely $6\%$ on ImageNet [18] and $4\%$ on VTAB [105]. However, when + +Self supervised methods can learn representations without labels, but natural language supervision learns better representations. Natural language supervised methods rely on quality of captions +2We note that what low sample means has changed substantially over the last few years + +we modify the pooling operation, the improvement substantially decreases to around $1\%$ on both ImageNet and VTAB. + +2. In the high sample size regime, there is virtually no difference in performance between MAE-CLIP and CLIP across ImageNet, VTAB, and VQA tasks. + +We believe our work is the first careful study of this form and contextualizes recent progress in both self-supervision and natural language supervision. + +The rest of the paper is organized as follows: In Section 2, we cover related work in the areas of self supervision and natural language supervision. In Section 3, we give an overview of the baseline methods we study, MAE, M3AE, CLIP and our new baseline MAE-CLIP. Then we present and analyse our small scale and large scale experimental findings in Sections 4 and 5. Finally, we discuss potential explanations for our findings and some future work in 6. + +# 2. Related Work + +Our work combines natural language supervision and self-supervision in a multi-task approach to visual encoding, and so research from these three areas is relevant. + +Natural language supervision for visual encoding covers a variety of approaches that assume access to datasets of images or videos associated with text. Some of the most successful use a contrastive pairwise alignment signal applied to very large batch and dataset sizes [68]. This large batch size means hard negative pairs can be produced by random sampling, avoiding the need for a memory bank [12, 13, 38] or momentum distillation [15]. FILIP [96] further improves zero-shot performance for some tasks by using a more fine-grained elementwise contrastive loss. Image captioning has also shown promise as a pre-training task [20, 90], with [20] in particular demonstrating strong data efficiency. UNITER [16] pursues a similar idea, applying a generative loss to both image and text modalities. [2, 16, 23, 41, 47, 51, 61, 81, 84, 97] apply the masked patch prediction problem from [21, 37, 50] to a joint image-text data space. These approaches either use a pre-trained convolutional neural network (CNN) to generate region of interest (RoI) proposal encodings, or predict the labelled class of the masked patch instead of the raw pixel values (or the quantized patch ID). UNIMO-2 [53] attempts to ground image and text patches for a single data example into a top-k quantized set of embeddings, before passing the grounded embeddings along with the raw non-grounded embeddings into a decoder, targeting a loss based on masked language modeling (MLM) and image-text-matching (ITM). + +Self-supervised approaches for visual encoding may be loosely categorized into those that target consistency constraints between multiple views of the same scene and those + +that attempt visual reconstruction on corrupted or conditioning representations of images. Consistency constraints are often derived through data-augmentation applied to a single real image [6, 10, 14, 35, 102], but these may also take the form of e.g. a temporal constraint if video data is available [89], or be applied at an image-patch level by making a smoothness assumption and selecting neighboring patches as 'consistent' whilst treating distant patches (or those sourced from another image) as negative examples [101]. Forms of denoising autoencoder (DAE) [87] are popular as a means of self-supervision, and have been investigated at a variety of scales; iGPT [9] learns to generate images as a row-major flattened sequence of pixels, whilst [65], [72], and [26] compress images into a short sequence of discrete codes before then regenerating and scoring at a pixel level. BEiT [3] first encodes an image into a sequence of discrete codes before masking and then predicting a subset of the codes. MAE [37] drops the discretization, instead predicting raw pixel loss for a subset of the encoded patches in a manner strongly reminiscent of BART [50]. SplitMask [25] applies both a masked image patch prediction loss and a pairwise contrastive loss, using the masked example as well as the inverted masked example to form each positive pair. Meanwhile, M3AE [24] and VL-BEiT [4] both propose a masked patch prediction problem applied to image and text modalities jointly. Recent concurrent work shows that while masking does not help generalization, the masking can be used to speed up training by dropping the masked tokens during the forward and backward pass [55]. EVA [29] shows that doing masked image modeling to predict the output embeddings (rather than pixels) of the masked image can be an effective pretraining task when combined with natural language supervision. We note for EVA the masked image modeling occurs only on smaller well curated datasets such as ImageNet, Conceptual Captions and COCO. + +Multi-task methods for pre-training visual encoders are a highly active area of research. CoCA [99] combines the cross-modality contrastive task from [68] with image captioning in the style of [20], and by pre-training at very large scale show that the resulting model is more performant than prior art across a very broad array of downstream visual understanding tasks. SLIP [64] combines a SimCLR style self supervision loss with the standard CLIP contrastive loss to train a jointy self-supervised and natural language supervised model, however while they show large performance gains they are all in a low accuracy regime (below $40\%$ Top-1 accuracy on ImageNet). Florence [100] employs a diverse array of tasks and datasets at pre-training time (including object detection, CLIP-style contrastive modeling, Visual Question Answering (VQA), and supervised classification). ALBEF [52] builds on a CLIP-style architecture by incorporating a single-stream encoder that consumes both + +![](images/b8d0073fe48fb99b79de939d95b11dfd5570288c17649fbc5ff4b69761c64648.jpg) +Figure 1. MAE-CLIP: CLIP [68] augmented with a generative decoder in the style of MAE [31]. + +modalities. It uses a masked language modeling loss for the text modality and an image-text matching loss for within-batch hard-negatives, mined according to contrastive similarity. FLAVA [79] also uses masked-image modeling, and uses both single and joint modality encoders, with additional single-modality datasets (i.e. not just image-text pairs) that allow it to generalize to longer text inputs and visual data which is unlikely to be captioned. X-VLM [103] utilizes image data which includes object-labelled bounding boxes to extend the image-text matching and contrastive losses to cover within-image regions as well as whole-image captions. They also include a masked language modelling objective, achieving impressive performance on many zero-shot downstream tasks—although the requirement for object bounding box data may not scale well for large pre-training datasets. SupMAE [57] adds a supervised labeling task to the MAE architecture, demonstrating that this results in better data efficiency. MVP [91], meanwhile, finetunes a pre-trained CLIP backbone using MIM, and demonstrates that starting with a pre-trained model improves downstream visual recognition performance. MAE-CLIP is reminiscent of M3AE (with the addition of a contrastive task), ALBEF (but applying the masked prediction task to both modalities, dropping the momentum distillation), and most of all FLAVA (without the single-modality masked-patch-prediction tasks, and with the use of high ratio random masking following MAE and M3AE). + +# 3. Background + +# 3.1. Contrastive language-image pre-training + +CLIP [68] and ALIGN [44] both demonstrated that contrastive image-language supervision applied at scale was capable of producing an image encoder that excels at a range of downstream tasks—often in a zero-shot setting. In our implementation, we begin by following the design outlined in [68]. Specifically, we use per-modality encoders to produce a single $\ell 2$ -normalized dense embedding vector for each image or text input, before applying a pairwise InfoNCE [65] loss to a large global batch of paired image-text embeddings, using all non-paired examples as hard nega + +tives. We define the image-to-text loss $L_{i2t}$ and the text-to-image loss $L_{t2i}$ as: + +$$ +L _ {i 2 t} = - \frac {1}{N} \sum_ {j} ^ {N} \log \frac {\exp \left(x _ {j} ^ {T} y _ {j} / \sigma\right)}{\sum_ {k = 1} ^ {N} \exp \left(x _ {j} ^ {T} y _ {k} / \sigma\right)} \tag {1} +$$ + +$$ +L _ {t 2 i} = - \frac {1}{N} \sum_ {j} ^ {N} \log \frac {\exp \left(y _ {j} ^ {T} x _ {j} / \sigma\right)}{\sum_ {k = 1} ^ {N} \exp \left(y _ {j} ^ {T} x _ {k} / \sigma\right)}, \tag {2} +$$ + +where $N$ is the global batch size, $x_{j}$ is the normalized embedding of the image for the $j$ -th pair and $y_{k}$ is the normalized embedding of the text in the $k$ -th pair in the batch. $\sigma$ is a learnable temperature parameter. The total contrastive loss $L_{c} = \frac{1}{2} (L_{t2i} + L_{i2t})$ is the average of these two losses. + +We use transformer encoders for both modalities, however, differently to [68] we explore several strategies for aggregating the image-patch or text-token outputs into a single embedding (see Section 4.4). We also eschew the autoregressive masking strategy for the text encoder, instead allowing full bi-directional self attention following [44]. + +# 3.1.1 Pooling + +In CLIP [68] the authors use a separate 'CLS' token projected through the network as the overall image representation. On the other hand, ALIGN [44] use global-average pooling over the encoded visual features. Whilst both approaches produced good downstream evaluation results, recent work [8, 71, 104], suggested that the choice of pooling strategy can strongly influence the quality of visual semantic embeddings. In particular, [71] demonstrated that this effect is present for CLIP-like visual encoders. Noting this, we opt to investigate three pooling strategies: 1. the default multihead-attention pooling (MAP), 2. global average pooling (GAP)—both described in [104]—as well as 3. non-maximal suppression pooling (MAX) as in [8, 71]. + +# 3.2. Masked Autoencoders + +In MAE [37] the authors demonstrate a simple technique for self-supervised image-encoder pre-training that—to our knowledge—is still considered state-of-the-art. They use a + +ViT [22] encoder-decoder architecture and apply it to heavily masked input images. The input to the encoder consists of the visible, unmasked image patches, first embedded via a linear projection before additive positional embeddings are applied, and then the result is fed through the encoder's transformer layers. They demonstrate that a very high patch masking ratio is critical to achieving good performance, and usually retain only $25\%$ of the image. The decoder consumes the output of the encoder, as well as a learned 'masked-patch' embedding, which is included to represent each masked token. Positional embeddings are also reapplied, to ensure that spatial information is passed through the decoder for the masked patches. The output of the decoder at the masked positions is then measured against ground truth using a simple mean-squared-error loss. As in [37] we experiment with predicting both normalized and un-normalized patch values, finding that predicting the normalized patch value slightly improves the performance of our MAE implementation. + +Masking across modalities: BART [50] applies a similar strategy to text data, but uses a cross-entropy loss over the masked-token output distribution and a far lower masking ratio (typically around $15\%$ ). M3AE [31] extends MAE and BART to incorporate inputs from both text and image modalities, using per-modality input and output projections. Otherwise the encoder-decoder architecture resembles MAE, with the addition of learned modality-indicating embeddings to each transformer input. + +The M3AE loss relies only on the contents of the image-text pairs, and comprises mean-squared error for the masked image patches $(L_{gen.i})$ and cross-entropy over the vocabulary for the masked text tokens $(L_{gen.t})$ : + +$$ +L _ {\text {g e n . i}} = \frac {1}{N} \sum_ {j} ^ {N} \left(p _ {j} - P _ {j}\right) ^ {2} \tag {3} +$$ + +$$ +L _ {\text {g e n . t}} = - \frac {1}{N} \sum_ {j} ^ {N} \log \frac {\exp \left(t _ {n , T _ {n}}\right)}{\sum_ {c = 1} ^ {C} \exp \left(t _ {n , c}\right)}, \tag {4} +$$ + +where $p_j$ and $P_j$ refer to the predicted and ground truth pixel value respectively and $t_n$ and $T_n$ refer to the predicted token distribution and the ground truth token, with $C$ as the number of unique tokens. For the image reconstruction loss, we normalize the ground-truth per-patch, following [37]. + +# 3.3. MAE-CLIP + +As the name suggests, MAE-CLIP attempts to incorporate aspects of MAE/M3AE into CLIP through the addition of a single dual-modality transformer decoder (as in M3AE) which consumes the output of the CLIP encoders. We show this in Figure 1. + +# 3.3.1 Components + +The model architecture is similar to ALBEF [52] and FLAVA [78], consisting of three components. + +An image encoder: Following ViT [22], we divide the input image into equally-sized, non-overlapping patches. After applying a linear projection and adding a 2-D position encoding, we feed the per-patch representations through a number of transformer layers [86]. + +A text encoder: Following ALIGN [43], our text encoder is based on BERT [21], in which text is first tokenized and embedded, and a 1-D trainable position encoding is added. Differently to BERT, we use pre-layernorm [94] and initialize the parameters according to the simplified scheme outlined in [70]. + +A cross-modality decoder: The decoder receives per-element encoded image and text representations from the encoders for both masked and un-masked elements. For the image modality, masked patches are added by replacing their values with a shared trainable mask token. Positional encodings are once again added to all elements, following [37]. We also add a per-modality trainable encoding to allow the decoder to easily distinguish between the two modalities. The decoder uses the same transformer implementation as the encoders. The output of the final decoder layer is then projected into a per-patch-modality output space and the overall loss is computed. + +# 3.3.2 Losses + +MAE-CLIP employs the losses of both CLIP (Eq. 1, 2) and M3AE (Eq. 3, 4). Our final loss is a weighted sum of the losses from each task, as follows: + +$$ +L = \frac {1}{2} \left(L _ {t 2 i} + L _ {i 2 t}\right) + w _ {i} \cdot L _ {\text {g e n . i}} + w _ {t} \cdot L _ {\text {g e n . t}} \tag {5} +$$ + +where $w_{i}$ and $w_{t}$ are scalars used to control the relative weight of the generative losses. We always provide paired image-text inputs to the model, leaving it to future work to explore the benefits of also incorporating single modality inputs and reconstruction losses, at scale. + +To avoid severely impacting the CLIP loss via computing it with masked inputs, we run the encoders twice: once with full unmasked input in order to compute the CLIP loss, before then making a second pass using only the unmasked input to compute the M3AE loss. Finally, we compute the weighted sum of the two losses, and use this to calculate the overall update. + +# 4. Experiments + +In this section, we present our experimental results to study how the addition of self-supervision affects the representation quality of natural language supervised models. We study standard masked auto-encoders (MAE), multimoddal masked auto-encoders (M3AE), contrastive language image pretraining (CLIP) and our newly introduced baseline method (MAE-CLIP), which combines mask based self-supervision with contrastive language-image learning. + +# 4.1. Experimental Setup + +Architecture: Experiments are performed using an image encoder based on the architecture described as ViT-B-16 in [93], in combination with a 12-layer, 512-wide text encoder with 8 attention heads. For the decoder, we use another 12-layer, 512-wide transformer with 8 attention heads, identical to the text encoder. We also use a byte-pair encoding tokenizer, fit to the OpenWebText corpus [32]. When training MAE-CLIP, we replace the random masking by a similarity masking strategy, which makes use of the element-wise CLIP similarity scores to select the masked element for the reconstruction task. + +Pretraining Datasets: Our analysis is divided into two sections. First, in sections 4.2 to 4.4, we present our study on the "low-sample" regime using the combination of CC12M [7] and CC3M [77] excluding all images whose text contains a [PERSON] tag5. The final dataset contains 11.3 million comparatively high quality image-text pairs. We refer to this dataset as CC or "small" scale. + +Pretraining configuration: We use a minibatch size of 16,384 samples and train using the AdamW optimizer [60] with decoupled weight decay. We use a base learning rate of $5 \times 10^{-4}$ and train using cosine learning rate decay [59] with an initial linear warmup of 200 steps. We reduce the $\beta_{2}$ decay factor to 0.98, and increase $\epsilon$ to $10^{-6}$ , finding that these changes improved rate of convergence for all models trained. To increase training convergence speed, we switch from a local (per-GPU) to a global contrastive loss after 500 steps. During the global contrastive loss phase, we set the generative text and image loss weights to $w_{t} = 0.05$ and $w_{i} = 0.1$ respectively. For all models which rely on masking, we use a $75\%$ masking ratio, consistent with [31,37], as we did not find an alternative that improved downstream results. In total, we train our models on CC for 32 epochs, which corresponds to $\sim 22,000$ steps. For our M3AE training runs, we use per-modality encoders so as to produce a baseline that is flop and parameter matched to our MAE-CLIP runs. + +
ModelsZero-shotLinear Probing
MAE-33.9
M3AE-52.5
CLIP29.752.6
MAE-CLIP33.858.9
+ +Table 1. ImageNet classification with zero-shot transfer or linear probing after pretraining on the $CC$ dataset (11.3M images). MAE-CLIP significantly improves the classification performance of CLIP in the small scale regime. + +# 4.2. Image Classification + +We evaluate the quality of the representations learned from pre-training on $CC$ by measuring the image classification performance, either by zero-shot transfer, or by training a linear classifier using the predicted visual features. For zero-shot classification, we use the average of 80 prompts and follow the routine prescribed in [69]. To train the linear classifier, we pre-compute features using the visual encoder6 and run AdamW for 20 to 80 epochs with a learning rate of 0.01. + +Table 1 shows the zero-shot and linear probing classification results on the ImageNet [19] dataset while Table 2 shows our linear probing results on most (17) of the VTAB [106] benchmark datasets7. Initially, we observe that the combination of self-supervision and natural language supervision provides a consistent and substantial improvement over either form of supervision by itself. This result concurs with previous works such as [31] and [64], which show that self-supervision aids natural language supervision. + +# 4.3.VQA + +Subsequently, we compare the performance of MAE-CLIP to the rest of the baselines on the visual question answering task using three datasets. CLEVR [45], VQAv2 [34] and GQA [42]. VQA assesses the model's multimodal reasoning capabilities as well as its visual grounding. In order to finetune our models for VQA, we freeze the image and text encoders and either randomly reinitialize the decoder in models such as MAE-CLIP or M3AE or add a new identical decoder for CLIP. During finetuning, we use a layer-wise learning rate decay, following [3, 17, 37]. For CLEVR and VQAv2 we finetune the decoder for 50 epochs, while for GQA we finetune for 5 epochs on the "train-all" split and 2 epochs on the "train-balanced" split following the protocol of [54]. In all cases, we treat the problem as a classification problem of selecting one answer out of the set of possible answers for each dataset. + +In Table 3 we observe that self-supervision combined + +Table 2. Linear probing accuracy (%) on classification tasks. All models are trained on the CC Dataset (11.3M images). (● VTAB/natural, ● VTAB/specialized and ● VTAB/structured.) + +
Caltech101CIFAR-100DTDFlowers102PetsSVHNEuroSATCamelyonResisc45Clevr/CloestClevr/CountDMLabdSprites/OridSprites/LocKITTI/DistsNORB/AzimsNORB/ElevAverage
MAE75.356.258.570.237.471.996.282.884.065.160.642.837.979.936.737.464.562.2
M3AE85.666.669.486.558.469.297.482.990.769.471.248.249.883.342.634.373.469.3
CLIP84.262.857.781.669.752.595.582.686.753.752.944.845.961.345.731.042.661.8
MAE-CLIP89.866.264.886.574.556.795.481.686.953.659.645.944.668.648.036.151.065.3
+ +
ModelCLEVRVQAv2GQA
MAE93.545.746.1
M3AE97.555.850.9
CLIP87.555.650.0
MAE-CLIP96.058.552.2
+ +with natural language supervision (MAE-CLIP) performs consistently better than either MAE or CLIP by themselves. Combining information from both modalities is evidently critical for effectively answering visual questions as the performance of MAE on VQAv2 and GQA is several percentage points lower than the rest of the methods. Interestingly, methods that employ a form of self-supervision perform significantly better on CLEVR. We argue that this discrepancy is due to CLEVR requiring object localization and spatial reasoning which benefits from this form of self-supervision. + +# 4.4. Pooling Analysis + +To better understand whether the gap between MAE-CLIP and CLIP we observed in the previous section is fundamental, we investigate the effect of the pooling operator on different downstream tasks. Table 4 compares the performance of CLIP and MAE-CLIP with the standard multi-head attention pooling (MAP), global average pooling (GAP) and max pooling (MAX) on all image classification tasks and visual question answering. We note that GAP and MAX pooling perform substantially better than the standard MAP pooling across ImageNet, VTAB and VQA tasks. A hypothesis posed by [71] is that the alternative pooling operators lead to better perceptual grouping in the visual representation. We study this with a qualitative analysis in Section 6.1. + +# 5. Experiments at Scale + +At small-scale, we showed that self-supervision combined with natural language supervision marginally im + +Table 3. VQA finetuning results for models pre-trained on CC Dataset (11.3M images). We train a new decoder for all methods while keeping the encoders frozen. MAE-CLIP performs significantly better than either only self-supervised or language supervised methods by themselves. + +
ModelPoolingImageNetVTABVQA
ZSLPLPFT
CLIPMAP29.752.661.864.4
MAE-CLIPMAP33.858.965.368.9
CLIPGAP29.359.870.665.1
MAE-CLIPGAP33.562.571.769.2
CLIPMAX33.162.368.763.5
MAE-CLIPMAX35.263.269.168.5
+ +Table 4. Results of CLIP and MAE-CLIP with different pooling options trained on CC dataset (11.3M images). We compare the models on ImageNet Zero-Shot (ZS), ImageNet Linear-Probing (LP), VTAB 17-task Linear-Probing average, and VQA 3 tasks Fine-Tuning (FT) average. + +proves the quality of the learned visual representations, and that the choice of image-encoder pooling strategy can heavily influence results. We now investigate whether these conclusions still hold when training on a much larger, 1.4B example dataset. + +# 5.1. Experimental Setup + +We follow the configuration described in Section 4.1, with a few key differences, noted below. + +Pretraining Dataset: We combine a 2.2B example web-crawled image-text pair dataset, termed the English-WebImage-Text dataset (EWIT-2.2B), with LAION-400M [74], CC3M [77], CC12M [7] and an internal high-quality image-text pair dataset containing approximately 134M image text pairs which we term the High Quality Image Text Pairs Dataset (HQITP-134M). We globally deduplicate this—keyed by image bytes—reducing the number of image-text pairs significantly, to yield a final 1.4B examples. Details provided in the supplementary material. In the rest of the paper we refer to this as the web-crawled or large-scale dataset. + +Pretraining configuration: We increase the learning rate warm-up to 1,000 steps, and compute a local contrastive loss for the first 10,000 steps, as during early training this improves the models rate of convergence. We train for a total of 480,000 steps, which corresponds to 6 full passes through the dataset. We do not increase the batch + +Table 5. Linear probing accuracy (%) on classification tasks. Models are all trained on our web-crawled dataset (1.4B images). (VTAB/natural, VTAB/specialized and VTAB/structured.) In the large scale pretraining regime, the difference between MAE-CLIP and CLIP is reduced to $< 1\%$ . * At evaluation time, our M3AE model was $50\%$ trained, so that performance may improve further. + +
Caltech101CIFAR-100DTDFlowers102PetsSVHNEuroSATCarnelyonResisc45Clevi/CloestClevi/CountDMLabdSprites/OridSprites/LocKITTI/DistsNORB/AzimsNORB/ElevAverage
M3AE*93.074.878.295.481.269.697.384.892.765.774.751.052.780.747.536.268.873.2
CLIP94.978.480.097.386.959.094.182.392.745.662.146.046.153.350.920.335.866.2
CLIPMAX96.181.080.997.389.965.796.083.294.152.867.849.959.567.641.223.445.870.1
MAE-CLIPMAX95.879.281.596.888.262.195.881.893.052.066.949.653.772.553.032.345.470.6
CLIPGAP95.880.581.697.688.766.097.084.493.356.771.453.358.070.150.638.355.172.9
MAE-CLIPGAP95.479.382.297.488.672.896.684.593.557.573.652.757.571.251.645.655.273.8
+ +
ModelsZero-shotLinear Probing
M3AE*-69.3
CLIPGAP61.875.9
CLIPMAX63.777.5
MAE-CLIPGAP57.475.7
MAE-CLIPMAX60.976.6
+ +Table 6. ImageNet classification after pretraining on web-crawled dataset (1.4B images). In the large scale regime, self-supervision does not complement natural language supervision. + +
ModelCLEVRVQAv2GQA
M3AE96.959.953.3
CLIPGAP87.861.855.0
CLIPMAX89.560.653.6
MAE-CLIPGAP92.861.955.3
MAE-CLIPMAX93.961.553.7
+ +Table 7. VQA finetuning results on web-crawled dataset (1.4B images). Similar to Table 15a, for large pretraining datasets, self-supervision is not improving the downstream VQA performance. + +size, and use the same learning rate schedule as for $CC$ . + +Model variants: We train M3AE as well as both GAP and MAX variants of CLIP and MAE-CLIP, having noted in the small-scale regime that these are the highest performing architectures and pre-training strategies. We do not train MAE at scale due to resource limitations, and because its performance was significantly lower than the models that utilize natural language supervision—a finding that is consistent with prior literature [31]. + +# 5.2. Image Classification + +Table 15a provides encoder performance for both zero-shot and linear-probing on ImageNet [19]. We note that at scale, MAE-CLIP typically shows worse performance than CLIP across both pooling strategies and tasks. We hypothesize that this may be due to the fact that masked-patch-prediction results in model capacity being lost to the task of predicting semantically irrelevant patches, meaning that we lose some of the benefits provided by the larger scale dataset (despite tuning the relative loss weights). If we are correct in this hypothesis then it is possible that a larger model capacity would help to alleviate this issue. We leave this to a future investigation. Meanwhile, in Table 14 we once again explore the encoder performance through linear-probing for VTAB [106]. We note that GAP provides a consistent improvement over MAX across both CLIP and MAE-CLIP models, and that M3AE is also a strong performer—as was + +the case in the small-scale regime. M3AE provides particular benefit to the "structured" VTAB tasks, however underperforms (sometimes by a large margin) on the "natural" tasks. MAE-CLIP ends up in between the two—doing better than CLIP but worse than M3AE on the structured tasks, and worse than CLIP but better than M3AE on the natural tasks. + +# 5.3.VQA + +We follow the same finetuning strategy as before, with results in Table 15b. Notably, we do not see a consistent difference across tasks for CLIP vs MAE-CLIP, with the possible exception of CLEVR [45], where it is clear that M3AE is by far the best performing approach—and its benefits improve the MAE-CLIP results on that task. GAP is perhaps a marginally stronger pooling strategy as measured against MAX, however the difference is smaller than it was in the small-scale regime. + +# 6. Analysis and Discussion + +While our results present a pessimistic view on combining self-supervision with natural language supervision, we don't have a clear reason as to why self-supervision fails at scale. Here we present two hypotheses that may be explored in future work. + +![](images/6a1803115410c8dae5d6e5cc382ef40d248a06721407883a4b035ecb0bd303ff.jpg) +(a) CLIP + +![](images/6e77dd4128ede60c24b9a7b39c084248839989742632b53f6c4c84779bd1e161.jpg) +(b) MAE-CLIP + +![](images/dc7d15102741e24869e262039d947f9be47c83657f0c2e2e6375acf8fc59b7ba.jpg) +(c) $\mathrm{CLIP}_{\mathrm{GAP}}$ +Figure 2. Zero-shot segmentation masks (top row) and GradCAM [75] visualizations (bottom row) for the prompt "a photo of a dog" for CLIP and MAE-CLIP. Self-supervision qualitatively improves the visual grounding of the model, however, the choice of pooling operator has the largest effect, with max pooling producing significantly better results [71]. + +![](images/240f6ae5dfd7ebbdc460c056fe14a4539dd23b1547e82e3e95d74d34f64ab11b.jpg) +(d) MAE-CLIPGAP + +![](images/e5ec95de2935b4a5d1c5dfacb2cc60c1a3231cc7a3b9121a07a911cf4a9e887b.jpg) +(e) $\mathrm{CLIP}_{\mathrm{MAX}}$ + +![](images/aa409b56104c878a95724ad919adaa7268c1203d1b6b0ca1179052a8fe931aa2.jpg) +(f) MAE-CLIPMAX + +# 6.1. Visual Grounding + +Visual grounding measures how well a particular representation or network can localize objects within an image. It is thought that representations with better visual grounding will excel at more general purpose tasks. In section 4, we found that pooling operator had a larger effect on performance than the addition of MAE. We qualitatively study the effect of both pooling and MAE on the visual grounding of the CLIP image and text encoders. We extract zero-shot segmentation masks and GradCAM [75] visualizations for the prompt "a photo of a dog", shown in Figure 2. Details are provided in the supplementary material. We observe that, with respect to visual grounding, the pooling operator plays a much more significant role than self-supervision [71]. In particular, $\mathrm{CLIP}_{\mathrm{MAX}}$ correctly segments the dog in the picture while CLIP focuses only in the background. However, both for CLIP and $\mathrm{CLIP}_{\mathrm{MAX}}$ we observe only an incremental improvement of the localization of the dog in the image when self-supervision is added. Unfortunately, even though the GAP variants of MAE-CLIP and CLIP achieve the best performance on VTAB, they do not also exhibit the most visual grounding. We hope that a more thorough future analysis can elucidate the relationship between self supervision and visual grounding. + +# 6.2. Dataset Diversity + +Alternatively, it may be that self-supervision and natural language supervision excel for entirely different parts of the dataset diversity-size spectrum. Historically, the strongest self-supervised visual-encoder baselines (when compared to supervised methods) include methods such as: SimCLR [11], SimCLRV2 [12] and masked auto encoders [36], which all excel on ImageNet—a somewhat diverse but a relatively "clean" and object-centric dataset with 1000 disjoint classes. On less diverse datasets such as Cifar-10 [49], specialized self-supervised methods such as ReMixMatch [5] + +have shown excellent performance in the extremely low data regime. Outside of vision, MAEs have shown near state-of-the-art performance on datasets such as Audioset [30,40], a clean 2M example, 632 class audio/video dataset. Recent works which perform controlled experiments scaling self supervised methods to massive datasets [62] show that self-supervised methods achieve numbers roughly $10\%$ worse (as measured by ImageNet linear probe) as compared to natural language supervised methods at similar data scale. + +On the other hand, natural language supervised models seem to only show competitive performance on massive and diverse datasets, CLIP [69] was trained on 400M examples, ALIGN [43] was trained on 1.8B examples and BASIC [66] was trained on 6.6B examples. Recently, [28] performs controlled natural language supervision at a smaller scale, showing that natural language supervised models are $5 - 10\%$ worse (on ImageNet) than their supervised counterparts. + +While all of these past results only pose circumstantial and not rigorous experimental evidence, we believe it poses a fruitful line of future work on studying the scaling trends of self supervised methods and contrasting them with other forms of supervision. + +# 7. Acknowledgement + +We are grateful to many colleagues for helpful advice and guidance throughout this project. We especially thank Jonathan Shlens, Brandon McKinzie, and Jason Ramuparam for feedback on paper drafts, Ruoming Pang for useful comments and suggestions early in the project, Tom Nickson for his training infrastructure support and Albin Madappally Jose for his assistance in building the training dataset. We also extend thanks to Dr. Nicola Strisciuglio for his clarifying input on earlier incarnations of this work, and Dr. Samuel Albanie for several engaging discussions and helpful pointers to related literature. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a Visual Language Model for Few-Shot Learning. Technical Report arXiv:2204.14198, arXiv, Apr. 2022. arXiv:2204.14198 [cs] type: article. 4 +[2] Tarik Arici, Mehmet Saygin Seyfioglu, Tal Neiman, Yi Xu, Son Train, Trishul Chilimbi, Belinda Zeng, and Ismail Tutar. MLIM: Vision-and-Language Model Pre-training with Masked Language and Image Modeling. arXiv:2109.12178 [cs], Sept. 2021. arXiv: 2109.12178. 2 +[3] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT Pre-Training of Image Transformers. arXiv:2106.08254 [cs], June 2021. arXiv: 2106.08254. 2, 5 +[4] Hangbo Bao, Wenhui Wang, Li Dong, and Furu Wei. VL-BEiT: Generative Vision-Language Pretraining. Technical Report arXiv:2206.01127, arXiv, June 2022. arXiv:2206.01127 [cs] type: article. 2 +[5] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. CoRR, abs/1911.09785, 2019. 8 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. DINO: Emerging Properties in Self-Supervised Vision Transformers. arXiv:2104.14294 [cs], May 2021. arXiv: 2104.14294. 2, 16 +[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts. Technical Report arXiv:2102.08981, arXiv, Mar. 2021. arXiv:2102.08981 [cs] type: article. 5, 6, 14 +[8] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. iGPT: Generative Pretraining From Pixels. In Proceedings of the 37th International Conference on Machine Learning, pages 1691-1703. PMLR, Nov. 2020. ISSN: 2640-3498. 2 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. SimCLR: A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597-1607. PMLR, Nov. 2020. ISSN: 2640-3498. 2 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. CoRR, abs/2002.05709, 2020. 1, 8 + +[12] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. SimCLRv2: Big Self-Supervised Models are Strong Semi-Supervised Learners. arXiv:2006.10029 [cs, stat], Oct. 2020. arXiv: 2006.10029. 2, 8 +[13] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. MoCov2: Improved Baselines with Momentum Contrastive Learning. arXiv:2003.04297 [cs], Mar. 2020. arXiv: 2003.04297. 2 +[14] Xinlei Chen and Kaiming He. SimSiam: Exploring Simple Siamese Representation Learning. arXiv:2011.10566 [cs], Nov. 2020. arXiv: 2011.10566. 2 +[15] Xinlei Chen, Saining Xie, and Kaiming He. MoCov3: An Empirical Study of Training Self-Supervised Vision Transformers. arXiv:2104.02057 [cs], Aug. 2021. arXiv: 2104.02057. 2 +[16] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: UNiversal Image-TExt Representation Learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, Lecture Notes in Computer Science, pages 104-120, Cham, Sept. 2020. Springer International Publishing. 2 +[17] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. Technical Report arXiv:2003.10555, arXiv, Mar. 2020. arXiv:2003.10555 [cs] type: article. 5 +[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1 +[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, June 2009. ISSN: 1063-6919. 5, 7 +[20] Karan Desai and Justin Johnson. VirTex: Learning Visual Representations from Textual Annotations. arXiv:2006.06666 [cs], Sept. 2021. arXiv:2006.06666. 2 +[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 [cs], May 2019. arXiv: 1810.04805. 2, 4 +[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. ViT: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv:2010.11929 [cs], June 2021. arXiv: 2010.11929.4 +[23] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuo-hang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. METER: An Empirical Study of Training End-to-End Vision-and-Language Transformers. arXiv:2111.02387 [cs], Nov. 2021. arXiv: 2111.02387. 2 + +[24] Jiali Duan, Liquin Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal Alignment using Representation Codebook. arXiv:2203.00048 [cs], Mar. 2022. arXiv: 2203.00048. 2 +[25] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herve Jegou, and Edouard Grave. SplitMask: Are Large-scale Datasets Necessary for Self-Supervised Pretraining? arXiv:2112.10740 [cs], Dec. 2021. arXiv: 2112.10740. 2 +[26] Patrick Esser, Robin Rombach, and Björn Ommer. VQGAN: Taming Transformers for High-Resolution Image Synthesis. arXiv:2012.09841 [cs], June 2021. arXiv: 2012.09841.2 +[27] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2):303-338, June 2010. 19 +[28] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip), 2022. 8 +[29] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale, 2022. 2 +[30] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 8 +[31] Xinyang Geng, Hao Liu, Lisa Lee, Dale Schuurmans, Sergey Levine, and Pieter Abbeel. M3AE: Multimodal Masked Autoencoders Learn Transferable Representations. Technical Report arXiv:2205.14204, arXiv, May 2022. arXiv:2205.14204 [cs] type: article. 1, 3, 4, 5, 7, 14, 21 +[32] Aaron Gokaslan and Vanya Cohen. Openwebtext corpus. http://Skylion007.github.io/OpenWebTextCorpus, 2019.5 +[33] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. arXiv:1706.02677 [cs], Apr. 2018. arXiv: 1706.02677. 15 +[34] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. VQA2: Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. pages 6904-6913, 2017. 5, 17 +[35] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. BYOL: Bootstrap your own latent: A new approach to self-supervised Learning. arXiv:2006.07733 [cs, stat], Sept. 2020. arXiv: 2006.07733. 2 + +[36] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. 1, 8 +[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. MAE: Masked Autoencoders Are Scalable Vision Learners. arXiv:2111.06377 [cs], Nov. 2021. arXiv: 2111.06377. 1, 2, 3, 4, 5 +[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. MoCov1: Momentum Contrast for Unsupervised Visual Representation Learning. pages 9729-9738, 2020. 2 +[39] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9068-9077, June 2022. 1 +[40] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen, 2022. 8 +[41] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing Out of the Box: End-to-End Pre-Training for Vision-Language Representation Learning. pages 12976-12985, Apr. 2021. 2 +[42] Drew A. Hudson and Christopher D. Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. pages 6700-6709, 2019. 5, 17 +[43] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. CoRR, abs/2102.05918, 2021. 1, 4, 8 +[44] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. ALIGN: Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. arXiv:2102.05918 [cs], June 2021. arXiv: 2102.05918.3, 14 +[45] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1988–1997, 2017. 5, 7, 17 +[46] Kaggle and EyePacs. Kaggle diabetic retinopathy detection. https://www.kaggle.com/c/diabetic-retinopathy-detection/data, July 2015. 14 +[47] Wonjae Kim, Bokyung Son, and Ildoo Kim. ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision. arXiv:2102.03334 [cs, stat], June 2021. arXiv: 2102.03334. 2 +[48] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. Technical Report arXiv:1412.6980, arXiv, Jan. 2017. arXiv:1412.6980 [cs] type: article. 15 + +[49] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 8 +[50] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. arXiv:1910.13461 [cs, stat], Oct. 2019. arXiv: 1910.13461.2, 4 +[51] Chenliang Li, Ming Yan, Haiyang Xu, Fuli Luo, Wei Wang, Bin Bi, and Songfang Huang. SemVLP: Vision-Language Pre-training by Aligning Semantics at Multiple Levels. arXiv:2103.07829 [cs], Mar. 2021. arXiv: 2103.07829. 2 +[52] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. arXiv:2107.07651 [cs], Oct. 2021. arXiv: 2107.07651. 2, 4 +[53] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO-2: End-to-End Unified Vision-Language Grounded Learning. arXiv:2203.09067 [cs], Mar. 2022. arXiv: 2203.09067. 2 +[54] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, Lecture Notes in Computer Science, pages 121–137, Cham, July 2020. Springer International Publishing. 5 +[55] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pretraining via masking, 2022. 2 +[56] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. CoRR, abs/2110.05208, 2021. 1 +[57] Feng Liang, Yangguang Li, and Diana Marculescu. SupMAE: Supervised Masked Autoencoders Are Efficient Vision Learners. Technical Report arXiv:2205.14540, arXiv, May 2022. arXiv:2205.14540 [cs] type: article. 3 +[58] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 16, 19 +[59] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic Gradient Descent with Warm Restarts. arXiv:1608.03983 [cs, math], May 2017. arXiv: 1608.03983. 5, 15 +[60] Ilya Loshchilov and Frank Hutter. AdamW: Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs, math], Jan. 2019. arXiv: 1711.05101. 5, 15 +[61] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. arXiv:1908.02265 [cs], Aug. 2019. arXiv: 1908.02265. 2 + +[62] Shlok Mishra, Joshua Robinson, Huiwen Chang, David Jacobs, Aaron Sarna, Aaron Maschinot, and Dilip Krishnan. A simple, efficient and scalable contrastive masked autoencoder for learning visual representations, 2022. 8 +[63] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. SLIP: Self-supervision meets Language-Image Pre-training. arXiv:2112.12750 [cs], Dec. 2021. arXiv: 2112.12750.1 +[64] Norman Mu, Alexander Kirillov, David A. Wagner, and Saining Xie. SLIP: self-supervision meets language-image pre-training. CoRR, abs/2112.12750, 2021. 1, 2, 5 +[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748 [cs, stat], Jan. 2019. arXiv: 1807.03748. 2, 3 +[66] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. CoRR, abs/2111.10050, 2021. 1, 8 +[67] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting Region-to-Phrase Correspondences for Richer Image-to-Sentence Models. pages 2641–2649, 2015. 16 +[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. CLIP: Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020 [cs], Feb. 2021. arXiv: 2103.00020. 1, 2, 3, 14, 15 +[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1, 5, 8 +[70] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. GPT-2: Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4, 15 +[71] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander T Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models, 2023. 3, 6, 8 +[72] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating Diverse High-Fidelity Images with VQ-VAE-2. Technical Report arXiv:1906.00446, arXiv, June 2019. arXiv:1906.00446 [cs, stat] type: article. 2 +[73] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenna Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 1 +[74] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo + +Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION400M: Open Dataset of CLIP-Filtered 400 Million ImageText Pairs. arXiv:2111.02114 [cs], Nov. 2021. arXiv:2111.02114.6, 14 +[75] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision, 128(2):336-359, Feb. 2020. arXiv:1610.02391 [cs]. 8 +[76] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. 14 +[77] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. GCC: Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 5, 6 +[78] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. CoRR, abs/2112.04482, 2021. 1, 4 +[79] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A Foundational Language And Vision Alignment Model. arXiv:2112.04482 [cs], Feb. 2022. arXiv: 2112.04482. 3 +[80] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross B. Girshick, Piotr Dólar, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. CoRR, abs/2201.08371, 2022. 1 +[81] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of Generic Visual-Linguistic Representations. arXiv:1908.08530 [cs], Feb. 2020. arXiv: 1908.08530. 2 +[82] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1 +[83] Yihong Sun, Adam Kortylewski, and Alan Yuille. Amodal Segmentation through Out-of-Task and Out-of-Distribution Generalization with a Bayesian Model. Technical Report arXiv:2010.13175, arXiv, July 2022. arXiv:2010.13175 [cs] type: article. 16 +[84] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. arXiv:1908.07490 [cs], Dec. 2019. arXiv: 1908.07490. 2 +[85] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine learning research, 9(11), 2008. 17 +[86] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, $\backslash$ Lukasz Kaiser, + +and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 4 +[87] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning - ICML '08, pages 1096-1103, Helsinki, Finland, 2008. ACM Press. 2 +[88] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for all vision and vision-language tasks, 2022. 1 +[89] Xiaolong Wang and Abhinav Gupta. Unsupervised Learning of Visual Representations using Videos. Technical Report arXiv:1505.00687, arXiv, Oct. 2015. arXiv:1505.00687 [cs] type: article. 2 +[90] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple Visual Language Model Pretraining with Weak Supervision. arXiv:2108.10904 [cs], Aug. 2021. arXiv: 2108.10904. 2 +[91] Longhui Wei, Lingxi Xie, Wengang Zhou, Houqiang Li, and Qi Tian. MVP: Multimodality-guided Visual Pretraining. arXiv:2203.05175 [cs], Mar. 2022. arXiv: 2203.05175. 3 +[92] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 14 +[93] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dólar, and Ross Girshick. Early Convolutions Help Transformers See Better. Technical Report arXiv:2106.14881, arXiv, Oct. 2021. arXiv:2106.14881 [cs] type: article. 5 +[94] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. Pre-layernorm: On Layer Normalization in the Transformer Architecture. In Proceedings of the 37th International Conference on Machine Learning, pages 10524-10533. PMLR, Nov. 2020. ISSN: 2640-3498. 4, 15 +[95] Jianwei Yang, Chunyuan Li, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Focal modulation networks, 2022. 1 +[96] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained Interactive Language-Image Pre-Training. arXiv:2111.07783 [cs], Nov. 2021. arXiv: 2111.07783. 2 +[97] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. ERNIE-ViL: Knowledge Enhanced Vision-Language Representations Through Scene Graph. June 2020. 2 +[98] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models, 2022. 1 + +[99] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive Captioners are Image-Text Foundation Models. arXiv:2205.01917 [cs], May 2022. arXiv: 2205.01917. 2 +[100] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A New Foundation Model for Computer Vision. arXiv:2111.11432 [cs], Nov. 2021. arXiv: 2111.11432. 2 +[101] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-level Representation Learning for Self-supervised Vision Transformers. Technical Report arXiv:2206.07990, arXiv, June 2022. arXiv:2206.07990 [cs] type: article. 2 +[102] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. arXiv:2103.03230 [cs, q-bio], June 2021. arXiv: 2103.03230. 2 +[103] Yan Zeng, Xinsong Zhang, and Hang Li. X:VLM: MultiGrained Vision Language Pre-Training: Aligning Texts with Visual Concepts. arXiv:2111.08276 [cs], Feb. 2022. arXiv: 2111.08276. 3 +[104] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12104-12113, June 2022. 3 +[105] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, André Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, Lucas Beyer, Olivier Bachem, Michael Tschannen, Marcin Michalski, Olivier Bousquet, Sylvain Gelly, and Neil Houlsby. The visual task adaptation benchmark. CoRR, abs/1910.04867, 2019. 1, 14 +[106] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djo-longa, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, Lucas Beyer, Olivier Bachem, Michael Tschannen, Marcin Michalski, Olivier Bousquet, Sylvain Gelly, and Neil Houlsby. VTAB: A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark. arXiv:1910.04867 [cs, stat], Feb. 2020. arXiv: 1910.04867.5,7,14 +[107] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. ADE20K: Semantic Understanding of Scenes through the ADE20K Dataset. arXiv:1608.05442 [cs], Oct. 2018. arXiv: 1608.05442. 19 \ No newline at end of file diff --git a/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/images.zip b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba531be39129f6f1ff6e8e8c2ca33356e9c877a2 --- /dev/null +++ b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66a5adaf80b10836a7913a23f2a7eb387b55184060dc1900ffe910e0758174d5 +size 404824 diff --git a/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/layout.json b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00e32ee6497220999c1a46aa262e57d861b673f1 --- /dev/null +++ b/2023/Masked Autoencoding Does Not Help Natural Language Supervision at Scale/layout.json @@ -0,0 +1,9871 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "text", + "content": "Masked Autoencoding Does Not Help Natural Language Supervision at Scale" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 143, + 534, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 143, + 534, + 157 + ], + "spans": [ + { + "bbox": [ + 59, + 143, + 534, + 157 + ], + "type": "text", + "content": "Floris Weers Vaishaal Shankar Angelos Katharopoulos Yinfei Yang Tom Gunter" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 159, + 481, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 159, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 109, + 159, + 481, + 171 + ], + "type": "text", + "content": "{fweers,vaishaal_shankar,a_katharopoulos,yinfeiy,tom_gunter}@apple.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 207, + 192, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 207, + 192, + 220 + ], + "spans": [ + { + "bbox": [ + 143, + 207, + 192, + 220 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 232, + 290, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 290, + 462 + ], + "type": "text", + "content": "Self supervision and natural language supervision have emerged as two exciting ways to train general purpose image encoders which excel at a variety of downstream tasks. Recent works such as M3AE [31] and SLIP [63] have suggested that these approaches can be effectively combined, but most notably their results use small (<20M examples) pre-training datasets and don't effectively reflect the large-scale regime (>100M samples) that is commonly used for these approaches. Here we investigate whether a similar approach can be effective when trained with a much larger amount of data. We find that a combination of two state of the art approaches: masked auto-encoders, MAE [37] and contrastive language image pre-training, CLIP [68] provides a benefit over CLIP when trained on a corpus of 11.3M image-text pairs, but little to no benefit (as evaluated on a suite of common vision tasks) over CLIP when trained on a large corpus of 1.4B images. Our work provides some much needed clarity into the effectiveness (or lack thereof) of self supervision for large-scale image-text training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "type": "text", + "content": "Large scale pretraining has become a powerful tool in the arsenal of computer vision researchers to produce state of the art results across a wider variety of tasks [39,88,95,98]. However, when pre-training on tens of millions to billions of images it is difficult to rely on standard supervised methods to train models, as datasets of this size often lack reliable labels. In the presence of these massive but largely under-curated datasets, two general classes of methods to train general purpose image encoders have emerged:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 622, + 287, + 692 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 622, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 287, + 647 + ], + "type": "text", + "content": "1. Self Supervised techniques that learn visual representations from the image data alone [11,36]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 655, + 287, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 655, + 287, + 692 + ], + "spans": [ + { + "bbox": [ + 54, + 655, + 287, + 692 + ], + "type": "text", + "content": "2. Natural Language Supervised methods that utilize paired free-form text data to learn visual representations [43, 69]" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Due to the unique strengths and weaknesses of each ap-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 208, + 547, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 547, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 547, + 400 + ], + "type": "text", + "content": "proach" + }, + { + "bbox": [ + 304, + 208, + 547, + 400 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 208, + 547, + 400 + ], + "type": "text", + "content": ", a recent flurry of work has introduced methods that combine both forms of supervision [31, 56, 64, 78] to varying degrees of success. While each of these methods establishes some regime where the additional supervision helps, none of these \"joint-supervision\" methods advance state of the art in any meaningful way. Additionally, to our knowledge none of these methods have shown comparative results at the scale many large scale vision models are currently trained at (" + }, + { + "bbox": [ + 304, + 208, + 547, + 400 + ], + "type": "inline_equation", + "content": ">100\\mathrm{M}" + }, + { + "bbox": [ + 304, + 208, + 547, + 400 + ], + "type": "text", + "content": " examples) [43, 66, 69, 73, 80, 82, 98]. Furthermore, methods that use both forms of supervision start with the presumption that the additional supervision is helpful and either often lack clean ablations or lack evaluations in a \"high accuracy\" regime—leading to further confusion regarding whether a combination of these methods can actually improve the state of the art. To clarify this issue, in this work, we investigate a simple question:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 407, + 526, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 407, + 526, + 443 + ], + "spans": [ + { + "bbox": [ + 325, + 407, + 526, + 443 + ], + "type": "text", + "content": "Does a combination of self supervision and natural language supervision actually lead to higher quality visual representations?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 605 + ], + "type": "text", + "content": "In order to answer this, we first introduce a straightforward baseline approach that combines standard self-supervision and language supervision techniques. We combine masked auto-encoders (MAE) and contrastive language image-pretraining (CLIP) to make MAE-CLIP. We then present a careful study of the performance of MAE, M3AE, CLIP and MAE-CLIP across a wide variety of tasks in two distinct regimes: a \"low-sample\" " + }, + { + "bbox": [ + 304, + 449, + 547, + 605 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 304, + 449, + 547, + 605 + ], + "type": "text", + "content": " 11.3 million example regime and a \"high-sample\" 1.4 billion example regime. We train self-supervised and language-supervised methods using the same pre-training datasets under the assumption that we have no knowledge about downstream tasks. Our experiments show:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 610, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 547, + 658 + ], + "type": "text", + "content": "1. In the low sample size regime, without changing the final pooling operation in the network, we observe a large performance improvement, namely " + }, + { + "bbox": [ + 313, + 610, + 547, + 658 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 610, + 547, + 658 + ], + "type": "text", + "content": " on ImageNet [18] and " + }, + { + "bbox": [ + 313, + 610, + 547, + 658 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 610, + 547, + 658 + ], + "type": "text", + "content": " on VTAB [105]. However, when" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 664, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 306, + 664, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 545, + 693 + ], + "type": "text", + "content": "Self supervised methods can learn representations without labels, but natural language supervision learns better representations. Natural language supervised methods rely on quality of captions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "2We note that what low sample means has changed substantially over the last few years" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23432" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "type": "text", + "content": "we modify the pooling operation, the improvement substantially decreases to around " + }, + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "type": "text", + "content": " on both ImageNet and VTAB." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 286, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 286, + 153 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 286, + 153 + ], + "type": "text", + "content": "2. In the high sample size regime, there is virtually no difference in performance between MAE-CLIP and CLIP across ImageNet, VTAB, and VQA tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 164, + 286, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 286, + 199 + ], + "type": "text", + "content": "We believe our work is the first careful study of this form and contextualizes recent progress in both self-supervision and natural language supervision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 201, + 286, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 286, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 286, + 295 + ], + "type": "text", + "content": "The rest of the paper is organized as follows: In Section 2, we cover related work in the areas of self supervision and natural language supervision. In Section 3, we give an overview of the baseline methods we study, MAE, M3AE, CLIP and our new baseline MAE-CLIP. Then we present and analyse our small scale and large scale experimental findings in Sections 4 and 5. Finally, we discuss potential explanations for our findings and some future work in 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 308, + 133, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 133, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 133, + 320 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 330, + 286, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 286, + 364 + ], + "type": "text", + "content": "Our work combines natural language supervision and self-supervision in a multi-task approach to visual encoding, and so research from these three areas is relevant." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 366, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 286, + 675 + ], + "type": "text", + "content": "Natural language supervision for visual encoding covers a variety of approaches that assume access to datasets of images or videos associated with text. Some of the most successful use a contrastive pairwise alignment signal applied to very large batch and dataset sizes [68]. This large batch size means hard negative pairs can be produced by random sampling, avoiding the need for a memory bank [12, 13, 38] or momentum distillation [15]. FILIP [96] further improves zero-shot performance for some tasks by using a more fine-grained elementwise contrastive loss. Image captioning has also shown promise as a pre-training task [20, 90], with [20] in particular demonstrating strong data efficiency. UNITER [16] pursues a similar idea, applying a generative loss to both image and text modalities. [2, 16, 23, 41, 47, 51, 61, 81, 84, 97] apply the masked patch prediction problem from [21, 37, 50] to a joint image-text data space. These approaches either use a pre-trained convolutional neural network (CNN) to generate region of interest (RoI) proposal encodings, or predict the labelled class of the masked patch instead of the raw pixel values (or the quantized patch ID). UNIMO-2 [53] attempts to ground image and text patches for a single data example into a top-k quantized set of embeddings, before passing the grounded embeddings along with the raw non-grounded embeddings into a decoder, targeting a loss based on masked language modeling (MLM) and image-text-matching (ITM)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "Self-supervised approaches for visual encoding may be loosely categorized into those that target consistency constraints between multiple views of the same scene and those" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 503 + ], + "type": "text", + "content": "that attempt visual reconstruction on corrupted or conditioning representations of images. Consistency constraints are often derived through data-augmentation applied to a single real image [6, 10, 14, 35, 102], but these may also take the form of e.g. a temporal constraint if video data is available [89], or be applied at an image-patch level by making a smoothness assumption and selecting neighboring patches as 'consistent' whilst treating distant patches (or those sourced from another image) as negative examples [101]. Forms of denoising autoencoder (DAE) [87] are popular as a means of self-supervision, and have been investigated at a variety of scales; iGPT [9] learns to generate images as a row-major flattened sequence of pixels, whilst [65], [72], and [26] compress images into a short sequence of discrete codes before then regenerating and scoring at a pixel level. BEiT [3] first encodes an image into a sequence of discrete codes before masking and then predicting a subset of the codes. MAE [37] drops the discretization, instead predicting raw pixel loss for a subset of the encoded patches in a manner strongly reminiscent of BART [50]. SplitMask [25] applies both a masked image patch prediction loss and a pairwise contrastive loss, using the masked example as well as the inverted masked example to form each positive pair. Meanwhile, M3AE [24] and VL-BEiT [4] both propose a masked patch prediction problem applied to image and text modalities jointly. Recent concurrent work shows that while masking does not help generalization, the masking can be used to speed up training by dropping the masked tokens during the forward and backward pass [55]. EVA [29] shows that doing masked image modeling to predict the output embeddings (rather than pixels) of the masked image can be an effective pretraining task when combined with natural language supervision. We note for EVA the masked image modeling occurs only on smaller well curated datasets such as ImageNet, Conceptual Captions and COCO." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Multi-task methods for pre-training visual encoders are a highly active area of research. CoCA [99] combines the cross-modality contrastive task from [68] with image captioning in the style of [20], and by pre-training at very large scale show that the resulting model is more performant than prior art across a very broad array of downstream visual understanding tasks. SLIP [64] combines a SimCLR style self supervision loss with the standard CLIP contrastive loss to train a jointy self-supervised and natural language supervised model, however while they show large performance gains they are all in a low accuracy regime (below " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " Top-1 accuracy on ImageNet). Florence [100] employs a diverse array of tasks and datasets at pre-training time (including object detection, CLIP-style contrastive modeling, Visual Question Answering (VQA), and supervised classification). ALBEF [52] builds on a CLIP-style architecture by incorporating a single-stream encoder that consumes both" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "type": "text", + "content": "23433" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 61, + 500, + 194 + ], + "blocks": [ + { + "bbox": [ + 97, + 61, + 500, + 194 + ], + "lines": [ + { + "bbox": [ + 97, + 61, + 500, + 194 + ], + "spans": [ + { + "bbox": [ + 97, + 61, + 500, + 194 + ], + "type": "image", + "image_path": "b8d0073fe48fb99b79de939d95b11dfd5570288c17649fbc5ff4b69761c64648.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 201, + 474, + 213 + ], + "lines": [ + { + "bbox": [ + 118, + 201, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 118, + 201, + 474, + 213 + ], + "type": "text", + "content": "Figure 1. MAE-CLIP: CLIP [68] augmented with a generative decoder in the style of MAE [31]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 219, + 289, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 219, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 289, + 542 + ], + "type": "text", + "content": "modalities. It uses a masked language modeling loss for the text modality and an image-text matching loss for within-batch hard-negatives, mined according to contrastive similarity. FLAVA [79] also uses masked-image modeling, and uses both single and joint modality encoders, with additional single-modality datasets (i.e. not just image-text pairs) that allow it to generalize to longer text inputs and visual data which is unlikely to be captioned. X-VLM [103] utilizes image data which includes object-labelled bounding boxes to extend the image-text matching and contrastive losses to cover within-image regions as well as whole-image captions. They also include a masked language modelling objective, achieving impressive performance on many zero-shot downstream tasks—although the requirement for object bounding box data may not scale well for large pre-training datasets. SupMAE [57] adds a supervised labeling task to the MAE architecture, demonstrating that this results in better data efficiency. MVP [91], meanwhile, finetunes a pre-trained CLIP backbone using MIM, and demonstrates that starting with a pre-trained model improves downstream visual recognition performance. MAE-CLIP is reminiscent of M3AE (with the addition of a contrastive task), ALBEF (but applying the masked prediction task to both modalities, dropping the momentum distillation), and most of all FLAVA (without the single-modality masked-patch-prediction tasks, and with the use of high ratio random masking following MAE and M3AE)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 552, + 126, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 126, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 126, + 565 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 574, + 261, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 261, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 261, + 586 + ], + "type": "text", + "content": "3.1. Contrastive language-image pre-training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "CLIP [68] and ALIGN [44] both demonstrated that contrastive image-language supervision applied at scale was capable of producing an image encoder that excels at a range of downstream tasks—often in a zero-shot setting. In our implementation, we begin by following the design outlined in [68]. Specifically, we use per-modality encoders to produce a single " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\ell 2" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "-normalized dense embedding vector for each image or text input, before applying a pairwise InfoNCE [65] loss to a large global batch of paired image-text embeddings, using all non-paired examples as hard nega" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 219, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 545, + 243 + ], + "type": "text", + "content": "tives. We define the image-to-text loss " + }, + { + "bbox": [ + 305, + 219, + 545, + 243 + ], + "type": "inline_equation", + "content": "L_{i2t}" + }, + { + "bbox": [ + 305, + 219, + 545, + 243 + ], + "type": "text", + "content": " and the text-to-image loss " + }, + { + "bbox": [ + 305, + 219, + 545, + 243 + ], + "type": "inline_equation", + "content": "L_{t2i}" + }, + { + "bbox": [ + 305, + 219, + 545, + 243 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 340, + 251, + 545, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 251, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 340, + 251, + 545, + 285 + ], + "type": "interline_equation", + "content": "L _ {i 2 t} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(x _ {j} ^ {T} y _ {j} / \\sigma\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(x _ {j} ^ {T} y _ {k} / \\sigma\\right)} \\tag {1}", + "image_path": "dfd9075d69d3ce2ebad93d36afc1285be996fe58a2d14e27bd2085f765c9a48b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 294, + 545, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 339, + 294, + 545, + 327 + ], + "type": "interline_equation", + "content": "L _ {t 2 i} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(y _ {j} ^ {T} x _ {j} / \\sigma\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(y _ {j} ^ {T} x _ {k} / \\sigma\\right)}, \\tag {2}", + "image_path": "364914e4b0da0543c529afb3dd96ab0195c51f486aa8ec6dfa80c4c8a7bd2e15.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": " is the global batch size, " + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": " is the normalized embedding of the image for the " + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": "-th pair and " + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": " is the normalized embedding of the text in the " + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": "-th pair in the batch. " + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": " is a learnable temperature parameter. The total contrastive loss " + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "inline_equation", + "content": "L_{c} = \\frac{1}{2} (L_{t2i} + L_{i2t})" + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": " is the average of these two losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 392, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 463 + ], + "type": "text", + "content": "We use transformer encoders for both modalities, however, differently to [68] we explore several strategies for aggregating the image-patch or text-token outputs into a single embedding (see Section 4.4). We also eschew the autoregressive masking strategy for the text encoder, instead allowing full bi-directional self attention following [44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 478, + 372, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 372, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 372, + 490 + ], + "type": "text", + "content": "3.1.1 Pooling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 496, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 652 + ], + "type": "text", + "content": "In CLIP [68] the authors use a separate 'CLS' token projected through the network as the overall image representation. On the other hand, ALIGN [44] use global-average pooling over the encoded visual features. Whilst both approaches produced good downstream evaluation results, recent work [8, 71, 104], suggested that the choice of pooling strategy can strongly influence the quality of visual semantic embeddings. In particular, [71] demonstrated that this effect is present for CLIP-like visual encoders. Noting this, we opt to investigate three pooling strategies: 1. the default multihead-attention pooling (MAP), 2. global average pooling (GAP)—both described in [104]—as well as 3. non-maximal suppression pooling (MAX) as in [8, 71]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 659, + 432, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 432, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 432, + 670 + ], + "type": "text", + "content": "3.2. Masked Autoencoders" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In MAE [37] the authors demonstrate a simple technique for self-supervised image-encoder pre-training that—to our knowledge—is still considered state-of-the-art. They use a" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "23434" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": "ViT [22] encoder-decoder architecture and apply it to heavily masked input images. The input to the encoder consists of the visible, unmasked image patches, first embedded via a linear projection before additive positional embeddings are applied, and then the result is fed through the encoder's transformer layers. They demonstrate that a very high patch masking ratio is critical to achieving good performance, and usually retain only " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": " of the image. The decoder consumes the output of the encoder, as well as a learned 'masked-patch' embedding, which is included to represent each masked token. Positional embeddings are also reapplied, to ensure that spatial information is passed through the decoder for the masked patches. The output of the decoder at the masked positions is then measured against ground truth using a simple mean-squared-error loss. As in [37] we experiment with predicting both normalized and un-normalized patch values, finding that predicting the normalized patch value slightly improves the performance of our MAE implementation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 288, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 409 + ], + "type": "text", + "content": "Masking across modalities: BART [50] applies a similar strategy to text data, but uses a cross-entropy loss over the masked-token output distribution and a far lower masking ratio (typically around " + }, + { + "bbox": [ + 46, + 300, + 288, + 409 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 300, + 288, + 409 + ], + "type": "text", + "content": "). M3AE [31] extends MAE and BART to incorporate inputs from both text and image modalities, using per-modality input and output projections. Otherwise the encoder-decoder architecture resembles MAE, with the addition of learned modality-indicating embeddings to each transformer input." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 410, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 288, + 460 + ], + "type": "text", + "content": "The M3AE loss relies only on the contents of the image-text pairs, and comprises mean-squared error for the masked image patches " + }, + { + "bbox": [ + 47, + 410, + 288, + 460 + ], + "type": "inline_equation", + "content": "(L_{gen.i})" + }, + { + "bbox": [ + 47, + 410, + 288, + 460 + ], + "type": "text", + "content": " and cross-entropy over the vocabulary for the masked text tokens " + }, + { + "bbox": [ + 47, + 410, + 288, + 460 + ], + "type": "inline_equation", + "content": "(L_{gen.t})" + }, + { + "bbox": [ + 47, + 410, + 288, + 460 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 480, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 480, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 87, + 480, + 287, + 514 + ], + "type": "interline_equation", + "content": "L _ {\\text {g e n . i}} = \\frac {1}{N} \\sum_ {j} ^ {N} \\left(p _ {j} - P _ {j}\\right) ^ {2} \\tag {3}", + "image_path": "3bb2a22c0e51e471aa3587e518613a1df9045c81cdc72366264d1a20e81fd9bf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 517, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 517, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 87, + 517, + 287, + 552 + ], + "type": "interline_equation", + "content": "L _ {\\text {g e n . t}} = - \\frac {1}{N} \\sum_ {j} ^ {N} \\log \\frac {\\exp \\left(t _ {n , T _ {n}}\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(t _ {n , c}\\right)}, \\tag {4}", + "image_path": "29e78d2820d7eeac95f35e2d7ce89c915d2cde2cc11e87371274ac9b50ba6b11.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "P_j" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": " refer to the predicted and ground truth pixel value respectively and " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "T_n" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": " refer to the predicted token distribution and the ground truth token, with " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": " as the number of unique tokens. For the image reconstruction loss, we normalize the ground-truth per-patch, following [37]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 126, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 126, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 126, + 645 + ], + "type": "text", + "content": "3.3. MAE-CLIP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": "As the name suggests, MAE-CLIP attempts to incorporate aspects of MAE/M3AE into CLIP through the addition of a single dual-modality transformer decoder (as in M3AE) which consumes the output of the CLIP encoders. We show this in Figure 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 394, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 394, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 394, + 85 + ], + "type": "text", + "content": "3.3.1 Components" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 93, + 545, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 93, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 93, + 545, + 118 + ], + "type": "text", + "content": "The model architecture is similar to ALBEF [52] and FLAVA [78], consisting of three components." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 118, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 118, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 547, + 179 + ], + "type": "text", + "content": "An image encoder: Following ViT [22], we divide the input image into equally-sized, non-overlapping patches. After applying a linear projection and adding a 2-D position encoding, we feed the per-patch representations through a number of transformer layers [86]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 180, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 547, + 252 + ], + "type": "text", + "content": "A text encoder: Following ALIGN [43], our text encoder is based on BERT [21], in which text is first tokenized and embedded, and a 1-D trainable position encoding is added. Differently to BERT, we use pre-layernorm [94] and initialize the parameters according to the simplified scheme outlined in [70]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 253, + 546, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 253, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 253, + 546, + 397 + ], + "type": "text", + "content": "A cross-modality decoder: The decoder receives per-element encoded image and text representations from the encoders for both masked and un-masked elements. For the image modality, masked patches are added by replacing their values with a shared trainable mask token. Positional encodings are once again added to all elements, following [37]. We also add a per-modality trainable encoding to allow the decoder to easily distinguish between the two modalities. The decoder uses the same transformer implementation as the encoders. The output of the final decoder layer is then projected into a per-patch-modality output space and the overall loss is computed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 415, + 367, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 367, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 367, + 426 + ], + "type": "text", + "content": "3.3.2 Losses" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 437, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 473 + ], + "type": "text", + "content": "MAE-CLIP employs the losses of both CLIP (Eq. 1, 2) and M3AE (Eq. 3, 4). Our final loss is a weighted sum of the losses from each task, as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 330, + 483, + 545, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 483, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 330, + 483, + 545, + 507 + ], + "type": "interline_equation", + "content": "L = \\frac {1}{2} \\left(L _ {t 2 i} + L _ {i 2 t}\\right) + w _ {i} \\cdot L _ {\\text {g e n . i}} + w _ {t} \\cdot L _ {\\text {g e n . t}} \\tag {5}", + "image_path": "1e11abbecf538c694e4f03696febb680f54f24692e335c0c6d90801a38afd0ac.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 517, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 545, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 517, + 545, + 576 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 305, + 517, + 545, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 517, + 545, + 576 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 305, + 517, + 545, + 576 + ], + "type": "text", + "content": " are scalars used to control the relative weight of the generative losses. We always provide paired image-text inputs to the model, leaving it to future work to explore the benefits of also incorporating single modality inputs and reconstruction losses, at scale." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 578, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 545, + 662 + ], + "type": "text", + "content": "To avoid severely impacting the CLIP loss via computing it with masked inputs, we run the encoders twice: once with full unmasked input in order to compute the CLIP loss, before then making a second pass using only the unmasked input to compute the M3AE loss. Finally, we compute the weighted sum of the two losses, and use this to calculate the overall update." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 673, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 673, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 547, + 713 + ], + "type": "text", + "content": "3 Alternative, more memory efficient, strategies exist for computing the overall updates (e.g. calculate per-loss gradients, or perform round-robin updates), but prior work [1] has demonstrated that some of these techniques may negatively impact the overall training result." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23435" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "text", + "content": "In this section, we present our experimental results to study how the addition of self-supervision affects the representation quality of natural language supervised models. We study standard masked auto-encoders (MAE), multimoddal masked auto-encoders (M3AE), contrastive language image pretraining (CLIP) and our newly introduced baseline method (MAE-CLIP), which combines mask based self-supervision with contrastive language-image learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 207, + 163, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 163, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 163, + 220 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 225, + 287, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 356 + ], + "type": "text", + "content": "Architecture: Experiments are performed using an image encoder based on the architecture described as ViT-B-16 in [93], in combination with a 12-layer, 512-wide text encoder with 8 attention heads. For the decoder, we use another 12-layer, 512-wide transformer with 8 attention heads, identical to the text encoder. We also use a byte-pair encoding tokenizer, fit to the OpenWebText corpus [32]. When training MAE-CLIP, we replace the random masking by a similarity masking strategy, which makes use of the element-wise CLIP similarity scores to select the masked element for the reconstruction task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": "Pretraining Datasets: Our analysis is divided into two sections. First, in sections 4.2 to 4.4, we present our study on the \"low-sample\" regime using the combination of CC12M [7] and CC3M [77] excluding all images whose text contains a [PERSON] tag5. The final dataset contains 11.3 million comparatively high quality image-text pairs. We refer to this dataset as CC or \"small\" scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": "Pretraining configuration: We use a minibatch size of 16,384 samples and train using the AdamW optimizer [60] with decoupled weight decay. We use a base learning rate of " + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": " and train using cosine learning rate decay [59] with an initial linear warmup of 200 steps. We reduce the " + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": " decay factor to 0.98, and increase " + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": ", finding that these changes improved rate of convergence for all models trained. To increase training convergence speed, we switch from a local (per-GPU) to a global contrastive loss after 500 steps. During the global contrastive loss phase, we set the generative text and image loss weights to " + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "inline_equation", + "content": "w_{t} = 0.05" + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "inline_equation", + "content": "w_{i} = 0.1" + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": " respectively. For all models which rely on masking, we use a " + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": " masking ratio, consistent with [31,37], as we did not find an alternative that improved downstream results. In total, we train our models on CC for 32 epochs, which corresponds to " + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "inline_equation", + "content": "\\sim 22,000" + }, + { + "bbox": [ + 46, + 441, + 288, + 668 + ], + "type": "text", + "content": " steps. For our M3AE training runs, we use per-modality encoders so as to produce a baseline that is flop and parameter matched to our MAE-CLIP runs." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 337, + 70, + 515, + 144 + ], + "blocks": [ + { + "bbox": [ + 337, + 70, + 515, + 144 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 515, + 144 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 515, + 144 + ], + "type": "table", + "html": "
ModelsZero-shotLinear Probing
MAE-33.9
M3AE-52.5
CLIP29.752.6
MAE-CLIP33.858.9
", + "image_path": "661f772e83024212c326dd63c6f2dc2a9a0f70d4b96ed4e45c8ac31c93dbaf4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "type": "text", + "content": "Table 1. ImageNet classification with zero-shot transfer or linear probing after pretraining on the " + }, + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "type": "inline_equation", + "content": "CC" + }, + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "type": "text", + "content": " dataset (11.3M images). MAE-CLIP significantly improves the classification performance of CLIP in the small scale regime." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 202, + 423, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 202, + 423, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 423, + 216 + ], + "type": "text", + "content": "4.2. Image Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 221, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 327 + ], + "type": "text", + "content": "We evaluate the quality of the representations learned from pre-training on " + }, + { + "bbox": [ + 304, + 221, + 545, + 327 + ], + "type": "inline_equation", + "content": "CC" + }, + { + "bbox": [ + 304, + 221, + 545, + 327 + ], + "type": "text", + "content": " by measuring the image classification performance, either by zero-shot transfer, or by training a linear classifier using the predicted visual features. For zero-shot classification, we use the average of 80 prompts and follow the routine prescribed in [69]. To train the linear classifier, we pre-compute features using the visual encoder6 and run AdamW for 20 to 80 epochs with a learning rate of 0.01." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 329, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 437 + ], + "type": "text", + "content": "Table 1 shows the zero-shot and linear probing classification results on the ImageNet [19] dataset while Table 2 shows our linear probing results on most (17) of the VTAB [106] benchmark datasets7. Initially, we observe that the combination of self-supervision and natural language supervision provides a consistent and substantial improvement over either form of supervision by itself. This result concurs with previous works such as [31] and [64], which show that self-supervision aids natural language supervision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 444, + 353, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 353, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 353, + 456 + ], + "type": "text", + "content": "4.3.VQA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 463, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 653 + ], + "type": "text", + "content": "Subsequently, we compare the performance of MAE-CLIP to the rest of the baselines on the visual question answering task using three datasets. CLEVR [45], VQAv2 [34] and GQA [42]. VQA assesses the model's multimodal reasoning capabilities as well as its visual grounding. In order to finetune our models for VQA, we freeze the image and text encoders and either randomly reinitialize the decoder in models such as MAE-CLIP or M3AE or add a new identical decoder for CLIP. During finetuning, we use a layer-wise learning rate decay, following [3, 17, 37]. For CLEVR and VQAv2 we finetune the decoder for 50 epochs, while for GQA we finetune for 5 epochs on the \"train-all\" split and 2 epochs on the \"train-balanced\" split following the protocol of [54]. In all cases, we treat the problem as a classification problem of selecting one answer out of the set of possible answers for each dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 654, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 654, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 654, + 545, + 666 + ], + "type": "text", + "content": "In Table 3 we observe that self-supervision combined" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 286, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 286, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 286, + 693 + ], + "type": "text", + "content": "4We compare the similarity masking strategy with the original random masking strategy on various tasks, the results show they perform similarly." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "5 Preliminary experiments on CLIP suggest removing such data leads to better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 673, + 433, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 673, + 433, + 684 + ], + "spans": [ + { + "bbox": [ + 317, + 673, + 433, + 684 + ], + "type": "text", + "content": "6We use no data augmentation here" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 684, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 684, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 684, + 545, + 712 + ], + "type": "text", + "content": "7We do not include results for either the diabetic retinopathy or the Sun397 tasks, due to licensing issues. More details in the supplementary material" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "23436" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 82, + 520, + 196 + ], + "blocks": [ + { + "bbox": [ + 47, + 59, + 545, + 80 + ], + "lines": [ + { + "bbox": [ + 47, + 59, + 545, + 80 + ], + "spans": [ + { + "bbox": [ + 47, + 59, + 545, + 80 + ], + "type": "text", + "content": "Table 2. Linear probing accuracy (%) on classification tasks. All models are trained on the CC Dataset (11.3M images). (● VTAB/natural, ● VTAB/specialized and ● VTAB/structured.)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 82, + 520, + 196 + ], + "lines": [ + { + "bbox": [ + 75, + 82, + 520, + 196 + ], + "spans": [ + { + "bbox": [ + 75, + 82, + 520, + 196 + ], + "type": "table", + "html": "
Caltech101CIFAR-100DTDFlowers102PetsSVHNEuroSATCamelyonResisc45Clevr/CloestClevr/CountDMLabdSprites/OridSprites/LocKITTI/DistsNORB/AzimsNORB/ElevAverage
MAE75.356.258.570.237.471.996.282.884.065.160.642.837.979.936.737.464.562.2
M3AE85.666.669.486.558.469.297.482.990.769.471.248.249.883.342.634.373.469.3
CLIP84.262.857.781.669.752.595.582.686.753.752.944.845.961.345.731.042.661.8
MAE-CLIP89.866.264.886.574.556.795.481.686.953.659.645.944.668.648.036.151.065.3
", + "image_path": "c1d7ee7eb17a7ec93fe7519a1da9a1b48a23b0f187043ceb61de04ee95cf43b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 81, + 200, + 254, + 268 + ], + "blocks": [ + { + "bbox": [ + 81, + 200, + 254, + 268 + ], + "lines": [ + { + "bbox": [ + 81, + 200, + 254, + 268 + ], + "spans": [ + { + "bbox": [ + 81, + 200, + 254, + 268 + ], + "type": "table", + "html": "
ModelCLEVRVQAv2GQA
MAE93.545.746.1
M3AE97.555.850.9
CLIP87.555.650.0
MAE-CLIP96.058.552.2
", + "image_path": "e0d22ff91b7f6efdf204110fecec6e56a63c89dd5770da3f5c5388aae68c3fbd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "text", + "content": "with natural language supervision (MAE-CLIP) performs consistently better than either MAE or CLIP by themselves. Combining information from both modalities is evidently critical for effectively answering visual questions as the performance of MAE on VQAv2 and GQA is several percentage points lower than the rest of the methods. Interestingly, methods that employ a form of self-supervision perform significantly better on CLEVR. We argue that this discrepancy is due to CLEVR requiring object localization and spatial reasoning which benefits from this form of self-supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 476, + 147, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 147, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 147, + 489 + ], + "type": "text", + "content": "4.4. Pooling Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 662 + ], + "type": "text", + "content": "To better understand whether the gap between MAE-CLIP and CLIP we observed in the previous section is fundamental, we investigate the effect of the pooling operator on different downstream tasks. Table 4 compares the performance of CLIP and MAE-CLIP with the standard multi-head attention pooling (MAP), global average pooling (GAP) and max pooling (MAX) on all image classification tasks and visual question answering. We note that GAP and MAX pooling perform substantially better than the standard MAP pooling across ImageNet, VTAB and VQA tasks. A hypothesis posed by [71] is that the alternative pooling operators lead to better perceptual grouping in the visual representation. We study this with a qualitative analysis in Section 6.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 669, + 170, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 170, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 170, + 682 + ], + "type": "text", + "content": "5. Experiments at Scale" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "At small-scale, we showed that self-supervision combined with natural language supervision marginally im" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 200, + 537, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 276, + 287, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 331 + ], + "type": "text", + "content": "Table 3. VQA finetuning results for models pre-trained on CC Dataset (11.3M images). We train a new decoder for all methods while keeping the encoders frozen. MAE-CLIP performs significantly better than either only self-supervised or language supervised methods by themselves." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 200, + 537, + 312 + ], + "lines": [ + { + "bbox": [ + 315, + 200, + 537, + 312 + ], + "spans": [ + { + "bbox": [ + 315, + 200, + 537, + 312 + ], + "type": "table", + "html": "
ModelPoolingImageNetVTABVQA
ZSLPLPFT
CLIPMAP29.752.661.864.4
MAE-CLIPMAP33.858.965.368.9
CLIPGAP29.359.870.665.1
MAE-CLIPGAP33.562.571.769.2
CLIPMAX33.162.368.763.5
MAE-CLIPMAX35.263.269.168.5
", + "image_path": "41c8ec98937388f3fd0fb5f0fe93ca15a6ae3bfd16c3dbb0aed025d4a1d9c538.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 319, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 374 + ], + "type": "text", + "content": "Table 4. Results of CLIP and MAE-CLIP with different pooling options trained on CC dataset (11.3M images). We compare the models on ImageNet Zero-Shot (ZS), ImageNet Linear-Probing (LP), VTAB 17-task Linear-Probing average, and VQA 3 tasks Fine-Tuning (FT) average." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 383, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 383, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 545, + 443 + ], + "type": "text", + "content": "proves the quality of the learned visual representations, and that the choice of image-encoder pooling strategy can heavily influence results. We now investigate whether these conclusions still hold when training on a much larger, 1.4B example dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 453, + 422, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 422, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 422, + 466 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 472, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 495 + ], + "type": "text", + "content": "We follow the configuration described in Section 4.1, with a few key differences, noted below." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 497, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 639 + ], + "type": "text", + "content": "Pretraining Dataset: We combine a 2.2B example web-crawled image-text pair dataset, termed the English-WebImage-Text dataset (EWIT-2.2B), with LAION-400M [74], CC3M [77], CC12M [7] and an internal high-quality image-text pair dataset containing approximately 134M image text pairs which we term the High Quality Image Text Pairs Dataset (HQITP-134M). We globally deduplicate this—keyed by image bytes—reducing the number of image-text pairs significantly, to yield a final 1.4B examples. Details provided in the supplementary material. In the rest of the paper we refer to this as the web-crawled or large-scale dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Pretraining configuration: We increase the learning rate warm-up to 1,000 steps, and compute a local contrastive loss for the first 10,000 steps, as during early training this improves the models rate of convergence. We train for a total of 480,000 steps, which corresponds to 6 full passes through the dataset. We do not increase the batch" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "23437" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 105, + 526, + 240 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 103 + ], + "type": "text", + "content": "Table 5. Linear probing accuracy (%) on classification tasks. Models are all trained on our web-crawled dataset (1.4B images). (VTAB/natural, VTAB/specialized and VTAB/structured.) In the large scale pretraining regime, the difference between MAE-CLIP and CLIP is reduced to " + }, + { + "bbox": [ + 46, + 70, + 545, + 103 + ], + "type": "inline_equation", + "content": "< 1\\%" + }, + { + "bbox": [ + 46, + 70, + 545, + 103 + ], + "type": "text", + "content": ". * At evaluation time, our M3AE model was " + }, + { + "bbox": [ + 46, + 70, + 545, + 103 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 70, + 545, + 103 + ], + "type": "text", + "content": " trained, so that performance may improve further." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 105, + 526, + 240 + ], + "lines": [ + { + "bbox": [ + 66, + 105, + 526, + 240 + ], + "spans": [ + { + "bbox": [ + 66, + 105, + 526, + 240 + ], + "type": "table", + "html": "
Caltech101CIFAR-100DTDFlowers102PetsSVHNEuroSATCarnelyonResisc45Clevi/CloestClevi/CountDMLabdSprites/OridSprites/LocKITTI/DistsNORB/AzimsNORB/ElevAverage
M3AE*93.074.878.295.481.269.697.384.892.765.774.751.052.780.747.536.268.873.2
CLIP94.978.480.097.386.959.094.182.392.745.662.146.046.153.350.920.335.866.2
CLIPMAX96.181.080.997.389.965.796.083.294.152.867.849.959.567.641.223.445.870.1
MAE-CLIPMAX95.879.281.596.888.262.195.881.893.052.066.949.653.772.553.032.345.470.6
CLIPGAP95.880.581.697.688.766.097.084.493.356.771.453.358.070.150.638.355.172.9
MAE-CLIPGAP95.479.382.297.488.672.896.684.593.557.573.652.757.571.251.645.655.273.8
", + "image_path": "f913c492a463eb4ffd78eb5c63f193421a441055e531c55213cdf3bb403c19b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 250, + 264, + 330 + ], + "blocks": [ + { + "bbox": [ + 72, + 250, + 264, + 330 + ], + "lines": [ + { + "bbox": [ + 72, + 250, + 264, + 330 + ], + "spans": [ + { + "bbox": [ + 72, + 250, + 264, + 330 + ], + "type": "table", + "html": "
ModelsZero-shotLinear Probing
M3AE*-69.3
CLIPGAP61.875.9
CLIPMAX63.777.5
MAE-CLIPGAP57.475.7
MAE-CLIPMAX60.976.6
", + "image_path": "789ccc830566b8b28285a755baa8e461510f24c89b255279abfbd270a313ffd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 332, + 250, + 520, + 329 + ], + "blocks": [ + { + "bbox": [ + 46, + 337, + 288, + 371 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 288, + 371 + ], + "type": "text", + "content": "Table 6. ImageNet classification after pretraining on web-crawled dataset (1.4B images). In the large scale regime, self-supervision does not complement natural language supervision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 250, + 520, + 329 + ], + "lines": [ + { + "bbox": [ + 332, + 250, + 520, + 329 + ], + "spans": [ + { + "bbox": [ + 332, + 250, + 520, + 329 + ], + "type": "table", + "html": "
ModelCLEVRVQAv2GQA
M3AE96.959.953.3
CLIPGAP87.861.855.0
CLIPMAX89.560.653.6
MAE-CLIPGAP92.861.955.3
MAE-CLIPMAX93.961.553.7
", + "image_path": "dbf42315a86d4c3464c93980804567fc537880018ab6aac4cfc45f4358c09232.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 337, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 305, + 337, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 545, + 371 + ], + "type": "text", + "content": "Table 7. VQA finetuning results on web-crawled dataset (1.4B images). Similar to Table 15a, for large pretraining datasets, self-supervision is not improving the downstream VQA performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 386, + 272, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 272, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 272, + 397 + ], + "type": "text", + "content": "size, and use the same learning rate schedule as for " + }, + { + "bbox": [ + 47, + 386, + 272, + 397 + ], + "type": "inline_equation", + "content": "CC" + }, + { + "bbox": [ + 47, + 386, + 272, + 397 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 398, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 495 + ], + "type": "text", + "content": "Model variants: We train M3AE as well as both GAP and MAX variants of CLIP and MAE-CLIP, having noted in the small-scale regime that these are the highest performing architectures and pre-training strategies. We do not train MAE at scale due to resource limitations, and because its performance was significantly lower than the models that utilize natural language supervision—a finding that is consistent with prior literature [31]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 503, + 164, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 164, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 164, + 516 + ], + "type": "text", + "content": "5.2. Image Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "Table 15a provides encoder performance for both zero-shot and linear-probing on ImageNet [19]. We note that at scale, MAE-CLIP typically shows worse performance than CLIP across both pooling strategies and tasks. We hypothesize that this may be due to the fact that masked-patch-prediction results in model capacity being lost to the task of predicting semantically irrelevant patches, meaning that we lose some of the benefits provided by the larger scale dataset (despite tuning the relative loss weights). If we are correct in this hypothesis then it is possible that a larger model capacity would help to alleviate this issue. We leave this to a future investigation. Meanwhile, in Table 14 we once again explore the encoder performance through linear-probing for VTAB [106]. We note that GAP provides a consistent improvement over MAX across both CLIP and MAE-CLIP models, and that M3AE is also a strong performer—as was" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 385, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 468 + ], + "type": "text", + "content": "the case in the small-scale regime. M3AE provides particular benefit to the \"structured\" VTAB tasks, however underperforms (sometimes by a large margin) on the \"natural\" tasks. MAE-CLIP ends up in between the two—doing better than CLIP but worse than M3AE on the structured tasks, and worse than CLIP but better than M3AE on the natural tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 484, + 353, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 484, + 353, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 353, + 496 + ], + "type": "text", + "content": "5.3.VQA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 505, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 613 + ], + "type": "text", + "content": "We follow the same finetuning strategy as before, with results in Table 15b. Notably, we do not see a consistent difference across tasks for CLIP vs MAE-CLIP, with the possible exception of CLEVR [45], where it is clear that M3AE is by far the best performing approach—and its benefits improve the MAE-CLIP results on that task. GAP is perhaps a marginally stronger pooling strategy as measured against MAX, however the difference is smaller than it was in the small-scale regime." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 631, + 443, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 631, + 443, + 644 + ], + "spans": [ + { + "bbox": [ + 306, + 631, + 443, + 644 + ], + "type": "text", + "content": "6. Analysis and Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "While our results present a pessimistic view on combining self-supervision with natural language supervision, we don't have a clear reason as to why self-supervision fails at scale. Here we present two hypotheses that may be explored in future work." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "23438" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 81, + 153, + 222 + ], + "blocks": [ + { + "bbox": [ + 102, + 71, + 132, + 79 + ], + "lines": [ + { + "bbox": [ + 102, + 71, + 132, + 79 + ], + "spans": [ + { + "bbox": [ + 102, + 71, + 132, + 79 + ], + "type": "text", + "content": "(a) CLIP" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 81, + 153, + 222 + ], + "lines": [ + { + "bbox": [ + 82, + 81, + 153, + 222 + ], + "spans": [ + { + "bbox": [ + 82, + 81, + 153, + 222 + ], + "type": "image", + "image_path": "6a1803115410c8dae5d6e5cc382ef40d248a06721407883a4b035ecb0bd303ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 154, + 81, + 225, + 222 + ], + "blocks": [ + { + "bbox": [ + 164, + 71, + 214, + 79 + ], + "lines": [ + { + "bbox": [ + 164, + 71, + 214, + 79 + ], + "spans": [ + { + "bbox": [ + 164, + 71, + 214, + 79 + ], + "type": "text", + "content": "(b) MAE-CLIP" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 81, + 225, + 222 + ], + "lines": [ + { + "bbox": [ + 154, + 81, + 225, + 222 + ], + "spans": [ + { + "bbox": [ + 154, + 81, + 225, + 222 + ], + "type": "image", + "image_path": "6e77dd4128ede60c24b9a7b39c084248839989742632b53f6c4c84779bd1e161.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 81, + 296, + 222 + ], + "blocks": [ + { + "bbox": [ + 240, + 72, + 282, + 80 + ], + "lines": [ + { + "bbox": [ + 240, + 72, + 282, + 80 + ], + "spans": [ + { + "bbox": [ + 240, + 72, + 282, + 80 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 240, + 72, + 282, + 80 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{GAP}}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 81, + 296, + 222 + ], + "lines": [ + { + "bbox": [ + 225, + 81, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 225, + 81, + 296, + 222 + ], + "type": "image", + "image_path": "dc7d15102741e24869e262039d947f9be47c83657f0c2e2e6375acf8fc59b7ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 547, + 265 + ], + "type": "text", + "content": "Figure 2. Zero-shot segmentation masks (top row) and GradCAM [75] visualizations (bottom row) for the prompt \"a photo of a dog\" for CLIP and MAE-CLIP. Self-supervision qualitatively improves the visual grounding of the model, however, the choice of pooling operator has the largest effect, with max pooling producing significantly better results [71]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 298, + 81, + 369, + 222 + ], + "blocks": [ + { + "bbox": [ + 301, + 71, + 364, + 80 + ], + "lines": [ + { + "bbox": [ + 301, + 71, + 364, + 80 + ], + "spans": [ + { + "bbox": [ + 301, + 71, + 364, + 80 + ], + "type": "text", + "content": "(d) MAE-CLIPGAP" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 81, + 369, + 222 + ], + "lines": [ + { + "bbox": [ + 298, + 81, + 369, + 222 + ], + "spans": [ + { + "bbox": [ + 298, + 81, + 369, + 222 + ], + "type": "image", + "image_path": "240f6ae5dfd7ebbdc460c056fe14a4539dd23b1547e82e3e95d74d34f64ab11b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 370, + 81, + 440, + 222 + ], + "blocks": [ + { + "bbox": [ + 383, + 72, + 426, + 79 + ], + "lines": [ + { + "bbox": [ + 383, + 72, + 426, + 79 + ], + "spans": [ + { + "bbox": [ + 383, + 72, + 426, + 79 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 383, + 72, + 426, + 79 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{MAX}}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 81, + 440, + 222 + ], + "lines": [ + { + "bbox": [ + 370, + 81, + 440, + 222 + ], + "spans": [ + { + "bbox": [ + 370, + 81, + 440, + 222 + ], + "type": "image", + "image_path": "e5ec95de2935b4a5d1c5dfacb2cc60c1a3231cc7a3b9121a07a911cf4a9e887b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 441, + 81, + 511, + 222 + ], + "blocks": [ + { + "bbox": [ + 444, + 72, + 507, + 79 + ], + "lines": [ + { + "bbox": [ + 444, + 72, + 507, + 79 + ], + "spans": [ + { + "bbox": [ + 444, + 72, + 507, + 79 + ], + "type": "text", + "content": "(f) MAE-CLIPMAX" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 81, + 511, + 222 + ], + "lines": [ + { + "bbox": [ + 441, + 81, + 511, + 222 + ], + "spans": [ + { + "bbox": [ + 441, + 81, + 511, + 222 + ], + "type": "image", + "image_path": "aa409b56104c878a95724ad919adaa7268c1203d1b6b0ca1179052a8fe931aa2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 277, + 153, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 153, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 153, + 289 + ], + "type": "text", + "content": "6.1. Visual Grounding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 294, + 289, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 289, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 289, + 570 + ], + "type": "text", + "content": "Visual grounding measures how well a particular representation or network can localize objects within an image. It is thought that representations with better visual grounding will excel at more general purpose tasks. In section 4, we found that pooling operator had a larger effect on performance than the addition of MAE. We qualitatively study the effect of both pooling and MAE on the visual grounding of the CLIP image and text encoders. We extract zero-shot segmentation masks and GradCAM [75] visualizations for the prompt \"a photo of a dog\", shown in Figure 2. Details are provided in the supplementary material. We observe that, with respect to visual grounding, the pooling operator plays a much more significant role than self-supervision [71]. In particular, " + }, + { + "bbox": [ + 46, + 294, + 289, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{MAX}}" + }, + { + "bbox": [ + 46, + 294, + 289, + 570 + ], + "type": "text", + "content": " correctly segments the dog in the picture while CLIP focuses only in the background. However, both for CLIP and " + }, + { + "bbox": [ + 46, + 294, + 289, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{MAX}}" + }, + { + "bbox": [ + 46, + 294, + 289, + 570 + ], + "type": "text", + "content": " we observe only an incremental improvement of the localization of the dog in the image when self-supervision is added. Unfortunately, even though the GAP variants of MAE-CLIP and CLIP achieve the best performance on VTAB, they do not also exhibit the most visual grounding. We hope that a more thorough future analysis can elucidate the relationship between self supervision and visual grounding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 575, + 149, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 149, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 149, + 588 + ], + "type": "text", + "content": "6.2. Dataset Diversity" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Alternatively, it may be that self-supervision and natural language supervision excel for entirely different parts of the dataset diversity-size spectrum. Historically, the strongest self-supervised visual-encoder baselines (when compared to supervised methods) include methods such as: SimCLR [11], SimCLRV2 [12] and masked auto encoders [36], which all excel on ImageNet—a somewhat diverse but a relatively \"clean\" and object-centric dataset with 1000 disjoint classes. On less diverse datasets such as Cifar-10 [49], specialized self-supervised methods such as ReMixMatch [5]" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "type": "text", + "content": "have shown excellent performance in the extremely low data regime. Outside of vision, MAEs have shown near state-of-the-art performance on datasets such as Audioset [30,40], a clean 2M example, 632 class audio/video dataset. Recent works which perform controlled experiments scaling self supervised methods to massive datasets [62] show that self-supervised methods achieve numbers roughly " + }, + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "type": "text", + "content": " worse (as measured by ImageNet linear probe) as compared to natural language supervised methods at similar data scale." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 396, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 492 + ], + "type": "text", + "content": "On the other hand, natural language supervised models seem to only show competitive performance on massive and diverse datasets, CLIP [69] was trained on 400M examples, ALIGN [43] was trained on 1.8B examples and BASIC [66] was trained on 6.6B examples. Recently, [28] performs controlled natural language supervision at a smaller scale, showing that natural language supervised models are " + }, + { + "bbox": [ + 304, + 396, + 545, + 492 + ], + "type": "inline_equation", + "content": "5 - 10\\%" + }, + { + "bbox": [ + 304, + 396, + 545, + 492 + ], + "type": "text", + "content": " worse (on ImageNet) than their supervised counterparts." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 492, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 552 + ], + "type": "text", + "content": "While all of these past results only pose circumstantial and not rigorous experimental evidence, we believe it poses a fruitful line of future work on studying the scaling trends of self supervised methods and contrasting them with other forms of supervision." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 562, + 416, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 416, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 416, + 574 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We are grateful to many colleagues for helpful advice and guidance throughout this project. We especially thank Jonathan Shlens, Brandon McKinzie, and Jason Ramuparam for feedback on paper drafts, Ruoming Pang for useful comments and suggestions early in the project, Tom Nickson for his training infrastructure support and Albin Madappally Jose for his assistance in building the training dataset. We also extend thanks to Dr. Nicola Strisciuglio for his clarifying input on earlier incarnations of this work, and Dr. Samuel Albanie for several engaging discussions and helpful pointers to related literature." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "23439" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 91, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 287, + 212 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a Visual Language Model for Few-Shot Learning. Technical Report arXiv:2204.14198, arXiv, Apr. 2022. arXiv:2204.14198 [cs] type: article. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 213, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 213, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 288, + 267 + ], + "type": "text", + "content": "[2] Tarik Arici, Mehmet Saygin Seyfioglu, Tal Neiman, Yi Xu, Son Train, Trishul Chilimbi, Belinda Zeng, and Ismail Tutar. MLIM: Vision-and-Language Model Pre-training with Masked Language and Image Modeling. arXiv:2109.12178 [cs], Sept. 2021. arXiv: 2109.12178. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 268, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 268, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 57, + 268, + 288, + 300 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT Pre-Training of Image Transformers. arXiv:2106.08254 [cs], June 2021. arXiv: 2106.08254. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 301, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 301, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 57, + 301, + 288, + 346 + ], + "type": "text", + "content": "[4] Hangbo Bao, Wenhui Wang, Li Dong, and Furu Wei. VL-BEiT: Generative Vision-Language Pretraining. Technical Report arXiv:2206.01127, arXiv, June 2022. arXiv:2206.01127 [cs] type: article. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 346, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 346, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 57, + 346, + 288, + 401 + ], + "type": "text", + "content": "[5] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. CoRR, abs/1911.09785, 2019. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 402, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 402, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 57, + 402, + 288, + 456 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. DINO: Emerging Properties in Self-Supervised Vision Transformers. arXiv:2104.14294 [cs], May 2021. arXiv: 2104.14294. 2, 16" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 458, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 458, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 458, + 288, + 513 + ], + "type": "text", + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts. Technical Report arXiv:2102.08981, arXiv, Mar. 2021. arXiv:2102.08981 [cs] type: article. 5, 6, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 514, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 288, + 557 + ], + "type": "text", + "content": "[8] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 558, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 288, + 612 + ], + "type": "text", + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. iGPT: Generative Pretraining From Pixels. In Proceedings of the 37th International Conference on Machine Learning, pages 1691-1703. PMLR, Nov. 2020. ISSN: 2640-3498. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 613, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 288, + 669 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. SimCLR: A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597-1607. PMLR, Nov. 2020. ISSN: 2640-3498. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. CoRR, abs/2002.05709, 2020. 1, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "type": "text", + "content": "[12] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. SimCLRv2: Big Self-Supervised Models are Strong Semi-Supervised Learners. arXiv:2006.10029 [cs, stat], Oct. 2020. arXiv: 2006.10029. 2, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 128, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 128, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 128, + 545, + 171 + ], + "type": "text", + "content": "[13] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. MoCov2: Improved Baselines with Momentum Contrastive Learning. arXiv:2003.04297 [cs], Mar. 2020. arXiv: 2003.04297. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 172, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 172, + 545, + 205 + ], + "type": "text", + "content": "[14] Xinlei Chen and Kaiming He. SimSiam: Exploring Simple Siamese Representation Learning. arXiv:2011.10566 [cs], Nov. 2020. arXiv: 2011.10566. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 205, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 205, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 205, + 545, + 249 + ], + "type": "text", + "content": "[15] Xinlei Chen, Saining Xie, and Kaiming He. MoCov3: An Empirical Study of Training Self-Supervised Vision Transformers. arXiv:2104.02057 [cs], Aug. 2021. arXiv: 2104.02057. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 250, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 250, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 545, + 327 + ], + "type": "text", + "content": "[16] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: UNiversal Image-TExt Representation Learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, Lecture Notes in Computer Science, pages 104-120, Cham, Sept. 2020. Springer International Publishing. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 545, + 382 + ], + "type": "text", + "content": "[17] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. Technical Report arXiv:2003.10555, arXiv, Mar. 2020. arXiv:2003.10555 [cs] type: article. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 545, + 426 + ], + "type": "text", + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 427, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 427, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 427, + 545, + 480 + ], + "type": "text", + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, June 2009. ISSN: 1063-6919. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 482, + 545, + 514 + ], + "type": "text", + "content": "[20] Karan Desai and Justin Johnson. VirTex: Learning Visual Representations from Textual Annotations. arXiv:2006.06666 [cs], Sept. 2021. arXiv:2006.06666. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "type": "text", + "content": "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 [cs], May 2019. arXiv: 1810.04805. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 570, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 545, + 646 + ], + "type": "text", + "content": "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. ViT: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv:2010.11929 [cs], June 2021. arXiv: 2010.11929.4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 712 + ], + "type": "text", + "content": "[23] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuo-hang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. METER: An Empirical Study of Training End-to-End Vision-and-Language Transformers. arXiv:2111.02387 [cs], Nov. 2021. arXiv: 2111.02387. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "23440" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "type": "text", + "content": "[24] Jiali Duan, Liquin Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal Alignment using Representation Codebook. arXiv:2203.00048 [cs], Mar. 2022. arXiv: 2203.00048. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 119, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 119, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 119, + 288, + 174 + ], + "type": "text", + "content": "[25] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herve Jegou, and Edouard Grave. SplitMask: Are Large-scale Datasets Necessary for Self-Supervised Pretraining? arXiv:2112.10740 [cs], Dec. 2021. arXiv: 2112.10740. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 176, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 176, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 176, + 288, + 219 + ], + "type": "text", + "content": "[26] Patrick Esser, Robin Rombach, and Björn Ommer. VQGAN: Taming Transformers for High-Resolution Image Synthesis. arXiv:2012.09841 [cs], June 2021. arXiv: 2012.09841.2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 222, + 288, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 222, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 53, + 222, + 288, + 266 + ], + "type": "text", + "content": "[27] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2):303-338, June 2010. 19" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "type": "text", + "content": "[28] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip), 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 316, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 288, + 358 + ], + "type": "text", + "content": "[29] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 361, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 288, + 426 + ], + "type": "text", + "content": "[30] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 430, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 288, + 485 + ], + "type": "text", + "content": "[31] Xinyang Geng, Hao Liu, Lisa Lee, Dale Schuurmans, Sergey Levine, and Pieter Abbeel. M3AE: Multimodal Masked Autoencoders Learn Transferable Representations. Technical Report arXiv:2205.14204, arXiv, May 2022. arXiv:2205.14204 [cs] type: article. 1, 3, 4, 5, 7, 14, 21" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 487, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 520 + ], + "type": "text", + "content": "[32] Aaron Gokaslan and Vanya Cohen. Openwebtext corpus. http://Skylion007.github.io/OpenWebTextCorpus, 2019.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 522, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 288, + 576 + ], + "type": "text", + "content": "[33] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. arXiv:1706.02677 [cs], Apr. 2018. arXiv: 1706.02677. 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 579, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 579, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 579, + 288, + 624 + ], + "type": "text", + "content": "[34] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. VQA2: Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. pages 6904-6913, 2017. 5, 17" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 625, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 288, + 712 + ], + "type": "text", + "content": "[35] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. BYOL: Bootstrap your own latent: A new approach to self-supervised Learning. arXiv:2006.07733 [cs, stat], Sept. 2020. arXiv: 2006.07733. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 106 + ], + "type": "text", + "content": "[36] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. 1, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 150 + ], + "type": "text", + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. MAE: Masked Autoencoders Are Scalable Vision Learners. arXiv:2111.06377 [cs], Nov. 2021. arXiv: 2111.06377. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 194 + ], + "type": "text", + "content": "[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. MoCov1: Momentum Contrast for Unsupervised Visual Representation Learning. pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "type": "text", + "content": "[39] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9068-9077, June 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "type": "text", + "content": "[40] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 312, + 310, + 545, + 354 + ], + "type": "text", + "content": "[41] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing Out of the Box: End-to-End Pre-Training for Vision-Language Representation Learning. pages 12976-12985, Apr. 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 355, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 355, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 312, + 355, + 545, + 397 + ], + "type": "text", + "content": "[42] Drew A. Hudson and Christopher D. Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. pages 6700-6709, 2019. 5, 17" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 399, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 545, + 454 + ], + "type": "text", + "content": "[43] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. CoRR, abs/2102.05918, 2021. 1, 4, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 456, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 545, + 521 + ], + "type": "text", + "content": "[44] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. ALIGN: Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. arXiv:2102.05918 [cs], June 2021. arXiv: 2102.05918.3, 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 523, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 545, + 588 + ], + "type": "text", + "content": "[45] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1988–1997, 2017. 5, 7, 17" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 312, + 590, + 545, + 623 + ], + "type": "text", + "content": "[46] Kaggle and EyePacs. Kaggle diabetic retinopathy detection. https://www.kaggle.com/c/diabetic-retinopathy-detection/data, July 2015. 14" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "type": "text", + "content": "[47] Wonjae Kim, Bokyung Son, and Ildoo Kim. ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision. arXiv:2102.03334 [cs, stat], June 2021. arXiv: 2102.03334. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "text", + "content": "[48] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. Technical Report arXiv:1412.6980, arXiv, Jan. 2017. arXiv:1412.6980 [cs] type: article. 15" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "text", + "content": "23441" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 95 + ], + "type": "text", + "content": "[49] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 173 + ], + "type": "text", + "content": "[50] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. arXiv:1910.13461 [cs, stat], Oct. 2019. arXiv: 1910.13461.2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "type": "text", + "content": "[51] Chenliang Li, Ming Yan, Haiyang Xu, Fuli Luo, Wei Wang, Bin Bi, and Songfang Huang. SemVLP: Vision-Language Pre-training by Aligning Semantics at Multiple Levels. arXiv:2103.07829 [cs], Mar. 2021. arXiv: 2103.07829. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 220, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 274 + ], + "type": "text", + "content": "[52] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. arXiv:2107.07651 [cs], Oct. 2021. arXiv: 2107.07651. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "type": "text", + "content": "[53] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO-2: End-to-End Unified Vision-Language Grounded Learning. arXiv:2203.09067 [cs], Mar. 2022. arXiv: 2203.09067. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 320, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 288, + 408 + ], + "type": "text", + "content": "[54] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, Lecture Notes in Computer Science, pages 121–137, Cham, July 2020. Springer International Publishing. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 410, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 287, + 442 + ], + "type": "text", + "content": "[55] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pretraining via masking, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 287, + 498 + ], + "type": "text", + "content": "[56] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. CoRR, abs/2110.05208, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 287, + 544 + ], + "type": "text", + "content": "[57] Feng Liang, Yangguang Li, and Diana Marculescu. SupMAE: Supervised Masked Autoencoders Are Efficient Vision Learners. Technical Report arXiv:2205.14540, arXiv, May 2022. arXiv:2205.14540 [cs] type: article. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 545, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 287, + 599 + ], + "type": "text", + "content": "[58] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 16, 19" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 287, + 634 + ], + "type": "text", + "content": "[59] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic Gradient Descent with Warm Restarts. arXiv:1608.03983 [cs, math], May 2017. arXiv: 1608.03983. 5, 15" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "type": "text", + "content": "[60] Ilya Loshchilov and Frank Hutter. AdamW: Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs, math], Jan. 2019. arXiv: 1711.05101. 5, 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[61] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. arXiv:1908.02265 [cs], Aug. 2019. arXiv: 1908.02265. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[62] Shlok Mishra, Joshua Robinson, Huiwen Chang, David Jacobs, Aaron Sarna, Aaron Maschinot, and Dilip Krishnan. A simple, efficient and scalable contrastive masked autoencoder for learning visual representations, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "type": "text", + "content": "[63] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. SLIP: Self-supervision meets Language-Image Pre-training. arXiv:2112.12750 [cs], Dec. 2021. arXiv: 2112.12750.1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 194 + ], + "type": "text", + "content": "[64] Norman Mu, Alexander Kirillov, David A. Wagner, and Saining Xie. SLIP: self-supervision meets language-image pre-training. CoRR, abs/2112.12750, 2021. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 545, + 237 + ], + "type": "text", + "content": "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748 [cs, stat], Jan. 2019. arXiv: 1807.03748. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 545, + 282 + ], + "type": "text", + "content": "[66] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. CoRR, abs/2111.10050, 2021. 1, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 284, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 284, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 545, + 337 + ], + "type": "text", + "content": "[67] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k Entities: Collecting Region-to-Phrase Correspondences for Richer Image-to-Sentence Models. pages 2641–2649, 2015. 16" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 338, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 545, + 414 + ], + "type": "text", + "content": "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. CLIP: Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020 [cs], Feb. 2021. arXiv: 2103.00020. 1, 2, 3, 14, 15" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 415, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 545, + 480 + ], + "type": "text", + "content": "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 312, + 482, + 545, + 525 + ], + "type": "text", + "content": "[70] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. GPT-2: Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 526, + 545, + 568 + ], + "type": "text", + "content": "[71] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander T Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models, 2023. 3, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 545, + 613 + ], + "type": "text", + "content": "[72] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating Diverse High-Fidelity Images with VQ-VAE-2. Technical Report arXiv:1906.00446, arXiv, June 2019. arXiv:1906.00446 [cs, stat] type: article. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 615, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 615, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 312, + 615, + 545, + 690 + ], + "type": "text", + "content": "[73] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenna Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "text", + "content": "[74] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23442" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 116 + ], + "type": "text", + "content": "Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION400M: Open Dataset of CLIP-Filtered 400 Million ImageText Pairs. arXiv:2111.02114 [cs], Nov. 2021. arXiv:2111.02114.6, 14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 184 + ], + "type": "text", + "content": "[75] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision, 128(2):336-359, Feb. 2020. arXiv:1610.02391 [cs]. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 54, + 186, + 288, + 229 + ], + "type": "text", + "content": "[76] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 231, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 231, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 231, + 287, + 307 + ], + "type": "text", + "content": "[77] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. GCC: Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 287, + 352 + ], + "type": "text", + "content": "[78] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. CoRR, abs/2112.04482, 2021. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 354, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 407 + ], + "type": "text", + "content": "[79] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A Foundational Language And Vision Alignment Model. arXiv:2112.04482 [cs], Feb. 2022. arXiv: 2112.04482. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 410, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 287, + 473 + ], + "type": "text", + "content": "[80] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross B. Girshick, Piotr Dólar, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. CoRR, abs/2201.08371, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "type": "text", + "content": "[81] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of Generic Visual-Linguistic Representations. arXiv:1908.08530 [cs], Feb. 2020. arXiv: 1908.08530. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 522, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 287, + 565 + ], + "type": "text", + "content": "[82] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 567, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 287, + 621 + ], + "type": "text", + "content": "[83] Yihong Sun, Adam Kortylewski, and Alan Yuille. Amodal Segmentation through Out-of-Task and Out-of-Distribution Generalization with a Bayesian Model. Technical Report arXiv:2010.13175, arXiv, July 2022. arXiv:2010.13175 [cs] type: article. 16" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 624, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 655 + ], + "type": "text", + "content": "[84] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. arXiv:1908.07490 [cs], Dec. 2019. arXiv: 1908.07490. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 287, + 689 + ], + "type": "text", + "content": "[85] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine learning research, 9(11), 2008. 17" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[86] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, " + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\backslash" + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": " Lukasz Kaiser," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "text", + "content": "and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 107, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 172 + ], + "type": "text", + "content": "[87] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning - ICML '08, pages 1096-1103, Helsinki, Finland, 2008. ACM Press. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 545, + 229 + ], + "type": "text", + "content": "[88] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for all vision and vision-language tasks, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "type": "text", + "content": "[89] Xiaolong Wang and Abhinav Gupta. Unsupervised Learning of Visual Representations using Videos. Technical Report arXiv:1505.00687, arXiv, Oct. 2015. arXiv:1505.00687 [cs] type: article. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "type": "text", + "content": "[90] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple Visual Language Model Pretraining with Weak Supervision. arXiv:2108.10904 [cs], Aug. 2021. arXiv: 2108.10904. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 321, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 321, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 312, + 321, + 545, + 363 + ], + "type": "text", + "content": "[91] Longhui Wei, Lingxi Xie, Wengang Zhou, Houqiang Li, and Qi Tian. MVP: Multimodality-guided Visual Pretraining. arXiv:2203.05175 [cs], Mar. 2022. arXiv: 2203.05175. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 312, + 365, + 545, + 420 + ], + "type": "text", + "content": "[92] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 312, + 422, + 545, + 476 + ], + "type": "text", + "content": "[93] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dólar, and Ross Girshick. Early Convolutions Help Transformers See Better. Technical Report arXiv:2106.14881, arXiv, Oct. 2021. arXiv:2106.14881 [cs] type: article. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 478, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 545, + 553 + ], + "type": "text", + "content": "[94] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. Pre-layernorm: On Layer Normalization in the Transformer Architecture. In Proceedings of the 37th International Conference on Machine Learning, pages 10524-10533. PMLR, Nov. 2020. ISSN: 2640-3498. 4, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 556, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 545, + 578 + ], + "type": "text", + "content": "[95] Jianwei Yang, Chunyuan Li, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Focal modulation networks, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 312, + 579, + 545, + 633 + ], + "type": "text", + "content": "[96] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained Interactive Language-Image Pre-Training. arXiv:2111.07783 [cs], Nov. 2021. arXiv: 2111.07783. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 678 + ], + "type": "text", + "content": "[97] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. ERNIE-ViL: Knowledge Enhanced Vision-Language Representations Through Scene Graph. June 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[98] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models, 2022. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23443" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 73, + 286, + 638 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 52, + 73, + 286, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 73, + 286, + 116 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 286, + 116 + ], + "type": "text", + "content": "[99] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive Captioners are Image-Text Foundation Models. arXiv:2205.01917 [cs], May 2022. arXiv: 2205.01917. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 204 + ], + "type": "text", + "content": "[100] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A New Foundation Model for Computer Vision. arXiv:2111.11432 [cs], Nov. 2021. arXiv: 2111.11432. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 207, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 207, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 207, + 286, + 261 + ], + "type": "text", + "content": "[101] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-level Representation Learning for Self-supervised Vision Transformers. Technical Report arXiv:2206.07990, arXiv, June 2022. arXiv:2206.07990 [cs] type: article. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 262, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 262, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 262, + 286, + 304 + ], + "type": "text", + "content": "[102] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. arXiv:2103.03230 [cs, q-bio], June 2021. arXiv: 2103.03230. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 308, + 286, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 286, + 349 + ], + "type": "text", + "content": "[103] Yan Zeng, Xinsong Zhang, and Hang Li. X:VLM: MultiGrained Vision Language Pre-Training: Aligning Texts with Visual Concepts. arXiv:2111.08276 [cs], Feb. 2022. arXiv: 2111.08276. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 353, + 286, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 353, + 286, + 404 + ], + "spans": [ + { + "bbox": [ + 49, + 353, + 286, + 404 + ], + "type": "text", + "content": "[104] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12104-12113, June 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 408, + 286, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 408, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 49, + 408, + 286, + 483 + ], + "type": "text", + "content": "[105] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, André Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, Lucas Beyer, Olivier Bachem, Michael Tschannen, Marcin Michalski, Olivier Bousquet, Sylvain Gelly, and Neil Houlsby. The visual task adaptation benchmark. CoRR, abs/1910.04867, 2019. 1, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 486, + 286, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 486, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 49, + 486, + 286, + 582 + ], + "type": "text", + "content": "[106] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djo-longa, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, Lucas Beyer, Olivier Bachem, Michael Tschannen, Marcin Michalski, Olivier Bousquet, Sylvain Gelly, and Neil Houlsby. VTAB: A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark. arXiv:1910.04867 [cs, stat], Feb. 2020. arXiv: 1910.04867.5,7,14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 586, + 286, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 586, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 49, + 586, + 286, + 638 + ], + "type": "text", + "content": "[107] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. ADE20K: Semantic Understanding of Scenes through the ADE20K Dataset. arXiv:1608.05442 [cs], Oct. 2018. arXiv: 1608.05442. 19" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23444" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_content_list.json b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6f0bad3df2d14c87008fad5f02e0d01d2f5c8db0 --- /dev/null +++ b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_content_list.json @@ -0,0 +1,1326 @@ +[ + { + "type": "text", + "text": "Masked Image Modeling with Local Multi-Scale Reconstruction", + "text_level": 1, + "bbox": [ + 161, + 130, + 808, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoqing Wang $^{1}$ , Yehui Tang $^{1,2}$ , Yunhe Wang $^{2*}$ , Jianyuan Guo $^{2}$ , Zhi-Hong Deng $^{1*}$ , Kai Han $^{2}$ $^{1}$ School of Intelligence Science and Technology, Peking University $^{2}$ Huawei Noah's Ark Lab {wanghaoqing,yhtang,zhdeng} $@$ pku.edu.cn, {yunhe.wang, kai.han} $@$ huawei.com", + "bbox": [ + 117, + 179, + 851, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masked Image Modeling (MIM) achieves outstanding success in self-supervised representation learning. Unfortunately, MIM models typically have huge computational burden and slow learning process, which is an inevitable obstacle for their industrial applications. Although the lower layers play the key role in MIM, existing MIM models conduct reconstruction task only at the top layer of encoder. The lower layers are not explicitly guided and the interaction among their patches is only used for calculating new activations. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to multiple local layers including lower and upper layers. Further, since the multiple layers expect to learn the information of different scales, we design local multiscale reconstruction, where the lower and upper layers reconstruct fine-scale and coarse-scale supervision signals respectively. This design not only accelerates the representation learning process by explicitly guiding multiple layers, but also facilitates multi-scale semantical understanding to the input. Extensive experiments show that with significantly less pre-training burden, our model achieves comparable or better performance on classification, detection and segmentation tasks than existing MIM models. Code is available with both MindSpore and PyTorch.", + "bbox": [ + 76, + 301, + 472, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Masked Image Modeling (MIM) [2, 21, 50] achieves outstanding success in the field of self-supervised visual representation learning, which is inspired by the Masked Language Modeling (MLM) [4, 29] in natural language processing and benefits from the development of vision transformers [16, 33, 45]. MIM learns semantic representations by first masking some parts of the input and then predicting their signals based on the unmasked parts, e.g., normalized pixels [21, 50], discrete tokens [2, 15], HOG feature [47], deep features [1, 54] or frequencies [32, 49].", + "bbox": [ + 75, + 722, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite superior performance on various downstream tasks, these models have huge computational burden and slow learning process [26]. They typically require thousands of GPU Hours for pre-training on ImageNet-1K to get generalizing representations. Since we expect to pretrain these models on more massive amount of unlabeled data (e.g., free Internet data) to obtain more generalizing representations in practice, the pre-training efficiency is an inevitable bottleneck limiting the industrial applications of MIM. How to accelerate the representation learning in MIM is an important topic. To this end, MAE [21] pioneered the asymmetric encoder-decoder strategy, where the costly encoder only operates few visible patches and the lightweight decoder takes all the patches as input for prediction. Further, GreenMIM [26] extends the asymmetric encoder-decoder strategy to hierarchical vision transformers (e.g., Swin [33]). Besides, [8, 19, 30] shrinks the input resolution to lessen the input patches, thereby reducing the computational burden. However, they all aim to accelerate the encoding process rather than the representation learning.", + "bbox": [ + 500, + 270, + 893, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In MIM, the learning of upper layers depends on that of lower ones during pre-training, since the upper-layer features are calculated from the lower layers. Besides, during fine-tuning the upper layers are typically tuned quickly to adapt to the downstream task while the lower ones change more slowly and need to be well-learned [2, 25, 52]. Even fine-tuning only the several upper layers and freezing the others can obtain similar performance [21]. Therefore, the lower layers of encoder play the key role in MIM. However, all existing MIM models only conduct reconstruction task at the top layer of encoder and the lower ones are not explicitly guide, thus the interaction among their patches is only used for calculating the activations of the next layer. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to both lower and upper layers to explicitly guide them and thus accelerate the overall learning process. Using tiny decoder is sufficient for each local reconstruction task and does not significantly increase the computational burden.", + "bbox": [ + 498, + 577, + 893, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How to properly conduct reconstruction tasks at multiple local layers is a non-trivial problem. For example, ap-", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2122", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e09ee8397bd0e0523b8b82cb7af81c8dc4c76d7d07bdc7ec05c2d89699bdf6a0.jpg", + "image_caption": [ + "Figure 1. Top-1 fine-tuning accuracy on ImageNet-1K vs. Pre-training duration. The duration is estimated on a machine with one Tesla V100-32G GPU, CUDA 10.2 and PyTorch 1.8. 'GPU Hours' is the running time on single GPU." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 480, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b956d76e31182bdd429661f85072584ea7fa4bc4f50977fad39967018b60440e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 856, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "plying the top-layer reconstruction task to carefully chosen local layers of ViT [16] can not achieve meaningful improvement. In general, the lower layers exploit low-level information and the upper ones learn high-level information [18, 36], so it is not appropriate to use the supervision signals of same scale for multiple local reconstruction tasks. Here 'scale' is the spatial size of the supervision signals calculated from the divided input regions, e.g., the signals from the $p \\times p$ regions in an input of $H \\times W$ resolution has the scale of $\\frac{H}{p} \\times \\frac{W}{p}$ . The fine-scale and coarse-scale supervisions typically contain low-level and high-level information of the input respectively, and these multi-scale supervisions from input are widely ignored by existing MIM models. To this end, we propose local multi-scale reconstruction where the lower and upper layers reconstruct fine-scale and coarse-scale supervisions respectively. This design not only accelerates the representation learning process, but also facilitates multi-scale semantical understanding to the input. When the decoded predictions have different scale with the supervisions (e.g., on ViT), we use the deconvolution/pool operations to rescale them. We also apply the asymmetric encoder-decoder strategy [21, 26] for quick encoding. Our model, dubbed as LocalMIM, are illustrated in Fig. 2 (a). Overall, we summarize our contributions as follows.", + "bbox": [ + 75, + 310, + 472, + 674 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, this is the first work in MIM to conduct local reconstructions and use multiscale supervisions from the input.", + "- Our model is architecture-agnostic and can be used in both columnar and pyramid architectures.", + "- From extensive experiments, we find that 1) LocalMIM is more efficient than existing MIM models, as shown in Fig. 1 and Table 1. For example, LocalMIM achieves the best MAE result with $3.1 \\times$ acceleration on ViT-B and the best GreenMIM result with $6.4 \\times$ acceleration on Swin-B. 2) In terms of top-1 fine-tuning accuracy on ImageNet-1K, LocalMIM achieves $84.0\\%$ using ViT-B and $84.1\\%$ using Swin" + ], + "bbox": [ + 94, + 683, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "B with significantly less pre-training duration than existing MIM models. The obtained representations also achieve better generalization on detection and segmentation downstream tasks, as shown in Table 2 and 3.", + "bbox": [ + 529, + 310, + 890, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 382, + 650, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Image Modeling. With the development of vision transformers [16, 20, 33, 42, 45], Masked Image Modeling (MIM) gradually replaces the dominant position of contrastive learning [10, 22, 44] in visual self-supervised representation learning due to its superior fine-tuning performance in various visual downstream tasks. Many target signals have been designed for the mask-prediction pretext task in MIM, such as normalized pixels [21, 50], discrete tokens [2, 15], HOG feature [47], deep features [1, 54] or frequencies [32, 49]. However, they are all only applied as single-scale supervisions for reconstruction. An inevitable bottleneck for the industrial applications of MIM is that these models typically require huge computational resources and long pre-training duration. To this end, some works accelerate the encoding process via the asymmetric encoder-decoder strategy [21, 26] or lessening the input patches [8, 30]. Only accelerating the encoding process sometimes doesn't really speed up the representation learning, like GreenMIM vs. SimMIM192 in Fig. 1. In this work, we use local multi-scale reconstructions to explicitly guide multiple lower layers and thus accelerate the overall learning process. Our method is compatible with the above quick encoding approaches. ConvMAE [18] fuses the feature of local layers for the final reconstruction to explicitly guide them, but still applies single-scale supervision signals.", + "bbox": [ + 496, + 412, + 890, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Locally supervised learning. Considering the biological brain learns mainly based on local information [6], some works [34, 35, 37, 46] used local error signals to greedily optimize individual blocks of the backbone without global back-propagation. This greedy training procedure significantly reduces memory overhead, but can not accelerate model learning. In this work, we also optimize error at", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2123", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d688c406129a1b7f8e1cba475142c9340370cd1956635d60d515f2f0fec09a86.jpg", + "image_caption": [ + "Figure 2. Illustration of local multi-scale reconstruction. a) We randomly mask some input patches and then predict their supervision signals of different scales at multiple local layers. The multi-scale supervisions are obtained by first dividing the input under different scales and then extracting signals with some feature descriptors. The lower layers are responsible for fine-scale reconstruction and the upper ones are responsible for coarse-scale reconstruction. We also use the asymmetric encoder-decoder strategy for quick encoding. b) The decoder consists of three parts: Transformer blocks, Deconvolution/Pool (optional) and Multilayer perceptron." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "multiple local layers but still use global back-propagation for end-to-end training. Compared with existing MIM models, our local reconstruction is more biologically plausible since the human brain prefers local learning rules [3, 13, 14]. Feature distillation [24, 39] also explicitly guides the local layers, but needs costly pre-trained or momentum teacher. In fact, the multi-scale supervisions from original input are sufficient to guide local layers and also readily available.", + "bbox": [ + 75, + 383, + 472, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-scale property. Biological visual perception is hierarchical and multi-scale [28]. Moreover, multi-scale features are also useful for many visual tasks [9,23,38]. To this end, many advanced vision transformers [17,33,43,45,51] hard-code the multi-scale property to the architectures and boost the performance. However, global single-scale reconstruction is not enough for guiding multiple local layers to learn multi-scale information. In this work, beyond the multi-scale features, we further introduce multi-scale supervisions to soft-code this property. This new design is not limited to specific architectures.", + "bbox": [ + 75, + 508, + 472, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Model", + "text_level": 1, + "bbox": [ + 76, + 691, + 156, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Masked Image Modeling", + "text_level": 1, + "bbox": [ + 76, + 718, + 305, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In MIM, the models first mask some parts of the input image and then predict their information based on the un-masked observation. For quick encoding, MAE [21] and GreenMIM [26] use the asymmetric encoder-decoder strategy. The major designs are described below.", + "bbox": [ + 75, + 742, + 468, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image presentation. Generally, we present an image to the sequence of visual patches as the input to vision transformers [16, 33]. The input image $x \\in \\mathbb{R}^{H \\times W \\times C}$ is reshaped to $N = HW / p^2$ non-overlapping patches $x^p \\in \\mathbb{R}^{N \\times (p^2C)}$ , where $p$ is the patch size, $(H,W)$ is the resolution of in", + "bbox": [ + 75, + 823, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "put image and $C$ is the number of channels. The patches $\\{x_i^p\\}_{i = 1}^N$ are then linearly mapped to the patch embeddings. For retaining the positional information, the patches are typically added with positional embeddings.", + "bbox": [ + 498, + 383, + 892, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision transformers. Unlike the grid operations in the convolutional neural networks, vision transformers [16,33,45] use the stacked multi-head self-attention modules. To obtain the multi-scale feature maps for dense prediction tasks, some advanced vision transformers [33, 45] extend the columnar ViT [16] to the pyramidal structure, where the feature maps change from fine-scale to coarse-scale and are expected to capture the multi-scale information from the input. The patches $[x_1^p,\\dots ,x_N^p ]$ are fed to a vision transformer with $L$ layers and obtain the output feature at $l$ -th layer $z^{l} = [z_{1}^{l},\\dots ,z_{N_{l}}^{l}],l = 1,\\dots ,L.$", + "bbox": [ + 496, + 454, + 892, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masking. Given the patch sequence $\\{x_i^p\\}_{i = 1}^N$ , MIM constructs a random mask $m\\in \\{0,1\\} ^N$ to indicate the masked patches that correspond to $m_{i} = 1$ . There are two main masking strategies, random masking [21] and block-wise masking [2]. Specially, since the number of patches $N_{l}$ decreases through layers in the pyramidal architectures, some works [18,26] first construct the random mask $m^L$ with the size of $N_{L}$ , and then up-sample it to size $N$ .", + "bbox": [ + 496, + 632, + 893, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder. Only the visible patches $x^{v} = \\{x_{i}^{p}|m_{i} = 0\\}$ are fed to the encoder and mapped to the latent features $z_{v}^{l}, l = 1,\\dots ,L$ , which significantly reduces compute and memory. Specially, since the local window attention in the pyramidal architectures is not compatible with the incomplete input patches (i.e., only visible patches), GreenMIM [26] proposes the Optimal Grouping algorithm and the Group Window Attention scheme. After pre-training, the encoder is used in various downstream tasks.", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2124", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoder. The decoder takes as the input of both encoded visible patches $z_v^L$ and mask tokens $\\{e_{[\\mathbb{M}]}|m_i^L = 1\\}$ , where the mask token $e_{[\\mathbb{M}]}$ is a shared and learnable vector. The positional embeddings are also added to them for providing the location information. Since the decoder is only used during pre-training to output the prediction $\\hat{y}^L$ , it can be any architectures which support the global information propagation among patches, e.g., a series of ViT or Swin blocks. Many works [21, 26, 30] show that using lightweight decoders is sufficient to learn generalizing representations.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global reconstruction. All existing MIM models predict the supervision signal $y$ based on the final output feature $z_v^L$ of the encoder and minimize a global reconstruction loss", + "bbox": [ + 75, + 246, + 468, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M I M} = - \\sum_ {i = 1} ^ {N _ {L}} m _ {i} ^ {L} \\cdot \\ln P \\left(y _ {i} \\mid \\hat {y} _ {i} ^ {L}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 301, + 468, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the loss is calculated on the masked patches with $m_{i}^{L} = 1$ , and $P(\\cdot |\\cdot)$ can be the Gaussian or Dirichlet distributions for regression or classification losses respectively.", + "bbox": [ + 76, + 354, + 468, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Analysis and motivation", + "text_level": 1, + "bbox": [ + 76, + 409, + 299, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For pre-training, the upper-layer features are calculated from the lower layers, so the well-learned lower layers can propagate semantical knowledge to the upper ones and facilitate their learning. For fine-tuning, the upper layers typically adapt quickly to specific downstream tasks, while the lower ones change more slowly and need to be sufficient learned during pre-training [2,25,52]. Even fine-tuning only the several upper layers and freezing the others can obtain comparable performance [21]. Therefore, the lower layers of the encoder play the key role in MIM.", + "bbox": [ + 75, + 431, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After patchification and linearly projection, the initial patch embeddings lose the inter-patch semantical relations. The self-attention mechanism in vision transformers is responsible for learning these relations by inter-patch interaction and build a better representation space than pixel space [5]. Further, since the self-attention mechanism has the computational complexity with a quadratic dependence on patch number $N$ , it is difficult to learn the inter-patch interactions, especially for the lower layers of pyramidal architectures where the small patch size $p$ leads to huge $N$ . However, under the global reconstruction loss, the interpatch interaction at lower layers is not explicitly guided, and the simple task of calculating new activations is not sufficient to guide it. As a result, it is hard for the patches at lower layers to learn inter-patch relations. Concretely, we examine Normalized Mutual Information (NMI) [41] between query and key patches at each layer, where the high NMI value means the attention maps strongly depend on the query patches. Intuitively, the semantical representations should have highly query-adaptive attentions, i.e., the different query patches faithfully attend to their semantically", + "bbox": [ + 75, + 584, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/05f26d05ff73645e0850ba2457b7c48d9276b00002b2710f815a495d27815808.jpg", + "image_caption": [ + "Figure 3. Normalized Mutual Information (NMI) between query and key patches at each layer of a pre-trained ViT-B." + ], + "image_footnote": [], + "bbox": [ + 529, + 90, + 862, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "related regions. This is an advantage of the self-attention mechanism. As shown in Fig. 3, existing MIM models with global loss have small NMI value at lower layers, which means their patches have less query-adaptive attentions.", + "bbox": [ + 496, + 316, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reconstruction task requires holistic reasoning among patches to predict the masked signals and thus obtains semantic understanding to the input. Since this challenging task facilitates non-trivial inter-patch interactions, we apply it at multiple local layers, including both lower and upper ones to explicitly guide them all.", + "bbox": [ + 496, + 378, + 890, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Local multi-scale reconstruction", + "text_level": 1, + "bbox": [ + 500, + 479, + 782, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In MIM, the supervision signals for a reconstruction task are directly calculated from the input. Concretely, we evenly divide the input image $x \\in \\mathbb{R}^{H \\times W \\times C}$ into nonoverlapping regions $\\{x_i \\in \\mathbb{R}^{p \\times p \\times C}\\}_{i=1}^{HW/p^2}$ and use some feature descriptor $\\pi$ to extract the supervision signal $y_i = \\pi(x_i)$ . To learn generalizing representations, many feature descriptors have been designed, such as pixel normalization [21, 50], HOG [47] and the pre-trained or momentum teacher [1,2,15,54]. We define the scale of supervision $y$ as its spatial size $\\frac{H}{p} \\times \\frac{W}{p}$ . For a given input, the fine-scale supervisions from finely-divided input regions typically contain the low-level semantical information of the input, like corners, edges or textures. Relatively, the coarse-scale supervisions capture high-level semantical information of the input, like the shape of partial or whole object. Intuitively, multi-scale supervisions can guide representation learning better than the common single-scale ones due to their richer semantic information. In this work, we mainly consider the feature descriptors which are readily available without extra pre-training burden and costly forward inference of teacher networks, e.g., pixel normalization and HOG.", + "bbox": [ + 496, + 503, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 4f, directly applying the top-layer reconstruction task at carefully chosen local layers of ViT can not achieve meaningful improvement, where each local task predicts the supervisions of same scale. In fact, the lower and upper layers expect to learning low-level and high-level", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2125", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/79b79e9cfc6aefbf76aae0f564ac6d7d1ac7389ec37819036ab325042d57806e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBackbone# ParamsPT EpochGPU Hours/Ep.Total GPU HoursAcc
Scratch, ViTViT-B86M01.5-82.3
Scratch, SwinSwin-B88M02.4-83.5
MoCo v3 [12]ViT-B86M600--83.2
DINO [7]ViT-B86M300--82.8
BEiT [2]ViT-B86M8002.4192083.2
iBOT [54]ViT-B86M40010.1404083.8
MAE [21]ViT-B86M8001.188083.3
MAE [21]ViT-B86M16001.1176083.6
MAE [21]ViT-L307M16001.7272085.9
MaskFeat [47]ViT-B86M16003.9624084.0
CAE [11]ViT-B86M8002.8224083.6
\\( LoMaR^† \\) [8]ViT-B86M16001.4224084.1
data2Vec† [1]ViT-B86M8003.0240084.2
PeCo [15]ViT-B86M800--84.5
LocalMIM-HOGViT-B86M1000.77083.3
LocalMIM-HOGViT-B86M16000.7112084.0
LocalMIM-HOGViT-L307M8001.080085.8
\\( SimMIM_{192} \\) [50]Swin-B88M8001.8144084.0
\\( SimMIM_{192} \\) [50]Swin-L197M8003.0240085.4
GreenMIM [26]Swin-B88M8000.864083.7
GreenMIM [26]Swin-L197M8001.4112085.1
LocalMIM-PixelSwin-B88M1001.010083.7
LocalMIM-HOGSwin-B88M1001.111083.8
LocalMIM-PixelSwin-B88M4001.040084.0
LocalMIM-HOGSwin-B88M4001.144084.1
LocalMIM-HOGSwin-L197M8001.6128085.6
", + "bbox": [ + 122, + 88, + 848, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Top-1 fine-tuning accuracy on ImageNet-1K. All models are pre-trained and fine-tuned under $224 \\times 224$ resolution except that $\\mathrm{SimMIM}_{192}$ uses $192 \\times 192$ resolution for pre-training. $\\dagger$ means using the relative positional encoding.", + "bbox": [ + 75, + 523, + 893, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "information respectively [18, 36], so it is not appropriate to use the single-scale supervisions to guide multiple local layers, even for the columnar architectures with feature maps of the same scale at all layers. To this end, we make the lower layers to reconstruct fine-scale supervisions and the upper ones to reconstruct coarse-scale supervisions. Specially, for the pyramidal architectures which already hardcode the multi-scale property to the features by setting their spatial sizes, we use the supervisions with same scale as the feature maps at the chosen layers for compatibility.", + "bbox": [ + 75, + 568, + 470, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The decoding process under a specific scale is illustrated in Fig. 2 (b). The decoder consists of three parts: Transformer blocks for reasoning, (optional) Deconvolution/Pool for rescaling and Multilayer perceptron for prediction. Concretely, based on the encoded visible patches $z_{v}^{l}$ from $l$ -th layer and the mask tokens $\\{e_{[\\mathbb{M}]}^{l}|m_{i}^{l} = 1\\}$ , a decoder outputs prediction $\\hat{y}^{l}$ that has the same scale as feature map $z_{v}^{l}$ . When the supervision $y^{l}$ has different scale with the feature map $z_{v}^{l}$ (e.g., in the columnar architectures), the decoded prediction can not match the supervision. To this end, we use the deconvolution/pool operations to rescale", + "bbox": [ + 75, + 729, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the prediction $\\hat{y}^l$ for matching the supervision $y^{l}$ . For example, we can rescale the $14\\times 14$ prediction to the scale of $56\\times 56$ with twice deconvolution operations or to the scale of $7\\times 7$ with average pool. To avoid excessive computational overhead, we use tiny decoders containing one Transformer block with small embedding dimension.", + "bbox": [ + 496, + 568, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training loss is the weighted summation of the reconstruction losses at the chosen layers", + "bbox": [ + 496, + 659, + 890, + 690 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {L o c a l M I M}} = - \\sum_ {l \\in \\mathcal {I}} w _ {l} \\cdot \\sum_ {i = 1} ^ {N _ {l}} m _ {i} ^ {l} \\cdot \\ln P \\left(y _ {i} ^ {l} \\mid \\hat {y} _ {i} ^ {l}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 698, + 890, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ is the set of chosen layers, $w_{l}$ is the coefficient of each local loss, and mask $m^l$ is calculated by up/down sampling the initial mask $m$ . These local losses guide the patches at multiple chosen layers to conduct semantic interactions under different scales, which not only accelerates the learning of multiple layers but also facilitates multiscale semantical understanding to the input. As shown in Fig. 3, compared with existing models, our LocalMIM has larger NMI values at lower layers, which means the attention maps depend more strongly on the query patches.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2126", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/695ef56dee092594a25538672d8a9e4edbc3118c47ba8c7d6e8ca0be332a808c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPT EpochPT HoursmIoU
Supervised--47.4
MoCo v3 [12]300-47.3
BEiT [2]800192047.1
MAE [21]1600176048.1
MaskFeat [47]1600624048.8
PeCo [15]800-48.5
CAE [11]800224048.8
LocalMIM-HOG1600112049.5
", + "bbox": [ + 86, + 88, + 459, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 311, + 209, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first evaluate our LocalMIM model on classification, detection and segmentation tasks, and then provide some ablation studies for deep understanding.", + "bbox": [ + 76, + 338, + 468, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Classification on ImageNet-1K", + "text_level": 1, + "bbox": [ + 76, + 397, + 346, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. Both pre-training and fine-tuning are conducted on ImageNet-1K [40] dataset under the $224 \\times 224$ resolution. We mainly examine two representative architectures, columnar ViT [16] and pyramidal Swin [33]. The input images are patchified with patch size $p = 16$ for ViT and $p = 4$ for Swin, and the obtained patches are randomly masked with ratio $r = 0.75$ by default. We use both HOG feature [47] and normalized pixels [21] as the supervision signals. For Swin, we introduce the reconstruction task after each stage and the supervision signals have the same scale with the output feature maps, e.g., the chosen layers $\\mathcal{I} = \\{2,4,22,24\\}$ and the scales of supervision are $\\{28^2,14^2,7^2,7^2\\}$ on Swin-B. Each decoder contains one transformer block with the embedding dimension of 128 and 4 attention heads. For ViT, the chosen layers $\\mathcal{I} = \\{2,4,4 + n,6 + n\\}$ , e.g., $n = 6$ on ViT-B and $n = 18$ on ViT-L. The scales of supervision are $\\{56^2,28^2,14^2,7^2\\}$ . Each decoder contains one transformer block with the embedding dimension of 256 and 8 attention heads. The pretraining and fine-tuning schedules mostly follow [21, 26], and more detailed settings can be found in the Appendix.", + "bbox": [ + 75, + 426, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We compare our LocalMIM with existing MIM models and examine both pre-training efficiency and top-1 fine-tuning accuracy on ImageNet-1K. The results are provided in Fig. 1 and Table 1. For fair comparison, we estimate the pre-training duration of each model on the same machine with one Tesla V100-32G GPU, CUDA 10.2 and Pytorch 1.8. We report the running time on single GPU, denoted as 'GPU Hours', see the Appendix for more details. On ViT-B, LocalMIM achieves the best results of MAE [21] and MaskFeat [47] with $3.1 \\times$ and $5.6 \\times$ accel", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b580ab01b5ea958f9b892c47a8a841babd98329019b6c55bce87185cabde3785.jpg", + "table_caption": [ + "Table 2. Semantic segmentation on ADE20K using UperNet with ViT-B backbone. Our LocalMIM achieves better results than previous MIM models with less pre-training duration." + ], + "table_footnote": [], + "table_body": "
ModelPT EpochPT HoursAPbAPm
Supervised30084048.543.2
SimMIM192 [50]800144050.444.4
GreenMIM [26]80064050.044.1
LocalMIM-HOG40044050.744.9
", + "bbox": [ + 504, + 88, + 879, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Object detection and instance segmentation on COCO. We fine-tune Mask R-CNN end-to-end with Swin-B backbone.", + "bbox": [ + 500, + 181, + 890, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "eration respectively. On Swin-B, LocalMIM achieves those of $\\mathrm{SimMIM}_{192}$ [50] and GreenMIM [26] with $3.6\\times$ and $6.4\\times$ acceleration respectively. In terms of final top-1 finetuning accuracy, LocalMIM achieves $84.0\\%$ with ViT-B and $84.1\\%$ with Swin-B. Compared with previous best results, LocalMIM achieves comparable performance with significantly less pre-training duration. Note that PeCo [15] uses more advanced feature descriptor, a pre-trained codebook, which introduces additional pre-training burden. Besides, on Swin-B, with 100 epochs pre-training and 100 epochs supervised fine-tuning LocalMIM achieves $83.8\\%$ top-1 accuracy and takes about 350 GPU Hours, while 300 epochs supervised training from scratch only achieves $83.5\\%$ top-1 accuracy and takes about 720 GPU Hours. This means even for the high-level classification task, self-supervision learning is more efficient and effective than supervised learning. Similar results can also be found on ViT.", + "bbox": [ + 496, + 224, + 890, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Downstream tasks", + "text_level": 1, + "bbox": [ + 500, + 491, + 678, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We transfer our pre-trained backbone to semantic segmentation on ADE20K [53] and object detection and segmentation on COCO [31].", + "bbox": [ + 500, + 513, + 890, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic segmentation on ADE20K. We conduct semantic segmentation on ADE20K using UperNet [48] and following the code in [2, 21]. See the Appendix for fine-tuning details. The results are shown in Table 2. With significantly less computation burden, LocalMIM outperforms the state-of-the-art result by 0.7. Besides, LocalMIM has the same top-1 fine-tuning accuracy with MaskFeat [47] on ImageNet-1K, but has better semantic segmentation performance, which means our local multi-scale reconstruction facilitates the learning of multi-scale semantic knowledge.", + "bbox": [ + 496, + 564, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object detection and instance segmentation on COCO. We fine-tune Mask R-CNN [23] on COCO with Swin-B backbone. Following [26], we also use the code base and $3 \\times$ fine-tuning schedule from the supervised Swin, where the model is fine-tuned for 36 epochs. See the Appendix for fine-tuning details. We report box AP $(\\mathrm{AP}^b)$ for object detection and mask AP $(\\mathrm{AP}^m)$ for instance segmentation. The results are shown in Table 3. With no labeling burden, our LocalMIM outperforms supervised pre-training by 2.2 $\\mathrm{AP}^b$ and $1.7\\mathrm{AP}^m$ . Compared to SimMIM192 and GreenMIM, our LocalMIM achieves $+0.3$ and $+0.7$ gain respectively for detection, and $+0.5$ and $+0.8$ gain for segmenta", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2127", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f3891c1c327ff13db4d7b0f2a0f40b53b7c21a2f2b4ade20e2905190987a68b2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
targetViT-BSwin-B
baselines82.983.2
Pixels83.083.7
HOG83.383.8
", + "bbox": [ + 99, + 88, + 313, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6f66e962161f3c0241c40bef955f5ea06778852e4f8a674d37b48be77f99e76d.jpg", + "table_caption": [ + "(a) Reconstruction target. HOG is more suitable as multi-scale supervisions than pixels." + ], + "table_footnote": [], + "table_body": "
locationsbackboneacc
GreenMIMSwin-B83.2
[24]83.3
[22, 24]83.4
[4, 22, 24]Swin-B83.6
[2, 4, 22, 24]83.8
fusion83.5
", + "bbox": [ + 86, + 181, + 326, + 271 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6fc144d7ada750056c732467376291f12cfd13a91236d854fba41da5fb2dbb19.jpg", + "table_caption": [ + "(d) Locations for Swin. Applying reconstructions at all stages mostly accelerates learning." + ], + "table_footnote": [], + "table_body": "
ratioViT-BSwin-B
0.4083.283.7
0.6083.183.7
0.7583.383.8
0.9083.083.5
", + "bbox": [ + 377, + 88, + 596, + 152 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1a6da0248cb28dc40a1e5625e82ef0209d07c661cd8edeeddb7f5821b4d0d6f7.jpg", + "table_caption": [ + "(b) Mask ratio. Masking $75\\%$ patches works the best for both ViT and Swin." + ], + "table_footnote": [], + "table_body": "
locationsbackboneacc
MAEViT-B82.9
[12]83.0
[3, 6, 9, 12]ViT-B83.1
[2, 4, 10, 12]83.3
[1, 2, 11, 12]82.9
", + "bbox": [ + 369, + 183, + 599, + 262 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1b9aa3a189b499410275ac870bab570f76c8d02528096849f2877bf4414622d5.jpg", + "table_caption": [ + "(e) Locations for ViT. The chosen layers for local losses obviously affect performance." + ], + "table_footnote": [], + "table_body": "
decoderViT-BSwin-B
512D - 16H83.3 (1.0h)83.8 (1.6h)
256D - 8H83.3 (0.7h)83.7 (1.3h)
128D - 4H83.2 (0.6h)83.8 (1.1h)
", + "bbox": [ + 650, + 88, + 879, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e7aba84e904cd8f4edb6f4b901c41a72489ac3728ffbff97f111cd5ab2267255.jpg", + "table_caption": [ + "(c) Decoder design. A tiny decoder performs as well as the larger one but is more efficient." + ], + "table_footnote": [], + "table_body": "
scalesbackboneacc
MAEViT-B82.9
[142, 142, 142, 142]83.0
[282, 142, 72, 72]ViT-B83.2
[562, 282, 142, 72]83.3
[72, 142, 282, 562]83.1
", + "bbox": [ + 645, + 183, + 885, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(f) Scales for ViT. Using the supervisions from fine-scale to coarse-scale works the best.", + "bbox": [ + 640, + 275, + 885, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation studies with ViT-B and Swin-B on ImageNet-1K. We report the top-1 fine-tuning accuracy (%) and the default setting is: the reconstruction target is HOG feature, the mask ratio is $75\\%$ , the decoder contains one Transformer block with dimension 256 and 8 heads for ViT and dimension 128 and 4 heads for Swin, the local reconstructions are applied at all stages of Swin and [2, 4, 10, 12]-th layer of ViT, the scales of supervision are $[28^2, 14^2, 7^2, 7^2]$ for Swin and $[56^2, 28^2, 14^2, 7^2]$ for ViT, and the pre-training length is 100 epochs.", + "bbox": [ + 75, + 313, + 893, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion with significantly less pre-training duration.", + "bbox": [ + 76, + 390, + 395, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 419, + 236, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We ablate our LocalMIM using the default setting in Table 4. MAE [21] and GreenMIM [26] are our main baselines on ViT and Swin respectively, since we all adopt the asymmetric encoder-decoder strategy.", + "bbox": [ + 75, + 444, + 468, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction target. We examine two types of signal as supervisions: HOG feature [47] and normalized pixel [21], and the results are shown in Table 4a. HOG feature is more suitable as multi-scale supervisions than normalized pixel on both ViT and Swin, since it contains more refined semantic information. Compared to the columnar ViT with single-scale feature maps, our multi-scale supervisions achieve more improvement on the pyramidal Swin. Since we need to rescale the predictions in the columnar architectures, it is harder to guide the learning of multi-scale information, e.g., we need more information-refined signals (HOG) and obtain relatively less improvement.", + "bbox": [ + 75, + 511, + 468, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask ratio. We explore different mask ratios, ranging from 0.4 to 0.9, and the results are provided in Table 4b. Similar to [21, 26], our LocalMIM is robust to the mask ratio from 0.4 to 0.75, and excessively large one (0.9) harms the performance since it over-complicates the reconstruction tasks.", + "bbox": [ + 75, + 698, + 468, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decoder design. As shown in [21, 26], a small single-block decoder can achieve the similar fine-tuning accuracy to the heavier one with multiple blocks. However, since we need to use the decoder for each reconstruction task to process the signals of specified scale, their decoder which has the embedding dimension of 512 and 16 attention heads is still heavy. To this end, we explore whether a smaller decoder can still handle the local reconstruction tasks for", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "learning good representations. One Transformer block is the minimal requirement to propagate information from visible patches to mask tokens, so we still use one Transformer block but narrow the embedding dimension and reduce the number of attention heads. The results (top-1 fine-tuning accuracy and GPU Hours per epoch) are shown in Table 4c, where $dD - hH'$ represents the decoder with the embedding dimension of $d$ and $h$ attention heads. Surprisingly, a tiny decoder can achieve the same even better fine-tuning performance, but is significantly more efficient.", + "bbox": [ + 496, + 388, + 890, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Locations for local reconstructions. An important design of our LocalMIM is the locations in the encoder where we introduce local reconstructions. Many pyramidal architectures like Swin already divide the entire backbone into multiple stages, so we introduce multi-scale reconstructions at their output locations. From the last stage, we gradually include the lower stages and the results are shown in Table 4d. Note that the 2-th, 4-th, 22-th and 24-th layer correspond to the output location of the four stages respectively. We also provide the result of fusion method [18] which fuses the output features from each stage for reconstruction. As we can see, introducing reconstructions at local stages can accelerate representation learning, and applying to all stages works the best. Our LocalMIM also outperforms the fusion method which only uses the single-scale reconstruction, and this verifies the importance of multi-scale reconstructions. Further, for the columnar ViT, we follow the experience from Swin and also select four local layers. Concretely, we consider the uniform division [3, 6, 9, 12], Swin-style division [2, 4, 10, 12] and a extreme division [1, 2, 11, 12]. The results are shown in Table 4e. The performance on columnar ViT is sensitive to the chosen local layers, and the Swin-style division $[2, 4, 4 + n, 6 + n]$ works the best.", + "bbox": [ + 496, + 553, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2128", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/384b5deb4a41f29823644f672ba46519d5b558711e78c2b84a9710e7e718db81.jpg", + "image_caption": [ + "Figure 4. Visualization of the attention maps for different query points, marked with red boxes. LocalMIM can distinguish semantical regions using self-attention mechanism." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 153, + 448 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe658e280a009b0a448b6b52f5b8ce8eed792aa4fd48e4011b46ea451eb91f9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 88, + 468, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scales of supervisions. Since the pyramidal architectures hard-code the multi-scale property of the feature maps, we directly use the supervisions of same scale as the feature maps for compatibility, i.e., we use $[28^2, 14^2, 7^2, 7^2]$ for Swin. For columnar ViT, we consider single-scale supervisions $[14^2, 14^2, 14^2, 14^2]$ , fine-to-coarse multi-scale supervisions $[28^2, 14^2, 7^2, 7^2]$ and $[56^2, 28^2, 14^2, 7^2]$ , and coarse-to-fine multi-scale supervisions $[7^2, 14^2, 28^2, 56^2]$ . The results are shown in Table 4f. The multi-scale supervisions achieve better fine-tuning accuracy than the single-scale ones, which means the multiple local layers require the supervision of different scales. Further, the fine-to-coarse supervisions outperform the coarse-to-fine ones, which is consistent with the design priors for visual architectures [33] and biological visual processing [28].", + "bbox": [ + 75, + 517, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of self-attention mechanism. To qualitatively illustrate the behavior of self-attention mechanism in LocalMIM, we visualize the attention maps for different query positions within an image, as shown in Fig. 4. We use each chosen patch as query and show the patches it attends to. We select the attention maps at 2-th, 4-th, 10-th and 12-th layers of a pre-trained ViT-B/16 backbone to show their changes during forward inference. For object-centric images, LocalMIM can distinguish the foreground object from the background. For more complex multi-object images,", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c95eb8b1e964061812a21bfe0271e8721fe0a7e55771c84b24ce798ccb5b8cac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modelbackboneGPU Hours/Ep.acc
LocalMIMViT-B0.783.3
w/ isolated grad0.783.0
LocalMIMSwin-B1.183.8
w/ isolated grad1.183.7
", + "bbox": [ + 506, + 88, + 887, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Training LocalMIM with isolated gradients achieves similar performance with global back-propagation.", + "bbox": [ + 498, + 178, + 890, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LocalMIM can effectively separate different objects without any task-specific supervision, which means the attention maps are query-adaptive. On the other hand, the patches at lower layers typically more focus on their neighboring regions and capture low-level information, while those at upper layers attend to a wide range of semantically related regions and capture high-level shape information.", + "bbox": [ + 496, + 224, + 890, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gradient-isolated training. Inspired by locally supervised learning [34, 35], we remove global back-propagation and stop the gradients after each chosen local layer used for reconstruction. The entire backbone is thus divided into multiple gradient-isolated parts. The results are shown in Table 5. Surprisingly, the gradient-isolated training achieves similar performance to global back-propagation, which further verifies the effectiveness of our multi-scale reconstructions for guiding the local layers. It even requires no gradient information from the upper layers. This observation also shows the promise of our local multi-scale reconstruction for the decoupled training of neural networks, which allows for training very deep networks without memory concern and reduces explosive or vanishing gradients.", + "bbox": [ + 496, + 335, + 892, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 566, + 625, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel and efficient pretext task, local multi-scale reconstruction, where the lower layers and upper layers reconstruct the fine-scale and coarse-scale supervisions from the input respectively. We obtain the multi-scale supervisions by first dividing the input under different scales and then extracting supervisions with appropriate feature descriptors. Our model needs no extra pre-trained codebook and no costly forward inference of teacher networks during pre-training. The asymmetric encoder-decoder strategy with tiny encoders also have small computational burden, our model thus can be trained quickly. The novel pretext task further accelerates the representation learning, especially for the pyramidal architectures. All these designs allow our model to achieve comparable performance to existing models but with significantly less pre-training burden. The gradient-isolated training also verifies the effectiveness of our novel task design in guiding the local layers.", + "bbox": [ + 496, + 592, + 890, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. We gratefully acknowledge the support of MindSpore [27], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2129", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022. 1, 2, 4, 5", + "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2022. 1, 2, 3, 4, 5, 6", + "[3] Yoshua Bengio, Dong-Hyun Lee, Jorg Bornschein, Thomas Mesnard, and Zhouhan Lin. Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156, 2015. 3", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[5] Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022.4", + "[6] Natalia Caporale, Yang Dan, et al. Spike timing-dependent plasticity: a hebbian learning rule. Annual review of neuroscience, 31(1):25-46, 2008. 2", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 5", + "[8] Jun Chen, Ming Hu, Boyang Li, and Mohamed Elhoseiny. Efficient self-supervised vision pretraining with local masked reconstruction. arXiv preprint arXiv:2206.00790, 2022. 1, 2, 5", + "[9] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 3", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[11] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.5, 6", + "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 5, 6", + "[13] Francis Crick. The recent excitement about neural networks. Nature, 337(6203):129-132, 1989. 3", + "[14] Yang Dan and Mu-ming Poo. Spike timing-dependent plasticity of neural circuits. Neuron, 44(1):23-30, 2004. 3", + "[15] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. 1, 2, 4, 5, 6", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 2, 3, 6", + "[17] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6824-6835, 2021. 3", + "[18] Peng Gao, Teli Ma, Hongsheng Li, Jifeng Dai, and Yu Qiao. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892, 2022. 2, 3, 5, 7", + "[19] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Yunhe Wang, and Chang Xu. Fastmim: Expediting masked image modeling pre-training for vision. arXiv preprint arXiv:2212.06593, 2022. 1", + "[20] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 2", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1, 2, 3, 4, 5, 6, 7", + "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[23] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 6", + "[24] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921-1930, 2019. 3", + "[25] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification. In ACL, 2018. 1, 4", + "[26] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling. arXiv preprint arXiv:2205.13515, 2022. 1, 2, 3, 4, 5, 6, 7", + "[27] Huawei. Mindspace. https://www.mindspore.cn/, 2020.8", + "[28] DH Hubel and TN Wiesel. Receptive fields of optic nerve fibres in the spider monkey. The Journal of physiology, 154(3):572, 1960. 3, 8" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2130", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 1", + "[30] Xiang Li, Wenhai Wang, Lingfeng Yang, and Jian Yang. Uniform masking: Enabling mae pre-training for pyramid-based vision transformers with locality. arXiv preprint arXiv:2205.10063, 2022. 1, 2, 4", + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6", + "[32] Hao Liu, Xinghua Jiang, Xin Li, Antai Guo, Deqiang Jiang, and Bo Ren. The devil is in the frequency: Geminated gestalt autoencoder for self-supervised visual pre-training. arXiv preprint arXiv:2204.08227, 2022. 1, 2", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 3, 6, 8", + "[34] Sindy Löwe, Peter O'Connor, and Bastiaan Veeling. Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems, 32, 2019. 2, 8", + "[35] Arild Nøkland and Lars Hiller Eidnes. Training neural networks with local error signals. In International conference on machine learning, pages 4839-4850. PMLR, 2019. 2, 8", + "[36] Namuk Park and Songkuk Kim. How do vision transformers work? In International Conference on Learning Representations, 2022. 2, 5", + "[37] Myeongjang Pyeon, Jihwan Moon, Taeyoung Hahn, and Gunhee Kim. Sedona: Search for decoupled neural networks toward greedy block-wise learning. In International Conference on Learning Representations, 2021. 2", + "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3", + "[39] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 3", + "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6", + "[41] Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for combining multiple partitions. Journal of machine learning research, 3(Dec):583-617, 2002. 4", + "[42] Zhiqing Sun, Zhuohan Li, Haoqing Wang, Di He, Zi Lin, and Zhihong Deng. Fast structured decoding for sequence models. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[43] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In Zhi-Hua" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1075-1081. International Joint Conferences on Artificial Intelligence Organization, 2021. Main Track. 3", + "[44] Haoqing Wang, Xun Guo, Zhi-Hong Deng, and Yan Lu. Rethinking minimal sufficient representation in contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16041-16050, 2022. 2", + "[45] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 1, 2, 3", + "[46] Yulin Wang, Zanlin Ni, Shiji Song, Le Yang, and Gao Huang. Revisiting locally supervised learning: an alternative to end-to-end training. In International Conference on Learning Representations, 2021. 2", + "[47] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 1, 2, 4, 5, 6, 7", + "[48] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6", + "[49] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 1, 2", + "[50] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 1, 2, 4, 5, 6", + "[51] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3008, 2021. 3", + "[52] Tianyi Zhang, Felix Wu, Arzoo Katiyar, Kilian Q Weinberger, and Yoav Artzi. Revisiting few-sample bert finetuning. In International Conference on Learning Representations, 2021. 1, 4", + "[53] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6", + "[54] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 1, 2, 4, 5" + ], + "bbox": [ + 503, + 92, + 890, + 878 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2131", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_model.json b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99fb4e0d63d29f10bffb07db5fd67b65e2e9a10e --- /dev/null +++ b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.809, + 0.152 + ], + "angle": 0, + "content": "Masked Image Modeling with Local Multi-Scale Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.18, + 0.852, + 0.235 + ], + "angle": 0, + "content": "Haoqing Wang\\(^{1}\\), Yehui Tang\\(^{1,2}\\), Yunhe Wang\\(^{2*}\\), Jianyuan Guo\\(^{2}\\), Zhi-Hong Deng\\(^{1*}\\), Kai Han\\(^{2}\\) \n\\(^{1}\\)School of Intelligence Science and Technology, Peking University \\(^{2}\\)Huawei Noah's Ark Lab {wanghaoqing,yhtang,zhdeng} \\(@\\) pku.edu.cn, {yunhe.wang, kai.han} \\(@\\) huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.286 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.473, + 0.664 + ], + "angle": 0, + "content": "Masked Image Modeling (MIM) achieves outstanding success in self-supervised representation learning. Unfortunately, MIM models typically have huge computational burden and slow learning process, which is an inevitable obstacle for their industrial applications. Although the lower layers play the key role in MIM, existing MIM models conduct reconstruction task only at the top layer of encoder. The lower layers are not explicitly guided and the interaction among their patches is only used for calculating new activations. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to multiple local layers including lower and upper layers. Further, since the multiple layers expect to learn the information of different scales, we design local multiscale reconstruction, where the lower and upper layers reconstruct fine-scale and coarse-scale supervision signals respectively. This design not only accelerates the representation learning process by explicitly guiding multiple layers, but also facilitates multi-scale semantical understanding to the input. Extensive experiments show that with significantly less pre-training burden, our model achieves comparable or better performance on classification, detection and segmentation tasks than existing MIM models. Code is available with both MindSpore and PyTorch." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.21, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Recently, Masked Image Modeling (MIM) [2, 21, 50] achieves outstanding success in the field of self-supervised visual representation learning, which is inspired by the Masked Language Modeling (MLM) [4, 29] in natural language processing and benefits from the development of vision transformers [16, 33, 45]. MIM learns semantic representations by first masking some parts of the input and then predicting their signals based on the unmasked parts, e.g., normalized pixels [21, 50], discrete tokens [2, 15], HOG feature [47], deep features [1, 54] or frequencies [32, 49]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.271, + 0.895, + 0.573 + ], + "angle": 0, + "content": "Despite superior performance on various downstream tasks, these models have huge computational burden and slow learning process [26]. They typically require thousands of GPU Hours for pre-training on ImageNet-1K to get generalizing representations. Since we expect to pretrain these models on more massive amount of unlabeled data (e.g., free Internet data) to obtain more generalizing representations in practice, the pre-training efficiency is an inevitable bottleneck limiting the industrial applications of MIM. How to accelerate the representation learning in MIM is an important topic. To this end, MAE [21] pioneered the asymmetric encoder-decoder strategy, where the costly encoder only operates few visible patches and the lightweight decoder takes all the patches as input for prediction. Further, GreenMIM [26] extends the asymmetric encoder-decoder strategy to hierarchical vision transformers (e.g., Swin [33]). Besides, [8, 19, 30] shrinks the input resolution to lessen the input patches, thereby reducing the computational burden. However, they all aim to accelerate the encoding process rather than the representation learning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.895, + 0.866 + ], + "angle": 0, + "content": "In MIM, the learning of upper layers depends on that of lower ones during pre-training, since the upper-layer features are calculated from the lower layers. Besides, during fine-tuning the upper layers are typically tuned quickly to adapt to the downstream task while the lower ones change more slowly and need to be well-learned [2, 25, 52]. Even fine-tuning only the several upper layers and freezing the others can obtain similar performance [21]. Therefore, the lower layers of encoder play the key role in MIM. However, all existing MIM models only conduct reconstruction task at the top layer of encoder and the lower ones are not explicitly guide, thus the interaction among their patches is only used for calculating the activations of the next layer. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to both lower and upper layers to explicitly guide them and thus accelerate the overall learning process. Using tiny decoder is sufficient for each local reconstruction task and does not significantly increase the computational burden." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "How to properly conduct reconstruction tasks at multiple local layers is a non-trivial problem. For example, ap-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2122" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.482, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.089, + 0.857, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 1. Top-1 fine-tuning accuracy on ImageNet-1K vs. Pre-training duration. The duration is estimated on a machine with one Tesla V100-32G GPU, CUDA 10.2 and PyTorch 1.8. 'GPU Hours' is the running time on single GPU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.473, + 0.675 + ], + "angle": 0, + "content": "plying the top-layer reconstruction task to carefully chosen local layers of ViT [16] can not achieve meaningful improvement. In general, the lower layers exploit low-level information and the upper ones learn high-level information [18, 36], so it is not appropriate to use the supervision signals of same scale for multiple local reconstruction tasks. Here 'scale' is the spatial size of the supervision signals calculated from the divided input regions, e.g., the signals from the \\( p \\times p \\) regions in an input of \\( H \\times W \\) resolution has the scale of \\( \\frac{H}{p} \\times \\frac{W}{p} \\). The fine-scale and coarse-scale supervisions typically contain low-level and high-level information of the input respectively, and these multi-scale supervisions from input are widely ignored by existing MIM models. To this end, we propose local multi-scale reconstruction where the lower and upper layers reconstruct fine-scale and coarse-scale supervisions respectively. This design not only accelerates the representation learning process, but also facilitates multi-scale semantical understanding to the input. When the decoded predictions have different scale with the supervisions (e.g., on ViT), we use the deconvolution/pool operations to rescale them. We also apply the asymmetric encoder-decoder strategy [21, 26] for quick encoding. Our model, dubbed as LocalMIM, are illustrated in Fig. 2 (a). Overall, we summarize our contributions as follows." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.684, + 0.47, + 0.73 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first work in MIM to conduct local reconstructions and use multiscale supervisions from the input." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.74, + 0.47, + 0.77 + ], + "angle": 0, + "content": "- Our model is architecture-agnostic and can be used in both columnar and pyramid architectures." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- From extensive experiments, we find that 1) LocalMIM is more efficient than existing MIM models, as shown in Fig. 1 and Table 1. For example, LocalMIM achieves the best MAE result with \\(3.1 \\times\\) acceleration on ViT-B and the best GreenMIM result with \\(6.4 \\times\\) acceleration on Swin-B. 2) In terms of top-1 fine-tuning accuracy on ImageNet-1K, LocalMIM achieves \\(84.0\\%\\) using ViT-B and \\(84.1\\%\\) using Swin" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.684, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.311, + 0.892, + 0.371 + ], + "angle": 0, + "content": "B with significantly less pre-training duration than existing MIM models. The obtained representations also achieve better generalization on detection and segmentation downstream tasks, as shown in Table 2 and 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.383, + 0.651, + 0.4 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Masked Image Modeling. With the development of vision transformers [16, 20, 33, 42, 45], Masked Image Modeling (MIM) gradually replaces the dominant position of contrastive learning [10, 22, 44] in visual self-supervised representation learning due to its superior fine-tuning performance in various visual downstream tasks. Many target signals have been designed for the mask-prediction pretext task in MIM, such as normalized pixels [21, 50], discrete tokens [2, 15], HOG feature [47], deep features [1, 54] or frequencies [32, 49]. However, they are all only applied as single-scale supervisions for reconstruction. An inevitable bottleneck for the industrial applications of MIM is that these models typically require huge computational resources and long pre-training duration. To this end, some works accelerate the encoding process via the asymmetric encoder-decoder strategy [21, 26] or lessening the input patches [8, 30]. Only accelerating the encoding process sometimes doesn't really speed up the representation learning, like GreenMIM vs. SimMIM192 in Fig. 1. In this work, we use local multi-scale reconstructions to explicitly guide multiple lower layers and thus accelerate the overall learning process. Our method is compatible with the above quick encoding approaches. ConvMAE [18] fuses the feature of local layers for the final reconstruction to explicitly guide them, but still applies single-scale supervision signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Locally supervised learning. Considering the biological brain learns mainly based on local information [6], some works [34, 35, 37, 46] used local error signals to greedily optimize individual blocks of the backbone without global back-propagation. This greedy training procedure significantly reduces memory overhead, but can not accelerate model learning. In this work, we also optimize error at" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2123" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.891, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.897, + 0.37 + ], + "angle": 0, + "content": "Figure 2. Illustration of local multi-scale reconstruction. a) We randomly mask some input patches and then predict their supervision signals of different scales at multiple local layers. The multi-scale supervisions are obtained by first dividing the input under different scales and then extracting signals with some feature descriptors. The lower layers are responsible for fine-scale reconstruction and the upper ones are responsible for coarse-scale reconstruction. We also use the asymmetric encoder-decoder strategy for quick encoding. b) The decoder consists of three parts: Transformer blocks, Deconvolution/Pool (optional) and Multilayer perceptron." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.473, + 0.506 + ], + "angle": 0, + "content": "multiple local layers but still use global back-propagation for end-to-end training. Compared with existing MIM models, our local reconstruction is more biologically plausible since the human brain prefers local learning rules [3, 13, 14]. Feature distillation [24, 39] also explicitly guides the local layers, but needs costly pre-trained or momentum teacher. In fact, the multi-scale supervisions from original input are sufficient to guide local layers and also readily available." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.678 + ], + "angle": 0, + "content": "Multi-scale property. Biological visual perception is hierarchical and multi-scale [28]. Moreover, multi-scale features are also useful for many visual tasks [9,23,38]. To this end, many advanced vision transformers [17,33,43,45,51] hard-code the multi-scale property to the architectures and boost the performance. However, global single-scale reconstruction is not enough for guiding multiple local layers to learn multi-scale information. In this work, beyond the multi-scale features, we further introduce multi-scale supervisions to soft-code this property. This new design is not limited to specific architectures." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.693, + 0.158, + 0.709 + ], + "angle": 0, + "content": "3. Model" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.719, + 0.307, + 0.736 + ], + "angle": 0, + "content": "3.1. Masked Image Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.82 + ], + "angle": 0, + "content": "In MIM, the models first mask some parts of the input image and then predict their information based on the un-masked observation. For quick encoding, MAE [21] and GreenMIM [26] use the asymmetric encoder-decoder strategy. The major designs are described below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Image presentation. Generally, we present an image to the sequence of visual patches as the input to vision transformers [16, 33]. The input image \\( x \\in \\mathbb{R}^{H \\times W \\times C} \\) is reshaped to \\( N = HW / p^2 \\) non-overlapping patches \\( x^p \\in \\mathbb{R}^{N \\times (p^2C)} \\), where \\( p \\) is the patch size, \\( (H,W) \\) is the resolution of in" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.384, + 0.893, + 0.445 + ], + "angle": 0, + "content": "put image and \\( C \\) is the number of channels. The patches \\( \\{x_i^p\\}_{i = 1}^N \\) are then linearly mapped to the patch embeddings. For retaining the positional information, the patches are typically added with positional embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.624 + ], + "angle": 0, + "content": "Vision transformers. Unlike the grid operations in the convolutional neural networks, vision transformers [16,33,45] use the stacked multi-head self-attention modules. To obtain the multi-scale feature maps for dense prediction tasks, some advanced vision transformers [33, 45] extend the columnar ViT [16] to the pyramidal structure, where the feature maps change from fine-scale to coarse-scale and are expected to capture the multi-scale information from the input. The patches \\([x_1^p,\\dots ,x_N^p ]\\) are fed to a vision transformer with \\(L\\) layers and obtain the output feature at \\(l\\)-th layer \\(z^{l} = [z_{1}^{l},\\dots ,z_{N_{l}}^{l}],l = 1,\\dots ,L.\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.895, + 0.756 + ], + "angle": 0, + "content": "Masking. Given the patch sequence \\(\\{x_i^p\\}_{i = 1}^N\\), MIM constructs a random mask \\(m\\in \\{0,1\\} ^N\\) to indicate the masked patches that correspond to \\(m_{i} = 1\\). There are two main masking strategies, random masking [21] and block-wise masking [2]. Specially, since the number of patches \\(N_{l}\\) decreases through layers in the pyramidal architectures, some works [18,26] first construct the random mask \\(m^L\\) with the size of \\(N_{L}\\), and then up-sample it to size \\(N\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Encoder. Only the visible patches \\( x^{v} = \\{x_{i}^{p}|m_{i} = 0\\} \\) are fed to the encoder and mapped to the latent features \\( z_{v}^{l}, l = 1,\\dots ,L \\), which significantly reduces compute and memory. Specially, since the local window attention in the pyramidal architectures is not compatible with the incomplete input patches (i.e., only visible patches), GreenMIM [26] proposes the Optimal Grouping algorithm and the Group Window Attention scheme. After pre-training, the encoder is used in various downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2124" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "Decoder. The decoder takes as the input of both encoded visible patches \\( z_v^L \\) and mask tokens \\( \\{e_{[\\mathbb{M}]}|m_i^L = 1\\} \\), where the mask token \\( e_{[\\mathbb{M}]} \\) is a shared and learnable vector. The positional embeddings are also added to them for providing the location information. Since the decoder is only used during pre-training to output the prediction \\( \\hat{y}^L \\), it can be any architectures which support the global information propagation among patches, e.g., a series of ViT or Swin blocks. Many works [21, 26, 30] show that using lightweight decoders is sufficient to learn generalizing representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.47, + 0.293 + ], + "angle": 0, + "content": "Global reconstruction. All existing MIM models predict the supervision signal \\(y\\) based on the final output feature \\(z_v^L\\) of the encoder and minimize a global reconstruction loss" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.303, + 0.47, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M I M} = - \\sum_ {i = 1} ^ {N _ {L}} m _ {i} ^ {L} \\cdot \\ln P \\left(y _ {i} \\mid \\hat {y} _ {i} ^ {L}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.469, + 0.402 + ], + "angle": 0, + "content": "where the loss is calculated on the masked patches with \\( m_{i}^{L} = 1 \\), and \\( P(\\cdot |\\cdot) \\) can be the Gaussian or Dirichlet distributions for regression or classification losses respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.41, + 0.3, + 0.426 + ], + "angle": 0, + "content": "3.2. Analysis and motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.584 + ], + "angle": 0, + "content": "For pre-training, the upper-layer features are calculated from the lower layers, so the well-learned lower layers can propagate semantical knowledge to the upper ones and facilitate their learning. For fine-tuning, the upper layers typically adapt quickly to specific downstream tasks, while the lower ones change more slowly and need to be sufficient learned during pre-training [2,25,52]. Even fine-tuning only the several upper layers and freezing the others can obtain comparable performance [21]. Therefore, the lower layers of the encoder play the key role in MIM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.471, + 0.902 + ], + "angle": 0, + "content": "After patchification and linearly projection, the initial patch embeddings lose the inter-patch semantical relations. The self-attention mechanism in vision transformers is responsible for learning these relations by inter-patch interaction and build a better representation space than pixel space [5]. Further, since the self-attention mechanism has the computational complexity with a quadratic dependence on patch number \\( N \\), it is difficult to learn the inter-patch interactions, especially for the lower layers of pyramidal architectures where the small patch size \\( p \\) leads to huge \\( N \\). However, under the global reconstruction loss, the interpatch interaction at lower layers is not explicitly guided, and the simple task of calculating new activations is not sufficient to guide it. As a result, it is hard for the patches at lower layers to learn inter-patch relations. Concretely, we examine Normalized Mutual Information (NMI) [41] between query and key patches at each layer, where the high NMI value means the attention maps strongly depend on the query patches. Intuitively, the semantical representations should have highly query-adaptive attentions, i.e., the different query patches faithfully attend to their semantically" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.091, + 0.863, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Figure 3. Normalized Mutual Information (NMI) between query and key patches at each layer of a pre-trained ViT-B." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.379 + ], + "angle": 0, + "content": "related regions. This is an advantage of the self-attention mechanism. As shown in Fig. 3, existing MIM models with global loss have small NMI value at lower layers, which means their patches have less query-adaptive attentions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.47 + ], + "angle": 0, + "content": "The reconstruction task requires holistic reasoning among patches to predict the masked signals and thus obtains semantic understanding to the input. Since this challenging task facilitates non-trivial inter-patch interactions, we apply it at multiple local layers, including both lower and upper ones to explicitly guide them all." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.783, + 0.495 + ], + "angle": 0, + "content": "3.3. Local multi-scale reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.825 + ], + "angle": 0, + "content": "In MIM, the supervision signals for a reconstruction task are directly calculated from the input. Concretely, we evenly divide the input image \\( x \\in \\mathbb{R}^{H \\times W \\times C} \\) into nonoverlapping regions \\( \\{x_i \\in \\mathbb{R}^{p \\times p \\times C}\\}_{i=1}^{HW/p^2} \\) and use some feature descriptor \\( \\pi \\) to extract the supervision signal \\( y_i = \\pi(x_i) \\). To learn generalizing representations, many feature descriptors have been designed, such as pixel normalization [21, 50], HOG [47] and the pre-trained or momentum teacher [1,2,15,54]. We define the scale of supervision \\( y \\) as its spatial size \\( \\frac{H}{p} \\times \\frac{W}{p} \\). For a given input, the fine-scale supervisions from finely-divided input regions typically contain the low-level semantical information of the input, like corners, edges or textures. Relatively, the coarse-scale supervisions capture high-level semantical information of the input, like the shape of partial or whole object. Intuitively, multi-scale supervisions can guide representation learning better than the common single-scale ones due to their richer semantic information. In this work, we mainly consider the feature descriptors which are readily available without extra pre-training burden and costly forward inference of teacher networks, e.g., pixel normalization and HOG." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 4f, directly applying the top-layer reconstruction task at carefully chosen local layers of ViT can not achieve meaningful improvement, where each local task predicts the supervisions of same scale. In fact, the lower and upper layers expect to learning low-level and high-level" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2125" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.849, + 0.515 + ], + "angle": 0, + "content": "
ModelBackbone# ParamsPT EpochGPU Hours/Ep.Total GPU HoursAcc
Scratch, ViTViT-B86M01.5-82.3
Scratch, SwinSwin-B88M02.4-83.5
MoCo v3 [12]ViT-B86M600--83.2
DINO [7]ViT-B86M300--82.8
BEiT [2]ViT-B86M8002.4192083.2
iBOT [54]ViT-B86M40010.1404083.8
MAE [21]ViT-B86M8001.188083.3
MAE [21]ViT-B86M16001.1176083.6
MAE [21]ViT-L307M16001.7272085.9
MaskFeat [47]ViT-B86M16003.9624084.0
CAE [11]ViT-B86M8002.8224083.6
\\( LoMaR^† \\) [8]ViT-B86M16001.4224084.1
data2Vec† [1]ViT-B86M8003.0240084.2
PeCo [15]ViT-B86M800--84.5
LocalMIM-HOGViT-B86M1000.77083.3
LocalMIM-HOGViT-B86M16000.7112084.0
LocalMIM-HOGViT-L307M8001.080085.8
\\( SimMIM_{192} \\) [50]Swin-B88M8001.8144084.0
\\( SimMIM_{192} \\) [50]Swin-L197M8003.0240085.4
GreenMIM [26]Swin-B88M8000.864083.7
GreenMIM [26]Swin-L197M8001.4112085.1
LocalMIM-PixelSwin-B88M1001.010083.7
LocalMIM-HOGSwin-B88M1001.111083.8
LocalMIM-PixelSwin-B88M4001.040084.0
LocalMIM-HOGSwin-B88M4001.144084.1
LocalMIM-HOGSwin-L197M8001.6128085.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.524, + 0.895, + 0.555 + ], + "angle": 0, + "content": "Table 1. Top-1 fine-tuning accuracy on ImageNet-1K. All models are pre-trained and fine-tuned under \\(224 \\times 224\\) resolution except that \\(\\mathrm{SimMIM}_{192}\\) uses \\(192 \\times 192\\) resolution for pre-training. \\(\\dagger\\) means using the relative positional encoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.721 + ], + "angle": 0, + "content": "information respectively [18, 36], so it is not appropriate to use the single-scale supervisions to guide multiple local layers, even for the columnar architectures with feature maps of the same scale at all layers. To this end, we make the lower layers to reconstruct fine-scale supervisions and the upper ones to reconstruct coarse-scale supervisions. Specially, for the pyramidal architectures which already hardcode the multi-scale property to the features by setting their spatial sizes, we use the supervisions with same scale as the feature maps at the chosen layers for compatibility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The decoding process under a specific scale is illustrated in Fig. 2 (b). The decoder consists of three parts: Transformer blocks for reasoning, (optional) Deconvolution/Pool for rescaling and Multilayer perceptron for prediction. Concretely, based on the encoded visible patches \\( z_{v}^{l} \\) from \\( l \\)-th layer and the mask tokens \\( \\{e_{[\\mathbb{M}]}^{l}|m_{i}^{l} = 1\\} \\), a decoder outputs prediction \\( \\hat{y}^{l} \\) that has the same scale as feature map \\( z_{v}^{l} \\). When the supervision \\( y^{l} \\) has different scale with the feature map \\( z_{v}^{l} \\) (e.g., in the columnar architectures), the decoded prediction can not match the supervision. To this end, we use the deconvolution/pool operations to rescale" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.659 + ], + "angle": 0, + "content": "the prediction \\(\\hat{y}^l\\) for matching the supervision \\(y^{l}\\). For example, we can rescale the \\(14\\times 14\\) prediction to the scale of \\(56\\times 56\\) with twice deconvolution operations or to the scale of \\(7\\times 7\\) with average pool. To avoid excessive computational overhead, we use tiny decoders containing one Transformer block with small embedding dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.691 + ], + "angle": 0, + "content": "The training loss is the weighted summation of the reconstruction losses at the chosen layers" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.699, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {L o c a l M I M}} = - \\sum_ {l \\in \\mathcal {I}} w _ {l} \\cdot \\sum_ {i = 1} ^ {N _ {l}} m _ {i} ^ {l} \\cdot \\ln P \\left(y _ {i} ^ {l} \\mid \\hat {y} _ {i} ^ {l}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) is the set of chosen layers, \\(w_{l}\\) is the coefficient of each local loss, and mask \\(m^l\\) is calculated by up/down sampling the initial mask \\(m\\). These local losses guide the patches at multiple chosen layers to conduct semantic interactions under different scales, which not only accelerates the learning of multiple layers but also facilitates multiscale semantical understanding to the input. As shown in Fig. 3, compared with existing models, our LocalMIM has larger NMI values at lower layers, which means the attention maps depend more strongly on the query patches." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "2126" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.46, + 0.242 + ], + "angle": 0, + "content": "
ModelPT EpochPT HoursmIoU
Supervised--47.4
MoCo v3 [12]300-47.3
BEiT [2]800192047.1
MAE [21]1600176048.1
MaskFeat [47]1600624048.8
PeCo [15]800-48.5
CAE [11]800224048.8
LocalMIM-HOG1600112049.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.253, + 0.47, + 0.296 + ], + "angle": 0, + "content": "Table 2. Semantic segmentation on ADE20K using UperNet with ViT-B backbone. Our LocalMIM achieves better results than previous MIM models with less pre-training duration." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.312, + 0.21, + 0.33 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.339, + 0.47, + 0.386 + ], + "angle": 0, + "content": "In this section, we first evaluate our LocalMIM model on classification, detection and segmentation tasks, and then provide some ablation studies for deep understanding." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.398, + 0.348, + 0.415 + ], + "angle": 0, + "content": "4.1. Classification on ImageNet-1K" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.745 + ], + "angle": 0, + "content": "Settings. Both pre-training and fine-tuning are conducted on ImageNet-1K [40] dataset under the \\(224 \\times 224\\) resolution. We mainly examine two representative architectures, columnar ViT [16] and pyramidal Swin [33]. The input images are patchified with patch size \\(p = 16\\) for ViT and \\(p = 4\\) for Swin, and the obtained patches are randomly masked with ratio \\(r = 0.75\\) by default. We use both HOG feature [47] and normalized pixels [21] as the supervision signals. For Swin, we introduce the reconstruction task after each stage and the supervision signals have the same scale with the output feature maps, e.g., the chosen layers \\(\\mathcal{I} = \\{2,4,22,24\\}\\) and the scales of supervision are \\(\\{28^2,14^2,7^2,7^2\\}\\) on Swin-B. Each decoder contains one transformer block with the embedding dimension of 128 and 4 attention heads. For ViT, the chosen layers \\(\\mathcal{I} = \\{2,4,4 + n,6 + n\\}\\), e.g., \\(n = 6\\) on ViT-B and \\(n = 18\\) on ViT-L. The scales of supervision are \\(\\{56^2,28^2,14^2,7^2\\}\\). Each decoder contains one transformer block with the embedding dimension of 256 and 8 attention heads. The pretraining and fine-tuning schedules mostly follow [21, 26], and more detailed settings can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Results. We compare our LocalMIM with existing MIM models and examine both pre-training efficiency and top-1 fine-tuning accuracy on ImageNet-1K. The results are provided in Fig. 1 and Table 1. For fair comparison, we estimate the pre-training duration of each model on the same machine with one Tesla V100-32G GPU, CUDA 10.2 and Pytorch 1.8. We report the running time on single GPU, denoted as 'GPU Hours', see the Appendix for more details. On ViT-B, LocalMIM achieves the best results of MAE [21] and MaskFeat [47] with \\(3.1 \\times\\) and \\(5.6 \\times\\) accel" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.88, + 0.173 + ], + "angle": 0, + "content": "
ModelPT EpochPT HoursAPbAPm
Supervised30084048.543.2
SimMIM192 [50]800144050.444.4
GreenMIM [26]80064050.044.1
LocalMIM-HOG40044050.744.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.183, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Table 3. Object detection and instance segmentation on COCO. We fine-tune Mask R-CNN end-to-end with Swin-B backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.226, + 0.892, + 0.482 + ], + "angle": 0, + "content": "eration respectively. On Swin-B, LocalMIM achieves those of \\(\\mathrm{SimMIM}_{192}\\) [50] and GreenMIM [26] with \\(3.6\\times\\) and \\(6.4\\times\\) acceleration respectively. In terms of final top-1 finetuning accuracy, LocalMIM achieves \\(84.0\\%\\) with ViT-B and \\(84.1\\%\\) with Swin-B. Compared with previous best results, LocalMIM achieves comparable performance with significantly less pre-training duration. Note that PeCo [15] uses more advanced feature descriptor, a pre-trained codebook, which introduces additional pre-training burden. Besides, on Swin-B, with 100 epochs pre-training and 100 epochs supervised fine-tuning LocalMIM achieves \\(83.8\\%\\) top-1 accuracy and takes about 350 GPU Hours, while 300 epochs supervised training from scratch only achieves \\(83.5\\%\\) top-1 accuracy and takes about 720 GPU Hours. This means even for the high-level classification task, self-supervision learning is more efficient and effective than supervised learning. Similar results can also be found on ViT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.492, + 0.679, + 0.506 + ], + "angle": 0, + "content": "4.2. Downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.515, + 0.892, + 0.56 + ], + "angle": 0, + "content": "We transfer our pre-trained backbone to semantic segmentation on ADE20K [53] and object detection and segmentation on COCO [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Semantic segmentation on ADE20K. We conduct semantic segmentation on ADE20K using UperNet [48] and following the code in [2, 21]. See the Appendix for fine-tuning details. The results are shown in Table 2. With significantly less computation burden, LocalMIM outperforms the state-of-the-art result by 0.7. Besides, LocalMIM has the same top-1 fine-tuning accuracy with MaskFeat [47] on ImageNet-1K, but has better semantic segmentation performance, which means our local multi-scale reconstruction facilitates the learning of multi-scale semantic knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Object detection and instance segmentation on COCO. We fine-tune Mask R-CNN [23] on COCO with Swin-B backbone. Following [26], we also use the code base and \\(3 \\times\\) fine-tuning schedule from the supervised Swin, where the model is fine-tuned for 36 epochs. See the Appendix for fine-tuning details. We report box AP \\((\\mathrm{AP}^b)\\) for object detection and mask AP \\((\\mathrm{AP}^m)\\) for instance segmentation. The results are shown in Table 3. With no labeling burden, our LocalMIM outperforms supervised pre-training by 2.2 \\(\\mathrm{AP}^b\\) and \\(1.7\\mathrm{AP}^m\\). Compared to SimMIM192 and GreenMIM, our LocalMIM achieves \\(+0.3\\) and \\(+0.7\\) gain respectively for detection, and \\(+0.5\\) and \\(+0.8\\) gain for segmenta" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2127" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.314, + 0.144 + ], + "angle": 0, + "content": "
targetViT-BSwin-B
baselines82.983.2
Pixels83.083.7
HOG83.383.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.157, + 0.33, + 0.181 + ], + "angle": 0, + "content": "(a) Reconstruction target. HOG is more suitable as multi-scale supervisions than pixels." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.183, + 0.327, + 0.272 + ], + "angle": 0, + "content": "
locationsbackboneacc
GreenMIMSwin-B83.2
[24]83.3
[22, 24]83.4
[4, 22, 24]Swin-B83.6
[2, 4, 22, 24]83.8
fusion83.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.276, + 0.33, + 0.301 + ], + "angle": 0, + "content": "(d) Locations for Swin. Applying reconstructions at all stages mostly accelerates learning." + }, + { + "type": "table", + "bbox": [ + 0.379, + 0.089, + 0.597, + 0.154 + ], + "angle": 0, + "content": "
ratioViT-BSwin-B
0.4083.283.7
0.6083.183.7
0.7583.383.8
0.9083.083.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.157, + 0.608, + 0.181 + ], + "angle": 0, + "content": "(b) Mask ratio. Masking \\(75\\%\\) patches works the best for both ViT and Swin." + }, + { + "type": "table", + "bbox": [ + 0.37, + 0.184, + 0.601, + 0.263 + ], + "angle": 0, + "content": "
locationsbackboneacc
MAEViT-B82.9
[12]83.0
[3, 6, 9, 12]ViT-B83.1
[2, 4, 10, 12]83.3
[1, 2, 11, 12]82.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.276, + 0.608, + 0.301 + ], + "angle": 0, + "content": "(e) Locations for ViT. The chosen layers for local losses obviously affect performance." + }, + { + "type": "table", + "bbox": [ + 0.651, + 0.089, + 0.88, + 0.144 + ], + "angle": 0, + "content": "
decoderViT-BSwin-B
512D - 16H83.3 (1.0h)83.8 (1.6h)
256D - 8H83.3 (0.7h)83.7 (1.3h)
128D - 4H83.2 (0.6h)83.8 (1.1h)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.641, + 0.157, + 0.887, + 0.181 + ], + "angle": 0, + "content": "(c) Decoder design. A tiny decoder performs as well as the larger one but is more efficient." + }, + { + "type": "table", + "bbox": [ + 0.647, + 0.184, + 0.887, + 0.263 + ], + "angle": 0, + "content": "
scalesbackboneacc
MAEViT-B82.9
[142, 142, 142, 142]83.0
[282, 142, 72, 72]ViT-B83.2
[562, 282, 142, 72]83.3
[72, 142, 282, 562]83.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.641, + 0.276, + 0.887, + 0.3 + ], + "angle": 0, + "content": "(f) Scales for ViT. Using the supervisions from fine-scale to coarse-scale works the best." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.894, + 0.371 + ], + "angle": 0, + "content": "Table 4. Ablation studies with ViT-B and Swin-B on ImageNet-1K. We report the top-1 fine-tuning accuracy (%) and the default setting is: the reconstruction target is HOG feature, the mask ratio is \\(75\\%\\), the decoder contains one Transformer block with dimension 256 and 8 heads for ViT and dimension 128 and 4 heads for Swin, the local reconstructions are applied at all stages of Swin and [2, 4, 10, 12]-th layer of ViT, the scales of supervision are \\([28^2, 14^2, 7^2, 7^2]\\) for Swin and \\([56^2, 28^2, 14^2, 7^2]\\) for ViT, and the pre-training length is 100 epochs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.391, + 0.396, + 0.406 + ], + "angle": 0, + "content": "tion with significantly less pre-training duration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.42, + 0.238, + 0.434 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.469, + 0.506 + ], + "angle": 0, + "content": "We ablate our LocalMIM using the default setting in Table 4. MAE [21] and GreenMIM [26] are our main baselines on ViT and Swin respectively, since we all adopt the asymmetric encoder-decoder strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.693 + ], + "angle": 0, + "content": "Reconstruction target. We examine two types of signal as supervisions: HOG feature [47] and normalized pixel [21], and the results are shown in Table 4a. HOG feature is more suitable as multi-scale supervisions than normalized pixel on both ViT and Swin, since it contains more refined semantic information. Compared to the columnar ViT with single-scale feature maps, our multi-scale supervisions achieve more improvement on the pyramidal Swin. Since we need to rescale the predictions in the columnar architectures, it is harder to guide the learning of multi-scale information, e.g., we need more information-refined signals (HOG) and obtain relatively less improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Mask ratio. We explore different mask ratios, ranging from 0.4 to 0.9, and the results are provided in Table 4b. Similar to [21, 26], our LocalMIM is robust to the mask ratio from 0.4 to 0.75, and excessively large one (0.9) harms the performance since it over-complicates the reconstruction tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Decoder design. As shown in [21, 26], a small single-block decoder can achieve the similar fine-tuning accuracy to the heavier one with multiple blocks. However, since we need to use the decoder for each reconstruction task to process the signals of specified scale, their decoder which has the embedding dimension of 512 and 16 attention heads is still heavy. To this end, we explore whether a smaller decoder can still handle the local reconstruction tasks for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.542 + ], + "angle": 0, + "content": "learning good representations. One Transformer block is the minimal requirement to propagate information from visible patches to mask tokens, so we still use one Transformer block but narrow the embedding dimension and reduce the number of attention heads. The results (top-1 fine-tuning accuracy and GPU Hours per epoch) are shown in Table 4c, where \\( dD - hH' \\) represents the decoder with the embedding dimension of \\( d \\) and \\( h \\) attention heads. Surprisingly, a tiny decoder can achieve the same even better fine-tuning performance, but is significantly more efficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Locations for local reconstructions. An important design of our LocalMIM is the locations in the encoder where we introduce local reconstructions. Many pyramidal architectures like Swin already divide the entire backbone into multiple stages, so we introduce multi-scale reconstructions at their output locations. From the last stage, we gradually include the lower stages and the results are shown in Table 4d. Note that the 2-th, 4-th, 22-th and 24-th layer correspond to the output location of the four stages respectively. We also provide the result of fusion method [18] which fuses the output features from each stage for reconstruction. As we can see, introducing reconstructions at local stages can accelerate representation learning, and applying to all stages works the best. Our LocalMIM also outperforms the fusion method which only uses the single-scale reconstruction, and this verifies the importance of multi-scale reconstructions. Further, for the columnar ViT, we follow the experience from Swin and also select four local layers. Concretely, we consider the uniform division [3, 6, 9, 12], Swin-style division [2, 4, 10, 12] and a extreme division [1, 2, 11, 12]. The results are shown in Table 4e. The performance on columnar ViT is sensitive to the chosen local layers, and the Swin-style division \\([2, 4, 4 + n, 6 + n]\\) works the best." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2128" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.155, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.089, + 0.47, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Figure 4. Visualization of the attention maps for different query points, marked with red boxes. LocalMIM can distinguish semantical regions using self-attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Scales of supervisions. Since the pyramidal architectures hard-code the multi-scale property of the feature maps, we directly use the supervisions of same scale as the feature maps for compatibility, i.e., we use \\([28^2, 14^2, 7^2, 7^2]\\) for Swin. For columnar ViT, we consider single-scale supervisions \\([14^2, 14^2, 14^2, 14^2]\\), fine-to-coarse multi-scale supervisions \\([28^2, 14^2, 7^2, 7^2]\\) and \\([56^2, 28^2, 14^2, 7^2]\\), and coarse-to-fine multi-scale supervisions \\([7^2, 14^2, 28^2, 56^2]\\). The results are shown in Table 4f. The multi-scale supervisions achieve better fine-tuning accuracy than the single-scale ones, which means the multiple local layers require the supervision of different scales. Further, the fine-to-coarse supervisions outperform the coarse-to-fine ones, which is consistent with the design priors for visual architectures [33] and biological visual processing [28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Analysis of self-attention mechanism. To qualitatively illustrate the behavior of self-attention mechanism in LocalMIM, we visualize the attention maps for different query positions within an image, as shown in Fig. 4. We use each chosen patch as query and show the patches it attends to. We select the attention maps at 2-th, 4-th, 10-th and 12-th layers of a pre-trained ViT-B/16 backbone to show their changes during forward inference. For object-centric images, LocalMIM can distinguish the foreground object from the background. For more complex multi-object images," + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.888, + 0.169 + ], + "angle": 0, + "content": "
modelbackboneGPU Hours/Ep.acc
LocalMIMViT-B0.783.3
w/ isolated grad0.783.0
LocalMIMSwin-B1.183.8
w/ isolated grad1.183.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.179, + 0.892, + 0.208 + ], + "angle": 0, + "content": "Table 5. Training LocalMIM with isolated gradients achieves similar performance with global back-propagation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.226, + 0.892, + 0.332 + ], + "angle": 0, + "content": "LocalMIM can effectively separate different objects without any task-specific supervision, which means the attention maps are query-adaptive. On the other hand, the patches at lower layers typically more focus on their neighboring regions and capture low-level information, while those at upper layers attend to a wide range of semantically related regions and capture high-level shape information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.893, + 0.55 + ], + "angle": 0, + "content": "Gradient-isolated training. Inspired by locally supervised learning [34, 35], we remove global back-propagation and stop the gradients after each chosen local layer used for reconstruction. The entire backbone is thus divided into multiple gradient-isolated parts. The results are shown in Table 5. Surprisingly, the gradient-isolated training achieves similar performance to global back-propagation, which further verifies the effectiveness of our multi-scale reconstructions for guiding the local layers. It even requires no gradient information from the upper layers. This observation also shows the promise of our local multi-scale reconstruction for the decoupled training of neural networks, which allows for training very deep networks without memory concern and reduces explosive or vanishing gradients." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.627, + 0.582 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.851 + ], + "angle": 0, + "content": "We present a novel and efficient pretext task, local multi-scale reconstruction, where the lower layers and upper layers reconstruct the fine-scale and coarse-scale supervisions from the input respectively. We obtain the multi-scale supervisions by first dividing the input under different scales and then extracting supervisions with appropriate feature descriptors. Our model needs no extra pre-trained codebook and no costly forward inference of teacher networks during pre-training. The asymmetric encoder-decoder strategy with tiny encoders also have small computational burden, our model thus can be trained quickly. The novel pretext task further accelerates the representation learning, especially for the pyramidal architectures. All these designs allow our model to achieve comparable performance to existing models but with significantly less pre-training burden. The gradient-isolated training also verifies the effectiveness of our novel task design in guiding the local layers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. We gratefully acknowledge the support of MindSpore [27], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2129" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Dong-Hyun Lee, Jorg Bornschein, Thomas Mesnard, and Zhouhan Lin. Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[5] Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.429 + ], + "angle": 0, + "content": "[6] Natalia Caporale, Yang Dan, et al. Spike timing-dependent plasticity: a hebbian learning rule. Annual review of neuroscience, 31(1):25-46, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.471, + 0.5 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.556 + ], + "angle": 0, + "content": "[8] Jun Chen, Ming Hu, Boyang Li, and Mohamed Elhoseiny. Efficient self-supervised vision pretraining with local masked reconstruction. arXiv preprint arXiv:2206.00790, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.558, + 0.471, + 0.627 + ], + "angle": 0, + "content": "[9] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.629, + 0.471, + 0.685 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.687, + 0.471, + 0.755 + ], + "angle": 0, + "content": "[11] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.757, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.814, + 0.471, + 0.842 + ], + "angle": 0, + "content": "[13] Francis Crick. The recent excitement about neural networks. Nature, 337(6203):129-132, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.844, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[14] Yang Dan and Mu-ming Poo. Spike timing-dependent plasticity of neural circuits. Neuron, 44(1):23-30, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.895, + 0.305 + ], + "angle": 0, + "content": "[17] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6824-6835, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[18] Peng Gao, Teli Ma, Hongsheng Li, Jifeng Dai, and Yu Qiao. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.351, + 0.892, + 0.406 + ], + "angle": 0, + "content": "[19] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Yunhe Wang, and Chang Xu. Fastmim: Expediting masked image modeling pre-training for vision. arXiv preprint arXiv:2212.06593, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.409, + 0.892, + 0.464 + ], + "angle": 0, + "content": "[20] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.467, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.61, + 0.892, + 0.666 + ], + "angle": 0, + "content": "[23] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.669, + 0.892, + 0.737 + ], + "angle": 0, + "content": "[24] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921-1930, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.74, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[25] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification. In ACL, 2018. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.77, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[26] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling. arXiv preprint arXiv:2205.13515, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[27] Huawei. Mindspace. https://www.mindspore.cn/, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] DH Hubel and TN Wiesel. Receptive fields of optic nerve fibres in the spider monkey. The Journal of physiology, 154(3):572, 1960. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2130" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[30] Xiang Li, Wenhai Wang, Lingfeng Yang, and Jian Yang. Uniform masking: Enabling mae pre-training for pyramid-based vision transformers with locality. arXiv preprint arXiv:2205.10063, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[32] Hao Liu, Xinghua Jiang, Xin Li, Antai Guo, Deqiang Jiang, and Bo Ren. The devil is in the frequency: Geminated gestalt autoencoder for self-supervised visual pre-training. arXiv preprint arXiv:2204.08227, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[34] Sindy Löwe, Peter O'Connor, and Bastiaan Veeling. Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems, 32, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.455, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[35] Arild Nøkland and Lars Hiller Eidnes. Training neural networks with local error signals. In International conference on machine learning, pages 4839-4850. PMLR, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[36] Namuk Park and Songkuk Kim. How do vision transformers work? In International Conference on Learning Representations, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[37] Myeongjang Pyeon, Jihwan Moon, Taeyoung Hahn, and Gunhee Kim. Sedona: Search for decoupled neural networks toward greedy block-wise learning. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.651, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[39] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.693, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.762, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[41] Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for combining multiple partitions. Journal of machine learning research, 3(Dec):583-617, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[42] Zhiqing Sun, Zhuohan Li, Haoqing Wang, Di He, Zi Lin, and Zhihong Deng. Fast structured decoding for sequence models. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[43] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In Zhi-Hua" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1075-1081. International Joint Conferences on Artificial Intelligence Organization, 2021. Main Track. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[44] Haoqing Wang, Xun Guo, Zhi-Hong Deng, and Yan Lu. Rethinking minimal sufficient representation in contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16041-16050, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[45] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[46] Yulin Wang, Zanlin Ni, Shiji Song, Le Yang, and Gao Huang. Revisiting locally supervised learning: an alternative to end-to-end training. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[47] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.445, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[48] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[49] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.558, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[50] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.629, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[51] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3008, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[52] Tianyi Zhang, Felix Wu, Arzoo Katiyar, Kilian Q Weinberger, and Yoav Artzi. Revisiting few-sample bert finetuning. In International Conference on Learning Representations, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.756, + 0.892, + 0.823 + ], + "angle": 0, + "content": "[53] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.826, + 0.892, + 0.88 + ], + "angle": 0, + "content": "[54] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 1, 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "2131" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_origin.pdf b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..000858dd36311d2e876bf8d5fcc6b9b75d08cd65 --- /dev/null +++ b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/cc2ef537-2f02-42c5-9105-77e1170379e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b06e6cc9a51cd378ceb3ad5afa18c5520262988432d1ff8c26a0a2e5d1907e39 +size 2339333 diff --git a/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/full.md b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..91bddfb90586d27cb8ceadc82c61392592a333b0 --- /dev/null +++ b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/full.md @@ -0,0 +1,261 @@ +# Masked Image Modeling with Local Multi-Scale Reconstruction + +Haoqing Wang $^{1}$ , Yehui Tang $^{1,2}$ , Yunhe Wang $^{2*}$ , Jianyuan Guo $^{2}$ , Zhi-Hong Deng $^{1*}$ , Kai Han $^{2}$ $^{1}$ School of Intelligence Science and Technology, Peking University $^{2}$ Huawei Noah's Ark Lab {wanghaoqing,yhtang,zhdeng} $@$ pku.edu.cn, {yunhe.wang, kai.han} $@$ huawei.com + +# Abstract + +Masked Image Modeling (MIM) achieves outstanding success in self-supervised representation learning. Unfortunately, MIM models typically have huge computational burden and slow learning process, which is an inevitable obstacle for their industrial applications. Although the lower layers play the key role in MIM, existing MIM models conduct reconstruction task only at the top layer of encoder. The lower layers are not explicitly guided and the interaction among their patches is only used for calculating new activations. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to multiple local layers including lower and upper layers. Further, since the multiple layers expect to learn the information of different scales, we design local multiscale reconstruction, where the lower and upper layers reconstruct fine-scale and coarse-scale supervision signals respectively. This design not only accelerates the representation learning process by explicitly guiding multiple layers, but also facilitates multi-scale semantical understanding to the input. Extensive experiments show that with significantly less pre-training burden, our model achieves comparable or better performance on classification, detection and segmentation tasks than existing MIM models. Code is available with both MindSpore and PyTorch. + +# 1. Introduction + +Recently, Masked Image Modeling (MIM) [2, 21, 50] achieves outstanding success in the field of self-supervised visual representation learning, which is inspired by the Masked Language Modeling (MLM) [4, 29] in natural language processing and benefits from the development of vision transformers [16, 33, 45]. MIM learns semantic representations by first masking some parts of the input and then predicting their signals based on the unmasked parts, e.g., normalized pixels [21, 50], discrete tokens [2, 15], HOG feature [47], deep features [1, 54] or frequencies [32, 49]. + +Despite superior performance on various downstream tasks, these models have huge computational burden and slow learning process [26]. They typically require thousands of GPU Hours for pre-training on ImageNet-1K to get generalizing representations. Since we expect to pretrain these models on more massive amount of unlabeled data (e.g., free Internet data) to obtain more generalizing representations in practice, the pre-training efficiency is an inevitable bottleneck limiting the industrial applications of MIM. How to accelerate the representation learning in MIM is an important topic. To this end, MAE [21] pioneered the asymmetric encoder-decoder strategy, where the costly encoder only operates few visible patches and the lightweight decoder takes all the patches as input for prediction. Further, GreenMIM [26] extends the asymmetric encoder-decoder strategy to hierarchical vision transformers (e.g., Swin [33]). Besides, [8, 19, 30] shrinks the input resolution to lessen the input patches, thereby reducing the computational burden. However, they all aim to accelerate the encoding process rather than the representation learning. + +In MIM, the learning of upper layers depends on that of lower ones during pre-training, since the upper-layer features are calculated from the lower layers. Besides, during fine-tuning the upper layers are typically tuned quickly to adapt to the downstream task while the lower ones change more slowly and need to be well-learned [2, 25, 52]. Even fine-tuning only the several upper layers and freezing the others can obtain similar performance [21]. Therefore, the lower layers of encoder play the key role in MIM. However, all existing MIM models only conduct reconstruction task at the top layer of encoder and the lower ones are not explicitly guide, thus the interaction among their patches is only used for calculating the activations of the next layer. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to both lower and upper layers to explicitly guide them and thus accelerate the overall learning process. Using tiny decoder is sufficient for each local reconstruction task and does not significantly increase the computational burden. + +How to properly conduct reconstruction tasks at multiple local layers is a non-trivial problem. For example, ap- + +![](images/e09ee8397bd0e0523b8b82cb7af81c8dc4c76d7d07bdc7ec05c2d89699bdf6a0.jpg) +Figure 1. Top-1 fine-tuning accuracy on ImageNet-1K vs. Pre-training duration. The duration is estimated on a machine with one Tesla V100-32G GPU, CUDA 10.2 and PyTorch 1.8. 'GPU Hours' is the running time on single GPU. + +![](images/b956d76e31182bdd429661f85072584ea7fa4bc4f50977fad39967018b60440e.jpg) + +plying the top-layer reconstruction task to carefully chosen local layers of ViT [16] can not achieve meaningful improvement. In general, the lower layers exploit low-level information and the upper ones learn high-level information [18, 36], so it is not appropriate to use the supervision signals of same scale for multiple local reconstruction tasks. Here 'scale' is the spatial size of the supervision signals calculated from the divided input regions, e.g., the signals from the $p \times p$ regions in an input of $H \times W$ resolution has the scale of $\frac{H}{p} \times \frac{W}{p}$ . The fine-scale and coarse-scale supervisions typically contain low-level and high-level information of the input respectively, and these multi-scale supervisions from input are widely ignored by existing MIM models. To this end, we propose local multi-scale reconstruction where the lower and upper layers reconstruct fine-scale and coarse-scale supervisions respectively. This design not only accelerates the representation learning process, but also facilitates multi-scale semantical understanding to the input. When the decoded predictions have different scale with the supervisions (e.g., on ViT), we use the deconvolution/pool operations to rescale them. We also apply the asymmetric encoder-decoder strategy [21, 26] for quick encoding. Our model, dubbed as LocalMIM, are illustrated in Fig. 2 (a). Overall, we summarize our contributions as follows. + +- To the best of our knowledge, this is the first work in MIM to conduct local reconstructions and use multiscale supervisions from the input. +- Our model is architecture-agnostic and can be used in both columnar and pyramid architectures. +- From extensive experiments, we find that 1) LocalMIM is more efficient than existing MIM models, as shown in Fig. 1 and Table 1. For example, LocalMIM achieves the best MAE result with $3.1 \times$ acceleration on ViT-B and the best GreenMIM result with $6.4 \times$ acceleration on Swin-B. 2) In terms of top-1 fine-tuning accuracy on ImageNet-1K, LocalMIM achieves $84.0\%$ using ViT-B and $84.1\%$ using Swin + +B with significantly less pre-training duration than existing MIM models. The obtained representations also achieve better generalization on detection and segmentation downstream tasks, as shown in Table 2 and 3. + +# 2. Related Works + +Masked Image Modeling. With the development of vision transformers [16, 20, 33, 42, 45], Masked Image Modeling (MIM) gradually replaces the dominant position of contrastive learning [10, 22, 44] in visual self-supervised representation learning due to its superior fine-tuning performance in various visual downstream tasks. Many target signals have been designed for the mask-prediction pretext task in MIM, such as normalized pixels [21, 50], discrete tokens [2, 15], HOG feature [47], deep features [1, 54] or frequencies [32, 49]. However, they are all only applied as single-scale supervisions for reconstruction. An inevitable bottleneck for the industrial applications of MIM is that these models typically require huge computational resources and long pre-training duration. To this end, some works accelerate the encoding process via the asymmetric encoder-decoder strategy [21, 26] or lessening the input patches [8, 30]. Only accelerating the encoding process sometimes doesn't really speed up the representation learning, like GreenMIM vs. SimMIM192 in Fig. 1. In this work, we use local multi-scale reconstructions to explicitly guide multiple lower layers and thus accelerate the overall learning process. Our method is compatible with the above quick encoding approaches. ConvMAE [18] fuses the feature of local layers for the final reconstruction to explicitly guide them, but still applies single-scale supervision signals. + +Locally supervised learning. Considering the biological brain learns mainly based on local information [6], some works [34, 35, 37, 46] used local error signals to greedily optimize individual blocks of the backbone without global back-propagation. This greedy training procedure significantly reduces memory overhead, but can not accelerate model learning. In this work, we also optimize error at + +![](images/d688c406129a1b7f8e1cba475142c9340370cd1956635d60d515f2f0fec09a86.jpg) +Figure 2. Illustration of local multi-scale reconstruction. a) We randomly mask some input patches and then predict their supervision signals of different scales at multiple local layers. The multi-scale supervisions are obtained by first dividing the input under different scales and then extracting signals with some feature descriptors. The lower layers are responsible for fine-scale reconstruction and the upper ones are responsible for coarse-scale reconstruction. We also use the asymmetric encoder-decoder strategy for quick encoding. b) The decoder consists of three parts: Transformer blocks, Deconvolution/Pool (optional) and Multilayer perceptron. + +multiple local layers but still use global back-propagation for end-to-end training. Compared with existing MIM models, our local reconstruction is more biologically plausible since the human brain prefers local learning rules [3, 13, 14]. Feature distillation [24, 39] also explicitly guides the local layers, but needs costly pre-trained or momentum teacher. In fact, the multi-scale supervisions from original input are sufficient to guide local layers and also readily available. + +Multi-scale property. Biological visual perception is hierarchical and multi-scale [28]. Moreover, multi-scale features are also useful for many visual tasks [9,23,38]. To this end, many advanced vision transformers [17,33,43,45,51] hard-code the multi-scale property to the architectures and boost the performance. However, global single-scale reconstruction is not enough for guiding multiple local layers to learn multi-scale information. In this work, beyond the multi-scale features, we further introduce multi-scale supervisions to soft-code this property. This new design is not limited to specific architectures. + +# 3. Model + +# 3.1. Masked Image Modeling + +In MIM, the models first mask some parts of the input image and then predict their information based on the un-masked observation. For quick encoding, MAE [21] and GreenMIM [26] use the asymmetric encoder-decoder strategy. The major designs are described below. + +Image presentation. Generally, we present an image to the sequence of visual patches as the input to vision transformers [16, 33]. The input image $x \in \mathbb{R}^{H \times W \times C}$ is reshaped to $N = HW / p^2$ non-overlapping patches $x^p \in \mathbb{R}^{N \times (p^2C)}$ , where $p$ is the patch size, $(H,W)$ is the resolution of in + +put image and $C$ is the number of channels. The patches $\{x_i^p\}_{i = 1}^N$ are then linearly mapped to the patch embeddings. For retaining the positional information, the patches are typically added with positional embeddings. + +Vision transformers. Unlike the grid operations in the convolutional neural networks, vision transformers [16,33,45] use the stacked multi-head self-attention modules. To obtain the multi-scale feature maps for dense prediction tasks, some advanced vision transformers [33, 45] extend the columnar ViT [16] to the pyramidal structure, where the feature maps change from fine-scale to coarse-scale and are expected to capture the multi-scale information from the input. The patches $[x_1^p,\dots ,x_N^p ]$ are fed to a vision transformer with $L$ layers and obtain the output feature at $l$ -th layer $z^{l} = [z_{1}^{l},\dots ,z_{N_{l}}^{l}],l = 1,\dots ,L.$ + +Masking. Given the patch sequence $\{x_i^p\}_{i = 1}^N$ , MIM constructs a random mask $m\in \{0,1\} ^N$ to indicate the masked patches that correspond to $m_{i} = 1$ . There are two main masking strategies, random masking [21] and block-wise masking [2]. Specially, since the number of patches $N_{l}$ decreases through layers in the pyramidal architectures, some works [18,26] first construct the random mask $m^L$ with the size of $N_{L}$ , and then up-sample it to size $N$ . + +Encoder. Only the visible patches $x^{v} = \{x_{i}^{p}|m_{i} = 0\}$ are fed to the encoder and mapped to the latent features $z_{v}^{l}, l = 1,\dots ,L$ , which significantly reduces compute and memory. Specially, since the local window attention in the pyramidal architectures is not compatible with the incomplete input patches (i.e., only visible patches), GreenMIM [26] proposes the Optimal Grouping algorithm and the Group Window Attention scheme. After pre-training, the encoder is used in various downstream tasks. + +Decoder. The decoder takes as the input of both encoded visible patches $z_v^L$ and mask tokens $\{e_{[\mathbb{M}]}|m_i^L = 1\}$ , where the mask token $e_{[\mathbb{M}]}$ is a shared and learnable vector. The positional embeddings are also added to them for providing the location information. Since the decoder is only used during pre-training to output the prediction $\hat{y}^L$ , it can be any architectures which support the global information propagation among patches, e.g., a series of ViT or Swin blocks. Many works [21, 26, 30] show that using lightweight decoders is sufficient to learn generalizing representations. + +Global reconstruction. All existing MIM models predict the supervision signal $y$ based on the final output feature $z_v^L$ of the encoder and minimize a global reconstruction loss + +$$ +\mathcal {L} _ {M I M} = - \sum_ {i = 1} ^ {N _ {L}} m _ {i} ^ {L} \cdot \ln P \left(y _ {i} \mid \hat {y} _ {i} ^ {L}\right) \tag {1} +$$ + +where the loss is calculated on the masked patches with $m_{i}^{L} = 1$ , and $P(\cdot |\cdot)$ can be the Gaussian or Dirichlet distributions for regression or classification losses respectively. + +# 3.2. Analysis and motivation + +For pre-training, the upper-layer features are calculated from the lower layers, so the well-learned lower layers can propagate semantical knowledge to the upper ones and facilitate their learning. For fine-tuning, the upper layers typically adapt quickly to specific downstream tasks, while the lower ones change more slowly and need to be sufficient learned during pre-training [2,25,52]. Even fine-tuning only the several upper layers and freezing the others can obtain comparable performance [21]. Therefore, the lower layers of the encoder play the key role in MIM. + +After patchification and linearly projection, the initial patch embeddings lose the inter-patch semantical relations. The self-attention mechanism in vision transformers is responsible for learning these relations by inter-patch interaction and build a better representation space than pixel space [5]. Further, since the self-attention mechanism has the computational complexity with a quadratic dependence on patch number $N$ , it is difficult to learn the inter-patch interactions, especially for the lower layers of pyramidal architectures where the small patch size $p$ leads to huge $N$ . However, under the global reconstruction loss, the interpatch interaction at lower layers is not explicitly guided, and the simple task of calculating new activations is not sufficient to guide it. As a result, it is hard for the patches at lower layers to learn inter-patch relations. Concretely, we examine Normalized Mutual Information (NMI) [41] between query and key patches at each layer, where the high NMI value means the attention maps strongly depend on the query patches. Intuitively, the semantical representations should have highly query-adaptive attentions, i.e., the different query patches faithfully attend to their semantically + +![](images/05f26d05ff73645e0850ba2457b7c48d9276b00002b2710f815a495d27815808.jpg) +Figure 3. Normalized Mutual Information (NMI) between query and key patches at each layer of a pre-trained ViT-B. + +related regions. This is an advantage of the self-attention mechanism. As shown in Fig. 3, existing MIM models with global loss have small NMI value at lower layers, which means their patches have less query-adaptive attentions. + +The reconstruction task requires holistic reasoning among patches to predict the masked signals and thus obtains semantic understanding to the input. Since this challenging task facilitates non-trivial inter-patch interactions, we apply it at multiple local layers, including both lower and upper ones to explicitly guide them all. + +# 3.3. Local multi-scale reconstruction + +In MIM, the supervision signals for a reconstruction task are directly calculated from the input. Concretely, we evenly divide the input image $x \in \mathbb{R}^{H \times W \times C}$ into nonoverlapping regions $\{x_i \in \mathbb{R}^{p \times p \times C}\}_{i=1}^{HW/p^2}$ and use some feature descriptor $\pi$ to extract the supervision signal $y_i = \pi(x_i)$ . To learn generalizing representations, many feature descriptors have been designed, such as pixel normalization [21, 50], HOG [47] and the pre-trained or momentum teacher [1,2,15,54]. We define the scale of supervision $y$ as its spatial size $\frac{H}{p} \times \frac{W}{p}$ . For a given input, the fine-scale supervisions from finely-divided input regions typically contain the low-level semantical information of the input, like corners, edges or textures. Relatively, the coarse-scale supervisions capture high-level semantical information of the input, like the shape of partial or whole object. Intuitively, multi-scale supervisions can guide representation learning better than the common single-scale ones due to their richer semantic information. In this work, we mainly consider the feature descriptors which are readily available without extra pre-training burden and costly forward inference of teacher networks, e.g., pixel normalization and HOG. + +As shown in Table 4f, directly applying the top-layer reconstruction task at carefully chosen local layers of ViT can not achieve meaningful improvement, where each local task predicts the supervisions of same scale. In fact, the lower and upper layers expect to learning low-level and high-level + +
ModelBackbone# ParamsPT EpochGPU Hours/Ep.Total GPU HoursAcc
Scratch, ViTViT-B86M01.5-82.3
Scratch, SwinSwin-B88M02.4-83.5
MoCo v3 [12]ViT-B86M600--83.2
DINO [7]ViT-B86M300--82.8
BEiT [2]ViT-B86M8002.4192083.2
iBOT [54]ViT-B86M40010.1404083.8
MAE [21]ViT-B86M8001.188083.3
MAE [21]ViT-B86M16001.1176083.6
MAE [21]ViT-L307M16001.7272085.9
MaskFeat [47]ViT-B86M16003.9624084.0
CAE [11]ViT-B86M8002.8224083.6
\( LoMaR^† \) [8]ViT-B86M16001.4224084.1
data2Vec† [1]ViT-B86M8003.0240084.2
PeCo [15]ViT-B86M800--84.5
LocalMIM-HOGViT-B86M1000.77083.3
LocalMIM-HOGViT-B86M16000.7112084.0
LocalMIM-HOGViT-L307M8001.080085.8
\( SimMIM_{192} \) [50]Swin-B88M8001.8144084.0
\( SimMIM_{192} \) [50]Swin-L197M8003.0240085.4
GreenMIM [26]Swin-B88M8000.864083.7
GreenMIM [26]Swin-L197M8001.4112085.1
LocalMIM-PixelSwin-B88M1001.010083.7
LocalMIM-HOGSwin-B88M1001.111083.8
LocalMIM-PixelSwin-B88M4001.040084.0
LocalMIM-HOGSwin-B88M4001.144084.1
LocalMIM-HOGSwin-L197M8001.6128085.6
+ +Table 1. Top-1 fine-tuning accuracy on ImageNet-1K. All models are pre-trained and fine-tuned under $224 \times 224$ resolution except that $\mathrm{SimMIM}_{192}$ uses $192 \times 192$ resolution for pre-training. $\dagger$ means using the relative positional encoding. + +information respectively [18, 36], so it is not appropriate to use the single-scale supervisions to guide multiple local layers, even for the columnar architectures with feature maps of the same scale at all layers. To this end, we make the lower layers to reconstruct fine-scale supervisions and the upper ones to reconstruct coarse-scale supervisions. Specially, for the pyramidal architectures which already hardcode the multi-scale property to the features by setting their spatial sizes, we use the supervisions with same scale as the feature maps at the chosen layers for compatibility. + +The decoding process under a specific scale is illustrated in Fig. 2 (b). The decoder consists of three parts: Transformer blocks for reasoning, (optional) Deconvolution/Pool for rescaling and Multilayer perceptron for prediction. Concretely, based on the encoded visible patches $z_{v}^{l}$ from $l$ -th layer and the mask tokens $\{e_{[\mathbb{M}]}^{l}|m_{i}^{l} = 1\}$ , a decoder outputs prediction $\hat{y}^{l}$ that has the same scale as feature map $z_{v}^{l}$ . When the supervision $y^{l}$ has different scale with the feature map $z_{v}^{l}$ (e.g., in the columnar architectures), the decoded prediction can not match the supervision. To this end, we use the deconvolution/pool operations to rescale + +the prediction $\hat{y}^l$ for matching the supervision $y^{l}$ . For example, we can rescale the $14\times 14$ prediction to the scale of $56\times 56$ with twice deconvolution operations or to the scale of $7\times 7$ with average pool. To avoid excessive computational overhead, we use tiny decoders containing one Transformer block with small embedding dimension. + +The training loss is the weighted summation of the reconstruction losses at the chosen layers + +$$ +\mathcal {L} _ {\text {L o c a l M I M}} = - \sum_ {l \in \mathcal {I}} w _ {l} \cdot \sum_ {i = 1} ^ {N _ {l}} m _ {i} ^ {l} \cdot \ln P \left(y _ {i} ^ {l} \mid \hat {y} _ {i} ^ {l}\right) \tag {2} +$$ + +where $\mathcal{I}$ is the set of chosen layers, $w_{l}$ is the coefficient of each local loss, and mask $m^l$ is calculated by up/down sampling the initial mask $m$ . These local losses guide the patches at multiple chosen layers to conduct semantic interactions under different scales, which not only accelerates the learning of multiple layers but also facilitates multiscale semantical understanding to the input. As shown in Fig. 3, compared with existing models, our LocalMIM has larger NMI values at lower layers, which means the attention maps depend more strongly on the query patches. + +
ModelPT EpochPT HoursmIoU
Supervised--47.4
MoCo v3 [12]300-47.3
BEiT [2]800192047.1
MAE [21]1600176048.1
MaskFeat [47]1600624048.8
PeCo [15]800-48.5
CAE [11]800224048.8
LocalMIM-HOG1600112049.5
+ +# 4. Experiments + +In this section, we first evaluate our LocalMIM model on classification, detection and segmentation tasks, and then provide some ablation studies for deep understanding. + +# 4.1. Classification on ImageNet-1K + +Settings. Both pre-training and fine-tuning are conducted on ImageNet-1K [40] dataset under the $224 \times 224$ resolution. We mainly examine two representative architectures, columnar ViT [16] and pyramidal Swin [33]. The input images are patchified with patch size $p = 16$ for ViT and $p = 4$ for Swin, and the obtained patches are randomly masked with ratio $r = 0.75$ by default. We use both HOG feature [47] and normalized pixels [21] as the supervision signals. For Swin, we introduce the reconstruction task after each stage and the supervision signals have the same scale with the output feature maps, e.g., the chosen layers $\mathcal{I} = \{2,4,22,24\}$ and the scales of supervision are $\{28^2,14^2,7^2,7^2\}$ on Swin-B. Each decoder contains one transformer block with the embedding dimension of 128 and 4 attention heads. For ViT, the chosen layers $\mathcal{I} = \{2,4,4 + n,6 + n\}$ , e.g., $n = 6$ on ViT-B and $n = 18$ on ViT-L. The scales of supervision are $\{56^2,28^2,14^2,7^2\}$ . Each decoder contains one transformer block with the embedding dimension of 256 and 8 attention heads. The pretraining and fine-tuning schedules mostly follow [21, 26], and more detailed settings can be found in the Appendix. + +Results. We compare our LocalMIM with existing MIM models and examine both pre-training efficiency and top-1 fine-tuning accuracy on ImageNet-1K. The results are provided in Fig. 1 and Table 1. For fair comparison, we estimate the pre-training duration of each model on the same machine with one Tesla V100-32G GPU, CUDA 10.2 and Pytorch 1.8. We report the running time on single GPU, denoted as 'GPU Hours', see the Appendix for more details. On ViT-B, LocalMIM achieves the best results of MAE [21] and MaskFeat [47] with $3.1 \times$ and $5.6 \times$ accel + +Table 2. Semantic segmentation on ADE20K using UperNet with ViT-B backbone. Our LocalMIM achieves better results than previous MIM models with less pre-training duration. + +
ModelPT EpochPT HoursAPbAPm
Supervised30084048.543.2
SimMIM192 [50]800144050.444.4
GreenMIM [26]80064050.044.1
LocalMIM-HOG40044050.744.9
+ +Table 3. Object detection and instance segmentation on COCO. We fine-tune Mask R-CNN end-to-end with Swin-B backbone. + +eration respectively. On Swin-B, LocalMIM achieves those of $\mathrm{SimMIM}_{192}$ [50] and GreenMIM [26] with $3.6\times$ and $6.4\times$ acceleration respectively. In terms of final top-1 finetuning accuracy, LocalMIM achieves $84.0\%$ with ViT-B and $84.1\%$ with Swin-B. Compared with previous best results, LocalMIM achieves comparable performance with significantly less pre-training duration. Note that PeCo [15] uses more advanced feature descriptor, a pre-trained codebook, which introduces additional pre-training burden. Besides, on Swin-B, with 100 epochs pre-training and 100 epochs supervised fine-tuning LocalMIM achieves $83.8\%$ top-1 accuracy and takes about 350 GPU Hours, while 300 epochs supervised training from scratch only achieves $83.5\%$ top-1 accuracy and takes about 720 GPU Hours. This means even for the high-level classification task, self-supervision learning is more efficient and effective than supervised learning. Similar results can also be found on ViT. + +# 4.2. Downstream tasks + +We transfer our pre-trained backbone to semantic segmentation on ADE20K [53] and object detection and segmentation on COCO [31]. + +Semantic segmentation on ADE20K. We conduct semantic segmentation on ADE20K using UperNet [48] and following the code in [2, 21]. See the Appendix for fine-tuning details. The results are shown in Table 2. With significantly less computation burden, LocalMIM outperforms the state-of-the-art result by 0.7. Besides, LocalMIM has the same top-1 fine-tuning accuracy with MaskFeat [47] on ImageNet-1K, but has better semantic segmentation performance, which means our local multi-scale reconstruction facilitates the learning of multi-scale semantic knowledge. + +Object detection and instance segmentation on COCO. We fine-tune Mask R-CNN [23] on COCO with Swin-B backbone. Following [26], we also use the code base and $3 \times$ fine-tuning schedule from the supervised Swin, where the model is fine-tuned for 36 epochs. See the Appendix for fine-tuning details. We report box AP $(\mathrm{AP}^b)$ for object detection and mask AP $(\mathrm{AP}^m)$ for instance segmentation. The results are shown in Table 3. With no labeling burden, our LocalMIM outperforms supervised pre-training by 2.2 $\mathrm{AP}^b$ and $1.7\mathrm{AP}^m$ . Compared to SimMIM192 and GreenMIM, our LocalMIM achieves $+0.3$ and $+0.7$ gain respectively for detection, and $+0.5$ and $+0.8$ gain for segmenta + +
targetViT-BSwin-B
baselines82.983.2
Pixels83.083.7
HOG83.383.8
+ +(a) Reconstruction target. HOG is more suitable as multi-scale supervisions than pixels. + +
locationsbackboneacc
GreenMIMSwin-B83.2
[24]83.3
[22, 24]83.4
[4, 22, 24]Swin-B83.6
[2, 4, 22, 24]83.8
fusion83.5
+ +(d) Locations for Swin. Applying reconstructions at all stages mostly accelerates learning. + +
ratioViT-BSwin-B
0.4083.283.7
0.6083.183.7
0.7583.383.8
0.9083.083.5
+ +(b) Mask ratio. Masking $75\%$ patches works the best for both ViT and Swin. + +
locationsbackboneacc
MAEViT-B82.9
[12]83.0
[3, 6, 9, 12]ViT-B83.1
[2, 4, 10, 12]83.3
[1, 2, 11, 12]82.9
+ +(e) Locations for ViT. The chosen layers for local losses obviously affect performance. + +
decoderViT-BSwin-B
512D - 16H83.3 (1.0h)83.8 (1.6h)
256D - 8H83.3 (0.7h)83.7 (1.3h)
128D - 4H83.2 (0.6h)83.8 (1.1h)
+ +(c) Decoder design. A tiny decoder performs as well as the larger one but is more efficient. + +
scalesbackboneacc
MAEViT-B82.9
[142, 142, 142, 142]83.0
[282, 142, 72, 72]ViT-B83.2
[562, 282, 142, 72]83.3
[72, 142, 282, 562]83.1
+ +(f) Scales for ViT. Using the supervisions from fine-scale to coarse-scale works the best. + +Table 4. Ablation studies with ViT-B and Swin-B on ImageNet-1K. We report the top-1 fine-tuning accuracy (%) and the default setting is: the reconstruction target is HOG feature, the mask ratio is $75\%$ , the decoder contains one Transformer block with dimension 256 and 8 heads for ViT and dimension 128 and 4 heads for Swin, the local reconstructions are applied at all stages of Swin and [2, 4, 10, 12]-th layer of ViT, the scales of supervision are $[28^2, 14^2, 7^2, 7^2]$ for Swin and $[56^2, 28^2, 14^2, 7^2]$ for ViT, and the pre-training length is 100 epochs. + +tion with significantly less pre-training duration. + +# 4.3. Ablation studies + +We ablate our LocalMIM using the default setting in Table 4. MAE [21] and GreenMIM [26] are our main baselines on ViT and Swin respectively, since we all adopt the asymmetric encoder-decoder strategy. + +Reconstruction target. We examine two types of signal as supervisions: HOG feature [47] and normalized pixel [21], and the results are shown in Table 4a. HOG feature is more suitable as multi-scale supervisions than normalized pixel on both ViT and Swin, since it contains more refined semantic information. Compared to the columnar ViT with single-scale feature maps, our multi-scale supervisions achieve more improvement on the pyramidal Swin. Since we need to rescale the predictions in the columnar architectures, it is harder to guide the learning of multi-scale information, e.g., we need more information-refined signals (HOG) and obtain relatively less improvement. + +Mask ratio. We explore different mask ratios, ranging from 0.4 to 0.9, and the results are provided in Table 4b. Similar to [21, 26], our LocalMIM is robust to the mask ratio from 0.4 to 0.75, and excessively large one (0.9) harms the performance since it over-complicates the reconstruction tasks. + +Decoder design. As shown in [21, 26], a small single-block decoder can achieve the similar fine-tuning accuracy to the heavier one with multiple blocks. However, since we need to use the decoder for each reconstruction task to process the signals of specified scale, their decoder which has the embedding dimension of 512 and 16 attention heads is still heavy. To this end, we explore whether a smaller decoder can still handle the local reconstruction tasks for + +learning good representations. One Transformer block is the minimal requirement to propagate information from visible patches to mask tokens, so we still use one Transformer block but narrow the embedding dimension and reduce the number of attention heads. The results (top-1 fine-tuning accuracy and GPU Hours per epoch) are shown in Table 4c, where $dD - hH'$ represents the decoder with the embedding dimension of $d$ and $h$ attention heads. Surprisingly, a tiny decoder can achieve the same even better fine-tuning performance, but is significantly more efficient. + +Locations for local reconstructions. An important design of our LocalMIM is the locations in the encoder where we introduce local reconstructions. Many pyramidal architectures like Swin already divide the entire backbone into multiple stages, so we introduce multi-scale reconstructions at their output locations. From the last stage, we gradually include the lower stages and the results are shown in Table 4d. Note that the 2-th, 4-th, 22-th and 24-th layer correspond to the output location of the four stages respectively. We also provide the result of fusion method [18] which fuses the output features from each stage for reconstruction. As we can see, introducing reconstructions at local stages can accelerate representation learning, and applying to all stages works the best. Our LocalMIM also outperforms the fusion method which only uses the single-scale reconstruction, and this verifies the importance of multi-scale reconstructions. Further, for the columnar ViT, we follow the experience from Swin and also select four local layers. Concretely, we consider the uniform division [3, 6, 9, 12], Swin-style division [2, 4, 10, 12] and a extreme division [1, 2, 11, 12]. The results are shown in Table 4e. The performance on columnar ViT is sensitive to the chosen local layers, and the Swin-style division $[2, 4, 4 + n, 6 + n]$ works the best. + +![](images/384b5deb4a41f29823644f672ba46519d5b558711e78c2b84a9710e7e718db81.jpg) +Figure 4. Visualization of the attention maps for different query points, marked with red boxes. LocalMIM can distinguish semantical regions using self-attention mechanism. + +![](images/fe658e280a009b0a448b6b52f5b8ce8eed792aa4fd48e4011b46ea451eb91f9d.jpg) + +Scales of supervisions. Since the pyramidal architectures hard-code the multi-scale property of the feature maps, we directly use the supervisions of same scale as the feature maps for compatibility, i.e., we use $[28^2, 14^2, 7^2, 7^2]$ for Swin. For columnar ViT, we consider single-scale supervisions $[14^2, 14^2, 14^2, 14^2]$ , fine-to-coarse multi-scale supervisions $[28^2, 14^2, 7^2, 7^2]$ and $[56^2, 28^2, 14^2, 7^2]$ , and coarse-to-fine multi-scale supervisions $[7^2, 14^2, 28^2, 56^2]$ . The results are shown in Table 4f. The multi-scale supervisions achieve better fine-tuning accuracy than the single-scale ones, which means the multiple local layers require the supervision of different scales. Further, the fine-to-coarse supervisions outperform the coarse-to-fine ones, which is consistent with the design priors for visual architectures [33] and biological visual processing [28]. + +Analysis of self-attention mechanism. To qualitatively illustrate the behavior of self-attention mechanism in LocalMIM, we visualize the attention maps for different query positions within an image, as shown in Fig. 4. We use each chosen patch as query and show the patches it attends to. We select the attention maps at 2-th, 4-th, 10-th and 12-th layers of a pre-trained ViT-B/16 backbone to show their changes during forward inference. For object-centric images, LocalMIM can distinguish the foreground object from the background. For more complex multi-object images, + +
modelbackboneGPU Hours/Ep.acc
LocalMIMViT-B0.783.3
w/ isolated grad0.783.0
LocalMIMSwin-B1.183.8
w/ isolated grad1.183.7
+ +Table 5. Training LocalMIM with isolated gradients achieves similar performance with global back-propagation. + +LocalMIM can effectively separate different objects without any task-specific supervision, which means the attention maps are query-adaptive. On the other hand, the patches at lower layers typically more focus on their neighboring regions and capture low-level information, while those at upper layers attend to a wide range of semantically related regions and capture high-level shape information. + +Gradient-isolated training. Inspired by locally supervised learning [34, 35], we remove global back-propagation and stop the gradients after each chosen local layer used for reconstruction. The entire backbone is thus divided into multiple gradient-isolated parts. The results are shown in Table 5. Surprisingly, the gradient-isolated training achieves similar performance to global back-propagation, which further verifies the effectiveness of our multi-scale reconstructions for guiding the local layers. It even requires no gradient information from the upper layers. This observation also shows the promise of our local multi-scale reconstruction for the decoupled training of neural networks, which allows for training very deep networks without memory concern and reduces explosive or vanishing gradients. + +# 5. Conclusions + +We present a novel and efficient pretext task, local multi-scale reconstruction, where the lower layers and upper layers reconstruct the fine-scale and coarse-scale supervisions from the input respectively. We obtain the multi-scale supervisions by first dividing the input under different scales and then extracting supervisions with appropriate feature descriptors. Our model needs no extra pre-trained codebook and no costly forward inference of teacher networks during pre-training. The asymmetric encoder-decoder strategy with tiny encoders also have small computational burden, our model thus can be trained quickly. The novel pretext task further accelerates the representation learning, especially for the pyramidal architectures. All these designs allow our model to achieve comparable performance to existing models but with significantly less pre-training burden. The gradient-isolated training also verifies the effectiveness of our novel task design in guiding the local layers. + +Acknowledgment. We gratefully acknowledge the support of MindSpore [27], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor. + +# References + +[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022. 1, 2, 4, 5 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2022. 1, 2, 3, 4, 5, 6 +[3] Yoshua Bengio, Dong-Hyun Lee, Jorg Bornschein, Thomas Mesnard, and Zhouhan Lin. Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156, 2015. 3 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[5] Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022.4 +[6] Natalia Caporale, Yang Dan, et al. Spike timing-dependent plasticity: a hebbian learning rule. Annual review of neuroscience, 31(1):25-46, 2008. 2 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 5 +[8] Jun Chen, Ming Hu, Boyang Li, and Mohamed Elhoseiny. Efficient self-supervised vision pretraining with local masked reconstruction. arXiv preprint arXiv:2206.00790, 2022. 1, 2, 5 +[9] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 3 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[11] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.5, 6 +[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 5, 6 +[13] Francis Crick. The recent excitement about neural networks. Nature, 337(6203):129-132, 1989. 3 +[14] Yang Dan and Mu-ming Poo. Spike timing-dependent plasticity of neural circuits. Neuron, 44(1):23-30, 2004. 3 +[15] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and + +Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. 1, 2, 4, 5, 6 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 2, 3, 6 +[17] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6824-6835, 2021. 3 +[18] Peng Gao, Teli Ma, Hongsheng Li, Jifeng Dai, and Yu Qiao. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892, 2022. 2, 3, 5, 7 +[19] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Yunhe Wang, and Chang Xu. Fastmim: Expediting masked image modeling pre-training for vision. arXiv preprint arXiv:2212.06593, 2022. 1 +[20] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 2 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1, 2, 3, 4, 5, 6, 7 +[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[23] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 6 +[24] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921-1930, 2019. 3 +[25] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification. In ACL, 2018. 1, 4 +[26] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling. arXiv preprint arXiv:2205.13515, 2022. 1, 2, 3, 4, 5, 6, 7 +[27] Huawei. Mindspace. https://www.mindspore.cn/, 2020.8 +[28] DH Hubel and TN Wiesel. Receptive fields of optic nerve fibres in the spider monkey. The Journal of physiology, 154(3):572, 1960. 3, 8 + +[29] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 1 +[30] Xiang Li, Wenhai Wang, Lingfeng Yang, and Jian Yang. Uniform masking: Enabling mae pre-training for pyramid-based vision transformers with locality. arXiv preprint arXiv:2205.10063, 2022. 1, 2, 4 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6 +[32] Hao Liu, Xinghua Jiang, Xin Li, Antai Guo, Deqiang Jiang, and Bo Ren. The devil is in the frequency: Geminated gestalt autoencoder for self-supervised visual pre-training. arXiv preprint arXiv:2204.08227, 2022. 1, 2 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 3, 6, 8 +[34] Sindy Löwe, Peter O'Connor, and Bastiaan Veeling. Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems, 32, 2019. 2, 8 +[35] Arild Nøkland and Lars Hiller Eidnes. Training neural networks with local error signals. In International conference on machine learning, pages 4839-4850. PMLR, 2019. 2, 8 +[36] Namuk Park and Songkuk Kim. How do vision transformers work? In International Conference on Learning Representations, 2022. 2, 5 +[37] Myeongjang Pyeon, Jihwan Moon, Taeyoung Hahn, and Gunhee Kim. Sedona: Search for decoupled neural networks toward greedy block-wise learning. In International Conference on Learning Representations, 2021. 2 +[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3 +[39] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 3 +[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6 +[41] Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for combining multiple partitions. Journal of machine learning research, 3(Dec):583-617, 2002. 4 +[42] Zhiqing Sun, Zhuohan Li, Haoqing Wang, Di He, Zi Lin, and Zhihong Deng. Fast structured decoding for sequence models. Advances in Neural Information Processing Systems, 32, 2019. 2 +[43] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In Zhi-Hua + +Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1075-1081. International Joint Conferences on Artificial Intelligence Organization, 2021. Main Track. 3 +[44] Haoqing Wang, Xun Guo, Zhi-Hong Deng, and Yan Lu. Rethinking minimal sufficient representation in contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16041-16050, 2022. 2 +[45] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 1, 2, 3 +[46] Yulin Wang, Zanlin Ni, Shiji Song, Le Yang, and Gao Huang. Revisiting locally supervised learning: an alternative to end-to-end training. In International Conference on Learning Representations, 2021. 2 +[47] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 1, 2, 4, 5, 6, 7 +[48] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6 +[49] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 1, 2 +[50] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 1, 2, 4, 5, 6 +[51] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3008, 2021. 3 +[52] Tianyi Zhang, Felix Wu, Arzoo Katiyar, Kilian Q Weinberger, and Yoav Artzi. Revisiting few-sample bert finetuning. In International Conference on Learning Representations, 2021. 1, 4 +[53] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6 +[54] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 1, 2, 4, 5 \ No newline at end of file diff --git a/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/images.zip b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..09aaeb0221244ae6ba1731d657e6f8712b24d5aa --- /dev/null +++ b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98e43ee12d6133aec9860810296c6b3a0488714b6e7958d4a9bc4bc7f43629f3 +size 493780 diff --git a/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/layout.json b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e14d50476532f79ad8502be5183e909201481ace --- /dev/null +++ b/2023/Masked Image Modeling With Local Multi-Scale Reconstruction/layout.json @@ -0,0 +1,8313 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 495, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 495, + 120 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 495, + 120 + ], + "type": "text", + "content": "Masked Image Modeling with Local Multi-Scale Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "spans": [ + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": "Haoqing Wang" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": ", Yehui Tang" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": ", Yunhe Wang" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": ", Jianyuan Guo" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": ", Zhi-Hong Deng" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": ", Kai Han" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": "School of Intelligence Science and Technology, Peking University " + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab {wanghaoqing,yhtang,zhdeng} " + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": " pku.edu.cn, {yunhe.wang, kai.han} " + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 72, + 142, + 521, + 186 + ], + "type": "text", + "content": " huawei.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 239, + 289, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 289, + 525 + ], + "type": "text", + "content": "Masked Image Modeling (MIM) achieves outstanding success in self-supervised representation learning. Unfortunately, MIM models typically have huge computational burden and slow learning process, which is an inevitable obstacle for their industrial applications. Although the lower layers play the key role in MIM, existing MIM models conduct reconstruction task only at the top layer of encoder. The lower layers are not explicitly guided and the interaction among their patches is only used for calculating new activations. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to multiple local layers including lower and upper layers. Further, since the multiple layers expect to learn the information of different scales, we design local multiscale reconstruction, where the lower and upper layers reconstruct fine-scale and coarse-scale supervision signals respectively. This design not only accelerates the representation learning process by explicitly guiding multiple layers, but also facilitates multi-scale semantical understanding to the input. Extensive experiments show that with significantly less pre-training burden, our model achieves comparable or better performance on classification, detection and segmentation tasks than existing MIM models. Code is available with both MindSpore and PyTorch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "content": "Recently, Masked Image Modeling (MIM) [2, 21, 50] achieves outstanding success in the field of self-supervised visual representation learning, which is inspired by the Masked Language Modeling (MLM) [4, 29] in natural language processing and benefits from the development of vision transformers [16, 33, 45]. MIM learns semantic representations by first masking some parts of the input and then predicting their signals based on the unmasked parts, e.g., normalized pixels [21, 50], discrete tokens [2, 15], HOG feature [47], deep features [1, 54] or frequencies [32, 49]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 214, + 547, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 547, + 453 + ], + "type": "text", + "content": "Despite superior performance on various downstream tasks, these models have huge computational burden and slow learning process [26]. They typically require thousands of GPU Hours for pre-training on ImageNet-1K to get generalizing representations. Since we expect to pretrain these models on more massive amount of unlabeled data (e.g., free Internet data) to obtain more generalizing representations in practice, the pre-training efficiency is an inevitable bottleneck limiting the industrial applications of MIM. How to accelerate the representation learning in MIM is an important topic. To this end, MAE [21] pioneered the asymmetric encoder-decoder strategy, where the costly encoder only operates few visible patches and the lightweight decoder takes all the patches as input for prediction. Further, GreenMIM [26] extends the asymmetric encoder-decoder strategy to hierarchical vision transformers (e.g., Swin [33]). Besides, [8, 19, 30] shrinks the input resolution to lessen the input patches, thereby reducing the computational burden. However, they all aim to accelerate the encoding process rather than the representation learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 457, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 547, + 685 + ], + "type": "text", + "content": "In MIM, the learning of upper layers depends on that of lower ones during pre-training, since the upper-layer features are calculated from the lower layers. Besides, during fine-tuning the upper layers are typically tuned quickly to adapt to the downstream task while the lower ones change more slowly and need to be well-learned [2, 25, 52]. Even fine-tuning only the several upper layers and freezing the others can obtain similar performance [21]. Therefore, the lower layers of encoder play the key role in MIM. However, all existing MIM models only conduct reconstruction task at the top layer of encoder and the lower ones are not explicitly guide, thus the interaction among their patches is only used for calculating the activations of the next layer. Considering the reconstruction task requires non-trivial inter-patch interactions to reason target signals, we apply it to both lower and upper layers to explicitly guide them and thus accelerate the overall learning process. Using tiny decoder is sufficient for each local reconstruction task and does not significantly increase the computational burden." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "How to properly conduct reconstruction tasks at multiple local layers is a non-trivial problem. For example, ap-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2122" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 294, + 202 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 294, + 202 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 294, + 202 + ], + "type": "image", + "image_path": "e09ee8397bd0e0523b8b82cb7af81c8dc4c76d7d07bdc7ec05c2d89699bdf6a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 547, + 234 + ], + "type": "text", + "content": "Figure 1. Top-1 fine-tuning accuracy on ImageNet-1K vs. Pre-training duration. The duration is estimated on a machine with one Tesla V100-32G GPU, CUDA 10.2 and PyTorch 1.8. 'GPU Hours' is the running time on single GPU." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 524, + 202 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 524, + 202 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 524, + 202 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 524, + 202 + ], + "type": "image", + "image_path": "b956d76e31182bdd429661f85072584ea7fa4bc4f50977fad39967018b60440e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "type": "text", + "content": "plying the top-layer reconstruction task to carefully chosen local layers of ViT [16] can not achieve meaningful improvement. In general, the lower layers exploit low-level information and the upper ones learn high-level information [18, 36], so it is not appropriate to use the supervision signals of same scale for multiple local reconstruction tasks. Here 'scale' is the spatial size of the supervision signals calculated from the divided input regions, e.g., the signals from the " + }, + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "type": "inline_equation", + "content": "p \\times p" + }, + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "type": "text", + "content": " regions in an input of " + }, + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "type": "text", + "content": " resolution has the scale of " + }, + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\frac{H}{p} \\times \\frac{W}{p}" + }, + { + "bbox": [ + 46, + 246, + 289, + 534 + ], + "type": "text", + "content": ". The fine-scale and coarse-scale supervisions typically contain low-level and high-level information of the input respectively, and these multi-scale supervisions from input are widely ignored by existing MIM models. To this end, we propose local multi-scale reconstruction where the lower and upper layers reconstruct fine-scale and coarse-scale supervisions respectively. This design not only accelerates the representation learning process, but also facilitates multi-scale semantical understanding to the input. When the decoded predictions have different scale with the supervisions (e.g., on ViT), we use the deconvolution/pool operations to rescale them. We also apply the asymmetric encoder-decoder strategy [21, 26] for quick encoding. Our model, dubbed as LocalMIM, are illustrated in Fig. 2 (a). Overall, we summarize our contributions as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 541, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 541, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 541, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 58, + 541, + 287, + 578 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first work in MIM to conduct local reconstructions and use multiscale supervisions from the input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 586, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 586, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 58, + 586, + 287, + 609 + ], + "type": "text", + "content": "- Our model is architecture-agnostic and can be used in both columnar and pyramid architectures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "text", + "content": "- From extensive experiments, we find that 1) LocalMIM is more efficient than existing MIM models, as shown in Fig. 1 and Table 1. For example, LocalMIM achieves the best MAE result with " + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "3.1 \\times" + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "text", + "content": " acceleration on ViT-B and the best GreenMIM result with " + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "6.4 \\times" + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "text", + "content": " acceleration on Swin-B. 2) In terms of top-1 fine-tuning accuracy on ImageNet-1K, LocalMIM achieves " + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "text", + "content": " using ViT-B and " + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "84.1\\%" + }, + { + "bbox": [ + 58, + 618, + 288, + 714 + ], + "type": "text", + "content": " using Swin" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 246, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 246, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 324, + 246, + 545, + 293 + ], + "type": "text", + "content": "B with significantly less pre-training duration than existing MIM models. The obtained representations also achieve better generalization on detection and segmentation downstream tasks, as shown in Table 2 and 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 303, + 398, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 398, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 398, + 316 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 626 + ], + "type": "text", + "content": "Masked Image Modeling. With the development of vision transformers [16, 20, 33, 42, 45], Masked Image Modeling (MIM) gradually replaces the dominant position of contrastive learning [10, 22, 44] in visual self-supervised representation learning due to its superior fine-tuning performance in various visual downstream tasks. Many target signals have been designed for the mask-prediction pretext task in MIM, such as normalized pixels [21, 50], discrete tokens [2, 15], HOG feature [47], deep features [1, 54] or frequencies [32, 49]. However, they are all only applied as single-scale supervisions for reconstruction. An inevitable bottleneck for the industrial applications of MIM is that these models typically require huge computational resources and long pre-training duration. To this end, some works accelerate the encoding process via the asymmetric encoder-decoder strategy [21, 26] or lessening the input patches [8, 30]. Only accelerating the encoding process sometimes doesn't really speed up the representation learning, like GreenMIM vs. SimMIM192 in Fig. 1. In this work, we use local multi-scale reconstructions to explicitly guide multiple lower layers and thus accelerate the overall learning process. Our method is compatible with the above quick encoding approaches. ConvMAE [18] fuses the feature of local layers for the final reconstruction to explicitly guide them, but still applies single-scale supervision signals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Locally supervised learning. Considering the biological brain learns mainly based on local information [6], some works [34, 35, 37, 46] used local error signals to greedily optimize individual blocks of the backbone without global back-propagation. This greedy training procedure significantly reduces memory overhead, but can not accelerate model learning. In this work, we also optimize error at" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2123" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 232 + ], + "type": "image", + "image_path": "d688c406129a1b7f8e1cba475142c9340370cd1956635d60d515f2f0fec09a86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 548, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 548, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 548, + 293 + ], + "type": "text", + "content": "Figure 2. Illustration of local multi-scale reconstruction. a) We randomly mask some input patches and then predict their supervision signals of different scales at multiple local layers. The multi-scale supervisions are obtained by first dividing the input under different scales and then extracting signals with some feature descriptors. The lower layers are responsible for fine-scale reconstruction and the upper ones are responsible for coarse-scale reconstruction. We also use the asymmetric encoder-decoder strategy for quick encoding. b) The decoder consists of three parts: Transformer blocks, Deconvolution/Pool (optional) and Multilayer perceptron." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 289, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 400 + ], + "type": "text", + "content": "multiple local layers but still use global back-propagation for end-to-end training. Compared with existing MIM models, our local reconstruction is more biologically plausible since the human brain prefers local learning rules [3, 13, 14]. Feature distillation [24, 39] also explicitly guides the local layers, but needs costly pre-trained or momentum teacher. In fact, the multi-scale supervisions from original input are sufficient to guide local layers and also readily available." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "type": "text", + "content": "Multi-scale property. Biological visual perception is hierarchical and multi-scale [28]. Moreover, multi-scale features are also useful for many visual tasks [9,23,38]. To this end, many advanced vision transformers [17,33,43,45,51] hard-code the multi-scale property to the architectures and boost the performance. However, global single-scale reconstruction is not enough for guiding multiple local layers to learn multi-scale information. In this work, beyond the multi-scale features, we further introduce multi-scale supervisions to soft-code this property. This new design is not limited to specific architectures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 548, + 96, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 96, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 96, + 561 + ], + "type": "text", + "content": "3. Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 569, + 187, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 187, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 187, + 582 + ], + "type": "text", + "content": "3.1. Masked Image Modeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 588, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 649 + ], + "type": "text", + "content": "In MIM, the models first mask some parts of the input image and then predict their information based on the un-masked observation. For quick encoding, MAE [21] and GreenMIM [26] use the asymmetric encoder-decoder strategy. The major designs are described below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "text", + "content": "Image presentation. Generally, we present an image to the sequence of visual patches as the input to vision transformers [16, 33]. The input image " + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "text", + "content": " is reshaped to " + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "inline_equation", + "content": "N = HW / p^2" + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "text", + "content": " non-overlapping patches " + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "inline_equation", + "content": "x^p \\in \\mathbb{R}^{N \\times (p^2C)}" + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "text", + "content": " is the patch size, " + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 46, + 652, + 288, + 715 + ], + "type": "text", + "content": " is the resolution of in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "text", + "content": "put image and " + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "text", + "content": " is the number of channels. The patches " + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "inline_equation", + "content": "\\{x_i^p\\}_{i = 1}^N" + }, + { + "bbox": [ + 305, + 304, + 546, + 352 + ], + "type": "text", + "content": " are then linearly mapped to the patch embeddings. For retaining the positional information, the patches are typically added with positional embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "text", + "content": "Vision transformers. Unlike the grid operations in the convolutional neural networks, vision transformers [16,33,45] use the stacked multi-head self-attention modules. To obtain the multi-scale feature maps for dense prediction tasks, some advanced vision transformers [33, 45] extend the columnar ViT [16] to the pyramidal structure, where the feature maps change from fine-scale to coarse-scale and are expected to capture the multi-scale information from the input. The patches " + }, + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "inline_equation", + "content": "[x_1^p,\\dots ,x_N^p ]" + }, + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "text", + "content": " are fed to a vision transformer with " + }, + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "text", + "content": " layers and obtain the output feature at " + }, + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "text", + "content": "-th layer " + }, + { + "bbox": [ + 304, + 360, + 546, + 494 + ], + "type": "inline_equation", + "content": "z^{l} = [z_{1}^{l},\\dots ,z_{N_{l}}^{l}],l = 1,\\dots ,L." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "content": "Masking. Given the patch sequence " + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\{x_i^p\\}_{i = 1}^N" + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "content": ", MIM constructs a random mask " + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "inline_equation", + "content": "m\\in \\{0,1\\} ^N" + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "content": " to indicate the masked patches that correspond to " + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "inline_equation", + "content": "m_{i} = 1" + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "content": ". There are two main masking strategies, random masking [21] and block-wise masking [2]. Specially, since the number of patches " + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "content": " decreases through layers in the pyramidal architectures, some works [18,26] first construct the random mask " + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "inline_equation", + "content": "m^L" + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "content": " with the size of " + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "content": ", and then up-sample it to size " + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 501, + 547, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Encoder. Only the visible patches " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "x^{v} = \\{x_{i}^{p}|m_{i} = 0\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " are fed to the encoder and mapped to the latent features " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "z_{v}^{l}, l = 1,\\dots ,L" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ", which significantly reduces compute and memory. Specially, since the local window attention in the pyramidal architectures is not compatible with the incomplete input patches (i.e., only visible patches), GreenMIM [26] proposes the Optimal Grouping algorithm and the Group Window Attention scheme. After pre-training, the encoder is used in various downstream tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2124" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "Decoder. The decoder takes as the input of both encoded visible patches " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "z_v^L" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " and mask tokens " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\{e_{[\\mathbb{M}]}|m_i^L = 1\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ", where the mask token " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "e_{[\\mathbb{M}]}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " is a shared and learnable vector. The positional embeddings are also added to them for providing the location information. Since the decoder is only used during pre-training to output the prediction " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\hat{y}^L" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ", it can be any architectures which support the global information propagation among patches, e.g., a series of ViT or Swin blocks. Many works [21, 26, 30] show that using lightweight decoders is sufficient to learn generalizing representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 287, + 232 + ], + "type": "text", + "content": "Global reconstruction. All existing MIM models predict the supervision signal " + }, + { + "bbox": [ + 46, + 195, + 287, + 232 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 195, + 287, + 232 + ], + "type": "text", + "content": " based on the final output feature " + }, + { + "bbox": [ + 46, + 195, + 287, + 232 + ], + "type": "inline_equation", + "content": "z_v^L" + }, + { + "bbox": [ + 46, + 195, + 287, + 232 + ], + "type": "text", + "content": " of the encoder and minimize a global reconstruction loss" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 239, + 287, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 239, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 96, + 239, + 287, + 273 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M I M} = - \\sum_ {i = 1} ^ {N _ {L}} m _ {i} ^ {L} \\cdot \\ln P \\left(y _ {i} \\mid \\hat {y} _ {i} ^ {L}\\right) \\tag {1}", + "image_path": "f4ca151682137ef26f1300492fce8fda56e2f3d24d53dfdf5c1440d73188ac1b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "type": "text", + "content": "where the loss is calculated on the masked patches with " + }, + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "type": "inline_equation", + "content": "m_{i}^{L} = 1" + }, + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "type": "inline_equation", + "content": "P(\\cdot |\\cdot)" + }, + { + "bbox": [ + 47, + 281, + 287, + 318 + ], + "type": "text", + "content": " can be the Gaussian or Dirichlet distributions for regression or classification losses respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 324, + 183, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 183, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 183, + 337 + ], + "type": "text", + "content": "3.2. Analysis and motivation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 462 + ], + "type": "text", + "content": "For pre-training, the upper-layer features are calculated from the lower layers, so the well-learned lower layers can propagate semantical knowledge to the upper ones and facilitate their learning. For fine-tuning, the upper layers typically adapt quickly to specific downstream tasks, while the lower ones change more slowly and need to be sufficient learned during pre-training [2,25,52]. Even fine-tuning only the several upper layers and freezing the others can obtain comparable performance [21]. Therefore, the lower layers of the encoder play the key role in MIM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "content": "After patchification and linearly projection, the initial patch embeddings lose the inter-patch semantical relations. The self-attention mechanism in vision transformers is responsible for learning these relations by inter-patch interaction and build a better representation space than pixel space [5]. Further, since the self-attention mechanism has the computational complexity with a quadratic dependence on patch number " + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "content": ", it is difficult to learn the inter-patch interactions, especially for the lower layers of pyramidal architectures where the small patch size " + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "content": " leads to huge " + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "content": ". However, under the global reconstruction loss, the interpatch interaction at lower layers is not explicitly guided, and the simple task of calculating new activations is not sufficient to guide it. As a result, it is hard for the patches at lower layers to learn inter-patch relations. Concretely, we examine Normalized Mutual Information (NMI) [41] between query and key patches at each layer, where the high NMI value means the attention maps strongly depend on the query patches. Intuitively, the semantical representations should have highly query-adaptive attentions, i.e., the different query patches faithfully attend to their semantically" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 324, + 72, + 528, + 205 + ], + "blocks": [ + { + "bbox": [ + 324, + 72, + 528, + 205 + ], + "lines": [ + { + "bbox": [ + 324, + 72, + 528, + 205 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 528, + 205 + ], + "type": "image", + "image_path": "05f26d05ff73645e0850ba2457b7c48d9276b00002b2710f815a495d27815808.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 216, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 239 + ], + "type": "text", + "content": "Figure 3. Normalized Mutual Information (NMI) between query and key patches at each layer of a pre-trained ViT-B." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "text", + "content": "related regions. This is an advantage of the self-attention mechanism. As shown in Fig. 3, existing MIM models with global loss have small NMI value at lower layers, which means their patches have less query-adaptive attentions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 300, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 372 + ], + "type": "text", + "content": "The reconstruction task requires holistic reasoning among patches to predict the masked signals and thus obtains semantic understanding to the input. Since this challenging task facilitates non-trivial inter-patch interactions, we apply it at multiple local layers, including both lower and upper ones to explicitly guide them all." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 380, + 479, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 479, + 392 + ], + "type": "text", + "content": "3.3. Local multi-scale reconstruction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "text", + "content": "In MIM, the supervision signals for a reconstruction task are directly calculated from the input. Concretely, we evenly divide the input image " + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "text", + "content": " into nonoverlapping regions " + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\{x_i \\in \\mathbb{R}^{p \\times p \\times C}\\}_{i=1}^{HW/p^2}" + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "text", + "content": " and use some feature descriptor " + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "text", + "content": " to extract the supervision signal " + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "inline_equation", + "content": "y_i = \\pi(x_i)" + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "text", + "content": ". To learn generalizing representations, many feature descriptors have been designed, such as pixel normalization [21, 50], HOG [47] and the pre-trained or momentum teacher [1,2,15,54]. We define the scale of supervision " + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "text", + "content": " as its spatial size " + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\frac{H}{p} \\times \\frac{W}{p}" + }, + { + "bbox": [ + 304, + 399, + 545, + 653 + ], + "type": "text", + "content": ". For a given input, the fine-scale supervisions from finely-divided input regions typically contain the low-level semantical information of the input, like corners, edges or textures. Relatively, the coarse-scale supervisions capture high-level semantical information of the input, like the shape of partial or whole object. Intuitively, multi-scale supervisions can guide representation learning better than the common single-scale ones due to their richer semantic information. In this work, we mainly consider the feature descriptors which are readily available without extra pre-training burden and costly forward inference of teacher networks, e.g., pixel normalization and HOG." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "As shown in Table 4f, directly applying the top-layer reconstruction task at carefully chosen local layers of ViT can not achieve meaningful improvement, where each local task predicts the supervisions of same scale. In fact, the lower and upper layers expect to learning low-level and high-level" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2125" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 519, + 407 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 519, + 407 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 519, + 407 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 519, + 407 + ], + "type": "table", + "html": "
ModelBackbone# ParamsPT EpochGPU Hours/Ep.Total GPU HoursAcc
Scratch, ViTViT-B86M01.5-82.3
Scratch, SwinSwin-B88M02.4-83.5
MoCo v3 [12]ViT-B86M600--83.2
DINO [7]ViT-B86M300--82.8
BEiT [2]ViT-B86M8002.4192083.2
iBOT [54]ViT-B86M40010.1404083.8
MAE [21]ViT-B86M8001.188083.3
MAE [21]ViT-B86M16001.1176083.6
MAE [21]ViT-L307M16001.7272085.9
MaskFeat [47]ViT-B86M16003.9624084.0
CAE [11]ViT-B86M8002.8224083.6
\\( LoMaR^† \\) [8]ViT-B86M16001.4224084.1
data2Vec† [1]ViT-B86M8003.0240084.2
PeCo [15]ViT-B86M800--84.5
LocalMIM-HOGViT-B86M1000.77083.3
LocalMIM-HOGViT-B86M16000.7112084.0
LocalMIM-HOGViT-L307M8001.080085.8
\\( SimMIM_{192} \\) [50]Swin-B88M8001.8144084.0
\\( SimMIM_{192} \\) [50]Swin-L197M8003.0240085.4
GreenMIM [26]Swin-B88M8000.864083.7
GreenMIM [26]Swin-L197M8001.4112085.1
LocalMIM-PixelSwin-B88M1001.010083.7
LocalMIM-HOGSwin-B88M1001.111083.8
LocalMIM-PixelSwin-B88M4001.040084.0
LocalMIM-HOGSwin-B88M4001.144084.1
LocalMIM-HOGSwin-L197M8001.6128085.6
", + "image_path": "79b79e9cfc6aefbf76aae0f564ac6d7d1ac7389ec37819036ab325042d57806e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "text", + "content": "Table 1. Top-1 fine-tuning accuracy on ImageNet-1K. All models are pre-trained and fine-tuned under " + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "text", + "content": " resolution except that " + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "inline_equation", + "content": "\\mathrm{SimMIM}_{192}" + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "text", + "content": " uses " + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "inline_equation", + "content": "192 \\times 192" + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "text", + "content": " resolution for pre-training. " + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 415, + 547, + 439 + ], + "type": "text", + "content": " means using the relative positional encoding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 450, + 288, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 571 + ], + "type": "text", + "content": "information respectively [18, 36], so it is not appropriate to use the single-scale supervisions to guide multiple local layers, even for the columnar architectures with feature maps of the same scale at all layers. To this end, we make the lower layers to reconstruct fine-scale supervisions and the upper ones to reconstruct coarse-scale supervisions. Specially, for the pyramidal architectures which already hardcode the multi-scale property to the features by setting their spatial sizes, we use the supervisions with same scale as the feature maps at the chosen layers for compatibility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "content": "The decoding process under a specific scale is illustrated in Fig. 2 (b). The decoder consists of three parts: Transformer blocks for reasoning, (optional) Deconvolution/Pool for rescaling and Multilayer perceptron for prediction. Concretely, based on the encoded visible patches " + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{v}^{l}" + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "content": "-th layer and the mask tokens " + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{e_{[\\mathbb{M}]}^{l}|m_{i}^{l} = 1\\}" + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "content": ", a decoder outputs prediction " + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}^{l}" + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "content": " that has the same scale as feature map " + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{v}^{l}" + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "content": ". When the supervision " + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "inline_equation", + "content": "y^{l}" + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "content": " has different scale with the feature map " + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{v}^{l}" + }, + { + "bbox": [ + 46, + 578, + 289, + 715 + ], + "type": "text", + "content": " (e.g., in the columnar architectures), the decoded prediction can not match the supervision. To this end, we use the deconvolution/pool operations to rescale" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "content": "the prediction " + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\hat{y}^l" + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "content": " for matching the supervision " + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "inline_equation", + "content": "y^{l}" + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "content": ". For example, we can rescale the " + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "inline_equation", + "content": "14\\times 14" + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "content": " prediction to the scale of " + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "inline_equation", + "content": "56\\times 56" + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "content": " with twice deconvolution operations or to the scale of " + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "inline_equation", + "content": "7\\times 7" + }, + { + "bbox": [ + 304, + 450, + 545, + 521 + ], + "type": "text", + "content": " with average pool. To avoid excessive computational overhead, we use tiny decoders containing one Transformer block with small embedding dimension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 522, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 547 + ], + "type": "text", + "content": "The training loss is the weighted summation of the reconstruction losses at the chosen layers" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 331, + 553, + 545, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 553, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 331, + 553, + 545, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {L o c a l M I M}} = - \\sum_ {l \\in \\mathcal {I}} w _ {l} \\cdot \\sum_ {i = 1} ^ {N _ {l}} m _ {i} ^ {l} \\cdot \\ln P \\left(y _ {i} ^ {l} \\mid \\hat {y} _ {i} ^ {l}\\right) \\tag {2}", + "image_path": "109c95b8663c6f27096f65402fad888fa10b7b7d260777eb09b878c67f25c661.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is the set of chosen layers, " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "w_{l}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is the coefficient of each local loss, and mask " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "m^l" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is calculated by up/down sampling the initial mask " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". These local losses guide the patches at multiple chosen layers to conduct semantic interactions under different scales, which not only accelerates the learning of multiple layers but also facilitates multiscale semantical understanding to the input. As shown in Fig. 3, compared with existing models, our LocalMIM has larger NMI values at lower layers, which means the attention maps depend more strongly on the query patches." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "2126" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 281, + 191 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 281, + 191 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 281, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 281, + 191 + ], + "type": "table", + "html": "
ModelPT EpochPT HoursmIoU
Supervised--47.4
MoCo v3 [12]300-47.3
BEiT [2]800192047.1
MAE [21]1600176048.1
MaskFeat [47]1600624048.8
PeCo [15]800-48.5
CAE [11]800224048.8
LocalMIM-HOG1600112049.5
", + "image_path": "695ef56dee092594a25538672d8a9e4edbc3118c47ba8c7d6e8ca0be332a808c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 247, + 128, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 128, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 128, + 261 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 268, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 287, + 305 + ], + "type": "text", + "content": "In this section, we first evaluate our LocalMIM model on classification, detection and segmentation tasks, and then provide some ablation studies for deep understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 315, + 212, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 212, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 212, + 328 + ], + "type": "text", + "content": "4.1. Classification on ImageNet-1K" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": "Settings. Both pre-training and fine-tuning are conducted on ImageNet-1K [40] dataset under the " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": " resolution. We mainly examine two representative architectures, columnar ViT [16] and pyramidal Swin [33]. The input images are patchified with patch size " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "p = 16" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": " for ViT and " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "p = 4" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": " for Swin, and the obtained patches are randomly masked with ratio " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "r = 0.75" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": " by default. We use both HOG feature [47] and normalized pixels [21] as the supervision signals. For Swin, we introduce the reconstruction task after each stage and the supervision signals have the same scale with the output feature maps, e.g., the chosen layers " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{2,4,22,24\\}" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": " and the scales of supervision are " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\{28^2,14^2,7^2,7^2\\}" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": " on Swin-B. Each decoder contains one transformer block with the embedding dimension of 128 and 4 attention heads. For ViT, the chosen layers " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{2,4,4 + n,6 + n\\}" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "n = 6" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": " on ViT-B and " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "n = 18" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": " on ViT-L. The scales of supervision are " + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\{56^2,28^2,14^2,7^2\\}" + }, + { + "bbox": [ + 46, + 338, + 287, + 590 + ], + "type": "text", + "content": ". Each decoder contains one transformer block with the embedding dimension of 256 and 8 attention heads. The pretraining and fine-tuning schedules mostly follow [21, 26], and more detailed settings can be found in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Results. We compare our LocalMIM with existing MIM models and examine both pre-training efficiency and top-1 fine-tuning accuracy on ImageNet-1K. The results are provided in Fig. 1 and Table 1. For fair comparison, we estimate the pre-training duration of each model on the same machine with one Tesla V100-32G GPU, CUDA 10.2 and Pytorch 1.8. We report the running time on single GPU, denoted as 'GPU Hours', see the Appendix for more details. On ViT-B, LocalMIM achieves the best results of MAE [21] and MaskFeat [47] with " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "3.1 \\times" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "5.6 \\times" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " accel" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 538, + 137 + ], + "blocks": [ + { + "bbox": [ + 47, + 200, + 287, + 234 + ], + "lines": [ + { + "bbox": [ + 47, + 200, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 287, + 234 + ], + "type": "text", + "content": "Table 2. Semantic segmentation on ADE20K using UperNet with ViT-B backbone. Our LocalMIM achieves better results than previous MIM models with less pre-training duration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 538, + 137 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 538, + 137 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 538, + 137 + ], + "type": "table", + "html": "
ModelPT EpochPT HoursAPbAPm
Supervised30084048.543.2
SimMIM192 [50]800144050.444.4
GreenMIM [26]80064050.044.1
LocalMIM-HOG40044050.744.9
", + "image_path": "b580ab01b5ea958f9b892c47a8a841babd98329019b6c55bce87185cabde3785.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "text", + "content": "Table 3. Object detection and instance segmentation on COCO. We fine-tune Mask R-CNN end-to-end with Swin-B backbone." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "content": "eration respectively. On Swin-B, LocalMIM achieves those of " + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{SimMIM}_{192}" + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "content": " [50] and GreenMIM [26] with " + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "inline_equation", + "content": "3.6\\times" + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "inline_equation", + "content": "6.4\\times" + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "content": " acceleration respectively. In terms of final top-1 finetuning accuracy, LocalMIM achieves " + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "content": " with ViT-B and " + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "inline_equation", + "content": "84.1\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "content": " with Swin-B. Compared with previous best results, LocalMIM achieves comparable performance with significantly less pre-training duration. Note that PeCo [15] uses more advanced feature descriptor, a pre-trained codebook, which introduces additional pre-training burden. Besides, on Swin-B, with 100 epochs pre-training and 100 epochs supervised fine-tuning LocalMIM achieves " + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "inline_equation", + "content": "83.8\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "content": " top-1 accuracy and takes about 350 GPU Hours, while 300 epochs supervised training from scratch only achieves " + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "inline_equation", + "content": "83.5\\%" + }, + { + "bbox": [ + 304, + 178, + 545, + 381 + ], + "type": "text", + "content": " top-1 accuracy and takes about 720 GPU Hours. This means even for the high-level classification task, self-supervision learning is more efficient and effective than supervised learning. Similar results can also be found on ViT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 389, + 415, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 415, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 415, + 400 + ], + "type": "text", + "content": "4.2. Downstream tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 407, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 443 + ], + "type": "text", + "content": "We transfer our pre-trained backbone to semantic segmentation on ADE20K [53] and object detection and segmentation on COCO [31]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": "Semantic segmentation on ADE20K. We conduct semantic segmentation on ADE20K using UperNet [48] and following the code in [2, 21]. See the Appendix for fine-tuning details. The results are shown in Table 2. With significantly less computation burden, LocalMIM outperforms the state-of-the-art result by 0.7. Besides, LocalMIM has the same top-1 fine-tuning accuracy with MaskFeat [47] on ImageNet-1K, but has better semantic segmentation performance, which means our local multi-scale reconstruction facilitates the learning of multi-scale semantic knowledge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Object detection and instance segmentation on COCO. We fine-tune Mask R-CNN [23] on COCO with Swin-B backbone. Following [26], we also use the code base and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " fine-tuning schedule from the supervised Swin, where the model is fine-tuned for 36 epochs. See the Appendix for fine-tuning details. We report box AP " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^b)" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " for object detection and mask AP " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^m)" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " for instance segmentation. The results are shown in Table 3. With no labeling burden, our LocalMIM outperforms supervised pre-training by 2.2 " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{AP}^m" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". Compared to SimMIM192 and GreenMIM, our LocalMIM achieves " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "+0.3" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "+0.7" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " gain respectively for detection, and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "+0.5" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "+0.8" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " gain for segmenta" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2127" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 192, + 114 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 192, + 114 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 192, + 114 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 192, + 114 + ], + "type": "table", + "html": "
targetViT-BSwin-B
baselines82.983.2
Pixels83.083.7
HOG83.383.8
", + "image_path": "f3891c1c327ff13db4d7b0f2a0f40b53b7c21a2f2b4ade20e2905190987a68b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 144, + 200, + 215 + ], + "blocks": [ + { + "bbox": [ + 50, + 124, + 201, + 143 + ], + "lines": [ + { + "bbox": [ + 50, + 124, + 201, + 143 + ], + "spans": [ + { + "bbox": [ + 50, + 124, + 201, + 143 + ], + "type": "text", + "content": "(a) Reconstruction target. HOG is more suitable as multi-scale supervisions than pixels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 144, + 200, + 215 + ], + "lines": [ + { + "bbox": [ + 53, + 144, + 200, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 144, + 200, + 215 + ], + "type": "table", + "html": "
locationsbackboneacc
GreenMIMSwin-B83.2
[24]83.3
[22, 24]83.4
[4, 22, 24]Swin-B83.6
[2, 4, 22, 24]83.8
fusion83.5
", + "image_path": "6f66e962161f3c0241c40bef955f5ea06778852e4f8a674d37b48be77f99e76d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 231, + 70, + 365, + 121 + ], + "blocks": [ + { + "bbox": [ + 50, + 218, + 201, + 238 + ], + "lines": [ + { + "bbox": [ + 50, + 218, + 201, + 238 + ], + "spans": [ + { + "bbox": [ + 50, + 218, + 201, + 238 + ], + "type": "text", + "content": "(d) Locations for Swin. Applying reconstructions at all stages mostly accelerates learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 231, + 70, + 365, + 121 + ], + "lines": [ + { + "bbox": [ + 231, + 70, + 365, + 121 + ], + "spans": [ + { + "bbox": [ + 231, + 70, + 365, + 121 + ], + "type": "table", + "html": "
ratioViT-BSwin-B
0.4083.283.7
0.6083.183.7
0.7583.383.8
0.9083.083.5
", + "image_path": "6fc144d7ada750056c732467376291f12cfd13a91236d854fba41da5fb2dbb19.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 226, + 145, + 367, + 208 + ], + "blocks": [ + { + "bbox": [ + 220, + 124, + 372, + 143 + ], + "lines": [ + { + "bbox": [ + 220, + 124, + 372, + 143 + ], + "spans": [ + { + "bbox": [ + 220, + 124, + 372, + 143 + ], + "type": "text", + "content": "(b) Mask ratio. Masking " + }, + { + "bbox": [ + 220, + 124, + 372, + 143 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 220, + 124, + 372, + 143 + ], + "type": "text", + "content": " patches works the best for both ViT and Swin." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 226, + 145, + 367, + 208 + ], + "lines": [ + { + "bbox": [ + 226, + 145, + 367, + 208 + ], + "spans": [ + { + "bbox": [ + 226, + 145, + 367, + 208 + ], + "type": "table", + "html": "
locationsbackboneacc
MAEViT-B82.9
[12]83.0
[3, 6, 9, 12]ViT-B83.1
[2, 4, 10, 12]83.3
[1, 2, 11, 12]82.9
", + "image_path": "1a6da0248cb28dc40a1e5625e82ef0209d07c661cd8edeeddb7f5821b4d0d6f7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 398, + 70, + 538, + 114 + ], + "blocks": [ + { + "bbox": [ + 220, + 218, + 372, + 238 + ], + "lines": [ + { + "bbox": [ + 220, + 218, + 372, + 238 + ], + "spans": [ + { + "bbox": [ + 220, + 218, + 372, + 238 + ], + "type": "text", + "content": "(e) Locations for ViT. The chosen layers for local losses obviously affect performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 398, + 70, + 538, + 114 + ], + "lines": [ + { + "bbox": [ + 398, + 70, + 538, + 114 + ], + "spans": [ + { + "bbox": [ + 398, + 70, + 538, + 114 + ], + "type": "table", + "html": "
decoderViT-BSwin-B
512D - 16H83.3 (1.0h)83.8 (1.6h)
256D - 8H83.3 (0.7h)83.7 (1.3h)
128D - 4H83.2 (0.6h)83.8 (1.1h)
", + "image_path": "1b9aa3a189b499410275ac870bab570f76c8d02528096849f2877bf4414622d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 395, + 145, + 542, + 208 + ], + "blocks": [ + { + "bbox": [ + 392, + 124, + 542, + 143 + ], + "lines": [ + { + "bbox": [ + 392, + 124, + 542, + 143 + ], + "spans": [ + { + "bbox": [ + 392, + 124, + 542, + 143 + ], + "type": "text", + "content": "(c) Decoder design. A tiny decoder performs as well as the larger one but is more efficient." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 395, + 145, + 542, + 208 + ], + "lines": [ + { + "bbox": [ + 395, + 145, + 542, + 208 + ], + "spans": [ + { + "bbox": [ + 395, + 145, + 542, + 208 + ], + "type": "table", + "html": "
scalesbackboneacc
MAEViT-B82.9
[142, 142, 142, 142]83.0
[282, 142, 72, 72]ViT-B83.2
[562, 282, 142, 72]83.3
[72, 142, 282, 562]83.1
", + "image_path": "e7aba84e904cd8f4edb6f4b901c41a72489ac3728ffbff97f111cd5ab2267255.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 392, + 218, + 542, + 237 + ], + "lines": [ + { + "bbox": [ + 392, + 218, + 542, + 237 + ], + "spans": [ + { + "bbox": [ + 392, + 218, + 542, + 237 + ], + "type": "text", + "content": "(f) Scales for ViT. Using the supervisions from fine-scale to coarse-scale works the best." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "text", + "content": "Table 4. Ablation studies with ViT-B and Swin-B on ImageNet-1K. We report the top-1 fine-tuning accuracy (%) and the default setting is: the reconstruction target is HOG feature, the mask ratio is " + }, + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "text", + "content": ", the decoder contains one Transformer block with dimension 256 and 8 heads for ViT and dimension 128 and 4 heads for Swin, the local reconstructions are applied at all stages of Swin and [2, 4, 10, 12]-th layer of ViT, the scales of supervision are " + }, + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "inline_equation", + "content": "[28^2, 14^2, 7^2, 7^2]" + }, + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "text", + "content": " for Swin and " + }, + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "inline_equation", + "content": "[56^2, 28^2, 14^2, 7^2]" + }, + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "text", + "content": " for ViT, and the pre-training length is 100 epochs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 309, + 242, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 242, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 242, + 321 + ], + "type": "text", + "content": "tion with significantly less pre-training duration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 332, + 145, + 343 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 145, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 145, + 343 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 400 + ], + "type": "text", + "content": "We ablate our LocalMIM using the default setting in Table 4. MAE [21] and GreenMIM [26] are our main baselines on ViT and Swin respectively, since we all adopt the asymmetric encoder-decoder strategy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "text", + "content": "Reconstruction target. We examine two types of signal as supervisions: HOG feature [47] and normalized pixel [21], and the results are shown in Table 4a. HOG feature is more suitable as multi-scale supervisions than normalized pixel on both ViT and Swin, since it contains more refined semantic information. Compared to the columnar ViT with single-scale feature maps, our multi-scale supervisions achieve more improvement on the pyramidal Swin. Since we need to rescale the predictions in the columnar architectures, it is harder to guide the learning of multi-scale information, e.g., we need more information-refined signals (HOG) and obtain relatively less improvement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 553, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 613 + ], + "type": "text", + "content": "Mask ratio. We explore different mask ratios, ranging from 0.4 to 0.9, and the results are provided in Table 4b. Similar to [21, 26], our LocalMIM is robust to the mask ratio from 0.4 to 0.75, and excessively large one (0.9) harms the performance since it over-complicates the reconstruction tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "Decoder design. As shown in [21, 26], a small single-block decoder can achieve the similar fine-tuning accuracy to the heavier one with multiple blocks. However, since we need to use the decoder for each reconstruction task to process the signals of specified scale, their decoder which has the embedding dimension of 512 and 16 attention heads is still heavy. To this end, we explore whether a smaller decoder can still handle the local reconstruction tasks for" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": "learning good representations. One Transformer block is the minimal requirement to propagate information from visible patches to mask tokens, so we still use one Transformer block but narrow the embedding dimension and reduce the number of attention heads. The results (top-1 fine-tuning accuracy and GPU Hours per epoch) are shown in Table 4c, where " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "dD - hH'" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": " represents the decoder with the embedding dimension of " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": " attention heads. Surprisingly, a tiny decoder can achieve the same even better fine-tuning performance, but is significantly more efficient." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "text", + "content": "Locations for local reconstructions. An important design of our LocalMIM is the locations in the encoder where we introduce local reconstructions. Many pyramidal architectures like Swin already divide the entire backbone into multiple stages, so we introduce multi-scale reconstructions at their output locations. From the last stage, we gradually include the lower stages and the results are shown in Table 4d. Note that the 2-th, 4-th, 22-th and 24-th layer correspond to the output location of the four stages respectively. We also provide the result of fusion method [18] which fuses the output features from each stage for reconstruction. As we can see, introducing reconstructions at local stages can accelerate representation learning, and applying to all stages works the best. Our LocalMIM also outperforms the fusion method which only uses the single-scale reconstruction, and this verifies the importance of multi-scale reconstructions. Further, for the columnar ViT, we follow the experience from Swin and also select four local layers. Concretely, we consider the uniform division [3, 6, 9, 12], Swin-style division [2, 4, 10, 12] and a extreme division [1, 2, 11, 12]. The results are shown in Table 4e. The performance on columnar ViT is sensitive to the chosen local layers, and the Swin-style division " + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "inline_equation", + "content": "[2, 4, 4 + n, 6 + n]" + }, + { + "bbox": [ + 304, + 438, + 546, + 715 + ], + "type": "text", + "content": " works the best." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2128" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 94, + 355 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 94, + 355 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 94, + 355 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 94, + 355 + ], + "type": "image", + "image_path": "384b5deb4a41f29823644f672ba46519d5b558711e78c2b84a9710e7e718db81.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "type": "text", + "content": "Figure 4. Visualization of the attention maps for different query points, marked with red boxes. LocalMIM can distinguish semantical regions using self-attention mechanism." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 104, + 70, + 287, + 356 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 287, + 356 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 287, + 356 + ], + "type": "image", + "image_path": "fe658e280a009b0a448b6b52f5b8ce8eed792aa4fd48e4011b46ea451eb91f9d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "text", + "content": "Scales of supervisions. Since the pyramidal architectures hard-code the multi-scale property of the feature maps, we directly use the supervisions of same scale as the feature maps for compatibility, i.e., we use " + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "inline_equation", + "content": "[28^2, 14^2, 7^2, 7^2]" + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "text", + "content": " for Swin. For columnar ViT, we consider single-scale supervisions " + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "inline_equation", + "content": "[14^2, 14^2, 14^2, 14^2]" + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "text", + "content": ", fine-to-coarse multi-scale supervisions " + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "inline_equation", + "content": "[28^2, 14^2, 7^2, 7^2]" + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "inline_equation", + "content": "[56^2, 28^2, 14^2, 7^2]" + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "text", + "content": ", and coarse-to-fine multi-scale supervisions " + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "inline_equation", + "content": "[7^2, 14^2, 28^2, 56^2]" + }, + { + "bbox": [ + 46, + 410, + 287, + 590 + ], + "type": "text", + "content": ". The results are shown in Table 4f. The multi-scale supervisions achieve better fine-tuning accuracy than the single-scale ones, which means the multiple local layers require the supervision of different scales. Further, the fine-to-coarse supervisions outperform the coarse-to-fine ones, which is consistent with the design priors for visual architectures [33] and biological visual processing [28]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Analysis of self-attention mechanism. To qualitatively illustrate the behavior of self-attention mechanism in LocalMIM, we visualize the attention maps for different query positions within an image, as shown in Fig. 4. We use each chosen patch as query and show the patches it attends to. We select the attention maps at 2-th, 4-th, 10-th and 12-th layers of a pre-trained ViT-B/16 backbone to show their changes during forward inference. For object-centric images, LocalMIM can distinguish the foreground object from the background. For more complex multi-object images," + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 543, + 133 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 543, + 133 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 543, + 133 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 543, + 133 + ], + "type": "table", + "html": "
modelbackboneGPU Hours/Ep.acc
LocalMIMViT-B0.783.3
w/ isolated grad0.783.0
LocalMIMSwin-B1.183.8
w/ isolated grad1.183.7
", + "image_path": "c95eb8b1e964061812a21bfe0271e8721fe0a7e55771c84b24ce798ccb5b8cac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 141, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 305, + 141, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 305, + 141, + 545, + 164 + ], + "type": "text", + "content": "Table 5. Training LocalMIM with isolated gradients achieves similar performance with global back-propagation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 178, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 262 + ], + "type": "text", + "content": "LocalMIM can effectively separate different objects without any task-specific supervision, which means the attention maps are query-adaptive. On the other hand, the patches at lower layers typically more focus on their neighboring regions and capture low-level information, while those at upper layers attend to a wide range of semantically related regions and capture high-level shape information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 266, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 546, + 435 + ], + "type": "text", + "content": "Gradient-isolated training. Inspired by locally supervised learning [34, 35], we remove global back-propagation and stop the gradients after each chosen local layer used for reconstruction. The entire backbone is thus divided into multiple gradient-isolated parts. The results are shown in Table 5. Surprisingly, the gradient-isolated training achieves similar performance to global back-propagation, which further verifies the effectiveness of our multi-scale reconstructions for guiding the local layers. It even requires no gradient information from the upper layers. This observation also shows the promise of our local multi-scale reconstruction for the decoupled training of neural networks, which allows for training very deep networks without memory concern and reduces explosive or vanishing gradients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 449, + 383, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 383, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 383, + 460 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 469, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 673 + ], + "type": "text", + "content": "We present a novel and efficient pretext task, local multi-scale reconstruction, where the lower layers and upper layers reconstruct the fine-scale and coarse-scale supervisions from the input respectively. We obtain the multi-scale supervisions by first dividing the input under different scales and then extracting supervisions with appropriate feature descriptors. Our model needs no extra pre-trained codebook and no costly forward inference of teacher networks during pre-training. The asymmetric encoder-decoder strategy with tiny encoders also have small computational burden, our model thus can be trained quickly. The novel pretext task further accelerates the representation learning, especially for the pyramidal architectures. All these designs allow our model to achieve comparable performance to existing models but with significantly less pre-training burden. The gradient-isolated training also verifies the effectiveness of our novel task design in guiding the local layers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment. We gratefully acknowledge the support of MindSpore [27], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2129" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Dong-Hyun Lee, Jorg Bornschein, Thomas Mesnard, and Zhouhan Lin. Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "type": "text", + "content": "[5] Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 339 + ], + "type": "text", + "content": "[6] Natalia Caporale, Yang Dan, et al. Spike timing-dependent plasticity: a hebbian learning rule. Annual review of neuroscience, 31(1):25-46, 2008. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 288, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 440 + ], + "type": "text", + "content": "[8] Jun Chen, Ming Hu, Boyang Li, and Mohamed Elhoseiny. Efficient self-supervised vision pretraining with local masked reconstruction. arXiv preprint arXiv:2206.00790, 2022. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 441, + 288, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 288, + 496 + ], + "type": "text", + "content": "[9] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 288, + 542 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 597 + ], + "type": "text", + "content": "[11] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 599, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 288, + 643 + ], + "type": "text", + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 644, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 288, + 666 + ], + "type": "text", + "content": "[13] Francis Crick. The recent excitement about neural networks. Nature, 337(6203):129-132, 1989. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 668, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 668, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 668, + 288, + 689 + ], + "type": "text", + "content": "[14] Yang Dan and Mu-ming Poo. Spike timing-dependent plasticity of neural circuits. Neuron, 44(1):23-30, 2004. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[15] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 184 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "type": "text", + "content": "[17] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6824-6835, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "type": "text", + "content": "[18] Peng Gao, Teli Ma, Hongsheng Li, Jifeng Dai, and Yu Qiao. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 277, + 545, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 321 + ], + "type": "text", + "content": "[19] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Yunhe Wang, and Chang Xu. Fastmim: Expediting masked image modeling pre-training for vision. arXiv preprint arXiv:2212.06593, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 323, + 545, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 323, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 545, + 367 + ], + "type": "text", + "content": "[20] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 369, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 369, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 369, + 545, + 425 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "text", + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 483, + 545, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 483, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 307, + 483, + 545, + 527 + ], + "type": "text", + "content": "[23] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 529, + 545, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 529, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 307, + 529, + 545, + 583 + ], + "type": "text", + "content": "[24] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921-1930, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 586, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 545, + 609 + ], + "type": "text", + "content": "[25] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification. In ACL, 2018. 1, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "type": "text", + "content": "[26] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling. arXiv preprint arXiv:2205.13515, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 656, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 656, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 656, + 545, + 677 + ], + "type": "text", + "content": "[27] Huawei. Mindspace. https://www.mindspore.cn/, 2020.8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[28] DH Hubel and TN Wiesel. Receptive fields of optic nerve fibres in the spider monkey. The Journal of physiology, 154(3):572, 1960. 3, 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2130" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "text", + "content": "[30] Xiang Li, Wenhai Wang, Lingfeng Yang, and Jian Yang. Uniform masking: Enabling mae pre-training for pyramid-based vision transformers with locality. arXiv preprint arXiv:2205.10063, 2022. 1, 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 215 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "text", + "content": "[32] Hao Liu, Xinghua Jiang, Xin Li, Antai Guo, Deqiang Jiang, and Bo Ren. The devil is in the frequency: Geminated gestalt autoencoder for self-supervised visual pre-training. arXiv preprint arXiv:2204.08227, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 261, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 315 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 317, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 358 + ], + "type": "text", + "content": "[34] Sindy Löwe, Peter O'Connor, and Bastiaan Veeling. Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems, 32, 2019. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 360, + 287, + 392 + ], + "type": "text", + "content": "[35] Arild Nøkland and Lars Hiller Eidnes. Training neural networks with local error signals. In International conference on machine learning, pages 4839-4850. PMLR, 2019. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 394, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 287, + 424 + ], + "type": "text", + "content": "[36] Namuk Park and Songkuk Kim. How do vision transformers work? In International Conference on Learning Representations, 2022. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 287, + 469 + ], + "type": "text", + "content": "[37] Myeongjang Pyeon, Jihwan Moon, Taeyoung Hahn, and Gunhee Kim. Sedona: Search for decoupled neural networks toward greedy block-wise learning. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "type": "text", + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 515, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 515, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 49, + 515, + 287, + 547 + ], + "type": "text", + "content": "[39] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "type": "text", + "content": "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 603, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 603, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 603, + 287, + 645 + ], + "type": "text", + "content": "[41] Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for combining multiple partitions. Journal of machine learning research, 3(Dec):583-617, 2002. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[42] Zhiqing Sun, Zhuohan Li, Haoqing Wang, Di He, Zi Lin, and Zhihong Deng. Fast structured decoding for sequence models. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "text", + "content": "[43] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In Zhi-Hua" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 696 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1075-1081. International Joint Conferences on Artificial Intelligence Organization, 2021. Main Track. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "text", + "content": "[44] Haoqing Wang, Xun Guo, Zhi-Hong Deng, and Yan Lu. Rethinking minimal sufficient representation in contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16041-16050, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 174, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 545, + 239 + ], + "type": "text", + "content": "[45] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 240, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 283 + ], + "type": "text", + "content": "[46] Yulin Wang, Zanlin Ni, Shiji Song, Le Yang, and Gao Huang. Revisiting locally supervised learning: an alternative to end-to-end training. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 285, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 349 + ], + "type": "text", + "content": "[47] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 352, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 352, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 352, + 545, + 396 + ], + "type": "text", + "content": "[48] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 397, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 545, + 439 + ], + "type": "text", + "content": "[49] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 441, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 545, + 496 + ], + "type": "text", + "content": "[50] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 498, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 552 + ], + "type": "text", + "content": "[51] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3008, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 554, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 596 + ], + "type": "text", + "content": "[52] Tianyi Zhang, Felix Wu, Arzoo Katiyar, Kilian Q Weinberger, and Yoav Artzi. Revisiting few-sample bert finetuning. In International Conference on Learning Representations, 2021. 1, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 598, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 598, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 308, + 598, + 545, + 651 + ], + "type": "text", + "content": "[53] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 654, + 545, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 654, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 308, + 654, + 545, + 696 + ], + "type": "text", + "content": "[54] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "2131" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_content_list.json b/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f2ed9bc2611c670165ca4aaf889a9e7875ccf9b0 --- /dev/null +++ b/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_content_list.json @@ -0,0 +1,2232 @@ +[ + { + "type": "text", + "text": "Masked Image Training for Generalizable Deep Image Denoising", + "text_level": 1, + "bbox": [ + 156, + 130, + 813, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyu Chen $^{1*}$ , Jinjin Gu $^{2,3*}$ , Yihao Liu $^{2,4,5}$ , Salma Abdel Magid $^{7}$ , Chao Dong $^{2,4}$ , Qiong Wang $^{6}$ , Hanspeter Pfister $^{7}$ , Lei Zhu $^{1,8\\dagger}$", + "bbox": [ + 200, + 179, + 767, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ The Hong Kong University of Science and Technology (Guangzhou) $^{2}$ Shanghai AI Lab $^{3}$ The University of Sydney \n $^{4}$ ShenZhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n $^{5}$ University of Chinese Academy of Sciences $^{6}$ Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n $^{7}$ Harvard University $^{8}$ The Hong Kong University of Science and Technology \nProject page: https://github.com/haoyuc/MaskedDenoising", + "bbox": [ + 111, + 222, + 861, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 354, + 312, + 371 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When capturing and storing images, devices inevitably introduce noise. Reducing this noise is a critical task called image denoising. Deep learning has become the de facto method for image denoising, especially with the emergence of Transformer-based models that have achieved notable state-of-the-art results on various image tasks. However, deep learning-based methods often suffer from a lack of generalization ability. For example, deep models trained on Gaussian noise may perform poorly when tested on other noise distributions. To address this issue, we present a novel approach to enhance the generalization performance of denoising networks, known as masked training. Our method involves masking random pixels of the input image and reconstructing the missing information during training. We also mask out the features in the self-attention layers to avoid the impact of training-testing inconsistency. Our approach exhibits better generalization ability than other deep learning models and is directly applicable to real-world scenarios. Additionally, our interpretability analysis demonstrates the superiority of our method.", + "bbox": [ + 76, + 387, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 705, + 209, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image denoising is a crucial research area that aims to recover clean images from noisy observations. Due to the rapid advancements in deep learning, many promising image denoising networks have been developed. These networks are typically trained using images synthesized from a pre-defined noise distribution and can achieve remarkable performance in removing the corresponding noise. However, a significant challenge in applying these deep models to real-world scenarios is their generalization ability. Since", + "bbox": [ + 75, + 728, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9373b1ea7f1ac7f1ac1b5de7a7335e2ba3aa06c42d3869ed80afea526de310f8.jpg", + "image_caption": [ + "Gaussian 15" + ], + "image_footnote": [], + "bbox": [ + 501, + 354, + 578, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ea9d39ecce52330d8179ea961126a6e04da3fdcb31e8cf9445a6e037e27e4de.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 578, + 354, + 656, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/227f37d923133f63d960230aee7b8e77dca56c453de3cad0cb49f90c722f890c.jpg", + "image_caption": [ + "Mixture noise", + "a. In-distribution", + "b. Out-of-distribution" + ], + "image_footnote": [], + "bbox": [ + 656, + 354, + 735, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbc519b4c42f9b813edae967851288d4913ba7353480551b7bc81f3ff01d66a6.jpg", + "image_caption": [ + "SwinIR", + "Figure 1. We illustrate the generalization problem of denoising networks. We train a SwinIR model on Gaussian noise with $\\sigma = 15$ . When tested on the same noise, SwinIR demonstrates outstanding performance. However, when applied to out-of-distribution noise, e.g., the mixture of various noise. SwinIR suffers from a huge performance drop. The model trained by the proposed masked training method maintains a reasonable denoising effect, despite aldo being trained on Gaussian noise." + ], + "image_footnote": [], + "bbox": [ + 735, + 354, + 816, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e9ee4c5c146a6f7baa388e2200e87b4be8f1467ee23d795d746b21f9602854a.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 816, + 354, + 890, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the real-world noise distribution can differ from that observed during training, these models often struggle to generalize to such scenarios.", + "bbox": [ + 496, + 594, + 890, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "More specifically, most existing denoising works train and evaluate models on images corrupted with Gaussian noise, limiting their performance to a single noise distribution. When these models are applied to remove noise drawn from other distributions, their performance drastically drops. Figure 1 shows an example. The research community has become increasingly aware of this generalization issue of deep models in recent years. As a countermeasure, some methods [80] assume that the noise level of a particular noise type is unknown, while others [5, 68] attempt to improve the performance in real-world scenarios by synthesizing or collecting training data closer to the target noise or directly performing unsupervised training on the target noise [11, 71]. However, none of these methods substantially improve the generalization performance of denoising networks, and they still struggle when the noise distribution is mismatched [1]. The generalization issue of", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Haoyu Chen and Jinjin Gu contribute equally to this work.", + "bbox": [ + 94, + 875, + 410, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Lei Zhu (leizhu@ust.hk) is the corresponding author.", + "bbox": [ + 96, + 887, + 382, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1692", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "deep denoising still poses challenges to making these methods broadly applicable.", + "bbox": [ + 78, + 90, + 465, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on improving the generalization ability of deep denoising models. We define generalization ability as the model's performance on noise different from what it observed during training. We argue that the generalization issue of deep denoising is due to the overfitting of training noise. The existing training strategy directly optimizes the similarity between the denoised image and the ground truth. The intention behind this is that the network should learn to reconstruct the texture and semantics of natural images correctly. However, what is often overlooked is that the network can also reduce the loss simply by overfitting the noise pattern, which is easier than learning the image content. This is at the heart of the generalization problem. Even many popular deep learning methods exacerbate this overfitting problem. When it comes to noise different from that observed during training, the network exhibits this same behavior, resulting in poor performance.", + "bbox": [ + 80, + 122, + 467, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of the preceding discussion, our study seeks to improve the generalization performance of deep denoising networks by directing them to learn image content reconstruction instead of overfitting to training noise. Drawing inspiration from recent masked modeling methods [4, 20, 34, 69], we employ a masked training strategy to explicitly learn representations for image content reconstruction, as opposed to training noise. Leveraging the properties of image processing Transformers [15, 45, 78], we introduce two masking mechanisms: the input mask and the attention mask. During training, the input mask removes input image pixels randomly, and the network reconstructs the removed pixels. The attention mask is implemented in each self-attention layer of the Transformer, enabling it to learn the completion of masked features dynamically and mitigate the distribution shift between training and testing in masked learning. Although we use Gaussian noise for training - similar to previous works - our method demonstrates significant performance improvements on various noise types, such as speckle noise, Poisson noise, salt and pepper noise, spatially correlated Gaussian noise, Monte Carlo-rendered image noise, ISP noise, and complex mixtures of multiple noise sources. Existing methods and models have yet to effectively and accurately remove all these diverse noise patterns.", + "bbox": [ + 80, + 378, + 467, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 80, + 768, + 223, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Denoising approaches very broadly lie in two categories: traditional model-based and data-driven deep-learning-based. Traditional methods are usually based on modeling image priors to recover image content contaminated by noise [7, 19, 23, 32, 53]. These methods usually do not impose too many constraints on the type of noise, and have been proven to be applicable to a variety of noise,", + "bbox": [ + 80, + 795, + 465, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with good generalization performance [1]. However, these methods are not satisfactory for the reconstruction of image content. In recent years, the paradigm of denoising has gradually shifted to data-driven methods based on deep learning methods [13]. Many techniques have been proposed to improve the capabilities of the denoising networks continuously, e.g., residual networks [39, 80, 81], dense networks [37, 86], recursive networks [9, 48, 63], multi-scale [21, 31, 76], encoder-decoder [16, 54, 73], attention operations [84, 85], self-similarity [35], and non-local operations [43, 44, 58]. Since 2020, the paradigm of vision network design has gradually shifted from CNNs to Transformers [22]. Vision Transformers treat input pixels as tokens and use self-attention operations to process interactions between these tokens. Inspired by the success of vision Transformers, many attempts have been made to employ Transformers for low-level vision tasks [10, 14, 15, 45, 62, 67, 70, 74, 77, 78] During the development of these models, the noise pattern used for training is often consistent with the testing one. The factor that determines its denoising performance is the fitting ability of the network, in other words, the ability of the network to overfit to the training noise. However, a better network does not mean a better generalization ability of the denoising model. As we will show in the experiment section, a more efficient network even indicates worse generalization performance.", + "bbox": [ + 503, + 90, + 890, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalization Problem in low-level vision often arises when the testing degradation does not match the training degradation, e.g., different downsampling kernel in superresolution [30, 40, 47]. We typically develop deep denoising models based on Gaussian noise in the laboratory setting. However, noise in the real-world is mostly non-Gaussian. Models trained on Gaussian noise fail in these non-Gaussian scenarios. There are two main categories of solutions to this problem. The first is to make training datasets with noise modeling as close to reality as possible during development, e.g., synthesizing real noise according to physical system modeling [5, 68], learning to generate real noise [11, 24, 71], collecting real noise - clean image pairs for training [1, 33, 41, 57]. Although the models obtained by these methods can improve the effect on the target noise, they still cannot generalize to out-of-distribution noise. Another category of solutions is to develop \"blind\" denoising models, which are supposed to deal with unknown noise [41, 72, 80]. These methods usually simply assume that the noise level is unknown, or train on a large amount of noise types [79], which also fails to generalize to other noise not present in the training set. Few workd have been proposed to study the reasons for the lack of generalization ability in low-level vision [40]. Liu et al. [49] argue that networks tend to overfit to degradations and show degradation \"semantics\" inside the network. The presence of these representations often means a decrease in general-", + "bbox": [ + 503, + 493, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1693", + "bbox": [ + 485, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/119cbf8907f4a27a639fefac363dafdb34aac61e4e18f48fafccc7b877012ad4.jpg", + "image_caption": [ + "Training image" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 228, + 202 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c6249b64558173b20eb0d0fb8b5897f49b7631dcadcc69f81dbb2c7c7d745870.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 232, + 88, + 346, + 141 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0443ac2881d07b01bd05fa1f007c43f5739b7aa8f0e63771833f60fcfdfc9df7.jpg", + "image_caption": [ + "Denoised by SwinIR" + ], + "image_footnote": [], + "bbox": [ + 351, + 88, + 468, + 140 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/223a24012884b990720efd14fe331f7e34fc443579bd20fc4685ae36faebc9ae.jpg", + "image_caption": [ + "Denoised by our method", + "Figure 2. SwinIR, when trained solely on immunohistochemistry images with Gaussian noise, can still denoise natural images. This observation supports the assertion that most existing methods perform denoising primarily through overfitting the training noise. In contrast, our approach emphasizes reconstructing natural image textures and edges observed in the training set on natural images, rather than relying on noise overfitting for denoising. This distinction underlines the fundamental difference between our method and previous approaches. \"Our reconstruction result\" refers to using our model but taking masked images as input." + ], + "image_footnote": [], + "bbox": [ + 232, + 151, + 346, + 203 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c6fb6ce865e0cd411b08d131e9f683530059f9985c7881988374e6ab7dd0540a.jpg", + "image_caption": [ + "Our reconstruction result" + ], + "image_footnote": [], + "bbox": [ + 351, + 151, + 468, + 203 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ization ability. The utilization of this knowledge can guide us to analyze and evaluate the generalization performance [50]. Apart from that, few works have been proposed to improve the generalization ability of denoising models.", + "bbox": [ + 75, + 375, + 468, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked modeling for language [6, 20, 59, 60] is successful for learning pre-trained representations that generalize well to various downstream tasks. These methods mask out a portion of the input sequence and train models to predict the missing content. A similar approach can also be applied to the vision model pre-training. Masked image models learn representations from corrupted images. The earliest attempts in this regard can be traced back at least to the denoising auto-encoder [66]. Since then, many works have used predicting missing parts of images to learn efficient image representations [4, 12, 34, 56, 69]. However, there have been few successful attempts to apply masked image modeling to low-level vision, even though the masked pretraining method is in the form of low-level vision tasks.", + "bbox": [ + 75, + 443, + 468, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 670, + 166, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to create denoising models capable of generalizing to noise not encountered in the training set. In this section, we first discuss our motivation before delving into the specifics of our masked training method.", + "bbox": [ + 75, + 696, + 468, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. When training a deep network on a large number of images, the expectation is for the network to learn to discern the rich semantics of natural images from noise-contaminated test cases. However, several studies have noted that the semantics and knowledge acquired by low-level vision networks differ significantly from our expectations [29, 49, 50, 52]. We argue that the poor generalization ability of denoising models results from our training method, which leads the model to focus on overfit", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ffa6648d79692cf7f4f6022cdec0f9b44a4dcdddd8f25b5afec223ac9b4c4e62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 578, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/da218d2b5d65f12e702bf9fa9087d80f4c0f54fbae698624ecd46aba03169d36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 88, + 733, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/994bc65861c5a73bf79b972e3af048355fa19873f59bd3845f74fbe8e9c30aab.jpg", + "image_caption": [ + "original", + "Figure 3. The illustration of the proposed mask-and-complete training strategy. Even if a large number of pixels are masked, the model can still reconstruct the input to some extent." + ], + "image_footnote": [], + "bbox": [ + 501, + 150, + 580, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cc82582f51af3712b10d74d3463ced7992f47c55b3baa5752ca81156a583ba22.jpg", + "image_caption": [ + "mask $70\\%$" + ], + "image_footnote": [], + "bbox": [ + 583, + 150, + 733, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/39c20430259683b0bfb1f88ed520c47cda6fea1fe3c145168f516ef2fb6e238d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 88, + 890, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4e754983b462d15d191d5106df24018fbb2bc7cd5c0c30c823d405eb77be583f.jpg", + "image_caption": [ + "mask $90\\%$" + ], + "image_footnote": [], + "bbox": [ + 736, + 150, + 890, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ting the training noise rather than learning image reconstruction. We conduct a simple experiment for verification. We trained a SwinIR denoising network [45] using images that greatly differ from natural images (immunohistochemistry images [65]). We synthesized training data pairs using Gaussian noise, and then assessed the model's performance on natural images with Gaussian noise. According to our hypothesis, if the model learns the content and reconstruction of image semantics from the training set, it should not perform well on natural images, as it has not been exposed to any. If the model is simply overfitting the noise, the model can remove the noise even if the images are different, as the model mainly relies on detecting the noise for denoising.", + "bbox": [ + 496, + 290, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The results are presented in Figure 2. As observed, the SwinIR trained on immunohistochemistry images can still denoise and reproduce the natural image. This supports our conjecture regarding generalization ability, indicating that most existing methods perform denoising by overfitting the training noise. Consequently, when the noise deviates from the training conditions, the denoising performance of these models declines significantly.", + "bbox": [ + 496, + 503, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This observation also inspires our approach to developing deep denoising models with improved generalization ability. We aim for the model to learn the reconstruction of image textures and structures, rather than focusing only on noise. In this paper, we propose a new masked training strategy for denoising networks. During training, we mask out a portion of the input pixels and then train the deep network to complete them, as shown in Figure 3. Our approach emphasizes reconstructing natural image textures and edges observed in the image, rather than overfitting noise. In Figure 2 we also show the results of our method. It is evident that our approach seeks to reconstruct the immunohistochemistry image texture from the training set on the testing natural image, instead of relying on noise overfitting for denoising. This demonstrates the potential of this idea in improving generalization performance. By training our method on natural images, it will concentrate on reconstructing the content of natural images, aligning with our", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1694", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b808158284f112b13b4111bafd433e2c8159155b2d4497a81d32a9878f08d546.jpg", + "image_caption": [ + "Figure 4. The transformer architecture of our proposed masked image training. We make a minimal change to the original SwinIR architecture – the input mask operation and the attention masks. Other micro-designs are not essentially different from other Transformers." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 188 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/34182b779ea0773cb66a339789d0f4b7a2cd36037c31611530058b0e0aa07bfe.jpg", + "image_caption": [ + "Figure 5. Quantitative effect of the attention mask. The histogram differences are also shown above." + ], + "image_footnote": [], + "bbox": [ + 76, + 239, + 472, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "core concept of employing deep learning for low-level vision tasks.", + "bbox": [ + 75, + 422, + 468, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Transformer Architecture. Our approach exploits the excellent properties of visual Transformers, so we first describe the basic Transformer backbone used in this study. The shifted window mechanism is proven to be flexible and effective for image processing tasks [15, 45, 78]. We only make minimal changes when applying it to the proposed masked training method without the loss of generality. This model is illustrated in Figure 4. Transformers divide the input signal into tokens and process spatial information using self-attention layers. In our method, a convolution layer with kernel size 1 is used as the feature embedding module to project the 3-channel pixel values into $C$ -dimensional feature tokens. The $1 \\times 1$ convolution layer ensures that pixels do not affect each other during feature embedding, which facilitates subsequent masking operations. These feature tokens are gathered with shape $H \\times W \\times C$ , where $H$ , $W$ and $C$ are the height, width and feature dimension. The shifted window mechanism first reshapes the feature maps of each frame to $\\frac{HW}{M^2} \\times M^2 \\times C$ features by partitioning the input into non-overlapping $M \\times M$ local windows, where $\\frac{HW}{M^2}$ is the total number of windows. We calculate self-attention on the feature tokens within the same window. Therefore, $M^2$ tokens are involved in each standard self-attention operation, and we produce the local window feature $X \\in \\mathbb{R}^{M^2 \\times C}$ . In each self-attention layer, the query $Q$ , key $K$ and value $V$ are calculated as $Q = XW^Q$ , $K = XW^K$ , $V = XW^V$ , where $W^Q, W^K, W^V \\in \\mathbb{R}^{C \\times D}$ are weight matrices, and $D$ is the dimension of projected vectors. Then, we use $Q$ to query $K$ to generate the attenuated feature vectors.", + "bbox": [ + 75, + 460, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4fd7a18c82c187b7c35d34645a9401784867b438bcbd0b53bdbd44584a74ced8.jpg", + "image_caption": [ + "Figure 6. The effectiveness of the input mask and attention mask. Note that the brightness of the image is wrong w/o attention mask." + ], + "image_footnote": [], + "bbox": [ + 500, + 241, + 893, + 313 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3ee6f75f716ecdaef68e3a92c211a139eb964c6e47682e47f954d1a10fe983fa.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. The importance of using Table 2. Ablation on the at-different mask operations. tention mask ratio." + ], + "table_body": "
Input MaskAttention MaskPSNRSSIMMix. noise on CBSD68 [55]
Ratio (%)PSNRSSIM
✓29.170.82276529.570.8657
✓26.960.82027529.760.8678
✓✓29.740.86728528.840.8548
", + "bbox": [ + 506, + 349, + 879, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion map $A = \\operatorname{softmax}\\left(\\frac{QK^T}{\\sqrt{D}} + B\\right) \\in \\mathbb{R}^{M^2 \\times M^2}$ , where $B$ is the learnable relative positional encoding. This attention map $A$ is then used for the weighted sum of $M^2$ vectors in $V$ . The multi-head settings are aligned with SwinIR [45] and ViT [22].", + "bbox": [ + 498, + 459, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked Training. Our masked training mainly consists of two aspects, the input mask and the attention mask. Although both are mask operations, the purpose of these two masks is different. We describe them separately.", + "bbox": [ + 496, + 551, + 890, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Input Mask randomly masks out the feature tokens embedded by the first convolution layer, and encourages the network to complete the masked information during training. The input mask explicitly constructs a very challenging inpainting problem, as shown in Figure 3. It can be seen that even if up to $90\\%$ of the pixel information is destroyed, the network can still reconstruct the target image to a certain extent. The method is very simple. Given the feature token tensor $\\mathbf{f} \\in \\mathbb{R}^{H \\times W \\times C}$ , we randomly replace the token with a [mask token] $\\in \\mathbb{R}^C$ with a probability $p_{\\mathrm{IM}}$ , where $p_{\\mathrm{IM}}$ is called the input mask ratio. The network is trained under the supervision of the $l_1$ -norm of the reconstructed image and the ground truth. The [mask token] can be learnable and initialized with a 0 vector. But we actually found that the 0 vector itself is already a suitable choice. The existence of the input mask forces the network to learn to recognize and reconstruct the content of the image from very limited information.", + "bbox": [ + 496, + 612, + 892, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Attention Mask. We cannot build usable image pro", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1695", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/524c77de7b8197cdf6e0befff3128ee7cb3758936020e495ae311e07cdc240de.jpg", + "image_caption": [ + "Figure 7. The trade-off of choosing different mask ratios. The performance drop on training noise is not significant until $75\\%$ masking ratio. Our performance gain on the noise outside the training set is greater than the performance loss on the training set." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cessing networks relying solely on the input mask operation. Because during testing, we will input uncorrupted images to retain enough information. At this time, due to the inconsistency between training and testing, the network will tend to increase the brightness of the output image. Such as the example in Figure 5. Since Transformer uses the self-attention operation to process spatial information, we can narrow the gap between training and testing by performing the same mask operation during the self-attention process. The specific mask operation is similar to the input mask, but a different attention mask ratio $p_{\\mathrm{AM}}$ and [mask token] are used. When some tokens in the self-attention are masked, the attention operation will adjust to the fact that the information of these tokens is no longer reliable. Self-attention will focus on unmasked tokens in each layer and complete the masked information. This operation is difficult to implement on convolutional networks. Figure 5 shows the effect of the attention mask. As can be seen, the attention mask successfully makes the masked trained network work on the unmasked input image.", + "bbox": [ + 75, + 234, + 470, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 551, + 209, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Settings. For synthesizing training data, we sample the clean images from DIV2K [64], Flickr2K [46], BSD500 [3], and WED [51] during training. In our work, all the networks are trained using Gaussian noise with standard deviation $\\sigma = 15$ . Each input image is randomly cropped to a spatial resolution of $64 \\times 64$ , and the number of the total training iteration is $200\\mathrm{K}$ . We adopt the Adam optimizer [38] with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ to minimize the $L_{1}$ pixel loss. The initial learning rate is set as $1 \\times 10^{-4}$ and reduced by half at the milestone of $100\\mathrm{K}$ iterations and $150\\mathrm{K}$ iterations. The batch size is set to 64.", + "bbox": [ + 75, + 577, + 468, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing Noise. Since the training process utilizes the Gaussian noise, we evaluate the generalization performance of the models on six other synthetic noise: (1) Speckle noise, a type of noise that occurs during the acquisition of medical images or tomography images. (2) Poisson noise, a type of signal-dependent noise that occurs during the acquisition of digital images.(3) Spatially-correlated noise. This is to synthesize the complex artifact after denoising using a flawed algorithm. It is produced by filtering Gaussian noise with a $3 \\times 3$ average kernel. Different standard deviations of", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the Gaussian noise indicate different noise levels. (4) Salt & pepper noise. (5) Image signal processing (ISP) noise. [5] proposes a method to synthesize realistic ISP noise during digital imaging. (6) Mixture noise obtained by mixing the above different types of noise with different levels [79]. The clean images are sampled from the benchmark datasets, including CBSD68 [55], Kodak24 [26], McMaster [82], and Urban100 [36]. We also include two real noise types in this work: the Smartphone Image Denoising Dataset (SIDD) [1] and Monte Carlo (MC) rendered image noise. For evaluation, we follow [27,28] and use the metrics PSNR, SSIM [51], and LPIPS [83] to evaluate the results. Since PSNR and SSIM are questioned in assessing the perceptual quality of images [27,28], we also use the LPIPS as an additional metric.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Results", + "text_level": 1, + "bbox": [ + 500, + 327, + 591, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ablation Study. Table 1 and Figure 6 show the effectiveness of using different mask operations. As we can see, without the input mask, the model will lose its generalization ability, and cannot effectively remove the noise outside the training set. Without the attention mask, due to the training-testing inconsistency, the quantitative performance degrades significantly, and the output image will have the wrong brightness. In addition, even without the attention mask, the generalization ability of the model is not significantly affected, and most of the noise is still effectively removed. The input mask is the crucial factor in improving the model the generalization ability.", + "bbox": [ + 496, + 351, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3a shows the impact of the different input mask ratios. We test fixed ratios and random ratios from a uniform distribution. From our experiments, fixed ratios are less stable for training than randomly chosen from a range, and the performance is also worse. The best quantitative performance is achieved with random sampling ratios between $75\\% \\sim 85\\%$ . This is a trade-off between denoising generalization ability and the preservation of image details. As shown in Figure 7, smaller ratios are not enough for the network to learn the distribution of images because more noise patterns are preserved. The larger ratio improves the model generalization, as the model focuses more on reconstruction. But at the same time, some image details may be lost. For attention mask ratio, we show the effects in Table 2. The optimal ratios are around $75\\%$ .", + "bbox": [ + 496, + 532, + 890, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Generalization Performance. We evaluate our deep denoising method on synthetic noise, where our training noise follows a Gaussian distribution with a single noise level, but we test on multiple types of non-Gaussian noise to assess the model's generalization performance. In Figure 11, we compare our method with other state-of-the-art models based on their PSNR and SSIM scores. The results show that our model outperforms all the other models in terms of generalization performance. Particularly, as", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1696", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/911690adc5bd52c632fe96531223a16b82cc3fbb98f44da00ff1b268b5e8c7e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 88, + 289, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba7f643d21300770df5ca47047bba32ac76ad798c4f7f196f8b80ef503b6e9f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 88, + 439, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a6005a27047b8f23abbcc6e9bdfec653843f05fd4c3f0804118ecfa282abd56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 88, + 588, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c4b9a663bf3ab8496e5ca7a1fcf16436e4f5e116f001da3610a6efaf5acfb88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 88, + 736, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0570e030ec1c76f1d33664aa960ecfd077b2f3af9e68b7ebd12ea0358c81cf8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 88, + 887, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/839898d8e0f03601ba079d6d2b986935f2f3bcab26072b3614d7eca537b29e2e.jpg", + "image_caption": [ + "CBSD68: img_053", + "CBSD68: img_0046" + ], + "image_footnote": [], + "bbox": [ + 83, + 227, + 289, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd4b664a95ee70c130168eb35db5df4ad5a930aa2e08b231753c000edbe31593.jpg", + "image_caption": [ + "Spatially correlated Gaussian, $\\sigma = 50$" + ], + "image_footnote": [], + "bbox": [ + 290, + 159, + 439, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0273b063d46fe4758fe5ce12293c32ca44f553f2465974718760d71fb35fcb1d.jpg", + "image_caption": [ + "DnCNN [80]" + ], + "image_footnote": [], + "bbox": [ + 441, + 159, + 586, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb8fa6a8e05f737d30a295f1adf893c55a7ed0b4b9eae0d4673b876aba467137.jpg", + "image_caption": [ + "RIDNet [2]" + ], + "image_footnote": [], + "bbox": [ + 589, + 159, + 736, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f57e5576cb301b3a729d010190f1a677ac968b903c6bcc9638c7329783fdce54.jpg", + "image_caption": [ + "RNAN [85]" + ], + "image_footnote": [], + "bbox": [ + 738, + 159, + 885, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 83, + 366, + 290, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CBSD68: img_0067", + "bbox": [ + 140, + 494, + 233, + 505 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77a48262e2831525ae5adca328406795d828389886d178614aff82f16c768002.jpg", + "image_caption": [ + "Restormer [75]" + ], + "image_footnote": [], + "bbox": [ + 290, + 227, + 439, + 284 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de034ae40e8ffbed275120b127d610984c406f64ab3cd1534417f5877377f784.jpg", + "image_caption": [ + "SwinIR [45]" + ], + "image_footnote": [], + "bbox": [ + 441, + 227, + 586, + 284 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6b2b2a3068d5d0f5a2d6f3eb35c4fa9a753835ef077d3589cf3570a639de835.jpg", + "image_caption": [ + "baseline" + ], + "image_footnote": [], + "bbox": [ + 589, + 227, + 736, + 284 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f3256261b64aea66cab5d72c3291484542cfa7f84a2d3942006d75a98f990cc.jpg", + "image_caption": [ + "Masked Training" + ], + "image_footnote": [], + "bbox": [ + 738, + 227, + 885, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CBSD68: img_0067", + "bbox": [ + 83, + 494, + 289, + 505 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d94ceccba28fffb236c9a57076403168602f47af7e593981afc4dcdcbeca6d0d.jpg", + "image_caption": [ + "Speckle noise, $\\sigma^2 = 0.013$" + ], + "image_footnote": [], + "bbox": [ + 290, + 297, + 439, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30a4277d6c51acc68461321ab2df739e9052363bc89966da788e9db4d86aae3c.jpg", + "image_caption": [ + "DnCNN [80]" + ], + "image_footnote": [], + "bbox": [ + 441, + 297, + 586, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5078b2a66d132996a58c2364a0b12f5a6509e0a5cc5e18ef80f9ba8076902491.jpg", + "image_caption": [ + "RIDNet [2]" + ], + "image_footnote": [], + "bbox": [ + 589, + 297, + 736, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9a1fc154dd1a6f4079ad25bc013cab9c6cb444785c626a959c745b3aa0c1d23.jpg", + "image_caption": [ + "RNAN [85]" + ], + "image_footnote": [], + "bbox": [ + 738, + 297, + 885, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 83, + 366, + 290, + 493 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/275a77bf7eb8d5948aa0882874985f1d45bfeca27f00d0f347fe3b15e22c941b.jpg", + "image_caption": [ + "Restormer [75]" + ], + "image_footnote": [], + "bbox": [ + 290, + 366, + 439, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f231ed7b81df15f5433d229b3ef1cf9357accf20c1c37a72b62ed7cd12f2e1f5.jpg", + "image_caption": [ + "SwinIR [45]" + ], + "image_footnote": [], + "bbox": [ + 441, + 366, + 586, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3cc84996467047107870bed410e4105e4fa7009dc9751b8e14f08e8f5b8d41e.jpg", + "image_caption": [ + "baseline" + ], + "image_footnote": [], + "bbox": [ + 589, + 366, + 736, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aecd15b639520ea0ba5ceca1a1c30f2dae901e744bc71bec120169d84cc8cd2b.jpg", + "image_caption": [ + "Masked Training" + ], + "image_footnote": [], + "bbox": [ + 738, + 366, + 885, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CBSD68: img_0067", + "bbox": [ + 83, + 494, + 289, + 505 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd73fd5c98e236ab194eaacbe68e86a819523df5001785f7884a02a8844ae7eb.jpg", + "image_caption": [ + "Salt-and-pepper noise, $d = 0.02$", + "Restormer [75]" + ], + "image_footnote": [], + "bbox": [ + 290, + 436, + 439, + 493 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/703ed3f6c493a10ad68348621588476a65c32bb9688cbcbccfb4d8d6bc09ea0e.jpg", + "image_caption": [ + "DnCNN [80]", + "SwinIR [45]" + ], + "image_footnote": [], + "bbox": [ + 441, + 436, + 586, + 493 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/828c5e8f8e3b3a1110e1f3a72b9202b3d3962576fce339911b29ffb6325f6b1c.jpg", + "image_caption": [ + "RIDNet [2]", + "baseline" + ], + "image_footnote": [], + "bbox": [ + 589, + 436, + 736, + 493 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec12561b78ec63a2fc1c9fa1d6532e9bddc1238f5e842786262a2847d6896d2e.jpg", + "image_caption": [ + "RNAN [85]", + "Masked Training" + ], + "image_footnote": [], + "bbox": [ + 738, + 436, + 885, + 493 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7bae78ae4f6ea8a6835f49474e7c8ab0ebf0eddf330036be55d6b147e9a1bb29.jpg", + "image_caption": [ + "Figure 8. Visual comparison on out-of-distribution noise. When all other methods fail completely, our method is still able to denoise effectively. Please refer to the supplementary material to see more visual results.", + "Rendered image (256 spp)" + ], + "image_footnote": [], + "bbox": [ + 78, + 546, + 210, + 680 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eeaaf3455adea2a9d6c20e797d80f0ae70bed4051b990536a8bd8c0378540b72.jpg", + "image_caption": [ + "Noisy" + ], + "image_footnote": [], + "bbox": [ + 212, + 546, + 336, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2ceb774440941b0b2a2ff63c653ab416ea817c274d24b5c70d3f9e716646a63c.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 339, + 546, + 462, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d983ce7d996eea169e7dbf638f5d573581a10ec55a3dcb33d22a60563922a98.jpg", + "image_caption": [ + "DnCNN" + ], + "image_footnote": [], + "bbox": [ + 465, + 546, + 591, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ca69d2eafa1f35c082f44c3b8a8e156a79bb17ef4014054b21c2f93f8d1aedd.jpg", + "image_caption": [ + "Noisy" + ], + "image_footnote": [], + "bbox": [ + 607, + 546, + 700, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfaf45f3c4f702d8cb73a92f4b39f63cf3c948f60e6cb75839a66ab787e2113a.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 702, + 546, + 792, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5a6acba261e5342dbdb79af8919c0e24fe52720256980c08a728a7825cef5eb.jpg", + "image_caption": [ + "DnCNN" + ], + "image_footnote": [], + "bbox": [ + 795, + 546, + 887, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76899e35721f963c496c81386757bfb0d70a51e04d57560d6d4071adabea1950.jpg", + "image_caption": [ + "Figure 9. Visual results of denoising a Monte Carlo rendered image.", + "RNAN" + ], + "image_footnote": [], + "bbox": [ + 212, + 611, + 336, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5676def3d9a50554072b314f2ad9b7a555b5e2767dd601dc52800c951b8ec94.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 339, + 611, + 462, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d84c270097afb20f203ae674985c08cb020fbfad2b4dcf114b24718a5feb4afb.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 465, + 611, + 591, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3c9210cc68eb76f09476e6f4f04d844ed8db80c12dda4bcc93fef942831d1d6.jpg", + "image_caption": [ + "RNAN", + "Figure 10. Results of ISP noise removal." + ], + "image_footnote": [], + "bbox": [ + 607, + 619, + 700, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4714e91bddfc967421fba23d9ce47892e57d1e8a5d85b77b29b6703ee43baa18.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 702, + 619, + 792, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08d89ec8452a46ba2e6caf57e5eaa8e727aa57adbabab4fa7405427fb270f45b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 795, + 619, + 885, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the noise level increases, our model exhibits a slower performance degradation and thus demonstrates better generalization. In contrast, other models suffer from significant performance drops when dealing with more severe noise. We also provide visual comparisons in Figure 8, where our model achieves remarkable denoising results even though it is trained only on Gaussian noise with a fixed standard deviation. In contrast, existing models tend to overfit the training noise and fail when facing unseen noise. More quantitative and qualitative results can be found in the supplementary material.", + "bbox": [ + 75, + 712, + 468, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on ISP noise. The removal of the ISP noise is of great application value. Brooks et al. [5] present a systematic approach for generating realistic raw data with ISP noise that can facilitate our research. We use the default parameter settings of the method proposed in [5] to synthesize ISP noise on the Kodak24 [26] dataset for testing. The results are shown in Figure 10 and Table 3c. Our method achieves superior results compared to all other methods. Notably, our method achieves a significant lead in LPIPS, indicating that our results exhibit better perceptual quality. Although DnCNN and our method obtain the same PSNR, our method still outperforms DnCNN in terms of SSIM and", + "bbox": [ + 496, + 712, + 890, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1697", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5f55ebf41b33728309cdd5926849b1336c3e6d58893c14e4625eb81fc7ed7ba5.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Abl. of input mask ratios." + ], + "table_body": "
Mix. noise on CBSD68 [55]
Ratio (%)PSNRSSIM
7529.170.8132
8529.440.8545
9519.600.7273
70–8029.860.8593
75–8530.040.8756
75–9029.870.8728
75–9529.260.8607
80–9029.740.8672
", + "bbox": [ + 89, + 88, + 251, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f81cbdfb9e8f5e5c1484586df006f88e0ec725cac216748187ae369e2d12661.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Quantitative comparison on Monte Carlo rendered image denoising." + ], + "table_body": "
Method128 samples per pixel64 samples per pixel
PSNRSSIMLPIPSPSNRSSIMLPIPS
DnCNN [80]29.940.78830.267126.280.67790.4216
RIDNet [2]29.960.79210.254826.270.67880.4122
RNAN [85]29.860.78250.270226.260.67430.4290
SwinIR [45]29.320.76270.294326.140.66510.4485
Restormer [75]24.980.65980.457524.590.58800.5375
Dropout [40]28.850.77530.294126.100.66960.4443
baseline29.680.77380.285125.910.65350.4564
Ours30.620.85000.225428.250.76940.3348
", + "bbox": [ + 256, + 88, + 638, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/87702f7385f9396681c6d81ec3ed9f5f756c32df982ac580742156b2a97fb973.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Comparison on synthetic ISP noise." + ], + "table_body": "
Synthetic ISP noise [5]
MethodPSNRSSIMLPIPS
DnCNN [80]29.440.78570.3083
RIDNet [2]28.750.74460.3696
RNAN [85]28.470.72430.3601
SwinIR [45]28.390.70790.3346
Restormer [75]19.310.49820.6556
Dropout [40]28.390.78160.2621
baseline28.890.75950.2917
Ours29.440.79200.2368
", + "bbox": [ + 645, + 88, + 883, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. We train all the models on Gaussian noise, $\\sigma = {15}$ . All the testing noise is out of the training set, therefore the results can show the models' generalization performance on different unseen noise.", + "bbox": [ + 75, + 239, + 890, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56f384c2f38d7cd075f0db2482381134bab882d2519be3d38fc60066c6b6b61e.jpg", + "image_caption": [ + "Figure 11. Performance comparisons on four noise types with different levels on the Kodak24 dataset [26]. All models are trained only on Gaussian noise. Our masked training approach demonstrates good generalization performance across different noise types. We involve multiple types and levels of noise in testing, the results cannot be shown here. More results are shown in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 76, + 276, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LPIPS. Furthermore, as evident from Figure 10, DnCNN's results still contain visible noise, while our method effectively removes the noise.", + "bbox": [ + 75, + 507, + 468, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Monte Carlo rendering noise. Monte Carlo denoising is a vital component of the rendering process since the widespread use in the industry of Monte Carlo rendering algorithms [8, 17, 42]. We use the test dataset proposed by [25] for Monte Carlo rendered image denoising. The test images were rendered in 128 samples-per-pixel (spp) and 64 spp. The lower the spp, the more severe the noise of the image. In order to adapt the test set to our model, we first convert the data set to sRGB color space by tone mapping. Figure 9 and Table 3b show the denoising results. Our method outperforms all methods on both 128spp and 64spp settings. In Figure 9, the existing methods fail completely because of poor generalization. Our model is still able to remove this noise, demonstrating the wide applicability of our method.", + "bbox": [ + 75, + 561, + 468, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Generalization Analysis", + "text_level": 1, + "bbox": [ + 76, + 800, + 295, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training curve. Figure 13 shows the training curves of the model with and without the proposed masked training. The models are trained using only Gaussian noise. The baseline method has a significant overfitting problem. The performance of our method gradually improves with train-", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing without overfitting.", + "bbox": [ + 500, + 507, + 653, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CKA analysis. To investigate how masked training differs from normal training strategy, we utilize the centered kernel alignment (CKA) [18, 61] to analyze the differences between network representations obtained from those two training methods. Due to the limited space, we describe the detail of CKA in supplementary. In Figure 12, we present our key findings. Specifically, Figure 12 (a) shows the cross-model comparison between the baseline model and our masked training model. We observe a significant difference between the two models in terms of their feature correlations in the deeper layers. Specifically, the features of the deeper layers of the baseline model exhibit low correlations with all layers of our model. This finding suggests that these two training methods exhibit inconsistent learning patterns for features, especially for the deeper layers.", + "bbox": [ + 496, + 536, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To explore how the models perform on different noise types, Figure 12 (b) shows the cross-noise comparison between in-distribution noise and out-of-distribution noise, such as Gaussian and Poisson noise. For the baseline model, we observe a low correlation between different noise types in the deep layers, indicating that the network processes these two types of noise in different ways for the deep layers. This trend holds for other types of noise as well. This phenomenon may be due to the baseline approach causing", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1698", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3fddd29972407b25eb0ba26388ae16cf4ca0d4987c003dc878fd14a44fc23e3.jpg", + "image_caption": [ + "a. cross-model comparison on Gaussian 15", + "comparison on Gaussian 15" + ], + "image_footnote": [], + "bbox": [ + 81, + 95, + 254, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4929cb376de2f96ded1959e0264c7e399669e051f3e929323fd353344a590813.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 95, + 406, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbde860c33764b295d60c7fdb749738c9007f3b0a32055d8f11082fe4e4e5a48.jpg", + "image_caption": [ + "b. cross-model comparison between different noise" + ], + "image_footnote": [], + "bbox": [ + 428, + 89, + 730, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/575e359a04a23253b4fceb74e6dd6c94d57d558d1b82b7c46b6694096c9cfe8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 89, + 888, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd31056147495e4addd7698a002f31719a18a82e0294e3436227511547e51041.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 250, + 225, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24843181b752ba6cb26f219c182b874367b3f289b82de08d2b1d6dc354d9c340.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 330, + 223, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd9cd3d1e91ead4be9d2800c3ddefea49778b3c8844ac77460d30b615c27d707.jpg", + "image_caption": [ + "Figure 13. The testing curves on different noise types and levels." + ], + "image_footnote": [], + "bbox": [ + 80, + 402, + 223, + 497 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d901f586b46a50b95c734cc308f2b2eb0f458ceaa71c55efe7c0b94b38680f10.jpg", + "image_caption": [ + "Figure 12. CKA similarity to analyze the representation similarity of network layers." + ], + "image_footnote": [], + "bbox": [ + 225, + 250, + 375, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/706b6c0ba059f4dffb6c278418bf7d06dc4d9eab04aeeda61e76c75dc24e3a64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 330, + 375, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/126ee05a5548aa7e50e43737b33d4467de7efb88576397bf905f42be9b99f5fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 409, + 375, + 497 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26e6debaec8127ba4ca711dc9b726b75bc00918754ae9418f4a915a0f50d72c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 258, + 465, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/087d703e65e83232dee7dc7c9a57f52d68049a61204acab5b94b1cad1dc025ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 333, + 465, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f86d31ff0b155f92efa52cc796205bed8e4fd14f726dd549240d33d90840c1eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 411, + 464, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5782e4e4e87baf48f4bc27b7bf04607a901bcb521c7133923e3b5cb7460b5fad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 532, + 272, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6efdd3301d8b9d52835f72e402abe1214857688ea456390470346894ca6560e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 532, + 467, + 640 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/120fec10a589d031aed46f891b0e7313202a2f177e53ad9a6a969a1f8760c5e2.jpg", + "image_caption": [ + "Figure 14. Comparing generalization ability with the SRGA metric. A lower SRGA value indicates better generalization ability.", + "Figure 15. The distribution of baseline model features is biased across different noise types. Our method produces similar feature distributions across different noise." + ], + "image_footnote": [], + "bbox": [ + 86, + 676, + 276, + 787 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd0ea18c0d79baeac8e0ea5749127e47d2a7e71ea49cf52da8ae1fbd73c6242e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 676, + 455, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the deep layers of the model to overfit to the patterns of the training set, thereby limiting their generalization capabilities to handle different noise types. In contrast, the high correlation between adjacent layers in our masked training", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model suggests that the model's representation of two different noise types is similar. Our masked training forces the network to learn the underlying distribution of the images themselves, which makes the model more robust to different noise types and enhances its generalization capability.", + "bbox": [ + 496, + 253, + 890, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantification of generalization performance. Liu et al. [49, 50] suggest that model generalization ability can be measured by measuring the consistency of the model's representations across different types of noise. They also propose a generalization assessment index for low-level vision networks called SRGA [50]. It is a non-parametric and nonlearning metric which exploits the statistical characteristics of internal features of deep networks. The lower the value of SRGA, the better the generalization ability. In our case, we use Gaussian noise as the reference and other types of noise for testing. Figure 14 shows the SRGA results. Inspired by [50], we visualize the distributions of deep features on different noise types, shown in Figure 15. We can see that for the baseline model, the feature distributions under different noise types deviate from each other significantly. For the model $w$ / masked training, the deep feature distributions of different noise types are close to each other. This confirms the effectiveness of our method.", + "bbox": [ + 496, + 334, + 892, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 500, + 619, + 756, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents a masked training method to improve the generalization performance of deep learning-based image denoising models. The limitation of our method is that the mask operation inevitably loses information. How to preserve more details needs to be explored in future work. Our approach is a step towards developing more robust models for real-world applications.", + "bbox": [ + 496, + 643, + 890, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work is supported in part by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), the National Natural Science Foundation of China under Grant (62276251), NSFC General Project 62072452, Guangdong Provincial Basic and Applied Basic Research Fund-Regional Joint Fund (Project reference number 2020B1515130004), the Joint Lab of CAS-HK, and the Youth Innovation Promotion Association of Chinese Academy of Sciences (No. 2020356).", + "bbox": [ + 496, + 753, + 892, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1699", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1692-1700, 2018. 1, 2, 5", + "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3155-3164, 2019. 6, 7", + "[3] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 2010. 5", + "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 3", + "[5] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11036-11045, 2019. 1, 2, 5, 6, 7", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[7] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 2, pages 60-65. IEEE, 2005. 2", + "[8] Brent Burley, David Adler, Matt Jen-Yuan Chiang, Hank Driskill, Ralf Habel, Patrick Kelly, Peter Kutz, Yining Karl Li, and Daniel Teece. The design and evolution of disney's hyperion renderer. ACM Transactions on Graphics (TOG), 37(3):1-22, 2018. 7", + "[9] Chang Chen, Zhiwei Xiong, Xinmei Tian, and Feng Wu. Deep boosting for image denoising. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2", + "[10] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2", + "[11] Jingwen Chen, Jiawei Chen, Hongyang Chao, and Ming Yang. Image blind denoising with generative adversarial network based noise modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3155-3164, 2018. 1, 2", + "[12] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 3", + "[13] Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE transactions on pattern analysis and machine intelligence, 39(6):1256-1272, 2016. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, and Xiaokang Yang. Recursive generalization transformer for image super-resolution. arXiv preprint arXiv:2303.06373, 2023. 2", + "[15] Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. In NIPS, 2022. 2, 4", + "[16] Shen Cheng, Yuzhi Wang, Haibin Huang, Donghao Liu, Haoqiang Fan, and Shuaicheng Liu. Nbnet: Noise basis learning for image denoising with subspace projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4896-4906, 2021. 2", + "[17] Per Christensen, Julian Fong, Jonathan Shade, Wayne Wooten, Brenden Schubert, Andrew Kensler, Stephen Friedman, Charlie Kilpatrick, Cliff Ramshaw, Marc Bannister, et al. Renderman: An advanced path-tracing architecture for movie rendering. ACM Transactions on Graphics (TOG), 37(3):1-21, 2018. 7", + "[18] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. The Journal of Machine Learning Research, 13:795-828, 2012. 7", + "[19] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2", + "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3", + "[21] Nithish Divakar and R Venkatesh Babu. Image denoising via cnns: An adversarial approach. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 80-87, 2017. 2", + "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 4", + "[23] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image processing, 15(12):3736-3745, 2006. 2", + "[24] Ruicheng Feng, Jinjin Gu, Yu Qiao, and Chao Dong. Suppressing model overfitting for image super-resolution networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 2", + "[25] Arthur Firmino, Jeppe Revall Frisvad, and Henrik Wann Jensen. Progressive denoising of monte carlo rendered images. In Computer Graphics Forum, volume 41, pages 1-11. Wiley Online Library, 2022. 7", + "[26] Rich Franzen. Kodak lossless true color image suite. source: http://r0k.us/graphics/kodak/, 1999. 5, 6, 7", + "[27] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for percep" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1700", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 5", + "[28] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 5", + "[29] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9199-9208, 2021. 3", + "[30] Jinjin Gu, Hannan Lu, Wangmeng Zuo, and Chao Dong. Blind super-resolution with iterative kernel correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1604-1613, 2019. 2", + "[31] Shuhang Gu, Yawei Li, Luc Van Gool, and Radu Timofte. Self-guided network for fast image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2511-2520, 2019. 2", + "[32] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2", + "[33] Shi Guo, Zifeei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1722, 2019. 2", + "[34] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3", + "[35] Xiaowan Hu, Ruijun Ma, Zhihong Liu, Yuanhao Cai, Xiaole Zhao, Yulun Zhang, and Haoqian Wang. Pseudo 3d autocorrelation network for real image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16175-16184, 2021. 2", + "[36] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5", + "[37] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6054-6063, 2019. 2", + "[38] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[39] Filippos Kokkinos and Stamatios Lefkimmiatis. Deep image demosaicking using a cascade of convolutional residual denoising networks. In Proceedings of the European conference on computer vision (ECCV), pages 303-319, 2018. 2", + "[40] Xiangtao Kong, Xina Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Reflash dropout in image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6002-6012, 2022. 2, 7" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 2", + "[42] Christopher Kulla, Alejandro Conty, Clifford Stein, and Larry Gritz. Sony pictures imageworks arnold. ACM Transactions on Graphics (TOG), 37(3):1-18, 2018. 7", + "[43] Stamatios Lefkimmiatis. Non-local color image denoising with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3587-3596, 2017. 2", + "[44] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3204-3213, 2018. 2", + "[45] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In CVPR, 2021. 2, 3, 4, 6, 7", + "[46] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 5", + "[47] Anran Liu, Yihao Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Blind image super-resolution: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[48] Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. Advances in neural information processing systems, 31, 2018. 2", + "[49] Yihao Liu, Anran Liu, Jinjin Gu, Zhipeng Zhang, Henhao Wu, Yu Qiao, and Chao Dong. “Discovering” semantics” in super-resolution networks. arXiv preprint arXiv:2108.00406, 2021. 2, 3, 8", + "[50] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. arXiv preprint arXiv:2205.07019, 2022. 3, 8", + "[51] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 2016. 5", + "[52] Salma Abdel Magid, Zudi Lin, Donglai Wei, Yulun Zhang, Jinjin Gu, and Hanspeter Pfister. Texture-based error analysis for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2118-2127, 2022. 3", + "[53] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In 2009 IEEE 12th international conference on computer vision, pages 2272-2279. IEEE, 2009. 2", + "[54] Xiaojiao Mao, Chunhua Shen, and Yu-Bin Yang. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections. Advances in neural information processing systems, 29, 2016. 2", + "[55] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1701", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 4, 5, 7", + "[56] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3", + "[57] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1586-1595, 2017. 2", + "[58] Tobias Plötz and Stefan Roth. Neural nearest neighbors networks. Advances in Neural information processing systems, 31, 2018. 2", + "[59] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3", + "[60] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3", + "[61] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 7", + "[62] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. arXiv preprint arXiv:2207.08494, 2022. 2", + "[63] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 2", + "[64] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, Kyoung Mu Lee, et al. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, 2017. 5", + "[65] Mathias Uhlen, Per Oksvold, Linn Fagerberg, Emma Lundberg, Kalle Jonasson, Mattias Forsberg, Martin Zwahlen, Caroline Kampf, Kenneth Wester, Sophia Hober, et al. Towards a knowledge-based human protein atlas. Nature biotechnology, 28(12):1248-1250, 2010. 3", + "[66] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3", + "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 2", + "[68] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758-2767, 2020. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2, 3", + "[70] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In CVPR, 2020. 2", + "[71] Yuan Yuan, Siyuan Liu, Jiawei Zhang, Yongbing Zhang, Chao Dong, and Liang Lin. Unsupervised image superresolution using cycle-in-cycle generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 701-710, 2018. 1, 2", + "[72] Zongsheng Yue, Hongwei Yong, Qian Zhao, Deyu Meng, and Lei Zhang. Variational denoising network: Toward blind noise modeling and removal. Advances in neural information processing systems, 32, 2019. 2", + "[73] Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. In European Conference on Computer Vision, pages 41-58. Springer, 2020. 2", + "[74] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2", + "[75] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 6, 7", + "[76] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2", + "[77] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. arXiv preprint arXiv:2303.06440, 2023. 2", + "[78] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. arXiv preprint arXiv:2210.01427, 2022. 2, 4", + "[79] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swim-conv-unet and data synthesis. arXiv preprint arXiv:2203.13278, 2022. 2, 5", + "[80] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 2, 6, 7", + "[81] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 2" + ], + "bbox": [ + 501, + 92, + 892, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1702", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2):023016, 2011. 5", + "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[84] Yulun Zhang, Kunpeng Li, Kai Li, Gan Sun, Yu Kong, and Yun Fu. Accurate and fast image denoising via attention guided scaling. IEEE Transactions on Image Processing, 30:6255-6265, 2021. 2", + "[85] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2, 6, 7", + "[86] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480-2495, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 345 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1703", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_model.json b/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..efcc9e048c905d8696adb8b15feb85e15c45c502 --- /dev/null +++ b/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_model.json @@ -0,0 +1,3656 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.815, + 0.154 + ], + "angle": 0, + "content": "Masked Image Training for Generalizable Deep Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.18, + 0.768, + 0.218 + ], + "angle": 0, + "content": "Haoyu Chen\\(^{1*}\\), Jinjin Gu\\(^{2,3*}\\), Yihao Liu\\(^{2,4,5}\\), Salma Abdel Magid\\(^{7}\\), Chao Dong\\(^{2,4}\\), Qiong Wang\\(^{6}\\), Hanspeter Pfister\\(^{7}\\), Lei Zhu\\(^{1,8\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.223, + 0.862, + 0.321 + ], + "angle": 0, + "content": "\\(^{1}\\)The Hong Kong University of Science and Technology (Guangzhou) \\(^{2}\\)Shanghai AI Lab \\(^{3}\\)The University of Sydney \n\\(^{4}\\)ShenZhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n\\(^{5}\\)University of Chinese Academy of Sciences \\(^{6}\\)Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n\\(^{7}\\)Harvard University \\(^{8}\\)The Hong Kong University of Science and Technology \nProject page: https://github.com/haoyuc/MaskedDenoising" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.356, + 0.313, + 0.372 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.388, + 0.474, + 0.691 + ], + "angle": 0, + "content": "When capturing and storing images, devices inevitably introduce noise. Reducing this noise is a critical task called image denoising. Deep learning has become the de facto method for image denoising, especially with the emergence of Transformer-based models that have achieved notable state-of-the-art results on various image tasks. However, deep learning-based methods often suffer from a lack of generalization ability. For example, deep models trained on Gaussian noise may perform poorly when tested on other noise distributions. To address this issue, we present a novel approach to enhance the generalization performance of denoising networks, known as masked training. Our method involves masking random pixels of the input image and reconstructing the missing information during training. We also mask out the features in the self-attention layers to avoid the impact of training-testing inconsistency. Our approach exhibits better generalization ability than other deep learning models and is directly applicable to real-world scenarios. Additionally, our interpretability analysis demonstrates the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.706, + 0.21, + 0.722 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Image denoising is a crucial research area that aims to recover clean images from noisy observations. Due to the rapid advancements in deep learning, many promising image denoising networks have been developed. These networks are typically trained using images synthesized from a pre-defined noise distribution and can achieve remarkable performance in removing the corresponding noise. However, a significant challenge in applying these deep models to real-world scenarios is their generalization ability. Since" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.355, + 0.58, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.437, + 0.571, + 0.446 + ], + "angle": 0, + "content": "Gaussian 15" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.355, + 0.657, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.438, + 0.636, + 0.446 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.355, + 0.736, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.438, + 0.732, + 0.446 + ], + "angle": 0, + "content": "Mixture noise" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.355, + 0.818, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.438, + 0.795, + 0.446 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.355, + 0.892, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.438, + 0.865, + 0.446 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.449, + 0.627, + 0.459 + ], + "angle": 0, + "content": "a. In-distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.449, + 0.834, + 0.459 + ], + "angle": 0, + "content": "b. Out-of-distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.465, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Figure 1. We illustrate the generalization problem of denoising networks. We train a SwinIR model on Gaussian noise with \\(\\sigma = 15\\). When tested on the same noise, SwinIR demonstrates outstanding performance. However, when applied to out-of-distribution noise, e.g., the mixture of various noise. SwinIR suffers from a huge performance drop. The model trained by the proposed masked training method maintains a reasonable denoising effect, despite aldo being trained on Gaussian noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.891, + 0.64 + ], + "angle": 0, + "content": "the real-world noise distribution can differ from that observed during training, these models often struggle to generalize to such scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "More specifically, most existing denoising works train and evaluate models on images corrupted with Gaussian noise, limiting their performance to a single noise distribution. When these models are applied to remove noise drawn from other distributions, their performance drastically drops. Figure 1 shows an example. The research community has become increasingly aware of this generalization issue of deep models in recent years. As a countermeasure, some methods [80] assume that the noise level of a particular noise type is unknown, while others [5, 68] attempt to improve the performance in real-world scenarios by synthesizing or collecting training data closer to the target noise or directly performing unsupervised training on the target noise [11, 71]. However, none of these methods substantially improve the generalization performance of denoising networks, and they still struggle when the noise distribution is mismatched [1]. The generalization issue of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.411, + 0.887 + ], + "angle": 0, + "content": "*Haoyu Chen and Jinjin Gu contribute equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.383, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) Lei Zhu (leizhu@ust.hk) is the corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.411, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1692" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.467, + 0.121 + ], + "angle": 0, + "content": "deep denoising still poses challenges to making these methods broadly applicable." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.123, + 0.468, + 0.378 + ], + "angle": 0, + "content": "In this work, we focus on improving the generalization ability of deep denoising models. We define generalization ability as the model's performance on noise different from what it observed during training. We argue that the generalization issue of deep denoising is due to the overfitting of training noise. The existing training strategy directly optimizes the similarity between the denoised image and the ground truth. The intention behind this is that the network should learn to reconstruct the texture and semantics of natural images correctly. However, what is often overlooked is that the network can also reduce the loss simply by overfitting the noise pattern, which is easier than learning the image content. This is at the heart of the generalization problem. Even many popular deep learning methods exacerbate this overfitting problem. When it comes to noise different from that observed during training, the network exhibits this same behavior, resulting in poor performance." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.38, + 0.468, + 0.755 + ], + "angle": 0, + "content": "In light of the preceding discussion, our study seeks to improve the generalization performance of deep denoising networks by directing them to learn image content reconstruction instead of overfitting to training noise. Drawing inspiration from recent masked modeling methods [4, 20, 34, 69], we employ a masked training strategy to explicitly learn representations for image content reconstruction, as opposed to training noise. Leveraging the properties of image processing Transformers [15, 45, 78], we introduce two masking mechanisms: the input mask and the attention mask. During training, the input mask removes input image pixels randomly, and the network reconstructs the removed pixels. The attention mask is implemented in each self-attention layer of the Transformer, enabling it to learn the completion of masked features dynamically and mitigate the distribution shift between training and testing in masked learning. Although we use Gaussian noise for training - similar to previous works - our method demonstrates significant performance improvements on various noise types, such as speckle noise, Poisson noise, salt and pepper noise, spatially correlated Gaussian noise, Monte Carlo-rendered image noise, ISP noise, and complex mixtures of multiple noise sources. Existing methods and models have yet to effectively and accurately remove all these diverse noise patterns." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.77, + 0.225, + 0.785 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.796, + 0.467, + 0.9 + ], + "angle": 0, + "content": "Image Denoising approaches very broadly lie in two categories: traditional model-based and data-driven deep-learning-based. Traditional methods are usually based on modeling image priors to recover image content contaminated by noise [7, 19, 23, 32, 53]. These methods usually do not impose too many constraints on the type of noise, and have been proven to be applicable to a variety of noise," + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.483 + ], + "angle": 0, + "content": "with good generalization performance [1]. However, these methods are not satisfactory for the reconstruction of image content. In recent years, the paradigm of denoising has gradually shifted to data-driven methods based on deep learning methods [13]. Many techniques have been proposed to improve the capabilities of the denoising networks continuously, e.g., residual networks [39, 80, 81], dense networks [37, 86], recursive networks [9, 48, 63], multi-scale [21, 31, 76], encoder-decoder [16, 54, 73], attention operations [84, 85], self-similarity [35], and non-local operations [43, 44, 58]. Since 2020, the paradigm of vision network design has gradually shifted from CNNs to Transformers [22]. Vision Transformers treat input pixels as tokens and use self-attention operations to process interactions between these tokens. Inspired by the success of vision Transformers, many attempts have been made to employ Transformers for low-level vision tasks [10, 14, 15, 45, 62, 67, 70, 74, 77, 78] During the development of these models, the noise pattern used for training is often consistent with the testing one. The factor that determines its denoising performance is the fitting ability of the network, in other words, the ability of the network to overfit to the training noise. However, a better network does not mean a better generalization ability of the denoising model. As we will show in the experiment section, a more efficient network even indicates worse generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.494, + 0.891, + 0.899 + ], + "angle": 0, + "content": "Generalization Problem in low-level vision often arises when the testing degradation does not match the training degradation, e.g., different downsampling kernel in superresolution [30, 40, 47]. We typically develop deep denoising models based on Gaussian noise in the laboratory setting. However, noise in the real-world is mostly non-Gaussian. Models trained on Gaussian noise fail in these non-Gaussian scenarios. There are two main categories of solutions to this problem. The first is to make training datasets with noise modeling as close to reality as possible during development, e.g., synthesizing real noise according to physical system modeling [5, 68], learning to generate real noise [11, 24, 71], collecting real noise - clean image pairs for training [1, 33, 41, 57]. Although the models obtained by these methods can improve the effect on the target noise, they still cannot generalize to out-of-distribution noise. Another category of solutions is to develop \"blind\" denoising models, which are supposed to deal with unknown noise [41, 72, 80]. These methods usually simply assume that the noise level is unknown, or train on a large amount of noise types [79], which also fails to generalize to other noise not present in the training set. Few workd have been proposed to study the reasons for the lack of generalization ability in low-level vision [40]. Liu et al. [49] argue that networks tend to overfit to degradations and show degradation \"semantics\" inside the network. The presence of these representations often means a decrease in general-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1693" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.229, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.204, + 0.192, + 0.215 + ], + "angle": 0, + "content": "Training image" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.089, + 0.348, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.142, + 0.32, + 0.151 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.469, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.142, + 0.457, + 0.151 + ], + "angle": 0, + "content": "Denoised by SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.152, + 0.348, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.204, + 0.346, + 0.214 + ], + "angle": 0, + "content": "Denoised by our method" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.152, + 0.469, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.205, + 0.467, + 0.214 + ], + "angle": 0, + "content": "Our reconstruction result" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.47, + 0.362 + ], + "angle": 0, + "content": "Figure 2. SwinIR, when trained solely on immunohistochemistry images with Gaussian noise, can still denoise natural images. This observation supports the assertion that most existing methods perform denoising primarily through overfitting the training noise. In contrast, our approach emphasizes reconstructing natural image textures and edges observed in the training set on natural images, rather than relying on noise overfitting for denoising. This distinction underlines the fundamental difference between our method and previous approaches. \"Our reconstruction result\" refers to using our model but taking masked images as input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.469, + 0.437 + ], + "angle": 0, + "content": "ization ability. The utilization of this knowledge can guide us to analyze and evaluate the generalization performance [50]. Apart from that, few works have been proposed to improve the generalization ability of denoising models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.655 + ], + "angle": 0, + "content": "Masked modeling for language [6, 20, 59, 60] is successful for learning pre-trained representations that generalize well to various downstream tasks. These methods mask out a portion of the input sequence and train models to predict the missing content. A similar approach can also be applied to the vision model pre-training. Masked image models learn representations from corrupted images. The earliest attempts in this regard can be traced back at least to the denoising auto-encoder [66]. Since then, many works have used predicting missing parts of images to learn efficient image representations [4, 12, 34, 56, 69]. However, there have been few successful attempts to apply masked image modeling to low-level vision, even though the masked pretraining method is in the form of low-level vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.168, + 0.687 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.758 + ], + "angle": 0, + "content": "Our objective is to create denoising models capable of generalizing to noise not encountered in the training set. In this section, we first discuss our motivation before delving into the specifics of our masked training method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Motivation. When training a deep network on a large number of images, the expectation is for the network to learn to discern the rich semantics of natural images from noise-contaminated test cases. However, several studies have noted that the semantics and knowledge acquired by low-level vision networks differ significantly from our expectations [29, 49, 50, 52]. We argue that the poor generalization ability of denoising models results from our training method, which leads the model to focus on overfit" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.58, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.089, + 0.735, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.151, + 0.581, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.213, + 0.564, + 0.223 + ], + "angle": 0, + "content": "original" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.151, + 0.735, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.213, + 0.689, + 0.223 + ], + "angle": 0, + "content": "mask \\(70\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.089, + 0.892, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.151, + 0.892, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.213, + 0.843, + 0.222 + ], + "angle": 0, + "content": "mask \\(90\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.232, + 0.892, + 0.274 + ], + "angle": 0, + "content": "Figure 3. The illustration of the proposed mask-and-complete training strategy. Even if a large number of pixels are masked, the model can still reconstruct the input to some extent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.501 + ], + "angle": 0, + "content": "ting the training noise rather than learning image reconstruction. We conduct a simple experiment for verification. We trained a SwinIR denoising network [45] using images that greatly differ from natural images (immunohistochemistry images [65]). We synthesized training data pairs using Gaussian noise, and then assessed the model's performance on natural images with Gaussian noise. According to our hypothesis, if the model learns the content and reconstruction of image semantics from the training set, it should not perform well on natural images, as it has not been exposed to any. If the model is simply overfitting the noise, the model can remove the noise even if the images are different, as the model mainly relies on detecting the noise for denoising." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.626 + ], + "angle": 0, + "content": "The results are presented in Figure 2. As observed, the SwinIR trained on immunohistochemistry images can still denoise and reproduce the natural image. This supports our conjecture regarding generalization ability, indicating that most existing methods perform denoising by overfitting the training noise. Consequently, when the noise deviates from the training conditions, the denoising performance of these models declines significantly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This observation also inspires our approach to developing deep denoising models with improved generalization ability. We aim for the model to learn the reconstruction of image textures and structures, rather than focusing only on noise. In this paper, we propose a new masked training strategy for denoising networks. During training, we mask out a portion of the input pixels and then train the deep network to complete them, as shown in Figure 3. Our approach emphasizes reconstructing natural image textures and edges observed in the image, rather than overfitting noise. In Figure 2 we also show the results of our method. It is evident that our approach seeks to reconstruct the immunohistochemistry image texture from the training set on the testing natural image, instead of relying on noise overfitting for denoising. This demonstrates the potential of this idea in improving generalization performance. By training our method on natural images, it will concentrate on reconstructing the content of natural images, aligning with our" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.2, + 0.895, + 0.23 + ], + "angle": 0, + "content": "Figure 4. The transformer architecture of our proposed masked image training. We make a minimal change to the original SwinIR architecture – the input mask operation and the attention masks. Other micro-designs are not essentially different from other Transformers." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.241, + 0.473, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.414 + ], + "angle": 0, + "content": "Figure 5. Quantitative effect of the attention mask. The histogram differences are also shown above." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.455 + ], + "angle": 0, + "content": "core concept of employing deep learning for low-level vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The Transformer Architecture. Our approach exploits the excellent properties of visual Transformers, so we first describe the basic Transformer backbone used in this study. The shifted window mechanism is proven to be flexible and effective for image processing tasks [15, 45, 78]. We only make minimal changes when applying it to the proposed masked training method without the loss of generality. This model is illustrated in Figure 4. Transformers divide the input signal into tokens and process spatial information using self-attention layers. In our method, a convolution layer with kernel size 1 is used as the feature embedding module to project the 3-channel pixel values into \\(C\\)-dimensional feature tokens. The \\(1 \\times 1\\) convolution layer ensures that pixels do not affect each other during feature embedding, which facilitates subsequent masking operations. These feature tokens are gathered with shape \\(H \\times W \\times C\\), where \\(H\\), \\(W\\) and \\(C\\) are the height, width and feature dimension. The shifted window mechanism first reshapes the feature maps of each frame to \\(\\frac{HW}{M^2} \\times M^2 \\times C\\) features by partitioning the input into non-overlapping \\(M \\times M\\) local windows, where \\(\\frac{HW}{M^2}\\) is the total number of windows. We calculate self-attention on the feature tokens within the same window. Therefore, \\(M^2\\) tokens are involved in each standard self-attention operation, and we produce the local window feature \\(X \\in \\mathbb{R}^{M^2 \\times C}\\). In each self-attention layer, the query \\(Q\\), key \\(K\\) and value \\(V\\) are calculated as \\(Q = XW^Q\\), \\(K = XW^K\\), \\(V = XW^V\\), where \\(W^Q, W^K, W^V \\in \\mathbb{R}^{C \\times D}\\) are weight matrices, and \\(D\\) is the dimension of projected vectors. Then, we use \\(Q\\) to query \\(K\\) to generate the attenuated feature vectors." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.242, + 0.895, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.315, + 0.892, + 0.343 + ], + "angle": 0, + "content": "Figure 6. The effectiveness of the input mask and attention mask. Note that the brightness of the image is wrong w/o attention mask." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.351, + 0.88, + 0.422 + ], + "angle": 0, + "content": "
Input MaskAttention MaskPSNRSSIMMix. noise on CBSD68 [55]
Ratio (%)PSNRSSIM
✓29.170.82276529.570.8657
✓26.960.82027529.760.8678
✓✓29.740.86728528.840.8548
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.423, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Table 1. The importance of using Table 2. Ablation on the at-different mask operations. tention mask ratio." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.892, + 0.538 + ], + "angle": 0, + "content": "tion map \\(A = \\operatorname{softmax}\\left(\\frac{QK^T}{\\sqrt{D}} + B\\right) \\in \\mathbb{R}^{M^2 \\times M^2}\\), where \\(B\\) is the learnable relative positional encoding. This attention map \\(A\\) is then used for the weighted sum of \\(M^2\\) vectors in \\(V\\). The multi-head settings are aligned with SwinIR [45] and ViT [22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Masked Training. Our masked training mainly consists of two aspects, the input mask and the attention mask. Although both are mask operations, the purpose of these two masks is different. We describe them separately." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.883 + ], + "angle": 0, + "content": "The Input Mask randomly masks out the feature tokens embedded by the first convolution layer, and encourages the network to complete the masked information during training. The input mask explicitly constructs a very challenging inpainting problem, as shown in Figure 3. It can be seen that even if up to \\(90\\%\\) of the pixel information is destroyed, the network can still reconstruct the target image to a certain extent. The method is very simple. Given the feature token tensor \\(\\mathbf{f} \\in \\mathbb{R}^{H \\times W \\times C}\\), we randomly replace the token with a [mask token] \\(\\in \\mathbb{R}^C\\) with a probability \\(p_{\\mathrm{IM}}\\), where \\(p_{\\mathrm{IM}}\\) is called the input mask ratio. The network is trained under the supervision of the \\(l_1\\)-norm of the reconstructed image and the ground truth. The [mask token] can be learnable and initialized with a 0 vector. But we actually found that the 0 vector itself is already a suitable choice. The existence of the input mask forces the network to learn to recognize and reconstruct the content of the image from very limited information." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The Attention Mask. We cannot build usable image pro" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1695" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.167, + 0.47, + 0.224 + ], + "angle": 0, + "content": "Figure 7. The trade-off of choosing different mask ratios. The performance drop on training noise is not significant until \\(75\\%\\) masking ratio. Our performance gain on the noise outside the training set is greater than the performance loss on the training set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.471, + 0.538 + ], + "angle": 0, + "content": "cessing networks relying solely on the input mask operation. Because during testing, we will input uncorrupted images to retain enough information. At this time, due to the inconsistency between training and testing, the network will tend to increase the brightness of the output image. Such as the example in Figure 5. Since Transformer uses the self-attention operation to process spatial information, we can narrow the gap between training and testing by performing the same mask operation during the self-attention process. The specific mask operation is similar to the input mask, but a different attention mask ratio \\( p_{\\mathrm{AM}} \\) and [mask token] are used. When some tokens in the self-attention are masked, the attention operation will adjust to the fact that the information of these tokens is no longer reliable. Self-attention will focus on unmasked tokens in each layer and complete the masked information. This operation is difficult to implement on convolutional networks. Figure 5 shows the effect of the attention mask. As can be seen, the attention mask successfully makes the masked trained network work on the unmasked input image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.552, + 0.21, + 0.57 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.744 + ], + "angle": 0, + "content": "Training Settings. For synthesizing training data, we sample the clean images from DIV2K [64], Flickr2K [46], BSD500 [3], and WED [51] during training. In our work, all the networks are trained using Gaussian noise with standard deviation \\(\\sigma = 15\\). Each input image is randomly cropped to a spatial resolution of \\(64 \\times 64\\), and the number of the total training iteration is \\(200\\mathrm{K}\\). We adopt the Adam optimizer [38] with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\) to minimize the \\(L_{1}\\) pixel loss. The initial learning rate is set as \\(1 \\times 10^{-4}\\) and reduced by half at the milestone of \\(100\\mathrm{K}\\) iterations and \\(150\\mathrm{K}\\) iterations. The batch size is set to 64." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Testing Noise. Since the training process utilizes the Gaussian noise, we evaluate the generalization performance of the models on six other synthetic noise: (1) Speckle noise, a type of noise that occurs during the acquisition of medical images or tomography images. (2) Poisson noise, a type of signal-dependent noise that occurs during the acquisition of digital images.(3) Spatially-correlated noise. This is to synthesize the complex artifact after denoising using a flawed algorithm. It is produced by filtering Gaussian noise with a \\(3 \\times 3\\) average kernel. Different standard deviations of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "the Gaussian noise indicate different noise levels. (4) Salt & pepper noise. (5) Image signal processing (ISP) noise. [5] proposes a method to synthesize realistic ISP noise during digital imaging. (6) Mixture noise obtained by mixing the above different types of noise with different levels [79]. The clean images are sampled from the benchmark datasets, including CBSD68 [55], Kodak24 [26], McMaster [82], and Urban100 [36]. We also include two real noise types in this work: the Smartphone Image Denoising Dataset (SIDD) [1] and Monte Carlo (MC) rendered image noise. For evaluation, we follow [27,28] and use the metrics PSNR, SSIM [51], and LPIPS [83] to evaluate the results. Since PSNR and SSIM are questioned in assessing the perceptual quality of images [27,28], we also use the LPIPS as an additional metric." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.328, + 0.593, + 0.344 + ], + "angle": 0, + "content": "4.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Ablation Study. Table 1 and Figure 6 show the effectiveness of using different mask operations. As we can see, without the input mask, the model will lose its generalization ability, and cannot effectively remove the noise outside the training set. Without the attention mask, due to the training-testing inconsistency, the quantitative performance degrades significantly, and the output image will have the wrong brightness. In addition, even without the attention mask, the generalization ability of the model is not significantly affected, and most of the noise is still effectively removed. The input mask is the crucial factor in improving the model the generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Table 3a shows the impact of the different input mask ratios. We test fixed ratios and random ratios from a uniform distribution. From our experiments, fixed ratios are less stable for training than randomly chosen from a range, and the performance is also worse. The best quantitative performance is achieved with random sampling ratios between \\(75\\% \\sim 85\\%\\). This is a trade-off between denoising generalization ability and the preservation of image details. As shown in Figure 7, smaller ratios are not enough for the network to learn the distribution of images because more noise patterns are preserved. The larger ratio improves the model generalization, as the model focuses more on reconstruction. But at the same time, some image details may be lost. For attention mask ratio, we show the effects in Table 2. The optimal ratios are around \\(75\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The Generalization Performance. We evaluate our deep denoising method on synthetic noise, where our training noise follows a Gaussian distribution with a single noise level, but we test on multiple types of non-Gaussian noise to assess the model's generalization performance. In Figure 11, we compare our method with other state-of-the-art models based on their PSNR and SSIM scores. The results show that our model outperforms all the other models in terms of generalization performance. Particularly, as" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1696" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.29, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.218, + 0.232, + 0.228 + ], + "angle": 0, + "content": "CBSD68: img_053" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.089, + 0.44, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.15, + 0.437, + 0.159 + ], + "angle": 0, + "content": "Spatially correlated Gaussian, \\(\\sigma = 50\\)" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.089, + 0.589, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.149, + 0.546, + 0.159 + ], + "angle": 0, + "content": "DnCNN [80]" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.089, + 0.738, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.149, + 0.691, + 0.159 + ], + "angle": 0, + "content": "RIDNet [2]" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.089, + 0.888, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.149, + 0.842, + 0.159 + ], + "angle": 0, + "content": "RNAN [85]" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.228, + 0.29, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.356, + 0.235, + 0.367 + ], + "angle": 0, + "content": "CBSD68: img_0046" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.16, + 0.44, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.218, + 0.402, + 0.229 + ], + "angle": 0, + "content": "Restormer [75]" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.16, + 0.588, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.218, + 0.545, + 0.229 + ], + "angle": 0, + "content": "SwinIR [45]" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.16, + 0.738, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.218, + 0.684, + 0.228 + ], + "angle": 0, + "content": "baseline" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.16, + 0.887, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.218, + 0.853, + 0.229 + ], + "angle": 0, + "content": "Masked Training" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.367, + 0.291, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.495, + 0.235, + 0.506 + ], + "angle": 0, + "content": "CBSD68: img_0067" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.228, + 0.44, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.286, + 0.433, + 0.298 + ], + "angle": 0, + "content": "Speckle noise, \\(\\sigma^2 = 0.013\\)" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.228, + 0.588, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.287, + 0.546, + 0.299 + ], + "angle": 0, + "content": "DnCNN [80]" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.228, + 0.738, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.288, + 0.691, + 0.299 + ], + "angle": 0, + "content": "RIDNet [2]" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.228, + 0.887, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.288, + 0.842, + 0.299 + ], + "angle": 0, + "content": "RNAN [85]" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.495, + 0.29, + 0.506 + ], + "angle": 0, + "content": "CBSD68: img_0067" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.298, + 0.44, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.357, + 0.402, + 0.367 + ], + "angle": 0, + "content": "Restormer [75]" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.298, + 0.588, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.357, + 0.545, + 0.367 + ], + "angle": 0, + "content": "SwinIR [45]" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.298, + 0.738, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.357, + 0.684, + 0.367 + ], + "angle": 0, + "content": "baseline" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.298, + 0.887, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.357, + 0.853, + 0.367 + ], + "angle": 0, + "content": "Masked Training" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.367, + 0.291, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.367, + 0.44, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.427, + 0.435, + 0.436 + ], + "angle": 0, + "content": "Salt-and-pepper noise, \\(d = 0.02\\)" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.367, + 0.588, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.427, + 0.545, + 0.436 + ], + "angle": 0, + "content": "DnCNN [80]" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.367, + 0.738, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.427, + 0.691, + 0.437 + ], + "angle": 0, + "content": "RIDNet [2]" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.367, + 0.887, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.427, + 0.842, + 0.437 + ], + "angle": 0, + "content": "RNAN [85]" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.495, + 0.29, + 0.506 + ], + "angle": 0, + "content": "CBSD68: img_0067" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.437, + 0.44, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.495, + 0.402, + 0.506 + ], + "angle": 0, + "content": "Restormer [75]" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.437, + 0.588, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.495, + 0.545, + 0.506 + ], + "angle": 0, + "content": "SwinIR [45]" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.437, + 0.738, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.495, + 0.684, + 0.505 + ], + "angle": 0, + "content": "baseline" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.437, + 0.887, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.495, + 0.853, + 0.506 + ], + "angle": 0, + "content": "Masked Training" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.509, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Figure 8. Visual comparison on out-of-distribution noise. When all other methods fail completely, our method is still able to denoise effectively. Please refer to the supplementary material to see more visual results." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.547, + 0.212, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.682, + 0.204, + 0.692 + ], + "angle": 0, + "content": "Rendered image (256 spp)" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.547, + 0.338, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.611, + 0.291, + 0.62 + ], + "angle": 0, + "content": "Noisy" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.547, + 0.464, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.611, + 0.425, + 0.62 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.547, + 0.593, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.611, + 0.548, + 0.62 + ], + "angle": 0, + "content": "DnCNN" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.547, + 0.702, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.611, + 0.67, + 0.621 + ], + "angle": 0, + "content": "Noisy" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.547, + 0.794, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.611, + 0.77, + 0.62 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.547, + 0.888, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.611, + 0.859, + 0.62 + ], + "angle": 0, + "content": "DnCNN" + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.694, + 0.536, + 0.708 + ], + "angle": 0, + "content": "Figure 9. Visual results of denoising a Monte Carlo rendered image." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.612, + 0.338, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.683, + 0.293, + 0.692 + ], + "angle": 0, + "content": "RNAN" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.612, + 0.464, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.683, + 0.421, + 0.692 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.612, + 0.593, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.683, + 0.541, + 0.692 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.621, + 0.702, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.683, + 0.671, + 0.692 + ], + "angle": 0, + "content": "RNAN" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.621, + 0.794, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.683, + 0.765, + 0.692 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.621, + 0.887, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.683, + 0.853, + 0.692 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.694, + 0.868, + 0.708 + ], + "angle": 0, + "content": "Figure 10. Results of ISP noise removal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.879 + ], + "angle": 0, + "content": "the noise level increases, our model exhibits a slower performance degradation and thus demonstrates better generalization. In contrast, other models suffer from significant performance drops when dealing with more severe noise. We also provide visual comparisons in Figure 8, where our model achieves remarkable denoising results even though it is trained only on Gaussian noise with a fixed standard deviation. In contrast, existing models tend to overfit the training noise and fail when facing unseen noise. More quantitative and qualitative results can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.894 + ], + "angle": 0, + "content": "Evaluation on ISP noise. The removal of the ISP noise is of great application value. Brooks et al. [5] present a systematic approach for generating realistic raw data with ISP noise that can facilitate our research. We use the default parameter settings of the method proposed in [5] to synthesize ISP noise on the Kodak24 [26] dataset for testing. The results are shown in Figure 10 and Table 3c. Our method achieves superior results compared to all other methods. Notably, our method achieves a significant lead in LPIPS, indicating that our results exhibit better perceptual quality. Although DnCNN and our method obtain the same PSNR, our method still outperforms DnCNN in terms of SSIM and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1697" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.09, + 0.089, + 0.253, + 0.221 + ], + "angle": 0, + "content": "
Mix. noise on CBSD68 [55]
Ratio (%)PSNRSSIM
7529.170.8132
8529.440.8545
9519.600.7273
70–8029.860.8593
75–8530.040.8756
75–9029.870.8728
75–9529.260.8607
80–9029.740.8672
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.223, + 0.246, + 0.236 + ], + "angle": 0, + "content": "(a) Abl. of input mask ratios." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.089, + 0.64, + 0.221 + ], + "angle": 0, + "content": "
Method128 samples per pixel64 samples per pixel
PSNRSSIMLPIPSPSNRSSIMLPIPS
DnCNN [80]29.940.78830.267126.280.67790.4216
RIDNet [2]29.960.79210.254826.270.67880.4122
RNAN [85]29.860.78250.270226.260.67430.4290
SwinIR [45]29.320.76270.294326.140.66510.4485
Restormer [75]24.980.65980.457524.590.58800.5375
Dropout [40]28.850.77530.294126.100.66960.4443
baseline29.680.77380.285125.910.65350.4564
Ours30.620.85000.225428.250.76940.3348
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.259, + 0.223, + 0.636, + 0.236 + ], + "angle": 0, + "content": "(b) Quantitative comparison on Monte Carlo rendered image denoising." + }, + { + "type": "table", + "bbox": [ + 0.647, + 0.089, + 0.885, + 0.221 + ], + "angle": 0, + "content": "
Synthetic ISP noise [5]
MethodPSNRSSIMLPIPS
DnCNN [80]29.440.78570.3083
RIDNet [2]28.750.74460.3696
RNAN [85]28.470.72430.3601
SwinIR [45]28.390.70790.3346
Restormer [75]19.310.49820.6556
Dropout [40]28.390.78160.2621
baseline28.890.75950.2917
Ours29.440.79200.2368
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.658, + 0.223, + 0.867, + 0.236 + ], + "angle": 0, + "content": "(c) Comparison on synthetic ISP noise." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.241, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Table 3. We train all the models on Gaussian noise, \\( \\sigma = {15} \\) . All the testing noise is out of the training set, therefore the results can show the models' generalization performance on different unseen noise." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.277, + 0.892, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.458, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Figure 11. Performance comparisons on four noise types with different levels on the Kodak24 dataset [26]. All models are trained only on Gaussian noise. Our masked training approach demonstrates good generalization performance across different noise types. We involve multiple types and levels of noise in testing, the results cannot be shown here. More results are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.554 + ], + "angle": 0, + "content": "LPIPS. Furthermore, as evident from Figure 10, DnCNN's results still contain visible noise, while our method effectively removes the noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Evaluation on Monte Carlo rendering noise. Monte Carlo denoising is a vital component of the rendering process since the widespread use in the industry of Monte Carlo rendering algorithms [8, 17, 42]. We use the test dataset proposed by [25] for Monte Carlo rendered image denoising. The test images were rendered in 128 samples-per-pixel (spp) and 64 spp. The lower the spp, the more severe the noise of the image. In order to adapt the test set to our model, we first convert the data set to sRGB color space by tone mapping. Figure 9 and Table 3b show the denoising results. Our method outperforms all methods on both 128spp and 64spp settings. In Figure 9, the existing methods fail completely because of poor generalization. Our model is still able to remove this noise, demonstrating the wide applicability of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.296, + 0.817 + ], + "angle": 0, + "content": "4.2. Generalization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Training curve. Figure 13 shows the training curves of the model with and without the proposed masked training. The models are trained using only Gaussian noise. The baseline method has a significant overfitting problem. The performance of our method gradually improves with train-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.508, + 0.655, + 0.525 + ], + "angle": 0, + "content": "ing without overfitting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.764 + ], + "angle": 0, + "content": "CKA analysis. To investigate how masked training differs from normal training strategy, we utilize the centered kernel alignment (CKA) [18, 61] to analyze the differences between network representations obtained from those two training methods. Due to the limited space, we describe the detail of CKA in supplementary. In Figure 12, we present our key findings. Specifically, Figure 12 (a) shows the cross-model comparison between the baseline model and our masked training model. We observe a significant difference between the two models in terms of their feature correlations in the deeper layers. Specifically, the features of the deeper layers of the baseline model exhibit low correlations with all layers of our model. This finding suggests that these two training methods exhibit inconsistent learning patterns for features, especially for the deeper layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "To explore how the models perform on different noise types, Figure 12 (b) shows the cross-noise comparison between in-distribution noise and out-of-distribution noise, such as Gaussian and Poisson noise. For the baseline model, we observe a low correlation between different noise types in the deep layers, indicating that the network processes these two types of noise in different ways for the deep layers. This trend holds for other types of noise as well. This phenomenon may be due to the baseline approach causing" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1698" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.096, + 0.255, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.211, + 0.366, + 0.222 + ], + "angle": 0, + "content": "a. cross-model comparison on Gaussian 15" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.096, + 0.408, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.209, + 0.366, + 0.222 + ], + "angle": 0, + "content": "comparison on Gaussian 15" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.09, + 0.731, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.212, + 0.813, + 0.222 + ], + "angle": 0, + "content": "b. cross-model comparison between different noise" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.09, + 0.89, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.228, + 0.735, + 0.243 + ], + "angle": 0, + "content": "Figure 12. CKA similarity to analyze the representation similarity of network layers." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.251, + 0.226, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.332, + 0.225, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.403, + 0.225, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.251, + 0.376, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.332, + 0.376, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.41, + 0.376, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.259, + 0.466, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.334, + 0.466, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.412, + 0.465, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.503, + 0.465, + 0.518 + ], + "angle": 0, + "content": "Figure 13. The testing curves on different noise types and levels." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.533, + 0.273, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.534, + 0.468, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.646, + 0.469, + 0.674 + ], + "angle": 0, + "content": "Figure 14. Comparing generalization ability with the SRGA metric. A lower SRGA value indicates better generalization ability." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.678, + 0.277, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.678, + 0.457, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.792, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Figure 15. The distribution of baseline model features is biased across different noise types. Our method produces similar feature distributions across different noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "the deep layers of the model to overfit to the patterns of the training set, thereby limiting their generalization capabilities to handle different noise types. In contrast, the high correlation between adjacent layers in our masked training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.331 + ], + "angle": 0, + "content": "model suggests that the model's representation of two different noise types is similar. Our masked training forces the network to learn the underlying distribution of the images themselves, which makes the model more robust to different noise types and enhances its generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.893, + 0.606 + ], + "angle": 0, + "content": "Quantification of generalization performance. Liu et al. [49, 50] suggest that model generalization ability can be measured by measuring the consistency of the model's representations across different types of noise. They also propose a generalization assessment index for low-level vision networks called SRGA [50]. It is a non-parametric and nonlearning metric which exploits the statistical characteristics of internal features of deep networks. The lower the value of SRGA, the better the generalization ability. In our case, we use Gaussian noise as the reference and other types of noise for testing. Figure 14 shows the SRGA results. Inspired by [50], we visualize the distributions of deep features on different noise types, shown in Figure 15. We can see that for the baseline model, the feature distributions under different noise types deviate from each other significantly. For the model \\( w \\) / masked training, the deep feature distributions of different noise types are close to each other. This confirms the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.757, + 0.636 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.752 + ], + "angle": 0, + "content": "This work presents a masked training method to improve the generalization performance of deep learning-based image denoising models. The limitation of our method is that the mask operation inevitably loses information. How to preserve more details needs to be explored in future work. Our approach is a step towards developing more robust models for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.893, + 0.891 + ], + "angle": 0, + "content": "Acknowledgment. This work is supported in part by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), the National Natural Science Foundation of China under Grant (62276251), NSFC General Project 62072452, Guangdong Provincial Basic and Applied Basic Research Fund-Regional Joint Fund (Project reference number 2020B1515130004), the Joint Lab of CAS-HK, and the Youth Innovation Promotion Association of Chinese Academy of Sciences (No. 2020356)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1699" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1692-1700, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3155-3164, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[5] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11036-11045, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.472 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.53 + ], + "angle": 0, + "content": "[7] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 2, pages 60-65. IEEE, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.471, + 0.6 + ], + "angle": 0, + "content": "[8] Brent Burley, David Adler, Matt Jen-Yuan Chiang, Hank Driskill, Ralf Habel, Patrick Kelly, Peter Kutz, Yining Karl Li, and Daniel Teece. The design and evolution of disney's hyperion renderer. ACM Transactions on Graphics (TOG), 37(3):1-22, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Chang Chen, Zhiwei Xiong, Xinmei Tian, and Feng Wu. Deep boosting for image denoising. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[10] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[11] Jingwen Chen, Jiawei Chen, Hongyang Chao, and Ming Yang. Image blind denoising with generative adversarial network based noise modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3155-3164, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE transactions on pattern analysis and machine intelligence, 39(6):1256-1272, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, and Xiaokang Yang. Recursive generalization transformer for image super-resolution. arXiv preprint arXiv:2303.06373, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. In NIPS, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Shen Cheng, Yuzhi Wang, Haibin Huang, Donghao Liu, Haoqiang Fan, and Shuaicheng Liu. Nbnet: Noise basis learning for image denoising with subspace projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4896-4906, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[17] Per Christensen, Julian Fong, Jonathan Shade, Wayne Wooten, Brenden Schubert, Andrew Kensler, Stephen Friedman, Charlie Kilpatrick, Cliff Ramshaw, Marc Bannister, et al. Renderman: An advanced path-tracing architecture for movie rendering. ACM Transactions on Graphics (TOG), 37(3):1-21, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[18] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. The Journal of Machine Learning Research, 13:795-828, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[19] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[21] Nithish Divakar and R Venkatesh Babu. Image denoising via cnns: An adversarial approach. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 80-87, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[23] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image processing, 15(12):3736-3745, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[24] Ruicheng Feng, Jinjin Gu, Yu Qiao, and Chao Dong. Suppressing model overfitting for image super-resolution networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[25] Arthur Firmino, Jeppe Revall Frisvad, and Henrik Wann Jensen. Progressive denoising of monte carlo rendered images. In Computer Graphics Forum, volume 41, pages 1-11. Wiley Online Library, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] Rich Franzen. Kodak lossless true color image suite. source: http://r0k.us/graphics/kodak/, 1999. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for percep" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1700" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "tual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[28] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[29] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9199-9208, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[30] Jinjin Gu, Hannan Lu, Wangmeng Zuo, and Chao Dong. Blind super-resolution with iterative kernel correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1604-1613, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[31] Shuhang Gu, Yawei Li, Luc Van Gool, and Radu Timofte. Self-guided network for fast image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2511-2520, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[32] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[33] Shi Guo, Zifeei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1722, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[34] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[35] Xiaowan Hu, Ruijun Ma, Zhihong Liu, Yuanhao Cai, Xiaole Zhao, Yulun Zhang, and Haoqian Wang. Pseudo 3d autocorrelation network for real image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16175-16184, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[36] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[37] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6054-6063, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[38] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[39] Filippos Kokkinos and Stamatios Lefkimmiatis. Deep image demosaicking using a cascade of convolutional residual denoising networks. In Proceedings of the European conference on computer vision (ECCV), pages 303-319, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Xiangtao Kong, Xina Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Reflash dropout in image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6002-6012, 2022. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[41] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Christopher Kulla, Alejandro Conty, Clifford Stein, and Larry Gritz. Sony pictures imageworks arnold. ACM Transactions on Graphics (TOG), 37(3):1-18, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[43] Stamatios Lefkimmiatis. Non-local color image denoising with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3587-3596, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[44] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3204-3213, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[45] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In CVPR, 2021. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[46] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[47] Anran Liu, Yihao Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Blind image super-resolution: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[48] Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[49] Yihao Liu, Anran Liu, Jinjin Gu, Zhipeng Zhang, Henhao Wu, Yu Qiao, and Chao Dong. “Discovering” semantics” in super-resolution networks. arXiv preprint arXiv:2108.00406, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[50] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. arXiv preprint arXiv:2205.07019, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[51] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[52] Salma Abdel Magid, Zudi Lin, Donglai Wei, Yulun Zhang, Jinjin Gu, and Hanspeter Pfister. Texture-based error analysis for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2118-2127, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[53] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In 2009 IEEE 12th international conference on computer vision, pages 2272-2279. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[54] Xiaojiao Mao, Chunhua Shen, and Yu-Bin Yang. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[55] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1701" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[56] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[57] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1586-1595, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[58] Tobias Plötz and Stefan Roth. Neural nearest neighbors networks. Advances in Neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[59] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[60] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[61] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[62] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. arXiv preprint arXiv:2207.08494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[63] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[64] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, Kyoung Mu Lee, et al. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[65] Mathias Uhlen, Per Oksvold, Linn Fagerberg, Emma Lundberg, Kalle Jonasson, Mattias Forsberg, Martin Zwahlen, Caroline Kampf, Kenneth Wester, Sophia Hober, et al. Towards a knowledge-based human protein atlas. Nature biotechnology, 28(12):1248-1250, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[66] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[68] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758-2767, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[69] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.204 + ], + "angle": 0, + "content": "[70] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.893, + 0.289 + ], + "angle": 0, + "content": "[71] Yuan Yuan, Siyuan Liu, Jiawei Zhang, Yongbing Zhang, Chao Dong, and Liang Lin. Unsupervised image superresolution using cycle-in-cycle generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 701-710, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.893, + 0.347 + ], + "angle": 0, + "content": "[72] Zongsheng Yue, Hongwei Yong, Qian Zhao, Deyu Meng, and Lei Zhang. Variational denoising network: Toward blind noise modeling and removal. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.893, + 0.404 + ], + "angle": 0, + "content": "[73] Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. In European Conference on Computer Vision, pages 41-58. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[74] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[75] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.893, + 0.616 + ], + "angle": 0, + "content": "[76] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[77] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. arXiv preprint arXiv:2303.06440, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.893, + 0.729 + ], + "angle": 0, + "content": "[78] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. arXiv preprint arXiv:2210.01427, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.893, + 0.787 + ], + "angle": 0, + "content": "[79] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swim-conv-unet and data synthesis. arXiv preprint arXiv:2203.13278, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.893, + 0.844 + ], + "angle": 0, + "content": "[80] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[81] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1702" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[82] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2):023016, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[84] Yulun Zhang, Kunpeng Li, Kai Li, Gan Sun, Yu Kong, and Yun Fu. Accurate and fast image denoising via attention guided scaling. IEEE Transactions on Image Processing, 30:6255-6265, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[85] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[86] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480-2495, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1703" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_origin.pdf b/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e5a5c8f73030b12380acf0e72f0f4cc7a593d989 --- /dev/null +++ b/2023/Masked Image Training for Generalizable Deep Image Denoising/1b515262-1a18-4a38-b492-e80c892d4e84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3abca9409c0f4cdcad04b8f99e36c9b523a02a62d0a69e3986baec729796430a +size 8021364 diff --git a/2023/Masked Image Training for Generalizable Deep Image Denoising/full.md b/2023/Masked Image Training for Generalizable Deep Image Denoising/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ec1fccf0c65a9f611aa78dbcf5bffcbfe7053687 --- /dev/null +++ b/2023/Masked Image Training for Generalizable Deep Image Denoising/full.md @@ -0,0 +1,458 @@ +# Masked Image Training for Generalizable Deep Image Denoising + +Haoyu Chen $^{1*}$ , Jinjin Gu $^{2,3*}$ , Yihao Liu $^{2,4,5}$ , Salma Abdel Magid $^{7}$ , Chao Dong $^{2,4}$ , Qiong Wang $^{6}$ , Hanspeter Pfister $^{7}$ , Lei Zhu $^{1,8\dagger}$ + +$^{1}$ The Hong Kong University of Science and Technology (Guangzhou) $^{2}$ Shanghai AI Lab $^{3}$ The University of Sydney + $^{4}$ ShenZhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences + $^{5}$ University of Chinese Academy of Sciences $^{6}$ Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences + $^{7}$ Harvard University $^{8}$ The Hong Kong University of Science and Technology +Project page: https://github.com/haoyuc/MaskedDenoising + +# Abstract + +When capturing and storing images, devices inevitably introduce noise. Reducing this noise is a critical task called image denoising. Deep learning has become the de facto method for image denoising, especially with the emergence of Transformer-based models that have achieved notable state-of-the-art results on various image tasks. However, deep learning-based methods often suffer from a lack of generalization ability. For example, deep models trained on Gaussian noise may perform poorly when tested on other noise distributions. To address this issue, we present a novel approach to enhance the generalization performance of denoising networks, known as masked training. Our method involves masking random pixels of the input image and reconstructing the missing information during training. We also mask out the features in the self-attention layers to avoid the impact of training-testing inconsistency. Our approach exhibits better generalization ability than other deep learning models and is directly applicable to real-world scenarios. Additionally, our interpretability analysis demonstrates the superiority of our method. + +# 1. Introduction + +Image denoising is a crucial research area that aims to recover clean images from noisy observations. Due to the rapid advancements in deep learning, many promising image denoising networks have been developed. These networks are typically trained using images synthesized from a pre-defined noise distribution and can achieve remarkable performance in removing the corresponding noise. However, a significant challenge in applying these deep models to real-world scenarios is their generalization ability. Since + +![](images/9373b1ea7f1ac7f1ac1b5de7a7335e2ba3aa06c42d3869ed80afea526de310f8.jpg) +Gaussian 15 + +![](images/2ea9d39ecce52330d8179ea961126a6e04da3fdcb31e8cf9445a6e037e27e4de.jpg) +SwinIR + +![](images/227f37d923133f63d960230aee7b8e77dca56c453de3cad0cb49f90c722f890c.jpg) +Mixture noise +a. In-distribution +b. Out-of-distribution + +![](images/bbc519b4c42f9b813edae967851288d4913ba7353480551b7bc81f3ff01d66a6.jpg) +SwinIR +Figure 1. We illustrate the generalization problem of denoising networks. We train a SwinIR model on Gaussian noise with $\sigma = 15$ . When tested on the same noise, SwinIR demonstrates outstanding performance. However, when applied to out-of-distribution noise, e.g., the mixture of various noise. SwinIR suffers from a huge performance drop. The model trained by the proposed masked training method maintains a reasonable denoising effect, despite aldo being trained on Gaussian noise. + +![](images/3e9ee4c5c146a6f7baa388e2200e87b4be8f1467ee23d795d746b21f9602854a.jpg) +Ours + +the real-world noise distribution can differ from that observed during training, these models often struggle to generalize to such scenarios. + +More specifically, most existing denoising works train and evaluate models on images corrupted with Gaussian noise, limiting their performance to a single noise distribution. When these models are applied to remove noise drawn from other distributions, their performance drastically drops. Figure 1 shows an example. The research community has become increasingly aware of this generalization issue of deep models in recent years. As a countermeasure, some methods [80] assume that the noise level of a particular noise type is unknown, while others [5, 68] attempt to improve the performance in real-world scenarios by synthesizing or collecting training data closer to the target noise or directly performing unsupervised training on the target noise [11, 71]. However, none of these methods substantially improve the generalization performance of denoising networks, and they still struggle when the noise distribution is mismatched [1]. The generalization issue of + +deep denoising still poses challenges to making these methods broadly applicable. + +In this work, we focus on improving the generalization ability of deep denoising models. We define generalization ability as the model's performance on noise different from what it observed during training. We argue that the generalization issue of deep denoising is due to the overfitting of training noise. The existing training strategy directly optimizes the similarity between the denoised image and the ground truth. The intention behind this is that the network should learn to reconstruct the texture and semantics of natural images correctly. However, what is often overlooked is that the network can also reduce the loss simply by overfitting the noise pattern, which is easier than learning the image content. This is at the heart of the generalization problem. Even many popular deep learning methods exacerbate this overfitting problem. When it comes to noise different from that observed during training, the network exhibits this same behavior, resulting in poor performance. + +In light of the preceding discussion, our study seeks to improve the generalization performance of deep denoising networks by directing them to learn image content reconstruction instead of overfitting to training noise. Drawing inspiration from recent masked modeling methods [4, 20, 34, 69], we employ a masked training strategy to explicitly learn representations for image content reconstruction, as opposed to training noise. Leveraging the properties of image processing Transformers [15, 45, 78], we introduce two masking mechanisms: the input mask and the attention mask. During training, the input mask removes input image pixels randomly, and the network reconstructs the removed pixels. The attention mask is implemented in each self-attention layer of the Transformer, enabling it to learn the completion of masked features dynamically and mitigate the distribution shift between training and testing in masked learning. Although we use Gaussian noise for training - similar to previous works - our method demonstrates significant performance improvements on various noise types, such as speckle noise, Poisson noise, salt and pepper noise, spatially correlated Gaussian noise, Monte Carlo-rendered image noise, ISP noise, and complex mixtures of multiple noise sources. Existing methods and models have yet to effectively and accurately remove all these diverse noise patterns. + +# 2. Related Works + +Image Denoising approaches very broadly lie in two categories: traditional model-based and data-driven deep-learning-based. Traditional methods are usually based on modeling image priors to recover image content contaminated by noise [7, 19, 23, 32, 53]. These methods usually do not impose too many constraints on the type of noise, and have been proven to be applicable to a variety of noise, + +with good generalization performance [1]. However, these methods are not satisfactory for the reconstruction of image content. In recent years, the paradigm of denoising has gradually shifted to data-driven methods based on deep learning methods [13]. Many techniques have been proposed to improve the capabilities of the denoising networks continuously, e.g., residual networks [39, 80, 81], dense networks [37, 86], recursive networks [9, 48, 63], multi-scale [21, 31, 76], encoder-decoder [16, 54, 73], attention operations [84, 85], self-similarity [35], and non-local operations [43, 44, 58]. Since 2020, the paradigm of vision network design has gradually shifted from CNNs to Transformers [22]. Vision Transformers treat input pixels as tokens and use self-attention operations to process interactions between these tokens. Inspired by the success of vision Transformers, many attempts have been made to employ Transformers for low-level vision tasks [10, 14, 15, 45, 62, 67, 70, 74, 77, 78] During the development of these models, the noise pattern used for training is often consistent with the testing one. The factor that determines its denoising performance is the fitting ability of the network, in other words, the ability of the network to overfit to the training noise. However, a better network does not mean a better generalization ability of the denoising model. As we will show in the experiment section, a more efficient network even indicates worse generalization performance. + +Generalization Problem in low-level vision often arises when the testing degradation does not match the training degradation, e.g., different downsampling kernel in superresolution [30, 40, 47]. We typically develop deep denoising models based on Gaussian noise in the laboratory setting. However, noise in the real-world is mostly non-Gaussian. Models trained on Gaussian noise fail in these non-Gaussian scenarios. There are two main categories of solutions to this problem. The first is to make training datasets with noise modeling as close to reality as possible during development, e.g., synthesizing real noise according to physical system modeling [5, 68], learning to generate real noise [11, 24, 71], collecting real noise - clean image pairs for training [1, 33, 41, 57]. Although the models obtained by these methods can improve the effect on the target noise, they still cannot generalize to out-of-distribution noise. Another category of solutions is to develop "blind" denoising models, which are supposed to deal with unknown noise [41, 72, 80]. These methods usually simply assume that the noise level is unknown, or train on a large amount of noise types [79], which also fails to generalize to other noise not present in the training set. Few workd have been proposed to study the reasons for the lack of generalization ability in low-level vision [40]. Liu et al. [49] argue that networks tend to overfit to degradations and show degradation "semantics" inside the network. The presence of these representations often means a decrease in general- + +![](images/119cbf8907f4a27a639fefac363dafdb34aac61e4e18f48fafccc7b877012ad4.jpg) +Training image + +![](images/c6249b64558173b20eb0d0fb8b5897f49b7631dcadcc69f81dbb2c7c7d745870.jpg) +Ground truth + +![](images/0443ac2881d07b01bd05fa1f007c43f5739b7aa8f0e63771833f60fcfdfc9df7.jpg) +Denoised by SwinIR + +![](images/223a24012884b990720efd14fe331f7e34fc443579bd20fc4685ae36faebc9ae.jpg) +Denoised by our method +Figure 2. SwinIR, when trained solely on immunohistochemistry images with Gaussian noise, can still denoise natural images. This observation supports the assertion that most existing methods perform denoising primarily through overfitting the training noise. In contrast, our approach emphasizes reconstructing natural image textures and edges observed in the training set on natural images, rather than relying on noise overfitting for denoising. This distinction underlines the fundamental difference between our method and previous approaches. "Our reconstruction result" refers to using our model but taking masked images as input. + +![](images/c6fb6ce865e0cd411b08d131e9f683530059f9985c7881988374e6ab7dd0540a.jpg) +Our reconstruction result + +ization ability. The utilization of this knowledge can guide us to analyze and evaluate the generalization performance [50]. Apart from that, few works have been proposed to improve the generalization ability of denoising models. + +Masked modeling for language [6, 20, 59, 60] is successful for learning pre-trained representations that generalize well to various downstream tasks. These methods mask out a portion of the input sequence and train models to predict the missing content. A similar approach can also be applied to the vision model pre-training. Masked image models learn representations from corrupted images. The earliest attempts in this regard can be traced back at least to the denoising auto-encoder [66]. Since then, many works have used predicting missing parts of images to learn efficient image representations [4, 12, 34, 56, 69]. However, there have been few successful attempts to apply masked image modeling to low-level vision, even though the masked pretraining method is in the form of low-level vision tasks. + +# 3. Method + +Our objective is to create denoising models capable of generalizing to noise not encountered in the training set. In this section, we first discuss our motivation before delving into the specifics of our masked training method. + +Motivation. When training a deep network on a large number of images, the expectation is for the network to learn to discern the rich semantics of natural images from noise-contaminated test cases. However, several studies have noted that the semantics and knowledge acquired by low-level vision networks differ significantly from our expectations [29, 49, 50, 52]. We argue that the poor generalization ability of denoising models results from our training method, which leads the model to focus on overfit + +![](images/ffa6648d79692cf7f4f6022cdec0f9b44a4dcdddd8f25b5afec223ac9b4c4e62.jpg) + +![](images/da218d2b5d65f12e702bf9fa9087d80f4c0f54fbae698624ecd46aba03169d36.jpg) + +![](images/994bc65861c5a73bf79b972e3af048355fa19873f59bd3845f74fbe8e9c30aab.jpg) +original +Figure 3. The illustration of the proposed mask-and-complete training strategy. Even if a large number of pixels are masked, the model can still reconstruct the input to some extent. + +![](images/cc82582f51af3712b10d74d3463ced7992f47c55b3baa5752ca81156a583ba22.jpg) +mask $70\%$ + +![](images/39c20430259683b0bfb1f88ed520c47cda6fea1fe3c145168f516ef2fb6e238d.jpg) + +![](images/4e754983b462d15d191d5106df24018fbb2bc7cd5c0c30c823d405eb77be583f.jpg) +mask $90\%$ + +ting the training noise rather than learning image reconstruction. We conduct a simple experiment for verification. We trained a SwinIR denoising network [45] using images that greatly differ from natural images (immunohistochemistry images [65]). We synthesized training data pairs using Gaussian noise, and then assessed the model's performance on natural images with Gaussian noise. According to our hypothesis, if the model learns the content and reconstruction of image semantics from the training set, it should not perform well on natural images, as it has not been exposed to any. If the model is simply overfitting the noise, the model can remove the noise even if the images are different, as the model mainly relies on detecting the noise for denoising. + +The results are presented in Figure 2. As observed, the SwinIR trained on immunohistochemistry images can still denoise and reproduce the natural image. This supports our conjecture regarding generalization ability, indicating that most existing methods perform denoising by overfitting the training noise. Consequently, when the noise deviates from the training conditions, the denoising performance of these models declines significantly. + +This observation also inspires our approach to developing deep denoising models with improved generalization ability. We aim for the model to learn the reconstruction of image textures and structures, rather than focusing only on noise. In this paper, we propose a new masked training strategy for denoising networks. During training, we mask out a portion of the input pixels and then train the deep network to complete them, as shown in Figure 3. Our approach emphasizes reconstructing natural image textures and edges observed in the image, rather than overfitting noise. In Figure 2 we also show the results of our method. It is evident that our approach seeks to reconstruct the immunohistochemistry image texture from the training set on the testing natural image, instead of relying on noise overfitting for denoising. This demonstrates the potential of this idea in improving generalization performance. By training our method on natural images, it will concentrate on reconstructing the content of natural images, aligning with our + +![](images/b808158284f112b13b4111bafd433e2c8159155b2d4497a81d32a9878f08d546.jpg) +Figure 4. The transformer architecture of our proposed masked image training. We make a minimal change to the original SwinIR architecture – the input mask operation and the attention masks. Other micro-designs are not essentially different from other Transformers. + +![](images/34182b779ea0773cb66a339789d0f4b7a2cd36037c31611530058b0e0aa07bfe.jpg) +Figure 5. Quantitative effect of the attention mask. The histogram differences are also shown above. + +core concept of employing deep learning for low-level vision tasks. + +The Transformer Architecture. Our approach exploits the excellent properties of visual Transformers, so we first describe the basic Transformer backbone used in this study. The shifted window mechanism is proven to be flexible and effective for image processing tasks [15, 45, 78]. We only make minimal changes when applying it to the proposed masked training method without the loss of generality. This model is illustrated in Figure 4. Transformers divide the input signal into tokens and process spatial information using self-attention layers. In our method, a convolution layer with kernel size 1 is used as the feature embedding module to project the 3-channel pixel values into $C$ -dimensional feature tokens. The $1 \times 1$ convolution layer ensures that pixels do not affect each other during feature embedding, which facilitates subsequent masking operations. These feature tokens are gathered with shape $H \times W \times C$ , where $H$ , $W$ and $C$ are the height, width and feature dimension. The shifted window mechanism first reshapes the feature maps of each frame to $\frac{HW}{M^2} \times M^2 \times C$ features by partitioning the input into non-overlapping $M \times M$ local windows, where $\frac{HW}{M^2}$ is the total number of windows. We calculate self-attention on the feature tokens within the same window. Therefore, $M^2$ tokens are involved in each standard self-attention operation, and we produce the local window feature $X \in \mathbb{R}^{M^2 \times C}$ . In each self-attention layer, the query $Q$ , key $K$ and value $V$ are calculated as $Q = XW^Q$ , $K = XW^K$ , $V = XW^V$ , where $W^Q, W^K, W^V \in \mathbb{R}^{C \times D}$ are weight matrices, and $D$ is the dimension of projected vectors. Then, we use $Q$ to query $K$ to generate the attenuated feature vectors. + +![](images/4fd7a18c82c187b7c35d34645a9401784867b438bcbd0b53bdbd44584a74ced8.jpg) +Figure 6. The effectiveness of the input mask and attention mask. Note that the brightness of the image is wrong w/o attention mask. + +
Input MaskAttention MaskPSNRSSIMMix. noise on CBSD68 [55]
Ratio (%)PSNRSSIM
✓29.170.82276529.570.8657
✓26.960.82027529.760.8678
✓✓29.740.86728528.840.8548
+ +Table 1. The importance of using Table 2. Ablation on the at-different mask operations. tention mask ratio. + +tion map $A = \operatorname{softmax}\left(\frac{QK^T}{\sqrt{D}} + B\right) \in \mathbb{R}^{M^2 \times M^2}$ , where $B$ is the learnable relative positional encoding. This attention map $A$ is then used for the weighted sum of $M^2$ vectors in $V$ . The multi-head settings are aligned with SwinIR [45] and ViT [22]. + +Masked Training. Our masked training mainly consists of two aspects, the input mask and the attention mask. Although both are mask operations, the purpose of these two masks is different. We describe them separately. + +The Input Mask randomly masks out the feature tokens embedded by the first convolution layer, and encourages the network to complete the masked information during training. The input mask explicitly constructs a very challenging inpainting problem, as shown in Figure 3. It can be seen that even if up to $90\%$ of the pixel information is destroyed, the network can still reconstruct the target image to a certain extent. The method is very simple. Given the feature token tensor $\mathbf{f} \in \mathbb{R}^{H \times W \times C}$ , we randomly replace the token with a [mask token] $\in \mathbb{R}^C$ with a probability $p_{\mathrm{IM}}$ , where $p_{\mathrm{IM}}$ is called the input mask ratio. The network is trained under the supervision of the $l_1$ -norm of the reconstructed image and the ground truth. The [mask token] can be learnable and initialized with a 0 vector. But we actually found that the 0 vector itself is already a suitable choice. The existence of the input mask forces the network to learn to recognize and reconstruct the content of the image from very limited information. + +The Attention Mask. We cannot build usable image pro + +![](images/524c77de7b8197cdf6e0befff3128ee7cb3758936020e495ae311e07cdc240de.jpg) +Figure 7. The trade-off of choosing different mask ratios. The performance drop on training noise is not significant until $75\%$ masking ratio. Our performance gain on the noise outside the training set is greater than the performance loss on the training set. + +cessing networks relying solely on the input mask operation. Because during testing, we will input uncorrupted images to retain enough information. At this time, due to the inconsistency between training and testing, the network will tend to increase the brightness of the output image. Such as the example in Figure 5. Since Transformer uses the self-attention operation to process spatial information, we can narrow the gap between training and testing by performing the same mask operation during the self-attention process. The specific mask operation is similar to the input mask, but a different attention mask ratio $p_{\mathrm{AM}}$ and [mask token] are used. When some tokens in the self-attention are masked, the attention operation will adjust to the fact that the information of these tokens is no longer reliable. Self-attention will focus on unmasked tokens in each layer and complete the masked information. This operation is difficult to implement on convolutional networks. Figure 5 shows the effect of the attention mask. As can be seen, the attention mask successfully makes the masked trained network work on the unmasked input image. + +# 4. Experiments + +Training Settings. For synthesizing training data, we sample the clean images from DIV2K [64], Flickr2K [46], BSD500 [3], and WED [51] during training. In our work, all the networks are trained using Gaussian noise with standard deviation $\sigma = 15$ . Each input image is randomly cropped to a spatial resolution of $64 \times 64$ , and the number of the total training iteration is $200\mathrm{K}$ . We adopt the Adam optimizer [38] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ to minimize the $L_{1}$ pixel loss. The initial learning rate is set as $1 \times 10^{-4}$ and reduced by half at the milestone of $100\mathrm{K}$ iterations and $150\mathrm{K}$ iterations. The batch size is set to 64. + +Testing Noise. Since the training process utilizes the Gaussian noise, we evaluate the generalization performance of the models on six other synthetic noise: (1) Speckle noise, a type of noise that occurs during the acquisition of medical images or tomography images. (2) Poisson noise, a type of signal-dependent noise that occurs during the acquisition of digital images.(3) Spatially-correlated noise. This is to synthesize the complex artifact after denoising using a flawed algorithm. It is produced by filtering Gaussian noise with a $3 \times 3$ average kernel. Different standard deviations of + +the Gaussian noise indicate different noise levels. (4) Salt & pepper noise. (5) Image signal processing (ISP) noise. [5] proposes a method to synthesize realistic ISP noise during digital imaging. (6) Mixture noise obtained by mixing the above different types of noise with different levels [79]. The clean images are sampled from the benchmark datasets, including CBSD68 [55], Kodak24 [26], McMaster [82], and Urban100 [36]. We also include two real noise types in this work: the Smartphone Image Denoising Dataset (SIDD) [1] and Monte Carlo (MC) rendered image noise. For evaluation, we follow [27,28] and use the metrics PSNR, SSIM [51], and LPIPS [83] to evaluate the results. Since PSNR and SSIM are questioned in assessing the perceptual quality of images [27,28], we also use the LPIPS as an additional metric. + +# 4.1. Results + +Ablation Study. Table 1 and Figure 6 show the effectiveness of using different mask operations. As we can see, without the input mask, the model will lose its generalization ability, and cannot effectively remove the noise outside the training set. Without the attention mask, due to the training-testing inconsistency, the quantitative performance degrades significantly, and the output image will have the wrong brightness. In addition, even without the attention mask, the generalization ability of the model is not significantly affected, and most of the noise is still effectively removed. The input mask is the crucial factor in improving the model the generalization ability. + +Table 3a shows the impact of the different input mask ratios. We test fixed ratios and random ratios from a uniform distribution. From our experiments, fixed ratios are less stable for training than randomly chosen from a range, and the performance is also worse. The best quantitative performance is achieved with random sampling ratios between $75\% \sim 85\%$ . This is a trade-off between denoising generalization ability and the preservation of image details. As shown in Figure 7, smaller ratios are not enough for the network to learn the distribution of images because more noise patterns are preserved. The larger ratio improves the model generalization, as the model focuses more on reconstruction. But at the same time, some image details may be lost. For attention mask ratio, we show the effects in Table 2. The optimal ratios are around $75\%$ . + +The Generalization Performance. We evaluate our deep denoising method on synthetic noise, where our training noise follows a Gaussian distribution with a single noise level, but we test on multiple types of non-Gaussian noise to assess the model's generalization performance. In Figure 11, we compare our method with other state-of-the-art models based on their PSNR and SSIM scores. The results show that our model outperforms all the other models in terms of generalization performance. Particularly, as + +![](images/911690adc5bd52c632fe96531223a16b82cc3fbb98f44da00ff1b268b5e8c7e7.jpg) + +![](images/ba7f643d21300770df5ca47047bba32ac76ad798c4f7f196f8b80ef503b6e9f6.jpg) + +![](images/1a6005a27047b8f23abbcc6e9bdfec653843f05fd4c3f0804118ecfa282abd56.jpg) + +![](images/3c4b9a663bf3ab8496e5ca7a1fcf16436e4f5e116f001da3610a6efaf5acfb88.jpg) + +![](images/0570e030ec1c76f1d33664aa960ecfd077b2f3af9e68b7ebd12ea0358c81cf8c.jpg) + +![](images/839898d8e0f03601ba079d6d2b986935f2f3bcab26072b3614d7eca537b29e2e.jpg) +CBSD68: img_053 +CBSD68: img_0046 + +![](images/dd4b664a95ee70c130168eb35db5df4ad5a930aa2e08b231753c000edbe31593.jpg) +Spatially correlated Gaussian, $\sigma = 50$ + +![](images/0273b063d46fe4758fe5ce12293c32ca44f553f2465974718760d71fb35fcb1d.jpg) +DnCNN [80] + +![](images/fb8fa6a8e05f737d30a295f1adf893c55a7ed0b4b9eae0d4673b876aba467137.jpg) +RIDNet [2] + +![](images/f57e5576cb301b3a729d010190f1a677ac968b903c6bcc9638c7329783fdce54.jpg) +RNAN [85] + +CBSD68: img_0067 + +![](images/77a48262e2831525ae5adca328406795d828389886d178614aff82f16c768002.jpg) +Restormer [75] + +![](images/de034ae40e8ffbed275120b127d610984c406f64ab3cd1534417f5877377f784.jpg) +SwinIR [45] + +![](images/f6b2b2a3068d5d0f5a2d6f3eb35c4fa9a753835ef077d3589cf3570a639de835.jpg) +baseline + +![](images/9f3256261b64aea66cab5d72c3291484542cfa7f84a2d3942006d75a98f990cc.jpg) +Masked Training + +CBSD68: img_0067 + +![](images/d94ceccba28fffb236c9a57076403168602f47af7e593981afc4dcdcbeca6d0d.jpg) +Speckle noise, $\sigma^2 = 0.013$ + +![](images/30a4277d6c51acc68461321ab2df739e9052363bc89966da788e9db4d86aae3c.jpg) +DnCNN [80] + +![](images/5078b2a66d132996a58c2364a0b12f5a6509e0a5cc5e18ef80f9ba8076902491.jpg) +RIDNet [2] + +![](images/f9a1fc154dd1a6f4079ad25bc013cab9c6cb444785c626a959c745b3aa0c1d23.jpg) +RNAN [85] + +![](images/275a77bf7eb8d5948aa0882874985f1d45bfeca27f00d0f347fe3b15e22c941b.jpg) +Restormer [75] + +![](images/f231ed7b81df15f5433d229b3ef1cf9357accf20c1c37a72b62ed7cd12f2e1f5.jpg) +SwinIR [45] + +![](images/d3cc84996467047107870bed410e4105e4fa7009dc9751b8e14f08e8f5b8d41e.jpg) +baseline + +![](images/aecd15b639520ea0ba5ceca1a1c30f2dae901e744bc71bec120169d84cc8cd2b.jpg) +Masked Training + +CBSD68: img_0067 + +![](images/cd73fd5c98e236ab194eaacbe68e86a819523df5001785f7884a02a8844ae7eb.jpg) +Salt-and-pepper noise, $d = 0.02$ +Restormer [75] + +![](images/703ed3f6c493a10ad68348621588476a65c32bb9688cbcbccfb4d8d6bc09ea0e.jpg) +DnCNN [80] +SwinIR [45] + +![](images/828c5e8f8e3b3a1110e1f3a72b9202b3d3962576fce339911b29ffb6325f6b1c.jpg) +RIDNet [2] +baseline + +![](images/ec12561b78ec63a2fc1c9fa1d6532e9bddc1238f5e842786262a2847d6896d2e.jpg) +RNAN [85] +Masked Training + +![](images/7bae78ae4f6ea8a6835f49474e7c8ab0ebf0eddf330036be55d6b147e9a1bb29.jpg) +Figure 8. Visual comparison on out-of-distribution noise. When all other methods fail completely, our method is still able to denoise effectively. Please refer to the supplementary material to see more visual results. +Rendered image (256 spp) + +![](images/eeaaf3455adea2a9d6c20e797d80f0ae70bed4051b990536a8bd8c0378540b72.jpg) +Noisy + +![](images/2ceb774440941b0b2a2ff63c653ab416ea817c274d24b5c70d3f9e716646a63c.jpg) +Reference + +![](images/7d983ce7d996eea169e7dbf638f5d573581a10ec55a3dcb33d22a60563922a98.jpg) +DnCNN + +![](images/6ca69d2eafa1f35c082f44c3b8a8e156a79bb17ef4014054b21c2f93f8d1aedd.jpg) +Noisy + +![](images/dfaf45f3c4f702d8cb73a92f4b39f63cf3c948f60e6cb75839a66ab787e2113a.jpg) +Reference + +![](images/b5a6acba261e5342dbdb79af8919c0e24fe52720256980c08a728a7825cef5eb.jpg) +DnCNN + +![](images/76899e35721f963c496c81386757bfb0d70a51e04d57560d6d4071adabea1950.jpg) +Figure 9. Visual results of denoising a Monte Carlo rendered image. +RNAN + +![](images/c5676def3d9a50554072b314f2ad9b7a555b5e2767dd601dc52800c951b8ec94.jpg) +SwinIR + +![](images/d84c270097afb20f203ae674985c08cb020fbfad2b4dcf114b24718a5feb4afb.jpg) +Ours + +![](images/a3c9210cc68eb76f09476e6f4f04d844ed8db80c12dda4bcc93fef942831d1d6.jpg) +RNAN +Figure 10. Results of ISP noise removal. + +![](images/4714e91bddfc967421fba23d9ce47892e57d1e8a5d85b77b29b6703ee43baa18.jpg) +SwinIR + +![](images/08d89ec8452a46ba2e6caf57e5eaa8e727aa57adbabab4fa7405427fb270f45b.jpg) +Ours + +the noise level increases, our model exhibits a slower performance degradation and thus demonstrates better generalization. In contrast, other models suffer from significant performance drops when dealing with more severe noise. We also provide visual comparisons in Figure 8, where our model achieves remarkable denoising results even though it is trained only on Gaussian noise with a fixed standard deviation. In contrast, existing models tend to overfit the training noise and fail when facing unseen noise. More quantitative and qualitative results can be found in the supplementary material. + +Evaluation on ISP noise. The removal of the ISP noise is of great application value. Brooks et al. [5] present a systematic approach for generating realistic raw data with ISP noise that can facilitate our research. We use the default parameter settings of the method proposed in [5] to synthesize ISP noise on the Kodak24 [26] dataset for testing. The results are shown in Figure 10 and Table 3c. Our method achieves superior results compared to all other methods. Notably, our method achieves a significant lead in LPIPS, indicating that our results exhibit better perceptual quality. Although DnCNN and our method obtain the same PSNR, our method still outperforms DnCNN in terms of SSIM and + +
Mix. noise on CBSD68 [55]
Ratio (%)PSNRSSIM
7529.170.8132
8529.440.8545
9519.600.7273
70–8029.860.8593
75–8530.040.8756
75–9029.870.8728
75–9529.260.8607
80–9029.740.8672
+ +(a) Abl. of input mask ratios. + +
Method128 samples per pixel64 samples per pixel
PSNRSSIMLPIPSPSNRSSIMLPIPS
DnCNN [80]29.940.78830.267126.280.67790.4216
RIDNet [2]29.960.79210.254826.270.67880.4122
RNAN [85]29.860.78250.270226.260.67430.4290
SwinIR [45]29.320.76270.294326.140.66510.4485
Restormer [75]24.980.65980.457524.590.58800.5375
Dropout [40]28.850.77530.294126.100.66960.4443
baseline29.680.77380.285125.910.65350.4564
Ours30.620.85000.225428.250.76940.3348
+ +(b) Quantitative comparison on Monte Carlo rendered image denoising. + +
Synthetic ISP noise [5]
MethodPSNRSSIMLPIPS
DnCNN [80]29.440.78570.3083
RIDNet [2]28.750.74460.3696
RNAN [85]28.470.72430.3601
SwinIR [45]28.390.70790.3346
Restormer [75]19.310.49820.6556
Dropout [40]28.390.78160.2621
baseline28.890.75950.2917
Ours29.440.79200.2368
+ +(c) Comparison on synthetic ISP noise. + +Table 3. We train all the models on Gaussian noise, $\sigma = {15}$ . All the testing noise is out of the training set, therefore the results can show the models' generalization performance on different unseen noise. + +![](images/56f384c2f38d7cd075f0db2482381134bab882d2519be3d38fc60066c6b6b61e.jpg) +Figure 11. Performance comparisons on four noise types with different levels on the Kodak24 dataset [26]. All models are trained only on Gaussian noise. Our masked training approach demonstrates good generalization performance across different noise types. We involve multiple types and levels of noise in testing, the results cannot be shown here. More results are shown in the supplementary material. + +LPIPS. Furthermore, as evident from Figure 10, DnCNN's results still contain visible noise, while our method effectively removes the noise. + +Evaluation on Monte Carlo rendering noise. Monte Carlo denoising is a vital component of the rendering process since the widespread use in the industry of Monte Carlo rendering algorithms [8, 17, 42]. We use the test dataset proposed by [25] for Monte Carlo rendered image denoising. The test images were rendered in 128 samples-per-pixel (spp) and 64 spp. The lower the spp, the more severe the noise of the image. In order to adapt the test set to our model, we first convert the data set to sRGB color space by tone mapping. Figure 9 and Table 3b show the denoising results. Our method outperforms all methods on both 128spp and 64spp settings. In Figure 9, the existing methods fail completely because of poor generalization. Our model is still able to remove this noise, demonstrating the wide applicability of our method. + +# 4.2. Generalization Analysis + +Training curve. Figure 13 shows the training curves of the model with and without the proposed masked training. The models are trained using only Gaussian noise. The baseline method has a significant overfitting problem. The performance of our method gradually improves with train- + +ing without overfitting. + +CKA analysis. To investigate how masked training differs from normal training strategy, we utilize the centered kernel alignment (CKA) [18, 61] to analyze the differences between network representations obtained from those two training methods. Due to the limited space, we describe the detail of CKA in supplementary. In Figure 12, we present our key findings. Specifically, Figure 12 (a) shows the cross-model comparison between the baseline model and our masked training model. We observe a significant difference between the two models in terms of their feature correlations in the deeper layers. Specifically, the features of the deeper layers of the baseline model exhibit low correlations with all layers of our model. This finding suggests that these two training methods exhibit inconsistent learning patterns for features, especially for the deeper layers. + +To explore how the models perform on different noise types, Figure 12 (b) shows the cross-noise comparison between in-distribution noise and out-of-distribution noise, such as Gaussian and Poisson noise. For the baseline model, we observe a low correlation between different noise types in the deep layers, indicating that the network processes these two types of noise in different ways for the deep layers. This trend holds for other types of noise as well. This phenomenon may be due to the baseline approach causing + +![](images/d3fddd29972407b25eb0ba26388ae16cf4ca0d4987c003dc878fd14a44fc23e3.jpg) +a. cross-model comparison on Gaussian 15 +comparison on Gaussian 15 + +![](images/4929cb376de2f96ded1959e0264c7e399669e051f3e929323fd353344a590813.jpg) + +![](images/cbde860c33764b295d60c7fdb749738c9007f3b0a32055d8f11082fe4e4e5a48.jpg) +b. cross-model comparison between different noise + +![](images/575e359a04a23253b4fceb74e6dd6c94d57d558d1b82b7c46b6694096c9cfe8d.jpg) + +![](images/fd31056147495e4addd7698a002f31719a18a82e0294e3436227511547e51041.jpg) + +![](images/24843181b752ba6cb26f219c182b874367b3f289b82de08d2b1d6dc354d9c340.jpg) + +![](images/cd9cd3d1e91ead4be9d2800c3ddefea49778b3c8844ac77460d30b615c27d707.jpg) +Figure 13. The testing curves on different noise types and levels. + +![](images/d901f586b46a50b95c734cc308f2b2eb0f458ceaa71c55efe7c0b94b38680f10.jpg) +Figure 12. CKA similarity to analyze the representation similarity of network layers. + +![](images/706b6c0ba059f4dffb6c278418bf7d06dc4d9eab04aeeda61e76c75dc24e3a64.jpg) + +![](images/126ee05a5548aa7e50e43737b33d4467de7efb88576397bf905f42be9b99f5fe.jpg) + +![](images/26e6debaec8127ba4ca711dc9b726b75bc00918754ae9418f4a915a0f50d72c0.jpg) + +![](images/087d703e65e83232dee7dc7c9a57f52d68049a61204acab5b94b1cad1dc025ff.jpg) + +![](images/f86d31ff0b155f92efa52cc796205bed8e4fd14f726dd549240d33d90840c1eb.jpg) + +![](images/5782e4e4e87baf48f4bc27b7bf04607a901bcb521c7133923e3b5cb7460b5fad.jpg) + +![](images/6efdd3301d8b9d52835f72e402abe1214857688ea456390470346894ca6560e5.jpg) + +![](images/120fec10a589d031aed46f891b0e7313202a2f177e53ad9a6a969a1f8760c5e2.jpg) +Figure 14. Comparing generalization ability with the SRGA metric. A lower SRGA value indicates better generalization ability. +Figure 15. The distribution of baseline model features is biased across different noise types. Our method produces similar feature distributions across different noise. + +![](images/bd0ea18c0d79baeac8e0ea5749127e47d2a7e71ea49cf52da8ae1fbd73c6242e.jpg) + +the deep layers of the model to overfit to the patterns of the training set, thereby limiting their generalization capabilities to handle different noise types. In contrast, the high correlation between adjacent layers in our masked training + +model suggests that the model's representation of two different noise types is similar. Our masked training forces the network to learn the underlying distribution of the images themselves, which makes the model more robust to different noise types and enhances its generalization capability. + +Quantification of generalization performance. Liu et al. [49, 50] suggest that model generalization ability can be measured by measuring the consistency of the model's representations across different types of noise. They also propose a generalization assessment index for low-level vision networks called SRGA [50]. It is a non-parametric and nonlearning metric which exploits the statistical characteristics of internal features of deep networks. The lower the value of SRGA, the better the generalization ability. In our case, we use Gaussian noise as the reference and other types of noise for testing. Figure 14 shows the SRGA results. Inspired by [50], we visualize the distributions of deep features on different noise types, shown in Figure 15. We can see that for the baseline model, the feature distributions under different noise types deviate from each other significantly. For the model $w$ / masked training, the deep feature distributions of different noise types are close to each other. This confirms the effectiveness of our method. + +# 5. Conclusion and Limitations + +This work presents a masked training method to improve the generalization performance of deep learning-based image denoising models. The limitation of our method is that the mask operation inevitably loses information. How to preserve more details needs to be explored in future work. Our approach is a step towards developing more robust models for real-world applications. + +Acknowledgment. This work is supported in part by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), the National Natural Science Foundation of China under Grant (62276251), NSFC General Project 62072452, Guangdong Provincial Basic and Applied Basic Research Fund-Regional Joint Fund (Project reference number 2020B1515130004), the Joint Lab of CAS-HK, and the Youth Innovation Promotion Association of Chinese Academy of Sciences (No. 2020356). + +# References + +[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1692-1700, 2018. 1, 2, 5 +[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3155-3164, 2019. 6, 7 +[3] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 2010. 5 +[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 3 +[5] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11036-11045, 2019. 1, 2, 5, 6, 7 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[7] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 2, pages 60-65. IEEE, 2005. 2 +[8] Brent Burley, David Adler, Matt Jen-Yuan Chiang, Hank Driskill, Ralf Habel, Patrick Kelly, Peter Kutz, Yining Karl Li, and Daniel Teece. The design and evolution of disney's hyperion renderer. ACM Transactions on Graphics (TOG), 37(3):1-22, 2018. 7 +[9] Chang Chen, Zhiwei Xiong, Xinmei Tian, and Feng Wu. Deep boosting for image denoising. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2 +[10] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2 +[11] Jingwen Chen, Jiawei Chen, Hongyang Chao, and Ming Yang. Image blind denoising with generative adversarial network based noise modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3155-3164, 2018. 1, 2 +[12] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 3 +[13] Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE transactions on pattern analysis and machine intelligence, 39(6):1256-1272, 2016. 2 + +[14] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, and Xiaokang Yang. Recursive generalization transformer for image super-resolution. arXiv preprint arXiv:2303.06373, 2023. 2 +[15] Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. In NIPS, 2022. 2, 4 +[16] Shen Cheng, Yuzhi Wang, Haibin Huang, Donghao Liu, Haoqiang Fan, and Shuaicheng Liu. Nbnet: Noise basis learning for image denoising with subspace projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4896-4906, 2021. 2 +[17] Per Christensen, Julian Fong, Jonathan Shade, Wayne Wooten, Brenden Schubert, Andrew Kensler, Stephen Friedman, Charlie Kilpatrick, Cliff Ramshaw, Marc Bannister, et al. Renderman: An advanced path-tracing architecture for movie rendering. ACM Transactions on Graphics (TOG), 37(3):1-21, 2018. 7 +[18] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. The Journal of Machine Learning Research, 13:795-828, 2012. 7 +[19] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2 +[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3 +[21] Nithish Divakar and R Venkatesh Babu. Image denoising via cnns: An adversarial approach. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 80-87, 2017. 2 +[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 4 +[23] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image processing, 15(12):3736-3745, 2006. 2 +[24] Ruicheng Feng, Jinjin Gu, Yu Qiao, and Chao Dong. Suppressing model overfitting for image super-resolution networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 2 +[25] Arthur Firmino, Jeppe Revall Frisvad, and Henrik Wann Jensen. Progressive denoising of monte carlo rendered images. In Computer Graphics Forum, volume 41, pages 1-11. Wiley Online Library, 2022. 7 +[26] Rich Franzen. Kodak lossless true color image suite. source: http://r0k.us/graphics/kodak/, 1999. 5, 6, 7 +[27] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for percep + +tual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 5 +[28] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 5 +[29] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9199-9208, 2021. 3 +[30] Jinjin Gu, Hannan Lu, Wangmeng Zuo, and Chao Dong. Blind super-resolution with iterative kernel correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1604-1613, 2019. 2 +[31] Shuhang Gu, Yawei Li, Luc Van Gool, and Radu Timofte. Self-guided network for fast image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2511-2520, 2019. 2 +[32] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2 +[33] Shi Guo, Zifeei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1722, 2019. 2 +[34] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3 +[35] Xiaowan Hu, Ruijun Ma, Zhihong Liu, Yuanhao Cai, Xiaole Zhao, Yulun Zhang, and Haoqian Wang. Pseudo 3d autocorrelation network for real image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16175-16184, 2021. 2 +[36] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5 +[37] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6054-6063, 2019. 2 +[38] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[39] Filippos Kokkinos and Stamatios Lefkimmiatis. Deep image demosaicking using a cascade of convolutional residual denoising networks. In Proceedings of the European conference on computer vision (ECCV), pages 303-319, 2018. 2 +[40] Xiangtao Kong, Xina Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Reflash dropout in image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6002-6012, 2022. 2, 7 + +[41] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 2 +[42] Christopher Kulla, Alejandro Conty, Clifford Stein, and Larry Gritz. Sony pictures imageworks arnold. ACM Transactions on Graphics (TOG), 37(3):1-18, 2018. 7 +[43] Stamatios Lefkimmiatis. Non-local color image denoising with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3587-3596, 2017. 2 +[44] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3204-3213, 2018. 2 +[45] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In CVPR, 2021. 2, 3, 4, 6, 7 +[46] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 5 +[47] Anran Liu, Yihao Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Blind image super-resolution: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[48] Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. Advances in neural information processing systems, 31, 2018. 2 +[49] Yihao Liu, Anran Liu, Jinjin Gu, Zhipeng Zhang, Henhao Wu, Yu Qiao, and Chao Dong. “Discovering” semantics” in super-resolution networks. arXiv preprint arXiv:2108.00406, 2021. 2, 3, 8 +[50] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. arXiv preprint arXiv:2205.07019, 2022. 3, 8 +[51] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 2016. 5 +[52] Salma Abdel Magid, Zudi Lin, Donglai Wei, Yulun Zhang, Jinjin Gu, and Hanspeter Pfister. Texture-based error analysis for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2118-2127, 2022. 3 +[53] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In 2009 IEEE 12th international conference on computer vision, pages 2272-2279. IEEE, 2009. 2 +[54] Xiaojiao Mao, Chunhua Shen, and Yu-Bin Yang. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections. Advances in neural information processing systems, 29, 2016. 2 +[55] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and + +measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 4, 5, 7 +[56] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3 +[57] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1586-1595, 2017. 2 +[58] Tobias Plötz and Stefan Roth. Neural nearest neighbors networks. Advances in Neural information processing systems, 31, 2018. 2 +[59] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3 +[60] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 +[61] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 7 +[62] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. arXiv preprint arXiv:2207.08494, 2022. 2 +[63] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 2 +[64] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, Kyoung Mu Lee, et al. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, 2017. 5 +[65] Mathias Uhlen, Per Oksvold, Linn Fagerberg, Emma Lundberg, Kalle Jonasson, Mattias Forsberg, Martin Zwahlen, Caroline Kampf, Kenneth Wester, Sophia Hober, et al. Towards a knowledge-based human protein atlas. Nature biotechnology, 28(12):1248-1250, 2010. 3 +[66] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3 +[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 2 +[68] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758-2767, 2020. 1, 2 + +[69] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2, 3 +[70] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In CVPR, 2020. 2 +[71] Yuan Yuan, Siyuan Liu, Jiawei Zhang, Yongbing Zhang, Chao Dong, and Liang Lin. Unsupervised image superresolution using cycle-in-cycle generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 701-710, 2018. 1, 2 +[72] Zongsheng Yue, Hongwei Yong, Qian Zhao, Deyu Meng, and Lei Zhang. Variational denoising network: Toward blind noise modeling and removal. Advances in neural information processing systems, 32, 2019. 2 +[73] Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. In European Conference on Computer Vision, pages 41-58. Springer, 2020. 2 +[74] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2 +[75] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 6, 7 +[76] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2 +[77] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. arXiv preprint arXiv:2303.06440, 2023. 2 +[78] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. arXiv preprint arXiv:2210.01427, 2022. 2, 4 +[79] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swim-conv-unet and data synthesis. arXiv preprint arXiv:2203.13278, 2022. 2, 5 +[80] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 2, 6, 7 +[81] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 2 + +[82] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2):023016, 2011. 5 +[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[84] Yulun Zhang, Kunpeng Li, Kai Li, Gan Sun, Yu Kong, and Yun Fu. Accurate and fast image denoising via attention guided scaling. IEEE Transactions on Image Processing, 30:6255-6265, 2021. 2 +[85] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2, 6, 7 +[86] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480-2495, 2020. 2 \ No newline at end of file diff --git a/2023/Masked Image Training for Generalizable Deep Image Denoising/images.zip b/2023/Masked Image Training for Generalizable Deep Image Denoising/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3adb616ac7070232664215deccecc1e86f088628 --- /dev/null +++ b/2023/Masked Image Training for Generalizable Deep Image Denoising/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6559c74c1112187e2970831c2367481fa1301fcfcbc23a768ef40ca289aa635f +size 939355 diff --git a/2023/Masked Image Training for Generalizable Deep Image Denoising/layout.json b/2023/Masked Image Training for Generalizable Deep Image Denoising/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..33357e8d44308f878dee6faaa34d4eb7ec19e25a --- /dev/null +++ b/2023/Masked Image Training for Generalizable Deep Image Denoising/layout.json @@ -0,0 +1,13178 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "text", + "content": "Masked Image Training for Generalizable Deep Image Denoising" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "spans": [ + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "content": "Haoyu Chen" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "content": ", Jinjin Gu" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "content": ", Yihao Liu" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{2,4,5}" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "content": ", Salma Abdel Magid" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "content": ", Chao Dong" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "content": ", Qiong Wang" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "content": ", Hanspeter Pfister" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "text", + "content": ", Lei Zhu" + }, + { + "bbox": [ + 123, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1,8\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "spans": [ + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology (Guangzhou) " + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "content": "Shanghai AI Lab " + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "content": "The University of Sydney \n" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "content": "ShenZhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "content": "Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "content": "Harvard University " + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 68, + 176, + 527, + 254 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology \nProject page: https://github.com/haoyuc/MaskedDenoising" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 281, + 191, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 281, + 191, + 294 + ], + "spans": [ + { + "bbox": [ + 143, + 281, + 191, + 294 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 307, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 290, + 547 + ], + "type": "text", + "content": "When capturing and storing images, devices inevitably introduce noise. Reducing this noise is a critical task called image denoising. Deep learning has become the de facto method for image denoising, especially with the emergence of Transformer-based models that have achieved notable state-of-the-art results on various image tasks. However, deep learning-based methods often suffer from a lack of generalization ability. For example, deep models trained on Gaussian noise may perform poorly when tested on other noise distributions. To address this issue, we present a novel approach to enhance the generalization performance of denoising networks, known as masked training. Our method involves masking random pixels of the input image and reconstructing the missing information during training. We also mask out the features in the self-attention layers to avoid the impact of training-testing inconsistency. Our approach exhibits better generalization ability than other deep learning models and is directly applicable to real-world scenarios. Additionally, our interpretability analysis demonstrates the superiority of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 559, + 128, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 128, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 128, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 577, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 684 + ], + "type": "text", + "content": "Image denoising is a crucial research area that aims to recover clean images from noisy observations. Due to the rapid advancements in deep learning, many promising image denoising networks have been developed. These networks are typically trained using images synthesized from a pre-defined noise distribution and can achieve remarkable performance in removing the corresponding noise. However, a significant challenge in applying these deep models to real-world scenarios is their generalization ability. Since" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 281, + 354, + 345 + ], + "blocks": [ + { + "bbox": [ + 307, + 281, + 354, + 345 + ], + "lines": [ + { + "bbox": [ + 307, + 281, + 354, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 354, + 345 + ], + "type": "image", + "image_path": "9373b1ea7f1ac7f1ac1b5de7a7335e2ba3aa06c42d3869ed80afea526de310f8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 346, + 349, + 353 + ], + "lines": [ + { + "bbox": [ + 310, + 346, + 349, + 353 + ], + "spans": [ + { + "bbox": [ + 310, + 346, + 349, + 353 + ], + "type": "text", + "content": "Gaussian 15" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 354, + 281, + 402, + 345 + ], + "blocks": [ + { + "bbox": [ + 354, + 281, + 402, + 345 + ], + "lines": [ + { + "bbox": [ + 354, + 281, + 402, + 345 + ], + "spans": [ + { + "bbox": [ + 354, + 281, + 402, + 345 + ], + "type": "image", + "image_path": "2ea9d39ecce52330d8179ea961126a6e04da3fdcb31e8cf9445a6e037e27e4de.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 346, + 389, + 353 + ], + "lines": [ + { + "bbox": [ + 367, + 346, + 389, + 353 + ], + "spans": [ + { + "bbox": [ + 367, + 346, + 389, + 353 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 402, + 281, + 450, + 345 + ], + "blocks": [ + { + "bbox": [ + 402, + 281, + 450, + 345 + ], + "lines": [ + { + "bbox": [ + 402, + 281, + 450, + 345 + ], + "spans": [ + { + "bbox": [ + 402, + 281, + 450, + 345 + ], + "type": "image", + "image_path": "227f37d923133f63d960230aee7b8e77dca56c453de3cad0cb49f90c722f890c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 346, + 447, + 353 + ], + "lines": [ + { + "bbox": [ + 406, + 346, + 447, + 353 + ], + "spans": [ + { + "bbox": [ + 406, + 346, + 447, + 353 + ], + "type": "text", + "content": "Mixture noise" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 355, + 383, + 363 + ], + "lines": [ + { + "bbox": [ + 326, + 355, + 383, + 363 + ], + "spans": [ + { + "bbox": [ + 326, + 355, + 383, + 363 + ], + "type": "text", + "content": "a. In-distribution" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 355, + 510, + 363 + ], + "lines": [ + { + "bbox": [ + 437, + 355, + 510, + 363 + ], + "spans": [ + { + "bbox": [ + 437, + 355, + 510, + 363 + ], + "type": "text", + "content": "b. Out-of-distribution" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 450, + 281, + 500, + 345 + ], + "blocks": [ + { + "bbox": [ + 450, + 281, + 500, + 345 + ], + "lines": [ + { + "bbox": [ + 450, + 281, + 500, + 345 + ], + "spans": [ + { + "bbox": [ + 450, + 281, + 500, + 345 + ], + "type": "image", + "image_path": "bbc519b4c42f9b813edae967851288d4913ba7353480551b7bc81f3ff01d66a6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 346, + 486, + 353 + ], + "lines": [ + { + "bbox": [ + 465, + 346, + 486, + 353 + ], + "spans": [ + { + "bbox": [ + 465, + 346, + 486, + 353 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 368, + 545, + 456 + ], + "lines": [ + { + "bbox": [ + 305, + 368, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 545, + 456 + ], + "type": "text", + "content": "Figure 1. We illustrate the generalization problem of denoising networks. We train a SwinIR model on Gaussian noise with " + }, + { + "bbox": [ + 305, + 368, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\sigma = 15" + }, + { + "bbox": [ + 305, + 368, + 545, + 456 + ], + "type": "text", + "content": ". When tested on the same noise, SwinIR demonstrates outstanding performance. However, when applied to out-of-distribution noise, e.g., the mixture of various noise. SwinIR suffers from a huge performance drop. The model trained by the proposed masked training method maintains a reasonable denoising effect, despite aldo being trained on Gaussian noise." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 500, + 281, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 500, + 281, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 500, + 281, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 500, + 281, + 545, + 345 + ], + "type": "image", + "image_path": "3e9ee4c5c146a6f7baa388e2200e87b4be8f1467ee23d795d746b21f9602854a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 346, + 529, + 353 + ], + "lines": [ + { + "bbox": [ + 514, + 346, + 529, + 353 + ], + "spans": [ + { + "bbox": [ + 514, + 346, + 529, + 353 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 471, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 506 + ], + "type": "text", + "content": "the real-world noise distribution can differ from that observed during training, these models often struggle to generalize to such scenarios." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "More specifically, most existing denoising works train and evaluate models on images corrupted with Gaussian noise, limiting their performance to a single noise distribution. When these models are applied to remove noise drawn from other distributions, their performance drastically drops. Figure 1 shows an example. The research community has become increasingly aware of this generalization issue of deep models in recent years. As a countermeasure, some methods [80] assume that the noise level of a particular noise type is unknown, while others [5, 68] attempt to improve the performance in real-world scenarios by synthesizing or collecting training data closer to the target noise or directly performing unsupervised training on the target noise [11, 71]. However, none of these methods substantially improve the generalization performance of denoising networks, and they still struggle when the noise distribution is mismatched [1]. The generalization issue of" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 251, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 251, + 702 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 251, + 702 + ], + "type": "text", + "content": "*Haoyu Chen and Jinjin Gu contribute equally to this work." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 59, + 703, + 234, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 234, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 234, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 59, + 703, + 234, + 712 + ], + "type": "text", + "content": " Lei Zhu (leizhu@ust.hk) is the corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1692" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 285, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 285, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 285, + 95 + ], + "type": "text", + "content": "deep denoising still poses challenges to making these methods broadly applicable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 97, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 97, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 49, + 97, + 286, + 299 + ], + "type": "text", + "content": "In this work, we focus on improving the generalization ability of deep denoising models. We define generalization ability as the model's performance on noise different from what it observed during training. We argue that the generalization issue of deep denoising is due to the overfitting of training noise. The existing training strategy directly optimizes the similarity between the denoised image and the ground truth. The intention behind this is that the network should learn to reconstruct the texture and semantics of natural images correctly. However, what is often overlooked is that the network can also reduce the loss simply by overfitting the noise pattern, which is easier than learning the image content. This is at the heart of the generalization problem. Even many popular deep learning methods exacerbate this overfitting problem. When it comes to noise different from that observed during training, the network exhibits this same behavior, resulting in poor performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 300, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 300, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 300, + 286, + 597 + ], + "type": "text", + "content": "In light of the preceding discussion, our study seeks to improve the generalization performance of deep denoising networks by directing them to learn image content reconstruction instead of overfitting to training noise. Drawing inspiration from recent masked modeling methods [4, 20, 34, 69], we employ a masked training strategy to explicitly learn representations for image content reconstruction, as opposed to training noise. Leveraging the properties of image processing Transformers [15, 45, 78], we introduce two masking mechanisms: the input mask and the attention mask. During training, the input mask removes input image pixels randomly, and the network reconstructs the removed pixels. The attention mask is implemented in each self-attention layer of the Transformer, enabling it to learn the completion of masked features dynamically and mitigate the distribution shift between training and testing in masked learning. Although we use Gaussian noise for training - similar to previous works - our method demonstrates significant performance improvements on various noise types, such as speckle noise, Poisson noise, salt and pepper noise, spatially correlated Gaussian noise, Monte Carlo-rendered image noise, ISP noise, and complex mixtures of multiple noise sources. Existing methods and models have yet to effectively and accurately remove all these diverse noise patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 609, + 137, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 609, + 137, + 621 + ], + "spans": [ + { + "bbox": [ + 49, + 609, + 137, + 621 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 630, + 285, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 630, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 630, + 285, + 712 + ], + "type": "text", + "content": "Image Denoising approaches very broadly lie in two categories: traditional model-based and data-driven deep-learning-based. Traditional methods are usually based on modeling image priors to recover image content contaminated by noise [7, 19, 23, 32, 53]. These methods usually do not impose too many constraints on the type of noise, and have been proven to be applicable to a variety of noise," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 382 + ], + "type": "text", + "content": "with good generalization performance [1]. However, these methods are not satisfactory for the reconstruction of image content. In recent years, the paradigm of denoising has gradually shifted to data-driven methods based on deep learning methods [13]. Many techniques have been proposed to improve the capabilities of the denoising networks continuously, e.g., residual networks [39, 80, 81], dense networks [37, 86], recursive networks [9, 48, 63], multi-scale [21, 31, 76], encoder-decoder [16, 54, 73], attention operations [84, 85], self-similarity [35], and non-local operations [43, 44, 58]. Since 2020, the paradigm of vision network design has gradually shifted from CNNs to Transformers [22]. Vision Transformers treat input pixels as tokens and use self-attention operations to process interactions between these tokens. Inspired by the success of vision Transformers, many attempts have been made to employ Transformers for low-level vision tasks [10, 14, 15, 45, 62, 67, 70, 74, 77, 78] During the development of these models, the noise pattern used for training is often consistent with the testing one. The factor that determines its denoising performance is the fitting ability of the network, in other words, the ability of the network to overfit to the training noise. However, a better network does not mean a better generalization ability of the denoising model. As we will show in the experiment section, a more efficient network even indicates worse generalization performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 391, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 391, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 391, + 545, + 712 + ], + "type": "text", + "content": "Generalization Problem in low-level vision often arises when the testing degradation does not match the training degradation, e.g., different downsampling kernel in superresolution [30, 40, 47]. We typically develop deep denoising models based on Gaussian noise in the laboratory setting. However, noise in the real-world is mostly non-Gaussian. Models trained on Gaussian noise fail in these non-Gaussian scenarios. There are two main categories of solutions to this problem. The first is to make training datasets with noise modeling as close to reality as possible during development, e.g., synthesizing real noise according to physical system modeling [5, 68], learning to generate real noise [11, 24, 71], collecting real noise - clean image pairs for training [1, 33, 41, 57]. Although the models obtained by these methods can improve the effect on the target noise, they still cannot generalize to out-of-distribution noise. Another category of solutions is to develop \"blind\" denoising models, which are supposed to deal with unknown noise [41, 72, 80]. These methods usually simply assume that the noise level is unknown, or train on a large amount of noise types [79], which also fails to generalize to other noise not present in the training set. Few workd have been proposed to study the reasons for the lack of generalization ability in low-level vision [40]. Liu et al. [49] argue that networks tend to overfit to degradations and show degradation \"semantics\" inside the network. The presence of these representations often means a decrease in general-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 750, + 314, + 757 + ], + "type": "text", + "content": "1693" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 140, + 160 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 140, + 160 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 140, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 140, + 160 + ], + "type": "image", + "image_path": "119cbf8907f4a27a639fefac363dafdb34aac61e4e18f48fafccc7b877012ad4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 161, + 117, + 170 + ], + "lines": [ + { + "bbox": [ + 71, + 161, + 117, + 170 + ], + "spans": [ + { + "bbox": [ + 71, + 161, + 117, + 170 + ], + "type": "text", + "content": "Training image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 142, + 70, + 212, + 112 + ], + "blocks": [ + { + "bbox": [ + 142, + 70, + 212, + 112 + ], + "lines": [ + { + "bbox": [ + 142, + 70, + 212, + 112 + ], + "spans": [ + { + "bbox": [ + 142, + 70, + 212, + 112 + ], + "type": "image", + "image_path": "c6249b64558173b20eb0d0fb8b5897f49b7631dcadcc69f81dbb2c7c7d745870.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 112, + 195, + 119 + ], + "lines": [ + { + "bbox": [ + 159, + 112, + 195, + 119 + ], + "spans": [ + { + "bbox": [ + 159, + 112, + 195, + 119 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 287, + 111 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 287, + 111 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 287, + 111 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 287, + 111 + ], + "type": "image", + "image_path": "0443ac2881d07b01bd05fa1f007c43f5739b7aa8f0e63771833f60fcfdfc9df7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 112, + 279, + 119 + ], + "lines": [ + { + "bbox": [ + 221, + 112, + 279, + 119 + ], + "spans": [ + { + "bbox": [ + 221, + 112, + 279, + 119 + ], + "type": "text", + "content": "Denoised by SwinIR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 142, + 120, + 212, + 161 + ], + "blocks": [ + { + "bbox": [ + 142, + 120, + 212, + 161 + ], + "lines": [ + { + "bbox": [ + 142, + 120, + 212, + 161 + ], + "spans": [ + { + "bbox": [ + 142, + 120, + 212, + 161 + ], + "type": "image", + "image_path": "223a24012884b990720efd14fe331f7e34fc443579bd20fc4685ae36faebc9ae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 161, + 211, + 169 + ], + "lines": [ + { + "bbox": [ + 143, + 161, + 211, + 169 + ], + "spans": [ + { + "bbox": [ + 143, + 161, + 211, + 169 + ], + "type": "text", + "content": "Denoised by our method" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 176, + 287, + 286 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 286 + ], + "type": "text", + "content": "Figure 2. SwinIR, when trained solely on immunohistochemistry images with Gaussian noise, can still denoise natural images. This observation supports the assertion that most existing methods perform denoising primarily through overfitting the training noise. In contrast, our approach emphasizes reconstructing natural image textures and edges observed in the training set on natural images, rather than relying on noise overfitting for denoising. This distinction underlines the fundamental difference between our method and previous approaches. \"Our reconstruction result\" refers to using our model but taking masked images as input." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 215, + 120, + 287, + 161 + ], + "blocks": [ + { + "bbox": [ + 215, + 120, + 287, + 161 + ], + "lines": [ + { + "bbox": [ + 215, + 120, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 215, + 120, + 287, + 161 + ], + "type": "image", + "image_path": "c6fb6ce865e0cd411b08d131e9f683530059f9985c7881988374e6ab7dd0540a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 162, + 285, + 169 + ], + "lines": [ + { + "bbox": [ + 216, + 162, + 285, + 169 + ], + "spans": [ + { + "bbox": [ + 216, + 162, + 285, + 169 + ], + "type": "text", + "content": "Our reconstruction result" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 297, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 346 + ], + "type": "text", + "content": "ization ability. The utilization of this knowledge can guide us to analyze and evaluate the generalization performance [50]. Apart from that, few works have been proposed to improve the generalization ability of denoising models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 518 + ], + "type": "text", + "content": "Masked modeling for language [6, 20, 59, 60] is successful for learning pre-trained representations that generalize well to various downstream tasks. These methods mask out a portion of the input sequence and train models to predict the missing content. A similar approach can also be applied to the vision model pre-training. Masked image models learn representations from corrupted images. The earliest attempts in this regard can be traced back at least to the denoising auto-encoder [66]. Since then, many works have used predicting missing parts of images to learn efficient image representations [4, 12, 34, 56, 69]. However, there have been few successful attempts to apply masked image modeling to low-level vision, even though the masked pretraining method is in the form of low-level vision tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 531, + 102, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 102, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 102, + 544 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 552, + 287, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 600 + ], + "type": "text", + "content": "Our objective is to create denoising models capable of generalizing to noise not encountered in the training set. In this section, we first discuss our motivation before delving into the specifics of our masked training method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Motivation. When training a deep network on a large number of images, the expectation is for the network to learn to discern the rich semantics of natural images from noise-contaminated test cases. However, several studies have noted that the semantics and knowledge acquired by low-level vision networks differ significantly from our expectations [29, 49, 50, 52]. We argue that the poor generalization ability of denoising models results from our training method, which leads the model to focus on overfit" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 354, + 118 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 354, + 118 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 354, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 354, + 118 + ], + "type": "image", + "image_path": "ffa6648d79692cf7f4f6022cdec0f9b44a4dcdddd8f25b5afec223ac9b4c4e62.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 357, + 70, + 449, + 118 + ], + "blocks": [ + { + "bbox": [ + 357, + 70, + 449, + 118 + ], + "lines": [ + { + "bbox": [ + 357, + 70, + 449, + 118 + ], + "spans": [ + { + "bbox": [ + 357, + 70, + 449, + 118 + ], + "type": "image", + "image_path": "da218d2b5d65f12e702bf9fa9087d80f4c0f54fbae698624ecd46aba03169d36.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 119, + 355, + 167 + ], + "blocks": [ + { + "bbox": [ + 307, + 119, + 355, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 119, + 355, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 355, + 167 + ], + "type": "image", + "image_path": "994bc65861c5a73bf79b972e3af048355fa19873f59bd3845f74fbe8e9c30aab.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 168, + 345, + 176 + ], + "lines": [ + { + "bbox": [ + 318, + 168, + 345, + 176 + ], + "spans": [ + { + "bbox": [ + 318, + 168, + 345, + 176 + ], + "type": "text", + "content": "original" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 183, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 305, + 183, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 183, + 545, + 217 + ], + "type": "text", + "content": "Figure 3. The illustration of the proposed mask-and-complete training strategy. Even if a large number of pixels are masked, the model can still reconstruct the input to some extent." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 357, + 119, + 449, + 167 + ], + "blocks": [ + { + "bbox": [ + 357, + 119, + 449, + 167 + ], + "lines": [ + { + "bbox": [ + 357, + 119, + 449, + 167 + ], + "spans": [ + { + "bbox": [ + 357, + 119, + 449, + 167 + ], + "type": "image", + "image_path": "cc82582f51af3712b10d74d3463ced7992f47c55b3baa5752ca81156a583ba22.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 168, + 421, + 176 + ], + "lines": [ + { + "bbox": [ + 385, + 168, + 421, + 176 + ], + "spans": [ + { + "bbox": [ + 385, + 168, + 421, + 176 + ], + "type": "text", + "content": "mask " + }, + { + "bbox": [ + 385, + 168, + 421, + 176 + ], + "type": "inline_equation", + "content": "70\\%" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 451, + 70, + 545, + 118 + ], + "blocks": [ + { + "bbox": [ + 451, + 70, + 545, + 118 + ], + "lines": [ + { + "bbox": [ + 451, + 70, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 451, + 70, + 545, + 118 + ], + "type": "image", + "image_path": "39c20430259683b0bfb1f88ed520c47cda6fea1fe3c145168f516ef2fb6e238d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 451, + 119, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 451, + 119, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 451, + 119, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 451, + 119, + 545, + 167 + ], + "type": "image", + "image_path": "4e754983b462d15d191d5106df24018fbb2bc7cd5c0c30c823d405eb77be583f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 168, + 515, + 175 + ], + "lines": [ + { + "bbox": [ + 480, + 168, + 515, + 175 + ], + "spans": [ + { + "bbox": [ + 480, + 168, + 515, + 175 + ], + "type": "text", + "content": "mask " + }, + { + "bbox": [ + 480, + 168, + 515, + 175 + ], + "type": "inline_equation", + "content": "90\\%" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 230, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 396 + ], + "type": "text", + "content": "ting the training noise rather than learning image reconstruction. We conduct a simple experiment for verification. We trained a SwinIR denoising network [45] using images that greatly differ from natural images (immunohistochemistry images [65]). We synthesized training data pairs using Gaussian noise, and then assessed the model's performance on natural images with Gaussian noise. According to our hypothesis, if the model learns the content and reconstruction of image semantics from the training set, it should not perform well on natural images, as it has not been exposed to any. If the model is simply overfitting the noise, the model can remove the noise even if the images are different, as the model mainly relies on detecting the noise for denoising." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 399, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 495 + ], + "type": "text", + "content": "The results are presented in Figure 2. As observed, the SwinIR trained on immunohistochemistry images can still denoise and reproduce the natural image. This supports our conjecture regarding generalization ability, indicating that most existing methods perform denoising by overfitting the training noise. Consequently, when the noise deviates from the training conditions, the denoising performance of these models declines significantly." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "This observation also inspires our approach to developing deep denoising models with improved generalization ability. We aim for the model to learn the reconstruction of image textures and structures, rather than focusing only on noise. In this paper, we propose a new masked training strategy for denoising networks. During training, we mask out a portion of the input pixels and then train the deep network to complete them, as shown in Figure 3. Our approach emphasizes reconstructing natural image textures and edges observed in the image, rather than overfitting noise. In Figure 2 we also show the results of our method. It is evident that our approach seeks to reconstruct the immunohistochemistry image texture from the training set on the testing natural image, instead of relying on noise overfitting for denoising. This demonstrates the potential of this idea in improving generalization performance. By training our method on natural images, it will concentrate on reconstructing the content of natural images, aligning with our" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1694" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 149 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 149 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 149 + ], + "type": "image", + "image_path": "b808158284f112b13b4111bafd433e2c8159155b2d4497a81d32a9878f08d546.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 158, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 547, + 182 + ], + "type": "text", + "content": "Figure 4. The transformer architecture of our proposed masked image training. We make a minimal change to the original SwinIR architecture – the input mask operation and the attention masks. Other micro-designs are not essentially different from other Transformers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 190, + 289, + 304 + ], + "blocks": [ + { + "bbox": [ + 47, + 190, + 289, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 289, + 304 + ], + "type": "image", + "image_path": "34182b779ea0773cb66a339789d0f4b7a2cd36037c31611530058b0e0aa07bfe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 305, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 327 + ], + "type": "text", + "content": "Figure 5. Quantitative effect of the attention mask. The histogram differences are also shown above." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 360 + ], + "type": "text", + "content": "core concept of employing deep learning for low-level vision tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": "The Transformer Architecture. Our approach exploits the excellent properties of visual Transformers, so we first describe the basic Transformer backbone used in this study. The shifted window mechanism is proven to be flexible and effective for image processing tasks [15, 45, 78]. We only make minimal changes when applying it to the proposed masked training method without the loss of generality. This model is illustrated in Figure 4. Transformers divide the input signal into tokens and process spatial information using self-attention layers. In our method, a convolution layer with kernel size 1 is used as the feature embedding module to project the 3-channel pixel values into " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": "-dimensional feature tokens. The " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " convolution layer ensures that pixels do not affect each other during feature embedding, which facilitates subsequent masking operations. These feature tokens are gathered with shape " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "H \\times W \\times C" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " are the height, width and feature dimension. The shifted window mechanism first reshapes the feature maps of each frame to " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\frac{HW}{M^2} \\times M^2 \\times C" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " features by partitioning the input into non-overlapping " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "M \\times M" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " local windows, where " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\frac{HW}{M^2}" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " is the total number of windows. We calculate self-attention on the feature tokens within the same window. Therefore, " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " tokens are involved in each standard self-attention operation, and we produce the local window feature " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{M^2 \\times C}" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": ". In each self-attention layer, the query " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " are calculated as " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "Q = XW^Q" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "K = XW^K" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "V = XW^V" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "W^Q, W^K, W^V \\in \\mathbb{R}^{C \\times D}" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " are weight matrices, and " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " is the dimension of projected vectors. Then, we use " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " to query " + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 365, + 289, + 715 + ], + "type": "text", + "content": " to generate the attenuated feature vectors." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 191, + 547, + 248 + ], + "blocks": [ + { + "bbox": [ + 306, + 191, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 306, + 191, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 547, + 248 + ], + "type": "image", + "image_path": "4fd7a18c82c187b7c35d34645a9401784867b438bcbd0b53bdbd44584a74ced8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 249, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 306, + 249, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 545, + 271 + ], + "type": "text", + "content": "Figure 6. The effectiveness of the input mask and attention mask. Note that the brightness of the image is wrong w/o attention mask." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 277, + 538, + 334 + ], + "blocks": [ + { + "bbox": [ + 310, + 277, + 538, + 334 + ], + "lines": [ + { + "bbox": [ + 310, + 277, + 538, + 334 + ], + "spans": [ + { + "bbox": [ + 310, + 277, + 538, + 334 + ], + "type": "table", + "html": "
Input MaskAttention MaskPSNRSSIMMix. noise on CBSD68 [55]
Ratio (%)PSNRSSIM
✓29.170.82276529.570.8657
✓26.960.82027529.760.8678
✓✓29.740.86728528.840.8548
", + "image_path": "3ee6f75f716ecdaef68e3a92c211a139eb964c6e47682e47f954d1a10fe983fa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 335, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 306, + 335, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 545, + 357 + ], + "type": "text", + "content": "Table 1. The importance of using Table 2. Ablation on the at-different mask operations. tention mask ratio." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "text", + "content": "tion map " + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "inline_equation", + "content": "A = \\operatorname{softmax}\\left(\\frac{QK^T}{\\sqrt{D}} + B\\right) \\in \\mathbb{R}^{M^2 \\times M^2}" + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "text", + "content": " is the learnable relative positional encoding. This attention map " + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "text", + "content": " is then used for the weighted sum of " + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "text", + "content": " vectors in " + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 364, + 545, + 426 + ], + "type": "text", + "content": ". The multi-head settings are aligned with SwinIR [45] and ViT [22]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 484 + ], + "type": "text", + "content": "Masked Training. Our masked training mainly consists of two aspects, the input mask and the attention mask. Although both are mask operations, the purpose of these two masks is different. We describe them separately." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "text", + "content": "The Input Mask randomly masks out the feature tokens embedded by the first convolution layer, and encourages the network to complete the masked information during training. The input mask explicitly constructs a very challenging inpainting problem, as shown in Figure 3. It can be seen that even if up to " + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "text", + "content": " of the pixel information is destroyed, the network can still reconstruct the target image to a certain extent. The method is very simple. Given the feature token tensor " + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "text", + "content": ", we randomly replace the token with a [mask token] " + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "text", + "content": " with a probability " + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{IM}}" + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{IM}}" + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "text", + "content": " is called the input mask ratio. The network is trained under the supervision of the " + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 304, + 485, + 546, + 699 + ], + "type": "text", + "content": "-norm of the reconstructed image and the ground truth. The [mask token] can be learnable and initialized with a 0 vector. But we actually found that the 0 vector itself is already a suitable choice. The existence of the input mask forces the network to learn to recognize and reconstruct the content of the image from very limited information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "The Attention Mask. We cannot build usable image pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1695" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 128 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 128 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 128 + ], + "type": "image", + "image_path": "524c77de7b8197cdf6e0befff3128ee7cb3758936020e495ae311e07cdc240de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 132, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 46, + 132, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 287, + 177 + ], + "type": "text", + "content": "Figure 7. The trade-off of choosing different mask ratios. The performance drop on training noise is not significant until " + }, + { + "bbox": [ + 46, + 132, + 287, + 177 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 132, + 287, + 177 + ], + "type": "text", + "content": " masking ratio. Our performance gain on the noise outside the training set is greater than the performance loss on the training set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 426 + ], + "type": "text", + "content": "cessing networks relying solely on the input mask operation. Because during testing, we will input uncorrupted images to retain enough information. At this time, due to the inconsistency between training and testing, the network will tend to increase the brightness of the output image. Such as the example in Figure 5. Since Transformer uses the self-attention operation to process spatial information, we can narrow the gap between training and testing by performing the same mask operation during the self-attention process. The specific mask operation is similar to the input mask, but a different attention mask ratio " + }, + { + "bbox": [ + 46, + 186, + 288, + 426 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{AM}}" + }, + { + "bbox": [ + 46, + 186, + 288, + 426 + ], + "type": "text", + "content": " and [mask token] are used. When some tokens in the self-attention are masked, the attention operation will adjust to the fact that the information of these tokens is no longer reliable. Self-attention will focus on unmasked tokens in each layer and complete the masked information. This operation is difficult to implement on convolutional networks. Figure 5 shows the effect of the attention mask. As can be seen, the attention mask successfully makes the masked trained network work on the unmasked input image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 437, + 128, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 128, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 128, + 451 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": "Training Settings. For synthesizing training data, we sample the clean images from DIV2K [64], Flickr2K [46], BSD500 [3], and WED [51] during training. In our work, all the networks are trained using Gaussian noise with standard deviation " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\sigma = 15" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": ". Each input image is randomly cropped to a spatial resolution of " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": ", and the number of the total training iteration is " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "200\\mathrm{K}" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": ". We adopt the Adam optimizer [38] with " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": " to minimize the " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": " pixel loss. The initial learning rate is set as " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": " and reduced by half at the milestone of " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "100\\mathrm{K}" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": " iterations and " + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "inline_equation", + "content": "150\\mathrm{K}" + }, + { + "bbox": [ + 46, + 457, + 287, + 589 + ], + "type": "text", + "content": " iterations. The batch size is set to 64." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Testing Noise. Since the training process utilizes the Gaussian noise, we evaluate the generalization performance of the models on six other synthetic noise: (1) Speckle noise, a type of noise that occurs during the acquisition of medical images or tomography images. (2) Poisson noise, a type of signal-dependent noise that occurs during the acquisition of digital images.(3) Spatially-correlated noise. This is to synthesize the complex artifact after denoising using a flawed algorithm. It is produced by filtering Gaussian noise with a " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " average kernel. Different standard deviations of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "the Gaussian noise indicate different noise levels. (4) Salt & pepper noise. (5) Image signal processing (ISP) noise. [5] proposes a method to synthesize realistic ISP noise during digital imaging. (6) Mixture noise obtained by mixing the above different types of noise with different levels [79]. The clean images are sampled from the benchmark datasets, including CBSD68 [55], Kodak24 [26], McMaster [82], and Urban100 [36]. We also include two real noise types in this work: the Smartphone Image Denoising Dataset (SIDD) [1] and Monte Carlo (MC) rendered image noise. For evaluation, we follow [27,28] and use the metrics PSNR, SSIM [51], and LPIPS [83] to evaluate the results. Since PSNR and SSIM are questioned in assessing the perceptual quality of images [27,28], we also use the LPIPS as an additional metric." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 259, + 362, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 362, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 362, + 272 + ], + "type": "text", + "content": "4.1. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 278, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 422 + ], + "type": "text", + "content": "Ablation Study. Table 1 and Figure 6 show the effectiveness of using different mask operations. As we can see, without the input mask, the model will lose its generalization ability, and cannot effectively remove the noise outside the training set. Without the attention mask, due to the training-testing inconsistency, the quantitative performance degrades significantly, and the output image will have the wrong brightness. In addition, even without the attention mask, the generalization ability of the model is not significantly affected, and most of the noise is still effectively removed. The input mask is the crucial factor in improving the model the generalization ability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "type": "text", + "content": "Table 3a shows the impact of the different input mask ratios. We test fixed ratios and random ratios from a uniform distribution. From our experiments, fixed ratios are less stable for training than randomly chosen from a range, and the performance is also worse. The best quantitative performance is achieved with random sampling ratios between " + }, + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "type": "inline_equation", + "content": "75\\% \\sim 85\\%" + }, + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "type": "text", + "content": ". This is a trade-off between denoising generalization ability and the preservation of image details. As shown in Figure 7, smaller ratios are not enough for the network to learn the distribution of images because more noise patterns are preserved. The larger ratio improves the model generalization, as the model focuses more on reconstruction. But at the same time, some image details may be lost. For attention mask ratio, we show the effects in Table 2. The optimal ratios are around " + }, + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "The Generalization Performance. We evaluate our deep denoising method on synthetic noise, where our training noise follows a Gaussian distribution with a single noise level, but we test on multiple types of non-Gaussian noise to assess the model's generalization performance. In Figure 11, we compare our method with other state-of-the-art models based on their PSNR and SSIM scores. The results show that our model outperforms all the other models in terms of generalization performance. Particularly, as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1696" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 177, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 177, + 171 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 177, + 171 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 177, + 171 + ], + "type": "image", + "image_path": "911690adc5bd52c632fe96531223a16b82cc3fbb98f44da00ff1b268b5e8c7e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 70, + 269, + 116 + ], + "blocks": [ + { + "bbox": [ + 179, + 70, + 269, + 116 + ], + "lines": [ + { + "bbox": [ + 179, + 70, + 269, + 116 + ], + "spans": [ + { + "bbox": [ + 179, + 70, + 269, + 116 + ], + "type": "image", + "image_path": "ba7f643d21300770df5ca47047bba32ac76ad798c4f7f196f8b80ef503b6e9f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 270, + 70, + 360, + 116 + ], + "blocks": [ + { + "bbox": [ + 270, + 70, + 360, + 116 + ], + "lines": [ + { + "bbox": [ + 270, + 70, + 360, + 116 + ], + "spans": [ + { + "bbox": [ + 270, + 70, + 360, + 116 + ], + "type": "image", + "image_path": "1a6005a27047b8f23abbcc6e9bdfec653843f05fd4c3f0804118ecfa282abd56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 361, + 70, + 451, + 116 + ], + "blocks": [ + { + "bbox": [ + 361, + 70, + 451, + 116 + ], + "lines": [ + { + "bbox": [ + 361, + 70, + 451, + 116 + ], + "spans": [ + { + "bbox": [ + 361, + 70, + 451, + 116 + ], + "type": "image", + "image_path": "3c4b9a663bf3ab8496e5ca7a1fcf16436e4f5e116f001da3610a6efaf5acfb88.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 452, + 70, + 543, + 116 + ], + "blocks": [ + { + "bbox": [ + 452, + 70, + 543, + 116 + ], + "lines": [ + { + "bbox": [ + 452, + 70, + 543, + 116 + ], + "spans": [ + { + "bbox": [ + 452, + 70, + 543, + 116 + ], + "type": "image", + "image_path": "0570e030ec1c76f1d33664aa960ecfd077b2f3af9e68b7ebd12ea0358c81cf8c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 51, + 180, + 177, + 281 + ], + "blocks": [ + { + "bbox": [ + 87, + 172, + 141, + 180 + ], + "lines": [ + { + "bbox": [ + 87, + 172, + 141, + 180 + ], + "spans": [ + { + "bbox": [ + 87, + 172, + 141, + 180 + ], + "type": "text", + "content": "CBSD68: img_053" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 180, + 177, + 281 + ], + "lines": [ + { + "bbox": [ + 51, + 180, + 177, + 281 + ], + "spans": [ + { + "bbox": [ + 51, + 180, + 177, + 281 + ], + "type": "image", + "image_path": "839898d8e0f03601ba079d6d2b986935f2f3bcab26072b3614d7eca537b29e2e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 281, + 143, + 290 + ], + "lines": [ + { + "bbox": [ + 85, + 281, + 143, + 290 + ], + "spans": [ + { + "bbox": [ + 85, + 281, + 143, + 290 + ], + "type": "text", + "content": "CBSD68: img_0046" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 178, + 126, + 269, + 171 + ], + "blocks": [ + { + "bbox": [ + 182, + 118, + 267, + 125 + ], + "lines": [ + { + "bbox": [ + 182, + 118, + 267, + 125 + ], + "spans": [ + { + "bbox": [ + 182, + 118, + 267, + 125 + ], + "type": "text", + "content": "Spatially correlated Gaussian, " + }, + { + "bbox": [ + 182, + 118, + 267, + 125 + ], + "type": "inline_equation", + "content": "\\sigma = 50" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 126, + 269, + 171 + ], + "lines": [ + { + "bbox": [ + 178, + 126, + 269, + 171 + ], + "spans": [ + { + "bbox": [ + 178, + 126, + 269, + 171 + ], + "type": "image", + "image_path": "dd4b664a95ee70c130168eb35db5df4ad5a930aa2e08b231753c000edbe31593.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 270, + 126, + 359, + 171 + ], + "blocks": [ + { + "bbox": [ + 296, + 118, + 334, + 125 + ], + "lines": [ + { + "bbox": [ + 296, + 118, + 334, + 125 + ], + "spans": [ + { + "bbox": [ + 296, + 118, + 334, + 125 + ], + "type": "text", + "content": "DnCNN [80]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 126, + 359, + 171 + ], + "lines": [ + { + "bbox": [ + 270, + 126, + 359, + 171 + ], + "spans": [ + { + "bbox": [ + 270, + 126, + 359, + 171 + ], + "type": "image", + "image_path": "0273b063d46fe4758fe5ce12293c32ca44f553f2465974718760d71fb35fcb1d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 361, + 126, + 451, + 171 + ], + "blocks": [ + { + "bbox": [ + 390, + 118, + 422, + 125 + ], + "lines": [ + { + "bbox": [ + 390, + 118, + 422, + 125 + ], + "spans": [ + { + "bbox": [ + 390, + 118, + 422, + 125 + ], + "type": "text", + "content": "RIDNet [2]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 126, + 451, + 171 + ], + "lines": [ + { + "bbox": [ + 361, + 126, + 451, + 171 + ], + "spans": [ + { + "bbox": [ + 361, + 126, + 451, + 171 + ], + "type": "image", + "image_path": "fb8fa6a8e05f737d30a295f1adf893c55a7ed0b4b9eae0d4673b876aba467137.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 452, + 126, + 542, + 171 + ], + "blocks": [ + { + "bbox": [ + 481, + 118, + 515, + 125 + ], + "lines": [ + { + "bbox": [ + 481, + 118, + 515, + 125 + ], + "spans": [ + { + "bbox": [ + 481, + 118, + 515, + 125 + ], + "type": "text", + "content": "RNAN [85]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 126, + 542, + 171 + ], + "lines": [ + { + "bbox": [ + 452, + 126, + 542, + 171 + ], + "spans": [ + { + "bbox": [ + 452, + 126, + 542, + 171 + ], + "type": "image", + "image_path": "f57e5576cb301b3a729d010190f1a677ac968b903c6bcc9638c7329783fdce54.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 51, + 290, + 178, + 391 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 290, + 178, + 391 + ], + "spans": [ + { + "bbox": [ + 51, + 290, + 178, + 391 + ], + "type": "image", + "image_path": "ba44400244d24b897af75374de660e99abe46155661b69455f57037e8120bd65.jpg" + } + ] + } + ], + "index": 20, + "type": "text" + }, + { + "bbox": [ + 86, + 392, + 143, + 400 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 392, + 143, + 400 + ], + "spans": [ + { + "bbox": [ + 86, + 392, + 143, + 400 + ], + "type": "text", + "content": "CBSD68: img_0067" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 178, + 180, + 269, + 225 + ], + "blocks": [ + { + "bbox": [ + 202, + 172, + 246, + 181 + ], + "lines": [ + { + "bbox": [ + 202, + 172, + 246, + 181 + ], + "spans": [ + { + "bbox": [ + 202, + 172, + 246, + 181 + ], + "type": "text", + "content": "Restormer [75]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 180, + 269, + 225 + ], + "lines": [ + { + "bbox": [ + 178, + 180, + 269, + 225 + ], + "spans": [ + { + "bbox": [ + 178, + 180, + 269, + 225 + ], + "type": "image", + "image_path": "77a48262e2831525ae5adca328406795d828389886d178614aff82f16c768002.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 270, + 180, + 359, + 225 + ], + "blocks": [ + { + "bbox": [ + 297, + 172, + 333, + 181 + ], + "lines": [ + { + "bbox": [ + 297, + 172, + 333, + 181 + ], + "spans": [ + { + "bbox": [ + 297, + 172, + 333, + 181 + ], + "type": "text", + "content": "SwinIR [45]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 180, + 359, + 225 + ], + "lines": [ + { + "bbox": [ + 270, + 180, + 359, + 225 + ], + "spans": [ + { + "bbox": [ + 270, + 180, + 359, + 225 + ], + "type": "image", + "image_path": "de034ae40e8ffbed275120b127d610984c406f64ab3cd1534417f5877377f784.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 361, + 180, + 451, + 225 + ], + "blocks": [ + { + "bbox": [ + 394, + 172, + 418, + 180 + ], + "lines": [ + { + "bbox": [ + 394, + 172, + 418, + 180 + ], + "spans": [ + { + "bbox": [ + 394, + 172, + 418, + 180 + ], + "type": "text", + "content": "baseline" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 180, + 451, + 225 + ], + "lines": [ + { + "bbox": [ + 361, + 180, + 451, + 225 + ], + "spans": [ + { + "bbox": [ + 361, + 180, + 451, + 225 + ], + "type": "image", + "image_path": "f6b2b2a3068d5d0f5a2d6f3eb35c4fa9a753835ef077d3589cf3570a639de835.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 452, + 180, + 542, + 225 + ], + "blocks": [ + { + "bbox": [ + 473, + 172, + 522, + 181 + ], + "lines": [ + { + "bbox": [ + 473, + 172, + 522, + 181 + ], + "spans": [ + { + "bbox": [ + 473, + 172, + 522, + 181 + ], + "type": "text", + "content": "Masked Training" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 180, + 542, + 225 + ], + "lines": [ + { + "bbox": [ + 452, + 180, + 542, + 225 + ], + "spans": [ + { + "bbox": [ + 452, + 180, + 542, + 225 + ], + "type": "image", + "image_path": "9f3256261b64aea66cab5d72c3291484542cfa7f84a2d3942006d75a98f990cc.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 51, + 392, + 177, + 400 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 392, + 177, + 400 + ], + "spans": [ + { + "bbox": [ + 51, + 392, + 177, + 400 + ], + "type": "text", + "content": "CBSD68: img_0067" + } + ] + } + ], + "index": 30, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 178, + 236, + 269, + 281 + ], + "blocks": [ + { + "bbox": [ + 183, + 226, + 264, + 236 + ], + "lines": [ + { + "bbox": [ + 183, + 226, + 264, + 236 + ], + "spans": [ + { + "bbox": [ + 183, + 226, + 264, + 236 + ], + "type": "text", + "content": "Speckle noise, " + }, + { + "bbox": [ + 183, + 226, + 264, + 236 + ], + "type": "inline_equation", + "content": "\\sigma^2 = 0.013" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 236, + 269, + 281 + ], + "lines": [ + { + "bbox": [ + 178, + 236, + 269, + 281 + ], + "spans": [ + { + "bbox": [ + 178, + 236, + 269, + 281 + ], + "type": "image", + "image_path": "d94ceccba28fffb236c9a57076403168602f47af7e593981afc4dcdcbeca6d0d.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 270, + 236, + 359, + 281 + ], + "blocks": [ + { + "bbox": [ + 296, + 227, + 334, + 236 + ], + "lines": [ + { + "bbox": [ + 296, + 227, + 334, + 236 + ], + "spans": [ + { + "bbox": [ + 296, + 227, + 334, + 236 + ], + "type": "text", + "content": "DnCNN [80]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 236, + 359, + 281 + ], + "lines": [ + { + "bbox": [ + 270, + 236, + 359, + 281 + ], + "spans": [ + { + "bbox": [ + 270, + 236, + 359, + 281 + ], + "type": "image", + "image_path": "30a4277d6c51acc68461321ab2df739e9052363bc89966da788e9db4d86aae3c.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 361, + 236, + 451, + 281 + ], + "blocks": [ + { + "bbox": [ + 390, + 228, + 422, + 236 + ], + "lines": [ + { + "bbox": [ + 390, + 228, + 422, + 236 + ], + "spans": [ + { + "bbox": [ + 390, + 228, + 422, + 236 + ], + "type": "text", + "content": "RIDNet [2]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 236, + 451, + 281 + ], + "lines": [ + { + "bbox": [ + 361, + 236, + 451, + 281 + ], + "spans": [ + { + "bbox": [ + 361, + 236, + 451, + 281 + ], + "type": "image", + "image_path": "5078b2a66d132996a58c2364a0b12f5a6509e0a5cc5e18ef80f9ba8076902491.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 452, + 236, + 542, + 281 + ], + "blocks": [ + { + "bbox": [ + 481, + 228, + 515, + 236 + ], + "lines": [ + { + "bbox": [ + 481, + 228, + 515, + 236 + ], + "spans": [ + { + "bbox": [ + 481, + 228, + 515, + 236 + ], + "type": "text", + "content": "RNAN [85]" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 236, + 542, + 281 + ], + "lines": [ + { + "bbox": [ + 452, + 236, + 542, + 281 + ], + "spans": [ + { + "bbox": [ + 452, + 236, + 542, + 281 + ], + "type": "image", + "image_path": "f9a1fc154dd1a6f4079ad25bc013cab9c6cb444785c626a959c745b3aa0c1d23.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 51, + 290, + 178, + 391 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 290, + 178, + 391 + ], + "spans": [ + { + "bbox": [ + 51, + 290, + 178, + 391 + ], + "type": "image", + "image_path": "ba44400244d24b897af75374de660e99abe46155661b69455f57037e8120bd65.jpg" + } + ] + } + ], + "index": 39, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 178, + 290, + 269, + 335 + ], + "blocks": [ + { + "bbox": [ + 202, + 282, + 246, + 290 + ], + "lines": [ + { + "bbox": [ + 202, + 282, + 246, + 290 + ], + "spans": [ + { + "bbox": [ + 202, + 282, + 246, + 290 + ], + "type": "text", + "content": "Restormer [75]" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 290, + 269, + 335 + ], + "lines": [ + { + "bbox": [ + 178, + 290, + 269, + 335 + ], + "spans": [ + { + "bbox": [ + 178, + 290, + 269, + 335 + ], + "type": "image", + "image_path": "275a77bf7eb8d5948aa0882874985f1d45bfeca27f00d0f347fe3b15e22c941b.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 270, + 290, + 359, + 335 + ], + "blocks": [ + { + "bbox": [ + 296, + 282, + 333, + 290 + ], + "lines": [ + { + "bbox": [ + 296, + 282, + 333, + 290 + ], + "spans": [ + { + "bbox": [ + 296, + 282, + 333, + 290 + ], + "type": "text", + "content": "SwinIR [45]" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 290, + 359, + 335 + ], + "lines": [ + { + "bbox": [ + 270, + 290, + 359, + 335 + ], + "spans": [ + { + "bbox": [ + 270, + 290, + 359, + 335 + ], + "type": "image", + "image_path": "f231ed7b81df15f5433d229b3ef1cf9357accf20c1c37a72b62ed7cd12f2e1f5.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 361, + 290, + 451, + 335 + ], + "blocks": [ + { + "bbox": [ + 394, + 282, + 418, + 290 + ], + "lines": [ + { + "bbox": [ + 394, + 282, + 418, + 290 + ], + "spans": [ + { + "bbox": [ + 394, + 282, + 418, + 290 + ], + "type": "text", + "content": "baseline" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 290, + 451, + 335 + ], + "lines": [ + { + "bbox": [ + 361, + 290, + 451, + 335 + ], + "spans": [ + { + "bbox": [ + 361, + 290, + 451, + 335 + ], + "type": "image", + "image_path": "d3cc84996467047107870bed410e4105e4fa7009dc9751b8e14f08e8f5b8d41e.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 452, + 290, + 542, + 335 + ], + "blocks": [ + { + "bbox": [ + 473, + 282, + 522, + 290 + ], + "lines": [ + { + "bbox": [ + 473, + 282, + 522, + 290 + ], + "spans": [ + { + "bbox": [ + 473, + 282, + 522, + 290 + ], + "type": "text", + "content": "Masked Training" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 290, + 542, + 335 + ], + "lines": [ + { + "bbox": [ + 452, + 290, + 542, + 335 + ], + "spans": [ + { + "bbox": [ + 452, + 290, + 542, + 335 + ], + "type": "image", + "image_path": "aecd15b639520ea0ba5ceca1a1c30f2dae901e744bc71bec120169d84cc8cd2b.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 51, + 392, + 177, + 400 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 392, + 177, + 400 + ], + "spans": [ + { + "bbox": [ + 51, + 392, + 177, + 400 + ], + "type": "text", + "content": "CBSD68: img_0067" + } + ] + } + ], + "index": 48, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 178, + 346, + 269, + 391 + ], + "blocks": [ + { + "bbox": [ + 181, + 338, + 266, + 345 + ], + "lines": [ + { + "bbox": [ + 181, + 338, + 266, + 345 + ], + "spans": [ + { + "bbox": [ + 181, + 338, + 266, + 345 + ], + "type": "text", + "content": "Salt-and-pepper noise, " + }, + { + "bbox": [ + 181, + 338, + 266, + 345 + ], + "type": "inline_equation", + "content": "d = 0.02" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 346, + 269, + 391 + ], + "lines": [ + { + "bbox": [ + 178, + 346, + 269, + 391 + ], + "spans": [ + { + "bbox": [ + 178, + 346, + 269, + 391 + ], + "type": "image", + "image_path": "cd73fd5c98e236ab194eaacbe68e86a819523df5001785f7884a02a8844ae7eb.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 392, + 246, + 400 + ], + "lines": [ + { + "bbox": [ + 202, + 392, + 246, + 400 + ], + "spans": [ + { + "bbox": [ + 202, + 392, + 246, + 400 + ], + "type": "text", + "content": "Restormer [75]" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 270, + 346, + 359, + 391 + ], + "blocks": [ + { + "bbox": [ + 296, + 338, + 333, + 345 + ], + "lines": [ + { + "bbox": [ + 296, + 338, + 333, + 345 + ], + "spans": [ + { + "bbox": [ + 296, + 338, + 333, + 345 + ], + "type": "text", + "content": "DnCNN [80]" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 346, + 359, + 391 + ], + "lines": [ + { + "bbox": [ + 270, + 346, + 359, + 391 + ], + "spans": [ + { + "bbox": [ + 270, + 346, + 359, + 391 + ], + "type": "image", + "image_path": "703ed3f6c493a10ad68348621588476a65c32bb9688cbcbccfb4d8d6bc09ea0e.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 392, + 333, + 400 + ], + "lines": [ + { + "bbox": [ + 296, + 392, + 333, + 400 + ], + "spans": [ + { + "bbox": [ + 296, + 392, + 333, + 400 + ], + "type": "text", + "content": "SwinIR [45]" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 361, + 346, + 451, + 391 + ], + "blocks": [ + { + "bbox": [ + 390, + 338, + 422, + 346 + ], + "lines": [ + { + "bbox": [ + 390, + 338, + 422, + 346 + ], + "spans": [ + { + "bbox": [ + 390, + 338, + 422, + 346 + ], + "type": "text", + "content": "RIDNet [2]" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 346, + 451, + 391 + ], + "lines": [ + { + "bbox": [ + 361, + 346, + 451, + 391 + ], + "spans": [ + { + "bbox": [ + 361, + 346, + 451, + 391 + ], + "type": "image", + "image_path": "828c5e8f8e3b3a1110e1f3a72b9202b3d3962576fce339911b29ffb6325f6b1c.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 392, + 418, + 399 + ], + "lines": [ + { + "bbox": [ + 394, + 392, + 418, + 399 + ], + "spans": [ + { + "bbox": [ + 394, + 392, + 418, + 399 + ], + "type": "text", + "content": "baseline" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 452, + 346, + 542, + 391 + ], + "blocks": [ + { + "bbox": [ + 481, + 338, + 515, + 346 + ], + "lines": [ + { + "bbox": [ + 481, + 338, + 515, + 346 + ], + "spans": [ + { + "bbox": [ + 481, + 338, + 515, + 346 + ], + "type": "text", + "content": "RNAN [85]" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 346, + 542, + 391 + ], + "lines": [ + { + "bbox": [ + 452, + 346, + 542, + 391 + ], + "spans": [ + { + "bbox": [ + 452, + 346, + 542, + 391 + ], + "type": "image", + "image_path": "ec12561b78ec63a2fc1c9fa1d6532e9bddc1238f5e842786262a2847d6896d2e.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 392, + 522, + 400 + ], + "lines": [ + { + "bbox": [ + 473, + 392, + 522, + 400 + ], + "spans": [ + { + "bbox": [ + 473, + 392, + 522, + 400 + ], + "type": "text", + "content": "Masked Training" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 48, + 433, + 129, + 539 + ], + "blocks": [ + { + "bbox": [ + 46, + 403, + 545, + 425 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 545, + 425 + ], + "type": "text", + "content": "Figure 8. Visual comparison on out-of-distribution noise. When all other methods fail completely, our method is still able to denoise effectively. Please refer to the supplementary material to see more visual results." + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 433, + 129, + 539 + ], + "lines": [ + { + "bbox": [ + 48, + 433, + 129, + 539 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 129, + 539 + ], + "type": "image", + "image_path": "7bae78ae4f6ea8a6835f49474e7c8ab0ebf0eddf330036be55d6b147e9a1bb29.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 540, + 124, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 540, + 124, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 124, + 548 + ], + "type": "text", + "content": "Rendered image (256 spp)" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 130, + 433, + 206, + 483 + ], + "blocks": [ + { + "bbox": [ + 130, + 433, + 206, + 483 + ], + "lines": [ + { + "bbox": [ + 130, + 433, + 206, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 206, + 483 + ], + "type": "image", + "image_path": "eeaaf3455adea2a9d6c20e797d80f0ae70bed4051b990536a8bd8c0378540b72.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 483, + 178, + 491 + ], + "lines": [ + { + "bbox": [ + 161, + 483, + 178, + 491 + ], + "spans": [ + { + "bbox": [ + 161, + 483, + 178, + 491 + ], + "type": "text", + "content": "Noisy" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 208, + 433, + 283, + 483 + ], + "blocks": [ + { + "bbox": [ + 208, + 433, + 283, + 483 + ], + "lines": [ + { + "bbox": [ + 208, + 433, + 283, + 483 + ], + "spans": [ + { + "bbox": [ + 208, + 433, + 283, + 483 + ], + "type": "image", + "image_path": "2ceb774440941b0b2a2ff63c653ab416ea817c274d24b5c70d3f9e716646a63c.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 483, + 260, + 491 + ], + "lines": [ + { + "bbox": [ + 233, + 483, + 260, + 491 + ], + "spans": [ + { + "bbox": [ + 233, + 483, + 260, + 491 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 285, + 433, + 362, + 483 + ], + "blocks": [ + { + "bbox": [ + 285, + 433, + 362, + 483 + ], + "lines": [ + { + "bbox": [ + 285, + 433, + 362, + 483 + ], + "spans": [ + { + "bbox": [ + 285, + 433, + 362, + 483 + ], + "type": "image", + "image_path": "7d983ce7d996eea169e7dbf638f5d573581a10ec55a3dcb33d22a60563922a98.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 483, + 335, + 491 + ], + "lines": [ + { + "bbox": [ + 312, + 483, + 335, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 483, + 335, + 491 + ], + "type": "text", + "content": "DnCNN" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 372, + 433, + 429, + 483 + ], + "blocks": [ + { + "bbox": [ + 372, + 433, + 429, + 483 + ], + "lines": [ + { + "bbox": [ + 372, + 433, + 429, + 483 + ], + "spans": [ + { + "bbox": [ + 372, + 433, + 429, + 483 + ], + "type": "image", + "image_path": "6ca69d2eafa1f35c082f44c3b8a8e156a79bb17ef4014054b21c2f93f8d1aedd.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 483, + 410, + 491 + ], + "lines": [ + { + "bbox": [ + 392, + 483, + 410, + 491 + ], + "spans": [ + { + "bbox": [ + 392, + 483, + 410, + 491 + ], + "type": "text", + "content": "Noisy" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 430, + 433, + 485, + 483 + ], + "blocks": [ + { + "bbox": [ + 430, + 433, + 485, + 483 + ], + "lines": [ + { + "bbox": [ + 430, + 433, + 485, + 483 + ], + "spans": [ + { + "bbox": [ + 430, + 433, + 485, + 483 + ], + "type": "image", + "image_path": "dfaf45f3c4f702d8cb73a92f4b39f63cf3c948f60e6cb75839a66ab787e2113a.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 483, + 471, + 491 + ], + "lines": [ + { + "bbox": [ + 444, + 483, + 471, + 491 + ], + "spans": [ + { + "bbox": [ + 444, + 483, + 471, + 491 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 487, + 433, + 543, + 483 + ], + "blocks": [ + { + "bbox": [ + 487, + 433, + 543, + 483 + ], + "lines": [ + { + "bbox": [ + 487, + 433, + 543, + 483 + ], + "spans": [ + { + "bbox": [ + 487, + 433, + 543, + 483 + ], + "type": "image", + "image_path": "b5a6acba261e5342dbdb79af8919c0e24fe52720256980c08a728a7825cef5eb.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 483, + 525, + 491 + ], + "lines": [ + { + "bbox": [ + 503, + 483, + 525, + 491 + ], + "spans": [ + { + "bbox": [ + 503, + 483, + 525, + 491 + ], + "type": "text", + "content": "DnCNN" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 130, + 484, + 206, + 540 + ], + "blocks": [ + { + "bbox": [ + 48, + 549, + 328, + 560 + ], + "lines": [ + { + "bbox": [ + 48, + 549, + 328, + 560 + ], + "spans": [ + { + "bbox": [ + 48, + 549, + 328, + 560 + ], + "type": "text", + "content": "Figure 9. Visual results of denoising a Monte Carlo rendered image." + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 484, + 206, + 540 + ], + "lines": [ + { + "bbox": [ + 130, + 484, + 206, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 206, + 540 + ], + "type": "image", + "image_path": "76899e35721f963c496c81386757bfb0d70a51e04d57560d6d4071adabea1950.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 540, + 179, + 548 + ], + "lines": [ + { + "bbox": [ + 160, + 540, + 179, + 548 + ], + "spans": [ + { + "bbox": [ + 160, + 540, + 179, + 548 + ], + "type": "text", + "content": "RNAN" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 208, + 484, + 283, + 540 + ], + "blocks": [ + { + "bbox": [ + 208, + 484, + 283, + 540 + ], + "lines": [ + { + "bbox": [ + 208, + 484, + 283, + 540 + ], + "spans": [ + { + "bbox": [ + 208, + 484, + 283, + 540 + ], + "type": "image", + "image_path": "c5676def3d9a50554072b314f2ad9b7a555b5e2767dd601dc52800c951b8ec94.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 540, + 257, + 548 + ], + "lines": [ + { + "bbox": [ + 236, + 540, + 257, + 548 + ], + "spans": [ + { + "bbox": [ + 236, + 540, + 257, + 548 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_caption" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 285, + 484, + 362, + 540 + ], + "blocks": [ + { + "bbox": [ + 285, + 484, + 362, + 540 + ], + "lines": [ + { + "bbox": [ + 285, + 484, + 362, + 540 + ], + "spans": [ + { + "bbox": [ + 285, + 484, + 362, + 540 + ], + "type": "image", + "image_path": "d84c270097afb20f203ae674985c08cb020fbfad2b4dcf114b24718a5feb4afb.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 540, + 331, + 548 + ], + "lines": [ + { + "bbox": [ + 316, + 540, + 331, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 540, + 331, + 548 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_caption" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 372, + 491, + 429, + 540 + ], + "blocks": [ + { + "bbox": [ + 372, + 491, + 429, + 540 + ], + "lines": [ + { + "bbox": [ + 372, + 491, + 429, + 540 + ], + "spans": [ + { + "bbox": [ + 372, + 491, + 429, + 540 + ], + "type": "image", + "image_path": "a3c9210cc68eb76f09476e6f4f04d844ed8db80c12dda4bcc93fef942831d1d6.jpg" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 540, + 410, + 548 + ], + "lines": [ + { + "bbox": [ + 391, + 540, + 410, + 548 + ], + "spans": [ + { + "bbox": [ + 391, + 540, + 410, + 548 + ], + "type": "text", + "content": "RNAN" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 549, + 531, + 560 + ], + "lines": [ + { + "bbox": [ + 383, + 549, + 531, + 560 + ], + "spans": [ + { + "bbox": [ + 383, + 549, + 531, + 560 + ], + "type": "text", + "content": "Figure 10. Results of ISP noise removal." + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_caption" + } + ], + "index": 79 + }, + { + "type": "image", + "bbox": [ + 430, + 491, + 485, + 540 + ], + "blocks": [ + { + "bbox": [ + 430, + 491, + 485, + 540 + ], + "lines": [ + { + "bbox": [ + 430, + 491, + 485, + 540 + ], + "spans": [ + { + "bbox": [ + 430, + 491, + 485, + 540 + ], + "type": "image", + "image_path": "4714e91bddfc967421fba23d9ce47892e57d1e8a5d85b77b29b6703ee43baa18.jpg" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 540, + 468, + 548 + ], + "lines": [ + { + "bbox": [ + 447, + 540, + 468, + 548 + ], + "spans": [ + { + "bbox": [ + 447, + 540, + 468, + 548 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_caption" + } + ], + "index": 81 + }, + { + "type": "image", + "bbox": [ + 487, + 491, + 542, + 540 + ], + "blocks": [ + { + "bbox": [ + 487, + 491, + 542, + 540 + ], + "lines": [ + { + "bbox": [ + 487, + 491, + 542, + 540 + ], + "spans": [ + { + "bbox": [ + 487, + 491, + 542, + 540 + ], + "type": "image", + "image_path": "08d89ec8452a46ba2e6caf57e5eaa8e727aa57adbabab4fa7405427fb270f45b.jpg" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 540, + 522, + 548 + ], + "lines": [ + { + "bbox": [ + 507, + 540, + 522, + 548 + ], + "spans": [ + { + "bbox": [ + 507, + 540, + 522, + 548 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_caption" + } + ], + "index": 83 + }, + { + "bbox": [ + 46, + 564, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 696 + ], + "type": "text", + "content": "the noise level increases, our model exhibits a slower performance degradation and thus demonstrates better generalization. In contrast, other models suffer from significant performance drops when dealing with more severe noise. We also provide visual comparisons in Figure 8, where our model achieves remarkable denoising results even though it is trained only on Gaussian noise with a fixed standard deviation. In contrast, existing models tend to overfit the training noise and fail when facing unseen noise. More quantitative and qualitative results can be found in the supplementary material." + } + ] + } + ], + "index": 86 + }, + { + "bbox": [ + 304, + 564, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 708 + ], + "type": "text", + "content": "Evaluation on ISP noise. The removal of the ISP noise is of great application value. Brooks et al. [5] present a systematic approach for generating realistic raw data with ISP noise that can facilitate our research. We use the default parameter settings of the method proposed in [5] to synthesize ISP noise on the Kodak24 [26] dataset for testing. The results are shown in Figure 10 and Table 3c. Our method achieves superior results compared to all other methods. Notably, our method achieves a significant lead in LPIPS, indicating that our results exhibit better perceptual quality. Although DnCNN and our method obtain the same PSNR, our method still outperforms DnCNN in terms of SSIM and" + } + ] + } + ], + "index": 87 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1697" + } + ] + } + ], + "index": 88 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 154, + 175 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 154, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 154, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 154, + 175 + ], + "type": "table", + "html": "
Mix. noise on CBSD68 [55]
Ratio (%)PSNRSSIM
7529.170.8132
8529.440.8545
9519.600.7273
70–8029.860.8593
75–8530.040.8756
75–9029.870.8728
75–9529.260.8607
80–9029.740.8672
", + "image_path": "5f55ebf41b33728309cdd5926849b1336c3e6d58893c14e4625eb81fc7ed7ba5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 176, + 150, + 186 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 150, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 150, + 186 + ], + "type": "text", + "content": "(a) Abl. of input mask ratios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 157, + 70, + 391, + 175 + ], + "blocks": [ + { + "bbox": [ + 157, + 70, + 391, + 175 + ], + "lines": [ + { + "bbox": [ + 157, + 70, + 391, + 175 + ], + "spans": [ + { + "bbox": [ + 157, + 70, + 391, + 175 + ], + "type": "table", + "html": "
Method128 samples per pixel64 samples per pixel
PSNRSSIMLPIPSPSNRSSIMLPIPS
DnCNN [80]29.940.78830.267126.280.67790.4216
RIDNet [2]29.960.79210.254826.270.67880.4122
RNAN [85]29.860.78250.270226.260.67430.4290
SwinIR [45]29.320.76270.294326.140.66510.4485
Restormer [75]24.980.65980.457524.590.58800.5375
Dropout [40]28.850.77530.294126.100.66960.4443
baseline29.680.77380.285125.910.65350.4564
Ours30.620.85000.225428.250.76940.3348
", + "image_path": "3f81cbdfb9e8f5e5c1484586df006f88e0ec725cac216748187ae369e2d12661.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 158, + 176, + 389, + 186 + ], + "lines": [ + { + "bbox": [ + 158, + 176, + 389, + 186 + ], + "spans": [ + { + "bbox": [ + 158, + 176, + 389, + 186 + ], + "type": "text", + "content": "(b) Quantitative comparison on Monte Carlo rendered image denoising." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 395, + 70, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 395, + 70, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 395, + 70, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 395, + 70, + 541, + 175 + ], + "type": "table", + "html": "
Synthetic ISP noise [5]
MethodPSNRSSIMLPIPS
DnCNN [80]29.440.78570.3083
RIDNet [2]28.750.74460.3696
RNAN [85]28.470.72430.3601
SwinIR [45]28.390.70790.3346
Restormer [75]19.310.49820.6556
Dropout [40]28.390.78160.2621
baseline28.890.75950.2917
Ours29.440.79200.2368
", + "image_path": "87702f7385f9396681c6d81ec3ed9f5f756c32df982ac580742156b2a97fb973.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 402, + 176, + 530, + 186 + ], + "lines": [ + { + "bbox": [ + 402, + 176, + 530, + 186 + ], + "spans": [ + { + "bbox": [ + 402, + 176, + 530, + 186 + ], + "type": "text", + "content": "(c) Comparison on synthetic ISP noise." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 190, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 545, + 213 + ], + "type": "text", + "content": "Table 3. We train all the models on Gaussian noise, " + }, + { + "bbox": [ + 46, + 190, + 545, + 213 + ], + "type": "inline_equation", + "content": "\\sigma = {15}" + }, + { + "bbox": [ + 46, + 190, + 545, + 213 + ], + "type": "text", + "content": " . All the testing noise is out of the training set, therefore the results can show the models' generalization performance on different unseen noise." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 47, + 219, + 545, + 361 + ], + "blocks": [ + { + "bbox": [ + 47, + 219, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 47, + 219, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 545, + 361 + ], + "type": "image", + "image_path": "56f384c2f38d7cd075f0db2482381134bab882d2519be3d38fc60066c6b6b61e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 362, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 362, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 545, + 396 + ], + "type": "text", + "content": "Figure 11. Performance comparisons on four noise types with different levels on the Kodak24 dataset [26]. All models are trained only on Gaussian noise. Our masked training approach demonstrates good generalization performance across different noise types. We involve multiple types and levels of noise in testing, the results cannot be shown here. More results are shown in the supplementary material." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 402, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 438 + ], + "type": "text", + "content": "LPIPS. Furthermore, as evident from Figure 10, DnCNN's results still contain visible noise, while our method effectively removes the noise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 445, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 625 + ], + "type": "text", + "content": "Evaluation on Monte Carlo rendering noise. Monte Carlo denoising is a vital component of the rendering process since the widespread use in the industry of Monte Carlo rendering algorithms [8, 17, 42]. We use the test dataset proposed by [25] for Monte Carlo rendered image denoising. The test images were rendered in 128 samples-per-pixel (spp) and 64 spp. The lower the spp, the more severe the noise of the image. In order to adapt the test set to our model, we first convert the data set to sRGB color space by tone mapping. Figure 9 and Table 3b show the denoising results. Our method outperforms all methods on both 128spp and 64spp settings. In Figure 9, the existing methods fail completely because of poor generalization. Our model is still able to remove this noise, demonstrating the wide applicability of our method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 634, + 181, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 181, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 181, + 647 + ], + "type": "text", + "content": "4.2. Generalization Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Training curve. Figure 13 shows the training curves of the model with and without the proposed masked training. The models are trained using only Gaussian noise. The baseline method has a significant overfitting problem. The performance of our method gradually improves with train-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 402, + 400, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 400, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 400, + 415 + ], + "type": "text", + "content": "ing without overfitting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 425, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 605 + ], + "type": "text", + "content": "CKA analysis. To investigate how masked training differs from normal training strategy, we utilize the centered kernel alignment (CKA) [18, 61] to analyze the differences between network representations obtained from those two training methods. Due to the limited space, we describe the detail of CKA in supplementary. In Figure 12, we present our key findings. Specifically, Figure 12 (a) shows the cross-model comparison between the baseline model and our masked training model. We observe a significant difference between the two models in terms of their feature correlations in the deeper layers. Specifically, the features of the deeper layers of the baseline model exhibit low correlations with all layers of our model. This finding suggests that these two training methods exhibit inconsistent learning patterns for features, especially for the deeper layers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "To explore how the models perform on different noise types, Figure 12 (b) shows the cross-noise comparison between in-distribution noise and out-of-distribution noise, such as Gaussian and Poisson noise. For the baseline model, we observe a low correlation between different noise types in the deep layers, indicating that the network processes these two types of noise in different ways for the deep layers. This trend holds for other types of noise as well. This phenomenon may be due to the baseline approach causing" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1698" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 76, + 156, + 165 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 156, + 165 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 156, + 165 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 156, + 165 + ], + "type": "image", + "image_path": "d3fddd29972407b25eb0ba26388ae16cf4ca0d4987c003dc878fd14a44fc23e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 167, + 223, + 175 + ], + "lines": [ + { + "bbox": [ + 90, + 167, + 223, + 175 + ], + "spans": [ + { + "bbox": [ + 90, + 167, + 223, + 175 + ], + "type": "text", + "content": "a. cross-model comparison on Gaussian 15" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 165, + 223, + 175 + ], + "lines": [ + { + "bbox": [ + 143, + 165, + 223, + 175 + ], + "spans": [ + { + "bbox": [ + 143, + 165, + 223, + 175 + ], + "type": "text", + "content": "comparison on Gaussian 15" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 76, + 249, + 161 + ], + "blocks": [ + { + "bbox": [ + 156, + 76, + 249, + 161 + ], + "lines": [ + { + "bbox": [ + 156, + 76, + 249, + 161 + ], + "spans": [ + { + "bbox": [ + 156, + 76, + 249, + 161 + ], + "type": "image", + "image_path": "4929cb376de2f96ded1959e0264c7e399669e051f3e929323fd353344a590813.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 262, + 71, + 447, + 167 + ], + "blocks": [ + { + "bbox": [ + 262, + 71, + 447, + 167 + ], + "lines": [ + { + "bbox": [ + 262, + 71, + 447, + 167 + ], + "spans": [ + { + "bbox": [ + 262, + 71, + 447, + 167 + ], + "type": "image", + "image_path": "cbde860c33764b295d60c7fdb749738c9007f3b0a32055d8f11082fe4e4e5a48.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 167, + 497, + 175 + ], + "lines": [ + { + "bbox": [ + 339, + 167, + 497, + 175 + ], + "spans": [ + { + "bbox": [ + 339, + 167, + 497, + 175 + ], + "type": "text", + "content": "b. cross-model comparison between different noise" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 447, + 71, + 544, + 166 + ], + "blocks": [ + { + "bbox": [ + 447, + 71, + 544, + 166 + ], + "lines": [ + { + "bbox": [ + 447, + 71, + 544, + 166 + ], + "spans": [ + { + "bbox": [ + 447, + 71, + 544, + 166 + ], + "type": "image", + "image_path": "575e359a04a23253b4fceb74e6dd6c94d57d558d1b82b7c46b6694096c9cfe8d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 198, + 138, + 261 + ], + "blocks": [ + { + "bbox": [ + 49, + 198, + 138, + 261 + ], + "lines": [ + { + "bbox": [ + 49, + 198, + 138, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 138, + 261 + ], + "type": "image", + "image_path": "fd31056147495e4addd7698a002f31719a18a82e0294e3436227511547e51041.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 262, + 137, + 318 + ], + "blocks": [ + { + "bbox": [ + 49, + 262, + 137, + 318 + ], + "lines": [ + { + "bbox": [ + 49, + 262, + 137, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 262, + 137, + 318 + ], + "type": "image", + "image_path": "24843181b752ba6cb26f219c182b874367b3f289b82de08d2b1d6dc354d9c340.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 49, + 319, + 137, + 394 + ], + "blocks": [ + { + "bbox": [ + 49, + 319, + 137, + 394 + ], + "lines": [ + { + "bbox": [ + 49, + 319, + 137, + 394 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 137, + 394 + ], + "type": "image", + "image_path": "cd9cd3d1e91ead4be9d2800c3ddefea49778b3c8844ac77460d30b615c27d707.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 398, + 284, + 410 + ], + "lines": [ + { + "bbox": [ + 49, + 398, + 284, + 410 + ], + "spans": [ + { + "bbox": [ + 49, + 398, + 284, + 410 + ], + "type": "text", + "content": "Figure 13. The testing curves on different noise types and levels." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 138, + 198, + 230, + 261 + ], + "blocks": [ + { + "bbox": [ + 142, + 180, + 449, + 192 + ], + "lines": [ + { + "bbox": [ + 142, + 180, + 449, + 192 + ], + "spans": [ + { + "bbox": [ + 142, + 180, + 449, + 192 + ], + "type": "text", + "content": "Figure 12. CKA similarity to analyze the representation similarity of network layers." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 198, + 230, + 261 + ], + "lines": [ + { + "bbox": [ + 138, + 198, + 230, + 261 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 230, + 261 + ], + "type": "image", + "image_path": "d901f586b46a50b95c734cc308f2b2eb0f458ceaa71c55efe7c0b94b38680f10.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 138, + 262, + 230, + 323 + ], + "blocks": [ + { + "bbox": [ + 138, + 262, + 230, + 323 + ], + "lines": [ + { + "bbox": [ + 138, + 262, + 230, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 230, + 323 + ], + "type": "image", + "image_path": "706b6c0ba059f4dffb6c278418bf7d06dc4d9eab04aeeda61e76c75dc24e3a64.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 138, + 324, + 230, + 394 + ], + "blocks": [ + { + "bbox": [ + 138, + 324, + 230, + 394 + ], + "lines": [ + { + "bbox": [ + 138, + 324, + 230, + 394 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 230, + 394 + ], + "type": "image", + "image_path": "126ee05a5548aa7e50e43737b33d4467de7efb88576397bf905f42be9b99f5fe.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 231, + 205, + 285, + 250 + ], + "blocks": [ + { + "bbox": [ + 231, + 205, + 285, + 250 + ], + "lines": [ + { + "bbox": [ + 231, + 205, + 285, + 250 + ], + "spans": [ + { + "bbox": [ + 231, + 205, + 285, + 250 + ], + "type": "image", + "image_path": "26e6debaec8127ba4ca711dc9b726b75bc00918754ae9418f4a915a0f50d72c0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 231, + 264, + 285, + 314 + ], + "blocks": [ + { + "bbox": [ + 231, + 264, + 285, + 314 + ], + "lines": [ + { + "bbox": [ + 231, + 264, + 285, + 314 + ], + "spans": [ + { + "bbox": [ + 231, + 264, + 285, + 314 + ], + "type": "image", + "image_path": "087d703e65e83232dee7dc7c9a57f52d68049a61204acab5b94b1cad1dc025ff.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 231, + 326, + 284, + 376 + ], + "blocks": [ + { + "bbox": [ + 231, + 326, + 284, + 376 + ], + "lines": [ + { + "bbox": [ + 231, + 326, + 284, + 376 + ], + "spans": [ + { + "bbox": [ + 231, + 326, + 284, + 376 + ], + "type": "image", + "image_path": "f86d31ff0b155f92efa52cc796205bed8e4fd14f726dd549240d33d90840c1eb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 47, + 422, + 167, + 506 + ], + "blocks": [ + { + "bbox": [ + 47, + 422, + 167, + 506 + ], + "lines": [ + { + "bbox": [ + 47, + 422, + 167, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 167, + 506 + ], + "type": "image", + "image_path": "5782e4e4e87baf48f4bc27b7bf04607a901bcb521c7133923e3b5cb7460b5fad.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 167, + 422, + 286, + 507 + ], + "blocks": [ + { + "bbox": [ + 167, + 422, + 286, + 507 + ], + "lines": [ + { + "bbox": [ + 167, + 422, + 286, + 507 + ], + "spans": [ + { + "bbox": [ + 167, + 422, + 286, + 507 + ], + "type": "image", + "image_path": "6efdd3301d8b9d52835f72e402abe1214857688ea456390470346894ca6560e5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 53, + 536, + 169, + 624 + ], + "blocks": [ + { + "bbox": [ + 47, + 511, + 287, + 533 + ], + "lines": [ + { + "bbox": [ + 47, + 511, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 287, + 533 + ], + "type": "text", + "content": "Figure 14. Comparing generalization ability with the SRGA metric. A lower SRGA value indicates better generalization ability." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 536, + 169, + 624 + ], + "lines": [ + { + "bbox": [ + 53, + 536, + 169, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 169, + 624 + ], + "type": "image", + "image_path": "120fec10a589d031aed46f891b0e7313202a2f177e53ad9a6a969a1f8760c5e2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 627, + 287, + 659 + ], + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 659 + ], + "type": "text", + "content": "Figure 15. The distribution of baseline model features is biased across different noise types. Our method produces similar feature distributions across different noise." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 170, + 536, + 279, + 624 + ], + "blocks": [ + { + "bbox": [ + 170, + 536, + 279, + 624 + ], + "lines": [ + { + "bbox": [ + 170, + 536, + 279, + 624 + ], + "spans": [ + { + "bbox": [ + 170, + 536, + 279, + 624 + ], + "type": "image", + "image_path": "bd0ea18c0d79baeac8e0ea5749127e47d2a7e71ea49cf52da8ae1fbd73c6242e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "the deep layers of the model to overfit to the patterns of the training set, thereby limiting their generalization capabilities to handle different noise types. In contrast, the high correlation between adjacent layers in our masked training" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "text", + "content": "model suggests that the model's representation of two different noise types is similar. Our masked training forces the network to learn the underlying distribution of the images themselves, which makes the model more robust to different noise types and enhances its generalization capability." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 265, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 546, + 479 + ], + "type": "text", + "content": "Quantification of generalization performance. Liu et al. [49, 50] suggest that model generalization ability can be measured by measuring the consistency of the model's representations across different types of noise. They also propose a generalization assessment index for low-level vision networks called SRGA [50]. It is a non-parametric and nonlearning metric which exploits the statistical characteristics of internal features of deep networks. The lower the value of SRGA, the better the generalization ability. In our case, we use Gaussian noise as the reference and other types of noise for testing. Figure 14 shows the SRGA results. Inspired by [50], we visualize the distributions of deep features on different noise types, shown in Figure 15. We can see that for the baseline model, the feature distributions under different noise types deviate from each other significantly. For the model " + }, + { + "bbox": [ + 304, + 265, + 546, + 479 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 265, + 546, + 479 + ], + "type": "text", + "content": " / masked training, the deep feature distributions of different noise types are close to each other. This confirms the effectiveness of our method." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 491, + 463, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 463, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 463, + 503 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 510, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 595 + ], + "type": "text", + "content": "This work presents a masked training method to improve the generalization performance of deep learning-based image denoising models. The limitation of our method is that the mask operation inevitably loses information. How to preserve more details needs to be explored in future work. Our approach is a step towards developing more robust models for real-world applications." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 597, + 546, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 546, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 546, + 705 + ], + "type": "text", + "content": "Acknowledgment. This work is supported in part by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), the National Natural Science Foundation of China under Grant (62276251), NSFC General Project 62072452, Guangdong Provincial Basic and Applied Basic Research Fund-Regional Joint Fund (Project reference number 2020B1515130004), the Joint Lab of CAS-HK, and the Youth Innovation Promotion Association of Chinese Academy of Sciences (No. 2020356)." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1699" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1692-1700, 2018. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3155-3164, 2019. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "text", + "content": "[3] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 2010. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 317 + ], + "type": "text", + "content": "[5] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11036-11045, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 419 + ], + "type": "text", + "content": "[7] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 2, pages 60-65. IEEE, 2005. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 475 + ], + "type": "text", + "content": "[8] Brent Burley, David Adler, Matt Jen-Yuan Chiang, Hank Driskill, Ralf Habel, Patrick Kelly, Peter Kutz, Yining Karl Li, and Daniel Teece. The design and evolution of disney's hyperion renderer. ACM Transactions on Graphics (TOG), 37(3):1-22, 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "type": "text", + "content": "[9] Chang Chen, Zhiwei Xiong, Xinmei Tian, and Feng Wu. Deep boosting for image denoising. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "text", + "content": "[10] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "text", + "content": "[11] Jingwen Chen, Jiawei Chen, Hongyang Chao, and Ming Yang. Image blind denoising with generative adversarial network based noise modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3155-3164, 2018. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[12] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[13] Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE transactions on pattern analysis and machine intelligence, 39(6):1256-1272, 2016. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, and Xiaokang Yang. Recursive generalization transformer for image super-resolution. arXiv preprint arXiv:2303.06373, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[15] Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. In NIPS, 2022. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[16] Shen Cheng, Yuzhi Wang, Haibin Huang, Donghao Liu, Haoqiang Fan, and Shuaicheng Liu. Nbnet: Noise basis learning for image denoising with subspace projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4896-4906, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "text", + "content": "[17] Per Christensen, Julian Fong, Jonathan Shade, Wayne Wooten, Brenden Schubert, Andrew Kensler, Stephen Friedman, Charlie Kilpatrick, Cliff Ramshaw, Marc Bannister, et al. Renderman: An advanced path-tracing architecture for movie rendering. ACM Transactions on Graphics (TOG), 37(3):1-21, 2018. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "text", + "content": "[18] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. The Journal of Machine Learning Research, 13:795-828, 2012. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "type": "text", + "content": "[19] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "text", + "content": "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[21] Nithish Divakar and R Venkatesh Babu. Image denoising via cnns: An adversarial approach. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 80-87, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 520 + ], + "type": "text", + "content": "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[23] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image processing, 15(12):3736-3745, 2006. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 567, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 621 + ], + "type": "text", + "content": "[24] Ruicheng Feng, Jinjin Gu, Yu Qiao, and Chao Dong. Suppressing model overfitting for image super-resolution networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[25] Arthur Firmino, Jeppe Revall Frisvad, and Henrik Wann Jensen. Progressive denoising of monte carlo rendered images. In Computer Graphics Forum, volume 41, pages 1-11. Wiley Online Library, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "type": "text", + "content": "[26] Rich Franzen. Kodak lossless true color image suite. source: http://r0k.us/graphics/kodak/, 1999. 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[27] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for percep" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1700" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "tual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "text", + "content": "[28] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "type": "text", + "content": "[29] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9199-9208, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "text", + "content": "[30] Jinjin Gu, Hannan Lu, Wangmeng Zuo, and Chao Dong. Blind super-resolution with iterative kernel correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1604-1613, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "text", + "content": "[31] Shuhang Gu, Yawei Li, Luc Van Gool, and Radu Timofte. Self-guided network for fast image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2511-2520, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "type": "text", + "content": "[32] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[33] Shi Guo, Zifeei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1722, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "type": "text", + "content": "[34] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[35] Xiaowan Hu, Ruijun Ma, Zhihong Liu, Yuanhao Cai, Xiaole Zhao, Yulun Zhang, and Haoqian Wang. Pseudo 3d autocorrelation network for real image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16175-16184, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 535 + ], + "type": "text", + "content": "[36] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "text", + "content": "[37] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6054-6063, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 613 + ], + "type": "text", + "content": "[38] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 615, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 287, + 667 + ], + "type": "text", + "content": "[39] Filippos Kokkinos and Stamatios Lefkimmiatis. Deep image demosaicking using a cascade of convolutional residual denoising networks. In Proceedings of the European conference on computer vision (ECCV), pages 303-319, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[40] Xiangtao Kong, Xina Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Reflash dropout in image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6002-6012, 2022. 2, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[41] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[42] Christopher Kulla, Alejandro Conty, Clifford Stein, and Larry Gritz. Sony pictures imageworks arnold. ACM Transactions on Graphics (TOG), 37(3):1-18, 2018. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[43] Stamatios Lefkimmiatis. Non-local color image denoising with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3587-3596, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[44] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3204-3213, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[45] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In CVPR, 2021. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[46] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 310, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 352 + ], + "type": "text", + "content": "[47] Anran Liu, Yihao Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Blind image super-resolution: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[48] Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[49] Yihao Liu, Anran Liu, Jinjin Gu, Zhipeng Zhang, Henhao Wu, Yu Qiao, and Chao Dong. “Discovering” semantics” in super-resolution networks. arXiv preprint arXiv:2108.00406, 2021. 2, 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "text", + "content": "[50] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. arXiv preprint arXiv:2205.07019, 2022. 3, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "text", + "content": "[51] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 2016. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 590 + ], + "type": "text", + "content": "[52] Salma Abdel Magid, Zudi Lin, Donglai Wei, Yulun Zhang, Jinjin Gu, and Hanspeter Pfister. Texture-based error analysis for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2118-2127, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "text", + "content": "[53] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In 2009 IEEE 12th international conference on computer vision, pages 2272-2279. IEEE, 2009. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[54] Xiaojiao Mao, Chunhua Shen, and Yu-Bin Yang. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[55] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1701" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 4, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[56] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "text", + "content": "[57] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1586-1595, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 237 + ], + "type": "text", + "content": "[58] Tobias Plötz and Stefan Roth. Neural nearest neighbors networks. Advances in Neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "type": "text", + "content": "[59] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[60] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "text", + "content": "[61] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "text", + "content": "[62] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. arXiv preprint arXiv:2207.08494, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[63] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[64] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, Kyoung Mu Lee, et al. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "type": "text", + "content": "[65] Mathias Uhlen, Per Oksvold, Linn Fagerberg, Emma Lundberg, Kalle Jonasson, Mattias Forsberg, Martin Zwahlen, Caroline Kampf, Kenneth Wester, Sophia Hober, et al. Towards a knowledge-based human protein atlas. Nature biotechnology, 28(12):1248-1250, 2010. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[66] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "type": "text", + "content": "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[68] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758-2767, 2020. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[69] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "type": "text", + "content": "[70] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In CVPR, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 228 + ], + "type": "text", + "content": "[71] Yuan Yuan, Siyuan Liu, Jiawei Zhang, Yongbing Zhang, Chao Dong, and Liang Lin. Unsupervised image superresolution using cycle-in-cycle generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 701-710, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "type": "text", + "content": "[72] Zongsheng Yue, Hongwei Yong, Qian Zhao, Deyu Meng, and Lei Zhang. Variational denoising network: Toward blind noise modeling and removal. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 546, + 319 + ], + "type": "text", + "content": "[73] Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. In European Conference on Computer Vision, pages 41-58. Springer, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 546, + 364 + ], + "type": "text", + "content": "[74] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 431 + ], + "type": "text", + "content": "[75] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 433, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 546, + 487 + ], + "type": "text", + "content": "[76] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 488, + 546, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 546, + 532 + ], + "type": "text", + "content": "[77] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. arXiv preprint arXiv:2303.06440, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 533, + 546, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 546, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 546, + 577 + ], + "type": "text", + "content": "[78] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. arXiv preprint arXiv:2210.01427, 2022. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 546, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 546, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 546, + 623 + ], + "type": "text", + "content": "[79] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Radu Timofte, and Luc Van Gool. Practical blind denoising via swim-conv-unet and data synthesis. arXiv preprint arXiv:2203.13278, 2022. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 546, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 546, + 668 + ], + "type": "text", + "content": "[80] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "text", + "content": "[81] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1702" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[82] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2):023016, 2011. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "text", + "content": "[84] Yulun Zhang, Kunpeng Li, Kai Li, Gan Sun, Yu Kong, and Yun Fu. Accurate and fast image denoising via attention guided scaling. IEEE Transactions on Image Processing, 30:6255-6265, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 229 + ], + "type": "text", + "content": "[85] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "text", + "content": "[86] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480-2495, 2020. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1703" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_content_list.json b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2856ed440946dd9394b55019fe0a67467fa03ab8 --- /dev/null +++ b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_content_list.json @@ -0,0 +1,1525 @@ +[ + { + "type": "text", + "text": "Masked Images Are Counterfactual Samples for Robust Fine-tuning", + "text_level": 1, + "bbox": [ + 140, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yao Xiao", + "bbox": [ + 192, + 180, + 272, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyi Tang", + "bbox": [ + 310, + 181, + 392, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengxu Wei *", + "bbox": [ + 431, + 181, + 542, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cong Liu", + "bbox": [ + 575, + 181, + 653, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liang Lin", + "bbox": [ + 692, + 181, + 774, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sun Yat-sen University", + "bbox": [ + 392, + 199, + 576, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xiaoy99,tangzy27}@mail2.sysu.edu.cn", + "bbox": [ + 153, + 219, + 472, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{weipx3, liucong3}@mail.sysu.edu.cn", + "bbox": [ + 506, + 219, + 815, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "linliang@ieee.org", + "bbox": [ + 406, + 237, + 558, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models are challenged by the distribution shift between the training data and test data. Recently, the large models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts. However, fine-tuning these models can lead to a trade-off between in-distribution (ID) performance and out-of-distribution (OOD) robustness. Existing methods for tackling this trade-off do not explicitly address the OOD robustness problem. In this paper, based on causal analysis of the aforementioned problems, we propose a novel fine-tuning method, which uses masked images as counterfactual samples that help improve the robustness of the fine-tuning model. Specifically, we mask either the semantics-related or semantics-unrelated patches of the images based on class activation map to break the spurious correlation, and refill the masked patches with patches from other images. The resulting counterfactual samples are used in feature-based distillation with the pre-trained model. Extensive experiments verify that regularizing the fine-tuning with the proposed masked images can achieve a better trade-off between ID and OOD performance, surpassing previous methods on the OOD performance. Our code is available at https://github.com/Coxy7/robust-finetuning.", + "bbox": [ + 76, + 318, + 472, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 683, + 209, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has made impressive advances in various tasks on computer vision. Despite the remarkable performance achieved on benchmark datasets, deep models are often challenged by the distribution shift between the training data and test data [5, 15, 16, 33]. It is commonly assumed that the training and test samples follow the same distribution, which may not hold in real-world applications due to the unpredictable change of lighting conditions, viewpoints, backgrounds, etc. Although there are attempts to improve the robustness of deep models to the distribution shift (or OOD robustness), it is still rather under-explored [29, 38].", + "bbox": [ + 75, + 709, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8bba6db4d514f606b8f66ae39d5fae798849202cb53728a56987d9e001bdd4f4.jpg", + "image_caption": [ + "Figure 1. Illustration of our work. Vanilla fine-tuned models tend to learn spurious correlations that degrade the OOD robustness. To tackle this issue, our model learns from the pre-trained model on the counterfactual CAM-based masked images." + ], + "image_footnote": [], + "bbox": [ + 519, + 286, + 879, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large-scale models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts [8, 19, 32]. Fine-tuning these pre-trained models on downstream tasks can be a promising approach to building robust models for different applications. However, it is found that while fine-tuning improves the performance on in-distribution (ID) data, it may reduce that on out-of-distribution (OOD) data [23, 43]. To tackle this trade-off, several methods [23, 42, 43] have been proposed to improve both ID and OOD performance in fine-tuning. However, they do not explicitly address the OOD robustness problem; instead, they implicitly preserve the robustness of the pre-trained model by constraining the distortion of pre-trained weights or using model ensembles.", + "bbox": [ + 496, + 580, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we revisit the issue of robustness degradation in fine-tuning from a causal perspective. A large-scale pre-trained model somewhat shows properties in causality and stays robust to OOD samples [41]. However, when fine-tuning on downstream tasks, a majority of the model parameters tend to be adjusted for the downstream task in fine-tuning due to the highly entangled representation of", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20301", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images, arguably destructive to the generalizable knowledge [20, 35]. In contrast, distribution shifts are usually sparse in the underlying causal factorization of the data generation process [7, 35]. In this low-dimensional case, if we know which variables vary with different data distributions in this factorization (i.e., the non-stationary factors), we can achieve the OOD robustness by simply excluding their influence on the final predictions of the model.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we consider a Structural Causal Model (SCM) [31] for the object-centric image generation process, as depicted in Fig. 2. In this SCM, images are generated according to a non-stationary domain-relevant factor and a stationary semantic factor. Between them is a spurious correlation caused by a hidden non-stationary confounder that influences how the domain-relevant factor changes with the semantic one. To retain the OOD robustness, a fin-tuning model should avoid mapping non-stationary domain-relevant features to the predicted semantics.", + "bbox": [ + 75, + 214, + 470, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose to fine-tune the models with masked images, which serve as counterfactual samples breaking the spurious correlation. Training on these samples helps preserve the stationary and generalizable knowledge of the pre-trained model. Concretely, we either mask the patches that contribute most to the label (i.e., the main object) or mask those with the least contribution (e.g., the context), which can be implemented based on class activation map (CAM) [9,46]. Such image masking forms an manipulation of a factual image and produces a counterfactual sample. Since the pre-trained model can better disentangle invariant features across domains, we require the finetuning model to learn from the pre-trained model on these counterfactual samples, as illustrated in Fig. 1. Furthermore, we argue that simply dropping the masked patches may be insufficient to alleviate the risk of fitting spurious correlations, and we propose to refill the masked patches with those from other images.", + "bbox": [ + 75, + 368, + 470, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We study different combinations of masking strategies (e.g., masking the object or the context) and refilling strategies (e.g., filling with patches from single or multiple images). Experimental results suggest that most of the strategies are applicable to the construction of counterfactual samples that help improve the robustness in fine-tuning, while masking the object generally achieves the best robustness. Compared with existing methods [23, 42, 43] on fine-tuning CLIP [32] models, our approach achieves better average accuracy on various OOD datasets without relying on model ensembles or weight constraints. We also find that, taking the weight-space ensemble of the zero-shot model and our fine-tuned model following WiSE-FT [43], hardly improves the trade-off between ID and OOD accuracy, which contradicts previous observations and implies that our approach may produce essentially different models in comparison with conventional fine-tuning.", + "bbox": [ + 75, + 643, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 498, + 89, + 650, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Causal Perspective on OOD Robustness", + "text_level": 1, + "bbox": [ + 500, + 114, + 841, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Causality holds the promise of OOD robustness as it has the property of robustness under interventions that underlie distribution shifts [2, 35]. To learn the latent causal mechanism in visual data and attain OOD robustness, a branch of methods [21, 28, 36] target at learning domain-invariant causal representations and seek guarantees on generalization. Some approaches approximates this goal based on invariant risk minimization [1, 3], risk extrapolation [22], adaptation speed [7], or variational Bayes [25]. Ilse et al. [18] help models escape from spurious correlations by training on handcrafted analog interventional data. In this work, we aim to explore the preservation of the OOD robustness in pre-trained models equipped with generalizable knowledge during transfer, based on the analysis of an underlying data generation causal mechanism.", + "bbox": [ + 496, + 138, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. ID-OD Trade-off in Fine-tuning", + "text_level": 1, + "bbox": [ + 500, + 375, + 795, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While a series of large pre-trained models exhibit strong OOD robustness [8,19,32], empirical evidence suggests that fine-tuning these models on downstream data may degrade the robustness [23, 32, 43]. It is also theoretically justified that compared with linear probing, vanilla fine-tuning may increase the ID performance while decreasing the OOD performance due to feature distortion [23]. To tackle this ID-OOD trade-off, LP-FT [23] applies linear probing before fine-tuning; Calibrated Ensemble [24] and WiSE-FT [43] propose to take the ensemble of the pre-trained (zero-shot) model and vanilla fine-tuned model in output-space and weight-space, respectively. Model Soup [42] also improves the trade-off via weight-space ensembles, but it utilizes multiple models fine-tuned with different hyper-parameters. While these methods are empirically effective in improving both ID and OOD performance, they only implicitly preserve the robustness of the pre-trained model by constraining the deviation of the downstream model from the pre-trained one. Instead, our approach is based on explicit causal modeling of the OOD robustness problem and the use of counterfactual samples in fine-tuning.", + "bbox": [ + 496, + 398, + 890, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Learning with Masked Images", + "text_level": 1, + "bbox": [ + 500, + 726, + 769, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked image modeling [4,14,40,44] has been proven to be an effective approach to vision model pre-training, where random sampling is a common strategy for masking. In our task, we find that learning with CAM-based masking can better address the spurious correlation in images. Similar to our approach, CSS [10] synthesizes counterfactual images for visual question answering (VQA) by masking out critical objects, for which the corresponding answer is changed to its negative; SwapMix [13] swaps the context objects in the feature space to reduce the reliance of VQA models on", + "bbox": [ + 496, + 750, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20302", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visual context. In this paper, we apply masking on images and refill the masked regions with content from other images, which is inspired by CutMix [45]. Besides, instead of manually constructing the labels for masked images like CSS or CutMix, we take the feature representations of the pre-trained model on the masked images as the supervision for the fine-tuning model.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 212, + 166, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisit OOD Robustness of Fine-tuned Models from a Causal Perspective", + "text_level": 1, + "bbox": [ + 76, + 238, + 467, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When fine-tuning a pre-trained model on downstream tasks in real-world applications, it is usually reliable to generalize only to in-distribution (ID) data. There are also increasing demands for robustness to distribution shifts or generalization to out-of-distribution (OOD) data. However, for a pre-trained large model that has been trained on data from diverse distributions and obtained generalizable knowledge (e.g., CLIP [32]), fine-tuning can lead to dramatic degradation of robustness [23].", + "bbox": [ + 75, + 277, + 467, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This problem can be interpreted from a causal perspective. Suppose that we correctly factorize the data generation process in which variables are causally connected. Then, the discrepancy between ID and OOD (or sub-population) data generation process is usually sparse in this causal factorization [35, 36], e.g., few variables' priors change. The joint distribution that we can observe thereupon shifts. However, without modeling the causal hierarchy in data, the pre-trained models naturally learn dense and entangled features to represent the concepts in the causal factorization. Although distribution shifts are sparse in causal factorization, these models have to densely adapt their parameters in transfer to the downstream data [20]. In this case, the loss of generalizable knowledge is inevitable, which explains the significant decrease of robustness to distribution shifts.", + "bbox": [ + 75, + 414, + 467, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core challenge in modeling the underlying causal factorization of visual data is the dense changes between ID and OOD data, despite the sparsity in causal factorization. For better robustness, a possible solution is to identify which visual features remain unchanged under distribution shifts, also known as stationary features [35]. Without OOD test data in hand to identify stationary features, we seek to model the most important stationary features as the semantic features in the downstream object-centric task, and reduce non-stationary features to domain-relevant features. Finally, we establish our causal model as in Fig. 2.", + "bbox": [ + 75, + 641, + 467, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 SCM for Image Generation Across Domains", + "text_level": 1, + "bbox": [ + 76, + 830, + 437, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To illustrate the proposed method, we first introduce in Fig. 2 a Structural Causal Model (SCM) [31] to model the underlying mechanism of object-centric image data genera", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/36189ee5f657509a3d3166fc27e3598d934330392b08df0dabc95deaac4903ce.jpg", + "image_caption": [ + "Figure 2. The causal graph of underlying object-centric image generation process across domains. $C$ : confounder; $D$ : domain; $S$ : object semantics; $H_d$ : (non-semantic) domain representation; $H_s$ : semantic representation; $X$ : image." + ], + "image_footnote": [], + "bbox": [ + 627, + 89, + 761, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion across diverse domains, which is consistent with previous works [18, 37] despite some slight modifications.", + "bbox": [ + 498, + 277, + 888, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, an SCM describes a directed acyclic graph composed of a set of endogenous variables $V$ and a set of exogenous variables $U$ . An endogenous variable is a variable whose value is determined by other variables. Exogenous variables correspond to unobserved influences, usually considered independent of each other. In Fig. 2, a set of endogenous variables are involved in the SCM, defined as $V = \\{D, S, H_d, H_s, X\\}$ . There is also an exogenous variable for each endogenous variable, e.g., $U_{H_s}$ that possibly corresponds to the object pose and influences the generation of $H_s$ . Here, we do not show them for simplicity, but only consider an important exogenous variable, the confounder $C$ . We denote the set of exogenous variables by $U = \\{C, U_D, U_S, U_{H_s}, U_{H_d}, U_X\\}$ .", + "bbox": [ + 498, + 308, + 890, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this SCM instance, $C$ denotes an unobservable confounder variable, which is an exogenous variable shown in gray. For instance, it can be some specific time or space. $D$ is a domain variable containing information varied with the domain. $S$ is a semantic variable, e.g., the category of the object. In our setting, $D$ and $S$ are observable endogenous variables shown in white. Correspondingly, $H_{d}$ is the domain representation and $H_{s}$ is the semantic representation, i.e., they indicate how $D$ and $S$ are displayed in the image space, respectively. Finally, $X$ is the image generated by the interaction of $H_{d}$ and $H_{s}$ .", + "bbox": [ + 498, + 520, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Spurious Correlation and OOD Robustness", + "text_level": 1, + "bbox": [ + 500, + 709, + 854, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, there exists a backdoor path between $H_{d}$ and $H_{s}$ , i.e. $H_{d} \\gets D \\gets C \\to S \\to H_{s}$ , making them spuriously correlated. Collecting data for a downstream task from a single source or environment may lead to strong spurious correlations between the semantic part $H_{s}$ and domain-relevant part $H_{d}$ . In other words, there can be strong selection bias [28]. For example, in a downstream insect classification dataset, the red admiral butterflies may all be captured sitting on flowers, as shown in Fig. 1. However, this spurious correlation does not necessarily hold in OOD data [6]. Moreover, in reality, what makes the label", + "bbox": [ + 500, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7ec26745c8d609b2ccac31706f7fede5671f97682d276e0160914ce5c51f80b6.jpg", + "image_caption": [ + "Figure 3. Conceptual comparison of pre-trained and fine-tuned models on what they depend on for predicting the semantic label of different samples. The robust pre-trained model is expected to depend only on $h_s$ , while the fine-tuned model may associate both $h_d$ and $h_s$ to the semantics $s$ . This may result in different predictions on counterfactual samples for the two models." + ], + "image_footnote": [], + "bbox": [ + 192, + 80, + 774, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "red admiral butterfly is the butterfly itself in the image space $(H_{s})$ rather than the flowers involved in $H_{d}$ . Therefore, the model is required to depend only on $H_{s}$ for the prediction of $S$ to attain OOD robustness.", + "bbox": [ + 75, + 422, + 468, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the reasons we have discussed earlier, even if a pre-trained large model has been endowed with reasonable knowledge to distinguish $H_{s}$ and $H_{d}$ (Fig. 3 (a)), the model is still susceptible to the spurious correlations during finetuning (Fig. 3 (d)). As a result, the transfer dramatically degenerates the pre-trained model's OOD robustness.", + "bbox": [ + 75, + 484, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From this perspective, a possible way to tackle the issue of robustness degradation is to break the spurious correlations in downstream data via specific manipulation, and explicitly requires the fine-tuning model to distinguish $H_{s}$ and $H_{d}$ following the pre-trained model. In the following sections, we propose a masking-based image manipulation method and indicate why it can be used to hinder the fine-tuning model from depending on $H_{d}$ .", + "bbox": [ + 75, + 577, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Masked Images as Counterfactual Samples", + "text_level": 1, + "bbox": [ + 76, + 709, + 444, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To break the spurious correlation in training images, we propose to mask out or replace certain regions of the images so that $H_{d}$ or $H_{s}$ is (partially) manipulated, which results in counterfactual samples. Formally, given an observational sample $x$ and the SCM $M$ , if we assign $d'$ to variable $D$ , the resultant counterfactual sample can be denoted as:", + "bbox": [ + 75, + 734, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {D = d ^ {\\prime}} (u) = x _ {M _ {D = d ^ {\\prime}}} (u) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 839, + 468, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{D = d'}$ is the SCM instance in which $d'$ is assigned to $D$ , and $u$ denotes the values of the exogenous variables.", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We illustrate how the pre-trained and fine-tuned models may process two kinds of counterfactual samples differently in Fig. 3. First, for counterfactual samples whose domain-relevant representations $h_d$ are replaced by some other context $p$ (i.e., $x_{H_d = p}$ in Fig. 3 (b,e)), the fine-tuned model tends to attribute its prediction on both $p$ and $h_s$ since it has learned the spurious correlations. Hence, its prediction can be misled by $p$ . Differently, the robust pre-trained model which can distinguish the semantic and non-semantic factors is not affected by $p$ . Second, for counterfactual samples whose semantic representations $h_s$ are replaced by $p$ (i.e., $x_{H_d = p}$ in Fig. 3 (c,f)), the fine-tuned model can still predict the original semantics $s$ from $h_d$ , while the pre-trained model cannot due to the missing semantic representations $h_s$ . In both cases, counterfactual samples may lead to different predictions between the two models.", + "bbox": [ + 496, + 422, + 893, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now the question is how to leverage these counterfactual samples to preserve the OOD robustness in fine-tuning. Directly training on these counterfactual samples with the labels of corresponding factual samples is unsuitable, since the original semantic information may be distorted. Given that the pre-trained models can have substantial power to capture semantic cues, their image-level feature representations usually contain rich semantic information. We thereby seek an alternative solution in which these samples are used in distillation. This way, the fine-tuning model learns to mimic the pre-trained model in feature representation. Formally, we denote the image encoders of the pre-trained and fine-tuning model by $\\hat{f}$ and $f$ , respectively, and denote the classification head of the fine-tuning model by $g$ . Then, the", + "bbox": [ + 496, + 689, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20304", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f836923693b44fc814c04943552bdc6b1bfb1112cf075d368541b2acc6ca4508.jpg", + "image_caption": [ + "Figure 4. Illustration of the mechanism of masking and refilling." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 468, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "overall training objective can be written as follows:", + "bbox": [ + 76, + 252, + 416, + 266 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} (g (f (x)), y) + \\beta \\mathcal {L} _ {\\mathrm {M S E}} (\\hat {f} (x _ {c f}), f (x _ {c f})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 275, + 468, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{CE}}$ is the cross-entropy loss, $\\mathcal{L}_{\\mathrm{MSE}}$ is the mean squared error, $x$ is the raw image, $y$ is the label for $x$ , $x_{cf}$ is the counterfactual image, and $\\beta$ is a constant factor.", + "bbox": [ + 76, + 300, + 468, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the remaining parts of this section, we will discuss the proper design of masked images and show that it is nontrivial to construct the counterfactual samples $x_{cf}$ that can be used to effectively improve the OOD robustness.", + "bbox": [ + 75, + 345, + 468, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Choosing the Masked Regions", + "text_level": 1, + "bbox": [ + 76, + 424, + 341, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To generate counterfactual samples, we need to select the masked regions in an image where the contents can be manipulated. Recent large-scale pre-trained models [19, 32] usually adopt a Vision Transformer-based architecture [12], dividing input images into a series of patches. Hence, we consider patch-based masking in this paper. A basic strategy is to randomly sample a proportion of patches (random-mask), which is adopted by recent pre-training methods like MAE [14]. However, random sampling is content-agnostic. Although both $H_{s}$ and $H_{d}$ may be manipulated in this scheme, it tends to leave some spurious correlations in the unmasked regions. It is found that an MAE model can achieve semantically plausible reconstruction from the randomly masked images with a high masking rate (e.g., 75%), suggesting that such random masking may preserve a significant amount of the information of the whole images. Therefore, randomly masked images may be less suitable for improving OOD robustness.", + "bbox": [ + 75, + 446, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alternatively, we investigate two content-based strategies that are complementary: (1) context-mask: masking the patches that are least relevant to the label (i.e., $H_{d}$ , usually the context); (2) object-mask: masking the patches that are most relevant to the label (i.e., $H_{s}$ , usually the main object). To measure the contribution of each image patch to the label, we generate the class activation map (CAM) [9, 46] based on the fine-tuning model. Then, image patches are separated into two groups by a constant threshold $t$ . In such a scheme, the semantic-relevant regions (where their activation values are larger than the threshold $t$ ) can coarsely serve as $H_{s}$ , and non-semantic regions as $H_{d}$ .", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Refilling the Masked Images", + "text_level": 1, + "bbox": [ + 500, + 90, + 751, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After choosing the patches to mask, simply dropping them following MAE [14] can lead to a counterfactual image, i.e., $x_{H_d = \\emptyset}(u)$ . However, we argue that this strategy (abbreviated as no-fill) can be insufficient for the construction of effective counterfactual samples for improving the robustness, and refilling the masked regions can tackle this issue.", + "bbox": [ + 498, + 114, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concretely, an effective counterfactual sample should cause contradictions between the pre-trained and fine-tuning model, so that the latter is regularized by the distillation loss in Eq. (2) to depend less on non-semantic parts of images for the prediction of semantics. However, masking without refilling may not construct such samples, as depicted in Fig. 4. Specifically, context-mask may not produce contradictions since both models can predict semantics from the object, so we need refilling to bring some conflicting context that disturbs the prediction of the fine-tuning model. Object-mask alone can cause contradictions in theory, as the fine-tuning model could still predict the original semantics from the context, while the pre-trained model could not. However, due to the imperfect masking in practice, the pre-trained model may still recognize the object from its unmasked parts. Then, refilling can further distort the original semantics to ensure contradictions. Hence, refilling can help to construct more effective counterfactual samples for the proposed fine-tuning approach.", + "bbox": [ + 496, + 205, + 892, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider two basic strategies to select the patches to fill: (1) single-fill: select the patches in the corresponding positions of a single image randomly sampled from the training batch; (2) multi-fill: for each patch to fill, independently sample a source image from the batch and select the patch in the corresponding position. All combinations of masking and refilling strategies are illustrated in Fig. 5.", + "bbox": [ + 496, + 492, + 890, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 609, + 632, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 635, + 581, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We focus on fine-tuning the pre-trained model on ImageNet [34], and evaluate the OOD robustness of the models on five datasets: ImageNet-V2 [33] is a new test set for ImageNet collected following the original protocol. ImageNet-R [15] consists of various renditions (e.g., art, cartoons) of 200 ImageNet classes; ImageNet-Sketch [39] contains sketches for each of the 1000 ImageNet classes; ObjectNet [5] is a test set with 113 overlapping classes with ImageNet, which is collected to show objects from new viewpoints on new backgrounds; ImageNet-A [16] contains natural images that are misclassified by models trained on ImageNet, covering 200 ImageNet classes.", + "bbox": [ + 496, + 657, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. Following [43], we use the top-1 accuracy as the metric of performance on ID and OOD data. For the five OOD datasets, we report the top-1 accuracy on each dataset, as well as the average OOD accuracy computed by", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20305", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5e66debb04a44932077755da89fd69b2accafdd4e4e57989112de546e9e6b17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 119, + 89, + 212, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16528456bc8117d2fc9a73451916eeda68b30190defff6d38ba2c68bebeb976f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 90, + 316, + 180 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4056c59f0ad36fd0a33d28cd0e885e5fb9a58933883a6ca5880f30d7cbed3797.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 90, + 421, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ed7675d8a9636bb18e5fe93ab0ecdb62775ab2febadd1b277553818b3a418f18.jpg", + "image_caption": [ + "random-mask single-fill", + "random-mask multi-fill" + ], + "image_footnote": [], + "bbox": [ + 120, + 199, + 212, + 270 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be36d09ebd18212dc3bb47c4524360d515b8ce1e83dd9350b214f212a6c85456.jpg", + "image_caption": [ + "context-mask single-fill" + ], + "image_footnote": [], + "bbox": [ + 223, + 199, + 316, + 270 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a68287cb18062e8b67bf016cf4ec9e83f709e4b705fb181c86232e1dd06e6ee9.jpg", + "image_caption": [ + "object-mask single-fill", + "object-mask multi-fill" + ], + "image_footnote": [], + "bbox": [ + 328, + 199, + 421, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2cd7f4be29dc5618e97ee13073f325031381753103e1317d5ecd537cdc84f1d.jpg", + "image_caption": [ + "Figure 5. Illustration of different masking and refilling strategies. The raw (factual) image is displayed in Fig. 4." + ], + "image_footnote": [], + "bbox": [ + 120, + 291, + 212, + 362 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4065a590ae2a407c5b55176930670a65697b92dfd6fdbdbe5dd87ab44f121465.jpg", + "image_caption": [ + "context-mask multi-fill" + ], + "image_footnote": [], + "bbox": [ + 223, + 291, + 316, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16ebef063fd467716171fbbf13ea5ed0bd2591db45c143b481e95ac1769e80f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 291, + 421, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "averaging the accuracy on the five datasets. For datasets covering a subset of the ImageNet classes, the top-1 prediction of a model is taken as the class with the highest probability among the subset (instead of all ImageNet classes).", + "bbox": [ + 75, + 411, + 467, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Unless otherwise specified, we use the ViT-B/32 model [12] pre-trained via CLIP [32]. The classification head of the zero-shot model is constructed from the pre-defined text prompts used by CLIP, and we adopt the implementation given by [43]. To obtain accurate CAM scores for image masking, we apply the method proposed in [9], which is designed for attention-based models, including ViT. For fine-tuning on ImageNet, we mainly follow the routine of WiSE-FT [43]. Specifically, we use the AdamW optimizer [27] with a batch size of 512, and finetune for 10 epochs. The learning rate is set to $3 \\times 10^{-5}$ for all parameters and follows a cosine-annealing schedule [26] with 500 warm-up steps. No data augmentation is applied apart from the necessary resizing and cropping, following the training of CLIP. We split out a validation set of 10240 samples from the ImageNet training set to perform early stopping and model selection based on the validation accuracy. $\\beta$ in Eq. (2) is fixed to be 30 for all experiments. Additional details are in the appendix.", + "bbox": [ + 75, + 472, + 467, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Masking and Refilling Strategies", + "text_level": 1, + "bbox": [ + 76, + 770, + 362, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we validate the effectiveness of using masked images for improving OOD robustness, and investigate the different masking and refilling strategies discussed in Sec. 3.2. Concretely, we consider three masking strategies: (1) random-mask: randomly select a fixed proportion of patches; (2) context-mask: masking the patches with CAM score lower than the threshold; (3)", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0cb293c28428178cc1e24b68b14c758fdc6c0a3c693d5794029c5b9fc76d49dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 92, + 874, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3ecd8847d9438167d8cb8efee1f5a69415feb65b8738b64b6c956d184624f2b.jpg", + "image_caption": [ + "Figure 6. ID and average OOD accuracy of our approach using different masking and refilling strategies. For each combination of masking and refilling strategies, we plot multiple results based on different settings of hyper-parameters (i.e., masking rate or CAM score threshold). Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 516, + 157, + 859, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "object-mask: masking the patches with CAM score higher than the threshold. The masking rate for random-mask is selected from $\\{0.25, 0.5, 0.75\\}$ , and the CAM score thresholds for the latter two strategies are selected from $\\{0.3, 0.4, 0.5, 0.6\\}$ . The refilling strategies include: (1) no-fill: baseline strategy that drops the masked patches without refilling; (2) single-fill: refill with patches from one other image; (3) multi-fill: refill with patches from multiple other images. We test all the combinations of masking and refilling strategies, and compare them with the baseline where no masking is applied (i.e., $x_{cf} = x$ in Eq. (2)).", + "bbox": [ + 496, + 443, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 6, we plot the average OOD accuracy against ID accuracy for each model trained with the above strategies. Besides, for each combination of masking and refilling strategies, we select the model with the highest validation accuracy, and report their results on each dataset in Tab. 1. Based on these results, we have the following conclusions. (1) Most of the combinations of masking and refilling strategies achieve better ID-OOD trade-off than the no-masking baseline, which suggests the effectiveness of image masking in our proposed approach. (2) Refilling masked images with patches from other images (i.e., single-fill or multi-fill) is better than solely dropping the masked patches. (3) Comparing the two proposed refilling strategies, single-fill generally results in better OOD accuracy, while multi-fill may yield better ID accuracy. (4) Comparing the masking strategies, object-mask is generally better than random-mask and context-mask in terms of OOD accuracy.", + "bbox": [ + 496, + 611, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The superiority of object-mask can be explained following our analysis in Sec. 3.2.2. Concretely, while both mask", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20306", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac3c7e8a2a7cfb3a52344f2ea209aaf4dcb36cdcfde1119262f39a38171ef948.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMaskingRefillingININ-V2IN-RIN-SketchObjectNetIN-AOOD avg.
Zero-shot [32]//63.455.969.342.344.531.448.7
Vanilla fine-tuning//75.964.757.039.839.520.044.2
Oursno masking/76.966.569.245.645.329.851.3
Oursrandom-maskno-fill77.566.966.445.746.530.851.2
context-mask77.867.466.745.645.930.051.1
object-mask77.767.167.646.246.831.551.9
Oursrandom-masksingle-fill77.667.169.046.447.833.452.7
context-mask77.266.968.846.547.832.952.6
object-mask77.567.169.746.948.033.853.1
Oursrandom-maskmulti-fill78.067.467.446.146.731.951.9
context-mask78.267.466.545.545.930.051.1
object-mask77.967.768.146.647.533.052.6
", + "bbox": [ + 81, + 88, + 888, + 275 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/129fa29191596f427493a07932d70e8aa37484a48b23476121ca580a7299610f.jpg", + "table_caption": [ + "Table 1. Comparison of different masking and refilling strategies. Accuracy on ImageNet (IN) and the five OOD datasets are reported. OOD avg. is the average OOD accuracy on the five OOD datasets. The best accuracy is bold-faced, and the second best accuracy is underlined. Results are averaged over three runs with different seeds." + ], + "table_footnote": [], + "table_body": "
Threshold0.70.60.50.40.30.2
Image MR0.060.100.150.250.390.55
Object MR0.170.230.330.480.650.80
IoU0.150.210.280.380.450.48
", + "bbox": [ + 81, + 339, + 465, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Validation of CAM-based object masking with different thresholds. Image masking rate (MR): masking rate of the whole image. Object masking rate (MR): masking rate concerning the object area. IoU: Intersection over Union regarding the object. Averaged over a subset of masked samples constructed in training.", + "bbox": [ + 75, + 404, + 468, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing strategies are theoretically valid, context-mask is more dependent on refilling, since it supposes that the refilled context can effectively lead to contradictions. This may not be satisfied by the proposed refilling strategies, as they are not aware of the content of the patches taken from other images. Hence, context-mask can be less effective in practice.", + "bbox": [ + 75, + 489, + 468, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, to verify that the superiority of our CAM-based object masking is due to the effective location of objects, we calculate the average masking rates of the objects during training. Specifically, we evaluate the object-mask and single-fill strategy with different CAM score thresholds. To obtain the masking rate of the object for an ImageNet image, we take the pixel-level mask produced by a pre-trained segmentation model [11] as the approximation of ground truth. In addition, we compute the masking rate of the whole image and the Intersection over Union (IoU) regarding the object. Details are provided in the appendix. As shown in Tab. 2, the object masking rate is significantly higher than the masking rate of the whole image, which suggests that the object-mask mainly masks the objects. Besides, as the threshold for masking decreases, both the object masking rate and IoU increase. Hence, the CAM-based object masking is empirically sound.", + "bbox": [ + 73, + 580, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with Existing Approaches", + "text_level": 1, + "bbox": [ + 76, + 847, + 410, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with three existing approaches, namely LP-FT [23], WiSE-FT [43] and Model Soup [42],", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which also aim to improve both ID and OOD accuracy of fine-tuning. For WiSE-FT, we report the results with the default mixing coefficient, i.e., $\\alpha = 0.5$ . For Model Soup, we take the uniform soup as the representative method, which achieves the best average OOD accuracy on CLIP fine-tuning as reported in [42]. For our approach, we consider both refilling strategies, and adopt the object masking with the CAM score threshold that yields the highest validation accuracy. The results are shown in Tab. 3. First, our approach surpasses the previous approaches in terms of average OOD accuracy. Second, our approach achieves better ID accuracy than other approaches except for Model Soup. Note that Model Soup takes the ensemble of many finetuned models with different hyper-parameters, including those trained with strong data augmentation. Conversely, we do not dive into hyper-parameters tuning and do not use data augmentation in our experiments. Third, compared with WiSE-FT, our approach achieves superior performance on ObjectNet and ImageNet-A, but is inferior on ImageNet-R and ImageNet-Sketch, which suggests that the two approaches may improve different perspectives of robustness. This motivates us to consider the integration of WiSE-FT.", + "bbox": [ + 496, + 343, + 890, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Integrating WiSE-FT", + "text_level": 1, + "bbox": [ + 500, + 693, + 700, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "WiSE-FT ensemble the weights of the zero-shot model and the vanilla fine-tuned model, and the ensemble model may achieve better accuracy on both ID and OOD data. Concretely, the weight ensemble is computed by $\\theta_{e} = (1 - \\alpha)\\cdot \\theta_{0} + \\alpha \\theta_{1}$ , where $\\theta_0$ is the weights of the zero-shot model, and $\\theta_{1}$ is the weights of the fine-tuned model. This weight-space ensemble method is agnostic of the finetuning method in form. Therefore, we may expect that the ensemble of the zero-shot model and the model fine-tuned with our method can achieve even better results. However, as shown in Fig. 7, such direct integration with WiSE-FT may not improve ID and OOD accuracy simultaneously.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20307", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/52e73d08cf9ca89a8dd8a2df318242601538c2bcac4cb3e8268e13c613fe3bca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMethodININ-V2IN-RIN-SketchObjectNetIN-AOOD avg.
CLIP ViT-B/32Zero-shot [32]63.455.969.342.344.531.448.7
Vanilla fine-tuning75.964.757.039.839.520.044.2
WiSE-FT†[43]76.666.670.247.146.331.952.4
Uniform soup‡[42]80.068.666.647.746.129.251.6
Ours (multi-fill)77.967.768.146.647.533.052.6
Ours (single-fill)77.567.169.746.948.033.853.1
CLIP ViT-B/16Zero-shot [32]68.361.977.648.354.050.158.4
Vanilla fine-tuning80.770.464.045.149.135.252.8
LP-FT [23]81.771.672.948.4/49.1/
WiSE-FT [43]81.772.878.753.957.352.263.0
Ours (multi-fill)82.573.476.452.756.852.062.3
Ours (single-fill)82.473.478.153.457.953.563.3
", + "bbox": [ + 81, + 88, + 888, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Accuracy of different methods for fine-tuning CLIP models on ImageNet (IN). OOD avg. is the average OOD accuracy on the five OOD datasets. The best accuracy is bold-faced, and the second best accuracy is underlined. Our results are averaged over three runs with different seeds. $\\dagger$ : our implementation. $\\ddagger$ : our evaluation on the official model.)", + "bbox": [ + 75, + 284, + 892, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/842de7f3f96a027dc579f70fbc9e5d41cfaa5e66e4c71126c2952eddaa8474c6.jpg", + "image_caption": [ + "Figure 7. Integration of WiSE-FT and our approach. WiSE-Ours: weight-space ensembles of the zero-shot model and our fine-tuned model. WiSE-KD: knowledge distillation from the WiSE-FT model ( $\\alpha = 0.5$ ). Vanilla FT: vanilla fine-tuning. For WiSE-FT and WiSE-Ours, each point corresponds to a value of $\\alpha \\in [0,1]$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 347, + 447, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, only for $\\alpha$ close to 1, the resulting model may achieve better average OOD accuracy, but still at the cost of lower ID accuracy. This is different from the observations of WiSE-FT that using a medium value of $\\alpha$ (e.g., 0.5) usually yields near-optimal results and surpasses the fine-tuned model on both ID and OOD accuracy [43]. It is suggested that our proposed method may produce a significantly different model as compared with vanilla fine-tuning. In other words, our fine-tuned weights may leave the basin of the loss landscape where the weights of most vanilla fine-tuned models and the pre-trained model lie in [30,42,43].", + "bbox": [ + 75, + 702, + 468, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another way to integrate WiSE-FT into our method is to utilize the ensemble model produced by WiSE-FT to guide", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the fine-tuning model via knowledge distillation (KD) [17]. For brevity, we name this method \"WiSE-KD\". Specifically, we consider using the WiSE-FT model as another teacher model, and add the vanilla knowledge distillation loss [17] to our training objective Eq. (2). Details are presented in the appendix. As presented in Fig. 7, applying WiSE-KD to our approach slightly improves both ID and OOD accuracy. As a comparison, while applying WiSE-KD to vanilla fin-tuning results in significantly better performance, it is still inferior to our approach, especially for OOD accuracy. This again validates the effectiveness of our approach in improving the OOD robustness in fine-tuning.", + "bbox": [ + 496, + 347, + 890, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 498, + 544, + 750, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we analyze the issue of robustness degradation in fine-tuning from a causal perspective, and find that masked images can be effective counterfactual samples for breaking the spurious correlation between semantic and non-semantic factors and improving the OOD robustness of fine-tuning models. Experiments suggest that our approach surpasses previous methods on the OOD performance with competitive ID performance.", + "bbox": [ + 496, + 571, + 890, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. As stated in Sec. 4.2, our refilling strategies are unaware of the content of the patches to fill, which limits the effectiveness of the resulting counterfactual samples. Besides, the feature-based distillation may not regularize the learning of the classification head of the fine-tuning model. Future works can devote to better refilling and distillation methods conforming to the causal modeling.", + "bbox": [ + 496, + 691, + 890, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 814, + 658, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by NSFC (No. U21A20470), National Key R&D Program of China (No. 2021ZD0111601), and the Science and Technology Program of Guangzhou, China (No. 202201011550).", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20308", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kartik Ahuja, Karthikeyan Shanmugam, Kush Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning, pages 145-155. PMLR, 2020. 2", + "[2] Martin Arjovsky. Out of distribution generalization in machine learning. PhD thesis, New York University, 2020. 2", + "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2", + "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 2", + "[5] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 1, 5", + "[6] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. 3", + "[7] Yoshua Bengio, Tristan Deleu, Nasim Rahaman, Nan Rosemary Ke, Sebastien Lachapelle, Olexa Bilaniuk, Anirudh Goyal, and Christopher Pal. A meta-transfer objective for learning to disentangle causal mechanisms. In International Conference on Learning Representations, 2019. 2", + "[8] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1, 2", + "[9] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 2, 5, 6", + "[10] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800-10809, 2020. 2", + "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 7", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 5, 6", + "[13] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5078-5088, 2022. 2", + "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 5", + "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 1, 5", + "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 1, 5", + "[17] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 8", + "[18] Maximilian Ilse, Jakub M Tomczak, and Patrick Forre. Selecting data augmentation for simulating interventions. In International Conference on Machine Learning, pages 4555-4562. PMLR, 2021. 2, 3", + "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2, 5", + "[20] Nan Rosemary Ke, Aniket Didolkar, Sarthak Mittal, Anirudh Goyal, Guillaume Lajoie, Stefan Bauer, Danilo Rezende, Yoshua Bengio, Michael Mozer, and Christopher Pal. Systematic evaluation of causal discovery in visual model based reinforcement learning. arXiv preprint arXiv:2107.00848, 2021. 2, 3", + "[21] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 2", + "[22] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In International Conference on Machine Learning, pages 5815-5826. PMLR, 2021. 2", + "[23] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In International Conference on Learning Representations, 2021. 1, 2, 3, 7, 8", + "[24] Ananya Kumar, Aditi Raghunathan, Tengyu Ma, and Percy Liang. Calibrated ensembles: A simple way to mitigate id-ood accuracy tradeoffs. In NeurIPS 2021 Workshop on" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20309", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Distribution Shifts: Connecting Methods and Applications, 2021. 2", + "[25] Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. Advances in Neural Information Processing Systems, 34:6155-6170, 2021. 2", + "[26] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. 2016. 6", + "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6", + "[28] Chaochao Lu, Yuhuai Wu, José Miguel Hernández-Lobato, and Bernhard Schölkopf. Invariant causal representation learning for out-of-distribution generalization. In International Conference on Learning Representations, 2021. 2, 3", + "[29] John P Miller, Rohan Taori, Aditi Raghunathan, Shiori Sagawa, Pang Wei Koh, Vaishaal Shankar, Percy Liang, Yair Carmon, and Ludwig Schmidt. Accuracy on the line: on the strong correlation between out-of-distribution and indistribution generalization. In International Conference on Machine Learning, pages 7721-7735. PMLR, 2021. 1", + "[30] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512-523, 2020. 8", + "[31] Judea Pearl. Causality. Cambridge university press, 2009. 2, 3", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 7, 8", + "[33] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 1, 5", + "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 5", + "[35] Nino Scherrer, Anirudh Goyal, Stefan Bauer, Yoshua Bengio, and Nan Rosemary Ke. On the generalization and adaption performance of causal models. arXiv preprint arXiv:2206.04620, 2022. 2, 3", + "[36] Bernhard Scholkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the IEEE, 109:612-634, 2021. 2, 3", + "[37] Adarsh Subbaswamy, Peter Schulam, and Suchi Saria. Preventing failures due to dataset shift: Learning predictive models that transport. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 3118-3127. PMLR, 2019. 3", + "[38] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring ro" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "bustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33:18583-18599, 2020. 1", + "[39] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 5", + "[40] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 2", + "[41] Moritz Willig, Matej Zečević, Devendra Singh Dhami, and Kristian Kersting. Can foundation models talk causality? arXiv preprint arXiv:2206.10591, 2022. 1", + "[42] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 1, 2, 7, 8", + "[43] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 1, 2, 5, 6, 7, 8", + "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2", + "[45] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 3", + "[46] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 709 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20310", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_model.json b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5a8f814e73e84d8abc34cc1f3622caca168dfb82 --- /dev/null +++ b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "Masked Images Are Counterfactual Samples for Robust Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.181, + 0.273, + 0.198 + ], + "angle": 0, + "content": "Yao Xiao" + }, + { + "type": "text", + "bbox": [ + 0.311, + 0.182, + 0.393, + 0.199 + ], + "angle": 0, + "content": "Ziyi Tang" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.182, + 0.543, + 0.199 + ], + "angle": 0, + "content": "Pengxu Wei *" + }, + { + "type": "text", + "bbox": [ + 0.576, + 0.182, + 0.654, + 0.199 + ], + "angle": 0, + "content": "Cong Liu" + }, + { + "type": "text", + "bbox": [ + 0.693, + 0.182, + 0.776, + 0.199 + ], + "angle": 0, + "content": "Liang Lin" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.2, + 0.577, + 0.218 + ], + "angle": 0, + "content": "Sun Yat-sen University" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.22, + 0.473, + 0.234 + ], + "angle": 0, + "content": "{xiaoy99,tangzy27}@mail2.sysu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.22, + 0.816, + 0.234 + ], + "angle": 0, + "content": "{weipx3, liucong3}@mail.sysu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.238, + 0.56, + 0.251 + ], + "angle": 0, + "content": "linliang@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.473, + 0.668 + ], + "angle": 0, + "content": "Deep learning models are challenged by the distribution shift between the training data and test data. Recently, the large models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts. However, fine-tuning these models can lead to a trade-off between in-distribution (ID) performance and out-of-distribution (OOD) robustness. Existing methods for tackling this trade-off do not explicitly address the OOD robustness problem. In this paper, based on causal analysis of the aforementioned problems, we propose a novel fine-tuning method, which uses masked images as counterfactual samples that help improve the robustness of the fine-tuning model. Specifically, we mask either the semantics-related or semantics-unrelated patches of the images based on class activation map to break the spurious correlation, and refill the masked patches with patches from other images. The resulting counterfactual samples are used in feature-based distillation with the pre-trained model. Extensive experiments verify that regularizing the fine-tuning with the proposed masked images can achieve a better trade-off between ID and OOD performance, surpassing previous methods on the OOD performance. Our code is available at https://github.com/Coxy7/robust-finetuning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.684, + 0.21, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Deep learning has made impressive advances in various tasks on computer vision. Despite the remarkable performance achieved on benchmark datasets, deep models are often challenged by the distribution shift between the training data and test data [5, 15, 16, 33]. It is commonly assumed that the training and test samples follow the same distribution, which may not hold in real-world applications due to the unpredictable change of lighting conditions, viewpoints, backgrounds, etc. Although there are attempts to improve the robustness of deep models to the distribution shift (or OOD robustness), it is still rather under-explored [29, 38]." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.287, + 0.88, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.505, + 0.895, + 0.561 + ], + "angle": 0, + "content": "Figure 1. Illustration of our work. Vanilla fine-tuned models tend to learn spurious correlations that degrade the OOD robustness. To tackle this issue, our model learns from the pre-trained model on the counterfactual CAM-based masked images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.893, + 0.793 + ], + "angle": 0, + "content": "Recently, large-scale models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts [8, 19, 32]. Fine-tuning these pre-trained models on downstream tasks can be a promising approach to building robust models for different applications. However, it is found that while fine-tuning improves the performance on in-distribution (ID) data, it may reduce that on out-of-distribution (OOD) data [23, 43]. To tackle this trade-off, several methods [23, 42, 43] have been proposed to improve both ID and OOD performance in fine-tuning. However, they do not explicitly address the OOD robustness problem; instead, they implicitly preserve the robustness of the pre-trained model by constraining the distortion of pre-trained weights or using model ensembles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we revisit the issue of robustness degradation in fine-tuning from a causal perspective. A large-scale pre-trained model somewhat shows properties in causality and stays robust to OOD samples [41]. However, when fine-tuning on downstream tasks, a majority of the model parameters tend to be adjusted for the downstream task in fine-tuning due to the highly entangled representation of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20301" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "images, arguably destructive to the generalizable knowledge [20, 35]. In contrast, distribution shifts are usually sparse in the underlying causal factorization of the data generation process [7, 35]. In this low-dimensional case, if we know which variables vary with different data distributions in this factorization (i.e., the non-stationary factors), we can achieve the OOD robustness by simply excluding their influence on the final predictions of the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.367 + ], + "angle": 0, + "content": "Specifically, we consider a Structural Causal Model (SCM) [31] for the object-centric image generation process, as depicted in Fig. 2. In this SCM, images are generated according to a non-stationary domain-relevant factor and a stationary semantic factor. Between them is a spurious correlation caused by a hidden non-stationary confounder that influences how the domain-relevant factor changes with the semantic one. To retain the OOD robustness, a fin-tuning model should avoid mapping non-stationary domain-relevant features to the predicted semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.471, + 0.642 + ], + "angle": 0, + "content": "To this end, we propose to fine-tune the models with masked images, which serve as counterfactual samples breaking the spurious correlation. Training on these samples helps preserve the stationary and generalizable knowledge of the pre-trained model. Concretely, we either mask the patches that contribute most to the label (i.e., the main object) or mask those with the least contribution (e.g., the context), which can be implemented based on class activation map (CAM) [9,46]. Such image masking forms an manipulation of a factual image and produces a counterfactual sample. Since the pre-trained model can better disentangle invariant features across domains, we require the finetuning model to learn from the pre-trained model on these counterfactual samples, as illustrated in Fig. 1. Furthermore, we argue that simply dropping the masked patches may be insufficient to alleviate the risk of fitting spurious correlations, and we propose to refill the masked patches with those from other images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We study different combinations of masking strategies (e.g., masking the object or the context) and refilling strategies (e.g., filling with patches from single or multiple images). Experimental results suggest that most of the strategies are applicable to the construction of counterfactual samples that help improve the robustness in fine-tuning, while masking the object generally achieves the best robustness. Compared with existing methods [23, 42, 43] on fine-tuning CLIP [32] models, our approach achieves better average accuracy on various OOD datasets without relying on model ensembles or weight constraints. We also find that, taking the weight-space ensemble of the zero-shot model and our fine-tuned model following WiSE-FT [43], hardly improves the trade-off between ID and OOD accuracy, which contradicts previous observations and implies that our approach may produce essentially different models in comparison with conventional fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.651, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.842, + 0.132 + ], + "angle": 0, + "content": "2.1. Causal Perspective on OOD Robustness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Causality holds the promise of OOD robustness as it has the property of robustness under interventions that underlie distribution shifts [2, 35]. To learn the latent causal mechanism in visual data and attain OOD robustness, a branch of methods [21, 28, 36] target at learning domain-invariant causal representations and seek guarantees on generalization. Some approaches approximates this goal based on invariant risk minimization [1, 3], risk extrapolation [22], adaptation speed [7], or variational Bayes [25]. Ilse et al. [18] help models escape from spurious correlations by training on handcrafted analog interventional data. In this work, we aim to explore the preservation of the OOD robustness in pre-trained models equipped with generalizable knowledge during transfer, based on the analysis of an underlying data generation causal mechanism." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.796, + 0.392 + ], + "angle": 0, + "content": "2.2. ID-OD Trade-off in Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.717 + ], + "angle": 0, + "content": "While a series of large pre-trained models exhibit strong OOD robustness [8,19,32], empirical evidence suggests that fine-tuning these models on downstream data may degrade the robustness [23, 32, 43]. It is also theoretically justified that compared with linear probing, vanilla fine-tuning may increase the ID performance while decreasing the OOD performance due to feature distortion [23]. To tackle this ID-OOD trade-off, LP-FT [23] applies linear probing before fine-tuning; Calibrated Ensemble [24] and WiSE-FT [43] propose to take the ensemble of the pre-trained (zero-shot) model and vanilla fine-tuned model in output-space and weight-space, respectively. Model Soup [42] also improves the trade-off via weight-space ensembles, but it utilizes multiple models fine-tuned with different hyper-parameters. While these methods are empirically effective in improving both ID and OOD performance, they only implicitly preserve the robustness of the pre-trained model by constraining the deviation of the downstream model from the pre-trained one. Instead, our approach is based on explicit causal modeling of the OOD robustness problem and the use of counterfactual samples in fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.771, + 0.743 + ], + "angle": 0, + "content": "2.3. Learning with Masked Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Masked image modeling [4,14,40,44] has been proven to be an effective approach to vision model pre-training, where random sampling is a common strategy for masking. In our task, we find that learning with CAM-based masking can better address the spurious correlation in images. Similar to our approach, CSS [10] synthesizes counterfactual images for visual question answering (VQA) by masking out critical objects, for which the corresponding answer is changed to its negative; SwapMix [13] swaps the context objects in the feature space to reduce the reliance of VQA models on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20302" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "visual context. In this paper, we apply masking on images and refill the masked regions with content from other images, which is inspired by CutMix [45]. Besides, instead of manually constructing the labels for masked images like CSS or CutMix, we take the feature representations of the pre-trained model on the masked images as the supervision for the fine-tuning model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.213, + 0.168, + 0.229 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.239, + 0.468, + 0.27 + ], + "angle": 0, + "content": "3.1. Revisit OOD Robustness of Fine-tuned Models from a Causal Perspective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.468, + 0.413 + ], + "angle": 0, + "content": "When fine-tuning a pre-trained model on downstream tasks in real-world applications, it is usually reliable to generalize only to in-distribution (ID) data. There are also increasing demands for robustness to distribution shifts or generalization to out-of-distribution (OOD) data. However, for a pre-trained large model that has been trained on data from diverse distributions and obtained generalizable knowledge (e.g., CLIP [32]), fine-tuning can lead to dramatic degradation of robustness [23]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.468, + 0.64 + ], + "angle": 0, + "content": "This problem can be interpreted from a causal perspective. Suppose that we correctly factorize the data generation process in which variables are causally connected. Then, the discrepancy between ID and OOD (or sub-population) data generation process is usually sparse in this causal factorization [35, 36], e.g., few variables' priors change. The joint distribution that we can observe thereupon shifts. However, without modeling the causal hierarchy in data, the pre-trained models naturally learn dense and entangled features to represent the concepts in the causal factorization. Although distribution shifts are sparse in causal factorization, these models have to densely adapt their parameters in transfer to the downstream data [20]. In this case, the loss of generalizable knowledge is inevitable, which explains the significant decrease of robustness to distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.468, + 0.807 + ], + "angle": 0, + "content": "The core challenge in modeling the underlying causal factorization of visual data is the dense changes between ID and OOD data, despite the sparsity in causal factorization. For better robustness, a possible solution is to identify which visual features remain unchanged under distribution shifts, also known as stationary features [35]. Without OOD test data in hand to identify stationary features, we seek to model the most important stationary features as the semantic features in the downstream object-centric task, and reduce non-stationary features to domain-relevant features. Finally, we establish our causal model as in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.439, + 0.845 + ], + "angle": 0, + "content": "3.1.1 SCM for Image Generation Across Domains" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To illustrate the proposed method, we first introduce in Fig. 2 a Structural Causal Model (SCM) [31] to model the underlying mechanism of object-centric image data genera" + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.09, + 0.763, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.206, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Figure 2. The causal graph of underlying object-centric image generation process across domains. \\( C \\): confounder; \\( D \\): domain; \\( S \\): object semantics; \\( H_d \\): (non-semantic) domain representation; \\( H_s \\): semantic representation; \\( X \\): image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.278, + 0.89, + 0.308 + ], + "angle": 0, + "content": "tion across diverse domains, which is consistent with previous works [18, 37] despite some slight modifications." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.309, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Formally, an SCM describes a directed acyclic graph composed of a set of endogenous variables \\(V\\) and a set of exogenous variables \\(U\\). An endogenous variable is a variable whose value is determined by other variables. Exogenous variables correspond to unobserved influences, usually considered independent of each other. In Fig. 2, a set of endogenous variables are involved in the SCM, defined as \\(V = \\{D, S, H_d, H_s, X\\}\\). There is also an exogenous variable for each endogenous variable, e.g., \\(U_{H_s}\\) that possibly corresponds to the object pose and influences the generation of \\(H_s\\). Here, we do not show them for simplicity, but only consider an important exogenous variable, the confounder \\(C\\). We denote the set of exogenous variables by \\(U = \\{C, U_D, U_S, U_{H_s}, U_{H_d}, U_X\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.892, + 0.687 + ], + "angle": 0, + "content": "In this SCM instance, \\( C \\) denotes an unobservable confounder variable, which is an exogenous variable shown in gray. For instance, it can be some specific time or space. \\( D \\) is a domain variable containing information varied with the domain. \\( S \\) is a semantic variable, e.g., the category of the object. In our setting, \\( D \\) and \\( S \\) are observable endogenous variables shown in white. Correspondingly, \\( H_{d} \\) is the domain representation and \\( H_{s} \\) is the semantic representation, i.e., they indicate how \\( D \\) and \\( S \\) are displayed in the image space, respectively. Finally, \\( X \\) is the image generated by the interaction of \\( H_{d} \\) and \\( H_{s} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.855, + 0.725 + ], + "angle": 0, + "content": "3.1.2 Spurious Correlation and OOD Robustness" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As depicted in Fig. 2, there exists a backdoor path between \\( H_{d} \\) and \\( H_{s} \\), i.e. \\( H_{d} \\gets D \\gets C \\to S \\to H_{s} \\), making them spuriously correlated. Collecting data for a downstream task from a single source or environment may lead to strong spurious correlations between the semantic part \\( H_{s} \\) and domain-relevant part \\( H_{d} \\). In other words, there can be strong selection bias [28]. For example, in a downstream insect classification dataset, the red admiral butterflies may all be captured sitting on flowers, as shown in Fig. 1. However, this spurious correlation does not necessarily hold in OOD data [6]. Moreover, in reality, what makes the label" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.193, + 0.082, + 0.775, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.355, + 0.895, + 0.398 + ], + "angle": 0, + "content": "Figure 3. Conceptual comparison of pre-trained and fine-tuned models on what they depend on for predicting the semantic label of different samples. The robust pre-trained model is expected to depend only on \\( h_s \\), while the fine-tuned model may associate both \\( h_d \\) and \\( h_s \\) to the semantics \\( s \\). This may result in different predictions on counterfactual samples for the two models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.483 + ], + "angle": 0, + "content": "red admiral butterfly is the butterfly itself in the image space \\((H_{s})\\) rather than the flowers involved in \\(H_{d}\\). Therefore, the model is required to depend only on \\(H_{s}\\) for the prediction of \\(S\\) to attain OOD robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.576 + ], + "angle": 0, + "content": "For the reasons we have discussed earlier, even if a pre-trained large model has been endowed with reasonable knowledge to distinguish \\( H_{s} \\) and \\( H_{d} \\) (Fig. 3 (a)), the model is still susceptible to the spurious correlations during finetuning (Fig. 3 (d)). As a result, the transfer dramatically degenerates the pre-trained model's OOD robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.699 + ], + "angle": 0, + "content": "From this perspective, a possible way to tackle the issue of robustness degradation is to break the spurious correlations in downstream data via specific manipulation, and explicitly requires the fine-tuning model to distinguish \\( H_{s} \\) and \\( H_{d} \\) following the pre-trained model. In the following sections, we propose a masking-based image manipulation method and indicate why it can be used to hinder the fine-tuning model from depending on \\( H_{d} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.445, + 0.728 + ], + "angle": 0, + "content": "3.2. Masked Images as Counterfactual Samples" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.827 + ], + "angle": 0, + "content": "To break the spurious correlation in training images, we propose to mask out or replace certain regions of the images so that \\( H_{d} \\) or \\( H_{s} \\) is (partially) manipulated, which results in counterfactual samples. Formally, given an observational sample \\( x \\) and the SCM \\( M \\), if we assign \\( d' \\) to variable \\( D \\), the resultant counterfactual sample can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.84, + 0.47, + 0.86 + ], + "angle": 0, + "content": "\\[\nx _ {D = d ^ {\\prime}} (u) = x _ {M _ {D = d ^ {\\prime}}} (u) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\( M_{D = d'} \\) is the SCM instance in which \\( d' \\) is assigned to \\( D \\), and \\( u \\) denotes the values of the exogenous variables." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.895, + 0.667 + ], + "angle": 0, + "content": "We illustrate how the pre-trained and fine-tuned models may process two kinds of counterfactual samples differently in Fig. 3. First, for counterfactual samples whose domain-relevant representations \\( h_d \\) are replaced by some other context \\( p \\) (i.e., \\( x_{H_d = p} \\) in Fig. 3 (b,e)), the fine-tuned model tends to attribute its prediction on both \\( p \\) and \\( h_s \\) since it has learned the spurious correlations. Hence, its prediction can be misled by \\( p \\). Differently, the robust pre-trained model which can distinguish the semantic and non-semantic factors is not affected by \\( p \\). Second, for counterfactual samples whose semantic representations \\( h_s \\) are replaced by \\( p \\) (i.e., \\( x_{H_d = p} \\) in Fig. 3 (c,f)), the fine-tuned model can still predict the original semantics \\( s \\) from \\( h_d \\), while the pre-trained model cannot due to the missing semantic representations \\( h_s \\). In both cases, counterfactual samples may lead to different predictions between the two models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Now the question is how to leverage these counterfactual samples to preserve the OOD robustness in fine-tuning. Directly training on these counterfactual samples with the labels of corresponding factual samples is unsuitable, since the original semantic information may be distorted. Given that the pre-trained models can have substantial power to capture semantic cues, their image-level feature representations usually contain rich semantic information. We thereby seek an alternative solution in which these samples are used in distillation. This way, the fine-tuning model learns to mimic the pre-trained model in feature representation. Formally, we denote the image encoders of the pre-trained and fine-tuning model by \\(\\hat{f}\\) and \\(f\\), respectively, and denote the classification head of the fine-tuning model by \\(g\\). Then, the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20304" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.47, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.227, + 0.465, + 0.241 + ], + "angle": 0, + "content": "Figure 4. Illustration of the mechanism of masking and refilling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.253, + 0.417, + 0.267 + ], + "angle": 0, + "content": "overall training objective can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.276, + 0.469, + 0.294 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} (g (f (x)), y) + \\beta \\mathcal {L} _ {\\mathrm {M S E}} (\\hat {f} (x _ {c f}), f (x _ {c f})), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.469, + 0.346 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{CE}}\\) is the cross-entropy loss, \\(\\mathcal{L}_{\\mathrm{MSE}}\\) is the mean squared error, \\(x\\) is the raw image, \\(y\\) is the label for \\(x\\), \\(x_{cf}\\) is the counterfactual image, and \\(\\beta\\) is a constant factor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.469, + 0.406 + ], + "angle": 0, + "content": "In the remaining parts of this section, we will discuss the proper design of masked images and show that it is nontrivial to construct the counterfactual samples \\( x_{cf} \\) that can be used to effectively improve the OOD robustness." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.342, + 0.439 + ], + "angle": 0, + "content": "3.2.1 Choosing the Masked Regions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.719 + ], + "angle": 0, + "content": "To generate counterfactual samples, we need to select the masked regions in an image where the contents can be manipulated. Recent large-scale pre-trained models [19, 32] usually adopt a Vision Transformer-based architecture [12], dividing input images into a series of patches. Hence, we consider patch-based masking in this paper. A basic strategy is to randomly sample a proportion of patches (random-mask), which is adopted by recent pre-training methods like MAE [14]. However, random sampling is content-agnostic. Although both \\( H_{s} \\) and \\( H_{d} \\) may be manipulated in this scheme, it tends to leave some spurious correlations in the unmasked regions. It is found that an MAE model can achieve semantically plausible reconstruction from the randomly masked images with a high masking rate (e.g., 75%), suggesting that such random masking may preserve a significant amount of the information of the whole images. Therefore, randomly masked images may be less suitable for improving OOD robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Alternatively, we investigate two content-based strategies that are complementary: (1) context-mask: masking the patches that are least relevant to the label (i.e., \\( H_{d} \\), usually the context); (2) object-mask: masking the patches that are most relevant to the label (i.e., \\( H_{s} \\), usually the main object). To measure the contribution of each image patch to the label, we generate the class activation map (CAM) [9, 46] based on the fine-tuning model. Then, image patches are separated into two groups by a constant threshold \\( t \\). In such a scheme, the semantic-relevant regions (where their activation values are larger than the threshold \\( t \\)) can coarsely serve as \\( H_{s} \\), and non-semantic regions as \\( H_{d} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.753, + 0.107 + ], + "angle": 0, + "content": "3.2.2 Refilling the Masked Images" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.206 + ], + "angle": 0, + "content": "After choosing the patches to mask, simply dropping them following MAE [14] can lead to a counterfactual image, i.e., \\( x_{H_d = \\emptyset}(u) \\). However, we argue that this strategy (abbreviated as no-fill) can be insufficient for the construction of effective counterfactual samples for improving the robustness, and refilling the masked regions can tackle this issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Concretely, an effective counterfactual sample should cause contradictions between the pre-trained and fine-tuning model, so that the latter is regularized by the distillation loss in Eq. (2) to depend less on non-semantic parts of images for the prediction of semantics. However, masking without refilling may not construct such samples, as depicted in Fig. 4. Specifically, context-mask may not produce contradictions since both models can predict semantics from the object, so we need refilling to bring some conflicting context that disturbs the prediction of the fine-tuning model. Object-mask alone can cause contradictions in theory, as the fine-tuning model could still predict the original semantics from the context, while the pre-trained model could not. However, due to the imperfect masking in practice, the pre-trained model may still recognize the object from its unmasked parts. Then, refilling can further distort the original semantics to ensure contradictions. Hence, refilling can help to construct more effective counterfactual samples for the proposed fine-tuning approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.6 + ], + "angle": 0, + "content": "We consider two basic strategies to select the patches to fill: (1) single-fill: select the patches in the corresponding positions of a single image randomly sampled from the training batch; (2) multi-fill: for each patch to fill, independently sample a source image from the batch and select the patch in the corresponding position. All combinations of masking and refilling strategies are illustrated in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.633, + 0.629 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.582, + 0.653 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Datasets. We focus on fine-tuning the pre-trained model on ImageNet [34], and evaluate the OOD robustness of the models on five datasets: ImageNet-V2 [33] is a new test set for ImageNet collected following the original protocol. ImageNet-R [15] consists of various renditions (e.g., art, cartoons) of 200 ImageNet classes; ImageNet-Sketch [39] contains sketches for each of the 1000 ImageNet classes; ObjectNet [5] is a test set with 113 overlapping classes with ImageNet, which is collected to show objects from new viewpoints on new backgrounds; ImageNet-A [16] contains natural images that are misclassified by models trained on ImageNet, covering 200 ImageNet classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation. Following [43], we use the top-1 accuracy as the metric of performance on ID and OOD data. For the five OOD datasets, we report the top-1 accuracy on each dataset, as well as the average OOD accuracy computed by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20305" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.214, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.184, + 0.2, + 0.199 + ], + "angle": 0, + "content": "random-mask single-fill" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.092, + 0.318, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.184, + 0.304, + 0.2 + ], + "angle": 0, + "content": "context-mask single-fill" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.091, + 0.422, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.183, + 0.406, + 0.2 + ], + "angle": 0, + "content": "object-mask single-fill" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.2, + 0.214, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.275, + 0.2, + 0.289 + ], + "angle": 0, + "content": "random-mask multi-fill" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.2, + 0.318, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.275, + 0.304, + 0.29 + ], + "angle": 0, + "content": "context-mask multi-fill" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.2, + 0.423, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.275, + 0.405, + 0.29 + ], + "angle": 0, + "content": "object-mask multi-fill" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.292, + 0.214, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.292, + 0.318, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.292, + 0.423, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.371, + 0.468, + 0.4 + ], + "angle": 0, + "content": "Figure 5. Illustration of different masking and refilling strategies. The raw (factual) image is displayed in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.468, + 0.472 + ], + "angle": 0, + "content": "averaging the accuracy on the five datasets. For datasets covering a subset of the ImageNet classes, the top-1 prediction of a model is taken as the class with the highest probability among the subset (instead of all ImageNet classes)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.468, + 0.76 + ], + "angle": 0, + "content": "Implementation details. Unless otherwise specified, we use the ViT-B/32 model [12] pre-trained via CLIP [32]. The classification head of the zero-shot model is constructed from the pre-defined text prompts used by CLIP, and we adopt the implementation given by [43]. To obtain accurate CAM scores for image masking, we apply the method proposed in [9], which is designed for attention-based models, including ViT. For fine-tuning on ImageNet, we mainly follow the routine of WiSE-FT [43]. Specifically, we use the AdamW optimizer [27] with a batch size of 512, and finetune for 10 epochs. The learning rate is set to \\(3 \\times 10^{-5}\\) for all parameters and follows a cosine-annealing schedule [26] with 500 warm-up steps. No data augmentation is applied apart from the necessary resizing and cropping, following the training of CLIP. We split out a validation set of 10240 samples from the ImageNet training set to perform early stopping and model selection based on the validation accuracy. \\(\\beta\\) in Eq. (2) is fixed to be 30 for all experiments. Additional details are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.364, + 0.789 + ], + "angle": 0, + "content": "4.2. Masking and Refilling Strategies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this section, we validate the effectiveness of using masked images for improving OOD robustness, and investigate the different masking and refilling strategies discussed in Sec. 3.2. Concretely, we consider three masking strategies: (1) random-mask: randomly select a fixed proportion of patches; (2) context-mask: masking the patches with CAM score lower than the threshold; (3)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.875, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.158, + 0.861, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.359, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Figure 6. ID and average OOD accuracy of our approach using different masking and refilling strategies. For each combination of masking and refilling strategies, we plot multiple results based on different settings of hyper-parameters (i.e., masking rate or CAM score threshold). Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.61 + ], + "angle": 0, + "content": "object-mask: masking the patches with CAM score higher than the threshold. The masking rate for random-mask is selected from \\(\\{0.25, 0.5, 0.75\\}\\), and the CAM score thresholds for the latter two strategies are selected from \\(\\{0.3, 0.4, 0.5, 0.6\\}\\). The refilling strategies include: (1) no-fill: baseline strategy that drops the masked patches without refilling; (2) single-fill: refill with patches from one other image; (3) multi-fill: refill with patches from multiple other images. We test all the combinations of masking and refilling strategies, and compare them with the baseline where no masking is applied (i.e., \\(x_{cf} = x\\) in Eq. (2))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.868 + ], + "angle": 0, + "content": "In Fig. 6, we plot the average OOD accuracy against ID accuracy for each model trained with the above strategies. Besides, for each combination of masking and refilling strategies, we select the model with the highest validation accuracy, and report their results on each dataset in Tab. 1. Based on these results, we have the following conclusions. (1) Most of the combinations of masking and refilling strategies achieve better ID-OOD trade-off than the no-masking baseline, which suggests the effectiveness of image masking in our proposed approach. (2) Refilling masked images with patches from other images (i.e., single-fill or multi-fill) is better than solely dropping the masked patches. (3) Comparing the two proposed refilling strategies, single-fill generally results in better OOD accuracy, while multi-fill may yield better ID accuracy. (4) Comparing the masking strategies, object-mask is generally better than random-mask and context-mask in terms of OOD accuracy." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The superiority of object-mask can be explained following our analysis in Sec. 3.2.2. Concretely, while both mask" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20306" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.89, + 0.276 + ], + "angle": 0, + "content": "
MethodMaskingRefillingININ-V2IN-RIN-SketchObjectNetIN-AOOD avg.
Zero-shot [32]//63.455.969.342.344.531.448.7
Vanilla fine-tuning//75.964.757.039.839.520.044.2
Oursno masking/76.966.569.245.645.329.851.3
Oursrandom-maskno-fill77.566.966.445.746.530.851.2
context-mask77.867.466.745.645.930.051.1
object-mask77.767.167.646.246.831.551.9
Oursrandom-masksingle-fill77.667.169.046.447.833.452.7
context-mask77.266.968.846.547.832.952.6
object-mask77.567.169.746.948.033.853.1
Oursrandom-maskmulti-fill78.067.467.446.146.731.951.9
context-mask78.267.466.545.545.930.051.1
object-mask77.967.768.146.647.533.052.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.28, + 0.893, + 0.325 + ], + "angle": 0, + "content": "Table 1. Comparison of different masking and refilling strategies. Accuracy on ImageNet (IN) and the five OOD datasets are reported. OOD avg. is the average OOD accuracy on the five OOD datasets. The best accuracy is bold-faced, and the second best accuracy is underlined. Results are averaged over three runs with different seeds." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.34, + 0.466, + 0.4 + ], + "angle": 0, + "content": "
Threshold0.70.60.50.40.30.2
Image MR0.060.100.150.250.390.55
Object MR0.170.230.330.480.650.80
IoU0.150.210.280.380.450.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.405, + 0.47, + 0.477 + ], + "angle": 0, + "content": "Table 2. Validation of CAM-based object masking with different thresholds. Image masking rate (MR): masking rate of the whole image. Object masking rate (MR): masking rate concerning the object area. IoU: Intersection over Union regarding the object. Averaged over a subset of masked samples constructed in training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.469, + 0.58 + ], + "angle": 0, + "content": "ing strategies are theoretically valid, context-mask is more dependent on refilling, since it supposes that the refilled context can effectively lead to contradictions. This may not be satisfied by the proposed refilling strategies, as they are not aware of the content of the patches taken from other images. Hence, context-mask can be less effective in practice." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.581, + 0.469, + 0.837 + ], + "angle": 0, + "content": "In addition, to verify that the superiority of our CAM-based object masking is due to the effective location of objects, we calculate the average masking rates of the objects during training. Specifically, we evaluate the object-mask and single-fill strategy with different CAM score thresholds. To obtain the masking rate of the object for an ImageNet image, we take the pixel-level mask produced by a pre-trained segmentation model [11] as the approximation of ground truth. In addition, we compute the masking rate of the whole image and the Intersection over Union (IoU) regarding the object. Details are provided in the appendix. As shown in Tab. 2, the object masking rate is significantly higher than the masking rate of the whole image, which suggests that the object-mask mainly masks the objects. Besides, as the threshold for masking decreases, both the object masking rate and IoU increase. Hence, the CAM-based object masking is empirically sound." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.411, + 0.865 + ], + "angle": 0, + "content": "4.3. Comparison with Existing Approaches" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "We compare our method with three existing approaches, namely LP-FT [23], WiSE-FT [43] and Model Soup [42]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.678 + ], + "angle": 0, + "content": "which also aim to improve both ID and OOD accuracy of fine-tuning. For WiSE-FT, we report the results with the default mixing coefficient, i.e., \\(\\alpha = 0.5\\). For Model Soup, we take the uniform soup as the representative method, which achieves the best average OOD accuracy on CLIP fine-tuning as reported in [42]. For our approach, we consider both refilling strategies, and adopt the object masking with the CAM score threshold that yields the highest validation accuracy. The results are shown in Tab. 3. First, our approach surpasses the previous approaches in terms of average OOD accuracy. Second, our approach achieves better ID accuracy than other approaches except for Model Soup. Note that Model Soup takes the ensemble of many finetuned models with different hyper-parameters, including those trained with strong data augmentation. Conversely, we do not dive into hyper-parameters tuning and do not use data augmentation in our experiments. Third, compared with WiSE-FT, our approach achieves superior performance on ObjectNet and ImageNet-A, but is inferior on ImageNet-R and ImageNet-Sketch, which suggests that the two approaches may improve different perspectives of robustness. This motivates us to consider the integration of WiSE-FT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.702, + 0.71 + ], + "angle": 0, + "content": "4.4. Integrating WiSE-FT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "WiSE-FT ensemble the weights of the zero-shot model and the vanilla fine-tuned model, and the ensemble model may achieve better accuracy on both ID and OOD data. Concretely, the weight ensemble is computed by \\(\\theta_{e} = (1 - \\alpha)\\cdot \\theta_{0} + \\alpha \\theta_{1}\\), where \\(\\theta_0\\) is the weights of the zero-shot model, and \\(\\theta_{1}\\) is the weights of the fine-tuned model. This weight-space ensemble method is agnostic of the finetuning method in form. Therefore, we may expect that the ensemble of the zero-shot model and the model fine-tuned with our method can achieve even better results. However, as shown in Fig. 7, such direct integration with WiSE-FT may not improve ID and OOD accuracy simultaneously." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20307" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.89, + 0.281 + ], + "angle": 0, + "content": "
ModelMethodININ-V2IN-RIN-SketchObjectNetIN-AOOD avg.
CLIP ViT-B/32Zero-shot [32]63.455.969.342.344.531.448.7
Vanilla fine-tuning75.964.757.039.839.520.044.2
WiSE-FT†[43]76.666.670.247.146.331.952.4
Uniform soup‡[42]80.068.666.647.746.129.251.6
Ours (multi-fill)77.967.768.146.647.533.052.6
Ours (single-fill)77.567.169.746.948.033.853.1
CLIP ViT-B/16Zero-shot [32]68.361.977.648.354.050.158.4
Vanilla fine-tuning80.770.464.045.149.135.252.8
LP-FT [23]81.771.672.948.4/49.1/
WiSE-FT [43]81.772.878.753.957.352.263.0
Ours (multi-fill)82.573.476.452.756.852.062.3
Ours (single-fill)82.473.478.153.457.953.563.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.285, + 0.893, + 0.327 + ], + "angle": 0, + "content": "Table 3. Accuracy of different methods for fine-tuning CLIP models on ImageNet (IN). OOD avg. is the average OOD accuracy on the five OOD datasets. The best accuracy is bold-faced, and the second best accuracy is underlined. Our results are averaged over three runs with different seeds. \\(\\dagger\\): our implementation. \\(\\ddagger\\): our evaluation on the official model.)" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.348, + 0.449, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Figure 7. Integration of WiSE-FT and our approach. WiSE-Ours: weight-space ensembles of the zero-shot model and our fine-tuned model. WiSE-KD: knowledge distillation from the WiSE-FT model (\\(\\alpha = 0.5\\)). Vanilla FT: vanilla fine-tuning. For WiSE-FT and WiSE-Ours, each point corresponds to a value of \\(\\alpha \\in [0,1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Specifically, only for \\(\\alpha\\) close to 1, the resulting model may achieve better average OOD accuracy, but still at the cost of lower ID accuracy. This is different from the observations of WiSE-FT that using a medium value of \\(\\alpha\\) (e.g., 0.5) usually yields near-optimal results and surpasses the fine-tuned model on both ID and OOD accuracy [43]. It is suggested that our proposed method may produce a significantly different model as compared with vanilla fine-tuning. In other words, our fine-tuned weights may leave the basin of the loss landscape where the weights of most vanilla fine-tuned models and the pre-trained model lie in [30,42,43]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Another way to integrate WiSE-FT into our method is to utilize the ensemble model produced by WiSE-FT to guide" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.53 + ], + "angle": 0, + "content": "the fine-tuning model via knowledge distillation (KD) [17]. For brevity, we name this method \"WiSE-KD\". Specifically, we consider using the WiSE-FT model as another teacher model, and add the vanilla knowledge distillation loss [17] to our training objective Eq. (2). Details are presented in the appendix. As presented in Fig. 7, applying WiSE-KD to our approach slightly improves both ID and OOD accuracy. As a comparison, while applying WiSE-KD to vanilla fin-tuning results in significantly better performance, it is still inferior to our approach, especially for OOD accuracy. This again validates the effectiveness of our approach in improving the OOD robustness in fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.545, + 0.75, + 0.56 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.692 + ], + "angle": 0, + "content": "In this paper, we analyze the issue of robustness degradation in fine-tuning from a causal perspective, and find that masked images can be effective counterfactual samples for breaking the spurious correlation between semantic and non-semantic factors and improving the OOD robustness of fine-tuning models. Experiments suggest that our approach surpasses previous methods on the OOD performance with competitive ID performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.799 + ], + "angle": 0, + "content": "Limitations. As stated in Sec. 4.2, our refilling strategies are unaware of the content of the patches to fill, which limits the effectiveness of the resulting counterfactual samples. Besides, the feature-based distillation may not regularize the learning of the classification head of the fine-tuning model. Future works can devote to better refilling and distillation methods conforming to the causal modeling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.66, + 0.831 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by NSFC (No. U21A20470), National Key R&D Program of China (No. 2021ZD0111601), and the Science and Technology Program of Guangzhou, China (No. 202201011550)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20308" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Kartik Ahuja, Karthikeyan Shanmugam, Kush Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning, pages 145-155. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[2] Martin Arjovsky. Out of distribution generalization in machine learning. PhD thesis, New York University, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.469, + 0.243 + ], + "angle": 0, + "content": "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.29, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[5] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[6] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[7] Yoshua Bengio, Tristan Deleu, Nasim Rahaman, Nan Rosemary Ke, Sebastien Lachapelle, Olexa Bilaniuk, Anirudh Goyal, and Christopher Pal. A meta-transfer objective for learning to disentangle causal mechanisms. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[8] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[9] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[10] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800-10809, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5078-5088, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[18] Maximilian Ilse, Jakub M Tomczak, and Patrick Forre. Selecting data augmentation for simulating interventions. In International Conference on Machine Learning, pages 4555-4562. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[20] Nan Rosemary Ke, Aniket Didolkar, Sarthak Mittal, Anirudh Goyal, Guillaume Lajoie, Stefan Bauer, Danilo Rezende, Yoshua Bengio, Michael Mozer, and Christopher Pal. Systematic evaluation of causal discovery in visual model based reinforcement learning. arXiv preprint arXiv:2107.00848, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[21] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[22] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In International Conference on Machine Learning, pages 5815-5826. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[23] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In International Conference on Learning Representations, 2021. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Ananya Kumar, Aditi Raghunathan, Tengyu Ma, and Percy Liang. Calibrated ensembles: A simple way to mitigate id-ood accuracy tradeoffs. In NeurIPS 2021 Workshop on" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20309" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "Distribution Shifts: Connecting Methods and Applications, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[25] Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. Advances in Neural Information Processing Systems, 34:6155-6170, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[28] Chaochao Lu, Yuhuai Wu, José Miguel Hernández-Lobato, and Bernhard Schölkopf. Invariant causal representation learning for out-of-distribution generalization. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[29] John P Miller, Rohan Taori, Aditi Raghunathan, Shiori Sagawa, Pang Wei Koh, Vaishaal Shankar, Percy Liang, Yair Carmon, and Ludwig Schmidt. Accuracy on the line: on the strong correlation between out-of-distribution and indistribution generalization. In International Conference on Machine Learning, pages 7721-7735. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[30] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512-523, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.468, + 0.474 + ], + "angle": 0, + "content": "[31] Judea Pearl. Causality. Cambridge university press, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[33] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[35] Nino Scherrer, Anirudh Goyal, Stefan Bauer, Yoshua Bengio, and Nan Rosemary Ke. On the generalization and adaption performance of causal models. arXiv preprint arXiv:2206.04620, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[36] Bernhard Scholkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the IEEE, 109:612-634, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[37] Adarsh Subbaswamy, Peter Schulam, and Suchi Saria. Preventing failures due to dataset shift: Learning predictive models that transport. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 3118-3127. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring ro" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "bustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33:18583-18599, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[39] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[40] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[41] Moritz Willig, Matej Zečević, Devendra Singh Dhami, and Kristian Kersting. Can foundation models talk causality? arXiv preprint arXiv:2206.10591, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[42] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[43] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.573, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[45] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.892, + 0.71 + ], + "angle": 0, + "content": "[46] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20310" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_origin.pdf b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d89111de2fcda272ad12c844aa77b8977fd2c560 --- /dev/null +++ b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/7801f6b3-3ee6-46d7-a157-f81dc90da6f9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:832949058e4793eb5416d7c804a64e23e9d90b4fb9e171e056b6552c527a0402 +size 2260833 diff --git a/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/full.md b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4905d7f6a920699963cd24d8572eb1e818b82e6 --- /dev/null +++ b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/full.md @@ -0,0 +1,287 @@ +# Masked Images Are Counterfactual Samples for Robust Fine-tuning + +Yao Xiao + +Ziyi Tang + +Pengxu Wei * + +Cong Liu + +Liang Lin + +Sun Yat-sen University + +{xiaoy99,tangzy27}@mail2.sysu.edu.cn + +{weipx3, liucong3}@mail.sysu.edu.cn + +linliang@ieee.org + +# Abstract + +Deep learning models are challenged by the distribution shift between the training data and test data. Recently, the large models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts. However, fine-tuning these models can lead to a trade-off between in-distribution (ID) performance and out-of-distribution (OOD) robustness. Existing methods for tackling this trade-off do not explicitly address the OOD robustness problem. In this paper, based on causal analysis of the aforementioned problems, we propose a novel fine-tuning method, which uses masked images as counterfactual samples that help improve the robustness of the fine-tuning model. Specifically, we mask either the semantics-related or semantics-unrelated patches of the images based on class activation map to break the spurious correlation, and refill the masked patches with patches from other images. The resulting counterfactual samples are used in feature-based distillation with the pre-trained model. Extensive experiments verify that regularizing the fine-tuning with the proposed masked images can achieve a better trade-off between ID and OOD performance, surpassing previous methods on the OOD performance. Our code is available at https://github.com/Coxy7/robust-finetuning. + +# 1. Introduction + +Deep learning has made impressive advances in various tasks on computer vision. Despite the remarkable performance achieved on benchmark datasets, deep models are often challenged by the distribution shift between the training data and test data [5, 15, 16, 33]. It is commonly assumed that the training and test samples follow the same distribution, which may not hold in real-world applications due to the unpredictable change of lighting conditions, viewpoints, backgrounds, etc. Although there are attempts to improve the robustness of deep models to the distribution shift (or OOD robustness), it is still rather under-explored [29, 38]. + +![](images/8bba6db4d514f606b8f66ae39d5fae798849202cb53728a56987d9e001bdd4f4.jpg) +Figure 1. Illustration of our work. Vanilla fine-tuned models tend to learn spurious correlations that degrade the OOD robustness. To tackle this issue, our model learns from the pre-trained model on the counterfactual CAM-based masked images. + +Recently, large-scale models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts [8, 19, 32]. Fine-tuning these pre-trained models on downstream tasks can be a promising approach to building robust models for different applications. However, it is found that while fine-tuning improves the performance on in-distribution (ID) data, it may reduce that on out-of-distribution (OOD) data [23, 43]. To tackle this trade-off, several methods [23, 42, 43] have been proposed to improve both ID and OOD performance in fine-tuning. However, they do not explicitly address the OOD robustness problem; instead, they implicitly preserve the robustness of the pre-trained model by constraining the distortion of pre-trained weights or using model ensembles. + +In this paper, we revisit the issue of robustness degradation in fine-tuning from a causal perspective. A large-scale pre-trained model somewhat shows properties in causality and stays robust to OOD samples [41]. However, when fine-tuning on downstream tasks, a majority of the model parameters tend to be adjusted for the downstream task in fine-tuning due to the highly entangled representation of + +images, arguably destructive to the generalizable knowledge [20, 35]. In contrast, distribution shifts are usually sparse in the underlying causal factorization of the data generation process [7, 35]. In this low-dimensional case, if we know which variables vary with different data distributions in this factorization (i.e., the non-stationary factors), we can achieve the OOD robustness by simply excluding their influence on the final predictions of the model. + +Specifically, we consider a Structural Causal Model (SCM) [31] for the object-centric image generation process, as depicted in Fig. 2. In this SCM, images are generated according to a non-stationary domain-relevant factor and a stationary semantic factor. Between them is a spurious correlation caused by a hidden non-stationary confounder that influences how the domain-relevant factor changes with the semantic one. To retain the OOD robustness, a fin-tuning model should avoid mapping non-stationary domain-relevant features to the predicted semantics. + +To this end, we propose to fine-tune the models with masked images, which serve as counterfactual samples breaking the spurious correlation. Training on these samples helps preserve the stationary and generalizable knowledge of the pre-trained model. Concretely, we either mask the patches that contribute most to the label (i.e., the main object) or mask those with the least contribution (e.g., the context), which can be implemented based on class activation map (CAM) [9,46]. Such image masking forms an manipulation of a factual image and produces a counterfactual sample. Since the pre-trained model can better disentangle invariant features across domains, we require the finetuning model to learn from the pre-trained model on these counterfactual samples, as illustrated in Fig. 1. Furthermore, we argue that simply dropping the masked patches may be insufficient to alleviate the risk of fitting spurious correlations, and we propose to refill the masked patches with those from other images. + +We study different combinations of masking strategies (e.g., masking the object or the context) and refilling strategies (e.g., filling with patches from single or multiple images). Experimental results suggest that most of the strategies are applicable to the construction of counterfactual samples that help improve the robustness in fine-tuning, while masking the object generally achieves the best robustness. Compared with existing methods [23, 42, 43] on fine-tuning CLIP [32] models, our approach achieves better average accuracy on various OOD datasets without relying on model ensembles or weight constraints. We also find that, taking the weight-space ensemble of the zero-shot model and our fine-tuned model following WiSE-FT [43], hardly improves the trade-off between ID and OOD accuracy, which contradicts previous observations and implies that our approach may produce essentially different models in comparison with conventional fine-tuning. + +# 2. Related Works + +# 2.1. Causal Perspective on OOD Robustness + +Causality holds the promise of OOD robustness as it has the property of robustness under interventions that underlie distribution shifts [2, 35]. To learn the latent causal mechanism in visual data and attain OOD robustness, a branch of methods [21, 28, 36] target at learning domain-invariant causal representations and seek guarantees on generalization. Some approaches approximates this goal based on invariant risk minimization [1, 3], risk extrapolation [22], adaptation speed [7], or variational Bayes [25]. Ilse et al. [18] help models escape from spurious correlations by training on handcrafted analog interventional data. In this work, we aim to explore the preservation of the OOD robustness in pre-trained models equipped with generalizable knowledge during transfer, based on the analysis of an underlying data generation causal mechanism. + +# 2.2. ID-OD Trade-off in Fine-tuning + +While a series of large pre-trained models exhibit strong OOD robustness [8,19,32], empirical evidence suggests that fine-tuning these models on downstream data may degrade the robustness [23, 32, 43]. It is also theoretically justified that compared with linear probing, vanilla fine-tuning may increase the ID performance while decreasing the OOD performance due to feature distortion [23]. To tackle this ID-OOD trade-off, LP-FT [23] applies linear probing before fine-tuning; Calibrated Ensemble [24] and WiSE-FT [43] propose to take the ensemble of the pre-trained (zero-shot) model and vanilla fine-tuned model in output-space and weight-space, respectively. Model Soup [42] also improves the trade-off via weight-space ensembles, but it utilizes multiple models fine-tuned with different hyper-parameters. While these methods are empirically effective in improving both ID and OOD performance, they only implicitly preserve the robustness of the pre-trained model by constraining the deviation of the downstream model from the pre-trained one. Instead, our approach is based on explicit causal modeling of the OOD robustness problem and the use of counterfactual samples in fine-tuning. + +# 2.3. Learning with Masked Images + +Masked image modeling [4,14,40,44] has been proven to be an effective approach to vision model pre-training, where random sampling is a common strategy for masking. In our task, we find that learning with CAM-based masking can better address the spurious correlation in images. Similar to our approach, CSS [10] synthesizes counterfactual images for visual question answering (VQA) by masking out critical objects, for which the corresponding answer is changed to its negative; SwapMix [13] swaps the context objects in the feature space to reduce the reliance of VQA models on + +visual context. In this paper, we apply masking on images and refill the masked regions with content from other images, which is inspired by CutMix [45]. Besides, instead of manually constructing the labels for masked images like CSS or CutMix, we take the feature representations of the pre-trained model on the masked images as the supervision for the fine-tuning model. + +# 3. Method + +# 3.1. Revisit OOD Robustness of Fine-tuned Models from a Causal Perspective + +When fine-tuning a pre-trained model on downstream tasks in real-world applications, it is usually reliable to generalize only to in-distribution (ID) data. There are also increasing demands for robustness to distribution shifts or generalization to out-of-distribution (OOD) data. However, for a pre-trained large model that has been trained on data from diverse distributions and obtained generalizable knowledge (e.g., CLIP [32]), fine-tuning can lead to dramatic degradation of robustness [23]. + +This problem can be interpreted from a causal perspective. Suppose that we correctly factorize the data generation process in which variables are causally connected. Then, the discrepancy between ID and OOD (or sub-population) data generation process is usually sparse in this causal factorization [35, 36], e.g., few variables' priors change. The joint distribution that we can observe thereupon shifts. However, without modeling the causal hierarchy in data, the pre-trained models naturally learn dense and entangled features to represent the concepts in the causal factorization. Although distribution shifts are sparse in causal factorization, these models have to densely adapt their parameters in transfer to the downstream data [20]. In this case, the loss of generalizable knowledge is inevitable, which explains the significant decrease of robustness to distribution shifts. + +The core challenge in modeling the underlying causal factorization of visual data is the dense changes between ID and OOD data, despite the sparsity in causal factorization. For better robustness, a possible solution is to identify which visual features remain unchanged under distribution shifts, also known as stationary features [35]. Without OOD test data in hand to identify stationary features, we seek to model the most important stationary features as the semantic features in the downstream object-centric task, and reduce non-stationary features to domain-relevant features. Finally, we establish our causal model as in Fig. 2. + +# 3.1.1 SCM for Image Generation Across Domains + +To illustrate the proposed method, we first introduce in Fig. 2 a Structural Causal Model (SCM) [31] to model the underlying mechanism of object-centric image data genera + +![](images/36189ee5f657509a3d3166fc27e3598d934330392b08df0dabc95deaac4903ce.jpg) +Figure 2. The causal graph of underlying object-centric image generation process across domains. $C$ : confounder; $D$ : domain; $S$ : object semantics; $H_d$ : (non-semantic) domain representation; $H_s$ : semantic representation; $X$ : image. + +tion across diverse domains, which is consistent with previous works [18, 37] despite some slight modifications. + +Formally, an SCM describes a directed acyclic graph composed of a set of endogenous variables $V$ and a set of exogenous variables $U$ . An endogenous variable is a variable whose value is determined by other variables. Exogenous variables correspond to unobserved influences, usually considered independent of each other. In Fig. 2, a set of endogenous variables are involved in the SCM, defined as $V = \{D, S, H_d, H_s, X\}$ . There is also an exogenous variable for each endogenous variable, e.g., $U_{H_s}$ that possibly corresponds to the object pose and influences the generation of $H_s$ . Here, we do not show them for simplicity, but only consider an important exogenous variable, the confounder $C$ . We denote the set of exogenous variables by $U = \{C, U_D, U_S, U_{H_s}, U_{H_d}, U_X\}$ . + +In this SCM instance, $C$ denotes an unobservable confounder variable, which is an exogenous variable shown in gray. For instance, it can be some specific time or space. $D$ is a domain variable containing information varied with the domain. $S$ is a semantic variable, e.g., the category of the object. In our setting, $D$ and $S$ are observable endogenous variables shown in white. Correspondingly, $H_{d}$ is the domain representation and $H_{s}$ is the semantic representation, i.e., they indicate how $D$ and $S$ are displayed in the image space, respectively. Finally, $X$ is the image generated by the interaction of $H_{d}$ and $H_{s}$ . + +# 3.1.2 Spurious Correlation and OOD Robustness + +As depicted in Fig. 2, there exists a backdoor path between $H_{d}$ and $H_{s}$ , i.e. $H_{d} \gets D \gets C \to S \to H_{s}$ , making them spuriously correlated. Collecting data for a downstream task from a single source or environment may lead to strong spurious correlations between the semantic part $H_{s}$ and domain-relevant part $H_{d}$ . In other words, there can be strong selection bias [28]. For example, in a downstream insect classification dataset, the red admiral butterflies may all be captured sitting on flowers, as shown in Fig. 1. However, this spurious correlation does not necessarily hold in OOD data [6]. Moreover, in reality, what makes the label + +![](images/7ec26745c8d609b2ccac31706f7fede5671f97682d276e0160914ce5c51f80b6.jpg) +Figure 3. Conceptual comparison of pre-trained and fine-tuned models on what they depend on for predicting the semantic label of different samples. The robust pre-trained model is expected to depend only on $h_s$ , while the fine-tuned model may associate both $h_d$ and $h_s$ to the semantics $s$ . This may result in different predictions on counterfactual samples for the two models. + +red admiral butterfly is the butterfly itself in the image space $(H_{s})$ rather than the flowers involved in $H_{d}$ . Therefore, the model is required to depend only on $H_{s}$ for the prediction of $S$ to attain OOD robustness. + +For the reasons we have discussed earlier, even if a pre-trained large model has been endowed with reasonable knowledge to distinguish $H_{s}$ and $H_{d}$ (Fig. 3 (a)), the model is still susceptible to the spurious correlations during finetuning (Fig. 3 (d)). As a result, the transfer dramatically degenerates the pre-trained model's OOD robustness. + +From this perspective, a possible way to tackle the issue of robustness degradation is to break the spurious correlations in downstream data via specific manipulation, and explicitly requires the fine-tuning model to distinguish $H_{s}$ and $H_{d}$ following the pre-trained model. In the following sections, we propose a masking-based image manipulation method and indicate why it can be used to hinder the fine-tuning model from depending on $H_{d}$ . + +# 3.2. Masked Images as Counterfactual Samples + +To break the spurious correlation in training images, we propose to mask out or replace certain regions of the images so that $H_{d}$ or $H_{s}$ is (partially) manipulated, which results in counterfactual samples. Formally, given an observational sample $x$ and the SCM $M$ , if we assign $d'$ to variable $D$ , the resultant counterfactual sample can be denoted as: + +$$ +x _ {D = d ^ {\prime}} (u) = x _ {M _ {D = d ^ {\prime}}} (u) \tag {1} +$$ + +where $M_{D = d'}$ is the SCM instance in which $d'$ is assigned to $D$ , and $u$ denotes the values of the exogenous variables. + +We illustrate how the pre-trained and fine-tuned models may process two kinds of counterfactual samples differently in Fig. 3. First, for counterfactual samples whose domain-relevant representations $h_d$ are replaced by some other context $p$ (i.e., $x_{H_d = p}$ in Fig. 3 (b,e)), the fine-tuned model tends to attribute its prediction on both $p$ and $h_s$ since it has learned the spurious correlations. Hence, its prediction can be misled by $p$ . Differently, the robust pre-trained model which can distinguish the semantic and non-semantic factors is not affected by $p$ . Second, for counterfactual samples whose semantic representations $h_s$ are replaced by $p$ (i.e., $x_{H_d = p}$ in Fig. 3 (c,f)), the fine-tuned model can still predict the original semantics $s$ from $h_d$ , while the pre-trained model cannot due to the missing semantic representations $h_s$ . In both cases, counterfactual samples may lead to different predictions between the two models. + +Now the question is how to leverage these counterfactual samples to preserve the OOD robustness in fine-tuning. Directly training on these counterfactual samples with the labels of corresponding factual samples is unsuitable, since the original semantic information may be distorted. Given that the pre-trained models can have substantial power to capture semantic cues, their image-level feature representations usually contain rich semantic information. We thereby seek an alternative solution in which these samples are used in distillation. This way, the fine-tuning model learns to mimic the pre-trained model in feature representation. Formally, we denote the image encoders of the pre-trained and fine-tuning model by $\hat{f}$ and $f$ , respectively, and denote the classification head of the fine-tuning model by $g$ . Then, the + +![](images/f836923693b44fc814c04943552bdc6b1bfb1112cf075d368541b2acc6ca4508.jpg) +Figure 4. Illustration of the mechanism of masking and refilling. + +overall training objective can be written as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {C E}} (g (f (x)), y) + \beta \mathcal {L} _ {\mathrm {M S E}} (\hat {f} (x _ {c f}), f (x _ {c f})), \tag {2} +$$ + +where $\mathcal{L}_{\mathrm{CE}}$ is the cross-entropy loss, $\mathcal{L}_{\mathrm{MSE}}$ is the mean squared error, $x$ is the raw image, $y$ is the label for $x$ , $x_{cf}$ is the counterfactual image, and $\beta$ is a constant factor. + +In the remaining parts of this section, we will discuss the proper design of masked images and show that it is nontrivial to construct the counterfactual samples $x_{cf}$ that can be used to effectively improve the OOD robustness. + +# 3.2.1 Choosing the Masked Regions + +To generate counterfactual samples, we need to select the masked regions in an image where the contents can be manipulated. Recent large-scale pre-trained models [19, 32] usually adopt a Vision Transformer-based architecture [12], dividing input images into a series of patches. Hence, we consider patch-based masking in this paper. A basic strategy is to randomly sample a proportion of patches (random-mask), which is adopted by recent pre-training methods like MAE [14]. However, random sampling is content-agnostic. Although both $H_{s}$ and $H_{d}$ may be manipulated in this scheme, it tends to leave some spurious correlations in the unmasked regions. It is found that an MAE model can achieve semantically plausible reconstruction from the randomly masked images with a high masking rate (e.g., 75%), suggesting that such random masking may preserve a significant amount of the information of the whole images. Therefore, randomly masked images may be less suitable for improving OOD robustness. + +Alternatively, we investigate two content-based strategies that are complementary: (1) context-mask: masking the patches that are least relevant to the label (i.e., $H_{d}$ , usually the context); (2) object-mask: masking the patches that are most relevant to the label (i.e., $H_{s}$ , usually the main object). To measure the contribution of each image patch to the label, we generate the class activation map (CAM) [9, 46] based on the fine-tuning model. Then, image patches are separated into two groups by a constant threshold $t$ . In such a scheme, the semantic-relevant regions (where their activation values are larger than the threshold $t$ ) can coarsely serve as $H_{s}$ , and non-semantic regions as $H_{d}$ . + +# 3.2.2 Refilling the Masked Images + +After choosing the patches to mask, simply dropping them following MAE [14] can lead to a counterfactual image, i.e., $x_{H_d = \emptyset}(u)$ . However, we argue that this strategy (abbreviated as no-fill) can be insufficient for the construction of effective counterfactual samples for improving the robustness, and refilling the masked regions can tackle this issue. + +Concretely, an effective counterfactual sample should cause contradictions between the pre-trained and fine-tuning model, so that the latter is regularized by the distillation loss in Eq. (2) to depend less on non-semantic parts of images for the prediction of semantics. However, masking without refilling may not construct such samples, as depicted in Fig. 4. Specifically, context-mask may not produce contradictions since both models can predict semantics from the object, so we need refilling to bring some conflicting context that disturbs the prediction of the fine-tuning model. Object-mask alone can cause contradictions in theory, as the fine-tuning model could still predict the original semantics from the context, while the pre-trained model could not. However, due to the imperfect masking in practice, the pre-trained model may still recognize the object from its unmasked parts. Then, refilling can further distort the original semantics to ensure contradictions. Hence, refilling can help to construct more effective counterfactual samples for the proposed fine-tuning approach. + +We consider two basic strategies to select the patches to fill: (1) single-fill: select the patches in the corresponding positions of a single image randomly sampled from the training batch; (2) multi-fill: for each patch to fill, independently sample a source image from the batch and select the patch in the corresponding position. All combinations of masking and refilling strategies are illustrated in Fig. 5. + +# 4. Experiments + +# 4.1. Setup + +Datasets. We focus on fine-tuning the pre-trained model on ImageNet [34], and evaluate the OOD robustness of the models on five datasets: ImageNet-V2 [33] is a new test set for ImageNet collected following the original protocol. ImageNet-R [15] consists of various renditions (e.g., art, cartoons) of 200 ImageNet classes; ImageNet-Sketch [39] contains sketches for each of the 1000 ImageNet classes; ObjectNet [5] is a test set with 113 overlapping classes with ImageNet, which is collected to show objects from new viewpoints on new backgrounds; ImageNet-A [16] contains natural images that are misclassified by models trained on ImageNet, covering 200 ImageNet classes. + +Evaluation. Following [43], we use the top-1 accuracy as the metric of performance on ID and OOD data. For the five OOD datasets, we report the top-1 accuracy on each dataset, as well as the average OOD accuracy computed by + +![](images/a5e66debb04a44932077755da89fd69b2accafdd4e4e57989112de546e9e6b17.jpg) + +![](images/16528456bc8117d2fc9a73451916eeda68b30190defff6d38ba2c68bebeb976f.jpg) + +![](images/4056c59f0ad36fd0a33d28cd0e885e5fb9a58933883a6ca5880f30d7cbed3797.jpg) + +![](images/ed7675d8a9636bb18e5fe93ab0ecdb62775ab2febadd1b277553818b3a418f18.jpg) +random-mask single-fill +random-mask multi-fill + +![](images/be36d09ebd18212dc3bb47c4524360d515b8ce1e83dd9350b214f212a6c85456.jpg) +context-mask single-fill + +![](images/a68287cb18062e8b67bf016cf4ec9e83f709e4b705fb181c86232e1dd06e6ee9.jpg) +object-mask single-fill +object-mask multi-fill + +![](images/f2cd7f4be29dc5618e97ee13073f325031381753103e1317d5ecd537cdc84f1d.jpg) +Figure 5. Illustration of different masking and refilling strategies. The raw (factual) image is displayed in Fig. 4. + +![](images/4065a590ae2a407c5b55176930670a65697b92dfd6fdbdbe5dd87ab44f121465.jpg) +context-mask multi-fill + +![](images/16ebef063fd467716171fbbf13ea5ed0bd2591db45c143b481e95ac1769e80f7.jpg) + +averaging the accuracy on the five datasets. For datasets covering a subset of the ImageNet classes, the top-1 prediction of a model is taken as the class with the highest probability among the subset (instead of all ImageNet classes). + +Implementation details. Unless otherwise specified, we use the ViT-B/32 model [12] pre-trained via CLIP [32]. The classification head of the zero-shot model is constructed from the pre-defined text prompts used by CLIP, and we adopt the implementation given by [43]. To obtain accurate CAM scores for image masking, we apply the method proposed in [9], which is designed for attention-based models, including ViT. For fine-tuning on ImageNet, we mainly follow the routine of WiSE-FT [43]. Specifically, we use the AdamW optimizer [27] with a batch size of 512, and finetune for 10 epochs. The learning rate is set to $3 \times 10^{-5}$ for all parameters and follows a cosine-annealing schedule [26] with 500 warm-up steps. No data augmentation is applied apart from the necessary resizing and cropping, following the training of CLIP. We split out a validation set of 10240 samples from the ImageNet training set to perform early stopping and model selection based on the validation accuracy. $\beta$ in Eq. (2) is fixed to be 30 for all experiments. Additional details are in the appendix. + +# 4.2. Masking and Refilling Strategies + +In this section, we validate the effectiveness of using masked images for improving OOD robustness, and investigate the different masking and refilling strategies discussed in Sec. 3.2. Concretely, we consider three masking strategies: (1) random-mask: randomly select a fixed proportion of patches; (2) context-mask: masking the patches with CAM score lower than the threshold; (3) + +![](images/0cb293c28428178cc1e24b68b14c758fdc6c0a3c693d5794029c5b9fc76d49dc.jpg) + +![](images/c3ecd8847d9438167d8cb8efee1f5a69415feb65b8738b64b6c956d184624f2b.jpg) +Figure 6. ID and average OOD accuracy of our approach using different masking and refilling strategies. For each combination of masking and refilling strategies, we plot multiple results based on different settings of hyper-parameters (i.e., masking rate or CAM score threshold). Best viewed in color. + +object-mask: masking the patches with CAM score higher than the threshold. The masking rate for random-mask is selected from $\{0.25, 0.5, 0.75\}$ , and the CAM score thresholds for the latter two strategies are selected from $\{0.3, 0.4, 0.5, 0.6\}$ . The refilling strategies include: (1) no-fill: baseline strategy that drops the masked patches without refilling; (2) single-fill: refill with patches from one other image; (3) multi-fill: refill with patches from multiple other images. We test all the combinations of masking and refilling strategies, and compare them with the baseline where no masking is applied (i.e., $x_{cf} = x$ in Eq. (2)). + +In Fig. 6, we plot the average OOD accuracy against ID accuracy for each model trained with the above strategies. Besides, for each combination of masking and refilling strategies, we select the model with the highest validation accuracy, and report their results on each dataset in Tab. 1. Based on these results, we have the following conclusions. (1) Most of the combinations of masking and refilling strategies achieve better ID-OOD trade-off than the no-masking baseline, which suggests the effectiveness of image masking in our proposed approach. (2) Refilling masked images with patches from other images (i.e., single-fill or multi-fill) is better than solely dropping the masked patches. (3) Comparing the two proposed refilling strategies, single-fill generally results in better OOD accuracy, while multi-fill may yield better ID accuracy. (4) Comparing the masking strategies, object-mask is generally better than random-mask and context-mask in terms of OOD accuracy. + +The superiority of object-mask can be explained following our analysis in Sec. 3.2.2. Concretely, while both mask + +
MethodMaskingRefillingININ-V2IN-RIN-SketchObjectNetIN-AOOD avg.
Zero-shot [32]//63.455.969.342.344.531.448.7
Vanilla fine-tuning//75.964.757.039.839.520.044.2
Oursno masking/76.966.569.245.645.329.851.3
Oursrandom-maskno-fill77.566.966.445.746.530.851.2
context-mask77.867.466.745.645.930.051.1
object-mask77.767.167.646.246.831.551.9
Oursrandom-masksingle-fill77.667.169.046.447.833.452.7
context-mask77.266.968.846.547.832.952.6
object-mask77.567.169.746.948.033.853.1
Oursrandom-maskmulti-fill78.067.467.446.146.731.951.9
context-mask78.267.466.545.545.930.051.1
object-mask77.967.768.146.647.533.052.6
+ +Table 1. Comparison of different masking and refilling strategies. Accuracy on ImageNet (IN) and the five OOD datasets are reported. OOD avg. is the average OOD accuracy on the five OOD datasets. The best accuracy is bold-faced, and the second best accuracy is underlined. Results are averaged over three runs with different seeds. + +
Threshold0.70.60.50.40.30.2
Image MR0.060.100.150.250.390.55
Object MR0.170.230.330.480.650.80
IoU0.150.210.280.380.450.48
+ +Table 2. Validation of CAM-based object masking with different thresholds. Image masking rate (MR): masking rate of the whole image. Object masking rate (MR): masking rate concerning the object area. IoU: Intersection over Union regarding the object. Averaged over a subset of masked samples constructed in training. + +ing strategies are theoretically valid, context-mask is more dependent on refilling, since it supposes that the refilled context can effectively lead to contradictions. This may not be satisfied by the proposed refilling strategies, as they are not aware of the content of the patches taken from other images. Hence, context-mask can be less effective in practice. + +In addition, to verify that the superiority of our CAM-based object masking is due to the effective location of objects, we calculate the average masking rates of the objects during training. Specifically, we evaluate the object-mask and single-fill strategy with different CAM score thresholds. To obtain the masking rate of the object for an ImageNet image, we take the pixel-level mask produced by a pre-trained segmentation model [11] as the approximation of ground truth. In addition, we compute the masking rate of the whole image and the Intersection over Union (IoU) regarding the object. Details are provided in the appendix. As shown in Tab. 2, the object masking rate is significantly higher than the masking rate of the whole image, which suggests that the object-mask mainly masks the objects. Besides, as the threshold for masking decreases, both the object masking rate and IoU increase. Hence, the CAM-based object masking is empirically sound. + +# 4.3. Comparison with Existing Approaches + +We compare our method with three existing approaches, namely LP-FT [23], WiSE-FT [43] and Model Soup [42], + +which also aim to improve both ID and OOD accuracy of fine-tuning. For WiSE-FT, we report the results with the default mixing coefficient, i.e., $\alpha = 0.5$ . For Model Soup, we take the uniform soup as the representative method, which achieves the best average OOD accuracy on CLIP fine-tuning as reported in [42]. For our approach, we consider both refilling strategies, and adopt the object masking with the CAM score threshold that yields the highest validation accuracy. The results are shown in Tab. 3. First, our approach surpasses the previous approaches in terms of average OOD accuracy. Second, our approach achieves better ID accuracy than other approaches except for Model Soup. Note that Model Soup takes the ensemble of many finetuned models with different hyper-parameters, including those trained with strong data augmentation. Conversely, we do not dive into hyper-parameters tuning and do not use data augmentation in our experiments. Third, compared with WiSE-FT, our approach achieves superior performance on ObjectNet and ImageNet-A, but is inferior on ImageNet-R and ImageNet-Sketch, which suggests that the two approaches may improve different perspectives of robustness. This motivates us to consider the integration of WiSE-FT. + +# 4.4. Integrating WiSE-FT + +WiSE-FT ensemble the weights of the zero-shot model and the vanilla fine-tuned model, and the ensemble model may achieve better accuracy on both ID and OOD data. Concretely, the weight ensemble is computed by $\theta_{e} = (1 - \alpha)\cdot \theta_{0} + \alpha \theta_{1}$ , where $\theta_0$ is the weights of the zero-shot model, and $\theta_{1}$ is the weights of the fine-tuned model. This weight-space ensemble method is agnostic of the finetuning method in form. Therefore, we may expect that the ensemble of the zero-shot model and the model fine-tuned with our method can achieve even better results. However, as shown in Fig. 7, such direct integration with WiSE-FT may not improve ID and OOD accuracy simultaneously. + +
ModelMethodININ-V2IN-RIN-SketchObjectNetIN-AOOD avg.
CLIP ViT-B/32Zero-shot [32]63.455.969.342.344.531.448.7
Vanilla fine-tuning75.964.757.039.839.520.044.2
WiSE-FT†[43]76.666.670.247.146.331.952.4
Uniform soup‡[42]80.068.666.647.746.129.251.6
Ours (multi-fill)77.967.768.146.647.533.052.6
Ours (single-fill)77.567.169.746.948.033.853.1
CLIP ViT-B/16Zero-shot [32]68.361.977.648.354.050.158.4
Vanilla fine-tuning80.770.464.045.149.135.252.8
LP-FT [23]81.771.672.948.4/49.1/
WiSE-FT [43]81.772.878.753.957.352.263.0
Ours (multi-fill)82.573.476.452.756.852.062.3
Ours (single-fill)82.473.478.153.457.953.563.3
+ +Table 3. Accuracy of different methods for fine-tuning CLIP models on ImageNet (IN). OOD avg. is the average OOD accuracy on the five OOD datasets. The best accuracy is bold-faced, and the second best accuracy is underlined. Our results are averaged over three runs with different seeds. $\dagger$ : our implementation. $\ddagger$ : our evaluation on the official model.) + +![](images/842de7f3f96a027dc579f70fbc9e5d41cfaa5e66e4c71126c2952eddaa8474c6.jpg) +Figure 7. Integration of WiSE-FT and our approach. WiSE-Ours: weight-space ensembles of the zero-shot model and our fine-tuned model. WiSE-KD: knowledge distillation from the WiSE-FT model ( $\alpha = 0.5$ ). Vanilla FT: vanilla fine-tuning. For WiSE-FT and WiSE-Ours, each point corresponds to a value of $\alpha \in [0,1]$ . + +Specifically, only for $\alpha$ close to 1, the resulting model may achieve better average OOD accuracy, but still at the cost of lower ID accuracy. This is different from the observations of WiSE-FT that using a medium value of $\alpha$ (e.g., 0.5) usually yields near-optimal results and surpasses the fine-tuned model on both ID and OOD accuracy [43]. It is suggested that our proposed method may produce a significantly different model as compared with vanilla fine-tuning. In other words, our fine-tuned weights may leave the basin of the loss landscape where the weights of most vanilla fine-tuned models and the pre-trained model lie in [30,42,43]. + +Another way to integrate WiSE-FT into our method is to utilize the ensemble model produced by WiSE-FT to guide + +the fine-tuning model via knowledge distillation (KD) [17]. For brevity, we name this method "WiSE-KD". Specifically, we consider using the WiSE-FT model as another teacher model, and add the vanilla knowledge distillation loss [17] to our training objective Eq. (2). Details are presented in the appendix. As presented in Fig. 7, applying WiSE-KD to our approach slightly improves both ID and OOD accuracy. As a comparison, while applying WiSE-KD to vanilla fin-tuning results in significantly better performance, it is still inferior to our approach, especially for OOD accuracy. This again validates the effectiveness of our approach in improving the OOD robustness in fine-tuning. + +# 5. Conclusion and Limitation + +In this paper, we analyze the issue of robustness degradation in fine-tuning from a causal perspective, and find that masked images can be effective counterfactual samples for breaking the spurious correlation between semantic and non-semantic factors and improving the OOD robustness of fine-tuning models. Experiments suggest that our approach surpasses previous methods on the OOD performance with competitive ID performance. + +Limitations. As stated in Sec. 4.2, our refilling strategies are unaware of the content of the patches to fill, which limits the effectiveness of the resulting counterfactual samples. Besides, the feature-based distillation may not regularize the learning of the classification head of the fine-tuning model. Future works can devote to better refilling and distillation methods conforming to the causal modeling. + +# Acknowledgments + +This work was supported in part by NSFC (No. U21A20470), National Key R&D Program of China (No. 2021ZD0111601), and the Science and Technology Program of Guangzhou, China (No. 202201011550). + +# References + +[1] Kartik Ahuja, Karthikeyan Shanmugam, Kush Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning, pages 145-155. PMLR, 2020. 2 +[2] Martin Arjovsky. Out of distribution generalization in machine learning. PhD thesis, New York University, 2020. 2 +[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2 +[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 2 +[5] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 1, 5 +[6] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. 3 +[7] Yoshua Bengio, Tristan Deleu, Nasim Rahaman, Nan Rosemary Ke, Sebastien Lachapelle, Olexa Bilaniuk, Anirudh Goyal, and Christopher Pal. A meta-transfer objective for learning to disentangle causal mechanisms. In International Conference on Learning Representations, 2019. 2 +[8] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1, 2 +[9] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 2, 5, 6 +[10] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800-10809, 2020. 2 +[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 7 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 5, 6 +[13] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual + +question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5078-5088, 2022. 2 +[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 5 +[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 1, 5 +[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 1, 5 +[17] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 8 +[18] Maximilian Ilse, Jakub M Tomczak, and Patrick Forre. Selecting data augmentation for simulating interventions. In International Conference on Machine Learning, pages 4555-4562. PMLR, 2021. 2, 3 +[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2, 5 +[20] Nan Rosemary Ke, Aniket Didolkar, Sarthak Mittal, Anirudh Goyal, Guillaume Lajoie, Stefan Bauer, Danilo Rezende, Yoshua Bengio, Michael Mozer, and Christopher Pal. Systematic evaluation of causal discovery in visual model based reinforcement learning. arXiv preprint arXiv:2107.00848, 2021. 2, 3 +[21] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 2 +[22] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In International Conference on Machine Learning, pages 5815-5826. PMLR, 2021. 2 +[23] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In International Conference on Learning Representations, 2021. 1, 2, 3, 7, 8 +[24] Ananya Kumar, Aditi Raghunathan, Tengyu Ma, and Percy Liang. Calibrated ensembles: A simple way to mitigate id-ood accuracy tradeoffs. In NeurIPS 2021 Workshop on + +Distribution Shifts: Connecting Methods and Applications, 2021. 2 +[25] Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. Advances in Neural Information Processing Systems, 34:6155-6170, 2021. 2 +[26] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. 2016. 6 +[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6 +[28] Chaochao Lu, Yuhuai Wu, José Miguel Hernández-Lobato, and Bernhard Schölkopf. Invariant causal representation learning for out-of-distribution generalization. In International Conference on Learning Representations, 2021. 2, 3 +[29] John P Miller, Rohan Taori, Aditi Raghunathan, Shiori Sagawa, Pang Wei Koh, Vaishaal Shankar, Percy Liang, Yair Carmon, and Ludwig Schmidt. Accuracy on the line: on the strong correlation between out-of-distribution and indistribution generalization. In International Conference on Machine Learning, pages 7721-7735. PMLR, 2021. 1 +[30] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512-523, 2020. 8 +[31] Judea Pearl. Causality. Cambridge university press, 2009. 2, 3 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 7, 8 +[33] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 1, 5 +[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 5 +[35] Nino Scherrer, Anirudh Goyal, Stefan Bauer, Yoshua Bengio, and Nan Rosemary Ke. On the generalization and adaption performance of causal models. arXiv preprint arXiv:2206.04620, 2022. 2, 3 +[36] Bernhard Scholkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the IEEE, 109:612-634, 2021. 2, 3 +[37] Adarsh Subbaswamy, Peter Schulam, and Suchi Saria. Preventing failures due to dataset shift: Learning predictive models that transport. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 3118-3127. PMLR, 2019. 3 +[38] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring ro + +bustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33:18583-18599, 2020. 1 +[39] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 5 +[40] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 2 +[41] Moritz Willig, Matej Zečević, Devendra Singh Dhami, and Kristian Kersting. Can foundation models talk causality? arXiv preprint arXiv:2206.10591, 2022. 1 +[42] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 1, 2, 7, 8 +[43] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 1, 2, 5, 6, 7, 8 +[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2 +[45] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 3 +[46] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 5 \ No newline at end of file diff --git a/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/images.zip b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..34480a32bc542bda9ba01596412397e44530b186 --- /dev/null +++ b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4aef58997d78363df3bd77a06f24dea1e01caa2000afd7f81e9a5dc601837445 +size 463185 diff --git a/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/layout.json b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c7bff795a791e61b0080e8eb730594745c7cb48 --- /dev/null +++ b/2023/Masked Images Are Counterfactual Samples for Robust Fine-Tuning/layout.json @@ -0,0 +1,8497 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "type": "text", + "content": "Masked Images Are Counterfactual Samples for Robust Fine-tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 143, + 167, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 143, + 167, + 156 + ], + "spans": [ + { + "bbox": [ + 118, + 143, + 167, + 156 + ], + "type": "text", + "content": "Yao Xiao" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 144, + 240, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 144, + 240, + 157 + ], + "spans": [ + { + "bbox": [ + 190, + 144, + 240, + 157 + ], + "type": "text", + "content": "Ziyi Tang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 264, + 144, + 332, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 144, + 332, + 157 + ], + "spans": [ + { + "bbox": [ + 264, + 144, + 332, + 157 + ], + "type": "text", + "content": "Pengxu Wei *" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 352, + 144, + 400, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 144, + 400, + 157 + ], + "spans": [ + { + "bbox": [ + 352, + 144, + 400, + 157 + ], + "type": "text", + "content": "Cong Liu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 424, + 144, + 474, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 144, + 474, + 157 + ], + "spans": [ + { + "bbox": [ + 424, + 144, + 474, + 157 + ], + "type": "text", + "content": "Liang Lin" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 240, + 158, + 353, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 158, + 353, + 172 + ], + "spans": [ + { + "bbox": [ + 240, + 158, + 353, + 172 + ], + "type": "text", + "content": "Sun Yat-sen University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 174, + 289, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 174, + 289, + 185 + ], + "spans": [ + { + "bbox": [ + 94, + 174, + 289, + 185 + ], + "type": "text", + "content": "{xiaoy99,tangzy27}@mail2.sysu.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 174, + 499, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 174, + 499, + 185 + ], + "spans": [ + { + "bbox": [ + 310, + 174, + 499, + 185 + ], + "type": "text", + "content": "{weipx3, liucong3}@mail.sysu.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 249, + 188, + 342, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 188, + 342, + 198 + ], + "spans": [ + { + "bbox": [ + 249, + 188, + 342, + 198 + ], + "type": "text", + "content": "linliang@ieee.org" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 252, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 289, + 529 + ], + "type": "text", + "content": "Deep learning models are challenged by the distribution shift between the training data and test data. Recently, the large models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts. However, fine-tuning these models can lead to a trade-off between in-distribution (ID) performance and out-of-distribution (OOD) robustness. Existing methods for tackling this trade-off do not explicitly address the OOD robustness problem. In this paper, based on causal analysis of the aforementioned problems, we propose a novel fine-tuning method, which uses masked images as counterfactual samples that help improve the robustness of the fine-tuning model. Specifically, we mask either the semantics-related or semantics-unrelated patches of the images based on class activation map to break the spurious correlation, and refill the masked patches with patches from other images. The resulting counterfactual samples are used in feature-based distillation with the pre-trained model. Extensive experiments verify that regularizing the fine-tuning with the proposed masked images can achieve a better trade-off between ID and OOD performance, surpassing previous methods on the OOD performance. Our code is available at https://github.com/Coxy7/robust-finetuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 541, + 128, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 128, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 128, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 562, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 693 + ], + "type": "text", + "content": "Deep learning has made impressive advances in various tasks on computer vision. Despite the remarkable performance achieved on benchmark datasets, deep models are often challenged by the distribution shift between the training data and test data [5, 15, 16, 33]. It is commonly assumed that the training and test samples follow the same distribution, which may not hold in real-world applications due to the unpredictable change of lighting conditions, viewpoints, backgrounds, etc. Although there are attempts to improve the robustness of deep models to the distribution shift (or OOD robustness), it is still rather under-explored [29, 38]." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 318, + 227, + 538, + 392 + ], + "blocks": [ + { + "bbox": [ + 318, + 227, + 538, + 392 + ], + "lines": [ + { + "bbox": [ + 318, + 227, + 538, + 392 + ], + "spans": [ + { + "bbox": [ + 318, + 227, + 538, + 392 + ], + "type": "image", + "image_path": "8bba6db4d514f606b8f66ae39d5fae798849202cb53728a56987d9e001bdd4f4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 399, + 547, + 444 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 547, + 444 + ], + "type": "text", + "content": "Figure 1. Illustration of our work. Vanilla fine-tuned models tend to learn spurious correlations that degrade the OOD robustness. To tackle this issue, our model learns from the pre-trained model on the counterfactual CAM-based masked images." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 460, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 628 + ], + "type": "text", + "content": "Recently, large-scale models pre-trained on diverse data have demonstrated unprecedented robustness to various distribution shifts [8, 19, 32]. Fine-tuning these pre-trained models on downstream tasks can be a promising approach to building robust models for different applications. However, it is found that while fine-tuning improves the performance on in-distribution (ID) data, it may reduce that on out-of-distribution (OOD) data [23, 43]. To tackle this trade-off, several methods [23, 42, 43] have been proposed to improve both ID and OOD performance in fine-tuning. However, they do not explicitly address the OOD robustness problem; instead, they implicitly preserve the robustness of the pre-trained model by constraining the distortion of pre-trained weights or using model ensembles." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we revisit the issue of robustness degradation in fine-tuning from a causal perspective. A large-scale pre-trained model somewhat shows properties in causality and stays robust to OOD samples [41]. However, when fine-tuning on downstream tasks, a majority of the model parameters tend to be adjusted for the downstream task in fine-tuning due to the highly entangled representation of" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20301" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "images, arguably destructive to the generalizable knowledge [20, 35]. In contrast, distribution shifts are usually sparse in the underlying causal factorization of the data generation process [7, 35]. In this low-dimensional case, if we know which variables vary with different data distributions in this factorization (i.e., the non-stationary factors), we can achieve the OOD robustness by simply excluding their influence on the final predictions of the model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 288, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 290 + ], + "type": "text", + "content": "Specifically, we consider a Structural Causal Model (SCM) [31] for the object-centric image generation process, as depicted in Fig. 2. In this SCM, images are generated according to a non-stationary domain-relevant factor and a stationary semantic factor. Between them is a spurious correlation caused by a hidden non-stationary confounder that influences how the domain-relevant factor changes with the semantic one. To retain the OOD robustness, a fin-tuning model should avoid mapping non-stationary domain-relevant features to the predicted semantics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 292, + 288, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 288, + 508 + ], + "type": "text", + "content": "To this end, we propose to fine-tune the models with masked images, which serve as counterfactual samples breaking the spurious correlation. Training on these samples helps preserve the stationary and generalizable knowledge of the pre-trained model. Concretely, we either mask the patches that contribute most to the label (i.e., the main object) or mask those with the least contribution (e.g., the context), which can be implemented based on class activation map (CAM) [9,46]. Such image masking forms an manipulation of a factual image and produces a counterfactual sample. Since the pre-trained model can better disentangle invariant features across domains, we require the finetuning model to learn from the pre-trained model on these counterfactual samples, as illustrated in Fig. 1. Furthermore, we argue that simply dropping the masked patches may be insufficient to alleviate the risk of fitting spurious correlations, and we propose to refill the masked patches with those from other images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "We study different combinations of masking strategies (e.g., masking the object or the context) and refilling strategies (e.g., filling with patches from single or multiple images). Experimental results suggest that most of the strategies are applicable to the construction of counterfactual samples that help improve the robustness in fine-tuning, while masking the object generally achieves the best robustness. Compared with existing methods [23, 42, 43] on fine-tuning CLIP [32] models, our approach achieves better average accuracy on various OOD datasets without relying on model ensembles or weight constraints. We also find that, taking the weight-space ensemble of the zero-shot model and our fine-tuned model following WiSE-FT [43], hardly improves the trade-off between ID and OOD accuracy, which contradicts previous observations and implies that our approach may produce essentially different models in comparison with conventional fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 71, + 398, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 398, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 398, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 91, + 515, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 515, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 515, + 104 + ], + "type": "text", + "content": "2.1. Causal Perspective on OOD Robustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 110, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 289 + ], + "type": "text", + "content": "Causality holds the promise of OOD robustness as it has the property of robustness under interventions that underlie distribution shifts [2, 35]. To learn the latent causal mechanism in visual data and attain OOD robustness, a branch of methods [21, 28, 36] target at learning domain-invariant causal representations and seek guarantees on generalization. Some approaches approximates this goal based on invariant risk minimization [1, 3], risk extrapolation [22], adaptation speed [7], or variational Bayes [25]. Ilse et al. [18] help models escape from spurious correlations by training on handcrafted analog interventional data. In this work, we aim to explore the preservation of the OOD robustness in pre-trained models equipped with generalizable knowledge during transfer, based on the analysis of an underlying data generation causal mechanism." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 297, + 487, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 487, + 310 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 487, + 310 + ], + "type": "text", + "content": "2.2. ID-OD Trade-off in Fine-tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 316, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 567 + ], + "type": "text", + "content": "While a series of large pre-trained models exhibit strong OOD robustness [8,19,32], empirical evidence suggests that fine-tuning these models on downstream data may degrade the robustness [23, 32, 43]. It is also theoretically justified that compared with linear probing, vanilla fine-tuning may increase the ID performance while decreasing the OOD performance due to feature distortion [23]. To tackle this ID-OOD trade-off, LP-FT [23] applies linear probing before fine-tuning; Calibrated Ensemble [24] and WiSE-FT [43] propose to take the ensemble of the pre-trained (zero-shot) model and vanilla fine-tuned model in output-space and weight-space, respectively. Model Soup [42] also improves the trade-off via weight-space ensembles, but it utilizes multiple models fine-tuned with different hyper-parameters. While these methods are empirically effective in improving both ID and OOD performance, they only implicitly preserve the robustness of the pre-trained model by constraining the deviation of the downstream model from the pre-trained one. Instead, our approach is based on explicit causal modeling of the OOD robustness problem and the use of counterfactual samples in fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 575, + 471, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 471, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 471, + 588 + ], + "type": "text", + "content": "2.3. Learning with Masked Images" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": "Masked image modeling [4,14,40,44] has been proven to be an effective approach to vision model pre-training, where random sampling is a common strategy for masking. In our task, we find that learning with CAM-based masking can better address the spurious correlation in images. Similar to our approach, CSS [10] synthesizes counterfactual images for visual question answering (VQA) by masking out critical objects, for which the corresponding answer is changed to its negative; SwapMix [13] swaps the context objects in the feature space to reduce the reliance of VQA models on" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20302" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "visual context. In this paper, we apply masking on images and refill the masked regions with content from other images, which is inspired by CutMix [45]. Besides, instead of manually constructing the labels for masked images like CSS or CutMix, we take the feature representations of the pre-trained model on the masked images as the supervision for the fine-tuning model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 168, + 102, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 102, + 181 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 102, + 181 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 189, + 286, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 189, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 189, + 286, + 213 + ], + "type": "text", + "content": "3.1. Revisit OOD Robustness of Fine-tuned Models from a Causal Perspective" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 286, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 286, + 327 + ], + "type": "text", + "content": "When fine-tuning a pre-trained model on downstream tasks in real-world applications, it is usually reliable to generalize only to in-distribution (ID) data. There are also increasing demands for robustness to distribution shifts or generalization to out-of-distribution (OOD) data. However, for a pre-trained large model that has been trained on data from diverse distributions and obtained generalizable knowledge (e.g., CLIP [32]), fine-tuning can lead to dramatic degradation of robustness [23]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 328, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 286, + 506 + ], + "type": "text", + "content": "This problem can be interpreted from a causal perspective. Suppose that we correctly factorize the data generation process in which variables are causally connected. Then, the discrepancy between ID and OOD (or sub-population) data generation process is usually sparse in this causal factorization [35, 36], e.g., few variables' priors change. The joint distribution that we can observe thereupon shifts. However, without modeling the causal hierarchy in data, the pre-trained models naturally learn dense and entangled features to represent the concepts in the causal factorization. Although distribution shifts are sparse in causal factorization, these models have to densely adapt their parameters in transfer to the downstream data [20]. In this case, the loss of generalizable knowledge is inevitable, which explains the significant decrease of robustness to distribution shifts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "text", + "content": "The core challenge in modeling the underlying causal factorization of visual data is the dense changes between ID and OOD data, despite the sparsity in causal factorization. For better robustness, a possible solution is to identify which visual features remain unchanged under distribution shifts, also known as stationary features [35]. Without OOD test data in hand to identify stationary features, we seek to model the most important stationary features as the semantic features in the downstream object-centric task, and reduce non-stationary features to domain-relevant features. Finally, we establish our causal model as in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 268, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 268, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 268, + 669 + ], + "type": "text", + "content": "3.1.1 SCM for Image Generation Across Domains" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "To illustrate the proposed method, we first introduce in Fig. 2 a Structural Causal Model (SCM) [31] to model the underlying mechanism of object-centric image data genera" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 384, + 71, + 466, + 158 + ], + "blocks": [ + { + "bbox": [ + 384, + 71, + 466, + 158 + ], + "lines": [ + { + "bbox": [ + 384, + 71, + 466, + 158 + ], + "spans": [ + { + "bbox": [ + 384, + 71, + 466, + 158 + ], + "type": "image", + "image_path": "36189ee5f657509a3d3166fc27e3598d934330392b08df0dabc95deaac4903ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "text", + "content": "Figure 2. The causal graph of underlying object-centric image generation process across domains. " + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "text", + "content": ": confounder; " + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "text", + "content": ": domain; " + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "text", + "content": ": object semantics; " + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "inline_equation", + "content": "H_d" + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "text", + "content": ": (non-semantic) domain representation; " + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "inline_equation", + "content": "H_s" + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "text", + "content": ": semantic representation; " + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 306, + 163, + 545, + 208 + ], + "type": "text", + "content": ": image." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 220, + 544, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 544, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 544, + 243 + ], + "type": "text", + "content": "tion across diverse domains, which is consistent with previous works [18, 37] despite some slight modifications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "content": "Formally, an SCM describes a directed acyclic graph composed of a set of endogenous variables " + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "content": " and a set of exogenous variables " + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "content": ". An endogenous variable is a variable whose value is determined by other variables. Exogenous variables correspond to unobserved influences, usually considered independent of each other. In Fig. 2, a set of endogenous variables are involved in the SCM, defined as " + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "inline_equation", + "content": "V = \\{D, S, H_d, H_s, X\\}" + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "content": ". There is also an exogenous variable for each endogenous variable, e.g., " + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "inline_equation", + "content": "U_{H_s}" + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "content": " that possibly corresponds to the object pose and influences the generation of " + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "inline_equation", + "content": "H_s" + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "content": ". Here, we do not show them for simplicity, but only consider an important exogenous variable, the confounder " + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "content": ". We denote the set of exogenous variables by " + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "inline_equation", + "content": "U = \\{C, U_D, U_S, U_{H_s}, U_{H_d}, U_X\\}" + }, + { + "bbox": [ + 305, + 244, + 545, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": "In this SCM instance, " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " denotes an unobservable confounder variable, which is an exogenous variable shown in gray. For instance, it can be some specific time or space. " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " is a domain variable containing information varied with the domain. " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " is a semantic variable, e.g., the category of the object. In our setting, " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " are observable endogenous variables shown in white. Correspondingly, " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " is the domain representation and " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " is the semantic representation, i.e., they indicate how " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " are displayed in the image space, respectively. Finally, " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " is the image generated by the interaction of " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 305, + 412, + 545, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 562, + 523, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 523, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 523, + 574 + ], + "type": "text", + "content": "3.1.2 Spurious Correlation and OOD Robustness" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "text", + "content": "As depicted in Fig. 2, there exists a backdoor path between " + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_{d} \\gets D \\gets C \\to S \\to H_{s}" + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "text", + "content": ", making them spuriously correlated. Collecting data for a downstream task from a single source or environment may lead to strong spurious correlations between the semantic part " + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "text", + "content": " and domain-relevant part " + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 306, + 582, + 545, + 713 + ], + "type": "text", + "content": ". In other words, there can be strong selection bias [28]. For example, in a downstream insect classification dataset, the red admiral butterflies may all be captured sitting on flowers, as shown in Fig. 1. However, this spurious correlation does not necessarily hold in OOD data [6]. Moreover, in reality, what makes the label" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20303" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 64, + 474, + 277 + ], + "blocks": [ + { + "bbox": [ + 118, + 64, + 474, + 277 + ], + "lines": [ + { + "bbox": [ + 118, + 64, + 474, + 277 + ], + "spans": [ + { + "bbox": [ + 118, + 64, + 474, + 277 + ], + "type": "image", + "image_path": "7ec26745c8d609b2ccac31706f7fede5671f97682d276e0160914ce5c51f80b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "text", + "content": "Figure 3. Conceptual comparison of pre-trained and fine-tuned models on what they depend on for predicting the semantic label of different samples. The robust pre-trained model is expected to depend only on " + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "text", + "content": ", while the fine-tuned model may associate both " + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "inline_equation", + "content": "h_d" + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "text", + "content": " to the semantics " + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 281, + 547, + 315 + ], + "type": "text", + "content": ". This may result in different predictions on counterfactual samples for the two models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "text", + "content": "red admiral butterfly is the butterfly itself in the image space " + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "inline_equation", + "content": "(H_{s})" + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "text", + "content": " rather than the flowers involved in " + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "text", + "content": ". Therefore, the model is required to depend only on " + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "text", + "content": " for the prediction of " + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "text", + "content": " to attain OOD robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": "For the reasons we have discussed earlier, even if a pre-trained large model has been endowed with reasonable knowledge to distinguish " + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": " (Fig. 3 (a)), the model is still susceptible to the spurious correlations during finetuning (Fig. 3 (d)). As a result, the transfer dramatically degenerates the pre-trained model's OOD robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "text", + "content": "From this perspective, a possible way to tackle the issue of robustness degradation is to break the spurious correlations in downstream data via specific manipulation, and explicitly requires the fine-tuning model to distinguish " + }, + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "text", + "content": " following the pre-trained model. In the following sections, we propose a masking-based image manipulation method and indicate why it can be used to hinder the fine-tuning model from depending on " + }, + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 46, + 457, + 287, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 562, + 272, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 272, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 272, + 576 + ], + "type": "text", + "content": "3.2. Masked Images as Counterfactual Samples" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": "To break the spurious correlation in training images, we propose to mask out or replace certain regions of the images so that " + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": " is (partially) manipulated, which results in counterfactual samples. Formally, given an observational sample " + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": " and the SCM " + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": ", if we assign " + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": " to variable " + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": ", the resultant counterfactual sample can be denoted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 665, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 665, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 665, + 287, + 681 + ], + "type": "interline_equation", + "content": "x _ {D = d ^ {\\prime}} (u) = x _ {M _ {D = d ^ {\\prime}}} (u) \\tag {1}", + "image_path": "44828b56c9db8c421791e73debb0e077d32411ccfac37b077a9f6a52074eba1d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "M_{D = d'}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the SCM instance in which " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is assigned to " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " denotes the values of the exogenous variables." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": "We illustrate how the pre-trained and fine-tuned models may process two kinds of counterfactual samples differently in Fig. 3. First, for counterfactual samples whose domain-relevant representations " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "h_d" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " are replaced by some other context " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "x_{H_d = p}" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " in Fig. 3 (b,e)), the fine-tuned model tends to attribute its prediction on both " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " since it has learned the spurious correlations. Hence, its prediction can be misled by " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": ". Differently, the robust pre-trained model which can distinguish the semantic and non-semantic factors is not affected by " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": ". Second, for counterfactual samples whose semantic representations " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " are replaced by " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "x_{H_d = p}" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " in Fig. 3 (c,f)), the fine-tuned model can still predict the original semantics " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "h_d" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": ", while the pre-trained model cannot due to the missing semantic representations " + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 304, + 335, + 547, + 528 + ], + "type": "text", + "content": ". In both cases, counterfactual samples may lead to different predictions between the two models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": "Now the question is how to leverage these counterfactual samples to preserve the OOD robustness in fine-tuning. Directly training on these counterfactual samples with the labels of corresponding factual samples is unsuitable, since the original semantic information may be distorted. Given that the pre-trained models can have substantial power to capture semantic cues, their image-level feature representations usually contain rich semantic information. We thereby seek an alternative solution in which these samples are used in distillation. This way, the fine-tuning model learns to mimic the pre-trained model in feature representation. Formally, we denote the image encoders of the pre-trained and fine-tuning model by " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": ", respectively, and denote the classification head of the fine-tuning model by " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": ". Then, the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20304" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 287, + 173 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 287, + 173 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 287, + 173 + ], + "type": "image", + "image_path": "f836923693b44fc814c04943552bdc6b1bfb1112cf075d368541b2acc6ca4508.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 179, + 284, + 190 + ], + "lines": [ + { + "bbox": [ + 49, + 179, + 284, + 190 + ], + "spans": [ + { + "bbox": [ + 49, + 179, + 284, + 190 + ], + "type": "text", + "content": "Figure 4. Illustration of the mechanism of masking and refilling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 200, + 255, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 255, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 255, + 211 + ], + "type": "text", + "content": "overall training objective can be written as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 218, + 287, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 218, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 59, + 218, + 287, + 232 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} (g (f (x)), y) + \\beta \\mathcal {L} _ {\\mathrm {M S E}} (\\hat {f} (x _ {c f}), f (x _ {c f})), \\tag {2}", + "image_path": "404e035f0880bb83bfb7a073bb1957fef09b1d75bdbca861ebf9d862b3b1f1ea.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "content": " is the cross-entropy loss, " + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "content": " is the mean squared error, " + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "content": " is the raw image, " + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "content": " is the label for " + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "x_{cf}" + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "content": " is the counterfactual image, and " + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 238, + 287, + 274 + ], + "type": "text", + "content": " is a constant factor." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 274, + 287, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 321 + ], + "type": "text", + "content": "In the remaining parts of this section, we will discuss the proper design of masked images and show that it is nontrivial to construct the counterfactual samples " + }, + { + "bbox": [ + 46, + 274, + 287, + 321 + ], + "type": "inline_equation", + "content": "x_{cf}" + }, + { + "bbox": [ + 46, + 274, + 287, + 321 + ], + "type": "text", + "content": " that can be used to effectively improve the OOD robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 336, + 209, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 209, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 209, + 347 + ], + "type": "text", + "content": "3.2.1 Choosing the Masked Regions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 354, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 569 + ], + "type": "text", + "content": "To generate counterfactual samples, we need to select the masked regions in an image where the contents can be manipulated. Recent large-scale pre-trained models [19, 32] usually adopt a Vision Transformer-based architecture [12], dividing input images into a series of patches. Hence, we consider patch-based masking in this paper. A basic strategy is to randomly sample a proportion of patches (random-mask), which is adopted by recent pre-training methods like MAE [14]. However, random sampling is content-agnostic. Although both " + }, + { + "bbox": [ + 46, + 354, + 287, + 569 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 46, + 354, + 287, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 287, + 569 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 46, + 354, + 287, + 569 + ], + "type": "text", + "content": " may be manipulated in this scheme, it tends to leave some spurious correlations in the unmasked regions. It is found that an MAE model can achieve semantically plausible reconstruction from the randomly masked images with a high masking rate (e.g., 75%), suggesting that such random masking may preserve a significant amount of the information of the whole images. Therefore, randomly masked images may be less suitable for improving OOD robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Alternatively, we investigate two content-based strategies that are complementary: (1) context-mask: masking the patches that are least relevant to the label (i.e., " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ", usually the context); (2) object-mask: masking the patches that are most relevant to the label (i.e., " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ", usually the main object). To measure the contribution of each image patch to the label, we generate the class activation map (CAM) [9, 46] based on the fine-tuning model. Then, image patches are separated into two groups by a constant threshold " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ". In such a scheme, the semantic-relevant regions (where their activation values are larger than the threshold " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ") can coarsely serve as " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ", and non-semantic regions as " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 460, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 460, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 460, + 84 + ], + "type": "text", + "content": "3.2.2 Refilling the Masked Images" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "type": "text", + "content": "After choosing the patches to mask, simply dropping them following MAE [14] can lead to a counterfactual image, i.e., " + }, + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "type": "inline_equation", + "content": "x_{H_d = \\emptyset}(u)" + }, + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "type": "text", + "content": ". However, we argue that this strategy (abbreviated as no-fill) can be insufficient for the construction of effective counterfactual samples for improving the robustness, and refilling the masked regions can tackle this issue." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 163, + 546, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 546, + 390 + ], + "type": "text", + "content": "Concretely, an effective counterfactual sample should cause contradictions between the pre-trained and fine-tuning model, so that the latter is regularized by the distillation loss in Eq. (2) to depend less on non-semantic parts of images for the prediction of semantics. However, masking without refilling may not construct such samples, as depicted in Fig. 4. Specifically, context-mask may not produce contradictions since both models can predict semantics from the object, so we need refilling to bring some conflicting context that disturbs the prediction of the fine-tuning model. Object-mask alone can cause contradictions in theory, as the fine-tuning model could still predict the original semantics from the context, while the pre-trained model could not. However, due to the imperfect masking in practice, the pre-trained model may still recognize the object from its unmasked parts. Then, refilling can further distort the original semantics to ensure contradictions. Hence, refilling can help to construct more effective counterfactual samples for the proposed fine-tuning approach." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 475 + ], + "type": "text", + "content": "We consider two basic strategies to select the patches to fill: (1) single-fill: select the patches in the corresponding positions of a single image randomly sampled from the training batch; (2) multi-fill: for each patch to fill, independently sample a source image from the batch and select the patch in the corresponding position. All combinations of masking and refilling strategies are illustrated in Fig. 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 483, + 387, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 387, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 387, + 498 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 503, + 356, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 356, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 356, + 517 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": "Datasets. We focus on fine-tuning the pre-trained model on ImageNet [34], and evaluate the OOD robustness of the models on five datasets: ImageNet-V2 [33] is a new test set for ImageNet collected following the original protocol. ImageNet-R [15] consists of various renditions (e.g., art, cartoons) of 200 ImageNet classes; ImageNet-Sketch [39] contains sketches for each of the 1000 ImageNet classes; ObjectNet [5] is a test set with 113 overlapping classes with ImageNet, which is collected to show objects from new viewpoints on new backgrounds; ImageNet-A [16] contains natural images that are misclassified by models trained on ImageNet, covering 200 ImageNet classes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Evaluation. Following [43], we use the top-1 accuracy as the metric of performance on ID and OOD data. For the five OOD datasets, we report the top-1 accuracy on each dataset, as well as the average OOD accuracy computed by" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20305" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 130, + 142 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 130, + 142 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 130, + 142 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 130, + 142 + ], + "type": "image", + "image_path": "a5e66debb04a44932077755da89fd69b2accafdd4e4e57989112de546e9e6b17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 72, + 194, + 143 + ], + "blocks": [ + { + "bbox": [ + 137, + 72, + 194, + 143 + ], + "lines": [ + { + "bbox": [ + 137, + 72, + 194, + 143 + ], + "spans": [ + { + "bbox": [ + 137, + 72, + 194, + 143 + ], + "type": "image", + "image_path": "16528456bc8117d2fc9a73451916eeda68b30190defff6d38ba2c68bebeb976f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 201, + 72, + 258, + 142 + ], + "blocks": [ + { + "bbox": [ + 201, + 72, + 258, + 142 + ], + "lines": [ + { + "bbox": [ + 201, + 72, + 258, + 142 + ], + "spans": [ + { + "bbox": [ + 201, + 72, + 258, + 142 + ], + "type": "image", + "image_path": "4056c59f0ad36fd0a33d28cd0e885e5fb9a58933883a6ca5880f30d7cbed3797.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 74, + 158, + 130, + 214 + ], + "blocks": [ + { + "bbox": [ + 84, + 145, + 122, + 157 + ], + "lines": [ + { + "bbox": [ + 84, + 145, + 122, + 157 + ], + "spans": [ + { + "bbox": [ + 84, + 145, + 122, + 157 + ], + "type": "text", + "content": "random-mask single-fill" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 158, + 130, + 214 + ], + "lines": [ + { + "bbox": [ + 74, + 158, + 130, + 214 + ], + "spans": [ + { + "bbox": [ + 74, + 158, + 130, + 214 + ], + "type": "image", + "image_path": "ed7675d8a9636bb18e5fe93ab0ecdb62775ab2febadd1b277553818b3a418f18.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 217, + 122, + 228 + ], + "lines": [ + { + "bbox": [ + 83, + 217, + 122, + 228 + ], + "spans": [ + { + "bbox": [ + 83, + 217, + 122, + 228 + ], + "type": "text", + "content": "random-mask multi-fill" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 158, + 194, + 214 + ], + "blocks": [ + { + "bbox": [ + 147, + 145, + 186, + 158 + ], + "lines": [ + { + "bbox": [ + 147, + 145, + 186, + 158 + ], + "spans": [ + { + "bbox": [ + 147, + 145, + 186, + 158 + ], + "type": "text", + "content": "context-mask single-fill" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 158, + 194, + 214 + ], + "lines": [ + { + "bbox": [ + 137, + 158, + 194, + 214 + ], + "spans": [ + { + "bbox": [ + 137, + 158, + 194, + 214 + ], + "type": "image", + "image_path": "be36d09ebd18212dc3bb47c4524360d515b8ce1e83dd9350b214f212a6c85456.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 201, + 158, + 258, + 215 + ], + "blocks": [ + { + "bbox": [ + 212, + 144, + 248, + 158 + ], + "lines": [ + { + "bbox": [ + 212, + 144, + 248, + 158 + ], + "spans": [ + { + "bbox": [ + 212, + 144, + 248, + 158 + ], + "type": "text", + "content": "object-mask single-fill" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 158, + 258, + 215 + ], + "lines": [ + { + "bbox": [ + 201, + 158, + 258, + 215 + ], + "spans": [ + { + "bbox": [ + 201, + 158, + 258, + 215 + ], + "type": "image", + "image_path": "a68287cb18062e8b67bf016cf4ec9e83f709e4b705fb181c86232e1dd06e6ee9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 217, + 247, + 229 + ], + "lines": [ + { + "bbox": [ + 212, + 217, + 247, + 229 + ], + "spans": [ + { + "bbox": [ + 212, + 217, + 247, + 229 + ], + "type": "text", + "content": "object-mask multi-fill" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 74, + 231, + 130, + 287 + ], + "blocks": [ + { + "bbox": [ + 74, + 231, + 130, + 287 + ], + "lines": [ + { + "bbox": [ + 74, + 231, + 130, + 287 + ], + "spans": [ + { + "bbox": [ + 74, + 231, + 130, + 287 + ], + "type": "image", + "image_path": "f2cd7f4be29dc5618e97ee13073f325031381753103e1317d5ecd537cdc84f1d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 293, + 286, + 316 + ], + "lines": [ + { + "bbox": [ + 47, + 293, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 286, + 316 + ], + "type": "text", + "content": "Figure 5. Illustration of different masking and refilling strategies. The raw (factual) image is displayed in Fig. 4." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 137, + 231, + 194, + 286 + ], + "blocks": [ + { + "bbox": [ + 146, + 217, + 186, + 229 + ], + "lines": [ + { + "bbox": [ + 146, + 217, + 186, + 229 + ], + "spans": [ + { + "bbox": [ + 146, + 217, + 186, + 229 + ], + "type": "text", + "content": "context-mask multi-fill" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 231, + 194, + 286 + ], + "lines": [ + { + "bbox": [ + 137, + 231, + 194, + 286 + ], + "spans": [ + { + "bbox": [ + 137, + 231, + 194, + 286 + ], + "type": "image", + "image_path": "4065a590ae2a407c5b55176930670a65697b92dfd6fdbdbe5dd87ab44f121465.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 201, + 231, + 258, + 287 + ], + "blocks": [ + { + "bbox": [ + 201, + 231, + 258, + 287 + ], + "lines": [ + { + "bbox": [ + 201, + 231, + 258, + 287 + ], + "spans": [ + { + "bbox": [ + 201, + 231, + 258, + 287 + ], + "type": "image", + "image_path": "16ebef063fd467716171fbbf13ea5ed0bd2591db45c143b481e95ac1769e80f7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 326, + 286, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 286, + 373 + ], + "type": "text", + "content": "averaging the accuracy on the five datasets. For datasets covering a subset of the ImageNet classes, the top-1 prediction of a model is taken as the class with the highest probability among the subset (instead of all ImageNet classes)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 374, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 286, + 601 + ], + "type": "text", + "content": "Implementation details. Unless otherwise specified, we use the ViT-B/32 model [12] pre-trained via CLIP [32]. The classification head of the zero-shot model is constructed from the pre-defined text prompts used by CLIP, and we adopt the implementation given by [43]. To obtain accurate CAM scores for image masking, we apply the method proposed in [9], which is designed for attention-based models, including ViT. For fine-tuning on ImageNet, we mainly follow the routine of WiSE-FT [43]. Specifically, we use the AdamW optimizer [27] with a batch size of 512, and finetune for 10 epochs. The learning rate is set to " + }, + { + "bbox": [ + 46, + 374, + 286, + 601 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 374, + 286, + 601 + ], + "type": "text", + "content": " for all parameters and follows a cosine-annealing schedule [26] with 500 warm-up steps. No data augmentation is applied apart from the necessary resizing and cropping, following the training of CLIP. We split out a validation set of 10240 samples from the ImageNet training set to perform early stopping and model selection based on the validation accuracy. " + }, + { + "bbox": [ + 46, + 374, + 286, + 601 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 374, + 286, + 601 + ], + "type": "text", + "content": " in Eq. (2) is fixed to be 30 for all experiments. Additional details are in the appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 610, + 222, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 222, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 222, + 624 + ], + "type": "text", + "content": "4.2. Masking and Refilling Strategies" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "In this section, we validate the effectiveness of using masked images for improving OOD robustness, and investigate the different masking and refilling strategies discussed in Sec. 3.2. Concretely, we consider three masking strategies: (1) random-mask: randomly select a fixed proportion of patches; (2) context-mask: masking the patches with CAM score lower than the threshold; (3)" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 535, + 123 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 535, + 123 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 535, + 123 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 535, + 123 + ], + "type": "image", + "image_path": "0cb293c28428178cc1e24b68b14c758fdc6c0a3c693d5794029c5b9fc76d49dc.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 316, + 125, + 526, + 277 + ], + "blocks": [ + { + "bbox": [ + 316, + 125, + 526, + 277 + ], + "lines": [ + { + "bbox": [ + 316, + 125, + 526, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 125, + 526, + 277 + ], + "type": "image", + "image_path": "c3ecd8847d9438167d8cb8efee1f5a69415feb65b8738b64b6c956d184624f2b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 284, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 284, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 545, + 338 + ], + "type": "text", + "content": "Figure 6. ID and average OOD accuracy of our approach using different masking and refilling strategies. For each combination of masking and refilling strategies, we plot multiple results based on different settings of hyper-parameters (i.e., masking rate or CAM score threshold). Best viewed in color." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "content": "object-mask: masking the patches with CAM score higher than the threshold. The masking rate for random-mask is selected from " + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\{0.25, 0.5, 0.75\\}" + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "content": ", and the CAM score thresholds for the latter two strategies are selected from " + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\{0.3, 0.4, 0.5, 0.6\\}" + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "content": ". The refilling strategies include: (1) no-fill: baseline strategy that drops the masked patches without refilling; (2) single-fill: refill with patches from one other image; (3) multi-fill: refill with patches from multiple other images. We test all the combinations of masking and refilling strategies, and compare them with the baseline where no masking is applied (i.e., " + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "inline_equation", + "content": "x_{cf} = x" + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "content": " in Eq. (2))." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 484, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 687 + ], + "type": "text", + "content": "In Fig. 6, we plot the average OOD accuracy against ID accuracy for each model trained with the above strategies. Besides, for each combination of masking and refilling strategies, we select the model with the highest validation accuracy, and report their results on each dataset in Tab. 1. Based on these results, we have the following conclusions. (1) Most of the combinations of masking and refilling strategies achieve better ID-OOD trade-off than the no-masking baseline, which suggests the effectiveness of image masking in our proposed approach. (2) Refilling masked images with patches from other images (i.e., single-fill or multi-fill) is better than solely dropping the masked patches. (3) Comparing the two proposed refilling strategies, single-fill generally results in better OOD accuracy, while multi-fill may yield better ID accuracy. (4) Comparing the masking strategies, object-mask is generally better than random-mask and context-mask in terms of OOD accuracy." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "The superiority of object-mask can be explained following our analysis in Sec. 3.2.2. Concretely, while both mask" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20306" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 544, + 218 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 544, + 218 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 544, + 218 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 544, + 218 + ], + "type": "table", + "html": "
MethodMaskingRefillingININ-V2IN-RIN-SketchObjectNetIN-AOOD avg.
Zero-shot [32]//63.455.969.342.344.531.448.7
Vanilla fine-tuning//75.964.757.039.839.520.044.2
Oursno masking/76.966.569.245.645.329.851.3
Oursrandom-maskno-fill77.566.966.445.746.530.851.2
context-mask77.867.466.745.645.930.051.1
object-mask77.767.167.646.246.831.551.9
Oursrandom-masksingle-fill77.667.169.046.447.833.452.7
context-mask77.266.968.846.547.832.952.6
object-mask77.567.169.746.948.033.853.1
Oursrandom-maskmulti-fill78.067.467.446.146.731.951.9
context-mask78.267.466.545.545.930.051.1
object-mask77.967.768.146.647.533.052.6
", + "image_path": "ac3c7e8a2a7cfb3a52344f2ea209aaf4dcb36cdcfde1119262f39a38171ef948.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 269, + 285, + 316 + ], + "blocks": [ + { + "bbox": [ + 46, + 221, + 546, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 546, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 546, + 257 + ], + "type": "text", + "content": "Table 1. Comparison of different masking and refilling strategies. Accuracy on ImageNet (IN) and the five OOD datasets are reported. OOD avg. is the average OOD accuracy on the five OOD datasets. The best accuracy is bold-faced, and the second best accuracy is underlined. Results are averaged over three runs with different seeds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 269, + 285, + 316 + ], + "lines": [ + { + "bbox": [ + 50, + 269, + 285, + 316 + ], + "spans": [ + { + "bbox": [ + 50, + 269, + 285, + 316 + ], + "type": "table", + "html": "
Threshold0.70.60.50.40.30.2
Image MR0.060.100.150.250.390.55
Object MR0.170.230.330.480.650.80
IoU0.150.210.280.380.450.48
", + "image_path": "129fa29191596f427493a07932d70e8aa37484a48b23476121ca580a7299610f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 320, + 287, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 377 + ], + "type": "text", + "content": "Table 2. Validation of CAM-based object masking with different thresholds. Image masking rate (MR): masking rate of the whole image. Object masking rate (MR): masking rate concerning the object area. IoU: Intersection over Union regarding the object. Averaged over a subset of masked samples constructed in training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 388, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 459 + ], + "type": "text", + "content": "ing strategies are theoretically valid, context-mask is more dependent on refilling, since it supposes that the refilled context can effectively lead to contradictions. This may not be satisfied by the proposed refilling strategies, as they are not aware of the content of the patches taken from other images. Hence, context-mask can be less effective in practice." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 460, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 460, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 45, + 460, + 287, + 662 + ], + "type": "text", + "content": "In addition, to verify that the superiority of our CAM-based object masking is due to the effective location of objects, we calculate the average masking rates of the objects during training. Specifically, we evaluate the object-mask and single-fill strategy with different CAM score thresholds. To obtain the masking rate of the object for an ImageNet image, we take the pixel-level mask produced by a pre-trained segmentation model [11] as the approximation of ground truth. In addition, we compute the masking rate of the whole image and the Intersection over Union (IoU) regarding the object. Details are provided in the appendix. As shown in Tab. 2, the object masking rate is significantly higher than the masking rate of the whole image, which suggests that the object-mask mainly masks the objects. Besides, as the threshold for masking decreases, both the object masking rate and IoU increase. Hence, the CAM-based object masking is empirically sound." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 251, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 251, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 251, + 685 + ], + "type": "text", + "content": "4.3. Comparison with Existing Approaches" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "We compare our method with three existing approaches, namely LP-FT [23], WiSE-FT [43] and Model Soup [42]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 272, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 536 + ], + "type": "text", + "content": "which also aim to improve both ID and OOD accuracy of fine-tuning. For WiSE-FT, we report the results with the default mixing coefficient, i.e., " + }, + { + "bbox": [ + 304, + 272, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 304, + 272, + 545, + 536 + ], + "type": "text", + "content": ". For Model Soup, we take the uniform soup as the representative method, which achieves the best average OOD accuracy on CLIP fine-tuning as reported in [42]. For our approach, we consider both refilling strategies, and adopt the object masking with the CAM score threshold that yields the highest validation accuracy. The results are shown in Tab. 3. First, our approach surpasses the previous approaches in terms of average OOD accuracy. Second, our approach achieves better ID accuracy than other approaches except for Model Soup. Note that Model Soup takes the ensemble of many finetuned models with different hyper-parameters, including those trained with strong data augmentation. Conversely, we do not dive into hyper-parameters tuning and do not use data augmentation in our experiments. Third, compared with WiSE-FT, our approach achieves superior performance on ObjectNet and ImageNet-A, but is inferior on ImageNet-R and ImageNet-Sketch, which suggests that the two approaches may improve different perspectives of robustness. This motivates us to consider the integration of WiSE-FT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 549, + 429, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 429, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 429, + 562 + ], + "type": "text", + "content": "4.4. Integrating WiSE-FT" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "WiSE-FT ensemble the weights of the zero-shot model and the vanilla fine-tuned model, and the ensemble model may achieve better accuracy on both ID and OOD data. Concretely, the weight ensemble is computed by " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{e} = (1 - \\alpha)\\cdot \\theta_{0} + \\alpha \\theta_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " is the weights of the zero-shot model, and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " is the weights of the fine-tuned model. This weight-space ensemble method is agnostic of the finetuning method in form. Therefore, we may expect that the ensemble of the zero-shot model and the model fine-tuned with our method can achieve even better results. However, as shown in Fig. 7, such direct integration with WiSE-FT may not improve ID and OOD accuracy simultaneously." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20307" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 544, + 222 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 544, + 222 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 544, + 222 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 544, + 222 + ], + "type": "table", + "html": "
ModelMethodININ-V2IN-RIN-SketchObjectNetIN-AOOD avg.
CLIP ViT-B/32Zero-shot [32]63.455.969.342.344.531.448.7
Vanilla fine-tuning75.964.757.039.839.520.044.2
WiSE-FT†[43]76.666.670.247.146.331.952.4
Uniform soup‡[42]80.068.666.647.746.129.251.6
Ours (multi-fill)77.967.768.146.647.533.052.6
Ours (single-fill)77.567.169.746.948.033.853.1
CLIP ViT-B/16Zero-shot [32]68.361.977.648.354.050.158.4
Vanilla fine-tuning80.770.464.045.149.135.252.8
LP-FT [23]81.771.672.948.4/49.1/
WiSE-FT [43]81.772.878.753.957.352.263.0
Ours (multi-fill)82.573.476.452.756.852.062.3
Ours (single-fill)82.473.478.153.457.953.563.3
", + "image_path": "52e73d08cf9ca89a8dd8a2df318242601538c2bcac4cb3e8268e13c613fe3bca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 225, + 546, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 546, + 258 + ], + "type": "text", + "content": "Table 3. Accuracy of different methods for fine-tuning CLIP models on ImageNet (IN). OOD avg. is the average OOD accuracy on the five OOD datasets. The best accuracy is bold-faced, and the second best accuracy is underlined. Our results are averaged over three runs with different seeds. " + }, + { + "bbox": [ + 46, + 225, + 546, + 258 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 225, + 546, + 258 + ], + "type": "text", + "content": ": our implementation. " + }, + { + "bbox": [ + 46, + 225, + 546, + 258 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 225, + 546, + 258 + ], + "type": "text", + "content": ": our evaluation on the official model.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 275, + 274, + 482 + ], + "blocks": [ + { + "bbox": [ + 58, + 275, + 274, + 482 + ], + "lines": [ + { + "bbox": [ + 58, + 275, + 274, + 482 + ], + "spans": [ + { + "bbox": [ + 58, + 275, + 274, + 482 + ], + "type": "image", + "image_path": "842de7f3f96a027dc579f70fbc9e5d41cfaa5e66e4c71126c2952eddaa8474c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 487, + 287, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 544 + ], + "type": "text", + "content": "Figure 7. Integration of WiSE-FT and our approach. WiSE-Ours: weight-space ensembles of the zero-shot model and our fine-tuned model. WiSE-KD: knowledge distillation from the WiSE-FT model (" + }, + { + "bbox": [ + 46, + 487, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 46, + 487, + 287, + 544 + ], + "type": "text", + "content": "). Vanilla FT: vanilla fine-tuning. For WiSE-FT and WiSE-Ours, each point corresponds to a value of " + }, + { + "bbox": [ + 46, + 487, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 46, + 487, + 287, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 556, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 688 + ], + "type": "text", + "content": "Specifically, only for " + }, + { + "bbox": [ + 46, + 556, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 556, + 287, + 688 + ], + "type": "text", + "content": " close to 1, the resulting model may achieve better average OOD accuracy, but still at the cost of lower ID accuracy. This is different from the observations of WiSE-FT that using a medium value of " + }, + { + "bbox": [ + 46, + 556, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 556, + 287, + 688 + ], + "type": "text", + "content": " (e.g., 0.5) usually yields near-optimal results and surpasses the fine-tuned model on both ID and OOD accuracy [43]. It is suggested that our proposed method may produce a significantly different model as compared with vanilla fine-tuning. In other words, our fine-tuned weights may leave the basin of the loss landscape where the weights of most vanilla fine-tuned models and the pre-trained model lie in [30,42,43]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Another way to integrate WiSE-FT into our method is to utilize the ensemble model produced by WiSE-FT to guide" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "text", + "content": "the fine-tuning model via knowledge distillation (KD) [17]. For brevity, we name this method \"WiSE-KD\". Specifically, we consider using the WiSE-FT model as another teacher model, and add the vanilla knowledge distillation loss [17] to our training objective Eq. (2). Details are presented in the appendix. As presented in Fig. 7, applying WiSE-KD to our approach slightly improves both ID and OOD accuracy. As a comparison, while applying WiSE-KD to vanilla fin-tuning results in significantly better performance, it is still inferior to our approach, especially for OOD accuracy. This again validates the effectiveness of our approach in improving the OOD robustness in fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 431, + 459, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 459, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 459, + 443 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "type": "text", + "content": "In this paper, we analyze the issue of robustness degradation in fine-tuning from a causal perspective, and find that masked images can be effective counterfactual samples for breaking the spurious correlation between semantic and non-semantic factors and improving the OOD robustness of fine-tuning models. Experiments suggest that our approach surpasses previous methods on the OOD performance with competitive ID performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "type": "text", + "content": "Limitations. As stated in Sec. 4.2, our refilling strategies are unaware of the content of the patches to fill, which limits the effectiveness of the resulting counterfactual samples. Besides, the feature-based distillation may not regularize the learning of the classification head of the fine-tuning model. Future works can devote to better refilling and distillation methods conforming to the causal modeling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by NSFC (No. U21A20470), National Key R&D Program of China (No. 2021ZD0111601), and the Science and Technology Program of Guangzhou, China (No. 202201011550)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20308" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Kartik Ahuja, Karthikeyan Shanmugam, Kush Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning, pages 145-155. PMLR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 158 + ], + "type": "text", + "content": "[2] Martin Arjovsky. Out of distribution generalization in machine learning. PhD thesis, New York University, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 287, + 192 + ], + "type": "text", + "content": "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 227 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 229, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 229, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 229, + 287, + 293 + ], + "type": "text", + "content": "[5] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 297, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 287, + 329 + ], + "type": "text", + "content": "[6] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 385 + ], + "type": "text", + "content": "[7] Yoshua Bengio, Tristan Deleu, Nasim Rahaman, Nan Rosemary Ke, Sebastien Lachapelle, Olexa Bilaniuk, Anirudh Goyal, and Christopher Pal. A meta-transfer objective for learning to disentangle causal mechanisms. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 388, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 287, + 441 + ], + "type": "text", + "content": "[8] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 444, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 287, + 497 + ], + "type": "text", + "content": "[9] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "type": "text", + "content": "[10] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800-10809, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "text", + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[13] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5078-5088, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "text", + "content": "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 231, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 283 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 1, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "type": "text", + "content": "[17] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "text", + "content": "[18] Maximilian Ilse, Jakub M Tomczak, and Patrick Forre. Selecting data augmentation for simulating interventions. In International Conference on Machine Learning, pages 4555-4562. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 366, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 430 + ], + "type": "text", + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "type": "text", + "content": "[20] Nan Rosemary Ke, Aniket Didolkar, Sarthak Mittal, Anirudh Goyal, Guillaume Lajoie, Stefan Bauer, Danilo Rezende, Yoshua Bengio, Michael Mozer, and Christopher Pal. Systematic evaluation of causal discovery in visual model based reinforcement learning. arXiv preprint arXiv:2107.00848, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 554 + ], + "type": "text", + "content": "[21] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 621 + ], + "type": "text", + "content": "[22] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In International Conference on Machine Learning, pages 5815-5826. PMLR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "text", + "content": "[23] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In International Conference on Learning Representations, 2021. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[24] Ananya Kumar, Aditi Raghunathan, Tengyu Ma, and Percy Liang. Calibrated ensembles: A simple way to mitigate id-ood accuracy tradeoffs. In NeurIPS 2021 Workshop on" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20309" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "text", + "content": "Distribution Shifts: Connecting Methods and Applications, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[25] Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. Advances in Neural Information Processing Systems, 34:6155-6170, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "type": "text", + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "type": "text", + "content": "[28] Chaochao Lu, Yuhuai Wu, José Miguel Hernández-Lobato, and Bernhard Schölkopf. Invariant causal representation learning for out-of-distribution generalization. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 319 + ], + "type": "text", + "content": "[29] John P Miller, Rohan Taori, Aditi Raghunathan, Shiori Sagawa, Pang Wei Koh, Vaishaal Shankar, Percy Liang, Yair Carmon, and Ludwig Schmidt. Accuracy on the line: on the strong correlation between out-of-distribution and indistribution generalization. In International Conference on Machine Learning, pages 7721-7735. PMLR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "text", + "content": "[30] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512-523, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 286, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 286, + 375 + ], + "type": "text", + "content": "[31] Judea Pearl. Causality. Cambridge university press, 2009. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 378, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 443 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "text", + "content": "[33] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[35] Nino Scherrer, Anirudh Goyal, Stefan Bauer, Yoshua Bengio, and Nan Rosemary Ke. On the generalization and adaption performance of causal models. arXiv preprint arXiv:2206.04620, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[36] Bernhard Scholkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the IEEE, 109:612-634, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[37] Adarsh Subbaswamy, Peter Schulam, and Suchi Saria. Preventing failures due to dataset shift: Learning predictive models that transport. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 3118-3127. PMLR, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[38] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring ro" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 562 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "bustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33:18583-18599, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[39] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[40] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "type": "text", + "content": "[41] Moritz Willig, Matej Zečević, Devendra Singh Dhami, and Kristian Kersting. Can foundation models talk causality? arXiv preprint arXiv:2206.10591, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 241, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 318 + ], + "type": "text", + "content": "[42] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 319, + 545, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 395 + ], + "type": "text", + "content": "[43] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 397, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 545, + 452 + ], + "type": "text", + "content": "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 453, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 453, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 453, + 545, + 507 + ], + "type": "text", + "content": "[45] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 509, + 545, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 545, + 562 + ], + "type": "text", + "content": "[46] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20310" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_content_list.json b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..891933a73c427f1330fa423a7c18ab0b478a53f9 --- /dev/null +++ b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_content_list.json @@ -0,0 +1,1615 @@ +[ + { + "type": "text", + "text": "Masked Jigsaw Puzzle: A Versatile Position Embedding for Vision Transformers", + "text_level": 1, + "bbox": [ + 80, + 130, + 888, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Ren $^{1,2*}$ Yahui Liu $^{2*}$ Yue Song $^{2}$ Wei Bi $^{3}$ Rita Cucchiara $^{4}$ Nicu Sebe $^{2}$ Wei Wang $^{5\\dagger}$ $^{1}$ University of Pisa, Italy $^{2}$ University of Trento, Italy \n $^{3}$ Tencent AI Lab, China $^{5}$ Beijing Jiaotong University, China \n $^{4}$ University of Modena and Reggio Emilia, Italy", + "bbox": [ + 117, + 178, + 857, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 285, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Position Embeddings (PEs), an arguably indispensable component in Vision Transformers (ViTs), have been shown to improve the performance of ViTs on many vision tasks. However, PEs have a potentially high risk of privacy leakage since the spatial information of the input patches is exposed. This caveat naturally raises a series of interesting questions about the impact of PEs on accuracy, privacy, prediction consistency, etc. To tackle these issues, we propose a Masked Jigsaw Puzzle (MJP) position embedding method. In particular, MJP first shuffles the selected patches via our block-wise random jigsaw puzzle shuffle algorithm, and their corresponding PEs are occluded. Meanwhile, for the nonoccluded patches, the PEs remain the original ones but their spatial relation is strengthened via our dense absolute localization regressor. The experimental results reveal that 1) PEs explicitly encode the 2D spatial relationship and lead to severe privacy leakage problems under gradient inversion attack; 2) Training ViTs with the naively shuffled patches can alleviate the problem, but it harms the accuracy; 3) Under a certain shuffle ratio, the proposed MJP not only boosts the performance and robustness on large-scale datasets (i.e., ImageNet-1K and ImageNet-C, -A/O) but also improves the privacy preservation ability under typical gradient attacks by a large margin. The source code and trained models are available at https://github.com/yh1leo/MJP.", + "bbox": [ + 76, + 318, + 473, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 702, + 209, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformers [38] demonstrated their overwhelming power on a broad range of language tasks (e.g., text classification, machine translation, or question answering [22, 38]), and the vision community follows it closely and extends it for vision tasks, such as image classification [7,37], object detection [2, 51], segmentation [47], and image generation [3, 24]. Most of the previous ViT-based methods focus on designing different pre-training objectives [9, 11, 12] or variants of self-attention mechanisms [28, 39, 40]. By contrast, PEs", + "bbox": [ + 75, + 727, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/187412b629ee88a0d519ae7486148841bf1c2706c67e299ff596f60225b9fcdd.jpg", + "image_caption": [ + "Figure 1. Low-dimensional projection of position embeddings from DeiT-S [37]. (a) The 2D UMAP projection, it shows that reverse diagonal indices have the same order as the input patch positions. (b) The 3D PCA projection, it also shows that the position information is well captured with PEs. Note that the embedding of index 1 (highlighted in red) corresponds to the [CLS] embedding that does not embed any positional information." + ], + "image_footnote": [], + "bbox": [ + 503, + 284, + 694, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74102392c2115e8f800f0639c893106513209ebac9fc0455140eaa6551e6c9e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 282, + 893, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "receive less attention from the research community and have not been well studied yet. In fact, apart from the attention mechanism, how to embed the position information into the self-attention mechanism is also one indispensable research topic in Transformer. It has been demonstrated that without the PEs, the pure language Transformer encoders (e.g., BERT [6] and RoBERTa [25]) may not well capture the meaning of positions [43]. As a consequence, the meaning of a sentence can not be well represented [8]. Similar phenomenon of PEs could also be observed in vision community. Dosovitskiy et al. [7] reveals that removing PEs causes performance degradation. Moreover, Lu et al. [29] analyzed this issue from the perspective of user privacy and demonstrated that the PEs place the model at severe privacy risk since it leaks the clues of reconstructing sequential patches back to images. Hence, it is very interesting and necessary to understand how the PEs affect the accuracy, privacy, and consistency in vision tasks. Here the consistency means whether the predictions of the transformed/shuffled image are consistent with the ones of the original image.", + "bbox": [ + 496, + 535, + 893, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To study the aforementioned effects of PEs, the key is to figure out what explicitly PEs learn about positions from input patches. To answer this question, we project the high-dimensional PEs into the 2D and 3D spaces using Uni-", + "bbox": [ + 496, + 839, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Email: {bin.ren, yahui.liu} @unitn.it", + "bbox": [ + 94, + 875, + 401, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author. Email: wei.wang@bjtu.edu.cn", + "bbox": [ + 96, + 887, + 383, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20382", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b5c1df59a6fd2e20c0cfba72b50c21c339b6fcad9fd5fd3999bd08718f6f806.jpg", + "image_caption": [ + "(a) $\\mathbf{x}$" + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 277, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e8cf282277e3393e50d87cd69d69998ac94c48fb332535d84a129b64afc6d4f2.jpg", + "image_caption": [ + "(b) $\\widetilde{\\mathbf{x}} (\\gamma = 1)$" + ], + "image_footnote": [], + "bbox": [ + 281, + 88, + 470, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4b91c42dd0362a306dfa83289a418a28a3c731bd986ca25fa9ee599beea44186.jpg", + "image_caption": [ + "(c) $\\widetilde{\\mathbf{x}} (\\gamma = 0.4)$", + "Figure 2. (a) The original input patches; (b) Totally random shuffled input patches; (c) Partially random shuffled input patches; (d) An overview of the proposed MJP. Note that we show the random shuffled patches and its corresponding unknown position embedding in green and the rest part in blue. DAL means the self-supervised dense absolute localization regression constraint." + ], + "image_footnote": [], + "bbox": [ + 475, + 88, + 666, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/513b953e5d3546cd6099861512244379cb0a5ac27e1481a30b0f4db305f0da6a.jpg", + "image_caption": [ + "(d) Overview of our MJP" + ], + "image_footnote": [], + "bbox": [ + 666, + 88, + 875, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "form Manifold Approximation & Projection (UMAP) [30] and PCA, respectively. Then for the first time, we visually demonstrate that the PEs can learn the 2D spatial relationship very well from the input image patches (the relation is visualized in Fig. 1). We can see that the PEs are distributed in the same order as the input patch positions. Therefore, we can easily obtain the actual spatial position of the input patches by analyzing the PEs. Now it explains why PEs can bring the performance gain for ViTs [7]. This is because the spatial relation in ViTs works similar as the inherent intrinsic inductive bias in CNNs (i.e., it models the local visual structure) [45]. However, these correctly learned spatial relations are unfortunately the exact key factor resulting in the privacy leakage [29].", + "bbox": [ + 75, + 308, + 472, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these observations, one straightforward idea to protect the user privacy is to provide ViTs with the randomly transformed (i.e., shuffled) input data. The underlying intuition is that the original correct spatial relation within input patches will be violated via such a transformation. Therefore, we transform the previous visually recognizable input image patches $\\mathbf{x}$ shown in Fig. 2(a) to its unrecognizable counterpart $\\widetilde{\\mathbf{x}}$ depicted in Fig. 2(b) during training. The experimental results show that such a strategy can effectively alleviate the privacy leakage problem. This is reasonable since the reconstruction of the original input data during the attack is misled by the incorrect spatial relation. However, the side-effect is that this leads to a severe accuracy drop.", + "bbox": [ + 75, + 521, + 470, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, we noticed that such a naive transformation strategy actually boosts the consistency [31, 34, 44] albeit the accuracy drops. Note that here the consistency represents if the predictions of the original and transformed (i.e., shuffled) images are consistent. Given the original input patches $\\mathbf{x}$ and its corresponding transformed (i.e., shuffled) counterpart, we say that the predictions are consistent if $\\arg \\max P(\\mathcal{F}(\\mathbf{x})) = \\arg \\max P(\\mathcal{F}(\\widetilde{\\mathbf{x}}))$ , where $\\mathcal{F}$ refers to the ViT models, and $P$ denotes the predicted logits.", + "bbox": [ + 75, + 718, + 470, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These observations hint that there might be a trade-off solution that makes ViTs take the best from both worlds (i.e., both the accuracy and the consistency). Hence, we pro", + "bbox": [ + 75, + 854, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pose the Masked Jigsaw Puzzle (MJP) position embedding method. Specifically, there are four core procedures in the MVP: (1) We first utilize a block-wise masking method [1] to randomly select a partial of the input sequential patches; (2) Next, we apply jigsaw puzzle to the selected patches (i.e., shuffle the orders); (3) After that, we use a shared unknown position embedding for the shuffled patches instead of using their original PEs; (4) To well maintain the position prior of the unshuffled patches, we introduce a dense absolute localization (DAL) regressor to strengthen their spatial relationship in a self-supervised manner. We simply demonstrate the idea of the first two procedures in Fig. 2(c), and an overview of the proposed MVP method is available in Fig. 2(d). In summary, our main contributions are:", + "bbox": [ + 496, + 308, + 893, + 520 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate that although PEs can boost the accuracy, the consistency against image patch shuffling is harmed. Therefore, we argue that studying PEs is a valuable research topic for the community.", + "- We propose a simple yet efficient Masked Jigsaw Puzzle (MJP) position embedding method which is able to find a balance among accuracy, privacy, and consistency.", + "- Extensive experimental results show that MJP boosts the accuracy on regular large-scale datasets (e.g., ImageNet-1K [32]) and the robustness largely on ImageNet-C [18], -A/O [19]. One additional bonus of MJP is that it can improve the privacy preservation ability under typical gradient attacks by a large margin." + ], + "bbox": [ + 500, + 523, + 893, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 744, + 640, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision Transformers", + "text_level": 1, + "bbox": [ + 500, + 771, + 694, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers [38], originally designed for Nature Language Processing (NLP) tasks, have recently shown promising performance on computer vision tasks [13, 22]. Benefiting from the strong representation power of modelling global relations between image patches, Vision Transformers (ViTs) [7] have achieved superior performance than their counterpart CNNs on image classification and various other", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20383", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "downstream tasks (e.g., object detection [2, 51], object re-identification [17], dense prediction [41, 42, 46, 50] and image generation [3, 4, 20, 24]).", + "bbox": [ + 75, + 90, + 470, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a core module in ViTs, multi-head self-attention (MSAs) [7, 38] aggregate sequential tokens with normalized attentions as: $z_{j} = \\sum_{i}\\mathrm{Softmax}(\\frac{QK}{\\sqrt{d}})_iV_{i,j}$ where $Q$ , $K$ and $V$ are query, key and value matrices, respectively. $d$ is the dimension of query and key, and $z_{j}$ is the $j$ -th output token. In theory, when the position information is not considered, the outputs of MSAs should be strictly invariant to the input sequence order (i.e., position-insensitivity). This indicates that a visually recognizable image can be transformed into its unrecognizable counterpart by permuting the order of image patches while maintaining the performance delivered by ViTs compared with the ViTs trained on the original non-permuted image. However, the usage of PEs hinders such implementations, where the outputs of the ViTs vary dramatically with the mentioned naive transformations shown in Fig. 2(b). In this work, our main focus is to explicitly figure out what PEs actually learn from input patches about positions, and how the PEs affect the accuracy, privacy, and the consistency properties of ViTs.", + "bbox": [ + 75, + 137, + 473, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Position Embeddings", + "text_level": 1, + "bbox": [ + 76, + 438, + 274, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Transformer networks, both the attention and the (individual token based) feed-forward layers are permutation invariant when the position information is not considered. In this way, the spatial relationships between image patches could not be modeled as the position information is completely discarded. As compensation, PEs are naturally introduced into ViTs to provide information about the token order during the learning process, since it offers possibilities for dependency modeling between elements at different positions. For example, previous works [10,33,38] indicated that PEs are useful to give the model a sense of which portion of the sequence in the input/output it is currently dealing with. Inspired by this, some works [15,21,23,27,35] showed diverse application scenarios that benefit from the usage of suitable PEs. In addition, Chu et al. [5] proposed correlating the PEs with their local neighborhood of the input sequence. Liu et al. [26] proposed to enhance the spatial prior (i.e., relative localization) in the final content embedding to indirectly enrich the inductive bias. Obviously, although these methods enhance the position information learnt by PEs, they indeed degenerate the position-insensitive property of MSAs.", + "bbox": [ + 75, + 460, + 470, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Especially, Wang et al. [43] revealed that Transformer encoders (e.g., BERT [6] and RoBERTa [25]) may not well capture the meaning of positions (absolute and relative positions). They showed that Transformer encoders learn the local position information that can only be effective in masked language modeling. In contrast, there does not exist such a similar \"masked language modeling\" procedure in ViTs (and VTs) in the typically supervised pre-training. Moreover, Lu", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "et al. [29] revealed that the learnable PEs place the model at severe privacy risk, which leaks the clues of reconstructing sequential patches to images. In this paper, we dive into the usage of PEs and propose an efficient position embedding method, MJP, to improve the position-insensitive property of ViTs without hurting the positive effects of PEs.", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries: 2D/3D Spatial Priors in PEs", + "text_level": 1, + "bbox": [ + 498, + 196, + 883, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although numerous works claim that the PEs can learn the 2-D spatial relationship of image patches, this claim has not been demonstrated visually or mathematically. To visualize the concrete relation of image patches captured in the high-dimensional position embedding, we project them into the 2-D and 3-D spaces using Uniform Manifold Approximation and Projection (UMAP) [30] and PCA, respectively. Fig. 1 displays the projections of PEs from DeiT-S [37]. Note that [7] only shows the cosine similarity between the PEs, which is not exactly the 2D spatial information mentioned in our paper. Here, we explicitly specify the spatial relationship as the relative distance in the 2D/3D coordinate space. As shown in Fig. 1, the spatial relationship can be projected into a 2D/3D coordinate system and indicate the localizations of these embeddings. For the UMAP, the projected positional embedding emerges as grid-like structures with the distances between each point roughly the same, which is coherent with the relation of embedded image patches. For the PCA, the position embedding that corresponds to neighboring image patches groups together and aggregates to the stick-like form. This phenomenon demonstrates that the spatial relationship is indeed learned by the PEs. Moreover, this also implies that the spatial relationship learned in the high-dimensional space still manifests in the low-dimensional space.", + "bbox": [ + 496, + 222, + 893, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The learnable PEs, which work as a lookup table of a dictionary, maps the 1-dimensional data into the sparse high-dimensional space. By nature, it is a sparse large matrix. Dimensionality reduction techniques can capture the structural information of sparse matrices using low dimensionalities. For the PCA, the amount of information retained in the projection can be measured by the ratio of explained variance. Formally, we have the definition as follows:", + "bbox": [ + 496, + 585, + 893, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Explained Variance) Let $\\mathbf{P}$ and $\\mathbf{V}$ denote the data matrix and the PCA projection matrix. The ratio of explained variance that $\\mathbf{PV}$ accounts for is defined as $\\sum \\sigma (\\mathbf{PV})^2 /\\sum \\sigma (\\mathbf{P})^2$ where $\\sigma (\\cdot)$ denotes the singular value.", + "bbox": [ + 498, + 712, + 893, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, we observe that the 3-dimensional PCA projection explains $54.6\\%$ of the total variance of the DeiT-S embedding matrix. Given that a large amount of information can be captured by the low-dimensional projection, we propose that explicitly enforcing low-dimensional positional prior can help the positional learning in the high-dimensional space, which might accelerate the convergence rate of training and improve the performance (See Sec. 4.2).", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20384", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Block-wise Random Jigsaw Puzzle Shuffle" + ], + "code_body": "Input: Input image: $\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times C}$ Shuffle Ratio: $\\gamma$ Patch Size: $P$ \nOutput: Shuffled image patches: $\\widetilde{\\pmb{x}}_p$ \n1: $\\mathbf{x}_p\\in \\mathbb{R}^{N\\times (P^2\\cdot C)}\\gets$ Patchlize(x,P) \n2: $\\mathbf{m}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}\\gets$ BinaryInitialize(xp,0) \n3: $\\widetilde{\\mathbf{m}}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}\\gets$ BlockwiseMask(m,γ) [25] \n4: $\\widetilde{\\pmb{x}}_p\\in \\mathbb{R}^{N\\times (P^2\\cdot C)}\\gets$ Jigsaw Puzzle(xp, $\\widetilde{\\mathbf{m}}$ 5: return $\\widetilde{\\pmb{x}}_p$", + "bbox": [ + 76, + 109, + 473, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 282, + 166, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the design principle of MSAs, the outputs of MSAs should be entirely position-agnostic. However, PEs hinder such a property because it can learn the strong 2-D spatial relation of the input patches (as we demonstrated in Sec. 3). To this end, we propose the block-wise random jigsaw puzzle shuffle algorithm (See Alg. 1) to transform the input patches with different shuffle ratios $\\gamma$ for intermingling the original correct spatial relation.", + "bbox": [ + 75, + 306, + 468, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we experimentally demonstrate that the totally shuffled strategy (i.e., $\\gamma = 1.0$ ) will degenerate the accuracy a lot albeit the consistency increases. As a remedy, we only shuffle portion of the sequence patches, and strengthen the spatial relation of the rest part with a dense absolute localization regression. Finally, a versatile position embedding method MJP is proposed. The detailed analysis of each module is available in the ablation study in Sec. 5.3. In the following sub-sections, we will introduce the Jigsaw Puzzle Shuffle method (Sec. 4.1), the spatial relation strengthen method (Sec. 4.2), and our final MJP method (Sec. 4.3).", + "bbox": [ + 75, + 429, + 470, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Block-wise Random Jigsaw Puzzle Shuffle", + "text_level": 1, + "bbox": [ + 76, + 603, + 434, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given an input image $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}$ , we first reshape it into a sequence of flattened 2D patches $\\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot C)}$ , where $(H, W)$ is the resolution of the original image, $C$ is the number of channels, $(P, P)$ is the resolution of each image patch, and then we have $N = HW / P^2$ , which denotes the number of patches. Then instead of directly applying block-wise masking method [1] to the image, we first initialize a binary mask matrix $\\mathbf{m} \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P}}$ with the same size as the image patches in $\\mathbf{x}_p$ . Next, we use [1] to update the binary mask $\\mathbf{m}$ in which the masked positions will be set to 1 and the rest untouched positions remains 0. The hyper-parameter $\\gamma$ is used to control the ratio of selected positions. After that, a jigsaw puzzle shuffle operation is applied to $\\mathbf{x}_p$ conditioned on the updated binary mask $\\widetilde{\\mathbf{m}}$ .", + "bbox": [ + 75, + 625, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we get the shuffled patch sequence $\\widetilde{\\mathbf{x}}_p$ , where the shuffled patches are actually $\\{\\mathbf{x}_p^i |\\widetilde{\\mathbf{m}}_i = 1\\}$ . Notably, in [1], the patches covered by the sampled mask are not visible to the encoder module, while in our method, the masking", + "bbox": [ + 75, + 840, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "strategy [1] is only used to mask out the positions/indices of the selected patches. These patches are still visible to the encoder module and they are randomly shuffled in a jigsaw puzzle manner. One intuitive shuffled toy example is shown in Fig. 2(c), where the green part means the selected shuffled region while the blue part remains unchanged. Actually, this algorithm shares a similar idea with the data augmentation method as the training images are always changing as their patches are randomly shuffled during each iteration.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Strengthening Spatial Prior in PEs", + "text_level": 1, + "bbox": [ + 498, + 234, + 802, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Liu et al. [26] noticed that by enhancing the 2-D spatial information of the output embeddings of the last layer of ViTs, the training convergence speed can be accelerated. Inspired by their work, we observe that similar gains can be obtained by applying a low-dimensional spatial prior in PEs. Different from the dense relative localization constraint in [26] which samples relative pairs from the whole output sequential embeddings, we propose a much simpler dense absolute localization (DAL) regression method to directly use self-supervised absolute location to enhance the spatial information in PEs, which avoids sampling relative pairs.", + "bbox": [ + 496, + 257, + 892, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the PEs capture the absolute position of the input patches, to some extent, the position information could be reconstructed via a reversed mapping function $g(\\cdot):\\mathcal{E}\\to \\mathcal{P}$ where $\\mathcal{E}$ and $\\mathcal{P}$ are embedding space and position space, respectively. Given that PEs have one-to-one correspondence with the sequential image patches, we can reshape them into $\\mathbf{E}_{\\mathrm{pos}}\\in \\mathbb{R}^{K\\times K\\times D}$ , where $K$ refers to height/weight of the grid and $D$ refers to latent vector size. Then we can compute the reverse mapping from $\\mathcal{E}\\rightarrow \\mathcal{P}$ via:", + "bbox": [ + 496, + 424, + 892, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\widetilde {i}, \\widetilde {j}\\right) ^ {T} = g \\left(\\mathbf {E} _ {\\text {p o s}} ^ {i, j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 565, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $(\\widetilde{i},\\widetilde{j})^T$ is the predicted patch position, and $\\mathbf{E}_{\\mathrm{pos}}^{i,j}$ is the position embedding of the patch $(i,j)$ in the $K\\times K$ grid. The dense absolute localization (DAL) loss is:", + "bbox": [ + 496, + 593, + 893, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D A L}} = \\mathbb {E} _ {\\mathbf {E} _ {\\text {p o s}}, 1 \\leq i, j \\leq K} ^ {i, j} [ \\| (i, j) ^ {T} - (\\widetilde {i}, \\widetilde {j}) ^ {T} \\| _ {1} ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 645, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the expectation is computed by averaging the $\\ell_1$ loss between the correspond $(i,j)^T$ and $(\\widetilde{i},\\widetilde{j})^T$ . Then, $\\mathcal{L}_{\\mathrm{DAL}}$ is added to the standard cross-entropy loss $(\\mathcal{L}_{\\mathrm{CE}})$ of the native ViTs. The final loss is: $\\mathcal{L}_{\\mathrm{all}} = \\mathcal{L}_{\\mathrm{CE}} + \\lambda \\mathcal{L}_{\\mathrm{DAL}}$ , where we set $\\lambda = 0.01$ for all experiments. Note that the mapping function can be either linear or nonlinear. Throughout this work, we mainly discuss three implementations, including non-parametric PCA, learnable linear (LN), and nonlinear (NLN) projection layers. They will be discussed in details in Sec. 5.3.", + "bbox": [ + 496, + 672, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. MJP Position Embedding", + "text_level": 1, + "bbox": [ + 500, + 832, + 733, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The main goal of MJP position embedding in our work is to enhance the consistency (i.e., position-insensitive property) of ViTs and preserve the accuracy (keeping the spatial", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 The pipeline of the proposed MJP." + ], + "code_body": "Input: Input image: $\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times C}$ Shuffle Ratio: $\\gamma$ ; Patch Size: $P$ 1: $\\widetilde{\\pmb{x}}_p\\gets Alg.1(\\mathbf{x},P,\\gamma)$ // 1st & 2nd procedures 2: $\\mathbf{E}_{\\mathrm{unk}}(\\widetilde{\\pmb{x}}_p)$ // 3rd procedure 3: $\\mathbf{DAL}(\\mathbf{x} - \\mathbf{x}\\cap \\widetilde{\\pmb{x}}_p)$ // 4th procedure, only for train", + "bbox": [ + 76, + 109, + 413, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "relation modeling functionality of PEs) for the standard classification tasks. Usually, with original input image patches $\\mathbf{x}$ we can formulate the function of the input layer which is located before the Transformer block as follows:", + "bbox": [ + 75, + 210, + 470, + 270 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {C L S}}; \\mathbf {x} _ {p} ^ {1} \\mathbf {E}; \\mathbf {x} _ {p} ^ {2} \\mathbf {E}, \\dots , \\mathbf {x} _ {p} ^ {N} \\mathbf {E} \\right] + \\mathbf {E} _ {\\text {p o s}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 281, + 468, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{E} \\in \\mathbb{R}^{(P^2 \\cdot C) \\times D}$ is a trainable linear projection layer and $\\mathbf{x}_p^k\\mathbf{E}$ refers to the output of linear projection (i.e., the patch embeddings), and $\\mathbf{E}_{\\mathrm{pos}} \\in \\mathbb{R}^{(N + 1) \\times D}$ refers to PEs (the additional one is applied to the [CLS] embedding).", + "bbox": [ + 75, + 311, + 468, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we apply the proposed block-wise random jigsaw puzzle shuffle algorithm to $\\mathbf{x}_p$ and produce the transformed patch sequences $\\widetilde{\\mathbf{x}}_p$ with the shuffle ratio $\\gamma$ . In this scenario, if we maintain the original position embedding sequence, it leads to a mismatch issue between the shuffled patches and the position embedding sequence. Therefore, we introduce a shared unknown position embedding to the shuffled positions to alleviate the mismatching issue. With the corresponding updated mask $\\widetilde{\\mathbf{m}}$ from Alg. 1 (1st & 2nd procedures), we propose our MJP PEs:", + "bbox": [ + 75, + 376, + 468, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathbf {E}} _ {\\text {p o s}} ^ {i} = \\left\\{ \\begin{array}{l l} \\mathbf {E} _ {\\text {p o s}} ^ {i}, & \\text {i f} \\widetilde {\\mathbf {m}} _ {i} = 0 \\\\ \\mathbf {E} _ {\\text {u n k}}, & \\text {i f} \\widetilde {\\mathbf {m}} _ {i} = 1 \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 537, + 468, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{E}_{\\mathrm{unk}} \\in \\mathbb{R}^{1 \\times D}$ refers to a share learnable embedding (i.e., unknown position embedding). It represents that the image patch in this position has random permutation and its position should be occluded (3rd procedure). $\\mathbf{E}_{\\mathrm{pos}}$ is the original position embedding for the rest image patches.", + "bbox": [ + 75, + 588, + 468, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides, we revisit the remaining PEs (i.e., corresponding to the non-selected patches) and apply low-dimensional prior on them. The low-dimensional prior is imposed by the proposed DAL (Sec. 4.2) regression method for strengthening the spatial relation (4th procedure). A toy illustration of the proposed MJP is shown in Fig. 2(d), where the green color represents the randomly permuted patches and its corresponding unknown PEs, while the blue color indicates the rest regular patches and its related original PEs. We also formalize these procedures as an algorithm in Alg. 2.", + "bbox": [ + 75, + 665, + 468, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, we replace the input layer with a new formulation:", + "bbox": [ + 96, + 816, + 464, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathbf {z}} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {C L S}}; \\widetilde {\\mathbf {x}} _ {p} ^ {1} \\mathbf {E}; \\widetilde {\\mathbf {x}} _ {p} ^ {2} \\mathbf {E}, \\dots , \\widetilde {\\mathbf {x}} _ {p} ^ {N} \\mathbf {E} \\right] + \\widetilde {\\mathbf {E}} _ {\\text {p o s}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 840, + 468, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the following procedures and modules are exactly the same as the ones in the original ViTs.", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1f9a78a72085550f74c5434903573a5ab39bafabe1653d58f015618b0eaadb57.jpg", + "table_caption": [ + "Table 1. Comparisons of different backbones on ImageNet-1K classification. Note that the image size here are all set to $224 \\times 224$ ." + ], + "table_footnote": [], + "table_body": "
MethodParam. Top-1 Acc.↑ Diff.Norm.↓ Consistency
ResNet-50 [16]2579.311.7751.5
ResNet-50 + MJP2579.47.1169.3
DeiT-S [37]2279.816.2164.3
DeiT-S + MJP2280.58.9682.9
Swin-T [28]2981.315.4941.5
Swin-T + MJP2981.312.3666.9
", + "bbox": [ + 506, + 118, + 885, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 228, + 630, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow the typical supervised pre-training procedure, where all the compared models are trained on ImageNet1K [32] to show the capacity of our proposed MJP method. We also benchmark the proposed MJP method on ImageNet1K, which contains 1.28M training images and 50K validation images of 1,000 classes. The training details mostly follow the training protocols from Touvron et al. [37].", + "bbox": [ + 496, + 253, + 893, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Regular ImageNet-1K training", + "text_level": 1, + "bbox": [ + 498, + 369, + 772, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We mainly compare with three typical existing methods, including two state-of-the-art Visual Transformers (i.e., DeiT [37] and Swin [28]) and one widely-used CNN-based ResNet-50 [16]. All these methods are of comparable sizes (i.e., less than 30M parameters). Besides the common Top-1 accuracy (Top-1 Acc.), we also report another two evaluation metrics (i.e. Diff. Norm. and Consistency) to show the position invariance of a model to the jigsaw puzzle transformation. For a given input image $\\mathbf{x}$ , we collect its counterpart $\\widetilde{\\mathbf{x}}$ by applying masked jigsaw puzzle to shuffle a portion of selected patches. Next, we calculate the difference $\\ell_2$ -norm (i.e., Diff. Norm.) between the [CLS] embedding inferred from $\\mathbf{x}$ and $\\widetilde{\\mathbf{x}}$ :", + "bbox": [ + 496, + 393, + 893, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\mathcal {F} _ {\\mathrm {C L S}} (\\mathbf {x}) - \\mathcal {F} _ {\\mathrm {C L S}} (\\widetilde {\\mathbf {x}}) \\right\\| _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 594, + 890, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For Consistency, we measure how many classification results stay the same (\\%) given the perturbation:", + "bbox": [ + 496, + 625, + 893, + 655 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathbf {x}} \\left[ \\mathbb {I} \\left\\{\\arg \\max P (\\mathcal {F} (\\mathbf {x})) = \\arg \\max P (\\mathcal {F} (\\widetilde {\\mathbf {x}})) \\right\\} \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 667, + 890, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$P(\\mathcal{F}(\\mathbf{x}))$ denotes the predicted classification probability of image $\\mathbf{x}$ , and $\\arg \\max P(\\mathcal{F}(\\mathbf{x}))$ represents the predicted class index. For a fair comparison, we use $\\gamma = 0.15$ to create $\\widetilde{\\mathbf{x}}$ in Table 1. According to these results, the proposed MJP does not have a negative effect on the top-1 accuracy (MJP even brings a marginal improvement for DeiT-S). More importantly, it improves Diff. Norm. and Consistency by a large margin. Besides, MJP works well in the variants of ViTs (e.g., Swin [28]), which shows a good potential generalization ability. Note that we add MJP to ResNet50 by only shuffling the image patches with Alg. 1 (No PEs). For Swin-T, we add MJP to the absolute PEs of Swin-T.", + "bbox": [ + 496, + 695, + 893, + 876 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 517, + 886, + 839, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20386", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/edfe94efef2ab9d518d4ef5f73138de4ef5ac5f4e16cf4b10710f6512243d02e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 75, + 316, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/94b9817bd27d7a2a49267dad77b19b2a1a115dc3340f7a9e6a0c810ac77ec96f.jpg", + "image_caption": [ + "(b)", + "Figure 3. Ablation on the mask ratio $\\gamma$ during inference: (a) comparisons among ViT-S, DeiT-S and our method (trained with $\\gamma = 0.03$ ); (b) comparisons between Swin-T and our method (trained with $\\gamma = 0.03$ ); (c) comparisons of our method on DeiT-S trained with different $\\gamma$ ." + ], + "image_footnote": [], + "bbox": [ + 370, + 75, + 594, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d99d083d74bafd54874aea96b7c1df4fab1377996c00b8e9d4dd94c1508b8d2.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 653, + 75, + 875, + 218 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/381c2bc5934fd57ca90bbba014f806805817aa8573cf9f29703dc705ce7647f5.jpg", + "table_caption": [ + "Table 2. Comparisons on robustness to common corruptions and adversarial examples." + ], + "table_footnote": [], + "table_body": "
MethodImageNet-CImageNet-AImageNet-O
mCE ↓Acc ↑AURRA ↑AUPR ↑
DeiT-S54.619.225.120.9
DeiT-S + MJP40.7821.629.822.6
", + "bbox": [ + 83, + 297, + 464, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Robustness on Challenging Sets", + "text_level": 1, + "bbox": [ + 76, + 380, + 356, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides the strength on standard classification, we also observe an auxiliary benefit of the proposed MJP on robustness. ImageNet-C [18] benchmarks a classifier's robustness to common corruptions2. The mean Corruption Error (mCE) is used to measure the generalization of a model at corrupted distributions (the lower the better).", + "bbox": [ + 75, + 405, + 470, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ImageNet-A/O [19] focus on adversarial examples, and it enables us to either test image classification performance when the input data distribution shifts (i.e., Acc and AU-RRA), or test out-of-distribution detection performance when the label distribution shifts (i.e., AUPR) $^3$ . As shown in Table 2, the proposed MJP not only boosts the accuracy on the standard evaluation on ImageNet-1K validation set but also improves the robustness largely on the adversarial samples.", + "bbox": [ + 75, + 496, + 470, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The underlying reason might be that MJP enforces the ViTs aware of both local and global context features, and it helps ViTs to get rid of some unnecessary sample-specific local features during the training. This has been verified by the visualization maps (see Fig. 3 of our Supp. Mat.).", + "bbox": [ + 75, + 632, + 468, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Analysis", + "text_level": 1, + "bbox": [ + 76, + 720, + 250, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results with different MJP ratios. As shown in Table 3, we test different masking ratios used in the block-wise masking strategy during the training. Obviously, Diff. Norm. has the inverse tendencies compared to Consistency, where a smaller Diff. Norm. usually indicates a larger (better) consistency score. Comparing the accuracy trained with $\\gamma >0$ , it shows our model is not sensitive to different $\\gamma$ . In particular, a", + "bbox": [ + 75, + 744, + 468, + 851 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f3939951ac98fb651168cd5ba0652c0123391cca2957d0cbe19046a8d2da5531.jpg", + "table_caption": [ + "Table 3. Ablation study on the proposed MJP method trained with different masking ratio $\\gamma$ ." + ], + "table_footnote": [], + "table_body": "
MetricMasking Ratio
00.030.090.150.210.27
Top-1 Acc.80.080.580.380.480.280.3
Diff. Norm.16.568.966.365.234.393.97
Consistency64.082.988.190.592.393.1
", + "bbox": [ + 504, + 300, + 883, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "small ratio (e.g., $\\gamma = 0.03$ ) is sufficient for boosting the accuracy. In addition, a large ratio can reduce the Diff. Norm. and improves the consistency by a large margin. Moreover, the model trained with a larger $\\gamma$ is inclined to be more consistent as shown in Fig. 3 (c). In addition, we observe that the consistency keeps increasing when the mask ratio increases from $3\\%$ to $27\\%$ . However, further increasing the mask ratio will not bring consistency improvement and what is worse is that the accuracy marginally decreases.", + "bbox": [ + 496, + 411, + 893, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a model trained with a fixed $\\gamma$ , we also test its accuracy with different masking ratios during the inference. As shown in Fig. 3 (a) and (b), the performances of original models drop significantly when we shuffle more image patches. In contrast, our proposed method shows more consistent performances.", + "bbox": [ + 496, + 549, + 893, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of variants of MJP. We test several variants of the proposed MJP, including (1) Removing the PEs from the original DeiT-S; (2) SPP: shuffling both the $16 \\times 16$ patches and pixels within the patches, which is tested in MLP-Mixer [36] to validate the invariance to permutations; (3) JP: applying masked jigsaw puzzle to the sequence patches; (4) IDX: using an additional collection of embeddings to indicate the global indexes for the input patches; (5) UNK: replacing the PEs in the masked positions with a shared unknown position embedding; (6) DAL: jointly learning dense absolute localization regression in an self-supervised manner during the pretraining, where we provide PCA, linear (LN) and nonlinear (NLN) projections, respectively.", + "bbox": [ + 496, + 641, + 895, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4 shows the detailed ablation studies on the variants of our proposed MJP method. First, as expected, when we remove the PEs from the original DeiT-S model, the accuracy decreases by $2.3\\%$ , but the consistency achieves $100\\%$ . It", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2https://github.com/hendrycks/robustness", + "bbox": [ + 91, + 861, + 413, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://github.com/hendrycks/natural-adv-examples", + "bbox": [ + 78, + 875, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/10e5889814a275ebfdb31517ba78c7a30635c7a7a7b7e62e6e58669ed2d2c762.jpg", + "table_caption": [ + "Table 4. Ablation study on the variants of the proposed MJP." + ], + "table_footnote": [], + "table_body": "
MethodTop-1 Acc. ↑Consistency ↑
A: DeiT-S [37]79.864.3
B: A - PEs77.5 (-2.3)100.0
C: A + SPP [36]74.9 (-4.9)74.8
D: A + DAL (NLN)80.0 (+0.2)64.0
E: A + JP79.2 (-0.6)73.8
F: A + JP + IDX79.9 (+0.1)79.6
G: A + JP + UNK80.1 (+0.3)83.8
H: A + JP + UNK + DAL (PCA)79.9 (+0.1)83.4
I: A + JP + UNK + DAL (LN)80.0 (+0.2)83.8
J: A + JP + UNK + DAL (NLN)80.5 (+0.7)82.9
", + "bbox": [ + 83, + 108, + 460, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a6780d7bf30e61dbb646ffc547ec6ebca2ddd8fddce8cb0f5c51780bab1498f.jpg", + "image_caption": [ + "Figure 4. Distributions of accumulated eigenvalues of PEs." + ], + "image_footnote": [], + "bbox": [ + 101, + 282, + 439, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "verifies that the ViTs are naturally position-insensitive once without using PEs.", + "bbox": [ + 75, + 448, + 468, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observe that the previous SPP strategy harms the accuracy of the model (i.e., $-4.9\\%$ ), which indicates it is insufficient to simply shuffle both patches and pixels in the input image. As we expect, the usage of UNK embedding alleviates the confusion between shuffled and unshuffled positions, which boosts both the accuracy and consistency.", + "bbox": [ + 75, + 479, + 468, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we can find that using dense absolute localization regression on the unmasked PEs can marginally boost the accuracy. Although PCA does not involve more parameters to the model, it significantly increases the computation latency (e.g., $+46\\%$ with only 5 iterations). In general, it is better to use a nonlinear projection (e.g., a 3-layers MLP with negligible computation latency) to allow the learned PEs to aggregate more additional information.", + "bbox": [ + 75, + 573, + 468, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Informativeness of the PEs. For a position space $\\mathcal{P}$ (i.e., a 1-dim or 2-dim space), we may not need a high-dimension embedding space $\\mathcal{X}$ to model the positions. To measure the informativeness of the learned PEs, we apply singular value decomposition (SVD) to PEs and analyze their eigenvalue distributions. Fig. 4 plots the curves of accumulated energy/sum of top- $n$ eigenvalues versus the energy of all the eigenvalues. Supposing that we are using PCA to project the PEs of both matrices, to achieve the same explained variance ratio, our MJP needs more singular values (i.e., large dimensionality) than DeiT-S. This indicates that our positional embedding matrix is more informative. Similar observation is also revealed by Wang et al. [43].", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5bcb5f22693f22bacc98e7a6515e250e0c1b5c3baee7bacaca70845239b9c848.jpg", + "table_caption": [ + "Table 5. Comparisons on gradient leakage by analytic attack [29] with ImageNet-1K validation set, where we test (1) ViT-S, DeiT-S and our model in the setting (a); (2) ViT-S, DeiT-S and our model in the setting (b) (i.e., MJP with $\\gamma = 0.27$ ); (3) ablation on without (w/o) using $\\mathbf{E}_{\\mathrm{unk}}$ in setting (a); and (4) Our model in setting (c)." + ], + "table_footnote": [], + "table_body": "
ModelSet.Acc.↑MSE↑\\( FFT_{2D} \\)↑PSNR↓SSIM↓LPIPS↑
(1)ViT-S [7]78.1.0278.003919.27.5203.3623
DeiT-S [37]a79.8.0350.005718.94.5182.3767
DeiT-S (w/o PEs)77.5.0379.008220.22.5912.2692
DeiT-S+MJP80.5.1055.016611.52.4053.6545
(2)ViT-S [7]18.7.0327.001618.44.6065.2836
DeiT-S [37]b36.0.0391.002417.60.5991.3355
DeiT-S (w/o PEs)77.5.0379.002520.25.6655.2370
DeiT-S+MJP62.9.1043.005911.66.4493.6519
(3)DeiT-S+MJP (w/o)a40.6.1043.005911.66.4493.6519
(4)DeiT-S+MJPc62.9.1706.03388.07.0875.8945
", + "bbox": [ + 506, + 159, + 890, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Privacy Preservation", + "text_level": 1, + "bbox": [ + 500, + 319, + 696, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The fundamental principle of the gradient attack methods in federated learning is that each sample activates only a portion of content-related neurons in the deep neural networks, leading to one specific backward gradients for one related samples (i.e., 1-to-1 mapping). Based on such an observation, we argue that feeding ViTs with input patches with permuted sequences may intuitively mislead the attack. This is because now both the original and the transformed inputs may be matched to the same backward gradients (i.e., $n$ -to-1 mapping).", + "bbox": [ + 496, + 343, + 893, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate such an assumption, we utilize the public protocols4 to recover image with gradient updates in the privacy attack. In this privacy attack, we apply the Analytic Attack proposed in APRIL [29], which is designed for attacking the ViTs. We randomly sample 1K images from the validation set of ImageNet-1K (i.e., one image per category). To evaluate the anti-attack performance of a model, we introduce image similarity metrics to account for pixel-wise mismatch, including Mean Square Error (MSE), Peak Signal-to-Noise Ratio (PSNR), cosine similarity in the Fourier space $(\\mathrm{FFT}_{2\\mathrm{D}})$ , and Learned Perceptual Image Patch Similarity (LPIPS) [49]. Different from the evaluation in gradient attacks [14, 29, 48], we suppose a model is with better capacity of privacy preservation when the recovered images from its gradient updates are less similar to the ground truth images.", + "bbox": [ + 496, + 494, + 893, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given an image $\\mathbf{x}$ and its transformed (i.e., patch shuffled) version $\\tilde{\\mathbf{x}}$ , a ViT model $\\mathcal{M}$ , and automatic evaluation metrics $\\phi$ , we conduct three different settings for fair comparisons: (a) $\\phi(\\nabla \\mathcal{M}(\\mathbf{x}), \\mathbf{x})$ , (b) $\\phi(\\nabla \\mathcal{M}(\\tilde{\\mathbf{x}}), \\tilde{\\mathbf{x}})$ , and (c) $\\phi(\\nabla \\mathcal{M}(\\tilde{\\mathbf{x}}), \\mathbf{x})$ , where $\\nabla$ refers to recovering input image through gradient attacks. Table 5 shows the quantitative comparisons between our method and the original ViTs for batch gradient inversion on ImageNet-1K. APRIL [29] enables a viable, complete recovery of original images from the gradient updates of the original ViTs. However, it per", + "bbox": [ + 496, + 722, + 893, + 875 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4https://github.com/JonasGeiping/breaching", + "bbox": [ + 514, + 886, + 849, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20388", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8d90202ddd8e6794a1450fbfbec90fc3e2444738bc2403962663ff53e141104.jpg", + "image_caption": [ + "Figure 5. Visual comparisons on image recovery with gradient updates [29]. Our proposed DeiT-S+MJP model significantly outperforms the original ViT-S [7] and DeiT-S [37] models." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 851, + 310 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f2a64ff88b9b557e10d61a528e1c5419c88e75cc30c072e6c63fb7d8e905b2ea.jpg", + "table_caption": [ + "Table 6. Explained variance versus PCA projected dimensionality." + ], + "table_footnote": [], + "table_body": "
Projected Dimension34567
DeiT-S EV (%)54.6168.5577.9585.5490.74
DeiT-S+MJP EV (%)46.7458.3669.1078.1384.55
", + "bbox": [ + 80, + 378, + 460, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "forms worse in recovering from \"DeiT-S+MJP\", leading to best performances on all evaluation metrics and outperform others by a large margin.", + "bbox": [ + 75, + 435, + 468, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More surprisingly, our proposed method makes APRIL yield unrecognizable images and fail in recovering the details in the original images (i.e., noisy patches in the outputs), as shown in Fig. 5. The left four columns in Fig. 5 are tested on original images, where all PEs are standard and correspond to their patch embeddings. Meanwhile, the right four columns are tested with transformed ones, where the shuffled patches are with the shared unknown PEs. Both the visual and quantitative comparisons verify that our MJP alleviates the gradient leakage problem. We also notice that DeiT-S without using PEs is inclined to be at higher risk of privacy leakage (i.e., easier to be attacked by gradients). These promising results indicate that our MJP is a promising strategy to protect user privacy in federated learning.", + "bbox": [ + 75, + 482, + 472, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 707, + 194, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PCA Projected Dimensionality. Table 6 presents the explained variance (EV) of our DeiT-S+MJP and DeiT-S versus different projection dimension. A low dimensionality can explain a large amount of information, which proves that the embedding matrix is sparse in nature. Moreover, to achieve the same explained variance ratio, our DeiT-S+MJP needs a large dimensionality than DeiT-S. This indicates that the positional embedding matrix of DeiT-S+MJP is less sparse but more informative.", + "bbox": [ + 75, + 732, + 468, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Accuracy Vs. Shuffle Ratio. Intuitively, with the increase of the shuffle ratio from the proposed MJP method, the orig-", + "bbox": [ + 75, + 869, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inal intrinsic inductive bias will be undermined. However, from the experimental results, the performance of ViTs is actually boosted via the proposed MJP. To figure out the reason behind such counter-intuitive phenomenon, we visualize the last self-attention of our proposed method in Fig. 3 of our Supp. Mat. It shows that the attention heads of our method present more diverse and content-aware attentions than original DeiT-S. We think that it might be attributed to the proposed DAL, which strengthens the spatial information of the unmasked PEs in a more efficient manner. To this end, such learned content-aware attention becomes more meaningful and results in the better accuracy.", + "bbox": [ + 496, + 361, + 893, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 556, + 619, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we first visually demonstrate that PEs can explicitly learn the 2D spatial relationship from the input patch sequences. By feeding ViTs with transformed input, we identify the issue that PEs may weaken the position-insensitive property. Based on this observation, we propose an easy-to-reproduce yet effective Masked Jigsaw Puzzle (MJP) position embedding method to alleviate the conflict in PEs (preserving the consistency versus maintaining the accuracy). Experimental results show the proposed MJP can bring the the position-insensitive property back to ViTs without degenerating the accuracy on the large-scale dataset (i.e., ImageNet-1K). In a certain sense, the proposed MJP is also a data augmentation technique, which boosts the robustness to the common corruptions (e.g., ImageNet-C) and adversarial examples (e.g., ImageNet-A/O). Surprisingly, MJP can improve the privacy preservation capacity of ViTs under typical gradient attacks by a large margin, which may pilot a new direction for privacy preservation.", + "bbox": [ + 496, + 583, + 893, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was partly supported by the National AI Ph.D. for Society Program of Italy, and the EU H2020 project AI4Media (No. 951911).", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20389", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations (ICLR), 2022. 2, 4", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 1, 3", + "[3] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3", + "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning (ICML), 2020. 3", + "[5] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 3", + "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 1, 3", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2020. 1, 2, 3, 7, 8", + "[8] Philipp Dufter, Martin Schmitt, and Hinrich Schütze. Position information in transformers: An overview. Computational Linguistics, 48(3):733-763, 2022. 1", + "[9] Yuxin Fang, Benchcheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection. Advances in Neural Information Processing Systems, 34:26183-26197, 2021. 1", + "[10] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning (ICML), 2017. 3", + "[11] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 1", + "[12] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 244-253, 2019. 1", + "[13] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on visual transformer. arXiv, 2020. 2" + ], + "bbox": [ + 76, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ali Hatamizadeh, Hongxu Yin, Holger Roth, Wenqi Li, Jan Kautz, Daguang Xu, and Pavlo Molchanov. Gradvit: Gradient inversion of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7", + "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[17] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[18] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 2, 6", + "[19] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6", + "[20] Yifan Jiang, Shiyu Chang, and Zhangyang Wang. Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[21] Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. Global-and-local relative position embedding for unsupervised video summarization. In European Conference on Computer Vision (ECCV), 2020. 3", + "[22] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 54(10s):1-41, 2022. 1, 2", + "[23] Shun Kiyono, Sosuke Kobayashi, Jun Suzuki, and Kentaro Inui. Shape: Shifted absolute position embedding for transformers. arXiv preprint arXiv:2109.05644, 2021. 3", + "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 3", + "[25] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 1, 3, 4", + "[26] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4", + "[27] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. 3" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20390", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 5", + "[29] Jiahao Lu, Xi Sheryl Zhang, Tianli Zhao, Xiangyu He, and Jian Cheng. April: Finding the achilles' heel on privacy for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10051-10060, 2022. 1, 2, 3, 7, 8", + "[30] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 2, 3", + "[31] Samrudhdhi B Rangrej, Chetan L Srinidhi, and James J Clark. Consistency driven sequential transformers attention model for partially observable scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2518-2527, 2022. 2", + "[32] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5", + "[33] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 3", + "[34] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2", + "[35] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021. 3", + "[36] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems (NeurIPS), 2021. 6, 7", + "[37] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. 1, 3, 5, 7, 8", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 3", + "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 1" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 1", + "[41] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[43] Yu-An Wang and Yun-Nung Chen. What do position embeddings learn? an empirical study of pre-trained language model positional encoding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 1, 3, 7", + "[44] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 2", + "[45] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. Advances in Neural Information Processing Systems, 34:28522-28535, 2021. 2", + "[46] Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[47] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1", + "[48] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7", + "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7", + "[50] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021. 1, 3" + ], + "bbox": [ + 501, + 92, + 893, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20391", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_model.json b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c76d9d0a821dedeba0dc5afdc400b91154919b72 --- /dev/null +++ b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.131, + 0.89, + 0.152 + ], + "angle": 0, + "content": "Masked Jigsaw Puzzle: A Versatile Position Embedding for Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.179, + 0.858, + 0.253 + ], + "angle": 0, + "content": "Bin Ren\\(^{1,2*}\\) Yahui Liu\\(^{2*}\\) Yue Song\\(^{2}\\) Wei Bi\\(^{3}\\) Rita Cucchiara\\(^{4}\\) Nicu Sebe\\(^{2}\\) Wei Wang\\(^{5\\dagger}\\) \n\\(^{1}\\)University of Pisa, Italy \\(^{2}\\)University of Trento, Italy \n\\(^{3}\\)Tencent AI Lab, China \\(^{5}\\)Beijing Jiaotong University, China \n\\(^{4}\\)University of Modena and Reggio Emilia, Italy" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.286, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.474, + 0.698 + ], + "angle": 0, + "content": "Position Embeddings (PEs), an arguably indispensable component in Vision Transformers (ViTs), have been shown to improve the performance of ViTs on many vision tasks. However, PEs have a potentially high risk of privacy leakage since the spatial information of the input patches is exposed. This caveat naturally raises a series of interesting questions about the impact of PEs on accuracy, privacy, prediction consistency, etc. To tackle these issues, we propose a Masked Jigsaw Puzzle (MJP) position embedding method. In particular, MJP first shuffles the selected patches via our block-wise random jigsaw puzzle shuffle algorithm, and their corresponding PEs are occluded. Meanwhile, for the nonoccluded patches, the PEs remain the original ones but their spatial relation is strengthened via our dense absolute localization regressor. The experimental results reveal that 1) PEs explicitly encode the 2D spatial relationship and lead to severe privacy leakage problems under gradient inversion attack; 2) Training ViTs with the naively shuffled patches can alleviate the problem, but it harms the accuracy; 3) Under a certain shuffle ratio, the proposed MJP not only boosts the performance and robustness on large-scale datasets (i.e., ImageNet-1K and ImageNet-C, -A/O) but also improves the privacy preservation ability under typical gradient attacks by a large margin. The source code and trained models are available at https://github.com/yh1leo/MJP." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.703, + 0.21, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.471, + 0.864 + ], + "angle": 0, + "content": "Transformers [38] demonstrated their overwhelming power on a broad range of language tasks (e.g., text classification, machine translation, or question answering [22, 38]), and the vision community follows it closely and extends it for vision tasks, such as image classification [7,37], object detection [2, 51], segmentation [47], and image generation [3, 24]. Most of the previous ViT-based methods focus on designing different pre-training objectives [9, 11, 12] or variants of self-attention mechanisms [28, 39, 40]. By contrast, PEs" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.285, + 0.695, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.284, + 0.895, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.429, + 0.896, + 0.527 + ], + "angle": 0, + "content": "Figure 1. Low-dimensional projection of position embeddings from DeiT-S [37]. (a) The 2D UMAP projection, it shows that reverse diagonal indices have the same order as the input patch positions. (b) The 3D PCA projection, it also shows that the position information is well captured with PEs. Note that the embedding of index 1 (highlighted in red) corresponds to the [CLS] embedding that does not embed any positional information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.839 + ], + "angle": 0, + "content": "receive less attention from the research community and have not been well studied yet. In fact, apart from the attention mechanism, how to embed the position information into the self-attention mechanism is also one indispensable research topic in Transformer. It has been demonstrated that without the PEs, the pure language Transformer encoders (e.g., BERT [6] and RoBERTa [25]) may not well capture the meaning of positions [43]. As a consequence, the meaning of a sentence can not be well represented [8]. Similar phenomenon of PEs could also be observed in vision community. Dosovitskiy et al. [7] reveals that removing PEs causes performance degradation. Moreover, Lu et al. [29] analyzed this issue from the perspective of user privacy and demonstrated that the PEs place the model at severe privacy risk since it leaks the clues of reconstructing sequential patches back to images. Hence, it is very interesting and necessary to understand how the PEs affect the accuracy, privacy, and consistency in vision tasks. Here the consistency means whether the predictions of the transformed/shuffled image are consistent with the ones of the original image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To study the aforementioned effects of PEs, the key is to figure out what explicitly PEs learn about positions from input patches. To answer this question, we project the high-dimensional PEs into the 2D and 3D spaces using Uni-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.402, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution. Email: {bin.ren, yahui.liu} @unitn.it" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.385, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author. Email: wei.wang@bjtu.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.402, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20382" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.088, + 0.278, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.239, + 0.201, + 0.251 + ], + "angle": 0, + "content": "(a) \\(\\mathbf{x}\\)" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.089, + 0.472, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.238, + 0.42, + 0.251 + ], + "angle": 0, + "content": "(b) \\(\\widetilde{\\mathbf{x}} (\\gamma = 1)\\)" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.089, + 0.667, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.238, + 0.619, + 0.251 + ], + "angle": 0, + "content": "(c) \\(\\widetilde{\\mathbf{x}} (\\gamma = 0.4)\\)" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.089, + 0.877, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.238, + 0.827, + 0.25 + ], + "angle": 0, + "content": "(d) Overview of our MJP" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.894, + 0.297 + ], + "angle": 0, + "content": "Figure 2. (a) The original input patches; (b) Totally random shuffled input patches; (c) Partially random shuffled input patches; (d) An overview of the proposed MJP. Note that we show the random shuffled patches and its corresponding unknown position embedding in green and the rest part in blue. DAL means the self-supervised dense absolute localization regression constraint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.473, + 0.52 + ], + "angle": 0, + "content": "form Manifold Approximation & Projection (UMAP) [30] and PCA, respectively. Then for the first time, we visually demonstrate that the PEs can learn the 2D spatial relationship very well from the input image patches (the relation is visualized in Fig. 1). We can see that the PEs are distributed in the same order as the input patch positions. Therefore, we can easily obtain the actual spatial position of the input patches by analyzing the PEs. Now it explains why PEs can bring the performance gain for ViTs [7]. This is because the spatial relation in ViTs works similar as the inherent intrinsic inductive bias in CNNs (i.e., it models the local visual structure) [45]. However, these correctly learned spatial relations are unfortunately the exact key factor resulting in the privacy leakage [29]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.472, + 0.718 + ], + "angle": 0, + "content": "Based on these observations, one straightforward idea to protect the user privacy is to provide ViTs with the randomly transformed (i.e., shuffled) input data. The underlying intuition is that the original correct spatial relation within input patches will be violated via such a transformation. Therefore, we transform the previous visually recognizable input image patches \\(\\mathbf{x}\\) shown in Fig. 2(a) to its unrecognizable counterpart \\(\\widetilde{\\mathbf{x}}\\) depicted in Fig. 2(b) during training. The experimental results show that such a strategy can effectively alleviate the privacy leakage problem. This is reasonable since the reconstruction of the original input data during the attack is misled by the incorrect spatial relation. However, the side-effect is that this leads to a severe accuracy drop." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.472, + 0.856 + ], + "angle": 0, + "content": "Meanwhile, we noticed that such a naive transformation strategy actually boosts the consistency [31, 34, 44] albeit the accuracy drops. Note that here the consistency represents if the predictions of the original and transformed (i.e., shuffled) images are consistent. Given the original input patches \\(\\mathbf{x}\\) and its corresponding transformed (i.e., shuffled) counterpart, we say that the predictions are consistent if \\(\\arg \\max P(\\mathcal{F}(\\mathbf{x})) = \\arg \\max P(\\mathcal{F}(\\widetilde{\\mathbf{x}}))\\), where \\(\\mathcal{F}\\) refers to the ViT models, and \\(P\\) denotes the predicted logits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.902 + ], + "angle": 0, + "content": "These observations hint that there might be a trade-off solution that makes ViTs take the best from both worlds (i.e., both the accuracy and the consistency). Hence, we pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.895, + 0.521 + ], + "angle": 0, + "content": "pose the Masked Jigsaw Puzzle (MJP) position embedding method. Specifically, there are four core procedures in the MVP: (1) We first utilize a block-wise masking method [1] to randomly select a partial of the input sequential patches; (2) Next, we apply jigsaw puzzle to the selected patches (i.e., shuffle the orders); (3) After that, we use a shared unknown position embedding for the shuffled patches instead of using their original PEs; (4) To well maintain the position prior of the unshuffled patches, we introduce a dense absolute localization (DAL) regressor to strengthen their spatial relationship in a self-supervised manner. We simply demonstrate the idea of the first two procedures in Fig. 2(c), and an overview of the proposed MVP method is available in Fig. 2(d). In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.525, + 0.894, + 0.586 + ], + "angle": 0, + "content": "- We demonstrate that although PEs can boost the accuracy, the consistency against image patch shuffling is harmed. Therefore, we argue that studying PEs is a valuable research topic for the community." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.591, + 0.894, + 0.637 + ], + "angle": 0, + "content": "- We propose a simple yet efficient Masked Jigsaw Puzzle (MJP) position embedding method which is able to find a balance among accuracy, privacy, and consistency." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.64, + 0.895, + 0.732 + ], + "angle": 0, + "content": "- Extensive experimental results show that MJP boosts the accuracy on regular large-scale datasets (e.g., ImageNet-1K [32]) and the robustness largely on ImageNet-C [18], -A/O [19]. One additional bonus of MJP is that it can improve the privacy preservation ability under typical gradient attacks by a large margin." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.525, + 0.895, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.746, + 0.642, + 0.763 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.695, + 0.787 + ], + "angle": 0, + "content": "2.1. Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Transformers [38], originally designed for Nature Language Processing (NLP) tasks, have recently shown promising performance on computer vision tasks [13, 22]. Benefiting from the strong representation power of modelling global relations between image patches, Vision Transformers (ViTs) [7] have achieved superior performance than their counterpart CNNs on image classification and various other" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20383" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "downstream tasks (e.g., object detection [2, 51], object re-identification [17], dense prediction [41, 42, 46, 50] and image generation [3, 4, 20, 24])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.474, + 0.429 + ], + "angle": 0, + "content": "As a core module in ViTs, multi-head self-attention (MSAs) [7, 38] aggregate sequential tokens with normalized attentions as: \\( z_{j} = \\sum_{i}\\mathrm{Softmax}(\\frac{QK}{\\sqrt{d}})_iV_{i,j} \\) where \\( Q \\), \\( K \\) and \\( V \\) are query, key and value matrices, respectively. \\( d \\) is the dimension of query and key, and \\( z_{j} \\) is the \\( j \\)-th output token. In theory, when the position information is not considered, the outputs of MSAs should be strictly invariant to the input sequence order (i.e., position-insensitivity). This indicates that a visually recognizable image can be transformed into its unrecognizable counterpart by permuting the order of image patches while maintaining the performance delivered by ViTs compared with the ViTs trained on the original non-permuted image. However, the usage of PEs hinders such implementations, where the outputs of the ViTs vary dramatically with the mentioned naive transformations shown in Fig. 2(b). In this work, our main focus is to explicitly figure out what PEs actually learn from input patches about positions, and how the PEs affect the accuracy, privacy, and the consistency properties of ViTs." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.439, + 0.276, + 0.455 + ], + "angle": 0, + "content": "2.2. Position Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.471, + 0.78 + ], + "angle": 0, + "content": "In Transformer networks, both the attention and the (individual token based) feed-forward layers are permutation invariant when the position information is not considered. In this way, the spatial relationships between image patches could not be modeled as the position information is completely discarded. As compensation, PEs are naturally introduced into ViTs to provide information about the token order during the learning process, since it offers possibilities for dependency modeling between elements at different positions. For example, previous works [10,33,38] indicated that PEs are useful to give the model a sense of which portion of the sequence in the input/output it is currently dealing with. Inspired by this, some works [15,21,23,27,35] showed diverse application scenarios that benefit from the usage of suitable PEs. In addition, Chu et al. [5] proposed correlating the PEs with their local neighborhood of the input sequence. Liu et al. [26] proposed to enhance the spatial prior (i.e., relative localization) in the final content embedding to indirectly enrich the inductive bias. Obviously, although these methods enhance the position information learnt by PEs, they indeed degenerate the position-insensitive property of MSAs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Especially, Wang et al. [43] revealed that Transformer encoders (e.g., BERT [6] and RoBERTa [25]) may not well capture the meaning of positions (absolute and relative positions). They showed that Transformer encoders learn the local position information that can only be effective in masked language modeling. In contrast, there does not exist such a similar \"masked language modeling\" procedure in ViTs (and VTs) in the typically supervised pre-training. Moreover, Lu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "et al. [29] revealed that the learnable PEs place the model at severe privacy risk, which leaks the clues of reconstructing sequential patches to images. In this paper, we dive into the usage of PEs and propose an efficient position embedding method, MJP, to improve the position-insensitive property of ViTs without hurting the positive effects of PEs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.197, + 0.884, + 0.214 + ], + "angle": 0, + "content": "3. Preliminaries: 2D/3D Spatial Priors in PEs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.895, + 0.585 + ], + "angle": 0, + "content": "Although numerous works claim that the PEs can learn the 2-D spatial relationship of image patches, this claim has not been demonstrated visually or mathematically. To visualize the concrete relation of image patches captured in the high-dimensional position embedding, we project them into the 2-D and 3-D spaces using Uniform Manifold Approximation and Projection (UMAP) [30] and PCA, respectively. Fig. 1 displays the projections of PEs from DeiT-S [37]. Note that [7] only shows the cosine similarity between the PEs, which is not exactly the 2D spatial information mentioned in our paper. Here, we explicitly specify the spatial relationship as the relative distance in the 2D/3D coordinate space. As shown in Fig. 1, the spatial relationship can be projected into a 2D/3D coordinate system and indicate the localizations of these embeddings. For the UMAP, the projected positional embedding emerges as grid-like structures with the distances between each point roughly the same, which is coherent with the relation of embedded image patches. For the PCA, the position embedding that corresponds to neighboring image patches groups together and aggregates to the stick-like form. This phenomenon demonstrates that the spatial relationship is indeed learned by the PEs. Moreover, this also implies that the spatial relationship learned in the high-dimensional space still manifests in the low-dimensional space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.895, + 0.706 + ], + "angle": 0, + "content": "The learnable PEs, which work as a lookup table of a dictionary, maps the 1-dimensional data into the sparse high-dimensional space. By nature, it is a sparse large matrix. Dimensionality reduction techniques can capture the structural information of sparse matrices using low dimensionalities. For the PCA, the amount of information retained in the projection can be measured by the ratio of explained variance. Formally, we have the definition as follows:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.713, + 0.894, + 0.775 + ], + "angle": 0, + "content": "Definition 1 (Explained Variance) Let \\(\\mathbf{P}\\) and \\(\\mathbf{V}\\) denote the data matrix and the PCA projection matrix. The ratio of explained variance that \\(\\mathbf{PV}\\) accounts for is defined as \\(\\sum \\sigma (\\mathbf{PV})^2 /\\sum \\sigma (\\mathbf{P})^2\\) where \\(\\sigma (\\cdot)\\) denotes the singular value." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In practice, we observe that the 3-dimensional PCA projection explains \\(54.6\\%\\) of the total variance of the DeiT-S embedding matrix. Given that a large amount of information can be captured by the low-dimensional projection, we propose that explicitly enforcing low-dimensional positional prior can help the positional learning in the high-dimensional space, which might accelerate the convergence rate of training and improve the performance (See Sec. 4.2)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20384" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.092, + 0.45, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Block-wise Random Jigsaw Puzzle Shuffle" + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.111, + 0.475, + 0.258 + ], + "angle": 0, + "content": "Input: Input image: \\(\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times C}\\) Shuffle Ratio: \\(\\gamma\\) Patch Size: \\(P\\) \nOutput: Shuffled image patches: \\(\\widetilde{\\pmb{x}}_p\\) \n1: \\(\\mathbf{x}_p\\in \\mathbb{R}^{N\\times (P^2\\cdot C)}\\gets\\) Patchlize(x,P) \n2: \\(\\mathbf{m}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}\\gets\\) BinaryInitialize(xp,0) \n3: \\(\\widetilde{\\mathbf{m}}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}\\gets\\) BlockwiseMask(m,γ) [25] \n4: \\(\\widetilde{\\pmb{x}}_p\\in \\mathbb{R}^{N\\times (P^2\\cdot C)}\\gets\\) Jigsaw Puzzle(xp, \\(\\widetilde{\\mathbf{m}}\\) 5: return \\(\\widetilde{\\pmb{x}}_p\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.283, + 0.168, + 0.299 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.429 + ], + "angle": 0, + "content": "Based on the design principle of MSAs, the outputs of MSAs should be entirely position-agnostic. However, PEs hinder such a property because it can learn the strong 2-D spatial relation of the input patches (as we demonstrated in Sec. 3). To this end, we propose the block-wise random jigsaw puzzle shuffle algorithm (See Alg. 1) to transform the input patches with different shuffle ratios \\(\\gamma\\) for intermingling the original correct spatial relation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.595 + ], + "angle": 0, + "content": "Since we experimentally demonstrate that the totally shuffled strategy (i.e., \\(\\gamma = 1.0\\)) will degenerate the accuracy a lot albeit the consistency increases. As a remedy, we only shuffle portion of the sequence patches, and strengthen the spatial relation of the rest part with a dense absolute localization regression. Finally, a versatile position embedding method MJP is proposed. The detailed analysis of each module is available in the ablation study in Sec. 5.3. In the following sub-sections, we will introduce the Jigsaw Puzzle Shuffle method (Sec. 4.1), the spatial relation strengthen method (Sec. 4.2), and our final MJP method (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.436, + 0.62 + ], + "angle": 0, + "content": "4.1. Block-wise Random Jigsaw Puzzle Shuffle" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.841 + ], + "angle": 0, + "content": "Specifically, given an input image \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}\\), we first reshape it into a sequence of flattened 2D patches \\(\\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot C)}\\), where \\((H, W)\\) is the resolution of the original image, \\(C\\) is the number of channels, \\((P, P)\\) is the resolution of each image patch, and then we have \\(N = HW / P^2\\), which denotes the number of patches. Then instead of directly applying block-wise masking method [1] to the image, we first initialize a binary mask matrix \\(\\mathbf{m} \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P}}\\) with the same size as the image patches in \\(\\mathbf{x}_p\\). Next, we use [1] to update the binary mask \\(\\mathbf{m}\\) in which the masked positions will be set to 1 and the rest untouched positions remains 0. The hyper-parameter \\(\\gamma\\) is used to control the ratio of selected positions. After that, a jigsaw puzzle shuffle operation is applied to \\(\\mathbf{x}_p\\) conditioned on the updated binary mask \\(\\widetilde{\\mathbf{m}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Finally, we get the shuffled patch sequence \\(\\widetilde{\\mathbf{x}}_p\\), where the shuffled patches are actually \\(\\{\\mathbf{x}_p^i |\\widetilde{\\mathbf{m}}_i = 1\\}\\). Notably, in [1], the patches covered by the sampled mask are not visible to the encoder module, while in our method, the masking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "strategy [1] is only used to mask out the positions/indices of the selected patches. These patches are still visible to the encoder module and they are randomly shuffled in a jigsaw puzzle manner. One intuitive shuffled toy example is shown in Fig. 2(c), where the green part means the selected shuffled region while the blue part remains unchanged. Actually, this algorithm shares a similar idea with the data augmentation method as the training images are always changing as their patches are randomly shuffled during each iteration." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.235, + 0.803, + 0.252 + ], + "angle": 0, + "content": "4.2. Strengthening Spatial Prior in PEs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Liu et al. [26] noticed that by enhancing the 2-D spatial information of the output embeddings of the last layer of ViTs, the training convergence speed can be accelerated. Inspired by their work, we observe that similar gains can be obtained by applying a low-dimensional spatial prior in PEs. Different from the dense relative localization constraint in [26] which samples relative pairs from the whole output sequential embeddings, we propose a much simpler dense absolute localization (DAL) regression method to directly use self-supervised absolute location to enhance the spatial information in PEs, which avoids sampling relative pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.893, + 0.56 + ], + "angle": 0, + "content": "Since the PEs capture the absolute position of the input patches, to some extent, the position information could be reconstructed via a reversed mapping function \\( g(\\cdot):\\mathcal{E}\\to \\mathcal{P} \\) where \\( \\mathcal{E} \\) and \\( \\mathcal{P} \\) are embedding space and position space, respectively. Given that PEs have one-to-one correspondence with the sequential image patches, we can reshape them into \\( \\mathbf{E}_{\\mathrm{pos}}\\in \\mathbb{R}^{K\\times K\\times D} \\), where \\( K \\) refers to height/weight of the grid and \\( D \\) refers to latent vector size. Then we can compute the reverse mapping from \\( \\mathcal{E}\\rightarrow \\mathcal{P} \\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.566, + 0.892, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\left(\\widetilde {i}, \\widetilde {j}\\right) ^ {T} = g \\left(\\mathbf {E} _ {\\text {p o s}} ^ {i, j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.894, + 0.64 + ], + "angle": 0, + "content": "in which \\((\\widetilde{i},\\widetilde{j})^T\\) is the predicted patch position, and \\(\\mathbf{E}_{\\mathrm{pos}}^{i,j}\\) is the position embedding of the patch \\((i,j)\\) in the \\(K\\times K\\) grid. The dense absolute localization (DAL) loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.646, + 0.892, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D A L}} = \\mathbb {E} _ {\\mathbf {E} _ {\\text {p o s}}, 1 \\leq i, j \\leq K} ^ {i, j} [ \\| (i, j) ^ {T} - (\\widetilde {i}, \\widetilde {j}) ^ {T} \\| _ {1} ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.824 + ], + "angle": 0, + "content": "where the expectation is computed by averaging the \\(\\ell_1\\) loss between the correspond \\((i,j)^T\\) and \\((\\widetilde{i},\\widetilde{j})^T\\). Then, \\(\\mathcal{L}_{\\mathrm{DAL}}\\) is added to the standard cross-entropy loss \\((\\mathcal{L}_{\\mathrm{CE}})\\) of the native ViTs. The final loss is: \\(\\mathcal{L}_{\\mathrm{all}} = \\mathcal{L}_{\\mathrm{CE}} + \\lambda \\mathcal{L}_{\\mathrm{DAL}}\\), where we set \\(\\lambda = 0.01\\) for all experiments. Note that the mapping function can be either linear or nonlinear. Throughout this work, we mainly discuss three implementations, including non-parametric PCA, learnable linear (LN), and nonlinear (NLN) projection layers. They will be discussed in details in Sec. 5.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.734, + 0.849 + ], + "angle": 0, + "content": "4.3. MJP Position Embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The main goal of MJP position embedding in our work is to enhance the consistency (i.e., position-insensitive property) of ViTs and preserve the accuracy (keeping the spatial" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20385" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.091, + 0.394, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 The pipeline of the proposed MJP." + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.111, + 0.415, + 0.185 + ], + "angle": 0, + "content": "Input: Input image: \\(\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times C}\\) Shuffle Ratio: \\(\\gamma\\) ; Patch Size: \\(P\\) 1: \\(\\widetilde{\\pmb{x}}_p\\gets Alg.1(\\mathbf{x},P,\\gamma)\\) // 1st & 2nd procedures 2: \\(\\mathbf{E}_{\\mathrm{unk}}(\\widetilde{\\pmb{x}}_p)\\) // 3rd procedure 3: \\(\\mathbf{DAL}(\\mathbf{x} - \\mathbf{x}\\cap \\widetilde{\\pmb{x}}_p)\\) // 4th procedure, only for train" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.211, + 0.471, + 0.271 + ], + "angle": 0, + "content": "relation modeling functionality of PEs) for the standard classification tasks. Usually, with original input image patches \\(\\mathbf{x}\\) we can formulate the function of the input layer which is located before the Transformer block as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.282, + 0.47, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {C L S}}; \\mathbf {x} _ {p} ^ {1} \\mathbf {E}; \\mathbf {x} _ {p} ^ {2} \\mathbf {E}, \\dots , \\mathbf {x} _ {p} ^ {N} \\mathbf {E} \\right] + \\mathbf {E} _ {\\text {p o s}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.469, + 0.377 + ], + "angle": 0, + "content": "where \\(\\mathbf{E} \\in \\mathbb{R}^{(P^2 \\cdot C) \\times D}\\) is a trainable linear projection layer and \\(\\mathbf{x}_p^k\\mathbf{E}\\) refers to the output of linear projection (i.e., the patch embeddings), and \\(\\mathbf{E}_{\\mathrm{pos}} \\in \\mathbb{R}^{(N + 1) \\times D}\\) refers to PEs (the additional one is applied to the [CLS] embedding)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.47, + 0.527 + ], + "angle": 0, + "content": "Next, we apply the proposed block-wise random jigsaw puzzle shuffle algorithm to \\(\\mathbf{x}_p\\) and produce the transformed patch sequences \\(\\widetilde{\\mathbf{x}}_p\\) with the shuffle ratio \\(\\gamma\\). In this scenario, if we maintain the original position embedding sequence, it leads to a mismatch issue between the shuffled patches and the position embedding sequence. Therefore, we introduce a shared unknown position embedding to the shuffled positions to alleviate the mismatching issue. With the corresponding updated mask \\(\\widetilde{\\mathbf{m}}\\) from Alg. 1 (1st & 2nd procedures), we propose our MJP PEs:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.538, + 0.47, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathbf {E}} _ {\\text {p o s}} ^ {i} = \\left\\{ \\begin{array}{l l} \\mathbf {E} _ {\\text {p o s}} ^ {i}, & \\text {i f} \\widetilde {\\mathbf {m}} _ {i} = 0 \\\\ \\mathbf {E} _ {\\text {u n k}}, & \\text {i f} \\widetilde {\\mathbf {m}} _ {i} = 1 \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.469, + 0.666 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}_{\\mathrm{unk}} \\in \\mathbb{R}^{1 \\times D}\\) refers to a share learnable embedding (i.e., unknown position embedding). It represents that the image patch in this position has random permutation and its position should be occluded (3rd procedure). \\(\\mathbf{E}_{\\mathrm{pos}}\\) is the original position embedding for the rest image patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.816 + ], + "angle": 0, + "content": "Besides, we revisit the remaining PEs (i.e., corresponding to the non-selected patches) and apply low-dimensional prior on them. The low-dimensional prior is imposed by the proposed DAL (Sec. 4.2) regression method for strengthening the spatial relation (4th procedure). A toy illustration of the proposed MJP is shown in Fig. 2(d), where the green color represents the randomly permuted patches and its corresponding unknown PEs, while the blue color indicates the rest regular patches and its related original PEs. We also formalize these procedures as an algorithm in Alg. 2." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.817, + 0.465, + 0.831 + ], + "angle": 0, + "content": "Thus, we replace the input layer with a new formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.841, + 0.47, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathbf {z}} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {C L S}}; \\widetilde {\\mathbf {x}} _ {p} ^ {1} \\mathbf {E}; \\widetilde {\\mathbf {x}} _ {p} ^ {2} \\mathbf {E}, \\dots , \\widetilde {\\mathbf {x}} _ {p} ^ {N} \\mathbf {E} \\right] + \\widetilde {\\mathbf {E}} _ {\\text {p o s}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "the following procedures and modules are exactly the same as the ones in the original ViTs." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 1. Comparisons of different backbones on ImageNet-1K classification. Note that the image size here are all set to \\(224 \\times 224\\)." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.119, + 0.886, + 0.216 + ], + "angle": 0, + "content": "
MethodParam. Top-1 Acc.↑ Diff.Norm.↓ Consistency
ResNet-50 [16]2579.311.7751.5
ResNet-50 + MJP2579.47.1169.3
DeiT-S [37]2279.816.2164.3
DeiT-S + MJP2280.58.9682.9
Swin-T [28]2981.315.4941.5
Swin-T + MJP2981.312.3666.9
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.229, + 0.632, + 0.246 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.894, + 0.36 + ], + "angle": 0, + "content": "We follow the typical supervised pre-training procedure, where all the compared models are trained on ImageNet1K [32] to show the capacity of our proposed MJP method. We also benchmark the proposed MJP method on ImageNet1K, which contains 1.28M training images and 50K validation images of 1,000 classes. The training details mostly follow the training protocols from Touvron et al. [37]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.371, + 0.774, + 0.387 + ], + "angle": 0, + "content": "5.1. Regular ImageNet-1K training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.895, + 0.59 + ], + "angle": 0, + "content": "We mainly compare with three typical existing methods, including two state-of-the-art Visual Transformers (i.e., DeiT [37] and Swin [28]) and one widely-used CNN-based ResNet-50 [16]. All these methods are of comparable sizes (i.e., less than 30M parameters). Besides the common Top-1 accuracy (Top-1 Acc.), we also report another two evaluation metrics (i.e. Diff. Norm. and Consistency) to show the position invariance of a model to the jigsaw puzzle transformation. For a given input image \\(\\mathbf{x}\\), we collect its counterpart \\(\\widetilde{\\mathbf{x}}\\) by applying masked jigsaw puzzle to shuffle a portion of selected patches. Next, we calculate the difference \\(\\ell_2\\)-norm (i.e., Diff. Norm.) between the [CLS] embedding inferred from \\(\\mathbf{x}\\) and \\(\\widetilde{\\mathbf{x}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.595, + 0.892, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\mathcal {F} _ {\\mathrm {C L S}} (\\mathbf {x}) - \\mathcal {F} _ {\\mathrm {C L S}} (\\widetilde {\\mathbf {x}}) \\right\\| _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.656 + ], + "angle": 0, + "content": "For Consistency, we measure how many classification results stay the same (\\%) given the perturbation:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.668, + 0.892, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathbf {x}} \\left[ \\mathbb {I} \\left\\{\\arg \\max P (\\mathcal {F} (\\mathbf {x})) = \\arg \\max P (\\mathcal {F} (\\widetilde {\\mathbf {x}})) \\right\\} \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.894, + 0.877 + ], + "angle": 0, + "content": "\\(P(\\mathcal{F}(\\mathbf{x}))\\) denotes the predicted classification probability of image \\(\\mathbf{x}\\), and \\(\\arg \\max P(\\mathcal{F}(\\mathbf{x}))\\) represents the predicted class index. For a fair comparison, we use \\(\\gamma = 0.15\\) to create \\(\\widetilde{\\mathbf{x}}\\) in Table 1. According to these results, the proposed MJP does not have a negative effect on the top-1 accuracy (MJP even brings a marginal improvement for DeiT-S). More importantly, it improves Diff. Norm. and Consistency by a large margin. Besides, MJP works well in the variants of ViTs (e.g., Swin [28]), which shows a good potential generalization ability. Note that we add MJP to ResNet50 by only shuffling the image patches with Alg. 1 (No PEs). For Swin-T, we add MJP to the absolute PEs of Swin-T." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.841, + 0.9 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.077, + 0.318, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.224, + 0.212, + 0.235 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.077, + 0.596, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.224, + 0.493, + 0.235 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.077, + 0.877, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.224, + 0.774, + 0.235 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.895, + 0.266 + ], + "angle": 0, + "content": "Figure 3. Ablation on the mask ratio \\(\\gamma\\) during inference: (a) comparisons among ViT-S, DeiT-S and our method (trained with \\(\\gamma = 0.03\\)); (b) comparisons between Swin-T and our method (trained with \\(\\gamma = 0.03\\)); (c) comparisons of our method on DeiT-S trained with different \\(\\gamma\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.271, + 0.471, + 0.299 + ], + "angle": 0, + "content": "Table 2. Comparisons on robustness to common corruptions and adversarial examples." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.298, + 0.465, + 0.368 + ], + "angle": 0, + "content": "
MethodImageNet-CImageNet-AImageNet-O
mCE ↓Acc ↑AURRA ↑AUPR ↑
DeiT-S54.619.225.120.9
DeiT-S + MJP40.7821.629.822.6
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.381, + 0.357, + 0.398 + ], + "angle": 0, + "content": "5.2. Robustness on Challenging Sets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.471, + 0.496 + ], + "angle": 0, + "content": "Besides the strength on standard classification, we also observe an auxiliary benefit of the proposed MJP on robustness. ImageNet-C [18] benchmarks a classifier's robustness to common corruptions2. The mean Corruption Error (mCE) is used to measure the generalization of a model at corrupted distributions (the lower the better)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.471, + 0.633 + ], + "angle": 0, + "content": "ImageNet-A/O [19] focus on adversarial examples, and it enables us to either test image classification performance when the input data distribution shifts (i.e., Acc and AU-RRA), or test out-of-distribution detection performance when the label distribution shifts (i.e., AUPR)\\(^3\\). As shown in Table 2, the proposed MJP not only boosts the accuracy on the standard evaluation on ImageNet-1K validation set but also improves the robustness largely on the adversarial samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.71 + ], + "angle": 0, + "content": "The underlying reason might be that MJP enforces the ViTs aware of both local and global context features, and it helps ViTs to get rid of some unnecessary sample-specific local features during the training. This has been verified by the visualization maps (see Fig. 3 of our Supp. Mat.)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.721, + 0.251, + 0.737 + ], + "angle": 0, + "content": "5.3. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Results with different MJP ratios. As shown in Table 3, we test different masking ratios used in the block-wise masking strategy during the training. Obviously, Diff. Norm. has the inverse tendencies compared to Consistency, where a smaller Diff. Norm. usually indicates a larger (better) consistency score. Comparing the accuracy trained with \\(\\gamma >0\\), it shows our model is not sensitive to different \\(\\gamma\\). In particular, a" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.271, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Table 3. Ablation study on the proposed MJP method trained with different masking ratio \\( \\gamma \\) ." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.301, + 0.885, + 0.395 + ], + "angle": 0, + "content": "
MetricMasking Ratio
00.030.090.150.210.27
Top-1 Acc.80.080.580.380.480.280.3
Diff. Norm.16.568.966.365.234.393.97
Consistency64.082.988.190.592.393.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.548 + ], + "angle": 0, + "content": "small ratio (e.g., \\(\\gamma = 0.03\\)) is sufficient for boosting the accuracy. In addition, a large ratio can reduce the Diff. Norm. and improves the consistency by a large margin. Moreover, the model trained with a larger \\(\\gamma\\) is inclined to be more consistent as shown in Fig. 3 (c). In addition, we observe that the consistency keeps increasing when the mask ratio increases from \\(3\\%\\) to \\(27\\%\\). However, further increasing the mask ratio will not bring consistency improvement and what is worse is that the accuracy marginally decreases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.64 + ], + "angle": 0, + "content": "For a model trained with a fixed \\(\\gamma\\), we also test its accuracy with different masking ratios during the inference. As shown in Fig. 3 (a) and (b), the performances of original models drop significantly when we shuffle more image patches. In contrast, our proposed method shows more consistent performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.896, + 0.839 + ], + "angle": 0, + "content": "Comparison of variants of MJP. We test several variants of the proposed MJP, including (1) Removing the PEs from the original DeiT-S; (2) SPP: shuffling both the \\(16 \\times 16\\) patches and pixels within the patches, which is tested in MLP-Mixer [36] to validate the invariance to permutations; (3) JP: applying masked jigsaw puzzle to the sequence patches; (4) IDX: using an additional collection of embeddings to indicate the global indexes for the input patches; (5) UNK: replacing the PEs in the masked positions with a shared unknown position embedding; (6) DAL: jointly learning dense absolute localization regression in an self-supervised manner during the pretraining, where we provide PCA, linear (LN) and nonlinear (NLN) projections, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Table 4 shows the detailed ablation studies on the variants of our proposed MJP method. First, as expected, when we remove the PEs from the original DeiT-S model, the accuracy decreases by \\(2.3\\%\\), but the consistency achieves \\(100\\%\\). It" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.862, + 0.414, + 0.876 + ], + "angle": 0, + "content": "2https://github.com/hendrycks/robustness" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.876, + 0.473, + 0.901 + ], + "angle": 0, + "content": "3https://github.com/hendrycks/natural-adv-examples" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.862, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20387" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.09, + 0.454, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation study on the variants of the proposed MJP." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.109, + 0.462, + 0.27 + ], + "angle": 0, + "content": "
MethodTop-1 Acc. ↑Consistency ↑
A: DeiT-S [37]79.864.3
B: A - PEs77.5 (-2.3)100.0
C: A + SPP [36]74.9 (-4.9)74.8
D: A + DAL (NLN)80.0 (+0.2)64.0
E: A + JP79.2 (-0.6)73.8
F: A + JP + IDX79.9 (+0.1)79.6
G: A + JP + UNK80.1 (+0.3)83.8
H: A + JP + UNK + DAL (PCA)79.9 (+0.1)83.4
I: A + JP + UNK + DAL (LN)80.0 (+0.2)83.8
J: A + JP + UNK + DAL (NLN)80.5 (+0.7)82.9
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.283, + 0.441, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.422, + 0.447, + 0.435 + ], + "angle": 0, + "content": "Figure 4. Distributions of accumulated eigenvalues of PEs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.478 + ], + "angle": 0, + "content": "verifies that the ViTs are naturally position-insensitive once without using PEs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.469, + 0.572 + ], + "angle": 0, + "content": "We observe that the previous SPP strategy harms the accuracy of the model (i.e., \\(-4.9\\%\\)), which indicates it is insufficient to simply shuffle both patches and pixels in the input image. As we expect, the usage of UNK embedding alleviates the confusion between shuffled and unshuffled positions, which boosts both the accuracy and consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.695 + ], + "angle": 0, + "content": "Finally, we can find that using dense absolute localization regression on the unmasked PEs can marginally boost the accuracy. Although PCA does not involve more parameters to the model, it significantly increases the computation latency (e.g., \\(+46\\%\\) with only 5 iterations). In general, it is better to use a nonlinear projection (e.g., a 3-layers MLP with negligible computation latency) to allow the learned PEs to aggregate more additional information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Informativeness of the PEs. For a position space \\(\\mathcal{P}\\) (i.e., a 1-dim or 2-dim space), we may not need a high-dimension embedding space \\(\\mathcal{X}\\) to model the positions. To measure the informativeness of the learned PEs, we apply singular value decomposition (SVD) to PEs and analyze their eigenvalue distributions. Fig. 4 plots the curves of accumulated energy/sum of top-\\(n\\) eigenvalues versus the energy of all the eigenvalues. Supposing that we are using PCA to project the PEs of both matrices, to achieve the same explained variance ratio, our MJP needs more singular values (i.e., large dimensionality) than DeiT-S. This indicates that our positional embedding matrix is more informative. Similar observation is also revealed by Wang et al. [43]." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.159 + ], + "angle": 0, + "content": "Table 5. Comparisons on gradient leakage by analytic attack [29] with ImageNet-1K validation set, where we test (1) ViT-S, DeiT-S and our model in the setting (a); (2) ViT-S, DeiT-S and our model in the setting (b) (i.e., MJP with \\(\\gamma = 0.27\\)); (3) ablation on without (w/o) using \\(\\mathbf{E}_{\\mathrm{unk}}\\) in setting (a); and (4) Our model in setting (c)." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.16, + 0.892, + 0.305 + ], + "angle": 0, + "content": "
ModelSet.Acc.↑MSE↑\\( FFT_{2D} \\)↑PSNR↓SSIM↓LPIPS↑
(1)ViT-S [7]78.1.0278.003919.27.5203.3623
DeiT-S [37]a79.8.0350.005718.94.5182.3767
DeiT-S (w/o PEs)77.5.0379.008220.22.5912.2692
DeiT-S+MJP80.5.1055.016611.52.4053.6545
(2)ViT-S [7]18.7.0327.001618.44.6065.2836
DeiT-S [37]b36.0.0391.002417.60.5991.3355
DeiT-S (w/o PEs)77.5.0379.002520.25.6655.2370
DeiT-S+MJP62.9.1043.005911.66.4493.6519
(3)DeiT-S+MJP (w/o)a40.6.1043.005911.66.4493.6519
(4)DeiT-S+MJPc62.9.1706.03388.07.0875.8945
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.697, + 0.336 + ], + "angle": 0, + "content": "5.4. Privacy Preservation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.895, + 0.495 + ], + "angle": 0, + "content": "The fundamental principle of the gradient attack methods in federated learning is that each sample activates only a portion of content-related neurons in the deep neural networks, leading to one specific backward gradients for one related samples (i.e., 1-to-1 mapping). Based on such an observation, we argue that feeding ViTs with input patches with permuted sequences may intuitively mislead the attack. This is because now both the original and the transformed inputs may be matched to the same backward gradients (i.e., \\( n \\)-to-1 mapping)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.895, + 0.722 + ], + "angle": 0, + "content": "To validate such an assumption, we utilize the public protocols4 to recover image with gradient updates in the privacy attack. In this privacy attack, we apply the Analytic Attack proposed in APRIL [29], which is designed for attacking the ViTs. We randomly sample 1K images from the validation set of ImageNet-1K (i.e., one image per category). To evaluate the anti-attack performance of a model, we introduce image similarity metrics to account for pixel-wise mismatch, including Mean Square Error (MSE), Peak Signal-to-Noise Ratio (PSNR), cosine similarity in the Fourier space \\((\\mathrm{FFT}_{2\\mathrm{D}})\\), and Learned Perceptual Image Patch Similarity (LPIPS) [49]. Different from the evaluation in gradient attacks [14, 29, 48], we suppose a model is with better capacity of privacy preservation when the recovered images from its gradient updates are less similar to the ground truth images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.895, + 0.875 + ], + "angle": 0, + "content": "Given an image \\(\\mathbf{x}\\) and its transformed (i.e., patch shuffled) version \\(\\tilde{\\mathbf{x}}\\), a ViT model \\(\\mathcal{M}\\), and automatic evaluation metrics \\(\\phi\\), we conduct three different settings for fair comparisons: (a) \\(\\phi(\\nabla \\mathcal{M}(\\mathbf{x}), \\mathbf{x})\\), (b) \\(\\phi(\\nabla \\mathcal{M}(\\tilde{\\mathbf{x}}), \\tilde{\\mathbf{x}})\\), and (c) \\(\\phi(\\nabla \\mathcal{M}(\\tilde{\\mathbf{x}}), \\mathbf{x})\\), where \\(\\nabla\\) refers to recovering input image through gradient attacks. Table 5 shows the quantitative comparisons between our method and the original ViTs for batch gradient inversion on ImageNet-1K. APRIL [29] enables a viable, complete recovery of original images from the gradient updates of the original ViTs. However, it per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.85, + 0.901 + ], + "angle": 0, + "content": "4https://github.com/JonasGeiping/breaching" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20388" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.852, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.321, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons on image recovery with gradient updates [29]. Our proposed DeiT-S+MJP model significantly outperforms the original ViT-S [7] and DeiT-S [37] models." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.361, + 0.468, + 0.376 + ], + "angle": 0, + "content": "Table 6. Explained variance versus PCA projected dimensionality." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.379, + 0.462, + 0.432 + ], + "angle": 0, + "content": "
Projected Dimension34567
DeiT-S EV (%)54.6168.5577.9585.5490.74
DeiT-S+MJP EV (%)46.7458.3669.1078.1384.55
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.47, + 0.482 + ], + "angle": 0, + "content": "forms worse in recovering from \"DeiT-S+MJP\", leading to best performances on all evaluation metrics and outperform others by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.473, + 0.696 + ], + "angle": 0, + "content": "More surprisingly, our proposed method makes APRIL yield unrecognizable images and fail in recovering the details in the original images (i.e., noisy patches in the outputs), as shown in Fig. 5. The left four columns in Fig. 5 are tested on original images, where all PEs are standard and correspond to their patch embeddings. Meanwhile, the right four columns are tested with transformed ones, where the shuffled patches are with the shared unknown PEs. Both the visual and quantitative comparisons verify that our MJP alleviates the gradient leakage problem. We also notice that DeiT-S without using PEs is inclined to be at higher risk of privacy leakage (i.e., easier to be attacked by gradients). These promising results indicate that our MJP is a promising strategy to protect user privacy in federated learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.196, + 0.723 + ], + "angle": 0, + "content": "5.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.868 + ], + "angle": 0, + "content": "PCA Projected Dimensionality. Table 6 presents the explained variance (EV) of our DeiT-S+MJP and DeiT-S versus different projection dimension. A low dimensionality can explain a large amount of information, which proves that the embedding matrix is sparse in nature. Moreover, to achieve the same explained variance ratio, our DeiT-S+MJP needs a large dimensionality than DeiT-S. This indicates that the positional embedding matrix of DeiT-S+MJP is less sparse but more informative." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Accuracy Vs. Shuffle Ratio. Intuitively, with the increase of the shuffle ratio from the proposed MJP method, the orig-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.895, + 0.545 + ], + "angle": 0, + "content": "inal intrinsic inductive bias will be undermined. However, from the experimental results, the performance of ViTs is actually boosted via the proposed MJP. To figure out the reason behind such counter-intuitive phenomenon, we visualize the last self-attention of our proposed method in Fig. 3 of our Supp. Mat. It shows that the attention heads of our method present more diverse and content-aware attentions than original DeiT-S. We think that it might be attributed to the proposed DAL, which strengthens the spatial information of the unmasked PEs in a more efficient manner. To this end, such learned content-aware attention becomes more meaningful and results in the better accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.62, + 0.574 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.856 + ], + "angle": 0, + "content": "In this paper, we first visually demonstrate that PEs can explicitly learn the 2D spatial relationship from the input patch sequences. By feeding ViTs with transformed input, we identify the issue that PEs may weaken the position-insensitive property. Based on this observation, we propose an easy-to-reproduce yet effective Masked Jigsaw Puzzle (MJP) position embedding method to alleviate the conflict in PEs (preserving the consistency versus maintaining the accuracy). Experimental results show the proposed MJP can bring the the position-insensitive property back to ViTs without degenerating the accuracy on the large-scale dataset (i.e., ImageNet-1K). In a certain sense, the proposed MJP is also a data augmentation technique, which boosts the robustness to the common corruptions (e.g., ImageNet-C) and adversarial examples (e.g., ImageNet-A/O). Surprisingly, MJP can improve the privacy preservation capacity of ViTs under typical gradient attacks by a large margin, which may pilot a new direction for privacy preservation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was partly supported by the National AI Ph.D. for Society Program of Italy, and the EU H2020 project AI4Media (No. 951911)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20389" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations (ICLR), 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning (ICML), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[5] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.472, + 0.471 + ], + "angle": 0, + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2020. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.471, + 0.6 + ], + "angle": 0, + "content": "[8] Philipp Dufter, Martin Schmitt, and Hinrich Schütze. Position information in transformers: An overview. Computational Linguistics, 48(3):733-763, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.471, + 0.671 + ], + "angle": 0, + "content": "[9] Yuxin Fang, Benchcheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection. Advances in Neural Information Processing Systems, 34:26183-26197, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.673, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[10] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning (ICML), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.731, + 0.471, + 0.785 + ], + "angle": 0, + "content": "[11] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.788, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[12] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 244-253, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[13] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on visual transformer. arXiv, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[14] Ali Hatamizadeh, Hongxu Yin, Holger Roth, Wenqi Li, Jan Kautz, Daguang Xu, and Pavlo Molchanov. Gradvit: Gradient inversion of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.894, + 0.346 + ], + "angle": 0, + "content": "[17] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[18] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.893, + 0.446 + ], + "angle": 0, + "content": "[19] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[20] Yifan Jiang, Shiyu Chang, and Zhangyang Wang. Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.893, + 0.56 + ], + "angle": 0, + "content": "[21] Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. Global-and-local relative position embedding for unsupervised video summarization. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[22] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 54(10s):1-41, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[23] Shun Kiyono, Sosuke Kobayashi, Jun Suzuki, and Kentaro Inui. Shape: Shifted absolute position embedding for transformers. arXiv preprint arXiv:2109.05644, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[25] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[26] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[27] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20390" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[29] Jiahao Lu, Xi Sheryl Zhang, Tianli Zhao, Xiangyu He, and Jian Cheng. April: Finding the achilles' heel on privacy for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10051-10060, 2022. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[30] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.361 + ], + "angle": 0, + "content": "[31] Samrudhdhi B Rangrej, Chetan L Srinidhi, and James J Clark. Consistency driven sequential transformers attention model for partially observable scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2518-2527, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[32] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[33] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[34] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.472, + 0.617 + ], + "angle": 0, + "content": "[35] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.472, + 0.688 + ], + "angle": 0, + "content": "[36] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems (NeurIPS), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[37] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. 1, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[40] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[41] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.894, + 0.359 + ], + "angle": 0, + "content": "[43] Yu-An Wang and Yun-Nung Chen. What do position embeddings learn? an empirical study of pre-trained language model positional encoding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.894, + 0.414 + ], + "angle": 0, + "content": "[44] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.417, + 0.894, + 0.471 + ], + "angle": 0, + "content": "[45] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. Advances in Neural Information Processing Systems, 34:28522-28535, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.473, + 0.894, + 0.541 + ], + "angle": 0, + "content": "[46] Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.894, + 0.611 + ], + "angle": 0, + "content": "[47] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.894, + 0.682 + ], + "angle": 0, + "content": "[48] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.684, + 0.894, + 0.753 + ], + "angle": 0, + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.755, + 0.894, + 0.837 + ], + "angle": 0, + "content": "[50] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.839, + 0.894, + 0.893 + ], + "angle": 0, + "content": "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20391" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_origin.pdf b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3672716ba9b0cfdefda1a7ee925963490634738e --- /dev/null +++ b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/e08baece-e2f2-4620-85a3-120ecb725f4e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c093421fb0875ab82a0d8ba49d6ad96c37cf1de1c094b00557f987e6a8fb55b +size 5164969 diff --git a/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/full.md b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5f0840b02d214e483e612a662ac61066a2a00fbb --- /dev/null +++ b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/full.md @@ -0,0 +1,321 @@ +# Masked Jigsaw Puzzle: A Versatile Position Embedding for Vision Transformers + +Bin Ren $^{1,2*}$ Yahui Liu $^{2*}$ Yue Song $^{2}$ Wei Bi $^{3}$ Rita Cucchiara $^{4}$ Nicu Sebe $^{2}$ Wei Wang $^{5\dagger}$ $^{1}$ University of Pisa, Italy $^{2}$ University of Trento, Italy + $^{3}$ Tencent AI Lab, China $^{5}$ Beijing Jiaotong University, China + $^{4}$ University of Modena and Reggio Emilia, Italy + +# Abstract + +Position Embeddings (PEs), an arguably indispensable component in Vision Transformers (ViTs), have been shown to improve the performance of ViTs on many vision tasks. However, PEs have a potentially high risk of privacy leakage since the spatial information of the input patches is exposed. This caveat naturally raises a series of interesting questions about the impact of PEs on accuracy, privacy, prediction consistency, etc. To tackle these issues, we propose a Masked Jigsaw Puzzle (MJP) position embedding method. In particular, MJP first shuffles the selected patches via our block-wise random jigsaw puzzle shuffle algorithm, and their corresponding PEs are occluded. Meanwhile, for the nonoccluded patches, the PEs remain the original ones but their spatial relation is strengthened via our dense absolute localization regressor. The experimental results reveal that 1) PEs explicitly encode the 2D spatial relationship and lead to severe privacy leakage problems under gradient inversion attack; 2) Training ViTs with the naively shuffled patches can alleviate the problem, but it harms the accuracy; 3) Under a certain shuffle ratio, the proposed MJP not only boosts the performance and robustness on large-scale datasets (i.e., ImageNet-1K and ImageNet-C, -A/O) but also improves the privacy preservation ability under typical gradient attacks by a large margin. The source code and trained models are available at https://github.com/yh1leo/MJP. + +# 1. Introduction + +Transformers [38] demonstrated their overwhelming power on a broad range of language tasks (e.g., text classification, machine translation, or question answering [22, 38]), and the vision community follows it closely and extends it for vision tasks, such as image classification [7,37], object detection [2, 51], segmentation [47], and image generation [3, 24]. Most of the previous ViT-based methods focus on designing different pre-training objectives [9, 11, 12] or variants of self-attention mechanisms [28, 39, 40]. By contrast, PEs + +![](images/187412b629ee88a0d519ae7486148841bf1c2706c67e299ff596f60225b9fcdd.jpg) +Figure 1. Low-dimensional projection of position embeddings from DeiT-S [37]. (a) The 2D UMAP projection, it shows that reverse diagonal indices have the same order as the input patch positions. (b) The 3D PCA projection, it also shows that the position information is well captured with PEs. Note that the embedding of index 1 (highlighted in red) corresponds to the [CLS] embedding that does not embed any positional information. + +![](images/74102392c2115e8f800f0639c893106513209ebac9fc0455140eaa6551e6c9e0.jpg) + +receive less attention from the research community and have not been well studied yet. In fact, apart from the attention mechanism, how to embed the position information into the self-attention mechanism is also one indispensable research topic in Transformer. It has been demonstrated that without the PEs, the pure language Transformer encoders (e.g., BERT [6] and RoBERTa [25]) may not well capture the meaning of positions [43]. As a consequence, the meaning of a sentence can not be well represented [8]. Similar phenomenon of PEs could also be observed in vision community. Dosovitskiy et al. [7] reveals that removing PEs causes performance degradation. Moreover, Lu et al. [29] analyzed this issue from the perspective of user privacy and demonstrated that the PEs place the model at severe privacy risk since it leaks the clues of reconstructing sequential patches back to images. Hence, it is very interesting and necessary to understand how the PEs affect the accuracy, privacy, and consistency in vision tasks. Here the consistency means whether the predictions of the transformed/shuffled image are consistent with the ones of the original image. + +To study the aforementioned effects of PEs, the key is to figure out what explicitly PEs learn about positions from input patches. To answer this question, we project the high-dimensional PEs into the 2D and 3D spaces using Uni- + +![](images/7b5c1df59a6fd2e20c0cfba72b50c21c339b6fcad9fd5fd3999bd08718f6f806.jpg) +(a) $\mathbf{x}$ + +![](images/e8cf282277e3393e50d87cd69d69998ac94c48fb332535d84a129b64afc6d4f2.jpg) +(b) $\widetilde{\mathbf{x}} (\gamma = 1)$ + +![](images/4b91c42dd0362a306dfa83289a418a28a3c731bd986ca25fa9ee599beea44186.jpg) +(c) $\widetilde{\mathbf{x}} (\gamma = 0.4)$ +Figure 2. (a) The original input patches; (b) Totally random shuffled input patches; (c) Partially random shuffled input patches; (d) An overview of the proposed MJP. Note that we show the random shuffled patches and its corresponding unknown position embedding in green and the rest part in blue. DAL means the self-supervised dense absolute localization regression constraint. + +![](images/513b953e5d3546cd6099861512244379cb0a5ac27e1481a30b0f4db305f0da6a.jpg) +(d) Overview of our MJP + +form Manifold Approximation & Projection (UMAP) [30] and PCA, respectively. Then for the first time, we visually demonstrate that the PEs can learn the 2D spatial relationship very well from the input image patches (the relation is visualized in Fig. 1). We can see that the PEs are distributed in the same order as the input patch positions. Therefore, we can easily obtain the actual spatial position of the input patches by analyzing the PEs. Now it explains why PEs can bring the performance gain for ViTs [7]. This is because the spatial relation in ViTs works similar as the inherent intrinsic inductive bias in CNNs (i.e., it models the local visual structure) [45]. However, these correctly learned spatial relations are unfortunately the exact key factor resulting in the privacy leakage [29]. + +Based on these observations, one straightforward idea to protect the user privacy is to provide ViTs with the randomly transformed (i.e., shuffled) input data. The underlying intuition is that the original correct spatial relation within input patches will be violated via such a transformation. Therefore, we transform the previous visually recognizable input image patches $\mathbf{x}$ shown in Fig. 2(a) to its unrecognizable counterpart $\widetilde{\mathbf{x}}$ depicted in Fig. 2(b) during training. The experimental results show that such a strategy can effectively alleviate the privacy leakage problem. This is reasonable since the reconstruction of the original input data during the attack is misled by the incorrect spatial relation. However, the side-effect is that this leads to a severe accuracy drop. + +Meanwhile, we noticed that such a naive transformation strategy actually boosts the consistency [31, 34, 44] albeit the accuracy drops. Note that here the consistency represents if the predictions of the original and transformed (i.e., shuffled) images are consistent. Given the original input patches $\mathbf{x}$ and its corresponding transformed (i.e., shuffled) counterpart, we say that the predictions are consistent if $\arg \max P(\mathcal{F}(\mathbf{x})) = \arg \max P(\mathcal{F}(\widetilde{\mathbf{x}}))$ , where $\mathcal{F}$ refers to the ViT models, and $P$ denotes the predicted logits. + +These observations hint that there might be a trade-off solution that makes ViTs take the best from both worlds (i.e., both the accuracy and the consistency). Hence, we pro + +pose the Masked Jigsaw Puzzle (MJP) position embedding method. Specifically, there are four core procedures in the MVP: (1) We first utilize a block-wise masking method [1] to randomly select a partial of the input sequential patches; (2) Next, we apply jigsaw puzzle to the selected patches (i.e., shuffle the orders); (3) After that, we use a shared unknown position embedding for the shuffled patches instead of using their original PEs; (4) To well maintain the position prior of the unshuffled patches, we introduce a dense absolute localization (DAL) regressor to strengthen their spatial relationship in a self-supervised manner. We simply demonstrate the idea of the first two procedures in Fig. 2(c), and an overview of the proposed MVP method is available in Fig. 2(d). In summary, our main contributions are: + +- We demonstrate that although PEs can boost the accuracy, the consistency against image patch shuffling is harmed. Therefore, we argue that studying PEs is a valuable research topic for the community. +- We propose a simple yet efficient Masked Jigsaw Puzzle (MJP) position embedding method which is able to find a balance among accuracy, privacy, and consistency. +- Extensive experimental results show that MJP boosts the accuracy on regular large-scale datasets (e.g., ImageNet-1K [32]) and the robustness largely on ImageNet-C [18], -A/O [19]. One additional bonus of MJP is that it can improve the privacy preservation ability under typical gradient attacks by a large margin. + +# 2. Related Work + +# 2.1. Vision Transformers + +Transformers [38], originally designed for Nature Language Processing (NLP) tasks, have recently shown promising performance on computer vision tasks [13, 22]. Benefiting from the strong representation power of modelling global relations between image patches, Vision Transformers (ViTs) [7] have achieved superior performance than their counterpart CNNs on image classification and various other + +downstream tasks (e.g., object detection [2, 51], object re-identification [17], dense prediction [41, 42, 46, 50] and image generation [3, 4, 20, 24]). + +As a core module in ViTs, multi-head self-attention (MSAs) [7, 38] aggregate sequential tokens with normalized attentions as: $z_{j} = \sum_{i}\mathrm{Softmax}(\frac{QK}{\sqrt{d}})_iV_{i,j}$ where $Q$ , $K$ and $V$ are query, key and value matrices, respectively. $d$ is the dimension of query and key, and $z_{j}$ is the $j$ -th output token. In theory, when the position information is not considered, the outputs of MSAs should be strictly invariant to the input sequence order (i.e., position-insensitivity). This indicates that a visually recognizable image can be transformed into its unrecognizable counterpart by permuting the order of image patches while maintaining the performance delivered by ViTs compared with the ViTs trained on the original non-permuted image. However, the usage of PEs hinders such implementations, where the outputs of the ViTs vary dramatically with the mentioned naive transformations shown in Fig. 2(b). In this work, our main focus is to explicitly figure out what PEs actually learn from input patches about positions, and how the PEs affect the accuracy, privacy, and the consistency properties of ViTs. + +# 2.2. Position Embeddings + +In Transformer networks, both the attention and the (individual token based) feed-forward layers are permutation invariant when the position information is not considered. In this way, the spatial relationships between image patches could not be modeled as the position information is completely discarded. As compensation, PEs are naturally introduced into ViTs to provide information about the token order during the learning process, since it offers possibilities for dependency modeling between elements at different positions. For example, previous works [10,33,38] indicated that PEs are useful to give the model a sense of which portion of the sequence in the input/output it is currently dealing with. Inspired by this, some works [15,21,23,27,35] showed diverse application scenarios that benefit from the usage of suitable PEs. In addition, Chu et al. [5] proposed correlating the PEs with their local neighborhood of the input sequence. Liu et al. [26] proposed to enhance the spatial prior (i.e., relative localization) in the final content embedding to indirectly enrich the inductive bias. Obviously, although these methods enhance the position information learnt by PEs, they indeed degenerate the position-insensitive property of MSAs. + +Especially, Wang et al. [43] revealed that Transformer encoders (e.g., BERT [6] and RoBERTa [25]) may not well capture the meaning of positions (absolute and relative positions). They showed that Transformer encoders learn the local position information that can only be effective in masked language modeling. In contrast, there does not exist such a similar "masked language modeling" procedure in ViTs (and VTs) in the typically supervised pre-training. Moreover, Lu + +et al. [29] revealed that the learnable PEs place the model at severe privacy risk, which leaks the clues of reconstructing sequential patches to images. In this paper, we dive into the usage of PEs and propose an efficient position embedding method, MJP, to improve the position-insensitive property of ViTs without hurting the positive effects of PEs. + +# 3. Preliminaries: 2D/3D Spatial Priors in PEs + +Although numerous works claim that the PEs can learn the 2-D spatial relationship of image patches, this claim has not been demonstrated visually or mathematically. To visualize the concrete relation of image patches captured in the high-dimensional position embedding, we project them into the 2-D and 3-D spaces using Uniform Manifold Approximation and Projection (UMAP) [30] and PCA, respectively. Fig. 1 displays the projections of PEs from DeiT-S [37]. Note that [7] only shows the cosine similarity between the PEs, which is not exactly the 2D spatial information mentioned in our paper. Here, we explicitly specify the spatial relationship as the relative distance in the 2D/3D coordinate space. As shown in Fig. 1, the spatial relationship can be projected into a 2D/3D coordinate system and indicate the localizations of these embeddings. For the UMAP, the projected positional embedding emerges as grid-like structures with the distances between each point roughly the same, which is coherent with the relation of embedded image patches. For the PCA, the position embedding that corresponds to neighboring image patches groups together and aggregates to the stick-like form. This phenomenon demonstrates that the spatial relationship is indeed learned by the PEs. Moreover, this also implies that the spatial relationship learned in the high-dimensional space still manifests in the low-dimensional space. + +The learnable PEs, which work as a lookup table of a dictionary, maps the 1-dimensional data into the sparse high-dimensional space. By nature, it is a sparse large matrix. Dimensionality reduction techniques can capture the structural information of sparse matrices using low dimensionalities. For the PCA, the amount of information retained in the projection can be measured by the ratio of explained variance. Formally, we have the definition as follows: + +Definition 1 (Explained Variance) Let $\mathbf{P}$ and $\mathbf{V}$ denote the data matrix and the PCA projection matrix. The ratio of explained variance that $\mathbf{PV}$ accounts for is defined as $\sum \sigma (\mathbf{PV})^2 /\sum \sigma (\mathbf{P})^2$ where $\sigma (\cdot)$ denotes the singular value. + +In practice, we observe that the 3-dimensional PCA projection explains $54.6\%$ of the total variance of the DeiT-S embedding matrix. Given that a large amount of information can be captured by the low-dimensional projection, we propose that explicitly enforcing low-dimensional positional prior can help the positional learning in the high-dimensional space, which might accelerate the convergence rate of training and improve the performance (See Sec. 4.2). + +Algorithm 1 Block-wise Random Jigsaw Puzzle Shuffle +Input: Input image: $\mathbf{x}\in \mathbb{R}^{H\times W\times C}$ Shuffle Ratio: $\gamma$ Patch Size: $P$ +Output: Shuffled image patches: $\widetilde{\pmb{x}}_p$ +1: $\mathbf{x}_p\in \mathbb{R}^{N\times (P^2\cdot C)}\gets$ Patchlize(x,P) +2: $\mathbf{m}\in \mathbb{R}^{\frac{H}{P}\times \frac{W}{P}}\gets$ BinaryInitialize(xp,0) +3: $\widetilde{\mathbf{m}}\in \mathbb{R}^{\frac{H}{P}\times \frac{W}{P}}\gets$ BlockwiseMask(m,γ) [25] +4: $\widetilde{\pmb{x}}_p\in \mathbb{R}^{N\times (P^2\cdot C)}\gets$ Jigsaw Puzzle(xp, $\widetilde{\mathbf{m}}$ 5: return $\widetilde{\pmb{x}}_p$ + +# 4. Method + +Based on the design principle of MSAs, the outputs of MSAs should be entirely position-agnostic. However, PEs hinder such a property because it can learn the strong 2-D spatial relation of the input patches (as we demonstrated in Sec. 3). To this end, we propose the block-wise random jigsaw puzzle shuffle algorithm (See Alg. 1) to transform the input patches with different shuffle ratios $\gamma$ for intermingling the original correct spatial relation. + +Since we experimentally demonstrate that the totally shuffled strategy (i.e., $\gamma = 1.0$ ) will degenerate the accuracy a lot albeit the consistency increases. As a remedy, we only shuffle portion of the sequence patches, and strengthen the spatial relation of the rest part with a dense absolute localization regression. Finally, a versatile position embedding method MJP is proposed. The detailed analysis of each module is available in the ablation study in Sec. 5.3. In the following sub-sections, we will introduce the Jigsaw Puzzle Shuffle method (Sec. 4.1), the spatial relation strengthen method (Sec. 4.2), and our final MJP method (Sec. 4.3). + +# 4.1. Block-wise Random Jigsaw Puzzle Shuffle + +Specifically, given an input image $\mathbf{x} \in \mathbb{R}^{H \times W \times C}$ , we first reshape it into a sequence of flattened 2D patches $\mathbf{x}_p \in \mathbb{R}^{N \times (P^2 \cdot C)}$ , where $(H, W)$ is the resolution of the original image, $C$ is the number of channels, $(P, P)$ is the resolution of each image patch, and then we have $N = HW / P^2$ , which denotes the number of patches. Then instead of directly applying block-wise masking method [1] to the image, we first initialize a binary mask matrix $\mathbf{m} \in \mathbb{R}^{\frac{H}{P} \times \frac{W}{P}}$ with the same size as the image patches in $\mathbf{x}_p$ . Next, we use [1] to update the binary mask $\mathbf{m}$ in which the masked positions will be set to 1 and the rest untouched positions remains 0. The hyper-parameter $\gamma$ is used to control the ratio of selected positions. After that, a jigsaw puzzle shuffle operation is applied to $\mathbf{x}_p$ conditioned on the updated binary mask $\widetilde{\mathbf{m}}$ . + +Finally, we get the shuffled patch sequence $\widetilde{\mathbf{x}}_p$ , where the shuffled patches are actually $\{\mathbf{x}_p^i |\widetilde{\mathbf{m}}_i = 1\}$ . Notably, in [1], the patches covered by the sampled mask are not visible to the encoder module, while in our method, the masking + +strategy [1] is only used to mask out the positions/indices of the selected patches. These patches are still visible to the encoder module and they are randomly shuffled in a jigsaw puzzle manner. One intuitive shuffled toy example is shown in Fig. 2(c), where the green part means the selected shuffled region while the blue part remains unchanged. Actually, this algorithm shares a similar idea with the data augmentation method as the training images are always changing as their patches are randomly shuffled during each iteration. + +# 4.2. Strengthening Spatial Prior in PEs + +Liu et al. [26] noticed that by enhancing the 2-D spatial information of the output embeddings of the last layer of ViTs, the training convergence speed can be accelerated. Inspired by their work, we observe that similar gains can be obtained by applying a low-dimensional spatial prior in PEs. Different from the dense relative localization constraint in [26] which samples relative pairs from the whole output sequential embeddings, we propose a much simpler dense absolute localization (DAL) regression method to directly use self-supervised absolute location to enhance the spatial information in PEs, which avoids sampling relative pairs. + +Since the PEs capture the absolute position of the input patches, to some extent, the position information could be reconstructed via a reversed mapping function $g(\cdot):\mathcal{E}\to \mathcal{P}$ where $\mathcal{E}$ and $\mathcal{P}$ are embedding space and position space, respectively. Given that PEs have one-to-one correspondence with the sequential image patches, we can reshape them into $\mathbf{E}_{\mathrm{pos}}\in \mathbb{R}^{K\times K\times D}$ , where $K$ refers to height/weight of the grid and $D$ refers to latent vector size. Then we can compute the reverse mapping from $\mathcal{E}\rightarrow \mathcal{P}$ via: + +$$ +\left(\widetilde {i}, \widetilde {j}\right) ^ {T} = g \left(\mathbf {E} _ {\text {p o s}} ^ {i, j}\right), \tag {1} +$$ + +in which $(\widetilde{i},\widetilde{j})^T$ is the predicted patch position, and $\mathbf{E}_{\mathrm{pos}}^{i,j}$ is the position embedding of the patch $(i,j)$ in the $K\times K$ grid. The dense absolute localization (DAL) loss is: + +$$ +\mathcal {L} _ {\mathrm {D A L}} = \mathbb {E} _ {\mathbf {E} _ {\text {p o s}}, 1 \leq i, j \leq K} ^ {i, j} [ \| (i, j) ^ {T} - (\widetilde {i}, \widetilde {j}) ^ {T} \| _ {1} ], \tag {2} +$$ + +where the expectation is computed by averaging the $\ell_1$ loss between the correspond $(i,j)^T$ and $(\widetilde{i},\widetilde{j})^T$ . Then, $\mathcal{L}_{\mathrm{DAL}}$ is added to the standard cross-entropy loss $(\mathcal{L}_{\mathrm{CE}})$ of the native ViTs. The final loss is: $\mathcal{L}_{\mathrm{all}} = \mathcal{L}_{\mathrm{CE}} + \lambda \mathcal{L}_{\mathrm{DAL}}$ , where we set $\lambda = 0.01$ for all experiments. Note that the mapping function can be either linear or nonlinear. Throughout this work, we mainly discuss three implementations, including non-parametric PCA, learnable linear (LN), and nonlinear (NLN) projection layers. They will be discussed in details in Sec. 5.3. + +# 4.3. MJP Position Embedding + +The main goal of MJP position embedding in our work is to enhance the consistency (i.e., position-insensitive property) of ViTs and preserve the accuracy (keeping the spatial + +Algorithm 2 The pipeline of the proposed MJP. +Input: Input image: $\mathbf{x}\in \mathbb{R}^{H\times W\times C}$ Shuffle Ratio: $\gamma$ ; Patch Size: $P$ 1: $\widetilde{\pmb{x}}_p\gets Alg.1(\mathbf{x},P,\gamma)$ // 1st & 2nd procedures 2: $\mathbf{E}_{\mathrm{unk}}(\widetilde{\pmb{x}}_p)$ // 3rd procedure 3: $\mathbf{DAL}(\mathbf{x} - \mathbf{x}\cap \widetilde{\pmb{x}}_p)$ // 4th procedure, only for train + +relation modeling functionality of PEs) for the standard classification tasks. Usually, with original input image patches $\mathbf{x}$ we can formulate the function of the input layer which is located before the Transformer block as follows: + +$$ +\mathbf {z} _ {0} = \left[ \mathbf {x} _ {\mathrm {C L S}}; \mathbf {x} _ {p} ^ {1} \mathbf {E}; \mathbf {x} _ {p} ^ {2} \mathbf {E}, \dots , \mathbf {x} _ {p} ^ {N} \mathbf {E} \right] + \mathbf {E} _ {\text {p o s}}, \tag {3} +$$ + +where $\mathbf{E} \in \mathbb{R}^{(P^2 \cdot C) \times D}$ is a trainable linear projection layer and $\mathbf{x}_p^k\mathbf{E}$ refers to the output of linear projection (i.e., the patch embeddings), and $\mathbf{E}_{\mathrm{pos}} \in \mathbb{R}^{(N + 1) \times D}$ refers to PEs (the additional one is applied to the [CLS] embedding). + +Next, we apply the proposed block-wise random jigsaw puzzle shuffle algorithm to $\mathbf{x}_p$ and produce the transformed patch sequences $\widetilde{\mathbf{x}}_p$ with the shuffle ratio $\gamma$ . In this scenario, if we maintain the original position embedding sequence, it leads to a mismatch issue between the shuffled patches and the position embedding sequence. Therefore, we introduce a shared unknown position embedding to the shuffled positions to alleviate the mismatching issue. With the corresponding updated mask $\widetilde{\mathbf{m}}$ from Alg. 1 (1st & 2nd procedures), we propose our MJP PEs: + +$$ +\widetilde {\mathbf {E}} _ {\text {p o s}} ^ {i} = \left\{ \begin{array}{l l} \mathbf {E} _ {\text {p o s}} ^ {i}, & \text {i f} \widetilde {\mathbf {m}} _ {i} = 0 \\ \mathbf {E} _ {\text {u n k}}, & \text {i f} \widetilde {\mathbf {m}} _ {i} = 1 \end{array} \right. \tag {4} +$$ + +where $\mathbf{E}_{\mathrm{unk}} \in \mathbb{R}^{1 \times D}$ refers to a share learnable embedding (i.e., unknown position embedding). It represents that the image patch in this position has random permutation and its position should be occluded (3rd procedure). $\mathbf{E}_{\mathrm{pos}}$ is the original position embedding for the rest image patches. + +Besides, we revisit the remaining PEs (i.e., corresponding to the non-selected patches) and apply low-dimensional prior on them. The low-dimensional prior is imposed by the proposed DAL (Sec. 4.2) regression method for strengthening the spatial relation (4th procedure). A toy illustration of the proposed MJP is shown in Fig. 2(d), where the green color represents the randomly permuted patches and its corresponding unknown PEs, while the blue color indicates the rest regular patches and its related original PEs. We also formalize these procedures as an algorithm in Alg. 2. + +Thus, we replace the input layer with a new formulation: + +$$ +\widetilde {\mathbf {z}} _ {0} = \left[ \mathbf {x} _ {\mathrm {C L S}}; \widetilde {\mathbf {x}} _ {p} ^ {1} \mathbf {E}; \widetilde {\mathbf {x}} _ {p} ^ {2} \mathbf {E}, \dots , \widetilde {\mathbf {x}} _ {p} ^ {N} \mathbf {E} \right] + \widetilde {\mathbf {E}} _ {\text {p o s}}. \tag {5} +$$ + +the following procedures and modules are exactly the same as the ones in the original ViTs. + +Table 1. Comparisons of different backbones on ImageNet-1K classification. Note that the image size here are all set to $224 \times 224$ . + +
MethodParam. Top-1 Acc.↑ Diff.Norm.↓ Consistency
ResNet-50 [16]2579.311.7751.5
ResNet-50 + MJP2579.47.1169.3
DeiT-S [37]2279.816.2164.3
DeiT-S + MJP2280.58.9682.9
Swin-T [28]2981.315.4941.5
Swin-T + MJP2981.312.3666.9
+ +# 5. Experiments + +We follow the typical supervised pre-training procedure, where all the compared models are trained on ImageNet1K [32] to show the capacity of our proposed MJP method. We also benchmark the proposed MJP method on ImageNet1K, which contains 1.28M training images and 50K validation images of 1,000 classes. The training details mostly follow the training protocols from Touvron et al. [37]. + +# 5.1. Regular ImageNet-1K training + +We mainly compare with three typical existing methods, including two state-of-the-art Visual Transformers (i.e., DeiT [37] and Swin [28]) and one widely-used CNN-based ResNet-50 [16]. All these methods are of comparable sizes (i.e., less than 30M parameters). Besides the common Top-1 accuracy (Top-1 Acc.), we also report another two evaluation metrics (i.e. Diff. Norm. and Consistency) to show the position invariance of a model to the jigsaw puzzle transformation. For a given input image $\mathbf{x}$ , we collect its counterpart $\widetilde{\mathbf{x}}$ by applying masked jigsaw puzzle to shuffle a portion of selected patches. Next, we calculate the difference $\ell_2$ -norm (i.e., Diff. Norm.) between the [CLS] embedding inferred from $\mathbf{x}$ and $\widetilde{\mathbf{x}}$ : + +$$ +\left\| \mathcal {F} _ {\mathrm {C L S}} (\mathbf {x}) - \mathcal {F} _ {\mathrm {C L S}} (\widetilde {\mathbf {x}}) \right\| _ {2} ^ {2}. \tag {6} +$$ + +For Consistency, we measure how many classification results stay the same (\%) given the perturbation: + +$$ +\mathbb {E} _ {\mathbf {x}} \left[ \mathbb {I} \left\{\arg \max P (\mathcal {F} (\mathbf {x})) = \arg \max P (\mathcal {F} (\widetilde {\mathbf {x}})) \right\} \right]. \tag {7} +$$ + +$P(\mathcal{F}(\mathbf{x}))$ denotes the predicted classification probability of image $\mathbf{x}$ , and $\arg \max P(\mathcal{F}(\mathbf{x}))$ represents the predicted class index. For a fair comparison, we use $\gamma = 0.15$ to create $\widetilde{\mathbf{x}}$ in Table 1. According to these results, the proposed MJP does not have a negative effect on the top-1 accuracy (MJP even brings a marginal improvement for DeiT-S). More importantly, it improves Diff. Norm. and Consistency by a large margin. Besides, MJP works well in the variants of ViTs (e.g., Swin [28]), which shows a good potential generalization ability. Note that we add MJP to ResNet50 by only shuffling the image patches with Alg. 1 (No PEs). For Swin-T, we add MJP to the absolute PEs of Swin-T. + +![](images/edfe94efef2ab9d518d4ef5f73138de4ef5ac5f4e16cf4b10710f6512243d02e.jpg) +(a) + +![](images/94b9817bd27d7a2a49267dad77b19b2a1a115dc3340f7a9e6a0c810ac77ec96f.jpg) +(b) +Figure 3. Ablation on the mask ratio $\gamma$ during inference: (a) comparisons among ViT-S, DeiT-S and our method (trained with $\gamma = 0.03$ ); (b) comparisons between Swin-T and our method (trained with $\gamma = 0.03$ ); (c) comparisons of our method on DeiT-S trained with different $\gamma$ . + +![](images/4d99d083d74bafd54874aea96b7c1df4fab1377996c00b8e9d4dd94c1508b8d2.jpg) +(c) + +Table 2. Comparisons on robustness to common corruptions and adversarial examples. + +
MethodImageNet-CImageNet-AImageNet-O
mCE ↓Acc ↑AURRA ↑AUPR ↑
DeiT-S54.619.225.120.9
DeiT-S + MJP40.7821.629.822.6
+ +# 5.2. Robustness on Challenging Sets + +Besides the strength on standard classification, we also observe an auxiliary benefit of the proposed MJP on robustness. ImageNet-C [18] benchmarks a classifier's robustness to common corruptions2. The mean Corruption Error (mCE) is used to measure the generalization of a model at corrupted distributions (the lower the better). + +ImageNet-A/O [19] focus on adversarial examples, and it enables us to either test image classification performance when the input data distribution shifts (i.e., Acc and AU-RRA), or test out-of-distribution detection performance when the label distribution shifts (i.e., AUPR) $^3$ . As shown in Table 2, the proposed MJP not only boosts the accuracy on the standard evaluation on ImageNet-1K validation set but also improves the robustness largely on the adversarial samples. + +The underlying reason might be that MJP enforces the ViTs aware of both local and global context features, and it helps ViTs to get rid of some unnecessary sample-specific local features during the training. This has been verified by the visualization maps (see Fig. 3 of our Supp. Mat.). + +# 5.3. Ablation Analysis + +Results with different MJP ratios. As shown in Table 3, we test different masking ratios used in the block-wise masking strategy during the training. Obviously, Diff. Norm. has the inverse tendencies compared to Consistency, where a smaller Diff. Norm. usually indicates a larger (better) consistency score. Comparing the accuracy trained with $\gamma >0$ , it shows our model is not sensitive to different $\gamma$ . In particular, a + +Table 3. Ablation study on the proposed MJP method trained with different masking ratio $\gamma$ . + +
MetricMasking Ratio
00.030.090.150.210.27
Top-1 Acc.80.080.580.380.480.280.3
Diff. Norm.16.568.966.365.234.393.97
Consistency64.082.988.190.592.393.1
+ +small ratio (e.g., $\gamma = 0.03$ ) is sufficient for boosting the accuracy. In addition, a large ratio can reduce the Diff. Norm. and improves the consistency by a large margin. Moreover, the model trained with a larger $\gamma$ is inclined to be more consistent as shown in Fig. 3 (c). In addition, we observe that the consistency keeps increasing when the mask ratio increases from $3\%$ to $27\%$ . However, further increasing the mask ratio will not bring consistency improvement and what is worse is that the accuracy marginally decreases. + +For a model trained with a fixed $\gamma$ , we also test its accuracy with different masking ratios during the inference. As shown in Fig. 3 (a) and (b), the performances of original models drop significantly when we shuffle more image patches. In contrast, our proposed method shows more consistent performances. + +Comparison of variants of MJP. We test several variants of the proposed MJP, including (1) Removing the PEs from the original DeiT-S; (2) SPP: shuffling both the $16 \times 16$ patches and pixels within the patches, which is tested in MLP-Mixer [36] to validate the invariance to permutations; (3) JP: applying masked jigsaw puzzle to the sequence patches; (4) IDX: using an additional collection of embeddings to indicate the global indexes for the input patches; (5) UNK: replacing the PEs in the masked positions with a shared unknown position embedding; (6) DAL: jointly learning dense absolute localization regression in an self-supervised manner during the pretraining, where we provide PCA, linear (LN) and nonlinear (NLN) projections, respectively. + +Table 4 shows the detailed ablation studies on the variants of our proposed MJP method. First, as expected, when we remove the PEs from the original DeiT-S model, the accuracy decreases by $2.3\%$ , but the consistency achieves $100\%$ . It + +Table 4. Ablation study on the variants of the proposed MJP. + +
MethodTop-1 Acc. ↑Consistency ↑
A: DeiT-S [37]79.864.3
B: A - PEs77.5 (-2.3)100.0
C: A + SPP [36]74.9 (-4.9)74.8
D: A + DAL (NLN)80.0 (+0.2)64.0
E: A + JP79.2 (-0.6)73.8
F: A + JP + IDX79.9 (+0.1)79.6
G: A + JP + UNK80.1 (+0.3)83.8
H: A + JP + UNK + DAL (PCA)79.9 (+0.1)83.4
I: A + JP + UNK + DAL (LN)80.0 (+0.2)83.8
J: A + JP + UNK + DAL (NLN)80.5 (+0.7)82.9
+ +![](images/8a6780d7bf30e61dbb646ffc547ec6ebca2ddd8fddce8cb0f5c51780bab1498f.jpg) +Figure 4. Distributions of accumulated eigenvalues of PEs. + +verifies that the ViTs are naturally position-insensitive once without using PEs. + +We observe that the previous SPP strategy harms the accuracy of the model (i.e., $-4.9\%$ ), which indicates it is insufficient to simply shuffle both patches and pixels in the input image. As we expect, the usage of UNK embedding alleviates the confusion between shuffled and unshuffled positions, which boosts both the accuracy and consistency. + +Finally, we can find that using dense absolute localization regression on the unmasked PEs can marginally boost the accuracy. Although PCA does not involve more parameters to the model, it significantly increases the computation latency (e.g., $+46\%$ with only 5 iterations). In general, it is better to use a nonlinear projection (e.g., a 3-layers MLP with negligible computation latency) to allow the learned PEs to aggregate more additional information. + +Informativeness of the PEs. For a position space $\mathcal{P}$ (i.e., a 1-dim or 2-dim space), we may not need a high-dimension embedding space $\mathcal{X}$ to model the positions. To measure the informativeness of the learned PEs, we apply singular value decomposition (SVD) to PEs and analyze their eigenvalue distributions. Fig. 4 plots the curves of accumulated energy/sum of top- $n$ eigenvalues versus the energy of all the eigenvalues. Supposing that we are using PCA to project the PEs of both matrices, to achieve the same explained variance ratio, our MJP needs more singular values (i.e., large dimensionality) than DeiT-S. This indicates that our positional embedding matrix is more informative. Similar observation is also revealed by Wang et al. [43]. + +Table 5. Comparisons on gradient leakage by analytic attack [29] with ImageNet-1K validation set, where we test (1) ViT-S, DeiT-S and our model in the setting (a); (2) ViT-S, DeiT-S and our model in the setting (b) (i.e., MJP with $\gamma = 0.27$ ); (3) ablation on without (w/o) using $\mathbf{E}_{\mathrm{unk}}$ in setting (a); and (4) Our model in setting (c). + +
ModelSet.Acc.↑MSE↑\( FFT_{2D} \)↑PSNR↓SSIM↓LPIPS↑
(1)ViT-S [7]78.1.0278.003919.27.5203.3623
DeiT-S [37]a79.8.0350.005718.94.5182.3767
DeiT-S (w/o PEs)77.5.0379.008220.22.5912.2692
DeiT-S+MJP80.5.1055.016611.52.4053.6545
(2)ViT-S [7]18.7.0327.001618.44.6065.2836
DeiT-S [37]b36.0.0391.002417.60.5991.3355
DeiT-S (w/o PEs)77.5.0379.002520.25.6655.2370
DeiT-S+MJP62.9.1043.005911.66.4493.6519
(3)DeiT-S+MJP (w/o)a40.6.1043.005911.66.4493.6519
(4)DeiT-S+MJPc62.9.1706.03388.07.0875.8945
+ +# 5.4. Privacy Preservation + +The fundamental principle of the gradient attack methods in federated learning is that each sample activates only a portion of content-related neurons in the deep neural networks, leading to one specific backward gradients for one related samples (i.e., 1-to-1 mapping). Based on such an observation, we argue that feeding ViTs with input patches with permuted sequences may intuitively mislead the attack. This is because now both the original and the transformed inputs may be matched to the same backward gradients (i.e., $n$ -to-1 mapping). + +To validate such an assumption, we utilize the public protocols4 to recover image with gradient updates in the privacy attack. In this privacy attack, we apply the Analytic Attack proposed in APRIL [29], which is designed for attacking the ViTs. We randomly sample 1K images from the validation set of ImageNet-1K (i.e., one image per category). To evaluate the anti-attack performance of a model, we introduce image similarity metrics to account for pixel-wise mismatch, including Mean Square Error (MSE), Peak Signal-to-Noise Ratio (PSNR), cosine similarity in the Fourier space $(\mathrm{FFT}_{2\mathrm{D}})$ , and Learned Perceptual Image Patch Similarity (LPIPS) [49]. Different from the evaluation in gradient attacks [14, 29, 48], we suppose a model is with better capacity of privacy preservation when the recovered images from its gradient updates are less similar to the ground truth images. + +Given an image $\mathbf{x}$ and its transformed (i.e., patch shuffled) version $\tilde{\mathbf{x}}$ , a ViT model $\mathcal{M}$ , and automatic evaluation metrics $\phi$ , we conduct three different settings for fair comparisons: (a) $\phi(\nabla \mathcal{M}(\mathbf{x}), \mathbf{x})$ , (b) $\phi(\nabla \mathcal{M}(\tilde{\mathbf{x}}), \tilde{\mathbf{x}})$ , and (c) $\phi(\nabla \mathcal{M}(\tilde{\mathbf{x}}), \mathbf{x})$ , where $\nabla$ refers to recovering input image through gradient attacks. Table 5 shows the quantitative comparisons between our method and the original ViTs for batch gradient inversion on ImageNet-1K. APRIL [29] enables a viable, complete recovery of original images from the gradient updates of the original ViTs. However, it per + +![](images/a8d90202ddd8e6794a1450fbfbec90fc3e2444738bc2403962663ff53e141104.jpg) +Figure 5. Visual comparisons on image recovery with gradient updates [29]. Our proposed DeiT-S+MJP model significantly outperforms the original ViT-S [7] and DeiT-S [37] models. + +Table 6. Explained variance versus PCA projected dimensionality. + +
Projected Dimension34567
DeiT-S EV (%)54.6168.5577.9585.5490.74
DeiT-S+MJP EV (%)46.7458.3669.1078.1384.55
+ +forms worse in recovering from "DeiT-S+MJP", leading to best performances on all evaluation metrics and outperform others by a large margin. + +More surprisingly, our proposed method makes APRIL yield unrecognizable images and fail in recovering the details in the original images (i.e., noisy patches in the outputs), as shown in Fig. 5. The left four columns in Fig. 5 are tested on original images, where all PEs are standard and correspond to their patch embeddings. Meanwhile, the right four columns are tested with transformed ones, where the shuffled patches are with the shared unknown PEs. Both the visual and quantitative comparisons verify that our MJP alleviates the gradient leakage problem. We also notice that DeiT-S without using PEs is inclined to be at higher risk of privacy leakage (i.e., easier to be attacked by gradients). These promising results indicate that our MJP is a promising strategy to protect user privacy in federated learning. + +# 5.5. Discussion + +PCA Projected Dimensionality. Table 6 presents the explained variance (EV) of our DeiT-S+MJP and DeiT-S versus different projection dimension. A low dimensionality can explain a large amount of information, which proves that the embedding matrix is sparse in nature. Moreover, to achieve the same explained variance ratio, our DeiT-S+MJP needs a large dimensionality than DeiT-S. This indicates that the positional embedding matrix of DeiT-S+MJP is less sparse but more informative. + +Accuracy Vs. Shuffle Ratio. Intuitively, with the increase of the shuffle ratio from the proposed MJP method, the orig- + +inal intrinsic inductive bias will be undermined. However, from the experimental results, the performance of ViTs is actually boosted via the proposed MJP. To figure out the reason behind such counter-intuitive phenomenon, we visualize the last self-attention of our proposed method in Fig. 3 of our Supp. Mat. It shows that the attention heads of our method present more diverse and content-aware attentions than original DeiT-S. We think that it might be attributed to the proposed DAL, which strengthens the spatial information of the unmasked PEs in a more efficient manner. To this end, such learned content-aware attention becomes more meaningful and results in the better accuracy. + +# 6. Conclusion + +In this paper, we first visually demonstrate that PEs can explicitly learn the 2D spatial relationship from the input patch sequences. By feeding ViTs with transformed input, we identify the issue that PEs may weaken the position-insensitive property. Based on this observation, we propose an easy-to-reproduce yet effective Masked Jigsaw Puzzle (MJP) position embedding method to alleviate the conflict in PEs (preserving the consistency versus maintaining the accuracy). Experimental results show the proposed MJP can bring the the position-insensitive property back to ViTs without degenerating the accuracy on the large-scale dataset (i.e., ImageNet-1K). In a certain sense, the proposed MJP is also a data augmentation technique, which boosts the robustness to the common corruptions (e.g., ImageNet-C) and adversarial examples (e.g., ImageNet-A/O). Surprisingly, MJP can improve the privacy preservation capacity of ViTs under typical gradient attacks by a large margin, which may pilot a new direction for privacy preservation. + +Acknowledgments. This work was partly supported by the National AI Ph.D. for Society Program of Italy, and the EU H2020 project AI4Media (No. 951911). + +# References + +[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations (ICLR), 2022. 2, 4 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 1, 3 +[3] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3 +[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning (ICML), 2020. 3 +[5] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 3 +[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 1, 3 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2020. 1, 2, 3, 7, 8 +[8] Philipp Dufter, Martin Schmitt, and Hinrich Schütze. Position information in transformers: An overview. Computational Linguistics, 48(3):733-763, 2022. 1 +[9] Yuxin Fang, Benchcheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection. Advances in Neural Information Processing Systems, 34:26183-26197, 2021. 1 +[10] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning (ICML), 2017. 3 +[11] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 1 +[12] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 244-253, 2019. 1 +[13] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on visual transformer. arXiv, 2020. 2 + +[14] Ali Hatamizadeh, Hongxu Yin, Holger Roth, Wenqi Li, Jan Kautz, Daguang Xu, and Pavlo Molchanov. Gradvit: Gradient inversion of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7 +[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[17] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[18] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 2, 6 +[19] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6 +[20] Yifan Jiang, Shiyu Chang, and Zhangyang Wang. Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[21] Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. Global-and-local relative position embedding for unsupervised video summarization. In European Conference on Computer Vision (ECCV), 2020. 3 +[22] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 54(10s):1-41, 2022. 1, 2 +[23] Shun Kiyono, Sosuke Kobayashi, Jun Suzuki, and Kentaro Inui. Shape: Shifted absolute position embedding for transformers. arXiv preprint arXiv:2109.05644, 2021. 3 +[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 3 +[25] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 1, 3, 4 +[26] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4 +[27] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. 3 + +[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 5 +[29] Jiahao Lu, Xi Sheryl Zhang, Tianli Zhao, Xiangyu He, and Jian Cheng. April: Finding the achilles' heel on privacy for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10051-10060, 2022. 1, 2, 3, 7, 8 +[30] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 2, 3 +[31] Samrudhdhi B Rangrej, Chetan L Srinidhi, and James J Clark. Consistency driven sequential transformers attention model for partially observable scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2518-2527, 2022. 2 +[32] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5 +[33] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 3 +[34] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2 +[35] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021. 3 +[36] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems (NeurIPS), 2021. 6, 7 +[37] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. 1, 3, 5, 7, 8 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 3 +[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 1 + +[40] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 1 +[41] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[43] Yu-An Wang and Yun-Nung Chen. What do position embeddings learn? an empirical study of pre-trained language model positional encoding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 1, 3, 7 +[44] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 2 +[45] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. Advances in Neural Information Processing Systems, 34:28522-28535, 2021. 2 +[46] Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[47] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1 +[48] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7 +[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7 +[50] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021. 1, 3 \ No newline at end of file diff --git a/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/images.zip b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e4c984bbab1b8aacf0449de80ce03b7c0479d934 --- /dev/null +++ b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66ff7288d2fb2cb57e10be2f65f202e1d2de7e735a59629d930ae0d12ca8b4b2 +size 552768 diff --git a/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/layout.json b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..973dad7eb4ce44644a1d7053c5a1362b9cc6a5a0 --- /dev/null +++ b/2023/Masked Jigsaw Puzzle_ A Versatile Position Embedding for Vision Transformers/layout.json @@ -0,0 +1,9729 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 49, + 103, + 544, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 103, + 544, + 120 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 544, + 120 + ], + "type": "text", + "content": "Masked Jigsaw Puzzle: A Versatile Position Embedding for Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "spans": [ + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": "Bin Ren" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": " Yahui Liu" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": " Yue Song" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": " Wei Bi" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": " Rita Cucchiara" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": " Nicu Sebe" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": " Wei Wang" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{5\\dagger}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": "University of Pisa, Italy " + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": "University of Trento, Italy \n" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": "Tencent AI Lab, China " + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": "Beijing Jiaotong University, China \n" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 72, + 141, + 525, + 200 + ], + "type": "text", + "content": "University of Modena and Reggio Emilia, Italy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 252, + 290, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 290, + 552 + ], + "type": "text", + "content": "Position Embeddings (PEs), an arguably indispensable component in Vision Transformers (ViTs), have been shown to improve the performance of ViTs on many vision tasks. However, PEs have a potentially high risk of privacy leakage since the spatial information of the input patches is exposed. This caveat naturally raises a series of interesting questions about the impact of PEs on accuracy, privacy, prediction consistency, etc. To tackle these issues, we propose a Masked Jigsaw Puzzle (MJP) position embedding method. In particular, MJP first shuffles the selected patches via our block-wise random jigsaw puzzle shuffle algorithm, and their corresponding PEs are occluded. Meanwhile, for the nonoccluded patches, the PEs remain the original ones but their spatial relation is strengthened via our dense absolute localization regressor. The experimental results reveal that 1) PEs explicitly encode the 2D spatial relationship and lead to severe privacy leakage problems under gradient inversion attack; 2) Training ViTs with the naively shuffled patches can alleviate the problem, but it harms the accuracy; 3) Under a certain shuffle ratio, the proposed MJP not only boosts the performance and robustness on large-scale datasets (i.e., ImageNet-1K and ImageNet-C, -A/O) but also improves the privacy preservation ability under typical gradient attacks by a large margin. The source code and trained models are available at https://github.com/yh1leo/MJP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 556, + 128, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 128, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 128, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 576, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 288, + 684 + ], + "type": "text", + "content": "Transformers [38] demonstrated their overwhelming power on a broad range of language tasks (e.g., text classification, machine translation, or question answering [22, 38]), and the vision community follows it closely and extends it for vision tasks, such as image classification [7,37], object detection [2, 51], segmentation [47], and image generation [3, 24]. Most of the previous ViT-based methods focus on designing different pre-training objectives [9, 11, 12] or variants of self-attention mechanisms [28, 39, 40]. By contrast, PEs" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 225, + 425, + 336 + ], + "blocks": [ + { + "bbox": [ + 308, + 225, + 425, + 336 + ], + "lines": [ + { + "bbox": [ + 308, + 225, + 425, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 225, + 425, + 336 + ], + "type": "image", + "image_path": "187412b629ee88a0d519ae7486148841bf1c2706c67e299ff596f60225b9fcdd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 339, + 548, + 417 + ], + "lines": [ + { + "bbox": [ + 304, + 339, + 548, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 548, + 417 + ], + "type": "text", + "content": "Figure 1. Low-dimensional projection of position embeddings from DeiT-S [37]. (a) The 2D UMAP projection, it shows that reverse diagonal indices have the same order as the input patch positions. (b) The 3D PCA projection, it also shows that the position information is well captured with PEs. Note that the embedding of index 1 (highlighted in red) corresponds to the [CLS] embedding that does not embed any positional information." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 224, + 547, + 336 + ], + "blocks": [ + { + "bbox": [ + 427, + 224, + 547, + 336 + ], + "lines": [ + { + "bbox": [ + 427, + 224, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 427, + 224, + 547, + 336 + ], + "type": "image", + "image_path": "74102392c2115e8f800f0639c893106513209ebac9fc0455140eaa6551e6c9e0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 424, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 664 + ], + "type": "text", + "content": "receive less attention from the research community and have not been well studied yet. In fact, apart from the attention mechanism, how to embed the position information into the self-attention mechanism is also one indispensable research topic in Transformer. It has been demonstrated that without the PEs, the pure language Transformer encoders (e.g., BERT [6] and RoBERTa [25]) may not well capture the meaning of positions [43]. As a consequence, the meaning of a sentence can not be well represented [8]. Similar phenomenon of PEs could also be observed in vision community. Dosovitskiy et al. [7] reveals that removing PEs causes performance degradation. Moreover, Lu et al. [29] analyzed this issue from the perspective of user privacy and demonstrated that the PEs place the model at severe privacy risk since it leaks the clues of reconstructing sequential patches back to images. Hence, it is very interesting and necessary to understand how the PEs affect the accuracy, privacy, and consistency in vision tasks. Here the consistency means whether the predictions of the transformed/shuffled image are consistent with the ones of the original image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "type": "text", + "content": "To study the aforementioned effects of PEs, the key is to figure out what explicitly PEs learn about positions from input patches. To answer this question, we project the high-dimensional PEs into the 2D and 3D spaces using Uni-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 246, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 246, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 246, + 703 + ], + "type": "text", + "content": "*Equal contribution. Email: {bin.ren, yahui.liu} @unitn.it" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 703, + 235, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 235, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 235, + 712 + ], + "type": "text", + "content": "† Corresponding author. Email: wei.wang@bjtu.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20382" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 170, + 185 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 170, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 170, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 170, + 185 + ], + "type": "image", + "image_path": "7b5c1df59a6fd2e20c0cfba72b50c21c339b6fcad9fd5fd3999bd08718f6f806.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 189, + 123, + 198 + ], + "lines": [ + { + "bbox": [ + 100, + 189, + 123, + 198 + ], + "spans": [ + { + "bbox": [ + 100, + 189, + 123, + 198 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 100, + 189, + 123, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 70, + 288, + 185 + ], + "blocks": [ + { + "bbox": [ + 172, + 70, + 288, + 185 + ], + "lines": [ + { + "bbox": [ + 172, + 70, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 172, + 70, + 288, + 185 + ], + "type": "image", + "image_path": "e8cf282277e3393e50d87cd69d69998ac94c48fb332535d84a129b64afc6d4f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 188, + 257, + 198 + ], + "lines": [ + { + "bbox": [ + 208, + 188, + 257, + 198 + ], + "spans": [ + { + "bbox": [ + 208, + 188, + 257, + 198 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 208, + 188, + 257, + 198 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}} (\\gamma = 1)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 291, + 70, + 408, + 185 + ], + "blocks": [ + { + "bbox": [ + 291, + 70, + 408, + 185 + ], + "lines": [ + { + "bbox": [ + 291, + 70, + 408, + 185 + ], + "spans": [ + { + "bbox": [ + 291, + 70, + 408, + 185 + ], + "type": "image", + "image_path": "4b91c42dd0362a306dfa83289a418a28a3c731bd986ca25fa9ee599beea44186.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 188, + 378, + 198 + ], + "lines": [ + { + "bbox": [ + 324, + 188, + 378, + 198 + ], + "spans": [ + { + "bbox": [ + 324, + 188, + 378, + 198 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 324, + 188, + 378, + 198 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}} (\\gamma = 0.4)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 201, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 235 + ], + "type": "text", + "content": "Figure 2. (a) The original input patches; (b) Totally random shuffled input patches; (c) Partially random shuffled input patches; (d) An overview of the proposed MJP. Note that we show the random shuffled patches and its corresponding unknown position embedding in green and the rest part in blue. DAL means the self-supervised dense absolute localization regression constraint." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 70, + 536, + 185 + ], + "blocks": [ + { + "bbox": [ + 408, + 70, + 536, + 185 + ], + "lines": [ + { + "bbox": [ + 408, + 70, + 536, + 185 + ], + "spans": [ + { + "bbox": [ + 408, + 70, + 536, + 185 + ], + "type": "image", + "image_path": "513b953e5d3546cd6099861512244379cb0a5ac27e1481a30b0f4db305f0da6a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 188, + 506, + 198 + ], + "lines": [ + { + "bbox": [ + 429, + 188, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 429, + 188, + 506, + 198 + ], + "type": "text", + "content": "(d) Overview of our MJP" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 244, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 289, + 411 + ], + "type": "text", + "content": "form Manifold Approximation & Projection (UMAP) [30] and PCA, respectively. Then for the first time, we visually demonstrate that the PEs can learn the 2D spatial relationship very well from the input image patches (the relation is visualized in Fig. 1). We can see that the PEs are distributed in the same order as the input patch positions. Therefore, we can easily obtain the actual spatial position of the input patches by analyzing the PEs. Now it explains why PEs can bring the performance gain for ViTs [7]. This is because the spatial relation in ViTs works similar as the inherent intrinsic inductive bias in CNNs (i.e., it models the local visual structure) [45]. However, these correctly learned spatial relations are unfortunately the exact key factor resulting in the privacy leakage [29]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 413, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 288, + 568 + ], + "type": "text", + "content": "Based on these observations, one straightforward idea to protect the user privacy is to provide ViTs with the randomly transformed (i.e., shuffled) input data. The underlying intuition is that the original correct spatial relation within input patches will be violated via such a transformation. Therefore, we transform the previous visually recognizable input image patches " + }, + { + "bbox": [ + 46, + 413, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 413, + 288, + 568 + ], + "type": "text", + "content": " shown in Fig. 2(a) to its unrecognizable counterpart " + }, + { + "bbox": [ + 46, + 413, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 413, + 288, + 568 + ], + "type": "text", + "content": " depicted in Fig. 2(b) during training. The experimental results show that such a strategy can effectively alleviate the privacy leakage problem. This is reasonable since the reconstruction of the original input data during the attack is misled by the incorrect spatial relation. However, the side-effect is that this leads to a severe accuracy drop." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": "Meanwhile, we noticed that such a naive transformation strategy actually boosts the consistency [31, 34, 44] albeit the accuracy drops. Note that here the consistency represents if the predictions of the original and transformed (i.e., shuffled) images are consistent. Given the original input patches " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": " and its corresponding transformed (i.e., shuffled) counterpart, we say that the predictions are consistent if " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\arg \\max P(\\mathcal{F}(\\mathbf{x})) = \\arg \\max P(\\mathcal{F}(\\widetilde{\\mathbf{x}}))" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": " refers to the ViT models, and " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": " denotes the predicted logits." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 714 + ], + "type": "text", + "content": "These observations hint that there might be a trade-off solution that makes ViTs take the best from both worlds (i.e., both the accuracy and the consistency). Hence, we pro" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 244, + 547, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 412 + ], + "type": "text", + "content": "pose the Masked Jigsaw Puzzle (MJP) position embedding method. Specifically, there are four core procedures in the MVP: (1) We first utilize a block-wise masking method [1] to randomly select a partial of the input sequential patches; (2) Next, we apply jigsaw puzzle to the selected patches (i.e., shuffle the orders); (3) After that, we use a shared unknown position embedding for the shuffled patches instead of using their original PEs; (4) To well maintain the position prior of the unshuffled patches, we introduce a dense absolute localization (DAL) regressor to strengthen their spatial relationship in a self-supervised manner. We simply demonstrate the idea of the first two procedures in Fig. 2(c), and an overview of the proposed MVP method is available in Fig. 2(d). In summary, our main contributions are:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 415, + 547, + 579 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 306, + 415, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 547, + 464 + ], + "type": "text", + "content": "- We demonstrate that although PEs can boost the accuracy, the consistency against image patch shuffling is harmed. Therefore, we argue that studying PEs is a valuable research topic for the community." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 468, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 547, + 504 + ], + "type": "text", + "content": "- We propose a simple yet efficient Masked Jigsaw Puzzle (MJP) position embedding method which is able to find a balance among accuracy, privacy, and consistency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 506, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 547, + 579 + ], + "type": "text", + "content": "- Extensive experimental results show that MJP boosts the accuracy on regular large-scale datasets (e.g., ImageNet-1K [32]) and the robustness largely on ImageNet-C [18], -A/O [19]. One additional bonus of MJP is that it can improve the privacy preservation ability under typical gradient attacks by a large margin." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 590, + 392, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 392, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 392, + 604 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 611, + 425, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 425, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 425, + 623 + ], + "type": "text", + "content": "2.1. Vision Transformers" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Transformers [38], originally designed for Nature Language Processing (NLP) tasks, have recently shown promising performance on computer vision tasks [13, 22]. Benefiting from the strong representation power of modelling global relations between image patches, Vision Transformers (ViTs) [7] have achieved superior performance than their counterpart CNNs on image classification and various other" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20383" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "content": "downstream tasks (e.g., object detection [2, 51], object re-identification [17], dense prediction [41, 42, 46, 50] and image generation [3, 4, 20, 24])." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "content": "As a core module in ViTs, multi-head self-attention (MSAs) [7, 38] aggregate sequential tokens with normalized attentions as: " + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "inline_equation", + "content": "z_{j} = \\sum_{i}\\mathrm{Softmax}(\\frac{QK}{\\sqrt{d}})_iV_{i,j}" + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "content": " are query, key and value matrices, respectively. " + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "content": " is the dimension of query and key, and " + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 109, + 290, + 339 + ], + "type": "text", + "content": "-th output token. In theory, when the position information is not considered, the outputs of MSAs should be strictly invariant to the input sequence order (i.e., position-insensitivity). This indicates that a visually recognizable image can be transformed into its unrecognizable counterpart by permuting the order of image patches while maintaining the performance delivered by ViTs compared with the ViTs trained on the original non-permuted image. However, the usage of PEs hinders such implementations, where the outputs of the ViTs vary dramatically with the mentioned naive transformations shown in Fig. 2(b). In this work, our main focus is to explicitly figure out what PEs actually learn from input patches about positions, and how the PEs affect the accuracy, privacy, and the consistency properties of ViTs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 347, + 168, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 168, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 168, + 360 + ], + "type": "text", + "content": "2.2. Position Embeddings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 288, + 617 + ], + "type": "text", + "content": "In Transformer networks, both the attention and the (individual token based) feed-forward layers are permutation invariant when the position information is not considered. In this way, the spatial relationships between image patches could not be modeled as the position information is completely discarded. As compensation, PEs are naturally introduced into ViTs to provide information about the token order during the learning process, since it offers possibilities for dependency modeling between elements at different positions. For example, previous works [10,33,38] indicated that PEs are useful to give the model a sense of which portion of the sequence in the input/output it is currently dealing with. Inspired by this, some works [15,21,23,27,35] showed diverse application scenarios that benefit from the usage of suitable PEs. In addition, Chu et al. [5] proposed correlating the PEs with their local neighborhood of the input sequence. Liu et al. [26] proposed to enhance the spatial prior (i.e., relative localization) in the final content embedding to indirectly enrich the inductive bias. Obviously, although these methods enhance the position information learnt by PEs, they indeed degenerate the position-insensitive property of MSAs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Especially, Wang et al. [43] revealed that Transformer encoders (e.g., BERT [6] and RoBERTa [25]) may not well capture the meaning of positions (absolute and relative positions). They showed that Transformer encoders learn the local position information that can only be effective in masked language modeling. In contrast, there does not exist such a similar \"masked language modeling\" procedure in ViTs (and VTs) in the typically supervised pre-training. Moreover, Lu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "et al. [29] revealed that the learnable PEs place the model at severe privacy risk, which leaks the clues of reconstructing sequential patches to images. In this paper, we dive into the usage of PEs and propose an efficient position embedding method, MJP, to improve the position-insensitive property of ViTs without hurting the positive effects of PEs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 156, + 541, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 156, + 541, + 169 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 541, + 169 + ], + "type": "text", + "content": "3. Preliminaries: 2D/3D Spatial Priors in PEs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 176, + 547, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 547, + 463 + ], + "type": "text", + "content": "Although numerous works claim that the PEs can learn the 2-D spatial relationship of image patches, this claim has not been demonstrated visually or mathematically. To visualize the concrete relation of image patches captured in the high-dimensional position embedding, we project them into the 2-D and 3-D spaces using Uniform Manifold Approximation and Projection (UMAP) [30] and PCA, respectively. Fig. 1 displays the projections of PEs from DeiT-S [37]. Note that [7] only shows the cosine similarity between the PEs, which is not exactly the 2D spatial information mentioned in our paper. Here, we explicitly specify the spatial relationship as the relative distance in the 2D/3D coordinate space. As shown in Fig. 1, the spatial relationship can be projected into a 2D/3D coordinate system and indicate the localizations of these embeddings. For the UMAP, the projected positional embedding emerges as grid-like structures with the distances between each point roughly the same, which is coherent with the relation of embedded image patches. For the PCA, the position embedding that corresponds to neighboring image patches groups together and aggregates to the stick-like form. This phenomenon demonstrates that the spatial relationship is indeed learned by the PEs. Moreover, this also implies that the spatial relationship learned in the high-dimensional space still manifests in the low-dimensional space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 464, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 559 + ], + "type": "text", + "content": "The learnable PEs, which work as a lookup table of a dictionary, maps the 1-dimensional data into the sparse high-dimensional space. By nature, it is a sparse large matrix. Dimensionality reduction techniques can capture the structural information of sparse matrices using low dimensionalities. For the PCA, the amount of information retained in the projection can be measured by the ratio of explained variance. Formally, we have the definition as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "text", + "content": "Definition 1 (Explained Variance) Let " + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "text", + "content": " denote the data matrix and the PCA projection matrix. The ratio of explained variance that " + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{PV}" + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "text", + "content": " accounts for is defined as " + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\sum \\sigma (\\mathbf{PV})^2 /\\sum \\sigma (\\mathbf{P})^2" + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 305, + 564, + 547, + 613 + ], + "type": "text", + "content": " denotes the singular value." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "In practice, we observe that the 3-dimensional PCA projection explains " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "54.6\\%" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " of the total variance of the DeiT-S embedding matrix. Given that a large amount of information can be captured by the low-dimensional projection, we propose that explicitly enforcing low-dimensional positional prior can help the positional learning in the high-dimensional space, which might accelerate the convergence rate of training and improve the performance (See Sec. 4.2)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20384" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 87, + 290, + 204 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 275, + 83 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 275, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 275, + 83 + ], + "type": "text", + "content": "Algorithm 1 Block-wise Random Jigsaw Puzzle Shuffle" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "lines": [ + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": "Input: Input image: " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " Shuffle Ratio: " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " Patch Size: " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " \nOutput: Shuffled image patches: " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{x}}_p" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p\\in \\mathbb{R}^{N\\times (P^2\\cdot C)}\\gets" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " Patchlize(x,P) \n2: " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{m}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}\\gets" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " BinaryInitialize(xp,0) \n3: " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{m}}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}\\gets" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " BlockwiseMask(m,γ) [25] \n4: " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{x}}_p\\in \\mathbb{R}^{N\\times (P^2\\cdot C)}\\gets" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " Jigsaw Puzzle(xp, " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{m}}" + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "text", + "content": " 5: return " + }, + { + "bbox": [ + 47, + 87, + 290, + 204 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{x}}_p" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 224, + 102, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 102, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 102, + 236 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "content": "Based on the design principle of MSAs, the outputs of MSAs should be entirely position-agnostic. However, PEs hinder such a property because it can learn the strong 2-D spatial relation of the input patches (as we demonstrated in Sec. 3). To this end, we propose the block-wise random jigsaw puzzle shuffle algorithm (See Alg. 1) to transform the input patches with different shuffle ratios " + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "content": " for intermingling the original correct spatial relation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 340, + 288, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 471 + ], + "type": "text", + "content": "Since we experimentally demonstrate that the totally shuffled strategy (i.e., " + }, + { + "bbox": [ + 46, + 340, + 288, + 471 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 46, + 340, + 288, + 471 + ], + "type": "text", + "content": ") will degenerate the accuracy a lot albeit the consistency increases. As a remedy, we only shuffle portion of the sequence patches, and strengthen the spatial relation of the rest part with a dense absolute localization regression. Finally, a versatile position embedding method MJP is proposed. The detailed analysis of each module is available in the ablation study in Sec. 5.3. In the following sub-sections, we will introduce the Jigsaw Puzzle Shuffle method (Sec. 4.1), the spatial relation strengthen method (Sec. 4.2), and our final MJP method (Sec. 4.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 478, + 266, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 266, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 266, + 491 + ], + "type": "text", + "content": "4.1. Block-wise Random Jigsaw Puzzle Shuffle" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": "Specifically, given an input image " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": ", we first reshape it into a sequence of flattened 2D patches " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot C)}" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "(H, W)" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": " is the resolution of the original image, " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": " is the number of channels, " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "(P, P)" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": " is the resolution of each image patch, and then we have " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "N = HW / P^2" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": ", which denotes the number of patches. Then instead of directly applying block-wise masking method [1] to the image, we first initialize a binary mask matrix " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P}}" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": " with the same size as the image patches in " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": ". Next, we use [1] to update the binary mask " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": " in which the masked positions will be set to 1 and the rest untouched positions remains 0. The hyper-parameter " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": " is used to control the ratio of selected positions. After that, a jigsaw puzzle shuffle operation is applied to " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": " conditioned on the updated binary mask " + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{m}}" + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "type": "text", + "content": "Finally, we get the shuffled patch sequence " + }, + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}}_p" + }, + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "type": "text", + "content": ", where the shuffled patches are actually " + }, + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_p^i |\\widetilde{\\mathbf{m}}_i = 1\\}" + }, + { + "bbox": [ + 46, + 666, + 288, + 713 + ], + "type": "text", + "content": ". Notably, in [1], the patches covered by the sampled mask are not visible to the encoder module, while in our method, the masking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "strategy [1] is only used to mask out the positions/indices of the selected patches. These patches are still visible to the encoder module and they are randomly shuffled in a jigsaw puzzle manner. One intuitive shuffled toy example is shown in Fig. 2(c), where the green part means the selected shuffled region while the blue part remains unchanged. Actually, this algorithm shares a similar idea with the data augmentation method as the training images are always changing as their patches are randomly shuffled during each iteration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 186, + 491, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 491, + 199 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 491, + 199 + ], + "type": "text", + "content": "4.2. Strengthening Spatial Prior in PEs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 204, + 546, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 546, + 335 + ], + "type": "text", + "content": "Liu et al. [26] noticed that by enhancing the 2-D spatial information of the output embeddings of the last layer of ViTs, the training convergence speed can be accelerated. Inspired by their work, we observe that similar gains can be obtained by applying a low-dimensional spatial prior in PEs. Different from the dense relative localization constraint in [26] which samples relative pairs from the whole output sequential embeddings, we propose a much simpler dense absolute localization (DAL) regression method to directly use self-supervised absolute location to enhance the spatial information in PEs, which avoids sampling relative pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "content": "Since the PEs capture the absolute position of the input patches, to some extent, the position information could be reconstructed via a reversed mapping function " + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "inline_equation", + "content": "g(\\cdot):\\mathcal{E}\\to \\mathcal{P}" + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "content": " are embedding space and position space, respectively. Given that PEs have one-to-one correspondence with the sequential image patches, we can reshape them into " + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pos}}\\in \\mathbb{R}^{K\\times K\\times D}" + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "content": " refers to height/weight of the grid and " + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "content": " refers to latent vector size. Then we can compute the reverse mapping from " + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\rightarrow \\mathcal{P}" + }, + { + "bbox": [ + 304, + 336, + 546, + 443 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 448, + 545, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 448, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 388, + 448, + 545, + 464 + ], + "type": "interline_equation", + "content": "\\left(\\widetilde {i}, \\widetilde {j}\\right) ^ {T} = g \\left(\\mathbf {E} _ {\\text {p o s}} ^ {i, j}\\right), \\tag {1}", + "image_path": "c321cc4887938dbdd839c97242f09275a5edbac5fc73c5d6f40c8209436aa088.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "inline_equation", + "content": "(\\widetilde{i},\\widetilde{j})^T" + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "content": " is the predicted patch position, and " + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pos}}^{i,j}" + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "content": " is the position embedding of the patch " + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "inline_equation", + "content": "K\\times K" + }, + { + "bbox": [ + 304, + 470, + 547, + 506 + ], + "type": "text", + "content": " grid. The dense absolute localization (DAL) loss is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 334, + 511, + 545, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 511, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 334, + 511, + 545, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D A L}} = \\mathbb {E} _ {\\mathbf {E} _ {\\text {p o s}}, 1 \\leq i, j \\leq K} ^ {i, j} [ \\| (i, j) ^ {T} - (\\widetilde {i}, \\widetilde {j}) ^ {T} \\| _ {1} ], \\tag {2}", + "image_path": "20f1d0a7f209ede7af6dc6f2f8607be90252cd6f438bc90559ea066f21fa918e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": "where the expectation is computed by averaging the " + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": " loss between the correspond " + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "inline_equation", + "content": "(i,j)^T" + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "inline_equation", + "content": "(\\widetilde{i},\\widetilde{j})^T" + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DAL}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": " is added to the standard cross-entropy loss " + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{CE}})" + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": " of the native ViTs. The final loss is: " + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{all}} = \\mathcal{L}_{\\mathrm{CE}} + \\lambda \\mathcal{L}_{\\mathrm{DAL}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": ", where we set " + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": " for all experiments. Note that the mapping function can be either linear or nonlinear. Throughout this work, we mainly discuss three implementations, including non-parametric PCA, learnable linear (LN), and nonlinear (NLN) projection layers. They will be discussed in details in Sec. 5.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 659, + 449, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 449, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 449, + 672 + ], + "type": "text", + "content": "4.3. MJP Position Embedding" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "The main goal of MJP position embedding in our work is to enhance the consistency (i.e., position-insensitive property) of ViTs and preserve the accuracy (keeping the spatial" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20385" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 87, + 253, + 146 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 241, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 241, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 241, + 84 + ], + "type": "text", + "content": "Algorithm 2 The pipeline of the proposed MJP." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "lines": [ + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "spans": [ + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "text", + "content": "Input: Input image: " + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "text", + "content": " Shuffle Ratio: " + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "text", + "content": " ; Patch Size: " + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "text", + "content": " 1: " + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{x}}_p\\gets Alg.1(\\mathbf{x},P,\\gamma)" + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "text", + "content": " // 1st & 2nd procedures 2: " + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{unk}}(\\widetilde{\\pmb{x}}_p)" + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "text", + "content": " // 3rd procedure 3: " + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "inline_equation", + "content": "\\mathbf{DAL}(\\mathbf{x} - \\mathbf{x}\\cap \\widetilde{\\pmb{x}}_p)" + }, + { + "bbox": [ + 47, + 87, + 253, + 146 + ], + "type": "text", + "content": " // 4th procedure, only for train" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 167, + 288, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 214 + ], + "type": "text", + "content": "relation modeling functionality of PEs) for the standard classification tasks. Usually, with original input image patches " + }, + { + "bbox": [ + 46, + 167, + 288, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 167, + 288, + 214 + ], + "type": "text", + "content": " we can formulate the function of the input layer which is located before the Transformer block as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 223, + 287, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 223, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 79, + 223, + 287, + 239 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {C L S}}; \\mathbf {x} _ {p} ^ {1} \\mathbf {E}; \\mathbf {x} _ {p} ^ {2} \\mathbf {E}, \\dots , \\mathbf {x} _ {p} ^ {N} \\mathbf {E} \\right] + \\mathbf {E} _ {\\text {p o s}}, \\tag {3}", + "image_path": "1564e90185be90604e50339a4ab2d08cc2d89e6850b08f16b1b10bb374af0fcd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{(P^2 \\cdot C) \\times D}" + }, + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "type": "text", + "content": " is a trainable linear projection layer and " + }, + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p^k\\mathbf{E}" + }, + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "type": "text", + "content": " refers to the output of linear projection (i.e., the patch embeddings), and " + }, + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pos}} \\in \\mathbb{R}^{(N + 1) \\times D}" + }, + { + "bbox": [ + 46, + 247, + 287, + 298 + ], + "type": "text", + "content": " refers to PEs (the additional one is applied to the [CLS] embedding)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": "Next, we apply the proposed block-wise random jigsaw puzzle shuffle algorithm to " + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p" + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": " and produce the transformed patch sequences " + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}}_p" + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": " with the shuffle ratio " + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": ". In this scenario, if we maintain the original position embedding sequence, it leads to a mismatch issue between the shuffled patches and the position embedding sequence. Therefore, we introduce a shared unknown position embedding to the shuffled positions to alleviate the mismatching issue. With the corresponding updated mask " + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{m}}" + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": " from Alg. 1 (1st & 2nd procedures), we propose our MJP PEs:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 426, + 287, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 426, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 287, + 458 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathbf {E}} _ {\\text {p o s}} ^ {i} = \\left\\{ \\begin{array}{l l} \\mathbf {E} _ {\\text {p o s}} ^ {i}, & \\text {i f} \\widetilde {\\mathbf {m}} _ {i} = 0 \\\\ \\mathbf {E} _ {\\text {u n k}}, & \\text {i f} \\widetilde {\\mathbf {m}} _ {i} = 1 \\end{array} \\right. \\tag {4}", + "image_path": "02efbb20bb67a3d494c6570a55311b4a01313e274b70d28a7d6075ef17fdc36e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 466, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 466, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{unk}} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 46, + 466, + 287, + 527 + ], + "type": "text", + "content": " refers to a share learnable embedding (i.e., unknown position embedding). It represents that the image patch in this position has random permutation and its position should be occluded (3rd procedure). " + }, + { + "bbox": [ + 46, + 466, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 527 + ], + "type": "text", + "content": " is the original position embedding for the rest image patches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 527, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 646 + ], + "type": "text", + "content": "Besides, we revisit the remaining PEs (i.e., corresponding to the non-selected patches) and apply low-dimensional prior on them. The low-dimensional prior is imposed by the proposed DAL (Sec. 4.2) regression method for strengthening the spatial relation (4th procedure). A toy illustration of the proposed MJP is shown in Fig. 2(d), where the green color represents the randomly permuted patches and its corresponding unknown PEs, while the blue color indicates the rest regular patches and its related original PEs. We also formalize these procedures as an algorithm in Alg. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 647, + 284, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 647, + 284, + 658 + ], + "spans": [ + { + "bbox": [ + 59, + 647, + 284, + 658 + ], + "type": "text", + "content": "Thus, we replace the input layer with a new formulation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 79, + 666, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 666, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 79, + 666, + 287, + 682 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathbf {z}} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {C L S}}; \\widetilde {\\mathbf {x}} _ {p} ^ {1} \\mathbf {E}; \\widetilde {\\mathbf {x}} _ {p} ^ {2} \\mathbf {E}, \\dots , \\widetilde {\\mathbf {x}} _ {p} ^ {N} \\mathbf {E} \\right] + \\widetilde {\\mathbf {E}} _ {\\text {p o s}}. \\tag {5}", + "image_path": "1277f3b2c66f683d16bd0e2c3f25ef87208c7904d8b743c1a06461d33be6512e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "the following procedures and modules are exactly the same as the ones in the original ViTs." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 94, + 542, + 171 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "type": "text", + "content": "Table 1. Comparisons of different backbones on ImageNet-1K classification. Note that the image size here are all set to " + }, + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 94, + 542, + 171 + ], + "lines": [ + { + "bbox": [ + 310, + 94, + 542, + 171 + ], + "spans": [ + { + "bbox": [ + 310, + 94, + 542, + 171 + ], + "type": "table", + "html": "
MethodParam. Top-1 Acc.↑ Diff.Norm.↓ Consistency
ResNet-50 [16]2579.311.7751.5
ResNet-50 + MJP2579.47.1169.3
DeiT-S [37]2279.816.2164.3
DeiT-S + MJP2280.58.9682.9
Swin-T [28]2981.315.4941.5
Swin-T + MJP2981.312.3666.9
", + "image_path": "1f9a78a72085550f74c5434903573a5ab39bafabe1653d58f015618b0eaadb57.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 181, + 386, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 181, + 386, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 181, + 386, + 194 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 201, + 547, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 285 + ], + "type": "text", + "content": "We follow the typical supervised pre-training procedure, where all the compared models are trained on ImageNet1K [32] to show the capacity of our proposed MJP method. We also benchmark the proposed MJP method on ImageNet1K, which contains 1.28M training images and 50K validation images of 1,000 classes. The training details mostly follow the training protocols from Touvron et al. [37]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 293, + 473, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 473, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 473, + 306 + ], + "type": "text", + "content": "5.1. Regular ImageNet-1K training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "text", + "content": "We mainly compare with three typical existing methods, including two state-of-the-art Visual Transformers (i.e., DeiT [37] and Swin [28]) and one widely-used CNN-based ResNet-50 [16]. All these methods are of comparable sizes (i.e., less than 30M parameters). Besides the common Top-1 accuracy (Top-1 Acc.), we also report another two evaluation metrics (i.e. Diff. Norm. and Consistency) to show the position invariance of a model to the jigsaw puzzle transformation. For a given input image " + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "text", + "content": ", we collect its counterpart " + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "text", + "content": " by applying masked jigsaw puzzle to shuffle a portion of selected patches. Next, we calculate the difference " + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "text", + "content": "-norm (i.e., Diff. Norm.) between the [CLS] embedding inferred from " + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 312, + 547, + 467 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 375, + 471, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 471, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 375, + 471, + 545, + 485 + ], + "type": "interline_equation", + "content": "\\left\\| \\mathcal {F} _ {\\mathrm {C L S}} (\\mathbf {x}) - \\mathcal {F} _ {\\mathrm {C L S}} (\\widetilde {\\mathbf {x}}) \\right\\| _ {2} ^ {2}. \\tag {6}", + "image_path": "48baa4c9c9be4560395c80baa0a03d29534669b221b5a70e900e0bbbb3621f58.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 495, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 519 + ], + "type": "text", + "content": "For Consistency, we measure how many classification results stay the same (\\%) given the perturbation:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 529, + 545, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 529, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 320, + 529, + 545, + 543 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathbf {x}} \\left[ \\mathbb {I} \\left\\{\\arg \\max P (\\mathcal {F} (\\mathbf {x})) = \\arg \\max P (\\mathcal {F} (\\widetilde {\\mathbf {x}})) \\right\\} \\right]. \\tag {7}", + "image_path": "69076062b8ef718e35420f3e5d006db94177d3cb89586123fb453c1c74a4b332.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "inline_equation", + "content": "P(\\mathcal{F}(\\mathbf{x}))" + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "text", + "content": " denotes the predicted classification probability of image " + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "inline_equation", + "content": "\\arg \\max P(\\mathcal{F}(\\mathbf{x}))" + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "text", + "content": " represents the predicted class index. For a fair comparison, we use " + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "inline_equation", + "content": "\\gamma = 0.15" + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "text", + "content": " to create " + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 551, + 547, + 694 + ], + "type": "text", + "content": " in Table 1. According to these results, the proposed MJP does not have a negative effect on the top-1 accuracy (MJP even brings a marginal improvement for DeiT-S). More importantly, it improves Diff. Norm. and Consistency by a large margin. Besides, MJP works well in the variants of ViTs (e.g., Swin [28]), which shows a good potential generalization ability. Note that we add MJP to ResNet50 by only shuffling the image patches with Alg. 1 (No PEs). For Swin-T, we add MJP to the absolute PEs of Swin-T." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 514, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 514, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 514, + 712 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20386" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 60, + 194, + 173 + ], + "blocks": [ + { + "bbox": [ + 56, + 60, + 194, + 173 + ], + "lines": [ + { + "bbox": [ + 56, + 60, + 194, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 60, + 194, + 173 + ], + "type": "image", + "image_path": "edfe94efef2ab9d518d4ef5f73138de4ef5ac5f4e16cf4b10710f6512243d02e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 177, + 129, + 186 + ], + "lines": [ + { + "bbox": [ + 118, + 177, + 129, + 186 + ], + "spans": [ + { + "bbox": [ + 118, + 177, + 129, + 186 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 60, + 364, + 173 + ], + "blocks": [ + { + "bbox": [ + 227, + 60, + 364, + 173 + ], + "lines": [ + { + "bbox": [ + 227, + 60, + 364, + 173 + ], + "spans": [ + { + "bbox": [ + 227, + 60, + 364, + 173 + ], + "type": "image", + "image_path": "94b9817bd27d7a2a49267dad77b19b2a1a115dc3340f7a9e6a0c810ac77ec96f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 177, + 301, + 186 + ], + "lines": [ + { + "bbox": [ + 290, + 177, + 301, + 186 + ], + "spans": [ + { + "bbox": [ + 290, + 177, + 301, + 186 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "text", + "content": "Figure 3. Ablation on the mask ratio " + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "text", + "content": " during inference: (a) comparisons among ViT-S, DeiT-S and our method (trained with " + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\gamma = 0.03" + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "text", + "content": "); (b) comparisons between Swin-T and our method (trained with " + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\gamma = 0.03" + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "text", + "content": "); (c) comparisons of our method on DeiT-S trained with different " + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 400, + 60, + 536, + 173 + ], + "blocks": [ + { + "bbox": [ + 400, + 60, + 536, + 173 + ], + "lines": [ + { + "bbox": [ + 400, + 60, + 536, + 173 + ], + "spans": [ + { + "bbox": [ + 400, + 60, + 536, + 173 + ], + "type": "image", + "image_path": "4d99d083d74bafd54874aea96b7c1df4fab1377996c00b8e9d4dd94c1508b8d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 177, + 473, + 186 + ], + "lines": [ + { + "bbox": [ + 462, + 177, + 473, + 186 + ], + "spans": [ + { + "bbox": [ + 462, + 177, + 473, + 186 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 51, + 236, + 284, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 214, + 288, + 236 + ], + "lines": [ + { + "bbox": [ + 47, + 214, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 288, + 236 + ], + "type": "text", + "content": "Table 2. Comparisons on robustness to common corruptions and adversarial examples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 236, + 284, + 291 + ], + "lines": [ + { + "bbox": [ + 51, + 236, + 284, + 291 + ], + "spans": [ + { + "bbox": [ + 51, + 236, + 284, + 291 + ], + "type": "table", + "html": "
MethodImageNet-CImageNet-AImageNet-O
mCE ↓Acc ↑AURRA ↑AUPR ↑
DeiT-S54.619.225.120.9
DeiT-S + MJP40.7821.629.822.6
", + "image_path": "381c2bc5934fd57ca90bbba014f806805817aa8573cf9f29703dc705ce7647f5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 301, + 218, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 218, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 218, + 315 + ], + "type": "text", + "content": "5.2. Robustness on Challenging Sets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 321, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 392 + ], + "type": "text", + "content": "Besides the strength on standard classification, we also observe an auxiliary benefit of the proposed MJP on robustness. ImageNet-C [18] benchmarks a classifier's robustness to common corruptions2. The mean Corruption Error (mCE) is used to measure the generalization of a model at corrupted distributions (the lower the better)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 393, + 288, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 288, + 501 + ], + "type": "text", + "content": "ImageNet-A/O [19] focus on adversarial examples, and it enables us to either test image classification performance when the input data distribution shifts (i.e., Acc and AU-RRA), or test out-of-distribution detection performance when the label distribution shifts (i.e., AUPR)" + }, + { + "bbox": [ + 46, + 393, + 288, + 501 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 393, + 288, + 501 + ], + "type": "text", + "content": ". As shown in Table 2, the proposed MJP not only boosts the accuracy on the standard evaluation on ImageNet-1K validation set but also improves the robustness largely on the adversarial samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 501, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 562 + ], + "type": "text", + "content": "The underlying reason might be that MJP enforces the ViTs aware of both local and global context features, and it helps ViTs to get rid of some unnecessary sample-specific local features during the training. This has been verified by the visualization maps (see Fig. 3 of our Supp. Mat.)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 571, + 153, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 153, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 153, + 583 + ], + "type": "text", + "content": "5.3. Ablation Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": "Results with different MJP ratios. As shown in Table 3, we test different masking ratios used in the block-wise masking strategy during the training. Obviously, Diff. Norm. has the inverse tendencies compared to Consistency, where a smaller Diff. Norm. usually indicates a larger (better) consistency score. Comparing the accuracy trained with " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\gamma >0" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": ", it shows our model is not sensitive to different " + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 590, + 287, + 674 + ], + "type": "text", + "content": ". In particular, a" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 309, + 238, + 541, + 312 + ], + "blocks": [ + { + "bbox": [ + 306, + 214, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 306, + 214, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 545, + 236 + ], + "type": "text", + "content": "Table 3. Ablation study on the proposed MJP method trained with different masking ratio " + }, + { + "bbox": [ + 306, + 214, + 545, + 236 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 214, + 545, + 236 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 238, + 541, + 312 + ], + "lines": [ + { + "bbox": [ + 309, + 238, + 541, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 238, + 541, + 312 + ], + "type": "table", + "html": "
MetricMasking Ratio
00.030.090.150.210.27
Top-1 Acc.80.080.580.380.480.280.3
Diff. Norm.16.568.966.365.234.393.97
Consistency64.082.988.190.592.393.1
", + "image_path": "f3939951ac98fb651168cd5ba0652c0123391cca2957d0cbe19046a8d2da5531.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": "small ratio (e.g., " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "\\gamma = 0.03" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": ") is sufficient for boosting the accuracy. In addition, a large ratio can reduce the Diff. Norm. and improves the consistency by a large margin. Moreover, the model trained with a larger " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": " is inclined to be more consistent as shown in Fig. 3 (c). In addition, we observe that the consistency keeps increasing when the mask ratio increases from " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": ". However, further increasing the mask ratio will not bring consistency improvement and what is worse is that the accuracy marginally decreases." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 435, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 506 + ], + "type": "text", + "content": "For a model trained with a fixed " + }, + { + "bbox": [ + 304, + 435, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 435, + 547, + 506 + ], + "type": "text", + "content": ", we also test its accuracy with different masking ratios during the inference. As shown in Fig. 3 (a) and (b), the performances of original models drop significantly when we shuffle more image patches. In contrast, our proposed method shows more consistent performances." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 508, + 548, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 548, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 548, + 664 + ], + "type": "text", + "content": "Comparison of variants of MJP. We test several variants of the proposed MJP, including (1) Removing the PEs from the original DeiT-S; (2) SPP: shuffling both the " + }, + { + "bbox": [ + 304, + 508, + 548, + 664 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 508, + 548, + 664 + ], + "type": "text", + "content": " patches and pixels within the patches, which is tested in MLP-Mixer [36] to validate the invariance to permutations; (3) JP: applying masked jigsaw puzzle to the sequence patches; (4) IDX: using an additional collection of embeddings to indicate the global indexes for the input patches; (5) UNK: replacing the PEs in the masked positions with a shared unknown position embedding; (6) DAL: jointly learning dense absolute localization regression in an self-supervised manner during the pretraining, where we provide PCA, linear (LN) and nonlinear (NLN) projections, respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Table 4 shows the detailed ablation studies on the variants of our proposed MJP method. First, as expected, when we remove the PEs from the original DeiT-S model, the accuracy decreases by " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ", but the consistency achieves " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ". It" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 56, + 682, + 253, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 682, + 253, + 693 + ], + "spans": [ + { + "bbox": [ + 56, + 682, + 253, + 693 + ], + "type": "text", + "content": "2https://github.com/hendrycks/robustness" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 48, + 693, + 289, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 289, + 713 + ], + "type": "text", + "content": "3https://github.com/hendrycks/natural-adv-examples" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20387" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 86, + 282, + 213 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 277, + 82 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 277, + 82 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 277, + 82 + ], + "type": "text", + "content": "Table 4. Ablation study on the variants of the proposed MJP." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 86, + 282, + 213 + ], + "lines": [ + { + "bbox": [ + 51, + 86, + 282, + 213 + ], + "spans": [ + { + "bbox": [ + 51, + 86, + 282, + 213 + ], + "type": "table", + "html": "
MethodTop-1 Acc. ↑Consistency ↑
A: DeiT-S [37]79.864.3
B: A - PEs77.5 (-2.3)100.0
C: A + SPP [36]74.9 (-4.9)74.8
D: A + DAL (NLN)80.0 (+0.2)64.0
E: A + JP79.2 (-0.6)73.8
F: A + JP + IDX79.9 (+0.1)79.6
G: A + JP + UNK80.1 (+0.3)83.8
H: A + JP + UNK + DAL (PCA)79.9 (+0.1)83.4
I: A + JP + UNK + DAL (LN)80.0 (+0.2)83.8
J: A + JP + UNK + DAL (NLN)80.5 (+0.7)82.9
", + "image_path": "10e5889814a275ebfdb31517ba78c7a30635c7a7a7b7e62e6e58669ed2d2c762.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 224, + 269, + 331 + ], + "blocks": [ + { + "bbox": [ + 62, + 224, + 269, + 331 + ], + "lines": [ + { + "bbox": [ + 62, + 224, + 269, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 269, + 331 + ], + "type": "image", + "image_path": "8a6780d7bf30e61dbb646ffc547ec6ebca2ddd8fddce8cb0f5c51780bab1498f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 334, + 273, + 344 + ], + "lines": [ + { + "bbox": [ + 59, + 334, + 273, + 344 + ], + "spans": [ + { + "bbox": [ + 59, + 334, + 273, + 344 + ], + "type": "text", + "content": "Figure 4. Distributions of accumulated eigenvalues of PEs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 378 + ], + "type": "text", + "content": "verifies that the ViTs are naturally position-insensitive once without using PEs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 453 + ], + "type": "text", + "content": "We observe that the previous SPP strategy harms the accuracy of the model (i.e., " + }, + { + "bbox": [ + 46, + 380, + 287, + 453 + ], + "type": "inline_equation", + "content": "-4.9\\%" + }, + { + "bbox": [ + 46, + 380, + 287, + 453 + ], + "type": "text", + "content": "), which indicates it is insufficient to simply shuffle both patches and pixels in the input image. As we expect, the usage of UNK embedding alleviates the confusion between shuffled and unshuffled positions, which boosts both the accuracy and consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 454, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 550 + ], + "type": "text", + "content": "Finally, we can find that using dense absolute localization regression on the unmasked PEs can marginally boost the accuracy. Although PCA does not involve more parameters to the model, it significantly increases the computation latency (e.g., " + }, + { + "bbox": [ + 46, + 454, + 287, + 550 + ], + "type": "inline_equation", + "content": "+46\\%" + }, + { + "bbox": [ + 46, + 454, + 287, + 550 + ], + "type": "text", + "content": " with only 5 iterations). In general, it is better to use a nonlinear projection (e.g., a 3-layers MLP with negligible computation latency) to allow the learned PEs to aggregate more additional information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "Informativeness of the PEs. For a position space " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " (i.e., a 1-dim or 2-dim space), we may not need a high-dimension embedding space " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " to model the positions. To measure the informativeness of the learned PEs, we apply singular value decomposition (SVD) to PEs and analyze their eigenvalue distributions. Fig. 4 plots the curves of accumulated energy/sum of top-" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " eigenvalues versus the energy of all the eigenvalues. Supposing that we are using PCA to project the PEs of both matrices, to achieve the same explained variance ratio, our MJP needs more singular values (i.e., large dimensionality) than DeiT-S. This indicates that our positional embedding matrix is more informative. Similar observation is also revealed by Wang et al. [43]." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 126, + 545, + 241 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 125 + ], + "type": "text", + "content": "Table 5. Comparisons on gradient leakage by analytic attack [29] with ImageNet-1K validation set, where we test (1) ViT-S, DeiT-S and our model in the setting (a); (2) ViT-S, DeiT-S and our model in the setting (b) (i.e., MJP with " + }, + { + "bbox": [ + 305, + 70, + 545, + 125 + ], + "type": "inline_equation", + "content": "\\gamma = 0.27" + }, + { + "bbox": [ + 305, + 70, + 545, + 125 + ], + "type": "text", + "content": "); (3) ablation on without (w/o) using " + }, + { + "bbox": [ + 305, + 70, + 545, + 125 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{unk}}" + }, + { + "bbox": [ + 305, + 70, + 545, + 125 + ], + "type": "text", + "content": " in setting (a); and (4) Our model in setting (c)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 126, + 545, + 241 + ], + "lines": [ + { + "bbox": [ + 310, + 126, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 310, + 126, + 545, + 241 + ], + "type": "table", + "html": "
ModelSet.Acc.↑MSE↑\\( FFT_{2D} \\)↑PSNR↓SSIM↓LPIPS↑
(1)ViT-S [7]78.1.0278.003919.27.5203.3623
DeiT-S [37]a79.8.0350.005718.94.5182.3767
DeiT-S (w/o PEs)77.5.0379.008220.22.5912.2692
DeiT-S+MJP80.5.1055.016611.52.4053.6545
(2)ViT-S [7]18.7.0327.001618.44.6065.2836
DeiT-S [37]b36.0.0391.002417.60.5991.3355
DeiT-S (w/o PEs)77.5.0379.002520.25.6655.2370
DeiT-S+MJP62.9.1043.005911.66.4493.6519
(3)DeiT-S+MJP (w/o)a40.6.1043.005911.66.4493.6519
(4)DeiT-S+MJPc62.9.1706.03388.07.0875.8945
", + "image_path": "5bcb5f22693f22bacc98e7a6515e250e0c1b5c3baee7bacaca70845239b9c848.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 253, + 426, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 426, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 426, + 266 + ], + "type": "text", + "content": "5.4. Privacy Preservation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 272, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 547, + 392 + ], + "type": "text", + "content": "The fundamental principle of the gradient attack methods in federated learning is that each sample activates only a portion of content-related neurons in the deep neural networks, leading to one specific backward gradients for one related samples (i.e., 1-to-1 mapping). Based on such an observation, we argue that feeding ViTs with input patches with permuted sequences may intuitively mislead the attack. This is because now both the original and the transformed inputs may be matched to the same backward gradients (i.e., " + }, + { + "bbox": [ + 304, + 272, + 547, + 392 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 272, + 547, + 392 + ], + "type": "text", + "content": "-to-1 mapping)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 392, + 547, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 547, + 571 + ], + "type": "text", + "content": "To validate such an assumption, we utilize the public protocols4 to recover image with gradient updates in the privacy attack. In this privacy attack, we apply the Analytic Attack proposed in APRIL [29], which is designed for attacking the ViTs. We randomly sample 1K images from the validation set of ImageNet-1K (i.e., one image per category). To evaluate the anti-attack performance of a model, we introduce image similarity metrics to account for pixel-wise mismatch, including Mean Square Error (MSE), Peak Signal-to-Noise Ratio (PSNR), cosine similarity in the Fourier space " + }, + { + "bbox": [ + 304, + 392, + 547, + 571 + ], + "type": "inline_equation", + "content": "(\\mathrm{FFT}_{2\\mathrm{D}})" + }, + { + "bbox": [ + 304, + 392, + 547, + 571 + ], + "type": "text", + "content": ", and Learned Perceptual Image Patch Similarity (LPIPS) [49]. Different from the evaluation in gradient attacks [14, 29, 48], we suppose a model is with better capacity of privacy preservation when the recovered images from its gradient updates are less similar to the ground truth images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": " and its transformed (i.e., patch shuffled) version " + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": ", a ViT model " + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": ", and automatic evaluation metrics " + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": ", we conduct three different settings for fair comparisons: (a) " + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\phi(\\nabla \\mathcal{M}(\\mathbf{x}), \\mathbf{x})" + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": ", (b) " + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\phi(\\nabla \\mathcal{M}(\\tilde{\\mathbf{x}}), \\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": ", and (c) " + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\phi(\\nabla \\mathcal{M}(\\tilde{\\mathbf{x}}), \\mathbf{x})" + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": " refers to recovering input image through gradient attacks. Table 5 shows the quantitative comparisons between our method and the original ViTs for batch gradient inversion on ImageNet-1K. APRIL [29] enables a viable, complete recovery of original images from the gradient updates of the original ViTs. However, it per" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 520, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 520, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 520, + 713 + ], + "type": "text", + "content": "4https://github.com/JonasGeiping/breaching" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20388" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 521, + 246 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 521, + 246 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 521, + 246 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 521, + 246 + ], + "type": "image", + "image_path": "a8d90202ddd8e6794a1450fbfbec90fc3e2444738bc2403962663ff53e141104.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 254, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 254, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 547, + 277 + ], + "type": "text", + "content": "Figure 5. Visual comparisons on image recovery with gradient updates [29]. Our proposed DeiT-S+MJP model significantly outperforms the original ViT-S [7] and DeiT-S [37] models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 300, + 282, + 342 + ], + "blocks": [ + { + "bbox": [ + 47, + 285, + 286, + 297 + ], + "lines": [ + { + "bbox": [ + 47, + 285, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 286, + 297 + ], + "type": "text", + "content": "Table 6. Explained variance versus PCA projected dimensionality." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 300, + 282, + 342 + ], + "lines": [ + { + "bbox": [ + 49, + 300, + 282, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 300, + 282, + 342 + ], + "type": "table", + "html": "
Projected Dimension34567
DeiT-S EV (%)54.6168.5577.9585.5490.74
DeiT-S+MJP EV (%)46.7458.3669.1078.1384.55
", + "image_path": "f2a64ff88b9b557e10d61a528e1c5419c88e75cc30c072e6c63fb7d8e905b2ea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 345, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 381 + ], + "type": "text", + "content": "forms worse in recovering from \"DeiT-S+MJP\", leading to best performances on all evaluation metrics and outperform others by a large margin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 382, + 289, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 289, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 289, + 551 + ], + "type": "text", + "content": "More surprisingly, our proposed method makes APRIL yield unrecognizable images and fail in recovering the details in the original images (i.e., noisy patches in the outputs), as shown in Fig. 5. The left four columns in Fig. 5 are tested on original images, where all PEs are standard and correspond to their patch embeddings. Meanwhile, the right four columns are tested with transformed ones, where the shuffled patches are with the shared unknown PEs. Both the visual and quantitative comparisons verify that our MJP alleviates the gradient leakage problem. We also notice that DeiT-S without using PEs is inclined to be at higher risk of privacy leakage (i.e., easier to be attacked by gradients). These promising results indicate that our MJP is a promising strategy to protect user privacy in federated learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 560, + 119, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 119, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 119, + 572 + ], + "type": "text", + "content": "5.5. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 580, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 687 + ], + "type": "text", + "content": "PCA Projected Dimensionality. Table 6 presents the explained variance (EV) of our DeiT-S+MJP and DeiT-S versus different projection dimension. A low dimensionality can explain a large amount of information, which proves that the embedding matrix is sparse in nature. Moreover, to achieve the same explained variance ratio, our DeiT-S+MJP needs a large dimensionality than DeiT-S. This indicates that the positional embedding matrix of DeiT-S+MJP is less sparse but more informative." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": "Accuracy Vs. Shuffle Ratio. Intuitively, with the increase of the shuffle ratio from the proposed MJP method, the orig-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "type": "text", + "content": "inal intrinsic inductive bias will be undermined. However, from the experimental results, the performance of ViTs is actually boosted via the proposed MJP. To figure out the reason behind such counter-intuitive phenomenon, we visualize the last self-attention of our proposed method in Fig. 3 of our Supp. Mat. It shows that the attention heads of our method present more diverse and content-aware attentions than original DeiT-S. We think that it might be attributed to the proposed DAL, which strengthens the spatial information of the unmasked PEs in a more efficient manner. To this end, such learned content-aware attention becomes more meaningful and results in the better accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 441, + 379, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 379, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 379, + 454 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 677 + ], + "type": "text", + "content": "In this paper, we first visually demonstrate that PEs can explicitly learn the 2D spatial relationship from the input patch sequences. By feeding ViTs with transformed input, we identify the issue that PEs may weaken the position-insensitive property. Based on this observation, we propose an easy-to-reproduce yet effective Masked Jigsaw Puzzle (MJP) position embedding method to alleviate the conflict in PEs (preserving the consistency versus maintaining the accuracy). Experimental results show the proposed MJP can bring the the position-insensitive property back to ViTs without degenerating the accuracy on the large-scale dataset (i.e., ImageNet-1K). In a certain sense, the proposed MJP is also a data augmentation technique, which boosts the robustness to the common corruptions (e.g., ImageNet-C) and adversarial examples (e.g., ImageNet-A/O). Surprisingly, MJP can improve the privacy preservation capacity of ViTs under typical gradient attacks by a large margin, which may pilot a new direction for privacy preservation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was partly supported by the National AI Ph.D. for Society Program of Italy, and the EU H2020 project AI4Media (No. 951911)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20389" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations (ICLR), 2022. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 226 + ], + "type": "text", + "content": "[3] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning (ICML), 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "text", + "content": "[5] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 373 + ], + "type": "text", + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 441 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2020. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "type": "text", + "content": "[8] Philipp Dufter, Martin Schmitt, and Hinrich Schütze. Position information in transformers: An overview. Computational Linguistics, 48(3):733-763, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "type": "text", + "content": "[9] Yuxin Fang, Benchcheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection. Advances in Neural Information Processing Systems, 34:26183-26197, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "type": "text", + "content": "[10] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning (ICML), 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 578, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 288, + 621 + ], + "type": "text", + "content": "[11] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 624, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 288, + 667 + ], + "type": "text", + "content": "[12] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 244-253, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "type": "text", + "content": "[13] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on visual transformer. arXiv, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[14] Ali Hatamizadeh, Hongxu Yin, Holger Roth, Wenqi Li, Jan Kautz, Daguang Xu, and Pavlo Molchanov. Gradvit: Gradient inversion of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "text", + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 228 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "type": "text", + "content": "[17] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 275, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 547, + 308 + ], + "type": "text", + "content": "[18] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 546, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 546, + 353 + ], + "type": "text", + "content": "[19] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "type": "text", + "content": "[20] Yifan Jiang, Shiyu Chang, and Zhangyang Wang. Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 546, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 443 + ], + "type": "text", + "content": "[21] Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. Global-and-local relative position embedding for unsupervised video summarization. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 444, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 547, + 488 + ], + "type": "text", + "content": "[22] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 54(10s):1-41, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 547, + 521 + ], + "type": "text", + "content": "[23] Shun Kiyono, Sosuke Kobayashi, Jun Suzuki, and Kentaro Inui. Shape: Shifted absolute position embedding for transformers. arXiv preprint arXiv:2109.05644, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "type": "text", + "content": "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "type": "text", + "content": "[25] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 1, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "type": "text", + "content": "[26] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "text", + "content": "[27] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20390" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "text", + "content": "[29] Jiahao Lu, Xi Sheryl Zhang, Tianli Zhao, Xiangyu He, and Jian Cheng. April: Finding the achilles' heel on privacy for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10051-10060, 2022. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "type": "text", + "content": "[30] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "type": "text", + "content": "[31] Samrudhdhi B Rangrej, Chetan L Srinidhi, and James J Clark. Consistency driven sequential transformers attention model for partially observable scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2518-2527, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "type": "text", + "content": "[32] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "type": "text", + "content": "[33] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "type": "text", + "content": "[34] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "type": "text", + "content": "[35] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 288, + 544 + ], + "type": "text", + "content": "[36] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems (NeurIPS), 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 601 + ], + "type": "text", + "content": "[37] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. 1, 3, 5, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[40] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[41] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "text", + "content": "[42] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 229, + 547, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 547, + 284 + ], + "type": "text", + "content": "[43] Yu-An Wang and Yun-Nung Chen. What do position embeddings learn? an empirical study of pre-trained language model positional encoding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 1, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 285, + 547, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 547, + 327 + ], + "type": "text", + "content": "[44] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 330, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 330, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 547, + 373 + ], + "type": "text", + "content": "[45] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. Advances in Neural Information Processing Systems, 34:28522-28535, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 374, + 547, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 374, + 547, + 428 + ], + "spans": [ + { + "bbox": [ + 307, + 374, + 547, + 428 + ], + "type": "text", + "content": "[46] Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 430, + 547, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 547, + 483 + ], + "type": "text", + "content": "[47] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 486, + 547, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 547, + 540 + ], + "type": "text", + "content": "[48] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 541, + 547, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 541, + 547, + 596 + ], + "spans": [ + { + "bbox": [ + 307, + 541, + 547, + 596 + ], + "type": "text", + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 597, + 547, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 597, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 307, + 597, + 547, + 662 + ], + "type": "text", + "content": "[50] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 664, + 547, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 664, + 547, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 664, + 547, + 707 + ], + "type": "text", + "content": "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021. 1, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20391" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_content_list.json b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..06a966b55bd85a25347ecbc719249f91a1ca4195 --- /dev/null +++ b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_content_list.json @@ -0,0 +1,1605 @@ +[ + { + "type": "text", + "text": "Masked Motion Encoding for Self-Supervised Video Representation Learning", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyu Sun $^{12*}$ Peihao Chen $^{1*}$ Liangwei Chen $^{1}$ Changhao Li $^{1}$ Thomas H. Li $^{6}$ Mingkui Tan $^{157\\dagger}$ Chuang Gan $^{34}$", + "bbox": [ + 232, + 179, + 733, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ South China University of Technology, $^{2}$ Information Technology R&D Innovation Center of Peking University $^{3}$ UMass Amherst, $^{4}$ MIT-IBM Watson AI Lab, $^{5}$ Key Laboratory of Big Data and Intelligent Robot, Ministry of Education $^{6}$ Peking University Shenzhen Graduate School, $^{7}$ Pazhou Laboratory", + "bbox": [ + 125, + 217, + 841, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How to learn discriminative video representation from unlabeled videos is challenging but crucial for video analysis. The latest attempts seek to learn a representation model by predicting the appearance contents in the masked regions. However, simply masking and recovering appearance contents may not be sufficient to model temporal clues as the appearance contents can be easily reconstructed from a single frame. To overcome this limitation, we present Masked Motion Encoding (MME), a new pretraining paradigm that reconstructs both appearance and motion information to explore temporal clues. In MME, we focus on addressing two critical challenges to improve the representation performance: 1) how to well represent the possible long-term motion across multiple frames; and 2) how to obtain fine-grained temporal clues from sparsely sampled videos. Motivated by the fact that human is able to recognize an action by tracking objects' position changes and shape changes, we propose to reconstruct a motion trajectory that represents these two kinds of change in the masked regions. Besides, given the sparse video input, we enforce the model to reconstruct dense motion trajectories in both spatial and temporal dimensions. Pre-trained with our MME paradigm, the model is able to anticipate long-term and fine-grained motion details. Code is available at https://github.com/XinyuSun/MME.", + "bbox": [ + 76, + 335, + 472, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 744, + 209, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video representation learning plays a critical role in video analysis like action recognition [15,32,79], action localization [12,81], video retrieval [4,82], videoQA [40], etc. Learning video representation is very difficult for two reasons. Firstly, it is extremely difficult and labor-intensive to annotate videos, and thus relying on annotated data to learn", + "bbox": [ + 75, + 771, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "video representations is not scalable. Also, the complex spatial-temporal contents with a large data volume are difficult to be represented simultaneously. How to perform self-supervised videos representation learning only using unlabeled videos has been a prominent research topic [7,13,49].", + "bbox": [ + 500, + 305, + 890, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taking advantage of spatial-temporal modeling using a flexible attention mechanism, vision transformers [3, 8, 25, 26, 53] have shown their superiority in representing video. Prior works [5, 37, 84] have successfully introduced the mask-and-predict scheme in NLP [9, 23] to pre-train an image transformer. These methods vary in different reconstruction objectives, including raw RGB pixels [37], handcrafted local patterns [75], and VQ-VAE embedding [5], all above are static appearance information in images. Based on previous successes, some researchers [64, 72, 75] attempt to extend this scheme to the video domain, where they mask 3D video regions and reconstruct appearance information. However, these methods suffers from two limitations. First, as the appearance information can be well reconstructed in a single image with an extremely high masking ratio (85% in MAE [37]), it is also feasible to be reconstructed in the tube-masked video frame-by-frame and neglect to learn important temporal clues. This can be proved by our ablation study (cf. Section 4.2.1). Second, existing works [64, 75] often sample frames sparsely with a fixed stride, and then mask some regions in these sampled frames. The reconstruction objectives only contain information in the sparsely sampled frames, and thus are hard to provide supervision signals for learning fine-grained motion details, which is critical to distinguish different actions [3, 8].", + "bbox": [ + 496, + 383, + 892, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to design a new mask-and-predict paradigm to tackle these two issues. Fig. 1(a) shows two key factors to model an action, i.e., position change and shape change. By observing the position change of the person, we realize he is jumping in the air, and by observing the shape change that his head falls back and then tucks to his chest, we are aware that he is adjusting his posture to cross the bar. We believe that anticipating these changes helps the model better understand an action.", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Email: {csxinyusun, phchencs}@gmail.com", + "bbox": [ + 93, + 875, + 442, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. Email: mingkuitan@scut.edu.cn", + "bbox": [ + 98, + 887, + 395, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2235", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f272cf36b26507a7e48f4f7bcca41ae586621c1b8ea190f28b5134bbe0e3c822.jpg", + "image_caption": [ + "(a) Two key factors to recognize a high jump action." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 424, + 199 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bebe77796b7293aa8da91f9a76519ce7c7c8b7fab55d3eb53b9e5144016de71c.jpg", + "image_caption": [ + "(b) Appearance reconstruction vs. motion trajectory reconstruction.", + "Figure 1. Illustration of motion trajectory reconstruction for Masked Motion Encoding. (a) Position change and shape change over time are two key factors to recognize an action, we leverage them to represent the motion trajectory. (b) Compared with the current appearance reconstruction task, our motion trajectory reconstruction takes into account both appearance and motion information." + ], + "image_footnote": [], + "bbox": [ + 433, + 89, + 870, + 202 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this observation, instead of predicting the appearance contents, we propose to predict motion trajectory, which represents impending position and shape changes, for the mask-and-predict task. Specifically, we use a dense grid to sample points as different object parts, and then track these points using optical flow in adjacent frames to generate trajectories, as shown in Fig. 1(b). The motion trajectory contains information in two aspects: the position features that describe relative movement; and the shape features that describe shape changes of the tracked object along the trajectory. To predict this motion trajectory, the model has to learn to reason the semantics of masked objects based on the visible patches, and then learn the correlation of objects among different frames and try to estimate their accurate motions. We name the proposed mask-and-predict task as Masked Motion Encoding (MME).", + "bbox": [ + 75, + 291, + 468, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, to help the model learn fine-grained motion details, we further propose to interpolate the motion trajectory. Taking sparsely sampled video as input, the model is asked to reconstruct spatially and temporally dense motion trajectories. This is inspired by the video frame interpolation task [77] where a deep model can reconstruct dense video at the pixel level from sparse video input. Different from it, we aim to reconstruct the fine-grained motion details of moving objects, which has higher-level motion information and is helpful for understanding actions. Our main contributions are as follows:", + "bbox": [ + 75, + 535, + 468, + 700 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Existing mask-and-predict task based on appearance reconstruction is hard to learn important temporal clues, which is critical for representing video content. Our Masked Motion Encoding (MME) paradigm overcomes this limitation by asking the model to reconstruct motion trajectory.", + "- Our motion interpolation scheme takes a sparsely sampled video as input and then predicts dense motion trajectory in both spatial and temporal dimensions. This scheme endows the model to capture long-term and fine-grained motion clues from sparse video input." + ], + "bbox": [ + 94, + 719, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experimental results on multiple standard video recognition benchmarks prove that the representations learned from the proposed mask-and-predict task achieve state-of-the-art performance on downstream action recognition tasks. Specifically, pre-trained on Kinetics-400 [10], our MME brings the gain of $2.3\\%$ on Something-Something V2 [34], $0.9\\%$ on Kinetics-400, $0.4\\%$ on UCF101 [59], and $4.7\\%$ on HMDB51 [44].", + "bbox": [ + 496, + 291, + 890, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 434, + 640, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised Video Representation Learning. Self-supervised video representation learning aims to learn discriminative video features for various downstream tasks in the absence of accurate video labels. To this end, most of the existing methods try to design an advanced pretext task like predicting the temporal order of shuffled video crops [78], perceiving the video speediness [7, 13] or solving puzzles [43, 49]. In addition, contrastive learning is also widely used in this domain [14, 16, 36, 41, 46, 54, 57, 68, 69], which constrains the consistency between different augmentation views and brings significant improvement. In particular, ASCNet and CoCLR [36, 41] focus on mining hard positive samples in different perspectives. Optical flow has also been proven to be effective in capturing motion information [70, 76]. Besides, tracking video objects' movement is also used in self-supervised learning [17, 18, 65, 73, 74]. Among them, Wang et al. [74] only utilizes the spatial encoder to extract frame appearance information. CtP framework [65] and Siamese-triplet network [73] only require the model to figure out the position and size changes of a specific video patch. Different from these methods, our proposed MME trace the fine-grained movement and shape changes of different parts of objects in the video, hence resulting in a superior video representation. Tokmakov et al. [63] utilize Dense Trajectory to provide initial pseudo labels for video clustering. However, the model does not predict trajectory motion features explicitly. Instead, we consider long-term and fine-grained trajectory motion features as explicit reconstruction targets.", + "bbox": [ + 496, + 462, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2236", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a4b473440a755a23c6769f6da56003c8dbe69b01002fb2d4fc4c4647893b4768.jpg", + "image_caption": [ + "Figure 2. Overview of Masked Motion Encoding (MME). Given a sparsely sampled video, we first divide it into several patches and randomly mask out some of them. And then, we feed the remaining patches to a ViT encoder to extract video representation. Last, a lightweight ViT decoder is involved to predict the content in the masked region, i.e., a motion trajectory containing position changes and shape changes of moving objects." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 341, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44625b59b73d1a3c19cca9691457b126a13c597b7f1cc929e32566029637542d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 88, + 511, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0654f83de7f1007f192b32769cf0f8c4e5f6d8475731897f8b64b741ca50fe6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 88, + 890, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mask Modeling for Vision Transformer. Recently, BEiT and MAE [5, 37] show two excellent mask image modeling paradigms. BEVT [72] and VideoMAE [64] extend these two paradigms to the video domain. To learn visual representations, BEVT [72] predicts the discrete tokens generated by a pre-trained VQ-VAE tokenizer [56]. Nevertheless, pre-training such a tokenizer involves an unbearable amount of data and computation. In contrast, VideoMAE [64] pretrain the Vision Transformer by regressing the RGB pixels located in the masked tubes of videos. Due to asymmetric encoder-decoder architecture and extremely high masking ratio, model pre-training is more efficient with VideoMAE. Besides, MaskFeat [75] finds that predicting the Histogram of Gradient (HOG [21]) of masked video contents is a strong objective for the mask-and-predict paradigm. Existing methods only consider static information in each video frame, thus the model can speculate the masked area by watching the visible area in each frame independently and failed to learn important temporal clues (cf. Section 3.1). Different from the video prediction methods [35, 51, 58, 60] that predict the future frames in pixel or latent space, our Masked Motion Encoding paradigm predicts incoming fine-grained motion in masked video regions, including position changes and shape changes.", + "bbox": [ + 76, + 311, + 472, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 689, + 250, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first revisit the current masked video modeling task for video representation learning (cf. Section 3.1). Then, we introduce our masked motion encoding (MME), where we change the task from recovering appearance to recovering motion trajectory (cf. Section 3.2).", + "bbox": [ + 75, + 714, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rethinking Masked Video Modeling", + "text_level": 1, + "bbox": [ + 76, + 801, + 393, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video clip sampled from a video, self-supervised video representation learning aims to learn a feature encoder $f_{enc}(\\cdot)$ that maps the clip to its corresponding feature that best describes the video. Existing masked video modeling methods [64, 75] attempt to learn such a feature", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "encoder through a mask-and-predict task. Specifically, the input clip is first divided into multiple non-overlapped 3D patches. Some of these patches are randomly masked and the remaining patches are fed into the feature encoder, followed by a decoder $f_{dec}(\\cdot)$ to reconstruct the information in the masked patches. Different works aim to reconstruct different information (e.g., raw pixel in VideoMAE [64] and HOG in MaskFeat [75]).", + "bbox": [ + 496, + 311, + 890, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the existing works share a common characteristic where they all attempt to recover static appearance information of the masked patches. Since an image with a high masking ratio (85% in MAE [37]) can be well reconstructed [37,75], we conjecture that the masked appearance information of a video can also be reconstructed frame by frame independently. In this sense, the model may focus more on the contents in the same frame. This may hinder the models from learning important temporal clues, which is critical for video representation. We empirically study this conjecture in the ablation study (cf. Section 4.2.1).", + "bbox": [ + 496, + 434, + 892, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. General Scheme of MME", + "text_level": 1, + "bbox": [ + 500, + 609, + 732, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To better learn temporal clues of a video, our MME changes the reconstruction content from static appearance information to object motion information, including position and shape change of objects. As shown in Fig. 2, a video clip is sparsely sampled from a video and is divided into a number of non-overlapped 3D patches of size $t \\times h \\times w$ , corresponding to time, height, and width. We follow VideoMAE [64] to use the tube masking strategy, where the masking map is the same for all frames, to mask a subset of patches. For computation efficiency, we follow MAE [37] to only feed the unmasked patches (and their positions) to the encoder. The output representation together with learnable [MASK] tokens are fed to a decoder to reconstruct motion trajectory $\\mathbf{z}$ in the masked patches. The training loss for MME is", + "bbox": [ + 496, + 632, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i \\in \\mathcal {I}} \\left| \\mathbf {z} _ {i} - \\hat {\\mathbf {z}} _ {i} \\right| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 871, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2237", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{z}}$ is the predicted motion trajectory, and $\\mathcal{I}$ is the index set of motion trajectories in all masked patches.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the fact that we humans recognize actions by perceiving position changes and shape changes of moving objects, we leverage these two types of information to represent the motion trajectory. Through pre-training on MME task, the model is endowed with the ability to explore important temporal clues. Another important characteristic of the motion trajectory is that it contains fine-grained motion information extracted at raw video rate. This fine-grained motion information provides the model with a supervision signal to anticipate fine-grained action from sparse video input. In the following, we will introduce the proposed motion trajectory in detail.", + "bbox": [ + 75, + 122, + 470, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Motion Trajectory for MME", + "text_level": 1, + "bbox": [ + 76, + 314, + 333, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The motion of moving objects can be represented in various ways such as optical flow [27], histograms of optical flow (HOF) [45], and motion boundary histograms (MBH) [45]. However, these descriptors can only represent short-term motion between two adjacent frames. We hope our motion trajectory represents long-term motion, which is critical for video representation. To this end, inspired by DT [66], we first track the moving object in the following $L$ frames to cover a longer range of motion, resulting in a trajectory $\\mathbf{T}$ , i.e.,", + "bbox": [ + 75, + 338, + 468, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} = \\left(\\mathbf {p} _ {t}, \\mathbf {p} _ {t + 1}, \\dots , \\mathbf {p} _ {t + L}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 503, + 468, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_t = (x_t, y_t)$ represents a point located at $(x_t, y_t)$ of frame $t$ , and $(\\cdot, \\cdot)$ indicates the concatenation operation. Along this trajectory, we fetch the position features $\\mathbf{z}^p$ and shape features $\\mathbf{z}^s$ of this object to compose a motion trajectory $\\mathbf{z}$ , i.e.,", + "bbox": [ + 76, + 532, + 468, + 607 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\left(\\mathbf {z} ^ {p}, \\mathbf {z} ^ {s}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 609, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The position features are represented by the position transition relative to the last time step, while the shape features are the HOG descriptors of the tracked object in different time steps.", + "bbox": [ + 76, + 633, + 468, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tracking objects using spatially and temporally dense trajectories. Some previous works [2, 18, 19, 24] try to use one trajectory to represent the motion of an individual object. In contrast, DT [66] points out that tracking spatially dense feature points sampled on a grid space performs better since it ensures better coverage of different objects in a video. Following DT [66], we use spatially dense grid points as the initial position of each trajectory. Specifically, we uniformly sample $K$ points in a masked patch of size $t \\times h \\times w$ , where each point indicates a part of an object. For each point, we track it through temporally dense $L$ frames according to the dense optical flow, resulting in $K$ trajectories. In this way, the model is able to capture spatially and", + "bbox": [ + 75, + 704, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "temporally dense motion information of objects through the mask-and-predict task.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a comparison, the reconstruction contents in existing works [64, 75] often extract temporally sparse videos sampled with a large stride $s > 1$ . The model takes as input a sparse video and predicts these sparse contents for learning video representation. Different from these works, our model also takes as input sparse video but we push the model to interpolate motion trajectory containing fine-grained motion information. This simple trajectory interpolation task does not increase the computational cost of the video encoder but helps the model learn more fine-grained action information even given sparse video as input. More details about dense flow calculating and trajectory tracking can be found in Appendix.", + "bbox": [ + 496, + 121, + 890, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Representing position features. Given a trajectory $\\mathbf{T}$ consisting of the tracked object position at each frame, we are more interested in the related movement of objects instead of their absolute location. Consequently, we represent the position features with related movement between two adjacent points $\\Delta \\mathbf{p}_t = \\mathbf{p}_{t+1} - \\mathbf{p}_t$ , i.e.,", + "bbox": [ + 496, + 325, + 890, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {p} = \\left(\\Delta \\mathbf {p} _ {t}, \\dots , \\Delta \\mathbf {p} _ {t + L - 1}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 429, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}^p$ is a $L\\times 2$ dimensional feature. As each patch contains $K$ position features, we concatenate and normalize them as position features part of the motion trajectory.", + "bbox": [ + 496, + 455, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Representing shape features. Besides embedding the movement, the model also needs to be aware of the shape changes of objects to recognize actions. Inspired by HOG [21], we use histograms of oriented gradients (HOG) with 9 bins to describe the shape of objects.", + "bbox": [ + 496, + 511, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with existing works [64, 75] that reconstruct HOG in every single frame, we are more interested in the dynamic shape changes of an object, which can better represent action in a video. To this end, we follow DT [66] to calculate trajectory-aligned HOG, consisting of HOG features around all tracked points in a trajectory, i.e.,", + "bbox": [ + 496, + 585, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {s} = \\left(\\operatorname {H O G} \\left(\\mathbf {p} _ {t}\\right), \\dots , \\operatorname {H O G} \\left(\\mathbf {p} _ {t + L - 1}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 688, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{HOG}(\\cdot)$ is the HOG descriptor and $\\mathbf{z}^s$ is a $L\\times 9$ dimensional feature. Also, as one patch contains $K$ trajectories, we concatenate $K$ trajectory-aligned HOG features and normalize them to the standard normal distribution as the shape features part of the motion trajectory.", + "bbox": [ + 496, + 715, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 805, + 633, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details. We conduct experiments on Kinetics-400 (K400), Something-Something V2 (SSV2), UCF101 and HMDB51 datasets. Unless otherwise stated, we follow previous trails [64] and feed the model a", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2238", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/37a2c3ca831e259e043ce9953161e6a3787deb950f95d3d034a0bfc5f3e3359f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbonePre-trainLabelEpochFramesGFLOPsParam.Acc.@1
TimeSformer [8]ViT-BIN-21K✓158196×1×3121M59.5
ViViT-L [3]ViT-LIN-21K+K400✓35163992 × 3 × 4311M65.4
Motionformer [53]ViT-BIN-21K+K400✓3516370 × 1 × 3109M66.5
SlowFast [29]ResNet-101K400✓1968+32106 × 1 × 353M63.1
MViT [26]MViT-BK400✓10064455 × 1 × 337M67.7
SVT [57]ViT-BIN-21K+K400×208+64196×1×3121M59.6
LSTCL [68]Swin-BK400×20016360×5×188M67.0
BEVT [72]Swin-BK400+DALL-E×15032282×1×388M67.1
VIMPAC [62]ViT-LHowTo100M+DALLE×10010N/A×10×3307M68.1
VideoMAE [64]ViT-BSSv2×40016180×2×387M67.6
VideoMAE [64]ViT-BSSv2×80016180×2×387M69.3
VideoMAE [64]ViT-BK400×80016180×2×387M68.2
MME (Ours)ViT-BSSv2×40016180×2×387M69.2
MME (Ours)ViT-BSSv2×80016180×2×387M70.0
MME (Ours)ViT-BK400×40016180×2×387M69.3
MME (Ours)ViT-BK400×80016180×2×387M70.5
", + "bbox": [ + 122, + 88, + 851, + 334 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cee3848ff4e33e251d74b08968a87bdcd2096728681833df99439ed26e7eed90.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-arts on Something-Something V2. Our MME outperforms the previous best method. Remarkably, MME show better transferability ability (transfer from K400 to SSv2) compared with VideoMAE." + ], + "table_footnote": [], + "table_body": "
MethodBackbonePre-trainLabelEpochFramesGFLOPsParam.Acc.@1
TimeSformer [8]ViT-BIN-21K✓158196×1×3121M78.3
Motionformer [53]ViT-BIN-21K✓3516370×10×3121M79.7
Motionformer [53]ViT-HRIN-21K✓3516959×10×3382M81.1
VIMPAC [62]ViT-LHowTo100M+DALLE×10010N/A×10×3307M77.4
SVT [57]ViT-BIN-21K+K400×208+64196×1×3121M78.1
BEVT [72]Swin-BK400+DALL-E×15032282×1×388M76.2
BEVT [72]Swin-BIN21K+K400+DALL-E×15032282×1×388M80.6
LSTCL [68]Swin-BK400×20016360×5×188M81.5
OmniMAE [33]ViT-BIN1K+K400×160032180×5×387M80.6
VideoMAE [64]ViT-BK400×160016180×5×387M80.9
ST-MAE [28]ViT-BK400×160016180×7×387M81.3
MME (Ours)ViT-BK400×160016180×7×387M81.8
", + "bbox": [ + 122, + 387, + 849, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Comparison with state-of-the-arts on Kinetics-400. Our MME outperforms the previous best method.", + "bbox": [ + 148, + 597, + 818, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$224 \\times 224$ 16-frame clip with strides 4 and 3 for K400 and SSV2 respectively. We use vanilla ViT-Base [25] as the backbone in all experiments. As for the motion trajectory, we set the trajectory length to $L = 6$ and the number of trajectories in each patch to $K = 4$ . Our models are trained on 16 NVIDIA GeForce RTX 3090 GPUs. Note that for all the ablation studies, we use a smaller pre-training dataset and reduce the input patches of fine-tuning to speed up the experiments, details can be found in Appendix.", + "bbox": [ + 75, + 631, + 470, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparisons with State-of-the-arts", + "text_level": 1, + "bbox": [ + 76, + 784, + 382, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-tuning. On the temporal-heavy dataset SSv2, we conduct experiments using two settings, namely in-domain pre-training (i.e., pre-train and fine-tune on the same dataset) and cross-domain pre-training (i.e., pre-train and fine-tune on different datasets). In Tab. 1, we have two main observations. 1) Our MME significantly performs", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "better than previous state-of-the-arts VideoMAE [64] on both in-domain and cross-domain settings. Specifically, under cross-domain settings (transferring from K400 to SSv2) with 800 epochs, our MME outperforms it by $2.3\\%$ , increasing the accuracy from $68.2\\%$ to $70.5\\%$ . 2) Our MME performs better under cross-domain setting compared with indomain setting, which is contrary to the existing observation on VideoMAE [64] and ST-MAE [28]. We attribute it to our masked motion encoding pretext task that helps models focus more on learning temporal clues instead of learning domain-specific appearance information (e.g., color, lighting). The learned temporal clues are more universal across different domains for video representation. This also shows the potential for taking advantage of larger datasets for pretraining, which we leave for future work.", + "bbox": [ + 496, + 631, + 890, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 2, our MME also achieves state-of-the-art performance on another large-scale dataset K400 with a gain", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2239", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/723550ed243c0b9b8d7b77a4327fef33ee76300a37ec43f0e10f81901e91b342.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbonePre-trainUCF101HMDB51
XDC [1]R(2+1)DIG65M94.267.1
GDT [52]R(2+1)DIG65M95.272.8
CVRL [54]R50K40092.967.9
CORPf [39]R50K40093.568.0
ρBYOL [30]R50K40094.272.1
VideoMAE [64]ViT-BK40096.173.3
MME (Ours)ViT-BK40096.578.0
", + "bbox": [ + 81, + 88, + 470, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of $0.5\\%$ . In Tab. 3, we also evaluate the transferability of MME by transferring a K400 pre-trained model to two small-scale datasets, namely UCF101 and HMDB51. Our MME achieves a large gain of $4.7\\%$ on HMDB51 and $0.4\\%$ on UCF101 compared with VideoMAE. This further demonstrates the strong ability of MME to learn from the large-scale dataset and transfer it to a small-scale one.", + "bbox": [ + 75, + 267, + 468, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Linear probing. We further evaluate our MME under liner probing setting on the SSv2 dataset as this dataset focuses more on the action. We follow SVT [57] to fix the transformer backbone and train a linear layer for 20 epochs. In Tab. 4, our method significantly outperform previous contrastive-based method [57] by $10.9\\%$ and masked video modeling methods [64] by $4.9\\%$ . These results indicate that the representation learned from our MME task contains more motion clues for action recognition.", + "bbox": [ + 75, + 382, + 468, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study on Reconstruction Content", + "text_level": 1, + "bbox": [ + 76, + 529, + 441, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we investigate the effect of different reconstruction contents, including appearance, short-term motion, and the proposed long-term fine-grained motion.", + "bbox": [ + 76, + 553, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Evaluating Appearance Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 619, + 405, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In section 3.1, we conjecture that appearance reconstruction is finished in each video frame independently. To verify this conjecture, we corrupt the temporal information of a video by randomly shuffling its frames. Since the video frames are randomly permuted, it is hard for a model to leverage information from other frames to reconstruct the current frame through their temporal correlation. We mask $90\\%$ regions of these videos following VideoMAE [64]. In Fig. 3 (a), we find that the reconstruction error (i.e., L2 error between predicted and ground-truth pixel) converges to a low value. We also conduct the same experiment using raw video without shuffling. The model undergoes a similar convergence process. This demonstrates that the masked pixels are well reconstructed without temporal information.", + "bbox": [ + 75, + 643, + 468, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further evaluate whether the model pre-trained on the appearance reconstruction task (i.e., VideoMAE) can well capture important temporal clues, we transfer two", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5b62ca779f755a39e4f3055e3f7858054cac5fd80c6e571830b05fa753f1dffe.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-arts on UCF101 and HMDB51." + ], + "table_footnote": [], + "table_body": "
MethodBackbonePre-trainAcc.@1
TimeSformer [8]ViT-BIN-21K14.0
SVT [57]ViT-BIN-21K+K40018.3
VideoMAE [64]ViT-BSSv224.3
MME (Ours)ViT-BSSv229.2
", + "bbox": [ + 509, + 88, + 883, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Linear Probing on Something-Something V2.", + "bbox": [ + 524, + 191, + 864, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56be86abe4159245519ff5cb14764bb11b4932b0cc46e0cd4ae25e501ca43177.jpg", + "image_caption": [ + "(a) Masked video reconstruction error.", + "Figure 3. We perform mask-and-predict pre-training using temporal shuffled videos and raw videos and then transfer them to the downstream task. (a) Masked appearance (i.e., pixels) is well reconstructed without temporal information. (b) Our MME outperforms VideoMAE when both provided temporal information." + ], + "image_footnote": [], + "bbox": [ + 504, + 213, + 692, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e3244eb30176476d083000ccfc9011e7d083583d70fbb073b1d5d56688315eb.jpg", + "image_caption": [ + "(b) Downstream performance." + ], + "image_footnote": [], + "bbox": [ + 692, + 214, + 887, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VideoMAE models (pre-trained on shuffled videos and raw videos, respectively) to the downstream action recognition task. In Fig. 3 (b), these two models perform competitively, indicating that removing the temporal information from the pre-training videos barely affects the learned video representation. We speculate this is because the VideoMAE model pays little attention to temporal information when performing the appearance reconstruction pre-training task. As a comparison, our MME paradigm performs significantly better than VideoMAE when the temporal information is provided for pre-training $(64.1\\%$ vs. $60.9\\%)$ . Once the temporal information is removed, MME performs similarly with VideoMAE $(60.7\\%$ vs. $60.9)$ . This demonstrates our MME paradigm takes great advantage of temporal information in the pre-training phase.", + "bbox": [ + 496, + 469, + 892, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Evaluating Motion Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 717, + 795, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following ablations, we separately verify the effectiveness of two proposals of our methods:", + "bbox": [ + 498, + 741, + 890, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of motion trajectory. We compare our motion trajectory with two appearance reconstruction baselines and three short-term motion reconstruction baselines.", + "bbox": [ + 498, + 781, + 890, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Appearance reconstruction baselines: 1) HOG: predicting HOG features [21] in the middle frame of each 3D patch [75]. 2) Pixel: predicting all the pixels of each 3D patch [28, 37, 64].", + "bbox": [ + 517, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2240", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/12767c128a5d3e7632592433d86f3744e3cc4a13dafcf14f2b71e14b6ba20ec3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Reconstruction contentAcc.@1Acc.@5
Pixel61.186.6
HOG61.186.9
HOG + Flow60.986.4
HOG + HOF61.386.6
HOG + MBH61.486.8
HOG + Traj. w/o shape63.588.2
Traj. (ours)64.188.4
", + "bbox": [ + 124, + 88, + 423, + 227 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/308e7d11215a368eb63c95bb87d07581c2ce8fb2d4891b36f7743a927a775847.jpg", + "table_caption": [ + "Table 5. Comparison of different Reconstruction content. Our motion trajectory outperforms all the other contents." + ], + "table_footnote": [], + "table_body": "
Reconstruction contentInterpolationAcc.@1Acc.@5
HOGx61.186.9
HOG✓60.986.8
HOG + MBHx61.386.8
HOG + MBH✓61.486.8
Traj. (ours)x62.287.5
Traj. (ours)✓64.188.4
", + "bbox": [ + 78, + 275, + 468, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Effectiveness of trajectory interpolation. Interpolating the motion trajectory brings significant improvement compared to the baselines.", + "bbox": [ + 76, + 412, + 468, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Short-term motion reconstruction baselines: 1) Flow: predicting the dense optical flow map of the masked patches. 2) HOF: predicting the histogram of optical flow orientation. 3) MBH: predicting the motion boundary histogram, which is the second-order local statistic of optical flow.", + "bbox": [ + 96, + 476, + 468, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 5, two types of appearance reconstruction contents (i.e., Pixel and HOG) perform similarly. Based on the HOG baseline, predicting additional short-term motion contents brings slight improvement (at least by $0.2\\%$ ) except for directly predicting optical flow (decreases by $0.2\\%$ ). We suspect that predicting the dense flow map for each pixel is too hard for a model and thus harms the performance. Instead, the histogram version of flow (i.e., HOF and MBH) help to reduce the prediction ambiguity [75], resulting in better performance. Furthermore, after adding our slight variant of motion trajectory (i.e., with position features but without shape features shown in Eq. (3)), the performance improves significantly compared to both appearance reconstruction baselines (by $2.4\\%$ ) and short-term motion reconstruction baselines (by $2.1\\%$ ). On top of this variant, adding trajectory-aligned shape features further brings additional gain by $0.6\\%$ . Because tracking the trajectory of objects in multiple frames provides long-term motion compared to short-term motion reconstruction, we believe the learning of long-term motion clues is one of the keys to enhancing video representation.", + "bbox": [ + 75, + 583, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of trajectory interpolation. In MME, we feed temporally sparse video to the model and push it to interpolate temporally dense trajectories. We are interested in whether interpolating appearance (i.e., HOG) or short-term motion contents (i.e., MBH) benefit learning fine-grained motion details. To this end, we conduct experiments to predict dense HOG and MBH, i.e., predicting these features of all frames within a masked patch even though these frames are not sampled as input. In Tab. 6, interpolating HOG or MBH brings little improvement. We suspect this is because these dense features represent the continuous movement of objects poorly, as they have not explicitly tracked the moving objects. In contrast, our temporally dense motion trajectory tracks the object densely, providing temporal fine-grained information for learning. The trajectory interpolation improves our MME by $1.9\\%$ . This result proves that the learning of fine-grained motion details is another key to enhancing video representation.", + "bbox": [ + 496, + 90, + 890, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study on MME Design", + "text_level": 1, + "bbox": [ + 498, + 372, + 782, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to investigate the effect of different parameters for MME pre-training.", + "bbox": [ + 498, + 395, + 890, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Trajectory length. Trajectory length $L$ indicates how many frames we track an object for extracting the motion trajectory. If $L = 0$ , our method degrades to MaskFeat [75], which only predicts HOG features in the masked patch. In Tab. 7a, MME achieves the best performance when $L = 6$ . This is because a longer trajectory provides more motion information for learning. But if the trajectory is too long, the drift of flow introduces accumulated noise when tracking the objects, which harms the performance.", + "bbox": [ + 496, + 444, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Trajectory density. In our experiments, we spatially dense track $K = 4$ motion trajectories. We conduct experiments to evaluate whether it is necessary to densely track multiple trajectories in a patch. Specifically, we assume only one object exists in a patch, and thus the 4 trajectories are similar. We merge these 4 trajectories as one by averaging or selecting the most salient one. We then consider the merged trajectories as the trajectory motion. In Tab. 7c, using dense trajectories significantly outperform the merged baselines. We suspect the spatial dense trajectories provide more motion information for learning.", + "bbox": [ + 496, + 589, + 890, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Trajectory normalization. Due to the irregular distribution of motion and shape information in the video, the trajectory values in different patches vary a lot. To help the model better model the trajectory, we adopt the patch norm method to normalize the trajectory values in each patch into a standard normal distribution, as has also been done in MAE [37]. The mean and variance are calculated by all $K = 4$ trajectories in a patch. Tab. 7b shows the effectiveness of adopting the patch norm.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2241", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/360003841f901d4ac3e55ee1440110715b7c63848ce26848b0fda1e06ae1377b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LengthAcc.@1Acc.@5
L = 463.288.0
L = 664.188.4
L = 863.388.0
", + "bbox": [ + 86, + 88, + 287, + 160 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/776b33282d4c341504062f9acc5f457986666082f547bb8d1e60d18ac1927db2.jpg", + "table_caption": [ + "(a) Effect of trajectory length." + ], + "table_footnote": [], + "table_body": "
SettingAcc.@1Acc.@5
w/o patch norm.62.187.44
w/ patch norm.64.188.4
", + "bbox": [ + 315, + 88, + 570, + 157 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2245de0dfa20d557d11d9c8344a7aefc1e416471d7a9bdf7f780ef66388d0151.jpg", + "table_caption": [ + "(b) Effect of trajectory normalization." + ], + "table_footnote": [], + "table_body": "
Trajectory densityAcc.@1Acc.@5
Dense64.188.4
Sparse (Max)60.286.3
Sparse (Mean)61.887.1
", + "bbox": [ + 596, + 88, + 864, + 160 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/49fc1dd7758d43aa1f495bc0318bd64b24caf9c4ea3217c5b6a3ac9ea2d50af1.jpg", + "table_caption": [ + "(c) Effect of trajectories density." + ], + "table_footnote": [], + "table_body": "
StrideAcc.@1Acc.@5
s = 462.987.7
s = 364.188.4
s = 263.588.3
s = 162.787.7
", + "bbox": [ + 91, + 193, + 285, + 279 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2292f77e0c3a2c0b2816508042af21a6562c2dbcc3b6bc1ee3735e19578e0cc8.jpg", + "table_caption": [ + "(d) Effect of sampling stride." + ], + "table_footnote": [], + "table_body": "
RatioTypeAcc.@1Acc.@5
90%Tube62.787.8
70%Tube64.188.4
40%Tube61.386.9
40%Cube61.286.9
", + "bbox": [ + 318, + 193, + 558, + 279 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f4b51b609a4b94add680eb84d14b4b1da7405489f84a324fb03f5488887b81af.jpg", + "table_caption": [ + "(e) Effect of spatial masking." + ], + "table_footnote": [], + "table_body": "
TypeDepthFLOPsAcc.@1Acc.@5
Joint151G64.188.4
255G63.488.3
Sep.156G63.488.4
265G63.388.0
", + "bbox": [ + 581, + 189, + 887, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(f) Effect of decoder. Sep. means two decoders.", + "bbox": [ + 591, + 284, + 877, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Pre-training details. We explore the detailed setting of our MME pre-training. The default entry is marked in gray .", + "bbox": [ + 109, + 310, + 854, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sampling stride. In order to control the difficulty of interpolating the motion trajectory, we ablate different temporal sampling strides leading to different levels of the sparse input video. With a medium sampling stride ( $s = 2$ ), our MME achieves the best performance. We also found that the performance decreases when the input video shares the same stride as the dense motion trajectory ( $s = 1$ ). One possible reason is that it is difficult for the model to anticipate fine-grained motion details in this case.", + "bbox": [ + 75, + 345, + 468, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial masking strategy. We ablated study different spatial masking strategies and ratios in Tab. 7e. We start from the default setting of MaskFeat [75] that uses a cube masking strategy with a $40\\%$ masking ratio, then we increase the masking ratio and adopt the tube masking strategy proposed by VideoMAE [64]. Using tube masking with a medium masking ratio $(70\\%)$ performs the best.", + "bbox": [ + 75, + 491, + 468, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decoder setting. We use several transformer blocks as decoder to reconstruct motion trajectory. We ablate different architectures of the decoder in Tab. 7f. Using a slight decoder with 1 transformer block performs the best, with the lowest computational cost. Increasing the depth of the decoder or using two parallel decoders to separately predict the position and shape features introduces more computational cost but brings little improvement.", + "bbox": [ + 75, + 607, + 468, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualization Results", + "text_level": 1, + "bbox": [ + 76, + 739, + 274, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the predicted motion trajectory in Fig. 4. The model is able to reconstruct the spatial and temporally dense trajectories of different object parts from a temporally sparse input video. Even without seeing the object in the current frame (the hand being masked out), the model is able to locate it through contextual patches and accurately estimate its trajectory. We only visualize the position changes since our prediction of HOG features has similar quality to the previous work [75].", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26244ef12f2d3ac2bd9de8767aa9a07b365bc3d25aaa0d4adbdc80fe5a848d1e.jpg", + "image_caption": [ + "Figure 4. Visualization of predicted motion trajectory. Each column represents (from left to right): (a) Masked frames; (b) ground truth; (c) prediction; (d) original frame." + ], + "image_footnote": [], + "bbox": [ + 516, + 344, + 875, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 551, + 617, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Masked Motion Encoding, a simple mask-and-predict paradigm that learns video representation by reconstructing the motion contents of masked regions. In particular, we reconstruct motion trajectory, which tracks moving objects densely to record their position and shape changes. By reconstructing this motion trajectory, the model learns long-term and fine-grained motion clues from sparse input videos. Empirical results show that our method outperforms the state-of-the-art mask-and-predict methods on Something-Something V2, Kinetics-400, UCF-101, and HMDB-51 datasets for action recognition.", + "bbox": [ + 496, + 579, + 892, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 767, + 660, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by the National Natural Science Foundation of China (NSFC) (No. 62072190, 62072186), Ministry of Science and Technology Foundation Project (No. 2020AAA0106900), Program for Guangdong Introducing Innovative and Entrepreneurial Teams (No. 2017ZT07X183), Guangdong Basic and Applied Basic Research Foundation (No. 2019B1515130001).", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2242", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, pages 9758-9770, 2020. 6", + "[2] Nadeem Anjum and Andrea Cavallaro. Multifeature object trajectory clustering for video analysis. IEEE Transactions on Circuits and Systems for Video Technology, 18:1555-1564, 2008. 4", + "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 1, 5", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 1", + "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: BERT pre-training of image transformers. In ICLR, 2022. 1, 3, 13", + "[6] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417, 2006. 12", + "[7] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In CVPR, pages 9922-9931, 2020. 1, 2", + "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding. In ICML, pages 813-824, 2021. 1, 5, 6", + "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 1", + "[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 2", + "[11] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703, 2020. 13", + "[12] Peihao Chen, Chuang Gan, Guangyao Shen, Wenbing Huang, Runhao Zeng, and Mingkui Tan. Relation attention for temporal action localization. IEEE Transactions on Multimedia, 22:2723-2733, 2019. 1", + "[13] Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan. Rspnet: Relative speed perception for unsupervised video representation learning. In AAAI, pages 1045-1053, 2021. 1, 2", + "[14] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas H Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurlPS, 2022. 2", + "[15] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 29:8292-8302, 2020. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Zhenfang Chen, Lin Ma, Wenhan Luo, Peng Tang, and Kwan-Yee K Wong. Look closer to ground better: Weakly-supervised temporal grounding of sentence in video. arXiv, 2020. 2", + "[17] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In ACL, 2019. 2", + "[18] Zhenfang Chen, Jiayuan Mao, Jiajun Wu, Kwan-Yee Kenneth Wong, Joshua B Tenenbaum, and Chuang Gan. Grounding physical concepts of objects and events through dynamic visual reasoning. arXiv, 2021. 2, 4", + "[19] Zhenfang Chen, Kexin Yi, Yunzhu Li, Mingyu Ding, Antonio Torralba, Joshua B Tenenbaum, and Chuang Gan. Comphy: Compositional physical reasoning of objects and events from videos. In ICLR, 2022. 4", + "[20] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, pages 702-703, 2020. 13", + "[21] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, pages 886-893, 2005. 3, 4, 6, 13", + "[22] Navneet Dalal, Bill Triggs, and Cordelia Schmid. Human detection using oriented histograms of flow and appearance. In ECCV, pages 428-441, 2006. 13", + "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, pages 4171–4186, 2018. 1", + "[24] Mingyu Ding, Zhenfang Chen, Tao Du, Ping Luo, Josh Tenenbaum, and Chuang Gan. Dynamic visual reasoning by learning differentiable physics models from video and language. In NeurIPS, volume 34, pages 887-899, 2021. 4", + "[25] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 5", + "[26] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 1, 5", + "[27] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In *SCIA*, pages 363-370, 2003. 4", + "[28] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In NeurIPS, 2022. 5, 6", + "[29] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 5", + "[30] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross B. Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, pages 3299-3309, 2021. 6", + "[31] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2243", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image analysis and automated cartography. Communications of the ACM, 24:381-395, 1981. 12", + "[32] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In European Conference on Computer Vision (ECCV), pages 758-775, 2020. 1", + "[33] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 5", + "[34] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 2", + "[35] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martín-Martín, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. arXiv preprint arXiv:2206.11894, 2022. 3", + "[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In NeurIPS, pages 5679-5690, 2020. 2", + "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 3, 6, 7", + "[38] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. In CVPR, pages 8126-8135, 2020. 13", + "[39] Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. In ICCV, pages 7939-7949, 2021. 6", + "[40] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware Graph Convolutional Networks for Video Question Answering. In AAAI, 2020. 1", + "[41] Deng Huang, Wenhao Wu, Weiwen Hu, Xu Liu, Dongliang He, Zhihua Wu, Xiangmiao Wu, Mingkui Tan, and Errui Ding. Ascnet: Self-supervised video representation learning with appearance-speed consistency. In ICCV, pages 8096-8105, 2021. 2", + "[42] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q. Weinberger. Deep networks with stochastic depth. In ECCV, pages 646-661, 2016. 13", + "[43] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. 2", + "[44] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, pages 2556-2563, 2011. 2", + "[45] Ivan Laptev, Marcin Marszalek, Cordelia Schmid, and Benjamin Rozenfeld. Learning realistic human actions from movies. In CVPR, pages 1-8, 2008. 4, 13" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Yuan Liu, Jingyuan Chen, Zhenfang Chen, Bing Deng, Jianqiang Huang, and Hanwang Zhang. The blessings of unlabeled background in untrimmed videos. In CVPR, pages 6176-6185, 2021. 2", + "[47] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 13", + "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 13", + "[49] Dezhao Luo, Chang Liu, Yu Zhou, Dongbao Yang, Can Ma, Qixiang Ye, and Weiping Wang. Video cloze procedure for self-supervised spatio-temporal learning. In AAAI, pages 11701-11708, 2020. 1, 2", + "[50] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035, 2019. 13", + "[51] Viorica Patraucean, Ankur Handa, and Roberto Cipolla. Spatio-temporal video autoencoder with differentiable memory. In ICLR Workshop, 2016. 3", + "[52] Mandela Patrick, Yuki M Asano, Polina Kuznetsova, Ruth Fong, Joao F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. 6", + "[53] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In NeurIPS, pages 12493-12506, 2021. 1, 5", + "[54] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, pages 6964-6974, 2021. 2, 6", + "[55] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Xiang Wang, Yuehuan Wang, Yiliang Lv, Changxin Gao, and Nong Sang. Mar: Masked autoencoders for efficient action recognition. arXiv preprint arXiv:2207.11660, 2022. 13", + "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 3", + "[57] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael Ryoo. Self-supervised video transformer. In CVPR, pages 2874-2884, 2022. 2, 5, 6", + "[58] Adria Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătráucean, Florent Altché, Michal Valko, et al. Broaden your views for self-supervised video learning. In ICCV, pages 1255-1265, 2021. 3", + "[59] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2", + "[60] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, pages 843-852, 2015. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2244", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 13", + "[62] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 5", + "[63] Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. Unsupervised learning of video representations via dense trajectory clustering. In ECCV 2020 Workshops, pages 404-421. Springer, 2020. 2", + "[64] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 3, 4, 5, 6, 8, 13", + "[65] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. In CVPR, pages 2563-2572, 2021. 2", + "[66] Heng Wang, Alexander Kläser, Cordelia Schmid, and Cheng-Lin Liu. Dense trajectories and motion boundary descriptors for action recognition. International journal of computer vision, 103:60-79, 2013. 4, 14", + "[67] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In ICCV, pages 3551-3558, 2013. 12, 14", + "[68] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. arXiv preprint arXiv:2106.09212, 2021. 2, 5", + "[69] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In CVPR, pages 11804-11813, 2021. 2", + "[70] Jiangliu Wang, Jianbo Jiao, Linchao Bao, Shengfeng He, Yunhui Liu, and Wei Liu. Self-supervised spatio-temporal representation learning for videos by predicting motion and appearance statistics. In CVPR, pages 4006-4015, 2019. 2", + "[71] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41:2740-2755, 2018. 12", + "[72] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, pages 14733-14743, 2021. 1, 3, 5", + "[73] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, pages 2794-2802, 2015. 2", + "[74] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, pages 2566-2576, 2019. 2", + "[75] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 1, 3, 4, 6, 7, 8, 13" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Donglai Wei, Joseph J. Lim, Andrew Zisserman, and William T. Freeman. Learning and using the arrow of time. In CVPR, pages 8052-8060, 2018. 2", + "[77] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P Allebach, and Chenliang Xu. Zooming slow-mo: Fast and accurate one-stage space-time video super-resolution. In CVPR, pages 3370-3379, 2020. 2", + "[78] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In CVPR, pages 10334–10343, 2019. 2", + "[79] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 1", + "[80] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 13", + "[81] Runhao Zeng, Chuang Gan, Peihao Chen, Wenbing Huang, Qingyao Wu, and Mingkui Tan. Breaking winner-takes-all: Iterative-winners-out networks for weakly supervised temporal action localization. IEEE Transactions on Image Processing, 28(12):5797-5808, 2019. 1", + "[82] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1", + "[83] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 13", + "[84] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan L. Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 1" + ], + "bbox": [ + 501, + 92, + 890, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2245", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_model.json b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8d7b5a8e5e0e5797cff6d39f978be3cd8d3ffcbc --- /dev/null +++ b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.877, + 0.154 + ], + "angle": 0, + "content": "Masked Motion Encoding for Self-Supervised Video Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.18, + 0.735, + 0.217 + ], + "angle": 0, + "content": "Xinyu Sun\\(^{12*}\\) Peihao Chen\\(^{1*}\\) Liangwei Chen\\(^{1}\\) Changhao Li\\(^{1}\\) Thomas H. Li\\(^{6}\\) Mingkui Tan\\(^{157\\dagger}\\) Chuang Gan\\(^{34}\\)" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.218, + 0.842, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)South China University of Technology, \\(^{2}\\)Information Technology R&D Innovation Center of Peking University \\(^{3}\\)UMass Amherst, \\(^{4}\\)MIT-IBM Watson AI Lab, \\(^{5}\\)Key Laboratory of Big Data and Intelligent Robot, Ministry of Education \\(^{6}\\)Peking University Shenzhen Graduate School, \\(^{7}\\)Pazhou Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.473, + 0.715 + ], + "angle": 0, + "content": "How to learn discriminative video representation from unlabeled videos is challenging but crucial for video analysis. The latest attempts seek to learn a representation model by predicting the appearance contents in the masked regions. However, simply masking and recovering appearance contents may not be sufficient to model temporal clues as the appearance contents can be easily reconstructed from a single frame. To overcome this limitation, we present Masked Motion Encoding (MME), a new pretraining paradigm that reconstructs both appearance and motion information to explore temporal clues. In MME, we focus on addressing two critical challenges to improve the representation performance: 1) how to well represent the possible long-term motion across multiple frames; and 2) how to obtain fine-grained temporal clues from sparsely sampled videos. Motivated by the fact that human is able to recognize an action by tracking objects' position changes and shape changes, we propose to reconstruct a motion trajectory that represents these two kinds of change in the masked regions. Besides, given the sparse video input, we enforce the model to reconstruct dense motion trajectories in both spatial and temporal dimensions. Pre-trained with our MME paradigm, the model is able to anticipate long-term and fine-grained motion details. Code is available at https://github.com/XinyuSun/MME." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.746, + 0.21, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Video representation learning plays a critical role in video analysis like action recognition [15,32,79], action localization [12,81], video retrieval [4,82], videoQA [40], etc. Learning video representation is very difficult for two reasons. Firstly, it is extremely difficult and labor-intensive to annotate videos, and thus relying on annotated data to learn" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.892, + 0.382 + ], + "angle": 0, + "content": "video representations is not scalable. Also, the complex spatial-temporal contents with a large data volume are difficult to be represented simultaneously. How to perform self-supervised videos representation learning only using unlabeled videos has been a prominent research topic [7,13,49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.893, + 0.761 + ], + "angle": 0, + "content": "Taking advantage of spatial-temporal modeling using a flexible attention mechanism, vision transformers [3, 8, 25, 26, 53] have shown their superiority in representing video. Prior works [5, 37, 84] have successfully introduced the mask-and-predict scheme in NLP [9, 23] to pre-train an image transformer. These methods vary in different reconstruction objectives, including raw RGB pixels [37], handcrafted local patterns [75], and VQ-VAE embedding [5], all above are static appearance information in images. Based on previous successes, some researchers [64, 72, 75] attempt to extend this scheme to the video domain, where they mask 3D video regions and reconstruct appearance information. However, these methods suffers from two limitations. First, as the appearance information can be well reconstructed in a single image with an extremely high masking ratio (85% in MAE [37]), it is also feasible to be reconstructed in the tube-masked video frame-by-frame and neglect to learn important temporal clues. This can be proved by our ablation study (cf. Section 4.2.1). Second, existing works [64, 75] often sample frames sparsely with a fixed stride, and then mask some regions in these sampled frames. The reconstruction objectives only contain information in the sparsely sampled frames, and thus are hard to provide supervision signals for learning fine-grained motion details, which is critical to distinguish different actions [3, 8]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we aim to design a new mask-and-predict paradigm to tackle these two issues. Fig. 1(a) shows two key factors to model an action, i.e., position change and shape change. By observing the position change of the person, we realize he is jumping in the air, and by observing the shape change that his head falls back and then tucks to his chest, we are aware that he is adjusting his posture to cross the bar. We believe that anticipating these changes helps the model better understand an action." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.444, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution. Email: {csxinyusun, phchencs}@gmail.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.888, + 0.396, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author. Email: mingkuitan@scut.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.444, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2235" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.425, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.206, + 0.411, + 0.22 + ], + "angle": 0, + "content": "(a) Two key factors to recognize a high jump action." + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.09, + 0.872, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.206, + 0.851, + 0.22 + ], + "angle": 0, + "content": "(b) Appearance reconstruction vs. motion trajectory reconstruction." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Figure 1. Illustration of motion trajectory reconstruction for Masked Motion Encoding. (a) Position change and shape change over time are two key factors to recognize an action, we leverage them to represent the motion trajectory. (b) Compared with the current appearance reconstruction task, our motion trajectory reconstruction takes into account both appearance and motion information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Based on this observation, instead of predicting the appearance contents, we propose to predict motion trajectory, which represents impending position and shape changes, for the mask-and-predict task. Specifically, we use a dense grid to sample points as different object parts, and then track these points using optical flow in adjacent frames to generate trajectories, as shown in Fig. 1(b). The motion trajectory contains information in two aspects: the position features that describe relative movement; and the shape features that describe shape changes of the tracked object along the trajectory. To predict this motion trajectory, the model has to learn to reason the semantics of masked objects based on the visible patches, and then learn the correlation of objects among different frames and try to estimate their accurate motions. We name the proposed mask-and-predict task as Masked Motion Encoding (MME)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Moreover, to help the model learn fine-grained motion details, we further propose to interpolate the motion trajectory. Taking sparsely sampled video as input, the model is asked to reconstruct spatially and temporally dense motion trajectories. This is inspired by the video frame interpolation task [77] where a deep model can reconstruct dense video at the pixel level from sparse video input. Different from it, we aim to reconstruct the fine-grained motion details of moving objects, which has higher-level motion information and is helpful for understanding actions. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.72, + 0.47, + 0.811 + ], + "angle": 0, + "content": "- Existing mask-and-predict task based on appearance reconstruction is hard to learn important temporal clues, which is critical for representing video content. Our Masked Motion Encoding (MME) paradigm overcomes this limitation by asking the model to reconstruct motion trajectory." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Our motion interpolation scheme takes a sparsely sampled video as input and then predicts dense motion trajectory in both spatial and temporal dimensions. This scheme endows the model to capture long-term and fine-grained motion clues from sparse video input." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Extensive experimental results on multiple standard video recognition benchmarks prove that the representations learned from the proposed mask-and-predict task achieve state-of-the-art performance on downstream action recognition tasks. Specifically, pre-trained on Kinetics-400 [10], our MME brings the gain of \\(2.3\\%\\) on Something-Something V2 [34], \\(0.9\\%\\) on Kinetics-400, \\(0.4\\%\\) on UCF101 [59], and \\(4.7\\%\\) on HMDB51 [44]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.642, + 0.451 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Self-supervised Video Representation Learning. Self-supervised video representation learning aims to learn discriminative video features for various downstream tasks in the absence of accurate video labels. To this end, most of the existing methods try to design an advanced pretext task like predicting the temporal order of shuffled video crops [78], perceiving the video speediness [7, 13] or solving puzzles [43, 49]. In addition, contrastive learning is also widely used in this domain [14, 16, 36, 41, 46, 54, 57, 68, 69], which constrains the consistency between different augmentation views and brings significant improvement. In particular, ASCNet and CoCLR [36, 41] focus on mining hard positive samples in different perspectives. Optical flow has also been proven to be effective in capturing motion information [70, 76]. Besides, tracking video objects' movement is also used in self-supervised learning [17, 18, 65, 73, 74]. Among them, Wang et al. [74] only utilizes the spatial encoder to extract frame appearance information. CtP framework [65] and Siamese-triplet network [73] only require the model to figure out the position and size changes of a specific video patch. Different from these methods, our proposed MME trace the fine-grained movement and shape changes of different parts of objects in the video, hence resulting in a superior video representation. Tokmakov et al. [63] utilize Dense Trajectory to provide initial pseudo labels for video clustering. However, the model does not predict trajectory motion features explicitly. Instead, we consider long-term and fine-grained trajectory motion features as explicit reconstruction targets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2236" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.342, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.089, + 0.512, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.089, + 0.891, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.894, + 0.302 + ], + "angle": 0, + "content": "Figure 2. Overview of Masked Motion Encoding (MME). Given a sparsely sampled video, we first divide it into several patches and randomly mask out some of them. And then, we feed the remaining patches to a ViT encoder to extract video representation. Last, a lightweight ViT decoder is involved to predict the content in the masked region, i.e., a motion trajectory containing position changes and shape changes of moving objects." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.313, + 0.473, + 0.678 + ], + "angle": 0, + "content": "Mask Modeling for Vision Transformer. Recently, BEiT and MAE [5, 37] show two excellent mask image modeling paradigms. BEVT [72] and VideoMAE [64] extend these two paradigms to the video domain. To learn visual representations, BEVT [72] predicts the discrete tokens generated by a pre-trained VQ-VAE tokenizer [56]. Nevertheless, pre-training such a tokenizer involves an unbearable amount of data and computation. In contrast, VideoMAE [64] pretrain the Vision Transformer by regressing the RGB pixels located in the masked tubes of videos. Due to asymmetric encoder-decoder architecture and extremely high masking ratio, model pre-training is more efficient with VideoMAE. Besides, MaskFeat [75] finds that predicting the Histogram of Gradient (HOG [21]) of masked video contents is a strong objective for the mask-and-predict paradigm. Existing methods only consider static information in each video frame, thus the model can speculate the masked area by watching the visible area in each frame independently and failed to learn important temporal clues (cf. Section 3.1). Different from the video prediction methods [35, 51, 58, 60] that predict the future frames in pixel or latent space, our Masked Motion Encoding paradigm predicts incoming fine-grained motion in masked video regions, including position changes and shape changes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.69, + 0.251, + 0.708 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.793 + ], + "angle": 0, + "content": "We first revisit the current masked video modeling task for video representation learning (cf. Section 3.1). Then, we introduce our masked motion encoding (MME), where we change the task from recovering appearance to recovering motion trajectory (cf. Section 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.394, + 0.819 + ], + "angle": 0, + "content": "3.1. Rethinking Masked Video Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Given a video clip sampled from a video, self-supervised video representation learning aims to learn a feature encoder \\( f_{enc}(\\cdot) \\) that maps the clip to its corresponding feature that best describes the video. Existing masked video modeling methods [64, 75] attempt to learn such a feature" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.435 + ], + "angle": 0, + "content": "encoder through a mask-and-predict task. Specifically, the input clip is first divided into multiple non-overlapped 3D patches. Some of these patches are randomly masked and the remaining patches are fed into the feature encoder, followed by a decoder \\( f_{dec}(\\cdot) \\) to reconstruct the information in the masked patches. Different works aim to reconstruct different information (e.g., raw pixel in VideoMAE [64] and HOG in MaskFeat [75])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.893, + 0.603 + ], + "angle": 0, + "content": "However, the existing works share a common characteristic where they all attempt to recover static appearance information of the masked patches. Since an image with a high masking ratio (85% in MAE [37]) can be well reconstructed [37,75], we conjecture that the masked appearance information of a video can also be reconstructed frame by frame independently. In this sense, the model may focus more on the contents in the same frame. This may hinder the models from learning important temporal clues, which is critical for video representation. We empirically study this conjecture in the ablation study (cf. Section 4.2.1)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.61, + 0.733, + 0.625 + ], + "angle": 0, + "content": "3.2. General Scheme of MME" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.861 + ], + "angle": 0, + "content": "To better learn temporal clues of a video, our MME changes the reconstruction content from static appearance information to object motion information, including position and shape change of objects. As shown in Fig. 2, a video clip is sparsely sampled from a video and is divided into a number of non-overlapped 3D patches of size \\( t \\times h \\times w \\), corresponding to time, height, and width. We follow VideoMAE [64] to use the tube masking strategy, where the masking map is the same for all frames, to mask a subset of patches. For computation efficiency, we follow MAE [37] to only feed the unmasked patches (and their positions) to the encoder. The output representation together with learnable [MASK] tokens are fed to a decoder to reconstruct motion trajectory \\( \\mathbf{z} \\) in the masked patches. The training loss for MME is" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.872, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i \\in \\mathcal {I}} \\left| \\mathbf {z} _ {i} - \\hat {\\mathbf {z}} _ {i} \\right| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2237" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{z}}\\) is the predicted motion trajectory, and \\(\\mathcal{I}\\) is the index set of motion trajectories in all masked patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.471, + 0.304 + ], + "angle": 0, + "content": "Motivated by the fact that we humans recognize actions by perceiving position changes and shape changes of moving objects, we leverage these two types of information to represent the motion trajectory. Through pre-training on MME task, the model is endowed with the ability to explore important temporal clues. Another important characteristic of the motion trajectory is that it contains fine-grained motion information extracted at raw video rate. This fine-grained motion information provides the model with a supervision signal to anticipate fine-grained action from sparse video input. In the following, we will introduce the proposed motion trajectory in detail." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.315, + 0.334, + 0.331 + ], + "angle": 0, + "content": "3.3. Motion Trajectory for MME" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.491 + ], + "angle": 0, + "content": "The motion of moving objects can be represented in various ways such as optical flow [27], histograms of optical flow (HOF) [45], and motion boundary histograms (MBH) [45]. However, these descriptors can only represent short-term motion between two adjacent frames. We hope our motion trajectory represents long-term motion, which is critical for video representation. To this end, inspired by DT [66], we first track the moving object in the following \\(L\\) frames to cover a longer range of motion, resulting in a trajectory \\(\\mathbf{T}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.504, + 0.47, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} = \\left(\\mathbf {p} _ {t}, \\mathbf {p} _ {t + 1}, \\dots , \\mathbf {p} _ {t + L}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.533, + 0.469, + 0.608 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_t = (x_t, y_t)\\) represents a point located at \\((x_t, y_t)\\) of frame \\(t\\), and \\((\\cdot, \\cdot)\\) indicates the concatenation operation. Along this trajectory, we fetch the position features \\(\\mathbf{z}^p\\) and shape features \\(\\mathbf{z}^s\\) of this object to compose a motion trajectory \\(\\mathbf{z}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.61, + 0.469, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\left(\\mathbf {z} ^ {p}, \\mathbf {z} ^ {s}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.635, + 0.469, + 0.696 + ], + "angle": 0, + "content": "The position features are represented by the position transition relative to the last time step, while the shape features are the HOG descriptors of the tracked object in different time steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Tracking objects using spatially and temporally dense trajectories. Some previous works [2, 18, 19, 24] try to use one trajectory to represent the motion of an individual object. In contrast, DT [66] points out that tracking spatially dense feature points sampled on a grid space performs better since it ensures better coverage of different objects in a video. Following DT [66], we use spatially dense grid points as the initial position of each trajectory. Specifically, we uniformly sample \\(K\\) points in a masked patch of size \\(t \\times h \\times w\\), where each point indicates a part of an object. For each point, we track it through temporally dense \\(L\\) frames according to the dense optical flow, resulting in \\(K\\) trajectories. In this way, the model is able to capture spatially and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "temporally dense motion information of objects through the mask-and-predict task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.319 + ], + "angle": 0, + "content": "As a comparison, the reconstruction contents in existing works [64, 75] often extract temporally sparse videos sampled with a large stride \\( s > 1 \\). The model takes as input a sparse video and predicts these sparse contents for learning video representation. Different from these works, our model also takes as input sparse video but we push the model to interpolate motion trajectory containing fine-grained motion information. This simple trajectory interpolation task does not increase the computational cost of the video encoder but helps the model learn more fine-grained action information even given sparse video as input. More details about dense flow calculating and trajectory tracking can be found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.419 + ], + "angle": 0, + "content": "Representing position features. Given a trajectory \\(\\mathbf{T}\\) consisting of the tracked object position at each frame, we are more interested in the related movement of objects instead of their absolute location. Consequently, we represent the position features with related movement between two adjacent points \\(\\Delta \\mathbf{p}_t = \\mathbf{p}_{t+1} - \\mathbf{p}_t\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.43, + 0.892, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {p} = \\left(\\Delta \\mathbf {p} _ {t}, \\dots , \\Delta \\mathbf {p} _ {t + L - 1}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.504 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}^p\\) is a \\(L\\times 2\\) dimensional feature. As each patch contains \\(K\\) position features, we concatenate and normalize them as position features part of the motion trajectory." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Representing shape features. Besides embedding the movement, the model also needs to be aware of the shape changes of objects to recognize actions. Inspired by HOG [21], we use histograms of oriented gradients (HOG) with 9 bins to describe the shape of objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Compared with existing works [64, 75] that reconstruct HOG in every single frame, we are more interested in the dynamic shape changes of an object, which can better represent action in a video. To this end, we follow DT [66] to calculate trajectory-aligned HOG, consisting of HOG features around all tracked points in a trajectory, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.689, + 0.892, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {s} = \\left(\\operatorname {H O G} \\left(\\mathbf {p} _ {t}\\right), \\dots , \\operatorname {H O G} \\left(\\mathbf {p} _ {t + L - 1}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.793 + ], + "angle": 0, + "content": "where \\(\\mathrm{HOG}(\\cdot)\\) is the HOG descriptor and \\(\\mathbf{z}^s\\) is a \\(L\\times 9\\) dimensional feature. Also, as one patch contains \\(K\\) trajectories, we concatenate \\(K\\) trajectory-aligned HOG features and normalize them to the standard normal distribution as the shape features part of the motion trajectory." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.806, + 0.634, + 0.824 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We conduct experiments on Kinetics-400 (K400), Something-Something V2 (SSV2), UCF101 and HMDB51 datasets. Unless otherwise stated, we follow previous trails [64] and feed the model a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2238" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.852, + 0.335 + ], + "angle": 0, + "content": "
MethodBackbonePre-trainLabelEpochFramesGFLOPsParam.Acc.@1
TimeSformer [8]ViT-BIN-21K✓158196×1×3121M59.5
ViViT-L [3]ViT-LIN-21K+K400✓35163992 × 3 × 4311M65.4
Motionformer [53]ViT-BIN-21K+K400✓3516370 × 1 × 3109M66.5
SlowFast [29]ResNet-101K400✓1968+32106 × 1 × 353M63.1
MViT [26]MViT-BK400✓10064455 × 1 × 337M67.7
SVT [57]ViT-BIN-21K+K400×208+64196×1×3121M59.6
LSTCL [68]Swin-BK400×20016360×5×188M67.0
BEVT [72]Swin-BK400+DALL-E×15032282×1×388M67.1
VIMPAC [62]ViT-LHowTo100M+DALLE×10010N/A×10×3307M68.1
VideoMAE [64]ViT-BSSv2×40016180×2×387M67.6
VideoMAE [64]ViT-BSSv2×80016180×2×387M69.3
VideoMAE [64]ViT-BK400×80016180×2×387M68.2
MME (Ours)ViT-BSSv2×40016180×2×387M69.2
MME (Ours)ViT-BSSv2×80016180×2×387M70.0
MME (Ours)ViT-BK400×40016180×2×387M69.3
MME (Ours)ViT-BK400×80016180×2×387M70.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.347, + 0.892, + 0.375 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-arts on Something-Something V2. Our MME outperforms the previous best method. Remarkably, MME show better transferability ability (transfer from K400 to SSv2) compared with VideoMAE." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.388, + 0.851, + 0.588 + ], + "angle": 0, + "content": "
MethodBackbonePre-trainLabelEpochFramesGFLOPsParam.Acc.@1
TimeSformer [8]ViT-BIN-21K✓158196×1×3121M78.3
Motionformer [53]ViT-BIN-21K✓3516370×10×3121M79.7
Motionformer [53]ViT-HRIN-21K✓3516959×10×3382M81.1
VIMPAC [62]ViT-LHowTo100M+DALLE×10010N/A×10×3307M77.4
SVT [57]ViT-BIN-21K+K400×208+64196×1×3121M78.1
BEVT [72]Swin-BK400+DALL-E×15032282×1×388M76.2
BEVT [72]Swin-BIN21K+K400+DALL-E×15032282×1×388M80.6
LSTCL [68]Swin-BK400×20016360×5×188M81.5
OmniMAE [33]ViT-BIN1K+K400×160032180×5×387M80.6
VideoMAE [64]ViT-BK400×160016180×5×387M80.9
ST-MAE [28]ViT-BK400×160016180×7×387M81.3
MME (Ours)ViT-BK400×160016180×7×387M81.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.15, + 0.598, + 0.819, + 0.612 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-arts on Kinetics-400. Our MME outperforms the previous best method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.471, + 0.768 + ], + "angle": 0, + "content": "\\(224 \\times 224\\) 16-frame clip with strides 4 and 3 for K400 and SSV2 respectively. We use vanilla ViT-Base [25] as the backbone in all experiments. As for the motion trajectory, we set the trajectory length to \\(L = 6\\) and the number of trajectories in each patch to \\(K = 4\\). Our models are trained on 16 NVIDIA GeForce RTX 3090 GPUs. Note that for all the ablation studies, we use a smaller pre-training dataset and reduce the input patches of fine-tuning to speed up the experiments, details can be found in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.383, + 0.801 + ], + "angle": 0, + "content": "4.1. Comparisons with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Fine-tuning. On the temporal-heavy dataset SSv2, we conduct experiments using two settings, namely in-domain pre-training (i.e., pre-train and fine-tune on the same dataset) and cross-domain pre-training (i.e., pre-train and fine-tune on different datasets). In Tab. 1, we have two main observations. 1) Our MME significantly performs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.858 + ], + "angle": 0, + "content": "better than previous state-of-the-arts VideoMAE [64] on both in-domain and cross-domain settings. Specifically, under cross-domain settings (transferring from K400 to SSv2) with 800 epochs, our MME outperforms it by \\(2.3\\%\\), increasing the accuracy from \\(68.2\\%\\) to \\(70.5\\%\\). 2) Our MME performs better under cross-domain setting compared with indomain setting, which is contrary to the existing observation on VideoMAE [64] and ST-MAE [28]. We attribute it to our masked motion encoding pretext task that helps models focus more on learning temporal clues instead of learning domain-specific appearance information (e.g., color, lighting). The learned temporal clues are more universal across different domains for video representation. This also shows the potential for taking advantage of larger datasets for pretraining, which we leave for future work." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Tab. 2, our MME also achieves state-of-the-art performance on another large-scale dataset K400 with a gain" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2239" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.472, + 0.214 + ], + "angle": 0, + "content": "
MethodBackbonePre-trainUCF101HMDB51
XDC [1]R(2+1)DIG65M94.267.1
GDT [52]R(2+1)DIG65M95.272.8
CVRL [54]R50K40092.967.9
CORPf [39]R50K40093.568.0
ρBYOL [30]R50K40094.272.1
VideoMAE [64]ViT-BK40096.173.3
MME (Ours)ViT-BK40096.578.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.223, + 0.47, + 0.25 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-arts on UCF101 and HMDB51." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.47, + 0.375 + ], + "angle": 0, + "content": "of \\(0.5\\%\\). In Tab. 3, we also evaluate the transferability of MME by transferring a K400 pre-trained model to two small-scale datasets, namely UCF101 and HMDB51. Our MME achieves a large gain of \\(4.7\\%\\) on HMDB51 and \\(0.4\\%\\) on UCF101 compared with VideoMAE. This further demonstrates the strong ability of MME to learn from the large-scale dataset and transfer it to a small-scale one." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.519 + ], + "angle": 0, + "content": "Linear probing. We further evaluate our MME under liner probing setting on the SSv2 dataset as this dataset focuses more on the action. We follow SVT [57] to fix the transformer backbone and train a linear layer for 20 epochs. In Tab. 4, our method significantly outperform previous contrastive-based method [57] by \\(10.9\\%\\) and masked video modeling methods [64] by \\(4.9\\%\\). These results indicate that the representation learned from our MME task contains more motion clues for action recognition." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.442, + 0.546 + ], + "angle": 0, + "content": "4.2. Ablation Study on Reconstruction Content" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.554, + 0.47, + 0.6 + ], + "angle": 0, + "content": "In this section, we investigate the effect of different reconstruction contents, including appearance, short-term motion, and the proposed long-term fine-grained motion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.406, + 0.635 + ], + "angle": 0, + "content": "4.2.1 Evaluating Appearance Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.854 + ], + "angle": 0, + "content": "In section 3.1, we conjecture that appearance reconstruction is finished in each video frame independently. To verify this conjecture, we corrupt the temporal information of a video by randomly shuffling its frames. Since the video frames are randomly permuted, it is hard for a model to leverage information from other frames to reconstruct the current frame through their temporal correlation. We mask \\(90\\%\\) regions of these videos following VideoMAE [64]. In Fig. 3 (a), we find that the reconstruction error (i.e., L2 error between predicted and ground-truth pixel) converges to a low value. We also conduct the same experiment using raw video without shuffling. The model undergoes a similar convergence process. This demonstrates that the masked pixels are well reconstructed without temporal information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To further evaluate whether the model pre-trained on the appearance reconstruction task (i.e., VideoMAE) can well capture important temporal clues, we transfer two" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.089, + 0.885, + 0.182 + ], + "angle": 0, + "content": "
MethodBackbonePre-trainAcc.@1
TimeSformer [8]ViT-BIN-21K14.0
SVT [57]ViT-BIN-21K+K40018.3
VideoMAE [64]ViT-BSSv224.3
MME (Ours)ViT-BSSv229.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.192, + 0.866, + 0.206 + ], + "angle": 0, + "content": "Table 4. Linear Probing on Something-Something V2." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.214, + 0.693, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.356, + 0.704, + 0.368 + ], + "angle": 0, + "content": "(a) Masked video reconstruction error." + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.215, + 0.888, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.356, + 0.875, + 0.369 + ], + "angle": 0, + "content": "(b) Downstream performance." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.38, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 3. We perform mask-and-predict pre-training using temporal shuffled videos and raw videos and then transfer them to the downstream task. (a) Masked appearance (i.e., pixels) is well reconstructed without temporal information. (b) Our MME outperforms VideoMAE when both provided temporal information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.893, + 0.699 + ], + "angle": 0, + "content": "VideoMAE models (pre-trained on shuffled videos and raw videos, respectively) to the downstream action recognition task. In Fig. 3 (b), these two models perform competitively, indicating that removing the temporal information from the pre-training videos barely affects the learned video representation. We speculate this is because the VideoMAE model pays little attention to temporal information when performing the appearance reconstruction pre-training task. As a comparison, our MME paradigm performs significantly better than VideoMAE when the temporal information is provided for pre-training \\((64.1\\%\\) vs. \\(60.9\\%)\\). Once the temporal information is removed, MME performs similarly with VideoMAE \\((60.7\\%\\) vs. \\(60.9)\\). This demonstrates our MME paradigm takes great advantage of temporal information in the pre-training phase." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.718, + 0.796, + 0.733 + ], + "angle": 0, + "content": "4.2.2 Evaluating Motion Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.742, + 0.892, + 0.772 + ], + "angle": 0, + "content": "In the following ablations, we separately verify the effectiveness of two proposals of our methods:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.782, + 0.892, + 0.826 + ], + "angle": 0, + "content": "Effectiveness of motion trajectory. We compare our motion trajectory with two appearance reconstruction baselines and three short-term motion reconstruction baselines." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "- Appearance reconstruction baselines: 1) HOG: predicting HOG features [21] in the middle frame of each 3D patch [75]. 2) Pixel: predicting all the pixels of each 3D patch [28, 37, 64]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2240" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.125, + 0.089, + 0.424, + 0.228 + ], + "angle": 0, + "content": "
Reconstruction contentAcc.@1Acc.@5
Pixel61.186.6
HOG61.186.9
HOG + Flow60.986.4
HOG + HOF61.386.6
HOG + MBH61.486.8
HOG + Traj. w/o shape63.588.2
Traj. (ours)64.188.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.239, + 0.47, + 0.268 + ], + "angle": 0, + "content": "Table 5. Comparison of different Reconstruction content. Our motion trajectory outperforms all the other contents." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.276, + 0.47, + 0.403 + ], + "angle": 0, + "content": "
Reconstruction contentInterpolationAcc.@1Acc.@5
HOGx61.186.9
HOG✓60.986.8
HOG + MBHx61.386.8
HOG + MBH✓61.486.8
Traj. (ours)x62.287.5
Traj. (ours)✓64.188.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.413, + 0.469, + 0.455 + ], + "angle": 0, + "content": "Table 6. Effectiveness of trajectory interpolation. Interpolating the motion trajectory brings significant improvement compared to the baselines." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.477, + 0.47, + 0.568 + ], + "angle": 0, + "content": "- Short-term motion reconstruction baselines: 1) Flow: predicting the dense optical flow map of the masked patches. 2) HOF: predicting the histogram of optical flow orientation. 3) MBH: predicting the motion boundary histogram, which is the second-order local statistic of optical flow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Tab. 5, two types of appearance reconstruction contents (i.e., Pixel and HOG) perform similarly. Based on the HOG baseline, predicting additional short-term motion contents brings slight improvement (at least by \\(0.2\\%\\)) except for directly predicting optical flow (decreases by \\(0.2\\%\\)). We suspect that predicting the dense flow map for each pixel is too hard for a model and thus harms the performance. Instead, the histogram version of flow (i.e., HOF and MBH) help to reduce the prediction ambiguity [75], resulting in better performance. Furthermore, after adding our slight variant of motion trajectory (i.e., with position features but without shape features shown in Eq. (3)), the performance improves significantly compared to both appearance reconstruction baselines (by \\(2.4\\%\\)) and short-term motion reconstruction baselines (by \\(2.1\\%\\)). On top of this variant, adding trajectory-aligned shape features further brings additional gain by \\(0.6\\%\\). Because tracking the trajectory of objects in multiple frames provides long-term motion compared to short-term motion reconstruction, we believe the learning of long-term motion clues is one of the keys to enhancing video representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Effectiveness of trajectory interpolation. In MME, we feed temporally sparse video to the model and push it to interpolate temporally dense trajectories. We are interested in whether interpolating appearance (i.e., HOG) or short-term motion contents (i.e., MBH) benefit learning fine-grained motion details. To this end, we conduct experiments to predict dense HOG and MBH, i.e., predicting these features of all frames within a masked patch even though these frames are not sampled as input. In Tab. 6, interpolating HOG or MBH brings little improvement. We suspect this is because these dense features represent the continuous movement of objects poorly, as they have not explicitly tracked the moving objects. In contrast, our temporally dense motion trajectory tracks the object densely, providing temporal fine-grained information for learning. The trajectory interpolation improves our MME by \\(1.9\\%\\). This result proves that the learning of fine-grained motion details is another key to enhancing video representation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.373, + 0.783, + 0.389 + ], + "angle": 0, + "content": "4.3. Ablation Study on MME Design" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.396, + 0.892, + 0.427 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to investigate the effect of different parameters for MME pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Trajectory length. Trajectory length \\( L \\) indicates how many frames we track an object for extracting the motion trajectory. If \\( L = 0 \\), our method degrades to MaskFeat [75], which only predicts HOG features in the masked patch. In Tab. 7a, MME achieves the best performance when \\( L = 6 \\). This is because a longer trajectory provides more motion information for learning. But if the trajectory is too long, the drift of flow introduces accumulated noise when tracking the objects, which harms the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Trajectory density. In our experiments, we spatially dense track \\( K = 4 \\) motion trajectories. We conduct experiments to evaluate whether it is necessary to densely track multiple trajectories in a patch. Specifically, we assume only one object exists in a patch, and thus the 4 trajectories are similar. We merge these 4 trajectories as one by averaging or selecting the most salient one. We then consider the merged trajectories as the trajectory motion. In Tab. 7c, using dense trajectories significantly outperform the merged baselines. We suspect the spatial dense trajectories provide more motion information for learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Trajectory normalization. Due to the irregular distribution of motion and shape information in the video, the trajectory values in different patches vary a lot. To help the model better model the trajectory, we adopt the patch norm method to normalize the trajectory values in each patch into a standard normal distribution, as has also been done in MAE [37]. The mean and variance are calculated by all \\( K = 4 \\) trajectories in a patch. Tab. 7b shows the effectiveness of adopting the patch norm." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2241" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.289, + 0.161 + ], + "angle": 0, + "content": "
LengthAcc.@1Acc.@5
L = 463.288.0
L = 664.188.4
L = 863.388.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.164, + 0.283, + 0.178 + ], + "angle": 0, + "content": "(a) Effect of trajectory length." + }, + { + "type": "table", + "bbox": [ + 0.316, + 0.089, + 0.571, + 0.159 + ], + "angle": 0, + "content": "
SettingAcc.@1Acc.@5
w/o patch norm.62.187.44
w/ patch norm.64.188.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.321, + 0.162, + 0.557, + 0.177 + ], + "angle": 0, + "content": "(b) Effect of trajectory normalization." + }, + { + "type": "table", + "bbox": [ + 0.598, + 0.089, + 0.865, + 0.161 + ], + "angle": 0, + "content": "
Trajectory densityAcc.@1Acc.@5
Dense64.188.4
Sparse (Max)60.286.3
Sparse (Mean)61.887.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.631, + 0.164, + 0.833, + 0.179 + ], + "angle": 0, + "content": "(c) Effect of trajectories density." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.194, + 0.286, + 0.28 + ], + "angle": 0, + "content": "
StrideAcc.@1Acc.@5
s = 462.987.7
s = 364.188.4
s = 263.588.3
s = 162.787.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.285, + 0.279, + 0.299 + ], + "angle": 0, + "content": "(d) Effect of sampling stride." + }, + { + "type": "table", + "bbox": [ + 0.319, + 0.194, + 0.56, + 0.28 + ], + "angle": 0, + "content": "
RatioTypeAcc.@1Acc.@5
90%Tube62.787.8
70%Tube64.188.4
40%Tube61.386.9
40%Cube61.286.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.351, + 0.285, + 0.53, + 0.3 + ], + "angle": 0, + "content": "(e) Effect of spatial masking." + }, + { + "type": "table", + "bbox": [ + 0.582, + 0.19, + 0.888, + 0.282 + ], + "angle": 0, + "content": "
TypeDepthFLOPsAcc.@1Acc.@5
Joint151G64.188.4
255G63.488.3
Sep.156G63.488.4
265G63.388.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.592, + 0.285, + 0.879, + 0.299 + ], + "angle": 0, + "content": "(f) Effect of decoder. Sep. means two decoders." + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.311, + 0.856, + 0.327 + ], + "angle": 0, + "content": "Table 7. Pre-training details. We explore the detailed setting of our MME pre-training. The default entry is marked in gray ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.47, + 0.483 + ], + "angle": 0, + "content": "Sampling stride. In order to control the difficulty of interpolating the motion trajectory, we ablate different temporal sampling strides leading to different levels of the sparse input video. With a medium sampling stride (\\(s = 2\\)), our MME achieves the best performance. We also found that the performance decreases when the input video shares the same stride as the dense motion trajectory (\\(s = 1\\)). One possible reason is that it is difficult for the model to anticipate fine-grained motion details in this case." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.6 + ], + "angle": 0, + "content": "Spatial masking strategy. We ablated study different spatial masking strategies and ratios in Tab. 7e. We start from the default setting of MaskFeat [75] that uses a cube masking strategy with a \\(40\\%\\) masking ratio, then we increase the masking ratio and adopt the tube masking strategy proposed by VideoMAE [64]. Using tube masking with a medium masking ratio \\((70\\%)\\) performs the best." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.73 + ], + "angle": 0, + "content": "Decoder setting. We use several transformer blocks as decoder to reconstruct motion trajectory. We ablate different architectures of the decoder in Tab. 7f. Using a slight decoder with 1 transformer block performs the best, with the lowest computational cost. Increasing the depth of the decoder or using two parallel decoders to separately predict the position and shape features introduces more computational cost but brings little improvement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.275, + 0.756 + ], + "angle": 0, + "content": "4.4. Visualization Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We visualize the predicted motion trajectory in Fig. 4. The model is able to reconstruct the spatial and temporally dense trajectories of different object parts from a temporally sparse input video. Even without seeing the object in the current frame (the hand being masked out), the model is able to locate it through contextual patches and accurately estimate its trajectory. We only visualize the position changes since our prediction of HOG features has similar quality to the previous work [75]." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.345, + 0.877, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.494, + 0.893, + 0.537 + ], + "angle": 0, + "content": "Figure 4. Visualization of predicted motion trajectory. Each column represents (from left to right): (a) Masked frames; (b) ground truth; (c) prediction; (d) original frame." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.552, + 0.619, + 0.567 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.747 + ], + "angle": 0, + "content": "We present Masked Motion Encoding, a simple mask-and-predict paradigm that learns video representation by reconstructing the motion contents of masked regions. In particular, we reconstruct motion trajectory, which tracks moving objects densely to record their position and shape changes. By reconstructing this motion trajectory, the model learns long-term and fine-grained motion clues from sparse input videos. Empirical results show that our method outperforms the state-of-the-art mask-and-predict methods on Something-Something V2, Kinetics-400, UCF-101, and HMDB-51 datasets for action recognition." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.661, + 0.785 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was partially supported by the National Natural Science Foundation of China (NSFC) (No. 62072190, 62072186), Ministry of Science and Technology Foundation Project (No. 2020AAA0106900), Program for Guangdong Introducing Innovative and Entrepreneurial Teams (No. 2017ZT07X183), Guangdong Basic and Applied Basic Research Foundation (No. 2019B1515130001)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2242" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, pages 9758-9770, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Nadeem Anjum and Andrea Cavallaro. Multifeature object trajectory clustering for video analysis. IEEE Transactions on Circuits and Systems for Video Technology, 18:1555-1564, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: BERT pre-training of image transformers. In ICLR, 2022. 1, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[6] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417, 2006. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.47, + 0.458 + ], + "angle": 0, + "content": "[7] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In CVPR, pages 9922-9931, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding. In ICML, pages 813-824, 2021. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.503, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[11] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703, 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[12] Peihao Chen, Chuang Gan, Guangyao Shen, Wenbing Huang, Runhao Zeng, and Mingkui Tan. Relation attention for temporal action localization. IEEE Transactions on Multimedia, 22:2723-2733, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[13] Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan. Rspnet: Relative speed perception for unsupervised video representation learning. In AAAI, pages 1045-1053, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[14] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas H Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurlPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[15] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 29:8292-8302, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Zhenfang Chen, Lin Ma, Wenhan Luo, Peng Tang, and Kwan-Yee K Wong. Look closer to ground better: Weakly-supervised temporal grounding of sentence in video. arXiv, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In ACL, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[18] Zhenfang Chen, Jiayuan Mao, Jiajun Wu, Kwan-Yee Kenneth Wong, Joshua B Tenenbaum, and Chuang Gan. Grounding physical concepts of objects and events through dynamic visual reasoning. arXiv, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[19] Zhenfang Chen, Kexin Yi, Yunzhu Li, Mingyu Ding, Antonio Torralba, Joshua B Tenenbaum, and Chuang Gan. Comphy: Compositional physical reasoning of objects and events from videos. In ICLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[20] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, pages 702-703, 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[21] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, pages 886-893, 2005. 3, 4, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[22] Navneet Dalal, Bill Triggs, and Cordelia Schmid. Human detection using oriented histograms of flow and appearance. In ECCV, pages 428-441, 2006. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, pages 4171–4186, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[24] Mingyu Ding, Zhenfang Chen, Tao Du, Ping Luo, Josh Tenenbaum, and Chuang Gan. Dynamic visual reasoning by learning differentiable physics models from video and language. In NeurIPS, volume 34, pages 887-899, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[25] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[26] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[27] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In *SCIA*, pages 363-370, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[28] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In NeurIPS, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[29] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[30] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross B. Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, pages 3299-3309, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[31] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2243" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "image analysis and automated cartography. Communications of the ACM, 24:381-395, 1981. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[32] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In European Conference on Computer Vision (ECCV), pages 758-775, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[33] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.239, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[34] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.469, + 0.378 + ], + "angle": 0, + "content": "[35] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martín-Martín, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. arXiv preprint arXiv:2206.11894, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.381, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In NeurIPS, pages 5679-5690, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.425, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[38] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. In CVPR, pages 8126-8135, 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[39] Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. In ICCV, pages 7939-7949, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[40] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware Graph Convolutional Networks for Video Question Answering. In AAAI, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.71 + ], + "angle": 0, + "content": "[41] Deng Huang, Wenhao Wu, Weiwen Hu, Xu Liu, Dongliang He, Zhihua Wu, Xiangmiao Wu, Mingkui Tan, and Errui Ding. Ascnet: Self-supervised video representation learning with appearance-speed consistency. In ICCV, pages 8096-8105, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[42] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q. Weinberger. Deep networks with stochastic depth. In ECCV, pages 646-661, 2016. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[43] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[44] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, pages 2556-2563, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Ivan Laptev, Marcin Marszalek, Cordelia Schmid, and Benjamin Rozenfeld. Learning realistic human actions from movies. In CVPR, pages 1-8, 2008. 4, 13" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[46] Yuan Liu, Jingyuan Chen, Zhenfang Chen, Bing Deng, Jianqiang Huang, and Hanwang Zhang. The blessings of unlabeled background in untrimmed videos. In CVPR, pages 6176-6185, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[47] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.208 + ], + "angle": 0, + "content": "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.211, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[49] Dezhao Luo, Chang Liu, Yu Zhou, Dongbao Yang, Can Ma, Qixiang Ye, and Weiping Wang. Video cloze procedure for self-supervised spatio-temporal learning. In AAAI, pages 11701-11708, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.269, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[50] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035, 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.381 + ], + "angle": 0, + "content": "[51] Viorica Patraucean, Ankur Handa, and Roberto Cipolla. Spatio-temporal video autoencoder with differentiable memory. In ICLR Workshop, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[52] Mandela Patrick, Yuki M Asano, Polina Kuznetsova, Ruth Fong, Joao F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[53] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In NeurIPS, pages 12493-12506, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[54] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, pages 6964-6974, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[55] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Xiang Wang, Yuehuan Wang, Yiliang Lv, Changxin Gao, and Nong Sang. Mar: Masked autoencoders for efficient action recognition. arXiv preprint arXiv:2207.11660, 2022. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[57] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael Ryoo. Self-supervised video transformer. In CVPR, pages 2874-2884, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[58] Adria Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătráucean, Florent Altché, Michal Valko, et al. Broaden your views for self-supervised video learning. In ICCV, pages 1255-1265, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[59] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[60] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, pages 843-852, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2244" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[61] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[62] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[63] Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. Unsupervised learning of video representations via dense trajectory clustering. In ECCV 2020 Workshops, pages 404-421. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[64] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 3, 4, 5, 6, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[65] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. In CVPR, pages 2563-2572, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[66] Heng Wang, Alexander Kläser, Cordelia Schmid, and Cheng-Lin Liu. Dense trajectories and motion boundary descriptors for action recognition. International journal of computer vision, 103:60-79, 2013. 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[67] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In ICCV, pages 3551-3558, 2013. 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.511 + ], + "angle": 0, + "content": "[68] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. arXiv preprint arXiv:2106.09212, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[69] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In CVPR, pages 11804-11813, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[70] Jiangliu Wang, Jianbo Jiao, Linchao Bao, Shengfeng He, Yunhui Liu, and Wei Liu. Self-supervised spatio-temporal representation learning for videos by predicting motion and appearance statistics. In CVPR, pages 4006-4015, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[71] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41:2740-2755, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.47, + 0.762 + ], + "angle": 0, + "content": "[72] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, pages 14733-14743, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[73] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, pages 2794-2802, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[74] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, pages 2566-2576, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[75] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 1, 3, 4, 6, 7, 8, 13" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[76] Donglai Wei, Joseph J. Lim, Andrew Zisserman, and William T. Freeman. Learning and using the arrow of time. In CVPR, pages 8052-8060, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[77] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P Allebach, and Chenliang Xu. Zooming slow-mo: Fast and accurate one-stage space-time video super-resolution. In CVPR, pages 3370-3379, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[78] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In CVPR, pages 10334–10343, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[79] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[80] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[81] Runhao Zeng, Chuang Gan, Peihao Chen, Wenbing Huang, Qingyao Wu, and Mingkui Tan. Breaking winner-takes-all: Iterative-winners-out networks for weakly supervised temporal action localization. IEEE Transactions on Image Processing, 28(12):5797-5808, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[82] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[83] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[84] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan L. Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2245" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_origin.pdf b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e6228ccbe7cce6fa480880723f6c986563c96802 --- /dev/null +++ b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/ffb96eff-92ef-423e-9c85-df5684067148_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:713199410b51c399cbe55485e0a39878a01d6dace1f09e137f9d5cf2117b1552 +size 1074108 diff --git a/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/full.md b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..da612f4013743d40fb6d1a067bea681de57f6d79 --- /dev/null +++ b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/full.md @@ -0,0 +1,337 @@ +# Masked Motion Encoding for Self-Supervised Video Representation Learning + +Xinyu Sun $^{12*}$ Peihao Chen $^{1*}$ Liangwei Chen $^{1}$ Changhao Li $^{1}$ Thomas H. Li $^{6}$ Mingkui Tan $^{157\dagger}$ Chuang Gan $^{34}$ + +$^{1}$ South China University of Technology, $^{2}$ Information Technology R&D Innovation Center of Peking University $^{3}$ UMass Amherst, $^{4}$ MIT-IBM Watson AI Lab, $^{5}$ Key Laboratory of Big Data and Intelligent Robot, Ministry of Education $^{6}$ Peking University Shenzhen Graduate School, $^{7}$ Pazhou Laboratory + +# Abstract + +How to learn discriminative video representation from unlabeled videos is challenging but crucial for video analysis. The latest attempts seek to learn a representation model by predicting the appearance contents in the masked regions. However, simply masking and recovering appearance contents may not be sufficient to model temporal clues as the appearance contents can be easily reconstructed from a single frame. To overcome this limitation, we present Masked Motion Encoding (MME), a new pretraining paradigm that reconstructs both appearance and motion information to explore temporal clues. In MME, we focus on addressing two critical challenges to improve the representation performance: 1) how to well represent the possible long-term motion across multiple frames; and 2) how to obtain fine-grained temporal clues from sparsely sampled videos. Motivated by the fact that human is able to recognize an action by tracking objects' position changes and shape changes, we propose to reconstruct a motion trajectory that represents these two kinds of change in the masked regions. Besides, given the sparse video input, we enforce the model to reconstruct dense motion trajectories in both spatial and temporal dimensions. Pre-trained with our MME paradigm, the model is able to anticipate long-term and fine-grained motion details. Code is available at https://github.com/XinyuSun/MME. + +# 1. Introduction + +Video representation learning plays a critical role in video analysis like action recognition [15,32,79], action localization [12,81], video retrieval [4,82], videoQA [40], etc. Learning video representation is very difficult for two reasons. Firstly, it is extremely difficult and labor-intensive to annotate videos, and thus relying on annotated data to learn + +video representations is not scalable. Also, the complex spatial-temporal contents with a large data volume are difficult to be represented simultaneously. How to perform self-supervised videos representation learning only using unlabeled videos has been a prominent research topic [7,13,49]. + +Taking advantage of spatial-temporal modeling using a flexible attention mechanism, vision transformers [3, 8, 25, 26, 53] have shown their superiority in representing video. Prior works [5, 37, 84] have successfully introduced the mask-and-predict scheme in NLP [9, 23] to pre-train an image transformer. These methods vary in different reconstruction objectives, including raw RGB pixels [37], handcrafted local patterns [75], and VQ-VAE embedding [5], all above are static appearance information in images. Based on previous successes, some researchers [64, 72, 75] attempt to extend this scheme to the video domain, where they mask 3D video regions and reconstruct appearance information. However, these methods suffers from two limitations. First, as the appearance information can be well reconstructed in a single image with an extremely high masking ratio (85% in MAE [37]), it is also feasible to be reconstructed in the tube-masked video frame-by-frame and neglect to learn important temporal clues. This can be proved by our ablation study (cf. Section 4.2.1). Second, existing works [64, 75] often sample frames sparsely with a fixed stride, and then mask some regions in these sampled frames. The reconstruction objectives only contain information in the sparsely sampled frames, and thus are hard to provide supervision signals for learning fine-grained motion details, which is critical to distinguish different actions [3, 8]. + +In this paper, we aim to design a new mask-and-predict paradigm to tackle these two issues. Fig. 1(a) shows two key factors to model an action, i.e., position change and shape change. By observing the position change of the person, we realize he is jumping in the air, and by observing the shape change that his head falls back and then tucks to his chest, we are aware that he is adjusting his posture to cross the bar. We believe that anticipating these changes helps the model better understand an action. + +![](images/f272cf36b26507a7e48f4f7bcca41ae586621c1b8ea190f28b5134bbe0e3c822.jpg) +(a) Two key factors to recognize a high jump action. + +![](images/bebe77796b7293aa8da91f9a76519ce7c7c8b7fab55d3eb53b9e5144016de71c.jpg) +(b) Appearance reconstruction vs. motion trajectory reconstruction. +Figure 1. Illustration of motion trajectory reconstruction for Masked Motion Encoding. (a) Position change and shape change over time are two key factors to recognize an action, we leverage them to represent the motion trajectory. (b) Compared with the current appearance reconstruction task, our motion trajectory reconstruction takes into account both appearance and motion information. + +Based on this observation, instead of predicting the appearance contents, we propose to predict motion trajectory, which represents impending position and shape changes, for the mask-and-predict task. Specifically, we use a dense grid to sample points as different object parts, and then track these points using optical flow in adjacent frames to generate trajectories, as shown in Fig. 1(b). The motion trajectory contains information in two aspects: the position features that describe relative movement; and the shape features that describe shape changes of the tracked object along the trajectory. To predict this motion trajectory, the model has to learn to reason the semantics of masked objects based on the visible patches, and then learn the correlation of objects among different frames and try to estimate their accurate motions. We name the proposed mask-and-predict task as Masked Motion Encoding (MME). + +Moreover, to help the model learn fine-grained motion details, we further propose to interpolate the motion trajectory. Taking sparsely sampled video as input, the model is asked to reconstruct spatially and temporally dense motion trajectories. This is inspired by the video frame interpolation task [77] where a deep model can reconstruct dense video at the pixel level from sparse video input. Different from it, we aim to reconstruct the fine-grained motion details of moving objects, which has higher-level motion information and is helpful for understanding actions. Our main contributions are as follows: + +- Existing mask-and-predict task based on appearance reconstruction is hard to learn important temporal clues, which is critical for representing video content. Our Masked Motion Encoding (MME) paradigm overcomes this limitation by asking the model to reconstruct motion trajectory. +- Our motion interpolation scheme takes a sparsely sampled video as input and then predicts dense motion trajectory in both spatial and temporal dimensions. This scheme endows the model to capture long-term and fine-grained motion clues from sparse video input. + +Extensive experimental results on multiple standard video recognition benchmarks prove that the representations learned from the proposed mask-and-predict task achieve state-of-the-art performance on downstream action recognition tasks. Specifically, pre-trained on Kinetics-400 [10], our MME brings the gain of $2.3\%$ on Something-Something V2 [34], $0.9\%$ on Kinetics-400, $0.4\%$ on UCF101 [59], and $4.7\%$ on HMDB51 [44]. + +# 2. Related Work + +Self-supervised Video Representation Learning. Self-supervised video representation learning aims to learn discriminative video features for various downstream tasks in the absence of accurate video labels. To this end, most of the existing methods try to design an advanced pretext task like predicting the temporal order of shuffled video crops [78], perceiving the video speediness [7, 13] or solving puzzles [43, 49]. In addition, contrastive learning is also widely used in this domain [14, 16, 36, 41, 46, 54, 57, 68, 69], which constrains the consistency between different augmentation views and brings significant improvement. In particular, ASCNet and CoCLR [36, 41] focus on mining hard positive samples in different perspectives. Optical flow has also been proven to be effective in capturing motion information [70, 76]. Besides, tracking video objects' movement is also used in self-supervised learning [17, 18, 65, 73, 74]. Among them, Wang et al. [74] only utilizes the spatial encoder to extract frame appearance information. CtP framework [65] and Siamese-triplet network [73] only require the model to figure out the position and size changes of a specific video patch. Different from these methods, our proposed MME trace the fine-grained movement and shape changes of different parts of objects in the video, hence resulting in a superior video representation. Tokmakov et al. [63] utilize Dense Trajectory to provide initial pseudo labels for video clustering. However, the model does not predict trajectory motion features explicitly. Instead, we consider long-term and fine-grained trajectory motion features as explicit reconstruction targets. + +![](images/a4b473440a755a23c6769f6da56003c8dbe69b01002fb2d4fc4c4647893b4768.jpg) +Figure 2. Overview of Masked Motion Encoding (MME). Given a sparsely sampled video, we first divide it into several patches and randomly mask out some of them. And then, we feed the remaining patches to a ViT encoder to extract video representation. Last, a lightweight ViT decoder is involved to predict the content in the masked region, i.e., a motion trajectory containing position changes and shape changes of moving objects. + +![](images/44625b59b73d1a3c19cca9691457b126a13c597b7f1cc929e32566029637542d.jpg) + +![](images/0654f83de7f1007f192b32769cf0f8c4e5f6d8475731897f8b64b741ca50fe6c.jpg) + +Mask Modeling for Vision Transformer. Recently, BEiT and MAE [5, 37] show two excellent mask image modeling paradigms. BEVT [72] and VideoMAE [64] extend these two paradigms to the video domain. To learn visual representations, BEVT [72] predicts the discrete tokens generated by a pre-trained VQ-VAE tokenizer [56]. Nevertheless, pre-training such a tokenizer involves an unbearable amount of data and computation. In contrast, VideoMAE [64] pretrain the Vision Transformer by regressing the RGB pixels located in the masked tubes of videos. Due to asymmetric encoder-decoder architecture and extremely high masking ratio, model pre-training is more efficient with VideoMAE. Besides, MaskFeat [75] finds that predicting the Histogram of Gradient (HOG [21]) of masked video contents is a strong objective for the mask-and-predict paradigm. Existing methods only consider static information in each video frame, thus the model can speculate the masked area by watching the visible area in each frame independently and failed to learn important temporal clues (cf. Section 3.1). Different from the video prediction methods [35, 51, 58, 60] that predict the future frames in pixel or latent space, our Masked Motion Encoding paradigm predicts incoming fine-grained motion in masked video regions, including position changes and shape changes. + +# 3. Proposed Method + +We first revisit the current masked video modeling task for video representation learning (cf. Section 3.1). Then, we introduce our masked motion encoding (MME), where we change the task from recovering appearance to recovering motion trajectory (cf. Section 3.2). + +# 3.1. Rethinking Masked Video Modeling + +Given a video clip sampled from a video, self-supervised video representation learning aims to learn a feature encoder $f_{enc}(\cdot)$ that maps the clip to its corresponding feature that best describes the video. Existing masked video modeling methods [64, 75] attempt to learn such a feature + +encoder through a mask-and-predict task. Specifically, the input clip is first divided into multiple non-overlapped 3D patches. Some of these patches are randomly masked and the remaining patches are fed into the feature encoder, followed by a decoder $f_{dec}(\cdot)$ to reconstruct the information in the masked patches. Different works aim to reconstruct different information (e.g., raw pixel in VideoMAE [64] and HOG in MaskFeat [75]). + +However, the existing works share a common characteristic where they all attempt to recover static appearance information of the masked patches. Since an image with a high masking ratio (85% in MAE [37]) can be well reconstructed [37,75], we conjecture that the masked appearance information of a video can also be reconstructed frame by frame independently. In this sense, the model may focus more on the contents in the same frame. This may hinder the models from learning important temporal clues, which is critical for video representation. We empirically study this conjecture in the ablation study (cf. Section 4.2.1). + +# 3.2. General Scheme of MME + +To better learn temporal clues of a video, our MME changes the reconstruction content from static appearance information to object motion information, including position and shape change of objects. As shown in Fig. 2, a video clip is sparsely sampled from a video and is divided into a number of non-overlapped 3D patches of size $t \times h \times w$ , corresponding to time, height, and width. We follow VideoMAE [64] to use the tube masking strategy, where the masking map is the same for all frames, to mask a subset of patches. For computation efficiency, we follow MAE [37] to only feed the unmasked patches (and their positions) to the encoder. The output representation together with learnable [MASK] tokens are fed to a decoder to reconstruct motion trajectory $\mathbf{z}$ in the masked patches. The training loss for MME is + +$$ +\mathcal {L} = \sum_ {i \in \mathcal {I}} \left| \mathbf {z} _ {i} - \hat {\mathbf {z}} _ {i} \right| ^ {2}, \tag {1} +$$ + +where $\hat{\mathbf{z}}$ is the predicted motion trajectory, and $\mathcal{I}$ is the index set of motion trajectories in all masked patches. + +Motivated by the fact that we humans recognize actions by perceiving position changes and shape changes of moving objects, we leverage these two types of information to represent the motion trajectory. Through pre-training on MME task, the model is endowed with the ability to explore important temporal clues. Another important characteristic of the motion trajectory is that it contains fine-grained motion information extracted at raw video rate. This fine-grained motion information provides the model with a supervision signal to anticipate fine-grained action from sparse video input. In the following, we will introduce the proposed motion trajectory in detail. + +# 3.3. Motion Trajectory for MME + +The motion of moving objects can be represented in various ways such as optical flow [27], histograms of optical flow (HOF) [45], and motion boundary histograms (MBH) [45]. However, these descriptors can only represent short-term motion between two adjacent frames. We hope our motion trajectory represents long-term motion, which is critical for video representation. To this end, inspired by DT [66], we first track the moving object in the following $L$ frames to cover a longer range of motion, resulting in a trajectory $\mathbf{T}$ , i.e., + +$$ +\mathbf {T} = \left(\mathbf {p} _ {t}, \mathbf {p} _ {t + 1}, \dots , \mathbf {p} _ {t + L}\right), \tag {2} +$$ + +where $\mathbf{p}_t = (x_t, y_t)$ represents a point located at $(x_t, y_t)$ of frame $t$ , and $(\cdot, \cdot)$ indicates the concatenation operation. Along this trajectory, we fetch the position features $\mathbf{z}^p$ and shape features $\mathbf{z}^s$ of this object to compose a motion trajectory $\mathbf{z}$ , i.e., + +$$ +\mathbf {z} = \left(\mathbf {z} ^ {p}, \mathbf {z} ^ {s}\right). \tag {3} +$$ + +The position features are represented by the position transition relative to the last time step, while the shape features are the HOG descriptors of the tracked object in different time steps. + +Tracking objects using spatially and temporally dense trajectories. Some previous works [2, 18, 19, 24] try to use one trajectory to represent the motion of an individual object. In contrast, DT [66] points out that tracking spatially dense feature points sampled on a grid space performs better since it ensures better coverage of different objects in a video. Following DT [66], we use spatially dense grid points as the initial position of each trajectory. Specifically, we uniformly sample $K$ points in a masked patch of size $t \times h \times w$ , where each point indicates a part of an object. For each point, we track it through temporally dense $L$ frames according to the dense optical flow, resulting in $K$ trajectories. In this way, the model is able to capture spatially and + +temporally dense motion information of objects through the mask-and-predict task. + +As a comparison, the reconstruction contents in existing works [64, 75] often extract temporally sparse videos sampled with a large stride $s > 1$ . The model takes as input a sparse video and predicts these sparse contents for learning video representation. Different from these works, our model also takes as input sparse video but we push the model to interpolate motion trajectory containing fine-grained motion information. This simple trajectory interpolation task does not increase the computational cost of the video encoder but helps the model learn more fine-grained action information even given sparse video as input. More details about dense flow calculating and trajectory tracking can be found in Appendix. + +Representing position features. Given a trajectory $\mathbf{T}$ consisting of the tracked object position at each frame, we are more interested in the related movement of objects instead of their absolute location. Consequently, we represent the position features with related movement between two adjacent points $\Delta \mathbf{p}_t = \mathbf{p}_{t+1} - \mathbf{p}_t$ , i.e., + +$$ +\mathbf {z} ^ {p} = \left(\Delta \mathbf {p} _ {t}, \dots , \Delta \mathbf {p} _ {t + L - 1}\right), \tag {4} +$$ + +where $\mathbf{z}^p$ is a $L\times 2$ dimensional feature. As each patch contains $K$ position features, we concatenate and normalize them as position features part of the motion trajectory. + +Representing shape features. Besides embedding the movement, the model also needs to be aware of the shape changes of objects to recognize actions. Inspired by HOG [21], we use histograms of oriented gradients (HOG) with 9 bins to describe the shape of objects. + +Compared with existing works [64, 75] that reconstruct HOG in every single frame, we are more interested in the dynamic shape changes of an object, which can better represent action in a video. To this end, we follow DT [66] to calculate trajectory-aligned HOG, consisting of HOG features around all tracked points in a trajectory, i.e., + +$$ +\mathbf {z} ^ {s} = \left(\operatorname {H O G} \left(\mathbf {p} _ {t}\right), \dots , \operatorname {H O G} \left(\mathbf {p} _ {t + L - 1}\right)\right), \tag {5} +$$ + +where $\mathrm{HOG}(\cdot)$ is the HOG descriptor and $\mathbf{z}^s$ is a $L\times 9$ dimensional feature. Also, as one patch contains $K$ trajectories, we concatenate $K$ trajectory-aligned HOG features and normalize them to the standard normal distribution as the shape features part of the motion trajectory. + +# 4. Experiments + +Implementation details. We conduct experiments on Kinetics-400 (K400), Something-Something V2 (SSV2), UCF101 and HMDB51 datasets. Unless otherwise stated, we follow previous trails [64] and feed the model a + +
MethodBackbonePre-trainLabelEpochFramesGFLOPsParam.Acc.@1
TimeSformer [8]ViT-BIN-21K✓158196×1×3121M59.5
ViViT-L [3]ViT-LIN-21K+K400✓35163992 × 3 × 4311M65.4
Motionformer [53]ViT-BIN-21K+K400✓3516370 × 1 × 3109M66.5
SlowFast [29]ResNet-101K400✓1968+32106 × 1 × 353M63.1
MViT [26]MViT-BK400✓10064455 × 1 × 337M67.7
SVT [57]ViT-BIN-21K+K400×208+64196×1×3121M59.6
LSTCL [68]Swin-BK400×20016360×5×188M67.0
BEVT [72]Swin-BK400+DALL-E×15032282×1×388M67.1
VIMPAC [62]ViT-LHowTo100M+DALLE×10010N/A×10×3307M68.1
VideoMAE [64]ViT-BSSv2×40016180×2×387M67.6
VideoMAE [64]ViT-BSSv2×80016180×2×387M69.3
VideoMAE [64]ViT-BK400×80016180×2×387M68.2
MME (Ours)ViT-BSSv2×40016180×2×387M69.2
MME (Ours)ViT-BSSv2×80016180×2×387M70.0
MME (Ours)ViT-BK400×40016180×2×387M69.3
MME (Ours)ViT-BK400×80016180×2×387M70.5
+ +Table 1. Comparison with state-of-the-arts on Something-Something V2. Our MME outperforms the previous best method. Remarkably, MME show better transferability ability (transfer from K400 to SSv2) compared with VideoMAE. + +
MethodBackbonePre-trainLabelEpochFramesGFLOPsParam.Acc.@1
TimeSformer [8]ViT-BIN-21K✓158196×1×3121M78.3
Motionformer [53]ViT-BIN-21K✓3516370×10×3121M79.7
Motionformer [53]ViT-HRIN-21K✓3516959×10×3382M81.1
VIMPAC [62]ViT-LHowTo100M+DALLE×10010N/A×10×3307M77.4
SVT [57]ViT-BIN-21K+K400×208+64196×1×3121M78.1
BEVT [72]Swin-BK400+DALL-E×15032282×1×388M76.2
BEVT [72]Swin-BIN21K+K400+DALL-E×15032282×1×388M80.6
LSTCL [68]Swin-BK400×20016360×5×188M81.5
OmniMAE [33]ViT-BIN1K+K400×160032180×5×387M80.6
VideoMAE [64]ViT-BK400×160016180×5×387M80.9
ST-MAE [28]ViT-BK400×160016180×7×387M81.3
MME (Ours)ViT-BK400×160016180×7×387M81.8
+ +Table 2. Comparison with state-of-the-arts on Kinetics-400. Our MME outperforms the previous best method. + +$224 \times 224$ 16-frame clip with strides 4 and 3 for K400 and SSV2 respectively. We use vanilla ViT-Base [25] as the backbone in all experiments. As for the motion trajectory, we set the trajectory length to $L = 6$ and the number of trajectories in each patch to $K = 4$ . Our models are trained on 16 NVIDIA GeForce RTX 3090 GPUs. Note that for all the ablation studies, we use a smaller pre-training dataset and reduce the input patches of fine-tuning to speed up the experiments, details can be found in Appendix. + +# 4.1. Comparisons with State-of-the-arts + +Fine-tuning. On the temporal-heavy dataset SSv2, we conduct experiments using two settings, namely in-domain pre-training (i.e., pre-train and fine-tune on the same dataset) and cross-domain pre-training (i.e., pre-train and fine-tune on different datasets). In Tab. 1, we have two main observations. 1) Our MME significantly performs + +better than previous state-of-the-arts VideoMAE [64] on both in-domain and cross-domain settings. Specifically, under cross-domain settings (transferring from K400 to SSv2) with 800 epochs, our MME outperforms it by $2.3\%$ , increasing the accuracy from $68.2\%$ to $70.5\%$ . 2) Our MME performs better under cross-domain setting compared with indomain setting, which is contrary to the existing observation on VideoMAE [64] and ST-MAE [28]. We attribute it to our masked motion encoding pretext task that helps models focus more on learning temporal clues instead of learning domain-specific appearance information (e.g., color, lighting). The learned temporal clues are more universal across different domains for video representation. This also shows the potential for taking advantage of larger datasets for pretraining, which we leave for future work. + +In Tab. 2, our MME also achieves state-of-the-art performance on another large-scale dataset K400 with a gain + +
MethodBackbonePre-trainUCF101HMDB51
XDC [1]R(2+1)DIG65M94.267.1
GDT [52]R(2+1)DIG65M95.272.8
CVRL [54]R50K40092.967.9
CORPf [39]R50K40093.568.0
ρBYOL [30]R50K40094.272.1
VideoMAE [64]ViT-BK40096.173.3
MME (Ours)ViT-BK40096.578.0
+ +of $0.5\%$ . In Tab. 3, we also evaluate the transferability of MME by transferring a K400 pre-trained model to two small-scale datasets, namely UCF101 and HMDB51. Our MME achieves a large gain of $4.7\%$ on HMDB51 and $0.4\%$ on UCF101 compared with VideoMAE. This further demonstrates the strong ability of MME to learn from the large-scale dataset and transfer it to a small-scale one. + +Linear probing. We further evaluate our MME under liner probing setting on the SSv2 dataset as this dataset focuses more on the action. We follow SVT [57] to fix the transformer backbone and train a linear layer for 20 epochs. In Tab. 4, our method significantly outperform previous contrastive-based method [57] by $10.9\%$ and masked video modeling methods [64] by $4.9\%$ . These results indicate that the representation learned from our MME task contains more motion clues for action recognition. + +# 4.2. Ablation Study on Reconstruction Content + +In this section, we investigate the effect of different reconstruction contents, including appearance, short-term motion, and the proposed long-term fine-grained motion. + +# 4.2.1 Evaluating Appearance Reconstruction + +In section 3.1, we conjecture that appearance reconstruction is finished in each video frame independently. To verify this conjecture, we corrupt the temporal information of a video by randomly shuffling its frames. Since the video frames are randomly permuted, it is hard for a model to leverage information from other frames to reconstruct the current frame through their temporal correlation. We mask $90\%$ regions of these videos following VideoMAE [64]. In Fig. 3 (a), we find that the reconstruction error (i.e., L2 error between predicted and ground-truth pixel) converges to a low value. We also conduct the same experiment using raw video without shuffling. The model undergoes a similar convergence process. This demonstrates that the masked pixels are well reconstructed without temporal information. + +To further evaluate whether the model pre-trained on the appearance reconstruction task (i.e., VideoMAE) can well capture important temporal clues, we transfer two + +Table 3. Comparison with state-of-the-arts on UCF101 and HMDB51. + +
MethodBackbonePre-trainAcc.@1
TimeSformer [8]ViT-BIN-21K14.0
SVT [57]ViT-BIN-21K+K40018.3
VideoMAE [64]ViT-BSSv224.3
MME (Ours)ViT-BSSv229.2
+ +Table 4. Linear Probing on Something-Something V2. + +![](images/56be86abe4159245519ff5cb14764bb11b4932b0cc46e0cd4ae25e501ca43177.jpg) +(a) Masked video reconstruction error. +Figure 3. We perform mask-and-predict pre-training using temporal shuffled videos and raw videos and then transfer them to the downstream task. (a) Masked appearance (i.e., pixels) is well reconstructed without temporal information. (b) Our MME outperforms VideoMAE when both provided temporal information. + +![](images/3e3244eb30176476d083000ccfc9011e7d083583d70fbb073b1d5d56688315eb.jpg) +(b) Downstream performance. + +VideoMAE models (pre-trained on shuffled videos and raw videos, respectively) to the downstream action recognition task. In Fig. 3 (b), these two models perform competitively, indicating that removing the temporal information from the pre-training videos barely affects the learned video representation. We speculate this is because the VideoMAE model pays little attention to temporal information when performing the appearance reconstruction pre-training task. As a comparison, our MME paradigm performs significantly better than VideoMAE when the temporal information is provided for pre-training $(64.1\%$ vs. $60.9\%)$ . Once the temporal information is removed, MME performs similarly with VideoMAE $(60.7\%$ vs. $60.9)$ . This demonstrates our MME paradigm takes great advantage of temporal information in the pre-training phase. + +# 4.2.2 Evaluating Motion Reconstruction + +In the following ablations, we separately verify the effectiveness of two proposals of our methods: + +Effectiveness of motion trajectory. We compare our motion trajectory with two appearance reconstruction baselines and three short-term motion reconstruction baselines. + +- Appearance reconstruction baselines: 1) HOG: predicting HOG features [21] in the middle frame of each 3D patch [75]. 2) Pixel: predicting all the pixels of each 3D patch [28, 37, 64]. + +
Reconstruction contentAcc.@1Acc.@5
Pixel61.186.6
HOG61.186.9
HOG + Flow60.986.4
HOG + HOF61.386.6
HOG + MBH61.486.8
HOG + Traj. w/o shape63.588.2
Traj. (ours)64.188.4
+ +Table 5. Comparison of different Reconstruction content. Our motion trajectory outperforms all the other contents. + +
Reconstruction contentInterpolationAcc.@1Acc.@5
HOGx61.186.9
HOG✓60.986.8
HOG + MBHx61.386.8
HOG + MBH✓61.486.8
Traj. (ours)x62.287.5
Traj. (ours)✓64.188.4
+ +Table 6. Effectiveness of trajectory interpolation. Interpolating the motion trajectory brings significant improvement compared to the baselines. + +- Short-term motion reconstruction baselines: 1) Flow: predicting the dense optical flow map of the masked patches. 2) HOF: predicting the histogram of optical flow orientation. 3) MBH: predicting the motion boundary histogram, which is the second-order local statistic of optical flow. + +In Tab. 5, two types of appearance reconstruction contents (i.e., Pixel and HOG) perform similarly. Based on the HOG baseline, predicting additional short-term motion contents brings slight improvement (at least by $0.2\%$ ) except for directly predicting optical flow (decreases by $0.2\%$ ). We suspect that predicting the dense flow map for each pixel is too hard for a model and thus harms the performance. Instead, the histogram version of flow (i.e., HOF and MBH) help to reduce the prediction ambiguity [75], resulting in better performance. Furthermore, after adding our slight variant of motion trajectory (i.e., with position features but without shape features shown in Eq. (3)), the performance improves significantly compared to both appearance reconstruction baselines (by $2.4\%$ ) and short-term motion reconstruction baselines (by $2.1\%$ ). On top of this variant, adding trajectory-aligned shape features further brings additional gain by $0.6\%$ . Because tracking the trajectory of objects in multiple frames provides long-term motion compared to short-term motion reconstruction, we believe the learning of long-term motion clues is one of the keys to enhancing video representation. + +Effectiveness of trajectory interpolation. In MME, we feed temporally sparse video to the model and push it to interpolate temporally dense trajectories. We are interested in whether interpolating appearance (i.e., HOG) or short-term motion contents (i.e., MBH) benefit learning fine-grained motion details. To this end, we conduct experiments to predict dense HOG and MBH, i.e., predicting these features of all frames within a masked patch even though these frames are not sampled as input. In Tab. 6, interpolating HOG or MBH brings little improvement. We suspect this is because these dense features represent the continuous movement of objects poorly, as they have not explicitly tracked the moving objects. In contrast, our temporally dense motion trajectory tracks the object densely, providing temporal fine-grained information for learning. The trajectory interpolation improves our MME by $1.9\%$ . This result proves that the learning of fine-grained motion details is another key to enhancing video representation. + +# 4.3. Ablation Study on MME Design + +In this section, we conduct ablation studies to investigate the effect of different parameters for MME pre-training. + +Trajectory length. Trajectory length $L$ indicates how many frames we track an object for extracting the motion trajectory. If $L = 0$ , our method degrades to MaskFeat [75], which only predicts HOG features in the masked patch. In Tab. 7a, MME achieves the best performance when $L = 6$ . This is because a longer trajectory provides more motion information for learning. But if the trajectory is too long, the drift of flow introduces accumulated noise when tracking the objects, which harms the performance. + +Trajectory density. In our experiments, we spatially dense track $K = 4$ motion trajectories. We conduct experiments to evaluate whether it is necessary to densely track multiple trajectories in a patch. Specifically, we assume only one object exists in a patch, and thus the 4 trajectories are similar. We merge these 4 trajectories as one by averaging or selecting the most salient one. We then consider the merged trajectories as the trajectory motion. In Tab. 7c, using dense trajectories significantly outperform the merged baselines. We suspect the spatial dense trajectories provide more motion information for learning. + +Trajectory normalization. Due to the irregular distribution of motion and shape information in the video, the trajectory values in different patches vary a lot. To help the model better model the trajectory, we adopt the patch norm method to normalize the trajectory values in each patch into a standard normal distribution, as has also been done in MAE [37]. The mean and variance are calculated by all $K = 4$ trajectories in a patch. Tab. 7b shows the effectiveness of adopting the patch norm. + +
LengthAcc.@1Acc.@5
L = 463.288.0
L = 664.188.4
L = 863.388.0
+ +(a) Effect of trajectory length. + +
SettingAcc.@1Acc.@5
w/o patch norm.62.187.44
w/ patch norm.64.188.4
+ +(b) Effect of trajectory normalization. + +
Trajectory densityAcc.@1Acc.@5
Dense64.188.4
Sparse (Max)60.286.3
Sparse (Mean)61.887.1
+ +(c) Effect of trajectories density. + +
StrideAcc.@1Acc.@5
s = 462.987.7
s = 364.188.4
s = 263.588.3
s = 162.787.7
+ +(d) Effect of sampling stride. + +
RatioTypeAcc.@1Acc.@5
90%Tube62.787.8
70%Tube64.188.4
40%Tube61.386.9
40%Cube61.286.9
+ +(e) Effect of spatial masking. + +
TypeDepthFLOPsAcc.@1Acc.@5
Joint151G64.188.4
255G63.488.3
Sep.156G63.488.4
265G63.388.0
+ +(f) Effect of decoder. Sep. means two decoders. + +Table 7. Pre-training details. We explore the detailed setting of our MME pre-training. The default entry is marked in gray . + +Sampling stride. In order to control the difficulty of interpolating the motion trajectory, we ablate different temporal sampling strides leading to different levels of the sparse input video. With a medium sampling stride ( $s = 2$ ), our MME achieves the best performance. We also found that the performance decreases when the input video shares the same stride as the dense motion trajectory ( $s = 1$ ). One possible reason is that it is difficult for the model to anticipate fine-grained motion details in this case. + +Spatial masking strategy. We ablated study different spatial masking strategies and ratios in Tab. 7e. We start from the default setting of MaskFeat [75] that uses a cube masking strategy with a $40\%$ masking ratio, then we increase the masking ratio and adopt the tube masking strategy proposed by VideoMAE [64]. Using tube masking with a medium masking ratio $(70\%)$ performs the best. + +Decoder setting. We use several transformer blocks as decoder to reconstruct motion trajectory. We ablate different architectures of the decoder in Tab. 7f. Using a slight decoder with 1 transformer block performs the best, with the lowest computational cost. Increasing the depth of the decoder or using two parallel decoders to separately predict the position and shape features introduces more computational cost but brings little improvement. + +# 4.4. Visualization Results + +We visualize the predicted motion trajectory in Fig. 4. The model is able to reconstruct the spatial and temporally dense trajectories of different object parts from a temporally sparse input video. Even without seeing the object in the current frame (the hand being masked out), the model is able to locate it through contextual patches and accurately estimate its trajectory. We only visualize the position changes since our prediction of HOG features has similar quality to the previous work [75]. + +![](images/26244ef12f2d3ac2bd9de8767aa9a07b365bc3d25aaa0d4adbdc80fe5a848d1e.jpg) +Figure 4. Visualization of predicted motion trajectory. Each column represents (from left to right): (a) Masked frames; (b) ground truth; (c) prediction; (d) original frame. + +# 5. Conclusion + +We present Masked Motion Encoding, a simple mask-and-predict paradigm that learns video representation by reconstructing the motion contents of masked regions. In particular, we reconstruct motion trajectory, which tracks moving objects densely to record their position and shape changes. By reconstructing this motion trajectory, the model learns long-term and fine-grained motion clues from sparse input videos. Empirical results show that our method outperforms the state-of-the-art mask-and-predict methods on Something-Something V2, Kinetics-400, UCF-101, and HMDB-51 datasets for action recognition. + +# Acknowledgement + +This work was partially supported by the National Natural Science Foundation of China (NSFC) (No. 62072190, 62072186), Ministry of Science and Technology Foundation Project (No. 2020AAA0106900), Program for Guangdong Introducing Innovative and Entrepreneurial Teams (No. 2017ZT07X183), Guangdong Basic and Applied Basic Research Foundation (No. 2019B1515130001). + +# References + +[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, pages 9758-9770, 2020. 6 +[2] Nadeem Anjum and Andrea Cavallaro. Multifeature object trajectory clustering for video analysis. IEEE Transactions on Circuits and Systems for Video Technology, 18:1555-1564, 2008. 4 +[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 1, 5 +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 1 +[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: BERT pre-training of image transformers. In ICLR, 2022. 1, 3, 13 +[6] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417, 2006. 12 +[7] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In CVPR, pages 9922-9931, 2020. 1, 2 +[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding. In ICML, pages 813-824, 2021. 1, 5, 6 +[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 1 +[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 2 +[11] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703, 2020. 13 +[12] Peihao Chen, Chuang Gan, Guangyao Shen, Wenbing Huang, Runhao Zeng, and Mingkui Tan. Relation attention for temporal action localization. IEEE Transactions on Multimedia, 22:2723-2733, 2019. 1 +[13] Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan. Rspnet: Relative speed perception for unsupervised video representation learning. In AAAI, pages 1045-1053, 2021. 1, 2 +[14] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas H Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurlPS, 2022. 2 +[15] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 29:8292-8302, 2020. 1 + +[16] Zhenfang Chen, Lin Ma, Wenhan Luo, Peng Tang, and Kwan-Yee K Wong. Look closer to ground better: Weakly-supervised temporal grounding of sentence in video. arXiv, 2020. 2 +[17] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In ACL, 2019. 2 +[18] Zhenfang Chen, Jiayuan Mao, Jiajun Wu, Kwan-Yee Kenneth Wong, Joshua B Tenenbaum, and Chuang Gan. Grounding physical concepts of objects and events through dynamic visual reasoning. arXiv, 2021. 2, 4 +[19] Zhenfang Chen, Kexin Yi, Yunzhu Li, Mingyu Ding, Antonio Torralba, Joshua B Tenenbaum, and Chuang Gan. Comphy: Compositional physical reasoning of objects and events from videos. In ICLR, 2022. 4 +[20] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, pages 702-703, 2020. 13 +[21] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, pages 886-893, 2005. 3, 4, 6, 13 +[22] Navneet Dalal, Bill Triggs, and Cordelia Schmid. Human detection using oriented histograms of flow and appearance. In ECCV, pages 428-441, 2006. 13 +[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, pages 4171–4186, 2018. 1 +[24] Mingyu Ding, Zhenfang Chen, Tao Du, Ping Luo, Josh Tenenbaum, and Chuang Gan. Dynamic visual reasoning by learning differentiable physics models from video and language. In NeurIPS, volume 34, pages 887-899, 2021. 4 +[25] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 5 +[26] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 1, 5 +[27] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In *SCIA*, pages 363-370, 2003. 4 +[28] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In NeurIPS, 2022. 5, 6 +[29] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 5 +[30] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross B. Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, pages 3299-3309, 2021. 6 +[31] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to + +image analysis and automated cartography. Communications of the ACM, 24:381-395, 1981. 12 +[32] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In European Conference on Computer Vision (ECCV), pages 758-775, 2020. 1 +[33] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 5 +[34] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 2 +[35] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martín-Martín, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. arXiv preprint arXiv:2206.11894, 2022. 3 +[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In NeurIPS, pages 5679-5690, 2020. 2 +[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 3, 6, 7 +[38] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. In CVPR, pages 8126-8135, 2020. 13 +[39] Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. In ICCV, pages 7939-7949, 2021. 6 +[40] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware Graph Convolutional Networks for Video Question Answering. In AAAI, 2020. 1 +[41] Deng Huang, Wenhao Wu, Weiwen Hu, Xu Liu, Dongliang He, Zhihua Wu, Xiangmiao Wu, Mingkui Tan, and Errui Ding. Ascnet: Self-supervised video representation learning with appearance-speed consistency. In ICCV, pages 8096-8105, 2021. 2 +[42] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q. Weinberger. Deep networks with stochastic depth. In ECCV, pages 646-661, 2016. 13 +[43] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. 2 +[44] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, pages 2556-2563, 2011. 2 +[45] Ivan Laptev, Marcin Marszalek, Cordelia Schmid, and Benjamin Rozenfeld. Learning realistic human actions from movies. In CVPR, pages 1-8, 2008. 4, 13 + +[46] Yuan Liu, Jingyuan Chen, Zhenfang Chen, Bing Deng, Jianqiang Huang, and Hanwang Zhang. The blessings of unlabeled background in untrimmed videos. In CVPR, pages 6176-6185, 2021. 2 +[47] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 13 +[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 13 +[49] Dezhao Luo, Chang Liu, Yu Zhou, Dongbao Yang, Can Ma, Qixiang Ye, and Weiping Wang. Video cloze procedure for self-supervised spatio-temporal learning. In AAAI, pages 11701-11708, 2020. 1, 2 +[50] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035, 2019. 13 +[51] Viorica Patraucean, Ankur Handa, and Roberto Cipolla. Spatio-temporal video autoencoder with differentiable memory. In ICLR Workshop, 2016. 3 +[52] Mandela Patrick, Yuki M Asano, Polina Kuznetsova, Ruth Fong, Joao F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. 6 +[53] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In NeurIPS, pages 12493-12506, 2021. 1, 5 +[54] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, pages 6964-6974, 2021. 2, 6 +[55] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Xiang Wang, Yuehuan Wang, Yiliang Lv, Changxin Gao, and Nong Sang. Mar: Masked autoencoders for efficient action recognition. arXiv preprint arXiv:2207.11660, 2022. 13 +[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 3 +[57] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael Ryoo. Self-supervised video transformer. In CVPR, pages 2874-2884, 2022. 2, 5, 6 +[58] Adria Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătráucean, Florent Altché, Michal Valko, et al. Broaden your views for self-supervised video learning. In ICCV, pages 1255-1265, 2021. 3 +[59] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2 +[60] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, pages 843-852, 2015. 3 + +[61] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 13 +[62] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 5 +[63] Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. Unsupervised learning of video representations via dense trajectory clustering. In ECCV 2020 Workshops, pages 404-421. Springer, 2020. 2 +[64] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 3, 4, 5, 6, 8, 13 +[65] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. In CVPR, pages 2563-2572, 2021. 2 +[66] Heng Wang, Alexander Kläser, Cordelia Schmid, and Cheng-Lin Liu. Dense trajectories and motion boundary descriptors for action recognition. International journal of computer vision, 103:60-79, 2013. 4, 14 +[67] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In ICCV, pages 3551-3558, 2013. 12, 14 +[68] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. arXiv preprint arXiv:2106.09212, 2021. 2, 5 +[69] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In CVPR, pages 11804-11813, 2021. 2 +[70] Jiangliu Wang, Jianbo Jiao, Linchao Bao, Shengfeng He, Yunhui Liu, and Wei Liu. Self-supervised spatio-temporal representation learning for videos by predicting motion and appearance statistics. In CVPR, pages 4006-4015, 2019. 2 +[71] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41:2740-2755, 2018. 12 +[72] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, pages 14733-14743, 2021. 1, 3, 5 +[73] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, pages 2794-2802, 2015. 2 +[74] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, pages 2566-2576, 2019. 2 +[75] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 1, 3, 4, 6, 7, 8, 13 + +[76] Donglai Wei, Joseph J. Lim, Andrew Zisserman, and William T. Freeman. Learning and using the arrow of time. In CVPR, pages 8052-8060, 2018. 2 +[77] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P Allebach, and Chenliang Xu. Zooming slow-mo: Fast and accurate one-stage space-time video super-resolution. In CVPR, pages 3370-3379, 2020. 2 +[78] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In CVPR, pages 10334–10343, 2019. 2 +[79] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 1 +[80] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 13 +[81] Runhao Zeng, Chuang Gan, Peihao Chen, Wenbing Huang, Qingyao Wu, and Mingkui Tan. Breaking winner-takes-all: Iterative-winners-out networks for weakly supervised temporal action localization. IEEE Transactions on Image Processing, 28(12):5797-5808, 2019. 1 +[82] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1 +[83] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 13 +[84] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan L. Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 1 \ No newline at end of file diff --git a/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/images.zip b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9fde9e3e3f417225f67fcb198b546e375b74184f --- /dev/null +++ b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23e96e0fefcc44a6258e1439c277b8844da392b8831f8987259c484c4fac1341 +size 559220 diff --git a/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/layout.json b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..021d0b2cacd5585c202190beb22c4c86938941c2 --- /dev/null +++ b/2023/Masked Motion Encoding for Self-Supervised Video Representation Learning/layout.json @@ -0,0 +1,9391 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "type": "text", + "content": "Masked Motion Encoding for Self-Supervised Video Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": "Xinyu Sun" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{12*}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": " Peihao Chen" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": " Liangwei Chen" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": " Changhao Li" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": " Thomas H. Li" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": " Mingkui Tan" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{157\\dagger}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": " Chuang Gan" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{34}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "spans": [ + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "text", + "content": "South China University of Technology, " + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "text", + "content": "Information Technology R&D Innovation Center of Peking University " + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "text", + "content": "UMass Amherst, " + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "text", + "content": "MIT-IBM Watson AI Lab, " + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "text", + "content": "Key Laboratory of Big Data and Intelligent Robot, Ministry of Education " + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "text", + "content": "Peking University Shenzhen Graduate School, " + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 77, + 172, + 515, + 213 + ], + "type": "text", + "content": "Pazhou Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 266, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 289, + 566 + ], + "type": "text", + "content": "How to learn discriminative video representation from unlabeled videos is challenging but crucial for video analysis. The latest attempts seek to learn a representation model by predicting the appearance contents in the masked regions. However, simply masking and recovering appearance contents may not be sufficient to model temporal clues as the appearance contents can be easily reconstructed from a single frame. To overcome this limitation, we present Masked Motion Encoding (MME), a new pretraining paradigm that reconstructs both appearance and motion information to explore temporal clues. In MME, we focus on addressing two critical challenges to improve the representation performance: 1) how to well represent the possible long-term motion across multiple frames; and 2) how to obtain fine-grained temporal clues from sparsely sampled videos. Motivated by the fact that human is able to recognize an action by tracking objects' position changes and shape changes, we propose to reconstruct a motion trajectory that represents these two kinds of change in the masked regions. Besides, given the sparse video input, we enforce the model to reconstruct dense motion trajectories in both spatial and temporal dimensions. Pre-trained with our MME paradigm, the model is able to anticipate long-term and fine-grained motion details. Code is available at https://github.com/XinyuSun/MME." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 590, + 128, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 128, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 128, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 684 + ], + "type": "text", + "content": "Video representation learning plays a critical role in video analysis like action recognition [15,32,79], action localization [12,81], video retrieval [4,82], videoQA [40], etc. Learning video representation is very difficult for two reasons. Firstly, it is extremely difficult and labor-intensive to annotate videos, and thus relying on annotated data to learn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 242, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 545, + 302 + ], + "type": "text", + "content": "video representations is not scalable. Also, the complex spatial-temporal contents with a large data volume are difficult to be represented simultaneously. How to perform self-supervised videos representation learning only using unlabeled videos has been a prominent research topic [7,13,49]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 546, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 602 + ], + "type": "text", + "content": "Taking advantage of spatial-temporal modeling using a flexible attention mechanism, vision transformers [3, 8, 25, 26, 53] have shown their superiority in representing video. Prior works [5, 37, 84] have successfully introduced the mask-and-predict scheme in NLP [9, 23] to pre-train an image transformer. These methods vary in different reconstruction objectives, including raw RGB pixels [37], handcrafted local patterns [75], and VQ-VAE embedding [5], all above are static appearance information in images. Based on previous successes, some researchers [64, 72, 75] attempt to extend this scheme to the video domain, where they mask 3D video regions and reconstruct appearance information. However, these methods suffers from two limitations. First, as the appearance information can be well reconstructed in a single image with an extremely high masking ratio (85% in MAE [37]), it is also feasible to be reconstructed in the tube-masked video frame-by-frame and neglect to learn important temporal clues. This can be proved by our ablation study (cf. Section 4.2.1). Second, existing works [64, 75] often sample frames sparsely with a fixed stride, and then mask some regions in these sampled frames. The reconstruction objectives only contain information in the sparsely sampled frames, and thus are hard to provide supervision signals for learning fine-grained motion details, which is critical to distinguish different actions [3, 8]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we aim to design a new mask-and-predict paradigm to tackle these two issues. Fig. 1(a) shows two key factors to model an action, i.e., position change and shape change. By observing the position change of the person, we realize he is jumping in the air, and by observing the shape change that his head falls back and then tucks to his chest, we are aware that he is adjusting his posture to cross the bar. We believe that anticipating these changes helps the model better understand an action." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 271, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 271, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 271, + 703 + ], + "type": "text", + "content": "*Equal contribution. Email: {csxinyusun, phchencs}@gmail.com" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 60, + 703, + 242, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 703, + 242, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 703, + 242, + 712 + ], + "type": "text", + "content": "†Corresponding author. Email: mingkuitan@scut.edu.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2235" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 260, + 158 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 260, + 158 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 260, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 260, + 158 + ], + "type": "image", + "image_path": "f272cf36b26507a7e48f4f7bcca41ae586621c1b8ea190f28b5134bbe0e3c822.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 163, + 251, + 174 + ], + "lines": [ + { + "bbox": [ + 70, + 163, + 251, + 174 + ], + "spans": [ + { + "bbox": [ + 70, + 163, + 251, + 174 + ], + "type": "text", + "content": "(a) Two key factors to recognize a high jump action." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 265, + 71, + 533, + 160 + ], + "blocks": [ + { + "bbox": [ + 265, + 71, + 533, + 160 + ], + "lines": [ + { + "bbox": [ + 265, + 71, + 533, + 160 + ], + "spans": [ + { + "bbox": [ + 265, + 71, + 533, + 160 + ], + "type": "image", + "image_path": "bebe77796b7293aa8da91f9a76519ce7c7c8b7fab55d3eb53b9e5144016de71c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 163, + 520, + 174 + ], + "lines": [ + { + "bbox": [ + 281, + 163, + 520, + 174 + ], + "spans": [ + { + "bbox": [ + 281, + 163, + 520, + 174 + ], + "type": "text", + "content": "(b) Appearance reconstruction vs. motion trajectory reconstruction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "text", + "content": "Figure 1. Illustration of motion trajectory reconstruction for Masked Motion Encoding. (a) Position change and shape change over time are two key factors to recognize an action, we leverage them to represent the motion trajectory. (b) Compared with the current appearance reconstruction task, our motion trajectory reconstruction takes into account both appearance and motion information." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 231, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 422 + ], + "type": "text", + "content": "Based on this observation, instead of predicting the appearance contents, we propose to predict motion trajectory, which represents impending position and shape changes, for the mask-and-predict task. Specifically, we use a dense grid to sample points as different object parts, and then track these points using optical flow in adjacent frames to generate trajectories, as shown in Fig. 1(b). The motion trajectory contains information in two aspects: the position features that describe relative movement; and the shape features that describe shape changes of the tracked object along the trajectory. To predict this motion trajectory, the model has to learn to reason the semantics of masked objects based on the visible patches, and then learn the correlation of objects among different frames and try to estimate their accurate motions. We name the proposed mask-and-predict task as Masked Motion Encoding (MME)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "content": "Moreover, to help the model learn fine-grained motion details, we further propose to interpolate the motion trajectory. Taking sparsely sampled video as input, the model is asked to reconstruct spatially and temporally dense motion trajectories. This is inspired by the video frame interpolation task [77] where a deep model can reconstruct dense video at the pixel level from sparse video input. Different from it, we aim to reconstruct the fine-grained motion details of moving objects, which has higher-level motion information and is helpful for understanding actions. Our main contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 570, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 58, + 570, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 570, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 58, + 570, + 287, + 642 + ], + "type": "text", + "content": "- Existing mask-and-predict task based on appearance reconstruction is hard to learn important temporal clues, which is critical for representing video content. Our Masked Motion Encoding (MME) paradigm overcomes this limitation by asking the model to reconstruct motion trajectory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 654, + 287, + 713 + ], + "type": "text", + "content": "- Our motion interpolation scheme takes a sparsely sampled video as input and then predicts dense motion trajectory in both spatial and temporal dimensions. This scheme endows the model to capture long-term and fine-grained motion clues from sparse video input." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "text", + "content": "Extensive experimental results on multiple standard video recognition benchmarks prove that the representations learned from the proposed mask-and-predict task achieve state-of-the-art performance on downstream action recognition tasks. Specifically, pre-trained on Kinetics-400 [10], our MME brings the gain of " + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "text", + "content": " on Something-Something V2 [34], " + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "text", + "content": " on Kinetics-400, " + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "text", + "content": " on UCF101 [59], and " + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 304, + 231, + 545, + 327 + ], + "type": "text", + "content": " on HMDB51 [44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 344, + 392, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 392, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 392, + 357 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "type": "text", + "content": "Self-supervised Video Representation Learning. Self-supervised video representation learning aims to learn discriminative video features for various downstream tasks in the absence of accurate video labels. To this end, most of the existing methods try to design an advanced pretext task like predicting the temporal order of shuffled video crops [78], perceiving the video speediness [7, 13] or solving puzzles [43, 49]. In addition, contrastive learning is also widely used in this domain [14, 16, 36, 41, 46, 54, 57, 68, 69], which constrains the consistency between different augmentation views and brings significant improvement. In particular, ASCNet and CoCLR [36, 41] focus on mining hard positive samples in different perspectives. Optical flow has also been proven to be effective in capturing motion information [70, 76]. Besides, tracking video objects' movement is also used in self-supervised learning [17, 18, 65, 73, 74]. Among them, Wang et al. [74] only utilizes the spatial encoder to extract frame appearance information. CtP framework [65] and Siamese-triplet network [73] only require the model to figure out the position and size changes of a specific video patch. Different from these methods, our proposed MME trace the fine-grained movement and shape changes of different parts of objects in the video, hence resulting in a superior video representation. Tokmakov et al. [63] utilize Dense Trajectory to provide initial pseudo labels for video clustering. However, the model does not predict trajectory motion features explicitly. Instead, we consider long-term and fine-grained trajectory motion features as explicit reconstruction targets." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2236" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 209, + 185 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 209, + 185 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 209, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 209, + 185 + ], + "type": "image", + "image_path": "a4b473440a755a23c6769f6da56003c8dbe69b01002fb2d4fc4c4647893b4768.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 239 + ], + "type": "text", + "content": "Figure 2. Overview of Masked Motion Encoding (MME). Given a sparsely sampled video, we first divide it into several patches and randomly mask out some of them. And then, we feed the remaining patches to a ViT encoder to extract video representation. Last, a lightweight ViT decoder is involved to predict the content in the masked region, i.e., a motion trajectory containing position changes and shape changes of moving objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 70, + 313, + 185 + ], + "blocks": [ + { + "bbox": [ + 211, + 70, + 313, + 185 + ], + "lines": [ + { + "bbox": [ + 211, + 70, + 313, + 185 + ], + "spans": [ + { + "bbox": [ + 211, + 70, + 313, + 185 + ], + "type": "image", + "image_path": "44625b59b73d1a3c19cca9691457b126a13c597b7f1cc929e32566029637542d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 545, + 185 + ], + "type": "image", + "image_path": "0654f83de7f1007f192b32769cf0f8c4e5f6d8475731897f8b64b741ca50fe6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 247, + 289, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 289, + 536 + ], + "type": "text", + "content": "Mask Modeling for Vision Transformer. Recently, BEiT and MAE [5, 37] show two excellent mask image modeling paradigms. BEVT [72] and VideoMAE [64] extend these two paradigms to the video domain. To learn visual representations, BEVT [72] predicts the discrete tokens generated by a pre-trained VQ-VAE tokenizer [56]. Nevertheless, pre-training such a tokenizer involves an unbearable amount of data and computation. In contrast, VideoMAE [64] pretrain the Vision Transformer by regressing the RGB pixels located in the masked tubes of videos. Due to asymmetric encoder-decoder architecture and extremely high masking ratio, model pre-training is more efficient with VideoMAE. Besides, MaskFeat [75] finds that predicting the Histogram of Gradient (HOG [21]) of masked video contents is a strong objective for the mask-and-predict paradigm. Existing methods only consider static information in each video frame, thus the model can speculate the masked area by watching the visible area in each frame independently and failed to learn important temporal clues (cf. Section 3.1). Different from the video prediction methods [35, 51, 58, 60] that predict the future frames in pixel or latent space, our Masked Motion Encoding paradigm predicts incoming fine-grained motion in masked video regions, including position changes and shape changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 546, + 153, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 153, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 153, + 560 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 566, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 628 + ], + "type": "text", + "content": "We first revisit the current masked video modeling task for video representation learning (cf. Section 3.1). Then, we introduce our masked motion encoding (MME), where we change the task from recovering appearance to recovering motion trajectory (cf. Section 3.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 241, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 241, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 241, + 648 + ], + "type": "text", + "content": "3.1. Rethinking Masked Video Modeling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Given a video clip sampled from a video, self-supervised video representation learning aims to learn a feature encoder " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{enc}(\\cdot)" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " that maps the clip to its corresponding feature that best describes the video. Existing masked video modeling methods [64, 75] attempt to learn such a feature" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "type": "text", + "content": "encoder through a mask-and-predict task. Specifically, the input clip is first divided into multiple non-overlapped 3D patches. Some of these patches are randomly masked and the remaining patches are fed into the feature encoder, followed by a decoder " + }, + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "type": "inline_equation", + "content": "f_{dec}(\\cdot)" + }, + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "type": "text", + "content": " to reconstruct the information in the masked patches. Different works aim to reconstruct different information (e.g., raw pixel in VideoMAE [64] and HOG in MaskFeat [75])." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 344, + 546, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 546, + 477 + ], + "type": "text", + "content": "However, the existing works share a common characteristic where they all attempt to recover static appearance information of the masked patches. Since an image with a high masking ratio (85% in MAE [37]) can be well reconstructed [37,75], we conjecture that the masked appearance information of a video can also be reconstructed frame by frame independently. In this sense, the model may focus more on the contents in the same frame. This may hinder the models from learning important temporal clues, which is critical for video representation. We empirically study this conjecture in the ablation study (cf. Section 4.2.1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 483, + 448, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 448, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 448, + 495 + ], + "type": "text", + "content": "3.2. General Scheme of MME" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 501, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 545, + 681 + ], + "type": "text", + "content": "To better learn temporal clues of a video, our MME changes the reconstruction content from static appearance information to object motion information, including position and shape change of objects. As shown in Fig. 2, a video clip is sparsely sampled from a video and is divided into a number of non-overlapped 3D patches of size " + }, + { + "bbox": [ + 304, + 501, + 545, + 681 + ], + "type": "inline_equation", + "content": "t \\times h \\times w" + }, + { + "bbox": [ + 304, + 501, + 545, + 681 + ], + "type": "text", + "content": ", corresponding to time, height, and width. We follow VideoMAE [64] to use the tube masking strategy, where the masking map is the same for all frames, to mask a subset of patches. For computation efficiency, we follow MAE [37] to only feed the unmasked patches (and their positions) to the encoder. The output representation together with learnable [MASK] tokens are fed to a decoder to reconstruct motion trajectory " + }, + { + "bbox": [ + 304, + 501, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 501, + 545, + 681 + ], + "type": "text", + "content": " in the masked patches. The training loss for MME is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 690, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 690, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 384, + 690, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i \\in \\mathcal {I}} \\left| \\mathbf {z} _ {i} - \\hat {\\mathbf {z}} _ {i} \\right| ^ {2}, \\tag {1}", + "image_path": "fdb634affdece40c4c03be8079073314a9c9f51e03b51b37a41d888a5789f562.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2237" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " is the predicted motion trajectory, and " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " is the index set of motion trajectories in all masked patches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 288, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 288, + 240 + ], + "type": "text", + "content": "Motivated by the fact that we humans recognize actions by perceiving position changes and shape changes of moving objects, we leverage these two types of information to represent the motion trajectory. Through pre-training on MME task, the model is endowed with the ability to explore important temporal clues. Another important characteristic of the motion trajectory is that it contains fine-grained motion information extracted at raw video rate. This fine-grained motion information provides the model with a supervision signal to anticipate fine-grained action from sparse video input. In the following, we will introduce the proposed motion trajectory in detail." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 249, + 204, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 204, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 204, + 262 + ], + "type": "text", + "content": "3.3. Motion Trajectory for MME" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 268, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 388 + ], + "type": "text", + "content": "The motion of moving objects can be represented in various ways such as optical flow [27], histograms of optical flow (HOF) [45], and motion boundary histograms (MBH) [45]. However, these descriptors can only represent short-term motion between two adjacent frames. We hope our motion trajectory represents long-term motion, which is critical for video representation. To this end, inspired by DT [66], we first track the moving object in the following " + }, + { + "bbox": [ + 46, + 268, + 287, + 388 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 268, + 287, + 388 + ], + "type": "text", + "content": " frames to cover a longer range of motion, resulting in a trajectory " + }, + { + "bbox": [ + 46, + 268, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 268, + 287, + 388 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 399, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 399, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 110, + 399, + 287, + 411 + ], + "type": "interline_equation", + "content": "\\mathbf {T} = \\left(\\mathbf {p} _ {t}, \\mathbf {p} _ {t + 1}, \\dots , \\mathbf {p} _ {t + L}\\right), \\tag {2}", + "image_path": "7282b5b9c80b3256ddf0a7fa364a8d103e1a162e6af466aa81b5403e1e6fc156.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t = (x_t, y_t)" + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "content": " represents a point located at " + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "inline_equation", + "content": "(x_t, y_t)" + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "content": " of frame " + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "inline_equation", + "content": "(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "content": " indicates the concatenation operation. Along this trajectory, we fetch the position features " + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^p" + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "content": " and shape features " + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^s" + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "content": " of this object to compose a motion trajectory " + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 422, + 287, + 481 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 483, + 287, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 287, + 495 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\left(\\mathbf {z} ^ {p}, \\mathbf {z} ^ {s}\\right). \\tag {3}", + "image_path": "3121093e1187844ae7ba20d20c4fa08baff43e61d1b0cf23ee3527fb430dbb4e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 551 + ], + "type": "text", + "content": "The position features are represented by the position transition relative to the last time step, while the shape features are the HOG descriptors of the tracked object in different time steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": "Tracking objects using spatially and temporally dense trajectories. Some previous works [2, 18, 19, 24] try to use one trajectory to represent the motion of an individual object. In contrast, DT [66] points out that tracking spatially dense feature points sampled on a grid space performs better since it ensures better coverage of different objects in a video. Following DT [66], we use spatially dense grid points as the initial position of each trajectory. Specifically, we uniformly sample " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": " points in a masked patch of size " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "t \\times h \\times w" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": ", where each point indicates a part of an object. For each point, we track it through temporally dense " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": " frames according to the dense optical flow, resulting in " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": " trajectories. In this way, the model is able to capture spatially and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "temporally dense motion information of objects through the mask-and-predict task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "type": "text", + "content": "As a comparison, the reconstruction contents in existing works [64, 75] often extract temporally sparse videos sampled with a large stride " + }, + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "type": "inline_equation", + "content": "s > 1" + }, + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "type": "text", + "content": ". The model takes as input a sparse video and predicts these sparse contents for learning video representation. Different from these works, our model also takes as input sparse video but we push the model to interpolate motion trajectory containing fine-grained motion information. This simple trajectory interpolation task does not increase the computational cost of the video encoder but helps the model learn more fine-grained action information even given sparse video as input. More details about dense flow calculating and trajectory tracking can be found in Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 331 + ], + "type": "text", + "content": "Representing position features. Given a trajectory " + }, + { + "bbox": [ + 304, + 258, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 258, + 545, + 331 + ], + "type": "text", + "content": " consisting of the tracked object position at each frame, we are more interested in the related movement of objects instead of their absolute location. Consequently, we represent the position features with related movement between two adjacent points " + }, + { + "bbox": [ + 304, + 258, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_t = \\mathbf{p}_{t+1} - \\mathbf{p}_t" + }, + { + "bbox": [ + 304, + 258, + 545, + 331 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 369, + 340, + 545, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 340, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 369, + 340, + 545, + 353 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {p} = \\left(\\Delta \\mathbf {p} _ {t}, \\dots , \\Delta \\mathbf {p} _ {t + L - 1}\\right), \\tag {4}", + "image_path": "fbac46fad70d42ddbcaa233de85a26ded5911fa5f54e3745daaaaa921ae58266.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^p" + }, + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "inline_equation", + "content": "L\\times 2" + }, + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "text", + "content": " dimensional feature. As each patch contains " + }, + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 361, + 545, + 399 + ], + "type": "text", + "content": " position features, we concatenate and normalize them as position features part of the motion trajectory." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 405, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 545, + 464 + ], + "type": "text", + "content": "Representing shape features. Besides embedding the movement, the model also needs to be aware of the shape changes of objects to recognize actions. Inspired by HOG [21], we use histograms of oriented gradients (HOG) with 9 bins to describe the shape of objects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 464, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 536 + ], + "type": "text", + "content": "Compared with existing works [64, 75] that reconstruct HOG in every single frame, we are more interested in the dynamic shape changes of an object, which can better represent action in a video. To this end, we follow DT [66] to calculate trajectory-aligned HOG, consisting of HOG features around all tracked points in a trajectory, i.e.," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 545, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 545, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 347, + 545, + 545, + 559 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {s} = \\left(\\operatorname {H O G} \\left(\\mathbf {p} _ {t}\\right), \\dots , \\operatorname {H O G} \\left(\\mathbf {p} _ {t + L - 1}\\right)\\right), \\tag {5}", + "image_path": "c473d5094aae51708c15677133df2ef752d94199460092f243e4bc3aa75e1a20.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{HOG}(\\cdot)" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " is the HOG descriptor and " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^s" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "L\\times 9" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " dimensional feature. Also, as one patch contains " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " trajectories, we concatenate " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " trajectory-aligned HOG features and normalize them to the standard normal distribution as the shape features part of the motion trajectory." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 638, + 388, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 638, + 388, + 652 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 388, + 652 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. We conduct experiments on Kinetics-400 (K400), Something-Something V2 (SSV2), UCF101 and HMDB51 datasets. Unless otherwise stated, we follow previous trails [64] and feed the model a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2238" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 521, + 265 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 521, + 265 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 521, + 265 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 521, + 265 + ], + "type": "table", + "html": "
MethodBackbonePre-trainLabelEpochFramesGFLOPsParam.Acc.@1
TimeSformer [8]ViT-BIN-21K✓158196×1×3121M59.5
ViViT-L [3]ViT-LIN-21K+K400✓35163992 × 3 × 4311M65.4
Motionformer [53]ViT-BIN-21K+K400✓3516370 × 1 × 3109M66.5
SlowFast [29]ResNet-101K400✓1968+32106 × 1 × 353M63.1
MViT [26]MViT-BK400✓10064455 × 1 × 337M67.7
SVT [57]ViT-BIN-21K+K400×208+64196×1×3121M59.6
LSTCL [68]Swin-BK400×20016360×5×188M67.0
BEVT [72]Swin-BK400+DALL-E×15032282×1×388M67.1
VIMPAC [62]ViT-LHowTo100M+DALLE×10010N/A×10×3307M68.1
VideoMAE [64]ViT-BSSv2×40016180×2×387M67.6
VideoMAE [64]ViT-BSSv2×80016180×2×387M69.3
VideoMAE [64]ViT-BK400×80016180×2×387M68.2
MME (Ours)ViT-BSSv2×40016180×2×387M69.2
MME (Ours)ViT-BSSv2×80016180×2×387M70.0
MME (Ours)ViT-BK400×40016180×2×387M69.3
MME (Ours)ViT-BK400×80016180×2×387M70.5
", + "image_path": "37a2c3ca831e259e043ce9953161e6a3787deb950f95d3d034a0bfc5f3e3359f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 307, + 520, + 465 + ], + "blocks": [ + { + "bbox": [ + 46, + 274, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 545, + 297 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-arts on Something-Something V2. Our MME outperforms the previous best method. Remarkably, MME show better transferability ability (transfer from K400 to SSv2) compared with VideoMAE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 307, + 520, + 465 + ], + "lines": [ + { + "bbox": [ + 75, + 307, + 520, + 465 + ], + "spans": [ + { + "bbox": [ + 75, + 307, + 520, + 465 + ], + "type": "table", + "html": "
MethodBackbonePre-trainLabelEpochFramesGFLOPsParam.Acc.@1
TimeSformer [8]ViT-BIN-21K✓158196×1×3121M78.3
Motionformer [53]ViT-BIN-21K✓3516370×10×3121M79.7
Motionformer [53]ViT-HRIN-21K✓3516959×10×3382M81.1
VIMPAC [62]ViT-LHowTo100M+DALLE×10010N/A×10×3307M77.4
SVT [57]ViT-BIN-21K+K400×208+64196×1×3121M78.1
BEVT [72]Swin-BK400+DALL-E×15032282×1×388M76.2
BEVT [72]Swin-BIN21K+K400+DALL-E×15032282×1×388M80.6
LSTCL [68]Swin-BK400×20016360×5×188M81.5
OmniMAE [33]ViT-BIN1K+K400×160032180×5×387M80.6
VideoMAE [64]ViT-BK400×160016180×5×387M80.9
ST-MAE [28]ViT-BK400×160016180×7×387M81.3
MME (Ours)ViT-BK400×160016180×7×387M81.8
", + "image_path": "cee3848ff4e33e251d74b08968a87bdcd2096728681833df99439ed26e7eed90.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 473, + 501, + 484 + ], + "lines": [ + { + "bbox": [ + 91, + 473, + 501, + 484 + ], + "spans": [ + { + "bbox": [ + 91, + 473, + 501, + 484 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-arts on Kinetics-400. Our MME outperforms the previous best method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 500, + 288, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 288, + 608 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 500, + 288, + 608 + ], + "type": "text", + "content": " 16-frame clip with strides 4 and 3 for K400 and SSV2 respectively. We use vanilla ViT-Base [25] as the backbone in all experiments. As for the motion trajectory, we set the trajectory length to " + }, + { + "bbox": [ + 46, + 500, + 288, + 608 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 46, + 500, + 288, + 608 + ], + "type": "text", + "content": " and the number of trajectories in each patch to " + }, + { + "bbox": [ + 46, + 500, + 288, + 608 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 46, + 500, + 288, + 608 + ], + "type": "text", + "content": ". Our models are trained on 16 NVIDIA GeForce RTX 3090 GPUs. Note that for all the ablation studies, we use a smaller pre-training dataset and reduce the input patches of fine-tuning to speed up the experiments, details can be found in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 621, + 234, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 234, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 234, + 634 + ], + "type": "text", + "content": "4.1. Comparisons with State-of-the-arts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Fine-tuning. On the temporal-heavy dataset SSv2, we conduct experiments using two settings, namely in-domain pre-training (i.e., pre-train and fine-tune on the same dataset) and cross-domain pre-training (i.e., pre-train and fine-tune on different datasets). In Tab. 1, we have two main observations. 1) Our MME significantly performs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "type": "text", + "content": "better than previous state-of-the-arts VideoMAE [64] on both in-domain and cross-domain settings. Specifically, under cross-domain settings (transferring from K400 to SSv2) with 800 epochs, our MME outperforms it by " + }, + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "type": "text", + "content": ", increasing the accuracy from " + }, + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "type": "inline_equation", + "content": "68.2\\%" + }, + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "type": "inline_equation", + "content": "70.5\\%" + }, + { + "bbox": [ + 304, + 500, + 545, + 679 + ], + "type": "text", + "content": ". 2) Our MME performs better under cross-domain setting compared with indomain setting, which is contrary to the existing observation on VideoMAE [64] and ST-MAE [28]. We attribute it to our masked motion encoding pretext task that helps models focus more on learning temporal clues instead of learning domain-specific appearance information (e.g., color, lighting). The learned temporal clues are more universal across different domains for video representation. This also shows the potential for taking advantage of larger datasets for pretraining, which we leave for future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In Tab. 2, our MME also achieves state-of-the-art performance on another large-scale dataset K400 with a gain" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2239" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 288, + 169 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 288, + 169 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 288, + 169 + ], + "type": "table", + "html": "
MethodBackbonePre-trainUCF101HMDB51
XDC [1]R(2+1)DIG65M94.267.1
GDT [52]R(2+1)DIG65M95.272.8
CVRL [54]R50K40092.967.9
CORPf [39]R50K40093.568.0
ρBYOL [30]R50K40094.272.1
VideoMAE [64]ViT-BK40096.173.3
MME (Ours)ViT-BK40096.578.0
", + "image_path": "723550ed243c0b9b8d7b77a4327fef33ee76300a37ec43f0e10f81901e91b342.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "type": "text", + "content": ". In Tab. 3, we also evaluate the transferability of MME by transferring a K400 pre-trained model to two small-scale datasets, namely UCF101 and HMDB51. Our MME achieves a large gain of " + }, + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "type": "text", + "content": " on HMDB51 and " + }, + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 46, + 212, + 287, + 297 + ], + "type": "text", + "content": " on UCF101 compared with VideoMAE. This further demonstrates the strong ability of MME to learn from the large-scale dataset and transfer it to a small-scale one." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "text", + "content": "Linear probing. We further evaluate our MME under liner probing setting on the SSv2 dataset as this dataset focuses more on the action. We follow SVT [57] to fix the transformer backbone and train a linear layer for 20 epochs. In Tab. 4, our method significantly outperform previous contrastive-based method [57] by " + }, + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "inline_equation", + "content": "10.9\\%" + }, + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "text", + "content": " and masked video modeling methods [64] by " + }, + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "text", + "content": ". These results indicate that the representation learned from our MME task contains more motion clues for action recognition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 419, + 270, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 270, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 270, + 432 + ], + "type": "text", + "content": "4.2. Ablation Study on Reconstruction Content" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "content": "In this section, we investigate the effect of different reconstruction contents, including appearance, short-term motion, and the proposed long-term fine-grained motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 491, + 248, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 248, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 248, + 502 + ], + "type": "text", + "content": "4.2.1 Evaluating Appearance Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 676 + ], + "type": "text", + "content": "In section 3.1, we conjecture that appearance reconstruction is finished in each video frame independently. To verify this conjecture, we corrupt the temporal information of a video by randomly shuffling its frames. Since the video frames are randomly permuted, it is hard for a model to leverage information from other frames to reconstruct the current frame through their temporal correlation. We mask " + }, + { + "bbox": [ + 46, + 510, + 287, + 676 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 676 + ], + "type": "text", + "content": " regions of these videos following VideoMAE [64]. In Fig. 3 (a), we find that the reconstruction error (i.e., L2 error between predicted and ground-truth pixel) converges to a low value. We also conduct the same experiment using raw video without shuffling. The model undergoes a similar convergence process. This demonstrates that the masked pixels are well reconstructed without temporal information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "To further evaluate whether the model pre-trained on the appearance reconstruction task (i.e., VideoMAE) can well capture important temporal clues, we transfer two" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 541, + 144 + ], + "blocks": [ + { + "bbox": [ + 47, + 176, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 176, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 287, + 198 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-arts on UCF101 and HMDB51." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 70, + 541, + 144 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 144 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 144 + ], + "type": "table", + "html": "
MethodBackbonePre-trainAcc.@1
TimeSformer [8]ViT-BIN-21K14.0
SVT [57]ViT-BIN-21K+K40018.3
VideoMAE [64]ViT-BSSv224.3
MME (Ours)ViT-BSSv229.2
", + "image_path": "5b62ca779f755a39e4f3055e3f7858054cac5fd80c6e571830b05fa753f1dffe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 152, + 529, + 163 + ], + "lines": [ + { + "bbox": [ + 321, + 152, + 529, + 163 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 529, + 163 + ], + "type": "text", + "content": "Table 4. Linear Probing on Something-Something V2." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 169, + 424, + 277 + ], + "blocks": [ + { + "bbox": [ + 309, + 169, + 424, + 277 + ], + "lines": [ + { + "bbox": [ + 309, + 169, + 424, + 277 + ], + "spans": [ + { + "bbox": [ + 309, + 169, + 424, + 277 + ], + "type": "image", + "image_path": "56be86abe4159245519ff5cb14764bb11b4932b0cc46e0cd4ae25e501ca43177.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 281, + 430, + 291 + ], + "lines": [ + { + "bbox": [ + 310, + 281, + 430, + 291 + ], + "spans": [ + { + "bbox": [ + 310, + 281, + 430, + 291 + ], + "type": "text", + "content": "(a) Masked video reconstruction error." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 300, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 357 + ], + "type": "text", + "content": "Figure 3. We perform mask-and-predict pre-training using temporal shuffled videos and raw videos and then transfer them to the downstream task. (a) Masked appearance (i.e., pixels) is well reconstructed without temporal information. (b) Our MME outperforms VideoMAE when both provided temporal information." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 424, + 170, + 543, + 277 + ], + "blocks": [ + { + "bbox": [ + 424, + 170, + 543, + 277 + ], + "lines": [ + { + "bbox": [ + 424, + 170, + 543, + 277 + ], + "spans": [ + { + "bbox": [ + 424, + 170, + 543, + 277 + ], + "type": "image", + "image_path": "3e3244eb30176476d083000ccfc9011e7d083583d70fbb073b1d5d56688315eb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 281, + 535, + 292 + ], + "lines": [ + { + "bbox": [ + 441, + 281, + 535, + 292 + ], + "spans": [ + { + "bbox": [ + 441, + 281, + 535, + 292 + ], + "type": "text", + "content": "(b) Downstream performance." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "text", + "content": "VideoMAE models (pre-trained on shuffled videos and raw videos, respectively) to the downstream action recognition task. In Fig. 3 (b), these two models perform competitively, indicating that removing the temporal information from the pre-training videos barely affects the learned video representation. We speculate this is because the VideoMAE model pays little attention to temporal information when performing the appearance reconstruction pre-training task. As a comparison, our MME paradigm performs significantly better than VideoMAE when the temporal information is provided for pre-training " + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "inline_equation", + "content": "(64.1\\%" + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "inline_equation", + "content": "60.9\\%)" + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "text", + "content": ". Once the temporal information is removed, MME performs similarly with VideoMAE " + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "inline_equation", + "content": "(60.7\\%" + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "inline_equation", + "content": "60.9)" + }, + { + "bbox": [ + 304, + 372, + 546, + 553 + ], + "type": "text", + "content": ". This demonstrates our MME paradigm takes great advantage of temporal information in the pre-training phase." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 568, + 487, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 568, + 487, + 580 + ], + "spans": [ + { + "bbox": [ + 306, + 568, + 487, + 580 + ], + "type": "text", + "content": "4.2.2 Evaluating Motion Reconstruction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 587, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 545, + 611 + ], + "type": "text", + "content": "In the following ablations, we separately verify the effectiveness of two proposals of our methods:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 619, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 619, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 305, + 619, + 545, + 654 + ], + "type": "text", + "content": "Effectiveness of motion trajectory. We compare our motion trajectory with two appearance reconstruction baselines and three short-term motion reconstruction baselines." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 665, + 547, + 713 + ], + "type": "text", + "content": "- Appearance reconstruction baselines: 1) HOG: predicting HOG features [21] in the middle frame of each 3D patch [75]. 2) Pixel: predicting all the pixels of each 3D patch [28, 37, 64]." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2240" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 70, + 259, + 180 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 259, + 180 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 259, + 180 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 259, + 180 + ], + "type": "table", + "html": "
Reconstruction contentAcc.@1Acc.@5
Pixel61.186.6
HOG61.186.9
HOG + Flow60.986.4
HOG + HOF61.386.6
HOG + MBH61.486.8
HOG + Traj. w/o shape63.588.2
Traj. (ours)64.188.4
", + "image_path": "12767c128a5d3e7632592433d86f3744e3cc4a13dafcf14f2b71e14b6ba20ec3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 218, + 287, + 319 + ], + "blocks": [ + { + "bbox": [ + 47, + 189, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 47, + 189, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 189, + 287, + 212 + ], + "type": "text", + "content": "Table 5. Comparison of different Reconstruction content. Our motion trajectory outperforms all the other contents." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 218, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 48, + 218, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 287, + 319 + ], + "type": "table", + "html": "
Reconstruction contentInterpolationAcc.@1Acc.@5
HOGx61.186.9
HOG✓60.986.8
HOG + MBHx61.386.8
HOG + MBH✓61.486.8
Traj. (ours)x62.287.5
Traj. (ours)✓64.188.4
", + "image_path": "308e7d11215a368eb63c95bb87d07581c2ce8fb2d4891b36f7743a927a775847.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 327, + 287, + 360 + ], + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 360 + ], + "type": "text", + "content": "Table 6. Effectiveness of trajectory interpolation. Interpolating the motion trajectory brings significant improvement compared to the baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 59, + 377, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 377, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 59, + 377, + 287, + 449 + ], + "type": "text", + "content": "- Short-term motion reconstruction baselines: 1) Flow: predicting the dense optical flow map of the masked patches. 2) HOF: predicting the histogram of optical flow orientation. 3) MBH: predicting the motion boundary histogram, which is the second-order local statistic of optical flow." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": "In Tab. 5, two types of appearance reconstruction contents (i.e., Pixel and HOG) perform similarly. Based on the HOG baseline, predicting additional short-term motion contents brings slight improvement (at least by " + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": ") except for directly predicting optical flow (decreases by " + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": "). We suspect that predicting the dense flow map for each pixel is too hard for a model and thus harms the performance. Instead, the histogram version of flow (i.e., HOF and MBH) help to reduce the prediction ambiguity [75], resulting in better performance. Furthermore, after adding our slight variant of motion trajectory (i.e., with position features but without shape features shown in Eq. (3)), the performance improves significantly compared to both appearance reconstruction baselines (by " + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": ") and short-term motion reconstruction baselines (by " + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": "). On top of this variant, adding trajectory-aligned shape features further brings additional gain by " + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": ". Because tracking the trajectory of objects in multiple frames provides long-term motion compared to short-term motion reconstruction, we believe the learning of long-term motion clues is one of the keys to enhancing video representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": "Effectiveness of trajectory interpolation. In MME, we feed temporally sparse video to the model and push it to interpolate temporally dense trajectories. We are interested in whether interpolating appearance (i.e., HOG) or short-term motion contents (i.e., MBH) benefit learning fine-grained motion details. To this end, we conduct experiments to predict dense HOG and MBH, i.e., predicting these features of all frames within a masked patch even though these frames are not sampled as input. In Tab. 6, interpolating HOG or MBH brings little improvement. We suspect this is because these dense features represent the continuous movement of objects poorly, as they have not explicitly tracked the moving objects. In contrast, our temporally dense motion trajectory tracks the object densely, providing temporal fine-grained information for learning. The trajectory interpolation improves our MME by " + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": ". This result proves that the learning of fine-grained motion details is another key to enhancing video representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 295, + 479, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 295, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 479, + 308 + ], + "type": "text", + "content": "4.3. Ablation Study on MME Design" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to investigate the effect of different parameters for MME pre-training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "text", + "content": "Trajectory length. Trajectory length " + }, + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "text", + "content": " indicates how many frames we track an object for extracting the motion trajectory. If " + }, + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "inline_equation", + "content": "L = 0" + }, + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "text", + "content": ", our method degrades to MaskFeat [75], which only predicts HOG features in the masked patch. In Tab. 7a, MME achieves the best performance when " + }, + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "text", + "content": ". This is because a longer trajectory provides more motion information for learning. But if the trajectory is too long, the drift of flow introduces accumulated noise when tracking the objects, which harms the performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "type": "text", + "content": "Trajectory density. In our experiments, we spatially dense track " + }, + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "type": "text", + "content": " motion trajectories. We conduct experiments to evaluate whether it is necessary to densely track multiple trajectories in a patch. Specifically, we assume only one object exists in a patch, and thus the 4 trajectories are similar. We merge these 4 trajectories as one by averaging or selecting the most salient one. We then consider the merged trajectories as the trajectory motion. In Tab. 7c, using dense trajectories significantly outperform the merged baselines. We suspect the spatial dense trajectories provide more motion information for learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Trajectory normalization. Due to the irregular distribution of motion and shape information in the video, the trajectory values in different patches vary a lot. To help the model better model the trajectory, we adopt the patch norm method to normalize the trajectory values in each patch into a standard normal distribution, as has also been done in MAE [37]. The mean and variance are calculated by all " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " trajectories in a patch. Tab. 7b shows the effectiveness of adopting the patch norm." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2241" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 176, + 127 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 176, + 127 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 176, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 176, + 127 + ], + "type": "table", + "html": "
LengthAcc.@1Acc.@5
L = 463.288.0
L = 664.188.4
L = 863.388.0
", + "image_path": "360003841f901d4ac3e55ee1440110715b7c63848ce26848b0fda1e06ae1377b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 193, + 70, + 349, + 125 + ], + "blocks": [ + { + "bbox": [ + 59, + 129, + 173, + 140 + ], + "lines": [ + { + "bbox": [ + 59, + 129, + 173, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 129, + 173, + 140 + ], + "type": "text", + "content": "(a) Effect of trajectory length." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 193, + 70, + 349, + 125 + ], + "lines": [ + { + "bbox": [ + 193, + 70, + 349, + 125 + ], + "spans": [ + { + "bbox": [ + 193, + 70, + 349, + 125 + ], + "type": "table", + "html": "
SettingAcc.@1Acc.@5
w/o patch norm.62.187.44
w/ patch norm.64.188.4
", + "image_path": "776b33282d4c341504062f9acc5f457986666082f547bb8d1e60d18ac1927db2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 365, + 70, + 529, + 127 + ], + "blocks": [ + { + "bbox": [ + 196, + 128, + 340, + 140 + ], + "lines": [ + { + "bbox": [ + 196, + 128, + 340, + 140 + ], + "spans": [ + { + "bbox": [ + 196, + 128, + 340, + 140 + ], + "type": "text", + "content": "(b) Effect of trajectory normalization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 365, + 70, + 529, + 127 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 529, + 127 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 529, + 127 + ], + "type": "table", + "html": "
Trajectory densityAcc.@1Acc.@5
Dense64.188.4
Sparse (Max)60.286.3
Sparse (Mean)61.887.1
", + "image_path": "2245de0dfa20d557d11d9c8344a7aefc1e416471d7a9bdf7f780ef66388d0151.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 153, + 175, + 221 + ], + "blocks": [ + { + "bbox": [ + 386, + 129, + 509, + 141 + ], + "lines": [ + { + "bbox": [ + 386, + 129, + 509, + 141 + ], + "spans": [ + { + "bbox": [ + 386, + 129, + 509, + 141 + ], + "type": "text", + "content": "(c) Effect of trajectories density." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 153, + 175, + 221 + ], + "lines": [ + { + "bbox": [ + 56, + 153, + 175, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 175, + 221 + ], + "type": "table", + "html": "
StrideAcc.@1Acc.@5
s = 462.987.7
s = 364.188.4
s = 263.588.3
s = 162.787.7
", + "image_path": "49fc1dd7758d43aa1f495bc0318bd64b24caf9c4ea3217c5b6a3ac9ea2d50af1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 195, + 153, + 342, + 221 + ], + "blocks": [ + { + "bbox": [ + 61, + 225, + 170, + 236 + ], + "lines": [ + { + "bbox": [ + 61, + 225, + 170, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 170, + 236 + ], + "type": "text", + "content": "(d) Effect of sampling stride." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 153, + 342, + 221 + ], + "lines": [ + { + "bbox": [ + 195, + 153, + 342, + 221 + ], + "spans": [ + { + "bbox": [ + 195, + 153, + 342, + 221 + ], + "type": "table", + "html": "
RatioTypeAcc.@1Acc.@5
90%Tube62.787.8
70%Tube64.188.4
40%Tube61.386.9
40%Cube61.286.9
", + "image_path": "2292f77e0c3a2c0b2816508042af21a6562c2dbcc3b6bc1ee3735e19578e0cc8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 356, + 150, + 543, + 223 + ], + "blocks": [ + { + "bbox": [ + 214, + 225, + 324, + 237 + ], + "lines": [ + { + "bbox": [ + 214, + 225, + 324, + 237 + ], + "spans": [ + { + "bbox": [ + 214, + 225, + 324, + 237 + ], + "type": "text", + "content": "(e) Effect of spatial masking." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 356, + 150, + 543, + 223 + ], + "lines": [ + { + "bbox": [ + 356, + 150, + 543, + 223 + ], + "spans": [ + { + "bbox": [ + 356, + 150, + 543, + 223 + ], + "type": "table", + "html": "
TypeDepthFLOPsAcc.@1Acc.@5
Joint151G64.188.4
255G63.488.3
Sep.156G63.488.4
265G63.388.0
", + "image_path": "f4b51b609a4b94add680eb84d14b4b1da7405489f84a324fb03f5488887b81af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 362, + 225, + 537, + 236 + ], + "lines": [ + { + "bbox": [ + 362, + 225, + 537, + 236 + ], + "spans": [ + { + "bbox": [ + 362, + 225, + 537, + 236 + ], + "type": "text", + "content": "(f) Effect of decoder. Sep. means two decoders." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 246, + 523, + 258 + ], + "lines": [ + { + "bbox": [ + 67, + 246, + 523, + 258 + ], + "spans": [ + { + "bbox": [ + 67, + 246, + 523, + 258 + ], + "type": "text", + "content": "Table 7. Pre-training details. We explore the detailed setting of our MME pre-training. The default entry is marked in gray ." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "type": "text", + "content": "Sampling stride. In order to control the difficulty of interpolating the motion trajectory, we ablate different temporal sampling strides leading to different levels of the sparse input video. With a medium sampling stride (" + }, + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "type": "inline_equation", + "content": "s = 2" + }, + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "type": "text", + "content": "), our MME achieves the best performance. We also found that the performance decreases when the input video shares the same stride as the dense motion trajectory (" + }, + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "type": "text", + "content": "). One possible reason is that it is difficult for the model to anticipate fine-grained motion details in this case." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 389, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 475 + ], + "type": "text", + "content": "Spatial masking strategy. We ablated study different spatial masking strategies and ratios in Tab. 7e. We start from the default setting of MaskFeat [75] that uses a cube masking strategy with a " + }, + { + "bbox": [ + 46, + 389, + 287, + 475 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 389, + 287, + 475 + ], + "type": "text", + "content": " masking ratio, then we increase the masking ratio and adopt the tube masking strategy proposed by VideoMAE [64]. Using tube masking with a medium masking ratio " + }, + { + "bbox": [ + 46, + 389, + 287, + 475 + ], + "type": "inline_equation", + "content": "(70\\%)" + }, + { + "bbox": [ + 46, + 389, + 287, + 475 + ], + "type": "text", + "content": " performs the best." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 481, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 578 + ], + "type": "text", + "content": "Decoder setting. We use several transformer blocks as decoder to reconstruct motion trajectory. We ablate different architectures of the decoder in Tab. 7f. Using a slight decoder with 1 transformer block performs the best, with the lowest computational cost. Increasing the depth of the decoder or using two parallel decoders to separately predict the position and shape features introduces more computational cost but brings little improvement." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 586, + 168, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 168, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 168, + 598 + ], + "type": "text", + "content": "4.4. Visualization Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We visualize the predicted motion trajectory in Fig. 4. The model is able to reconstruct the spatial and temporally dense trajectories of different object parts from a temporally sparse input video. Even without seeing the object in the current frame (the hand being masked out), the model is able to locate it through contextual patches and accurately estimate its trajectory. We only visualize the position changes since our prediction of HOG features has similar quality to the previous work [75]." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 273, + 536, + 384 + ], + "blocks": [ + { + "bbox": [ + 316, + 273, + 536, + 384 + ], + "lines": [ + { + "bbox": [ + 316, + 273, + 536, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 536, + 384 + ], + "type": "image", + "image_path": "26244ef12f2d3ac2bd9de8767aa9a07b365bc3d25aaa0d4adbdc80fe5a848d1e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 391, + 546, + 425 + ], + "lines": [ + { + "bbox": [ + 304, + 391, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 546, + 425 + ], + "type": "text", + "content": "Figure 4. Visualization of predicted motion trajectory. Each column represents (from left to right): (a) Masked frames; (b) ground truth; (c) prediction; (d) original frame." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 437, + 378, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 378, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 378, + 449 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 459, + 546, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 591 + ], + "type": "text", + "content": "We present Masked Motion Encoding, a simple mask-and-predict paradigm that learns video representation by reconstructing the motion contents of masked regions. In particular, we reconstruct motion trajectory, which tracks moving objects densely to record their position and shape changes. By reconstructing this motion trajectory, the model learns long-term and fine-grained motion clues from sparse input videos. Empirical results show that our method outperforms the state-of-the-art mask-and-predict methods on Something-Something V2, Kinetics-400, UCF-101, and HMDB-51 datasets for action recognition." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 608, + 404, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 404, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 404, + 621 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "This work was partially supported by the National Natural Science Foundation of China (NSFC) (No. 62072190, 62072186), Ministry of Science and Technology Foundation Project (No. 2020AAA0106900), Program for Guangdong Introducing Innovative and Entrepreneurial Teams (No. 2017ZT07X183), Guangdong Basic and Applied Basic Research Foundation (No. 2019B1515130001)." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2242" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, pages 9758-9770, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Nadeem Anjum and Andrea Cavallaro. Multifeature object trajectory clustering for video analysis. IEEE Transactions on Circuits and Systems for Video Technology, 18:1555-1564, 2008. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "text", + "content": "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 1, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 282 + ], + "type": "text", + "content": "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: BERT pre-training of image transformers. In ICLR, 2022. 1, 3, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 316 + ], + "type": "text", + "content": "[6] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417, 2006. 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 319, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 362 + ], + "type": "text", + "content": "[7] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In CVPR, pages 9922-9931, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 364, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 396 + ], + "type": "text", + "content": "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding. In ICML, pages 813-824, 2021. 1, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 287, + 453 + ], + "type": "text", + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "type": "text", + "content": "[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "text", + "content": "[11] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703, 2020. 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "text", + "content": "[12] Peihao Chen, Chuang Gan, Guangyao Shen, Wenbing Huang, Runhao Zeng, and Mingkui Tan. Relation attention for temporal action localization. IEEE Transactions on Multimedia, 22:2723-2733, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "text", + "content": "[13] Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan. Rspnet: Relative speed perception for unsupervised video representation learning. In AAAI, pages 1045-1053, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[14] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas H Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurlPS, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[15] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 29:8292-8302, 2020. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Zhenfang Chen, Lin Ma, Wenhan Luo, Peng Tang, and Kwan-Yee K Wong. Look closer to ground better: Weakly-supervised temporal grounding of sentence in video. arXiv, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[17] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In ACL, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[18] Zhenfang Chen, Jiayuan Mao, Jiajun Wu, Kwan-Yee Kenneth Wong, Joshua B Tenenbaum, and Chuang Gan. Grounding physical concepts of objects and events through dynamic visual reasoning. arXiv, 2021. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[19] Zhenfang Chen, Kexin Yi, Yunzhu Li, Mingyu Ding, Antonio Torralba, Joshua B Tenenbaum, and Chuang Gan. Comphy: Compositional physical reasoning of objects and events from videos. In ICLR, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "text", + "content": "[20] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, pages 702-703, 2020. 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[21] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, pages 886-893, 2005. 3, 4, 6, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "text", + "content": "[22] Navneet Dalal, Bill Triggs, and Cordelia Schmid. Human detection using oriented histograms of flow and appearance. In ECCV, pages 428-441, 2006. 13" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, pages 4171–4186, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[24] Mingyu Ding, Zhenfang Chen, Tao Du, Ping Luo, Josh Tenenbaum, and Chuang Gan. Dynamic visual reasoning by learning differentiable physics models from video and language. In NeurIPS, volume 34, pages 887-899, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 509 + ], + "type": "text", + "content": "[25] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "text", + "content": "[26] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 1, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 555, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 578 + ], + "type": "text", + "content": "[27] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In *SCIA*, pages 363-370, 2003. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 579, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 611 + ], + "type": "text", + "content": "[28] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In NeurIPS, 2022. 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "text", + "content": "[29] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[30] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross B. Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, pages 3299-3309, 2021. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[31] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2243" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "image analysis and automated cartography. Communications of the ACM, 24:381-395, 1981. 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "text", + "content": "[32] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In European Conference on Computer Vision (ECCV), pages 758-775, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "type": "text", + "content": "[33] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 189, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 189, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 189, + 287, + 253 + ], + "type": "text", + "content": "[34] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 299 + ], + "type": "text", + "content": "[35] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martín-Martín, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. arXiv preprint arXiv:2206.11894, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 301, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 301, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 301, + 287, + 334 + ], + "type": "text", + "content": "[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In NeurIPS, pages 5679-5690, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 336, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 336, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 48, + 336, + 287, + 379 + ], + "type": "text", + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 425 + ], + "type": "text", + "content": "[38] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. In CVPR, pages 8126-8135, 2020. 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 428, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 470 + ], + "type": "text", + "content": "[39] Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. In ICCV, pages 7939-7949, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "text", + "content": "[40] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware Graph Convolutional Networks for Video Question Answering. In AAAI, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 509, + 287, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 562 + ], + "type": "text", + "content": "[41] Deng Huang, Wenhao Wu, Weiwen Hu, Xu Liu, Dongliang He, Zhihua Wu, Xiangmiao Wu, Mingkui Tan, and Errui Ding. Ascnet: Self-supervised video representation learning with appearance-speed consistency. In ICCV, pages 8096-8105, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "type": "text", + "content": "[42] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q. Weinberger. Deep networks with stochastic depth. In ECCV, pages 646-661, 2016. 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "text", + "content": "[43] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[44] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, pages 2556-2563, 2011. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[45] Ivan Laptev, Marcin Marszalek, Cordelia Schmid, and Benjamin Rozenfeld. Learning realistic human actions from movies. In CVPR, pages 1-8, 2008. 4, 13" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[46] Yuan Liu, Jingyuan Chen, Zhenfang Chen, Bing Deng, Jianqiang Huang, and Hanwang Zhang. The blessings of unlabeled background in untrimmed videos. In CVPR, pages 6176-6185, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 119, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 140 + ], + "type": "text", + "content": "[47] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 143, + 545, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 164 + ], + "type": "text", + "content": "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 167, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 167, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 545, + 209 + ], + "type": "text", + "content": "[49] Dezhao Luo, Chang Liu, Yu Zhou, Dongbao Yang, Can Ma, Qixiang Ye, and Weiping Wang. Video cloze procedure for self-supervised spatio-temporal learning. In AAAI, pages 11701-11708, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 213, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 266 + ], + "type": "text", + "content": "[50] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035, 2019. 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 269, + 545, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 301 + ], + "type": "text", + "content": "[51] Viorica Patraucean, Ankur Handa, and Roberto Cipolla. Spatio-temporal video autoencoder with differentiable memory. In ICLR Workshop, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "type": "text", + "content": "[52] Mandela Patrick, Yuki M Asano, Polina Kuznetsova, Ruth Fong, Joao F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "type": "text", + "content": "[53] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In NeurIPS, pages 12493-12506, 2021. 1, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 406, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 449 + ], + "type": "text", + "content": "[54] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, pages 6964-6974, 2021. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "type": "text", + "content": "[55] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Xiang Wang, Yuehuan Wang, Yiliang Lv, Changxin Gao, and Nong Sang. Mar: Masked autoencoders for efficient action recognition. arXiv preprint arXiv:2207.11660, 2022. 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 498, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 540 + ], + "type": "text", + "content": "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 544, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 585 + ], + "type": "text", + "content": "[57] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael Ryoo. Self-supervised video transformer. In CVPR, pages 2874-2884, 2022. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "type": "text", + "content": "[58] Adria Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătráucean, Florent Altché, Michal Valko, et al. Broaden your views for self-supervised video learning. In ICCV, pages 1255-1265, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[59] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[60] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, pages 843-852, 2015. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2244" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[61] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "text", + "content": "[62] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "text", + "content": "[63] Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. Unsupervised learning of video representations via dense trajectory clustering. In ECCV 2020 Workshops, pages 404-421. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "type": "text", + "content": "[64] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 3, 4, 5, 6, 8, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 294 + ], + "type": "text", + "content": "[65] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. In CVPR, pages 2563-2572, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "type": "text", + "content": "[66] Heng Wang, Alexander Kläser, Cordelia Schmid, and Cheng-Lin Liu. Dense trajectories and motion boundary descriptors for action recognition. International journal of computer vision, 103:60-79, 2013. 4, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "type": "text", + "content": "[67] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In ICCV, pages 3551-3558, 2013. 12, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "type": "text", + "content": "[68] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. arXiv preprint arXiv:2106.09212, 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[69] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In CVPR, pages 11804-11813, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "text", + "content": "[70] Jiangliu Wang, Jianbo Jiao, Linchao Bao, Shengfeng He, Yunhui Liu, and Wei Liu. Self-supervised spatio-temporal representation learning for videos by predicting motion and appearance statistics. In CVPR, pages 4006-4015, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "text", + "content": "[71] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41:2740-2755, 2018. 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "text", + "content": "[72] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, pages 14733-14743, 2021. 1, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[73] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, pages 2794-2802, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[74] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, pages 2566-2576, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[75] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 1, 3, 4, 6, 7, 8, 13" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 453 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[76] Donglai Wei, Joseph J. Lim, Andrew Zisserman, and William T. Freeman. Learning and using the arrow of time. In CVPR, pages 8052-8060, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[77] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P Allebach, and Chenliang Xu. Zooming slow-mo: Fast and accurate one-stage space-time video super-resolution. In CVPR, pages 3370-3379, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 151, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 195 + ], + "type": "text", + "content": "[78] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In CVPR, pages 10334–10343, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[79] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[80] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "type": "text", + "content": "[81] Runhao Zeng, Chuang Gan, Peihao Chen, Wenbing Huang, Qingyao Wu, and Mingkui Tan. Breaking winner-takes-all: Iterative-winners-out networks for weakly supervised temporal action localization. IEEE Transactions on Image Processing, 28(12):5797-5808, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 387 + ], + "type": "text", + "content": "[82] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 387, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 419 + ], + "type": "text", + "content": "[83] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 13" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 420, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 453 + ], + "type": "text", + "content": "[84] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan L. Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2245" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_content_list.json b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8469b4444aa05716c06e23cbc73c2859e1a3db7f --- /dev/null +++ b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_content_list.json @@ -0,0 +1,1752 @@ +[ + { + "type": "text", + "text": "Masked representation learning for domain generalized stereo matching", + "text_level": 1, + "bbox": [ + 122, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhibo Rao $^{1,2\\boxtimes}$ , Bangshu Xiong $^{1}$ , Mingyi He $^{2}$ , Yuchao Dai $^{2}$ , Renjie He $^{2}$ , Zhelun Shen $^{3}$ , Xing Li $^{2\\boxtimes}$ $^{1}$ Nanchang Hangkong University, Nanchang, China \n $^{2}$ Northwestern Polytechnical University, Xi'an, China \n $^{3}$ Baidu Research, Beijing, China", + "bbox": [ + 96, + 178, + 870, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "raoxi36@foxmail.com, xiongs@126.com, {myhe, daiyuchao, davidhrj} @nwpu.edu.cn, 1901213310@pku.edu.cn, lixing36@foxmail.com", + "bbox": [ + 138, + 253, + 825, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, many deep stereo matching methods have begun to focus on cross-domain performance, achieving impressive achievements. However, these methods did not deal with the significant volatility of generalization performance among different training epochs. Inspired by masked representation learning and multi-task learning, this paper designs a simple and effective masked representation for domain generalized stereo matching. First, we feed the masked left and complete right images as input into the models. Then, we add a lightweight and simple decoder following the feature extraction module to recover the original left image. Finally, we train the models with two tasks (stereo matching and image reconstruction) as a pseudomulti-task learning framework, promoting models to learn structure information and to improve generalization performance. We implement our method on two well-known architectures (CFNet and LacGwcNet) to demonstrate its effectiveness. Experimental results on multi-datasets show that: (1) our method can be easily plugged into the current various stereo matching models to improve generalization performance; (2) our method can reduce the significant volatility of generalization performance among different training epochs; (3) we find that the current methods prefer to choose the best results among different training epochs as generalization performance, but it is impossible to select the best performance by ground truth in practice.", + "bbox": [ + 75, + 353, + 473, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 775, + 209, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stereo matching is a challenging research topic of computer vision, which aims to obtain the corresponding pixels between two rectified stereo images [3, 30, 37, 39]. It is essential in many applications, including autonomous [11], augmented reality [35], virtual reality [2], etc.", + "bbox": [ + 75, + 801, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/affa0af305a6708276a5e279a71d68eb70622006b0651033db6c13149d67b3d0.jpg", + "image_caption": [ + "(a) CFNet", + "Figure 1. The generalization performance among different epochs on multi-datasets. When the models converge on the source domain (Sceneflow), the results are stable on the source domain. However, generalization performance has fluctuations on target datasets (KITTI 2012&2015, ETH3D, and Middlebury)." + ], + "image_footnote": [], + "bbox": [ + 511, + 327, + 689, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5329afccfc24509ed3eabff4361888eb8a703cf52f1b6dc6cf066c053b736101.jpg", + "image_caption": [ + "(b) LacGwcNet" + ], + "image_footnote": [], + "bbox": [ + 718, + 327, + 895, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In mainstream deep stereo matching methods, the main steps include four parts: feature extraction, cost volume, feature matching, and disparity regression [10]. To improve the accuracy or speed, researchers proposed many strategies to improve the above four parts [3, 25, 36]. For example, 1) scene awareness modules are applied to the feature extraction to improve accuracy [3, 21]; 2) similarity measurements are employed to enhance physical representation in cost volume construction [23, 26, 36]; 3) various feature matching modules are proposed to solve inefficient calculations or improve accuracy [15, 25, 39]; 4) loss functions of classification or other fields are joined to disparity regression to obtain unimodal results [42]. Although the above methods have achieved significant progress in accuracy or efficiency, they failed to obtain good generalization performance in unseen domains [30, 40].", + "bbox": [ + 496, + 597, + 893, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, many approaches learn domain-invariant representation features to achieve better generalization capability [40]. There are three solutions to accomplish this objective: the unsupervised matching method [24,32,33], domain", + "bbox": [ + 498, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$1 \\boxtimes$ is the corresponding author.", + "bbox": [ + 94, + 886, + 264, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5435", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adaptation techniques [12, 13, 17, 31], and domain generalization approaches [16, 30, 41]. The above solutions reveal that feature presentation is crucial for improving generalization capability [40, 41]. However, all domain generalized matching methods did not mention that the generalization performance varies significantly among different training epochs, as shown in Fig. 1. Namely, the results were unstable. These methods preferred to choose the best results to represent the generalization performance by testing the models of different training epochs. However, we can not employ ground truth (can not obtain) to select the best model among different training epochs in practice. Thus, it is crucial to keep a stable generalization performance.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, many researchers introduced multi-task learning into stereo matching [18, 27]. For instance, Rao et al. proposed a bidirectional guided attention network to cope with semantic segmentation and stereo matching simultaneously [27]. Liu et al. employed the task-shared and task-specific manner to obtain a generalizable representation and used the loss weights to balance all tasks [18]. The multi-task learning proved that feature sharing and weight sharing are essential in different tasks, promoting the learning process and obtaining a better feature representation. Inspired by these multi-task learning methods, we use masked left images as input to reconstruct the original left image as another task. By building pseudo-multi-task learning, we obtain better structural information features, urging stereo matching models to perform better generalization.", + "bbox": [ + 75, + 287, + 472, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper addresses the domain generalization for stereo matching methods by masked representation learning. Our solution considerably increases generalization accuracy while reducing the volatility of generalization performance among different training epochs. First, we randomly mask (or remove) the part of the left image with a fixed ratio following a uniform distribution. Second, we add a simple decoder to the feature extraction module's tail to recover the original left image. Finally, we train the model with two tasks (stereo matching and image reconstruction) as a pseudo-multi-task learning framework and test generalization performance on the multi-datasets. In this task, our strategy helps existing methods to achieve significant gains in generalization performance. Furthermore, we exhibit the generalization performance over different training epochs to demonstrate the superiority of our approach. We hope these observations will help other tasks explore better generalization results. The main contributions are as follows:", + "bbox": [ + 75, + 517, + 472, + 789 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We combine masked image modeling and stereo matching as a pseudo-multi-task learning framework to increase generalization accuracy in stereo matching.", + "- Our approach is employed for various stereo matching networks, significantly improving cross-domain accuracy and reducing the volatility of generalization performance among different training epochs." + ], + "bbox": [ + 94, + 794, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We find that the accuracy of the existing stereo matching domain generalization methods varies significantly among different training epochs. Thus, we advise that stability should be evaluated in cross-domain methods.", + "bbox": [ + 517, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 165, + 640, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning based stereo matching. Since Zbontar et al. used convolution neural networks (CNNs) to replace hand-crafted features and compute matching costs, many researchers started to apply this novel technology to update the traditional pipelines [38]. Mayer et al. built a cost volume in a correlation manner and created a sizeable synthetic dataset to meet the data requirements in stereo matching [19]. Kendall et al. first proposed 4D cost volume by stacking two view features and used 3D convolution to regularize costs, producing a profound impact in stereo matching [10]. Chang et al. introduced the scene awareness module to extract features and applied a 3D hourglass module to aggregate cost volume [3]. Zhang et al. utilized the traditional semi-global matching method to reform the feature matching module (3D convolution), capturing the local or global cost dependencies [39]. Shen et al. replaced the 3D hourglass module to cascade structure as a coarse to fine process, achieving a state-of-the-art generalization performance [30]. Li et al. employed a hierarchical network with recurrent refinement to modify a stacked cascaded architecture and proposed an adaptive group correlation layer to reduce the impact of erroneous rectification [11].", + "bbox": [ + 496, + 190, + 890, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain adaptation stereo matching. To improve generalization performance, many researchers converted the original domain (synthetic scenes) to the target domain (real scenes) [31]. Liu et al. employed an end-to-end training framework with domain translation to tackle the problem of pixel distortion and stereo mismatch after translation [17]. Song et al. designed a bottom-up domain adaptation method based on color transfer and cost regularization to improve generalization without any learnable parameters [31]. Li et al. proposed Fourier-based amplitude transform (FAT), mapping the source image to the target style without altering semantic content [12].", + "bbox": [ + 496, + 523, + 890, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain generalization stereo matching. Generalization performance is an important indicator that reveals the ability of networks to cope with the unseen domain. Tonioni et al. joint unsupervised and continuous online learning to preserve the model's accuracy in multiple environments [33]. Zhang et al. proposed a domain-invariant stereo matching network by regularizing the distribution of learned representations and extracting robust structural representations [40]. Cai et al. introduced matching functions and confidence measures to replace the learning-based feature extraction module, leading to superior generalization to unseen environments [1]. Li et al. introduced a transformer to the stereo matching task, which showed strong", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5436", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/478bf12b845d215f3146b28ad61236ae44ab9284e0c995f2e7afeee02ecfc296.jpg", + "image_caption": [ + "Figure 2. The overall pipeline of our method. Inspired by masked representation learning and multi-task learning, we build a pseudo-multi-task learning framework based on image reconstruction and stereo matching. The feature extraction module can learn better structure information by feature fusion in pseudo-multi-task learning, improving generalization performance and stability." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "robustness [14]. Liu et al. leveraged the feature of a model trained on the large-scale dataset to deal with the domain shift and used a cosine similarity-based cost volume to graft task-oriented features [16]. Zhang et al. argued maintaining feature consistency between matching pixels is a vital factor for the generalization capability and employed a simple pixel-wise contrastive learning across the viewpoints [41].", + "bbox": [ + 75, + 468, + 468, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked representation learning. Masked modeling is a highly successful way for pre-training in natural language processing (NLP) and computer vision (CV) [5, 9]. Vincent et al. proposed denoising autoencoders (DAE), which corrupted an input signal and learned to reconstruct the original, uncorrupted signal [34]. Pathak et al. presented an unsupervised visual feature learning algorithm driven by context-based pixel prediction, proving the effectiveness of CNN pre-training on classification, detection, and segmentation tasks [22]. Devlin et al. held out a portion of the input sequence and trained the transformer to predict the missing content [5]. Chen et al. inspired by progress in unsupervised representation learning in NLP and trained a sequence transformer to predict pixels without incorporating knowledge of structure information [4]. He et al. attempted to demonstrate that masked autoencoders (MAE) are scalable self-supervised learners for computer vision, which achieved the best accuracy among approaches that only used ImageNet-1K data [9]. Feichtenhofer et al. studied a simple extension of Masked Autoencoders (MAE) to spatiotemporal representation learning from videos [6].", + "bbox": [ + 75, + 583, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 467, + 589, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Pipeline", + "text_level": 1, + "bbox": [ + 500, + 494, + 660, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a typical stereo matching method based on deep learning [3, 10, 11, 30, 39], the rectified left and right images are fed to the feature extraction module. Then, these methods construct the cost volume by concatenating the left and traveled right features. Finally, they use 3D convolutions to aggregate the cost volume and regress to the disparity map. Therefore, the poor generalization ability of stereo matching must be caused by two learnable modules (feature extraction module and feature matching module).", + "bbox": [ + 496, + 520, + 890, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In previous studies [16, 40], many achievements have demonstrated that feature representation in the feature extraction module is a vital factor in the model's generalization ability. Meanwhile, we have noticed that multi-task learning helps the models obtain a better matching accuracy, especially when another task has a recognition attribute (e.g., semantic segmentation, object detection, classification, etc.) [18, 27]. Why can recognition algorithms help models to get better performance in multi-task learning? Because feature-sharing in two or more tasks can help the models learn better structural or scene information [16, 18]. Inspired by the above research, we elegantly employ masked representation learning to build a pseudomulti-task learning framework, promoting models to learn better structural information in the feature extraction module to improve generalization ability, as shown in Fig. 2.", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5437", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked Input image. Following MAE [6, 9], we follow a uniform distribution to randomly mask pixels as gray in the left image with the fixed size and ratio (e.g., 0.15, 0.3, 0.45), as shown in Fig. 3. Unlike MAE, we use a low ratio and small mask size to mask pixels and do not employ random positions of patches. Because 1) we hope the models extrapolate the missing pixel from visible neighboring patches to perceive structural information; 2) lacking too many pixels is not conducive to learning correlation in the stereo matching task, mainly feature matching module; 3) stereo matching is sensitive to location information, and the goal is to build the relationship of correlation, so we do not use random positions. This design helps the models to learn structural information in the feature extraction module and does not impact the training process of other learnable modules (e.g., the feature matching module.).", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ab67131582ec1dd77c91f24f2f8047d3bd30db09cf23c6e667ea23d88756f4b3.jpg", + "image_caption": [ + "(a) original left image" + ], + "image_footnote": [], + "bbox": [ + 89, + 347, + 269, + 426 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10520036269da1f7527a9558c046b2e8fe285d4f3d3799200f082c6adfa94732.jpg", + "image_caption": [ + "(b) ratio: 0.15" + ], + "image_footnote": [], + "bbox": [ + 277, + 347, + 455, + 426 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/90d25ae6e42177911a0b4892ed6e2b675d7a5dc8589d0d453133203be863cf0d.jpg", + "image_caption": [ + "(c) ratio: 0.3" + ], + "image_footnote": [], + "bbox": [ + 89, + 441, + 269, + 521 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e60cead34bad873e09bf5cc61e5f887cd99847ed56a67c5c897fe7a2f2ea7efb.jpg", + "image_caption": [ + "(d) ratio: 0.4", + "Figure 3. Illustration of mask sampling strategy. It determines the difficulty of the reconstruction task and affects the reconstruction quality. More masked pixels mean more difficult reconstruction tasks and less reference information in stereo matching." + ], + "image_footnote": [], + "bbox": [ + 277, + 441, + 455, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After masking the left image, we feed the masked left and unmasked right images to the feature extraction module to obtain the left and right features. Then, we design a simple and efficient decoder to predict the missing pixels by the left features, introduced next.", + "bbox": [ + 75, + 627, + 468, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoder Design. We treat the feature extraction module as an encoder to encode the masked left image and obtain the left features. In stereo matching, all methods use down-sampling in the feature extraction process to reduce computation and memory utilization. Therefore, we get the features with a compressed size (e.g., $H / 4 \\times W / 4 \\times F$ ) after the feature extraction module. To reconstruct the target image, we design a lightweight decoder to decode the features to the original size, as shown in Fig. 2.", + "bbox": [ + 75, + 703, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We only employ the decoder module to cope with the image reconstruction task during training. In other words, the decoder module does not work in the testing process. Therefore, the decoder module will not affect the runtime of the", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "existing stereo matching models in the testing process. In this paper, our decoder design is very lightweight and only consists of three 2D convolutions and two up-sampling operations. Thus, the training process will not be significantly prolonged in our design.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction Target. The decoder module reconstructs the left image by predicting the missing pixels. The last layer of the decoder module is a line projection without batch normalization (BN) or ReLU Activation Function (ReLU). Following MAE, we use normalized pixels as the reconstruction target to improve representation quality, and we only compute the loss on the masked pixel of the left image. Our idea is to use image reconstruction and stereo matching to build a pseudo-multi-task learning framework, promoting structural representation learning to improve cross-domain performance further.", + "bbox": [ + 496, + 167, + 892, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Loss Functions", + "text_level": 1, + "bbox": [ + 500, + 342, + 651, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply image reconstruction and stereo matching to construct pseudo-multi-task learning. Hence, the loss functions contain two parts, as follows:", + "bbox": [ + 496, + 366, + 890, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction loss function. We employ the mean squared error (MSE) to compute the loss $\\mathcal{L}_r$ between reconstructed and original left images in the pixel space. The reconstruction loss function $\\mathcal{L}_r$ can be defined as follows:", + "bbox": [ + 496, + 411, + 890, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(I _ {o} (i) - I _ {r} (i)\\right) ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 482, + 890, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ denotes the number of masked pixels, $i$ is the identifier of masked pixels, $I_{o}$ represents the original left image, and $I_{r}$ indicates the reconstructed masked pixels.", + "bbox": [ + 496, + 534, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Matching loss function. The matching loss $\\mathcal{L}_m$ in different papers is different. Most of these methods use the mean absolute error (MAE) or Smooth $L_{1}$ loss as based loss functions. Then, they add photometric or other loss functions to improve performance. In our method, we follow the loss functions of previous matching works and continue to adopt the same manner as the loss following their paper. Therefore, the total loss can be presented as follows:", + "bbox": [ + 496, + 580, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r} + \\mathcal {L} _ {m}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 713, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Algorithm applicability and advantage", + "text_level": 1, + "bbox": [ + 500, + 741, + 831, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm applicability. Although many stereo matching methods have been proposed recently, they do not break away from the existing framework (four parts mentioned in Sec. 1). All models have feature extraction modules, although the style of the modules is variable. Hence, our method can be plugged into the current matching methods by treating the feature extraction modules as encoders and adding a simple decoder. Our method works for all current matching algorithms, not just a few.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5438", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5ba0e301cd3617262d03ba9099226fdf03688dac97bdbde6c3d7931ba0dcc170.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 97, + 269, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d80a5d57abf3390af94b9b5faf3138d12328d1b7a51dd0a6cf644658b9d6de59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 97, + 475, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b88dc55b93ec861587fc6bf2bad67c45e89ac4ccbbe34182d3f59ac5eba67baa.jpg", + "image_caption": [ + "Figure 4. CFNet with different masking ratio. As the masking ratio rises, performance rises first and then declines." + ], + "image_footnote": [], + "bbox": [ + 86, + 246, + 269, + 388 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7153230bbd1d2183fe756480b52b9a96c4d016cb53321b9325a0ce715698d21b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 248, + 475, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm advantage. Compared with current domain adaptation and domain generalization methods in stereo matching, our approach has many advantages regarding runtime and convenience, as shown in follows. (1) Unlike domain adaptation technologies, our method does not need an additional training process or access to the target domain data. Meanwhile, our decoder module is a lightweight structure. Thus, it will not significantly prolong the training process of the existing algorithms. (2) Unlike domain generalization methods, our image reconstruction branch does not participate in testing. Therefore, our approach does not affect the runtime of the current matching methods.", + "bbox": [ + 75, + 458, + 470, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets & Evaluation metrics", + "text_level": 1, + "bbox": [ + 76, + 679, + 346, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The current domain generalized evaluation manner is that the models are trained on the source domain and tested on the target domain. Thus, we first briefly describe these datasets based on different domains and then introduce evaluation metrics on these datasets.", + "bbox": [ + 75, + 703, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Source domain. We train all models on the Sceneflow dataset [19] in the experiment with the same schedule.", + "bbox": [ + 75, + 777, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target domain. Following previous works [12, 16, 30], we evaluate all models on the KITTI 2012&2015 (KT-12 and KT-15) [7, 20], ETH3D (ET) [29], and Middlebury (MB) [28] without fine-tuning.", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metric. Following previous works [16, 40], we apply end-point-error (EPE) and the percentage of error", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pixels larger than $t$ pixels $(>t\\mathrm{px})$ as evaluation metrics. According to evaluated websites, we set $t = 1,2,3$ , which correspond to the default thresholds of $t = 1$ (bad 1.0) on the ETH3D, and $t = 2$ (bad 2.0) on the Middlebury, and $t = 3(D_1)$ on the KITTI 2012&2015, respectively.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 174, + 718, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we implement our strategy with state-of-the-art methods (CFNet [30] and LacGwcNet [15]) by PyTorch. According to our goal, we need to decide on the hyper-parameters and training schedule, as presented below.", + "bbox": [ + 496, + 198, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyper-parameters. We set the maximum disparity $D = 256$ in CFNet and LacGwcNet to ensure that nearly all possible disparity values in the images could be detected. We apply Adam Optimizer $(\\beta_{1} = 0.9, \\beta_{2} = 0.999)$ to optimize the model. In training, we set a mini-batch size of 1 image pair per GPU (six on 6 GPUs) and all loss weights following their papers.", + "bbox": [ + 496, + 258, + 892, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training schedule. We only train models on the Scene-Flow dataset. The learning rate is initially set to $1 \\times 10^{-3}$ for 30 epochs and then reduced to $1 \\times 10^{-4}$ for the other 10 epochs. After this process, we obtain the final models. For data augmentation, we use random cropping, random resize, random shift, and random chromatic transformations to cope with the input image on the SceneFlow dataset.", + "bbox": [ + 496, + 364, + 893, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Main Properties", + "text_level": 1, + "bbox": [ + 500, + 478, + 661, + 494 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a71d3b94046edcf1045441588ee04fc67361cc520e6dc6f47fcef70dee2b7813.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RatioTypeKT-12KT-15ETMB
0EPE1.111.550.692.45
0.15EPE1.051.440.561.86
0.25EPE1.051.450.571.85
0.35EPE1.131.530.652.14
0.45EPE1.271.630.752.57
0t-px error5.836.567.2515.17
0.15t-px error5.016.096.6412.82
0.25t-px error5.126.206.7412.60
0.35t-px error5.636.486.9414.81
0.45t-px error4.006.797.0415.32
", + "bbox": [ + 524, + 511, + 867, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The generalization performance of CFNet with different masking ratios. A low ratio improves generalization performance, while a high ratio hurts generalization performance.", + "bbox": [ + 496, + 691, + 890, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Masking ratio. Fig. 4 and Tab. 1 show the influence of the masking ratio (we use an average of ten epochs as generalization performance.). When the masking ratio is low, it does not affect the performance (source domain) but improves generalization performance. As the masking ratio increases, the performance (source domain) gradually declines, and generalization performance rises first and then falls. Thus, a high masking ratio is not unsuited for this pseudo-multi-task framework due to two factors: (1) a high masking ratio will increase the difficulty of reconstruction,", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5439", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is a disadvantage to the learning process; (2) excessive masking will reduce the learnable matching area in the following modules, such as the feature matching module. Some qualitative results of our final models on the four realistic datasets are displayed in Fig. 5.", + "bbox": [ + 76, + 90, + 472, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53df3f3a1d653cb1095b8a939731203a52148992d2cb234ebd1596ac38456b54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 180, + 205, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d93e7137ed45e36da76987ec7f393612ecb6fcde09fa762ca9b9f2e03c7c59dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 237, + 205, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53ae5ee96e489323508aefbe2b1b65fcd47e8a3f33fff356929df4f7c6651552.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 268, + 205, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9a45c45e1f6ee3fcfef2bc5cf8df1c6e416b15c2d2f3416ed0b8ca34d46d8dbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 300, + 205, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5ac5bb5ae135cbe847d9468f7624fbcba8b5412cd8686a26e918bc4e3915f62.jpg", + "image_caption": [ + "(a) Left image" + ], + "image_footnote": [], + "bbox": [ + 80, + 352, + 205, + 420 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f50043f96b8b94c1cc23db8c000fb50ea595c1ec5a4b9ef43bb4f7bd6eec61c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 180, + 334, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e1791ec47a9c7e3575890519a1ccc0ca7bb2233ee0f72e1551c12c7eaf01b22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 237, + 334, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd77e08ff1e9f1550da5c59e78ca92a4d3bec9c313b2779e15e2420d05019ddc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 268, + 334, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e14747a501d9db7b03f6ee60d4b1c09f2da7d5b071554a18d19de03eb1e5bf53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 300, + 334, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2a750cb2a2c649f08088eb9cec96c7a3e314231cf6fc8f5629956c0125de8a4.jpg", + "image_caption": [ + "(b) Mask-CFNet" + ], + "image_footnote": [], + "bbox": [ + 210, + 352, + 334, + 420 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7bbc7a5c3c66a40fbb8c3fa813b856795348ad0e92808fa087b0778b5b8dec4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 180, + 465, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec1400aacc0d6b6fb1ba97fc8e215ca318d0e97914229d23d2fa74a0ca293dd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 237, + 465, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd60ff35742e81a118a8372c98269fe03794a7a949812e01a1fcda8aae3b9bf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 268, + 465, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1eefab6a1ef7bd2bf012cae80d55d681812c8f3d2abdee87e2b096d4e68ba6fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 300, + 465, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f71af24dd6b301d1aaa63d31bcdcb7a6004541ba2382d92ac80cf54188da809c.jpg", + "image_caption": [ + "(C) Mask-LacGwcNet", + "Figure 6. The convergence process with or without masked representation learning. It proves that a low ratio does not change the convergence process and the matching accuracy, while a high ratio affects the learning process and reduces the matching accuracy." + ], + "image_footnote": [], + "bbox": [ + 339, + 352, + 465, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Convergence process. In Fig. 6, we compare the performance and the convergence process with or without masked representation learning. Because CFNet and LacGwcNet have many outputs, we use the matching accuracy of the final output as the performance on the training dataset. Meanwhile, we test the performance of these methods on the source domain. We conclude that: (1) the results are very stable on the source domain; (2) compared with baselines, our method does not change the convergence process for the low mask ratio; (3) a high mask ratio will affect the learning process and reduce the matching accuracy. Stereo matching is a task highly related to location information. Mask ratio means more missing pixels, and it will affect the models to learn the correlation based on adjacent pixels in the feature matching module.", + "bbox": [ + 75, + 522, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Runtime. As shown in Tab. 2, we list the runtime of two models at the training and testing processes. Our method only adds a lightweight decoder to reconstruct the left image in the training process, which does not participate in the test process. Therefore, we draw the following conclusions: (1) for the training process, our method does not significantly prolong training time compared with baselines; (2) for the testing process, our approach is no different from baselines.", + "bbox": [ + 75, + 750, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training epochs and generalization performance. As presented in Fig. 7, we draw two models' curves with or", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c222d2aa4af29bf66381cf56664c2eabf11890f1ba914664e4abe834c59e46ab.jpg", + "image_caption": [ + "(a) CFNet (training)" + ], + "image_footnote": [], + "bbox": [ + 513, + 97, + 687, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ff5c92d15978089e3b43d4348a8aa04fab1a8c16a34f4aea7ca928bb9c5fe3a.jpg", + "image_caption": [ + "(b) LacGwcNet (training)" + ], + "image_footnote": [], + "bbox": [ + 720, + 97, + 895, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ef88127f9058272805745f091242fdd84f3dfa7539e512b685df5a86e256cf7.jpg", + "image_caption": [ + "Figure 5. The examples of different models on the source and target datasets (without fine-tuning). From top to bottom are SceneFlow (source dataset), KITTI 2012&2015, ETH3D, and Middlebury (four target datasets).", + "(c) CFNet (testing)" + ], + "image_footnote": [], + "bbox": [ + 511, + 263, + 687, + 402 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/351a49878550615554a409b39437119b9e27610b5547f1a960ff8e4252d9622d.jpg", + "image_caption": [ + "(d) LacGwcNet (testing)" + ], + "image_footnote": [], + "bbox": [ + 720, + 263, + 895, + 402 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/78ba66fb7024cb54730b0b36b6f35269de133eb91b469e114a3a9d2d87a32e5c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMaskTrainingResolutionRuntime (s)
CFNet✓✓576 × 3200.89
CFNet✓X960 × 5760.052
CFNetX✓576 × 3200.84
CFNetXX960 × 5760.051
LacGwcNet✓✓576 × 3201.63
LacGwcNet✓X960 × 5760.264
LacGwcNetX✓576 × 3201.61
LacGwcNetXX960 × 5760.264
", + "bbox": [ + 501, + 507, + 890, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The runtime with different resolutions. It presents that our method does not significantly prolong training time and is no different from the baseline at the test runtime.", + "bbox": [ + 498, + 660, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "without masked representation learning. When the models converge ( $>30$ epochs, as shown in Fig. 6), the results are stable on the source dataset. However, the generalization performance varies significantly between adjacent training epochs. Compared with the baselines, the models with masked representation learning can perform better and more stable. As shown in Tab. 3, we apply mean and variance to measure these fluctuations among different training epochs. By quantifying the changes, we find that the structural information can help the models achieve a relatively stable and better generalization performance.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5440", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ce9ce9b17a138d063ff27b08d503b2bf1c5129ba1719742286e4411fc4dab84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 97, + 269, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96965d7e7eb28ba7fef44de5a0ddf4b306b40732c1447001439663e3098f6d4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 98, + 475, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acc6621ae0fc7c3b5dfad5e439c28fea0fc57c6b272afa1f8b9cc29d2c512d7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 98, + 686, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc848f0efa3f1eca1ae70656ad5f6a493ca4ede1866dbe63bf35ab251764934b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 90, + 893, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8fa2b09cd0791ce87142edcbed3bf83511b7cc3597b56caafc9508b0bd8900e5.jpg", + "image_caption": [ + "Figure 7. The generalization performance with or without masked representation among different epochs. It shows that the generalization performance varies significantly between adjacent training epochs, and our method achieves better performance and is more stable." + ], + "image_footnote": [], + "bbox": [ + 86, + 247, + 267, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71b3177b1fb2f24bcf84663db47df3dfacb00d4760808ce2638ca22b7ae273e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 247, + 475, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ac593ad6854eedede0074dd33457bb355807943d7e276f10639a9b3b13e7416.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 247, + 684, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49c79bc0b4c769097f525df37e18cefa0554012eb573d0224ab47f627287b1ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 241, + 893, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e1c24b83b73940e0403844674c66b545e5519af3efe0c1f4df2ec6d801d62ece.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
M.MaskDataEPE (Mean)EPE (Var.)D1 (Mean)D1 (Var.)
CFNet✓KT-121.440.045.030.03
XKT-121.550.055.820.13
✓KT-151.050.016.080.07
XKT-151.110.016.560.19
✓ET0.560.026.630.21
XET0.690.037.240.30
✓MB1.860.1312.820.37
XMB2.450.1315.160.90
LacGwcNet✓KT-121.430.036.570.30
XKT-121.830.329.1710.46
✓KT-151.410.026.080.23
XKT-151.780.118.379.24
✓ET1.000.376.571.03
XET2.180.847.991.37
✓MB2.400.0217.300.89
XMB2.490.1918.280.51
", + "bbox": [ + 81, + 455, + 464, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparisons", + "text_level": 1, + "bbox": [ + 76, + 801, + 215, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with generalization method. We compare our methods with other stereo matching methods, including well-known and generalized methods. Tab. 4 shows that our approach can help models improve generalization ability, indirectly proving that masked representation learn", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fcfc44f5ea15168d5ac49f31d0b50b492367fd24084c3a843b9e088d30dc8161.jpg", + "table_caption": [ + "Table 3. Volatility comparison of with or without masked representation. It demonstrates that our method can help the models achieve a relatively stable and better generalization performance." + ], + "table_footnote": [], + "table_body": "
MethodKT-12 > 3pxKT-15 > 3pxMB > 2pxET > 1px
PSMNet [3]15.116.326.923.8
GWCNet [8]12.012.234.211.0
GANet [39]10.111.720.314.1
DSMNet [40]6.26.521.86.2
FC-DSM [41]5.56.212.06.0
CFNet [30]4.75.815.35.8
GF-PSMNet [16]5.34.610.96.2
Mask-CFNet4.85.813.75.7
Mask-LacGwcNet5.75.616.95.3
", + "bbox": [ + 504, + 455, + 887, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Cross-domain generalization evaluation (peak results) on four target datasets. All methods are only trained on the Scene-Flow dataset and tested on training images of four real datasets.", + "bbox": [ + 498, + 665, + 890, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing can promote high-level representation feature learning. Note that all methods have volatility among training epochs, and we list the volatility as shown in the supplemental material. It prompted us to suggest that the scope of fluctuation should be attached to future work in stereo matching.", + "bbox": [ + 496, + 734, + 890, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with fine-tuning. We fine-tune the models (CFNet and LacGwcNet) on the KITTI datasets to test the performance after fine-tuning. Tab. 5 shows that the results are similar to the trend of the SceneFlow dataset. When the masking ratio is low (< 25%), our method does not affect the performance and produces nearly the same re", + "bbox": [ + 496, + 810, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5441", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ea6fd0b18f612cce2f460fc5bcb46ee0e36d602c12d6801284deb84acb61304.jpg", + "table_caption": [ + "sults. As the ratio rises, the performance gradually declines." + ], + "table_footnote": [], + "table_body": "
MethodRatioKT-12 (Out-Noc)KT-15 (D1-all)
LacGwcNet [15]01.131.77
LacGwcNet [15]0.151.151.78
LacGwcNet [15]0.251.161.77
LacGwcNet [15]0.351.271.95
LacGwcNet [15]0.451.392.21
CFNet [30]01.231.88
CFNet [30]0.151.231.89
CFNet [30]0.251.271.91
CFNet [30]0.351.362.05
CFNet [30]0.451.482.28
", + "bbox": [ + 99, + 119, + 447, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The fine-tuning results on the KITTI dataset. It indicates the results are similar to the trend of the SceneFlow dataset.", + "bbox": [ + 76, + 328, + 467, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 388, + 189, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper has proposed a simple approach to alleviate the significant volatility of generalization performance. This section will discuss two topics: (1) When does our approach fail? (2) What is the real unseen domain?", + "bbox": [ + 76, + 416, + 467, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When does our approach fail? (1) In our study, we have added the channels of the feature extraction module to get a better reconstruction image. However, the generalization performance did not obtain remarkable improvement. Although larger models can handle two tasks well, it is not conducive to the fusion of multi-task learning because larger modules may learn features of two tasks separately. Similarly, it will weaken the effect of masked representation. (2) Excessive masking will cause the learnable matching area to shrink in the feature matching module, affecting matching accuracy or generalization. Thus, large-size masking is not recommended to improve generalization.", + "bbox": [ + 75, + 477, + 467, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What is the real unseen domain? In previous studies, nearly all papers used KITTI, ETH 3D, and Middlebury datasets as the unseen domains. However, can these datasets be represented the unseen domain? Although the color distribution, illumination, and authenticity are different between the source domain and the target domains, similar objects are included in the source domain, such as the street scenes in the Driving dataset (sub-dataset of SceneFlow), pedestrian in the Driving dataset, daily supplies in FlyingThing3D (sub-dataset of SceneFlow), etc. Herefore, we can get good results without fine-tuning models on the target dataset. Does it mean we can use these algorithms in practice? The answer is no. We use these models that work well on target datasets to deal with daily images collected by the zed camera or remote sensing images collected by satellite. As shown in Fig. 8, we can not get the expected results", + "bbox": [ + 75, + 659, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29f260035131d743d16fce9ec6dc1cf4902f88ca4fd2a435bbdb82a2f3a3912a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 89, + 686, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/519ce6ec7e811d3f847ce80516a5fe7cfe8c0313c2fb142999d6b329f706cc22.jpg", + "image_caption": [ + "(a) remote sensing image", + "(c) daily image", + "Figure 8. The examples of failures in the real unseen domain. When elements do not appear in the source dataset, the models can not predict the expected results." + ], + "image_footnote": [], + "bbox": [ + 506, + 243, + 684, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba832e6837f33195faf029f0326bb3ddd795577129db0ee1d88b1773d274e78f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 89, + 883, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a525c3383e040ae9a5a4aa3ed0027ff51c270627c3162023e1138e78bef3357c.jpg", + "image_caption": [ + "(b) disparity map", + "(d) disparity map" + ], + "image_footnote": [], + "bbox": [ + 705, + 243, + 883, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for these images. We consider that the main reason is that these elements or scenes do not appear in the source dataset, such as the human body, buildings under the top view angle, etc. When the real unseen domain appears in the images, matching methods based on deep learning can not predict disparity correctly. Thus, we doubt the cross-domain performance of these datasets (KITTI, ETH 3D, and Middlebury) can represent actual cross-domain performance.", + "bbox": [ + 498, + 426, + 890, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 568, + 617, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a simple masked representation method to address the problem of unstable generalization performance among different training epochs for stereo matching. Our approach is inspired by masked representation learning and multi-task learning to construct a pseudo-multi-task learning framework, which helps models learn better features in the feature extraction module to improve generalization performance and stability. Extensive experiments have proved the effectiveness of our method. Meanwhile, it also demonstrated that the current evaluation manner is unsuitable for measuring generalization performance. In the last, we discussed the failure of our approach and doubted the current definition of the unseen domain.", + "bbox": [ + 498, + 595, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported in part by Jiangxi Natural Science Foundations (20224BAB212009, 20212BBE53017), National Natural Science Foundation of China (61866027, 62271410, and 62001396), the Foundation of Key Laboratory of Jiangxi Province for Image Processing and Pattern Recognition (ET202204304), and the Doctor's Foundation of Nanchang Hangkong University (EA202204228, EA202204257). We thank the anonymous area chairs and reviewers for their helpful feedback.", + "bbox": [ + 498, + 796, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5442", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Changjiang Cai, Matteo Poggi, Stefano Mattoccia, and Filippos Mordohai. Matching-space stereo networks for cross-domain generalization. In International Conference on 3D Vision (3DV), pages 364-373, 2020. 2", + "[2] Rohan Chabra, Julian Straub, Christopher Sweeney, Richard Newcombe, and Henry Fuchs. Stereodnet: Dilated residual stereonet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786-11795, 2019. 1", + "[3] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5410-5418, 2018. 1, 2, 3, 7", + "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proceedings of International Conference on Machine Learning (ICML), pages 1691-1703, 2020. 3", + "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 3", + "[6] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In arXiv preprint, 2022. 3, 4", + "[7] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. 5", + "[8] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3273-3282, 2019. 7", + "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 3, 4", + "[10] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 66-75, 2017. 1, 2, 3", + "[11] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16263-16272, 2022. 1, 2, 3", + "[12] Xing Li, Yangyu Fan, Zhibo Rao, Zhe Guo, and Guoyun Lv. Improving stereo matching generalization via fourier-based" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "amplitude transform. IEEE Signal Processing Letters (SPL), 2022. 2, 5", + "[13] Xing Li, Yangyu Fan, Zhibo Rao, Guoyun Lv, and Shiya Liu. Synthetic-to-real domain adaptation joint spatial feature transform for stereo matching. IEEE Signal Processing Letters (SPL), 29:60-64, 2021. 2", + "[14] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6197-6206, 2021. 3", + "[15] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 36, pages 1647-1655, 2022. 1, 5, 8", + "[16] Biyang Liu, Huimin Yu, and Guodong Qi. Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13012-13021, 2022. 2, 3, 5, 7", + "[17] Rui Liu, Chengxi Yang, Wenxiu Sun, Xiaogang Wang, and Hongsheng Li. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12757-12766, 2020. 2", + "[18] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1871-1880, 2019. 2, 3", + "[19] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 2, 5", + "[20] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3061-3070, 2015. 5", + "[21] Guang-Yu Nie, Ming-Ming Cheng, Yun Liu, Zhengfa Liang, Deng-Ping Fan, Yue Liu, and Yongtian Wang. Multi-level context ultra-aggregation for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3283-3291, 2019. 1", + "[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536-2544, 2016. 3", + "[23] Zhibo Rao, Yuchao Dai, Zhelun Shen, and Renjie He. Rethinking training strategy in stereo matching. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), pages 1-14, 2022. 1", + "[24] Zhibo Rao, Mingyi He, Yuchao Dai, and Zhelun Shen. Patch attention network with generative adversarial model" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5443", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for semi-supervised binocular disparity prediction. The Visual Computer, pages 1-17, 2020. 1", + "[25] Zhibo Rao, Mingyi He, Yuchao Dai, and Zhelun Shen. Sliding space-disparity transformer for stereo matching. Neural Computing and Applications (NCAA), pages 1-14, 2022. 1", + "[26] Zhibo Rao, Mingyi He, Yuchao Dai, Zhidong Zhu, Bo Li, and Renjie He. NLca-net: a non-local context attention network for stereo matching. APSIPA Transactions on Signal and Information Processing, 9, 2020. 1", + "[27] Zhibo Rao, Mingyi He, Zhidong Zhu, Yuchao Dai, and Renjie He. Bidirectional guided attention network for 3-d semantic detection of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(7):6138-6153, 2020. 2, 3", + "[28] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nešić, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In German conference on pattern recognition, pages 31–42, 2014. 5", + "[29] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3260-3269, 2017. 5", + "[30] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13906-13915, 2021. 1, 2, 3, 5, 7, 8", + "[31] Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. Adastereo: a simple and efficient approach for adaptive stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10328-10337, 2021. 2", + "[32] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9661-9670, 2019. 1", + "[33] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 195-204, 2019. 1, 2", + "[34] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of International Conference on Machine Learning (ICML), pages 1096-1103, 2008. 3", + "[35] Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12497-12506, 2021. 1", + "[36] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12981-12990, 2022. 1", + "[37] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1959-1968, 2020. 1", + "[38] Jure Zbontar and Yann LeCun. Computing the stereo matching cost with a convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1592-1599, 2015. 2", + "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 185-194, 2019. 1, 2, 3, 7", + "[40] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In European Conference on Computer Vision (ECCV), pages 420-439, 2020. 1, 2, 3, 5, 7", + "[41] Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13001-13011, 2022. 2, 3, 7", + "[42] Youmin Zhang, Yimin Chen, Xiao Bai, Suihanjin Yu, Kun Yu, Zhiwei Li, and Kuiyuan Yang. Adaptive unimodal cost volume filtering for deep stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 12926-12934, 2020. 1" + ], + "bbox": [ + 501, + 92, + 893, + 542 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5444", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_model.json b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f86c3642cf05dc729bc485eae101fff4e1c853b --- /dev/null +++ b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.131, + 0.85, + 0.154 + ], + "angle": 0, + "content": "Masked representation learning for domain generalized stereo matching" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.179, + 0.871, + 0.252 + ], + "angle": 0, + "content": "Zhibo Rao\\(^{1,2\\boxtimes}\\), Bangshu Xiong\\(^{1}\\), Mingyi He\\(^{2}\\), Yuchao Dai\\(^{2}\\), Renjie He\\(^{2}\\), Zhelun Shen\\(^{3}\\), Xing Li\\(^{2\\boxtimes}\\) \n\\(^{1}\\)Nanchang Hangkong University, Nanchang, China \n\\(^{2}\\)Northwestern Polytechnical University, Xi'an, China \n\\(^{3}\\)Baidu Research, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.255, + 0.826, + 0.287 + ], + "angle": 0, + "content": "raoxi36@foxmail.com, xiongs@126.com, {myhe, daiyuchao, davidhrj} @nwpu.edu.cn, 1901213310@pku.edu.cn, lixing36@foxmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.749 + ], + "angle": 0, + "content": "Recently, many deep stereo matching methods have begun to focus on cross-domain performance, achieving impressive achievements. However, these methods did not deal with the significant volatility of generalization performance among different training epochs. Inspired by masked representation learning and multi-task learning, this paper designs a simple and effective masked representation for domain generalized stereo matching. First, we feed the masked left and complete right images as input into the models. Then, we add a lightweight and simple decoder following the feature extraction module to recover the original left image. Finally, we train the models with two tasks (stereo matching and image reconstruction) as a pseudomulti-task learning framework, promoting models to learn structure information and to improve generalization performance. We implement our method on two well-known architectures (CFNet and LacGwcNet) to demonstrate its effectiveness. Experimental results on multi-datasets show that: (1) our method can be easily plugged into the current various stereo matching models to improve generalization performance; (2) our method can reduce the significant volatility of generalization performance among different training epochs; (3) we find that the current methods prefer to choose the best results among different training epochs as generalization performance, but it is impossible to select the best performance by ground truth in practice." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.776, + 0.21, + 0.792 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.802, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Stereo matching is a challenging research topic of computer vision, which aims to obtain the corresponding pixels between two rectified stereo images [3, 30, 37, 39]. It is essential in many applications, including autonomous [11], augmented reality [35], virtual reality [2], etc." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.328, + 0.691, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.473, + 0.64, + 0.486 + ], + "angle": 0, + "content": "(a) CFNet" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.328, + 0.897, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.473, + 0.864, + 0.486 + ], + "angle": 0, + "content": "(b) LacGwcNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.5, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Figure 1. The generalization performance among different epochs on multi-datasets. When the models converge on the source domain (Sceneflow), the results are stable on the source domain. However, generalization performance has fluctuations on target datasets (KITTI 2012&2015, ETH3D, and Middlebury)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.895, + 0.84 + ], + "angle": 0, + "content": "In mainstream deep stereo matching methods, the main steps include four parts: feature extraction, cost volume, feature matching, and disparity regression [10]. To improve the accuracy or speed, researchers proposed many strategies to improve the above four parts [3, 25, 36]. For example, 1) scene awareness modules are applied to the feature extraction to improve accuracy [3, 21]; 2) similarity measurements are employed to enhance physical representation in cost volume construction [23, 26, 36]; 3) various feature matching modules are proposed to solve inefficient calculations or improve accuracy [15, 25, 39]; 4) loss functions of classification or other fields are joined to disparity regression to obtain unimodal results [42]. Although the above methods have achieved significant progress in accuracy or efficiency, they failed to obtain good generalization performance in unseen domains [30, 40]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Therefore, many approaches learn domain-invariant representation features to achieve better generalization capability [40]. There are three solutions to accomplish this objective: the unsupervised matching method [24,32,33], domain" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.266, + 0.9 + ], + "angle": 0, + "content": "\\(1 \\boxtimes\\) is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5435" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "adaptation techniques [12, 13, 17, 31], and domain generalization approaches [16, 30, 41]. The above solutions reveal that feature presentation is crucial for improving generalization capability [40, 41]. However, all domain generalized matching methods did not mention that the generalization performance varies significantly among different training epochs, as shown in Fig. 1. Namely, the results were unstable. These methods preferred to choose the best results to represent the generalization performance by testing the models of different training epochs. However, we can not employ ground truth (can not obtain) to select the best model among different training epochs in practice. Thus, it is crucial to keep a stable generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.517 + ], + "angle": 0, + "content": "On the other hand, many researchers introduced multi-task learning into stereo matching [18, 27]. For instance, Rao et al. proposed a bidirectional guided attention network to cope with semantic segmentation and stereo matching simultaneously [27]. Liu et al. employed the task-shared and task-specific manner to obtain a generalizable representation and used the loss weights to balance all tasks [18]. The multi-task learning proved that feature sharing and weight sharing are essential in different tasks, promoting the learning process and obtaining a better feature representation. Inspired by these multi-task learning methods, we use masked left images as input to reconstruct the original left image as another task. By building pseudo-multi-task learning, we obtain better structural information features, urging stereo matching models to perform better generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.473, + 0.79 + ], + "angle": 0, + "content": "This paper addresses the domain generalization for stereo matching methods by masked representation learning. Our solution considerably increases generalization accuracy while reducing the volatility of generalization performance among different training epochs. First, we randomly mask (or remove) the part of the left image with a fixed ratio following a uniform distribution. Second, we add a simple decoder to the feature extraction module's tail to recover the original left image. Finally, we train the model with two tasks (stereo matching and image reconstruction) as a pseudo-multi-task learning framework and test generalization performance on the multi-datasets. In this task, our strategy helps existing methods to achieve significant gains in generalization performance. Furthermore, we exhibit the generalization performance over different training epochs to demonstrate the superiority of our approach. We hope these observations will help other tasks explore better generalization results. The main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.795, + 0.468, + 0.839 + ], + "angle": 0, + "content": "- We combine masked image modeling and stereo matching as a pseudo-multi-task learning framework to increase generalization accuracy in stereo matching." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.841, + 0.468, + 0.899 + ], + "angle": 0, + "content": "- Our approach is employed for various stereo matching networks, significantly improving cross-domain accuracy and reducing the volatility of generalization performance among different training epochs." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.795, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "- We find that the accuracy of the existing stereo matching domain generalization methods varies significantly among different training epochs. Thus, we advise that stability should be evaluated in cross-domain methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.166, + 0.642, + 0.182 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Deep learning based stereo matching. Since Zbontar et al. used convolution neural networks (CNNs) to replace hand-crafted features and compute matching costs, many researchers started to apply this novel technology to update the traditional pipelines [38]. Mayer et al. built a cost volume in a correlation manner and created a sizeable synthetic dataset to meet the data requirements in stereo matching [19]. Kendall et al. first proposed 4D cost volume by stacking two view features and used 3D convolution to regularize costs, producing a profound impact in stereo matching [10]. Chang et al. introduced the scene awareness module to extract features and applied a 3D hourglass module to aggregate cost volume [3]. Zhang et al. utilized the traditional semi-global matching method to reform the feature matching module (3D convolution), capturing the local or global cost dependencies [39]. Shen et al. replaced the 3D hourglass module to cascade structure as a coarse to fine process, achieving a state-of-the-art generalization performance [30]. Li et al. employed a hierarchical network with recurrent refinement to modify a stacked cascaded architecture and proposed an adaptive group correlation layer to reduce the impact of erroneous rectification [11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Domain adaptation stereo matching. To improve generalization performance, many researchers converted the original domain (synthetic scenes) to the target domain (real scenes) [31]. Liu et al. employed an end-to-end training framework with domain translation to tackle the problem of pixel distortion and stereo mismatch after translation [17]. Song et al. designed a bottom-up domain adaptation method based on color transfer and cost regularization to improve generalization without any learnable parameters [31]. Li et al. proposed Fourier-based amplitude transform (FAT), mapping the source image to the target style without altering semantic content [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Domain generalization stereo matching. Generalization performance is an important indicator that reveals the ability of networks to cope with the unseen domain. Tonioni et al. joint unsupervised and continuous online learning to preserve the model's accuracy in multiple environments [33]. Zhang et al. proposed a domain-invariant stereo matching network by regularizing the distribution of learned representations and extracting robust structural representations [40]. Cai et al. introduced matching functions and confidence measures to replace the learning-based feature extraction module, leading to superior generalization to unseen environments [1]. Li et al. introduced a transformer to the stereo matching task, which showed strong" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5436" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.895, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.893, + 0.445 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline of our method. Inspired by masked representation learning and multi-task learning, we build a pseudo-multi-task learning framework based on image reconstruction and stereo matching. The feature extraction module can learn better structure information by feature fusion in pseudo-multi-task learning, improving generalization performance and stability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.576 + ], + "angle": 0, + "content": "robustness [14]. Liu et al. leveraged the feature of a model trained on the large-scale dataset to deal with the domain shift and used a cosine similarity-based cost volume to graft task-oriented features [16]. Zhang et al. argued maintaining feature consistency between matching pixels is a vital factor for the generalization capability and employed a simple pixel-wise contrastive learning across the viewpoints [41]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Masked representation learning. Masked modeling is a highly successful way for pre-training in natural language processing (NLP) and computer vision (CV) [5, 9]. Vincent et al. proposed denoising autoencoders (DAE), which corrupted an input signal and learned to reconstruct the original, uncorrupted signal [34]. Pathak et al. presented an unsupervised visual feature learning algorithm driven by context-based pixel prediction, proving the effectiveness of CNN pre-training on classification, detection, and segmentation tasks [22]. Devlin et al. held out a portion of the input sequence and trained the transformer to predict the missing content [5]. Chen et al. inspired by progress in unsupervised representation learning in NLP and trained a sequence transformer to predict pixels without incorporating knowledge of structure information [4]. He et al. attempted to demonstrate that masked autoencoders (MAE) are scalable self-supervised learners for computer vision, which achieved the best accuracy among approaches that only used ImageNet-1K data [9]. Feichtenhofer et al. studied a simple extension of Masked Autoencoders (MAE) to spatiotemporal representation learning from videos [6]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.591, + 0.483 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.661, + 0.512 + ], + "angle": 0, + "content": "3.1. Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.657 + ], + "angle": 0, + "content": "In a typical stereo matching method based on deep learning [3, 10, 11, 30, 39], the rectified left and right images are fed to the feature extraction module. Then, these methods construct the cost volume by concatenating the left and traveled right features. Finally, they use 3D convolutions to aggregate the cost volume and regress to the disparity map. Therefore, the poor generalization ability of stereo matching must be caused by two learnable modules (feature extraction module and feature matching module)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In previous studies [16, 40], many achievements have demonstrated that feature representation in the feature extraction module is a vital factor in the model's generalization ability. Meanwhile, we have noticed that multi-task learning helps the models obtain a better matching accuracy, especially when another task has a recognition attribute (e.g., semantic segmentation, object detection, classification, etc.) [18, 27]. Why can recognition algorithms help models to get better performance in multi-task learning? Because feature-sharing in two or more tasks can help the models learn better structural or scene information [16, 18]. Inspired by the above research, we elegantly employ masked representation learning to build a pseudomulti-task learning framework, promoting models to learn better structural information in the feature extraction module to improve generalization ability, as shown in Fig. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5437" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "Masked Input image. Following MAE [6, 9], we follow a uniform distribution to randomly mask pixels as gray in the left image with the fixed size and ratio (e.g., 0.15, 0.3, 0.45), as shown in Fig. 3. Unlike MAE, we use a low ratio and small mask size to mask pixels and do not employ random positions of patches. Because 1) we hope the models extrapolate the missing pixel from visible neighboring patches to perceive structural information; 2) lacking too many pixels is not conducive to learning correlation in the stereo matching task, mainly feature matching module; 3) stereo matching is sensitive to location information, and the goal is to build the relationship of correlation, so we do not use random positions. This design helps the models to learn structural information in the feature extraction module and does not impact the training process of other learnable modules (e.g., the feature matching module.)." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.348, + 0.27, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.43, + 0.234, + 0.441 + ], + "angle": 0, + "content": "(a) original left image" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.348, + 0.456, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.43, + 0.402, + 0.44 + ], + "angle": 0, + "content": "(b) ratio: 0.15" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.442, + 0.27, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.525, + 0.211, + 0.535 + ], + "angle": 0, + "content": "(c) ratio: 0.3" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.442, + 0.456, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.525, + 0.399, + 0.535 + ], + "angle": 0, + "content": "(d) ratio: 0.4" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Figure 3. Illustration of mask sampling strategy. It determines the difficulty of the reconstruction task and affects the reconstruction quality. More masked pixels mean more difficult reconstruction tasks and less reference information in stereo matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.703 + ], + "angle": 0, + "content": "After masking the left image, we feed the masked left and unmasked right images to the feature extraction module to obtain the left and right features. Then, we design a simple and efficient decoder to predict the missing pixels by the left features, introduced next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Decoder Design. We treat the feature extraction module as an encoder to encode the masked left image and obtain the left features. In stereo matching, all methods use down-sampling in the feature extraction process to reduce computation and memory utilization. Therefore, we get the features with a compressed size (e.g., \\( H / 4 \\times W / 4 \\times F \\)) after the feature extraction module. To reconstruct the target image, we design a lightweight decoder to decode the features to the original size, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We only employ the decoder module to cope with the image reconstruction task during training. In other words, the decoder module does not work in the testing process. Therefore, the decoder module will not affect the runtime of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "existing stereo matching models in the testing process. In this paper, our decoder design is very lightweight and only consists of three 2D convolutions and two up-sampling operations. Thus, the training process will not be significantly prolonged in our design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.333 + ], + "angle": 0, + "content": "Reconstruction Target. The decoder module reconstructs the left image by predicting the missing pixels. The last layer of the decoder module is a line projection without batch normalization (BN) or ReLU Activation Function (ReLU). Following MAE, we use normalized pixels as the reconstruction target to improve representation quality, and we only compute the loss on the masked pixel of the left image. Our idea is to use image reconstruction and stereo matching to build a pseudo-multi-task learning framework, promoting structural representation learning to improve cross-domain performance further." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.343, + 0.652, + 0.358 + ], + "angle": 0, + "content": "3.2. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.412 + ], + "angle": 0, + "content": "We apply image reconstruction and stereo matching to construct pseudo-multi-task learning. Hence, the loss functions contain two parts, as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Reconstruction loss function. We employ the mean squared error (MSE) to compute the loss \\(\\mathcal{L}_r\\) between reconstructed and original left images in the pixel space. The reconstruction loss function \\(\\mathcal{L}_r\\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.483, + 0.892, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(I _ {o} (i) - I _ {r} (i)\\right) ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.581 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of masked pixels, \\(i\\) is the identifier of masked pixels, \\(I_{o}\\) represents the original left image, and \\(I_{r}\\) indicates the reconstructed masked pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Matching loss function. The matching loss \\(\\mathcal{L}_m\\) in different papers is different. Most of these methods use the mean absolute error (MAE) or Smooth \\(L_{1}\\) loss as based loss functions. Then, they add photometric or other loss functions to improve performance. In our method, we follow the loss functions of previous matching works and continue to adopt the same manner as the loss following their paper. Therefore, the total loss can be presented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.714, + 0.892, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r} + \\mathcal {L} _ {m}. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.833, + 0.759 + ], + "angle": 0, + "content": "3.3. Algorithm applicability and advantage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Algorithm applicability. Although many stereo matching methods have been proposed recently, they do not break away from the existing framework (four parts mentioned in Sec. 1). All models have feature extraction modules, although the style of the modules is variable. Hence, our method can be plugged into the current matching methods by treating the feature extraction modules as encoders and adding a simple decoder. Our method works for all current matching algorithms, not just a few." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5438" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.098, + 0.271, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.098, + 0.476, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.247, + 0.27, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.249, + 0.476, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Figure 4. CFNet with different masking ratio. As the masking ratio rises, performance rises first and then declines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.471, + 0.641 + ], + "angle": 0, + "content": "Algorithm advantage. Compared with current domain adaptation and domain generalization methods in stereo matching, our approach has many advantages regarding runtime and convenience, as shown in follows. (1) Unlike domain adaptation technologies, our method does not need an additional training process or access to the target domain data. Meanwhile, our decoder module is a lightweight structure. Thus, it will not significantly prolong the training process of the existing algorithms. (2) Unlike domain generalization methods, our image reconstruction branch does not participate in testing. Therefore, our approach does not affect the runtime of the current matching methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.21, + 0.672 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.348, + 0.696 + ], + "angle": 0, + "content": "4.1. Datasets & Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.779 + ], + "angle": 0, + "content": "The current domain generalized evaluation manner is that the models are trained on the source domain and tested on the target domain. Thus, we first briefly describe these datasets based on different domains and then introduce evaluation metrics on these datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Source domain. We train all models on the Sceneflow dataset [19] in the experiment with the same schedule." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Target domain. Following previous works [12, 16, 30], we evaluate all models on the KITTI 2012&2015 (KT-12 and KT-15) [7, 20], ETH3D (ET) [29], and Middlebury (MB) [28] without fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Evaluation metric. Following previous works [16, 40], we apply end-point-error (EPE) and the percentage of error" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "pixels larger than \\(t\\) pixels \\((>t\\mathrm{px})\\) as evaluation metrics. According to evaluated websites, we set \\(t = 1,2,3\\), which correspond to the default thresholds of \\(t = 1\\) (bad 1.0) on the ETH3D, and \\(t = 2\\) (bad 2.0) on the Middlebury, and \\(t = 3(D_1)\\) on the KITTI 2012&2015, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.175, + 0.719, + 0.192 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.259 + ], + "angle": 0, + "content": "In this paper, we implement our strategy with state-of-the-art methods (CFNet [30] and LacGwcNet [15]) by PyTorch. According to our goal, we need to decide on the hyper-parameters and training schedule, as presented below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.893, + 0.365 + ], + "angle": 0, + "content": "Hyper-parameters. We set the maximum disparity \\( D = 256 \\) in CFNet and LacGwcNet to ensure that nearly all possible disparity values in the images could be detected. We apply Adam Optimizer \\( (\\beta_{1} = 0.9, \\beta_{2} = 0.999) \\) to optimize the model. In training, we set a mini-batch size of 1 image pair per GPU (six on 6 GPUs) and all loss weights following their papers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.894, + 0.47 + ], + "angle": 0, + "content": "Training schedule. We only train models on the Scene-Flow dataset. The learning rate is initially set to \\(1 \\times 10^{-3}\\) for 30 epochs and then reduced to \\(1 \\times 10^{-4}\\) for the other 10 epochs. After this process, we obtain the final models. For data augmentation, we use random cropping, random resize, random shift, and random chromatic transformations to cope with the input image on the SceneFlow dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.479, + 0.663, + 0.495 + ], + "angle": 0, + "content": "4.3. Main Properties" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.512, + 0.869, + 0.683 + ], + "angle": 0, + "content": "
RatioTypeKT-12KT-15ETMB
0EPE1.111.550.692.45
0.15EPE1.051.440.561.86
0.25EPE1.051.450.571.85
0.35EPE1.131.530.652.14
0.45EPE1.271.630.752.57
0t-px error5.836.567.2515.17
0.15t-px error5.016.096.6412.82
0.25t-px error5.126.206.7412.60
0.35t-px error5.636.486.9414.81
0.45t-px error4.006.797.0415.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Table 1. The generalization performance of CFNet with different masking ratios. A low ratio improves generalization performance, while a high ratio hurts generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Masking ratio. Fig. 4 and Tab. 1 show the influence of the masking ratio (we use an average of ten epochs as generalization performance.). When the masking ratio is low, it does not affect the performance (source domain) but improves generalization performance. As the masking ratio increases, the performance (source domain) gradually declines, and generalization performance rises first and then falls. Thus, a high masking ratio is not unsuited for this pseudo-multi-task framework due to two factors: (1) a high masking ratio will increase the difficulty of reconstruction," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5439" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "which is a disadvantage to the learning process; (2) excessive masking will reduce the learnable matching area in the following modules, such as the feature matching module. Some qualitative results of our final models on the four realistic datasets are displayed in Fig. 5." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.181, + 0.207, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.238, + 0.206, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.269, + 0.206, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.301, + 0.206, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.353, + 0.206, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.423, + 0.177, + 0.434 + ], + "angle": 0, + "content": "(a) Left image" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.181, + 0.336, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.238, + 0.336, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.269, + 0.336, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.301, + 0.336, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.353, + 0.336, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.423, + 0.312, + 0.433 + ], + "angle": 0, + "content": "(b) Mask-CFNet" + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.181, + 0.466, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.238, + 0.466, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.269, + 0.466, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.301, + 0.466, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.353, + 0.466, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.423, + 0.455, + 0.433 + ], + "angle": 0, + "content": "(C) Mask-LacGwcNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.45, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Figure 5. The examples of different models on the source and target datasets (without fine-tuning). From top to bottom are SceneFlow (source dataset), KITTI 2012&2015, ETH3D, and Middlebury (four target datasets)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Convergence process. In Fig. 6, we compare the performance and the convergence process with or without masked representation learning. Because CFNet and LacGwcNet have many outputs, we use the matching accuracy of the final output as the performance on the training dataset. Meanwhile, we test the performance of these methods on the source domain. We conclude that: (1) the results are very stable on the source domain; (2) compared with baselines, our method does not change the convergence process for the low mask ratio; (3) a high mask ratio will affect the learning process and reduce the matching accuracy. Stereo matching is a task highly related to location information. Mask ratio means more missing pixels, and it will affect the models to learn the correlation based on adjacent pixels in the feature matching module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Runtime. As shown in Tab. 2, we list the runtime of two models at the training and testing processes. Our method only adds a lightweight decoder to reconstruct the left image in the training process, which does not participate in the test process. Therefore, we draw the following conclusions: (1) for the training process, our method does not significantly prolong training time compared with baselines; (2) for the testing process, our approach is no different from baselines." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Training epochs and generalization performance. As presented in Fig. 7, we draw two models' curves with or" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.098, + 0.689, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.241, + 0.675, + 0.256 + ], + "angle": 0, + "content": "(a) CFNet (training)" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.098, + 0.897, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.241, + 0.901, + 0.256 + ], + "angle": 0, + "content": "(b) LacGwcNet (training)" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.265, + 0.688, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.407, + 0.671, + 0.422 + ], + "angle": 0, + "content": "(c) CFNet (testing)" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.265, + 0.897, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.407, + 0.897, + 0.422 + ], + "angle": 0, + "content": "(d) LacGwcNet (testing)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.434, + 0.893, + 0.49 + ], + "angle": 0, + "content": "Figure 6. The convergence process with or without masked representation learning. It proves that a low ratio does not change the convergence process and the matching accuracy, while a high ratio affects the learning process and reduces the matching accuracy." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.508, + 0.892, + 0.651 + ], + "angle": 0, + "content": "
ModelMaskTrainingResolutionRuntime (s)
CFNet✓✓576 × 3200.89
CFNet✓X960 × 5760.052
CFNetX✓576 × 3200.84
CFNetXX960 × 5760.051
LacGwcNet✓✓576 × 3201.63
LacGwcNet✓X960 × 5760.264
LacGwcNetX✓576 × 3201.61
LacGwcNetXX960 × 5760.264
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Table 2. The runtime with different resolutions. It presents that our method does not significantly prolong training time and is no different from the baseline at the test runtime." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "without masked representation learning. When the models converge (\\(>30\\) epochs, as shown in Fig. 6), the results are stable on the source dataset. However, the generalization performance varies significantly between adjacent training epochs. Compared with the baselines, the models with masked representation learning can perform better and more stable. As shown in Tab. 3, we apply mean and variance to measure these fluctuations among different training epochs. By quantifying the changes, we find that the structural information can help the models achieve a relatively stable and better generalization performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5440" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.098, + 0.27, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.099, + 0.477, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.099, + 0.687, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.091, + 0.895, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.248, + 0.268, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.248, + 0.476, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.248, + 0.686, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.242, + 0.895, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.405, + 0.893, + 0.434 + ], + "angle": 0, + "content": "Figure 7. The generalization performance with or without masked representation among different epochs. It shows that the generalization performance varies significantly between adjacent training epochs, and our method achieves better performance and is more stable." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.457, + 0.465, + 0.723 + ], + "angle": 0, + "content": "
M.MaskDataEPE (Mean)EPE (Var.)D1 (Mean)D1 (Var.)
CFNet✓KT-121.440.045.030.03
XKT-121.550.055.820.13
✓KT-151.050.016.080.07
XKT-151.110.016.560.19
✓ET0.560.026.630.21
XET0.690.037.240.30
✓MB1.860.1312.820.37
XMB2.450.1315.160.90
LacGwcNet✓KT-121.430.036.570.30
XKT-121.830.329.1710.46
✓KT-151.410.026.080.23
XKT-151.780.118.379.24
✓ET1.000.376.571.03
XET2.180.847.991.37
✓MB2.400.0217.300.89
XMB2.490.1918.280.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Table 3. Volatility comparison of with or without masked representation. It demonstrates that our method can help the models achieve a relatively stable and better generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.217, + 0.818 + ], + "angle": 0, + "content": "4.4. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Comparisons with generalization method. We compare our methods with other stereo matching methods, including well-known and generalized methods. Tab. 4 shows that our approach can help models improve generalization ability, indirectly proving that masked representation learn" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.457, + 0.888, + 0.655 + ], + "angle": 0, + "content": "
MethodKT-12 > 3pxKT-15 > 3pxMB > 2pxET > 1px
PSMNet [3]15.116.326.923.8
GWCNet [8]12.012.234.211.0
GANet [39]10.111.720.314.1
DSMNet [40]6.26.521.86.2
FC-DSM [41]5.56.212.06.0
CFNet [30]4.75.815.35.8
GF-PSMNet [16]5.34.610.96.2
Mask-CFNet4.85.813.75.7
Mask-LacGwcNet5.75.616.95.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.666, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Table 4. Cross-domain generalization evaluation (peak results) on four target datasets. All methods are only trained on the Scene-Flow dataset and tested on training images of four real datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.81 + ], + "angle": 0, + "content": "ing can promote high-level representation feature learning. Note that all methods have volatility among training epochs, and we list the volatility as shown in the supplemental material. It prompted us to suggest that the scope of fluctuation should be attached to future work in stereo matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Comparisons with fine-tuning. We fine-tune the models (CFNet and LacGwcNet) on the KITTI datasets to test the performance after fine-tuning. Tab. 5 shows that the results are similar to the trend of the SceneFlow dataset. When the masking ratio is low (< 25%), our method does not affect the performance and produces nearly the same re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5441" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.092, + 0.468, + 0.107 + ], + "angle": 0, + "content": "sults. As the ratio rises, the performance gradually declines." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.12, + 0.449, + 0.32 + ], + "angle": 0, + "content": "
MethodRatioKT-12 (Out-Noc)KT-15 (D1-all)
LacGwcNet [15]01.131.77
LacGwcNet [15]0.151.151.78
LacGwcNet [15]0.251.161.77
LacGwcNet [15]0.351.271.95
LacGwcNet [15]0.451.392.21
CFNet [30]01.231.88
CFNet [30]0.151.231.89
CFNet [30]0.251.271.91
CFNet [30]0.351.362.05
CFNet [30]0.451.482.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.329, + 0.468, + 0.357 + ], + "angle": 0, + "content": "Table 5. The fine-tuning results on the KITTI dataset. It indicates the results are similar to the trend of the SceneFlow dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.39, + 0.191, + 0.405 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.468, + 0.476 + ], + "angle": 0, + "content": "This paper has proposed a simple approach to alleviate the significant volatility of generalization performance. This section will discuss two topics: (1) When does our approach fail? (2) What is the real unseen domain?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.468, + 0.658 + ], + "angle": 0, + "content": "When does our approach fail? (1) In our study, we have added the channels of the feature extraction module to get a better reconstruction image. However, the generalization performance did not obtain remarkable improvement. Although larger models can handle two tasks well, it is not conducive to the fusion of multi-task learning because larger modules may learn features of two tasks separately. Similarly, it will weaken the effect of masked representation. (2) Excessive masking will cause the learnable matching area to shrink in the feature matching module, affecting matching accuracy or generalization. Thus, large-size masking is not recommended to improve generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.468, + 0.901 + ], + "angle": 0, + "content": "What is the real unseen domain? In previous studies, nearly all papers used KITTI, ETH 3D, and Middlebury datasets as the unseen domains. However, can these datasets be represented the unseen domain? Although the color distribution, illumination, and authenticity are different between the source domain and the target domains, similar objects are included in the source domain, such as the street scenes in the Driving dataset (sub-dataset of SceneFlow), pedestrian in the Driving dataset, daily supplies in FlyingThing3D (sub-dataset of SceneFlow), etc. Herefore, we can get good results without fine-tuning models on the target dataset. Does it mean we can use these algorithms in practice? The answer is no. We use these models that work well on target datasets to deal with daily images collected by the zed camera or remote sensing images collected by satellite. As shown in Fig. 8, we can not get the expected results" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.687, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.231, + 0.664, + 0.243 + ], + "angle": 0, + "content": "(a) remote sensing image" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.244, + 0.686, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.324, + 0.64, + 0.337 + ], + "angle": 0, + "content": "(c) daily image" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.09, + 0.885, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.231, + 0.846, + 0.243 + ], + "angle": 0, + "content": "(b) disparity map" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.244, + 0.885, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.324, + 0.844, + 0.337 + ], + "angle": 0, + "content": "(d) disparity map" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.354, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Figure 8. The examples of failures in the real unseen domain. When elements do not appear in the source dataset, the models can not predict the expected results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.428, + 0.892, + 0.549 + ], + "angle": 0, + "content": "for these images. We consider that the main reason is that these elements or scenes do not appear in the source dataset, such as the human body, buildings under the top view angle, etc. When the real unseen domain appears in the images, matching methods based on deep learning can not predict disparity correctly. Thus, we doubt the cross-domain performance of these datasets (KITTI, ETH 3D, and Middlebury) can represent actual cross-domain performance." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.569, + 0.618, + 0.585 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.597, + 0.892, + 0.793 + ], + "angle": 0, + "content": "In this paper, we have proposed a simple masked representation method to address the problem of unstable generalization performance among different training epochs for stereo matching. Our approach is inspired by masked representation learning and multi-task learning to construct a pseudo-multi-task learning framework, which helps models learn better features in the feature extraction module to improve generalization performance and stability. Extensive experiments have proved the effectiveness of our method. Meanwhile, it also demonstrated that the current evaluation manner is unsuitable for measuring generalization performance. In the last, we discussed the failure of our approach and doubted the current definition of the unseen domain." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.797, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported in part by Jiangxi Natural Science Foundations (20224BAB212009, 20212BBE53017), National Natural Science Foundation of China (61866027, 62271410, and 62001396), the Foundation of Key Laboratory of Jiangxi Province for Image Processing and Pattern Recognition (ET202204304), and the Doctor's Foundation of Nanchang Hangkong University (EA202204228, EA202204257). We thank the anonymous area chairs and reviewers for their helpful feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5442" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Changjiang Cai, Matteo Poggi, Stefano Mattoccia, and Filippos Mordohai. Matching-space stereo networks for cross-domain generalization. In International Conference on 3D Vision (3DV), pages 364-373, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Rohan Chabra, Julian Straub, Christopher Sweeney, Richard Newcombe, and Henry Fuchs. Stereodnet: Dilated residual stereonet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786-11795, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[3] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5410-5418, 2018. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proceedings of International Conference on Machine Learning (ICML), pages 1691-1703, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.443 + ], + "angle": 0, + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.487 + ], + "angle": 0, + "content": "[6] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In arXiv preprint, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[7] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[8] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3273-3282, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.47, + 0.7 + ], + "angle": 0, + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.702, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[10] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 66-75, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.788, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[11] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16263-16272, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Xing Li, Yangyu Fan, Zhibo Rao, Zhe Guo, and Guoyun Lv. Improving stereo matching generalization via fourier-based" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "amplitude transform. IEEE Signal Processing Letters (SPL), 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Xing Li, Yangyu Fan, Zhibo Rao, Guoyun Lv, and Shiya Liu. Synthetic-to-real domain adaptation joint spatial feature transform for stereo matching. IEEE Signal Processing Letters (SPL), 29:60-64, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[14] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6197-6206, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[15] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 36, pages 1647-1655, 2022. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[16] Biyang Liu, Huimin Yu, and Guodong Qi. Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13012-13021, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[17] Rui Liu, Chengxi Yang, Wenxiu Sun, Xiaogang Wang, and Hongsheng Li. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12757-12766, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[18] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1871-1880, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[19] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[20] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3061-3070, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[21] Guang-Yu Nie, Ming-Ming Cheng, Yun Liu, Zhengfa Liang, Deng-Ping Fan, Yue Liu, and Yongtian Wang. Multi-level context ultra-aggregation for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3283-3291, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536-2544, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[23] Zhibo Rao, Yuchao Dai, Zhelun Shen, and Renjie He. Rethinking training strategy in stereo matching. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), pages 1-14, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Zhibo Rao, Mingyi He, Yuchao Dai, and Zhelun Shen. Patch attention network with generative adversarial model" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5443" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "for semi-supervised binocular disparity prediction. The Visual Computer, pages 1-17, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[25] Zhibo Rao, Mingyi He, Yuchao Dai, and Zhelun Shen. Sliding space-disparity transformer for stereo matching. Neural Computing and Applications (NCAA), pages 1-14, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.471, + 0.22 + ], + "angle": 0, + "content": "[26] Zhibo Rao, Mingyi He, Yuchao Dai, Zhidong Zhu, Bo Li, and Renjie He. NLca-net: a non-local context attention network for stereo matching. APSIPA Transactions on Signal and Information Processing, 9, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[27] Zhibo Rao, Mingyi He, Zhidong Zhu, Yuchao Dai, and Renjie He. Bidirectional guided attention network for 3-d semantic detection of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(7):6138-6153, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.471, + 0.362 + ], + "angle": 0, + "content": "[28] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nešić, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In German conference on pattern recognition, pages 31–42, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[29] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3260-3269, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[30] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13906-13915, 2021. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[31] Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. Adastereo: a simple and efficient approach for adaptive stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10328-10337, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[32] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9661-9670, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[33] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 195-204, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[34] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of International Conference on Machine Learning (ICML), pages 1096-1103, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[35] Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12497-12506, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[36] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12981-12990, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[37] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1959-1968, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[38] Jure Zbontar and Yann LeCun. Computing the stereo matching cost with a convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1592-1599, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 185-194, 2019. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[40] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In European Conference on Computer Vision (ECCV), pages 420-439, 2020. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[41] Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13001-13011, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.894, + 0.543 + ], + "angle": 0, + "content": "[42] Youmin Zhang, Yimin Chen, Xiao Bai, Suihanjin Yu, Kun Yu, Zhiwei Li, and Kuiyuan Yang. Adaptive unimodal cost volume filtering for deep stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 12926-12934, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5444" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_origin.pdf b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c441648697f27300c0663aad42086ecabe549964 --- /dev/null +++ b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/33890d9e-3bb5-4cca-822b-e87656eea803_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e74024ed42acf9ca713e667592fec10516dbf0b4695684e6648403f485c76532 +size 1098147 diff --git a/2023/Masked Representation Learning for Domain Generalized Stereo Matching/full.md b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..660c2462c1f3be6cd109c12a72ba1879cb79e6e1 --- /dev/null +++ b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/full.md @@ -0,0 +1,328 @@ +# Masked representation learning for domain generalized stereo matching + +Zhibo Rao $^{1,2\boxtimes}$ , Bangshu Xiong $^{1}$ , Mingyi He $^{2}$ , Yuchao Dai $^{2}$ , Renjie He $^{2}$ , Zhelun Shen $^{3}$ , Xing Li $^{2\boxtimes}$ $^{1}$ Nanchang Hangkong University, Nanchang, China + $^{2}$ Northwestern Polytechnical University, Xi'an, China + $^{3}$ Baidu Research, Beijing, China + +raoxi36@foxmail.com, xiongs@126.com, {myhe, daiyuchao, davidhrj} @nwpu.edu.cn, 1901213310@pku.edu.cn, lixing36@foxmail.com + +# Abstract + +Recently, many deep stereo matching methods have begun to focus on cross-domain performance, achieving impressive achievements. However, these methods did not deal with the significant volatility of generalization performance among different training epochs. Inspired by masked representation learning and multi-task learning, this paper designs a simple and effective masked representation for domain generalized stereo matching. First, we feed the masked left and complete right images as input into the models. Then, we add a lightweight and simple decoder following the feature extraction module to recover the original left image. Finally, we train the models with two tasks (stereo matching and image reconstruction) as a pseudomulti-task learning framework, promoting models to learn structure information and to improve generalization performance. We implement our method on two well-known architectures (CFNet and LacGwcNet) to demonstrate its effectiveness. Experimental results on multi-datasets show that: (1) our method can be easily plugged into the current various stereo matching models to improve generalization performance; (2) our method can reduce the significant volatility of generalization performance among different training epochs; (3) we find that the current methods prefer to choose the best results among different training epochs as generalization performance, but it is impossible to select the best performance by ground truth in practice. + +# 1. Introduction + +Stereo matching is a challenging research topic of computer vision, which aims to obtain the corresponding pixels between two rectified stereo images [3, 30, 37, 39]. It is essential in many applications, including autonomous [11], augmented reality [35], virtual reality [2], etc. + +![](images/affa0af305a6708276a5e279a71d68eb70622006b0651033db6c13149d67b3d0.jpg) +(a) CFNet +Figure 1. The generalization performance among different epochs on multi-datasets. When the models converge on the source domain (Sceneflow), the results are stable on the source domain. However, generalization performance has fluctuations on target datasets (KITTI 2012&2015, ETH3D, and Middlebury). + +![](images/5329afccfc24509ed3eabff4361888eb8a703cf52f1b6dc6cf066c053b736101.jpg) +(b) LacGwcNet + +In mainstream deep stereo matching methods, the main steps include four parts: feature extraction, cost volume, feature matching, and disparity regression [10]. To improve the accuracy or speed, researchers proposed many strategies to improve the above four parts [3, 25, 36]. For example, 1) scene awareness modules are applied to the feature extraction to improve accuracy [3, 21]; 2) similarity measurements are employed to enhance physical representation in cost volume construction [23, 26, 36]; 3) various feature matching modules are proposed to solve inefficient calculations or improve accuracy [15, 25, 39]; 4) loss functions of classification or other fields are joined to disparity regression to obtain unimodal results [42]. Although the above methods have achieved significant progress in accuracy or efficiency, they failed to obtain good generalization performance in unseen domains [30, 40]. + +Therefore, many approaches learn domain-invariant representation features to achieve better generalization capability [40]. There are three solutions to accomplish this objective: the unsupervised matching method [24,32,33], domain + +adaptation techniques [12, 13, 17, 31], and domain generalization approaches [16, 30, 41]. The above solutions reveal that feature presentation is crucial for improving generalization capability [40, 41]. However, all domain generalized matching methods did not mention that the generalization performance varies significantly among different training epochs, as shown in Fig. 1. Namely, the results were unstable. These methods preferred to choose the best results to represent the generalization performance by testing the models of different training epochs. However, we can not employ ground truth (can not obtain) to select the best model among different training epochs in practice. Thus, it is crucial to keep a stable generalization performance. + +On the other hand, many researchers introduced multi-task learning into stereo matching [18, 27]. For instance, Rao et al. proposed a bidirectional guided attention network to cope with semantic segmentation and stereo matching simultaneously [27]. Liu et al. employed the task-shared and task-specific manner to obtain a generalizable representation and used the loss weights to balance all tasks [18]. The multi-task learning proved that feature sharing and weight sharing are essential in different tasks, promoting the learning process and obtaining a better feature representation. Inspired by these multi-task learning methods, we use masked left images as input to reconstruct the original left image as another task. By building pseudo-multi-task learning, we obtain better structural information features, urging stereo matching models to perform better generalization. + +This paper addresses the domain generalization for stereo matching methods by masked representation learning. Our solution considerably increases generalization accuracy while reducing the volatility of generalization performance among different training epochs. First, we randomly mask (or remove) the part of the left image with a fixed ratio following a uniform distribution. Second, we add a simple decoder to the feature extraction module's tail to recover the original left image. Finally, we train the model with two tasks (stereo matching and image reconstruction) as a pseudo-multi-task learning framework and test generalization performance on the multi-datasets. In this task, our strategy helps existing methods to achieve significant gains in generalization performance. Furthermore, we exhibit the generalization performance over different training epochs to demonstrate the superiority of our approach. We hope these observations will help other tasks explore better generalization results. The main contributions are as follows: + +- We combine masked image modeling and stereo matching as a pseudo-multi-task learning framework to increase generalization accuracy in stereo matching. +- Our approach is employed for various stereo matching networks, significantly improving cross-domain accuracy and reducing the volatility of generalization performance among different training epochs. + +- We find that the accuracy of the existing stereo matching domain generalization methods varies significantly among different training epochs. Thus, we advise that stability should be evaluated in cross-domain methods. + +# 2. Related Work + +Deep learning based stereo matching. Since Zbontar et al. used convolution neural networks (CNNs) to replace hand-crafted features and compute matching costs, many researchers started to apply this novel technology to update the traditional pipelines [38]. Mayer et al. built a cost volume in a correlation manner and created a sizeable synthetic dataset to meet the data requirements in stereo matching [19]. Kendall et al. first proposed 4D cost volume by stacking two view features and used 3D convolution to regularize costs, producing a profound impact in stereo matching [10]. Chang et al. introduced the scene awareness module to extract features and applied a 3D hourglass module to aggregate cost volume [3]. Zhang et al. utilized the traditional semi-global matching method to reform the feature matching module (3D convolution), capturing the local or global cost dependencies [39]. Shen et al. replaced the 3D hourglass module to cascade structure as a coarse to fine process, achieving a state-of-the-art generalization performance [30]. Li et al. employed a hierarchical network with recurrent refinement to modify a stacked cascaded architecture and proposed an adaptive group correlation layer to reduce the impact of erroneous rectification [11]. + +Domain adaptation stereo matching. To improve generalization performance, many researchers converted the original domain (synthetic scenes) to the target domain (real scenes) [31]. Liu et al. employed an end-to-end training framework with domain translation to tackle the problem of pixel distortion and stereo mismatch after translation [17]. Song et al. designed a bottom-up domain adaptation method based on color transfer and cost regularization to improve generalization without any learnable parameters [31]. Li et al. proposed Fourier-based amplitude transform (FAT), mapping the source image to the target style without altering semantic content [12]. + +Domain generalization stereo matching. Generalization performance is an important indicator that reveals the ability of networks to cope with the unseen domain. Tonioni et al. joint unsupervised and continuous online learning to preserve the model's accuracy in multiple environments [33]. Zhang et al. proposed a domain-invariant stereo matching network by regularizing the distribution of learned representations and extracting robust structural representations [40]. Cai et al. introduced matching functions and confidence measures to replace the learning-based feature extraction module, leading to superior generalization to unseen environments [1]. Li et al. introduced a transformer to the stereo matching task, which showed strong + +![](images/478bf12b845d215f3146b28ad61236ae44ab9284e0c995f2e7afeee02ecfc296.jpg) +Figure 2. The overall pipeline of our method. Inspired by masked representation learning and multi-task learning, we build a pseudo-multi-task learning framework based on image reconstruction and stereo matching. The feature extraction module can learn better structure information by feature fusion in pseudo-multi-task learning, improving generalization performance and stability. + +robustness [14]. Liu et al. leveraged the feature of a model trained on the large-scale dataset to deal with the domain shift and used a cosine similarity-based cost volume to graft task-oriented features [16]. Zhang et al. argued maintaining feature consistency between matching pixels is a vital factor for the generalization capability and employed a simple pixel-wise contrastive learning across the viewpoints [41]. + +Masked representation learning. Masked modeling is a highly successful way for pre-training in natural language processing (NLP) and computer vision (CV) [5, 9]. Vincent et al. proposed denoising autoencoders (DAE), which corrupted an input signal and learned to reconstruct the original, uncorrupted signal [34]. Pathak et al. presented an unsupervised visual feature learning algorithm driven by context-based pixel prediction, proving the effectiveness of CNN pre-training on classification, detection, and segmentation tasks [22]. Devlin et al. held out a portion of the input sequence and trained the transformer to predict the missing content [5]. Chen et al. inspired by progress in unsupervised representation learning in NLP and trained a sequence transformer to predict pixels without incorporating knowledge of structure information [4]. He et al. attempted to demonstrate that masked autoencoders (MAE) are scalable self-supervised learners for computer vision, which achieved the best accuracy among approaches that only used ImageNet-1K data [9]. Feichtenhofer et al. studied a simple extension of Masked Autoencoders (MAE) to spatiotemporal representation learning from videos [6]. + +# 3. Method + +# 3.1. Overall Pipeline + +In a typical stereo matching method based on deep learning [3, 10, 11, 30, 39], the rectified left and right images are fed to the feature extraction module. Then, these methods construct the cost volume by concatenating the left and traveled right features. Finally, they use 3D convolutions to aggregate the cost volume and regress to the disparity map. Therefore, the poor generalization ability of stereo matching must be caused by two learnable modules (feature extraction module and feature matching module). + +In previous studies [16, 40], many achievements have demonstrated that feature representation in the feature extraction module is a vital factor in the model's generalization ability. Meanwhile, we have noticed that multi-task learning helps the models obtain a better matching accuracy, especially when another task has a recognition attribute (e.g., semantic segmentation, object detection, classification, etc.) [18, 27]. Why can recognition algorithms help models to get better performance in multi-task learning? Because feature-sharing in two or more tasks can help the models learn better structural or scene information [16, 18]. Inspired by the above research, we elegantly employ masked representation learning to build a pseudomulti-task learning framework, promoting models to learn better structural information in the feature extraction module to improve generalization ability, as shown in Fig. 2. + +Masked Input image. Following MAE [6, 9], we follow a uniform distribution to randomly mask pixels as gray in the left image with the fixed size and ratio (e.g., 0.15, 0.3, 0.45), as shown in Fig. 3. Unlike MAE, we use a low ratio and small mask size to mask pixels and do not employ random positions of patches. Because 1) we hope the models extrapolate the missing pixel from visible neighboring patches to perceive structural information; 2) lacking too many pixels is not conducive to learning correlation in the stereo matching task, mainly feature matching module; 3) stereo matching is sensitive to location information, and the goal is to build the relationship of correlation, so we do not use random positions. This design helps the models to learn structural information in the feature extraction module and does not impact the training process of other learnable modules (e.g., the feature matching module.). + +![](images/ab67131582ec1dd77c91f24f2f8047d3bd30db09cf23c6e667ea23d88756f4b3.jpg) +(a) original left image + +![](images/10520036269da1f7527a9558c046b2e8fe285d4f3d3799200f082c6adfa94732.jpg) +(b) ratio: 0.15 + +![](images/90d25ae6e42177911a0b4892ed6e2b675d7a5dc8589d0d453133203be863cf0d.jpg) +(c) ratio: 0.3 + +![](images/e60cead34bad873e09bf5cc61e5f887cd99847ed56a67c5c897fe7a2f2ea7efb.jpg) +(d) ratio: 0.4 +Figure 3. Illustration of mask sampling strategy. It determines the difficulty of the reconstruction task and affects the reconstruction quality. More masked pixels mean more difficult reconstruction tasks and less reference information in stereo matching. + +After masking the left image, we feed the masked left and unmasked right images to the feature extraction module to obtain the left and right features. Then, we design a simple and efficient decoder to predict the missing pixels by the left features, introduced next. + +Decoder Design. We treat the feature extraction module as an encoder to encode the masked left image and obtain the left features. In stereo matching, all methods use down-sampling in the feature extraction process to reduce computation and memory utilization. Therefore, we get the features with a compressed size (e.g., $H / 4 \times W / 4 \times F$ ) after the feature extraction module. To reconstruct the target image, we design a lightweight decoder to decode the features to the original size, as shown in Fig. 2. + +We only employ the decoder module to cope with the image reconstruction task during training. In other words, the decoder module does not work in the testing process. Therefore, the decoder module will not affect the runtime of the + +existing stereo matching models in the testing process. In this paper, our decoder design is very lightweight and only consists of three 2D convolutions and two up-sampling operations. Thus, the training process will not be significantly prolonged in our design. + +Reconstruction Target. The decoder module reconstructs the left image by predicting the missing pixels. The last layer of the decoder module is a line projection without batch normalization (BN) or ReLU Activation Function (ReLU). Following MAE, we use normalized pixels as the reconstruction target to improve representation quality, and we only compute the loss on the masked pixel of the left image. Our idea is to use image reconstruction and stereo matching to build a pseudo-multi-task learning framework, promoting structural representation learning to improve cross-domain performance further. + +# 3.2. Loss Functions + +We apply image reconstruction and stereo matching to construct pseudo-multi-task learning. Hence, the loss functions contain two parts, as follows: + +Reconstruction loss function. We employ the mean squared error (MSE) to compute the loss $\mathcal{L}_r$ between reconstructed and original left images in the pixel space. The reconstruction loss function $\mathcal{L}_r$ can be defined as follows: + +$$ +\mathcal {L} _ {r} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(I _ {o} (i) - I _ {r} (i)\right) ^ {2}, \tag {1} +$$ + +where $N$ denotes the number of masked pixels, $i$ is the identifier of masked pixels, $I_{o}$ represents the original left image, and $I_{r}$ indicates the reconstructed masked pixels. + +Matching loss function. The matching loss $\mathcal{L}_m$ in different papers is different. Most of these methods use the mean absolute error (MAE) or Smooth $L_{1}$ loss as based loss functions. Then, they add photometric or other loss functions to improve performance. In our method, we follow the loss functions of previous matching works and continue to adopt the same manner as the loss following their paper. Therefore, the total loss can be presented as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {r} + \mathcal {L} _ {m}. \tag {2} +$$ + +# 3.3. Algorithm applicability and advantage + +Algorithm applicability. Although many stereo matching methods have been proposed recently, they do not break away from the existing framework (four parts mentioned in Sec. 1). All models have feature extraction modules, although the style of the modules is variable. Hence, our method can be plugged into the current matching methods by treating the feature extraction modules as encoders and adding a simple decoder. Our method works for all current matching algorithms, not just a few. + +![](images/5ba0e301cd3617262d03ba9099226fdf03688dac97bdbde6c3d7931ba0dcc170.jpg) + +![](images/d80a5d57abf3390af94b9b5faf3138d12328d1b7a51dd0a6cf644658b9d6de59.jpg) + +![](images/b88dc55b93ec861587fc6bf2bad67c45e89ac4ccbbe34182d3f59ac5eba67baa.jpg) +Figure 4. CFNet with different masking ratio. As the masking ratio rises, performance rises first and then declines. + +![](images/7153230bbd1d2183fe756480b52b9a96c4d016cb53321b9325a0ce715698d21b.jpg) + +Algorithm advantage. Compared with current domain adaptation and domain generalization methods in stereo matching, our approach has many advantages regarding runtime and convenience, as shown in follows. (1) Unlike domain adaptation technologies, our method does not need an additional training process or access to the target domain data. Meanwhile, our decoder module is a lightweight structure. Thus, it will not significantly prolong the training process of the existing algorithms. (2) Unlike domain generalization methods, our image reconstruction branch does not participate in testing. Therefore, our approach does not affect the runtime of the current matching methods. + +# 4. Experiments + +# 4.1. Datasets & Evaluation metrics + +The current domain generalized evaluation manner is that the models are trained on the source domain and tested on the target domain. Thus, we first briefly describe these datasets based on different domains and then introduce evaluation metrics on these datasets. + +Source domain. We train all models on the Sceneflow dataset [19] in the experiment with the same schedule. + +Target domain. Following previous works [12, 16, 30], we evaluate all models on the KITTI 2012&2015 (KT-12 and KT-15) [7, 20], ETH3D (ET) [29], and Middlebury (MB) [28] without fine-tuning. + +Evaluation metric. Following previous works [16, 40], we apply end-point-error (EPE) and the percentage of error + +pixels larger than $t$ pixels $(>t\mathrm{px})$ as evaluation metrics. According to evaluated websites, we set $t = 1,2,3$ , which correspond to the default thresholds of $t = 1$ (bad 1.0) on the ETH3D, and $t = 2$ (bad 2.0) on the Middlebury, and $t = 3(D_1)$ on the KITTI 2012&2015, respectively. + +# 4.2. Implementation Details + +In this paper, we implement our strategy with state-of-the-art methods (CFNet [30] and LacGwcNet [15]) by PyTorch. According to our goal, we need to decide on the hyper-parameters and training schedule, as presented below. + +Hyper-parameters. We set the maximum disparity $D = 256$ in CFNet and LacGwcNet to ensure that nearly all possible disparity values in the images could be detected. We apply Adam Optimizer $(\beta_{1} = 0.9, \beta_{2} = 0.999)$ to optimize the model. In training, we set a mini-batch size of 1 image pair per GPU (six on 6 GPUs) and all loss weights following their papers. + +Training schedule. We only train models on the Scene-Flow dataset. The learning rate is initially set to $1 \times 10^{-3}$ for 30 epochs and then reduced to $1 \times 10^{-4}$ for the other 10 epochs. After this process, we obtain the final models. For data augmentation, we use random cropping, random resize, random shift, and random chromatic transformations to cope with the input image on the SceneFlow dataset. + +# 4.3. Main Properties + +
RatioTypeKT-12KT-15ETMB
0EPE1.111.550.692.45
0.15EPE1.051.440.561.86
0.25EPE1.051.450.571.85
0.35EPE1.131.530.652.14
0.45EPE1.271.630.752.57
0t-px error5.836.567.2515.17
0.15t-px error5.016.096.6412.82
0.25t-px error5.126.206.7412.60
0.35t-px error5.636.486.9414.81
0.45t-px error4.006.797.0415.32
+ +Table 1. The generalization performance of CFNet with different masking ratios. A low ratio improves generalization performance, while a high ratio hurts generalization performance. + +Masking ratio. Fig. 4 and Tab. 1 show the influence of the masking ratio (we use an average of ten epochs as generalization performance.). When the masking ratio is low, it does not affect the performance (source domain) but improves generalization performance. As the masking ratio increases, the performance (source domain) gradually declines, and generalization performance rises first and then falls. Thus, a high masking ratio is not unsuited for this pseudo-multi-task framework due to two factors: (1) a high masking ratio will increase the difficulty of reconstruction, + +which is a disadvantage to the learning process; (2) excessive masking will reduce the learnable matching area in the following modules, such as the feature matching module. Some qualitative results of our final models on the four realistic datasets are displayed in Fig. 5. + +![](images/53df3f3a1d653cb1095b8a939731203a52148992d2cb234ebd1596ac38456b54.jpg) + +![](images/d93e7137ed45e36da76987ec7f393612ecb6fcde09fa762ca9b9f2e03c7c59dc.jpg) + +![](images/53ae5ee96e489323508aefbe2b1b65fcd47e8a3f33fff356929df4f7c6651552.jpg) + +![](images/9a45c45e1f6ee3fcfef2bc5cf8df1c6e416b15c2d2f3416ed0b8ca34d46d8dbb.jpg) + +![](images/a5ac5bb5ae135cbe847d9468f7624fbcba8b5412cd8686a26e918bc4e3915f62.jpg) +(a) Left image + +![](images/7f50043f96b8b94c1cc23db8c000fb50ea595c1ec5a4b9ef43bb4f7bd6eec61c.jpg) + +![](images/5e1791ec47a9c7e3575890519a1ccc0ca7bb2233ee0f72e1551c12c7eaf01b22.jpg) + +![](images/cd77e08ff1e9f1550da5c59e78ca92a4d3bec9c313b2779e15e2420d05019ddc.jpg) + +![](images/e14747a501d9db7b03f6ee60d4b1c09f2da7d5b071554a18d19de03eb1e5bf53.jpg) + +![](images/e2a750cb2a2c649f08088eb9cec96c7a3e314231cf6fc8f5629956c0125de8a4.jpg) +(b) Mask-CFNet + +![](images/c7bbc7a5c3c66a40fbb8c3fa813b856795348ad0e92808fa087b0778b5b8dec4.jpg) + +![](images/ec1400aacc0d6b6fb1ba97fc8e215ca318d0e97914229d23d2fa74a0ca293dd5.jpg) + +![](images/fd60ff35742e81a118a8372c98269fe03794a7a949812e01a1fcda8aae3b9bf9.jpg) + +![](images/1eefab6a1ef7bd2bf012cae80d55d681812c8f3d2abdee87e2b096d4e68ba6fa.jpg) + +![](images/f71af24dd6b301d1aaa63d31bcdcb7a6004541ba2382d92ac80cf54188da809c.jpg) +(C) Mask-LacGwcNet +Figure 6. The convergence process with or without masked representation learning. It proves that a low ratio does not change the convergence process and the matching accuracy, while a high ratio affects the learning process and reduces the matching accuracy. + +Convergence process. In Fig. 6, we compare the performance and the convergence process with or without masked representation learning. Because CFNet and LacGwcNet have many outputs, we use the matching accuracy of the final output as the performance on the training dataset. Meanwhile, we test the performance of these methods on the source domain. We conclude that: (1) the results are very stable on the source domain; (2) compared with baselines, our method does not change the convergence process for the low mask ratio; (3) a high mask ratio will affect the learning process and reduce the matching accuracy. Stereo matching is a task highly related to location information. Mask ratio means more missing pixels, and it will affect the models to learn the correlation based on adjacent pixels in the feature matching module. + +Runtime. As shown in Tab. 2, we list the runtime of two models at the training and testing processes. Our method only adds a lightweight decoder to reconstruct the left image in the training process, which does not participate in the test process. Therefore, we draw the following conclusions: (1) for the training process, our method does not significantly prolong training time compared with baselines; (2) for the testing process, our approach is no different from baselines. + +Training epochs and generalization performance. As presented in Fig. 7, we draw two models' curves with or + +![](images/c222d2aa4af29bf66381cf56664c2eabf11890f1ba914664e4abe834c59e46ab.jpg) +(a) CFNet (training) + +![](images/4ff5c92d15978089e3b43d4348a8aa04fab1a8c16a34f4aea7ca928bb9c5fe3a.jpg) +(b) LacGwcNet (training) + +![](images/5ef88127f9058272805745f091242fdd84f3dfa7539e512b685df5a86e256cf7.jpg) +Figure 5. The examples of different models on the source and target datasets (without fine-tuning). From top to bottom are SceneFlow (source dataset), KITTI 2012&2015, ETH3D, and Middlebury (four target datasets). +(c) CFNet (testing) + +![](images/351a49878550615554a409b39437119b9e27610b5547f1a960ff8e4252d9622d.jpg) +(d) LacGwcNet (testing) + +
ModelMaskTrainingResolutionRuntime (s)
CFNet✓✓576 × 3200.89
CFNet✓X960 × 5760.052
CFNetX✓576 × 3200.84
CFNetXX960 × 5760.051
LacGwcNet✓✓576 × 3201.63
LacGwcNet✓X960 × 5760.264
LacGwcNetX✓576 × 3201.61
LacGwcNetXX960 × 5760.264
+ +Table 2. The runtime with different resolutions. It presents that our method does not significantly prolong training time and is no different from the baseline at the test runtime. + +without masked representation learning. When the models converge ( $>30$ epochs, as shown in Fig. 6), the results are stable on the source dataset. However, the generalization performance varies significantly between adjacent training epochs. Compared with the baselines, the models with masked representation learning can perform better and more stable. As shown in Tab. 3, we apply mean and variance to measure these fluctuations among different training epochs. By quantifying the changes, we find that the structural information can help the models achieve a relatively stable and better generalization performance. + +![](images/4ce9ce9b17a138d063ff27b08d503b2bf1c5129ba1719742286e4411fc4dab84.jpg) + +![](images/96965d7e7eb28ba7fef44de5a0ddf4b306b40732c1447001439663e3098f6d4f.jpg) + +![](images/acc6621ae0fc7c3b5dfad5e439c28fea0fc57c6b272afa1f8b9cc29d2c512d7b.jpg) + +![](images/cc848f0efa3f1eca1ae70656ad5f6a493ca4ede1866dbe63bf35ab251764934b.jpg) + +![](images/8fa2b09cd0791ce87142edcbed3bf83511b7cc3597b56caafc9508b0bd8900e5.jpg) +Figure 7. The generalization performance with or without masked representation among different epochs. It shows that the generalization performance varies significantly between adjacent training epochs, and our method achieves better performance and is more stable. + +![](images/71b3177b1fb2f24bcf84663db47df3dfacb00d4760808ce2638ca22b7ae273e0.jpg) + +![](images/3ac593ad6854eedede0074dd33457bb355807943d7e276f10639a9b3b13e7416.jpg) + +![](images/49c79bc0b4c769097f525df37e18cefa0554012eb573d0224ab47f627287b1ac.jpg) + +
M.MaskDataEPE (Mean)EPE (Var.)D1 (Mean)D1 (Var.)
CFNet✓KT-121.440.045.030.03
XKT-121.550.055.820.13
✓KT-151.050.016.080.07
XKT-151.110.016.560.19
✓ET0.560.026.630.21
XET0.690.037.240.30
✓MB1.860.1312.820.37
XMB2.450.1315.160.90
LacGwcNet✓KT-121.430.036.570.30
XKT-121.830.329.1710.46
✓KT-151.410.026.080.23
XKT-151.780.118.379.24
✓ET1.000.376.571.03
XET2.180.847.991.37
✓MB2.400.0217.300.89
XMB2.490.1918.280.51
+ +# 4.4. Comparisons + +Comparisons with generalization method. We compare our methods with other stereo matching methods, including well-known and generalized methods. Tab. 4 shows that our approach can help models improve generalization ability, indirectly proving that masked representation learn + +Table 3. Volatility comparison of with or without masked representation. It demonstrates that our method can help the models achieve a relatively stable and better generalization performance. + +
MethodKT-12 > 3pxKT-15 > 3pxMB > 2pxET > 1px
PSMNet [3]15.116.326.923.8
GWCNet [8]12.012.234.211.0
GANet [39]10.111.720.314.1
DSMNet [40]6.26.521.86.2
FC-DSM [41]5.56.212.06.0
CFNet [30]4.75.815.35.8
GF-PSMNet [16]5.34.610.96.2
Mask-CFNet4.85.813.75.7
Mask-LacGwcNet5.75.616.95.3
+ +Table 4. Cross-domain generalization evaluation (peak results) on four target datasets. All methods are only trained on the Scene-Flow dataset and tested on training images of four real datasets. + +ing can promote high-level representation feature learning. Note that all methods have volatility among training epochs, and we list the volatility as shown in the supplemental material. It prompted us to suggest that the scope of fluctuation should be attached to future work in stereo matching. + +Comparisons with fine-tuning. We fine-tune the models (CFNet and LacGwcNet) on the KITTI datasets to test the performance after fine-tuning. Tab. 5 shows that the results are similar to the trend of the SceneFlow dataset. When the masking ratio is low (< 25%), our method does not affect the performance and produces nearly the same re + +sults. As the ratio rises, the performance gradually declines. + +
MethodRatioKT-12 (Out-Noc)KT-15 (D1-all)
LacGwcNet [15]01.131.77
LacGwcNet [15]0.151.151.78
LacGwcNet [15]0.251.161.77
LacGwcNet [15]0.351.271.95
LacGwcNet [15]0.451.392.21
CFNet [30]01.231.88
CFNet [30]0.151.231.89
CFNet [30]0.251.271.91
CFNet [30]0.351.362.05
CFNet [30]0.451.482.28
+ +Table 5. The fine-tuning results on the KITTI dataset. It indicates the results are similar to the trend of the SceneFlow dataset. + +# 5. Discussion + +This paper has proposed a simple approach to alleviate the significant volatility of generalization performance. This section will discuss two topics: (1) When does our approach fail? (2) What is the real unseen domain? + +When does our approach fail? (1) In our study, we have added the channels of the feature extraction module to get a better reconstruction image. However, the generalization performance did not obtain remarkable improvement. Although larger models can handle two tasks well, it is not conducive to the fusion of multi-task learning because larger modules may learn features of two tasks separately. Similarly, it will weaken the effect of masked representation. (2) Excessive masking will cause the learnable matching area to shrink in the feature matching module, affecting matching accuracy or generalization. Thus, large-size masking is not recommended to improve generalization. + +What is the real unseen domain? In previous studies, nearly all papers used KITTI, ETH 3D, and Middlebury datasets as the unseen domains. However, can these datasets be represented the unseen domain? Although the color distribution, illumination, and authenticity are different between the source domain and the target domains, similar objects are included in the source domain, such as the street scenes in the Driving dataset (sub-dataset of SceneFlow), pedestrian in the Driving dataset, daily supplies in FlyingThing3D (sub-dataset of SceneFlow), etc. Herefore, we can get good results without fine-tuning models on the target dataset. Does it mean we can use these algorithms in practice? The answer is no. We use these models that work well on target datasets to deal with daily images collected by the zed camera or remote sensing images collected by satellite. As shown in Fig. 8, we can not get the expected results + +![](images/29f260035131d743d16fce9ec6dc1cf4902f88ca4fd2a435bbdb82a2f3a3912a.jpg) + +![](images/519ce6ec7e811d3f847ce80516a5fe7cfe8c0313c2fb142999d6b329f706cc22.jpg) +(a) remote sensing image +(c) daily image +Figure 8. The examples of failures in the real unseen domain. When elements do not appear in the source dataset, the models can not predict the expected results. + +![](images/ba832e6837f33195faf029f0326bb3ddd795577129db0ee1d88b1773d274e78f.jpg) + +![](images/a525c3383e040ae9a5a4aa3ed0027ff51c270627c3162023e1138e78bef3357c.jpg) +(b) disparity map +(d) disparity map + +for these images. We consider that the main reason is that these elements or scenes do not appear in the source dataset, such as the human body, buildings under the top view angle, etc. When the real unseen domain appears in the images, matching methods based on deep learning can not predict disparity correctly. Thus, we doubt the cross-domain performance of these datasets (KITTI, ETH 3D, and Middlebury) can represent actual cross-domain performance. + +# 6. Conclusion + +In this paper, we have proposed a simple masked representation method to address the problem of unstable generalization performance among different training epochs for stereo matching. Our approach is inspired by masked representation learning and multi-task learning to construct a pseudo-multi-task learning framework, which helps models learn better features in the feature extraction module to improve generalization performance and stability. Extensive experiments have proved the effectiveness of our method. Meanwhile, it also demonstrated that the current evaluation manner is unsuitable for measuring generalization performance. In the last, we discussed the failure of our approach and doubted the current definition of the unseen domain. + +Acknowledgements. This research was supported in part by Jiangxi Natural Science Foundations (20224BAB212009, 20212BBE53017), National Natural Science Foundation of China (61866027, 62271410, and 62001396), the Foundation of Key Laboratory of Jiangxi Province for Image Processing and Pattern Recognition (ET202204304), and the Doctor's Foundation of Nanchang Hangkong University (EA202204228, EA202204257). We thank the anonymous area chairs and reviewers for their helpful feedback. + +# References + +[1] Changjiang Cai, Matteo Poggi, Stefano Mattoccia, and Filippos Mordohai. Matching-space stereo networks for cross-domain generalization. In International Conference on 3D Vision (3DV), pages 364-373, 2020. 2 +[2] Rohan Chabra, Julian Straub, Christopher Sweeney, Richard Newcombe, and Henry Fuchs. Stereodnet: Dilated residual stereonet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786-11795, 2019. 1 +[3] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5410-5418, 2018. 1, 2, 3, 7 +[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proceedings of International Conference on Machine Learning (ICML), pages 1691-1703, 2020. 3 +[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 3 +[6] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In arXiv preprint, 2022. 3, 4 +[7] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. 5 +[8] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3273-3282, 2019. 7 +[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 3, 4 +[10] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 66-75, 2017. 1, 2, 3 +[11] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16263-16272, 2022. 1, 2, 3 +[12] Xing Li, Yangyu Fan, Zhibo Rao, Zhe Guo, and Guoyun Lv. Improving stereo matching generalization via fourier-based + +amplitude transform. IEEE Signal Processing Letters (SPL), 2022. 2, 5 +[13] Xing Li, Yangyu Fan, Zhibo Rao, Guoyun Lv, and Shiya Liu. Synthetic-to-real domain adaptation joint spatial feature transform for stereo matching. IEEE Signal Processing Letters (SPL), 29:60-64, 2021. 2 +[14] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6197-6206, 2021. 3 +[15] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 36, pages 1647-1655, 2022. 1, 5, 8 +[16] Biyang Liu, Huimin Yu, and Guodong Qi. Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13012-13021, 2022. 2, 3, 5, 7 +[17] Rui Liu, Chengxi Yang, Wenxiu Sun, Xiaogang Wang, and Hongsheng Li. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12757-12766, 2020. 2 +[18] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1871-1880, 2019. 2, 3 +[19] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 2, 5 +[20] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3061-3070, 2015. 5 +[21] Guang-Yu Nie, Ming-Ming Cheng, Yun Liu, Zhengfa Liang, Deng-Ping Fan, Yue Liu, and Yongtian Wang. Multi-level context ultra-aggregation for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3283-3291, 2019. 1 +[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536-2544, 2016. 3 +[23] Zhibo Rao, Yuchao Dai, Zhelun Shen, and Renjie He. Rethinking training strategy in stereo matching. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), pages 1-14, 2022. 1 +[24] Zhibo Rao, Mingyi He, Yuchao Dai, and Zhelun Shen. Patch attention network with generative adversarial model + +for semi-supervised binocular disparity prediction. The Visual Computer, pages 1-17, 2020. 1 +[25] Zhibo Rao, Mingyi He, Yuchao Dai, and Zhelun Shen. Sliding space-disparity transformer for stereo matching. Neural Computing and Applications (NCAA), pages 1-14, 2022. 1 +[26] Zhibo Rao, Mingyi He, Yuchao Dai, Zhidong Zhu, Bo Li, and Renjie He. NLca-net: a non-local context attention network for stereo matching. APSIPA Transactions on Signal and Information Processing, 9, 2020. 1 +[27] Zhibo Rao, Mingyi He, Zhidong Zhu, Yuchao Dai, and Renjie He. Bidirectional guided attention network for 3-d semantic detection of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(7):6138-6153, 2020. 2, 3 +[28] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nešić, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In German conference on pattern recognition, pages 31–42, 2014. 5 +[29] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3260-3269, 2017. 5 +[30] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13906-13915, 2021. 1, 2, 3, 5, 7, 8 +[31] Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. Adastereo: a simple and efficient approach for adaptive stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10328-10337, 2021. 2 +[32] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9661-9670, 2019. 1 +[33] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 195-204, 2019. 1, 2 +[34] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of International Conference on Machine Learning (ICML), pages 1096-1103, 2008. 3 +[35] Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12497-12506, 2021. 1 +[36] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo + +matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12981-12990, 2022. 1 +[37] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1959-1968, 2020. 1 +[38] Jure Zbontar and Yann LeCun. Computing the stereo matching cost with a convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1592-1599, 2015. 2 +[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 185-194, 2019. 1, 2, 3, 7 +[40] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In European Conference on Computer Vision (ECCV), pages 420-439, 2020. 1, 2, 3, 5, 7 +[41] Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13001-13011, 2022. 2, 3, 7 +[42] Youmin Zhang, Yimin Chen, Xiao Bai, Suihanjin Yu, Kun Yu, Zhiwei Li, and Kuiyuan Yang. Adaptive unimodal cost volume filtering for deep stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 12926-12934, 2020. 1 \ No newline at end of file diff --git a/2023/Masked Representation Learning for Domain Generalized Stereo Matching/images.zip b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..05f8363e8b90034d2b96e93e518b9d25a1e561db --- /dev/null +++ b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:452a4a2528dc90207739ebc4be036b660ae11059114413491052cf4cf6f7c480 +size 785003 diff --git a/2023/Masked Representation Learning for Domain Generalized Stereo Matching/layout.json b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c95a92b1e93ddce643a5d78ccb4db56e270a0a2b --- /dev/null +++ b/2023/Masked Representation Learning for Domain Generalized Stereo Matching/layout.json @@ -0,0 +1,8221 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 520, + 121 + ], + "type": "text", + "content": "Masked representation learning for domain generalized stereo matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": "Zhibo Rao" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{1,2\\boxtimes}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": ", Bangshu Xiong" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": ", Mingyi He" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": ", Yuchao Dai" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": ", Renjie He" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": ", Zhelun Shen" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": ", Xing Li" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": "Nanchang Hangkong University, Nanchang, China \n" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": "Northwestern Polytechnical University, Xi'an, China \n" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 59, + 141, + 533, + 199 + ], + "type": "text", + "content": "Baidu Research, Beijing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 201, + 505, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 201, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 85, + 201, + 505, + 227 + ], + "type": "text", + "content": "raoxi36@foxmail.com, xiongs@126.com, {myhe, daiyuchao, davidhrj} @nwpu.edu.cn, 1901213310@pku.edu.cn, lixing36@foxmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 280, + 290, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 593 + ], + "type": "text", + "content": "Recently, many deep stereo matching methods have begun to focus on cross-domain performance, achieving impressive achievements. However, these methods did not deal with the significant volatility of generalization performance among different training epochs. Inspired by masked representation learning and multi-task learning, this paper designs a simple and effective masked representation for domain generalized stereo matching. First, we feed the masked left and complete right images as input into the models. Then, we add a lightweight and simple decoder following the feature extraction module to recover the original left image. Finally, we train the models with two tasks (stereo matching and image reconstruction) as a pseudomulti-task learning framework, promoting models to learn structure information and to improve generalization performance. We implement our method on two well-known architectures (CFNet and LacGwcNet) to demonstrate its effectiveness. Experimental results on multi-datasets show that: (1) our method can be easily plugged into the current various stereo matching models to improve generalization performance; (2) our method can reduce the significant volatility of generalization performance among different training epochs; (3) we find that the current methods prefer to choose the best results among different training epochs as generalization performance, but it is impossible to select the best performance by ground truth in practice." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 614, + 128, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 128, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 128, + 627 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": "Stereo matching is a challenging research topic of computer vision, which aims to obtain the corresponding pixels between two rectified stereo images [3, 30, 37, 39]. It is essential in many applications, including autonomous [11], augmented reality [35], virtual reality [2], etc." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 259, + 422, + 371 + ], + "blocks": [ + { + "bbox": [ + 313, + 259, + 422, + 371 + ], + "lines": [ + { + "bbox": [ + 313, + 259, + 422, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 422, + 371 + ], + "type": "image", + "image_path": "affa0af305a6708276a5e279a71d68eb70622006b0651033db6c13149d67b3d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 374, + 391, + 384 + ], + "lines": [ + { + "bbox": [ + 353, + 374, + 391, + 384 + ], + "spans": [ + { + "bbox": [ + 353, + 374, + 391, + 384 + ], + "type": "text", + "content": "(a) CFNet" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "type": "text", + "content": "Figure 1. The generalization performance among different epochs on multi-datasets. When the models converge on the source domain (Sceneflow), the results are stable on the source domain. However, generalization performance has fluctuations on target datasets (KITTI 2012&2015, ETH3D, and Middlebury)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 440, + 259, + 548, + 371 + ], + "blocks": [ + { + "bbox": [ + 440, + 259, + 548, + 371 + ], + "lines": [ + { + "bbox": [ + 440, + 259, + 548, + 371 + ], + "spans": [ + { + "bbox": [ + 440, + 259, + 548, + 371 + ], + "type": "image", + "image_path": "5329afccfc24509ed3eabff4361888eb8a703cf52f1b6dc6cf066c053b736101.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 374, + 528, + 384 + ], + "lines": [ + { + "bbox": [ + 470, + 374, + 528, + 384 + ], + "spans": [ + { + "bbox": [ + 470, + 374, + 528, + 384 + ], + "type": "text", + "content": "(b) LacGwcNet" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 473, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 665 + ], + "type": "text", + "content": "In mainstream deep stereo matching methods, the main steps include four parts: feature extraction, cost volume, feature matching, and disparity regression [10]. To improve the accuracy or speed, researchers proposed many strategies to improve the above four parts [3, 25, 36]. For example, 1) scene awareness modules are applied to the feature extraction to improve accuracy [3, 21]; 2) similarity measurements are employed to enhance physical representation in cost volume construction [23, 26, 36]; 3) various feature matching modules are proposed to solve inefficient calculations or improve accuracy [15, 25, 39]; 4) loss functions of classification or other fields are joined to disparity regression to obtain unimodal results [42]. Although the above methods have achieved significant progress in accuracy or efficiency, they failed to obtain good generalization performance in unseen domains [30, 40]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": "Therefore, many approaches learn domain-invariant representation features to achieve better generalization capability [40]. There are three solutions to accomplish this objective: the unsupervised matching method [24,32,33], domain" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 162, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 162, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 162, + 712 + ], + "type": "inline_equation", + "content": "1 \\boxtimes" + }, + { + "bbox": [ + 58, + 702, + 162, + 712 + ], + "type": "text", + "content": " is the corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5435" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "adaptation techniques [12, 13, 17, 31], and domain generalization approaches [16, 30, 41]. The above solutions reveal that feature presentation is crucial for improving generalization capability [40, 41]. However, all domain generalized matching methods did not mention that the generalization performance varies significantly among different training epochs, as shown in Fig. 1. Namely, the results were unstable. These methods preferred to choose the best results to represent the generalization performance by testing the models of different training epochs. However, we can not employ ground truth (can not obtain) to select the best model among different training epochs in practice. Thus, it is crucial to keep a stable generalization performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 409 + ], + "type": "text", + "content": "On the other hand, many researchers introduced multi-task learning into stereo matching [18, 27]. For instance, Rao et al. proposed a bidirectional guided attention network to cope with semantic segmentation and stereo matching simultaneously [27]. Liu et al. employed the task-shared and task-specific manner to obtain a generalizable representation and used the loss weights to balance all tasks [18]. The multi-task learning proved that feature sharing and weight sharing are essential in different tasks, promoting the learning process and obtaining a better feature representation. Inspired by these multi-task learning methods, we use masked left images as input to reconstruct the original left image as another task. By building pseudo-multi-task learning, we obtain better structural information features, urging stereo matching models to perform better generalization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 410, + 289, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 289, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 289, + 625 + ], + "type": "text", + "content": "This paper addresses the domain generalization for stereo matching methods by masked representation learning. Our solution considerably increases generalization accuracy while reducing the volatility of generalization performance among different training epochs. First, we randomly mask (or remove) the part of the left image with a fixed ratio following a uniform distribution. Second, we add a simple decoder to the feature extraction module's tail to recover the original left image. Finally, we train the model with two tasks (stereo matching and image reconstruction) as a pseudo-multi-task learning framework and test generalization performance on the multi-datasets. In this task, our strategy helps existing methods to achieve significant gains in generalization performance. Furthermore, we exhibit the generalization performance over different training epochs to demonstrate the superiority of our approach. We hope these observations will help other tasks explore better generalization results. The main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 629, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 629, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 629, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 58, + 629, + 286, + 664 + ], + "type": "text", + "content": "- We combine masked image modeling and stereo matching as a pseudo-multi-task learning framework to increase generalization accuracy in stereo matching." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 666, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 666, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 666, + 286, + 712 + ], + "type": "text", + "content": "- Our approach is employed for various stereo matching networks, significantly improving cross-domain accuracy and reducing the volatility of generalization performance among different training epochs." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 120 + ], + "type": "text", + "content": "- We find that the accuracy of the existing stereo matching domain generalization methods varies significantly among different training epochs. Thus, we advise that stability should be evaluated in cross-domain methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 151, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 545, + 414 + ], + "type": "text", + "content": "Deep learning based stereo matching. Since Zbontar et al. used convolution neural networks (CNNs) to replace hand-crafted features and compute matching costs, many researchers started to apply this novel technology to update the traditional pipelines [38]. Mayer et al. built a cost volume in a correlation manner and created a sizeable synthetic dataset to meet the data requirements in stereo matching [19]. Kendall et al. first proposed 4D cost volume by stacking two view features and used 3D convolution to regularize costs, producing a profound impact in stereo matching [10]. Chang et al. introduced the scene awareness module to extract features and applied a 3D hourglass module to aggregate cost volume [3]. Zhang et al. utilized the traditional semi-global matching method to reform the feature matching module (3D convolution), capturing the local or global cost dependencies [39]. Shen et al. replaced the 3D hourglass module to cascade structure as a coarse to fine process, achieving a state-of-the-art generalization performance [30]. Li et al. employed a hierarchical network with recurrent refinement to modify a stacked cascaded architecture and proposed an adaptive group correlation layer to reduce the impact of erroneous rectification [11]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 415, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 557 + ], + "type": "text", + "content": "Domain adaptation stereo matching. To improve generalization performance, many researchers converted the original domain (synthetic scenes) to the target domain (real scenes) [31]. Liu et al. employed an end-to-end training framework with domain translation to tackle the problem of pixel distortion and stereo mismatch after translation [17]. Song et al. designed a bottom-up domain adaptation method based on color transfer and cost regularization to improve generalization without any learnable parameters [31]. Li et al. proposed Fourier-based amplitude transform (FAT), mapping the source image to the target style without altering semantic content [12]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Domain generalization stereo matching. Generalization performance is an important indicator that reveals the ability of networks to cope with the unseen domain. Tonioni et al. joint unsupervised and continuous online learning to preserve the model's accuracy in multiple environments [33]. Zhang et al. proposed a domain-invariant stereo matching network by regularizing the distribution of learned representations and extracting robust structural representations [40]. Cai et al. introduced matching functions and confidence measures to replace the learning-based feature extraction module, leading to superior generalization to unseen environments [1]. Li et al. introduced a transformer to the stereo matching task, which showed strong" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5436" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 305 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 305 + ], + "type": "image", + "image_path": "478bf12b845d215f3146b28ad61236ae44ab9284e0c995f2e7afeee02ecfc296.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 546, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 546, + 352 + ], + "type": "text", + "content": "Figure 2. The overall pipeline of our method. Inspired by masked representation learning and multi-task learning, we build a pseudo-multi-task learning framework based on image reconstruction and stereo matching. The feature extraction module can learn better structure information by feature fusion in pseudo-multi-task learning, improving generalization performance and stability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 371, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 456 + ], + "type": "text", + "content": "robustness [14]. Liu et al. leveraged the feature of a model trained on the large-scale dataset to deal with the domain shift and used a cosine similarity-based cost volume to graft task-oriented features [16]. Zhang et al. argued maintaining feature consistency between matching pixels is a vital factor for the generalization capability and employed a simple pixel-wise contrastive learning across the viewpoints [41]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": "Masked representation learning. Masked modeling is a highly successful way for pre-training in natural language processing (NLP) and computer vision (CV) [5, 9]. Vincent et al. proposed denoising autoencoders (DAE), which corrupted an input signal and learned to reconstruct the original, uncorrupted signal [34]. Pathak et al. presented an unsupervised visual feature learning algorithm driven by context-based pixel prediction, proving the effectiveness of CNN pre-training on classification, detection, and segmentation tasks [22]. Devlin et al. held out a portion of the input sequence and trained the transformer to predict the missing content [5]. Chen et al. inspired by progress in unsupervised representation learning in NLP and trained a sequence transformer to predict pixels without incorporating knowledge of structure information [4]. He et al. attempted to demonstrate that masked autoencoders (MAE) are scalable self-supervised learners for computer vision, which achieved the best accuracy among approaches that only used ImageNet-1K data [9]. Feichtenhofer et al. studied a simple extension of Masked Autoencoders (MAE) to spatiotemporal representation learning from videos [6]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 370, + 361, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 361, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 361, + 382 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 392, + 404, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 404, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 404, + 405 + ], + "type": "text", + "content": "3.1. Overall Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "content": "In a typical stereo matching method based on deep learning [3, 10, 11, 30, 39], the rectified left and right images are fed to the feature extraction module. Then, these methods construct the cost volume by concatenating the left and traveled right features. Finally, they use 3D convolutions to aggregate the cost volume and regress to the disparity map. Therefore, the poor generalization ability of stereo matching must be caused by two learnable modules (feature extraction module and feature matching module)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "In previous studies [16, 40], many achievements have demonstrated that feature representation in the feature extraction module is a vital factor in the model's generalization ability. Meanwhile, we have noticed that multi-task learning helps the models obtain a better matching accuracy, especially when another task has a recognition attribute (e.g., semantic segmentation, object detection, classification, etc.) [18, 27]. Why can recognition algorithms help models to get better performance in multi-task learning? Because feature-sharing in two or more tasks can help the models learn better structural or scene information [16, 18]. Inspired by the above research, we elegantly employ masked representation learning to build a pseudomulti-task learning framework, promoting models to learn better structural information in the feature extraction module to improve generalization ability, as shown in Fig. 2." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5437" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "Masked Input image. Following MAE [6, 9], we follow a uniform distribution to randomly mask pixels as gray in the left image with the fixed size and ratio (e.g., 0.15, 0.3, 0.45), as shown in Fig. 3. Unlike MAE, we use a low ratio and small mask size to mask pixels and do not employ random positions of patches. Because 1) we hope the models extrapolate the missing pixel from visible neighboring patches to perceive structural information; 2) lacking too many pixels is not conducive to learning correlation in the stereo matching task, mainly feature matching module; 3) stereo matching is sensitive to location information, and the goal is to build the relationship of correlation, so we do not use random positions. This design helps the models to learn structural information in the feature extraction module and does not impact the training process of other learnable modules (e.g., the feature matching module.)." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 275, + 165, + 338 + ], + "blocks": [ + { + "bbox": [ + 55, + 275, + 165, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 165, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 165, + 338 + ], + "type": "image", + "image_path": "ab67131582ec1dd77c91f24f2f8047d3bd30db09cf23c6e667ea23d88756f4b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 340, + 143, + 349 + ], + "lines": [ + { + "bbox": [ + 78, + 340, + 143, + 349 + ], + "spans": [ + { + "bbox": [ + 78, + 340, + 143, + 349 + ], + "type": "text", + "content": "(a) original left image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 275, + 279, + 338 + ], + "blocks": [ + { + "bbox": [ + 170, + 275, + 279, + 338 + ], + "lines": [ + { + "bbox": [ + 170, + 275, + 279, + 338 + ], + "spans": [ + { + "bbox": [ + 170, + 275, + 279, + 338 + ], + "type": "image", + "image_path": "10520036269da1f7527a9558c046b2e8fe285d4f3d3799200f082c6adfa94732.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 340, + 246, + 348 + ], + "lines": [ + { + "bbox": [ + 204, + 340, + 246, + 348 + ], + "spans": [ + { + "bbox": [ + 204, + 340, + 246, + 348 + ], + "type": "text", + "content": "(b) ratio: 0.15" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 55, + 350, + 165, + 413 + ], + "blocks": [ + { + "bbox": [ + 55, + 350, + 165, + 413 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 165, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 165, + 413 + ], + "type": "image", + "image_path": "90d25ae6e42177911a0b4892ed6e2b675d7a5dc8589d0d453133203be863cf0d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 415, + 129, + 423 + ], + "lines": [ + { + "bbox": [ + 91, + 415, + 129, + 423 + ], + "spans": [ + { + "bbox": [ + 91, + 415, + 129, + 423 + ], + "type": "text", + "content": "(c) ratio: 0.3" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 170, + 350, + 279, + 413 + ], + "blocks": [ + { + "bbox": [ + 170, + 350, + 279, + 413 + ], + "lines": [ + { + "bbox": [ + 170, + 350, + 279, + 413 + ], + "spans": [ + { + "bbox": [ + 170, + 350, + 279, + 413 + ], + "type": "image", + "image_path": "e60cead34bad873e09bf5cc61e5f887cd99847ed56a67c5c897fe7a2f2ea7efb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 415, + 244, + 423 + ], + "lines": [ + { + "bbox": [ + 206, + 415, + 244, + 423 + ], + "spans": [ + { + "bbox": [ + 206, + 415, + 244, + 423 + ], + "type": "text", + "content": "(d) ratio: 0.4" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 438, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 483 + ], + "type": "text", + "content": "Figure 3. Illustration of mask sampling strategy. It determines the difficulty of the reconstruction task and affects the reconstruction quality. More masked pixels mean more difficult reconstruction tasks and less reference information in stereo matching." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 497, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 556 + ], + "type": "text", + "content": "After masking the left image, we feed the masked left and unmasked right images to the feature extraction module to obtain the left and right features. Then, we design a simple and efficient decoder to predict the missing pixels by the left features, introduced next." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "content": "Decoder Design. We treat the feature extraction module as an encoder to encode the masked left image and obtain the left features. In stereo matching, all methods use down-sampling in the feature extraction process to reduce computation and memory utilization. Therefore, we get the features with a compressed size (e.g., " + }, + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "inline_equation", + "content": "H / 4 \\times W / 4 \\times F" + }, + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "content": ") after the feature extraction module. To reconstruct the target image, we design a lightweight decoder to decode the features to the original size, as shown in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "We only employ the decoder module to cope with the image reconstruction task during training. In other words, the decoder module does not work in the testing process. Therefore, the decoder module will not affect the runtime of the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "existing stereo matching models in the testing process. In this paper, our decoder design is very lightweight and only consists of three 2D convolutions and two up-sampling operations. Thus, the training process will not be significantly prolonged in our design." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 133, + 546, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 263 + ], + "type": "text", + "content": "Reconstruction Target. The decoder module reconstructs the left image by predicting the missing pixels. The last layer of the decoder module is a line projection without batch normalization (BN) or ReLU Activation Function (ReLU). Following MAE, we use normalized pixels as the reconstruction target to improve representation quality, and we only compute the loss on the masked pixel of the left image. Our idea is to use image reconstruction and stereo matching to build a pseudo-multi-task learning framework, promoting structural representation learning to improve cross-domain performance further." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 271, + 399, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 271, + 399, + 283 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 399, + 283 + ], + "type": "text", + "content": "3.2. Loss Functions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 290, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 326 + ], + "type": "text", + "content": "We apply image reconstruction and stereo matching to construct pseudo-multi-task learning. Hence, the loss functions contain two parts, as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 326, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 373 + ], + "type": "text", + "content": "Reconstruction loss function. We employ the mean squared error (MSE) to compute the loss " + }, + { + "bbox": [ + 304, + 326, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 304, + 326, + 545, + 373 + ], + "type": "text", + "content": " between reconstructed and original left images in the pixel space. The reconstruction loss function " + }, + { + "bbox": [ + 304, + 326, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 304, + 326, + 545, + 373 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 354, + 382, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 354, + 382, + 545, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(I _ {o} (i) - I _ {r} (i)\\right) ^ {2}, \\tag {1}", + "image_path": "bf7d61daeed2adc7c883198e082fc3f32bb0dd23a2a45c1e5dd51124526f9e06.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "content": " denotes the number of masked pixels, " + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "content": " is the identifier of masked pixels, " + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "content": " represents the original left image, and " + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "content": " indicates the reconstructed masked pixels." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": "Matching loss function. The matching loss " + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": " in different papers is different. Most of these methods use the mean absolute error (MAE) or Smooth " + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 460, + 545, + 555 + ], + "type": "text", + "content": " loss as based loss functions. Then, they add photometric or other loss functions to improve performance. In our method, we follow the loss functions of previous matching works and continue to adopt the same manner as the loss following their paper. Therefore, the total loss can be presented as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 383, + 565, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 565, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 383, + 565, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r} + \\mathcal {L} _ {m}. \\tag {2}", + "image_path": "124a6f86c2613e88b3e7b4abfa490819d43ab6bdc957626e1131a749f30c8b26.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 587, + 509, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 509, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 509, + 601 + ], + "type": "text", + "content": "3.3. Algorithm applicability and advantage" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Algorithm applicability. Although many stereo matching methods have been proposed recently, they do not break away from the existing framework (four parts mentioned in Sec. 1). All models have feature extraction modules, although the style of the modules is variable. Hence, our method can be plugged into the current matching methods by treating the feature extraction modules as encoders and adding a simple decoder. Our method works for all current matching algorithms, not just a few." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5438" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 77, + 165, + 188 + ], + "blocks": [ + { + "bbox": [ + 53, + 77, + 165, + 188 + ], + "lines": [ + { + "bbox": [ + 53, + 77, + 165, + 188 + ], + "spans": [ + { + "bbox": [ + 53, + 77, + 165, + 188 + ], + "type": "image", + "image_path": "5ba0e301cd3617262d03ba9099226fdf03688dac97bdbde6c3d7931ba0dcc170.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 183, + 77, + 291, + 188 + ], + "blocks": [ + { + "bbox": [ + 183, + 77, + 291, + 188 + ], + "lines": [ + { + "bbox": [ + 183, + 77, + 291, + 188 + ], + "spans": [ + { + "bbox": [ + 183, + 77, + 291, + 188 + ], + "type": "image", + "image_path": "d80a5d57abf3390af94b9b5faf3138d12328d1b7a51dd0a6cf644658b9d6de59.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 195, + 165, + 308 + ], + "blocks": [ + { + "bbox": [ + 53, + 195, + 165, + 308 + ], + "lines": [ + { + "bbox": [ + 53, + 195, + 165, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 165, + 308 + ], + "type": "image", + "image_path": "b88dc55b93ec861587fc6bf2bad67c45e89ac4ccbbe34182d3f59ac5eba67baa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 342 + ], + "type": "text", + "content": "Figure 4. CFNet with different masking ratio. As the masking ratio rises, performance rises first and then declines." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 182, + 197, + 291, + 308 + ], + "blocks": [ + { + "bbox": [ + 182, + 197, + 291, + 308 + ], + "lines": [ + { + "bbox": [ + 182, + 197, + 291, + 308 + ], + "spans": [ + { + "bbox": [ + 182, + 197, + 291, + 308 + ], + "type": "image", + "image_path": "7153230bbd1d2183fe756480b52b9a96c4d016cb53321b9325a0ce715698d21b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 363, + 288, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 288, + 507 + ], + "type": "text", + "content": "Algorithm advantage. Compared with current domain adaptation and domain generalization methods in stereo matching, our approach has many advantages regarding runtime and convenience, as shown in follows. (1) Unlike domain adaptation technologies, our method does not need an additional training process or access to the target domain data. Meanwhile, our decoder module is a lightweight structure. Thus, it will not significantly prolong the training process of the existing algorithms. (2) Unlike domain generalization methods, our image reconstruction branch does not participate in testing. Therefore, our approach does not affect the runtime of the current matching methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 538, + 212, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 212, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 212, + 551 + ], + "type": "text", + "content": "4.1. Datasets & Evaluation metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "type": "text", + "content": "The current domain generalized evaluation manner is that the models are trained on the source domain and tested on the target domain. Thus, we first briefly describe these datasets based on different domains and then introduce evaluation metrics on these datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "type": "text", + "content": "Source domain. We train all models on the Sceneflow dataset [19] in the experiment with the same schedule." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "Target domain. Following previous works [12, 16, 30], we evaluate all models on the KITTI 2012&2015 (KT-12 and KT-15) [7, 20], ETH3D (ET) [29], and Middlebury (MB) [28] without fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Evaluation metric. Following previous works [16, 40], we apply end-point-error (EPE) and the percentage of error" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "pixels larger than " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " pixels " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "(>t\\mathrm{px})" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " as evaluation metrics. According to evaluated websites, we set " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "t = 1,2,3" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", which correspond to the default thresholds of " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " (bad 1.0) on the ETH3D, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "t = 2" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " (bad 2.0) on the Middlebury, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "t = 3(D_1)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " on the KITTI 2012&2015, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 138, + 440, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 138, + 440, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 138, + 440, + 152 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 157, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 205 + ], + "type": "text", + "content": "In this paper, we implement our strategy with state-of-the-art methods (CFNet [30] and LacGwcNet [15]) by PyTorch. According to our goal, we need to decide on the hyper-parameters and training schedule, as presented below." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 205, + 546, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 546, + 289 + ], + "type": "text", + "content": "Hyper-parameters. We set the maximum disparity " + }, + { + "bbox": [ + 304, + 205, + 546, + 289 + ], + "type": "inline_equation", + "content": "D = 256" + }, + { + "bbox": [ + 304, + 205, + 546, + 289 + ], + "type": "text", + "content": " in CFNet and LacGwcNet to ensure that nearly all possible disparity values in the images could be detected. We apply Adam Optimizer " + }, + { + "bbox": [ + 304, + 205, + 546, + 289 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999)" + }, + { + "bbox": [ + 304, + 205, + 546, + 289 + ], + "type": "text", + "content": " to optimize the model. In training, we set a mini-batch size of 1 image pair per GPU (six on 6 GPUs) and all loss weights following their papers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 289, + 547, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 372 + ], + "type": "text", + "content": "Training schedule. We only train models on the Scene-Flow dataset. The learning rate is initially set to " + }, + { + "bbox": [ + 304, + 289, + 547, + 372 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 289, + 547, + 372 + ], + "type": "text", + "content": " for 30 epochs and then reduced to " + }, + { + "bbox": [ + 304, + 289, + 547, + 372 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 289, + 547, + 372 + ], + "type": "text", + "content": " for the other 10 epochs. After this process, we obtain the final models. For data augmentation, we use random cropping, random resize, random shift, and random chromatic transformations to cope with the input image on the SceneFlow dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 379, + 405, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 405, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 405, + 392 + ], + "type": "text", + "content": "4.3. Main Properties" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 321, + 405, + 531, + 540 + ], + "blocks": [ + { + "bbox": [ + 321, + 405, + 531, + 540 + ], + "lines": [ + { + "bbox": [ + 321, + 405, + 531, + 540 + ], + "spans": [ + { + "bbox": [ + 321, + 405, + 531, + 540 + ], + "type": "table", + "html": "
RatioTypeKT-12KT-15ETMB
0EPE1.111.550.692.45
0.15EPE1.051.440.561.86
0.25EPE1.051.450.571.85
0.35EPE1.131.530.652.14
0.45EPE1.271.630.752.57
0t-px error5.836.567.2515.17
0.15t-px error5.016.096.6412.82
0.25t-px error5.126.206.7412.60
0.35t-px error5.636.486.9414.81
0.45t-px error4.006.797.0415.32
", + "image_path": "a71d3b94046edcf1045441588ee04fc67361cc520e6dc6f47fcef70dee2b7813.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 548, + 545, + 583 + ], + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 583 + ], + "type": "text", + "content": "Table 1. The generalization performance of CFNet with different masking ratios. A low ratio improves generalization performance, while a high ratio hurts generalization performance." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Masking ratio. Fig. 4 and Tab. 1 show the influence of the masking ratio (we use an average of ten epochs as generalization performance.). When the masking ratio is low, it does not affect the performance (source domain) but improves generalization performance. As the masking ratio increases, the performance (source domain) gradually declines, and generalization performance rises first and then falls. Thus, a high masking ratio is not unsuited for this pseudo-multi-task framework due to two factors: (1) a high masking ratio will increase the difficulty of reconstruction," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5439" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": "which is a disadvantage to the learning process; (2) excessive masking will reduce the learnable matching area in the following modules, such as the feature matching module. Some qualitative results of our final models on the four realistic datasets are displayed in Fig. 5." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 143, + 126, + 187 + ], + "blocks": [ + { + "bbox": [ + 49, + 143, + 126, + 187 + ], + "lines": [ + { + "bbox": [ + 49, + 143, + 126, + 187 + ], + "spans": [ + { + "bbox": [ + 49, + 143, + 126, + 187 + ], + "type": "image", + "image_path": "53df3f3a1d653cb1095b8a939731203a52148992d2cb234ebd1596ac38456b54.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 188, + 126, + 212 + ], + "blocks": [ + { + "bbox": [ + 49, + 188, + 126, + 212 + ], + "lines": [ + { + "bbox": [ + 49, + 188, + 126, + 212 + ], + "spans": [ + { + "bbox": [ + 49, + 188, + 126, + 212 + ], + "type": "image", + "image_path": "d93e7137ed45e36da76987ec7f393612ecb6fcde09fa762ca9b9f2e03c7c59dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 213, + 126, + 237 + ], + "blocks": [ + { + "bbox": [ + 49, + 213, + 126, + 237 + ], + "lines": [ + { + "bbox": [ + 49, + 213, + 126, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 213, + 126, + 237 + ], + "type": "image", + "image_path": "53ae5ee96e489323508aefbe2b1b65fcd47e8a3f33fff356929df4f7c6651552.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 238, + 126, + 278 + ], + "blocks": [ + { + "bbox": [ + 49, + 238, + 126, + 278 + ], + "lines": [ + { + "bbox": [ + 49, + 238, + 126, + 278 + ], + "spans": [ + { + "bbox": [ + 49, + 238, + 126, + 278 + ], + "type": "image", + "image_path": "9a45c45e1f6ee3fcfef2bc5cf8df1c6e416b15c2d2f3416ed0b8ca34d46d8dbb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 279, + 126, + 333 + ], + "blocks": [ + { + "bbox": [ + 49, + 279, + 126, + 333 + ], + "lines": [ + { + "bbox": [ + 49, + 279, + 126, + 333 + ], + "spans": [ + { + "bbox": [ + 49, + 279, + 126, + 333 + ], + "type": "image", + "image_path": "a5ac5bb5ae135cbe847d9468f7624fbcba8b5412cd8686a26e918bc4e3915f62.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 335, + 108, + 343 + ], + "lines": [ + { + "bbox": [ + 67, + 335, + 108, + 343 + ], + "spans": [ + { + "bbox": [ + 67, + 335, + 108, + 343 + ], + "type": "text", + "content": "(a) Left image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 129, + 143, + 205, + 187 + ], + "blocks": [ + { + "bbox": [ + 129, + 143, + 205, + 187 + ], + "lines": [ + { + "bbox": [ + 129, + 143, + 205, + 187 + ], + "spans": [ + { + "bbox": [ + 129, + 143, + 205, + 187 + ], + "type": "image", + "image_path": "7f50043f96b8b94c1cc23db8c000fb50ea595c1ec5a4b9ef43bb4f7bd6eec61c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 129, + 188, + 205, + 212 + ], + "blocks": [ + { + "bbox": [ + 129, + 188, + 205, + 212 + ], + "lines": [ + { + "bbox": [ + 129, + 188, + 205, + 212 + ], + "spans": [ + { + "bbox": [ + 129, + 188, + 205, + 212 + ], + "type": "image", + "image_path": "5e1791ec47a9c7e3575890519a1ccc0ca7bb2233ee0f72e1551c12c7eaf01b22.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 129, + 213, + 205, + 236 + ], + "blocks": [ + { + "bbox": [ + 129, + 213, + 205, + 236 + ], + "lines": [ + { + "bbox": [ + 129, + 213, + 205, + 236 + ], + "spans": [ + { + "bbox": [ + 129, + 213, + 205, + 236 + ], + "type": "image", + "image_path": "cd77e08ff1e9f1550da5c59e78ca92a4d3bec9c313b2779e15e2420d05019ddc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 129, + 238, + 205, + 278 + ], + "blocks": [ + { + "bbox": [ + 129, + 238, + 205, + 278 + ], + "lines": [ + { + "bbox": [ + 129, + 238, + 205, + 278 + ], + "spans": [ + { + "bbox": [ + 129, + 238, + 205, + 278 + ], + "type": "image", + "image_path": "e14747a501d9db7b03f6ee60d4b1c09f2da7d5b071554a18d19de03eb1e5bf53.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 129, + 279, + 205, + 333 + ], + "blocks": [ + { + "bbox": [ + 129, + 279, + 205, + 333 + ], + "lines": [ + { + "bbox": [ + 129, + 279, + 205, + 333 + ], + "spans": [ + { + "bbox": [ + 129, + 279, + 205, + 333 + ], + "type": "image", + "image_path": "e2a750cb2a2c649f08088eb9cec96c7a3e314231cf6fc8f5629956c0125de8a4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 335, + 190, + 342 + ], + "lines": [ + { + "bbox": [ + 144, + 335, + 190, + 342 + ], + "spans": [ + { + "bbox": [ + 144, + 335, + 190, + 342 + ], + "type": "text", + "content": "(b) Mask-CFNet" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 208, + 143, + 285, + 187 + ], + "blocks": [ + { + "bbox": [ + 208, + 143, + 285, + 187 + ], + "lines": [ + { + "bbox": [ + 208, + 143, + 285, + 187 + ], + "spans": [ + { + "bbox": [ + 208, + 143, + 285, + 187 + ], + "type": "image", + "image_path": "c7bbc7a5c3c66a40fbb8c3fa813b856795348ad0e92808fa087b0778b5b8dec4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 208, + 188, + 285, + 212 + ], + "blocks": [ + { + "bbox": [ + 208, + 188, + 285, + 212 + ], + "lines": [ + { + "bbox": [ + 208, + 188, + 285, + 212 + ], + "spans": [ + { + "bbox": [ + 208, + 188, + 285, + 212 + ], + "type": "image", + "image_path": "ec1400aacc0d6b6fb1ba97fc8e215ca318d0e97914229d23d2fa74a0ca293dd5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 208, + 213, + 285, + 237 + ], + "blocks": [ + { + "bbox": [ + 208, + 213, + 285, + 237 + ], + "lines": [ + { + "bbox": [ + 208, + 213, + 285, + 237 + ], + "spans": [ + { + "bbox": [ + 208, + 213, + 285, + 237 + ], + "type": "image", + "image_path": "fd60ff35742e81a118a8372c98269fe03794a7a949812e01a1fcda8aae3b9bf9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 208, + 238, + 285, + 278 + ], + "blocks": [ + { + "bbox": [ + 208, + 238, + 285, + 278 + ], + "lines": [ + { + "bbox": [ + 208, + 238, + 285, + 278 + ], + "spans": [ + { + "bbox": [ + 208, + 238, + 285, + 278 + ], + "type": "image", + "image_path": "1eefab6a1ef7bd2bf012cae80d55d681812c8f3d2abdee87e2b096d4e68ba6fa.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 208, + 279, + 285, + 333 + ], + "blocks": [ + { + "bbox": [ + 208, + 279, + 285, + 333 + ], + "lines": [ + { + "bbox": [ + 208, + 279, + 285, + 333 + ], + "spans": [ + { + "bbox": [ + 208, + 279, + 285, + 333 + ], + "type": "image", + "image_path": "f71af24dd6b301d1aaa63d31bcdcb7a6004541ba2382d92ac80cf54188da809c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 335, + 278, + 342 + ], + "lines": [ + { + "bbox": [ + 215, + 335, + 278, + 342 + ], + "spans": [ + { + "bbox": [ + 215, + 335, + 278, + 342 + ], + "type": "text", + "content": "(C) Mask-LacGwcNet" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 343, + 546, + 388 + ], + "lines": [ + { + "bbox": [ + 305, + 343, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 546, + 388 + ], + "type": "text", + "content": "Figure 6. The convergence process with or without masked representation learning. It proves that a low ratio does not change the convergence process and the matching accuracy, while a high ratio affects the learning process and reduces the matching accuracy." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 414, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 593 + ], + "type": "text", + "content": "Convergence process. In Fig. 6, we compare the performance and the convergence process with or without masked representation learning. Because CFNet and LacGwcNet have many outputs, we use the matching accuracy of the final output as the performance on the training dataset. Meanwhile, we test the performance of these methods on the source domain. We conclude that: (1) the results are very stable on the source domain; (2) compared with baselines, our method does not change the convergence process for the low mask ratio; (3) a high mask ratio will affect the learning process and reduce the matching accuracy. Stereo matching is a task highly related to location information. Mask ratio means more missing pixels, and it will affect the models to learn the correlation based on adjacent pixels in the feature matching module." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "content": "Runtime. As shown in Tab. 2, we list the runtime of two models at the training and testing processes. Our method only adds a lightweight decoder to reconstruct the left image in the training process, which does not participate in the test process. Therefore, we draw the following conclusions: (1) for the training process, our method does not significantly prolong training time compared with baselines; (2) for the testing process, our approach is no different from baselines." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Training epochs and generalization performance. As presented in Fig. 7, we draw two models' curves with or" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 314, + 77, + 421, + 187 + ], + "blocks": [ + { + "bbox": [ + 314, + 77, + 421, + 187 + ], + "lines": [ + { + "bbox": [ + 314, + 77, + 421, + 187 + ], + "spans": [ + { + "bbox": [ + 314, + 77, + 421, + 187 + ], + "type": "image", + "image_path": "c222d2aa4af29bf66381cf56664c2eabf11890f1ba914664e4abe834c59e46ab.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 190, + 413, + 202 + ], + "lines": [ + { + "bbox": [ + 331, + 190, + 413, + 202 + ], + "spans": [ + { + "bbox": [ + 331, + 190, + 413, + 202 + ], + "type": "text", + "content": "(a) CFNet (training)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 441, + 77, + 548, + 187 + ], + "blocks": [ + { + "bbox": [ + 441, + 77, + 548, + 187 + ], + "lines": [ + { + "bbox": [ + 441, + 77, + 548, + 187 + ], + "spans": [ + { + "bbox": [ + 441, + 77, + 548, + 187 + ], + "type": "image", + "image_path": "4ff5c92d15978089e3b43d4348a8aa04fab1a8c16a34f4aea7ca928bb9c5fe3a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 190, + 551, + 202 + ], + "lines": [ + { + "bbox": [ + 447, + 190, + 551, + 202 + ], + "spans": [ + { + "bbox": [ + 447, + 190, + 551, + 202 + ], + "type": "text", + "content": "(b) LacGwcNet (training)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 313, + 209, + 421, + 319 + ], + "blocks": [ + { + "bbox": [ + 46, + 356, + 287, + 400 + ], + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 400 + ], + "type": "text", + "content": "Figure 5. The examples of different models on the source and target datasets (without fine-tuning). From top to bottom are SceneFlow (source dataset), KITTI 2012&2015, ETH3D, and Middlebury (four target datasets)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 209, + 421, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 421, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 421, + 319 + ], + "type": "image", + "image_path": "5ef88127f9058272805745f091242fdd84f3dfa7539e512b685df5a86e256cf7.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 322, + 410, + 334 + ], + "lines": [ + { + "bbox": [ + 333, + 322, + 410, + 334 + ], + "spans": [ + { + "bbox": [ + 333, + 322, + 410, + 334 + ], + "type": "text", + "content": "(c) CFNet (testing)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 441, + 209, + 548, + 319 + ], + "blocks": [ + { + "bbox": [ + 441, + 209, + 548, + 319 + ], + "lines": [ + { + "bbox": [ + 441, + 209, + 548, + 319 + ], + "spans": [ + { + "bbox": [ + 441, + 209, + 548, + 319 + ], + "type": "image", + "image_path": "351a49878550615554a409b39437119b9e27610b5547f1a960ff8e4252d9622d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 322, + 548, + 334 + ], + "lines": [ + { + "bbox": [ + 450, + 322, + 548, + 334 + ], + "spans": [ + { + "bbox": [ + 450, + 322, + 548, + 334 + ], + "type": "text", + "content": "(d) LacGwcNet (testing)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 307, + 402, + 545, + 515 + ], + "blocks": [ + { + "bbox": [ + 307, + 402, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 307, + 402, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 307, + 402, + 545, + 515 + ], + "type": "table", + "html": "
ModelMaskTrainingResolutionRuntime (s)
CFNet✓✓576 × 3200.89
CFNet✓X960 × 5760.052
CFNetX✓576 × 3200.84
CFNetXX960 × 5760.051
LacGwcNet✓✓576 × 3201.63
LacGwcNet✓X960 × 5760.264
LacGwcNetX✓576 × 3201.61
LacGwcNetXX960 × 5760.264
", + "image_path": "78ba66fb7024cb54730b0b36b6f35269de133eb91b469e114a3a9d2d87a32e5c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 523, + 545, + 555 + ], + "lines": [ + { + "bbox": [ + 305, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 523, + 545, + 555 + ], + "type": "text", + "content": "Table 2. The runtime with different resolutions. It presents that our method does not significantly prolong training time and is no different from the baseline at the test runtime." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "without masked representation learning. When the models converge (" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": ">30" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " epochs, as shown in Fig. 6), the results are stable on the source dataset. However, the generalization performance varies significantly between adjacent training epochs. Compared with the baselines, the models with masked representation learning can perform better and more stable. As shown in Tab. 3, we apply mean and variance to measure these fluctuations among different training epochs. By quantifying the changes, we find that the structural information can help the models achieve a relatively stable and better generalization performance." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5440" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 77, + 165, + 188 + ], + "blocks": [ + { + "bbox": [ + 53, + 77, + 165, + 188 + ], + "lines": [ + { + "bbox": [ + 53, + 77, + 165, + 188 + ], + "spans": [ + { + "bbox": [ + 53, + 77, + 165, + 188 + ], + "type": "image", + "image_path": "4ce9ce9b17a138d063ff27b08d503b2bf1c5129ba1719742286e4411fc4dab84.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 78, + 291, + 188 + ], + "blocks": [ + { + "bbox": [ + 184, + 78, + 291, + 188 + ], + "lines": [ + { + "bbox": [ + 184, + 78, + 291, + 188 + ], + "spans": [ + { + "bbox": [ + 184, + 78, + 291, + 188 + ], + "type": "image", + "image_path": "96965d7e7eb28ba7fef44de5a0ddf4b306b40732c1447001439663e3098f6d4f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 313, + 78, + 420, + 188 + ], + "blocks": [ + { + "bbox": [ + 313, + 78, + 420, + 188 + ], + "lines": [ + { + "bbox": [ + 313, + 78, + 420, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 78, + 420, + 188 + ], + "type": "image", + "image_path": "acc6621ae0fc7c3b5dfad5e439c28fea0fc57c6b272afa1f8b9cc29d2c512d7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 439, + 72, + 547, + 188 + ], + "blocks": [ + { + "bbox": [ + 439, + 72, + 547, + 188 + ], + "lines": [ + { + "bbox": [ + 439, + 72, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 439, + 72, + 547, + 188 + ], + "type": "image", + "image_path": "cc848f0efa3f1eca1ae70656ad5f6a493ca4ede1866dbe63bf35ab251764934b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 53, + 196, + 164, + 308 + ], + "blocks": [ + { + "bbox": [ + 53, + 196, + 164, + 308 + ], + "lines": [ + { + "bbox": [ + 53, + 196, + 164, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 164, + 308 + ], + "type": "image", + "image_path": "8fa2b09cd0791ce87142edcbed3bf83511b7cc3597b56caafc9508b0bd8900e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 320, + 546, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 546, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 546, + 343 + ], + "type": "text", + "content": "Figure 7. The generalization performance with or without masked representation among different epochs. It shows that the generalization performance varies significantly between adjacent training epochs, and our method achieves better performance and is more stable." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 182, + 196, + 291, + 308 + ], + "blocks": [ + { + "bbox": [ + 182, + 196, + 291, + 308 + ], + "lines": [ + { + "bbox": [ + 182, + 196, + 291, + 308 + ], + "spans": [ + { + "bbox": [ + 182, + 196, + 291, + 308 + ], + "type": "image", + "image_path": "71b3177b1fb2f24bcf84663db47df3dfacb00d4760808ce2638ca22b7ae273e0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 196, + 419, + 308 + ], + "blocks": [ + { + "bbox": [ + 313, + 196, + 419, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 196, + 419, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 419, + 308 + ], + "type": "image", + "image_path": "3ac593ad6854eedede0074dd33457bb355807943d7e276f10639a9b3b13e7416.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 440, + 191, + 547, + 308 + ], + "blocks": [ + { + "bbox": [ + 440, + 191, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 440, + 191, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 440, + 191, + 547, + 308 + ], + "type": "image", + "image_path": "49c79bc0b4c769097f525df37e18cefa0554012eb573d0224ab47f627287b1ac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 50, + 361, + 284, + 572 + ], + "blocks": [ + { + "bbox": [ + 50, + 361, + 284, + 572 + ], + "lines": [ + { + "bbox": [ + 50, + 361, + 284, + 572 + ], + "spans": [ + { + "bbox": [ + 50, + 361, + 284, + 572 + ], + "type": "table", + "html": "
M.MaskDataEPE (Mean)EPE (Var.)D1 (Mean)D1 (Var.)
CFNet✓KT-121.440.045.030.03
XKT-121.550.055.820.13
✓KT-151.050.016.080.07
XKT-151.110.016.560.19
✓ET0.560.026.630.21
XET0.690.037.240.30
✓MB1.860.1312.820.37
XMB2.450.1315.160.90
LacGwcNet✓KT-121.430.036.570.30
XKT-121.830.329.1710.46
✓KT-151.410.026.080.23
XKT-151.780.118.379.24
✓ET1.000.376.571.03
XET2.180.847.991.37
✓MB2.400.0217.300.89
XMB2.490.1918.280.51
", + "image_path": "e1c24b83b73940e0403844674c66b545e5519af3efe0c1f4df2ec6d801d62ece.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 132, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 132, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 132, + 647 + ], + "type": "text", + "content": "4.4. Comparisons" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Comparisons with generalization method. We compare our methods with other stereo matching methods, including well-known and generalized methods. Tab. 4 shows that our approach can help models improve generalization ability, indirectly proving that masked representation learn" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 361, + 543, + 518 + ], + "blocks": [ + { + "bbox": [ + 46, + 581, + 287, + 615 + ], + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 615 + ], + "type": "text", + "content": "Table 3. Volatility comparison of with or without masked representation. It demonstrates that our method can help the models achieve a relatively stable and better generalization performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 361, + 543, + 518 + ], + "lines": [ + { + "bbox": [ + 309, + 361, + 543, + 518 + ], + "spans": [ + { + "bbox": [ + 309, + 361, + 543, + 518 + ], + "type": "table", + "html": "
MethodKT-12 > 3pxKT-15 > 3pxMB > 2pxET > 1px
PSMNet [3]15.116.326.923.8
GWCNet [8]12.012.234.211.0
GANet [39]10.111.720.314.1
DSMNet [40]6.26.521.86.2
FC-DSM [41]5.56.212.06.0
CFNet [30]4.75.815.35.8
GF-PSMNet [16]5.34.610.96.2
Mask-CFNet4.85.813.75.7
Mask-LacGwcNet5.75.616.95.3
", + "image_path": "fcfc44f5ea15168d5ac49f31d0b50b492367fd24084c3a843b9e088d30dc8161.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 527, + 545, + 561 + ], + "lines": [ + { + "bbox": [ + 305, + 527, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 545, + 561 + ], + "type": "text", + "content": "Table 4. Cross-domain generalization evaluation (peak results) on four target datasets. All methods are only trained on the Scene-Flow dataset and tested on training images of four real datasets." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": "ing can promote high-level representation feature learning. Note that all methods have volatility among training epochs, and we list the volatility as shown in the supplemental material. It prompted us to suggest that the scope of fluctuation should be attached to future work in stereo matching." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": "Comparisons with fine-tuning. We fine-tune the models (CFNet and LacGwcNet) on the KITTI datasets to test the performance after fine-tuning. Tab. 5 shows that the results are similar to the trend of the SceneFlow dataset. When the masking ratio is low (< 25%), our method does not affect the performance and produces nearly the same re" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5441" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 95, + 274, + 253 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "type": "text", + "content": "sults. As the ratio rises, the performance gradually declines." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 95, + 274, + 253 + ], + "lines": [ + { + "bbox": [ + 61, + 95, + 274, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 95, + 274, + 253 + ], + "type": "table", + "html": "
MethodRatioKT-12 (Out-Noc)KT-15 (D1-all)
LacGwcNet [15]01.131.77
LacGwcNet [15]0.151.151.78
LacGwcNet [15]0.251.161.77
LacGwcNet [15]0.351.271.95
LacGwcNet [15]0.451.392.21
CFNet [30]01.231.88
CFNet [30]0.151.231.89
CFNet [30]0.251.271.91
CFNet [30]0.351.362.05
CFNet [30]0.451.482.28
", + "image_path": "6ea6fd0b18f612cce2f460fc5bcb46ee0e36d602c12d6801284deb84acb61304.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 260, + 286, + 282 + ], + "lines": [ + { + "bbox": [ + 47, + 260, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 286, + 282 + ], + "type": "text", + "content": "Table 5. The fine-tuning results on the KITTI dataset. It indicates the results are similar to the trend of the SceneFlow dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 308, + 116, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 116, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 116, + 320 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 330, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 286, + 376 + ], + "type": "text", + "content": "This paper has proposed a simple approach to alleviate the significant volatility of generalization performance. This section will discuss two topics: (1) When does our approach fail? (2) What is the real unseen domain?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 378, + 286, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 286, + 521 + ], + "type": "text", + "content": "When does our approach fail? (1) In our study, we have added the channels of the feature extraction module to get a better reconstruction image. However, the generalization performance did not obtain remarkable improvement. Although larger models can handle two tasks well, it is not conducive to the fusion of multi-task learning because larger modules may learn features of two tasks separately. Similarly, it will weaken the effect of masked representation. (2) Excessive masking will cause the learnable matching area to shrink in the feature matching module, affecting matching accuracy or generalization. Thus, large-size masking is not recommended to improve generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 286, + 713 + ], + "type": "text", + "content": "What is the real unseen domain? In previous studies, nearly all papers used KITTI, ETH 3D, and Middlebury datasets as the unseen domains. However, can these datasets be represented the unseen domain? Although the color distribution, illumination, and authenticity are different between the source domain and the target domains, similar objects are included in the source domain, such as the street scenes in the Driving dataset (sub-dataset of SceneFlow), pedestrian in the Driving dataset, daily supplies in FlyingThing3D (sub-dataset of SceneFlow), etc. Herefore, we can get good results without fine-tuning models on the target dataset. Does it mean we can use these algorithms in practice? The answer is no. We use these models that work well on target datasets to deal with daily images collected by the zed camera or remote sensing images collected by satellite. As shown in Fig. 8, we can not get the expected results" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 420, + 179 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 420, + 179 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 420, + 179 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 420, + 179 + ], + "type": "image", + "image_path": "29f260035131d743d16fce9ec6dc1cf4902f88ca4fd2a435bbdb82a2f3a3912a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 193, + 419, + 255 + ], + "blocks": [ + { + "bbox": [ + 318, + 182, + 406, + 192 + ], + "lines": [ + { + "bbox": [ + 318, + 182, + 406, + 192 + ], + "spans": [ + { + "bbox": [ + 318, + 182, + 406, + 192 + ], + "type": "text", + "content": "(a) remote sensing image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 193, + 419, + 255 + ], + "lines": [ + { + "bbox": [ + 310, + 193, + 419, + 255 + ], + "spans": [ + { + "bbox": [ + 310, + 193, + 419, + 255 + ], + "type": "image", + "image_path": "519ce6ec7e811d3f847ce80516a5fe7cfe8c0313c2fb142999d6b329f706cc22.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 256, + 391, + 266 + ], + "lines": [ + { + "bbox": [ + 337, + 256, + 391, + 266 + ], + "spans": [ + { + "bbox": [ + 337, + 256, + 391, + 266 + ], + "type": "text", + "content": "(c) daily image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 280, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 306, + 280, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 545, + 314 + ], + "type": "text", + "content": "Figure 8. The examples of failures in the real unseen domain. When elements do not appear in the source dataset, the models can not predict the expected results." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 71, + 541, + 179 + ], + "blocks": [ + { + "bbox": [ + 432, + 71, + 541, + 179 + ], + "lines": [ + { + "bbox": [ + 432, + 71, + 541, + 179 + ], + "spans": [ + { + "bbox": [ + 432, + 71, + 541, + 179 + ], + "type": "image", + "image_path": "ba832e6837f33195faf029f0326bb3ddd795577129db0ee1d88b1773d274e78f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 432, + 193, + 541, + 255 + ], + "blocks": [ + { + "bbox": [ + 457, + 182, + 517, + 192 + ], + "lines": [ + { + "bbox": [ + 457, + 182, + 517, + 192 + ], + "spans": [ + { + "bbox": [ + 457, + 182, + 517, + 192 + ], + "type": "text", + "content": "(b) disparity map" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 193, + 541, + 255 + ], + "lines": [ + { + "bbox": [ + 432, + 193, + 541, + 255 + ], + "spans": [ + { + "bbox": [ + 432, + 193, + 541, + 255 + ], + "type": "image", + "image_path": "a525c3383e040ae9a5a4aa3ed0027ff51c270627c3162023e1138e78bef3357c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 256, + 516, + 266 + ], + "lines": [ + { + "bbox": [ + 456, + 256, + 516, + 266 + ], + "spans": [ + { + "bbox": [ + 456, + 256, + 516, + 266 + ], + "type": "text", + "content": "(d) disparity map" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 338, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 434 + ], + "type": "text", + "content": "for these images. We consider that the main reason is that these elements or scenes do not appear in the source dataset, such as the human body, buildings under the top view angle, etc. When the real unseen domain appears in the images, matching methods based on deep learning can not predict disparity correctly. Thus, we doubt the cross-domain performance of these datasets (KITTI, ETH 3D, and Middlebury) can represent actual cross-domain performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 450, + 378, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 378, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 378, + 463 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 472, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 628 + ], + "type": "text", + "content": "In this paper, we have proposed a simple masked representation method to address the problem of unstable generalization performance among different training epochs for stereo matching. Our approach is inspired by masked representation learning and multi-task learning to construct a pseudo-multi-task learning framework, which helps models learn better features in the feature extraction module to improve generalization performance and stability. Extensive experiments have proved the effectiveness of our method. Meanwhile, it also demonstrated that the current evaluation manner is unsuitable for measuring generalization performance. In the last, we discussed the failure of our approach and doubted the current definition of the unseen domain." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 631, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 631, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 631, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This research was supported in part by Jiangxi Natural Science Foundations (20224BAB212009, 20212BBE53017), National Natural Science Foundation of China (61866027, 62271410, and 62001396), the Foundation of Key Laboratory of Jiangxi Province for Image Processing and Pattern Recognition (ET202204304), and the Doctor's Foundation of Nanchang Hangkong University (EA202204228, EA202204257). We thank the anonymous area chairs and reviewers for their helpful feedback." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5442" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Changjiang Cai, Matteo Poggi, Stefano Mattoccia, and Filippos Mordohai. Matching-space stereo networks for cross-domain generalization. In International Conference on 3D Vision (3DV), pages 364-373, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Rohan Chabra, Julian Straub, Christopher Sweeney, Richard Newcombe, and Henry Fuchs. Stereodnet: Dilated residual stereonet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786-11795, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "text", + "content": "[3] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5410-5418, 2018. 1, 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "type": "text", + "content": "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proceedings of International Conference on Machine Learning (ICML), pages 1691-1703, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "text", + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 385 + ], + "type": "text", + "content": "[6] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In arXiv preprint, 2022. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "type": "text", + "content": "[7] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "type": "text", + "content": "[8] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3273-3282, 2019. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 499, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 287, + 554 + ], + "type": "text", + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 621 + ], + "type": "text", + "content": "[10] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 66-75, 2017. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[11] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16263-16272, 2022. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Xing Li, Yangyu Fan, Zhibo Rao, Zhe Guo, and Guoyun Lv. Improving stereo matching generalization via fourier-based" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "amplitude transform. IEEE Signal Processing Letters (SPL), 2022. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[13] Xing Li, Yangyu Fan, Zhibo Rao, Guoyun Lv, and Shiya Liu. Synthetic-to-real domain adaptation joint spatial feature transform for stereo matching. IEEE Signal Processing Letters (SPL), 29:60-64, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "type": "text", + "content": "[14] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6197-6206, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[15] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 36, pages 1647-1655, 2022. 1, 5, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "text", + "content": "[16] Biyang Liu, Huimin Yu, and Guodong Qi. Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13012-13021, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 381 + ], + "type": "text", + "content": "[17] Rui Liu, Chengxi Yang, Wenxiu Sun, Xiaogang Wang, and Hongsheng Li. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12757-12766, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[18] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1871-1880, 2019. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "text", + "content": "[19] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[20] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3061-3070, 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "text", + "content": "[21] Guang-Yu Nie, Ming-Ming Cheng, Yun Liu, Zhengfa Liang, Deng-Ping Fan, Yue Liu, and Yongtian Wang. Multi-level context ultra-aggregation for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3283-3291, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 593, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 646 + ], + "type": "text", + "content": "[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536-2544, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[23] Zhibo Rao, Yuchao Dai, Zhelun Shen, and Renjie He. Rethinking training strategy in stereo matching. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), pages 1-14, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[24] Zhibo Rao, Mingyi He, Yuchao Dai, and Zhelun Shen. Patch attention network with generative adversarial model" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5443" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "for semi-supervised binocular disparity prediction. The Visual Computer, pages 1-17, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[25] Zhibo Rao, Mingyi He, Yuchao Dai, and Zhelun Shen. Sliding space-disparity transformer for stereo matching. Neural Computing and Applications (NCAA), pages 1-14, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "type": "text", + "content": "[26] Zhibo Rao, Mingyi He, Yuchao Dai, Zhidong Zhu, Bo Li, and Renjie He. NLca-net: a non-local context attention network for stereo matching. APSIPA Transactions on Signal and Information Processing, 9, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "text", + "content": "[27] Zhibo Rao, Mingyi He, Zhidong Zhu, Yuchao Dai, and Renjie He. Bidirectional guided attention network for 3-d semantic detection of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(7):6138-6153, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 288, + 286 + ], + "type": "text", + "content": "[28] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nešić, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In German conference on pattern recognition, pages 31–42, 2014. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 353 + ], + "type": "text", + "content": "[29] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3260-3269, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 409 + ], + "type": "text", + "content": "[30] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13906-13915, 2021. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "type": "text", + "content": "[31] Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. Adastereo: a simple and efficient approach for adaptive stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10328-10337, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "type": "text", + "content": "[32] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9661-9670, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 577 + ], + "type": "text", + "content": "[33] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 195-204, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[34] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of International Conference on Machine Learning (ICML), pages 1096-1103, 2008. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[35] Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12497-12506, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[36] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 430 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12981-12990, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[37] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1959-1968, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[38] Jure Zbontar and Yann LeCun. Computing the stereo matching cost with a convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1592-1599, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 185-194, 2019. 1, 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "text", + "content": "[40] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In European Conference on Computer Vision (ECCV), pages 420-439, 2020. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 373 + ], + "type": "text", + "content": "[41] Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13001-13011, 2022. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 375, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 547, + 430 + ], + "type": "text", + "content": "[42] Youmin Zhang, Yimin Chen, Xiao Bai, Suihanjin Yu, Kun Yu, Zhiwei Li, and Kuiyuan Yang. Adaptive unimodal cost volume filtering for deep stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 12926-12934, 2020. 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5444" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_content_list.json b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5de7fbc08cf24a623cd3ce26859e4d8f3b51fcd2 --- /dev/null +++ b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_content_list.json @@ -0,0 +1,1594 @@ +[ + { + "type": "text", + "text": "Masked Scene Contrast: A Scalable Framework for Unsupervised 3D Representation Learning", + "text_level": 1, + "bbox": [ + 223, + 130, + 748, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyang Wu Xin Wen Xihui Liu Hengshuang Zhao* The University of Hong Kong", + "bbox": [ + 248, + 203, + 727, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/Pointcept/Pointcept", + "bbox": [ + 312, + 241, + 651, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a pioneering work, PointContrast conducts unsupervised 3D representation learning via leveraging contrastive learning over raw RGB-D frames and proves its effectiveness on various downstream tasks. However, the trend of large-scale unsupervised learning in 3D has yet to emerge due to two stumbling blocks: the inefficiency of matching RGB-D frames as contrastive views and the annoying mode collapse phenomenon mentioned in previous works. Turning the two stumbling blocks into empirical stepping stones, we first propose an efficient and effective contrastive learning framework, which generates contrastive views directly on scene-level point clouds by a well-curated data augmentation pipeline and a practical view mixing strategy. Second, we introduce reconstructive learning on the contrastive learning framework with an exquisite design of contrastive cross masks, which targets the reconstruction of point color and surfel normal. Our Masked Scene Contrast (MSC) framework is capable of extracting comprehensive 3D representations more efficiently and effectively. It accelerates the pre-training procedure by at least $3 \\times$ and still achieves an uncompromised performance compared with previous work. Besides, MSC also enables large-scale 3D pre-training across multiple datasets, which further boosts the performance and achieves state-of-the-art fine-tuning results on several downstream tasks, e.g., $75.5\\%$ mIoU on ScanNet semantic segmentation validation set.", + "bbox": [ + 75, + 323, + 473, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 744, + 209, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised visual representation learning aims at learning visual representations from vast amounts of unlabeled data. The learned representations are proved to be beneficial for various downstream tasks like segmentation and detection. It has attracted lots of attention and achieved remarkable progress in 2D image understanding, exceeding the upper bound of human supervision [20, 25].", + "bbox": [ + 75, + 771, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5bca18c0c15d934a8a1c822d7d6380fefc53181b242f6aab28687fc5e35ebf9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 291, + 861, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b92e0692df24dfe8bc3140d03f792bcae10feb36a78021125d206a6261570a36.jpg", + "image_caption": [], + "image_footnote": [ + "RGB-D Camera Random Crop Visible Region", + "Figure 1. Comparison of unsupervised 3D representation learning. The previous method [56] (top) relies on raw RGB-D frames with restricted views for contrastive learning, resulting in low efficiency and inferior versatility. Our approach (bottom) directly operates on scene-level views with contrastive learning and masked point modeling, leading to high efficiency and superior generality, further enabling large-scale pre-training across multiple datasets." + ], + "bbox": [ + 524, + 407, + 861, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the impressive success of unsupervised visual representation learning in 2D, it is underexplored in 3D. Modern 3D scene understanding algorithms [10, 52] are focused on supervised learning, where models are trained directly from scratch on targeted datasets and tasks. Well-pretrained visual representations can undoubtedly boost the performance of these algorithms and are currently in urgent demand. Recent work PointContrast [56] conducts a preliminary exploration in 3D unsupervised learning. However, it is limited to raw RGB-D frames with an inefficient learning paradigm, which is not scalable and applicable to large-scale unsupervised learning. To address this essential and inevitable challenge, we focus on building a scalable framework for large-scale 3D unsupervised learning.", + "bbox": [ + 496, + 656, + 892, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One technical stumbling block towards large-scale pretraining is the inefficient learning strategy introduced by", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author. Email: hszhao@cs.hku.hk", + "bbox": [ + 94, + 886, + 410, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9415", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "matching RGB-D frames as contrastive views. PointContrast [56] opens the door to pre-training on real indoor scene datasets and proposes frame matching to generate contrastive views with natural camera views, as in Figure 1 top. However, frame matching is inefficient since duplicated encoding exists for matched frames, resulting in limited scene diversity in batch training and optimization. Meanwhile, not all of the 3D scene data contains raw RGB-D frames, leading to failure deployments of the algorithm. Inspired by the great success of SimCLR [7], we investigate generating strong contrastive views by directly applying a series of well-curated data augmentations to scene-level point clouds, eliminating the dependence on raw RGB-D frames, as in Figure 1 bottom. Combined with an effective mechanism that mixes up query views, our contrastive learning design accelerates the pre-training procedure by $4.4 \\times$ and achieves superior performance with purely point cloud data, compared to PointContast with raw data. The superior design also enables large-scale pre-training across multiple datasets like ScanNet [16] and ArkitScenes [5].", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another obstacle is the mode collapse phenomenon that occurs when scaling up the optimization iterations. We owe the culprit for this circumstance to the insufficient difficulty of unsupervised learning tasks. To further tackle the mode collapse challenge in unsupervised learning and scale up the optimization iterations, inspired by recent masked autoencoders [23, 57], we construct a masked point modeling paradigm where both point color reconstruction objective and surfel normal reconstruction objective are proposed to recover the masked color and geometric information of the point cloud respectively. We incorporate the mask point modeling strategy into our contrastive learning framework via an exquisite design of contrastive cross masks, leading towards a scalable unsupervised 3D representation learning framework, namely Masked Scene Contrast (MSC).", + "bbox": [ + 76, + 393, + 472, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework is efficient, effective, and scalable. We conduct extensive experimental evaluations to validate its capability. On the popular point cloud dataset ScanNet, our algorithm accelerates the pre-training procedure by more than $3 \\times$ , and achieves better performance on downstream tasks, when compared to the previous representative Point-Contrast. Besides, our method also enables large-scale 3D pre-training across multiple datasets, leading to state-of-the-art fine-tuning results on several downstream tasks, e.g. $75.5\\%$ mIoU on ScanNet semantic segmentation validation set. In conclusion, our work opens up new possibilities for large-scale unsupervised 3D representation learning.", + "bbox": [ + 76, + 619, + 472, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 814, + 218, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Image contrastive learning. Based on the instance discrimination [17] pretext task, and combined with the contrastive learning [27, 47] paradigm, modern variants of 2D image contrastive representation learning have shown", + "bbox": [ + 76, + 839, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "strong abilities in learning transferable visual representations [7, 24, 54]. With the learning objective built on the similarity between randomly augmented image views, this line of work strongly depends on a large batch size [7, 24] and a finely designed data augmentation pipeline [7, 20, 46] to achieve better performance. We find these two points also hold true for 3D contrastive learning.", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Image reconstructive learning. In 2D unsupervised learning, there is also a recent trend of switching the pretext task from instance discrimination [4,6,7,20,24] to masked image modeling [3,23,50,57,60]. Based on a denoising autoencoder [48]-style architecture, the task is to reconstruct the RGB value [23,57], discrete token [3,60], or feature [50] of masked pixels. This line of work has shown strong potential in learning representations on large-scale datasets and is less prone to model collapsing like instance-discrimination-based methods (e.g., contrastive learning). When combined with contrastive learning [2,18,53], the performance can be further boosted, yet with less dependence on data scale [18].", + "bbox": [ + 496, + 214, + 893, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Scene understanding. The deep neural architectures for understanding 3D scenes can be roughly categorized into three paradigms according to the way they model point clouds: projection-based, voxel-based, and point-based. Projection-based works project 3D points into various image planes and adopt 2D CNN-based backbones to extract features [9, 30, 31, 43]. On the other hand, the voxel-based stream transforms point clouds into regular voxel representations to operate 3D convolutions [33, 42]. Their efficiency is then improved thanks to sparse convolution [10, 13, 19]. Point-based methods, in contrast, directly operate on the point cloud [12, 37, 38, 44, 58], and see a recent transition towards transformer-based architectures [21, 52, 59]. Following [56], we mainly pre-train on the voxel-based method SparseUNet [10] implemented with SpConv [15].", + "bbox": [ + 496, + 414, + 893, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Representation learning. Unlike the 2D counterparts, where large-scale unsupervised pre-training has been a common choice for facilitating downstream tasks [6], 3D representation learning is still not mature, and most works still train from scratch on the target data directly [28]. While earlier works in 3D representation learning simply build on a single object [22, 40, 41, 49], recent works start to train on scene-centric point clouds [28, 56]. However, unlike in 2D that scene-centric representation learning has been well-studied [32, 51, 55], the pre-training on 3D scenes, which relies on raw frame data [28, 56], still faces inefficiency issues and finds it hard to scale up to larger scale datasets. In contrast, we explore directly learning at the scene level, which shows significantly higher efficiency in processing scene data, and opens the possibility for pre-training with larger-scale point clouds, for the first time ever.", + "bbox": [ + 496, + 657, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9416", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93f1b3604b750be5300d4a25bc343024ff70ac639363956ca61d69ab21ed4007.jpg", + "image_caption": [ + "(a) Frame matching [28, 56]. 1. Extract raw RGB-D frames and camera positions from raw data. 2. Project each 2D frame into 3D space produces frame-level point cloud views. 3. Calculate pairwise overlapping rates among each frame of a single view and select pairs with overlapping rates larger than $30\\%$ as pairs of contrastive views." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 478, + 223 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a4c6dd5d370f437fa7b9a46213ca0038cfed07102698112ac2cceb9b743195e.jpg", + "image_caption": [ + "(b) Scene augmentation (ours). 1. Apply spatial augmentations containing rotation, flipping, and scaling. 2. Apply photometric augmentations containing brightness, contrast, saturation, hue, and gaussian noise jittering. 3. Generate and apply contrastive cross masks to the two views after sampling augmentations containing cropping and voxelization.", + "Figure 2. View generation. Compared with frame matching (FM), our scene augmentation (SA) is efficient and effective. 1. SA can end-to-end produce contrastive views on the original point cloud with ignorable latency, while FM requires preprocessing devouring enormous storage resources (e.g. additional 1.5 TB storage for ScanNet) in step 2 and pairwise matching is time-consuming. 2. SA produces scene-level views, while FM can only produce frame-level views containing limited information. Benefiting from advanced photometric augmentations, SA has the capacity to simulate the same scene under different lighting." + ], + "image_footnote": [], + "bbox": [ + 496, + 89, + 887, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Pilot Study", + "text_level": 1, + "bbox": [ + 76, + 382, + 194, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section analyzes the two main obstacles towards large-scale pre-training with point clouds. Our proposed design is based on the conclusion of the pilot study.", + "bbox": [ + 75, + 407, + 470, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Is matching RGB-D frames a good choice?", + "text_level": 1, + "bbox": [ + 76, + 453, + 377, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a seminal work in 3D representation learning, Point-Contrast [56] first enables pre-training in real-world indoor scenes with matched raw RGB-D frames as contrastive views. A visualization of the frame matching procedure is illustrated in Figure 2a. This protocol seems natural for indoor scenes since point clouds of indoor scenes are usually derived from RGB-D videos [1,5,16], where the raw frames are extracted from. However, this framework has multiple drawbacks that can hinder the scalability of training:", + "bbox": [ + 75, + 468, + 468, + 604 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Redundant frame encoding. The pairwise matching strategy that PointContrast adopts allows one frame to be matched multiple times. As a result, each frame can be encoded multiple times in one step, adding to redundancy in training.", + "- Low learning efficiency. In one training step, the frame matching strategy only allows the framework to process several views of a single scene. Therefore the amount of information that PointContrast can process in one step is rather limited, and the overall time for one training cycle is also notably high.", + "- Dependency on raw RGB-D frames. The whole framework is built on the assumption that RGB-D videos are available, yet this is not true for every publicly-available point cloud dataset. Even when available, the storage cost of RGBD frames is also significantly higher than the reconstructed point cloud data." + ], + "bbox": [ + 94, + 606, + 468, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consequently, pre-training frameworks [28, 56] based on matching frames as contrastive samples require enormous", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "computing and storage resources. For example, Point-Contrast sub-samples RGB-D scans from the raw ScanNet videos every 25 frames, consuming $\\sim 30$ times the storage of the processed point clouds, and takes 80 GPU hours to process an epoch of 1500 scenes. Even at such cost, our experiments in Table 1a verify that the raw RGB-D frames cannot bring additional information over the processed point clouds to achieve a better representation.", + "bbox": [ + 496, + 383, + 890, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will explore the possibility of pre-training on the point clouds directly.", + "bbox": [ + 496, + 515, + 890, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "What's the revelation behind mode collapse?", + "text_level": 1, + "bbox": [ + 498, + 556, + 815, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mode collapse, defined as the phenomenon that all features collapse to a single vector, remains an unsolved problem accompanying the development of 3D representation learning [11, 56]. To alleviate this problem, PointContrast introduced InfoNCE loss [35], which has been shown to stabilize training, to replace the hardest-contrastive loss. Yet, the problem of mode collapse can still occur when the amount of training data and the length of the training schedule increase. Given the empirical conclusion of 2D contrastive learning [8], the occurrence of mode collapse is unusual under the premise that a large number of negative samples are already adopted. Interestingly, we notice that the mean negative pair cosine similarity of previous works is mostly close to 0, indicating that the negative samples are mostly easy and thus have little penalty towards the trivial solution. Although the InfoNCE loss alleviates this problem with an alternated optimization objective, we argue that a more desirable solution can be achieved by raising the difficulty of the unsupervised pretext task.", + "bbox": [ + 496, + 571, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will further raise the difficulty of the pretext task to solve the mode collapse problem.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9417", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e44cb99999a7badb233eeb7577b714c00c3aa521f877ed13ede5835b7ec4bd84.jpg", + "image_caption": [ + "Figure 3. Our MSC framework. (1) Generating a pair of contrastive views with a well-curated data augmentation pipeline consisting of photometric, spatial, and sampling augmentations. (2) Generating a pair of complementary masks and applying them to the pair of contrastive views. Replacing masked point features with a learnable mask token vector. (3) Extracting point representation with a given U-Net style backbone for point cloud understanding. (4) Reassembling masked contrastive views to masked points combination and unmasked points combination. (5) Matching points share similar positional relationships in the two views as positive sample pairs and computing InfoNCE loss to optimize contrastive objective. (6) Predicting masked point color and normal and computing Mean Squared Error loss and Cosine Similarity Loss with ground truth respectively, to optimize the reconstruction objective." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 877, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 76, + 465, + 187, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the analysis of the stumbling blocks for large-scale pre-training in Section 3, we first introduce our optimized contrastive learning design in Section 4.1 to make the process more efficient. Then we solve the long-term problem of mode collapse with an additional reconstructive learning design in Section 4.2. The final optimization target is described in Section 4.3. Combining these exquisite designs, we build the whole framework, namely Masked Scene Contrast (MSC), and a visual illustration of our MSC is available in Figure 3.", + "bbox": [ + 75, + 492, + 472, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 654, + 279, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Framework.** Different from the previous protocol of matching RGB-D frames decomposed from indoor scenes, our contrastive learning framework directly operates on the point cloud data. Given a point cloud $\\mathbf{X} = (P, C)$ , where $P \\in \\mathbb{R}^{n \\times 3}$ represents the spatial features (coordinate) of the points and $C \\in \\mathbb{R}^{n \\times 3}$ represents the photometric features (color) of the points, the contrastive learning framework can be summarized as follows:", + "bbox": [ + 75, + 676, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- View generation. For a given point cloud $\\mathbf{X}$ , we generate query view $\\mathbf{X}_r$ and key view $\\mathbf{X}_k$ of the original point cloud with a sequence of stochastic data augmentations, which includes photometric, spatial, and sampling augmentations.", + "- Feature extraction. Encoding point cloud features $\\mathbf{F}_r$" + ], + "bbox": [ + 94, + 805, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\pmb{F}_k$ with a U-Net style backbone $\\zeta(\\cdot)$ to $\\hat{\\pmb{F}}_r$ and $\\hat{\\pmb{F}}_k$ respectively.", + "bbox": [ + 529, + 465, + 888, + 497 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Point matching. The positive samples of contrastive learning are point pairs with close spatial positions in the two views. For each point belonging to the query view, we calculate the correspondence mapping $\\mathcal{P} = \\{(i,j)\\}_{n'}$ to points of the key view. If $(i,j) \\in \\mathcal{P}$ then point $(\\pmb{p}_i,\\pmb{c}_i)$ and point $(\\pmb{p}_j,\\pmb{c}_j)$ constructs a pair across two views.", + "- Loss computation. Computing the contrastive learning loss on the representation of two views $\\hat{F}_r$ and $\\hat{F}_k$ and the correspondence mapping $\\mathcal{P}$ . An encoded query view should be similar to its key view." + ], + "bbox": [ + 517, + 505, + 893, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data augmentation. As a pioneering work in image contrastive learning, SimCLR [7] reveals that a well-curated data augmentation pipeline is crucial for learning strong representations. Unlike supervised learning, contrastive learning requires much stronger data augmentations to prevent trivial solutions. However, an effective data augmentation recipe is still absent in 3D representation learning. The frame matching scheme in prior works [28,56] simply applies a randomly rotating operator to contrastive targets. Even if the RGB-D frames can be viewed as a natural random crop, the augmentation space is still far from diverse enough. The pretext task it forms is not yet challenging enough to facilitate the contrastive learning framework to learn robust representations for downstream tasks.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9418", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4edc151802c93bc6091ae903bffbc9a0be3f2d3db9856aa4240262f90de67822.jpg", + "image_caption": [ + "Figure 4. View Mixing. Randomly mix up query views while keeping key views unmixed for a given batch of pairwise contrastive views. Detaching mixed query view after feature extraction for contrastive comparison with matched key views." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As presented in Figure 2b, our well-designed stochastic data augmentation pipeline includes photometric augmentations, spatial augmentations, and sampling augmentations. Inspired by the advanced photometric augmentation validated by our 2D counterparts [7, 20], we further strengthen the photometric augmentation component introduced by Choy et al. [10] with random brightness, contrast, saturation, hue, and gaussian noise jittering for photometric augmentation. Besides that, random rotating, flipping, and scaling constitute our spatial augmentations, and the sampling augmentation is composed of random cropping and grid sampling.", + "bbox": [ + 75, + 354, + 468, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Empirically, the order of data augmentations is also a key component of our recipe. For example, grid sampling after random rotation leads to cross grids for sampling, which further increases the distinction between contrastive views and has a better augmentation effect. The specific data augmentation settings are available in the appendix, and a comparison with previous methods is presented in Figure 2.", + "bbox": [ + 75, + 540, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "View mixing. Recently, Nekrasov et al. [34] proposes a data augmentation technique for 3D understanding models by mixing two scenes as a hybrid training sample, which can significantly suppress model overfitting. Inspired by the mixing mechanism, we integrate the logic of mixing as part of the contrastive learning objective. As illustrated in Figure 4, for a batch of pairwise views, we randomly mix up the query views while maintaining the key views unchanged before the feature extraction process. The simple operation can effectively increase the robustness of the backbone and improve the robustness of the point cloud representation.", + "bbox": [ + 75, + 652, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive target. We follow the design of PointContrast on the contrast target and apply InfoNCE loss to the matched points. Given correspondence mapping $\\mathcal{P} = \\{(i,j)\\}_{n'}$ produced by point matching and points representation $\\hat{\\pmb{F}}_r$ and $\\hat{\\pmb{F}}_k$ embedded during feature extraction, the", + "bbox": [ + 76, + 823, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "contrastive loss is:", + "bbox": [ + 500, + 92, + 624, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} = \\frac {\\boldsymbol {f} _ {i ^ {\\prime}} ^ {r T} \\boldsymbol {f} _ {j ^ {\\prime}} ^ {k}}{| | \\boldsymbol {f} _ {i ^ {\\prime}} ^ {r T} | | \\cdot | | \\boldsymbol {f} _ {j ^ {\\prime}} ^ {k} | |}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 108, + 890, + 147 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {I n f o N C E}} = \\sum_ {i} ^ {n} - \\log \\frac {\\exp \\left(s _ {i i} / \\tau\\right)}{\\sum_ {j} ^ {n} \\exp \\left(s _ {i j} / \\tau\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 150, + 890, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "note that $S = \\{s_{ij}\\} \\in \\mathbb{R}^{n\\times n}$ is the pairwise cosine similarity matrix between positive samples and negative samples, while $\\tau$ is the temperature factor scaling cosine similarity. In practice, we control temperature factor $\\tau$ as 0.4, which is the same as previous works [28,56].", + "bbox": [ + 496, + 194, + 890, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Reconstructive Learning", + "text_level": 1, + "bbox": [ + 500, + 277, + 725, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As is mentioned in Section 3, one of the stumbling blocks for large-scale representation is mode collapse, and our solution is to scale up the difficulty of the unsupervised pre-training task. Motivated by the success of masked image modeling [23, 57] in 2D representations, we propose masked point modeling, which can be naturally integrated into our contrastive learning framework. Benefiting from this design, our framework can fully use non-overlapped regions of contrastive views that cannot be utilized by contrastive learning.", + "bbox": [ + 496, + 300, + 890, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive cross mask. The key design that enables additional construction learning in our contrastive learning framework is the contrastive cross mask. For a given query view and key view of a single point cloud, we partite the unioned point set into non-overlapping grid partitions by their original position before spatial augmentation. Given a mask rate $r$ range from 0 to 0.5, we randomly generate a pair of masks $M_r, M_k \\in \\mathbb{R}^{1 \\times n_{r,k}}$ , in which there are no shared masked patches. Then, we follow the practice of SimMIM [57] to apply the pair of masks to the two views respectively by replacing the input feature with a learnable mask token vector $t \\in \\mathbb{R}^c$ . Consequently, the feature extraction process can be rewritten as follows:", + "bbox": [ + 496, + 452, + 892, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {F}} _ {r, k} = \\zeta \\left(\\left(1 - \\boldsymbol {M} _ {r, k}\\right) \\boldsymbol {F} _ {r, k} + \\boldsymbol {M} _ {r, k} \\boldsymbol {T} _ {r, k}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 652, + 890, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T_{r,k} \\in \\mathbb{R}^{n_r, n_k \\times c}$ is the expand matrix of mask token vector $t$ to fit the feature dimensions.", + "bbox": [ + 496, + 676, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction target. The features of the point cloud are composed of two parts, the coordinates that determine the geometric structure and the colors that represent the texture features. We build up reconstruction targets for the two groups of features separately.", + "bbox": [ + 496, + 707, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reconstruction of point cloud texture is straightforward, we predict the photometric value of each point with a linear projection. We compute the mean squared error (MSE) between the reconstructed and original color of masked points as the color reconstruction loss:", + "bbox": [ + 496, + 782, + 890, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = \\frac {\\sum_ {i} ^ {n _ {r}} m _ {i} ^ {r} | | \\boldsymbol {x} _ {i} ^ {r} - \\hat {\\boldsymbol {x}} _ {i} ^ {r} | | _ {2} ^ {2} + \\sum_ {i} ^ {n _ {k}} m _ {i} ^ {k} | | \\boldsymbol {x} _ {i} ^ {k} - \\hat {\\boldsymbol {x}} _ {i} ^ {k} | | _ {2} ^ {2}}{n _ {r} ^ {\\prime} + n _ {k} ^ {\\prime}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 862, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9419", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_r'$ and $n_k'$ represent the number of mask points belonging to refer view and key view, $m_r^i$ and $m_k^i$ mean the $i$ -th element of $M_r$ and $M_k$ respectively.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Point coordinates play an important role in describing the geometric structure of point clouds, and it is worth noting that directly reconstructing the coordinates of masked points is not reasonable since masked points are only sampled from 3D object surface rather than the continuous surface itself. Reconstructing points coordination would lead to an overfitted representation. To overcome the challenge, we introduce the concept of surfel reconstruction. Surfel is an abbreviation for a surface element or surface voxel in the discrete topology literature [26] and primitives rendering [36]. For each masked point, we reconstruct the normal vector of the corresponding surfel and compute the mean cosine similarity between estimations and surfel normals as a contrastive loss:", + "bbox": [ + 75, + 137, + 470, + 347 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {n}} = \\frac {\\sum_ {i} ^ {n _ {r}} m _ {i} ^ {r} \\boldsymbol {x} _ {i} ^ {r T} \\hat {\\boldsymbol {x}} _ {i} ^ {r} + \\sum_ {i} ^ {n _ {k}} m _ {i} ^ {k} \\boldsymbol {x} _ {i} ^ {k T} \\hat {\\boldsymbol {x}} _ {i} ^ {k}}{n _ {r} ^ {\\prime} + n _ {k} ^ {\\prime}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 353, + 468, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n_r'$ and $n_k'$ represent the number of mask points belonging to refer view and key view, $m_r^i$ and $m_k^i$ mean the $i$ -th element of $M_r$ and $M_k$ respectively.", + "bbox": [ + 76, + 398, + 470, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 452, + 222, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our framework combines the contrastive target, the color reconstruction target, and the surfel reconstruction to make the unsupervised task more scalable. The overall loss function is a weighted sum of Eq. 2, Eq. 4, and Eq. 5 which is written as follow:", + "bbox": [ + 76, + 474, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\lambda_ {c} \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {n} \\mathcal {L} _ {\\mathrm {n}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 561, + 468, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{c}$ and $\\lambda_{n}$ are the weight parameters that balance the three loss components. Empirically we find that performance is robust to the choice of weight parameters, and we make $\\lambda_{c} = \\lambda_{n} = 1$ in practice.", + "bbox": [ + 76, + 588, + 468, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experimental evaluations to validate the capability of our framework, built upon the point cloud perception codebase Pointcept [14]. We first ablate our designs with an efficient pre-training pipeline that only utilizes ScanNet point cloud in Section 5.1, while without compromising performance. Then we explore large-scale pre-training across multiple datasets and compare our performance with previous results in Section 5.2.", + "bbox": [ + 75, + 686, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Main Properties", + "text_level": 1, + "bbox": [ + 76, + 816, + 238, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We ablate the main designs and intriguing properties of our MSC in Table 1, and the default setting is available in the caption. We enable efficient pre-training by introducing our view generation pipeline, which is ablated in Table 1a.", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd17a2c7fb786d0e34c0355be1dff800a9141f21a1531f19c287ea93361b948f.jpg", + "image_caption": [ + "Figure 5. Masked scenes and color reconstructions. We visualize one of the cross masks of each scene with a mask rate of $50\\%$ (left) and color reconstruction of the masked point combinations (right). We pre-train our MSC with a mask patch size of $10\\mathrm{cm}$ and generalize to results to different mask sizes. Compared with the original point clouds, the loss of detail cannot be avoided, while boundary and texture are well preserved by our model." + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 888, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All of our ablation experiments only require 20G ScanNet point cloud data and around 14 hours pertaining on a single machine containing 8 NVIDIA RTX3090.", + "bbox": [ + 496, + 460, + 890, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "View generation. In Table 1a, we show the results with different generation strategies of contrastive views. Unlike PointContrast [56], which builds on the raw RGB-D frames, our strategy directly utilizes the scene-level point cloud. This significantly reduces storage requirements (96% less) and allows more efficient use of the training data. With a 30% equivalent number of training iterations, our method can attain $120 \\times$ training epochs, making full use of the training data for the first time. This results in 0.4 points higher FT mIoU with $4.4 \\times$ speedup. When combined with an additional mask point modeling strategy, the performance can be further boosted by 0.6 points, yet still with a notable speedup.", + "bbox": [ + 496, + 507, + 892, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Number of positive pairs. In Table 1b, we show the effects of different numbers of positive pairs. Our method sees consistent improvements with an increasing number of positive pairs, while for PointContrast [56], the information in a large batch of positive pairs cannot be effectively utilized. Our intuition is that the ability to process scene-level views, which are larger in scale and contain much more information than frame-level views, enables our method to take advantage of a larger number of positive pairs.", + "bbox": [ + 496, + 704, + 893, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data augmentation. In Table 1c, we analyze the effect of different data augmentation combinations. Adopting either spatial augmentation or photometric augmentation leads to sub-optimal performance, and the combination of", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9420", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/85ccc8006b3f6009a22005288a2f477a94952cf1431453762e93cb6f9e77bfd4.jpg", + "table_caption": [], + "table_footnote": [ + "(a) View generation. Views produced by our enhanced data augmentation are stronger than the original RGB-D frames. Scene-level views can significantly speed up pre-training and make contrastive learning more effective. The performance can be further boosted with additional masked point modeling." + ], + "table_body": "
View generation methodsPre-train dataStorageBatch sizeItersEpochsFT mIoU (%)Hours (h)Speedup
Frame matching (PointContrast [56])ScanNet Raw500G32100k574.0481.0×
Scene augmentation w/o mask (ours)ScanNet v220G3230k60074.4114.4×
Scene augmentation w mask (ours)ScanNet v220G3230k60075.0143.4×
", + "bbox": [ + 99, + 89, + 870, + 152 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/548ba6a456ae41fa0bf5e891fe2b5353f6eae18878c89b5cd14d8a4916f9e765.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Number of positive pairs. A larger amount of sampled positive pairs are necessary for scene-level views." + ], + "table_body": "
#Pos pairsPC [56]MSC (ours)
102473.874.3
204874.074.5
409873.774.9
819273.975.0
", + "bbox": [ + 117, + 184, + 315, + 261 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e7f4fb5b65c6bc4975239c02bd9c4719a810d2d924a61c8376d45e31f57a9024.jpg", + "table_caption": [], + "table_footnote": [ + "(e) Cross mask. Masks containing shared masked patches have a negative influence on contrastive learning." + ], + "table_body": "
MaskTaskFT mIoU (%)
w/o crossw/o contrast74.1
w crossw/o contrast74.4
w/o crossw contrast74.7
w crossw contrast75.0
", + "bbox": [ + 101, + 305, + 333, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b733a3f17ea30a41c535faa4a43d69fb910787de1eace168c4adfa374d8095c.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Data augmentation. The combination of spatial and photometric augmentation makes the view generation pipeline comes to work." + ], + "table_body": "
SpatialPhotometricFT mIoU (%)
w/o augw/o aug72.1
w augw/o aug73.4
w/o augw aug72.8
w augw aug74.4
", + "bbox": [ + 375, + 184, + 596, + 261 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/28d148d3911b458d76e05fbb9755024f1fda780e60d71c72eebc40eb4eff62d6.jpg", + "table_caption": [], + "table_footnote": [ + "(f) Mask grid size. Our design works with mask patches with a grid size larger than $0.1\\mathrm{m}$ , and we consider $0.15\\mathrm{m}$ as a default setting." + ], + "table_body": "
Mask grid size (m)FT mIoU (%)
0.0574.3
0.175.0
0.1575.0
0.274.8
", + "bbox": [ + 385, + 305, + 588, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49f699e03d822bf7f3235449bca4d6612f46deedde691f2b45b48248a2aacdf8.jpg", + "table_caption": [], + "table_footnote": [ + "(d) View mixing. Randomly mixing query views while leaving key views unmixed is a sweet point." + ], + "table_body": "
Query viewKey viewFT mIoU (%)
w/o mixw/o mix74.1
w mixw/o mix74.4
w/o mixw mix74.2
w mixw mix73.7
", + "bbox": [ + 643, + 184, + 870, + 261 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f3dad773b42abe65f78111debc55cf58b79c93115605efe8afa9ec7b14beab9.jpg", + "table_caption": [], + "table_footnote": [ + "(g) Reconstruction target. Both targets have a positive effect, while color reconstruction has a dominant impact on indoor scenes." + ], + "table_body": "
ColorNormalFT mIoU (%)
w/ow/o74.4
ww/o74.9
w/ow74.6
ww75.0
", + "bbox": [ + 663, + 305, + 849, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Ablation experiments. We adopt SparseUNet and efficient pre-training on ScanNet [16] point cloud data to ablate our designs. We report fine-turning (FT) mIoU (%) results on ScanNet 20 classes semantic segmentation as the default metric. If not specified, the default setting is as follows: the pre-training period is 600 epochs, the masking ratio is $30\\%$ and the masked patch has a grid size of $0.15\\mathrm{m}$ in the real-world space, view mixing probability is 0.8. All of our designs are enabled by default. Default settings are marked in gray.", + "bbox": [ + 75, + 426, + 890, + 483 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6cab827345a95d521ba2da5e5633418f8e425c42492928a9b20b20a01886af92.jpg", + "image_caption": [ + "Figure 6. Masking ratio. A masking ratio ranging from $30\\%$ to $40\\%$ works well with our design, and a higher mask rate negatively influences contrastive learning. The y-axes represent ScanNet semantic segmentation validation mIoU $(\\%)$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 503, + 434, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both helps make our view generation pipeline come to work. Concerning relative significance, since the spatial augmentation is highly coupled with the masking strategy, the gain from it is higher than the photometric augmentation. But still, both are necessary.", + "bbox": [ + 75, + 686, + 467, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "View mixing. In Table 1d, we explore different config for the view mixing strategy. Randomly mixing the query views while leaving the key views unchanged yields the best performance. Our intuition is that the key views, which form the vocabulary of the InfoNCE loss, should be relatively stable, thus reducing the ambiguity of the learning target. In contrast, introducing more diversity to the queries can be more helpful.", + "bbox": [ + 75, + 763, + 467, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross mask. In Table 1e, we study the cross mask strategy.", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This strategy ensures that the two augmented views have no overlapping tokens. As reported in the table, whether with or without the contrastive learning target, this strategy ensures fewer shortcuts to the task and enables consistent downstream improvements.", + "bbox": [ + 496, + 500, + 890, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask grid size. In Table 1f, we show the results ablating the grid size for producing masks on point clouds. Our design works with a grid size larger than $0.1\\mathrm{m}$ , and we consider $0.15\\mathrm{m}$ as a default setting. As shown in Figure 5, our reconstruction module is robust to extending mask grid size, which indicates that high guilty representation is captured during pre-training.", + "bbox": [ + 496, + 579, + 890, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Masking ratio. In Figure 6, we depict the effect of the ratio of masked tokens. A masking ratio ranging from $30\\%$ to $40\\%$ works well with our design, and a higher masking rate has a negative impact on the overall performance. This varies from the conclusion of MAE [23], in which a higher masking ratio of $75\\%$ achieves top performance. Our hypothesis is that the contrastive learning objective built on the masked point clouds might favor a lower masking ratio, and the results in Figure 6 reflect a trade-off between the contrastive learning objective and the reconstructive objective. And pure reconstructive learning on point clouds is an interesting direction for future explorations.", + "bbox": [ + 496, + 686, + 890, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction target. In Table 1g, we ablate the effects of two components of our reconstruction target: color re", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9421", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c176e9657cf6fd5bba6dcb954ef85bbc23176a4f132e68ae49202482ad03b50d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsBackbonesSemantic Seg. (mIoU)
SCPC [56]CSC [28]MSC (ours)
ScanNetSparseUNet72.274.1 (+1.9)73.8 (+1.6)75.5 (+3.3)
ScanNet200SparseUNet25.026.2 (+1.2)26.4 (+1.4)28.8 (+3.8)
", + "bbox": [ + 94, + 88, + 455, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/53ce0712230a5a21f9b9ac487cf0b46de045648358901ff15246827caf7e529b.jpg", + "table_caption": [ + "(a) Semantic segmentation. We conduct pre-training on SparseUNet and compare semantic segmentation mIoU $(\\%)$ results on ScanNet and ScanNet200 [39] validation set." + ], + "table_footnote": [], + "table_body": "
DatasetsBackbonesInstance Seg. (mAP@0.5)
SCPC [56]CSC [28]MSC (ours)
ScanNetSparseUNet56.958.0 (+1.1)59.4 (+2.5)59.6 (+2.7)
ScanNet200SparseUNet24.524.9 (+0.4)25.2 (+0.7)26.8 (+2.3)
", + "bbox": [ + 94, + 202, + 457, + 273 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1757c0a1d0c54c311689652155cc0a1a8282a22f0b11694e6ad8e304fbbc8960.jpg", + "table_caption": [ + "(b) Instance segmentation. We conduct pre-training on SparseUNet and compare instance segmentation mAP@0.5 $(\\%)$ results driven by PointGroup [29] on ScanNet and ScanNet200 [39] validation set." + ], + "table_footnote": [], + "table_body": "
LRSemantic Seg.LASemantic Seg.
Pct.SCCSCVIBUSMSCPts.SCCSCVIBUSMSC
100%72.273.8-75.3Full72.273.8-75.3
1%26.028.928.629.22041.955.561.061.2
5%47.849.847.450.75053.960.565.666.8
10%56.759.460.561.010062.265.968.969.7
20%62.964.664.864.920065.568.269.670.7
", + "bbox": [ + 89, + 316, + 460, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) Data efficiency. We follow the ScanNet Data Efficient benchmark and compare the validation results SparseUNet with previous methods.", + "bbox": [ + 83, + 426, + 462, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Results comparison. We adopt cross-dataset pre-training utilizing ScanNet and ArkitScenes point cloud scenes for comparison of downstream task results. The pre-training setting is fixed as the default described in Table 1. More specific pre-training details are available in the Appendix. SC denotes train from scratch.", + "bbox": [ + 75, + 455, + 468, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "construction and normal reconstruction. Given the premise that indoor scenes are used, both targets show a positive effect on overall performance, while color reconstruction has a higher impact. The intuition is that the difference in texture reflected by color has a higher impact on the task of semantic segmentation, while the normal is helpful but has less influence (consider the same task on a 2D image).", + "bbox": [ + 75, + 535, + 468, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Results Comparison", + "text_level": 1, + "bbox": [ + 76, + 650, + 269, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we extend the scale of pre-training by merging multiple datasets and compare downstream task fine-turning performance with previous unsupervised pretraining frameworks [28, 56]. Specifically, we adopt the default model setting ablated in Section 5.1 and train on both ScanNet [16] and ArkitScenes [5] point clouds, extending pre-training assets from 1,513 scenes to 6,560 scenes.", + "bbox": [ + 75, + 674, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic segmentation. In Table 2a, we report the semantic segmentation results on ScanNet and ScanNet200 [39] benchmark with SparseUNet and compare them with previous results. Our improvements are consistent and significant with larger pre-training assets for both benchmarks. With SparseUNet backbone, we outperform the current state-of-art pretraining framework by 1.7 points on ScanNet and 2.4 points on ScanNet200. Meanwhile, it is worth", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "noting that driven by powerful MSC, we set a new best validation result on ScanNet semantic segmentation and pushed the previous SOTA to $75.5\\%$ with a baseline model.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instance segmentation. In Table 2b, we report the instance segmentation results on ScanNet and ScanNet200 [39] with SparseUNet backbone driven by PointGroup [29]. Comparing them with previous results, we still see consistent improvements. Specifically, our framework achieves $59.6\\%$ on the ScanNet validation set, which is 3.3 points higher than training from scratch and 0.2 points promotion compared with the previous state-of-art performance. The boost is more significant on ScanNet200, which is 1.6 points higher than the previous SOTA.", + "bbox": [ + 496, + 136, + 892, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data efficiency. In Table 2c, we compare the ScanNet Data Efficient [28] results with previous methods. Our MSC shows consistently superior performance even compared with the latest data-efficient learning framework [45].", + "bbox": [ + 496, + 287, + 893, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 498, + 364, + 748, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we tackle the problem of scalable unsupervised 3D representation learning. To this end, we present Masked Scene Contrast (MSC), an efficient, effective, and scalable framework that directly operates on scene-level views with contrastive learning and masked point modeling. Benefiting from the efficient scene-level point cloud processing pipeline and the effective training objectives, our method harvests high efficiency and superior generality, and enables large-scale pre-training across multiple datasets.", + "bbox": [ + 496, + 388, + 890, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The key factor that empowers our method's scalability to larger-scale pre-training lies in the efficient pipeline that can directly learn from point cloud data, rather than the raw RGB-D frames. The efficiency, however, does not only mean processing data at scale. When only the standard dataset ScanNet is used for pre-training, our method still achieves uncompromised performance, yet with at least $3 \\times$ speedup over previous works. This is especially meaningful considering the exhaustively long experimental time in the pre-training community, and it can better facilitate the verification of new ideas in future works.", + "bbox": [ + 496, + 526, + 892, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It should also be noted that given the limit in computing resources, and the absence of a pre-training dataset that is sustainably at scale, the scalability of our method is not fully presented. In other words, our method opens the possibility of large-scale pre-training on 3D point cloud data for the first time, and we call for a large-enough 3D scene dataset that can fully unleash this potential. We hope our method can inspire future works that take the first step to real large-scale 3D pre-training, as the 2D community does.", + "bbox": [ + 496, + 691, + 890, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 837, + 668, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China (No. 62201484), HKU Startup Fund, and HKU Seed Fund for Basic Research.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9422", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 3", + "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael G. Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 2", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In ICLR, 2022. 2", + "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 2", + "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurIPS Workshops, 2021. 2, 3, 8", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, 2021. 2", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4, 5", + "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3", + "[9] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 2", + "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 5", + "[11] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 3", + "[12] Ruihang Chu, Yukang Chen, Tao Kong, Lu Qi, and Lei Li. Icm-3d: Instantiated category modeling for 3d instance segmentation. RA-L, 2021. 2", + "[13] Ruihang Chu, Xiaoqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, and Jiaya Jia. Twist: Two-way inter-label self-training for semi-supervised 3d instance segmentation. In CVPR, 2022. 2", + "[14] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.6", + "[15] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.2", + "[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3, 7, 8", + "[17] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with exemplar convolutional neural networks. In TPAMI, 2015. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herve Jegou, and Edouard Grave. Are largescale datasets necessary for self-supervised pre-training? arXiv:2112.10740, 2021. 2", + "[19] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2", + "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020. 1, 2, 5", + "[21] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 2021. 2", + "[22] Kaveh Hassani and Mike Haley. Unsupervised multi-task feature learning on point clouds. In ICCV, 2019. 2", + "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 5, 7", + "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2", + "[25] Olivier J Henaff, Aravind Srinivas, Jeffrey De Fauw, Ali Razavi, Carl Doersch, SM Eslami, and Aaron van den Oord. Data-efficient image recognition with contrastive predictive coding. In ICML, 2020. 1", + "[26] Gabor T Herman. Discrete multidimensional jordan surfaces. CVGIP, 1992. 6", + "[27] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2018. 2", + "[28] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 3, 4, 5, 8", + "[29] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. CVPR, 2020. 8", + "[30] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 2", + "[31] Bo Li, Tianlei Zhang, and Tian Xia. Vehicle detection from 3d lidar using fully convolutional network. In RSS, 2016. 2", + "[32] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv:2011.13677, 2020. 2", + "[33] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015. 2", + "[34] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Augmentation for 3D Scenes. In 3DV, 2021. 5", + "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9423", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In SIGGRAPH, 2000. 6", + "[37] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2", + "[38] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2", + "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 8", + "[40] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In ECCV, 2020. 2", + "[41] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. In NeurIPS, 2019. 2", + "[42] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, 2017. 2", + "[43] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik G. Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. 2", + "[44] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2", + "[45] Beiwen Tian, Liyi Luo, Hao Zhao, and Guyue Zhou. Vibus: Data-efficient 3d scene parsing with viewpoint bottleneck and uncertainty-spectrum modeling. ISPRS, 2022. 8", + "[46] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, 2020. 2", + "[47] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 2", + "[48] Pascal Vincent, H. Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 2010. 2", + "[49] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, 2019. 2", + "[50] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. 2", + "[51] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. In NeurIPS, 2022. 2", + "[52] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1, 2", + "[53] Zhirong Wu, Zihang Lai, Xiao Sun, and Stephen Lin. Extreme masking for learning instance and distributed visual representations. arXiv:2206.04667, 2022. 2" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2", + "[55] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. In NeurIPS, 2021. 2", + "[56] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 1, 2, 3, 4, 5, 6, 7, 8", + "[57] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2, 5", + "[58] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2", + "[59] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2", + "[60] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 402 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9424", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_model.json b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aaff9af6a216730aa2dffbe4de8aec268b3de84e --- /dev/null +++ b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_model.json @@ -0,0 +1,2376 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.131, + 0.749, + 0.177 + ], + "angle": 0, + "content": "Masked Scene Contrast: A Scalable Framework for Unsupervised 3D Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.204, + 0.728, + 0.24 + ], + "angle": 0, + "content": "Xiaoyang Wu Xin Wen Xihui Liu Hengshuang Zhao* The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.242, + 0.653, + 0.257 + ], + "angle": 0, + "content": "https://github.com/Pointcept/Pointcept" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.718 + ], + "angle": 0, + "content": "As a pioneering work, PointContrast conducts unsupervised 3D representation learning via leveraging contrastive learning over raw RGB-D frames and proves its effectiveness on various downstream tasks. However, the trend of large-scale unsupervised learning in 3D has yet to emerge due to two stumbling blocks: the inefficiency of matching RGB-D frames as contrastive views and the annoying mode collapse phenomenon mentioned in previous works. Turning the two stumbling blocks into empirical stepping stones, we first propose an efficient and effective contrastive learning framework, which generates contrastive views directly on scene-level point clouds by a well-curated data augmentation pipeline and a practical view mixing strategy. Second, we introduce reconstructive learning on the contrastive learning framework with an exquisite design of contrastive cross masks, which targets the reconstruction of point color and surfel normal. Our Masked Scene Contrast (MSC) framework is capable of extracting comprehensive 3D representations more efficiently and effectively. It accelerates the pre-training procedure by at least \\(3 \\times\\) and still achieves an uncompromised performance compared with previous work. Besides, MSC also enables large-scale 3D pre-training across multiple datasets, which further boosts the performance and achieves state-of-the-art fine-tuning results on several downstream tasks, e.g., \\(75.5\\%\\) mIoU on ScanNet semantic segmentation validation set." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.746, + 0.21, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Unsupervised visual representation learning aims at learning visual representations from vast amounts of unlabeled data. The learned representations are proved to be beneficial for various downstream tasks like segmentation and detection. It has attracted lots of attention and achieved remarkable progress in 2D image understanding, exceeding the upper bound of human supervision [20, 25]." + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.292, + 0.862, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.409, + 0.862, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.528, + 0.533, + 0.856, + 0.549 + ], + "angle": 0, + "content": "RGB-D Camera Random Crop Visible Region" + }, + { + "type": "image_footnote", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Comparison of unsupervised 3D representation learning. The previous method [56] (top) relies on raw RGB-D frames with restricted views for contrastive learning, resulting in low efficiency and inferior versatility. Our approach (bottom) directly operates on scene-level views with contrastive learning and masked point modeling, leading to high efficiency and superior generality, further enabling large-scale pre-training across multiple datasets." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.893, + 0.869 + ], + "angle": 0, + "content": "Despite the impressive success of unsupervised visual representation learning in 2D, it is underexplored in 3D. Modern 3D scene understanding algorithms [10, 52] are focused on supervised learning, where models are trained directly from scratch on targeted datasets and tasks. Well-pretrained visual representations can undoubtedly boost the performance of these algorithms and are currently in urgent demand. Recent work PointContrast [56] conducts a preliminary exploration in 3D unsupervised learning. However, it is limited to raw RGB-D frames with an inefficient learning paradigm, which is not scalable and applicable to large-scale unsupervised learning. To address this essential and inevitable challenge, we focus on building a scalable framework for large-scale 3D unsupervised learning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "One technical stumbling block towards large-scale pretraining is the inefficient learning strategy introduced by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.411, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author. Email: hszhao@cs.hku.hk" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9415" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.394 + ], + "angle": 0, + "content": "matching RGB-D frames as contrastive views. PointContrast [56] opens the door to pre-training on real indoor scene datasets and proposes frame matching to generate contrastive views with natural camera views, as in Figure 1 top. However, frame matching is inefficient since duplicated encoding exists for matched frames, resulting in limited scene diversity in batch training and optimization. Meanwhile, not all of the 3D scene data contains raw RGB-D frames, leading to failure deployments of the algorithm. Inspired by the great success of SimCLR [7], we investigate generating strong contrastive views by directly applying a series of well-curated data augmentations to scene-level point clouds, eliminating the dependence on raw RGB-D frames, as in Figure 1 bottom. Combined with an effective mechanism that mixes up query views, our contrastive learning design accelerates the pre-training procedure by \\(4.4 \\times\\) and achieves superior performance with purely point cloud data, compared to PointContast with raw data. The superior design also enables large-scale pre-training across multiple datasets like ScanNet [16] and ArkitScenes [5]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.394, + 0.473, + 0.62 + ], + "angle": 0, + "content": "Another obstacle is the mode collapse phenomenon that occurs when scaling up the optimization iterations. We owe the culprit for this circumstance to the insufficient difficulty of unsupervised learning tasks. To further tackle the mode collapse challenge in unsupervised learning and scale up the optimization iterations, inspired by recent masked autoencoders [23, 57], we construct a masked point modeling paradigm where both point color reconstruction objective and surfel normal reconstruction objective are proposed to recover the masked color and geometric information of the point cloud respectively. We incorporate the mask point modeling strategy into our contrastive learning framework via an exquisite design of contrastive cross masks, leading towards a scalable unsupervised 3D representation learning framework, namely Masked Scene Contrast (MSC)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.62, + 0.473, + 0.803 + ], + "angle": 0, + "content": "Our framework is efficient, effective, and scalable. We conduct extensive experimental evaluations to validate its capability. On the popular point cloud dataset ScanNet, our algorithm accelerates the pre-training procedure by more than \\(3 \\times\\), and achieves better performance on downstream tasks, when compared to the previous representative Point-Contrast. Besides, our method also enables large-scale 3D pre-training across multiple datasets, leading to state-of-the-art fine-tuning results on several downstream tasks, e.g. \\(75.5\\%\\) mIoU on ScanNet semantic segmentation validation set. In conclusion, our work opens up new possibilities for large-scale unsupervised 3D representation learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.815, + 0.22, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "2D Image contrastive learning. Based on the instance discrimination [17] pretext task, and combined with the contrastive learning [27, 47] paradigm, modern variants of 2D image contrastive representation learning have shown" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "strong abilities in learning transferable visual representations [7, 24, 54]. With the learning objective built on the similarity between randomly augmented image views, this line of work strongly depends on a large batch size [7, 24] and a finely designed data augmentation pipeline [7, 20, 46] to achieve better performance. We find these two points also hold true for 3D contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.895, + 0.398 + ], + "angle": 0, + "content": "2D Image reconstructive learning. In 2D unsupervised learning, there is also a recent trend of switching the pretext task from instance discrimination [4,6,7,20,24] to masked image modeling [3,23,50,57,60]. Based on a denoising autoencoder [48]-style architecture, the task is to reconstruct the RGB value [23,57], discrete token [3,60], or feature [50] of masked pixels. This line of work has shown strong potential in learning representations on large-scale datasets and is less prone to model collapsing like instance-discrimination-based methods (e.g., contrastive learning). When combined with contrastive learning [2,18,53], the performance can be further boosted, yet with less dependence on data scale [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.644 + ], + "angle": 0, + "content": "3D Scene understanding. The deep neural architectures for understanding 3D scenes can be roughly categorized into three paradigms according to the way they model point clouds: projection-based, voxel-based, and point-based. Projection-based works project 3D points into various image planes and adopt 2D CNN-based backbones to extract features [9, 30, 31, 43]. On the other hand, the voxel-based stream transforms point clouds into regular voxel representations to operate 3D convolutions [33, 42]. Their efficiency is then improved thanks to sparse convolution [10, 13, 19]. Point-based methods, in contrast, directly operate on the point cloud [12, 37, 38, 44, 58], and see a recent transition towards transformer-based architectures [21, 52, 59]. Following [56], we mainly pre-train on the voxel-based method SparseUNet [10] implemented with SpConv [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.897, + 0.902 + ], + "angle": 0, + "content": "3D Representation learning. Unlike the 2D counterparts, where large-scale unsupervised pre-training has been a common choice for facilitating downstream tasks [6], 3D representation learning is still not mature, and most works still train from scratch on the target data directly [28]. While earlier works in 3D representation learning simply build on a single object [22, 40, 41, 49], recent works start to train on scene-centric point clouds [28, 56]. However, unlike in 2D that scene-centric representation learning has been well-studied [32, 51, 55], the pre-training on 3D scenes, which relies on raw frame data [28, 56], still faces inefficiency issues and finds it hard to scale up to larger scale datasets. In contrast, we explore directly learning at the scene level, which shows significantly higher efficiency in processing scene data, and opens the possibility for pre-training with larger-scale point clouds, for the first time ever." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "9416" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.48, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.226, + 0.483, + 0.287 + ], + "angle": 0, + "content": "(a) Frame matching [28, 56]. 1. Extract raw RGB-D frames and camera positions from raw data. 2. Project each 2D frame into 3D space produces frame-level point cloud views. 3. Calculate pairwise overlapping rates among each frame of a single view and select pairs with overlapping rates larger than \\(30\\%\\) as pairs of contrastive views." + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.09, + 0.888, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.226, + 0.892, + 0.287 + ], + "angle": 0, + "content": "(b) Scene augmentation (ours). 1. Apply spatial augmentations containing rotation, flipping, and scaling. 2. Apply photometric augmentations containing brightness, contrast, saturation, hue, and gaussian noise jittering. 3. Generate and apply contrastive cross masks to the two views after sampling augmentations containing cropping and voxelization." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Figure 2. View generation. Compared with frame matching (FM), our scene augmentation (SA) is efficient and effective. 1. SA can end-to-end produce contrastive views on the original point cloud with ignorable latency, while FM requires preprocessing devouring enormous storage resources (e.g. additional 1.5 TB storage for ScanNet) in step 2 and pairwise matching is time-consuming. 2. SA produces scene-level views, while FM can only produce frame-level views containing limited information. Benefiting from advanced photometric augmentations, SA has the capacity to simulate the same scene under different lighting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.195, + 0.401 + ], + "angle": 0, + "content": "3. Pilot Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.471, + 0.453 + ], + "angle": 0, + "content": "This section analyzes the two main obstacles towards large-scale pre-training with point clouds. Our proposed design is based on the conclusion of the pilot study." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.454, + 0.379, + 0.468 + ], + "angle": 0, + "content": "Is matching RGB-D frames a good choice?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.605 + ], + "angle": 0, + "content": "As a seminal work in 3D representation learning, Point-Contrast [56] first enables pre-training in real-world indoor scenes with matched raw RGB-D frames as contrastive views. A visualization of the frame matching procedure is illustrated in Figure 2a. This protocol seems natural for indoor scenes since point clouds of indoor scenes are usually derived from RGB-D videos [1,5,16], where the raw frames are extracted from. However, this framework has multiple drawbacks that can hinder the scalability of training:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.607, + 0.469, + 0.683 + ], + "angle": 0, + "content": "- Redundant frame encoding. The pairwise matching strategy that PointContrast adopts allows one frame to be matched multiple times. As a result, each frame can be encoded multiple times in one step, adding to redundancy in training." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.685, + 0.469, + 0.776 + ], + "angle": 0, + "content": "- Low learning efficiency. In one training step, the frame matching strategy only allows the framework to process several views of a single scene. Therefore the amount of information that PointContrast can process in one step is rather limited, and the overall time for one training cycle is also notably high." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.779, + 0.469, + 0.868 + ], + "angle": 0, + "content": "- Dependency on raw RGB-D frames. The whole framework is built on the assumption that RGB-D videos are available, yet this is not true for every publicly-available point cloud dataset. Even when available, the storage cost of RGBD frames is also significantly higher than the reconstructed point cloud data." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.607, + 0.469, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Consequently, pre-training frameworks [28, 56] based on matching frames as contrastive samples require enormous" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.506 + ], + "angle": 0, + "content": "computing and storage resources. For example, Point-Contrast sub-samples RGB-D scans from the raw ScanNet videos every 25 frames, consuming \\(\\sim 30\\) times the storage of the processed point clouds, and takes 80 GPU hours to process an epoch of 1500 scenes. Even at such cost, our experiments in Table 1a verify that the raw RGB-D frames cannot bring additional information over the processed point clouds to achieve a better representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.547 + ], + "angle": 0, + "content": "We will explore the possibility of pre-training on the point clouds directly." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.558, + 0.816, + 0.572 + ], + "angle": 0, + "content": "What's the revelation behind mode collapse?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.892, + 0.86 + ], + "angle": 0, + "content": "Mode collapse, defined as the phenomenon that all features collapse to a single vector, remains an unsolved problem accompanying the development of 3D representation learning [11, 56]. To alleviate this problem, PointContrast introduced InfoNCE loss [35], which has been shown to stabilize training, to replace the hardest-contrastive loss. Yet, the problem of mode collapse can still occur when the amount of training data and the length of the training schedule increase. Given the empirical conclusion of 2D contrastive learning [8], the occurrence of mode collapse is unusual under the premise that a large number of negative samples are already adopted. Interestingly, we notice that the mean negative pair cosine similarity of previous works is mostly close to 0, indicating that the negative samples are mostly easy and thus have little penalty towards the trivial solution. Although the InfoNCE loss alleviates this problem with an alternated optimization objective, we argue that a more desirable solution can be achieved by raising the difficulty of the unsupervised pretext task." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We will further raise the difficulty of the pretext task to solve the mode collapse problem." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9417" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.878, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.344, + 0.896, + 0.442 + ], + "angle": 0, + "content": "Figure 3. Our MSC framework. (1) Generating a pair of contrastive views with a well-curated data augmentation pipeline consisting of photometric, spatial, and sampling augmentations. (2) Generating a pair of complementary masks and applying them to the pair of contrastive views. Replacing masked point features with a learnable mask token vector. (3) Extracting point representation with a given U-Net style backbone for point cloud understanding. (4) Reassembling masked contrastive views to masked points combination and unmasked points combination. (5) Matching points share similar positional relationships in the two views as positive sample pairs and computing InfoNCE loss to optimize contrastive objective. (6) Predicting masked point color and normal and computing Mean Squared Error loss and Cosine Similarity Loss with ground truth respectively, to optimize the reconstruction objective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.467, + 0.188, + 0.485 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.473, + 0.645 + ], + "angle": 0, + "content": "Based on the analysis of the stumbling blocks for large-scale pre-training in Section 3, we first introduce our optimized contrastive learning design in Section 4.1 to make the process more efficient. Then we solve the long-term problem of mode collapse with an additional reconstructive learning design in Section 4.2. The final optimization target is described in Section 4.3. Combining these exquisite designs, we build the whole framework, namely Masked Scene Contrast (MSC), and a visual illustration of our MSC is available in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.281, + 0.672 + ], + "angle": 0, + "content": "4.1. Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.799 + ], + "angle": 0, + "content": "**Framework.** Different from the previous protocol of matching RGB-D frames decomposed from indoor scenes, our contrastive learning framework directly operates on the point cloud data. Given a point cloud \\( \\mathbf{X} = (P, C) \\), where \\( P \\in \\mathbb{R}^{n \\times 3} \\) represents the spatial features (coordinate) of the points and \\( C \\in \\mathbb{R}^{n \\times 3} \\) represents the photometric features (color) of the points, the contrastive learning framework can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.806, + 0.47, + 0.882 + ], + "angle": 0, + "content": "- View generation. For a given point cloud \\( \\mathbf{X} \\), we generate query view \\( \\mathbf{X}_r \\) and key view \\( \\mathbf{X}_k \\) of the original point cloud with a sequence of stochastic data augmentations, which includes photometric, spatial, and sampling augmentations." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "- Feature extraction. Encoding point cloud features \\( \\mathbf{F}_r \\)" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.806, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.467, + 0.89, + 0.498 + ], + "angle": 0, + "content": "and \\(\\pmb{F}_k\\) with a U-Net style backbone \\(\\zeta(\\cdot)\\) to \\(\\hat{\\pmb{F}}_r\\) and \\(\\hat{\\pmb{F}}_k\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.506, + 0.892, + 0.611 + ], + "angle": 0, + "content": "- Point matching. The positive samples of contrastive learning are point pairs with close spatial positions in the two views. For each point belonging to the query view, we calculate the correspondence mapping \\(\\mathcal{P} = \\{(i,j)\\}_{n'}\\) to points of the key view. If \\((i,j) \\in \\mathcal{P}\\) then point \\((\\pmb{p}_i,\\pmb{c}_i)\\) and point \\((\\pmb{p}_j,\\pmb{c}_j)\\) constructs a pair across two views." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.62, + 0.894, + 0.68 + ], + "angle": 0, + "content": "- Loss computation. Computing the contrastive learning loss on the representation of two views \\(\\hat{F}_r\\) and \\(\\hat{F}_k\\) and the correspondence mapping \\(\\mathcal{P}\\). An encoded query view should be similar to its key view." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.506, + 0.894, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Data augmentation. As a pioneering work in image contrastive learning, SimCLR [7] reveals that a well-curated data augmentation pipeline is crucial for learning strong representations. Unlike supervised learning, contrastive learning requires much stronger data augmentations to prevent trivial solutions. However, an effective data augmentation recipe is still absent in 3D representation learning. The frame matching scheme in prior works [28,56] simply applies a randomly rotating operator to contrastive targets. Even if the RGB-D frames can be viewed as a natural random crop, the augmentation space is still far from diverse enough. The pretext task it forms is not yet challenging enough to facilitate the contrastive learning framework to learn robust representations for downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9418" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.473, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.471, + 0.318 + ], + "angle": 0, + "content": "Figure 4. View Mixing. Randomly mix up query views while keeping key views unmixed for a given batch of pairwise contrastive views. Detaching mixed query view after feature extraction for contrastive comparison with matched key views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.537 + ], + "angle": 0, + "content": "As presented in Figure 2b, our well-designed stochastic data augmentation pipeline includes photometric augmentations, spatial augmentations, and sampling augmentations. Inspired by the advanced photometric augmentation validated by our 2D counterparts [7, 20], we further strengthen the photometric augmentation component introduced by Choy et al. [10] with random brightness, contrast, saturation, hue, and gaussian noise jittering for photometric augmentation. Besides that, random rotating, flipping, and scaling constitute our spatial augmentations, and the sampling augmentation is composed of random cropping and grid sampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.541, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Empirically, the order of data augmentations is also a key component of our recipe. For example, grid sampling after random rotation leads to cross grids for sampling, which further increases the distinction between contrastive views and has a better augmentation effect. The specific data augmentation settings are available in the appendix, and a comparison with previous methods is presented in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.82 + ], + "angle": 0, + "content": "View mixing. Recently, Nekrasov et al. [34] proposes a data augmentation technique for 3D understanding models by mixing two scenes as a hybrid training sample, which can significantly suppress model overfitting. Inspired by the mixing mechanism, we integrate the logic of mixing as part of the contrastive learning objective. As illustrated in Figure 4, for a batch of pairwise views, we randomly mix up the query views while maintaining the key views unchanged before the feature extraction process. The simple operation can effectively increase the robustness of the backbone and improve the robustness of the point cloud representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.824, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Contrastive target. We follow the design of PointContrast on the contrast target and apply InfoNCE loss to the matched points. Given correspondence mapping \\(\\mathcal{P} = \\{(i,j)\\}_{n'}\\) produced by point matching and points representation \\(\\hat{\\pmb{F}}_r\\) and \\(\\hat{\\pmb{F}}_k\\) embedded during feature extraction, the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.093, + 0.625, + 0.106 + ], + "angle": 0, + "content": "contrastive loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.109, + 0.892, + 0.148 + ], + "angle": 0, + "content": "\\[\ns _ {i j} = \\frac {\\boldsymbol {f} _ {i ^ {\\prime}} ^ {r T} \\boldsymbol {f} _ {j ^ {\\prime}} ^ {k}}{| | \\boldsymbol {f} _ {i ^ {\\prime}} ^ {r T} | | \\cdot | | \\boldsymbol {f} _ {j ^ {\\prime}} ^ {k} | |}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.151, + 0.892, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {I n f o N C E}} = \\sum_ {i} ^ {n} - \\log \\frac {\\exp \\left(s _ {i i} / \\tau\\right)}{\\sum_ {j} ^ {n} \\exp \\left(s _ {i j} / \\tau\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.195, + 0.892, + 0.271 + ], + "angle": 0, + "content": "note that \\(S = \\{s_{ij}\\} \\in \\mathbb{R}^{n\\times n}\\) is the pairwise cosine similarity matrix between positive samples and negative samples, while \\(\\tau\\) is the temperature factor scaling cosine similarity. In practice, we control temperature factor \\(\\tau\\) as 0.4, which is the same as previous works [28,56]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.278, + 0.727, + 0.294 + ], + "angle": 0, + "content": "4.2. Reconstructive Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.451 + ], + "angle": 0, + "content": "As is mentioned in Section 3, one of the stumbling blocks for large-scale representation is mode collapse, and our solution is to scale up the difficulty of the unsupervised pre-training task. Motivated by the success of masked image modeling [23, 57] in 2D representations, we propose masked point modeling, which can be naturally integrated into our contrastive learning framework. Benefiting from this design, our framework can fully use non-overlapped regions of contrastive views that cannot be utilized by contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.893, + 0.648 + ], + "angle": 0, + "content": "Contrastive cross mask. The key design that enables additional construction learning in our contrastive learning framework is the contrastive cross mask. For a given query view and key view of a single point cloud, we partite the unioned point set into non-overlapping grid partitions by their original position before spatial augmentation. Given a mask rate \\( r \\) range from 0 to 0.5, we randomly generate a pair of masks \\( M_r, M_k \\in \\mathbb{R}^{1 \\times n_{r,k}} \\), in which there are no shared masked patches. Then, we follow the practice of SimMIM [57] to apply the pair of masks to the two views respectively by replacing the input feature with a learnable mask token vector \\( t \\in \\mathbb{R}^c \\). Consequently, the feature extraction process can be rewritten as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.654, + 0.892, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {F}} _ {r, k} = \\zeta \\left(\\left(1 - \\boldsymbol {M} _ {r, k}\\right) \\boldsymbol {F} _ {r, k} + \\boldsymbol {M} _ {r, k} \\boldsymbol {T} _ {r, k}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.891, + 0.707 + ], + "angle": 0, + "content": "where \\( T_{r,k} \\in \\mathbb{R}^{n_r, n_k \\times c} \\) is the expand matrix of mask token vector \\( t \\) to fit the feature dimensions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.892, + 0.783 + ], + "angle": 0, + "content": "Reconstruction target. The features of the point cloud are composed of two parts, the coordinates that determine the geometric structure and the colors that represent the texture features. We build up reconstruction targets for the two groups of features separately." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.783, + 0.892, + 0.858 + ], + "angle": 0, + "content": "The reconstruction of point cloud texture is straightforward, we predict the photometric value of each point with a linear projection. We compute the mean squared error (MSE) between the reconstructed and original color of masked points as the color reconstruction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.863, + 0.892, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} = \\frac {\\sum_ {i} ^ {n _ {r}} m _ {i} ^ {r} | | \\boldsymbol {x} _ {i} ^ {r} - \\hat {\\boldsymbol {x}} _ {i} ^ {r} | | _ {2} ^ {2} + \\sum_ {i} ^ {n _ {k}} m _ {i} ^ {k} | | \\boldsymbol {x} _ {i} ^ {k} - \\hat {\\boldsymbol {x}} _ {i} ^ {k} | | _ {2} ^ {2}}{n _ {r} ^ {\\prime} + n _ {k} ^ {\\prime}}, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.137 + ], + "angle": 0, + "content": "where \\( n_r' \\) and \\( n_k' \\) represent the number of mask points belonging to refer view and key view, \\( m_r^i \\) and \\( m_k^i \\) mean the \\( i \\)-th element of \\( M_r \\) and \\( M_k \\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.348 + ], + "angle": 0, + "content": "Point coordinates play an important role in describing the geometric structure of point clouds, and it is worth noting that directly reconstructing the coordinates of masked points is not reasonable since masked points are only sampled from 3D object surface rather than the continuous surface itself. Reconstructing points coordination would lead to an overfitted representation. To overcome the challenge, we introduce the concept of surfel reconstruction. Surfel is an abbreviation for a surface element or surface voxel in the discrete topology literature [26] and primitives rendering [36]. For each masked point, we reconstruct the normal vector of the corresponding surfel and compute the mean cosine similarity between estimations and surfel normals as a contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.354, + 0.47, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {n}} = \\frac {\\sum_ {i} ^ {n _ {r}} m _ {i} ^ {r} \\boldsymbol {x} _ {i} ^ {r T} \\hat {\\boldsymbol {x}} _ {i} ^ {r} + \\sum_ {i} ^ {n _ {k}} m _ {i} ^ {k} \\boldsymbol {x} _ {i} ^ {k T} \\hat {\\boldsymbol {x}} _ {i} ^ {k}}{n _ {r} ^ {\\prime} + n _ {k} ^ {\\prime}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.399, + 0.471, + 0.446 + ], + "angle": 0, + "content": "where \\( n_r' \\) and \\( n_k' \\) represent the number of mask points belonging to refer view and key view, \\( m_r^i \\) and \\( m_k^i \\) mean the \\( i \\)-th element of \\( M_r \\) and \\( M_k \\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.223, + 0.468 + ], + "angle": 0, + "content": "4.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Our framework combines the contrastive target, the color reconstruction target, and the surfel reconstruction to make the unsupervised task more scalable. The overall loss function is a weighted sum of Eq. 2, Eq. 4, and Eq. 5 which is written as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.563, + 0.47, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\lambda_ {c} \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {n} \\mathcal {L} _ {\\mathrm {n}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.589, + 0.47, + 0.651 + ], + "angle": 0, + "content": "where \\(\\lambda_{c}\\) and \\(\\lambda_{n}\\) are the weight parameters that balance the three loss components. Empirically we find that performance is robust to the choice of weight parameters, and we make \\(\\lambda_{c} = \\lambda_{n} = 1\\) in practice." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.663, + 0.21, + 0.68 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.809 + ], + "angle": 0, + "content": "We conduct extensive experimental evaluations to validate the capability of our framework, built upon the point cloud perception codebase Pointcept [14]. We first ablate our designs with an efficient pre-training pipeline that only utilizes ScanNet point cloud in Section 5.1, while without compromising performance. Then we explore large-scale pre-training across multiple datasets and compare our performance with previous results in Section 5.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.24, + 0.834 + ], + "angle": 0, + "content": "5.1. Main Properties" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We ablate the main designs and intriguing properties of our MSC in Table 1, and the default setting is available in the caption. We enable efficient pre-training by introducing our view generation pipeline, which is ablated in Table 1a." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.889, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Figure 5. Masked scenes and color reconstructions. We visualize one of the cross masks of each scene with a mask rate of \\(50\\%\\) (left) and color reconstruction of the masked point combinations (right). We pre-train our MSC with a mask patch size of \\(10\\mathrm{cm}\\) and generalize to results to different mask sizes. Compared with the original point clouds, the loss of detail cannot be avoided, while boundary and texture are well preserved by our model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.507 + ], + "angle": 0, + "content": "All of our ablation experiments only require 20G ScanNet point cloud data and around 14 hours pertaining on a single machine containing 8 NVIDIA RTX3090." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.704 + ], + "angle": 0, + "content": "View generation. In Table 1a, we show the results with different generation strategies of contrastive views. Unlike PointContrast [56], which builds on the raw RGB-D frames, our strategy directly utilizes the scene-level point cloud. This significantly reduces storage requirements (96% less) and allows more efficient use of the training data. With a 30% equivalent number of training iterations, our method can attain \\(120 \\times\\) training epochs, making full use of the training data for the first time. This results in 0.4 points higher FT mIoU with \\(4.4 \\times\\) speedup. When combined with an additional mask point modeling strategy, the performance can be further boosted by 0.6 points, yet still with a notable speedup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Number of positive pairs. In Table 1b, we show the effects of different numbers of positive pairs. Our method sees consistent improvements with an increasing number of positive pairs, while for PointContrast [56], the information in a large batch of positive pairs cannot be effectively utilized. Our intuition is that the ability to process scene-level views, which are larger in scale and contain much more information than frame-level views, enables our method to take advantage of a larger number of positive pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Data augmentation. In Table 1c, we analyze the effect of different data augmentation combinations. Adopting either spatial augmentation or photometric augmentation leads to sub-optimal performance, and the combination of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9420" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.09, + 0.871, + 0.153 + ], + "angle": 0, + "content": "
View generation methodsPre-train dataStorageBatch sizeItersEpochsFT mIoU (%)Hours (h)Speedup
Frame matching (PointContrast [56])ScanNet Raw500G32100k574.0481.0×
Scene augmentation w/o mask (ours)ScanNet v220G3230k60074.4114.4×
Scene augmentation w mask (ours)ScanNet v220G3230k60075.0143.4×
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.08, + 0.157, + 0.885, + 0.181 + ], + "angle": 0, + "content": "(a) View generation. Views produced by our enhanced data augmentation are stronger than the original RGB-D frames. Scene-level views can significantly speed up pre-training and make contrastive learning more effective. The performance can be further boosted with additional masked point modeling." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.185, + 0.316, + 0.262 + ], + "angle": 0, + "content": "
#Pos pairsPC [56]MSC (ours)
102473.874.3
204874.074.5
409873.774.9
819273.975.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.264, + 0.338, + 0.3 + ], + "angle": 0, + "content": "(b) Number of positive pairs. A larger amount of sampled positive pairs are necessary for scene-level views." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.306, + 0.334, + 0.384 + ], + "angle": 0, + "content": "
MaskTaskFT mIoU (%)
w/o crossw/o contrast74.1
w crossw/o contrast74.4
w/o crossw contrast74.7
w crossw contrast75.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.386, + 0.338, + 0.423 + ], + "angle": 0, + "content": "(e) Cross mask. Masks containing shared masked patches have a negative influence on contrastive learning." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.185, + 0.597, + 0.262 + ], + "angle": 0, + "content": "
SpatialPhotometricFT mIoU (%)
w/o augw/o aug72.1
w augw/o aug73.4
w/o augw aug72.8
w augw aug74.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.36, + 0.264, + 0.609, + 0.301 + ], + "angle": 0, + "content": "(c) Data augmentation. The combination of spatial and photometric augmentation makes the view generation pipeline comes to work." + }, + { + "type": "table", + "bbox": [ + 0.386, + 0.306, + 0.589, + 0.384 + ], + "angle": 0, + "content": "
Mask grid size (m)FT mIoU (%)
0.0574.3
0.175.0
0.1575.0
0.274.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.36, + 0.386, + 0.608, + 0.423 + ], + "angle": 0, + "content": "(f) Mask grid size. Our design works with mask patches with a grid size larger than \\(0.1\\mathrm{m}\\), and we consider \\(0.15\\mathrm{m}\\) as a default setting." + }, + { + "type": "table", + "bbox": [ + 0.645, + 0.185, + 0.871, + 0.262 + ], + "angle": 0, + "content": "
Query viewKey viewFT mIoU (%)
w/o mixw/o mix74.1
w mixw/o mix74.4
w/o mixw mix74.2
w mixw mix73.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.631, + 0.264, + 0.879, + 0.3 + ], + "angle": 0, + "content": "(d) View mixing. Randomly mixing query views while leaving key views unmixed is a sweet point." + }, + { + "type": "table", + "bbox": [ + 0.665, + 0.306, + 0.85, + 0.384 + ], + "angle": 0, + "content": "
ColorNormalFT mIoU (%)
w/ow/o74.4
ww/o74.9
w/ow74.6
ww75.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.631, + 0.386, + 0.879, + 0.423 + ], + "angle": 0, + "content": "(g) Reconstruction target. Both targets have a positive effect, while color reconstruction has a dominant impact on indoor scenes." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.428, + 0.892, + 0.484 + ], + "angle": 0, + "content": "Table 1. Ablation experiments. We adopt SparseUNet and efficient pre-training on ScanNet [16] point cloud data to ablate our designs. We report fine-turning (FT) mIoU (%) results on ScanNet 20 classes semantic segmentation as the default metric. If not specified, the default setting is as follows: the pre-training period is 600 epochs, the masking ratio is \\(30\\%\\) and the masked patch has a grid size of \\(0.15\\mathrm{m}\\) in the real-world space, view mixing probability is 0.8. All of our designs are enabled by default. Default settings are marked in gray." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.505, + 0.435, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.68 + ], + "angle": 0, + "content": "Figure 6. Masking ratio. A masking ratio ranging from \\(30\\%\\) to \\(40\\%\\) works well with our design, and a higher mask rate negatively influences contrastive learning. The y-axes represent ScanNet semantic segmentation validation mIoU \\((\\%)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.468, + 0.763 + ], + "angle": 0, + "content": "both helps make our view generation pipeline come to work. Concerning relative significance, since the spatial augmentation is highly coupled with the masking strategy, the gain from it is higher than the photometric augmentation. But still, both are necessary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.468, + 0.884 + ], + "angle": 0, + "content": "View mixing. In Table 1d, we explore different config for the view mixing strategy. Randomly mixing the query views while leaving the key views unchanged yields the best performance. Our intuition is that the key views, which form the vocabulary of the InfoNCE loss, should be relatively stable, thus reducing the ambiguity of the learning target. In contrast, introducing more diversity to the queries can be more helpful." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Cross mask. In Table 1e, we study the cross mask strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.577 + ], + "angle": 0, + "content": "This strategy ensures that the two augmented views have no overlapping tokens. As reported in the table, whether with or without the contrastive learning target, this strategy ensures fewer shortcuts to the task and enables consistent downstream improvements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Mask grid size. In Table 1f, we show the results ablating the grid size for producing masks on point clouds. Our design works with a grid size larger than \\(0.1\\mathrm{m}\\), and we consider \\(0.15\\mathrm{m}\\) as a default setting. As shown in Figure 5, our reconstruction module is robust to extending mask grid size, which indicates that high guilty representation is captured during pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Masking ratio. In Figure 6, we depict the effect of the ratio of masked tokens. A masking ratio ranging from \\(30\\%\\) to \\(40\\%\\) works well with our design, and a higher masking rate has a negative impact on the overall performance. This varies from the conclusion of MAE [23], in which a higher masking ratio of \\(75\\%\\) achieves top performance. Our hypothesis is that the contrastive learning objective built on the masked point clouds might favor a lower masking ratio, and the results in Figure 6 reflect a trade-off between the contrastive learning objective and the reconstructive objective. And pure reconstructive learning on point clouds is an interesting direction for future explorations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Reconstruction target. In Table 1g, we ablate the effects of two components of our reconstruction target: color re" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9421" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.457, + 0.16 + ], + "angle": 0, + "content": "
DatasetsBackbonesSemantic Seg. (mIoU)
SCPC [56]CSC [28]MSC (ours)
ScanNetSparseUNet72.274.1 (+1.9)73.8 (+1.6)75.5 (+3.3)
ScanNet200SparseUNet25.026.2 (+1.2)26.4 (+1.4)28.8 (+3.8)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.162, + 0.465, + 0.199 + ], + "angle": 0, + "content": "(a) Semantic segmentation. We conduct pre-training on SparseUNet and compare semantic segmentation mIoU \\((\\%)\\) results on ScanNet and ScanNet200 [39] validation set." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.203, + 0.459, + 0.274 + ], + "angle": 0, + "content": "
DatasetsBackbonesInstance Seg. (mAP@0.5)
SCPC [56]CSC [28]MSC (ours)
ScanNetSparseUNet56.958.0 (+1.1)59.4 (+2.5)59.6 (+2.7)
ScanNet200SparseUNet24.524.9 (+0.4)25.2 (+0.7)26.8 (+2.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.277, + 0.465, + 0.314 + ], + "angle": 0, + "content": "(b) Instance segmentation. We conduct pre-training on SparseUNet and compare instance segmentation mAP@0.5 \\((\\%)\\) results driven by PointGroup [29] on ScanNet and ScanNet200 [39] validation set." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.318, + 0.462, + 0.426 + ], + "angle": 0, + "content": "
LRSemantic Seg.LASemantic Seg.
Pct.SCCSCVIBUSMSCPts.SCCSCVIBUSMSC
100%72.273.8-75.3Full72.273.8-75.3
1%26.028.928.629.22041.955.561.061.2
5%47.849.847.450.75053.960.565.666.8
10%56.759.460.561.010062.265.968.969.7
20%62.964.664.864.920065.568.269.670.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.428, + 0.464, + 0.453 + ], + "angle": 0, + "content": "(c) Data efficiency. We follow the ScanNet Data Efficient benchmark and compare the validation results SparseUNet with previous methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.469, + 0.527 + ], + "angle": 0, + "content": "Table 2. Results comparison. We adopt cross-dataset pre-training utilizing ScanNet and ArkitScenes point cloud scenes for comparison of downstream task results. The pre-training setting is fixed as the default described in Table 1. More specific pre-training details are available in the Appendix. SC denotes train from scratch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.643 + ], + "angle": 0, + "content": "construction and normal reconstruction. Given the premise that indoor scenes are used, both targets show a positive effect on overall performance, while color reconstruction has a higher impact. The intuition is that the difference in texture reflected by color has a higher impact on the task of semantic segmentation, while the normal is helpful but has less influence (consider the same task on a 2D image)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.27, + 0.668 + ], + "angle": 0, + "content": "5.2. Results Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.78 + ], + "angle": 0, + "content": "In this section, we extend the scale of pre-training by merging multiple datasets and compare downstream task fine-turning performance with previous unsupervised pretraining frameworks [28, 56]. Specifically, we adopt the default model setting ablated in Section 5.1 and train on both ScanNet [16] and ArkitScenes [5] point clouds, extending pre-training assets from 1,513 scenes to 6,560 scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Semantic segmentation. In Table 2a, we report the semantic segmentation results on ScanNet and ScanNet200 [39] benchmark with SparseUNet and compare them with previous results. Our improvements are consistent and significant with larger pre-training assets for both benchmarks. With SparseUNet backbone, we outperform the current state-of-art pretraining framework by 1.7 points on ScanNet and 2.4 points on ScanNet200. Meanwhile, it is worth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "noting that driven by powerful MSC, we set a new best validation result on ScanNet semantic segmentation and pushed the previous SOTA to \\(75.5\\%\\) with a baseline model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.288 + ], + "angle": 0, + "content": "Instance segmentation. In Table 2b, we report the instance segmentation results on ScanNet and ScanNet200 [39] with SparseUNet backbone driven by PointGroup [29]. Comparing them with previous results, we still see consistent improvements. Specifically, our framework achieves \\(59.6\\%\\) on the ScanNet validation set, which is 3.3 points higher than training from scratch and 0.2 points promotion compared with the previous state-of-art performance. The boost is more significant on ScanNet200, which is 1.6 points higher than the previous SOTA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.894, + 0.35 + ], + "angle": 0, + "content": "Data efficiency. In Table 2c, we compare the ScanNet Data Efficient [28] results with previous methods. Our MSC shows consistently superior performance even compared with the latest data-efficient learning framework [45]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.365, + 0.749, + 0.381 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.526 + ], + "angle": 0, + "content": "In this paper, we tackle the problem of scalable unsupervised 3D representation learning. To this end, we present Masked Scene Contrast (MSC), an efficient, effective, and scalable framework that directly operates on scene-level views with contrastive learning and masked point modeling. Benefiting from the efficient scene-level point cloud processing pipeline and the effective training objectives, our method harvests high efficiency and superior generality, and enables large-scale pre-training across multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.893, + 0.692 + ], + "angle": 0, + "content": "The key factor that empowers our method's scalability to larger-scale pre-training lies in the efficient pipeline that can directly learn from point cloud data, rather than the raw RGB-D frames. The efficiency, however, does not only mean processing data at scale. When only the standard dataset ScanNet is used for pre-training, our method still achieves uncompromised performance, yet with at least \\(3 \\times\\) speedup over previous works. This is especially meaningful considering the exhaustively long experimental time in the pre-training community, and it can better facilitate the verification of new ideas in future works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.83 + ], + "angle": 0, + "content": "It should also be noted that given the limit in computing resources, and the absence of a pre-training dataset that is sustainably at scale, the scalability of our method is not fully presented. In other words, our method opens the possibility of large-scale pre-training on 3D point cloud data for the first time, and we call for a large-enough 3D scene dataset that can fully unleash this potential. We hope our method can inspire future works that take the first step to real large-scale 3D pre-training, as the 2D community does." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.838, + 0.669, + 0.854 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China (No. 62201484), HKU Startup Fund, and HKU Seed Fund for Basic Research." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9422" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.468, + 0.212 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael G. Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.468, + 0.242 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.468, + 0.283 + ], + "angle": 0, + "content": "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurIPS Workshops, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.468, + 0.465 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[9] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.536, + 0.468, + 0.577 + ], + "angle": 0, + "content": "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[11] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.468, + 0.648 + ], + "angle": 0, + "content": "[12] Ruihang Chu, Yukang Chen, Tao Kong, Lu Qi, and Lei Li. Icm-3d: Instantiated category modeling for 3d instance segmentation. RA-L, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.703 + ], + "angle": 0, + "content": "[13] Ruihang Chu, Xiaoqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, and Jiaya Jia. Twist: Two-way inter-label self-training for semi-supervised 3d instance segmentation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[14] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[15] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with exemplar convolutional neural networks. In TPAMI, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[18] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herve Jegou, and Edouard Grave. Are largescale datasets necessary for self-supervised pre-training? arXiv:2112.10740, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[19] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[21] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[22] Kaveh Hassani and Mike Haley. Unsupervised multi-task feature learning on point clouds. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.338, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.381, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.424, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[25] Olivier J Henaff, Aravind Srinivas, Jeffrey De Fauw, Ali Razavi, Carl Doersch, SM Eslami, and Aaron van den Oord. Data-efficient image recognition with contrastive predictive coding. In ICML, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[26] Gabor T Herman. Discrete multidimensional jordan surfaces. CVGIP, 1992. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[27] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[28] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[29] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. CVPR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.655, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[30] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.699, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[31] Bo Li, Tianlei Zhang, and Tian Xia. Vehicle detection from 3d lidar using fully convolutional network. In RSS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[32] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv:2011.13677, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[33] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.815, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[34] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Augmentation for 3D Scenes. In 3DV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9423" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[36] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In SIGGRAPH, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.176 + ], + "angle": 0, + "content": "[37] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[38] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.468, + 0.304 + ], + "angle": 0, + "content": "[40] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[41] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.39 + ], + "angle": 0, + "content": "[42] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.468, + 0.432 + ], + "angle": 0, + "content": "[43] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik G. Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[44] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.468, + 0.532 + ], + "angle": 0, + "content": "[45] Beiwen Tian, Liyi Luo, Hao Zhao, and Guyue Zhou. Vibus: Data-efficient 3d scene parsing with viewpoint bottleneck and uncertainty-spectrum modeling. ISPRS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.534, + 0.468, + 0.574 + ], + "angle": 0, + "content": "[46] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.577, + 0.468, + 0.616 + ], + "angle": 0, + "content": "[47] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.619, + 0.468, + 0.673 + ], + "angle": 0, + "content": "[48] Pascal Vincent, H. Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.676, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[49] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[50] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.774, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[51] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[52] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[53] Zhirong Wu, Zihang Lai, Xiao Sun, and Stephen Lin. Extreme masking for learning instance and distributed visual representations. arXiv:2206.04667, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[54] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[55] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[56] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[57] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[58] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[59] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[60] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9424" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_origin.pdf b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3234c837178b415769fd83f3eca0d7c837b3779f --- /dev/null +++ b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/5c72a7be-cb6e-40e5-904d-21114e3e1daa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87c50bb1830b6d296ef0a925fa42db79daa482b5b50e18ab3054718d1ae7cec0 +size 6204960 diff --git a/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/full.md b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ee0e7d6d76cb9893e986bf4425108d2efd128da1 --- /dev/null +++ b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/full.md @@ -0,0 +1,328 @@ +# Masked Scene Contrast: A Scalable Framework for Unsupervised 3D Representation Learning + +Xiaoyang Wu Xin Wen Xihui Liu Hengshuang Zhao* The University of Hong Kong + +https://github.com/Pointcept/Pointcept + +# Abstract + +As a pioneering work, PointContrast conducts unsupervised 3D representation learning via leveraging contrastive learning over raw RGB-D frames and proves its effectiveness on various downstream tasks. However, the trend of large-scale unsupervised learning in 3D has yet to emerge due to two stumbling blocks: the inefficiency of matching RGB-D frames as contrastive views and the annoying mode collapse phenomenon mentioned in previous works. Turning the two stumbling blocks into empirical stepping stones, we first propose an efficient and effective contrastive learning framework, which generates contrastive views directly on scene-level point clouds by a well-curated data augmentation pipeline and a practical view mixing strategy. Second, we introduce reconstructive learning on the contrastive learning framework with an exquisite design of contrastive cross masks, which targets the reconstruction of point color and surfel normal. Our Masked Scene Contrast (MSC) framework is capable of extracting comprehensive 3D representations more efficiently and effectively. It accelerates the pre-training procedure by at least $3 \times$ and still achieves an uncompromised performance compared with previous work. Besides, MSC also enables large-scale 3D pre-training across multiple datasets, which further boosts the performance and achieves state-of-the-art fine-tuning results on several downstream tasks, e.g., $75.5\%$ mIoU on ScanNet semantic segmentation validation set. + +# 1. Introduction + +Unsupervised visual representation learning aims at learning visual representations from vast amounts of unlabeled data. The learned representations are proved to be beneficial for various downstream tasks like segmentation and detection. It has attracted lots of attention and achieved remarkable progress in 2D image understanding, exceeding the upper bound of human supervision [20, 25]. + +![](images/5bca18c0c15d934a8a1c822d7d6380fefc53181b242f6aab28687fc5e35ebf9f.jpg) + +![](images/b92e0692df24dfe8bc3140d03f792bcae10feb36a78021125d206a6261570a36.jpg) +RGB-D Camera Random Crop Visible Region +Figure 1. Comparison of unsupervised 3D representation learning. The previous method [56] (top) relies on raw RGB-D frames with restricted views for contrastive learning, resulting in low efficiency and inferior versatility. Our approach (bottom) directly operates on scene-level views with contrastive learning and masked point modeling, leading to high efficiency and superior generality, further enabling large-scale pre-training across multiple datasets. + +Despite the impressive success of unsupervised visual representation learning in 2D, it is underexplored in 3D. Modern 3D scene understanding algorithms [10, 52] are focused on supervised learning, where models are trained directly from scratch on targeted datasets and tasks. Well-pretrained visual representations can undoubtedly boost the performance of these algorithms and are currently in urgent demand. Recent work PointContrast [56] conducts a preliminary exploration in 3D unsupervised learning. However, it is limited to raw RGB-D frames with an inefficient learning paradigm, which is not scalable and applicable to large-scale unsupervised learning. To address this essential and inevitable challenge, we focus on building a scalable framework for large-scale 3D unsupervised learning. + +One technical stumbling block towards large-scale pretraining is the inefficient learning strategy introduced by + +matching RGB-D frames as contrastive views. PointContrast [56] opens the door to pre-training on real indoor scene datasets and proposes frame matching to generate contrastive views with natural camera views, as in Figure 1 top. However, frame matching is inefficient since duplicated encoding exists for matched frames, resulting in limited scene diversity in batch training and optimization. Meanwhile, not all of the 3D scene data contains raw RGB-D frames, leading to failure deployments of the algorithm. Inspired by the great success of SimCLR [7], we investigate generating strong contrastive views by directly applying a series of well-curated data augmentations to scene-level point clouds, eliminating the dependence on raw RGB-D frames, as in Figure 1 bottom. Combined with an effective mechanism that mixes up query views, our contrastive learning design accelerates the pre-training procedure by $4.4 \times$ and achieves superior performance with purely point cloud data, compared to PointContast with raw data. The superior design also enables large-scale pre-training across multiple datasets like ScanNet [16] and ArkitScenes [5]. + +Another obstacle is the mode collapse phenomenon that occurs when scaling up the optimization iterations. We owe the culprit for this circumstance to the insufficient difficulty of unsupervised learning tasks. To further tackle the mode collapse challenge in unsupervised learning and scale up the optimization iterations, inspired by recent masked autoencoders [23, 57], we construct a masked point modeling paradigm where both point color reconstruction objective and surfel normal reconstruction objective are proposed to recover the masked color and geometric information of the point cloud respectively. We incorporate the mask point modeling strategy into our contrastive learning framework via an exquisite design of contrastive cross masks, leading towards a scalable unsupervised 3D representation learning framework, namely Masked Scene Contrast (MSC). + +Our framework is efficient, effective, and scalable. We conduct extensive experimental evaluations to validate its capability. On the popular point cloud dataset ScanNet, our algorithm accelerates the pre-training procedure by more than $3 \times$ , and achieves better performance on downstream tasks, when compared to the previous representative Point-Contrast. Besides, our method also enables large-scale 3D pre-training across multiple datasets, leading to state-of-the-art fine-tuning results on several downstream tasks, e.g. $75.5\%$ mIoU on ScanNet semantic segmentation validation set. In conclusion, our work opens up new possibilities for large-scale unsupervised 3D representation learning. + +# 2. Related Work + +2D Image contrastive learning. Based on the instance discrimination [17] pretext task, and combined with the contrastive learning [27, 47] paradigm, modern variants of 2D image contrastive representation learning have shown + +strong abilities in learning transferable visual representations [7, 24, 54]. With the learning objective built on the similarity between randomly augmented image views, this line of work strongly depends on a large batch size [7, 24] and a finely designed data augmentation pipeline [7, 20, 46] to achieve better performance. We find these two points also hold true for 3D contrastive learning. + +2D Image reconstructive learning. In 2D unsupervised learning, there is also a recent trend of switching the pretext task from instance discrimination [4,6,7,20,24] to masked image modeling [3,23,50,57,60]. Based on a denoising autoencoder [48]-style architecture, the task is to reconstruct the RGB value [23,57], discrete token [3,60], or feature [50] of masked pixels. This line of work has shown strong potential in learning representations on large-scale datasets and is less prone to model collapsing like instance-discrimination-based methods (e.g., contrastive learning). When combined with contrastive learning [2,18,53], the performance can be further boosted, yet with less dependence on data scale [18]. + +3D Scene understanding. The deep neural architectures for understanding 3D scenes can be roughly categorized into three paradigms according to the way they model point clouds: projection-based, voxel-based, and point-based. Projection-based works project 3D points into various image planes and adopt 2D CNN-based backbones to extract features [9, 30, 31, 43]. On the other hand, the voxel-based stream transforms point clouds into regular voxel representations to operate 3D convolutions [33, 42]. Their efficiency is then improved thanks to sparse convolution [10, 13, 19]. Point-based methods, in contrast, directly operate on the point cloud [12, 37, 38, 44, 58], and see a recent transition towards transformer-based architectures [21, 52, 59]. Following [56], we mainly pre-train on the voxel-based method SparseUNet [10] implemented with SpConv [15]. + +3D Representation learning. Unlike the 2D counterparts, where large-scale unsupervised pre-training has been a common choice for facilitating downstream tasks [6], 3D representation learning is still not mature, and most works still train from scratch on the target data directly [28]. While earlier works in 3D representation learning simply build on a single object [22, 40, 41, 49], recent works start to train on scene-centric point clouds [28, 56]. However, unlike in 2D that scene-centric representation learning has been well-studied [32, 51, 55], the pre-training on 3D scenes, which relies on raw frame data [28, 56], still faces inefficiency issues and finds it hard to scale up to larger scale datasets. In contrast, we explore directly learning at the scene level, which shows significantly higher efficiency in processing scene data, and opens the possibility for pre-training with larger-scale point clouds, for the first time ever. + +![](images/93f1b3604b750be5300d4a25bc343024ff70ac639363956ca61d69ab21ed4007.jpg) +(a) Frame matching [28, 56]. 1. Extract raw RGB-D frames and camera positions from raw data. 2. Project each 2D frame into 3D space produces frame-level point cloud views. 3. Calculate pairwise overlapping rates among each frame of a single view and select pairs with overlapping rates larger than $30\%$ as pairs of contrastive views. + +![](images/0a4c6dd5d370f437fa7b9a46213ca0038cfed07102698112ac2cceb9b743195e.jpg) +(b) Scene augmentation (ours). 1. Apply spatial augmentations containing rotation, flipping, and scaling. 2. Apply photometric augmentations containing brightness, contrast, saturation, hue, and gaussian noise jittering. 3. Generate and apply contrastive cross masks to the two views after sampling augmentations containing cropping and voxelization. +Figure 2. View generation. Compared with frame matching (FM), our scene augmentation (SA) is efficient and effective. 1. SA can end-to-end produce contrastive views on the original point cloud with ignorable latency, while FM requires preprocessing devouring enormous storage resources (e.g. additional 1.5 TB storage for ScanNet) in step 2 and pairwise matching is time-consuming. 2. SA produces scene-level views, while FM can only produce frame-level views containing limited information. Benefiting from advanced photometric augmentations, SA has the capacity to simulate the same scene under different lighting. + +# 3. Pilot Study + +This section analyzes the two main obstacles towards large-scale pre-training with point clouds. Our proposed design is based on the conclusion of the pilot study. + +# Is matching RGB-D frames a good choice? + +As a seminal work in 3D representation learning, Point-Contrast [56] first enables pre-training in real-world indoor scenes with matched raw RGB-D frames as contrastive views. A visualization of the frame matching procedure is illustrated in Figure 2a. This protocol seems natural for indoor scenes since point clouds of indoor scenes are usually derived from RGB-D videos [1,5,16], where the raw frames are extracted from. However, this framework has multiple drawbacks that can hinder the scalability of training: + +- Redundant frame encoding. The pairwise matching strategy that PointContrast adopts allows one frame to be matched multiple times. As a result, each frame can be encoded multiple times in one step, adding to redundancy in training. +- Low learning efficiency. In one training step, the frame matching strategy only allows the framework to process several views of a single scene. Therefore the amount of information that PointContrast can process in one step is rather limited, and the overall time for one training cycle is also notably high. +- Dependency on raw RGB-D frames. The whole framework is built on the assumption that RGB-D videos are available, yet this is not true for every publicly-available point cloud dataset. Even when available, the storage cost of RGBD frames is also significantly higher than the reconstructed point cloud data. + +Consequently, pre-training frameworks [28, 56] based on matching frames as contrastive samples require enormous + +computing and storage resources. For example, Point-Contrast sub-samples RGB-D scans from the raw ScanNet videos every 25 frames, consuming $\sim 30$ times the storage of the processed point clouds, and takes 80 GPU hours to process an epoch of 1500 scenes. Even at such cost, our experiments in Table 1a verify that the raw RGB-D frames cannot bring additional information over the processed point clouds to achieve a better representation. + +We will explore the possibility of pre-training on the point clouds directly. + +# What's the revelation behind mode collapse? + +Mode collapse, defined as the phenomenon that all features collapse to a single vector, remains an unsolved problem accompanying the development of 3D representation learning [11, 56]. To alleviate this problem, PointContrast introduced InfoNCE loss [35], which has been shown to stabilize training, to replace the hardest-contrastive loss. Yet, the problem of mode collapse can still occur when the amount of training data and the length of the training schedule increase. Given the empirical conclusion of 2D contrastive learning [8], the occurrence of mode collapse is unusual under the premise that a large number of negative samples are already adopted. Interestingly, we notice that the mean negative pair cosine similarity of previous works is mostly close to 0, indicating that the negative samples are mostly easy and thus have little penalty towards the trivial solution. Although the InfoNCE loss alleviates this problem with an alternated optimization objective, we argue that a more desirable solution can be achieved by raising the difficulty of the unsupervised pretext task. + +We will further raise the difficulty of the pretext task to solve the mode collapse problem. + +![](images/e44cb99999a7badb233eeb7577b714c00c3aa521f877ed13ede5835b7ec4bd84.jpg) +Figure 3. Our MSC framework. (1) Generating a pair of contrastive views with a well-curated data augmentation pipeline consisting of photometric, spatial, and sampling augmentations. (2) Generating a pair of complementary masks and applying them to the pair of contrastive views. Replacing masked point features with a learnable mask token vector. (3) Extracting point representation with a given U-Net style backbone for point cloud understanding. (4) Reassembling masked contrastive views to masked points combination and unmasked points combination. (5) Matching points share similar positional relationships in the two views as positive sample pairs and computing InfoNCE loss to optimize contrastive objective. (6) Predicting masked point color and normal and computing Mean Squared Error loss and Cosine Similarity Loss with ground truth respectively, to optimize the reconstruction objective. + +# 4. Approach + +Based on the analysis of the stumbling blocks for large-scale pre-training in Section 3, we first introduce our optimized contrastive learning design in Section 4.1 to make the process more efficient. Then we solve the long-term problem of mode collapse with an additional reconstructive learning design in Section 4.2. The final optimization target is described in Section 4.3. Combining these exquisite designs, we build the whole framework, namely Masked Scene Contrast (MSC), and a visual illustration of our MSC is available in Figure 3. + +# 4.1. Contrastive Learning + +**Framework.** Different from the previous protocol of matching RGB-D frames decomposed from indoor scenes, our contrastive learning framework directly operates on the point cloud data. Given a point cloud $\mathbf{X} = (P, C)$ , where $P \in \mathbb{R}^{n \times 3}$ represents the spatial features (coordinate) of the points and $C \in \mathbb{R}^{n \times 3}$ represents the photometric features (color) of the points, the contrastive learning framework can be summarized as follows: + +- View generation. For a given point cloud $\mathbf{X}$ , we generate query view $\mathbf{X}_r$ and key view $\mathbf{X}_k$ of the original point cloud with a sequence of stochastic data augmentations, which includes photometric, spatial, and sampling augmentations. +- Feature extraction. Encoding point cloud features $\mathbf{F}_r$ + +and $\pmb{F}_k$ with a U-Net style backbone $\zeta(\cdot)$ to $\hat{\pmb{F}}_r$ and $\hat{\pmb{F}}_k$ respectively. + +- Point matching. The positive samples of contrastive learning are point pairs with close spatial positions in the two views. For each point belonging to the query view, we calculate the correspondence mapping $\mathcal{P} = \{(i,j)\}_{n'}$ to points of the key view. If $(i,j) \in \mathcal{P}$ then point $(\pmb{p}_i,\pmb{c}_i)$ and point $(\pmb{p}_j,\pmb{c}_j)$ constructs a pair across two views. +- Loss computation. Computing the contrastive learning loss on the representation of two views $\hat{F}_r$ and $\hat{F}_k$ and the correspondence mapping $\mathcal{P}$ . An encoded query view should be similar to its key view. + +Data augmentation. As a pioneering work in image contrastive learning, SimCLR [7] reveals that a well-curated data augmentation pipeline is crucial for learning strong representations. Unlike supervised learning, contrastive learning requires much stronger data augmentations to prevent trivial solutions. However, an effective data augmentation recipe is still absent in 3D representation learning. The frame matching scheme in prior works [28,56] simply applies a randomly rotating operator to contrastive targets. Even if the RGB-D frames can be viewed as a natural random crop, the augmentation space is still far from diverse enough. The pretext task it forms is not yet challenging enough to facilitate the contrastive learning framework to learn robust representations for downstream tasks. + +![](images/4edc151802c93bc6091ae903bffbc9a0be3f2d3db9856aa4240262f90de67822.jpg) +Figure 4. View Mixing. Randomly mix up query views while keeping key views unmixed for a given batch of pairwise contrastive views. Detaching mixed query view after feature extraction for contrastive comparison with matched key views. + +As presented in Figure 2b, our well-designed stochastic data augmentation pipeline includes photometric augmentations, spatial augmentations, and sampling augmentations. Inspired by the advanced photometric augmentation validated by our 2D counterparts [7, 20], we further strengthen the photometric augmentation component introduced by Choy et al. [10] with random brightness, contrast, saturation, hue, and gaussian noise jittering for photometric augmentation. Besides that, random rotating, flipping, and scaling constitute our spatial augmentations, and the sampling augmentation is composed of random cropping and grid sampling. + +Empirically, the order of data augmentations is also a key component of our recipe. For example, grid sampling after random rotation leads to cross grids for sampling, which further increases the distinction between contrastive views and has a better augmentation effect. The specific data augmentation settings are available in the appendix, and a comparison with previous methods is presented in Figure 2. + +View mixing. Recently, Nekrasov et al. [34] proposes a data augmentation technique for 3D understanding models by mixing two scenes as a hybrid training sample, which can significantly suppress model overfitting. Inspired by the mixing mechanism, we integrate the logic of mixing as part of the contrastive learning objective. As illustrated in Figure 4, for a batch of pairwise views, we randomly mix up the query views while maintaining the key views unchanged before the feature extraction process. The simple operation can effectively increase the robustness of the backbone and improve the robustness of the point cloud representation. + +Contrastive target. We follow the design of PointContrast on the contrast target and apply InfoNCE loss to the matched points. Given correspondence mapping $\mathcal{P} = \{(i,j)\}_{n'}$ produced by point matching and points representation $\hat{\pmb{F}}_r$ and $\hat{\pmb{F}}_k$ embedded during feature extraction, the + +contrastive loss is: + +$$ +s _ {i j} = \frac {\boldsymbol {f} _ {i ^ {\prime}} ^ {r T} \boldsymbol {f} _ {j ^ {\prime}} ^ {k}}{| | \boldsymbol {f} _ {i ^ {\prime}} ^ {r T} | | \cdot | | \boldsymbol {f} _ {j ^ {\prime}} ^ {k} | |}, \tag {1} +$$ + +$$ +\mathcal {L} _ {\text {I n f o N C E}} = \sum_ {i} ^ {n} - \log \frac {\exp \left(s _ {i i} / \tau\right)}{\sum_ {j} ^ {n} \exp \left(s _ {i j} / \tau\right)}, \tag {2} +$$ + +note that $S = \{s_{ij}\} \in \mathbb{R}^{n\times n}$ is the pairwise cosine similarity matrix between positive samples and negative samples, while $\tau$ is the temperature factor scaling cosine similarity. In practice, we control temperature factor $\tau$ as 0.4, which is the same as previous works [28,56]. + +# 4.2. Reconstructive Learning + +As is mentioned in Section 3, one of the stumbling blocks for large-scale representation is mode collapse, and our solution is to scale up the difficulty of the unsupervised pre-training task. Motivated by the success of masked image modeling [23, 57] in 2D representations, we propose masked point modeling, which can be naturally integrated into our contrastive learning framework. Benefiting from this design, our framework can fully use non-overlapped regions of contrastive views that cannot be utilized by contrastive learning. + +Contrastive cross mask. The key design that enables additional construction learning in our contrastive learning framework is the contrastive cross mask. For a given query view and key view of a single point cloud, we partite the unioned point set into non-overlapping grid partitions by their original position before spatial augmentation. Given a mask rate $r$ range from 0 to 0.5, we randomly generate a pair of masks $M_r, M_k \in \mathbb{R}^{1 \times n_{r,k}}$ , in which there are no shared masked patches. Then, we follow the practice of SimMIM [57] to apply the pair of masks to the two views respectively by replacing the input feature with a learnable mask token vector $t \in \mathbb{R}^c$ . Consequently, the feature extraction process can be rewritten as follows: + +$$ +\hat {\boldsymbol {F}} _ {r, k} = \zeta \left(\left(1 - \boldsymbol {M} _ {r, k}\right) \boldsymbol {F} _ {r, k} + \boldsymbol {M} _ {r, k} \boldsymbol {T} _ {r, k}\right), \tag {3} +$$ + +where $T_{r,k} \in \mathbb{R}^{n_r, n_k \times c}$ is the expand matrix of mask token vector $t$ to fit the feature dimensions. + +Reconstruction target. The features of the point cloud are composed of two parts, the coordinates that determine the geometric structure and the colors that represent the texture features. We build up reconstruction targets for the two groups of features separately. + +The reconstruction of point cloud texture is straightforward, we predict the photometric value of each point with a linear projection. We compute the mean squared error (MSE) between the reconstructed and original color of masked points as the color reconstruction loss: + +$$ +\mathcal {L} _ {\mathrm {c}} = \frac {\sum_ {i} ^ {n _ {r}} m _ {i} ^ {r} | | \boldsymbol {x} _ {i} ^ {r} - \hat {\boldsymbol {x}} _ {i} ^ {r} | | _ {2} ^ {2} + \sum_ {i} ^ {n _ {k}} m _ {i} ^ {k} | | \boldsymbol {x} _ {i} ^ {k} - \hat {\boldsymbol {x}} _ {i} ^ {k} | | _ {2} ^ {2}}{n _ {r} ^ {\prime} + n _ {k} ^ {\prime}}, \tag {4} +$$ + +where $n_r'$ and $n_k'$ represent the number of mask points belonging to refer view and key view, $m_r^i$ and $m_k^i$ mean the $i$ -th element of $M_r$ and $M_k$ respectively. + +Point coordinates play an important role in describing the geometric structure of point clouds, and it is worth noting that directly reconstructing the coordinates of masked points is not reasonable since masked points are only sampled from 3D object surface rather than the continuous surface itself. Reconstructing points coordination would lead to an overfitted representation. To overcome the challenge, we introduce the concept of surfel reconstruction. Surfel is an abbreviation for a surface element or surface voxel in the discrete topology literature [26] and primitives rendering [36]. For each masked point, we reconstruct the normal vector of the corresponding surfel and compute the mean cosine similarity between estimations and surfel normals as a contrastive loss: + +$$ +\mathcal {L} _ {\mathrm {n}} = \frac {\sum_ {i} ^ {n _ {r}} m _ {i} ^ {r} \boldsymbol {x} _ {i} ^ {r T} \hat {\boldsymbol {x}} _ {i} ^ {r} + \sum_ {i} ^ {n _ {k}} m _ {i} ^ {k} \boldsymbol {x} _ {i} ^ {k T} \hat {\boldsymbol {x}} _ {i} ^ {k}}{n _ {r} ^ {\prime} + n _ {k} ^ {\prime}}, \tag {5} +$$ + +where $n_r'$ and $n_k'$ represent the number of mask points belonging to refer view and key view, $m_r^i$ and $m_k^i$ mean the $i$ -th element of $M_r$ and $M_k$ respectively. + +# 4.3. Loss Function + +Our framework combines the contrastive target, the color reconstruction target, and the surfel reconstruction to make the unsupervised task more scalable. The overall loss function is a weighted sum of Eq. 2, Eq. 4, and Eq. 5 which is written as follow: + +$$ +\mathcal {L} _ {\text {o v e r a l l}} = \mathcal {L} _ {\text {I n f o N C E}} + \lambda_ {c} \mathcal {L} _ {\mathrm {c}} + \lambda_ {n} \mathcal {L} _ {\mathrm {n}}, \tag {6} +$$ + +where $\lambda_{c}$ and $\lambda_{n}$ are the weight parameters that balance the three loss components. Empirically we find that performance is robust to the choice of weight parameters, and we make $\lambda_{c} = \lambda_{n} = 1$ in practice. + +# 5. Experiments + +We conduct extensive experimental evaluations to validate the capability of our framework, built upon the point cloud perception codebase Pointcept [14]. We first ablate our designs with an efficient pre-training pipeline that only utilizes ScanNet point cloud in Section 5.1, while without compromising performance. Then we explore large-scale pre-training across multiple datasets and compare our performance with previous results in Section 5.2. + +# 5.1. Main Properties + +We ablate the main designs and intriguing properties of our MSC in Table 1, and the default setting is available in the caption. We enable efficient pre-training by introducing our view generation pipeline, which is ablated in Table 1a. + +![](images/fd17a2c7fb786d0e34c0355be1dff800a9141f21a1531f19c287ea93361b948f.jpg) +Figure 5. Masked scenes and color reconstructions. We visualize one of the cross masks of each scene with a mask rate of $50\%$ (left) and color reconstruction of the masked point combinations (right). We pre-train our MSC with a mask patch size of $10\mathrm{cm}$ and generalize to results to different mask sizes. Compared with the original point clouds, the loss of detail cannot be avoided, while boundary and texture are well preserved by our model. + +All of our ablation experiments only require 20G ScanNet point cloud data and around 14 hours pertaining on a single machine containing 8 NVIDIA RTX3090. + +View generation. In Table 1a, we show the results with different generation strategies of contrastive views. Unlike PointContrast [56], which builds on the raw RGB-D frames, our strategy directly utilizes the scene-level point cloud. This significantly reduces storage requirements (96% less) and allows more efficient use of the training data. With a 30% equivalent number of training iterations, our method can attain $120 \times$ training epochs, making full use of the training data for the first time. This results in 0.4 points higher FT mIoU with $4.4 \times$ speedup. When combined with an additional mask point modeling strategy, the performance can be further boosted by 0.6 points, yet still with a notable speedup. + +Number of positive pairs. In Table 1b, we show the effects of different numbers of positive pairs. Our method sees consistent improvements with an increasing number of positive pairs, while for PointContrast [56], the information in a large batch of positive pairs cannot be effectively utilized. Our intuition is that the ability to process scene-level views, which are larger in scale and contain much more information than frame-level views, enables our method to take advantage of a larger number of positive pairs. + +Data augmentation. In Table 1c, we analyze the effect of different data augmentation combinations. Adopting either spatial augmentation or photometric augmentation leads to sub-optimal performance, and the combination of + +
View generation methodsPre-train dataStorageBatch sizeItersEpochsFT mIoU (%)Hours (h)Speedup
Frame matching (PointContrast [56])ScanNet Raw500G32100k574.0481.0×
Scene augmentation w/o mask (ours)ScanNet v220G3230k60074.4114.4×
Scene augmentation w mask (ours)ScanNet v220G3230k60075.0143.4×
+ +(a) View generation. Views produced by our enhanced data augmentation are stronger than the original RGB-D frames. Scene-level views can significantly speed up pre-training and make contrastive learning more effective. The performance can be further boosted with additional masked point modeling. + +
#Pos pairsPC [56]MSC (ours)
102473.874.3
204874.074.5
409873.774.9
819273.975.0
+ +(b) Number of positive pairs. A larger amount of sampled positive pairs are necessary for scene-level views. + +
MaskTaskFT mIoU (%)
w/o crossw/o contrast74.1
w crossw/o contrast74.4
w/o crossw contrast74.7
w crossw contrast75.0
+ +(e) Cross mask. Masks containing shared masked patches have a negative influence on contrastive learning. + +
SpatialPhotometricFT mIoU (%)
w/o augw/o aug72.1
w augw/o aug73.4
w/o augw aug72.8
w augw aug74.4
+ +(c) Data augmentation. The combination of spatial and photometric augmentation makes the view generation pipeline comes to work. + +
Mask grid size (m)FT mIoU (%)
0.0574.3
0.175.0
0.1575.0
0.274.8
+ +(f) Mask grid size. Our design works with mask patches with a grid size larger than $0.1\mathrm{m}$ , and we consider $0.15\mathrm{m}$ as a default setting. + +
Query viewKey viewFT mIoU (%)
w/o mixw/o mix74.1
w mixw/o mix74.4
w/o mixw mix74.2
w mixw mix73.7
+ +(d) View mixing. Randomly mixing query views while leaving key views unmixed is a sweet point. + +
ColorNormalFT mIoU (%)
w/ow/o74.4
ww/o74.9
w/ow74.6
ww75.0
+ +(g) Reconstruction target. Both targets have a positive effect, while color reconstruction has a dominant impact on indoor scenes. + +Table 1. Ablation experiments. We adopt SparseUNet and efficient pre-training on ScanNet [16] point cloud data to ablate our designs. We report fine-turning (FT) mIoU (%) results on ScanNet 20 classes semantic segmentation as the default metric. If not specified, the default setting is as follows: the pre-training period is 600 epochs, the masking ratio is $30\%$ and the masked patch has a grid size of $0.15\mathrm{m}$ in the real-world space, view mixing probability is 0.8. All of our designs are enabled by default. Default settings are marked in gray. + +![](images/6cab827345a95d521ba2da5e5633418f8e425c42492928a9b20b20a01886af92.jpg) +Figure 6. Masking ratio. A masking ratio ranging from $30\%$ to $40\%$ works well with our design, and a higher mask rate negatively influences contrastive learning. The y-axes represent ScanNet semantic segmentation validation mIoU $(\%)$ . + +both helps make our view generation pipeline come to work. Concerning relative significance, since the spatial augmentation is highly coupled with the masking strategy, the gain from it is higher than the photometric augmentation. But still, both are necessary. + +View mixing. In Table 1d, we explore different config for the view mixing strategy. Randomly mixing the query views while leaving the key views unchanged yields the best performance. Our intuition is that the key views, which form the vocabulary of the InfoNCE loss, should be relatively stable, thus reducing the ambiguity of the learning target. In contrast, introducing more diversity to the queries can be more helpful. + +Cross mask. In Table 1e, we study the cross mask strategy. + +This strategy ensures that the two augmented views have no overlapping tokens. As reported in the table, whether with or without the contrastive learning target, this strategy ensures fewer shortcuts to the task and enables consistent downstream improvements. + +Mask grid size. In Table 1f, we show the results ablating the grid size for producing masks on point clouds. Our design works with a grid size larger than $0.1\mathrm{m}$ , and we consider $0.15\mathrm{m}$ as a default setting. As shown in Figure 5, our reconstruction module is robust to extending mask grid size, which indicates that high guilty representation is captured during pre-training. + +Masking ratio. In Figure 6, we depict the effect of the ratio of masked tokens. A masking ratio ranging from $30\%$ to $40\%$ works well with our design, and a higher masking rate has a negative impact on the overall performance. This varies from the conclusion of MAE [23], in which a higher masking ratio of $75\%$ achieves top performance. Our hypothesis is that the contrastive learning objective built on the masked point clouds might favor a lower masking ratio, and the results in Figure 6 reflect a trade-off between the contrastive learning objective and the reconstructive objective. And pure reconstructive learning on point clouds is an interesting direction for future explorations. + +Reconstruction target. In Table 1g, we ablate the effects of two components of our reconstruction target: color re + +
DatasetsBackbonesSemantic Seg. (mIoU)
SCPC [56]CSC [28]MSC (ours)
ScanNetSparseUNet72.274.1 (+1.9)73.8 (+1.6)75.5 (+3.3)
ScanNet200SparseUNet25.026.2 (+1.2)26.4 (+1.4)28.8 (+3.8)
+ +(a) Semantic segmentation. We conduct pre-training on SparseUNet and compare semantic segmentation mIoU $(\%)$ results on ScanNet and ScanNet200 [39] validation set. + +
DatasetsBackbonesInstance Seg. (mAP@0.5)
SCPC [56]CSC [28]MSC (ours)
ScanNetSparseUNet56.958.0 (+1.1)59.4 (+2.5)59.6 (+2.7)
ScanNet200SparseUNet24.524.9 (+0.4)25.2 (+0.7)26.8 (+2.3)
+ +(b) Instance segmentation. We conduct pre-training on SparseUNet and compare instance segmentation mAP@0.5 $(\%)$ results driven by PointGroup [29] on ScanNet and ScanNet200 [39] validation set. + +
LRSemantic Seg.LASemantic Seg.
Pct.SCCSCVIBUSMSCPts.SCCSCVIBUSMSC
100%72.273.8-75.3Full72.273.8-75.3
1%26.028.928.629.22041.955.561.061.2
5%47.849.847.450.75053.960.565.666.8
10%56.759.460.561.010062.265.968.969.7
20%62.964.664.864.920065.568.269.670.7
+ +(c) Data efficiency. We follow the ScanNet Data Efficient benchmark and compare the validation results SparseUNet with previous methods. + +Table 2. Results comparison. We adopt cross-dataset pre-training utilizing ScanNet and ArkitScenes point cloud scenes for comparison of downstream task results. The pre-training setting is fixed as the default described in Table 1. More specific pre-training details are available in the Appendix. SC denotes train from scratch. + +construction and normal reconstruction. Given the premise that indoor scenes are used, both targets show a positive effect on overall performance, while color reconstruction has a higher impact. The intuition is that the difference in texture reflected by color has a higher impact on the task of semantic segmentation, while the normal is helpful but has less influence (consider the same task on a 2D image). + +# 5.2. Results Comparison + +In this section, we extend the scale of pre-training by merging multiple datasets and compare downstream task fine-turning performance with previous unsupervised pretraining frameworks [28, 56]. Specifically, we adopt the default model setting ablated in Section 5.1 and train on both ScanNet [16] and ArkitScenes [5] point clouds, extending pre-training assets from 1,513 scenes to 6,560 scenes. + +Semantic segmentation. In Table 2a, we report the semantic segmentation results on ScanNet and ScanNet200 [39] benchmark with SparseUNet and compare them with previous results. Our improvements are consistent and significant with larger pre-training assets for both benchmarks. With SparseUNet backbone, we outperform the current state-of-art pretraining framework by 1.7 points on ScanNet and 2.4 points on ScanNet200. Meanwhile, it is worth + +noting that driven by powerful MSC, we set a new best validation result on ScanNet semantic segmentation and pushed the previous SOTA to $75.5\%$ with a baseline model. + +Instance segmentation. In Table 2b, we report the instance segmentation results on ScanNet and ScanNet200 [39] with SparseUNet backbone driven by PointGroup [29]. Comparing them with previous results, we still see consistent improvements. Specifically, our framework achieves $59.6\%$ on the ScanNet validation set, which is 3.3 points higher than training from scratch and 0.2 points promotion compared with the previous state-of-art performance. The boost is more significant on ScanNet200, which is 1.6 points higher than the previous SOTA. + +Data efficiency. In Table 2c, we compare the ScanNet Data Efficient [28] results with previous methods. Our MSC shows consistently superior performance even compared with the latest data-efficient learning framework [45]. + +# 6. Conclusion and Discussion + +In this paper, we tackle the problem of scalable unsupervised 3D representation learning. To this end, we present Masked Scene Contrast (MSC), an efficient, effective, and scalable framework that directly operates on scene-level views with contrastive learning and masked point modeling. Benefiting from the efficient scene-level point cloud processing pipeline and the effective training objectives, our method harvests high efficiency and superior generality, and enables large-scale pre-training across multiple datasets. + +The key factor that empowers our method's scalability to larger-scale pre-training lies in the efficient pipeline that can directly learn from point cloud data, rather than the raw RGB-D frames. The efficiency, however, does not only mean processing data at scale. When only the standard dataset ScanNet is used for pre-training, our method still achieves uncompromised performance, yet with at least $3 \times$ speedup over previous works. This is especially meaningful considering the exhaustively long experimental time in the pre-training community, and it can better facilitate the verification of new ideas in future works. + +It should also be noted that given the limit in computing resources, and the absence of a pre-training dataset that is sustainably at scale, the scalability of our method is not fully presented. In other words, our method opens the possibility of large-scale pre-training on 3D point cloud data for the first time, and we call for a large-enough 3D scene dataset that can fully unleash this potential. We hope our method can inspire future works that take the first step to real large-scale 3D pre-training, as the 2D community does. + +# Acknowledgements + +This work is supported in part by the National Natural Science Foundation of China (No. 62201484), HKU Startup Fund, and HKU Seed Fund for Basic Research. + +# References + +[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 3 +[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael G. Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 2 +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In ICLR, 2022. 2 +[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 2 +[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurIPS Workshops, 2021. 2, 3, 8 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, 2021. 2 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4, 5 +[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3 +[9] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 2 +[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 5 +[11] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 3 +[12] Ruihang Chu, Yukang Chen, Tao Kong, Lu Qi, and Lei Li. Icm-3d: Instantiated category modeling for 3d instance segmentation. RA-L, 2021. 2 +[13] Ruihang Chu, Xiaoqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, and Jiaya Jia. Twist: Two-way inter-label self-training for semi-supervised 3d instance segmentation. In CVPR, 2022. 2 +[14] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.6 +[15] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.2 +[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3, 7, 8 +[17] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with exemplar convolutional neural networks. In TPAMI, 2015. 2 + +[18] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herve Jegou, and Edouard Grave. Are largescale datasets necessary for self-supervised pre-training? arXiv:2112.10740, 2021. 2 +[19] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2 +[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020. 1, 2, 5 +[21] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 2021. 2 +[22] Kaveh Hassani and Mike Haley. Unsupervised multi-task feature learning on point clouds. In ICCV, 2019. 2 +[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 5, 7 +[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2 +[25] Olivier J Henaff, Aravind Srinivas, Jeffrey De Fauw, Ali Razavi, Carl Doersch, SM Eslami, and Aaron van den Oord. Data-efficient image recognition with contrastive predictive coding. In ICML, 2020. 1 +[26] Gabor T Herman. Discrete multidimensional jordan surfaces. CVGIP, 1992. 6 +[27] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2018. 2 +[28] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 3, 4, 5, 8 +[29] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. CVPR, 2020. 8 +[30] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 2 +[31] Bo Li, Tianlei Zhang, and Tian Xia. Vehicle detection from 3d lidar using fully convolutional network. In RSS, 2016. 2 +[32] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv:2011.13677, 2020. 2 +[33] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015. 2 +[34] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Augmentation for 3D Scenes. In 3DV, 2021. 5 +[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 3 + +[36] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In SIGGRAPH, 2000. 6 +[37] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2 +[38] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2 +[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 8 +[40] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In ECCV, 2020. 2 +[41] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. In NeurIPS, 2019. 2 +[42] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, 2017. 2 +[43] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik G. Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. 2 +[44] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2 +[45] Beiwen Tian, Liyi Luo, Hao Zhao, and Guyue Zhou. Vibus: Data-efficient 3d scene parsing with viewpoint bottleneck and uncertainty-spectrum modeling. ISPRS, 2022. 8 +[46] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, 2020. 2 +[47] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 2 +[48] Pascal Vincent, H. Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 2010. 2 +[49] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, 2019. 2 +[50] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. 2 +[51] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. In NeurIPS, 2022. 2 +[52] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1, 2 +[53] Zhirong Wu, Zihang Lai, Xiao Sun, and Stephen Lin. Extreme masking for learning instance and distributed visual representations. arXiv:2206.04667, 2022. 2 + +[54] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2 +[55] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. In NeurIPS, 2021. 2 +[56] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 1, 2, 3, 4, 5, 6, 7, 8 +[57] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2, 5 +[58] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2 +[59] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2 +[60] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 2 \ No newline at end of file diff --git a/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/images.zip b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f99f6b7a822208a53df08437e6e4ca8b8171fc1c --- /dev/null +++ b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d390210d463ed5fe21170c6c2e02f04205dfecf5c6dd82316ef8962c5757b50b +size 547699 diff --git a/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/layout.json b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..539aa0376152be56ae91a6fdc1536844977a9fd1 --- /dev/null +++ b/2023/Masked Scene Contrast_ A Scalable Framework for Unsupervised 3D Representation Learning/layout.json @@ -0,0 +1,8714 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 103, + 458, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 103, + 458, + 140 + ], + "spans": [ + { + "bbox": [ + 137, + 103, + 458, + 140 + ], + "type": "text", + "content": "Masked Scene Contrast: A Scalable Framework for Unsupervised 3D Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 152, + 161, + 445, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 161, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 152, + 161, + 445, + 190 + ], + "type": "text", + "content": "Xiaoyang Wu Xin Wen Xihui Liu Hengshuang Zhao* The University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 191, + 191, + 399, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 191, + 399, + 203 + ], + "spans": [ + { + "bbox": [ + 191, + 191, + 399, + 203 + ], + "type": "text", + "content": "https://github.com/Pointcept/Pointcept" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "text", + "content": "As a pioneering work, PointContrast conducts unsupervised 3D representation learning via leveraging contrastive learning over raw RGB-D frames and proves its effectiveness on various downstream tasks. However, the trend of large-scale unsupervised learning in 3D has yet to emerge due to two stumbling blocks: the inefficiency of matching RGB-D frames as contrastive views and the annoying mode collapse phenomenon mentioned in previous works. Turning the two stumbling blocks into empirical stepping stones, we first propose an efficient and effective contrastive learning framework, which generates contrastive views directly on scene-level point clouds by a well-curated data augmentation pipeline and a practical view mixing strategy. Second, we introduce reconstructive learning on the contrastive learning framework with an exquisite design of contrastive cross masks, which targets the reconstruction of point color and surfel normal. Our Masked Scene Contrast (MSC) framework is capable of extracting comprehensive 3D representations more efficiently and effectively. It accelerates the pre-training procedure by at least " + }, + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "text", + "content": " and still achieves an uncompromised performance compared with previous work. Besides, MSC also enables large-scale 3D pre-training across multiple datasets, which further boosts the performance and achieves state-of-the-art fine-tuning results on several downstream tasks, e.g., " + }, + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "inline_equation", + "content": "75.5\\%" + }, + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "text", + "content": " mIoU on ScanNet semantic segmentation validation set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 590, + 128, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 128, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 128, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "type": "text", + "content": "Unsupervised visual representation learning aims at learning visual representations from vast amounts of unlabeled data. The learned representations are proved to be beneficial for various downstream tasks like segmentation and detection. It has attracted lots of attention and achieved remarkable progress in 2D image understanding, exceeding the upper bound of human supervision [20, 25]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 231, + 527, + 322 + ], + "blocks": [ + { + "bbox": [ + 321, + 231, + 527, + 322 + ], + "lines": [ + { + "bbox": [ + 321, + 231, + 527, + 322 + ], + "spans": [ + { + "bbox": [ + 321, + 231, + 527, + 322 + ], + "type": "image", + "image_path": "5bca18c0c15d934a8a1c822d7d6380fefc53181b242f6aab28687fc5e35ebf9f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 323, + 527, + 421 + ], + "blocks": [ + { + "bbox": [ + 321, + 323, + 527, + 421 + ], + "lines": [ + { + "bbox": [ + 321, + 323, + 527, + 421 + ], + "spans": [ + { + "bbox": [ + 321, + 323, + 527, + 421 + ], + "type": "image", + "image_path": "b92e0692df24dfe8bc3140d03f792bcae10feb36a78021125d206a6261570a36.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 422, + 523, + 434 + ], + "lines": [ + { + "bbox": [ + 323, + 422, + 523, + 434 + ], + "spans": [ + { + "bbox": [ + 323, + 422, + 523, + 434 + ], + "type": "text", + "content": "RGB-D Camera Random Crop Visible Region" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 304, + 436, + 545, + 513 + ], + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 513 + ], + "type": "text", + "content": "Figure 1. Comparison of unsupervised 3D representation learning. The previous method [56] (top) relies on raw RGB-D frames with restricted views for contrastive learning, resulting in low efficiency and inferior versatility. Our approach (bottom) directly operates on scene-level views with contrastive learning and masked point modeling, leading to high efficiency and superior generality, further enabling large-scale pre-training across multiple datasets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 520, + 546, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 546, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 546, + 688 + ], + "type": "text", + "content": "Despite the impressive success of unsupervised visual representation learning in 2D, it is underexplored in 3D. Modern 3D scene understanding algorithms [10, 52] are focused on supervised learning, where models are trained directly from scratch on targeted datasets and tasks. Well-pretrained visual representations can undoubtedly boost the performance of these algorithms and are currently in urgent demand. Recent work PointContrast [56] conducts a preliminary exploration in 3D unsupervised learning. However, it is limited to raw RGB-D frames with an inefficient learning paradigm, which is not scalable and applicable to large-scale unsupervised learning. To address this essential and inevitable challenge, we focus on building a scalable framework for large-scale 3D unsupervised learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "One technical stumbling block towards large-scale pretraining is the inefficient learning strategy introduced by" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 251, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 251, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 251, + 712 + ], + "type": "text", + "content": "*Corresponding Author. Email: hszhao@cs.hku.hk" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9415" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "matching RGB-D frames as contrastive views. PointContrast [56] opens the door to pre-training on real indoor scene datasets and proposes frame matching to generate contrastive views with natural camera views, as in Figure 1 top. However, frame matching is inefficient since duplicated encoding exists for matched frames, resulting in limited scene diversity in batch training and optimization. Meanwhile, not all of the 3D scene data contains raw RGB-D frames, leading to failure deployments of the algorithm. Inspired by the great success of SimCLR [7], we investigate generating strong contrastive views by directly applying a series of well-curated data augmentations to scene-level point clouds, eliminating the dependence on raw RGB-D frames, as in Figure 1 bottom. Combined with an effective mechanism that mixes up query views, our contrastive learning design accelerates the pre-training procedure by " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "4.4 \\times" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": " and achieves superior performance with purely point cloud data, compared to PointContast with raw data. The superior design also enables large-scale pre-training across multiple datasets like ScanNet [16] and ArkitScenes [5]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 312, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 289, + 491 + ], + "type": "text", + "content": "Another obstacle is the mode collapse phenomenon that occurs when scaling up the optimization iterations. We owe the culprit for this circumstance to the insufficient difficulty of unsupervised learning tasks. To further tackle the mode collapse challenge in unsupervised learning and scale up the optimization iterations, inspired by recent masked autoencoders [23, 57], we construct a masked point modeling paradigm where both point color reconstruction objective and surfel normal reconstruction objective are proposed to recover the masked color and geometric information of the point cloud respectively. We incorporate the mask point modeling strategy into our contrastive learning framework via an exquisite design of contrastive cross masks, leading towards a scalable unsupervised 3D representation learning framework, namely Masked Scene Contrast (MSC)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 491, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 289, + 635 + ], + "type": "text", + "content": "Our framework is efficient, effective, and scalable. We conduct extensive experimental evaluations to validate its capability. On the popular point cloud dataset ScanNet, our algorithm accelerates the pre-training procedure by more than " + }, + { + "bbox": [ + 47, + 491, + 289, + 635 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 47, + 491, + 289, + 635 + ], + "type": "text", + "content": ", and achieves better performance on downstream tasks, when compared to the previous representative Point-Contrast. Besides, our method also enables large-scale 3D pre-training across multiple datasets, leading to state-of-the-art fine-tuning results on several downstream tasks, e.g. " + }, + { + "bbox": [ + 47, + 491, + 289, + 635 + ], + "type": "inline_equation", + "content": "75.5\\%" + }, + { + "bbox": [ + 47, + 491, + 289, + 635 + ], + "type": "text", + "content": " mIoU on ScanNet semantic segmentation validation set. In conclusion, our work opens up new possibilities for large-scale unsupervised 3D representation learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "content": "2D Image contrastive learning. Based on the instance discrimination [17] pretext task, and combined with the contrastive learning [27, 47] paradigm, modern variants of 2D image contrastive representation learning have shown" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "strong abilities in learning transferable visual representations [7, 24, 54]. With the learning objective built on the similarity between randomly augmented image views, this line of work strongly depends on a large batch size [7, 24] and a finely designed data augmentation pipeline [7, 20, 46] to achieve better performance. We find these two points also hold true for 3D contrastive learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 170, + 547, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 547, + 315 + ], + "type": "text", + "content": "2D Image reconstructive learning. In 2D unsupervised learning, there is also a recent trend of switching the pretext task from instance discrimination [4,6,7,20,24] to masked image modeling [3,23,50,57,60]. Based on a denoising autoencoder [48]-style architecture, the task is to reconstruct the RGB value [23,57], discrete token [3,60], or feature [50] of masked pixels. This line of work has shown strong potential in learning representations on large-scale datasets and is less prone to model collapsing like instance-discrimination-based methods (e.g., contrastive learning). When combined with contrastive learning [2,18,53], the performance can be further boosted, yet with less dependence on data scale [18]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 328, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 510 + ], + "type": "text", + "content": "3D Scene understanding. The deep neural architectures for understanding 3D scenes can be roughly categorized into three paradigms according to the way they model point clouds: projection-based, voxel-based, and point-based. Projection-based works project 3D points into various image planes and adopt 2D CNN-based backbones to extract features [9, 30, 31, 43]. On the other hand, the voxel-based stream transforms point clouds into regular voxel representations to operate 3D convolutions [33, 42]. Their efficiency is then improved thanks to sparse convolution [10, 13, 19]. Point-based methods, in contrast, directly operate on the point cloud [12, 37, 38, 44, 58], and see a recent transition towards transformer-based architectures [21, 52, 59]. Following [56], we mainly pre-train on the voxel-based method SparseUNet [10] implemented with SpConv [15]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 521, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 548, + 714 + ], + "type": "text", + "content": "3D Representation learning. Unlike the 2D counterparts, where large-scale unsupervised pre-training has been a common choice for facilitating downstream tasks [6], 3D representation learning is still not mature, and most works still train from scratch on the target data directly [28]. While earlier works in 3D representation learning simply build on a single object [22, 40, 41, 49], recent works start to train on scene-centric point clouds [28, 56]. However, unlike in 2D that scene-centric representation learning has been well-studied [32, 51, 55], the pre-training on 3D scenes, which relies on raw frame data [28, 56], still faces inefficiency issues and finds it hard to scale up to larger scale datasets. In contrast, we explore directly learning at the scene level, which shows significantly higher efficiency in processing scene data, and opens the possibility for pre-training with larger-scale point clouds, for the first time ever." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "9416" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 293, + 177 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 293, + 177 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 293, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 293, + 177 + ], + "type": "image", + "image_path": "93f1b3604b750be5300d4a25bc343024ff70ac639363956ca61d69ab21ed4007.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 178, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 295, + 227 + ], + "type": "text", + "content": "(a) Frame matching [28, 56]. 1. Extract raw RGB-D frames and camera positions from raw data. 2. Project each 2D frame into 3D space produces frame-level point cloud views. 3. Calculate pairwise overlapping rates among each frame of a single view and select pairs with overlapping rates larger than " + }, + { + "bbox": [ + 47, + 178, + 295, + 227 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 47, + 178, + 295, + 227 + ], + "type": "text", + "content": " as pairs of contrastive views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 71, + 543, + 177 + ], + "blocks": [ + { + "bbox": [ + 304, + 71, + 543, + 177 + ], + "lines": [ + { + "bbox": [ + 304, + 71, + 543, + 177 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 543, + 177 + ], + "type": "image", + "image_path": "0a4c6dd5d370f437fa7b9a46213ca0038cfed07102698112ac2cceb9b743195e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 178, + 545, + 227 + ], + "lines": [ + { + "bbox": [ + 297, + 178, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 297, + 178, + 545, + 227 + ], + "type": "text", + "content": "(b) Scene augmentation (ours). 1. Apply spatial augmentations containing rotation, flipping, and scaling. 2. Apply photometric augmentations containing brightness, contrast, saturation, hue, and gaussian noise jittering. 3. Generate and apply contrastive cross masks to the two views after sampling augmentations containing cropping and voxelization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 234, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 546, + 289 + ], + "type": "text", + "content": "Figure 2. View generation. Compared with frame matching (FM), our scene augmentation (SA) is efficient and effective. 1. SA can end-to-end produce contrastive views on the original point cloud with ignorable latency, while FM requires preprocessing devouring enormous storage resources (e.g. additional 1.5 TB storage for ScanNet) in step 2 and pairwise matching is time-consuming. 2. SA produces scene-level views, while FM can only produce frame-level views containing limited information. Benefiting from advanced photometric augmentations, SA has the capacity to simulate the same scene under different lighting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 119, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 119, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 119, + 317 + ], + "type": "text", + "content": "3. Pilot Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 323, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 358 + ], + "type": "text", + "content": "This section analyzes the two main obstacles towards large-scale pre-training with point clouds. Our proposed design is based on the conclusion of the pilot study." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 359, + 231, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 231, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 231, + 370 + ], + "type": "text", + "content": "Is matching RGB-D frames a good choice?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "type": "text", + "content": "As a seminal work in 3D representation learning, Point-Contrast [56] first enables pre-training in real-world indoor scenes with matched raw RGB-D frames as contrastive views. A visualization of the frame matching procedure is illustrated in Figure 2a. This protocol seems natural for indoor scenes since point clouds of indoor scenes are usually derived from RGB-D videos [1,5,16], where the raw frames are extracted from. However, this framework has multiple drawbacks that can hinder the scalability of training:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 480, + 287, + 687 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 58, + 480, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 480, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 58, + 480, + 287, + 540 + ], + "type": "text", + "content": "- Redundant frame encoding. The pairwise matching strategy that PointContrast adopts allows one frame to be matched multiple times. As a result, each frame can be encoded multiple times in one step, adding to redundancy in training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 542, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 542, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 58, + 542, + 287, + 614 + ], + "type": "text", + "content": "- Low learning efficiency. In one training step, the frame matching strategy only allows the framework to process several views of a single scene. Therefore the amount of information that PointContrast can process in one step is rather limited, and the overall time for one training cycle is also notably high." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 616, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 616, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 58, + 616, + 287, + 687 + ], + "type": "text", + "content": "- Dependency on raw RGB-D frames. The whole framework is built on the assumption that RGB-D videos are available, yet this is not true for every publicly-available point cloud dataset. Even when available, the storage cost of RGBD frames is also significantly higher than the reconstructed point cloud data." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Consequently, pre-training frameworks [28, 56] based on matching frames as contrastive samples require enormous" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 304, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 400 + ], + "type": "text", + "content": "computing and storage resources. For example, Point-Contrast sub-samples RGB-D scans from the raw ScanNet videos every 25 frames, consuming " + }, + { + "bbox": [ + 304, + 304, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\sim 30" + }, + { + "bbox": [ + 304, + 304, + 545, + 400 + ], + "type": "text", + "content": " times the storage of the processed point clouds, and takes 80 GPU hours to process an epoch of 1500 scenes. Even at such cost, our experiments in Table 1a verify that the raw RGB-D frames cannot bring additional information over the processed point clouds to achieve a better representation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 408, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 433 + ], + "type": "text", + "content": "We will explore the possibility of pre-training on the point clouds directly." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 441, + 499, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 499, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 499, + 453 + ], + "type": "text", + "content": "What's the revelation behind mode collapse?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 453, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 681 + ], + "type": "text", + "content": "Mode collapse, defined as the phenomenon that all features collapse to a single vector, remains an unsolved problem accompanying the development of 3D representation learning [11, 56]. To alleviate this problem, PointContrast introduced InfoNCE loss [35], which has been shown to stabilize training, to replace the hardest-contrastive loss. Yet, the problem of mode collapse can still occur when the amount of training data and the length of the training schedule increase. Given the empirical conclusion of 2D contrastive learning [8], the occurrence of mode collapse is unusual under the premise that a large number of negative samples are already adopted. Interestingly, we notice that the mean negative pair cosine similarity of previous works is mostly close to 0, indicating that the negative samples are mostly easy and thus have little penalty towards the trivial solution. Although the InfoNCE loss alleviates this problem with an alternated optimization objective, we argue that a more desirable solution can be achieved by raising the difficulty of the unsupervised pretext task." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We will further raise the difficulty of the pretext task to solve the mode collapse problem." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9417" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 537, + 261 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 537, + 261 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 537, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 537, + 261 + ], + "type": "image", + "image_path": "e44cb99999a7badb233eeb7577b714c00c3aa521f877ed13ede5835b7ec4bd84.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 272, + 548, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 272, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 548, + 350 + ], + "type": "text", + "content": "Figure 3. Our MSC framework. (1) Generating a pair of contrastive views with a well-curated data augmentation pipeline consisting of photometric, spatial, and sampling augmentations. (2) Generating a pair of complementary masks and applying them to the pair of contrastive views. Replacing masked point features with a learnable mask token vector. (3) Extracting point representation with a given U-Net style backbone for point cloud understanding. (4) Reassembling masked contrastive views to masked points combination and unmasked points combination. (5) Matching points share similar positional relationships in the two views as positive sample pairs and computing InfoNCE loss to optimize contrastive objective. (6) Predicting masked point color and normal and computing Mean Squared Error loss and Cosine Similarity Loss with ground truth respectively, to optimize the reconstruction objective." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 369, + 115, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 115, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 115, + 384 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 390, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 289, + 510 + ], + "type": "text", + "content": "Based on the analysis of the stumbling blocks for large-scale pre-training in Section 3, we first introduce our optimized contrastive learning design in Section 4.1 to make the process more efficient. Then we solve the long-term problem of mode collapse with an additional reconstructive learning design in Section 4.2. The final optimization target is described in Section 4.3. Combining these exquisite designs, we build the whole framework, namely Masked Scene Contrast (MSC), and a visual illustration of our MSC is available in Figure 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 518, + 171, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 171, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 171, + 532 + ], + "type": "text", + "content": "4.1. Contrastive Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "type": "text", + "content": "**Framework.** Different from the previous protocol of matching RGB-D frames decomposed from indoor scenes, our contrastive learning framework directly operates on the point cloud data. Given a point cloud " + }, + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = (P, C)" + }, + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{n \\times 3}" + }, + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "type": "text", + "content": " represents the spatial features (coordinate) of the points and " + }, + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{n \\times 3}" + }, + { + "bbox": [ + 46, + 536, + 287, + 632 + ], + "type": "text", + "content": " represents the photometric features (color) of the points, the contrastive learning framework can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 638, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "type": "text", + "content": "- View generation. For a given point cloud " + }, + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "type": "text", + "content": ", we generate query view " + }, + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_r" + }, + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "type": "text", + "content": " and key view " + }, + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_k" + }, + { + "bbox": [ + 58, + 638, + 287, + 698 + ], + "type": "text", + "content": " of the original point cloud with a sequence of stochastic data augmentations, which includes photometric, spatial, and sampling augmentations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "content": "- Feature extraction. Encoding point cloud features " + }, + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_r" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "spans": [ + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "inline_equation", + "content": "\\pmb{F}_k" + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "text", + "content": " with a U-Net style backbone " + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "inline_equation", + "content": "\\zeta(\\cdot)" + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}_r" + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}_k" + }, + { + "bbox": [ + 324, + 369, + 544, + 394 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 400, + 547, + 538 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "text", + "content": "- Point matching. The positive samples of contrastive learning are point pairs with close spatial positions in the two views. For each point belonging to the query view, we calculate the correspondence mapping " + }, + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{(i,j)\\}_{n'}" + }, + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "text", + "content": " to points of the key view. If " + }, + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "inline_equation", + "content": "(i,j) \\in \\mathcal{P}" + }, + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "text", + "content": " then point " + }, + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "inline_equation", + "content": "(\\pmb{p}_i,\\pmb{c}_i)" + }, + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "text", + "content": " and point " + }, + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "inline_equation", + "content": "(\\pmb{p}_j,\\pmb{c}_j)" + }, + { + "bbox": [ + 317, + 400, + 545, + 483 + ], + "type": "text", + "content": " constructs a pair across two views." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "type": "text", + "content": "- Loss computation. Computing the contrastive learning loss on the representation of two views " + }, + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\hat{F}_r" + }, + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\hat{F}_k" + }, + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "type": "text", + "content": " and the correspondence mapping " + }, + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 317, + 491, + 547, + 538 + ], + "type": "text", + "content": ". An encoded query view should be similar to its key view." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Data augmentation. As a pioneering work in image contrastive learning, SimCLR [7] reveals that a well-curated data augmentation pipeline is crucial for learning strong representations. Unlike supervised learning, contrastive learning requires much stronger data augmentations to prevent trivial solutions. However, an effective data augmentation recipe is still absent in 3D representation learning. The frame matching scheme in prior works [28,56] simply applies a randomly rotating operator to contrastive targets. Even if the RGB-D frames can be viewed as a natural random crop, the augmentation space is still far from diverse enough. The pretext task it forms is not yet challenging enough to facilitate the contrastive learning framework to learn robust representations for downstream tasks." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9418" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 198 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 198 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 198 + ], + "type": "image", + "image_path": "4edc151802c93bc6091ae903bffbc9a0be3f2d3db9856aa4240262f90de67822.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 288, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 288, + 251 + ], + "type": "text", + "content": "Figure 4. View Mixing. Randomly mix up query views while keeping key views unmixed for a given batch of pairwise contrastive views. Detaching mixed query view after feature extraction for contrastive comparison with matched key views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": "As presented in Figure 2b, our well-designed stochastic data augmentation pipeline includes photometric augmentations, spatial augmentations, and sampling augmentations. Inspired by the advanced photometric augmentation validated by our 2D counterparts [7, 20], we further strengthen the photometric augmentation component introduced by Choy et al. [10] with random brightness, contrast, saturation, hue, and gaussian noise jittering for photometric augmentation. Besides that, random rotating, flipping, and scaling constitute our spatial augmentations, and the sampling augmentation is composed of random cropping and grid sampling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 428, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 428, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 428, + 287, + 513 + ], + "type": "text", + "content": "Empirically, the order of data augmentations is also a key component of our recipe. For example, grid sampling after random rotation leads to cross grids for sampling, which further increases the distinction between contrastive views and has a better augmentation effect. The specific data augmentation settings are available in the appendix, and a comparison with previous methods is presented in Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 517, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 649 + ], + "type": "text", + "content": "View mixing. Recently, Nekrasov et al. [34] proposes a data augmentation technique for 3D understanding models by mixing two scenes as a hybrid training sample, which can significantly suppress model overfitting. Inspired by the mixing mechanism, we integrate the logic of mixing as part of the contrastive learning objective. As illustrated in Figure 4, for a batch of pairwise views, we randomly mix up the query views while maintaining the key views unchanged before the feature extraction process. The simple operation can effectively increase the robustness of the backbone and improve the robustness of the point cloud representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "type": "text", + "content": "Contrastive target. We follow the design of PointContrast on the contrast target and apply InfoNCE loss to the matched points. Given correspondence mapping " + }, + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{(i,j)\\}_{n'}" + }, + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "type": "text", + "content": " produced by point matching and points representation " + }, + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}_r" + }, + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}_k" + }, + { + "bbox": [ + 47, + 652, + 288, + 715 + ], + "type": "text", + "content": " embedded during feature extraction, the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 73, + 382, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 382, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 382, + 83 + ], + "type": "text", + "content": "contrastive loss is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 370, + 86, + 545, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 86, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 370, + 86, + 545, + 117 + ], + "type": "interline_equation", + "content": "s _ {i j} = \\frac {\\boldsymbol {f} _ {i ^ {\\prime}} ^ {r T} \\boldsymbol {f} _ {j ^ {\\prime}} ^ {k}}{| | \\boldsymbol {f} _ {i ^ {\\prime}} ^ {r T} | | \\cdot | | \\boldsymbol {f} _ {j ^ {\\prime}} ^ {k} | |}, \\tag {1}", + "image_path": "9c933e2249922183d1b88170a7388fdd2d0777e26aa248e762f2d394e378b8a1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 349, + 119, + 545, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 119, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 349, + 119, + 545, + 150 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {I n f o N C E}} = \\sum_ {i} ^ {n} - \\log \\frac {\\exp \\left(s _ {i i} / \\tau\\right)}{\\sum_ {j} ^ {n} \\exp \\left(s _ {i j} / \\tau\\right)}, \\tag {2}", + "image_path": "ed7e118ae540dfb0d095a53ac703bccc86ebd71d8d0a4b7afdb10cf8c6419e5a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "type": "text", + "content": "note that " + }, + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "S = \\{s_{ij}\\} \\in \\mathbb{R}^{n\\times n}" + }, + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "type": "text", + "content": " is the pairwise cosine similarity matrix between positive samples and negative samples, while " + }, + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "type": "text", + "content": " is the temperature factor scaling cosine similarity. In practice, we control temperature factor " + }, + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 154, + 545, + 214 + ], + "type": "text", + "content": " as 0.4, which is the same as previous works [28,56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 220, + 444, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 444, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 444, + 232 + ], + "type": "text", + "content": "4.2. Reconstructive Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 357 + ], + "type": "text", + "content": "As is mentioned in Section 3, one of the stumbling blocks for large-scale representation is mode collapse, and our solution is to scale up the difficulty of the unsupervised pre-training task. Motivated by the success of masked image modeling [23, 57] in 2D representations, we propose masked point modeling, which can be naturally integrated into our contrastive learning framework. Benefiting from this design, our framework can fully use non-overlapped regions of contrastive views that cannot be utilized by contrastive learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "type": "text", + "content": "Contrastive cross mask. The key design that enables additional construction learning in our contrastive learning framework is the contrastive cross mask. For a given query view and key view of a single point cloud, we partite the unioned point set into non-overlapping grid partitions by their original position before spatial augmentation. Given a mask rate " + }, + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "type": "text", + "content": " range from 0 to 0.5, we randomly generate a pair of masks " + }, + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "type": "inline_equation", + "content": "M_r, M_k \\in \\mathbb{R}^{1 \\times n_{r,k}}" + }, + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "type": "text", + "content": ", in which there are no shared masked patches. Then, we follow the practice of SimMIM [57] to apply the pair of masks to the two views respectively by replacing the input feature with a learnable mask token vector " + }, + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 304, + 358, + 546, + 513 + ], + "type": "text", + "content": ". Consequently, the feature extraction process can be rewritten as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 342, + 517, + 545, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 517, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 342, + 517, + 545, + 532 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {F}} _ {r, k} = \\zeta \\left(\\left(1 - \\boldsymbol {M} _ {r, k}\\right) \\boldsymbol {F} _ {r, k} + \\boldsymbol {M} _ {r, k} \\boldsymbol {T} _ {r, k}\\right), \\tag {3}", + "image_path": "e7f6f80e2283f64d5025815d38d94d58bf70c931d2d097ea55100e3985901e95.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 536, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 536, + 545, + 559 + ], + "type": "inline_equation", + "content": "T_{r,k} \\in \\mathbb{R}^{n_r, n_k \\times c}" + }, + { + "bbox": [ + 304, + 536, + 545, + 559 + ], + "type": "text", + "content": " is the expand matrix of mask token vector " + }, + { + "bbox": [ + 304, + 536, + 545, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 536, + 545, + 559 + ], + "type": "text", + "content": " to fit the feature dimensions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 560, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 545, + 620 + ], + "type": "text", + "content": "Reconstruction target. The features of the point cloud are composed of two parts, the coordinates that determine the geometric structure and the colors that represent the texture features. We build up reconstruction targets for the two groups of features separately." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 620, + 545, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 545, + 679 + ], + "type": "text", + "content": "The reconstruction of point cloud texture is straightforward, we predict the photometric value of each point with a linear projection. We compute the mean squared error (MSE) between the reconstructed and original color of masked points as the color reconstruction loss:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 683, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 683, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 683, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} = \\frac {\\sum_ {i} ^ {n _ {r}} m _ {i} ^ {r} | | \\boldsymbol {x} _ {i} ^ {r} - \\hat {\\boldsymbol {x}} _ {i} ^ {r} | | _ {2} ^ {2} + \\sum_ {i} ^ {n _ {k}} m _ {i} ^ {k} | | \\boldsymbol {x} _ {i} ^ {k} - \\hat {\\boldsymbol {x}} _ {i} ^ {k} | | _ {2} ^ {2}}{n _ {r} ^ {\\prime} + n _ {k} ^ {\\prime}}, \\tag {4}", + "image_path": "ed1442bcd89b63e44ce44e72c755a0d96608b14c2b0d359df086a4290b69bdbc.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9419" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "n_r'" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "n_k'" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " represent the number of mask points belonging to refer view and key view, " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "m_r^i" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "m_k^i" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " mean the " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "M_k" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 275 + ], + "type": "text", + "content": "Point coordinates play an important role in describing the geometric structure of point clouds, and it is worth noting that directly reconstructing the coordinates of masked points is not reasonable since masked points are only sampled from 3D object surface rather than the continuous surface itself. Reconstructing points coordination would lead to an overfitted representation. To overcome the challenge, we introduce the concept of surfel reconstruction. Surfel is an abbreviation for a surface element or surface voxel in the discrete topology literature [26] and primitives rendering [36]. For each masked point, we reconstruct the normal vector of the corresponding surfel and compute the mean cosine similarity between estimations and surfel normals as a contrastive loss:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 280, + 287, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 280, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 81, + 280, + 287, + 309 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {n}} = \\frac {\\sum_ {i} ^ {n _ {r}} m _ {i} ^ {r} \\boldsymbol {x} _ {i} ^ {r T} \\hat {\\boldsymbol {x}} _ {i} ^ {r} + \\sum_ {i} ^ {n _ {k}} m _ {i} ^ {k} \\boldsymbol {x} _ {i} ^ {k T} \\hat {\\boldsymbol {x}} _ {i} ^ {k}}{n _ {r} ^ {\\prime} + n _ {k} ^ {\\prime}}, \\tag {5}", + "image_path": "ad1d43972ebad60ceb227aa81f783bc076fa3b0eed7b32d992652c451dc3135c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "inline_equation", + "content": "n_r'" + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "inline_equation", + "content": "n_k'" + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "content": " represent the number of mask points belonging to refer view and key view, " + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "inline_equation", + "content": "m_r^i" + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "inline_equation", + "content": "m_k^i" + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "content": " mean the " + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "inline_equation", + "content": "M_k" + }, + { + "bbox": [ + 47, + 316, + 288, + 353 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 358, + 136, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 136, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 136, + 370 + ], + "type": "text", + "content": "4.3. Loss Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 376, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 437 + ], + "type": "text", + "content": "Our framework combines the contrastive target, the color reconstruction target, and the surfel reconstruction to make the unsupervised task more scalable. The overall loss function is a weighted sum of Eq. 2, Eq. 4, and Eq. 5 which is written as follow:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 445, + 287, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 445, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 94, + 445, + 287, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\lambda_ {c} \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {n} \\mathcal {L} _ {\\mathrm {n}}, \\tag {6}", + "image_path": "0e60175a585b4c11fb3625df2137978d4c149d9d62cfd9c9cd88a4edb97db0e9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\lambda_{n}" + }, + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "type": "text", + "content": " are the weight parameters that balance the three loss components. Empirically we find that performance is robust to the choice of weight parameters, and we make " + }, + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\lambda_{c} = \\lambda_{n} = 1" + }, + { + "bbox": [ + 47, + 466, + 287, + 515 + ], + "type": "text", + "content": " in practice." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "content": "We conduct extensive experimental evaluations to validate the capability of our framework, built upon the point cloud perception codebase Pointcept [14]. We first ablate our designs with an efficient pre-training pipeline that only utilizes ScanNet point cloud in Section 5.1, while without compromising performance. Then we explore large-scale pre-training across multiple datasets and compare our performance with previous results in Section 5.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 647, + 146, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 146, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 146, + 660 + ], + "type": "text", + "content": "5.1. Main Properties" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "We ablate the main designs and intriguing properties of our MSC in Table 1, and the default setting is available in the caption. We enable efficient pre-training by introducing our view generation pipeline, which is ablated in Table 1a." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 544, + 255 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 544, + 255 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 544, + 255 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 544, + 255 + ], + "type": "image", + "image_path": "fd17a2c7fb786d0e34c0355be1dff800a9141f21a1531f19c287ea93361b948f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "type": "text", + "content": "Figure 5. Masked scenes and color reconstructions. We visualize one of the cross masks of each scene with a mask rate of " + }, + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "type": "text", + "content": " (left) and color reconstruction of the masked point combinations (right). We pre-train our MSC with a mask patch size of " + }, + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 266, + 545, + 344 + ], + "type": "text", + "content": " and generalize to results to different mask sizes. Compared with the original point clouds, the loss of detail cannot be avoided, while boundary and texture are well preserved by our model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 365, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 401 + ], + "type": "text", + "content": "All of our ablation experiments only require 20G ScanNet point cloud data and around 14 hours pertaining on a single machine containing 8 NVIDIA RTX3090." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 402, + 546, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 557 + ], + "type": "text", + "content": "View generation. In Table 1a, we show the results with different generation strategies of contrastive views. Unlike PointContrast [56], which builds on the raw RGB-D frames, our strategy directly utilizes the scene-level point cloud. This significantly reduces storage requirements (96% less) and allows more efficient use of the training data. With a 30% equivalent number of training iterations, our method can attain " + }, + { + "bbox": [ + 304, + 402, + 546, + 557 + ], + "type": "inline_equation", + "content": "120 \\times" + }, + { + "bbox": [ + 304, + 402, + 546, + 557 + ], + "type": "text", + "content": " training epochs, making full use of the training data for the first time. This results in 0.4 points higher FT mIoU with " + }, + { + "bbox": [ + 304, + 402, + 546, + 557 + ], + "type": "inline_equation", + "content": "4.4 \\times" + }, + { + "bbox": [ + 304, + 402, + 546, + 557 + ], + "type": "text", + "content": " speedup. When combined with an additional mask point modeling strategy, the performance can be further boosted by 0.6 points, yet still with a notable speedup." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 665 + ], + "type": "text", + "content": "Number of positive pairs. In Table 1b, we show the effects of different numbers of positive pairs. Our method sees consistent improvements with an increasing number of positive pairs, while for PointContrast [56], the information in a large batch of positive pairs cannot be effectively utilized. Our intuition is that the ability to process scene-level views, which are larger in scale and contain much more information than frame-level views, enables our method to take advantage of a larger number of positive pairs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Data augmentation. In Table 1c, we analyze the effect of different data augmentation combinations. Adopting either spatial augmentation or photometric augmentation leads to sub-optimal performance, and the combination of" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9420" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 71, + 533, + 121 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 533, + 121 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 533, + 121 + ], + "type": "table", + "html": "
View generation methodsPre-train dataStorageBatch sizeItersEpochsFT mIoU (%)Hours (h)Speedup
Frame matching (PointContrast [56])ScanNet Raw500G32100k574.0481.0×
Scene augmentation w/o mask (ours)ScanNet v220G3230k60074.4114.4×
Scene augmentation w mask (ours)ScanNet v220G3230k60075.0143.4×
", + "image_path": "85ccc8006b3f6009a22005288a2f477a94952cf1431453762e93cb6f9e77bfd4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 48, + 124, + 541, + 143 + ], + "lines": [ + { + "bbox": [ + 48, + 124, + 541, + 143 + ], + "spans": [ + { + "bbox": [ + 48, + 124, + 541, + 143 + ], + "type": "text", + "content": "(a) View generation. Views produced by our enhanced data augmentation are stronger than the original RGB-D frames. Scene-level views can significantly speed up pre-training and make contrastive learning more effective. The performance can be further boosted with additional masked point modeling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 146, + 193, + 207 + ], + "blocks": [ + { + "bbox": [ + 72, + 146, + 193, + 207 + ], + "lines": [ + { + "bbox": [ + 72, + 146, + 193, + 207 + ], + "spans": [ + { + "bbox": [ + 72, + 146, + 193, + 207 + ], + "type": "table", + "html": "
#Pos pairsPC [56]MSC (ours)
102473.874.3
204874.074.5
409873.774.9
819273.975.0
", + "image_path": "548ba6a456ae41fa0bf5e891fe2b5353f6eae18878c89b5cd14d8a4916f9e765.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 209, + 206, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 206, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 206, + 237 + ], + "type": "text", + "content": "(b) Number of positive pairs. A larger amount of sampled positive pairs are necessary for scene-level views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 62, + 242, + 204, + 304 + ], + "blocks": [ + { + "bbox": [ + 62, + 242, + 204, + 304 + ], + "lines": [ + { + "bbox": [ + 62, + 242, + 204, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 242, + 204, + 304 + ], + "type": "table", + "html": "
MaskTaskFT mIoU (%)
w/o crossw/o contrast74.1
w crossw/o contrast74.4
w/o crossw contrast74.7
w crossw contrast75.0
", + "image_path": "e7f4fb5b65c6bc4975239c02bd9c4719a810d2d924a61c8376d45e31f57a9024.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 305, + 206, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 305, + 206, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 206, + 335 + ], + "type": "text", + "content": "(e) Cross mask. Masks containing shared masked patches have a negative influence on contrastive learning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 230, + 146, + 365, + 207 + ], + "blocks": [ + { + "bbox": [ + 230, + 146, + 365, + 207 + ], + "lines": [ + { + "bbox": [ + 230, + 146, + 365, + 207 + ], + "spans": [ + { + "bbox": [ + 230, + 146, + 365, + 207 + ], + "type": "table", + "html": "
SpatialPhotometricFT mIoU (%)
w/o augw/o aug72.1
w augw/o aug73.4
w/o augw aug72.8
w augw aug74.4
", + "image_path": "5b733a3f17ea30a41c535faa4a43d69fb910787de1eace168c4adfa374d8095c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 220, + 209, + 372, + 238 + ], + "lines": [ + { + "bbox": [ + 220, + 209, + 372, + 238 + ], + "spans": [ + { + "bbox": [ + 220, + 209, + 372, + 238 + ], + "type": "text", + "content": "(c) Data augmentation. The combination of spatial and photometric augmentation makes the view generation pipeline comes to work." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 236, + 242, + 360, + 304 + ], + "blocks": [ + { + "bbox": [ + 236, + 242, + 360, + 304 + ], + "lines": [ + { + "bbox": [ + 236, + 242, + 360, + 304 + ], + "spans": [ + { + "bbox": [ + 236, + 242, + 360, + 304 + ], + "type": "table", + "html": "
Mask grid size (m)FT mIoU (%)
0.0574.3
0.175.0
0.1575.0
0.274.8
", + "image_path": "28d148d3911b458d76e05fbb9755024f1fda780e60d71c72eebc40eb4eff62d6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 220, + 305, + 372, + 335 + ], + "lines": [ + { + "bbox": [ + 220, + 305, + 372, + 335 + ], + "spans": [ + { + "bbox": [ + 220, + 305, + 372, + 335 + ], + "type": "text", + "content": "(f) Mask grid size. Our design works with mask patches with a grid size larger than " + }, + { + "bbox": [ + 220, + 305, + 372, + 335 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 220, + 305, + 372, + 335 + ], + "type": "text", + "content": ", and we consider " + }, + { + "bbox": [ + 220, + 305, + 372, + 335 + ], + "type": "inline_equation", + "content": "0.15\\mathrm{m}" + }, + { + "bbox": [ + 220, + 305, + 372, + 335 + ], + "type": "text", + "content": " as a default setting." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 394, + 146, + 533, + 207 + ], + "blocks": [ + { + "bbox": [ + 394, + 146, + 533, + 207 + ], + "lines": [ + { + "bbox": [ + 394, + 146, + 533, + 207 + ], + "spans": [ + { + "bbox": [ + 394, + 146, + 533, + 207 + ], + "type": "table", + "html": "
Query viewKey viewFT mIoU (%)
w/o mixw/o mix74.1
w mixw/o mix74.4
w/o mixw mix74.2
w mixw mix73.7
", + "image_path": "49f699e03d822bf7f3235449bca4d6612f46deedde691f2b45b48248a2aacdf8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 386, + 209, + 537, + 237 + ], + "lines": [ + { + "bbox": [ + 386, + 209, + 537, + 237 + ], + "spans": [ + { + "bbox": [ + 386, + 209, + 537, + 237 + ], + "type": "text", + "content": "(d) View mixing. Randomly mixing query views while leaving key views unmixed is a sweet point." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 406, + 242, + 520, + 304 + ], + "blocks": [ + { + "bbox": [ + 406, + 242, + 520, + 304 + ], + "lines": [ + { + "bbox": [ + 406, + 242, + 520, + 304 + ], + "spans": [ + { + "bbox": [ + 406, + 242, + 520, + 304 + ], + "type": "table", + "html": "
ColorNormalFT mIoU (%)
w/ow/o74.4
ww/o74.9
w/ow74.6
ww75.0
", + "image_path": "2f3dad773b42abe65f78111debc55cf58b79c93115605efe8afa9ec7b14beab9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 386, + 305, + 537, + 335 + ], + "lines": [ + { + "bbox": [ + 386, + 305, + 537, + 335 + ], + "spans": [ + { + "bbox": [ + 386, + 305, + 537, + 335 + ], + "type": "text", + "content": "(g) Reconstruction target. Both targets have a positive effect, while color reconstruction has a dominant impact on indoor scenes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 338, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 545, + 383 + ], + "type": "text", + "content": "Table 1. Ablation experiments. We adopt SparseUNet and efficient pre-training on ScanNet [16] point cloud data to ablate our designs. We report fine-turning (FT) mIoU (%) results on ScanNet 20 classes semantic segmentation as the default metric. If not specified, the default setting is as follows: the pre-training period is 600 epochs, the masking ratio is " + }, + { + "bbox": [ + 46, + 338, + 545, + 383 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 338, + 545, + 383 + ], + "type": "text", + "content": " and the masked patch has a grid size of " + }, + { + "bbox": [ + 46, + 338, + 545, + 383 + ], + "type": "inline_equation", + "content": "0.15\\mathrm{m}" + }, + { + "bbox": [ + 46, + 338, + 545, + 383 + ], + "type": "text", + "content": " in the real-world space, view mixing probability is 0.8. All of our designs are enabled by default. Default settings are marked in gray." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 399, + 266, + 487 + ], + "blocks": [ + { + "bbox": [ + 58, + 399, + 266, + 487 + ], + "lines": [ + { + "bbox": [ + 58, + 399, + 266, + 487 + ], + "spans": [ + { + "bbox": [ + 58, + 399, + 266, + 487 + ], + "type": "image", + "image_path": "6cab827345a95d521ba2da5e5633418f8e425c42492928a9b20b20a01886af92.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "type": "text", + "content": "Figure 6. Masking ratio. A masking ratio ranging from " + }, + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "type": "text", + "content": " works well with our design, and a higher mask rate negatively influences contrastive learning. The y-axes represent ScanNet semantic segmentation validation mIoU " + }, + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 494, + 287, + 538 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 544, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 286, + 604 + ], + "type": "text", + "content": "both helps make our view generation pipeline come to work. Concerning relative significance, since the spatial augmentation is highly coupled with the masking strategy, the gain from it is higher than the photometric augmentation. But still, both are necessary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "type": "text", + "content": "View mixing. In Table 1d, we explore different config for the view mixing strategy. Randomly mixing the query views while leaving the key views unchanged yields the best performance. Our intuition is that the key views, which form the vocabulary of the InfoNCE loss, should be relatively stable, thus reducing the ambiguity of the learning target. In contrast, introducing more diversity to the queries can be more helpful." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "Cross mask. In Table 1e, we study the cross mask strategy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "type": "text", + "content": "This strategy ensures that the two augmented views have no overlapping tokens. As reported in the table, whether with or without the contrastive learning target, this strategy ensures fewer shortcuts to the task and enables consistent downstream improvements." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "text", + "content": "Mask grid size. In Table 1f, we show the results ablating the grid size for producing masks on point clouds. Our design works with a grid size larger than " + }, + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "text", + "content": ", and we consider " + }, + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "inline_equation", + "content": "0.15\\mathrm{m}" + }, + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "text", + "content": " as a default setting. As shown in Figure 5, our reconstruction module is robust to extending mask grid size, which indicates that high guilty representation is captured during pre-training." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "content": "Masking ratio. In Figure 6, we depict the effect of the ratio of masked tokens. A masking ratio ranging from " + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "content": " works well with our design, and a higher masking rate has a negative impact on the overall performance. This varies from the conclusion of MAE [23], in which a higher masking ratio of " + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "content": " achieves top performance. Our hypothesis is that the contrastive learning objective built on the masked point clouds might favor a lower masking ratio, and the results in Figure 6 reflect a trade-off between the contrastive learning objective and the reconstructive objective. And pure reconstructive learning on point clouds is an interesting direction for future explorations." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Reconstruction target. In Table 1g, we ablate the effects of two components of our reconstruction target: color re" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9421" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 279, + 126 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 279, + 126 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 279, + 126 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 279, + 126 + ], + "type": "table", + "html": "
DatasetsBackbonesSemantic Seg. (mIoU)
SCPC [56]CSC [28]MSC (ours)
ScanNetSparseUNet72.274.1 (+1.9)73.8 (+1.6)75.5 (+3.3)
ScanNet200SparseUNet25.026.2 (+1.2)26.4 (+1.4)28.8 (+3.8)
", + "image_path": "c176e9657cf6fd5bba6dcb954ef85bbc23176a4f132e68ae49202482ad03b50d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 160, + 280, + 217 + ], + "blocks": [ + { + "bbox": [ + 51, + 128, + 284, + 157 + ], + "lines": [ + { + "bbox": [ + 51, + 128, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 51, + 128, + 284, + 157 + ], + "type": "text", + "content": "(a) Semantic segmentation. We conduct pre-training on SparseUNet and compare semantic segmentation mIoU " + }, + { + "bbox": [ + 51, + 128, + 284, + 157 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 51, + 128, + 284, + 157 + ], + "type": "text", + "content": " results on ScanNet and ScanNet200 [39] validation set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 160, + 280, + 217 + ], + "lines": [ + { + "bbox": [ + 58, + 160, + 280, + 217 + ], + "spans": [ + { + "bbox": [ + 58, + 160, + 280, + 217 + ], + "type": "table", + "html": "
DatasetsBackbonesInstance Seg. (mAP@0.5)
SCPC [56]CSC [28]MSC (ours)
ScanNetSparseUNet56.958.0 (+1.1)59.4 (+2.5)59.6 (+2.7)
ScanNet200SparseUNet24.524.9 (+0.4)25.2 (+0.7)26.8 (+2.3)
", + "image_path": "53ce0712230a5a21f9b9ac487cf0b46de045648358901ff15246827caf7e529b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 251, + 282, + 337 + ], + "blocks": [ + { + "bbox": [ + 51, + 219, + 284, + 248 + ], + "lines": [ + { + "bbox": [ + 51, + 219, + 284, + 248 + ], + "spans": [ + { + "bbox": [ + 51, + 219, + 284, + 248 + ], + "type": "text", + "content": "(b) Instance segmentation. We conduct pre-training on SparseUNet and compare instance segmentation mAP@0.5 " + }, + { + "bbox": [ + 51, + 219, + 284, + 248 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 51, + 219, + 284, + 248 + ], + "type": "text", + "content": " results driven by PointGroup [29] on ScanNet and ScanNet200 [39] validation set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 251, + 282, + 337 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 282, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 282, + 337 + ], + "type": "table", + "html": "
LRSemantic Seg.LASemantic Seg.
Pct.SCCSCVIBUSMSCPts.SCCSCVIBUSMSC
100%72.273.8-75.3Full72.273.8-75.3
1%26.028.928.629.22041.955.561.061.2
5%47.849.847.450.75053.960.565.666.8
10%56.759.460.561.010062.265.968.969.7
20%62.964.664.864.920065.568.269.670.7
", + "image_path": "1757c0a1d0c54c311689652155cc0a1a8282a22f0b11694e6ad8e304fbbc8960.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 338, + 283, + 358 + ], + "lines": [ + { + "bbox": [ + 51, + 338, + 283, + 358 + ], + "spans": [ + { + "bbox": [ + 51, + 338, + 283, + 358 + ], + "type": "text", + "content": "(c) Data efficiency. We follow the ScanNet Data Efficient benchmark and compare the validation results SparseUNet with previous methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 361, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 417 + ], + "type": "text", + "content": "Table 2. Results comparison. We adopt cross-dataset pre-training utilizing ScanNet and ArkitScenes point cloud scenes for comparison of downstream task results. The pre-training setting is fixed as the default described in Table 1. More specific pre-training details are available in the Appendix. SC denotes train from scratch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 424, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 509 + ], + "type": "text", + "content": "construction and normal reconstruction. Given the premise that indoor scenes are used, both targets show a positive effect on overall performance, while color reconstruction has a higher impact. The intuition is that the difference in texture reflected by color has a higher impact on the task of semantic segmentation, while the normal is helpful but has less influence (consider the same task on a 2D image)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 515, + 165, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 165, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 165, + 529 + ], + "type": "text", + "content": "5.2. Results Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": "In this section, we extend the scale of pre-training by merging multiple datasets and compare downstream task fine-turning performance with previous unsupervised pretraining frameworks [28, 56]. Specifically, we adopt the default model setting ablated in Section 5.1 and train on both ScanNet [16] and ArkitScenes [5] point clouds, extending pre-training assets from 1,513 scenes to 6,560 scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Semantic segmentation. In Table 2a, we report the semantic segmentation results on ScanNet and ScanNet200 [39] benchmark with SparseUNet and compare them with previous results. Our improvements are consistent and significant with larger pre-training assets for both benchmarks. With SparseUNet backbone, we outperform the current state-of-art pretraining framework by 1.7 points on ScanNet and 2.4 points on ScanNet200. Meanwhile, it is worth" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "noting that driven by powerful MSC, we set a new best validation result on ScanNet semantic segmentation and pushed the previous SOTA to " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "75.5\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " with a baseline model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 108, + 546, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 228 + ], + "type": "text", + "content": "Instance segmentation. In Table 2b, we report the instance segmentation results on ScanNet and ScanNet200 [39] with SparseUNet backbone driven by PointGroup [29]. Comparing them with previous results, we still see consistent improvements. Specifically, our framework achieves " + }, + { + "bbox": [ + 304, + 108, + 546, + 228 + ], + "type": "inline_equation", + "content": "59.6\\%" + }, + { + "bbox": [ + 304, + 108, + 546, + 228 + ], + "type": "text", + "content": " on the ScanNet validation set, which is 3.3 points higher than training from scratch and 0.2 points promotion compared with the previous state-of-art performance. The boost is more significant on ScanNet200, which is 1.6 points higher than the previous SOTA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 228, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 277 + ], + "type": "text", + "content": "Data efficiency. In Table 2c, we compare the ScanNet Data Efficient [28] results with previous methods. Our MSC shows consistently superior performance even compared with the latest data-efficient learning framework [45]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 289, + 458, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 458, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 458, + 301 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 308, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 416 + ], + "type": "text", + "content": "In this paper, we tackle the problem of scalable unsupervised 3D representation learning. To this end, we present Masked Scene Contrast (MSC), an efficient, effective, and scalable framework that directly operates on scene-level views with contrastive learning and masked point modeling. Benefiting from the efficient scene-level point cloud processing pipeline and the effective training objectives, our method harvests high efficiency and superior generality, and enables large-scale pre-training across multiple datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 417, + 546, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 546, + 548 + ], + "type": "text", + "content": "The key factor that empowers our method's scalability to larger-scale pre-training lies in the efficient pipeline that can directly learn from point cloud data, rather than the raw RGB-D frames. The efficiency, however, does not only mean processing data at scale. When only the standard dataset ScanNet is used for pre-training, our method still achieves uncompromised performance, yet with at least " + }, + { + "bbox": [ + 304, + 417, + 546, + 548 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 417, + 546, + 548 + ], + "type": "text", + "content": " speedup over previous works. This is especially meaningful considering the exhaustively long experimental time in the pre-training community, and it can better facilitate the verification of new ideas in future works." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 548, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 657 + ], + "type": "text", + "content": "It should also be noted that given the limit in computing resources, and the absence of a pre-training dataset that is sustainably at scale, the scalability of our method is not fully presented. In other words, our method opens the possibility of large-scale pre-training on 3D point cloud data for the first time, and we call for a large-enough 3D scene dataset that can fully unleash this potential. We hope our method can inspire future works that take the first step to real large-scale 3D pre-training, as the 2D community does." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 663, + 409, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 663, + 409, + 676 + ], + "spans": [ + { + "bbox": [ + 306, + 663, + 409, + 676 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China (No. 62201484), HKU Startup Fund, and HKU Seed Fund for Basic Research." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9422" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 167 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael G. Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 286, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 286, + 191 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In ICLR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 286, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 286, + 224 + ], + "type": "text", + "content": "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 286, + 290 + ], + "type": "text", + "content": "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurIPS Workshops, 2021. 2, 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 286, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 286, + 334 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 286, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 368 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 286, + 390 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "type": "text", + "content": "[9] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 424, + 286, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 424, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 424, + 286, + 456 + ], + "type": "text", + "content": "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 457, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 286, + 479 + ], + "type": "text", + "content": "[11] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 480, + 286, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 286, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 286, + 513 + ], + "type": "text", + "content": "[12] Ruihang Chu, Yukang Chen, Tao Kong, Lu Qi, and Lei Li. Icm-3d: Instantiated category modeling for 3d instance segmentation. RA-L, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 514, + 286, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 556 + ], + "type": "text", + "content": "[13] Ruihang Chu, Xiaoqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, and Jiaya Jia. Twist: Two-way inter-label self-training for semi-supervised 3d instance segmentation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 558, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 590 + ], + "type": "text", + "content": "[14] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 591, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 286, + 624 + ], + "type": "text", + "content": "[15] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 624, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 668 + ], + "type": "text", + "content": "[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with exemplar convolutional neural networks. In TPAMI, 2015. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[18] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herve Jegou, and Edouard Grave. Are largescale datasets necessary for self-supervised pre-training? arXiv:2112.10740, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[19] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "type": "text", + "content": "[21] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 243, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 266 + ], + "type": "text", + "content": "[22] Kaveh Hassani and Mike Haley. Unsupervised multi-task feature learning on point clouds. In ICCV, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "type": "text", + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 301, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 301, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 301, + 545, + 334 + ], + "type": "text", + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 335, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 335, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 545, + 380 + ], + "type": "text", + "content": "[25] Olivier J Henaff, Aravind Srinivas, Jeffrey De Fauw, Ali Razavi, Carl Doersch, SM Eslami, and Aaron van den Oord. Data-efficient image recognition with contrastive predictive coding. In ICML, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 381, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 403 + ], + "type": "text", + "content": "[26] Gabor T Herman. Discrete multidimensional jordan surfaces. CVGIP, 1992. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 448 + ], + "type": "text", + "content": "[27] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 449, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 483 + ], + "type": "text", + "content": "[28] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "type": "text", + "content": "[29] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. CVPR, 2020. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 518, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 518, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 518, + 545, + 552 + ], + "type": "text", + "content": "[30] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 553, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 553, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 553, + 545, + 575 + ], + "type": "text", + "content": "[31] Bo Li, Tianlei Zhang, and Tian Xia. Vehicle detection from 3d lidar using fully convolutional network. In RSS, 2016. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "type": "text", + "content": "[32] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv:2011.13677, 2020. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "type": "text", + "content": "[33] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 645, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 645, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 645, + 545, + 678 + ], + "type": "text", + "content": "[34] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Augmentation for 3D Scenes. In 3DV, 2021. 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 3" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9423" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 105 + ], + "type": "text", + "content": "[36] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In SIGGRAPH, 2000. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 139 + ], + "type": "text", + "content": "[37] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 173 + ], + "type": "text", + "content": "[38] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 286, + 206 + ], + "type": "text", + "content": "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 209, + 286, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 286, + 240 + ], + "type": "text", + "content": "[40] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 243, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 286, + 274 + ], + "type": "text", + "content": "[41] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "text", + "content": "[42] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 310, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 286, + 342 + ], + "type": "text", + "content": "[43] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik G. Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 343, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 286, + 387 + ], + "type": "text", + "content": "[44] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 388, + 286, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 421 + ], + "type": "text", + "content": "[45] Beiwen Tian, Liyi Luo, Hao Zhao, and Guyue Zhou. Vibus: Data-efficient 3d scene parsing with viewpoint bottleneck and uncertainty-spectrum modeling. ISPRS, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 422, + 286, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 286, + 454 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 286, + 454 + ], + "type": "text", + "content": "[46] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 456, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 286, + 487 + ], + "type": "text", + "content": "[47] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 490, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 490, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 490, + 286, + 533 + ], + "type": "text", + "content": "[48] Pascal Vincent, H. Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 2010. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 535, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 535, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 535, + 286, + 566 + ], + "type": "text", + "content": "[49] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 568, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 286, + 610 + ], + "type": "text", + "content": "[50] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 613, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 286, + 644 + ], + "type": "text", + "content": "[51] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 647, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 286, + 679 + ], + "type": "text", + "content": "[52] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "text", + "content": "[53] Zhirong Wu, Zihang Lai, Xiao Sun, and Stephen Lin. Extreme masking for learning instance and distributed visual representations. arXiv:2206.04667, 2022. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 319 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[54] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[55] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[56] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "text", + "content": "[57] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "type": "text", + "content": "[58] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 264, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 285 + ], + "type": "text", + "content": "[59] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "type": "text", + "content": "[60] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9424" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_content_list.json b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f8fede9d83eb7bfa189bfb8e602b178f6a41648 --- /dev/null +++ b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_content_list.json @@ -0,0 +1,1315 @@ +[ + { + "type": "text", + "text": "Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning", + "text_level": 1, + "bbox": [ + 137, + 128, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Wang $^{1,2}$ Dongdong Chen $^{3}$ Zuxuan Wu $^{1,2\\dagger}$ Yinpeng Chen $^{3}$ Xiyang Dai $^{3}$ Mengchen Liu $^{3}$ Lu Yuan $^{3}$ Yu-Gang Jiang $^{1,2\\dagger}$ $^{1}$ Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University \n $^{2}$ Shanghai Collaborative Innovation Center of Intelligent Visual Computing \n $^{3}$ Microsoft Cloud + AI", + "bbox": [ + 150, + 200, + 816, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from masked visual modeling, self-supervised video representation learning has achieved remarkable progress. However, existing methods focus on learning representations from scratch through reconstructing low-level features like raw pixel values. In this paper, we propose masked video distillation (MVD), a simple yet effective two-stage masked feature modeling framework for video representation learning: firstly we pretrain an image (or video) model by recovering low-level features of masked patches, then we use the resulting features as targets for masked feature modeling. For the choice of teacher models, we observe that students taught by video teachers perform better on temporally-heavy video tasks, while image teachers transfer stronger spatial representations for spatially-heavy video tasks. Visualization analysis also indicates different teachers produce different learned patterns for students. To leverage the advantage of different teachers, we design a spatial-temporal co-teaching method for MVD. Specifically, we distill student models from both video teachers and image teachers by masked feature modeling. Extensive experimental results demonstrate that video transformers pretrained with spatial-temporal co-teaching outperform models distilled with a single teacher on a multitude of video datasets. Our MVD with vanilla ViT achieves state-of-the-art performance compared with previous methods on several challenging video downstream tasks. For example, with the ViT-Large model, our MVD achieves $86.4\\%$ and $76.7\\%$ Top-1 accuracy on Kinetics-400 and Something-Something-v2, outperforming VideoMAE by $1.2\\%$ and $2.4\\%$ respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with $77.3\\%$ Top-1 accuracy on Something-Something-v2. Code will be available at https://github.com/ruiwang2021/mvd.", + "bbox": [ + 75, + 358, + 472, + 859 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e64338a607bb510d4dd0847dfea409c7f50ef859da431d2a17f508f233496e57.jpg", + "image_caption": [ + "Figure 1. Comparisons of MVD with previous supervised or self-supervised methods on Something-Something v2. Each line represents the corresponding model of different sizes." + ], + "image_footnote": [], + "bbox": [ + 501, + 324, + 887, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 601, + 632, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For self-supervised visual representation learning, recent masked image modeling (MIM) methods like MAE [31] and BEiT [2] achieve promising results with vision transformers [17] on various vision downstream tasks. Such a pretraining paradigm has also been adapted to the video domain and boosts video transformers by clear margins compared with supervised pretraining on several video downstream tasks. Representative masked video modeling (MVM) works include BEVT [63], VideoMAE [57] and ST-MAE [21].", + "bbox": [ + 496, + 627, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following MAE [31] and BEiT [2], existing masked video modeling methods [21, 57, 63] pretrain video transformers through reconstructing low-level features, e.g., raw pixel values or low-level VQVAE tokens. However, using low-level features as reconstruction targets often incur much noise. And due to the high redundancy in video data, it is easy for masked video modeling to learn shortcuts, thus resulting in limited transfer performance on downstream", + "bbox": [ + 496, + 779, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors", + "bbox": [ + 101, + 886, + 230, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6312", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tasks. To alleviate this issue, masked video modeling [57] often uses larger masking ratios.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we observe that much better performance on video downstream tasks can be achieved by conducting masked feature prediction by using the high-level features of pretrained MIM and MVM models as masked prediction targets. This can be viewed as two-stage masked video modeling, where MIM pretrained image models (i.e., an image teacher) or MVM pretrained video models (i.e., an video teacher) are obtained in the first stage, and they further act as teachers in the second stage for the student model via providing the high-level feature targets. Therefore, we call this method Masked Video Distillation (MVD).", + "bbox": [ + 75, + 123, + 468, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More interestingly, we find that student models distilled with different teachers in MVD exhibit different properties on different video downstream tasks. Specifically, students distilled from the image teacher perform better on video tasks that mainly rely on spatial clues, while students distilled from the video teacher model perform better on the video downstream tasks where temporal dynamics are more necessary. We think during the pretraining process of masked video modeling in the first stage, video teachers have learned spatial-temporal context in their high-level features. Therefore, when employing such high-level representations as prediction targets of masked feature modeling, it will help encouraging the student model to learn stronger temporal dynamics. By analogy, image teachers provide high-level features as targets that include more spatial information, which can help the student model learn more spatially meaningful representations. We further analyze the feature targets provided by image teachers and video teachers, and calculate the cross-frame feature similarity. It shows that the features provided by the video teachers contain more temporal dynamics.", + "bbox": [ + 75, + 291, + 470, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above observation, to leverage the advantages of video teachers and image teachers, we propose a simple yet effective spatial-temporal co-teaching strategy for MVD. In detail, the student model is designed to reconstruct the features coming from both the image teacher and video teacher with two different decoders, so as to learn stronger spatial representation and temporal dynamics at the same time. Experiments demonstrate that MVD with co-teaching from both the image teacher and the video teacher significantly outperforms MVD only using one single teacher on several challenging downstream tasks.", + "bbox": [ + 75, + 611, + 468, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the simplicity, our MVD is super effective and achieves very strong performance on multiple standard video recognition benchmarks. For example, on Kinectics-400 and Something-Something-v2 datasets, compared to the baseline without distillation, MVD with 400 epochs using a teacher model of the same size achieves $1.2\\%$ , $2.8\\%$ Top-1 accuracy gain on ViT-B. If a larger teacher model ViT-L is used, more significant performance gains (i.e.,", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$1.9\\%$ , $4.0\\%$ can be obtained. When ViT-Large is the target student model, our method can achieve $86.4\\%$ and $76.7\\%$ Top-1 accuracy on these two datasets, surpassing existing state-of-the-art method VideoMAE [57] by $1.2\\%$ and $2.4\\%$ respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with $77.3\\%$ Top-1 accuracy on Something-Something-v2.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as below:", + "bbox": [ + 517, + 196, + 834, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We find that using MIM pretrained image models and MVM pretrained video models as teachers to provide the high-level features for continued masked feature prediction can learn better video representation. And representations learned with image teachers and video teachers show different properties on different downstream video datasets.", + "- We propose masked video distillation together with a simple yet effective co-teaching strategy, which enjoys the synergy of image and video teachers.", + "- We demonstrate strong performance on multiple standard video recognition benchmarks, surpassing both the baseline without MVD and prior state-of-the-art methods by clear margins." + ], + "bbox": [ + 517, + 220, + 890, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 463, + 640, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision transformers for video understanding. For video understanding tasks, modeling the spatial-temporal information is the most important factor to consider in the architecture design. In the early works of video understanding, common video architectures, e.g., 3D CNNs [7, 20, 22, 58, 60] and 2D CNNs with temporal modules [12, 39, 53, 62, 73], are designed by extending existing 2D CNNs on the temporal dimension. Recently, Vision Transformers [13, 17, 41] achieve significant progress on several computer vision tasks. Some works [4, 18, 42, 69] also adapt vision transformers to the video domain and achieve superior performance compared to previous CNN-based architectures. For example, TimeSformer [4] and ViViT [1] study several variants of space-time factorization for the plain ViT. Some works [5, 47, 50] further explore how to reduce computational cost of the space-time attention. VideoSwin [42] and MViT [18, 38] study the hierarchical architecture and introduce an inductive locality bias into video transformers. Uniformer [36] and Video Mobile-Former [64] propose to integrate 3D CNNs and spatial-temporal attention mechanism for efficiency consideration. For convincing performance on the video understanding tasks, most video transformers require model weights pretrained on the large-scale image datasets. In this paper, we explore the self-supervised pretraining of video transformers and show pretraining strategy will significantly influence the downstream performance, which is orthogonal to the transformer architecture design.", + "bbox": [ + 496, + 492, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6313", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised video representation learning. The early works [3, 45, 66, 70] of self-supervised video representation learning focus on designing the pretext tasks based on the temporal structure of videos. More recently, contrastive learning [8, 32, 37] that forces different views of the same image sample to be closer in the feature space while pushing the views of different images farther becomes a new paradigm of representation learning, and some works [11, 23, 24, 29, 30, 46, 49] design the contrastive learning methods on the video domain by exploring effective ways of spatial-temporal augmentations. However, as the learning supervision based on contrastive learning is applied on global representation, it cannot well model the local relationship or learn fine-grained local representation.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked visual modeling. Masked language modeling [10, 40] has been one of the dominant pretraining methods of language transformers. With the success of vision transformers, masked visual modeling [2] has been introduced to self-supervised visual pretraining and demonstrates to be also helpful to multimodal visual-language learning [16, 74]. Following BERT [10] pretraining, BEiT [2] and PeCo [15] pretrain ViT by predicting the discrete visual tokens of masked patches, which are encoded by a pretrained VQ-VAE. MAE [31] proposes an asymmetric encoder-decoder framework for the reconstruction of pixels. SimMIM [68] and MaskFeat [67] propose to recover low-level features of masked patches like pixels or HOG features with hierarchical ViT. In contrast, some works [9, 14, 75] adopt an exponential moving average of the student model as the online teacher model, which makes the target features bootstrapped during training. In the video domain, some pioneering works [21, 55, 57, 63] extend masked image modeling to masked video modeling. BEVT [63] proposes a two-stream pretraining joint pretraining framework by predicting the discrete tokens with both image transformer and video transformer. VideoMAE [57] and ST-MAE [21] follow MAE and reconstruct the pixels of masked video patches with an extremely high masking ratio. Unlike most previous works of masked video modeling, our MVD focuses on masked feature modeling with high-level features as targets, and finds that student models using image and video teacher models will have different properties and complement each other.", + "bbox": [ + 75, + 306, + 472, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge distillation. Knowledge distillation [26,33,48] aims to transfer the knowledge of the teacher model to the student model by adopting the output of the teacher model as the target for training the student model. Typical knowledge distillation works [33, 51, 52] mainly focus on supervised learning, e.g., image classification. The idea of knowledge distillation has also been leveraged in many self-supervised methods by learning from momentum encoders [6, 56] or different architectures [24]. Recently, self-supervised knowledge distillation [19, 71, 72] has also been", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "studied to learn representations from self-supervised pretrained models. In this paper, we present the first attempt that uses the masked image modeling pretrained image and video model as the masked feature prediction target in the video domain. It shows self-supervised MIM pretrained model can further bootstrap the mask video pretraining and bring significant performance gain.", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 210, + 591, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While masked video modeling has demonstrated promising performance for self-supervised learning, most existing approaches reconstruct relatively low-level information in the forms of raw pixels [57], low-level features like HOG [67] and VQVAE tokens [63]. In this paper, instead of reconstructing low-level information, we conduct masked video modeling at the feature-level. This is achieved by a two-stage framework, MVD, optimized to predict high-level features that derived from off-the-shelf MIM pretrained image models [31] and MVM pretrained video models [57] which are readily available. Below, we first given an overview of the masked feature modeling paradigm in Sec. 3.1 and then we introduce our proposed MVD in Sec. 3.2. Finally, we present the architectural design of MVD in Sec. 3.3.", + "bbox": [ + 496, + 237, + 893, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The Paradigm of Masked Feature Modeling", + "text_level": 1, + "bbox": [ + 498, + 474, + 872, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core of masked feature modeling is to train models to predict features of masked input regions. In this paper, we follow the decoupled encoder-decoder transformer architecture in MAE [31] due to its effectiveness and simplicity. An input $X$ (image $X_{img} \\in \\mathbb{R}^{H \\times W \\times 3}$ or video $X_{vid} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ ) is partitioned into several non-overlapping patches, and then each patch is mapped to a visual token with a linear projection layer. Before feeding tokens to the transformer encoder $f$ , a subset of tokens is masked and dropped from the token sequence. To reconstruct the information of masked tokens, the token sequence consisting of the visible tokens from the encoder and learnable mask tokens are input to a shallow transformer decoder $g$ :", + "bbox": [ + 496, + 497, + 893, + 695 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY = g \\left(\\operatorname {c o n c a t} \\left(f \\left(X _ {\\text {v i s}}\\right), T _ {m}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 709, + 890, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $X_{vis}$ denotes the visible input tokens, and $T_{m}$ denotes the mask tokens. The subset of output tokens from decoder corresponded to input mask tokens contains reconstructed information of masked tokens. The reconstruction target for each masked patch $X(p)$ is represented as a patch feature $h(X(p))$ . Here, $h$ is represents a function for generating the target features, e.g. $h$ produces low-level RGB values of pixels in the patch in [21, 57]. Then, to train the encoder and the decoder, a loss function that measures the distance $D$ between the ground-truth features of masked patches and reconstructed ones is defined as:", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6314", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/42876f42afae95eccbfa8b9eeba9ec9aa8139a1c1c013a96cca151ad7b2822d8.jpg", + "image_caption": [ + "Figure 2. An overview of MVD framework. Firstly the image teacher is pretrained by masked image modeling and the video teacher is pretrained by masked video modeling. Then the student model is trained from scratch to predict target high-level features encoded by the image teacher and the video teacher. The teacher models are fixed in the distillation stage." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 831, + 328 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m f m} (h) = \\frac {1}{| M |} \\sum_ {p \\in M} D (Y (p), h (X (p))) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 400, + 468, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p$ is the token index and $M$ is the set of masked tokens. For pixel regression in MAE [31] and VideoMAE [57], the L2 distance is used as the distance metric.", + "bbox": [ + 76, + 441, + 470, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Masked Video Distillation", + "text_level": 1, + "bbox": [ + 76, + 494, + 313, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we propose Masked Video Distillation (MVD), which performs masked feature modeling on videos using high-level features as opposed to low-level pixels. In particular, we simply use outputs generated by off-the-shelf self-supervised pretrained image or video models, which are readily available, as reconstruction targets. These high-level features, serving as targets of the mask & prediction tasks, are encoded by teacher models pretrained by masked visual modeling like MAE or VideoMAE. For video representation learning, the reconstruction targets can take the form of spatial features encoded by image teacher models, or spatial-temporal features encoded by video teacher models. More specifically, the image teachers is pretrained by masked image modeling, while the video teacher is pretrained with masked video modeling, both of which aim at reconstructing raw pixels. Once trained, we use the image encoder $h_{img}$ to generate the spatial targets, and the pretrained video transformer encoder $h_{vid}$ to generate spatial-temporal targets. The loss function of MVD with the image teacher and video teacher can be denoted by $L_{mfm}(h_{img})$ and $L_{mfm}(h_{vid})$ , respectively", + "bbox": [ + 75, + 518, + 470, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial-temporal Co-teaching. When performing MVD with a single teacher, we observe that students distilled from different teachers learn different video representations and perform well on different kinds of downstream video", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tasks. To improve the performance of MVD on different downstream video tasks, we propose spatial-temporal co-teaching that explores information from both image and video teachers such that the student model can handle videos of different types better. For instance, videos with fastly changing human actions require more temporal information while spatial clues might be sufficient for relatively static videos. To this end, MVD is trained to predict target high-level features produced by the image teacher and the video teacher at the same time. This is achieved by using two separated decoders to reconstruct different target features. The final loss of MVD with spatial-temporal co-teaching is:", + "bbox": [ + 496, + 390, + 893, + 588 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m v d} = \\lambda_ {1} L _ {m f m} \\left(h _ {i m g}\\right) + \\lambda_ {2} L _ {m f m} \\left(h _ {v i d}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 612, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ denote the hyper-parameters that balance the weights of the image teacher and the video teacher. The pseudo code of MVD is shown in Algorithm 1.", + "bbox": [ + 496, + 637, + 893, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Architectural Design", + "text_level": 1, + "bbox": [ + 500, + 691, + 696, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoder. For MVD, the vanilla transformer backbone is used as the encoder. For a video input $X_{vid} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ , we adopt 3D patch embedding with a patch size of $2 \\times 16 \\times 16$ . After patch partitioning and linear embedding, we obtain $T / 2 \\times H / 16 \\times W / 16$ tokens. For the masked feature modeling, the tokens are masked with a high masking ratio and the remaining tokens are fed into the transformer layers. For finetuning on downstream tasks, we input all tokens to the subsequent layers. In each layer, joint spatial-temporal self-attention is applied on the whole input token sequence.", + "bbox": [ + 496, + 714, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mask strategy. For MVD, we follow [57] and adopt tube masking for masked feature modeling. First a 2D random", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6315", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudocode of MVD in PyTorch style." + ], + "code_body": "f: student encoder \n# g.img: decoder for reconstructing spatial features \n# g_vid: decoder for reconstructing spatial-temporal features \n# t_m: learnable mask tokens \n# h.img: image teacher model \n# h_vid: video teacher model \nfor x, m in loader: # x: video data, m: mask \nx_pe = patch_emb(x) # patch embedding of input \nx.vis = mask_select(x_pe, 1 - m) # masking tokens \nq.vis = f(x.vis) # visible local patch features \n# reconstruction of target features \np IMG = g.img(concat(q.vis, t_m)) \np_vid = g_vid(concat(q.vis, t_m)) \n# compute target features with teacher models \nk IMG = h.img(x) # target spatial features \nk_vid = h_vid(x) # target spatial-temporal features \n# compute reconstruction loss \nloss IMG = smooth_L1_loss(p.img ⊙ m, k.img ⊙ m) \nloss_vid = smooth_L1_loss(p:vid ⊙ m, k:vid ⊙ m) \nloss = λ1 * loss IMG + λ2 * loss_vid \nloss.backup() \noptimizer_STEP() # optimizer update", + "guess_lang": "python", + "bbox": [ + 76, + 116, + 467, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mask is generated and then extended along the temporal dimension. Therefore, the spatial mask on each time slice is the same, which prevents information leakage between frames. Tube masking with a high masking ratio (e.g., $90\\%$ ) encourages the video transformer to model the high-level semantics during pretraining.", + "bbox": [ + 75, + 407, + 470, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoder. For MVD, shallow decoders consist of vanilla transformer layers and a linear projection layer. The transformer layers in decoders are the same as those in the encoder. Since spatial-temporal co-teaching introduces two different reconstruction targets for masked feature modeling, two separated decoders that share the same architecture but contain different weights are placed on the top of the encoder. Learnable masked tokens corresponded to masked patches are concatenated with visible tokens from the encoder before fed into the decoder. After jointly modeling the spatial-temporal relationships, the output tokens of transformer layers are mapped to final predictions by the linear projection layer.", + "bbox": [ + 75, + 511, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction targets. For generating spatial-temporal target features, the video teacher, which shares the same architecture as the student model, is pretrained by a VideoMAE [57] manner on the video dataset. For obtaining spatial targets, we adopt the vanilla image ViT pretrained by masked image modeling [31] on the image dataset (e.g., ImageNet-1K). It is worth noting that one 3D patch (with size of $2 \\times 16 \\times 16$ ) for the video transformer corresponds to two 2D patches (with size of $16 \\times 16$ ) for the image transformer. Following [21], we predict the spatial features of a single time slice (that is the front 2D patch), which reduce the prediction layer's size.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 633, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the experimental setup in Sec. 4.1, and then present the main results in Sec. 4.2, followed by an extensive analysis to verify the effectiveness of different components in Sec. 4.3.", + "bbox": [ + 496, + 114, + 893, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 186, + 689, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We pretrain the vanilla ViT with MVD on Kinetics-400 by default, and evaluate the learned model on four video recognition downstream tasks: (a) Kinetics-400 (K400) [7], which consists of $\\sim 240K$ training videos and $\\sim 20K$ validation videos with an average duration of 10 seconds. All video clips are labeled into 400 classes. (b) Something-Something V2 (SSv2) [27], which contains $\\sim 160K$ videos for training and $\\sim 20K$ videos for validation. The videos in SSv2 with an average duration of 4 seconds are labeled into 174 motion-centric categories. (c) UCF-101 [54] is a relatively small dataset, consisting of $\\sim 9.5K$ training videos and $\\sim 3.5K$ validation videos. (d) HMDB51 [35] is also a small video dataset that contains around $3.5K/1.5K$ train/val videos. We also transfer pretrained models to a challenging spatial-temporal action detection dataset AVA [28]. The results of AVA are included in supplementary materials.", + "bbox": [ + 496, + 210, + 893, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our MVD is performed on vanilla ViTs with different capacities (i.e., ViT-S, ViT-B, ViT-L). By default, image teacher models are pretrained on ImageNet-1K for 1600 epochs and video teachers are pretrained on K400 for 1600 epochs. We follow the training strategy in MAE [31] and VideoMAE [21] for image teachers and video teachers respectively. In the distillation stage, student models are first pretrained from scratch on K400 for 400 epochs unless mentioned otherwise. The resulting models are then finetuned on downstream video tasks. The video clip length is 16 for both pretraining and finetuning. We adopt AdamW optimizer [44] and Smooth L1 loss for the optimization of student models. More details are included in supplementary materials.", + "bbox": [ + 496, + 473, + 893, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 695, + 638, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Students distilled from different teachers. Unlike masked image modeling, masked feature modeling on video data has more choices on reconstruction targets. Besides spatial features, to include temporal dynamics in the reconstruction target, we can also adopt spatial-temporal features encoded by pretrained video models. In Table 1, we compare students distilled by the image teacher and the video teacher on K400, a downstream task that mainly relies on spatial clues, and SSv2, a temporally-heavy downstream task. Our observations are as follow: (a) Masked feature modeling with high-level features as targets achieves convincing performance on downstream video tasks, and outperforms Video", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6316", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/50df52d8e10c4c7ca7e120607bf30fde978ecd0cdb35b38a19e0a49cd89a075a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
studentteachersK400 top-1SSv2 top-1
imagevideo
ViT-S✓✗80.469.4
✗✓80.170.0
✓✓80.670.7
ViT-B✓✗82.371.4
✗✓82.171.8
✓✓82.772.5
", + "bbox": [ + 83, + 88, + 459, + 212 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1e1cc95c0054fd05efa40de8931eb3bf994b1eaf04465612ebb95e0d6cca9107.jpg", + "table_caption": [ + "Table 1. Co-teaching with the image teacher and the video teacher outperforms distillation with a single teacher in MVD. We adopt ViT-B as the architecture of teachers and distill students for 400 epochs here." + ], + "table_footnote": [], + "table_body": "
studentteacherK400 top-1SSv2 top-1
ViMAEMVDViMAEMVD
ViT-SViT-B79.080.6 ↑1.666.470.7 ↑4.3
ViT-SViT-L79.081.0 ↑2.066.470.9 ↑4.5
ViT-BViT-B81.582.7 ↑1.269.772.5 ↑2.8
ViT-BViT-L81.583.4 ↑1.969.773.7 ↑4.0
ViT-LViT-L85.286.0 ↑0.874.076.1 ↑2.1
", + "bbox": [ + 78, + 275, + 470, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MAE baseline significantly (compared with the baseline results in Table 2). In particular, with both image and video teachers using a ViT-S as the backbone, MVD achieves consistent gains over VideoMAE on both K400 (80.6% vs. 79.0%) and SSV2 (70.7% vs. 66.4%). (b) Students distilled from the image teacher achieve higher top-1 accuracy on K400, while students distilled from the video teacher perform better on SSv2. For example, ViT-S achieves an accuracy of 80.4% and 69.4% using an image teacher on K400 and SSv2 respectively. With a video teacher, on the other hand, ViT-S offers a top-1 accuracy of 80.1% and 70.0% respectively. As it has been shown that videos in K400 less sensitive to temporal modeling compared to SSv2, the results demonstrate students learn stronger spatial representation from the image teacher while the video teacher transfer more knowledge about temporal dynamics to students.", + "bbox": [ + 75, + 450, + 468, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Co-teaching outperforms distilling with a single teacher. To improve the performance on different kinds of downstream video tasks, we introduce spatial-temporal co-teaching in MVD, which trains the model to predict spatial features and spatial-temporal features of masked patches in a decoupled way. The results in Table 1 indicate that students distilled with spatial-temporal co-teaching outperforms students distilled from either single teacher on both spatially-heavy task and temporally-heavy task.", + "bbox": [ + 75, + 698, + 467, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MVD outperforms VideoMAE baseline significantly. In Table 2, MVD with spatial-temporal co-teaching is compared with VideoMAE pretrained on K400. When the size of teacher models is the same as that of students, MVD out", + "bbox": [ + 76, + 839, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/184939b9f85e0c9c294775bea22a135e2fcaa24dfe1c979ed9b8dace15f8d1ea.jpg", + "table_caption": [ + "Table 2. MVD achieves significant improvement compared with VideoMAE across different model scales. MVD is pretrained on K400 for 400 epochs, while VideoMAE is pretrained on K400 for 1600 epochs here." + ], + "table_footnote": [], + "table_body": "
methodextra datatop-1top-5GFLOPsParam
supervised
NL I3D R101 [65]-77.393.3359×3062
ip-CSN-152 [59]-77.892.8109×3033
SlowFast NL [22]-79.893.9234×3060
X3D-XL [20]-79.193.948×3011
MViTv1-B [18]-80.294.4170×537
VideoSwin-B [42]IN-1K80.694.6282×1288
Uniformer-B [36]IN-1K83.095.4259×1250
TimeSformer [4]IN-21K80.794.72380×3121
Mformer-B [47]IN-21K79.794.2370×30109
Mformer-L [47]IN-21K80.294.81185×30382
ViViT-L FE [1]IN-21K81.793.83980×3N/A
VideoSwin-L [42]IN-21K83.195.9604×12197
self-supervised
VIMPAC ViT-L [55]HowTo100M77.4N/AN/A×30307
BEVT Swin-B [63]IN-1K81.1N/A282×1288
MaskFeat MViT-S [67]-82.295.171×1036
VideoMAE ViT-S [57]-79.093.857×1522
VideoMAE ViT-B [57]-81.595.1180×1587
VideoMAE ViT-L [57]-85.296.8597×15305
VideoMAE ViT-H [57]-86.697.11192×15633
ST-MAE ViT-B [21]-81.394.9180×2187
ST-MAE ViT-L [21]-84.896.2598×21304
ST-MAE ViT-H [21]-85.196.61193×21632
OmniMAE ViT-B [25]IN-1K80.8N/A180×1587
OmniMAE ViT-L [25]IN1K+SSv284.0N/A597×15305
OmniMAE ViT-H [25]IN1K+SSv284.8N/A1192×15633
MVD-S (Teacher-B)IN-1K80.694.757×1522
MVD-S (Teacher-L)IN-1K81.094.857×1522
MVD-B (Teacher-B)IN-1K82.795.4180×1587
MVD-B (Teacher-L)IN-1K83.495.8180×1587
MVD-L (Teacher-L)IN-1K86.096.9597×15305
MVD-L (Teacher-L) †IN-1K86.497.0597×15305
MVD-H (Teacher-H) †IN-1K87.397.41192×15633
", + "bbox": [ + 496, + 88, + 893, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with previous works on Kinetics-400. \"MVD-S (Teacher-B)\" denotes that the student is ViT-S and the teacher is ViT-B. $\\dagger$ denotes that the student is distilled for 800 epochs instead of 400 epochs. The inference cost is reported with the cost of a single view $\\times$ the number of views. Note that IN-1K is only used for training image teachers in MVD.", + "bbox": [ + 496, + 595, + 890, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performs VideoMAE by a clear margin on both K400 and SSv2. Larger models as teachers can further boost the performance of MVD. It is worth mentioning that not only is our MVD particularly effective for relatively small models, but also it improves the performance of large vision models like ViT-L. For example, with a ViT-L model as the student, MVD achieves $86.0\\%$ and $76.1\\%$ on K400 and SSv2, surpassing VideoMAE by $0.8\\%$ and $2.1\\%$ , respectively.", + "bbox": [ + 496, + 696, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with state-of-the-art. We compare MVD with prior studies on four video recognition tasks. Results on K400 are shown in Table 3. Our MVD outperforms previous self-supervised methods with similar or less computational cost. Even compared with video transform", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6317", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cc113547bae6fbfcfc7a5afa7af8e5721c5948986eafec086e038c8cff931edf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodextra datatop-1GFLOPsParam
supervised
SlowFast R101 [22]K40063.1106×353
TSM-RGB R50 [39]IN-1K63.362×624
TAM R50 [43]IN-1K66.099×651
TDN R101 [61]IN-1K69.6198×388
MViTv1-B [18]-67.7455×337
MViTv2-B [38]K40070.5225×351
Uniformer-B [36]K40071.2259×350
TimeSformer-HR [4]IN-21K62.51703×3121
ViViT-L FE [1]IN21K+K40065.9995×12N/A
Mformer-B [47]IN21K+K40066.5370×3109
Mformer-L [47]IN21K+K40068.11185×3382
VideoSwin-B [42]IN21K+K40069.6321×388
MViTv2-L [38]IN21K+K40073.32828×3213
self-supervised
VIMPAC ViT-L [55]HowTo100M68.1N/A×30307
BEVT Swin-B [63]IN1K+K40071.4321×388
MaskFeat MViT-L [67]K40074.42828×3218
VideoMAE ViT-S [57]K40066.457×622
VideoMAE ViT-S [57]-66.857×622
VideoMAE ViT-B [57]K40069.7180×687
VideoMAE ViT-B [57]-70.8180×687
VideoMAE ViT-L [57]K40074.0597×6305
VideoMAE ViT-L [57]-74.3597×6305
ST-MAE ViT-L [21]K40072.1598×3304
ST-MAE ViT-H [21]K40074.11193×3632
OmniMAE ViT-B [25]IN1K+K40069.0180×687
OmniMAE ViT-L [25]IN-1K74.2597×6305
OmniMAE ViT-H [25]IN-1K75.31192×6632
MVD-S (Teacher-B)IN1K+K40070.757×622
MVD-S (Teacher-L)IN1K+K40070.957×622
MVD-B (Teacher-B)IN1K+K40072.5180×687
MVD-B (Teacher-L)IN1K+K40073.7180×687
MVD-L (Teacher-L)IN1K+K40076.1597×6305
MVD-L (Teacher-L) †IN1K+K40076.7597×6305
MVD-H (Teacher-H) †IN1K+K40077.31192×6633
", + "bbox": [ + 73, + 88, + 472, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ers pretrained on ImageNet-21K, MVD achieves superior performance. Particularly, MVD-H achieves an accuracy of $87.3\\%$ on K400, outperforming previous top-performing methods by clear margins. Table 4 presents comparisons to the state-of-the-art methods on SSv2. We observe that for downstream tasks depending on temporal relationship modeling, self-supervised methods based on MVM achieve better performance in comparison with supervised methods (cf. results in the middle group of Table 4 vs. results in the top group). Once again, our MVD, producing an accuracy of $76.1\\%$ with a large model, beats both supervised meth", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8707af0e1bd3a85c90c59211d9235345b4386207db2a6e41e023ea38a79e1747.jpg", + "table_caption": [ + "Table 4. Comparison with previous works on Something-Something V2.. † denotes that the student is distilled for 800 epochs." + ], + "table_footnote": [], + "table_body": "
methodextra dataParamUCF101HMDB51
VideoMoCo R2+1D [46]K4001578.749.2
MemDPC R2D3D [30]K4003286.154.5
Vi2CLR S3D [11]K400989.155.7
CORP Slow-R50 [34]K4003293.568.0
CVRL Slow-R50 [49]K4003292.967.9
CVRL Slow-R152 [49]K60032894.470.6
ρBYOL Slow-R50 [23]K4003294.272.1
VIMPAC ViT-L [55]HowTo100M30792.765.9
VideoMAE ViT-B [57]K4008796.173.3
MVD-B (Teacher-B)IN-1K+K4008797.076.4
MVD-B (Teacher-L)IN-1K+K4008797.579.7
", + "bbox": [ + 498, + 88, + 895, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Comparison with previous state-of-the-art methods on UCF101 and HMDB51.", + "bbox": [ + 498, + 262, + 890, + 289 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b85514161767a24e5ba9d65e7a6ebec3da134586a2592035b7fe27a40a8bdffc.jpg", + "image_caption": [ + "(a) image teacher." + ], + "image_footnote": [], + "bbox": [ + 503, + 300, + 681, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a74256dfcf3e265a50a769507214dc95a131cbb2672bc6fde43b3c3c6ba27ea.jpg", + "image_caption": [ + "(b) video teacher.", + "Figure 3. Feature similarity across different frames for different teacher models. Similarity matrices are computed on the Kinetics-400 validation set." + ], + "image_footnote": [], + "bbox": [ + 709, + 299, + 887, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods and self-supervised methods by clear margins. With more training epochs (i.e., 800 epochs), MVD with ViT-L achieves more significant performance gains (i.e., $1.2\\%$ , $2.4\\%$ on K400 and SSv2) compared with VideoMAE. When a huge model is used, the performance can still be boosted and MVD achieves $77.3\\%$ top-1 accuracy on SSv2.", + "bbox": [ + 496, + 482, + 890, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate the transfer learning ability of MVD on two relatively small datasets, UCF101 and HMDB51. As shown in Table 5, MVD with ViT-B obtains higher accuracy compared with prior works based on well-designed pretext tasks, contrastive learning and masked video modeling methods. Especially compared to the original VideoMAE [57] ViT-B teacher model, we achieve $0.9\\%$ and $3.1\\%$ higher points on these two datasets respectively. Additionally, when teachers with larger size are adopted, MVD achieves stronger transfer learning performance.", + "bbox": [ + 496, + 574, + 890, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 500, + 738, + 718, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide an in-depth analysis of the effectiveness of different components in MVD.", + "bbox": [ + 496, + 763, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of features encoded by different teachers. The properties of target features generated by different teachers may influence the performance of students on different downstream tasks. To quantify the temporal dynamics that teacher models capture from the input video, we study the similarity between feature maps across different frames of each input video clip via the cosine similarity. As similarity", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6318", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f21e6d752cc792304fcbe915ce427f16ea0e5c490d436db6be0fcf4db6bf9ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodepochtime(h)K400 top-1
VideoMAE80010781.0
VideoMAE160021481.5
MVD(800+) 400(107+) 5781.9
", + "bbox": [ + 84, + 88, + 455, + 151 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b8c785510b01f5b756208b0c5574af8377521103df4114f165e2f0490dc7d9c3.jpg", + "table_caption": [ + "Table 6. Training time comparison between MVD and VideoMAE on 32 NVIDIA V100 GPUs. The training time of teacher models is considered. we only adopt a video teacher pretrained for 800 epochs in MVD here. ViT-B is used here." + ], + "table_footnote": [], + "table_body": "
teachersreconstruct pixelSSv2 top-1
image modelX68.7
image model✓67.9 ↓0.8
image+video modelsX70.1
image+video models✓69.0 ↓1.1
", + "bbox": [ + 81, + 213, + 460, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0ea09233f5232f07a5f99176590c2d0252130f04302cdc66133a84493527826d.jpg", + "table_caption": [ + "Table 7. Ablations on reconstructing low-level features in masked video distillation. We distill ViT-S students with ViT-B teachers for 300 epochs here." + ], + "table_footnote": [], + "table_body": "
teacherinitepochtop-1
K400SSv2
momentum encoderX80080.570.4
momentum encoder✓80081.870.8
fixed image modelX40082.371.4
fixed video modelX40082.171.8
fixed co-teachingX40082.772.5
", + "bbox": [ + 88, + 337, + 454, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "matrices shown in Figure 3, for image teachers, the feature maps of different frames are almost the same. However, for video teachers, the features of different frames have larger difference. This indicates that video teachers capture more temporal difference. Therefore, students distilled from video teachers can learn stronger temporal dynamics and perform better on temporally-heavy downstream tasks.", + "bbox": [ + 75, + 483, + 467, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training time comparison. We study whether MVD is able to achieve better balance of accuracy and efficiency than VideoMAE. For fair comparisons, the training time of teacher models is also counted in the total training time of MVD. Results are shown in Table 6. We see that MVD can achieve better accuracy (i.e., $81.9\\%$ ) by a total of 164 hours of training which is 50 hours less than VideoMAE (producing an accuracy of $81.5\\%$ ) trained for 1600 epochs.", + "bbox": [ + 75, + 594, + 467, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reconstruction signals in MVD. In MVD, we first pretrain teacher models by recovering the pixels of masked patches in a MAE manner, then adopt features produced by teachers as the targets of masked feature modeling. In Table 7, we study whether to include an additional decoder branch for reconstructing pixels of masked patches in the distillation stage. The experimental results show that for both MVD with a single teacher and MVD with spatial-temporal co-teaching, the reconstruction of low-level feature targets degrades the performance for downstream tasks. Therefore, we only reconstruct high-level features of masked patches in the distillation stage of MVD.", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5aa368d100821e3939d980725e7f95ac7d35ed6a3972c8fd1c731cc71ac6b9e8.jpg", + "table_caption": [ + "Table 8. Comparison with bootstrapped teachers and students initialized with IN1K-pretrained models. ViT-B is used here." + ], + "table_footnote": [], + "table_body": "
distillation methodK400 top-1SSv2 top-1
per-token distillation80.970.5
masked reconstruction82.171.8
", + "bbox": [ + 514, + 88, + 872, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Comparison with feature distillation without masked reconstruction. We distill ViT-B from the video teacher here.", + "bbox": [ + 500, + 138, + 890, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with bootstrapped teachers. Some recent works of image representation learning [9, 14, 16, 75] adopt features of a momentum encoder as the targets of MIM, while we use frozen teacher models in MVD. In Table 8, we compare fixed teachers with bootstrapped teachers that are updated by an exponential moving average of the online encoder during pretraining. According to a MIM method [14], two strong baselines of bootstrapped teachers are built for video representation learning: (a) The student model is trained from scratch with masked feature modeling, and the target features are generated by a momentum encoder. (b) The framework of bootstrapped teachers is first pretrained on IN-1K for 800 epochs, then the pretrained weights are adopted to initialize the video pretraining. As shown in Table 8, MVD with the frozen teacher beats the method with a bootstrapped teacher on the downstream video tasks, even if only a single teacher is utilized.", + "bbox": [ + 496, + 179, + 890, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with feature distillation. In the previous work of self-supervised feature distillation [19, 26, 72], the distillation loss is directly computed upon the full feature maps between teachers and students. Accordingly, we use a baseline method named per-token distillation. Specifically, the output features of students are projected by a MLP, and then forced to mimicking the teacher's feature at each token with a Smooth L1 loss. As shown in Table 9, masked feature reconstruction in our MVD outperforms per-token distillation on both K400 and SSv2.", + "bbox": [ + 496, + 441, + 890, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 614, + 617, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study masked video distillation upon MIM pretrained image or video transformers. We have three interesting findings: 1) Using MIM pretrained image transformers or MVM pretrained video transformers as teachers supervise masked feature prediction can significantly boost the finetuning performance on video downstream tasks; 2) The representation distilled with image and video teachers will have different properties, i.e., image teachers will benefit spatial-heavy video tasks more while video teachers benefit temporal-heavy video tasks more; 3) Combining image and video teachers will enjoy the synergy and thus produce higher performance. Even though the proposed masked video distillation seems very straightforward, we hope such interesting findings can motivate more thinking about masked video pretraining.", + "bbox": [ + 496, + 641, + 890, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This project was supported by NSFC under Grant No. 62032006 and No. 62102092.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6319", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2, 6, 7", + "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. In ICLR, 2022. 1, 3", + "[3] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In CVPR, 2020. 3", + "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 2, 6, 7", + "[5] Adrian Bulat, Juan Manuel Perez Rua, Swathikiran Sudhakaran, Brais Martinez, and Georgios Tzimiropoulos. Space-time mixing attention for video transformer. In NeurIPS, 2021. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3", + "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 2, 5", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3", + "[9] Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Sdae: Self-distillated masked autoencoder. In ECCV, 2022. 3, 8", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3", + "[11] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saqib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2clr: Video and image for visual contrastive learning of representation. In ICCV, 2021. 3, 7", + "[12] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2", + "[13] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. 2", + "[14] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In ECCV, 2022. 3, 8", + "[15] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of vision transformers. In Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI), 2023. 3", + "[16] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3, 8", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2", + "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 2, 6, 7", + "[19] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation. arXiv preprint arXiv:2101.04731, 2021. 3, 8", + "[20] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In CVPR, 2020. 2, 6", + "[21] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. 1, 3, 5, 6, 7", + "[22] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2, 6, 7", + "[23] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, 2021. 3, 7", + "[24] Chongjian Ge, Youwei Liang, Yibing Song, Jianbo Jiao, Jue Wang, and Ping Luo. Revitalizing cnn attention via transformers in self-supervised visual representation learning. NeurIPS, 2021. 3", + "[25] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 6, 7", + "[26] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. IJCV, 2021. 3, 8", + "[27] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fründ, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 5", + "[28] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 5", + "[29] Sheng Guo, Zihua Xiong, Yujie Zhong, Limin Wang, Xiaobo Guo, Bing Han, and Weilin Huang. Cross-architecture self-supervised video representation learning. In CVPR, 2022. 3" + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6320", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Tengda Han, Weidi Xie, and Andrew Zisserman. Memory-augmented dense predictive coding for video representation learning. In ECCV, 2020. 3, 7", + "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 3, 4, 5", + "[32] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3", + "[33] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[34] Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. In ICCV, 2021. 7", + "[35] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 5", + "[36] Kunchang Li, Yali Wang, Gao Peng, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatial-temporal representation learning. In ICLR, 2022. 2, 6, 7", + "[37] Suichan Li, Dongdong Chen, Yinpeng Chen, Lu Yuan, Lei Zhang, Qi Chu, Bin Liu, and Nenghai Yu. Improve unsupervised pretraining for few-label transfer. In ICCV, 2021. 3", + "[38] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In CVPR, 2022. 2, 7", + "[39] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 2, 7", + "[40] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3", + "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2", + "[42] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 2, 6, 7", + "[43] Zhaoyang Liu, Limin Wang, Wayne Wu, Chen Qian, and Tong Lu. Tam: Temporal adaptive module for video recognition. In CVPR, 2021. 7", + "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018. 5", + "[45] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. 3", + "[46] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In CVPR, 2021. 3, 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In NeurIPS, 2021. 2, 6, 7", + "[48] Mary Phuong and Christoph Lampert. Towards understanding knowledge distillation. In ICML, 2019. 3", + "[49] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, 2021. 3, 7", + "[50] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In NeurIPS, 2021. 2", + "[51] Zhiqiang Shen and Marios Savvides. Meal v2: Boosting vanilla resnet-50 to $80\\%+$ top-1 accuracy on imagenet without tricks. arXiv preprint arXiv:2009.08453, 2020. 3", + "[52] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. arXiv preprint arXiv:2112.01528, 2021. 3", + "[53] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014. 2", + "[54] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5", + "[55] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 3, 6, 7", + "[56] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Piriavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, 2021. 3", + "[57] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7", + "[58] Du Tran, Lubomir D Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. C3d: Generic features for video analysis. In ICCV, 2015. 2", + "[59] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 6", + "[60] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 2", + "[61] Limin Wang, Zhan Tong, Bin Ji, and Gangshan Wu. Tdn: Temporal difference networks for efficient action recognition. In CVPR, 2021. 7", + "[62] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE TPAMI, 2018. 2", + "[63] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, 2022. 1, 3, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6321", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Rui Wang, Zuxuan Wu, Dongdong Chen, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Luowei Zhou, Lu Yuan, and YuGang Jiang. Video mobile-former: Video recognition with efficient global spatial-temporal modeling. arXiv preprint arXiv:2208.12257, 2022. 2", + "[65] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 6", + "[66] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, 2015. 3", + "[67] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. 3, 6, 7", + "[68] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 3", + "[69] Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. arXiv preprint arXiv:2211.13222, 2022. 2", + "[70] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019. 3", + "[71] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, 2020. 3", + "[72] Haohang Xu, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, Xinggang Wang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Bag of instances aggregation boosts self-supervised distillation. In ICLR, 2021. 3, 8", + "[73] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015. 2", + "[74] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. 3", + "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 3, 8" + ], + "bbox": [ + 78, + 90, + 468, + 741 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6322", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_model.json b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ef0a18e77b6134bd44c82e3da2df972c2838a898 --- /dev/null +++ b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.13, + 0.834, + 0.177 + ], + "angle": 0, + "content": "Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.202, + 0.818, + 0.293 + ], + "angle": 0, + "content": "Rui Wang\\(^{1,2}\\) Dongdong Chen\\(^{3}\\) Zuxuan Wu\\(^{1,2\\dagger}\\) Yinpeng Chen\\(^{3}\\) Xiyang Dai\\(^{3}\\) Mengchen Liu\\(^{3}\\) Lu Yuan\\(^{3}\\) Yu-Gang Jiang\\(^{1,2\\dagger}\\) \n\\(^{1}\\)Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University \n\\(^{2}\\)Shanghai Collaborative Innovation Center of Intelligent Visual Computing \n\\(^{3}\\)Microsoft Cloud + AI" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.86 + ], + "angle": 0, + "content": "Benefiting from masked visual modeling, self-supervised video representation learning has achieved remarkable progress. However, existing methods focus on learning representations from scratch through reconstructing low-level features like raw pixel values. In this paper, we propose masked video distillation (MVD), a simple yet effective two-stage masked feature modeling framework for video representation learning: firstly we pretrain an image (or video) model by recovering low-level features of masked patches, then we use the resulting features as targets for masked feature modeling. For the choice of teacher models, we observe that students taught by video teachers perform better on temporally-heavy video tasks, while image teachers transfer stronger spatial representations for spatially-heavy video tasks. Visualization analysis also indicates different teachers produce different learned patterns for students. To leverage the advantage of different teachers, we design a spatial-temporal co-teaching method for MVD. Specifically, we distill student models from both video teachers and image teachers by masked feature modeling. Extensive experimental results demonstrate that video transformers pretrained with spatial-temporal co-teaching outperform models distilled with a single teacher on a multitude of video datasets. Our MVD with vanilla ViT achieves state-of-the-art performance compared with previous methods on several challenging video downstream tasks. For example, with the ViT-Large model, our MVD achieves \\(86.4\\%\\) and \\(76.7\\%\\) Top-1 accuracy on Kinetics-400 and Something-Something-v2, outperforming VideoMAE by \\(1.2\\%\\) and \\(2.4\\%\\) respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with \\(77.3\\%\\) Top-1 accuracy on Something-Something-v2. Code will be available at https://github.com/ruiwang2021/mvd." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.325, + 0.888, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.545, + 0.894, + 0.587 + ], + "angle": 0, + "content": "Figure 1. Comparisons of MVD with previous supervised or self-supervised methods on Something-Something v2. Each line represents the corresponding model of different sizes." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.602, + 0.633, + 0.618 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.895, + 0.779 + ], + "angle": 0, + "content": "For self-supervised visual representation learning, recent masked image modeling (MIM) methods like MAE [31] and BEiT [2] achieve promising results with vision transformers [17] on various vision downstream tasks. Such a pretraining paradigm has also been adapted to the video domain and boosts video transformers by clear margins compared with supervised pretraining on several video downstream tasks. Representative masked video modeling (MVM) works include BEVT [63], VideoMAE [57] and ST-MAE [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Following MAE [31] and BEiT [2], existing masked video modeling methods [21, 57, 63] pretrain video transformers through reconstructing low-level features, e.g., raw pixel values or low-level VQVAE tokens. However, using low-level features as reconstruction targets often incur much noise. And due to the high redundancy in video data, it is easy for masked video modeling to learn shortcuts, thus resulting in limited transfer performance on downstream" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.232, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6312" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "tasks. To alleviate this issue, masked video modeling [57] often uses larger masking ratios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.47, + 0.29 + ], + "angle": 0, + "content": "In this paper, we observe that much better performance on video downstream tasks can be achieved by conducting masked feature prediction by using the high-level features of pretrained MIM and MVM models as masked prediction targets. This can be viewed as two-stage masked video modeling, where MIM pretrained image models (i.e., an image teacher) or MVM pretrained video models (i.e., an video teacher) are obtained in the first stage, and they further act as teachers in the second stage for the student model via providing the high-level feature targets. Therefore, we call this method Masked Video Distillation (MVD)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.609 + ], + "angle": 0, + "content": "More interestingly, we find that student models distilled with different teachers in MVD exhibit different properties on different video downstream tasks. Specifically, students distilled from the image teacher perform better on video tasks that mainly rely on spatial clues, while students distilled from the video teacher model perform better on the video downstream tasks where temporal dynamics are more necessary. We think during the pretraining process of masked video modeling in the first stage, video teachers have learned spatial-temporal context in their high-level features. Therefore, when employing such high-level representations as prediction targets of masked feature modeling, it will help encouraging the student model to learn stronger temporal dynamics. By analogy, image teachers provide high-level features as targets that include more spatial information, which can help the student model learn more spatially meaningful representations. We further analyze the feature targets provided by image teachers and video teachers, and calculate the cross-frame feature similarity. It shows that the features provided by the video teachers contain more temporal dynamics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Motivated by the above observation, to leverage the advantages of video teachers and image teachers, we propose a simple yet effective spatial-temporal co-teaching strategy for MVD. In detail, the student model is designed to reconstruct the features coming from both the image teacher and video teacher with two different decoders, so as to learn stronger spatial representation and temporal dynamics at the same time. Experiments demonstrate that MVD with co-teaching from both the image teacher and the video teacher significantly outperforms MVD only using one single teacher on several challenging downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Despite the simplicity, our MVD is super effective and achieves very strong performance on multiple standard video recognition benchmarks. For example, on Kinectics-400 and Something-Something-v2 datasets, compared to the baseline without distillation, MVD with 400 epochs using a teacher model of the same size achieves \\(1.2\\%\\), \\(2.8\\%\\) Top-1 accuracy gain on ViT-B. If a larger teacher model ViT-L is used, more significant performance gains (i.e.," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.197 + ], + "angle": 0, + "content": "\\(1.9\\%\\), \\(4.0\\%\\) can be obtained. When ViT-Large is the target student model, our method can achieve \\(86.4\\%\\) and \\(76.7\\%\\) Top-1 accuracy on these two datasets, surpassing existing state-of-the-art method VideoMAE [57] by \\(1.2\\%\\) and \\(2.4\\%\\) respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with \\(77.3\\%\\) Top-1 accuracy on Something-Something-v2." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.198, + 0.835, + 0.212 + ], + "angle": 0, + "content": "Our contributions can be summarized as below:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.221, + 0.892, + 0.326 + ], + "angle": 0, + "content": "- We find that using MIM pretrained image models and MVM pretrained video models as teachers to provide the high-level features for continued masked feature prediction can learn better video representation. And representations learned with image teachers and video teachers show different properties on different downstream video datasets." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.336, + 0.892, + 0.381 + ], + "angle": 0, + "content": "- We propose masked video distillation together with a simple yet effective co-teaching strategy, which enjoys the synergy of image and video teachers." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.391, + 0.892, + 0.451 + ], + "angle": 0, + "content": "- We demonstrate strong performance on multiple standard video recognition benchmarks, surpassing both the baseline without MVD and prior state-of-the-art methods by clear margins." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.221, + 0.892, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.464, + 0.642, + 0.48 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Vision transformers for video understanding. For video understanding tasks, modeling the spatial-temporal information is the most important factor to consider in the architecture design. In the early works of video understanding, common video architectures, e.g., 3D CNNs [7, 20, 22, 58, 60] and 2D CNNs with temporal modules [12, 39, 53, 62, 73], are designed by extending existing 2D CNNs on the temporal dimension. Recently, Vision Transformers [13, 17, 41] achieve significant progress on several computer vision tasks. Some works [4, 18, 42, 69] also adapt vision transformers to the video domain and achieve superior performance compared to previous CNN-based architectures. For example, TimeSformer [4] and ViViT [1] study several variants of space-time factorization for the plain ViT. Some works [5, 47, 50] further explore how to reduce computational cost of the space-time attention. VideoSwin [42] and MViT [18, 38] study the hierarchical architecture and introduce an inductive locality bias into video transformers. Uniformer [36] and Video Mobile-Former [64] propose to integrate 3D CNNs and spatial-temporal attention mechanism for efficiency consideration. For convincing performance on the video understanding tasks, most video transformers require model weights pretrained on the large-scale image datasets. In this paper, we explore the self-supervised pretraining of video transformers and show pretraining strategy will significantly influence the downstream performance, which is orthogonal to the transformer architecture design." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6313" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "Self-supervised video representation learning. The early works [3, 45, 66, 70] of self-supervised video representation learning focus on designing the pretext tasks based on the temporal structure of videos. More recently, contrastive learning [8, 32, 37] that forces different views of the same image sample to be closer in the feature space while pushing the views of different images farther becomes a new paradigm of representation learning, and some works [11, 23, 24, 29, 30, 46, 49] design the contrastive learning methods on the video domain by exploring effective ways of spatial-temporal augmentations. However, as the learning supervision based on contrastive learning is applied on global representation, it cannot well model the local relationship or learn fine-grained local representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.473, + 0.746 + ], + "angle": 0, + "content": "Masked visual modeling. Masked language modeling [10, 40] has been one of the dominant pretraining methods of language transformers. With the success of vision transformers, masked visual modeling [2] has been introduced to self-supervised visual pretraining and demonstrates to be also helpful to multimodal visual-language learning [16, 74]. Following BERT [10] pretraining, BEiT [2] and PeCo [15] pretrain ViT by predicting the discrete visual tokens of masked patches, which are encoded by a pretrained VQ-VAE. MAE [31] proposes an asymmetric encoder-decoder framework for the reconstruction of pixels. SimMIM [68] and MaskFeat [67] propose to recover low-level features of masked patches like pixels or HOG features with hierarchical ViT. In contrast, some works [9, 14, 75] adopt an exponential moving average of the student model as the online teacher model, which makes the target features bootstrapped during training. In the video domain, some pioneering works [21, 55, 57, 63] extend masked image modeling to masked video modeling. BEVT [63] proposes a two-stream pretraining joint pretraining framework by predicting the discrete tokens with both image transformer and video transformer. VideoMAE [57] and ST-MAE [21] follow MAE and reconstruct the pixels of masked video patches with an extremely high masking ratio. Unlike most previous works of masked video modeling, our MVD focuses on masked feature modeling with high-level features as targets, and finds that student models using image and video teacher models will have different properties and complement each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Knowledge distillation. Knowledge distillation [26,33,48] aims to transfer the knowledge of the teacher model to the student model by adopting the output of the teacher model as the target for training the student model. Typical knowledge distillation works [33, 51, 52] mainly focus on supervised learning, e.g., image classification. The idea of knowledge distillation has also been leveraged in many self-supervised methods by learning from momentum encoders [6, 56] or different architectures [24]. Recently, self-supervised knowledge distillation [19, 71, 72] has also been" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "studied to learn representations from self-supervised pretrained models. In this paper, we present the first attempt that uses the masked image modeling pretrained image and video model as the masked feature prediction target in the video domain. It shows self-supervised MIM pretrained model can further bootstrap the mask video pretraining and bring significant performance gain." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.212, + 0.593, + 0.228 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.238, + 0.895, + 0.465 + ], + "angle": 0, + "content": "While masked video modeling has demonstrated promising performance for self-supervised learning, most existing approaches reconstruct relatively low-level information in the forms of raw pixels [57], low-level features like HOG [67] and VQVAE tokens [63]. In this paper, instead of reconstructing low-level information, we conduct masked video modeling at the feature-level. This is achieved by a two-stage framework, MVD, optimized to predict high-level features that derived from off-the-shelf MIM pretrained image models [31] and MVM pretrained video models [57] which are readily available. Below, we first given an overview of the masked feature modeling paradigm in Sec. 3.1 and then we introduce our proposed MVD in Sec. 3.2. Finally, we present the architectural design of MVD in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.475, + 0.874, + 0.492 + ], + "angle": 0, + "content": "3.1. The Paradigm of Masked Feature Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.895, + 0.696 + ], + "angle": 0, + "content": "The core of masked feature modeling is to train models to predict features of masked input regions. In this paper, we follow the decoupled encoder-decoder transformer architecture in MAE [31] due to its effectiveness and simplicity. An input \\(X\\) (image \\(X_{img} \\in \\mathbb{R}^{H \\times W \\times 3}\\) or video \\(X_{vid} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\)) is partitioned into several non-overlapping patches, and then each patch is mapped to a visual token with a linear projection layer. Before feeding tokens to the transformer encoder \\(f\\), a subset of tokens is masked and dropped from the token sequence. To reconstruct the information of masked tokens, the token sequence consisting of the visible tokens from the encoder and learnable mask tokens are input to a shallow transformer decoder \\(g\\):" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.71, + 0.892, + 0.728 + ], + "angle": 0, + "content": "\\[\nY = g \\left(\\operatorname {c o n c a t} \\left(f \\left(X _ {\\text {v i s}}\\right), T _ {m}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\( X_{vis} \\) denotes the visible input tokens, and \\( T_{m} \\) denotes the mask tokens. The subset of output tokens from decoder corresponded to input mask tokens contains reconstructed information of masked tokens. The reconstruction target for each masked patch \\( X(p) \\) is represented as a patch feature \\( h(X(p)) \\). Here, \\( h \\) is represents a function for generating the target features, e.g. \\( h \\) produces low-level RGB values of pixels in the patch in [21, 57]. Then, to train the encoder and the decoder, a loss function that measures the distance \\( D \\) between the ground-truth features of masked patches and reconstructed ones is defined as:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6314" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.833, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.895, + 0.38 + ], + "angle": 0, + "content": "Figure 2. An overview of MVD framework. Firstly the image teacher is pretrained by masked image modeling and the video teacher is pretrained by masked video modeling. Then the student model is trained from scratch to predict target high-level features encoded by the image teacher and the video teacher. The teacher models are fixed in the distillation stage." + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.401, + 0.47, + 0.439 + ], + "angle": 0, + "content": "\\[\nL _ {m f m} (h) = \\frac {1}{| M |} \\sum_ {p \\in M} D (Y (p), h (X (p))) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.471, + 0.489 + ], + "angle": 0, + "content": "where \\( p \\) is the token index and \\( M \\) is the set of masked tokens. For pixel regression in MAE [31] and VideoMAE [57], the L2 distance is used as the distance metric." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.495, + 0.314, + 0.51 + ], + "angle": 0, + "content": "3.2. Masked Video Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.471, + 0.838 + ], + "angle": 0, + "content": "In this paper, we propose Masked Video Distillation (MVD), which performs masked feature modeling on videos using high-level features as opposed to low-level pixels. In particular, we simply use outputs generated by off-the-shelf self-supervised pretrained image or video models, which are readily available, as reconstruction targets. These high-level features, serving as targets of the mask & prediction tasks, are encoded by teacher models pretrained by masked visual modeling like MAE or VideoMAE. For video representation learning, the reconstruction targets can take the form of spatial features encoded by image teacher models, or spatial-temporal features encoded by video teacher models. More specifically, the image teachers is pretrained by masked image modeling, while the video teacher is pretrained with masked video modeling, both of which aim at reconstructing raw pixels. Once trained, we use the image encoder \\( h_{img} \\) to generate the spatial targets, and the pretrained video transformer encoder \\( h_{vid} \\) to generate spatial-temporal targets. The loss function of MVD with the image teacher and video teacher can be denoted by \\( L_{mfm}(h_{img}) \\) and \\( L_{mfm}(h_{vid}) \\), respectively" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Spatial-temporal Co-teaching. When performing MVD with a single teacher, we observe that students distilled from different teachers learn different video representations and perform well on different kinds of downstream video" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.895, + 0.589 + ], + "angle": 0, + "content": "tasks. To improve the performance of MVD on different downstream video tasks, we propose spatial-temporal co-teaching that explores information from both image and video teachers such that the student model can handle videos of different types better. For instance, videos with fastly changing human actions require more temporal information while spatial clues might be sufficient for relatively static videos. To this end, MVD is trained to predict target high-level features produced by the image teacher and the video teacher at the same time. This is achieved by using two separated decoders to reconstruct different target features. The final loss of MVD with spatial-temporal co-teaching is:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.613, + 0.892, + 0.63 + ], + "angle": 0, + "content": "\\[\nL _ {m v d} = \\lambda_ {1} L _ {m f m} \\left(h _ {i m g}\\right) + \\lambda_ {2} L _ {m f m} \\left(h _ {v i d}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.894, + 0.684 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) denote the hyper-parameters that balance the weights of the image teacher and the video teacher. The pseudo code of MVD is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.697, + 0.708 + ], + "angle": 0, + "content": "3.3. Architectural Design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Encoder. For MVD, the vanilla transformer backbone is used as the encoder. For a video input \\( X_{vid} \\in \\mathbb{R}^{T \\times H \\times W \\times 3} \\), we adopt 3D patch embedding with a patch size of \\( 2 \\times 16 \\times 16 \\). After patch partitioning and linear embedding, we obtain \\( T / 2 \\times H / 16 \\times W / 16 \\) tokens. For the masked feature modeling, the tokens are masked with a high masking ratio and the remaining tokens are fed into the transformer layers. For finetuning on downstream tasks, we input all tokens to the subsequent layers. In each layer, joint spatial-temporal self-attention is applied on the whole input token sequence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Mask strategy. For MVD, we follow [57] and adopt tube masking for masked feature modeling. First a 2D random" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6315" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.091, + 0.423, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of MVD in PyTorch style." + }, + { + "type": "code", + "bbox": [ + 0.078, + 0.117, + 0.468, + 0.377 + ], + "angle": 0, + "content": "f: student encoder \n# g.img: decoder for reconstructing spatial features \n# g_vid: decoder for reconstructing spatial-temporal features \n# t_m: learnable mask tokens \n# h.img: image teacher model \n# h_vid: video teacher model \nfor x, m in loader: # x: video data, m: mask \nx_pe = patch_emb(x) # patch embedding of input \nx.vis = mask_select(x_pe, 1 - m) # masking tokens \nq.vis = f(x.vis) # visible local patch features \n# reconstruction of target features \np IMG = g.img(concat(q.vis, t_m)) \np_vid = g_vid(concat(q.vis, t_m)) \n# compute target features with teacher models \nk IMG = h.img(x) # target spatial features \nk_vid = h_vid(x) # target spatial-temporal features \n# compute reconstruction loss \nloss IMG = smooth_L1_loss(p.img ⊙ m, k.img ⊙ m) \nloss_vid = smooth_L1_loss(p:vid ⊙ m, k:vid ⊙ m) \nloss = λ1 * loss IMG + λ2 * loss_vid \nloss.backup() \noptimizer_STEP() # optimizer update" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.471, + 0.501 + ], + "angle": 0, + "content": "mask is generated and then extended along the temporal dimension. Therefore, the spatial mask on each time slice is the same, which prevents information leakage between frames. Tube masking with a high masking ratio (e.g., \\(90\\%\\)) encourages the video transformer to model the high-level semantics during pretraining." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Decoder. For MVD, shallow decoders consist of vanilla transformer layers and a linear projection layer. The transformer layers in decoders are the same as those in the encoder. Since spatial-temporal co-teaching introduces two different reconstruction targets for masked feature modeling, two separated decoders that share the same architecture but contain different weights are placed on the top of the encoder. Learnable masked tokens corresponded to masked patches are concatenated with visible tokens from the encoder before fed into the decoder. After jointly modeling the spatial-temporal relationships, the output tokens of transformer layers are mapped to final predictions by the linear projection layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Reconstruction targets. For generating spatial-temporal target features, the video teacher, which shares the same architecture as the student model, is pretrained by a VideoMAE [57] manner on the video dataset. For obtaining spatial targets, we adopt the vanilla image ViT pretrained by masked image modeling [31] on the image dataset (e.g., ImageNet-1K). It is worth noting that one 3D patch (with size of \\(2 \\times 16 \\times 16\\)) for the video transformer corresponds to two 2D patches (with size of \\(16 \\times 16\\)) for the image transformer. Following [21], we predict the spatial features of a single time slice (that is the front 2D patch), which reduce the prediction layer's size." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.894, + 0.178 + ], + "angle": 0, + "content": "In this section, we first introduce the experimental setup in Sec. 4.1, and then present the main results in Sec. 4.2, followed by an extensive analysis to verify the effectiveness of different components in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.188, + 0.691, + 0.205 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.895, + 0.469 + ], + "angle": 0, + "content": "Dataset. We pretrain the vanilla ViT with MVD on Kinetics-400 by default, and evaluate the learned model on four video recognition downstream tasks: (a) Kinetics-400 (K400) [7], which consists of \\(\\sim 240K\\) training videos and \\(\\sim 20K\\) validation videos with an average duration of 10 seconds. All video clips are labeled into 400 classes. (b) Something-Something V2 (SSv2) [27], which contains \\(\\sim 160K\\) videos for training and \\(\\sim 20K\\) videos for validation. The videos in SSv2 with an average duration of 4 seconds are labeled into 174 motion-centric categories. (c) UCF-101 [54] is a relatively small dataset, consisting of \\(\\sim 9.5K\\) training videos and \\(\\sim 3.5K\\) validation videos. (d) HMDB51 [35] is also a small video dataset that contains around \\(3.5K/1.5K\\) train/val videos. We also transfer pretrained models to a challenging spatial-temporal action detection dataset AVA [28]. The results of AVA are included in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.895, + 0.685 + ], + "angle": 0, + "content": "Implementation details. Our MVD is performed on vanilla ViTs with different capacities (i.e., ViT-S, ViT-B, ViT-L). By default, image teacher models are pretrained on ImageNet-1K for 1600 epochs and video teachers are pretrained on K400 for 1600 epochs. We follow the training strategy in MAE [31] and VideoMAE [21] for image teachers and video teachers respectively. In the distillation stage, student models are first pretrained from scratch on K400 for 400 epochs unless mentioned otherwise. The resulting models are then finetuned on downstream video tasks. The video clip length is 16 for both pretraining and finetuning. We adopt AdamW optimizer [44] and Smooth L1 loss for the optimization of student models. More details are included in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.64, + 0.712 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Students distilled from different teachers. Unlike masked image modeling, masked feature modeling on video data has more choices on reconstruction targets. Besides spatial features, to include temporal dynamics in the reconstruction target, we can also adopt spatial-temporal features encoded by pretrained video models. In Table 1, we compare students distilled by the image teacher and the video teacher on K400, a downstream task that mainly relies on spatial clues, and SSv2, a temporally-heavy downstream task. Our observations are as follow: (a) Masked feature modeling with high-level features as targets achieves convincing performance on downstream video tasks, and outperforms Video" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6316" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.46, + 0.213 + ], + "angle": 0, + "content": "
studentteachersK400 top-1SSv2 top-1
imagevideo
ViT-S✓✗80.469.4
✗✓80.170.0
✓✓80.670.7
ViT-B✓✗82.371.4
✗✓82.171.8
✓✓82.772.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.215, + 0.47, + 0.271 + ], + "angle": 0, + "content": "Table 1. Co-teaching with the image teacher and the video teacher outperforms distillation with a single teacher in MVD. We adopt ViT-B as the architecture of teachers and distill students for 400 epochs here." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.276, + 0.472, + 0.382 + ], + "angle": 0, + "content": "
studentteacherK400 top-1SSv2 top-1
ViMAEMVDViMAEMVD
ViT-SViT-B79.080.6 ↑1.666.470.7 ↑4.3
ViT-SViT-L79.081.0 ↑2.066.470.9 ↑4.5
ViT-BViT-B81.582.7 ↑1.269.772.5 ↑2.8
ViT-BViT-L81.583.4 ↑1.969.773.7 ↑4.0
ViT-LViT-L85.286.0 ↑0.874.076.1 ↑2.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.384, + 0.47, + 0.44 + ], + "angle": 0, + "content": "Table 2. MVD achieves significant improvement compared with VideoMAE across different model scales. MVD is pretrained on K400 for 400 epochs, while VideoMAE is pretrained on K400 for 1600 epochs here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.47, + 0.692 + ], + "angle": 0, + "content": "MAE baseline significantly (compared with the baseline results in Table 2). In particular, with both image and video teachers using a ViT-S as the backbone, MVD achieves consistent gains over VideoMAE on both K400 (80.6% vs. 79.0%) and SSV2 (70.7% vs. 66.4%). (b) Students distilled from the image teacher achieve higher top-1 accuracy on K400, while students distilled from the video teacher perform better on SSv2. For example, ViT-S achieves an accuracy of 80.4% and 69.4% using an image teacher on K400 and SSv2 respectively. With a video teacher, on the other hand, ViT-S offers a top-1 accuracy of 80.1% and 70.0% respectively. As it has been shown that videos in K400 less sensitive to temporal modeling compared to SSv2, the results demonstrate students learn stronger spatial representation from the image teacher while the video teacher transfer more knowledge about temporal dynamics to students." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.468, + 0.835 + ], + "angle": 0, + "content": "Co-teaching outperforms distilling with a single teacher. To improve the performance on different kinds of downstream video tasks, we introduce spatial-temporal co-teaching in MVD, which trains the model to predict spatial features and spatial-temporal features of masked patches in a decoupled way. The results in Table 1 indicate that students distilled with spatial-temporal co-teaching outperforms students distilled from either single teacher on both spatially-heavy task and temporally-heavy task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.9 + ], + "angle": 0, + "content": "MVD outperforms VideoMAE baseline significantly. In Table 2, MVD with spatial-temporal co-teaching is compared with VideoMAE pretrained on K400. When the size of teacher models is the same as that of students, MVD out" + }, + { + "type": "table", + "bbox": [ + 0.498, + 0.089, + 0.895, + 0.595 + ], + "angle": 0, + "content": "
methodextra datatop-1top-5GFLOPsParam
supervised
NL I3D R101 [65]-77.393.3359×3062
ip-CSN-152 [59]-77.892.8109×3033
SlowFast NL [22]-79.893.9234×3060
X3D-XL [20]-79.193.948×3011
MViTv1-B [18]-80.294.4170×537
VideoSwin-B [42]IN-1K80.694.6282×1288
Uniformer-B [36]IN-1K83.095.4259×1250
TimeSformer [4]IN-21K80.794.72380×3121
Mformer-B [47]IN-21K79.794.2370×30109
Mformer-L [47]IN-21K80.294.81185×30382
ViViT-L FE [1]IN-21K81.793.83980×3N/A
VideoSwin-L [42]IN-21K83.195.9604×12197
self-supervised
VIMPAC ViT-L [55]HowTo100M77.4N/AN/A×30307
BEVT Swin-B [63]IN-1K81.1N/A282×1288
MaskFeat MViT-S [67]-82.295.171×1036
VideoMAE ViT-S [57]-79.093.857×1522
VideoMAE ViT-B [57]-81.595.1180×1587
VideoMAE ViT-L [57]-85.296.8597×15305
VideoMAE ViT-H [57]-86.697.11192×15633
ST-MAE ViT-B [21]-81.394.9180×2187
ST-MAE ViT-L [21]-84.896.2598×21304
ST-MAE ViT-H [21]-85.196.61193×21632
OmniMAE ViT-B [25]IN-1K80.8N/A180×1587
OmniMAE ViT-L [25]IN1K+SSv284.0N/A597×15305
OmniMAE ViT-H [25]IN1K+SSv284.8N/A1192×15633
MVD-S (Teacher-B)IN-1K80.694.757×1522
MVD-S (Teacher-L)IN-1K81.094.857×1522
MVD-B (Teacher-B)IN-1K82.795.4180×1587
MVD-B (Teacher-L)IN-1K83.495.8180×1587
MVD-L (Teacher-L)IN-1K86.096.9597×15305
MVD-L (Teacher-L) †IN-1K86.497.0597×15305
MVD-H (Teacher-H) †IN-1K87.397.41192×15633
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.68 + ], + "angle": 0, + "content": "Table 3. Comparison with previous works on Kinetics-400. \"MVD-S (Teacher-B)\" denotes that the student is ViT-S and the teacher is ViT-B. \\(\\dagger\\) denotes that the student is distilled for 800 epochs instead of 400 epochs. The inference cost is reported with the cost of a single view \\(\\times\\) the number of views. Note that IN-1K is only used for training image teachers in MVD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.817 + ], + "angle": 0, + "content": "performs VideoMAE by a clear margin on both K400 and SSv2. Larger models as teachers can further boost the performance of MVD. It is worth mentioning that not only is our MVD particularly effective for relatively small models, but also it improves the performance of large vision models like ViT-L. For example, with a ViT-L model as the student, MVD achieves \\(86.0\\%\\) and \\(76.1\\%\\) on K400 and SSv2, surpassing VideoMAE by \\(0.8\\%\\) and \\(2.1\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Comparison with state-of-the-art. We compare MVD with prior studies on four video recognition tasks. Results on K400 are shown in Table 3. Our MVD outperforms previous self-supervised methods with similar or less computational cost. Even compared with video transform" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6317" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.075, + 0.089, + 0.473, + 0.685 + ], + "angle": 0, + "content": "
methodextra datatop-1GFLOPsParam
supervised
SlowFast R101 [22]K40063.1106×353
TSM-RGB R50 [39]IN-1K63.362×624
TAM R50 [43]IN-1K66.099×651
TDN R101 [61]IN-1K69.6198×388
MViTv1-B [18]-67.7455×337
MViTv2-B [38]K40070.5225×351
Uniformer-B [36]K40071.2259×350
TimeSformer-HR [4]IN-21K62.51703×3121
ViViT-L FE [1]IN21K+K40065.9995×12N/A
Mformer-B [47]IN21K+K40066.5370×3109
Mformer-L [47]IN21K+K40068.11185×3382
VideoSwin-B [42]IN21K+K40069.6321×388
MViTv2-L [38]IN21K+K40073.32828×3213
self-supervised
VIMPAC ViT-L [55]HowTo100M68.1N/A×30307
BEVT Swin-B [63]IN1K+K40071.4321×388
MaskFeat MViT-L [67]K40074.42828×3218
VideoMAE ViT-S [57]K40066.457×622
VideoMAE ViT-S [57]-66.857×622
VideoMAE ViT-B [57]K40069.7180×687
VideoMAE ViT-B [57]-70.8180×687
VideoMAE ViT-L [57]K40074.0597×6305
VideoMAE ViT-L [57]-74.3597×6305
ST-MAE ViT-L [21]K40072.1598×3304
ST-MAE ViT-H [21]K40074.11193×3632
OmniMAE ViT-B [25]IN1K+K40069.0180×687
OmniMAE ViT-L [25]IN-1K74.2597×6305
OmniMAE ViT-H [25]IN-1K75.31192×6632
MVD-S (Teacher-B)IN1K+K40070.757×622
MVD-S (Teacher-L)IN1K+K40070.957×622
MVD-B (Teacher-B)IN1K+K40072.5180×687
MVD-B (Teacher-L)IN1K+K40073.7180×687
MVD-L (Teacher-L)IN1K+K40076.1597×6305
MVD-L (Teacher-L) †IN1K+K40076.7597×6305
MVD-H (Teacher-H) †IN1K+K40077.31192×6633
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Table 4. Comparison with previous works on Something-Something V2.. † denotes that the student is distilled for 800 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "ers pretrained on ImageNet-21K, MVD achieves superior performance. Particularly, MVD-H achieves an accuracy of \\(87.3\\%\\) on K400, outperforming previous top-performing methods by clear margins. Table 4 presents comparisons to the state-of-the-art methods on SSv2. We observe that for downstream tasks depending on temporal relationship modeling, self-supervised methods based on MVM achieve better performance in comparison with supervised methods (cf. results in the middle group of Table 4 vs. results in the top group). Once again, our MVD, producing an accuracy of \\(76.1\\%\\) with a large model, beats both supervised meth" + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.089, + 0.897, + 0.261 + ], + "angle": 0, + "content": "
methodextra dataParamUCF101HMDB51
VideoMoCo R2+1D [46]K4001578.749.2
MemDPC R2D3D [30]K4003286.154.5
Vi2CLR S3D [11]K400989.155.7
CORP Slow-R50 [34]K4003293.568.0
CVRL Slow-R50 [49]K4003292.967.9
CVRL Slow-R152 [49]K60032894.470.6
ρBYOL Slow-R50 [23]K4003294.272.1
VIMPAC ViT-L [55]HowTo100M30792.765.9
VideoMAE ViT-B [57]K4008796.173.3
MVD-B (Teacher-B)IN-1K+K4008797.076.4
MVD-B (Teacher-L)IN-1K+K4008797.579.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Table 5. Comparison with previous state-of-the-art methods on UCF101 and HMDB51." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.301, + 0.682, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.417, + 0.642, + 0.429 + ], + "angle": 0, + "content": "(a) image teacher." + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.3, + 0.888, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.417, + 0.847, + 0.428 + ], + "angle": 0, + "content": "(b) video teacher." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.432, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Figure 3. Feature similarity across different frames for different teacher models. Similarity matrices are computed on the Kinetics-400 validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.574 + ], + "angle": 0, + "content": "ods and self-supervised methods by clear margins. With more training epochs (i.e., 800 epochs), MVD with ViT-L achieves more significant performance gains (i.e., \\(1.2\\%\\), \\(2.4\\%\\) on K400 and SSv2) compared with VideoMAE. When a huge model is used, the performance can still be boosted and MVD achieves \\(77.3\\%\\) top-1 accuracy on SSv2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.727 + ], + "angle": 0, + "content": "We also evaluate the transfer learning ability of MVD on two relatively small datasets, UCF101 and HMDB51. As shown in Table 5, MVD with ViT-B obtains higher accuracy compared with prior works based on well-designed pretext tasks, contrastive learning and masked video modeling methods. Especially compared to the original VideoMAE [57] ViT-B teacher model, we achieve \\(0.9\\%\\) and \\(3.1\\%\\) higher points on these two datasets respectively. Additionally, when teachers with larger size are adopted, MVD achieves stronger transfer learning performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.72, + 0.755 + ], + "angle": 0, + "content": "4.3. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.891, + 0.794 + ], + "angle": 0, + "content": "In this section, we provide an in-depth analysis of the effectiveness of different components in MVD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Analysis of features encoded by different teachers. The properties of target features generated by different teachers may influence the performance of students on different downstream tasks. To quantify the temporal dynamics that teacher models capture from the input video, we study the similarity between feature maps across different frames of each input video clip via the cosine similarity. As similarity" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6318" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.457, + 0.152 + ], + "angle": 0, + "content": "
methodepochtime(h)K400 top-1
VideoMAE80010781.0
VideoMAE160021481.5
MVD(800+) 400(107+) 5781.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.154, + 0.47, + 0.209 + ], + "angle": 0, + "content": "Table 6. Training time comparison between MVD and VideoMAE on 32 NVIDIA V100 GPUs. The training time of teacher models is considered. we only adopt a video teacher pretrained for 800 epochs in MVD here. ViT-B is used here." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.214, + 0.462, + 0.293 + ], + "angle": 0, + "content": "
teachersreconstruct pixelSSv2 top-1
image modelX68.7
image model✓67.9 ↓0.8
image+video modelsX70.1
image+video models✓69.0 ↓1.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.295, + 0.468, + 0.336 + ], + "angle": 0, + "content": "Table 7. Ablations on reconstructing low-level features in masked video distillation. We distill ViT-S students with ViT-B teachers for 300 epochs here." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.338, + 0.455, + 0.447 + ], + "angle": 0, + "content": "
teacherinitepochtop-1
K400SSv2
momentum encoderX80080.570.4
momentum encoder✓80081.870.8
fixed image modelX40082.371.4
fixed video modelX40082.171.8
fixed co-teachingX40082.772.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.448, + 0.468, + 0.476 + ], + "angle": 0, + "content": "Table 8. Comparison with bootstrapped teachers and students initialized with IN1K-pretrained models. ViT-B is used here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.468, + 0.59 + ], + "angle": 0, + "content": "matrices shown in Figure 3, for image teachers, the feature maps of different frames are almost the same. However, for video teachers, the features of different frames have larger difference. This indicates that video teachers capture more temporal difference. Therefore, students distilled from video teachers can learn stronger temporal dynamics and perform better on temporally-heavy downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.468, + 0.716 + ], + "angle": 0, + "content": "Training time comparison. We study whether MVD is able to achieve better balance of accuracy and efficiency than VideoMAE. For fair comparisons, the training time of teacher models is also counted in the total training time of MVD. Results are shown in Table 6. We see that MVD can achieve better accuracy (i.e., \\(81.9\\%\\)) by a total of 164 hours of training which is 50 hours less than VideoMAE (producing an accuracy of \\(81.5\\%\\)) trained for 1600 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Reconstruction signals in MVD. In MVD, we first pretrain teacher models by recovering the pixels of masked patches in a MAE manner, then adopt features produced by teachers as the targets of masked feature modeling. In Table 7, we study whether to include an additional decoder branch for reconstructing pixels of masked patches in the distillation stage. The experimental results show that for both MVD with a single teacher and MVD with spatial-temporal co-teaching, the reconstruction of low-level feature targets degrades the performance for downstream tasks. Therefore, we only reconstruct high-level features of masked patches in the distillation stage of MVD." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.874, + 0.137 + ], + "angle": 0, + "content": "
distillation methodK400 top-1SSv2 top-1
per-token distillation80.970.5
masked reconstruction82.171.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.139, + 0.892, + 0.166 + ], + "angle": 0, + "content": "Table 9. Comparison with feature distillation without masked reconstruction. We distill ViT-B from the video teacher here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.18, + 0.892, + 0.436 + ], + "angle": 0, + "content": "Comparison with bootstrapped teachers. Some recent works of image representation learning [9, 14, 16, 75] adopt features of a momentum encoder as the targets of MIM, while we use frozen teacher models in MVD. In Table 8, we compare fixed teachers with bootstrapped teachers that are updated by an exponential moving average of the online encoder during pretraining. According to a MIM method [14], two strong baselines of bootstrapped teachers are built for video representation learning: (a) The student model is trained from scratch with masked feature modeling, and the target features are generated by a momentum encoder. (b) The framework of bootstrapped teachers is first pretrained on IN-1K for 800 epochs, then the pretrained weights are adopted to initialize the video pretraining. As shown in Table 8, MVD with the frozen teacher beats the method with a bootstrapped teacher on the downstream video tasks, even if only a single teacher is utilized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Comparison with feature distillation. In the previous work of self-supervised feature distillation [19, 26, 72], the distillation loss is directly computed upon the full feature maps between teachers and students. Accordingly, we use a baseline method named per-token distillation. Specifically, the output features of students are projected by a MLP, and then forced to mimicking the teacher's feature at each token with a Smooth L1 loss. As shown in Table 9, masked feature reconstruction in our MVD outperforms per-token distillation on both K400 and SSv2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.615, + 0.619, + 0.63 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.868 + ], + "angle": 0, + "content": "In this paper, we study masked video distillation upon MIM pretrained image or video transformers. We have three interesting findings: 1) Using MIM pretrained image transformers or MVM pretrained video transformers as teachers supervise masked feature prediction can significantly boost the finetuning performance on video downstream tasks; 2) The representation distilled with image and video teachers will have different properties, i.e., image teachers will benefit spatial-heavy video tasks more while video teachers benefit temporal-heavy video tasks more; 3) Combining image and video teachers will enjoy the synergy and thus produce higher performance. Even though the proposed masked video distillation seems very straightforward, we hope such interesting findings can motivate more thinking about masked video pretraining." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement This project was supported by NSFC under Grant No. 62032006 and No. 62102092." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6319" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.198 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. In ICLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.468, + 0.257 + ], + "angle": 0, + "content": "[3] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[5] Adrian Bulat, Juan Manuel Perez Rua, Swathikiran Sudhakaran, Brais Martinez, and Georgios Tzimiropoulos. Space-time mixing attention for video transformer. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.468, + 0.413 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.468, + 0.457 + ], + "angle": 0, + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.468, + 0.5 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.468, + 0.544 + ], + "angle": 0, + "content": "[9] Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Sdae: Self-distillated masked autoencoder. In ECCV, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.468, + 0.587 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.468, + 0.643 + ], + "angle": 0, + "content": "[11] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saqib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2clr: Video and image for visual contrastive learning of representation. In ICCV, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[12] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[13] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[14] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In ECCV, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "of vision transformers. In Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[16] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.271 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[19] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation. arXiv preprint arXiv:2101.04731, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[20] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In CVPR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[21] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[22] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[23] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.527, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[24] Chongjian Ge, Youwei Liang, Yibing Song, Jianbo Jiao, Jue Wang, and Ping Luo. Revitalizing cnn attention via transformers in self-supervised visual representation learning. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[25] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[26] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. IJCV, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.68, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[27] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fründ, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[28] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[29] Sheng Guo, Zihua Xiong, Yujie Zhong, Limin Wang, Xiaobo Guo, Bing Han, and Weilin Huang. Cross-architecture self-supervised video representation learning. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6320" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[30] Tengda Han, Weidi Xie, and Andrew Zisserman. Memory-augmented dense predictive coding for video representation learning. In ECCV, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[32] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[33] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[34] Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[35] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[36] Kunchang Li, Yali Wang, Gao Peng, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatial-temporal representation learning. In ICLR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[37] Suichan Li, Dongdong Chen, Yinpeng Chen, Lu Yuan, Lei Zhang, Qi Chu, Bin Liu, and Nenghai Yu. Improve unsupervised pretraining for few-label transfer. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[38] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[39] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[40] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[42] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[43] Zhaoyang Liu, Limin Wang, Wayne Wu, Chen Qian, and Tong Lu. Tam: Temporal adaptive module for video recognition. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[45] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[46] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In CVPR, 2021. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[47] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In NeurIPS, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[48] Mary Phuong and Christoph Lampert. Towards understanding knowledge distillation. In ICML, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[49] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[50] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[51] Zhiqiang Shen and Marios Savvides. Meal v2: Boosting vanilla resnet-50 to \\(80\\%+\\) top-1 accuracy on imagenet without tricks. arXiv preprint arXiv:2009.08453, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[52] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. arXiv preprint arXiv:2112.01528, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[53] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[54] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[55] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[56] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Piriavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[57] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[58] Du Tran, Lubomir D Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. C3d: Generic features for video analysis. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[59] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[60] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[61] Limin Wang, Zhan Tong, Bin Ji, and Gangshan Wu. Tdn: Temporal difference networks for efficient action recognition. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[62] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE TPAMI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[63] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, 2022. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6321" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[64] Rui Wang, Zuxuan Wu, Dongdong Chen, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Luowei Zhou, Lu Yuan, and YuGang Jiang. Video mobile-former: Video recognition with efficient global spatial-temporal modeling. arXiv preprint arXiv:2208.12257, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[65] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[66] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[67] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[68] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[69] Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. arXiv preprint arXiv:2211.13222, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[70] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[71] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[72] Haohang Xu, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, Xinggang Wang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Bag of instances aggregation boosts self-supervised distillation. In ICLR, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[73] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[74] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6322" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_origin.pdf b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..152c567091c5478b66a87c0e6368a1db00c95acf --- /dev/null +++ b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/5401dde5-fc89-4f9e-a63d-583f6d1c59d3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e20dafe36ff3f2fd082577b31d1eb793f7964a52ea1b676556148aec8db04c0 +size 1699143 diff --git a/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/full.md b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..11ef658b5705fa581bf5671809ffc48b28406a67 --- /dev/null +++ b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/full.md @@ -0,0 +1,301 @@ +# Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning + +Rui Wang $^{1,2}$ Dongdong Chen $^{3}$ Zuxuan Wu $^{1,2\dagger}$ Yinpeng Chen $^{3}$ Xiyang Dai $^{3}$ Mengchen Liu $^{3}$ Lu Yuan $^{3}$ Yu-Gang Jiang $^{1,2\dagger}$ $^{1}$ Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University + $^{2}$ Shanghai Collaborative Innovation Center of Intelligent Visual Computing + $^{3}$ Microsoft Cloud + AI + +# Abstract + +Benefiting from masked visual modeling, self-supervised video representation learning has achieved remarkable progress. However, existing methods focus on learning representations from scratch through reconstructing low-level features like raw pixel values. In this paper, we propose masked video distillation (MVD), a simple yet effective two-stage masked feature modeling framework for video representation learning: firstly we pretrain an image (or video) model by recovering low-level features of masked patches, then we use the resulting features as targets for masked feature modeling. For the choice of teacher models, we observe that students taught by video teachers perform better on temporally-heavy video tasks, while image teachers transfer stronger spatial representations for spatially-heavy video tasks. Visualization analysis also indicates different teachers produce different learned patterns for students. To leverage the advantage of different teachers, we design a spatial-temporal co-teaching method for MVD. Specifically, we distill student models from both video teachers and image teachers by masked feature modeling. Extensive experimental results demonstrate that video transformers pretrained with spatial-temporal co-teaching outperform models distilled with a single teacher on a multitude of video datasets. Our MVD with vanilla ViT achieves state-of-the-art performance compared with previous methods on several challenging video downstream tasks. For example, with the ViT-Large model, our MVD achieves $86.4\%$ and $76.7\%$ Top-1 accuracy on Kinetics-400 and Something-Something-v2, outperforming VideoMAE by $1.2\%$ and $2.4\%$ respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with $77.3\%$ Top-1 accuracy on Something-Something-v2. Code will be available at https://github.com/ruiwang2021/mvd. + +![](images/e64338a607bb510d4dd0847dfea409c7f50ef859da431d2a17f508f233496e57.jpg) +Figure 1. Comparisons of MVD with previous supervised or self-supervised methods on Something-Something v2. Each line represents the corresponding model of different sizes. + +# 1. Introduction + +For self-supervised visual representation learning, recent masked image modeling (MIM) methods like MAE [31] and BEiT [2] achieve promising results with vision transformers [17] on various vision downstream tasks. Such a pretraining paradigm has also been adapted to the video domain and boosts video transformers by clear margins compared with supervised pretraining on several video downstream tasks. Representative masked video modeling (MVM) works include BEVT [63], VideoMAE [57] and ST-MAE [21]. + +Following MAE [31] and BEiT [2], existing masked video modeling methods [21, 57, 63] pretrain video transformers through reconstructing low-level features, e.g., raw pixel values or low-level VQVAE tokens. However, using low-level features as reconstruction targets often incur much noise. And due to the high redundancy in video data, it is easy for masked video modeling to learn shortcuts, thus resulting in limited transfer performance on downstream + +tasks. To alleviate this issue, masked video modeling [57] often uses larger masking ratios. + +In this paper, we observe that much better performance on video downstream tasks can be achieved by conducting masked feature prediction by using the high-level features of pretrained MIM and MVM models as masked prediction targets. This can be viewed as two-stage masked video modeling, where MIM pretrained image models (i.e., an image teacher) or MVM pretrained video models (i.e., an video teacher) are obtained in the first stage, and they further act as teachers in the second stage for the student model via providing the high-level feature targets. Therefore, we call this method Masked Video Distillation (MVD). + +More interestingly, we find that student models distilled with different teachers in MVD exhibit different properties on different video downstream tasks. Specifically, students distilled from the image teacher perform better on video tasks that mainly rely on spatial clues, while students distilled from the video teacher model perform better on the video downstream tasks where temporal dynamics are more necessary. We think during the pretraining process of masked video modeling in the first stage, video teachers have learned spatial-temporal context in their high-level features. Therefore, when employing such high-level representations as prediction targets of masked feature modeling, it will help encouraging the student model to learn stronger temporal dynamics. By analogy, image teachers provide high-level features as targets that include more spatial information, which can help the student model learn more spatially meaningful representations. We further analyze the feature targets provided by image teachers and video teachers, and calculate the cross-frame feature similarity. It shows that the features provided by the video teachers contain more temporal dynamics. + +Motivated by the above observation, to leverage the advantages of video teachers and image teachers, we propose a simple yet effective spatial-temporal co-teaching strategy for MVD. In detail, the student model is designed to reconstruct the features coming from both the image teacher and video teacher with two different decoders, so as to learn stronger spatial representation and temporal dynamics at the same time. Experiments demonstrate that MVD with co-teaching from both the image teacher and the video teacher significantly outperforms MVD only using one single teacher on several challenging downstream tasks. + +Despite the simplicity, our MVD is super effective and achieves very strong performance on multiple standard video recognition benchmarks. For example, on Kinectics-400 and Something-Something-v2 datasets, compared to the baseline without distillation, MVD with 400 epochs using a teacher model of the same size achieves $1.2\%$ , $2.8\%$ Top-1 accuracy gain on ViT-B. If a larger teacher model ViT-L is used, more significant performance gains (i.e., + +$1.9\%$ , $4.0\%$ can be obtained. When ViT-Large is the target student model, our method can achieve $86.4\%$ and $76.7\%$ Top-1 accuracy on these two datasets, surpassing existing state-of-the-art method VideoMAE [57] by $1.2\%$ and $2.4\%$ respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with $77.3\%$ Top-1 accuracy on Something-Something-v2. + +Our contributions can be summarized as below: + +- We find that using MIM pretrained image models and MVM pretrained video models as teachers to provide the high-level features for continued masked feature prediction can learn better video representation. And representations learned with image teachers and video teachers show different properties on different downstream video datasets. +- We propose masked video distillation together with a simple yet effective co-teaching strategy, which enjoys the synergy of image and video teachers. +- We demonstrate strong performance on multiple standard video recognition benchmarks, surpassing both the baseline without MVD and prior state-of-the-art methods by clear margins. + +# 2. Related Work + +Vision transformers for video understanding. For video understanding tasks, modeling the spatial-temporal information is the most important factor to consider in the architecture design. In the early works of video understanding, common video architectures, e.g., 3D CNNs [7, 20, 22, 58, 60] and 2D CNNs with temporal modules [12, 39, 53, 62, 73], are designed by extending existing 2D CNNs on the temporal dimension. Recently, Vision Transformers [13, 17, 41] achieve significant progress on several computer vision tasks. Some works [4, 18, 42, 69] also adapt vision transformers to the video domain and achieve superior performance compared to previous CNN-based architectures. For example, TimeSformer [4] and ViViT [1] study several variants of space-time factorization for the plain ViT. Some works [5, 47, 50] further explore how to reduce computational cost of the space-time attention. VideoSwin [42] and MViT [18, 38] study the hierarchical architecture and introduce an inductive locality bias into video transformers. Uniformer [36] and Video Mobile-Former [64] propose to integrate 3D CNNs and spatial-temporal attention mechanism for efficiency consideration. For convincing performance on the video understanding tasks, most video transformers require model weights pretrained on the large-scale image datasets. In this paper, we explore the self-supervised pretraining of video transformers and show pretraining strategy will significantly influence the downstream performance, which is orthogonal to the transformer architecture design. + +Self-supervised video representation learning. The early works [3, 45, 66, 70] of self-supervised video representation learning focus on designing the pretext tasks based on the temporal structure of videos. More recently, contrastive learning [8, 32, 37] that forces different views of the same image sample to be closer in the feature space while pushing the views of different images farther becomes a new paradigm of representation learning, and some works [11, 23, 24, 29, 30, 46, 49] design the contrastive learning methods on the video domain by exploring effective ways of spatial-temporal augmentations. However, as the learning supervision based on contrastive learning is applied on global representation, it cannot well model the local relationship or learn fine-grained local representation. + +Masked visual modeling. Masked language modeling [10, 40] has been one of the dominant pretraining methods of language transformers. With the success of vision transformers, masked visual modeling [2] has been introduced to self-supervised visual pretraining and demonstrates to be also helpful to multimodal visual-language learning [16, 74]. Following BERT [10] pretraining, BEiT [2] and PeCo [15] pretrain ViT by predicting the discrete visual tokens of masked patches, which are encoded by a pretrained VQ-VAE. MAE [31] proposes an asymmetric encoder-decoder framework for the reconstruction of pixels. SimMIM [68] and MaskFeat [67] propose to recover low-level features of masked patches like pixels or HOG features with hierarchical ViT. In contrast, some works [9, 14, 75] adopt an exponential moving average of the student model as the online teacher model, which makes the target features bootstrapped during training. In the video domain, some pioneering works [21, 55, 57, 63] extend masked image modeling to masked video modeling. BEVT [63] proposes a two-stream pretraining joint pretraining framework by predicting the discrete tokens with both image transformer and video transformer. VideoMAE [57] and ST-MAE [21] follow MAE and reconstruct the pixels of masked video patches with an extremely high masking ratio. Unlike most previous works of masked video modeling, our MVD focuses on masked feature modeling with high-level features as targets, and finds that student models using image and video teacher models will have different properties and complement each other. + +Knowledge distillation. Knowledge distillation [26,33,48] aims to transfer the knowledge of the teacher model to the student model by adopting the output of the teacher model as the target for training the student model. Typical knowledge distillation works [33, 51, 52] mainly focus on supervised learning, e.g., image classification. The idea of knowledge distillation has also been leveraged in many self-supervised methods by learning from momentum encoders [6, 56] or different architectures [24]. Recently, self-supervised knowledge distillation [19, 71, 72] has also been + +studied to learn representations from self-supervised pretrained models. In this paper, we present the first attempt that uses the masked image modeling pretrained image and video model as the masked feature prediction target in the video domain. It shows self-supervised MIM pretrained model can further bootstrap the mask video pretraining and bring significant performance gain. + +# 3. Method + +While masked video modeling has demonstrated promising performance for self-supervised learning, most existing approaches reconstruct relatively low-level information in the forms of raw pixels [57], low-level features like HOG [67] and VQVAE tokens [63]. In this paper, instead of reconstructing low-level information, we conduct masked video modeling at the feature-level. This is achieved by a two-stage framework, MVD, optimized to predict high-level features that derived from off-the-shelf MIM pretrained image models [31] and MVM pretrained video models [57] which are readily available. Below, we first given an overview of the masked feature modeling paradigm in Sec. 3.1 and then we introduce our proposed MVD in Sec. 3.2. Finally, we present the architectural design of MVD in Sec. 3.3. + +# 3.1. The Paradigm of Masked Feature Modeling + +The core of masked feature modeling is to train models to predict features of masked input regions. In this paper, we follow the decoupled encoder-decoder transformer architecture in MAE [31] due to its effectiveness and simplicity. An input $X$ (image $X_{img} \in \mathbb{R}^{H \times W \times 3}$ or video $X_{vid} \in \mathbb{R}^{T \times H \times W \times 3}$ ) is partitioned into several non-overlapping patches, and then each patch is mapped to a visual token with a linear projection layer. Before feeding tokens to the transformer encoder $f$ , a subset of tokens is masked and dropped from the token sequence. To reconstruct the information of masked tokens, the token sequence consisting of the visible tokens from the encoder and learnable mask tokens are input to a shallow transformer decoder $g$ : + +$$ +Y = g \left(\operatorname {c o n c a t} \left(f \left(X _ {\text {v i s}}\right), T _ {m}\right)\right), \tag {1} +$$ + +where $X_{vis}$ denotes the visible input tokens, and $T_{m}$ denotes the mask tokens. The subset of output tokens from decoder corresponded to input mask tokens contains reconstructed information of masked tokens. The reconstruction target for each masked patch $X(p)$ is represented as a patch feature $h(X(p))$ . Here, $h$ is represents a function for generating the target features, e.g. $h$ produces low-level RGB values of pixels in the patch in [21, 57]. Then, to train the encoder and the decoder, a loss function that measures the distance $D$ between the ground-truth features of masked patches and reconstructed ones is defined as: + +![](images/42876f42afae95eccbfa8b9eeba9ec9aa8139a1c1c013a96cca151ad7b2822d8.jpg) +Figure 2. An overview of MVD framework. Firstly the image teacher is pretrained by masked image modeling and the video teacher is pretrained by masked video modeling. Then the student model is trained from scratch to predict target high-level features encoded by the image teacher and the video teacher. The teacher models are fixed in the distillation stage. + +$$ +L _ {m f m} (h) = \frac {1}{| M |} \sum_ {p \in M} D (Y (p), h (X (p))) \tag {2} +$$ + +where $p$ is the token index and $M$ is the set of masked tokens. For pixel regression in MAE [31] and VideoMAE [57], the L2 distance is used as the distance metric. + +# 3.2. Masked Video Distillation + +In this paper, we propose Masked Video Distillation (MVD), which performs masked feature modeling on videos using high-level features as opposed to low-level pixels. In particular, we simply use outputs generated by off-the-shelf self-supervised pretrained image or video models, which are readily available, as reconstruction targets. These high-level features, serving as targets of the mask & prediction tasks, are encoded by teacher models pretrained by masked visual modeling like MAE or VideoMAE. For video representation learning, the reconstruction targets can take the form of spatial features encoded by image teacher models, or spatial-temporal features encoded by video teacher models. More specifically, the image teachers is pretrained by masked image modeling, while the video teacher is pretrained with masked video modeling, both of which aim at reconstructing raw pixels. Once trained, we use the image encoder $h_{img}$ to generate the spatial targets, and the pretrained video transformer encoder $h_{vid}$ to generate spatial-temporal targets. The loss function of MVD with the image teacher and video teacher can be denoted by $L_{mfm}(h_{img})$ and $L_{mfm}(h_{vid})$ , respectively + +Spatial-temporal Co-teaching. When performing MVD with a single teacher, we observe that students distilled from different teachers learn different video representations and perform well on different kinds of downstream video + +tasks. To improve the performance of MVD on different downstream video tasks, we propose spatial-temporal co-teaching that explores information from both image and video teachers such that the student model can handle videos of different types better. For instance, videos with fastly changing human actions require more temporal information while spatial clues might be sufficient for relatively static videos. To this end, MVD is trained to predict target high-level features produced by the image teacher and the video teacher at the same time. This is achieved by using two separated decoders to reconstruct different target features. The final loss of MVD with spatial-temporal co-teaching is: + +$$ +L _ {m v d} = \lambda_ {1} L _ {m f m} \left(h _ {i m g}\right) + \lambda_ {2} L _ {m f m} \left(h _ {v i d}\right) \tag {3} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ denote the hyper-parameters that balance the weights of the image teacher and the video teacher. The pseudo code of MVD is shown in Algorithm 1. + +# 3.3. Architectural Design + +Encoder. For MVD, the vanilla transformer backbone is used as the encoder. For a video input $X_{vid} \in \mathbb{R}^{T \times H \times W \times 3}$ , we adopt 3D patch embedding with a patch size of $2 \times 16 \times 16$ . After patch partitioning and linear embedding, we obtain $T / 2 \times H / 16 \times W / 16$ tokens. For the masked feature modeling, the tokens are masked with a high masking ratio and the remaining tokens are fed into the transformer layers. For finetuning on downstream tasks, we input all tokens to the subsequent layers. In each layer, joint spatial-temporal self-attention is applied on the whole input token sequence. + +Mask strategy. For MVD, we follow [57] and adopt tube masking for masked feature modeling. First a 2D random + +Algorithm 1 Pseudocode of MVD in PyTorch style. +```python +f: student encoder +# g.img: decoder for reconstructing spatial features +# g_vid: decoder for reconstructing spatial-temporal features +# t_m: learnable mask tokens +# h.img: image teacher model +# h_vid: video teacher model +for x, m in loader: # x: video data, m: mask +x_pe = patch_emb(x) # patch embedding of input +x.vis = mask_select(x_pe, 1 - m) # masking tokens +q.vis = f(x.vis) # visible local patch features +# reconstruction of target features +p IMG = g.img(concat(q.vis, t_m)) +p_vid = g_vid(concat(q.vis, t_m)) +# compute target features with teacher models +k IMG = h.img(x) # target spatial features +k_vid = h_vid(x) # target spatial-temporal features +# compute reconstruction loss +loss IMG = smooth_L1_loss(p.img ⊙ m, k.img ⊙ m) +loss_vid = smooth_L1_loss(p:vid ⊙ m, k:vid ⊙ m) +loss = λ1 * loss IMG + λ2 * loss_vid +loss.backup() +optimizer_STEP() # optimizer update +``` + +mask is generated and then extended along the temporal dimension. Therefore, the spatial mask on each time slice is the same, which prevents information leakage between frames. Tube masking with a high masking ratio (e.g., $90\%$ ) encourages the video transformer to model the high-level semantics during pretraining. + +Decoder. For MVD, shallow decoders consist of vanilla transformer layers and a linear projection layer. The transformer layers in decoders are the same as those in the encoder. Since spatial-temporal co-teaching introduces two different reconstruction targets for masked feature modeling, two separated decoders that share the same architecture but contain different weights are placed on the top of the encoder. Learnable masked tokens corresponded to masked patches are concatenated with visible tokens from the encoder before fed into the decoder. After jointly modeling the spatial-temporal relationships, the output tokens of transformer layers are mapped to final predictions by the linear projection layer. + +Reconstruction targets. For generating spatial-temporal target features, the video teacher, which shares the same architecture as the student model, is pretrained by a VideoMAE [57] manner on the video dataset. For obtaining spatial targets, we adopt the vanilla image ViT pretrained by masked image modeling [31] on the image dataset (e.g., ImageNet-1K). It is worth noting that one 3D patch (with size of $2 \times 16 \times 16$ ) for the video transformer corresponds to two 2D patches (with size of $16 \times 16$ ) for the image transformer. Following [21], we predict the spatial features of a single time slice (that is the front 2D patch), which reduce the prediction layer's size. + +# 4. Experiments + +In this section, we first introduce the experimental setup in Sec. 4.1, and then present the main results in Sec. 4.2, followed by an extensive analysis to verify the effectiveness of different components in Sec. 4.3. + +# 4.1. Experimental Setup + +Dataset. We pretrain the vanilla ViT with MVD on Kinetics-400 by default, and evaluate the learned model on four video recognition downstream tasks: (a) Kinetics-400 (K400) [7], which consists of $\sim 240K$ training videos and $\sim 20K$ validation videos with an average duration of 10 seconds. All video clips are labeled into 400 classes. (b) Something-Something V2 (SSv2) [27], which contains $\sim 160K$ videos for training and $\sim 20K$ videos for validation. The videos in SSv2 with an average duration of 4 seconds are labeled into 174 motion-centric categories. (c) UCF-101 [54] is a relatively small dataset, consisting of $\sim 9.5K$ training videos and $\sim 3.5K$ validation videos. (d) HMDB51 [35] is also a small video dataset that contains around $3.5K/1.5K$ train/val videos. We also transfer pretrained models to a challenging spatial-temporal action detection dataset AVA [28]. The results of AVA are included in supplementary materials. + +Implementation details. Our MVD is performed on vanilla ViTs with different capacities (i.e., ViT-S, ViT-B, ViT-L). By default, image teacher models are pretrained on ImageNet-1K for 1600 epochs and video teachers are pretrained on K400 for 1600 epochs. We follow the training strategy in MAE [31] and VideoMAE [21] for image teachers and video teachers respectively. In the distillation stage, student models are first pretrained from scratch on K400 for 400 epochs unless mentioned otherwise. The resulting models are then finetuned on downstream video tasks. The video clip length is 16 for both pretraining and finetuning. We adopt AdamW optimizer [44] and Smooth L1 loss for the optimization of student models. More details are included in supplementary materials. + +# 4.2. Main Results + +Students distilled from different teachers. Unlike masked image modeling, masked feature modeling on video data has more choices on reconstruction targets. Besides spatial features, to include temporal dynamics in the reconstruction target, we can also adopt spatial-temporal features encoded by pretrained video models. In Table 1, we compare students distilled by the image teacher and the video teacher on K400, a downstream task that mainly relies on spatial clues, and SSv2, a temporally-heavy downstream task. Our observations are as follow: (a) Masked feature modeling with high-level features as targets achieves convincing performance on downstream video tasks, and outperforms Video + +
studentteachersK400 top-1SSv2 top-1
imagevideo
ViT-S✓✗80.469.4
✗✓80.170.0
✓✓80.670.7
ViT-B✓✗82.371.4
✗✓82.171.8
✓✓82.772.5
+ +Table 1. Co-teaching with the image teacher and the video teacher outperforms distillation with a single teacher in MVD. We adopt ViT-B as the architecture of teachers and distill students for 400 epochs here. + +
studentteacherK400 top-1SSv2 top-1
ViMAEMVDViMAEMVD
ViT-SViT-B79.080.6 ↑1.666.470.7 ↑4.3
ViT-SViT-L79.081.0 ↑2.066.470.9 ↑4.5
ViT-BViT-B81.582.7 ↑1.269.772.5 ↑2.8
ViT-BViT-L81.583.4 ↑1.969.773.7 ↑4.0
ViT-LViT-L85.286.0 ↑0.874.076.1 ↑2.1
+ +MAE baseline significantly (compared with the baseline results in Table 2). In particular, with both image and video teachers using a ViT-S as the backbone, MVD achieves consistent gains over VideoMAE on both K400 (80.6% vs. 79.0%) and SSV2 (70.7% vs. 66.4%). (b) Students distilled from the image teacher achieve higher top-1 accuracy on K400, while students distilled from the video teacher perform better on SSv2. For example, ViT-S achieves an accuracy of 80.4% and 69.4% using an image teacher on K400 and SSv2 respectively. With a video teacher, on the other hand, ViT-S offers a top-1 accuracy of 80.1% and 70.0% respectively. As it has been shown that videos in K400 less sensitive to temporal modeling compared to SSv2, the results demonstrate students learn stronger spatial representation from the image teacher while the video teacher transfer more knowledge about temporal dynamics to students. + +Co-teaching outperforms distilling with a single teacher. To improve the performance on different kinds of downstream video tasks, we introduce spatial-temporal co-teaching in MVD, which trains the model to predict spatial features and spatial-temporal features of masked patches in a decoupled way. The results in Table 1 indicate that students distilled with spatial-temporal co-teaching outperforms students distilled from either single teacher on both spatially-heavy task and temporally-heavy task. + +MVD outperforms VideoMAE baseline significantly. In Table 2, MVD with spatial-temporal co-teaching is compared with VideoMAE pretrained on K400. When the size of teacher models is the same as that of students, MVD out + +Table 2. MVD achieves significant improvement compared with VideoMAE across different model scales. MVD is pretrained on K400 for 400 epochs, while VideoMAE is pretrained on K400 for 1600 epochs here. + +
methodextra datatop-1top-5GFLOPsParam
supervised
NL I3D R101 [65]-77.393.3359×3062
ip-CSN-152 [59]-77.892.8109×3033
SlowFast NL [22]-79.893.9234×3060
X3D-XL [20]-79.193.948×3011
MViTv1-B [18]-80.294.4170×537
VideoSwin-B [42]IN-1K80.694.6282×1288
Uniformer-B [36]IN-1K83.095.4259×1250
TimeSformer [4]IN-21K80.794.72380×3121
Mformer-B [47]IN-21K79.794.2370×30109
Mformer-L [47]IN-21K80.294.81185×30382
ViViT-L FE [1]IN-21K81.793.83980×3N/A
VideoSwin-L [42]IN-21K83.195.9604×12197
self-supervised
VIMPAC ViT-L [55]HowTo100M77.4N/AN/A×30307
BEVT Swin-B [63]IN-1K81.1N/A282×1288
MaskFeat MViT-S [67]-82.295.171×1036
VideoMAE ViT-S [57]-79.093.857×1522
VideoMAE ViT-B [57]-81.595.1180×1587
VideoMAE ViT-L [57]-85.296.8597×15305
VideoMAE ViT-H [57]-86.697.11192×15633
ST-MAE ViT-B [21]-81.394.9180×2187
ST-MAE ViT-L [21]-84.896.2598×21304
ST-MAE ViT-H [21]-85.196.61193×21632
OmniMAE ViT-B [25]IN-1K80.8N/A180×1587
OmniMAE ViT-L [25]IN1K+SSv284.0N/A597×15305
OmniMAE ViT-H [25]IN1K+SSv284.8N/A1192×15633
MVD-S (Teacher-B)IN-1K80.694.757×1522
MVD-S (Teacher-L)IN-1K81.094.857×1522
MVD-B (Teacher-B)IN-1K82.795.4180×1587
MVD-B (Teacher-L)IN-1K83.495.8180×1587
MVD-L (Teacher-L)IN-1K86.096.9597×15305
MVD-L (Teacher-L) †IN-1K86.497.0597×15305
MVD-H (Teacher-H) †IN-1K87.397.41192×15633
+ +Table 3. Comparison with previous works on Kinetics-400. "MVD-S (Teacher-B)" denotes that the student is ViT-S and the teacher is ViT-B. $\dagger$ denotes that the student is distilled for 800 epochs instead of 400 epochs. The inference cost is reported with the cost of a single view $\times$ the number of views. Note that IN-1K is only used for training image teachers in MVD. + +performs VideoMAE by a clear margin on both K400 and SSv2. Larger models as teachers can further boost the performance of MVD. It is worth mentioning that not only is our MVD particularly effective for relatively small models, but also it improves the performance of large vision models like ViT-L. For example, with a ViT-L model as the student, MVD achieves $86.0\%$ and $76.1\%$ on K400 and SSv2, surpassing VideoMAE by $0.8\%$ and $2.1\%$ , respectively. + +Comparison with state-of-the-art. We compare MVD with prior studies on four video recognition tasks. Results on K400 are shown in Table 3. Our MVD outperforms previous self-supervised methods with similar or less computational cost. Even compared with video transform + +
methodextra datatop-1GFLOPsParam
supervised
SlowFast R101 [22]K40063.1106×353
TSM-RGB R50 [39]IN-1K63.362×624
TAM R50 [43]IN-1K66.099×651
TDN R101 [61]IN-1K69.6198×388
MViTv1-B [18]-67.7455×337
MViTv2-B [38]K40070.5225×351
Uniformer-B [36]K40071.2259×350
TimeSformer-HR [4]IN-21K62.51703×3121
ViViT-L FE [1]IN21K+K40065.9995×12N/A
Mformer-B [47]IN21K+K40066.5370×3109
Mformer-L [47]IN21K+K40068.11185×3382
VideoSwin-B [42]IN21K+K40069.6321×388
MViTv2-L [38]IN21K+K40073.32828×3213
self-supervised
VIMPAC ViT-L [55]HowTo100M68.1N/A×30307
BEVT Swin-B [63]IN1K+K40071.4321×388
MaskFeat MViT-L [67]K40074.42828×3218
VideoMAE ViT-S [57]K40066.457×622
VideoMAE ViT-S [57]-66.857×622
VideoMAE ViT-B [57]K40069.7180×687
VideoMAE ViT-B [57]-70.8180×687
VideoMAE ViT-L [57]K40074.0597×6305
VideoMAE ViT-L [57]-74.3597×6305
ST-MAE ViT-L [21]K40072.1598×3304
ST-MAE ViT-H [21]K40074.11193×3632
OmniMAE ViT-B [25]IN1K+K40069.0180×687
OmniMAE ViT-L [25]IN-1K74.2597×6305
OmniMAE ViT-H [25]IN-1K75.31192×6632
MVD-S (Teacher-B)IN1K+K40070.757×622
MVD-S (Teacher-L)IN1K+K40070.957×622
MVD-B (Teacher-B)IN1K+K40072.5180×687
MVD-B (Teacher-L)IN1K+K40073.7180×687
MVD-L (Teacher-L)IN1K+K40076.1597×6305
MVD-L (Teacher-L) †IN1K+K40076.7597×6305
MVD-H (Teacher-H) †IN1K+K40077.31192×6633
+ +ers pretrained on ImageNet-21K, MVD achieves superior performance. Particularly, MVD-H achieves an accuracy of $87.3\%$ on K400, outperforming previous top-performing methods by clear margins. Table 4 presents comparisons to the state-of-the-art methods on SSv2. We observe that for downstream tasks depending on temporal relationship modeling, self-supervised methods based on MVM achieve better performance in comparison with supervised methods (cf. results in the middle group of Table 4 vs. results in the top group). Once again, our MVD, producing an accuracy of $76.1\%$ with a large model, beats both supervised meth + +Table 4. Comparison with previous works on Something-Something V2.. † denotes that the student is distilled for 800 epochs. + +
methodextra dataParamUCF101HMDB51
VideoMoCo R2+1D [46]K4001578.749.2
MemDPC R2D3D [30]K4003286.154.5
Vi2CLR S3D [11]K400989.155.7
CORP Slow-R50 [34]K4003293.568.0
CVRL Slow-R50 [49]K4003292.967.9
CVRL Slow-R152 [49]K60032894.470.6
ρBYOL Slow-R50 [23]K4003294.272.1
VIMPAC ViT-L [55]HowTo100M30792.765.9
VideoMAE ViT-B [57]K4008796.173.3
MVD-B (Teacher-B)IN-1K+K4008797.076.4
MVD-B (Teacher-L)IN-1K+K4008797.579.7
+ +Table 5. Comparison with previous state-of-the-art methods on UCF101 and HMDB51. + +![](images/b85514161767a24e5ba9d65e7a6ebec3da134586a2592035b7fe27a40a8bdffc.jpg) +(a) image teacher. + +![](images/7a74256dfcf3e265a50a769507214dc95a131cbb2672bc6fde43b3c3c6ba27ea.jpg) +(b) video teacher. +Figure 3. Feature similarity across different frames for different teacher models. Similarity matrices are computed on the Kinetics-400 validation set. + +ods and self-supervised methods by clear margins. With more training epochs (i.e., 800 epochs), MVD with ViT-L achieves more significant performance gains (i.e., $1.2\%$ , $2.4\%$ on K400 and SSv2) compared with VideoMAE. When a huge model is used, the performance can still be boosted and MVD achieves $77.3\%$ top-1 accuracy on SSv2. + +We also evaluate the transfer learning ability of MVD on two relatively small datasets, UCF101 and HMDB51. As shown in Table 5, MVD with ViT-B obtains higher accuracy compared with prior works based on well-designed pretext tasks, contrastive learning and masked video modeling methods. Especially compared to the original VideoMAE [57] ViT-B teacher model, we achieve $0.9\%$ and $3.1\%$ higher points on these two datasets respectively. Additionally, when teachers with larger size are adopted, MVD achieves stronger transfer learning performance. + +# 4.3. Analysis and Discussion + +In this section, we provide an in-depth analysis of the effectiveness of different components in MVD. + +Analysis of features encoded by different teachers. The properties of target features generated by different teachers may influence the performance of students on different downstream tasks. To quantify the temporal dynamics that teacher models capture from the input video, we study the similarity between feature maps across different frames of each input video clip via the cosine similarity. As similarity + +
methodepochtime(h)K400 top-1
VideoMAE80010781.0
VideoMAE160021481.5
MVD(800+) 400(107+) 5781.9
+ +Table 6. Training time comparison between MVD and VideoMAE on 32 NVIDIA V100 GPUs. The training time of teacher models is considered. we only adopt a video teacher pretrained for 800 epochs in MVD here. ViT-B is used here. + +
teachersreconstruct pixelSSv2 top-1
image modelX68.7
image model✓67.9 ↓0.8
image+video modelsX70.1
image+video models✓69.0 ↓1.1
+ +Table 7. Ablations on reconstructing low-level features in masked video distillation. We distill ViT-S students with ViT-B teachers for 300 epochs here. + +
teacherinitepochtop-1
K400SSv2
momentum encoderX80080.570.4
momentum encoder✓80081.870.8
fixed image modelX40082.371.4
fixed video modelX40082.171.8
fixed co-teachingX40082.772.5
+ +matrices shown in Figure 3, for image teachers, the feature maps of different frames are almost the same. However, for video teachers, the features of different frames have larger difference. This indicates that video teachers capture more temporal difference. Therefore, students distilled from video teachers can learn stronger temporal dynamics and perform better on temporally-heavy downstream tasks. + +Training time comparison. We study whether MVD is able to achieve better balance of accuracy and efficiency than VideoMAE. For fair comparisons, the training time of teacher models is also counted in the total training time of MVD. Results are shown in Table 6. We see that MVD can achieve better accuracy (i.e., $81.9\%$ ) by a total of 164 hours of training which is 50 hours less than VideoMAE (producing an accuracy of $81.5\%$ ) trained for 1600 epochs. + +Reconstruction signals in MVD. In MVD, we first pretrain teacher models by recovering the pixels of masked patches in a MAE manner, then adopt features produced by teachers as the targets of masked feature modeling. In Table 7, we study whether to include an additional decoder branch for reconstructing pixels of masked patches in the distillation stage. The experimental results show that for both MVD with a single teacher and MVD with spatial-temporal co-teaching, the reconstruction of low-level feature targets degrades the performance for downstream tasks. Therefore, we only reconstruct high-level features of masked patches in the distillation stage of MVD. + +Table 8. Comparison with bootstrapped teachers and students initialized with IN1K-pretrained models. ViT-B is used here. + +
distillation methodK400 top-1SSv2 top-1
per-token distillation80.970.5
masked reconstruction82.171.8
+ +Table 9. Comparison with feature distillation without masked reconstruction. We distill ViT-B from the video teacher here. + +Comparison with bootstrapped teachers. Some recent works of image representation learning [9, 14, 16, 75] adopt features of a momentum encoder as the targets of MIM, while we use frozen teacher models in MVD. In Table 8, we compare fixed teachers with bootstrapped teachers that are updated by an exponential moving average of the online encoder during pretraining. According to a MIM method [14], two strong baselines of bootstrapped teachers are built for video representation learning: (a) The student model is trained from scratch with masked feature modeling, and the target features are generated by a momentum encoder. (b) The framework of bootstrapped teachers is first pretrained on IN-1K for 800 epochs, then the pretrained weights are adopted to initialize the video pretraining. As shown in Table 8, MVD with the frozen teacher beats the method with a bootstrapped teacher on the downstream video tasks, even if only a single teacher is utilized. + +Comparison with feature distillation. In the previous work of self-supervised feature distillation [19, 26, 72], the distillation loss is directly computed upon the full feature maps between teachers and students. Accordingly, we use a baseline method named per-token distillation. Specifically, the output features of students are projected by a MLP, and then forced to mimicking the teacher's feature at each token with a Smooth L1 loss. As shown in Table 9, masked feature reconstruction in our MVD outperforms per-token distillation on both K400 and SSv2. + +# 5. Conclusion + +In this paper, we study masked video distillation upon MIM pretrained image or video transformers. We have three interesting findings: 1) Using MIM pretrained image transformers or MVM pretrained video transformers as teachers supervise masked feature prediction can significantly boost the finetuning performance on video downstream tasks; 2) The representation distilled with image and video teachers will have different properties, i.e., image teachers will benefit spatial-heavy video tasks more while video teachers benefit temporal-heavy video tasks more; 3) Combining image and video teachers will enjoy the synergy and thus produce higher performance. Even though the proposed masked video distillation seems very straightforward, we hope such interesting findings can motivate more thinking about masked video pretraining. + +Acknowledgement This project was supported by NSFC under Grant No. 62032006 and No. 62102092. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2, 6, 7 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. In ICLR, 2022. 1, 3 +[3] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In CVPR, 2020. 3 +[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 2, 6, 7 +[5] Adrian Bulat, Juan Manuel Perez Rua, Swathikiran Sudhakaran, Brais Martinez, and Georgios Tzimiropoulos. Space-time mixing attention for video transformer. In NeurIPS, 2021. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3 +[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 2, 5 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3 +[9] Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Sdae: Self-distillated masked autoencoder. In ECCV, 2022. 3, 8 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3 +[11] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saqib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2clr: Video and image for visual contrastive learning of representation. In ICCV, 2021. 3, 7 +[12] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2 +[13] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. 2 +[14] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In ECCV, 2022. 3, 8 +[15] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training + +of vision transformers. In Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI), 2023. 3 +[16] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3, 8 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2 +[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 2, 6, 7 +[19] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation. arXiv preprint arXiv:2101.04731, 2021. 3, 8 +[20] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In CVPR, 2020. 2, 6 +[21] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. 1, 3, 5, 6, 7 +[22] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2, 6, 7 +[23] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, 2021. 3, 7 +[24] Chongjian Ge, Youwei Liang, Yibing Song, Jianbo Jiao, Jue Wang, and Ping Luo. Revitalizing cnn attention via transformers in self-supervised visual representation learning. NeurIPS, 2021. 3 +[25] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 6, 7 +[26] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. IJCV, 2021. 3, 8 +[27] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fründ, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The "something something" video database for learning and evaluating visual common sense. In ICCV, 2017. 5 +[28] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 5 +[29] Sheng Guo, Zihua Xiong, Yujie Zhong, Limin Wang, Xiaobo Guo, Bing Han, and Weilin Huang. Cross-architecture self-supervised video representation learning. In CVPR, 2022. 3 + +[30] Tengda Han, Weidi Xie, and Andrew Zisserman. Memory-augmented dense predictive coding for video representation learning. In ECCV, 2020. 3, 7 +[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 3, 4, 5 +[32] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3 +[33] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[34] Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. In ICCV, 2021. 7 +[35] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 5 +[36] Kunchang Li, Yali Wang, Gao Peng, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatial-temporal representation learning. In ICLR, 2022. 2, 6, 7 +[37] Suichan Li, Dongdong Chen, Yinpeng Chen, Lu Yuan, Lei Zhang, Qi Chu, Bin Liu, and Nenghai Yu. Improve unsupervised pretraining for few-label transfer. In ICCV, 2021. 3 +[38] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In CVPR, 2022. 2, 7 +[39] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 2, 7 +[40] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3 +[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2 +[42] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 2, 6, 7 +[43] Zhaoyang Liu, Limin Wang, Wayne Wu, Chen Qian, and Tong Lu. Tam: Temporal adaptive module for video recognition. In CVPR, 2021. 7 +[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018. 5 +[45] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. 3 +[46] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In CVPR, 2021. 3, 7 + +[47] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In NeurIPS, 2021. 2, 6, 7 +[48] Mary Phuong and Christoph Lampert. Towards understanding knowledge distillation. In ICML, 2019. 3 +[49] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, 2021. 3, 7 +[50] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In NeurIPS, 2021. 2 +[51] Zhiqiang Shen and Marios Savvides. Meal v2: Boosting vanilla resnet-50 to $80\%+$ top-1 accuracy on imagenet without tricks. arXiv preprint arXiv:2009.08453, 2020. 3 +[52] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. arXiv preprint arXiv:2112.01528, 2021. 3 +[53] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014. 2 +[54] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[55] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 3, 6, 7 +[56] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Piriavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, 2021. 3 +[57] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7 +[58] Du Tran, Lubomir D Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. C3d: Generic features for video analysis. In ICCV, 2015. 2 +[59] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 6 +[60] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 2 +[61] Limin Wang, Zhan Tong, Bin Ji, and Gangshan Wu. Tdn: Temporal difference networks for efficient action recognition. In CVPR, 2021. 7 +[62] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE TPAMI, 2018. 2 +[63] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, 2022. 1, 3, 6, 7 + +[64] Rui Wang, Zuxuan Wu, Dongdong Chen, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Luowei Zhou, Lu Yuan, and YuGang Jiang. Video mobile-former: Video recognition with efficient global spatial-temporal modeling. arXiv preprint arXiv:2208.12257, 2022. 2 +[65] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 6 +[66] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, 2015. 3 +[67] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. 3, 6, 7 +[68] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 3 +[69] Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. arXiv preprint arXiv:2211.13222, 2022. 2 +[70] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019. 3 +[71] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, 2020. 3 +[72] Haohang Xu, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, Xinggang Wang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Bag of instances aggregation boosts self-supervised distillation. In ICLR, 2021. 3, 8 +[73] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015. 2 +[74] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. 3 +[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 3, 8 \ No newline at end of file diff --git a/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/images.zip b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44b932b42508e75eecf4b254a26531e068f8caae --- /dev/null +++ b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aefabe46162f5c601162ca934bc9ac81678d49904dfa088ed0c272695c85d292 +size 656139 diff --git a/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/layout.json b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5b6e637974a880e7b065ad44f3d26344b4eb9be2 --- /dev/null +++ b/2023/Masked Video Distillation_ Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning/layout.json @@ -0,0 +1,8090 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 102, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 102, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 510, + 140 + ], + "type": "text", + "content": "Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "spans": [ + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": "Rui Wang" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": " Dongdong Chen" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": " Zuxuan Wu" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": " Yinpeng Chen" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": " Xiyang Dai" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": " Mengchen Liu" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": " Lu Yuan" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": " Yu-Gang Jiang" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": "Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University \n" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": "Shanghai Collaborative Innovation Center of Intelligent Visual Computing \n" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 159, + 500, + 232 + ], + "type": "text", + "content": "Microsoft Cloud + AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "text", + "content": "Benefiting from masked visual modeling, self-supervised video representation learning has achieved remarkable progress. However, existing methods focus on learning representations from scratch through reconstructing low-level features like raw pixel values. In this paper, we propose masked video distillation (MVD), a simple yet effective two-stage masked feature modeling framework for video representation learning: firstly we pretrain an image (or video) model by recovering low-level features of masked patches, then we use the resulting features as targets for masked feature modeling. For the choice of teacher models, we observe that students taught by video teachers perform better on temporally-heavy video tasks, while image teachers transfer stronger spatial representations for spatially-heavy video tasks. Visualization analysis also indicates different teachers produce different learned patterns for students. To leverage the advantage of different teachers, we design a spatial-temporal co-teaching method for MVD. Specifically, we distill student models from both video teachers and image teachers by masked feature modeling. Extensive experimental results demonstrate that video transformers pretrained with spatial-temporal co-teaching outperform models distilled with a single teacher on a multitude of video datasets. Our MVD with vanilla ViT achieves state-of-the-art performance compared with previous methods on several challenging video downstream tasks. For example, with the ViT-Large model, our MVD achieves " + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "inline_equation", + "content": "86.4\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "inline_equation", + "content": "76.7\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "text", + "content": " Top-1 accuracy on Kinetics-400 and Something-Something-v2, outperforming VideoMAE by " + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "text", + "content": " respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with " + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "inline_equation", + "content": "77.3\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 681 + ], + "type": "text", + "content": " Top-1 accuracy on Something-Something-v2. Code will be available at https://github.com/ruiwang2021/mvd." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 257, + 543, + 429 + ], + "blocks": [ + { + "bbox": [ + 307, + 257, + 543, + 429 + ], + "lines": [ + { + "bbox": [ + 307, + 257, + 543, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 543, + 429 + ], + "type": "image", + "image_path": "e64338a607bb510d4dd0847dfea409c7f50ef859da431d2a17f508f233496e57.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 431, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 431, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 547, + 464 + ], + "type": "text", + "content": "Figure 1. Comparisons of MVD with previous supervised or self-supervised methods on Something-Something v2. Each line represents the corresponding model of different sizes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 476, + 387, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 387, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 387, + 489 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 497, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 616 + ], + "type": "text", + "content": "For self-supervised visual representation learning, recent masked image modeling (MIM) methods like MAE [31] and BEiT [2] achieve promising results with vision transformers [17] on various vision downstream tasks. Such a pretraining paradigm has also been adapted to the video domain and boosts video transformers by clear margins compared with supervised pretraining on several video downstream tasks. Representative masked video modeling (MVM) works include BEVT [63], VideoMAE [57] and ST-MAE [21]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": "Following MAE [31] and BEiT [2], existing masked video modeling methods [21, 57, 63] pretrain video transformers through reconstructing low-level features, e.g., raw pixel values or low-level VQVAE tokens. However, using low-level features as reconstruction targets often incur much noise. And due to the high redundancy in video data, it is easy for masked video modeling to learn shortcuts, thus resulting in limited transfer performance on downstream" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 141, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 141, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 141, + 712 + ], + "type": "text", + "content": "†Corresponding authors" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6312" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "tasks. To alleviate this issue, masked video modeling [57] often uses larger masking ratios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 287, + 229 + ], + "type": "text", + "content": "In this paper, we observe that much better performance on video downstream tasks can be achieved by conducting masked feature prediction by using the high-level features of pretrained MIM and MVM models as masked prediction targets. This can be viewed as two-stage masked video modeling, where MIM pretrained image models (i.e., an image teacher) or MVM pretrained video models (i.e., an video teacher) are obtained in the first stage, and they further act as teachers in the second stage for the student model via providing the high-level feature targets. Therefore, we call this method Masked Video Distillation (MVD)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 288, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 482 + ], + "type": "text", + "content": "More interestingly, we find that student models distilled with different teachers in MVD exhibit different properties on different video downstream tasks. Specifically, students distilled from the image teacher perform better on video tasks that mainly rely on spatial clues, while students distilled from the video teacher model perform better on the video downstream tasks where temporal dynamics are more necessary. We think during the pretraining process of masked video modeling in the first stage, video teachers have learned spatial-temporal context in their high-level features. Therefore, when employing such high-level representations as prediction targets of masked feature modeling, it will help encouraging the student model to learn stronger temporal dynamics. By analogy, image teachers provide high-level features as targets that include more spatial information, which can help the student model learn more spatially meaningful representations. We further analyze the feature targets provided by image teachers and video teachers, and calculate the cross-frame feature similarity. It shows that the features provided by the video teachers contain more temporal dynamics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 616 + ], + "type": "text", + "content": "Motivated by the above observation, to leverage the advantages of video teachers and image teachers, we propose a simple yet effective spatial-temporal co-teaching strategy for MVD. In detail, the student model is designed to reconstruct the features coming from both the image teacher and video teacher with two different decoders, so as to learn stronger spatial representation and temporal dynamics at the same time. Experiments demonstrate that MVD with co-teaching from both the image teacher and the video teacher significantly outperforms MVD only using one single teacher on several challenging downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Despite the simplicity, our MVD is super effective and achieves very strong performance on multiple standard video recognition benchmarks. For example, on Kinectics-400 and Something-Something-v2 datasets, compared to the baseline without distillation, MVD with 400 epochs using a teacher model of the same size achieves " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " Top-1 accuracy gain on ViT-B. If a larger teacher model ViT-L is used, more significant performance gains (i.e.," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " can be obtained. When ViT-Large is the target student model, our method can achieve " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "86.4\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "76.7\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " Top-1 accuracy on these two datasets, surpassing existing state-of-the-art method VideoMAE [57] by " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "77.3\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " Top-1 accuracy on Something-Something-v2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 156, + 511, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 156, + 511, + 167 + ], + "spans": [ + { + "bbox": [ + 317, + 156, + 511, + 167 + ], + "type": "text", + "content": "Our contributions can be summarized as below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 175, + 545, + 357 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 175, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 545, + 258 + ], + "type": "text", + "content": "- We find that using MIM pretrained image models and MVM pretrained video models as teachers to provide the high-level features for continued masked feature prediction can learn better video representation. And representations learned with image teachers and video teachers show different properties on different downstream video datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 266, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 545, + 301 + ], + "type": "text", + "content": "- We propose masked video distillation together with a simple yet effective co-teaching strategy, which enjoys the synergy of image and video teachers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 309, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 545, + 357 + ], + "type": "text", + "content": "- We demonstrate strong performance on multiple standard video recognition benchmarks, surpassing both the baseline without MVD and prior state-of-the-art methods by clear margins." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 367, + 392, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 392, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 392, + 380 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 713 + ], + "type": "text", + "content": "Vision transformers for video understanding. For video understanding tasks, modeling the spatial-temporal information is the most important factor to consider in the architecture design. In the early works of video understanding, common video architectures, e.g., 3D CNNs [7, 20, 22, 58, 60] and 2D CNNs with temporal modules [12, 39, 53, 62, 73], are designed by extending existing 2D CNNs on the temporal dimension. Recently, Vision Transformers [13, 17, 41] achieve significant progress on several computer vision tasks. Some works [4, 18, 42, 69] also adapt vision transformers to the video domain and achieve superior performance compared to previous CNN-based architectures. For example, TimeSformer [4] and ViViT [1] study several variants of space-time factorization for the plain ViT. Some works [5, 47, 50] further explore how to reduce computational cost of the space-time attention. VideoSwin [42] and MViT [18, 38] study the hierarchical architecture and introduce an inductive locality bias into video transformers. Uniformer [36] and Video Mobile-Former [64] propose to integrate 3D CNNs and spatial-temporal attention mechanism for efficiency consideration. For convincing performance on the video understanding tasks, most video transformers require model weights pretrained on the large-scale image datasets. In this paper, we explore the self-supervised pretraining of video transformers and show pretraining strategy will significantly influence the downstream performance, which is orthogonal to the transformer architecture design." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6313" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "Self-supervised video representation learning. The early works [3, 45, 66, 70] of self-supervised video representation learning focus on designing the pretext tasks based on the temporal structure of videos. More recently, contrastive learning [8, 32, 37] that forces different views of the same image sample to be closer in the feature space while pushing the views of different images farther becomes a new paradigm of representation learning, and some works [11, 23, 24, 29, 30, 46, 49] design the contrastive learning methods on the video domain by exploring effective ways of spatial-temporal augmentations. However, as the learning supervision based on contrastive learning is applied on global representation, it cannot well model the local relationship or learn fine-grained local representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 590 + ], + "type": "text", + "content": "Masked visual modeling. Masked language modeling [10, 40] has been one of the dominant pretraining methods of language transformers. With the success of vision transformers, masked visual modeling [2] has been introduced to self-supervised visual pretraining and demonstrates to be also helpful to multimodal visual-language learning [16, 74]. Following BERT [10] pretraining, BEiT [2] and PeCo [15] pretrain ViT by predicting the discrete visual tokens of masked patches, which are encoded by a pretrained VQ-VAE. MAE [31] proposes an asymmetric encoder-decoder framework for the reconstruction of pixels. SimMIM [68] and MaskFeat [67] propose to recover low-level features of masked patches like pixels or HOG features with hierarchical ViT. In contrast, some works [9, 14, 75] adopt an exponential moving average of the student model as the online teacher model, which makes the target features bootstrapped during training. In the video domain, some pioneering works [21, 55, 57, 63] extend masked image modeling to masked video modeling. BEVT [63] proposes a two-stream pretraining joint pretraining framework by predicting the discrete tokens with both image transformer and video transformer. VideoMAE [57] and ST-MAE [21] follow MAE and reconstruct the pixels of masked video patches with an extremely high masking ratio. Unlike most previous works of masked video modeling, our MVD focuses on masked feature modeling with high-level features as targets, and finds that student models using image and video teacher models will have different properties and complement each other." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Knowledge distillation. Knowledge distillation [26,33,48] aims to transfer the knowledge of the teacher model to the student model by adopting the output of the teacher model as the target for training the student model. Typical knowledge distillation works [33, 51, 52] mainly focus on supervised learning, e.g., image classification. The idea of knowledge distillation has also been leveraged in many self-supervised methods by learning from momentum encoders [6, 56] or different architectures [24]. Recently, self-supervised knowledge distillation [19, 71, 72] has also been" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "studied to learn representations from self-supervised pretrained models. In this paper, we present the first attempt that uses the masked image modeling pretrained image and video model as the masked feature prediction target in the video domain. It shows self-supervised MIM pretrained model can further bootstrap the mask video pretraining and bring significant performance gain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 167, + 362, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 362, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 362, + 180 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 188, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 547, + 368 + ], + "type": "text", + "content": "While masked video modeling has demonstrated promising performance for self-supervised learning, most existing approaches reconstruct relatively low-level information in the forms of raw pixels [57], low-level features like HOG [67] and VQVAE tokens [63]. In this paper, instead of reconstructing low-level information, we conduct masked video modeling at the feature-level. This is achieved by a two-stage framework, MVD, optimized to predict high-level features that derived from off-the-shelf MIM pretrained image models [31] and MVM pretrained video models [57] which are readily available. Below, we first given an overview of the masked feature modeling paradigm in Sec. 3.1 and then we introduce our proposed MVD in Sec. 3.2. Finally, we present the architectural design of MVD in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 376, + 534, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 534, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 534, + 389 + ], + "type": "text", + "content": "3.1. The Paradigm of Masked Feature Modeling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "text", + "content": "The core of masked feature modeling is to train models to predict features of masked input regions. In this paper, we follow the decoupled encoder-decoder transformer architecture in MAE [31] due to its effectiveness and simplicity. An input " + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "text", + "content": " (image " + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "inline_equation", + "content": "X_{img} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "text", + "content": " or video " + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "inline_equation", + "content": "X_{vid} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "text", + "content": ") is partitioned into several non-overlapping patches, and then each patch is mapped to a visual token with a linear projection layer. Before feeding tokens to the transformer encoder " + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "text", + "content": ", a subset of tokens is masked and dropped from the token sequence. To reconstruct the information of masked tokens, the token sequence consisting of the visible tokens from the encoder and learnable mask tokens are input to a shallow transformer decoder " + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 394, + 547, + 551 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 562, + 545, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 562, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 358, + 562, + 545, + 576 + ], + "type": "interline_equation", + "content": "Y = g \\left(\\operatorname {c o n c a t} \\left(f \\left(X _ {\\text {v i s}}\\right), T _ {m}\\right)\\right), \\tag {1}", + "image_path": "5099e0352b5fb85ddc7cd901f6d04bbc9f14da1c358809d6346ce6d4d746921f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "X_{vis}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " denotes the visible input tokens, and " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "T_{m}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " denotes the mask tokens. The subset of output tokens from decoder corresponded to input mask tokens contains reconstructed information of masked tokens. The reconstruction target for each masked patch " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "X(p)" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " is represented as a patch feature " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "h(X(p))" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " is represents a function for generating the target features, e.g. " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " produces low-level RGB values of pixels in the patch in [21, 57]. Then, to train the encoder and the decoder, a loss function that measures the distance " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " between the ground-truth features of masked patches and reconstructed ones is defined as:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6314" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 509, + 260 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 509, + 260 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 509, + 260 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 509, + 260 + ], + "type": "image", + "image_path": "42876f42afae95eccbfa8b9eeba9ec9aa8139a1c1c013a96cca151ad7b2822d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 547, + 300 + ], + "type": "text", + "content": "Figure 2. An overview of MVD framework. Firstly the image teacher is pretrained by masked image modeling and the video teacher is pretrained by masked video modeling. Then the student model is trained from scratch to predict target high-level features encoded by the image teacher and the video teacher. The teacher models are fixed in the distillation stage." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 317, + 287, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 317, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 80, + 317, + 287, + 347 + ], + "type": "interline_equation", + "content": "L _ {m f m} (h) = \\frac {1}{| M |} \\sum_ {p \\in M} D (Y (p), h (X (p))) \\tag {2}", + "image_path": "d2cd2883718a7f424680686953172c8b372dae530edbf53e26159f0b5942d095.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 350, + 288, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 288, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 350, + 288, + 387 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 350, + 288, + 387 + ], + "type": "text", + "content": " is the token index and " + }, + { + "bbox": [ + 47, + 350, + 288, + 387 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 350, + 288, + 387 + ], + "type": "text", + "content": " is the set of masked tokens. For pixel regression in MAE [31] and VideoMAE [57], the L2 distance is used as the distance metric." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 392, + 192, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 192, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 192, + 403 + ], + "type": "text", + "content": "3.2. Masked Video Distillation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "text", + "content": "In this paper, we propose Masked Video Distillation (MVD), which performs masked feature modeling on videos using high-level features as opposed to low-level pixels. In particular, we simply use outputs generated by off-the-shelf self-supervised pretrained image or video models, which are readily available, as reconstruction targets. These high-level features, serving as targets of the mask & prediction tasks, are encoded by teacher models pretrained by masked visual modeling like MAE or VideoMAE. For video representation learning, the reconstruction targets can take the form of spatial features encoded by image teacher models, or spatial-temporal features encoded by video teacher models. More specifically, the image teachers is pretrained by masked image modeling, while the video teacher is pretrained with masked video modeling, both of which aim at reconstructing raw pixels. Once trained, we use the image encoder " + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "inline_equation", + "content": "h_{img}" + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "text", + "content": " to generate the spatial targets, and the pretrained video transformer encoder " + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "inline_equation", + "content": "h_{vid}" + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "text", + "content": " to generate spatial-temporal targets. The loss function of MVD with the image teacher and video teacher can be denoted by " + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "inline_equation", + "content": "L_{mfm}(h_{img})" + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "inline_equation", + "content": "L_{mfm}(h_{vid})" + }, + { + "bbox": [ + 46, + 411, + 288, + 663 + ], + "type": "text", + "content": ", respectively" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Spatial-temporal Co-teaching. When performing MVD with a single teacher, we observe that students distilled from different teachers learn different video representations and perform well on different kinds of downstream video" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 309, + 547, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 547, + 466 + ], + "type": "text", + "content": "tasks. To improve the performance of MVD on different downstream video tasks, we propose spatial-temporal co-teaching that explores information from both image and video teachers such that the student model can handle videos of different types better. For instance, videos with fastly changing human actions require more temporal information while spatial clues might be sufficient for relatively static videos. To this end, MVD is trained to predict target high-level features produced by the image teacher and the video teacher at the same time. This is achieved by using two separated decoders to reconstruct different target features. The final loss of MVD with spatial-temporal co-teaching is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 337, + 485, + 545, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 485, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 337, + 485, + 545, + 498 + ], + "type": "interline_equation", + "content": "L _ {m v d} = \\lambda_ {1} L _ {m f m} \\left(h _ {i m g}\\right) + \\lambda_ {2} L _ {m f m} \\left(h _ {v i d}\\right) \\tag {3}", + "image_path": "515a3ff566239f676f010882c9d2cedceed592c283a6f6d3374b31f0a1af3943.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 505, + 547, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 505, + 547, + 541 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 304, + 505, + 547, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 505, + 547, + 541 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 304, + 505, + 547, + 541 + ], + "type": "text", + "content": " denote the hyper-parameters that balance the weights of the image teacher and the video teacher. The pseudo code of MVD is shown in Algorithm 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 548, + 426, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 426, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 426, + 560 + ], + "type": "text", + "content": "3.3. Architectural Design" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "content": "Encoder. For MVD, the vanilla transformer backbone is used as the encoder. For a video input " + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "inline_equation", + "content": "X_{vid} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "content": ", we adopt 3D patch embedding with a patch size of " + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "inline_equation", + "content": "2 \\times 16 \\times 16" + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "content": ". After patch partitioning and linear embedding, we obtain " + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "inline_equation", + "content": "T / 2 \\times H / 16 \\times W / 16" + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "content": " tokens. For the masked feature modeling, the tokens are masked with a high masking ratio and the remaining tokens are fed into the transformer layers. For finetuning on downstream tasks, we input all tokens to the subsequent layers. In each layer, joint spatial-temporal self-attention is applied on the whole input token sequence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Mask strategy. For MVD, we follow [57] and adopt tube masking for masked feature modeling. First a 2D random" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6315" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 92, + 286, + 298 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 258, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 258, + 84 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of MVD in PyTorch style." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 92, + 286, + 298 + ], + "lines": [ + { + "bbox": [ + 47, + 92, + 286, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 286, + 298 + ], + "type": "text", + "content": "f: student encoder \n# g.img: decoder for reconstructing spatial features \n# g_vid: decoder for reconstructing spatial-temporal features \n# t_m: learnable mask tokens \n# h.img: image teacher model \n# h_vid: video teacher model \nfor x, m in loader: # x: video data, m: mask \nx_pe = patch_emb(x) # patch embedding of input \nx.vis = mask_select(x_pe, 1 - m) # masking tokens \nq.vis = f(x.vis) # visible local patch features \n# reconstruction of target features \np IMG = g.img(concat(q.vis, t_m)) \np_vid = g_vid(concat(q.vis, t_m)) \n# compute target features with teacher models \nk IMG = h.img(x) # target spatial features \nk_vid = h_vid(x) # target spatial-temporal features \n# compute reconstruction loss \nloss IMG = smooth_L1_loss(p.img ⊙ m, k.img ⊙ m) \nloss_vid = smooth_L1_loss(p:vid ⊙ m, k:vid ⊙ m) \nloss = λ1 * loss IMG + λ2 * loss_vid \nloss.backup() \noptimizer_STEP() # optimizer update" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 46, + 323, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 396 + ], + "type": "text", + "content": "mask is generated and then extended along the temporal dimension. Therefore, the spatial mask on each time slice is the same, which prevents information leakage between frames. Tube masking with a high masking ratio (e.g., " + }, + { + "bbox": [ + 46, + 323, + 288, + 396 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 323, + 288, + 396 + ], + "type": "text", + "content": ") encourages the video transformer to model the high-level semantics during pretraining." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 405, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 561 + ], + "type": "text", + "content": "Decoder. For MVD, shallow decoders consist of vanilla transformer layers and a linear projection layer. The transformer layers in decoders are the same as those in the encoder. Since spatial-temporal co-teaching introduces two different reconstruction targets for masked feature modeling, two separated decoders that share the same architecture but contain different weights are placed on the top of the encoder. Learnable masked tokens corresponded to masked patches are concatenated with visible tokens from the encoder before fed into the decoder. After jointly modeling the spatial-temporal relationships, the output tokens of transformer layers are mapped to final predictions by the linear projection layer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Reconstruction targets. For generating spatial-temporal target features, the video teacher, which shares the same architecture as the student model, is pretrained by a VideoMAE [57] manner on the video dataset. For obtaining spatial targets, we adopt the vanilla image ViT pretrained by masked image modeling [31] on the image dataset (e.g., ImageNet-1K). It is worth noting that one 3D patch (with size of " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 16 \\times 16" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ") for the video transformer corresponds to two 2D patches (with size of " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ") for the image transformer. Following [21], we predict the spatial features of a single time slice (that is the front 2D patch), which reduce the prediction layer's size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 547, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 140 + ], + "type": "text", + "content": "In this section, we first introduce the experimental setup in Sec. 4.1, and then present the main results in Sec. 4.2, followed by an extensive analysis to verify the effectiveness of different components in Sec. 4.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 148, + 422, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 148, + 422, + 162 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 422, + 162 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "content": "Dataset. We pretrain the vanilla ViT with MVD on Kinetics-400 by default, and evaluate the learned model on four video recognition downstream tasks: (a) Kinetics-400 (K400) [7], which consists of " + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\sim 240K" + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "content": " training videos and " + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\sim 20K" + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "content": " validation videos with an average duration of 10 seconds. All video clips are labeled into 400 classes. (b) Something-Something V2 (SSv2) [27], which contains " + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\sim 160K" + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "content": " videos for training and " + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\sim 20K" + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "content": " videos for validation. The videos in SSv2 with an average duration of 4 seconds are labeled into 174 motion-centric categories. (c) UCF-101 [54] is a relatively small dataset, consisting of " + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\sim 9.5K" + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "content": " training videos and " + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\sim 3.5K" + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "content": " validation videos. (d) HMDB51 [35] is also a small video dataset that contains around " + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "inline_equation", + "content": "3.5K/1.5K" + }, + { + "bbox": [ + 304, + 167, + 547, + 371 + ], + "type": "text", + "content": " train/val videos. We also transfer pretrained models to a challenging spatial-temporal action detection dataset AVA [28]. The results of AVA are included in supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 375, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 547, + 542 + ], + "type": "text", + "content": "Implementation details. Our MVD is performed on vanilla ViTs with different capacities (i.e., ViT-S, ViT-B, ViT-L). By default, image teacher models are pretrained on ImageNet-1K for 1600 epochs and video teachers are pretrained on K400 for 1600 epochs. We follow the training strategy in MAE [31] and VideoMAE [21] for image teachers and video teachers respectively. In the distillation stage, student models are first pretrained from scratch on K400 for 400 epochs unless mentioned otherwise. The resulting models are then finetuned on downstream video tasks. The video clip length is 16 for both pretraining and finetuning. We adopt AdamW optimizer [44] and Smooth L1 loss for the optimization of student models. More details are included in supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 551, + 391, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 391, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 391, + 563 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Students distilled from different teachers. Unlike masked image modeling, masked feature modeling on video data has more choices on reconstruction targets. Besides spatial features, to include temporal dynamics in the reconstruction target, we can also adopt spatial-temporal features encoded by pretrained video models. In Table 1, we compare students distilled by the image teacher and the video teacher on K400, a downstream task that mainly relies on spatial clues, and SSv2, a temporally-heavy downstream task. Our observations are as follow: (a) Masked feature modeling with high-level features as targets achieves convincing performance on downstream video tasks, and outperforms Video" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6316" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 281, + 168 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 281, + 168 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 281, + 168 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 281, + 168 + ], + "type": "table", + "html": "
studentteachersK400 top-1SSv2 top-1
imagevideo
ViT-S✓✗80.469.4
✗✓80.170.0
✓✓80.670.7
ViT-B✓✗82.371.4
✗✓82.171.8
✓✓82.772.5
", + "image_path": "50df52d8e10c4c7ca7e120607bf30fde978ecd0cdb35b38a19e0a49cd89a075a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 218, + 288, + 302 + ], + "blocks": [ + { + "bbox": [ + 47, + 170, + 287, + 214 + ], + "lines": [ + { + "bbox": [ + 47, + 170, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 287, + 214 + ], + "type": "text", + "content": "Table 1. Co-teaching with the image teacher and the video teacher outperforms distillation with a single teacher in MVD. We adopt ViT-B as the architecture of teachers and distill students for 400 epochs here." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 218, + 288, + 302 + ], + "lines": [ + { + "bbox": [ + 48, + 218, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 288, + 302 + ], + "type": "table", + "html": "
studentteacherK400 top-1SSv2 top-1
ViMAEMVDViMAEMVD
ViT-SViT-B79.080.6 ↑1.666.470.7 ↑4.3
ViT-SViT-L79.081.0 ↑2.066.470.9 ↑4.5
ViT-BViT-B81.582.7 ↑1.269.772.5 ↑2.8
ViT-BViT-L81.583.4 ↑1.969.773.7 ↑4.0
ViT-LViT-L85.286.0 ↑0.874.076.1 ↑2.1
", + "image_path": "1e1cc95c0054fd05efa40de8931eb3bf994b1eaf04465612ebb95e0d6cca9107.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 548 + ], + "type": "text", + "content": "MAE baseline significantly (compared with the baseline results in Table 2). In particular, with both image and video teachers using a ViT-S as the backbone, MVD achieves consistent gains over VideoMAE on both K400 (80.6% vs. 79.0%) and SSV2 (70.7% vs. 66.4%). (b) Students distilled from the image teacher achieve higher top-1 accuracy on K400, while students distilled from the video teacher perform better on SSv2. For example, ViT-S achieves an accuracy of 80.4% and 69.4% using an image teacher on K400 and SSv2 respectively. With a video teacher, on the other hand, ViT-S offers a top-1 accuracy of 80.1% and 70.0% respectively. As it has been shown that videos in K400 less sensitive to temporal modeling compared to SSv2, the results demonstrate students learn stronger spatial representation from the image teacher while the video teacher transfer more knowledge about temporal dynamics to students." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 553, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 286, + 661 + ], + "type": "text", + "content": "Co-teaching outperforms distilling with a single teacher. To improve the performance on different kinds of downstream video tasks, we introduce spatial-temporal co-teaching in MVD, which trains the model to predict spatial features and spatial-temporal features of masked patches in a decoupled way. The results in Table 1 indicate that students distilled with spatial-temporal co-teaching outperforms students distilled from either single teacher on both spatially-heavy task and temporally-heavy task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "content": "MVD outperforms VideoMAE baseline significantly. In Table 2, MVD with spatial-temporal co-teaching is compared with VideoMAE pretrained on K400. When the size of teacher models is the same as that of students, MVD out" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 304, + 70, + 547, + 471 + ], + "blocks": [ + { + "bbox": [ + 47, + 304, + 287, + 348 + ], + "lines": [ + { + "bbox": [ + 47, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 287, + 348 + ], + "type": "text", + "content": "Table 2. MVD achieves significant improvement compared with VideoMAE across different model scales. MVD is pretrained on K400 for 400 epochs, while VideoMAE is pretrained on K400 for 1600 epochs here." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 70, + 547, + 471 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 471 + ], + "type": "table", + "html": "
methodextra datatop-1top-5GFLOPsParam
supervised
NL I3D R101 [65]-77.393.3359×3062
ip-CSN-152 [59]-77.892.8109×3033
SlowFast NL [22]-79.893.9234×3060
X3D-XL [20]-79.193.948×3011
MViTv1-B [18]-80.294.4170×537
VideoSwin-B [42]IN-1K80.694.6282×1288
Uniformer-B [36]IN-1K83.095.4259×1250
TimeSformer [4]IN-21K80.794.72380×3121
Mformer-B [47]IN-21K79.794.2370×30109
Mformer-L [47]IN-21K80.294.81185×30382
ViViT-L FE [1]IN-21K81.793.83980×3N/A
VideoSwin-L [42]IN-21K83.195.9604×12197
self-supervised
VIMPAC ViT-L [55]HowTo100M77.4N/AN/A×30307
BEVT Swin-B [63]IN-1K81.1N/A282×1288
MaskFeat MViT-S [67]-82.295.171×1036
VideoMAE ViT-S [57]-79.093.857×1522
VideoMAE ViT-B [57]-81.595.1180×1587
VideoMAE ViT-L [57]-85.296.8597×15305
VideoMAE ViT-H [57]-86.697.11192×15633
ST-MAE ViT-B [21]-81.394.9180×2187
ST-MAE ViT-L [21]-84.896.2598×21304
ST-MAE ViT-H [21]-85.196.61193×21632
OmniMAE ViT-B [25]IN-1K80.8N/A180×1587
OmniMAE ViT-L [25]IN1K+SSv284.0N/A597×15305
OmniMAE ViT-H [25]IN1K+SSv284.8N/A1192×15633
MVD-S (Teacher-B)IN-1K80.694.757×1522
MVD-S (Teacher-L)IN-1K81.094.857×1522
MVD-B (Teacher-B)IN-1K82.795.4180×1587
MVD-B (Teacher-L)IN-1K83.495.8180×1587
MVD-L (Teacher-L)IN-1K86.096.9597×15305
MVD-L (Teacher-L) †IN-1K86.497.0597×15305
MVD-H (Teacher-H) †IN-1K87.397.41192×15633
", + "image_path": "184939b9f85e0c9c294775bea22a135e2fcaa24dfe1c979ed9b8dace15f8d1ea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 545, + 538 + ], + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 538 + ], + "type": "text", + "content": "Table 3. Comparison with previous works on Kinetics-400. \"MVD-S (Teacher-B)\" denotes that the student is ViT-S and the teacher is ViT-B. " + }, + { + "bbox": [ + 304, + 472, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 472, + 545, + 538 + ], + "type": "text", + "content": " denotes that the student is distilled for 800 epochs instead of 400 epochs. The inference cost is reported with the cost of a single view " + }, + { + "bbox": [ + 304, + 472, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 472, + 545, + 538 + ], + "type": "text", + "content": " the number of views. Note that IN-1K is only used for training image teachers in MVD." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "content": "performs VideoMAE by a clear margin on both K400 and SSv2. Larger models as teachers can further boost the performance of MVD. It is worth mentioning that not only is our MVD particularly effective for relatively small models, but also it improves the performance of large vision models like ViT-L. For example, with a ViT-L model as the student, MVD achieves " + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "inline_equation", + "content": "86.0\\%" + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "inline_equation", + "content": "76.1\\%" + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "content": " on K400 and SSv2, surpassing VideoMAE by " + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Comparison with state-of-the-art. We compare MVD with prior studies on four video recognition tasks. Results on K400 are shown in Table 3. Our MVD outperforms previous self-supervised methods with similar or less computational cost. Even compared with video transform" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6317" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 45, + 70, + 289, + 542 + ], + "blocks": [ + { + "bbox": [ + 45, + 70, + 289, + 542 + ], + "lines": [ + { + "bbox": [ + 45, + 70, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 45, + 70, + 289, + 542 + ], + "type": "table", + "html": "
methodextra datatop-1GFLOPsParam
supervised
SlowFast R101 [22]K40063.1106×353
TSM-RGB R50 [39]IN-1K63.362×624
TAM R50 [43]IN-1K66.099×651
TDN R101 [61]IN-1K69.6198×388
MViTv1-B [18]-67.7455×337
MViTv2-B [38]K40070.5225×351
Uniformer-B [36]K40071.2259×350
TimeSformer-HR [4]IN-21K62.51703×3121
ViViT-L FE [1]IN21K+K40065.9995×12N/A
Mformer-B [47]IN21K+K40066.5370×3109
Mformer-L [47]IN21K+K40068.11185×3382
VideoSwin-B [42]IN21K+K40069.6321×388
MViTv2-L [38]IN21K+K40073.32828×3213
self-supervised
VIMPAC ViT-L [55]HowTo100M68.1N/A×30307
BEVT Swin-B [63]IN1K+K40071.4321×388
MaskFeat MViT-L [67]K40074.42828×3218
VideoMAE ViT-S [57]K40066.457×622
VideoMAE ViT-S [57]-66.857×622
VideoMAE ViT-B [57]K40069.7180×687
VideoMAE ViT-B [57]-70.8180×687
VideoMAE ViT-L [57]K40074.0597×6305
VideoMAE ViT-L [57]-74.3597×6305
ST-MAE ViT-L [21]K40072.1598×3304
ST-MAE ViT-H [21]K40074.11193×3632
OmniMAE ViT-B [25]IN1K+K40069.0180×687
OmniMAE ViT-L [25]IN-1K74.2597×6305
OmniMAE ViT-H [25]IN-1K75.31192×6632
MVD-S (Teacher-B)IN1K+K40070.757×622
MVD-S (Teacher-L)IN1K+K40070.957×622
MVD-B (Teacher-B)IN1K+K40072.5180×687
MVD-B (Teacher-L)IN1K+K40073.7180×687
MVD-L (Teacher-L)IN1K+K40076.1597×6305
MVD-L (Teacher-L) †IN1K+K40076.7597×6305
MVD-H (Teacher-H) †IN1K+K40077.31192×6633
", + "image_path": "cc113547bae6fbfcfc7a5afa7af8e5721c5948986eafec086e038c8cff931edf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "ers pretrained on ImageNet-21K, MVD achieves superior performance. Particularly, MVD-H achieves an accuracy of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "87.3\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " on K400, outperforming previous top-performing methods by clear margins. Table 4 presents comparisons to the state-of-the-art methods on SSv2. We observe that for downstream tasks depending on temporal relationship modeling, self-supervised methods based on MVM achieve better performance in comparison with supervised methods (cf. results in the middle group of Table 4 vs. results in the top group). Once again, our MVD, producing an accuracy of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "76.1\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " with a large model, beats both supervised meth" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 305, + 70, + 548, + 206 + ], + "blocks": [ + { + "bbox": [ + 46, + 544, + 287, + 577 + ], + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 577 + ], + "type": "text", + "content": "Table 4. Comparison with previous works on Something-Something V2.. † denotes that the student is distilled for 800 epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 70, + 548, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 548, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 548, + 206 + ], + "type": "table", + "html": "
methodextra dataParamUCF101HMDB51
VideoMoCo R2+1D [46]K4001578.749.2
MemDPC R2D3D [30]K4003286.154.5
Vi2CLR S3D [11]K400989.155.7
CORP Slow-R50 [34]K4003293.568.0
CVRL Slow-R50 [49]K4003292.967.9
CVRL Slow-R152 [49]K60032894.470.6
ρBYOL Slow-R50 [23]K4003294.272.1
VIMPAC ViT-L [55]HowTo100M30792.765.9
VideoMAE ViT-B [57]K4008796.173.3
MVD-B (Teacher-B)IN-1K+K4008797.076.4
MVD-B (Teacher-L)IN-1K+K4008797.579.7
", + "image_path": "8707af0e1bd3a85c90c59211d9235345b4386207db2a6e41e023ea38a79e1747.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 208, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 229 + ], + "type": "text", + "content": "Table 5. Comparison with previous state-of-the-art methods on UCF101 and HMDB51." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 308, + 238, + 417, + 328 + ], + "blocks": [ + { + "bbox": [ + 308, + 238, + 417, + 328 + ], + "lines": [ + { + "bbox": [ + 308, + 238, + 417, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 238, + 417, + 328 + ], + "type": "image", + "image_path": "b85514161767a24e5ba9d65e7a6ebec3da134586a2592035b7fe27a40a8bdffc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 330, + 392, + 339 + ], + "lines": [ + { + "bbox": [ + 333, + 330, + 392, + 339 + ], + "spans": [ + { + "bbox": [ + 333, + 330, + 392, + 339 + ], + "type": "text", + "content": "(a) image teacher." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 434, + 237, + 543, + 329 + ], + "blocks": [ + { + "bbox": [ + 434, + 237, + 543, + 329 + ], + "lines": [ + { + "bbox": [ + 434, + 237, + 543, + 329 + ], + "spans": [ + { + "bbox": [ + 434, + 237, + 543, + 329 + ], + "type": "image", + "image_path": "7a74256dfcf3e265a50a769507214dc95a131cbb2672bc6fde43b3c3c6ba27ea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 330, + 518, + 338 + ], + "lines": [ + { + "bbox": [ + 460, + 330, + 518, + 338 + ], + "spans": [ + { + "bbox": [ + 460, + 330, + 518, + 338 + ], + "type": "text", + "content": "(b) video teacher." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 342, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 373 + ], + "type": "text", + "content": "Figure 3. Feature similarity across different frames for different teacher models. Similarity matrices are computed on the Kinetics-400 validation set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "content": "ods and self-supervised methods by clear margins. With more training epochs (i.e., 800 epochs), MVD with ViT-L achieves more significant performance gains (i.e., " + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "content": " on K400 and SSv2) compared with VideoMAE. When a huge model is used, the performance can still be boosted and MVD achieves " + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "inline_equation", + "content": "77.3\\%" + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "content": " top-1 accuracy on SSv2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 455, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 575 + ], + "type": "text", + "content": "We also evaluate the transfer learning ability of MVD on two relatively small datasets, UCF101 and HMDB51. As shown in Table 5, MVD with ViT-B obtains higher accuracy compared with prior works based on well-designed pretext tasks, contrastive learning and masked video modeling methods. Especially compared to the original VideoMAE [57] ViT-B teacher model, we achieve " + }, + { + "bbox": [ + 304, + 455, + 545, + 575 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 455, + 545, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 455, + 545, + 575 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 304, + 455, + 545, + 575 + ], + "type": "text", + "content": " higher points on these two datasets respectively. Additionally, when teachers with larger size are adopted, MVD achieves stronger transfer learning performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 585, + 440, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 440, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 440, + 597 + ], + "type": "text", + "content": "4.3. Analysis and Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "text", + "content": "In this section, we provide an in-depth analysis of the effectiveness of different components in MVD." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Analysis of features encoded by different teachers. The properties of target features generated by different teachers may influence the performance of students on different downstream tasks. To quantify the temporal dynamics that teacher models capture from the input video, we study the similarity between feature maps across different frames of each input video clip via the cosine similarity. As similarity" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6318" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 279, + 120 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 279, + 120 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 279, + 120 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 279, + 120 + ], + "type": "table", + "html": "
methodepochtime(h)K400 top-1
VideoMAE80010781.0
VideoMAE160021481.5
MVD(800+) 400(107+) 5781.9
", + "image_path": "3f21e6d752cc792304fcbe915ce427f16ea0e5c490d436db6be0fcf4db6bf9ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 169, + 282, + 232 + ], + "blocks": [ + { + "bbox": [ + 47, + 121, + 287, + 165 + ], + "lines": [ + { + "bbox": [ + 47, + 121, + 287, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 287, + 165 + ], + "type": "text", + "content": "Table 6. Training time comparison between MVD and VideoMAE on 32 NVIDIA V100 GPUs. The training time of teacher models is considered. we only adopt a video teacher pretrained for 800 epochs in MVD here. ViT-B is used here." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 169, + 282, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 169, + 282, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 169, + 282, + 232 + ], + "type": "table", + "html": "
teachersreconstruct pixelSSv2 top-1
image modelX68.7
image model✓67.9 ↓0.8
image+video modelsX70.1
image+video models✓69.0 ↓1.1
", + "image_path": "b8c785510b01f5b756208b0c5574af8377521103df4114f165e2f0490dc7d9c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 54, + 267, + 278, + 354 + ], + "blocks": [ + { + "bbox": [ + 47, + 233, + 286, + 266 + ], + "lines": [ + { + "bbox": [ + 47, + 233, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 286, + 266 + ], + "type": "text", + "content": "Table 7. Ablations on reconstructing low-level features in masked video distillation. We distill ViT-S students with ViT-B teachers for 300 epochs here." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 267, + 278, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 267, + 278, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 278, + 354 + ], + "type": "table", + "html": "
teacherinitepochtop-1
K400SSv2
momentum encoderX80080.570.4
momentum encoder✓80081.870.8
fixed image modelX40082.371.4
fixed video modelX40082.171.8
fixed co-teachingX40082.772.5
", + "image_path": "0ea09233f5232f07a5f99176590c2d0252130f04302cdc66133a84493527826d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 383, + 286, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 286, + 467 + ], + "type": "text", + "content": "matrices shown in Figure 3, for image teachers, the feature maps of different frames are almost the same. However, for video teachers, the features of different frames have larger difference. This indicates that video teachers capture more temporal difference. Therefore, students distilled from video teachers can learn stronger temporal dynamics and perform better on temporally-heavy downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "text", + "content": "Training time comparison. We study whether MVD is able to achieve better balance of accuracy and efficiency than VideoMAE. For fair comparisons, the training time of teacher models is also counted in the total training time of MVD. Results are shown in Table 6. We see that MVD can achieve better accuracy (i.e., " + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "inline_equation", + "content": "81.9\\%" + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "text", + "content": ") by a total of 164 hours of training which is 50 hours less than VideoMAE (producing an accuracy of " + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "inline_equation", + "content": "81.5\\%" + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "text", + "content": ") trained for 1600 epochs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Reconstruction signals in MVD. In MVD, we first pretrain teacher models by recovering the pixels of masked patches in a MAE manner, then adopt features produced by teachers as the targets of masked feature modeling. In Table 7, we study whether to include an additional decoder branch for reconstructing pixels of masked patches in the distillation stage. The experimental results show that for both MVD with a single teacher and MVD with spatial-temporal co-teaching, the reconstruction of low-level feature targets degrades the performance for downstream tasks. Therefore, we only reconstruct high-level features of masked patches in the distillation stage of MVD." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 534, + 108 + ], + "blocks": [ + { + "bbox": [ + 47, + 354, + 286, + 376 + ], + "lines": [ + { + "bbox": [ + 47, + 354, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 286, + 376 + ], + "type": "text", + "content": "Table 8. Comparison with bootstrapped teachers and students initialized with IN1K-pretrained models. ViT-B is used here." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 534, + 108 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 534, + 108 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 534, + 108 + ], + "type": "table", + "html": "
distillation methodK400 top-1SSv2 top-1
per-token distillation80.970.5
masked reconstruction82.171.8
", + "image_path": "5aa368d100821e3939d980725e7f95ac7d35ed6a3972c8fd1c731cc71ac6b9e8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 110, + 545, + 131 + ], + "lines": [ + { + "bbox": [ + 306, + 110, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 545, + 131 + ], + "type": "text", + "content": "Table 9. Comparison with feature distillation without masked reconstruction. We distill ViT-B from the video teacher here." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 142, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 142, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 142, + 545, + 345 + ], + "type": "text", + "content": "Comparison with bootstrapped teachers. Some recent works of image representation learning [9, 14, 16, 75] adopt features of a momentum encoder as the targets of MIM, while we use frozen teacher models in MVD. In Table 8, we compare fixed teachers with bootstrapped teachers that are updated by an exponential moving average of the online encoder during pretraining. According to a MIM method [14], two strong baselines of bootstrapped teachers are built for video representation learning: (a) The student model is trained from scratch with masked feature modeling, and the target features are generated by a momentum encoder. (b) The framework of bootstrapped teachers is first pretrained on IN-1K for 800 epochs, then the pretrained weights are adopted to initialize the video pretraining. As shown in Table 8, MVD with the frozen teacher beats the method with a bootstrapped teacher on the downstream video tasks, even if only a single teacher is utilized." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 350, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 469 + ], + "type": "text", + "content": "Comparison with feature distillation. In the previous work of self-supervised feature distillation [19, 26, 72], the distillation loss is directly computed upon the full feature maps between teachers and students. Accordingly, we use a baseline method named per-token distillation. Specifically, the output features of students are projected by a MLP, and then forced to mimicking the teacher's feature at each token with a Smooth L1 loss. As shown in Table 9, masked feature reconstruction in our MVD outperforms per-token distillation on both K400 and SSv2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 487, + 378, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 378, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 378, + 498 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 508, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 687 + ], + "type": "text", + "content": "In this paper, we study masked video distillation upon MIM pretrained image or video transformers. We have three interesting findings: 1) Using MIM pretrained image transformers or MVM pretrained video transformers as teachers supervise masked feature prediction can significantly boost the finetuning performance on video downstream tasks; 2) The representation distilled with image and video teachers will have different properties, i.e., image teachers will benefit spatial-heavy video tasks more while video teachers benefit temporal-heavy video tasks more; 3) Combining image and video teachers will enjoy the synergy and thus produce higher performance. Even though the proposed masked video distillation seems very straightforward, we hope such interesting findings can motivate more thinking about masked video pretraining." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement This project was supported by NSFC under Grant No. 62032006 and No. 62102092." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6319" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 156 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. In ICLR, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 286, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 286, + 203 + ], + "type": "text", + "content": "[3] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In CVPR, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "type": "text", + "content": "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "type": "text", + "content": "[5] Adrian Bulat, Juan Manuel Perez Rua, Swathikiran Sudhakaran, Brais Martinez, and Georgios Tzimiropoulos. Space-time mixing attention for video transformer. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 286, + 327 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 330, + 286, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 286, + 361 + ], + "type": "text", + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 364, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 286, + 396 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 399, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 286, + 430 + ], + "type": "text", + "content": "[9] Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Sdae: Self-distillated masked autoencoder. In ECCV, 2022. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 433, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 286, + 464 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 467, + 286, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 286, + 509 + ], + "type": "text", + "content": "[11] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saqib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2clr: Video and image for visual contrastive learning of representation. In ICCV, 2021. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 286, + 565 + ], + "type": "text", + "content": "[12] Jeffrey Donahue, Lisa Anne Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 632 + ], + "type": "text", + "content": "[13] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "text", + "content": "[14] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In ECCV, 2022. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "text", + "content": "[15] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "of vision transformers. In Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI), 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[16] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 152, + 545, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 214 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "type": "text", + "content": "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[19] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation. arXiv preprint arXiv:2101.04731, 2021. 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 294, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 315 + ], + "type": "text", + "content": "[20] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In CVPR, 2020. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "text", + "content": "[21] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "type": "text", + "content": "[22] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[23] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In CVPR, 2021. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 417, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 417, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 417, + 545, + 458 + ], + "type": "text", + "content": "[24] Chongjian Ge, Youwei Liang, Yibing Song, Jianbo Jiao, Jue Wang, and Ping Luo. Revitalizing cnn attention via transformers in self-supervised visual representation learning. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[25] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "text", + "content": "[26] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. IJCV, 2021. 3, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 538, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 538, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 538, + 545, + 612 + ], + "type": "text", + "content": "[27] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fründ, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "type": "text", + "content": "[28] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[29] Sheng Guo, Zihua Xiong, Yujie Zhong, Limin Wang, Xiaobo Guo, Bing Han, and Weilin Huang. Cross-architecture self-supervised video representation learning. In CVPR, 2022. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6320" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[30] Tengda Han, Weidi Xie, and Andrew Zisserman. Memory-augmented dense predictive coding for video representation learning. In ECCV, 2020. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 3, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[32] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "type": "text", + "content": "[33] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 216, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 287, + 250 + ], + "type": "text", + "content": "[34] Kai Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. In ICCV, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "text", + "content": "[35] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 284, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 326 + ], + "type": "text", + "content": "[36] Kunchang Li, Yali Wang, Gao Peng, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatial-temporal representation learning. In ICLR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "text", + "content": "[37] Suichan Li, Dongdong Chen, Yinpeng Chen, Lu Yuan, Lei Zhang, Qi Chu, Bin Liu, and Nenghai Yu. Improve unsupervised pretraining for few-label transfer. In ICCV, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "text", + "content": "[38] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "type": "text", + "content": "[39] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 448, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 287, + 502 + ], + "type": "text", + "content": "[40] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "type": "text", + "content": "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 581 + ], + "type": "text", + "content": "[42] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "text", + "content": "[43] Zhaoyang Liu, Limin Wang, Wayne Wu, Chen Qian, and Tong Lu. Tam: Temporal adaptive module for video recognition. In CVPR, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 614, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 635 + ], + "type": "text", + "content": "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "text", + "content": "[45] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[46] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In CVPR, 2021. 3, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[47] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In NeurIPS, 2021. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 128, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 545, + 149 + ], + "type": "text", + "content": "[48] Mary Phuong and Christoph Lampert. Towards understanding knowledge distillation. In ICML, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[49] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In CVPR, 2021. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "text", + "content": "[50] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "text", + "content": "[51] Zhiqiang Shen and Marios Savvides. Meal v2: Boosting vanilla resnet-50 to " + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "inline_equation", + "content": "80\\%+" + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "text", + "content": " top-1 accuracy on imagenet without tricks. arXiv preprint arXiv:2009.08453, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[52] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. arXiv preprint arXiv:2112.01528, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "type": "text", + "content": "[53] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "text", + "content": "[54] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "type": "text", + "content": "[55] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "type": "text", + "content": "[56] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Piriavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "text", + "content": "[57] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "type": "text", + "content": "[58] Du Tran, Lubomir D Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. C3d: Generic features for video analysis. In ICCV, 2015. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 525, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 558 + ], + "type": "text", + "content": "[59] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "text", + "content": "[60] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 592, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 624 + ], + "type": "text", + "content": "[61] Limin Wang, Zhan Tong, Bin Ji, and Gangshan Wu. Tdn: Temporal difference networks for efficient action recognition. In CVPR, 2021. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "text", + "content": "[62] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE TPAMI, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[63] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In CVPR, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6321" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 587 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[64] Rui Wang, Zuxuan Wu, Dongdong Chen, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Luowei Zhou, Lu Yuan, and YuGang Jiang. Video mobile-former: Video recognition with efficient global spatial-temporal modeling. arXiv preprint arXiv:2208.12257, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "type": "text", + "content": "[65] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "type": "text", + "content": "[66] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "text", + "content": "[67] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "text", + "content": "[68] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[69] Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. arXiv preprint arXiv:2211.13222, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "text", + "content": "[70] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 385 + ], + "type": "text", + "content": "[71] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "type": "text", + "content": "[72] Haohang Xu, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, Xinggang Wang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Bag of instances aggregation boosts self-supervised distillation. In ICLR, 2021. 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "type": "text", + "content": "[73] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 478, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 542 + ], + "type": "text", + "content": "[74] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2022), 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "type": "text", + "content": "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 3, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6322" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_content_list.json b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ff5e4ba4f6f7cbf54a592ad4149bf1f3e58bc5d8 --- /dev/null +++ b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_content_list.json @@ -0,0 +1,1778 @@ +[ + { + "type": "text", + "text": "Masked Wavelet Representation for Compact Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 135, + 130, + 834, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Rho1*", + "bbox": [ + 91, + 178, + 202, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Byeonghyeon Lee $^{2*}$", + "bbox": [ + 228, + 180, + 392, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seungtae Nam² \nEunbyung Park²,³† \n¹AI2XL, KT", + "bbox": [ + 416, + 180, + 563, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joo Chan Lee²", + "bbox": [ + 581, + 180, + 702, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jong Hwan $\\mathrm{Ko}^{2,3\\dagger}$", + "bbox": [ + 736, + 180, + 883, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Artificial Intelligence, Sungkyunkwan University", + "bbox": [ + 225, + 233, + 743, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Electrical and Computer Engineering, Sungkyunkwan University", + "bbox": [ + 163, + 251, + 805, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural radiance fields (NeRF) have demonstrated the potential of coordinate-based neural representation (neural fields or implicit neural representation) in neural rendering. However, using a multi-layer perceptron (MLP) to represent a 3D scene or object requires enormous computational resources and time. There have been recent studies on how to reduce these computational inefficiencies by using additional data structures, such as grids or trees. Despite the promising performance, the explicit data structure necessitates a substantial amount of memory. In this work, we present a method to reduce the size without compromising the advantages of having additional data structures. In detail, we propose using the wavelet transform on grid-based neural fields. Grid-based neural fields are for fast convergence, and the wavelet transform, whose efficiency has been demonstrated in high-performance standard codes, is to improve the parameter efficiency of grids. Furthermore, in order to achieve a higher sparsity of grid coefficients while maintaining reconstruction quality, we present a novel trainable masking approach. Experimental results demonstrate that non-spatial grid coefficients, such as wavelet coefficients, are capable of attaining a higher level of sparsity than spatial grid coefficients, resulting in a more compact representation. With our proposed mask and compression pipeline, we achieved state-of-the-art performance within a memory budget of 2 MB. Our code is available at https://github.com/daniel03c1/maskedwavelet_nerf.", + "bbox": [ + 76, + 335, + 473, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 775, + 209, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in coordinate-based neural representation (neural fields or implicit neural representation) have demonstrated remarkable performance in many applications. In particular, neural radiance fields (NeRF) have", + "bbox": [ + 76, + 801, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/959d47197dd2a619d446a16183c11c28e9004d14b86c2dd6dcac18aea2a9d6f2.jpg", + "image_caption": [ + "Figure 1. Rate-distortion curves on the NeRF synthetic dataset. The numbers inside parenthesis denote the axis resolution of grids." + ], + "image_footnote": [], + "bbox": [ + 504, + 304, + 890, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sparked interest by synthesizing high-quality images from novel viewpoints. It uses a multi-layer perceptron (MLP) with positional encoding to map coordinates to corresponding colors and opacities. Combined with the differentiable volumetric rendering and the neural network's architectural priors, it has shown great potential to be a new representation paradigm. However, the high computational costs (in both training and inference) have been a significant bottleneck, often taking several days to converge.", + "bbox": [ + 496, + 551, + 892, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several follow-up studies have been proposed to accelerate training and inference times [5, 9, 11, 14, 16, 31, 32, 40, 44]. To speed up inference, KiloNeRF [32] proposed splitting a 3D scene into thousands of partial scenes, each of which is assigned a tiny, distinct neural network. While achieving impressive speed-up, it requires a massive amount of memory storage. On the other hand, FastNeRF [14] suggested caching and factorizing the NeRF network to reduce computational costs. Meanwhile, meta-learning algorithms [12, 26, 29] have also been applied to accelerate the training time, and they have shown faster convergence with the learned initialization [34, 40]. However, it requires well-organized large-scale datasets and a pre-training process. Furthermore, the rendering costs re", + "bbox": [ + 496, + 688, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 96, + 887, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20680", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "main unchanged since they use the same network architecture [40], and meta-learning algorithms often suffer from poor out-of-distribution generalization performance.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternatively, there has been a surge of recent interest in incorporating classical data structures, such as grids or trees, into the NeRF framework [13, 22, 24, 33, 38, 39, 49]. Incorporating additional data structures has significantly reduced training and inference time (from days to a few minutes) without compromising reconstruction quality. However, the overall size dramatically increases due to these dense and volumetric structures. In order to reduce the spatial complexity, several methods have been proposed, including pruning areas [33, 37, 49], encodings [24, 38], and tensor decomposition [3, 4, 17].", + "bbox": [ + 75, + 138, + 468, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to further improve the spatial complexity while maintaining the rendering quality. Leveraging the decades of research on standard compression algorithms [35,36,46], we propose compressing grid-based neural fields using frequency-based transformations. In the frequency domain, a large portion of the coefficients can be discarded without considerably degrading the reconstruction quality, and most standard compression algorithms have exploited frequency domain representations. Thus, we propose using this property on grid-based neural fields to maximize parameter efficiency. Among other alternatives, we employ the discrete wavelet transform (DWT) due to its compactness and ability to efficiently capture both global and local information.", + "bbox": [ + 75, + 306, + 468, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Once we obtain sparse representations via frequency domain representations, we can take advantage of the existing compression techniques. Unlike conventional media data (e.g., image and audio), however, no off-the-shelf compression tools exist that we can leverage without complicated engineering efforts. In addition, since NeRF, or neural rendering networks in general, is a relatively new data format, the characteristics or patterns of their coefficients have not been thoroughly investigated. Thus, we present a compression pipeline for our purposes. To automatically filter out unnecessary coefficients, we propose a trainable binary mask. For each 3D scene, we jointly optimize grid parameters and their corresponding masks. This perscene optimization strategy can be more optimal than the global quantization table used in standard image compression codecs [43].", + "bbox": [ + 75, + 518, + 468, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To compress sparse grid representations, we first merge masks with wavelet coefficients to zero out coefficients and then apply standard compression algorithms to masked coefficients. We utilize the run-length encoding to encode binary information about which coefficients are non-zero. For further compression, we apply one of the entropy coding algorithms, Huffman encoding [19], to these encoded outputs.", + "bbox": [ + 75, + 762, + 468, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method incurs negligible computational costs at test time, requiring only one inverse DWT (IDWT) per grid. Af-", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ter the IDWT, the wavelet grids are transformed into spatial grids, eliminating the need for additional IDWT during rendering. As a result, the computing time and costs (including memory costs) are identical to the original spatial grid-based NeRF.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 517, + 167, + 818, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose using wavelet coefficients to improve parameter sparsity and reconstruction quality. Through experiments, we show that the wavelet coefficients can be more compact than the spatial domain coefficients in neural radiance fields.", + "- We propose a trainable mask that can be applied to any grid-based neural representation. Experimental results demonstrate that our proposed masking method can zero out more than $95\\%$ of the total grid parameters while maintaining high reconstruction quality.", + "- We achieve state-of-the-art performance in novel view synthesis under a memory budget of 2 MB." + ], + "bbox": [ + 517, + 193, + 890, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 412, + 648, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neural rendering", + "text_level": 1, + "bbox": [ + 500, + 438, + 669, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural radiance fields. NeRF [23] has demonstrated that 3D scenes and objects can be successfully represented using coordinate-based neural representation. NeRF renders a scene by casting a ray per pixel and sampling colors and opacity from points that lie on the ray. However, since this approach relies on the dense, point-wise sampling of color values and opacity, the computational costs are significant.", + "bbox": [ + 496, + 460, + 890, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to this inefficiency, there has been considerable interest in improving computational efficiency. A plethora of methods propose to adopt data structures such as trees [45, 49], point clouds [27, 28, 48, 50], and grids [4, 13, 15, 24, 33, 37, 38] to efficiently represent 3D scenes and objects. PlenOctrees [49] use an octree structure for real-time rendering. Despite the reduced rendering costs, the increased size of more than ten times the size of standard MLP-based neural fields, such as NeRF [23], limits their wide application. Point cloud-based approaches [27, 48] reduce the number of points sampled per ray by utilizing the scene geometry of a point cloud. Lastly, a number of studies have demonstrated the efficiency of using grid structures, including vectors, matrices, and tensors, on neural fields [4, 13, 24, 37]. With the use of these grid structures, training iterations that once required tens of hours can now be completed in a matter of minutes. Furthermore, real-time rendering at inference time is also possible. However, dense 3D grid structures require substantial amounts of memory, often exceeding 1 GB. To alleviate this memory requirement, several methods have been proposed to reduce the size of grids [4, 24, 38, 42].", + "bbox": [ + 496, + 568, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20681", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve near-instant rendering with reasonable memory requirements, Instant-NGP [24] proposed using hash-based multi-resolution grids. Using hash functions, each grid maps input coordinates to corresponding feature vectors. Similarly, VQAD [38] proposed using codebooks and vector quantization rather than the hash function. This enables control over the overall size of the neural fields. However, during training, a sizable amount of memory is required to learn which code from the codebook to use for each point on the grid.", + "bbox": [ + 75, + 90, + 470, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to alleviate the memory requirement of using 3D grids, another line of work decomposes 3D grids into lower dimensional representations, such as planes and vectors [3, 4, 17, 42]. EG3D [3] proposes a tri-plane approach, which represents a 3D scene with three perpendicular planes and uses feature vectors, extracted separately from each plane, as inputs for the following MLPs. TensoRF [4] is another parameter-efficient yet expressive method for decomposing dense three-dimensional grids into a smaller set of parameters, such as planes and lines. Although these methods dramatically reduce the time and space complexity of the 3D scene and object representation, their overall sizes are larger than those of MLP-only methods.", + "bbox": [ + 75, + 244, + 470, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frequency-based representation. Several studies have explored frequency-based parameterization for efficient neural field representations [17, 41, 45, 47]. Fourier PlenOctree [45] has demonstrated that the Fourier transform can improve both parameter efficiency and training speed of the PlenOctree structure [49]. Nevertheless, the overall size is larger than MLP-only methods. PREF [17] has also exploited the Fourier transform in 3D scene representation. It applies the Fourier transform to a tensor decomposition-based representation. However, it has not shown notable improvements in both the representation quality and the parameter efficiency compared to spatial grid representations.", + "bbox": [ + 75, + 459, + 468, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides the Fourier transform, there is another line of signal representations in the frequency domain: the wavelet transform. The wavelet transform decomposes a signal using a set of basis functions called wavelets. The wavelet transform can extract both frequency and temporal information, in contrast to the Fourier transform, which can only capture frequency information. Because each coefficient covers a different frequency and period in time, it is known that the wavelet transform can represent transient signals more compactly than Fourier-based transforms. This compactness contributes to standard compression CODECs, such as JPEG 2000 [35]. Motivated by the success of the standard CODECs, we aim to mimic their compression pipeline and apply it to the neural radiance fields. One disadvantage of the wavelet transform is that it can be less efficient than Fourier-based transforms for smooth signals. However, we believe that most situations in grid-based neural fields", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are not the case, considering that each grid's resolution is constrained to accurately depict detailed scenes and objects with limited memory budgets.", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Model compression", + "text_level": 1, + "bbox": [ + 500, + 156, + 686, + 171 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, many works have studied to downsize neural networks through various techniques such as weight pruning and quantization.", + "bbox": [ + 498, + 181, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pruning. For efficient representations, pruning methods have been explored in neural fields. Point-NeRF [48] proposes a pruning strategy for point cloud representation by imposing a sparsity loss on point confidence. Re:NeRF [8] has validated that simply applying a general pruning technique, which gradually removes coefficients throughout the course of training, to grid-based neural representations results in a significant performance drop. It proposed iterative parameter removal and conditional parameter inclusion and demonstrated comparable compression performance. However, it has not shown desirable performance under limited memory budgets. Similarly, KiloNeRF [32] prunes out parameters for unused areas using a fixed threshold. Instead, we propose a trainable mask that can remove a large number of coefficients while causing only minor degradation in the reconstruction quality.", + "bbox": [ + 496, + 229, + 890, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quantization. In terms of neural network quantization, there are two main approaches: post-quantization (PTQ) [21, 25] and quantization-aware training (QAT) [7, 18, 30, 51]. Post-quantization methods quantify network parameters after training is finished. PTQ has the advantage that it does not require additional training iterations or datasets. Furthermore, it supports arbitrary-bit quantization. Performance, however, might degrade because the network parameters were not optimized for the target bit precision. Unlike post-quantization, QAT optimizes neural networks for a certain bit precision during training iterations. Since quantization is not differentiable, it usually relies on the straight-through estimator (STE) [1] for network optimization [7]. QAT has been validated for its effectiveness in neural representation models. However, PTQ has the disadvantage of requiring additional training samples and iterations when quantizing the weights of a pretrained network. Since we train neural fields from scratch, we use PTQ for weight quantization.", + "bbox": [ + 496, + 474, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 786, + 598, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Neural radiance fields", + "text_level": 1, + "bbox": [ + 500, + 813, + 705, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a neural radiance field leveraging grid representations. It takes an input coordinate $x \\in \\mathbb{R}^3$ and viewing direction, $d \\in \\mathbb{R}^2$ , generating a four-dimensional vector consisting of a density and three-channel RGB colors, de", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20682", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7f76d1a1d654f1e03ca4723bbfb1f3ebf321d4586161904668b80bb5ce1ad621.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our model. Images with purple borders illustrate wavelet coefficients, while those with orange borders visualize spatial coefficients. The wavelet coefficients in each 2D grid are multiplied with a binarized mask to form a masked wavelet coefficient grid. Masked wavelet coefficients are then inverse-transformed to spatial features. We sample feature vectors for input coordinates using bilinear interpolation on the grids. Then, opacity and color for each input coordinate are estimated using the sampled feature vectors following the defined model." + ], + "image_footnote": [], + "bbox": [ + 86, + 93, + 893, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fined as follows.", + "bbox": [ + 76, + 304, + 189, + 318 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma (x) = f _ {\\sigma} \\left(x; \\gamma_ {\\sigma}, \\mathcal {M} _ {\\sigma}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 329, + 468, + 345 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc (x, d) = f _ {c} (x, d; \\theta , \\gamma_ {c}, \\mathcal {M} _ {c}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 348, + 468, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ is the parameter of an MLP, $\\gamma = \\{\\gamma_{\\sigma}, \\gamma_{c}\\}$ is a set of grid parameters, and $\\mathcal{M} = \\{\\mathcal{M}_{\\sigma}, \\mathcal{M}_{c}\\}$ is a set of masks for grid parameters, which will be described shortly (Sec. 3.3). The following volumetric rendering equation is used to synthesize novel views:", + "bbox": [ + 76, + 372, + 468, + 448 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (r (t)) c (r (t), d) d t, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 455, + 468, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (r (s)) d s\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 494, + 468, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r(t)$ is a ray from a camera viewpoint, and $C(r)$ is the expected color of the ray $r(t)$ . Two integral bounds (near and far) are denoted as $t_n$ and $t_f$ , respectively. Please refer to NeRF [23] for more details.", + "bbox": [ + 75, + 536, + 468, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Wavelet transform on decomposed tensors", + "text_level": 1, + "bbox": [ + 76, + 606, + 437, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frequency-based algorithms, such as discrete cosine transform (DCT) or discrete wavelet transform (DWT), have been developed and improved over the decades. For high compression performance, standard CODEs have used them to make sparse representations. However, in the case of 3D grid representation, their computational complexity increases cubically with grid resolution, making high-resolution 3D scene and object representations impractical. In order to use frequency-based algorithms for compact representation, we need lower-dimensional data, such as 2D planes or 1D lines.", + "bbox": [ + 75, + 628, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent studies have explored various decomposition methods to reduce the spatial complexity of 3D grid-based representations. Among them, plane-based representations have achieved remarkable success in reducing the number of parameters while maintaining the rendering performance [3, 4, 17]. They propose decomposing a 3D tensor into a set of 2D planes or 1D vectors.", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To combine the best of both worlds, we propose using the wavelet transform on 2D plane-based neural fields [4]. We use the wavelet transform because of its compactness, especially for non-repeating, non-smooth signals.", + "bbox": [ + 496, + 304, + 890, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For efficient 3D object and scene representation, recent studies proposed using a lower-dimensional grid (2D planes or 1D lines) [3, 4]. For example, EG3D [3] utilizes three 2D planes (tri-plane) and TensoRF [4] employs three sets, each consisting of a plane and a vector. We describe our method based on TensoRF. In detail, we use a set of 2D matrices and 1D vectors for grid representation, $\\gamma_{\\sigma} = \\{\\mathcal{W}_{\\sigma,r}^{x}, \\mathcal{W}_{\\sigma,r}^{y}, \\mathcal{W}_{\\sigma,r}^{z}, v_{\\sigma,r}^{x}, v_{\\sigma,r}^{y}, v_{\\sigma,r}^{z}\\}_{r=1}^{N_{\\sigma,r}}$ ( $\\sigma$ denotes density, and we will omit the subscript $\\sigma$ for brevity from now on) as proposed in TensoRF [4]. $N_{r}$ is the number of ranks in matrix-vector decomposition and $\\mathcal{W}_r^x \\in \\mathbb{R}^{H \\times W}$ , $\\mathcal{W}_r^y \\in \\mathbb{R}^{W \\times D}$ , $\\mathcal{W}_r^z \\in \\mathbb{R}^{H \\times D}$ are matrices, $v_r^x \\in \\mathbb{R}^D$ , $v_r^y \\in \\mathbb{R}^H$ , $v_r^z \\in \\mathbb{R}^W$ , are vectors in $x, y, z$ directions, respectively. $H, W, D$ are the resolution of the grid. More formally, a 3D grid representation $G$ can be defined as follows.", + "bbox": [ + 496, + 364, + 892, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} v _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\mathcal {W} _ {r} ^ {d}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 604, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes the outer product and $\\mathrm{idwt}(\\cdot)$ is a two-dimensional inverse discrete wavelet transform (IDWT).", + "bbox": [ + 496, + 651, + 890, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With a slight abuse of notation, a single level IDWT of a matrix $\\mathcal{W} \\in \\mathcal{R}^{m \\times n}$ can be written as follows.", + "bbox": [ + 496, + 681, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {i d w t} (\\mathcal {W}) = \\Phi \\star (\\uparrow 2) \\left(\\mathcal {W} _ {L L}\\right) + \\Psi_ {H L} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {H L}\\right) \\\\ + \\Psi_ {L H} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {L H}\\right) + \\Psi_ {H H} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {H H}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 715, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{W}_{LL},\\mathcal{W}_{HL},\\mathcal{W}_{LH},\\mathcal{W}_{HH}\\in \\mathbb{R}^{m / 2\\times n / 2}$ denote approximation, horizontal, vertical, and diagonal coefficients, respectively. $\\star$ denotes 2D cross correlation and $(\\uparrow 2):$ $\\mathbb{R}^{m / 2\\times n / 2}\\to \\mathbb{R}^{m\\times n}$ denotes (nearest neighbor) upscaling. $\\Psi$ and $\\Phi$ are wavelet function and scaling function, respectively. We use the biorthogonal 4.4 wavelet function and its corresponding scaling function [6]. Even though we follow the grid representation of TensoRF, our method is not limited to the existing 2D plane-based representations but can", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20683", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be applied to any other approaches that adopt 2D planes to represent higher-dimensional structures.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Eq. 5, we only apply IDWT over matrices $\\mathcal{W}$ since we observed that IDWT to vectors resulted in significant quality degradation. As this transformation is differentiable, the model can be trained end-to-end. We can also obtain the 3D grid representation for appearance through the same process and provide all formulations in the supplementary material.", + "bbox": [ + 75, + 121, + 468, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method incurs insignificant costs during inference. Once training is finished, only one IDWT per grid is needed to transform wavelet grids into spatial grids. Thus, the remaining computational costs and time are exactly the same as the original spatial grid-based neural fields.", + "bbox": [ + 75, + 227, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Multi-level wavelet transform", + "text_level": 1, + "bbox": [ + 76, + 324, + 336, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further improve parameter sparsity and reconstruction quality, we propose using a multi-level wavelet transform. We experimentally found that higher-level wavelet transformations result in higher sparsity in grids (Sec. 4.3.1). However, naively using a multi-level wavelet transform on grid parameters degrades the representation quality (Tab. 1). We hypothesize that this is because the range of wavelet coefficients and the range of their gradients vary depending on the level of decomposition. In more detail, high-pass coefficients typically have a larger gradient scale with a smaller weight range. Different weight scales and gradients for different frequencies seem to hinder the optimization process. As such, we propose multiplying wavelet coefficients by the scales proportional to the inverse of frequency. With our proposed scaling factors $s$ , Eq. 5 can be rewritten as follows.", + "bbox": [ + 75, + 348, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} v _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\mathcal {W} _ {r} ^ {d} \\odot s _ {r} ^ {d}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 599, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the Hadamard product. For example, a two-level DWT, as shown in Fig. 3, generates a total of seven groups of coefficients (LL2, HL2, LH2, HH2, HL1, LH1, HH1). The scaling factors $s$ for HL2, LH2, and HH2 will be set to $1/2$ , and for HL1, LH1, and HH1 to $1/3$ . This scaling method enhances the quality of the reconstruction; more experiments are described in Sec. 4.3.2.", + "bbox": [ + 75, + 655, + 468, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning to mask", + "text_level": 1, + "bbox": [ + 76, + 771, + 248, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Even though Wavelet coefficients can be sparse, we need additional methods to attain higher sparsity. Thus, we propose using element-wise masks to increase the portion of zero elements in grids. By jointly optimizing binary masks and grid parameters, we aim to zero out the majority of coefficients without significantly degrading the rendering quality. As we defined earlier,", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e0bf6b7659ae5bf5a24a39da257e851199bc516faada2a3a9a074c6d9d5d7dbe.jpg", + "image_caption": [ + "LL2", + "HL2" + ], + "image_footnote": [], + "bbox": [ + 521, + 88, + 694, + 123 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/70c07940a878603facf6acc54b1582d82250b83c0faf013bcbeb8176fc3a5f95.jpg", + "image_caption": [ + "HL2", + "HH2" + ], + "image_footnote": [], + "bbox": [ + 521, + 138, + 694, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a1537c997984619c22ce6ca166adb0c13877ab5fbac013c6a15bf2635b59e89.jpg", + "image_caption": [ + "HL1" + ], + "image_footnote": [], + "bbox": [ + 699, + 88, + 870, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/de9ae86b7d4249439733caf2b6e4e02910d9c4ae19e8cc6df599ad42ac5c2c29.jpg", + "image_caption": [ + "HL1", + "Figure 3. Multi-level discrete wavelet coefficients (2-level). LL, HL, LH, and HH denote approximation, horizontal, vertical, and diagonal coefficients, respectively. The numbers on the right denote the level of decomposition." + ], + "image_footnote": [], + "bbox": [ + 521, + 186, + 694, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f3e0dbf2714f67350e973b8c433c70828df89e78e2a414162e2d64c795efcc9b.jpg", + "image_caption": [ + "HH1" + ], + "image_footnote": [], + "bbox": [ + 697, + 186, + 870, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we have a set of trainable element-wise masks, $\\mathcal{M} = \\{\\mathcal{M}_r^x,\\mathcal{M}_r^y,\\mathcal{M}_r^z,m_r^x,m_r^y,m_r^z\\}_{r = 1}^{N_r}$ (note that we omitted the subscript $\\sigma$ for brevity). During training, the grid parameters and corresponding binarized masks are multiplied element by element. Since calculating gradients directly from binarized masks is not feasible, we used the straight-through-estimator technique [1] to train and use masks. Formally, the masked matrix parameters $\\widehat{\\mathcal{W}}_r$ for $\\mathcal{W}_r$ can be expressed as follows,", + "bbox": [ + 498, + 375, + 890, + 511 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {W}} _ {r} = \\operatorname {s g} \\left(\\left(\\mathcal {H} \\left(\\mathcal {M} _ {r}\\right) - \\sigma \\left(\\mathcal {M} _ {r}\\right)\\right) \\odot \\mathcal {W} _ {r}\\right) + \\sigma \\left(\\mathcal {M} _ {r}\\right) \\odot \\mathcal {W} _ {r}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 518, + 890, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathfrak{sg}(\\cdot)$ is the stop-gradient operator. $\\sigma (\\cdot)$ and $\\mathcal{H}(\\cdot)$ denote the element-wise sigmoid and Heaviside step function, respectively. We can also compute the masked vector parameters $\\widehat{v}_r$ similarly. By replacing grid parameters $(\\mathcal{W}_r,$ $v_{r})$ from Eq. (7) with masked grid parameters $(\\widehat{\\mathcal{W}}_r,\\widehat{v}_r)$ , we can represent a masked 3D grid representation as follows:", + "bbox": [ + 498, + 561, + 890, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} \\widehat {v} _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\widehat {\\mathcal {W}} _ {r} ^ {d} \\odot s _ {r} ^ {d}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 662, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To make binarized mask values sparse, we use the sum of all mask values as the additional loss term $\\mathcal{L}_m$ . The overall loss function is a sum of rendering loss $\\mathcal{L}_r$ and the mask regularization term $\\mathcal{L}_m$ .", + "bbox": [ + 498, + 714, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r} + \\lambda_ {m} \\mathcal {L} _ {m}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 784, + 890, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $\\lambda_{m}$ to control the sparsity of the parameters of the grid representation.", + "bbox": [ + 498, + 809, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Compression pipeline", + "text_level": 1, + "bbox": [ + 500, + 847, + 702, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With our proposed masking method and multi-level DWT, the ratio of zeros in grids can go up to more than", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20684", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/10b955614e42106c4b08fca6b6ef434184514211e81702172d5f6d94a0eb9f82.jpg", + "image_caption": [ + "Figure 4. Proposed bitmap compression pipeline. RLE and Huffman denote run-length encoding and Huffman encoding, respectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 472, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "95%. However, sparse representations themselves do not reduce the total size. In this section, we describe our proposed compression pipeline for sparse grid parameters. Instead of storing grids as they are, we separately store nonzero coefficients and bitmaps (or masks) that indicate which coefficients are non-zero. Despite using 1-bit bitmaps, the overall bitmap size is large due to the large number of parameters. To reduce the bitmap sizes, we propose a compression pipeline with the following three stages: n-bit casting, run-length encoding (RLE), and the Huffman encoding [19]. The compression pipeline is illustrated in Fig. 4.", + "bbox": [ + 75, + 310, + 468, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Before applying the compression pipeline, we split wavelet coefficient masks by level of decomposition. For the 2-level wavelet transform, for instance, we group HL1, LH1, and HH1, then HL2, LH2, and HH2, and finally LL2 (Fig. 3). This is based on our experimental results that wavelet coefficients with the same levels of sparsity have similar levels of sparsity (Fig. 6). We found that grouping coefficients with similar sparsity results in better compression performance. In addition, we found that directly applying RLE to 1-bit streams causes inefficient bit allocation due to the numerous repeating zeros. Therefore, we first cast the binary mask values to n-bit unsigned integers, and perform RLE afterward. In our experiments, we set n to 8. Finally, we apply the Huffman encoding algorithm to the RLE-encoded streams to map values with a high probability to shorter bits.", + "bbox": [ + 75, + 477, + 470, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 731, + 209, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental settings", + "text_level": 1, + "bbox": [ + 76, + 756, + 281, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared our proposed method with other baselines in four datasets; NeRF synthetic dataset [23], Neural Sparse Voxel Fields (NSVF) [22] dataset, TanksAndTemps dataset [20], and LLFF dataset [10].", + "bbox": [ + 75, + 779, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used TensoRF VM-192 as the baseline model and applied the proposed method. To control the size of our method, we adjusted the mask regularization weight $\\lambda_{m}$ the number of grid channels (from VM-192 to VM-384),", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the resolution of grids. Unless otherwise specified, we used 8-bit quantization for every method. We compared our method with quantized NeRF, TensoRF models (CP and VM), cNeRF [2] and CCNeRF [42]. We also tried to compare our approach with VQAD [38], which efficiently compresses 3D scene representations using codebooks and vector quantization. However, VQAD requires depth maps to prune out vacant areas in advance. Otherwise, the required memory exceeds the memory capacity of the Tesla A100 equipped with 40 GB of memory. Since the datasets we used do not provide depth maps, we did not compare VQAD with ours.", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We followed the experimental settings of TensoRF [4]. We trained models for 30,000 iterations, each of which is a minibatch of 4,096 rays. We used the Adam optimizer and an exponential learning rate decay scheduler. Following TensoRF, the initial learning rates of grid-related parameters and MLP-related parameters were set to 0.02 and 0.001, respectively. Final learning rates were set at 1/10 of the initial learning rate. We updated the alpha masks at the 2000th, 4000th, 6000th, 11000th, 16000th, 21000th, and 26000th iterations. We set $\\lambda_{m}$ in Eq. (10) to 1e-10 for high parameter sparsity and 5e-11 and 2.5e-11 for relatively lower sparsity. We set the initial values of masks $\\mathcal{M}$ to one, and set their initial learning rate to the same learning rate as grid parameters $\\phi$ . As a reconstruction quality measurement, we used PSNR.", + "bbox": [ + 496, + 273, + 892, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 500, + 511, + 591, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figs. 1 and 5 show the quantitative performances of our method and the baselines on various datasets. Each graph displays the average PSNR and size of methods. More detailed numbers are available in the supplementary materials.", + "bbox": [ + 496, + 536, + 890, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As the rate-distortion curves show, our method achieves higher efficiency than the baseline models and other methods. Even with the base network (VM-192), our method consistently outperforms other baselines on all novel view synthesis datasets, under a small memory budget of $2\\mathrm{MB}$ . We adjusted both $\\lambda_{m}$ and the network size to increase the output size. By adjusting $\\text{lambda}_{m}$ , we can control both size and quality while having no effect on computational costs, time, or memory requirements. Increasing network size (doubling the number of channels and increasing the resolution) can push the boundary even further. It is also worth mentioning that applying our method to a larger backbone model (VM-384) outperforms models with a smaller backbone model (VM-192), even with the same value of $\\lambda_{m}$ or a similar model size. However, as the network grows in size, the computational costs and time required for training and testing increase. Therefore, it is still reasonable to use a small backbone model when using our proposed method.", + "bbox": [ + 496, + 598, + 892, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 7 shows the qualitative results with and without our proposed masking method. Trainable masks removed more", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20685", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d9d79b8ea73f76a1eaf5fd70eb08c89818029238dbc68514cceb8d202447e47.jpg", + "image_caption": [ + "Figure 5. Rate distortion curves on the NSVF synthetic, TanksAndTemples, and LLFF dataset. We used 8-bit precision for every method." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 356, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25bcb6b6b6bbee4b92d2c296d685755351782de0d5bcbd51383be3f8ce81e2b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 90, + 612, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e50f7c0d0a55485ef572a64b096b41516b98767db7ce05aeeaab61b223533c74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 90, + 867, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8609fa5d6a211f7a711d10dc4fa5086413051cc09548871985880c858dcb7ca5.jpg", + "image_caption": [ + "(a) Raw mask", + "Figure 6. Qualitative analysis on mask $\\mathcal{M}$ (Chair in the NeRF synthetic dataset)." + ], + "image_footnote": [], + "bbox": [ + 98, + 292, + 254, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2067b199aa78fb46d8bb225d83e59f7cb9868a841cc091847681d9a2f74bc6b3.jpg", + "image_caption": [ + "(b) Binarized mask" + ], + "image_footnote": [], + "bbox": [ + 256, + 292, + 450, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "than $97\\%$ of the grid parameters; nevertheless, the rendered results are still accurate, and the qualitative difference is almost imperceptible. This demonstrates that our proposed method efficiently eliminates unnecessary wavelet coefficients by leveraging trainable masks.", + "bbox": [ + 75, + 518, + 468, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 illustrates both raw and binarized masks after training. First of all, sparsity varies depending on the level of the wavelet transform. As shown in Fig. 6b, coefficients with lower frequencies have lower sparsity, whereas coefficients with higher frequencies have higher sparsity. What is interesting is that the raw mask values seem to reflect the characteristics of corresponding wavelet coefficients. Vertical, horizontal, and diagonal patterns can be found in the raw mask values for the vertical, horizontal, and diagonal coefficients, respectively.", + "bbox": [ + 75, + 599, + 470, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 780, + 236, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze each component of our proposed method. Every experiment was conducted on the whole NeRF synthetic dataset, and we used the average PSNR as a measurement for representation quality. We used 4-level wavelet transform and set $\\lambda_{m}$ to 1e-10, unless otherwise specified.", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a97d65edb55746a3cb2ce9922e464730f5edf6cb7776c0a305694458175ecc8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPSNR
DWT31.949
DWT w/o level-wise scaling31.145
DCT31.325
", + "bbox": [ + 558, + 291, + 834, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Ablation study on level-wise scaling. Without our proposed scaling method, the performance of DWT is worse than DCT.", + "bbox": [ + 498, + 369, + 890, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Discrete wavelet transform", + "text_level": 1, + "bbox": [ + 500, + 439, + 741, + 453 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the compactness of wavelet coefficients and the performance improvements brought by the multi-level wavelet transform. Fig. 8 shows the rate-distortion curves of four different levels of wavelet transform and non-wavelet coefficients. As shown in the figure, wavelet-based grid representation outperforms the grid in the spatial domain, especially when the sparsity is high. Furthermore, rate-distortion curves demonstrate that higher level of wavelet transform improves visual quality and sparsity.", + "bbox": [ + 496, + 463, + 890, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Level-wise wavelet coefficient scaling", + "text_level": 1, + "bbox": [ + 498, + 633, + 810, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.2.1, we proposed multiplying wavelet coefficients by the scales proportional to the inverse of frequency. Thus, we compared two versions of wavelet transform; one with our proposed weight scaling (Sec. 3.2.1) and the other without it. As Tab. 1 shows, scaling wavelet coefficients improves reconstruction performance.", + "bbox": [ + 496, + 659, + 890, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3 Wavelet functions", + "text_level": 1, + "bbox": [ + 500, + 785, + 679, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze how different wavelet functions affect reconstruction quality. For comparison, we selected several wavelet functions: Haar, Coiflet 1, Daubechies 4, and reverse biorthogonal 4.4. As shown in Tab. 2, the type of wavelet function has little effect on reconstruction quality. Even Haar, the simplest wavelet function, performs", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "footer", + "text": "20686", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/763c28005c70c802cd1afc5bcb97d18903563016fbd1e36ffc828e36abcba959.jpg", + "image_caption": [ + "Figure 7. A qualitative analysis of grid sparsity. The sparsity refers to the ratio of zero elements in grids. With the help of trainable masks, removing more than $97\\%$ of grid parameters only results in an imperceptible difference." + ], + "image_footnote": [], + "bbox": [ + 127, + 88, + 509, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/971911b1e1283ea2eb744dc52f019a7cb5b3bb6c16faa846332575c31ae3a86a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 90, + 823, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7879939e62fbc4841de393219326c0d9f9dfe9087abd4ccf68f03155f704e6e.jpg", + "image_caption": [ + "Figure 8. The rate distortion curves of different signal representation schemes (spatial, DWT, and DCT). Sparsity on the $x$ axis refers to the ratio of zeros in grid parameters. The grid sparsity was controlled by $\\lambda_{m}$ . The numbers inside the parenthesis indicate the levels of the wavelet transform." + ], + "image_footnote": [], + "bbox": [ + 84, + 316, + 460, + 510 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4d06dd13213f24f24eb0fa013a931f48fe62ba19d605f1071b64d6efa4ce29bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Wavelet FunctionPSNR
Haar31.889
Coiflets 131.846
Daubechies 431.734
biorthogonal 4.4 (default)31.949
reverse biorthogonal 4.431.727
", + "bbox": [ + 138, + 609, + 408, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Reconstruction performances of different wavelet functions.", + "bbox": [ + 76, + 719, + 468, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fairly well. Still, the biorthogonal 4.4 that we used shows the best performance among selected wavelet functions.", + "bbox": [ + 76, + 776, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.4 Discrete cosine transform", + "text_level": 1, + "bbox": [ + 76, + 830, + 308, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we compare the performance of DWT with that of DCT, which represents a signal with respect to the sum of cosine functions with different frequencies. DCT,", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "similar to DWT, has an energy compaction property. That is, a small number of non-zero DCT coefficients are sufficient to represent a given signal. Because of its energy compactness, DCT or its variants are another widely used signal transformation for image, audio, and audio compression. However, as shown in Fig. 8 and Tab. 1, DWT shows superior compression and representation performance to DCT. We believe this is due to the non-repeating, non-smooth information in grids, for which DWT is more appropriate.", + "bbox": [ + 498, + 310, + 892, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 463, + 617, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a compact representation for grid-based neural fields, enabled by a novel masking strategy and multilevel wavelet transform. We demonstrate that these two components made it possible to remove more than $95\\%$ of grid parameters without a significant loss of visual quality. With our proposed compression pipeline and the sparse wavelet coefficients, we achieved state-of-the-art performance under a memory budget of 2 MB. We believe we can reduce the size further with more sophisticated compression pipelines.", + "bbox": [ + 496, + 488, + 892, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed method is naturally constrained by the limitations of grid-based neural fields that are developed for bounded scenes or objects. We believe that expanding this grid-based representation to encompass unbounded or large scenes would be an intriguing direction, given that we can now compactly represent 3D objects with our proposed representation scheme.", + "bbox": [ + 496, + 641, + 893, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 763, + 660, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Institute of Information and Communications Technology Planning and Evaluation (IITP) grants (2021-0-02068, IITP-2021-0-02052, IITP-2023-2020-0-01821, IITP-2019-0-00421) and National Research Foundation of Korea (NRF) grants (2022R1A4A3032913, 2022R1F1A1064184, 2022R1A4A3033571) funded by the Korea Government (MSIT).", + "bbox": [ + 496, + 787, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20687", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3, 5", + "[2] Thomas Bird, Johannes Balle, Saurabh Singh, and Philip A Chou. 3d scene compression through entropy penalized neural representation functions. In 2021 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2021. 6", + "[3] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, June 2022. 2, 3, 4", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision. Springer, 2022. 2, 3, 4, 6", + "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 1", + "[6] Albert Cohen, Ingrid Daubechies, and J-C Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on pure and applied mathematics, 45(5):485-560, 1992. 4", + "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. 3", + "[8] Chenxi Lola Deng and Enzo Tartaglione. Compressing explicit voxel grid representations: fast nerfs become also small. arXiv preprint arXiv:2210.12782, 2022. 3", + "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1", + "[10] Ben Mildenhall et al. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6", + "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, SA '22, New York, NY, USA, 2022. Association for Computing Machinery. 1", + "[12] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning," + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "volume 70 of Proceedings of Machine Learning Research, pages 1126-1135. PMLR, 06-11 Aug 2017. 1", + "[13] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, June 2022. 2", + "[14] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14346-14355, October 2021. 1", + "[15] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. pages 5875-5884, October 2021. 2", + "[16] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12902-12911, June 2022. 1", + "[17] Binbin Huang, Xinhao Yan, Anpei Chen, Shenghua Gao, and Jingyi Yu. Pref: Phasorial embedding fields for compact neural representations. arXiv preprint arXiv:2205.13524, 2022. 2, 3, 4", + "[18] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Binarized neural networks. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 3", + "[19] David A Huffman. A method for the construction of minimum-redundancy codes. Proceedings of the IRE, 40(9):1098-1101, 1952. 2, 6", + "[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), jul 2017. 6", + "[21] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. 3", + "[22] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 15651-15663. Curran Associates, Inc., 2020. 2, 6", + "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 4, 6", + "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), jul 2022. 2, 3", + "[25] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? Adaptive" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20688", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rounding for post-training quantization. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7197-7206. PMLR, 13-18 Jul 2020. 3", + "[26] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 1", + "[27] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. pages 18419-18429, June 2022. 2", + "[28] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 13032-13044. Curran Associates, Inc., 2021. 2", + "[29] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 1", + "[30] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision – ECCV 2016, pages 525–542, Cham, 2016. Springer International Publishing. 3", + "[31] Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, and Andrea Tagliasacchi. Derf: Decomposed radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14153-14161, 2021. 1", + "[32] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14335-14345, October 2021. 1, 3", + "[33] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. VoxGRAF: Fast 3d-aware image synthesis with sparse voxel grids. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2", + "[34] Vincent Sitzmann, Eric Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 10136-10147. Curran Associates, Inc., 2020. 1", + "[35] Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The jpeg 2000 still image compression standard. IEEE Signal processing magazine, 18(5):36-58, 2001. 2, 3", + "[36] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 2", + "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, June 2022. 2", + "[38] Towaki Takikawa, Alex Evans, Jonathan Tremblay, Thomas Müller, Morgan McGuire, Alec Jacobson, and Sanja Fidler. Variable bitrate neural fields. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 6", + "[39] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. arXiv preprint arXiv:2101.10994, 2021. 2", + "[40] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P. Srinivasan, Jonathan T. Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 1, 2", + "[41] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3", + "[42] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable neRF via rank-residual decomposition. In Advances in Neural Information Processing Systems, 2022. 2, 3, 6", + "[43] Gregory K. Wallace. TheJPEG still picture compression standard.Commun.ACM,34(4):30-44,apr 1991.2", + "[44] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 612-629, Cham, 2022. Springer Nature Switzerland. 1", + "[45] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13524-13534, June 2022. 2, 3", + "[46] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology, 13(7):560-576, 2003. 2", + "[47] Zhijie Wu, Yuhe Jin, and Kwang Moo Yi. Neural fourier filter bank. arXiv e-prints, pages arXiv-2212, 2022. 3", + "[48] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2, 3", + "[49] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5752-5761, October 2021. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20689", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. arXiv preprint arXiv:2205.14330, 2022. 2", + "[51] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained ternary quantization. In International Conference on Learning Representations, 2017. 3" + ], + "bbox": [ + 78, + 90, + 468, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20690", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_model.json b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d068b43f2a68a783799f1464521e1cbf322ec76 --- /dev/null +++ b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.836, + 0.154 + ], + "angle": 0, + "content": "Masked Wavelet Representation for Compact Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.179, + 0.203, + 0.199 + ], + "angle": 0, + "content": "Daniel Rho1*" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.181, + 0.393, + 0.199 + ], + "angle": 0, + "content": "Byeonghyeon Lee\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.181, + 0.564, + 0.233 + ], + "angle": 0, + "content": "Seungtae Nam² \nEunbyung Park²,³† \n¹AI2XL, KT" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.181, + 0.703, + 0.198 + ], + "angle": 0, + "content": "Joo Chan Lee²" + }, + { + "type": "text", + "bbox": [ + 0.738, + 0.181, + 0.884, + 0.199 + ], + "angle": 0, + "content": "Jong Hwan \\(\\mathrm{Ko}^{2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.234, + 0.744, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Artificial Intelligence, Sungkyunkwan University" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.252, + 0.806, + 0.271 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Electrical and Computer Engineering, Sungkyunkwan University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.474, + 0.745 + ], + "angle": 0, + "content": "Neural radiance fields (NeRF) have demonstrated the potential of coordinate-based neural representation (neural fields or implicit neural representation) in neural rendering. However, using a multi-layer perceptron (MLP) to represent a 3D scene or object requires enormous computational resources and time. There have been recent studies on how to reduce these computational inefficiencies by using additional data structures, such as grids or trees. Despite the promising performance, the explicit data structure necessitates a substantial amount of memory. In this work, we present a method to reduce the size without compromising the advantages of having additional data structures. In detail, we propose using the wavelet transform on grid-based neural fields. Grid-based neural fields are for fast convergence, and the wavelet transform, whose efficiency has been demonstrated in high-performance standard codes, is to improve the parameter efficiency of grids. Furthermore, in order to achieve a higher sparsity of grid coefficients while maintaining reconstruction quality, we present a novel trainable masking approach. Experimental results demonstrate that non-spatial grid coefficients, such as wavelet coefficients, are capable of attaining a higher level of sparsity than spatial grid coefficients, resulting in a more compact representation. With our proposed mask and compression pipeline, we achieved state-of-the-art performance within a memory budget of 2 MB. Our code is available at https://github.com/daniel03c1/maskedwavelet_nerf." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.776, + 0.21, + 0.792 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.802, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Recent advances in coordinate-based neural representation (neural fields or implicit neural representation) have demonstrated remarkable performance in many applications. In particular, neural radiance fields (NeRF) have" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.305, + 0.891, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.495, + 0.893, + 0.524 + ], + "angle": 0, + "content": "Figure 1. Rate-distortion curves on the NeRF synthetic dataset. The numbers inside parenthesis denote the axis resolution of grids." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.893, + 0.688 + ], + "angle": 0, + "content": "sparked interest by synthesizing high-quality images from novel viewpoints. It uses a multi-layer perceptron (MLP) with positional encoding to map coordinates to corresponding colors and opacities. Combined with the differentiable volumetric rendering and the neural network's architectural priors, it has shown great potential to be a new representation paradigm. However, the high computational costs (in both training and inference) have been a significant bottleneck, often taking several days to converge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Several follow-up studies have been proposed to accelerate training and inference times [5, 9, 11, 14, 16, 31, 32, 40, 44]. To speed up inference, KiloNeRF [32] proposed splitting a 3D scene into thousands of partial scenes, each of which is assigned a tiny, distinct neural network. While achieving impressive speed-up, it requires a massive amount of memory storage. On the other hand, FastNeRF [14] suggested caching and factorizing the NeRF network to reduce computational costs. Meanwhile, meta-learning algorithms [12, 26, 29] have also been applied to accelerate the training time, and they have shown faster convergence with the learned initialization [34, 40]. However, it requires well-organized large-scale datasets and a pre-training process. Furthermore, the rendering costs re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.204, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.225, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.225, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20680" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "main unchanged since they use the same network architecture [40], and meta-learning algorithms often suffer from poor out-of-distribution generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.305 + ], + "angle": 0, + "content": "Alternatively, there has been a surge of recent interest in incorporating classical data structures, such as grids or trees, into the NeRF framework [13, 22, 24, 33, 38, 39, 49]. Incorporating additional data structures has significantly reduced training and inference time (from days to a few minutes) without compromising reconstruction quality. However, the overall size dramatically increases due to these dense and volumetric structures. In order to reduce the spatial complexity, several methods have been proposed, including pruning areas [33, 37, 49], encodings [24, 38], and tensor decomposition [3, 4, 17]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.47, + 0.518 + ], + "angle": 0, + "content": "This paper aims to further improve the spatial complexity while maintaining the rendering quality. Leveraging the decades of research on standard compression algorithms [35,36,46], we propose compressing grid-based neural fields using frequency-based transformations. In the frequency domain, a large portion of the coefficients can be discarded without considerably degrading the reconstruction quality, and most standard compression algorithms have exploited frequency domain representations. Thus, we propose using this property on grid-based neural fields to maximize parameter efficiency. Among other alternatives, we employ the discrete wavelet transform (DWT) due to its compactness and ability to efficiently capture both global and local information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.762 + ], + "angle": 0, + "content": "Once we obtain sparse representations via frequency domain representations, we can take advantage of the existing compression techniques. Unlike conventional media data (e.g., image and audio), however, no off-the-shelf compression tools exist that we can leverage without complicated engineering efforts. In addition, since NeRF, or neural rendering networks in general, is a relatively new data format, the characteristics or patterns of their coefficients have not been thoroughly investigated. Thus, we present a compression pipeline for our purposes. To automatically filter out unnecessary coefficients, we propose a trainable binary mask. For each 3D scene, we jointly optimize grid parameters and their corresponding masks. This perscene optimization strategy can be more optimal than the global quantization table used in standard image compression codecs [43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.869 + ], + "angle": 0, + "content": "To compress sparse grid representations, we first merge masks with wavelet coefficients to zero out coefficients and then apply standard compression algorithms to masked coefficients. We utilize the run-length encoding to encode binary information about which coefficients are non-zero. For further compression, we apply one of the entropy coding algorithms, Huffman encoding [19], to these encoded outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our method incurs negligible computational costs at test time, requiring only one inverse DWT (IDWT) per grid. Af-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "ter the IDWT, the wavelet grids are transformed into spatial grids, eliminating the need for additional IDWT during rendering. As a result, the computing time and costs (including memory costs) are identical to the original spatial grid-based NeRF." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.168, + 0.82, + 0.183 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.194, + 0.892, + 0.269 + ], + "angle": 0, + "content": "- We propose using wavelet coefficients to improve parameter sparsity and reconstruction quality. Through experiments, we show that the wavelet coefficients can be more compact than the spatial domain coefficients in neural radiance fields." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.282, + 0.892, + 0.357 + ], + "angle": 0, + "content": "- We propose a trainable mask that can be applied to any grid-based neural representation. Experimental results demonstrate that our proposed masking method can zero out more than \\(95\\%\\) of the total grid parameters while maintaining high reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.369, + 0.892, + 0.399 + ], + "angle": 0, + "content": "- We achieve state-of-the-art performance in novel view synthesis under a memory budget of 2 MB." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.194, + 0.892, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.414, + 0.65, + 0.43 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.439, + 0.671, + 0.456 + ], + "angle": 0, + "content": "2.1. Neural rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Neural radiance fields. NeRF [23] has demonstrated that 3D scenes and objects can be successfully represented using coordinate-based neural representation. NeRF renders a scene by casting a ray per pixel and sampling colors and opacity from points that lie on the ray. However, since this approach relies on the dense, point-wise sampling of color values and opacity, the computational costs are significant." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Due to this inefficiency, there has been considerable interest in improving computational efficiency. A plethora of methods propose to adopt data structures such as trees [45, 49], point clouds [27, 28, 48, 50], and grids [4, 13, 15, 24, 33, 37, 38] to efficiently represent 3D scenes and objects. PlenOctrees [49] use an octree structure for real-time rendering. Despite the reduced rendering costs, the increased size of more than ten times the size of standard MLP-based neural fields, such as NeRF [23], limits their wide application. Point cloud-based approaches [27, 48] reduce the number of points sampled per ray by utilizing the scene geometry of a point cloud. Lastly, a number of studies have demonstrated the efficiency of using grid structures, including vectors, matrices, and tensors, on neural fields [4, 13, 24, 37]. With the use of these grid structures, training iterations that once required tens of hours can now be completed in a matter of minutes. Furthermore, real-time rendering at inference time is also possible. However, dense 3D grid structures require substantial amounts of memory, often exceeding 1 GB. To alleviate this memory requirement, several methods have been proposed to reduce the size of grids [4, 24, 38, 42]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20681" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.243 + ], + "angle": 0, + "content": "To achieve near-instant rendering with reasonable memory requirements, Instant-NGP [24] proposed using hash-based multi-resolution grids. Using hash functions, each grid maps input coordinates to corresponding feature vectors. Similarly, VQAD [38] proposed using codebooks and vector quantization rather than the hash function. This enables control over the overall size of the neural fields. However, during training, a sizable amount of memory is required to learn which code from the codebook to use for each point on the grid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.471, + 0.456 + ], + "angle": 0, + "content": "In order to alleviate the memory requirement of using 3D grids, another line of work decomposes 3D grids into lower dimensional representations, such as planes and vectors [3, 4, 17, 42]. EG3D [3] proposes a tri-plane approach, which represents a 3D scene with three perpendicular planes and uses feature vectors, extracted separately from each plane, as inputs for the following MLPs. TensoRF [4] is another parameter-efficient yet expressive method for decomposing dense three-dimensional grids into a smaller set of parameters, such as planes and lines. Although these methods dramatically reduce the time and space complexity of the 3D scene and object representation, their overall sizes are larger than those of MLP-only methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Frequency-based representation. Several studies have explored frequency-based parameterization for efficient neural field representations [17, 41, 45, 47]. Fourier PlenOctree [45] has demonstrated that the Fourier transform can improve both parameter efficiency and training speed of the PlenOctree structure [49]. Nevertheless, the overall size is larger than MLP-only methods. PREF [17] has also exploited the Fourier transform in 3D scene representation. It applies the Fourier transform to a tensor decomposition-based representation. However, it has not shown notable improvements in both the representation quality and the parameter efficiency compared to spatial grid representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Besides the Fourier transform, there is another line of signal representations in the frequency domain: the wavelet transform. The wavelet transform decomposes a signal using a set of basis functions called wavelets. The wavelet transform can extract both frequency and temporal information, in contrast to the Fourier transform, which can only capture frequency information. Because each coefficient covers a different frequency and period in time, it is known that the wavelet transform can represent transient signals more compactly than Fourier-based transforms. This compactness contributes to standard compression CODECs, such as JPEG 2000 [35]. Motivated by the success of the standard CODECs, we aim to mimic their compression pipeline and apply it to the neural radiance fields. One disadvantage of the wavelet transform is that it can be less efficient than Fourier-based transforms for smooth signals. However, we believe that most situations in grid-based neural fields" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "are not the case, considering that each grid's resolution is constrained to accurately depict detailed scenes and objects with limited memory budgets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.157, + 0.687, + 0.172 + ], + "angle": 0, + "content": "2.2. Model compression" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.183, + 0.892, + 0.228 + ], + "angle": 0, + "content": "In recent years, many works have studied to downsize neural networks through various techniques such as weight pruning and quantization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Pruning. For efficient representations, pruning methods have been explored in neural fields. Point-NeRF [48] proposes a pruning strategy for point cloud representation by imposing a sparsity loss on point confidence. Re:NeRF [8] has validated that simply applying a general pruning technique, which gradually removes coefficients throughout the course of training, to grid-based neural representations results in a significant performance drop. It proposed iterative parameter removal and conditional parameter inclusion and demonstrated comparable compression performance. However, it has not shown desirable performance under limited memory budgets. Similarly, KiloNeRF [32] prunes out parameters for unused areas using a fixed threshold. Instead, we propose a trainable mask that can remove a large number of coefficients while causing only minor degradation in the reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Quantization. In terms of neural network quantization, there are two main approaches: post-quantization (PTQ) [21, 25] and quantization-aware training (QAT) [7, 18, 30, 51]. Post-quantization methods quantify network parameters after training is finished. PTQ has the advantage that it does not require additional training iterations or datasets. Furthermore, it supports arbitrary-bit quantization. Performance, however, might degrade because the network parameters were not optimized for the target bit precision. Unlike post-quantization, QAT optimizes neural networks for a certain bit precision during training iterations. Since quantization is not differentiable, it usually relies on the straight-through estimator (STE) [1] for network optimization [7]. QAT has been validated for its effectiveness in neural representation models. However, PTQ has the disadvantage of requiring additional training samples and iterations when quantizing the weights of a pretrained network. Since we train neural fields from scratch, we use PTQ for weight quantization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.599, + 0.802 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.707, + 0.829 + ], + "angle": 0, + "content": "3.1. Neural radiance fields" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We consider a neural radiance field leveraging grid representations. It takes an input coordinate \\(x \\in \\mathbb{R}^3\\) and viewing direction, \\(d \\in \\mathbb{R}^2\\), generating a four-dimensional vector consisting of a density and three-channel RGB colors, de" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20682" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.094, + 0.895, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.209, + 0.895, + 0.28 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our model. Images with purple borders illustrate wavelet coefficients, while those with orange borders visualize spatial coefficients. The wavelet coefficients in each 2D grid are multiplied with a binarized mask to form a masked wavelet coefficient grid. Masked wavelet coefficients are then inverse-transformed to spatial features. We sample feature vectors for input coordinates using bilinear interpolation on the grids. Then, opacity and color for each input coordinate are estimated using the sampled feature vectors following the defined model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.305, + 0.191, + 0.319 + ], + "angle": 0, + "content": "fined as follows." + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.33, + 0.47, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\sigma (x) = f _ {\\sigma} \\left(x; \\gamma_ {\\sigma}, \\mathcal {M} _ {\\sigma}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.349, + 0.469, + 0.365 + ], + "angle": 0, + "content": "\\[\nc (x, d) = f _ {c} (x, d; \\theta , \\gamma_ {c}, \\mathcal {M} _ {c}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.373, + 0.47, + 0.449 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the parameter of an MLP, \\(\\gamma = \\{\\gamma_{\\sigma}, \\gamma_{c}\\}\\) is a set of grid parameters, and \\(\\mathcal{M} = \\{\\mathcal{M}_{\\sigma}, \\mathcal{M}_{c}\\}\\) is a set of masks for grid parameters, which will be described shortly (Sec. 3.3). The following volumetric rendering equation is used to synthesize novel views:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.457, + 0.469, + 0.493 + ], + "angle": 0, + "content": "\\[\nC (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (r (t)) c (r (t), d) d t, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.495, + 0.469, + 0.53 + ], + "angle": 0, + "content": "\\[\nT (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (r (s)) d s\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.598 + ], + "angle": 0, + "content": "where \\( r(t) \\) is a ray from a camera viewpoint, and \\( C(r) \\) is the expected color of the ray \\( r(t) \\). Two integral bounds (near and far) are denoted as \\( t_n \\) and \\( t_f \\), respectively. Please refer to NeRF [23] for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.607, + 0.438, + 0.623 + ], + "angle": 0, + "content": "3.2. Wavelet transform on decomposed tensors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Frequency-based algorithms, such as discrete cosine transform (DCT) or discrete wavelet transform (DWT), have been developed and improved over the decades. For high compression performance, standard CODEs have used them to make sparse representations. However, in the case of 3D grid representation, their computational complexity increases cubically with grid resolution, making high-resolution 3D scene and object representations impractical. In order to use frequency-based algorithms for compact representation, we need lower-dimensional data, such as 2D planes or 1D lines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Recent studies have explored various decomposition methods to reduce the spatial complexity of 3D grid-based representations. Among them, plane-based representations have achieved remarkable success in reducing the number of parameters while maintaining the rendering performance [3, 4, 17]. They propose decomposing a 3D tensor into a set of 2D planes or 1D vectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.892, + 0.365 + ], + "angle": 0, + "content": "To combine the best of both worlds, we propose using the wavelet transform on 2D plane-based neural fields [4]. We use the wavelet transform because of its compactness, especially for non-repeating, non-smooth signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.893, + 0.606 + ], + "angle": 0, + "content": "For efficient 3D object and scene representation, recent studies proposed using a lower-dimensional grid (2D planes or 1D lines) [3, 4]. For example, EG3D [3] utilizes three 2D planes (tri-plane) and TensoRF [4] employs three sets, each consisting of a plane and a vector. We describe our method based on TensoRF. In detail, we use a set of 2D matrices and 1D vectors for grid representation, \\(\\gamma_{\\sigma} = \\{\\mathcal{W}_{\\sigma,r}^{x}, \\mathcal{W}_{\\sigma,r}^{y}, \\mathcal{W}_{\\sigma,r}^{z}, v_{\\sigma,r}^{x}, v_{\\sigma,r}^{y}, v_{\\sigma,r}^{z}\\}_{r=1}^{N_{\\sigma,r}}\\) (\\(\\sigma\\) denotes density, and we will omit the subscript \\(\\sigma\\) for brevity from now on) as proposed in TensoRF [4]. \\(N_{r}\\) is the number of ranks in matrix-vector decomposition and \\(\\mathcal{W}_r^x \\in \\mathbb{R}^{H \\times W}\\), \\(\\mathcal{W}_r^y \\in \\mathbb{R}^{W \\times D}\\), \\(\\mathcal{W}_r^z \\in \\mathbb{R}^{H \\times D}\\) are matrices, \\(v_r^x \\in \\mathbb{R}^D\\), \\(v_r^y \\in \\mathbb{R}^H\\), \\(v_r^z \\in \\mathbb{R}^W\\), are vectors in \\(x, y, z\\) directions, respectively. \\(H, W, D\\) are the resolution of the grid. More formally, a 3D grid representation \\(G\\) can be defined as follows." + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.606, + 0.891, + 0.649 + ], + "angle": 0, + "content": "\\[\nG = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} v _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\mathcal {W} _ {r} ^ {d}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.891, + 0.681 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes the outer product and \\(\\mathrm{idwt}(\\cdot)\\) is a two-dimensional inverse discrete wavelet transform (IDWT)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.712 + ], + "angle": 0, + "content": "With a slight abuse of notation, a single level IDWT of a matrix \\(\\mathcal{W} \\in \\mathcal{R}^{m \\times n}\\) can be written as follows." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.716, + 0.891, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {i d w t} (\\mathcal {W}) = \\Phi \\star (\\uparrow 2) \\left(\\mathcal {W} _ {L L}\\right) + \\Psi_ {H L} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {H L}\\right) \\\\ + \\Psi_ {L H} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {L H}\\right) + \\Psi_ {H H} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {H H}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}_{LL},\\mathcal{W}_{HL},\\mathcal{W}_{LH},\\mathcal{W}_{HH}\\in \\mathbb{R}^{m / 2\\times n / 2}\\) denote approximation, horizontal, vertical, and diagonal coefficients, respectively. \\(\\star\\) denotes 2D cross correlation and \\((\\uparrow 2):\\) \\(\\mathbb{R}^{m / 2\\times n / 2}\\to \\mathbb{R}^{m\\times n}\\) denotes (nearest neighbor) upscaling. \\(\\Psi\\) and \\(\\Phi\\) are wavelet function and scaling function, respectively. We use the biorthogonal 4.4 wavelet function and its corresponding scaling function [6]. Even though we follow the grid representation of TensoRF, our method is not limited to the existing 2D plane-based representations but can" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "be applied to any other approaches that adopt 2D planes to represent higher-dimensional structures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.469, + 0.228 + ], + "angle": 0, + "content": "As shown in Eq. 5, we only apply IDWT over matrices \\(\\mathcal{W}\\) since we observed that IDWT to vectors resulted in significant quality degradation. As this transformation is differentiable, the model can be trained end-to-end. We can also obtain the 3D grid representation for appearance through the same process and provide all formulations in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.469, + 0.304 + ], + "angle": 0, + "content": "Our method incurs insignificant costs during inference. Once training is finished, only one IDWT per grid is needed to transform wavelet grids into spatial grids. Thus, the remaining computational costs and time are exactly the same as the original spatial grid-based neural fields." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.338, + 0.339 + ], + "angle": 0, + "content": "3.2.1 Multi-level wavelet transform" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.59 + ], + "angle": 0, + "content": "To further improve parameter sparsity and reconstruction quality, we propose using a multi-level wavelet transform. We experimentally found that higher-level wavelet transformations result in higher sparsity in grids (Sec. 4.3.1). However, naively using a multi-level wavelet transform on grid parameters degrades the representation quality (Tab. 1). We hypothesize that this is because the range of wavelet coefficients and the range of their gradients vary depending on the level of decomposition. In more detail, high-pass coefficients typically have a larger gradient scale with a smaller weight range. Different weight scales and gradients for different frequencies seem to hinder the optimization process. As such, we propose multiplying wavelet coefficients by the scales proportional to the inverse of frequency. With our proposed scaling factors \\( s \\), Eq. 5 can be rewritten as follows." + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.6, + 0.47, + 0.645 + ], + "angle": 0, + "content": "\\[\nG = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} v _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\mathcal {W} _ {r} ^ {d} \\odot s _ {r} ^ {d}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.761 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the Hadamard product. For example, a two-level DWT, as shown in Fig. 3, generates a total of seven groups of coefficients (LL2, HL2, LH2, HH2, HL1, LH1, HH1). The scaling factors \\(s\\) for HL2, LH2, and HH2 will be set to \\(1/2\\), and for HL1, LH1, and HH1 to \\(1/3\\). This scaling method enhances the quality of the reconstruction; more experiments are described in Sec. 4.3.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.249, + 0.787 + ], + "angle": 0, + "content": "3.3. Learning to mask" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Even though Wavelet coefficients can be sparse, we need additional methods to attain higher sparsity. Thus, we propose using element-wise masks to increase the portion of zero elements in grids. By jointly optimizing binary masks and grid parameters, we aim to zero out the majority of coefficients without significantly degrading the rendering quality. As we defined earlier," + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.089, + 0.696, + 0.124 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.125, + 0.578, + 0.134 + ], + "angle": 0, + "content": "LL2" + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.125, + 0.663, + 0.134 + ], + "angle": 0, + "content": "HL2" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.139, + 0.696, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.174, + 0.578, + 0.183 + ], + "angle": 0, + "content": "HL2" + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.174, + 0.663, + 0.183 + ], + "angle": 0, + "content": "HH2" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.089, + 0.871, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.157, + 0.796, + 0.166 + ], + "angle": 0, + "content": "HL1" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.187, + 0.696, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.256, + 0.619, + 0.265 + ], + "angle": 0, + "content": "HL1" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.187, + 0.871, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.256, + 0.796, + 0.265 + ], + "angle": 0, + "content": "HH1" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.295, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Figure 3. Multi-level discrete wavelet coefficients (2-level). LL, HL, LH, and HH denote approximation, horizontal, vertical, and diagonal coefficients, respectively. The numbers on the right denote the level of decomposition." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.375, + 0.892, + 0.512 + ], + "angle": 0, + "content": "we have a set of trainable element-wise masks, \\(\\mathcal{M} = \\{\\mathcal{M}_r^x,\\mathcal{M}_r^y,\\mathcal{M}_r^z,m_r^x,m_r^y,m_r^z\\}_{r = 1}^{N_r}\\) (note that we omitted the subscript \\(\\sigma\\) for brevity). During training, the grid parameters and corresponding binarized masks are multiplied element by element. Since calculating gradients directly from binarized masks is not feasible, we used the straight-through-estimator technique [1] to train and use masks. Formally, the masked matrix parameters \\(\\widehat{\\mathcal{W}}_r\\) for \\(\\mathcal{W}_r\\) can be expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.519, + 0.892, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {W}} _ {r} = \\operatorname {s g} \\left(\\left(\\mathcal {H} \\left(\\mathcal {M} _ {r}\\right) - \\sigma \\left(\\mathcal {M} _ {r}\\right)\\right) \\odot \\mathcal {W} _ {r}\\right) + \\sigma \\left(\\mathcal {M} _ {r}\\right) \\odot \\mathcal {W} _ {r}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.563, + 0.892, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathfrak{sg}(\\cdot)\\) is the stop-gradient operator. \\(\\sigma (\\cdot)\\) and \\(\\mathcal{H}(\\cdot)\\) denote the element-wise sigmoid and Heaviside step function, respectively. We can also compute the masked vector parameters \\(\\widehat{v}_r\\) similarly. By replacing grid parameters \\((\\mathcal{W}_r,\\) \\(v_{r})\\) from Eq. (7) with masked grid parameters \\((\\widehat{\\mathcal{W}}_r,\\widehat{v}_r)\\) , we can represent a masked 3D grid representation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.663, + 0.892, + 0.707 + ], + "angle": 0, + "content": "\\[\nG = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} \\widehat {v} _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\widehat {\\mathcal {W}} _ {r} ^ {d} \\odot s _ {r} ^ {d}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.715, + 0.892, + 0.775 + ], + "angle": 0, + "content": "To make binarized mask values sparse, we use the sum of all mask values as the additional loss term \\(\\mathcal{L}_m\\). The overall loss function is a sum of rendering loss \\(\\mathcal{L}_r\\) and the mask regularization term \\(\\mathcal{L}_m\\)." + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.785, + 0.891, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r} + \\lambda_ {m} \\mathcal {L} _ {m}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.891, + 0.84 + ], + "angle": 0, + "content": "We use \\(\\lambda_{m}\\) to control the sparsity of the parameters of the grid representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.704, + 0.864 + ], + "angle": 0, + "content": "3.4. Compression pipeline" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With our proposed masking method and multi-level DWT, the ratio of zeros in grids can go up to more than" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20684" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.242, + 0.472, + 0.286 + ], + "angle": 0, + "content": "Figure 4. Proposed bitmap compression pipeline. RLE and Huffman denote run-length encoding and Huffman encoding, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.477 + ], + "angle": 0, + "content": "95%. However, sparse representations themselves do not reduce the total size. In this section, we describe our proposed compression pipeline for sparse grid parameters. Instead of storing grids as they are, we separately store nonzero coefficients and bitmaps (or masks) that indicate which coefficients are non-zero. Despite using 1-bit bitmaps, the overall bitmap size is large due to the large number of parameters. To reduce the bitmap sizes, we propose a compression pipeline with the following three stages: n-bit casting, run-length encoding (RLE), and the Huffman encoding [19]. The compression pipeline is illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.471, + 0.718 + ], + "angle": 0, + "content": "Before applying the compression pipeline, we split wavelet coefficient masks by level of decomposition. For the 2-level wavelet transform, for instance, we group HL1, LH1, and HH1, then HL2, LH2, and HH2, and finally LL2 (Fig. 3). This is based on our experimental results that wavelet coefficients with the same levels of sparsity have similar levels of sparsity (Fig. 6). We found that grouping coefficients with similar sparsity results in better compression performance. In addition, we found that directly applying RLE to 1-bit streams causes inefficient bit allocation due to the numerous repeating zeros. Therefore, we first cast the binary mask values to n-bit unsigned integers, and perform RLE afterward. In our experiments, we set n to 8. Finally, we apply the Huffman encoding algorithm to the RLE-encoded streams to map values with a high probability to shorter bits." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.21, + 0.75 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.282, + 0.774 + ], + "angle": 0, + "content": "4.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.84 + ], + "angle": 0, + "content": "We compared our proposed method with other baselines in four datasets; NeRF synthetic dataset [23], Neural Sparse Voxel Fields (NSVF) [22] dataset, TanksAndTemps dataset [20], and LLFF dataset [10]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We used TensoRF VM-192 as the baseline model and applied the proposed method. To control the size of our method, we adjusted the mask regularization weight \\(\\lambda_{m}\\) the number of grid channels (from VM-192 to VM-384)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "and the resolution of grids. Unless otherwise specified, we used 8-bit quantization for every method. We compared our method with quantized NeRF, TensoRF models (CP and VM), cNeRF [2] and CCNeRF [42]. We also tried to compare our approach with VQAD [38], which efficiently compresses 3D scene representations using codebooks and vector quantization. However, VQAD requires depth maps to prune out vacant areas in advance. Otherwise, the required memory exceeds the memory capacity of the Tesla A100 equipped with 40 GB of memory. Since the datasets we used do not provide depth maps, we did not compare VQAD with ours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.893, + 0.5 + ], + "angle": 0, + "content": "We followed the experimental settings of TensoRF [4]. We trained models for 30,000 iterations, each of which is a minibatch of 4,096 rays. We used the Adam optimizer and an exponential learning rate decay scheduler. Following TensoRF, the initial learning rates of grid-related parameters and MLP-related parameters were set to 0.02 and 0.001, respectively. Final learning rates were set at 1/10 of the initial learning rate. We updated the alpha masks at the 2000th, 4000th, 6000th, 11000th, 16000th, 21000th, and 26000th iterations. We set \\(\\lambda_{m}\\) in Eq. (10) to 1e-10 for high parameter sparsity and 5e-11 and 2.5e-11 for relatively lower sparsity. We set the initial values of masks \\(\\mathcal{M}\\) to one, and set their initial learning rate to the same learning rate as grid parameters \\(\\phi\\). As a reconstruction quality measurement, we used PSNR." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.593, + 0.527 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Figs. 1 and 5 show the quantitative performances of our method and the baselines on various datasets. Each graph displays the average PSNR and size of methods. More detailed numbers are available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.869 + ], + "angle": 0, + "content": "As the rate-distortion curves show, our method achieves higher efficiency than the baseline models and other methods. Even with the base network (VM-192), our method consistently outperforms other baselines on all novel view synthesis datasets, under a small memory budget of \\(2\\mathrm{MB}\\). We adjusted both \\(\\lambda_{m}\\) and the network size to increase the output size. By adjusting \\(\\text{lambda}_{m}\\), we can control both size and quality while having no effect on computational costs, time, or memory requirements. Increasing network size (doubling the number of channels and increasing the resolution) can push the boundary even further. It is also worth mentioning that applying our method to a larger backbone model (VM-384) outperforms models with a smaller backbone model (VM-192), even with the same value of \\(\\lambda_{m}\\) or a similar model size. However, as the network grows in size, the computational costs and time required for training and testing increase. Therefore, it is still reasonable to use a small backbone model when using our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Fig. 7 shows the qualitative results with and without our proposed masking method. Trainable masks removed more" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20685" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.357, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.091, + 0.613, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.092, + 0.868, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.255, + 0.89, + 0.27 + ], + "angle": 0, + "content": "Figure 5. Rate distortion curves on the NSVF synthetic, TanksAndTemples, and LLFF dataset. We used 8-bit precision for every method." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.294, + 0.256, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.427, + 0.23, + 0.438 + ], + "angle": 0, + "content": "(a) Raw mask" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.293, + 0.451, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.427, + 0.422, + 0.438 + ], + "angle": 0, + "content": "(b) Binarized mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.451, + 0.47, + 0.48 + ], + "angle": 0, + "content": "Figure 6. Qualitative analysis on mask \\(\\mathcal{M}\\) (Chair in the NeRF synthetic dataset)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.595 + ], + "angle": 0, + "content": "than \\(97\\%\\) of the grid parameters; nevertheless, the rendered results are still accurate, and the qualitative difference is almost imperceptible. This demonstrates that our proposed method efficiently eliminates unnecessary wavelet coefficients by leveraging trainable masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.471, + 0.753 + ], + "angle": 0, + "content": "Fig. 6 illustrates both raw and binarized masks after training. First of all, sparsity varies depending on the level of the wavelet transform. As shown in Fig. 6b, coefficients with lower frequencies have lower sparsity, whereas coefficients with higher frequencies have higher sparsity. What is interesting is that the raw mask values seem to reflect the characteristics of corresponding wavelet coefficients. Vertical, horizontal, and diagonal patterns can be found in the raw mask values for the vertical, horizontal, and diagonal coefficients, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.781, + 0.238, + 0.795 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, we analyze each component of our proposed method. Every experiment was conducted on the whole NeRF synthetic dataset, and we used the average PSNR as a measurement for representation quality. We used 4-level wavelet transform and set \\(\\lambda_{m}\\) to 1e-10, unless otherwise specified." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.292, + 0.835, + 0.361 + ], + "angle": 0, + "content": "
MethodsPSNR
DWT31.949
DWT w/o level-wise scaling31.145
DCT31.325
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Table 1. Ablation study on level-wise scaling. Without our proposed scaling method, the performance of DWT is worse than DCT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.742, + 0.454 + ], + "angle": 0, + "content": "4.3.1 Discrete wavelet transform" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.617 + ], + "angle": 0, + "content": "In this section, we evaluate the compactness of wavelet coefficients and the performance improvements brought by the multi-level wavelet transform. Fig. 8 shows the rate-distortion curves of four different levels of wavelet transform and non-wavelet coefficients. As shown in the figure, wavelet-based grid representation outperforms the grid in the spatial domain, especially when the sparsity is high. Furthermore, rate-distortion curves demonstrate that higher level of wavelet transform improves visual quality and sparsity." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.812, + 0.652 + ], + "angle": 0, + "content": "4.3.2 Level-wise wavelet coefficient scaling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.766 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.2.1, we proposed multiplying wavelet coefficients by the scales proportional to the inverse of frequency. Thus, we compared two versions of wavelet transform; one with our proposed weight scaling (Sec. 3.2.1) and the other without it. As Tab. 1 shows, scaling wavelet coefficients improves reconstruction performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.68, + 0.8 + ], + "angle": 0, + "content": "4.3.3 Wavelet functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this section, we analyze how different wavelet functions affect reconstruction quality. For comparison, we selected several wavelet functions: Haar, Coiflet 1, Daubechies 4, and reverse biorthogonal 4.4. As shown in Tab. 2, the type of wavelet function has little effect on reconstruction quality. Even Haar, the simplest wavelet function, performs" + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.089, + 0.51, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.092, + 0.825, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.893, + 0.286 + ], + "angle": 0, + "content": "Figure 7. A qualitative analysis of grid sparsity. The sparsity refers to the ratio of zero elements in grids. With the help of trainable masks, removing more than \\(97\\%\\) of grid parameters only results in an imperceptible difference." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.317, + 0.462, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.596 + ], + "angle": 0, + "content": "Figure 8. The rate distortion curves of different signal representation schemes (spatial, DWT, and DCT). Sparsity on the \\( x \\) axis refers to the ratio of zeros in grid parameters. The grid sparsity was controlled by \\( \\lambda_{m} \\). The numbers inside the parenthesis indicate the levels of the wavelet transform." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.611, + 0.409, + 0.711 + ], + "angle": 0, + "content": "
Wavelet FunctionPSNR
Haar31.889
Coiflets 131.846
Daubechies 431.734
biorthogonal 4.4 (default)31.949
reverse biorthogonal 4.431.727
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.72, + 0.47, + 0.748 + ], + "angle": 0, + "content": "Table 2. Reconstruction performances of different wavelet functions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.777, + 0.47, + 0.809 + ], + "angle": 0, + "content": "fairly well. Still, the biorthogonal 4.4 that we used shows the best performance among selected wavelet functions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.31, + 0.845 + ], + "angle": 0, + "content": "4.3.4 Discrete cosine transform" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this section, we compare the performance of DWT with that of DCT, which represents a signal with respect to the sum of cosine functions with different frequencies. DCT," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.311, + 0.893, + 0.448 + ], + "angle": 0, + "content": "similar to DWT, has an energy compaction property. That is, a small number of non-zero DCT coefficients are sufficient to represent a given signal. Because of its energy compactness, DCT or its variants are another widely used signal transformation for image, audio, and audio compression. However, as shown in Fig. 8 and Tab. 1, DWT shows superior compression and representation performance to DCT. We believe this is due to the non-repeating, non-smooth information in grids, for which DWT is more appropriate." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.464, + 0.619, + 0.479 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.893, + 0.64 + ], + "angle": 0, + "content": "We propose a compact representation for grid-based neural fields, enabled by a novel masking strategy and multilevel wavelet transform. We demonstrate that these two components made it possible to remove more than \\(95\\%\\) of grid parameters without a significant loss of visual quality. With our proposed compression pipeline and the sparse wavelet coefficients, we achieved state-of-the-art performance under a memory budget of 2 MB. We believe we can reduce the size further with more sophisticated compression pipelines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.894, + 0.748 + ], + "angle": 0, + "content": "Our proposed method is naturally constrained by the limitations of grid-based neural fields that are developed for bounded scenes or objects. We believe that expanding this grid-based representation to encompass unbounded or large scenes would be an intriguing direction, given that we can now compactly represent 3D objects with our proposed representation scheme." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.764, + 0.661, + 0.781 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the Institute of Information and Communications Technology Planning and Evaluation (IITP) grants (2021-0-02068, IITP-2021-0-02052, IITP-2023-2020-0-01821, IITP-2019-0-00421) and National Research Foundation of Korea (NRF) grants (2022R1A4A3032913, 2022R1F1A1064184, 2022R1A4A3033571) funded by the Korea Government (MSIT)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20687" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.229 + ], + "angle": 0, + "content": "[2] Thomas Bird, Johannes Balle, Saurabh Singh, and Philip A Chou. 3d scene compression through entropy penalized neural representation functions. In 2021 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[3] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, June 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision. Springer, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Albert Cohen, Ingrid Daubechies, and J-C Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on pure and applied mathematics, 45(5):485-560, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[8] Chenxi Lola Deng and Enzo Tartaglione. Compressing explicit voxel grid representations: fast nerfs become also small. arXiv preprint arXiv:2210.12782, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.716, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[10] Ben Mildenhall et al. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.76, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, SA '22, New York, NY, USA, 2022. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning," + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "volume 70 of Proceedings of Machine Learning Research, pages 1126-1135. PMLR, 06-11 Aug 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[13] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[14] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14346-14355, October 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[15] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. pages 5875-5884, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[16] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12902-12911, June 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[17] Binbin Huang, Xinhao Yan, Anpei Chen, Shenghua Gao, and Jingyi Yu. Pref: Phasorial embedding fields for compact neural representations. arXiv preprint arXiv:2205.13524, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[18] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Binarized neural networks. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[19] David A Huffman. A method for the construction of minimum-redundancy codes. Proceedings of the IRE, 40(9):1098-1101, 1952. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), jul 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[21] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[22] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 15651-15663. Curran Associates, Inc., 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), jul 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[25] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? Adaptive" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20688" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "rounding for post-training quantization. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7197-7206. PMLR, 13-18 Jul 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[26] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[27] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. pages 18419-18429, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[28] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 13032-13044. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[29] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[30] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision – ECCV 2016, pages 525–542, Cham, 2016. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[31] Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, and Andrea Tagliasacchi. Derf: Decomposed radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14153-14161, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.621 + ], + "angle": 0, + "content": "[32] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14335-14345, October 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[33] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. VoxGRAF: Fast 3d-aware image synthesis with sparse voxel grids. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[34] Vincent Sitzmann, Eric Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 10136-10147. Curran Associates, Inc., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[35] Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The jpeg 2000 still image compression standard. IEEE Signal processing magazine, 18(5):36-58, 2001. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[36] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[38] Towaki Takikawa, Alex Evans, Jonathan Tremblay, Thomas Müller, Morgan McGuire, Alec Jacobson, and Sanja Fidler. Variable bitrate neural fields. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[39] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. arXiv preprint arXiv:2101.10994, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[40] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P. Srinivasan, Jonathan T. Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[41] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[42] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable neRF via rank-residual decomposition. In Advances in Neural Information Processing Systems, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[43] Gregory K. Wallace. TheJPEG still picture compression standard.Commun.ACM,34(4):30-44,apr 1991.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[44] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 612-629, Cham, 2022. Springer Nature Switzerland. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[45] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13524-13534, June 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[46] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology, 13(7):560-576, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[47] Zhijie Wu, Yuhe Jin, and Kwang Moo Yi. Neural fourier filter bank. arXiv e-prints, pages arXiv-2212, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[48] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[49] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5752-5761, October 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20689" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[50] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. arXiv preprint arXiv:2205.14330, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[51] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained ternary quantization. In International Conference on Learning Representations, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20690" + } + ] +] \ No newline at end of file diff --git a/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_origin.pdf b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8f3db1c58f20c3d3a858e5784d3fdcc11dd5071b --- /dev/null +++ b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/45d0202b-5376-4366-b728-470ed1f9b25b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ec51c6cd088115eb6125c0e43115995c4bc16be370f2eef319979e3009676c6 +size 4301657 diff --git a/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/full.md b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..12f1c36a59516b1b9f41f8d45a10675219f34931 --- /dev/null +++ b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/full.md @@ -0,0 +1,351 @@ +# Masked Wavelet Representation for Compact Neural Radiance Fields + +Daniel Rho1* + +Byeonghyeon Lee $^{2*}$ + +Seungtae Nam² +Eunbyung Park²,³† +¹AI2XL, KT + +Joo Chan Lee² + +Jong Hwan $\mathrm{Ko}^{2,3\dagger}$ + +$^{2}$ Department of Artificial Intelligence, Sungkyunkwan University + +$^{3}$ Department of Electrical and Computer Engineering, Sungkyunkwan University + +# Abstract + +Neural radiance fields (NeRF) have demonstrated the potential of coordinate-based neural representation (neural fields or implicit neural representation) in neural rendering. However, using a multi-layer perceptron (MLP) to represent a 3D scene or object requires enormous computational resources and time. There have been recent studies on how to reduce these computational inefficiencies by using additional data structures, such as grids or trees. Despite the promising performance, the explicit data structure necessitates a substantial amount of memory. In this work, we present a method to reduce the size without compromising the advantages of having additional data structures. In detail, we propose using the wavelet transform on grid-based neural fields. Grid-based neural fields are for fast convergence, and the wavelet transform, whose efficiency has been demonstrated in high-performance standard codes, is to improve the parameter efficiency of grids. Furthermore, in order to achieve a higher sparsity of grid coefficients while maintaining reconstruction quality, we present a novel trainable masking approach. Experimental results demonstrate that non-spatial grid coefficients, such as wavelet coefficients, are capable of attaining a higher level of sparsity than spatial grid coefficients, resulting in a more compact representation. With our proposed mask and compression pipeline, we achieved state-of-the-art performance within a memory budget of 2 MB. Our code is available at https://github.com/daniel03c1/maskedwavelet_nerf. + +# 1. Introduction + +Recent advances in coordinate-based neural representation (neural fields or implicit neural representation) have demonstrated remarkable performance in many applications. In particular, neural radiance fields (NeRF) have + +![](images/959d47197dd2a619d446a16183c11c28e9004d14b86c2dd6dcac18aea2a9d6f2.jpg) +Figure 1. Rate-distortion curves on the NeRF synthetic dataset. The numbers inside parenthesis denote the axis resolution of grids. + +sparked interest by synthesizing high-quality images from novel viewpoints. It uses a multi-layer perceptron (MLP) with positional encoding to map coordinates to corresponding colors and opacities. Combined with the differentiable volumetric rendering and the neural network's architectural priors, it has shown great potential to be a new representation paradigm. However, the high computational costs (in both training and inference) have been a significant bottleneck, often taking several days to converge. + +Several follow-up studies have been proposed to accelerate training and inference times [5, 9, 11, 14, 16, 31, 32, 40, 44]. To speed up inference, KiloNeRF [32] proposed splitting a 3D scene into thousands of partial scenes, each of which is assigned a tiny, distinct neural network. While achieving impressive speed-up, it requires a massive amount of memory storage. On the other hand, FastNeRF [14] suggested caching and factorizing the NeRF network to reduce computational costs. Meanwhile, meta-learning algorithms [12, 26, 29] have also been applied to accelerate the training time, and they have shown faster convergence with the learned initialization [34, 40]. However, it requires well-organized large-scale datasets and a pre-training process. Furthermore, the rendering costs re + +main unchanged since they use the same network architecture [40], and meta-learning algorithms often suffer from poor out-of-distribution generalization performance. + +Alternatively, there has been a surge of recent interest in incorporating classical data structures, such as grids or trees, into the NeRF framework [13, 22, 24, 33, 38, 39, 49]. Incorporating additional data structures has significantly reduced training and inference time (from days to a few minutes) without compromising reconstruction quality. However, the overall size dramatically increases due to these dense and volumetric structures. In order to reduce the spatial complexity, several methods have been proposed, including pruning areas [33, 37, 49], encodings [24, 38], and tensor decomposition [3, 4, 17]. + +This paper aims to further improve the spatial complexity while maintaining the rendering quality. Leveraging the decades of research on standard compression algorithms [35,36,46], we propose compressing grid-based neural fields using frequency-based transformations. In the frequency domain, a large portion of the coefficients can be discarded without considerably degrading the reconstruction quality, and most standard compression algorithms have exploited frequency domain representations. Thus, we propose using this property on grid-based neural fields to maximize parameter efficiency. Among other alternatives, we employ the discrete wavelet transform (DWT) due to its compactness and ability to efficiently capture both global and local information. + +Once we obtain sparse representations via frequency domain representations, we can take advantage of the existing compression techniques. Unlike conventional media data (e.g., image and audio), however, no off-the-shelf compression tools exist that we can leverage without complicated engineering efforts. In addition, since NeRF, or neural rendering networks in general, is a relatively new data format, the characteristics or patterns of their coefficients have not been thoroughly investigated. Thus, we present a compression pipeline for our purposes. To automatically filter out unnecessary coefficients, we propose a trainable binary mask. For each 3D scene, we jointly optimize grid parameters and their corresponding masks. This perscene optimization strategy can be more optimal than the global quantization table used in standard image compression codecs [43]. + +To compress sparse grid representations, we first merge masks with wavelet coefficients to zero out coefficients and then apply standard compression algorithms to masked coefficients. We utilize the run-length encoding to encode binary information about which coefficients are non-zero. For further compression, we apply one of the entropy coding algorithms, Huffman encoding [19], to these encoded outputs. + +Our method incurs negligible computational costs at test time, requiring only one inverse DWT (IDWT) per grid. Af- + +ter the IDWT, the wavelet grids are transformed into spatial grids, eliminating the need for additional IDWT during rendering. As a result, the computing time and costs (including memory costs) are identical to the original spatial grid-based NeRF. + +In summary, our contributions are as follows: + +- We propose using wavelet coefficients to improve parameter sparsity and reconstruction quality. Through experiments, we show that the wavelet coefficients can be more compact than the spatial domain coefficients in neural radiance fields. +- We propose a trainable mask that can be applied to any grid-based neural representation. Experimental results demonstrate that our proposed masking method can zero out more than $95\%$ of the total grid parameters while maintaining high reconstruction quality. +- We achieve state-of-the-art performance in novel view synthesis under a memory budget of 2 MB. + +# 2. Related Works + +# 2.1. Neural rendering + +Neural radiance fields. NeRF [23] has demonstrated that 3D scenes and objects can be successfully represented using coordinate-based neural representation. NeRF renders a scene by casting a ray per pixel and sampling colors and opacity from points that lie on the ray. However, since this approach relies on the dense, point-wise sampling of color values and opacity, the computational costs are significant. + +Due to this inefficiency, there has been considerable interest in improving computational efficiency. A plethora of methods propose to adopt data structures such as trees [45, 49], point clouds [27, 28, 48, 50], and grids [4, 13, 15, 24, 33, 37, 38] to efficiently represent 3D scenes and objects. PlenOctrees [49] use an octree structure for real-time rendering. Despite the reduced rendering costs, the increased size of more than ten times the size of standard MLP-based neural fields, such as NeRF [23], limits their wide application. Point cloud-based approaches [27, 48] reduce the number of points sampled per ray by utilizing the scene geometry of a point cloud. Lastly, a number of studies have demonstrated the efficiency of using grid structures, including vectors, matrices, and tensors, on neural fields [4, 13, 24, 37]. With the use of these grid structures, training iterations that once required tens of hours can now be completed in a matter of minutes. Furthermore, real-time rendering at inference time is also possible. However, dense 3D grid structures require substantial amounts of memory, often exceeding 1 GB. To alleviate this memory requirement, several methods have been proposed to reduce the size of grids [4, 24, 38, 42]. + +To achieve near-instant rendering with reasonable memory requirements, Instant-NGP [24] proposed using hash-based multi-resolution grids. Using hash functions, each grid maps input coordinates to corresponding feature vectors. Similarly, VQAD [38] proposed using codebooks and vector quantization rather than the hash function. This enables control over the overall size of the neural fields. However, during training, a sizable amount of memory is required to learn which code from the codebook to use for each point on the grid. + +In order to alleviate the memory requirement of using 3D grids, another line of work decomposes 3D grids into lower dimensional representations, such as planes and vectors [3, 4, 17, 42]. EG3D [3] proposes a tri-plane approach, which represents a 3D scene with three perpendicular planes and uses feature vectors, extracted separately from each plane, as inputs for the following MLPs. TensoRF [4] is another parameter-efficient yet expressive method for decomposing dense three-dimensional grids into a smaller set of parameters, such as planes and lines. Although these methods dramatically reduce the time and space complexity of the 3D scene and object representation, their overall sizes are larger than those of MLP-only methods. + +Frequency-based representation. Several studies have explored frequency-based parameterization for efficient neural field representations [17, 41, 45, 47]. Fourier PlenOctree [45] has demonstrated that the Fourier transform can improve both parameter efficiency and training speed of the PlenOctree structure [49]. Nevertheless, the overall size is larger than MLP-only methods. PREF [17] has also exploited the Fourier transform in 3D scene representation. It applies the Fourier transform to a tensor decomposition-based representation. However, it has not shown notable improvements in both the representation quality and the parameter efficiency compared to spatial grid representations. + +Besides the Fourier transform, there is another line of signal representations in the frequency domain: the wavelet transform. The wavelet transform decomposes a signal using a set of basis functions called wavelets. The wavelet transform can extract both frequency and temporal information, in contrast to the Fourier transform, which can only capture frequency information. Because each coefficient covers a different frequency and period in time, it is known that the wavelet transform can represent transient signals more compactly than Fourier-based transforms. This compactness contributes to standard compression CODECs, such as JPEG 2000 [35]. Motivated by the success of the standard CODECs, we aim to mimic their compression pipeline and apply it to the neural radiance fields. One disadvantage of the wavelet transform is that it can be less efficient than Fourier-based transforms for smooth signals. However, we believe that most situations in grid-based neural fields + +are not the case, considering that each grid's resolution is constrained to accurately depict detailed scenes and objects with limited memory budgets. + +# 2.2. Model compression + +In recent years, many works have studied to downsize neural networks through various techniques such as weight pruning and quantization. + +Pruning. For efficient representations, pruning methods have been explored in neural fields. Point-NeRF [48] proposes a pruning strategy for point cloud representation by imposing a sparsity loss on point confidence. Re:NeRF [8] has validated that simply applying a general pruning technique, which gradually removes coefficients throughout the course of training, to grid-based neural representations results in a significant performance drop. It proposed iterative parameter removal and conditional parameter inclusion and demonstrated comparable compression performance. However, it has not shown desirable performance under limited memory budgets. Similarly, KiloNeRF [32] prunes out parameters for unused areas using a fixed threshold. Instead, we propose a trainable mask that can remove a large number of coefficients while causing only minor degradation in the reconstruction quality. + +Quantization. In terms of neural network quantization, there are two main approaches: post-quantization (PTQ) [21, 25] and quantization-aware training (QAT) [7, 18, 30, 51]. Post-quantization methods quantify network parameters after training is finished. PTQ has the advantage that it does not require additional training iterations or datasets. Furthermore, it supports arbitrary-bit quantization. Performance, however, might degrade because the network parameters were not optimized for the target bit precision. Unlike post-quantization, QAT optimizes neural networks for a certain bit precision during training iterations. Since quantization is not differentiable, it usually relies on the straight-through estimator (STE) [1] for network optimization [7]. QAT has been validated for its effectiveness in neural representation models. However, PTQ has the disadvantage of requiring additional training samples and iterations when quantizing the weights of a pretrained network. Since we train neural fields from scratch, we use PTQ for weight quantization. + +# 3. Methods + +# 3.1. Neural radiance fields + +We consider a neural radiance field leveraging grid representations. It takes an input coordinate $x \in \mathbb{R}^3$ and viewing direction, $d \in \mathbb{R}^2$ , generating a four-dimensional vector consisting of a density and three-channel RGB colors, de + +![](images/7f76d1a1d654f1e03ca4723bbfb1f3ebf321d4586161904668b80bb5ce1ad621.jpg) +Figure 2. The overall architecture of our model. Images with purple borders illustrate wavelet coefficients, while those with orange borders visualize spatial coefficients. The wavelet coefficients in each 2D grid are multiplied with a binarized mask to form a masked wavelet coefficient grid. Masked wavelet coefficients are then inverse-transformed to spatial features. We sample feature vectors for input coordinates using bilinear interpolation on the grids. Then, opacity and color for each input coordinate are estimated using the sampled feature vectors following the defined model. + +fined as follows. + +$$ +\sigma (x) = f _ {\sigma} \left(x; \gamma_ {\sigma}, \mathcal {M} _ {\sigma}\right), \tag {1} +$$ + +$$ +c (x, d) = f _ {c} (x, d; \theta , \gamma_ {c}, \mathcal {M} _ {c}), \tag {2} +$$ + +where $\theta$ is the parameter of an MLP, $\gamma = \{\gamma_{\sigma}, \gamma_{c}\}$ is a set of grid parameters, and $\mathcal{M} = \{\mathcal{M}_{\sigma}, \mathcal{M}_{c}\}$ is a set of masks for grid parameters, which will be described shortly (Sec. 3.3). The following volumetric rendering equation is used to synthesize novel views: + +$$ +C (r) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (r (t)) c (r (t), d) d t, \tag {3} +$$ + +$$ +T (t) = \exp \left(- \int_ {t _ {n}} ^ {t} \sigma (r (s)) d s\right), \tag {4} +$$ + +where $r(t)$ is a ray from a camera viewpoint, and $C(r)$ is the expected color of the ray $r(t)$ . Two integral bounds (near and far) are denoted as $t_n$ and $t_f$ , respectively. Please refer to NeRF [23] for more details. + +# 3.2. Wavelet transform on decomposed tensors + +Frequency-based algorithms, such as discrete cosine transform (DCT) or discrete wavelet transform (DWT), have been developed and improved over the decades. For high compression performance, standard CODEs have used them to make sparse representations. However, in the case of 3D grid representation, their computational complexity increases cubically with grid resolution, making high-resolution 3D scene and object representations impractical. In order to use frequency-based algorithms for compact representation, we need lower-dimensional data, such as 2D planes or 1D lines. + +Recent studies have explored various decomposition methods to reduce the spatial complexity of 3D grid-based representations. Among them, plane-based representations have achieved remarkable success in reducing the number of parameters while maintaining the rendering performance [3, 4, 17]. They propose decomposing a 3D tensor into a set of 2D planes or 1D vectors. + +To combine the best of both worlds, we propose using the wavelet transform on 2D plane-based neural fields [4]. We use the wavelet transform because of its compactness, especially for non-repeating, non-smooth signals. + +For efficient 3D object and scene representation, recent studies proposed using a lower-dimensional grid (2D planes or 1D lines) [3, 4]. For example, EG3D [3] utilizes three 2D planes (tri-plane) and TensoRF [4] employs three sets, each consisting of a plane and a vector. We describe our method based on TensoRF. In detail, we use a set of 2D matrices and 1D vectors for grid representation, $\gamma_{\sigma} = \{\mathcal{W}_{\sigma,r}^{x}, \mathcal{W}_{\sigma,r}^{y}, \mathcal{W}_{\sigma,r}^{z}, v_{\sigma,r}^{x}, v_{\sigma,r}^{y}, v_{\sigma,r}^{z}\}_{r=1}^{N_{\sigma,r}}$ ( $\sigma$ denotes density, and we will omit the subscript $\sigma$ for brevity from now on) as proposed in TensoRF [4]. $N_{r}$ is the number of ranks in matrix-vector decomposition and $\mathcal{W}_r^x \in \mathbb{R}^{H \times W}$ , $\mathcal{W}_r^y \in \mathbb{R}^{W \times D}$ , $\mathcal{W}_r^z \in \mathbb{R}^{H \times D}$ are matrices, $v_r^x \in \mathbb{R}^D$ , $v_r^y \in \mathbb{R}^H$ , $v_r^z \in \mathbb{R}^W$ , are vectors in $x, y, z$ directions, respectively. $H, W, D$ are the resolution of the grid. More formally, a 3D grid representation $G$ can be defined as follows. + +$$ +G = \sum_ {r = 1} ^ {N _ {r}} \sum_ {d \in \{x, y, z \}} v _ {r} ^ {d} \otimes \mathrm {i d w t} \left(\mathcal {W} _ {r} ^ {d}\right), \tag {5} +$$ + +where $\otimes$ denotes the outer product and $\mathrm{idwt}(\cdot)$ is a two-dimensional inverse discrete wavelet transform (IDWT). + +With a slight abuse of notation, a single level IDWT of a matrix $\mathcal{W} \in \mathcal{R}^{m \times n}$ can be written as follows. + +$$ +\begin{array}{l} \mathrm {i d w t} (\mathcal {W}) = \Phi \star (\uparrow 2) \left(\mathcal {W} _ {L L}\right) + \Psi_ {H L} \star (\uparrow 2) \left(\mathcal {W} _ {H L}\right) \\ + \Psi_ {L H} \star (\uparrow 2) \left(\mathcal {W} _ {L H}\right) + \Psi_ {H H} \star (\uparrow 2) \left(\mathcal {W} _ {H H}\right), \tag {6} \\ \end{array} +$$ + +where $\mathcal{W}_{LL},\mathcal{W}_{HL},\mathcal{W}_{LH},\mathcal{W}_{HH}\in \mathbb{R}^{m / 2\times n / 2}$ denote approximation, horizontal, vertical, and diagonal coefficients, respectively. $\star$ denotes 2D cross correlation and $(\uparrow 2):$ $\mathbb{R}^{m / 2\times n / 2}\to \mathbb{R}^{m\times n}$ denotes (nearest neighbor) upscaling. $\Psi$ and $\Phi$ are wavelet function and scaling function, respectively. We use the biorthogonal 4.4 wavelet function and its corresponding scaling function [6]. Even though we follow the grid representation of TensoRF, our method is not limited to the existing 2D plane-based representations but can + +be applied to any other approaches that adopt 2D planes to represent higher-dimensional structures. + +As shown in Eq. 5, we only apply IDWT over matrices $\mathcal{W}$ since we observed that IDWT to vectors resulted in significant quality degradation. As this transformation is differentiable, the model can be trained end-to-end. We can also obtain the 3D grid representation for appearance through the same process and provide all formulations in the supplementary material. + +Our method incurs insignificant costs during inference. Once training is finished, only one IDWT per grid is needed to transform wavelet grids into spatial grids. Thus, the remaining computational costs and time are exactly the same as the original spatial grid-based neural fields. + +# 3.2.1 Multi-level wavelet transform + +To further improve parameter sparsity and reconstruction quality, we propose using a multi-level wavelet transform. We experimentally found that higher-level wavelet transformations result in higher sparsity in grids (Sec. 4.3.1). However, naively using a multi-level wavelet transform on grid parameters degrades the representation quality (Tab. 1). We hypothesize that this is because the range of wavelet coefficients and the range of their gradients vary depending on the level of decomposition. In more detail, high-pass coefficients typically have a larger gradient scale with a smaller weight range. Different weight scales and gradients for different frequencies seem to hinder the optimization process. As such, we propose multiplying wavelet coefficients by the scales proportional to the inverse of frequency. With our proposed scaling factors $s$ , Eq. 5 can be rewritten as follows. + +$$ +G = \sum_ {r = 1} ^ {N _ {r}} \sum_ {d \in \{x, y, z \}} v _ {r} ^ {d} \otimes \mathrm {i d w t} \left(\mathcal {W} _ {r} ^ {d} \odot s _ {r} ^ {d}\right), \tag {7} +$$ + +where $\odot$ denotes the Hadamard product. For example, a two-level DWT, as shown in Fig. 3, generates a total of seven groups of coefficients (LL2, HL2, LH2, HH2, HL1, LH1, HH1). The scaling factors $s$ for HL2, LH2, and HH2 will be set to $1/2$ , and for HL1, LH1, and HH1 to $1/3$ . This scaling method enhances the quality of the reconstruction; more experiments are described in Sec. 4.3.2. + +# 3.3. Learning to mask + +Even though Wavelet coefficients can be sparse, we need additional methods to attain higher sparsity. Thus, we propose using element-wise masks to increase the portion of zero elements in grids. By jointly optimizing binary masks and grid parameters, we aim to zero out the majority of coefficients without significantly degrading the rendering quality. As we defined earlier, + +![](images/e0bf6b7659ae5bf5a24a39da257e851199bc516faada2a3a9a074c6d9d5d7dbe.jpg) +LL2 +HL2 + +![](images/70c07940a878603facf6acc54b1582d82250b83c0faf013bcbeb8176fc3a5f95.jpg) +HL2 +HH2 + +![](images/4a1537c997984619c22ce6ca166adb0c13877ab5fbac013c6a15bf2635b59e89.jpg) +HL1 + +![](images/de9ae86b7d4249439733caf2b6e4e02910d9c4ae19e8cc6df599ad42ac5c2c29.jpg) +HL1 +Figure 3. Multi-level discrete wavelet coefficients (2-level). LL, HL, LH, and HH denote approximation, horizontal, vertical, and diagonal coefficients, respectively. The numbers on the right denote the level of decomposition. + +![](images/f3e0dbf2714f67350e973b8c433c70828df89e78e2a414162e2d64c795efcc9b.jpg) +HH1 + +we have a set of trainable element-wise masks, $\mathcal{M} = \{\mathcal{M}_r^x,\mathcal{M}_r^y,\mathcal{M}_r^z,m_r^x,m_r^y,m_r^z\}_{r = 1}^{N_r}$ (note that we omitted the subscript $\sigma$ for brevity). During training, the grid parameters and corresponding binarized masks are multiplied element by element. Since calculating gradients directly from binarized masks is not feasible, we used the straight-through-estimator technique [1] to train and use masks. Formally, the masked matrix parameters $\widehat{\mathcal{W}}_r$ for $\mathcal{W}_r$ can be expressed as follows, + +$$ +\widehat {\mathcal {W}} _ {r} = \operatorname {s g} \left(\left(\mathcal {H} \left(\mathcal {M} _ {r}\right) - \sigma \left(\mathcal {M} _ {r}\right)\right) \odot \mathcal {W} _ {r}\right) + \sigma \left(\mathcal {M} _ {r}\right) \odot \mathcal {W} _ {r}, \tag {8} +$$ + +where $\mathfrak{sg}(\cdot)$ is the stop-gradient operator. $\sigma (\cdot)$ and $\mathcal{H}(\cdot)$ denote the element-wise sigmoid and Heaviside step function, respectively. We can also compute the masked vector parameters $\widehat{v}_r$ similarly. By replacing grid parameters $(\mathcal{W}_r,$ $v_{r})$ from Eq. (7) with masked grid parameters $(\widehat{\mathcal{W}}_r,\widehat{v}_r)$ , we can represent a masked 3D grid representation as follows: + +$$ +G = \sum_ {r = 1} ^ {N _ {r}} \sum_ {d \in \{x, y, z \}} \widehat {v} _ {r} ^ {d} \otimes \mathrm {i d w t} \left(\widehat {\mathcal {W}} _ {r} ^ {d} \odot s _ {r} ^ {d}\right) \tag {9} +$$ + +To make binarized mask values sparse, we use the sum of all mask values as the additional loss term $\mathcal{L}_m$ . The overall loss function is a sum of rendering loss $\mathcal{L}_r$ and the mask regularization term $\mathcal{L}_m$ . + +$$ +\mathcal {L} = \mathcal {L} _ {r} + \lambda_ {m} \mathcal {L} _ {m}. \tag {10} +$$ + +We use $\lambda_{m}$ to control the sparsity of the parameters of the grid representation. + +# 3.4. Compression pipeline + +With our proposed masking method and multi-level DWT, the ratio of zeros in grids can go up to more than + +![](images/10b955614e42106c4b08fca6b6ef434184514211e81702172d5f6d94a0eb9f82.jpg) +Figure 4. Proposed bitmap compression pipeline. RLE and Huffman denote run-length encoding and Huffman encoding, respectively. + +95%. However, sparse representations themselves do not reduce the total size. In this section, we describe our proposed compression pipeline for sparse grid parameters. Instead of storing grids as they are, we separately store nonzero coefficients and bitmaps (or masks) that indicate which coefficients are non-zero. Despite using 1-bit bitmaps, the overall bitmap size is large due to the large number of parameters. To reduce the bitmap sizes, we propose a compression pipeline with the following three stages: n-bit casting, run-length encoding (RLE), and the Huffman encoding [19]. The compression pipeline is illustrated in Fig. 4. + +Before applying the compression pipeline, we split wavelet coefficient masks by level of decomposition. For the 2-level wavelet transform, for instance, we group HL1, LH1, and HH1, then HL2, LH2, and HH2, and finally LL2 (Fig. 3). This is based on our experimental results that wavelet coefficients with the same levels of sparsity have similar levels of sparsity (Fig. 6). We found that grouping coefficients with similar sparsity results in better compression performance. In addition, we found that directly applying RLE to 1-bit streams causes inefficient bit allocation due to the numerous repeating zeros. Therefore, we first cast the binary mask values to n-bit unsigned integers, and perform RLE afterward. In our experiments, we set n to 8. Finally, we apply the Huffman encoding algorithm to the RLE-encoded streams to map values with a high probability to shorter bits. + +# 4. Experiments + +# 4.1. Experimental settings + +We compared our proposed method with other baselines in four datasets; NeRF synthetic dataset [23], Neural Sparse Voxel Fields (NSVF) [22] dataset, TanksAndTemps dataset [20], and LLFF dataset [10]. + +We used TensoRF VM-192 as the baseline model and applied the proposed method. To control the size of our method, we adjusted the mask regularization weight $\lambda_{m}$ the number of grid channels (from VM-192 to VM-384), + +and the resolution of grids. Unless otherwise specified, we used 8-bit quantization for every method. We compared our method with quantized NeRF, TensoRF models (CP and VM), cNeRF [2] and CCNeRF [42]. We also tried to compare our approach with VQAD [38], which efficiently compresses 3D scene representations using codebooks and vector quantization. However, VQAD requires depth maps to prune out vacant areas in advance. Otherwise, the required memory exceeds the memory capacity of the Tesla A100 equipped with 40 GB of memory. Since the datasets we used do not provide depth maps, we did not compare VQAD with ours. + +We followed the experimental settings of TensoRF [4]. We trained models for 30,000 iterations, each of which is a minibatch of 4,096 rays. We used the Adam optimizer and an exponential learning rate decay scheduler. Following TensoRF, the initial learning rates of grid-related parameters and MLP-related parameters were set to 0.02 and 0.001, respectively. Final learning rates were set at 1/10 of the initial learning rate. We updated the alpha masks at the 2000th, 4000th, 6000th, 11000th, 16000th, 21000th, and 26000th iterations. We set $\lambda_{m}$ in Eq. (10) to 1e-10 for high parameter sparsity and 5e-11 and 2.5e-11 for relatively lower sparsity. We set the initial values of masks $\mathcal{M}$ to one, and set their initial learning rate to the same learning rate as grid parameters $\phi$ . As a reconstruction quality measurement, we used PSNR. + +# 4.2. Results + +Figs. 1 and 5 show the quantitative performances of our method and the baselines on various datasets. Each graph displays the average PSNR and size of methods. More detailed numbers are available in the supplementary materials. + +As the rate-distortion curves show, our method achieves higher efficiency than the baseline models and other methods. Even with the base network (VM-192), our method consistently outperforms other baselines on all novel view synthesis datasets, under a small memory budget of $2\mathrm{MB}$ . We adjusted both $\lambda_{m}$ and the network size to increase the output size. By adjusting $\text{lambda}_{m}$ , we can control both size and quality while having no effect on computational costs, time, or memory requirements. Increasing network size (doubling the number of channels and increasing the resolution) can push the boundary even further. It is also worth mentioning that applying our method to a larger backbone model (VM-384) outperforms models with a smaller backbone model (VM-192), even with the same value of $\lambda_{m}$ or a similar model size. However, as the network grows in size, the computational costs and time required for training and testing increase. Therefore, it is still reasonable to use a small backbone model when using our proposed method. + +Fig. 7 shows the qualitative results with and without our proposed masking method. Trainable masks removed more + +![](images/4d9d79b8ea73f76a1eaf5fd70eb08c89818029238dbc68514cceb8d202447e47.jpg) +Figure 5. Rate distortion curves on the NSVF synthetic, TanksAndTemples, and LLFF dataset. We used 8-bit precision for every method. + +![](images/25bcb6b6b6bbee4b92d2c296d685755351782de0d5bcbd51383be3f8ce81e2b5.jpg) + +![](images/e50f7c0d0a55485ef572a64b096b41516b98767db7ce05aeeaab61b223533c74.jpg) + +![](images/8609fa5d6a211f7a711d10dc4fa5086413051cc09548871985880c858dcb7ca5.jpg) +(a) Raw mask +Figure 6. Qualitative analysis on mask $\mathcal{M}$ (Chair in the NeRF synthetic dataset). + +![](images/2067b199aa78fb46d8bb225d83e59f7cb9868a841cc091847681d9a2f74bc6b3.jpg) +(b) Binarized mask + +than $97\%$ of the grid parameters; nevertheless, the rendered results are still accurate, and the qualitative difference is almost imperceptible. This demonstrates that our proposed method efficiently eliminates unnecessary wavelet coefficients by leveraging trainable masks. + +Fig. 6 illustrates both raw and binarized masks after training. First of all, sparsity varies depending on the level of the wavelet transform. As shown in Fig. 6b, coefficients with lower frequencies have lower sparsity, whereas coefficients with higher frequencies have higher sparsity. What is interesting is that the raw mask values seem to reflect the characteristics of corresponding wavelet coefficients. Vertical, horizontal, and diagonal patterns can be found in the raw mask values for the vertical, horizontal, and diagonal coefficients, respectively. + +# 4.3. Ablation studies + +In this section, we analyze each component of our proposed method. Every experiment was conducted on the whole NeRF synthetic dataset, and we used the average PSNR as a measurement for representation quality. We used 4-level wavelet transform and set $\lambda_{m}$ to 1e-10, unless otherwise specified. + +
MethodsPSNR
DWT31.949
DWT w/o level-wise scaling31.145
DCT31.325
+ +Table 1. Ablation study on level-wise scaling. Without our proposed scaling method, the performance of DWT is worse than DCT. + +# 4.3.1 Discrete wavelet transform + +In this section, we evaluate the compactness of wavelet coefficients and the performance improvements brought by the multi-level wavelet transform. Fig. 8 shows the rate-distortion curves of four different levels of wavelet transform and non-wavelet coefficients. As shown in the figure, wavelet-based grid representation outperforms the grid in the spatial domain, especially when the sparsity is high. Furthermore, rate-distortion curves demonstrate that higher level of wavelet transform improves visual quality and sparsity. + +# 4.3.2 Level-wise wavelet coefficient scaling + +As mentioned in Sec. 3.2.1, we proposed multiplying wavelet coefficients by the scales proportional to the inverse of frequency. Thus, we compared two versions of wavelet transform; one with our proposed weight scaling (Sec. 3.2.1) and the other without it. As Tab. 1 shows, scaling wavelet coefficients improves reconstruction performance. + +# 4.3.3 Wavelet functions + +In this section, we analyze how different wavelet functions affect reconstruction quality. For comparison, we selected several wavelet functions: Haar, Coiflet 1, Daubechies 4, and reverse biorthogonal 4.4. As shown in Tab. 2, the type of wavelet function has little effect on reconstruction quality. Even Haar, the simplest wavelet function, performs + +![](images/763c28005c70c802cd1afc5bcb97d18903563016fbd1e36ffc828e36abcba959.jpg) +Figure 7. A qualitative analysis of grid sparsity. The sparsity refers to the ratio of zero elements in grids. With the help of trainable masks, removing more than $97\%$ of grid parameters only results in an imperceptible difference. + +![](images/971911b1e1283ea2eb744dc52f019a7cb5b3bb6c16faa846332575c31ae3a86a.jpg) + +![](images/b7879939e62fbc4841de393219326c0d9f9dfe9087abd4ccf68f03155f704e6e.jpg) +Figure 8. The rate distortion curves of different signal representation schemes (spatial, DWT, and DCT). Sparsity on the $x$ axis refers to the ratio of zeros in grid parameters. The grid sparsity was controlled by $\lambda_{m}$ . The numbers inside the parenthesis indicate the levels of the wavelet transform. + +
Wavelet FunctionPSNR
Haar31.889
Coiflets 131.846
Daubechies 431.734
biorthogonal 4.4 (default)31.949
reverse biorthogonal 4.431.727
+ +Table 2. Reconstruction performances of different wavelet functions. + +fairly well. Still, the biorthogonal 4.4 that we used shows the best performance among selected wavelet functions. + +# 4.3.4 Discrete cosine transform + +In this section, we compare the performance of DWT with that of DCT, which represents a signal with respect to the sum of cosine functions with different frequencies. DCT, + +similar to DWT, has an energy compaction property. That is, a small number of non-zero DCT coefficients are sufficient to represent a given signal. Because of its energy compactness, DCT or its variants are another widely used signal transformation for image, audio, and audio compression. However, as shown in Fig. 8 and Tab. 1, DWT shows superior compression and representation performance to DCT. We believe this is due to the non-repeating, non-smooth information in grids, for which DWT is more appropriate. + +# 5. Conclusion + +We propose a compact representation for grid-based neural fields, enabled by a novel masking strategy and multilevel wavelet transform. We demonstrate that these two components made it possible to remove more than $95\%$ of grid parameters without a significant loss of visual quality. With our proposed compression pipeline and the sparse wavelet coefficients, we achieved state-of-the-art performance under a memory budget of 2 MB. We believe we can reduce the size further with more sophisticated compression pipelines. + +Our proposed method is naturally constrained by the limitations of grid-based neural fields that are developed for bounded scenes or objects. We believe that expanding this grid-based representation to encompass unbounded or large scenes would be an intriguing direction, given that we can now compactly represent 3D objects with our proposed representation scheme. + +# Acknowledgments + +This work was supported in part by the Institute of Information and Communications Technology Planning and Evaluation (IITP) grants (2021-0-02068, IITP-2021-0-02052, IITP-2023-2020-0-01821, IITP-2019-0-00421) and National Research Foundation of Korea (NRF) grants (2022R1A4A3032913, 2022R1F1A1064184, 2022R1A4A3033571) funded by the Korea Government (MSIT). + +# References + +[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3, 5 +[2] Thomas Bird, Johannes Balle, Saurabh Singh, and Philip A Chou. 3d scene compression through entropy penalized neural representation functions. In 2021 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2021. 6 +[3] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, June 2022. 2, 3, 4 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision. Springer, 2022. 2, 3, 4, 6 +[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 1 +[6] Albert Cohen, Ingrid Daubechies, and J-C Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on pure and applied mathematics, 45(5):485-560, 1992. 4 +[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. 3 +[8] Chenxi Lola Deng and Enzo Tartaglione. Compressing explicit voxel grid representations: fast nerfs become also small. arXiv preprint arXiv:2210.12782, 2022. 3 +[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1 +[10] Ben Mildenhall et al. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6 +[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, SA '22, New York, NY, USA, 2022. Association for Computing Machinery. 1 +[12] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, + +volume 70 of Proceedings of Machine Learning Research, pages 1126-1135. PMLR, 06-11 Aug 2017. 1 +[13] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, June 2022. 2 +[14] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14346-14355, October 2021. 1 +[15] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. pages 5875-5884, October 2021. 2 +[16] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12902-12911, June 2022. 1 +[17] Binbin Huang, Xinhao Yan, Anpei Chen, Shenghua Gao, and Jingyi Yu. Pref: Phasorial embedding fields for compact neural representations. arXiv preprint arXiv:2205.13524, 2022. 2, 3, 4 +[18] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Binarized neural networks. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 3 +[19] David A Huffman. A method for the construction of minimum-redundancy codes. Proceedings of the IRE, 40(9):1098-1101, 1952. 2, 6 +[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), jul 2017. 6 +[21] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. 3 +[22] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 15651-15663. Curran Associates, Inc., 2020. 2, 6 +[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 4, 6 +[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), jul 2022. 2, 3 +[25] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? Adaptive + +rounding for post-training quantization. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7197-7206. PMLR, 13-18 Jul 2020. 3 +[26] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 1 +[27] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. pages 18419-18429, June 2022. 2 +[28] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 13032-13044. Curran Associates, Inc., 2021. 2 +[29] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 1 +[30] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision – ECCV 2016, pages 525–542, Cham, 2016. Springer International Publishing. 3 +[31] Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, and Andrea Tagliasacchi. Derf: Decomposed radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14153-14161, 2021. 1 +[32] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14335-14345, October 2021. 1, 3 +[33] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. VoxGRAF: Fast 3d-aware image synthesis with sparse voxel grids. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2 +[34] Vincent Sitzmann, Eric Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 10136-10147. Curran Associates, Inc., 2020. 1 +[35] Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The jpeg 2000 still image compression standard. IEEE Signal processing magazine, 18(5):36-58, 2001. 2, 3 +[36] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 2 +[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields + +reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, June 2022. 2 +[38] Towaki Takikawa, Alex Evans, Jonathan Tremblay, Thomas Müller, Morgan McGuire, Alec Jacobson, and Sanja Fidler. Variable bitrate neural fields. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 6 +[39] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. arXiv preprint arXiv:2101.10994, 2021. 2 +[40] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P. Srinivasan, Jonathan T. Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 1, 2 +[41] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3 +[42] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable neRF via rank-residual decomposition. In Advances in Neural Information Processing Systems, 2022. 2, 3, 6 +[43] Gregory K. Wallace. TheJPEG still picture compression standard.Commun.ACM,34(4):30-44,apr 1991.2 +[44] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 612-629, Cham, 2022. Springer Nature Switzerland. 1 +[45] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13524-13534, June 2022. 2, 3 +[46] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology, 13(7):560-576, 2003. 2 +[47] Zhijie Wu, Yuhe Jin, and Kwang Moo Yi. Neural fourier filter bank. arXiv e-prints, pages arXiv-2212, 2022. 3 +[48] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2, 3 +[49] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5752-5761, October 2021. 2, 3 + +[50] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. arXiv preprint arXiv:2205.14330, 2022. 2 +[51] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained ternary quantization. In International Conference on Learning Representations, 2017. 3 \ No newline at end of file diff --git a/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/images.zip b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..956ac407411ed37680a4997c51c062abd3812ff0 --- /dev/null +++ b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f53010a7d45296349479d2b979a0960000455849dc6a4d9d8b9733de4ef0d37 +size 362164 diff --git a/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/layout.json b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5e35c20eee2e0573127f78849753dd4cabbad315 --- /dev/null +++ b/2023/Masked Wavelet Representation for Compact Neural Radiance Fields/layout.json @@ -0,0 +1,8901 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "type": "text", + "content": "Masked Wavelet Representation for Compact Neural Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 141, + 124, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 124, + 157 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 124, + 157 + ], + "type": "text", + "content": "Daniel Rho1*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 143, + 240, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 143, + 240, + 157 + ], + "spans": [ + { + "bbox": [ + 140, + 143, + 240, + 157 + ], + "type": "text", + "content": "Byeonghyeon Lee" + }, + { + "bbox": [ + 140, + 143, + 240, + 157 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 143, + 345, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 143, + 345, + 184 + ], + "spans": [ + { + "bbox": [ + 255, + 143, + 345, + 184 + ], + "type": "text", + "content": "Seungtae Nam² \nEunbyung Park²,³† \n¹AI2XL, KT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 356, + 143, + 430, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 143, + 430, + 156 + ], + "spans": [ + { + "bbox": [ + 356, + 143, + 430, + 156 + ], + "type": "text", + "content": "Joo Chan Lee²" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 451, + 143, + 541, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 143, + 541, + 157 + ], + "spans": [ + { + "bbox": [ + 451, + 143, + 541, + 157 + ], + "type": "text", + "content": "Jong Hwan " + }, + { + "bbox": [ + 451, + 143, + 541, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{Ko}^{2,3\\dagger}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 185, + 455, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 455, + 200 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 455, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 185, + 455, + 200 + ], + "type": "text", + "content": "Department of Artificial Intelligence, Sungkyunkwan University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 199, + 493, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 199, + 493, + 214 + ], + "spans": [ + { + "bbox": [ + 100, + 199, + 493, + 214 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 199, + 493, + 214 + ], + "type": "text", + "content": "Department of Electrical and Computer Engineering, Sungkyunkwan University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 266, + 290, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 590 + ], + "type": "text", + "content": "Neural radiance fields (NeRF) have demonstrated the potential of coordinate-based neural representation (neural fields or implicit neural representation) in neural rendering. However, using a multi-layer perceptron (MLP) to represent a 3D scene or object requires enormous computational resources and time. There have been recent studies on how to reduce these computational inefficiencies by using additional data structures, such as grids or trees. Despite the promising performance, the explicit data structure necessitates a substantial amount of memory. In this work, we present a method to reduce the size without compromising the advantages of having additional data structures. In detail, we propose using the wavelet transform on grid-based neural fields. Grid-based neural fields are for fast convergence, and the wavelet transform, whose efficiency has been demonstrated in high-performance standard codes, is to improve the parameter efficiency of grids. Furthermore, in order to achieve a higher sparsity of grid coefficients while maintaining reconstruction quality, we present a novel trainable masking approach. Experimental results demonstrate that non-spatial grid coefficients, such as wavelet coefficients, are capable of attaining a higher level of sparsity than spatial grid coefficients, resulting in a more compact representation. With our proposed mask and compression pipeline, we achieved state-of-the-art performance within a memory budget of 2 MB. Our code is available at https://github.com/daniel03c1/maskedwavelet_nerf." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 614, + 128, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 128, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 128, + 627 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "content": "Recent advances in coordinate-based neural representation (neural fields or implicit neural representation) have demonstrated remarkable performance in many applications. In particular, neural radiance fields (NeRF) have" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 241, + 545, + 386 + ], + "blocks": [ + { + "bbox": [ + 309, + 241, + 545, + 386 + ], + "lines": [ + { + "bbox": [ + 309, + 241, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 309, + 241, + 545, + 386 + ], + "type": "image", + "image_path": "959d47197dd2a619d446a16183c11c28e9004d14b86c2dd6dcac18aea2a9d6f2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 392, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 306, + 392, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 546, + 415 + ], + "type": "text", + "content": "Figure 1. Rate-distortion curves on the NeRF synthetic dataset. The numbers inside parenthesis denote the axis resolution of grids." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 544 + ], + "type": "text", + "content": "sparked interest by synthesizing high-quality images from novel viewpoints. It uses a multi-layer perceptron (MLP) with positional encoding to map coordinates to corresponding colors and opacities. Combined with the differentiable volumetric rendering and the neural network's architectural priors, it has shown great potential to be a new representation paradigm. However, the high computational costs (in both training and inference) have been a significant bottleneck, often taking several days to converge." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": "Several follow-up studies have been proposed to accelerate training and inference times [5, 9, 11, 14, 16, 31, 32, 40, 44]. To speed up inference, KiloNeRF [32] proposed splitting a 3D scene into thousands of partial scenes, each of which is assigned a tiny, distinct neural network. While achieving impressive speed-up, it requires a massive amount of memory storage. On the other hand, FastNeRF [14] suggested caching and factorizing the NeRF network to reduce computational costs. Meanwhile, meta-learning algorithms [12, 26, 29] have also been applied to accelerate the training time, and they have shown faster convergence with the learned initialization [34, 40]. However, it requires well-organized large-scale datasets and a pre-training process. Furthermore, the rendering costs re" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20680" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "main unchanged since they use the same network architecture [40], and meta-learning algorithms often suffer from poor out-of-distribution generalization performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 241 + ], + "type": "text", + "content": "Alternatively, there has been a surge of recent interest in incorporating classical data structures, such as grids or trees, into the NeRF framework [13, 22, 24, 33, 38, 39, 49]. Incorporating additional data structures has significantly reduced training and inference time (from days to a few minutes) without compromising reconstruction quality. However, the overall size dramatically increases due to these dense and volumetric structures. In order to reduce the spatial complexity, several methods have been proposed, including pruning areas [33, 37, 49], encodings [24, 38], and tensor decomposition [3, 4, 17]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 410 + ], + "type": "text", + "content": "This paper aims to further improve the spatial complexity while maintaining the rendering quality. Leveraging the decades of research on standard compression algorithms [35,36,46], we propose compressing grid-based neural fields using frequency-based transformations. In the frequency domain, a large portion of the coefficients can be discarded without considerably degrading the reconstruction quality, and most standard compression algorithms have exploited frequency domain representations. Thus, we propose using this property on grid-based neural fields to maximize parameter efficiency. Among other alternatives, we employ the discrete wavelet transform (DWT) due to its compactness and ability to efficiently capture both global and local information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 603 + ], + "type": "text", + "content": "Once we obtain sparse representations via frequency domain representations, we can take advantage of the existing compression techniques. Unlike conventional media data (e.g., image and audio), however, no off-the-shelf compression tools exist that we can leverage without complicated engineering efforts. In addition, since NeRF, or neural rendering networks in general, is a relatively new data format, the characteristics or patterns of their coefficients have not been thoroughly investigated. Thus, we present a compression pipeline for our purposes. To automatically filter out unnecessary coefficients, we propose a trainable binary mask. For each 3D scene, we jointly optimize grid parameters and their corresponding masks. This perscene optimization strategy can be more optimal than the global quantization table used in standard image compression codecs [43]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 604, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 688 + ], + "type": "text", + "content": "To compress sparse grid representations, we first merge masks with wavelet coefficients to zero out coefficients and then apply standard compression algorithms to masked coefficients. We utilize the run-length encoding to encode binary information about which coefficients are non-zero. For further compression, we apply one of the entropy coding algorithms, Huffman encoding [19], to these encoded outputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Our method incurs negligible computational costs at test time, requiring only one inverse DWT (IDWT) per grid. Af-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "ter the IDWT, the wavelet grids are transformed into spatial grids, eliminating the need for additional IDWT during rendering. As a result, the computing time and costs (including memory costs) are identical to the original spatial grid-based NeRF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 133, + 501, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 133, + 501, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 133, + 501, + 144 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 153, + 545, + 316 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 153, + 545, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 545, + 213 + ], + "type": "text", + "content": "- We propose using wavelet coefficients to improve parameter sparsity and reconstruction quality. Through experiments, we show that the wavelet coefficients can be more compact than the spatial domain coefficients in neural radiance fields." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 223, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 223, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 223, + 545, + 282 + ], + "type": "text", + "content": "- We propose a trainable mask that can be applied to any grid-based neural representation. Experimental results demonstrate that our proposed masking method can zero out more than " + }, + { + "bbox": [ + 317, + 223, + 545, + 282 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 317, + 223, + 545, + 282 + ], + "type": "text", + "content": " of the total grid parameters while maintaining high reconstruction quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 292, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 292, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 292, + 545, + 316 + ], + "type": "text", + "content": "- We achieve state-of-the-art performance in novel view synthesis under a memory budget of 2 MB." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 327, + 397, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 397, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 397, + 340 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 347, + 410, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 410, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 410, + 361 + ], + "type": "text", + "content": "2.1. Neural rendering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 365, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 449 + ], + "type": "text", + "content": "Neural radiance fields. NeRF [23] has demonstrated that 3D scenes and objects can be successfully represented using coordinate-based neural representation. NeRF renders a scene by casting a ray per pixel and sampling colors and opacity from points that lie on the ray. However, since this approach relies on the dense, point-wise sampling of color values and opacity, the computational costs are significant." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": "Due to this inefficiency, there has been considerable interest in improving computational efficiency. A plethora of methods propose to adopt data structures such as trees [45, 49], point clouds [27, 28, 48, 50], and grids [4, 13, 15, 24, 33, 37, 38] to efficiently represent 3D scenes and objects. PlenOctrees [49] use an octree structure for real-time rendering. Despite the reduced rendering costs, the increased size of more than ten times the size of standard MLP-based neural fields, such as NeRF [23], limits their wide application. Point cloud-based approaches [27, 48] reduce the number of points sampled per ray by utilizing the scene geometry of a point cloud. Lastly, a number of studies have demonstrated the efficiency of using grid structures, including vectors, matrices, and tensors, on neural fields [4, 13, 24, 37]. With the use of these grid structures, training iterations that once required tens of hours can now be completed in a matter of minutes. Furthermore, real-time rendering at inference time is also possible. However, dense 3D grid structures require substantial amounts of memory, often exceeding 1 GB. To alleviate this memory requirement, several methods have been proposed to reduce the size of grids [4, 24, 38, 42]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20681" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 192 + ], + "type": "text", + "content": "To achieve near-instant rendering with reasonable memory requirements, Instant-NGP [24] proposed using hash-based multi-resolution grids. Using hash functions, each grid maps input coordinates to corresponding feature vectors. Similarly, VQAD [38] proposed using codebooks and vector quantization rather than the hash function. This enables control over the overall size of the neural fields. However, during training, a sizable amount of memory is required to learn which code from the codebook to use for each point on the grid." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "type": "text", + "content": "In order to alleviate the memory requirement of using 3D grids, another line of work decomposes 3D grids into lower dimensional representations, such as planes and vectors [3, 4, 17, 42]. EG3D [3] proposes a tri-plane approach, which represents a 3D scene with three perpendicular planes and uses feature vectors, extracted separately from each plane, as inputs for the following MLPs. TensoRF [4] is another parameter-efficient yet expressive method for decomposing dense three-dimensional grids into a smaller set of parameters, such as planes and lines. Although these methods dramatically reduce the time and space complexity of the 3D scene and object representation, their overall sizes are larger than those of MLP-only methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 364, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 508 + ], + "type": "text", + "content": "Frequency-based representation. Several studies have explored frequency-based parameterization for efficient neural field representations [17, 41, 45, 47]. Fourier PlenOctree [45] has demonstrated that the Fourier transform can improve both parameter efficiency and training speed of the PlenOctree structure [49]. Nevertheless, the overall size is larger than MLP-only methods. PREF [17] has also exploited the Fourier transform in 3D scene representation. It applies the Fourier transform to a tensor decomposition-based representation. However, it has not shown notable improvements in both the representation quality and the parameter efficiency compared to spatial grid representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Besides the Fourier transform, there is another line of signal representations in the frequency domain: the wavelet transform. The wavelet transform decomposes a signal using a set of basis functions called wavelets. The wavelet transform can extract both frequency and temporal information, in contrast to the Fourier transform, which can only capture frequency information. Because each coefficient covers a different frequency and period in time, it is known that the wavelet transform can represent transient signals more compactly than Fourier-based transforms. This compactness contributes to standard compression CODECs, such as JPEG 2000 [35]. Motivated by the success of the standard CODECs, we aim to mimic their compression pipeline and apply it to the neural radiance fields. One disadvantage of the wavelet transform is that it can be less efficient than Fourier-based transforms for smooth signals. However, we believe that most situations in grid-based neural fields" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "are not the case, considering that each grid's resolution is constrained to accurately depict detailed scenes and objects with limited memory budgets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 124, + 420, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 124, + 420, + 136 + ], + "spans": [ + { + "bbox": [ + 306, + 124, + 420, + 136 + ], + "type": "text", + "content": "2.2. Model compression" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 144, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 180 + ], + "type": "text", + "content": "In recent years, many works have studied to downsize neural networks through various techniques such as weight pruning and quantization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 182, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 374 + ], + "type": "text", + "content": "Pruning. For efficient representations, pruning methods have been explored in neural fields. Point-NeRF [48] proposes a pruning strategy for point cloud representation by imposing a sparsity loss on point confidence. Re:NeRF [8] has validated that simply applying a general pruning technique, which gradually removes coefficients throughout the course of training, to grid-based neural representations results in a significant performance drop. It proposed iterative parameter removal and conditional parameter inclusion and demonstrated comparable compression performance. However, it has not shown desirable performance under limited memory budgets. Similarly, KiloNeRF [32] prunes out parameters for unused areas using a fixed threshold. Instead, we propose a trainable mask that can remove a large number of coefficients while causing only minor degradation in the reconstruction quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 376, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 604 + ], + "type": "text", + "content": "Quantization. In terms of neural network quantization, there are two main approaches: post-quantization (PTQ) [21, 25] and quantization-aware training (QAT) [7, 18, 30, 51]. Post-quantization methods quantify network parameters after training is finished. PTQ has the advantage that it does not require additional training iterations or datasets. Furthermore, it supports arbitrary-bit quantization. Performance, however, might degrade because the network parameters were not optimized for the target bit precision. Unlike post-quantization, QAT optimizes neural networks for a certain bit precision during training iterations. Since quantization is not differentiable, it usually relies on the straight-through estimator (STE) [1] for network optimization [7]. QAT has been validated for its effectiveness in neural representation models. However, PTQ has the disadvantage of requiring additional training samples and iterations when quantizing the weights of a pretrained network. Since we train neural fields from scratch, we use PTQ for weight quantization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 623, + 366, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 366, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 366, + 635 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 644, + 432, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 432, + 656 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 432, + 656 + ], + "type": "text", + "content": "3.1. Neural radiance fields" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "We consider a neural radiance field leveraging grid representations. It takes an input coordinate " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " and viewing direction, " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": ", generating a four-dimensional vector consisting of a density and three-channel RGB colors, de" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20682" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 74, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 53, + 74, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 53, + 74, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 547, + 148 + ], + "type": "image", + "image_path": "7f76d1a1d654f1e03ca4723bbfb1f3ebf321d4586161904668b80bb5ce1ad621.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 165, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 46, + 165, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 547, + 221 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our model. Images with purple borders illustrate wavelet coefficients, while those with orange borders visualize spatial coefficients. The wavelet coefficients in each 2D grid are multiplied with a binarized mask to form a masked wavelet coefficient grid. Masked wavelet coefficients are then inverse-transformed to spatial features. We sample feature vectors for input coordinates using bilinear interpolation on the grids. Then, opacity and color for each input coordinate are estimated using the sampled feature vectors following the defined model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 241, + 116, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 116, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 116, + 252 + ], + "type": "text", + "content": "fined as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 261, + 287, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 261, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 287, + 274 + ], + "type": "interline_equation", + "content": "\\sigma (x) = f _ {\\sigma} \\left(x; \\gamma_ {\\sigma}, \\mathcal {M} _ {\\sigma}\\right), \\tag {1}", + "image_path": "79df702dd98e3cc447532ab7d1f2dda5291c0585590170b2b4be9e52e67b9fa1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 276, + 287, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 276, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 287, + 289 + ], + "type": "interline_equation", + "content": "c (x, d) = f _ {c} (x, d; \\theta , \\gamma_ {c}, \\mathcal {M} _ {c}), \\tag {2}", + "image_path": "9511a0fd486cb6b8399d618ad7c8d4e25cbdd853a7c22039170c345a3d774fed.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "type": "text", + "content": " is the parameter of an MLP, " + }, + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\gamma = \\{\\gamma_{\\sigma}, \\gamma_{c}\\}" + }, + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "type": "text", + "content": " is a set of grid parameters, and " + }, + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathcal{M}_{\\sigma}, \\mathcal{M}_{c}\\}" + }, + { + "bbox": [ + 47, + 295, + 287, + 355 + ], + "type": "text", + "content": " is a set of masks for grid parameters, which will be described shortly (Sec. 3.3). The following volumetric rendering equation is used to synthesize novel views:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 361, + 287, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 361, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 89, + 361, + 287, + 390 + ], + "type": "interline_equation", + "content": "C (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (r (t)) c (r (t), d) d t, \\tag {3}", + "image_path": "419c1f253fa982e8b6b9bb4cd77a95b5ffa5dd87e512284db3cd1041ba6f4dfd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 392, + 287, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 392, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 118, + 392, + 287, + 419 + ], + "type": "interline_equation", + "content": "T (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (r (s)) d s\\right), \\tag {4}", + "image_path": "e81410dc118a73a525325129c9de5ca0ec7f463005245d52fd1d05f6d5c23ea7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "inline_equation", + "content": "r(t)" + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "text", + "content": " is a ray from a camera viewpoint, and " + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "inline_equation", + "content": "C(r)" + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "text", + "content": " is the expected color of the ray " + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "inline_equation", + "content": "r(t)" + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "text", + "content": ". Two integral bounds (near and far) are denoted as " + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 46, + 425, + 287, + 473 + ], + "type": "text", + "content": ", respectively. Please refer to NeRF [23] for more details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 480, + 268, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 268, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 268, + 493 + ], + "type": "text", + "content": "3.2. Wavelet transform on decomposed tensors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": "Frequency-based algorithms, such as discrete cosine transform (DCT) or discrete wavelet transform (DWT), have been developed and improved over the decades. For high compression performance, standard CODEs have used them to make sparse representations. However, in the case of 3D grid representation, their computational complexity increases cubically with grid resolution, making high-resolution 3D scene and object representations impractical. In order to use frequency-based algorithms for compact representation, we need lower-dimensional data, such as 2D planes or 1D lines." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Recent studies have explored various decomposition methods to reduce the spatial complexity of 3D grid-based representations. Among them, plane-based representations have achieved remarkable success in reducing the number of parameters while maintaining the rendering performance [3, 4, 17]. They propose decomposing a 3D tensor into a set of 2D planes or 1D vectors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "content": "To combine the best of both worlds, we propose using the wavelet transform on 2D plane-based neural fields [4]. We use the wavelet transform because of its compactness, especially for non-repeating, non-smooth signals." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": "For efficient 3D object and scene representation, recent studies proposed using a lower-dimensional grid (2D planes or 1D lines) [3, 4]. For example, EG3D [3] utilizes three 2D planes (tri-plane) and TensoRF [4] employs three sets, each consisting of a plane and a vector. We describe our method based on TensoRF. In detail, we use a set of 2D matrices and 1D vectors for grid representation, " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\gamma_{\\sigma} = \\{\\mathcal{W}_{\\sigma,r}^{x}, \\mathcal{W}_{\\sigma,r}^{y}, \\mathcal{W}_{\\sigma,r}^{z}, v_{\\sigma,r}^{x}, v_{\\sigma,r}^{y}, v_{\\sigma,r}^{z}\\}_{r=1}^{N_{\\sigma,r}}" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": " denotes density, and we will omit the subscript " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": " for brevity from now on) as proposed in TensoRF [4]. " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": " is the number of ranks in matrix-vector decomposition and " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_r^x \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_r^y \\in \\mathbb{R}^{W \\times D}" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_r^z \\in \\mathbb{R}^{H \\times D}" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": " are matrices, " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "v_r^x \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "v_r^y \\in \\mathbb{R}^H" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "v_r^z \\in \\mathbb{R}^W" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": ", are vectors in " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": " directions, respectively. " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "H, W, D" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": " are the resolution of the grid. More formally, a 3D grid representation " + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 289, + 546, + 479 + ], + "type": "text", + "content": " can be defined as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 479, + 545, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 479, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 351, + 479, + 545, + 514 + ], + "type": "interline_equation", + "content": "G = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} v _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\mathcal {W} _ {r} ^ {d}\\right), \\tag {5}", + "image_path": "a0e08ffa46984c1c93a7c2e843aa99436645eb168f0c5cf3d243ee86ed6c5837.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "type": "text", + "content": " denotes the outer product and " + }, + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\mathrm{idwt}(\\cdot)" + }, + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "type": "text", + "content": " is a two-dimensional inverse discrete wavelet transform (IDWT)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 540, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 563 + ], + "type": "text", + "content": "With a slight abuse of notation, a single level IDWT of a matrix " + }, + { + "bbox": [ + 304, + 540, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{W} \\in \\mathcal{R}^{m \\times n}" + }, + { + "bbox": [ + 304, + 540, + 545, + 563 + ], + "type": "text", + "content": " can be written as follows." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 567, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 545, + 605 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {i d w t} (\\mathcal {W}) = \\Phi \\star (\\uparrow 2) \\left(\\mathcal {W} _ {L L}\\right) + \\Psi_ {H L} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {H L}\\right) \\\\ + \\Psi_ {L H} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {L H}\\right) + \\Psi_ {H H} \\star (\\uparrow 2) \\left(\\mathcal {W} _ {H H}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "0596a85f6f8668e5b5458f668e64a08ece5b56e7dc50a76b1e08fa806181e238.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{LL},\\mathcal{W}_{HL},\\mathcal{W}_{LH},\\mathcal{W}_{HH}\\in \\mathbb{R}^{m / 2\\times n / 2}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " denote approximation, horizontal, vertical, and diagonal coefficients, respectively. " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " denotes 2D cross correlation and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\uparrow 2):" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{m / 2\\times n / 2}\\to \\mathbb{R}^{m\\times n}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " denotes (nearest neighbor) upscaling. " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " are wavelet function and scaling function, respectively. We use the biorthogonal 4.4 wavelet function and its corresponding scaling function [6]. Even though we follow the grid representation of TensoRF, our method is not limited to the existing 2D plane-based representations but can" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20683" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "be applied to any other approaches that adopt 2D planes to represent higher-dimensional structures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 180 + ], + "type": "text", + "content": "As shown in Eq. 5, we only apply IDWT over matrices " + }, + { + "bbox": [ + 46, + 96, + 287, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 96, + 287, + 180 + ], + "type": "text", + "content": " since we observed that IDWT to vectors resulted in significant quality degradation. As this transformation is differentiable, the model can be trained end-to-end. We can also obtain the 3D grid representation for appearance through the same process and provide all formulations in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "content": "Our method incurs insignificant costs during inference. Once training is finished, only one IDWT per grid is needed to transform wavelet grids into spatial grids. Thus, the remaining computational costs and time are exactly the same as the original spatial grid-based neural fields." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 206, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 206, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 206, + 268 + ], + "type": "text", + "content": "3.2.1 Multi-level wavelet transform" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 276, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 467 + ], + "type": "text", + "content": "To further improve parameter sparsity and reconstruction quality, we propose using a multi-level wavelet transform. We experimentally found that higher-level wavelet transformations result in higher sparsity in grids (Sec. 4.3.1). However, naively using a multi-level wavelet transform on grid parameters degrades the representation quality (Tab. 1). We hypothesize that this is because the range of wavelet coefficients and the range of their gradients vary depending on the level of decomposition. In more detail, high-pass coefficients typically have a larger gradient scale with a smaller weight range. Different weight scales and gradients for different frequencies seem to hinder the optimization process. As such, we propose multiplying wavelet coefficients by the scales proportional to the inverse of frequency. With our proposed scaling factors " + }, + { + "bbox": [ + 46, + 276, + 287, + 467 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 276, + 287, + 467 + ], + "type": "text", + "content": ", Eq. 5 can be rewritten as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 475, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 475, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 81, + 475, + 287, + 510 + ], + "type": "interline_equation", + "content": "G = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} v _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\mathcal {W} _ {r} ^ {d} \\odot s _ {r} ^ {d}\\right), \\tag {7}", + "image_path": "2630b48cf8efbe5a7a97824a1f834dba7284f562483efbbc9e05e1c704e9d824.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "content": " denotes the Hadamard product. For example, a two-level DWT, as shown in Fig. 3, generates a total of seven groups of coefficients (LL2, HL2, LH2, HH2, HL1, LH1, HH1). The scaling factors " + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "content": " for HL2, LH2, and HH2 will be set to " + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "content": ", and for HL1, LH1, and HH1 to " + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 46, + 519, + 287, + 602 + ], + "type": "text", + "content": ". This scaling method enhances the quality of the reconstruction; more experiments are described in Sec. 4.3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 152, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 152, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 152, + 623 + ], + "type": "text", + "content": "3.3. Learning to mask" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Even though Wavelet coefficients can be sparse, we need additional methods to attain higher sparsity. Thus, we propose using element-wise masks to increase the portion of zero elements in grids. By jointly optimizing binary masks and grid parameters, we aim to zero out the majority of coefficients without significantly degrading the rendering quality. As we defined earlier," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 70, + 425, + 98 + ], + "blocks": [ + { + "bbox": [ + 319, + 70, + 425, + 98 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 425, + 98 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 425, + 98 + ], + "type": "image", + "image_path": "e0bf6b7659ae5bf5a24a39da257e851199bc516faada2a3a9a074c6d9d5d7dbe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 99, + 353, + 106 + ], + "lines": [ + { + "bbox": [ + 340, + 99, + 353, + 106 + ], + "spans": [ + { + "bbox": [ + 340, + 99, + 353, + 106 + ], + "type": "text", + "content": "LL2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 99, + 405, + 106 + ], + "lines": [ + { + "bbox": [ + 375, + 99, + 405, + 106 + ], + "spans": [ + { + "bbox": [ + 375, + 99, + 405, + 106 + ], + "type": "text", + "content": "HL2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 110, + 425, + 137 + ], + "blocks": [ + { + "bbox": [ + 319, + 110, + 425, + 137 + ], + "lines": [ + { + "bbox": [ + 319, + 110, + 425, + 137 + ], + "spans": [ + { + "bbox": [ + 319, + 110, + 425, + 137 + ], + "type": "image", + "image_path": "70c07940a878603facf6acc54b1582d82250b83c0faf013bcbeb8176fc3a5f95.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 137, + 353, + 144 + ], + "lines": [ + { + "bbox": [ + 339, + 137, + 353, + 144 + ], + "spans": [ + { + "bbox": [ + 339, + 137, + 353, + 144 + ], + "type": "text", + "content": "HL2" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 137, + 405, + 144 + ], + "lines": [ + { + "bbox": [ + 375, + 137, + 405, + 144 + ], + "spans": [ + { + "bbox": [ + 375, + 137, + 405, + 144 + ], + "type": "text", + "content": "HH2" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 428, + 70, + 533, + 123 + ], + "blocks": [ + { + "bbox": [ + 428, + 70, + 533, + 123 + ], + "lines": [ + { + "bbox": [ + 428, + 70, + 533, + 123 + ], + "spans": [ + { + "bbox": [ + 428, + 70, + 533, + 123 + ], + "type": "image", + "image_path": "4a1537c997984619c22ce6ca166adb0c13877ab5fbac013c6a15bf2635b59e89.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 124, + 487, + 131 + ], + "lines": [ + { + "bbox": [ + 473, + 124, + 487, + 131 + ], + "spans": [ + { + "bbox": [ + 473, + 124, + 487, + 131 + ], + "type": "text", + "content": "HL1" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 319, + 148, + 425, + 201 + ], + "blocks": [ + { + "bbox": [ + 319, + 148, + 425, + 201 + ], + "lines": [ + { + "bbox": [ + 319, + 148, + 425, + 201 + ], + "spans": [ + { + "bbox": [ + 319, + 148, + 425, + 201 + ], + "type": "image", + "image_path": "de9ae86b7d4249439733caf2b6e4e02910d9c4ae19e8cc6df599ad42ac5c2c29.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 202, + 378, + 209 + ], + "lines": [ + { + "bbox": [ + 365, + 202, + 378, + 209 + ], + "spans": [ + { + "bbox": [ + 365, + 202, + 378, + 209 + ], + "type": "text", + "content": "HL1" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 233, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 305, + 233, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 545, + 277 + ], + "type": "text", + "content": "Figure 3. Multi-level discrete wavelet coefficients (2-level). LL, HL, LH, and HH denote approximation, horizontal, vertical, and diagonal coefficients, respectively. The numbers on the right denote the level of decomposition." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 427, + 148, + 533, + 201 + ], + "blocks": [ + { + "bbox": [ + 427, + 148, + 533, + 201 + ], + "lines": [ + { + "bbox": [ + 427, + 148, + 533, + 201 + ], + "spans": [ + { + "bbox": [ + 427, + 148, + 533, + 201 + ], + "type": "image", + "image_path": "f3e0dbf2714f67350e973b8c433c70828df89e78e2a414162e2d64c795efcc9b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 202, + 487, + 209 + ], + "lines": [ + { + "bbox": [ + 473, + 202, + 487, + 209 + ], + "spans": [ + { + "bbox": [ + 473, + 202, + 487, + 209 + ], + "type": "text", + "content": "HH1" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "text", + "content": "we have a set of trainable element-wise masks, " + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathcal{M}_r^x,\\mathcal{M}_r^y,\\mathcal{M}_r^z,m_r^x,m_r^y,m_r^z\\}_{r = 1}^{N_r}" + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "text", + "content": " (note that we omitted the subscript " + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "text", + "content": " for brevity). During training, the grid parameters and corresponding binarized masks are multiplied element by element. Since calculating gradients directly from binarized masks is not feasible, we used the straight-through-estimator technique [1] to train and use masks. Formally, the masked matrix parameters " + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{W}}_r" + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_r" + }, + { + "bbox": [ + 305, + 297, + 545, + 405 + ], + "type": "text", + "content": " can be expressed as follows," + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 309, + 411, + 545, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 411, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 309, + 411, + 545, + 438 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {W}} _ {r} = \\operatorname {s g} \\left(\\left(\\mathcal {H} \\left(\\mathcal {M} _ {r}\\right) - \\sigma \\left(\\mathcal {M} _ {r}\\right)\\right) \\odot \\mathcal {W} _ {r}\\right) + \\sigma \\left(\\mathcal {M} _ {r}\\right) \\odot \\mathcal {W} _ {r}, \\tag {8}", + "image_path": "04226e346cf94f7f4b028d41301c236017d4dc35d651c5a1d09332ae0436b289.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathfrak{sg}(\\cdot)" + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "text", + "content": " is the stop-gradient operator. " + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "text", + "content": " denote the element-wise sigmoid and Heaviside step function, respectively. We can also compute the masked vector parameters " + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\widehat{v}_r" + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "text", + "content": " similarly. By replacing grid parameters " + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "inline_equation", + "content": "(\\mathcal{W}_r," + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "inline_equation", + "content": "v_{r})" + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "text", + "content": " from Eq. (7) with masked grid parameters " + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "inline_equation", + "content": "(\\widehat{\\mathcal{W}}_r,\\widehat{v}_r)" + }, + { + "bbox": [ + 305, + 445, + 545, + 518 + ], + "type": "text", + "content": " , we can represent a masked 3D grid representation as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 341, + 525, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 525, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 341, + 525, + 545, + 559 + ], + "type": "interline_equation", + "content": "G = \\sum_ {r = 1} ^ {N _ {r}} \\sum_ {d \\in \\{x, y, z \\}} \\widehat {v} _ {r} ^ {d} \\otimes \\mathrm {i d w t} \\left(\\widehat {\\mathcal {W}} _ {r} ^ {d} \\odot s _ {r} ^ {d}\\right) \\tag {9}", + "image_path": "f0e54fc9233aa4a12f727e444ec4e499cfef22be1fa4d6cd1314e444b09519fd.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "type": "text", + "content": "To make binarized mask values sparse, we use the sum of all mask values as the additional loss term " + }, + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "type": "text", + "content": ". The overall loss function is a sum of rendering loss " + }, + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "type": "text", + "content": " and the mask regularization term " + }, + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 305, + 566, + 545, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 387, + 621, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 621, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 387, + 621, + 545, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r} + \\lambda_ {m} \\mathcal {L} _ {m}. \\tag {10}", + "image_path": "5d41f43897ba1ccd38ad4e0fe214112ff74492ac902df473271da3a084e69429.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "text", + "content": " to control the sparsity of the parameters of the grid representation." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 671, + 430, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 430, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 430, + 684 + ], + "type": "text", + "content": "3.4. Compression pipeline" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "With our proposed masking method and multi-level DWT, the ratio of zeros in grids can go up to more than" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20684" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 289, + 173 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 173 + ], + "type": "image", + "image_path": "10b955614e42106c4b08fca6b6ef434184514211e81702172d5f6d94a0eb9f82.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 191, + 288, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 288, + 226 + ], + "type": "text", + "content": "Figure 4. Proposed bitmap compression pipeline. RLE and Huffman denote run-length encoding and Huffman encoding, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 246, + 287, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 377 + ], + "type": "text", + "content": "95%. However, sparse representations themselves do not reduce the total size. In this section, we describe our proposed compression pipeline for sparse grid parameters. Instead of storing grids as they are, we separately store nonzero coefficients and bitmaps (or masks) that indicate which coefficients are non-zero. Despite using 1-bit bitmaps, the overall bitmap size is large due to the large number of parameters. To reduce the bitmap sizes, we propose a compression pipeline with the following three stages: n-bit casting, run-length encoding (RLE), and the Huffman encoding [19]. The compression pipeline is illustrated in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 378, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 288, + 568 + ], + "type": "text", + "content": "Before applying the compression pipeline, we split wavelet coefficient masks by level of decomposition. For the 2-level wavelet transform, for instance, we group HL1, LH1, and HH1, then HL2, LH2, and HH2, and finally LL2 (Fig. 3). This is based on our experimental results that wavelet coefficients with the same levels of sparsity have similar levels of sparsity (Fig. 6). We found that grouping coefficients with similar sparsity results in better compression performance. In addition, we found that directly applying RLE to 1-bit streams causes inefficient bit allocation due to the numerous repeating zeros. Therefore, we first cast the binary mask values to n-bit unsigned integers, and perform RLE afterward. In our experiments, we set n to 8. Finally, we apply the Huffman encoding algorithm to the RLE-encoded streams to map values with a high probability to shorter bits." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 579, + 128, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 128, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 128, + 594 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 172, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 172, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 172, + 613 + ], + "type": "text", + "content": "4.1. Experimental settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": "We compared our proposed method with other baselines in four datasets; NeRF synthetic dataset [23], Neural Sparse Voxel Fields (NSVF) [22] dataset, TanksAndTemps dataset [20], and LLFF dataset [10]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "We used TensoRF VM-192 as the baseline model and applied the proposed method. To control the size of our method, we adjusted the mask regularization weight " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " the number of grid channels (from VM-192 to VM-384)," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "and the resolution of grids. Unless otherwise specified, we used 8-bit quantization for every method. We compared our method with quantized NeRF, TensoRF models (CP and VM), cNeRF [2] and CCNeRF [42]. We also tried to compare our approach with VQAD [38], which efficiently compresses 3D scene representations using codebooks and vector quantization. However, VQAD requires depth maps to prune out vacant areas in advance. Otherwise, the required memory exceeds the memory capacity of the Tesla A100 equipped with 40 GB of memory. Since the datasets we used do not provide depth maps, we did not compare VQAD with ours." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "type": "text", + "content": "We followed the experimental settings of TensoRF [4]. We trained models for 30,000 iterations, each of which is a minibatch of 4,096 rays. We used the Adam optimizer and an exponential learning rate decay scheduler. Following TensoRF, the initial learning rates of grid-related parameters and MLP-related parameters were set to 0.02 and 0.001, respectively. Final learning rates were set at 1/10 of the initial learning rate. We updated the alpha masks at the 2000th, 4000th, 6000th, 11000th, 16000th, 21000th, and 26000th iterations. We set " + }, + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "type": "text", + "content": " in Eq. (10) to 1e-10 for high parameter sparsity and 5e-11 and 2.5e-11 for relatively lower sparsity. We set the initial values of masks " + }, + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "type": "text", + "content": " to one, and set their initial learning rate to the same learning rate as grid parameters " + }, + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 217, + 546, + 396 + ], + "type": "text", + "content": ". As a reconstruction quality measurement, we used PSNR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 405, + 362, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 362, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 362, + 417 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 425, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 472 + ], + "type": "text", + "content": "Figs. 1 and 5 show the quantitative performances of our method and the baselines on various datasets. Each graph displays the average PSNR and size of methods. More detailed numbers are available in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "text", + "content": "As the rate-distortion curves show, our method achieves higher efficiency than the baseline models and other methods. Even with the base network (VM-192), our method consistently outperforms other baselines on all novel view synthesis datasets, under a small memory budget of " + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "inline_equation", + "content": "2\\mathrm{MB}" + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "text", + "content": ". We adjusted both " + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "text", + "content": " and the network size to increase the output size. By adjusting " + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "inline_equation", + "content": "\\text{lambda}_{m}" + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "text", + "content": ", we can control both size and quality while having no effect on computational costs, time, or memory requirements. Increasing network size (doubling the number of channels and increasing the resolution) can push the boundary even further. It is also worth mentioning that applying our method to a larger backbone model (VM-384) outperforms models with a smaller backbone model (VM-192), even with the same value of " + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 304, + 474, + 546, + 688 + ], + "type": "text", + "content": " or a similar model size. However, as the network grows in size, the computational costs and time required for training and testing increase. Therefore, it is still reasonable to use a small backbone model when using our proposed method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Fig. 7 shows the qualitative results with and without our proposed masking method. Trainable masks removed more" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20685" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 218, + 194 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 218, + 194 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 218, + 194 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 218, + 194 + ], + "type": "image", + "image_path": "4d9d79b8ea73f76a1eaf5fd70eb08c89818029238dbc68514cceb8d202447e47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 201, + 544, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 201, + 544, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 544, + 213 + ], + "type": "text", + "content": "Figure 5. Rate distortion curves on the NSVF synthetic, TanksAndTemples, and LLFF dataset. We used 8-bit precision for every method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 72, + 375, + 194 + ], + "blocks": [ + { + "bbox": [ + 220, + 72, + 375, + 194 + ], + "lines": [ + { + "bbox": [ + 220, + 72, + 375, + 194 + ], + "spans": [ + { + "bbox": [ + 220, + 72, + 375, + 194 + ], + "type": "image", + "image_path": "25bcb6b6b6bbee4b92d2c296d685755351782de0d5bcbd51383be3f8ce81e2b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 72, + 531, + 194 + ], + "blocks": [ + { + "bbox": [ + 376, + 72, + 531, + 194 + ], + "lines": [ + { + "bbox": [ + 376, + 72, + 531, + 194 + ], + "spans": [ + { + "bbox": [ + 376, + 72, + 531, + 194 + ], + "type": "image", + "image_path": "e50f7c0d0a55485ef572a64b096b41516b98767db7ce05aeeaab61b223533c74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 232, + 156, + 335 + ], + "blocks": [ + { + "bbox": [ + 60, + 232, + 156, + 335 + ], + "lines": [ + { + "bbox": [ + 60, + 232, + 156, + 335 + ], + "spans": [ + { + "bbox": [ + 60, + 232, + 156, + 335 + ], + "type": "image", + "image_path": "8609fa5d6a211f7a711d10dc4fa5086413051cc09548871985880c858dcb7ca5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 338, + 140, + 346 + ], + "lines": [ + { + "bbox": [ + 95, + 338, + 140, + 346 + ], + "spans": [ + { + "bbox": [ + 95, + 338, + 140, + 346 + ], + "type": "text", + "content": "(a) Raw mask" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 357, + 287, + 380 + ], + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 380 + ], + "type": "text", + "content": "Figure 6. Qualitative analysis on mask " + }, + { + "bbox": [ + 47, + 357, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 357, + 287, + 380 + ], + "type": "text", + "content": " (Chair in the NeRF synthetic dataset)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 157, + 232, + 276, + 335 + ], + "blocks": [ + { + "bbox": [ + 157, + 232, + 276, + 335 + ], + "lines": [ + { + "bbox": [ + 157, + 232, + 276, + 335 + ], + "spans": [ + { + "bbox": [ + 157, + 232, + 276, + 335 + ], + "type": "image", + "image_path": "2067b199aa78fb46d8bb225d83e59f7cb9868a841cc091847681d9a2f74bc6b3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 338, + 258, + 346 + ], + "lines": [ + { + "bbox": [ + 195, + 338, + 258, + 346 + ], + "spans": [ + { + "bbox": [ + 195, + 338, + 258, + 346 + ], + "type": "text", + "content": "(b) Binarized mask" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 411, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 471 + ], + "type": "text", + "content": "than " + }, + { + "bbox": [ + 46, + 411, + 287, + 471 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 46, + 411, + 287, + 471 + ], + "type": "text", + "content": " of the grid parameters; nevertheless, the rendered results are still accurate, and the qualitative difference is almost imperceptible. This demonstrates that our proposed method efficiently eliminates unnecessary wavelet coefficients by leveraging trainable masks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 475, + 288, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 288, + 596 + ], + "type": "text", + "content": "Fig. 6 illustrates both raw and binarized masks after training. First of all, sparsity varies depending on the level of the wavelet transform. As shown in Fig. 6b, coefficients with lower frequencies have lower sparsity, whereas coefficients with higher frequencies have higher sparsity. What is interesting is that the raw mask values seem to reflect the characteristics of corresponding wavelet coefficients. Vertical, horizontal, and diagonal patterns can be found in the raw mask values for the vertical, horizontal, and diagonal coefficients, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 618, + 145, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 145, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 145, + 629 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "In this section, we analyze each component of our proposed method. Every experiment was conducted on the whole NeRF synthetic dataset, and we used the average PSNR as a measurement for representation quality. We used 4-level wavelet transform and set " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " to 1e-10, unless otherwise specified." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 342, + 231, + 511, + 285 + ], + "blocks": [ + { + "bbox": [ + 342, + 231, + 511, + 285 + ], + "lines": [ + { + "bbox": [ + 342, + 231, + 511, + 285 + ], + "spans": [ + { + "bbox": [ + 342, + 231, + 511, + 285 + ], + "type": "table", + "html": "
MethodsPSNR
DWT31.949
DWT w/o level-wise scaling31.145
DCT31.325
", + "image_path": "9a97d65edb55746a3cb2ce9922e464730f5edf6cb7776c0a305694458175ecc8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 293, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 326 + ], + "type": "text", + "content": "Table 1. Ablation study on level-wise scaling. Without our proposed scaling method, the performance of DWT is worse than DCT." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 348, + 454, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 454, + 359 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 454, + 359 + ], + "type": "text", + "content": "4.3.1 Discrete wavelet transform" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 367, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 488 + ], + "type": "text", + "content": "In this section, we evaluate the compactness of wavelet coefficients and the performance improvements brought by the multi-level wavelet transform. Fig. 8 shows the rate-distortion curves of four different levels of wavelet transform and non-wavelet coefficients. As shown in the figure, wavelet-based grid representation outperforms the grid in the spatial domain, especially when the sparsity is high. Furthermore, rate-distortion curves demonstrate that higher level of wavelet transform improves visual quality and sparsity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 502, + 496, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 496, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 496, + 516 + ], + "type": "text", + "content": "4.3.2 Level-wise wavelet coefficient scaling" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 522, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 606 + ], + "type": "text", + "content": "As mentioned in Sec. 3.2.1, we proposed multiplying wavelet coefficients by the scales proportional to the inverse of frequency. Thus, we compared two versions of wavelet transform; one with our proposed weight scaling (Sec. 3.2.1) and the other without it. As Tab. 1 shows, scaling wavelet coefficients improves reconstruction performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 622, + 416, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 416, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 416, + 633 + ], + "type": "text", + "content": "4.3.3 Wavelet functions" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "In this section, we analyze how different wavelet functions affect reconstruction quality. For comparison, we selected several wavelet functions: Haar, Coiflet 1, Daubechies 4, and reverse biorthogonal 4.4. As shown in Tab. 2, the type of wavelet function has little effect on reconstruction quality. Even Haar, the simplest wavelet function, performs" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20686" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 70, + 312, + 182 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 312, + 182 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 312, + 182 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 312, + 182 + ], + "type": "image", + "image_path": "763c28005c70c802cd1afc5bcb97d18903563016fbd1e36ffc828e36abcba959.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 546, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 546, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 546, + 226 + ], + "type": "text", + "content": "Figure 7. A qualitative analysis of grid sparsity. The sparsity refers to the ratio of zero elements in grids. With the help of trainable masks, removing more than " + }, + { + "bbox": [ + 46, + 202, + 546, + 226 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 46, + 202, + 546, + 226 + ], + "type": "text", + "content": " of grid parameters only results in an imperceptible difference." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 504, + 179 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 504, + 179 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 504, + 179 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 504, + 179 + ], + "type": "image", + "image_path": "971911b1e1283ea2eb744dc52f019a7cb5b3bb6c16faa846332575c31ae3a86a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 251, + 282, + 404 + ], + "blocks": [ + { + "bbox": [ + 52, + 251, + 282, + 404 + ], + "lines": [ + { + "bbox": [ + 52, + 251, + 282, + 404 + ], + "spans": [ + { + "bbox": [ + 52, + 251, + 282, + 404 + ], + "type": "image", + "image_path": "b7879939e62fbc4841de393219326c0d9f9dfe9087abd4ccf68f03155f704e6e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 416, + 287, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 472 + ], + "type": "text", + "content": "Figure 8. The rate distortion curves of different signal representation schemes (spatial, DWT, and DCT). Sparsity on the " + }, + { + "bbox": [ + 46, + 416, + 287, + 472 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 416, + 287, + 472 + ], + "type": "text", + "content": " axis refers to the ratio of zeros in grid parameters. The grid sparsity was controlled by " + }, + { + "bbox": [ + 46, + 416, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 46, + 416, + 287, + 472 + ], + "type": "text", + "content": ". The numbers inside the parenthesis indicate the levels of the wavelet transform." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 85, + 483, + 250, + 563 + ], + "blocks": [ + { + "bbox": [ + 85, + 483, + 250, + 563 + ], + "lines": [ + { + "bbox": [ + 85, + 483, + 250, + 563 + ], + "spans": [ + { + "bbox": [ + 85, + 483, + 250, + 563 + ], + "type": "table", + "html": "
Wavelet FunctionPSNR
Haar31.889
Coiflets 131.846
Daubechies 431.734
biorthogonal 4.4 (default)31.949
reverse biorthogonal 4.431.727
", + "image_path": "4d06dd13213f24f24eb0fa013a931f48fe62ba19d605f1071b64d6efa4ce29bf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 570, + 287, + 592 + ], + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 592 + ], + "type": "text", + "content": "Table 2. Reconstruction performances of different wavelet functions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 615, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 287, + 640 + ], + "type": "text", + "content": "fairly well. Still, the biorthogonal 4.4 that we used shows the best performance among selected wavelet functions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 189, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 189, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 189, + 669 + ], + "type": "text", + "content": "4.3.4 Discrete cosine transform" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "In this section, we compare the performance of DWT with that of DCT, which represents a signal with respect to the sum of cosine functions with different frequencies. DCT," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 246, + 546, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 246, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 546, + 354 + ], + "type": "text", + "content": "similar to DWT, has an energy compaction property. That is, a small number of non-zero DCT coefficients are sufficient to represent a given signal. Because of its energy compactness, DCT or its variants are another widely used signal transformation for image, audio, and audio compression. However, as shown in Fig. 8 and Tab. 1, DWT shows superior compression and representation performance to DCT. We believe this is due to the non-repeating, non-smooth information in grids, for which DWT is more appropriate." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 367, + 378, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 378, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 378, + 379 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "content": "We propose a compact representation for grid-based neural fields, enabled by a novel masking strategy and multilevel wavelet transform. We demonstrate that these two components made it possible to remove more than " + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "content": " of grid parameters without a significant loss of visual quality. With our proposed compression pipeline and the sparse wavelet coefficients, we achieved state-of-the-art performance under a memory budget of 2 MB. We believe we can reduce the size further with more sophisticated compression pipelines." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 508, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 592 + ], + "type": "text", + "content": "Our proposed method is naturally constrained by the limitations of grid-based neural fields that are developed for bounded scenes or objects. We believe that expanding this grid-based representation to encompass unbounded or large scenes would be an intriguing direction, given that we can now compactly represent 3D objects with our proposed representation scheme." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 605, + 404, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 404, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 404, + 618 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 624, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 547, + 713 + ], + "type": "text", + "content": "This work was supported in part by the Institute of Information and Communications Technology Planning and Evaluation (IITP) grants (2021-0-02068, IITP-2021-0-02052, IITP-2023-2020-0-01821, IITP-2019-0-00421) and National Research Foundation of Korea (NRF) grants (2022R1A4A3032913, 2022R1F1A1064184, 2022R1A4A3033571) funded by the Korea Government (MSIT)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20687" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "type": "text", + "content": "[2] Thomas Bird, Johannes Balle, Saurabh Singh, and Philip A Chou. 3d scene compression through entropy penalized neural representation functions. In 2021 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 270 + ], + "type": "text", + "content": "[3] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, June 2022. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 305 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision. Springer, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "text", + "content": "[6] Albert Cohen, Ingrid Daubechies, and J-C Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on pure and applied mathematics, 45(5):485-560, 1992. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 474 + ], + "type": "text", + "content": "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 476, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 287, + 509 + ], + "type": "text", + "content": "[8] Chenxi Lola Deng and Enzo Tartaglione. Compressing explicit voxel grid representations: fast nerfs become also small. arXiv preprint arXiv:2210.12782, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "type": "text", + "content": "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "text", + "content": "[10] Ben Mildenhall et al. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 668 + ], + "type": "text", + "content": "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, SA '22, New York, NY, USA, 2022. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "text", + "content": "[12] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "volume 70 of Proceedings of Machine Learning Research, pages 1126-1135. PMLR, 06-11 Aug 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 97, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 547, + 152 + ], + "type": "text", + "content": "[13] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, June 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 547, + 206 + ], + "type": "text", + "content": "[14] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14346-14355, October 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "text", + "content": "[15] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. pages 5875-5884, October 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "text", + "content": "[16] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12902-12911, June 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[17] Binbin Huang, Xinhao Yan, Anpei Chen, Shenghua Gao, and Jingyi Yu. Pref: Phasorial embedding fields for compact neural representations. arXiv preprint arXiv:2205.13524, 2022. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "type": "text", + "content": "[18] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Binarized neural networks. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "type": "text", + "content": "[19] David A Huffman. A method for the construction of minimum-redundancy codes. Proceedings of the IRE, 40(9):1098-1101, 1952. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 477 + ], + "type": "text", + "content": "[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), jul 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "type": "text", + "content": "[21] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 523, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 588 + ], + "type": "text", + "content": "[22] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 15651-15663. Curran Associates, Inc., 2020. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 590, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 645 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), jul 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[25] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? Adaptive" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20688" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "text", + "content": "rounding for post-training quantization. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7197-7206. PMLR, 13-18 Jul 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 160 + ], + "type": "text", + "content": "[26] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 194 + ], + "type": "text", + "content": "[27] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. pages 18419-18429, June 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 288, + 270 + ], + "type": "text", + "content": "[28] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 13032-13044. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "text", + "content": "[29] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 381 + ], + "type": "text", + "content": "[30] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision – ECCV 2016, pages 525–542, Cham, 2016. Springer International Publishing. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "text", + "content": "[31] Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, and Andrea Tagliasacchi. Derf: Decomposed radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14153-14161, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 491 + ], + "type": "text", + "content": "[32] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14335-14345, October 2021. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 547 + ], + "type": "text", + "content": "[33] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. VoxGRAF: Fast 3d-aware image synthesis with sparse voxel grids. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 613 + ], + "type": "text", + "content": "[34] Vincent Sitzmann, Eric Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 10136-10147. Curran Associates, Inc., 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 647 + ], + "type": "text", + "content": "[35] Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The jpeg 2000 still image compression standard. IEEE Signal processing magazine, 18(5):36-58, 2001. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "text", + "content": "[36] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[37] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, June 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 162 + ], + "type": "text", + "content": "[38] Towaki Takikawa, Alex Evans, Jonathan Tremblay, Thomas Müller, Morgan McGuire, Alec Jacobson, and Sanja Fidler. Variable bitrate neural fields. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 217 + ], + "type": "text", + "content": "[39] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. arXiv preprint arXiv:2101.10994, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "type": "text", + "content": "[40] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P. Srinivasan, Jonathan T. Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 261, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 327 + ], + "type": "text", + "content": "[41] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "text", + "content": "[42] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable neRF via rank-residual decomposition. In Advances in Neural Information Processing Systems, 2022. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 372, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 393 + ], + "type": "text", + "content": "[43] Gregory K. Wallace. TheJPEG still picture compression standard.Commun.ACM,34(4):30-44,apr 1991.2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 394, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 470 + ], + "type": "text", + "content": "[44] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 612-629, Cham, 2022. Springer Nature Switzerland. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 471, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 536 + ], + "type": "text", + "content": "[45] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13524-13534, June 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "type": "text", + "content": "[46] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology, 13(7):560-576, 2003. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 582, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 603 + ], + "type": "text", + "content": "[47] Zhijie Wu, Yuhe Jin, and Kwang Moo Yi. Neural fourier filter bank. arXiv e-prints, pages arXiv-2212, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 545, + 658 + ], + "type": "text", + "content": "[48] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[49] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5752-5761, October 2021. 2, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20689" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 150 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[50] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. arXiv preprint arXiv:2205.14330, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "text", + "content": "[51] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained ternary quantization. In International Conference on Learning Representations, 2017. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20690" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_content_list.json b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bedaaf6c7219efd060409e3e4b17062fe40b5af4 --- /dev/null +++ b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_content_list.json @@ -0,0 +1,1700 @@ +[ + { + "type": "text", + "text": "Masked and Adaptive Transformer for Exemplar Based Image Translation", + "text_level": 1, + "bbox": [ + 107, + 130, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chang Jiang $^{1}$ , Fei Gao $^{1,2*}$ , Biao Ma $^{1}$ , Yuhao Lin $^{1}$ , Nannan Wang $^{3}$ , Gang Xu $^{1}$", + "bbox": [ + 183, + 179, + 790, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Computer Science and Technology, Hangzhou Dianzi University", + "bbox": [ + 183, + 198, + 789, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Hangzhou Institute of Technology, Xidian University $^{3}$ ISN State Key Laboratory, Xidian University", + "bbox": [ + 86, + 215, + 888, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jc233, aiartma, linyh, gxu}@hdu.edu.cn, {fgao, nnwang}@xidian.edu.cn", + "bbox": [ + 181, + 234, + 789, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a novel framework for exemplar based image translation. Recent advanced methods for this task mainly focus on establishing cross-domain semantic correspondence, which sequentially dominates image generation in the manner of local style control. Unfortunately, cross-domain semantic matching is challenging; and matching errors ultimately degrade the quality of generated images. To overcome this challenge, we improve the accuracy of matching on the one hand, and diminish the role of matching in image generation on the other hand. To achieve the former, we propose a masked and adaptive transformer (MAT) for learning accurate cross-domain correspondence, and executing context-aware feature augmentation. To achieve the latter, we use source features of the input and global style codes of the exemplar, as supplementary information, for decoding an image. Besides, we devise a novel contrastive style learning method, for acquire quality-discriminative style representations, which in turn benefit high-quality image generation. Experimental results show that our method, dubbed MATEBIT, performs considerably better than state-of-the-art methods, in diverse image translation tasks. The codes are available at https://github.com/AiArt-HDU/MATEBIT.", + "bbox": [ + 75, + 319, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-to-image translation aims at transfer images in a source domain to a target domain [16, 50]. Early studies learn mappings directly by Generating Adversarial Networks (GANs), and have shown great success in various applications [2, 42]. Recently, exemplar based image translation [29, 30, 45], where an exemplar image is used to control the style of translated images, has attracted a lot of attention. Such methods allow high flexibility and controllability, and have a wide range of potential applications in social networks and metaverse. For example, people can transfer a", + "bbox": [ + 75, + 723, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73cb608b03c9debbdc14df50dd542ab2a27abb3dfaed97834c96cb8fed1dec58.jpg", + "image_caption": [ + "Figure 1. Visualization of correspondence maps. The red point is the query position. Full Corr. and Masked Corr. denote the full correspondence [45] and masked one in our method, respectively. CAM denotes visualization by Class Activation Mapping [48]." + ], + "image_footnote": [], + "bbox": [ + 522, + 287, + 869, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "facial sketch to an artistic portrait, in the style of oil paintings or avatars. Despite the remarkable progress, yielding high-fidelity images with consistent semantic and faithful styles remains a grand challenge.", + "bbox": [ + 496, + 546, + 893, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early pioneering works [15, 21, 35] attempt to globally control the style of generated images. However, such methods ignore spatial correlations between an input image and an exemplar, and may fail to produce faithful details. Recently, some advanced methods [25, 44, 45, 49] first establish the cross-domain semantic correspondence between an input image and an exemplar, and then use it to warp the exemplar for controlling local style patterns. In these methods, the quality of generated images relies heavily on the learned correspondence [39]. Unfortunately, cross-domain semantic matching is challenging, since there is no reliable supervision on correspondence learning [45]. As a result, potential matching errors ultimately lead to degraded artifacts in generated images.", + "bbox": [ + 496, + 609, + 893, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To combat this challenge, we propose to boost the matching accuracy on one hand, and to diminish the role of matching in image generation on the other hand. Inspired by the great success of Transformers [6, 10, 26, 41], we first devise a Masked and Adaptive Transformer (MAT) for learning ac", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 101, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22418", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "curate cross-domain correspondence and executing context-aware feature augmentation. Previous works [44, 45, 49] have used the vanilla attention mechanism [41] for learning full correspondence. However, the initial attention typically involves ambiguous correspondences (2nd row in Fig. 1). To mitigate these limitations, in MAT, we use a masked attention to distinguish the correspondence as reliable or not, and then reliability-adaptively aggregate representations. Besides, the Feed-Forward Network (FFN) [41] in vanilla transformers neglects contextual correlations inside an image. We thus replace FFN by an adaptive convolution block [28], where the coordinate attention [12] and depth-wise separable convolution [5] are used to capture contextual correlations and to improve efficiency. With a joint consideration of matching reliability and contextual correlations, MAT gradually focuses on accurate correspondences and emphasizes on features of interest (3rd row in Fig. 1).", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, to boost both the semantic consistency and style faithfulness, we supplementally use semantic features of the input image and global style codes of the exemplar for decoding an image. To this end, we first design our whole network following the U-Net architecture [16]. Besides, we devise a novel contrastive style learning (CSL) framework for acquiring discriminative style representations. Recently, Zhang et al. [47] propose a similar CSL method, where the target exemplar is used as a positive sample, and the other exemplars as negative ones. Differently, we use low-quality images, generated during early training stages, as negative samples. In this way, our style codes are desired to discriminate not only subtle differences in style, but also those in perceptual quality. Ultimately, the learned global style codes, cooperating with the local style control induced by MAT, in turn benefit high-quality image generation.", + "bbox": [ + 75, + 347, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the proposed techniques above, our full model, dubbed MATEBIT, diminishes the impact of position-wise matching on image quality, and integrates both local and global style control for image generation. Experimental results show that MATEBIT generates considerably more plausible images than previous state-of-the-art methods, in diverse image translation tasks. In addition, comprehensive ablation studies demonstrate the effectiveness of our proposed components. Finally, we perform interesting applications of photo-to-painting translation and Chinese ink paintings generation.", + "bbox": [ + 75, + 589, + 472, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Relate Work", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Exemplar Based Image Translation. Recently, exemplar based image translation has attracted increasing attention. For example, Park et al. [35] learn an encoder to map the exemplar image into a global style vector, and use it to guide image generation. Such a global style control strategy enables style consistency in whole, but fails to produce subtle details. Most recently, researchers propose a", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "matching-then-generation framework [39]. Specially, they first establish dense correspondence between an input and an exemplar, and then reshuffle the exemplar for locally control the style of synthesize images. For example, Zhang et al. [45] establish position-wise correspondence based on the Cosine attention mechanism and warp the exemplar correspondingly. Afterwards, the warped image dominates the generation of images in the manner of SPADE [35]. To reduce the cost of matching in high-resolution image generation, Zhou et al. [49] introduce a hierarchical refinement of semantic correspondence from ConvGRU-PatchMatch. Besides, Liu et al. [25] used a dynamic pruning method for learning hierarchical sparse correspondence. They also use reliability-adaptive feature integration to improve the quality of generated images.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous methods merely use global or local style control, and the latter relies heavily on the learned correspondence. Besides, they consider little about contextual correlations inside an image. In this paper, we use both global and local style control to boost the style consistency. Besides, we take contextual correlations into consideration and execute reliability-adaptive feature augmentation.", + "bbox": [ + 496, + 320, + 893, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers. Transformers [41] have shown incredible success from the field of natural language processing (NLP) [19] to computer vision (CV) [6, 26]. Multihead attention (MHA) and FFN are key components in a Transformer, and have been used in exemplar based image translation. However, they induce unreliable matching results and neglect context correlations in feature translation. In our MAT, we combat these limitations by replacing them with a masked attention and a context-aware convolution block, respectively. Recently, researchers use semantic masks to facilitate representation learning [4, 8, 37], where a mask predictor is required. Differently, we use a ReLU function to mask over the attention layer, for distinguishing correspondence as reliable or not (Sec. 3.1). In general, MAT follows a concise and efficient architecture.", + "bbox": [ + 496, + 429, + 895, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Learning. Contrastive learning has shown its effectiveness in various computer vision tasks [9, 13, 34]. The basic idea is to learn a representation by pushing positive samples toward an anchor, and moving negative samples away from it. Different sampling strategies and contrastive losses have been extensively explored in various downstream tasks. For example, Chen et al. [3] and He et al. [9] obtain positive samples by augmenting original data. In the field of image translation, Park et al. [34] propose patchwise contrastive learning by maximizing the mutual information between cross-domain patches. Similarly, Zhang et al. [47] use contrastive learning for acquiring discriminative style representations. In the task of exemplar based image translation, Zhan et al. [44] use contrastive learning to align cross-domain images to a consistent semantic feature space, so as to boost the accuracy of matching. Differently, we", + "bbox": [ + 496, + 659, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22419", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/87cb6bd4342918ba51234ed1a6d836c2b00741bcfcd42f8778a9cb1df09b4e05.jpg", + "image_caption": [ + "Figure 2. Overview of our image translation network, MATEBIT." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "use early generated images as negative samples, so that the learned style representations can discriminate subtle differences in both style and perceptual quality (Sec. 3.2).", + "bbox": [ + 76, + 273, + 468, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 337, + 287, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $x_{A}$ in domain $\\mathcal{A}$ and an exemplar image $y_{B}$ in domain $\\mathcal{B}$ , our goal is to generate a target image $x_{B}$ which preserves semantic structures in $x_{A}$ but resembles the style of similar parts in $y_{B}$ . Fig. 2 shows an overview of our translation network $\\mathcal{G}$ . Specially, we first align $x_{A}$ and $y_{B}$ to an intermediate feature space by encoders $\\mathcal{E}_{A}$ and $\\mathcal{E}_{B}$ , respectively. Afterwards, we use a Masked and Adaptive Transformer (MAT) for correspondence learning and feature augmentation. Finally, a decoder $\\mathcal{D}_B$ produces an output image $\\hat{x}_B$ based on the augmented features, as well as the source features and target style codes. Details are described below.", + "bbox": [ + 75, + 363, + 472, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Masked and Adaptive Transformer (MAT)", + "text_level": 1, + "bbox": [ + 76, + 556, + 441, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to establish accurate cross-domain correspondence, we propose a novel and concise Transformer architecture, i.e. MAT. In general, the architecture of MAT (Fig. 3b) follows that of vanilla Transformers (Fig. 3a) [41]. Differently, we use masked attention to distinguish reliable and unreliable correspondence, instead of using multi-head attention. Besides, we use Positional Normalization (PONO) [23] and an Adaptive Convolution (AdaConv) block [28], instead of LN and MLP-based FFN, respectively. MAT is desired to gradually concentrate on accurate matching, and to reliability-adaptively augment representations with contextual correlations.", + "bbox": [ + 75, + 582, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked Correspondence Learning. Let $\\mathbf{X}_A \\in \\mathbb{R}^{H \\times W \\times C}$ and $\\mathbf{Y}_B \\in \\mathbb{R}^{H \\times W \\times C}$ be the representations of $x_A$ and $y_B$ in the intermediate feature space, with height $H$ , width $W$ , and $C$ channels. We first map $\\mathbf{X}_A$ to the query $\\mathbf{Q} \\in \\mathbb{R}^{HW \\times C}$ , and $\\mathbf{Y}_B$ to the key $\\mathbf{K} \\in \\mathbb{R}^{HW \\times C}$ and value $\\mathbf{V} \\in \\mathbb{R}^{HW \\times C}$ , by using $1 \\times 1$ convolutions, respectively. As shown in Fig. 3d, we add positional encoding (PE) to $\\mathbf{X}_A$ and $\\mathbf{Y}_B$ , for embedding spatial correlations. Afterwards, we learn the initial correspondence $\\mathbf{A} \\in \\mathbb{R}^{HW \\times HW}$ fol", + "bbox": [ + 76, + 763, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de343606c827cbfaccca0f7d431dccf574b28466ed6aa0ab5a2d4facc7f444d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 94, + 625, + 316 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/16099285aba2a1f5025a9aa51abd3a44a96f4e9e4d85a700a372bfd3c1377d79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 94, + 750, + 316 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee95e5984517f9afcf7e7686b851fc64c55b65c7ed5d4fbbfb4862faea5a3984.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 95, + 867, + 316 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cce8d1a41fd26826f80744de106158f1377f8e06284d1222f113314602c08c06.jpg", + "image_caption": [ + "Figure 3. Detailed architectures. (a) Vanilla Transformer block, (b) MAT block, (c) AdaConv block, and (d) Masked Attention." + ], + "image_footnote": [], + "bbox": [ + 524, + 318, + 867, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lowing the Cosine attention mechanism [45], i.e.", + "bbox": [ + 498, + 553, + 821, + 568 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} (u, v) = \\frac {\\tilde {\\mathbf {Q}} (u) \\tilde {\\mathbf {K}} (v) ^ {T}}{| | \\tilde {\\mathbf {Q}} (u) | | \\cdot | | \\tilde {\\mathbf {K}} (v) | |}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 578, + 890, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\tilde{\\mathbf{Q}}(u) = \\mathbf{Q}(u) - \\bar{\\mathbf{Q}}(u)$ , $\\tilde{\\mathbf{K}}(v) = \\mathbf{K}(v) - \\bar{\\mathbf{K}}(v)$ , where $u, v \\in [1, \\dots, HW]$ are position indices; $\\bar{\\mathbf{Q}}(u)$ and $\\bar{\\mathbf{K}}(v)$ are the means of $\\mathbf{Q}(u)$ and $\\mathbf{K}(v)$ , respectively. $\\mathbf{A}(u, v)$ is the matching score between $\\mathbf{Q}(u)$ and $\\mathbf{K}(v)$ .", + "bbox": [ + 498, + 625, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous methods [44, 45] typically use the initial correspondence map $\\mathbf{A}$ to reshuffle an exemplar for controlling local patterns in image synthesis. However, induced by the difficulties in cross-domain correspondence learning, $\\mathbf{A}$ involves unreliable match scores (Fig. 3d). As a result, the reshuffled image will lead to implausible artifacts in generated images. To combat this limitation, we distinguish initial matching scores as reliable or not, according to their signs [32]. The masked correspondence map becomes:", + "bbox": [ + 496, + 688, + 892, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {\\text {m a s k}} = \\operatorname {R e L U} (\\mathbf {A}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 835, + 890, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In DynaST [25], two networks are used to predict the reliability mask of correspondence. However, it's challenging to effectively train the network, because there is no super", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22420", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vision on matching during training. In contrast, ReLU contains no learnable parameters and ultimately leads to superior performance over DynaST (Sec. 4.1).", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reliability-Adaptive Feature Aggregation. For regions with reliable correspondence in $x_{A}$ , we use $\\mathbf{A}_{mask}$ to warp the value features, $\\mathbf{V}$ , derived from the exemplar:", + "bbox": [ + 76, + 136, + 467, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\text {c o r}} = \\tilde {\\mathbf {A}} _ {\\text {m a s k}} \\mathbf {V}, \\text {w i t h} \\tilde {\\mathbf {A}} _ {\\text {m a s k}} = \\operatorname {s o f t m a x} (\\alpha \\cdot \\mathbf {A} _ {\\text {m a s k}}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 189, + 467, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scaling coefficient to control the sharpness of the softmax function. In default, we set its value as 100.", + "bbox": [ + 76, + 222, + 467, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For regions with unreliable correspondence in $x_{A}$ , $\\mathbf{X}_{cor}$ provides an average style representation of $\\mathbf{V}$ . We further extract complementary information from the query, $\\mathbf{Q}$ , derived from the input. Inspired by SPADE [35], we first transfer $\\mathbf{Q}$ to the target domain by using pixel-wise modulation parameters (i.e., $\\gamma$ for scale and $\\beta$ for bias) learned from $x_{A}$ . The modulation is formulated by:", + "bbox": [ + 75, + 252, + 467, + 357 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {\\text {n o r m}} = \\gamma (x _ {A}) \\frac {\\mathbf {Q} - \\mu (\\mathbf {Q})}{\\sigma (\\mathbf {Q})} + \\beta (x _ {A}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 366, + 468, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu (\\mathbf{Q})$ and $\\sigma (\\mathbf{Q})$ are the mean value and standard deviance of $\\mathbf{Q}$ . Afterwards, we select the translated features of unreliable corresponded regions in $x_{A}$ by:", + "bbox": [ + 76, + 407, + 467, + 453 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {u n c o r} = \\left(1 - \\sum_ {j} \\mathbf {A} _ {m a s k}\\right) \\odot \\mathbf {Q} _ {n o r m}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 462, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the summation is along the second dimension; $\\odot$ denotes point-wise production with broadcasting. Since $\\gamma$ and $\\beta$ are learned from the input image $x_{A}$ , the modulated features preserve the semantic information of $x_{A}$ . Besides, constraints on the generated image will push the selected features convey to the style of $y_{B}$ .", + "bbox": [ + 76, + 488, + 467, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ideally, $\\mathbf{X}_{cor}$ and $\\mathbf{X}_{uncor}$ would complement each other and facilitate both semantic consistency and style relevance in image generation. To this end, we integrate $\\mathbf{X}_{cor}$ , $\\mathbf{X}_{uncor}$ , and $\\mathbf{Q}$ by:", + "bbox": [ + 76, + 578, + 467, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {a g g} = \\operatorname {P O N O} \\left(\\mathbf {X} _ {c o r} + \\mathbf {X} _ {u n c o r} + \\mathbf {Q}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 648, + 467, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In PONO [23], features at each position are normalized dependently. Compared to LN in vanilla transformers and DynaST [25], PONO boosts the flexibility in reliability-adaptive feature aggregation.", + "bbox": [ + 76, + 672, + 467, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Context-Aware Feature Augmentation. Inspired by ConvNeXT [28], we replace FFN by an AdaConv block to position-adaptively emphasize informative representations. Besides, we use the coordinate attention (CoordAten) module [12] to capture contextual correlations.", + "bbox": [ + 76, + 734, + 467, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of the AdaConv block is as shown in Fig. 3c. We fist use the depthwise convolution (Dwise) to update representations in each channel separately; and then use two pointwise convolutions (Pwise) to automatically emphasize representations of interest, at every position. The Gaussian Error Linear Unit (GELU) activation", + "bbox": [ + 75, + 810, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/92a1c60f1844b473bacc6475f76c30344beaf7bd2d8d2b1c76a8e42dcfd3c97f.jpg", + "image_caption": [ + "Figure 4. Contrastive style learning. The memory bank consists of divergent low-quality images generated in early training stages." + ], + "image_footnote": [], + "bbox": [ + 513, + 89, + 874, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "function and Layer Norm (LN) are used after the first Pwise layer [28]. Notably, Coordinate is used after the Dwise layer for modeling long-range dependencies in an image. Specially, Coordinate produces cross-channel and position-sensitive attention maps, which helps our model to more accurately locate the representations of interest [12].", + "bbox": [ + 496, + 310, + 890, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the output of a MAT block is obtained with a residual connection, i.e. $\\mathbf{X}_{\\mathrm{MAT}} = \\mathrm{AdaConv}(\\mathbf{X}_{agg}) + \\mathbf{X}_{agg}$ . In the implementation, we stack three MAT blocks in default to gradually refine the correspondence and to augment informative representations (Fig. 1). Empirical verifications will be given in Sec. 4.2.", + "bbox": [ + 496, + 401, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Benefits of MAT. Fig. 3d illustrates the impact of MAT. The query point locates over the left eye of the source image. Here we show the magnitudes of its correspondence over the exemplar, in the third layer of MAT. Obviously, the original correspondence $\\mathbf{A}$ covers both eyes of the exemplar. In contrast, the masked correspondence $\\mathbf{A}_{mask}$ accurately concentrates over the left eye. Such superiority significantly boost the quality of ultimate images.", + "bbox": [ + 496, + 492, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Contrastive Style Learning (CSL)", + "text_level": 1, + "bbox": [ + 500, + 619, + 794, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In MATEBIT, we use the encoder $\\mathcal{E}_B$ to extract local style information $\\mathbf{X}_B$ , and then a MLP to extract global style codes $\\mathbf{z}$ . $\\mathbf{X}_B$ and $\\mathbf{z}$ perform local and global style control on generated images, respectively (Sec. 3.3). To boost the discriminative capacity of style representations, as well as the quality of generated images, we propose a novel contrastive style learning (CSL) method (as shown in Fig 4).", + "bbox": [ + 496, + 642, + 890, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our settings, the exemplars are drawn by human artists and thus considered as high-quality. In contrast, the images generated in early training stages are typically low-quality. Inspired by the idea of contrastive learning [9], we use the exemplar $y_{B}$ as the positive sample, while a collection of early generated images as negative. Let $\\mathbf{z}$ denotes style codes of the generated image $\\hat{x}_B$ , $\\mathbf{z}^+$ that of exemplar $y_{B}$ , and $\\{\\mathbf{z}_1^-, \\mathbf{z}_2^-, \\dots, \\mathbf{z}_m^-\\}$ the style codes of $m$ negative samples. CSL learns style representations by maximizing the mutual information between anchors and positive samples,", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22421", + "bbox": [ + 478, + 944, + 516, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while minimizing that between anchors and negative samples. Our contrastive style loss is computed by:", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s t y l e} = - \\log \\frac {\\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} ^ {+}}{\\tau}\\right)}{\\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} ^ {+}}{\\tau}\\right) + \\sum_ {j = 1} ^ {m} \\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} _ {j} ^ {-}}{\\tau}\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 128, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau = 0.07$ and $m = 1024$ . In the implementation, we use a queue to cache negative style vectors.", + "bbox": [ + 76, + 183, + 468, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Translation network", + "text_level": 1, + "bbox": [ + 76, + 222, + 269, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To boost both the semantic consistency and style faithfulness, we additionally use source semantic features and global style codes for decoding an image. Specially, we design our whole translation network following U-Net (Fig. 2), where the multi-level features in $\\mathcal{E}_A$ are skip-connected to the decoder $\\mathcal{D}_B$ , for supplementing informative semantic structures of the input image $x_A$ . Besides, we use the style codes $\\mathbf{z}$ to globally control the style of generated images, in the manner of AdaIN [14]. Specially, $\\mathbf{z}$ is mapped to channel-wise modulating factors by fully-connected (FC) layers. In this way, we diminish the impact of correspondence learning on image generation, and provide reliable style control for even unmatched regions.", + "bbox": [ + 75, + 244, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, our translation network allows both local and global style control, and reuses the semantic features of input images. As a result, the generated image is desired to present consistent semantic to the input $x_{A}$ and faithful style to the exemplar $y_{B}$ . More details of our network are available in the supplementary material.", + "bbox": [ + 75, + 441, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss functions", + "text_level": 1, + "bbox": [ + 76, + 541, + 223, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our whole network is end-to-end optimized to jointly achieve high-fidelity image generation and accurate correspondence. Following [45], we obtain training triplets $\\{x_A, y_B, x_B\\}$ from the ready-made data pair $\\{x_A, x_B\\}$ , where $y_B$ is a geometrically warped version of $x_B$ . The generated image is denoted by $\\hat{x}_B = \\mathcal{G}(x_A, y_B)$ . Our loss functions are similar to [45], except for the previous contrastive style loss $\\mathcal{L}_{style}$ and the structural loss $\\mathcal{L}_{str}$ below.", + "bbox": [ + 75, + 564, + 468, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic alignment Loss. For accurate cross-domain correspondence learning, the encoders $\\mathcal{E}_A$ and $\\mathcal{E}_B$ should align $x_A$ and $x_B$ to consistent representations. The corresponding semantic alignment loss is:", + "bbox": [ + 75, + 686, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\left\\| \\mathcal {E} _ {A} \\left(x _ {A}\\right) - \\mathcal {E} _ {B} \\left(x _ {B}\\right) \\right\\| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 756, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correspondence Loss. Ideally, if we warp $y_{B}$ in the same way as Eq.3, the resulting image should be exactly $x_{B}$ . We thus constrain the learned correspondence by:", + "bbox": [ + 76, + 781, + 468, + 827 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o r r}} = \\left\\| \\tilde {\\mathbf {A}} _ {\\text {m a s k}} ^ {T} y _ {B} \\downarrow - x _ {B} \\downarrow \\right\\| _ {1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 835, + 468, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\downarrow$ indicates down-sampling $y_{B}$ and $x_{B}$ to the size (i.e. width and height) of $\\mathbf{X}_A$ .", + "bbox": [ + 76, + 869, + 467, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perceptual Loss. The generated image $\\hat{x}_B$ should be semantic-consistent with the ground truth $x_B$ in term of semantic. We thus use the perceptual loss:", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p e r c}} = \\left\\| \\varphi_ {l} \\left(\\hat {x} _ {B}\\right) - \\varphi_ {l} \\left(x _ {B}\\right) \\right\\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 146, + 890, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi_{l}$ denotes the activations after layer relu4_2 in pretrained VGG19 [40], which represent high-level semantics.", + "bbox": [ + 498, + 171, + 890, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contextual Loss. In addition, the generated image should be in the same style as the exemplar. In addition to the previous contrastive style loss (Eq.7), we additionally use the contextual loss (CX) [31] to constrain on local style consistency. The contextual loss is computed by:", + "bbox": [ + 498, + 202, + 890, + 277 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c t x} = - \\log \\left(\\sum_ {l} w _ {l} \\mathrm {C X} \\left(\\varphi_ {l} \\left(\\hat {x} _ {B}\\right), \\varphi_ {l} \\left(y _ {B}\\right)\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 285, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{l}$ balances the terms of different VGG19 layers.", + "bbox": [ + 500, + 327, + 867, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Structural Loss. The generated image should preserve semantic structures in the input image. Correspondingly, we use the Learned Perceptual Image Patch Similarity (LPIPS) [46] between their boundaries as the structural loss:", + "bbox": [ + 498, + 343, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s t r} = \\operatorname {L P I P S} \\left(\\mathcal {H} \\left(\\hat {x} _ {B}\\right), \\mathcal {H} \\left(x _ {B}\\right)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 412, + 890, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{H}$ is the HED algorithm [43], which has been widely used for extracting semantic boundaries in an image.", + "bbox": [ + 498, + 438, + 890, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial loss. Finally, we add a discriminator $\\mathcal{D}$ to distinguish real images in domain $\\mathcal{B}$ and the generated images [7]. The adversarial loss is:", + "bbox": [ + 498, + 469, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} ^ {\\mathcal {D}} = - \\mathbb {E} [ h (\\mathcal {D} (y _ {B})) ] - \\mathbb {E} [ h (- \\mathcal {D} (\\hat {x} _ {B})) ], \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 521, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} ^ {\\mathcal {G}} = - \\mathbb {E} [ \\mathcal {D} (\\hat {x} _ {B}) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 542, + 681, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h(t) = \\min (0, - 1 + t)$ is the hinge loss function [1].", + "bbox": [ + 500, + 570, + 888, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total loss. In summary, our overall objective function is,", + "bbox": [ + 519, + 585, + 888, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\mathcal {G}} \\max _ {\\mathcal {D}} \\lambda_ {1} \\mathcal {L} _ {\\text {s t y l e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c o r r}} + \\lambda_ {4} \\mathcal {L} _ {\\text {s t r}} \\tag {14} \\\\ + \\lambda_ {5} \\left(\\mathcal {L} _ {p e r c} + \\mathcal {L} _ {c t x}\\right) + \\lambda_ {6} \\left(\\mathcal {L} _ {a d v} ^ {\\mathcal {G}} + \\mathcal {L} _ {a d v} ^ {\\mathcal {D}}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 622, + 890, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ denotes the weight parameters.", + "bbox": [ + 500, + 674, + 763, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 694, + 624, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We apply spectral normalization [33] to all the layers in the translation network and discriminator. We use the Adam [20] solver with $\\beta_{1} = 0$ and $\\beta_{2} = 0.999$ . The learning rates for the generator and discriminator are set as $1e - 4$ and $4e - 4$ respectively, following TTUR [11]. The experiments are conducted using 4 24GB RTX3090 GPUs. Limited by the computation load, we restrict the resolution of generated images to $256\\times 256$ in all translation tasks.", + "bbox": [ + 498, + 719, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We mainly conduct experiments on the following datasets. (1) CelebA-HQ [22] contains 30,000 facial photos. We chose 24,000 samples as the training set", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22422", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9a154237a79f08c46aa09a363fec2ffdbc51b3f30c873ea3297acd6ca7c04e47.jpg", + "table_caption": [ + "Table 1. Comparison on the Metfaces [18], CelebA-HQ [22], Ukiyo-e [36], Cartoon [36], AAHQ [24], and DeepFashion [27] datasets." + ], + "table_footnote": [], + "table_body": "
CelebA-HQMetfacesCartoonUkiyo-eAAHQDeepFashionTime ↓
FID ↓SWD ↓Texture ↑Color ↑Semantic↑FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓(s)
SPADE [35]31.526.90.9270.9550.92245.626.997.530.545.626.979.432.136.227.80.196
CoCosNet [45]14.315.20.9580.9770.94925.624.366.827.138.313.962.621.914.417.20.321
CoCosNet-v2 [49]13.214.00.9540.9750.94823.322.466.427.032.111.062.422.813.016.71.573
MCL-Net [44]12.814.20.9510.9760.95323.824.567.927.932.412.464.422.212.916.20.309
DynaST [25]12.012.40.9590.9780.95229.228.662.826.538.914.267.224.08.411.80.214
MATEBIT (ours)11.513.20.9660.9860.94926.019.164.427.630.311.556.019.58.210.00.185
", + "bbox": [ + 81, + 106, + 890, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/702a17053195b99e87c84d12b12e78855cb6452160ec0e42efbfcfbbed3d057a.jpg", + "image_caption": [ + "Figure 5. Results on the Metfaces [18], CelebA-HQ [22], Ukiyo-e [36], Cartoon [36], AAHQ [24], and DeepFashion [27] datasets." + ], + "image_footnote": [], + "bbox": [ + 101, + 223, + 867, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and 3000 as the test set. (2) Metfaces [18] consists of 1336 high-quality artistic facial portraits. (3) AAHQ [24] consists of high-quality facial avatars. We randomly select 1500 samples for training and 1000 samples for testing. (4) Ukiyo-e [36] consists of high-quality Ukiyo-e faces. We randomly select 3000 and 1000 samples for training and testing, respectively. (5) Cartoon [36] consists of 317 cartoon faces. (6) DeepFashion [27] consists of 800,00 fashion images. On CelebA-HQ, we connect the face landmarks for face region, and use Canny edge detector to detect edges in the background. On DeepFashion, we use the officially provided landmarks as input. On the other datasets, we use HED [43] to obtain semantic edges.", + "bbox": [ + 75, + 638, + 470, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 838, + 366, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We select several advanced models, including SPADE [35], CoCosNet [45], CoCosNet-v2 [49], MCL-Net [44], and DynaST [25], for comparison. For a fair comparison,", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we retrain their models at resolution $256 \\times 256$ under the same settings as ours.", + "bbox": [ + 498, + 638, + 890, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative evaluation. We adopt several criteria to fully evaluate the generation results. (1) Fréchet Inception Score (FID) [38] and Sliced Wasserstein distance (SWD) [17] are used to evaluate the image perceptual quality. (2) To assess style relevance and semantic consistency of translated images [45], we compute the color, texture, and semantic metrics based on VGG19 [40]. Specifically, the cosine similarities between low-level features (i.e. $relu1\\_2$ and $relu2\\_2$ ) are used to measure color and texture relevance, respectively; the average cosine similarity between high-level features (i.e. $relu3\\_2$ , $relu4\\_2$ , and $relu5\\_2$ ) measures the semantic consistency.", + "bbox": [ + 496, + 671, + 892, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative comparison results are shown in Table 1. Compared to existing methods, our model consistently achieves superior or highly competitive performance across", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22423", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82cbc5b4d8f3d6e09a10fbeee5607b7de4e024ff83dcd007bee5ad55eb025ed8.jpg", + "image_caption": [ + "Figure 6. More results generated by MATEBIT. Images shown in the left-bottom corner are source images or exemplars." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 446, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "all the datasets. Especially, MATEBIT significantly improves the style relevance in both texture and color. On the complicated AAHQ dataset, which contains diverse styles of avatars, MATEBIT dramatically decreases both FID and SWD. Such superiority indicates that our generated images are of better perceptual quality; and present consistent appearance to similar parts in exemplars. We additionally report the average time each method costs for generating an image. Our method shows the best efficiency and is significantly faster than previous methods.", + "bbox": [ + 75, + 439, + 468, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparison. Fig 5 illustrates images generated by different methods. Obviously, previous methods present geometric distortions, blurring artifacts, inconsistent colors, or identity inconsistency. In contrast, MATEBIT consistently produces appealing results, including more results shown in Fig. 6. Specially, our results preserve the semantic structure of input images, and present consistent appearance with semantically similar regions in exemplars. Previous methods suffer serious degradations mainly due to the matching errors in full correspondence learning. In our method, we distinct reliable and unreliable correspondence, and release the role of matching in image generation. As a result, our method stably transfers a source image to the target style of an exemplar.", + "bbox": [ + 75, + 592, + 468, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 816, + 225, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impacts of MAT. We present a comprehensive analysis to justify the important component in our architecture, i.e. MAT. We here modify our full model by (1) removing the MAT module (i.e. w/o MAT), (2) removing ReLU in", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/202e9d72426fd8a879453b556e7bd83ccebd505dd8eed053b88a7a08b54b5432.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 689, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1aa06cd1a4d6a58884dd41fa220dfeb8b76705ac2b172a9325b15123d222ba6c.jpg", + "image_caption": [ + "Figure 7. Impact of the number of MAT blocks on performance." + ], + "image_footnote": [], + "bbox": [ + 696, + 90, + 888, + 179 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ceefe2e5d338cc6c84f870bf797182fde28d5cb37cbdf5c57f7b6926f6c20c5.jpg", + "table_caption": [ + "Table 2. Results of ablation studies on MetaFaces." + ], + "table_footnote": [], + "table_body": "
FID ↓SWD ↓Texture ↑Color ↑Semantic↑
w/o MAT27.722.30.8830.8520.856
w/o ReLU30.120.00.9160.9560.928
w/o MAT (→ Full Corr. [45])34.119.70.8720.9690.902
w/o AdaConv (→ FFN [41])34.320.10.8410.9710.896
w/o Lstyle30.320.80.8740.8480.908
w/o Lstyle (→ LCAST [47])31.019.80.9040.9830.921
w/o Lstr28.421.80.9150.9830.911
w/o skip connection47.027.70.9250.9580.902
w/o global style z27.621.20.9270.9610.925
Full Model26.019.10.9380.9840.927
", + "bbox": [ + 503, + 208, + 888, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MAT (i.e. w/o ReLU), (3) replacing MAT with three-layer full correspondence learning modules [45] (i.e. Full Corr.), and (4) replacing the AdaConv with FFN [45] (i.e. w/ AdaConv). The results in Table 2 show that removing MAT or ReLU dramatically hurts the performance. Besides, using the full correspondence learning in [45] or using FFN also significantly decreases the texture relevance and semantic consistency. Correspondingly, these model variants leads to inferior results in terms of textures or colors, compared to our full model (Fig. 8). Recall the visualized correspondence in Fig. 1, our method learns remarkably accurate correspondence, which ultimately benefits the quality of generated images. In addition, Fig. 7 shows that both the semantic consistency and style realism broadly improve with the number of MAT blocks and peak at three. All these observations demonstrate our motivation that MAT gradually refines cross-domain correspondence and augments informative representations for generating high-quality images.", + "bbox": [ + 496, + 371, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contrastive Style Loss. To verify the effectiveness of the proposed contrastive style learning methodology, we train our model by (1) removing the style loss (i.e. w/o $\\mathcal{L}_{\\text{style}}$ ) and (2) replacing $\\mathcal{L}_{\\text{style}}$ with the loss used in CAST [47] (i.e. w/ $\\mathcal{L}_{\\text{CAST}}$ ). In CAST, only high-quality exemplars in different styles are used as negative samples. Differently, we use low-quality generated images in diverse styles as negative samples. From both Table 2 and Fig. 8, we observe that: (1) without $\\mathcal{L}_{\\text{style}}$ , although the generated images show high semantic consistency, they present low style relevance; (2) $\\mathcal{L}_{\\text{CAST}}$ benefits the style relevance, but leads to inferior performance to $\\mathcal{L}_{\\text{style}}$ . These comparison results meet our expectation that: our CSL methodology enables the learned style codes to discriminate subtle divergences between images with different perceptual qualities. Such discriminability facilitates pushes the network to generate high-quality images.", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22424", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13f1705bb5463520e19bff57e698365e044b700db4c94da9ab6f774b55814bb6.jpg", + "image_caption": [ + "Figure 8. Comparison of generated images by different variants of our model, on Metfaces [18]." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1180542e4d15edc9093f885a172b0d4844e0d428d69c5388191c0dcd65d46671.jpg", + "image_caption": [ + "Figure 9. Our method can transfer facial photos to artistic portraits in the style of exemplars." + ], + "image_footnote": [], + "bbox": [ + 96, + 179, + 444, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Skip connections & global style control. In MATEBIT, we use skip connections to supplement input semantic information. Removing skip connections dramatically hurts the semantic inconsistency and the quantitative results. Besides, using global style vector $\\mathbf{z}$ increases subtle details, e.g. the colors over the mouth, rings, and hairs.", + "bbox": [ + 76, + 429, + 467, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, MAT learns accurate correspondence and enables context-aware feature augmentation; the contrastive style learning benefits the style control and high-quality image generation; and the U-Net architecture helps the preservation of semantic information. Ultimately, all these benefits make our model significantly outperform previous state-of-the-art methods in generating plausible images.", + "bbox": [ + 75, + 521, + 467, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Applications", + "text_level": 1, + "bbox": [ + 76, + 633, + 210, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Artistic Portrait Generation. An potential application of our method is transferring a facial photo to an artistic portrait, in the style of an exemplar. We here apply the previously learned models to randomly selected facial photos from CelebA-HQ [22]. As illustrated in Fig. 9, our method can generate appealing portraits with consistent identity and faithful style appearance.", + "bbox": [ + 75, + 657, + 467, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Chinese Ink Painting Generation. To verify the capacity of our model in generating complex images, we additionally apply it to generate Chinese Ink paintings. Specifically, we collect paintings of landscapes and facial portraits from the web, and then train and test our model on each subset respectively. Fig. 10 illustrates the results of painting generation and photo-to-painting translation. Obviously, all the generated images show remarkably high quality. Besides, our model successfully captures subtle differences", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8856d757dc719f66384509178a9a7088c7264a071a6c5ba2acc263f040c362b2.jpg", + "image_caption": [ + "Figure 10. Chinese ink paintings generation (1st & 3rd rows), as well as photo-to-painting translation (2nd & 4th rows)." + ], + "image_footnote": [], + "bbox": [ + 524, + 179, + 864, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "between different exemplars, demonstrating its remarkable capacity in style control.", + "bbox": [ + 498, + 430, + 890, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 469, + 625, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a novel exemplar-guided image translation method, dubbed MATEBIT. Both quantitative and qualitative experiments show that MATEBIT is capable of generating high-fidelity images in a number of tasks. Besides, ablation studies demonstrate the effectiveness of MAT and contrastive style learning. Despite such achievements, the artistic portraits transferred from facial photos (Fig. 9) are inferior to those shown in Fig. 6. This may be due to the subtle differences in edge maps between photos and artistic paintings. In the near future, we will explore to solve this issue via semi-supervised learning or domain transfer technologies.", + "bbox": [ + 498, + 496, + 890, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 693, + 666, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China under Grant 2018AAA0103202; in part by the National Natural Science Foundation of China under Grants 61971172, 61971339, 62176230, U22A2033, U22A2096; in part by the Zhejiang Provincial Science and Technology Program in China under Grant 2021C01108; in part by the NSFC-Zhejiang Joint Fund for the Integration of Industrialization and Informatization under Grant U1909210; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; and in part by Open Research Projects of Zhejiang Lab under Grant 2021KG0AB01.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22425", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. 2018. 5", + "[2] Huiwen Chang, Jingwan Lu, Fisher Yu, and Adam Finkelstein. Pairedcyclegan: Asymmetric style transfer for applying and removing makeup. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 40-48, 2018. 1", + "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2", + "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2", + "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 5", + "[8] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2", + "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 4", + "[10] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11936-11945, 2021. 1", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[12] Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13713-13722, 2021. 2, 4" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Xueqi Hu, Xinyue Zhou, Qiusheng Huang, Zhengyi Shi, Li Sun, and Qingli Li. Qs-attn: Query-selected attention for contrastive learning in i2i translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18291-18300, 2022. 2", + "[14] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5", + "[15] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1", + "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1, 2", + "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6", + "[18] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 6, 8", + "[19] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 2", + "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[21] Thomas N Kipf and Max Welling. Variational graph autoencoders. arXiv preprint arXiv:1611.07308, 2016. 1", + "[22] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549-5558, 2020. 5, 6, 8", + "[23] Boyi Li, Felix Wu, Kilian Q Weinberger, and Serge Belongie. Positional normalization. In Advances in Neural Information Processing Systems, pages 1620-1632, 2019. 3, 4", + "[24] Mingcong Liu, Qiang Li, Zekui Qin, Guoxin Zhang, Pengfei Wan, and Wen Zheng. Blendgan: implicitly gan blending for arbitrary stylized face generation. Advances in Neural Information Processing Systems, 34:29710-29722, 2021. 6", + "[25] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In European Conference on Computer Vision, 2022. 1, 2, 3, 4, 6", + "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22426", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6", + "[28] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 2, 3, 4", + "[29] Mandi Luo, Haoxue Wu, Huaibo Huang, Weizan He, and Ran He. Memory-modulated transformer network for heterogeneous face recognition. IEEE Transactions on Information Forensics and Security, 2022. 1", + "[30] Liqian Ma, Xu Jia, Stamatios Georgoulis, Tinne Tuytelaars, and Luc Van Gool. Exemplar guided unsupervised image-to-image translation with semantic consistency. Proceedings ICLR 2019, 2019. 1", + "[31] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In Proceedings of the European conference on computer vision (ECCV), pages 768-783, 2018. 5", + "[32] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel flow: Semantic correspondence with multi-layer neural features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3395-3404, 2019. 3", + "[33] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In International Conference on Learning Representations, 2018. 5", + "[34] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In European conference on computer vision, pages 319-345. Springer, 2020. 2", + "[35] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 1, 2, 4, 6", + "[36] Justin N. M. Pinkney. Aligned ukiyo-e faces dataset, 2020. 6", + "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2", + "[38] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, August 2020. version 0.2.1. 6", + "[39] Junyoung Seo, Gyuseong Lee, Seokju Cho, Jiyoung Lee, and Seungryong Kim. Midms: Matching interleaved diffusion models for exemplar-based image translation. arXiv preprint arXiv:2209.11047, 2022. 1, 2", + "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR, 2015. 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 7", + "[42] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision, pages 63-79. Springer, 2018. 1", + "[43] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 5, 6", + "[44] Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, and Changgong Zhang. Marginal contrastive correspondence for guided image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10663-10672, 2022. 1, 2, 3, 6", + "[45] Pan Zhang, Bo Zhang, Dong Chen, Lu Yuan, and Fang Wen. Cross-domain correspondence learning for exemplar-based image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5143-5153, 2020. 1, 2, 3, 5, 6, 7", + "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[47] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH, 2022. 2, 7", + "[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1", + "[49] Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocosnet v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11465–11475, 2021. 1, 2, 6", + "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 1" + ], + "bbox": [ + 501, + 92, + 893, + 753 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_model.json b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c6a7bc9cc58b4d4f8e0e8e3382e422c148109d47 --- /dev/null +++ b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.862, + 0.154 + ], + "angle": 0, + "content": "Masked and Adaptive Transformer for Exemplar Based Image Translation" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.18, + 0.791, + 0.2 + ], + "angle": 0, + "content": "Chang Jiang\\(^{1}\\), Fei Gao\\(^{1,2*}\\), Biao Ma\\(^{1}\\), Yuhao Lin\\(^{1}\\), Nannan Wang\\(^{3}\\), Gang Xu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.199, + 0.79, + 0.217 + ], + "angle": 0, + "content": "1 School of Computer Science and Technology, Hangzhou Dianzi University" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.216, + 0.89, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\) Hangzhou Institute of Technology, Xidian University \\(^{3}\\) ISN State Key Laboratory, Xidian University" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.236, + 0.79, + 0.253 + ], + "angle": 0, + "content": "{jc233, aiartma, linyh, gxu}@hdu.edu.cn, {fgao, nnwang}@xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.668 + ], + "angle": 0, + "content": "We present a novel framework for exemplar based image translation. Recent advanced methods for this task mainly focus on establishing cross-domain semantic correspondence, which sequentially dominates image generation in the manner of local style control. Unfortunately, cross-domain semantic matching is challenging; and matching errors ultimately degrade the quality of generated images. To overcome this challenge, we improve the accuracy of matching on the one hand, and diminish the role of matching in image generation on the other hand. To achieve the former, we propose a masked and adaptive transformer (MAT) for learning accurate cross-domain correspondence, and executing context-aware feature augmentation. To achieve the latter, we use source features of the input and global style codes of the exemplar, as supplementary information, for decoding an image. Besides, we devise a novel contrastive style learning method, for acquire quality-discriminative style representations, which in turn benefit high-quality image generation. Experimental results show that our method, dubbed MATEBIT, performs considerably better than state-of-the-art methods, in diverse image translation tasks. The codes are available at https://github.com/AiArt-HDU/MATEBIT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.698, + 0.21, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Image-to-image translation aims at transfer images in a source domain to a target domain [16, 50]. Early studies learn mappings directly by Generating Adversarial Networks (GANs), and have shown great success in various applications [2, 42]. Recently, exemplar based image translation [29, 30, 45], where an exemplar image is used to control the style of translated images, has attracted a lot of attention. Such methods allow high flexibility and controllability, and have a wide range of potential applications in social networks and metaverse. For example, people can transfer a" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.288, + 0.87, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.483, + 0.895, + 0.54 + ], + "angle": 0, + "content": "Figure 1. Visualization of correspondence maps. The red point is the query position. Full Corr. and Masked Corr. denote the full correspondence [45] and masked one in our method, respectively. CAM denotes visualization by Class Activation Mapping [48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.895, + 0.608 + ], + "angle": 0, + "content": "facial sketch to an artistic portrait, in the style of oil paintings or avatars. Despite the remarkable progress, yielding high-fidelity images with consistent semantic and faithful styles remains a grand challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.822 + ], + "angle": 0, + "content": "Early pioneering works [15, 21, 35] attempt to globally control the style of generated images. However, such methods ignore spatial correlations between an input image and an exemplar, and may fail to produce faithful details. Recently, some advanced methods [25, 44, 45, 49] first establish the cross-domain semantic correspondence between an input image and an exemplar, and then use it to warp the exemplar for controlling local style patterns. In these methods, the quality of generated images relies heavily on the learned correspondence [39]. Unfortunately, cross-domain semantic matching is challenging, since there is no reliable supervision on correspondence learning [45]. As a result, potential matching errors ultimately lead to degraded artifacts in generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To combat this challenge, we propose to boost the matching accuracy on one hand, and to diminish the role of matching in image generation on the other hand. Inspired by the great success of Transformers [6, 10, 26, 41], we first devise a Masked and Adaptive Transformer (MAT) for learning ac" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22418" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "curate cross-domain correspondence and executing context-aware feature augmentation. Previous works [44, 45, 49] have used the vanilla attention mechanism [41] for learning full correspondence. However, the initial attention typically involves ambiguous correspondences (2nd row in Fig. 1). To mitigate these limitations, in MAT, we use a masked attention to distinguish the correspondence as reliable or not, and then reliability-adaptively aggregate representations. Besides, the Feed-Forward Network (FFN) [41] in vanilla transformers neglects contextual correlations inside an image. We thus replace FFN by an adaptive convolution block [28], where the coordinate attention [12] and depth-wise separable convolution [5] are used to capture contextual correlations and to improve efficiency. With a joint consideration of matching reliability and contextual correlations, MAT gradually focuses on accurate correspondences and emphasizes on features of interest (3rd row in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.59 + ], + "angle": 0, + "content": "In addition, to boost both the semantic consistency and style faithfulness, we supplementally use semantic features of the input image and global style codes of the exemplar for decoding an image. To this end, we first design our whole network following the U-Net architecture [16]. Besides, we devise a novel contrastive style learning (CSL) framework for acquiring discriminative style representations. Recently, Zhang et al. [47] propose a similar CSL method, where the target exemplar is used as a positive sample, and the other exemplars as negative ones. Differently, we use low-quality images, generated during early training stages, as negative samples. In this way, our style codes are desired to discriminate not only subtle differences in style, but also those in perceptual quality. Ultimately, the learned global style codes, cooperating with the local style control induced by MAT, in turn benefit high-quality image generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.473, + 0.756 + ], + "angle": 0, + "content": "With the proposed techniques above, our full model, dubbed MATEBIT, diminishes the impact of position-wise matching on image quality, and integrates both local and global style control for image generation. Experimental results show that MATEBIT generates considerably more plausible images than previous state-of-the-art methods, in diverse image translation tasks. In addition, comprehensive ablation studies demonstrate the effectiveness of our proposed components. Finally, we perform interesting applications of photo-to-painting translation and Chinese ink paintings generation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.21, + 0.785 + ], + "angle": 0, + "content": "2. Relate Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Exemplar Based Image Translation. Recently, exemplar based image translation has attracted increasing attention. For example, Park et al. [35] learn an encoder to map the exemplar image into a global style vector, and use it to guide image generation. Such a global style control strategy enables style consistency in whole, but fails to produce subtle details. Most recently, researchers propose a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "matching-then-generation framework [39]. Specially, they first establish dense correspondence between an input and an exemplar, and then reshuffle the exemplar for locally control the style of synthesize images. For example, Zhang et al. [45] establish position-wise correspondence based on the Cosine attention mechanism and warp the exemplar correspondingly. Afterwards, the warped image dominates the generation of images in the manner of SPADE [35]. To reduce the cost of matching in high-resolution image generation, Zhou et al. [49] introduce a hierarchical refinement of semantic correspondence from ConvGRU-PatchMatch. Besides, Liu et al. [25] used a dynamic pruning method for learning hierarchical sparse correspondence. They also use reliability-adaptive feature integration to improve the quality of generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.428 + ], + "angle": 0, + "content": "Previous methods merely use global or local style control, and the latter relies heavily on the learned correspondence. Besides, they consider little about contextual correlations inside an image. In this paper, we use both global and local style control to boost the style consistency. Besides, we take contextual correlations into consideration and execute reliability-adaptive feature augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.896, + 0.656 + ], + "angle": 0, + "content": "Transformers. Transformers [41] have shown incredible success from the field of natural language processing (NLP) [19] to computer vision (CV) [6, 26]. Multihead attention (MHA) and FFN are key components in a Transformer, and have been used in exemplar based image translation. However, they induce unreliable matching results and neglect context correlations in feature translation. In our MAT, we combat these limitations by replacing them with a masked attention and a context-aware convolution block, respectively. Recently, researchers use semantic masks to facilitate representation learning [4, 8, 37], where a mask predictor is required. Differently, we use a ReLU function to mask over the attention layer, for distinguishing correspondence as reliable or not (Sec. 3.1). In general, MAT follows a concise and efficient architecture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Contrastive Learning. Contrastive learning has shown its effectiveness in various computer vision tasks [9, 13, 34]. The basic idea is to learn a representation by pushing positive samples toward an anchor, and moving negative samples away from it. Different sampling strategies and contrastive losses have been extensively explored in various downstream tasks. For example, Chen et al. [3] and He et al. [9] obtain positive samples by augmenting original data. In the field of image translation, Park et al. [34] propose patchwise contrastive learning by maximizing the mutual information between cross-domain patches. Similarly, Zhang et al. [47] use contrastive learning for acquiring discriminative style representations. In the task of exemplar based image translation, Zhan et al. [44] use contrastive learning to align cross-domain images to a consistent semantic feature space, so as to boost the accuracy of matching. Differently, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22419" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.249, + 0.468, + 0.264 + ], + "angle": 0, + "content": "Figure 2. Overview of our image translation network, MATEBIT." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.275, + 0.47, + 0.321 + ], + "angle": 0, + "content": "use early generated images as negative samples, so that the learned style representations can discriminate subtle differences in both style and perceptual quality (Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.338, + 0.289, + 0.355 + ], + "angle": 0, + "content": "3. The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.545 + ], + "angle": 0, + "content": "Given an input image \\( x_{A} \\) in domain \\( \\mathcal{A} \\) and an exemplar image \\( y_{B} \\) in domain \\( \\mathcal{B} \\), our goal is to generate a target image \\( x_{B} \\) which preserves semantic structures in \\( x_{A} \\) but resembles the style of similar parts in \\( y_{B} \\). Fig. 2 shows an overview of our translation network \\( \\mathcal{G} \\). Specially, we first align \\( x_{A} \\) and \\( y_{B} \\) to an intermediate feature space by encoders \\( \\mathcal{E}_{A} \\) and \\( \\mathcal{E}_{B} \\), respectively. Afterwards, we use a Masked and Adaptive Transformer (MAT) for correspondence learning and feature augmentation. Finally, a decoder \\( \\mathcal{D}_B \\) produces an output image \\( \\hat{x}_B \\) based on the augmented features, as well as the source features and target style codes. Details are described below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.442, + 0.575 + ], + "angle": 0, + "content": "3.1. Masked and Adaptive Transformer (MAT)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.763 + ], + "angle": 0, + "content": "In order to establish accurate cross-domain correspondence, we propose a novel and concise Transformer architecture, i.e. MAT. In general, the architecture of MAT (Fig. 3b) follows that of vanilla Transformers (Fig. 3a) [41]. Differently, we use masked attention to distinguish reliable and unreliable correspondence, instead of using multi-head attention. Besides, we use Positional Normalization (PONO) [23] and an Adaptive Convolution (AdaConv) block [28], instead of LN and MLP-based FFN, respectively. MAT is desired to gradually concentrate on accurate matching, and to reliability-adaptively augment representations with contextual correlations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Masked Correspondence Learning. Let \\(\\mathbf{X}_A \\in \\mathbb{R}^{H \\times W \\times C}\\) and \\(\\mathbf{Y}_B \\in \\mathbb{R}^{H \\times W \\times C}\\) be the representations of \\(x_A\\) and \\(y_B\\) in the intermediate feature space, with height \\(H\\), width \\(W\\), and \\(C\\) channels. We first map \\(\\mathbf{X}_A\\) to the query \\(\\mathbf{Q} \\in \\mathbb{R}^{HW \\times C}\\), and \\(\\mathbf{Y}_B\\) to the key \\(\\mathbf{K} \\in \\mathbb{R}^{HW \\times C}\\) and value \\(\\mathbf{V} \\in \\mathbb{R}^{HW \\times C}\\), by using \\(1 \\times 1\\) convolutions, respectively. As shown in Fig. 3d, we add positional encoding (PE) to \\(\\mathbf{X}_A\\) and \\(\\mathbf{Y}_B\\), for embedding spatial correlations. Afterwards, we learn the initial correspondence \\(\\mathbf{A} \\in \\mathbb{R}^{HW \\times HW}\\) fol" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.095, + 0.627, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.095, + 0.751, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.096, + 0.868, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.319, + 0.868, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.514, + 0.892, + 0.542 + ], + "angle": 0, + "content": "Figure 3. Detailed architectures. (a) Vanilla Transformer block, (b) MAT block, (c) AdaConv block, and (d) Masked Attention." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.822, + 0.569 + ], + "angle": 0, + "content": "lowing the Cosine attention mechanism [45], i.e." + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.579, + 0.892, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} (u, v) = \\frac {\\tilde {\\mathbf {Q}} (u) \\tilde {\\mathbf {K}} (v) ^ {T}}{| | \\tilde {\\mathbf {Q}} (u) | | \\cdot | | \\tilde {\\mathbf {K}} (v) | |}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.626, + 0.892, + 0.688 + ], + "angle": 0, + "content": "with \\(\\tilde{\\mathbf{Q}}(u) = \\mathbf{Q}(u) - \\bar{\\mathbf{Q}}(u)\\), \\(\\tilde{\\mathbf{K}}(v) = \\mathbf{K}(v) - \\bar{\\mathbf{K}}(v)\\), where \\(u, v \\in [1, \\dots, HW]\\) are position indices; \\(\\bar{\\mathbf{Q}}(u)\\) and \\(\\bar{\\mathbf{K}}(v)\\) are the means of \\(\\mathbf{Q}(u)\\) and \\(\\mathbf{K}(v)\\), respectively. \\(\\mathbf{A}(u, v)\\) is the matching score between \\(\\mathbf{Q}(u)\\) and \\(\\mathbf{K}(v)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.825 + ], + "angle": 0, + "content": "Previous methods [44, 45] typically use the initial correspondence map \\(\\mathbf{A}\\) to reshuffle an exemplar for controlling local patterns in image synthesis. However, induced by the difficulties in cross-domain correspondence learning, \\(\\mathbf{A}\\) involves unreliable match scores (Fig. 3d). As a result, the reshuffled image will lead to implausible artifacts in generated images. To combat this limitation, we distinguish initial matching scores as reliable or not, according to their signs [32]. The masked correspondence map becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.836, + 0.892, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {\\text {m a s k}} = \\operatorname {R e L U} (\\mathbf {A}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In DynaST [25], two networks are used to predict the reliability mask of correspondence. However, it's challenging to effectively train the network, because there is no super" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "vision on matching during training. In contrast, ReLU contains no learnable parameters and ultimately leads to superior performance over DynaST (Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.137, + 0.468, + 0.184 + ], + "angle": 0, + "content": "Reliability-Adaptive Feature Aggregation. For regions with reliable correspondence in \\( x_{A} \\), we use \\( \\mathbf{A}_{mask} \\) to warp the value features, \\( \\mathbf{V} \\), derived from the exemplar:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.19, + 0.468, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\text {c o r}} = \\tilde {\\mathbf {A}} _ {\\text {m a s k}} \\mathbf {V}, \\text {w i t h} \\tilde {\\mathbf {A}} _ {\\text {m a s k}} = \\operatorname {s o f t m a x} (\\alpha \\cdot \\mathbf {A} _ {\\text {m a s k}}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.223, + 0.468, + 0.252 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scaling coefficient to control the sharpness of the softmax function. In default, we set its value as 100." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.253, + 0.468, + 0.358 + ], + "angle": 0, + "content": "For regions with unreliable correspondence in \\( x_{A} \\), \\( \\mathbf{X}_{cor} \\) provides an average style representation of \\( \\mathbf{V} \\). We further extract complementary information from the query, \\( \\mathbf{Q} \\), derived from the input. Inspired by SPADE [35], we first transfer \\( \\mathbf{Q} \\) to the target domain by using pixel-wise modulation parameters (i.e., \\( \\gamma \\) for scale and \\( \\beta \\) for bias) learned from \\( x_{A} \\). The modulation is formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.367, + 0.469, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {\\text {n o r m}} = \\gamma (x _ {A}) \\frac {\\mathbf {Q} - \\mu (\\mathbf {Q})}{\\sigma (\\mathbf {Q})} + \\beta (x _ {A}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.408, + 0.468, + 0.454 + ], + "angle": 0, + "content": "where \\(\\mu (\\mathbf{Q})\\) and \\(\\sigma (\\mathbf{Q})\\) are the mean value and standard deviance of \\(\\mathbf{Q}\\). Afterwards, we select the translated features of unreliable corresponded regions in \\(x_{A}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.463, + 0.469, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {u n c o r} = \\left(1 - \\sum_ {j} \\mathbf {A} _ {m a s k}\\right) \\odot \\mathbf {Q} _ {n o r m}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.468, + 0.579 + ], + "angle": 0, + "content": "where the summation is along the second dimension; \\(\\odot\\) denotes point-wise production with broadcasting. Since \\(\\gamma\\) and \\(\\beta\\) are learned from the input image \\(x_{A}\\), the modulated features preserve the semantic information of \\(x_{A}\\). Besides, constraints on the generated image will push the selected features convey to the style of \\(y_{B}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.579, + 0.468, + 0.64 + ], + "angle": 0, + "content": "Ideally, \\(\\mathbf{X}_{cor}\\) and \\(\\mathbf{X}_{uncor}\\) would complement each other and facilitate both semantic consistency and style relevance in image generation. To this end, we integrate \\(\\mathbf{X}_{cor}\\), \\(\\mathbf{X}_{uncor}\\), and \\(\\mathbf{Q}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.649, + 0.468, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {a g g} = \\operatorname {P O N O} \\left(\\mathbf {X} _ {c o r} + \\mathbf {X} _ {u n c o r} + \\mathbf {Q}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.468, + 0.734 + ], + "angle": 0, + "content": "In PONO [23], features at each position are normalized dependently. Compared to LN in vanilla transformers and DynaST [25], PONO boosts the flexibility in reliability-adaptive feature aggregation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Context-Aware Feature Augmentation. Inspired by ConvNeXT [28], we replace FFN by an AdaConv block to position-adaptively emphasize informative representations. Besides, we use the coordinate attention (CoordAten) module [12] to capture contextual correlations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The architecture of the AdaConv block is as shown in Fig. 3c. We fist use the depthwise convolution (Dwise) to update representations in each channel separately; and then use two pointwise convolutions (Pwise) to automatically emphasize representations of interest, at every position. The Gaussian Error Linear Unit (GELU) activation" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.09, + 0.875, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.297 + ], + "angle": 0, + "content": "Figure 4. Contrastive style learning. The memory bank consists of divergent low-quality images generated in early training stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.401 + ], + "angle": 0, + "content": "function and Layer Norm (LN) are used after the first Pwise layer [28]. Notably, Coordinate is used after the Dwise layer for modeling long-range dependencies in an image. Specially, Coordinate produces cross-channel and position-sensitive attention maps, which helps our model to more accurately locate the representations of interest [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Finally, the output of a MAT block is obtained with a residual connection, i.e. \\(\\mathbf{X}_{\\mathrm{MAT}} = \\mathrm{AdaConv}(\\mathbf{X}_{agg}) + \\mathbf{X}_{agg}\\). In the implementation, we stack three MAT blocks in default to gradually refine the correspondence and to augment informative representations (Fig. 1). Empirical verifications will be given in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Benefits of MAT. Fig. 3d illustrates the impact of MAT. The query point locates over the left eye of the source image. Here we show the magnitudes of its correspondence over the exemplar, in the third layer of MAT. Obviously, the original correspondence \\(\\mathbf{A}\\) covers both eyes of the exemplar. In contrast, the masked correspondence \\(\\mathbf{A}_{mask}\\) accurately concentrates over the left eye. Such superiority significantly boost the quality of ultimate images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.795, + 0.635 + ], + "angle": 0, + "content": "3.2. Contrastive Style Learning (CSL)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.749 + ], + "angle": 0, + "content": "In MATEBIT, we use the encoder \\(\\mathcal{E}_B\\) to extract local style information \\(\\mathbf{X}_B\\), and then a MLP to extract global style codes \\(\\mathbf{z}\\). \\(\\mathbf{X}_B\\) and \\(\\mathbf{z}\\) perform local and global style control on generated images, respectively (Sec. 3.3). To boost the discriminative capacity of style representations, as well as the quality of generated images, we propose a novel contrastive style learning (CSL) method (as shown in Fig 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our settings, the exemplars are drawn by human artists and thus considered as high-quality. In contrast, the images generated in early training stages are typically low-quality. Inspired by the idea of contrastive learning [9], we use the exemplar \\( y_{B} \\) as the positive sample, while a collection of early generated images as negative. Let \\( \\mathbf{z} \\) denotes style codes of the generated image \\( \\hat{x}_B \\), \\( \\mathbf{z}^+ \\) that of exemplar \\( y_{B} \\), and \\( \\{\\mathbf{z}_1^-, \\mathbf{z}_2^-, \\dots, \\mathbf{z}_m^-\\} \\) the style codes of \\( m \\) negative samples. CSL learns style representations by maximizing the mutual information between anchors and positive samples," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "22421" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "while minimizing that between anchors and negative samples. Our contrastive style loss is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.13, + 0.47, + 0.177 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s t y l e} = - \\log \\frac {\\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} ^ {+}}{\\tau}\\right)}{\\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} ^ {+}}{\\tau}\\right) + \\sum_ {j = 1} ^ {m} \\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} _ {j} ^ {-}}{\\tau}\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.184, + 0.469, + 0.215 + ], + "angle": 0, + "content": "where \\(\\tau = 0.07\\) and \\(m = 1024\\). In the implementation, we use a queue to cache negative style vectors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.223, + 0.271, + 0.238 + ], + "angle": 0, + "content": "3.3. Translation network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.469, + 0.442 + ], + "angle": 0, + "content": "To boost both the semantic consistency and style faithfulness, we additionally use source semantic features and global style codes for decoding an image. Specially, we design our whole translation network following U-Net (Fig. 2), where the multi-level features in \\(\\mathcal{E}_A\\) are skip-connected to the decoder \\(\\mathcal{D}_B\\), for supplementing informative semantic structures of the input image \\(x_A\\). Besides, we use the style codes \\(\\mathbf{z}\\) to globally control the style of generated images, in the manner of AdaIN [14]. Specially, \\(\\mathbf{z}\\) is mapped to channel-wise modulating factors by fully-connected (FC) layers. In this way, we diminish the impact of correspondence learning on image generation, and provide reliable style control for even unmatched regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.469, + 0.534 + ], + "angle": 0, + "content": "In summary, our translation network allows both local and global style control, and reuses the semantic features of input images. As a result, the generated image is desired to present consistent semantic to the input \\( x_{A} \\) and faithful style to the exemplar \\( y_{B} \\). More details of our network are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.542, + 0.225, + 0.556 + ], + "angle": 0, + "content": "3.4. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.469, + 0.686 + ], + "angle": 0, + "content": "Our whole network is end-to-end optimized to jointly achieve high-fidelity image generation and accurate correspondence. Following [45], we obtain training triplets \\(\\{x_A, y_B, x_B\\}\\) from the ready-made data pair \\(\\{x_A, x_B\\}\\), where \\(y_B\\) is a geometrically warped version of \\(x_B\\). The generated image is denoted by \\(\\hat{x}_B = \\mathcal{G}(x_A, y_B)\\). Our loss functions are similar to [45], except for the previous contrastive style loss \\(\\mathcal{L}_{style}\\) and the structural loss \\(\\mathcal{L}_{str}\\) below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.746 + ], + "angle": 0, + "content": "Semantic alignment Loss. For accurate cross-domain correspondence learning, the encoders \\(\\mathcal{E}_A\\) and \\(\\mathcal{E}_B\\) should align \\(x_A\\) and \\(x_B\\) to consistent representations. The corresponding semantic alignment loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.757, + 0.469, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\left\\| \\mathcal {E} _ {A} \\left(x _ {A}\\right) - \\mathcal {E} _ {B} \\left(x _ {B}\\right) \\right\\| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.782, + 0.469, + 0.828 + ], + "angle": 0, + "content": "Correspondence Loss. Ideally, if we warp \\( y_{B} \\) in the same way as Eq.3, the resulting image should be exactly \\( x_{B} \\). We thus constrain the learned correspondence by:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.837, + 0.469, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o r r}} = \\left\\| \\tilde {\\mathbf {A}} _ {\\text {m a s k}} ^ {T} y _ {B} \\downarrow - x _ {B} \\downarrow \\right\\| _ {1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "where \\(\\downarrow\\) indicates down-sampling \\(y_{B}\\) and \\(x_{B}\\) to the size (i.e. width and height) of \\(\\mathbf{X}_A\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Perceptual Loss. The generated image \\(\\hat{x}_B\\) should be semantic-consistent with the ground truth \\(x_B\\) in term of semantic. We thus use the perceptual loss:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.147, + 0.892, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p e r c}} = \\left\\| \\varphi_ {l} \\left(\\hat {x} _ {B}\\right) - \\varphi_ {l} \\left(x _ {B}\\right) \\right\\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.172, + 0.892, + 0.202 + ], + "angle": 0, + "content": "where \\(\\varphi_{l}\\) denotes the activations after layer relu4_2 in pretrained VGG19 [40], which represent high-level semantics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.203, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Contextual Loss. In addition, the generated image should be in the same style as the exemplar. In addition to the previous contrastive style loss (Eq.7), we additionally use the contextual loss (CX) [31] to constrain on local style consistency. The contextual loss is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.286, + 0.892, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c t x} = - \\log \\left(\\sum_ {l} w _ {l} \\mathrm {C X} \\left(\\varphi_ {l} \\left(\\hat {x} _ {B}\\right), \\varphi_ {l} \\left(y _ {B}\\right)\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.328, + 0.869, + 0.343 + ], + "angle": 0, + "content": "where \\(w_{l}\\) balances the terms of different VGG19 layers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.344, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Structural Loss. The generated image should preserve semantic structures in the input image. Correspondingly, we use the Learned Perceptual Image Patch Similarity (LPIPS) [46] between their boundaries as the structural loss:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.413, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s t r} = \\operatorname {L P I P S} \\left(\\mathcal {H} \\left(\\hat {x} _ {B}\\right), \\mathcal {H} \\left(x _ {B}\\right)\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.892, + 0.469 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}\\) is the HED algorithm [43], which has been widely used for extracting semantic boundaries in an image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Adversarial loss. Finally, we add a discriminator \\(\\mathcal{D}\\) to distinguish real images in domain \\(\\mathcal{B}\\) and the generated images [7]. The adversarial loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.522, + 0.892, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} ^ {\\mathcal {D}} = - \\mathbb {E} [ h (\\mathcal {D} (y _ {B})) ] - \\mathbb {E} [ h (- \\mathcal {D} (\\hat {x} _ {B})) ], \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.544, + 0.682, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} ^ {\\mathcal {G}} = - \\mathbb {E} [ \\mathcal {D} (\\hat {x} _ {B}) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.571, + 0.89, + 0.586 + ], + "angle": 0, + "content": "where \\( h(t) = \\min (0, - 1 + t) \\) is the hinge loss function [1]." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.587, + 0.89, + 0.601 + ], + "angle": 0, + "content": "Total loss. In summary, our overall objective function is," + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.623, + 0.892, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\mathcal {G}} \\max _ {\\mathcal {D}} \\lambda_ {1} \\mathcal {L} _ {\\text {s t y l e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c o r r}} + \\lambda_ {4} \\mathcal {L} _ {\\text {s t r}} \\tag {14} \\\\ + \\lambda_ {5} \\left(\\mathcal {L} _ {p e r c} + \\mathcal {L} _ {c t x}\\right) + \\lambda_ {6} \\left(\\mathcal {L} _ {a d v} ^ {\\mathcal {G}} + \\mathcal {L} _ {a d v} ^ {\\mathcal {D}}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.675, + 0.764, + 0.689 + ], + "angle": 0, + "content": "where \\(\\lambda\\) denotes the weight parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.625, + 0.711 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.72, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Implementation details. We apply spectral normalization [33] to all the layers in the translation network and discriminator. We use the Adam [20] solver with \\(\\beta_{1} = 0\\) and \\(\\beta_{2} = 0.999\\). The learning rates for the generator and discriminator are set as \\(1e - 4\\) and \\(4e - 4\\) respectively, following TTUR [11]. The experiments are conducted using 4 24GB RTX3090 GPUs. Limited by the computation load, we restrict the resolution of generated images to \\(256\\times 256\\) in all translation tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. We mainly conduct experiments on the following datasets. (1) CelebA-HQ [22] contains 30,000 facial photos. We chose 24,000 samples as the training set" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22422" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.879, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison on the Metfaces [18], CelebA-HQ [22], Ukiyo-e [36], Cartoon [36], AAHQ [24], and DeepFashion [27] datasets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.107, + 0.891, + 0.219 + ], + "angle": 0, + "content": "
CelebA-HQMetfacesCartoonUkiyo-eAAHQDeepFashionTime ↓
FID ↓SWD ↓Texture ↑Color ↑Semantic↑FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓(s)
SPADE [35]31.526.90.9270.9550.92245.626.997.530.545.626.979.432.136.227.80.196
CoCosNet [45]14.315.20.9580.9770.94925.624.366.827.138.313.962.621.914.417.20.321
CoCosNet-v2 [49]13.214.00.9540.9750.94823.322.466.427.032.111.062.422.813.016.71.573
MCL-Net [44]12.814.20.9510.9760.95323.824.567.927.932.412.464.422.212.916.20.309
DynaST [25]12.012.40.9590.9780.95229.228.662.826.538.914.267.224.08.411.80.214
MATEBIT (ours)11.513.20.9660.9860.94926.019.164.427.630.311.556.019.58.210.00.185
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.224, + 0.868, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.621, + 0.868, + 0.635 + ], + "angle": 0, + "content": "Figure 5. Results on the Metfaces [18], CelebA-HQ [22], Ukiyo-e [36], Cartoon [36], AAHQ [24], and DeepFashion [27] datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.471, + 0.837 + ], + "angle": 0, + "content": "and 3000 as the test set. (2) Metfaces [18] consists of 1336 high-quality artistic facial portraits. (3) AAHQ [24] consists of high-quality facial avatars. We randomly select 1500 samples for training and 1000 samples for testing. (4) Ukiyo-e [36] consists of high-quality Ukiyo-e faces. We randomly select 3000 and 1000 samples for training and testing, respectively. (5) Cartoon [36] consists of 317 cartoon faces. (6) DeepFashion [27] consists of 800,00 fashion images. On CelebA-HQ, we connect the face landmarks for face region, and use Canny edge detector to detect edges in the background. On DeepFashion, we use the officially provided landmarks as input. On the other datasets, we use HED [43] to obtain semantic edges." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.839, + 0.367, + 0.855 + ], + "angle": 0, + "content": "4.1. Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We select several advanced models, including SPADE [35], CoCosNet [45], CoCosNet-v2 [49], MCL-Net [44], and DynaST [25], for comparison. For a fair comparison," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.64, + 0.892, + 0.671 + ], + "angle": 0, + "content": "we retrain their models at resolution \\(256 \\times 256\\) under the same settings as ours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.893, + 0.853 + ], + "angle": 0, + "content": "Quantitative evaluation. We adopt several criteria to fully evaluate the generation results. (1) Fréchet Inception Score (FID) [38] and Sliced Wasserstein distance (SWD) [17] are used to evaluate the image perceptual quality. (2) To assess style relevance and semantic consistency of translated images [45], we compute the color, texture, and semantic metrics based on VGG19 [40]. Specifically, the cosine similarities between low-level features (i.e. \\(relu1\\_2\\) and \\(relu2\\_2\\)) are used to measure color and texture relevance, respectively; the average cosine similarity between high-level features (i.e. \\(relu3\\_2\\), \\(relu4\\_2\\), and \\(relu5\\_2\\)) measures the semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The quantitative comparison results are shown in Table 1. Compared to existing methods, our model consistently achieves superior or highly competitive performance across" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22423" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.091, + 0.447, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.47, + 0.434 + ], + "angle": 0, + "content": "Figure 6. More results generated by MATEBIT. Images shown in the left-bottom corner are source images or exemplars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.59 + ], + "angle": 0, + "content": "all the datasets. Especially, MATEBIT significantly improves the style relevance in both texture and color. On the complicated AAHQ dataset, which contains diverse styles of avatars, MATEBIT dramatically decreases both FID and SWD. Such superiority indicates that our generated images are of better perceptual quality; and present consistent appearance to similar parts in exemplars. We additionally report the average time each method costs for generating an image. Our method shows the best efficiency and is significantly faster than previous methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.804 + ], + "angle": 0, + "content": "Qualitative comparison. Fig 5 illustrates images generated by different methods. Obviously, previous methods present geometric distortions, blurring artifacts, inconsistent colors, or identity inconsistency. In contrast, MATEBIT consistently produces appealing results, including more results shown in Fig. 6. Specially, our results preserve the semantic structure of input images, and present consistent appearance with semantically similar regions in exemplars. Previous methods suffer serious degradations mainly due to the matching errors in full correspondence learning. In our method, we distinct reliable and unreliable correspondence, and release the role of matching in image generation. As a result, our method stably transfers a source image to the target style of an exemplar." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.227, + 0.832 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Impacts of MAT. We present a comprehensive analysis to justify the important component in our architecture, i.e. MAT. We here modify our full model by (1) removing the MAT module (i.e. w/o MAT), (2) removing ReLU in" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.691, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.091, + 0.89, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.18, + 0.885, + 0.193 + ], + "angle": 0, + "content": "Figure 7. Impact of the number of MAT blocks on performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.195, + 0.845, + 0.207 + ], + "angle": 0, + "content": "Table 2. Results of ablation studies on MetaFaces." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.209, + 0.89, + 0.366 + ], + "angle": 0, + "content": "
FID ↓SWD ↓Texture ↑Color ↑Semantic↑
w/o MAT27.722.30.8830.8520.856
w/o ReLU30.120.00.9160.9560.928
w/o MAT (→ Full Corr. [45])34.119.70.8720.9690.902
w/o AdaConv (→ FFN [41])34.320.10.8410.9710.896
w/o Lstyle30.320.80.8740.8480.908
w/o Lstyle (→ LCAST [47])31.019.80.9040.9830.921
w/o Lstr28.421.80.9150.9830.911
w/o skip connection47.027.70.9250.9580.902
w/o global style z27.621.20.9270.9610.925
Full Model26.019.10.9380.9840.927
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.644 + ], + "angle": 0, + "content": "MAT (i.e. w/o ReLU), (3) replacing MAT with three-layer full correspondence learning modules [45] (i.e. Full Corr.), and (4) replacing the AdaConv with FFN [45] (i.e. w/ AdaConv). The results in Table 2 show that removing MAT or ReLU dramatically hurts the performance. Besides, using the full correspondence learning in [45] or using FFN also significantly decreases the texture relevance and semantic consistency. Correspondingly, these model variants leads to inferior results in terms of textures or colors, compared to our full model (Fig. 8). Recall the visualized correspondence in Fig. 1, our method learns remarkably accurate correspondence, which ultimately benefits the quality of generated images. In addition, Fig. 7 shows that both the semantic consistency and style realism broadly improve with the number of MAT blocks and peak at three. All these observations demonstrate our motivation that MAT gradually refines cross-domain correspondence and augments informative representations for generating high-quality images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Contrastive Style Loss. To verify the effectiveness of the proposed contrastive style learning methodology, we train our model by (1) removing the style loss (i.e. w/o \\(\\mathcal{L}_{\\text{style}}\\)) and (2) replacing \\(\\mathcal{L}_{\\text{style}}\\) with the loss used in CAST [47] (i.e. w/ \\(\\mathcal{L}_{\\text{CAST}}\\)). In CAST, only high-quality exemplars in different styles are used as negative samples. Differently, we use low-quality generated images in diverse styles as negative samples. From both Table 2 and Fig. 8, we observe that: (1) without \\(\\mathcal{L}_{\\text{style}}\\), although the generated images show high semantic consistency, they present low style relevance; (2) \\(\\mathcal{L}_{\\text{CAST}}\\) benefits the style relevance, but leads to inferior performance to \\(\\mathcal{L}_{\\text{style}}\\). These comparison results meet our expectation that: our CSL methodology enables the learned style codes to discriminate subtle divergences between images with different perceptual qualities. Such discriminability facilitates pushes the network to generate high-quality images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22424" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.162, + 0.768, + 0.175 + ], + "angle": 0, + "content": "Figure 8. Comparison of generated images by different variants of our model, on Metfaces [18]." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.18, + 0.446, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.395, + 0.468, + 0.422 + ], + "angle": 0, + "content": "Figure 9. Our method can transfer facial photos to artistic portraits in the style of exemplars." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.43, + 0.468, + 0.52 + ], + "angle": 0, + "content": "Skip connections & global style control. In MATEBIT, we use skip connections to supplement input semantic information. Removing skip connections dramatically hurts the semantic inconsistency and the quantitative results. Besides, using global style vector \\(\\mathbf{z}\\) increases subtle details, e.g. the colors over the mouth, rings, and hairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.468, + 0.628 + ], + "angle": 0, + "content": "In summary, MAT learns accurate correspondence and enables context-aware feature augmentation; the contrastive style learning benefits the style control and high-quality image generation; and the U-Net architecture helps the preservation of semantic information. Ultimately, all these benefits make our model significantly outperform previous state-of-the-art methods in generating plausible images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.212, + 0.65 + ], + "angle": 0, + "content": "4.3. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.468, + 0.763 + ], + "angle": 0, + "content": "Artistic Portrait Generation. An potential application of our method is transferring a facial photo to an artistic portrait, in the style of an exemplar. We here apply the previously learned models to randomly selected facial photos from CelebA-HQ [22]. As illustrated in Fig. 9, our method can generate appealing portraits with consistent identity and faithful style appearance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Chinese Ink Painting Generation. To verify the capacity of our model in generating complex images, we additionally apply it to generate Chinese Ink paintings. Specifically, we collect paintings of landscapes and facial portraits from the web, and then train and test our model on each subset respectively. Fig. 10 illustrates the results of painting generation and photo-to-painting translation. Obviously, all the generated images show remarkably high quality. Besides, our model successfully captures subtle differences" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.18, + 0.865, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.396, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Figure 10. Chinese ink paintings generation (1st & 3rd rows), as well as photo-to-painting translation (2nd & 4th rows)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.431, + 0.892, + 0.461 + ], + "angle": 0, + "content": "between different exemplars, demonstrating its remarkable capacity in style control." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.627, + 0.485 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.497, + 0.892, + 0.678 + ], + "angle": 0, + "content": "This paper presents a novel exemplar-guided image translation method, dubbed MATEBIT. Both quantitative and qualitative experiments show that MATEBIT is capable of generating high-fidelity images in a number of tasks. Besides, ablation studies demonstrate the effectiveness of MAT and contrastive style learning. Despite such achievements, the artistic portraits transferred from facial photos (Fig. 9) are inferior to those shown in Fig. 6. This may be due to the subtle differences in edge maps between photos and artistic paintings. In the near future, we will explore to solve this issue via semi-supervised learning or domain transfer technologies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.667, + 0.71 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2018AAA0103202; in part by the National Natural Science Foundation of China under Grants 61971172, 61971339, 62176230, U22A2033, U22A2096; in part by the Zhejiang Provincial Science and Technology Program in China under Grant 2021C01108; in part by the NSFC-Zhejiang Joint Fund for the Integration of Industrialization and Informatization under Grant U1909210; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; and in part by Open Research Projects of Zhejiang Lab under Grant 2021KG0AB01." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22425" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[2] Huiwen Chang, Jingwan Lu, Fisher Yu, and Adam Finkelstein. Pairedcyclegan: Asymmetric style transfer for applying and removing makeup. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 40-48, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.503, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[8] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.632, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[10] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11936-11945, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13713-13722, 2021. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Xueqi Hu, Xinyue Zhou, Qiusheng Huang, Zhengyi Shi, Li Sun, and Qingli Li. Qs-attn: Query-selected attention for contrastive learning in i2i translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18291-18300, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[14] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[15] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.275, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[18] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[19] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[21] Thomas N Kipf and Max Welling. Variational graph autoencoders. arXiv preprint arXiv:1611.07308, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549-5558, 2020. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[23] Boyi Li, Felix Wu, Kilian Q Weinberger, and Serge Belongie. Positional normalization. In Advances in Neural Information Processing Systems, pages 1620-1632, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[24] Mingcong Liu, Qiang Li, Zekui Qin, Guoxin Zhang, Pengfei Wan, and Wen Zheng. Blendgan: implicitly gan blending for arbitrary stylized face generation. Advances in Neural Information Processing Systems, 34:29710-29722, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.777, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[25] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In European Conference on Computer Vision, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22426" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[28] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[29] Mandi Luo, Haoxue Wu, Huaibo Huang, Weizan He, and Ran He. Memory-modulated transformer network for heterogeneous face recognition. IEEE Transactions on Information Forensics and Security, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[30] Liqian Ma, Xu Jia, Stamatios Georgoulis, Tinne Tuytelaars, and Luc Van Gool. Exemplar guided unsupervised image-to-image translation with semantic consistency. Proceedings ICLR 2019, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[31] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In Proceedings of the European conference on computer vision (ECCV), pages 768-783, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[32] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel flow: Semantic correspondence with multi-layer neural features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3395-3404, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[33] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[34] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In European conference on computer vision, pages 319-345. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.649 + ], + "angle": 0, + "content": "[35] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[36] Justin N. M. Pinkney. Aligned ukiyo-e faces dataset, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[38] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, August 2020. version 0.2.1. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[39] Junyoung Seo, Gyuseong Lee, Seokju Cho, Jiyoung Lee, and Seungryong Kim. Midms: Matching interleaved diffusion models for exemplar-based image translation. arXiv preprint arXiv:2209.11047, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR, 2015. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[42] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision, pages 63-79. Springer, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[43] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[44] Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, and Changgong Zhang. Marginal contrastive correspondence for guided image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10663-10672, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.893, + 0.402 + ], + "angle": 0, + "content": "[45] Pan Zhang, Bo Zhang, Dong Chen, Lu Yuan, and Fang Wen. Cross-domain correspondence learning for exemplar-based image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5143-5153, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.893, + 0.472 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.473, + 0.893, + 0.53 + ], + "angle": 0, + "content": "[47] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.893, + 0.599 + ], + "angle": 0, + "content": "[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.893, + 0.683 + ], + "angle": 0, + "content": "[49] Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocosnet v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11465–11475, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.685, + 0.893, + 0.754 + ], + "angle": 0, + "content": "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22427" + } + ] +] \ No newline at end of file diff --git a/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_origin.pdf b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a78ac29bc41c5d489905d8c33bc252a3ff81dbb7 --- /dev/null +++ b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/122f9675-10c2-4e31-827b-15924283ce1b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7aeb5f39563fb5bff56a9fee1e22eb8bd2103bde8413bef30e883bb23e7078b +size 3998710 diff --git a/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/full.md b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7adf88a90a240a40afb083a047f8e52c2c2eecf --- /dev/null +++ b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/full.md @@ -0,0 +1,345 @@ +# Masked and Adaptive Transformer for Exemplar Based Image Translation + +Chang Jiang $^{1}$ , Fei Gao $^{1,2*}$ , Biao Ma $^{1}$ , Yuhao Lin $^{1}$ , Nannan Wang $^{3}$ , Gang Xu $^{1}$ + +1 School of Computer Science and Technology, Hangzhou Dianzi University + +$^{2}$ Hangzhou Institute of Technology, Xidian University $^{3}$ ISN State Key Laboratory, Xidian University + +{jc233, aiartma, linyh, gxu}@hdu.edu.cn, {fgao, nnwang}@xidian.edu.cn + +# Abstract + +We present a novel framework for exemplar based image translation. Recent advanced methods for this task mainly focus on establishing cross-domain semantic correspondence, which sequentially dominates image generation in the manner of local style control. Unfortunately, cross-domain semantic matching is challenging; and matching errors ultimately degrade the quality of generated images. To overcome this challenge, we improve the accuracy of matching on the one hand, and diminish the role of matching in image generation on the other hand. To achieve the former, we propose a masked and adaptive transformer (MAT) for learning accurate cross-domain correspondence, and executing context-aware feature augmentation. To achieve the latter, we use source features of the input and global style codes of the exemplar, as supplementary information, for decoding an image. Besides, we devise a novel contrastive style learning method, for acquire quality-discriminative style representations, which in turn benefit high-quality image generation. Experimental results show that our method, dubbed MATEBIT, performs considerably better than state-of-the-art methods, in diverse image translation tasks. The codes are available at https://github.com/AiArt-HDU/MATEBIT. + +# 1. Introduction + +Image-to-image translation aims at transfer images in a source domain to a target domain [16, 50]. Early studies learn mappings directly by Generating Adversarial Networks (GANs), and have shown great success in various applications [2, 42]. Recently, exemplar based image translation [29, 30, 45], where an exemplar image is used to control the style of translated images, has attracted a lot of attention. Such methods allow high flexibility and controllability, and have a wide range of potential applications in social networks and metaverse. For example, people can transfer a + +![](images/73cb608b03c9debbdc14df50dd542ab2a27abb3dfaed97834c96cb8fed1dec58.jpg) +Figure 1. Visualization of correspondence maps. The red point is the query position. Full Corr. and Masked Corr. denote the full correspondence [45] and masked one in our method, respectively. CAM denotes visualization by Class Activation Mapping [48]. + +facial sketch to an artistic portrait, in the style of oil paintings or avatars. Despite the remarkable progress, yielding high-fidelity images with consistent semantic and faithful styles remains a grand challenge. + +Early pioneering works [15, 21, 35] attempt to globally control the style of generated images. However, such methods ignore spatial correlations between an input image and an exemplar, and may fail to produce faithful details. Recently, some advanced methods [25, 44, 45, 49] first establish the cross-domain semantic correspondence between an input image and an exemplar, and then use it to warp the exemplar for controlling local style patterns. In these methods, the quality of generated images relies heavily on the learned correspondence [39]. Unfortunately, cross-domain semantic matching is challenging, since there is no reliable supervision on correspondence learning [45]. As a result, potential matching errors ultimately lead to degraded artifacts in generated images. + +To combat this challenge, we propose to boost the matching accuracy on one hand, and to diminish the role of matching in image generation on the other hand. Inspired by the great success of Transformers [6, 10, 26, 41], we first devise a Masked and Adaptive Transformer (MAT) for learning ac + +curate cross-domain correspondence and executing context-aware feature augmentation. Previous works [44, 45, 49] have used the vanilla attention mechanism [41] for learning full correspondence. However, the initial attention typically involves ambiguous correspondences (2nd row in Fig. 1). To mitigate these limitations, in MAT, we use a masked attention to distinguish the correspondence as reliable or not, and then reliability-adaptively aggregate representations. Besides, the Feed-Forward Network (FFN) [41] in vanilla transformers neglects contextual correlations inside an image. We thus replace FFN by an adaptive convolution block [28], where the coordinate attention [12] and depth-wise separable convolution [5] are used to capture contextual correlations and to improve efficiency. With a joint consideration of matching reliability and contextual correlations, MAT gradually focuses on accurate correspondences and emphasizes on features of interest (3rd row in Fig. 1). + +In addition, to boost both the semantic consistency and style faithfulness, we supplementally use semantic features of the input image and global style codes of the exemplar for decoding an image. To this end, we first design our whole network following the U-Net architecture [16]. Besides, we devise a novel contrastive style learning (CSL) framework for acquiring discriminative style representations. Recently, Zhang et al. [47] propose a similar CSL method, where the target exemplar is used as a positive sample, and the other exemplars as negative ones. Differently, we use low-quality images, generated during early training stages, as negative samples. In this way, our style codes are desired to discriminate not only subtle differences in style, but also those in perceptual quality. Ultimately, the learned global style codes, cooperating with the local style control induced by MAT, in turn benefit high-quality image generation. + +With the proposed techniques above, our full model, dubbed MATEBIT, diminishes the impact of position-wise matching on image quality, and integrates both local and global style control for image generation. Experimental results show that MATEBIT generates considerably more plausible images than previous state-of-the-art methods, in diverse image translation tasks. In addition, comprehensive ablation studies demonstrate the effectiveness of our proposed components. Finally, we perform interesting applications of photo-to-painting translation and Chinese ink paintings generation. + +# 2. Relate Work + +Exemplar Based Image Translation. Recently, exemplar based image translation has attracted increasing attention. For example, Park et al. [35] learn an encoder to map the exemplar image into a global style vector, and use it to guide image generation. Such a global style control strategy enables style consistency in whole, but fails to produce subtle details. Most recently, researchers propose a + +matching-then-generation framework [39]. Specially, they first establish dense correspondence between an input and an exemplar, and then reshuffle the exemplar for locally control the style of synthesize images. For example, Zhang et al. [45] establish position-wise correspondence based on the Cosine attention mechanism and warp the exemplar correspondingly. Afterwards, the warped image dominates the generation of images in the manner of SPADE [35]. To reduce the cost of matching in high-resolution image generation, Zhou et al. [49] introduce a hierarchical refinement of semantic correspondence from ConvGRU-PatchMatch. Besides, Liu et al. [25] used a dynamic pruning method for learning hierarchical sparse correspondence. They also use reliability-adaptive feature integration to improve the quality of generated images. + +Previous methods merely use global or local style control, and the latter relies heavily on the learned correspondence. Besides, they consider little about contextual correlations inside an image. In this paper, we use both global and local style control to boost the style consistency. Besides, we take contextual correlations into consideration and execute reliability-adaptive feature augmentation. + +Transformers. Transformers [41] have shown incredible success from the field of natural language processing (NLP) [19] to computer vision (CV) [6, 26]. Multihead attention (MHA) and FFN are key components in a Transformer, and have been used in exemplar based image translation. However, they induce unreliable matching results and neglect context correlations in feature translation. In our MAT, we combat these limitations by replacing them with a masked attention and a context-aware convolution block, respectively. Recently, researchers use semantic masks to facilitate representation learning [4, 8, 37], where a mask predictor is required. Differently, we use a ReLU function to mask over the attention layer, for distinguishing correspondence as reliable or not (Sec. 3.1). In general, MAT follows a concise and efficient architecture. + +Contrastive Learning. Contrastive learning has shown its effectiveness in various computer vision tasks [9, 13, 34]. The basic idea is to learn a representation by pushing positive samples toward an anchor, and moving negative samples away from it. Different sampling strategies and contrastive losses have been extensively explored in various downstream tasks. For example, Chen et al. [3] and He et al. [9] obtain positive samples by augmenting original data. In the field of image translation, Park et al. [34] propose patchwise contrastive learning by maximizing the mutual information between cross-domain patches. Similarly, Zhang et al. [47] use contrastive learning for acquiring discriminative style representations. In the task of exemplar based image translation, Zhan et al. [44] use contrastive learning to align cross-domain images to a consistent semantic feature space, so as to boost the accuracy of matching. Differently, we + +![](images/87cb6bd4342918ba51234ed1a6d836c2b00741bcfcd42f8778a9cb1df09b4e05.jpg) +Figure 2. Overview of our image translation network, MATEBIT. + +use early generated images as negative samples, so that the learned style representations can discriminate subtle differences in both style and perceptual quality (Sec. 3.2). + +# 3. The Proposed Method + +Given an input image $x_{A}$ in domain $\mathcal{A}$ and an exemplar image $y_{B}$ in domain $\mathcal{B}$ , our goal is to generate a target image $x_{B}$ which preserves semantic structures in $x_{A}$ but resembles the style of similar parts in $y_{B}$ . Fig. 2 shows an overview of our translation network $\mathcal{G}$ . Specially, we first align $x_{A}$ and $y_{B}$ to an intermediate feature space by encoders $\mathcal{E}_{A}$ and $\mathcal{E}_{B}$ , respectively. Afterwards, we use a Masked and Adaptive Transformer (MAT) for correspondence learning and feature augmentation. Finally, a decoder $\mathcal{D}_B$ produces an output image $\hat{x}_B$ based on the augmented features, as well as the source features and target style codes. Details are described below. + +# 3.1. Masked and Adaptive Transformer (MAT) + +In order to establish accurate cross-domain correspondence, we propose a novel and concise Transformer architecture, i.e. MAT. In general, the architecture of MAT (Fig. 3b) follows that of vanilla Transformers (Fig. 3a) [41]. Differently, we use masked attention to distinguish reliable and unreliable correspondence, instead of using multi-head attention. Besides, we use Positional Normalization (PONO) [23] and an Adaptive Convolution (AdaConv) block [28], instead of LN and MLP-based FFN, respectively. MAT is desired to gradually concentrate on accurate matching, and to reliability-adaptively augment representations with contextual correlations. + +Masked Correspondence Learning. Let $\mathbf{X}_A \in \mathbb{R}^{H \times W \times C}$ and $\mathbf{Y}_B \in \mathbb{R}^{H \times W \times C}$ be the representations of $x_A$ and $y_B$ in the intermediate feature space, with height $H$ , width $W$ , and $C$ channels. We first map $\mathbf{X}_A$ to the query $\mathbf{Q} \in \mathbb{R}^{HW \times C}$ , and $\mathbf{Y}_B$ to the key $\mathbf{K} \in \mathbb{R}^{HW \times C}$ and value $\mathbf{V} \in \mathbb{R}^{HW \times C}$ , by using $1 \times 1$ convolutions, respectively. As shown in Fig. 3d, we add positional encoding (PE) to $\mathbf{X}_A$ and $\mathbf{Y}_B$ , for embedding spatial correlations. Afterwards, we learn the initial correspondence $\mathbf{A} \in \mathbb{R}^{HW \times HW}$ fol + +![](images/de343606c827cbfaccca0f7d431dccf574b28466ed6aa0ab5a2d4facc7f444d3.jpg) + +![](images/16099285aba2a1f5025a9aa51abd3a44a96f4e9e4d85a700a372bfd3c1377d79.jpg) + +![](images/ee95e5984517f9afcf7e7686b851fc64c55b65c7ed5d4fbbfb4862faea5a3984.jpg) + +![](images/cce8d1a41fd26826f80744de106158f1377f8e06284d1222f113314602c08c06.jpg) +Figure 3. Detailed architectures. (a) Vanilla Transformer block, (b) MAT block, (c) AdaConv block, and (d) Masked Attention. + +lowing the Cosine attention mechanism [45], i.e. + +$$ +\mathbf {A} (u, v) = \frac {\tilde {\mathbf {Q}} (u) \tilde {\mathbf {K}} (v) ^ {T}}{| | \tilde {\mathbf {Q}} (u) | | \cdot | | \tilde {\mathbf {K}} (v) | |}, \tag {1} +$$ + +with $\tilde{\mathbf{Q}}(u) = \mathbf{Q}(u) - \bar{\mathbf{Q}}(u)$ , $\tilde{\mathbf{K}}(v) = \mathbf{K}(v) - \bar{\mathbf{K}}(v)$ , where $u, v \in [1, \dots, HW]$ are position indices; $\bar{\mathbf{Q}}(u)$ and $\bar{\mathbf{K}}(v)$ are the means of $\mathbf{Q}(u)$ and $\mathbf{K}(v)$ , respectively. $\mathbf{A}(u, v)$ is the matching score between $\mathbf{Q}(u)$ and $\mathbf{K}(v)$ . + +Previous methods [44, 45] typically use the initial correspondence map $\mathbf{A}$ to reshuffle an exemplar for controlling local patterns in image synthesis. However, induced by the difficulties in cross-domain correspondence learning, $\mathbf{A}$ involves unreliable match scores (Fig. 3d). As a result, the reshuffled image will lead to implausible artifacts in generated images. To combat this limitation, we distinguish initial matching scores as reliable or not, according to their signs [32]. The masked correspondence map becomes: + +$$ +\mathbf {A} _ {\text {m a s k}} = \operatorname {R e L U} (\mathbf {A}), \tag {2} +$$ + +In DynaST [25], two networks are used to predict the reliability mask of correspondence. However, it's challenging to effectively train the network, because there is no super + +vision on matching during training. In contrast, ReLU contains no learnable parameters and ultimately leads to superior performance over DynaST (Sec. 4.1). + +Reliability-Adaptive Feature Aggregation. For regions with reliable correspondence in $x_{A}$ , we use $\mathbf{A}_{mask}$ to warp the value features, $\mathbf{V}$ , derived from the exemplar: + +$$ +\mathbf {X} _ {\text {c o r}} = \tilde {\mathbf {A}} _ {\text {m a s k}} \mathbf {V}, \text {w i t h} \tilde {\mathbf {A}} _ {\text {m a s k}} = \operatorname {s o f t m a x} (\alpha \cdot \mathbf {A} _ {\text {m a s k}}), \tag {3} +$$ + +where $\alpha$ is a scaling coefficient to control the sharpness of the softmax function. In default, we set its value as 100. + +For regions with unreliable correspondence in $x_{A}$ , $\mathbf{X}_{cor}$ provides an average style representation of $\mathbf{V}$ . We further extract complementary information from the query, $\mathbf{Q}$ , derived from the input. Inspired by SPADE [35], we first transfer $\mathbf{Q}$ to the target domain by using pixel-wise modulation parameters (i.e., $\gamma$ for scale and $\beta$ for bias) learned from $x_{A}$ . The modulation is formulated by: + +$$ +\mathbf {Q} _ {\text {n o r m}} = \gamma (x _ {A}) \frac {\mathbf {Q} - \mu (\mathbf {Q})}{\sigma (\mathbf {Q})} + \beta (x _ {A}), \tag {4} +$$ + +where $\mu (\mathbf{Q})$ and $\sigma (\mathbf{Q})$ are the mean value and standard deviance of $\mathbf{Q}$ . Afterwards, we select the translated features of unreliable corresponded regions in $x_{A}$ by: + +$$ +\mathbf {X} _ {u n c o r} = \left(1 - \sum_ {j} \mathbf {A} _ {m a s k}\right) \odot \mathbf {Q} _ {n o r m}, \tag {5} +$$ + +where the summation is along the second dimension; $\odot$ denotes point-wise production with broadcasting. Since $\gamma$ and $\beta$ are learned from the input image $x_{A}$ , the modulated features preserve the semantic information of $x_{A}$ . Besides, constraints on the generated image will push the selected features convey to the style of $y_{B}$ . + +Ideally, $\mathbf{X}_{cor}$ and $\mathbf{X}_{uncor}$ would complement each other and facilitate both semantic consistency and style relevance in image generation. To this end, we integrate $\mathbf{X}_{cor}$ , $\mathbf{X}_{uncor}$ , and $\mathbf{Q}$ by: + +$$ +\mathbf {X} _ {a g g} = \operatorname {P O N O} \left(\mathbf {X} _ {c o r} + \mathbf {X} _ {u n c o r} + \mathbf {Q}\right). \tag {6} +$$ + +In PONO [23], features at each position are normalized dependently. Compared to LN in vanilla transformers and DynaST [25], PONO boosts the flexibility in reliability-adaptive feature aggregation. + +Context-Aware Feature Augmentation. Inspired by ConvNeXT [28], we replace FFN by an AdaConv block to position-adaptively emphasize informative representations. Besides, we use the coordinate attention (CoordAten) module [12] to capture contextual correlations. + +The architecture of the AdaConv block is as shown in Fig. 3c. We fist use the depthwise convolution (Dwise) to update representations in each channel separately; and then use two pointwise convolutions (Pwise) to automatically emphasize representations of interest, at every position. The Gaussian Error Linear Unit (GELU) activation + +![](images/92a1c60f1844b473bacc6475f76c30344beaf7bd2d8d2b1c76a8e42dcfd3c97f.jpg) +Figure 4. Contrastive style learning. The memory bank consists of divergent low-quality images generated in early training stages. + +function and Layer Norm (LN) are used after the first Pwise layer [28]. Notably, Coordinate is used after the Dwise layer for modeling long-range dependencies in an image. Specially, Coordinate produces cross-channel and position-sensitive attention maps, which helps our model to more accurately locate the representations of interest [12]. + +Finally, the output of a MAT block is obtained with a residual connection, i.e. $\mathbf{X}_{\mathrm{MAT}} = \mathrm{AdaConv}(\mathbf{X}_{agg}) + \mathbf{X}_{agg}$ . In the implementation, we stack three MAT blocks in default to gradually refine the correspondence and to augment informative representations (Fig. 1). Empirical verifications will be given in Sec. 4.2. + +Benefits of MAT. Fig. 3d illustrates the impact of MAT. The query point locates over the left eye of the source image. Here we show the magnitudes of its correspondence over the exemplar, in the third layer of MAT. Obviously, the original correspondence $\mathbf{A}$ covers both eyes of the exemplar. In contrast, the masked correspondence $\mathbf{A}_{mask}$ accurately concentrates over the left eye. Such superiority significantly boost the quality of ultimate images. + +# 3.2. Contrastive Style Learning (CSL) + +In MATEBIT, we use the encoder $\mathcal{E}_B$ to extract local style information $\mathbf{X}_B$ , and then a MLP to extract global style codes $\mathbf{z}$ . $\mathbf{X}_B$ and $\mathbf{z}$ perform local and global style control on generated images, respectively (Sec. 3.3). To boost the discriminative capacity of style representations, as well as the quality of generated images, we propose a novel contrastive style learning (CSL) method (as shown in Fig 4). + +In our settings, the exemplars are drawn by human artists and thus considered as high-quality. In contrast, the images generated in early training stages are typically low-quality. Inspired by the idea of contrastive learning [9], we use the exemplar $y_{B}$ as the positive sample, while a collection of early generated images as negative. Let $\mathbf{z}$ denotes style codes of the generated image $\hat{x}_B$ , $\mathbf{z}^+$ that of exemplar $y_{B}$ , and $\{\mathbf{z}_1^-, \mathbf{z}_2^-, \dots, \mathbf{z}_m^-\}$ the style codes of $m$ negative samples. CSL learns style representations by maximizing the mutual information between anchors and positive samples, + +while minimizing that between anchors and negative samples. Our contrastive style loss is computed by: + +$$ +\mathcal {L} _ {s t y l e} = - \log \frac {\exp \left(\frac {\mathbf {z} ^ {T} \mathbf {z} ^ {+}}{\tau}\right)}{\exp \left(\frac {\mathbf {z} ^ {T} \mathbf {z} ^ {+}}{\tau}\right) + \sum_ {j = 1} ^ {m} \exp \left(\frac {\mathbf {z} ^ {T} \mathbf {z} _ {j} ^ {-}}{\tau}\right)}, \tag {7} +$$ + +where $\tau = 0.07$ and $m = 1024$ . In the implementation, we use a queue to cache negative style vectors. + +# 3.3. Translation network + +To boost both the semantic consistency and style faithfulness, we additionally use source semantic features and global style codes for decoding an image. Specially, we design our whole translation network following U-Net (Fig. 2), where the multi-level features in $\mathcal{E}_A$ are skip-connected to the decoder $\mathcal{D}_B$ , for supplementing informative semantic structures of the input image $x_A$ . Besides, we use the style codes $\mathbf{z}$ to globally control the style of generated images, in the manner of AdaIN [14]. Specially, $\mathbf{z}$ is mapped to channel-wise modulating factors by fully-connected (FC) layers. In this way, we diminish the impact of correspondence learning on image generation, and provide reliable style control for even unmatched regions. + +In summary, our translation network allows both local and global style control, and reuses the semantic features of input images. As a result, the generated image is desired to present consistent semantic to the input $x_{A}$ and faithful style to the exemplar $y_{B}$ . More details of our network are available in the supplementary material. + +# 3.4. Loss functions + +Our whole network is end-to-end optimized to jointly achieve high-fidelity image generation and accurate correspondence. Following [45], we obtain training triplets $\{x_A, y_B, x_B\}$ from the ready-made data pair $\{x_A, x_B\}$ , where $y_B$ is a geometrically warped version of $x_B$ . The generated image is denoted by $\hat{x}_B = \mathcal{G}(x_A, y_B)$ . Our loss functions are similar to [45], except for the previous contrastive style loss $\mathcal{L}_{style}$ and the structural loss $\mathcal{L}_{str}$ below. + +Semantic alignment Loss. For accurate cross-domain correspondence learning, the encoders $\mathcal{E}_A$ and $\mathcal{E}_B$ should align $x_A$ and $x_B$ to consistent representations. The corresponding semantic alignment loss is: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \left\| \mathcal {E} _ {A} \left(x _ {A}\right) - \mathcal {E} _ {B} \left(x _ {B}\right) \right\| _ {1}. \tag {8} +$$ + +Correspondence Loss. Ideally, if we warp $y_{B}$ in the same way as Eq.3, the resulting image should be exactly $x_{B}$ . We thus constrain the learned correspondence by: + +$$ +\mathcal {L} _ {\text {c o r r}} = \left\| \tilde {\mathbf {A}} _ {\text {m a s k}} ^ {T} y _ {B} \downarrow - x _ {B} \downarrow \right\| _ {1}, \tag {9} +$$ + +where $\downarrow$ indicates down-sampling $y_{B}$ and $x_{B}$ to the size (i.e. width and height) of $\mathbf{X}_A$ . + +Perceptual Loss. The generated image $\hat{x}_B$ should be semantic-consistent with the ground truth $x_B$ in term of semantic. We thus use the perceptual loss: + +$$ +\mathcal {L} _ {\text {p e r c}} = \left\| \varphi_ {l} \left(\hat {x} _ {B}\right) - \varphi_ {l} \left(x _ {B}\right) \right\| _ {1}, \tag {10} +$$ + +where $\varphi_{l}$ denotes the activations after layer relu4_2 in pretrained VGG19 [40], which represent high-level semantics. + +Contextual Loss. In addition, the generated image should be in the same style as the exemplar. In addition to the previous contrastive style loss (Eq.7), we additionally use the contextual loss (CX) [31] to constrain on local style consistency. The contextual loss is computed by: + +$$ +\mathcal {L} _ {c t x} = - \log \left(\sum_ {l} w _ {l} \mathrm {C X} \left(\varphi_ {l} \left(\hat {x} _ {B}\right), \varphi_ {l} \left(y _ {B}\right)\right)\right) \tag {11} +$$ + +where $w_{l}$ balances the terms of different VGG19 layers. + +Structural Loss. The generated image should preserve semantic structures in the input image. Correspondingly, we use the Learned Perceptual Image Patch Similarity (LPIPS) [46] between their boundaries as the structural loss: + +$$ +\mathcal {L} _ {s t r} = \operatorname {L P I P S} \left(\mathcal {H} \left(\hat {x} _ {B}\right), \mathcal {H} \left(x _ {B}\right)\right), \tag {12} +$$ + +where $\mathcal{H}$ is the HED algorithm [43], which has been widely used for extracting semantic boundaries in an image. + +Adversarial loss. Finally, we add a discriminator $\mathcal{D}$ to distinguish real images in domain $\mathcal{B}$ and the generated images [7]. The adversarial loss is: + +$$ +\mathcal {L} _ {a d v} ^ {\mathcal {D}} = - \mathbb {E} [ h (\mathcal {D} (y _ {B})) ] - \mathbb {E} [ h (- \mathcal {D} (\hat {x} _ {B})) ], \tag {13} +$$ + +$$ +\mathcal {L} _ {a d v} ^ {\mathcal {G}} = - \mathbb {E} [ \mathcal {D} (\hat {x} _ {B}) ], +$$ + +where $h(t) = \min (0, - 1 + t)$ is the hinge loss function [1]. + +Total loss. In summary, our overall objective function is, + +$$ +\begin{array}{l} \min _ {\mathcal {G}} \max _ {\mathcal {D}} \lambda_ {1} \mathcal {L} _ {\text {s t y l e}} + \lambda_ {2} \mathcal {L} _ {\text {a l i g n}} + \lambda_ {3} \mathcal {L} _ {\text {c o r r}} + \lambda_ {4} \mathcal {L} _ {\text {s t r}} \tag {14} \\ + \lambda_ {5} \left(\mathcal {L} _ {p e r c} + \mathcal {L} _ {c t x}\right) + \lambda_ {6} \left(\mathcal {L} _ {a d v} ^ {\mathcal {G}} + \mathcal {L} _ {a d v} ^ {\mathcal {D}}\right) \\ \end{array} +$$ + +where $\lambda$ denotes the weight parameters. + +# 4. Experiment + +Implementation details. We apply spectral normalization [33] to all the layers in the translation network and discriminator. We use the Adam [20] solver with $\beta_{1} = 0$ and $\beta_{2} = 0.999$ . The learning rates for the generator and discriminator are set as $1e - 4$ and $4e - 4$ respectively, following TTUR [11]. The experiments are conducted using 4 24GB RTX3090 GPUs. Limited by the computation load, we restrict the resolution of generated images to $256\times 256$ in all translation tasks. + +Datasets. We mainly conduct experiments on the following datasets. (1) CelebA-HQ [22] contains 30,000 facial photos. We chose 24,000 samples as the training set + +Table 1. Comparison on the Metfaces [18], CelebA-HQ [22], Ukiyo-e [36], Cartoon [36], AAHQ [24], and DeepFashion [27] datasets. + +
CelebA-HQMetfacesCartoonUkiyo-eAAHQDeepFashionTime ↓
FID ↓SWD ↓Texture ↑Color ↑Semantic↑FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓(s)
SPADE [35]31.526.90.9270.9550.92245.626.997.530.545.626.979.432.136.227.80.196
CoCosNet [45]14.315.20.9580.9770.94925.624.366.827.138.313.962.621.914.417.20.321
CoCosNet-v2 [49]13.214.00.9540.9750.94823.322.466.427.032.111.062.422.813.016.71.573
MCL-Net [44]12.814.20.9510.9760.95323.824.567.927.932.412.464.422.212.916.20.309
DynaST [25]12.012.40.9590.9780.95229.228.662.826.538.914.267.224.08.411.80.214
MATEBIT (ours)11.513.20.9660.9860.94926.019.164.427.630.311.556.019.58.210.00.185
+ +![](images/702a17053195b99e87c84d12b12e78855cb6452160ec0e42efbfcfbbed3d057a.jpg) +Figure 5. Results on the Metfaces [18], CelebA-HQ [22], Ukiyo-e [36], Cartoon [36], AAHQ [24], and DeepFashion [27] datasets. + +and 3000 as the test set. (2) Metfaces [18] consists of 1336 high-quality artistic facial portraits. (3) AAHQ [24] consists of high-quality facial avatars. We randomly select 1500 samples for training and 1000 samples for testing. (4) Ukiyo-e [36] consists of high-quality Ukiyo-e faces. We randomly select 3000 and 1000 samples for training and testing, respectively. (5) Cartoon [36] consists of 317 cartoon faces. (6) DeepFashion [27] consists of 800,00 fashion images. On CelebA-HQ, we connect the face landmarks for face region, and use Canny edge detector to detect edges in the background. On DeepFashion, we use the officially provided landmarks as input. On the other datasets, we use HED [43] to obtain semantic edges. + +# 4.1. Comparison with state-of-the-art + +We select several advanced models, including SPADE [35], CoCosNet [45], CoCosNet-v2 [49], MCL-Net [44], and DynaST [25], for comparison. For a fair comparison, + +we retrain their models at resolution $256 \times 256$ under the same settings as ours. + +Quantitative evaluation. We adopt several criteria to fully evaluate the generation results. (1) Fréchet Inception Score (FID) [38] and Sliced Wasserstein distance (SWD) [17] are used to evaluate the image perceptual quality. (2) To assess style relevance and semantic consistency of translated images [45], we compute the color, texture, and semantic metrics based on VGG19 [40]. Specifically, the cosine similarities between low-level features (i.e. $relu1\_2$ and $relu2\_2$ ) are used to measure color and texture relevance, respectively; the average cosine similarity between high-level features (i.e. $relu3\_2$ , $relu4\_2$ , and $relu5\_2$ ) measures the semantic consistency. + +The quantitative comparison results are shown in Table 1. Compared to existing methods, our model consistently achieves superior or highly competitive performance across + +![](images/82cbc5b4d8f3d6e09a10fbeee5607b7de4e024ff83dcd007bee5ad55eb025ed8.jpg) +Figure 6. More results generated by MATEBIT. Images shown in the left-bottom corner are source images or exemplars. + +all the datasets. Especially, MATEBIT significantly improves the style relevance in both texture and color. On the complicated AAHQ dataset, which contains diverse styles of avatars, MATEBIT dramatically decreases both FID and SWD. Such superiority indicates that our generated images are of better perceptual quality; and present consistent appearance to similar parts in exemplars. We additionally report the average time each method costs for generating an image. Our method shows the best efficiency and is significantly faster than previous methods. + +Qualitative comparison. Fig 5 illustrates images generated by different methods. Obviously, previous methods present geometric distortions, blurring artifacts, inconsistent colors, or identity inconsistency. In contrast, MATEBIT consistently produces appealing results, including more results shown in Fig. 6. Specially, our results preserve the semantic structure of input images, and present consistent appearance with semantically similar regions in exemplars. Previous methods suffer serious degradations mainly due to the matching errors in full correspondence learning. In our method, we distinct reliable and unreliable correspondence, and release the role of matching in image generation. As a result, our method stably transfers a source image to the target style of an exemplar. + +# 4.2. Ablation study + +Impacts of MAT. We present a comprehensive analysis to justify the important component in our architecture, i.e. MAT. We here modify our full model by (1) removing the MAT module (i.e. w/o MAT), (2) removing ReLU in + +![](images/202e9d72426fd8a879453b556e7bd83ccebd505dd8eed053b88a7a08b54b5432.jpg) + +![](images/1aa06cd1a4d6a58884dd41fa220dfeb8b76705ac2b172a9325b15123d222ba6c.jpg) +Figure 7. Impact of the number of MAT blocks on performance. + +Table 2. Results of ablation studies on MetaFaces. + +
FID ↓SWD ↓Texture ↑Color ↑Semantic↑
w/o MAT27.722.30.8830.8520.856
w/o ReLU30.120.00.9160.9560.928
w/o MAT (→ Full Corr. [45])34.119.70.8720.9690.902
w/o AdaConv (→ FFN [41])34.320.10.8410.9710.896
w/o Lstyle30.320.80.8740.8480.908
w/o Lstyle (→ LCAST [47])31.019.80.9040.9830.921
w/o Lstr28.421.80.9150.9830.911
w/o skip connection47.027.70.9250.9580.902
w/o global style z27.621.20.9270.9610.925
Full Model26.019.10.9380.9840.927
+ +MAT (i.e. w/o ReLU), (3) replacing MAT with three-layer full correspondence learning modules [45] (i.e. Full Corr.), and (4) replacing the AdaConv with FFN [45] (i.e. w/ AdaConv). The results in Table 2 show that removing MAT or ReLU dramatically hurts the performance. Besides, using the full correspondence learning in [45] or using FFN also significantly decreases the texture relevance and semantic consistency. Correspondingly, these model variants leads to inferior results in terms of textures or colors, compared to our full model (Fig. 8). Recall the visualized correspondence in Fig. 1, our method learns remarkably accurate correspondence, which ultimately benefits the quality of generated images. In addition, Fig. 7 shows that both the semantic consistency and style realism broadly improve with the number of MAT blocks and peak at three. All these observations demonstrate our motivation that MAT gradually refines cross-domain correspondence and augments informative representations for generating high-quality images. + +Contrastive Style Loss. To verify the effectiveness of the proposed contrastive style learning methodology, we train our model by (1) removing the style loss (i.e. w/o $\mathcal{L}_{\text{style}}$ ) and (2) replacing $\mathcal{L}_{\text{style}}$ with the loss used in CAST [47] (i.e. w/ $\mathcal{L}_{\text{CAST}}$ ). In CAST, only high-quality exemplars in different styles are used as negative samples. Differently, we use low-quality generated images in diverse styles as negative samples. From both Table 2 and Fig. 8, we observe that: (1) without $\mathcal{L}_{\text{style}}$ , although the generated images show high semantic consistency, they present low style relevance; (2) $\mathcal{L}_{\text{CAST}}$ benefits the style relevance, but leads to inferior performance to $\mathcal{L}_{\text{style}}$ . These comparison results meet our expectation that: our CSL methodology enables the learned style codes to discriminate subtle divergences between images with different perceptual qualities. Such discriminability facilitates pushes the network to generate high-quality images. + +![](images/13f1705bb5463520e19bff57e698365e044b700db4c94da9ab6f774b55814bb6.jpg) +Figure 8. Comparison of generated images by different variants of our model, on Metfaces [18]. + +![](images/1180542e4d15edc9093f885a172b0d4844e0d428d69c5388191c0dcd65d46671.jpg) +Figure 9. Our method can transfer facial photos to artistic portraits in the style of exemplars. + +Skip connections & global style control. In MATEBIT, we use skip connections to supplement input semantic information. Removing skip connections dramatically hurts the semantic inconsistency and the quantitative results. Besides, using global style vector $\mathbf{z}$ increases subtle details, e.g. the colors over the mouth, rings, and hairs. + +In summary, MAT learns accurate correspondence and enables context-aware feature augmentation; the contrastive style learning benefits the style control and high-quality image generation; and the U-Net architecture helps the preservation of semantic information. Ultimately, all these benefits make our model significantly outperform previous state-of-the-art methods in generating plausible images. + +# 4.3. Applications + +Artistic Portrait Generation. An potential application of our method is transferring a facial photo to an artistic portrait, in the style of an exemplar. We here apply the previously learned models to randomly selected facial photos from CelebA-HQ [22]. As illustrated in Fig. 9, our method can generate appealing portraits with consistent identity and faithful style appearance. + +Chinese Ink Painting Generation. To verify the capacity of our model in generating complex images, we additionally apply it to generate Chinese Ink paintings. Specifically, we collect paintings of landscapes and facial portraits from the web, and then train and test our model on each subset respectively. Fig. 10 illustrates the results of painting generation and photo-to-painting translation. Obviously, all the generated images show remarkably high quality. Besides, our model successfully captures subtle differences + +![](images/8856d757dc719f66384509178a9a7088c7264a071a6c5ba2acc263f040c362b2.jpg) +Figure 10. Chinese ink paintings generation (1st & 3rd rows), as well as photo-to-painting translation (2nd & 4th rows). + +between different exemplars, demonstrating its remarkable capacity in style control. + +# 5. Conclusions + +This paper presents a novel exemplar-guided image translation method, dubbed MATEBIT. Both quantitative and qualitative experiments show that MATEBIT is capable of generating high-fidelity images in a number of tasks. Besides, ablation studies demonstrate the effectiveness of MAT and contrastive style learning. Despite such achievements, the artistic portraits transferred from facial photos (Fig. 9) are inferior to those shown in Fig. 6. This may be due to the subtle differences in edge maps between photos and artistic paintings. In the near future, we will explore to solve this issue via semi-supervised learning or domain transfer technologies. + +# Acknowledgements + +This work was supported in part by the National Key Research and Development Program of China under Grant 2018AAA0103202; in part by the National Natural Science Foundation of China under Grants 61971172, 61971339, 62176230, U22A2033, U22A2096; in part by the Zhejiang Provincial Science and Technology Program in China under Grant 2021C01108; in part by the NSFC-Zhejiang Joint Fund for the Integration of Industrialization and Informatization under Grant U1909210; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; and in part by Open Research Projects of Zhejiang Lab under Grant 2021KG0AB01. + +# References + +[1] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. 2018. 5 +[2] Huiwen Chang, Jingwan Lu, Fisher Yu, and Adam Finkelstein. Pairedcyclegan: Asymmetric style transfer for applying and removing makeup. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 40-48, 2018. 1 +[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2 +[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2 +[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 5 +[8] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2 +[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 4 +[10] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11936-11945, 2021. 1 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[12] Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13713-13722, 2021. 2, 4 + +[13] Xueqi Hu, Xinyue Zhou, Qiusheng Huang, Zhengyi Shi, Li Sun, and Qingli Li. Qs-attn: Query-selected attention for contrastive learning in i2i translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18291-18300, 2022. 2 +[14] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5 +[15] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1 +[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1, 2 +[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6 +[18] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 6, 8 +[19] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 2 +[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[21] Thomas N Kipf and Max Welling. Variational graph autoencoders. arXiv preprint arXiv:1611.07308, 2016. 1 +[22] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549-5558, 2020. 5, 6, 8 +[23] Boyi Li, Felix Wu, Kilian Q Weinberger, and Serge Belongie. Positional normalization. In Advances in Neural Information Processing Systems, pages 1620-1632, 2019. 3, 4 +[24] Mingcong Liu, Qiang Li, Zekui Qin, Guoxin Zhang, Pengfei Wan, and Wen Zheng. Blendgan: implicitly gan blending for arbitrary stylized face generation. Advances in Neural Information Processing Systems, 34:29710-29722, 2021. 6 +[25] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In European Conference on Computer Vision, 2022. 1, 2, 3, 4, 6 +[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2 + +[27] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6 +[28] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 2, 3, 4 +[29] Mandi Luo, Haoxue Wu, Huaibo Huang, Weizan He, and Ran He. Memory-modulated transformer network for heterogeneous face recognition. IEEE Transactions on Information Forensics and Security, 2022. 1 +[30] Liqian Ma, Xu Jia, Stamatios Georgoulis, Tinne Tuytelaars, and Luc Van Gool. Exemplar guided unsupervised image-to-image translation with semantic consistency. Proceedings ICLR 2019, 2019. 1 +[31] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In Proceedings of the European conference on computer vision (ECCV), pages 768-783, 2018. 5 +[32] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel flow: Semantic correspondence with multi-layer neural features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3395-3404, 2019. 3 +[33] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In International Conference on Learning Representations, 2018. 5 +[34] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In European conference on computer vision, pages 319-345. Springer, 2020. 2 +[35] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 1, 2, 4, 6 +[36] Justin N. M. Pinkney. Aligned ukiyo-e faces dataset, 2020. 6 +[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2 +[38] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, August 2020. version 0.2.1. 6 +[39] Junyoung Seo, Gyuseong Lee, Seokju Cho, Jiyoung Lee, and Seungryong Kim. Midms: Matching interleaved diffusion models for exemplar-based image translation. arXiv preprint arXiv:2209.11047, 2022. 1, 2 +[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR, 2015. 5, 6 + +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 7 +[42] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision, pages 63-79. Springer, 2018. 1 +[43] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 5, 6 +[44] Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, and Changgong Zhang. Marginal contrastive correspondence for guided image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10663-10672, 2022. 1, 2, 3, 6 +[45] Pan Zhang, Bo Zhang, Dong Chen, Lu Yuan, and Fang Wen. Cross-domain correspondence learning for exemplar-based image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5143-5153, 2020. 1, 2, 3, 5, 6, 7 +[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[47] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH, 2022. 2, 7 +[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1 +[49] Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocosnet v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11465–11475, 2021. 1, 2, 6 +[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 1 \ No newline at end of file diff --git a/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/images.zip b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d2dab15c406754b0cc5f19a1db00a4b84efc4c78 --- /dev/null +++ b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4131e377ef8e6e6a123807aa043871054254e5c1bc46f388a38ccae8b971f16 +size 818558 diff --git a/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/layout.json b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c8e308df217fa1ee5c9243787ae3f0af677b629 --- /dev/null +++ b/2023/Masked and Adaptive Transformer for Exemplar Based Image Translation/layout.json @@ -0,0 +1,9828 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "text", + "content": "Masked and Adaptive Transformer for Exemplar Based Image Translation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "text", + "content": "Chang Jiang" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "text", + "content": ", Fei Gao" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "text", + "content": ", Biao Ma" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "text", + "content": ", Yuhao Lin" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "text", + "content": ", Nannan Wang" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "text", + "content": ", Gang Xu" + }, + { + "bbox": [ + 112, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 157, + 483, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 157, + 483, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 483, + 171 + ], + "type": "text", + "content": "1 School of Computer Science and Technology, Hangzhou Dianzi University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 171, + 544, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 171, + 544, + 186 + ], + "type": "text", + "content": " Hangzhou Institute of Technology, Xidian University " + }, + { + "bbox": [ + 53, + 171, + 544, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 171, + 544, + 186 + ], + "type": "text", + "content": " ISN State Key Laboratory, Xidian University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 186, + 483, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 186, + 483, + 200 + ], + "spans": [ + { + "bbox": [ + 111, + 186, + 483, + 200 + ], + "type": "text", + "content": "{jc233, aiartma, linyh, gxu}@hdu.edu.cn, {fgao, nnwang}@xidian.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "type": "text", + "content": "We present a novel framework for exemplar based image translation. Recent advanced methods for this task mainly focus on establishing cross-domain semantic correspondence, which sequentially dominates image generation in the manner of local style control. Unfortunately, cross-domain semantic matching is challenging; and matching errors ultimately degrade the quality of generated images. To overcome this challenge, we improve the accuracy of matching on the one hand, and diminish the role of matching in image generation on the other hand. To achieve the former, we propose a masked and adaptive transformer (MAT) for learning accurate cross-domain correspondence, and executing context-aware feature augmentation. To achieve the latter, we use source features of the input and global style codes of the exemplar, as supplementary information, for decoding an image. Besides, we devise a novel contrastive style learning method, for acquire quality-discriminative style representations, which in turn benefit high-quality image generation. Experimental results show that our method, dubbed MATEBIT, performs considerably better than state-of-the-art methods, in diverse image translation tasks. The codes are available at https://github.com/AiArt-HDU/MATEBIT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "type": "text", + "content": "Image-to-image translation aims at transfer images in a source domain to a target domain [16, 50]. Early studies learn mappings directly by Generating Adversarial Networks (GANs), and have shown great success in various applications [2, 42]. Recently, exemplar based image translation [29, 30, 45], where an exemplar image is used to control the style of translated images, has attracted a lot of attention. Such methods allow high flexibility and controllability, and have a wide range of potential applications in social networks and metaverse. For example, people can transfer a" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 228, + 532, + 380 + ], + "blocks": [ + { + "bbox": [ + 320, + 228, + 532, + 380 + ], + "lines": [ + { + "bbox": [ + 320, + 228, + 532, + 380 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 532, + 380 + ], + "type": "image", + "image_path": "73cb608b03c9debbdc14df50dd542ab2a27abb3dfaed97834c96cb8fed1dec58.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 382, + 547, + 427 + ], + "lines": [ + { + "bbox": [ + 305, + 382, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 547, + 427 + ], + "type": "text", + "content": "Figure 1. Visualization of correspondence maps. The red point is the query position. Full Corr. and Masked Corr. denote the full correspondence [45] and masked one in our method, respectively. CAM denotes visualization by Class Activation Mapping [48]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 433, + 547, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 481 + ], + "type": "text", + "content": "facial sketch to an artistic portrait, in the style of oil paintings or avatars. Despite the remarkable progress, yielding high-fidelity images with consistent semantic and faithful styles remains a grand challenge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 483, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 651 + ], + "type": "text", + "content": "Early pioneering works [15, 21, 35] attempt to globally control the style of generated images. However, such methods ignore spatial correlations between an input image and an exemplar, and may fail to produce faithful details. Recently, some advanced methods [25, 44, 45, 49] first establish the cross-domain semantic correspondence between an input image and an exemplar, and then use it to warp the exemplar for controlling local style patterns. In these methods, the quality of generated images relies heavily on the learned correspondence [39]. Unfortunately, cross-domain semantic matching is challenging, since there is no reliable supervision on correspondence learning [45]. As a result, potential matching errors ultimately lead to degraded artifacts in generated images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "To combat this challenge, we propose to boost the matching accuracy on one hand, and to diminish the role of matching in image generation on the other hand. Inspired by the great success of Transformers [6, 10, 26, 41], we first devise a Masked and Adaptive Transformer (MAT) for learning ac" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22418" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "curate cross-domain correspondence and executing context-aware feature augmentation. Previous works [44, 45, 49] have used the vanilla attention mechanism [41] for learning full correspondence. However, the initial attention typically involves ambiguous correspondences (2nd row in Fig. 1). To mitigate these limitations, in MAT, we use a masked attention to distinguish the correspondence as reliable or not, and then reliability-adaptively aggregate representations. Besides, the Feed-Forward Network (FFN) [41] in vanilla transformers neglects contextual correlations inside an image. We thus replace FFN by an adaptive convolution block [28], where the coordinate attention [12] and depth-wise separable convolution [5] are used to capture contextual correlations and to improve efficiency. With a joint consideration of matching reliability and contextual correlations, MAT gradually focuses on accurate correspondences and emphasizes on features of interest (3rd row in Fig. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 467 + ], + "type": "text", + "content": "In addition, to boost both the semantic consistency and style faithfulness, we supplementally use semantic features of the input image and global style codes of the exemplar for decoding an image. To this end, we first design our whole network following the U-Net architecture [16]. Besides, we devise a novel contrastive style learning (CSL) framework for acquiring discriminative style representations. Recently, Zhang et al. [47] propose a similar CSL method, where the target exemplar is used as a positive sample, and the other exemplars as negative ones. Differently, we use low-quality images, generated during early training stages, as negative samples. In this way, our style codes are desired to discriminate not only subtle differences in style, but also those in perceptual quality. Ultimately, the learned global style codes, cooperating with the local style control induced by MAT, in turn benefit high-quality image generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 289, + 598 + ], + "type": "text", + "content": "With the proposed techniques above, our full model, dubbed MATEBIT, diminishes the impact of position-wise matching on image quality, and integrates both local and global style control for image generation. Experimental results show that MATEBIT generates considerably more plausible images than previous state-of-the-art methods, in diverse image translation tasks. In addition, comprehensive ablation studies demonstrate the effectiveness of our proposed components. Finally, we perform interesting applications of photo-to-painting translation and Chinese ink paintings generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "type": "text", + "content": "2. Relate Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Exemplar Based Image Translation. Recently, exemplar based image translation has attracted increasing attention. For example, Park et al. [35] learn an encoder to map the exemplar image into a global style vector, and use it to guide image generation. Such a global style control strategy enables style consistency in whole, but fails to produce subtle details. Most recently, researchers propose a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "matching-then-generation framework [39]. Specially, they first establish dense correspondence between an input and an exemplar, and then reshuffle the exemplar for locally control the style of synthesize images. For example, Zhang et al. [45] establish position-wise correspondence based on the Cosine attention mechanism and warp the exemplar correspondingly. Afterwards, the warped image dominates the generation of images in the manner of SPADE [35]. To reduce the cost of matching in high-resolution image generation, Zhou et al. [49] introduce a hierarchical refinement of semantic correspondence from ConvGRU-PatchMatch. Besides, Liu et al. [25] used a dynamic pruning method for learning hierarchical sparse correspondence. They also use reliability-adaptive feature integration to improve the quality of generated images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "type": "text", + "content": "Previous methods merely use global or local style control, and the latter relies heavily on the learned correspondence. Besides, they consider little about contextual correlations inside an image. In this paper, we use both global and local style control to boost the style consistency. Besides, we take contextual correlations into consideration and execute reliability-adaptive feature augmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 340, + 548, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 548, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 548, + 519 + ], + "type": "text", + "content": "Transformers. Transformers [41] have shown incredible success from the field of natural language processing (NLP) [19] to computer vision (CV) [6, 26]. Multihead attention (MHA) and FFN are key components in a Transformer, and have been used in exemplar based image translation. However, they induce unreliable matching results and neglect context correlations in feature translation. In our MAT, we combat these limitations by replacing them with a masked attention and a context-aware convolution block, respectively. Recently, researchers use semantic masks to facilitate representation learning [4, 8, 37], where a mask predictor is required. Differently, we use a ReLU function to mask over the attention layer, for distinguishing correspondence as reliable or not (Sec. 3.1). In general, MAT follows a concise and efficient architecture." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": "Contrastive Learning. Contrastive learning has shown its effectiveness in various computer vision tasks [9, 13, 34]. The basic idea is to learn a representation by pushing positive samples toward an anchor, and moving negative samples away from it. Different sampling strategies and contrastive losses have been extensively explored in various downstream tasks. For example, Chen et al. [3] and He et al. [9] obtain positive samples by augmenting original data. In the field of image translation, Park et al. [34] propose patchwise contrastive learning by maximizing the mutual information between cross-domain patches. Similarly, Zhang et al. [47] use contrastive learning for acquiring discriminative style representations. In the task of exemplar based image translation, Zhan et al. [44] use contrastive learning to align cross-domain images to a consistent semantic feature space, so as to boost the accuracy of matching. Differently, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22419" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 193 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 193 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 193 + ], + "type": "image", + "image_path": "87cb6bd4342918ba51234ed1a6d836c2b00741bcfcd42f8778a9cb1df09b4e05.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 197, + 286, + 209 + ], + "lines": [ + { + "bbox": [ + 47, + 197, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 286, + 209 + ], + "type": "text", + "content": "Figure 2. Overview of our image translation network, MATEBIT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 287, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 287, + 254 + ], + "type": "text", + "content": "use early generated images as negative samples, so that the learned style representations can discriminate subtle differences in both style and perceptual quality (Sec. 3.2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 267, + 176, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 267, + 176, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 176, + 281 + ], + "type": "text", + "content": "3. The Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " in domain " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " and an exemplar image " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " in domain " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": ", our goal is to generate a target image " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "x_{B}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " which preserves semantic structures in " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " but resembles the style of similar parts in " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": ". Fig. 2 shows an overview of our translation network " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": ". Specially, we first align " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " to an intermediate feature space by encoders " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{A}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{B}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": ", respectively. Afterwards, we use a Masked and Adaptive Transformer (MAT) for correspondence learning and feature augmentation. Finally, a decoder " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_B" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " produces an output image " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\hat{x}_B" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " based on the augmented features, as well as the source features and target style codes. Details are described below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 441, + 270, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 270, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 270, + 455 + ], + "type": "text", + "content": "3.1. Masked and Adaptive Transformer (MAT)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 604 + ], + "type": "text", + "content": "In order to establish accurate cross-domain correspondence, we propose a novel and concise Transformer architecture, i.e. MAT. In general, the architecture of MAT (Fig. 3b) follows that of vanilla Transformers (Fig. 3a) [41]. Differently, we use masked attention to distinguish reliable and unreliable correspondence, instead of using multi-head attention. Besides, we use Positional Normalization (PONO) [23] and an Adaptive Convolution (AdaConv) block [28], instead of LN and MLP-based FFN, respectively. MAT is desired to gradually concentrate on accurate matching, and to reliability-adaptively augment representations with contextual correlations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": "Masked Correspondence Learning. Let " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_B \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " be the representations of " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_A" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "y_B" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " in the intermediate feature space, with height " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " channels. We first map " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " to the query " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_B" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " to the key " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": ", by using " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " convolutions, respectively. As shown in Fig. 3d, we add positional encoding (PE) to " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_B" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": ", for embedding spatial correlations. Afterwards, we learn the initial correspondence " + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{HW \\times HW}" + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "text", + "content": " fol" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 75, + 383, + 251 + ], + "blocks": [ + { + "bbox": [ + 321, + 75, + 383, + 251 + ], + "lines": [ + { + "bbox": [ + 321, + 75, + 383, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 75, + 383, + 251 + ], + "type": "image", + "image_path": "de343606c827cbfaccca0f7d431dccf574b28466ed6aa0ab5a2d4facc7f444d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 385, + 75, + 459, + 251 + ], + "blocks": [ + { + "bbox": [ + 385, + 75, + 459, + 251 + ], + "lines": [ + { + "bbox": [ + 385, + 75, + 459, + 251 + ], + "spans": [ + { + "bbox": [ + 385, + 75, + 459, + 251 + ], + "type": "image", + "image_path": "16099285aba2a1f5025a9aa51abd3a44a96f4e9e4d85a700a372bfd3c1377d79.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 462, + 76, + 531, + 251 + ], + "blocks": [ + { + "bbox": [ + 462, + 76, + 531, + 251 + ], + "lines": [ + { + "bbox": [ + 462, + 76, + 531, + 251 + ], + "spans": [ + { + "bbox": [ + 462, + 76, + 531, + 251 + ], + "type": "image", + "image_path": "ee95e5984517f9afcf7e7686b851fc64c55b65c7ed5d4fbbfb4862faea5a3984.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 252, + 531, + 405 + ], + "blocks": [ + { + "bbox": [ + 321, + 252, + 531, + 405 + ], + "lines": [ + { + "bbox": [ + 321, + 252, + 531, + 405 + ], + "spans": [ + { + "bbox": [ + 321, + 252, + 531, + 405 + ], + "type": "image", + "image_path": "cce8d1a41fd26826f80744de106158f1377f8e06284d1222f113314602c08c06.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 407, + 545, + 429 + ], + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 429 + ], + "type": "text", + "content": "Figure 3. Detailed architectures. (a) Vanilla Transformer block, (b) MAT block, (c) AdaConv block, and (d) Masked Attention." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 438, + 503, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 503, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 503, + 450 + ], + "type": "text", + "content": "lowing the Cosine attention mechanism [45], i.e." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 362, + 458, + 545, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 458, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 362, + 458, + 545, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {A} (u, v) = \\frac {\\tilde {\\mathbf {Q}} (u) \\tilde {\\mathbf {K}} (v) ^ {T}}{| | \\tilde {\\mathbf {Q}} (u) | | \\cdot | | \\tilde {\\mathbf {K}} (v) | |}, \\tag {1}", + "image_path": "29f2aa4ea8c3d3f8915351a5c6eca43f0fb83e29aa9c7c659ac27e016af71b80.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}(u) = \\mathbf{Q}(u) - \\bar{\\mathbf{Q}}(u)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{K}}(v) = \\mathbf{K}(v) - \\bar{\\mathbf{K}}(v)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "u, v \\in [1, \\dots, HW]" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": " are position indices; " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Q}}(u)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{K}}(v)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": " are the means of " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}(u)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{K}(v)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(u, v)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": " is the matching score between " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}(u)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{K}(v)" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "type": "text", + "content": "Previous methods [44, 45] typically use the initial correspondence map " + }, + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "type": "text", + "content": " to reshuffle an exemplar for controlling local patterns in image synthesis. However, induced by the difficulties in cross-domain correspondence learning, " + }, + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 545, + 546, + 653 + ], + "type": "text", + "content": " involves unreliable match scores (Fig. 3d). As a result, the reshuffled image will lead to implausible artifacts in generated images. To combat this limitation, we distinguish initial matching scores as reliable or not, according to their signs [32]. The masked correspondence map becomes:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 380, + 662, + 545, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 662, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 380, + 662, + 545, + 674 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {\\text {m a s k}} = \\operatorname {R e L U} (\\mathbf {A}), \\tag {2}", + "image_path": "d633ff3c2268d5ce313562a5c4147793fcb5a585bdf2cb187c0d0cfb12e13431.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "In DynaST [25], two networks are used to predict the reliability mask of correspondence. However, it's challenging to effectively train the network, because there is no super" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22420" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "vision on matching during training. In contrast, ReLU contains no learnable parameters and ultimately leads to superior performance over DynaST (Sec. 4.1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "type": "text", + "content": "Reliability-Adaptive Feature Aggregation. For regions with reliable correspondence in " + }, + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{mask}" + }, + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "type": "text", + "content": " to warp the value features, " + }, + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 47, + 108, + 286, + 145 + ], + "type": "text", + "content": ", derived from the exemplar:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 150, + 286, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 150, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 51, + 150, + 286, + 175 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\text {c o r}} = \\tilde {\\mathbf {A}} _ {\\text {m a s k}} \\mathbf {V}, \\text {w i t h} \\tilde {\\mathbf {A}} _ {\\text {m a s k}} = \\operatorname {s o f t m a x} (\\alpha \\cdot \\mathbf {A} _ {\\text {m a s k}}), \\tag {3}", + "image_path": "265c2e5d63462473c42f206fae04e9ebe4edc772fe2bad543128c8cfc4b1859c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 176, + 286, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 286, + 199 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 176, + 286, + 199 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 176, + 286, + 199 + ], + "type": "text", + "content": " is a scaling coefficient to control the sharpness of the softmax function. In default, we set its value as 100." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": "For regions with unreliable correspondence in " + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{cor}" + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": " provides an average style representation of " + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": ". We further extract complementary information from the query, " + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": ", derived from the input. Inspired by SPADE [35], we first transfer " + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": " to the target domain by using pixel-wise modulation parameters (i.e., " + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": " for scale and " + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": " for bias) learned from " + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 46, + 200, + 286, + 283 + ], + "type": "text", + "content": ". The modulation is formulated by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 290, + 287, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 290, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 86, + 290, + 287, + 316 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {\\text {n o r m}} = \\gamma (x _ {A}) \\frac {\\mathbf {Q} - \\mu (\\mathbf {Q})}{\\sigma (\\mathbf {Q})} + \\beta (x _ {A}), \\tag {4}", + "image_path": "dd4fa79fb8db8fc5d023a21acedec224107c4e2395d0ec92dd16c7ec5edfa12f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "inline_equation", + "content": "\\mu (\\mathbf{Q})" + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "inline_equation", + "content": "\\sigma (\\mathbf{Q})" + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "text", + "content": " are the mean value and standard deviance of " + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "text", + "content": ". Afterwards, we select the translated features of unreliable corresponded regions in " + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 47, + 323, + 286, + 359 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 366, + 287, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 366, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 84, + 366, + 287, + 386 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {u n c o r} = \\left(1 - \\sum_ {j} \\mathbf {A} _ {m a s k}\\right) \\odot \\mathbf {Q} _ {n o r m}, \\tag {5}", + "image_path": "f0000ac697c018220561f4de4a10d34f2ee803763bdc0af27f1d25292d16903a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "text", + "content": "where the summation is along the second dimension; " + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "text", + "content": " denotes point-wise production with broadcasting. Since " + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "text", + "content": " are learned from the input image " + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "text", + "content": ", the modulated features preserve the semantic information of " + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "text", + "content": ". Besides, constraints on the generated image will push the selected features convey to the style of " + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 47, + 387, + 286, + 458 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "text", + "content": "Ideally, " + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{cor}" + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{uncor}" + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "text", + "content": " would complement each other and facilitate both semantic consistency and style relevance in image generation. To this end, we integrate " + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{cor}" + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{uncor}" + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 47, + 458, + 286, + 506 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 514, + 286, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 514, + 286, + 527 + ], + "spans": [ + { + "bbox": [ + 85, + 514, + 286, + 527 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {a g g} = \\operatorname {P O N O} \\left(\\mathbf {X} _ {c o r} + \\mathbf {X} _ {u n c o r} + \\mathbf {Q}\\right). \\tag {6}", + "image_path": "e2286875b8ea70e3d2a56afcfee43b976fbea04dc42e4d0dee1914b87f438bd6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 533, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 286, + 581 + ], + "type": "text", + "content": "In PONO [23], features at each position are normalized dependently. Compared to LN in vanilla transformers and DynaST [25], PONO boosts the flexibility in reliability-adaptive feature aggregation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 582, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 286, + 641 + ], + "type": "text", + "content": "Context-Aware Feature Augmentation. Inspired by ConvNeXT [28], we replace FFN by an AdaConv block to position-adaptively emphasize informative representations. Besides, we use the coordinate attention (CoordAten) module [12] to capture contextual correlations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": "The architecture of the AdaConv block is as shown in Fig. 3c. We fist use the depthwise convolution (Dwise) to update representations in each channel separately; and then use two pointwise convolutions (Pwise) to automatically emphasize representations of interest, at every position. The Gaussian Error Linear Unit (GELU) activation" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 314, + 71, + 535, + 208 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 535, + 208 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 535, + 208 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 535, + 208 + ], + "type": "image", + "image_path": "92a1c60f1844b473bacc6475f76c30344beaf7bd2d8d2b1c76a8e42dcfd3c97f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 213, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 235 + ], + "type": "text", + "content": "Figure 4. Contrastive style learning. The memory bank consists of divergent low-quality images generated in early training stages." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 246, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 317 + ], + "type": "text", + "content": "function and Layer Norm (LN) are used after the first Pwise layer [28]. Notably, Coordinate is used after the Dwise layer for modeling long-range dependencies in an image. Specially, Coordinate produces cross-channel and position-sensitive attention maps, which helps our model to more accurately locate the representations of interest [12]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 318, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 389 + ], + "type": "text", + "content": "Finally, the output of a MAT block is obtained with a residual connection, i.e. " + }, + { + "bbox": [ + 304, + 318, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{MAT}} = \\mathrm{AdaConv}(\\mathbf{X}_{agg}) + \\mathbf{X}_{agg}" + }, + { + "bbox": [ + 304, + 318, + 545, + 389 + ], + "type": "text", + "content": ". In the implementation, we stack three MAT blocks in default to gradually refine the correspondence and to augment informative representations (Fig. 1). Empirical verifications will be given in Sec. 4.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": "Benefits of MAT. Fig. 3d illustrates the impact of MAT. The query point locates over the left eye of the source image. Here we show the magnitudes of its correspondence over the exemplar, in the third layer of MAT. Obviously, the original correspondence " + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": " covers both eyes of the exemplar. In contrast, the masked correspondence " + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{mask}" + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": " accurately concentrates over the left eye. Such superiority significantly boost the quality of ultimate images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 491, + 486, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 486, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 486, + 502 + ], + "type": "text", + "content": "3.2. Contrastive Style Learning (CSL)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "content": "In MATEBIT, we use the encoder " + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_B" + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "content": " to extract local style information " + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_B" + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "content": ", and then a MLP to extract global style codes " + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_B" + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "content": " perform local and global style control on generated images, respectively (Sec. 3.3). To boost the discriminative capacity of style representations, as well as the quality of generated images, we propose a novel contrastive style learning (CSL) method (as shown in Fig 4)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "In our settings, the exemplars are drawn by human artists and thus considered as high-quality. In contrast, the images generated in early training stages are typically low-quality. Inspired by the idea of contrastive learning [9], we use the exemplar " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " as the positive sample, while a collection of early generated images as negative. Let " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " denotes style codes of the generated image " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}_B" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^+" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " that of exemplar " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_1^-, \\mathbf{z}_2^-, \\dots, \\mathbf{z}_m^-\\}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " the style codes of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " negative samples. CSL learns style representations by maximizing the mutual information between anchors and positive samples," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 316, + 757 + ], + "type": "text", + "content": "22421" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "while minimizing that between anchors and negative samples. Our contrastive style loss is computed by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 102, + 287, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 287, + 140 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s t y l e} = - \\log \\frac {\\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} ^ {+}}{\\tau}\\right)}{\\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} ^ {+}}{\\tau}\\right) + \\sum_ {j = 1} ^ {m} \\exp \\left(\\frac {\\mathbf {z} ^ {T} \\mathbf {z} _ {j} ^ {-}}{\\tau}\\right)}, \\tag {7}", + "image_path": "7c48530bc1dfdd2e5a6158a5ca46d8f5f741d8c8ddd0eb7d6999314cb1143e08.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 145, + 287, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 145, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 145, + 287, + 170 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 145, + 287, + 170 + ], + "type": "inline_equation", + "content": "\\tau = 0.07" + }, + { + "bbox": [ + 47, + 145, + 287, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 145, + 287, + 170 + ], + "type": "inline_equation", + "content": "m = 1024" + }, + { + "bbox": [ + 47, + 145, + 287, + 170 + ], + "type": "text", + "content": ". In the implementation, we use a queue to cache negative style vectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 176, + 165, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 165, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 165, + 188 + ], + "type": "text", + "content": "3.3. Translation network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "content": "To boost both the semantic consistency and style faithfulness, we additionally use source semantic features and global style codes for decoding an image. Specially, we design our whole translation network following U-Net (Fig. 2), where the multi-level features in " + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_A" + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "content": " are skip-connected to the decoder " + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_B" + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "content": ", for supplementing informative semantic structures of the input image " + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "inline_equation", + "content": "x_A" + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "content": ". Besides, we use the style codes " + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "content": " to globally control the style of generated images, in the manner of AdaIN [14]. Specially, " + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "content": " is mapped to channel-wise modulating factors by fully-connected (FC) layers. In this way, we diminish the impact of correspondence learning on image generation, and provide reliable style control for even unmatched regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "content": "In summary, our translation network allows both local and global style control, and reuses the semantic features of input images. As a result, the generated image is desired to present consistent semantic to the input " + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "content": " and faithful style to the exemplar " + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "content": ". More details of our network are available in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 429, + 137, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 137, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 137, + 440 + ], + "type": "text", + "content": "3.4. Loss functions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "content": "Our whole network is end-to-end optimized to jointly achieve high-fidelity image generation and accurate correspondence. Following [45], we obtain training triplets " + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\{x_A, y_B, x_B\\}" + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "content": " from the ready-made data pair " + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\{x_A, x_B\\}" + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "inline_equation", + "content": "y_B" + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "content": " is a geometrically warped version of " + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "inline_equation", + "content": "x_B" + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "content": ". The generated image is denoted by " + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\hat{x}_B = \\mathcal{G}(x_A, y_B)" + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "content": ". Our loss functions are similar to [45], except for the previous contrastive style loss " + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{style}" + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "content": " and the structural loss " + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{str}" + }, + { + "bbox": [ + 46, + 447, + 287, + 543 + ], + "type": "text", + "content": " below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "text", + "content": "Semantic alignment Loss. For accurate cross-domain correspondence learning, the encoders " + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_A" + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_B" + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "text", + "content": " should align " + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "inline_equation", + "content": "x_A" + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "inline_equation", + "content": "x_B" + }, + { + "bbox": [ + 46, + 544, + 287, + 590 + ], + "type": "text", + "content": " to consistent representations. The corresponding semantic alignment loss is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 98, + 599, + 287, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 599, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 98, + 599, + 287, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\left\\| \\mathcal {E} _ {A} \\left(x _ {A}\\right) - \\mathcal {E} _ {B} \\left(x _ {B}\\right) \\right\\| _ {1}. \\tag {8}", + "image_path": "ace73574563c17d588d8af6f00a140ebab29f09d9c52ee9b05c6b699fd2b5a81.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 619, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 287, + 655 + ], + "type": "text", + "content": "Correspondence Loss. Ideally, if we warp " + }, + { + "bbox": [ + 47, + 619, + 287, + 655 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 47, + 619, + 287, + 655 + ], + "type": "text", + "content": " in the same way as Eq.3, the resulting image should be exactly " + }, + { + "bbox": [ + 47, + 619, + 287, + 655 + ], + "type": "inline_equation", + "content": "x_{B}" + }, + { + "bbox": [ + 47, + 619, + 287, + 655 + ], + "type": "text", + "content": ". We thus constrain the learned correspondence by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 97, + 662, + 287, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 662, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 97, + 662, + 287, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o r r}} = \\left\\| \\tilde {\\mathbf {A}} _ {\\text {m a s k}} ^ {T} y _ {B} \\downarrow - x _ {B} \\downarrow \\right\\| _ {1}, \\tag {9}", + "image_path": "f2e8ff2686c055b9ef94e10ddcc38813dc532875f481004ca6b4758b4f8551e8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": " indicates down-sampling " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "x_{B}" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": " to the size (i.e. width and height) of " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "Perceptual Loss. The generated image " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\hat{x}_B" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " should be semantic-consistent with the ground truth " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "x_B" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " in term of semantic. We thus use the perceptual loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 116, + 545, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 116, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 361, + 116, + 545, + 129 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p e r c}} = \\left\\| \\varphi_ {l} \\left(\\hat {x} _ {B}\\right) - \\varphi_ {l} \\left(x _ {B}\\right) \\right\\| _ {1}, \\tag {10}", + "image_path": "5dbeffa33de71e4d4a4cd236e663c1c6da480252b770f48b9c2e399a001f9c3e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 136, + 545, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 136, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 136, + 545, + 159 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 136, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\varphi_{l}" + }, + { + "bbox": [ + 305, + 136, + 545, + 159 + ], + "type": "text", + "content": " denotes the activations after layer relu4_2 in pretrained VGG19 [40], which represent high-level semantics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 160, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 160, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 160, + 545, + 220 + ], + "type": "text", + "content": "Contextual Loss. In addition, the generated image should be in the same style as the exemplar. In addition to the previous contrastive style loss (Eq.7), we additionally use the contextual loss (CX) [31] to constrain on local style consistency. The contextual loss is computed by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 226, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 226, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 324, + 226, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c t x} = - \\log \\left(\\sum_ {l} w _ {l} \\mathrm {C X} \\left(\\varphi_ {l} \\left(\\hat {x} _ {B}\\right), \\varphi_ {l} \\left(y _ {B}\\right)\\right)\\right) \\tag {11}", + "image_path": "1b387175b78ef8157ecf733a0919129cf902964f329307dc31562ff2ea38c015.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 259, + 531, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 531, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 531, + 271 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 259, + 531, + 271 + ], + "type": "inline_equation", + "content": "w_{l}" + }, + { + "bbox": [ + 306, + 259, + 531, + 271 + ], + "type": "text", + "content": " balances the terms of different VGG19 layers." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 272, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 272, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 545, + 319 + ], + "type": "text", + "content": "Structural Loss. The generated image should preserve semantic structures in the input image. Correspondingly, we use the Learned Perceptual Image Patch Similarity (LPIPS) [46] between their boundaries as the structural loss:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 327, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 327, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 358, + 327, + 545, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s t r} = \\operatorname {L P I P S} \\left(\\mathcal {H} \\left(\\hat {x} _ {B}\\right), \\mathcal {H} \\left(x _ {B}\\right)\\right), \\tag {12}", + "image_path": "80a04276d88f8a6b8e73e398772f7f710f15a286ea3d4daffcaa0ad4c50b8617.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 347, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 545, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 347, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 305, + 347, + 545, + 371 + ], + "type": "text", + "content": " is the HED algorithm [43], which has been widely used for extracting semantic boundaries in an image." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "text", + "content": "Adversarial loss. Finally, we add a discriminator " + }, + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "text", + "content": " to distinguish real images in domain " + }, + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "text", + "content": " and the generated images [7]. The adversarial loss is:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 330, + 413, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 413, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 330, + 413, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} ^ {\\mathcal {D}} = - \\mathbb {E} [ h (\\mathcal {D} (y _ {B})) ] - \\mathbb {E} [ h (- \\mathcal {D} (\\hat {x} _ {B})) ], \\tag {13}", + "image_path": "ce7c2d7b6078158afd1535bcfa7ea787e77a5b1bebb370d53ca49f2dbb51a93c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 332, + 430, + 417, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 430, + 417, + 445 + ], + "spans": [ + { + "bbox": [ + 332, + 430, + 417, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} ^ {\\mathcal {G}} = - \\mathbb {E} [ \\mathcal {D} (\\hat {x} _ {B}) ],", + "image_path": "55413c6d999578c477c93b88a6ed5c82e4a3796fd7c01b397df21ef14952ead0.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 452, + 544, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 544, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 544, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 452, + 544, + 464 + ], + "type": "inline_equation", + "content": "h(t) = \\min (0, - 1 + t)" + }, + { + "bbox": [ + 306, + 452, + 544, + 464 + ], + "type": "text", + "content": " is the hinge loss function [1]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 464, + 544, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 464, + 544, + 475 + ], + "spans": [ + { + "bbox": [ + 318, + 464, + 544, + 475 + ], + "type": "text", + "content": "Total loss. In summary, our overall objective function is," + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 310, + 493, + 545, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 493, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 310, + 493, + 545, + 527 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\mathcal {G}} \\max _ {\\mathcal {D}} \\lambda_ {1} \\mathcal {L} _ {\\text {s t y l e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c o r r}} + \\lambda_ {4} \\mathcal {L} _ {\\text {s t r}} \\tag {14} \\\\ + \\lambda_ {5} \\left(\\mathcal {L} _ {p e r c} + \\mathcal {L} _ {c t x}\\right) + \\lambda_ {6} \\left(\\mathcal {L} _ {a d v} ^ {\\mathcal {G}} + \\mathcal {L} _ {a d v} ^ {\\mathcal {D}}\\right) \\\\ \\end{array}", + "image_path": "db6b0a770465857f58608173172e39bb27d4fe203f6e512540a02f7e44ceb553.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 534, + 467, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 467, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 467, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 534, + 467, + 545 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 534, + 467, + 545 + ], + "type": "text", + "content": " denotes the weight parameters." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 550, + 382, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 382, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 382, + 563 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "text", + "content": "Implementation details. We apply spectral normalization [33] to all the layers in the translation network and discriminator. We use the Adam [20] solver with " + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0" + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "text", + "content": ". The learning rates for the generator and discriminator are set as " + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "inline_equation", + "content": "4e - 4" + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "text", + "content": " respectively, following TTUR [11]. The experiments are conducted using 4 24GB RTX3090 GPUs. Limited by the computation load, we restrict the resolution of generated images to " + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 305, + 570, + 545, + 676 + ], + "type": "text", + "content": " in all translation tasks." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Datasets. We mainly conduct experiments on the following datasets. (1) CelebA-HQ [22] contains 30,000 facial photos. We chose 24,000 samples as the training set" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22422" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 84, + 545, + 173 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 537, + 82 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 537, + 82 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 537, + 82 + ], + "type": "text", + "content": "Table 1. Comparison on the Metfaces [18], CelebA-HQ [22], Ukiyo-e [36], Cartoon [36], AAHQ [24], and DeepFashion [27] datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 84, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 84, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 84, + 545, + 173 + ], + "type": "table", + "html": "
CelebA-HQMetfacesCartoonUkiyo-eAAHQDeepFashionTime ↓
FID ↓SWD ↓Texture ↑Color ↑Semantic↑FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓FID ↓SWD ↓(s)
SPADE [35]31.526.90.9270.9550.92245.626.997.530.545.626.979.432.136.227.80.196
CoCosNet [45]14.315.20.9580.9770.94925.624.366.827.138.313.962.621.914.417.20.321
CoCosNet-v2 [49]13.214.00.9540.9750.94823.322.466.427.032.111.062.422.813.016.71.573
MCL-Net [44]12.814.20.9510.9760.95323.824.567.927.932.412.464.422.212.916.20.309
DynaST [25]12.012.40.9590.9780.95229.228.662.826.538.914.267.224.08.411.80.214
MATEBIT (ours)11.513.20.9660.9860.94926.019.164.427.630.311.556.019.58.210.00.185
", + "image_path": "9a154237a79f08c46aa09a363fec2ffdbc51b3f30c873ea3297acd6ca7c04e47.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 177, + 531, + 490 + ], + "blocks": [ + { + "bbox": [ + 62, + 177, + 531, + 490 + ], + "lines": [ + { + "bbox": [ + 62, + 177, + 531, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 177, + 531, + 490 + ], + "type": "image", + "image_path": "702a17053195b99e87c84d12b12e78855cb6452160ec0e42efbfcfbbed3d057a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 491, + 531, + 502 + ], + "lines": [ + { + "bbox": [ + 61, + 491, + 531, + 502 + ], + "spans": [ + { + "bbox": [ + 61, + 491, + 531, + 502 + ], + "type": "text", + "content": "Figure 5. Results on the Metfaces [18], CelebA-HQ [22], Ukiyo-e [36], Cartoon [36], AAHQ [24], and DeepFashion [27] datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "text", + "content": "and 3000 as the test set. (2) Metfaces [18] consists of 1336 high-quality artistic facial portraits. (3) AAHQ [24] consists of high-quality facial avatars. We randomly select 1500 samples for training and 1000 samples for testing. (4) Ukiyo-e [36] consists of high-quality Ukiyo-e faces. We randomly select 3000 and 1000 samples for training and testing, respectively. (5) Cartoon [36] consists of 317 cartoon faces. (6) DeepFashion [27] consists of 800,00 fashion images. On CelebA-HQ, we connect the face landmarks for face region, and use Canny edge detector to detect edges in the background. On DeepFashion, we use the officially provided landmarks as input. On the other datasets, we use HED [43] to obtain semantic edges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 664, + 224, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 224, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 224, + 677 + ], + "type": "text", + "content": "4.1. Comparison with state-of-the-art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "We select several advanced models, including SPADE [35], CoCosNet [45], CoCosNet-v2 [49], MCL-Net [44], and DynaST [25], for comparison. For a fair comparison," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 506, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 545, + 531 + ], + "type": "text", + "content": "we retrain their models at resolution " + }, + { + "bbox": [ + 305, + 506, + 545, + 531 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 305, + 506, + 545, + 531 + ], + "type": "text", + "content": " under the same settings as ours." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "text", + "content": "Quantitative evaluation. We adopt several criteria to fully evaluate the generation results. (1) Fréchet Inception Score (FID) [38] and Sliced Wasserstein distance (SWD) [17] are used to evaluate the image perceptual quality. (2) To assess style relevance and semantic consistency of translated images [45], we compute the color, texture, and semantic metrics based on VGG19 [40]. Specifically, the cosine similarities between low-level features (i.e. " + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "inline_equation", + "content": "relu1\\_2" + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "inline_equation", + "content": "relu2\\_2" + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "text", + "content": ") are used to measure color and texture relevance, respectively; the average cosine similarity between high-level features (i.e. " + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "inline_equation", + "content": "relu3\\_2" + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "inline_equation", + "content": "relu4\\_2" + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "inline_equation", + "content": "relu5\\_2" + }, + { + "bbox": [ + 304, + 532, + 546, + 675 + ], + "type": "text", + "content": ") measures the semantic consistency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "The quantitative comparison results are shown in Table 1. Compared to existing methods, our model consistently achieves superior or highly competitive performance across" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22423" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 273, + 319 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 273, + 319 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 273, + 319 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 273, + 319 + ], + "type": "image", + "image_path": "82cbc5b4d8f3d6e09a10fbeee5607b7de4e024ff83dcd007bee5ad55eb025ed8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 321, + 287, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 287, + 343 + ], + "type": "text", + "content": "Figure 6. More results generated by MATEBIT. Images shown in the left-bottom corner are source images or exemplars." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 348, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 467 + ], + "type": "text", + "content": "all the datasets. Especially, MATEBIT significantly improves the style relevance in both texture and color. On the complicated AAHQ dataset, which contains diverse styles of avatars, MATEBIT dramatically decreases both FID and SWD. Such superiority indicates that our generated images are of better perceptual quality; and present consistent appearance to similar parts in exemplars. We additionally report the average time each method costs for generating an image. Our method shows the best efficiency and is significantly faster than previous methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 469, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 636 + ], + "type": "text", + "content": "Qualitative comparison. Fig 5 illustrates images generated by different methods. Obviously, previous methods present geometric distortions, blurring artifacts, inconsistent colors, or identity inconsistency. In contrast, MATEBIT consistently produces appealing results, including more results shown in Fig. 6. Specially, our results preserve the semantic structure of input images, and present consistent appearance with semantically similar regions in exemplars. Previous methods suffer serious degradations mainly due to the matching errors in full correspondence learning. In our method, we distinct reliable and unreliable correspondence, and release the role of matching in image generation. As a result, our method stably transfers a source image to the target style of an exemplar." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 647, + 138, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 138, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 138, + 658 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Impacts of MAT. We present a comprehensive analysis to justify the important component in our architecture, i.e. MAT. We here modify our full model by (1) removing the MAT module (i.e. w/o MAT), (2) removing ReLU in" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 422, + 141 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 422, + 141 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 422, + 141 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 422, + 141 + ], + "type": "image", + "image_path": "202e9d72426fd8a879453b556e7bd83ccebd505dd8eed053b88a7a08b54b5432.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 426, + 72, + 544, + 142 + ], + "blocks": [ + { + "bbox": [ + 426, + 72, + 544, + 142 + ], + "lines": [ + { + "bbox": [ + 426, + 72, + 544, + 142 + ], + "spans": [ + { + "bbox": [ + 426, + 72, + 544, + 142 + ], + "type": "image", + "image_path": "1aa06cd1a4d6a58884dd41fa220dfeb8b76705ac2b172a9325b15123d222ba6c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 142, + 541, + 152 + ], + "lines": [ + { + "bbox": [ + 309, + 142, + 541, + 152 + ], + "spans": [ + { + "bbox": [ + 309, + 142, + 541, + 152 + ], + "type": "text", + "content": "Figure 7. Impact of the number of MAT blocks on performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 165, + 544, + 289 + ], + "blocks": [ + { + "bbox": [ + 334, + 154, + 517, + 163 + ], + "lines": [ + { + "bbox": [ + 334, + 154, + 517, + 163 + ], + "spans": [ + { + "bbox": [ + 334, + 154, + 517, + 163 + ], + "type": "text", + "content": "Table 2. Results of ablation studies on MetaFaces." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 165, + 544, + 289 + ], + "lines": [ + { + "bbox": [ + 308, + 165, + 544, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 544, + 289 + ], + "type": "table", + "html": "
FID ↓SWD ↓Texture ↑Color ↑Semantic↑
w/o MAT27.722.30.8830.8520.856
w/o ReLU30.120.00.9160.9560.928
w/o MAT (→ Full Corr. [45])34.119.70.8720.9690.902
w/o AdaConv (→ FFN [41])34.320.10.8410.9710.896
w/o Lstyle30.320.80.8740.8480.908
w/o Lstyle (→ LCAST [47])31.019.80.9040.9830.921
w/o Lstr28.421.80.9150.9830.911
w/o skip connection47.027.70.9250.9580.902
w/o global style z27.621.20.9270.9610.925
Full Model26.019.10.9380.9840.927
", + "image_path": "4ceefe2e5d338cc6c84f870bf797182fde28d5cb37cbdf5c57f7b6926f6c20c5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 294, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 510 + ], + "type": "text", + "content": "MAT (i.e. w/o ReLU), (3) replacing MAT with three-layer full correspondence learning modules [45] (i.e. Full Corr.), and (4) replacing the AdaConv with FFN [45] (i.e. w/ AdaConv). The results in Table 2 show that removing MAT or ReLU dramatically hurts the performance. Besides, using the full correspondence learning in [45] or using FFN also significantly decreases the texture relevance and semantic consistency. Correspondingly, these model variants leads to inferior results in terms of textures or colors, compared to our full model (Fig. 8). Recall the visualized correspondence in Fig. 1, our method learns remarkably accurate correspondence, which ultimately benefits the quality of generated images. In addition, Fig. 7 shows that both the semantic consistency and style realism broadly improve with the number of MAT blocks and peak at three. All these observations demonstrate our motivation that MAT gradually refines cross-domain correspondence and augments informative representations for generating high-quality images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Contrastive Style Loss. To verify the effectiveness of the proposed contrastive style learning methodology, we train our model by (1) removing the style loss (i.e. w/o " + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{style}}" + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": ") and (2) replacing " + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{style}}" + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": " with the loss used in CAST [47] (i.e. w/ " + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{CAST}}" + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "). In CAST, only high-quality exemplars in different styles are used as negative samples. Differently, we use low-quality generated images in diverse styles as negative samples. From both Table 2 and Fig. 8, we observe that: (1) without " + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{style}}" + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": ", although the generated images show high semantic consistency, they present low style relevance; (2) " + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{CAST}}" + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": " benefits the style relevance, but leads to inferior performance to " + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{style}}" + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": ". These comparison results meet our expectation that: our CSL methodology enables the learned style codes to discriminate subtle divergences between images with different perceptual qualities. Such discriminability facilitates pushes the network to generate high-quality images." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22424" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 126 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "type": "image", + "image_path": "13f1705bb5463520e19bff57e698365e044b700db4c94da9ab6f774b55814bb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 128, + 470, + 138 + ], + "lines": [ + { + "bbox": [ + 122, + 128, + 470, + 138 + ], + "spans": [ + { + "bbox": [ + 122, + 128, + 470, + 138 + ], + "type": "text", + "content": "Figure 8. Comparison of generated images by different variants of our model, on Metfaces [18]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 142, + 272, + 312 + ], + "blocks": [ + { + "bbox": [ + 59, + 142, + 272, + 312 + ], + "lines": [ + { + "bbox": [ + 59, + 142, + 272, + 312 + ], + "spans": [ + { + "bbox": [ + 59, + 142, + 272, + 312 + ], + "type": "image", + "image_path": "1180542e4d15edc9093f885a172b0d4844e0d428d69c5388191c0dcd65d46671.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 312, + 286, + 334 + ], + "lines": [ + { + "bbox": [ + 47, + 312, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 286, + 334 + ], + "type": "text", + "content": "Figure 9. Our method can transfer facial photos to artistic portraits in the style of exemplars." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 340, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 340, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 286, + 411 + ], + "type": "text", + "content": "Skip connections & global style control. In MATEBIT, we use skip connections to supplement input semantic information. Removing skip connections dramatically hurts the semantic inconsistency and the quantitative results. Besides, using global style vector " + }, + { + "bbox": [ + 47, + 340, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 340, + 286, + 411 + ], + "type": "text", + "content": " increases subtle details, e.g. the colors over the mouth, rings, and hairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 286, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 286, + 497 + ], + "type": "text", + "content": "In summary, MAT learns accurate correspondence and enables context-aware feature augmentation; the contrastive style learning benefits the style control and high-quality image generation; and the U-Net architecture helps the preservation of semantic information. Ultimately, all these benefits make our model significantly outperform previous state-of-the-art methods in generating plausible images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 502, + 129, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 129, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 129, + 514 + ], + "type": "text", + "content": "4.3. Applications" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 604 + ], + "type": "text", + "content": "Artistic Portrait Generation. An potential application of our method is transferring a facial photo to an artistic portrait, in the style of an exemplar. We here apply the previously learned models to randomly selected facial photos from CelebA-HQ [22]. As illustrated in Fig. 9, our method can generate appealing portraits with consistent identity and faithful style appearance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Chinese Ink Painting Generation. To verify the capacity of our model in generating complex images, we additionally apply it to generate Chinese Ink paintings. Specifically, we collect paintings of landscapes and facial portraits from the web, and then train and test our model on each subset respectively. Fig. 10 illustrates the results of painting generation and photo-to-painting translation. Obviously, all the generated images show remarkably high quality. Besides, our model successfully captures subtle differences" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 142, + 529, + 312 + ], + "blocks": [ + { + "bbox": [ + 321, + 142, + 529, + 312 + ], + "lines": [ + { + "bbox": [ + 321, + 142, + 529, + 312 + ], + "spans": [ + { + "bbox": [ + 321, + 142, + 529, + 312 + ], + "type": "image", + "image_path": "8856d757dc719f66384509178a9a7088c7264a071a6c5ba2acc263f040c362b2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 313, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 306, + 313, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 313, + 545, + 335 + ], + "type": "text", + "content": "Figure 10. Chinese ink paintings generation (1st & 3rd rows), as well as photo-to-painting translation (2nd & 4th rows)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 341, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 341, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 545, + 365 + ], + "type": "text", + "content": "between different exemplars, demonstrating its remarkable capacity in style control." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 372, + 383, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 383, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 383, + 384 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 393, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 393, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 545, + 536 + ], + "type": "text", + "content": "This paper presents a novel exemplar-guided image translation method, dubbed MATEBIT. Both quantitative and qualitative experiments show that MATEBIT is capable of generating high-fidelity images in a number of tasks. Besides, ablation studies demonstrate the effectiveness of MAT and contrastive style learning. Despite such achievements, the artistic portraits transferred from facial photos (Fig. 9) are inferior to those shown in Fig. 6. This may be due to the subtle differences in edge maps between photos and artistic paintings. In the near future, we will explore to solve this issue via semi-supervised learning or domain transfer technologies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 549, + 408, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 408, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 408, + 562 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2018AAA0103202; in part by the National Natural Science Foundation of China under Grants 61971172, 61971339, 62176230, U22A2033, U22A2096; in part by the Zhejiang Provincial Science and Technology Program in China under Grant 2021C01108; in part by the NSFC-Zhejiang Joint Fund for the Integration of Industrialization and Informatization under Grant U1909210; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; and in part by Open Research Projects of Zhejiang Lab under Grant 2021KG0AB01." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22425" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 180 + ], + "type": "text", + "content": "[2] Huiwen Chang, Jingwan Lu, Fisher Yu, and Adam Finkelstein. Pairedcyclegan: Asymmetric style transfer for applying and removing makeup. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 40-48, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 227 + ], + "type": "text", + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 283 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 328 + ], + "type": "text", + "content": "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 396 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 398, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 398, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 287, + 441 + ], + "type": "text", + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 444, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 287, + 497 + ], + "type": "text", + "content": "[8] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 500, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 287, + 554 + ], + "type": "text", + "content": "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "text", + "content": "[10] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11936-11945, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[12] Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13713-13722, 2021. 2, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Xueqi Hu, Xinyue Zhou, Qiusheng Huang, Zhengyi Shi, Li Sun, and Qingli Li. Qs-attn: Query-selected attention for contrastive learning in i2i translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18291-18300, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "text", + "content": "[14] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "text", + "content": "[15] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "text", + "content": "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "type": "text", + "content": "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "type": "text", + "content": "[18] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[19] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 416, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 446 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 470 + ], + "type": "text", + "content": "[21] Thomas N Kipf and Max Welling. Variational graph autoencoders. arXiv preprint arXiv:1611.07308, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "type": "text", + "content": "[22] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549-5558, 2020. 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "text", + "content": "[23] Boyi Li, Felix Wu, Kilian Q Weinberger, and Serge Belongie. Positional normalization. In Advances in Neural Information Processing Systems, pages 1620-1632, 2019. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "text", + "content": "[24] Mingcong Liu, Qiang Li, Zekui Qin, Guoxin Zhang, Pengfei Wan, and Wen Zheng. Blendgan: implicitly gan blending for arbitrary stylized face generation. Advances in Neural Information Processing Systems, 34:29710-29722, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 615, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 615, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 615, + 545, + 657 + ], + "type": "text", + "content": "[25] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In European Conference on Computer Vision, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22426" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "text", + "content": "[28] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "text", + "content": "[29] Mandi Luo, Haoxue Wu, Huaibo Huang, Weizan He, and Ran He. Memory-modulated transformer network for heterogeneous face recognition. IEEE Transactions on Information Forensics and Security, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[30] Liqian Ma, Xu Jia, Stamatios Georgoulis, Tinne Tuytelaars, and Luc Van Gool. Exemplar guided unsupervised image-to-image translation with semantic consistency. Proceedings ICLR 2019, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "text", + "content": "[31] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In Proceedings of the European conference on computer vision (ECCV), pages 768-783, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 369 + ], + "type": "text", + "content": "[32] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel flow: Semantic correspondence with multi-layer neural features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3395-3404, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "text", + "content": "[33] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[34] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In European conference on computer vision, pages 319-345. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[35] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 1, 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 535 + ], + "type": "text", + "content": "[36] Justin N. M. Pinkney. Aligned ukiyo-e faces dataset, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "text", + "content": "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "text", + "content": "[38] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, August 2020. version 0.2.1. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[39] Junyoung Seo, Gyuseong Lee, Seokju Cho, Jiyoung Lee, and Seungryong Kim. Midms: Matching interleaved diffusion models for exemplar-based image translation. arXiv preprint arXiv:2209.11047, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR, 2015. 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 597 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "type": "text", + "content": "[42] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision, pages 63-79. Springer, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 546, + 206 + ], + "type": "text", + "content": "[43] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 207, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 547, + 262 + ], + "type": "text", + "content": "[44] Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, and Changgong Zhang. Marginal contrastive correspondence for guided image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10663-10672, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 263, + 546, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 546, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 546, + 318 + ], + "type": "text", + "content": "[45] Pan Zhang, Bo Zhang, Dong Chen, Lu Yuan, and Fang Wen. Cross-domain correspondence learning for exemplar-based image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5143-5153, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 319, + 546, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 546, + 373 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 374, + 546, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 374, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 374, + 546, + 419 + ], + "type": "text", + "content": "[47] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH, 2022. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 420, + 546, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 546, + 474 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 546, + 474 + ], + "type": "text", + "content": "[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 475, + 546, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 546, + 540 + ], + "type": "text", + "content": "[49] Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocosnet v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11465–11475, 2021. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 542, + 546, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 542, + 546, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 542, + 546, + 597 + ], + "type": "text", + "content": "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22427" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_content_list.json b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3099a68603ac764f028553930f30aa746dd4d59c --- /dev/null +++ b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_content_list.json @@ -0,0 +1,1805 @@ +[ + { + "type": "text", + "text": "Master: Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer", + "text_level": 1, + "bbox": [ + 109, + 128, + 861, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Tang $^{1*}$ Songhua Liu $^{2*}$ Tianwei Lin $^{3}$ Shaoli Huang $^{4}$ Fu Li $^{3}$ Dongliang He $^{3}$ Xinchao Wang $^{2\\dagger}$", + "bbox": [ + 241, + 202, + 738, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Center for Data Science, Peking University $^{2}$ National University of Singapore", + "bbox": [ + 227, + 239, + 754, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ VIS, Baidu Inc. $^{4}$ Tencent AI Lab", + "bbox": [ + 362, + 257, + 601, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tanghao@stu.pku.edu.cn songhua.liu@u.nus.edu", + "bbox": [ + 285, + 277, + 696, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lintianwei01,lifu,hedongliang01}@baidu.com shaolihuang@tencent.com xinchao@nus.edu.sg", + "bbox": [ + 86, + 294, + 880, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2265e982334aa3d38d302ca2cd0acb11c8f855a706e3c1f58c5e686b5be6a7d3.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 76, + 316, + 212, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7116b00247e4e9aa669574d1c349a4f34f15acb837a839f8ca2d4d8ea67e370.jpg", + "image_caption": [ + "(b)", + "Figure 1. Left: With more Transformer layers, the number of parameters and training difficulty increase significantly. Right: Our method supports test-time controlling of stylization degree via tuning the number of adopted Transformer layers. It is also readily adaptable to few-shot style transfer, where stylization with only 1 layer can be further improved. Text-guided few-shot style transfer is achievable." + ], + "image_footnote": [], + "bbox": [ + 217, + 316, + 359, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4729d7188c270f0708fa04f15441404551f99646e6687f632d04da8c3946734a.jpg", + "image_caption": [ + "Content/Style" + ], + "image_footnote": [], + "bbox": [ + 364, + 320, + 447, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a002410cbc987d143e156f13f952eac456cc3a0c7b3485606bd80dc4d775d76.jpg", + "image_caption": [ + "Layer 1" + ], + "image_footnote": [], + "bbox": [ + 450, + 321, + 537, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0397542bebbcff4821cc5b411605abc84da086049ea6fbee22ed950c4a0291e.jpg", + "image_caption": [ + "Layer 3" + ], + "image_footnote": [], + "bbox": [ + 540, + 321, + 625, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40324f219537218ca4ed4779076be71593839fa09484551872352291beca21b5.jpg", + "image_caption": [ + "Layer 5" + ], + "image_footnote": [], + "bbox": [ + 630, + 321, + 714, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35755527d19bfc364ca5ffc42f6860c3ab7720adba3e4c155eb979032abf5e41.jpg", + "image_caption": [ + "Few Shot" + ], + "image_footnote": [], + "bbox": [ + 717, + 321, + 802, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da914c77b9f2de5ee7f43e92906edb89a2753d0ba9afa1d21147b3b26e2559cb.jpg", + "image_caption": [ + "\"The scream by Edvard Munch\"" + ], + "image_footnote": [], + "bbox": [ + 805, + 321, + 890, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 465, + 313, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformer-based models achieve favorable performance in artistic style transfer recently thanks to its global receptive field and powerful multi-head/layer attention operations. Nevertheless, the over-paramerized multilayer structure increases parameters significantly and thus presents a heavy burden for training. Moreover, for the task of style transfer, vanilla Transformer that fuses content and style features by residual connections is prone to content-wise distortion. In this paper, we devise a novel Transformer model termed as Master specifically for style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters, which (1) reduces the total number of parameters, (2) leads to more robust training convergence, and (3) is readily to control the degree of stylization via tuning the number of stacked layers freely during inference. On the other hand, different from the vanilla version, we adopt a learnable scaling operation on content features before content-style feature interaction, which better preserves the original similarity between a pair of content features while ensuring the stylization quality. We also propose a novel meta learning scheme for the proposed model so that it can not only work in the typical setting of arbitrary style transfer, but also adaptable to the few-shot setting, by only fine-tuning the Transformer encoder layer in the few-shot stage for one", + "bbox": [ + 75, + 487, + 472, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "specific style. Text-guided few-shot style transfer is firstly achieved with the proposed framework. Extensive experiments demonstrate the superiority of Master under both zero-shot and few-shot style transfer settings.", + "bbox": [ + 500, + 467, + 892, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 556, + 632, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Artistic style transfer aims at applying style patterns like colors and textures of a reference image to a given content image while preserving the semantic structure of the content. In contrast to the pioneering optimization method [9] and early per-style-per-model methods like [17, 26], arbitrary style transfer methods [3, 13, 16, 21, 22, 24, 31] enable real time style transfer for any style image in the test time in a zero-shot manner. The flexibility has led to this arbitrary-style-per-model fashion to dominate style transfer research.", + "bbox": [ + 496, + 582, + 890, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, to enhance the representation of global information in arbitrary style transfer, Transformer [40] is introduced to this area [4], leveraging the global receptive field and powerful multi-head/layer structure, and achieves superior performance. Nevertheless, the over-parameterized multi-layer structure increases model parameters significantly. As shown in Fig. 1(a), there are 25.94M learnable parameters for a 3-layer Transformer structure in StyTr2 [4], v.s. 3.50M in AdaIN [13], a simple but effective baseline in arbitrary style transfer. Such a large model for standard Transformer inevitably presents a heavy burden for training. As shown in Fig. 1(b), when there are more than", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18329", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acf3d3dbaedee70ad481c29fb68b35357f3ce3b3eac58ecc396221abd620540a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 88, + 468, + 145 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c7782a5691269e99e26b011c6c4e1f635c1de7b3e3bfc1a23aafda69325db50a.jpg", + "image_caption": [ + "Style /", + "Content", + "Figure 2. Residual connection in the vanilla Transformer tends to destroy the original similarity relationship on content structure in style transfer task. Our model is designed to address this problem with learnable scaling parameters. The top row shows a simple 2D visualization and the bottom one provides a qualitative example." + ], + "image_footnote": [], + "bbox": [ + 80, + 148, + 194, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/334709ab8081e5446453ce540b589cd652765d759204889f2c21e15448f6b559.jpg", + "image_caption": [ + "Vanilla Transformer", + "w Residual Connection" + ], + "image_footnote": [], + "bbox": [ + 215, + 148, + 326, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/07eaa7e3cd6c4e7e2d6df88b917127ce9d1f0ece8a6651896c3c817781428a86.jpg", + "image_caption": [ + "Our Transformer", + "w Learnable Scaling" + ], + "image_footnote": [], + "bbox": [ + 351, + 148, + 462, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4 layers, vanilla Transformer even fails to get convergent in training, which limits the scalability of the Transformer model in style transfer.", + "bbox": [ + 75, + 352, + 468, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, vanilla Transformer relies on residual connections [12] to stylize content features, which suffers from the content-distortion problem. We illustrate this effect with a 2D visualization in Fig. 2(top), where residual connections lead the transformation results of two content feature vectors to move towards the dominated style features and thereby tend to eliminate their original distinction. The visual effect is that the stylized images would be dominated by some strong style features, such as salient edges, with the original self-(dis)similarity of content structures destroyed, as the example shown in Fig. 2(bottom).", + "bbox": [ + 75, + 398, + 468, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Focusing on these drawbacks, in this paper, we are dedicated to devising a novel Transformer architecture specifically for artistic style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters and a random number of stacked layers are adopted for each training iteration. Compared with the original version, sharing parameters across different layers reduces the total number of parameters significantly and leads to more convenient training convergence. As a byproduct, it is also readily for our model to control the degree of stylization via tuning the number of stacked layers freely in the inference time, as shown in Fig. 1(right). On the other hand, we equip Transformer with learnable scale parameters for content-style interactions instead of residual connections, which alleviates content distortion to a large degree and better preserves content structures while rendering vivid style patterns simultaneously, as shown in the 2D visualization and the qualitative example in Fig. 2.", + "bbox": [ + 75, + 566, + 468, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, beyond the typical zero-shot arbitrary style transfer, leveraging a meta learning scheme, our method is adaptable to the few-shot setting. By only fine-tuning the Transformer encoder layer in the few-shot stage, rapid", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adaptation to the model for a specific style within a limited number of updates is possible, where the stylization with only 1 layer can be further improved, as shown in Fig. 1. Beyond that, we first achieve text-guided few-shot style transfer with this framework, which largely alleviates the training burden of previous per-text-per-model solution. In this sense, we term the overall pipeline Meta Style Transformer (Master). Our contributions are summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Transformer architecture specifically for artistic style transfer. It shares parameters between different layers, which not only helps training convergence, but also allows convenient control over the stylization effect.", + "- We identify the content distortion problem of residual connections in Transformer and propose learnable scale parameters as an option to alleviate the problem.", + "- We introduce a meta learning framework for adapting original training setting of zero-shot style transfer to the few-shot scenario, with which our Master achieves very good trade-off between flexibility and quality.", + "- Experiments show that our model achieves results better than those of arbitrary-style-per-model methods. Furthermore, under the few-shot setting, either conditioned on image or text, Master can even yield performance on par with that of per-style-per-model methods with significantly less training cost." + ], + "bbox": [ + 517, + 239, + 890, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 571, + 648, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Arbitrary-Style-Per-Model Methods. A lot of works achieve arbitrary style transfer via global feature transformation, e.g., WCT [22], AdaIN [13], Linear style transfer [21], DIN [15], MCCNet [2], and MAST [14]. In general, they can achieve the most attainable style transfer speed but dismiss stylized effects for local details a lot. In order to add more consideration of local details to global transformation based methods, there are patch-similarity based solution [1, 10, 37] and attention based method [3, 24, 31, 44]. While paying attention to local details, it is still hard to transfer complex style patterns and prone to unsatisfactory distortions due to the simple swap and fusion strategies.", + "bbox": [ + 496, + 597, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer in Style Transfer. Transformer proposed in [40] is widely used in the natural language processing and has become a powerful baseline. Recently, Transformer receives better performance than CNN models in many vision tasks and are involved in various model zoos [6, 11, 18, 42, 43, 45]. Specifically, in the field of style transfer, [41] propose StyFormer, where the transformation is driven by the", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18330", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/660f6ed346876c1f9579582f50f1dbc7664726a31894ee5ac71a3b2cf7b661c1.jpg", + "image_caption": [ + "Figure 3. Overview of our model architecture." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 895, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cross-attention module in Transformer. [4] develop a Transformer model based on ViT [5]. However, the number of parameters and training difficulty would increase considerably with the increasing of Transformer layers. Also, following residual connections in original Transformer, it suffers from the distortion problem on content structures as shown in Fig. 2. Similar problems also exist in [27].", + "bbox": [ + 75, + 330, + 468, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta Learning in Style Transfer. On this routine, [36] train a meta network to predict parameters of the generator model for one style reference image. [47] employ MAML algorithm [7] to find a style-free model for fast adaptation. Our method is different from theirs in three aspects: (1) our framework works for both few-shot and arbitrary style transfer thanks to the cross-attention in Transformer; (2) our meta learning algorithm is based on Reptile [30], a first-order meta learning algorithm without the necessity to operate gradients of higher levels, which is more efficient in training; and (3) only style encoder instead of the whole model needs to be updated during the few-shot learning stage, which makes it more convenient in practice.", + "bbox": [ + 75, + 436, + 468, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-Guided Synthesis. The emergence of CLIP model [34] bridges text and image domain, which supports a series of works on text-guided synthesis [8, 32, 35]. In the field of style transfer, CLIPstyler [20] proposes a patch-wise CLIP loss and achieves text-guided style transfer. Nevertheless, as a per-text-per-model solution, it is inconvenient in practice to handle a large number of text inputs. Despite the recent dataset distillation scheme may alleviate this issue, existing approaches, however, largely focus on images as input [25, 28, 46]. In this paper, we first achieve text-guided few-shot style transfer with the proposed meta learning algorithm, which improves the flexibility of CLIPstyler significantly.", + "bbox": [ + 75, + 633, + 470, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 844, + 174, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we give details of the proposed Meta Style Transformer (Master) for zero-shot and few-shot style", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transfer. We first introduce the network architecture of the proposed model, then illustrate how we train our model in a meta-learning fashion, and finally describe loss function.", + "bbox": [ + 498, + 330, + 892, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Network Architecture", + "text_level": 1, + "bbox": [ + 500, + 386, + 705, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed model comprises an encoder, a feature modification module, and a decoder, as demonstrated in Fig. 3. We employ the first 2 stages of Swin Transformer [29] as the encoder Enc to extract common image features for both content and style images. The decoder Dec follows the setting of [13] with 3 upsampling convolutional blocks. For feature modification, we propose a Style Transformer module for transferring complex style patterns, which will be introduced later. Taking a style image $I_{s}$ and a content image $I_{c}$ as input, we first divide them into $4 \\times 4$ patches and extract their corresponding feature maps $F_{c}$ and $F_{s}$ with the Swin encoder Enc:", + "bbox": [ + 498, + 410, + 892, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {c} = \\operatorname {E n c} (I _ {c}), \\\\ F _ {s} = \\operatorname {E n c} \\left(I _ {s}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 601, + 890, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the spatial scales for $F_{c}$ and $F_{s}$ are 1/8 of those for $I_{c}$ and $I_{s}$ . Then, the embedded feature $F_{cs}$ is derived by:", + "bbox": [ + 500, + 648, + 892, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} = \\operatorname {S t y l e T r a n s} \\left(F _ {s}, F _ {c}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 691, + 890, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where StyleTrans denotes the Style Transformer module, and $F_{cs}$ has the same shape as $F_{c}$ . Finally, we can synthesize the stylized image $I_{cs}$ with the decoder as:", + "bbox": [ + 498, + 720, + 890, + 767 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c s} = \\operatorname {D e c} \\left(F _ {c s}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 780, + 890, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Style Transformer. Similar to previous Transformer-based style transfer models like StyTr2 [4], the Transformer encoder is used to encode style information while the Transformer decoder takes charge of content-style interaction. In our model, there is only one copy of parameters shared by all Transformer encoder and decoder layers. Also, different", + "bbox": [ + 498, + 809, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18331", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from standard Transformer where decoder layers would follow all encoder layers, encoder and decoder layers would be executed in an alternate fashion. The next layer would take the output of the current layer as input.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our Transformer decoder layer is composed of self-attention, cross-attention, and non-linear blocks, for content encoding and content-style interaction. Specifically, content features are first processed with a self-attention step. Then, cross-attention is conducted by taking the content encoding as query and the style encoding as key and value, followed by a MLP for non-linear transformation.", + "bbox": [ + 75, + 152, + 467, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, in vanilla Transformer, features before and after the cross-attention step are fused by a residual connection, which is harmful for content structures as analyzed in Fig. 2. We thereby replace the residual connection with dynamic and learnable scaling and shifting steps, whose parameters are determined by the style encoder. In this way, the output of style encoder should consist of three parts: key for the following cross-attention $K_{s}$ , scaling parameters $V_{\\sigma}$ , and shifting parameters $V_{\\mu}$ . The prediction of each part shares a same self-attention map to save memory but uses independent non-linear transformation. The process in the Transformer encoder layer can be formulated as:", + "bbox": [ + 75, + 260, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M H A} (Q, K, V) = [ h e a d _ {1}, \\dots , h e a d _ {h} ] W ^ {O},\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 455, + 387, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {h e a d} _ {i} = \\operatorname {A t t} \\left(Q W ^ {Q _ {i}}, K W ^ {K _ {i}}, V W ^ {V _ {i}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 476, + 426, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t} (Q, K, V) = \\operatorname {s o f t m a x} (\\frac {Q K ^ {\\top}}{\\sqrt {d _ {k}}} V),\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 497, + 361, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} K _ {s} ^ {\\prime} = K _ {s} + \\operatorname {M H A} \\left(K _ {s}, K _ {s}, K _ {s}\\right), \\\\ \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i l t e} \\end{array} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 534, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {s} = K _ {s} ^ {\\prime} + \\operatorname {M L P} \\left(K _ {s} ^ {\\prime}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 553, + 344, + 569 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\sigma} ^ {\\prime} = V _ {\\sigma} + \\operatorname {M H A} \\left(K _ {s}, K _ {s}, V _ {\\sigma}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 571, + 398, + 588 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\sigma} = V _ {\\sigma} ^ {\\prime} + \\operatorname {M L P} \\left(V _ {\\sigma} ^ {\\prime}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 590, + 339, + 607 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\mu} ^ {\\prime} = V _ {\\mu} + \\mathrm {M H A} \\left(K _ {s}, K _ {s}, V _ {\\mu}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 609, + 398, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\mu} = V _ {\\mu} ^ {\\prime} + \\operatorname {M L P} \\left(V _ {\\mu} ^ {\\prime}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 628, + 339, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K_{s}, V_{\\sigma}$ , and $V_{\\mu}$ are initialized as the style feature $F_{s}$ before the first Transformer encoder layer. We do not incorporate normalization to encode style features since second-order statistics can largely represent style information.", + "bbox": [ + 75, + 660, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, in the cross-attention of Transformer decoder, scaling and shifting parameters for each content feature point is aggregated from $V_{\\sigma}$ and $V_{\\mu}$ respectively according to the cross-attention map. The process in one Transformer decoder block can be written as:", + "bbox": [ + 75, + 723, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} ^ {\\prime} = F _ {c s} + \\mathrm {M H A} \\left(F _ {c s}, F _ {c s}, F _ {c s}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 810, + 383, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma = \\operatorname {M H A} \\left(\\operatorname {I N} \\left(F _ {c s} ^ {\\prime}\\right), \\operatorname {I N} \\left(K _ {s}\\right), V _ {\\sigma}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 829, + 393, + 847 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\mathrm {M H A} (\\mathrm {I N} \\left(F _ {c s} ^ {\\prime}\\right), \\mathrm {I N} \\left(K _ {s}\\right), V _ {\\mu}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 849, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} ^ {\\prime \\prime} = \\sigma \\odot F _ {c s} ^ {\\prime} + \\mu ,\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 867, + 287, + 883 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} = F _ {c s} ^ {\\prime \\prime} + \\operatorname {M L P} \\left(F _ {c s} ^ {\\prime \\prime}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 886, + 318, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Meta Training", + "bbox": [ + 500, + 90, + 687, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: $\\mathcal{D}_c$ : content dataset; $\\mathcal{D}_s$ : style dataset; $\\delta$ : inner learning rate; $\\eta$ : outer learning rate; $k$ : number of inner updates; $T$ : maximal number of stacked layers;", + "bbox": [ + 500, + 109, + 890, + 155 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ensure: trained meta generator parameters $\\theta$", + "bbox": [ + 500, + 155, + 802, + 170 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: initialize $\\theta$ randomly", + "2: for iteration 1, 2, 3, ... do", + "3: sample a style image $I_{s}$ from $\\mathcal{D}_s$", + "4: $\\omega \\gets \\theta$", + "5: for $k$ times do", + "6: sample a batch of content image $I_{c}$ from $\\mathcal{D}_c$", + "7: sample the number of layers from $1\\sim T$", + "8: forward propagation using Eq. 1-5", + "9: compute inner loss $L$ using Eq. 8", + "10: $\\omega \\gets \\omega -\\delta \\nabla L$", + "11: end for", + "12: $\\theta \\gets \\theta +\\eta (\\omega -\\theta)$", + "13: end for" + ], + "bbox": [ + 504, + 171, + 867, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where IN denotes instance normalization [39] and $\\odot$ represents element-wise multiplication. $F_{cs}$ is initialized as $F_{c}$ before the first Transformer decoder layer.", + "bbox": [ + 498, + 396, + 890, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training Pipeline", + "text_level": 1, + "bbox": [ + 500, + 453, + 668, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve high-quality style transfer, we introduce a two-stage training strategy that comprises meta training and fast adaptation. The meta training stage is designed to learn a generic model initialization, while the fast adaptation adapts the network for a single style in a few iterations for few-shot style transfer. Note that zero-shot style transfer is a special case of the overall training configuration.", + "bbox": [ + 496, + 477, + 890, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meta Training: Inspired by Reptile [30], a first-order meta learning algorithm, rendering style patterns of a specific reference style image can be viewed as a task. We seek an optimal initialization for neural networks in this stage, so that the networks can be rapidly adapted for a new task in only a few shots. The main training procedure for this stage is shown in Algorithm 1. In each iteration, we sample 1 style and $k$ batches of contents to perform inner optimization to obtain \"fast weights\" $\\omega$ , which would later guide the update of \"slow weights\" $\\theta$ , to move a step in the direction of $\\omega - \\theta$ . Notably, as there is only one group of parameters for Transformer encoder and decoder layers, we randomly choose a number as the number of stacked layers for the Style Transformer in each iteration.", + "bbox": [ + 496, + 583, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fast Adaptation: The trained model after the first stage would serve as an initialization and will be adapted for a single style. With the same objective, this stage behaves almost identically to the internal loop in Algorithm 1. The only difference is that only parameters of the Transformer encoder layer are necessary to be updated, since (1) the Transformer encoder layer is the main component to extract style pat", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18332", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "terns and has the most significant impact on stylization, and (2) it would save memory and speed up the adaptation.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-Shot Style Transfer: \"Zero-Shot\" means that, once the meta training is done, there is no fast adaptation stage needed and the meta model itself can support arbitrary style transfer. To encourage the model to produce satisfactory results in the zero-shot setup, we set the inner optimization time $k$ to 1, where the algorithm is reduced to the typical training paradigm of existing arbitrary style transfer methods. In this sense, Algorithm 1 provides a more general setting for style transfer in both zero-shot and few-shot cases.", + "bbox": [ + 75, + 121, + 467, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-Guided Style Transfer. We perform text-guided style transfer based on our image style transfer model with slight modifications. Following the common practice, a pretrained CLIP encoder is imported to extract text features. To be consistent with image input, we use a StyleGAN-like mapping network to convert text features into pseudo image features. In the meta training stage, since CLIP unifies the feature space, we use image instead of text as style input to avoid additional text dataset. In the fast adaptation stage, the mapping network along with the Transformer encoder is updated, considering there is still a gap between CLIP image feature and text feature.", + "bbox": [ + 75, + 257, + 467, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 446, + 220, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training objective in both meta-training and fast adaptation stages follows many works in arbitrary style transfer, which consists of content loss and style loss. Let $F^x$ represent features on ReLU- $x$ -1 layer of a pre-trained VGG19 network [38] for loss computation. The content loss is defined by the normalized perceptual loss [17]:", + "bbox": [ + 75, + 470, + 467, + 560 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t}} = \\sum_ {x = 2} ^ {5} \\| \\mathrm {I N} \\left(F _ {c} ^ {x}\\right) - \\mathrm {I N} \\left(F _ {c s} ^ {x}\\right) \\| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 561, + 467, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "while the style loss adopts the mean-variance loss [13]:", + "bbox": [ + 80, + 609, + 444, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s t y} = \\sum_ {x = 2} ^ {5} \\left(\\| \\mu \\left(F _ {s} ^ {x}\\right) - \\mu \\left(F _ {c s} ^ {x}\\right) \\| _ {2} + \\| \\sigma \\left(F _ {s} ^ {x}\\right) - \\sigma \\left(F _ {c s} ^ {x}\\right) \\| _ {2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 627, + 467, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\sigma$ calculate the channel-wise mean and standard deviation separately. The overall objective is given by the weighted summation of the two losses:", + "bbox": [ + 75, + 681, + 467, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t}} + \\lambda \\mathcal {L} _ {\\text {s t y}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 729, + 467, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ controls the balance between two terms.", + "bbox": [ + 75, + 750, + 398, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In text-guided style transfer, the loss functions are the same as CLIPstyler, including global CLIP loss, directional CLIP loss and PatchCLIP loss.", + "bbox": [ + 75, + 763, + 467, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 821, + 207, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 847, + 292, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use MS-COCO [23] as our content dataset and WikiArt test set [33] as our style dataset. The content", + "bbox": [ + 76, + 869, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4ac380fb9dd8fe54cac32269b9325afc0bb32536d34f11d12d33e05a2ebe8f99.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative comparisons. ZS and FS for our model denote zero-shot and few-shot modes. Vanilla denote replacing our architecture with original Transformer. Norm means adding layer normalization in the Transformer encoder layer. L1/L3 means using 1/3 Transformer layers in the test time. * and † denote few-shot and per-style-per-model methods." + ], + "table_body": "
MethodLcontLsimLsty
AdaIN4.88±0.700.53±0.231.51±0.78
Linear3.93±0.760.44±0.171.97±0.96
AvatarNet5.76±0.530.56±0.193.19±1.82
SANet4.72±0.610.50±0.171.15±0.58
MANet4.93±0.570.50±0.171.41±0.77
MCCNet4.22±0.690.47±0.171.56±0.84
AdaAttN4.46±0.700.43±0.162.20±1.19
StyTr23.78±0.990.48±0.221.50±0.69
StyFormer4.94±0.790.43±0.152.20±1.44
MetaNet3.48±0.850.45±0.192.47±1.69
MetaStyle*3.64±1.120.42±0.182.47±1.06
Johnson†4.60±0.760.59±0.201.02±0.34
Ours-Vanilla5.50±0.630.59±0.260.85±0.38
Ours-Norm4.70±0.750.43±0.130.93±0.33
Ours-ZS-L14.13±0.680.41±0.140.92±0.40
Ours-ZS-L34.20±0.680.41±0.130.81±0.31
Ours-FS*4.24±0.820.38±0.120.79±0.25
", + "bbox": [ + 519, + 88, + 867, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset contains roughly 80,000 images and the style dataset has about 20,000 images. The optimizer is Adam [19] with learning rates of both inner and outer loops set as 0.0001 and the batch size is 4. In training, we first resize the content and style image to $512 \\times 512$ and then randomly crop to $256 \\times 256$ resolution. During inference, our model can handle inputs of any size. The update times for inner optimization $k$ is set as 2 for few-shot case and 1 for zero-shot case. In training, the maximal number of stacked layers is 4. All the multi-head attention blocks are instantiated as shifted window attention in [29], with window size 8 and shift size 4. The model is trained on a Nvidia 3090 with 9k iterations in the meta training stage for convergence while only 100 steps for image input and 20 steps for text input in the fast adaptation for few-shot style transfer, which takes less than 1 minute. Hyper-parameter $\\lambda$ is set as 10.", + "bbox": [ + 496, + 470, + 890, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with Prior Works", + "text_level": 1, + "bbox": [ + 500, + 724, + 769, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare results by our Master with 13 state-of-the-art style transfer methods, including 3 global transformation based methods (AdaIN [13], Linear style transfer [21], and MCCNet [2]), 1 patch swap based method (Avatar-Net [37]), 3 attention based methods (SANet [31], MANet [3], and AdaAttN [24]), 2 transformer based methods (StyTr2 [4] and StyFromer [41]), 2 meta learning based methods (MetaNet [36] and MetaStyle [47]), 1 per-style-per-model method by Johnson et al. [17] and 1 text-guided style transfer method (CLIPstyler).", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18333", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Comparison. We adopt content loss $\\mathcal{L}_{cont}$ in Eq. 6 and style loss $\\mathcal{L}_{sty}$ in Eq. 7 as evaluation metrics to reflect effects of learning by different methods. We also design a metric $\\mathcal{L}_{sim}$ to reflect the preservation of the spatial-wise self cosine similarity of content structures:", + "bbox": [ + 76, + 90, + 468, + 165 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {*, i j} ^ {x} = 1 - \\frac {F _ {*, i} ^ {x} \\cdot F _ {*, j} ^ {x}}{\\| F _ {*, i} ^ {x} \\| \\| F _ {*, j} ^ {x} \\|}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 167, + 346, + 203 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s i m} = \\sum_ {x = 3} ^ {4} \\frac {1}{n _ {x} ^ {2}} \\sum_ {i, j} \\left| \\frac {D _ {c , i j} ^ {x}}{\\sum_ {k} D _ {c , k j} ^ {x}} - \\frac {D _ {c s , i j} ^ {x}}{\\sum_ {k} D _ {c s , k j} ^ {x}} \\right|, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 200, + 467, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n_x$ is the number of spatial locations for the current feature map and the second foot script denotes the spatial index. Smaller $\\mathcal{L}_{sim}$ means that the original spatial-wise relationship is better preserved during style transfer, i.e., less content distortion. We use the test dataset in the code page of [13] for evaluation, with 11 content images and 20 style images to form 220 content-style pairs. We report the mean and standard deviation over the 220 cases in Tab. 1. Notably, the lowest style loss and content similarity loss are achieved by Master simultaneously compared with previous methods, even better than the per-style-per-model solution by Johnson et al. [17], and can be further reduced in the few-shot case with comparable content loss values, which demonstrates the joint advantages of our method in content preserving and style rendering. Meanwhile, the lower standard deviations demonstrate the robustness of our model.", + "bbox": [ + 76, + 250, + 467, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. Qualitative examples by different methods are shown in Fig. 4. Full comparisons with more methods can be found in the supplementary material.", + "bbox": [ + 76, + 492, + 467, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first give discussion on the global-transformation based methods. As shown in the 3rd and 4th rows, in AdaIN and Linear, features in all positions share the same transformation matrix, which fails to migrate detail textures in style images and has poor color saturation.", + "bbox": [ + 76, + 537, + 467, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for the attention-based methods, SANet in the 5th row introduces local focus to style images. However, due to the simple fusion of one-layer attention results and original content features, it brings textures and content distortion in many cases, e.g., background areas of the 4th and 5th columns. AdaAttN in the 6th row dedicates to addressing the distortion problem in SANet, but it sacrifices the ability to render salient style patterns and textures.", + "bbox": [ + 76, + 613, + 467, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then compare our results with those by Transformer-based methods. StyFormer in the 7th row adopts a design of the cross-attention module in Transformer without exploring the effect of self-attention to enhance style rendering. Thus, the results appear under-stylized compared with those by our method. StyTr2 in the 8th row is a Transformer architecture with a pure self-attention mechanism, which makes the performance on the extraction and migration of local textures not satisfactory enough, e.g., strokes of the 2nd column and waves of the 4th column. Moreover, since it follows the design of residual connections in", + "bbox": [ + 76, + 734, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5dabdc7252bdd13c46365be822eca167bb8b12a04afc528073f9d2fb5489fe9.jpg", + "image_caption": [ + "Figure 4. Comparison with previous state-of-the-art style transfer methods. Zoom in for better details." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 887, + 854 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18334", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1721071f92e1d26376da096c2d9d544732fa1a71d395d93c7abc1eb63b3f4852.jpg", + "image_caption": [ + "Figure 5. Comparison with CLIPstyler in text-guided style transfer. Zoom in for better details." + ], + "image_footnote": [], + "bbox": [ + 78, + 93, + 468, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vanilla Transformer, the issue of content distortion in Fig. 2 is still evident, as shown in the background areas of the 4th and 5th columns. Compared with these methods, Master renders more vivid global and local style patterns. Meanwhile, it well maintains content structures simultaneously with learnable scaling and shifting parameters. Furthermore, in the few-shot case, comparable performance can be achieved with only 1 Transformer layer.", + "bbox": [ + 75, + 454, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MetaNet and MetaStyle also involve the concept of meta learning into their style transfer frameworks. On the one hand, MetaNet predicts parameters of a generator network given one style image via a meta network, which is an ambitious goal and makes the training more difficult. Thus, as shown in the 9th row, the stylized effects of their results still appear weak. On the other hand, MetaStyle achieves inferior performance in terms of local details as demonstrated in the 10th row, since only parameters of global normalization layers are learnable in the fast adaptation stage.", + "bbox": [ + 75, + 578, + 468, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The single style transfer method by Johnson et al. in the 11th row tends to arrange patterns learned by the network arbitrarily, which may produce undesired and distorted effects. By contrast, the complex content-style dependence is constructed in our Transformer model and thus achieves more remarkable performance on transferring style patterns to proper positions in content images.", + "bbox": [ + 75, + 731, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Some qualitative comparisons on text-guided style transfer with CLIPstyler are shown in Fig. 5, which uses per-text-per-model fashion. Through these results, we observe that despite significantly more training steps, CLIPstyler", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae3e93bf6b1addfd77baed52ef7ab4a7c4f0b911e56234fde9800a6353b26662.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 99, + 848, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a50828409e302f2615288a1c83c3243beb6c2661eac1362b7bd05dac496bfc1.jpg", + "image_caption": [ + "Figure 6. Results of user study. Left: Our zero-shot stylization results compared with SOTA methods. Right: Our few-shot stylization comparison results.", + "Figure 7. Ablation study on key designs in Master model." + ], + "image_footnote": [], + "bbox": [ + 500, + 250, + 890, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "still suffers from weak stylization. By contrast, our method generates more vivid results in only 20 adaptation steps, which demonstrates the advantages of the proposed Master architecture and the versatility of our training pipeline.", + "bbox": [ + 496, + 338, + 890, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. Following most style transfer works, we conduct user study and report user preference. We choose 20 content images and 20 style images to form 400 content-style pairs. We involve 100 people and randomly assign 20 stylized results from compared methods to each subject. Our method showcases zero-shot results with 3 Transformer layers in 10 tests and few-shot results in the remaining 10. For each pair, the order of results is randomized, and participants choose their favorite. With 2,000 votes in total, Fig. 6 demonstrates Master's superior style transfer quality.", + "bbox": [ + 496, + 400, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 560, + 653, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Architecture. We conduct ablation studies on key designs in Master to illustrate their effects on the stylization quality. As shown in Fig. 7, on the one hand, if vanilla Transformer model is used, without learnable scaling and shifting parameters, noisy textures can be introduced significantly, which distorts the original content structures and affects style transfer quality. On the other hand, if layer normalization operations in the standard Transformer are used in the style encoder, style patterns would become less saturated, since normalization removes second-order feature statistics which represents the style information to a large degree. Quantitative studies in Tab. 1 also support our analysis and come to the same conclusion.", + "bbox": [ + 496, + 583, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meta Training and Fast Adaptation. We provide training loss visualization in Fig. 8, to demonstrate the effectiveness of our meta training and fast adaptation algorithm. Our meta training stage learns an appropriate initial state, which enables fast decent and convergence of loss values in the fast adaptation stage. We show the same visualization in the per-style-per-model setting as a benchmark, using the same loss function and network architecture. As shown in Fig. 8,", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18335", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c18a13ed7d1b1d1a8d60849660f554dee74e4b644a5f92bebaf63eb8e21614ca.jpg", + "image_caption": [ + "Figure 8. Training loss visualization during meta training and fast adaptation of three transformer based models: StyTr2, StyFormer, and our Master. The loss curve under per-style-per-model setting by our model is used as a reference. Ours-wo-Fix means that all the parameters need to be updated during the fast adaptation stage." + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 424, + 261 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "it requires roughly 3k iterations for the model to be convergent in the per-style-per-model case (Single Style), while the meta model (Ours) after 9k iterations can be adapted for any style image in only a few shots with competitive training results. Thus, the total number of iterations for our method is significantly smaller than per-style-per-model ones with the growth number of required styles.", + "bbox": [ + 75, + 339, + 468, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By default, all the parameters except those in the style encoder module are fixed during the fast adaptation stage. We also experiment with training the whole model in this stage (Ours wo Fix in Fig. 8) and find that the training and convergence would become difficult and require more time and memory. Moreover, updating all the parameters may result in insufficient adaptation given a limited number of training steps. As demonstrated in Fig. 7, there is a gap on the global tone between the result and the style image.", + "bbox": [ + 75, + 446, + 468, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Base Model. We evaluate StyFormer and StyTr2 as base models, presenting loss visualizations. Heavier Transformers raise training difficulty for earlier models, resulting in inferior convergence. When the number of Transformer layers is 5, StyTr2 even fails to converge. By contrast, our model adopts parameter-sharing across all Transformer layers, which results in an overall light-weight structure. Consequently, it enjoys better training effects in the meta training stage as shown in Fig. 8. Moreover, as shown in the zoom-in part, Master also enables overall lower and more stable training in fast adaptation. Qualitative examples by different base models can be found in the supplement.", + "bbox": [ + 75, + 582, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Controllable Style Transfer. We compare with some widely-adopted approaches for the similar goal to control the degree of stylization, including (1) Recursion, which treats stylized results as input of content images in the next iteration, and (2) Linear Mixup, which conducts linear combinations between content features and stylized features before the decoder. Vanilla Transformer is adopted for these approaches. The major difference between Recursion and ours is that our method takes the iterative sty", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a3aef678cececb99a53609e418444464b1b2a588529b74f8e8784f6a6283a0a.jpg", + "image_caption": [ + "Figure 9. Comparisons with widely-adopted approaches to control the degree of stylization." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 870, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ization into consideration in the training time so that the model would learn to transfer style patterns layer-by-layer. Thus, as shown in Fig. 9, Recursion merely uses more intensive colors with the increasing of repeat times. Also, for Linear Mixup, it is typically hard to generate reasonable style transfer results especially when the weight of contents is high. Compared with them, our method is capable of rendering more style patterns and increasing the artistic abstraction when more Transformer layers are stacked.", + "bbox": [ + 496, + 325, + 890, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 474, + 617, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel Transformer model specifically for artistic style transfer, termed as Meta Style Transformer (Master). On one hand, parameters of different Transformer layers are shared, which reduces the total number of parameters significantly and thus enables easier convergence. It is also convenient for Master to control the degree of stylization via customizing the number of stacked layers in the inference time. On the other hand, different from standard Transformer, our model adopts dynamic and learnable scaling and shifting operations instead of original residual connections, which helps preserve similarity relationship in content structures while migrating remarkable style patterns. Specifically, Master is trained using a meta learning algorithm for few-shot style transfer. Only parameters of the Transformer encoder layer need to be updated in the few-shot stage, which benefits the fast and robust adaptation. Zero-shot arbitrary style transfer is a special case of the training configuration. Experiments suggest that Master outperforms previous state-of-the-art arbitrary style transfer methods on both content preserving and style rendering in the zero-shot case.", + "bbox": [ + 496, + 500, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 830, + 650, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project is supported by the Singapore Ministry of Education Academic Research Fund Tier 1 (WBS: A-0009440-01-00).", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18336", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2", + "[2] Yingying Deng, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, and Changsheng Xu. Arbitrary video style transfer via multi-channel correlation. arXiv preprint arXiv:2009.08003, 2020. 2, 5", + "[3] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2719-2727, 2020. 1, 2, 5", + "[4] Yingying Deng, Fan Tang, Xingjia Pan, Weiming Dong, Chongyang Ma, and Changsheng Xu. $stytr^2$ : Unbiased image style transfer with transformers, 2021. 1, 3, 5", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2020. 3", + "[6] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks, 2017. 3", + "[8] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3", + "[9] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1", + "[10] Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. Arbitrary style transfer with deep feature reshuffle. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8222–8231, 2018. 2", + "[11] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2", + "[13] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 1, 2, 3, 5, 6", + "[14] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 14861-14869, 2021. 2", + "[15] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4369-4376, 2020. 2", + "[16] Yongcheng Jing, Yining Mao, Yiding Yang, Yibing Zhan, Mingli Song, Xinchao Wang, and Dacheng Tao. Learning graph neural networks for image style transfer. In ECCV, 2022. 1", + "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 1, 5, 6", + "[18] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey, 2021. 2", + "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3", + "[21] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast arbitrary style transfer. arXiv preprint arXiv:1808.04537, 2018. 1, 2, 5", + "[22] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. arXiv preprint arXiv:1705.08086, 2017. 1, 2", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5", + "[24] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6649-6658, October 2021. 1, 2, 5", + "[25] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. In Conference on Neural Information Processing Systems, 2022. 3", + "[26] Songhua Liu, Hao Wu, Shoutong Luo, and Zhengxing Sun. Stable video style transfer based on partial convolution with depth-aware supervision. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2445-2453, 2020. 1", + "[27] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In European Conference on Computer Vision, 2022. 3", + "[28] Songhua Liu, Jingwen Ye, Runpeng Yu, and Xinchao Wang. Slimmable dataset condensation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18337", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3, 5", + "[30] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms, 2018. 3, 4", + "[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5880–5888, 2019. 1, 2, 5", + "[32] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 3", + "[33] Fred Phillips and Brandy Mackintosh. Wiki art gallery, inc.: A case for critical thinking. Issues in Accounting Education, 26(3):593-608, 2011. 5", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3", + "[36] Falong Shen, Shuicheng Yan, and Gang Zeng. Meta networks for neural style transfer, 2017. 3, 5", + "[37] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8242-8250, 2018. 2, 5", + "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 5", + "[39] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4", + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 1, 2", + "[41] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 2, 5", + "[42] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, 2022. 2", + "[43] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Conference on Neural Information Processing Systems, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1467-1475, 2019. 2", + "[45] Jingwen Ye, Songhua Liu, and Xinchao Wang. Partial network cloning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[46] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. 3", + "[47] Chi Zhang, Yixin Zhu, and Song-Chun Zhu. Metastyle: Three-way trade-off among speed, flexibility, and quality in neural style transfer, 2019. 3, 5" + ], + "bbox": [ + 501, + 92, + 890, + 289 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18338", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_model.json b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_model.json new file mode 100644 index 0000000000000000000000000000000000000000..67e2cb7b2f538ff2d9bc7ae21256659113a3814a --- /dev/null +++ b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_model.json @@ -0,0 +1,2618 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.862, + 0.176 + ], + "angle": 0, + "content": "Master: Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.203, + 0.739, + 0.24 + ], + "angle": 0, + "content": "Hao Tang\\(^{1*}\\) Songhua Liu\\(^{2*}\\) Tianwei Lin\\(^{3}\\) Shaoli Huang\\(^{4}\\) Fu Li\\(^{3}\\) Dongliang He\\(^{3}\\) Xinchao Wang\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.24, + 0.756, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Center for Data Science, Peking University \\(^{2}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.258, + 0.602, + 0.273 + ], + "angle": 0, + "content": "\\(^{3}\\)VIS, Baidu Inc. \\(^{4}\\)Tencent AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.278, + 0.697, + 0.292 + ], + "angle": 0, + "content": "tanghao@stu.pku.edu.cn songhua.liu@u.nus.edu" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.295, + 0.882, + 0.31 + ], + "angle": 0, + "content": "{lintianwei01,lifu,hedongliang01}@baidu.com shaolihuang@tencent.com xinchao@nus.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.318, + 0.214, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.392, + 0.154, + 0.402 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.318, + 0.361, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.392, + 0.294, + 0.402 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.321, + 0.449, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.392, + 0.44, + 0.404 + ], + "angle": 0, + "content": "Content/Style" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.322, + 0.538, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.392, + 0.514, + 0.403 + ], + "angle": 0, + "content": "Layer 1" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.322, + 0.627, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.392, + 0.604, + 0.403 + ], + "angle": 0, + "content": "Layer 3" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.322, + 0.715, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.392, + 0.691, + 0.403 + ], + "angle": 0, + "content": "Layer 5" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.322, + 0.803, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.392, + 0.784, + 0.402 + ], + "angle": 0, + "content": "Few Shot" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.322, + 0.892, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.389, + 0.887, + 0.41 + ], + "angle": 0, + "content": "\"The scream by Edvard Munch\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.41, + 0.893, + 0.452 + ], + "angle": 0, + "content": "Figure 1. Left: With more Transformer layers, the number of parameters and training difficulty increase significantly. Right: Our method supports test-time controlling of stylization degree via tuning the number of adopted Transformer layers. It is also readily adaptable to few-shot style transfer, where stylization with only 1 layer can be further improved. Text-guided few-shot style transfer is achievable." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.467, + 0.314, + 0.482 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.473, + 0.866 + ], + "angle": 0, + "content": "Transformer-based models achieve favorable performance in artistic style transfer recently thanks to its global receptive field and powerful multi-head/layer attention operations. Nevertheless, the over-paramerized multilayer structure increases parameters significantly and thus presents a heavy burden for training. Moreover, for the task of style transfer, vanilla Transformer that fuses content and style features by residual connections is prone to content-wise distortion. In this paper, we devise a novel Transformer model termed as Master specifically for style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters, which (1) reduces the total number of parameters, (2) leads to more robust training convergence, and (3) is readily to control the degree of stylization via tuning the number of stacked layers freely during inference. On the other hand, different from the vanilla version, we adopt a learnable scaling operation on content features before content-style feature interaction, which better preserves the original similarity between a pair of content features while ensuring the stylization quality. We also propose a novel meta learning scheme for the proposed model so that it can not only work in the typical setting of arbitrary style transfer, but also adaptable to the few-shot setting, by only fine-tuning the Transformer encoder layer in the few-shot stage for one" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.468, + 0.893, + 0.529 + ], + "angle": 0, + "content": "specific style. Text-guided few-shot style transfer is firstly achieved with the proposed framework. Extensive experiments demonstrate the superiority of Master under both zero-shot and few-shot style transfer settings." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.558, + 0.633, + 0.573 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Artistic style transfer aims at applying style patterns like colors and textures of a reference image to a given content image while preserving the semantic structure of the content. In contrast to the pioneering optimization method [9] and early per-style-per-model methods like [17, 26], arbitrary style transfer methods [3, 13, 16, 21, 22, 24, 31] enable real time style transfer for any style image in the test time in a zero-shot manner. The flexibility has led to this arbitrary-style-per-model fashion to dominate style transfer research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Recently, to enhance the representation of global information in arbitrary style transfer, Transformer [40] is introduced to this area [4], leveraging the global receptive field and powerful multi-head/layer structure, and achieves superior performance. Nevertheless, the over-parameterized multi-layer structure increases model parameters significantly. As shown in Fig. 1(a), there are 25.94M learnable parameters for a 3-layer Transformer structure in StyTr2 [4], v.s. 3.50M in AdaIN [13], a simple but effective baseline in arbitrary style transfer. Such a large model for standard Transformer inevitably presents a heavy burden for training. As shown in Fig. 1(b), when there are more than" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18329" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.47, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.15, + 0.195, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.237, + 0.155, + 0.248 + ], + "angle": 0, + "content": "Style /" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.249, + 0.158, + 0.259 + ], + "angle": 0, + "content": "Content" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.15, + 0.328, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.237, + 0.324, + 0.248 + ], + "angle": 0, + "content": "Vanilla Transformer" + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.249, + 0.332, + 0.259 + ], + "angle": 0, + "content": "w Residual Connection" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.15, + 0.463, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.237, + 0.45, + 0.248 + ], + "angle": 0, + "content": "Our Transformer" + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.249, + 0.46, + 0.26 + ], + "angle": 0, + "content": "w Learnable Scaling" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.271, + 0.47, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Residual connection in the vanilla Transformer tends to destroy the original similarity relationship on content structure in style transfer task. Our model is designed to address this problem with learnable scaling parameters. The top row shows a simple 2D visualization and the bottom one provides a qualitative example." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.398 + ], + "angle": 0, + "content": "4 layers, vanilla Transformer even fails to get convergent in training, which limits the scalability of the Transformer model in style transfer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.469, + 0.566 + ], + "angle": 0, + "content": "Moreover, vanilla Transformer relies on residual connections [12] to stylize content features, which suffers from the content-distortion problem. We illustrate this effect with a 2D visualization in Fig. 2(top), where residual connections lead the transformation results of two content feature vectors to move towards the dominated style features and thereby tend to eliminate their original distinction. The visual effect is that the stylized images would be dominated by some strong style features, such as salient edges, with the original self-(dis)similarity of content structures destroyed, as the example shown in Fig. 2(bottom)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Focusing on these drawbacks, in this paper, we are dedicated to devising a novel Transformer architecture specifically for artistic style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters and a random number of stacked layers are adopted for each training iteration. Compared with the original version, sharing parameters across different layers reduces the total number of parameters significantly and leads to more convenient training convergence. As a byproduct, it is also readily for our model to control the degree of stylization via tuning the number of stacked layers freely in the inference time, as shown in Fig. 1(right). On the other hand, we equip Transformer with learnable scale parameters for content-style interactions instead of residual connections, which alleviates content distortion to a large degree and better preserves content structures while rendering vivid style patterns simultaneously, as shown in the 2D visualization and the qualitative example in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Furthermore, beyond the typical zero-shot arbitrary style transfer, leveraging a meta learning scheme, our method is adaptable to the few-shot setting. By only fine-tuning the Transformer encoder layer in the few-shot stage, rapid" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "adaptation to the model for a specific style within a limited number of updates is possible, where the stylization with only 1 layer can be further improved, as shown in Fig. 1. Beyond that, we first achieve text-guided few-shot style transfer with this framework, which largely alleviates the training burden of previous per-text-per-model solution. In this sense, we term the overall pipeline Meta Style Transformer (Master). Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.241, + 0.892, + 0.316 + ], + "angle": 0, + "content": "- We propose a novel Transformer architecture specifically for artistic style transfer. It shares parameters between different layers, which not only helps training convergence, but also allows convenient control over the stylization effect." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.33, + 0.892, + 0.376 + ], + "angle": 0, + "content": "- We identify the content distortion problem of residual connections in Transformer and propose learnable scale parameters as an option to alleviate the problem." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.39, + 0.892, + 0.451 + ], + "angle": 0, + "content": "- We introduce a meta learning framework for adapting original training setting of zero-shot style transfer to the few-shot scenario, with which our Master achieves very good trade-off between flexibility and quality." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.465, + 0.892, + 0.555 + ], + "angle": 0, + "content": "- Experiments show that our model achieves results better than those of arbitrary-style-per-model methods. Furthermore, under the few-shot setting, either conditioned on image or text, Master can even yield performance on par with that of per-style-per-model methods with significantly less training cost." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.241, + 0.892, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.65, + 0.587 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Arbitrary-Style-Per-Model Methods. A lot of works achieve arbitrary style transfer via global feature transformation, e.g., WCT [22], AdaIN [13], Linear style transfer [21], DIN [15], MCCNet [2], and MAST [14]. In general, they can achieve the most attainable style transfer speed but dismiss stylized effects for local details a lot. In order to add more consideration of local details to global transformation based methods, there are patch-similarity based solution [1, 10, 37] and attention based method [3, 24, 31, 44]. While paying attention to local details, it is still hard to transfer complex style patterns and prone to unsatisfactory distortions due to the simple swap and fusion strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Transformer in Style Transfer. Transformer proposed in [40] is widely used in the natural language processing and has become a powerful baseline. Recently, Transformer receives better performance than CNN models in many vision tasks and are involved in various model zoos [6, 11, 18, 42, 43, 45]. Specifically, in the field of style transfer, [41] propose StyFormer, where the transformation is driven by the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18330" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.896, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.308, + 0.625, + 0.323 + ], + "angle": 0, + "content": "Figure 3. Overview of our model architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.47, + 0.436 + ], + "angle": 0, + "content": "cross-attention module in Transformer. [4] develop a Transformer model based on ViT [5]. However, the number of parameters and training difficulty would increase considerably with the increasing of Transformer layers. Also, following residual connections in original Transformer, it suffers from the distortion problem on content structures as shown in Fig. 2. Similar problems also exist in [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Meta Learning in Style Transfer. On this routine, [36] train a meta network to predict parameters of the generator model for one style reference image. [47] employ MAML algorithm [7] to find a style-free model for fast adaptation. Our method is different from theirs in three aspects: (1) our framework works for both few-shot and arbitrary style transfer thanks to the cross-attention in Transformer; (2) our meta learning algorithm is based on Reptile [30], a first-order meta learning algorithm without the necessity to operate gradients of higher levels, which is more efficient in training; and (3) only style encoder instead of the whole model needs to be updated during the few-shot learning stage, which makes it more convenient in practice." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.471, + 0.83 + ], + "angle": 0, + "content": "Text-Guided Synthesis. The emergence of CLIP model [34] bridges text and image domain, which supports a series of works on text-guided synthesis [8, 32, 35]. In the field of style transfer, CLIPstyler [20] proposes a patch-wise CLIP loss and achieves text-guided style transfer. Nevertheless, as a per-text-per-model solution, it is inconvenient in practice to handle a large number of text inputs. Despite the recent dataset distillation scheme may alleviate this issue, existing approaches, however, largely focus on images as input [25, 28, 46]. In this paper, we first achieve text-guided few-shot style transfer with the proposed meta learning algorithm, which improves the flexibility of CLIPstyler significantly." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.176, + 0.86 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we give details of the proposed Meta Style Transformer (Master) for zero-shot and few-shot style" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.331, + 0.893, + 0.378 + ], + "angle": 0, + "content": "transfer. We first introduce the network architecture of the proposed model, then illustrate how we train our model in a meta-learning fashion, and finally describe loss function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.387, + 0.706, + 0.402 + ], + "angle": 0, + "content": "3.1. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.893, + 0.593 + ], + "angle": 0, + "content": "The proposed model comprises an encoder, a feature modification module, and a decoder, as demonstrated in Fig. 3. We employ the first 2 stages of Swin Transformer [29] as the encoder Enc to extract common image features for both content and style images. The decoder Dec follows the setting of [13] with 3 upsampling convolutional blocks. For feature modification, we propose a Style Transformer module for transferring complex style patterns, which will be introduced later. Taking a style image \\( I_{s} \\) and a content image \\( I_{c} \\) as input, we first divide them into \\( 4 \\times 4 \\) patches and extract their corresponding feature maps \\( F_{c} \\) and \\( F_{s} \\) with the Swin encoder Enc:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.602, + 0.891, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {c} = \\operatorname {E n c} (I _ {c}), \\\\ F _ {s} = \\operatorname {E n c} \\left(I _ {s}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.65, + 0.893, + 0.681 + ], + "angle": 0, + "content": "where the spatial scales for \\( F_{c} \\) and \\( F_{s} \\) are 1/8 of those for \\( I_{c} \\) and \\( I_{s} \\). Then, the embedded feature \\( F_{cs} \\) is derived by:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.693, + 0.891, + 0.71 + ], + "angle": 0, + "content": "\\[\nF _ {c s} = \\operatorname {S t y l e T r a n s} \\left(F _ {s}, F _ {c}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.722, + 0.892, + 0.768 + ], + "angle": 0, + "content": "where StyleTrans denotes the Style Transformer module, and \\( F_{cs} \\) has the same shape as \\( F_{c} \\). Finally, we can synthesize the stylized image \\( I_{cs} \\) with the decoder as:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.781, + 0.891, + 0.797 + ], + "angle": 0, + "content": "\\[\nI _ {c s} = \\operatorname {D e c} \\left(F _ {c s}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Style Transformer. Similar to previous Transformer-based style transfer models like StyTr2 [4], the Transformer encoder is used to encode style information while the Transformer decoder takes charge of content-style interaction. In our model, there is only one copy of parameters shared by all Transformer encoder and decoder layers. Also, different" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18331" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "from standard Transformer where decoder layers would follow all encoder layers, encoder and decoder layers would be executed in an alternate fashion. The next layer would take the output of the current layer as input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.468, + 0.259 + ], + "angle": 0, + "content": "Our Transformer decoder layer is composed of self-attention, cross-attention, and non-linear blocks, for content encoding and content-style interaction. Specifically, content features are first processed with a self-attention step. Then, cross-attention is conducted by taking the content encoding as query and the style encoding as key and value, followed by a MLP for non-linear transformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Notably, in vanilla Transformer, features before and after the cross-attention step are fused by a residual connection, which is harmful for content structures as analyzed in Fig. 2. We thereby replace the residual connection with dynamic and learnable scaling and shifting steps, whose parameters are determined by the style encoder. In this way, the output of style encoder should consist of three parts: key for the following cross-attention \\( K_{s} \\), scaling parameters \\( V_{\\sigma} \\), and shifting parameters \\( V_{\\mu} \\). The prediction of each part shares a same self-attention map to save memory but uses independent non-linear transformation. The process in the Transformer encoder layer can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.457, + 0.388, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M H A} (Q, K, V) = [ h e a d _ {1}, \\dots , h e a d _ {h} ] W ^ {O},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.477, + 0.428, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\operatorname {h e a d} _ {i} = \\operatorname {A t t} \\left(Q W ^ {Q _ {i}}, K W ^ {K _ {i}}, V W ^ {V _ {i}}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.498, + 0.362, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t} (Q, K, V) = \\operatorname {s o f t m a x} (\\frac {Q K ^ {\\top}}{\\sqrt {d _ {k}}} V),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.535, + 0.469, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} K _ {s} ^ {\\prime} = K _ {s} + \\operatorname {M H A} \\left(K _ {s}, K _ {s}, K _ {s}\\right), \\\\ \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i l t e} \\end{array} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.554, + 0.346, + 0.57 + ], + "angle": 0, + "content": "\\[\nK _ {s} = K _ {s} ^ {\\prime} + \\operatorname {M L P} \\left(K _ {s} ^ {\\prime}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.572, + 0.4, + 0.589 + ], + "angle": 0, + "content": "\\[\nV _ {\\sigma} ^ {\\prime} = V _ {\\sigma} + \\operatorname {M H A} \\left(K _ {s}, K _ {s}, V _ {\\sigma}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.591, + 0.341, + 0.608 + ], + "angle": 0, + "content": "\\[\nV _ {\\sigma} = V _ {\\sigma} ^ {\\prime} + \\operatorname {M L P} \\left(V _ {\\sigma} ^ {\\prime}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.61, + 0.4, + 0.628 + ], + "angle": 0, + "content": "\\[\nV _ {\\mu} ^ {\\prime} = V _ {\\mu} + \\mathrm {M H A} \\left(K _ {s}, K _ {s}, V _ {\\mu}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.63, + 0.341, + 0.648 + ], + "angle": 0, + "content": "\\[\nV _ {\\mu} = V _ {\\mu} ^ {\\prime} + \\operatorname {M L P} \\left(V _ {\\mu} ^ {\\prime}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.469, + 0.722 + ], + "angle": 0, + "content": "where \\( K_{s}, V_{\\sigma} \\), and \\( V_{\\mu} \\) are initialized as the style feature \\( F_{s} \\) before the first Transformer encoder layer. We do not incorporate normalization to encode style features since second-order statistics can largely represent style information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.469, + 0.798 + ], + "angle": 0, + "content": "Then, in the cross-attention of Transformer decoder, scaling and shifting parameters for each content feature point is aggregated from \\( V_{\\sigma} \\) and \\( V_{\\mu} \\) respectively according to the cross-attention map. The process in one Transformer decoder block can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.811, + 0.384, + 0.829 + ], + "angle": 0, + "content": "\\[\nF _ {c s} ^ {\\prime} = F _ {c s} + \\mathrm {M H A} \\left(F _ {c s}, F _ {c s}, F _ {c s}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.83, + 0.395, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\sigma = \\operatorname {M H A} \\left(\\operatorname {I N} \\left(F _ {c s} ^ {\\prime}\\right), \\operatorname {I N} \\left(K _ {s}\\right), V _ {\\sigma}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.85, + 0.469, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\mu = \\mathrm {M H A} (\\mathrm {I N} \\left(F _ {c s} ^ {\\prime}\\right), \\mathrm {I N} \\left(K _ {s}\\right), V _ {\\mu}), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.868, + 0.288, + 0.885 + ], + "angle": 0, + "content": "\\[\nF _ {c s} ^ {\\prime \\prime} = \\sigma \\odot F _ {c s} ^ {\\prime} + \\mu ,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.887, + 0.32, + 0.904 + ], + "angle": 0, + "content": "\\[\nF _ {c s} = F _ {c s} ^ {\\prime \\prime} + \\operatorname {M L P} \\left(F _ {c s} ^ {\\prime \\prime}\\right),\n\\]" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.688, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Meta Training" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.11, + 0.892, + 0.156 + ], + "angle": 0, + "content": "Require: \\(\\mathcal{D}_c\\): content dataset; \\(\\mathcal{D}_s\\): style dataset; \\(\\delta\\): inner learning rate; \\(\\eta\\): outer learning rate; \\(k\\): number of inner updates; \\(T\\): maximal number of stacked layers;" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.156, + 0.803, + 0.171 + ], + "angle": 0, + "content": "Ensure: trained meta generator parameters \\(\\theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.172, + 0.67, + 0.187 + ], + "angle": 0, + "content": "1: initialize \\(\\theta\\) randomly" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.187, + 0.708, + 0.201 + ], + "angle": 0, + "content": "2: for iteration 1, 2, 3, ... do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.202, + 0.77, + 0.217 + ], + "angle": 0, + "content": "3: sample a style image \\(I_{s}\\) from \\(\\mathcal{D}_s\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.218, + 0.602, + 0.23 + ], + "angle": 0, + "content": "4: \\(\\omega \\gets \\theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.232, + 0.652, + 0.246 + ], + "angle": 0, + "content": "5: for \\(k\\) times do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.247, + 0.869, + 0.262 + ], + "angle": 0, + "content": "6: sample a batch of content image \\( I_{c} \\) from \\( \\mathcal{D}_c \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.263, + 0.848, + 0.277 + ], + "angle": 0, + "content": "7: sample the number of layers from \\(1\\sim T\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.278, + 0.807, + 0.292 + ], + "angle": 0, + "content": "8: forward propagation using Eq. 1-5" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.293, + 0.799, + 0.307 + ], + "angle": 0, + "content": "9: compute inner loss \\(L\\) using Eq. 8" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.308, + 0.681, + 0.321 + ], + "angle": 0, + "content": "10: \\(\\omega \\gets \\omega -\\delta \\nabla L\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.323, + 0.608, + 0.336 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.337, + 0.678, + 0.353 + ], + "angle": 0, + "content": "12: \\(\\theta \\gets \\theta +\\eta (\\omega -\\theta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.354, + 0.584, + 0.366 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.172, + 0.869, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.892, + 0.443 + ], + "angle": 0, + "content": "where IN denotes instance normalization [39] and \\(\\odot\\) represents element-wise multiplication. \\(F_{cs}\\) is initialized as \\(F_{c}\\) before the first Transformer decoder layer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.454, + 0.669, + 0.47 + ], + "angle": 0, + "content": "3.2. Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.584 + ], + "angle": 0, + "content": "To achieve high-quality style transfer, we introduce a two-stage training strategy that comprises meta training and fast adaptation. The meta training stage is designed to learn a generic model initialization, while the fast adaptation adapts the network for a single style in a few iterations for few-shot style transfer. Note that zero-shot style transfer is a special case of the overall training configuration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Meta Training: Inspired by Reptile [30], a first-order meta learning algorithm, rendering style patterns of a specific reference style image can be viewed as a task. We seek an optimal initialization for neural networks in this stage, so that the networks can be rapidly adapted for a new task in only a few shots. The main training procedure for this stage is shown in Algorithm 1. In each iteration, we sample 1 style and \\( k \\) batches of contents to perform inner optimization to obtain \"fast weights\" \\( \\omega \\), which would later guide the update of \"slow weights\" \\( \\theta \\), to move a step in the direction of \\( \\omega - \\theta \\). Notably, as there is only one group of parameters for Transformer encoder and decoder layers, we randomly choose a number as the number of stacked layers for the Style Transformer in each iteration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fast Adaptation: The trained model after the first stage would serve as an initialization and will be adapted for a single style. With the same objective, this stage behaves almost identically to the internal loop in Algorithm 1. The only difference is that only parameters of the Transformer encoder layer are necessary to be updated, since (1) the Transformer encoder layer is the main component to extract style pat" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18332" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "terns and has the most significant impact on stylization, and (2) it would save memory and speed up the adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.257 + ], + "angle": 0, + "content": "Zero-Shot Style Transfer: \"Zero-Shot\" means that, once the meta training is done, there is no fast adaptation stage needed and the meta model itself can support arbitrary style transfer. To encourage the model to produce satisfactory results in the zero-shot setup, we set the inner optimization time \\( k \\) to 1, where the algorithm is reduced to the typical training paradigm of existing arbitrary style transfer methods. In this sense, Algorithm 1 provides a more general setting for style transfer in both zero-shot and few-shot cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.468, + 0.438 + ], + "angle": 0, + "content": "Text-Guided Style Transfer. We perform text-guided style transfer based on our image style transfer model with slight modifications. Following the common practice, a pretrained CLIP encoder is imported to extract text features. To be consistent with image input, we use a StyleGAN-like mapping network to convert text features into pseudo image features. In the meta training stage, since CLIP unifies the feature space, we use image instead of text as style input to avoid additional text dataset. In the fast adaptation stage, the mapping network along with the Transformer encoder is updated, considering there is still a gap between CLIP image feature and text feature." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.222, + 0.462 + ], + "angle": 0, + "content": "3.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.468, + 0.561 + ], + "angle": 0, + "content": "The training objective in both meta-training and fast adaptation stages follows many works in arbitrary style transfer, which consists of content loss and style loss. Let \\( F^x \\) represent features on ReLU-\\( x \\)-1 layer of a pre-trained VGG19 network [38] for loss computation. The content loss is defined by the normalized perceptual loss [17]:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.562, + 0.468, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t}} = \\sum_ {x = 2} ^ {5} \\| \\mathrm {I N} \\left(F _ {c} ^ {x}\\right) - \\mathrm {I N} \\left(F _ {c s} ^ {x}\\right) \\| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.61, + 0.445, + 0.625 + ], + "angle": 0, + "content": "while the style loss adopts the mean-variance loss [13]:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.628, + 0.468, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s t y} = \\sum_ {x = 2} ^ {5} \\left(\\| \\mu \\left(F _ {s} ^ {x}\\right) - \\mu \\left(F _ {c s} ^ {x}\\right) \\| _ {2} + \\| \\sigma \\left(F _ {s} ^ {x}\\right) - \\sigma \\left(F _ {c s} ^ {x}\\right) \\| _ {2}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.468, + 0.725 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\sigma\\) calculate the channel-wise mean and standard deviation separately. The overall objective is given by the weighted summation of the two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.731, + 0.468, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t}} + \\lambda \\mathcal {L} _ {\\text {s t y}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.4, + 0.763 + ], + "angle": 0, + "content": "where \\(\\lambda\\) controls the balance between two terms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.809 + ], + "angle": 0, + "content": "In text-guided style transfer, the loss functions are the same as CLIPstyler, including global CLIP loss, directional CLIP loss and PatchCLIP loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.208, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.293, + 0.863 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "We use MS-COCO [23] as our content dataset and WikiArt test set [33] as our style dataset. The content" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.869, + 0.377 + ], + "angle": 0, + "content": "
MethodLcontLsimLsty
AdaIN4.88±0.700.53±0.231.51±0.78
Linear3.93±0.760.44±0.171.97±0.96
AvatarNet5.76±0.530.56±0.193.19±1.82
SANet4.72±0.610.50±0.171.15±0.58
MANet4.93±0.570.50±0.171.41±0.77
MCCNet4.22±0.690.47±0.171.56±0.84
AdaAttN4.46±0.700.43±0.162.20±1.19
StyTr23.78±0.990.48±0.221.50±0.69
StyFormer4.94±0.790.43±0.152.20±1.44
MetaNet3.48±0.850.45±0.192.47±1.69
MetaStyle*3.64±1.120.42±0.182.47±1.06
Johnson†4.60±0.760.59±0.201.02±0.34
Ours-Vanilla5.50±0.630.59±0.260.85±0.38
Ours-Norm4.70±0.750.43±0.130.93±0.33
Ours-ZS-L14.13±0.680.41±0.140.92±0.40
Ours-ZS-L34.20±0.680.41±0.130.81±0.31
Ours-FS*4.24±0.820.38±0.120.79±0.25
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons. ZS and FS for our model denote zero-shot and few-shot modes. Vanilla denote replacing our architecture with original Transformer. Norm means adding layer normalization in the Transformer encoder layer. L1/L3 means using 1/3 Transformer layers in the test time. * and † denote few-shot and per-style-per-model methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.713 + ], + "angle": 0, + "content": "dataset contains roughly 80,000 images and the style dataset has about 20,000 images. The optimizer is Adam [19] with learning rates of both inner and outer loops set as 0.0001 and the batch size is 4. In training, we first resize the content and style image to \\(512 \\times 512\\) and then randomly crop to \\(256 \\times 256\\) resolution. During inference, our model can handle inputs of any size. The update times for inner optimization \\(k\\) is set as 2 for few-shot case and 1 for zero-shot case. In training, the maximal number of stacked layers is 4. All the multi-head attention blocks are instantiated as shifted window attention in [29], with window size 8 and shift size 4. The model is trained on a Nvidia 3090 with 9k iterations in the meta training stage for convergence while only 100 steps for image input and 20 steps for text input in the fast adaptation for few-shot style transfer, which takes less than 1 minute. Hyper-parameter \\(\\lambda\\) is set as 10." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.77, + 0.741 + ], + "angle": 0, + "content": "4.2. Comparison with Prior Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we compare results by our Master with 13 state-of-the-art style transfer methods, including 3 global transformation based methods (AdaIN [13], Linear style transfer [21], and MCCNet [2]), 1 patch swap based method (Avatar-Net [37]), 3 attention based methods (SANet [31], MANet [3], and AdaAttN [24]), 2 transformer based methods (StyTr2 [4] and StyFromer [41]), 2 meta learning based methods (MetaNet [36] and MetaStyle [47]), 1 per-style-per-model method by Johnson et al. [17] and 1 text-guided style transfer method (CLIPstyler)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18333" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "Quantitative Comparison. We adopt content loss \\(\\mathcal{L}_{cont}\\) in Eq. 6 and style loss \\(\\mathcal{L}_{sty}\\) in Eq. 7 as evaluation metrics to reflect effects of learning by different methods. We also design a metric \\(\\mathcal{L}_{sim}\\) to reflect the preservation of the spatial-wise self cosine similarity of content structures:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.168, + 0.348, + 0.204 + ], + "angle": 0, + "content": "\\[\nD _ {*, i j} ^ {x} = 1 - \\frac {F _ {*, i} ^ {x} \\cdot F _ {*, j} ^ {x}}{\\| F _ {*, i} ^ {x} \\| \\| F _ {*, j} ^ {x} \\|}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.202, + 0.468, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s i m} = \\sum_ {x = 3} ^ {4} \\frac {1}{n _ {x} ^ {2}} \\sum_ {i, j} \\left| \\frac {D _ {c , i j} ^ {x}}{\\sum_ {k} D _ {c , k j} ^ {x}} - \\frac {D _ {c s , i j} ^ {x}}{\\sum_ {k} D _ {c s , k j} ^ {x}} \\right|, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.251, + 0.468, + 0.492 + ], + "angle": 0, + "content": "where \\( n_x \\) is the number of spatial locations for the current feature map and the second foot script denotes the spatial index. Smaller \\( \\mathcal{L}_{sim} \\) means that the original spatial-wise relationship is better preserved during style transfer, i.e., less content distortion. We use the test dataset in the code page of [13] for evaluation, with 11 content images and 20 style images to form 220 content-style pairs. We report the mean and standard deviation over the 220 cases in Tab. 1. Notably, the lowest style loss and content similarity loss are achieved by Master simultaneously compared with previous methods, even better than the per-style-per-model solution by Johnson et al. [17], and can be further reduced in the few-shot case with comparable content loss values, which demonstrates the joint advantages of our method in content preserving and style rendering. Meanwhile, the lower standard deviations demonstrate the robustness of our model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.493, + 0.468, + 0.538 + ], + "angle": 0, + "content": "Qualitative Comparison. Qualitative examples by different methods are shown in Fig. 4. Full comparisons with more methods can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.468, + 0.613 + ], + "angle": 0, + "content": "We first give discussion on the global-transformation based methods. As shown in the 3rd and 4th rows, in AdaIN and Linear, features in all positions share the same transformation matrix, which fails to migrate detail textures in style images and has poor color saturation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.614, + 0.468, + 0.734 + ], + "angle": 0, + "content": "As for the attention-based methods, SANet in the 5th row introduces local focus to style images. However, due to the simple fusion of one-layer attention results and original content features, it brings textures and content distortion in many cases, e.g., background areas of the 4th and 5th columns. AdaAttN in the 6th row dedicates to addressing the distortion problem in SANet, but it sacrifices the ability to render salient style patterns and textures." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.468, + 0.9 + ], + "angle": 0, + "content": "We then compare our results with those by Transformer-based methods. StyFormer in the 7th row adopts a design of the cross-attention module in Transformer without exploring the effect of self-attention to enhance style rendering. Thus, the results appear under-stylized compared with those by our method. StyTr2 in the 8th row is a Transformer architecture with a pure self-attention mechanism, which makes the performance on the extraction and migration of local textures not satisfactory enough, e.g., strokes of the 2nd column and waves of the 4th column. Moreover, since it follows the design of residual connections in" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.888, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.867, + 0.892, + 0.894 + ], + "angle": 0, + "content": "Figure 4. Comparison with previous state-of-the-art style transfer methods. Zoom in for better details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18334" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.094, + 0.47, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.47, + 0.448 + ], + "angle": 0, + "content": "Figure 5. Comparison with CLIPstyler in text-guided style transfer. Zoom in for better details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.576 + ], + "angle": 0, + "content": "vanilla Transformer, the issue of content distortion in Fig. 2 is still evident, as shown in the background areas of the 4th and 5th columns. Compared with these methods, Master renders more vivid global and local style patterns. Meanwhile, it well maintains content structures simultaneously with learnable scaling and shifting parameters. Furthermore, in the few-shot case, comparable performance can be achieved with only 1 Transformer layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.731 + ], + "angle": 0, + "content": "MetaNet and MetaStyle also involve the concept of meta learning into their style transfer frameworks. On the one hand, MetaNet predicts parameters of a generator network given one style image via a meta network, which is an ambitious goal and makes the training more difficult. Thus, as shown in the 9th row, the stylized effects of their results still appear weak. On the other hand, MetaStyle achieves inferior performance in terms of local details as demonstrated in the 10th row, since only parameters of global normalization layers are learnable in the fast adaptation stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.469, + 0.839 + ], + "angle": 0, + "content": "The single style transfer method by Johnson et al. in the 11th row tends to arrange patterns learned by the network arbitrarily, which may produce undesired and distorted effects. By contrast, the complex content-style dependence is constructed in our Transformer model and thus achieves more remarkable performance on transferring style patterns to proper positions in content images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Some qualitative comparisons on text-guided style transfer with CLIPstyler are shown in Fig. 5, which uses per-text-per-model fashion. Through these results, we observe that despite significantly more training steps, CLIPstyler" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.1, + 0.849, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.205, + 0.892, + 0.247 + ], + "angle": 0, + "content": "Figure 6. Results of user study. Left: Our zero-shot stylization results compared with SOTA methods. Right: Our few-shot stylization comparison results." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.25, + 0.892, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.321, + 0.867, + 0.335 + ], + "angle": 0, + "content": "Figure 7. Ablation study on key designs in Master model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.401 + ], + "angle": 0, + "content": "still suffers from weak stylization. By contrast, our method generates more vivid results in only 20 adaptation steps, which demonstrates the advantages of the proposed Master architecture and the versatility of our training pipeline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.552 + ], + "angle": 0, + "content": "User Study. Following most style transfer works, we conduct user study and report user preference. We choose 20 content images and 20 style images to form 400 content-style pairs. We involve 100 people and randomly assign 20 stylized results from compared methods to each subject. Our method showcases zero-shot results with 3 Transformer layers in 10 tests and few-shot results in the remaining 10. For each pair, the order of results is randomized, and participants choose their favorite. With 2,000 votes in total, Fig. 6 demonstrates Master's superior style transfer quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.654, + 0.577 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Architecture. We conduct ablation studies on key designs in Master to illustrate their effects on the stylization quality. As shown in Fig. 7, on the one hand, if vanilla Transformer model is used, without learnable scaling and shifting parameters, noisy textures can be introduced significantly, which distorts the original content structures and affects style transfer quality. On the other hand, if layer normalization operations in the standard Transformer are used in the style encoder, style patterns would become less saturated, since normalization removes second-order feature statistics which represents the style information to a large degree. Quantitative studies in Tab. 1 also support our analysis and come to the same conclusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Meta Training and Fast Adaptation. We provide training loss visualization in Fig. 8, to demonstrate the effectiveness of our meta training and fast adaptation algorithm. Our meta training stage learns an appropriate initial state, which enables fast decent and convergence of loss values in the fast adaptation stage. We show the same visualization in the per-style-per-model setting as a benchmark, using the same loss function and network architecture. As shown in Fig. 8," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18335" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.426, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.334 + ], + "angle": 0, + "content": "Figure 8. Training loss visualization during meta training and fast adaptation of three transformer based models: StyTr2, StyFormer, and our Master. The loss curve under per-style-per-model setting by our model is used as a reference. Ours-wo-Fix means that all the parameters need to be updated during the fast adaptation stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.469, + 0.446 + ], + "angle": 0, + "content": "it requires roughly 3k iterations for the model to be convergent in the per-style-per-model case (Single Style), while the meta model (Ours) after 9k iterations can be adapted for any style image in only a few shots with competitive training results. Thus, the total number of iterations for our method is significantly smaller than per-style-per-model ones with the growth number of required styles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.469, + 0.582 + ], + "angle": 0, + "content": "By default, all the parameters except those in the style encoder module are fixed during the fast adaptation stage. We also experiment with training the whole model in this stage (Ours wo Fix in Fig. 8) and find that the training and convergence would become difficult and require more time and memory. Moreover, updating all the parameters may result in insufficient adaptation given a limited number of training steps. As demonstrated in Fig. 7, there is a gap on the global tone between the result and the style image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Base Model. We evaluate StyFormer and StyTr2 as base models, presenting loss visualizations. Heavier Transformers raise training difficulty for earlier models, resulting in inferior convergence. When the number of Transformer layers is 5, StyTr2 even fails to converge. By contrast, our model adopts parameter-sharing across all Transformer layers, which results in an overall light-weight structure. Consequently, it enjoys better training effects in the meta training stage as shown in Fig. 8. Moreover, as shown in the zoom-in part, Master also enables overall lower and more stable training in fast adaptation. Qualitative examples by different base models can be found in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Controllable Style Transfer. We compare with some widely-adopted approaches for the similar goal to control the degree of stylization, including (1) Recursion, which treats stylized results as input of content images in the next iteration, and (2) Linear Mixup, which conducts linear combinations between content features and stylized features before the decoder. Vanilla Transformer is adopted for these approaches. The major difference between Recursion and ours is that our method takes the iterative sty" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.089, + 0.872, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.295, + 0.891, + 0.322 + ], + "angle": 0, + "content": "Figure 9. Comparisons with widely-adopted approaches to control the degree of stylization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.463 + ], + "angle": 0, + "content": "ization into consideration in the training time so that the model would learn to transfer style patterns layer-by-layer. Thus, as shown in Fig. 9, Recursion merely uses more intensive colors with the increasing of repeat times. Also, for Linear Mixup, it is typically hard to generate reasonable style transfer results especially when the weight of contents is high. Compared with them, our method is capable of rendering more style patterns and increasing the artistic abstraction when more Transformer layers are stacked." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.475, + 0.618, + 0.49 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.817 + ], + "angle": 0, + "content": "In this paper, we propose a novel Transformer model specifically for artistic style transfer, termed as Meta Style Transformer (Master). On one hand, parameters of different Transformer layers are shared, which reduces the total number of parameters significantly and thus enables easier convergence. It is also convenient for Master to control the degree of stylization via customizing the number of stacked layers in the inference time. On the other hand, different from standard Transformer, our model adopts dynamic and learnable scaling and shifting operations instead of original residual connections, which helps preserve similarity relationship in content structures while migrating remarkable style patterns. Specifically, Master is trained using a meta learning algorithm for few-shot style transfer. Only parameters of the Transformer encoder layer need to be updated in the few-shot stage, which benefits the fast and robust adaptation. Zero-shot arbitrary style transfer is a special case of the training configuration. Experiments suggest that Master outperforms previous state-of-the-art arbitrary style transfer methods on both content preserving and style rendering in the zero-shot case." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.651, + 0.847 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.891, + 0.9 + ], + "angle": 0, + "content": "This project is supported by the Singapore Ministry of Education Academic Research Fund Tier 1 (WBS: A-0009440-01-00)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18336" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Yingying Deng, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, and Changsheng Xu. Arbitrary video style transfer via multi-channel correlation. arXiv preprint arXiv:2009.08003, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2719-2727, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[4] Yingying Deng, Fan Tang, Xingjia Pan, Weiming Dong, Chongyang Ma, and Changsheng Xu. \\(stytr^2\\): Unbiased image style transfer with transformers, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.408 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[6] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.471, + 0.676 + ], + "angle": 0, + "content": "[10] Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. Arbitrary style transfer with deep feature reshuffle. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8222–8231, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[11] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.746, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.804, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[13] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.86, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 14861-14869, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[15] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4369-4376, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Yongcheng Jing, Yining Mao, Yiding Yang, Yibing Zhan, Mingli Song, Xinchao Wang, and Dacheng Tao. Learning graph neural networks for image style transfer. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[18] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[21] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast arbitrary style transfer. arXiv preprint arXiv:1808.04537, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[22] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. arXiv preprint arXiv:1705.08086, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[24] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6649-6658, October 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[25] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. In Conference on Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[26] Songhua Liu, Hao Wu, Shoutong Luo, and Zhengxing Sun. Stable video style transfer based on partial convolution with depth-aware supervision. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2445-2453, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[27] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[28] Songhua Liu, Jingwen Ye, Runpeng Yu, and Xinchao Wang. Slimmable dataset condensation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18337" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[30] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5880–5888, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[32] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[33] Fred Phillips and Brandy Mackintosh. Wiki art gallery, inc.: A case for critical thinking. Issues in Accounting Education, 26(3):593-608, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[36] Falong Shen, Shuicheng Yan, and Gang Zeng. Meta networks for neural style transfer, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[37] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8242-8250, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[39] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[41] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[42] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[43] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Conference on Neural Information Processing Systems, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[44] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1467-1475, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[45] Jingwen Ye, Songhua Liu, and Xinchao Wang. Partial network cloning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[46] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[47] Chi Zhang, Yixin Zhu, and Song-Chun Zhu. Metastyle: Three-way trade-off among speed, flexibility, and quality in neural style transfer, 2019. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18338" + } + ] +] \ No newline at end of file diff --git a/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_origin.pdf b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..619989ccbfb0f1b7330e3543edefac524929d3a0 --- /dev/null +++ b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/936c6f6e-56a4-479a-88a9-b546f91bc886_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20b313dcd0d5f27a8c829dcf7fad0a00e23ebd643c6b51bd8052eecbe4cbf662 +size 6146949 diff --git a/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/full.md b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..069271a7d84a55f831fccff14e6aa72b50dcf401 --- /dev/null +++ b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/full.md @@ -0,0 +1,391 @@ +# Master: Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer + +Hao Tang $^{1*}$ Songhua Liu $^{2*}$ Tianwei Lin $^{3}$ Shaoli Huang $^{4}$ Fu Li $^{3}$ Dongliang He $^{3}$ Xinchao Wang $^{2\dagger}$ + +$^{1}$ Center for Data Science, Peking University $^{2}$ National University of Singapore + +$^{3}$ VIS, Baidu Inc. $^{4}$ Tencent AI Lab + +tanghao@stu.pku.edu.cn songhua.liu@u.nus.edu + +{lintianwei01,lifu,hedongliang01}@baidu.com shaolihuang@tencent.com xinchao@nus.edu.sg + +![](images/2265e982334aa3d38d302ca2cd0acb11c8f855a706e3c1f58c5e686b5be6a7d3.jpg) +(a) + +![](images/b7116b00247e4e9aa669574d1c349a4f34f15acb837a839f8ca2d4d8ea67e370.jpg) +(b) +Figure 1. Left: With more Transformer layers, the number of parameters and training difficulty increase significantly. Right: Our method supports test-time controlling of stylization degree via tuning the number of adopted Transformer layers. It is also readily adaptable to few-shot style transfer, where stylization with only 1 layer can be further improved. Text-guided few-shot style transfer is achievable. + +![](images/4729d7188c270f0708fa04f15441404551f99646e6687f632d04da8c3946734a.jpg) +Content/Style + +![](images/1a002410cbc987d143e156f13f952eac456cc3a0c7b3485606bd80dc4d775d76.jpg) +Layer 1 + +![](images/c0397542bebbcff4821cc5b411605abc84da086049ea6fbee22ed950c4a0291e.jpg) +Layer 3 + +![](images/40324f219537218ca4ed4779076be71593839fa09484551872352291beca21b5.jpg) +Layer 5 + +![](images/35755527d19bfc364ca5ffc42f6860c3ab7720adba3e4c155eb979032abf5e41.jpg) +Few Shot + +![](images/da914c77b9f2de5ee7f43e92906edb89a2753d0ba9afa1d21147b3b26e2559cb.jpg) +"The scream by Edvard Munch" + +# Abstract + +Transformer-based models achieve favorable performance in artistic style transfer recently thanks to its global receptive field and powerful multi-head/layer attention operations. Nevertheless, the over-paramerized multilayer structure increases parameters significantly and thus presents a heavy burden for training. Moreover, for the task of style transfer, vanilla Transformer that fuses content and style features by residual connections is prone to content-wise distortion. In this paper, we devise a novel Transformer model termed as Master specifically for style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters, which (1) reduces the total number of parameters, (2) leads to more robust training convergence, and (3) is readily to control the degree of stylization via tuning the number of stacked layers freely during inference. On the other hand, different from the vanilla version, we adopt a learnable scaling operation on content features before content-style feature interaction, which better preserves the original similarity between a pair of content features while ensuring the stylization quality. We also propose a novel meta learning scheme for the proposed model so that it can not only work in the typical setting of arbitrary style transfer, but also adaptable to the few-shot setting, by only fine-tuning the Transformer encoder layer in the few-shot stage for one + +specific style. Text-guided few-shot style transfer is firstly achieved with the proposed framework. Extensive experiments demonstrate the superiority of Master under both zero-shot and few-shot style transfer settings. + +# 1. Introduction + +Artistic style transfer aims at applying style patterns like colors and textures of a reference image to a given content image while preserving the semantic structure of the content. In contrast to the pioneering optimization method [9] and early per-style-per-model methods like [17, 26], arbitrary style transfer methods [3, 13, 16, 21, 22, 24, 31] enable real time style transfer for any style image in the test time in a zero-shot manner. The flexibility has led to this arbitrary-style-per-model fashion to dominate style transfer research. + +Recently, to enhance the representation of global information in arbitrary style transfer, Transformer [40] is introduced to this area [4], leveraging the global receptive field and powerful multi-head/layer structure, and achieves superior performance. Nevertheless, the over-parameterized multi-layer structure increases model parameters significantly. As shown in Fig. 1(a), there are 25.94M learnable parameters for a 3-layer Transformer structure in StyTr2 [4], v.s. 3.50M in AdaIN [13], a simple but effective baseline in arbitrary style transfer. Such a large model for standard Transformer inevitably presents a heavy burden for training. As shown in Fig. 1(b), when there are more than + +![](images/acf3d3dbaedee70ad481c29fb68b35357f3ce3b3eac58ecc396221abd620540a.jpg) + +![](images/c7782a5691269e99e26b011c6c4e1f635c1de7b3e3bfc1a23aafda69325db50a.jpg) +Style / +Content +Figure 2. Residual connection in the vanilla Transformer tends to destroy the original similarity relationship on content structure in style transfer task. Our model is designed to address this problem with learnable scaling parameters. The top row shows a simple 2D visualization and the bottom one provides a qualitative example. + +![](images/334709ab8081e5446453ce540b589cd652765d759204889f2c21e15448f6b559.jpg) +Vanilla Transformer +w Residual Connection + +![](images/07eaa7e3cd6c4e7e2d6df88b917127ce9d1f0ece8a6651896c3c817781428a86.jpg) +Our Transformer +w Learnable Scaling + +4 layers, vanilla Transformer even fails to get convergent in training, which limits the scalability of the Transformer model in style transfer. + +Moreover, vanilla Transformer relies on residual connections [12] to stylize content features, which suffers from the content-distortion problem. We illustrate this effect with a 2D visualization in Fig. 2(top), where residual connections lead the transformation results of two content feature vectors to move towards the dominated style features and thereby tend to eliminate their original distinction. The visual effect is that the stylized images would be dominated by some strong style features, such as salient edges, with the original self-(dis)similarity of content structures destroyed, as the example shown in Fig. 2(bottom). + +Focusing on these drawbacks, in this paper, we are dedicated to devising a novel Transformer architecture specifically for artistic style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters and a random number of stacked layers are adopted for each training iteration. Compared with the original version, sharing parameters across different layers reduces the total number of parameters significantly and leads to more convenient training convergence. As a byproduct, it is also readily for our model to control the degree of stylization via tuning the number of stacked layers freely in the inference time, as shown in Fig. 1(right). On the other hand, we equip Transformer with learnable scale parameters for content-style interactions instead of residual connections, which alleviates content distortion to a large degree and better preserves content structures while rendering vivid style patterns simultaneously, as shown in the 2D visualization and the qualitative example in Fig. 2. + +Furthermore, beyond the typical zero-shot arbitrary style transfer, leveraging a meta learning scheme, our method is adaptable to the few-shot setting. By only fine-tuning the Transformer encoder layer in the few-shot stage, rapid + +adaptation to the model for a specific style within a limited number of updates is possible, where the stylization with only 1 layer can be further improved, as shown in Fig. 1. Beyond that, we first achieve text-guided few-shot style transfer with this framework, which largely alleviates the training burden of previous per-text-per-model solution. In this sense, we term the overall pipeline Meta Style Transformer (Master). Our contributions are summarized as follows: + +- We propose a novel Transformer architecture specifically for artistic style transfer. It shares parameters between different layers, which not only helps training convergence, but also allows convenient control over the stylization effect. +- We identify the content distortion problem of residual connections in Transformer and propose learnable scale parameters as an option to alleviate the problem. +- We introduce a meta learning framework for adapting original training setting of zero-shot style transfer to the few-shot scenario, with which our Master achieves very good trade-off between flexibility and quality. +- Experiments show that our model achieves results better than those of arbitrary-style-per-model methods. Furthermore, under the few-shot setting, either conditioned on image or text, Master can even yield performance on par with that of per-style-per-model methods with significantly less training cost. + +# 2. Related Works + +Arbitrary-Style-Per-Model Methods. A lot of works achieve arbitrary style transfer via global feature transformation, e.g., WCT [22], AdaIN [13], Linear style transfer [21], DIN [15], MCCNet [2], and MAST [14]. In general, they can achieve the most attainable style transfer speed but dismiss stylized effects for local details a lot. In order to add more consideration of local details to global transformation based methods, there are patch-similarity based solution [1, 10, 37] and attention based method [3, 24, 31, 44]. While paying attention to local details, it is still hard to transfer complex style patterns and prone to unsatisfactory distortions due to the simple swap and fusion strategies. + +Transformer in Style Transfer. Transformer proposed in [40] is widely used in the natural language processing and has become a powerful baseline. Recently, Transformer receives better performance than CNN models in many vision tasks and are involved in various model zoos [6, 11, 18, 42, 43, 45]. Specifically, in the field of style transfer, [41] propose StyFormer, where the transformation is driven by the + +![](images/660f6ed346876c1f9579582f50f1dbc7664726a31894ee5ac71a3b2cf7b661c1.jpg) +Figure 3. Overview of our model architecture. + +cross-attention module in Transformer. [4] develop a Transformer model based on ViT [5]. However, the number of parameters and training difficulty would increase considerably with the increasing of Transformer layers. Also, following residual connections in original Transformer, it suffers from the distortion problem on content structures as shown in Fig. 2. Similar problems also exist in [27]. + +Meta Learning in Style Transfer. On this routine, [36] train a meta network to predict parameters of the generator model for one style reference image. [47] employ MAML algorithm [7] to find a style-free model for fast adaptation. Our method is different from theirs in three aspects: (1) our framework works for both few-shot and arbitrary style transfer thanks to the cross-attention in Transformer; (2) our meta learning algorithm is based on Reptile [30], a first-order meta learning algorithm without the necessity to operate gradients of higher levels, which is more efficient in training; and (3) only style encoder instead of the whole model needs to be updated during the few-shot learning stage, which makes it more convenient in practice. + +Text-Guided Synthesis. The emergence of CLIP model [34] bridges text and image domain, which supports a series of works on text-guided synthesis [8, 32, 35]. In the field of style transfer, CLIPstyler [20] proposes a patch-wise CLIP loss and achieves text-guided style transfer. Nevertheless, as a per-text-per-model solution, it is inconvenient in practice to handle a large number of text inputs. Despite the recent dataset distillation scheme may alleviate this issue, existing approaches, however, largely focus on images as input [25, 28, 46]. In this paper, we first achieve text-guided few-shot style transfer with the proposed meta learning algorithm, which improves the flexibility of CLIPstyler significantly. + +# 3. Methods + +In this section, we give details of the proposed Meta Style Transformer (Master) for zero-shot and few-shot style + +transfer. We first introduce the network architecture of the proposed model, then illustrate how we train our model in a meta-learning fashion, and finally describe loss function. + +# 3.1. Network Architecture + +The proposed model comprises an encoder, a feature modification module, and a decoder, as demonstrated in Fig. 3. We employ the first 2 stages of Swin Transformer [29] as the encoder Enc to extract common image features for both content and style images. The decoder Dec follows the setting of [13] with 3 upsampling convolutional blocks. For feature modification, we propose a Style Transformer module for transferring complex style patterns, which will be introduced later. Taking a style image $I_{s}$ and a content image $I_{c}$ as input, we first divide them into $4 \times 4$ patches and extract their corresponding feature maps $F_{c}$ and $F_{s}$ with the Swin encoder Enc: + +$$ +\begin{array}{l} F _ {c} = \operatorname {E n c} (I _ {c}), \\ F _ {s} = \operatorname {E n c} \left(I _ {s}\right), \tag {1} \\ \end{array} +$$ + +where the spatial scales for $F_{c}$ and $F_{s}$ are 1/8 of those for $I_{c}$ and $I_{s}$ . Then, the embedded feature $F_{cs}$ is derived by: + +$$ +F _ {c s} = \operatorname {S t y l e T r a n s} \left(F _ {s}, F _ {c}\right), \tag {2} +$$ + +where StyleTrans denotes the Style Transformer module, and $F_{cs}$ has the same shape as $F_{c}$ . Finally, we can synthesize the stylized image $I_{cs}$ with the decoder as: + +$$ +I _ {c s} = \operatorname {D e c} \left(F _ {c s}\right). \tag {3} +$$ + +Style Transformer. Similar to previous Transformer-based style transfer models like StyTr2 [4], the Transformer encoder is used to encode style information while the Transformer decoder takes charge of content-style interaction. In our model, there is only one copy of parameters shared by all Transformer encoder and decoder layers. Also, different + +from standard Transformer where decoder layers would follow all encoder layers, encoder and decoder layers would be executed in an alternate fashion. The next layer would take the output of the current layer as input. + +Our Transformer decoder layer is composed of self-attention, cross-attention, and non-linear blocks, for content encoding and content-style interaction. Specifically, content features are first processed with a self-attention step. Then, cross-attention is conducted by taking the content encoding as query and the style encoding as key and value, followed by a MLP for non-linear transformation. + +Notably, in vanilla Transformer, features before and after the cross-attention step are fused by a residual connection, which is harmful for content structures as analyzed in Fig. 2. We thereby replace the residual connection with dynamic and learnable scaling and shifting steps, whose parameters are determined by the style encoder. In this way, the output of style encoder should consist of three parts: key for the following cross-attention $K_{s}$ , scaling parameters $V_{\sigma}$ , and shifting parameters $V_{\mu}$ . The prediction of each part shares a same self-attention map to save memory but uses independent non-linear transformation. The process in the Transformer encoder layer can be formulated as: + +$$ +\mathrm {M H A} (Q, K, V) = [ h e a d _ {1}, \dots , h e a d _ {h} ] W ^ {O}, +$$ + +$$ +\operatorname {h e a d} _ {i} = \operatorname {A t t} \left(Q W ^ {Q _ {i}}, K W ^ {K _ {i}}, V W ^ {V _ {i}}\right), +$$ + +$$ +\operatorname {A t t} (Q, K, V) = \operatorname {s o f t m a x} (\frac {Q K ^ {\top}}{\sqrt {d _ {k}}} V), +$$ + +$$ +\begin{array}{l} K _ {s} ^ {\prime} = K _ {s} + \operatorname {M H A} \left(K _ {s}, K _ {s}, K _ {s}\right), \\ \text {H a r m i n f i g u r a t i o n} \quad \text {H a r m i n f i g u r a t i o n} \quad \text {H a r m i n f i g u r a t i o n} \quad \text {H a r m i n f i g u r a t i o n} \quad \text {H a r m i n f i g u r a t i o n} \quad \text {H a r m i n f i l t e} \end{array} \tag {4} +$$ + +$$ +K _ {s} = K _ {s} ^ {\prime} + \operatorname {M L P} \left(K _ {s} ^ {\prime}\right), +$$ + +$$ +V _ {\sigma} ^ {\prime} = V _ {\sigma} + \operatorname {M H A} \left(K _ {s}, K _ {s}, V _ {\sigma}\right), +$$ + +$$ +V _ {\sigma} = V _ {\sigma} ^ {\prime} + \operatorname {M L P} \left(V _ {\sigma} ^ {\prime}\right), +$$ + +$$ +V _ {\mu} ^ {\prime} = V _ {\mu} + \mathrm {M H A} \left(K _ {s}, K _ {s}, V _ {\mu}\right), +$$ + +$$ +V _ {\mu} = V _ {\mu} ^ {\prime} + \operatorname {M L P} \left(V _ {\mu} ^ {\prime}\right), +$$ + +where $K_{s}, V_{\sigma}$ , and $V_{\mu}$ are initialized as the style feature $F_{s}$ before the first Transformer encoder layer. We do not incorporate normalization to encode style features since second-order statistics can largely represent style information. + +Then, in the cross-attention of Transformer decoder, scaling and shifting parameters for each content feature point is aggregated from $V_{\sigma}$ and $V_{\mu}$ respectively according to the cross-attention map. The process in one Transformer decoder block can be written as: + +$$ +F _ {c s} ^ {\prime} = F _ {c s} + \mathrm {M H A} \left(F _ {c s}, F _ {c s}, F _ {c s}\right), +$$ + +$$ +\sigma = \operatorname {M H A} \left(\operatorname {I N} \left(F _ {c s} ^ {\prime}\right), \operatorname {I N} \left(K _ {s}\right), V _ {\sigma}\right), +$$ + +$$ +\mu = \mathrm {M H A} (\mathrm {I N} \left(F _ {c s} ^ {\prime}\right), \mathrm {I N} \left(K _ {s}\right), V _ {\mu}), \tag {5} +$$ + +$$ +F _ {c s} ^ {\prime \prime} = \sigma \odot F _ {c s} ^ {\prime} + \mu , +$$ + +$$ +F _ {c s} = F _ {c s} ^ {\prime \prime} + \operatorname {M L P} \left(F _ {c s} ^ {\prime \prime}\right), +$$ + +Algorithm 1 Meta Training + +Require: $\mathcal{D}_c$ : content dataset; $\mathcal{D}_s$ : style dataset; $\delta$ : inner learning rate; $\eta$ : outer learning rate; $k$ : number of inner updates; $T$ : maximal number of stacked layers; + +Ensure: trained meta generator parameters $\theta$ + +1: initialize $\theta$ randomly +2: for iteration 1, 2, 3, ... do +3: sample a style image $I_{s}$ from $\mathcal{D}_s$ +4: $\omega \gets \theta$ +5: for $k$ times do +6: sample a batch of content image $I_{c}$ from $\mathcal{D}_c$ +7: sample the number of layers from $1\sim T$ +8: forward propagation using Eq. 1-5 +9: compute inner loss $L$ using Eq. 8 +10: $\omega \gets \omega -\delta \nabla L$ +11: end for +12: $\theta \gets \theta +\eta (\omega -\theta)$ +13: end for + +where IN denotes instance normalization [39] and $\odot$ represents element-wise multiplication. $F_{cs}$ is initialized as $F_{c}$ before the first Transformer decoder layer. + +# 3.2. Training Pipeline + +To achieve high-quality style transfer, we introduce a two-stage training strategy that comprises meta training and fast adaptation. The meta training stage is designed to learn a generic model initialization, while the fast adaptation adapts the network for a single style in a few iterations for few-shot style transfer. Note that zero-shot style transfer is a special case of the overall training configuration. + +Meta Training: Inspired by Reptile [30], a first-order meta learning algorithm, rendering style patterns of a specific reference style image can be viewed as a task. We seek an optimal initialization for neural networks in this stage, so that the networks can be rapidly adapted for a new task in only a few shots. The main training procedure for this stage is shown in Algorithm 1. In each iteration, we sample 1 style and $k$ batches of contents to perform inner optimization to obtain "fast weights" $\omega$ , which would later guide the update of "slow weights" $\theta$ , to move a step in the direction of $\omega - \theta$ . Notably, as there is only one group of parameters for Transformer encoder and decoder layers, we randomly choose a number as the number of stacked layers for the Style Transformer in each iteration. + +Fast Adaptation: The trained model after the first stage would serve as an initialization and will be adapted for a single style. With the same objective, this stage behaves almost identically to the internal loop in Algorithm 1. The only difference is that only parameters of the Transformer encoder layer are necessary to be updated, since (1) the Transformer encoder layer is the main component to extract style pat + +terns and has the most significant impact on stylization, and (2) it would save memory and speed up the adaptation. + +Zero-Shot Style Transfer: "Zero-Shot" means that, once the meta training is done, there is no fast adaptation stage needed and the meta model itself can support arbitrary style transfer. To encourage the model to produce satisfactory results in the zero-shot setup, we set the inner optimization time $k$ to 1, where the algorithm is reduced to the typical training paradigm of existing arbitrary style transfer methods. In this sense, Algorithm 1 provides a more general setting for style transfer in both zero-shot and few-shot cases. + +Text-Guided Style Transfer. We perform text-guided style transfer based on our image style transfer model with slight modifications. Following the common practice, a pretrained CLIP encoder is imported to extract text features. To be consistent with image input, we use a StyleGAN-like mapping network to convert text features into pseudo image features. In the meta training stage, since CLIP unifies the feature space, we use image instead of text as style input to avoid additional text dataset. In the fast adaptation stage, the mapping network along with the Transformer encoder is updated, considering there is still a gap between CLIP image feature and text feature. + +# 3.3. Loss Function + +The training objective in both meta-training and fast adaptation stages follows many works in arbitrary style transfer, which consists of content loss and style loss. Let $F^x$ represent features on ReLU- $x$ -1 layer of a pre-trained VGG19 network [38] for loss computation. The content loss is defined by the normalized perceptual loss [17]: + +$$ +\mathcal {L} _ {\text {c o n t}} = \sum_ {x = 2} ^ {5} \| \mathrm {I N} \left(F _ {c} ^ {x}\right) - \mathrm {I N} \left(F _ {c s} ^ {x}\right) \| _ {2}, \tag {6} +$$ + +while the style loss adopts the mean-variance loss [13]: + +$$ +\mathcal {L} _ {s t y} = \sum_ {x = 2} ^ {5} \left(\| \mu \left(F _ {s} ^ {x}\right) - \mu \left(F _ {c s} ^ {x}\right) \| _ {2} + \| \sigma \left(F _ {s} ^ {x}\right) - \sigma \left(F _ {c s} ^ {x}\right) \| _ {2}\right), \tag {7} +$$ + +where $\mu$ and $\sigma$ calculate the channel-wise mean and standard deviation separately. The overall objective is given by the weighted summation of the two losses: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o n t}} + \lambda \mathcal {L} _ {\text {s t y}}, \tag {8} +$$ + +where $\lambda$ controls the balance between two terms. + +In text-guided style transfer, the loss functions are the same as CLIPstyler, including global CLIP loss, directional CLIP loss and PatchCLIP loss. + +# 4. Experiments + +# 4.1. Implementation Details + +We use MS-COCO [23] as our content dataset and WikiArt test set [33] as our style dataset. The content + +
MethodLcontLsimLsty
AdaIN4.88±0.700.53±0.231.51±0.78
Linear3.93±0.760.44±0.171.97±0.96
AvatarNet5.76±0.530.56±0.193.19±1.82
SANet4.72±0.610.50±0.171.15±0.58
MANet4.93±0.570.50±0.171.41±0.77
MCCNet4.22±0.690.47±0.171.56±0.84
AdaAttN4.46±0.700.43±0.162.20±1.19
StyTr23.78±0.990.48±0.221.50±0.69
StyFormer4.94±0.790.43±0.152.20±1.44
MetaNet3.48±0.850.45±0.192.47±1.69
MetaStyle*3.64±1.120.42±0.182.47±1.06
Johnson†4.60±0.760.59±0.201.02±0.34
Ours-Vanilla5.50±0.630.59±0.260.85±0.38
Ours-Norm4.70±0.750.43±0.130.93±0.33
Ours-ZS-L14.13±0.680.41±0.140.92±0.40
Ours-ZS-L34.20±0.680.41±0.130.81±0.31
Ours-FS*4.24±0.820.38±0.120.79±0.25
+ +Table 1. Quantitative comparisons. ZS and FS for our model denote zero-shot and few-shot modes. Vanilla denote replacing our architecture with original Transformer. Norm means adding layer normalization in the Transformer encoder layer. L1/L3 means using 1/3 Transformer layers in the test time. * and † denote few-shot and per-style-per-model methods. + +dataset contains roughly 80,000 images and the style dataset has about 20,000 images. The optimizer is Adam [19] with learning rates of both inner and outer loops set as 0.0001 and the batch size is 4. In training, we first resize the content and style image to $512 \times 512$ and then randomly crop to $256 \times 256$ resolution. During inference, our model can handle inputs of any size. The update times for inner optimization $k$ is set as 2 for few-shot case and 1 for zero-shot case. In training, the maximal number of stacked layers is 4. All the multi-head attention blocks are instantiated as shifted window attention in [29], with window size 8 and shift size 4. The model is trained on a Nvidia 3090 with 9k iterations in the meta training stage for convergence while only 100 steps for image input and 20 steps for text input in the fast adaptation for few-shot style transfer, which takes less than 1 minute. Hyper-parameter $\lambda$ is set as 10. + +# 4.2. Comparison with Prior Works + +In this section, we compare results by our Master with 13 state-of-the-art style transfer methods, including 3 global transformation based methods (AdaIN [13], Linear style transfer [21], and MCCNet [2]), 1 patch swap based method (Avatar-Net [37]), 3 attention based methods (SANet [31], MANet [3], and AdaAttN [24]), 2 transformer based methods (StyTr2 [4] and StyFromer [41]), 2 meta learning based methods (MetaNet [36] and MetaStyle [47]), 1 per-style-per-model method by Johnson et al. [17] and 1 text-guided style transfer method (CLIPstyler). + +Quantitative Comparison. We adopt content loss $\mathcal{L}_{cont}$ in Eq. 6 and style loss $\mathcal{L}_{sty}$ in Eq. 7 as evaluation metrics to reflect effects of learning by different methods. We also design a metric $\mathcal{L}_{sim}$ to reflect the preservation of the spatial-wise self cosine similarity of content structures: + +$$ +D _ {*, i j} ^ {x} = 1 - \frac {F _ {*, i} ^ {x} \cdot F _ {*, j} ^ {x}}{\| F _ {*, i} ^ {x} \| \| F _ {*, j} ^ {x} \|} +$$ + +$$ +\mathcal {L} _ {s i m} = \sum_ {x = 3} ^ {4} \frac {1}{n _ {x} ^ {2}} \sum_ {i, j} \left| \frac {D _ {c , i j} ^ {x}}{\sum_ {k} D _ {c , k j} ^ {x}} - \frac {D _ {c s , i j} ^ {x}}{\sum_ {k} D _ {c s , k j} ^ {x}} \right|, \tag {9} +$$ + +where $n_x$ is the number of spatial locations for the current feature map and the second foot script denotes the spatial index. Smaller $\mathcal{L}_{sim}$ means that the original spatial-wise relationship is better preserved during style transfer, i.e., less content distortion. We use the test dataset in the code page of [13] for evaluation, with 11 content images and 20 style images to form 220 content-style pairs. We report the mean and standard deviation over the 220 cases in Tab. 1. Notably, the lowest style loss and content similarity loss are achieved by Master simultaneously compared with previous methods, even better than the per-style-per-model solution by Johnson et al. [17], and can be further reduced in the few-shot case with comparable content loss values, which demonstrates the joint advantages of our method in content preserving and style rendering. Meanwhile, the lower standard deviations demonstrate the robustness of our model. + +Qualitative Comparison. Qualitative examples by different methods are shown in Fig. 4. Full comparisons with more methods can be found in the supplementary material. + +We first give discussion on the global-transformation based methods. As shown in the 3rd and 4th rows, in AdaIN and Linear, features in all positions share the same transformation matrix, which fails to migrate detail textures in style images and has poor color saturation. + +As for the attention-based methods, SANet in the 5th row introduces local focus to style images. However, due to the simple fusion of one-layer attention results and original content features, it brings textures and content distortion in many cases, e.g., background areas of the 4th and 5th columns. AdaAttN in the 6th row dedicates to addressing the distortion problem in SANet, but it sacrifices the ability to render salient style patterns and textures. + +We then compare our results with those by Transformer-based methods. StyFormer in the 7th row adopts a design of the cross-attention module in Transformer without exploring the effect of self-attention to enhance style rendering. Thus, the results appear under-stylized compared with those by our method. StyTr2 in the 8th row is a Transformer architecture with a pure self-attention mechanism, which makes the performance on the extraction and migration of local textures not satisfactory enough, e.g., strokes of the 2nd column and waves of the 4th column. Moreover, since it follows the design of residual connections in + +![](images/a5dabdc7252bdd13c46365be822eca167bb8b12a04afc528073f9d2fb5489fe9.jpg) +Figure 4. Comparison with previous state-of-the-art style transfer methods. Zoom in for better details. + +![](images/1721071f92e1d26376da096c2d9d544732fa1a71d395d93c7abc1eb63b3f4852.jpg) +Figure 5. Comparison with CLIPstyler in text-guided style transfer. Zoom in for better details. + +vanilla Transformer, the issue of content distortion in Fig. 2 is still evident, as shown in the background areas of the 4th and 5th columns. Compared with these methods, Master renders more vivid global and local style patterns. Meanwhile, it well maintains content structures simultaneously with learnable scaling and shifting parameters. Furthermore, in the few-shot case, comparable performance can be achieved with only 1 Transformer layer. + +MetaNet and MetaStyle also involve the concept of meta learning into their style transfer frameworks. On the one hand, MetaNet predicts parameters of a generator network given one style image via a meta network, which is an ambitious goal and makes the training more difficult. Thus, as shown in the 9th row, the stylized effects of their results still appear weak. On the other hand, MetaStyle achieves inferior performance in terms of local details as demonstrated in the 10th row, since only parameters of global normalization layers are learnable in the fast adaptation stage. + +The single style transfer method by Johnson et al. in the 11th row tends to arrange patterns learned by the network arbitrarily, which may produce undesired and distorted effects. By contrast, the complex content-style dependence is constructed in our Transformer model and thus achieves more remarkable performance on transferring style patterns to proper positions in content images. + +Some qualitative comparisons on text-guided style transfer with CLIPstyler are shown in Fig. 5, which uses per-text-per-model fashion. Through these results, we observe that despite significantly more training steps, CLIPstyler + +![](images/ae3e93bf6b1addfd77baed52ef7ab4a7c4f0b911e56234fde9800a6353b26662.jpg) + +![](images/5a50828409e302f2615288a1c83c3243beb6c2661eac1362b7bd05dac496bfc1.jpg) +Figure 6. Results of user study. Left: Our zero-shot stylization results compared with SOTA methods. Right: Our few-shot stylization comparison results. +Figure 7. Ablation study on key designs in Master model. + +still suffers from weak stylization. By contrast, our method generates more vivid results in only 20 adaptation steps, which demonstrates the advantages of the proposed Master architecture and the versatility of our training pipeline. + +User Study. Following most style transfer works, we conduct user study and report user preference. We choose 20 content images and 20 style images to form 400 content-style pairs. We involve 100 people and randomly assign 20 stylized results from compared methods to each subject. Our method showcases zero-shot results with 3 Transformer layers in 10 tests and few-shot results in the remaining 10. For each pair, the order of results is randomized, and participants choose their favorite. With 2,000 votes in total, Fig. 6 demonstrates Master's superior style transfer quality. + +# 4.3. Ablation Study + +Architecture. We conduct ablation studies on key designs in Master to illustrate their effects on the stylization quality. As shown in Fig. 7, on the one hand, if vanilla Transformer model is used, without learnable scaling and shifting parameters, noisy textures can be introduced significantly, which distorts the original content structures and affects style transfer quality. On the other hand, if layer normalization operations in the standard Transformer are used in the style encoder, style patterns would become less saturated, since normalization removes second-order feature statistics which represents the style information to a large degree. Quantitative studies in Tab. 1 also support our analysis and come to the same conclusion. + +Meta Training and Fast Adaptation. We provide training loss visualization in Fig. 8, to demonstrate the effectiveness of our meta training and fast adaptation algorithm. Our meta training stage learns an appropriate initial state, which enables fast decent and convergence of loss values in the fast adaptation stage. We show the same visualization in the per-style-per-model setting as a benchmark, using the same loss function and network architecture. As shown in Fig. 8, + +![](images/c18a13ed7d1b1d1a8d60849660f554dee74e4b644a5f92bebaf63eb8e21614ca.jpg) +Figure 8. Training loss visualization during meta training and fast adaptation of three transformer based models: StyTr2, StyFormer, and our Master. The loss curve under per-style-per-model setting by our model is used as a reference. Ours-wo-Fix means that all the parameters need to be updated during the fast adaptation stage. + +it requires roughly 3k iterations for the model to be convergent in the per-style-per-model case (Single Style), while the meta model (Ours) after 9k iterations can be adapted for any style image in only a few shots with competitive training results. Thus, the total number of iterations for our method is significantly smaller than per-style-per-model ones with the growth number of required styles. + +By default, all the parameters except those in the style encoder module are fixed during the fast adaptation stage. We also experiment with training the whole model in this stage (Ours wo Fix in Fig. 8) and find that the training and convergence would become difficult and require more time and memory. Moreover, updating all the parameters may result in insufficient adaptation given a limited number of training steps. As demonstrated in Fig. 7, there is a gap on the global tone between the result and the style image. + +Base Model. We evaluate StyFormer and StyTr2 as base models, presenting loss visualizations. Heavier Transformers raise training difficulty for earlier models, resulting in inferior convergence. When the number of Transformer layers is 5, StyTr2 even fails to converge. By contrast, our model adopts parameter-sharing across all Transformer layers, which results in an overall light-weight structure. Consequently, it enjoys better training effects in the meta training stage as shown in Fig. 8. Moreover, as shown in the zoom-in part, Master also enables overall lower and more stable training in fast adaptation. Qualitative examples by different base models can be found in the supplement. + +Controllable Style Transfer. We compare with some widely-adopted approaches for the similar goal to control the degree of stylization, including (1) Recursion, which treats stylized results as input of content images in the next iteration, and (2) Linear Mixup, which conducts linear combinations between content features and stylized features before the decoder. Vanilla Transformer is adopted for these approaches. The major difference between Recursion and ours is that our method takes the iterative sty + +![](images/4a3aef678cececb99a53609e418444464b1b2a588529b74f8e8784f6a6283a0a.jpg) +Figure 9. Comparisons with widely-adopted approaches to control the degree of stylization. + +ization into consideration in the training time so that the model would learn to transfer style patterns layer-by-layer. Thus, as shown in Fig. 9, Recursion merely uses more intensive colors with the increasing of repeat times. Also, for Linear Mixup, it is typically hard to generate reasonable style transfer results especially when the weight of contents is high. Compared with them, our method is capable of rendering more style patterns and increasing the artistic abstraction when more Transformer layers are stacked. + +# 5. Conclusion + +In this paper, we propose a novel Transformer model specifically for artistic style transfer, termed as Meta Style Transformer (Master). On one hand, parameters of different Transformer layers are shared, which reduces the total number of parameters significantly and thus enables easier convergence. It is also convenient for Master to control the degree of stylization via customizing the number of stacked layers in the inference time. On the other hand, different from standard Transformer, our model adopts dynamic and learnable scaling and shifting operations instead of original residual connections, which helps preserve similarity relationship in content structures while migrating remarkable style patterns. Specifically, Master is trained using a meta learning algorithm for few-shot style transfer. Only parameters of the Transformer encoder layer need to be updated in the few-shot stage, which benefits the fast and robust adaptation. Zero-shot arbitrary style transfer is a special case of the training configuration. Experiments suggest that Master outperforms previous state-of-the-art arbitrary style transfer methods on both content preserving and style rendering in the zero-shot case. + +# Acknowledgment + +This project is supported by the Singapore Ministry of Education Academic Research Fund Tier 1 (WBS: A-0009440-01-00). + +# References + +[1] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2 +[2] Yingying Deng, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, and Changsheng Xu. Arbitrary video style transfer via multi-channel correlation. arXiv preprint arXiv:2009.08003, 2020. 2, 5 +[3] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2719-2727, 2020. 1, 2, 5 +[4] Yingying Deng, Fan Tang, Xingjia Pan, Weiming Dong, Chongyang Ma, and Changsheng Xu. $stytr^2$ : Unbiased image style transfer with transformers, 2021. 1, 3, 5 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2020. 3 +[6] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks, 2017. 3 +[8] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3 +[9] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1 +[10] Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. Arbitrary style transfer with deep feature reshuffle. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8222–8231, 2018. 2 +[11] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2 +[13] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 1, 2, 3, 5, 6 +[14] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 14861-14869, 2021. 2 +[15] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4369-4376, 2020. 2 +[16] Yongcheng Jing, Yining Mao, Yiding Yang, Yibing Zhan, Mingli Song, Xinchao Wang, and Dacheng Tao. Learning graph neural networks for image style transfer. In ECCV, 2022. 1 +[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 1, 5, 6 +[18] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey, 2021. 2 +[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3 +[21] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast arbitrary style transfer. arXiv preprint arXiv:1808.04537, 2018. 1, 2, 5 +[22] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. arXiv preprint arXiv:1705.08086, 2017. 1, 2 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5 +[24] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6649-6658, October 2021. 1, 2, 5 +[25] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. In Conference on Neural Information Processing Systems, 2022. 3 +[26] Songhua Liu, Hao Wu, Shoutong Luo, and Zhengxing Sun. Stable video style transfer based on partial convolution with depth-aware supervision. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2445-2453, 2020. 1 +[27] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In European Conference on Computer Vision, 2022. 3 +[28] Songhua Liu, Jingwen Ye, Runpeng Yu, and Xinchao Wang. Slimmable dataset condensation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3 + +[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3, 5 +[30] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms, 2018. 3, 4 +[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5880–5888, 2019. 1, 2, 5 +[32] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 3 +[33] Fred Phillips and Brandy Mackintosh. Wiki art gallery, inc.: A case for critical thinking. Issues in Accounting Education, 26(3):593-608, 2011. 5 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3 +[36] Falong Shen, Shuicheng Yan, and Gang Zeng. Meta networks for neural style transfer, 2017. 3, 5 +[37] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8242-8250, 2018. 2, 5 +[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 5 +[39] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4 +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 1, 2 +[41] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 2, 5 +[42] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, 2022. 2 +[43] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Conference on Neural Information Processing Systems, 2022. 2 + +[44] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1467-1475, 2019. 2 +[45] Jingwen Ye, Songhua Liu, and Xinchao Wang. Partial network cloning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[46] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. 3 +[47] Chi Zhang, Yixin Zhu, and Song-Chun Zhu. Metastyle: Three-way trade-off among speed, flexibility, and quality in neural style transfer, 2019. 3, 5 \ No newline at end of file diff --git a/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/images.zip b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7895583f1d0dbea8a874be56e8d5fd7ad4bbc2a1 --- /dev/null +++ b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4f44e62dab2e092a2be1f849885758c4c3ec213a41454f23c37c61d10103b11 +size 850956 diff --git a/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/layout.json b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6d0b560d88170a3cfb5817b71f0a4bbd713e37f2 --- /dev/null +++ b/2023/Master_ Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer/layout.json @@ -0,0 +1,9380 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 527, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 527, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 527, + 139 + ], + "type": "text", + "content": "Master: Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "spans": [ + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "text", + "content": "Hao Tang" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "text", + "content": " Songhua Liu" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "text", + "content": " Tianwei Lin" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "text", + "content": " Shaoli Huang" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "text", + "content": " Fu Li" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "text", + "content": " Dongliang He" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "text", + "content": " Xinchao Wang" + }, + { + "bbox": [ + 148, + 160, + 452, + 190 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 190, + 462, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 190, + 462, + 204 + ], + "spans": [ + { + "bbox": [ + 139, + 190, + 462, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 190, + 462, + 204 + ], + "type": "text", + "content": "Center for Data Science, Peking University " + }, + { + "bbox": [ + 139, + 190, + 462, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 190, + 462, + 204 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 204, + 368, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 204, + 368, + 216 + ], + "spans": [ + { + "bbox": [ + 222, + 204, + 368, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 222, + 204, + 368, + 216 + ], + "type": "text", + "content": "VIS, Baidu Inc. " + }, + { + "bbox": [ + 222, + 204, + 368, + 216 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 222, + 204, + 368, + 216 + ], + "type": "text", + "content": "Tencent AI Lab" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 175, + 220, + 426, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 220, + 426, + 231 + ], + "spans": [ + { + "bbox": [ + 175, + 220, + 426, + 231 + ], + "type": "text", + "content": "tanghao@stu.pku.edu.cn songhua.liu@u.nus.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 233, + 539, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 233, + 539, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 233, + 539, + 245 + ], + "type": "text", + "content": "{lintianwei01,lifu,hedongliang01}@baidu.com shaolihuang@tencent.com xinchao@nus.edu.sg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 47, + 251, + 130, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 251, + 130, + 308 + ], + "lines": [ + { + "bbox": [ + 47, + 251, + 130, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 130, + 308 + ], + "type": "image", + "image_path": "2265e982334aa3d38d302ca2cd0acb11c8f855a706e3c1f58c5e686b5be6a7d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 310, + 94, + 318 + ], + "lines": [ + { + "bbox": [ + 84, + 310, + 94, + 318 + ], + "spans": [ + { + "bbox": [ + 84, + 310, + 94, + 318 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 251, + 220, + 309 + ], + "blocks": [ + { + "bbox": [ + 133, + 251, + 220, + 309 + ], + "lines": [ + { + "bbox": [ + 133, + 251, + 220, + 309 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 220, + 309 + ], + "type": "image", + "image_path": "b7116b00247e4e9aa669574d1c349a4f34f15acb837a839f8ca2d4d8ea67e370.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 310, + 179, + 318 + ], + "lines": [ + { + "bbox": [ + 169, + 310, + 179, + 318 + ], + "spans": [ + { + "bbox": [ + 169, + 310, + 179, + 318 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 324, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 324, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 546, + 357 + ], + "type": "text", + "content": "Figure 1. Left: With more Transformer layers, the number of parameters and training difficulty increase significantly. Right: Our method supports test-time controlling of stylization degree via tuning the number of adopted Transformer layers. It is also readily adaptable to few-shot style transfer, where stylization with only 1 layer can be further improved. Text-guided few-shot style transfer is achievable." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 223, + 254, + 274, + 308 + ], + "blocks": [ + { + "bbox": [ + 223, + 254, + 274, + 308 + ], + "lines": [ + { + "bbox": [ + 223, + 254, + 274, + 308 + ], + "spans": [ + { + "bbox": [ + 223, + 254, + 274, + 308 + ], + "type": "image", + "image_path": "4729d7188c270f0708fa04f15441404551f99646e6687f632d04da8c3946734a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 310, + 269, + 319 + ], + "lines": [ + { + "bbox": [ + 228, + 310, + 269, + 319 + ], + "spans": [ + { + "bbox": [ + 228, + 310, + 269, + 319 + ], + "type": "text", + "content": "Content/Style" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 276, + 255, + 329, + 306 + ], + "blocks": [ + { + "bbox": [ + 276, + 255, + 329, + 306 + ], + "lines": [ + { + "bbox": [ + 276, + 255, + 329, + 306 + ], + "spans": [ + { + "bbox": [ + 276, + 255, + 329, + 306 + ], + "type": "image", + "image_path": "1a002410cbc987d143e156f13f952eac456cc3a0c7b3485606bd80dc4d775d76.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 310, + 314, + 319 + ], + "lines": [ + { + "bbox": [ + 290, + 310, + 314, + 319 + ], + "spans": [ + { + "bbox": [ + 290, + 310, + 314, + 319 + ], + "type": "text", + "content": "Layer 1" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 331, + 255, + 383, + 307 + ], + "blocks": [ + { + "bbox": [ + 331, + 255, + 383, + 307 + ], + "lines": [ + { + "bbox": [ + 331, + 255, + 383, + 307 + ], + "spans": [ + { + "bbox": [ + 331, + 255, + 383, + 307 + ], + "type": "image", + "image_path": "c0397542bebbcff4821cc5b411605abc84da086049ea6fbee22ed950c4a0291e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 310, + 369, + 319 + ], + "lines": [ + { + "bbox": [ + 345, + 310, + 369, + 319 + ], + "spans": [ + { + "bbox": [ + 345, + 310, + 369, + 319 + ], + "type": "text", + "content": "Layer 3" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 386, + 255, + 437, + 307 + ], + "blocks": [ + { + "bbox": [ + 386, + 255, + 437, + 307 + ], + "lines": [ + { + "bbox": [ + 386, + 255, + 437, + 307 + ], + "spans": [ + { + "bbox": [ + 386, + 255, + 437, + 307 + ], + "type": "image", + "image_path": "40324f219537218ca4ed4779076be71593839fa09484551872352291beca21b5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 310, + 422, + 319 + ], + "lines": [ + { + "bbox": [ + 399, + 310, + 422, + 319 + ], + "spans": [ + { + "bbox": [ + 399, + 310, + 422, + 319 + ], + "type": "text", + "content": "Layer 5" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 439, + 255, + 491, + 307 + ], + "blocks": [ + { + "bbox": [ + 439, + 255, + 491, + 307 + ], + "lines": [ + { + "bbox": [ + 439, + 255, + 491, + 307 + ], + "spans": [ + { + "bbox": [ + 439, + 255, + 491, + 307 + ], + "type": "image", + "image_path": "35755527d19bfc364ca5ffc42f6860c3ab7720adba3e4c155eb979032abf5e41.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 310, + 479, + 318 + ], + "lines": [ + { + "bbox": [ + 451, + 310, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 451, + 310, + 479, + 318 + ], + "type": "text", + "content": "Few Shot" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 493, + 255, + 545, + 307 + ], + "blocks": [ + { + "bbox": [ + 493, + 255, + 545, + 307 + ], + "lines": [ + { + "bbox": [ + 493, + 255, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 493, + 255, + 545, + 307 + ], + "type": "image", + "image_path": "da914c77b9f2de5ee7f43e92906edb89a2753d0ba9afa1d21147b3b26e2559cb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 308, + 542, + 324 + ], + "lines": [ + { + "bbox": [ + 495, + 308, + 542, + 324 + ], + "spans": [ + { + "bbox": [ + 495, + 308, + 542, + 324 + ], + "type": "text", + "content": "\"The scream by Edvard Munch\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 143, + 369, + 192, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 369, + 192, + 381 + ], + "spans": [ + { + "bbox": [ + 143, + 369, + 192, + 381 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 386, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 289, + 685 + ], + "type": "text", + "content": "Transformer-based models achieve favorable performance in artistic style transfer recently thanks to its global receptive field and powerful multi-head/layer attention operations. Nevertheless, the over-paramerized multilayer structure increases parameters significantly and thus presents a heavy burden for training. Moreover, for the task of style transfer, vanilla Transformer that fuses content and style features by residual connections is prone to content-wise distortion. In this paper, we devise a novel Transformer model termed as Master specifically for style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters, which (1) reduces the total number of parameters, (2) leads to more robust training convergence, and (3) is readily to control the degree of stylization via tuning the number of stacked layers freely during inference. On the other hand, different from the vanilla version, we adopt a learnable scaling operation on content features before content-style feature interaction, which better preserves the original similarity between a pair of content features while ensuring the stylization quality. We also propose a novel meta learning scheme for the proposed model so that it can not only work in the typical setting of arbitrary style transfer, but also adaptable to the few-shot setting, by only fine-tuning the Transformer encoder layer in the few-shot stage for one" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 370, + 546, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 546, + 418 + ], + "type": "text", + "content": "specific style. Text-guided few-shot style transfer is firstly achieved with the proposed framework. Extensive experiments demonstrate the superiority of Master under both zero-shot and few-shot style transfer settings." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 441, + 387, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 387, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 387, + 453 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "type": "text", + "content": "Artistic style transfer aims at applying style patterns like colors and textures of a reference image to a given content image while preserving the semantic structure of the content. In contrast to the pioneering optimization method [9] and early per-style-per-model methods like [17, 26], arbitrary style transfer methods [3, 13, 16, 21, 22, 24, 31] enable real time style transfer for any style image in the test time in a zero-shot manner. The flexibility has led to this arbitrary-style-per-model fashion to dominate style transfer research." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Recently, to enhance the representation of global information in arbitrary style transfer, Transformer [40] is introduced to this area [4], leveraging the global receptive field and powerful multi-head/layer structure, and achieves superior performance. Nevertheless, the over-parameterized multi-layer structure increases model parameters significantly. As shown in Fig. 1(a), there are 25.94M learnable parameters for a 3-layer Transformer structure in StyTr2 [4], v.s. 3.50M in AdaIN [13], a simple but effective baseline in arbitrary style transfer. Such a large model for standard Transformer inevitably presents a heavy burden for training. As shown in Fig. 1(b), when there are more than" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18329" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 287, + 115 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 115 + ], + "type": "image", + "image_path": "acf3d3dbaedee70ad481c29fb68b35357f3ce3b3eac58ecc396221abd620540a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 118, + 119, + 186 + ], + "blocks": [ + { + "bbox": [ + 49, + 118, + 119, + 186 + ], + "lines": [ + { + "bbox": [ + 49, + 118, + 119, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 119, + 186 + ], + "type": "image", + "image_path": "c7782a5691269e99e26b011c6c4e1f635c1de7b3e3bfc1a23aafda69325db50a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 187, + 94, + 196 + ], + "lines": [ + { + "bbox": [ + 71, + 187, + 94, + 196 + ], + "spans": [ + { + "bbox": [ + 71, + 187, + 94, + 196 + ], + "type": "text", + "content": "Style /" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 197, + 96, + 205 + ], + "lines": [ + { + "bbox": [ + 69, + 197, + 96, + 205 + ], + "spans": [ + { + "bbox": [ + 69, + 197, + 96, + 205 + ], + "type": "text", + "content": "Content" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 214, + 287, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 270 + ], + "type": "text", + "content": "Figure 2. Residual connection in the vanilla Transformer tends to destroy the original similarity relationship on content structure in style transfer task. Our model is designed to address this problem with learnable scaling parameters. The top row shows a simple 2D visualization and the bottom one provides a qualitative example." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 132, + 118, + 200, + 186 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 200, + 186 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 200, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 200, + 186 + ], + "type": "image", + "image_path": "334709ab8081e5446453ce540b589cd652765d759204889f2c21e15448f6b559.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 187, + 198, + 196 + ], + "lines": [ + { + "bbox": [ + 134, + 187, + 198, + 196 + ], + "spans": [ + { + "bbox": [ + 134, + 187, + 198, + 196 + ], + "type": "text", + "content": "Vanilla Transformer" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 197, + 203, + 205 + ], + "lines": [ + { + "bbox": [ + 129, + 197, + 203, + 205 + ], + "spans": [ + { + "bbox": [ + 129, + 197, + 203, + 205 + ], + "type": "text", + "content": "w Residual Connection" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 118, + 283, + 186 + ], + "blocks": [ + { + "bbox": [ + 215, + 118, + 283, + 186 + ], + "lines": [ + { + "bbox": [ + 215, + 118, + 283, + 186 + ], + "spans": [ + { + "bbox": [ + 215, + 118, + 283, + 186 + ], + "type": "image", + "image_path": "07eaa7e3cd6c4e7e2d6df88b917127ce9d1f0ece8a6651896c3c817781428a86.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 187, + 275, + 196 + ], + "lines": [ + { + "bbox": [ + 219, + 187, + 275, + 196 + ], + "spans": [ + { + "bbox": [ + 219, + 187, + 275, + 196 + ], + "type": "text", + "content": "Our Transformer" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 197, + 281, + 205 + ], + "lines": [ + { + "bbox": [ + 216, + 197, + 281, + 205 + ], + "spans": [ + { + "bbox": [ + 216, + 197, + 281, + 205 + ], + "type": "text", + "content": "w Learnable Scaling" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 279, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 315 + ], + "type": "text", + "content": "4 layers, vanilla Transformer even fails to get convergent in training, which limits the scalability of the Transformer model in style transfer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "type": "text", + "content": "Moreover, vanilla Transformer relies on residual connections [12] to stylize content features, which suffers from the content-distortion problem. We illustrate this effect with a 2D visualization in Fig. 2(top), where residual connections lead the transformation results of two content feature vectors to move towards the dominated style features and thereby tend to eliminate their original distinction. The visual effect is that the stylized images would be dominated by some strong style features, such as salient edges, with the original self-(dis)similarity of content structures destroyed, as the example shown in Fig. 2(bottom)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 449, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 664 + ], + "type": "text", + "content": "Focusing on these drawbacks, in this paper, we are dedicated to devising a novel Transformer architecture specifically for artistic style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters and a random number of stacked layers are adopted for each training iteration. Compared with the original version, sharing parameters across different layers reduces the total number of parameters significantly and leads to more convenient training convergence. As a byproduct, it is also readily for our model to control the degree of stylization via tuning the number of stacked layers freely in the inference time, as shown in Fig. 1(right). On the other hand, we equip Transformer with learnable scale parameters for content-style interactions instead of residual connections, which alleviates content distortion to a large degree and better preserves content structures while rendering vivid style patterns simultaneously, as shown in the 2D visualization and the qualitative example in Fig. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Furthermore, beyond the typical zero-shot arbitrary style transfer, leveraging a meta learning scheme, our method is adaptable to the few-shot setting. By only fine-tuning the Transformer encoder layer in the few-shot stage, rapid" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "adaptation to the model for a specific style within a limited number of updates is possible, where the stylization with only 1 layer can be further improved, as shown in Fig. 1. Beyond that, we first achieve text-guided few-shot style transfer with this framework, which largely alleviates the training burden of previous per-text-per-model solution. In this sense, we term the overall pipeline Meta Style Transformer (Master). Our contributions are summarized as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 190, + 545, + 439 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 317, + 190, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 190, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 317, + 190, + 545, + 250 + ], + "type": "text", + "content": "- We propose a novel Transformer architecture specifically for artistic style transfer. It shares parameters between different layers, which not only helps training convergence, but also allows convenient control over the stylization effect." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 261, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 261, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 545, + 297 + ], + "type": "text", + "content": "- We identify the content distortion problem of residual connections in Transformer and propose learnable scale parameters as an option to alleviate the problem." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 308, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 308, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 308, + 545, + 357 + ], + "type": "text", + "content": "- We introduce a meta learning framework for adapting original training setting of zero-shot style transfer to the few-shot scenario, with which our Master achieves very good trade-off between flexibility and quality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 368, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 368, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 317, + 368, + 545, + 439 + ], + "type": "text", + "content": "- Experiments show that our model achieves results better than those of arbitrary-style-per-model methods. Furthermore, under the few-shot setting, either conditioned on image or text, Master can even yield performance on par with that of per-style-per-model methods with significantly less training cost." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 453, + 397, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 397, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 397, + 464 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "content": "Arbitrary-Style-Per-Model Methods. A lot of works achieve arbitrary style transfer via global feature transformation, e.g., WCT [22], AdaIN [13], Linear style transfer [21], DIN [15], MCCNet [2], and MAST [14]. In general, they can achieve the most attainable style transfer speed but dismiss stylized effects for local details a lot. In order to add more consideration of local details to global transformation based methods, there are patch-similarity based solution [1, 10, 37] and attention based method [3, 24, 31, 44]. While paying attention to local details, it is still hard to transfer complex style patterns and prone to unsatisfactory distortions due to the simple swap and fusion strategies." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Transformer in Style Transfer. Transformer proposed in [40] is widely used in the natural language processing and has become a powerful baseline. Recently, Transformer receives better performance than CNN models in many vision tasks and are involved in various model zoos [6, 11, 18, 42, 43, 45]. Specifically, in the field of style transfer, [41] propose StyFormer, where the transformation is driven by the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18330" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 548, + 235 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 548, + 235 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 548, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 548, + 235 + ], + "type": "image", + "image_path": "660f6ed346876c1f9579582f50f1dbc7664726a31894ee5ac71a3b2cf7b661c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 243, + 382, + 255 + ], + "lines": [ + { + "bbox": [ + 211, + 243, + 382, + 255 + ], + "spans": [ + { + "bbox": [ + 211, + 243, + 382, + 255 + ], + "type": "text", + "content": "Figure 3. Overview of our model architecture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 287, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 345 + ], + "type": "text", + "content": "cross-attention module in Transformer. [4] develop a Transformer model based on ViT [5]. However, the number of parameters and training difficulty would increase considerably with the increasing of Transformer layers. Also, following residual connections in original Transformer, it suffers from the distortion problem on content structures as shown in Fig. 2. Similar problems also exist in [27]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 501 + ], + "type": "text", + "content": "Meta Learning in Style Transfer. On this routine, [36] train a meta network to predict parameters of the generator model for one style reference image. [47] employ MAML algorithm [7] to find a style-free model for fast adaptation. Our method is different from theirs in three aspects: (1) our framework works for both few-shot and arbitrary style transfer thanks to the cross-attention in Transformer; (2) our meta learning algorithm is based on Reptile [30], a first-order meta learning algorithm without the necessity to operate gradients of higher levels, which is more efficient in training; and (3) only style encoder instead of the whole model needs to be updated during the few-shot learning stage, which makes it more convenient in practice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 502, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 288, + 657 + ], + "type": "text", + "content": "Text-Guided Synthesis. The emergence of CLIP model [34] bridges text and image domain, which supports a series of works on text-guided synthesis [8, 32, 35]. In the field of style transfer, CLIPstyler [20] proposes a patch-wise CLIP loss and achieves text-guided style transfer. Nevertheless, as a per-text-per-model solution, it is inconvenient in practice to handle a large number of text inputs. Despite the recent dataset distillation scheme may alleviate this issue, existing approaches, however, largely focus on images as input [25, 28, 46]. In this paper, we first achieve text-guided few-shot style transfer with the proposed meta learning algorithm, which improves the flexibility of CLIPstyler significantly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 107, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 107, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 107, + 681 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "In this section, we give details of the proposed Meta Style Transformer (Master) for zero-shot and few-shot style" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 262, + 546, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 546, + 299 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 546, + 299 + ], + "type": "text", + "content": "transfer. We first introduce the network architecture of the proposed model, then illustrate how we train our model in a meta-learning fashion, and finally describe loss function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 306, + 432, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 432, + 318 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 432, + 318 + ], + "type": "text", + "content": "3.1. Network Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "text", + "content": "The proposed model comprises an encoder, a feature modification module, and a decoder, as demonstrated in Fig. 3. We employ the first 2 stages of Swin Transformer [29] as the encoder Enc to extract common image features for both content and style images. The decoder Dec follows the setting of [13] with 3 upsampling convolutional blocks. For feature modification, we propose a Style Transformer module for transferring complex style patterns, which will be introduced later. Taking a style image " + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "text", + "content": " and a content image " + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "text", + "content": " as input, we first divide them into " + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "text", + "content": " patches and extract their corresponding feature maps " + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 305, + 325, + 546, + 469 + ], + "type": "text", + "content": " with the Swin encoder Enc:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 394, + 476, + 545, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 476, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 394, + 476, + 545, + 504 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {c} = \\operatorname {E n c} (I _ {c}), \\\\ F _ {s} = \\operatorname {E n c} \\left(I _ {s}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "f860e62852106b5a8a1555a004f33e5b55f1a620ca8c68d1e393e668e2ff1f4a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "text", + "content": "where the spatial scales for " + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "text", + "content": " are 1/8 of those for " + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "text", + "content": ". Then, the embedded feature " + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 306, + 514, + 546, + 539 + ], + "type": "text", + "content": " is derived by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 369, + 548, + 545, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 548, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 369, + 548, + 545, + 562 + ], + "type": "interline_equation", + "content": "F _ {c s} = \\operatorname {S t y l e T r a n s} \\left(F _ {s}, F _ {c}\\right), \\tag {2}", + "image_path": "97218442b987b4b10cd802cfc8cccd416ffcb1224a49d6253b39f87e9ecbca91.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "type": "text", + "content": "where StyleTrans denotes the Style Transformer module, and " + }, + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "type": "text", + "content": " has the same shape as " + }, + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "type": "text", + "content": ". Finally, we can synthesize the stylized image " + }, + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 305, + 571, + 545, + 608 + ], + "type": "text", + "content": " with the decoder as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 391, + 618, + 545, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 618, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 391, + 618, + 545, + 631 + ], + "type": "interline_equation", + "content": "I _ {c s} = \\operatorname {D e c} \\left(F _ {c s}\\right). \\tag {3}", + "image_path": "8a3b8d12e1ae6cc6438108e4304c2bc00cefd3b9402bfb95bd4588aebeef7b7a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 641, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 546, + 713 + ], + "type": "text", + "content": "Style Transformer. Similar to previous Transformer-based style transfer models like StyTr2 [4], the Transformer encoder is used to encode style information while the Transformer decoder takes charge of content-style interaction. In our model, there is only one copy of parameters shared by all Transformer encoder and decoder layers. Also, different" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18331" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "from standard Transformer where decoder layers would follow all encoder layers, encoder and decoder layers would be executed in an alternate fashion. The next layer would take the output of the current layer as input." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 205 + ], + "type": "text", + "content": "Our Transformer decoder layer is composed of self-attention, cross-attention, and non-linear blocks, for content encoding and content-style interaction. Specifically, content features are first processed with a self-attention step. Then, cross-attention is conducted by taking the content encoding as query and the style encoding as key and value, followed by a MLP for non-linear transformation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "type": "text", + "content": "Notably, in vanilla Transformer, features before and after the cross-attention step are fused by a residual connection, which is harmful for content structures as analyzed in Fig. 2. We thereby replace the residual connection with dynamic and learnable scaling and shifting steps, whose parameters are determined by the style encoder. In this way, the output of style encoder should consist of three parts: key for the following cross-attention " + }, + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "type": "text", + "content": ", scaling parameters " + }, + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "type": "inline_equation", + "content": "V_{\\sigma}" + }, + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "type": "text", + "content": ", and shifting parameters " + }, + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "type": "inline_equation", + "content": "V_{\\mu}" + }, + { + "bbox": [ + 46, + 206, + 287, + 350 + ], + "type": "text", + "content": ". The prediction of each part shares a same self-attention map to save memory but uses independent non-linear transformation. The process in the Transformer encoder layer can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 361, + 237, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 361, + 237, + 376 + ], + "spans": [ + { + "bbox": [ + 59, + 361, + 237, + 376 + ], + "type": "interline_equation", + "content": "\\mathrm {M H A} (Q, K, V) = [ h e a d _ {1}, \\dots , h e a d _ {h} ] W ^ {O},", + "image_path": "2a9557b34fe194a5e43696a1a9916c8a556bc902fc3884f5ec422d3ee7cc6bb1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 377, + 261, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 377, + 261, + 392 + ], + "spans": [ + { + "bbox": [ + 102, + 377, + 261, + 392 + ], + "type": "interline_equation", + "content": "\\operatorname {h e a d} _ {i} = \\operatorname {A t t} \\left(Q W ^ {Q _ {i}}, K W ^ {K _ {i}}, V W ^ {V _ {i}}\\right),", + "image_path": "bdd85ed9ac95bbd4c37a51378477a8924ae00ce7e42cd3a0789d2c6ec0f91444.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 394, + 221, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 394, + 221, + 422 + ], + "spans": [ + { + "bbox": [ + 70, + 394, + 221, + 422 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t} (Q, K, V) = \\operatorname {s o f t m a x} (\\frac {Q K ^ {\\top}}{\\sqrt {d _ {k}}} V),", + "image_path": "4202b78db31acc344e4737918c430f54676ec8addb0c89f4d5559152a493abdf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 423, + 287, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 423, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 113, + 423, + 287, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} K _ {s} ^ {\\prime} = K _ {s} + \\operatorname {M H A} \\left(K _ {s}, K _ {s}, K _ {s}\\right), \\\\ \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i g u r a t i o n} \\quad \\text {H a r m i n f i l t e} \\end{array} \\tag {4}", + "image_path": "dce11d20bd31def37fe3e214e4cd133dcf1ef7e40fb0523950c0f11e75b39b81.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 438, + 211, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 438, + 211, + 451 + ], + "spans": [ + { + "bbox": [ + 114, + 438, + 211, + 451 + ], + "type": "interline_equation", + "content": "K _ {s} = K _ {s} ^ {\\prime} + \\operatorname {M L P} \\left(K _ {s} ^ {\\prime}\\right),", + "image_path": "44680bb8c0da39a2b08760f5a8b366500c865245e31774b74fd29be90ac7f059.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 453, + 244, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 453, + 244, + 466 + ], + "spans": [ + { + "bbox": [ + 115, + 453, + 244, + 466 + ], + "type": "interline_equation", + "content": "V _ {\\sigma} ^ {\\prime} = V _ {\\sigma} + \\operatorname {M H A} \\left(K _ {s}, K _ {s}, V _ {\\sigma}\\right),", + "image_path": "c7089d6ea72b48a8728950691cbe4665fb0cd33435d159b519a4232814a33f66.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 468, + 208, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 468, + 208, + 481 + ], + "spans": [ + { + "bbox": [ + 116, + 468, + 208, + 481 + ], + "type": "interline_equation", + "content": "V _ {\\sigma} = V _ {\\sigma} ^ {\\prime} + \\operatorname {M L P} \\left(V _ {\\sigma} ^ {\\prime}\\right),", + "image_path": "b0ddb16db4759177dff2faf6e21b29c3fbd27d46c7a595cf7951844ccb7b6bbd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 483, + 244, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 483, + 244, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 483, + 244, + 497 + ], + "type": "interline_equation", + "content": "V _ {\\mu} ^ {\\prime} = V _ {\\mu} + \\mathrm {M H A} \\left(K _ {s}, K _ {s}, V _ {\\mu}\\right),", + "image_path": "3a700ace5f20d132c7ccfba8019fdd9b763408d4287187d42ee2685a0ca6c30e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 498, + 208, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 498, + 208, + 513 + ], + "spans": [ + { + "bbox": [ + 116, + 498, + 208, + 513 + ], + "type": "interline_equation", + "content": "V _ {\\mu} = V _ {\\mu} ^ {\\prime} + \\operatorname {M L P} \\left(V _ {\\mu} ^ {\\prime}\\right),", + "image_path": "c1a72f904e0b4d4002029f71d72dd5b6c8427562863ad2d55fbc75e2be165b0e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "type": "inline_equation", + "content": "K_{s}, V_{\\sigma}" + }, + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "type": "inline_equation", + "content": "V_{\\mu}" + }, + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "type": "text", + "content": " are initialized as the style feature " + }, + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 46, + 523, + 287, + 571 + ], + "type": "text", + "content": " before the first Transformer encoder layer. We do not incorporate normalization to encode style features since second-order statistics can largely represent style information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 573, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 632 + ], + "type": "text", + "content": "Then, in the cross-attention of Transformer decoder, scaling and shifting parameters for each content feature point is aggregated from " + }, + { + "bbox": [ + 46, + 573, + 287, + 632 + ], + "type": "inline_equation", + "content": "V_{\\sigma}" + }, + { + "bbox": [ + 46, + 573, + 287, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 573, + 287, + 632 + ], + "type": "inline_equation", + "content": "V_{\\mu}" + }, + { + "bbox": [ + 46, + 573, + 287, + 632 + ], + "type": "text", + "content": " respectively according to the cross-attention map. The process in one Transformer decoder block can be written as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 91, + 642, + 235, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 642, + 235, + 656 + ], + "spans": [ + { + "bbox": [ + 91, + 642, + 235, + 656 + ], + "type": "interline_equation", + "content": "F _ {c s} ^ {\\prime} = F _ {c s} + \\mathrm {M H A} \\left(F _ {c s}, F _ {c s}, F _ {c s}\\right),", + "image_path": "4623159d79fdaa1266900dd6b5c7a3a64420ed003706bdcd2a84f631e8b84bd5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 100, + 657, + 241, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 657, + 241, + 671 + ], + "spans": [ + { + "bbox": [ + 100, + 657, + 241, + 671 + ], + "type": "interline_equation", + "content": "\\sigma = \\operatorname {M H A} \\left(\\operatorname {I N} \\left(F _ {c s} ^ {\\prime}\\right), \\operatorname {I N} \\left(K _ {s}\\right), V _ {\\sigma}\\right),", + "image_path": "f867e6ada1316dd6041f208c96c3b627a653e798a623b516d415bd84b4a3eea3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 100, + 673, + 287, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 673, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 100, + 673, + 287, + 686 + ], + "type": "interline_equation", + "content": "\\mu = \\mathrm {M H A} (\\mathrm {I N} \\left(F _ {c s} ^ {\\prime}\\right), \\mathrm {I N} \\left(K _ {s}\\right), V _ {\\mu}), \\tag {5}", + "image_path": "4e8181d724732bddf693b346c22a4276cfa3c750b3151bfb8487a7ae7c18dfa0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 93, + 687, + 176, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 687, + 176, + 700 + ], + "spans": [ + { + "bbox": [ + 93, + 687, + 176, + 700 + ], + "type": "interline_equation", + "content": "F _ {c s} ^ {\\prime \\prime} = \\sigma \\odot F _ {c s} ^ {\\prime} + \\mu ,", + "image_path": "593e65fba8bcfa135fe5476dffc4186a92323d03bc6dcfa9b797ded76b1191b9.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 93, + 702, + 195, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 702, + 195, + 715 + ], + "spans": [ + { + "bbox": [ + 93, + 702, + 195, + 715 + ], + "type": "interline_equation", + "content": "F _ {c s} = F _ {c s} ^ {\\prime \\prime} + \\operatorname {M L P} \\left(F _ {c s} ^ {\\prime \\prime}\\right),", + "image_path": "69f2d062c3ce5f4b9932d7278e96ee2037d82e22b0ec9daed75339505bc37ad5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 72, + 421, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 421, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 421, + 84 + ], + "type": "text", + "content": "Algorithm 1 Meta Training" + } + ] + } + ], + "index": 19, + "type": "text" + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "text", + "content": ": content dataset; " + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "text", + "content": ": style dataset; " + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "text", + "content": ": inner learning rate; " + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "text", + "content": ": outer learning rate; " + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "text", + "content": ": number of inner updates; " + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 87, + 545, + 123 + ], + "type": "text", + "content": ": maximal number of stacked layers;" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 123, + 491, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 123, + 491, + 135 + ], + "spans": [ + { + "bbox": [ + 306, + 123, + 491, + 135 + ], + "type": "text", + "content": "Ensure: trained meta generator parameters " + }, + { + "bbox": [ + 306, + 123, + 491, + 135 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 309, + 136, + 531, + 289 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 136, + 410, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 136, + 410, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 136, + 410, + 148 + ], + "type": "text", + "content": "1: initialize " + }, + { + "bbox": [ + 312, + 136, + 410, + 148 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 312, + 136, + 410, + 148 + ], + "type": "text", + "content": " randomly" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 148, + 433, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 148, + 433, + 159 + ], + "spans": [ + { + "bbox": [ + 312, + 148, + 433, + 159 + ], + "type": "text", + "content": "2: for iteration 1, 2, 3, ... do" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 159, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 159, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 159, + 471, + 171 + ], + "type": "text", + "content": "3: sample a style image " + }, + { + "bbox": [ + 312, + 159, + 471, + 171 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 312, + 159, + 471, + 171 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 312, + 159, + 471, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 172, + 368, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 172, + 368, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 172, + 368, + 182 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 312, + 172, + 368, + 182 + ], + "type": "inline_equation", + "content": "\\omega \\gets \\theta" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 183, + 399, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 183, + 399, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 183, + 399, + 194 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 312, + 183, + 399, + 194 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 312, + 183, + 399, + 194 + ], + "type": "text", + "content": " times do" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 195, + 531, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 531, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 531, + 207 + ], + "type": "text", + "content": "6: sample a batch of content image " + }, + { + "bbox": [ + 312, + 195, + 531, + 207 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 312, + 195, + 531, + 207 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 312, + 195, + 531, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 208, + 518, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 208, + 518, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 208, + 518, + 219 + ], + "type": "text", + "content": "7: sample the number of layers from " + }, + { + "bbox": [ + 312, + 208, + 518, + 219 + ], + "type": "inline_equation", + "content": "1\\sim T" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 220, + 493, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 493, + 231 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 493, + 231 + ], + "type": "text", + "content": "8: forward propagation using Eq. 1-5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 232, + 488, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 232, + 488, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 488, + 243 + ], + "type": "text", + "content": "9: compute inner loss " + }, + { + "bbox": [ + 312, + 232, + 488, + 243 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 312, + 232, + 488, + 243 + ], + "type": "text", + "content": " using Eq. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 309, + 243, + 416, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 243, + 416, + 254 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 416, + 254 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 309, + 243, + 416, + 254 + ], + "type": "inline_equation", + "content": "\\omega \\gets \\omega -\\delta \\nabla L" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 309, + 255, + 372, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 255, + 372, + 266 + ], + "spans": [ + { + "bbox": [ + 309, + 255, + 372, + 266 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 309, + 266, + 414, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 266, + 414, + 279 + ], + "spans": [ + { + "bbox": [ + 309, + 266, + 414, + 279 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 309, + 266, + 414, + 279 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta +\\eta (\\omega -\\theta)" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 309, + 280, + 357, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 280, + 357, + 289 + ], + "spans": [ + { + "bbox": [ + 309, + 280, + 357, + 289 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "content": "where IN denotes instance normalization [39] and " + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "content": " represents element-wise multiplication. " + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "content": " is initialized as " + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "content": " before the first Transformer decoder layer." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 359, + 409, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 359, + 409, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 359, + 409, + 372 + ], + "type": "text", + "content": "3.2. Training Pipeline" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 378, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 462 + ], + "type": "text", + "content": "To achieve high-quality style transfer, we introduce a two-stage training strategy that comprises meta training and fast adaptation. The meta training stage is designed to learn a generic model initialization, while the fast adaptation adapts the network for a single style in a few iterations for few-shot style transfer. Note that zero-shot style transfer is a special case of the overall training configuration." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": "Meta Training: Inspired by Reptile [30], a first-order meta learning algorithm, rendering style patterns of a specific reference style image can be viewed as a task. We seek an optimal initialization for neural networks in this stage, so that the networks can be rapidly adapted for a new task in only a few shots. The main training procedure for this stage is shown in Algorithm 1. In each iteration, we sample 1 style and " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": " batches of contents to perform inner optimization to obtain \"fast weights\" " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": ", which would later guide the update of \"slow weights\" " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": ", to move a step in the direction of " + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\omega - \\theta" + }, + { + "bbox": [ + 304, + 462, + 545, + 628 + ], + "type": "text", + "content": ". Notably, as there is only one group of parameters for Transformer encoder and decoder layers, we randomly choose a number as the number of stacked layers for the Style Transformer in each iteration." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Fast Adaptation: The trained model after the first stage would serve as an initialization and will be adapted for a single style. With the same objective, this stage behaves almost identically to the internal loop in Algorithm 1. The only difference is that only parameters of the Transformer encoder layer are necessary to be updated, since (1) the Transformer encoder layer is the main component to extract style pat" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18332" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "terns and has the most significant impact on stylization, and (2) it would save memory and speed up the adaptation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "type": "text", + "content": "Zero-Shot Style Transfer: \"Zero-Shot\" means that, once the meta training is done, there is no fast adaptation stage needed and the meta model itself can support arbitrary style transfer. To encourage the model to produce satisfactory results in the zero-shot setup, we set the inner optimization time " + }, + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "type": "text", + "content": " to 1, where the algorithm is reduced to the typical training paradigm of existing arbitrary style transfer methods. In this sense, Algorithm 1 provides a more general setting for style transfer in both zero-shot and few-shot cases." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 286, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 286, + 346 + ], + "type": "text", + "content": "Text-Guided Style Transfer. We perform text-guided style transfer based on our image style transfer model with slight modifications. Following the common practice, a pretrained CLIP encoder is imported to extract text features. To be consistent with image input, we use a StyleGAN-like mapping network to convert text features into pseudo image features. In the meta training stage, since CLIP unifies the feature space, we use image instead of text as style input to avoid additional text dataset. In the fast adaptation stage, the mapping network along with the Transformer encoder is updated, considering there is still a gap between CLIP image feature and text feature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 354, + 135, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 135, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 135, + 365 + ], + "type": "text", + "content": "3.3. Loss Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 286, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 286, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 286, + 444 + ], + "type": "text", + "content": "The training objective in both meta-training and fast adaptation stages follows many works in arbitrary style transfer, which consists of content loss and style loss. Let " + }, + { + "bbox": [ + 46, + 373, + 286, + 444 + ], + "type": "inline_equation", + "content": "F^x" + }, + { + "bbox": [ + 46, + 373, + 286, + 444 + ], + "type": "text", + "content": " represent features on ReLU-" + }, + { + "bbox": [ + 46, + 373, + 286, + 444 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 373, + 286, + 444 + ], + "type": "text", + "content": "-1 layer of a pre-trained VGG19 network [38] for loss computation. The content loss is defined by the normalized perceptual loss [17]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 445, + 286, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 445, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 91, + 445, + 286, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t}} = \\sum_ {x = 2} ^ {5} \\| \\mathrm {I N} \\left(F _ {c} ^ {x}\\right) - \\mathrm {I N} \\left(F _ {c s} ^ {x}\\right) \\| _ {2}, \\tag {6}", + "image_path": "439b33690b16182caeb481761c5cee9fd0fbb32d014d565e8dade30630cc02ef.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 483, + 272, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 483, + 272, + 495 + ], + "spans": [ + { + "bbox": [ + 49, + 483, + 272, + 495 + ], + "type": "text", + "content": "while the style loss adopts the mean-variance loss [13]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 497, + 286, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 497, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 49, + 497, + 286, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s t y} = \\sum_ {x = 2} ^ {5} \\left(\\| \\mu \\left(F _ {s} ^ {x}\\right) - \\mu \\left(F _ {c s} ^ {x}\\right) \\| _ {2} + \\| \\sigma \\left(F _ {s} ^ {x}\\right) - \\sigma \\left(F _ {c s} ^ {x}\\right) \\| _ {2}\\right), \\tag {7}", + "image_path": "b236481d4d11b9513866ceee8a917fd9cd62d9049728cb22cec15cd55e8c3f97.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 540, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 286, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 540, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 540, + 286, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 540, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 540, + 286, + 574 + ], + "type": "text", + "content": " calculate the channel-wise mean and standard deviation separately. The overall objective is given by the weighted summation of the two losses:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 124, + 578, + 286, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 578, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 124, + 578, + 286, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t}} + \\lambda \\mathcal {L} _ {\\text {s t y}}, \\tag {8}", + "image_path": "fdf22fa31301820248ce8322abde2785d4db122b30c81c8a9130a9a6a521dbc3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 244, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 244, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 244, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 244, + 604 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 594, + 244, + 604 + ], + "type": "text", + "content": " controls the balance between two terms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 640 + ], + "type": "text", + "content": "In text-guided style transfer, the loss functions are the same as CLIPstyler, including global CLIP loss, directional CLIP loss and PatchCLIP loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 651, + 127, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 127, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 127, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 671, + 179, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 179, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 179, + 683 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "content": "We use MS-COCO [23] as our content dataset and WikiArt test set [33] as our style dataset. The content" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 531, + 298 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 531, + 298 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 531, + 298 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 531, + 298 + ], + "type": "table", + "html": "
MethodLcontLsimLsty
AdaIN4.88±0.700.53±0.231.51±0.78
Linear3.93±0.760.44±0.171.97±0.96
AvatarNet5.76±0.530.56±0.193.19±1.82
SANet4.72±0.610.50±0.171.15±0.58
MANet4.93±0.570.50±0.171.41±0.77
MCCNet4.22±0.690.47±0.171.56±0.84
AdaAttN4.46±0.700.43±0.162.20±1.19
StyTr23.78±0.990.48±0.221.50±0.69
StyFormer4.94±0.790.43±0.152.20±1.44
MetaNet3.48±0.850.45±0.192.47±1.69
MetaStyle*3.64±1.120.42±0.182.47±1.06
Johnson†4.60±0.760.59±0.201.02±0.34
Ours-Vanilla5.50±0.630.59±0.260.85±0.38
Ours-Norm4.70±0.750.43±0.130.93±0.33
Ours-ZS-L14.13±0.680.41±0.140.92±0.40
Ours-ZS-L34.20±0.680.41±0.130.81±0.31
Ours-FS*4.24±0.820.38±0.120.79±0.25
", + "image_path": "4ac380fb9dd8fe54cac32269b9325afc0bb32536d34f11d12d33e05a2ebe8f99.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 300, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 365 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons. ZS and FS for our model denote zero-shot and few-shot modes. Vanilla denote replacing our architecture with original Transformer. Norm means adding layer normalization in the Transformer encoder layer. L1/L3 means using 1/3 Transformer layers in the test time. * and † denote few-shot and per-style-per-model methods." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "text", + "content": "dataset contains roughly 80,000 images and the style dataset has about 20,000 images. The optimizer is Adam [19] with learning rates of both inner and outer loops set as 0.0001 and the batch size is 4. In training, we first resize the content and style image to " + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "text", + "content": " and then randomly crop to " + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "text", + "content": " resolution. During inference, our model can handle inputs of any size. The update times for inner optimization " + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "text", + "content": " is set as 2 for few-shot case and 1 for zero-shot case. In training, the maximal number of stacked layers is 4. All the multi-head attention blocks are instantiated as shifted window attention in [29], with window size 8 and shift size 4. The model is trained on a Nvidia 3090 with 9k iterations in the meta training stage for convergence while only 100 steps for image input and 20 steps for text input in the fast adaptation for few-shot style transfer, which takes less than 1 minute. Hyper-parameter " + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 373, + 545, + 564 + ], + "type": "text", + "content": " is set as 10." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 574, + 471, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 471, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 471, + 586 + ], + "type": "text", + "content": "4.2. Comparison with Prior Works" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "In this section, we compare results by our Master with 13 state-of-the-art style transfer methods, including 3 global transformation based methods (AdaIN [13], Linear style transfer [21], and MCCNet [2]), 1 patch swap based method (Avatar-Net [37]), 3 attention based methods (SANet [31], MANet [3], and AdaAttN [24]), 2 transformer based methods (StyTr2 [4] and StyFromer [41]), 2 meta learning based methods (MetaNet [36] and MetaStyle [47]), 1 per-style-per-model method by Johnson et al. [17] and 1 text-guided style transfer method (CLIPstyler)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18333" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": "Quantitative Comparison. We adopt content loss " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cont}" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": " in Eq. 6 and style loss " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sty}" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": " in Eq. 7 as evaluation metrics to reflect effects of learning by different methods. We also design a metric " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sim}" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": " to reflect the preservation of the spatial-wise self cosine similarity of content structures:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 133, + 212, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 133, + 212, + 161 + ], + "spans": [ + { + "bbox": [ + 104, + 133, + 212, + 161 + ], + "type": "interline_equation", + "content": "D _ {*, i j} ^ {x} = 1 - \\frac {F _ {*, i} ^ {x} \\cdot F _ {*, j} ^ {x}}{\\| F _ {*, i} ^ {x} \\| \\| F _ {*, j} ^ {x} \\|}", + "image_path": "86869d5dbf55ac4ae5499b973ea478236aed66525c64af4ff97637eba5895a52.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 159, + 286, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 159, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 286, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s i m} = \\sum_ {x = 3} ^ {4} \\frac {1}{n _ {x} ^ {2}} \\sum_ {i, j} \\left| \\frac {D _ {c , i j} ^ {x}}{\\sum_ {k} D _ {c , k j} ^ {x}} - \\frac {D _ {c s , i j} ^ {x}}{\\sum_ {k} D _ {c s , k j} ^ {x}} \\right|, \\tag {9}", + "image_path": "65e1ab68f7ff4a7614780b7e2c3ab68c04a8d84d51981c234fa7226b4ee2f5fb.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 198, + 286, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 286, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 286, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 198, + 286, + 389 + ], + "type": "inline_equation", + "content": "n_x" + }, + { + "bbox": [ + 47, + 198, + 286, + 389 + ], + "type": "text", + "content": " is the number of spatial locations for the current feature map and the second foot script denotes the spatial index. Smaller " + }, + { + "bbox": [ + 47, + 198, + 286, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sim}" + }, + { + "bbox": [ + 47, + 198, + 286, + 389 + ], + "type": "text", + "content": " means that the original spatial-wise relationship is better preserved during style transfer, i.e., less content distortion. We use the test dataset in the code page of [13] for evaluation, with 11 content images and 20 style images to form 220 content-style pairs. We report the mean and standard deviation over the 220 cases in Tab. 1. Notably, the lowest style loss and content similarity loss are achieved by Master simultaneously compared with previous methods, even better than the per-style-per-model solution by Johnson et al. [17], and can be further reduced in the few-shot case with comparable content loss values, which demonstrates the joint advantages of our method in content preserving and style rendering. Meanwhile, the lower standard deviations demonstrate the robustness of our model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 390, + 286, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 286, + 426 + ], + "type": "text", + "content": "Qualitative Comparison. Qualitative examples by different methods are shown in Fig. 4. Full comparisons with more methods can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 426, + 286, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 286, + 485 + ], + "type": "text", + "content": "We first give discussion on the global-transformation based methods. As shown in the 3rd and 4th rows, in AdaIN and Linear, features in all positions share the same transformation matrix, which fails to migrate detail textures in style images and has poor color saturation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 486, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 286, + 581 + ], + "type": "text", + "content": "As for the attention-based methods, SANet in the 5th row introduces local focus to style images. However, due to the simple fusion of one-layer attention results and original content features, it brings textures and content distortion in many cases, e.g., background areas of the 4th and 5th columns. AdaAttN in the 6th row dedicates to addressing the distortion problem in SANet, but it sacrifices the ability to render salient style patterns and textures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 582, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 286, + 712 + ], + "type": "text", + "content": "We then compare our results with those by Transformer-based methods. StyFormer in the 7th row adopts a design of the cross-attention module in Transformer without exploring the effect of self-attention to enhance style rendering. Thus, the results appear under-stylized compared with those by our method. StyTr2 in the 8th row is a Transformer architecture with a pure self-attention mechanism, which makes the performance on the extraction and migration of local textures not satisfactory enough, e.g., strokes of the 2nd column and waves of the 4th column. Moreover, since it follows the design of residual connections in" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 543, + 677 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 543, + 677 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 543, + 677 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 543, + 677 + ], + "type": "image", + "image_path": "a5dabdc7252bdd13c46365be822eca167bb8b12a04afc528073f9d2fb5489fe9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 686, + 545, + 708 + ], + "lines": [ + { + "bbox": [ + 306, + 686, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 306, + 686, + 545, + 708 + ], + "type": "text", + "content": "Figure 4. Comparison with previous state-of-the-art style transfer methods. Zoom in for better details." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18334" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 74, + 287, + 323 + ], + "blocks": [ + { + "bbox": [ + 48, + 74, + 287, + 323 + ], + "lines": [ + { + "bbox": [ + 48, + 74, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 74, + 287, + 323 + ], + "type": "image", + "image_path": "1721071f92e1d26376da096c2d9d544732fa1a71d395d93c7abc1eb63b3f4852.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 287, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 287, + 354 + ], + "type": "text", + "content": "Figure 5. Comparison with CLIPstyler in text-guided style transfer. Zoom in for better details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 360, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 456 + ], + "type": "text", + "content": "vanilla Transformer, the issue of content distortion in Fig. 2 is still evident, as shown in the background areas of the 4th and 5th columns. Compared with these methods, Master renders more vivid global and local style patterns. Meanwhile, it well maintains content structures simultaneously with learnable scaling and shifting parameters. Furthermore, in the few-shot case, comparable performance can be achieved with only 1 Transformer layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "type": "text", + "content": "MetaNet and MetaStyle also involve the concept of meta learning into their style transfer frameworks. On the one hand, MetaNet predicts parameters of a generator network given one style image via a meta network, which is an ambitious goal and makes the training more difficult. Thus, as shown in the 9th row, the stylized effects of their results still appear weak. On the other hand, MetaStyle achieves inferior performance in terms of local details as demonstrated in the 10th row, since only parameters of global normalization layers are learnable in the fast adaptation stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": "The single style transfer method by Johnson et al. in the 11th row tends to arrange patterns learned by the network arbitrarily, which may produce undesired and distorted effects. By contrast, the complex content-style dependence is constructed in our Transformer model and thus achieves more remarkable performance on transferring style patterns to proper positions in content images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Some qualitative comparisons on text-guided style transfer with CLIPstyler are shown in Fig. 5, which uses per-text-per-model fashion. Through these results, we observe that despite significantly more training steps, CLIPstyler" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 340, + 79, + 519, + 149 + ], + "blocks": [ + { + "bbox": [ + 340, + 79, + 519, + 149 + ], + "lines": [ + { + "bbox": [ + 340, + 79, + 519, + 149 + ], + "spans": [ + { + "bbox": [ + 340, + 79, + 519, + 149 + ], + "type": "image", + "image_path": "ae3e93bf6b1addfd77baed52ef7ab4a7c4f0b911e56234fde9800a6353b26662.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 198, + 545, + 251 + ], + "blocks": [ + { + "bbox": [ + 305, + 162, + 545, + 195 + ], + "lines": [ + { + "bbox": [ + 305, + 162, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 305, + 162, + 545, + 195 + ], + "type": "text", + "content": "Figure 6. Results of user study. Left: Our zero-shot stylization results compared with SOTA methods. Right: Our few-shot stylization comparison results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 198, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 306, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 545, + 251 + ], + "type": "image", + "image_path": "5a50828409e302f2615288a1c83c3243beb6c2661eac1362b7bd05dac496bfc1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 254, + 530, + 265 + ], + "lines": [ + { + "bbox": [ + 320, + 254, + 530, + 265 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 530, + 265 + ], + "type": "text", + "content": "Figure 7. Ablation study on key designs in Master model." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 317 + ], + "type": "text", + "content": "still suffers from weak stylization. By contrast, our method generates more vivid results in only 20 adaptation steps, which demonstrates the advantages of the proposed Master architecture and the versatility of our training pipeline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": "User Study. Following most style transfer works, we conduct user study and report user preference. We choose 20 content images and 20 style images to form 400 content-style pairs. We involve 100 people and randomly assign 20 stylized results from compared methods to each subject. Our method showcases zero-shot results with 3 Transformer layers in 10 tests and few-shot results in the remaining 10. For each pair, the order of results is randomized, and participants choose their favorite. With 2,000 votes in total, Fig. 6 demonstrates Master's superior style transfer quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": "Architecture. We conduct ablation studies on key designs in Master to illustrate their effects on the stylization quality. As shown in Fig. 7, on the one hand, if vanilla Transformer model is used, without learnable scaling and shifting parameters, noisy textures can be introduced significantly, which distorts the original content structures and affects style transfer quality. On the other hand, if layer normalization operations in the standard Transformer are used in the style encoder, style patterns would become less saturated, since normalization removes second-order feature statistics which represents the style information to a large degree. Quantitative studies in Tab. 1 also support our analysis and come to the same conclusion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Meta Training and Fast Adaptation. We provide training loss visualization in Fig. 8, to demonstrate the effectiveness of our meta training and fast adaptation algorithm. Our meta training stage learns an appropriate initial state, which enables fast decent and convergence of loss values in the fast adaptation stage. We show the same visualization in the per-style-per-model setting as a benchmark, using the same loss function and network architecture. As shown in Fig. 8," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18335" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 260, + 207 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 260, + 207 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 260, + 207 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 260, + 207 + ], + "type": "image", + "image_path": "c18a13ed7d1b1d1a8d60849660f554dee74e4b644a5f92bebaf63eb8e21614ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 287, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 264 + ], + "type": "text", + "content": "Figure 8. Training loss visualization during meta training and fast adaptation of three transformer based models: StyTr2, StyFormer, and our Master. The loss curve under per-style-per-model setting by our model is used as a reference. Ours-wo-Fix means that all the parameters need to be updated during the fast adaptation stage." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 353 + ], + "type": "text", + "content": "it requires roughly 3k iterations for the model to be convergent in the per-style-per-model case (Single Style), while the meta model (Ours) after 9k iterations can be adapted for any style image in only a few shots with competitive training results. Thus, the total number of iterations for our method is significantly smaller than per-style-per-model ones with the growth number of required styles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "text", + "content": "By default, all the parameters except those in the style encoder module are fixed during the fast adaptation stage. We also experiment with training the whole model in this stage (Ours wo Fix in Fig. 8) and find that the training and convergence would become difficult and require more time and memory. Moreover, updating all the parameters may result in insufficient adaptation given a limited number of training steps. As demonstrated in Fig. 7, there is a gap on the global tone between the result and the style image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 461, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 605 + ], + "type": "text", + "content": "Base Model. We evaluate StyFormer and StyTr2 as base models, presenting loss visualizations. Heavier Transformers raise training difficulty for earlier models, resulting in inferior convergence. When the number of Transformer layers is 5, StyTr2 even fails to converge. By contrast, our model adopts parameter-sharing across all Transformer layers, which results in an overall light-weight structure. Consequently, it enjoys better training effects in the meta training stage as shown in Fig. 8. Moreover, as shown in the zoom-in part, Master also enables overall lower and more stable training in fast adaptation. Qualitative examples by different base models can be found in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Controllable Style Transfer. We compare with some widely-adopted approaches for the similar goal to control the degree of stylization, including (1) Recursion, which treats stylized results as input of content images in the next iteration, and (2) Linear Mixup, which conducts linear combinations between content features and stylized features before the decoder. Vanilla Transformer is adopted for these approaches. The major difference between Recursion and ours is that our method takes the iterative sty" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 533, + 232 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 533, + 232 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 533, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 533, + 232 + ], + "type": "image", + "image_path": "4a3aef678cececb99a53609e418444464b1b2a588529b74f8e8784f6a6283a0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "type": "text", + "content": "Figure 9. Comparisons with widely-adopted approaches to control the degree of stylization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 258, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 366 + ], + "type": "text", + "content": "ization into consideration in the training time so that the model would learn to transfer style patterns layer-by-layer. Thus, as shown in Fig. 9, Recursion merely uses more intensive colors with the increasing of repeat times. Also, for Linear Mixup, it is typically hard to generate reasonable style transfer results especially when the weight of contents is high. Compared with them, our method is capable of rendering more style patterns and increasing the artistic abstraction when more Transformer layers are stacked." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 376, + 378, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 378, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 378, + 388 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 647 + ], + "type": "text", + "content": "In this paper, we propose a novel Transformer model specifically for artistic style transfer, termed as Meta Style Transformer (Master). On one hand, parameters of different Transformer layers are shared, which reduces the total number of parameters significantly and thus enables easier convergence. It is also convenient for Master to control the degree of stylization via customizing the number of stacked layers in the inference time. On the other hand, different from standard Transformer, our model adopts dynamic and learnable scaling and shifting operations instead of original residual connections, which helps preserve similarity relationship in content structures while migrating remarkable style patterns. Specifically, Master is trained using a meta learning algorithm for few-shot style transfer. Only parameters of the Transformer encoder layer need to be updated in the few-shot stage, which benefits the fast and robust adaptation. Zero-shot arbitrary style transfer is a special case of the training configuration. Experiments suggest that Master outperforms previous state-of-the-art arbitrary style transfer methods on both content preserving and style rendering in the zero-shot case." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 658, + 398, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 398, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 398, + 670 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "This project is supported by the Singapore Ministry of Education Academic Research Fund Tier 1 (WBS: A-0009440-01-00)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18336" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Yingying Deng, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, and Changsheng Xu. Arbitrary video style transfer via multi-channel correlation. arXiv preprint arXiv:2009.08003, 2020. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "text", + "content": "[3] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2719-2727, 2020. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "text", + "content": "[4] Yingying Deng, Fan Tang, Xingjia Pan, Weiming Dong, Chongyang Ma, and Changsheng Xu. " + }, + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "inline_equation", + "content": "stytr^2" + }, + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "text", + "content": ": Unbiased image style transfer with transformers, 2021. 1, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 368 + ], + "type": "text", + "content": "[6] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "text", + "content": "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "text", + "content": "[8] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "text", + "content": "[10] Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. Arbitrary style transfer with deep feature reshuffle. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8222–8231, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "text", + "content": "[11] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 635 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "text", + "content": "[13] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "text", + "content": "[14] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 14861-14869, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "text", + "content": "[15] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4369-4376, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[16] Yongcheng Jing, Yining Mao, Yiding Yang, Yibing Zhan, Mingli Song, Xinchao Wang, and Dacheng Tao. Learning graph neural networks for image style transfer. In ECCV, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 241 + ], + "type": "text", + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 1, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 242, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 275 + ], + "type": "text", + "content": "[18] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 354 + ], + "type": "text", + "content": "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "text", + "content": "[21] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast arbitrary style transfer. arXiv preprint arXiv:1808.04537, 2018. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 388, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 422 + ], + "type": "text", + "content": "[22] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. arXiv preprint arXiv:1705.08086, 2017. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 422, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 478 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "text", + "content": "[24] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6649-6658, October 2021. 1, 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "text", + "content": "[25] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. In Conference on Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "text", + "content": "[26] Songhua Liu, Hao Wu, Shoutong Luo, and Zhengxing Sun. Stable video style transfer based on partial convolution with depth-aware supervision. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2445-2453, 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[27] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "text", + "content": "[28] Songhua Liu, Jingwen Ye, Runpeng Yu, and Xinchao Wang. Slimmable dataset condensation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18337" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "type": "text", + "content": "[30] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms, 2018. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5880–5888, 2019. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "text", + "content": "[32] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "type": "text", + "content": "[33] Fred Phillips and Brandy Mackintosh. Wiki art gallery, inc.: A case for critical thinking. Issues in Accounting Education, 26(3):593-608, 2011. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "type": "text", + "content": "[36] Falong Shen, Shuicheng Yan, and Gang Zeng. Meta networks for neural style transfer, 2017. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "text", + "content": "[37] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8242-8250, 2018. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 509 + ], + "type": "text", + "content": "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "text", + "content": "[39] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "type": "text", + "content": "[41] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[42] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[43] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Conference on Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 229 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[44] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1467-1475, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[45] Jingwen Ye, Songhua Liu, and Xinchao Wang. Partial network cloning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[46] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[47] Chi Zhang, Yixin Zhu, and Song-Chun Zhu. Metastyle: Three-way trade-off among speed, flexibility, and quality in neural style transfer, 2019. 3, 5" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18338" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_content_list.json b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..08febde3328c2319c6dcb61e22b8545decd874c1 --- /dev/null +++ b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_content_list.json @@ -0,0 +1,1651 @@ +[ + { + "type": "text", + "text": "Matching Is Not Enough: A Two-Stage Framework for Category-Agnostic Pose Estimation", + "text_level": 1, + "bbox": [ + 225, + 130, + 743, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Min Shi $^{1,2,*}$ Zihao Huang $^{1,*}$ Xianzheng Ma $^{2}$ Xiaowei Hu $^{2}$ Zhiguo Cao $^{1,\\dagger}$ $^{1}$ School of AIA, Huazhong University of Science and Technology \n $^{2}$ Shanghai AI Laboratory", + "bbox": [ + 161, + 208, + 803, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{min_shi, zihaohuang, zgcao}@hust.edu.cn, {maxianzheng, huxiaowei}@pjlab.org.cn", + "bbox": [ + 133, + 266, + 825, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 316, + 313, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Category-agnostic pose estimation (CAPE) aims to predict keypoints for arbitrary categories given support images with keypoint annotations. Existing approaches match the keypoints across the image for localization. However, such a one-stage matching paradigm shows inferior accuracy: the prediction heavily relies on the matching results, which can be noisy due to the open set nature in CAPE. For example, two mirror-symmetric keypoints (e.g., left and right eyes) in the query image can both trigger high similarity on certain support keypoints (eyes), which leads to duplicated or opposite predictions. To calibrate the inaccurate matching results, we introduce a two-stage framework, where matched keypoints from the first stage are viewed as similarity-aware position proposals. Then, the model learns to fetch relevant features to correct the initial proposals in the second stage. We instantiate the framework with a transformer model tailored for CAPE. The transformer encoder incorporates specific designs to improve the representation and similarity modeling in the first matching stage. In the second stage, similarity-aware proposals are packed as queries in the decoder for refinement via cross-attention. Our method surpasses the previous best approach by large margins on CAPE benchmark MP-100 on both accuracy and efficiency. Code available at github.com/flyinglynx/CapeFormer", + "bbox": [ + 75, + 349, + 473, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans can quickly grasp the essentials of a keypoint for arbitrary objects (e.g., the nose of an animal) and then pinpoint the same keypoint on another object in the same category. However, most pose estimation models are still category-specific, which cannot be applied to a new category unless they are trained with sufficient category-", + "bbox": [ + 75, + 782, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba94522789afd8436d3a534b5462e9eb7ffab4494e9cfd7fc56f804c76778680.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 315, + 885, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c8ed89129cf639a07d41f0c75eb1a7b99aaa02c95276caf0232300b712a50a3.jpg", + "image_caption": [ + "Figure 1. Comparison with existing methods. (a) Compared with previous one-stage matching paradigm that entirely relies on keypoint matching, we propose a novel two-stage framework, which learns to correct unreliable matching results in the second stage. (b) The proposed approach attains better accuracy and efficiency compared with the previous best approach." + ], + "image_footnote": [], + "bbox": [ + 508, + 479, + 875, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "specific data. Recently, category-agnostic pose estimation (CAPE) [40] is proposed to address such a generic pose estimation problem. CAPE aims to localize instance keypoints for arbitrary categories given one or few support images with keypoint annotations (support keypoints).", + "bbox": [ + 496, + 731, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike category-specific pose estimation models that learn to recognize specific keypoints for a fixed category, CAPE models learn to represent and compare keypoints for open-world categories. As shown in Fig. 1 (a), existing CAPE approaches [40] match the support keypoints across the query image in an embedding space. Then, the match-", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: Equal contribution. †: Corresponding author.", + "bbox": [ + 99, + 886, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7308", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing results are used as clues for keypoint localization. Although a sophisticated feature embedding pipeline is elaborated [40], we find that such a one-stage pipeline shows inferior accuracy and efficiency for CAPE tasks. Specifically, the prediction heavily relies on matching results. However, due to the open set nature in CAPE, the similarity results often contain some noises: 1) insufficient similarity when support and query instances differ significantly in poses, textures, or styles; 2) false positive similarity is triggered when points in the query image share similar appearance with support keypoints. For example, one-stage matching paradigm has difficulties distinguishing the left and right of keypoints due to similar visual characteristics. Thus, the left eye in support keypoints can be indiscriminately predicted to the right eye in the query image.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we remedy these issues from two aspects. First, to refine the matching results, we introduce a two-stage framework for CAPE, where matched keypoints in the first stage are viewed as similarity-aware position proposals. The model learns to refine the proposals in the second stage. In addition, to construct a robust matching process, we also improve the representation quality and similarity metrics.", + "bbox": [ + 75, + 321, + 472, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we instantiate the two-stage framework with a transformer, termed CAPE transformer (CapeFormer). The transformer encoder encodes the support keypoints and query images in the first stage, with specific designs to improve the representation quality. We first design a query-support joint refine encoder to mutually transfer information among support keypoints and the query images, thus narrowing the gap between the support and query instances (poses or styles). Then, we add a support keypoint identifier to alleviate the ambiguity between two support keypoints when they are geometrically close or with similar appearance. After the feature encoding, similarity-aware position proposals are generated for the second stage. We directly match the support keypoint with query image features using a learnable inner-product [34]. The similarity peaks are selected as the similarity-aware position proposals for support keypoints. In the second stage, the position proposals are packed with support keypoint features as the queries in the transformer decoder. In each decoder layer, queries extract relevant features with cross-attention, which are used as clues to update the proposals.", + "bbox": [ + 75, + 429, + 472, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on a category-agnostic pose estimation benchmark MP-100 [40]. Our method outperforms the previous best approach POMNet [40] by large margins, with an improvement of $5.6\\%$ and $8.6\\%$ under 1-shot and 5-shot settings, respectively. When training and test categories almost have no properties in common, a more significant improvement (up to $10.5\\%$ ) can be obtained compared with POMNet. Note that, our method is trained end-to-end and shows better efficiency compared with previous best approach, as in Fig. 1 (b).", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 517, + 90, + 846, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a two-stage framework for category-agnostic pose estimation tackling the noisy matching results.", + "- We instantiate the two-stage framework with a transformer model, termed CapeFormer, with specific designs to enhance the representation and similarity modeling in the matching pipeline.", + "- CapeFormer outperforms the previous best approach significantly in both accuracy and efficiency." + ], + "bbox": [ + 517, + 111, + 893, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 273, + 640, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 297, + 660, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pose estimation aims to localize the semantic keypoints of instances. Most pose estimation approaches are targeted for one specific category, e.g., human [1, 19, 45, 46], animals [3, 17, 20], or vehicles [29, 36]. According to how points are localized, pose estimation methods can be categorized into regression-based [18, 21, 28, 37, 38], heatmap-based [4, 44], offset-based [12] and query-based [33] ones. In this work, we address category-agnostic pose estimation, which has not been well-studied yet. Compared with category-specific pose estimation, category-agnostic pose estimation focuses on constructing a generic representation and similarity metrics: keypoints are predicted by comparing the support keypoints and the query images in the embedding space.", + "bbox": [ + 496, + 321, + 893, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Category-Agnostic Vision Models", + "text_level": 1, + "bbox": [ + 500, + 544, + 794, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Category-agnostic vision models aim to build a generic model that can be applied to any categories on a particular task, e.g., object detection [14], segmentation [43], object counting and localization [24, 34]. This setting is also known as the few-shot learning. Recently, POMNet [40] first proposes the category-agnostic pose estimation task and elaborates a dataset MP-100. Most category-agnostic models can be classified into meta-learning and metric-learning approaches. Meta-learning approaches [11] aim to initialize the model with a set of optimal parameters, where the model can quickly fit the new tasks by learning from a few support samples. Metric learning-based approaches match the support samples with query images in an embedding space. Then, the matching results are used as clues for subsequent tasks. For pose estimation, POMNet [40] encodes query images and support keypoints with a transformer. Then, a regression head directly infers the similarity from the concatenated support keypoint and query image features. We find that simply using matching results for point localization is challenging. Hence, we extend such a matching paradigm with a two-stage framework to correct the unreliable matching results for accurate predictions.", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7309", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9e963a7607f3a087c4daa28ca0973cc5c1f08a2bdbae363c315cdb737ca54e22.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed CapeFormer. CapeFormer consists of a matching stage where similarity-aware position proposals are generated, and a refine stage that further calibrates the proposals. A support identifier and query-support joint refine encoder are proposed to improve the representation quality." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 883, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Semantic Correspondence", + "text_level": 1, + "bbox": [ + 76, + 380, + 316, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic correspondence (SC) [8, 15] aims to match the points that have identical semantic meaning with a dense correspondence field. Hence, SC models can be directly applied on the CAPE: localize the corresponding points for the support keypoints with the estimated correspondence field. Typical SC models first match each pair of pixels and use the matching score to construct correspondence. Similar to our two-stage design that alleviates the noise on similarity matching, a series of SC models refine the matching scores with geometry constraints [26], high-dimensional convolutions [31], and attention mechanism [6, 7]. However, we find directly applying SC models for CAPE is inefficient and yields inferior performance. Besides, most techniques for the matching refine in SC are based on the dense matching volume, which is not available in CAPE, as only the sparse support keypoints are matched.", + "bbox": [ + 75, + 404, + 473, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. CAPE Transformer", + "text_level": 1, + "bbox": [ + 76, + 654, + 269, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-agnostic pose estimation (CAPE) aims to estimate instance keypoints given one or a few support images with keypoint annotations. Specifically, we denote the query image as $I_{q}$ and the support image as $I_{s}$ , which has $K$ predefined support keypoints. The CAPE model will infer the positions of the corresponding $K$ keypoints in the query images. For simplicity, we use the 1-shot setting for illustration (the number of support images equals one).", + "bbox": [ + 75, + 680, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Two-Stage Framework for CAPE", + "text_level": 1, + "bbox": [ + 76, + 804, + 372, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the previous one-stage paradigm, Cape-Former features a two-stage framework. As in Fig. 2, the first stage consists of the input embedding, a support-query joint refine encoder, and a similarity-aware proposal generator. A proposal refine decoder forms the second stage.", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input embedding encodes the query image and support keypoints into the embedding space (Sec. 3.2). We denote the features for the query image and support keypoints as $F_{q}$ and $F_{s}$ , which are further refined with the query-support joint refine encoder (Sec. 3.3). The proposal generator measures the similarity between $F_{s}$ and $F_{q}$ and the similarity peaks are selected as the similarity-aware position proposals (Sec. 3.4). Then, in the second stage, the similarity-aware proposals are gradually refined into the final predictions (Sec. 3.5).", + "bbox": [ + 496, + 381, + 893, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Input Embedding", + "text_level": 1, + "bbox": [ + 498, + 544, + 671, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input embedding is illustrated in Fig. 3. We first project the support and query image into feature maps with a shared backbone. The flattened feature map of the query image is used as its representation, denoted by $F_{q} \\in \\mathbb{R}^{c \\times n}$ , where $n$ equals the number of pixels in the feature map and $c$ denotes the embedded dimension. To obtain the features of the support keypoint, we conduct a \"soft\" ROI pooling [30] on the support image feature map. $K$ soft masks are generated by placing a Gaussian kernel centered at support keypoints, and then multiplied with the support feature map. The summation of masked support feature map is adopted as the support keypoint feature $F_{s} \\in \\mathbb{R}^{K \\times c}$ . For multiple support images, e.g., 5-shot settings, the support keypoint features from different images are averaged.", + "bbox": [ + 496, + 566, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Support keypoint identifier. When encoding support keypoints, each support keypoint's positional and context information is discarded during ROI pooling. Hence, when two support keypoints are geometrically close or with similar appearance (e.g., left and right eyes), they can be indistinguishable in the embedding space. Such ambiguity can lead to inaccurate predictions that are duplicated or mirror-symmetric to the ground truth.", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7310", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f3db0a27cd08c797aedf302aafb97a14adf813dbb4972189e33f669638f5ca65.jpg", + "image_caption": [ + "Figure 3. Illustration of the input embedding and the support-query joint refine encoder." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 310, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a6a8220e256e82432ffb0c951708c315268af0fe7a2fcc638cb9884e7e826fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 90, + 464, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate such ambiguity, we propose a simple but effective support keypoint identifier that gives an \"identity\" to each support keypoint. The implementation is embarrassingly simple: $K$ embedding vectors are generated via sinusoidal encoding, denoted by $I_{s} \\in \\mathbb{R}^{K \\times c}$ . This identifier is added to $F_{s}$ ahead of attention calculation, which will be discussed later in this section. Such a simple modification can lead to over 3% improvement with negligible computation overhead.", + "bbox": [ + 75, + 375, + 472, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Query-Support Refine Encoder", + "text_level": 1, + "bbox": [ + 76, + 532, + 354, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Directly matching the support keypoints with the query image right after the input embedding is challenging, as instances in the support and query images can be different, such as different colors, textures, and poses. Thus, the target and support keypoints may not possess sufficient similarity in the embedding space [10, 34]. An intuitive way to alleviate this is to fuse relevant query and support features. Hence, the query and support features can be pulled closer in the embedding space.", + "bbox": [ + 75, + 560, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we implement information fusion and transfer among support and query with self-attention. As illustrated in Fig. 3, the inputs of the $l^{th}$ encoder layer are query and support features, denoted as $F_{q}^{l}$ and $F_{s}^{l}$ . We augment the $F_{q}^{l}$ with positional embedding as in DETR [5], and add the support keypoint identifier into the $F_{s}^{l}$ . The augmented $F_{q}^{l}$ and $F_{s}^{l}$ are concatenated into a feature sequence with the length of $n + K$ . The sequence is then fed into the standard multi-head self-attention layer [39]. The output sequence from the self-attention layer is processed by a feed forward network (FFN). Then, the sequence are re-split to the support keypoint and query image features, which are used as the inputs for the next encoder layer.", + "bbox": [ + 75, + 700, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Similarity-Aware Proposal Generator", + "text_level": 1, + "bbox": [ + 500, + 90, + 823, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The similarity-aware proposal generator matches the support keypoint features with the query features to obtain the similarity maps, whose peaks are selected as the similarity-aware proposals. Previous state-of-the-art approach POMNet [40] measures similarity implicitly: a regression head directly infers similarity maps (heatmaps) from the concatenation of support keypoint and query features. However, regressing similarity without explicit similarity modeling has the risk of overfitting in class-agnostic setting [34]. Moreover, considerable computation overhead is introduced. For example, given 100 support keypoint features with 256 channels as in our case, the similarity maps are obtained by forwarding 512-channel feature maps with an FCN for 100 times.", + "bbox": [ + 496, + 113, + 890, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To embrace both efficiency and generality, we use a learnable inner-product [34] to model similarity explicitly. Specifically, we first calculate a channel attention weight $a \\in \\mathbb{R}^{K \\times c}$ conditioned on $F_s \\in \\mathbb{R}^{K \\times c}$ . Then, the similarity map $M$ is obtained as:", + "bbox": [ + 496, + 324, + 890, + 401 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM = \\left(W _ {q} F _ {q}\\right) \\left(\\boldsymbol {a} \\circ W _ {s} F _ {s}\\right) ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 409, + 890, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{q}$ and $W_{s}$ are two learnable projection matrices in $\\mathbb{R}^{c\\times c}$ and $\\circ$ denotes the Hadamard product. Then, similarity maps $M\\in \\mathbb{R}^{n\\times K}$ are reshaped into 2D with the shape of $H\\times W$ . The channel attention helps stressing the key patterns during inner product computation [34].", + "bbox": [ + 496, + 438, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The maximums from the similarity maps are selected as the similarity-aware position proposals for support keypoints. To reduce the round-off errors on the down-scaled resolutions, instead of directly selecting the maximums, we first generate a $3 \\times 3$ local window around the similarity peaks and normalize this local window with softmax. Then, we take the expectation of the coordinates in each local window as the proposal, which can be formulated as:", + "bbox": [ + 496, + 513, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {k} = \\sum_ {i, j \\in W _ {k}} i \\cdot \\frac {\\exp (M _ {k , i j})}{\\sum_ {i , j \\in W} \\exp (M _ {i j})}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 656, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {k} = \\sum_ {i, j \\in W _ {k}} j \\cdot \\frac {\\exp (M _ {k , i j})}{\\sum_ {i , j \\in W} \\exp (M _ {i j})}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 696, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{k}$ denotes the local window for the $k^{th}$ support keypoint and $i, j$ indexes the horizontal and vertical coordinates. Here we denote the similarity-aware position proposals for all the support keypoints as $P \\in \\mathbb{R}^{K \\times 2}$ .", + "bbox": [ + 496, + 747, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Proposal Refine Decoder", + "text_level": 1, + "bbox": [ + 500, + 816, + 725, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned in Sec. 1, due to the unreliable matching process, directly using the matched keypoints from the first stage as predictions is prone to error. Intuitively, each support keypoint can be corrected by jointly considering other", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7311", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5bf9a158d390d6d1573f62937d46ec31a518925e7d268599ce64393d6c375625.jpg", + "image_caption": [ + "Figure 4. Illustration of the proposal refine decoder. The decoder consists of a self-attention and a cross-attention layer" + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 428, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "support keypoint predictions and keypoint-related information in the query image. Hence, we devise the second stage to unleash the potential of correcting the raw similarity-aware position proposals.", + "bbox": [ + 75, + 401, + 468, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 4, we design a proposal refine decoder, where the support keypoint embedding together with the keypoint positions are fed into the decoder layer. Each layer consists of two steps: 1) self-attention among support keypoints makes each proposal aware of other keypoints' positions and contents; 2) cross-attention that helps extract relevant content from the query feature for each support keypoint. By integrating these two attention mechanisms, the decoder can further refine the position proposals and produce more accurate predictions.", + "bbox": [ + 75, + 462, + 468, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-attention among support keypoints. Self-attention enables interactions among the support keypoints, which makes each proposal aware of other keypoints' positions and contents. To prepare the inputs for the self-attention layer, the support keypoint positions are first converted into sinusoidal positional encoding $E_{s}^{l} \\in \\mathbb{R}^{K \\times c}$ . Then, the support keypoint identifier $I_{s}$ and position embedding $E_{s}^{l}$ are added into $F_{s}^{l}$ forming the query and key for the first self-attention layer. The support keypoint feature $F_{s}^{l}$ is directly used as the value. The multi-head self-attention layer is followed by a transformer-style feed-forward network (FFN) with residual connection and layer normalization [2]. We denote the support feature processed by the FFN as $F_{s}^{l'}$ .", + "bbox": [ + 75, + 613, + 468, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-attention between support and query features. Cross-attention helps extract relevant contents from the query feature $F_{q}$ for each support keypoint. As illustrated in Fig. 4, we use the concatenation of $F_{s}^{\\prime}$ and the summation of $I_{s}$ and $E_{s}^{l}$ as the query for the cross-attention layer. Correspondingly, query image features $F_{q}$ and positional em", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bedding $P_{q}$ are concatenated as the key, and $F_{q}$ are used as the value. Instead of addition, we concatenate the positional embedding and support identifier for key and value, following DAB-DETR [23]. An FFN processes the output from the cross-attention layer to form the support keypoint features for the next decoder layer, i.e., $F_{s}^{l + 1}$ .", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Position update. We use the refined support feature $F_{s}^{l + 1}$ to update the keypoint positions $P^l$ to $P^{l + 1}$ . An MLP takes $F_{s}^{l + 1}$ as input and infers the offset $O^{l + 1}\\in \\mathbb{R}^{K\\times 2}$ pointing to the ground truth. Then, the keypoint positions are updated by:", + "bbox": [ + 496, + 181, + 890, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {l + 1} = \\sigma \\left(O ^ {l} + \\sigma^ {- 1} \\left(P ^ {l}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 267, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ and $\\sigma^{-1}$ denote the sigmoid and its inverse function. Note that the similarity-aware proposals are used as the keypoint positions for the first decoder layer, i.e., $P^1$ . The updated keypoint positions from the last decoder layer are used as the final prediction.", + "bbox": [ + 496, + 297, + 890, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Supervision Signal", + "text_level": 1, + "bbox": [ + 500, + 382, + 679, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CapeFormer incorporates two supervision signals: the similarity loss and the offset loss. The similarity loss facilitates the learning of representation and the similarity metric in the first stage. The offset loss supervises the learning of proposal refinement.", + "bbox": [ + 496, + 405, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarity loss. The similarity loss is imposed on the similarity maps $M$ from the similarity-aware proposal generator. We simultaneously constrain the shape of similarity map and the accuracy of similarity-aware proposals with a heatmap term $\\mathcal{L}_h$ and an expectation term $\\mathcal{L}_e$ .", + "bbox": [ + 496, + 479, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For heatmap loss, we use sigmoid to normalize similarity map $M$ . Then, $\\mathcal{L}_h$ are computed with $\\ell_2$ loss as:", + "bbox": [ + 496, + 556, + 890, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {1}{H \\cdot W} \\| \\text {s i g m o i d} (M _ {i}) - H _ {i} \\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 598, + 890, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{i}$ denotes the ground truth heatmap for the $i^{th}$ keypoint. For $\\mathcal{L}_e$ , we first normalize the similarity map with softmax and then take the expectation of the coordinates as:", + "bbox": [ + 496, + 651, + 890, + 709 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE \\left(P _ {k}\\right) = \\sum_ {j = 1} ^ {H} \\sum_ {i = 1} ^ {W} \\frac {\\exp \\left(M _ {k , i j}\\right)}{\\sum \\exp \\left(M _ {k}\\right)} \\cdot (i, j), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 709, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E(P_k)$ denotes the coordinate expectation for the $k^{th}$ keypoint. $H$ and $W$ denote the height and width for similarity map. Then we use $\\ell_1$ loss to calculate $\\mathcal{L}_e$ as:", + "bbox": [ + 496, + 757, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\left| E \\left(P _ {i}\\right) - \\hat {P} _ {i} \\right| _ {1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 815, + 890, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{P}_i$ denotes the ground truth coordinate of the $i^{th}$ point.", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7312", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dc617a87cd9bf815751701d2c9c224f84606be1356e13b62dc708c6edf015d21.jpg", + "table_caption": [ + "Table 1. Results on the MP-100 dataset. Performance (PCK) under 1-shot and 5-shot settings. The proposed CapeFormer achieves the best PCK on all the splits under 1-shot and 5-shot settings." + ], + "table_footnote": [], + "table_body": "
Method1-shot5-shot
Split 1Split 2Split 3Split 4Split 5AverageSplit 1Split 2Split 3Split 4Split 5Average
ProtoNet [35]46.0540.8449.1343.3444.5444.7860.3153.5161.9258.4458.6158.56
MAML [11]68.1454.7264.1963.2457.2061.5070.0355.9863.2164.7958.4762.50
Fine-tune [27]70.6057.0466.0665.0059.2063.5871.6757.8466.7666.5360.2464.61
POMNet [40]84.2378.2578.1778.6879.1779.7084.7279.6178.0080.3880.8580.71
CapeFormer89.4584.8883.5983.5385.0985.3191.9488.9289.4088.0188.2589.30
", + "bbox": [ + 89, + 122, + 877, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Offset regression loss. Supposing a total of $L$ decoder layers are used, we define the offset regression loss $\\mathcal{L}_o$ as:", + "bbox": [ + 75, + 253, + 470, + 286 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\sum_ {i = 1} ^ {K} \\left| P _ {i} ^ {l + 1} - \\hat {P} _ {i} \\right| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 295, + 468, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall training loss is calculated as $\\lambda_h \\cdot \\mathcal{L}_h + \\mathcal{L}_e + \\mathcal{L}_o$ , where $\\lambda_h$ is a pre-defined weight for the heatmap loss.", + "bbox": [ + 76, + 347, + 468, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 392, + 209, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 416, + 295, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network architecture. We use the ResNet-50 [13] pretrained from ImageNet [9] as the backbone following POMNet [40]. Different from POMNet, the backbone is shared by query and support images in our cases for efficiency. Both the query-support joint refine encoder and the proposal refine decoder have three layers. Refer to supplementary for more details on network architectures.", + "bbox": [ + 75, + 439, + 468, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training strategy. We use Adam [16] optimizer to train the model for 200 epochs with the batch size of 16. The learning rate is set as $1e^{-5}$ and decays by $10 \\times$ on the $160^{th}$ and $180^{th}$ epoch. The weight for heatmap loss is empirically set as 2.0 to make different loss terms lie in the same magnitude. Data augmentation and pre-processing are kept the same with POMNet [40] for fair comparisons.", + "bbox": [ + 75, + 547, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and metric. We use MP-100 dataset [40] to train and evaluate our method. Samples are collected from the existing category-specific pose estimation datasets [22, 32, 41]. MP-100 dataset encompasses over 18k images from 100 different categories. Keypoint numbers of different categories vary from 8 to 68. To train and evaluate the model, the samples are organized into five splits. On each split, the training, validation, and test categories have no overlaps, i.e., the categories for evaluation are not accessed during training. We use the Probability of Correct Keypoint (PCK) [42] as the quantitative metric. The threshold for PCK is set to 0.2 following POMNet [40].", + "bbox": [ + 75, + 656, + 470, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on MP-100", + "text_level": 1, + "bbox": [ + 76, + 847, + 259, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with the previous best CAPE method POMNet [40] and three baselines: ProtoNet [35],", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/005199e105a2abf5589f9c540e07b240a3e9659290296f93c08d83b92c88964a.jpg", + "table_caption": [ + "Table 2. Cross super-category evaluation on MP-100 dataset. Experiments are conducted under the 1-shot setting." + ], + "table_footnote": [], + "table_body": "
MethodHuman BodyHuman FaceVehicleFurniture
ProtoNet [35]37.6157.8028.3542.64
MAML [11]51.9325.7217.6820.09
Fine-tune [27]52.1125.5317.4620.76
POMNet [40]73.8279.6334.9247.27
CapeFormer83.4480.9645.4052.49
", + "bbox": [ + 501, + 286, + 890, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MAML [11], and Fine-tune [27]. PCKs of different approaches on the MP-100 dataset under 1-shot and 5-shot settings are reported in Table 1, where we can see our method significantly outperforms POMNet [40], and improves the average PCK by $5.6\\%$ under the 1-shot setting and $8.6\\%$ under the 5-shot setting.", + "bbox": [ + 496, + 407, + 890, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1-shot CapeFormer can surpass other methods with five support images, demonstrating the robustness of the proposed two-stage framework when the support information is limited. When increasing the shots from one to five, CapeFormer obtains a significant improvement of $4.0\\%$ on PCK. For POMNet, the improvement from 1-shot to 5-shot is $1.0\\%$ . This shows that CapeFormer can better utilize the additional information when more support samples are available. Note that we use the same way as POMNet [40] to encode the support keypoint features for fair comparisons. We also include a comparison with the semantic correspondence approaches in the supplementary.", + "bbox": [ + 496, + 500, + 892, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Cross Super-Category Pose Estimation", + "text_level": 1, + "bbox": [ + 498, + 694, + 834, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the MP-100 dataset, training, validation, and test categories do not overlap. However, some categories may still share similar characteristics, such as the faces of different animals. To better evaluate the generalization ability on significantly different categories, we conduct a cross-supercategory evaluation following POMNet [40]. Four supercategories: human face, human body, vehicle, and furniture, are collected from the MP-100 dataset as the test categories, while the other categories are used as the training samples.", + "bbox": [ + 496, + 717, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As in Table 2, CapeFormer achieves the state-of-the-art performance. The advantages on the human body and vehicle are particularly remarkable, with an improvement", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7313", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29308e8337988c5ac90f05e95e6f46b8c56bf2d4a4405b1312a5f1bfdf253317.jpg", + "image_caption": [ + "Figure 5. Qualitative results of our method. We visualize the keypoint predictions under 1-shot setting. The left column denotes the support image with corresponding keypoints. The middle column denotes the results without proposal refine decoder, and the right column are results from the full method." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 885, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of $9.62\\%$ and $10.48\\%$ on PCK, respectively. Compared with the standard setting, the improvement in cross supercategory setting is more significant, which further demonstrates the generalization ability and robustness of the proposed method. The improvement of the human face is limited $(1.36\\%)$ . A plausible reason is that the human face has much more keypoints (68) than the training categories (the largest keypoint number in training is 38). Such discrepancy renders great difficulties for generalization: the keypoints are densely arranged, and the training data does not contain such a fine-grained matching case.", + "bbox": [ + 75, + 450, + 472, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 630, + 230, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we first ablate the two-stage framework and the proposed modules in Table 3. Then, we validate some detailed design choices, including the implementation of training loss and support keypoint identifier. We perform experiments on the test set of MP-100 split1 under the 1-shot setting following [40]. We set some variants of the encoder for ablations on network components. (i) DETR [5] encoder removes the query-support interactions, i.e., the information fusion among support keypoints and query images is disabled; (ii) QSR denotes the proposed query-support joint refine encoder.", + "bbox": [ + 75, + 654, + 468, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two-stage framework. First we focus on the effectiveness of the second stage. Comparing No. 3 and No. 5 in Table 3, the decoder brings an improvement of $4.13\\%$ on PCK, demonstrating the necessity of the second stage.", + "bbox": [ + 75, + 824, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare one-stage and two-stage paradigms qualita-", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/599076dd8674ea8dbc2e88a3728cfced3f00773ee02305cef15ef806b17f7051.jpg", + "table_caption": [ + "Table 3. Ablation study on different components in Cape-Former. Support ID stands for the support keypoint identifier in Sec. 3.2. QSR denotes our query-support joint refine encoder." + ], + "table_footnote": [], + "table_body": "
Support IDEncoderDecoderParadigmPCK
No.1--DETR--one-stage80.32
No.2--QSR--one-stage82.86
No.3✓QSR--one-stage85.32
No.4--QSR✓two-stage85.81
No.5✓QSR✓two-stage89.45
", + "bbox": [ + 506, + 494, + 883, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tively in Fig. 5. \"One-Stage\" denotes the direct matching results from No. 3 setting in Table 3, here and after. The direct matching results are prone to error when query and support instances are different, e.g., the dog on the right part of the first row. We also visualize how the decoder corrects the predictions in Fig. 6. For the alpaca face, the inner left canthus (red point) is close to the annotation of the right inner canthus (yellow point) due to the occlusion. Hence, the similarity-aware position proposal for the left inner canthus is wrongly generated at the right inner canthus. The decoder gradually pushes it to the correct position.", + "bbox": [ + 496, + 622, + 890, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Representation. Here we analyze the effectiveness of query-support joint refine encoder and the support keypoint identifier. i) Comparing No.1 with No.2 in Table 3, the QSR encoder brings an absolute improvement of $2.54\\%$ on PCK over DETR encoder, which suggests the effectiveness of support-query mutual information fusion. 2) Comparing No.2 with No.3 in Table 3, introducing the support iden", + "bbox": [ + 496, + 795, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7314", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49655a490a2fc6a8b7f47d8b0e479d4b7a84423acdc3f3a2f6c55247ad36eb47.jpg", + "image_caption": [ + "Support Image" + ], + "image_footnote": [], + "bbox": [ + 106, + 102, + 218, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b8f5e90a1c83c3ae768f86ef5ce50be6820ef883b1ad9b2177ad7d9c712ea30.jpg", + "image_caption": [ + "Figure 6. The decoder can help correct unreliable proposals. The left column shows the support keypoint (red point) and the right column shows the refining process in the second stage. Note that in the first row, the left inner canthus is invisible due to occlusion, which looks to be on the right." + ], + "image_footnote": [], + "bbox": [ + 135, + 186, + 187, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5926882196251118edc0e7c495bba333263908760c7d441f7bcc445067669b55.jpg", + "image_caption": [ + "Query Image" + ], + "image_footnote": [], + "bbox": [ + 232, + 102, + 334, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee5e1e0fb09c78a3f0542a202893398dfb18b151f32513b116ab76901f010a3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 184, + 310, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/149c9adb2e71a4a7c378bd6b8b44a82e8de76413e78ce8929fa25c4f4f3b9a52.jpg", + "image_caption": [ + "Figure 7. Visualization of attention map. We visualize the encoder and decoder attention for the support keypoint at the left inner canthus. The support keypoint coarsely attends to keypoints with similar appearance in the encoder while focusing more on the corresponding keypoints for position update in the decoder." + ], + "image_footnote": [], + "bbox": [ + 334, + 122, + 431, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e04505fc44fc6d67e18b4ebb08f2f0ffa01b2f278aff78eb3dde28622c343c57.jpg", + "table_caption": [ + "Table 4. Analysis of different terms in the similarity loss." + ], + "table_footnote": [], + "table_body": "
LhLePCK
✓--88.74
--✓89.41
✓✓89.45
", + "bbox": [ + 84, + 371, + 246, + 443 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7a4f0a1b31b9c4385002c5a3f5cd03cdc0d891be3c3cab5f4ba6104fd2c38e25.jpg", + "table_caption": [ + "Table 5. Comparison of different support keypoint identifiers." + ], + "table_footnote": [], + "table_body": "
TypePCK
--85.81
Learnable88.96
Fixed89.45
", + "bbox": [ + 290, + 371, + 439, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tifier improves the PCK by $2.46\\%$ . Note that the No. 3 model excludes the decoder, which is one-stage. Its performance can still achieve state-of-the-art on split1 test set, which suggests that the support identifier and QSR decoder significantly improve representation quality.", + "bbox": [ + 75, + 455, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To intuitively show how the encoder and decoder work, we visualize the encoder/decoder attention maps in Fig. 7. The attention shows which location the support keypoint will focus during the information fusion (first stage) and the proposal correction (second stage). As can be observed, in the encoder, the support keypoint attends to keypoints with similar appearance, while in the decoder, the attention becomes more focused on the corresponding keypoints for proposal correction.", + "bbox": [ + 75, + 532, + 468, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss function. We analyze the similarity loss in Table 4. i) Only using the heatmap supervision (row 1) yields the worst performance, indicating that the expectation of loss is beneficial for precise localization. ii) As shown in row 3, the combination of heatmap loss and expectation attains the best performance, as both the shape of similarity map and the proposal position can be supervised.", + "bbox": [ + 75, + 671, + 467, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Support keypoint identifier. We compare two different implementations of support keypoint identifier: learnable embedding and fixed sinusoidal embedding in Table 5. Using fixed sinusoidal embedding improves PCK by $0.49\\%$ compared with learnable embedding. The learnable embedding is similar to the learnable query embedding in DETR-based detection methods [5,25], which can be viewed as a learnable content embedding or anchors [23]. However, our", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/571c800b7eee5e00e91b5c70f8dae652f57ede0ba2af04b4939fa5204e838e43.jpg", + "image_caption": [ + "Support Image &Keypoint" + ], + "image_footnote": [], + "bbox": [ + 529, + 90, + 638, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/868daeb6585332de87de355dfa487dd85070ee3761b2bdef7802a95cfee5014a.jpg", + "image_caption": [ + "Encoder Attention" + ], + "image_footnote": [], + "bbox": [ + 640, + 90, + 746, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/021f9e381df4d11926700c097dd789fdeb9f28c6e215a050f0aafbc4735f254c.jpg", + "image_caption": [ + "Decoder Cross-Attention" + ], + "image_footnote": [], + "bbox": [ + 750, + 90, + 857, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f93d7c6753a77c5d2757695e15f8fbfe1e284f9e43758e91fe58eddbe409623b.jpg", + "table_caption": [ + "Table 6. Inference and Training Efficiency. We test the efficiency on RTX3090 under 1-shot setting." + ], + "table_footnote": [], + "table_body": "
MethodParams (M)GFLOPsFPSMemory (GB)
POMNet [40]48.2138.016.8013.8 × 2
One-Stage26.8622.6536.907.5
CapeFormer31.1423.6826.097.8
", + "bbox": [ + 503, + 316, + 885, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "support keypoint identifier is fixed and serves a different purpose: distinguishing different support keypoints. Learnable support identifier deteriorates the performance as they may overfit the training categories more easily.", + "bbox": [ + 496, + 400, + 890, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Performance and Efficiency Trade-off", + "text_level": 1, + "bbox": [ + 498, + 470, + 826, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the efficiency under 1-shot setting for POMNet and CapeFormer in Table 6. The training memory is evaluated with a batchsize of 16. i) Comparing row 1 and 3, CapeFormer significantly reduces the parameters $(-35.4\\%)$ and FLOPs $(-37.7\\%)$ compared with POMNet due to the shared backbone and efficient similarity metric. ii) Comparing rows 2 and 3, the second stage only introduces limited parameters and FLOPs, which shows a satisfying trade-off between performance and efficiency. iii) CapeFormer is efficient for training, which consumes only 8GB memory on a single GPU (almost $1/4$ of POMNet).", + "bbox": [ + 496, + 494, + 890, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 676, + 617, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we formulate a two-stage framework for category-agnostic pose estimation (CAPE) and instantiate it with a novel transformer model. Unlike previous one-stage approaches that entirely rely on matching results, the matching results in the first stage are viewed as similarity-aware position proposals which are refined in the second stage. We also propose specific designs to improve the representation and similarity modeling. On the CAPE benchmark MP-100, our method outperforms the previous best method by large margins on both accuracy and efficiency.", + "bbox": [ + 496, + 700, + 890, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (Grant No. U1913602 and 61876211).", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7315", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter V. Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 3686-3693, 2014. 2", + "[2] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv Computer Research Repository, abs/1607.06450, 2016. 5", + "[3] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9497-9506, 2019. 2", + "[4] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1):172-186, 2018. 2", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of European Conference on Computer Vision (ECCV), pages 213-229, 2020. 4, 7, 8", + "[6] Seokju Cho, Sunghwan Hong, Sangryul Jeon, Yunsung Lee, Kwanghoon Sohn, and Seungryong Kim. Cats: Cost aggregation transformers for visual correspondence. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 9011-9023, 2021. 3", + "[7] Seokju Cho, Sunghwan Hong, and Seungryong Kim. Cats++: Boosting cost aggregation with convolutions and transformers. arXiv Computer Research Repository, abs/2202.06817, 2022. 3", + "[8] Christopher B. Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Krishna Chandraker. Universal correspondence network. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 2406-2414, 2016. 3", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 248-255, 2009. 6", + "[10] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 4", + "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of International Conference on Machine Learning (ICML), pages 1126–1135, 2017. 2, 6", + "[12] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 14671-14681, 2021. 2", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 770-778, 2016. 6" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Shuqiang Jiang, Sisi Liang, Chengpeng Chen, Yaohui Zhu, and Xiangyang Li. Class agnostic image common object detection. IEEE Trans. Image Process., 28(6):2836-2846, 2019. 2", + "[15] Seungryong Kim, Dongbo Min, Bumsub Ham, Stephen Lin, and Kwanghoon Sohn. FCSS: fully convolutional self-similarity for dense semantic correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):581-595, 2019. 3", + "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of International Conference on Learning Representations, 2014. 6", + "[17] Rollyn Labuguen, Jumpei Matsumoto, Salvador Blanco Negrete, Hiroshi Nishimaru, Hisao Nishijo, Masahiko Takada, Yasuhiro Go, Ken ichi Inoue, and Tomohiro Shibata. Macaquepose: A novel \"in the wild\" macaque monkey pose dataset for markerless motion capture. Frontiers in behavioral neuroscience (2021), 2021. 2", + "[18] Jiefeng Li, Siyuan Bian, Ailing Zeng, Can Wang, Bo Pang, Wentao Liu, and Cewu Lu. Human pose regression with residual log-likelihood estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 11005–11014, 2021. 2", + "[19] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 10863-10872, 2018. 2", + "[20] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: A benchmark for amur tiger re-identification in the wild. In Proceedings of ACM International Conference on Multimedia, pages 2590–2598, 2020. 2", + "[21] Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Tokenpose: Learning keypoint tokens for human pose estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 11293-11302, 2021. 2", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), pages 740-755, 2014. 6", + "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In Proceedings of International Conference on Learning Representations, 2022. 5, 8", + "[24] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Proceedings of Asian Conference on Computer Vision (ACCV), pages 669–684, 2019. 2", + "[25] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3631-3640, 2021. 8", + "[26] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel flow: Semantic correspondence with multi-layer" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7316", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural features. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 3394-3403, 2019. 3", + "[27] Akihiro Nakamura and Tatsuya Harada. Revisiting finetuning for few-shot learning. arXiv Computer Research Repository, abs/1910.00216, 2019. 6", + "[28] Xuecheng Nie, Jianfeng Zhang, Shuicheng Yan, and Jiashi Feng. Single-stage multi-person pose machines. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 6950-6959, 2019. 2", + "[29] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 1906-1915, 2018. 2", + "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 3", + "[31] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomás Pajdla, and Josef Sivic. Neighbourhood consensus networks. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 1658-1669, 2018. 3", + "[32] Christos Sagonas, Epameinondas Antonakos, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: database and results. Image Vis. Comput., 47:3-18, 2016. 6", + "[33] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 11059-11068, 2022. 2", + "[34] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 9529–9538, 2022. 2, 4", + "[35] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4077-4087, 2017. 6", + "[36] Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5452-5462, 2019. 2", + "[37] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 2621-2630, 2017. 2", + "[38] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 1653-1660, 2014. 2", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 4", + "[40] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Proceedings of European Conference on Computer Vision (ECCV), volume abs/2207.10387, 2022. 1, 2, 4, 6, 7, 8", + "[41] Shuo Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5525-5533, 2016. 6", + "[42] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2878-2890, 2013. 6", + "[43] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5217–5226, 2019. 2", + "[44] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 7091-7100, 2020. 2", + "[45] Song-Hai Zhang, Ruilong Li, Xin Dong, Paul L. Rosin, Zixi Cai, Han Xi, Dingcheng Yang, Haozhi Huang, and Shi-Min Hu. Pose2seg: Detection free human instance segmentation. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 889-898, 2019. 2", + "[46] Jian Zhao, Jianshu Li, Yu Cheng, Li Zhou, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In Proceedings of ACM International Conference on Multimedia, pages 792-800, 2018. 2" + ], + "bbox": [ + 501, + 92, + 893, + 611 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7317", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_model.json b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cbd78c0b6a96bc77783242ff57ea485d2f74a836 --- /dev/null +++ b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.226, + 0.131, + 0.745, + 0.177 + ], + "angle": 0, + "content": "Matching Is Not Enough: A Two-Stage Framework for Category-Agnostic Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.209, + 0.804, + 0.265 + ], + "angle": 0, + "content": "Min Shi\\(^{1,2,*}\\) Zihao Huang\\(^{1,*}\\) Xianzheng Ma\\(^{2}\\) Xiaowei Hu\\(^{2}\\) Zhiguo Cao\\(^{1,\\dagger}\\) \n\\(^{1}\\) School of AIA, Huazhong University of Science and Technology \n\\(^{2}\\) Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.267, + 0.826, + 0.283 + ], + "angle": 0, + "content": "{min_shi, zihaohuang, zgcao}@hust.edu.cn, {maxianzheng, huxiaowei}@pjlab.org.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.317, + 0.314, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.474, + 0.729 + ], + "angle": 0, + "content": "Category-agnostic pose estimation (CAPE) aims to predict keypoints for arbitrary categories given support images with keypoint annotations. Existing approaches match the keypoints across the image for localization. However, such a one-stage matching paradigm shows inferior accuracy: the prediction heavily relies on the matching results, which can be noisy due to the open set nature in CAPE. For example, two mirror-symmetric keypoints (e.g., left and right eyes) in the query image can both trigger high similarity on certain support keypoints (eyes), which leads to duplicated or opposite predictions. To calibrate the inaccurate matching results, we introduce a two-stage framework, where matched keypoints from the first stage are viewed as similarity-aware position proposals. Then, the model learns to fetch relevant features to correct the initial proposals in the second stage. We instantiate the framework with a transformer model tailored for CAPE. The transformer encoder incorporates specific designs to improve the representation and similarity modeling in the first matching stage. In the second stage, similarity-aware proposals are packed as queries in the decoder for refinement via cross-attention. Our method surpasses the previous best approach by large margins on CAPE benchmark MP-100 on both accuracy and efficiency. Code available at github.com/flyinglynx/CapeFormer" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.759, + 0.21, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Humans can quickly grasp the essentials of a keypoint for arbitrary objects (e.g., the nose of an animal) and then pinpoint the same keypoint on another object in the same category. However, most pose estimation models are still category-specific, which cannot be applied to a new category unless they are trained with sufficient category-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.316, + 0.887, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.48, + 0.877, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.632, + 0.895, + 0.716 + ], + "angle": 0, + "content": "Figure 1. Comparison with existing methods. (a) Compared with previous one-stage matching paradigm that entirely relies on keypoint matching, we propose a novel two-stage framework, which learns to correct unreliable matching results in the second stage. (b) The proposed approach attains better accuracy and efficiency compared with the previous best approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.894, + 0.809 + ], + "angle": 0, + "content": "specific data. Recently, category-agnostic pose estimation (CAPE) [40] is proposed to address such a generic pose estimation problem. CAPE aims to localize instance keypoints for arbitrary categories given one or few support images with keypoint annotations (support keypoints)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Unlike category-specific pose estimation models that learn to recognize specific keypoints for a fixed category, CAPE models learn to represent and compare keypoints for open-world categories. As shown in Fig. 1 (a), existing CAPE approaches [40] match the support keypoints across the query image in an embedding space. Then, the match-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.355, + 0.901 + ], + "angle": 0, + "content": "*: Equal contribution. †: Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7308" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "ing results are used as clues for keypoint localization. Although a sophisticated feature embedding pipeline is elaborated [40], we find that such a one-stage pipeline shows inferior accuracy and efficiency for CAPE tasks. Specifically, the prediction heavily relies on matching results. However, due to the open set nature in CAPE, the similarity results often contain some noises: 1) insufficient similarity when support and query instances differ significantly in poses, textures, or styles; 2) false positive similarity is triggered when points in the query image share similar appearance with support keypoints. For example, one-stage matching paradigm has difficulties distinguishing the left and right of keypoints due to similar visual characteristics. Thus, the left eye in support keypoints can be indiscriminately predicted to the right eye in the query image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.473, + 0.429 + ], + "angle": 0, + "content": "In this paper, we remedy these issues from two aspects. First, to refine the matching results, we introduce a two-stage framework for CAPE, where matched keypoints in the first stage are viewed as similarity-aware position proposals. The model learns to refine the proposals in the second stage. In addition, to construct a robust matching process, we also improve the representation quality and similarity metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.748 + ], + "angle": 0, + "content": "Specifically, we instantiate the two-stage framework with a transformer, termed CAPE transformer (CapeFormer). The transformer encoder encodes the support keypoints and query images in the first stage, with specific designs to improve the representation quality. We first design a query-support joint refine encoder to mutually transfer information among support keypoints and the query images, thus narrowing the gap between the support and query instances (poses or styles). Then, we add a support keypoint identifier to alleviate the ambiguity between two support keypoints when they are geometrically close or with similar appearance. After the feature encoding, similarity-aware position proposals are generated for the second stage. We directly match the support keypoint with query image features using a learnable inner-product [34]. The similarity peaks are selected as the similarity-aware position proposals for support keypoints. In the second stage, the position proposals are packed with support keypoint features as the queries in the transformer decoder. In each decoder layer, queries extract relevant features with cross-attention, which are used as clues to update the proposals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We evaluate our method on a category-agnostic pose estimation benchmark MP-100 [40]. Our method outperforms the previous best approach POMNet [40] by large margins, with an improvement of \\(5.6\\%\\) and \\(8.6\\%\\) under 1-shot and 5-shot settings, respectively. When training and test categories almost have no properties in common, a more significant improvement (up to \\(10.5\\%\\)) can be obtained compared with POMNet. Note that, our method is trained end-to-end and shows better efficiency compared with previous best approach, as in Fig. 1 (b)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.847, + 0.108 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.112, + 0.894, + 0.156 + ], + "angle": 0, + "content": "- We propose a two-stage framework for category-agnostic pose estimation tackling the noisy matching results." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.163, + 0.894, + 0.224 + ], + "angle": 0, + "content": "- We instantiate the two-stage framework with a transformer model, termed CapeFormer, with specific designs to enhance the representation and similarity modeling in the matching pipeline." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.229, + 0.894, + 0.261 + ], + "angle": 0, + "content": "- CapeFormer outperforms the previous best approach significantly in both accuracy and efficiency." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.112, + 0.894, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.274, + 0.642, + 0.29 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.299, + 0.661, + 0.315 + ], + "angle": 0, + "content": "2.1. Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.895, + 0.535 + ], + "angle": 0, + "content": "Pose estimation aims to localize the semantic keypoints of instances. Most pose estimation approaches are targeted for one specific category, e.g., human [1, 19, 45, 46], animals [3, 17, 20], or vehicles [29, 36]. According to how points are localized, pose estimation methods can be categorized into regression-based [18, 21, 28, 37, 38], heatmap-based [4, 44], offset-based [12] and query-based [33] ones. In this work, we address category-agnostic pose estimation, which has not been well-studied yet. Compared with category-specific pose estimation, category-agnostic pose estimation focuses on constructing a generic representation and similarity metrics: keypoints are predicted by comparing the support keypoints and the query images in the embedding space." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.795, + 0.562 + ], + "angle": 0, + "content": "2.2. Category-Agnostic Vision Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Category-agnostic vision models aim to build a generic model that can be applied to any categories on a particular task, e.g., object detection [14], segmentation [43], object counting and localization [24, 34]. This setting is also known as the few-shot learning. Recently, POMNet [40] first proposes the category-agnostic pose estimation task and elaborates a dataset MP-100. Most category-agnostic models can be classified into meta-learning and metric-learning approaches. Meta-learning approaches [11] aim to initialize the model with a set of optimal parameters, where the model can quickly fit the new tasks by learning from a few support samples. Metric learning-based approaches match the support samples with query images in an embedding space. Then, the matching results are used as clues for subsequent tasks. For pose estimation, POMNet [40] encodes query images and support keypoints with a transformer. Then, a regression head directly infers the similarity from the concatenated support keypoint and query image features. We find that simply using matching results for point localization is challenging. Hence, we extend such a matching paradigm with a two-stage framework to correct the unreliable matching results for accurate predictions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7309" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.885, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.895, + 0.365 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed CapeFormer. CapeFormer consists of a matching stage where similarity-aware position proposals are generated, and a refine stage that further calibrates the proposals. A support identifier and query-support joint refine encoder are proposed to improve the representation quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.381, + 0.317, + 0.399 + ], + "angle": 0, + "content": "2.3. Semantic Correspondence" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.474, + 0.648 + ], + "angle": 0, + "content": "Semantic correspondence (SC) [8, 15] aims to match the points that have identical semantic meaning with a dense correspondence field. Hence, SC models can be directly applied on the CAPE: localize the corresponding points for the support keypoints with the estimated correspondence field. Typical SC models first match each pair of pixels and use the matching score to construct correspondence. Similar to our two-stage design that alleviates the noise on similarity matching, a series of SC models refine the matching scores with geometry constraints [26], high-dimensional convolutions [31], and attention mechanism [6, 7]. However, we find directly applying SC models for CAPE is inefficient and yields inferior performance. Besides, most techniques for the matching refine in SC are based on the dense matching volume, which is not available in CAPE, as only the sparse support keypoints are matched." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.27, + 0.673 + ], + "angle": 0, + "content": "3. CAPE Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Class-agnostic pose estimation (CAPE) aims to estimate instance keypoints given one or a few support images with keypoint annotations. Specifically, we denote the query image as \\( I_{q} \\) and the support image as \\( I_{s} \\), which has \\( K \\) predefined support keypoints. The CAPE model will infer the positions of the corresponding \\( K \\) keypoints in the query images. For simplicity, we use the 1-shot setting for illustration (the number of support images equals one)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.805, + 0.373, + 0.822 + ], + "angle": 0, + "content": "3.1. Two-Stage Framework for CAPE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Different from the previous one-stage paradigm, Cape-Former features a two-stage framework. As in Fig. 2, the first stage consists of the input embedding, a support-query joint refine encoder, and a similarity-aware proposal generator. A proposal refine decoder forms the second stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.895, + 0.534 + ], + "angle": 0, + "content": "The input embedding encodes the query image and support keypoints into the embedding space (Sec. 3.2). We denote the features for the query image and support keypoints as \\( F_{q} \\) and \\( F_{s} \\), which are further refined with the query-support joint refine encoder (Sec. 3.3). The proposal generator measures the similarity between \\( F_{s} \\) and \\( F_{q} \\) and the similarity peaks are selected as the similarity-aware position proposals (Sec. 3.4). Then, in the second stage, the similarity-aware proposals are gradually refined into the final predictions (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.545, + 0.673, + 0.562 + ], + "angle": 0, + "content": "3.2. Input Embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.78 + ], + "angle": 0, + "content": "The input embedding is illustrated in Fig. 3. We first project the support and query image into feature maps with a shared backbone. The flattened feature map of the query image is used as its representation, denoted by \\( F_{q} \\in \\mathbb{R}^{c \\times n} \\), where \\( n \\) equals the number of pixels in the feature map and \\( c \\) denotes the embedded dimension. To obtain the features of the support keypoint, we conduct a \"soft\" ROI pooling [30] on the support image feature map. \\( K \\) soft masks are generated by placing a Gaussian kernel centered at support keypoints, and then multiplied with the support feature map. The summation of masked support feature map is adopted as the support keypoint feature \\( F_{s} \\in \\mathbb{R}^{K \\times c} \\). For multiple support images, e.g., 5-shot settings, the support keypoint features from different images are averaged." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Support keypoint identifier. When encoding support keypoints, each support keypoint's positional and context information is discarded during ROI pooling. Hence, when two support keypoints are geometrically close or with similar appearance (e.g., left and right eyes), they can be indistinguishable in the embedding space. Such ambiguity can lead to inaccurate predictions that are duplicated or mirror-symmetric to the ground truth." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7310" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.312, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.091, + 0.465, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.311, + 0.47, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Illustration of the input embedding and the support-query joint refine encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.473, + 0.512 + ], + "angle": 0, + "content": "To alleviate such ambiguity, we propose a simple but effective support keypoint identifier that gives an \"identity\" to each support keypoint. The implementation is embarrassingly simple: \\( K \\) embedding vectors are generated via sinusoidal encoding, denoted by \\( I_{s} \\in \\mathbb{R}^{K \\times c} \\). This identifier is added to \\( F_{s} \\) ahead of attention calculation, which will be discussed later in this section. Such a simple modification can lead to over 3% improvement with negligible computation overhead." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.534, + 0.355, + 0.55 + ], + "angle": 0, + "content": "3.3. Query-Support Refine Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Directly matching the support keypoints with the query image right after the input embedding is challenging, as instances in the support and query images can be different, such as different colors, textures, and poses. Thus, the target and support keypoints may not possess sufficient similarity in the embedding space [10, 34]. An intuitive way to alleviate this is to fuse relevant query and support features. Hence, the query and support features can be pulled closer in the embedding space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To this end, we implement information fusion and transfer among support and query with self-attention. As illustrated in Fig. 3, the inputs of the \\( l^{th} \\) encoder layer are query and support features, denoted as \\( F_{q}^{l} \\) and \\( F_{s}^{l} \\). We augment the \\( F_{q}^{l} \\) with positional embedding as in DETR [5], and add the support keypoint identifier into the \\( F_{s}^{l} \\). The augmented \\( F_{q}^{l} \\) and \\( F_{s}^{l} \\) are concatenated into a feature sequence with the length of \\( n + K \\). The sequence is then fed into the standard multi-head self-attention layer [39]. The output sequence from the self-attention layer is processed by a feed forward network (FFN). Then, the sequence are re-split to the support keypoint and query image features, which are used as the inputs for the next encoder layer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.825, + 0.108 + ], + "angle": 0, + "content": "3.4. Similarity-Aware Proposal Generator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.324 + ], + "angle": 0, + "content": "The similarity-aware proposal generator matches the support keypoint features with the query features to obtain the similarity maps, whose peaks are selected as the similarity-aware proposals. Previous state-of-the-art approach POMNet [40] measures similarity implicitly: a regression head directly infers similarity maps (heatmaps) from the concatenation of support keypoint and query features. However, regressing similarity without explicit similarity modeling has the risk of overfitting in class-agnostic setting [34]. Moreover, considerable computation overhead is introduced. For example, given 100 support keypoint features with 256 channels as in our case, the similarity maps are obtained by forwarding 512-channel feature maps with an FCN for 100 times." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.402 + ], + "angle": 0, + "content": "To embrace both efficiency and generality, we use a learnable inner-product [34] to model similarity explicitly. Specifically, we first calculate a channel attention weight \\(a \\in \\mathbb{R}^{K \\times c}\\) conditioned on \\(F_s \\in \\mathbb{R}^{K \\times c}\\). Then, the similarity map \\(M\\) is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.41, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\nM = \\left(W _ {q} F _ {q}\\right) \\left(\\boldsymbol {a} \\circ W _ {s} F _ {s}\\right) ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.515 + ], + "angle": 0, + "content": "where \\( W_{q} \\) and \\( W_{s} \\) are two learnable projection matrices in \\( \\mathbb{R}^{c\\times c} \\) and \\( \\circ \\) denotes the Hadamard product. Then, similarity maps \\( M\\in \\mathbb{R}^{n\\times K} \\) are reshaped into 2D with the shape of \\( H\\times W \\). The channel attention helps stressing the key patterns during inner product computation [34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.636 + ], + "angle": 0, + "content": "The maximums from the similarity maps are selected as the similarity-aware position proposals for support keypoints. To reduce the round-off errors on the down-scaled resolutions, instead of directly selecting the maximums, we first generate a \\(3 \\times 3\\) local window around the similarity peaks and normalize this local window with softmax. Then, we take the expectation of the coordinates in each local window as the proposal, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.657, + 0.891, + 0.695 + ], + "angle": 0, + "content": "\\[\nx _ {k} = \\sum_ {i, j \\in W _ {k}} i \\cdot \\frac {\\exp (M _ {k , i j})}{\\sum_ {i , j \\in W} \\exp (M _ {i j})}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.698, + 0.891, + 0.737 + ], + "angle": 0, + "content": "\\[\ny _ {k} = \\sum_ {i, j \\in W _ {k}} j \\cdot \\frac {\\exp (M _ {k , i j})}{\\sum_ {i , j \\in W} \\exp (M _ {i j})}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.809 + ], + "angle": 0, + "content": "where \\( W_{k} \\) denotes the local window for the \\( k^{th} \\) support keypoint and \\( i, j \\) indexes the horizontal and vertical coordinates. Here we denote the similarity-aware position proposals for all the support keypoints as \\( P \\in \\mathbb{R}^{K \\times 2} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.726, + 0.833 + ], + "angle": 0, + "content": "3.5. Proposal Refine Decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "As mentioned in Sec. 1, due to the unreliable matching process, directly using the matched keypoints from the first stage as predictions is prone to error. Intuitively, each support keypoint can be corrected by jointly considering other" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7311" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.091, + 0.429, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.389 + ], + "angle": 0, + "content": "Figure 4. Illustration of the proposal refine decoder. The decoder consists of a self-attention and a cross-attention layer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.463 + ], + "angle": 0, + "content": "support keypoint predictions and keypoint-related information in the query image. Hence, we devise the second stage to unleash the potential of correcting the raw similarity-aware position proposals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 4, we design a proposal refine decoder, where the support keypoint embedding together with the keypoint positions are fed into the decoder layer. Each layer consists of two steps: 1) self-attention among support keypoints makes each proposal aware of other keypoints' positions and contents; 2) cross-attention that helps extract relevant content from the query feature for each support keypoint. By integrating these two attention mechanisms, the decoder can further refine the position proposals and produce more accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.811 + ], + "angle": 0, + "content": "Self-attention among support keypoints. Self-attention enables interactions among the support keypoints, which makes each proposal aware of other keypoints' positions and contents. To prepare the inputs for the self-attention layer, the support keypoint positions are first converted into sinusoidal positional encoding \\( E_{s}^{l} \\in \\mathbb{R}^{K \\times c} \\). Then, the support keypoint identifier \\( I_{s} \\) and position embedding \\( E_{s}^{l} \\) are added into \\( F_{s}^{l} \\) forming the query and key for the first self-attention layer. The support keypoint feature \\( F_{s}^{l} \\) is directly used as the value. The multi-head self-attention layer is followed by a transformer-style feed-forward network (FFN) with residual connection and layer normalization [2]. We denote the support feature processed by the FFN as \\( F_{s}^{l'} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.904 + ], + "angle": 0, + "content": "Cross-attention between support and query features. Cross-attention helps extract relevant contents from the query feature \\( F_{q} \\) for each support keypoint. As illustrated in Fig. 4, we use the concatenation of \\( F_{s}^{\\prime} \\) and the summation of \\( I_{s} \\) and \\( E_{s}^{l} \\) as the query for the cross-attention layer. Correspondingly, query image features \\( F_{q} \\) and positional em" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "bedding \\(P_{q}\\) are concatenated as the key, and \\(F_{q}\\) are used as the value. Instead of addition, we concatenate the positional embedding and support identifier for key and value, following DAB-DETR [23]. An FFN processes the output from the cross-attention layer to form the support keypoint features for the next decoder layer, i.e., \\(F_{s}^{l + 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Position update. We use the refined support feature \\( F_{s}^{l + 1} \\) to update the keypoint positions \\( P^l \\) to \\( P^{l + 1} \\). An MLP takes \\( F_{s}^{l + 1} \\) as input and infers the offset \\( O^{l + 1}\\in \\mathbb{R}^{K\\times 2} \\) pointing to the ground truth. Then, the keypoint positions are updated by:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.268, + 0.892, + 0.288 + ], + "angle": 0, + "content": "\\[\nP ^ {l + 1} = \\sigma \\left(O ^ {l} + \\sigma^ {- 1} \\left(P ^ {l}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.892, + 0.374 + ], + "angle": 0, + "content": "where \\(\\sigma\\) and \\(\\sigma^{-1}\\) denote the sigmoid and its inverse function. Note that the similarity-aware proposals are used as the keypoint positions for the first decoder layer, i.e., \\(P^1\\). The updated keypoint positions from the last decoder layer are used as the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.383, + 0.681, + 0.4 + ], + "angle": 0, + "content": "3.6. Supervision Signal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.481 + ], + "angle": 0, + "content": "CapeFormer incorporates two supervision signals: the similarity loss and the offset loss. The similarity loss facilitates the learning of representation and the similarity metric in the first stage. The offset loss supervises the learning of proposal refinement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.557 + ], + "angle": 0, + "content": "Similarity loss. The similarity loss is imposed on the similarity maps \\( M \\) from the similarity-aware proposal generator. We simultaneously constrain the shape of similarity map and the accuracy of similarity-aware proposals with a heatmap term \\( \\mathcal{L}_h \\) and an expectation term \\( \\mathcal{L}_e \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.589 + ], + "angle": 0, + "content": "For heatmap loss, we use sigmoid to normalize similarity map \\(M\\). Then, \\(\\mathcal{L}_h\\) are computed with \\(\\ell_2\\) loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.599, + 0.892, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {1}{H \\cdot W} \\| \\text {s i g m o i d} (M _ {i}) - H _ {i} \\| _ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.71 + ], + "angle": 0, + "content": "where \\(H_{i}\\) denotes the ground truth heatmap for the \\(i^{th}\\) keypoint. For \\(\\mathcal{L}_e\\), we first normalize the similarity map with softmax and then take the expectation of the coordinates as:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.71, + 0.892, + 0.751 + ], + "angle": 0, + "content": "\\[\nE \\left(P _ {k}\\right) = \\sum_ {j = 1} ^ {H} \\sum_ {i = 1} ^ {W} \\frac {\\exp \\left(M _ {k , i j}\\right)}{\\sum \\exp \\left(M _ {k}\\right)} \\cdot (i, j), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.806 + ], + "angle": 0, + "content": "where \\( E(P_k) \\) denotes the coordinate expectation for the \\( k^{th} \\) keypoint. \\( H \\) and \\( W \\) denote the height and width for similarity map. Then we use \\( \\ell_1 \\) loss to calculate \\( \\mathcal{L}_e \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\left| E \\left(P _ {i}\\right) - \\hat {P} _ {i} \\right| _ {1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{P}_i\\) denotes the ground truth coordinate of the \\(i^{th}\\) point." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7312" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 1. Results on the MP-100 dataset. Performance (PCK) under 1-shot and 5-shot settings. The proposed CapeFormer achieves the best PCK on all the splits under 1-shot and 5-shot settings." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.123, + 0.878, + 0.238 + ], + "angle": 0, + "content": "
Method1-shot5-shot
Split 1Split 2Split 3Split 4Split 5AverageSplit 1Split 2Split 3Split 4Split 5Average
ProtoNet [35]46.0540.8449.1343.3444.5444.7860.3153.5161.9258.4458.6158.56
MAML [11]68.1454.7264.1963.2457.2061.5070.0355.9863.2164.7958.4762.50
Fine-tune [27]70.6057.0466.0665.0059.2063.5871.6757.8466.7666.5360.2464.61
POMNet [40]84.2378.2578.1778.6879.1779.7084.7279.6178.0080.3880.8580.71
CapeFormer89.4584.8883.5983.5385.0985.3191.9488.9289.4088.0188.2589.30
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.471, + 0.287 + ], + "angle": 0, + "content": "Offset regression loss. Supposing a total of \\(L\\) decoder layers are used, we define the offset regression loss \\(\\mathcal{L}_o\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.296, + 0.47, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\sum_ {i = 1} ^ {K} \\left| P _ {i} ^ {l + 1} - \\hat {P} _ {i} \\right| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.47, + 0.381 + ], + "angle": 0, + "content": "The overall training loss is calculated as \\(\\lambda_h \\cdot \\mathcal{L}_h + \\mathcal{L}_e + \\mathcal{L}_o\\), where \\(\\lambda_h\\) is a pre-defined weight for the heatmap loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.21, + 0.41 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.417, + 0.296, + 0.433 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.546 + ], + "angle": 0, + "content": "Network architecture. We use the ResNet-50 [13] pretrained from ImageNet [9] as the backbone following POMNet [40]. Different from POMNet, the backbone is shared by query and support images in our cases for efficiency. Both the query-support joint refine encoder and the proposal refine decoder have three layers. Refer to supplementary for more details on network architectures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Training strategy. We use Adam [16] optimizer to train the model for 200 epochs with the batch size of 16. The learning rate is set as \\( 1e^{-5} \\) and decays by \\( 10 \\times \\) on the \\( 160^{th} \\) and \\( 180^{th} \\) epoch. The weight for heatmap loss is empirically set as 2.0 to make different loss terms lie in the same magnitude. Data augmentation and pre-processing are kept the same with POMNet [40] for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Dataset and metric. We use MP-100 dataset [40] to train and evaluate our method. Samples are collected from the existing category-specific pose estimation datasets [22, 32, 41]. MP-100 dataset encompasses over 18k images from 100 different categories. Keypoint numbers of different categories vary from 8 to 68. To train and evaluate the model, the samples are organized into five splits. On each split, the training, validation, and test categories have no overlaps, i.e., the categories for evaluation are not accessed during training. We use the Probability of Correct Keypoint (PCK) [42] as the quantitative metric. The threshold for PCK is set to 0.2 following POMNet [40]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.26, + 0.863 + ], + "angle": 0, + "content": "4.2. Results on MP-100" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We compare our method with the previous best CAPE method POMNet [40] and three baselines: ProtoNet [35]," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.253, + 0.892, + 0.282 + ], + "angle": 0, + "content": "Table 2. Cross super-category evaluation on MP-100 dataset. Experiments are conducted under the 1-shot setting." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.287, + 0.892, + 0.387 + ], + "angle": 0, + "content": "
MethodHuman BodyHuman FaceVehicleFurniture
ProtoNet [35]37.6157.8028.3542.64
MAML [11]51.9325.7217.6820.09
Fine-tune [27]52.1125.5317.4620.76
POMNet [40]73.8279.6334.9247.27
CapeFormer83.4480.9645.4052.49
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.499 + ], + "angle": 0, + "content": "MAML [11], and Fine-tune [27]. PCKs of different approaches on the MP-100 dataset under 1-shot and 5-shot settings are reported in Table 1, where we can see our method significantly outperforms POMNet [40], and improves the average PCK by \\(5.6\\%\\) under the 1-shot setting and \\(8.6\\%\\) under the 5-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.893, + 0.682 + ], + "angle": 0, + "content": "1-shot CapeFormer can surpass other methods with five support images, demonstrating the robustness of the proposed two-stage framework when the support information is limited. When increasing the shots from one to five, CapeFormer obtains a significant improvement of \\(4.0\\%\\) on PCK. For POMNet, the improvement from 1-shot to 5-shot is \\(1.0\\%\\). This shows that CapeFormer can better utilize the additional information when more support samples are available. Note that we use the same way as POMNet [40] to encode the support keypoint features for fair comparisons. We also include a comparison with the semantic correspondence approaches in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.695, + 0.836, + 0.71 + ], + "angle": 0, + "content": "4.3. Cross Super-Category Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.854 + ], + "angle": 0, + "content": "In the MP-100 dataset, training, validation, and test categories do not overlap. However, some categories may still share similar characteristics, such as the faces of different animals. To better evaluate the generalization ability on significantly different categories, we conduct a cross-supercategory evaluation following POMNet [40]. Four supercategories: human face, human body, vehicle, and furniture, are collected from the MP-100 dataset as the test categories, while the other categories are used as the training samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "As in Table 2, CapeFormer achieves the state-of-the-art performance. The advantages on the human body and vehicle are particularly remarkable, with an improvement" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7313" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.887, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our method. We visualize the keypoint predictions under 1-shot setting. The left column denotes the support image with corresponding keypoints. The middle column denotes the results without proposal refine decoder, and the right column are results from the full method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.473, + 0.618 + ], + "angle": 0, + "content": "of \\(9.62\\%\\) and \\(10.48\\%\\) on PCK, respectively. Compared with the standard setting, the improvement in cross supercategory setting is more significant, which further demonstrates the generalization ability and robustness of the proposed method. The improvement of the human face is limited \\((1.36\\%)\\). A plausible reason is that the human face has much more keypoints (68) than the training categories (the largest keypoint number in training is 38). Such discrepancy renders great difficulties for generalization: the keypoints are densely arranged, and the training data does not contain such a fine-grained matching case." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.231, + 0.647 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.821 + ], + "angle": 0, + "content": "In this section, we first ablate the two-stage framework and the proposed modules in Table 3. Then, we validate some detailed design choices, including the implementation of training loss and support keypoint identifier. We perform experiments on the test set of MP-100 split1 under the 1-shot setting following [40]. We set some variants of the encoder for ablations on network components. (i) DETR [5] encoder removes the query-support interactions, i.e., the information fusion among support keypoints and query images is disabled; (ii) QSR denotes the proposed query-support joint refine encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Two-stage framework. First we focus on the effectiveness of the second stage. Comparing No. 3 and No. 5 in Table 3, the decoder brings an improvement of \\(4.13\\%\\) on PCK, demonstrating the necessity of the second stage." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We compare one-stage and two-stage paradigms qualita-" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.449, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Table 3. Ablation study on different components in Cape-Former. Support ID stands for the support keypoint identifier in Sec. 3.2. QSR denotes our query-support joint refine encoder." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.495, + 0.884, + 0.597 + ], + "angle": 0, + "content": "
Support IDEncoderDecoderParadigmPCK
No.1--DETR--one-stage80.32
No.2--QSR--one-stage82.86
No.3✓QSR--one-stage85.32
No.4--QSR✓two-stage85.81
No.5✓QSR✓two-stage89.45
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.79 + ], + "angle": 0, + "content": "tively in Fig. 5. \"One-Stage\" denotes the direct matching results from No. 3 setting in Table 3, here and after. The direct matching results are prone to error when query and support instances are different, e.g., the dog on the right part of the first row. We also visualize how the decoder corrects the predictions in Fig. 6. For the alpaca face, the inner left canthus (red point) is close to the annotation of the right inner canthus (yellow point) due to the occlusion. Hence, the similarity-aware position proposal for the left inner canthus is wrongly generated at the right inner canthus. The decoder gradually pushes it to the correct position." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Representation. Here we analyze the effectiveness of query-support joint refine encoder and the support keypoint identifier. i) Comparing No.1 with No.2 in Table 3, the QSR encoder brings an absolute improvement of \\(2.54\\%\\) on PCK over DETR encoder, which suggests the effectiveness of support-query mutual information fusion. 2) Comparing No.2 with No.3 in Table 3, introducing the support iden" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7314" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.09, + 0.203, + 0.102 + ], + "angle": 0, + "content": "Support Image" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.103, + 0.219, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.187, + 0.188, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.091, + 0.319, + 0.102 + ], + "angle": 0, + "content": "Query Image" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.103, + 0.335, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.185, + 0.311, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.123, + 0.432, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.333 + ], + "angle": 0, + "content": "Figure 6. The decoder can help correct unreliable proposals. The left column shows the support keypoint (red point) and the right column shows the refining process in the second stage. Note that in the first row, the left inner canthus is invisible due to occlusion, which looks to be on the right." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.338, + 0.257, + 0.366 + ], + "angle": 0, + "content": "Table 4. Analysis of different terms in the similarity loss." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.372, + 0.248, + 0.444 + ], + "angle": 0, + "content": "
LhLePCK
✓--88.74
--✓89.41
✓✓89.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.268, + 0.338, + 0.466, + 0.366 + ], + "angle": 0, + "content": "Table 5. Comparison of different support keypoint identifiers." + }, + { + "type": "table", + "bbox": [ + 0.291, + 0.372, + 0.441, + 0.444 + ], + "angle": 0, + "content": "
TypePCK
--85.81
Learnable88.96
Fixed89.45
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.533 + ], + "angle": 0, + "content": "tifier improves the PCK by \\(2.46\\%\\). Note that the No. 3 model excludes the decoder, which is one-stage. Its performance can still achieve state-of-the-art on split1 test set, which suggests that the support identifier and QSR decoder significantly improve representation quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.669 + ], + "angle": 0, + "content": "To intuitively show how the encoder and decoder work, we visualize the encoder/decoder attention maps in Fig. 7. The attention shows which location the support keypoint will focus during the information fusion (first stage) and the proposal correction (second stage). As can be observed, in the encoder, the support keypoint attends to keypoints with similar appearance, while in the decoder, the attention becomes more focused on the corresponding keypoints for proposal correction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.468, + 0.778 + ], + "angle": 0, + "content": "Loss function. We analyze the similarity loss in Table 4. i) Only using the heatmap supervision (row 1) yields the worst performance, indicating that the expectation of loss is beneficial for precise localization. ii) As shown in row 3, the combination of heatmap loss and expectation attains the best performance, as both the shape of similarity map and the proposal position can be supervised." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Support keypoint identifier. We compare two different implementations of support keypoint identifier: learnable embedding and fixed sinusoidal embedding in Table 5. Using fixed sinusoidal embedding improves PCK by \\(0.49\\%\\) compared with learnable embedding. The learnable embedding is similar to the learnable query embedding in DETR-based detection methods [5,25], which can be viewed as a learnable content embedding or anchors [23]. However, our" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.091, + 0.64, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.178, + 0.624, + 0.201 + ], + "angle": 0, + "content": "Support Image &Keypoint" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.091, + 0.747, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.178, + 0.719, + 0.199 + ], + "angle": 0, + "content": "Encoder Attention" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.091, + 0.859, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.179, + 0.845, + 0.199 + ], + "angle": 0, + "content": "Decoder Cross-Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.204, + 0.892, + 0.274 + ], + "angle": 0, + "content": "Figure 7. Visualization of attention map. We visualize the encoder and decoder attention for the support keypoint at the left inner canthus. The support keypoint coarsely attends to keypoints with similar appearance in the encoder while focusing more on the corresponding keypoints for position update in the decoder." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.283, + 0.891, + 0.311 + ], + "angle": 0, + "content": "Table 6. Inference and Training Efficiency. We test the efficiency on RTX3090 under 1-shot setting." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.317, + 0.887, + 0.389 + ], + "angle": 0, + "content": "
MethodParams (M)GFLOPsFPSMemory (GB)
POMNet [40]48.2138.016.8013.8 × 2
One-Stage26.8622.6536.907.5
CapeFormer31.1423.6826.097.8
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.463 + ], + "angle": 0, + "content": "support keypoint identifier is fixed and serves a different purpose: distinguishing different support keypoints. Learnable support identifier deteriorates the performance as they may overfit the training categories more easily." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.472, + 0.828, + 0.489 + ], + "angle": 0, + "content": "4.5. Performance and Efficiency Trade-off" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.662 + ], + "angle": 0, + "content": "We evaluate the efficiency under 1-shot setting for POMNet and CapeFormer in Table 6. The training memory is evaluated with a batchsize of 16. i) Comparing row 1 and 3, CapeFormer significantly reduces the parameters \\((-35.4\\%)\\) and FLOPs \\((-37.7\\%)\\) compared with POMNet due to the shared backbone and efficient similarity metric. ii) Comparing rows 2 and 3, the second stage only introduces limited parameters and FLOPs, which shows a satisfying trade-off between performance and efficiency. iii) CapeFormer is efficient for training, which consumes only 8GB memory on a single GPU (almost \\(1/4\\) of POMNet)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.619, + 0.692 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.853 + ], + "angle": 0, + "content": "In this work, we formulate a two-stage framework for category-agnostic pose estimation (CAPE) and instantiate it with a novel transformer model. Unlike previous one-stage approaches that entirely rely on matching results, the matching results in the first stage are viewed as similarity-aware position proposals which are refined in the second stage. We also propose specific designs to improve the representation and similarity modeling. On the CAPE benchmark MP-100, our method outperforms the previous best method by large margins on both accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (Grant No. U1913602 and 61876211)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7315" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter V. Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 3686-3693, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[2] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv Computer Research Repository, abs/1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.296 + ], + "angle": 0, + "content": "[3] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9497-9506, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.366 + ], + "angle": 0, + "content": "[4] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1):172-186, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of European Conference on Computer Vision (ECCV), pages 213-229, 2020. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[6] Seokju Cho, Sunghwan Hong, Sangryul Jeon, Yunsung Lee, Kwanghoon Sohn, and Seungryong Kim. Cats: Cost aggregation transformers for visual correspondence. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 9011-9023, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.563 + ], + "angle": 0, + "content": "[7] Seokju Cho, Sunghwan Hong, and Seungryong Kim. Cats++: Boosting cost aggregation with convolutions and transformers. arXiv Computer Research Repository, abs/2202.06817, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[8] Christopher B. Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Krishna Chandraker. Universal correspondence network. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 2406-2414, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of International Conference on Machine Learning (ICML), pages 1126–1135, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 14671-14681, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Shuqiang Jiang, Sisi Liang, Chengpeng Chen, Yaohui Zhu, and Xiangyang Li. Class agnostic image common object detection. IEEE Trans. Image Process., 28(6):2836-2846, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[15] Seungryong Kim, Dongbo Min, Bumsub Ham, Stephen Lin, and Kwanghoon Sohn. FCSS: fully convolutional self-similarity for dense semantic correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):581-595, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of International Conference on Learning Representations, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[17] Rollyn Labuguen, Jumpei Matsumoto, Salvador Blanco Negrete, Hiroshi Nishimaru, Hisao Nishijo, Masahiko Takada, Yasuhiro Go, Ken ichi Inoue, and Tomohiro Shibata. Macaquepose: A novel \"in the wild\" macaque monkey pose dataset for markerless motion capture. Frontiers in behavioral neuroscience (2021), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[18] Jiefeng Li, Siyuan Bian, Ailing Zeng, Can Wang, Bo Pang, Wentao Liu, and Cewu Lu. Human pose regression with residual log-likelihood estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 11005–11014, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[19] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 10863-10872, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[20] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: A benchmark for amur tiger re-identification in the wild. In Proceedings of ACM International Conference on Multimedia, pages 2590–2598, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[21] Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Tokenpose: Learning keypoint tokens for human pose estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 11293-11302, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), pages 740-755, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In Proceedings of International Conference on Learning Representations, 2022. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Proceedings of Asian Conference on Computer Vision (ACCV), pages 669–684, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3631-3640, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel flow: Semantic correspondence with multi-layer" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7316" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "neural features. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 3394-3403, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[27] Akihiro Nakamura and Tatsuya Harada. Revisiting finetuning for few-shot learning. arXiv Computer Research Repository, abs/1910.00216, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[28] Xuecheng Nie, Jianfeng Zhang, Shuicheng Yan, and Jiashi Feng. Single-stage multi-person pose machines. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 6950-6959, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[29] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 1906-1915, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[31] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomás Pajdla, and Josef Sivic. Neighbourhood consensus networks. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 1658-1669, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[32] Christos Sagonas, Epameinondas Antonakos, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: database and results. Image Vis. Comput., 47:3-18, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[33] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 11059-11068, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[34] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 9529–9538, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[35] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4077-4087, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[36] Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5452-5462, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[37] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 2621-2630, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[38] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 1653-1660, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Proceedings of European Conference on Computer Vision (ECCV), volume abs/2207.10387, 2022. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[41] Shuo Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5525-5533, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[42] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2878-2890, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[43] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5217–5226, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[44] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 7091-7100, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[45] Song-Hai Zhang, Ruilong Li, Xin Dong, Paul L. Rosin, Zixi Cai, Han Xi, Dingcheng Yang, Haozhi Huang, and Shi-Min Hu. Pose2seg: Detection free human instance segmentation. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 889-898, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.894, + 0.612 + ], + "angle": 0, + "content": "[46] Jian Zhao, Jianshu Li, Yu Cheng, Li Zhou, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In Proceedings of ACM International Conference on Multimedia, pages 792-800, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7317" + } + ] +] \ No newline at end of file diff --git a/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_origin.pdf b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eec499d49299f26022103995c81ed0baf00a8531 --- /dev/null +++ b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/dd58280d-b5a5-4fde-a3e8-4c9ec2674829_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f27bedc6d8905752724ab2eabea700ff81232610ecc55321329728ce8c1de5f4 +size 5910148 diff --git a/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/full.md b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..33a3a6c22f3b23530af47cd28a2d78849e5dad35 --- /dev/null +++ b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/full.md @@ -0,0 +1,328 @@ +# Matching Is Not Enough: A Two-Stage Framework for Category-Agnostic Pose Estimation + +Min Shi $^{1,2,*}$ Zihao Huang $^{1,*}$ Xianzheng Ma $^{2}$ Xiaowei Hu $^{2}$ Zhiguo Cao $^{1,\dagger}$ $^{1}$ School of AIA, Huazhong University of Science and Technology + $^{2}$ Shanghai AI Laboratory + +{min_shi, zihaohuang, zgcao}@hust.edu.cn, {maxianzheng, huxiaowei}@pjlab.org.cn + +# Abstract + +Category-agnostic pose estimation (CAPE) aims to predict keypoints for arbitrary categories given support images with keypoint annotations. Existing approaches match the keypoints across the image for localization. However, such a one-stage matching paradigm shows inferior accuracy: the prediction heavily relies on the matching results, which can be noisy due to the open set nature in CAPE. For example, two mirror-symmetric keypoints (e.g., left and right eyes) in the query image can both trigger high similarity on certain support keypoints (eyes), which leads to duplicated or opposite predictions. To calibrate the inaccurate matching results, we introduce a two-stage framework, where matched keypoints from the first stage are viewed as similarity-aware position proposals. Then, the model learns to fetch relevant features to correct the initial proposals in the second stage. We instantiate the framework with a transformer model tailored for CAPE. The transformer encoder incorporates specific designs to improve the representation and similarity modeling in the first matching stage. In the second stage, similarity-aware proposals are packed as queries in the decoder for refinement via cross-attention. Our method surpasses the previous best approach by large margins on CAPE benchmark MP-100 on both accuracy and efficiency. Code available at github.com/flyinglynx/CapeFormer + +# 1. Introduction + +Humans can quickly grasp the essentials of a keypoint for arbitrary objects (e.g., the nose of an animal) and then pinpoint the same keypoint on another object in the same category. However, most pose estimation models are still category-specific, which cannot be applied to a new category unless they are trained with sufficient category- + +![](images/ba94522789afd8436d3a534b5462e9eb7ffab4494e9cfd7fc56f804c76778680.jpg) + +![](images/2c8ed89129cf639a07d41f0c75eb1a7b99aaa02c95276caf0232300b712a50a3.jpg) +Figure 1. Comparison with existing methods. (a) Compared with previous one-stage matching paradigm that entirely relies on keypoint matching, we propose a novel two-stage framework, which learns to correct unreliable matching results in the second stage. (b) The proposed approach attains better accuracy and efficiency compared with the previous best approach. + +specific data. Recently, category-agnostic pose estimation (CAPE) [40] is proposed to address such a generic pose estimation problem. CAPE aims to localize instance keypoints for arbitrary categories given one or few support images with keypoint annotations (support keypoints). + +Unlike category-specific pose estimation models that learn to recognize specific keypoints for a fixed category, CAPE models learn to represent and compare keypoints for open-world categories. As shown in Fig. 1 (a), existing CAPE approaches [40] match the support keypoints across the query image in an embedding space. Then, the match- + +ing results are used as clues for keypoint localization. Although a sophisticated feature embedding pipeline is elaborated [40], we find that such a one-stage pipeline shows inferior accuracy and efficiency for CAPE tasks. Specifically, the prediction heavily relies on matching results. However, due to the open set nature in CAPE, the similarity results often contain some noises: 1) insufficient similarity when support and query instances differ significantly in poses, textures, or styles; 2) false positive similarity is triggered when points in the query image share similar appearance with support keypoints. For example, one-stage matching paradigm has difficulties distinguishing the left and right of keypoints due to similar visual characteristics. Thus, the left eye in support keypoints can be indiscriminately predicted to the right eye in the query image. + +In this paper, we remedy these issues from two aspects. First, to refine the matching results, we introduce a two-stage framework for CAPE, where matched keypoints in the first stage are viewed as similarity-aware position proposals. The model learns to refine the proposals in the second stage. In addition, to construct a robust matching process, we also improve the representation quality and similarity metrics. + +Specifically, we instantiate the two-stage framework with a transformer, termed CAPE transformer (CapeFormer). The transformer encoder encodes the support keypoints and query images in the first stage, with specific designs to improve the representation quality. We first design a query-support joint refine encoder to mutually transfer information among support keypoints and the query images, thus narrowing the gap between the support and query instances (poses or styles). Then, we add a support keypoint identifier to alleviate the ambiguity between two support keypoints when they are geometrically close or with similar appearance. After the feature encoding, similarity-aware position proposals are generated for the second stage. We directly match the support keypoint with query image features using a learnable inner-product [34]. The similarity peaks are selected as the similarity-aware position proposals for support keypoints. In the second stage, the position proposals are packed with support keypoint features as the queries in the transformer decoder. In each decoder layer, queries extract relevant features with cross-attention, which are used as clues to update the proposals. + +We evaluate our method on a category-agnostic pose estimation benchmark MP-100 [40]. Our method outperforms the previous best approach POMNet [40] by large margins, with an improvement of $5.6\%$ and $8.6\%$ under 1-shot and 5-shot settings, respectively. When training and test categories almost have no properties in common, a more significant improvement (up to $10.5\%$ ) can be obtained compared with POMNet. Note that, our method is trained end-to-end and shows better efficiency compared with previous best approach, as in Fig. 1 (b). + +Our contributions can be summarized as follows: + +- We propose a two-stage framework for category-agnostic pose estimation tackling the noisy matching results. +- We instantiate the two-stage framework with a transformer model, termed CapeFormer, with specific designs to enhance the representation and similarity modeling in the matching pipeline. +- CapeFormer outperforms the previous best approach significantly in both accuracy and efficiency. + +# 2. Related Work + +# 2.1. Pose Estimation + +Pose estimation aims to localize the semantic keypoints of instances. Most pose estimation approaches are targeted for one specific category, e.g., human [1, 19, 45, 46], animals [3, 17, 20], or vehicles [29, 36]. According to how points are localized, pose estimation methods can be categorized into regression-based [18, 21, 28, 37, 38], heatmap-based [4, 44], offset-based [12] and query-based [33] ones. In this work, we address category-agnostic pose estimation, which has not been well-studied yet. Compared with category-specific pose estimation, category-agnostic pose estimation focuses on constructing a generic representation and similarity metrics: keypoints are predicted by comparing the support keypoints and the query images in the embedding space. + +# 2.2. Category-Agnostic Vision Models + +Category-agnostic vision models aim to build a generic model that can be applied to any categories on a particular task, e.g., object detection [14], segmentation [43], object counting and localization [24, 34]. This setting is also known as the few-shot learning. Recently, POMNet [40] first proposes the category-agnostic pose estimation task and elaborates a dataset MP-100. Most category-agnostic models can be classified into meta-learning and metric-learning approaches. Meta-learning approaches [11] aim to initialize the model with a set of optimal parameters, where the model can quickly fit the new tasks by learning from a few support samples. Metric learning-based approaches match the support samples with query images in an embedding space. Then, the matching results are used as clues for subsequent tasks. For pose estimation, POMNet [40] encodes query images and support keypoints with a transformer. Then, a regression head directly infers the similarity from the concatenated support keypoint and query image features. We find that simply using matching results for point localization is challenging. Hence, we extend such a matching paradigm with a two-stage framework to correct the unreliable matching results for accurate predictions. + +![](images/9e963a7607f3a087c4daa28ca0973cc5c1f08a2bdbae363c315cdb737ca54e22.jpg) +Figure 2. Illustration of the proposed CapeFormer. CapeFormer consists of a matching stage where similarity-aware position proposals are generated, and a refine stage that further calibrates the proposals. A support identifier and query-support joint refine encoder are proposed to improve the representation quality. + +# 2.3. Semantic Correspondence + +Semantic correspondence (SC) [8, 15] aims to match the points that have identical semantic meaning with a dense correspondence field. Hence, SC models can be directly applied on the CAPE: localize the corresponding points for the support keypoints with the estimated correspondence field. Typical SC models first match each pair of pixels and use the matching score to construct correspondence. Similar to our two-stage design that alleviates the noise on similarity matching, a series of SC models refine the matching scores with geometry constraints [26], high-dimensional convolutions [31], and attention mechanism [6, 7]. However, we find directly applying SC models for CAPE is inefficient and yields inferior performance. Besides, most techniques for the matching refine in SC are based on the dense matching volume, which is not available in CAPE, as only the sparse support keypoints are matched. + +# 3. CAPE Transformer + +Class-agnostic pose estimation (CAPE) aims to estimate instance keypoints given one or a few support images with keypoint annotations. Specifically, we denote the query image as $I_{q}$ and the support image as $I_{s}$ , which has $K$ predefined support keypoints. The CAPE model will infer the positions of the corresponding $K$ keypoints in the query images. For simplicity, we use the 1-shot setting for illustration (the number of support images equals one). + +# 3.1. Two-Stage Framework for CAPE + +Different from the previous one-stage paradigm, Cape-Former features a two-stage framework. As in Fig. 2, the first stage consists of the input embedding, a support-query joint refine encoder, and a similarity-aware proposal generator. A proposal refine decoder forms the second stage. + +The input embedding encodes the query image and support keypoints into the embedding space (Sec. 3.2). We denote the features for the query image and support keypoints as $F_{q}$ and $F_{s}$ , which are further refined with the query-support joint refine encoder (Sec. 3.3). The proposal generator measures the similarity between $F_{s}$ and $F_{q}$ and the similarity peaks are selected as the similarity-aware position proposals (Sec. 3.4). Then, in the second stage, the similarity-aware proposals are gradually refined into the final predictions (Sec. 3.5). + +# 3.2. Input Embedding + +The input embedding is illustrated in Fig. 3. We first project the support and query image into feature maps with a shared backbone. The flattened feature map of the query image is used as its representation, denoted by $F_{q} \in \mathbb{R}^{c \times n}$ , where $n$ equals the number of pixels in the feature map and $c$ denotes the embedded dimension. To obtain the features of the support keypoint, we conduct a "soft" ROI pooling [30] on the support image feature map. $K$ soft masks are generated by placing a Gaussian kernel centered at support keypoints, and then multiplied with the support feature map. The summation of masked support feature map is adopted as the support keypoint feature $F_{s} \in \mathbb{R}^{K \times c}$ . For multiple support images, e.g., 5-shot settings, the support keypoint features from different images are averaged. + +Support keypoint identifier. When encoding support keypoints, each support keypoint's positional and context information is discarded during ROI pooling. Hence, when two support keypoints are geometrically close or with similar appearance (e.g., left and right eyes), they can be indistinguishable in the embedding space. Such ambiguity can lead to inaccurate predictions that are duplicated or mirror-symmetric to the ground truth. + +![](images/f3db0a27cd08c797aedf302aafb97a14adf813dbb4972189e33f669638f5ca65.jpg) +Figure 3. Illustration of the input embedding and the support-query joint refine encoder. + +![](images/9a6a8220e256e82432ffb0c951708c315268af0fe7a2fcc638cb9884e7e826fd.jpg) + +To alleviate such ambiguity, we propose a simple but effective support keypoint identifier that gives an "identity" to each support keypoint. The implementation is embarrassingly simple: $K$ embedding vectors are generated via sinusoidal encoding, denoted by $I_{s} \in \mathbb{R}^{K \times c}$ . This identifier is added to $F_{s}$ ahead of attention calculation, which will be discussed later in this section. Such a simple modification can lead to over 3% improvement with negligible computation overhead. + +# 3.3. Query-Support Refine Encoder + +Directly matching the support keypoints with the query image right after the input embedding is challenging, as instances in the support and query images can be different, such as different colors, textures, and poses. Thus, the target and support keypoints may not possess sufficient similarity in the embedding space [10, 34]. An intuitive way to alleviate this is to fuse relevant query and support features. Hence, the query and support features can be pulled closer in the embedding space. + +To this end, we implement information fusion and transfer among support and query with self-attention. As illustrated in Fig. 3, the inputs of the $l^{th}$ encoder layer are query and support features, denoted as $F_{q}^{l}$ and $F_{s}^{l}$ . We augment the $F_{q}^{l}$ with positional embedding as in DETR [5], and add the support keypoint identifier into the $F_{s}^{l}$ . The augmented $F_{q}^{l}$ and $F_{s}^{l}$ are concatenated into a feature sequence with the length of $n + K$ . The sequence is then fed into the standard multi-head self-attention layer [39]. The output sequence from the self-attention layer is processed by a feed forward network (FFN). Then, the sequence are re-split to the support keypoint and query image features, which are used as the inputs for the next encoder layer. + +# 3.4. Similarity-Aware Proposal Generator + +The similarity-aware proposal generator matches the support keypoint features with the query features to obtain the similarity maps, whose peaks are selected as the similarity-aware proposals. Previous state-of-the-art approach POMNet [40] measures similarity implicitly: a regression head directly infers similarity maps (heatmaps) from the concatenation of support keypoint and query features. However, regressing similarity without explicit similarity modeling has the risk of overfitting in class-agnostic setting [34]. Moreover, considerable computation overhead is introduced. For example, given 100 support keypoint features with 256 channels as in our case, the similarity maps are obtained by forwarding 512-channel feature maps with an FCN for 100 times. + +To embrace both efficiency and generality, we use a learnable inner-product [34] to model similarity explicitly. Specifically, we first calculate a channel attention weight $a \in \mathbb{R}^{K \times c}$ conditioned on $F_s \in \mathbb{R}^{K \times c}$ . Then, the similarity map $M$ is obtained as: + +$$ +M = \left(W _ {q} F _ {q}\right) \left(\boldsymbol {a} \circ W _ {s} F _ {s}\right) ^ {T}, \tag {1} +$$ + +where $W_{q}$ and $W_{s}$ are two learnable projection matrices in $\mathbb{R}^{c\times c}$ and $\circ$ denotes the Hadamard product. Then, similarity maps $M\in \mathbb{R}^{n\times K}$ are reshaped into 2D with the shape of $H\times W$ . The channel attention helps stressing the key patterns during inner product computation [34]. + +The maximums from the similarity maps are selected as the similarity-aware position proposals for support keypoints. To reduce the round-off errors on the down-scaled resolutions, instead of directly selecting the maximums, we first generate a $3 \times 3$ local window around the similarity peaks and normalize this local window with softmax. Then, we take the expectation of the coordinates in each local window as the proposal, which can be formulated as: + +$$ +x _ {k} = \sum_ {i, j \in W _ {k}} i \cdot \frac {\exp (M _ {k , i j})}{\sum_ {i , j \in W} \exp (M _ {i j})}, \tag {2} +$$ + +$$ +y _ {k} = \sum_ {i, j \in W _ {k}} j \cdot \frac {\exp (M _ {k , i j})}{\sum_ {i , j \in W} \exp (M _ {i j})}, \tag {3} +$$ + +where $W_{k}$ denotes the local window for the $k^{th}$ support keypoint and $i, j$ indexes the horizontal and vertical coordinates. Here we denote the similarity-aware position proposals for all the support keypoints as $P \in \mathbb{R}^{K \times 2}$ . + +# 3.5. Proposal Refine Decoder + +As mentioned in Sec. 1, due to the unreliable matching process, directly using the matched keypoints from the first stage as predictions is prone to error. Intuitively, each support keypoint can be corrected by jointly considering other + +![](images/5bf9a158d390d6d1573f62937d46ec31a518925e7d268599ce64393d6c375625.jpg) +Figure 4. Illustration of the proposal refine decoder. The decoder consists of a self-attention and a cross-attention layer + +support keypoint predictions and keypoint-related information in the query image. Hence, we devise the second stage to unleash the potential of correcting the raw similarity-aware position proposals. + +Specifically, as shown in Fig. 4, we design a proposal refine decoder, where the support keypoint embedding together with the keypoint positions are fed into the decoder layer. Each layer consists of two steps: 1) self-attention among support keypoints makes each proposal aware of other keypoints' positions and contents; 2) cross-attention that helps extract relevant content from the query feature for each support keypoint. By integrating these two attention mechanisms, the decoder can further refine the position proposals and produce more accurate predictions. + +Self-attention among support keypoints. Self-attention enables interactions among the support keypoints, which makes each proposal aware of other keypoints' positions and contents. To prepare the inputs for the self-attention layer, the support keypoint positions are first converted into sinusoidal positional encoding $E_{s}^{l} \in \mathbb{R}^{K \times c}$ . Then, the support keypoint identifier $I_{s}$ and position embedding $E_{s}^{l}$ are added into $F_{s}^{l}$ forming the query and key for the first self-attention layer. The support keypoint feature $F_{s}^{l}$ is directly used as the value. The multi-head self-attention layer is followed by a transformer-style feed-forward network (FFN) with residual connection and layer normalization [2]. We denote the support feature processed by the FFN as $F_{s}^{l'}$ . + +Cross-attention between support and query features. Cross-attention helps extract relevant contents from the query feature $F_{q}$ for each support keypoint. As illustrated in Fig. 4, we use the concatenation of $F_{s}^{\prime}$ and the summation of $I_{s}$ and $E_{s}^{l}$ as the query for the cross-attention layer. Correspondingly, query image features $F_{q}$ and positional em + +bedding $P_{q}$ are concatenated as the key, and $F_{q}$ are used as the value. Instead of addition, we concatenate the positional embedding and support identifier for key and value, following DAB-DETR [23]. An FFN processes the output from the cross-attention layer to form the support keypoint features for the next decoder layer, i.e., $F_{s}^{l + 1}$ . + +Position update. We use the refined support feature $F_{s}^{l + 1}$ to update the keypoint positions $P^l$ to $P^{l + 1}$ . An MLP takes $F_{s}^{l + 1}$ as input and infers the offset $O^{l + 1}\in \mathbb{R}^{K\times 2}$ pointing to the ground truth. Then, the keypoint positions are updated by: + +$$ +P ^ {l + 1} = \sigma \left(O ^ {l} + \sigma^ {- 1} \left(P ^ {l}\right)\right), \tag {4} +$$ + +where $\sigma$ and $\sigma^{-1}$ denote the sigmoid and its inverse function. Note that the similarity-aware proposals are used as the keypoint positions for the first decoder layer, i.e., $P^1$ . The updated keypoint positions from the last decoder layer are used as the final prediction. + +# 3.6. Supervision Signal + +CapeFormer incorporates two supervision signals: the similarity loss and the offset loss. The similarity loss facilitates the learning of representation and the similarity metric in the first stage. The offset loss supervises the learning of proposal refinement. + +Similarity loss. The similarity loss is imposed on the similarity maps $M$ from the similarity-aware proposal generator. We simultaneously constrain the shape of similarity map and the accuracy of similarity-aware proposals with a heatmap term $\mathcal{L}_h$ and an expectation term $\mathcal{L}_e$ . + +For heatmap loss, we use sigmoid to normalize similarity map $M$ . Then, $\mathcal{L}_h$ are computed with $\ell_2$ loss as: + +$$ +\mathcal {L} _ {h} = \frac {1}{K} \sum_ {i = 1} ^ {K} \frac {1}{H \cdot W} \| \text {s i g m o i d} (M _ {i}) - H _ {i} \| _ {2}, \tag {5} +$$ + +where $H_{i}$ denotes the ground truth heatmap for the $i^{th}$ keypoint. For $\mathcal{L}_e$ , we first normalize the similarity map with softmax and then take the expectation of the coordinates as: + +$$ +E \left(P _ {k}\right) = \sum_ {j = 1} ^ {H} \sum_ {i = 1} ^ {W} \frac {\exp \left(M _ {k , i j}\right)}{\sum \exp \left(M _ {k}\right)} \cdot (i, j), \tag {6} +$$ + +where $E(P_k)$ denotes the coordinate expectation for the $k^{th}$ keypoint. $H$ and $W$ denote the height and width for similarity map. Then we use $\ell_1$ loss to calculate $\mathcal{L}_e$ as: + +$$ +\mathcal {L} _ {e} = \frac {1}{K} \sum_ {i = 1} ^ {K} \left| E \left(P _ {i}\right) - \hat {P} _ {i} \right| _ {1}, \tag {7} +$$ + +where $\hat{P}_i$ denotes the ground truth coordinate of the $i^{th}$ point. + +Table 1. Results on the MP-100 dataset. Performance (PCK) under 1-shot and 5-shot settings. The proposed CapeFormer achieves the best PCK on all the splits under 1-shot and 5-shot settings. + +
Method1-shot5-shot
Split 1Split 2Split 3Split 4Split 5AverageSplit 1Split 2Split 3Split 4Split 5Average
ProtoNet [35]46.0540.8449.1343.3444.5444.7860.3153.5161.9258.4458.6158.56
MAML [11]68.1454.7264.1963.2457.2061.5070.0355.9863.2164.7958.4762.50
Fine-tune [27]70.6057.0466.0665.0059.2063.5871.6757.8466.7666.5360.2464.61
POMNet [40]84.2378.2578.1778.6879.1779.7084.7279.6178.0080.3880.8580.71
CapeFormer89.4584.8883.5983.5385.0985.3191.9488.9289.4088.0188.2589.30
+ +Offset regression loss. Supposing a total of $L$ decoder layers are used, we define the offset regression loss $\mathcal{L}_o$ as: + +$$ +\mathcal {L} _ {o} = \frac {1}{L} \sum_ {l = 1} ^ {L} \sum_ {i = 1} ^ {K} \left| P _ {i} ^ {l + 1} - \hat {P} _ {i} \right| _ {1}. \tag {8} +$$ + +The overall training loss is calculated as $\lambda_h \cdot \mathcal{L}_h + \mathcal{L}_e + \mathcal{L}_o$ , where $\lambda_h$ is a pre-defined weight for the heatmap loss. + +# 4. Experiments + +# 4.1. Implementation Details + +Network architecture. We use the ResNet-50 [13] pretrained from ImageNet [9] as the backbone following POMNet [40]. Different from POMNet, the backbone is shared by query and support images in our cases for efficiency. Both the query-support joint refine encoder and the proposal refine decoder have three layers. Refer to supplementary for more details on network architectures. + +Training strategy. We use Adam [16] optimizer to train the model for 200 epochs with the batch size of 16. The learning rate is set as $1e^{-5}$ and decays by $10 \times$ on the $160^{th}$ and $180^{th}$ epoch. The weight for heatmap loss is empirically set as 2.0 to make different loss terms lie in the same magnitude. Data augmentation and pre-processing are kept the same with POMNet [40] for fair comparisons. + +Dataset and metric. We use MP-100 dataset [40] to train and evaluate our method. Samples are collected from the existing category-specific pose estimation datasets [22, 32, 41]. MP-100 dataset encompasses over 18k images from 100 different categories. Keypoint numbers of different categories vary from 8 to 68. To train and evaluate the model, the samples are organized into five splits. On each split, the training, validation, and test categories have no overlaps, i.e., the categories for evaluation are not accessed during training. We use the Probability of Correct Keypoint (PCK) [42] as the quantitative metric. The threshold for PCK is set to 0.2 following POMNet [40]. + +# 4.2. Results on MP-100 + +We compare our method with the previous best CAPE method POMNet [40] and three baselines: ProtoNet [35], + +Table 2. Cross super-category evaluation on MP-100 dataset. Experiments are conducted under the 1-shot setting. + +
MethodHuman BodyHuman FaceVehicleFurniture
ProtoNet [35]37.6157.8028.3542.64
MAML [11]51.9325.7217.6820.09
Fine-tune [27]52.1125.5317.4620.76
POMNet [40]73.8279.6334.9247.27
CapeFormer83.4480.9645.4052.49
+ +MAML [11], and Fine-tune [27]. PCKs of different approaches on the MP-100 dataset under 1-shot and 5-shot settings are reported in Table 1, where we can see our method significantly outperforms POMNet [40], and improves the average PCK by $5.6\%$ under the 1-shot setting and $8.6\%$ under the 5-shot setting. + +1-shot CapeFormer can surpass other methods with five support images, demonstrating the robustness of the proposed two-stage framework when the support information is limited. When increasing the shots from one to five, CapeFormer obtains a significant improvement of $4.0\%$ on PCK. For POMNet, the improvement from 1-shot to 5-shot is $1.0\%$ . This shows that CapeFormer can better utilize the additional information when more support samples are available. Note that we use the same way as POMNet [40] to encode the support keypoint features for fair comparisons. We also include a comparison with the semantic correspondence approaches in the supplementary. + +# 4.3. Cross Super-Category Pose Estimation + +In the MP-100 dataset, training, validation, and test categories do not overlap. However, some categories may still share similar characteristics, such as the faces of different animals. To better evaluate the generalization ability on significantly different categories, we conduct a cross-supercategory evaluation following POMNet [40]. Four supercategories: human face, human body, vehicle, and furniture, are collected from the MP-100 dataset as the test categories, while the other categories are used as the training samples. + +As in Table 2, CapeFormer achieves the state-of-the-art performance. The advantages on the human body and vehicle are particularly remarkable, with an improvement + +![](images/29308e8337988c5ac90f05e95e6f46b8c56bf2d4a4405b1312a5f1bfdf253317.jpg) +Figure 5. Qualitative results of our method. We visualize the keypoint predictions under 1-shot setting. The left column denotes the support image with corresponding keypoints. The middle column denotes the results without proposal refine decoder, and the right column are results from the full method. + +of $9.62\%$ and $10.48\%$ on PCK, respectively. Compared with the standard setting, the improvement in cross supercategory setting is more significant, which further demonstrates the generalization ability and robustness of the proposed method. The improvement of the human face is limited $(1.36\%)$ . A plausible reason is that the human face has much more keypoints (68) than the training categories (the largest keypoint number in training is 38). Such discrepancy renders great difficulties for generalization: the keypoints are densely arranged, and the training data does not contain such a fine-grained matching case. + +# 4.4. Ablation Study + +In this section, we first ablate the two-stage framework and the proposed modules in Table 3. Then, we validate some detailed design choices, including the implementation of training loss and support keypoint identifier. We perform experiments on the test set of MP-100 split1 under the 1-shot setting following [40]. We set some variants of the encoder for ablations on network components. (i) DETR [5] encoder removes the query-support interactions, i.e., the information fusion among support keypoints and query images is disabled; (ii) QSR denotes the proposed query-support joint refine encoder. + +Two-stage framework. First we focus on the effectiveness of the second stage. Comparing No. 3 and No. 5 in Table 3, the decoder brings an improvement of $4.13\%$ on PCK, demonstrating the necessity of the second stage. + +We compare one-stage and two-stage paradigms qualita- + +Table 3. Ablation study on different components in Cape-Former. Support ID stands for the support keypoint identifier in Sec. 3.2. QSR denotes our query-support joint refine encoder. + +
Support IDEncoderDecoderParadigmPCK
No.1--DETR--one-stage80.32
No.2--QSR--one-stage82.86
No.3✓QSR--one-stage85.32
No.4--QSR✓two-stage85.81
No.5✓QSR✓two-stage89.45
+ +tively in Fig. 5. "One-Stage" denotes the direct matching results from No. 3 setting in Table 3, here and after. The direct matching results are prone to error when query and support instances are different, e.g., the dog on the right part of the first row. We also visualize how the decoder corrects the predictions in Fig. 6. For the alpaca face, the inner left canthus (red point) is close to the annotation of the right inner canthus (yellow point) due to the occlusion. Hence, the similarity-aware position proposal for the left inner canthus is wrongly generated at the right inner canthus. The decoder gradually pushes it to the correct position. + +Representation. Here we analyze the effectiveness of query-support joint refine encoder and the support keypoint identifier. i) Comparing No.1 with No.2 in Table 3, the QSR encoder brings an absolute improvement of $2.54\%$ on PCK over DETR encoder, which suggests the effectiveness of support-query mutual information fusion. 2) Comparing No.2 with No.3 in Table 3, introducing the support iden + +![](images/49655a490a2fc6a8b7f47d8b0e479d4b7a84423acdc3f3a2f6c55247ad36eb47.jpg) +Support Image + +![](images/8b8f5e90a1c83c3ae768f86ef5ce50be6820ef883b1ad9b2177ad7d9c712ea30.jpg) +Figure 6. The decoder can help correct unreliable proposals. The left column shows the support keypoint (red point) and the right column shows the refining process in the second stage. Note that in the first row, the left inner canthus is invisible due to occlusion, which looks to be on the right. + +![](images/5926882196251118edc0e7c495bba333263908760c7d441f7bcc445067669b55.jpg) +Query Image + +![](images/ee5e1e0fb09c78a3f0542a202893398dfb18b151f32513b116ab76901f010a3a.jpg) + +![](images/149c9adb2e71a4a7c378bd6b8b44a82e8de76413e78ce8929fa25c4f4f3b9a52.jpg) +Figure 7. Visualization of attention map. We visualize the encoder and decoder attention for the support keypoint at the left inner canthus. The support keypoint coarsely attends to keypoints with similar appearance in the encoder while focusing more on the corresponding keypoints for position update in the decoder. + +Table 4. Analysis of different terms in the similarity loss. + +
LhLePCK
✓--88.74
--✓89.41
✓✓89.45
+ +Table 5. Comparison of different support keypoint identifiers. + +
TypePCK
--85.81
Learnable88.96
Fixed89.45
+ +tifier improves the PCK by $2.46\%$ . Note that the No. 3 model excludes the decoder, which is one-stage. Its performance can still achieve state-of-the-art on split1 test set, which suggests that the support identifier and QSR decoder significantly improve representation quality. + +To intuitively show how the encoder and decoder work, we visualize the encoder/decoder attention maps in Fig. 7. The attention shows which location the support keypoint will focus during the information fusion (first stage) and the proposal correction (second stage). As can be observed, in the encoder, the support keypoint attends to keypoints with similar appearance, while in the decoder, the attention becomes more focused on the corresponding keypoints for proposal correction. + +Loss function. We analyze the similarity loss in Table 4. i) Only using the heatmap supervision (row 1) yields the worst performance, indicating that the expectation of loss is beneficial for precise localization. ii) As shown in row 3, the combination of heatmap loss and expectation attains the best performance, as both the shape of similarity map and the proposal position can be supervised. + +Support keypoint identifier. We compare two different implementations of support keypoint identifier: learnable embedding and fixed sinusoidal embedding in Table 5. Using fixed sinusoidal embedding improves PCK by $0.49\%$ compared with learnable embedding. The learnable embedding is similar to the learnable query embedding in DETR-based detection methods [5,25], which can be viewed as a learnable content embedding or anchors [23]. However, our + +![](images/571c800b7eee5e00e91b5c70f8dae652f57ede0ba2af04b4939fa5204e838e43.jpg) +Support Image &Keypoint + +![](images/868daeb6585332de87de355dfa487dd85070ee3761b2bdef7802a95cfee5014a.jpg) +Encoder Attention + +![](images/021f9e381df4d11926700c097dd789fdeb9f28c6e215a050f0aafbc4735f254c.jpg) +Decoder Cross-Attention + +Table 6. Inference and Training Efficiency. We test the efficiency on RTX3090 under 1-shot setting. + +
MethodParams (M)GFLOPsFPSMemory (GB)
POMNet [40]48.2138.016.8013.8 × 2
One-Stage26.8622.6536.907.5
CapeFormer31.1423.6826.097.8
+ +support keypoint identifier is fixed and serves a different purpose: distinguishing different support keypoints. Learnable support identifier deteriorates the performance as they may overfit the training categories more easily. + +# 4.5. Performance and Efficiency Trade-off + +We evaluate the efficiency under 1-shot setting for POMNet and CapeFormer in Table 6. The training memory is evaluated with a batchsize of 16. i) Comparing row 1 and 3, CapeFormer significantly reduces the parameters $(-35.4\%)$ and FLOPs $(-37.7\%)$ compared with POMNet due to the shared backbone and efficient similarity metric. ii) Comparing rows 2 and 3, the second stage only introduces limited parameters and FLOPs, which shows a satisfying trade-off between performance and efficiency. iii) CapeFormer is efficient for training, which consumes only 8GB memory on a single GPU (almost $1/4$ of POMNet). + +# 5. Conclusion + +In this work, we formulate a two-stage framework for category-agnostic pose estimation (CAPE) and instantiate it with a novel transformer model. Unlike previous one-stage approaches that entirely rely on matching results, the matching results in the first stage are viewed as similarity-aware position proposals which are refined in the second stage. We also propose specific designs to improve the representation and similarity modeling. On the CAPE benchmark MP-100, our method outperforms the previous best method by large margins on both accuracy and efficiency. + +Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (Grant No. U1913602 and 61876211). + +# References + +[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter V. Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 3686-3693, 2014. 2 +[2] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv Computer Research Repository, abs/1607.06450, 2016. 5 +[3] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9497-9506, 2019. 2 +[4] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1):172-186, 2018. 2 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of European Conference on Computer Vision (ECCV), pages 213-229, 2020. 4, 7, 8 +[6] Seokju Cho, Sunghwan Hong, Sangryul Jeon, Yunsung Lee, Kwanghoon Sohn, and Seungryong Kim. Cats: Cost aggregation transformers for visual correspondence. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 9011-9023, 2021. 3 +[7] Seokju Cho, Sunghwan Hong, and Seungryong Kim. Cats++: Boosting cost aggregation with convolutions and transformers. arXiv Computer Research Repository, abs/2202.06817, 2022. 3 +[8] Christopher B. Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Krishna Chandraker. Universal correspondence network. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 2406-2414, 2016. 3 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 248-255, 2009. 6 +[10] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 4 +[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of International Conference on Machine Learning (ICML), pages 1126–1135, 2017. 2, 6 +[12] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 14671-14681, 2021. 2 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 770-778, 2016. 6 + +[14] Shuqiang Jiang, Sisi Liang, Chengpeng Chen, Yaohui Zhu, and Xiangyang Li. Class agnostic image common object detection. IEEE Trans. Image Process., 28(6):2836-2846, 2019. 2 +[15] Seungryong Kim, Dongbo Min, Bumsub Ham, Stephen Lin, and Kwanghoon Sohn. FCSS: fully convolutional self-similarity for dense semantic correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):581-595, 2019. 3 +[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of International Conference on Learning Representations, 2014. 6 +[17] Rollyn Labuguen, Jumpei Matsumoto, Salvador Blanco Negrete, Hiroshi Nishimaru, Hisao Nishijo, Masahiko Takada, Yasuhiro Go, Ken ichi Inoue, and Tomohiro Shibata. Macaquepose: A novel "in the wild" macaque monkey pose dataset for markerless motion capture. Frontiers in behavioral neuroscience (2021), 2021. 2 +[18] Jiefeng Li, Siyuan Bian, Ailing Zeng, Can Wang, Bo Pang, Wentao Liu, and Cewu Lu. Human pose regression with residual log-likelihood estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 11005–11014, 2021. 2 +[19] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 10863-10872, 2018. 2 +[20] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: A benchmark for amur tiger re-identification in the wild. In Proceedings of ACM International Conference on Multimedia, pages 2590–2598, 2020. 2 +[21] Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Tokenpose: Learning keypoint tokens for human pose estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 11293-11302, 2021. 2 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), pages 740-755, 2014. 6 +[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In Proceedings of International Conference on Learning Representations, 2022. 5, 8 +[24] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Proceedings of Asian Conference on Computer Vision (ACCV), pages 669–684, 2019. 2 +[25] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3631-3640, 2021. 8 +[26] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel flow: Semantic correspondence with multi-layer + +neural features. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 3394-3403, 2019. 3 +[27] Akihiro Nakamura and Tatsuya Harada. Revisiting finetuning for few-shot learning. arXiv Computer Research Repository, abs/1910.00216, 2019. 6 +[28] Xuecheng Nie, Jianfeng Zhang, Shuicheng Yan, and Jiashi Feng. Single-stage multi-person pose machines. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 6950-6959, 2019. 2 +[29] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 1906-1915, 2018. 2 +[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 3 +[31] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomás Pajdla, and Josef Sivic. Neighbourhood consensus networks. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 1658-1669, 2018. 3 +[32] Christos Sagonas, Epameinondas Antonakos, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: database and results. Image Vis. Comput., 47:3-18, 2016. 6 +[33] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 11059-11068, 2022. 2 +[34] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 9529–9538, 2022. 2, 4 +[35] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4077-4087, 2017. 6 +[36] Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5452-5462, 2019. 2 +[37] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 2621-2630, 2017. 2 +[38] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 1653-1660, 2014. 2 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia + +Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 4 +[40] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Proceedings of European Conference on Computer Vision (ECCV), volume abs/2207.10387, 2022. 1, 2, 4, 6, 7, 8 +[41] Shuo Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5525-5533, 2016. 6 +[42] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2878-2890, 2013. 6 +[43] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5217–5226, 2019. 2 +[44] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 7091-7100, 2020. 2 +[45] Song-Hai Zhang, Ruilong Li, Xin Dong, Paul L. Rosin, Zixi Cai, Han Xi, Dingcheng Yang, Haozhi Huang, and Shi-Min Hu. Pose2seg: Detection free human instance segmentation. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 889-898, 2019. 2 +[46] Jian Zhao, Jianshu Li, Yu Cheng, Li Zhou, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In Proceedings of ACM International Conference on Multimedia, pages 792-800, 2018. 2 \ No newline at end of file diff --git a/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/images.zip b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ce3b2e8c738f7091437dca1daf5200087dc10698 --- /dev/null +++ b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ecc6e2d7cefefcf39a222c939d5d322ba22fcdc31f8f127e1fd23f858b89dec +size 656155 diff --git a/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/layout.json b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..187f465435eef10a4346bee71c235ed762800dd1 --- /dev/null +++ b/2023/Matching Is Not Enough_ A Two-Stage Framework for Category-Agnostic Pose Estimation/layout.json @@ -0,0 +1,9156 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 103, + 455, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 103, + 455, + 140 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 455, + 140 + ], + "type": "text", + "content": "Matching Is Not Enough: A Two-Stage Framework for Category-Agnostic Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "spans": [ + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "text", + "content": "Min Shi" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "text", + "content": " Zihao Huang" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "text", + "content": " Xianzheng Ma" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "text", + "content": " Xiaowei Hu" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "text", + "content": " Zhiguo Cao" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "text", + "content": " School of AIA, Huazhong University of Science and Technology \n" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 165, + 492, + 209 + ], + "type": "text", + "content": " Shanghai AI Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 211, + 505, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 211, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 82, + 211, + 505, + 224 + ], + "type": "text", + "content": "{min_shi, zihaohuang, zgcao}@hust.edu.cn, {maxianzheng, huxiaowei}@pjlab.org.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 277, + 290, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 290, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 290, + 577 + ], + "type": "text", + "content": "Category-agnostic pose estimation (CAPE) aims to predict keypoints for arbitrary categories given support images with keypoint annotations. Existing approaches match the keypoints across the image for localization. However, such a one-stage matching paradigm shows inferior accuracy: the prediction heavily relies on the matching results, which can be noisy due to the open set nature in CAPE. For example, two mirror-symmetric keypoints (e.g., left and right eyes) in the query image can both trigger high similarity on certain support keypoints (eyes), which leads to duplicated or opposite predictions. To calibrate the inaccurate matching results, we introduce a two-stage framework, where matched keypoints from the first stage are viewed as similarity-aware position proposals. Then, the model learns to fetch relevant features to correct the initial proposals in the second stage. We instantiate the framework with a transformer model tailored for CAPE. The transformer encoder incorporates specific designs to improve the representation and similarity modeling in the first matching stage. In the second stage, similarity-aware proposals are packed as queries in the decoder for refinement via cross-attention. Our method surpasses the previous best approach by large margins on CAPE benchmark MP-100 on both accuracy and efficiency. Code available at github.com/flyinglynx/CapeFormer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "content": "Humans can quickly grasp the essentials of a keypoint for arbitrary objects (e.g., the nose of an animal) and then pinpoint the same keypoint on another object in the same category. However, most pose estimation models are still category-specific, which cannot be applied to a new category unless they are trained with sufficient category-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 250, + 542, + 378 + ], + "blocks": [ + { + "bbox": [ + 307, + 250, + 542, + 378 + ], + "lines": [ + { + "bbox": [ + 307, + 250, + 542, + 378 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 542, + 378 + ], + "type": "image", + "image_path": "ba94522789afd8436d3a534b5462e9eb7ffab4494e9cfd7fc56f804c76778680.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 311, + 380, + 536, + 497 + ], + "blocks": [ + { + "bbox": [ + 311, + 380, + 536, + 497 + ], + "lines": [ + { + "bbox": [ + 311, + 380, + 536, + 497 + ], + "spans": [ + { + "bbox": [ + 311, + 380, + 536, + 497 + ], + "type": "image", + "image_path": "2c8ed89129cf639a07d41f0c75eb1a7b99aaa02c95276caf0232300b712a50a3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 500, + 547, + 567 + ], + "lines": [ + { + "bbox": [ + 304, + 500, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 547, + 567 + ], + "type": "text", + "content": "Figure 1. Comparison with existing methods. (a) Compared with previous one-stage matching paradigm that entirely relies on keypoint matching, we propose a novel two-stage framework, which learns to correct unreliable matching results in the second stage. (b) The proposed approach attains better accuracy and efficiency compared with the previous best approach." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 579, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 547, + 640 + ], + "type": "text", + "content": "specific data. Recently, category-agnostic pose estimation (CAPE) [40] is proposed to address such a generic pose estimation problem. CAPE aims to localize instance keypoints for arbitrary categories given one or few support images with keypoint annotations (support keypoints)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Unlike category-specific pose estimation models that learn to recognize specific keypoints for a fixed category, CAPE models learn to represent and compare keypoints for open-world categories. As shown in Fig. 1 (a), existing CAPE approaches [40] match the support keypoints across the query image in an embedding space. Then, the match-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 217, + 713 + ], + "type": "text", + "content": "*: Equal contribution. †: Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7308" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "ing results are used as clues for keypoint localization. Although a sophisticated feature embedding pipeline is elaborated [40], we find that such a one-stage pipeline shows inferior accuracy and efficiency for CAPE tasks. Specifically, the prediction heavily relies on matching results. However, due to the open set nature in CAPE, the similarity results often contain some noises: 1) insufficient similarity when support and query instances differ significantly in poses, textures, or styles; 2) false positive similarity is triggered when points in the query image share similar appearance with support keypoints. For example, one-stage matching paradigm has difficulties distinguishing the left and right of keypoints due to similar visual characteristics. Thus, the left eye in support keypoints can be indiscriminately predicted to the right eye in the query image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 289, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 339 + ], + "type": "text", + "content": "In this paper, we remedy these issues from two aspects. First, to refine the matching results, we introduce a two-stage framework for CAPE, where matched keypoints in the first stage are viewed as similarity-aware position proposals. The model learns to refine the proposals in the second stage. In addition, to construct a robust matching process, we also improve the representation quality and similarity metrics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 289, + 592 + ], + "type": "text", + "content": "Specifically, we instantiate the two-stage framework with a transformer, termed CAPE transformer (CapeFormer). The transformer encoder encodes the support keypoints and query images in the first stage, with specific designs to improve the representation quality. We first design a query-support joint refine encoder to mutually transfer information among support keypoints and the query images, thus narrowing the gap between the support and query instances (poses or styles). Then, we add a support keypoint identifier to alleviate the ambiguity between two support keypoints when they are geometrically close or with similar appearance. After the feature encoding, similarity-aware position proposals are generated for the second stage. We directly match the support keypoint with query image features using a learnable inner-product [34]. The similarity peaks are selected as the similarity-aware position proposals for support keypoints. In the second stage, the position proposals are packed with support keypoint features as the queries in the transformer decoder. In each decoder layer, queries extract relevant features with cross-attention, which are used as clues to update the proposals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "We evaluate our method on a category-agnostic pose estimation benchmark MP-100 [40]. Our method outperforms the previous best approach POMNet [40] by large margins, with an improvement of " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " under 1-shot and 5-shot settings, respectively. When training and test categories almost have no properties in common, a more significant improvement (up to " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ") can be obtained compared with POMNet. Note that, our method is trained end-to-end and shows better efficiency compared with previous best approach, as in Fig. 1 (b)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 72, + 518, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 518, + 85 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 518, + 85 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 88, + 547, + 206 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 88, + 547, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 88, + 547, + 123 + ], + "spans": [ + { + "bbox": [ + 317, + 88, + 547, + 123 + ], + "type": "text", + "content": "- We propose a two-stage framework for category-agnostic pose estimation tackling the noisy matching results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 129, + 547, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 547, + 177 + ], + "type": "text", + "content": "- We instantiate the two-stage framework with a transformer model, termed CapeFormer, with specific designs to enhance the representation and similarity modeling in the matching pipeline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 181, + 547, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 181, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 181, + 547, + 206 + ], + "type": "text", + "content": "- CapeFormer outperforms the previous best approach significantly in both accuracy and efficiency." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 236, + 404, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 404, + 249 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 404, + 249 + ], + "type": "text", + "content": "2.1. Pose Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 255, + 547, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 423 + ], + "type": "text", + "content": "Pose estimation aims to localize the semantic keypoints of instances. Most pose estimation approaches are targeted for one specific category, e.g., human [1, 19, 45, 46], animals [3, 17, 20], or vehicles [29, 36]. According to how points are localized, pose estimation methods can be categorized into regression-based [18, 21, 28, 37, 38], heatmap-based [4, 44], offset-based [12] and query-based [33] ones. In this work, we address category-agnostic pose estimation, which has not been well-studied yet. Compared with category-specific pose estimation, category-agnostic pose estimation focuses on constructing a generic representation and similarity metrics: keypoints are predicted by comparing the support keypoints and the query images in the embedding space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 431, + 486, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 486, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 486, + 445 + ], + "type": "text", + "content": "2.2. Category-Agnostic Vision Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "Category-agnostic vision models aim to build a generic model that can be applied to any categories on a particular task, e.g., object detection [14], segmentation [43], object counting and localization [24, 34]. This setting is also known as the few-shot learning. Recently, POMNet [40] first proposes the category-agnostic pose estimation task and elaborates a dataset MP-100. Most category-agnostic models can be classified into meta-learning and metric-learning approaches. Meta-learning approaches [11] aim to initialize the model with a set of optimal parameters, where the model can quickly fit the new tasks by learning from a few support samples. Metric learning-based approaches match the support samples with query images in an embedding space. Then, the matching results are used as clues for subsequent tasks. For pose estimation, POMNet [40] encodes query images and support keypoints with a transformer. Then, a regression head directly infers the similarity from the concatenated support keypoint and query image features. We find that simply using matching results for point localization is challenging. Hence, we extend such a matching paradigm with a two-stage framework to correct the unreliable matching results for accurate predictions." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7309" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 541, + 247 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 541, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 541, + 247 + ], + "type": "image", + "image_path": "9e963a7607f3a087c4daa28ca0973cc5c1f08a2bdbae363c315cdb737ca54e22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 547, + 289 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed CapeFormer. CapeFormer consists of a matching stage where similarity-aware position proposals are generated, and a refine stage that further calibrates the proposals. A support identifier and query-support joint refine encoder are proposed to improve the representation quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 301, + 194, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 194, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 194, + 316 + ], + "type": "text", + "content": "2.3. Semantic Correspondence" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 320, + 290, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 290, + 513 + ], + "type": "text", + "content": "Semantic correspondence (SC) [8, 15] aims to match the points that have identical semantic meaning with a dense correspondence field. Hence, SC models can be directly applied on the CAPE: localize the corresponding points for the support keypoints with the estimated correspondence field. Typical SC models first match each pair of pixels and use the matching score to construct correspondence. Similar to our two-stage design that alleviates the noise on similarity matching, a series of SC models refine the matching scores with geometry constraints [26], high-dimensional convolutions [31], and attention mechanism [6, 7]. However, we find directly applying SC models for CAPE is inefficient and yields inferior performance. Besides, most techniques for the matching refine in SC are based on the dense matching volume, which is not available in CAPE, as only the sparse support keypoints are matched." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 518, + 165, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 165, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 165, + 533 + ], + "type": "text", + "content": "3. CAPE Transformer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "text", + "content": "Class-agnostic pose estimation (CAPE) aims to estimate instance keypoints given one or a few support images with keypoint annotations. Specifically, we denote the query image as " + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "text", + "content": " and the support image as " + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "text", + "content": ", which has " + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "text", + "content": " predefined support keypoints. The CAPE model will infer the positions of the corresponding " + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 539, + 287, + 635 + ], + "type": "text", + "content": " keypoints in the query images. For simplicity, we use the 1-shot setting for illustration (the number of support images equals one)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 637, + 228, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 228, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 228, + 651 + ], + "type": "text", + "content": "3.1. Two-Stage Framework for CAPE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "Different from the previous one-stage paradigm, Cape-Former features a two-stage framework. As in Fig. 2, the first stage consists of the input embedding, a support-query joint refine encoder, and a similarity-aware proposal generator. A proposal refine decoder forms the second stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": "The input embedding encodes the query image and support keypoints into the embedding space (Sec. 3.2). We denote the features for the query image and support keypoints as " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": ", which are further refined with the query-support joint refine encoder (Sec. 3.3). The proposal generator measures the similarity between " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " and the similarity peaks are selected as the similarity-aware position proposals (Sec. 3.4). Then, in the second stage, the similarity-aware proposals are gradually refined into the final predictions (Sec. 3.5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 431, + 411, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 411, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 411, + 445 + ], + "type": "text", + "content": "3.2. Input Embedding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "content": "The input embedding is illustrated in Fig. 3. We first project the support and query image into feature maps with a shared backbone. The flattened feature map of the query image is used as its representation, denoted by " + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "inline_equation", + "content": "F_{q} \\in \\mathbb{R}^{c \\times n}" + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "content": " equals the number of pixels in the feature map and " + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "content": " denotes the embedded dimension. To obtain the features of the support keypoint, we conduct a \"soft\" ROI pooling [30] on the support image feature map. " + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "content": " soft masks are generated by placing a Gaussian kernel centered at support keypoints, and then multiplied with the support feature map. The summation of masked support feature map is adopted as the support keypoint feature " + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "inline_equation", + "content": "F_{s} \\in \\mathbb{R}^{K \\times c}" + }, + { + "bbox": [ + 304, + 449, + 547, + 617 + ], + "type": "text", + "content": ". For multiple support images, e.g., 5-shot settings, the support keypoint features from different images are averaged." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Support keypoint identifier. When encoding support keypoints, each support keypoint's positional and context information is discarded during ROI pooling. Hence, when two support keypoints are geometrically close or with similar appearance (e.g., left and right eyes), they can be indistinguishable in the embedding space. Such ambiguity can lead to inaccurate predictions that are duplicated or mirror-symmetric to the ground truth." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7310" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 190, + 243 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 190, + 243 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 190, + 243 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 190, + 243 + ], + "type": "image", + "image_path": "f3db0a27cd08c797aedf302aafb97a14adf813dbb4972189e33f669638f5ca65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 246, + 287, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 246, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 287, + 269 + ], + "type": "text", + "content": "Figure 3. Illustration of the input embedding and the support-query joint refine encoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 194, + 72, + 284, + 243 + ], + "blocks": [ + { + "bbox": [ + 194, + 72, + 284, + 243 + ], + "lines": [ + { + "bbox": [ + 194, + 72, + 284, + 243 + ], + "spans": [ + { + "bbox": [ + 194, + 72, + 284, + 243 + ], + "type": "image", + "image_path": "9a6a8220e256e82432ffb0c951708c315268af0fe7a2fcc638cb9884e7e826fd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "type": "text", + "content": "To alleviate such ambiguity, we propose a simple but effective support keypoint identifier that gives an \"identity\" to each support keypoint. The implementation is embarrassingly simple: " + }, + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "type": "text", + "content": " embedding vectors are generated via sinusoidal encoding, denoted by " + }, + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "type": "inline_equation", + "content": "I_{s} \\in \\mathbb{R}^{K \\times c}" + }, + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "type": "text", + "content": ". This identifier is added to " + }, + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 46, + 297, + 289, + 405 + ], + "type": "text", + "content": " ahead of attention calculation, which will be discussed later in this section. Such a simple modification can lead to over 3% improvement with negligible computation overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 422, + 217, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 217, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 217, + 435 + ], + "type": "text", + "content": "3.3. Query-Support Refine Encoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": "Directly matching the support keypoints with the query image right after the input embedding is challenging, as instances in the support and query images can be different, such as different colors, textures, and poses. Thus, the target and support keypoints may not possess sufficient similarity in the embedding space [10, 34]. An intuitive way to alleviate this is to fuse relevant query and support features. Hence, the query and support features can be pulled closer in the embedding space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": "To this end, we implement information fusion and transfer among support and query with self-attention. As illustrated in Fig. 3, the inputs of the " + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": " encoder layer are query and support features, denoted as " + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{q}^{l}" + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{s}^{l}" + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": ". We augment the " + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{q}^{l}" + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": " with positional embedding as in DETR [5], and add the support keypoint identifier into the " + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{s}^{l}" + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": ". The augmented " + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{q}^{l}" + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{s}^{l}" + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": " are concatenated into a feature sequence with the length of " + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "inline_equation", + "content": "n + K" + }, + { + "bbox": [ + 46, + 555, + 288, + 714 + ], + "type": "text", + "content": ". The sequence is then fed into the standard multi-head self-attention layer [39]. The output sequence from the self-attention layer is processed by a feed forward network (FFN). Then, the sequence are re-split to the support keypoint and query image features, which are used as the inputs for the next encoder layer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 504, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 504, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 504, + 85 + ], + "type": "text", + "content": "3.4. Similarity-Aware Proposal Generator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 90, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 256 + ], + "type": "text", + "content": "The similarity-aware proposal generator matches the support keypoint features with the query features to obtain the similarity maps, whose peaks are selected as the similarity-aware proposals. Previous state-of-the-art approach POMNet [40] measures similarity implicitly: a regression head directly infers similarity maps (heatmaps) from the concatenation of support keypoint and query features. However, regressing similarity without explicit similarity modeling has the risk of overfitting in class-agnostic setting [34]. Moreover, considerable computation overhead is introduced. For example, given 100 support keypoint features with 256 channels as in our case, the similarity maps are obtained by forwarding 512-channel feature maps with an FCN for 100 times." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "type": "text", + "content": "To embrace both efficiency and generality, we use a learnable inner-product [34] to model similarity explicitly. Specifically, we first calculate a channel attention weight " + }, + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{R}^{K \\times c}" + }, + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "type": "inline_equation", + "content": "F_s \\in \\mathbb{R}^{K \\times c}" + }, + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "type": "text", + "content": ". Then, the similarity map " + }, + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 257, + 545, + 318 + ], + "type": "text", + "content": " is obtained as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 324, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 324, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 367, + 324, + 545, + 340 + ], + "type": "interline_equation", + "content": "M = \\left(W _ {q} F _ {q}\\right) \\left(\\boldsymbol {a} \\circ W _ {s} F _ {s}\\right) ^ {T}, \\tag {1}", + "image_path": "27bf2905441171cd20ed12b3eabb9edc0c3bd325e780a7e062c0ac85847d5f09.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "W_{s}" + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "text", + "content": " are two learnable projection matrices in " + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{c\\times c}" + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "text", + "content": " denotes the Hadamard product. Then, similarity maps " + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "M\\in \\mathbb{R}^{n\\times K}" + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "text", + "content": " are reshaped into 2D with the shape of " + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 347, + 545, + 407 + ], + "type": "text", + "content": ". The channel attention helps stressing the key patterns during inner product computation [34]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "text", + "content": "The maximums from the similarity maps are selected as the similarity-aware position proposals for support keypoints. To reduce the round-off errors on the down-scaled resolutions, instead of directly selecting the maximums, we first generate a " + }, + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "text", + "content": " local window around the similarity peaks and normalize this local window with softmax. Then, we take the expectation of the coordinates in each local window as the proposal, which can be formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 520, + 545, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 520, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 350, + 520, + 545, + 550 + ], + "type": "interline_equation", + "content": "x _ {k} = \\sum_ {i, j \\in W _ {k}} i \\cdot \\frac {\\exp (M _ {k , i j})}{\\sum_ {i , j \\in W} \\exp (M _ {i j})}, \\tag {2}", + "image_path": "cd16bdabe79dd0dcb4552765757da0664a3d17c39e5bb356bf5186317d52a899.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 352, + 552, + 545, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 552, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 352, + 552, + 545, + 583 + ], + "type": "interline_equation", + "content": "y _ {k} = \\sum_ {i, j \\in W _ {k}} j \\cdot \\frac {\\exp (M _ {k , i j})}{\\sum_ {i , j \\in W} \\exp (M _ {i j})}, \\tag {3}", + "image_path": "8dbb9892518909580959a8bcbc0a9e47a6aeae7f1e44488367dd9f09a451aaee.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "text", + "content": " denotes the local window for the " + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "text", + "content": " support keypoint and " + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "text", + "content": " indexes the horizontal and vertical coordinates. Here we denote the similarity-aware position proposals for all the support keypoints as " + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{K \\times 2}" + }, + { + "bbox": [ + 304, + 592, + 545, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 647, + 444, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 444, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 444, + 659 + ], + "type": "text", + "content": "3.5. Proposal Refine Decoder" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "As mentioned in Sec. 1, due to the unreliable matching process, directly using the matched keypoints from the first stage as predictions is prone to error. Intuitively, each support keypoint can be corrected by jointly considering other" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7311" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 262, + 280 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 262, + 280 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 262, + 280 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 262, + 280 + ], + "type": "image", + "image_path": "5bf9a158d390d6d1573f62937d46ec31a518925e7d268599ce64393d6c375625.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 287, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 308 + ], + "type": "text", + "content": "Figure 4. Illustration of the proposal refine decoder. The decoder consists of a self-attention and a cross-attention layer" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 318, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 366 + ], + "type": "text", + "content": "support keypoint predictions and keypoint-related information in the query image. Hence, we devise the second stage to unleash the potential of correcting the raw similarity-aware position proposals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 4, we design a proposal refine decoder, where the support keypoint embedding together with the keypoint positions are fed into the decoder layer. Each layer consists of two steps: 1) self-attention among support keypoints makes each proposal aware of other keypoints' positions and contents; 2) cross-attention that helps extract relevant content from the query feature for each support keypoint. By integrating these two attention mechanisms, the decoder can further refine the position proposals and produce more accurate predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "text", + "content": "Self-attention among support keypoints. Self-attention enables interactions among the support keypoints, which makes each proposal aware of other keypoints' positions and contents. To prepare the inputs for the self-attention layer, the support keypoint positions are first converted into sinusoidal positional encoding " + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "inline_equation", + "content": "E_{s}^{l} \\in \\mathbb{R}^{K \\times c}" + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "text", + "content": ". Then, the support keypoint identifier " + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "text", + "content": " and position embedding " + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "inline_equation", + "content": "E_{s}^{l}" + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "text", + "content": " are added into " + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "inline_equation", + "content": "F_{s}^{l}" + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "text", + "content": " forming the query and key for the first self-attention layer. The support keypoint feature " + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "inline_equation", + "content": "F_{s}^{l}" + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "text", + "content": " is directly used as the value. The multi-head self-attention layer is followed by a transformer-style feed-forward network (FFN) with residual connection and layer normalization [2]. We denote the support feature processed by the FFN as " + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "inline_equation", + "content": "F_{s}^{l'}" + }, + { + "bbox": [ + 46, + 486, + 287, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "Cross-attention between support and query features. Cross-attention helps extract relevant contents from the query feature " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " for each support keypoint. As illustrated in Fig. 4, we use the concatenation of " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "F_{s}^{\\prime}" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " and the summation of " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "E_{s}^{l}" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " as the query for the cross-attention layer. Correspondingly, query image features " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " and positional em" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "bedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "P_{q}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " are concatenated as the key, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " are used as the value. Instead of addition, we concatenate the positional embedding and support identifier for key and value, following DAB-DETR [23]. An FFN processes the output from the cross-attention layer to form the support keypoint features for the next decoder layer, i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "F_{s}^{l + 1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "text", + "content": "Position update. We use the refined support feature " + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "inline_equation", + "content": "F_{s}^{l + 1}" + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "text", + "content": " to update the keypoint positions " + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "inline_equation", + "content": "P^l" + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "inline_equation", + "content": "P^{l + 1}" + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "text", + "content": ". An MLP takes " + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "inline_equation", + "content": "F_{s}^{l + 1}" + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "text", + "content": " as input and infers the offset " + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "inline_equation", + "content": "O^{l + 1}\\in \\mathbb{R}^{K\\times 2}" + }, + { + "bbox": [ + 304, + 144, + 545, + 204 + ], + "type": "text", + "content": " pointing to the ground truth. Then, the keypoint positions are updated by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 365, + 212, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 212, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 365, + 212, + 545, + 228 + ], + "type": "interline_equation", + "content": "P ^ {l + 1} = \\sigma \\left(O ^ {l} + \\sigma^ {- 1} \\left(P ^ {l}\\right)\\right), \\tag {4}", + "image_path": "bb56dd96fa8b1d5c5186fbe50ad9d49d29a5aa7313dd73d1b0dacba8217a95f9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\sigma^{-1}" + }, + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "type": "text", + "content": " denote the sigmoid and its inverse function. Note that the similarity-aware proposals are used as the keypoint positions for the first decoder layer, i.e., " + }, + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "type": "inline_equation", + "content": "P^1" + }, + { + "bbox": [ + 304, + 236, + 545, + 296 + ], + "type": "text", + "content": ". The updated keypoint positions from the last decoder layer are used as the final prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 303, + 416, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 416, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 416, + 316 + ], + "type": "text", + "content": "3.6. Supervision Signal" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 321, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 380 + ], + "type": "text", + "content": "CapeFormer incorporates two supervision signals: the similarity loss and the offset loss. The similarity loss facilitates the learning of representation and the similarity metric in the first stage. The offset loss supervises the learning of proposal refinement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "text", + "content": "Similarity loss. The similarity loss is imposed on the similarity maps " + }, + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "text", + "content": " from the similarity-aware proposal generator. We simultaneously constrain the shape of similarity map and the accuracy of similarity-aware proposals with a heatmap term " + }, + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "text", + "content": " and an expectation term " + }, + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "type": "text", + "content": "For heatmap loss, we use sigmoid to normalize similarity map " + }, + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "type": "text", + "content": " are computed with " + }, + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 441, + 545, + 466 + ], + "type": "text", + "content": " loss as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 474, + 545, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 474, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 331, + 474, + 545, + 507 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {1}{H \\cdot W} \\| \\text {s i g m o i d} (M _ {i}) - H _ {i} \\| _ {2}, \\tag {5}", + "image_path": "e24c77d3e4571094ade236bd273b2ff371d93b88009ad6a16334fbfa365a227a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "type": "text", + "content": " denotes the ground truth heatmap for the " + }, + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "type": "text", + "content": " keypoint. For " + }, + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 304, + 516, + 545, + 562 + ], + "type": "text", + "content": ", we first normalize the similarity map with softmax and then take the expectation of the coordinates as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 562, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 562, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 342, + 562, + 545, + 594 + ], + "type": "interline_equation", + "content": "E \\left(P _ {k}\\right) = \\sum_ {j = 1} ^ {H} \\sum_ {i = 1} ^ {W} \\frac {\\exp \\left(M _ {k , i j}\\right)}{\\sum \\exp \\left(M _ {k}\\right)} \\cdot (i, j), \\tag {6}", + "image_path": "995d5a52e05a23ed698a8fd3ebca6987f1a8a915154662febb93503faa7caa89.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "inline_equation", + "content": "E(P_k)" + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "text", + "content": " denotes the coordinate expectation for the " + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "text", + "content": " keypoint. " + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "text", + "content": " denote the height and width for similarity map. Then we use " + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "text", + "content": " loss to calculate " + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 304, + 600, + 545, + 638 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 364, + 646, + 545, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 364, + 646, + 545, + 678 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\left| E \\left(P _ {i}\\right) - \\hat {P} _ {i} \\right| _ {1}, \\tag {7}", + "image_path": "5ddc685bbaba30e888ad9644fc6c6472d1b274a3d76d8752ff53b85c79abb786.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{P}_i" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " denotes the ground truth coordinate of the " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " point." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7312" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 97, + 537, + 188 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Results on the MP-100 dataset. Performance (PCK) under 1-shot and 5-shot settings. The proposed CapeFormer achieves the best PCK on all the splits under 1-shot and 5-shot settings." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 97, + 537, + 188 + ], + "lines": [ + { + "bbox": [ + 55, + 97, + 537, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 537, + 188 + ], + "type": "table", + "html": "
Method1-shot5-shot
Split 1Split 2Split 3Split 4Split 5AverageSplit 1Split 2Split 3Split 4Split 5Average
ProtoNet [35]46.0540.8449.1343.3444.5444.7860.3153.5161.9258.4458.6158.56
MAML [11]68.1454.7264.1963.2457.2061.5070.0355.9863.2164.7958.4762.50
Fine-tune [27]70.6057.0466.0665.0059.2063.5871.6757.8466.7666.5360.2464.61
POMNet [40]84.2378.2578.1778.6879.1779.7084.7279.6178.0080.3880.8580.71
CapeFormer89.4584.8883.5983.5385.0985.3191.9488.9289.4088.0188.2589.30
", + "image_path": "dc617a87cd9bf815751701d2c9c224f84606be1356e13b62dc708c6edf015d21.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 201, + 288, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 288, + 227 + ], + "type": "text", + "content": "Offset regression loss. Supposing a total of " + }, + { + "bbox": [ + 46, + 201, + 288, + 227 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 201, + 288, + 227 + ], + "type": "text", + "content": " decoder layers are used, we define the offset regression loss " + }, + { + "bbox": [ + 46, + 201, + 288, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_o" + }, + { + "bbox": [ + 46, + 201, + 288, + 227 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 234, + 287, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 234, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 101, + 234, + 287, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\sum_ {i = 1} ^ {K} \\left| P _ {i} ^ {l + 1} - \\hat {P} _ {i} \\right| _ {1}. \\tag {8}", + "image_path": "fe48073945313531c047b1fca67fa3c94fb92b8ad71806cf087cadbc1aedd831.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 275, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 287, + 301 + ], + "type": "text", + "content": "The overall training loss is calculated as " + }, + { + "bbox": [ + 47, + 275, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\lambda_h \\cdot \\mathcal{L}_h + \\mathcal{L}_e + \\mathcal{L}_o" + }, + { + "bbox": [ + 47, + 275, + 287, + 301 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 275, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\lambda_h" + }, + { + "bbox": [ + 47, + 275, + 287, + 301 + ], + "type": "text", + "content": " is a pre-defined weight for the heatmap loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 311, + 128, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 128, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 128, + 324 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 330, + 181, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 181, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 181, + 342 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": "Network architecture. We use the ResNet-50 [13] pretrained from ImageNet [9] as the backbone following POMNet [40]. Different from POMNet, the backbone is shared by query and support images in our cases for efficiency. Both the query-support joint refine encoder and the proposal refine decoder have three layers. Refer to supplementary for more details on network architectures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "content": "Training strategy. We use Adam [16] optimizer to train the model for 200 epochs with the batch size of 16. The learning rate is set as " + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "content": " and decays by " + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "inline_equation", + "content": "160^{th}" + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "inline_equation", + "content": "180^{th}" + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "content": " epoch. The weight for heatmap loss is empirically set as 2.0 to make different loss terms lie in the same magnitude. Data augmentation and pre-processing are kept the same with POMNet [40] for fair comparisons." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 520, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 288, + 664 + ], + "type": "text", + "content": "Dataset and metric. We use MP-100 dataset [40] to train and evaluate our method. Samples are collected from the existing category-specific pose estimation datasets [22, 32, 41]. MP-100 dataset encompasses over 18k images from 100 different categories. Keypoint numbers of different categories vary from 8 to 68. To train and evaluate the model, the samples are organized into five splits. On each split, the training, validation, and test categories have no overlaps, i.e., the categories for evaluation are not accessed during training. We use the Probability of Correct Keypoint (PCK) [42] as the quantitative metric. The threshold for PCK is set to 0.2 following POMNet [40]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 671, + 159, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 159, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 159, + 683 + ], + "type": "text", + "content": "4.2. Results on MP-100" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "We compare our method with the previous best CAPE method POMNet [40] and three baselines: ProtoNet [35]," + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 227, + 545, + 306 + ], + "blocks": [ + { + "bbox": [ + 305, + 200, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 305, + 200, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 545, + 223 + ], + "type": "text", + "content": "Table 2. Cross super-category evaluation on MP-100 dataset. Experiments are conducted under the 1-shot setting." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 227, + 545, + 306 + ], + "lines": [ + { + "bbox": [ + 307, + 227, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 545, + 306 + ], + "type": "table", + "html": "
MethodHuman BodyHuman FaceVehicleFurniture
ProtoNet [35]37.6157.8028.3542.64
MAML [11]51.9325.7217.6820.09
Fine-tune [27]52.1125.5317.4620.76
POMNet [40]73.8279.6334.9247.27
CapeFormer83.4480.9645.4052.49
", + "image_path": "005199e105a2abf5589f9c540e07b240a3e9659290296f93c08d83b92c88964a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": "MAML [11], and Fine-tune [27]. PCKs of different approaches on the MP-100 dataset under 1-shot and 5-shot settings are reported in Table 1, where we can see our method significantly outperforms POMNet [40], and improves the average PCK by " + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": " under the 1-shot setting and " + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": " under the 5-shot setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 396, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 540 + ], + "type": "text", + "content": "1-shot CapeFormer can surpass other methods with five support images, demonstrating the robustness of the proposed two-stage framework when the support information is limited. When increasing the shots from one to five, CapeFormer obtains a significant improvement of " + }, + { + "bbox": [ + 304, + 396, + 546, + 540 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 304, + 396, + 546, + 540 + ], + "type": "text", + "content": " on PCK. For POMNet, the improvement from 1-shot to 5-shot is " + }, + { + "bbox": [ + 304, + 396, + 546, + 540 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 304, + 396, + 546, + 540 + ], + "type": "text", + "content": ". This shows that CapeFormer can better utilize the additional information when more support samples are available. Note that we use the same way as POMNet [40] to encode the support keypoint features for fair comparisons. We also include a comparison with the semantic correspondence approaches in the supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 550, + 511, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 550, + 511, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 550, + 511, + 562 + ], + "type": "text", + "content": "4.3. Cross Super-Category Pose Estimation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 568, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 676 + ], + "type": "text", + "content": "In the MP-100 dataset, training, validation, and test categories do not overlap. However, some categories may still share similar characteristics, such as the faces of different animals. To better evaluate the generalization ability on significantly different categories, we conduct a cross-supercategory evaluation following POMNet [40]. Four supercategories: human face, human body, vehicle, and furniture, are collected from the MP-100 dataset as the test categories, while the other categories are used as the training samples." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "As in Table 2, CapeFormer achieves the state-of-the-art performance. The advantages on the human body and vehicle are particularly remarkable, with an improvement" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7313" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 542, + 299 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 542, + 299 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 542, + 299 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 542, + 299 + ], + "type": "image", + "image_path": "29308e8337988c5ac90f05e95e6f46b8c56bf2d4a4405b1312a5f1bfdf253317.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 546, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 546, + 335 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our method. We visualize the keypoint predictions under 1-shot setting. The left column denotes the support image with corresponding keypoints. The middle column denotes the results without proposal refine decoder, and the right column are results from the full method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "9.62\\%" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "10.48\\%" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " on PCK, respectively. Compared with the standard setting, the improvement in cross supercategory setting is more significant, which further demonstrates the generalization ability and robustness of the proposed method. The improvement of the human face is limited " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "(1.36\\%)" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": ". A plausible reason is that the human face has much more keypoints (68) than the training categories (the largest keypoint number in training is 38). Such discrepancy renders great difficulties for generalization: the keypoints are densely arranged, and the training data does not contain such a fine-grained matching case." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 499, + 141, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 141, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 141, + 512 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 518, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 650 + ], + "type": "text", + "content": "In this section, we first ablate the two-stage framework and the proposed modules in Table 3. Then, we validate some detailed design choices, including the implementation of training loss and support keypoint identifier. We perform experiments on the test set of MP-100 split1 under the 1-shot setting following [40]. We set some variants of the encoder for ablations on network components. (i) DETR [5] encoder removes the query-support interactions, i.e., the information fusion among support keypoints and query images is disabled; (ii) QSR denotes the proposed query-support joint refine encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "content": "Two-stage framework. First we focus on the effectiveness of the second stage. Comparing No. 3 and No. 5 in Table 3, the decoder brings an improvement of " + }, + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "inline_equation", + "content": "4.13\\%" + }, + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "content": " on PCK, demonstrating the necessity of the second stage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "We compare one-stage and two-stage paradigms qualita-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 392, + 541, + 472 + ], + "blocks": [ + { + "bbox": [ + 305, + 355, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 355, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 545, + 389 + ], + "type": "text", + "content": "Table 3. Ablation study on different components in Cape-Former. Support ID stands for the support keypoint identifier in Sec. 3.2. QSR denotes our query-support joint refine encoder." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 392, + 541, + 472 + ], + "lines": [ + { + "bbox": [ + 310, + 392, + 541, + 472 + ], + "spans": [ + { + "bbox": [ + 310, + 392, + 541, + 472 + ], + "type": "table", + "html": "
Support IDEncoderDecoderParadigmPCK
No.1--DETR--one-stage80.32
No.2--QSR--one-stage82.86
No.3✓QSR--one-stage85.32
No.4--QSR✓two-stage85.81
No.5✓QSR✓two-stage89.45
", + "image_path": "599076dd8674ea8dbc2e88a3728cfced3f00773ee02305cef15ef806b17f7051.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 493, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 625 + ], + "type": "text", + "content": "tively in Fig. 5. \"One-Stage\" denotes the direct matching results from No. 3 setting in Table 3, here and after. The direct matching results are prone to error when query and support instances are different, e.g., the dog on the right part of the first row. We also visualize how the decoder corrects the predictions in Fig. 6. For the alpaca face, the inner left canthus (red point) is close to the annotation of the right inner canthus (yellow point) due to the occlusion. Hence, the similarity-aware position proposal for the left inner canthus is wrongly generated at the right inner canthus. The decoder gradually pushes it to the correct position." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "content": "Representation. Here we analyze the effectiveness of query-support joint refine encoder and the support keypoint identifier. i) Comparing No.1 with No.2 in Table 3, the QSR encoder brings an absolute improvement of " + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "inline_equation", + "content": "2.54\\%" + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "content": " on PCK over DETR encoder, which suggests the effectiveness of support-query mutual information fusion. 2) Comparing No.2 with No.3 in Table 3, introducing the support iden" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7314" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 81, + 134, + 144 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 124, + 80 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 124, + 80 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 124, + 80 + ], + "type": "text", + "content": "Support Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 81, + 134, + 144 + ], + "lines": [ + { + "bbox": [ + 65, + 81, + 134, + 144 + ], + "spans": [ + { + "bbox": [ + 65, + 81, + 134, + 144 + ], + "type": "image", + "image_path": "49655a490a2fc6a8b7f47d8b0e479d4b7a84423acdc3f3a2f6c55247ad36eb47.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 83, + 148, + 115, + 200 + ], + "blocks": [ + { + "bbox": [ + 83, + 148, + 115, + 200 + ], + "lines": [ + { + "bbox": [ + 83, + 148, + 115, + 200 + ], + "spans": [ + { + "bbox": [ + 83, + 148, + 115, + 200 + ], + "type": "image", + "image_path": "8b8f5e90a1c83c3ae768f86ef5ce50be6820ef883b1ad9b2177ad7d9c712ea30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 208, + 287, + 263 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 263 + ], + "type": "text", + "content": "Figure 6. The decoder can help correct unreliable proposals. The left column shows the support keypoint (red point) and the right column shows the refining process in the second stage. Note that in the first row, the left inner canthus is invisible due to occlusion, which looks to be on the right." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 81, + 205, + 143 + ], + "blocks": [ + { + "bbox": [ + 151, + 72, + 195, + 80 + ], + "lines": [ + { + "bbox": [ + 151, + 72, + 195, + 80 + ], + "spans": [ + { + "bbox": [ + 151, + 72, + 195, + 80 + ], + "type": "text", + "content": "Query Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 142, + 81, + 205, + 143 + ], + "lines": [ + { + "bbox": [ + 142, + 81, + 205, + 143 + ], + "spans": [ + { + "bbox": [ + 142, + 81, + 205, + 143 + ], + "type": "image", + "image_path": "5926882196251118edc0e7c495bba333263908760c7d441f7bcc445067669b55.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 142, + 146, + 190, + 202 + ], + "blocks": [ + { + "bbox": [ + 142, + 146, + 190, + 202 + ], + "lines": [ + { + "bbox": [ + 142, + 146, + 190, + 202 + ], + "spans": [ + { + "bbox": [ + 142, + 146, + 190, + 202 + ], + "type": "image", + "image_path": "ee5e1e0fb09c78a3f0542a202893398dfb18b151f32513b116ab76901f010a3a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 205, + 97, + 264, + 196 + ], + "blocks": [ + { + "bbox": [ + 205, + 97, + 264, + 196 + ], + "lines": [ + { + "bbox": [ + 205, + 97, + 264, + 196 + ], + "spans": [ + { + "bbox": [ + 205, + 97, + 264, + 196 + ], + "type": "image", + "image_path": "149c9adb2e71a4a7c378bd6b8b44a82e8de76413e78ce8929fa25c4f4f3b9a52.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 161, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 304, + 161, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 545, + 217 + ], + "type": "text", + "content": "Figure 7. Visualization of attention map. We visualize the encoder and decoder attention for the support keypoint at the left inner canthus. The support keypoint coarsely attends to keypoints with similar appearance in the encoder while focusing more on the corresponding keypoints for position update in the decoder." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 52, + 294, + 151, + 351 + ], + "blocks": [ + { + "bbox": [ + 47, + 267, + 157, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 267, + 157, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 157, + 289 + ], + "type": "text", + "content": "Table 4. Analysis of different terms in the similarity loss." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 294, + 151, + 351 + ], + "lines": [ + { + "bbox": [ + 52, + 294, + 151, + 351 + ], + "spans": [ + { + "bbox": [ + 52, + 294, + 151, + 351 + ], + "type": "table", + "html": "
LhLePCK
✓--88.74
--✓89.41
✓✓89.45
", + "image_path": "e04505fc44fc6d67e18b4ebb08f2f0ffa01b2f278aff78eb3dde28622c343c57.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 178, + 294, + 269, + 351 + ], + "blocks": [ + { + "bbox": [ + 164, + 267, + 285, + 289 + ], + "lines": [ + { + "bbox": [ + 164, + 267, + 285, + 289 + ], + "spans": [ + { + "bbox": [ + 164, + 267, + 285, + 289 + ], + "type": "text", + "content": "Table 5. Comparison of different support keypoint identifiers." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 178, + 294, + 269, + 351 + ], + "lines": [ + { + "bbox": [ + 178, + 294, + 269, + 351 + ], + "spans": [ + { + "bbox": [ + 178, + 294, + 269, + 351 + ], + "type": "table", + "html": "
TypePCK
--85.81
Learnable88.96
Fixed89.45
", + "image_path": "7a4f0a1b31b9c4385002c5a3f5cd03cdc0d891be3c3cab5f4ba6104fd2c38e25.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "type": "text", + "content": "tifier improves the PCK by " + }, + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "type": "inline_equation", + "content": "2.46\\%" + }, + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "type": "text", + "content": ". Note that the No. 3 model excludes the decoder, which is one-stage. Its performance can still achieve state-of-the-art on split1 test set, which suggests that the support identifier and QSR decoder significantly improve representation quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 422, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 529 + ], + "type": "text", + "content": "To intuitively show how the encoder and decoder work, we visualize the encoder/decoder attention maps in Fig. 7. The attention shows which location the support keypoint will focus during the information fusion (first stage) and the proposal correction (second stage). As can be observed, in the encoder, the support keypoint attends to keypoints with similar appearance, while in the decoder, the attention becomes more focused on the corresponding keypoints for proposal correction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 532, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 286, + 616 + ], + "type": "text", + "content": "Loss function. We analyze the similarity loss in Table 4. i) Only using the heatmap supervision (row 1) yields the worst performance, indicating that the expectation of loss is beneficial for precise localization. ii) As shown in row 3, the combination of heatmap loss and expectation attains the best performance, as both the shape of similarity map and the proposal position can be supervised." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Support keypoint identifier. We compare two different implementations of support keypoint identifier: learnable embedding and fixed sinusoidal embedding in Table 5. Using fixed sinusoidal embedding improves PCK by " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.49\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " compared with learnable embedding. The learnable embedding is similar to the learnable query embedding in DETR-based detection methods [5,25], which can be viewed as a learnable content embedding or anchors [23]. However, our" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 324, + 72, + 391, + 140 + ], + "blocks": [ + { + "bbox": [ + 324, + 72, + 391, + 140 + ], + "lines": [ + { + "bbox": [ + 324, + 72, + 391, + 140 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 391, + 140 + ], + "type": "image", + "image_path": "571c800b7eee5e00e91b5c70f8dae652f57ede0ba2af04b4939fa5204e838e43.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 140, + 381, + 159 + ], + "lines": [ + { + "bbox": [ + 334, + 140, + 381, + 159 + ], + "spans": [ + { + "bbox": [ + 334, + 140, + 381, + 159 + ], + "type": "text", + "content": "Support Image &Keypoint" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 392, + 72, + 457, + 139 + ], + "blocks": [ + { + "bbox": [ + 392, + 72, + 457, + 139 + ], + "lines": [ + { + "bbox": [ + 392, + 72, + 457, + 139 + ], + "spans": [ + { + "bbox": [ + 392, + 72, + 457, + 139 + ], + "type": "image", + "image_path": "868daeb6585332de87de355dfa487dd85070ee3761b2bdef7802a95cfee5014a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 140, + 440, + 157 + ], + "lines": [ + { + "bbox": [ + 410, + 140, + 440, + 157 + ], + "spans": [ + { + "bbox": [ + 410, + 140, + 440, + 157 + ], + "type": "text", + "content": "Encoder Attention" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 459, + 72, + 525, + 140 + ], + "blocks": [ + { + "bbox": [ + 459, + 72, + 525, + 140 + ], + "lines": [ + { + "bbox": [ + 459, + 72, + 525, + 140 + ], + "spans": [ + { + "bbox": [ + 459, + 72, + 525, + 140 + ], + "type": "image", + "image_path": "021f9e381df4d11926700c097dd789fdeb9f28c6e215a050f0aafbc4735f254c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 141, + 517, + 157 + ], + "lines": [ + { + "bbox": [ + 467, + 141, + 517, + 157 + ], + "spans": [ + { + "bbox": [ + 467, + 141, + 517, + 157 + ], + "type": "text", + "content": "Decoder Cross-Attention" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 308, + 251, + 542, + 308 + ], + "blocks": [ + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "type": "text", + "content": "Table 6. Inference and Training Efficiency. We test the efficiency on RTX3090 under 1-shot setting." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 251, + 542, + 308 + ], + "lines": [ + { + "bbox": [ + 308, + 251, + 542, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 542, + 308 + ], + "type": "table", + "html": "
MethodParams (M)GFLOPsFPSMemory (GB)
POMNet [40]48.2138.016.8013.8 × 2
One-Stage26.8622.6536.907.5
CapeFormer31.1423.6826.097.8
", + "image_path": "f93d7c6753a77c5d2757695e15f8fbfe1e284f9e43758e91fe58eddbe409623b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 317, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 366 + ], + "type": "text", + "content": "support keypoint identifier is fixed and serves a different purpose: distinguishing different support keypoints. Learnable support identifier deteriorates the performance as they may overfit the training categories more easily." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 373, + 506, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 506, + 387 + ], + "type": "text", + "content": "4.5. Performance and Efficiency Trade-off" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "type": "text", + "content": "We evaluate the efficiency under 1-shot setting for POMNet and CapeFormer in Table 6. The training memory is evaluated with a batchsize of 16. i) Comparing row 1 and 3, CapeFormer significantly reduces the parameters " + }, + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "(-35.4\\%)" + }, + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "type": "text", + "content": " and FLOPs " + }, + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "(-37.7\\%)" + }, + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "type": "text", + "content": " compared with POMNet due to the shared backbone and efficient similarity metric. ii) Comparing rows 2 and 3, the second stage only introduces limited parameters and FLOPs, which shows a satisfying trade-off between performance and efficiency. iii) CapeFormer is efficient for training, which consumes only 8GB memory on a single GPU (almost " + }, + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 304, + 392, + 545, + 524 + ], + "type": "text", + "content": " of POMNet)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 536, + 378, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 378, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 378, + 548 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 555, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 675 + ], + "type": "text", + "content": "In this work, we formulate a two-stage framework for category-agnostic pose estimation (CAPE) and instantiate it with a novel transformer model. Unlike previous one-stage approaches that entirely rely on matching results, the matching results in the first stage are viewed as similarity-aware position proposals which are refined in the second stage. We also propose specific designs to improve the representation and similarity modeling. On the CAPE benchmark MP-100, our method outperforms the previous best method by large margins on both accuracy and efficiency." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (Grant No. U1913602 and 61876211)." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7315" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter V. Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 3686-3693, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 287, + 178 + ], + "type": "text", + "content": "[2] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv Computer Research Repository, abs/1607.06450, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 234 + ], + "type": "text", + "content": "[3] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9497-9506, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 289 + ], + "type": "text", + "content": "[4] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1):172-186, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of European Conference on Computer Vision (ECCV), pages 213-229, 2020. 4, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 401 + ], + "type": "text", + "content": "[6] Seokju Cho, Sunghwan Hong, Sangryul Jeon, Yunsung Lee, Kwanghoon Sohn, and Seungryong Kim. Cats: Cost aggregation transformers for visual correspondence. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 9011-9023, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "type": "text", + "content": "[7] Seokju Cho, Sunghwan Hong, and Seungryong Kim. Cats++: Boosting cost aggregation with convolutions and transformers. arXiv Computer Research Repository, abs/2202.06817, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "text", + "content": "[8] Christopher B. Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Krishna Chandraker. Universal correspondence network. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 2406-2414, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 248-255, 2009. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[10] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of International Conference on Machine Learning (ICML), pages 1126–1135, 2017. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[12] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 14671-14681, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Shuqiang Jiang, Sisi Liang, Chengpeng Chen, Yaohui Zhu, and Xiangyang Li. Class agnostic image common object detection. IEEE Trans. Image Process., 28(6):2836-2846, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[15] Seungryong Kim, Dongbo Min, Bumsub Ham, Stephen Lin, and Kwanghoon Sohn. FCSS: fully convolutional self-similarity for dense semantic correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):581-595, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of International Conference on Learning Representations, 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "text", + "content": "[17] Rollyn Labuguen, Jumpei Matsumoto, Salvador Blanco Negrete, Hiroshi Nishimaru, Hisao Nishijo, Masahiko Takada, Yasuhiro Go, Ken ichi Inoue, and Tomohiro Shibata. Macaquepose: A novel \"in the wild\" macaque monkey pose dataset for markerless motion capture. Frontiers in behavioral neuroscience (2021), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "text", + "content": "[18] Jiefeng Li, Siyuan Bian, Ailing Zeng, Can Wang, Bo Pang, Wentao Liu, and Cewu Lu. Human pose regression with residual log-likelihood estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 11005–11014, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 332, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 386 + ], + "type": "text", + "content": "[19] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 10863-10872, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "text", + "content": "[20] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: A benchmark for amur tiger re-identification in the wild. In Proceedings of ACM International Conference on Multimedia, pages 2590–2598, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "type": "text", + "content": "[21] Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Tokenpose: Learning keypoint tokens for human pose estimation. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 11293-11302, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), pages 740-755, 2014. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 599 + ], + "type": "text", + "content": "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In Proceedings of International Conference on Learning Representations, 2022. 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "text", + "content": "[24] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Proceedings of Asian Conference on Computer Vision (ACCV), pages 669–684, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[25] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3631-3640, 2021. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "type": "text", + "content": "[26] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel flow: Semantic correspondence with multi-layer" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7316" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "neural features. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 3394-3403, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[27] Akihiro Nakamura and Tatsuya Harada. Revisiting finetuning for few-shot learning. arXiv Computer Research Repository, abs/1910.00216, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "text", + "content": "[28] Xuecheng Nie, Jianfeng Zhang, Shuicheng Yan, and Jiashi Feng. Single-stage multi-person pose machines. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 6950-6959, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "type": "text", + "content": "[29] N. Dinesh Reddy, Minh Vo, and Srinivasa G. Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 1906-1915, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "text", + "content": "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "type": "text", + "content": "[31] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomás Pajdla, and Josef Sivic. Neighbourhood consensus networks. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 1658-1669, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "text", + "content": "[32] Christos Sagonas, Epameinondas Antonakos, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: database and results. Image Vis. Comput., 47:3-18, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "text", + "content": "[33] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 11059-11068, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "type": "text", + "content": "[34] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 9529–9538, 2022. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[35] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4077-4087, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 600 + ], + "type": "text", + "content": "[36] Xibin Song, Peng Wang, Dingfu Zhou, Rui Zhu, Chenye Guan, Yuchao Dai, Hao Su, Hongdong Li, and Ruigang Yang. Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5452-5462, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[37] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 2621-2630, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[38] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 1653-1660, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 484 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[40] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Proceedings of European Conference on Computer Vision (ECCV), volume abs/2207.10387, 2022. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "text", + "content": "[41] Shuo Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5525-5533, 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[42] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2878-2890, 2013. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "text", + "content": "[43] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 5217–5226, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "type": "text", + "content": "[44] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 7091-7100, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 418 + ], + "type": "text", + "content": "[45] Song-Hai Zhang, Ruilong Li, Xin Dong, Paul L. Rosin, Zixi Cai, Han Xi, Dingcheng Yang, Haozhi Huang, and Shi-Min Hu. Pose2seg: Detection free human instance segmentation. In Proceedings of IEEE Conference on Computer Vision Pattern Recognition (CVPR), pages 889-898, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 420, + 547, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 547, + 484 + ], + "type": "text", + "content": "[46] Jian Zhao, Jianshu Li, Yu Cheng, Li Zhou, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In Proceedings of ACM International Conference on Multimedia, pages 792-800, 2018. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7317" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_content_list.json b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d014a850dfd554f2818c2092f9100e152567efd5 --- /dev/null +++ b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_content_list.json @@ -0,0 +1,1798 @@ +[ + { + "type": "text", + "text": "MeMaHand: Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction", + "text_level": 1, + "bbox": [ + 238, + 130, + 733, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Congyi Wang* Feida Zhu* Shilei Wen† ByteDance", + "bbox": [ + 299, + 202, + 650, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kongyi.1990，zhufeida，zhengmin.666}@bytedance.com", + "bbox": [ + 259, + 241, + 710, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods proposed for hand reconstruction tasks usually parameterize a generic 3D hand model or predict hand mesh positions directly. The parametric representations consisting of hand shapes and rotational poses are more stable, while the non-parametric methods can predict more accurate mesh positions. In this paper, we propose to reconstruct meshes and estimate MANO parameters of two hands from a single RGB image simultaneously to utilize the merits of two kinds of hand representations. To fulfill this target, we propose novel Mesh-Mano interaction blocks (MMIBs), which take mesh vertices positions and MANO parameters as two kinds of query tokens. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. The transformer encoders are equipped with different asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Moreover, we introduce the mesh alignment refinement module to further enhance the mesh-image alignment. Extensive experiments on the InterHand2.6M benchmark demonstrate promising results over the state-of-the-art hand reconstruction methods.", + "bbox": [ + 75, + 321, + 473, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 670, + 209, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-based 3D hand analysis plays an important role in many applications such as virtual reality (VR) and augmented reality (AR). Two-hand reconstruction from a single RGB image is more challenging due to complex mutual interactions and occlusions. Besides, the skin appearance similarity makes it difficult for the network to align image features to the corresponding hand.", + "bbox": [ + 75, + 695, + 468, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous hand reconstruction works can be divided into two categories, parametric methods [3, 7, 32, 33] and nonparametric methods [4-6, 10, 18-20]. Parametric methods typically learn to regress pose and shape parameters of MANO model [23], where pose represents joint rota", + "bbox": [ + 75, + 801, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/628c321180901f9334be6329b8db67b50dbb40e122d8f523a49885606cccd297.jpg", + "image_caption": [ + "Figure 1. Comparison with the state-of-the-art method IntagHand [16] for single-image two-hand reconstruction. The integration of parametric and non-parametric hand representations allows us to achieve better performance in hard cases such as severe occlusions and challenging viewpoints." + ], + "image_footnote": [], + "bbox": [ + 519, + 290, + 870, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions in axis-angle representation and shape represents the coefficients of shape PCA bases. The MANO prior can yield plausible hand shapes from a single monocular image. However, they can not produce fine-grained hand meshes due to their limited capacity.", + "bbox": [ + 496, + 612, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid progress of graph convolutional network (GCN) and transformer techniques [10, 16, 18, 19], it is observed that direct mesh reconstruction can achieve state-of-the-art performance towards the Euclidean distances between the ground truth vertices and the predicted vertices. Nonetheless, the non-parametric methods are less robust in handling challenging viewpoints or severe occlusions.", + "bbox": [ + 496, + 688, + 893, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel single-image two-hand reconstruction method designed to predict mesh vertices positions and estimate MANO parameters simultaneously to utilize the merits of two kinds of hand representations. The proposed Mesh-Mano interaction Hand reconstruction architecture (MeMaHand) consists of three modules: 1) the image encoder-decoder module, 2) the", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 94, + 886, + 339, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "564", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mesh-mano interaction module, 3) and the mesh alignment refinement module. To extract contextually meaningful image features, we pre-train a classical image encoder-decoder network on auxiliary tasks including hand segmentation, hand 2D joints and dense mapping encodings. The low-resolution features encode more global knowledge, while the high-resolution features contain more local details. Secondly, the mesh-mano interaction module stacks three mesh-mano interaction blocks (MMIBs) to transform the mesh vertices and MANO parameters queries initialized by the global image feature vector. We observe that the hand prior embedded in the MANO parameters is valuable for predicting stable hand meshes in challenging situations such as severe occlusions. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. The transformer encoders are equipped with different asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Each MMIB is followed by an upsampling operation to upsample the mesh vertices tokens in a coarse-to-fine manner. Finally, the mesh alignment refinement module utilizes one MMIB to predict offsets for mesh vertices and MANO parameters to enhance mesh-image alignment. To improve the reliability of image evidence, we project mesh vertices predicted by the Mesh-Mano interaction module onto the 2D image plane. The explicit mesh-aligned image features are concatenated to the transformer input tokens.", + "bbox": [ + 76, + 90, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The whole network, including the pre-trained image feature encoder-decoder, is jointly optimized such that the image features better adapt to our hand mesh reconstruction task. Benefiting from the mesh-mano interaction mechanism and mesh alignment refinement stage, extensive experiments demonstrate that our method outperforms existing both parametric and non-parametric methods on Inter-Hand2.6M [22] dataset. In summary, the contributions of our approach are as follows:", + "bbox": [ + 76, + 513, + 472, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose MeMaHand to integrate the merits of parametric and non-parametric hand representation. The mesh vertices and MANO parameters are mutually reinforced to achieve better performance in mesh recovery and parameter regression.", + "- A mesh-image alignment feedback loop is utilized to improve the reliability of image evidence. Therefore, more accurate predictions are obtained by rectifying the mesh-image misalignment.", + "- Our method achieves superior performance on the InterHand2.6M dataset, compared with both non-parametric and parametric methods." + ], + "bbox": [ + 94, + 654, + 468, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 845, + 227, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parametric Hand Reconstruction. Parametric approaches [1-3, 7, 11, 12, 23, 32, 33] use a parametric hand model such", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as MANO [23] and focus on regressing the pose and shape parameters from a single image. The rich embedded prior information (e.g. the geometric dynamic constraints of joint rotations) can assist deep model learning when 3D annotations are insufficient. Weak 2D joints supervision [3] and motion caption data [33] are utilized to train convolutional neural networks (CNN) to predict MANO parameters. The reliance on 3D manual annotations is further totally alleviated by S²HAND [7]. However, the reconstructed mesh can not fully express the local details of variable 3D hand shapes due to the limited capacity of the MANO model.", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-parametric Hand Reconstruction. Non-parametric methods aim to reconstruct the hand mesh directly [5, 6, 8, 10, 16-21, 26, 29]. To explicitly encode mesh topology, graph convolutional network (GCN) is adopted for aggregating adjacent vertices features. Hierarchical architectures [10, 16] are designed for mesh generation from coarse to fine. To model long-range dependencies, multi-layer transformer encoders are introduced such that the global interactions can be modeled without being limited by any mesh topology [18, 19]. Most recently, IntagHand [16] achieves the state-of-the-art performance on InterHand2.6M dataset [22]. However, we observe that IntagHand is less robust in handling challenging viewpoints or severe occlusions. The reconstructed mesh may be corrupted into unnatural shapes.", + "bbox": [ + 496, + 262, + 893, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interacting Two-Hand Reconstruction. Although single-hand methods can extend to two-hand reconstruction, the correlation between the left and right hands is not considered. Besides, the performance deteriorates for the close-interacting hands. Depth cameras [15] are sensitive to tracking accuracy and multi-view images [24, 27] are expensive to acquire. Based on a large-scale interacting hand dataset named InterHand2.6M [22], deep learning based single image methods either estimate hand joint positions [14, 22], or employ a 2.5D heatmap to predict the MANO parameters [30], or directly reconstruct meshes [16]. In summary, existing methods only employ a single specific hand representation. In contrast, our method not only combines the merits of MANO representations and mesh representations but also proposes a dedicated transformer encoder with asymmetric attention masks to make them collaborate seamlessly. Besides, the joint rotation and precise mesh vertices positions can be used for different applications, such as driving CG characters or virtual try-on.", + "bbox": [ + 496, + 479, + 893, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 782, + 635, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overall architecture of MeMaHand is depicted in Fig. 2. Given a single image $I$ , the proposed method can predict the mesh vertices positions $V \\in \\mathbb{R}^{778 \\times 3}$ , MANO pose parameter $\\theta \\in \\mathbb{R}^{48}$ and shape parameter $\\beta \\in \\mathbb{R}^{10}$ of two hands simultaneously in one forward pass. In this section, we first describe the overall architecture of MeMa", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "565", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9cff5c1c98fd285d7f2cf0186dd56bc1357acdc7bda3818b244ad0550f425b7f.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our proposed MeMaHand. We pretrain an image encoder-decoder module on auxiliary tasks to extract multi-scale contextually meaningful image features. The Mesh-Mano interaction module stacks three Mesh-Mano interaction blocks (MMIBs) to predict the mesh vertices positions and MANO parameters simultaneously. The mesh alignment refinement module rectifies the mesh-image misalignment by explicitly utilizing the mesh-align image features." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 883, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hand in Sec. 3.1. Then, we illustrate the mesh-mano interaction module and mesh-alignment refinement module in Sec. 3.2 and Sec. 3.3, respectively. Finally, we introduce the model objectives used to train our network in Sec. 3.4.", + "bbox": [ + 75, + 500, + 468, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. System Overview", + "text_level": 1, + "bbox": [ + 76, + 571, + 246, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extract contextually meaningful image features, we pre-train a classical image encoder-decoder module on auxiliary tasks including hand segmentation, hand 2D joints, and dense mapping encodings. The 2D image conditions are generated from the reconstructed mesh of ground-truth MANO parameters. ResNet50 [13] is leveraged as the backbone. Global average pooling feature $F_{gap}$ and multi-scaled image feature $\\{\\phi_i \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}, i = 1, 2, 3\\}$ are extracted from the image encoder (E) and decoder (D).", + "bbox": [ + 75, + 597, + 468, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Afterward, we propose Mesh-Mano Interaction Module to predict mesh vertices positions $V_{MMIM}$ and MANO parameters $\\theta_{MMIM}, \\beta_{MMIM}$ of both left (L) and right (R) hands. We use the global feature $F_{gap}$ to initialize the vertex and MANO queries. The grid-sampled image features $\\{\\phi_i^{grid}, i = 1,2,3\\}$ are tokenized as well following the practice of Mesh Graphformer [19].", + "bbox": [ + 76, + 733, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mesh alignment refinement module further improves the predictions generated by the mesh-mano interaction module. The explicit mesh-aligned image features extracted from the multi-scale feature maps $\\phi_{i}$ are more informative", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "than grid-sampled image features, which helps rectify the predictions. Details will be elaborated in Sec. 3.3.", + "bbox": [ + 498, + 500, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Mesh-Mano Interaction Module", + "text_level": 1, + "bbox": [ + 498, + 537, + 782, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mesh-mano interaction module reconstructs hand mesh in a coarse-to-fine manner with three Mesh-Mano Interaction Blocks (MMIBs). We leverage the graph coarsening method [16] to build three-level sub meshes with vertex number $N_{1} = 63$ , $N_{2} = 126$ , $N_{3} = 252$ . Each MMIB is followed by an upsample operation that reverses the topological relationship between adjacent sub-meshes. The full mesh vertices positions ( $N = 778$ ) are obtained with a simple $1 \\times 1$ Conv from the final output vertex token.", + "bbox": [ + 496, + 561, + 890, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 3 shows the detailed structure of MMIB. At level $i$ , the input tokens of MMIB include two kinds of queries: vertex queries $f_{v}^{h}\\in \\mathbb{R}^{N_{i}\\times D_{i}}$ and MANO parameter queries $f_{m}^{h}\\in \\mathbb{R}^{2\\times D_{i}}$ , where $h$ indicates left $(L)$ or right $(R)$ hand, and $D_{i}$ represents the feature dimensions. Therefore, the total sequence length is $(2\\times N_i + 4)$ .", + "bbox": [ + 496, + 698, + 890, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Graph Residual Block. The design of the graph residual block is similar to [10, 16]. The Chebyshev spectral graph CNN [9] is adopted to transform the vertex token to intermediate graph features,", + "bbox": [ + 496, + 792, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {g r a p h}} = \\mathbf {G r a p h C o n v} \\left(f _ {v} ^ {L}, f _ {v} ^ {R}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 859, + 890, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the operation GraphConv denotes the graph con", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "566", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/53d71bfedb9ec3aebdce209f4c65a95584025888a1a899d4dc58284b0f4f3e70.jpg", + "image_caption": [ + "Figure 3. The detailed structure of Mesh-Mano Interaction Block (MMIB). MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 436, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "volution. More details can be found in [9]. For simplicity, we use $f_{graph}$ to represent the concatenation of updated left-hand and right-hand vertex tokens.", + "bbox": [ + 75, + 446, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intra-hand Transformer Encoder. While graph CNN is useful for extracting local information, it is less efficient at capturing long-range dependencies. We use transformer encoders to model dependencies not only between long-range vertices but also between mesh vertices and MANO parameters. The spatial image features at resolution $i$ are also tokenized as $\\phi_i^{grid}$ by grid sampling following the practice of Mesh Graphformer [19]. Therefore, the input tokens of intra-hand transformer encoder (TFEintra) consist of graph features $(f_{graph})$ , MANO queries $(f_m^L, f_m^R)$ and grid image feature $(\\phi_i^{grid})$ . Based on the observation that the mesh vertices regression can lead to more precise mesh reconstruction and the MANO parameters are more stable, we propose an asymmetric attention mask excluding the manoto-mesh attention such that the MANO parameter tokens will not directly affect the mesh vertex tokens. The mesh-to-mano attention remains. Therefore, the mesh vertex tokens are implicitly regularized by the MANO parameters loss. TFEintra focuses on modeling the intra-hand dependencies. The inter-hand attention is also excluded. The resulting asymmetric attention mask $M_{intra}$ is shown in Fig. 3. Finally, the intra-hand transformer encoder is formulated as:", + "bbox": [ + 75, + 494, + 470, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i n t r a} = T F E _ {i n t r a} \\left(\\left[ f _ {g r a p h}, f _ {m} ^ {L}, f _ {m} ^ {R}, \\phi_ {i} ^ {g r i d} \\right], M _ {i n t r a}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 837, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\ldots]$ denotes the concatenation operation. For simplicity, $f_{\\text{intra}}$ represents the concatenation of updated mesh", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vertex and MANO parameter tokens.", + "bbox": [ + 500, + 90, + 746, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inter-hand Transformer Encoder. In contrast to $TFE_{intra}$ , the inter-hand transformer encoder $(TFE_{inter})$ focuses on modeling the inter-hand correlations. Fig. 3 presents the resulting asymmetric attention mask $M_{inter}$ . The image feature tokens are not used in $TFE_{inter}$ . Finally, the inter-hand transformer encoder is formulated as:", + "bbox": [ + 498, + 111, + 890, + 202 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i n t e r} = T F E _ {i n t e r} \\left(f _ {i n t r a}, M _ {i n t e r}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 219, + 890, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{inter}$ can be further split into updated vertex tokens $\\bar{f}_v^L,\\bar{f}_v^R$ and MANO parameter tokens $\\bar{f}_m^L,\\bar{f}_m^R$ .", + "bbox": [ + 498, + 251, + 890, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed MMIB predicts the mesh vertices positions and MANO parameters simultaneously in a unified architecture. Intra-hand and inter-hand mesh-mano dependencies are modeled by two cascaded transformer encoders with different attention masks. After three MMIBs, a simple $1 \\times 1$ Conv is applied to obtain the mesh vertices positions and MANO parameters from output tokens,", + "bbox": [ + 496, + 284, + 890, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {M M I M} ^ {h} = \\operatorname {C o n v} _ {v} \\left(\\bar {f} _ {v} ^ {h}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 402, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {M M I M} ^ {h}, \\beta_ {M M I M} ^ {h} = C o n v _ {m} \\left(\\bar {f} _ {m} ^ {h}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 425, + 890, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h \\in \\{L, R\\}$ indicates the left or right hand. To tackle the noticeable misalignment between the estimated meshes and image evidence, we propose a mesh alignment refinement module to rectify the results.", + "bbox": [ + 496, + 458, + 890, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Mesh Alignment Refinement Module", + "text_level": 1, + "bbox": [ + 500, + 535, + 818, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further improve mesh-image alignment, we propose a novel mesh-alignment refinement module inspired by PyMAF [31]. Specifically, we project the mesh vertices $V_{MMIM}$ predicted by the mesh-mano interaction module onto the multi-scale image features explicitly. Three-scale mesh-aligned image features are concatenated together for each corresponding vertex token. A simple multi-layer perceptron (MLP) is adopted to reduce the dimensions, resulting in a fused image feature $\\phi_v$ .", + "bbox": [ + 496, + 560, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize one MMIB to refine the mesh vertex and MANO parameter tokens $\\bar{f}_v^h, \\bar{f}_m^h$ . To effectively utilize the mesh-aligned image evidence, we made some modifications to the tokens. Specifically, we concatenate the image feature $\\phi_v$ to vertex token $\\bar{f}_v^h$ along the channel dimension. For MANO parameter tokens, we perform a global average pooling operation to obtain global feature vector $\\phi_m$ before concatenation. The rectified tokens are formulated as,", + "bbox": [ + 496, + 698, + 890, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {v} ^ {h}, \\hat {f} _ {m} ^ {h} = M M I B \\left(\\bar {f} _ {v} ^ {h}, \\bar {f} _ {m} ^ {h}, \\phi_ {v}, \\phi_ {m}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 833, + 890, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One simple $1 \\times 1$ Conv layer is applied to obtain the mesh vertices positions offsets and MANO parameter off", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "567", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sets for rectification,", + "bbox": [ + 76, + 90, + 217, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {h} = V _ {M M I M} ^ {h} + \\operatorname {C o n v} _ {v} \\left(\\hat {f} _ {v} ^ {h}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 114, + 468, + 133 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {h} = \\theta_ {M M I M} ^ {h} + C o n v _ {m} \\left(\\hat {f} _ {m} ^ {h}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 136, + 468, + 154 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta^ {h} = \\beta_ {M M I M} ^ {h} + \\operatorname {C o n v} _ {m} \\left(\\hat {f} _ {m} ^ {h}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 157, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V^h$ , $\\theta^h$ and $\\beta^h$ are the final results. Note that we can stack several MMIBs to predict the offsets iteratively. In our experiments, one single MMIB is enough to achieve satisfactory results. PyMAF [31] fuses the mesh-align features into one global feature, which ignores the inherent spatial relationship of mesh vertices. In contrast, our MMIB adopts GCN to model the spatial relation between adjacent mesh vertices. The transformer encoder is responsible for modeling the long-range dependencies.", + "bbox": [ + 76, + 186, + 468, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Objectives", + "text_level": 1, + "bbox": [ + 76, + 332, + 246, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method predicts mesh vertices positions and MANO parameters simultaneously. The learning objectives can be divided into three categories: mesh vertex loss, MANO parameter loss and mesh-mano consistency loss.", + "bbox": [ + 76, + 354, + 468, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh Vertex Loss: The widely-used L1 loss is adopted to supervise the vertex positions and the 2D projections:", + "bbox": [ + 76, + 417, + 468, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V} = \\sum_ {h = L, R} \\| V ^ {h} - V _ {G T} ^ {h} \\| _ {1} + \\| \\Pi (V _ {h}) - \\Pi (V _ {G T} ^ {h}) \\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 458, + 468, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V_{GT}^{h}$ represents the ground-truth vertex positions and $\\Pi$ denotes the projection function. Given vertex positions, the joint positions can be regressed by multiplying the predefined regression matrix $\\mathcal{J}$ . Joint losses are formulated as:", + "bbox": [ + 76, + 501, + 468, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {J} = \\sum_ {h = L, R} \\| \\mathcal {J} V ^ {h} - \\mathcal {J} V _ {G T} ^ {h} \\| _ {1} \\\\ + \\sum_ {h = L, R} \\| \\Pi \\mathcal {J} \\left(V _ {h}\\right) - \\Pi \\left(\\mathcal {J} V _ {G T} ^ {h}\\right) \\| _ {1}. \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 571, + 468, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The face normal loss is introduced to regularize the surface normal consistency:", + "bbox": [ + 76, + 650, + 468, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N} = \\sum_ {h = L, R} \\sum_ {f = 1} ^ {F} \\sum_ {i = 1} ^ {3} \\| e _ {f, i} ^ {h} \\cdot n _ {f, G T} ^ {h} \\| _ {1}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 690, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_{f,i}^{h}$ represents the $i$ th edge of face $f$ at hand $h$ and $n_{f,GT}^{h}$ is the normal vector of this face from the ground truth mesh. The edge length consistency loss is to enforce the edge length consistency:", + "bbox": [ + 76, + 743, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E} = \\sum_ {h = L, R} \\sum_ {i = 1} ^ {E} \\| e _ {i} ^ {h} - e _ {i, G T} ^ {h} \\| _ {1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 816, + 468, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_i^h$ represents the $i$ th edge of hand $h$ and $E$ denotes the total edge number, respectively.", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9166ca1547682e815e9a053e83522cbb067607afd04abe2ccc1e5c9cf8e20523.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParams(M)Inference Time (ms)
InterShape [30]139.647.77
Intaghand [16]37.2851.86
Ours38.3160.52
", + "bbox": [ + 539, + 88, + 857, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Parameters and inference speed on Tesla V100", + "bbox": [ + 529, + 161, + 861, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MANO Parameter Loss: Given the ground truth MANO parameters, L1 loss is used to regress parameters,", + "bbox": [ + 498, + 202, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P} = \\sum_ {h = L, R} \\| \\theta^ {h} - \\theta_ {G T} ^ {h} \\| _ {1} + \\| \\beta^ {h} - \\beta_ {G T} ^ {h} \\| _ {1}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 244, + 890, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, we reconstruct hand mesh based on the MANO model. The reconstructed mesh should be close to the ground-truth mesh,", + "bbox": [ + 498, + 289, + 890, + 334 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V, \\text {m a n o}} = \\sum_ {h = L, R} \\| \\mathcal {M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - V _ {G T} ^ {h} \\| _ {1}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 345, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh-Mano Consistency Loss. The predicted mesh and the reconstructed mesh from MANO parameters should be consistent with each other:", + "bbox": [ + 498, + 393, + 890, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s i s t}} = \\sum_ {h = L, R} \\| \\mathcal {M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - V ^ {h} \\| _ {1} \\tag {16} \\\\ + \\left\\| \\mathcal {J M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - \\mathcal {J} V ^ {h} \\right\\| _ {1}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 448, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the overall training loss $\\mathcal{L}$ is:", + "bbox": [ + 517, + 513, + 795, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {V} \\mathcal {L} _ {V} + \\lambda_ {J} \\mathcal {L} _ {J} + \\lambda_ {N} \\mathcal {L} _ {N} + \\lambda_ {E} \\mathcal {L} _ {E} \\tag {17} \\\\ + \\lambda_ {P} \\mathcal {L} _ {P} + \\lambda_ {V, m a n o} \\mathcal {L} _ {V, m a n o} + \\lambda_ {\\text {c o n s i s t}} \\mathcal {L} _ {\\text {c o n s i s t}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 541, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{V} = 40$ , $\\lambda_{J} = 40$ , $\\lambda_{N} = 5$ , $\\lambda_{E} = 40$ , $\\lambda_{P} = 10$ , $\\lambda_{V,mano} = 10$ and $\\lambda_{consist} = 40$ . The whole network including the pre-trained image feature encoder-decoder is jointly optimized such that the image features better adapt to our hand mesh reconstruction task.", + "bbox": [ + 498, + 588, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 678, + 624, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation", + "text_level": 1, + "bbox": [ + 500, + 703, + 761, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Dataset. Interhand2.6M [22] is a large-scale hand dataset with ground truth mesh annotations including both single-hand and interacting two-hand images. We pick out the interacting hand (IH) data for training and testing. For a fair comparison, we follow the preprocessing steps of [16] which produces 366K training and 261K testing samples.", + "bbox": [ + 498, + 729, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The image encoder uses ResNet50 [13] as the backbone. The decoder contains four simple deconvolutional layers. We first pretrain our image encoder-decoder module on auxiliary tasks. Then, the whole network is jointly optimized with Adam optimizer at", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "568", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d01db973666f0219532d9ae561a584bfbafbb2680e4736f23f58e7541f44b90f.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with SOTA parametric method InterShape [30] and SOTA non-parametric method IntagHand [16] on InterHand2.6M dataset. Our method performs better on close-interacting two-hand reconstruction (first row). Besides, our method is robust in hard cases such as severe occlusions and challenging viewpoints (second row)." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 483, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb5876f9ba615edb94708ccff8196db4b01e0a695811ed507b616b3b88e224d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 883, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fecd2dd1be525dc5ad17d7a9f80acf9e6828cd819a68b5970ac4f2e3f4d94c07.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 86, + 308, + 168, + 433 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/25a5505b0a9f7a1d25d38c2a80c88bcf5c7281c3aae6a9f9bd7d1089c38be9e1.jpg", + "image_caption": [ + "GT", + "Ours", + "Figure 5. Our method produces accurate hand reconstructions for various two-hand interactions." + ], + "image_footnote": [], + "bbox": [ + 173, + 308, + 336, + 433 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8db48590528402b273c7a72ae3840a497a8547c6eb040a7375277de90deb696.jpg", + "image_caption": [ + "Alternative Views" + ], + "image_footnote": [], + "bbox": [ + 341, + 316, + 486, + 429 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de4242ab860c49e4f0baa8f44193dc83ec2cea1fc17c298b6e90fe3e3d09c921.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 486, + 308, + 571, + 433 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0531ff7aa698fb3d91f278aa1e0ac520501ff30441245a56318502f846f3f95.jpg", + "image_caption": [ + "GT", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 573, + 308, + 738, + 433 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b9034f4a6fd92e1c166ed031211c71447b235b395822787787c388a663a0daa.jpg", + "image_caption": [ + "Alternative Views" + ], + "image_footnote": [], + "bbox": [ + 741, + 318, + 883, + 429 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fa1935afd96fdd24072ca4b800825b0971a4009bac8d9c3433436dd74efac0d1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPE↓MPVPE↓AUC ↑PROJ2D↓
zimmermann et al. [34]36.36---
Spurr et al. [25]15.40---
InterNet [22]16.00-0.711-
Intaghand [16]8.799.030.8066.47
Ours8.658.890.8326.22
", + "bbox": [ + 91, + 494, + 459, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison with state-of-art non-parametric methods on InterHand2.6M dataset.", + "bbox": [ + 76, + 592, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a learning rate $1 \\times 10^{-4}$ . Data augmentation includes random rotation, random translation, and random scaling. The batch size is set to 32. It takes around 48 hours to train MeMaHand with 4 Tesla V100 GPUs.", + "bbox": [ + 75, + 662, + 468, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Mean Per Joint Position Error (MPJPE) and Mean Per Vertex Position Error (MPVPE) in millimeters are adopted to evaluate the mesh reconstruction accuracy. Additionally, we report the percentage of correct keypoints (PCK) curve and Area Under the Curve (AUC) across the thresholds between 0 and 50 millimeters. To evaluate the mesh-image alignment accuracy, the reconstructed mesh vertices are projected onto the 2D image plane. The mesh-image alignment accuracy is measured by PROJ2D, which calculates the distance in image pixels between the projected ground truth vertices and the predicted vertices.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 500, + 496, + 864, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our model with state-of-the-art non-parametric methods including InterNet [22] and Intaghand [16] and parametric methods including [3], MinimalHand [33] and InterShape [30]. The officially released weights are used to obtain the results. The model parameters and inference time of our method and other SOTA methods are reported in Tab. 1. The inference of our model can be completed in $60.52ms$ , comparable to other SOTAs but with better mesh reconstruction results. Besides, the mesh vertex positions as well as the MANO parameters are predicted simultaneously in one single forward pass.", + "bbox": [ + 496, + 525, + 890, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Non-parametric Methods. Non-parametric methods directly generate mesh vertex positions, which can express the local details of variable 3D hand shapes. The quantitative results of non-parametric methods are shown in Tab. 2. It can be seen that our method achieves the best performance on all evaluation metrics. Fig. 6(a) presents the PCK curve, which further demonstrates the superiority of our method at all threshold levels.", + "bbox": [ + 496, + 699, + 890, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 presents the qualitative comparison with SOTA non-parametric method IntagHand [16]. Our method performs better on close interacting two-hand reconstruction (first row). Besides, we observe that IntagHand produces collapsed meshes in hard cases such as severe occlusions", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "569", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8832cb057416e9ea798c88119611ef0f99c5d11bb8530769c27b12b7366407ec.jpg", + "image_caption": [ + "(a) Comparison with non-parametric method" + ], + "image_footnote": [], + "bbox": [ + 117, + 111, + 441, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e065247fc52936c06f16d9272c9bd3a5ef806c3e8f2544598fbb6f7ddedf6d0.jpg", + "image_caption": [ + "(b) Comparison with parametric method" + ], + "image_footnote": [], + "bbox": [ + 504, + 111, + 826, + 296 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f274a1405a6d4d23cd0f90b71e9f8119268f93dfa2f73c9d44a11a1c3595c102.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPE↓MPVPE ↓AUC↑PROJ2D↓
Boukhayma et al. [3]16.9317.980.672-
MinimalHand [33]23.4823.890.543-
InterShape [30]13.4813.950.7359.95
Ours10.8510.920.7906.90
", + "bbox": [ + 86, + 375, + 464, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison with state-of-art parametric methods on InterHand2.6M dataset.", + "bbox": [ + 76, + 457, + 468, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and challenging viewpoints (second row). In contrast, our method is more robust to such situations. We attribute this success to the integration of MANO representations, which predicts stable hand meshes. More results with alternative views are shown in Figure 5.", + "bbox": [ + 75, + 516, + 468, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Parametric Methods. For a fair comparison, the reconstructed hand meshes from our predicted MANO parameters are utilized for evaluation. The quantitative comparisons are listed in Tab. 3. The PCK curve is shown in Fig. 6(a). The MANO parameter estimation of our method outperforms other SOTA parametric methods. Thanks to the mesh-mano interaction block, the MANO token is conditioned on the mesh vertex tokens where spatial relations are retained. Such dependencies are more informative than one-dimensional global feature utilized in competing methods [3, 30, 33].", + "bbox": [ + 75, + 597, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Tab. 2 and Tab. 3, we can find the parametric method InterShape [30] has higher statistical errors on joint and vertex positions compared with the non-parametric method IntagHand [16] due to the limited capacity of the parametric model. However, from the second row of Fig. 4, we can see InterShape produces reasonable hand shapes in these challenging cases. By combining the merits of parametric and non-parametric hand representations, our method achieves the best performance both quantitatively", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5de1be686022aa4fc1520cfc845f8fd2a2113d7fcae4aebacf137b61acb6221.jpg", + "image_caption": [ + "Figure 6. Comparison with the state-of-the-art methods on InterHand2.6M dataset. Left: 3D PCK of non-parametric methods. Right: 3D PCK of parametric methods. The AUC values are shown in parentheses.", + "Figure 7. Two-hand reconstruction of images in the wild, which are taken from the RGB2Hands [28]." + ], + "image_footnote": [], + "bbox": [ + 501, + 376, + 893, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and qualitatively.", + "bbox": [ + 500, + 593, + 617, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Extension to Images in the Wild", + "text_level": 1, + "bbox": [ + 500, + 617, + 782, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further present the hand reconstruction results on images in the wild. As shown in Fig. 7, our method performs well on real-life images taken from the RGB2Hands dataset [28], which demonstrates the generalization ability of our approach. Recall that our method is designed to predict accurate mesh vertices positions and MANO parameters simultaneously in one forward pass. The MANO pose parameter represents the joint rotations in axis-angle representation, which is useful for animating 3D hands in computer graphics. Accurate hand mesh vertices positions and MANO parameters for hands in the wild can facilitate different human-computer-interaction (HCI) applications.", + "bbox": [ + 496, + 641, + 890, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 832, + 653, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the proposed modules in our framework, we conduct an ablation study on several variants of our method. The quantitative results of all vari-", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "570", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f66d90fb1b4ee50f4a4a1bcbbf5d376824a2c7078433105cd793dbe1391b48fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MPJPE↓MPVPE↓AUC↑PROJ2D↓
A: w/o MANO token8.879.090.8186.85
B: w/o Mesh-Align8.799.010.8266.64
C: Mesh-Align scale-168.738.970.8296.38
D: Mesh-Align scale-648.778.990.8266.58
E: Regress MANO from Fgap8.839.040.8286.86
F: w/o asymmetric attention9.179.370.8226.67
G: w/o auxiliary tasks9.119.280.8236.78
Full Model8.658.890.8326.22
", + "bbox": [ + 84, + 88, + 475, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparisons of different variants of our method.", + "bbox": [ + 99, + 224, + 442, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25bdbd62e90d44a4afff8232342ae0d4d54d2b8eca08fbdc12d147e4111e0092.jpg", + "image_caption": [ + "Figure 8. Ablation study on MANO tokens." + ], + "image_footnote": [], + "bbox": [ + 133, + 258, + 225, + 415 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d4ecd5b8b81954f6dd360626b7169be44b210982978b02cbdebf4ace3171fb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 258, + 320, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/039af84d3432e6e7a86bfdcb531e00b66575811ccd126e0eae9469f666e915bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 258, + 415, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29d567eecff4aedd51d7cf561fe6b8e79395b1dd7fc147a7d90b0f9664286e13.jpg", + "image_caption": [ + "Figure 9. Ablation study on mesh alignment refinement." + ], + "image_footnote": [], + "bbox": [ + 78, + 457, + 158, + 592 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f0258ac679720dde1cbc7824c524cddb8f4294acfcfae84f9361a6c0c5a57e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 458, + 316, + 592 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/785f616a298a9505e8160e4a93ac19d1ff03dcfaf0ce6ab12d0b7c570cc0fac2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 458, + 468, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ants are presented in Table 4.", + "bbox": [ + 76, + 648, + 271, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of MANO token. Variant A (w/o MANO token) represents removing the MANO tokens. The transformer encoder only models the dependencies between the mesh vertex tokens and image features. Fig. 8 show the qualitative comparison. Without hand prior information, the reconstructed mesh may be corrupted when parts of the hands are occluded. In contrast, our full model can generate reasonable meshes in these challenging situations.", + "bbox": [ + 75, + 667, + 468, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Mesh Alignment Refinement. Variant B (w/o Mesh-Align) represents removing the mesh alignment refinement module. The predictions generated by the Mesh-Mano interaction module are taken as the final outputs. Variant C and D denote extracting the single-scale image features at resolution 16 and 64, respectively. The performances degrade without using mesh align refinement", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9c1cc86008180787f43161818858f9d25ad7400d0928f6a3124b2d550b8f9c3.jpg", + "image_caption": [ + "Figure 10. Failures cases caused by severe occlusions." + ], + "image_footnote": [], + "bbox": [ + 555, + 89, + 839, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "or using single-scale image features. As shown in Fig. 9, our full model produces better-aligned results.", + "bbox": [ + 498, + 301, + 888, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of MMIB. Parametric and non-parametric representations are widely used for many years. However, combining them is not trivial. In variant $\\mathbf{E}$ (Regress MANO from $F_{gap}$ ), MMIB is responsible for updating vertex tokens. The MANO parameters are regressed from the global average pooling features $F_{gap}$ with simple fully connected layers rather than through MMIB. The performance gain of variant $\\mathbf{E}$ is limited as shown in Tab. 4. On the other hand, we propose an asymmetric attention mask to exclude the mano-to-mesh attention. To verify the choice, variant $\\mathbf{F}$ (w/o asymmetric attention mask) denotes removing the asymmetric attention mechanism. The performance degrades compared with our full model.", + "bbox": [ + 498, + 335, + 890, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Pretraining on Auxiliary Tasks. Variant G represents the backbone is pre-trained on ImageNet. Our full model performs better since pre-training on auxiliary tasks are essential to extract semantically-meaningful image features. The auxiliary 2D image conditions are generated from the ground-truth mesh of Interhand2.6M training split. We did not use extra datasets.", + "bbox": [ + 498, + 536, + 890, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 657, + 617, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel approach MeMaHand for two-hand mesh reconstruction from a single image. The mesh-mano interaction module combines the merits of non-parametric and parametric representations. Then, the mesh alignment refinement module further rectifies the results with an explicit mesh alignment feedback loop. Extensive experiments on the InterHand2.6M benchmark demonstrate that our proposed MeMaHand is superior to both existing parametric and non-parametric methods.", + "bbox": [ + 496, + 684, + 890, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although our method generates promising results, it still fails in cases of severe occlusions. Fig. 10 shows some failure cases, where inter-penetration occurs. Taking the physical plausibility into consideration will be our future work.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "571", + "bbox": [ + 485, + 945, + 508, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1067-1076, 2019. 2", + "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interacting objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6121-6131, 2020. 2", + "[3] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10843-10852, 2019. 1, 2, 6, 7", + "[4] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 666-682, 2018. 1", + "[5] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12929–12938, 2021. 1, 2", + "[6] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274–13283, 2021. 1, 2", + "[7] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10451-10460, 2021. 1, 2", + "[8] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. 2", + "[9] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016. 3, 4", + "[10] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10833-10842, 2019. 1, 2, 3", + "[11] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 571-580, 2020. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 5", + "[14] Dong Uk Kim, Kwang In Kim, and Seunglyul Baek. End-to-end detection and pose estimation of two interacting hands. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11189-11198, 2021. 2", + "[15] Nikolaos Kyriazis and Antonis Argyros. Scalable 3d tracking of multiple interacting objects. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3430-3437, 2014. 2", + "[16] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 1, 2, 3, 5, 6, 7", + "[17] Guan Ming Lim, Prayook Jatesiktat, and Wei Tech Ang. Mobilehand: Real-time 3d hand shape and pose estimation from color image. In International Conference on Neural Information Processing, pages 450-459. Springer, 2020. 2", + "[18] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. 1, 2", + "[19] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. 1, 2, 3, 4", + "[20] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 1, 2", + "[21] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 2", + "[22] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision, pages 548-564. Springer, 2020. 2, 5, 6", + "[23] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Trans. Graph., 36(6), nov 2017. 1, 2", + "[24] Breannan Smith, Chenglei Wu, He Wen, Patrick Peluse, Yaser Sheikh, Jessica K Hodgins, and Takaaki Shiratori. Constraining dense hand surface tracking with elasticity. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "572", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 6", + "[26] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11698-11707, October 2021. 2", + "[27] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. International Journal of Computer Vision, 118(2):172-193, 2016. 2", + "[28] Jiayi Wang, Franziska Mueller, Florian Bernard, Suzanne Sorli, Oleksandr Sotnychenko, Neng Qian, Miguel A Otaduy, Dan Casas, and Christian Theobalt. Rgb2hands: real-time tracking of 3d hand interactions from monocular rgb video. ACM Transactions on Graphics (ToG), 39(6):1-16, 2020. 7", + "[29] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2", + "[30] Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma, and Hongan Wang. Interacting two-hand 3d pose and shape reconstruction from single color image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11354-11363, 2021. 2, 5, 6, 7", + "[31] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. 4, 5", + "[32] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 1, 2", + "[33] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 1, 2, 6, 7", + "[34] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 6" + ], + "bbox": [ + 78, + 90, + 468, + 821 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "573", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_model.json b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..877b25757288799592959f41934ab3a60625a993 --- /dev/null +++ b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.24, + 0.131, + 0.734, + 0.177 + ], + "angle": 0, + "content": "MeMaHand: Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.203, + 0.651, + 0.24 + ], + "angle": 0, + "content": "Congyi Wang* Feida Zhu* Shilei Wen† ByteDance" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.242, + 0.712, + 0.258 + ], + "angle": 0, + "content": "{kongyi.1990，zhufeida，zhengmin.666}@bytedance.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.474, + 0.642 + ], + "angle": 0, + "content": "Existing methods proposed for hand reconstruction tasks usually parameterize a generic 3D hand model or predict hand mesh positions directly. The parametric representations consisting of hand shapes and rotational poses are more stable, while the non-parametric methods can predict more accurate mesh positions. In this paper, we propose to reconstruct meshes and estimate MANO parameters of two hands from a single RGB image simultaneously to utilize the merits of two kinds of hand representations. To fulfill this target, we propose novel Mesh-Mano interaction blocks (MMIBs), which take mesh vertices positions and MANO parameters as two kinds of query tokens. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. The transformer encoders are equipped with different asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Moreover, we introduce the mesh alignment refinement module to further enhance the mesh-image alignment. Extensive experiments on the InterHand2.6M benchmark demonstrate promising results over the state-of-the-art hand reconstruction methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.671, + 0.21, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.802 + ], + "angle": 0, + "content": "Vision-based 3D hand analysis plays an important role in many applications such as virtual reality (VR) and augmented reality (AR). Two-hand reconstruction from a single RGB image is more challenging due to complex mutual interactions and occlusions. Besides, the skin appearance similarity makes it difficult for the network to align image features to the corresponding hand." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.802, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Previous hand reconstruction works can be divided into two categories, parametric methods [3, 7, 32, 33] and nonparametric methods [4-6, 10, 18-20]. Parametric methods typically learn to regress pose and shape parameters of MANO model [23], where pose represents joint rota" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.291, + 0.872, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.515, + 0.895, + 0.586 + ], + "angle": 0, + "content": "Figure 1. Comparison with the state-of-the-art method IntagHand [16] for single-image two-hand reconstruction. The integration of parametric and non-parametric hand representations allows us to achieve better performance in hard cases such as severe occlusions and challenging viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.894, + 0.688 + ], + "angle": 0, + "content": "tions in axis-angle representation and shape represents the coefficients of shape PCA bases. The MANO prior can yield plausible hand shapes from a single monocular image. However, they can not produce fine-grained hand meshes due to their limited capacity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.894, + 0.795 + ], + "angle": 0, + "content": "With the rapid progress of graph convolutional network (GCN) and transformer techniques [10, 16, 18, 19], it is observed that direct mesh reconstruction can achieve state-of-the-art performance towards the Euclidean distances between the ground truth vertices and the predicted vertices. Nonetheless, the non-parametric methods are less robust in handling challenging viewpoints or severe occlusions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this paper, we introduce a novel single-image two-hand reconstruction method designed to predict mesh vertices positions and estimate MANO parameters simultaneously to utilize the merits of two kinds of hand representations. The proposed Mesh-Mano interaction Hand reconstruction architecture (MeMaHand) consists of three modules: 1) the image encoder-decoder module, 2) the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.341, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "564" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.514 + ], + "angle": 0, + "content": "mesh-mano interaction module, 3) and the mesh alignment refinement module. To extract contextually meaningful image features, we pre-train a classical image encoder-decoder network on auxiliary tasks including hand segmentation, hand 2D joints and dense mapping encodings. The low-resolution features encode more global knowledge, while the high-resolution features contain more local details. Secondly, the mesh-mano interaction module stacks three mesh-mano interaction blocks (MMIBs) to transform the mesh vertices and MANO parameters queries initialized by the global image feature vector. We observe that the hand prior embedded in the MANO parameters is valuable for predicting stable hand meshes in challenging situations such as severe occlusions. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. The transformer encoders are equipped with different asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Each MMIB is followed by an upsampling operation to upsample the mesh vertices tokens in a coarse-to-fine manner. Finally, the mesh alignment refinement module utilizes one MMIB to predict offsets for mesh vertices and MANO parameters to enhance mesh-image alignment. To improve the reliability of image evidence, we project mesh vertices predicted by the Mesh-Mano interaction module onto the 2D image plane. The explicit mesh-aligned image features are concatenated to the transformer input tokens." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.473, + 0.65 + ], + "angle": 0, + "content": "The whole network, including the pre-trained image feature encoder-decoder, is jointly optimized such that the image features better adapt to our hand mesh reconstruction task. Benefiting from the mesh-mano interaction mechanism and mesh alignment refinement stage, extensive experiments demonstrate that our method outperforms existing both parametric and non-parametric methods on Inter-Hand2.6M [22] dataset. In summary, the contributions of our approach are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.655, + 0.47, + 0.729 + ], + "angle": 0, + "content": "- We propose MeMaHand to integrate the merits of parametric and non-parametric hand representation. The mesh vertices and MANO parameters are mutually reinforced to achieve better performance in mesh recovery and parameter regression." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.731, + 0.47, + 0.79 + ], + "angle": 0, + "content": "- A mesh-image alignment feedback loop is utilized to improve the reliability of image evidence. Therefore, more accurate predictions are obtained by rectifying the mesh-image misalignment." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.792, + 0.47, + 0.836 + ], + "angle": 0, + "content": "- Our method achieves superior performance on the InterHand2.6M dataset, compared with both non-parametric and parametric methods." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.655, + 0.47, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.228, + 0.86 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Parametric Hand Reconstruction. Parametric approaches [1-3, 7, 11, 12, 23, 32, 33] use a parametric hand model such" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "as MANO [23] and focus on regressing the pose and shape parameters from a single image. The rich embedded prior information (e.g. the geometric dynamic constraints of joint rotations) can assist deep model learning when 3D annotations are insufficient. Weak 2D joints supervision [3] and motion caption data [33] are utilized to train convolutional neural networks (CNN) to predict MANO parameters. The reliance on 3D manual annotations is further totally alleviated by S²HAND [7]. However, the reconstructed mesh can not fully express the local details of variable 3D hand shapes due to the limited capacity of the MANO model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.895, + 0.476 + ], + "angle": 0, + "content": "Non-parametric Hand Reconstruction. Non-parametric methods aim to reconstruct the hand mesh directly [5, 6, 8, 10, 16-21, 26, 29]. To explicitly encode mesh topology, graph convolutional network (GCN) is adopted for aggregating adjacent vertices features. Hierarchical architectures [10, 16] are designed for mesh generation from coarse to fine. To model long-range dependencies, multi-layer transformer encoders are introduced such that the global interactions can be modeled without being limited by any mesh topology [18, 19]. Most recently, IntagHand [16] achieves the state-of-the-art performance on InterHand2.6M dataset [22]. However, we observe that IntagHand is less robust in handling challenging viewpoints or severe occlusions. The reconstructed mesh may be corrupted into unnatural shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.895, + 0.768 + ], + "angle": 0, + "content": "Interacting Two-Hand Reconstruction. Although single-hand methods can extend to two-hand reconstruction, the correlation between the left and right hands is not considered. Besides, the performance deteriorates for the close-interacting hands. Depth cameras [15] are sensitive to tracking accuracy and multi-view images [24, 27] are expensive to acquire. Based on a large-scale interacting hand dataset named InterHand2.6M [22], deep learning based single image methods either estimate hand joint positions [14, 22], or employ a 2.5D heatmap to predict the MANO parameters [30], or directly reconstruct meshes [16]. In summary, existing methods only employ a single specific hand representation. In contrast, our method not only combines the merits of MANO representations and mesh representations but also proposes a dedicated transformer encoder with asymmetric attention masks to make them collaborate seamlessly. Besides, the joint rotation and precise mesh vertices positions can be used for different applications, such as driving CG characters or virtual try-on." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.637, + 0.801 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The overall architecture of MeMaHand is depicted in Fig. 2. Given a single image \\(I\\), the proposed method can predict the mesh vertices positions \\(V \\in \\mathbb{R}^{778 \\times 3}\\), MANO pose parameter \\(\\theta \\in \\mathbb{R}^{48}\\) and shape parameter \\(\\beta \\in \\mathbb{R}^{10}\\) of two hands simultaneously in one forward pass. In this section, we first describe the overall architecture of MeMa" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "565" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.885, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.418, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our proposed MeMaHand. We pretrain an image encoder-decoder module on auxiliary tasks to extract multi-scale contextually meaningful image features. The Mesh-Mano interaction module stacks three Mesh-Mano interaction blocks (MMIBs) to predict the mesh vertices positions and MANO parameters simultaneously. The mesh alignment refinement module rectifies the mesh-image misalignment by explicitly utilizing the mesh-align image features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.562 + ], + "angle": 0, + "content": "Hand in Sec. 3.1. Then, we illustrate the mesh-mano interaction module and mesh-alignment refinement module in Sec. 3.2 and Sec. 3.3, respectively. Finally, we introduce the model objectives used to train our network in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.247, + 0.589 + ], + "angle": 0, + "content": "3.1. System Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.734 + ], + "angle": 0, + "content": "To extract contextually meaningful image features, we pre-train a classical image encoder-decoder module on auxiliary tasks including hand segmentation, hand 2D joints, and dense mapping encodings. The 2D image conditions are generated from the reconstructed mesh of ground-truth MANO parameters. ResNet50 [13] is leveraged as the backbone. Global average pooling feature \\( F_{gap} \\) and multi-scaled image feature \\( \\{\\phi_i \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}, i = 1, 2, 3\\} \\) are extracted from the image encoder (E) and decoder (D)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.734, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Afterward, we propose Mesh-Mano Interaction Module to predict mesh vertices positions \\( V_{MMIM} \\) and MANO parameters \\( \\theta_{MMIM}, \\beta_{MMIM} \\) of both left (L) and right (R) hands. We use the global feature \\( F_{gap} \\) to initialize the vertex and MANO queries. The grid-sampled image features \\( \\{\\phi_i^{grid}, i = 1,2,3\\} \\) are tokenized as well following the practice of Mesh Graphformer [19]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The mesh alignment refinement module further improves the predictions generated by the mesh-mano interaction module. The explicit mesh-aligned image features extracted from the multi-scale feature maps \\(\\phi_{i}\\) are more informative" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.501, + 0.892, + 0.533 + ], + "angle": 0, + "content": "than grid-sampled image features, which helps rectify the predictions. Details will be elaborated in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.539, + 0.783, + 0.554 + ], + "angle": 0, + "content": "3.2. Mesh-Mano Interaction Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.699 + ], + "angle": 0, + "content": "The mesh-mano interaction module reconstructs hand mesh in a coarse-to-fine manner with three Mesh-Mano Interaction Blocks (MMIBs). We leverage the graph coarsening method [16] to build three-level sub meshes with vertex number \\( N_{1} = 63 \\), \\( N_{2} = 126 \\), \\( N_{3} = 252 \\). Each MMIB is followed by an upsample operation that reverses the topological relationship between adjacent sub-meshes. The full mesh vertices positions (\\( N = 778 \\)) are obtained with a simple \\( 1 \\times 1 \\) Conv from the final output vertex token." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Fig. 3 shows the detailed structure of MMIB. At level \\(i\\), the input tokens of MMIB include two kinds of queries: vertex queries \\(f_{v}^{h}\\in \\mathbb{R}^{N_{i}\\times D_{i}}\\) and MANO parameter queries \\(f_{m}^{h}\\in \\mathbb{R}^{2\\times D_{i}}\\), where \\(h\\) indicates left \\((L)\\) or right \\((R)\\) hand, and \\(D_{i}\\) represents the feature dimensions. Therefore, the total sequence length is \\((2\\times N_i + 4)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Graph Residual Block. The design of the graph residual block is similar to [10, 16]. The Chebyshev spectral graph CNN [9] is adopted to transform the vertex token to intermediate graph features," + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.86, + 0.891, + 0.88 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {g r a p h}} = \\mathbf {G r a p h C o n v} \\left(f _ {v} ^ {L}, f _ {v} ^ {R}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where the operation GraphConv denotes the graph con" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "566" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.09, + 0.437, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.422 + ], + "angle": 0, + "content": "Figure 3. The detailed structure of Mesh-Mano Interaction Block (MMIB). MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.492 + ], + "angle": 0, + "content": "volution. More details can be found in [9]. For simplicity, we use \\( f_{graph} \\) to represent the concatenation of updated left-hand and right-hand vertex tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.471, + 0.83 + ], + "angle": 0, + "content": "Intra-hand Transformer Encoder. While graph CNN is useful for extracting local information, it is less efficient at capturing long-range dependencies. We use transformer encoders to model dependencies not only between long-range vertices but also between mesh vertices and MANO parameters. The spatial image features at resolution \\(i\\) are also tokenized as \\(\\phi_i^{grid}\\) by grid sampling following the practice of Mesh Graphformer [19]. Therefore, the input tokens of intra-hand transformer encoder (TFEintra) consist of graph features \\((f_{graph})\\), MANO queries \\((f_m^L, f_m^R)\\) and grid image feature \\((\\phi_i^{grid})\\). Based on the observation that the mesh vertices regression can lead to more precise mesh reconstruction and the MANO parameters are more stable, we propose an asymmetric attention mask excluding the manoto-mesh attention such that the MANO parameter tokens will not directly affect the mesh vertex tokens. The mesh-to-mano attention remains. Therefore, the mesh vertex tokens are implicitly regularized by the MANO parameters loss. TFEintra focuses on modeling the intra-hand dependencies. The inter-hand attention is also excluded. The resulting asymmetric attention mask \\(M_{intra}\\) is shown in Fig. 3. Finally, the intra-hand transformer encoder is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.838, + 0.47, + 0.871 + ], + "angle": 0, + "content": "\\[\nf _ {i n t r a} = T F E _ {i n t r a} \\left(\\left[ f _ {g r a p h}, f _ {m} ^ {L}, f _ {m} ^ {R}, \\phi_ {i} ^ {g r i d} \\right], M _ {i n t r a}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\([\\ldots]\\) denotes the concatenation operation. For simplicity, \\(f_{\\text{intra}}\\) represents the concatenation of updated mesh" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.747, + 0.107 + ], + "angle": 0, + "content": "vertex and MANO parameter tokens." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.112, + 0.892, + 0.203 + ], + "angle": 0, + "content": "Inter-hand Transformer Encoder. In contrast to \\(TFE_{intra}\\), the inter-hand transformer encoder \\((TFE_{inter})\\) focuses on modeling the inter-hand correlations. Fig. 3 presents the resulting asymmetric attention mask \\(M_{inter}\\). The image feature tokens are not used in \\(TFE_{inter}\\). Finally, the inter-hand transformer encoder is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.22, + 0.892, + 0.236 + ], + "angle": 0, + "content": "\\[\nf _ {i n t e r} = T F E _ {i n t e r} \\left(f _ {i n t r a}, M _ {i n t e r}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.252, + 0.891, + 0.283 + ], + "angle": 0, + "content": "where \\(f_{inter}\\) can be further split into updated vertex tokens \\(\\bar{f}_v^L,\\bar{f}_v^R\\) and MANO parameter tokens \\(\\bar{f}_m^L,\\bar{f}_m^R\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.39 + ], + "angle": 0, + "content": "Our proposed MMIB predicts the mesh vertices positions and MANO parameters simultaneously in a unified architecture. Intra-hand and inter-hand mesh-mano dependencies are modeled by two cascaded transformer encoders with different attention masks. After three MMIBs, a simple \\(1 \\times 1\\) Conv is applied to obtain the mesh vertices positions and MANO parameters from output tokens," + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.404, + 0.891, + 0.423 + ], + "angle": 0, + "content": "\\[\nV _ {M M I M} ^ {h} = \\operatorname {C o n v} _ {v} \\left(\\bar {f} _ {v} ^ {h}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.426, + 0.891, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\theta_ {M M I M} ^ {h}, \\beta_ {M M I M} ^ {h} = C o n v _ {m} \\left(\\bar {f} _ {m} ^ {h}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.521 + ], + "angle": 0, + "content": "where \\( h \\in \\{L, R\\} \\) indicates the left or right hand. To tackle the noticeable misalignment between the estimated meshes and image evidence, we propose a mesh alignment refinement module to rectify the results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.536, + 0.82, + 0.551 + ], + "angle": 0, + "content": "3.3. Mesh Alignment Refinement Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.697 + ], + "angle": 0, + "content": "To further improve mesh-image alignment, we propose a novel mesh-alignment refinement module inspired by PyMAF [31]. Specifically, we project the mesh vertices \\( V_{MMIM} \\) predicted by the mesh-mano interaction module onto the multi-scale image features explicitly. Three-scale mesh-aligned image features are concatenated together for each corresponding vertex token. A simple multi-layer perceptron (MLP) is adopted to reduce the dimensions, resulting in a fused image feature \\( \\phi_v \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.82 + ], + "angle": 0, + "content": "We utilize one MMIB to refine the mesh vertex and MANO parameter tokens \\(\\bar{f}_v^h, \\bar{f}_m^h\\). To effectively utilize the mesh-aligned image evidence, we made some modifications to the tokens. Specifically, we concatenate the image feature \\(\\phi_v\\) to vertex token \\(\\bar{f}_v^h\\) along the channel dimension. For MANO parameter tokens, we perform a global average pooling operation to obtain global feature vector \\(\\phi_m\\) before concatenation. The rectified tokens are formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.834, + 0.891, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {v} ^ {h}, \\hat {f} _ {m} ^ {h} = M M I B \\left(\\bar {f} _ {v} ^ {h}, \\bar {f} _ {m} ^ {h}, \\phi_ {v}, \\phi_ {m}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "One simple \\(1 \\times 1\\) Conv layer is applied to obtain the mesh vertices positions offsets and MANO parameter off" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "567" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.218, + 0.106 + ], + "angle": 0, + "content": "sets for rectification," + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.115, + 0.469, + 0.134 + ], + "angle": 0, + "content": "\\[\nV ^ {h} = V _ {M M I M} ^ {h} + \\operatorname {C o n v} _ {v} \\left(\\hat {f} _ {v} ^ {h}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.137, + 0.469, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\theta^ {h} = \\theta_ {M M I M} ^ {h} + C o n v _ {m} \\left(\\hat {f} _ {m} ^ {h}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.159, + 0.469, + 0.177 + ], + "angle": 0, + "content": "\\[\n\\beta^ {h} = \\beta_ {M M I M} ^ {h} + \\operatorname {C o n v} _ {m} \\left(\\hat {f} _ {m} ^ {h}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.187, + 0.47, + 0.324 + ], + "angle": 0, + "content": "where \\(V^h\\), \\(\\theta^h\\) and \\(\\beta^h\\) are the final results. Note that we can stack several MMIBs to predict the offsets iteratively. In our experiments, one single MMIB is enough to achieve satisfactory results. PyMAF [31] fuses the mesh-align features into one global feature, which ignores the inherent spatial relationship of mesh vertices. In contrast, our MMIB adopts GCN to model the spatial relation between adjacent mesh vertices. The transformer encoder is responsible for modeling the long-range dependencies." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.333, + 0.248, + 0.349 + ], + "angle": 0, + "content": "3.4. Model Objectives" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.469, + 0.416 + ], + "angle": 0, + "content": "Our method predicts mesh vertices positions and MANO parameters simultaneously. The learning objectives can be divided into three categories: mesh vertex loss, MANO parameter loss and mesh-mano consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.419, + 0.469, + 0.45 + ], + "angle": 0, + "content": "Mesh Vertex Loss: The widely-used L1 loss is adopted to supervise the vertex positions and the 2D projections:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.459, + 0.469, + 0.491 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V} = \\sum_ {h = L, R} \\| V ^ {h} - V _ {G T} ^ {h} \\| _ {1} + \\| \\Pi (V _ {h}) - \\Pi (V _ {G T} ^ {h}) \\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.502, + 0.469, + 0.564 + ], + "angle": 0, + "content": "where \\( V_{GT}^{h} \\) represents the ground-truth vertex positions and \\( \\Pi \\) denotes the projection function. Given vertex positions, the joint positions can be regressed by multiplying the predefined regression matrix \\( \\mathcal{J} \\). Joint losses are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.573, + 0.469, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {J} = \\sum_ {h = L, R} \\| \\mathcal {J} V ^ {h} - \\mathcal {J} V _ {G T} ^ {h} \\| _ {1} \\\\ + \\sum_ {h = L, R} \\| \\Pi \\mathcal {J} \\left(V _ {h}\\right) - \\Pi \\left(\\mathcal {J} V _ {G T} ^ {h}\\right) \\| _ {1}. \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.651, + 0.469, + 0.681 + ], + "angle": 0, + "content": "The face normal loss is introduced to regularize the surface normal consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.691, + 0.469, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N} = \\sum_ {h = L, R} \\sum_ {f = 1} ^ {F} \\sum_ {i = 1} ^ {3} \\| e _ {f, i} ^ {h} \\cdot n _ {f, G T} ^ {h} \\| _ {1}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.744, + 0.469, + 0.807 + ], + "angle": 0, + "content": "where \\( e_{f,i}^{h} \\) represents the \\( i \\)th edge of face \\( f \\) at hand \\( h \\) and \\( n_{f,GT}^{h} \\) is the normal vector of this face from the ground truth mesh. The edge length consistency loss is to enforce the edge length consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.817, + 0.469, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E} = \\sum_ {h = L, R} \\sum_ {i = 1} ^ {E} \\| e _ {i} ^ {h} - e _ {i, G T} ^ {h} \\| _ {1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\( e_i^h \\) represents the \\( i \\)th edge of hand \\( h \\) and \\( E \\) denotes the total edge number, respectively." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.089, + 0.858, + 0.151 + ], + "angle": 0, + "content": "
MethodParams(M)Inference Time (ms)
InterShape [30]139.647.77
Intaghand [16]37.2851.86
Ours38.3160.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.162, + 0.862, + 0.176 + ], + "angle": 0, + "content": "Table 1. Parameters and inference speed on Tesla V100" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.203, + 0.892, + 0.234 + ], + "angle": 0, + "content": "MANO Parameter Loss: Given the ground truth MANO parameters, L1 loss is used to regress parameters," + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.245, + 0.892, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P} = \\sum_ {h = L, R} \\| \\theta^ {h} - \\theta_ {G T} ^ {h} \\| _ {1} + \\| \\beta^ {h} - \\beta_ {G T} ^ {h} \\| _ {1}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.29, + 0.892, + 0.335 + ], + "angle": 0, + "content": "In addition, we reconstruct hand mesh based on the MANO model. The reconstructed mesh should be close to the ground-truth mesh," + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.347, + 0.892, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V, \\text {m a n o}} = \\sum_ {h = L, R} \\| \\mathcal {M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - V _ {G T} ^ {h} \\| _ {1}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Mesh-Mano Consistency Loss. The predicted mesh and the reconstructed mesh from MANO parameters should be consistent with each other:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.449, + 0.891, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s i s t}} = \\sum_ {h = L, R} \\| \\mathcal {M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - V ^ {h} \\| _ {1} \\tag {16} \\\\ + \\left\\| \\mathcal {J M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - \\mathcal {J} V ^ {h} \\right\\| _ {1}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.515, + 0.797, + 0.53 + ], + "angle": 0, + "content": "In summary, the overall training loss \\(\\mathcal{L}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.542, + 0.891, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {V} \\mathcal {L} _ {V} + \\lambda_ {J} \\mathcal {L} _ {J} + \\lambda_ {N} \\mathcal {L} _ {N} + \\lambda_ {E} \\mathcal {L} _ {E} \\tag {17} \\\\ + \\lambda_ {P} \\mathcal {L} _ {P} + \\lambda_ {V, m a n o} \\mathcal {L} _ {V, m a n o} + \\lambda_ {\\text {c o n s i s t}} \\mathcal {L} _ {\\text {c o n s i s t}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.589, + 0.892, + 0.664 + ], + "angle": 0, + "content": "where \\(\\lambda_{V} = 40\\), \\(\\lambda_{J} = 40\\), \\(\\lambda_{N} = 5\\), \\(\\lambda_{E} = 40\\), \\(\\lambda_{P} = 10\\), \\(\\lambda_{V,mano} = 10\\) and \\(\\lambda_{consist} = 40\\). The whole network including the pre-trained image feature encoder-decoder is jointly optimized such that the image features better adapt to our hand mesh reconstruction task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.625, + 0.696 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.704, + 0.762, + 0.72 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.731, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Training Dataset. Interhand2.6M [22] is a large-scale hand dataset with ground truth mesh annotations including both single-hand and interacting two-hand images. We pick out the interacting hand (IH) data for training and testing. For a fair comparison, we follow the preprocessing steps of [16] which produces 366K training and 261K testing samples." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. The image encoder uses ResNet50 [13] as the backbone. The decoder contains four simple deconvolutional layers. We first pretrain our image encoder-decoder module on auxiliary tasks. Then, the whole network is jointly optimized with Adam optimizer at" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "568" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.484, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.091, + 0.885, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with SOTA parametric method InterShape [30] and SOTA non-parametric method IntagHand [16] on InterHand2.6M dataset. Our method performs better on close-interacting two-hand reconstruction (first row). Besides, our method is robust in hard cases such as severe occlusions and challenging viewpoints (second row)." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.309, + 0.169, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.435, + 0.16, + 0.445 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.309, + 0.337, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.435, + 0.223, + 0.444 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.435, + 0.306, + 0.444 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.318, + 0.487, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.435, + 0.456, + 0.444 + ], + "angle": 0, + "content": "Alternative Views" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.309, + 0.572, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.435, + 0.562, + 0.445 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.309, + 0.74, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.435, + 0.626, + 0.444 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.435, + 0.708, + 0.444 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.319, + 0.884, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.435, + 0.864, + 0.444 + ], + "angle": 0, + "content": "Alternative Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.457, + 0.769, + 0.471 + ], + "angle": 0, + "content": "Figure 5. Our method produces accurate hand reconstructions for various two-hand interactions." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.495, + 0.46, + 0.582 + ], + "angle": 0, + "content": "
MethodMPJPE↓MPVPE↓AUC ↑PROJ2D↓
zimmermann et al. [34]36.36---
Spurr et al. [25]15.40---
InterNet [22]16.00-0.711-
Intaghand [16]8.799.030.8066.47
Ours8.658.890.8326.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.593, + 0.47, + 0.62 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with state-of-art non-parametric methods on InterHand2.6M dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.723 + ], + "angle": 0, + "content": "a learning rate \\(1 \\times 10^{-4}\\). Data augmentation includes random rotation, random translation, and random scaling. The batch size is set to 32. It takes around 48 hours to train MeMaHand with 4 Tesla V100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Metrics. Mean Per Joint Position Error (MPJPE) and Mean Per Vertex Position Error (MPVPE) in millimeters are adopted to evaluate the mesh reconstruction accuracy. Additionally, we report the percentage of correct keypoints (PCK) curve and Area Under the Curve (AUC) across the thresholds between 0 and 50 millimeters. To evaluate the mesh-image alignment accuracy, the reconstructed mesh vertices are projected onto the 2D image plane. The mesh-image alignment accuracy is measured by PROJ2D, which calculates the distance in image pixels between the projected ground truth vertices and the predicted vertices." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.866, + 0.514 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.692 + ], + "angle": 0, + "content": "We compare our model with state-of-the-art non-parametric methods including InterNet [22] and Intaghand [16] and parametric methods including [3], MinimalHand [33] and InterShape [30]. The officially released weights are used to obtain the results. The model parameters and inference time of our method and other SOTA methods are reported in Tab. 1. The inference of our model can be completed in \\(60.52ms\\), comparable to other SOTAs but with better mesh reconstruction results. Besides, the mesh vertex positions as well as the MANO parameters are predicted simultaneously in one single forward pass." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Comparison with Non-parametric Methods. Non-parametric methods directly generate mesh vertex positions, which can express the local details of variable 3D hand shapes. The quantitative results of non-parametric methods are shown in Tab. 2. It can be seen that our method achieves the best performance on all evaluation metrics. Fig. 6(a) presents the PCK curve, which further demonstrates the superiority of our method at all threshold levels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 presents the qualitative comparison with SOTA non-parametric method IntagHand [16]. Our method performs better on close interacting two-hand reconstruction (first row). Besides, we observe that IntagHand produces collapsed meshes in hard cases such as severe occlusions" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "569" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.112, + 0.442, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.3, + 0.411, + 0.312 + ], + "angle": 0, + "content": "(a) Comparison with non-parametric method" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.112, + 0.828, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.3, + 0.786, + 0.312 + ], + "angle": 0, + "content": "(b) Comparison with parametric method" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.894, + 0.355 + ], + "angle": 0, + "content": "Figure 6. Comparison with the state-of-the-art methods on InterHand2.6M dataset. Left: 3D PCK of non-parametric methods. Right: 3D PCK of parametric methods. The AUC values are shown in parentheses." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.376, + 0.465, + 0.449 + ], + "angle": 0, + "content": "
MethodMPJPE↓MPVPE ↓AUC↑PROJ2D↓
Boukhayma et al. [3]16.9317.980.672-
MinimalHand [33]23.4823.890.543-
InterShape [30]13.4813.950.7359.95
Ours10.8510.920.7906.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.458, + 0.47, + 0.486 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison with state-of-art parametric methods on InterHand2.6M dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.592 + ], + "angle": 0, + "content": "and challenging viewpoints (second row). In contrast, our method is more robust to such situations. We attribute this success to the integration of MANO representations, which predicts stable hand meshes. More results with alternative views are shown in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Comparison with Parametric Methods. For a fair comparison, the reconstructed hand meshes from our predicted MANO parameters are utilized for evaluation. The quantitative comparisons are listed in Tab. 3. The PCK curve is shown in Fig. 6(a). The MANO parameter estimation of our method outperforms other SOTA parametric methods. Thanks to the mesh-mano interaction block, the MANO token is conditioned on the mesh vertex tokens where spatial relations are retained. Such dependencies are more informative than one-dimensional global feature utilized in competing methods [3, 30, 33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "From Tab. 2 and Tab. 3, we can find the parametric method InterShape [30] has higher statistical errors on joint and vertex positions compared with the non-parametric method IntagHand [16] due to the limited capacity of the parametric model. However, from the second row of Fig. 4, we can see InterShape produces reasonable hand shapes in these challenging cases. By combining the merits of parametric and non-parametric hand representations, our method achieves the best performance both quantitatively" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.377, + 0.895, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.538, + 0.894, + 0.568 + ], + "angle": 0, + "content": "Figure 7. Two-hand reconstruction of images in the wild, which are taken from the RGB2Hands [28]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.594, + 0.618, + 0.609 + ], + "angle": 0, + "content": "and qualitatively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.783, + 0.635 + ], + "angle": 0, + "content": "4.3. Extension to Images in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.823 + ], + "angle": 0, + "content": "We further present the hand reconstruction results on images in the wild. As shown in Fig. 7, our method performs well on real-life images taken from the RGB2Hands dataset [28], which demonstrates the generalization ability of our approach. Recall that our method is designed to predict accurate mesh vertices positions and MANO parameters simultaneously in one forward pass. The MANO pose parameter represents the joint rotations in axis-angle representation, which is useful for animating 3D hands in computer graphics. Accurate hand mesh vertices positions and MANO parameters for hands in the wild can facilitate different human-computer-interaction (HCI) applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.655, + 0.849 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the proposed modules in our framework, we conduct an ablation study on several variants of our method. The quantitative results of all vari-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "570" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.476, + 0.215 + ], + "angle": 0, + "content": "
MPJPE↓MPVPE↓AUC↑PROJ2D↓
A: w/o MANO token8.879.090.8186.85
B: w/o Mesh-Align8.799.010.8266.64
C: Mesh-Align scale-168.738.970.8296.38
D: Mesh-Align scale-648.778.990.8266.58
E: Regress MANO from Fgap8.839.040.8286.86
F: w/o asymmetric attention9.179.370.8226.67
G: w/o auxiliary tasks9.119.280.8236.78
Full Model8.658.890.8326.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.226, + 0.444, + 0.24 + ], + "angle": 0, + "content": "Table 4. Comparisons of different variants of our method." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.259, + 0.227, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.259, + 0.321, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.259, + 0.416, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.427, + 0.403, + 0.441 + ], + "angle": 0, + "content": "Figure 8. Ablation study on MANO tokens." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.458, + 0.159, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.459, + 0.317, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.459, + 0.469, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.604, + 0.441, + 0.618 + ], + "angle": 0, + "content": "Figure 9. Ablation study on mesh alignment refinement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.649, + 0.272, + 0.664 + ], + "angle": 0, + "content": "ants are presented in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.469, + 0.79 + ], + "angle": 0, + "content": "Effectiveness of MANO token. Variant A (w/o MANO token) represents removing the MANO tokens. The transformer encoder only models the dependencies between the mesh vertex tokens and image features. Fig. 8 show the qualitative comparison. Without hand prior information, the reconstructed mesh may be corrupted when parts of the hands are occluded. In contrast, our full model can generate reasonable meshes in these challenging situations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Mesh Alignment Refinement. Variant B (w/o Mesh-Align) represents removing the mesh alignment refinement module. The predictions generated by the Mesh-Mano interaction module are taken as the final outputs. Variant C and D denote extracting the single-scale image features at resolution 16 and 64, respectively. The performances degrade without using mesh align refinement" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.09, + 0.84, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.259, + 0.856, + 0.273 + ], + "angle": 0, + "content": "Figure 10. Failures cases caused by severe occlusions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.302, + 0.89, + 0.332 + ], + "angle": 0, + "content": "or using single-scale image features. As shown in Fig. 9, our full model produces better-aligned results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.336, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Effectiveness of MMIB. Parametric and non-parametric representations are widely used for many years. However, combining them is not trivial. In variant \\(\\mathbf{E}\\) (Regress MANO from \\(F_{gap}\\)), MMIB is responsible for updating vertex tokens. The MANO parameters are regressed from the global average pooling features \\(F_{gap}\\) with simple fully connected layers rather than through MMIB. The performance gain of variant \\(\\mathbf{E}\\) is limited as shown in Tab. 4. On the other hand, we propose an asymmetric attention mask to exclude the mano-to-mesh attention. To verify the choice, variant \\(\\mathbf{F}\\) (w/o asymmetric attention mask) denotes removing the asymmetric attention mechanism. The performance degrades compared with our full model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.642 + ], + "angle": 0, + "content": "Effectiveness of Pretraining on Auxiliary Tasks. Variant G represents the backbone is pre-trained on ImageNet. Our full model performs better since pre-training on auxiliary tasks are essential to extract semantically-meaningful image features. The auxiliary 2D image conditions are generated from the ground-truth mesh of Interhand2.6M training split. We did not use extra datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.659, + 0.618, + 0.674 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.821 + ], + "angle": 0, + "content": "In this paper, we propose a novel approach MeMaHand for two-hand mesh reconstruction from a single image. The mesh-mano interaction module combines the merits of non-parametric and parametric representations. Then, the mesh alignment refinement module further rectifies the results with an explicit mesh alignment feedback loop. Extensive experiments on the InterHand2.6M benchmark demonstrate that our proposed MeMaHand is superior to both existing parametric and non-parametric methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Limitations. Although our method generates promising results, it still fails in cases of severe occlusions. Fig. 10 shows some failure cases, where inter-penetration occurs. Taking the physical plausibility into consideration will be our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.509, + 0.957 + ], + "angle": 0, + "content": "571" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1067-1076, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interacting objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6121-6131, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[3] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10843-10852, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.47, + 0.368 + ], + "angle": 0, + "content": "[4] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 666-682, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[5] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12929–12938, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[6] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274–13283, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.619 + ], + "angle": 0, + "content": "[7] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10451-10460, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[8] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.691, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[9] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[10] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10833-10842, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[11] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 571-580, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[12] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[14] Dong Uk Kim, Kwang In Kim, and Seunglyul Baek. End-to-end detection and pose estimation of two interacting hands. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11189-11198, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[15] Nikolaos Kyriazis and Antonis Argyros. Scalable 3d tracking of multiple interacting objects. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3430-3437, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[16] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[17] Guan Ming Lim, Prayook Jatesiktat, and Wei Tech Ang. Mobilehand: Real-time 3d hand shape and pose estimation from color image. In International Conference on Neural Information Processing, pages 450-459. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[18] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.581 + ], + "angle": 0, + "content": "[19] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[20] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[21] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[22] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision, pages 548-564. Springer, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[23] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Trans. Graph., 36(6), nov 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[24] Breannan Smith, Chenglei Wu, He Wen, Patrick Peluse, Yaser Sheikh, Jessica K Hodgins, and Takaaki Shiratori. Constraining dense hand surface tracking with elasticity. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "572" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[25] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[26] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11698-11707, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.288 + ], + "angle": 0, + "content": "[27] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. International Journal of Computer Vision, 118(2):172-193, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[28] Jiayi Wang, Franziska Mueller, Florian Bernard, Suzanne Sorli, Oleksandr Sotnychenko, Neng Qian, Miguel A Otaduy, Dan Casas, and Christian Theobalt. Rgb2hands: real-time tracking of 3d hand interactions from monocular rgb video. ACM Transactions on Graphics (ToG), 39(6):1-16, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[29] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.445, + 0.469, + 0.527 + ], + "angle": 0, + "content": "[30] Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma, and Hongan Wang. Interacting two-hand 3d pose and shape reconstruction from single color image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11354-11363, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.529, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[31] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.681 + ], + "angle": 0, + "content": "[32] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.684, + 0.469, + 0.766 + ], + "angle": 0, + "content": "[33] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.768, + 0.469, + 0.823 + ], + "angle": 0, + "content": "[34] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.823 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "573" + } + ] +] \ No newline at end of file diff --git a/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_origin.pdf b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b936999b6629ee051d5d5a5aa7b9be7f94423564 --- /dev/null +++ b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/867425da-f5a0-40aa-92d2-6339df5b377e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c061b3365351d7e72f281cc04bfce234f100128c7bea2da9c4b29543cde3d2be +size 5187705 diff --git a/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/full.md b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..20ffbabe1e7b368378429c9bf9f37d0ec0949daf --- /dev/null +++ b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/full.md @@ -0,0 +1,356 @@ +# MeMaHand: Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction + +Congyi Wang* Feida Zhu* Shilei Wen† ByteDance + +{kongyi.1990，zhufeida，zhengmin.666}@bytedance.com + +# Abstract + +Existing methods proposed for hand reconstruction tasks usually parameterize a generic 3D hand model or predict hand mesh positions directly. The parametric representations consisting of hand shapes and rotational poses are more stable, while the non-parametric methods can predict more accurate mesh positions. In this paper, we propose to reconstruct meshes and estimate MANO parameters of two hands from a single RGB image simultaneously to utilize the merits of two kinds of hand representations. To fulfill this target, we propose novel Mesh-Mano interaction blocks (MMIBs), which take mesh vertices positions and MANO parameters as two kinds of query tokens. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. The transformer encoders are equipped with different asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Moreover, we introduce the mesh alignment refinement module to further enhance the mesh-image alignment. Extensive experiments on the InterHand2.6M benchmark demonstrate promising results over the state-of-the-art hand reconstruction methods. + +# 1. Introduction + +Vision-based 3D hand analysis plays an important role in many applications such as virtual reality (VR) and augmented reality (AR). Two-hand reconstruction from a single RGB image is more challenging due to complex mutual interactions and occlusions. Besides, the skin appearance similarity makes it difficult for the network to align image features to the corresponding hand. + +Previous hand reconstruction works can be divided into two categories, parametric methods [3, 7, 32, 33] and nonparametric methods [4-6, 10, 18-20]. Parametric methods typically learn to regress pose and shape parameters of MANO model [23], where pose represents joint rota + +![](images/628c321180901f9334be6329b8db67b50dbb40e122d8f523a49885606cccd297.jpg) +Figure 1. Comparison with the state-of-the-art method IntagHand [16] for single-image two-hand reconstruction. The integration of parametric and non-parametric hand representations allows us to achieve better performance in hard cases such as severe occlusions and challenging viewpoints. + +tions in axis-angle representation and shape represents the coefficients of shape PCA bases. The MANO prior can yield plausible hand shapes from a single monocular image. However, they can not produce fine-grained hand meshes due to their limited capacity. + +With the rapid progress of graph convolutional network (GCN) and transformer techniques [10, 16, 18, 19], it is observed that direct mesh reconstruction can achieve state-of-the-art performance towards the Euclidean distances between the ground truth vertices and the predicted vertices. Nonetheless, the non-parametric methods are less robust in handling challenging viewpoints or severe occlusions. + +In this paper, we introduce a novel single-image two-hand reconstruction method designed to predict mesh vertices positions and estimate MANO parameters simultaneously to utilize the merits of two kinds of hand representations. The proposed Mesh-Mano interaction Hand reconstruction architecture (MeMaHand) consists of three modules: 1) the image encoder-decoder module, 2) the + +mesh-mano interaction module, 3) and the mesh alignment refinement module. To extract contextually meaningful image features, we pre-train a classical image encoder-decoder network on auxiliary tasks including hand segmentation, hand 2D joints and dense mapping encodings. The low-resolution features encode more global knowledge, while the high-resolution features contain more local details. Secondly, the mesh-mano interaction module stacks three mesh-mano interaction blocks (MMIBs) to transform the mesh vertices and MANO parameters queries initialized by the global image feature vector. We observe that the hand prior embedded in the MANO parameters is valuable for predicting stable hand meshes in challenging situations such as severe occlusions. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. The transformer encoders are equipped with different asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Each MMIB is followed by an upsampling operation to upsample the mesh vertices tokens in a coarse-to-fine manner. Finally, the mesh alignment refinement module utilizes one MMIB to predict offsets for mesh vertices and MANO parameters to enhance mesh-image alignment. To improve the reliability of image evidence, we project mesh vertices predicted by the Mesh-Mano interaction module onto the 2D image plane. The explicit mesh-aligned image features are concatenated to the transformer input tokens. + +The whole network, including the pre-trained image feature encoder-decoder, is jointly optimized such that the image features better adapt to our hand mesh reconstruction task. Benefiting from the mesh-mano interaction mechanism and mesh alignment refinement stage, extensive experiments demonstrate that our method outperforms existing both parametric and non-parametric methods on Inter-Hand2.6M [22] dataset. In summary, the contributions of our approach are as follows: + +- We propose MeMaHand to integrate the merits of parametric and non-parametric hand representation. The mesh vertices and MANO parameters are mutually reinforced to achieve better performance in mesh recovery and parameter regression. +- A mesh-image alignment feedback loop is utilized to improve the reliability of image evidence. Therefore, more accurate predictions are obtained by rectifying the mesh-image misalignment. +- Our method achieves superior performance on the InterHand2.6M dataset, compared with both non-parametric and parametric methods. + +# 2. Related Works + +Parametric Hand Reconstruction. Parametric approaches [1-3, 7, 11, 12, 23, 32, 33] use a parametric hand model such + +as MANO [23] and focus on regressing the pose and shape parameters from a single image. The rich embedded prior information (e.g. the geometric dynamic constraints of joint rotations) can assist deep model learning when 3D annotations are insufficient. Weak 2D joints supervision [3] and motion caption data [33] are utilized to train convolutional neural networks (CNN) to predict MANO parameters. The reliance on 3D manual annotations is further totally alleviated by S²HAND [7]. However, the reconstructed mesh can not fully express the local details of variable 3D hand shapes due to the limited capacity of the MANO model. + +Non-parametric Hand Reconstruction. Non-parametric methods aim to reconstruct the hand mesh directly [5, 6, 8, 10, 16-21, 26, 29]. To explicitly encode mesh topology, graph convolutional network (GCN) is adopted for aggregating adjacent vertices features. Hierarchical architectures [10, 16] are designed for mesh generation from coarse to fine. To model long-range dependencies, multi-layer transformer encoders are introduced such that the global interactions can be modeled without being limited by any mesh topology [18, 19]. Most recently, IntagHand [16] achieves the state-of-the-art performance on InterHand2.6M dataset [22]. However, we observe that IntagHand is less robust in handling challenging viewpoints or severe occlusions. The reconstructed mesh may be corrupted into unnatural shapes. + +Interacting Two-Hand Reconstruction. Although single-hand methods can extend to two-hand reconstruction, the correlation between the left and right hands is not considered. Besides, the performance deteriorates for the close-interacting hands. Depth cameras [15] are sensitive to tracking accuracy and multi-view images [24, 27] are expensive to acquire. Based on a large-scale interacting hand dataset named InterHand2.6M [22], deep learning based single image methods either estimate hand joint positions [14, 22], or employ a 2.5D heatmap to predict the MANO parameters [30], or directly reconstruct meshes [16]. In summary, existing methods only employ a single specific hand representation. In contrast, our method not only combines the merits of MANO representations and mesh representations but also proposes a dedicated transformer encoder with asymmetric attention masks to make them collaborate seamlessly. Besides, the joint rotation and precise mesh vertices positions can be used for different applications, such as driving CG characters or virtual try-on. + +# 3. Methodology + +The overall architecture of MeMaHand is depicted in Fig. 2. Given a single image $I$ , the proposed method can predict the mesh vertices positions $V \in \mathbb{R}^{778 \times 3}$ , MANO pose parameter $\theta \in \mathbb{R}^{48}$ and shape parameter $\beta \in \mathbb{R}^{10}$ of two hands simultaneously in one forward pass. In this section, we first describe the overall architecture of MeMa + +![](images/9cff5c1c98fd285d7f2cf0186dd56bc1357acdc7bda3818b244ad0550f425b7f.jpg) +Figure 2. The overall architecture of our proposed MeMaHand. We pretrain an image encoder-decoder module on auxiliary tasks to extract multi-scale contextually meaningful image features. The Mesh-Mano interaction module stacks three Mesh-Mano interaction blocks (MMIBs) to predict the mesh vertices positions and MANO parameters simultaneously. The mesh alignment refinement module rectifies the mesh-image misalignment by explicitly utilizing the mesh-align image features. + +Hand in Sec. 3.1. Then, we illustrate the mesh-mano interaction module and mesh-alignment refinement module in Sec. 3.2 and Sec. 3.3, respectively. Finally, we introduce the model objectives used to train our network in Sec. 3.4. + +# 3.1. System Overview + +To extract contextually meaningful image features, we pre-train a classical image encoder-decoder module on auxiliary tasks including hand segmentation, hand 2D joints, and dense mapping encodings. The 2D image conditions are generated from the reconstructed mesh of ground-truth MANO parameters. ResNet50 [13] is leveraged as the backbone. Global average pooling feature $F_{gap}$ and multi-scaled image feature $\{\phi_i \in \mathbb{R}^{C_i \times H_i \times W_i}, i = 1, 2, 3\}$ are extracted from the image encoder (E) and decoder (D). + +Afterward, we propose Mesh-Mano Interaction Module to predict mesh vertices positions $V_{MMIM}$ and MANO parameters $\theta_{MMIM}, \beta_{MMIM}$ of both left (L) and right (R) hands. We use the global feature $F_{gap}$ to initialize the vertex and MANO queries. The grid-sampled image features $\{\phi_i^{grid}, i = 1,2,3\}$ are tokenized as well following the practice of Mesh Graphformer [19]. + +The mesh alignment refinement module further improves the predictions generated by the mesh-mano interaction module. The explicit mesh-aligned image features extracted from the multi-scale feature maps $\phi_{i}$ are more informative + +than grid-sampled image features, which helps rectify the predictions. Details will be elaborated in Sec. 3.3. + +# 3.2. Mesh-Mano Interaction Module + +The mesh-mano interaction module reconstructs hand mesh in a coarse-to-fine manner with three Mesh-Mano Interaction Blocks (MMIBs). We leverage the graph coarsening method [16] to build three-level sub meshes with vertex number $N_{1} = 63$ , $N_{2} = 126$ , $N_{3} = 252$ . Each MMIB is followed by an upsample operation that reverses the topological relationship between adjacent sub-meshes. The full mesh vertices positions ( $N = 778$ ) are obtained with a simple $1 \times 1$ Conv from the final output vertex token. + +Fig. 3 shows the detailed structure of MMIB. At level $i$ , the input tokens of MMIB include two kinds of queries: vertex queries $f_{v}^{h}\in \mathbb{R}^{N_{i}\times D_{i}}$ and MANO parameter queries $f_{m}^{h}\in \mathbb{R}^{2\times D_{i}}$ , where $h$ indicates left $(L)$ or right $(R)$ hand, and $D_{i}$ represents the feature dimensions. Therefore, the total sequence length is $(2\times N_i + 4)$ . + +Graph Residual Block. The design of the graph residual block is similar to [10, 16]. The Chebyshev spectral graph CNN [9] is adopted to transform the vertex token to intermediate graph features, + +$$ +f _ {\text {g r a p h}} = \mathbf {G r a p h C o n v} \left(f _ {v} ^ {L}, f _ {v} ^ {R}\right), \tag {1} +$$ + +where the operation GraphConv denotes the graph con + +![](images/53d71bfedb9ec3aebdce209f4c65a95584025888a1a899d4dc58284b0f4f3e70.jpg) +Figure 3. The detailed structure of Mesh-Mano Interaction Block (MMIB). MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. + +volution. More details can be found in [9]. For simplicity, we use $f_{graph}$ to represent the concatenation of updated left-hand and right-hand vertex tokens. + +Intra-hand Transformer Encoder. While graph CNN is useful for extracting local information, it is less efficient at capturing long-range dependencies. We use transformer encoders to model dependencies not only between long-range vertices but also between mesh vertices and MANO parameters. The spatial image features at resolution $i$ are also tokenized as $\phi_i^{grid}$ by grid sampling following the practice of Mesh Graphformer [19]. Therefore, the input tokens of intra-hand transformer encoder (TFEintra) consist of graph features $(f_{graph})$ , MANO queries $(f_m^L, f_m^R)$ and grid image feature $(\phi_i^{grid})$ . Based on the observation that the mesh vertices regression can lead to more precise mesh reconstruction and the MANO parameters are more stable, we propose an asymmetric attention mask excluding the manoto-mesh attention such that the MANO parameter tokens will not directly affect the mesh vertex tokens. The mesh-to-mano attention remains. Therefore, the mesh vertex tokens are implicitly regularized by the MANO parameters loss. TFEintra focuses on modeling the intra-hand dependencies. The inter-hand attention is also excluded. The resulting asymmetric attention mask $M_{intra}$ is shown in Fig. 3. Finally, the intra-hand transformer encoder is formulated as: + +$$ +f _ {i n t r a} = T F E _ {i n t r a} \left(\left[ f _ {g r a p h}, f _ {m} ^ {L}, f _ {m} ^ {R}, \phi_ {i} ^ {g r i d} \right], M _ {i n t r a}\right), \tag {2} +$$ + +where $[\ldots]$ denotes the concatenation operation. For simplicity, $f_{\text{intra}}$ represents the concatenation of updated mesh + +vertex and MANO parameter tokens. + +Inter-hand Transformer Encoder. In contrast to $TFE_{intra}$ , the inter-hand transformer encoder $(TFE_{inter})$ focuses on modeling the inter-hand correlations. Fig. 3 presents the resulting asymmetric attention mask $M_{inter}$ . The image feature tokens are not used in $TFE_{inter}$ . Finally, the inter-hand transformer encoder is formulated as: + +$$ +f _ {i n t e r} = T F E _ {i n t e r} \left(f _ {i n t r a}, M _ {i n t e r}\right), \tag {3} +$$ + +where $f_{inter}$ can be further split into updated vertex tokens $\bar{f}_v^L,\bar{f}_v^R$ and MANO parameter tokens $\bar{f}_m^L,\bar{f}_m^R$ . + +Our proposed MMIB predicts the mesh vertices positions and MANO parameters simultaneously in a unified architecture. Intra-hand and inter-hand mesh-mano dependencies are modeled by two cascaded transformer encoders with different attention masks. After three MMIBs, a simple $1 \times 1$ Conv is applied to obtain the mesh vertices positions and MANO parameters from output tokens, + +$$ +V _ {M M I M} ^ {h} = \operatorname {C o n v} _ {v} \left(\bar {f} _ {v} ^ {h}\right), \tag {4} +$$ + +$$ +\theta_ {M M I M} ^ {h}, \beta_ {M M I M} ^ {h} = C o n v _ {m} \left(\bar {f} _ {m} ^ {h}\right), \tag {5} +$$ + +where $h \in \{L, R\}$ indicates the left or right hand. To tackle the noticeable misalignment between the estimated meshes and image evidence, we propose a mesh alignment refinement module to rectify the results. + +# 3.3. Mesh Alignment Refinement Module + +To further improve mesh-image alignment, we propose a novel mesh-alignment refinement module inspired by PyMAF [31]. Specifically, we project the mesh vertices $V_{MMIM}$ predicted by the mesh-mano interaction module onto the multi-scale image features explicitly. Three-scale mesh-aligned image features are concatenated together for each corresponding vertex token. A simple multi-layer perceptron (MLP) is adopted to reduce the dimensions, resulting in a fused image feature $\phi_v$ . + +We utilize one MMIB to refine the mesh vertex and MANO parameter tokens $\bar{f}_v^h, \bar{f}_m^h$ . To effectively utilize the mesh-aligned image evidence, we made some modifications to the tokens. Specifically, we concatenate the image feature $\phi_v$ to vertex token $\bar{f}_v^h$ along the channel dimension. For MANO parameter tokens, we perform a global average pooling operation to obtain global feature vector $\phi_m$ before concatenation. The rectified tokens are formulated as, + +$$ +\hat {f} _ {v} ^ {h}, \hat {f} _ {m} ^ {h} = M M I B \left(\bar {f} _ {v} ^ {h}, \bar {f} _ {m} ^ {h}, \phi_ {v}, \phi_ {m}\right). \tag {6} +$$ + +One simple $1 \times 1$ Conv layer is applied to obtain the mesh vertices positions offsets and MANO parameter off + +sets for rectification, + +$$ +V ^ {h} = V _ {M M I M} ^ {h} + \operatorname {C o n v} _ {v} \left(\hat {f} _ {v} ^ {h}\right), \tag {7} +$$ + +$$ +\theta^ {h} = \theta_ {M M I M} ^ {h} + C o n v _ {m} \left(\hat {f} _ {m} ^ {h}\right), \tag {8} +$$ + +$$ +\beta^ {h} = \beta_ {M M I M} ^ {h} + \operatorname {C o n v} _ {m} \left(\hat {f} _ {m} ^ {h}\right), \tag {9} +$$ + +where $V^h$ , $\theta^h$ and $\beta^h$ are the final results. Note that we can stack several MMIBs to predict the offsets iteratively. In our experiments, one single MMIB is enough to achieve satisfactory results. PyMAF [31] fuses the mesh-align features into one global feature, which ignores the inherent spatial relationship of mesh vertices. In contrast, our MMIB adopts GCN to model the spatial relation between adjacent mesh vertices. The transformer encoder is responsible for modeling the long-range dependencies. + +# 3.4. Model Objectives + +Our method predicts mesh vertices positions and MANO parameters simultaneously. The learning objectives can be divided into three categories: mesh vertex loss, MANO parameter loss and mesh-mano consistency loss. + +Mesh Vertex Loss: The widely-used L1 loss is adopted to supervise the vertex positions and the 2D projections: + +$$ +\mathcal {L} _ {V} = \sum_ {h = L, R} \| V ^ {h} - V _ {G T} ^ {h} \| _ {1} + \| \Pi (V _ {h}) - \Pi (V _ {G T} ^ {h}) \| _ {1}, \tag {10} +$$ + +where $V_{GT}^{h}$ represents the ground-truth vertex positions and $\Pi$ denotes the projection function. Given vertex positions, the joint positions can be regressed by multiplying the predefined regression matrix $\mathcal{J}$ . Joint losses are formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {J} = \sum_ {h = L, R} \| \mathcal {J} V ^ {h} - \mathcal {J} V _ {G T} ^ {h} \| _ {1} \\ + \sum_ {h = L, R} \| \Pi \mathcal {J} \left(V _ {h}\right) - \Pi \left(\mathcal {J} V _ {G T} ^ {h}\right) \| _ {1}. \tag {11} \\ \end{array} +$$ + +The face normal loss is introduced to regularize the surface normal consistency: + +$$ +\mathcal {L} _ {N} = \sum_ {h = L, R} \sum_ {f = 1} ^ {F} \sum_ {i = 1} ^ {3} \| e _ {f, i} ^ {h} \cdot n _ {f, G T} ^ {h} \| _ {1}, \tag {12} +$$ + +where $e_{f,i}^{h}$ represents the $i$ th edge of face $f$ at hand $h$ and $n_{f,GT}^{h}$ is the normal vector of this face from the ground truth mesh. The edge length consistency loss is to enforce the edge length consistency: + +$$ +\mathcal {L} _ {E} = \sum_ {h = L, R} \sum_ {i = 1} ^ {E} \| e _ {i} ^ {h} - e _ {i, G T} ^ {h} \| _ {1}, \tag {13} +$$ + +where $e_i^h$ represents the $i$ th edge of hand $h$ and $E$ denotes the total edge number, respectively. + +
MethodParams(M)Inference Time (ms)
InterShape [30]139.647.77
Intaghand [16]37.2851.86
Ours38.3160.52
+ +Table 1. Parameters and inference speed on Tesla V100 + +MANO Parameter Loss: Given the ground truth MANO parameters, L1 loss is used to regress parameters, + +$$ +\mathcal {L} _ {P} = \sum_ {h = L, R} \| \theta^ {h} - \theta_ {G T} ^ {h} \| _ {1} + \| \beta^ {h} - \beta_ {G T} ^ {h} \| _ {1}. \tag {14} +$$ + +In addition, we reconstruct hand mesh based on the MANO model. The reconstructed mesh should be close to the ground-truth mesh, + +$$ +\mathcal {L} _ {V, \text {m a n o}} = \sum_ {h = L, R} \| \mathcal {M A N O} \left(\theta^ {h}, \beta^ {h}\right) - V _ {G T} ^ {h} \| _ {1}. \tag {15} +$$ + +Mesh-Mano Consistency Loss. The predicted mesh and the reconstructed mesh from MANO parameters should be consistent with each other: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n s i s t}} = \sum_ {h = L, R} \| \mathcal {M A N O} \left(\theta^ {h}, \beta^ {h}\right) - V ^ {h} \| _ {1} \tag {16} \\ + \left\| \mathcal {J M A N O} \left(\theta^ {h}, \beta^ {h}\right) - \mathcal {J} V ^ {h} \right\| _ {1}. \\ \end{array} +$$ + +In summary, the overall training loss $\mathcal{L}$ is: + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {V} \mathcal {L} _ {V} + \lambda_ {J} \mathcal {L} _ {J} + \lambda_ {N} \mathcal {L} _ {N} + \lambda_ {E} \mathcal {L} _ {E} \tag {17} \\ + \lambda_ {P} \mathcal {L} _ {P} + \lambda_ {V, m a n o} \mathcal {L} _ {V, m a n o} + \lambda_ {\text {c o n s i s t}} \mathcal {L} _ {\text {c o n s i s t}}. \\ \end{array} +$$ + +where $\lambda_{V} = 40$ , $\lambda_{J} = 40$ , $\lambda_{N} = 5$ , $\lambda_{E} = 40$ , $\lambda_{P} = 10$ , $\lambda_{V,mano} = 10$ and $\lambda_{consist} = 40$ . The whole network including the pre-trained image feature encoder-decoder is jointly optimized such that the image features better adapt to our hand mesh reconstruction task. + +# 4. Experiment + +# 4.1. Datasets and Implementation + +Training Dataset. Interhand2.6M [22] is a large-scale hand dataset with ground truth mesh annotations including both single-hand and interacting two-hand images. We pick out the interacting hand (IH) data for training and testing. For a fair comparison, we follow the preprocessing steps of [16] which produces 366K training and 261K testing samples. + +Implementation Details. The image encoder uses ResNet50 [13] as the backbone. The decoder contains four simple deconvolutional layers. We first pretrain our image encoder-decoder module on auxiliary tasks. Then, the whole network is jointly optimized with Adam optimizer at + +![](images/d01db973666f0219532d9ae561a584bfbafbb2680e4736f23f58e7541f44b90f.jpg) +Figure 4. Qualitative comparison with SOTA parametric method InterShape [30] and SOTA non-parametric method IntagHand [16] on InterHand2.6M dataset. Our method performs better on close-interacting two-hand reconstruction (first row). Besides, our method is robust in hard cases such as severe occlusions and challenging viewpoints (second row). + +![](images/eb5876f9ba615edb94708ccff8196db4b01e0a695811ed507b616b3b88e224d6.jpg) + +![](images/fecd2dd1be525dc5ad17d7a9f80acf9e6828cd819a68b5970ac4f2e3f4d94c07.jpg) +Input Image + +![](images/25a5505b0a9f7a1d25d38c2a80c88bcf5c7281c3aae6a9f9bd7d1089c38be9e1.jpg) +GT +Ours +Figure 5. Our method produces accurate hand reconstructions for various two-hand interactions. + +![](images/c8db48590528402b273c7a72ae3840a497a8547c6eb040a7375277de90deb696.jpg) +Alternative Views + +![](images/de4242ab860c49e4f0baa8f44193dc83ec2cea1fc17c298b6e90fe3e3d09c921.jpg) +Input Image + +![](images/e0531ff7aa698fb3d91f278aa1e0ac520501ff30441245a56318502f846f3f95.jpg) +GT +Ours + +![](images/2b9034f4a6fd92e1c166ed031211c71447b235b395822787787c388a663a0daa.jpg) +Alternative Views + +
MethodMPJPE↓MPVPE↓AUC ↑PROJ2D↓
zimmermann et al. [34]36.36---
Spurr et al. [25]15.40---
InterNet [22]16.00-0.711-
Intaghand [16]8.799.030.8066.47
Ours8.658.890.8326.22
+ +Table 2. Quantitative comparison with state-of-art non-parametric methods on InterHand2.6M dataset. + +a learning rate $1 \times 10^{-4}$ . Data augmentation includes random rotation, random translation, and random scaling. The batch size is set to 32. It takes around 48 hours to train MeMaHand with 4 Tesla V100 GPUs. + +Metrics. Mean Per Joint Position Error (MPJPE) and Mean Per Vertex Position Error (MPVPE) in millimeters are adopted to evaluate the mesh reconstruction accuracy. Additionally, we report the percentage of correct keypoints (PCK) curve and Area Under the Curve (AUC) across the thresholds between 0 and 50 millimeters. To evaluate the mesh-image alignment accuracy, the reconstructed mesh vertices are projected onto the 2D image plane. The mesh-image alignment accuracy is measured by PROJ2D, which calculates the distance in image pixels between the projected ground truth vertices and the predicted vertices. + +# 4.2. Comparison with State-of-the-art Methods + +We compare our model with state-of-the-art non-parametric methods including InterNet [22] and Intaghand [16] and parametric methods including [3], MinimalHand [33] and InterShape [30]. The officially released weights are used to obtain the results. The model parameters and inference time of our method and other SOTA methods are reported in Tab. 1. The inference of our model can be completed in $60.52ms$ , comparable to other SOTAs but with better mesh reconstruction results. Besides, the mesh vertex positions as well as the MANO parameters are predicted simultaneously in one single forward pass. + +Comparison with Non-parametric Methods. Non-parametric methods directly generate mesh vertex positions, which can express the local details of variable 3D hand shapes. The quantitative results of non-parametric methods are shown in Tab. 2. It can be seen that our method achieves the best performance on all evaluation metrics. Fig. 6(a) presents the PCK curve, which further demonstrates the superiority of our method at all threshold levels. + +Fig. 4 presents the qualitative comparison with SOTA non-parametric method IntagHand [16]. Our method performs better on close interacting two-hand reconstruction (first row). Besides, we observe that IntagHand produces collapsed meshes in hard cases such as severe occlusions + +![](images/8832cb057416e9ea798c88119611ef0f99c5d11bb8530769c27b12b7366407ec.jpg) +(a) Comparison with non-parametric method + +![](images/3e065247fc52936c06f16d9272c9bd3a5ef806c3e8f2544598fbb6f7ddedf6d0.jpg) +(b) Comparison with parametric method + +
MethodMPJPE↓MPVPE ↓AUC↑PROJ2D↓
Boukhayma et al. [3]16.9317.980.672-
MinimalHand [33]23.4823.890.543-
InterShape [30]13.4813.950.7359.95
Ours10.8510.920.7906.90
+ +Table 3. Quantitative comparison with state-of-art parametric methods on InterHand2.6M dataset. + +and challenging viewpoints (second row). In contrast, our method is more robust to such situations. We attribute this success to the integration of MANO representations, which predicts stable hand meshes. More results with alternative views are shown in Figure 5. + +Comparison with Parametric Methods. For a fair comparison, the reconstructed hand meshes from our predicted MANO parameters are utilized for evaluation. The quantitative comparisons are listed in Tab. 3. The PCK curve is shown in Fig. 6(a). The MANO parameter estimation of our method outperforms other SOTA parametric methods. Thanks to the mesh-mano interaction block, the MANO token is conditioned on the mesh vertex tokens where spatial relations are retained. Such dependencies are more informative than one-dimensional global feature utilized in competing methods [3, 30, 33]. + +From Tab. 2 and Tab. 3, we can find the parametric method InterShape [30] has higher statistical errors on joint and vertex positions compared with the non-parametric method IntagHand [16] due to the limited capacity of the parametric model. However, from the second row of Fig. 4, we can see InterShape produces reasonable hand shapes in these challenging cases. By combining the merits of parametric and non-parametric hand representations, our method achieves the best performance both quantitatively + +![](images/b5de1be686022aa4fc1520cfc845f8fd2a2113d7fcae4aebacf137b61acb6221.jpg) +Figure 6. Comparison with the state-of-the-art methods on InterHand2.6M dataset. Left: 3D PCK of non-parametric methods. Right: 3D PCK of parametric methods. The AUC values are shown in parentheses. +Figure 7. Two-hand reconstruction of images in the wild, which are taken from the RGB2Hands [28]. + +and qualitatively. + +# 4.3. Extension to Images in the Wild + +We further present the hand reconstruction results on images in the wild. As shown in Fig. 7, our method performs well on real-life images taken from the RGB2Hands dataset [28], which demonstrates the generalization ability of our approach. Recall that our method is designed to predict accurate mesh vertices positions and MANO parameters simultaneously in one forward pass. The MANO pose parameter represents the joint rotations in axis-angle representation, which is useful for animating 3D hands in computer graphics. Accurate hand mesh vertices positions and MANO parameters for hands in the wild can facilitate different human-computer-interaction (HCI) applications. + +# 4.4. Ablation Study + +To evaluate the effectiveness of the proposed modules in our framework, we conduct an ablation study on several variants of our method. The quantitative results of all vari- + +
MPJPE↓MPVPE↓AUC↑PROJ2D↓
A: w/o MANO token8.879.090.8186.85
B: w/o Mesh-Align8.799.010.8266.64
C: Mesh-Align scale-168.738.970.8296.38
D: Mesh-Align scale-648.778.990.8266.58
E: Regress MANO from Fgap8.839.040.8286.86
F: w/o asymmetric attention9.179.370.8226.67
G: w/o auxiliary tasks9.119.280.8236.78
Full Model8.658.890.8326.22
+ +Table 4. Comparisons of different variants of our method. + +![](images/25bdbd62e90d44a4afff8232342ae0d4d54d2b8eca08fbdc12d147e4111e0092.jpg) +Figure 8. Ablation study on MANO tokens. + +![](images/4d4ecd5b8b81954f6dd360626b7169be44b210982978b02cbdebf4ace3171fb5.jpg) + +![](images/039af84d3432e6e7a86bfdcb531e00b66575811ccd126e0eae9469f666e915bc.jpg) + +![](images/29d567eecff4aedd51d7cf561fe6b8e79395b1dd7fc147a7d90b0f9664286e13.jpg) +Figure 9. Ablation study on mesh alignment refinement. + +![](images/2f0258ac679720dde1cbc7824c524cddb8f4294acfcfae84f9361a6c0c5a57e4.jpg) + +![](images/785f616a298a9505e8160e4a93ac19d1ff03dcfaf0ce6ab12d0b7c570cc0fac2.jpg) + +ants are presented in Table 4. + +Effectiveness of MANO token. Variant A (w/o MANO token) represents removing the MANO tokens. The transformer encoder only models the dependencies between the mesh vertex tokens and image features. Fig. 8 show the qualitative comparison. Without hand prior information, the reconstructed mesh may be corrupted when parts of the hands are occluded. In contrast, our full model can generate reasonable meshes in these challenging situations. + +Effectiveness of Mesh Alignment Refinement. Variant B (w/o Mesh-Align) represents removing the mesh alignment refinement module. The predictions generated by the Mesh-Mano interaction module are taken as the final outputs. Variant C and D denote extracting the single-scale image features at resolution 16 and 64, respectively. The performances degrade without using mesh align refinement + +![](images/f9c1cc86008180787f43161818858f9d25ad7400d0928f6a3124b2d550b8f9c3.jpg) +Figure 10. Failures cases caused by severe occlusions. + +or using single-scale image features. As shown in Fig. 9, our full model produces better-aligned results. + +Effectiveness of MMIB. Parametric and non-parametric representations are widely used for many years. However, combining them is not trivial. In variant $\mathbf{E}$ (Regress MANO from $F_{gap}$ ), MMIB is responsible for updating vertex tokens. The MANO parameters are regressed from the global average pooling features $F_{gap}$ with simple fully connected layers rather than through MMIB. The performance gain of variant $\mathbf{E}$ is limited as shown in Tab. 4. On the other hand, we propose an asymmetric attention mask to exclude the mano-to-mesh attention. To verify the choice, variant $\mathbf{F}$ (w/o asymmetric attention mask) denotes removing the asymmetric attention mechanism. The performance degrades compared with our full model. + +Effectiveness of Pretraining on Auxiliary Tasks. Variant G represents the backbone is pre-trained on ImageNet. Our full model performs better since pre-training on auxiliary tasks are essential to extract semantically-meaningful image features. The auxiliary 2D image conditions are generated from the ground-truth mesh of Interhand2.6M training split. We did not use extra datasets. + +# 5. Conclusion + +In this paper, we propose a novel approach MeMaHand for two-hand mesh reconstruction from a single image. The mesh-mano interaction module combines the merits of non-parametric and parametric representations. Then, the mesh alignment refinement module further rectifies the results with an explicit mesh alignment feedback loop. Extensive experiments on the InterHand2.6M benchmark demonstrate that our proposed MeMaHand is superior to both existing parametric and non-parametric methods. + +Limitations. Although our method generates promising results, it still fails in cases of severe occlusions. Fig. 10 shows some failure cases, where inter-penetration occurs. Taking the physical plausibility into consideration will be our future work. + +# References + +[1] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1067-1076, 2019. 2 +[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interacting objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6121-6131, 2020. 2 +[3] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10843-10852, 2019. 1, 2, 6, 7 +[4] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 666-682, 2018. 1 +[5] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12929–12938, 2021. 1, 2 +[6] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274–13283, 2021. 1, 2 +[7] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10451-10460, 2021. 1, 2 +[8] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. 2 +[9] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016. 3, 4 +[10] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10833-10842, 2019. 1, 2, 3 +[11] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 571-580, 2020. 2 + +[12] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 5 +[14] Dong Uk Kim, Kwang In Kim, and Seunglyul Baek. End-to-end detection and pose estimation of two interacting hands. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11189-11198, 2021. 2 +[15] Nikolaos Kyriazis and Antonis Argyros. Scalable 3d tracking of multiple interacting objects. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3430-3437, 2014. 2 +[16] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 1, 2, 3, 5, 6, 7 +[17] Guan Ming Lim, Prayook Jatesiktat, and Wei Tech Ang. Mobilehand: Real-time 3d hand shape and pose estimation from color image. In International Conference on Neural Information Processing, pages 450-459. Springer, 2020. 2 +[18] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. 1, 2 +[19] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. 1, 2, 3, 4 +[20] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 1, 2 +[21] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 2 +[22] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision, pages 548-564. Springer, 2020. 2, 5, 6 +[23] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Trans. Graph., 36(6), nov 2017. 1, 2 +[24] Breannan Smith, Chenglei Wu, He Wen, Patrick Peluse, Yaser Sheikh, Jessica K Hodgins, and Takaaki Shiratori. Constraining dense hand surface tracking with elasticity. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2 + +[25] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 6 +[26] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11698-11707, October 2021. 2 +[27] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. International Journal of Computer Vision, 118(2):172-193, 2016. 2 +[28] Jiayi Wang, Franziska Mueller, Florian Bernard, Suzanne Sorli, Oleksandr Sotnychenko, Neng Qian, Miguel A Otaduy, Dan Casas, and Christian Theobalt. Rgb2hands: real-time tracking of 3d hand interactions from monocular rgb video. ACM Transactions on Graphics (ToG), 39(6):1-16, 2020. 7 +[29] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2 +[30] Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma, and Hongan Wang. Interacting two-hand 3d pose and shape reconstruction from single color image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11354-11363, 2021. 2, 5, 6, 7 +[31] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. 4, 5 +[32] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 1, 2 +[33] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 1, 2, 6, 7 +[34] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 6 \ No newline at end of file diff --git a/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/images.zip b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..662f33ef323a3a9e1f3099b38b2401c0c7c71c0d --- /dev/null +++ b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2211779a53e5860860bdf3b4630de440918c51579bfef0658721862e33a0477b +size 648391 diff --git a/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/layout.json b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ebb952c2ab56d7c728c70c53de99813463f135ef --- /dev/null +++ b/2023/MeMaHand_ Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction/layout.json @@ -0,0 +1,8706 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "type": "text", + "content": "MeMaHand: Exploiting Mesh-Mano Interaction for Single Image Two-Hand Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 183, + 160, + 398, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 160, + 398, + 190 + ], + "spans": [ + { + "bbox": [ + 183, + 160, + 398, + 190 + ], + "type": "text", + "content": "Congyi Wang* Feida Zhu* Shilei Wen† ByteDance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 191, + 435, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 191, + 435, + 204 + ], + "spans": [ + { + "bbox": [ + 159, + 191, + 435, + 204 + ], + "type": "text", + "content": "{kongyi.1990，zhufeida，zhengmin.666}@bytedance.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 255, + 290, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 290, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 290, + 508 + ], + "type": "text", + "content": "Existing methods proposed for hand reconstruction tasks usually parameterize a generic 3D hand model or predict hand mesh positions directly. The parametric representations consisting of hand shapes and rotational poses are more stable, while the non-parametric methods can predict more accurate mesh positions. In this paper, we propose to reconstruct meshes and estimate MANO parameters of two hands from a single RGB image simultaneously to utilize the merits of two kinds of hand representations. To fulfill this target, we propose novel Mesh-Mano interaction blocks (MMIBs), which take mesh vertices positions and MANO parameters as two kinds of query tokens. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. The transformer encoders are equipped with different asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Moreover, we introduce the mesh alignment refinement module to further enhance the mesh-image alignment. Extensive experiments on the InterHand2.6M benchmark demonstrate promising results over the state-of-the-art hand reconstruction methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 551, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 635 + ], + "type": "text", + "content": "Vision-based 3D hand analysis plays an important role in many applications such as virtual reality (VR) and augmented reality (AR). Two-hand reconstruction from a single RGB image is more challenging due to complex mutual interactions and occlusions. Besides, the skin appearance similarity makes it difficult for the network to align image features to the corresponding hand." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 635, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 288, + 696 + ], + "type": "text", + "content": "Previous hand reconstruction works can be divided into two categories, parametric methods [3, 7, 32, 33] and nonparametric methods [4-6, 10, 18-20]. Parametric methods typically learn to regress pose and shape parameters of MANO model [23], where pose represents joint rota" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 230, + 533, + 403 + ], + "blocks": [ + { + "bbox": [ + 318, + 230, + 533, + 403 + ], + "lines": [ + { + "bbox": [ + 318, + 230, + 533, + 403 + ], + "spans": [ + { + "bbox": [ + 318, + 230, + 533, + 403 + ], + "type": "image", + "image_path": "628c321180901f9334be6329b8db67b50dbb40e122d8f523a49885606cccd297.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 407, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 407, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 547, + 464 + ], + "type": "text", + "content": "Figure 1. Comparison with the state-of-the-art method IntagHand [16] for single-image two-hand reconstruction. The integration of parametric and non-parametric hand representations allows us to achieve better performance in hard cases such as severe occlusions and challenging viewpoints." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 485, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 544 + ], + "type": "text", + "content": "tions in axis-angle representation and shape represents the coefficients of shape PCA bases. The MANO prior can yield plausible hand shapes from a single monocular image. However, they can not produce fine-grained hand meshes due to their limited capacity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 629 + ], + "type": "text", + "content": "With the rapid progress of graph convolutional network (GCN) and transformer techniques [10, 16, 18, 19], it is observed that direct mesh reconstruction can achieve state-of-the-art performance towards the Euclidean distances between the ground truth vertices and the predicted vertices. Nonetheless, the non-parametric methods are less robust in handling challenging viewpoints or severe occlusions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "In this paper, we introduce a novel single-image two-hand reconstruction method designed to predict mesh vertices positions and estimate MANO parameters simultaneously to utilize the merits of two kinds of hand representations. The proposed Mesh-Mano interaction Hand reconstruction architecture (MeMaHand) consists of three modules: 1) the image encoder-decoder module, 2) the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "564" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "mesh-mano interaction module, 3) and the mesh alignment refinement module. To extract contextually meaningful image features, we pre-train a classical image encoder-decoder network on auxiliary tasks including hand segmentation, hand 2D joints and dense mapping encodings. The low-resolution features encode more global knowledge, while the high-resolution features contain more local details. Secondly, the mesh-mano interaction module stacks three mesh-mano interaction blocks (MMIBs) to transform the mesh vertices and MANO parameters queries initialized by the global image feature vector. We observe that the hand prior embedded in the MANO parameters is valuable for predicting stable hand meshes in challenging situations such as severe occlusions. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies. The transformer encoders are equipped with different asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Each MMIB is followed by an upsampling operation to upsample the mesh vertices tokens in a coarse-to-fine manner. Finally, the mesh alignment refinement module utilizes one MMIB to predict offsets for mesh vertices and MANO parameters to enhance mesh-image alignment. To improve the reliability of image evidence, we project mesh vertices predicted by the Mesh-Mano interaction module onto the 2D image plane. The explicit mesh-aligned image features are concatenated to the transformer input tokens." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 407, + 289, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 289, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 289, + 514 + ], + "type": "text", + "content": "The whole network, including the pre-trained image feature encoder-decoder, is jointly optimized such that the image features better adapt to our hand mesh reconstruction task. Benefiting from the mesh-mano interaction mechanism and mesh alignment refinement stage, extensive experiments demonstrate that our method outperforms existing both parametric and non-parametric methods on Inter-Hand2.6M [22] dataset. In summary, the contributions of our approach are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 518, + 287, + 662 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 518, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 518, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 58, + 518, + 287, + 577 + ], + "type": "text", + "content": "- We propose MeMaHand to integrate the merits of parametric and non-parametric hand representation. The mesh vertices and MANO parameters are mutually reinforced to achieve better performance in mesh recovery and parameter regression." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 578, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 578, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 58, + 578, + 287, + 625 + ], + "type": "text", + "content": "- A mesh-image alignment feedback loop is utilized to improve the reliability of image evidence. Therefore, more accurate predictions are obtained by rectifying the mesh-image misalignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 627, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 627, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 58, + 627, + 287, + 662 + ], + "type": "text", + "content": "- Our method achieves superior performance on the InterHand2.6M dataset, compared with both non-parametric and parametric methods." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 670, + 139, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 139, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 139, + 681 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Parametric Hand Reconstruction. Parametric approaches [1-3, 7, 11, 12, 23, 32, 33] use a parametric hand model such" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "as MANO [23] and focus on regressing the pose and shape parameters from a single image. The rich embedded prior information (e.g. the geometric dynamic constraints of joint rotations) can assist deep model learning when 3D annotations are insufficient. Weak 2D joints supervision [3] and motion caption data [33] are utilized to train convolutional neural networks (CNN) to predict MANO parameters. The reliance on 3D manual annotations is further totally alleviated by S²HAND [7]. However, the reconstructed mesh can not fully express the local details of variable 3D hand shapes due to the limited capacity of the MANO model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 208, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 547, + 376 + ], + "type": "text", + "content": "Non-parametric Hand Reconstruction. Non-parametric methods aim to reconstruct the hand mesh directly [5, 6, 8, 10, 16-21, 26, 29]. To explicitly encode mesh topology, graph convolutional network (GCN) is adopted for aggregating adjacent vertices features. Hierarchical architectures [10, 16] are designed for mesh generation from coarse to fine. To model long-range dependencies, multi-layer transformer encoders are introduced such that the global interactions can be modeled without being limited by any mesh topology [18, 19]. Most recently, IntagHand [16] achieves the state-of-the-art performance on InterHand2.6M dataset [22]. However, we observe that IntagHand is less robust in handling challenging viewpoints or severe occlusions. The reconstructed mesh may be corrupted into unnatural shapes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 380, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 608 + ], + "type": "text", + "content": "Interacting Two-Hand Reconstruction. Although single-hand methods can extend to two-hand reconstruction, the correlation between the left and right hands is not considered. Besides, the performance deteriorates for the close-interacting hands. Depth cameras [15] are sensitive to tracking accuracy and multi-view images [24, 27] are expensive to acquire. Based on a large-scale interacting hand dataset named InterHand2.6M [22], deep learning based single image methods either estimate hand joint positions [14, 22], or employ a 2.5D heatmap to predict the MANO parameters [30], or directly reconstruct meshes [16]. In summary, existing methods only employ a single specific hand representation. In contrast, our method not only combines the merits of MANO representations and mesh representations but also proposes a dedicated transformer encoder with asymmetric attention masks to make them collaborate seamlessly. Besides, the joint rotation and precise mesh vertices positions can be used for different applications, such as driving CG characters or virtual try-on." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 620, + 389, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 389, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 389, + 634 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "The overall architecture of MeMaHand is depicted in Fig. 2. Given a single image " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": ", the proposed method can predict the mesh vertices positions " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{778 \\times 3}" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": ", MANO pose parameter " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{48}" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " and shape parameter " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " of two hands simultaneously in one forward pass. In this section, we first describe the overall architecture of MeMa" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "565" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 541, + 322 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 541, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 541, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 541, + 322 + ], + "type": "image", + "image_path": "9cff5c1c98fd285d7f2cf0186dd56bc1357acdc7bda3818b244ad0550f425b7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 546, + 377 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our proposed MeMaHand. We pretrain an image encoder-decoder module on auxiliary tasks to extract multi-scale contextually meaningful image features. The Mesh-Mano interaction module stacks three Mesh-Mano interaction blocks (MMIBs) to predict the mesh vertices positions and MANO parameters simultaneously. The mesh alignment refinement module rectifies the mesh-image misalignment by explicitly utilizing the mesh-align image features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": "Hand in Sec. 3.1. Then, we illustrate the mesh-mano interaction module and mesh-alignment refinement module in Sec. 3.2 and Sec. 3.3, respectively. Finally, we introduce the model objectives used to train our network in Sec. 3.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 151, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 151, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 151, + 466 + ], + "type": "text", + "content": "3.1. System Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": "To extract contextually meaningful image features, we pre-train a classical image encoder-decoder module on auxiliary tasks including hand segmentation, hand 2D joints, and dense mapping encodings. The 2D image conditions are generated from the reconstructed mesh of ground-truth MANO parameters. ResNet50 [13] is leveraged as the backbone. Global average pooling feature " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "F_{gap}" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": " and multi-scaled image feature " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\{\\phi_i \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}, i = 1, 2, 3\\}" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": " are extracted from the image encoder (E) and decoder (D)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "text", + "content": "Afterward, we propose Mesh-Mano Interaction Module to predict mesh vertices positions " + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "inline_equation", + "content": "V_{MMIM}" + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "text", + "content": " and MANO parameters " + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\theta_{MMIM}, \\beta_{MMIM}" + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "text", + "content": " of both left (L) and right (R) hands. We use the global feature " + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "inline_equation", + "content": "F_{gap}" + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "text", + "content": " to initialize the vertex and MANO queries. The grid-sampled image features " + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\{\\phi_i^{grid}, i = 1,2,3\\}" + }, + { + "bbox": [ + 47, + 581, + 287, + 665 + ], + "type": "text", + "content": " are tokenized as well following the practice of Mesh Graphformer [19]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "The mesh alignment refinement module further improves the predictions generated by the mesh-mano interaction module. The explicit mesh-aligned image features extracted from the multi-scale feature maps " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " are more informative" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 396, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 422 + ], + "type": "text", + "content": "than grid-sampled image features, which helps rectify the predictions. Details will be elaborated in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 426, + 479, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 479, + 438 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 479, + 438 + ], + "type": "text", + "content": "3.2. Mesh-Mano Interaction Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "text", + "content": "The mesh-mano interaction module reconstructs hand mesh in a coarse-to-fine manner with three Mesh-Mano Interaction Blocks (MMIBs). We leverage the graph coarsening method [16] to build three-level sub meshes with vertex number " + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "inline_equation", + "content": "N_{1} = 63" + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "inline_equation", + "content": "N_{2} = 126" + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "inline_equation", + "content": "N_{3} = 252" + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "text", + "content": ". Each MMIB is followed by an upsample operation that reverses the topological relationship between adjacent sub-meshes. The full mesh vertices positions (" + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "inline_equation", + "content": "N = 778" + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "text", + "content": ") are obtained with a simple " + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 445, + 545, + 553 + ], + "type": "text", + "content": " Conv from the final output vertex token." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": "Fig. 3 shows the detailed structure of MMIB. At level " + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": ", the input tokens of MMIB include two kinds of queries: vertex queries " + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "inline_equation", + "content": "f_{v}^{h}\\in \\mathbb{R}^{N_{i}\\times D_{i}}" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": " and MANO parameter queries " + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "inline_equation", + "content": "f_{m}^{h}\\in \\mathbb{R}^{2\\times D_{i}}" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": " indicates left " + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "inline_equation", + "content": "(L)" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": " or right " + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": " hand, and " + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": " represents the feature dimensions. Therefore, the total sequence length is " + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "inline_equation", + "content": "(2\\times N_i + 4)" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "text", + "content": "Graph Residual Block. The design of the graph residual block is similar to [10, 16]. The Chebyshev spectral graph CNN [9] is adopted to transform the vertex token to intermediate graph features," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 681, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 681, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 354, + 681, + 545, + 696 + ], + "type": "interline_equation", + "content": "f _ {\\text {g r a p h}} = \\mathbf {G r a p h C o n v} \\left(f _ {v} ^ {L}, f _ {v} ^ {R}\\right), \\tag {1}", + "image_path": "1786c5171a8dc1ad3aabb52a67e6929fbb03071e3320378c268275e45c56d710.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "where the operation GraphConv denotes the graph con" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "566" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 267, + 284 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 267, + 284 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 267, + 284 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 267, + 284 + ], + "type": "image", + "image_path": "53d71bfedb9ec3aebdce209f4c65a95584025888a1a899d4dc58284b0f4f3e70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "type": "text", + "content": "Figure 3. The detailed structure of Mesh-Mano Interaction Block (MMIB). MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range dependencies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 389 + ], + "type": "text", + "content": "volution. More details can be found in [9]. For simplicity, we use " + }, + { + "bbox": [ + 46, + 354, + 287, + 389 + ], + "type": "inline_equation", + "content": "f_{graph}" + }, + { + "bbox": [ + 46, + 354, + 287, + 389 + ], + "type": "text", + "content": " to represent the concatenation of updated left-hand and right-hand vertex tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "text", + "content": "Intra-hand Transformer Encoder. While graph CNN is useful for extracting local information, it is less efficient at capturing long-range dependencies. We use transformer encoders to model dependencies not only between long-range vertices but also between mesh vertices and MANO parameters. The spatial image features at resolution " + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "text", + "content": " are also tokenized as " + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "inline_equation", + "content": "\\phi_i^{grid}" + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "text", + "content": " by grid sampling following the practice of Mesh Graphformer [19]. Therefore, the input tokens of intra-hand transformer encoder (TFEintra) consist of graph features " + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "inline_equation", + "content": "(f_{graph})" + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "text", + "content": ", MANO queries " + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "inline_equation", + "content": "(f_m^L, f_m^R)" + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "text", + "content": " and grid image feature " + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "inline_equation", + "content": "(\\phi_i^{grid})" + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "text", + "content": ". Based on the observation that the mesh vertices regression can lead to more precise mesh reconstruction and the MANO parameters are more stable, we propose an asymmetric attention mask excluding the manoto-mesh attention such that the MANO parameter tokens will not directly affect the mesh vertex tokens. The mesh-to-mano attention remains. Therefore, the mesh vertex tokens are implicitly regularized by the MANO parameters loss. TFEintra focuses on modeling the intra-hand dependencies. The inter-hand attention is also excluded. The resulting asymmetric attention mask " + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "inline_equation", + "content": "M_{intra}" + }, + { + "bbox": [ + 46, + 392, + 288, + 657 + ], + "type": "text", + "content": " is shown in Fig. 3. Finally, the intra-hand transformer encoder is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 663, + 287, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 663, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 663, + 287, + 689 + ], + "type": "interline_equation", + "content": "f _ {i n t r a} = T F E _ {i n t r a} \\left(\\left[ f _ {g r a p h}, f _ {m} ^ {L}, f _ {m} ^ {R}, \\phi_ {i} ^ {g r i d} \\right], M _ {i n t r a}\\right), \\tag {2}", + "image_path": "842f2f179197ca4c33bd1b870a84ff2f57f1e5edebb76dbaf5af3e9e8f7b5642.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "[\\ldots]" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " denotes the concatenation operation. For simplicity, " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{\\text{intra}}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " represents the concatenation of updated mesh" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 457, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 457, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 457, + 84 + ], + "type": "text", + "content": "vertex and MANO parameter tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "text", + "content": "Inter-hand Transformer Encoder. In contrast to " + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "inline_equation", + "content": "TFE_{intra}" + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "text", + "content": ", the inter-hand transformer encoder " + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "inline_equation", + "content": "(TFE_{inter})" + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "text", + "content": " focuses on modeling the inter-hand correlations. Fig. 3 presents the resulting asymmetric attention mask " + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "inline_equation", + "content": "M_{inter}" + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "text", + "content": ". The image feature tokens are not used in " + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "inline_equation", + "content": "TFE_{inter}" + }, + { + "bbox": [ + 305, + 88, + 545, + 160 + ], + "type": "text", + "content": ". Finally, the inter-hand transformer encoder is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 174, + 545, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 174, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 351, + 174, + 545, + 186 + ], + "type": "interline_equation", + "content": "f _ {i n t e r} = T F E _ {i n t e r} \\left(f _ {i n t r a}, M _ {i n t e r}\\right), \\tag {3}", + "image_path": "43a66de9f225def412945d7988cbdbe70c6fd39037c82486a8ecd40b08fad893.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "type": "inline_equation", + "content": "f_{inter}" + }, + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "type": "text", + "content": " can be further split into updated vertex tokens " + }, + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\bar{f}_v^L,\\bar{f}_v^R" + }, + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "type": "text", + "content": " and MANO parameter tokens " + }, + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\bar{f}_m^L,\\bar{f}_m^R" + }, + { + "bbox": [ + 305, + 199, + 545, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 225, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 308 + ], + "type": "text", + "content": "Our proposed MMIB predicts the mesh vertices positions and MANO parameters simultaneously in a unified architecture. Intra-hand and inter-hand mesh-mano dependencies are modeled by two cascaded transformer encoders with different attention masks. After three MMIBs, a simple " + }, + { + "bbox": [ + 304, + 225, + 545, + 308 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 225, + 545, + 308 + ], + "type": "text", + "content": " Conv is applied to obtain the mesh vertices positions and MANO parameters from output tokens," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 393, + 319, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 319, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 393, + 319, + 545, + 335 + ], + "type": "interline_equation", + "content": "V _ {M M I M} ^ {h} = \\operatorname {C o n v} _ {v} \\left(\\bar {f} _ {v} ^ {h}\\right), \\tag {4}", + "image_path": "666967060a603c4f4e00bcaad5b2a86327d001b09155ad8d5fe2fc850c2c22dd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 337, + 545, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 337, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 354, + 337, + 545, + 352 + ], + "type": "interline_equation", + "content": "\\theta_ {M M I M} ^ {h}, \\beta_ {M M I M} ^ {h} = C o n v _ {m} \\left(\\bar {f} _ {m} ^ {h}\\right), \\tag {5}", + "image_path": "cd62cb2084e3ac42e791e9de8e68e117ab499d60174c5f3ed58f8cf74c8c62e6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 363, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 412 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 363, + 545, + 412 + ], + "type": "inline_equation", + "content": "h \\in \\{L, R\\}" + }, + { + "bbox": [ + 304, + 363, + 545, + 412 + ], + "type": "text", + "content": " indicates the left or right hand. To tackle the noticeable misalignment between the estimated meshes and image evidence, we propose a mesh alignment refinement module to rectify the results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 424, + 501, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 501, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 501, + 436 + ], + "type": "text", + "content": "3.3. Mesh Alignment Refinement Module" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 444, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 552 + ], + "type": "text", + "content": "To further improve mesh-image alignment, we propose a novel mesh-alignment refinement module inspired by PyMAF [31]. Specifically, we project the mesh vertices " + }, + { + "bbox": [ + 304, + 444, + 545, + 552 + ], + "type": "inline_equation", + "content": "V_{MMIM}" + }, + { + "bbox": [ + 304, + 444, + 545, + 552 + ], + "type": "text", + "content": " predicted by the mesh-mano interaction module onto the multi-scale image features explicitly. Three-scale mesh-aligned image features are concatenated together for each corresponding vertex token. A simple multi-layer perceptron (MLP) is adopted to reduce the dimensions, resulting in a fused image feature " + }, + { + "bbox": [ + 304, + 444, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\phi_v" + }, + { + "bbox": [ + 304, + 444, + 545, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": "We utilize one MMIB to refine the mesh vertex and MANO parameter tokens " + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\bar{f}_v^h, \\bar{f}_m^h" + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": ". To effectively utilize the mesh-aligned image evidence, we made some modifications to the tokens. Specifically, we concatenate the image feature " + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\phi_v" + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": " to vertex token " + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\bar{f}_v^h" + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": " along the channel dimension. For MANO parameter tokens, we perform a global average pooling operation to obtain global feature vector " + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\phi_m" + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": " before concatenation. The rectified tokens are formulated as," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 350, + 660, + 545, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 660, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 350, + 660, + 545, + 675 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {v} ^ {h}, \\hat {f} _ {m} ^ {h} = M M I B \\left(\\bar {f} _ {v} ^ {h}, \\bar {f} _ {m} ^ {h}, \\phi_ {v}, \\phi_ {m}\\right). \\tag {6}", + "image_path": "315dacf7cf41de6fdb794822415d9ba39532c9464afab9fa6c250c3cb0ada6d5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "One simple " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " Conv layer is applied to obtain the mesh vertices positions offsets and MANO parameter off" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "567" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 133, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 133, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 133, + 83 + ], + "type": "text", + "content": "sets for rectification," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 91, + 287, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 91, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 102, + 91, + 287, + 106 + ], + "type": "interline_equation", + "content": "V ^ {h} = V _ {M M I M} ^ {h} + \\operatorname {C o n v} _ {v} \\left(\\hat {f} _ {v} ^ {h}\\right), \\tag {7}", + "image_path": "7e83cb62dca68d160c8d4677e200b5d96dcbda83848321ae5ed4d4fdabe3cef8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 108, + 287, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 108, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 287, + 122 + ], + "type": "interline_equation", + "content": "\\theta^ {h} = \\theta_ {M M I M} ^ {h} + C o n v _ {m} \\left(\\hat {f} _ {m} ^ {h}\\right), \\tag {8}", + "image_path": "3f7f6a198514fd82ac9479ec1356429b981cbe8ae109f2f7e8d1b6d22eb32402.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 287, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 125, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 287, + 140 + ], + "type": "interline_equation", + "content": "\\beta^ {h} = \\beta_ {M M I M} ^ {h} + \\operatorname {C o n v} _ {m} \\left(\\hat {f} _ {m} ^ {h}\\right), \\tag {9}", + "image_path": "f5b153f8f0c7d50ad791462ccd5d6aea2ce9443b6e8d6482a9144d9ffdc7f4d3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "type": "inline_equation", + "content": "V^h" + }, + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "type": "inline_equation", + "content": "\\theta^h" + }, + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "type": "inline_equation", + "content": "\\beta^h" + }, + { + "bbox": [ + 47, + 148, + 287, + 256 + ], + "type": "text", + "content": " are the final results. Note that we can stack several MMIBs to predict the offsets iteratively. In our experiments, one single MMIB is enough to achieve satisfactory results. PyMAF [31] fuses the mesh-align features into one global feature, which ignores the inherent spatial relationship of mesh vertices. In contrast, our MMIB adopts GCN to model the spatial relation between adjacent mesh vertices. The transformer encoder is responsible for modeling the long-range dependencies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 263, + 151, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 151, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 151, + 276 + ], + "type": "text", + "content": "3.4. Model Objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "text", + "content": "Our method predicts mesh vertices positions and MANO parameters simultaneously. The learning objectives can be divided into three categories: mesh vertex loss, MANO parameter loss and mesh-mano consistency loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 331, + 287, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 287, + 356 + ], + "type": "text", + "content": "Mesh Vertex Loss: The widely-used L1 loss is adopted to supervise the vertex positions and the 2D projections:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 363, + 287, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 363, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 52, + 363, + 287, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V} = \\sum_ {h = L, R} \\| V ^ {h} - V _ {G T} ^ {h} \\| _ {1} + \\| \\Pi (V _ {h}) - \\Pi (V _ {G T} ^ {h}) \\| _ {1}, \\tag {10}", + "image_path": "414222dbb8899ef21d9dd7b08fe993d094ab326f2b7513a266f6dd1c9764d476.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "type": "inline_equation", + "content": "V_{GT}^{h}" + }, + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "type": "text", + "content": " represents the ground-truth vertex positions and " + }, + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "type": "text", + "content": " denotes the projection function. Given vertex positions, the joint positions can be regressed by multiplying the predefined regression matrix " + }, + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 47, + 397, + 287, + 446 + ], + "type": "text", + "content": ". Joint losses are formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 453, + 287, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 453, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 85, + 453, + 287, + 508 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {J} = \\sum_ {h = L, R} \\| \\mathcal {J} V ^ {h} - \\mathcal {J} V _ {G T} ^ {h} \\| _ {1} \\\\ + \\sum_ {h = L, R} \\| \\Pi \\mathcal {J} \\left(V _ {h}\\right) - \\Pi \\left(\\mathcal {J} V _ {G T} ^ {h}\\right) \\| _ {1}. \\tag {11} \\\\ \\end{array}", + "image_path": "754d5d1ba2d337ff4dd906b57c7170710077bcc680feb05bcab12cd18d6bcb05.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 515, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 287, + 539 + ], + "type": "text", + "content": "The face normal loss is introduced to regularize the surface normal consistency:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 547, + 287, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 547, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 89, + 547, + 287, + 581 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N} = \\sum_ {h = L, R} \\sum_ {f = 1} ^ {F} \\sum_ {i = 1} ^ {3} \\| e _ {f, i} ^ {h} \\cdot n _ {f, G T} ^ {h} \\| _ {1}, \\tag {12}", + "image_path": "83d4ce50bb7a929e2f107a6165b0bc9c18129a0f3387a57f18f05097601ba2f6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "inline_equation", + "content": "e_{f,i}^{h}" + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "text", + "content": "th edge of face " + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "text", + "content": " at hand " + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "inline_equation", + "content": "n_{f,GT}^{h}" + }, + { + "bbox": [ + 47, + 589, + 287, + 639 + ], + "type": "text", + "content": " is the normal vector of this face from the ground truth mesh. The edge length consistency loss is to enforce the edge length consistency:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 99, + 647, + 287, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 99, + 647, + 287, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E} = \\sum_ {h = L, R} \\sum_ {i = 1} ^ {E} \\| e _ {i} ^ {h} - e _ {i, G T} ^ {h} \\| _ {1}, \\tag {13}", + "image_path": "24f6dfef1c5ec3a19ad6b008b019be8b4a967e63e9bd82cfafa5dad720476a62.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "e_i^h" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "th edge of hand " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " denotes the total edge number, respectively." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 330, + 70, + 525, + 119 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 525, + 119 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 525, + 119 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 525, + 119 + ], + "type": "table", + "html": "
MethodParams(M)Inference Time (ms)
InterShape [30]139.647.77
Intaghand [16]37.2851.86
Ours38.3160.52
", + "image_path": "9166ca1547682e815e9a053e83522cbb067607afd04abe2ccc1e5c9cf8e20523.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 128, + 527, + 139 + ], + "lines": [ + { + "bbox": [ + 324, + 128, + 527, + 139 + ], + "spans": [ + { + "bbox": [ + 324, + 128, + 527, + 139 + ], + "type": "text", + "content": "Table 1. Parameters and inference speed on Tesla V100" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 160, + 545, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 160, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 160, + 545, + 185 + ], + "type": "text", + "content": "MANO Parameter Loss: Given the ground truth MANO parameters, L1 loss is used to regress parameters," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 194, + 545, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 194, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 326, + 194, + 545, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P} = \\sum_ {h = L, R} \\| \\theta^ {h} - \\theta_ {G T} ^ {h} \\| _ {1} + \\| \\beta^ {h} - \\beta_ {G T} ^ {h} \\| _ {1}. \\tag {14}", + "image_path": "9a56e8b5aaaf6b9b8f9255cc71d8e69c15d6c31a86e10de0379b0d0c7c60e49b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 229, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 229, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 305, + 229, + 545, + 265 + ], + "type": "text", + "content": "In addition, we reconstruct hand mesh based on the MANO model. The reconstructed mesh should be close to the ground-truth mesh," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 274, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 274, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 318, + 274, + 545, + 300 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V, \\text {m a n o}} = \\sum_ {h = L, R} \\| \\mathcal {M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - V _ {G T} ^ {h} \\| _ {1}. \\tag {15}", + "image_path": "b7c6c4d835ad8f3774f19619a0c136b4d2624a20ecb8424a0a107b43309809e6.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 312, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 347 + ], + "type": "text", + "content": "Mesh-Mano Consistency Loss. The predicted mesh and the reconstructed mesh from MANO parameters should be consistent with each other:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 323, + 355, + 545, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 355, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 323, + 355, + 545, + 399 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s i s t}} = \\sum_ {h = L, R} \\| \\mathcal {M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - V ^ {h} \\| _ {1} \\tag {16} \\\\ + \\left\\| \\mathcal {J M A N O} \\left(\\theta^ {h}, \\beta^ {h}\\right) - \\mathcal {J} V ^ {h} \\right\\| _ {1}. \\\\ \\end{array}", + "image_path": "e31a6be49e552c3c00e7bb6bf37563c68d30c9b672f5d42c678e0771deab36c5.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 407, + 487, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 487, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 487, + 419 + ], + "type": "text", + "content": "In summary, the overall training loss " + }, + { + "bbox": [ + 317, + 407, + 487, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 317, + 407, + 487, + 419 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 429, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {V} \\mathcal {L} _ {V} + \\lambda_ {J} \\mathcal {L} _ {J} + \\lambda_ {N} \\mathcal {L} _ {N} + \\lambda_ {E} \\mathcal {L} _ {E} \\tag {17} \\\\ + \\lambda_ {P} \\mathcal {L} _ {P} + \\lambda_ {V, m a n o} \\mathcal {L} _ {V, m a n o} + \\lambda_ {\\text {c o n s i s t}} \\mathcal {L} _ {\\text {c o n s i s t}}. \\\\ \\end{array}", + "image_path": "21aa76191cf7d078752a1af23bfac1b8d93920f6122ec277d191e6c1325aa2e3.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{V} = 40" + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{J} = 40" + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{N} = 5" + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{E} = 40" + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{P} = 10" + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{V,mano} = 10" + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{consist} = 40" + }, + { + "bbox": [ + 305, + 466, + 545, + 525 + ], + "type": "text", + "content": ". The whole network including the pre-trained image feature encoder-decoder is jointly optimized such that the image features better adapt to our hand mesh reconstruction task." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 537, + 382, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 382, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 382, + 551 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 557, + 466, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 466, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 466, + 570 + ], + "type": "text", + "content": "4.1. Datasets and Implementation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 578, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 545, + 651 + ], + "type": "text", + "content": "Training Dataset. Interhand2.6M [22] is a large-scale hand dataset with ground truth mesh annotations including both single-hand and interacting two-hand images. We pick out the interacting hand (IH) data for training and testing. For a fair comparison, we follow the preprocessing steps of [16] which produces 366K training and 261K testing samples." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. The image encoder uses ResNet50 [13] as the backbone. The decoder contains four simple deconvolutional layers. We first pretrain our image encoder-decoder module on auxiliary tasks. Then, the whole network is jointly optimized with Adam optimizer at" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "568" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 296, + 181 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 296, + 181 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 296, + 181 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 296, + 181 + ], + "type": "image", + "image_path": "d01db973666f0219532d9ae561a584bfbafbb2680e4736f23f58e7541f44b90f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with SOTA parametric method InterShape [30] and SOTA non-parametric method IntagHand [16] on InterHand2.6M dataset. Our method performs better on close-interacting two-hand reconstruction (first row). Besides, our method is robust in hard cases such as severe occlusions and challenging viewpoints (second row)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 541, + 182 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 541, + 182 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 541, + 182 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 541, + 182 + ], + "type": "image", + "image_path": "eb5876f9ba615edb94708ccff8196db4b01e0a695811ed507b616b3b88e224d6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 244, + 103, + 343 + ], + "blocks": [ + { + "bbox": [ + 53, + 244, + 103, + 343 + ], + "lines": [ + { + "bbox": [ + 53, + 244, + 103, + 343 + ], + "spans": [ + { + "bbox": [ + 53, + 244, + 103, + 343 + ], + "type": "image", + "image_path": "fecd2dd1be525dc5ad17d7a9f80acf9e6828cd819a68b5970ac4f2e3f4d94c07.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 344, + 97, + 352 + ], + "lines": [ + { + "bbox": [ + 57, + 344, + 97, + 352 + ], + "spans": [ + { + "bbox": [ + 57, + 344, + 97, + 352 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 106, + 244, + 206, + 343 + ], + "blocks": [ + { + "bbox": [ + 106, + 244, + 206, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 206, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 206, + 343 + ], + "type": "image", + "image_path": "25a5505b0a9f7a1d25d38c2a80c88bcf5c7281c3aae6a9f9bd7d1089c38be9e1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 344, + 136, + 351 + ], + "lines": [ + { + "bbox": [ + 125, + 344, + 136, + 351 + ], + "spans": [ + { + "bbox": [ + 125, + 344, + 136, + 351 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 344, + 187, + 351 + ], + "lines": [ + { + "bbox": [ + 170, + 344, + 187, + 351 + ], + "spans": [ + { + "bbox": [ + 170, + 344, + 187, + 351 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 122, + 361, + 470, + 373 + ], + "lines": [ + { + "bbox": [ + 122, + 361, + 470, + 373 + ], + "spans": [ + { + "bbox": [ + 122, + 361, + 470, + 373 + ], + "type": "text", + "content": "Figure 5. Our method produces accurate hand reconstructions for various two-hand interactions." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 209, + 251, + 298, + 340 + ], + "blocks": [ + { + "bbox": [ + 209, + 251, + 298, + 340 + ], + "lines": [ + { + "bbox": [ + 209, + 251, + 298, + 340 + ], + "spans": [ + { + "bbox": [ + 209, + 251, + 298, + 340 + ], + "type": "image", + "image_path": "c8db48590528402b273c7a72ae3840a497a8547c6eb040a7375277de90deb696.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 344, + 279, + 351 + ], + "lines": [ + { + "bbox": [ + 222, + 344, + 279, + 351 + ], + "spans": [ + { + "bbox": [ + 222, + 344, + 279, + 351 + ], + "type": "text", + "content": "Alternative Views" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 298, + 244, + 350, + 343 + ], + "blocks": [ + { + "bbox": [ + 298, + 244, + 350, + 343 + ], + "lines": [ + { + "bbox": [ + 298, + 244, + 350, + 343 + ], + "spans": [ + { + "bbox": [ + 298, + 244, + 350, + 343 + ], + "type": "image", + "image_path": "de4242ab860c49e4f0baa8f44193dc83ec2cea1fc17c298b6e90fe3e3d09c921.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 344, + 343, + 352 + ], + "lines": [ + { + "bbox": [ + 304, + 344, + 343, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 343, + 352 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 351, + 244, + 452, + 343 + ], + "blocks": [ + { + "bbox": [ + 351, + 244, + 452, + 343 + ], + "lines": [ + { + "bbox": [ + 351, + 244, + 452, + 343 + ], + "spans": [ + { + "bbox": [ + 351, + 244, + 452, + 343 + ], + "type": "image", + "image_path": "e0531ff7aa698fb3d91f278aa1e0ac520501ff30441245a56318502f846f3f95.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 344, + 383, + 351 + ], + "lines": [ + { + "bbox": [ + 371, + 344, + 383, + 351 + ], + "spans": [ + { + "bbox": [ + 371, + 344, + 383, + 351 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 344, + 433, + 351 + ], + "lines": [ + { + "bbox": [ + 416, + 344, + 433, + 351 + ], + "spans": [ + { + "bbox": [ + 416, + 344, + 433, + 351 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 454, + 252, + 541, + 340 + ], + "blocks": [ + { + "bbox": [ + 454, + 252, + 541, + 340 + ], + "lines": [ + { + "bbox": [ + 454, + 252, + 541, + 340 + ], + "spans": [ + { + "bbox": [ + 454, + 252, + 541, + 340 + ], + "type": "image", + "image_path": "2b9034f4a6fd92e1c166ed031211c71447b235b395822787787c388a663a0daa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 344, + 528, + 351 + ], + "lines": [ + { + "bbox": [ + 471, + 344, + 528, + 351 + ], + "spans": [ + { + "bbox": [ + 471, + 344, + 528, + 351 + ], + "type": "text", + "content": "Alternative Views" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 56, + 392, + 281, + 460 + ], + "blocks": [ + { + "bbox": [ + 56, + 392, + 281, + 460 + ], + "lines": [ + { + "bbox": [ + 56, + 392, + 281, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 281, + 460 + ], + "type": "table", + "html": "
MethodMPJPE↓MPVPE↓AUC ↑PROJ2D↓
zimmermann et al. [34]36.36---
Spurr et al. [25]15.40---
InterNet [22]16.00-0.711-
Intaghand [16]8.799.030.8066.47
Ours8.658.890.8326.22
", + "image_path": "fa1935afd96fdd24072ca4b800825b0971a4009bac8d9c3433436dd74efac0d1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with state-of-art non-parametric methods on InterHand2.6M dataset." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 525, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 572 + ], + "type": "text", + "content": "a learning rate " + }, + { + "bbox": [ + 46, + 525, + 287, + 572 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 525, + 287, + 572 + ], + "type": "text", + "content": ". Data augmentation includes random rotation, random translation, and random scaling. The batch size is set to 32. It takes around 48 hours to train MeMaHand with 4 Tesla V100 GPUs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Metrics. Mean Per Joint Position Error (MPJPE) and Mean Per Vertex Position Error (MPVPE) in millimeters are adopted to evaluate the mesh reconstruction accuracy. Additionally, we report the percentage of correct keypoints (PCK) curve and Area Under the Curve (AUC) across the thresholds between 0 and 50 millimeters. To evaluate the mesh-image alignment accuracy, the reconstructed mesh vertices are projected onto the 2D image plane. The mesh-image alignment accuracy is measured by PROJ2D, which calculates the distance in image pixels between the projected ground truth vertices and the predicted vertices." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 393, + 529, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 529, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 529, + 407 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "content": "We compare our model with state-of-the-art non-parametric methods including InterNet [22] and Intaghand [16] and parametric methods including [3], MinimalHand [33] and InterShape [30]. The officially released weights are used to obtain the results. The model parameters and inference time of our method and other SOTA methods are reported in Tab. 1. The inference of our model can be completed in " + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "inline_equation", + "content": "60.52ms" + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "content": ", comparable to other SOTAs but with better mesh reconstruction results. Besides, the mesh vertex positions as well as the MANO parameters are predicted simultaneously in one single forward pass." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 554, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 650 + ], + "type": "text", + "content": "Comparison with Non-parametric Methods. Non-parametric methods directly generate mesh vertex positions, which can express the local details of variable 3D hand shapes. The quantitative results of non-parametric methods are shown in Tab. 2. It can be seen that our method achieves the best performance on all evaluation metrics. Fig. 6(a) presents the PCK curve, which further demonstrates the superiority of our method at all threshold levels." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Fig. 4 presents the qualitative comparison with SOTA non-parametric method IntagHand [16]. Our method performs better on close interacting two-hand reconstruction (first row). Besides, we observe that IntagHand produces collapsed meshes in hard cases such as severe occlusions" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "569" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 88, + 270, + 235 + ], + "blocks": [ + { + "bbox": [ + 72, + 88, + 270, + 235 + ], + "lines": [ + { + "bbox": [ + 72, + 88, + 270, + 235 + ], + "spans": [ + { + "bbox": [ + 72, + 88, + 270, + 235 + ], + "type": "image", + "image_path": "8832cb057416e9ea798c88119611ef0f99c5d11bb8530769c27b12b7366407ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 237, + 251, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 251, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 251, + 247 + ], + "type": "text", + "content": "(a) Comparison with non-parametric method" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 88, + 506, + 235 + ], + "blocks": [ + { + "bbox": [ + 309, + 88, + 506, + 235 + ], + "lines": [ + { + "bbox": [ + 309, + 88, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 309, + 88, + 506, + 235 + ], + "type": "image", + "image_path": "3e065247fc52936c06f16d9272c9bd3a5ef806c3e8f2544598fbb6f7ddedf6d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 237, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 350, + 237, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 350, + 237, + 481, + 247 + ], + "type": "text", + "content": "(b) Comparison with parametric method" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 53, + 297, + 284, + 355 + ], + "blocks": [ + { + "bbox": [ + 53, + 297, + 284, + 355 + ], + "lines": [ + { + "bbox": [ + 53, + 297, + 284, + 355 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 284, + 355 + ], + "type": "table", + "html": "
MethodMPJPE↓MPVPE ↓AUC↑PROJ2D↓
Boukhayma et al. [3]16.9317.980.672-
MinimalHand [33]23.4823.890.543-
InterShape [30]13.4813.950.7359.95
Ours10.8510.920.7906.90
", + "image_path": "f274a1405a6d4d23cd0f90b71e9f8119268f93dfa2f73c9d44a11a1c3595c102.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 362, + 287, + 384 + ], + "lines": [ + { + "bbox": [ + 47, + 362, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 287, + 384 + ], + "type": "text", + "content": "Table 3. Quantitative comparison with state-of-art parametric methods on InterHand2.6M dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 409, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 468 + ], + "type": "text", + "content": "and challenging viewpoints (second row). In contrast, our method is more robust to such situations. We attribute this success to the integration of MANO representations, which predicts stable hand meshes. More results with alternative views are shown in Figure 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 473, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 604 + ], + "type": "text", + "content": "Comparison with Parametric Methods. For a fair comparison, the reconstructed hand meshes from our predicted MANO parameters are utilized for evaluation. The quantitative comparisons are listed in Tab. 3. The PCK curve is shown in Fig. 6(a). The MANO parameter estimation of our method outperforms other SOTA parametric methods. Thanks to the mesh-mano interaction block, the MANO token is conditioned on the mesh vertex tokens where spatial relations are retained. Such dependencies are more informative than one-dimensional global feature utilized in competing methods [3, 30, 33]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "From Tab. 2 and Tab. 3, we can find the parametric method InterShape [30] has higher statistical errors on joint and vertex positions compared with the non-parametric method IntagHand [16] due to the limited capacity of the parametric model. However, from the second row of Fig. 4, we can see InterShape produces reasonable hand shapes in these challenging cases. By combining the merits of parametric and non-parametric hand representations, our method achieves the best performance both quantitatively" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 298, + 547, + 418 + ], + "blocks": [ + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "type": "text", + "content": "Figure 6. Comparison with the state-of-the-art methods on InterHand2.6M dataset. Left: 3D PCK of non-parametric methods. Right: 3D PCK of parametric methods. The AUC values are shown in parentheses." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 298, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 307, + 298, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 547, + 418 + ], + "type": "image", + "image_path": "b5de1be686022aa4fc1520cfc845f8fd2a2113d7fcae4aebacf137b61acb6221.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 426, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 449 + ], + "type": "text", + "content": "Figure 7. Two-hand reconstruction of images in the wild, which are taken from the RGB2Hands [28]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 470, + 378, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 378, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 378, + 482 + ], + "type": "text", + "content": "and qualitatively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 489, + 479, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 479, + 502 + ], + "type": "text", + "content": "4.3. Extension to Images in the Wild" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": "We further present the hand reconstruction results on images in the wild. As shown in Fig. 7, our method performs well on real-life images taken from the RGB2Hands dataset [28], which demonstrates the generalization ability of our approach. Recall that our method is designed to predict accurate mesh vertices positions and MANO parameters simultaneously in one forward pass. The MANO pose parameter represents the joint rotations in axis-angle representation, which is useful for animating 3D hands in computer graphics. Accurate hand mesh vertices positions and MANO parameters for hands in the wild can facilitate different human-computer-interaction (HCI) applications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 659, + 400, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 400, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 400, + 672 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "To evaluate the effectiveness of the proposed modules in our framework, we conduct an ablation study on several variants of our method. The quantitative results of all vari-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "570" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 291, + 170 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 291, + 170 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 291, + 170 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 291, + 170 + ], + "type": "table", + "html": "
MPJPE↓MPVPE↓AUC↑PROJ2D↓
A: w/o MANO token8.879.090.8186.85
B: w/o Mesh-Align8.799.010.8266.64
C: Mesh-Align scale-168.738.970.8296.38
D: Mesh-Align scale-648.778.990.8266.58
E: Regress MANO from Fgap8.839.040.8286.86
F: w/o asymmetric attention9.179.370.8226.67
G: w/o auxiliary tasks9.119.280.8236.78
Full Model8.658.890.8326.22
", + "image_path": "f66d90fb1b4ee50f4a4a1bcbbf5d376824a2c7078433105cd793dbe1391b48fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 178, + 271, + 190 + ], + "lines": [ + { + "bbox": [ + 61, + 178, + 271, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 271, + 190 + ], + "type": "text", + "content": "Table 4. Comparisons of different variants of our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 82, + 205, + 138, + 329 + ], + "blocks": [ + { + "bbox": [ + 82, + 205, + 138, + 329 + ], + "lines": [ + { + "bbox": [ + 82, + 205, + 138, + 329 + ], + "spans": [ + { + "bbox": [ + 82, + 205, + 138, + 329 + ], + "type": "image", + "image_path": "25bdbd62e90d44a4afff8232342ae0d4d54d2b8eca08fbdc12d147e4111e0092.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 338, + 246, + 349 + ], + "lines": [ + { + "bbox": [ + 87, + 338, + 246, + 349 + ], + "spans": [ + { + "bbox": [ + 87, + 338, + 246, + 349 + ], + "type": "text", + "content": "Figure 8. Ablation study on MANO tokens." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 139, + 205, + 196, + 328 + ], + "blocks": [ + { + "bbox": [ + 139, + 205, + 196, + 328 + ], + "lines": [ + { + "bbox": [ + 139, + 205, + 196, + 328 + ], + "spans": [ + { + "bbox": [ + 139, + 205, + 196, + 328 + ], + "type": "image", + "image_path": "4d4ecd5b8b81954f6dd360626b7169be44b210982978b02cbdebf4ace3171fb5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 197, + 205, + 254, + 328 + ], + "blocks": [ + { + "bbox": [ + 197, + 205, + 254, + 328 + ], + "lines": [ + { + "bbox": [ + 197, + 205, + 254, + 328 + ], + "spans": [ + { + "bbox": [ + 197, + 205, + 254, + 328 + ], + "type": "image", + "image_path": "039af84d3432e6e7a86bfdcb531e00b66575811ccd126e0eae9469f666e915bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 362, + 97, + 469 + ], + "blocks": [ + { + "bbox": [ + 48, + 362, + 97, + 469 + ], + "lines": [ + { + "bbox": [ + 48, + 362, + 97, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 362, + 97, + 469 + ], + "type": "image", + "image_path": "29d567eecff4aedd51d7cf561fe6b8e79395b1dd7fc147a7d90b0f9664286e13.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 478, + 269, + 489 + ], + "lines": [ + { + "bbox": [ + 64, + 478, + 269, + 489 + ], + "spans": [ + { + "bbox": [ + 64, + 478, + 269, + 489 + ], + "type": "text", + "content": "Figure 9. Ablation study on mesh alignment refinement." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 97, + 363, + 194, + 469 + ], + "blocks": [ + { + "bbox": [ + 97, + 363, + 194, + 469 + ], + "lines": [ + { + "bbox": [ + 97, + 363, + 194, + 469 + ], + "spans": [ + { + "bbox": [ + 97, + 363, + 194, + 469 + ], + "type": "image", + "image_path": "2f0258ac679720dde1cbc7824c524cddb8f4294acfcfae84f9361a6c0c5a57e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 194, + 363, + 287, + 469 + ], + "blocks": [ + { + "bbox": [ + 194, + 363, + 287, + 469 + ], + "lines": [ + { + "bbox": [ + 194, + 363, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 194, + 363, + 287, + 469 + ], + "type": "image", + "image_path": "785f616a298a9505e8160e4a93ac19d1ff03dcfaf0ce6ab12d0b7c570cc0fac2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 514, + 166, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 166, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 166, + 525 + ], + "type": "text", + "content": "ants are presented in Table 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 529, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 625 + ], + "type": "text", + "content": "Effectiveness of MANO token. Variant A (w/o MANO token) represents removing the MANO tokens. The transformer encoder only models the dependencies between the mesh vertex tokens and image features. Fig. 8 show the qualitative comparison. Without hand prior information, the reconstructed mesh may be corrupted when parts of the hands are occluded. In contrast, our full model can generate reasonable meshes in these challenging situations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Effectiveness of Mesh Alignment Refinement. Variant B (w/o Mesh-Align) represents removing the mesh alignment refinement module. The predictions generated by the Mesh-Mano interaction module are taken as the final outputs. Variant C and D denote extracting the single-scale image features at resolution 16 and 64, respectively. The performances degrade without using mesh align refinement" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 340, + 71, + 514, + 196 + ], + "blocks": [ + { + "bbox": [ + 340, + 71, + 514, + 196 + ], + "lines": [ + { + "bbox": [ + 340, + 71, + 514, + 196 + ], + "spans": [ + { + "bbox": [ + 340, + 71, + 514, + 196 + ], + "type": "image", + "image_path": "f9c1cc86008180787f43161818858f9d25ad7400d0928f6a3124b2d550b8f9c3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 205, + 523, + 216 + ], + "lines": [ + { + "bbox": [ + 326, + 205, + 523, + 216 + ], + "spans": [ + { + "bbox": [ + 326, + 205, + 523, + 216 + ], + "type": "text", + "content": "Figure 10. Failures cases caused by severe occlusions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 239, + 544, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 544, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 544, + 262 + ], + "type": "text", + "content": "or using single-scale image features. As shown in Fig. 9, our full model produces better-aligned results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "text", + "content": "Effectiveness of MMIB. Parametric and non-parametric representations are widely used for many years. However, combining them is not trivial. In variant " + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "text", + "content": " (Regress MANO from " + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "F_{gap}" + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "text", + "content": "), MMIB is responsible for updating vertex tokens. The MANO parameters are regressed from the global average pooling features " + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "F_{gap}" + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "text", + "content": " with simple fully connected layers rather than through MMIB. The performance gain of variant " + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "text", + "content": " is limited as shown in Tab. 4. On the other hand, we propose an asymmetric attention mask to exclude the mano-to-mesh attention. To verify the choice, variant " + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 305, + 266, + 545, + 422 + ], + "type": "text", + "content": " (w/o asymmetric attention mask) denotes removing the asymmetric attention mechanism. The performance degrades compared with our full model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 425, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 508 + ], + "type": "text", + "content": "Effectiveness of Pretraining on Auxiliary Tasks. Variant G represents the backbone is pre-trained on ImageNet. Our full model performs better since pre-training on auxiliary tasks are essential to extract semantically-meaningful image features. The auxiliary 2D image conditions are generated from the ground-truth mesh of Interhand2.6M training split. We did not use extra datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 521, + 378, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 378, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 378, + 533 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "type": "text", + "content": "In this paper, we propose a novel approach MeMaHand for two-hand mesh reconstruction from a single image. The mesh-mano interaction module combines the merits of non-parametric and parametric representations. Then, the mesh alignment refinement module further rectifies the results with an explicit mesh alignment feedback loop. Extensive experiments on the InterHand2.6M benchmark demonstrate that our proposed MeMaHand is superior to both existing parametric and non-parametric methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Limitations. Although our method generates promising results, it still fails in cases of severe occlusions. Fig. 10 shows some failure cases, where inter-penetration occurs. Taking the physical plausibility into consideration will be our future work." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 311, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 311, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 311, + 757 + ], + "type": "text", + "content": "571" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1067-1076, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interacting objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6121-6131, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "type": "text", + "content": "[3] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10843-10852, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "type": "text", + "content": "[4] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 666-682, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 292, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 287, + 357 + ], + "type": "text", + "content": "[5] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12929–12938, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 424 + ], + "type": "text", + "content": "[6] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274–13283, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 425, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 490 + ], + "type": "text", + "content": "[7] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10451-10460, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 544 + ], + "type": "text", + "content": "[8] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "type": "text", + "content": "[9] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "text", + "content": "[10] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10833-10842, 2019. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[11] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 571-580, 2020. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[12] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "type": "text", + "content": "[14] Dong Uk Kim, Kwang In Kim, and Seunglyul Baek. End-to-end detection and pose estimation of two interacting hands. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11189-11198, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "text", + "content": "[15] Nikolaos Kyriazis and Antonis Argyros. Scalable 3d tracking of multiple interacting objects. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3430-3437, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 268, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 323 + ], + "type": "text", + "content": "[16] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "type": "text", + "content": "[17] Guan Ming Lim, Prayook Jatesiktat, and Wei Tech Ang. Mobilehand: Real-time 3d hand shape and pose estimation from color image. In International Conference on Neural Information Processing, pages 450-459. Springer, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[18] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "text", + "content": "[19] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 463, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 517 + ], + "type": "text", + "content": "[20] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "type": "text", + "content": "[21] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "text", + "content": "[22] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision, pages 548-564. Springer, 2020. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 623, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 656 + ], + "type": "text", + "content": "[23] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Trans. Graph., 36(6), nov 2017. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] Breannan Smith, Chenglei Wu, He Wen, Patrick Peluse, Yaser Sheikh, Jessica K Hodgins, and Takaaki Shiratori. Constraining dense hand surface tracking with elasticity. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "572" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 651 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[25] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[26] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11698-11707, October 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "text", + "content": "[27] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. International Journal of Computer Vision, 118(2):172-193, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 287, + 294 + ], + "type": "text", + "content": "[28] Jiayi Wang, Franziska Mueller, Florian Bernard, Suzanne Sorli, Oleksandr Sotnychenko, Neng Qian, Miguel A Otaduy, Dan Casas, and Christian Theobalt. Rgb2hands: real-time tracking of 3d hand interactions from monocular rgb video. ACM Transactions on Graphics (ToG), 39(6):1-16, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 350 + ], + "type": "text", + "content": "[29] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 352, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 352, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 287, + 417 + ], + "type": "text", + "content": "[30] Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma, and Hongan Wang. Interacting two-hand 3d pose and shape reconstruction from single color image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11354-11363, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 418, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 483 + ], + "type": "text", + "content": "[31] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 485, + 287, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 539 + ], + "type": "text", + "content": "[32] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 541, + 287, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 541, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 48, + 541, + 287, + 606 + ], + "type": "text", + "content": "[33] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 608, + 287, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 608, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 48, + 608, + 287, + 651 + ], + "type": "text", + "content": "[34] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "573" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_content_list.json b/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..42a01a779094500845f52ca6002ca727894c82e1 --- /dev/null +++ b/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_content_list.json @@ -0,0 +1,1647 @@ +[ + { + "type": "text", + "text": "Megahertz Light Steering without Moving Parts", + "text_level": 1, + "bbox": [ + 241, + 130, + 728, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adithya Pediredla $^{1,2}$ , Srinivasa G. Narasimhan $^{2}$ , Maysamreza Chamanzar $^{2}$ , Ioannis Gkioulekas $^{2}$ , Dartmouth College, Carnegie Mellon University", + "bbox": [ + 107, + 156, + 861, + 193 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 227, + 313, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a light steering technology that operates at megahertz frequencies, has no moving parts, and costs less than a hundred dollars. Our technology can benefit many projector and imaging systems that critically rely on high-speed, reliable, low-cost, and wavelength-independent light steering, including laser scanning projectors, LiDAR sensors, and fluorescence microscopes. Our technology uses ultrasound waves to generate a spatiotemporally-varying refractive index field inside a compressible medium, such as water, turning the medium into a dynamic traveling lens. By controlling the electrical input of the ultrasound transducers that generate the waves, we can change the lens, and thus steer light, at the speed of sound (1.5 km/s in water). We build a physical prototype of this technology, use it to realize different scanning techniques at megahertz rates (three orders of magnitude faster than commercial alternatives such as galvo mirror scanners), and demonstrate proof-of-concept projector and LiDAR applications. To encourage further innovation towards this new technology, we derive theory for its fundamental limits and develop a physically-accurate simulator for virtual design. Our technology offers a promising solution for achieving high-speed and low-cost light steering in a variety of applications.", + "bbox": [ + 75, + 252, + 473, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 606, + 209, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many imaging systems rely on the ability to steer light, either as it leaves a source or as it reaches a sensor. Examples include laser scanning projectors [28, 56], LiDAR depth sensors [38, 74, 75], and microscopy techniques (confocal microscopy [23, 50], light-sheet microscopy [55, 68], multiphoton microscopy [17, 82]). Compared to full-field lighting and imaging, light steering systems help improve light efficiency [47], counter indirect illumination [27, 42], and enhance illumination and imaging contrast [9, 55]. However, these advantages come at the cost of reduced acquisition speed, bulky moving hardware, and motion artifacts. To alleviate these costs, we introduce a new light steering technology that, through the use of ultrasonic sculpting, makes it possible to scan light both transversally and axially at megahertz (MHz) rates. Additionally, our technology achieves these high scanning rates without any moving parts. Lastly, prototypes of our technology cost no more than a hundred dollars. Altogether, these characteristics represent significant", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98bdc621d43009f9672422cd7f71f230268f5498cc29391576d98d6a0f44e677.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 229, + 890, + 371 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6ae1c32c570040b3786380601a27bba3c646318667c13fbe932773cae2d1cff.jpg", + "image_caption": [ + "Figure 1. Megahertz light steering. (a) Light steering systems in LiDAR systems and laser projectors have moving mechanical components, limiting them to kHz scanning rates. (b) Our technology uses the acousto-optic effect to enable MHz light steering without moving parts. The insets show that, before galvo mirrors could scan even a few points, our prototype scanned a thousand points to project the letter \"A\" on the wall." + ], + "image_footnote": [], + "bbox": [ + 501, + 371, + 890, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "advances over previous light steering technologies (Table 1).", + "bbox": [ + 498, + 619, + 893, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our technology uses the acousto-optic effect to turn a transparent medium, such as water, into a programmable optic that steers an incident light beam. Sound is a pressure wave that travels inside a medium by compressing and rarefying it, spatiotemporally changing the medium density. In turn, this changes the refractive index of the medium, which is proportional to the density [63,79]. We design the pressure profile of the sound wave so that, at any time instant, the spatially-varying refractive index makes the medium behave as a periodic set of virtual gradient-index (GRIN) lenses, each with an aperture equal to the sound wavelength. The GRIN lenses bend light beams incident on the medium, with the GRIN profile determining the beam trajectory. These lenses travel at the speed of sound (1.5 km/s in water) and are reconfigurable at MHz frequencies, allowing us to steer light faster than mechanical devices. To enable flexible steering", + "bbox": [ + 496, + 635, + 893, + 876 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This is different from Bragg's diffraction in acousto-optic deflectors.", + "bbox": [ + 517, + 886, + 883, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1", + "bbox": [ + 493, + 945, + 503, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "patterns, we combine this optic with a pulsed laser with a programmable pulse rate. By synchronizing the laser source with the sound waveform, and modulating the phase of the sound waveform, we control both the speed of beam steering and the location of the beam.", + "bbox": [ + 75, + 90, + 470, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Sections 3 and 4, we explain the physical and mathematical details of our technology. We introduce a new design that uses two linear transducers to generate traveling acoustic waves, and discuss how different synchronization choices between ultrasound and pulsed laser result in different scanning patterns. To facilitate the exploration of design parameters (ultrasound speed and frequency, laser frequency) and configurations (transducer geometry), we also develop a physics-based simulator for our technology. In Section 6, we also discuss the fundamental limits of our technology due to diffraction and the uncertainty principle in wave physics.", + "bbox": [ + 75, + 167, + 470, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Section 5, we experimentally demonstrate these fast programmable light steering techniques for various applications. In particular, in Section 5.2, we demonstrate an arbitrary point projector that can scan arbitrary and programmable light patterns. Compared to raster scanning projectors, which can project billions of points per second in a grid pattern but only a few thousand arbitrary points per second, our prototype can project a million arbitrary points per second, an acceleration by three orders of magnitude. In Section 5.3, we demonstrate a LiDAR prototype that combines our light steering technology with a single-photon avalanche diode (SPAD). We show 3D scans of $100 \\times 100$ resolution at 5000 frames per second (50 million points per second) with a single-pixel SPAD, which is not feasible with scanning galvo mirrors.", + "bbox": [ + 75, + 333, + 470, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Our main contributions are:", + "text_level": 1, + "bbox": [ + 76, + 560, + 372, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A new light steering technology based on the acoustooptic effect that is three orders of magnitude faster than state-of-the-art mechanical steering technologies.", + "2. A new hardware design with planar transducers generating of traveling waves.", + "3. An experimental prototype demonstrating ultrafast arbitrary point projection and LiDAR scanning.", + "4. A physics-based renderer to simulate digital twins of our prototype and evaluate different designs.", + "5. The derivation of limits due to fundamental restrictions from wave physics (diffraction limit, scanning speed vs. aperture tradeoff, and uncertainty principle)." + ], + "bbox": [ + 89, + 575, + 470, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide our open-source simulator, data, and additional details in the supplement and project website.2", + "bbox": [ + 75, + 756, + 468, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Limitations. Our prototype has a diffraction-limited point spread function (PSF) with a large spatial extent and a “+” shape, due to the use of two linear transducers that create a rectangular aperture. This limits spatial resolution, and introduces structured blur artifacts. These limitations are not fundamental to our core technology, and can be overcome", + "bbox": [ + 75, + 786, + 470, + 877 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/5b6152def0fd06b2d6b008faedf5694165160ebbb7a949e662d44b8fd2b09c23.jpg", + "table_caption": [ + "Table 1. Comparison of light steering technologies. MEMS is microelectromechanical systems, OPA is optical phased arrays, BW is optical bandwidth. For the arrows, red is bad, green is good, down is low, and up is high, and more arrows imply a bigger effect. Our method is superior in terms of cost, speed and supported bandwidth, with no moving parts or fabrication." + ], + "table_footnote": [], + "table_body": "
Tech.speedfab.BWcostmoving
galvo↓↓×↑↓✓
liquid lenses↓↓↓×↑↓×
MEMS↑✓↑↑✓
OPA↑↑↑✓↓↑↑×
ours↑↑×↑↓↓↓×
", + "bbox": [ + 521, + 174, + 867, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with improved designs and better engineering (Section 6).", + "bbox": [ + 498, + 292, + 880, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 319, + 637, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Light steering in imaging systems. Light steering is a core component in scanning-based active imaging systems. For example, fluorescence microscopy techniques such as confocal microscopy [23, 50], multiphoton microscopy [17, 82], light sheet microscopy [55, 68], and superresolution microscopy [7], use scanning to decrease scattered light and improve light efficiency and imaging contrast. Another example is LiDAR sensors [84] found in commercial applications, such as autonomous cars. In these sensors, scanning decreases multipath interference and helps reduce hardware cost, removing the need for two-dimensional LiDAR arrays. Our technology improves the speed and reliability of scanning-based LiDAR, while further reducing the cost.", + "bbox": [ + 496, + 340, + 893, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Light steering is also used in scanning-based laser projectors, to achieve high light efficiency and contrast by illuminating only where necessary. Laser projectors typically perform a 2D raster scan the field of view, with a fast and a slow scanning axis. The speed of the fast scanning axis limits the frame rate to the order of hundreds of kHz. In contrast, our technique can operate at tens of MHz, improving the raster scan rate by two orders of magnitude. Additionally, raster scanners cannot project arbitrary point sequences at a fast rate, and are limited to the frame rate of the projector (60-120 points per second) even when the projection pattern is very sparse. We demonstrate the projection of a million arbitrary points per second, four orders of magnitude faster.", + "bbox": [ + 496, + 537, + 893, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computational imaging techniques that use laser projectors or LiDAR are generally also scanning-based. Examples include structured light [26, 70], light-transport probing [47, 48], motion contrast 3D [44], epipolar gating [3, 46], light curtains [11, 76], slope-disparity gating [12, 37, 73], and non-line-of-sight imaging [39, 41, 49, 51, 81]. As a consequence, all these techniques can become orders of magnitude faster if combined with our light steering device.", + "bbox": [ + 496, + 734, + 893, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Light steering technologies We can distinguish between mechanical and non-mechanical light steering techniques. The former include rotating prisms [2, 78], mirrors [35, 45],", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2https://imaging.cs.cmu.edu/ultrafast_steering", + "bbox": [ + 93, + 886, + 457, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 493, + 946, + 503, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ab681afbc93c5a77a6cc1d26875f8d3c39cd192483aebc184c27ba38f91e335.jpg", + "image_caption": [ + "Figure 2. Comparison to previous work. (a) Previous ultrasonically sculpted lenses [10, 32] change the focal length at high speed (b) Our lens can change both focal length (transparencies) and the spatial location (colors) at high speed. We achieve this using traveling waves and synchronizing sound and light signals." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 468, + 170 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and digital micromechanical systems (MEMS) [61, 65]. The latter include acousto-optic (AO) [20, 83] or electro-optic (EO) [69] deflectors, liquid crystal devices [64, 77], and optical phased arrays (OPA) [29, 58]. Table 1 compares these techniques to our MHz steering technology: Mechanical techniques are slow due to the need to physically move optical elements. AO and EO deflectors operate at kHz rates. Liquid lenses and crystals typically have a long settling time, making them the slowest among these techniques.", + "bbox": [ + 75, + 280, + 472, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OPAs [29, 58] are solid-state on-chip devices that can steer light at even GHz rates. However, they require on-chip microscopic coherent laser generators [71] and cannot easily be coupled to external lasers. They are also expensive to fabricate and currently limited to low-resolution angles. Cheng et al. [15] and Spector [67] have detailed reviews.", + "bbox": [ + 75, + 426, + 472, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Acousto-optic devices. Many acousto-optic devices are commercially available, including tunable filters [21], modulators [5], frequency shifters [85], and deflectors [30]. Among these, the acousto-optic deflector can serve as a light steering device. However, acousto-optic deflectors use a different physical phenomenon, Bragg's diffraction [72], where beam deviation is proportional to the acoustic wave frequency, and are thus orders of magnitude slower than our technology. Electro-optic deflectors [40, 66] operate on a similar phenomenon and are thus similarly slow.", + "bbox": [ + 75, + 529, + 472, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tunable acoustic gradient-index (TAG) lenses [16,32] and ultrasonically-sculpted virtual optical waveguides [10, 34] use the same physical principles as our technology. TAG lenses can change the focus depth of an incident beam at $\\mathrm{kHz}$ rates; however, unlike our technology, they cannot steer it in the transverse axis (Figure 2). Following their recent commercialization, TAG lenses fostered innovation in scientific and application fields such as laser micromachining [14, 18], three-dimensional biomedical imaging [36, 80], microscopy [13, 19, 33], optical coherence tomography [22], high-throughput industrial inspection [31], and adaptive optics [62, 86]. We believe that the additional transverse steering capabilities from our technology will similarly help stimulate significant further innovation.", + "bbox": [ + 75, + 690, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/3b2567a658298eb0b54fe18492af84e3643ca8c7f38da82284254c63a602ab2e.jpg", + "table_caption": [ + "Table 2. Notation and parameters we use in the paper." + ], + "table_footnote": [], + "table_body": "
quantitysymbol
speedCUS
wavelengthλus
frequencyfus
angular frequencyωus = 2πfus
wavenumberkus = 2πλus-1
timeperiodTus = fus-1
", + "bbox": [ + 571, + 104, + 818, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Traveling-wave acousto-optic lenses", + "text_level": 1, + "bbox": [ + 500, + 220, + 821, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We generate ultrasonic waves inside a transparent medium, such as water, by submerging inside it a planar transducer that we drive with a single harmonic voltage (Figure 3(a)). The resulting pressure $P(x,t)$ inside the medium (Figure 3(b)) equals", + "bbox": [ + 496, + 242, + 893, + 316 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP (x, t) = P _ {o} + P _ {s} \\cos \\left(k _ {\\mathrm {u s}} x - \\omega_ {\\mathrm {u s}} t\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 323, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where: $t$ is time, $x$ is distance normal to the transducer plane—the pressure is independent of $y$ and $z$ coordinates— $P_{o}$ is the medium pressure without ultrasound, $P_{s}$ is proportional to the transducer voltage amplitude, and the remaining parameters are in Table 2.", + "bbox": [ + 496, + 344, + 893, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The refractive index $n(x,t)$ of the medium changes proportionally to the pressure (Figure 3(c)):", + "bbox": [ + 496, + 420, + 893, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nn (x, t) = n _ {o} + n _ {s} \\cos \\left(k _ {\\mathrm {u s}} x - \\omega_ {\\mathrm {u s}} t\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 457, + 890, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where: $n_{o}$ is the refractive index of the medium without the transducer, $n_{s} = kP_{s}$ , and $k$ is an empirical coefficient ( $k = 1.402 \\times 10^{-5}$ bar $^{-1}$ for water) [63].", + "bbox": [ + 496, + 478, + 890, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At time $t = 0$ , the convex lobes of this refractive index profile (i.e., regions $x = [l\\lambda_{\\mathrm{us}} - \\lambda_{\\mathrm{us}} / 2, l\\lambda_{\\mathrm{us}} + \\lambda_{\\mathrm{us}} / 2]$ ; $l \\in \\mathbb{Z}$ ) act as GRIN lenses. Each of these lenses focuses light rays traveling parallel to the transducer onto a line (Figure 3(d)). We can change the lens focal length and aperture by varying the amplitude and frequency of the transducer voltage.", + "bbox": [ + 496, + 523, + 893, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traveling lenses. The pressure wave and refractive index profile propagate along the $x$ -direction. Thus, the convex lobes vary as: $x(t) = [c_{\\mathrm{us}}t + l\\lambda_{\\mathrm{us}} - \\lambda_{\\mathrm{us}} / 2, c_{\\mathrm{us}}t + l\\lambda_{\\mathrm{us}} + \\lambda_{\\mathrm{us}} / 2]$ ; $l \\in \\mathbb{Z}$ . As a result, ultrasonically-sculpted cylindrical GRIN lenses are dynamic, and focused lines travel normal to the transducer at the speed of ultrasound (Figure 4).", + "bbox": [ + 496, + 614, + 893, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Scanning techniques", + "text_level": 1, + "bbox": [ + 500, + 713, + 694, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Even though the ultrasonically-sculpted GRIN lenses travel at the speed of ultrasound, we cannot control the speed or location of the lens focus. To enable such control, we use a pulsed laser with the same repetition frequency as the ultrasound frequency, and programmable phase modulation for the transducer voltage.", + "bbox": [ + 496, + 734, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single transducer. For intuition, we first describe the case of a single transducer. Due to phase modulation, the pressure pattern from Equation (1) becomes $P(x,t) = P_{o} + P_{s}\\cos (k_{\\mathrm{us}}x - \\omega_{\\mathrm{us}}t - \\phi (t))$ . If the illumination is continuous, the position of the focused light is $x_{n}(t) =$", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 493, + 945, + 503, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/40988d26d26223b134844c679c926106c1abb10a6215a4b23212ed4ac56a1262.jpg", + "image_caption": [ + "(a) setup" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 501, + 310 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51c2eb9d2618f335ca43a673e622f741703dabce171da9a6e98728fa14e01b06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 95, + 712, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3636dae10e2b138887d1e277d34c70281560b43a100ba41f937d24e68096b914.jpg", + "image_caption": [ + "(c) refractive index" + ], + "image_footnote": [], + "bbox": [ + 514, + 224, + 681, + 311 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/38c6a23d4b523f8f12b7d314651fce1898a145dbfb2eb39986b6bf5521e22f20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 102, + 885, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/94e2c41db8f77d2a41076a6aa8a84b6ab379d8cefc8a2044e1ba1c43923e12dd.jpg", + "image_caption": [ + "(d) ray diagram" + ], + "image_footnote": [], + "bbox": [ + 720, + 222, + 887, + 310 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1105b76db2e71f8f6798e2fc9445ec39b815f1d68e300f1f1a94092cf5cf495f.jpg", + "image_caption": [ + "Figure 3. Ultrasonic sculpting. (a) By placing a planar piezoelectric transducer inside a compressible medium and vibrating the transducer with a sinusoidal voltage, we generate ultrasound inside the medium. (b) Sound is a pressure wave, and at any given time, the pressure inside the medium varies spatially based on the voltage waveform applied to the transducer. (c) The change in the refractive index of the medium is proportional to the pressure. Therefore, the refractive index of the medium also varies spatially, turning the medium into a GRIN lens. (d) Light rays that pass through this GRIN lens will curve continuously and focus on a set of lines. The focal length of these waveguides/lenses is a function of the voltage and frequency applied to the transducer.", + "(a) ray diagrams", + "Figure 4. Planar transducer steers the light. As sound propagates, the spatial refractive index pattern also propagates at the speed of sound. The focussed light, which is a set of lines, also travels normal to the transducer plane at the speed of sound. Apart from the ray diagrams, the images rendered with our simulator (submitted as supplementary) show that the planar transducer moves the focus region (in this case, a line) at the speed of sound." + ], + "image_footnote": [], + "bbox": [ + 81, + 419, + 261, + 553 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d7ddfa99370e77f8ceece793d423c2e77dc9eba3b02a24ac34496def9b8030aa.jpg", + "image_caption": [ + "(b) images" + ], + "image_footnote": [], + "bbox": [ + 267, + 425, + 468, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$c_{\\mathrm{us}}t - n\\lambda_{\\mathrm{us}} + \\frac{\\phi(t)}{2\\pi}\\lambda ;n\\in \\mathbb{Z},$ which is a set of lines moving continuously at the speed of sound. By pulsing the illumination at ultrasound frequency, we get a set of lines flickering at the ultrasound frequency: $x_{n}(mT_{\\mathrm{us}}) = (m - n)\\lambda_{\\mathrm{us}} + \\frac{\\phi(mT_{\\mathrm{us}})}{2\\pi}\\lambda_{\\mathrm{us}}.$ If we modulate the phase linearly $(\\phi (t) = k\\omega_{\\mathrm{us}}t;k < 1)$ ,we have $x(mT_{\\mathrm{us}}) = mk c_{\\mathrm{us}}T_{\\mathrm{us}}$ for $m = n$ (i.e., the line due to the same laser and ultrasound pulse). Therefore, the lines travel at a reduced speed $kc_{\\mathrm{us}}$", + "bbox": [ + 75, + 672, + 472, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Point scanning with two transducers", + "text_level": 1, + "bbox": [ + 76, + 805, + 393, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To focus light to a point, we use two orthogonal planar transducers (Figure 5). We describe two extensions of the above approach for focus point control: one for scanning arbitrary point locations at the ultrasound frequency (i.e., MHz), and another for raster scanning at the laser repetition", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frequency, which is higher than the ultrasound frequency.", + "bbox": [ + 500, + 419, + 877, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Arbitrary point scanning. To scan arbitrary points $(x(mT_{\\mathrm{us}}),y(mT_{\\mathrm{us}}))$ for each laser pulse $m$ , we modulate the phases $\\phi_x(t)$ and $\\phi_y(t)$ of both transducers. The focus point location within the region $[0,\\lambda_{\\mathrm{us}}]\\times [0,\\lambda_{\\mathrm{us}}]$ is:", + "bbox": [ + 498, + 441, + 890, + 502 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx (t) = \\left(\\phi_ {x} (t) / k _ {\\mathrm {u s}} + c _ {\\mathrm {u s}} t\\right) \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 507, + 890, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny (t) = \\left(\\phi_ {y} (t) / k _ {\\mathrm {u s}} + c _ {\\mathrm {u s}} t\\right) \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 526, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 6 shows the refractive index and ray diagram for four sets of $\\phi_{x}$ and $\\phi_{y}$ values. To scan a set of points $(x(mT_{\\mathrm{us}}), y(mT_{\\mathrm{us}}))$ , we compute the phases $(\\phi_{x}(mT_{\\mathrm{us}}), \\phi_{y}(mT_{\\mathrm{us}}))$ using Equations (3)-(4), and interpolate to compute $(\\phi_{x}(t), \\phi_{y}(t))$ .", + "bbox": [ + 496, + 547, + 893, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Raster scanning In theory, we could use arbitrary point scanning to raster scan a two-dimensional grid of points. In that case, the phase modulation for raster scanning would be linear, $(\\phi_x(t),\\phi_y(t)) = (k_x\\omega_{\\mathrm{us}}t,k_y\\omega_{\\mathrm{us}}t)$ , where $k_{x}$ and $k_{y}$ are phase modulation rates. The phase modulation rate for the faster axis would be equal to the product of the number of scan points and the modulation rate of the slower axis.", + "bbox": [ + 496, + 630, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, this approach would limit raster scanning frequency to the ultrasound frequency. If the laser repetition frequency is higher, we can scan more points by running the laser at its highest frequency. Their locations will be:", + "bbox": [ + 496, + 742, + 893, + 803 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx \\left(m T _ {L}\\right) = \\left(m k _ {x}\\right) \\frac {\\lambda_ {\\mathrm {u s}}}{s} \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 808, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny \\left(m T _ {L}\\right) = \\left(m k _ {y}\\right) \\frac {\\lambda_ {\\mathrm {u s}}}{s} \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 839, + 890, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s = f_{L} / f_{\\mathrm{us}}$ is the ratio of laser $(f_{L})$ and ultrasound $(f_{\\mathrm{us}})$ frequencies, and $T_{L} = 1 / f_{L}$ is the inter-pulse time.", + "bbox": [ + 496, + 871, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 493, + 945, + 503, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32959e457d339467566da9b521888dd6181d6aa93cc15a4f1c13cc68e7314ede.jpg", + "image_caption": [ + "Figure 5. Two planar transducers for focusing light at a point. (a) We place two planar piezoelectric transducers orthogonal to each other inside a medium. We drive both transducers independently with a sinusoidal voltage. (b) The pressure wave inside the medium is a superposition of the pressure waves generated by the transducers. (c) The change in the refractive index is proportional to the net pressure. (d) Light rays from a wide beam focus on a set of points. We restrict the illumination beam size to focus light on a single point." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 277, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5360dc2c60743b082c541688f52bdec0aa513a6bb66871da834d5aa8cba93455.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 95, + 483, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3dd7296e25b0a72a7cc2ababa16f439758b45b2e5b9dd6f2d64208c6ab097911.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 95, + 704, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36c5c4cd976deb5cf9d74e7e0577039c46a24c13fc64b0c543ea225f03b874b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 104, + 890, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/699a4cdfd2aa2513ab725ab114b9627be0b13130f2f7a12123fbc751b40d1377.jpg", + "image_caption": [ + "Figure 6. Point steering. By controlling the phases of the sinusoidal voltages applied to the transducers, we control the location of the focus position. To continuously steer the focus location, we phase modulate the voltages applied to both transducers." + ], + "image_footnote": [], + "bbox": [ + 94, + 319, + 447, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 592, + 209, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We discuss an experimental prototype implementing our acousto-optic light steering technology, and combining it with a pulsed laser, single-pixel SPAD, and galvo mirrors (for comparison). We use this prototype to demonstrate projector and LiDAR applications. We compare our light steering system with commercially available galvo mirrors, to demonstrate the speed and the new capabilities our system enables. We keep the field-of-view and aperture same for both systems. This comparison is not the most favorable for galvo mirrors, as they typically have a larger field of view, but it is a fair one for evaluating the system's speed.", + "bbox": [ + 75, + 614, + 470, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Prototype", + "text_level": 1, + "bbox": [ + 76, + 789, + 189, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 7 shows our experimental prototype. We place two planar piezo transducers (P-25.40mm-25.40mm-2.10mm-880-WFB, APC International, Ltd) orthogonal to each other and at an inclination of $45^{\\circ}$ relative to an acrylic tank containing water, to minimize interference from interreflections. A signal generator (SDG6022x, Siglent Tech.) drives the", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "transducers via a power amplifier (ENI A300, Bell electronics). We colocate a pulsed laser (ALPHALAS GmbH, PICOPOWER-LD-510) and a gated SPAD (Microphoton Devices s.r.l.) using a beamsplitter, similar to previous techniques [24, 25, 57]. However, we do not place a lens in front of the SPAD, as the ultrasonically-sculpted lens focuses light from the object onto the SPAD. Instead, we place a lens in front of the laser to create a diverging ray that undergoes the same focusing by the ultrasonically-sculpted lens. We place an aperture after the ultrasonically-sculpted lens to limit the scanning area to only one ultrasonic period. A $45^{\\circ}$ mirror directs the beam to a pair of galvo mirrors (GVS-212, Thorlabs Inc.). We use the galvo mirrors only for comparisons.", + "bbox": [ + 496, + 321, + 893, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To synchronize the transducers, laser, SPAD gate, and SPAD timing circuit, we use two signal generators and synchronize their clock and trigger signals. We use one signal generator to drive the transducers. The two channels of the second signal generator run the laser and the SPAD gate. We explain the SPAD timing synchronization details in Section 5.3. We provide more details about our prototype, including design and alignment, in the supplement.", + "bbox": [ + 496, + 518, + 893, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Arbitrary point projector", + "text_level": 1, + "bbox": [ + 500, + 650, + 733, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the technique in Section 4.1 to compute the phases and synthesize the transducer voltage waveforms required to project an arbitrary target sequence of points. We drive the transducers with this waveform and the laser at the same frequency as ultrasound (1 MHz).", + "bbox": [ + 496, + 671, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To project the same sequence of points with the galvo mirrors, we drive the laser and transducers at a fixed frequency without any phase modulation, which results in a single-point focus. We steer this point with the galvo mirrors to scan the same desired sequence of points. We drive the galvo mirrors at $1\\mathrm{kHz}$ , $2\\mathrm{kHz}$ , $5\\mathrm{kHz}$ frequencies (points per second). The galvo mirrors are rated for $1\\mathrm{kHz}$ , and driving it at frequencies higher than $5\\mathrm{kHz}$ leads to higher motor current and failure of the fuse.", + "bbox": [ + 496, + 747, + 893, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We project the patterns on a white cardboard screen, and", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 493, + 945, + 503, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1b4d290c54087f5ebd83caa17f5ec71a646c59938edeb6dda6efb00e72d2acc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 90, + 459, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d4f1682f72860d93b0ca61c9e94fda7c4c556446b9db283b1ba02c8fb30cc7b.jpg", + "image_caption": [ + "(a) schematic", + "(b) experimental prototype", + "Figure 7. Hardware setup. We show the (a) schematic and (b) prototype built for showing the proof-of-concept applications and comparing them with a galvo mirror system. We diverge the laser beam, and the expanded beam is focused by the ultrasonically-sculpted refractive index. The beam passes through the galvo mirrors onto the scene. We steer the beam either with the ultrasonically-sculpted lens or the galvo mirrors, but not both, depending on the experiment. The reflected light from the object takes the same path back to the sensor. The SPAD sensor, which is colocated with the laser, does not have any optics in front of it other than the ultrasonically-sculpted lenses. This setup allows us to compare the scanning speed of our system with galvo mirrors, while keeping the aperture the same." + ], + "image_footnote": [], + "bbox": [ + 86, + 263, + 457, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "capture images with a camera (Allied Vision PRO-GT3400-09) for two exposures (1 ms and 50 ms) that we show in Figure 8. In this case, we are projecting 100 points that form the letter \"A\". At 1 ms exposure, the galvo mirrors only scan a few points, whereas our technique scans the entire shape ten times. At 100 ms or higher exposure, the galvo mirrors can project all the points without distortion.", + "bbox": [ + 75, + 628, + 470, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. LiDAR", + "text_level": 1, + "bbox": [ + 76, + 743, + 169, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a gated SPAD for depth and transient measurements. The gate helps reject the backscattered photons from various optics. After gating, our system does not suffer from pile-up [54, 59, 60]. We use a signal generator to drive the SPAD gate instead of the picosecond delayer (PSD) common in SPAD-based LiDAR systems [8, 41, 49, 53]. Our approach is inexpensive and generates programmable delays at much higher resolution ( $1\\mu s$ ) than the PSD ( $50\\mathrm{ns}$ ).", + "bbox": [ + 75, + 763, + 468, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We run both the transducer and SPAD signal genera", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tors in burst mode with the trigger running at $100\\mathrm{Hz}$ for synchronization. We use picoharp (Microphoton Devices s.r.l.) to measure the time-of-flight of the photons and to synchronize the scanning position. Several previous techniques [39,43,49,57] achieve this synchronization by sending a synchronization signal to picoharp, which reports a marker event when the synchronization signal is detected. Unfortunately, as our scanning runs at a much higher rate than galvo mirror scanning, we cannot use an external synchronization signal for picoharp due to the massive number of events that would be generated. Fortunately, the picoharp reports the pulse index for each laser pulse it detects. We use the pulse index to compute when the photon is fired, and use Equations (5)-(6) to compute the scan position.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first use the pulsed raster scanning technique in Section 4.1 to raster scan the scene. We set $f_{L} = 50\\mathrm{MHz}$ , $f_{\\mathrm{us}} = 1\\mathrm{MHz}$ , $k_{x} = 1.0001$ , and $k_{y} = 1.01$ . These settings result in a spatial resolution of $100\\times 100$ and scan rate of 5000 frames per second. Figure 9 shows results for four scenes, each with two letters at various depths. Our single channel (single sensor and laser) system scans 50 million points per second. This is three orders of magnitude faster than the 128-channel high-end Velodyne VLS-128™ LiDAR [1], which scans 5 million points per second.", + "bbox": [ + 496, + 303, + 893, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also use the arbitrary point scanning technique in Section 4.1 to selectively scan a few points in the field of view. An advantage of such an approach is the potential for adaptive scanning [6]. In Figure 10, we create a three-point scene and image it to compute their spatial locations. We then scan these three points using both galvo mirrors and our technique for $3\\mathrm{ms}$ (1 ms per point). When scanning with the galvo mirrors, a lot of time is spent moving the galvo mirrors from one scan point to another. As a result, many pulses fired by the laser are wasted during the travel time of the galvo mirrors. By contrast, ultrasonic beam steering shifts the focus within the time period of the two laser photons ( $1\\mu \\mathrm{s}$ ) and does not waste any pulses. As a result, our technique's depth estimation accuracy (computed using mean squared error averaged over a hundred experiments) is $250\\times$ higher than that of the galvo mirrors.", + "bbox": [ + 496, + 455, + 895, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 712, + 612, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduced a high-speed, low-cost (the cost of each transducer is only $26) technology to steer light. Our technology uses the physics of traveling sound waves synchronized with illumination and imaging sensors to enable multiple scanning applications. We built an experimental prototype to demonstrate proof-of-concept applications, such as the ability to perform arbitrary point light projection and raster scanning for LiDAR at MHz rates. In the rest of this section, we discuss some fundamental limits and tradeoffs inherent in our light steering technology due to wave physics, and directions for further exploration they suggest. We summa", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 493, + 946, + 504, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7e37173e651e71338d2ebe1160887581b25cfeeb92c5c082581557d0b3f98182.jpg", + "image_caption": [ + "Figure 8. Acousto-optic vs. galvo-mirror projection. We compare our beam steering technique with commercially available Thorlabs galvo mirrors (GVS-212). The \"A\" shape is made up of 100 points. With a $1\\mathrm{MHz}$ transducer, we are able to project a million points per second (pps), and hence, project ten thousand \"A\"s per second. Constrained by the laser's low beam power $(20\\mu \\mathrm{W}$ at $1\\mathrm{MHz})$ , practically we can only capture \"A\" at 1 ms exposure. The commercially available galvo mirrors, which are only rated at 1 kpps, only project a streak when driven at $1\\mathrm{kHz}$ , $2\\mathrm{kHz}$ , and $5\\mathrm{kHz}$ . At 50 ms exposure and $1\\mathrm{kHz}$ scan rate, the galvo mirrors only project half the pattern at the rated $1\\mathrm{kHz}$ , and at higher frequencies, the galvo mirrors project a corrupted pattern as we are operating them well beyond their 1 kpps rating." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 885, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36e855f27d82270e5352235e958fe28ce9fe4d71c9e986f741a4727af36020db.jpg", + "image_caption": [ + "Figure 9. Depth raster scanning. Each of the four scenes has two characters, \"CV\", \"PR\", \"20\", \"23\". \"C\" and \"P\" are at approximately $160\\mathrm{cm}$ depth, \"2\" at $170\\mathrm{cm}$ , and the remaining at $180\\mathrm{cm}$ . The top row shows the peak of the transient measured by the SPAD, and the bottom row shows the depth map in cm. We scan the scene at $100 \\times 100$ resolution using the raster scanning technique in Section 4.1 for an exposure of one second. The Thorlabs galvo mirrors are not capable of scanning these scenes in under a second." + ], + "image_footnote": [], + "bbox": [ + 78, + 515, + 888, + 806 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 493, + 945, + 504, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9de0c97472f8a3f82ca7e9b78cbd563742e6ac3437ef19e08fc99da0236d84c.jpg", + "image_caption": [ + "(a) galvo mirrors" + ], + "image_footnote": [], + "bbox": [ + 99, + 74, + 269, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f61e00bf1299a817d6553d9f7c863afdb35c09a86400f1ae554f29ffefb47359.jpg", + "image_caption": [ + "(depth error $= 51.3\\mathrm{cm}$", + "(b) our technique", + "(depth error $= 3.28\\mathrm{cm}$" + ], + "image_footnote": [], + "bbox": [ + 282, + 74, + 439, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37b73fd585935e410e6dbe17f0382c158996069afcb8021bf883ef92aaf9f5ce.jpg", + "image_caption": [ + "Figure 10. Transient and depth measurement. We scan three small objects placed at different depths using commercial galvo mirrors and our technique. The total exposure duration is $3\\mathrm{ms}$ , corresponding to the fastest speed (1kpps) at which the galvo mirrors can scan the scene. The laser repetition rate is $1\\mathrm{MHz}$ . We show three transients measured by (a) the galvo mirrors and (b) our technique when the scan location is at the first, second, and third point. The galvo mirrors spend a significant amount of time moving the focus point from one object to another, wasting pulses emitted during that time. In contrast, our technique does not waste any pulses, as it moves the focus point in just $1\\mu \\mathrm{s}$ , the time between two laser pulse emissions. This results in significant improvement in depth estimation accuracy (computed from the highest peak of the transient) compared to the galvo mirrors. The mean squared depth error of our technique is $16\\times$ smaller than galvos.", + "Figure 11. Beam shape. We show an HDR image of the measured beam. It has a cross shape, as the ultrasonic lens has a square aperture. Using more transducers can result in the beam shape to be more Gaussian-like. The center peak vs. the crosshairs intensity ratio is 20, and the center peak vs. the background ratio is 400." + ], + "image_footnote": [], + "bbox": [ + 155, + 448, + 390, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rize some of these tradeoffs in Table 3. In the supplement, we provide a physics-based renderer that can help virtually evaluate improved designs of our technology.", + "bbox": [ + 75, + 656, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scanning speed vs. aperture tradeoff. For the arbitrary point projector, the scanning speed (points per second, pps) is equal to the frequency of the ultrasonic transducer voltage. The aperture of the ultrasonically-sculpted lens is equal to the ultrasound wavelength $\\lambda_{\\mathrm{us}} = c_{\\mathrm{us}} / f_{\\mathrm{us}}$ . Therefore, the product of scanning frequency and the lens aperture is always less than or equal to the speed of sound. Increasing the scanning speed decreases the aperture of the ultrasonic lens. This decrease in the aperture size is not a problem for projector applications, but for LiDAR applications, the decrease in aperture decreases light throughput.", + "bbox": [ + 75, + 703, + 470, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diffraction limit. The numerical aperture of the ultrasonically-sculpted lens is approximately $n\\lambda_{\\mathrm{us}} / 2F$ , where", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3a248ed2b49ca89c04b6732858de24968899cfec6bddc6ca28eebfbde12952eb.jpg", + "table_caption": [ + "Table 3. Tradeoffs between various system parameters. $\\uparrow$ is better." + ], + "table_footnote": [], + "table_body": "
parametercusfusλn
scan speed↑↑↓-
aperture↑↓--
spatial resolution↑↓↓↑
", + "bbox": [ + 558, + 104, + 828, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$F$ is the focal length of the ultrasonically-sculped waveguide. Therefore, the diffraction-limited spot size is", + "bbox": [ + 500, + 190, + 893, + 222 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta x = \\frac {\\lambda}{2 N A} \\approx \\frac {\\lambda}{\\lambda_ {\\mathrm {u s}}} \\frac {F}{n} = \\frac {\\lambda f _ {\\mathrm {u s}}}{c _ {\\mathrm {u s}}} \\frac {F}{n}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 224, + 890, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Uncertainty principle. Spatial resolution is inversely proportional to the diffraction-limited spot size, and as mentioned earlier, temporal resolution is determined by the frequency of the transducer. So, we have the following uncertainty principle between spatial and temporal resolution:", + "bbox": [ + 498, + 267, + 893, + 343 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta x \\Delta T \\leq \\frac {\\lambda}{c _ {\\mathrm {u s}}} \\frac {F}{n}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 345, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To improve spatiotemporal resolution, we can decrease the focal length of the waveguide by increasing the voltage applied to the transducer. However, increasing the voltage makes the system non-linear and increases the probability of cavitation in the medium [52].", + "bbox": [ + 498, + 381, + 893, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another approach to improve the system's overall quality (increase aperture size, spatial and temporal resolutions) is to use a medium with a higher speed of sound. For example, tellurium dioxide $\\mathrm{(TeO_2)}$ glass has three times higher speed of sound and $50\\%$ higher refractive index than water. Using tellurium dioxide glass will improve light efficiency by an order of magnitude, and simultaneously improve spatial resolution by five times. This material is also part of existing acousto-optic devices [30], but we found them hard to modify. Tellurium dioxide, being a solid, would additionally be a more stable medium than liquid water.", + "bbox": [ + 496, + 464, + 893, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shape of the blur kernel. We used two transducers in our system, and each one of them creates a cylindrical lens. The net effect of these two cylindrical lenses is a square aperture, whose Fourier transform is the product of two 1D sincs. Therefore, our blur kernel has a cross-shape, as we can see in Fig. 11. We can make this blur kernel closer to a Gaussian-like blur kernel by using multiple transducers arranged around a circular path and synchronized. Based on the application, we can also use deconvolution techniques to improve the results in post-processing [4].", + "bbox": [ + 496, + 636, + 893, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank Hossein Baktash and Lloyd Lobo for help with the hardware prototype, and Andre Nascimento and Amy Lee for help with the physics-based renderer. This work was supported by NSF awards 1730147, 1900821, 1900849, 1935849, gifts from AWS Cloud Credits for Research and the Sybiel Berkman Foundation, and a Sloan Research Fellowship for Ioannis Gkioulekas.", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 493, + 945, + 503, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Technical specifications of Alpha Prime (VLS-128). https://www.mapix.com/lidar-scanner-sensors/velodyne/alpha-prime/. Accessed: 2023-03-27. 6", + "[2] Mohammad Abdo, Vlad Badilita, and Jan Korvink. Spatial scanning hyperspectral imaging combining a rotating slit with a dove prism. Optics Express, 27(15):20290-20304, 2019. 2", + "[3] Supreeth Achar, Joseph R. Bartels, William L. 'Red' Whittaker, Kiriakos N. Kutulakos, and Srinivasa G. Narasimhan. Epipolar time-of-flight imaging. ACM Trans. Graph., 36(4), jul 2017. 2", + "[4] Hossein Baktash, Yash Belhe, Matteo Giuseppe Scopelliti, Yi Hua, Aswin C Sankaranarayanan, and Maysamreza Chamanzar. Computational imaging using ultrasonically-sculpted virtual lenses. In 2022 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2022. 8", + "[5] Jared Beller and Linbo Shao. Acousto-optic modulators integrated on-chip. Light: Science & Applications, 11(1):1-2, 2022. 3", + "[6] Alexander W Bergman, David B Lindell, and Gordon Wetzstein. Deep adaptive lidar: End-to-end optimization of sampling and depth completion at low sampling rates. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2020. 6", + "[7] Charles Bond, Adriana N Santiago-Ruiz, Qing Tang, and Melike Lakadamyali. Technological advances in super-resolution microscopy to study cellular processes. Molecular Cell, 82(2):315-332, 2022. 2", + "[8] Mauro Buttafava, Jessica Zeman, Alberto Tosi, Kevin Eliceiri, and Andreas Velten. Non-line-of-sight imaging using a time-gated single photon avalanche diode. Optics express, 23(16):20997-21011, 2015. 6", + "[9] Walter A Carrington, Kevin E Fogarty, Larry Lifschitz, and Fredric S Fay. Three-dimensional imaging on confocal and wide-field microscopes. Handbook of biological confocal microscopy, pages 151-161, 1990. 1", + "[10] Maysamreza Chamanzar, Matteo Giuseppe Scopelliti, Julien Bloch, Ninh Do, Minyoung Huh, Dongjin Seo, Jillian Iafrati, Vikaaas S Sohal, Mohammad-Reza Alam, and Michel M Maharbiz. Ultrasonic sculpting of virtual optical waveguides in tissue. Nature communications, 10(1):1-10, 2019. 3", + "[11] Dorian Chan, Srinivasa G Narasimhan, and Matthew O'Toole. Holocurtains: Programming light curtains via binary holography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17886-17895, 2022. 2", + "[12] Sreenithy Chandran, Hiroyuki Kubo, Tomoki Ueda, Takuya Funatomi, Yasuhiro Mukaigawa, and Suren Jayasuriya. Slopedisparity gating: System and applications. IEEE Transactions on Computational Imaging, 8:317-332, 2022. 2", + "[13] Ting-Hsuan Chen, Jesse Ault, Howard A Stone, and Craig B Arnold. High-speed axial-scanning wide-field microscopy for volumetric particle tracking velocimetry. *Experiments in Fluids*, 58(5):1-7, 2017. 3" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ting-Hsuan Chen, Romain Fardel, and Craig B Arnold. Ultrafast z-scanning for high-efficiency laser micro-machining. Light: Science & Applications, 7(4):17181-17181, 2018. 3", + "[15] Yang Cheng, Jie Cao, and Qun Hao. Optical beam steering using liquid-based devices. Optics and Lasers in Engineering, 146:106700, 2021. 3", + "[16] Maxim N Cherkashin, Carsten Brenner, Georg Schmitz, and Martin R Hofmann. Transversally travelling ultrasound for light guiding deep into scattering media. Communications Physics, 3(1):1-11, 2020. 3", + "[17] Winifried Denk, James H Strickler, and Watt W Webb. Two-photon laser scanning fluorescence microscopy. Science, 248(4951):73-76, 1990. 1, 2", + "[18] Xiaohan Du, SeungYeon Kang, and Craig B Arnold. Optimization of ultrafast axial scanning parameters for efficient pulsed laser micro-machining. Journal of Materials Processing Technology, 288:116850, 2021. 3", + "[19] Martí Duocastella, Craig B Arnold, and Jason Puchalla. S electable light-sheet uniformity using tuned axial scanning. Microscopy research and technique, 80(2):250-259, 2017. 3", + "[20] Vladislav Gavyusev, Giuseppe Sancataldo, Pietro Ricci, Alberto Montalbano, Chiara Fornetto, Lapo Turrini, Annunziatina Laurino, Luca Pesce, Giuseppe de Vito, Natascia Tiso, et al. Dual-beam confocal light-sheet microscopy via flexible acousto-optic deflector. Journal of biomedical optics, 24(10):106504, 2019. 3", + "[21] Milton S Gottlieb. Acousto-optic tunable filters. In Design and fabrication of acousto-optic devices, pages 197-283. CRC Press, 2021. 3", + "[22] Ireneusz Grulkowski, Krzysztof Szulzycki, and Maciej Wojtkowski. Microscopic oct imaging with focus extension by ultrahigh-speed acousto-optic tunable lens and stroboscopic illumination. Optics Express, 22(26):31746-31760, 2014. 3", + "[23] Min Gu. Principles of three dimensional imaging in confocal microscopes. World Scientific, 1996. 1, 2", + "[24] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3d imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7918, 2019. 5", + "[25] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6770-6779, 2019. 5", + "[26] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G Narasimhan. Structured light 3d scanning in the presence of global illumination. In CVPR 2011, pages 713-720. IEEE, 2011. 2", + "[27] Kristofer Henderson, Xiaomeng Liu, Justin Folden, Brevin Tilmon, Suren Jayasuriya, and Sanjeev Koppal. Design and calibration of a fast flying-dot projector for dynamic light transport acquisition. IEEE Transactions on Computational Imaging, 6:529-543, 2020. 1", + "[28] Sven TS Holmström, Utku Baran, and Hakan Urey. Mems laser scanners: a review. Journal of Microelectromechanical Systems, 23(2):259-275, 2014. 1", + "[29] Ching-Pai Hsu, Boda Li, Braulio Solano-Rivas, Amar R Gohil, Pak Hung Chan, Andrew D Moore, and Valentina" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 493, + 945, + 504, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Donzella. A review and perspective on optical phased array for automotive lidar. IEEE Journal of Selected Topics in Quantum Electronics, 27(1):1-16, 2020. 3", + "[30] Dawoon Jeong, Hansol Jang, Min Uk Jung, and Chang-Seok Kim. Angular resolution variable fmcw lidar with acoustooptic deflector. In Imaging Systems and Applications, pages ITh4D-3. Optica Publishing Group, 2022. 3, 8", + "[31] S Kang, E Dotsenko, D Amrhein, C Theriault, and Craig B Arnold. Ultra-high-speed variable focus optics for novel applications in advanced imaging. In Photonic Instrumentation Engineering V, volume 10539, page 1053902. SPIE, 2018. 3", + "[32] SeungYeon Kang, Martí Duocastella, and Craig B Arnold. Variable optical elements for fast focus control. Nature Photonics, 14(9):533-542, 2020. 3", + "[33] Doycho Karagyozov, Mirna Mihovilovic Skanata, Amanda Lesar, and Marc Gershow. Recording neural activity in unrestrained animals with three-dimensional tracking two-photon microscopy. Cell reports, 25(5):1371-1383, 2018. 3", + "[34] Yasin Karimi, Matteo Giuseppe Scopelliti, Ninh Do, Mohammad-Reza Alam, and Maysamreza Chamanzar. In situ 3d reconfigurable ultrasonically sculpted optical beam paths. Optics express, 27(5):7249-7265, 2019. 3", + "[35] Byung-Sub Kim, Steve Gibson, and Tsu-Chin Tsao. Adaptive control of a tilt mirror for laser beam steering. In Proceedings of the 2004 American Control Conference, volume 4, pages 3417–3421. IEEE, 2004. 2", + "[36] Lingjie Kong, Jianyong Tang, Justin P Little, Yang Yu, Tim Lammermann, Charles P Lin, Ronald N Germain, and Meng Cui. Continuous volumetric imaging via an optical phase-locked ultrasound lens. Nature methods, 12(8):759-762, 2015. 3", + "[37] Hiroyuki Kubo, Suren Jayasuriya, Takafumi Iwaguchi, Takuya Funatomi, Yasuhiro Mukaigawa, and Srinivasa G Narasimhan. Programmable non-epipolar indirect light transport: Capture and analysis. IEEE Transactions on Visualization and Computer Graphics, 27(4):2421-2436, 2019. 2", + "[38] Nanxi Li, Chong Pei Ho, Jin Xue, Leh Woon Lim, Guanyu Chen, Yuan Hsing Fu, and Lennon Yao Ting Lee. A progress review on solid-state lidar and nanophotonics-based lidar sensors. *Laser & Photonics Reviews*, page 2100511, 2022. 1", + "[39] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fkmigration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2, 6", + "[40] Qiyu Liu, Huan Li, and Mo Li. Electromechanical Brillouin scattering in integrated optomechanical waveguides. Optica, 6(6):778-785, 2019. 3", + "[41] Xiaochun Liu, Ibón Guillén, Marco La Manna, Ji Hyun Nam, Syed Azer Reza, Toan Huu Le, Adrian Jarabo, Diego Gutierrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572(7771):620-623, 2019. 2, 6", + "[42] Xiaomeng Liu, Kristofer Henderson, Joshua Rego, Suren Jayasuriya, and Sanjeev Koppal. Dense lissajous sampling and interpolation for dynamic light-transport. Optics Express, 29(12):18362–18381, 2021. 1" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Julio Marco, Adrian Jarabo, Ji Hyun Nam, Xiaochun Liu, Miguel Ángel Cosculluela, Andreas Velten, and Diego Gutierrez. Virtual light transport matrices for non-line-of-sight imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2440-2449, 2021. 6", + "[44] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2", + "[45] Aongus McCarthy, Nils J Krichel, Nathan R Gemmell, Ximing Ren, Michael G Tanner, Sander N Dorenbos, Val Zwiller, Robert H Hadfield, and Gerald S Buller. Kilometer-range, high resolution depth imaging via 1560 nm wavelength single-photon detection. Optics express, 21(7):8904-8915, 2013. 2", + "[46] Srinivasa Narasimhan, Supreeth Achar, Kiriakos Kutulakos, Joe Bartels, and William Whittaker. Method for epipolar time of flight imaging, Mar. 19 2020. US Patent App. 16/468,617. 2", + "[47] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 1, 2", + "[48] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual coding to probe light transport. ACM Trans. Graph., 31(4):39-1, 2012. 2", + "[49] Matthew O'Toole, David B Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555(7696):338-341, 2018. 2, 6", + "[50] James Pawley. Handbook of biological confocal microscopy, volume 236. Springer Science & Business Media, 2006. 1, 2", + "[51] Adithya Pediredla, Akshit Dave, and Ashok Veeraraghavan. Snlos: Non-line-of-sight scanning through temporal focusing. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2019. 2", + "[52] Adithya Pediredla, Matteo Giuseppe Scopelliti, Srinivasa Narasimhan, Maysamreza Chamanzar, and Ioannis Gkioulekas. Optimized virtual optical waveguides enhance light throughput in scattering media. preprint, 2021. 8", + "[53] Adithya Kumar Pediredla, Mauro Buttafava, Alberto Tosi, Oliver Cossairt, and Ashok Veeraraghavan. Reconstructing rooms using photon echoes: A plane based model and reconstruction algorithm for looking around the corner. In 2017 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2017. 6", + "[54] Adithya K Pediredla, Aswin C Sankaranarayanan, Mauro Buttafava, Alberto Tosi, and Ashok Veeraraghavan. Signal processing based pile-up compensation for gated single-photon avalanche diodes. arXiv preprint arXiv:1806.07437, 2018. 6", + "[55] Adithya Kumar Pediredla, Shizheng Zhang, Ben Avants, Fan Ye, Shin Nagayama, Ziying Chen, Caleb Kemere, Jacob T Robinson, and Ashok Veeraraghavan. Deep imaging in scattering media with selective plane illumination microscopy. Journal of biomedical optics, 21(12):126009, 2016. 1, 2", + "[56] Kurt E Petersen. Silicon torsional scanning mirror. IBM Journal of Research and Development, 24(5):631-637, 1980. 1" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 491, + 945, + 506, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Ryan Po, Adithya Pediredla, and Ioannis Gkioulekas. Adaptive gating for single-photon 3d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16354–16363, 2022. 5, 6", + "[58] Christopher V Poulton, Matthew J Byrd, Benjamin Moss, Erman Timurdogan, Ron Millman, and Michael R Watts. 8192-element optical phased array with $100^{\\circ}$ steering range and flip-chip cmos. In CLEO: QELS_Fundamental Science, pages JTh4A-3. Optical Society of America, 2020. 3", + "[59] Joshua Rapp, Yanting Ma, Robin MA Dawson, and Vivek K Goyal. Dead time compensation for high-flux ranging. IEEE Transactions on Signal Processing, 67(13):3471-3486, 2019. 6", + "[60] Joshua Rapp, Yanting Ma, Robin MA Dawson, and Vivek K Goyal. High-flux single-photon lidar. Optica, 8(1):30-39, 2021. 6", + "[61] Joshua Rodriguez, Braden Smith, Brandon Hellman, Adley Gin, Alonzo Espinoza, and Yuzuru Takashima. Multi-beam and single-chip lidar with discrete beam steering by digital micromirror device. In Physics and Simulation of Optoelectronic Devices XXVI, volume 10526, pages 89-94. SPIE, 2018. 3", + "[62] Patrick S Salter, Z Iqbal, and Martin J Booth. Analysis of the three-dimensional focal positioning capability of adaptive optic elements. International Journal of Optomechatronics, 7(1):1-14, 2013. 3", + "[63] Matteo Giuseppe Scopelliti and Maysamreza Chamanzar. Ultrasonically sculpted virtual relay lens for in situ microimaging. Light: Science & Applications, 8(1):1-15, 2019. 1, 3", + "[64] Xiaobing Shang, Jin-Yi Tan, Oliver Willekens, Jelle De Smet, Pankaj Joshi, Dieter Cuypers, Esma Islamaj, Jeroen Beeckman, Kristiaan Neyts, Michael Vervaekte, et al. Electrically controllable liquid crystal component for efficient light steering. IEEE Photonics Journal, 7(2):1-13, 2015. 3", + "[65] Braden Smith, Brandon Hellman, Adley Gin, Alonzo Espinoza, and Yuzuru Takashima. Single chip lidar with discrete beam steering by digital micromirror device. Optics Express, 25(13):14732-14745, 2017. 3", + "[66] Qi Wang Song, Xu-Ming Wang, Rebecca Bussjager, and Joseph Osman. Electro-optic beam-steering device based on a lanthanum-modified lead zirconate titanate ceramic wafer. Applied optics, 35(17):3155-3162, 1996. 3", + "[67] Steven J Spector. Review of lens-assisted beam steering methods. Journal of Optical Microsystems, 2(1):011003, 2022. 3", + "[68] Ernst HK Stelzer, Frederic Strobl, Bo-Jui Chang, Friedrich Preusser, Stephan Preibisch, Katie McDole, and Reto Fiolka. Light sheet fluorescence microscopy. Nature Reviews Methods Primers, 1(1):1-25, 2021. 1, 2", + "[69] Lin Sun, Jin-ha Kim, Chiou-hung Jang, Dechang An, Xue-jun Lu, Qingjun Zhou, John Martin Taboada, Ray T Chen, Jeffery J Maki, Suning Tang, et al. Polymeric waveguide prism-based electro-optic beam deflector. Optical Engineering, 40(7):1217-1222, 2001. 3", + "[70] Varun Sundar, Sizhuo Ma, Aswin C Sankaranarayanan, and Mohit Gupta. Single-photon structured light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17865-17875, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Opsys Tech. Opsys tech: The only scanning lidar with absolutely no moving parts. https://www.opsys-tech.com/post/opsys-tech-the-only-scanning-lidar-with-absolutely-no-moving-parts. Accessed: 2022-10-20. 3", + "[72] Chen S Tsai. Guided-wave acousto-optics: interactions, devices, and applications, volume 23. Springer Science & Business Media, 2013. 3", + "[73] Tomoki Ueda, Hiroyuki Kubo, Suren Jayasuriya, Takuya Funatomi, and Yasuhiro Mukaigawa. Slope disparity gating using a synchronized projector-camera system. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-9. IEEE, 2019. 2", + "[74] Federica Villa, Fabio Severini, Francesca Madonini, and Franco Zappa. Spads and sipms arrays for long-range high-speed light detection and ranging (lidar). Sensors, 21(11):3839, 2021. 1", + "[75] Dingkang Wang, Connor Watkins, and Huikai Xie. Mems mirrors for lidar: a review. Micromachines, 11(5):456, 2020. 1", + "[76] Jian Wang, Joseph Bartels, William Whittaker, Aswin C Sankaranarayanan, and Srinivasa G Narasimhan. Programmable triangulation light curtains. In Proceedings of the European Conference on Computer Vision (ECCV), pages 19-34, 2018. 2", + "[77] Xiangru Wang, Liang Wu, Caidong Xiong, Man Li, Qinggui Tan, Jiyang Shang, Shuanghong Wu, and Qi Qiu. Agile laser beam deflection with high steering precision and angular resolution using liquid crystal optical phased array. IEEE Transactions on Nanotechnology, 17(1):26-28, 2016. 3", + "[78] Zihan Wang, Jie Cao, Qun Hao, Fanghua Zhang, Yang Cheng, and Xianyue Kong. Super-resolution imaging and field of view extension using a single camera with risley prisms. Review of Scientific Instruments, 90(3):033701, 2019. 2", + "[79] RM Waxler and CE Weir. Effect of pressure and temperature on the refractive indices of benzene, carbon tetrachloride, and water. Journal of research of the National Bureau of Standards. Section A, Physics and chemistry, 67(2):163, 1963. 1", + "[80] Ming-Tzo Wei, Shana Elbaum-Garfinkle, Alex S Holehouse, Carlos Chih-Hsiung Chen, Marina Feric, Craig B Arnold, Rodney D Priestley, Rohit V Pappu, and Clifford P Brangwynne. Phase behaviour of disordered proteins underlying low density and high permeability of liquid organelles. Nature chemistry, 9(11):1118-1125, 2017. 3", + "[81] Shumian Xin, Sotiris Nousias, Kiriakos N Kutulakos, Aswin C Sankaranarayanan, Srinivasa G Narasimhan, and Ioannis Gkioulekas. A theory of fermat paths for non-line-of-sight shape reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6800-6809, 2019. 2", + "[82] Chris Xu, Warren Zipfel, Jason B Shear, Rebecca M Williams, and Watt W Webb. Multiphoton fluorescence excitation: new spectral windows for biological nonlinear microscopy. Proceedings of the National Academy of Sciences, 93(20):10763-10768, 1996. 1, 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 491, + 946, + 504, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Jiawen Xu and Jiong Tang. Tunable prism based on piezoelectric metamaterial for acoustic beam steering. Applied Physics Letters, 110(18):181902, 2017. 3", + "[84] Donghai Yang, Yifan Liu, Qingjiu Chen, Meng Chen, Shaodong Zhan, Nim-kwan Cheung, Ho-Yin Chan, Zhidong Wang, and Wen Jung Li. Development of the high angular resolution $360^{\\circ}$ lidar based on scanning mem's mirror. Scientific Reports, 13(1):1540, 2023. 2", + "[85] Zejie Yu and Xiankai Sun. Gigahertz acousto-optic modulation and frequency shifting on etchless lithium niobate integrated platform. ACS Photonics, 8(3):798-803, 2021. 3", + "[86] Mantas Žurauskas, Oliver Barnstedt, Maria Frade-Rodriguez, Scott Waddell, and Martin J Booth. Rapid adaptive remote focusing microscope for sensing of volumetric neural activity. Biomedical optics express, 8(10):4369-4379, 2017. 3" + ], + "bbox": [ + 78, + 90, + 470, + 304 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 491, + 946, + 506, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_model.json b/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8767f2d48f608f42683ff40eae60c1aa3dd5c008 --- /dev/null +++ b/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_model.json @@ -0,0 +1,2732 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.242, + 0.131, + 0.73, + 0.153 + ], + "angle": 0, + "content": "Megahertz Light Steering without Moving Parts" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.157, + 0.862, + 0.194 + ], + "angle": 0, + "content": "Adithya Pediredla\\(^{1,2}\\), Srinivasa G. Narasimhan\\(^{2}\\), Maysamreza Chamanzar\\(^{2}\\), Ioannis Gkioulekas\\(^{2}\\), Dartmouth College, Carnegie Mellon University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.228, + 0.314, + 0.244 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.253, + 0.474, + 0.601 + ], + "angle": 0, + "content": "We introduce a light steering technology that operates at megahertz frequencies, has no moving parts, and costs less than a hundred dollars. Our technology can benefit many projector and imaging systems that critically rely on high-speed, reliable, low-cost, and wavelength-independent light steering, including laser scanning projectors, LiDAR sensors, and fluorescence microscopes. Our technology uses ultrasound waves to generate a spatiotemporally-varying refractive index field inside a compressible medium, such as water, turning the medium into a dynamic traveling lens. By controlling the electrical input of the ultrasound transducers that generate the waves, we can change the lens, and thus steer light, at the speed of sound (1.5 km/s in water). We build a physical prototype of this technology, use it to realize different scanning techniques at megahertz rates (three orders of magnitude faster than commercial alternatives such as galvo mirror scanners), and demonstrate proof-of-concept projector and LiDAR applications. To encourage further innovation towards this new technology, we derive theory for its fundamental limits and develop a physically-accurate simulator for virtual design. Our technology offers a promising solution for achieving high-speed and low-cost light steering in a variety of applications." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.607, + 0.21, + 0.622 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Many imaging systems rely on the ability to steer light, either as it leaves a source or as it reaches a sensor. Examples include laser scanning projectors [28, 56], LiDAR depth sensors [38, 74, 75], and microscopy techniques (confocal microscopy [23, 50], light-sheet microscopy [55, 68], multiphoton microscopy [17, 82]). Compared to full-field lighting and imaging, light steering systems help improve light efficiency [47], counter indirect illumination [27, 42], and enhance illumination and imaging contrast [9, 55]. However, these advantages come at the cost of reduced acquisition speed, bulky moving hardware, and motion artifacts. To alleviate these costs, we introduce a new light steering technology that, through the use of ultrasonic sculpting, makes it possible to scan light both transversally and axially at megahertz (MHz) rates. Additionally, our technology achieves these high scanning rates without any moving parts. Lastly, prototypes of our technology cost no more than a hundred dollars. Altogether, these characteristics represent significant" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.23, + 0.892, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.516, + 0.895, + 0.613 + ], + "angle": 0, + "content": "Figure 1. Megahertz light steering. (a) Light steering systems in LiDAR systems and laser projectors have moving mechanical components, limiting them to kHz scanning rates. (b) Our technology uses the acousto-optic effect to enable MHz light steering without moving parts. The insets show that, before galvo mirrors could scan even a few points, our prototype scanned a thousand points to project the letter \"A\" on the wall." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.62, + 0.894, + 0.635 + ], + "angle": 0, + "content": "advances over previous light steering technologies (Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.895, + 0.877 + ], + "angle": 0, + "content": "Our technology uses the acousto-optic effect to turn a transparent medium, such as water, into a programmable optic that steers an incident light beam. Sound is a pressure wave that travels inside a medium by compressing and rarefying it, spatiotemporally changing the medium density. In turn, this changes the refractive index of the medium, which is proportional to the density [63,79]. We design the pressure profile of the sound wave so that, at any time instant, the spatially-varying refractive index makes the medium behave as a periodic set of virtual gradient-index (GRIN) lenses, each with an aperture equal to the sound wavelength. The GRIN lenses bend light beams incident on the medium, with the GRIN profile determining the beam trajectory. These lenses travel at the speed of sound (1.5 km/s in water) and are reconfigurable at MHz frequencies, allowing us to steer light faster than mechanical devices. To enable flexible steering" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.885, + 0.9 + ], + "angle": 0, + "content": "This is different from Bragg's diffraction in acousto-optic deflectors." + }, + { + "type": "page_number", + "bbox": [ + 0.495, + 0.946, + 0.504, + 0.956 + ], + "angle": 0, + "content": "1" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.166 + ], + "angle": 0, + "content": "patterns, we combine this optic with a pulsed laser with a programmable pulse rate. By synchronizing the laser source with the sound waveform, and modulating the phase of the sound waveform, we control both the speed of beam steering and the location of the beam." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.472, + 0.333 + ], + "angle": 0, + "content": "In Sections 3 and 4, we explain the physical and mathematical details of our technology. We introduce a new design that uses two linear transducers to generate traveling acoustic waves, and discuss how different synchronization choices between ultrasound and pulsed laser result in different scanning patterns. To facilitate the exploration of design parameters (ultrasound speed and frequency, laser frequency) and configurations (transducer geometry), we also develop a physics-based simulator for our technology. In Section 6, we also discuss the fundamental limits of our technology due to diffraction and the uncertainty principle in wave physics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.472, + 0.56 + ], + "angle": 0, + "content": "In Section 5, we experimentally demonstrate these fast programmable light steering techniques for various applications. In particular, in Section 5.2, we demonstrate an arbitrary point projector that can scan arbitrary and programmable light patterns. Compared to raster scanning projectors, which can project billions of points per second in a grid pattern but only a few thousand arbitrary points per second, our prototype can project a million arbitrary points per second, an acceleration by three orders of magnitude. In Section 5.3, we demonstrate a LiDAR prototype that combines our light steering technology with a single-photon avalanche diode (SPAD). We show 3D scans of \\(100 \\times 100\\) resolution at 5000 frames per second (50 million points per second) with a single-pixel SPAD, which is not feasible with scanning galvo mirrors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.561, + 0.373, + 0.574 + ], + "angle": 0, + "content": "Contributions. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.47, + 0.619 + ], + "angle": 0, + "content": "1. A new light steering technology based on the acoustooptic effect that is three orders of magnitude faster than state-of-the-art mechanical steering technologies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.471, + 0.65 + ], + "angle": 0, + "content": "2. A new hardware design with planar transducers generating of traveling waves." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.471, + 0.681 + ], + "angle": 0, + "content": "3. An experimental prototype demonstrating ultrafast arbitrary point projection and LiDAR scanning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.471, + 0.711 + ], + "angle": 0, + "content": "4. A physics-based renderer to simulate digital twins of our prototype and evaluate different designs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.471, + 0.756 + ], + "angle": 0, + "content": "5. The derivation of limits due to fundamental restrictions from wave physics (diffraction limit, scanning speed vs. aperture tradeoff, and uncertainty principle)." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.576, + 0.471, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.787 + ], + "angle": 0, + "content": "We provide our open-source simulator, data, and additional details in the supplement and project website.2" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.472, + 0.878 + ], + "angle": 0, + "content": "Limitations. Our prototype has a diffraction-limited point spread function (PSF) with a large spatial extent and a “+” shape, due to the use of two linear transducers that create a rectangular aperture. This limits spatial resolution, and introduces structured blur artifacts. These limitations are not fundamental to our core technology, and can be overcome" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.894, + 0.173 + ], + "angle": 0, + "content": "Table 1. Comparison of light steering technologies. MEMS is microelectromechanical systems, OPA is optical phased arrays, BW is optical bandwidth. For the arrows, red is bad, green is good, down is low, and up is high, and more arrows imply a bigger effect. Our method is superior in terms of cost, speed and supported bandwidth, with no moving parts or fabrication." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.175, + 0.868, + 0.285 + ], + "angle": 0, + "content": "
Tech.speedfab.BWcostmoving
galvo↓↓×↑↓✓
liquid lenses↓↓↓×↑↓×
MEMS↑✓↑↑✓
OPA↑↑↑✓↓↑↑×
ours↑↑×↑↓↓↓×
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.294, + 0.882, + 0.309 + ], + "angle": 0, + "content": "with improved designs and better engineering (Section 6)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.638, + 0.334 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.895, + 0.537 + ], + "angle": 0, + "content": "Light steering in imaging systems. Light steering is a core component in scanning-based active imaging systems. For example, fluorescence microscopy techniques such as confocal microscopy [23, 50], multiphoton microscopy [17, 82], light sheet microscopy [55, 68], and superresolution microscopy [7], use scanning to decrease scattered light and improve light efficiency and imaging contrast. Another example is LiDAR sensors [84] found in commercial applications, such as autonomous cars. In these sensors, scanning decreases multipath interference and helps reduce hardware cost, removing the need for two-dimensional LiDAR arrays. Our technology improves the speed and reliability of scanning-based LiDAR, while further reducing the cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.734 + ], + "angle": 0, + "content": "Light steering is also used in scanning-based laser projectors, to achieve high light efficiency and contrast by illuminating only where necessary. Laser projectors typically perform a 2D raster scan the field of view, with a fast and a slow scanning axis. The speed of the fast scanning axis limits the frame rate to the order of hundreds of kHz. In contrast, our technique can operate at tens of MHz, improving the raster scan rate by two orders of magnitude. Additionally, raster scanners cannot project arbitrary point sequences at a fast rate, and are limited to the frame rate of the projector (60-120 points per second) even when the projection pattern is very sparse. We demonstrate the projection of a million arbitrary points per second, four orders of magnitude faster." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.855 + ], + "angle": 0, + "content": "Computational imaging techniques that use laser projectors or LiDAR are generally also scanning-based. Examples include structured light [26, 70], light-transport probing [47, 48], motion contrast 3D [44], epipolar gating [3, 46], light curtains [11, 76], slope-disparity gating [12, 37, 73], and non-line-of-sight imaging [39, 41, 49, 51, 81]. As a consequence, all these techniques can become orders of magnitude faster if combined with our light steering device." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Light steering technologies We can distinguish between mechanical and non-mechanical light steering techniques. The former include rotating prisms [2, 78], mirrors [35, 45]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.887, + 0.458, + 0.901 + ], + "angle": 0, + "content": "2https://imaging.cs.cmu.edu/ultrafast_steering" + }, + { + "type": "page_number", + "bbox": [ + 0.494, + 0.947, + 0.504, + 0.957 + ], + "angle": 0, + "content": "2" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.47, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.18, + 0.471, + 0.25 + ], + "angle": 0, + "content": "Figure 2. Comparison to previous work. (a) Previous ultrasonically sculpted lenses [10, 32] change the focal length at high speed (b) Our lens can change both focal length (transparencies) and the spatial location (colors) at high speed. We achieve this using traveling waves and synchronizing sound and light signals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.473, + 0.417 + ], + "angle": 0, + "content": "and digital micromechanical systems (MEMS) [61, 65]. The latter include acousto-optic (AO) [20, 83] or electro-optic (EO) [69] deflectors, liquid crystal devices [64, 77], and optical phased arrays (OPA) [29, 58]. Table 1 compares these techniques to our MHz steering technology: Mechanical techniques are slow due to the need to physically move optical elements. AO and EO deflectors operate at kHz rates. Liquid lenses and crystals typically have a long settling time, making them the slowest among these techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.473, + 0.518 + ], + "angle": 0, + "content": "OPAs [29, 58] are solid-state on-chip devices that can steer light at even GHz rates. However, they require on-chip microscopic coherent laser generators [71] and cannot easily be coupled to external lasers. They are also expensive to fabricate and currently limited to low-resolution angles. Cheng et al. [15] and Spector [67] have detailed reviews." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.681 + ], + "angle": 0, + "content": "Acousto-optic devices. Many acousto-optic devices are commercially available, including tunable filters [21], modulators [5], frequency shifters [85], and deflectors [30]. Among these, the acousto-optic deflector can serve as a light steering device. However, acousto-optic deflectors use a different physical phenomenon, Bragg's diffraction [72], where beam deviation is proportional to the acoustic wave frequency, and are thus orders of magnitude slower than our technology. Electro-optic deflectors [40, 66] operate on a similar phenomenon and are thus similarly slow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Tunable acoustic gradient-index (TAG) lenses [16,32] and ultrasonically-sculpted virtual optical waveguides [10, 34] use the same physical principles as our technology. TAG lenses can change the focus depth of an incident beam at \\(\\mathrm{kHz}\\) rates; however, unlike our technology, they cannot steer it in the transverse axis (Figure 2). Following their recent commercialization, TAG lenses fostered innovation in scientific and application fields such as laser micromachining [14, 18], three-dimensional biomedical imaging [36, 80], microscopy [13, 19, 33], optical coherence tomography [22], high-throughput industrial inspection [31], and adaptive optics [62, 86]. We believe that the additional transverse steering capabilities from our technology will similarly help stimulate significant further innovation." + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.09, + 0.855, + 0.104 + ], + "angle": 0, + "content": "Table 2. Notation and parameters we use in the paper." + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.106, + 0.819, + 0.216 + ], + "angle": 0, + "content": "
quantitysymbol
speedCUS
wavelengthλus
frequencyfus
angular frequencyωus = 2πfus
wavenumberkus = 2πλus-1
timeperiodTus = fus-1
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.222, + 0.822, + 0.239 + ], + "angle": 0, + "content": "3. Traveling-wave acousto-optic lenses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.895, + 0.318 + ], + "angle": 0, + "content": "We generate ultrasonic waves inside a transparent medium, such as water, by submerging inside it a planar transducer that we drive with a single harmonic voltage (Figure 3(a)). The resulting pressure \\( P(x,t) \\) inside the medium (Figure 3(b)) equals" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.324, + 0.892, + 0.34 + ], + "angle": 0, + "content": "\\[\nP (x, t) = P _ {o} + P _ {s} \\cos \\left(k _ {\\mathrm {u s}} x - \\omega_ {\\mathrm {u s}} t\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.895, + 0.42 + ], + "angle": 0, + "content": "where: \\( t \\) is time, \\( x \\) is distance normal to the transducer plane—the pressure is independent of \\( y \\) and \\( z \\) coordinates—\\( P_{o} \\) is the medium pressure without ultrasound, \\( P_{s} \\) is proportional to the transducer voltage amplitude, and the remaining parameters are in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.894, + 0.451 + ], + "angle": 0, + "content": "The refractive index \\( n(x,t) \\) of the medium changes proportionally to the pressure (Figure 3(c)):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.458, + 0.892, + 0.473 + ], + "angle": 0, + "content": "\\[\nn (x, t) = n _ {o} + n _ {s} \\cos \\left(k _ {\\mathrm {u s}} x - \\omega_ {\\mathrm {u s}} t\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.524 + ], + "angle": 0, + "content": "where: \\( n_{o} \\) is the refractive index of the medium without the transducer, \\( n_{s} = kP_{s} \\), and \\( k \\) is an empirical coefficient (\\( k = 1.402 \\times 10^{-5} \\) bar\\(^{-1}\\) for water) [63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.894, + 0.614 + ], + "angle": 0, + "content": "At time \\( t = 0 \\), the convex lobes of this refractive index profile (i.e., regions \\( x = [l\\lambda_{\\mathrm{us}} - \\lambda_{\\mathrm{us}} / 2, l\\lambda_{\\mathrm{us}} + \\lambda_{\\mathrm{us}} / 2] \\); \\( l \\in \\mathbb{Z} \\)) act as GRIN lenses. Each of these lenses focuses light rays traveling parallel to the transducer onto a line (Figure 3(d)). We can change the lens focal length and aperture by varying the amplitude and frequency of the transducer voltage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.894, + 0.705 + ], + "angle": 0, + "content": "Traveling lenses. The pressure wave and refractive index profile propagate along the \\(x\\)-direction. Thus, the convex lobes vary as: \\(x(t) = [c_{\\mathrm{us}}t + l\\lambda_{\\mathrm{us}} - \\lambda_{\\mathrm{us}} / 2, c_{\\mathrm{us}}t + l\\lambda_{\\mathrm{us}} + \\lambda_{\\mathrm{us}} / 2]\\); \\(l \\in \\mathbb{Z}\\). As a result, ultrasonically-sculpted cylindrical GRIN lenses are dynamic, and focused lines travel normal to the transducer at the speed of ultrasound (Figure 4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.714, + 0.696, + 0.731 + ], + "angle": 0, + "content": "4. Scanning techniques" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Even though the ultrasonically-sculpted GRIN lenses travel at the speed of ultrasound, we cannot control the speed or location of the lens focus. To enable such control, we use a pulsed laser with the same repetition frequency as the ultrasound frequency, and programmable phase modulation for the transducer voltage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Single transducer. For intuition, we first describe the case of a single transducer. Due to phase modulation, the pressure pattern from Equation (1) becomes \\( P(x,t) = P_{o} + P_{s}\\cos (k_{\\mathrm{us}}x - \\omega_{\\mathrm{us}}t - \\phi (t)) \\). If the illumination is continuous, the position of the focused light is \\( x_{n}(t) = \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.494, + 0.946, + 0.504, + 0.956 + ], + "angle": 0, + "content": "3" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.502, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.314, + 0.229, + 0.327 + ], + "angle": 0, + "content": "(a) setup" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.097, + 0.714, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.226, + 0.683, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.315, + 0.651, + 0.327 + ], + "angle": 0, + "content": "(c) refractive index" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.103, + 0.887, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.223, + 0.888, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.313, + 0.852, + 0.327 + ], + "angle": 0, + "content": "(d) ray diagram" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.894, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Ultrasonic sculpting. (a) By placing a planar piezoelectric transducer inside a compressible medium and vibrating the transducer with a sinusoidal voltage, we generate ultrasound inside the medium. (b) Sound is a pressure wave, and at any given time, the pressure inside the medium varies spatially based on the voltage waveform applied to the transducer. (c) The change in the refractive index of the medium is proportional to the pressure. Therefore, the refractive index of the medium also varies spatially, turning the medium into a GRIN lens. (d) Light rays that pass through this GRIN lens will curve continuously and focus on a set of lines. The focal length of these waveguides/lenses is a function of the voltage and frequency applied to the transducer." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.42, + 0.263, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.554, + 0.212, + 0.566 + ], + "angle": 0, + "content": "(a) ray diagrams" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.426, + 0.469, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.555, + 0.399, + 0.567 + ], + "angle": 0, + "content": "(b) images" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.665 + ], + "angle": 0, + "content": "Figure 4. Planar transducer steers the light. As sound propagates, the spatial refractive index pattern also propagates at the speed of sound. The focussed light, which is a set of lines, also travels normal to the transducer plane at the speed of sound. Apart from the ray diagrams, the images rendered with our simulator (submitted as supplementary) show that the planar transducer moves the focus region (in this case, a line) at the speed of sound." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.799 + ], + "angle": 0, + "content": "\\(c_{\\mathrm{us}}t - n\\lambda_{\\mathrm{us}} + \\frac{\\phi(t)}{2\\pi}\\lambda ;n\\in \\mathbb{Z},\\) which is a set of lines moving continuously at the speed of sound. By pulsing the illumination at ultrasound frequency, we get a set of lines flickering at the ultrasound frequency: \\(x_{n}(mT_{\\mathrm{us}}) = (m - n)\\lambda_{\\mathrm{us}} + \\frac{\\phi(mT_{\\mathrm{us}})}{2\\pi}\\lambda_{\\mathrm{us}}.\\) If we modulate the phase linearly \\((\\phi (t) = k\\omega_{\\mathrm{us}}t;k < 1)\\) ,we have \\(x(mT_{\\mathrm{us}}) = mk c_{\\mathrm{us}}T_{\\mathrm{us}}\\) for \\(m = n\\) (i.e., the line due to the same laser and ultrasound pulse). Therefore, the lines travel at a reduced speed \\(kc_{\\mathrm{us}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.394, + 0.822 + ], + "angle": 0, + "content": "4.1. Point scanning with two transducers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To focus light to a point, we use two orthogonal planar transducers (Figure 5). We describe two extensions of the above approach for focus point control: one for scanning arbitrary point locations at the ultrasound frequency (i.e., MHz), and another for raster scanning at the laser repetition" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.42, + 0.879, + 0.436 + ], + "angle": 0, + "content": "frequency, which is higher than the ultrasound frequency." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.442, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Arbitrary point scanning. To scan arbitrary points \\((x(mT_{\\mathrm{us}}),y(mT_{\\mathrm{us}}))\\) for each laser pulse \\(m\\), we modulate the phases \\(\\phi_x(t)\\) and \\(\\phi_y(t)\\) of both transducers. The focus point location within the region \\([0,\\lambda_{\\mathrm{us}}]\\times [0,\\lambda_{\\mathrm{us}}]\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.508, + 0.892, + 0.524 + ], + "angle": 0, + "content": "\\[\nx (t) = \\left(\\phi_ {x} (t) / k _ {\\mathrm {u s}} + c _ {\\mathrm {u s}} t\\right) \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.527, + 0.892, + 0.544 + ], + "angle": 0, + "content": "\\[\ny (t) = \\left(\\phi_ {y} (t) / k _ {\\mathrm {u s}} + c _ {\\mathrm {u s}} t\\right) \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.894, + 0.626 + ], + "angle": 0, + "content": "Figure 6 shows the refractive index and ray diagram for four sets of \\(\\phi_{x}\\) and \\(\\phi_{y}\\) values. To scan a set of points \\((x(mT_{\\mathrm{us}}), y(mT_{\\mathrm{us}}))\\), we compute the phases \\((\\phi_{x}(mT_{\\mathrm{us}}), \\phi_{y}(mT_{\\mathrm{us}}))\\) using Equations (3)-(4), and interpolate to compute \\((\\phi_{x}(t), \\phi_{y}(t))\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Raster scanning In theory, we could use arbitrary point scanning to raster scan a two-dimensional grid of points. In that case, the phase modulation for raster scanning would be linear, \\((\\phi_x(t),\\phi_y(t)) = (k_x\\omega_{\\mathrm{us}}t,k_y\\omega_{\\mathrm{us}}t)\\), where \\(k_{x}\\) and \\(k_{y}\\) are phase modulation rates. The phase modulation rate for the faster axis would be equal to the product of the number of scan points and the modulation rate of the slower axis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.894, + 0.804 + ], + "angle": 0, + "content": "However, this approach would limit raster scanning frequency to the ultrasound frequency. If the laser repetition frequency is higher, we can scan more points by running the laser at its highest frequency. Their locations will be:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.809, + 0.892, + 0.839 + ], + "angle": 0, + "content": "\\[\nx \\left(m T _ {L}\\right) = \\left(m k _ {x}\\right) \\frac {\\lambda_ {\\mathrm {u s}}}{s} \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.84, + 0.892, + 0.868 + ], + "angle": 0, + "content": "\\[\ny \\left(m T _ {L}\\right) = \\left(m k _ {y}\\right) \\frac {\\lambda_ {\\mathrm {u s}}}{s} \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.872, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where \\( s = f_{L} / f_{\\mathrm{us}} \\) is the ratio of laser \\( (f_{L}) \\) and ultrasound \\( (f_{\\mathrm{us}}) \\) frequencies, and \\( T_{L} = 1 / f_{L} \\) is the inter-pulse time." + }, + { + "type": "page_number", + "bbox": [ + 0.494, + 0.946, + 0.504, + 0.956 + ], + "angle": 0, + "content": "4" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.278, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.096, + 0.484, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.096, + 0.705, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.106, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.894, + 0.314 + ], + "angle": 0, + "content": "Figure 5. Two planar transducers for focusing light at a point. (a) We place two planar piezoelectric transducers orthogonal to each other inside a medium. We drive both transducers independently with a sinusoidal voltage. (b) The pressure wave inside the medium is a superposition of the pressure waves generated by the transducers. (c) The change in the refractive index is proportional to the net pressure. (d) Light rays from a wide beam focus on a set of points. We restrict the illumination beam size to focus light on a single point." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.32, + 0.449, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.527, + 0.471, + 0.584 + ], + "angle": 0, + "content": "Figure 6. Point steering. By controlling the phases of the sinusoidal voltages applied to the transducers, we control the location of the focus position. To continuously steer the focus location, we phase modulate the voltages applied to both transducers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.593, + 0.21, + 0.61 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.471, + 0.782 + ], + "angle": 0, + "content": "We discuss an experimental prototype implementing our acousto-optic light steering technology, and combining it with a pulsed laser, single-pixel SPAD, and galvo mirrors (for comparison). We use this prototype to demonstrate projector and LiDAR applications. We compare our light steering system with commercially available galvo mirrors, to demonstrate the speed and the new capabilities our system enables. We keep the field-of-view and aperture same for both systems. This comparison is not the most favorable for galvo mirrors, as they typically have a larger field of view, but it is a fair one for evaluating the system's speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.191, + 0.807 + ], + "angle": 0, + "content": "5.1. Prototype" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Figure 7 shows our experimental prototype. We place two planar piezo transducers (P-25.40mm-25.40mm-2.10mm-880-WFB, APC International, Ltd) orthogonal to each other and at an inclination of \\(45^{\\circ}\\) relative to an acrylic tank containing water, to minimize interference from interreflections. A signal generator (SDG6022x, Siglent Tech.) drives the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.518 + ], + "angle": 0, + "content": "transducers via a power amplifier (ENI A300, Bell electronics). We colocate a pulsed laser (ALPHALAS GmbH, PICOPOWER-LD-510) and a gated SPAD (Microphoton Devices s.r.l.) using a beamsplitter, similar to previous techniques [24, 25, 57]. However, we do not place a lens in front of the SPAD, as the ultrasonically-sculpted lens focuses light from the object onto the SPAD. Instead, we place a lens in front of the laser to create a diverging ray that undergoes the same focusing by the ultrasonically-sculpted lens. We place an aperture after the ultrasonically-sculpted lens to limit the scanning area to only one ultrasonic period. A \\(45^{\\circ}\\) mirror directs the beam to a pair of galvo mirrors (GVS-212, Thorlabs Inc.). We use the galvo mirrors only for comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.642 + ], + "angle": 0, + "content": "To synchronize the transducers, laser, SPAD gate, and SPAD timing circuit, we use two signal generators and synchronize their clock and trigger signals. We use one signal generator to drive the transducers. The two channels of the second signal generator run the laser and the SPAD gate. We explain the SPAD timing synchronization details in Section 5.3. We provide more details about our prototype, including design and alignment, in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.734, + 0.667 + ], + "angle": 0, + "content": "5.2. Arbitrary point projector" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.746 + ], + "angle": 0, + "content": "We use the technique in Section 4.1 to compute the phases and synthesize the transducer voltage waveforms required to project an arbitrary target sequence of points. We drive the transducers with this waveform and the laser at the same frequency as ultrasound (1 MHz)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.895, + 0.883 + ], + "angle": 0, + "content": "To project the same sequence of points with the galvo mirrors, we drive the laser and transducers at a fixed frequency without any phase modulation, which results in a single-point focus. We steer this point with the galvo mirrors to scan the same desired sequence of points. We drive the galvo mirrors at \\(1\\mathrm{kHz}\\), \\(2\\mathrm{kHz}\\), \\(5\\mathrm{kHz}\\) frequencies (points per second). The galvo mirrors are rated for \\(1\\mathrm{kHz}\\), and driving it at frequencies higher than \\(5\\mathrm{kHz}\\) leads to higher motor current and failure of the fuse." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We project the patterns on a white cardboard screen, and" + }, + { + "type": "page_number", + "bbox": [ + 0.494, + 0.946, + 0.504, + 0.956 + ], + "angle": 0, + "content": "5" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.46, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.251, + 0.32, + 0.264 + ], + "angle": 0, + "content": "(a) schematic" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.265, + 0.458, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.424, + 0.362, + 0.438 + ], + "angle": 0, + "content": "(b) experimental prototype" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.438, + 0.473, + 0.619 + ], + "angle": 0, + "content": "Figure 7. Hardware setup. We show the (a) schematic and (b) prototype built for showing the proof-of-concept applications and comparing them with a galvo mirror system. We diverge the laser beam, and the expanded beam is focused by the ultrasonically-sculpted refractive index. The beam passes through the galvo mirrors onto the scene. We steer the beam either with the ultrasonically-sculpted lens or the galvo mirrors, but not both, depending on the experiment. The reflected light from the object takes the same path back to the sensor. The SPAD sensor, which is colocated with the laser, does not have any optics in front of it other than the ultrasonically-sculpted lenses. This setup allows us to compare the scanning speed of our system with galvo mirrors, while keeping the aperture the same." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.472, + 0.738 + ], + "angle": 0, + "content": "capture images with a camera (Allied Vision PRO-GT3400-09) for two exposures (1 ms and 50 ms) that we show in Figure 8. In this case, we are projecting 100 points that form the letter \"A\". At 1 ms exposure, the galvo mirrors only scan a few points, whereas our technique scans the entire shape ten times. At 100 ms or higher exposure, the galvo mirrors can project all the points without distortion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.744, + 0.17, + 0.758 + ], + "angle": 0, + "content": "5.3. LiDAR" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.884 + ], + "angle": 0, + "content": "We use a gated SPAD for depth and transient measurements. The gate helps reject the backscattered photons from various optics. After gating, our system does not suffer from pile-up [54, 59, 60]. We use a signal generator to drive the SPAD gate instead of the picosecond delayer (PSD) common in SPAD-based LiDAR systems [8, 41, 49, 53]. Our approach is inexpensive and generates programmable delays at much higher resolution (\\(1\\mu s\\)) than the PSD (\\(50\\mathrm{ns}\\))." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We run both the transducer and SPAD signal genera" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "tors in burst mode with the trigger running at \\(100\\mathrm{Hz}\\) for synchronization. We use picoharp (Microphoton Devices s.r.l.) to measure the time-of-flight of the photons and to synchronize the scanning position. Several previous techniques [39,43,49,57] achieve this synchronization by sending a synchronization signal to picoharp, which reports a marker event when the synchronization signal is detected. Unfortunately, as our scanning runs at a much higher rate than galvo mirror scanning, we cannot use an external synchronization signal for picoharp due to the massive number of events that would be generated. Fortunately, the picoharp reports the pulse index for each laser pulse it detects. We use the pulse index to compute when the photon is fired, and use Equations (5)-(6) to compute the scan position." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.457 + ], + "angle": 0, + "content": "We first use the pulsed raster scanning technique in Section 4.1 to raster scan the scene. We set \\( f_{L} = 50\\mathrm{MHz} \\), \\( f_{\\mathrm{us}} = 1\\mathrm{MHz} \\), \\( k_{x} = 1.0001 \\), and \\( k_{y} = 1.01 \\). These settings result in a spatial resolution of \\( 100\\times 100 \\) and scan rate of 5000 frames per second. Figure 9 shows results for four scenes, each with two letters at various depths. Our single channel (single sensor and laser) system scans 50 million points per second. This is three orders of magnitude faster than the 128-channel high-end Velodyne VLS-128™ LiDAR [1], which scans 5 million points per second." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.896, + 0.7 + ], + "angle": 0, + "content": "We also use the arbitrary point scanning technique in Section 4.1 to selectively scan a few points in the field of view. An advantage of such an approach is the potential for adaptive scanning [6]. In Figure 10, we create a three-point scene and image it to compute their spatial locations. We then scan these three points using both galvo mirrors and our technique for \\(3\\mathrm{ms}\\) (1 ms per point). When scanning with the galvo mirrors, a lot of time is spent moving the galvo mirrors from one scan point to another. As a result, many pulses fired by the laser are wasted during the travel time of the galvo mirrors. By contrast, ultrasonic beam steering shifts the focus within the time period of the two laser photons (\\(1\\mu \\mathrm{s}\\)) and does not waste any pulses. As a result, our technique's depth estimation accuracy (computed using mean squared error averaged over a hundred experiments) is \\(250\\times\\) higher than that of the galvo mirrors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.713, + 0.614, + 0.728 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We introduced a high-speed, low-cost (the cost of each transducer is only $26) technology to steer light. Our technology uses the physics of traveling sound waves synchronized with illumination and imaging sensors to enable multiple scanning applications. We built an experimental prototype to demonstrate proof-of-concept applications, such as the ability to perform arbitrary point light projection and raster scanning for LiDAR at MHz rates. In the rest of this section, we discuss some fundamental limits and tradeoffs inherent in our light steering technology due to wave physics, and directions for further exploration they suggest. We summa" + }, + { + "type": "page_number", + "bbox": [ + 0.494, + 0.947, + 0.505, + 0.957 + ], + "angle": 0, + "content": "6" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.887, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.424, + 0.896, + 0.509 + ], + "angle": 0, + "content": "Figure 8. Acousto-optic vs. galvo-mirror projection. We compare our beam steering technique with commercially available Thorlabs galvo mirrors (GVS-212). The \"A\" shape is made up of 100 points. With a \\(1\\mathrm{MHz}\\) transducer, we are able to project a million points per second (pps), and hence, project ten thousand \"A\"s per second. Constrained by the laser's low beam power \\((20\\mu \\mathrm{W}\\) at \\(1\\mathrm{MHz})\\), practically we can only capture \"A\" at 1 ms exposure. The commercially available galvo mirrors, which are only rated at 1 kpps, only project a streak when driven at \\(1\\mathrm{kHz}\\), \\(2\\mathrm{kHz}\\), and \\(5\\mathrm{kHz}\\). At 50 ms exposure and \\(1\\mathrm{kHz}\\) scan rate, the galvo mirrors only project half the pattern at the rated \\(1\\mathrm{kHz}\\), and at higher frequencies, the galvo mirrors project a corrupted pattern as we are operating them well beyond their 1 kpps rating." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.516, + 0.889, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.81, + 0.895, + 0.867 + ], + "angle": 0, + "content": "Figure 9. Depth raster scanning. Each of the four scenes has two characters, \"CV\", \"PR\", \"20\", \"23\". \"C\" and \"P\" are at approximately \\(160\\mathrm{cm}\\) depth, \"2\" at \\(170\\mathrm{cm}\\), and the remaining at \\(180\\mathrm{cm}\\). The top row shows the peak of the transient measured by the SPAD, and the bottom row shows the depth map in cm. We scan the scene at \\(100 \\times 100\\) resolution using the raster scanning technique in Section 4.1 for an exposure of one second. The Thorlabs galvo mirrors are not capable of scanning these scenes in under a second." + }, + { + "type": "page_number", + "bbox": [ + 0.494, + 0.946, + 0.506, + 0.956 + ], + "angle": 0, + "content": "7" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.075, + 0.27, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.214, + 0.245, + 0.226 + ], + "angle": 0, + "content": "(a) galvo mirrors" + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.226, + 0.262, + 0.238 + ], + "angle": 0, + "content": "(depth error \\(= 51.3\\mathrm{cm}\\)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.075, + 0.44, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.214, + 0.413, + 0.226 + ], + "angle": 0, + "content": "(b) our technique" + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.226, + 0.43, + 0.238 + ], + "angle": 0, + "content": "(depth error \\(= 3.28\\mathrm{cm}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.473, + 0.446 + ], + "angle": 0, + "content": "Figure 10. Transient and depth measurement. We scan three small objects placed at different depths using commercial galvo mirrors and our technique. The total exposure duration is \\(3\\mathrm{ms}\\), corresponding to the fastest speed (1kpps) at which the galvo mirrors can scan the scene. The laser repetition rate is \\(1\\mathrm{MHz}\\). We show three transients measured by (a) the galvo mirrors and (b) our technique when the scan location is at the first, second, and third point. The galvo mirrors spend a significant amount of time moving the focus point from one object to another, wasting pulses emitted during that time. In contrast, our technique does not waste any pulses, as it moves the focus point in just \\(1\\mu \\mathrm{s}\\), the time between two laser pulse emissions. This results in significant improvement in depth estimation accuracy (computed from the highest peak of the transient) compared to the galvo mirrors. The mean squared depth error of our technique is \\(16\\times\\) smaller than galvos." + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.449, + 0.391, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Figure 11. Beam shape. We show an HDR image of the measured beam. It has a cross shape, as the ultrasonic lens has a square aperture. Using more transducers can result in the beam shape to be more Gaussian-like. The center peak vs. the crosshairs intensity ratio is 20, and the center peak vs. the background ratio is 400." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.702 + ], + "angle": 0, + "content": "rize some of these tradeoffs in Table 3. In the supplement, we provide a physics-based renderer that can help virtually evaluate improved designs of our technology." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.869 + ], + "angle": 0, + "content": "Scanning speed vs. aperture tradeoff. For the arbitrary point projector, the scanning speed (points per second, pps) is equal to the frequency of the ultrasonic transducer voltage. The aperture of the ultrasonically-sculpted lens is equal to the ultrasound wavelength \\(\\lambda_{\\mathrm{us}} = c_{\\mathrm{us}} / f_{\\mathrm{us}}\\). Therefore, the product of scanning frequency and the lens aperture is always less than or equal to the speed of sound. Increasing the scanning speed decreases the aperture of the ultrasonic lens. This decrease in the aperture size is not a problem for projector applications, but for LiDAR applications, the decrease in aperture decreases light throughput." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Diffraction limit. The numerical aperture of the ultrasonically-sculpted lens is approximately \\(n\\lambda_{\\mathrm{us}} / 2F\\), where" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.104 + ], + "angle": 0, + "content": "Table 3. Tradeoffs between various system parameters. \\( \\uparrow \\) is better." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.106, + 0.83, + 0.171 + ], + "angle": 0, + "content": "
parametercusfusλn
scan speed↑↑↓-
aperture↑↓--
spatial resolution↑↓↓↑
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.191, + 0.894, + 0.223 + ], + "angle": 0, + "content": "\\(F\\) is the focal length of the ultrasonically-sculped waveguide. Therefore, the diffraction-limited spot size is" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.225, + 0.892, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\Delta x = \\frac {\\lambda}{2 N A} \\approx \\frac {\\lambda}{\\lambda_ {\\mathrm {u s}}} \\frac {F}{n} = \\frac {\\lambda f _ {\\mathrm {u s}}}{c _ {\\mathrm {u s}}} \\frac {F}{n}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.268, + 0.895, + 0.344 + ], + "angle": 0, + "content": "Uncertainty principle. Spatial resolution is inversely proportional to the diffraction-limited spot size, and as mentioned earlier, temporal resolution is determined by the frequency of the transducer. So, we have the following uncertainty principle between spatial and temporal resolution:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.347, + 0.892, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\Delta x \\Delta T \\leq \\frac {\\lambda}{c _ {\\mathrm {u s}}} \\frac {F}{n}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.894, + 0.457 + ], + "angle": 0, + "content": "To improve spatiotemporal resolution, we can decrease the focal length of the waveguide by increasing the voltage applied to the transducer. However, increasing the voltage makes the system non-linear and increases the probability of cavitation in the medium [52]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.895, + 0.631 + ], + "angle": 0, + "content": "Another approach to improve the system's overall quality (increase aperture size, spatial and temporal resolutions) is to use a medium with a higher speed of sound. For example, tellurium dioxide \\(\\mathrm{(TeO_2)}\\) glass has three times higher speed of sound and \\(50\\%\\) higher refractive index than water. Using tellurium dioxide glass will improve light efficiency by an order of magnitude, and simultaneously improve spatial resolution by five times. This material is also part of existing acousto-optic devices [30], but we found them hard to modify. Tellurium dioxide, being a solid, would additionally be a more stable medium than liquid water." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.894, + 0.789 + ], + "angle": 0, + "content": "Shape of the blur kernel. We used two transducers in our system, and each one of them creates a cylindrical lens. The net effect of these two cylindrical lenses is a square aperture, whose Fourier transform is the product of two 1D sincs. Therefore, our blur kernel has a cross-shape, as we can see in Fig. 11. We can make this blur kernel closer to a Gaussian-like blur kernel by using multiple transducers arranged around a circular path and synchronized. Based on the application, we can also use deconvolution techniques to improve the results in post-processing [4]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. We thank Hossein Baktash and Lloyd Lobo for help with the hardware prototype, and Andre Nascimento and Amy Lee for help with the physics-based renderer. This work was supported by NSF awards 1730147, 1900821, 1900849, 1935849, gifts from AWS Cloud Credits for Research and the Sybiel Berkman Foundation, and a Sloan Research Fellowship for Ioannis Gkioulekas." + }, + { + "type": "page_number", + "bbox": [ + 0.494, + 0.946, + 0.504, + 0.956 + ], + "angle": 0, + "content": "8" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.472, + 0.159 + ], + "angle": 0, + "content": "[1] Technical specifications of Alpha Prime (VLS-128). https://www.mapix.com/lidar-scanner-sensors/velodyne/alpha-prime/. Accessed: 2023-03-27. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.161, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[2] Mohammad Abdo, Vlad Badilita, and Jan Korvink. Spatial scanning hyperspectral imaging combining a rotating slit with a dove prism. Optics Express, 27(15):20290-20304, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.206, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[3] Supreeth Achar, Joseph R. Bartels, William L. 'Red' Whittaker, Kiriakos N. Kutulakos, and Srinivasa G. Narasimhan. Epipolar time-of-flight imaging. ACM Trans. Graph., 36(4), jul 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.265, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[4] Hossein Baktash, Yash Belhe, Matteo Giuseppe Scopelliti, Yi Hua, Aswin C Sankaranarayanan, and Maysamreza Chamanzar. Computational imaging using ultrasonically-sculpted virtual lenses. In 2022 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.351, + 0.472, + 0.392 + ], + "angle": 0, + "content": "[5] Jared Beller and Linbo Shao. Acousto-optic modulators integrated on-chip. Light: Science & Applications, 11(1):1-2, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.395, + 0.472, + 0.465 + ], + "angle": 0, + "content": "[6] Alexander W Bergman, David B Lindell, and Gordon Wetzstein. Deep adaptive lidar: End-to-end optimization of sampling and depth completion at low sampling rates. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.472, + 0.523 + ], + "angle": 0, + "content": "[7] Charles Bond, Adriana N Santiago-Ruiz, Qing Tang, and Melike Lakadamyali. Technological advances in super-resolution microscopy to study cellular processes. Molecular Cell, 82(2):315-332, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.526, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[8] Mauro Buttafava, Jessica Zeman, Alberto Tosi, Kevin Eliceiri, and Andreas Velten. Non-line-of-sight imaging using a time-gated single photon avalanche diode. Optics express, 23(16):20997-21011, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.584, + 0.472, + 0.64 + ], + "angle": 0, + "content": "[9] Walter A Carrington, Kevin E Fogarty, Larry Lifschitz, and Fredric S Fay. Three-dimensional imaging on confocal and wide-field microscopes. Handbook of biological confocal microscopy, pages 151-161, 1990. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.472, + 0.712 + ], + "angle": 0, + "content": "[10] Maysamreza Chamanzar, Matteo Giuseppe Scopelliti, Julien Bloch, Ninh Do, Minyoung Huh, Dongjin Seo, Jillian Iafrati, Vikaaas S Sohal, Mohammad-Reza Alam, and Michel M Maharbiz. Ultrasonic sculpting of virtual optical waveguides in tissue. Nature communications, 10(1):1-10, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.472, + 0.783 + ], + "angle": 0, + "content": "[11] Dorian Chan, Srinivasa G Narasimhan, and Matthew O'Toole. Holocurtains: Programming light curtains via binary holography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17886-17895, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[12] Sreenithy Chandran, Hiroyuki Kubo, Tomoki Ueda, Takuya Funatomi, Yasuhiro Mukaigawa, and Suren Jayasuriya. Slopedisparity gating: System and applications. IEEE Transactions on Computational Imaging, 8:317-332, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] Ting-Hsuan Chen, Jesse Ault, Howard A Stone, and Craig B Arnold. High-speed axial-scanning wide-field microscopy for volumetric particle tracking velocimetry. *Experiments in Fluids*, 58(5):1-7, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[14] Ting-Hsuan Chen, Romain Fardel, and Craig B Arnold. Ultrafast z-scanning for high-efficiency laser micro-machining. Light: Science & Applications, 7(4):17181-17181, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[15] Yang Cheng, Jie Cao, and Qun Hao. Optical beam steering using liquid-based devices. Optics and Lasers in Engineering, 146:106700, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[16] Maxim N Cherkashin, Carsten Brenner, Georg Schmitz, and Martin R Hofmann. Transversally travelling ultrasound for light guiding deep into scattering media. Communications Physics, 3(1):1-11, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[17] Winifried Denk, James H Strickler, and Watt W Webb. Two-photon laser scanning fluorescence microscopy. Science, 248(4951):73-76, 1990. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.894, + 0.333 + ], + "angle": 0, + "content": "[18] Xiaohan Du, SeungYeon Kang, and Craig B Arnold. Optimization of ultrafast axial scanning parameters for efficient pulsed laser micro-machining. Journal of Materials Processing Technology, 288:116850, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.894, + 0.376 + ], + "angle": 0, + "content": "[19] Martí Duocastella, Craig B Arnold, and Jason Puchalla. S electable light-sheet uniformity using tuned axial scanning. Microscopy research and technique, 80(2):250-259, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[20] Vladislav Gavyusev, Giuseppe Sancataldo, Pietro Ricci, Alberto Montalbano, Chiara Fornetto, Lapo Turrini, Annunziatina Laurino, Luca Pesce, Giuseppe de Vito, Natascia Tiso, et al. Dual-beam confocal light-sheet microscopy via flexible acousto-optic deflector. Journal of biomedical optics, 24(10):106504, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[21] Milton S Gottlieb. Acousto-optic tunable filters. In Design and fabrication of acousto-optic devices, pages 197-283. CRC Press, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[22] Ireneusz Grulkowski, Krzysztof Szulzycki, and Maciej Wojtkowski. Microscopic oct imaging with focus extension by ultrahigh-speed acousto-optic tunable lens and stroboscopic illumination. Optics Express, 22(26):31746-31760, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.893, + 0.588 + ], + "angle": 0, + "content": "[23] Min Gu. Principles of three dimensional imaging in confocal microscopes. World Scientific, 1996. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[24] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3d imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7918, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[25] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6770-6779, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[26] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G Narasimhan. Structured light 3d scanning in the presence of global illumination. In CVPR 2011, pages 713-720. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.893, + 0.829 + ], + "angle": 0, + "content": "[27] Kristofer Henderson, Xiaomeng Liu, Justin Folden, Brevin Tilmon, Suren Jayasuriya, and Sanjeev Koppal. Design and calibration of a fast flying-dot projector for dynamic light transport acquisition. IEEE Transactions on Computational Imaging, 6:529-543, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[28] Sven TS Holmström, Utku Baran, and Hakan Urey. Mems laser scanners: a review. Journal of Microelectromechanical Systems, 23(2):259-275, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[29] Ching-Pai Hsu, Boda Li, Braulio Solano-Rivas, Amar R Gohil, Pak Hung Chan, Andrew D Moore, and Valentina" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.494, + 0.946, + 0.505, + 0.956 + ], + "angle": 0, + "content": "9" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.471, + 0.134 + ], + "angle": 0, + "content": "Donzella. A review and perspective on optical phased array for automotive lidar. IEEE Journal of Selected Topics in Quantum Electronics, 27(1):1-16, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.471, + 0.193 + ], + "angle": 0, + "content": "[30] Dawoon Jeong, Hansol Jang, Min Uk Jung, and Chang-Seok Kim. Angular resolution variable fmcw lidar with acoustooptic deflector. In Imaging Systems and Applications, pages ITh4D-3. Optica Publishing Group, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.195, + 0.472, + 0.25 + ], + "angle": 0, + "content": "[31] S Kang, E Dotsenko, D Amrhein, C Theriault, and Craig B Arnold. Ultra-high-speed variable focus optics for novel applications in advanced imaging. In Photonic Instrumentation Engineering V, volume 10539, page 1053902. SPIE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.253, + 0.472, + 0.294 + ], + "angle": 0, + "content": "[32] SeungYeon Kang, Martí Duocastella, and Craig B Arnold. Variable optical elements for fast focus control. Nature Photonics, 14(9):533-542, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.472, + 0.352 + ], + "angle": 0, + "content": "[33] Doycho Karagyozov, Mirna Mihovilovic Skanata, Amanda Lesar, and Marc Gershow. Recording neural activity in unrestrained animals with three-dimensional tracking two-photon microscopy. Cell reports, 25(5):1371-1383, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.472, + 0.41 + ], + "angle": 0, + "content": "[34] Yasin Karimi, Matteo Giuseppe Scopelliti, Ninh Do, Mohammad-Reza Alam, and Maysamreza Chamanzar. In situ 3d reconfigurable ultrasonically sculpted optical beam paths. Optics express, 27(5):7249-7265, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.472, + 0.467 + ], + "angle": 0, + "content": "[35] Byung-Sub Kim, Steve Gibson, and Tsu-Chin Tsao. Adaptive control of a tilt mirror for laser beam steering. In Proceedings of the 2004 American Control Conference, volume 4, pages 3417–3421. IEEE, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.472, + 0.538 + ], + "angle": 0, + "content": "[36] Lingjie Kong, Jianyong Tang, Justin P Little, Yang Yu, Tim Lammermann, Charles P Lin, Ronald N Germain, and Meng Cui. Continuous volumetric imaging via an optical phase-locked ultrasound lens. Nature methods, 12(8):759-762, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.472, + 0.611 + ], + "angle": 0, + "content": "[37] Hiroyuki Kubo, Suren Jayasuriya, Takafumi Iwaguchi, Takuya Funatomi, Yasuhiro Mukaigawa, and Srinivasa G Narasimhan. Programmable non-epipolar indirect light transport: Capture and analysis. IEEE Transactions on Visualization and Computer Graphics, 27(4):2421-2436, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.472, + 0.669 + ], + "angle": 0, + "content": "[38] Nanxi Li, Chong Pei Ho, Jin Xue, Leh Woon Lim, Guanyu Chen, Yuan Hsing Fu, and Lennon Yao Ting Lee. A progress review on solid-state lidar and nanophotonics-based lidar sensors. *Laser & Photonics Reviews*, page 2100511, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.472, + 0.726 + ], + "angle": 0, + "content": "[39] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fkmigration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.472, + 0.771 + ], + "angle": 0, + "content": "[40] Qiyu Liu, Huan Li, and Mo Li. Electromechanical Brillouin scattering in integrated optomechanical waveguides. Optica, 6(6):778-785, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[41] Xiaochun Liu, Ibón Guillén, Marco La Manna, Ji Hyun Nam, Syed Azer Reza, Toan Huu Le, Adrian Jarabo, Diego Gutierrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572(7771):620-623, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[42] Xiaomeng Liu, Kristofer Henderson, Joshua Rego, Suren Jayasuriya, and Sanjeev Koppal. Dense lissajous sampling and interpolation for dynamic light-transport. Optics Express, 29(12):18362–18381, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[43] Julio Marco, Adrian Jarabo, Ji Hyun Nam, Xiaochun Liu, Miguel Ángel Cosculluela, Andreas Velten, and Diego Gutierrez. Virtual light transport matrices for non-line-of-sight imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2440-2449, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[44] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.894, + 0.299 + ], + "angle": 0, + "content": "[45] Aongus McCarthy, Nils J Krichel, Nathan R Gemmell, Ximing Ren, Michael G Tanner, Sander N Dorenbos, Val Zwiller, Robert H Hadfield, and Gerald S Buller. Kilometer-range, high resolution depth imaging via 1560 nm wavelength single-photon detection. Optics express, 21(7):8904-8915, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.894, + 0.355 + ], + "angle": 0, + "content": "[46] Srinivasa Narasimhan, Supreeth Achar, Kiriakos Kutulakos, Joe Bartels, and William Whittaker. Method for epipolar time of flight imaging, Mar. 19 2020. US Patent App. 16/468,617. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[47] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[48] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual coding to probe light transport. ACM Trans. Graph., 31(4):39-1, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[49] Matthew O'Toole, David B Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555(7696):338-341, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[50] James Pawley. Handbook of biological confocal microscopy, volume 236. Springer Science & Business Media, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[51] Adithya Pediredla, Akshit Dave, and Ashok Veeraraghavan. Snlos: Non-line-of-sight scanning through temporal focusing. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.894, + 0.637 + ], + "angle": 0, + "content": "[52] Adithya Pediredla, Matteo Giuseppe Scopelliti, Srinivasa Narasimhan, Maysamreza Chamanzar, and Ioannis Gkioulekas. Optimized virtual optical waveguides enhance light throughput in scattering media. preprint, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.894, + 0.72 + ], + "angle": 0, + "content": "[53] Adithya Kumar Pediredla, Mauro Buttafava, Alberto Tosi, Oliver Cossairt, and Ashok Veeraraghavan. Reconstructing rooms using photon echoes: A plane based model and reconstruction algorithm for looking around the corner. In 2017 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[54] Adithya K Pediredla, Aswin C Sankaranarayanan, Mauro Buttafava, Alberto Tosi, and Ashok Veeraraghavan. Signal processing based pile-up compensation for gated single-photon avalanche diodes. arXiv preprint arXiv:1806.07437, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[55] Adithya Kumar Pediredla, Shizheng Zhang, Ben Avants, Fan Ye, Shin Nagayama, Ziying Chen, Caleb Kemere, Jacob T Robinson, and Ashok Veeraraghavan. Deep imaging in scattering media with selective plane illumination microscopy. Journal of biomedical optics, 21(12):126009, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[56] Kurt E Petersen. Silicon torsional scanning mirror. IBM Journal of Research and Development, 24(5):631-637, 1980. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.492, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "10" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[57] Ryan Po, Adithya Pediredla, and Ioannis Gkioulekas. Adaptive gating for single-photon 3d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16354–16363, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[58] Christopher V Poulton, Matthew J Byrd, Benjamin Moss, Erman Timurdogan, Ron Millman, and Michael R Watts. 8192-element optical phased array with \\(100^{\\circ}\\) steering range and flip-chip cmos. In CLEO: QELS_Fundamental Science, pages JTh4A-3. Optical Society of America, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[59] Joshua Rapp, Yanting Ma, Robin MA Dawson, and Vivek K Goyal. Dead time compensation for high-flux ranging. IEEE Transactions on Signal Processing, 67(13):3471-3486, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[60] Joshua Rapp, Yanting Ma, Robin MA Dawson, and Vivek K Goyal. High-flux single-photon lidar. Optica, 8(1):30-39, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[61] Joshua Rodriguez, Braden Smith, Brandon Hellman, Adley Gin, Alonzo Espinoza, and Yuzuru Takashima. Multi-beam and single-chip lidar with discrete beam steering by digital micromirror device. In Physics and Simulation of Optoelectronic Devices XXVI, volume 10526, pages 89-94. SPIE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.385, + 0.471, + 0.439 + ], + "angle": 0, + "content": "[62] Patrick S Salter, Z Iqbal, and Martin J Booth. Analysis of the three-dimensional focal positioning capability of adaptive optic elements. International Journal of Optomechatronics, 7(1):1-14, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.441, + 0.471, + 0.494 + ], + "angle": 0, + "content": "[63] Matteo Giuseppe Scopelliti and Maysamreza Chamanzar. Ultrasonically sculpted virtual relay lens for in situ microimaging. Light: Science & Applications, 8(1):1-15, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.497, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[64] Xiaobing Shang, Jin-Yi Tan, Oliver Willekens, Jelle De Smet, Pankaj Joshi, Dieter Cuypers, Esma Islamaj, Jeroen Beeckman, Kristiaan Neyts, Michael Vervaekte, et al. Electrically controllable liquid crystal component for efficient light steering. IEEE Photonics Journal, 7(2):1-13, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.567, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[65] Braden Smith, Brandon Hellman, Adley Gin, Alonzo Espinoza, and Yuzuru Takashima. Single chip lidar with discrete beam steering by digital micromirror device. Optics Express, 25(13):14732-14745, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.623, + 0.471, + 0.677 + ], + "angle": 0, + "content": "[66] Qi Wang Song, Xu-Ming Wang, Rebecca Bussjager, and Joseph Osman. Electro-optic beam-steering device based on a lanthanum-modified lead zirconate titanate ceramic wafer. Applied optics, 35(17):3155-3162, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[67] Steven J Spector. Review of lens-assisted beam steering methods. Journal of Optical Microsystems, 2(1):011003, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.72, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[68] Ernst HK Stelzer, Frederic Strobl, Bo-Jui Chang, Friedrich Preusser, Stephan Preibisch, Katie McDole, and Reto Fiolka. Light sheet fluorescence microscopy. Nature Reviews Methods Primers, 1(1):1-25, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[69] Lin Sun, Jin-ha Kim, Chiou-hung Jang, Dechang An, Xue-jun Lu, Qingjun Zhou, John Martin Taboada, Ray T Chen, Jeffery J Maki, Suning Tang, et al. Polymeric waveguide prism-based electro-optic beam deflector. Optical Engineering, 40(7):1217-1222, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[70] Varun Sundar, Sizhuo Ma, Aswin C Sankaranarayanan, and Mohit Gupta. Single-photon structured light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17865-17875, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.161 + ], + "angle": 0, + "content": "[71] Opsys Tech. Opsys tech: The only scanning lidar with absolutely no moving parts. https://www.opsys-tech.com/post/opsys-tech-the-only-scanning-lidar-with-absolutely-no-moving-parts. Accessed: 2022-10-20. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[72] Chen S Tsai. Guided-wave acousto-optics: interactions, devices, and applications, volume 23. Springer Science & Business Media, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.894, + 0.278 + ], + "angle": 0, + "content": "[73] Tomoki Ueda, Hiroyuki Kubo, Suren Jayasuriya, Takuya Funatomi, and Yasuhiro Mukaigawa. Slope disparity gating using a synchronized projector-camera system. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-9. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.893, + 0.336 + ], + "angle": 0, + "content": "[74] Federica Villa, Fabio Severini, Francesca Madonini, and Franco Zappa. Spads and sipms arrays for long-range high-speed light detection and ranging (lidar). Sensors, 21(11):3839, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.893, + 0.38 + ], + "angle": 0, + "content": "[75] Dingkang Wang, Connor Watkins, and Huikai Xie. Mems mirrors for lidar: a review. Micromachines, 11(5):456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.894, + 0.452 + ], + "angle": 0, + "content": "[76] Jian Wang, Joseph Bartels, William Whittaker, Aswin C Sankaranarayanan, and Srinivasa G Narasimhan. Programmable triangulation light curtains. In Proceedings of the European Conference on Computer Vision (ECCV), pages 19-34, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.893, + 0.526 + ], + "angle": 0, + "content": "[77] Xiangru Wang, Liang Wu, Caidong Xiong, Man Li, Qinggui Tan, Jiyang Shang, Shuanghong Wu, and Qi Qiu. Agile laser beam deflection with high steering precision and angular resolution using liquid crystal optical phased array. IEEE Transactions on Nanotechnology, 17(1):26-28, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.529, + 0.894, + 0.584 + ], + "angle": 0, + "content": "[78] Zihan Wang, Jie Cao, Qun Hao, Fanghua Zhang, Yang Cheng, and Xianyue Kong. Super-resolution imaging and field of view extension using a single camera with risley prisms. Review of Scientific Instruments, 90(3):033701, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.894, + 0.655 + ], + "angle": 0, + "content": "[79] RM Waxler and CE Weir. Effect of pressure and temperature on the refractive indices of benzene, carbon tetrachloride, and water. Journal of research of the National Bureau of Standards. Section A, Physics and chemistry, 67(2):163, 1963. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.894, + 0.742 + ], + "angle": 0, + "content": "[80] Ming-Tzo Wei, Shana Elbaum-Garfinkle, Alex S Holehouse, Carlos Chih-Hsiung Chen, Marina Feric, Craig B Arnold, Rodney D Priestley, Rohit V Pappu, and Clifford P Brangwynne. Phase behaviour of disordered proteins underlying low density and high permeability of liquid organelles. Nature chemistry, 9(11):1118-1125, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.894, + 0.827 + ], + "angle": 0, + "content": "[81] Shumian Xin, Sotiris Nousias, Kiriakos N Kutulakos, Aswin C Sankaranarayanan, Srinivasa G Narasimhan, and Ioannis Gkioulekas. A theory of fermat paths for non-line-of-sight shape reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6800-6809, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[82] Chris Xu, Warren Zipfel, Jason B Shear, Rebecca M Williams, and Watt W Webb. Multiphoton fluorescence excitation: new spectral windows for biological nonlinear microscopy. Proceedings of the National Academy of Sciences, 93(20):10763-10768, 1996. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.492, + 0.947, + 0.506, + 0.956 + ], + "angle": 0, + "content": "11" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "[83] Jiawen Xu and Jiong Tang. Tunable prism based on piezoelectric metamaterial for acoustic beam steering. Applied Physics Letters, 110(18):181902, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[84] Donghai Yang, Yifan Liu, Qingjiu Chen, Meng Chen, Shaodong Zhan, Nim-kwan Cheung, Ho-Yin Chan, Zhidong Wang, and Wen Jung Li. Development of the high angular resolution \\(360^{\\circ}\\) lidar based on scanning mem's mirror. Scientific Reports, 13(1):1540, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[85] Zejie Yu and Xiankai Sun. Gigahertz acousto-optic modulation and frequency shifting on etchless lithium niobate integrated platform. ACS Photonics, 8(3):798-803, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[86] Mantas Žurauskas, Oliver Barnstedt, Maria Frade-Rodriguez, Scott Waddell, and Martin J Booth. Rapid adaptive remote focusing microscope for sensing of volumetric neural activity. Biomedical optics express, 8(10):4369-4379, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.493, + 0.947, + 0.508, + 0.956 + ], + "angle": 0, + "content": "12" + } + ] +] \ No newline at end of file diff --git a/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_origin.pdf b/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd8dd7732036adff1f67271eb1c18fea57a28821 --- /dev/null +++ b/2023/Megahertz Light Steering Without Moving Parts/dc876064-04d9-4c6f-9ea4-3a0fbd5711b7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:558227c07c64d9eb4a20f5986c07cd205761d82cbb0bf74c63aadabbac560d49 +size 3688767 diff --git a/2023/Megahertz Light Steering Without Moving Parts/full.md b/2023/Megahertz Light Steering Without Moving Parts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..792abffd18e8fb7b18c2576e3badc0e907f63497 --- /dev/null +++ b/2023/Megahertz Light Steering Without Moving Parts/full.md @@ -0,0 +1,353 @@ +# Megahertz Light Steering without Moving Parts + +Adithya Pediredla $^{1,2}$ , Srinivasa G. Narasimhan $^{2}$ , Maysamreza Chamanzar $^{2}$ , Ioannis Gkioulekas $^{2}$ , Dartmouth College, Carnegie Mellon University + +# Abstract + +We introduce a light steering technology that operates at megahertz frequencies, has no moving parts, and costs less than a hundred dollars. Our technology can benefit many projector and imaging systems that critically rely on high-speed, reliable, low-cost, and wavelength-independent light steering, including laser scanning projectors, LiDAR sensors, and fluorescence microscopes. Our technology uses ultrasound waves to generate a spatiotemporally-varying refractive index field inside a compressible medium, such as water, turning the medium into a dynamic traveling lens. By controlling the electrical input of the ultrasound transducers that generate the waves, we can change the lens, and thus steer light, at the speed of sound (1.5 km/s in water). We build a physical prototype of this technology, use it to realize different scanning techniques at megahertz rates (three orders of magnitude faster than commercial alternatives such as galvo mirror scanners), and demonstrate proof-of-concept projector and LiDAR applications. To encourage further innovation towards this new technology, we derive theory for its fundamental limits and develop a physically-accurate simulator for virtual design. Our technology offers a promising solution for achieving high-speed and low-cost light steering in a variety of applications. + +# 1. Introduction + +Many imaging systems rely on the ability to steer light, either as it leaves a source or as it reaches a sensor. Examples include laser scanning projectors [28, 56], LiDAR depth sensors [38, 74, 75], and microscopy techniques (confocal microscopy [23, 50], light-sheet microscopy [55, 68], multiphoton microscopy [17, 82]). Compared to full-field lighting and imaging, light steering systems help improve light efficiency [47], counter indirect illumination [27, 42], and enhance illumination and imaging contrast [9, 55]. However, these advantages come at the cost of reduced acquisition speed, bulky moving hardware, and motion artifacts. To alleviate these costs, we introduce a new light steering technology that, through the use of ultrasonic sculpting, makes it possible to scan light both transversally and axially at megahertz (MHz) rates. Additionally, our technology achieves these high scanning rates without any moving parts. Lastly, prototypes of our technology cost no more than a hundred dollars. Altogether, these characteristics represent significant + +![](images/98bdc621d43009f9672422cd7f71f230268f5498cc29391576d98d6a0f44e677.jpg) + +![](images/f6ae1c32c570040b3786380601a27bba3c646318667c13fbe932773cae2d1cff.jpg) +Figure 1. Megahertz light steering. (a) Light steering systems in LiDAR systems and laser projectors have moving mechanical components, limiting them to kHz scanning rates. (b) Our technology uses the acousto-optic effect to enable MHz light steering without moving parts. The insets show that, before galvo mirrors could scan even a few points, our prototype scanned a thousand points to project the letter "A" on the wall. + +advances over previous light steering technologies (Table 1). + +Our technology uses the acousto-optic effect to turn a transparent medium, such as water, into a programmable optic that steers an incident light beam. Sound is a pressure wave that travels inside a medium by compressing and rarefying it, spatiotemporally changing the medium density. In turn, this changes the refractive index of the medium, which is proportional to the density [63,79]. We design the pressure profile of the sound wave so that, at any time instant, the spatially-varying refractive index makes the medium behave as a periodic set of virtual gradient-index (GRIN) lenses, each with an aperture equal to the sound wavelength. The GRIN lenses bend light beams incident on the medium, with the GRIN profile determining the beam trajectory. These lenses travel at the speed of sound (1.5 km/s in water) and are reconfigurable at MHz frequencies, allowing us to steer light faster than mechanical devices. To enable flexible steering + +patterns, we combine this optic with a pulsed laser with a programmable pulse rate. By synchronizing the laser source with the sound waveform, and modulating the phase of the sound waveform, we control both the speed of beam steering and the location of the beam. + +In Sections 3 and 4, we explain the physical and mathematical details of our technology. We introduce a new design that uses two linear transducers to generate traveling acoustic waves, and discuss how different synchronization choices between ultrasound and pulsed laser result in different scanning patterns. To facilitate the exploration of design parameters (ultrasound speed and frequency, laser frequency) and configurations (transducer geometry), we also develop a physics-based simulator for our technology. In Section 6, we also discuss the fundamental limits of our technology due to diffraction and the uncertainty principle in wave physics. + +In Section 5, we experimentally demonstrate these fast programmable light steering techniques for various applications. In particular, in Section 5.2, we demonstrate an arbitrary point projector that can scan arbitrary and programmable light patterns. Compared to raster scanning projectors, which can project billions of points per second in a grid pattern but only a few thousand arbitrary points per second, our prototype can project a million arbitrary points per second, an acceleration by three orders of magnitude. In Section 5.3, we demonstrate a LiDAR prototype that combines our light steering technology with a single-photon avalanche diode (SPAD). We show 3D scans of $100 \times 100$ resolution at 5000 frames per second (50 million points per second) with a single-pixel SPAD, which is not feasible with scanning galvo mirrors. + +# Contributions. Our main contributions are: + +1. A new light steering technology based on the acoustooptic effect that is three orders of magnitude faster than state-of-the-art mechanical steering technologies. +2. A new hardware design with planar transducers generating of traveling waves. +3. An experimental prototype demonstrating ultrafast arbitrary point projection and LiDAR scanning. +4. A physics-based renderer to simulate digital twins of our prototype and evaluate different designs. +5. The derivation of limits due to fundamental restrictions from wave physics (diffraction limit, scanning speed vs. aperture tradeoff, and uncertainty principle). + +We provide our open-source simulator, data, and additional details in the supplement and project website.2 + +Limitations. Our prototype has a diffraction-limited point spread function (PSF) with a large spatial extent and a “+” shape, due to the use of two linear transducers that create a rectangular aperture. This limits spatial resolution, and introduces structured blur artifacts. These limitations are not fundamental to our core technology, and can be overcome + +Table 1. Comparison of light steering technologies. MEMS is microelectromechanical systems, OPA is optical phased arrays, BW is optical bandwidth. For the arrows, red is bad, green is good, down is low, and up is high, and more arrows imply a bigger effect. Our method is superior in terms of cost, speed and supported bandwidth, with no moving parts or fabrication. + +
Tech.speedfab.BWcostmoving
galvo↓↓×↑↓✓
liquid lenses↓↓↓×↑↓×
MEMS↑✓↑↑✓
OPA↑↑↑✓↓↑↑×
ours↑↑×↑↓↓↓×
+ +with improved designs and better engineering (Section 6). + +# 2. Related work + +Light steering in imaging systems. Light steering is a core component in scanning-based active imaging systems. For example, fluorescence microscopy techniques such as confocal microscopy [23, 50], multiphoton microscopy [17, 82], light sheet microscopy [55, 68], and superresolution microscopy [7], use scanning to decrease scattered light and improve light efficiency and imaging contrast. Another example is LiDAR sensors [84] found in commercial applications, such as autonomous cars. In these sensors, scanning decreases multipath interference and helps reduce hardware cost, removing the need for two-dimensional LiDAR arrays. Our technology improves the speed and reliability of scanning-based LiDAR, while further reducing the cost. + +Light steering is also used in scanning-based laser projectors, to achieve high light efficiency and contrast by illuminating only where necessary. Laser projectors typically perform a 2D raster scan the field of view, with a fast and a slow scanning axis. The speed of the fast scanning axis limits the frame rate to the order of hundreds of kHz. In contrast, our technique can operate at tens of MHz, improving the raster scan rate by two orders of magnitude. Additionally, raster scanners cannot project arbitrary point sequences at a fast rate, and are limited to the frame rate of the projector (60-120 points per second) even when the projection pattern is very sparse. We demonstrate the projection of a million arbitrary points per second, four orders of magnitude faster. + +Computational imaging techniques that use laser projectors or LiDAR are generally also scanning-based. Examples include structured light [26, 70], light-transport probing [47, 48], motion contrast 3D [44], epipolar gating [3, 46], light curtains [11, 76], slope-disparity gating [12, 37, 73], and non-line-of-sight imaging [39, 41, 49, 51, 81]. As a consequence, all these techniques can become orders of magnitude faster if combined with our light steering device. + +Light steering technologies We can distinguish between mechanical and non-mechanical light steering techniques. The former include rotating prisms [2, 78], mirrors [35, 45], + +![](images/4ab681afbc93c5a77a6cc1d26875f8d3c39cd192483aebc184c27ba38f91e335.jpg) +Figure 2. Comparison to previous work. (a) Previous ultrasonically sculpted lenses [10, 32] change the focal length at high speed (b) Our lens can change both focal length (transparencies) and the spatial location (colors) at high speed. We achieve this using traveling waves and synchronizing sound and light signals. + +and digital micromechanical systems (MEMS) [61, 65]. The latter include acousto-optic (AO) [20, 83] or electro-optic (EO) [69] deflectors, liquid crystal devices [64, 77], and optical phased arrays (OPA) [29, 58]. Table 1 compares these techniques to our MHz steering technology: Mechanical techniques are slow due to the need to physically move optical elements. AO and EO deflectors operate at kHz rates. Liquid lenses and crystals typically have a long settling time, making them the slowest among these techniques. + +OPAs [29, 58] are solid-state on-chip devices that can steer light at even GHz rates. However, they require on-chip microscopic coherent laser generators [71] and cannot easily be coupled to external lasers. They are also expensive to fabricate and currently limited to low-resolution angles. Cheng et al. [15] and Spector [67] have detailed reviews. + +Acousto-optic devices. Many acousto-optic devices are commercially available, including tunable filters [21], modulators [5], frequency shifters [85], and deflectors [30]. Among these, the acousto-optic deflector can serve as a light steering device. However, acousto-optic deflectors use a different physical phenomenon, Bragg's diffraction [72], where beam deviation is proportional to the acoustic wave frequency, and are thus orders of magnitude slower than our technology. Electro-optic deflectors [40, 66] operate on a similar phenomenon and are thus similarly slow. + +Tunable acoustic gradient-index (TAG) lenses [16,32] and ultrasonically-sculpted virtual optical waveguides [10, 34] use the same physical principles as our technology. TAG lenses can change the focus depth of an incident beam at $\mathrm{kHz}$ rates; however, unlike our technology, they cannot steer it in the transverse axis (Figure 2). Following their recent commercialization, TAG lenses fostered innovation in scientific and application fields such as laser micromachining [14, 18], three-dimensional biomedical imaging [36, 80], microscopy [13, 19, 33], optical coherence tomography [22], high-throughput industrial inspection [31], and adaptive optics [62, 86]. We believe that the additional transverse steering capabilities from our technology will similarly help stimulate significant further innovation. + +Table 2. Notation and parameters we use in the paper. + +
quantitysymbol
speedCUS
wavelengthλus
frequencyfus
angular frequencyωus = 2πfus
wavenumberkus = 2πλus-1
timeperiodTus = fus-1
+ +# 3. Traveling-wave acousto-optic lenses + +We generate ultrasonic waves inside a transparent medium, such as water, by submerging inside it a planar transducer that we drive with a single harmonic voltage (Figure 3(a)). The resulting pressure $P(x,t)$ inside the medium (Figure 3(b)) equals + +$$ +P (x, t) = P _ {o} + P _ {s} \cos \left(k _ {\mathrm {u s}} x - \omega_ {\mathrm {u s}} t\right), \tag {1} +$$ + +where: $t$ is time, $x$ is distance normal to the transducer plane—the pressure is independent of $y$ and $z$ coordinates— $P_{o}$ is the medium pressure without ultrasound, $P_{s}$ is proportional to the transducer voltage amplitude, and the remaining parameters are in Table 2. + +The refractive index $n(x,t)$ of the medium changes proportionally to the pressure (Figure 3(c)): + +$$ +n (x, t) = n _ {o} + n _ {s} \cos \left(k _ {\mathrm {u s}} x - \omega_ {\mathrm {u s}} t\right), \tag {2} +$$ + +where: $n_{o}$ is the refractive index of the medium without the transducer, $n_{s} = kP_{s}$ , and $k$ is an empirical coefficient ( $k = 1.402 \times 10^{-5}$ bar $^{-1}$ for water) [63]. + +At time $t = 0$ , the convex lobes of this refractive index profile (i.e., regions $x = [l\lambda_{\mathrm{us}} - \lambda_{\mathrm{us}} / 2, l\lambda_{\mathrm{us}} + \lambda_{\mathrm{us}} / 2]$ ; $l \in \mathbb{Z}$ ) act as GRIN lenses. Each of these lenses focuses light rays traveling parallel to the transducer onto a line (Figure 3(d)). We can change the lens focal length and aperture by varying the amplitude and frequency of the transducer voltage. + +Traveling lenses. The pressure wave and refractive index profile propagate along the $x$ -direction. Thus, the convex lobes vary as: $x(t) = [c_{\mathrm{us}}t + l\lambda_{\mathrm{us}} - \lambda_{\mathrm{us}} / 2, c_{\mathrm{us}}t + l\lambda_{\mathrm{us}} + \lambda_{\mathrm{us}} / 2]$ ; $l \in \mathbb{Z}$ . As a result, ultrasonically-sculpted cylindrical GRIN lenses are dynamic, and focused lines travel normal to the transducer at the speed of ultrasound (Figure 4). + +# 4. Scanning techniques + +Even though the ultrasonically-sculpted GRIN lenses travel at the speed of ultrasound, we cannot control the speed or location of the lens focus. To enable such control, we use a pulsed laser with the same repetition frequency as the ultrasound frequency, and programmable phase modulation for the transducer voltage. + +Single transducer. For intuition, we first describe the case of a single transducer. Due to phase modulation, the pressure pattern from Equation (1) becomes $P(x,t) = P_{o} + P_{s}\cos (k_{\mathrm{us}}x - \omega_{\mathrm{us}}t - \phi (t))$ . If the illumination is continuous, the position of the focused light is $x_{n}(t) =$ + +![](images/40988d26d26223b134844c679c926106c1abb10a6215a4b23212ed4ac56a1262.jpg) +(a) setup + +![](images/51c2eb9d2618f335ca43a673e622f741703dabce171da9a6e98728fa14e01b06.jpg) + +![](images/3636dae10e2b138887d1e277d34c70281560b43a100ba41f937d24e68096b914.jpg) +(c) refractive index + +![](images/38c6a23d4b523f8f12b7d314651fce1898a145dbfb2eb39986b6bf5521e22f20.jpg) + +![](images/94e2c41db8f77d2a41076a6aa8a84b6ab379d8cefc8a2044e1ba1c43923e12dd.jpg) +(d) ray diagram + +![](images/1105b76db2e71f8f6798e2fc9445ec39b815f1d68e300f1f1a94092cf5cf495f.jpg) +Figure 3. Ultrasonic sculpting. (a) By placing a planar piezoelectric transducer inside a compressible medium and vibrating the transducer with a sinusoidal voltage, we generate ultrasound inside the medium. (b) Sound is a pressure wave, and at any given time, the pressure inside the medium varies spatially based on the voltage waveform applied to the transducer. (c) The change in the refractive index of the medium is proportional to the pressure. Therefore, the refractive index of the medium also varies spatially, turning the medium into a GRIN lens. (d) Light rays that pass through this GRIN lens will curve continuously and focus on a set of lines. The focal length of these waveguides/lenses is a function of the voltage and frequency applied to the transducer. +(a) ray diagrams +Figure 4. Planar transducer steers the light. As sound propagates, the spatial refractive index pattern also propagates at the speed of sound. The focussed light, which is a set of lines, also travels normal to the transducer plane at the speed of sound. Apart from the ray diagrams, the images rendered with our simulator (submitted as supplementary) show that the planar transducer moves the focus region (in this case, a line) at the speed of sound. + +![](images/d7ddfa99370e77f8ceece793d423c2e77dc9eba3b02a24ac34496def9b8030aa.jpg) +(b) images + +$c_{\mathrm{us}}t - n\lambda_{\mathrm{us}} + \frac{\phi(t)}{2\pi}\lambda ;n\in \mathbb{Z},$ which is a set of lines moving continuously at the speed of sound. By pulsing the illumination at ultrasound frequency, we get a set of lines flickering at the ultrasound frequency: $x_{n}(mT_{\mathrm{us}}) = (m - n)\lambda_{\mathrm{us}} + \frac{\phi(mT_{\mathrm{us}})}{2\pi}\lambda_{\mathrm{us}}.$ If we modulate the phase linearly $(\phi (t) = k\omega_{\mathrm{us}}t;k < 1)$ ,we have $x(mT_{\mathrm{us}}) = mk c_{\mathrm{us}}T_{\mathrm{us}}$ for $m = n$ (i.e., the line due to the same laser and ultrasound pulse). Therefore, the lines travel at a reduced speed $kc_{\mathrm{us}}$ + +# 4.1. Point scanning with two transducers + +To focus light to a point, we use two orthogonal planar transducers (Figure 5). We describe two extensions of the above approach for focus point control: one for scanning arbitrary point locations at the ultrasound frequency (i.e., MHz), and another for raster scanning at the laser repetition + +frequency, which is higher than the ultrasound frequency. + +Arbitrary point scanning. To scan arbitrary points $(x(mT_{\mathrm{us}}),y(mT_{\mathrm{us}}))$ for each laser pulse $m$ , we modulate the phases $\phi_x(t)$ and $\phi_y(t)$ of both transducers. The focus point location within the region $[0,\lambda_{\mathrm{us}}]\times [0,\lambda_{\mathrm{us}}]$ is: + +$$ +x (t) = \left(\phi_ {x} (t) / k _ {\mathrm {u s}} + c _ {\mathrm {u s}} t\right) \bmod \lambda_ {\mathrm {u s}}, \tag {3} +$$ + +$$ +y (t) = \left(\phi_ {y} (t) / k _ {\mathrm {u s}} + c _ {\mathrm {u s}} t\right) \bmod \lambda_ {\mathrm {u s}}, \tag {4} +$$ + +Figure 6 shows the refractive index and ray diagram for four sets of $\phi_{x}$ and $\phi_{y}$ values. To scan a set of points $(x(mT_{\mathrm{us}}), y(mT_{\mathrm{us}}))$ , we compute the phases $(\phi_{x}(mT_{\mathrm{us}}), \phi_{y}(mT_{\mathrm{us}}))$ using Equations (3)-(4), and interpolate to compute $(\phi_{x}(t), \phi_{y}(t))$ . + +Raster scanning In theory, we could use arbitrary point scanning to raster scan a two-dimensional grid of points. In that case, the phase modulation for raster scanning would be linear, $(\phi_x(t),\phi_y(t)) = (k_x\omega_{\mathrm{us}}t,k_y\omega_{\mathrm{us}}t)$ , where $k_{x}$ and $k_{y}$ are phase modulation rates. The phase modulation rate for the faster axis would be equal to the product of the number of scan points and the modulation rate of the slower axis. + +However, this approach would limit raster scanning frequency to the ultrasound frequency. If the laser repetition frequency is higher, we can scan more points by running the laser at its highest frequency. Their locations will be: + +$$ +x \left(m T _ {L}\right) = \left(m k _ {x}\right) \frac {\lambda_ {\mathrm {u s}}}{s} \bmod \lambda_ {\mathrm {u s}}, \tag {5} +$$ + +$$ +y \left(m T _ {L}\right) = \left(m k _ {y}\right) \frac {\lambda_ {\mathrm {u s}}}{s} \bmod \lambda_ {\mathrm {u s}}, \tag {6} +$$ + +where $s = f_{L} / f_{\mathrm{us}}$ is the ratio of laser $(f_{L})$ and ultrasound $(f_{\mathrm{us}})$ frequencies, and $T_{L} = 1 / f_{L}$ is the inter-pulse time. + +![](images/32959e457d339467566da9b521888dd6181d6aa93cc15a4f1c13cc68e7314ede.jpg) +Figure 5. Two planar transducers for focusing light at a point. (a) We place two planar piezoelectric transducers orthogonal to each other inside a medium. We drive both transducers independently with a sinusoidal voltage. (b) The pressure wave inside the medium is a superposition of the pressure waves generated by the transducers. (c) The change in the refractive index is proportional to the net pressure. (d) Light rays from a wide beam focus on a set of points. We restrict the illumination beam size to focus light on a single point. + +![](images/5360dc2c60743b082c541688f52bdec0aa513a6bb66871da834d5aa8cba93455.jpg) + +![](images/3dd7296e25b0a72a7cc2ababa16f439758b45b2e5b9dd6f2d64208c6ab097911.jpg) + +![](images/36c5c4cd976deb5cf9d74e7e0577039c46a24c13fc64b0c543ea225f03b874b1.jpg) + +![](images/699a4cdfd2aa2513ab725ab114b9627be0b13130f2f7a12123fbc751b40d1377.jpg) +Figure 6. Point steering. By controlling the phases of the sinusoidal voltages applied to the transducers, we control the location of the focus position. To continuously steer the focus location, we phase modulate the voltages applied to both transducers. + +# 5. Experiments + +We discuss an experimental prototype implementing our acousto-optic light steering technology, and combining it with a pulsed laser, single-pixel SPAD, and galvo mirrors (for comparison). We use this prototype to demonstrate projector and LiDAR applications. We compare our light steering system with commercially available galvo mirrors, to demonstrate the speed and the new capabilities our system enables. We keep the field-of-view and aperture same for both systems. This comparison is not the most favorable for galvo mirrors, as they typically have a larger field of view, but it is a fair one for evaluating the system's speed. + +# 5.1. Prototype + +Figure 7 shows our experimental prototype. We place two planar piezo transducers (P-25.40mm-25.40mm-2.10mm-880-WFB, APC International, Ltd) orthogonal to each other and at an inclination of $45^{\circ}$ relative to an acrylic tank containing water, to minimize interference from interreflections. A signal generator (SDG6022x, Siglent Tech.) drives the + +transducers via a power amplifier (ENI A300, Bell electronics). We colocate a pulsed laser (ALPHALAS GmbH, PICOPOWER-LD-510) and a gated SPAD (Microphoton Devices s.r.l.) using a beamsplitter, similar to previous techniques [24, 25, 57]. However, we do not place a lens in front of the SPAD, as the ultrasonically-sculpted lens focuses light from the object onto the SPAD. Instead, we place a lens in front of the laser to create a diverging ray that undergoes the same focusing by the ultrasonically-sculpted lens. We place an aperture after the ultrasonically-sculpted lens to limit the scanning area to only one ultrasonic period. A $45^{\circ}$ mirror directs the beam to a pair of galvo mirrors (GVS-212, Thorlabs Inc.). We use the galvo mirrors only for comparisons. + +To synchronize the transducers, laser, SPAD gate, and SPAD timing circuit, we use two signal generators and synchronize their clock and trigger signals. We use one signal generator to drive the transducers. The two channels of the second signal generator run the laser and the SPAD gate. We explain the SPAD timing synchronization details in Section 5.3. We provide more details about our prototype, including design and alignment, in the supplement. + +# 5.2. Arbitrary point projector + +We use the technique in Section 4.1 to compute the phases and synthesize the transducer voltage waveforms required to project an arbitrary target sequence of points. We drive the transducers with this waveform and the laser at the same frequency as ultrasound (1 MHz). + +To project the same sequence of points with the galvo mirrors, we drive the laser and transducers at a fixed frequency without any phase modulation, which results in a single-point focus. We steer this point with the galvo mirrors to scan the same desired sequence of points. We drive the galvo mirrors at $1\mathrm{kHz}$ , $2\mathrm{kHz}$ , $5\mathrm{kHz}$ frequencies (points per second). The galvo mirrors are rated for $1\mathrm{kHz}$ , and driving it at frequencies higher than $5\mathrm{kHz}$ leads to higher motor current and failure of the fuse. + +We project the patterns on a white cardboard screen, and + +![](images/a1b4d290c54087f5ebd83caa17f5ec71a646c59938edeb6dda6efb00e72d2acc.jpg) + +![](images/3d4f1682f72860d93b0ca61c9e94fda7c4c556446b9db283b1ba02c8fb30cc7b.jpg) +(a) schematic +(b) experimental prototype +Figure 7. Hardware setup. We show the (a) schematic and (b) prototype built for showing the proof-of-concept applications and comparing them with a galvo mirror system. We diverge the laser beam, and the expanded beam is focused by the ultrasonically-sculpted refractive index. The beam passes through the galvo mirrors onto the scene. We steer the beam either with the ultrasonically-sculpted lens or the galvo mirrors, but not both, depending on the experiment. The reflected light from the object takes the same path back to the sensor. The SPAD sensor, which is colocated with the laser, does not have any optics in front of it other than the ultrasonically-sculpted lenses. This setup allows us to compare the scanning speed of our system with galvo mirrors, while keeping the aperture the same. + +capture images with a camera (Allied Vision PRO-GT3400-09) for two exposures (1 ms and 50 ms) that we show in Figure 8. In this case, we are projecting 100 points that form the letter "A". At 1 ms exposure, the galvo mirrors only scan a few points, whereas our technique scans the entire shape ten times. At 100 ms or higher exposure, the galvo mirrors can project all the points without distortion. + +# 5.3. LiDAR + +We use a gated SPAD for depth and transient measurements. The gate helps reject the backscattered photons from various optics. After gating, our system does not suffer from pile-up [54, 59, 60]. We use a signal generator to drive the SPAD gate instead of the picosecond delayer (PSD) common in SPAD-based LiDAR systems [8, 41, 49, 53]. Our approach is inexpensive and generates programmable delays at much higher resolution ( $1\mu s$ ) than the PSD ( $50\mathrm{ns}$ ). + +We run both the transducer and SPAD signal genera + +tors in burst mode with the trigger running at $100\mathrm{Hz}$ for synchronization. We use picoharp (Microphoton Devices s.r.l.) to measure the time-of-flight of the photons and to synchronize the scanning position. Several previous techniques [39,43,49,57] achieve this synchronization by sending a synchronization signal to picoharp, which reports a marker event when the synchronization signal is detected. Unfortunately, as our scanning runs at a much higher rate than galvo mirror scanning, we cannot use an external synchronization signal for picoharp due to the massive number of events that would be generated. Fortunately, the picoharp reports the pulse index for each laser pulse it detects. We use the pulse index to compute when the photon is fired, and use Equations (5)-(6) to compute the scan position. + +We first use the pulsed raster scanning technique in Section 4.1 to raster scan the scene. We set $f_{L} = 50\mathrm{MHz}$ , $f_{\mathrm{us}} = 1\mathrm{MHz}$ , $k_{x} = 1.0001$ , and $k_{y} = 1.01$ . These settings result in a spatial resolution of $100\times 100$ and scan rate of 5000 frames per second. Figure 9 shows results for four scenes, each with two letters at various depths. Our single channel (single sensor and laser) system scans 50 million points per second. This is three orders of magnitude faster than the 128-channel high-end Velodyne VLS-128™ LiDAR [1], which scans 5 million points per second. + +We also use the arbitrary point scanning technique in Section 4.1 to selectively scan a few points in the field of view. An advantage of such an approach is the potential for adaptive scanning [6]. In Figure 10, we create a three-point scene and image it to compute their spatial locations. We then scan these three points using both galvo mirrors and our technique for $3\mathrm{ms}$ (1 ms per point). When scanning with the galvo mirrors, a lot of time is spent moving the galvo mirrors from one scan point to another. As a result, many pulses fired by the laser are wasted during the travel time of the galvo mirrors. By contrast, ultrasonic beam steering shifts the focus within the time period of the two laser photons ( $1\mu \mathrm{s}$ ) and does not waste any pulses. As a result, our technique's depth estimation accuracy (computed using mean squared error averaged over a hundred experiments) is $250\times$ higher than that of the galvo mirrors. + +# 6. Discussion + +We introduced a high-speed, low-cost (the cost of each transducer is only $26) technology to steer light. Our technology uses the physics of traveling sound waves synchronized with illumination and imaging sensors to enable multiple scanning applications. We built an experimental prototype to demonstrate proof-of-concept applications, such as the ability to perform arbitrary point light projection and raster scanning for LiDAR at MHz rates. In the rest of this section, we discuss some fundamental limits and tradeoffs inherent in our light steering technology due to wave physics, and directions for further exploration they suggest. We summa + +![](images/7e37173e651e71338d2ebe1160887581b25cfeeb92c5c082581557d0b3f98182.jpg) +Figure 8. Acousto-optic vs. galvo-mirror projection. We compare our beam steering technique with commercially available Thorlabs galvo mirrors (GVS-212). The "A" shape is made up of 100 points. With a $1\mathrm{MHz}$ transducer, we are able to project a million points per second (pps), and hence, project ten thousand "A"s per second. Constrained by the laser's low beam power $(20\mu \mathrm{W}$ at $1\mathrm{MHz})$ , practically we can only capture "A" at 1 ms exposure. The commercially available galvo mirrors, which are only rated at 1 kpps, only project a streak when driven at $1\mathrm{kHz}$ , $2\mathrm{kHz}$ , and $5\mathrm{kHz}$ . At 50 ms exposure and $1\mathrm{kHz}$ scan rate, the galvo mirrors only project half the pattern at the rated $1\mathrm{kHz}$ , and at higher frequencies, the galvo mirrors project a corrupted pattern as we are operating them well beyond their 1 kpps rating. + +![](images/36e855f27d82270e5352235e958fe28ce9fe4d71c9e986f741a4727af36020db.jpg) +Figure 9. Depth raster scanning. Each of the four scenes has two characters, "CV", "PR", "20", "23". "C" and "P" are at approximately $160\mathrm{cm}$ depth, "2" at $170\mathrm{cm}$ , and the remaining at $180\mathrm{cm}$ . The top row shows the peak of the transient measured by the SPAD, and the bottom row shows the depth map in cm. We scan the scene at $100 \times 100$ resolution using the raster scanning technique in Section 4.1 for an exposure of one second. The Thorlabs galvo mirrors are not capable of scanning these scenes in under a second. + +![](images/d9de0c97472f8a3f82ca7e9b78cbd563742e6ac3437ef19e08fc99da0236d84c.jpg) +(a) galvo mirrors + +![](images/f61e00bf1299a817d6553d9f7c863afdb35c09a86400f1ae554f29ffefb47359.jpg) +(depth error $= 51.3\mathrm{cm}$ +(b) our technique +(depth error $= 3.28\mathrm{cm}$ + +![](images/37b73fd585935e410e6dbe17f0382c158996069afcb8021bf883ef92aaf9f5ce.jpg) +Figure 10. Transient and depth measurement. We scan three small objects placed at different depths using commercial galvo mirrors and our technique. The total exposure duration is $3\mathrm{ms}$ , corresponding to the fastest speed (1kpps) at which the galvo mirrors can scan the scene. The laser repetition rate is $1\mathrm{MHz}$ . We show three transients measured by (a) the galvo mirrors and (b) our technique when the scan location is at the first, second, and third point. The galvo mirrors spend a significant amount of time moving the focus point from one object to another, wasting pulses emitted during that time. In contrast, our technique does not waste any pulses, as it moves the focus point in just $1\mu \mathrm{s}$ , the time between two laser pulse emissions. This results in significant improvement in depth estimation accuracy (computed from the highest peak of the transient) compared to the galvo mirrors. The mean squared depth error of our technique is $16\times$ smaller than galvos. +Figure 11. Beam shape. We show an HDR image of the measured beam. It has a cross shape, as the ultrasonic lens has a square aperture. Using more transducers can result in the beam shape to be more Gaussian-like. The center peak vs. the crosshairs intensity ratio is 20, and the center peak vs. the background ratio is 400. + +rize some of these tradeoffs in Table 3. In the supplement, we provide a physics-based renderer that can help virtually evaluate improved designs of our technology. + +Scanning speed vs. aperture tradeoff. For the arbitrary point projector, the scanning speed (points per second, pps) is equal to the frequency of the ultrasonic transducer voltage. The aperture of the ultrasonically-sculpted lens is equal to the ultrasound wavelength $\lambda_{\mathrm{us}} = c_{\mathrm{us}} / f_{\mathrm{us}}$ . Therefore, the product of scanning frequency and the lens aperture is always less than or equal to the speed of sound. Increasing the scanning speed decreases the aperture of the ultrasonic lens. This decrease in the aperture size is not a problem for projector applications, but for LiDAR applications, the decrease in aperture decreases light throughput. + +Diffraction limit. The numerical aperture of the ultrasonically-sculpted lens is approximately $n\lambda_{\mathrm{us}} / 2F$ , where + +Table 3. Tradeoffs between various system parameters. $\uparrow$ is better. + +
parametercusfusλn
scan speed↑↑↓-
aperture↑↓--
spatial resolution↑↓↓↑
+ +$F$ is the focal length of the ultrasonically-sculped waveguide. Therefore, the diffraction-limited spot size is + +$$ +\Delta x = \frac {\lambda}{2 N A} \approx \frac {\lambda}{\lambda_ {\mathrm {u s}}} \frac {F}{n} = \frac {\lambda f _ {\mathrm {u s}}}{c _ {\mathrm {u s}}} \frac {F}{n}. \tag {7} +$$ + +Uncertainty principle. Spatial resolution is inversely proportional to the diffraction-limited spot size, and as mentioned earlier, temporal resolution is determined by the frequency of the transducer. So, we have the following uncertainty principle between spatial and temporal resolution: + +$$ +\Delta x \Delta T \leq \frac {\lambda}{c _ {\mathrm {u s}}} \frac {F}{n}. \tag {8} +$$ + +To improve spatiotemporal resolution, we can decrease the focal length of the waveguide by increasing the voltage applied to the transducer. However, increasing the voltage makes the system non-linear and increases the probability of cavitation in the medium [52]. + +Another approach to improve the system's overall quality (increase aperture size, spatial and temporal resolutions) is to use a medium with a higher speed of sound. For example, tellurium dioxide $\mathrm{(TeO_2)}$ glass has three times higher speed of sound and $50\%$ higher refractive index than water. Using tellurium dioxide glass will improve light efficiency by an order of magnitude, and simultaneously improve spatial resolution by five times. This material is also part of existing acousto-optic devices [30], but we found them hard to modify. Tellurium dioxide, being a solid, would additionally be a more stable medium than liquid water. + +Shape of the blur kernel. We used two transducers in our system, and each one of them creates a cylindrical lens. The net effect of these two cylindrical lenses is a square aperture, whose Fourier transform is the product of two 1D sincs. Therefore, our blur kernel has a cross-shape, as we can see in Fig. 11. We can make this blur kernel closer to a Gaussian-like blur kernel by using multiple transducers arranged around a circular path and synchronized. Based on the application, we can also use deconvolution techniques to improve the results in post-processing [4]. + +Acknowledgments. We thank Hossein Baktash and Lloyd Lobo for help with the hardware prototype, and Andre Nascimento and Amy Lee for help with the physics-based renderer. This work was supported by NSF awards 1730147, 1900821, 1900849, 1935849, gifts from AWS Cloud Credits for Research and the Sybiel Berkman Foundation, and a Sloan Research Fellowship for Ioannis Gkioulekas. + +# References + +[1] Technical specifications of Alpha Prime (VLS-128). https://www.mapix.com/lidar-scanner-sensors/velodyne/alpha-prime/. Accessed: 2023-03-27. 6 +[2] Mohammad Abdo, Vlad Badilita, and Jan Korvink. Spatial scanning hyperspectral imaging combining a rotating slit with a dove prism. Optics Express, 27(15):20290-20304, 2019. 2 +[3] Supreeth Achar, Joseph R. Bartels, William L. 'Red' Whittaker, Kiriakos N. Kutulakos, and Srinivasa G. Narasimhan. Epipolar time-of-flight imaging. ACM Trans. Graph., 36(4), jul 2017. 2 +[4] Hossein Baktash, Yash Belhe, Matteo Giuseppe Scopelliti, Yi Hua, Aswin C Sankaranarayanan, and Maysamreza Chamanzar. Computational imaging using ultrasonically-sculpted virtual lenses. In 2022 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2022. 8 +[5] Jared Beller and Linbo Shao. Acousto-optic modulators integrated on-chip. Light: Science & Applications, 11(1):1-2, 2022. 3 +[6] Alexander W Bergman, David B Lindell, and Gordon Wetzstein. Deep adaptive lidar: End-to-end optimization of sampling and depth completion at low sampling rates. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2020. 6 +[7] Charles Bond, Adriana N Santiago-Ruiz, Qing Tang, and Melike Lakadamyali. Technological advances in super-resolution microscopy to study cellular processes. Molecular Cell, 82(2):315-332, 2022. 2 +[8] Mauro Buttafava, Jessica Zeman, Alberto Tosi, Kevin Eliceiri, and Andreas Velten. Non-line-of-sight imaging using a time-gated single photon avalanche diode. Optics express, 23(16):20997-21011, 2015. 6 +[9] Walter A Carrington, Kevin E Fogarty, Larry Lifschitz, and Fredric S Fay. Three-dimensional imaging on confocal and wide-field microscopes. Handbook of biological confocal microscopy, pages 151-161, 1990. 1 +[10] Maysamreza Chamanzar, Matteo Giuseppe Scopelliti, Julien Bloch, Ninh Do, Minyoung Huh, Dongjin Seo, Jillian Iafrati, Vikaaas S Sohal, Mohammad-Reza Alam, and Michel M Maharbiz. Ultrasonic sculpting of virtual optical waveguides in tissue. Nature communications, 10(1):1-10, 2019. 3 +[11] Dorian Chan, Srinivasa G Narasimhan, and Matthew O'Toole. Holocurtains: Programming light curtains via binary holography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17886-17895, 2022. 2 +[12] Sreenithy Chandran, Hiroyuki Kubo, Tomoki Ueda, Takuya Funatomi, Yasuhiro Mukaigawa, and Suren Jayasuriya. Slopedisparity gating: System and applications. IEEE Transactions on Computational Imaging, 8:317-332, 2022. 2 +[13] Ting-Hsuan Chen, Jesse Ault, Howard A Stone, and Craig B Arnold. High-speed axial-scanning wide-field microscopy for volumetric particle tracking velocimetry. *Experiments in Fluids*, 58(5):1-7, 2017. 3 + +[14] Ting-Hsuan Chen, Romain Fardel, and Craig B Arnold. Ultrafast z-scanning for high-efficiency laser micro-machining. Light: Science & Applications, 7(4):17181-17181, 2018. 3 +[15] Yang Cheng, Jie Cao, and Qun Hao. Optical beam steering using liquid-based devices. Optics and Lasers in Engineering, 146:106700, 2021. 3 +[16] Maxim N Cherkashin, Carsten Brenner, Georg Schmitz, and Martin R Hofmann. Transversally travelling ultrasound for light guiding deep into scattering media. Communications Physics, 3(1):1-11, 2020. 3 +[17] Winifried Denk, James H Strickler, and Watt W Webb. Two-photon laser scanning fluorescence microscopy. Science, 248(4951):73-76, 1990. 1, 2 +[18] Xiaohan Du, SeungYeon Kang, and Craig B Arnold. Optimization of ultrafast axial scanning parameters for efficient pulsed laser micro-machining. Journal of Materials Processing Technology, 288:116850, 2021. 3 +[19] Martí Duocastella, Craig B Arnold, and Jason Puchalla. S electable light-sheet uniformity using tuned axial scanning. Microscopy research and technique, 80(2):250-259, 2017. 3 +[20] Vladislav Gavyusev, Giuseppe Sancataldo, Pietro Ricci, Alberto Montalbano, Chiara Fornetto, Lapo Turrini, Annunziatina Laurino, Luca Pesce, Giuseppe de Vito, Natascia Tiso, et al. Dual-beam confocal light-sheet microscopy via flexible acousto-optic deflector. Journal of biomedical optics, 24(10):106504, 2019. 3 +[21] Milton S Gottlieb. Acousto-optic tunable filters. In Design and fabrication of acousto-optic devices, pages 197-283. CRC Press, 2021. 3 +[22] Ireneusz Grulkowski, Krzysztof Szulzycki, and Maciej Wojtkowski. Microscopic oct imaging with focus extension by ultrahigh-speed acousto-optic tunable lens and stroboscopic illumination. Optics Express, 22(26):31746-31760, 2014. 3 +[23] Min Gu. Principles of three dimensional imaging in confocal microscopes. World Scientific, 1996. 1, 2 +[24] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3d imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7918, 2019. 5 +[25] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6770-6779, 2019. 5 +[26] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G Narasimhan. Structured light 3d scanning in the presence of global illumination. In CVPR 2011, pages 713-720. IEEE, 2011. 2 +[27] Kristofer Henderson, Xiaomeng Liu, Justin Folden, Brevin Tilmon, Suren Jayasuriya, and Sanjeev Koppal. Design and calibration of a fast flying-dot projector for dynamic light transport acquisition. IEEE Transactions on Computational Imaging, 6:529-543, 2020. 1 +[28] Sven TS Holmström, Utku Baran, and Hakan Urey. Mems laser scanners: a review. Journal of Microelectromechanical Systems, 23(2):259-275, 2014. 1 +[29] Ching-Pai Hsu, Boda Li, Braulio Solano-Rivas, Amar R Gohil, Pak Hung Chan, Andrew D Moore, and Valentina + +Donzella. A review and perspective on optical phased array for automotive lidar. IEEE Journal of Selected Topics in Quantum Electronics, 27(1):1-16, 2020. 3 +[30] Dawoon Jeong, Hansol Jang, Min Uk Jung, and Chang-Seok Kim. Angular resolution variable fmcw lidar with acoustooptic deflector. In Imaging Systems and Applications, pages ITh4D-3. Optica Publishing Group, 2022. 3, 8 +[31] S Kang, E Dotsenko, D Amrhein, C Theriault, and Craig B Arnold. Ultra-high-speed variable focus optics for novel applications in advanced imaging. In Photonic Instrumentation Engineering V, volume 10539, page 1053902. SPIE, 2018. 3 +[32] SeungYeon Kang, Martí Duocastella, and Craig B Arnold. Variable optical elements for fast focus control. Nature Photonics, 14(9):533-542, 2020. 3 +[33] Doycho Karagyozov, Mirna Mihovilovic Skanata, Amanda Lesar, and Marc Gershow. Recording neural activity in unrestrained animals with three-dimensional tracking two-photon microscopy. Cell reports, 25(5):1371-1383, 2018. 3 +[34] Yasin Karimi, Matteo Giuseppe Scopelliti, Ninh Do, Mohammad-Reza Alam, and Maysamreza Chamanzar. In situ 3d reconfigurable ultrasonically sculpted optical beam paths. Optics express, 27(5):7249-7265, 2019. 3 +[35] Byung-Sub Kim, Steve Gibson, and Tsu-Chin Tsao. Adaptive control of a tilt mirror for laser beam steering. In Proceedings of the 2004 American Control Conference, volume 4, pages 3417–3421. IEEE, 2004. 2 +[36] Lingjie Kong, Jianyong Tang, Justin P Little, Yang Yu, Tim Lammermann, Charles P Lin, Ronald N Germain, and Meng Cui. Continuous volumetric imaging via an optical phase-locked ultrasound lens. Nature methods, 12(8):759-762, 2015. 3 +[37] Hiroyuki Kubo, Suren Jayasuriya, Takafumi Iwaguchi, Takuya Funatomi, Yasuhiro Mukaigawa, and Srinivasa G Narasimhan. Programmable non-epipolar indirect light transport: Capture and analysis. IEEE Transactions on Visualization and Computer Graphics, 27(4):2421-2436, 2019. 2 +[38] Nanxi Li, Chong Pei Ho, Jin Xue, Leh Woon Lim, Guanyu Chen, Yuan Hsing Fu, and Lennon Yao Ting Lee. A progress review on solid-state lidar and nanophotonics-based lidar sensors. *Laser & Photonics Reviews*, page 2100511, 2022. 1 +[39] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fkmigration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2, 6 +[40] Qiyu Liu, Huan Li, and Mo Li. Electromechanical Brillouin scattering in integrated optomechanical waveguides. Optica, 6(6):778-785, 2019. 3 +[41] Xiaochun Liu, Ibón Guillén, Marco La Manna, Ji Hyun Nam, Syed Azer Reza, Toan Huu Le, Adrian Jarabo, Diego Gutierrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572(7771):620-623, 2019. 2, 6 +[42] Xiaomeng Liu, Kristofer Henderson, Joshua Rego, Suren Jayasuriya, and Sanjeev Koppal. Dense lissajous sampling and interpolation for dynamic light-transport. Optics Express, 29(12):18362–18381, 2021. 1 + +[43] Julio Marco, Adrian Jarabo, Ji Hyun Nam, Xiaochun Liu, Miguel Ángel Cosculluela, Andreas Velten, and Diego Gutierrez. Virtual light transport matrices for non-line-of-sight imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2440-2449, 2021. 6 +[44] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2 +[45] Aongus McCarthy, Nils J Krichel, Nathan R Gemmell, Ximing Ren, Michael G Tanner, Sander N Dorenbos, Val Zwiller, Robert H Hadfield, and Gerald S Buller. Kilometer-range, high resolution depth imaging via 1560 nm wavelength single-photon detection. Optics express, 21(7):8904-8915, 2013. 2 +[46] Srinivasa Narasimhan, Supreeth Achar, Kiriakos Kutulakos, Joe Bartels, and William Whittaker. Method for epipolar time of flight imaging, Mar. 19 2020. US Patent App. 16/468,617. 2 +[47] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 1, 2 +[48] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual coding to probe light transport. ACM Trans. Graph., 31(4):39-1, 2012. 2 +[49] Matthew O'Toole, David B Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555(7696):338-341, 2018. 2, 6 +[50] James Pawley. Handbook of biological confocal microscopy, volume 236. Springer Science & Business Media, 2006. 1, 2 +[51] Adithya Pediredla, Akshit Dave, and Ashok Veeraraghavan. Snlos: Non-line-of-sight scanning through temporal focusing. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2019. 2 +[52] Adithya Pediredla, Matteo Giuseppe Scopelliti, Srinivasa Narasimhan, Maysamreza Chamanzar, and Ioannis Gkioulekas. Optimized virtual optical waveguides enhance light throughput in scattering media. preprint, 2021. 8 +[53] Adithya Kumar Pediredla, Mauro Buttafava, Alberto Tosi, Oliver Cossairt, and Ashok Veeraraghavan. Reconstructing rooms using photon echoes: A plane based model and reconstruction algorithm for looking around the corner. In 2017 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2017. 6 +[54] Adithya K Pediredla, Aswin C Sankaranarayanan, Mauro Buttafava, Alberto Tosi, and Ashok Veeraraghavan. Signal processing based pile-up compensation for gated single-photon avalanche diodes. arXiv preprint arXiv:1806.07437, 2018. 6 +[55] Adithya Kumar Pediredla, Shizheng Zhang, Ben Avants, Fan Ye, Shin Nagayama, Ziying Chen, Caleb Kemere, Jacob T Robinson, and Ashok Veeraraghavan. Deep imaging in scattering media with selective plane illumination microscopy. Journal of biomedical optics, 21(12):126009, 2016. 1, 2 +[56] Kurt E Petersen. Silicon torsional scanning mirror. IBM Journal of Research and Development, 24(5):631-637, 1980. 1 + +[57] Ryan Po, Adithya Pediredla, and Ioannis Gkioulekas. Adaptive gating for single-photon 3d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16354–16363, 2022. 5, 6 +[58] Christopher V Poulton, Matthew J Byrd, Benjamin Moss, Erman Timurdogan, Ron Millman, and Michael R Watts. 8192-element optical phased array with $100^{\circ}$ steering range and flip-chip cmos. In CLEO: QELS_Fundamental Science, pages JTh4A-3. Optical Society of America, 2020. 3 +[59] Joshua Rapp, Yanting Ma, Robin MA Dawson, and Vivek K Goyal. Dead time compensation for high-flux ranging. IEEE Transactions on Signal Processing, 67(13):3471-3486, 2019. 6 +[60] Joshua Rapp, Yanting Ma, Robin MA Dawson, and Vivek K Goyal. High-flux single-photon lidar. Optica, 8(1):30-39, 2021. 6 +[61] Joshua Rodriguez, Braden Smith, Brandon Hellman, Adley Gin, Alonzo Espinoza, and Yuzuru Takashima. Multi-beam and single-chip lidar with discrete beam steering by digital micromirror device. In Physics and Simulation of Optoelectronic Devices XXVI, volume 10526, pages 89-94. SPIE, 2018. 3 +[62] Patrick S Salter, Z Iqbal, and Martin J Booth. Analysis of the three-dimensional focal positioning capability of adaptive optic elements. International Journal of Optomechatronics, 7(1):1-14, 2013. 3 +[63] Matteo Giuseppe Scopelliti and Maysamreza Chamanzar. Ultrasonically sculpted virtual relay lens for in situ microimaging. Light: Science & Applications, 8(1):1-15, 2019. 1, 3 +[64] Xiaobing Shang, Jin-Yi Tan, Oliver Willekens, Jelle De Smet, Pankaj Joshi, Dieter Cuypers, Esma Islamaj, Jeroen Beeckman, Kristiaan Neyts, Michael Vervaekte, et al. Electrically controllable liquid crystal component for efficient light steering. IEEE Photonics Journal, 7(2):1-13, 2015. 3 +[65] Braden Smith, Brandon Hellman, Adley Gin, Alonzo Espinoza, and Yuzuru Takashima. Single chip lidar with discrete beam steering by digital micromirror device. Optics Express, 25(13):14732-14745, 2017. 3 +[66] Qi Wang Song, Xu-Ming Wang, Rebecca Bussjager, and Joseph Osman. Electro-optic beam-steering device based on a lanthanum-modified lead zirconate titanate ceramic wafer. Applied optics, 35(17):3155-3162, 1996. 3 +[67] Steven J Spector. Review of lens-assisted beam steering methods. Journal of Optical Microsystems, 2(1):011003, 2022. 3 +[68] Ernst HK Stelzer, Frederic Strobl, Bo-Jui Chang, Friedrich Preusser, Stephan Preibisch, Katie McDole, and Reto Fiolka. Light sheet fluorescence microscopy. Nature Reviews Methods Primers, 1(1):1-25, 2021. 1, 2 +[69] Lin Sun, Jin-ha Kim, Chiou-hung Jang, Dechang An, Xue-jun Lu, Qingjun Zhou, John Martin Taboada, Ray T Chen, Jeffery J Maki, Suning Tang, et al. Polymeric waveguide prism-based electro-optic beam deflector. Optical Engineering, 40(7):1217-1222, 2001. 3 +[70] Varun Sundar, Sizhuo Ma, Aswin C Sankaranarayanan, and Mohit Gupta. Single-photon structured light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17865-17875, 2022. 2 + +[71] Opsys Tech. Opsys tech: The only scanning lidar with absolutely no moving parts. https://www.opsys-tech.com/post/opsys-tech-the-only-scanning-lidar-with-absolutely-no-moving-parts. Accessed: 2022-10-20. 3 +[72] Chen S Tsai. Guided-wave acousto-optics: interactions, devices, and applications, volume 23. Springer Science & Business Media, 2013. 3 +[73] Tomoki Ueda, Hiroyuki Kubo, Suren Jayasuriya, Takuya Funatomi, and Yasuhiro Mukaigawa. Slope disparity gating using a synchronized projector-camera system. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-9. IEEE, 2019. 2 +[74] Federica Villa, Fabio Severini, Francesca Madonini, and Franco Zappa. Spads and sipms arrays for long-range high-speed light detection and ranging (lidar). Sensors, 21(11):3839, 2021. 1 +[75] Dingkang Wang, Connor Watkins, and Huikai Xie. Mems mirrors for lidar: a review. Micromachines, 11(5):456, 2020. 1 +[76] Jian Wang, Joseph Bartels, William Whittaker, Aswin C Sankaranarayanan, and Srinivasa G Narasimhan. Programmable triangulation light curtains. In Proceedings of the European Conference on Computer Vision (ECCV), pages 19-34, 2018. 2 +[77] Xiangru Wang, Liang Wu, Caidong Xiong, Man Li, Qinggui Tan, Jiyang Shang, Shuanghong Wu, and Qi Qiu. Agile laser beam deflection with high steering precision and angular resolution using liquid crystal optical phased array. IEEE Transactions on Nanotechnology, 17(1):26-28, 2016. 3 +[78] Zihan Wang, Jie Cao, Qun Hao, Fanghua Zhang, Yang Cheng, and Xianyue Kong. Super-resolution imaging and field of view extension using a single camera with risley prisms. Review of Scientific Instruments, 90(3):033701, 2019. 2 +[79] RM Waxler and CE Weir. Effect of pressure and temperature on the refractive indices of benzene, carbon tetrachloride, and water. Journal of research of the National Bureau of Standards. Section A, Physics and chemistry, 67(2):163, 1963. 1 +[80] Ming-Tzo Wei, Shana Elbaum-Garfinkle, Alex S Holehouse, Carlos Chih-Hsiung Chen, Marina Feric, Craig B Arnold, Rodney D Priestley, Rohit V Pappu, and Clifford P Brangwynne. Phase behaviour of disordered proteins underlying low density and high permeability of liquid organelles. Nature chemistry, 9(11):1118-1125, 2017. 3 +[81] Shumian Xin, Sotiris Nousias, Kiriakos N Kutulakos, Aswin C Sankaranarayanan, Srinivasa G Narasimhan, and Ioannis Gkioulekas. A theory of fermat paths for non-line-of-sight shape reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6800-6809, 2019. 2 +[82] Chris Xu, Warren Zipfel, Jason B Shear, Rebecca M Williams, and Watt W Webb. Multiphoton fluorescence excitation: new spectral windows for biological nonlinear microscopy. Proceedings of the National Academy of Sciences, 93(20):10763-10768, 1996. 1, 2 + +[83] Jiawen Xu and Jiong Tang. Tunable prism based on piezoelectric metamaterial for acoustic beam steering. Applied Physics Letters, 110(18):181902, 2017. 3 +[84] Donghai Yang, Yifan Liu, Qingjiu Chen, Meng Chen, Shaodong Zhan, Nim-kwan Cheung, Ho-Yin Chan, Zhidong Wang, and Wen Jung Li. Development of the high angular resolution $360^{\circ}$ lidar based on scanning mem's mirror. Scientific Reports, 13(1):1540, 2023. 2 +[85] Zejie Yu and Xiankai Sun. Gigahertz acousto-optic modulation and frequency shifting on etchless lithium niobate integrated platform. ACS Photonics, 8(3):798-803, 2021. 3 +[86] Mantas Žurauskas, Oliver Barnstedt, Maria Frade-Rodriguez, Scott Waddell, and Martin J Booth. Rapid adaptive remote focusing microscope for sensing of volumetric neural activity. Biomedical optics express, 8(10):4369-4379, 2017. 3 \ No newline at end of file diff --git a/2023/Megahertz Light Steering Without Moving Parts/images.zip b/2023/Megahertz Light Steering Without Moving Parts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..32a3fd30357fb1607511c86f822bd7568455bea7 --- /dev/null +++ b/2023/Megahertz Light Steering Without Moving Parts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89fc3672ba59b8c9b0854aa49671d80b140ec402cc2db557a6967420b2361ac6 +size 507848 diff --git a/2023/Megahertz Light Steering Without Moving Parts/layout.json b/2023/Megahertz Light Steering Without Moving Parts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..06afa723e505e7f6e9fd9bdc241a36efec2e9a71 --- /dev/null +++ b/2023/Megahertz Light Steering Without Moving Parts/layout.json @@ -0,0 +1,10235 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "type": "text", + "content": "Megahertz Light Steering without Moving Parts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "spans": [ + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "text", + "content": "Adithya Pediredla" + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "text", + "content": ", Srinivasa G. Narasimhan" + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "text", + "content": ", Maysamreza Chamanzar" + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "text", + "content": ", Ioannis Gkioulekas" + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 124, + 527, + 153 + ], + "type": "text", + "content": ", Dartmouth College, Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 180, + 192, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 180, + 192, + 193 + ], + "spans": [ + { + "bbox": [ + 143, + 180, + 192, + 193 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 200, + 290, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 290, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 290, + 475 + ], + "type": "text", + "content": "We introduce a light steering technology that operates at megahertz frequencies, has no moving parts, and costs less than a hundred dollars. Our technology can benefit many projector and imaging systems that critically rely on high-speed, reliable, low-cost, and wavelength-independent light steering, including laser scanning projectors, LiDAR sensors, and fluorescence microscopes. Our technology uses ultrasound waves to generate a spatiotemporally-varying refractive index field inside a compressible medium, such as water, turning the medium into a dynamic traveling lens. By controlling the electrical input of the ultrasound transducers that generate the waves, we can change the lens, and thus steer light, at the speed of sound (1.5 km/s in water). We build a physical prototype of this technology, use it to realize different scanning techniques at megahertz rates (three orders of magnitude faster than commercial alternatives such as galvo mirror scanners), and demonstrate proof-of-concept projector and LiDAR applications. To encourage further innovation towards this new technology, we derive theory for its fundamental limits and develop a physically-accurate simulator for virtual design. Our technology offers a promising solution for achieving high-speed and low-cost light steering in a variety of applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 480, + 128, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 128, + 492 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 128, + 492 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "Many imaging systems rely on the ability to steer light, either as it leaves a source or as it reaches a sensor. Examples include laser scanning projectors [28, 56], LiDAR depth sensors [38, 74, 75], and microscopy techniques (confocal microscopy [23, 50], light-sheet microscopy [55, 68], multiphoton microscopy [17, 82]). Compared to full-field lighting and imaging, light steering systems help improve light efficiency [47], counter indirect illumination [27, 42], and enhance illumination and imaging contrast [9, 55]. However, these advantages come at the cost of reduced acquisition speed, bulky moving hardware, and motion artifacts. To alleviate these costs, we introduce a new light steering technology that, through the use of ultrasonic sculpting, makes it possible to scan light both transversally and axially at megahertz (MHz) rates. Additionally, our technology achieves these high scanning rates without any moving parts. Lastly, prototypes of our technology cost no more than a hundred dollars. Altogether, these characteristics represent significant" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 182, + 545, + 294 + ], + "blocks": [ + { + "bbox": [ + 307, + 182, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 307, + 182, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 545, + 294 + ], + "type": "image", + "image_path": "98bdc621d43009f9672422cd7f71f230268f5498cc29391576d98d6a0f44e677.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 294, + 545, + 407 + ], + "blocks": [ + { + "bbox": [ + 307, + 294, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 407 + ], + "type": "image", + "image_path": "f6ae1c32c570040b3786380601a27bba3c646318667c13fbe932773cae2d1cff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 408, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 485 + ], + "type": "text", + "content": "Figure 1. Megahertz light steering. (a) Light steering systems in LiDAR systems and laser projectors have moving mechanical components, limiting them to kHz scanning rates. (b) Our technology uses the acousto-optic effect to enable MHz light steering without moving parts. The insets show that, before galvo mirrors could scan even a few points, our prototype scanned a thousand points to project the letter \"A\" on the wall." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 491, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 547, + 502 + ], + "type": "text", + "content": "advances over previous light steering technologies (Table 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 503, + 547, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 547, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 547, + 694 + ], + "type": "text", + "content": "Our technology uses the acousto-optic effect to turn a transparent medium, such as water, into a programmable optic that steers an incident light beam. Sound is a pressure wave that travels inside a medium by compressing and rarefying it, spatiotemporally changing the medium density. In turn, this changes the refractive index of the medium, which is proportional to the density [63,79]. We design the pressure profile of the sound wave so that, at any time instant, the spatially-varying refractive index makes the medium behave as a periodic set of virtual gradient-index (GRIN) lenses, each with an aperture equal to the sound wavelength. The GRIN lenses bend light beams incident on the medium, with the GRIN profile determining the beam trajectory. These lenses travel at the speed of sound (1.5 km/s in water) and are reconfigurable at MHz frequencies, allowing us to steer light faster than mechanical devices. To enable flexible steering" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 317, + 702, + 541, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 541, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 541, + 712 + ], + "type": "text", + "content": "This is different from Bragg's diffraction in acousto-optic deflectors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "content": "patterns, we combine this optic with a pulsed laser with a programmable pulse rate. By synchronizing the laser source with the sound waveform, and modulating the phase of the sound waveform, we control both the speed of beam steering and the location of the beam." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 263 + ], + "type": "text", + "content": "In Sections 3 and 4, we explain the physical and mathematical details of our technology. We introduce a new design that uses two linear transducers to generate traveling acoustic waves, and discuss how different synchronization choices between ultrasound and pulsed laser result in different scanning patterns. To facilitate the exploration of design parameters (ultrasound speed and frequency, laser frequency) and configurations (transducer geometry), we also develop a physics-based simulator for our technology. In Section 6, we also discuss the fundamental limits of our technology due to diffraction and the uncertainty principle in wave physics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 264, + 288, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 288, + 443 + ], + "type": "text", + "content": "In Section 5, we experimentally demonstrate these fast programmable light steering techniques for various applications. In particular, in Section 5.2, we demonstrate an arbitrary point projector that can scan arbitrary and programmable light patterns. Compared to raster scanning projectors, which can project billions of points per second in a grid pattern but only a few thousand arbitrary points per second, our prototype can project a million arbitrary points per second, an acceleration by three orders of magnitude. In Section 5.3, we demonstrate a LiDAR prototype that combines our light steering technology with a single-photon avalanche diode (SPAD). We show 3D scans of " + }, + { + "bbox": [ + 46, + 264, + 288, + 443 + ], + "type": "inline_equation", + "content": "100 \\times 100" + }, + { + "bbox": [ + 46, + 264, + 288, + 443 + ], + "type": "text", + "content": " resolution at 5000 frames per second (50 million points per second) with a single-pixel SPAD, which is not feasible with scanning galvo mirrors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 444, + 228, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 228, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 228, + 454 + ], + "type": "text", + "content": "Contributions. Our main contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 288, + 598 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 456, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 287, + 490 + ], + "type": "text", + "content": "1. A new light steering technology based on the acoustooptic effect that is three orders of magnitude faster than state-of-the-art mechanical steering technologies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 288, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 288, + 514 + ], + "type": "text", + "content": "2. A new hardware design with planar transducers generating of traveling waves." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 288, + 539 + ], + "type": "text", + "content": "3. An experimental prototype demonstrating ultrafast arbitrary point projection and LiDAR scanning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 539, + 288, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 288, + 563 + ], + "type": "text", + "content": "4. A physics-based renderer to simulate digital twins of our prototype and evaluate different designs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 288, + 598 + ], + "type": "text", + "content": "5. The derivation of limits due to fundamental restrictions from wave physics (diffraction limit, scanning speed vs. aperture tradeoff, and uncertainty principle)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 599, + 287, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 623 + ], + "type": "text", + "content": "We provide our open-source simulator, data, and additional details in the supplement and project website.2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 623, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 288, + 695 + ], + "type": "text", + "content": "Limitations. Our prototype has a diffraction-limited point spread function (PSF) with a large spatial extent and a “+” shape, due to the use of two linear transducers that create a rectangular aperture. This limits spatial resolution, and introduces structured blur artifacts. These limitations are not fundamental to our core technology, and can be overcome" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 319, + 138, + 531, + 225 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 137 + ], + "type": "text", + "content": "Table 1. Comparison of light steering technologies. MEMS is microelectromechanical systems, OPA is optical phased arrays, BW is optical bandwidth. For the arrows, red is bad, green is good, down is low, and up is high, and more arrows imply a bigger effect. Our method is superior in terms of cost, speed and supported bandwidth, with no moving parts or fabrication." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 138, + 531, + 225 + ], + "lines": [ + { + "bbox": [ + 319, + 138, + 531, + 225 + ], + "spans": [ + { + "bbox": [ + 319, + 138, + 531, + 225 + ], + "type": "table", + "html": "
Tech.speedfab.BWcostmoving
galvo↓↓×↑↓✓
liquid lenses↓↓↓×↑↓×
MEMS↑✓↑↑✓
OPA↑↑↑✓↓↑↑×
ours↑↑×↑↓↓↓×
", + "image_path": "5b6152def0fd06b2d6b008faedf5694165160ebbb7a949e662d44b8fd2b09c23.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 232, + 539, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 539, + 244 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 539, + 244 + ], + "type": "text", + "content": "with improved designs and better engineering (Section 6)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 253, + 390, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 390, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 390, + 264 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 270, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 547, + 425 + ], + "type": "text", + "content": "Light steering in imaging systems. Light steering is a core component in scanning-based active imaging systems. For example, fluorescence microscopy techniques such as confocal microscopy [23, 50], multiphoton microscopy [17, 82], light sheet microscopy [55, 68], and superresolution microscopy [7], use scanning to decrease scattered light and improve light efficiency and imaging contrast. Another example is LiDAR sensors [84] found in commercial applications, such as autonomous cars. In these sensors, scanning decreases multipath interference and helps reduce hardware cost, removing the need for two-dimensional LiDAR arrays. Our technology improves the speed and reliability of scanning-based LiDAR, while further reducing the cost." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 426, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 581 + ], + "type": "text", + "content": "Light steering is also used in scanning-based laser projectors, to achieve high light efficiency and contrast by illuminating only where necessary. Laser projectors typically perform a 2D raster scan the field of view, with a fast and a slow scanning axis. The speed of the fast scanning axis limits the frame rate to the order of hundreds of kHz. In contrast, our technique can operate at tens of MHz, improving the raster scan rate by two orders of magnitude. Additionally, raster scanners cannot project arbitrary point sequences at a fast rate, and are limited to the frame rate of the projector (60-120 points per second) even when the projection pattern is very sparse. We demonstrate the projection of a million arbitrary points per second, four orders of magnitude faster." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 677 + ], + "type": "text", + "content": "Computational imaging techniques that use laser projectors or LiDAR are generally also scanning-based. Examples include structured light [26, 70], light-transport probing [47, 48], motion contrast 3D [44], epipolar gating [3, 46], light curtains [11, 76], slope-disparity gating [12, 37, 73], and non-line-of-sight imaging [39, 41, 49, 51, 81]. As a consequence, all these techniques can become orders of magnitude faster if combined with our light steering device." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Light steering technologies We can distinguish between mechanical and non-mechanical light steering techniques. The former include rotating prisms [2, 78], mirrors [35, 45]," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 702, + 280, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 280, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 280, + 713 + ], + "type": "text", + "content": "2https://imaging.cs.cmu.edu/ultrafast_steering" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 750, + 308, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 750, + 308, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 308, + 757 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 287, + 135 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 287, + 135 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 287, + 135 + ], + "type": "image", + "image_path": "4ab681afbc93c5a77a6cc1d26875f8d3c39cd192483aebc184c27ba38f91e335.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 142, + 288, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 142, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 142, + 288, + 198 + ], + "type": "text", + "content": "Figure 2. Comparison to previous work. (a) Previous ultrasonically sculpted lenses [10, 32] change the focal length at high speed (b) Our lens can change both focal length (transparencies) and the spatial location (colors) at high speed. We achieve this using traveling waves and synchronizing sound and light signals." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 222, + 289, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 289, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 289, + 330 + ], + "type": "text", + "content": "and digital micromechanical systems (MEMS) [61, 65]. The latter include acousto-optic (AO) [20, 83] or electro-optic (EO) [69] deflectors, liquid crystal devices [64, 77], and optical phased arrays (OPA) [29, 58]. Table 1 compares these techniques to our MHz steering technology: Mechanical techniques are slow due to the need to physically move optical elements. AO and EO deflectors operate at kHz rates. Liquid lenses and crystals typically have a long settling time, making them the slowest among these techniques." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 410 + ], + "type": "text", + "content": "OPAs [29, 58] are solid-state on-chip devices that can steer light at even GHz rates. However, they require on-chip microscopic coherent laser generators [71] and cannot easily be coupled to external lasers. They are also expensive to fabricate and currently limited to low-resolution angles. Cheng et al. [15] and Spector [67] have detailed reviews." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 419, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 539 + ], + "type": "text", + "content": "Acousto-optic devices. Many acousto-optic devices are commercially available, including tunable filters [21], modulators [5], frequency shifters [85], and deflectors [30]. Among these, the acousto-optic deflector can serve as a light steering device. However, acousto-optic deflectors use a different physical phenomenon, Bragg's diffraction [72], where beam deviation is proportional to the acoustic wave frequency, and are thus orders of magnitude slower than our technology. Electro-optic deflectors [40, 66] operate on a similar phenomenon and are thus similarly slow." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 547, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 289, + 715 + ], + "type": "text", + "content": "Tunable acoustic gradient-index (TAG) lenses [16,32] and ultrasonically-sculpted virtual optical waveguides [10, 34] use the same physical principles as our technology. TAG lenses can change the focus depth of an incident beam at " + }, + { + "bbox": [ + 46, + 547, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 547, + 289, + 715 + ], + "type": "text", + "content": " rates; however, unlike our technology, they cannot steer it in the transverse axis (Figure 2). Following their recent commercialization, TAG lenses fostered innovation in scientific and application fields such as laser micromachining [14, 18], three-dimensional biomedical imaging [36, 80], microscopy [13, 19, 33], optical coherence tomography [22], high-throughput industrial inspection [31], and adaptive optics [62, 86]. We believe that the additional transverse steering capabilities from our technology will similarly help stimulate significant further innovation." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 350, + 83, + 501, + 171 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 523, + 82 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 523, + 82 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 523, + 82 + ], + "type": "text", + "content": "Table 2. Notation and parameters we use in the paper." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 350, + 83, + 501, + 171 + ], + "lines": [ + { + "bbox": [ + 350, + 83, + 501, + 171 + ], + "spans": [ + { + "bbox": [ + 350, + 83, + 501, + 171 + ], + "type": "table", + "html": "
quantitysymbol
speedCUS
wavelengthλus
frequencyfus
angular frequencyωus = 2πfus
wavenumberkus = 2πλus-1
timeperiodTus = fus-1
", + "image_path": "3b2567a658298eb0b54fe18492af84e3643ca8c7f38da82284254c63a602ab2e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 175, + 503, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 175, + 503, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 503, + 189 + ], + "type": "text", + "content": "3. Traveling-wave acousto-optic lenses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 192, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 547, + 251 + ], + "type": "text", + "content": "We generate ultrasonic waves inside a transparent medium, such as water, by submerging inside it a planar transducer that we drive with a single harmonic voltage (Figure 3(a)). The resulting pressure " + }, + { + "bbox": [ + 304, + 192, + 547, + 251 + ], + "type": "inline_equation", + "content": "P(x,t)" + }, + { + "bbox": [ + 304, + 192, + 547, + 251 + ], + "type": "text", + "content": " inside the medium (Figure 3(b)) equals" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 256, + 545, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 256, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 350, + 256, + 545, + 269 + ], + "type": "interline_equation", + "content": "P (x, t) = P _ {o} + P _ {s} \\cos \\left(k _ {\\mathrm {u s}} x - \\omega_ {\\mathrm {u s}} t\\right), \\tag {1}", + "image_path": "9090448048b8ca40965a8dfa7d638721c650af78a64ea57d5abe7ce9edda55e7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "text", + "content": "where: " + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "text", + "content": " is time, " + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "text", + "content": " is distance normal to the transducer plane—the pressure is independent of " + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "text", + "content": " coordinates—" + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "inline_equation", + "content": "P_{o}" + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "text", + "content": " is the medium pressure without ultrasound, " + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 304, + 273, + 547, + 332 + ], + "type": "text", + "content": " is proportional to the transducer voltage amplitude, and the remaining parameters are in Table 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 333, + 547, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 547, + 357 + ], + "type": "text", + "content": "The refractive index " + }, + { + "bbox": [ + 304, + 333, + 547, + 357 + ], + "type": "inline_equation", + "content": "n(x,t)" + }, + { + "bbox": [ + 304, + 333, + 547, + 357 + ], + "type": "text", + "content": " of the medium changes proportionally to the pressure (Figure 3(c)):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 362, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 362, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 351, + 362, + 545, + 374 + ], + "type": "interline_equation", + "content": "n (x, t) = n _ {o} + n _ {s} \\cos \\left(k _ {\\mathrm {u s}} x - \\omega_ {\\mathrm {u s}} t\\right), \\tag {2}", + "image_path": "5061cbed42271b48f9e39060218082f073e0553fdab5633090bacb246d0779d9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "text", + "content": "where: " + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "inline_equation", + "content": "n_{o}" + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "text", + "content": " is the refractive index of the medium without the transducer, " + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "inline_equation", + "content": "n_{s} = kP_{s}" + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "text", + "content": " is an empirical coefficient (" + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "inline_equation", + "content": "k = 1.402 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "text", + "content": " bar" + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "inline_equation", + "content": "^{-1}" + }, + { + "bbox": [ + 304, + 379, + 545, + 415 + ], + "type": "text", + "content": " for water) [63]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "type": "text", + "content": "At time " + }, + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "type": "text", + "content": ", the convex lobes of this refractive index profile (i.e., regions " + }, + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "type": "inline_equation", + "content": "x = [l\\lambda_{\\mathrm{us}} - \\lambda_{\\mathrm{us}} / 2, l\\lambda_{\\mathrm{us}} + \\lambda_{\\mathrm{us}} / 2]" + }, + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "type": "inline_equation", + "content": "l \\in \\mathbb{Z}" + }, + { + "bbox": [ + 304, + 415, + 547, + 486 + ], + "type": "text", + "content": ") act as GRIN lenses. Each of these lenses focuses light rays traveling parallel to the transducer onto a line (Figure 3(d)). We can change the lens focal length and aperture by varying the amplitude and frequency of the transducer voltage." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "type": "text", + "content": "Traveling lenses. The pressure wave and refractive index profile propagate along the " + }, + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "type": "text", + "content": "-direction. Thus, the convex lobes vary as: " + }, + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "type": "inline_equation", + "content": "x(t) = [c_{\\mathrm{us}}t + l\\lambda_{\\mathrm{us}} - \\lambda_{\\mathrm{us}} / 2, c_{\\mathrm{us}}t + l\\lambda_{\\mathrm{us}} + \\lambda_{\\mathrm{us}} / 2]" + }, + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "type": "inline_equation", + "content": "l \\in \\mathbb{Z}" + }, + { + "bbox": [ + 304, + 487, + 547, + 558 + ], + "type": "text", + "content": ". As a result, ultrasonically-sculpted cylindrical GRIN lenses are dynamic, and focused lines travel normal to the transducer at the speed of ultrasound (Figure 4)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 565, + 425, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 565, + 425, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 565, + 425, + 578 + ], + "type": "text", + "content": "4. Scanning techniques" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": "Even though the ultrasonically-sculpted GRIN lenses travel at the speed of ultrasound, we cannot control the speed or location of the lens focus. To enable such control, we use a pulsed laser with the same repetition frequency as the ultrasound frequency, and programmable phase modulation for the transducer voltage." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Single transducer. For intuition, we first describe the case of a single transducer. Due to phase modulation, the pressure pattern from Equation (1) becomes " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "P(x,t) = P_{o} + P_{s}\\cos (k_{\\mathrm{us}}x - \\omega_{\\mathrm{us}}t - \\phi (t))" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". If the illumination is continuous, the position of the focused light is " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{n}(t) =" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 307, + 246 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 307, + 246 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 307, + 246 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 307, + 246 + ], + "type": "image", + "image_path": "40988d26d26223b134844c679c926106c1abb10a6215a4b23212ed4ac56a1262.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 248, + 140, + 258 + ], + "lines": [ + { + "bbox": [ + 107, + 248, + 140, + 258 + ], + "spans": [ + { + "bbox": [ + 107, + 248, + 140, + 258 + ], + "type": "text", + "content": "(a) setup" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 76, + 436, + 178 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 436, + 178 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 436, + 178 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 436, + 178 + ], + "type": "image", + "image_path": "51c2eb9d2618f335ca43a673e622f741703dabce171da9a6e98728fa14e01b06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 178, + 417, + 247 + ], + "blocks": [ + { + "bbox": [ + 315, + 178, + 417, + 247 + ], + "lines": [ + { + "bbox": [ + 315, + 178, + 417, + 247 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 417, + 247 + ], + "type": "image", + "image_path": "3636dae10e2b138887d1e277d34c70281560b43a100ba41f937d24e68096b914.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 249, + 398, + 258 + ], + "lines": [ + { + "bbox": [ + 327, + 249, + 398, + 258 + ], + "spans": [ + { + "bbox": [ + 327, + 249, + 398, + 258 + ], + "type": "text", + "content": "(c) refractive index" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 440, + 81, + 542, + 175 + ], + "blocks": [ + { + "bbox": [ + 440, + 81, + 542, + 175 + ], + "lines": [ + { + "bbox": [ + 440, + 81, + 542, + 175 + ], + "spans": [ + { + "bbox": [ + 440, + 81, + 542, + 175 + ], + "type": "image", + "image_path": "38c6a23d4b523f8f12b7d314651fce1898a145dbfb2eb39986b6bf5521e22f20.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 441, + 176, + 543, + 246 + ], + "blocks": [ + { + "bbox": [ + 441, + 176, + 543, + 246 + ], + "lines": [ + { + "bbox": [ + 441, + 176, + 543, + 246 + ], + "spans": [ + { + "bbox": [ + 441, + 176, + 543, + 246 + ], + "type": "image", + "image_path": "94e2c41db8f77d2a41076a6aa8a84b6ab379d8cefc8a2044e1ba1c43923e12dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 247, + 521, + 258 + ], + "lines": [ + { + "bbox": [ + 462, + 247, + 521, + 258 + ], + "spans": [ + { + "bbox": [ + 462, + 247, + 521, + 258 + ], + "type": "text", + "content": "(d) ray diagram" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 332, + 160, + 438 + ], + "blocks": [ + { + "bbox": [ + 46, + 259, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 547, + 326 + ], + "type": "text", + "content": "Figure 3. Ultrasonic sculpting. (a) By placing a planar piezoelectric transducer inside a compressible medium and vibrating the transducer with a sinusoidal voltage, we generate ultrasound inside the medium. (b) Sound is a pressure wave, and at any given time, the pressure inside the medium varies spatially based on the voltage waveform applied to the transducer. (c) The change in the refractive index of the medium is proportional to the pressure. Therefore, the refractive index of the medium also varies spatially, turning the medium into a GRIN lens. (d) Light rays that pass through this GRIN lens will curve continuously and focus on a set of lines. The focal length of these waveguides/lenses is a function of the voltage and frequency applied to the transducer." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 332, + 160, + 438 + ], + "lines": [ + { + "bbox": [ + 50, + 332, + 160, + 438 + ], + "spans": [ + { + "bbox": [ + 50, + 332, + 160, + 438 + ], + "type": "image", + "image_path": "1105b76db2e71f8f6798e2fc9445ec39b815f1d68e300f1f1a94092cf5cf495f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 438, + 129, + 448 + ], + "lines": [ + { + "bbox": [ + 75, + 438, + 129, + 448 + ], + "spans": [ + { + "bbox": [ + 75, + 438, + 129, + 448 + ], + "type": "text", + "content": "(a) ray diagrams" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 449, + 287, + 526 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 526 + ], + "type": "text", + "content": "Figure 4. Planar transducer steers the light. As sound propagates, the spatial refractive index pattern also propagates at the speed of sound. The focussed light, which is a set of lines, also travels normal to the transducer plane at the speed of sound. Apart from the ray diagrams, the images rendered with our simulator (submitted as supplementary) show that the planar transducer moves the focus region (in this case, a line) at the speed of sound." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 164, + 337, + 287, + 436 + ], + "blocks": [ + { + "bbox": [ + 164, + 337, + 287, + 436 + ], + "lines": [ + { + "bbox": [ + 164, + 337, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 164, + 337, + 287, + 436 + ], + "type": "image", + "image_path": "d7ddfa99370e77f8ceece793d423c2e77dc9eba3b02a24ac34496def9b8030aa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 439, + 244, + 449 + ], + "lines": [ + { + "bbox": [ + 208, + 439, + 244, + 449 + ], + "spans": [ + { + "bbox": [ + 208, + 439, + 244, + 449 + ], + "type": "text", + "content": "(b) images" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{us}}t - n\\lambda_{\\mathrm{us}} + \\frac{\\phi(t)}{2\\pi}\\lambda ;n\\in \\mathbb{Z}," + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "text", + "content": " which is a set of lines moving continuously at the speed of sound. By pulsing the illumination at ultrasound frequency, we get a set of lines flickering at the ultrasound frequency: " + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "inline_equation", + "content": "x_{n}(mT_{\\mathrm{us}}) = (m - n)\\lambda_{\\mathrm{us}} + \\frac{\\phi(mT_{\\mathrm{us}})}{2\\pi}\\lambda_{\\mathrm{us}}." + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "text", + "content": " If we modulate the phase linearly " + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "inline_equation", + "content": "(\\phi (t) = k\\omega_{\\mathrm{us}}t;k < 1)" + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "text", + "content": " ,we have " + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "inline_equation", + "content": "x(mT_{\\mathrm{us}}) = mk c_{\\mathrm{us}}T_{\\mathrm{us}}" + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "inline_equation", + "content": "m = n" + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "text", + "content": " (i.e., the line due to the same laser and ultrasound pulse). Therefore, the lines travel at a reduced speed " + }, + { + "bbox": [ + 46, + 533, + 289, + 632 + ], + "type": "inline_equation", + "content": "kc_{\\mathrm{us}}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 638, + 241, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 241, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 241, + 651 + ], + "type": "text", + "content": "4.1. Point scanning with two transducers" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "To focus light to a point, we use two orthogonal planar transducers (Figure 5). We describe two extensions of the above approach for focus point control: one for scanning arbitrary point locations at the ultrasound frequency (i.e., MHz), and another for raster scanning at the laser repetition" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 332, + 537, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 332, + 537, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 332, + 537, + 345 + ], + "type": "text", + "content": "frequency, which is higher than the ultrasound frequency." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "text", + "content": "Arbitrary point scanning. To scan arbitrary points " + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "inline_equation", + "content": "(x(mT_{\\mathrm{us}}),y(mT_{\\mathrm{us}}))" + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "text", + "content": " for each laser pulse " + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "text", + "content": ", we modulate the phases " + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\phi_x(t)" + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\phi_y(t)" + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "text", + "content": " of both transducers. The focus point location within the region " + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "inline_equation", + "content": "[0,\\lambda_{\\mathrm{us}}]\\times [0,\\lambda_{\\mathrm{us}}]" + }, + { + "bbox": [ + 305, + 350, + 545, + 398 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 351, + 402, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 402, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 351, + 402, + 545, + 415 + ], + "type": "interline_equation", + "content": "x (t) = \\left(\\phi_ {x} (t) / k _ {\\mathrm {u s}} + c _ {\\mathrm {u s}} t\\right) \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {3}", + "image_path": "9ff13734e2b86bae67b3dd47d50fc1666d1ec2e06f7a2813a2174f59ccbda4cb.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 352, + 417, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 417, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 352, + 417, + 545, + 430 + ], + "type": "interline_equation", + "content": "y (t) = \\left(\\phi_ {y} (t) / k _ {\\mathrm {u s}} + c _ {\\mathrm {u s}} t\\right) \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {4}", + "image_path": "686d4a22e9d3b558a3e2e6e344ee670d6ead0b596995fc82346146b32b9c6cde.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "content": "Figure 6 shows the refractive index and ray diagram for four sets of " + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "inline_equation", + "content": "\\phi_{x}" + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "inline_equation", + "content": "\\phi_{y}" + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "content": " values. To scan a set of points " + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "inline_equation", + "content": "(x(mT_{\\mathrm{us}}), y(mT_{\\mathrm{us}}))" + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "content": ", we compute the phases " + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "inline_equation", + "content": "(\\phi_{x}(mT_{\\mathrm{us}}), \\phi_{y}(mT_{\\mathrm{us}}))" + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "content": " using Equations (3)-(4), and interpolate to compute " + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "inline_equation", + "content": "(\\phi_{x}(t), \\phi_{y}(t))" + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "type": "text", + "content": "Raster scanning In theory, we could use arbitrary point scanning to raster scan a two-dimensional grid of points. In that case, the phase modulation for raster scanning would be linear, " + }, + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "type": "inline_equation", + "content": "(\\phi_x(t),\\phi_y(t)) = (k_x\\omega_{\\mathrm{us}}t,k_y\\omega_{\\mathrm{us}}t)" + }, + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "type": "inline_equation", + "content": "k_{x}" + }, + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "type": "inline_equation", + "content": "k_{y}" + }, + { + "bbox": [ + 304, + 499, + 545, + 583 + ], + "type": "text", + "content": " are phase modulation rates. The phase modulation rate for the faster axis would be equal to the product of the number of scan points and the modulation rate of the slower axis." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 588, + 547, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 547, + 636 + ], + "type": "text", + "content": "However, this approach would limit raster scanning frequency to the ultrasound frequency. If the laser repetition frequency is higher, we can scan more points by running the laser at its highest frequency. Their locations will be:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 359, + 640, + 545, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 640, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 359, + 640, + 545, + 664 + ], + "type": "interline_equation", + "content": "x \\left(m T _ {L}\\right) = \\left(m k _ {x}\\right) \\frac {\\lambda_ {\\mathrm {u s}}}{s} \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {5}", + "image_path": "928f3e74befc068c9595a73b149d665221ee2944d850997ecc7970d8d06d7bfc.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 359, + 665, + 545, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 665, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 359, + 665, + 545, + 687 + ], + "type": "interline_equation", + "content": "y \\left(m T _ {L}\\right) = \\left(m k _ {y}\\right) \\frac {\\lambda_ {\\mathrm {u s}}}{s} \\bmod \\lambda_ {\\mathrm {u s}}, \\tag {6}", + "image_path": "7f8a1f216c7243c82e5b4ca06e92db220abf7f2a4b1a191b770900f126cb5feb.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "inline_equation", + "content": "s = f_{L} / f_{\\mathrm{us}}" + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "text", + "content": " is the ratio of laser " + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "inline_equation", + "content": "(f_{L})" + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "text", + "content": " and ultrasound " + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{us}})" + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "text", + "content": " frequencies, and " + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "inline_equation", + "content": "T_{L} = 1 / f_{L}" + }, + { + "bbox": [ + 304, + 690, + 545, + 715 + ], + "type": "text", + "content": " is the inter-pulse time." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 170, + 202 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 170, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 170, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 170, + 202 + ], + "type": "image", + "image_path": "32959e457d339467566da9b521888dd6181d6aa93cc15a4f1c13cc68e7314ede.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "type": "text", + "content": "Figure 5. Two planar transducers for focusing light at a point. (a) We place two planar piezoelectric transducers orthogonal to each other inside a medium. We drive both transducers independently with a sinusoidal voltage. (b) The pressure wave inside the medium is a superposition of the pressure waves generated by the transducers. (c) The change in the refractive index is proportional to the net pressure. (d) Light rays from a wide beam focus on a set of points. We restrict the illumination beam size to focus light on a single point." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 76, + 296, + 202 + ], + "blocks": [ + { + "bbox": [ + 170, + 76, + 296, + 202 + ], + "lines": [ + { + "bbox": [ + 170, + 76, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 170, + 76, + 296, + 202 + ], + "type": "image", + "image_path": "5360dc2c60743b082c541688f52bdec0aa513a6bb66871da834d5aa8cba93455.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 301, + 76, + 431, + 202 + ], + "blocks": [ + { + "bbox": [ + 301, + 76, + 431, + 202 + ], + "lines": [ + { + "bbox": [ + 301, + 76, + 431, + 202 + ], + "spans": [ + { + "bbox": [ + 301, + 76, + 431, + 202 + ], + "type": "image", + "image_path": "3dd7296e25b0a72a7cc2ababa16f439758b45b2e5b9dd6f2d64208c6ab097911.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 434, + 83, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 434, + 83, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 434, + 83, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 434, + 83, + 545, + 202 + ], + "type": "image", + "image_path": "36c5c4cd976deb5cf9d74e7e0577039c46a24c13fc64b0c543ea225f03b874b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 253, + 274, + 415 + ], + "blocks": [ + { + "bbox": [ + 58, + 253, + 274, + 415 + ], + "lines": [ + { + "bbox": [ + 58, + 253, + 274, + 415 + ], + "spans": [ + { + "bbox": [ + 58, + 253, + 274, + 415 + ], + "type": "image", + "image_path": "699a4cdfd2aa2513ab725ab114b9627be0b13130f2f7a12123fbc751b40d1377.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 417, + 288, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 417, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 288, + 462 + ], + "type": "text", + "content": "Figure 6. Point steering. By controlling the phases of the sinusoidal voltages applied to the transducers, we control the location of the focus position. To continuously steer the focus location, we phase modulate the voltages applied to both transducers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 469, + 128, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 128, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 128, + 483 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 487, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 619 + ], + "type": "text", + "content": "We discuss an experimental prototype implementing our acousto-optic light steering technology, and combining it with a pulsed laser, single-pixel SPAD, and galvo mirrors (for comparison). We use this prototype to demonstrate projector and LiDAR applications. We compare our light steering system with commercially available galvo mirrors, to demonstrate the speed and the new capabilities our system enables. We keep the field-of-view and aperture same for both systems. This comparison is not the most favorable for galvo mirrors, as they typically have a larger field of view, but it is a fair one for evaluating the system's speed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 625, + 116, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 116, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 116, + 639 + ], + "type": "text", + "content": "5.1. Prototype" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Figure 7 shows our experimental prototype. We place two planar piezo transducers (P-25.40mm-25.40mm-2.10mm-880-WFB, APC International, Ltd) orthogonal to each other and at an inclination of " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " relative to an acrylic tank containing water, to minimize interference from interreflections. A signal generator (SDG6022x, Siglent Tech.) drives the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 255, + 547, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 410 + ], + "type": "text", + "content": "transducers via a power amplifier (ENI A300, Bell electronics). We colocate a pulsed laser (ALPHALAS GmbH, PICOPOWER-LD-510) and a gated SPAD (Microphoton Devices s.r.l.) using a beamsplitter, similar to previous techniques [24, 25, 57]. However, we do not place a lens in front of the SPAD, as the ultrasonically-sculpted lens focuses light from the object onto the SPAD. Instead, we place a lens in front of the laser to create a diverging ray that undergoes the same focusing by the ultrasonically-sculpted lens. We place an aperture after the ultrasonically-sculpted lens to limit the scanning area to only one ultrasonic period. A " + }, + { + "bbox": [ + 304, + 255, + 547, + 410 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 304, + 255, + 547, + 410 + ], + "type": "text", + "content": " mirror directs the beam to a pair of galvo mirrors (GVS-212, Thorlabs Inc.). We use the galvo mirrors only for comparisons." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 411, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 508 + ], + "type": "text", + "content": "To synchronize the transducers, laser, SPAD gate, and SPAD timing circuit, we use two signal generators and synchronize their clock and trigger signals. We use one signal generator to drive the transducers. The two channels of the second signal generator run the laser and the SPAD gate. We explain the SPAD timing synchronization details in Section 5.3. We provide more details about our prototype, including design and alignment, in the supplement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 515, + 449, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 449, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 449, + 528 + ], + "type": "text", + "content": "5.2. Arbitrary point projector" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 532, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 590 + ], + "type": "text", + "content": "We use the technique in Section 4.1 to compute the phases and synthesize the transducer voltage waveforms required to project an arbitrary target sequence of points. We drive the transducers with this waveform and the laser at the same frequency as ultrasound (1 MHz)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "text", + "content": "To project the same sequence of points with the galvo mirrors, we drive the laser and transducers at a fixed frequency without any phase modulation, which results in a single-point focus. We steer this point with the galvo mirrors to scan the same desired sequence of points. We drive the galvo mirrors at " + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "inline_equation", + "content": "1\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "inline_equation", + "content": "2\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "inline_equation", + "content": "5\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "text", + "content": " frequencies (points per second). The galvo mirrors are rated for " + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "inline_equation", + "content": "1\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "text", + "content": ", and driving it at frequencies higher than " + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "inline_equation", + "content": "5\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 592, + 547, + 699 + ], + "type": "text", + "content": " leads to higher motor current and failure of the fuse." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "We project the patterns on a white cardboard screen, and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 281, + 196 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 281, + 196 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 281, + 196 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 281, + 196 + ], + "type": "image", + "image_path": "a1b4d290c54087f5ebd83caa17f5ec71a646c59938edeb6dda6efb00e72d2acc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 209, + 280, + 334 + ], + "blocks": [ + { + "bbox": [ + 140, + 198, + 195, + 209 + ], + "lines": [ + { + "bbox": [ + 140, + 198, + 195, + 209 + ], + "spans": [ + { + "bbox": [ + 140, + 198, + 195, + 209 + ], + "type": "text", + "content": "(a) schematic" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 209, + 280, + 334 + ], + "lines": [ + { + "bbox": [ + 53, + 209, + 280, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 280, + 334 + ], + "type": "image", + "image_path": "3d4f1682f72860d93b0ca61c9e94fda7c4c556446b9db283b1ba02c8fb30cc7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 335, + 221, + 346 + ], + "lines": [ + { + "bbox": [ + 113, + 335, + 221, + 346 + ], + "spans": [ + { + "bbox": [ + 113, + 335, + 221, + 346 + ], + "type": "text", + "content": "(b) experimental prototype" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 346, + 289, + 490 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 289, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 289, + 490 + ], + "type": "text", + "content": "Figure 7. Hardware setup. We show the (a) schematic and (b) prototype built for showing the proof-of-concept applications and comparing them with a galvo mirror system. We diverge the laser beam, and the expanded beam is focused by the ultrasonically-sculpted refractive index. The beam passes through the galvo mirrors onto the scene. We steer the beam either with the ultrasonically-sculpted lens or the galvo mirrors, but not both, depending on the experiment. The reflected light from the object takes the same path back to the sensor. The SPAD sensor, which is colocated with the laser, does not have any optics in front of it other than the ultrasonically-sculpted lenses. This setup allows us to compare the scanning speed of our system with galvo mirrors, while keeping the aperture the same." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 288, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 584 + ], + "type": "text", + "content": "capture images with a camera (Allied Vision PRO-GT3400-09) for two exposures (1 ms and 50 ms) that we show in Figure 8. In this case, we are projecting 100 points that form the letter \"A\". At 1 ms exposure, the galvo mirrors only scan a few points, whereas our technique scans the entire shape ten times. At 100 ms or higher exposure, the galvo mirrors can project all the points without distortion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 589, + 104, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 104, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 104, + 600 + ], + "type": "text", + "content": "5.3. LiDAR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": "We use a gated SPAD for depth and transient measurements. The gate helps reject the backscattered photons from various optics. After gating, our system does not suffer from pile-up [54, 59, 60]. We use a signal generator to drive the SPAD gate instead of the picosecond delayer (PSD) common in SPAD-based LiDAR systems [8, 41, 49, 53]. Our approach is inexpensive and generates programmable delays at much higher resolution (" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "1\\mu s" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": ") than the PSD (" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "50\\mathrm{ns}" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "We run both the transducer and SPAD signal genera" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "tors in burst mode with the trigger running at " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "100\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " for synchronization. We use picoharp (Microphoton Devices s.r.l.) to measure the time-of-flight of the photons and to synchronize the scanning position. Several previous techniques [39,43,49,57] achieve this synchronization by sending a synchronization signal to picoharp, which reports a marker event when the synchronization signal is detected. Unfortunately, as our scanning runs at a much higher rate than galvo mirror scanning, we cannot use an external synchronization signal for picoharp due to the massive number of events that would be generated. Fortunately, the picoharp reports the pulse index for each laser pulse it detects. We use the pulse index to compute when the photon is fired, and use Equations (5)-(6) to compute the scan position." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": "We first use the pulsed raster scanning technique in Section 4.1 to raster scan the scene. We set " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "f_{L} = 50\\mathrm{MHz}" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{us}} = 1\\mathrm{MHz}" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "k_{x} = 1.0001" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "k_{y} = 1.01" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": ". These settings result in a spatial resolution of " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "100\\times 100" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": " and scan rate of 5000 frames per second. Figure 9 shows results for four scenes, each with two letters at various depths. Our single channel (single sensor and laser) system scans 50 million points per second. This is three orders of magnitude faster than the 128-channel high-end Velodyne VLS-128™ LiDAR [1], which scans 5 million points per second." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "type": "text", + "content": "We also use the arbitrary point scanning technique in Section 4.1 to selectively scan a few points in the field of view. An advantage of such an approach is the potential for adaptive scanning [6]. In Figure 10, we create a three-point scene and image it to compute their spatial locations. We then scan these three points using both galvo mirrors and our technique for " + }, + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "type": "inline_equation", + "content": "3\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "type": "text", + "content": " (1 ms per point). When scanning with the galvo mirrors, a lot of time is spent moving the galvo mirrors from one scan point to another. As a result, many pulses fired by the laser are wasted during the travel time of the galvo mirrors. By contrast, ultrasonic beam steering shifts the focus within the time period of the two laser photons (" + }, + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "type": "inline_equation", + "content": "1\\mu \\mathrm{s}" + }, + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "type": "text", + "content": ") and does not waste any pulses. As a result, our technique's depth estimation accuracy (computed using mean squared error averaged over a hundred experiments) is " + }, + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "type": "inline_equation", + "content": "250\\times" + }, + { + "bbox": [ + 304, + 361, + 548, + 554 + ], + "type": "text", + "content": " higher than that of the galvo mirrors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 564, + 375, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 564, + 375, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 564, + 375, + 576 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "We introduced a high-speed, low-cost (the cost of each transducer is only $26) technology to steer light. Our technology uses the physics of traveling sound waves synchronized with illumination and imaging sensors to enable multiple scanning applications. We built an experimental prototype to demonstrate proof-of-concept applications, such as the ability to perform arbitrary point light projection and raster scanning for LiDAR at MHz rates. In the rest of this section, we discuss some fundamental limits and tradeoffs inherent in our light steering technology due to wave physics, and directions for further exploration they suggest. We summa" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 750, + 309, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 757 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 542, + 328 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 542, + 328 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 542, + 328 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 542, + 328 + ], + "type": "image", + "image_path": "7e37173e651e71338d2ebe1160887581b25cfeeb92c5c082581557d0b3f98182.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": "Figure 8. Acousto-optic vs. galvo-mirror projection. We compare our beam steering technique with commercially available Thorlabs galvo mirrors (GVS-212). The \"A\" shape is made up of 100 points. With a " + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "inline_equation", + "content": "1\\mathrm{MHz}" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": " transducer, we are able to project a million points per second (pps), and hence, project ten thousand \"A\"s per second. Constrained by the laser's low beam power " + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "inline_equation", + "content": "(20\\mu \\mathrm{W}" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "inline_equation", + "content": "1\\mathrm{MHz})" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": ", practically we can only capture \"A\" at 1 ms exposure. The commercially available galvo mirrors, which are only rated at 1 kpps, only project a streak when driven at " + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "inline_equation", + "content": "1\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "inline_equation", + "content": "2\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "inline_equation", + "content": "5\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": ". At 50 ms exposure and " + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "inline_equation", + "content": "1\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": " scan rate, the galvo mirrors only project half the pattern at the rated " + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "inline_equation", + "content": "1\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 335, + 548, + 403 + ], + "type": "text", + "content": ", and at higher frequencies, the galvo mirrors project a corrupted pattern as we are operating them well beyond their 1 kpps rating." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 408, + 544, + 639 + ], + "blocks": [ + { + "bbox": [ + 48, + 408, + 544, + 639 + ], + "lines": [ + { + "bbox": [ + 48, + 408, + 544, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 544, + 639 + ], + "type": "image", + "image_path": "36e855f27d82270e5352235e958fe28ce9fe4d71c9e986f741a4727af36020db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "lines": [ + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "text", + "content": "Figure 9. Depth raster scanning. Each of the four scenes has two characters, \"CV\", \"PR\", \"20\", \"23\". \"C\" and \"P\" are at approximately " + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "inline_equation", + "content": "160\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "text", + "content": " depth, \"2\" at " + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "inline_equation", + "content": "170\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "text", + "content": ", and the remaining at " + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "inline_equation", + "content": "180\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "text", + "content": ". The top row shows the peak of the transient measured by the SPAD, and the bottom row shows the depth map in cm. We scan the scene at " + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "inline_equation", + "content": "100 \\times 100" + }, + { + "bbox": [ + 46, + 641, + 547, + 686 + ], + "type": "text", + "content": " resolution using the raster scanning technique in Section 4.1 for an exposure of one second. The Thorlabs galvo mirrors are not capable of scanning these scenes in under a second." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 749, + 309, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 749, + 309, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 749, + 309, + 757 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 59, + 165, + 167 + ], + "blocks": [ + { + "bbox": [ + 61, + 59, + 165, + 167 + ], + "lines": [ + { + "bbox": [ + 61, + 59, + 165, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 59, + 165, + 167 + ], + "type": "image", + "image_path": "d9de0c97472f8a3f82ca7e9b78cbd563742e6ac3437ef19e08fc99da0236d84c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 169, + 149, + 178 + ], + "lines": [ + { + "bbox": [ + 89, + 169, + 149, + 178 + ], + "spans": [ + { + "bbox": [ + 89, + 169, + 149, + 178 + ], + "type": "text", + "content": "(a) galvo mirrors" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 59, + 269, + 167 + ], + "blocks": [ + { + "bbox": [ + 78, + 178, + 160, + 188 + ], + "lines": [ + { + "bbox": [ + 78, + 178, + 160, + 188 + ], + "spans": [ + { + "bbox": [ + 78, + 178, + 160, + 188 + ], + "type": "text", + "content": "(depth error " + }, + { + "bbox": [ + 78, + 178, + 160, + 188 + ], + "type": "inline_equation", + "content": "= 51.3\\mathrm{cm}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 59, + 269, + 167 + ], + "lines": [ + { + "bbox": [ + 173, + 59, + 269, + 167 + ], + "spans": [ + { + "bbox": [ + 173, + 59, + 269, + 167 + ], + "type": "image", + "image_path": "f61e00bf1299a817d6553d9f7c863afdb35c09a86400f1ae554f29ffefb47359.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 169, + 252, + 178 + ], + "lines": [ + { + "bbox": [ + 192, + 169, + 252, + 178 + ], + "spans": [ + { + "bbox": [ + 192, + 169, + 252, + 178 + ], + "type": "text", + "content": "(b) our technique" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 178, + 263, + 188 + ], + "lines": [ + { + "bbox": [ + 182, + 178, + 263, + 188 + ], + "spans": [ + { + "bbox": [ + 182, + 178, + 263, + 188 + ], + "type": "text", + "content": "(depth error " + }, + { + "bbox": [ + 182, + 178, + 263, + 188 + ], + "type": "inline_equation", + "content": "= 3.28\\mathrm{cm}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 95, + 355, + 239, + 456 + ], + "blocks": [ + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "text", + "content": "Figure 10. Transient and depth measurement. We scan three small objects placed at different depths using commercial galvo mirrors and our technique. The total exposure duration is " + }, + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "inline_equation", + "content": "3\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "text", + "content": ", corresponding to the fastest speed (1kpps) at which the galvo mirrors can scan the scene. The laser repetition rate is " + }, + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "inline_equation", + "content": "1\\mathrm{MHz}" + }, + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "text", + "content": ". We show three transients measured by (a) the galvo mirrors and (b) our technique when the scan location is at the first, second, and third point. The galvo mirrors spend a significant amount of time moving the focus point from one object to another, wasting pulses emitted during that time. In contrast, our technique does not waste any pulses, as it moves the focus point in just " + }, + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "inline_equation", + "content": "1\\mu \\mathrm{s}" + }, + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "text", + "content": ", the time between two laser pulse emissions. This results in significant improvement in depth estimation accuracy (computed from the highest peak of the transient) compared to the galvo mirrors. The mean squared depth error of our technique is " + }, + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "inline_equation", + "content": "16\\times" + }, + { + "bbox": [ + 46, + 188, + 289, + 353 + ], + "type": "text", + "content": " smaller than galvos." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 355, + 239, + 456 + ], + "lines": [ + { + "bbox": [ + 95, + 355, + 239, + 456 + ], + "spans": [ + { + "bbox": [ + 95, + 355, + 239, + 456 + ], + "type": "image", + "image_path": "37b73fd585935e410e6dbe17f0382c158996069afcb8021bf883ef92aaf9f5ce.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 459, + 287, + 515 + ], + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 515 + ], + "type": "text", + "content": "Figure 11. Beam shape. We show an HDR image of the measured beam. It has a cross shape, as the ultrasonic lens has a square aperture. Using more transducers can result in the beam shape to be more Gaussian-like. The center peak vs. the crosshairs intensity ratio is 20, and the center peak vs. the background ratio is 400." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 520, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 555 + ], + "type": "text", + "content": "rize some of these tradeoffs in Table 3. In the supplement, we provide a physics-based renderer that can help virtually evaluate improved designs of our technology." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "content": "Scanning speed vs. aperture tradeoff. For the arbitrary point projector, the scanning speed (points per second, pps) is equal to the frequency of the ultrasonic transducer voltage. The aperture of the ultrasonically-sculpted lens is equal to the ultrasound wavelength " + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{us}} = c_{\\mathrm{us}} / f_{\\mathrm{us}}" + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "content": ". Therefore, the product of scanning frequency and the lens aperture is always less than or equal to the speed of sound. Increasing the scanning speed decreases the aperture of the ultrasonic lens. This decrease in the aperture size is not a problem for projector applications, but for LiDAR applications, the decrease in aperture decreases light throughput." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Diffraction limit. The numerical aperture of the ultrasonically-sculpted lens is approximately " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "n\\lambda_{\\mathrm{us}} / 2F" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 342, + 83, + 507, + 135 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "type": "text", + "content": "Table 3. Tradeoffs between various system parameters. " + }, + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "type": "text", + "content": " is better." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 83, + 507, + 135 + ], + "lines": [ + { + "bbox": [ + 342, + 83, + 507, + 135 + ], + "spans": [ + { + "bbox": [ + 342, + 83, + 507, + 135 + ], + "type": "table", + "html": "
parametercusfusλn
scan speed↑↑↓-
aperture↑↓--
spatial resolution↑↓↓↑
", + "image_path": "3a248ed2b49ca89c04b6732858de24968899cfec6bddc6ca28eebfbde12952eb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 151, + 547, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 151, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 306, + 151, + 547, + 176 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 151, + 547, + 176 + ], + "type": "text", + "content": " is the focal length of the ultrasonically-sculped waveguide. Therefore, the diffraction-limited spot size is" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 178, + 545, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 178, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 356, + 178, + 545, + 204 + ], + "type": "interline_equation", + "content": "\\Delta x = \\frac {\\lambda}{2 N A} \\approx \\frac {\\lambda}{\\lambda_ {\\mathrm {u s}}} \\frac {F}{n} = \\frac {\\lambda f _ {\\mathrm {u s}}}{c _ {\\mathrm {u s}}} \\frac {F}{n}. \\tag {7}", + "image_path": "c27ead3cd34b7596d0459e1e836ad0d07a52adfb8475607cc9f64e832666cd1f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 212, + 547, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 212, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 547, + 272 + ], + "type": "text", + "content": "Uncertainty principle. Spatial resolution is inversely proportional to the diffraction-limited spot size, and as mentioned earlier, temporal resolution is determined by the frequency of the transducer. So, we have the following uncertainty principle between spatial and temporal resolution:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 389, + 274, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 274, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 389, + 274, + 545, + 300 + ], + "type": "interline_equation", + "content": "\\Delta x \\Delta T \\leq \\frac {\\lambda}{c _ {\\mathrm {u s}}} \\frac {F}{n}. \\tag {8}", + "image_path": "6d7ae0197457fdce4ca8a9a56193322640dd6302f8b49af9ee15be97c843dcca.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 302, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 547, + 361 + ], + "type": "text", + "content": "To improve spatiotemporal resolution, we can decrease the focal length of the waveguide by increasing the voltage applied to the transducer. However, increasing the voltage makes the system non-linear and increases the probability of cavitation in the medium [52]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 368, + 547, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 547, + 499 + ], + "type": "text", + "content": "Another approach to improve the system's overall quality (increase aperture size, spatial and temporal resolutions) is to use a medium with a higher speed of sound. For example, tellurium dioxide " + }, + { + "bbox": [ + 304, + 368, + 547, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{(TeO_2)}" + }, + { + "bbox": [ + 304, + 368, + 547, + 499 + ], + "type": "text", + "content": " glass has three times higher speed of sound and " + }, + { + "bbox": [ + 304, + 368, + 547, + 499 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 368, + 547, + 499 + ], + "type": "text", + "content": " higher refractive index than water. Using tellurium dioxide glass will improve light efficiency by an order of magnitude, and simultaneously improve spatial resolution by five times. This material is also part of existing acousto-optic devices [30], but we found them hard to modify. Tellurium dioxide, being a solid, would additionally be a more stable medium than liquid water." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 504, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 547, + 624 + ], + "type": "text", + "content": "Shape of the blur kernel. We used two transducers in our system, and each one of them creates a cylindrical lens. The net effect of these two cylindrical lenses is a square aperture, whose Fourier transform is the product of two 1D sincs. Therefore, our blur kernel has a cross-shape, as we can see in Fig. 11. We can make this blur kernel closer to a Gaussian-like blur kernel by using multiple transducers arranged around a circular path and synchronized. Based on the application, we can also use deconvolution techniques to improve the results in post-processing [4]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments. We thank Hossein Baktash and Lloyd Lobo for help with the hardware prototype, and Andre Nascimento and Amy Lee for help with the physics-based renderer. This work was supported by NSF awards 1730147, 1900821, 1900849, 1935849, gifts from AWS Cloud Credits for Research and the Sybiel Berkman Foundation, and a Sloan Research Fellowship for Ioannis Gkioulekas." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 749, + 308, + 757 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 92, + 288, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 288, + 125 + ], + "type": "text", + "content": "[1] Technical specifications of Alpha Prime (VLS-128). https://www.mapix.com/lidar-scanner-sensors/velodyne/alpha-prime/. Accessed: 2023-03-27. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 127, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 127, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 53, + 127, + 288, + 161 + ], + "type": "text", + "content": "[2] Mohammad Abdo, Vlad Badilita, and Jan Korvink. Spatial scanning hyperspectral imaging combining a rotating slit with a dove prism. Optics Express, 27(15):20290-20304, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 207 + ], + "type": "text", + "content": "[3] Supreeth Achar, Joseph R. Bartels, William L. 'Red' Whittaker, Kiriakos N. Kutulakos, and Srinivasa G. Narasimhan. Epipolar time-of-flight imaging. ACM Trans. Graph., 36(4), jul 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 209, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 274 + ], + "type": "text", + "content": "[4] Hossein Baktash, Yash Belhe, Matteo Giuseppe Scopelliti, Yi Hua, Aswin C Sankaranarayanan, and Maysamreza Chamanzar. Computational imaging using ultrasonically-sculpted virtual lenses. In 2022 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 277, + 288, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 288, + 310 + ], + "type": "text", + "content": "[5] Jared Beller and Linbo Shao. Acousto-optic modulators integrated on-chip. Light: Science & Applications, 11(1):1-2, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 312, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 288, + 368 + ], + "type": "text", + "content": "[6] Alexander W Bergman, David B Lindell, and Gordon Wetzstein. Deep adaptive lidar: End-to-end optimization of sampling and depth completion at low sampling rates. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 370, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 288, + 414 + ], + "type": "text", + "content": "[7] Charles Bond, Adriana N Santiago-Ruiz, Qing Tang, and Melike Lakadamyali. Technological advances in super-resolution microscopy to study cellular processes. Molecular Cell, 82(2):315-332, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 416, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 416, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 416, + 288, + 459 + ], + "type": "text", + "content": "[8] Mauro Buttafava, Jessica Zeman, Alberto Tosi, Kevin Eliceiri, and Andreas Velten. Non-line-of-sight imaging using a time-gated single photon avalanche diode. Optics express, 23(16):20997-21011, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 462, + 288, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 462, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 462, + 288, + 506 + ], + "type": "text", + "content": "[9] Walter A Carrington, Kevin E Fogarty, Larry Lifschitz, and Fredric S Fay. Three-dimensional imaging on confocal and wide-field microscopes. Handbook of biological confocal microscopy, pages 151-161, 1990. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "type": "text", + "content": "[10] Maysamreza Chamanzar, Matteo Giuseppe Scopelliti, Julien Bloch, Ninh Do, Minyoung Huh, Dongjin Seo, Jillian Iafrati, Vikaaas S Sohal, Mohammad-Reza Alam, and Michel M Maharbiz. Ultrasonic sculpting of virtual optical waveguides in tissue. Nature communications, 10(1):1-10, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "type": "text", + "content": "[11] Dorian Chan, Srinivasa G Narasimhan, and Matthew O'Toole. Holocurtains: Programming light curtains via binary holography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17886-17895, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 622, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 288, + 667 + ], + "type": "text", + "content": "[12] Sreenithy Chandran, Hiroyuki Kubo, Tomoki Ueda, Takuya Funatomi, Yasuhiro Mukaigawa, and Suren Jayasuriya. Slopedisparity gating: System and applications. IEEE Transactions on Computational Imaging, 8:317-332, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[13] Ting-Hsuan Chen, Jesse Ault, Howard A Stone, and Craig B Arnold. High-speed axial-scanning wide-field microscopy for volumetric particle tracking velocimetry. *Experiments in Fluids*, 58(5):1-7, 2017. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[14] Ting-Hsuan Chen, Romain Fardel, and Craig B Arnold. Ultrafast z-scanning for high-efficiency laser micro-machining. Light: Science & Applications, 7(4):17181-17181, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 139 + ], + "type": "text", + "content": "[15] Yang Cheng, Jie Cao, and Qun Hao. Optical beam steering using liquid-based devices. Optics and Lasers in Engineering, 146:106700, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 546, + 184 + ], + "type": "text", + "content": "[16] Maxim N Cherkashin, Carsten Brenner, Georg Schmitz, and Martin R Hofmann. Transversally travelling ultrasound for light guiding deep into scattering media. Communications Physics, 3(1):1-11, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 217 + ], + "type": "text", + "content": "[17] Winifried Denk, James H Strickler, and Watt W Webb. Two-photon laser scanning fluorescence microscopy. Science, 248(4951):73-76, 1990. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 263 + ], + "type": "text", + "content": "[18] Xiaohan Du, SeungYeon Kang, and Craig B Arnold. Optimization of ultrafast axial scanning parameters for efficient pulsed laser micro-machining. Journal of Materials Processing Technology, 288:116850, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 265, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 547, + 297 + ], + "type": "text", + "content": "[19] Martí Duocastella, Craig B Arnold, and Jason Puchalla. S electable light-sheet uniformity using tuned axial scanning. Microscopy research and technique, 80(2):250-259, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 299, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 547, + 363 + ], + "type": "text", + "content": "[20] Vladislav Gavyusev, Giuseppe Sancataldo, Pietro Ricci, Alberto Montalbano, Chiara Fornetto, Lapo Turrini, Annunziatina Laurino, Luca Pesce, Giuseppe de Vito, Natascia Tiso, et al. Dual-beam confocal light-sheet microscopy via flexible acousto-optic deflector. Journal of biomedical optics, 24(10):106504, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 397 + ], + "type": "text", + "content": "[21] Milton S Gottlieb. Acousto-optic tunable filters. In Design and fabrication of acousto-optic devices, pages 197-283. CRC Press, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "type": "text", + "content": "[22] Ireneusz Grulkowski, Krzysztof Szulzycki, and Maciej Wojtkowski. Microscopic oct imaging with focus extension by ultrahigh-speed acousto-optic tunable lens and stroboscopic illumination. Optics Express, 22(26):31746-31760, 2014. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 546, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 546, + 465 + ], + "type": "text", + "content": "[23] Min Gu. Principles of three dimensional imaging in confocal microscopes. World Scientific, 1996. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 468, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 547, + 510 + ], + "type": "text", + "content": "[24] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3d imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7918, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "type": "text", + "content": "[25] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6770-6779, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 557, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 547, + 600 + ], + "type": "text", + "content": "[26] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G Narasimhan. Structured light 3d scanning in the presence of global illumination. In CVPR 2011, pages 713-720. IEEE, 2011. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 601, + 546, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 546, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 546, + 656 + ], + "type": "text", + "content": "[27] Kristofer Henderson, Xiaomeng Liu, Justin Folden, Brevin Tilmon, Suren Jayasuriya, and Sanjeev Koppal. Design and calibration of a fast flying-dot projector for dynamic light transport acquisition. IEEE Transactions on Computational Imaging, 6:529-543, 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 689 + ], + "type": "text", + "content": "[28] Sven TS Holmström, Utku Baran, and Hakan Urey. Mems laser scanners: a review. Journal of Microelectromechanical Systems, 23(2):259-275, 2014. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "text", + "content": "[29] Ching-Pai Hsu, Boda Li, Braulio Solano-Rivas, Amar R Gohil, Pak Hung Chan, Andrew D Moore, and Valentina" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 749, + 309, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 749, + 309, + 757 + ], + "spans": [ + { + "bbox": [ + 302, + 749, + 309, + 757 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "Donzella. A review and perspective on optical phased array for automotive lidar. IEEE Journal of Selected Topics in Quantum Electronics, 27(1):1-16, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 288, + 152 + ], + "type": "text", + "content": "[30] Dawoon Jeong, Hansol Jang, Min Uk Jung, and Chang-Seok Kim. Angular resolution variable fmcw lidar with acoustooptic deflector. In Imaging Systems and Applications, pages ITh4D-3. Optica Publishing Group, 2022. 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 154, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 154, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 154, + 288, + 198 + ], + "type": "text", + "content": "[31] S Kang, E Dotsenko, D Amrhein, C Theriault, and Craig B Arnold. Ultra-high-speed variable focus optics for novel applications in advanced imaging. In Photonic Instrumentation Engineering V, volume 10539, page 1053902. SPIE, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 200, + 288, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 200, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 288, + 232 + ], + "type": "text", + "content": "[32] SeungYeon Kang, Martí Duocastella, and Craig B Arnold. Variable optical elements for fast focus control. Nature Photonics, 14(9):533-542, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 288, + 278 + ], + "type": "text", + "content": "[33] Doycho Karagyozov, Mirna Mihovilovic Skanata, Amanda Lesar, and Marc Gershow. Recording neural activity in unrestrained animals with three-dimensional tracking two-photon microscopy. Cell reports, 25(5):1371-1383, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 280, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 288, + 324 + ], + "type": "text", + "content": "[34] Yasin Karimi, Matteo Giuseppe Scopelliti, Ninh Do, Mohammad-Reza Alam, and Maysamreza Chamanzar. In situ 3d reconfigurable ultrasonically sculpted optical beam paths. Optics express, 27(5):7249-7265, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 326, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 288, + 369 + ], + "type": "text", + "content": "[35] Byung-Sub Kim, Steve Gibson, and Tsu-Chin Tsao. Adaptive control of a tilt mirror for laser beam steering. In Proceedings of the 2004 American Control Conference, volume 4, pages 3417–3421. IEEE, 2004. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "type": "text", + "content": "[36] Lingjie Kong, Jianyong Tang, Justin P Little, Yang Yu, Tim Lammermann, Charles P Lin, Ronald N Germain, and Meng Cui. Continuous volumetric imaging via an optical phase-locked ultrasound lens. Nature methods, 12(8):759-762, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 428, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 288, + 483 + ], + "type": "text", + "content": "[37] Hiroyuki Kubo, Suren Jayasuriya, Takafumi Iwaguchi, Takuya Funatomi, Yasuhiro Mukaigawa, and Srinivasa G Narasimhan. Programmable non-epipolar indirect light transport: Capture and analysis. IEEE Transactions on Visualization and Computer Graphics, 27(4):2421-2436, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 485, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 288, + 529 + ], + "type": "text", + "content": "[38] Nanxi Li, Chong Pei Ho, Jin Xue, Leh Woon Lim, Guanyu Chen, Yuan Hsing Fu, and Lennon Yao Ting Lee. A progress review on solid-state lidar and nanophotonics-based lidar sensors. *Laser & Photonics Reviews*, page 2100511, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 288, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 574 + ], + "type": "text", + "content": "[39] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fkmigration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 577, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 288, + 610 + ], + "type": "text", + "content": "[40] Qiyu Liu, Huan Li, and Mo Li. Electromechanical Brillouin scattering in integrated optomechanical waveguides. Optica, 6(6):778-785, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[41] Xiaochun Liu, Ibón Guillén, Marco La Manna, Ji Hyun Nam, Syed Azer Reza, Toan Huu Le, Adrian Jarabo, Diego Gutierrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572(7771):620-623, 2019. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[42] Xiaomeng Liu, Kristofer Henderson, Joshua Rego, Suren Jayasuriya, and Sanjeev Koppal. Dense lissajous sampling and interpolation for dynamic light-transport. Optics Express, 29(12):18362–18381, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[43] Julio Marco, Adrian Jarabo, Ji Hyun Nam, Xiaochun Liu, Miguel Ángel Cosculluela, Andreas Velten, and Diego Gutierrez. Virtual light transport matrices for non-line-of-sight imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2440-2449, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "type": "text", + "content": "[44] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 547, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 547, + 236 + ], + "type": "text", + "content": "[45] Aongus McCarthy, Nils J Krichel, Nathan R Gemmell, Ximing Ren, Michael G Tanner, Sander N Dorenbos, Val Zwiller, Robert H Hadfield, and Gerald S Buller. Kilometer-range, high resolution depth imaging via 1560 nm wavelength single-photon detection. Optics express, 21(7):8904-8915, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 238, + 547, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 547, + 281 + ], + "type": "text", + "content": "[46] Srinivasa Narasimhan, Supreeth Achar, Kiriakos Kutulakos, Joe Bartels, and William Whittaker. Method for epipolar time of flight imaging, Mar. 19 2020. US Patent App. 16/468,617. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 282, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 547, + 326 + ], + "type": "text", + "content": "[47] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 547, + 360 + ], + "type": "text", + "content": "[48] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual coding to probe light transport. ACM Trans. Graph., 31(4):39-1, 2012. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 361, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 547, + 392 + ], + "type": "text", + "content": "[49] Matthew O'Toole, David B Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555(7696):338-341, 2018. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 547, + 415 + ], + "type": "text", + "content": "[50] James Pawley. Handbook of biological confocal microscopy, volume 236. Springer Science & Business Media, 2006. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "type": "text", + "content": "[51] Adithya Pediredla, Akshit Dave, and Ashok Veeraraghavan. Snlos: Non-line-of-sight scanning through temporal focusing. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 460, + 547, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 547, + 504 + ], + "type": "text", + "content": "[52] Adithya Pediredla, Matteo Giuseppe Scopelliti, Srinivasa Narasimhan, Maysamreza Chamanzar, and Ioannis Gkioulekas. Optimized virtual optical waveguides enhance light throughput in scattering media. preprint, 2021. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 504, + 547, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 547, + 570 + ], + "type": "text", + "content": "[53] Adithya Kumar Pediredla, Mauro Buttafava, Alberto Tosi, Oliver Cossairt, and Ashok Veeraraghavan. Reconstructing rooms using photon echoes: A plane based model and reconstruction algorithm for looking around the corner. In 2017 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 571, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 547, + 624 + ], + "type": "text", + "content": "[54] Adithya K Pediredla, Aswin C Sankaranarayanan, Mauro Buttafava, Alberto Tosi, and Ashok Veeraraghavan. Signal processing based pile-up compensation for gated single-photon avalanche diodes. arXiv preprint arXiv:1806.07437, 2018. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 625, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 547, + 680 + ], + "type": "text", + "content": "[55] Adithya Kumar Pediredla, Shizheng Zhang, Ben Avants, Fan Ye, Shin Nagayama, Ziying Chen, Caleb Kemere, Jacob T Robinson, and Ashok Veeraraghavan. Deep imaging in scattering media with selective plane illumination microscopy. Journal of biomedical optics, 21(12):126009, 2016. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 681, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 712 + ], + "type": "text", + "content": "[56] Kurt E Petersen. Silicon torsional scanning mirror. IBM Journal of Research and Development, 24(5):631-637, 1980. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 301, + 749, + 310, + 757 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[57] Ryan Po, Adithya Pediredla, and Ioannis Gkioulekas. Adaptive gating for single-photon 3d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16354–16363, 2022. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 172 + ], + "type": "text", + "content": "[58] Christopher V Poulton, Matthew J Byrd, Benjamin Moss, Erman Timurdogan, Ron Millman, and Michael R Watts. 8192-element optical phased array with " + }, + { + "bbox": [ + 49, + 118, + 288, + 172 + ], + "type": "inline_equation", + "content": "100^{\\circ}" + }, + { + "bbox": [ + 49, + 118, + 288, + 172 + ], + "type": "text", + "content": " steering range and flip-chip cmos. In CLEO: QELS_Fundamental Science, pages JTh4A-3. Optical Society of America, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 173, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 288, + 215 + ], + "type": "text", + "content": "[59] Joshua Rapp, Yanting Ma, Robin MA Dawson, and Vivek K Goyal. Dead time compensation for high-flux ranging. IEEE Transactions on Signal Processing, 67(13):3471-3486, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 288, + 248 + ], + "type": "text", + "content": "[60] Joshua Rapp, Yanting Ma, Robin MA Dawson, and Vivek K Goyal. High-flux single-photon lidar. Optica, 8(1):30-39, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 288, + 304 + ], + "type": "text", + "content": "[61] Joshua Rodriguez, Braden Smith, Brandon Hellman, Adley Gin, Alonzo Espinoza, and Yuzuru Takashima. Multi-beam and single-chip lidar with discrete beam steering by digital micromirror device. In Physics and Simulation of Optoelectronic Devices XXVI, volume 10526, pages 89-94. SPIE, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 304, + 288, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 304, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 288, + 347 + ], + "type": "text", + "content": "[62] Patrick S Salter, Z Iqbal, and Martin J Booth. Analysis of the three-dimensional focal positioning capability of adaptive optic elements. International Journal of Optomechatronics, 7(1):1-14, 2013. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 349, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 349, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 49, + 349, + 288, + 391 + ], + "type": "text", + "content": "[63] Matteo Giuseppe Scopelliti and Maysamreza Chamanzar. Ultrasonically sculpted virtual relay lens for in situ microimaging. Light: Science & Applications, 8(1):1-15, 2019. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 393, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 393, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 393, + 288, + 448 + ], + "type": "text", + "content": "[64] Xiaobing Shang, Jin-Yi Tan, Oliver Willekens, Jelle De Smet, Pankaj Joshi, Dieter Cuypers, Esma Islamaj, Jeroen Beeckman, Kristiaan Neyts, Michael Vervaekte, et al. Electrically controllable liquid crystal component for efficient light steering. IEEE Photonics Journal, 7(2):1-13, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 449, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 288, + 491 + ], + "type": "text", + "content": "[65] Braden Smith, Brandon Hellman, Adley Gin, Alonzo Espinoza, and Yuzuru Takashima. Single chip lidar with discrete beam steering by digital micromirror device. Optics Express, 25(13):14732-14745, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 493, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 493, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 493, + 288, + 536 + ], + "type": "text", + "content": "[66] Qi Wang Song, Xu-Ming Wang, Rebecca Bussjager, and Joseph Osman. Electro-optic beam-steering device based on a lanthanum-modified lead zirconate titanate ceramic wafer. Applied optics, 35(17):3155-3162, 1996. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 537, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 288, + 568 + ], + "type": "text", + "content": "[67] Steven J Spector. Review of lens-assisted beam steering methods. Journal of Optical Microsystems, 2(1):011003, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 570, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 570, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 570, + 288, + 613 + ], + "type": "text", + "content": "[68] Ernst HK Stelzer, Frederic Strobl, Bo-Jui Chang, Friedrich Preusser, Stephan Preibisch, Katie McDole, and Reto Fiolka. Light sheet fluorescence microscopy. Nature Reviews Methods Primers, 1(1):1-25, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 614, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 288, + 668 + ], + "type": "text", + "content": "[69] Lin Sun, Jin-ha Kim, Chiou-hung Jang, Dechang An, Xue-jun Lu, Qingjun Zhou, John Martin Taboada, Ray T Chen, Jeffery J Maki, Suning Tang, et al. Polymeric waveguide prism-based electro-optic beam deflector. Optical Engineering, 40(7):1217-1222, 2001. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 288, + 712 + ], + "type": "text", + "content": "[70] Varun Sundar, Sizhuo Ma, Aswin C Sankaranarayanan, and Mohit Gupta. Single-photon structured light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17865-17875, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[71] Opsys Tech. Opsys tech: The only scanning lidar with absolutely no moving parts. https://www.opsys-tech.com/post/opsys-tech-the-only-scanning-lidar-with-absolutely-no-moving-parts. Accessed: 2022-10-20. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 130, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 547, + 162 + ], + "type": "text", + "content": "[72] Chen S Tsai. Guided-wave acousto-optics: interactions, devices, and applications, volume 23. Springer Science & Business Media, 2013. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 165, + 547, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 547, + 220 + ], + "type": "text", + "content": "[73] Tomoki Ueda, Hiroyuki Kubo, Suren Jayasuriya, Takuya Funatomi, and Yasuhiro Mukaigawa. Slope disparity gating using a synchronized projector-camera system. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-9. IEEE, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 222, + 546, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 546, + 266 + ], + "type": "text", + "content": "[74] Federica Villa, Fabio Severini, Francesca Madonini, and Franco Zappa. Spads and sipms arrays for long-range high-speed light detection and ranging (lidar). Sensors, 21(11):3839, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 269, + 546, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 546, + 300 + ], + "type": "text", + "content": "[75] Dingkang Wang, Connor Watkins, and Huikai Xie. Mems mirrors for lidar: a review. Micromachines, 11(5):456, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 547, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 547, + 357 + ], + "type": "text", + "content": "[76] Jian Wang, Joseph Bartels, William Whittaker, Aswin C Sankaranarayanan, and Srinivasa G Narasimhan. Programmable triangulation light curtains. In Proceedings of the European Conference on Computer Vision (ECCV), pages 19-34, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 361, + 546, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 546, + 416 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 546, + 416 + ], + "type": "text", + "content": "[77] Xiangru Wang, Liang Wu, Caidong Xiong, Man Li, Qinggui Tan, Jiyang Shang, Shuanghong Wu, and Qi Qiu. Agile laser beam deflection with high steering precision and angular resolution using liquid crystal optical phased array. IEEE Transactions on Nanotechnology, 17(1):26-28, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 418, + 547, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 547, + 462 + ], + "type": "text", + "content": "[78] Zihan Wang, Jie Cao, Qun Hao, Fanghua Zhang, Yang Cheng, and Xianyue Kong. Super-resolution imaging and field of view extension using a single camera with risley prisms. Review of Scientific Instruments, 90(3):033701, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 464, + 547, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 547, + 518 + ], + "type": "text", + "content": "[79] RM Waxler and CE Weir. Effect of pressure and temperature on the refractive indices of benzene, carbon tetrachloride, and water. Journal of research of the National Bureau of Standards. Section A, Physics and chemistry, 67(2):163, 1963. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 521, + 547, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 547, + 587 + ], + "type": "text", + "content": "[80] Ming-Tzo Wei, Shana Elbaum-Garfinkle, Alex S Holehouse, Carlos Chih-Hsiung Chen, Marina Feric, Craig B Arnold, Rodney D Priestley, Rohit V Pappu, and Clifford P Brangwynne. Phase behaviour of disordered proteins underlying low density and high permeability of liquid organelles. Nature chemistry, 9(11):1118-1125, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 547, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 654 + ], + "type": "text", + "content": "[81] Shumian Xin, Sotiris Nousias, Kiriakos N Kutulakos, Aswin C Sankaranarayanan, Srinivasa G Narasimhan, and Ioannis Gkioulekas. A theory of fermat paths for non-line-of-sight shape reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6800-6809, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "text", + "content": "[82] Chris Xu, Warren Zipfel, Jason B Shear, Rebecca M Williams, and Watt W Webb. Multiphoton fluorescence excitation: new spectral windows for biological nonlinear microscopy. Proceedings of the National Academy of Sciences, 93(20):10763-10768, 1996. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 750, + 309, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 750, + 309, + 757 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 309, + 757 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 241 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[83] Jiawen Xu and Jiong Tang. Tunable prism based on piezoelectric metamaterial for acoustic beam steering. Applied Physics Letters, 110(18):181902, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "text", + "content": "[84] Donghai Yang, Yifan Liu, Qingjiu Chen, Meng Chen, Shaodong Zhan, Nim-kwan Cheung, Ho-Yin Chan, Zhidong Wang, and Wen Jung Li. Development of the high angular resolution " + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "text", + "content": " lidar based on scanning mem's mirror. Scientific Reports, 13(1):1540, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "text", + "content": "[85] Zejie Yu and Xiankai Sun. Gigahertz acousto-optic modulation and frequency shifting on etchless lithium niobate integrated platform. ACS Photonics, 8(3):798-803, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "type": "text", + "content": "[86] Mantas Žurauskas, Oliver Barnstedt, Maria Frade-Rodriguez, Scott Waddell, and Martin J Booth. Rapid adaptive remote focusing microscope for sensing of volumetric neural activity. Biomedical optics express, 8(10):4369-4379, 2017. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 750, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 757 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_content_list.json b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..442f36ddbf29b6bdcd00f53b6784a58b892401b7 --- /dev/null +++ b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_content_list.json @@ -0,0 +1,1447 @@ +[ + { + "type": "text", + "text": "Memory-friendly Scalable Super-resolution via Rewinding Lottery Ticket Hypothesis", + "text_level": 1, + "bbox": [ + 191, + 161, + 783, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jin Lin $^{1*}$ , Xiaotong Luo $^{1*}$ , Ming Hong $^{1}$ , Yanyun Qu $^{1\\dagger}$ , Yuan Xie $^{2\\dagger}$ , Zongze Wu $^{3}$", + "bbox": [ + 179, + 226, + 786, + 245 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Informatics, Xiamen University, Fujian, China", + "bbox": [ + 257, + 245, + 712, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Computer Science and Technology, East China Normal University, Shanghai, China \n $^{3}$ School of Mechatronics and Control Engineering, Shenzhen University, Shenzhen, China", + "bbox": [ + 109, + 261, + 858, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{linjin, xiaotluo, mingh}@stu.xmu.edu.cn, yyqu@xmu.edu.cn,", + "bbox": [ + 223, + 296, + 734, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yxie@cs.ecnu.edu.cn, zzwu@szu.edu.cn", + "bbox": [ + 315, + 314, + 636, + 326 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 359, + 312, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scalable deep Super-Resolution (SR) models are increasingly in demand, whose memory can be customized and tuned to the computational recourse of the platform. The existing dynamic scalable SR methods are not memory-friendly enough because multi-scale models have to be saved with a fixed size for each model. Inspired by the success of Lottery Tickets Hypothesis (LTH) on image classification, we explore the existence of unstructured scalable SR deep models, that is, we find gradual shrinkage subnetworks of extreme sparsity named winning tickets. In this paper, we propose a Memory-friendly Scalable SR framework (MSSR). The advantage is that only a single scalable model covers multiple SR models with different sizes, instead of reloading SR models of different sizes. Concretely, MSSR consists of the forward and backward stages, the former for model compression and the latter for model expansion. In the forward stage, we take advantage of LTH with rewinding weights to progressively shrink the SR model and the pruning-out masks that form nested sets. Moreover, stochastic self-distillation (SSD) is conducted to boost the performance of sub-networks. By stochastically selecting multiple depths, the current model inputs the selected features into the corresponding parts in the larger model and improves the performance of the current model based on the feedback results of the larger model. In the backward stage, the smaller SR model could be expanded by recovering and fine-tuning the pruned parameters according to the pruning-out masks obtained in the forward. Extensive experiments show the effectiveness of MMSR. The smallest-scale sub-network could achieve the sparsity of $94\\%$ and outperforms the compared lightweight SR methods.", + "bbox": [ + 75, + 389, + 472, + 818 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5cc0ce30a2990a07f9b58171cda1cd176768b8e9c3c0d6c539f47ed29203d2e.jpg", + "image_caption": [ + "Figure 1. The flowchart of scalable SR network. S_n and S_w denote the neurons and the neural connections (weights) of the simplest subnetwork; M_n and M_w denote the intermediate neurons and neural connections; Cur_n and Cur_w denote the specific neurons and neural connections belonging to the current subnetwork. Free_n denotes the pruning-out neurons. The final model (with $100\\%$ recovered parameters) reaches the original size. The scalable model is adjustable to the memory resource allocation." + ], + "image_footnote": [], + "bbox": [ + 509, + 358, + 884, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 633, + 630, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image super-resolution (SISR) aims to reconstruct a high-resolution (HR) image from the corresponding low-resolution (LR) one. With the rising of deep learning, deep SR methods have made incredible progress. However, the existing SR models mostly require computational and memory resources, so they do not favor resource-limited devices such as mobile phones, robotics, and some edge devices.", + "bbox": [ + 497, + 657, + 890, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The lightweight SR methods are attracting more attention for better application to resource-limited devices. The existing lightweight SR methods mainly focus on designing compact architectures [17, 20] with a fixed size, such as multi-scale pyramid [20], multiple-level receptive fields [17, 18], and recursive learning [19]. However, most lightweight SR models with fixed sizes are not flexible in applications. If one model does not match the resources of", + "bbox": [ + 497, + 755, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 86 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 95, + 842, + 201, + 853 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 95, + 853, + 223, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14398", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the platform, it has to be retrained by compression methods to match the resources and then reloaded onto the devices.", + "bbox": [ + 75, + 125, + 467, + 153 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The urgent demand to customize models based on deployment resources is increasing. Dynamic neural networks for SR [14,22] are proposed to adjust the network architecture according to different computational resources. The existing dynamic deep SR models often explore dynamic depth or width [22, 26], but they either require large memory resources or are not convenient for users to wait for retraining another SR model. The former leads to saving the multi-scale SR models of different sizes and the latter leads to retraining the model before being reused again. The limitation lies in that they are not memory-friendly. In many edge-device SR applications, the devices may be scalable, that is, their memories may be small in the beginning and be expanded later. Thus, we discuss two issues in this paper: 1) how to make a scalable lightweight model for the multi-scale computational recourse. 2) how to make the lightweight model expand to a larger-size model for better performance if the computational recourse is increased.", + "bbox": [ + 75, + 155, + 467, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for the first issue, inspired by the success of Lottery Ticket Hypothesis [10] which points out that there could exist a properly pruned sub-network named winning tickets to achieve comparable performance against the original dense network in model compression of classification, it is used to find the sub-network for SR. We are the first to study the existence of scalable winning tickets for SR. Iterative pruning and rewinding weights in LTH are beneficial to the scalable lightweight SR model. Iterative pruning may compress the SR model according to an arbitrary size. It is observed in [24] that the winning tickets are related to an insufficient DNN, and rewinding LTH outperforms the original LTH. That is, the initial weights in LTH are replaced with the T-iteration weights during pruning and fine-tuning. The scalable deep SR model is shown in Fig. 1.", + "bbox": [ + 75, + 406, + 467, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for the second issue, the scalable SR model can customize parameters to adapt to different memory resources rather than load or offload different models for different devices. In other words, during real applications, there will be only one simple model to be employed for inference whose size is decided by the computational resource.", + "bbox": [ + 75, + 615, + 467, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a memory-friendly scalable deep SR model (MSSR) via rewinding LTH. We use the rewinding LTH [10] to generate our unstructured scalable mask. MSSR is backtracking and contains forward and backward stages. The former focuses on model compression by rewinding LTH with iterative pruning and finetuning, and the latter focuses on iterative model expansion until it goes back to its original size. Multi-scale winning tickets together with the pruning-out masks are obtained by rewinding LTH in the forward stage with the decrease in the number of parameters. The pruning-out masks are nested. In order to make the compressed SR model not degrade sig", + "bbox": [ + 75, + 700, + 467, + 865 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nificantly, stochastic self-distillation (SSD) is used to improve the representation of the small-scale SR model, and knowledge is transferred from the last-scale model to the current scale model. In the backward stage, the smallest model is expanded gradually to the model with the original size with the expanded mask.", + "bbox": [ + 497, + 125, + 890, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are three-fold:", + "bbox": [ + 519, + 210, + 853, + 223 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A memory-friendly scalable dynamic SR lightweight model via rewinding LTH is proposed. MSSR is re-configurable and switchable to sub-networks with different sizes according to on-device resource constraints on the fly.", + "- MSSR is backtracking, which contains forward and backward stages. Multi-scale winning tickets form nested masks for the multi-scale models. SSD is conducted by replacing the features in randomly selected layers between Teacher and Student to improve the performance of the scalable SR lightweight models.", + "- Extensive experiments demonstrate that MSSR can generalize to different SR models as well as state-of-the-art attention-based models, ENLCN [1]." + ], + "bbox": [ + 517, + 231, + 890, + 442 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 499, + 459, + 640, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Lightweight Deep Learning-based SR", + "text_level": 1, + "bbox": [ + 499, + 482, + 823, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SISR aims to generate a high-resolution image from the low-resolution version. Recently, efficient and lightweight SISR networks have attracted increasing interest in the computer vision community. The lightweight SR methods consist of static and dynamic neural networks for SR. The earlier lightweight SR methods mainly focus on the static network structure design. IDN [18] utilizes group convolution and combined the local long and short path features effectively. CARN [19] utilizes a recursive cascading mechanism for learning multi-level feature representations. IMDN [17] introduces information multi-distillation blocks for extracting hierarchical features. FALSR [6] applies neural architecture search for efficient network design automatically. LatticeNet [27] designs a lattice block to combine two residual blocks adaptively. However, the aforementioned lightweight SR models need elaborate designs.", + "bbox": [ + 497, + 505, + 890, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, dynamic lightweight SR models are demanded increasingly whose size is adjustable according to different computational resources. DS-Net [22] makes the width of the SR model adjustable, and IKS [26] makes the sparsity of individual kernels adjustable. However, they require saving lightweight SR models with different shrinkage sizes, which is not convenient for users to offload the old model and then load a new one. Our method uses a scalable SR model adjustable to the changing computational recourse to avoid the loading and offloading models.", + "bbox": [ + 497, + 726, + 890, + 865 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14399", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4cbf2900c4189795dcdd7620250b00d900293e7ce223b8daa04e70e62a4b6e42.jpg", + "image_caption": [ + "Figure 2. The architecture of memory-friendly scalable deep SR model (MSSR). It contains two stages: the forward stage for model compression and the backward stage for model expansion. In the forward stage, the original SR network is firstly processed by IMP to obtain winning tickets and the pruning-out mask $LM_{1}$ , and then the winning tickets are processed by SSD and the lightweight subnetwork is obtained. So do the other gradual shrinkage sub-networks and $\\{LM_{i}\\}$ forms the nested sets. In the backward stage, according to the pruning-out masks, the smaller lightweight SR model is gradually expanded and finely tuned to obtain a larger lightweight SR model." + ], + "image_footnote": [], + "bbox": [ + 82, + 124, + 884, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Pruning and Lottery Ticket Hypothesis", + "text_level": 1, + "bbox": [ + 75, + 524, + 413, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model pruning is to prune away the unimportant connections (parameters) [21] in the neural network which trades off between the model size and performance. The existing pruning methods can be classified into unstructured [6, 13, 21] and structured methods [15, 25, 39]. The structured pruning methods remove the substructures of the model, such as channels, filters, and layers, while the unstructured pruning methods are magnitude-based, where the weights are removed below the threshold. However, magnitude-based pruning methods are not hardware-friendly and computationally efficient. Lottery Ticket Hypothesis (LTH) [10] opened a pandas's box of immense possibilities in the field of pruning and sparse models. It assumes that there exist sparse sub-networks independently training from scratch that can match the performance of dense networks. Some literatures discuss the effectiveness of pruning, fine-tuning [24], and rewinding weights [29] for LTH.", + "bbox": [ + 75, + 547, + 468, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LTH has been widely applied in image classification [9-12, 30, 32, 33], natural language processing [5, 28, 29, 34], and so on. However, few works discuss the existence of LTH in the deep model for image restoration. We are the first to investigate the existence of scalable winning tickets for lightweight deep SR models.", + "bbox": [ + 75, + 783, + 470, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 499, + 523, + 589, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of MSSR", + "text_level": 1, + "bbox": [ + 499, + 549, + 682, + 562 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the memory-friendly scalable SR model which is an all-in-one model with multiscale SR models sharing a model. As shown in Fig. 2, our MSSR consists of two stages: the forward stage for model compression, and the backward stage for model expansion.", + "bbox": [ + 497, + 573, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the forward stage, we use rewinding LTH (RLTH) to iteratively shrink a dense SR model to obtain a sub-network with the small size not significantly scarifying the performance. During pruning by LTH each time, we treat the current lightweight SR model as Student and its last ancestor lightweight SR model as Teacher, and conduct stochastic self-distillation between Teacher and Student. In Fig. 2, the dark-blue block denotes the winning tickets, and the white block denotes the pruning part. We also record the mask of pruning part and use a matrix with \"1\" to represent the pruning-out weights and with \"0\" for others. It is observed that the masks are nested and the initial mask is continuously contained in the later mask. In the backward stage, we iteratively expand the lightweight sub-network according to the saved nested masks. The most lightweight network marked by 1 is expanded iteratively to the gradual less", + "bbox": [ + 497, + 644, + 892, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14400", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lightweight network, the expanding order is shown as \"1\", \"1-2\", \"1-2-3\", and \"1-2-3-4\".", + "bbox": [ + 75, + 125, + 468, + 153 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. The Forward Stage: Model Compression", + "text_level": 1, + "bbox": [ + 77, + 163, + 428, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The forward streamline is shown in Fig. 3. Let a network denoted by $f(\\theta_0; x)$ and the pruned masks which form the \"nested sets\" denoted by $LM = \\{\\}$ , where $\\theta_0$ denotes the initial weights. The flowchart shows a whole process in an iteration for shrinking a model and it contains three components: SR model pre-train, IMP with rewinding weights, and SSD. Firstly, an SR network is trained to reach a steady state, and its weights are named the first nonpruning weights $\\theta$ . After that, Iterative Magnitude Pruning is conducted on the SR model $\\theta$ , and the weights below the pre-set threshold are pruned out which forms a pruning mask $m \\in \\{0,1\\}^n$ . Following [24], we rewind the remaining weights to their values in $\\theta$ , i.e., $\\theta' = \\theta \\odot m$ . Then, we retrain the pruned weights $\\theta'$ by SSD arriving at $\\hat{\\theta}$ . So do the next more lightweight SR model with the model weights $\\hat{\\theta}$ as the initiation.", + "bbox": [ + 75, + 185, + 468, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LTH with Rewinding for SR. Benefiting from LTHs, we implement iterative magnitude pruning with rewinding to gradually shrink the SR model. The motivation lies in the controllability and flexibility of LTH. We can shrink the model arbitrarily by LTH, which favors the scalable lightweight SR model.", + "bbox": [ + 75, + 407, + 468, + 490 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We mainly consider two issues: the model pruning process and the pruning rules. The pruning process corresponds to the process of gradually shrinking the model, and the pruning methods need to meet the following rules: 1) the amount of prune-out parameters is arbitrarily adjustable; 2) the model is shrunk gradually, which leads to the nested sets and the larger model contains its smaller descendant. LTH fits well with the needs of our scalable mask generator and the appropriate use of its task settings to implement the model compression process, in particular discarding its over parameterization, a feature that is not suitable for intensive prediction tasks.", + "bbox": [ + 75, + 491, + 468, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IMP is proven to be effective for LTH. In [24], it is proved that weight rewinding is better than the original weight initiation. That is, unpruned weights are rewound to their values from $T$ -iteration training and retrained from there using the original training schedule.", + "bbox": [ + 75, + 657, + 468, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with one-shot magnitude pruning (OMP) which reaches the goal model at one time, IMP does not shrink a model to reach the pre-set size at one time but shrinks a model with several iterations and the model size is evenly decreased at each iteration. Let the network be pruned for $T$ times and be expected to discard $p\\%$ of the parameters. The weights/parameters of the network are represented by $\\omega \\in \\mathbb{R}^n$ and the pruning mask by $m \\in \\{0,1\\}^n$ where $n$ represents all the capacity of the network.", + "bbox": [ + 75, + 726, + 468, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weight rewinding does not initiate the model with the", + "bbox": [ + 95, + 852, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c4dc5c98f9c2e90ab20862a08e1e97e9d7a22e8c2c9a4de97d2fe75e732194cd.jpg", + "image_caption": [ + "Figure 3. The flowchart of the whole process of IMP and SSD. The initial model $\\theta_0$ is pre-trained to obtain a stable model $\\theta$ . IMP is implemented on $\\theta$ and the winning ticket (sub-network) is obtained. The sub-network is initiated by $\\theta \\odot m$ and is processed by SSD for the better model $\\theta$ ." + ], + "image_footnote": [], + "bbox": [ + 502, + 123, + 878, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "original initiation weights $\\theta_0\\odot m$ , but with the $t$ -iteration $\\theta_t\\odot m$ where $0 < t\\leq T$ , resulting in weights $\\theta_{t}$ and the network function $f(x;\\theta_t\\odot m)$ . We highlight that the original LTH always chooses $\\theta_0$ as dense model, while in this work we choose $\\theta_{t}$ from the last pruning iteration. After \"Pruning\" and \"Rewinding and Retraining\", $\\theta = \\hat{\\theta}$ is updated and the mask at the $t$ -th iteration is obtained which is denoted as $LM_{t} = LM_{t - 1}\\cup \\{m\\}$ . In other words, the mask $LM_{t}$ at $t$ -th iteration is contained in the mask $LM_{t - 1}$ at $(t - 1)$ -th. \"Pruning\" and \"Rewinding and Retraining\" are repeated until the model reaches the end condition, which is the iteration steps $T$ or target sparsity $p$", + "bbox": [ + 497, + 366, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stochastic Self-distillation for SR. In the forward stage, due to the progressive shrinking of the SR model, it is definitely difficult to maintain the performance as same as the original model. Especially, when only $1\\%$ of the parameters are kept, most magnitude-pruning methods do not work.", + "bbox": [ + 497, + 533, + 890, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to mitigate the degradation of lightweight SR models, we adopt self-distillation [36] to boost the performance of the pruned model. Unlike the existing self-distillation which uses the function of features obtained by the model at previous iterations such as the average and summation function, we introduce attribution analysis to self-distillation. We randomly select several layers in Student, and the feature maps of selected layers replace the counterpart of Teacher, which is named Stochastic self-distillation (SSD). Concretely, the current lightweight SR model is regarded as the student network (Student), and its last ancestor network is regarded as the teacher network (Teacher), and Student is obtained by pruning Teacher. SSD is conducted between Teacher and Student. Student is the pruned model of Teacher. Let $X$ and $Y$ denote the network input and ground truth, respectively.", + "bbox": [ + 497, + 603, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of SSD is shown in Fig. 4, which contains two streamlines: Teacher and Student. The knowledge in feature maps is transferred from Teacher to Student.", + "bbox": [ + 497, + 824, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14401", + "bbox": [ + 465, + 887, + 502, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3fa501b40f78f274e3ace7e22932cd8d736c9cb22d1f3ef6857b4b25ed35364.jpg", + "image_caption": [ + "Figure 4. The flowchart of SSD. Student is the sub-network by pruning Teacher. Several layers of Student are selected randomly. The feature maps in a selected layer replace the counterpart of Teacher which outputs the SR result, and so do other selected layers. The total loss is the summation of the knowledge distillation loss $L_{skd}$ and the fidelity loss $L_{1}$ ." + ], + "image_footnote": [], + "bbox": [ + 102, + 123, + 436, + 302 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If Student contains $d$ layers, at each-iteration distillation, $s$ layers randomly are selected. In realization, a one-hot vector is randomly generated, named Switch Vector, with a length of $d$ in each training, $\\alpha \\in \\{0,1\\}^d$ . If the element is equal to 1 in $\\alpha$ , the feature maps in the layer corresponding to this position are selected, and are named Lucky feature maps. Then, Lucky feature maps in the selected layers to replace the counterparts of Teacher.", + "bbox": [ + 75, + 401, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In details, Student firstly is trained for a stable SR model. After that, SSD is conducted. At each iteration, Switch vector is dynamically generated, which is formulated as:", + "bbox": [ + 75, + 514, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD C = \\left[ \\alpha_ {1}, \\alpha_ {2}, \\dots , \\alpha_ {d} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 558, + 468, + 572 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ denotes the $i$ -th element corresponding to the $i$ -th layer and $d$ is the layer number in Student. Then, lucky feature maps of Student are selected according to $\\alpha_{i}$ , and they are formulated as:", + "bbox": [ + 75, + 574, + 468, + 629 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L F} = \\left[ f _ {\\alpha_ {1}} ^ {s}, f _ {\\alpha_ {2}} ^ {s}, \\dots , f _ {\\alpha_ {d}} ^ {s} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 631, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\alpha}^{s}$ denotes the feature obtained from Student. Note that all elements in Switch Vector are independent and identically distributed.", + "bbox": [ + 75, + 650, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lucky feature maps replace the counterpart of Teacher each time, and the corresponding output of new Teacher is denoted by $Y_{\\alpha_i}^t$ , and the corresponding weights of new Teacher are denoted by $\\theta_t$ , which is formulated as:", + "bbox": [ + 75, + 695, + 468, + 749 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} Y _ {\\alpha_ {i}} ^ {t} = F \\left(\\theta^ {t}; x, f _ {\\alpha_ {i}} ^ {s}\\right), i = 1, \\dots , d, \\\\ \\mathrm {L Y} = \\left[ Y _ {\\alpha_ {1}} ^ {t}, Y _ {\\alpha_ {2}} ^ {t}, \\dots , Y _ {\\alpha_ {d}} ^ {t} \\right], \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 751, + 468, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $LY$ denotes the set of SR outputs, $Y_{\\alpha}^{t}$ is the SSD output of lucky maps, $f_{\\alpha}^{s}$ . All the outputs between Student and Teacher are supervised by HR image $Y$ . The knowledge of Teacher takes more rationality to be distilled into Student.", + "bbox": [ + 75, + 783, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the total loss function utilized to train the SR model consists of the supervised loss $L_{1}$ and self-distillation", + "bbox": [ + 75, + 839, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Memory-friendlyly scalable lightweight framework for image SR.", + "bbox": [ + 499, + 125, + 889, + 153 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data: HR and LR image datasets: $D_{hr}$ , $D_{lr}$", + "bbox": [ + 499, + 155, + 791, + 168 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Dense network with randomly initial weights: $f(\\theta_0)$ , pruning rounds: $T$ , and pruning target rate: $p$", + "bbox": [ + 499, + 169, + 850, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Scalable lightweight SR model", + "bbox": [ + 500, + 195, + 768, + 209 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: /* The forward stage: model compression */", + "2: Initializing: Pruned mask $m = \\{1\\}^n$ , lottery mask pool $LM = \\{m\\}$ and epochs of each training $N$", + "3: Pre-training: Train network for $N$ epochs $f(\\theta_0;x)\\longrightarrow f(\\theta ;x)$", + "4: while $i < T - 1$ do", + "5: Prune bottom $p\\% \\times \\frac{1}{T}$ of $m \\odot \\theta$ and update $m$ ;", + "6: Re-train the pruned sub-network $\\theta^{\\prime} = m\\odot \\theta$ for $N$ epochs $f(\\theta^{\\prime};x)\\longrightarrow f(\\hat{\\theta};x)$ with $SSD(\\theta ,\\theta^{\\prime};x)$", + "7: Update mask pool $LM = LM \\cup \\{m\\}$ and weights $\\theta = \\hat{\\theta}$ ;", + "8: $i\\gets i + 1$", + "9: end while", + "10: Generate the sparse sub-network matching the target sparsity, and the mask pool $LM$ ;", + "11: /* The backward stage: model expansion */", + "12: Initializing: Current step $t = T$ and sparsest sub-network $\\theta_T = \\theta$", + "13: while $t > 1$ do", + "14: Calculate the grad mask $m = LM_{t - 1} - LM_t$ to determine the trainable weights;", + "15: Fine-tune/re-train the trainable weights for $N$ epochs $f(\\theta_t;x)\\longrightarrow f(\\theta_t';x)$", + "16: Update $\\theta_t = \\theta_t'$ ;", + "17: $t\\gets t - 1;$", + "18: end while" + ], + "bbox": [ + 505, + 212, + 887, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss $L_{skd}$ , which is formulated as:", + "bbox": [ + 499, + 598, + 726, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = L _ {1} + \\lambda L _ {s k d} = \\left\\| Y ^ {s} - Y \\right\\| _ {1} + \\frac {\\lambda}{d} \\sum_ {i = 1} ^ {d} \\alpha_ {i} \\left\\| Y _ {\\alpha_ {i}} ^ {t} - Y \\right\\| _ {1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 512, + 622, + 889, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ denotes the regularization parameter, which is empirically set as 0.1 in our experiments. During inference, the SR model $\\theta_{p}$ can be customized for various resources according to different sparsity $p$ . After the network training, a lottery mask pool can be built for the backward stage.", + "bbox": [ + 497, + 668, + 890, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. The Backward Stage: Model Expansion", + "text_level": 1, + "bbox": [ + 499, + 747, + 840, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our intention is to obtain a scalable network that can deploy according to customized memory resources. In this subsection, we introduce how to gradually recover a lightweight SR model, which is a backtracking process of IMP. We start from the most lightweight SR model and end with the original SR model. Corresponding to gradually shrinking the SR model with LTH, we solve two issues in", + "bbox": [ + 497, + 769, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14402", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76a605c86e31ff150ac11bb41a27637194670e3b2fd7677e655658c7b2156c43.jpg", + "image_caption": [ + "Network Sparsity (\\% pruned)", + "Figure 5. Comparisons of IMP and OMP on various famous SR methods, which are evaluated on Set14 for $4 \\times$ SR. The performance of the baseline dense network, the sparse sub-network obtained by IMP, as well as the sub-network obtained by OMP. The sub-network is with only $20\\%$ of the weights." + ], + "image_footnote": [], + "bbox": [ + 90, + 119, + 457, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the model expansion: iterative model expansion (IME) and the expansion rule, the former refers to the requirement of the IME process and the latter refers to the requirement of the learning rule.", + "bbox": [ + 75, + 340, + 468, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let's focus on an expansion from the $t$ -th lightweight model to the $(t - 1)$ -th lightweight model where $t$ comes from the definition in the forward stage. The larger $t$ is, the smaller model is. During IMP, we obtain two pruned-out masks corresponding to the $t$ -th and $(t - 1)$ -th lightweight models, denoted by $\\mathrm{LM}_t$ and $\\mathrm{LM}_{t - 1}$ , respectively. Their difference represents the expanded positions, i.e., $m = LM_{t - 1} - LM_t$ . Without changing the $t$ -th model, we recover the weights according to the mask $m$ . After that, we fine-tune the recovered model and obtain the $(t - 1)$ -th model. So do the next model expansion. The detailed procedure is described in Algorithm 1.", + "bbox": [ + 75, + 397, + 468, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 75, + 577, + 208, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 75, + 600, + 179, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DIV2K [2] is adopted to train the SR model, which is a high-quality natural scene dataset widely used for image SR tasks. The whole dataset includes 800 training images and 100 validation images totally with abundant texture and contents. The bicubic degraded LR images are synthesized following the previous works [17, 37]. Besides, the trained models are evaluated on four public SR benchmark datasets: Set5 [4], Set14 [35], B100 [3], and Urban100 [16].", + "bbox": [ + 75, + 621, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b7e0307b616471c5b885e7fff9b128d110f9da70aa1f83e31206c6501585966b.jpg", + "table_caption": [ + "Table 1. Ablation studies about the iterative magnitude pruning (IMP) and the stochastic self-distillation learning (SSD). All these models are validated on Set14 for $4 \\times$ image SR." + ], + "table_footnote": [], + "table_body": "
EDSR基礎IMP (LTH)IMP RewindingSSDSparsityPSNR
✓×××0%28.55
✓✓××94%27.66
✓×✓×94%28.12
✓×✓✓94%28.33
", + "bbox": [ + 78, + 789, + 468, + 855 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5d748414ff03a7ca9e7fcb2104ea3d23b11853521630b1920ce29b0298842639.jpg", + "table_caption": [ + "Table 2. The performance of the sparse sub-network pruned by IMP-SSD on Set14 for $2 \\times, 3 \\times, 4 \\times$ image SR." + ], + "table_footnote": [], + "table_body": "
ScaleModelSparsityPSNR (dB)Params
2×EDSR基礎0%33.631370K
EDSR_IMP94%32.8885K
EDSR_IMP-SSD94%33.09 (+0.21)85K
3×EDSR基礎0%30.301558K
EDSR_IMP94%29.7397K
EDSR_IMP-SSD94%29.92 (+0.19)97K
4×EDSR基礎0%28.551518K
EDSR_IMP94%28.1295K
EDSR_IMP-SSD94%28.33 (+0.21)95K
", + "bbox": [ + 500, + 155, + 890, + 307 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation and Training Details", + "text_level": 1, + "bbox": [ + 497, + 325, + 821, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For training the proposed SR model, $48 \\times 48$ image patches with RGB channels are cropped as the model input. The data augmentation is implemented by random flipping and rotation. The Adam optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ is utilized to train the models. Besides, the mini-batch size is set to 16. The learning rate is initialized as $2e - 4$ . The pruned weights can be determined by IMPSSD each 300 epochs and the same setting is adopted in the model expansion stage. We use PyTorch to implement our models with a GTX TITAN GPU.", + "bbox": [ + 497, + 340, + 890, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Objective criteria, i.e., peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are adopted to evaluate the model performance. The two metrics are both calculated on the Y channel of the YCbCr space.", + "bbox": [ + 497, + 479, + 890, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 497, + 545, + 653, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we mainly give a detailed analysis of the proposed lottery-ticket stochastic self-distillation scalable framework (IMP-SSD) for image SR. We adopt EDSR [23] as the backbone network to implement our experiments.", + "bbox": [ + 497, + 561, + 890, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "IMP vs. OMP. To verify the superiority of IMP over OMP on the compression task of image SR, we conduct IMP and OMP on the forward stage of our method instantiated with three popular SR methods: EDSR [23], RDN [38], and RCAN [37]. As shown in Fig. 5, when the pruned network reaches a very small model size, usually as the sparsity of the model is higher than $80\\%$ , where the sparsity means the ratio of the pruning-out parameters, IMP does not significantly scarify the SR performance while OMP degrades a lot. Especially, it obtains 27.66dB in PSNR with $94\\%$ sparsity as shown in Table 1. OMP is obviously unable to make the RCAN model reach a very small model similar to that by IMP. That's why we pick IMP as our pruning algorithm, which demonstrates the advantages of LTH once again.", + "bbox": [ + 497, + 617, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of IMP-SSD. To demonstrate the effectiveness of IMP-SSD, we employ EDSR as the backbone and compare MSSR with EDSR. As shown in Table 2, our IMP-SSD method can not only compress a model to a very", + "bbox": [ + 497, + 809, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14403", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/35d0432593e88a5a019303e7df8500fa84f4f257b965d87108a297177e390f9f.jpg", + "table_caption": [ + "Table 3. Comparison of the quantitative results of average PSNR/SSIM with state-of-the-art efficient SR models." + ], + "table_footnote": [], + "table_body": "
ScaleMethodParamsSet5 PSNR/SSIMSet14 PSNR/SSIMB100 PSNR/SSIMUrban100 PSNR/SSIMDIV2K_valid PSNR/SSIM
4×EDSR (CVPRW' 2017)1518K32.09 / 0.893728.55 / 0.780627.54 / 0.734825.97 / 0.781930.38 / 0.8364
EDSR_IMP-SSD_88% (ours)190K31.89 / 0.891128.44 / 0.777727.46 / 0.732625.71 / 0.774330.25 / 0.8336
EDSR_IMP-SSD_94% (ours)95K31.74 / 0.887928.33 / 0.774927.39 / 0.729725.48 / 0.765730.12 / 0.8303
RCAN (ECCV' 2018)15592K32.19 / 0.895028.63 / 0.781927.57 / 0.735526.21 / 0.788830.40 / 0.8331
RCAN_IMP-SSD_88% (ours)1949K32.30 / 0.896328.72 / 0.784427.63 / 0.737926.32 / 0.793230.53 / 0.8396
RCAN_IMP-SSD_94% (ours)974K32.02 / 0.892428.49 / 0.778827.51 / 0.733425.87 / 0.780030.30 / 0.8347
RDN (CVPR' 2018)5579K32.24 / 0.896528.71 / 0.784127.64 / 0.738426.31 / 0.792830.55 / 0.8406
RDN_IMP-SSD_88% (ours)697K31.87 / 0.890228.38 / 0.776627.42 / 0.731325.53 / 0.768130.15 / 0.8315
RDN_IMP-SSD_97% (ours)174K31.36 / 0.882628.10 / 0.769527.23 / 0.724925.09 / 0.311629.86 / 0.8243
ENLCN (AAAI' 2022)1536K32.14 / 0.893928.59 / 0.781227.56 / 0.735226.05 / 0.784530.36 / 0.8360
ENLCN_IMP-SSD_88% (ours)192K31.97 / 0.891128.47 / 0.773127.47 / 0.732525.77 / 0.731330.22 / 0.8326
ENLCN_IMP-SSD_97% (ours)48K31.41 / 0.882528.09 / 0.769027.24 / 0.725025.15 / 0.313329.84 / 0.8239
", + "bbox": [ + 80, + 142, + 890, + 313 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5b194c0404c14e3f3a75261f68acf8bf1919e067c5c0fca31bc0d842eedeaa0.jpg", + "table_caption": [ + "Table 4. Comparison between the performance (PSNR) of the sparse sub-networks via IMP-SSD and the performance of the sparse sub-networks of the scalable network (SCL) for $4 \\times$ SR." + ], + "table_footnote": [], + "table_body": "
Prune (%)Set5Set14B100Urban100
IMP-SSDSCLIMP-SSDSCLIMP-SSDSCLIMP-SSDSCL
0 %32.0931.9428.5528.5327.5427.5325.9725.89
50 %32.0531.9628.5728.5127.5527.5025.9925.79
75 %32.0331.9228.5328.4627.5227.4825.9025.73
88 %31.8931.8228.4428.4027.4627.4325.7125.61
94 %31.7431.7428.3328.3327.3927.3925.4825.48
", + "bbox": [ + 80, + 372, + 465, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "small size but also mitigate the degradation in performance for different magnification factors $(2\\times, 3\\times,$ and $4\\times)$ . Besides, the results in Table 1 also illustrate that IMP-SSD further improves the PSNR to $28.33\\mathrm{dB}$ with the similar sparsity. Therefore, our method has significant effects on scalable model compression.", + "bbox": [ + 75, + 460, + 468, + 543 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "IMP-SSD SR vs. Scalable SR. In order to show the effectiveness of MSSR, we compare the scalable lightweight SR model obtained in the backward stage with the middle result by IMP-SSD in the forward stage. We use EDSR as the backbone, and make the comparison in five degrees of sparsities: $0\\%$ , $50\\%$ , $75\\%$ , $88\\%$ , $94\\%$ . In Table 4, IMP-SSD generates five pruned SR models corresponding to five sparsities, while scalable SR only has one model. With the increase of sparsity, the SR performance decreases, but the gap between two neighbor sparsities is not more than $0.2\\mathrm{dB}$ . Thus, the SR scalable model makes sense. As shown in Table 4, it is observed that scalable SR achieves comparable results to IMP-SSD on four datasets: Set5, Set14, B100, and Urban100. It shows that our method uses only one scalable model that can reach the level of the separately lightweight SR models at different sparse degrees. Therefore, MSSR is memory-friendly.", + "bbox": [ + 75, + 543, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Instantiating with Other SR methods", + "text_level": 1, + "bbox": [ + 75, + 788, + 396, + 802 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the proposed scalable lightweight framework (MSSR), we integrate it with four representative SR methods: EDSR [23], RCAN [37], RDN [38], and ENLCN [1]. All these SR models are re", + "bbox": [ + 75, + 809, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/19991bbb3be9b57595fd42f306a6ea0dfc245cb109d12da398bc5982b1132853.jpg", + "table_caption": [ + "Table 5. The analysis about IMP-SSD on different methods for $4 \\times$ image SR validated on Set14." + ], + "table_footnote": [], + "table_body": "
PSNR Sparsity\nModel0%50%75%88%94%97%99%
EDSR_IMP-SSD28.5528.5728.5328.4428.3328.0520.36
RDN_IMP-SSD28.7028.6828.5228.3828.2328.1027.46
RCAN_IMP-SSD28.6328.7728.7328.7228.4928.1527.68
ENLCN_IMP-SSD28.5928.6228.5628.4728.3028.0927.60
", + "bbox": [ + 502, + 363, + 890, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained under the same experimental settings. We show the results in Table 3. It shows that our method is universally applicable to the existing SR methods. Compared with the original models with the magnification factor $4 \\times$ , under the sparsities of $88 \\%$ and $94 \\% / 97 \\%$ , the maximum decreases of IMP-SSD instantiated with EDSR, RCAN, RDN, and ENLCN are (0.2dB, -0.1dB, 0.8dB, 0.3dB), and (0.5dB, 0.4dB, 0.9dB, 0.9dB), respectively. Note that, in RCAN, IMP-SSD is better than its original model. We guess that SSD could improve the SR results.", + "bbox": [ + 497, + 458, + 890, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 5, we give more comparisons of the SR performance with more different sparsity. It shows that all methods except EDSR, with the decrease of the model size, our method degrades gradually with not much range of decrease, their ranges are $1.3\\mathrm{dB}$ , 1dB, 0.9dB in RCAN, RDN, ENLCN, respectively. EDSR degrades significantly by $8.2\\mathrm{dB}$ with the sparsity of $99\\%$ .", + "bbox": [ + 497, + 597, + 890, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparison with State-of-the-art", + "text_level": 1, + "bbox": [ + 497, + 705, + 793, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of MSSR, we compare our approach (IMP-SSD) with several state-of-the-art lightweight SR methods: SRCNN [7], FSRCNN [8], CARN [19], and ESRN [31]. Note that IMP-SSD obtains the gradually shrinking lightweight SR models. The comparison results for $2 \\times$ , $3 \\times$ , and $4 \\times$ magnification factors are shown in Table 6. PSNR and SSIM are adopted as metrics to evaluate image quality by convention.", + "bbox": [ + 497, + 726, + 890, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It shows that our IMP-SSD achieves the best results in terms of the coordination of parameters and performance.", + "bbox": [ + 497, + 838, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14404", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ea85864f82dfc7c3e9deafe0b6e5d0f89a70d1d10c3652a0dc0baba3069a3093.jpg", + "table_caption": [ + "Table 6. Comparison of the quantitative results of average PSNR/SSIM with state-of-the-art efficient SR models." + ], + "table_footnote": [], + "table_body": "
ScaleMethodParamsSet5 PSNR/SSIMSet14 PSNR/SSIMB100 PSNR/SSIMUrban100 PSNR/SSIM
2×SRCNN [7] (ECCV' 2014)57K36.66 / 0.954232.42 / 0.906331.36 / 0.887929.50 / 0.8946
FSRCNN [8] (ECCV' 2016)12K37.00 / 0.955832.63 / 0.908831.53 / 0.892029.88 / 0.9020
CARN [19] (ECCV' 2018)415K37.68 / 0.959433.27 / 0.914931.97 / 0.896931.18 / 0.9186
ESRN [31] (AAAI' 2020)324K37.85 / 0.960033.42 / 0.916132.10 / 0.898731.79 / 0.9248
EDSR [23] (CVPRW' 2017)1370K37.98 / 0.960533.63 / 0.918032.17 / 0.899632.08 / 0.9280
EDSR_IMP-SSD.88% (ours)171K37.82 / 0.959933.44 / 0.916732.09 / 0.898731.69 / 0.9244
EDSR_IMP-SSD.94% (ours)85K37.51 / 0.958633.09 / 0.913731.88 / 0.896030.94 / 0.9163
3×SRCNN [7] (ECCV' 2014)57K32.75 / 0.909029.28 / 0.820928.41 / 0.786326.24 / 0.7989
FSRCNN [8] (ECCV' 2016)12K33.16 / 0.914029.43 / 0.824228.53 / 0.791026.43 / 0.8080
CARN [19] (ECCV' 2018)415K34.00 / 0.924030.07 / 0.836828.92 / 0.800727.53 / 0.8379
ESRN [31] (AAAI' 2020)324K34.23 / 0.926230.27 / 0.840029.03 / 0.803927.95 / 0.8481
EDSR [23] (CVPRW' 2017)1558K34.37 / 0.926730.31 / 0.840929.07 / 0.804328.05 / 0.8506
EDSR_IMP-SSD.88% (ours)195K34.15 / 0.925230.20 / 0.839229.00 / 0.802827.81 / 0.8449
EDSR_IMP-SSD.94% (ours)97K33.74 / 0.921629.92 / 0.834028.81 / 0.797627.16 / 0.8298
4×SRCNN [7] (ECCV' 2014)57K30.48 / 0.862827.49 / 0.750326.90 / 0.710124.52 / 0.7221
FSRCNN [8] (ECCV' 2016)12K30.71 / 0.865727.59 / 0.753526.98 / 0.715024.62 / 0.7280
CARN [19] (ECCV' 2018)415K31.78 / 0.889428.39 / 0.776227.42 / 0.730325.56 / 0.7674
ESRN [31] (AAAI' 2020)324K31.99 / 0.891928.49 / 0.777927.50 / 0.733125.87 / 0.7782
EDSR [23] (CVPRW' 2017)1518K32.09 / 0.893728.55 / 0.780627.54 / 0.734825.97 / 0.7819
EDSR_IMP-SSD.88% (ours)190K31.89 / 0.891128.44 / 0.777727.46 / 0.732625.71 / 0.7743
EDSR_IMP-SSD.94% (ours)95K31.74 / 0.887928.33 / 0.774927.39 / 0.729725.48 / 0.7657
", + "bbox": [ + 80, + 138, + 889, + 413 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d919ee3f6b8728f97d6403df8378b5a0642b399578e0f637a35a36ac50bfda7f.jpg", + "image_caption": [ + "Figure 6. Visual comparisons of IMP-SSD with other lightweight SR models on Urban100 for $4 \\times$ SR." + ], + "image_footnote": [], + "bbox": [ + 90, + 434, + 470, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In particular, compared with other lightweight SR models such as CARN and ESRN, our IMP-SSD in EDSR with the sparsity of $88\\%$ achieves comparable performance while our model size is about half or two-thirds of those. Compared with smaller SR models such as SRCNN and FSRCNN, IMP-SSD with the sparsity of $94\\%$ , IMP-SSD achieves significant gains in PSNR by at least 0.4dB, 0.3dB, 0.4dB for $2\\times$ , $3\\times$ , and $4\\times$ SR, respectively. The same conclusion is made in SSIM. Furthermore, compared with the original EDSR, IMP-SSD makes the SR model shrink to reach a very small size but their SR performance does not degrade very much. Compared with the original EDSR, the maximum decreases are 0.4dB, 0.2dB, 0.2dB for IMP-SSD with the sparsity of $88\\%$ and 0.6dB, 0.9dB, 0.5dB for IMP-SSD with the sparsity of $94\\%$ , respectively. It demonstrates that our method is effective on scalable SR.", + "bbox": [ + 75, + 644, + 468, + 865 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also give the comparison of our method in visual effect in Fig. 6. It is observed that our method achieves a better visual effect than other lightweight SR models even with larger model sparsity.", + "bbox": [ + 497, + 435, + 890, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 499, + 508, + 626, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel memory-friendly scalable lightweight framework (MSSR) with rewinding LTH, which can be deployed on arbitrary computational resources according to customization. MSSR is backtracking and includes two stages: the forward stage for model compression and the backward for model expansion. For the first time, LTH with rewinding weights is implemented to explore the winning tickets of deep SR models. We use IMP to gradually shrink the SR model and not only obtain the scalable lightweight model but also obtain the pruning-out masks which form the nested sets. In order to improve the performance of the shrinking model, SSD is conducted. Unlike the traditional SD, we replace the feature maps on several layers of Teacher with the counterparts of Students and conduct distillation. In the backward stage, we gradually recover the lightweight SR models depending on the pruning-out masks and then fine-tune the current model. We make instantiation on several state-of-the-art SR models. The extensive experimental results are effective.", + "bbox": [ + 497, + 533, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by grants from the National Key Research and Development Program of China (No.2020AAA0108301), National Natural Science Foundation of China (No.62176224, No.62222602), and CCF-Lenovo Blue Ocean Research Fund.", + "bbox": [ + 497, + 796, + 890, + 865 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14405", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 173, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Efficient non-local contrastive attention for image superresolution. In AAAI, 2022. 2, 7", + "[2] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017. 6", + "[3] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 2010. 6", + "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 6", + "[5] Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Zhangyang Wang, and Michael Carbin. The lottery ticket hypothesis for pre-trained bert networks. In NeurIPS, 2020. 3", + "[6] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2020. 2, 3", + "[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016. 7, 8", + "[8] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016. 7, 8", + "[9] Utku Evci, Fabian Pedregosa, Aidan Gomez, and Erich Elsen. The difficulty of training sparse neural networks. In ICLR, 2021. 3", + "[10] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 2, 3", + "[11] Jonathan Frankle, David J Schwab, and Ari S Morcos. The early phase of neural network training. In ICLR, 2020. 3", + "[12] Sharath Girish, Shishira R Maiya, Kamal Gupta, Hao Chen, Larry S Davis, and Abhinav Shrivastava. The lottery ticket hypothesis for object recognition. In CVPR, 2021. 3", + "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 3", + "[14] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. TPAMI, 2021. 2", + "[15] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In ICCV, 2017. 3", + "[16] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 6", + "[17] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM MM, 2019. 1, 2, 6", + "[18] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 1, 2" + ], + "bbox": [ + 78, + 147, + 468, + 865 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 1, 2, 7, 8", + "[20] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. 1", + "[21] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. In NeurIPS, 1989. 3", + "[22] Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Dynamic slimmable network. In CVPR, 2021. 2", + "[23] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 6, 7, 8", + "[24] Ning Liu, Geng Yuan, Zhengping Che, Xuan Shen, Xiaolong Ma, Qing Jin, Jian Ren, Jian Tang, Sijia Liu, and Yanzhi Wang. Lottery ticket preserves weight correlation: Is it desirable or not? In ICML, 2021. 2, 3, 4", + "[25] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 3", + "[26] Xiaotong Luo, Mingliang Dai, Yulun Zhang, Yuan Xie, Ding Liu, Yanyun Qu, Yun Fu, and Junping Zhang. Adjustable memory-efficient image super-resolution via individual kernel sparsity. In ACM MM, 2022. 2", + "[27] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, 2020. 2", + "[28] Sai Prasanna, Anna Rogers, and Anna Rumshisky. When bert plays the lottery, all tickets are winning. In ACL, 2020. 3", + "[29] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. In ICLR, 2020. 3", + "[30] Pedro Savarese, Hugo Silva, and Michael Maire. Winning the lottery with continuous sparsification. In NeurIPS, 2020. 3", + "[31] Dehua Song, Chang Xu, Xu Jia, Yiyi Chen, Chunjing Xu, and Yunhe Wang. Efficient residual dense block search for image super-resolution. In AAAI, 2020. 7, 8", + "[32] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. In ICLR, 2020. 3", + "[33] Shihui Yin, Kyu-Hyoun Kim, Jinwook Oh, Naigang Wang, Mauricio Serrano, Jae-Sun Seo, and Jungwook Choi. The sooner the better: Investigating structure of early winning lottery tickets. In ICLR, 2020. 3", + "[34] Haonan Yu, Sergey Edunov, Yuandong Tian, and Ari S Morcos. Playing the lottery with rewards and multiple languages: lottery tickets in rl and nlp. In ICLR, 2020. 3", + "[35] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In ICCV, 2010. 6", + "[36] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In ICCV, 2019. 4" + ], + "bbox": [ + 502, + 127, + 890, + 864 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14406", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 6, 7", + "[38] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 6, 7", + "[39] Hao Zhou, Jose M Alvarez, and Fatih Porikli. Less is more: Towards compact cnns. In ECCV, 2016. 3" + ], + "bbox": [ + 78, + 125, + 468, + 229 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14407", + "bbox": [ + 467, + 887, + 504, + 897 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_model.json b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..19a3bc752b10280e3933b794ce5b0f4803843a92 --- /dev/null +++ b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.792, + 0.087 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.193, + 0.162, + 0.784, + 0.205 + ], + "angle": 0, + "content": "Memory-friendly Scalable Super-resolution via Rewinding Lottery Ticket Hypothesis" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.228, + 0.788, + 0.246 + ], + "angle": 0, + "content": "Jin Lin\\(^{1*}\\), Xiaotong Luo\\(^{1*}\\), Ming Hong\\(^{1}\\), Yanyun Qu\\(^{1\\dagger}\\), Yuan Xie\\(^{2\\dagger}\\), Zongze Wu\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.246, + 0.714, + 0.262 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Informatics, Xiamen University, Fujian, China" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.262, + 0.86, + 0.297 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Computer Science and Technology, East China Normal University, Shanghai, China \n\\(^{3}\\)School of Mechatronics and Control Engineering, Shenzhen University, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.298, + 0.735, + 0.312 + ], + "angle": 0, + "content": "{linjin, xiaotluo, mingh}@stu.xmu.edu.cn, yyqu@xmu.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.315, + 0.638, + 0.327 + ], + "angle": 0, + "content": "yxie@cs.ecnu.edu.cn, zzwu@szu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.36, + 0.313, + 0.374 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.39, + 0.473, + 0.819 + ], + "angle": 0, + "content": "Scalable deep Super-Resolution (SR) models are increasingly in demand, whose memory can be customized and tuned to the computational recourse of the platform. The existing dynamic scalable SR methods are not memory-friendly enough because multi-scale models have to be saved with a fixed size for each model. Inspired by the success of Lottery Tickets Hypothesis (LTH) on image classification, we explore the existence of unstructured scalable SR deep models, that is, we find gradual shrinkage subnetworks of extreme sparsity named winning tickets. In this paper, we propose a Memory-friendly Scalable SR framework (MSSR). The advantage is that only a single scalable model covers multiple SR models with different sizes, instead of reloading SR models of different sizes. Concretely, MSSR consists of the forward and backward stages, the former for model compression and the latter for model expansion. In the forward stage, we take advantage of LTH with rewinding weights to progressively shrink the SR model and the pruning-out masks that form nested sets. Moreover, stochastic self-distillation (SSD) is conducted to boost the performance of sub-networks. By stochastically selecting multiple depths, the current model inputs the selected features into the corresponding parts in the larger model and improves the performance of the current model based on the feedback results of the larger model. In the backward stage, the smaller SR model could be expanded by recovering and fine-tuning the pruned parameters according to the pruning-out masks obtained in the forward. Extensive experiments show the effectiveness of MMSR. The smallest-scale sub-network could achieve the sparsity of \\(94\\%\\) and outperforms the compared lightweight SR methods." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.359, + 0.885, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.506, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Figure 1. The flowchart of scalable SR network. S_n and S_w denote the neurons and the neural connections (weights) of the simplest subnetwork; M_n and M_w denote the intermediate neurons and neural connections; Cur_n and Cur_w denote the specific neurons and neural connections belonging to the current subnetwork. Free_n denotes the pruning-out neurons. The final model (with \\(100\\%\\) recovered parameters) reaches the original size. The scalable model is adjustable to the memory resource allocation." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.634, + 0.631, + 0.648 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Single image super-resolution (SISR) aims to reconstruct a high-resolution (HR) image from the corresponding low-resolution (LR) one. With the rising of deep learning, deep SR methods have made incredible progress. However, the existing SR models mostly require computational and memory resources, so they do not favor resource-limited devices such as mobile phones, robotics, and some edge devices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.893, + 0.868 + ], + "angle": 0, + "content": "The lightweight SR methods are attracting more attention for better application to resource-limited devices. The existing lightweight SR methods mainly focus on designing compact architectures [17, 20] with a fixed size, such as multi-scale pyramid [20], multiple-level receptive fields [17, 18], and recursive learning [19]. However, most lightweight SR models with fixed sizes are not flexible in applications. If one model does not match the resources of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.843, + 0.203, + 0.854 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.855, + 0.225, + 0.867 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.843, + 0.225, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "14398" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.127, + 0.468, + 0.154 + ], + "angle": 0, + "content": "the platform, it has to be retrained by compression methods to match the resources and then reloaded onto the devices." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.156, + 0.468, + 0.405 + ], + "angle": 0, + "content": "The urgent demand to customize models based on deployment resources is increasing. Dynamic neural networks for SR [14,22] are proposed to adjust the network architecture according to different computational resources. The existing dynamic deep SR models often explore dynamic depth or width [22, 26], but they either require large memory resources or are not convenient for users to wait for retraining another SR model. The former leads to saving the multi-scale SR models of different sizes and the latter leads to retraining the model before being reused again. The limitation lies in that they are not memory-friendly. In many edge-device SR applications, the devices may be scalable, that is, their memories may be small in the beginning and be expanded later. Thus, we discuss two issues in this paper: 1) how to make a scalable lightweight model for the multi-scale computational recourse. 2) how to make the lightweight model expand to a larger-size model for better performance if the computational recourse is increased." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.407, + 0.468, + 0.614 + ], + "angle": 0, + "content": "As for the first issue, inspired by the success of Lottery Ticket Hypothesis [10] which points out that there could exist a properly pruned sub-network named winning tickets to achieve comparable performance against the original dense network in model compression of classification, it is used to find the sub-network for SR. We are the first to study the existence of scalable winning tickets for SR. Iterative pruning and rewinding weights in LTH are beneficial to the scalable lightweight SR model. Iterative pruning may compress the SR model according to an arbitrary size. It is observed in [24] that the winning tickets are related to an insufficient DNN, and rewinding LTH outperforms the original LTH. That is, the initial weights in LTH are replaced with the T-iteration weights during pruning and fine-tuning. The scalable deep SR model is shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.616, + 0.468, + 0.699 + ], + "angle": 0, + "content": "As for the second issue, the scalable SR model can customize parameters to adapt to different memory resources rather than load or offload different models for different devices. In other words, during real applications, there will be only one simple model to be employed for inference whose size is decided by the computational resource." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.701, + 0.468, + 0.866 + ], + "angle": 0, + "content": "In this paper, we propose a memory-friendly scalable deep SR model (MSSR) via rewinding LTH. We use the rewinding LTH [10] to generate our unstructured scalable mask. MSSR is backtracking and contains forward and backward stages. The former focuses on model compression by rewinding LTH with iterative pruning and finetuning, and the latter focuses on iterative model expansion until it goes back to its original size. Multi-scale winning tickets together with the pruning-out masks are obtained by rewinding LTH in the forward stage with the decrease in the number of parameters. The pruning-out masks are nested. In order to make the compressed SR model not degrade sig" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.892, + 0.209 + ], + "angle": 0, + "content": "nificantly, stochastic self-distillation (SSD) is used to improve the representation of the small-scale SR model, and knowledge is transferred from the last-scale model to the current scale model. In the backward stage, the smallest model is expanded gradually to the model with the original size with the expanded mask." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.211, + 0.855, + 0.224 + ], + "angle": 0, + "content": "The main contributions of this work are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.232, + 0.891, + 0.301 + ], + "angle": 0, + "content": "- A memory-friendly scalable dynamic SR lightweight model via rewinding LTH is proposed. MSSR is re-configurable and switchable to sub-networks with different sizes according to on-device resource constraints on the fly." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.31, + 0.892, + 0.393 + ], + "angle": 0, + "content": "- MSSR is backtracking, which contains forward and backward stages. Multi-scale winning tickets form nested masks for the multi-scale models. SSD is conducted by replacing the features in randomly selected layers between Teacher and Student to improve the performance of the scalable SR lightweight models." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.402, + 0.891, + 0.443 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that MSSR can generalize to different SR models as well as state-of-the-art attention-based models, ENLCN [1]." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.232, + 0.892, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.642, + 0.474 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.824, + 0.498 + ], + "angle": 0, + "content": "2.1. Lightweight Deep Learning-based SR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.727 + ], + "angle": 0, + "content": "SISR aims to generate a high-resolution image from the low-resolution version. Recently, efficient and lightweight SISR networks have attracted increasing interest in the computer vision community. The lightweight SR methods consist of static and dynamic neural networks for SR. The earlier lightweight SR methods mainly focus on the static network structure design. IDN [18] utilizes group convolution and combined the local long and short path features effectively. CARN [19] utilizes a recursive cascading mechanism for learning multi-level feature representations. IMDN [17] introduces information multi-distillation blocks for extracting hierarchical features. FALSR [6] applies neural architecture search for efficient network design automatically. LatticeNet [27] designs a lattice block to combine two residual blocks adaptively. However, the aforementioned lightweight SR models need elaborate designs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Recently, dynamic lightweight SR models are demanded increasingly whose size is adjustable according to different computational resources. DS-Net [22] makes the width of the SR model adjustable, and IKS [26] makes the sparsity of individual kernels adjustable. However, they require saving lightweight SR models with different shrinkage sizes, which is not convenient for users to offload the old model and then load a new one. Our method uses a scalable SR model adjustable to the changing computational recourse to avoid the loading and offloading models." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "14399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.125, + 0.885, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.444, + 0.893, + 0.509 + ], + "angle": 0, + "content": "Figure 2. The architecture of memory-friendly scalable deep SR model (MSSR). It contains two stages: the forward stage for model compression and the backward stage for model expansion. In the forward stage, the original SR network is firstly processed by IMP to obtain winning tickets and the pruning-out mask \\( LM_{1} \\), and then the winning tickets are processed by SSD and the lightweight subnetwork is obtained. So do the other gradual shrinkage sub-networks and \\( \\{LM_{i}\\} \\) forms the nested sets. In the backward stage, according to the pruning-out masks, the smaller lightweight SR model is gradually expanded and finely tuned to obtain a larger lightweight SR model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.525, + 0.414, + 0.541 + ], + "angle": 0, + "content": "2.2. Pruning and Lottery Ticket Hypothesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.783 + ], + "angle": 0, + "content": "Model pruning is to prune away the unimportant connections (parameters) [21] in the neural network which trades off between the model size and performance. The existing pruning methods can be classified into unstructured [6, 13, 21] and structured methods [15, 25, 39]. The structured pruning methods remove the substructures of the model, such as channels, filters, and layers, while the unstructured pruning methods are magnitude-based, where the weights are removed below the threshold. However, magnitude-based pruning methods are not hardware-friendly and computationally efficient. Lottery Ticket Hypothesis (LTH) [10] opened a pandas's box of immense possibilities in the field of pruning and sparse models. It assumes that there exist sparse sub-networks independently training from scratch that can match the performance of dense networks. Some literatures discuss the effectiveness of pruning, fine-tuning [24], and rewinding weights [29] for LTH." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.471, + 0.868 + ], + "angle": 0, + "content": "LTH has been widely applied in image classification [9-12, 30, 32, 33], natural language processing [5, 28, 29, 34], and so on. However, few works discuss the existence of LTH in the deep model for image restoration. We are the first to investigate the existence of scalable winning tickets for lightweight deep SR models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.591, + 0.538 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.684, + 0.563 + ], + "angle": 0, + "content": "3.1. Overview of MSSR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.644 + ], + "angle": 0, + "content": "In this section, we introduce the memory-friendly scalable SR model which is an all-in-one model with multiscale SR models sharing a model. As shown in Fig. 2, our MSSR consists of two stages: the forward stage for model compression, and the backward stage for model expansion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.893, + 0.868 + ], + "angle": 0, + "content": "In the forward stage, we use rewinding LTH (RLTH) to iteratively shrink a dense SR model to obtain a sub-network with the small size not significantly scarifying the performance. During pruning by LTH each time, we treat the current lightweight SR model as Student and its last ancestor lightweight SR model as Teacher, and conduct stochastic self-distillation between Teacher and Student. In Fig. 2, the dark-blue block denotes the winning tickets, and the white block denotes the pruning part. We also record the mask of pruning part and use a matrix with \"1\" to represent the pruning-out weights and with \"0\" for others. It is observed that the masks are nested and the initial mask is continuously contained in the later mask. In the backward stage, we iteratively expand the lightweight sub-network according to the saved nested masks. The most lightweight network marked by 1 is expanded iteratively to the gradual less" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "14400" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.47, + 0.154 + ], + "angle": 0, + "content": "lightweight network, the expanding order is shown as \"1\", \"1-2\", \"1-2-3\", and \"1-2-3-4\"." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.164, + 0.429, + 0.18 + ], + "angle": 0, + "content": "3.2. The Forward Stage: Model Compression" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.47, + 0.406 + ], + "angle": 0, + "content": "The forward streamline is shown in Fig. 3. Let a network denoted by \\( f(\\theta_0; x) \\) and the pruned masks which form the \"nested sets\" denoted by \\( LM = \\{\\} \\), where \\( \\theta_0 \\) denotes the initial weights. The flowchart shows a whole process in an iteration for shrinking a model and it contains three components: SR model pre-train, IMP with rewinding weights, and SSD. Firstly, an SR network is trained to reach a steady state, and its weights are named the first nonpruning weights \\( \\theta \\). After that, Iterative Magnitude Pruning is conducted on the SR model \\( \\theta \\), and the weights below the pre-set threshold are pruned out which forms a pruning mask \\( m \\in \\{0,1\\}^n \\). Following [24], we rewind the remaining weights to their values in \\( \\theta \\), i.e., \\( \\theta' = \\theta \\odot m \\). Then, we retrain the pruned weights \\( \\theta' \\) by SSD arriving at \\( \\hat{\\theta} \\). So do the next more lightweight SR model with the model weights \\( \\hat{\\theta} \\) as the initiation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.408, + 0.469, + 0.491 + ], + "angle": 0, + "content": "LTH with Rewinding for SR. Benefiting from LTHs, we implement iterative magnitude pruning with rewinding to gradually shrink the SR model. The motivation lies in the controllability and flexibility of LTH. We can shrink the model arbitrarily by LTH, which favors the scalable lightweight SR model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.656 + ], + "angle": 0, + "content": "We mainly consider two issues: the model pruning process and the pruning rules. The pruning process corresponds to the process of gradually shrinking the model, and the pruning methods need to meet the following rules: 1) the amount of prune-out parameters is arbitrarily adjustable; 2) the model is shrunk gradually, which leads to the nested sets and the larger model contains its smaller descendant. LTH fits well with the needs of our scalable mask generator and the appropriate use of its task settings to implement the model compression process, in particular discarding its over parameterization, a feature that is not suitable for intensive prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.469, + 0.727 + ], + "angle": 0, + "content": "IMP is proven to be effective for LTH. In [24], it is proved that weight rewinding is better than the original weight initiation. That is, unpruned weights are rewound to their values from \\( T \\)-iteration training and retrained from there using the original training schedule." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.469, + 0.852 + ], + "angle": 0, + "content": "Compared with one-shot magnitude pruning (OMP) which reaches the goal model at one time, IMP does not shrink a model to reach the pre-set size at one time but shrinks a model with several iterations and the model size is evenly decreased at each iteration. Let the network be pruned for \\( T \\) times and be expected to discard \\( p\\% \\) of the parameters. The weights/parameters of the network are represented by \\( \\omega \\in \\mathbb{R}^n \\) and the pruning mask by \\( m \\in \\{0,1\\}^n \\) where \\( n \\) represents all the capacity of the network." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.853, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Weight rewinding does not initiate the model with the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.124, + 0.88, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Figure 3. The flowchart of the whole process of IMP and SSD. The initial model \\(\\theta_0\\) is pre-trained to obtain a stable model \\(\\theta\\). IMP is implemented on \\(\\theta\\) and the winning ticket (sub-network) is obtained. The sub-network is initiated by \\(\\theta \\odot m\\) and is processed by SSD for the better model \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.533 + ], + "angle": 0, + "content": "original initiation weights \\(\\theta_0\\odot m\\) , but with the \\(t\\) -iteration \\(\\theta_t\\odot m\\) where \\(0 < t\\leq T\\) , resulting in weights \\(\\theta_{t}\\) and the network function \\(f(x;\\theta_t\\odot m)\\) . We highlight that the original LTH always chooses \\(\\theta_0\\) as dense model, while in this work we choose \\(\\theta_{t}\\) from the last pruning iteration. After \"Pruning\" and \"Rewinding and Retraining\", \\(\\theta = \\hat{\\theta}\\) is updated and the mask at the \\(t\\) -th iteration is obtained which is denoted as \\(LM_{t} = LM_{t - 1}\\cup \\{m\\}\\) . In other words, the mask \\(LM_{t}\\) at \\(t\\) -th iteration is contained in the mask \\(LM_{t - 1}\\) at \\((t - 1)\\) -th. \"Pruning\" and \"Rewinding and Retraining\" are repeated until the model reaches the end condition, which is the iteration steps \\(T\\) or target sparsity \\(p\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Stochastic Self-distillation for SR. In the forward stage, due to the progressive shrinking of the SR model, it is definitely difficult to maintain the performance as same as the original model. Especially, when only \\(1\\%\\) of the parameters are kept, most magnitude-pruning methods do not work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.824 + ], + "angle": 0, + "content": "In order to mitigate the degradation of lightweight SR models, we adopt self-distillation [36] to boost the performance of the pruned model. Unlike the existing self-distillation which uses the function of features obtained by the model at previous iterations such as the average and summation function, we introduce attribution analysis to self-distillation. We randomly select several layers in Student, and the feature maps of selected layers replace the counterpart of Teacher, which is named Stochastic self-distillation (SSD). Concretely, the current lightweight SR model is regarded as the student network (Student), and its last ancestor network is regarded as the teacher network (Teacher), and Student is obtained by pruning Teacher. SSD is conducted between Teacher and Student. Student is the pruned model of Teacher. Let \\( X \\) and \\( Y \\) denote the network input and ground truth, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.867 + ], + "angle": 0, + "content": "The architecture of SSD is shown in Fig. 4, which contains two streamlines: Teacher and Student. The knowledge in feature maps is transferred from Teacher to Student." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.504, + 0.898 + ], + "angle": 0, + "content": "14401" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.124, + 0.437, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.308, + 0.471, + 0.384 + ], + "angle": 0, + "content": "Figure 4. The flowchart of SSD. Student is the sub-network by pruning Teacher. Several layers of Student are selected randomly. The feature maps in a selected layer replace the counterpart of Teacher which outputs the SR result, and so do other selected layers. The total loss is the summation of the knowledge distillation loss \\( L_{skd} \\) and the fidelity loss \\( L_{1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.469, + 0.514 + ], + "angle": 0, + "content": "If Student contains \\(d\\) layers, at each-iteration distillation, \\(s\\) layers randomly are selected. In realization, a one-hot vector is randomly generated, named Switch Vector, with a length of \\(d\\) in each training, \\(\\alpha \\in \\{0,1\\}^d\\). If the element is equal to 1 in \\(\\alpha\\), the feature maps in the layer corresponding to this position are selected, and are named Lucky feature maps. Then, Lucky feature maps in the selected layers to replace the counterparts of Teacher." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.469, + 0.557 + ], + "angle": 0, + "content": "In details, Student firstly is trained for a stable SR model. After that, SSD is conducted. At each iteration, Switch vector is dynamically generated, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.559, + 0.469, + 0.573 + ], + "angle": 0, + "content": "\\[\nD C = \\left[ \\alpha_ {1}, \\alpha_ {2}, \\dots , \\alpha_ {d} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.575, + 0.469, + 0.63 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) denotes the \\(i\\)-th element corresponding to the \\(i\\)-th layer and \\(d\\) is the layer number in Student. Then, lucky feature maps of Student are selected according to \\(\\alpha_{i}\\), and they are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.632, + 0.469, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L F} = \\left[ f _ {\\alpha_ {1}} ^ {s}, f _ {\\alpha_ {2}} ^ {s}, \\dots , f _ {\\alpha_ {d}} ^ {s} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.652, + 0.469, + 0.692 + ], + "angle": 0, + "content": "where \\( f_{\\alpha}^{s} \\) denotes the feature obtained from Student. Note that all elements in Switch Vector are independent and identically distributed." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.696, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Lucky feature maps replace the counterpart of Teacher each time, and the corresponding output of new Teacher is denoted by \\( Y_{\\alpha_i}^t \\), and the corresponding weights of new Teacher are denoted by \\( \\theta_t \\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.752, + 0.469, + 0.783 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} Y _ {\\alpha_ {i}} ^ {t} = F \\left(\\theta^ {t}; x, f _ {\\alpha_ {i}} ^ {s}\\right), i = 1, \\dots , d, \\\\ \\mathrm {L Y} = \\left[ Y _ {\\alpha_ {1}} ^ {t}, Y _ {\\alpha_ {2}} ^ {t}, \\dots , Y _ {\\alpha_ {d}} ^ {t} \\right], \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.469, + 0.838 + ], + "angle": 0, + "content": "where \\(LY\\) denotes the set of SR outputs, \\(Y_{\\alpha}^{t}\\) is the SSD output of lucky maps, \\(f_{\\alpha}^{s}\\). All the outputs between Student and Teacher are supervised by HR image \\(Y\\). The knowledge of Teacher takes more rationality to be distilled into Student." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Therefore, the total loss function utilized to train the SR model consists of the supervised loss \\( L_{1} \\) and self-distillation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.126, + 0.89, + 0.154 + ], + "angle": 0, + "content": "Algorithm 1 Memory-friendlyly scalable lightweight framework for image SR." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.156, + 0.792, + 0.169 + ], + "angle": 0, + "content": "Data: HR and LR image datasets: \\(D_{hr}\\), \\(D_{lr}\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.17, + 0.852, + 0.196 + ], + "angle": 0, + "content": "Input: Dense network with randomly initial weights: \\( f(\\theta_0) \\), pruning rounds: \\( T \\), and pruning target rate: \\( p \\)" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.197, + 0.769, + 0.21 + ], + "angle": 0, + "content": "Output: Scalable lightweight SR model" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.213, + 0.824, + 0.226 + ], + "angle": 0, + "content": "1: /* The forward stage: model compression */" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.227, + 0.888, + 0.254 + ], + "angle": 0, + "content": "2: Initializing: Pruned mask \\( m = \\{1\\}^n \\), lottery mask pool \\( LM = \\{m\\} \\) and epochs of each training \\( N \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.255, + 0.805, + 0.282 + ], + "angle": 0, + "content": "3: Pre-training: Train network for \\(N\\) epochs \\(f(\\theta_0;x)\\longrightarrow f(\\theta ;x)\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.283, + 0.656, + 0.294 + ], + "angle": 0, + "content": "4: while \\( i < T - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.295, + 0.86, + 0.31 + ], + "angle": 0, + "content": "5: Prune bottom \\( p\\% \\times \\frac{1}{T} \\) of \\( m \\odot \\theta \\) and update \\( m \\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.311, + 0.884, + 0.337 + ], + "angle": 0, + "content": "6: Re-train the pruned sub-network \\(\\theta^{\\prime} = m\\odot \\theta\\) for \\(N\\) epochs \\(f(\\theta^{\\prime};x)\\longrightarrow f(\\hat{\\theta};x)\\) with \\(SSD(\\theta ,\\theta^{\\prime};x)\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.338, + 0.88, + 0.363 + ], + "angle": 0, + "content": "7: Update mask pool \\( LM = LM \\cup \\{m\\} \\) and weights \\( \\theta = \\hat{\\theta} \\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.365, + 0.629, + 0.377 + ], + "angle": 0, + "content": "8: \\(i\\gets i + 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.379, + 0.6, + 0.391 + ], + "angle": 0, + "content": "9: end while" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.393, + 0.875, + 0.42 + ], + "angle": 0, + "content": "10: Generate the sparse sub-network matching the target sparsity, and the mask pool \\(LM\\);" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.421, + 0.818, + 0.433 + ], + "angle": 0, + "content": "11: /* The backward stage: model expansion */" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.435, + 0.817, + 0.461 + ], + "angle": 0, + "content": "12: Initializing: Current step \\( t = T \\) and sparsest sub-network \\( \\theta_T = \\theta \\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.462, + 0.629, + 0.473 + ], + "angle": 0, + "content": "13: while \\( t > 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.476, + 0.859, + 0.503 + ], + "angle": 0, + "content": "14: Calculate the grad mask \\( m = LM_{t - 1} - LM_t \\) to determine the trainable weights;" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.504, + 0.851, + 0.531 + ], + "angle": 0, + "content": "15: Fine-tune/re-train the trainable weights for \\(N\\) epochs \\(f(\\theta_t;x)\\longrightarrow f(\\theta_t';x)\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.532, + 0.659, + 0.545 + ], + "angle": 0, + "content": "16: Update \\(\\theta_t = \\theta_t'\\);" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.546, + 0.632, + 0.557 + ], + "angle": 0, + "content": "17: \\(t\\gets t - 1;\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.559, + 0.599, + 0.57 + ], + "angle": 0, + "content": "18: end while" + }, + { + "type": "list", + "bbox": [ + 0.507, + 0.213, + 0.888, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.599, + 0.727, + 0.613 + ], + "angle": 0, + "content": "loss \\(L_{skd}\\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.623, + 0.89, + 0.668 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = L _ {1} + \\lambda L _ {s k d} = \\left\\| Y ^ {s} - Y \\right\\| _ {1} + \\frac {\\lambda}{d} \\sum_ {i = 1} ^ {d} \\alpha_ {i} \\left\\| Y _ {\\alpha_ {i}} ^ {t} - Y \\right\\| _ {1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.669, + 0.892, + 0.738 + ], + "angle": 0, + "content": "where \\(\\lambda\\) denotes the regularization parameter, which is empirically set as 0.1 in our experiments. During inference, the SR model \\(\\theta_{p}\\) can be customized for various resources according to different sparsity \\(p\\). After the network training, a lottery mask pool can be built for the backward stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.748, + 0.842, + 0.764 + ], + "angle": 0, + "content": "3.3. The Backward Stage: Model Expansion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Our intention is to obtain a scalable network that can deploy according to customized memory resources. In this subsection, we introduce how to gradually recover a lightweight SR model, which is a backtracking process of IMP. We start from the most lightweight SR model and end with the original SR model. Corresponding to gradually shrinking the SR model with LTH, we solve two issues in" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14402" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.121, + 0.458, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.248, + 0.378, + 0.261 + ], + "angle": 0, + "content": "Network Sparsity (\\% pruned)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.326 + ], + "angle": 0, + "content": "Figure 5. Comparisons of IMP and OMP on various famous SR methods, which are evaluated on Set14 for \\(4 \\times\\) SR. The performance of the baseline dense network, the sparse sub-network obtained by IMP, as well as the sub-network obtained by OMP. The sub-network is with only \\(20\\%\\) of the weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.398 + ], + "angle": 0, + "content": "the model expansion: iterative model expansion (IME) and the expansion rule, the former refers to the requirement of the IME process and the latter refers to the requirement of the learning rule." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.565 + ], + "angle": 0, + "content": "Let's focus on an expansion from the \\(t\\)-th lightweight model to the \\((t - 1)\\)-th lightweight model where \\(t\\) comes from the definition in the forward stage. The larger \\(t\\) is, the smaller model is. During IMP, we obtain two pruned-out masks corresponding to the \\(t\\)-th and \\((t - 1)\\)-th lightweight models, denoted by \\(\\mathrm{LM}_t\\) and \\(\\mathrm{LM}_{t - 1}\\), respectively. Their difference represents the expanded positions, i.e., \\(m = LM_{t - 1} - LM_t\\). Without changing the \\(t\\)-th model, we recover the weights according to the mask \\(m\\). After that, we fine-tune the recovered model and obtain the \\((t - 1)\\)-th model. So do the next model expansion. The detailed procedure is described in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.578, + 0.21, + 0.594 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.181, + 0.615 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.734 + ], + "angle": 0, + "content": "DIV2K [2] is adopted to train the SR model, which is a high-quality natural scene dataset widely used for image SR tasks. The whole dataset includes 800 training images and 100 validation images totally with abundant texture and contents. The bicubic degraded LR images are synthesized following the previous works [17, 37]. Besides, the trained models are evaluated on four public SR benchmark datasets: Set5 [4], Set14 [35], B100 [3], and Urban100 [16]." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.746, + 0.47, + 0.786 + ], + "angle": 0, + "content": "Table 1. Ablation studies about the iterative magnitude pruning (IMP) and the stochastic self-distillation learning (SSD). All these models are validated on Set14 for \\(4 \\times\\) image SR." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.79, + 0.47, + 0.856 + ], + "angle": 0, + "content": "
EDSR基礎IMP (LTH)IMP RewindingSSDSparsityPSNR
✓×××0%28.55
✓✓××94%27.66
✓×✓×94%28.12
✓×✓✓94%28.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.124, + 0.892, + 0.151 + ], + "angle": 0, + "content": "Table 2. The performance of the sparse sub-network pruned by IMP-SSD on Set14 for \\(2 \\times, 3 \\times, 4 \\times\\) image SR." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.156, + 0.892, + 0.308 + ], + "angle": 0, + "content": "
ScaleModelSparsityPSNR (dB)Params
2×EDSR基礎0%33.631370K
EDSR_IMP94%32.8885K
EDSR_IMP-SSD94%33.09 (+0.21)85K
3×EDSR基礎0%30.301558K
EDSR_IMP94%29.7397K
EDSR_IMP-SSD94%29.92 (+0.19)97K
4×EDSR基礎0%28.551518K
EDSR_IMP94%28.1295K
EDSR_IMP-SSD94%28.33 (+0.21)95K
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.326, + 0.822, + 0.34 + ], + "angle": 0, + "content": "4.2. Implementation and Training Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.479 + ], + "angle": 0, + "content": "For training the proposed SR model, \\(48 \\times 48\\) image patches with RGB channels are cropped as the model input. The data augmentation is implemented by random flipping and rotation. The Adam optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\) is utilized to train the models. Besides, the mini-batch size is set to 16. The learning rate is initialized as \\(2e - 4\\). The pruned weights can be determined by IMPSSD each 300 epochs and the same setting is adopted in the model expansion stage. We use PyTorch to implement our models with a GTX TITAN GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Objective criteria, i.e., peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are adopted to evaluate the model performance. The two metrics are both calculated on the Y channel of the YCbCr space." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.546, + 0.654, + 0.561 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Next, we mainly give a detailed analysis of the proposed lottery-ticket stochastic self-distillation scalable framework (IMP-SSD) for image SR. We adopt EDSR [23] as the backbone network to implement our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.811 + ], + "angle": 0, + "content": "IMP vs. OMP. To verify the superiority of IMP over OMP on the compression task of image SR, we conduct IMP and OMP on the forward stage of our method instantiated with three popular SR methods: EDSR [23], RDN [38], and RCAN [37]. As shown in Fig. 5, when the pruned network reaches a very small model size, usually as the sparsity of the model is higher than \\(80\\%\\), where the sparsity means the ratio of the pruning-out parameters, IMP does not significantly scarify the SR performance while OMP degrades a lot. Especially, it obtains 27.66dB in PSNR with \\(94\\%\\) sparsity as shown in Table 1. OMP is obviously unable to make the RCAN model reach a very small model similar to that by IMP. That's why we pick IMP as our pruning algorithm, which demonstrates the advantages of LTH once again." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Effectiveness of IMP-SSD. To demonstrate the effectiveness of IMP-SSD, we employ EDSR as the backbone and compare MSSR with EDSR. As shown in Table 2, our IMP-SSD method can not only compress a model to a very" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "14403" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.151, + 0.124, + 0.816, + 0.138 + ], + "angle": 0, + "content": "Table 3. Comparison of the quantitative results of average PSNR/SSIM with state-of-the-art efficient SR models." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.143, + 0.891, + 0.314 + ], + "angle": 0, + "content": "
ScaleMethodParamsSet5 PSNR/SSIMSet14 PSNR/SSIMB100 PSNR/SSIMUrban100 PSNR/SSIMDIV2K_valid PSNR/SSIM
4×EDSR (CVPRW' 2017)1518K32.09 / 0.893728.55 / 0.780627.54 / 0.734825.97 / 0.781930.38 / 0.8364
EDSR_IMP-SSD_88% (ours)190K31.89 / 0.891128.44 / 0.777727.46 / 0.732625.71 / 0.774330.25 / 0.8336
EDSR_IMP-SSD_94% (ours)95K31.74 / 0.887928.33 / 0.774927.39 / 0.729725.48 / 0.765730.12 / 0.8303
RCAN (ECCV' 2018)15592K32.19 / 0.895028.63 / 0.781927.57 / 0.735526.21 / 0.788830.40 / 0.8331
RCAN_IMP-SSD_88% (ours)1949K32.30 / 0.896328.72 / 0.784427.63 / 0.737926.32 / 0.793230.53 / 0.8396
RCAN_IMP-SSD_94% (ours)974K32.02 / 0.892428.49 / 0.778827.51 / 0.733425.87 / 0.780030.30 / 0.8347
RDN (CVPR' 2018)5579K32.24 / 0.896528.71 / 0.784127.64 / 0.738426.31 / 0.792830.55 / 0.8406
RDN_IMP-SSD_88% (ours)697K31.87 / 0.890228.38 / 0.776627.42 / 0.731325.53 / 0.768130.15 / 0.8315
RDN_IMP-SSD_97% (ours)174K31.36 / 0.882628.10 / 0.769527.23 / 0.724925.09 / 0.311629.86 / 0.8243
ENLCN (AAAI' 2022)1536K32.14 / 0.893928.59 / 0.781227.56 / 0.735226.05 / 0.784530.36 / 0.8360
ENLCN_IMP-SSD_88% (ours)192K31.97 / 0.891128.47 / 0.773127.47 / 0.732525.77 / 0.731330.22 / 0.8326
ENLCN_IMP-SSD_97% (ours)48K31.41 / 0.882528.09 / 0.769027.24 / 0.725025.15 / 0.313329.84 / 0.8239
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.334, + 0.47, + 0.372 + ], + "angle": 0, + "content": "Table 4. Comparison between the performance (PSNR) of the sparse sub-networks via IMP-SSD and the performance of the sparse sub-networks of the scalable network (SCL) for \\(4 \\times\\) SR." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.374, + 0.466, + 0.449 + ], + "angle": 0, + "content": "
Prune (%)Set5Set14B100Urban100
IMP-SSDSCLIMP-SSDSCLIMP-SSDSCLIMP-SSDSCL
0 %32.0931.9428.5528.5327.5427.5325.9725.89
50 %32.0531.9628.5728.5127.5527.5025.9925.79
75 %32.0331.9228.5328.4627.5227.4825.9025.73
88 %31.8931.8228.4428.4027.4627.4325.7125.61
94 %31.7431.7428.3328.3327.3927.3925.4825.48
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.469, + 0.544 + ], + "angle": 0, + "content": "small size but also mitigate the degradation in performance for different magnification factors \\((2\\times, 3\\times,\\) and \\(4\\times)\\). Besides, the results in Table 1 also illustrate that IMP-SSD further improves the PSNR to \\(28.33\\mathrm{dB}\\) with the similar sparsity. Therefore, our method has significant effects on scalable model compression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.78 + ], + "angle": 0, + "content": "IMP-SSD SR vs. Scalable SR. In order to show the effectiveness of MSSR, we compare the scalable lightweight SR model obtained in the backward stage with the middle result by IMP-SSD in the forward stage. We use EDSR as the backbone, and make the comparison in five degrees of sparsities: \\(0\\%\\), \\(50\\%\\), \\(75\\%\\), \\(88\\%\\), \\(94\\%\\). In Table 4, IMP-SSD generates five pruned SR models corresponding to five sparsities, while scalable SR only has one model. With the increase of sparsity, the SR performance decreases, but the gap between two neighbor sparsities is not more than \\(0.2\\mathrm{dB}\\). Thus, the SR scalable model makes sense. As shown in Table 4, it is observed that scalable SR achieves comparable results to IMP-SSD on four datasets: Set5, Set14, B100, and Urban100. It shows that our method uses only one scalable model that can reach the level of the separately lightweight SR models at different sparse degrees. Therefore, MSSR is memory-friendly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.789, + 0.398, + 0.804 + ], + "angle": 0, + "content": "4.4. Instantiating with Other SR methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.867 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the proposed scalable lightweight framework (MSSR), we integrate it with four representative SR methods: EDSR [23], RCAN [37], RDN [38], and ENLCN [1]. All these SR models are re" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.334, + 0.892, + 0.36 + ], + "angle": 0, + "content": "Table 5. The analysis about IMP-SSD on different methods for \\(4 \\times\\) image SR validated on Set14." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.364, + 0.891, + 0.44 + ], + "angle": 0, + "content": "
PSNR Sparsity\nModel0%50%75%88%94%97%99%
EDSR_IMP-SSD28.5528.5728.5328.4428.3328.0520.36
RDN_IMP-SSD28.7028.6828.5228.3828.2328.1027.46
RCAN_IMP-SSD28.6328.7728.7328.7228.4928.1527.68
ENLCN_IMP-SSD28.5928.6228.5628.4728.3028.0927.60
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.596 + ], + "angle": 0, + "content": "trained under the same experimental settings. We show the results in Table 3. It shows that our method is universally applicable to the existing SR methods. Compared with the original models with the magnification factor \\(4 \\times\\), under the sparsities of \\(88 \\%\\) and \\(94 \\% / 97 \\%\\), the maximum decreases of IMP-SSD instantiated with EDSR, RCAN, RDN, and ENLCN are (0.2dB, -0.1dB, 0.8dB, 0.3dB), and (0.5dB, 0.4dB, 0.9dB, 0.9dB), respectively. Note that, in RCAN, IMP-SSD is better than its original model. We guess that SSD could improve the SR results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.695 + ], + "angle": 0, + "content": "In Table 5, we give more comparisons of the SR performance with more different sparsity. It shows that all methods except EDSR, with the decrease of the model size, our method degrades gradually with not much range of decrease, their ranges are \\(1.3\\mathrm{dB}\\), 1dB, 0.9dB in RCAN, RDN, ENLCN, respectively. EDSR degrades significantly by \\(8.2\\mathrm{dB}\\) with the sparsity of \\(99\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.706, + 0.794, + 0.72 + ], + "angle": 0, + "content": "4.5. Comparison with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.892, + 0.838 + ], + "angle": 0, + "content": "To validate the effectiveness of MSSR, we compare our approach (IMP-SSD) with several state-of-the-art lightweight SR methods: SRCNN [7], FSRCNN [8], CARN [19], and ESRN [31]. Note that IMP-SSD obtains the gradually shrinking lightweight SR models. The comparison results for \\(2 \\times\\), \\(3 \\times\\), and \\(4 \\times\\) magnification factors are shown in Table 6. PSNR and SSIM are adopted as metrics to evaluate image quality by convention." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.892, + 0.867 + ], + "angle": 0, + "content": "It shows that our IMP-SSD achieves the best results in terms of the coordination of parameters and performance." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "14404" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.151, + 0.124, + 0.816, + 0.137 + ], + "angle": 0, + "content": "Table 6. Comparison of the quantitative results of average PSNR/SSIM with state-of-the-art efficient SR models." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.139, + 0.89, + 0.414 + ], + "angle": 0, + "content": "
ScaleMethodParamsSet5 PSNR/SSIMSet14 PSNR/SSIMB100 PSNR/SSIMUrban100 PSNR/SSIM
2×SRCNN [7] (ECCV' 2014)57K36.66 / 0.954232.42 / 0.906331.36 / 0.887929.50 / 0.8946
FSRCNN [8] (ECCV' 2016)12K37.00 / 0.955832.63 / 0.908831.53 / 0.892029.88 / 0.9020
CARN [19] (ECCV' 2018)415K37.68 / 0.959433.27 / 0.914931.97 / 0.896931.18 / 0.9186
ESRN [31] (AAAI' 2020)324K37.85 / 0.960033.42 / 0.916132.10 / 0.898731.79 / 0.9248
EDSR [23] (CVPRW' 2017)1370K37.98 / 0.960533.63 / 0.918032.17 / 0.899632.08 / 0.9280
EDSR_IMP-SSD.88% (ours)171K37.82 / 0.959933.44 / 0.916732.09 / 0.898731.69 / 0.9244
EDSR_IMP-SSD.94% (ours)85K37.51 / 0.958633.09 / 0.913731.88 / 0.896030.94 / 0.9163
3×SRCNN [7] (ECCV' 2014)57K32.75 / 0.909029.28 / 0.820928.41 / 0.786326.24 / 0.7989
FSRCNN [8] (ECCV' 2016)12K33.16 / 0.914029.43 / 0.824228.53 / 0.791026.43 / 0.8080
CARN [19] (ECCV' 2018)415K34.00 / 0.924030.07 / 0.836828.92 / 0.800727.53 / 0.8379
ESRN [31] (AAAI' 2020)324K34.23 / 0.926230.27 / 0.840029.03 / 0.803927.95 / 0.8481
EDSR [23] (CVPRW' 2017)1558K34.37 / 0.926730.31 / 0.840929.07 / 0.804328.05 / 0.8506
EDSR_IMP-SSD.88% (ours)195K34.15 / 0.925230.20 / 0.839229.00 / 0.802827.81 / 0.8449
EDSR_IMP-SSD.94% (ours)97K33.74 / 0.921629.92 / 0.834028.81 / 0.797627.16 / 0.8298
4×SRCNN [7] (ECCV' 2014)57K30.48 / 0.862827.49 / 0.750326.90 / 0.710124.52 / 0.7221
FSRCNN [8] (ECCV' 2016)12K30.71 / 0.865727.59 / 0.753526.98 / 0.715024.62 / 0.7280
CARN [19] (ECCV' 2018)415K31.78 / 0.889428.39 / 0.776227.42 / 0.730325.56 / 0.7674
ESRN [31] (AAAI' 2020)324K31.99 / 0.891928.49 / 0.777927.50 / 0.733125.87 / 0.7782
EDSR [23] (CVPRW' 2017)1518K32.09 / 0.893728.55 / 0.780627.54 / 0.734825.97 / 0.7819
EDSR_IMP-SSD.88% (ours)190K31.89 / 0.891128.44 / 0.777727.46 / 0.732625.71 / 0.7743
EDSR_IMP-SSD.94% (ours)95K31.74 / 0.887928.33 / 0.774927.39 / 0.729725.48 / 0.7657
" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.435, + 0.472, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.606, + 0.47, + 0.632 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons of IMP-SSD with other lightweight SR models on Urban100 for \\(4 \\times\\) SR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.866 + ], + "angle": 0, + "content": "In particular, compared with other lightweight SR models such as CARN and ESRN, our IMP-SSD in EDSR with the sparsity of \\(88\\%\\) achieves comparable performance while our model size is about half or two-thirds of those. Compared with smaller SR models such as SRCNN and FSRCNN, IMP-SSD with the sparsity of \\(94\\%\\), IMP-SSD achieves significant gains in PSNR by at least 0.4dB, 0.3dB, 0.4dB for \\(2\\times\\), \\(3\\times\\), and \\(4\\times\\) SR, respectively. The same conclusion is made in SSIM. Furthermore, compared with the original EDSR, IMP-SSD makes the SR model shrink to reach a very small size but their SR performance does not degrade very much. Compared with the original EDSR, the maximum decreases are 0.4dB, 0.2dB, 0.2dB for IMP-SSD with the sparsity of \\(88\\%\\) and 0.6dB, 0.9dB, 0.5dB for IMP-SSD with the sparsity of \\(94\\%\\), respectively. It demonstrates that our method is effective on scalable SR." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.436, + 0.892, + 0.493 + ], + "angle": 0, + "content": "We also give the comparison of our method in visual effect in Fig. 6. It is observed that our method achieves a better visual effect than other lightweight SR models even with larger model sparsity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.509, + 0.627, + 0.523 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.795 + ], + "angle": 0, + "content": "In this paper, we propose a novel memory-friendly scalable lightweight framework (MSSR) with rewinding LTH, which can be deployed on arbitrary computational resources according to customization. MSSR is backtracking and includes two stages: the forward stage for model compression and the backward for model expansion. For the first time, LTH with rewinding weights is implemented to explore the winning tickets of deep SR models. We use IMP to gradually shrink the SR model and not only obtain the scalable lightweight model but also obtain the pruning-out masks which form the nested sets. In order to improve the performance of the shrinking model, SSD is conducted. Unlike the traditional SD, we replace the feature maps on several layers of Teacher with the counterparts of Students and conduct distillation. In the backward stage, we gradually recover the lightweight SR models depending on the pruning-out masks and then fine-tune the current model. We make instantiation on several state-of-the-art SR models. The extensive experimental results are effective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by grants from the National Key Research and Development Program of China (No.2020AAA0108301), National Natural Science Foundation of China (No.62176224, No.62222602), and CCF-Lenovo Blue Ocean Research Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14405" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.469, + 0.173 + ], + "angle": 0, + "content": "[1] Efficient non-local contrastive attention for image superresolution. In AAAI, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.175, + 0.469, + 0.211 + ], + "angle": 0, + "content": "[2] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.469, + 0.251 + ], + "angle": 0, + "content": "[3] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.254, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.469, + 0.355 + ], + "angle": 0, + "content": "[5] Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Zhangyang Wang, and Michael Carbin. The lottery ticket hypothesis for pre-trained bert networks. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.357, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[6] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.436, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[8] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.469, + 0.512 + ], + "angle": 0, + "content": "[9] Utku Evci, Fabian Pedregosa, Aidan Gomez, and Erich Elsen. The difficulty of training sparse neural networks. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.515, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[10] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[11] Jonathan Frankle, David J Schwab, and Ari S Morcos. The early phase of neural network training. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.618 + ], + "angle": 0, + "content": "[12] Sharath Girish, Shishira R Maiya, Kamal Gupta, Hao Chen, Larry S Davis, and Abhinav Shrivastava. The lottery ticket hypothesis for object recognition. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.709 + ], + "angle": 0, + "content": "[14] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. TPAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.711, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[15] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.75, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[16] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.826 + ], + "angle": 0, + "content": "[17] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM MM, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.828, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[18] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.469, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.165 + ], + "angle": 0, + "content": "[19] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[20] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[21] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. In NeurIPS, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.23, + 0.892, + 0.266 + ], + "angle": 0, + "content": "[22] Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Dynamic slimmable network. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[23] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[24] Ning Liu, Geng Yuan, Zhengping Che, Xuan Shen, Xiaolong Ma, Qing Jin, Jian Ren, Jian Tang, Sijia Liu, and Yanzhi Wang. Lottery ticket preserves weight correlation: Is it desirable or not? In ICML, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.406 + ], + "angle": 0, + "content": "[25] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.408, + 0.892, + 0.458 + ], + "angle": 0, + "content": "[26] Xiaotong Luo, Mingliang Dai, Yulun Zhang, Yuan Xie, Ding Liu, Yanyun Qu, Yun Fu, and Junping Zhang. Adjustable memory-efficient image super-resolution via individual kernel sparsity. In ACM MM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[27] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[28] Sai Prasanna, Anna Rogers, and Anna Rumshisky. When bert plays the lottery, all tickets are winning. In ACL, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[29] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[30] Pedro Savarese, Hugo Silva, and Michael Maire. Winning the lottery with continuous sparsification. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[31] Dehua Song, Chang Xu, Xu Jia, Yiyi Chen, Chunjing Xu, and Yunhe Wang. Efficient residual dense block search for image super-resolution. In AAAI, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[32] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.738 + ], + "angle": 0, + "content": "[33] Shihui Yin, Kyu-Hyoun Kim, Jinwook Oh, Naigang Wang, Mauricio Serrano, Jae-Sun Seo, and Jungwook Choi. The sooner the better: Investigating structure of early winning lottery tickets. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.74, + 0.892, + 0.776 + ], + "angle": 0, + "content": "[34] Haonan Yu, Sergey Edunov, Yuandong Tian, and Ari S Morcos. Playing the lottery with rewards and multiple languages: lottery tickets in rl and nlp. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.778, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[35] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In ICCV, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.865 + ], + "angle": 0, + "content": "[36] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In ICCV, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14406" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.127, + 0.47, + 0.165 + ], + "angle": 0, + "content": "[37] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[38] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.468, + 0.23 + ], + "angle": 0, + "content": "[39] Hao Zhou, Jose M Alvarez, and Fatih Porikli. Less is more: Towards compact cnns. In ECCV, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.127, + 0.47, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.468, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14407" + } + ] +] \ No newline at end of file diff --git a/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_origin.pdf b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..49b67a692cd63d09a54a080787cc2e007a1e0210 --- /dev/null +++ b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/d7f650f5-0e90-4e4e-9e3b-372fb6561fe3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9290d3f6aea40f7180d93f1241737403dfaf47f8435866734ef1fca49dce9673 +size 640402 diff --git a/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/full.md b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..021e7b21a5a34fd8628786bdee2682b38d14bbd3 --- /dev/null +++ b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/full.md @@ -0,0 +1,291 @@ +# Memory-friendly Scalable Super-resolution via Rewinding Lottery Ticket Hypothesis + +Jin Lin $^{1*}$ , Xiaotong Luo $^{1*}$ , Ming Hong $^{1}$ , Yanyun Qu $^{1\dagger}$ , Yuan Xie $^{2\dagger}$ , Zongze Wu $^{3}$ + +$^{1}$ School of Informatics, Xiamen University, Fujian, China + +$^{2}$ School of Computer Science and Technology, East China Normal University, Shanghai, China + $^{3}$ School of Mechatronics and Control Engineering, Shenzhen University, Shenzhen, China + +{linjin, xiaotluo, mingh}@stu.xmu.edu.cn, yyqu@xmu.edu.cn, + +yxie@cs.ecnu.edu.cn, zzwu@szu.edu.cn + +# Abstract + +Scalable deep Super-Resolution (SR) models are increasingly in demand, whose memory can be customized and tuned to the computational recourse of the platform. The existing dynamic scalable SR methods are not memory-friendly enough because multi-scale models have to be saved with a fixed size for each model. Inspired by the success of Lottery Tickets Hypothesis (LTH) on image classification, we explore the existence of unstructured scalable SR deep models, that is, we find gradual shrinkage subnetworks of extreme sparsity named winning tickets. In this paper, we propose a Memory-friendly Scalable SR framework (MSSR). The advantage is that only a single scalable model covers multiple SR models with different sizes, instead of reloading SR models of different sizes. Concretely, MSSR consists of the forward and backward stages, the former for model compression and the latter for model expansion. In the forward stage, we take advantage of LTH with rewinding weights to progressively shrink the SR model and the pruning-out masks that form nested sets. Moreover, stochastic self-distillation (SSD) is conducted to boost the performance of sub-networks. By stochastically selecting multiple depths, the current model inputs the selected features into the corresponding parts in the larger model and improves the performance of the current model based on the feedback results of the larger model. In the backward stage, the smaller SR model could be expanded by recovering and fine-tuning the pruned parameters according to the pruning-out masks obtained in the forward. Extensive experiments show the effectiveness of MMSR. The smallest-scale sub-network could achieve the sparsity of $94\%$ and outperforms the compared lightweight SR methods. + +![](images/a5cc0ce30a2990a07f9b58171cda1cd176768b8e9c3c0d6c539f47ed29203d2e.jpg) +Figure 1. The flowchart of scalable SR network. S_n and S_w denote the neurons and the neural connections (weights) of the simplest subnetwork; M_n and M_w denote the intermediate neurons and neural connections; Cur_n and Cur_w denote the specific neurons and neural connections belonging to the current subnetwork. Free_n denotes the pruning-out neurons. The final model (with $100\%$ recovered parameters) reaches the original size. The scalable model is adjustable to the memory resource allocation. + +# 1. Introduction + +Single image super-resolution (SISR) aims to reconstruct a high-resolution (HR) image from the corresponding low-resolution (LR) one. With the rising of deep learning, deep SR methods have made incredible progress. However, the existing SR models mostly require computational and memory resources, so they do not favor resource-limited devices such as mobile phones, robotics, and some edge devices. + +The lightweight SR methods are attracting more attention for better application to resource-limited devices. The existing lightweight SR methods mainly focus on designing compact architectures [17, 20] with a fixed size, such as multi-scale pyramid [20], multiple-level receptive fields [17, 18], and recursive learning [19]. However, most lightweight SR models with fixed sizes are not flexible in applications. If one model does not match the resources of + +the platform, it has to be retrained by compression methods to match the resources and then reloaded onto the devices. + +The urgent demand to customize models based on deployment resources is increasing. Dynamic neural networks for SR [14,22] are proposed to adjust the network architecture according to different computational resources. The existing dynamic deep SR models often explore dynamic depth or width [22, 26], but they either require large memory resources or are not convenient for users to wait for retraining another SR model. The former leads to saving the multi-scale SR models of different sizes and the latter leads to retraining the model before being reused again. The limitation lies in that they are not memory-friendly. In many edge-device SR applications, the devices may be scalable, that is, their memories may be small in the beginning and be expanded later. Thus, we discuss two issues in this paper: 1) how to make a scalable lightweight model for the multi-scale computational recourse. 2) how to make the lightweight model expand to a larger-size model for better performance if the computational recourse is increased. + +As for the first issue, inspired by the success of Lottery Ticket Hypothesis [10] which points out that there could exist a properly pruned sub-network named winning tickets to achieve comparable performance against the original dense network in model compression of classification, it is used to find the sub-network for SR. We are the first to study the existence of scalable winning tickets for SR. Iterative pruning and rewinding weights in LTH are beneficial to the scalable lightweight SR model. Iterative pruning may compress the SR model according to an arbitrary size. It is observed in [24] that the winning tickets are related to an insufficient DNN, and rewinding LTH outperforms the original LTH. That is, the initial weights in LTH are replaced with the T-iteration weights during pruning and fine-tuning. The scalable deep SR model is shown in Fig. 1. + +As for the second issue, the scalable SR model can customize parameters to adapt to different memory resources rather than load or offload different models for different devices. In other words, during real applications, there will be only one simple model to be employed for inference whose size is decided by the computational resource. + +In this paper, we propose a memory-friendly scalable deep SR model (MSSR) via rewinding LTH. We use the rewinding LTH [10] to generate our unstructured scalable mask. MSSR is backtracking and contains forward and backward stages. The former focuses on model compression by rewinding LTH with iterative pruning and finetuning, and the latter focuses on iterative model expansion until it goes back to its original size. Multi-scale winning tickets together with the pruning-out masks are obtained by rewinding LTH in the forward stage with the decrease in the number of parameters. The pruning-out masks are nested. In order to make the compressed SR model not degrade sig + +nificantly, stochastic self-distillation (SSD) is used to improve the representation of the small-scale SR model, and knowledge is transferred from the last-scale model to the current scale model. In the backward stage, the smallest model is expanded gradually to the model with the original size with the expanded mask. + +The main contributions of this work are three-fold: + +- A memory-friendly scalable dynamic SR lightweight model via rewinding LTH is proposed. MSSR is re-configurable and switchable to sub-networks with different sizes according to on-device resource constraints on the fly. +- MSSR is backtracking, which contains forward and backward stages. Multi-scale winning tickets form nested masks for the multi-scale models. SSD is conducted by replacing the features in randomly selected layers between Teacher and Student to improve the performance of the scalable SR lightweight models. +- Extensive experiments demonstrate that MSSR can generalize to different SR models as well as state-of-the-art attention-based models, ENLCN [1]. + +# 2. Related work + +# 2.1. Lightweight Deep Learning-based SR + +SISR aims to generate a high-resolution image from the low-resolution version. Recently, efficient and lightweight SISR networks have attracted increasing interest in the computer vision community. The lightweight SR methods consist of static and dynamic neural networks for SR. The earlier lightweight SR methods mainly focus on the static network structure design. IDN [18] utilizes group convolution and combined the local long and short path features effectively. CARN [19] utilizes a recursive cascading mechanism for learning multi-level feature representations. IMDN [17] introduces information multi-distillation blocks for extracting hierarchical features. FALSR [6] applies neural architecture search for efficient network design automatically. LatticeNet [27] designs a lattice block to combine two residual blocks adaptively. However, the aforementioned lightweight SR models need elaborate designs. + +Recently, dynamic lightweight SR models are demanded increasingly whose size is adjustable according to different computational resources. DS-Net [22] makes the width of the SR model adjustable, and IKS [26] makes the sparsity of individual kernels adjustable. However, they require saving lightweight SR models with different shrinkage sizes, which is not convenient for users to offload the old model and then load a new one. Our method uses a scalable SR model adjustable to the changing computational recourse to avoid the loading and offloading models. + +![](images/4cbf2900c4189795dcdd7620250b00d900293e7ce223b8daa04e70e62a4b6e42.jpg) +Figure 2. The architecture of memory-friendly scalable deep SR model (MSSR). It contains two stages: the forward stage for model compression and the backward stage for model expansion. In the forward stage, the original SR network is firstly processed by IMP to obtain winning tickets and the pruning-out mask $LM_{1}$ , and then the winning tickets are processed by SSD and the lightweight subnetwork is obtained. So do the other gradual shrinkage sub-networks and $\{LM_{i}\}$ forms the nested sets. In the backward stage, according to the pruning-out masks, the smaller lightweight SR model is gradually expanded and finely tuned to obtain a larger lightweight SR model. + +# 2.2. Pruning and Lottery Ticket Hypothesis + +Model pruning is to prune away the unimportant connections (parameters) [21] in the neural network which trades off between the model size and performance. The existing pruning methods can be classified into unstructured [6, 13, 21] and structured methods [15, 25, 39]. The structured pruning methods remove the substructures of the model, such as channels, filters, and layers, while the unstructured pruning methods are magnitude-based, where the weights are removed below the threshold. However, magnitude-based pruning methods are not hardware-friendly and computationally efficient. Lottery Ticket Hypothesis (LTH) [10] opened a pandas's box of immense possibilities in the field of pruning and sparse models. It assumes that there exist sparse sub-networks independently training from scratch that can match the performance of dense networks. Some literatures discuss the effectiveness of pruning, fine-tuning [24], and rewinding weights [29] for LTH. + +LTH has been widely applied in image classification [9-12, 30, 32, 33], natural language processing [5, 28, 29, 34], and so on. However, few works discuss the existence of LTH in the deep model for image restoration. We are the first to investigate the existence of scalable winning tickets for lightweight deep SR models. + +# 3. Method + +# 3.1. Overview of MSSR + +In this section, we introduce the memory-friendly scalable SR model which is an all-in-one model with multiscale SR models sharing a model. As shown in Fig. 2, our MSSR consists of two stages: the forward stage for model compression, and the backward stage for model expansion. + +In the forward stage, we use rewinding LTH (RLTH) to iteratively shrink a dense SR model to obtain a sub-network with the small size not significantly scarifying the performance. During pruning by LTH each time, we treat the current lightweight SR model as Student and its last ancestor lightweight SR model as Teacher, and conduct stochastic self-distillation between Teacher and Student. In Fig. 2, the dark-blue block denotes the winning tickets, and the white block denotes the pruning part. We also record the mask of pruning part and use a matrix with "1" to represent the pruning-out weights and with "0" for others. It is observed that the masks are nested and the initial mask is continuously contained in the later mask. In the backward stage, we iteratively expand the lightweight sub-network according to the saved nested masks. The most lightweight network marked by 1 is expanded iteratively to the gradual less + +lightweight network, the expanding order is shown as "1", "1-2", "1-2-3", and "1-2-3-4". + +# 3.2. The Forward Stage: Model Compression + +The forward streamline is shown in Fig. 3. Let a network denoted by $f(\theta_0; x)$ and the pruned masks which form the "nested sets" denoted by $LM = \{\}$ , where $\theta_0$ denotes the initial weights. The flowchart shows a whole process in an iteration for shrinking a model and it contains three components: SR model pre-train, IMP with rewinding weights, and SSD. Firstly, an SR network is trained to reach a steady state, and its weights are named the first nonpruning weights $\theta$ . After that, Iterative Magnitude Pruning is conducted on the SR model $\theta$ , and the weights below the pre-set threshold are pruned out which forms a pruning mask $m \in \{0,1\}^n$ . Following [24], we rewind the remaining weights to their values in $\theta$ , i.e., $\theta' = \theta \odot m$ . Then, we retrain the pruned weights $\theta'$ by SSD arriving at $\hat{\theta}$ . So do the next more lightweight SR model with the model weights $\hat{\theta}$ as the initiation. + +LTH with Rewinding for SR. Benefiting from LTHs, we implement iterative magnitude pruning with rewinding to gradually shrink the SR model. The motivation lies in the controllability and flexibility of LTH. We can shrink the model arbitrarily by LTH, which favors the scalable lightweight SR model. + +We mainly consider two issues: the model pruning process and the pruning rules. The pruning process corresponds to the process of gradually shrinking the model, and the pruning methods need to meet the following rules: 1) the amount of prune-out parameters is arbitrarily adjustable; 2) the model is shrunk gradually, which leads to the nested sets and the larger model contains its smaller descendant. LTH fits well with the needs of our scalable mask generator and the appropriate use of its task settings to implement the model compression process, in particular discarding its over parameterization, a feature that is not suitable for intensive prediction tasks. + +IMP is proven to be effective for LTH. In [24], it is proved that weight rewinding is better than the original weight initiation. That is, unpruned weights are rewound to their values from $T$ -iteration training and retrained from there using the original training schedule. + +Compared with one-shot magnitude pruning (OMP) which reaches the goal model at one time, IMP does not shrink a model to reach the pre-set size at one time but shrinks a model with several iterations and the model size is evenly decreased at each iteration. Let the network be pruned for $T$ times and be expected to discard $p\%$ of the parameters. The weights/parameters of the network are represented by $\omega \in \mathbb{R}^n$ and the pruning mask by $m \in \{0,1\}^n$ where $n$ represents all the capacity of the network. + +Weight rewinding does not initiate the model with the + +![](images/c4dc5c98f9c2e90ab20862a08e1e97e9d7a22e8c2c9a4de97d2fe75e732194cd.jpg) +Figure 3. The flowchart of the whole process of IMP and SSD. The initial model $\theta_0$ is pre-trained to obtain a stable model $\theta$ . IMP is implemented on $\theta$ and the winning ticket (sub-network) is obtained. The sub-network is initiated by $\theta \odot m$ and is processed by SSD for the better model $\theta$ . + +original initiation weights $\theta_0\odot m$ , but with the $t$ -iteration $\theta_t\odot m$ where $0 < t\leq T$ , resulting in weights $\theta_{t}$ and the network function $f(x;\theta_t\odot m)$ . We highlight that the original LTH always chooses $\theta_0$ as dense model, while in this work we choose $\theta_{t}$ from the last pruning iteration. After "Pruning" and "Rewinding and Retraining", $\theta = \hat{\theta}$ is updated and the mask at the $t$ -th iteration is obtained which is denoted as $LM_{t} = LM_{t - 1}\cup \{m\}$ . In other words, the mask $LM_{t}$ at $t$ -th iteration is contained in the mask $LM_{t - 1}$ at $(t - 1)$ -th. "Pruning" and "Rewinding and Retraining" are repeated until the model reaches the end condition, which is the iteration steps $T$ or target sparsity $p$ + +Stochastic Self-distillation for SR. In the forward stage, due to the progressive shrinking of the SR model, it is definitely difficult to maintain the performance as same as the original model. Especially, when only $1\%$ of the parameters are kept, most magnitude-pruning methods do not work. + +In order to mitigate the degradation of lightweight SR models, we adopt self-distillation [36] to boost the performance of the pruned model. Unlike the existing self-distillation which uses the function of features obtained by the model at previous iterations such as the average and summation function, we introduce attribution analysis to self-distillation. We randomly select several layers in Student, and the feature maps of selected layers replace the counterpart of Teacher, which is named Stochastic self-distillation (SSD). Concretely, the current lightweight SR model is regarded as the student network (Student), and its last ancestor network is regarded as the teacher network (Teacher), and Student is obtained by pruning Teacher. SSD is conducted between Teacher and Student. Student is the pruned model of Teacher. Let $X$ and $Y$ denote the network input and ground truth, respectively. + +The architecture of SSD is shown in Fig. 4, which contains two streamlines: Teacher and Student. The knowledge in feature maps is transferred from Teacher to Student. + +![](images/d3fa501b40f78f274e3ace7e22932cd8d736c9cb22d1f3ef6857b4b25ed35364.jpg) +Figure 4. The flowchart of SSD. Student is the sub-network by pruning Teacher. Several layers of Student are selected randomly. The feature maps in a selected layer replace the counterpart of Teacher which outputs the SR result, and so do other selected layers. The total loss is the summation of the knowledge distillation loss $L_{skd}$ and the fidelity loss $L_{1}$ . + +If Student contains $d$ layers, at each-iteration distillation, $s$ layers randomly are selected. In realization, a one-hot vector is randomly generated, named Switch Vector, with a length of $d$ in each training, $\alpha \in \{0,1\}^d$ . If the element is equal to 1 in $\alpha$ , the feature maps in the layer corresponding to this position are selected, and are named Lucky feature maps. Then, Lucky feature maps in the selected layers to replace the counterparts of Teacher. + +In details, Student firstly is trained for a stable SR model. After that, SSD is conducted. At each iteration, Switch vector is dynamically generated, which is formulated as: + +$$ +D C = \left[ \alpha_ {1}, \alpha_ {2}, \dots , \alpha_ {d} \right], \tag {1} +$$ + +where $\alpha_{i}$ denotes the $i$ -th element corresponding to the $i$ -th layer and $d$ is the layer number in Student. Then, lucky feature maps of Student are selected according to $\alpha_{i}$ , and they are formulated as: + +$$ +\mathrm {L F} = \left[ f _ {\alpha_ {1}} ^ {s}, f _ {\alpha_ {2}} ^ {s}, \dots , f _ {\alpha_ {d}} ^ {s} \right], \tag {2} +$$ + +where $f_{\alpha}^{s}$ denotes the feature obtained from Student. Note that all elements in Switch Vector are independent and identically distributed. + +Lucky feature maps replace the counterpart of Teacher each time, and the corresponding output of new Teacher is denoted by $Y_{\alpha_i}^t$ , and the corresponding weights of new Teacher are denoted by $\theta_t$ , which is formulated as: + +$$ +\left\{ \begin{array}{l} Y _ {\alpha_ {i}} ^ {t} = F \left(\theta^ {t}; x, f _ {\alpha_ {i}} ^ {s}\right), i = 1, \dots , d, \\ \mathrm {L Y} = \left[ Y _ {\alpha_ {1}} ^ {t}, Y _ {\alpha_ {2}} ^ {t}, \dots , Y _ {\alpha_ {d}} ^ {t} \right], \end{array} \right. \tag {3} +$$ + +where $LY$ denotes the set of SR outputs, $Y_{\alpha}^{t}$ is the SSD output of lucky maps, $f_{\alpha}^{s}$ . All the outputs between Student and Teacher are supervised by HR image $Y$ . The knowledge of Teacher takes more rationality to be distilled into Student. + +Therefore, the total loss function utilized to train the SR model consists of the supervised loss $L_{1}$ and self-distillation + +Algorithm 1 Memory-friendlyly scalable lightweight framework for image SR. + +Data: HR and LR image datasets: $D_{hr}$ , $D_{lr}$ + +Input: Dense network with randomly initial weights: $f(\theta_0)$ , pruning rounds: $T$ , and pruning target rate: $p$ + +Output: Scalable lightweight SR model + +1: /* The forward stage: model compression */ +2: Initializing: Pruned mask $m = \{1\}^n$ , lottery mask pool $LM = \{m\}$ and epochs of each training $N$ +3: Pre-training: Train network for $N$ epochs $f(\theta_0;x)\longrightarrow f(\theta ;x)$ +4: while $i < T - 1$ do +5: Prune bottom $p\% \times \frac{1}{T}$ of $m \odot \theta$ and update $m$ ; +6: Re-train the pruned sub-network $\theta^{\prime} = m\odot \theta$ for $N$ epochs $f(\theta^{\prime};x)\longrightarrow f(\hat{\theta};x)$ with $SSD(\theta ,\theta^{\prime};x)$ +7: Update mask pool $LM = LM \cup \{m\}$ and weights $\theta = \hat{\theta}$ ; +8: $i\gets i + 1$ +9: end while +10: Generate the sparse sub-network matching the target sparsity, and the mask pool $LM$ ; +11: /* The backward stage: model expansion */ +12: Initializing: Current step $t = T$ and sparsest sub-network $\theta_T = \theta$ +13: while $t > 1$ do +14: Calculate the grad mask $m = LM_{t - 1} - LM_t$ to determine the trainable weights; +15: Fine-tune/re-train the trainable weights for $N$ epochs $f(\theta_t;x)\longrightarrow f(\theta_t';x)$ +16: Update $\theta_t = \theta_t'$ ; +17: $t\gets t - 1;$ +18: end while + +loss $L_{skd}$ , which is formulated as: + +$$ +L _ {t o t a l} = L _ {1} + \lambda L _ {s k d} = \left\| Y ^ {s} - Y \right\| _ {1} + \frac {\lambda}{d} \sum_ {i = 1} ^ {d} \alpha_ {i} \left\| Y _ {\alpha_ {i}} ^ {t} - Y \right\| _ {1}, \tag {4} +$$ + +where $\lambda$ denotes the regularization parameter, which is empirically set as 0.1 in our experiments. During inference, the SR model $\theta_{p}$ can be customized for various resources according to different sparsity $p$ . After the network training, a lottery mask pool can be built for the backward stage. + +# 3.3. The Backward Stage: Model Expansion + +Our intention is to obtain a scalable network that can deploy according to customized memory resources. In this subsection, we introduce how to gradually recover a lightweight SR model, which is a backtracking process of IMP. We start from the most lightweight SR model and end with the original SR model. Corresponding to gradually shrinking the SR model with LTH, we solve two issues in + +![](images/76a605c86e31ff150ac11bb41a27637194670e3b2fd7677e655658c7b2156c43.jpg) +Network Sparsity (\% pruned) +Figure 5. Comparisons of IMP and OMP on various famous SR methods, which are evaluated on Set14 for $4 \times$ SR. The performance of the baseline dense network, the sparse sub-network obtained by IMP, as well as the sub-network obtained by OMP. The sub-network is with only $20\%$ of the weights. + +the model expansion: iterative model expansion (IME) and the expansion rule, the former refers to the requirement of the IME process and the latter refers to the requirement of the learning rule. + +Let's focus on an expansion from the $t$ -th lightweight model to the $(t - 1)$ -th lightweight model where $t$ comes from the definition in the forward stage. The larger $t$ is, the smaller model is. During IMP, we obtain two pruned-out masks corresponding to the $t$ -th and $(t - 1)$ -th lightweight models, denoted by $\mathrm{LM}_t$ and $\mathrm{LM}_{t - 1}$ , respectively. Their difference represents the expanded positions, i.e., $m = LM_{t - 1} - LM_t$ . Without changing the $t$ -th model, we recover the weights according to the mask $m$ . After that, we fine-tune the recovered model and obtain the $(t - 1)$ -th model. So do the next model expansion. The detailed procedure is described in Algorithm 1. + +# 4. Experiments + +# 4.1. Datasets + +DIV2K [2] is adopted to train the SR model, which is a high-quality natural scene dataset widely used for image SR tasks. The whole dataset includes 800 training images and 100 validation images totally with abundant texture and contents. The bicubic degraded LR images are synthesized following the previous works [17, 37]. Besides, the trained models are evaluated on four public SR benchmark datasets: Set5 [4], Set14 [35], B100 [3], and Urban100 [16]. + +Table 1. Ablation studies about the iterative magnitude pruning (IMP) and the stochastic self-distillation learning (SSD). All these models are validated on Set14 for $4 \times$ image SR. + +
EDSR基礎IMP (LTH)IMP RewindingSSDSparsityPSNR
✓×××0%28.55
✓✓××94%27.66
✓×✓×94%28.12
✓×✓✓94%28.33
+ +Table 2. The performance of the sparse sub-network pruned by IMP-SSD on Set14 for $2 \times, 3 \times, 4 \times$ image SR. + +
ScaleModelSparsityPSNR (dB)Params
2×EDSR基礎0%33.631370K
EDSR_IMP94%32.8885K
EDSR_IMP-SSD94%33.09 (+0.21)85K
3×EDSR基礎0%30.301558K
EDSR_IMP94%29.7397K
EDSR_IMP-SSD94%29.92 (+0.19)97K
4×EDSR基礎0%28.551518K
EDSR_IMP94%28.1295K
EDSR_IMP-SSD94%28.33 (+0.21)95K
+ +# 4.2. Implementation and Training Details + +For training the proposed SR model, $48 \times 48$ image patches with RGB channels are cropped as the model input. The data augmentation is implemented by random flipping and rotation. The Adam optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ is utilized to train the models. Besides, the mini-batch size is set to 16. The learning rate is initialized as $2e - 4$ . The pruned weights can be determined by IMPSSD each 300 epochs and the same setting is adopted in the model expansion stage. We use PyTorch to implement our models with a GTX TITAN GPU. + +Objective criteria, i.e., peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are adopted to evaluate the model performance. The two metrics are both calculated on the Y channel of the YCbCr space. + +# 4.3. Ablation Study + +Next, we mainly give a detailed analysis of the proposed lottery-ticket stochastic self-distillation scalable framework (IMP-SSD) for image SR. We adopt EDSR [23] as the backbone network to implement our experiments. + +IMP vs. OMP. To verify the superiority of IMP over OMP on the compression task of image SR, we conduct IMP and OMP on the forward stage of our method instantiated with three popular SR methods: EDSR [23], RDN [38], and RCAN [37]. As shown in Fig. 5, when the pruned network reaches a very small model size, usually as the sparsity of the model is higher than $80\%$ , where the sparsity means the ratio of the pruning-out parameters, IMP does not significantly scarify the SR performance while OMP degrades a lot. Especially, it obtains 27.66dB in PSNR with $94\%$ sparsity as shown in Table 1. OMP is obviously unable to make the RCAN model reach a very small model similar to that by IMP. That's why we pick IMP as our pruning algorithm, which demonstrates the advantages of LTH once again. + +Effectiveness of IMP-SSD. To demonstrate the effectiveness of IMP-SSD, we employ EDSR as the backbone and compare MSSR with EDSR. As shown in Table 2, our IMP-SSD method can not only compress a model to a very + +Table 3. Comparison of the quantitative results of average PSNR/SSIM with state-of-the-art efficient SR models. + +
ScaleMethodParamsSet5 PSNR/SSIMSet14 PSNR/SSIMB100 PSNR/SSIMUrban100 PSNR/SSIMDIV2K_valid PSNR/SSIM
4×EDSR (CVPRW' 2017)1518K32.09 / 0.893728.55 / 0.780627.54 / 0.734825.97 / 0.781930.38 / 0.8364
EDSR_IMP-SSD_88% (ours)190K31.89 / 0.891128.44 / 0.777727.46 / 0.732625.71 / 0.774330.25 / 0.8336
EDSR_IMP-SSD_94% (ours)95K31.74 / 0.887928.33 / 0.774927.39 / 0.729725.48 / 0.765730.12 / 0.8303
RCAN (ECCV' 2018)15592K32.19 / 0.895028.63 / 0.781927.57 / 0.735526.21 / 0.788830.40 / 0.8331
RCAN_IMP-SSD_88% (ours)1949K32.30 / 0.896328.72 / 0.784427.63 / 0.737926.32 / 0.793230.53 / 0.8396
RCAN_IMP-SSD_94% (ours)974K32.02 / 0.892428.49 / 0.778827.51 / 0.733425.87 / 0.780030.30 / 0.8347
RDN (CVPR' 2018)5579K32.24 / 0.896528.71 / 0.784127.64 / 0.738426.31 / 0.792830.55 / 0.8406
RDN_IMP-SSD_88% (ours)697K31.87 / 0.890228.38 / 0.776627.42 / 0.731325.53 / 0.768130.15 / 0.8315
RDN_IMP-SSD_97% (ours)174K31.36 / 0.882628.10 / 0.769527.23 / 0.724925.09 / 0.311629.86 / 0.8243
ENLCN (AAAI' 2022)1536K32.14 / 0.893928.59 / 0.781227.56 / 0.735226.05 / 0.784530.36 / 0.8360
ENLCN_IMP-SSD_88% (ours)192K31.97 / 0.891128.47 / 0.773127.47 / 0.732525.77 / 0.731330.22 / 0.8326
ENLCN_IMP-SSD_97% (ours)48K31.41 / 0.882528.09 / 0.769027.24 / 0.725025.15 / 0.313329.84 / 0.8239
+ +Table 4. Comparison between the performance (PSNR) of the sparse sub-networks via IMP-SSD and the performance of the sparse sub-networks of the scalable network (SCL) for $4 \times$ SR. + +
Prune (%)Set5Set14B100Urban100
IMP-SSDSCLIMP-SSDSCLIMP-SSDSCLIMP-SSDSCL
0 %32.0931.9428.5528.5327.5427.5325.9725.89
50 %32.0531.9628.5728.5127.5527.5025.9925.79
75 %32.0331.9228.5328.4627.5227.4825.9025.73
88 %31.8931.8228.4428.4027.4627.4325.7125.61
94 %31.7431.7428.3328.3327.3927.3925.4825.48
+ +small size but also mitigate the degradation in performance for different magnification factors $(2\times, 3\times,$ and $4\times)$ . Besides, the results in Table 1 also illustrate that IMP-SSD further improves the PSNR to $28.33\mathrm{dB}$ with the similar sparsity. Therefore, our method has significant effects on scalable model compression. + +IMP-SSD SR vs. Scalable SR. In order to show the effectiveness of MSSR, we compare the scalable lightweight SR model obtained in the backward stage with the middle result by IMP-SSD in the forward stage. We use EDSR as the backbone, and make the comparison in five degrees of sparsities: $0\%$ , $50\%$ , $75\%$ , $88\%$ , $94\%$ . In Table 4, IMP-SSD generates five pruned SR models corresponding to five sparsities, while scalable SR only has one model. With the increase of sparsity, the SR performance decreases, but the gap between two neighbor sparsities is not more than $0.2\mathrm{dB}$ . Thus, the SR scalable model makes sense. As shown in Table 4, it is observed that scalable SR achieves comparable results to IMP-SSD on four datasets: Set5, Set14, B100, and Urban100. It shows that our method uses only one scalable model that can reach the level of the separately lightweight SR models at different sparse degrees. Therefore, MSSR is memory-friendly. + +# 4.4. Instantiating with Other SR methods + +To demonstrate the effectiveness of the proposed scalable lightweight framework (MSSR), we integrate it with four representative SR methods: EDSR [23], RCAN [37], RDN [38], and ENLCN [1]. All these SR models are re + +Table 5. The analysis about IMP-SSD on different methods for $4 \times$ image SR validated on Set14. + +
PSNR Sparsity +Model0%50%75%88%94%97%99%
EDSR_IMP-SSD28.5528.5728.5328.4428.3328.0520.36
RDN_IMP-SSD28.7028.6828.5228.3828.2328.1027.46
RCAN_IMP-SSD28.6328.7728.7328.7228.4928.1527.68
ENLCN_IMP-SSD28.5928.6228.5628.4728.3028.0927.60
+ +trained under the same experimental settings. We show the results in Table 3. It shows that our method is universally applicable to the existing SR methods. Compared with the original models with the magnification factor $4 \times$ , under the sparsities of $88 \%$ and $94 \% / 97 \%$ , the maximum decreases of IMP-SSD instantiated with EDSR, RCAN, RDN, and ENLCN are (0.2dB, -0.1dB, 0.8dB, 0.3dB), and (0.5dB, 0.4dB, 0.9dB, 0.9dB), respectively. Note that, in RCAN, IMP-SSD is better than its original model. We guess that SSD could improve the SR results. + +In Table 5, we give more comparisons of the SR performance with more different sparsity. It shows that all methods except EDSR, with the decrease of the model size, our method degrades gradually with not much range of decrease, their ranges are $1.3\mathrm{dB}$ , 1dB, 0.9dB in RCAN, RDN, ENLCN, respectively. EDSR degrades significantly by $8.2\mathrm{dB}$ with the sparsity of $99\%$ . + +# 4.5. Comparison with State-of-the-art + +To validate the effectiveness of MSSR, we compare our approach (IMP-SSD) with several state-of-the-art lightweight SR methods: SRCNN [7], FSRCNN [8], CARN [19], and ESRN [31]. Note that IMP-SSD obtains the gradually shrinking lightweight SR models. The comparison results for $2 \times$ , $3 \times$ , and $4 \times$ magnification factors are shown in Table 6. PSNR and SSIM are adopted as metrics to evaluate image quality by convention. + +It shows that our IMP-SSD achieves the best results in terms of the coordination of parameters and performance. + +Table 6. Comparison of the quantitative results of average PSNR/SSIM with state-of-the-art efficient SR models. + +
ScaleMethodParamsSet5 PSNR/SSIMSet14 PSNR/SSIMB100 PSNR/SSIMUrban100 PSNR/SSIM
2×SRCNN [7] (ECCV' 2014)57K36.66 / 0.954232.42 / 0.906331.36 / 0.887929.50 / 0.8946
FSRCNN [8] (ECCV' 2016)12K37.00 / 0.955832.63 / 0.908831.53 / 0.892029.88 / 0.9020
CARN [19] (ECCV' 2018)415K37.68 / 0.959433.27 / 0.914931.97 / 0.896931.18 / 0.9186
ESRN [31] (AAAI' 2020)324K37.85 / 0.960033.42 / 0.916132.10 / 0.898731.79 / 0.9248
EDSR [23] (CVPRW' 2017)1370K37.98 / 0.960533.63 / 0.918032.17 / 0.899632.08 / 0.9280
EDSR_IMP-SSD.88% (ours)171K37.82 / 0.959933.44 / 0.916732.09 / 0.898731.69 / 0.9244
EDSR_IMP-SSD.94% (ours)85K37.51 / 0.958633.09 / 0.913731.88 / 0.896030.94 / 0.9163
3×SRCNN [7] (ECCV' 2014)57K32.75 / 0.909029.28 / 0.820928.41 / 0.786326.24 / 0.7989
FSRCNN [8] (ECCV' 2016)12K33.16 / 0.914029.43 / 0.824228.53 / 0.791026.43 / 0.8080
CARN [19] (ECCV' 2018)415K34.00 / 0.924030.07 / 0.836828.92 / 0.800727.53 / 0.8379
ESRN [31] (AAAI' 2020)324K34.23 / 0.926230.27 / 0.840029.03 / 0.803927.95 / 0.8481
EDSR [23] (CVPRW' 2017)1558K34.37 / 0.926730.31 / 0.840929.07 / 0.804328.05 / 0.8506
EDSR_IMP-SSD.88% (ours)195K34.15 / 0.925230.20 / 0.839229.00 / 0.802827.81 / 0.8449
EDSR_IMP-SSD.94% (ours)97K33.74 / 0.921629.92 / 0.834028.81 / 0.797627.16 / 0.8298
4×SRCNN [7] (ECCV' 2014)57K30.48 / 0.862827.49 / 0.750326.90 / 0.710124.52 / 0.7221
FSRCNN [8] (ECCV' 2016)12K30.71 / 0.865727.59 / 0.753526.98 / 0.715024.62 / 0.7280
CARN [19] (ECCV' 2018)415K31.78 / 0.889428.39 / 0.776227.42 / 0.730325.56 / 0.7674
ESRN [31] (AAAI' 2020)324K31.99 / 0.891928.49 / 0.777927.50 / 0.733125.87 / 0.7782
EDSR [23] (CVPRW' 2017)1518K32.09 / 0.893728.55 / 0.780627.54 / 0.734825.97 / 0.7819
EDSR_IMP-SSD.88% (ours)190K31.89 / 0.891128.44 / 0.777727.46 / 0.732625.71 / 0.7743
EDSR_IMP-SSD.94% (ours)95K31.74 / 0.887928.33 / 0.774927.39 / 0.729725.48 / 0.7657
+ +![](images/d919ee3f6b8728f97d6403df8378b5a0642b399578e0f637a35a36ac50bfda7f.jpg) +Figure 6. Visual comparisons of IMP-SSD with other lightweight SR models on Urban100 for $4 \times$ SR. + +In particular, compared with other lightweight SR models such as CARN and ESRN, our IMP-SSD in EDSR with the sparsity of $88\%$ achieves comparable performance while our model size is about half or two-thirds of those. Compared with smaller SR models such as SRCNN and FSRCNN, IMP-SSD with the sparsity of $94\%$ , IMP-SSD achieves significant gains in PSNR by at least 0.4dB, 0.3dB, 0.4dB for $2\times$ , $3\times$ , and $4\times$ SR, respectively. The same conclusion is made in SSIM. Furthermore, compared with the original EDSR, IMP-SSD makes the SR model shrink to reach a very small size but their SR performance does not degrade very much. Compared with the original EDSR, the maximum decreases are 0.4dB, 0.2dB, 0.2dB for IMP-SSD with the sparsity of $88\%$ and 0.6dB, 0.9dB, 0.5dB for IMP-SSD with the sparsity of $94\%$ , respectively. It demonstrates that our method is effective on scalable SR. + +We also give the comparison of our method in visual effect in Fig. 6. It is observed that our method achieves a better visual effect than other lightweight SR models even with larger model sparsity. + +# 5. Conclusions + +In this paper, we propose a novel memory-friendly scalable lightweight framework (MSSR) with rewinding LTH, which can be deployed on arbitrary computational resources according to customization. MSSR is backtracking and includes two stages: the forward stage for model compression and the backward for model expansion. For the first time, LTH with rewinding weights is implemented to explore the winning tickets of deep SR models. We use IMP to gradually shrink the SR model and not only obtain the scalable lightweight model but also obtain the pruning-out masks which form the nested sets. In order to improve the performance of the shrinking model, SSD is conducted. Unlike the traditional SD, we replace the feature maps on several layers of Teacher with the counterparts of Students and conduct distillation. In the backward stage, we gradually recover the lightweight SR models depending on the pruning-out masks and then fine-tune the current model. We make instantiation on several state-of-the-art SR models. The extensive experimental results are effective. + +Acknowledgments. This work is supported by grants from the National Key Research and Development Program of China (No.2020AAA0108301), National Natural Science Foundation of China (No.62176224, No.62222602), and CCF-Lenovo Blue Ocean Research Fund. + +# References + +[1] Efficient non-local contrastive attention for image superresolution. In AAAI, 2022. 2, 7 +[2] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017. 6 +[3] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 2010. 6 +[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 6 +[5] Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Zhangyang Wang, and Michael Carbin. The lottery ticket hypothesis for pre-trained bert networks. In NeurIPS, 2020. 3 +[6] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2020. 2, 3 +[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016. 7, 8 +[8] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016. 7, 8 +[9] Utku Evci, Fabian Pedregosa, Aidan Gomez, and Erich Elsen. The difficulty of training sparse neural networks. In ICLR, 2021. 3 +[10] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 2, 3 +[11] Jonathan Frankle, David J Schwab, and Ari S Morcos. The early phase of neural network training. In ICLR, 2020. 3 +[12] Sharath Girish, Shishira R Maiya, Kamal Gupta, Hao Chen, Larry S Davis, and Abhinav Shrivastava. The lottery ticket hypothesis for object recognition. In CVPR, 2021. 3 +[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 3 +[14] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. TPAMI, 2021. 2 +[15] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In ICCV, 2017. 3 +[16] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 6 +[17] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM MM, 2019. 1, 2, 6 +[18] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 1, 2 + +[19] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 1, 2, 7, 8 +[20] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. 1 +[21] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. In NeurIPS, 1989. 3 +[22] Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Dynamic slimmable network. In CVPR, 2021. 2 +[23] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 6, 7, 8 +[24] Ning Liu, Geng Yuan, Zhengping Che, Xuan Shen, Xiaolong Ma, Qing Jin, Jian Ren, Jian Tang, Sijia Liu, and Yanzhi Wang. Lottery ticket preserves weight correlation: Is it desirable or not? In ICML, 2021. 2, 3, 4 +[25] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 3 +[26] Xiaotong Luo, Mingliang Dai, Yulun Zhang, Yuan Xie, Ding Liu, Yanyun Qu, Yun Fu, and Junping Zhang. Adjustable memory-efficient image super-resolution via individual kernel sparsity. In ACM MM, 2022. 2 +[27] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, 2020. 2 +[28] Sai Prasanna, Anna Rogers, and Anna Rumshisky. When bert plays the lottery, all tickets are winning. In ACL, 2020. 3 +[29] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. In ICLR, 2020. 3 +[30] Pedro Savarese, Hugo Silva, and Michael Maire. Winning the lottery with continuous sparsification. In NeurIPS, 2020. 3 +[31] Dehua Song, Chang Xu, Xu Jia, Yiyi Chen, Chunjing Xu, and Yunhe Wang. Efficient residual dense block search for image super-resolution. In AAAI, 2020. 7, 8 +[32] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. In ICLR, 2020. 3 +[33] Shihui Yin, Kyu-Hyoun Kim, Jinwook Oh, Naigang Wang, Mauricio Serrano, Jae-Sun Seo, and Jungwook Choi. The sooner the better: Investigating structure of early winning lottery tickets. In ICLR, 2020. 3 +[34] Haonan Yu, Sergey Edunov, Yuandong Tian, and Ari S Morcos. Playing the lottery with rewards and multiple languages: lottery tickets in rl and nlp. In ICLR, 2020. 3 +[35] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In ICCV, 2010. 6 +[36] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In ICCV, 2019. 4 + +[37] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 6, 7 +[38] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 6, 7 +[39] Hao Zhou, Jose M Alvarez, and Fatih Porikli. Less is more: Towards compact cnns. In ECCV, 2016. 3 \ No newline at end of file diff --git a/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/images.zip b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..63e37dab13bd9ada53988bff8d5b0df947112c8a --- /dev/null +++ b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:430ac1d8f5baec530b5a8751f28671aa23ee5f97d6c2bca08b9025d60012af88 +size 727048 diff --git a/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/layout.json b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d0fd6f2f09778670081908799db35415474e5f75 --- /dev/null +++ b/2023/Memory-Friendly Scalable Super-Resolution via Rewinding Lottery Ticket Hypothesis/layout.json @@ -0,0 +1,9528 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 136, + 466, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 136, + 466, + 172 + ], + "spans": [ + { + "bbox": [ + 114, + 136, + 466, + 172 + ], + "type": "text", + "content": "Memory-friendly Scalable Super-resolution via Rewinding Lottery Ticket Hypothesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "spans": [ + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "text", + "content": "Jin Lin" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "text", + "content": ", Xiaotong Luo" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "text", + "content": ", Ming Hong" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "text", + "content": ", Yanyun Qu" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "text", + "content": ", Yuan Xie" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "text", + "content": ", Zongze Wu" + }, + { + "bbox": [ + 107, + 191, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 207, + 424, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 207, + 424, + 220 + ], + "spans": [ + { + "bbox": [ + 153, + 207, + 424, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 207, + 424, + 220 + ], + "type": "text", + "content": "School of Informatics, Xiamen University, Fujian, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 220, + 511, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 220, + 511, + 250 + ], + "spans": [ + { + "bbox": [ + 65, + 220, + 511, + 250 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 220, + 511, + 250 + ], + "type": "text", + "content": "School of Computer Science and Technology, East China Normal University, Shanghai, China \n" + }, + { + "bbox": [ + 65, + 220, + 511, + 250 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 220, + 511, + 250 + ], + "type": "text", + "content": "School of Mechatronics and Control Engineering, Shenzhen University, Shenzhen, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 250, + 437, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 250, + 437, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 250, + 437, + 262 + ], + "type": "text", + "content": "{linjin, xiaotluo, mingh}@stu.xmu.edu.cn, yyqu@xmu.edu.cn," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 265, + 379, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 265, + 379, + 275 + ], + "spans": [ + { + "bbox": [ + 188, + 265, + 379, + 275 + ], + "type": "text", + "content": "yxie@cs.ecnu.edu.cn, zzwu@szu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 303, + 186, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 303, + 186, + 314 + ], + "spans": [ + { + "bbox": [ + 139, + 303, + 186, + 314 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 328, + 281, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 328, + 281, + 689 + ], + "spans": [ + { + "bbox": [ + 45, + 328, + 281, + 689 + ], + "type": "text", + "content": "Scalable deep Super-Resolution (SR) models are increasingly in demand, whose memory can be customized and tuned to the computational recourse of the platform. The existing dynamic scalable SR methods are not memory-friendly enough because multi-scale models have to be saved with a fixed size for each model. Inspired by the success of Lottery Tickets Hypothesis (LTH) on image classification, we explore the existence of unstructured scalable SR deep models, that is, we find gradual shrinkage subnetworks of extreme sparsity named winning tickets. In this paper, we propose a Memory-friendly Scalable SR framework (MSSR). The advantage is that only a single scalable model covers multiple SR models with different sizes, instead of reloading SR models of different sizes. Concretely, MSSR consists of the forward and backward stages, the former for model compression and the latter for model expansion. In the forward stage, we take advantage of LTH with rewinding weights to progressively shrink the SR model and the pruning-out masks that form nested sets. Moreover, stochastic self-distillation (SSD) is conducted to boost the performance of sub-networks. By stochastically selecting multiple depths, the current model inputs the selected features into the corresponding parts in the larger model and improves the performance of the current model based on the feedback results of the larger model. In the backward stage, the smaller SR model could be expanded by recovering and fine-tuning the pruned parameters according to the pruning-out masks obtained in the forward. Extensive experiments show the effectiveness of MMSR. The smallest-scale sub-network could achieve the sparsity of " + }, + { + "bbox": [ + 45, + 328, + 281, + 689 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 45, + 328, + 281, + 689 + ], + "type": "text", + "content": " and outperforms the compared lightweight SR methods." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 303, + 302, + 526, + 421 + ], + "blocks": [ + { + "bbox": [ + 303, + 302, + 526, + 421 + ], + "lines": [ + { + "bbox": [ + 303, + 302, + 526, + 421 + ], + "spans": [ + { + "bbox": [ + 303, + 302, + 526, + 421 + ], + "type": "image", + "image_path": "a5cc0ce30a2990a07f9b58171cda1cd176768b8e9c3c0d6c539f47ed29203d2e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 426, + 530, + 511 + ], + "lines": [ + { + "bbox": [ + 296, + 426, + 530, + 511 + ], + "spans": [ + { + "bbox": [ + 296, + 426, + 530, + 511 + ], + "type": "text", + "content": "Figure 1. The flowchart of scalable SR network. S_n and S_w denote the neurons and the neural connections (weights) of the simplest subnetwork; M_n and M_w denote the intermediate neurons and neural connections; Cur_n and Cur_w denote the specific neurons and neural connections belonging to the current subnetwork. Free_n denotes the pruning-out neurons. The final model (with " + }, + { + "bbox": [ + 296, + 426, + 530, + 511 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 296, + 426, + 530, + 511 + ], + "type": "text", + "content": " recovered parameters) reaches the original size. The scalable model is adjustable to the memory resource allocation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 298, + 533, + 375, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 533, + 375, + 545 + ], + "spans": [ + { + "bbox": [ + 298, + 533, + 375, + 545 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 554, + 530, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 554, + 530, + 636 + ], + "spans": [ + { + "bbox": [ + 296, + 554, + 530, + 636 + ], + "type": "text", + "content": "Single image super-resolution (SISR) aims to reconstruct a high-resolution (HR) image from the corresponding low-resolution (LR) one. With the rising of deep learning, deep SR methods have made incredible progress. However, the existing SR models mostly require computational and memory resources, so they do not favor resource-limited devices such as mobile phones, robotics, and some edge devices." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": "The lightweight SR methods are attracting more attention for better application to resource-limited devices. The existing lightweight SR methods mainly focus on designing compact architectures [17, 20] with a fixed size, such as multi-scale pyramid [20], multiple-level receptive fields [17, 18], and recursive learning [19]. However, most lightweight SR models with fixed sizes are not flexible in applications. If one model does not match the resources of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 709, + 120, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 709, + 120, + 719 + ], + "spans": [ + { + "bbox": [ + 57, + 709, + 120, + 719 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "spans": [ + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "14398" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "type": "text", + "content": "the platform, it has to be retrained by compression methods to match the resources and then reloaded onto the devices." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 131, + 278, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 131, + 278, + 341 + ], + "spans": [ + { + "bbox": [ + 45, + 131, + 278, + 341 + ], + "type": "text", + "content": "The urgent demand to customize models based on deployment resources is increasing. Dynamic neural networks for SR [14,22] are proposed to adjust the network architecture according to different computational resources. The existing dynamic deep SR models often explore dynamic depth or width [22, 26], but they either require large memory resources or are not convenient for users to wait for retraining another SR model. The former leads to saving the multi-scale SR models of different sizes and the latter leads to retraining the model before being reused again. The limitation lies in that they are not memory-friendly. In many edge-device SR applications, the devices may be scalable, that is, their memories may be small in the beginning and be expanded later. Thus, we discuss two issues in this paper: 1) how to make a scalable lightweight model for the multi-scale computational recourse. 2) how to make the lightweight model expand to a larger-size model for better performance if the computational recourse is increased." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 342, + 278, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 342, + 278, + 516 + ], + "spans": [ + { + "bbox": [ + 45, + 342, + 278, + 516 + ], + "type": "text", + "content": "As for the first issue, inspired by the success of Lottery Ticket Hypothesis [10] which points out that there could exist a properly pruned sub-network named winning tickets to achieve comparable performance against the original dense network in model compression of classification, it is used to find the sub-network for SR. We are the first to study the existence of scalable winning tickets for SR. Iterative pruning and rewinding weights in LTH are beneficial to the scalable lightweight SR model. Iterative pruning may compress the SR model according to an arbitrary size. It is observed in [24] that the winning tickets are related to an insufficient DNN, and rewinding LTH outperforms the original LTH. That is, the initial weights in LTH are replaced with the T-iteration weights during pruning and fine-tuning. The scalable deep SR model is shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 518, + 278, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 518, + 278, + 588 + ], + "spans": [ + { + "bbox": [ + 45, + 518, + 278, + 588 + ], + "type": "text", + "content": "As for the second issue, the scalable SR model can customize parameters to adapt to different memory resources rather than load or offload different models for different devices. In other words, during real applications, there will be only one simple model to be employed for inference whose size is decided by the computational resource." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 590, + 278, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 590, + 278, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 590, + 278, + 729 + ], + "type": "text", + "content": "In this paper, we propose a memory-friendly scalable deep SR model (MSSR) via rewinding LTH. We use the rewinding LTH [10] to generate our unstructured scalable mask. MSSR is backtracking and contains forward and backward stages. The former focuses on model compression by rewinding LTH with iterative pruning and finetuning, and the latter focuses on iterative model expansion until it goes back to its original size. Multi-scale winning tickets together with the pruning-out masks are obtained by rewinding LTH in the forward stage with the decrease in the number of parameters. The pruning-out masks are nested. In order to make the compressed SR model not degrade sig" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "content": "nificantly, stochastic self-distillation (SSD) is used to improve the representation of the small-scale SR model, and knowledge is transferred from the last-scale model to the current scale model. In the backward stage, the smallest model is expanded gradually to the model with the original size with the expanded mask." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 309, + 177, + 508, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 177, + 508, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 177, + 508, + 188 + ], + "type": "text", + "content": "The main contributions of this work are three-fold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 195, + 530, + 373 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 308, + 195, + 530, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 530, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 530, + 253 + ], + "type": "text", + "content": "- A memory-friendly scalable dynamic SR lightweight model via rewinding LTH is proposed. MSSR is re-configurable and switchable to sub-networks with different sizes according to on-device resource constraints on the fly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 261, + 530, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 530, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 530, + 330 + ], + "type": "text", + "content": "- MSSR is backtracking, which contains forward and backward stages. Multi-scale winning tickets form nested masks for the multi-scale models. SSD is conducted by replacing the features in randomly selected layers between Teacher and Student to improve the performance of the scalable SR lightweight models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 338, + 530, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 530, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 530, + 373 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that MSSR can generalize to different SR models as well as state-of-the-art attention-based models, ENLCN [1]." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 387, + 381, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 387, + 381, + 399 + ], + "spans": [ + { + "bbox": [ + 297, + 387, + 381, + 399 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 406, + 490, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 406, + 490, + 419 + ], + "spans": [ + { + "bbox": [ + 297, + 406, + 490, + 419 + ], + "type": "text", + "content": "2.1. Lightweight Deep Learning-based SR" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 426, + 530, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 426, + 530, + 612 + ], + "spans": [ + { + "bbox": [ + 296, + 426, + 530, + 612 + ], + "type": "text", + "content": "SISR aims to generate a high-resolution image from the low-resolution version. Recently, efficient and lightweight SISR networks have attracted increasing interest in the computer vision community. The lightweight SR methods consist of static and dynamic neural networks for SR. The earlier lightweight SR methods mainly focus on the static network structure design. IDN [18] utilizes group convolution and combined the local long and short path features effectively. CARN [19] utilizes a recursive cascading mechanism for learning multi-level feature representations. IMDN [17] introduces information multi-distillation blocks for extracting hierarchical features. FALSR [6] applies neural architecture search for efficient network design automatically. LatticeNet [27] designs a lattice block to combine two residual blocks adaptively. However, the aforementioned lightweight SR models need elaborate designs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 612, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 612, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 296, + 612, + 530, + 729 + ], + "type": "text", + "content": "Recently, dynamic lightweight SR models are demanded increasingly whose size is adjustable according to different computational resources. DS-Net [22] makes the width of the SR model adjustable, and IKS [26] makes the sparsity of individual kernels adjustable. However, they require saving lightweight SR models with different shrinkage sizes, which is not convenient for users to offload the old model and then load a new one. Our method uses a scalable SR model adjustable to the changing computational recourse to avoid the loading and offloading models." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "14399" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 105, + 526, + 372 + ], + "blocks": [ + { + "bbox": [ + 49, + 105, + 526, + 372 + ], + "lines": [ + { + "bbox": [ + 49, + 105, + 526, + 372 + ], + "spans": [ + { + "bbox": [ + 49, + 105, + 526, + 372 + ], + "type": "image", + "image_path": "4cbf2900c4189795dcdd7620250b00d900293e7ce223b8daa04e70e62a4b6e42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 373, + 531, + 428 + ], + "lines": [ + { + "bbox": [ + 45, + 373, + 531, + 428 + ], + "spans": [ + { + "bbox": [ + 45, + 373, + 531, + 428 + ], + "type": "text", + "content": "Figure 2. The architecture of memory-friendly scalable deep SR model (MSSR). It contains two stages: the forward stage for model compression and the backward stage for model expansion. In the forward stage, the original SR network is firstly processed by IMP to obtain winning tickets and the pruning-out mask " + }, + { + "bbox": [ + 45, + 373, + 531, + 428 + ], + "type": "inline_equation", + "content": "LM_{1}" + }, + { + "bbox": [ + 45, + 373, + 531, + 428 + ], + "type": "text", + "content": ", and then the winning tickets are processed by SSD and the lightweight subnetwork is obtained. So do the other gradual shrinkage sub-networks and " + }, + { + "bbox": [ + 45, + 373, + 531, + 428 + ], + "type": "inline_equation", + "content": "\\{LM_{i}\\}" + }, + { + "bbox": [ + 45, + 373, + 531, + 428 + ], + "type": "text", + "content": " forms the nested sets. In the backward stage, according to the pruning-out masks, the smaller lightweight SR model is gradually expanded and finely tuned to obtain a larger lightweight SR model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 442, + 246, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 442, + 246, + 455 + ], + "spans": [ + { + "bbox": [ + 45, + 442, + 246, + 455 + ], + "type": "text", + "content": "2.2. Pruning and Lottery Ticket Hypothesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 461, + 279, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 461, + 279, + 659 + ], + "spans": [ + { + "bbox": [ + 45, + 461, + 279, + 659 + ], + "type": "text", + "content": "Model pruning is to prune away the unimportant connections (parameters) [21] in the neural network which trades off between the model size and performance. The existing pruning methods can be classified into unstructured [6, 13, 21] and structured methods [15, 25, 39]. The structured pruning methods remove the substructures of the model, such as channels, filters, and layers, while the unstructured pruning methods are magnitude-based, where the weights are removed below the threshold. However, magnitude-based pruning methods are not hardware-friendly and computationally efficient. Lottery Ticket Hypothesis (LTH) [10] opened a pandas's box of immense possibilities in the field of pruning and sparse models. It assumes that there exist sparse sub-networks independently training from scratch that can match the performance of dense networks. Some literatures discuss the effectiveness of pruning, fine-tuning [24], and rewinding weights [29] for LTH." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 660, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 280, + 730 + ], + "type": "text", + "content": "LTH has been widely applied in image classification [9-12, 30, 32, 33], natural language processing [5, 28, 29, 34], and so on. However, few works discuss the existence of LTH in the deep model for image restoration. We are the first to investigate the existence of scalable winning tickets for lightweight deep SR models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 297, + 441, + 351, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 441, + 351, + 452 + ], + "spans": [ + { + "bbox": [ + 297, + 441, + 351, + 452 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 463, + 406, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 463, + 406, + 474 + ], + "spans": [ + { + "bbox": [ + 297, + 463, + 406, + 474 + ], + "type": "text", + "content": "3.1. Overview of MSSR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 483, + 530, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 483, + 530, + 542 + ], + "spans": [ + { + "bbox": [ + 296, + 483, + 530, + 542 + ], + "type": "text", + "content": "In this section, we introduce the memory-friendly scalable SR model which is an all-in-one model with multiscale SR models sharing a model. As shown in Fig. 2, our MSSR consists of two stages: the forward stage for model compression, and the backward stage for model expansion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 543, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 543, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 543, + 531, + 730 + ], + "type": "text", + "content": "In the forward stage, we use rewinding LTH (RLTH) to iteratively shrink a dense SR model to obtain a sub-network with the small size not significantly scarifying the performance. During pruning by LTH each time, we treat the current lightweight SR model as Student and its last ancestor lightweight SR model as Teacher, and conduct stochastic self-distillation between Teacher and Student. In Fig. 2, the dark-blue block denotes the winning tickets, and the white block denotes the pruning part. We also record the mask of pruning part and use a matrix with \"1\" to represent the pruning-out weights and with \"0\" for others. It is observed that the masks are nested and the initial mask is continuously contained in the later mask. In the backward stage, we iteratively expand the lightweight sub-network according to the saved nested masks. The most lightweight network marked by 1 is expanded iteratively to the gradual less" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "14400" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "content": "lightweight network, the expanding order is shown as \"1\", \"1-2\", \"1-2-3\", and \"1-2-3-4\"." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 138, + 255, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 138, + 255, + 151 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 255, + 151 + ], + "type": "text", + "content": "3.2. The Forward Stage: Model Compression" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "spans": [ + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": "The forward streamline is shown in Fig. 3. Let a network denoted by " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "f(\\theta_0; x)" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": " and the pruned masks which form the \"nested sets\" denoted by " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "LM = \\{\\}" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": " denotes the initial weights. The flowchart shows a whole process in an iteration for shrinking a model and it contains three components: SR model pre-train, IMP with rewinding weights, and SSD. Firstly, an SR network is trained to reach a steady state, and its weights are named the first nonpruning weights " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": ". After that, Iterative Magnitude Pruning is conducted on the SR model " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": ", and the weights below the pre-set threshold are pruned out which forms a pruning mask " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^n" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": ". Following [24], we rewind the remaining weights to their values in " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "\\theta' = \\theta \\odot m" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": ". Then, we retrain the pruned weights " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": " by SSD arriving at " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": ". So do the next more lightweight SR model with the model weights " + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 45, + 156, + 279, + 341 + ], + "type": "text", + "content": " as the initiation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 343, + 279, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 343, + 279, + 413 + ], + "spans": [ + { + "bbox": [ + 45, + 343, + 279, + 413 + ], + "type": "text", + "content": "LTH with Rewinding for SR. Benefiting from LTHs, we implement iterative magnitude pruning with rewinding to gradually shrink the SR model. The motivation lies in the controllability and flexibility of LTH. We can shrink the model arbitrarily by LTH, which favors the scalable lightweight SR model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 414, + 279, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 414, + 279, + 552 + ], + "spans": [ + { + "bbox": [ + 45, + 414, + 279, + 552 + ], + "type": "text", + "content": "We mainly consider two issues: the model pruning process and the pruning rules. The pruning process corresponds to the process of gradually shrinking the model, and the pruning methods need to meet the following rules: 1) the amount of prune-out parameters is arbitrarily adjustable; 2) the model is shrunk gradually, which leads to the nested sets and the larger model contains its smaller descendant. LTH fits well with the needs of our scalable mask generator and the appropriate use of its task settings to implement the model compression process, in particular discarding its over parameterization, a feature that is not suitable for intensive prediction tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 554, + 279, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 554, + 279, + 612 + ], + "spans": [ + { + "bbox": [ + 45, + 554, + 279, + 612 + ], + "type": "text", + "content": "IMP is proven to be effective for LTH. In [24], it is proved that weight rewinding is better than the original weight initiation. That is, unpruned weights are rewound to their values from " + }, + { + "bbox": [ + 45, + 554, + 279, + 612 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 554, + 279, + 612 + ], + "type": "text", + "content": "-iteration training and retrained from there using the original training schedule." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "spans": [ + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "text", + "content": "Compared with one-shot magnitude pruning (OMP) which reaches the goal model at one time, IMP does not shrink a model to reach the pre-set size at one time but shrinks a model with several iterations and the model size is evenly decreased at each iteration. Let the network be pruned for " + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "text", + "content": " times and be expected to discard " + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "text", + "content": " of the parameters. The weights/parameters of the network are represented by " + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "inline_equation", + "content": "\\omega \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "text", + "content": " and the pruning mask by " + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^n" + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 45, + 612, + 279, + 717 + ], + "type": "text", + "content": " represents all the capacity of the network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 718, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 718, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 57, + 718, + 279, + 730 + ], + "type": "text", + "content": "Weight rewinding does not initiate the model with the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 299, + 104, + 523, + 233 + ], + "blocks": [ + { + "bbox": [ + 299, + 104, + 523, + 233 + ], + "lines": [ + { + "bbox": [ + 299, + 104, + 523, + 233 + ], + "spans": [ + { + "bbox": [ + 299, + 104, + 523, + 233 + ], + "type": "image", + "image_path": "c4dc5c98f9c2e90ab20862a08e1e97e9d7a22e8c2c9a4de97d2fe75e732194cd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "lines": [ + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "spans": [ + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "text", + "content": "Figure 3. The flowchart of the whole process of IMP and SSD. The initial model " + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "text", + "content": " is pre-trained to obtain a stable model " + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "text", + "content": ". IMP is implemented on " + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "text", + "content": " and the winning ticket (sub-network) is obtained. The sub-network is initiated by " + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "inline_equation", + "content": "\\theta \\odot m" + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "text", + "content": " and is processed by SSD for the better model " + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 296, + 243, + 530, + 296 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "spans": [ + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": "original initiation weights " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "\\theta_0\\odot m" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " , but with the " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " -iteration " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "\\theta_t\\odot m" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "0 < t\\leq T" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " , resulting in weights " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " and the network function " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "f(x;\\theta_t\\odot m)" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " . We highlight that the original LTH always chooses " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " as dense model, while in this work we choose " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " from the last pruning iteration. After \"Pruning\" and \"Rewinding and Retraining\", " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "\\theta = \\hat{\\theta}" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " is updated and the mask at the " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " -th iteration is obtained which is denoted as " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "LM_{t} = LM_{t - 1}\\cup \\{m\\}" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " . In other words, the mask " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "LM_{t}" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " -th iteration is contained in the mask " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "LM_{t - 1}" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " -th. \"Pruning\" and \"Rewinding and Retraining\" are repeated until the model reaches the end condition, which is the iteration steps " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "text", + "content": " or target sparsity " + }, + { + "bbox": [ + 296, + 309, + 530, + 448 + ], + "type": "inline_equation", + "content": "p" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 449, + 530, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 449, + 530, + 507 + ], + "spans": [ + { + "bbox": [ + 296, + 449, + 530, + 507 + ], + "type": "text", + "content": "Stochastic Self-distillation for SR. In the forward stage, due to the progressive shrinking of the SR model, it is definitely difficult to maintain the performance as same as the original model. Especially, when only " + }, + { + "bbox": [ + 296, + 449, + 530, + 507 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 296, + 449, + 530, + 507 + ], + "type": "text", + "content": " of the parameters are kept, most magnitude-pruning methods do not work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 508, + 530, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 508, + 530, + 693 + ], + "spans": [ + { + "bbox": [ + 296, + 508, + 530, + 693 + ], + "type": "text", + "content": "In order to mitigate the degradation of lightweight SR models, we adopt self-distillation [36] to boost the performance of the pruned model. Unlike the existing self-distillation which uses the function of features obtained by the model at previous iterations such as the average and summation function, we introduce attribution analysis to self-distillation. We randomly select several layers in Student, and the feature maps of selected layers replace the counterpart of Teacher, which is named Stochastic self-distillation (SSD). Concretely, the current lightweight SR model is regarded as the student network (Student), and its last ancestor network is regarded as the teacher network (Teacher), and Student is obtained by pruning Teacher. SSD is conducted between Teacher and Student. Student is the pruned model of Teacher. Let " + }, + { + "bbox": [ + 296, + 508, + 530, + 693 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 296, + 508, + 530, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 508, + 530, + 693 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 296, + 508, + 530, + 693 + ], + "type": "text", + "content": " denote the network input and ground truth, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": "The architecture of SSD is shown in Fig. 4, which contains two streamlines: Teacher and Student. The knowledge in feature maps is transferred from Teacher to Student." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "type": "text", + "content": "14401" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 104, + 260, + 255 + ], + "blocks": [ + { + "bbox": [ + 61, + 104, + 260, + 255 + ], + "lines": [ + { + "bbox": [ + 61, + 104, + 260, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 104, + 260, + 255 + ], + "type": "image", + "image_path": "d3fa501b40f78f274e3ace7e22932cd8d736c9cb22d1f3ef6857b4b25ed35364.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 259, + 280, + 323 + ], + "lines": [ + { + "bbox": [ + 45, + 259, + 280, + 323 + ], + "spans": [ + { + "bbox": [ + 45, + 259, + 280, + 323 + ], + "type": "text", + "content": "Figure 4. The flowchart of SSD. Student is the sub-network by pruning Teacher. Several layers of Student are selected randomly. The feature maps in a selected layer replace the counterpart of Teacher which outputs the SR result, and so do other selected layers. The total loss is the summation of the knowledge distillation loss " + }, + { + "bbox": [ + 45, + 259, + 280, + 323 + ], + "type": "inline_equation", + "content": "L_{skd}" + }, + { + "bbox": [ + 45, + 259, + 280, + 323 + ], + "type": "text", + "content": " and the fidelity loss " + }, + { + "bbox": [ + 45, + 259, + 280, + 323 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 45, + 259, + 280, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "spans": [ + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "text", + "content": "If Student contains " + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "text", + "content": " layers, at each-iteration distillation, " + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "text", + "content": " layers randomly are selected. In realization, a one-hot vector is randomly generated, named Switch Vector, with a length of " + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "text", + "content": " in each training, " + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{0,1\\}^d" + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "text", + "content": ". If the element is equal to 1 in " + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 338, + 279, + 432 + ], + "type": "text", + "content": ", the feature maps in the layer corresponding to this position are selected, and are named Lucky feature maps. Then, Lucky feature maps in the selected layers to replace the counterparts of Teacher." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 433, + 279, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 433, + 279, + 468 + ], + "spans": [ + { + "bbox": [ + 45, + 433, + 279, + 468 + ], + "type": "text", + "content": "In details, Student firstly is trained for a stable SR model. After that, SSD is conducted. At each iteration, Switch vector is dynamically generated, which is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 470, + 279, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 470, + 279, + 482 + ], + "spans": [ + { + "bbox": [ + 117, + 470, + 279, + 482 + ], + "type": "interline_equation", + "content": "D C = \\left[ \\alpha_ {1}, \\alpha_ {2}, \\dots , \\alpha_ {d} \\right], \\tag {1}", + "image_path": "2260884400599868d703d2a727b6d184c3b4825020f158919d8286f5aeaa21fc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "spans": [ + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "text", + "content": "-th element corresponding to the " + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "text", + "content": "-th layer and " + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "text", + "content": " is the layer number in Student. Then, lucky feature maps of Student are selected according to " + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 45, + 484, + 279, + 530 + ], + "type": "text", + "content": ", and they are formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 532, + 279, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 532, + 279, + 546 + ], + "spans": [ + { + "bbox": [ + 113, + 532, + 279, + 546 + ], + "type": "interline_equation", + "content": "\\mathrm {L F} = \\left[ f _ {\\alpha_ {1}} ^ {s}, f _ {\\alpha_ {2}} ^ {s}, \\dots , f _ {\\alpha_ {d}} ^ {s} \\right], \\tag {2}", + "image_path": "bb1a5d9a5973bb70e417eb217926bcc107548f68aafa203314230901bfcebd13.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 548, + 279, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 548, + 279, + 582 + ], + "spans": [ + { + "bbox": [ + 45, + 548, + 279, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 548, + 279, + 582 + ], + "type": "inline_equation", + "content": "f_{\\alpha}^{s}" + }, + { + "bbox": [ + 45, + 548, + 279, + 582 + ], + "type": "text", + "content": " denotes the feature obtained from Student. Note that all elements in Switch Vector are independent and identically distributed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 586, + 279, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 586, + 279, + 631 + ], + "spans": [ + { + "bbox": [ + 45, + 586, + 279, + 631 + ], + "type": "text", + "content": "Lucky feature maps replace the counterpart of Teacher each time, and the corresponding output of new Teacher is denoted by " + }, + { + "bbox": [ + 45, + 586, + 279, + 631 + ], + "type": "inline_equation", + "content": "Y_{\\alpha_i}^t" + }, + { + "bbox": [ + 45, + 586, + 279, + 631 + ], + "type": "text", + "content": ", and the corresponding weights of new Teacher are denoted by " + }, + { + "bbox": [ + 45, + 586, + 279, + 631 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 45, + 586, + 279, + 631 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 633, + 279, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 633, + 279, + 659 + ], + "spans": [ + { + "bbox": [ + 93, + 633, + 279, + 659 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} Y _ {\\alpha_ {i}} ^ {t} = F \\left(\\theta^ {t}; x, f _ {\\alpha_ {i}} ^ {s}\\right), i = 1, \\dots , d, \\\\ \\mathrm {L Y} = \\left[ Y _ {\\alpha_ {1}} ^ {t}, Y _ {\\alpha_ {2}} ^ {t}, \\dots , Y _ {\\alpha_ {d}} ^ {t} \\right], \\end{array} \\right. \\tag {3}", + "image_path": "84436bde33329d08691a5a606eebd12d395b54be13542d8c3852fa7d49a8d9b3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "inline_equation", + "content": "LY" + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "text", + "content": " denotes the set of SR outputs, " + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "inline_equation", + "content": "Y_{\\alpha}^{t}" + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "text", + "content": " is the SSD output of lucky maps, " + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "inline_equation", + "content": "f_{\\alpha}^{s}" + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "text", + "content": ". All the outputs between Student and Teacher are supervised by HR image " + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 45, + 660, + 279, + 705 + ], + "type": "text", + "content": ". The knowledge of Teacher takes more rationality to be distilled into Student." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 707, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 707, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 707, + 279, + 730 + ], + "type": "text", + "content": "Therefore, the total loss function utilized to train the SR model consists of the supervised loss " + }, + { + "bbox": [ + 45, + 707, + 279, + 730 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 45, + 707, + 279, + 730 + ], + "type": "text", + "content": " and self-distillation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 106, + 529, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 106, + 529, + 129 + ], + "spans": [ + { + "bbox": [ + 297, + 106, + 529, + 129 + ], + "type": "text", + "content": "Algorithm 1 Memory-friendlyly scalable lightweight framework for image SR." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 131, + 471, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 131, + 471, + 142 + ], + "spans": [ + { + "bbox": [ + 297, + 131, + 471, + 142 + ], + "type": "text", + "content": "Data: HR and LR image datasets: " + }, + { + "bbox": [ + 297, + 131, + 471, + 142 + ], + "type": "inline_equation", + "content": "D_{hr}" + }, + { + "bbox": [ + 297, + 131, + 471, + 142 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 297, + 131, + 471, + 142 + ], + "type": "inline_equation", + "content": "D_{lr}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 143, + 506, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 143, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 297, + 143, + 506, + 165 + ], + "type": "text", + "content": "Input: Dense network with randomly initial weights: " + }, + { + "bbox": [ + 297, + 143, + 506, + 165 + ], + "type": "inline_equation", + "content": "f(\\theta_0)" + }, + { + "bbox": [ + 297, + 143, + 506, + 165 + ], + "type": "text", + "content": ", pruning rounds: " + }, + { + "bbox": [ + 297, + 143, + 506, + 165 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 297, + 143, + 506, + 165 + ], + "type": "text", + "content": ", and pruning target rate: " + }, + { + "bbox": [ + 297, + 143, + 506, + 165 + ], + "type": "inline_equation", + "content": "p" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 298, + 165, + 457, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 165, + 457, + 176 + ], + "spans": [ + { + "bbox": [ + 298, + 165, + 457, + 176 + ], + "type": "text", + "content": "Output: Scalable lightweight SR model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 301, + 179, + 528, + 479 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 304, + 179, + 490, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 490, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 490, + 190 + ], + "type": "text", + "content": "1: /* The forward stage: model compression */" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 191, + 528, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 528, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 528, + 213 + ], + "type": "text", + "content": "2: Initializing: Pruned mask " + }, + { + "bbox": [ + 304, + 191, + 528, + 213 + ], + "type": "inline_equation", + "content": "m = \\{1\\}^n" + }, + { + "bbox": [ + 304, + 191, + 528, + 213 + ], + "type": "text", + "content": ", lottery mask pool " + }, + { + "bbox": [ + 304, + 191, + 528, + 213 + ], + "type": "inline_equation", + "content": "LM = \\{m\\}" + }, + { + "bbox": [ + 304, + 191, + 528, + 213 + ], + "type": "text", + "content": " and epochs of each training " + }, + { + "bbox": [ + 304, + 191, + 528, + 213 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 214, + 478, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 478, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 478, + 237 + ], + "type": "text", + "content": "3: Pre-training: Train network for " + }, + { + "bbox": [ + 304, + 214, + 478, + 237 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 214, + 478, + 237 + ], + "type": "text", + "content": " epochs " + }, + { + "bbox": [ + 304, + 214, + 478, + 237 + ], + "type": "inline_equation", + "content": "f(\\theta_0;x)\\longrightarrow f(\\theta ;x)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 238, + 390, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 390, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 390, + 247 + ], + "type": "text", + "content": "4: while " + }, + { + "bbox": [ + 304, + 238, + 390, + 247 + ], + "type": "inline_equation", + "content": "i < T - 1" + }, + { + "bbox": [ + 304, + 238, + 390, + 247 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "type": "text", + "content": "5: Prune bottom " + }, + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "type": "inline_equation", + "content": "p\\% \\times \\frac{1}{T}" + }, + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "type": "inline_equation", + "content": "m \\odot \\theta" + }, + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "type": "text", + "content": " and update " + }, + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 248, + 511, + 261 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "text", + "content": "6: Re-train the pruned sub-network " + }, + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime} = m\\odot \\theta" + }, + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "text", + "content": " epochs " + }, + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "inline_equation", + "content": "f(\\theta^{\\prime};x)\\longrightarrow f(\\hat{\\theta};x)" + }, + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 261, + 525, + 283 + ], + "type": "inline_equation", + "content": "SSD(\\theta ,\\theta^{\\prime};x)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 284, + 523, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 523, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 523, + 305 + ], + "type": "text", + "content": "7: Update mask pool " + }, + { + "bbox": [ + 304, + 284, + 523, + 305 + ], + "type": "inline_equation", + "content": "LM = LM \\cup \\{m\\}" + }, + { + "bbox": [ + 304, + 284, + 523, + 305 + ], + "type": "text", + "content": " and weights " + }, + { + "bbox": [ + 304, + 284, + 523, + 305 + ], + "type": "inline_equation", + "content": "\\theta = \\hat{\\theta}" + }, + { + "bbox": [ + 304, + 284, + 523, + 305 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 307, + 374, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 374, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 374, + 317 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 304, + 307, + 374, + 317 + ], + "type": "inline_equation", + "content": "i\\gets i + 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 319, + 357, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 357, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 357, + 329 + ], + "type": "text", + "content": "9: end while" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 302, + 330, + 520, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 330, + 520, + 353 + ], + "spans": [ + { + "bbox": [ + 302, + 330, + 520, + 353 + ], + "type": "text", + "content": "10: Generate the sparse sub-network matching the target sparsity, and the mask pool " + }, + { + "bbox": [ + 302, + 330, + 520, + 353 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 302, + 330, + 520, + 353 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 302, + 354, + 486, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 354, + 486, + 364 + ], + "spans": [ + { + "bbox": [ + 302, + 354, + 486, + 364 + ], + "type": "text", + "content": "11: /* The backward stage: model expansion */" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 301, + 366, + 486, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 366, + 486, + 388 + ], + "spans": [ + { + "bbox": [ + 301, + 366, + 486, + 388 + ], + "type": "text", + "content": "12: Initializing: Current step " + }, + { + "bbox": [ + 301, + 366, + 486, + 388 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 301, + 366, + 486, + 388 + ], + "type": "text", + "content": " and sparsest sub-network " + }, + { + "bbox": [ + 301, + 366, + 486, + 388 + ], + "type": "inline_equation", + "content": "\\theta_T = \\theta" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 301, + 389, + 374, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 389, + 374, + 398 + ], + "spans": [ + { + "bbox": [ + 301, + 389, + 374, + 398 + ], + "type": "text", + "content": "13: while " + }, + { + "bbox": [ + 301, + 389, + 374, + 398 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 301, + 389, + 374, + 398 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 301, + 400, + 511, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 400, + 511, + 423 + ], + "spans": [ + { + "bbox": [ + 301, + 400, + 511, + 423 + ], + "type": "text", + "content": "14: Calculate the grad mask " + }, + { + "bbox": [ + 301, + 400, + 511, + 423 + ], + "type": "inline_equation", + "content": "m = LM_{t - 1} - LM_t" + }, + { + "bbox": [ + 301, + 400, + 511, + 423 + ], + "type": "text", + "content": " to determine the trainable weights;" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 301, + 424, + 506, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 424, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 301, + 424, + 506, + 447 + ], + "type": "text", + "content": "15: Fine-tune/re-train the trainable weights for " + }, + { + "bbox": [ + 301, + 424, + 506, + 447 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 301, + 424, + 506, + 447 + ], + "type": "text", + "content": " epochs " + }, + { + "bbox": [ + 301, + 424, + 506, + 447 + ], + "type": "inline_equation", + "content": "f(\\theta_t;x)\\longrightarrow f(\\theta_t';x)" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 301, + 447, + 392, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 447, + 392, + 458 + ], + "spans": [ + { + "bbox": [ + 301, + 447, + 392, + 458 + ], + "type": "text", + "content": "16: Update " + }, + { + "bbox": [ + 301, + 447, + 392, + 458 + ], + "type": "inline_equation", + "content": "\\theta_t = \\theta_t'" + }, + { + "bbox": [ + 301, + 447, + 392, + 458 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 301, + 459, + 376, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 459, + 376, + 468 + ], + "spans": [ + { + "bbox": [ + 301, + 459, + 376, + 468 + ], + "type": "text", + "content": "17: " + }, + { + "bbox": [ + 301, + 459, + 376, + 468 + ], + "type": "inline_equation", + "content": "t\\gets t - 1;" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 301, + 470, + 356, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 470, + 356, + 479 + ], + "spans": [ + { + "bbox": [ + 301, + 470, + 356, + 479 + ], + "type": "text", + "content": "18: end while" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 504, + 432, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 504, + 432, + 516 + ], + "spans": [ + { + "bbox": [ + 297, + 504, + 432, + 516 + ], + "type": "text", + "content": "loss " + }, + { + "bbox": [ + 297, + 504, + 432, + 516 + ], + "type": "inline_equation", + "content": "L_{skd}" + }, + { + "bbox": [ + 297, + 504, + 432, + 516 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 305, + 524, + 529, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 524, + 529, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 524, + 529, + 562 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = L _ {1} + \\lambda L _ {s k d} = \\left\\| Y ^ {s} - Y \\right\\| _ {1} + \\frac {\\lambda}{d} \\sum_ {i = 1} ^ {d} \\alpha_ {i} \\left\\| Y _ {\\alpha_ {i}} ^ {t} - Y \\right\\| _ {1}, \\tag {4}", + "image_path": "27fd3344e1d22c7fcf3bf3a6ee5bac146141d51963190a5f3f770f128297744f.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "spans": [ + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "type": "text", + "content": " denotes the regularization parameter, which is empirically set as 0.1 in our experiments. During inference, the SR model " + }, + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "type": "inline_equation", + "content": "\\theta_{p}" + }, + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "type": "text", + "content": " can be customized for various resources according to different sparsity " + }, + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 296, + 563, + 530, + 621 + ], + "type": "text", + "content": ". After the network training, a lottery mask pool can be built for the backward stage." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 297, + 629, + 500, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 629, + 500, + 643 + ], + "spans": [ + { + "bbox": [ + 297, + 629, + 500, + 643 + ], + "type": "text", + "content": "3.3. The Backward Stage: Model Expansion" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "type": "text", + "content": "Our intention is to obtain a scalable network that can deploy according to customized memory resources. In this subsection, we introduce how to gradually recover a lightweight SR model, which is a backtracking process of IMP. We start from the most lightweight SR model and end with the original SR model. Corresponding to gradually shrinking the SR model with LTH, we solve two issues in" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "14402" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 101, + 272, + 204 + ], + "blocks": [ + { + "bbox": [ + 54, + 101, + 272, + 204 + ], + "lines": [ + { + "bbox": [ + 54, + 101, + 272, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 101, + 272, + 204 + ], + "type": "image", + "image_path": "76a605c86e31ff150ac11bb41a27637194670e3b2fd7677e655658c7b2156c43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 208, + 224, + 219 + ], + "lines": [ + { + "bbox": [ + 108, + 208, + 224, + 219 + ], + "spans": [ + { + "bbox": [ + 108, + 208, + 224, + 219 + ], + "type": "text", + "content": "Network Sparsity (\\% pruned)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 220, + 279, + 274 + ], + "lines": [ + { + "bbox": [ + 45, + 220, + 279, + 274 + ], + "spans": [ + { + "bbox": [ + 45, + 220, + 279, + 274 + ], + "type": "text", + "content": "Figure 5. Comparisons of IMP and OMP on various famous SR methods, which are evaluated on Set14 for " + }, + { + "bbox": [ + 45, + 220, + 279, + 274 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 45, + 220, + 279, + 274 + ], + "type": "text", + "content": " SR. The performance of the baseline dense network, the sparse sub-network obtained by IMP, as well as the sub-network obtained by OMP. The sub-network is with only " + }, + { + "bbox": [ + 45, + 220, + 279, + 274 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 45, + 220, + 279, + 274 + ], + "type": "text", + "content": " of the weights." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 287, + 279, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 287, + 279, + 335 + ], + "spans": [ + { + "bbox": [ + 45, + 287, + 279, + 335 + ], + "type": "text", + "content": "the model expansion: iterative model expansion (IME) and the expansion rule, the former refers to the requirement of the IME process and the latter refers to the requirement of the learning rule." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "spans": [ + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": "Let's focus on an expansion from the " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": "-th lightweight model to the " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": "-th lightweight model where " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": " comes from the definition in the forward stage. The larger " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": " is, the smaller model is. During IMP, we obtain two pruned-out masks corresponding to the " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": "-th lightweight models, denoted by " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{LM}_t" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{LM}_{t - 1}" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": ", respectively. Their difference represents the expanded positions, i.e., " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "m = LM_{t - 1} - LM_t" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": ". Without changing the " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": "-th model, we recover the weights according to the mask " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": ". After that, we fine-tune the recovered model and obtain the " + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 45, + 335, + 279, + 475 + ], + "type": "text", + "content": "-th model. So do the next model expansion. The detailed procedure is described in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 486, + 124, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 486, + 124, + 500 + ], + "spans": [ + { + "bbox": [ + 45, + 486, + 124, + 500 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 506, + 107, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 506, + 107, + 517 + ], + "spans": [ + { + "bbox": [ + 45, + 506, + 107, + 517 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 523, + 279, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 523, + 279, + 618 + ], + "spans": [ + { + "bbox": [ + 45, + 523, + 279, + 618 + ], + "type": "text", + "content": "DIV2K [2] is adopted to train the SR model, which is a high-quality natural scene dataset widely used for image SR tasks. The whole dataset includes 800 training images and 100 validation images totally with abundant texture and contents. The bicubic degraded LR images are synthesized following the previous works [17, 37]. Besides, the trained models are evaluated on four public SR benchmark datasets: Set5 [4], Set14 [35], B100 [3], and Urban100 [16]." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 47, + 665, + 279, + 720 + ], + "blocks": [ + { + "bbox": [ + 45, + 628, + 279, + 661 + ], + "lines": [ + { + "bbox": [ + 45, + 628, + 279, + 661 + ], + "spans": [ + { + "bbox": [ + 45, + 628, + 279, + 661 + ], + "type": "text", + "content": "Table 1. Ablation studies about the iterative magnitude pruning (IMP) and the stochastic self-distillation learning (SSD). All these models are validated on Set14 for " + }, + { + "bbox": [ + 45, + 628, + 279, + 661 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 45, + 628, + 279, + 661 + ], + "type": "text", + "content": " image SR." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 665, + 279, + 720 + ], + "lines": [ + { + "bbox": [ + 47, + 665, + 279, + 720 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 279, + 720 + ], + "type": "table", + "html": "
EDSR基礎IMP (LTH)IMP RewindingSSDSparsityPSNR
✓×××0%28.55
✓✓××94%27.66
✓×✓×94%28.12
✓×✓✓94%28.33
", + "image_path": "b7e0307b616471c5b885e7fff9b128d110f9da70aa1f83e31206c6501585966b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 298, + 131, + 530, + 259 + ], + "blocks": [ + { + "bbox": [ + 296, + 104, + 530, + 127 + ], + "lines": [ + { + "bbox": [ + 296, + 104, + 530, + 127 + ], + "spans": [ + { + "bbox": [ + 296, + 104, + 530, + 127 + ], + "type": "text", + "content": "Table 2. The performance of the sparse sub-network pruned by IMP-SSD on Set14 for " + }, + { + "bbox": [ + 296, + 104, + 530, + 127 + ], + "type": "inline_equation", + "content": "2 \\times, 3 \\times, 4 \\times" + }, + { + "bbox": [ + 296, + 104, + 530, + 127 + ], + "type": "text", + "content": " image SR." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 131, + 530, + 259 + ], + "lines": [ + { + "bbox": [ + 298, + 131, + 530, + 259 + ], + "spans": [ + { + "bbox": [ + 298, + 131, + 530, + 259 + ], + "type": "table", + "html": "
ScaleModelSparsityPSNR (dB)Params
2×EDSR基礎0%33.631370K
EDSR_IMP94%32.8885K
EDSR_IMP-SSD94%33.09 (+0.21)85K
3×EDSR基礎0%30.301558K
EDSR_IMP94%29.7397K
EDSR_IMP-SSD94%29.92 (+0.19)97K
4×EDSR基礎0%28.551518K
EDSR_IMP94%28.1295K
EDSR_IMP-SSD94%28.33 (+0.21)95K
", + "image_path": "5d748414ff03a7ca9e7fcb2104ea3d23b11853521630b1920ce29b0298842639.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 274, + 489, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 274, + 489, + 286 + ], + "spans": [ + { + "bbox": [ + 296, + 274, + 489, + 286 + ], + "type": "text", + "content": "4.2. Implementation and Training Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "spans": [ + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "text", + "content": "For training the proposed SR model, " + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "text", + "content": " image patches with RGB channels are cropped as the model input. The data augmentation is implemented by random flipping and rotation. The Adam optimizer with " + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "text", + "content": " is utilized to train the models. Besides, the mini-batch size is set to 16. The learning rate is initialized as " + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "inline_equation", + "content": "2e - 4" + }, + { + "bbox": [ + 296, + 287, + 530, + 403 + ], + "type": "text", + "content": ". The pruned weights can be determined by IMPSSD each 300 epochs and the same setting is adopted in the model expansion stage. We use PyTorch to implement our models with a GTX TITAN GPU." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 404, + 530, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 404, + 530, + 452 + ], + "spans": [ + { + "bbox": [ + 296, + 404, + 530, + 452 + ], + "type": "text", + "content": "Objective criteria, i.e., peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are adopted to evaluate the model performance. The two metrics are both calculated on the Y channel of the YCbCr space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 459, + 389, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 459, + 389, + 472 + ], + "spans": [ + { + "bbox": [ + 296, + 459, + 389, + 472 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 473, + 530, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 473, + 530, + 519 + ], + "spans": [ + { + "bbox": [ + 296, + 473, + 530, + 519 + ], + "type": "text", + "content": "Next, we mainly give a detailed analysis of the proposed lottery-ticket stochastic self-distillation scalable framework (IMP-SSD) for image SR. We adopt EDSR [23] as the backbone network to implement our experiments." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 520, + 530, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 520, + 530, + 682 + ], + "spans": [ + { + "bbox": [ + 296, + 520, + 530, + 682 + ], + "type": "text", + "content": "IMP vs. OMP. To verify the superiority of IMP over OMP on the compression task of image SR, we conduct IMP and OMP on the forward stage of our method instantiated with three popular SR methods: EDSR [23], RDN [38], and RCAN [37]. As shown in Fig. 5, when the pruned network reaches a very small model size, usually as the sparsity of the model is higher than " + }, + { + "bbox": [ + 296, + 520, + 530, + 682 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 296, + 520, + 530, + 682 + ], + "type": "text", + "content": ", where the sparsity means the ratio of the pruning-out parameters, IMP does not significantly scarify the SR performance while OMP degrades a lot. Especially, it obtains 27.66dB in PSNR with " + }, + { + "bbox": [ + 296, + 520, + 530, + 682 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 296, + 520, + 530, + 682 + ], + "type": "text", + "content": " sparsity as shown in Table 1. OMP is obviously unable to make the RCAN model reach a very small model similar to that by IMP. That's why we pick IMP as our pruning algorithm, which demonstrates the advantages of LTH once again." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": "Effectiveness of IMP-SSD. To demonstrate the effectiveness of IMP-SSD, we employ EDSR as the backbone and compare MSSR with EDSR. As shown in Table 2, our IMP-SSD method can not only compress a model to a very" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "14403" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 120, + 530, + 264 + ], + "blocks": [ + { + "bbox": [ + 89, + 104, + 485, + 116 + ], + "lines": [ + { + "bbox": [ + 89, + 104, + 485, + 116 + ], + "spans": [ + { + "bbox": [ + 89, + 104, + 485, + 116 + ], + "type": "text", + "content": "Table 3. Comparison of the quantitative results of average PSNR/SSIM with state-of-the-art efficient SR models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 120, + 530, + 264 + ], + "lines": [ + { + "bbox": [ + 48, + 120, + 530, + 264 + ], + "spans": [ + { + "bbox": [ + 48, + 120, + 530, + 264 + ], + "type": "table", + "html": "
ScaleMethodParamsSet5 PSNR/SSIMSet14 PSNR/SSIMB100 PSNR/SSIMUrban100 PSNR/SSIMDIV2K_valid PSNR/SSIM
4×EDSR (CVPRW' 2017)1518K32.09 / 0.893728.55 / 0.780627.54 / 0.734825.97 / 0.781930.38 / 0.8364
EDSR_IMP-SSD_88% (ours)190K31.89 / 0.891128.44 / 0.777727.46 / 0.732625.71 / 0.774330.25 / 0.8336
EDSR_IMP-SSD_94% (ours)95K31.74 / 0.887928.33 / 0.774927.39 / 0.729725.48 / 0.765730.12 / 0.8303
RCAN (ECCV' 2018)15592K32.19 / 0.895028.63 / 0.781927.57 / 0.735526.21 / 0.788830.40 / 0.8331
RCAN_IMP-SSD_88% (ours)1949K32.30 / 0.896328.72 / 0.784427.63 / 0.737926.32 / 0.793230.53 / 0.8396
RCAN_IMP-SSD_94% (ours)974K32.02 / 0.892428.49 / 0.778827.51 / 0.733425.87 / 0.780030.30 / 0.8347
RDN (CVPR' 2018)5579K32.24 / 0.896528.71 / 0.784127.64 / 0.738426.31 / 0.792830.55 / 0.8406
RDN_IMP-SSD_88% (ours)697K31.87 / 0.890228.38 / 0.776627.42 / 0.731325.53 / 0.768130.15 / 0.8315
RDN_IMP-SSD_97% (ours)174K31.36 / 0.882628.10 / 0.769527.23 / 0.724925.09 / 0.311629.86 / 0.8243
ENLCN (AAAI' 2022)1536K32.14 / 0.893928.59 / 0.781227.56 / 0.735226.05 / 0.784530.36 / 0.8360
ENLCN_IMP-SSD_88% (ours)192K31.97 / 0.891128.47 / 0.773127.47 / 0.732525.77 / 0.731330.22 / 0.8326
ENLCN_IMP-SSD_97% (ours)48K31.41 / 0.882528.09 / 0.769027.24 / 0.725025.15 / 0.313329.84 / 0.8239
", + "image_path": "35d0432593e88a5a019303e7df8500fa84f4f257b965d87108a297177e390f9f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 314, + 277, + 378 + ], + "blocks": [ + { + "bbox": [ + 45, + 281, + 279, + 313 + ], + "lines": [ + { + "bbox": [ + 45, + 281, + 279, + 313 + ], + "spans": [ + { + "bbox": [ + 45, + 281, + 279, + 313 + ], + "type": "text", + "content": "Table 4. Comparison between the performance (PSNR) of the sparse sub-networks via IMP-SSD and the performance of the sparse sub-networks of the scalable network (SCL) for " + }, + { + "bbox": [ + 45, + 281, + 279, + 313 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 45, + 281, + 279, + 313 + ], + "type": "text", + "content": " SR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 314, + 277, + 378 + ], + "lines": [ + { + "bbox": [ + 48, + 314, + 277, + 378 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 277, + 378 + ], + "type": "table", + "html": "
Prune (%)Set5Set14B100Urban100
IMP-SSDSCLIMP-SSDSCLIMP-SSDSCLIMP-SSDSCL
0 %32.0931.9428.5528.5327.5427.5325.9725.89
50 %32.0531.9628.5728.5127.5527.5025.9925.79
75 %32.0331.9228.5328.4627.5227.4825.9025.73
88 %31.8931.8228.4428.4027.4627.4325.7125.61
94 %31.7431.7428.3328.3327.3927.3925.4825.48
", + "image_path": "d5b194c0404c14e3f3a75261f68acf8bf1919e067c5c0fca31bc0d842eedeaa0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "spans": [ + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "type": "text", + "content": "small size but also mitigate the degradation in performance for different magnification factors " + }, + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "type": "inline_equation", + "content": "(2\\times, 3\\times," + }, + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "type": "inline_equation", + "content": "4\\times)" + }, + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "type": "text", + "content": ". Besides, the results in Table 1 also illustrate that IMP-SSD further improves the PSNR to " + }, + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "type": "inline_equation", + "content": "28.33\\mathrm{dB}" + }, + { + "bbox": [ + 45, + 388, + 279, + 458 + ], + "type": "text", + "content": " with the similar sparsity. Therefore, our method has significant effects on scalable model compression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "spans": [ + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "text", + "content": "IMP-SSD SR vs. Scalable SR. In order to show the effectiveness of MSSR, we compare the scalable lightweight SR model obtained in the backward stage with the middle result by IMP-SSD in the forward stage. We use EDSR as the backbone, and make the comparison in five degrees of sparsities: " + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "text", + "content": ". In Table 4, IMP-SSD generates five pruned SR models corresponding to five sparsities, while scalable SR only has one model. With the increase of sparsity, the SR performance decreases, but the gap between two neighbor sparsities is not more than " + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{dB}" + }, + { + "bbox": [ + 45, + 458, + 279, + 656 + ], + "type": "text", + "content": ". Thus, the SR scalable model makes sense. As shown in Table 4, it is observed that scalable SR achieves comparable results to IMP-SSD on four datasets: Set5, Set14, B100, and Urban100. It shows that our method uses only one scalable model that can reach the level of the separately lightweight SR models at different sparse degrees. Therefore, MSSR is memory-friendly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 664, + 236, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 664, + 236, + 676 + ], + "spans": [ + { + "bbox": [ + 45, + 664, + 236, + 676 + ], + "type": "text", + "content": "4.4. Instantiating with Other SR methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the proposed scalable lightweight framework (MSSR), we integrate it with four representative SR methods: EDSR [23], RCAN [37], RDN [38], and ENLCN [1]. All these SR models are re" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 299, + 306, + 530, + 370 + ], + "blocks": [ + { + "bbox": [ + 296, + 281, + 530, + 303 + ], + "lines": [ + { + "bbox": [ + 296, + 281, + 530, + 303 + ], + "spans": [ + { + "bbox": [ + 296, + 281, + 530, + 303 + ], + "type": "text", + "content": "Table 5. The analysis about IMP-SSD on different methods for " + }, + { + "bbox": [ + 296, + 281, + 530, + 303 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 296, + 281, + 530, + 303 + ], + "type": "text", + "content": " image SR validated on Set14." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 299, + 306, + 530, + 370 + ], + "lines": [ + { + "bbox": [ + 299, + 306, + 530, + 370 + ], + "spans": [ + { + "bbox": [ + 299, + 306, + 530, + 370 + ], + "type": "table", + "html": "
PSNR Sparsity\nModel0%50%75%88%94%97%99%
EDSR_IMP-SSD28.5528.5728.5328.4428.3328.0520.36
RDN_IMP-SSD28.7028.6828.5228.3828.2328.1027.46
RCAN_IMP-SSD28.6328.7728.7328.7228.4928.1527.68
ENLCN_IMP-SSD28.5928.6228.5628.4728.3028.0927.60
", + "image_path": "19991bbb3be9b57595fd42f306a6ea0dfc245cb109d12da398bc5982b1132853.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "spans": [ + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "type": "text", + "content": "trained under the same experimental settings. We show the results in Table 3. It shows that our method is universally applicable to the existing SR methods. Compared with the original models with the magnification factor " + }, + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "type": "text", + "content": ", under the sparsities of " + }, + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "type": "inline_equation", + "content": "88 \\%" + }, + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "type": "inline_equation", + "content": "94 \\% / 97 \\%" + }, + { + "bbox": [ + 296, + 386, + 530, + 501 + ], + "type": "text", + "content": ", the maximum decreases of IMP-SSD instantiated with EDSR, RCAN, RDN, and ENLCN are (0.2dB, -0.1dB, 0.8dB, 0.3dB), and (0.5dB, 0.4dB, 0.9dB, 0.9dB), respectively. Note that, in RCAN, IMP-SSD is better than its original model. We guess that SSD could improve the SR results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "spans": [ + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "type": "text", + "content": "In Table 5, we give more comparisons of the SR performance with more different sparsity. It shows that all methods except EDSR, with the decrease of the model size, our method degrades gradually with not much range of decrease, their ranges are " + }, + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "type": "inline_equation", + "content": "1.3\\mathrm{dB}" + }, + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "type": "text", + "content": ", 1dB, 0.9dB in RCAN, RDN, ENLCN, respectively. EDSR degrades significantly by " + }, + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "type": "inline_equation", + "content": "8.2\\mathrm{dB}" + }, + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "type": "text", + "content": " with the sparsity of " + }, + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 296, + 503, + 530, + 585 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 594, + 472, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 594, + 472, + 606 + ], + "spans": [ + { + "bbox": [ + 296, + 594, + 472, + 606 + ], + "type": "text", + "content": "4.5. Comparison with State-of-the-art" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "spans": [ + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "type": "text", + "content": "To validate the effectiveness of MSSR, we compare our approach (IMP-SSD) with several state-of-the-art lightweight SR methods: SRCNN [7], FSRCNN [8], CARN [19], and ESRN [31]. Note that IMP-SSD obtains the gradually shrinking lightweight SR models. The comparison results for " + }, + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 296, + 612, + 530, + 705 + ], + "type": "text", + "content": " magnification factors are shown in Table 6. PSNR and SSIM are adopted as metrics to evaluate image quality by convention." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "type": "text", + "content": "It shows that our IMP-SSD achieves the best results in terms of the coordination of parameters and performance." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "14404" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 117, + 529, + 348 + ], + "blocks": [ + { + "bbox": [ + 89, + 104, + 485, + 115 + ], + "lines": [ + { + "bbox": [ + 89, + 104, + 485, + 115 + ], + "spans": [ + { + "bbox": [ + 89, + 104, + 485, + 115 + ], + "type": "text", + "content": "Table 6. Comparison of the quantitative results of average PSNR/SSIM with state-of-the-art efficient SR models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 117, + 529, + 348 + ], + "lines": [ + { + "bbox": [ + 48, + 117, + 529, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 529, + 348 + ], + "type": "table", + "html": "
ScaleMethodParamsSet5 PSNR/SSIMSet14 PSNR/SSIMB100 PSNR/SSIMUrban100 PSNR/SSIM
2×SRCNN [7] (ECCV' 2014)57K36.66 / 0.954232.42 / 0.906331.36 / 0.887929.50 / 0.8946
FSRCNN [8] (ECCV' 2016)12K37.00 / 0.955832.63 / 0.908831.53 / 0.892029.88 / 0.9020
CARN [19] (ECCV' 2018)415K37.68 / 0.959433.27 / 0.914931.97 / 0.896931.18 / 0.9186
ESRN [31] (AAAI' 2020)324K37.85 / 0.960033.42 / 0.916132.10 / 0.898731.79 / 0.9248
EDSR [23] (CVPRW' 2017)1370K37.98 / 0.960533.63 / 0.918032.17 / 0.899632.08 / 0.9280
EDSR_IMP-SSD.88% (ours)171K37.82 / 0.959933.44 / 0.916732.09 / 0.898731.69 / 0.9244
EDSR_IMP-SSD.94% (ours)85K37.51 / 0.958633.09 / 0.913731.88 / 0.896030.94 / 0.9163
3×SRCNN [7] (ECCV' 2014)57K32.75 / 0.909029.28 / 0.820928.41 / 0.786326.24 / 0.7989
FSRCNN [8] (ECCV' 2016)12K33.16 / 0.914029.43 / 0.824228.53 / 0.791026.43 / 0.8080
CARN [19] (ECCV' 2018)415K34.00 / 0.924030.07 / 0.836828.92 / 0.800727.53 / 0.8379
ESRN [31] (AAAI' 2020)324K34.23 / 0.926230.27 / 0.840029.03 / 0.803927.95 / 0.8481
EDSR [23] (CVPRW' 2017)1558K34.37 / 0.926730.31 / 0.840929.07 / 0.804328.05 / 0.8506
EDSR_IMP-SSD.88% (ours)195K34.15 / 0.925230.20 / 0.839229.00 / 0.802827.81 / 0.8449
EDSR_IMP-SSD.94% (ours)97K33.74 / 0.921629.92 / 0.834028.81 / 0.797627.16 / 0.8298
4×SRCNN [7] (ECCV' 2014)57K30.48 / 0.862827.49 / 0.750326.90 / 0.710124.52 / 0.7221
FSRCNN [8] (ECCV' 2016)12K30.71 / 0.865727.59 / 0.753526.98 / 0.715024.62 / 0.7280
CARN [19] (ECCV' 2018)415K31.78 / 0.889428.39 / 0.776227.42 / 0.730325.56 / 0.7674
ESRN [31] (AAAI' 2020)324K31.99 / 0.891928.49 / 0.777927.50 / 0.733125.87 / 0.7782
EDSR [23] (CVPRW' 2017)1518K32.09 / 0.893728.55 / 0.780627.54 / 0.734825.97 / 0.7819
EDSR_IMP-SSD.88% (ours)190K31.89 / 0.891128.44 / 0.777727.46 / 0.732625.71 / 0.7743
EDSR_IMP-SSD.94% (ours)95K31.74 / 0.887928.33 / 0.774927.39 / 0.729725.48 / 0.7657
", + "image_path": "ea85864f82dfc7c3e9deafe0b6e5d0f89a70d1d10c3652a0dc0baba3069a3093.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 54, + 366, + 280, + 508 + ], + "blocks": [ + { + "bbox": [ + 54, + 366, + 280, + 508 + ], + "lines": [ + { + "bbox": [ + 54, + 366, + 280, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 280, + 508 + ], + "type": "image", + "image_path": "d919ee3f6b8728f97d6403df8378b5a0642b399578e0f637a35a36ac50bfda7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 510, + 279, + 532 + ], + "lines": [ + { + "bbox": [ + 45, + 510, + 279, + 532 + ], + "spans": [ + { + "bbox": [ + 45, + 510, + 279, + 532 + ], + "type": "text", + "content": "Figure 6. Visual comparisons of IMP-SSD with other lightweight SR models on Urban100 for " + }, + { + "bbox": [ + 45, + 510, + 279, + 532 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 45, + 510, + 279, + 532 + ], + "type": "text", + "content": " SR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": "In particular, compared with other lightweight SR models such as CARN and ESRN, our IMP-SSD in EDSR with the sparsity of " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": " achieves comparable performance while our model size is about half or two-thirds of those. Compared with smaller SR models such as SRCNN and FSRCNN, IMP-SSD with the sparsity of " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": ", IMP-SSD achieves significant gains in PSNR by at least 0.4dB, 0.3dB, 0.4dB for " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": " SR, respectively. The same conclusion is made in SSIM. Furthermore, compared with the original EDSR, IMP-SSD makes the SR model shrink to reach a very small size but their SR performance does not degrade very much. Compared with the original EDSR, the maximum decreases are 0.4dB, 0.2dB, 0.2dB for IMP-SSD with the sparsity of " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": " and 0.6dB, 0.9dB, 0.5dB for IMP-SSD with the sparsity of " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": ", respectively. It demonstrates that our method is effective on scalable SR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 367, + 530, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 367, + 530, + 415 + ], + "spans": [ + { + "bbox": [ + 296, + 367, + 530, + 415 + ], + "type": "text", + "content": "We also give the comparison of our method in visual effect in Fig. 6. It is observed that our method achieves a better visual effect than other lightweight SR models even with larger model sparsity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 428, + 373, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 428, + 373, + 440 + ], + "spans": [ + { + "bbox": [ + 297, + 428, + 373, + 440 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 449, + 530, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 449, + 530, + 669 + ], + "spans": [ + { + "bbox": [ + 296, + 449, + 530, + 669 + ], + "type": "text", + "content": "In this paper, we propose a novel memory-friendly scalable lightweight framework (MSSR) with rewinding LTH, which can be deployed on arbitrary computational resources according to customization. MSSR is backtracking and includes two stages: the forward stage for model compression and the backward for model expansion. For the first time, LTH with rewinding weights is implemented to explore the winning tickets of deep SR models. We use IMP to gradually shrink the SR model and not only obtain the scalable lightweight model but also obtain the pruning-out masks which form the nested sets. In order to improve the performance of the shrinking model, SSD is conducted. Unlike the traditional SD, we replace the feature maps on several layers of Teacher with the counterparts of Students and conduct distillation. In the backward stage, we gradually recover the lightweight SR models depending on the pruning-out masks and then fine-tune the current model. We make instantiation on several state-of-the-art SR models. The extensive experimental results are effective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 671, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 671, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 296, + 671, + 530, + 729 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by grants from the National Key Research and Development Program of China (No.2020AAA0108301), National Natural Science Foundation of China (No.62176224, No.62222602), and CCF-Lenovo Blue Ocean Research Fund." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "14405" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 51, + 124, + 279, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 279, + 145 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 279, + 145 + ], + "type": "text", + "content": "[1] Efficient non-local contrastive attention for image superresolution. In AAAI, 2022. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 147, + 279, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 147, + 279, + 177 + ], + "spans": [ + { + "bbox": [ + 51, + 147, + 279, + 177 + ], + "type": "text", + "content": "[2] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 181, + 279, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 181, + 279, + 211 + ], + "spans": [ + { + "bbox": [ + 51, + 181, + 279, + 211 + ], + "type": "text", + "content": "[3] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 2010. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 213, + 279, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 213, + 279, + 255 + ], + "spans": [ + { + "bbox": [ + 51, + 213, + 279, + 255 + ], + "type": "text", + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 256, + 279, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 256, + 279, + 298 + ], + "spans": [ + { + "bbox": [ + 51, + 256, + 279, + 298 + ], + "type": "text", + "content": "[5] Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Zhangyang Wang, and Michael Carbin. The lottery ticket hypothesis for pre-trained bert networks. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 300, + 279, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 300, + 279, + 332 + ], + "spans": [ + { + "bbox": [ + 51, + 300, + 279, + 332 + ], + "type": "text", + "content": "[6] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 334, + 279, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 334, + 279, + 364 + ], + "spans": [ + { + "bbox": [ + 51, + 334, + 279, + 364 + ], + "type": "text", + "content": "[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 367, + 279, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 367, + 279, + 397 + ], + "spans": [ + { + "bbox": [ + 51, + 367, + 279, + 397 + ], + "type": "text", + "content": "[8] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 399, + 279, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 399, + 279, + 431 + ], + "spans": [ + { + "bbox": [ + 51, + 399, + 279, + 431 + ], + "type": "text", + "content": "[9] Utku Evci, Fabian Pedregosa, Aidan Gomez, and Erich Elsen. The difficulty of training sparse neural networks. In ICLR, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 433, + 279, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 279, + 463 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 279, + 463 + ], + "type": "text", + "content": "[10] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 466, + 279, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 279, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 279, + 487 + ], + "type": "text", + "content": "[11] Jonathan Frankle, David J Schwab, and Ari S Morcos. The early phase of neural network training. In ICLR, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 489, + 279, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 279, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 279, + 520 + ], + "type": "text", + "content": "[12] Sharath Girish, Shishira R Maiya, Kamal Gupta, Hao Chen, Larry S Davis, and Abhinav Shrivastava. The lottery ticket hypothesis for object recognition. In CVPR, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 522, + 279, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 279, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 279, + 563 + ], + "type": "text", + "content": "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 564, + 279, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 279, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 279, + 596 + ], + "type": "text", + "content": "[14] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. TPAMI, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 598, + 279, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 279, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 279, + 628 + ], + "type": "text", + "content": "[15] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In ICCV, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 631, + 279, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 279, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 279, + 662 + ], + "type": "text", + "content": "[16] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 664, + 279, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 279, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 279, + 695 + ], + "type": "text", + "content": "[17] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM MM, 2019. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 697, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 697, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 697, + 279, + 729 + ], + "type": "text", + "content": "[18] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 728 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "type": "text", + "content": "[19] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 1, 2, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 140, + 530, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 530, + 170 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 530, + 170 + ], + "type": "text", + "content": "[20] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 172, + 530, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 172, + 530, + 191 + ], + "spans": [ + { + "bbox": [ + 299, + 172, + 530, + 191 + ], + "type": "text", + "content": "[21] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. In NeurIPS, 1989. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 193, + 530, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 193, + 530, + 223 + ], + "spans": [ + { + "bbox": [ + 299, + 193, + 530, + 223 + ], + "type": "text", + "content": "[22] Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Dynamic slimmable network. In CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 225, + 530, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 225, + 530, + 256 + ], + "spans": [ + { + "bbox": [ + 299, + 225, + 530, + 256 + ], + "type": "text", + "content": "[23] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 258, + 530, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 258, + 530, + 299 + ], + "spans": [ + { + "bbox": [ + 299, + 258, + 530, + 299 + ], + "type": "text", + "content": "[24] Ning Liu, Geng Yuan, Zhengping Che, Xuan Shen, Xiaolong Ma, Qing Jin, Jian Ren, Jian Tang, Sijia Liu, and Yanzhi Wang. Lottery ticket preserves weight correlation: Is it desirable or not? In ICML, 2021. 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 300, + 530, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 300, + 530, + 341 + ], + "spans": [ + { + "bbox": [ + 299, + 300, + 530, + 341 + ], + "type": "text", + "content": "[25] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 343, + 530, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 343, + 530, + 385 + ], + "spans": [ + { + "bbox": [ + 299, + 343, + 530, + 385 + ], + "type": "text", + "content": "[26] Xiaotong Luo, Mingliang Dai, Yulun Zhang, Yuan Xie, Ding Liu, Yanyun Qu, Yun Fu, and Junping Zhang. Adjustable memory-efficient image super-resolution via individual kernel sparsity. In ACM MM, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 386, + 530, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 386, + 530, + 417 + ], + "spans": [ + { + "bbox": [ + 299, + 386, + 530, + 417 + ], + "type": "text", + "content": "[27] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 419, + 530, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 419, + 530, + 448 + ], + "spans": [ + { + "bbox": [ + 299, + 419, + 530, + 448 + ], + "type": "text", + "content": "[28] Sai Prasanna, Anna Rogers, and Anna Rumshisky. When bert plays the lottery, all tickets are winning. In ACL, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 450, + 530, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 450, + 530, + 481 + ], + "spans": [ + { + "bbox": [ + 299, + 450, + 530, + 481 + ], + "type": "text", + "content": "[29] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. In ICLR, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 483, + 530, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 483, + 530, + 513 + ], + "spans": [ + { + "bbox": [ + 299, + 483, + 530, + 513 + ], + "type": "text", + "content": "[30] Pedro Savarese, Hugo Silva, and Michael Maire. Winning the lottery with continuous sparsification. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 515, + 530, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 515, + 530, + 546 + ], + "spans": [ + { + "bbox": [ + 299, + 515, + 530, + 546 + ], + "type": "text", + "content": "[31] Dehua Song, Chang Xu, Xu Jia, Yiyi Chen, Chunjing Xu, and Yunhe Wang. Efficient residual dense block search for image super-resolution. In AAAI, 2020. 7, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 299, + 548, + 530, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 548, + 530, + 578 + ], + "spans": [ + { + "bbox": [ + 299, + 548, + 530, + 578 + ], + "type": "text", + "content": "[32] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. In ICLR, 2020. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 299, + 580, + 530, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 580, + 530, + 621 + ], + "spans": [ + { + "bbox": [ + 299, + 580, + 530, + 621 + ], + "type": "text", + "content": "[33] Shihui Yin, Kyu-Hyoun Kim, Jinwook Oh, Naigang Wang, Mauricio Serrano, Jae-Sun Seo, and Jungwook Choi. The sooner the better: Investigating structure of early winning lottery tickets. In ICLR, 2020. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 299, + 623, + 530, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 623, + 530, + 653 + ], + "spans": [ + { + "bbox": [ + 299, + 623, + 530, + 653 + ], + "type": "text", + "content": "[34] Haonan Yu, Sergey Edunov, Yuandong Tian, and Ari S Morcos. Playing the lottery with rewards and multiple languages: lottery tickets in rl and nlp. In ICLR, 2020. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 299, + 655, + 530, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 655, + 530, + 685 + ], + "spans": [ + { + "bbox": [ + 299, + 655, + 530, + 685 + ], + "type": "text", + "content": "[35] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In ICCV, 2010. 6" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 299, + 687, + 530, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 687, + 530, + 728 + ], + "spans": [ + { + "bbox": [ + 299, + 687, + 530, + 728 + ], + "type": "text", + "content": "[36] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In ICCV, 2019. 4" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "14406" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 193 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[37] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 279, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 279, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 279, + 170 + ], + "type": "text", + "content": "[38] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 173, + 278, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 278, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 278, + 193 + ], + "type": "text", + "content": "[39] Hao Zhou, Jose M Alvarez, and Fatih Porikli. Less is more: Towards compact cnns. In ECCV, 2016. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "type": "text", + "content": "14407" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_content_list.json b/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dcc41315965b76cc51f1d9424bbc91e7c9154382 --- /dev/null +++ b/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_content_list.json @@ -0,0 +1,1829 @@ +[ + { + "type": "text", + "text": "Meta Architecture for Point Cloud Analysis", + "text_level": 1, + "bbox": [ + 263, + 130, + 707, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haojia Lin $^{1}$ , Xiawu Zheng $^{2}$ , Lijiang Li $^{1}$ , Fei Chao $^{1}$ , Shanshan Wang $^{2,3}$ , Yan Wang $^{4}$ , Yonghong Tian $^{2,5}$ , Rongrong Ji $^{1,2,6*}$", + "bbox": [ + 200, + 179, + 764, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, School of Informatics, Xiamen University, 361005, P.R. China.", + "bbox": [ + 107, + 215, + 859, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Peng Cheng Laboratory. $^{3}$ Chinese Academy of Sciences. $^{4}$ Samsara Inc.", + "bbox": [ + 197, + 251, + 769, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ National Engineering Research Center for Visual Technology, Peking University. $^{6}$ Shenzhen Research Institute of Xiamen University.", + "bbox": [ + 161, + 268, + 805, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{linhj, lilijiang}@stu.xmu.edu.cn, zhengxw01@pcl.ac.cn, {rrji, feichao}@xmu.edu.cn", + "bbox": [ + 122, + 306, + 843, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ss.wang@siat.ac.cn, yan.wang@samsara.com, yhtian@pku.edu.cn", + "bbox": [ + 220, + 325, + 743, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 375, + 313, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in 3D point cloud analysis bring a diverse set of network architectures to the field. However, the lack of a unified framework to interpret those networks makes any systematic comparison, contrast, or analysis challenging, and practically limits healthy development of the field. In this paper, we take the initiative to explore and propose a unified framework called PointMeta, to which the popular 3D point cloud analysis approaches could fit. This brings three benefits. First, it allows us to compare different approaches in a fair manner, and use quick experiments to verify any empirical observations or assumptions summarized from the comparison. Second, the big picture brought by PointMeta enables us to think across different components, and revisit common beliefs and key design decisions made by the popular approaches. Third, based on the learnings from the previous two analyses, by doing simple tweaks on the existing approaches, we are able to derive a basic building block, termed PointMetaBase. It shows very strong performance in efficiency and effectiveness through extensive experiments on challenging benchmarks, and thus verifies the necessity and benefits of high-level interpretation, contrast, and comparison like PointMeta. In particular, PointMetaBase surpasses the previous state-of-the-art method by $0.7\\% /1.4\\% 2.1\\%$ mIoU with only $2\\% /11\\% /13\\%$ of the computation cost on the S3DIS datasets. The code and models are available at https://github.com/linhaojia13/PointMetaBase.", + "bbox": [ + 76, + 415, + 472, + 824 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2770bbcbf68355d70490c1ac7cd4a1976b518ec991a73ee3e7a8829f01a3532b.jpg", + "image_caption": [ + "Figure 1. Segmentation performance of PointMetaBase on S3DIS [1]. PointMetaBase surpasses the state-of-the-art method PointNeXt [16] significantly with a large FLOPs reduction." + ], + "image_footnote": [], + "bbox": [ + 504, + 376, + 887, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 675, + 630, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past two decades, the popularity of 3D data acquisition technology has led to great interest in point cloud analysis. Unlike images, point clouds are inherently sparse, unordered, and irregular. These characteristics make it challenging for the powerful convolutional neural networks (CNN) [4, 18] to extract useful information from point clouds. Early studies attempt to convert the point cloud into grids by either voxelization or 2D projections, such that the standard CNN can be applied, but the applications are limited by the extra computation and information loss.", + "bbox": [ + 496, + 702, + 890, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the introduction of permutation equivariance and invariance by PointNet [13] and PointNet++ [14], it is possible to use CNNs to process point clouds in their unstruc", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: rrji@xmu.edu.cn", + "bbox": [ + 94, + 886, + 316, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17682", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tured format. Following PointNet++ [14], many point-based networks have been introduced, most of which focus on the development of complex building blocks to extract local features, such as the $\\mathcal{X}$ - Conv convolution in PointCNN [22] and the self-attention layer in Point Transformer [32]. Although brought significant performance gains, these models are very complicated. On one hand, this makes the approaches computational demanding and thus limited in applications. On the other hand, it also gives extra challenges in systematic comparison and analysis among the models, and thus impacts efficient development of the field, which ideally would benefit from theoretic and empirical guidance.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The community is well aware of this issue, and research efforts have been put on looking for a unified perspective to compare these methods and identify the most crucial implementation details. PosPool [10] and PointNeXt [16] take a step toward this goal. Adopting a deep residual architecture as the base network, PosPool [10] evaluates the representative building blocks [8, 14, 22, 26] and finds that they perform similarly well. PointNeXt [16] revisits the improved training and scaling strategies adopted by previous SotA methods [7, 15, 22, 32] and tweaks the early model PointNet++ [14] with the learnings, which achieves state-of-the-art performance. Both of them give insights and inspirations to the community via dedicated experiment designs and extensive empirical analysis. However, neither of them provides a framework that is general enough to fit a broad range of point cloud analysis approaches, such that large scale comparison and contrast could be done.", + "bbox": [ + 75, + 287, + 472, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we argue that if viewed from the perspective of basic building blocks, the majority of existing approaches could be fit into a single meta architecture (Sec. 3). We call it PointMeta. PointMeta abstracts the computation pipeline of the building blocks for point cloud analysis into four meta functions: a neighbor update function, a neighbor aggregation function, a point update function and a position embedding function (implicit or explicit). As will be discussed in more details in the next sections, this framework brings three benefits, which are also our core technical contributions:", + "bbox": [ + 75, + 545, + 472, + 709 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In the dimension of models, it allows us to compare and contrast different models in a fair manner. So it becomes practical to observe and summarize learnings and assumptions, whose correctness could be further verified through experiments with variables controlled under the same framework. For example, among the systematic analysis on all the components across popular models, we found for the positional embedding function, explicit positional embedding is empirically the best choice (Sec. 4.2).", + "- In the dimension of components, it allows us to have" + ], + "bbox": [ + 94, + 723, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a higher level view across components, and thus to revisit the common beliefs and design decisions of the existing approaches. For example, despite the common perception, we find that the neighbor aggregation blocks may collaborate or compete with the learned neighbor features (Sec. 4.3), and thus should be designed carefully.", + "bbox": [ + 529, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Based on the learnings from the previous two dimensions, we are then able to do simple tweaks on the building blocks to apply the best practices (Sec. 4.5). The result building block, PointMetaBase, achieves very strong performance, surpassing the previous state-of-the-art method [16] by $0.7\\% / 1.4\\% / 2.1\\%$ mIoU with only $2\\% / 11\\% / 13\\%$ of the computation cost on S3DIS [1] (Fig. 1), and can act as a baseline for further research.", + "bbox": [ + 517, + 208, + 893, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 359, + 640, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Networks for Point Cloud Analysis. The unstructured data format of point cloud makes it difficult to apply CNN directly. To address this challenge, PointNet [13] introduces permutation equivariance and invariance via point-wise MLP and max pooling, which makes it possible to process point cloud directly. To better capture locality, PointNet++ [14] presents a Set Abstraction module to aggregate features from the neighbor points. Most works after PointNet++ focus on the design of build blocks. Graph-based methods [7, 25, 27] utilize graph neural networks to extract local features. Pseudo-grid-based methods [8,21,22] transform the local features onto pseudo grids which allow for regular convolutions. Adaptive weight-based methods [5,9] aggregate the neighbor features via weights adaptively learned from the locality. Recently, Transformer-based block [6, 32] are introduced to capture local information through self-attention. These proposed building blocks are so complicated that make challenges in systematic comparison and analysis, which suggests an urgent need of a common framework to unify existing models in this field.", + "bbox": [ + 496, + 385, + 893, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. PointMeta", + "text_level": 1, + "bbox": [ + 500, + 702, + 612, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Meta Architecture", + "text_level": 1, + "bbox": [ + 500, + 727, + 679, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the graph network framework [2], we reformulate existing building blocks for point cloud analysis into a single meta architecture, named PointMeta. As shown in Fig. 2 (a), a PointMeta block is composed of four meta functions: a neighbor update function $\\phi^n$ , a neighbor aggregation function $\\phi^a$ , a point update function $\\phi^p$ and a position embedding function $\\phi^e$ (implicit or explicit). The forward pass of a PointMeta block is as follows:", + "bbox": [ + 496, + 751, + 890, + 872 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\mathcal {N} (i)} = \\phi^ {n} \\circ \\phi^ {e} \\left(f _ {i}, p _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 885, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/41d674684abf894d1a8083199d7c3e93452c04390476ce8670060651356bd6d1.jpg", + "image_caption": [ + "(a) PointMeta" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 207, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/963b7cc4839de276bd2d8d73ed9e1838cbff50ba68b1e4b21dd87b497897bf39.jpg", + "image_caption": [ + "(b) PointNet++" + ], + "image_footnote": [], + "bbox": [ + 212, + 90, + 295, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/88b61868aae50f225cc781fa976ffb589a02a22bf569464011831c6616a699eb.jpg", + "image_caption": [ + "(c) DGCNN" + ], + "image_footnote": [], + "bbox": [ + 303, + 90, + 393, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6ebcad888433fced510b30228249046d70ff071a2ce60cdb5b18b29f5c5cd56.jpg", + "image_caption": [ + "(d) PointCNN", + "Figure 2. PointMeta and its instantiation examples. (a) In PointMeta, the position embedding function is usually combined with the neighbor update function or the aggregation function implicitly or explicitly. (b) $\\sim$ (f) Representative building blocks can naturally fit into PointMeta. (g) Applying the summarized best practices, we do simple tweaks on the building blocks and propose PointMetaBase." + ], + "image_footnote": [], + "bbox": [ + 405, + 90, + 516, + 256 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/902577f248e47bcfc2d08ae786f41f008e2e50dfb0cd7bb4c05ca083eab86704.jpg", + "image_caption": [ + "(e) Point Transformer" + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 653, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2e85920323a7b4a9fac7d04d08f53de16d208aa099da8c7a952a389be601048a.jpg", + "image_caption": [ + "(f) ASSANet" + ], + "image_footnote": [], + "bbox": [ + 655, + 90, + 763, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/46995296ac8f64a4b113e7208b8f058ce2d77058ec41277de09fb2158561b9d2.jpg", + "image_caption": [ + "(g) PointMetaBase" + ], + "image_footnote": [], + "bbox": [ + 772, + 90, + 893, + 256 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {(1)} = \\phi^ {a} \\circ \\phi^ {e} \\left(f _ {i}, p _ {i}, \\mathbf {f} _ {\\mathcal {N} (i)}, \\mathbf {p} _ {\\mathcal {N} (i)}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 354, + 468, + 376 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {(2)} = \\phi^ {p} \\left(f _ {i} ^ {(1)}, p _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 396, + 468, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{i} \\in \\mathbb{R}^{d}$ and $p_i \\in \\mathbb{R}^3$ represent the feature and the coordinate of point $i$ respectively, and $d$ denotes the feature length. Let $\\mathcal{N}(i)$ denote the neighbors of point $i$ , $f_{j}$ and $p_j$ be the feature and coordinate of each neighbor of point $i$ respectively, $\\pmb{f}_{\\mathcal{N}(i)} = \\{f_j | j \\in \\mathcal{N}(i)\\}$ represent all neighbor features of point $i$ , $\\pmb{p}_{\\mathcal{N}(i)} = \\{p_j | j \\in \\mathcal{N}(i)\\}$ represent all neighbor coordinates of point $i$ , and $k$ denote the neighbor number.", + "bbox": [ + 75, + 425, + 468, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neighbor update. The first step is to group and update neighbors for each point. Usually, the neighbor features $f_{\\mathcal{N}(i)}$ is produced by K-Nearest query as [27, 32] or ball query as [14-16]. The update on neighbor features is usually conducted by a multi-layer perceptions (MLP) to make sure the permutation equivariance of the neighbor update function $\\phi^n$ . Specifically, a permutation transformation on $f_{\\mathcal{N}(i)}$ results in the same permutation change on the output of the neighbor update.", + "bbox": [ + 75, + 547, + 468, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neighbor aggregation. Literally, the neighbor aggregation function is used to aggregate the neighbor features for a center point. Different from $\\phi^n$ , the aggregation function $\\phi^a$ must satisfy permutation invariance, i.e., a permutation transformation on the input $f_{\\mathcal{N}(i)}$ will have no influence on the output $f_i^{(1)}$ .", + "bbox": [ + 75, + 684, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point update. The point update usually contains an MLP, which is used to further enhance the point features.", + "bbox": [ + 76, + 779, + 468, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Position embedding. Incorporating representations of position information via position embedding is a common practice for Vision Transformer (ViT) [11, 23, 28], since ViT models must capture spatial relationships of tokens. We thus introduce the concept of position embedding into PointMeta. Within a point cloud, points are organized in", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an unstructured format, however the points are highly correlated in the 3D Euclidean space. Therefore, the building block of point cloud analysis should possess the position awareness, which is provided by the position embedding function $\\phi^e$ . The function $\\phi^e$ is usually combined with the neighbor update function $\\phi^n$ or the aggregation function $A$ implicitly or explicitly, thus we present them with an bundled notation, $\\phi^n\\circ \\phi^e$ or $\\phi^a\\circ \\phi^e$ .", + "bbox": [ + 496, + 340, + 890, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Instantiation Examples", + "text_level": 1, + "bbox": [ + 500, + 470, + 714, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we select three representative models from the literature as the instantiation examples and explain how these models implement the meta functions within PointMeta. We illustrate more instantiation examples [5, 8, 16, 22, 27, 29] in the supplementary due to the limited space. Sometimes these meta functions are implemented sophisticatedly, we will decompose them into several subfunctions for the sake of brevity and readability.", + "bbox": [ + 496, + 494, + 890, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PointNet++ [14]. The neighbor update is described as follows:", + "bbox": [ + 498, + 614, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\mathcal {N} (i)} ^ {\\prime}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 655, + 795, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {j} = \\operatorname {M L P} _ {1} \\left(\\operatorname {C o n c a t} \\left(f _ {j} ^ {\\prime}, p _ {j} - p _ {i}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 676, + 808, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{MLP}_1$ denotes the multi-layer perceptions applied on the concatenated neighbor features. The neighbor aggregation is implemented using a simple max pooling operation,", + "bbox": [ + 500, + 703, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {(1)} = \\operatorname {M a x} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 758, + 764, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The point update function is another MLP layer, thus we have", + "bbox": [ + 496, + 787, + 890, + 815 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {(2)} = \\mathrm {M L P} _ {2} (f _ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 814, + 754, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the position embedding function $\\phi_{e}$ is implicitly implemented by $\\mathrm{MLP}_1$ , which incorporates the concatenation of the relative positions $p_j - p_i$ and grouping features $f_j'$ as input and returns position-aware neighbor features $f_j$ .", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17684", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Transformer [32]. The neighbor update is described as follows:", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\prime} = \\operatorname {M L P} _ {1} (f _ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 128, + 328, + 146 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\mathcal {N} (i)}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 148, + 372, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The neighbor aggregation is implemented using vector self-attention, which is described as", + "bbox": [ + 76, + 172, + 468, + 200 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} e ^ {j} = \\operatorname {M L P} _ {2} \\left(p _ {j} - p _ {i}\\right), \\\\ M _ {j} = \\operatorname {S o f t m a x} \\left(\\mathrm {M L P} _ {5} \\left(\\mathrm {M L P} _ {3} \\left(f _ {i}\\right) - \\mathrm {M L P} _ {4} \\left(f _ {j}\\right) + e _ {j}\\right)\\right), \\\\ f _ {i} ^ {(1)} = \\sum_ {j \\in \\mathcal {N} (i)} \\left(M _ {j} \\odot \\left(f _ {j} + e _ {j}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 209, + 455, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\odot$ denotes the Hadamard product and $M\\in \\mathbb{R}^{k\\times d}$ denotes the vector self-attention (VSA) mask. Finally, the point feature is updated with an MLP layer and a shortcut, then we have", + "bbox": [ + 76, + 291, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {(2)} = \\mathrm {M L P} _ {6} \\left(f _ {i} ^ {(1)}\\right) + f _ {i}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 357, + 356, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ASSANet [15]. The neighbor update is described as:", + "bbox": [ + 96, + 383, + 468, + 400 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\prime} = \\operatorname {M L P} _ {1} (f _ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 422, + 328, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\mathcal {N} (i)} ^ {\\prime}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 441, + 372, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {j} = \\operatorname {R e p e a t} _ {1} \\left(f _ {j} ^ {\\prime}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 463, + 334, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the operation $\\mathrm{Repeat}_1$ means repeating the features three times to produce $f_{j}\\in \\mathbb{R}^{3d\\times 1}$ . The neighbor aggregation is implemented using an anistropic function, which is described as:", + "bbox": [ + 76, + 487, + 468, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {j} = \\operatorname {R e p e a t} _ {2} \\left(p _ {j} - p _ {i}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 554, + 351, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {(1)} = \\frac {1}{k} \\sum_ {j \\in \\mathcal {N} (i)} \\left(f _ {j} \\odot e _ {j}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 574, + 361, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the position embedding $e^j \\in \\mathbb{R}^{3d \\times 1}$ is produced by simply repeating the relative coordinates $d$ times. Note that, a similar position embedding and aggregation method are also adopted by PosPool [10], which is named position pooling. Similar with Point Transformer [32], the point feature is updated with an MLP layer and a shortcut.", + "bbox": [ + 76, + 621, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Explore Best Practices", + "text_level": 1, + "bbox": [ + 76, + 724, + 289, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given how many instantiations of PointMeta have appeared in the literature, we should push this general framework as far as possible in the architecture exploration to determine the most crucial implementation details. With the help of PointMeta, it is convenient to use quick experiments with variables controlled to verify any empirical observations or assumptions summarized from the comparison. In this section, we first compare and analyze the building blocks of existing models from the implementation details of these meta functions (Sec. 4.1, 4.2, 4.3, 4.4). Then,", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we summarize the best practices from the systematic analysis (Sec. 4.5). Finally, applying these practices, we do simple tweaks on the building blocks and propose a simple yet well-performing block, PointMetaBase (Sec. 4.6).", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Neighbor Update", + "text_level": 1, + "bbox": [ + 498, + 161, + 668, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the order of Group and MLP operation, the neighbor update function is categorized into two classes:", + "bbox": [ + 498, + 184, + 890, + 214 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Group before MLP [7, 14, 16, 27].", + "MLP before Group [6,8,10,15,32]." + ], + "bbox": [ + 517, + 224, + 769, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As pointed out by ASSANet [15], compared with the first order, adopting the second order reduces the FLOPs by $\\frac{d\\times d\\times N\\times K\\times L}{d\\times d\\times N\\times L} = K$ times, where $N$ denotes the point number, $K$ denotes the neighbor number, $L$ denotes the layer number of the MLP. Although applying MLP before Group brings a large efficiency gain for the neighbor update, a side effect is that the relative coordinates cannot be used as input for the MLP. However, this problem can be easily solved by other position embedding methods, which is specified in Sec. 4.2.", + "bbox": [ + 496, + 273, + 890, + 424 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4af9eac1eaf69dcdbb63ad9793f161cc82b165b0336fc3298737331a2d9f376c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantmIoU (%)Params (M)FLOPs (G)
Plain-Max47.3±0.72.01.4
Plain-PP65.1±0.22.01.5
Plain-PP-Max58.4±0.62.01.5
Plain-IPE-Max68.0±0.32.013.8
Plain-EPE-Max69.0±0.32.01.8
Plain-EPE-PP65.4±0.12.01.8
", + "bbox": [ + 526, + 435, + 864, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Performance of the blocks with different position embedding functions and neighbor aggregation function on Area 5 of S3DIS [1]. Plain-PP-Max means multiplying the neighbor features with relative coordinates as PP but aggregating via max pooling.", + "bbox": [ + 496, + 571, + 890, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Position Embedding", + "text_level": 1, + "bbox": [ + 500, + 656, + 691, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The manners of implementing the position embedding can be categorized into three classes:", + "bbox": [ + 498, + 679, + 890, + 709 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Implicit position embedding (IPE) [7, 14, 16, 27].", + "- Explicit position embedding (EPE) [6,8,32].", + "- Position pooling (PP) [10, 15]." + ], + "bbox": [ + 517, + 719, + 854, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implicit position embedding (IPE). Following PointNet++ [14], many methods [12, 16, 31] concatenate the relative coordinates with the neighbor features as inputs of the MLP layer in the neighbor update function. As going deeper through the networks, the dimension of the neighbor features become higher while the coordinates dimension still remains 3. This may cause the loss of position", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17685", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "information in the high-level features. In addition, IPE requires applying MLP before Group in the neighbor update function, which will increase the computation $K$ times.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Explicit position embedding (EPE). Incorporating explicit representations of position information is a common practice for ViT [11, 23, 28]. As mentioned in Sec. 3.2, Point Transformer [32] learn a position embedding $e^j$ using $\\mathrm{MLP}_2$ and add it to the neighbor features $f_j$ . Compare with IPE, EPE avoid the dimension imbalance between the position representations and the neighbor features. Because the input dimension of $\\mathrm{MLP}_2$ is only 3, the computation of EPE becomes very small by controlling the dimension of the hidden layers. PointCNN [8] also uses an MLP to implement the EPE function, but it merges the EPE by concatenation instead of addition, which incurs a computation increase. StratifiedFormer [6] maintains a learnable quantized lookup table as EPE. In this section, we select the EPE of Point Transformer as the representative EPE method for the later experiments due to its simplicity.", + "bbox": [ + 75, + 137, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Position pooling (PP). As mentioned in Sec. 3.2, AS-SANet [15] and PosPool [10] use relative coordinates as weights to pool neighbor features. This way can embed the position information into the neighbor features and make an aggregation function without learnable weights. The efficiency of position pooling is worth affirming, but embedding the position in a non-data-driven way may harm the learned neighbor features.", + "bbox": [ + 75, + 380, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Empirical analysis. We conduct empirical experiments to evaluate the different PE functions described above. The first step is to construct a plain block as the baseline. Following the suggestion given in Sec. 4.1, the neighbor update function adopts the order of MLP-before-Group. The layer number of the MLP here is set to 1. The neighbor aggregation function is a simple max pooling. The point update function is a 1-layer MLP. We denote this block as PlainMax. We equip Plain-Max with the three kinds of position embedding described above and compare their performance on Area 5 of S3DIS [1]. The MLP used respectively in EPE and IPE are both 1-layer. The macro-architecture is in level \"L\" (described in Sec. 4.6). As shown in Tab. 1, PlainEPE-Max achieves the best performance with a tiny computation increase against Plain-PP, Plain-PP-Max and Plain-IPE-Max. Because coupling the neighbor features transformation with position embedding, Plain-IPE incurs heavy redundant computation. In a summary, EPE is the optimal choice among the three kinds of position embeddings.", + "bbox": [ + 75, + 501, + 468, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Neighbor Aggregation", + "text_level": 1, + "bbox": [ + 76, + 801, + 284, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to whether learnable weights exist in the neighbor aggregation function, we divide the aggregations function into two types:", + "bbox": [ + 75, + 825, + 468, + 872 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Non-learnable aggregation [7, 10, 14-16, 27].", + "bbox": [ + 94, + 885, + 406, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/654de023b9701b49ffddef88e810d1fa38cf2158d7f925859b409147ca4771f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantsmIoU (%)Params (M)FLOPs (G)
Point Trans [32]70.5±0.37.85.6
Point Trans (-VSA+Max)70.3±0.25.13.3
", + "bbox": [ + 500, + 88, + 903, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Ablation of the vector self-attention of Point Transformer [32] on Area 5 of S3DIS [1]. -VSA+Max means replacing VSA with max pooling.", + "bbox": [ + 498, + 162, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Learnable aggregation [6,8,22,29,32].", + "bbox": [ + 517, + 231, + 790, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learnable aggregation. As mentioned in Sec. 3.1, the aggregation function should satisfy permutation invariance because the point cloud is unordered. The permutation invariance is acquired by learning a space transformation [8, 29], learning pointwise weights [32], or pairwise self-attention [6]. In this section, we select the vector self-attention of Point Transformer [32] as the representative implementation of the learnable aggregation function due to its excellent performance.", + "bbox": [ + 496, + 256, + 890, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Non-learnable aggregation. There are two types of non-learnable aggregation in the literatures: position pooling [10, 15] and max pooling [7, 14, 16, 27]. As mentioned in Sec. 4.2, position pooling embed the position in a nondata-driven way and may compete with the learned neighbor features. As shown in Tab. 1, Plain-EPE-Max surpasses Plain-EPE-PP significantly, which supports our assumption. Therefore, we select the max pooling as the representative non-learnable aggregation function.", + "bbox": [ + 496, + 392, + 890, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Empirical analysis. We conduct an ablation experiment on Point Transformer [32] to compare the aggregation power of both max pooling and vector self-attention. Specifically, we replace the VSA with a max pooling operation (denoted as -VSA+Max). As a result, the aggregation of the ablated Point Transformer is formulated as:", + "bbox": [ + 496, + 527, + 890, + 618 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {j} = \\operatorname {M L P} _ {2} \\left(p _ {j} - p _ {i}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 628, + 769, + 646 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {(1)} = \\operatorname {M a x} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)} + \\boldsymbol {e} _ {\\mathcal {N} (i)}\\right), j \\in \\mathcal {N} (i), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 670, + 890, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_{\\mathcal{N}(i)} = \\{e_j | j \\in \\mathcal{N}(i)\\}$ . As shown in Tab. 2, compared with VSA, max pooling achieves comparable performance with much fewer parameters and FLOPs. Here comes a question, why max pooling has a comparable power with the learnable aggregation? We explore the answer within our PointMeta framework and find that the max pooling is a special case of VSA but with binary and sparse attention. Specifically, the aggregation function Eq. 4 can be reformulated as follows:", + "bbox": [ + 496, + 694, + 890, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {j} = \\operatorname {S o f t m a x} _ {T \\rightarrow 0} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 843, + 787, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {(1)} = \\sum_ {j \\in \\mathcal {N} (i)} \\left(M _ {j} \\odot \\left(f _ {j} + e _ {j}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 863, + 803, + 896 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17686", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/323155c65b0e41b65c644c3c9eceaac8ada571f6efde3a1ea6de33dce4e4bc6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantsmIoU (%)Params (M)FLOPs (G)
N1P169.0±0.32.01.7
N2P068.7±0.32.01.7
N0P267.8±0.42.01.7
N1P269.5±0.32.72.0
N1P2-Inv69.7±0.37.13.8
N2P169.4±0.32.72.0
N2P1-Inv69.6±0.47.13.8
N1P369.5±0.33.42.3
N3P169.6±0.53.42.3
", + "bbox": [ + 125, + 88, + 421, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Semantic segmentation of the Plain-EPE with different MLP configuration on Area 5 of S3DIS [1]. \"N1P2\" means that the MLP in the neighbor update is 1-layer and 2-layer. \"Inv\" means the inverted bottleneck.", + "bbox": [ + 75, + 268, + 468, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{Softmax}_{T\\to 0}$ denotes a softmax function whose temperature approximates 0, which makes the softmax output $M_{j}$ approximates the binary masks. This formulation reveals the collaboration between max pooling and the neighbor update function, i.e., max pooling is an aggregation function with binary and sparse attention which is generated by applying softmax on the output of the neighbor update", + "bbox": [ + 75, + 351, + 467, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Point Update", + "text_level": 1, + "bbox": [ + 76, + 467, + 215, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The point update function is used to enhance the point features after the neighbor aggregation. Several existing studies do not implement this function in their building blocks [7, 27], and some other studies use a heavy point update function, such as a 2-layer MLP with an inverted bottleneck [6, 16]. A natural question arises: how much computation complexity should we allocate for the point update function? We answer this question via extensive ablated experiments.", + "bbox": [ + 75, + 489, + 467, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Empirical analysis. According to the exploration above, we adopt the Plain-EPE-Max as the baseline due to its excellent performance in the ablated experiments (Tab. 1). Since the computation are concentrated on the MLP of the neighbor update function and point update function, we can allocate computation between the two update functions by changing the configurations of the two MLP. We ablate the MLP from two aspects, i.e., the layer number and whether to use an inverted bottleneck that adopted by PointNeXt [16] and StratifiedFormer [6]. As shown in Tab. 3, the configuration \"N1P2\" can achieve an optimal balance between complexity and performance. More layers or inverted bottlenecks can only promote the model marginally but lead to much more computation.", + "bbox": [ + 75, + 626, + 467, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Best Practices", + "text_level": 1, + "bbox": [ + 76, + 847, + 220, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we summarize the best practices for the building block as follows:", + "bbox": [ + 75, + 869, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- **Position Embedding.** Explicit position embedding can endow the models with better position awareness with a low computation cost comparing with other PE manners.", + "bbox": [ + 517, + 90, + 890, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Point Update. Point update is important and the complexity allocation between the neighbor update and the point update should be properly set.", + "bbox": [ + 517, + 161, + 890, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Neighbor Aggregation. Max pooling is not a pure non-learnable aggregation function, its aggregation power is comparable with the learnable aggregation function but with much less computation complexity. Any exploration for a new aggregation function should set max pooling as the baseline.", + "bbox": [ + 517, + 217, + 890, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Neighbor Update. Applying the MLP before neighbors grouping will bring a $K$ times FLOPs reduction comparing with using the contrary order. However, the position embedding function must be carefully chosen for the position awareness.", + "bbox": [ + 517, + 318, + 890, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The first and the second practice are usually ignored by researchers but are vital for the effectiveness and efficiency of the building block, which is demonstrated in Sec. 4.2 and Sec. 4.4. The third shares a similar viewpoint with PosPool [10] which claims that different local aggregation operators perform similarly. However, from the perspective of PointMeta, we provide a deeper insight that max pooling can be reformulated as a special case of vector self-attention [32], which makes it have comparable aggregation power comparing with the learnable aggregation function. The fourth practice about neighbor update has been emphasized by ASSANet [15], we restate it here because it is important to notice its collaboration with position embedding function. We believe that these practices can provide a guideline for the future research.", + "bbox": [ + 496, + 404, + 890, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. PointMetaBase", + "text_level": 1, + "bbox": [ + 500, + 640, + 653, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply the best practices summarized in Sec. 4.5 to design the building block and implement the four meta functions as simply as possible. We name this block as PointMetaBase, which we hope to become a solid baseline for future research. Fig. 2 (g) shows the architecture of PointMetaBase. The layer number of the first MLP is set to 1 and that of the second MLP is set to 2. The neighbor aggregation function is a simple max pooling. We use the EPE adopted by Point Transformer [32] as the position embedding function. Applying EPE and the MLP-before-Group order, we tweak the set abstraction module [14, 16] as the reduction block, termed PointMetaSA. We adopt the same scaling strategies and decoder with PointNeXt [16] to construct our PointMetaBase family:", + "bbox": [ + 496, + 662, + 890, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- PointMetaBase-S: $C = 32$ , $B = 0$", + "bbox": [ + 517, + 885, + 764, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17687", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/140a4fe15b827e27b326c5862fbabbcb1a883b84730102789e9211d4f52eddda.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodS3DIS 6-FoldS3DIS Area-5ScanNet mIoUParams. MFLOPs GThroughput (ins./sec.)
mIoU (%)OA (%)mIoU (%)OA (%)Val (%)Test (%)
PointNet++ [14]54.581.053.583.053.555.71.07.2237
PointCNN [8]65.488.157.385.9-45.80.6--
DeepGCN [7]60.085.952.5---3.6--
KPConv [22]70.6-67.1-69.268.615.0--
RandLA-Net [5]70.088.0---64.51.35.8-
BAAF-Net [17]72.288.965.488.9--5.0--
Point Transformer [32]73.590.270.490.870.6-7.85.6-
CBL [20]73.189.669.490.6-70.518.6--
ASSANet [15]--65.888.9--2.42.5300
ASSANet-L [15]--68.089.7--115.636.2136
PointNeXt-S [16]68.087.463.4±0.887.9±0.364.5-0.83.6276
PointNeXt-B [16]71.588.867.3±0.289.4±0.168.4-3.88.9161
PointNeXt-L [16]73.989.869.0±0.590.0±0.169.4-7.115.2109
PointNeXt-XL [16]74.990.370.5±0.390.6±0.171.571.241.684.843
PointMetaBase-L75.690.669.5±0.390.5±0.171.0-2.72.0187
PointMetaBase-XL76.391.071.1±0.490.9±0.171.8-15.39.2104
PointMetaBase-XXL77.091.371.3±0.790.8±0.672.871.419.711.090
", + "bbox": [ + 86, + 88, + 883, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Semantic segmentation on S3DIS [1] (6-Fold and Area 5) and ScanNet V2 [3]. Our PointMetaBase family surpass the state-of-the-art method PointNeXt [16] significantly with large FLOPs reduction.", + "bbox": [ + 75, + 412, + 892, + 443 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- PointMetaBase-L: $C = 32$ , $B = (2,4,2,2)$", + "- PointMetaBase-XL: $C = 64$ , $B = (3,6,3,3)$", + "- PointMetaBase-XXL: $C = 64$ , $B = (4,8,4,4)$" + ], + "bbox": [ + 94, + 467, + 426, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$C$ denotes the channel size of the stem MLP and $B$ denotes the number of the PointMetaBase block in each stage. Note that when $B = 0$ , only one PointMetaSA block but no PointMetaBase blocks are used at each stage. Due to the excellent efficiency of PointMetaBase, we do not need to construct the network at the level \"B\" as done in Point-NeXt [16]. In contrast, we choose to scale the model to a larger level \"XXL\". The full details of network architecture are available in the supplementary materials.", + "bbox": [ + 75, + 553, + 472, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 707, + 209, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We studied the performance of PointMetaBase on S3DIS [1], ScanNet V2 [3], ScanObjectNN [24] and ShapeNetPart [30]. To enable a fair comparison, the same data processing and evaluation protocols adopted by the state-of-the-art method PointNeXt [16] were used in our experiments.", + "bbox": [ + 75, + 732, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental Setups. We optimized all our models using CrossEntropy loss with label smoothing [19], AdamW optimizer, an initial learning rate 0.001, and weight decay $10^{-4}$ , with Cosine Decay, with a 32G V100 GPU. Model parameters, FLOPs and inference throughput are provided for comparison. For a fair comparison, we use the open", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "codes provided by PointNext [16] to conduct the measurements. We also use MindSpore to validate the generalization on different deep learning frameworks. All the measurements are conducted using an NVIDIA Tesla A100 40GB GPU and a 12-core Intel Xeon @ 2.40GHz CPU.", + "bbox": [ + 496, + 467, + 892, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.3D Scene Segmentation", + "text_level": 1, + "bbox": [ + 498, + 551, + 718, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "S3DIS [1] is a challenging benchmark composed of 6 large-scale indoor areas, 271 rooms, and 13 semantic categories in total. The standard 6-fold cross-validation results and the Area-5 results on S3DIS are reported in Tab. 4. Since we adopt the same scaling strategies, PointMetaBase-L and PointMetaBase-XL have the same configuration on width and depth as PointNeXt-L and PointNeXt-XL. With only $13\\%$ FLOPs, our PointMetaBase-L outperforms PointNext-L by $1.7\\%$ and $0.8\\%$ in terms of mean IoU (mIoU) and overall accuracy (OA), respectively, and is much faster in terms of throughput. Compared with PointNext-XL, our PointMetaBase-XL uses only $11\\%$ of the FLOPs, but surpasses with $1.4\\%$ mIoU and $0.7\\%$ OA. The inference speed is $2.4\\times$ faster. The excellent efficiency and performance of PointMetaBase can be attributed to the the removal of the computation redundancy in the two update functions and the improved position awareness introduced by explicit position embedding. ScanNet [3] contains 3D indoor scenes of various rooms with 20 semantic categories. The dataset is divided into 3 part: training,", + "bbox": [ + 496, + 575, + 893, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/guochengqian/PointNeXt", + "bbox": [ + 517, + 886, + 759, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17688", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/edadc472c09b94023cc0a7769a5805f4f822370bbe2a5b9415736ce0b6673545.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodins. mIoU (%)cls. mIoU (%)Params (M)FLOPs (G)Throughput (ins./sec)
PointNet++ [14]85.181.91.04.9-
KPConv [22]86.485.1---
ASSANet [15]86.1----
Point Transformer [32]86.683.77.8--
PointMLP [12]86.184.6---
StratifiedFormer [6]86.685.1---
PointNeXt-S [16]86.7±0.084.4±0.21.04.5873
PointNeXt-S (C=64) [16]86.9±0.184.8±0.53.717.8391
PointNeXt-S (C=160) [16]87.0±0.185.2±0.122.5110.2118
PointMetaBase-S (ours)86.7±0.084.3±0.11.01.391194
PointMetaBase-S (C=64) (ours)86.9±0.184.9±0.23.83.85706
PointMetaBase-S (C=160) (ours)87.1±0.085.1±0.322.718.45271
", + "bbox": [ + 158, + 88, + 812, + 305 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/357e88776e7e3329831d6c65160e6ca6b7451f5e6ab604c51fe5c40f13600968.jpg", + "table_caption": [ + "Table 5. Part segmentation on ShapeNetPart [30]. Compaerd with the state-of-the-art method PointNeXt [16], our PointMetaBase family achieve comparable performance with large FLOPs reduction and faster inference speed." + ], + "table_footnote": [], + "table_body": "
VariantsOA (%)FLOPs (G)TP (ins./sec)
PointNet++ [14]77.91.72655
PointCNN [8]78.5--
DGCNN [27]78.14.8735
DRNet [13]80.3--
PointMLP [12]85.4±1.331.3405
PointNeXt-S [16]87.7±0.41.72230
PointMetaBase-S87.9±0.20.62674
", + "bbox": [ + 89, + 367, + 457, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Classification on ScanObjectNN [24]. Our model surpasses PointNeXt-S [16] both with much less computation cost.", + "bbox": [ + 75, + 525, + 468, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "validation, and test splits, with 1201, 312 and 100 scans, respectively. Following PointNeXt [16], we do not use any voting strategies. As shown in Tab. 4, our PointMetaBase-L surpasses PointNeXt-L $1.6\\%$ mIoU with only $14\\%$ FLOPs and $1.7\\times$ faster inference speed. PointMetaBase-XL surpasses PointNeXt-XL $0.3\\%$ mIoU with only $11.3\\%$ FLOPs and $2.4\\times$ faster inference speed.", + "bbox": [ + 75, + 579, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. 3D Object Classification", + "text_level": 1, + "bbox": [ + 76, + 693, + 302, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ScanObjectNN [24] contains about 15000 real scanning objects, which are grouped into 15 classes and have 2902 unique instances. Following [12, 16], we experiment on the hardest perturbed variant (PB_T50_RS). To make a fair comparison with PointNeXt [16], we did not use upscaled variants of PointMetaBase on this benchmark. As reported in Tab. 6, our model surpasses PointNeXt-S, while using much fewer model FLOPs and running much faster.", + "bbox": [ + 75, + 717, + 470, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. 3D Part Segmentation", + "text_level": 1, + "bbox": [ + 76, + 847, + 284, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ShapeNetPart [30] is a dataset for part segmentation of objects. It consists of 16880 models with 16 differ", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ent shape categories. Following the state-of-the-art method PointNeXt [16], we used voting by averaging the results of 10 randomly scaled input point clouds, with scaling factors equal to [0.8,1.2]. As shown in Tab. 5, compaerd with PointNeXt, our PointMetaBase family achieve comparable performance with large FLOPs reduction and faster inference speed. Specially, for $C = 32 / C = 64 / C = 160$ , the FLOPs of our PointMetaBase are only $30\\% 21\\% / 17\\%$ that of PointNeXt, and the inference speed of PointMetaBase is $1.4 / 1.8 / 2.3$ times faster than that of PointNeXt.", + "bbox": [ + 496, + 369, + 893, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 540, + 619, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we reformulate existing models for point cloud analysis into a general framework, PointMeta. Within this framework, we provide an in-depth analysis on the building blocks of existing models and summarize a few best practices. Furthermore, we follow these best practices and propose a simple building block, PointMetaBase, which enjoys excellent performance and efficiency. Our work can inspire new insights into building block design and future architecture exploration. Next, building on this work, we plan to explore neural architecture search methods [33-37] that are specifically tailored for point cloud analysis.", + "bbox": [ + 496, + 566, + 893, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by National Key R&D Program of China (No. 2022ZD0118202), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J01002, No. 2022J06001), Guangdong Basic and Applied Basic Research Foundation (No. 2019B1515120049), and CAAI-Huawei MindSpore Open Fund.", + "bbox": [ + 496, + 733, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17689", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 118, + 173, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Sasha Sax, Amir R. Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. CoRR, abs/1702.01105, 2017. 1, 2, 4, 5, 6, 7", + "[2] Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Flores Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gulçehre, H. Francis Song, Andrew J. Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, Kelsey R. Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matthew M. Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational inductive biases, deep learning, and graph networks. CoRR, abs/1806.01261, 2018. 2", + "[3] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 2432–2443. IEEE Computer Society, 2017. 7", + "[4] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1", + "[5] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11105-11114, 2020. 2, 3, 7", + "[6] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8490-8499. IEEE, 2022. 2, 4, 5, 6, 8", + "[7] Guohao Li, Matthias Müller, Ali K. Thabet, and Bernard Ghanem. Deep GCs: Can GCs go as deep as cnns? In ICCV, pages 9266-9275, 2019. 2, 4, 5, 6, 7", + "[8] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, pages 828-838, 2018. 2, 3, 4, 5, 7, 8", + "[9] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In CVPR, pages 8895-8904, 2019. 2", + "[10] Ze Liu, Han Hu, Yue Cao, Zheng Zhang, and Xin Tong. A closer look at local aggregation operators in point cloud analysis. In ECCV, pages 326-342, 2020. 2, 4, 5, 6", + "[11] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002, 2021. 3, 5", + "[12] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 4, 8", + "[13] Charles Ruizhongtai Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 77-85, 2017. 1, 2, 8", + "[14] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on" + ], + "bbox": [ + 78, + 143, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1, 2, 3, 4, 5, 6, 7, 8", + "[15] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali K. Thabet, and Bernard Ghanem. Assanet: An anisotropic separable set abstraction for efficient point cloud representation learning. In NeurIPS, pages 28119-28130, 2021. 2, 3, 4, 5, 6, 7, 8", + "[16] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Abed Al Kader Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[17] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 7", + "[18] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, ICLR, 2015. 1", + "[19] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7", + "[20] Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. In CVPR, pages 8479-8489, 2022. 7", + "[21] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, pages 3887-3896, 2018. 2", + "[22] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6410–6419, 2019. 2, 3, 5, 7, 8", + "[23] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 3, 5", + "[24] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, pages 1588–1597, 2019. 7, 8", + "[25] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, pages 10296-10305, 2019. 2", + "[26] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In CVPR, pages 2589-2597, 2018. 2", + "[27] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. ACM Trans. Graph., 38(5):146:1-146:12, 2019. 2, 3, 4, 5, 6, 8", + "[28] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In ICCV, pages 10013-10021, 2021. 3, 5" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17690", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Wenxuan Wu, Zhongang Qi, and Fuxin Li. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, pages 9621-9630, 2019. 3, 5", + "[30] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Trans. Graph., 35(6):210:1-210:12, 2016. 7, 8", + "[31] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, pages 5565-5573, 2019. 4", + "[32] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16239-16248, 2021. 2, 3, 4, 5, 6, 7, 8", + "[33] Xiawu Zheng, Xiang Fei, Lei Zhang, Chenglin Wu, Fei Chao, Jianzhuang Liu, Wei Zeng, Yonghong Tian, and Rongrong Ji. Neural architecture search with representation mutual information. In CVPR, pages 11912-11921, 2022. 8", + "[34] Xiawu Zheng, Rongrong Ji, Yuhang Chen, Qiang Wang, Baochang Zhang, Jie Chen, Qixiang Ye, Feiyue Huang, and Yonghong Tian. Migo-nas: Towards fast and generalizable neural architecture search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9):2936-2952, 2021. 8", + "[35] Xiawu Zheng, Rongrong Ji, Lang Tang, Baochang Zhang, Jianzhuang Liu, and Qi Tian. Multinomial distribution learning for effective neural architecture search. In ICCV, 2019. 8", + "[36] Xiawu Zheng, Rongrong Ji, Qiang Wang, Qixiang Ye, Zhenguo Li, Yonghong Tian, and Qi Tian. Rethinking performance estimation in neural architecture search. In CVPR, 2020. 8", + "[37] Xiawu Zheng, Yang Zhang, Sirui Hong, Huixia Li, Lang Tang, Youcheng Xiong, Jin Zhou, Yan Wang, Xiaoshuai Sun, Pengfei Zhu, Chenglin Wu, and Rongrong Ji. Evolving fully automated machine learning via life-long knowledge anchors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9):3091-3107, 2021. 8" + ], + "bbox": [ + 78, + 90, + 467, + 627 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17691", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_model.json b/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7a0774540c97f8e7e76f694395247d8cb67d5246 --- /dev/null +++ b/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_model.json @@ -0,0 +1,2332 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.264, + 0.131, + 0.709, + 0.154 + ], + "angle": 0, + "content": "Meta Architecture for Point Cloud Analysis" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.18, + 0.766, + 0.217 + ], + "angle": 0, + "content": "Haojia Lin\\(^{1}\\), Xiawu Zheng\\(^{2}\\), Lijiang Li\\(^{1}\\), Fei Chao\\(^{1}\\), Shanshan Wang\\(^{2,3}\\), Yan Wang\\(^{4}\\), Yonghong Tian\\(^{2,5}\\), Rongrong Ji\\(^{1,2,6*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.217, + 0.861, + 0.251 + ], + "angle": 0, + "content": "1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, School of Informatics, Xiamen University, 361005, P.R. China." + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.252, + 0.77, + 0.269 + ], + "angle": 0, + "content": "\\(^{2}\\)Peng Cheng Laboratory. \\(^{3}\\)Chinese Academy of Sciences. \\(^{4}\\)Samsara Inc." + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.27, + 0.807, + 0.305 + ], + "angle": 0, + "content": "\\(^{5}\\)National Engineering Research Center for Visual Technology, Peking University. \\(^{6}\\)Shenzhen Research Institute of Xiamen University." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.307, + 0.844, + 0.323 + ], + "angle": 0, + "content": "{linhj, lilijiang}@stu.xmu.edu.cn, zhengxw01@pcl.ac.cn, {rrji, feichao}@xmu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.326, + 0.744, + 0.34 + ], + "angle": 0, + "content": "ss.wang@siat.ac.cn, yan.wang@samsara.com, yhtian@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.375, + 0.314, + 0.391 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.416, + 0.473, + 0.825 + ], + "angle": 0, + "content": "Recent advances in 3D point cloud analysis bring a diverse set of network architectures to the field. However, the lack of a unified framework to interpret those networks makes any systematic comparison, contrast, or analysis challenging, and practically limits healthy development of the field. In this paper, we take the initiative to explore and propose a unified framework called PointMeta, to which the popular 3D point cloud analysis approaches could fit. This brings three benefits. First, it allows us to compare different approaches in a fair manner, and use quick experiments to verify any empirical observations or assumptions summarized from the comparison. Second, the big picture brought by PointMeta enables us to think across different components, and revisit common beliefs and key design decisions made by the popular approaches. Third, based on the learnings from the previous two analyses, by doing simple tweaks on the existing approaches, we are able to derive a basic building block, termed PointMetaBase. It shows very strong performance in efficiency and effectiveness through extensive experiments on challenging benchmarks, and thus verifies the necessity and benefits of high-level interpretation, contrast, and comparison like PointMeta. In particular, PointMetaBase surpasses the previous state-of-the-art method by \\(0.7\\% /1.4\\% 2.1\\%\\) mIoU with only \\(2\\% /11\\% /13\\%\\) of the computation cost on the S3DIS datasets. The code and models are available at https://github.com/linhaojia13/PointMetaBase." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.377, + 0.888, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.605, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Segmentation performance of PointMetaBase on S3DIS [1]. PointMetaBase surpasses the state-of-the-art method PointNeXt [16] significantly with a large FLOPs reduction." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.676, + 0.631, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.854 + ], + "angle": 0, + "content": "In the past two decades, the popularity of 3D data acquisition technology has led to great interest in point cloud analysis. Unlike images, point clouds are inherently sparse, unordered, and irregular. These characteristics make it challenging for the powerful convolutional neural networks (CNN) [4, 18] to extract useful information from point clouds. Early studies attempt to convert the point cloud into grids by either voxelization or 2D projections, such that the standard CNN can be applied, but the applications are limited by the extra computation and information loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With the introduction of permutation equivariance and invariance by PointNet [13] and PointNet++ [14], it is possible to use CNNs to process point clouds in their unstruc" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.318, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author: rrji@xmu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17682" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "tured format. Following PointNet++ [14], many point-based networks have been introduced, most of which focus on the development of complex building blocks to extract local features, such as the \\(\\mathcal{X}\\) - Conv convolution in PointCNN [22] and the self-attention layer in Point Transformer [32]. Although brought significant performance gains, these models are very complicated. On one hand, this makes the approaches computational demanding and thus limited in applications. On the other hand, it also gives extra challenges in systematic comparison and analysis among the models, and thus impacts efficient development of the field, which ideally would benefit from theoretic and empirical guidance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.544 + ], + "angle": 0, + "content": "The community is well aware of this issue, and research efforts have been put on looking for a unified perspective to compare these methods and identify the most crucial implementation details. PosPool [10] and PointNeXt [16] take a step toward this goal. Adopting a deep residual architecture as the base network, PosPool [10] evaluates the representative building blocks [8, 14, 22, 26] and finds that they perform similarly well. PointNeXt [16] revisits the improved training and scaling strategies adopted by previous SotA methods [7, 15, 22, 32] and tweaks the early model PointNet++ [14] with the learnings, which achieves state-of-the-art performance. Both of them give insights and inspirations to the community via dedicated experiment designs and extensive empirical analysis. However, neither of them provides a framework that is general enough to fit a broad range of point cloud analysis approaches, such that large scale comparison and contrast could be done." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.473, + 0.71 + ], + "angle": 0, + "content": "In this paper, we argue that if viewed from the perspective of basic building blocks, the majority of existing approaches could be fit into a single meta architecture (Sec. 3). We call it PointMeta. PointMeta abstracts the computation pipeline of the building blocks for point cloud analysis into four meta functions: a neighbor update function, a neighbor aggregation function, a point update function and a position embedding function (implicit or explicit). As will be discussed in more details in the next sections, this framework brings three benefits, which are also our core technical contributions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.724, + 0.473, + 0.876 + ], + "angle": 0, + "content": "- In the dimension of models, it allows us to compare and contrast different models in a fair manner. So it becomes practical to observe and summarize learnings and assumptions, whose correctness could be further verified through experiments with variables controlled under the same framework. For example, among the systematic analysis on all the components across popular models, we found for the positional embedding function, explicit positional embedding is empirically the best choice (Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- In the dimension of components, it allows us to have" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.724, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "a higher level view across components, and thus to revisit the common beliefs and design decisions of the existing approaches. For example, despite the common perception, we find that the neighbor aggregation blocks may collaborate or compete with the learned neighbor features (Sec. 4.3), and thus should be designed carefully." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.209, + 0.895, + 0.345 + ], + "angle": 0, + "content": "- Based on the learnings from the previous two dimensions, we are then able to do simple tweaks on the building blocks to apply the best practices (Sec. 4.5). The result building block, PointMetaBase, achieves very strong performance, surpassing the previous state-of-the-art method [16] by \\(0.7\\% / 1.4\\% / 2.1\\%\\) mIoU with only \\(2\\% / 11\\% / 13\\%\\) of the computation cost on S3DIS [1] (Fig. 1), and can act as a baseline for further research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.642, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.895, + 0.689 + ], + "angle": 0, + "content": "Neural Networks for Point Cloud Analysis. The unstructured data format of point cloud makes it difficult to apply CNN directly. To address this challenge, PointNet [13] introduces permutation equivariance and invariance via point-wise MLP and max pooling, which makes it possible to process point cloud directly. To better capture locality, PointNet++ [14] presents a Set Abstraction module to aggregate features from the neighbor points. Most works after PointNet++ focus on the design of build blocks. Graph-based methods [7, 25, 27] utilize graph neural networks to extract local features. Pseudo-grid-based methods [8,21,22] transform the local features onto pseudo grids which allow for regular convolutions. Adaptive weight-based methods [5,9] aggregate the neighbor features via weights adaptively learned from the locality. Recently, Transformer-based block [6, 32] are introduced to capture local information through self-attention. These proposed building blocks are so complicated that make challenges in systematic comparison and analysis, which suggests an urgent need of a common framework to unify existing models in this field." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.703, + 0.614, + 0.719 + ], + "angle": 0, + "content": "3. PointMeta" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.728, + 0.68, + 0.743 + ], + "angle": 0, + "content": "3.1. Meta Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Inspired by the graph network framework [2], we reformulate existing building blocks for point cloud analysis into a single meta architecture, named PointMeta. As shown in Fig. 2 (a), a PointMeta block is composed of four meta functions: a neighbor update function \\(\\phi^n\\), a neighbor aggregation function \\(\\phi^a\\), a point update function \\(\\phi^p\\) and a position embedding function \\(\\phi^e\\) (implicit or explicit). The forward pass of a PointMeta block is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.886, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\mathcal {N} (i)} = \\phi^ {n} \\circ \\phi^ {e} \\left(f _ {i}, p _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17683" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.209, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.257, + 0.168, + 0.268 + ], + "angle": 0, + "content": "(a) PointMeta" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.091, + 0.296, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.257, + 0.293, + 0.267 + ], + "angle": 0, + "content": "(b) PointNet++" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.091, + 0.395, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.257, + 0.386, + 0.268 + ], + "angle": 0, + "content": "(c) DGCNN" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.091, + 0.517, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.258, + 0.486, + 0.268 + ], + "angle": 0, + "content": "(d) PointCNN" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.091, + 0.655, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.257, + 0.64, + 0.268 + ], + "angle": 0, + "content": "(e) Point Transformer" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.091, + 0.764, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.257, + 0.738, + 0.268 + ], + "angle": 0, + "content": "(f) ASSANet" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.091, + 0.895, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.258, + 0.873, + 0.268 + ], + "angle": 0, + "content": "(g) PointMetaBase" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.894, + 0.316 + ], + "angle": 0, + "content": "Figure 2. PointMeta and its instantiation examples. (a) In PointMeta, the position embedding function is usually combined with the neighbor update function or the aggregation function implicitly or explicitly. (b) \\(\\sim\\) (f) Representative building blocks can naturally fit into PointMeta. (g) Applying the summarized best practices, we do simple tweaks on the building blocks and propose PointMetaBase." + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.355, + 0.469, + 0.377 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {(1)} = \\phi^ {a} \\circ \\phi^ {e} \\left(f _ {i}, p _ {i}, \\mathbf {f} _ {\\mathcal {N} (i)}, \\mathbf {p} _ {\\mathcal {N} (i)}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.397, + 0.469, + 0.418 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {(2)} = \\phi^ {p} \\left(f _ {i} ^ {(1)}, p _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.546 + ], + "angle": 0, + "content": "where \\( f_{i} \\in \\mathbb{R}^{d} \\) and \\( p_i \\in \\mathbb{R}^3 \\) represent the feature and the coordinate of point \\( i \\) respectively, and \\( d \\) denotes the feature length. Let \\( \\mathcal{N}(i) \\) denote the neighbors of point \\( i \\), \\( f_{j} \\) and \\( p_j \\) be the feature and coordinate of each neighbor of point \\( i \\) respectively, \\( \\pmb{f}_{\\mathcal{N}(i)} = \\{f_j | j \\in \\mathcal{N}(i)\\} \\) represent all neighbor features of point \\( i \\), \\( \\pmb{p}_{\\mathcal{N}(i)} = \\{p_j | j \\in \\mathcal{N}(i)\\} \\) represent all neighbor coordinates of point \\( i \\), and \\( k \\) denote the neighbor number." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.684 + ], + "angle": 0, + "content": "Neighbor update. The first step is to group and update neighbors for each point. Usually, the neighbor features \\( f_{\\mathcal{N}(i)} \\) is produced by K-Nearest query as [27, 32] or ball query as [14-16]. The update on neighbor features is usually conducted by a multi-layer perceptions (MLP) to make sure the permutation equivariance of the neighbor update function \\( \\phi^n \\). Specifically, a permutation transformation on \\( f_{\\mathcal{N}(i)} \\) results in the same permutation change on the output of the neighbor update." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Neighbor aggregation. Literally, the neighbor aggregation function is used to aggregate the neighbor features for a center point. Different from \\(\\phi^n\\), the aggregation function \\(\\phi^a\\) must satisfy permutation invariance, i.e., a permutation transformation on the input \\(f_{\\mathcal{N}(i)}\\) will have no influence on the output \\(f_i^{(1)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Point update. The point update usually contains an MLP, which is used to further enhance the point features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Position embedding. Incorporating representations of position information via position embedding is a common practice for Vision Transformer (ViT) [11, 23, 28], since ViT models must capture spatial relationships of tokens. We thus introduce the concept of position embedding into PointMeta. Within a point cloud, points are organized in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.463 + ], + "angle": 0, + "content": "an unstructured format, however the points are highly correlated in the 3D Euclidean space. Therefore, the building block of point cloud analysis should possess the position awareness, which is provided by the position embedding function \\(\\phi^e\\). The function \\(\\phi^e\\) is usually combined with the neighbor update function \\(\\phi^n\\) or the aggregation function \\(A\\) implicitly or explicitly, thus we present them with an bundled notation, \\(\\phi^n\\circ \\phi^e\\) or \\(\\phi^a\\circ \\phi^e\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.715, + 0.487 + ], + "angle": 0, + "content": "3.2. Instantiation Examples" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.615 + ], + "angle": 0, + "content": "In this section, we select three representative models from the literature as the instantiation examples and explain how these models implement the meta functions within PointMeta. We illustrate more instantiation examples [5, 8, 16, 22, 27, 29] in the supplementary due to the limited space. Sometimes these meta functions are implemented sophisticatedly, we will decompose them into several subfunctions for the sake of brevity and readability." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.615, + 0.892, + 0.644 + ], + "angle": 0, + "content": "PointNet++ [14]. The neighbor update is described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.656, + 0.797, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\mathcal {N} (i)} ^ {\\prime}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.677, + 0.81, + 0.696 + ], + "angle": 0, + "content": "\\[\nf _ {j} = \\operatorname {M L P} _ {1} \\left(\\operatorname {C o n c a t} \\left(f _ {j} ^ {\\prime}, p _ {j} - p _ {i}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.704, + 0.891, + 0.75 + ], + "angle": 0, + "content": "where \\(\\mathrm{MLP}_1\\) denotes the multi-layer perceptions applied on the concatenated neighbor features. The neighbor aggregation is implemented using a simple max pooling operation," + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.759, + 0.765, + 0.78 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {(1)} = \\operatorname {M a x} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.891, + 0.816 + ], + "angle": 0, + "content": "The point update function is another MLP layer, thus we have" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.815, + 0.756, + 0.836 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {(2)} = \\mathrm {M L P} _ {2} (f _ {i})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.904 + ], + "angle": 0, + "content": "Note that the position embedding function \\(\\phi_{e}\\) is implicitly implemented by \\(\\mathrm{MLP}_1\\), which incorporates the concatenation of the relative positions \\(p_j - p_i\\) and grouping features \\(f_j'\\) as input and returns position-aware neighbor features \\(f_j\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17684" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Point Transformer [32]. The neighbor update is described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.129, + 0.33, + 0.147 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\prime} = \\operatorname {M L P} _ {1} (f _ {i}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.149, + 0.374, + 0.167 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\mathcal {N} (i)}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.173, + 0.469, + 0.202 + ], + "angle": 0, + "content": "The neighbor aggregation is implemented using vector self-attention, which is described as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.21, + 0.456, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} e ^ {j} = \\operatorname {M L P} _ {2} \\left(p _ {j} - p _ {i}\\right), \\\\ M _ {j} = \\operatorname {S o f t m a x} \\left(\\mathrm {M L P} _ {5} \\left(\\mathrm {M L P} _ {3} \\left(f _ {i}\\right) - \\mathrm {M L P} _ {4} \\left(f _ {j}\\right) + e _ {j}\\right)\\right), \\\\ f _ {i} ^ {(1)} = \\sum_ {j \\in \\mathcal {N} (i)} \\left(M _ {j} \\odot \\left(f _ {j} + e _ {j}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.292, + 0.469, + 0.352 + ], + "angle": 0, + "content": "where the \\(\\odot\\) denotes the Hadamard product and \\(M\\in \\mathbb{R}^{k\\times d}\\) denotes the vector self-attention (VSA) mask. Finally, the point feature is updated with an MLP layer and a shortcut, then we have" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.358, + 0.357, + 0.379 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {(2)} = \\mathrm {M L P} _ {6} \\left(f _ {i} ^ {(1)}\\right) + f _ {i}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.385, + 0.469, + 0.401 + ], + "angle": 0, + "content": "ASSANet [15]. The neighbor update is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.423, + 0.329, + 0.44 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\prime} = \\operatorname {M L P} _ {1} (f _ {i}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.442, + 0.374, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\mathcal {N} (i)} ^ {\\prime}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.464, + 0.336, + 0.482 + ], + "angle": 0, + "content": "\\[\nf _ {j} = \\operatorname {R e p e a t} _ {1} \\left(f _ {j} ^ {\\prime}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.469, + 0.547 + ], + "angle": 0, + "content": "where the operation \\(\\mathrm{Repeat}_1\\) means repeating the features three times to produce \\(f_{j}\\in \\mathbb{R}^{3d\\times 1}\\). The neighbor aggregation is implemented using an anistropic function, which is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.555, + 0.352, + 0.573 + ], + "angle": 0, + "content": "\\[\ne ^ {j} = \\operatorname {R e p e a t} _ {2} \\left(p _ {j} - p _ {i}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.575, + 0.362, + 0.613 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {(1)} = \\frac {1}{k} \\sum_ {j \\in \\mathcal {N} (i)} \\left(f _ {j} \\odot e _ {j}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.622, + 0.469, + 0.713 + ], + "angle": 0, + "content": "where the position embedding \\(e^j \\in \\mathbb{R}^{3d \\times 1}\\) is produced by simply repeating the relative coordinates \\(d\\) times. Note that, a similar position embedding and aggregation method are also adopted by PosPool [10], which is named position pooling. Similar with Point Transformer [32], the point feature is updated with an MLP layer and a shortcut." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.29, + 0.741 + ], + "angle": 0, + "content": "4. Explore Best Practices" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Given how many instantiations of PointMeta have appeared in the literature, we should push this general framework as far as possible in the architecture exploration to determine the most crucial implementation details. With the help of PointMeta, it is convenient to use quick experiments with variables controlled to verify any empirical observations or assumptions summarized from the comparison. In this section, we first compare and analyze the building blocks of existing models from the implementation details of these meta functions (Sec. 4.1, 4.2, 4.3, 4.4). Then," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "we summarize the best practices from the systematic analysis (Sec. 4.5). Finally, applying these practices, we do simple tweaks on the building blocks and propose a simple yet well-performing block, PointMetaBase (Sec. 4.6)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.162, + 0.669, + 0.178 + ], + "angle": 0, + "content": "4.1. Neighbor Update" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.185, + 0.891, + 0.215 + ], + "angle": 0, + "content": "According to the order of Group and MLP operation, the neighbor update function is categorized into two classes:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.225, + 0.757, + 0.241 + ], + "angle": 0, + "content": "- Group before MLP [7, 14, 16, 27]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.25, + 0.771, + 0.265 + ], + "angle": 0, + "content": "MLP before Group [6,8,10,15,32]." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.225, + 0.771, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.425 + ], + "angle": 0, + "content": "As pointed out by ASSANet [15], compared with the first order, adopting the second order reduces the FLOPs by \\(\\frac{d\\times d\\times N\\times K\\times L}{d\\times d\\times N\\times L} = K\\) times, where \\(N\\) denotes the point number, \\(K\\) denotes the neighbor number, \\(L\\) denotes the layer number of the MLP. Although applying MLP before Group brings a large efficiency gain for the neighbor update, a side effect is that the relative coordinates cannot be used as input for the MLP. However, this problem can be easily solved by other position embedding methods, which is specified in Sec. 4.2." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.436, + 0.866, + 0.562 + ], + "angle": 0, + "content": "
VariantmIoU (%)Params (M)FLOPs (G)
Plain-Max47.3±0.72.01.4
Plain-PP65.1±0.22.01.5
Plain-PP-Max58.4±0.62.01.5
Plain-IPE-Max68.0±0.32.013.8
Plain-EPE-Max69.0±0.32.01.8
Plain-EPE-PP65.4±0.12.01.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Table 1. Performance of the blocks with different position embedding functions and neighbor aggregation function on Area 5 of S3DIS [1]. Plain-PP-Max means multiplying the neighbor features with relative coordinates as PP but aggregating via max pooling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.657, + 0.692, + 0.673 + ], + "angle": 0, + "content": "4.2. Position Embedding" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.68, + 0.891, + 0.71 + ], + "angle": 0, + "content": "The manners of implementing the position embedding can be categorized into three classes:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.72, + 0.855, + 0.736 + ], + "angle": 0, + "content": "- Implicit position embedding (IPE) [7, 14, 16, 27]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.745, + 0.83, + 0.761 + ], + "angle": 0, + "content": "- Explicit position embedding (EPE) [6,8,32]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.77, + 0.736, + 0.785 + ], + "angle": 0, + "content": "- Position pooling (PP) [10, 15]." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.72, + 0.855, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implicit position embedding (IPE). Following PointNet++ [14], many methods [12, 16, 31] concatenate the relative coordinates with the neighbor features as inputs of the MLP layer in the neighbor update function. As going deeper through the networks, the dimension of the neighbor features become higher while the coordinates dimension still remains 3. This may cause the loss of position" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17685" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.137 + ], + "angle": 0, + "content": "information in the high-level features. In addition, IPE requires applying MLP before Group in the neighbor update function, which will increase the computation \\( K \\) times." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.47, + 0.38 + ], + "angle": 0, + "content": "Explicit position embedding (EPE). Incorporating explicit representations of position information is a common practice for ViT [11, 23, 28]. As mentioned in Sec. 3.2, Point Transformer [32] learn a position embedding \\( e^j \\) using \\( \\mathrm{MLP}_2 \\) and add it to the neighbor features \\( f_j \\). Compare with IPE, EPE avoid the dimension imbalance between the position representations and the neighbor features. Because the input dimension of \\( \\mathrm{MLP}_2 \\) is only 3, the computation of EPE becomes very small by controlling the dimension of the hidden layers. PointCNN [8] also uses an MLP to implement the EPE function, but it merges the EPE by concatenation instead of addition, which incurs a computation increase. StratifiedFormer [6] maintains a learnable quantized lookup table as EPE. In this section, we select the EPE of Point Transformer as the representative EPE method for the later experiments due to its simplicity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.501 + ], + "angle": 0, + "content": "Position pooling (PP). As mentioned in Sec. 3.2, AS-SANet [15] and PosPool [10] use relative coordinates as weights to pool neighbor features. This way can embed the position information into the neighbor features and make an aggregation function without learnable weights. The efficiency of position pooling is worth affirming, but embedding the position in a non-data-driven way may harm the learned neighbor features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.469, + 0.79 + ], + "angle": 0, + "content": "Empirical analysis. We conduct empirical experiments to evaluate the different PE functions described above. The first step is to construct a plain block as the baseline. Following the suggestion given in Sec. 4.1, the neighbor update function adopts the order of MLP-before-Group. The layer number of the MLP here is set to 1. The neighbor aggregation function is a simple max pooling. The point update function is a 1-layer MLP. We denote this block as PlainMax. We equip Plain-Max with the three kinds of position embedding described above and compare their performance on Area 5 of S3DIS [1]. The MLP used respectively in EPE and IPE are both 1-layer. The macro-architecture is in level \"L\" (described in Sec. 4.6). As shown in Tab. 1, PlainEPE-Max achieves the best performance with a tiny computation increase against Plain-PP, Plain-PP-Max and Plain-IPE-Max. Because coupling the neighbor features transformation with position embedding, Plain-IPE incurs heavy redundant computation. In a summary, EPE is the optimal choice among the three kinds of position embeddings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.803, + 0.285, + 0.819 + ], + "angle": 0, + "content": "4.3. Neighbor Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.469, + 0.873 + ], + "angle": 0, + "content": "According to whether learnable weights exist in the neighbor aggregation function, we divide the aggregations function into two types:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.407, + 0.901 + ], + "angle": 0, + "content": "Non-learnable aggregation [7, 10, 14-16, 27]." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.905, + 0.155 + ], + "angle": 0, + "content": "
VariantsmIoU (%)Params (M)FLOPs (G)
Point Trans [32]70.5±0.37.85.6
Point Trans (-VSA+Max)70.3±0.25.13.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.164, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Table 2. Ablation of the vector self-attention of Point Transformer [32] on Area 5 of S3DIS [1]. -VSA+Max means replacing VSA with max pooling." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.232, + 0.792, + 0.248 + ], + "angle": 0, + "content": "- Learnable aggregation [6,8,22,29,32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.392 + ], + "angle": 0, + "content": "Learnable aggregation. As mentioned in Sec. 3.1, the aggregation function should satisfy permutation invariance because the point cloud is unordered. The permutation invariance is acquired by learning a space transformation [8, 29], learning pointwise weights [32], or pairwise self-attention [6]. In this section, we select the vector self-attention of Point Transformer [32] as the representative implementation of the learnable aggregation function due to its excellent performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Non-learnable aggregation. There are two types of non-learnable aggregation in the literatures: position pooling [10, 15] and max pooling [7, 14, 16, 27]. As mentioned in Sec. 4.2, position pooling embed the position in a nondata-driven way and may compete with the learned neighbor features. As shown in Tab. 1, Plain-EPE-Max surpasses Plain-EPE-PP significantly, which supports our assumption. Therefore, we select the max pooling as the representative non-learnable aggregation function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.619 + ], + "angle": 0, + "content": "Empirical analysis. We conduct an ablation experiment on Point Transformer [32] to compare the aggregation power of both max pooling and vector self-attention. Specifically, we replace the VSA with a max pooling operation (denoted as -VSA+Max). As a result, the aggregation of the ablated Point Transformer is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.63, + 0.77, + 0.647 + ], + "angle": 0, + "content": "\\[\ne ^ {j} = \\operatorname {M L P} _ {2} \\left(p _ {j} - p _ {i}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.671, + 0.891, + 0.69 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {(1)} = \\operatorname {M a x} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)} + \\boldsymbol {e} _ {\\mathcal {N} (i)}\\right), j \\in \\mathcal {N} (i), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.831 + ], + "angle": 0, + "content": "where \\( e_{\\mathcal{N}(i)} = \\{e_j | j \\in \\mathcal{N}(i)\\} \\). As shown in Tab. 2, compared with VSA, max pooling achieves comparable performance with much fewer parameters and FLOPs. Here comes a question, why max pooling has a comparable power with the learnable aggregation? We explore the answer within our PointMeta framework and find that the max pooling is a special case of VSA but with binary and sparse attention. Specifically, the aggregation function Eq. 4 can be reformulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.844, + 0.789, + 0.861 + ], + "angle": 0, + "content": "\\[\nM _ {j} = \\operatorname {S o f t m a x} _ {T \\rightarrow 0} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.864, + 0.804, + 0.897 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {(1)} = \\sum_ {j \\in \\mathcal {N} (i)} \\left(M _ {j} \\odot \\left(f _ {j} + e _ {j}\\right)\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17686" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.422, + 0.258 + ], + "angle": 0, + "content": "
VariantsmIoU (%)Params (M)FLOPs (G)
N1P169.0±0.32.01.7
N2P068.7±0.32.01.7
N0P267.8±0.42.01.7
N1P269.5±0.32.72.0
N1P2-Inv69.7±0.37.13.8
N2P169.4±0.32.72.0
N2P1-Inv69.6±0.47.13.8
N1P369.5±0.33.42.3
N3P169.6±0.53.42.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.27, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Table 3. Semantic segmentation of the Plain-EPE with different MLP configuration on Area 5 of S3DIS [1]. \"N1P2\" means that the MLP in the neighbor update is 1-layer and 2-layer. \"Inv\" means the inverted bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.468, + 0.458 + ], + "angle": 0, + "content": "where \\(\\mathrm{Softmax}_{T\\to 0}\\) denotes a softmax function whose temperature approximates 0, which makes the softmax output \\(M_{j}\\) approximates the binary masks. This formulation reveals the collaboration between max pooling and the neighbor update function, i.e., max pooling is an aggregation function with binary and sparse attention which is generated by applying softmax on the output of the neighbor update" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.216, + 0.484 + ], + "angle": 0, + "content": "4.4. Point Update" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.626 + ], + "angle": 0, + "content": "The point update function is used to enhance the point features after the neighbor aggregation. Several existing studies do not implement this function in their building blocks [7, 27], and some other studies use a heavy point update function, such as a 2-layer MLP with an inverted bottleneck [6, 16]. A natural question arises: how much computation complexity should we allocate for the point update function? We answer this question via extensive ablated experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.468, + 0.837 + ], + "angle": 0, + "content": "Empirical analysis. According to the exploration above, we adopt the Plain-EPE-Max as the baseline due to its excellent performance in the ablated experiments (Tab. 1). Since the computation are concentrated on the MLP of the neighbor update function and point update function, we can allocate computation between the two update functions by changing the configurations of the two MLP. We ablate the MLP from two aspects, i.e., the layer number and whether to use an inverted bottleneck that adopted by PointNeXt [16] and StratifiedFormer [6]. As shown in Tab. 3, the configuration \"N1P2\" can achieve an optimal balance between complexity and performance. More layers or inverted bottlenecks can only promote the model marginally but lead to much more computation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.221, + 0.862 + ], + "angle": 0, + "content": "4.5. Best Practices" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Here we summarize the best practices for the building block as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.891, + 0.15 + ], + "angle": 0, + "content": "- **Position Embedding.** Explicit position embedding can endow the models with better position awareness with a low computation cost comparing with other PE manners." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.162, + 0.891, + 0.208 + ], + "angle": 0, + "content": "- Point Update. Point update is important and the complexity allocation between the neighbor update and the point update should be properly set." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.218, + 0.891, + 0.308 + ], + "angle": 0, + "content": "- Neighbor Aggregation. Max pooling is not a pure non-learnable aggregation function, its aggregation power is comparable with the learnable aggregation function but with much less computation complexity. Any exploration for a new aggregation function should set max pooling as the baseline." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.319, + 0.891, + 0.394 + ], + "angle": 0, + "content": "- Neighbor Update. Applying the MLP before neighbors grouping will bring a \\(K\\) times FLOPs reduction comparing with using the contrary order. However, the position embedding function must be carefully chosen for the position awareness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.631 + ], + "angle": 0, + "content": "The first and the second practice are usually ignored by researchers but are vital for the effectiveness and efficiency of the building block, which is demonstrated in Sec. 4.2 and Sec. 4.4. The third shares a similar viewpoint with PosPool [10] which claims that different local aggregation operators perform similarly. However, from the perspective of PointMeta, we provide a deeper insight that max pooling can be reformulated as a special case of vector self-attention [32], which makes it have comparable aggregation power comparing with the learnable aggregation function. The fourth practice about neighbor update has been emphasized by ASSANet [15], we restate it here because it is important to notice its collaboration with position embedding function. We believe that these practices can provide a guideline for the future research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.641, + 0.655, + 0.656 + ], + "angle": 0, + "content": "4.6. PointMetaBase" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.875 + ], + "angle": 0, + "content": "We apply the best practices summarized in Sec. 4.5 to design the building block and implement the four meta functions as simply as possible. We name this block as PointMetaBase, which we hope to become a solid baseline for future research. Fig. 2 (g) shows the architecture of PointMetaBase. The layer number of the first MLP is set to 1 and that of the second MLP is set to 2. The neighbor aggregation function is a simple max pooling. We use the EPE adopted by Point Transformer [32] as the position embedding function. Applying EPE and the MLP-before-Group order, we tweak the set abstraction module [14, 16] as the reduction block, termed PointMetaSA. We adopt the same scaling strategies and decoder with PointNeXt [16] to construct our PointMetaBase family:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.766, + 0.9 + ], + "angle": 0, + "content": "- PointMetaBase-S: \\(C = 32\\), \\(B = 0\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17687" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.089, + 0.885, + 0.403 + ], + "angle": 0, + "content": "
MethodS3DIS 6-FoldS3DIS Area-5ScanNet mIoUParams. MFLOPs GThroughput (ins./sec.)
mIoU (%)OA (%)mIoU (%)OA (%)Val (%)Test (%)
PointNet++ [14]54.581.053.583.053.555.71.07.2237
PointCNN [8]65.488.157.385.9-45.80.6--
DeepGCN [7]60.085.952.5---3.6--
KPConv [22]70.6-67.1-69.268.615.0--
RandLA-Net [5]70.088.0---64.51.35.8-
BAAF-Net [17]72.288.965.488.9--5.0--
Point Transformer [32]73.590.270.490.870.6-7.85.6-
CBL [20]73.189.669.490.6-70.518.6--
ASSANet [15]--65.888.9--2.42.5300
ASSANet-L [15]--68.089.7--115.636.2136
PointNeXt-S [16]68.087.463.4±0.887.9±0.364.5-0.83.6276
PointNeXt-B [16]71.588.867.3±0.289.4±0.168.4-3.88.9161
PointNeXt-L [16]73.989.869.0±0.590.0±0.169.4-7.115.2109
PointNeXt-XL [16]74.990.370.5±0.390.6±0.171.571.241.684.843
PointMetaBase-L75.690.669.5±0.390.5±0.171.0-2.72.0187
PointMetaBase-XL76.391.071.1±0.490.9±0.171.8-15.39.2104
PointMetaBase-XXL77.091.371.3±0.790.8±0.672.871.419.711.090
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.413, + 0.893, + 0.444 + ], + "angle": 0, + "content": "Table 4. Semantic segmentation on S3DIS [1] (6-Fold and Area 5) and ScanNet V2 [3]. Our PointMetaBase family surpass the state-of-the-art method PointNeXt [16] significantly with large FLOPs reduction." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.468, + 0.404, + 0.484 + ], + "angle": 0, + "content": "- PointMetaBase-L: \\(C = 32\\), \\(B = (2,4,2,2)\\)" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.497, + 0.415, + 0.513 + ], + "angle": 0, + "content": "- PointMetaBase-XL: \\(C = 64\\), \\(B = (3,6,3,3)\\)" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.526, + 0.427, + 0.542 + ], + "angle": 0, + "content": "- PointMetaBase-XXL: \\( C = 64 \\), \\( B = (4,8,4,4) \\)" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.468, + 0.427, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.692 + ], + "angle": 0, + "content": "\\(C\\) denotes the channel size of the stem MLP and \\(B\\) denotes the number of the PointMetaBase block in each stage. Note that when \\(B = 0\\), only one PointMetaSA block but no PointMetaBase blocks are used at each stage. Due to the excellent efficiency of PointMetaBase, we do not need to construct the network at the level \"B\" as done in Point-NeXt [16]. In contrast, we choose to scale the model to a larger level \"XXL\". The full details of network architecture are available in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.21, + 0.725 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.809 + ], + "angle": 0, + "content": "We studied the performance of PointMetaBase on S3DIS [1], ScanNet V2 [3], ScanObjectNN [24] and ShapeNetPart [30]. To enable a fair comparison, the same data processing and evaluation protocols adopted by the state-of-the-art method PointNeXt [16] were used in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Experimental Setups. We optimized all our models using CrossEntropy loss with label smoothing [19], AdamW optimizer, an initial learning rate 0.001, and weight decay \\(10^{-4}\\), with Cosine Decay, with a 32G V100 GPU. Model parameters, FLOPs and inference throughput are provided for comparison. For a fair comparison, we use the open" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.893, + 0.545 + ], + "angle": 0, + "content": "codes provided by PointNext [16] to conduct the measurements. We also use MindSpore to validate the generalization on different deep learning frameworks. All the measurements are conducted using an NVIDIA Tesla A100 40GB GPU and a 12-core Intel Xeon @ 2.40GHz CPU." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.553, + 0.719, + 0.57 + ], + "angle": 0, + "content": "5.1.3D Scene Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.895, + 0.88 + ], + "angle": 0, + "content": "S3DIS [1] is a challenging benchmark composed of 6 large-scale indoor areas, 271 rooms, and 13 semantic categories in total. The standard 6-fold cross-validation results and the Area-5 results on S3DIS are reported in Tab. 4. Since we adopt the same scaling strategies, PointMetaBase-L and PointMetaBase-XL have the same configuration on width and depth as PointNeXt-L and PointNeXt-XL. With only \\(13\\%\\) FLOPs, our PointMetaBase-L outperforms PointNext-L by \\(1.7\\%\\) and \\(0.8\\%\\) in terms of mean IoU (mIoU) and overall accuracy (OA), respectively, and is much faster in terms of throughput. Compared with PointNext-XL, our PointMetaBase-XL uses only \\(11\\%\\) of the FLOPs, but surpasses with \\(1.4\\%\\) mIoU and \\(0.7\\%\\) OA. The inference speed is \\(2.4\\times\\) faster. The excellent efficiency and performance of PointMetaBase can be attributed to the the removal of the computation redundancy in the two update functions and the improved position awareness introduced by explicit position embedding. ScanNet [3] contains 3D indoor scenes of various rooms with 20 semantic categories. The dataset is divided into 3 part: training," + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.761, + 0.902 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/guochengqian/PointNeXt" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17688" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.813, + 0.306 + ], + "angle": 0, + "content": "
Methodins. mIoU (%)cls. mIoU (%)Params (M)FLOPs (G)Throughput (ins./sec)
PointNet++ [14]85.181.91.04.9-
KPConv [22]86.485.1---
ASSANet [15]86.1----
Point Transformer [32]86.683.77.8--
PointMLP [12]86.184.6---
StratifiedFormer [6]86.685.1---
PointNeXt-S [16]86.7±0.084.4±0.21.04.5873
PointNeXt-S (C=64) [16]86.9±0.184.8±0.53.717.8391
PointNeXt-S (C=160) [16]87.0±0.185.2±0.122.5110.2118
PointMetaBase-S (ours)86.7±0.084.3±0.11.01.391194
PointMetaBase-S (C=64) (ours)86.9±0.184.9±0.23.83.85706
PointMetaBase-S (C=160) (ours)87.1±0.085.1±0.322.718.45271
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.316, + 0.893, + 0.345 + ], + "angle": 0, + "content": "Table 5. Part segmentation on ShapeNetPart [30]. Compaerd with the state-of-the-art method PointNeXt [16], our PointMetaBase family achieve comparable performance with large FLOPs reduction and faster inference speed." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.368, + 0.459, + 0.516 + ], + "angle": 0, + "content": "
VariantsOA (%)FLOPs (G)TP (ins./sec)
PointNet++ [14]77.91.72655
PointCNN [8]78.5--
DGCNN [27]78.14.8735
DRNet [13]80.3--
PointMLP [12]85.4±1.331.3405
PointNeXt-S [16]87.7±0.41.72230
PointMetaBase-S87.9±0.20.62674
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Table 6. Classification on ScanObjectNN [24]. Our model surpasses PointNeXt-S [16] both with much less computation cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.687 + ], + "angle": 0, + "content": "validation, and test splits, with 1201, 312 and 100 scans, respectively. Following PointNeXt [16], we do not use any voting strategies. As shown in Tab. 4, our PointMetaBase-L surpasses PointNeXt-L \\(1.6\\%\\) mIoU with only \\(14\\%\\) FLOPs and \\(1.7\\times\\) faster inference speed. PointMetaBase-XL surpasses PointNeXt-XL \\(0.3\\%\\) mIoU with only \\(11.3\\%\\) FLOPs and \\(2.4\\times\\) faster inference speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.303, + 0.71 + ], + "angle": 0, + "content": "5.2. 3D Object Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.471, + 0.84 + ], + "angle": 0, + "content": "ScanObjectNN [24] contains about 15000 real scanning objects, which are grouped into 15 classes and have 2902 unique instances. Following [12, 16], we experiment on the hardest perturbed variant (PB_T50_RS). To make a fair comparison with PointNeXt [16], we did not use upscaled variants of PointMetaBase on this benchmark. As reported in Tab. 6, our model surpasses PointNeXt-S, while using much fewer model FLOPs and running much faster." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.285, + 0.864 + ], + "angle": 0, + "content": "5.3. 3D Part Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "ShapeNetPart [30] is a dataset for part segmentation of objects. It consists of 16880 models with 16 differ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.895, + 0.522 + ], + "angle": 0, + "content": "ent shape categories. Following the state-of-the-art method PointNeXt [16], we used voting by averaging the results of 10 randomly scaled input point clouds, with scaling factors equal to [0.8,1.2]. As shown in Tab. 5, compaerd with PointNeXt, our PointMetaBase family achieve comparable performance with large FLOPs reduction and faster inference speed. Specially, for \\( C = 32 / C = 64 / C = 160 \\), the FLOPs of our PointMetaBase are only \\( 30\\% 21\\% / 17\\% \\) that of PointNeXt, and the inference speed of PointMetaBase is \\( 1.4 / 1.8 / 2.3 \\) times faster than that of PointNeXt." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.62, + 0.557 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.894, + 0.734 + ], + "angle": 0, + "content": "In this paper, we reformulate existing models for point cloud analysis into a general framework, PointMeta. Within this framework, we provide an in-depth analysis on the building blocks of existing models and summarize a few best practices. Furthermore, we follow these best practices and propose a simple building block, PointMetaBase, which enjoys excellent performance and efficiency. Our work can inspire new insights into building block design and future architecture exploration. Next, building on this work, we plan to explore neural architecture search methods [33-37] that are specifically tailored for point cloud analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by National Key R&D Program of China (No. 2022ZD0118202), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J01002, No. 2022J06001), Guangdong Basic and Applied Basic Research Foundation (No. 2019B1515120049), and CAAI-Huawei MindSpore Open Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17689" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.119, + 0.174, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.468, + 0.185 + ], + "angle": 0, + "content": "[1] Iro Armeni, Sasha Sax, Amir R. Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. CoRR, abs/1702.01105, 2017. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.324 + ], + "angle": 0, + "content": "[2] Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Flores Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gulçehre, H. Francis Song, Andrew J. Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, Kelsey R. Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matthew M. Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational inductive biases, deep learning, and graph networks. CoRR, abs/1806.01261, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.468, + 0.393 + ], + "angle": 0, + "content": "[3] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 2432–2443. IEEE Computer Society, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.468, + 0.437 + ], + "angle": 0, + "content": "[4] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[5] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11105-11114, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.468, + 0.548 + ], + "angle": 0, + "content": "[6] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8490-8499. IEEE, 2022. 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.468, + 0.591 + ], + "angle": 0, + "content": "[7] Guohao Li, Matthias Müller, Ali K. Thabet, and Bernard Ghanem. Deep GCs: Can GCs go as deep as cnns? In ICCV, pages 9266-9275, 2019. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.468, + 0.634 + ], + "angle": 0, + "content": "[8] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, pages 828-838, 2018. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[9] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In CVPR, pages 8895-8904, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[10] Ze Liu, Han Hu, Yue Cao, Zheng Zhang, and Xin Tong. A closer look at local aggregation operators in point cloud analysis. In ECCV, pages 326-342, 2020. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[11] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[12] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[13] Charles Ruizhongtai Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 77-85, 2017. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.144, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[15] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali K. Thabet, and Bernard Ghanem. Assanet: An anisotropic separable set abstraction for efficient point cloud representation learning. In NeurIPS, pages 28119-28130, 2021. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[16] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Abed Al Kader Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.266, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[17] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[18] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, ICLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.354, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[19] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.411, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[20] Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. In CVPR, pages 8479-8489, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[21] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, pages 3887-3896, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.499, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[22] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6410–6419, 2019. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.557, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[23] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.615, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[24] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, pages 1588–1597, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[25] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, pages 10296-10305, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[26] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In CVPR, pages 2589-2597, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. ACM Trans. Graph., 38(5):146:1-146:12, 2019. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In ICCV, pages 10013-10021, 2021. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17690" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[29] Wenxuan Wu, Zhongang Qi, and Fuxin Li. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, pages 9621-9630, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[30] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Trans. Graph., 35(6):210:1-210:12, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[31] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, pages 5565-5573, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[32] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16239-16248, 2021. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[33] Xiawu Zheng, Xiang Fei, Lei Zhang, Chenglin Wu, Fei Chao, Jianzhuang Liu, Wei Zeng, Yonghong Tian, and Rongrong Ji. Neural architecture search with representation mutual information. In CVPR, pages 11912-11921, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.429 + ], + "angle": 0, + "content": "[34] Xiawu Zheng, Rongrong Ji, Yuhang Chen, Qiang Wang, Baochang Zhang, Jie Chen, Qixiang Ye, Feiyue Huang, and Yonghong Tian. Migo-nas: Towards fast and generalizable neural architecture search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9):2936-2952, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.468, + 0.486 + ], + "angle": 0, + "content": "[35] Xiawu Zheng, Rongrong Ji, Lang Tang, Baochang Zhang, Jianzhuang Liu, and Qi Tian. Multinomial distribution learning for effective neural architecture search. In ICCV, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.489, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[36] Xiawu Zheng, Rongrong Ji, Qiang Wang, Qixiang Ye, Zhenguo Li, Yonghong Tian, and Qi Tian. Rethinking performance estimation in neural architecture search. In CVPR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.546, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[37] Xiawu Zheng, Yang Zhang, Sirui Hong, Huixia Li, Lang Tang, Youcheng Xiong, Jin Zhou, Yan Wang, Xiaoshuai Sun, Pengfei Zhu, Chenglin Wu, and Rongrong Ji. Evolving fully automated machine learning via life-long knowledge anchors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9):3091-3107, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17691" + } + ] +] \ No newline at end of file diff --git a/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_origin.pdf b/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ef8c68606a9e9cd77fc3f6bebfcdb3b952077ec9 --- /dev/null +++ b/2023/Meta Architecture for Point Cloud Analysis/ef212f38-9097-46a0-9adc-735684e16116_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56ad868fef04e81f05b647a0268f6e017619c09501e58b589b5f084d7127796e +size 338429 diff --git a/2023/Meta Architecture for Point Cloud Analysis/full.md b/2023/Meta Architecture for Point Cloud Analysis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4ad847a5d5f2b0e41bf445198f3c7d1802394f42 --- /dev/null +++ b/2023/Meta Architecture for Point Cloud Analysis/full.md @@ -0,0 +1,367 @@ +# Meta Architecture for Point Cloud Analysis + +Haojia Lin $^{1}$ , Xiawu Zheng $^{2}$ , Lijiang Li $^{1}$ , Fei Chao $^{1}$ , Shanshan Wang $^{2,3}$ , Yan Wang $^{4}$ , Yonghong Tian $^{2,5}$ , Rongrong Ji $^{1,2,6*}$ + +1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, School of Informatics, Xiamen University, 361005, P.R. China. + +$^{2}$ Peng Cheng Laboratory. $^{3}$ Chinese Academy of Sciences. $^{4}$ Samsara Inc. + +$^{5}$ National Engineering Research Center for Visual Technology, Peking University. $^{6}$ Shenzhen Research Institute of Xiamen University. + +{linhj, lilijiang}@stu.xmu.edu.cn, zhengxw01@pcl.ac.cn, {rrji, feichao}@xmu.edu.cn + +ss.wang@siat.ac.cn, yan.wang@samsara.com, yhtian@pku.edu.cn + +# Abstract + +Recent advances in 3D point cloud analysis bring a diverse set of network architectures to the field. However, the lack of a unified framework to interpret those networks makes any systematic comparison, contrast, or analysis challenging, and practically limits healthy development of the field. In this paper, we take the initiative to explore and propose a unified framework called PointMeta, to which the popular 3D point cloud analysis approaches could fit. This brings three benefits. First, it allows us to compare different approaches in a fair manner, and use quick experiments to verify any empirical observations or assumptions summarized from the comparison. Second, the big picture brought by PointMeta enables us to think across different components, and revisit common beliefs and key design decisions made by the popular approaches. Third, based on the learnings from the previous two analyses, by doing simple tweaks on the existing approaches, we are able to derive a basic building block, termed PointMetaBase. It shows very strong performance in efficiency and effectiveness through extensive experiments on challenging benchmarks, and thus verifies the necessity and benefits of high-level interpretation, contrast, and comparison like PointMeta. In particular, PointMetaBase surpasses the previous state-of-the-art method by $0.7\% /1.4\% 2.1\%$ mIoU with only $2\% /11\% /13\%$ of the computation cost on the S3DIS datasets. The code and models are available at https://github.com/linhaojia13/PointMetaBase. + +![](images/2770bbcbf68355d70490c1ac7cd4a1976b518ec991a73ee3e7a8829f01a3532b.jpg) +Figure 1. Segmentation performance of PointMetaBase on S3DIS [1]. PointMetaBase surpasses the state-of-the-art method PointNeXt [16] significantly with a large FLOPs reduction. + +# 1. Introduction + +In the past two decades, the popularity of 3D data acquisition technology has led to great interest in point cloud analysis. Unlike images, point clouds are inherently sparse, unordered, and irregular. These characteristics make it challenging for the powerful convolutional neural networks (CNN) [4, 18] to extract useful information from point clouds. Early studies attempt to convert the point cloud into grids by either voxelization or 2D projections, such that the standard CNN can be applied, but the applications are limited by the extra computation and information loss. + +With the introduction of permutation equivariance and invariance by PointNet [13] and PointNet++ [14], it is possible to use CNNs to process point clouds in their unstruc + +tured format. Following PointNet++ [14], many point-based networks have been introduced, most of which focus on the development of complex building blocks to extract local features, such as the $\mathcal{X}$ - Conv convolution in PointCNN [22] and the self-attention layer in Point Transformer [32]. Although brought significant performance gains, these models are very complicated. On one hand, this makes the approaches computational demanding and thus limited in applications. On the other hand, it also gives extra challenges in systematic comparison and analysis among the models, and thus impacts efficient development of the field, which ideally would benefit from theoretic and empirical guidance. + +The community is well aware of this issue, and research efforts have been put on looking for a unified perspective to compare these methods and identify the most crucial implementation details. PosPool [10] and PointNeXt [16] take a step toward this goal. Adopting a deep residual architecture as the base network, PosPool [10] evaluates the representative building blocks [8, 14, 22, 26] and finds that they perform similarly well. PointNeXt [16] revisits the improved training and scaling strategies adopted by previous SotA methods [7, 15, 22, 32] and tweaks the early model PointNet++ [14] with the learnings, which achieves state-of-the-art performance. Both of them give insights and inspirations to the community via dedicated experiment designs and extensive empirical analysis. However, neither of them provides a framework that is general enough to fit a broad range of point cloud analysis approaches, such that large scale comparison and contrast could be done. + +In this paper, we argue that if viewed from the perspective of basic building blocks, the majority of existing approaches could be fit into a single meta architecture (Sec. 3). We call it PointMeta. PointMeta abstracts the computation pipeline of the building blocks for point cloud analysis into four meta functions: a neighbor update function, a neighbor aggregation function, a point update function and a position embedding function (implicit or explicit). As will be discussed in more details in the next sections, this framework brings three benefits, which are also our core technical contributions: + +- In the dimension of models, it allows us to compare and contrast different models in a fair manner. So it becomes practical to observe and summarize learnings and assumptions, whose correctness could be further verified through experiments with variables controlled under the same framework. For example, among the systematic analysis on all the components across popular models, we found for the positional embedding function, explicit positional embedding is empirically the best choice (Sec. 4.2). +- In the dimension of components, it allows us to have + +a higher level view across components, and thus to revisit the common beliefs and design decisions of the existing approaches. For example, despite the common perception, we find that the neighbor aggregation blocks may collaborate or compete with the learned neighbor features (Sec. 4.3), and thus should be designed carefully. + +- Based on the learnings from the previous two dimensions, we are then able to do simple tweaks on the building blocks to apply the best practices (Sec. 4.5). The result building block, PointMetaBase, achieves very strong performance, surpassing the previous state-of-the-art method [16] by $0.7\% / 1.4\% / 2.1\%$ mIoU with only $2\% / 11\% / 13\%$ of the computation cost on S3DIS [1] (Fig. 1), and can act as a baseline for further research. + +# 2. Related Work + +Neural Networks for Point Cloud Analysis. The unstructured data format of point cloud makes it difficult to apply CNN directly. To address this challenge, PointNet [13] introduces permutation equivariance and invariance via point-wise MLP and max pooling, which makes it possible to process point cloud directly. To better capture locality, PointNet++ [14] presents a Set Abstraction module to aggregate features from the neighbor points. Most works after PointNet++ focus on the design of build blocks. Graph-based methods [7, 25, 27] utilize graph neural networks to extract local features. Pseudo-grid-based methods [8,21,22] transform the local features onto pseudo grids which allow for regular convolutions. Adaptive weight-based methods [5,9] aggregate the neighbor features via weights adaptively learned from the locality. Recently, Transformer-based block [6, 32] are introduced to capture local information through self-attention. These proposed building blocks are so complicated that make challenges in systematic comparison and analysis, which suggests an urgent need of a common framework to unify existing models in this field. + +# 3. PointMeta + +# 3.1. Meta Architecture + +Inspired by the graph network framework [2], we reformulate existing building blocks for point cloud analysis into a single meta architecture, named PointMeta. As shown in Fig. 2 (a), a PointMeta block is composed of four meta functions: a neighbor update function $\phi^n$ , a neighbor aggregation function $\phi^a$ , a point update function $\phi^p$ and a position embedding function $\phi^e$ (implicit or explicit). The forward pass of a PointMeta block is as follows: + +$$ +\boldsymbol {f} _ {\mathcal {N} (i)} = \phi^ {n} \circ \phi^ {e} \left(f _ {i}, p _ {i}\right), \tag {1} +$$ + +![](images/41d674684abf894d1a8083199d7c3e93452c04390476ce8670060651356bd6d1.jpg) +(a) PointMeta + +![](images/963b7cc4839de276bd2d8d73ed9e1838cbff50ba68b1e4b21dd87b497897bf39.jpg) +(b) PointNet++ + +![](images/88b61868aae50f225cc781fa976ffb589a02a22bf569464011831c6616a699eb.jpg) +(c) DGCNN + +![](images/b6ebcad888433fced510b30228249046d70ff071a2ce60cdb5b18b29f5c5cd56.jpg) +(d) PointCNN +Figure 2. PointMeta and its instantiation examples. (a) In PointMeta, the position embedding function is usually combined with the neighbor update function or the aggregation function implicitly or explicitly. (b) $\sim$ (f) Representative building blocks can naturally fit into PointMeta. (g) Applying the summarized best practices, we do simple tweaks on the building blocks and propose PointMetaBase. + +![](images/902577f248e47bcfc2d08ae786f41f008e2e50dfb0cd7bb4c05ca083eab86704.jpg) +(e) Point Transformer + +![](images/2e85920323a7b4a9fac7d04d08f53de16d208aa099da8c7a952a389be601048a.jpg) +(f) ASSANet + +![](images/46995296ac8f64a4b113e7208b8f058ce2d77058ec41277de09fb2158561b9d2.jpg) +(g) PointMetaBase + +$$ +f _ {i} ^ {(1)} = \phi^ {a} \circ \phi^ {e} \left(f _ {i}, p _ {i}, \mathbf {f} _ {\mathcal {N} (i)}, \mathbf {p} _ {\mathcal {N} (i)}\right) \tag {2} +$$ + +$$ +f _ {i} ^ {(2)} = \phi^ {p} \left(f _ {i} ^ {(1)}, p _ {i}\right), \tag {3} +$$ + +where $f_{i} \in \mathbb{R}^{d}$ and $p_i \in \mathbb{R}^3$ represent the feature and the coordinate of point $i$ respectively, and $d$ denotes the feature length. Let $\mathcal{N}(i)$ denote the neighbors of point $i$ , $f_{j}$ and $p_j$ be the feature and coordinate of each neighbor of point $i$ respectively, $\pmb{f}_{\mathcal{N}(i)} = \{f_j | j \in \mathcal{N}(i)\}$ represent all neighbor features of point $i$ , $\pmb{p}_{\mathcal{N}(i)} = \{p_j | j \in \mathcal{N}(i)\}$ represent all neighbor coordinates of point $i$ , and $k$ denote the neighbor number. + +Neighbor update. The first step is to group and update neighbors for each point. Usually, the neighbor features $f_{\mathcal{N}(i)}$ is produced by K-Nearest query as [27, 32] or ball query as [14-16]. The update on neighbor features is usually conducted by a multi-layer perceptions (MLP) to make sure the permutation equivariance of the neighbor update function $\phi^n$ . Specifically, a permutation transformation on $f_{\mathcal{N}(i)}$ results in the same permutation change on the output of the neighbor update. + +Neighbor aggregation. Literally, the neighbor aggregation function is used to aggregate the neighbor features for a center point. Different from $\phi^n$ , the aggregation function $\phi^a$ must satisfy permutation invariance, i.e., a permutation transformation on the input $f_{\mathcal{N}(i)}$ will have no influence on the output $f_i^{(1)}$ . + +Point update. The point update usually contains an MLP, which is used to further enhance the point features. + +Position embedding. Incorporating representations of position information via position embedding is a common practice for Vision Transformer (ViT) [11, 23, 28], since ViT models must capture spatial relationships of tokens. We thus introduce the concept of position embedding into PointMeta. Within a point cloud, points are organized in + +an unstructured format, however the points are highly correlated in the 3D Euclidean space. Therefore, the building block of point cloud analysis should possess the position awareness, which is provided by the position embedding function $\phi^e$ . The function $\phi^e$ is usually combined with the neighbor update function $\phi^n$ or the aggregation function $A$ implicitly or explicitly, thus we present them with an bundled notation, $\phi^n\circ \phi^e$ or $\phi^a\circ \phi^e$ . + +# 3.2. Instantiation Examples + +In this section, we select three representative models from the literature as the instantiation examples and explain how these models implement the meta functions within PointMeta. We illustrate more instantiation examples [5, 8, 16, 22, 27, 29] in the supplementary due to the limited space. Sometimes these meta functions are implemented sophisticatedly, we will decompose them into several subfunctions for the sake of brevity and readability. + +PointNet++ [14]. The neighbor update is described as follows: + +$$ +\boldsymbol {f} _ {\mathcal {N} (i)} ^ {\prime}, \boldsymbol {p} _ {\mathcal {N} (i)} = \operatorname {G r o u p} \left(f _ {i}, p _ {i}\right), +$$ + +$$ +f _ {j} = \operatorname {M L P} _ {1} \left(\operatorname {C o n c a t} \left(f _ {j} ^ {\prime}, p _ {j} - p _ {i}\right)\right), +$$ + +where $\mathrm{MLP}_1$ denotes the multi-layer perceptions applied on the concatenated neighbor features. The neighbor aggregation is implemented using a simple max pooling operation, + +$$ +f _ {i} ^ {(1)} = \operatorname {M a x} \left(\boldsymbol {f} _ {\mathcal {N} (i)}\right). +$$ + +The point update function is another MLP layer, thus we have + +$$ +f _ {i} ^ {(2)} = \mathrm {M L P} _ {2} (f _ {i}) +$$ + +Note that the position embedding function $\phi_{e}$ is implicitly implemented by $\mathrm{MLP}_1$ , which incorporates the concatenation of the relative positions $p_j - p_i$ and grouping features $f_j'$ as input and returns position-aware neighbor features $f_j$ . + +Point Transformer [32]. The neighbor update is described as follows: + +$$ +f _ {i} ^ {\prime} = \operatorname {M L P} _ {1} (f _ {i}), +$$ + +$$ +\boldsymbol {f} _ {\mathcal {N} (i)}, \boldsymbol {p} _ {\mathcal {N} (i)} = \operatorname {G r o u p} \left(f _ {i}, p _ {i}\right). +$$ + +The neighbor aggregation is implemented using vector self-attention, which is described as + +$$ +\begin{array}{l} e ^ {j} = \operatorname {M L P} _ {2} \left(p _ {j} - p _ {i}\right), \\ M _ {j} = \operatorname {S o f t m a x} \left(\mathrm {M L P} _ {5} \left(\mathrm {M L P} _ {3} \left(f _ {i}\right) - \mathrm {M L P} _ {4} \left(f _ {j}\right) + e _ {j}\right)\right), \\ f _ {i} ^ {(1)} = \sum_ {j \in \mathcal {N} (i)} \left(M _ {j} \odot \left(f _ {j} + e _ {j}\right)\right), \\ \end{array} +$$ + +where the $\odot$ denotes the Hadamard product and $M\in \mathbb{R}^{k\times d}$ denotes the vector self-attention (VSA) mask. Finally, the point feature is updated with an MLP layer and a shortcut, then we have + +$$ +f _ {i} ^ {(2)} = \mathrm {M L P} _ {6} \left(f _ {i} ^ {(1)}\right) + f _ {i} +$$ + +ASSANet [15]. The neighbor update is described as: + +$$ +f _ {i} ^ {\prime} = \operatorname {M L P} _ {1} (f _ {i}), +$$ + +$$ +\boldsymbol {f} _ {\mathcal {N} (i)} ^ {\prime}, \boldsymbol {p} _ {\mathcal {N} (i)} = \operatorname {G r o u p} \left(f _ {i}, p _ {i}\right), +$$ + +$$ +f _ {j} = \operatorname {R e p e a t} _ {1} \left(f _ {j} ^ {\prime}\right), +$$ + +where the operation $\mathrm{Repeat}_1$ means repeating the features three times to produce $f_{j}\in \mathbb{R}^{3d\times 1}$ . The neighbor aggregation is implemented using an anistropic function, which is described as: + +$$ +e ^ {j} = \operatorname {R e p e a t} _ {2} \left(p _ {j} - p _ {i}\right), +$$ + +$$ +f _ {i} ^ {(1)} = \frac {1}{k} \sum_ {j \in \mathcal {N} (i)} \left(f _ {j} \odot e _ {j}\right), +$$ + +where the position embedding $e^j \in \mathbb{R}^{3d \times 1}$ is produced by simply repeating the relative coordinates $d$ times. Note that, a similar position embedding and aggregation method are also adopted by PosPool [10], which is named position pooling. Similar with Point Transformer [32], the point feature is updated with an MLP layer and a shortcut. + +# 4. Explore Best Practices + +Given how many instantiations of PointMeta have appeared in the literature, we should push this general framework as far as possible in the architecture exploration to determine the most crucial implementation details. With the help of PointMeta, it is convenient to use quick experiments with variables controlled to verify any empirical observations or assumptions summarized from the comparison. In this section, we first compare and analyze the building blocks of existing models from the implementation details of these meta functions (Sec. 4.1, 4.2, 4.3, 4.4). Then, + +we summarize the best practices from the systematic analysis (Sec. 4.5). Finally, applying these practices, we do simple tweaks on the building blocks and propose a simple yet well-performing block, PointMetaBase (Sec. 4.6). + +# 4.1. Neighbor Update + +According to the order of Group and MLP operation, the neighbor update function is categorized into two classes: + +- Group before MLP [7, 14, 16, 27]. +MLP before Group [6,8,10,15,32]. + +As pointed out by ASSANet [15], compared with the first order, adopting the second order reduces the FLOPs by $\frac{d\times d\times N\times K\times L}{d\times d\times N\times L} = K$ times, where $N$ denotes the point number, $K$ denotes the neighbor number, $L$ denotes the layer number of the MLP. Although applying MLP before Group brings a large efficiency gain for the neighbor update, a side effect is that the relative coordinates cannot be used as input for the MLP. However, this problem can be easily solved by other position embedding methods, which is specified in Sec. 4.2. + +
VariantmIoU (%)Params (M)FLOPs (G)
Plain-Max47.3±0.72.01.4
Plain-PP65.1±0.22.01.5
Plain-PP-Max58.4±0.62.01.5
Plain-IPE-Max68.0±0.32.013.8
Plain-EPE-Max69.0±0.32.01.8
Plain-EPE-PP65.4±0.12.01.8
+ +Table 1. Performance of the blocks with different position embedding functions and neighbor aggregation function on Area 5 of S3DIS [1]. Plain-PP-Max means multiplying the neighbor features with relative coordinates as PP but aggregating via max pooling. + +# 4.2. Position Embedding + +The manners of implementing the position embedding can be categorized into three classes: + +- Implicit position embedding (IPE) [7, 14, 16, 27]. +- Explicit position embedding (EPE) [6,8,32]. +- Position pooling (PP) [10, 15]. + +Implicit position embedding (IPE). Following PointNet++ [14], many methods [12, 16, 31] concatenate the relative coordinates with the neighbor features as inputs of the MLP layer in the neighbor update function. As going deeper through the networks, the dimension of the neighbor features become higher while the coordinates dimension still remains 3. This may cause the loss of position + +information in the high-level features. In addition, IPE requires applying MLP before Group in the neighbor update function, which will increase the computation $K$ times. + +Explicit position embedding (EPE). Incorporating explicit representations of position information is a common practice for ViT [11, 23, 28]. As mentioned in Sec. 3.2, Point Transformer [32] learn a position embedding $e^j$ using $\mathrm{MLP}_2$ and add it to the neighbor features $f_j$ . Compare with IPE, EPE avoid the dimension imbalance between the position representations and the neighbor features. Because the input dimension of $\mathrm{MLP}_2$ is only 3, the computation of EPE becomes very small by controlling the dimension of the hidden layers. PointCNN [8] also uses an MLP to implement the EPE function, but it merges the EPE by concatenation instead of addition, which incurs a computation increase. StratifiedFormer [6] maintains a learnable quantized lookup table as EPE. In this section, we select the EPE of Point Transformer as the representative EPE method for the later experiments due to its simplicity. + +Position pooling (PP). As mentioned in Sec. 3.2, AS-SANet [15] and PosPool [10] use relative coordinates as weights to pool neighbor features. This way can embed the position information into the neighbor features and make an aggregation function without learnable weights. The efficiency of position pooling is worth affirming, but embedding the position in a non-data-driven way may harm the learned neighbor features. + +Empirical analysis. We conduct empirical experiments to evaluate the different PE functions described above. The first step is to construct a plain block as the baseline. Following the suggestion given in Sec. 4.1, the neighbor update function adopts the order of MLP-before-Group. The layer number of the MLP here is set to 1. The neighbor aggregation function is a simple max pooling. The point update function is a 1-layer MLP. We denote this block as PlainMax. We equip Plain-Max with the three kinds of position embedding described above and compare their performance on Area 5 of S3DIS [1]. The MLP used respectively in EPE and IPE are both 1-layer. The macro-architecture is in level "L" (described in Sec. 4.6). As shown in Tab. 1, PlainEPE-Max achieves the best performance with a tiny computation increase against Plain-PP, Plain-PP-Max and Plain-IPE-Max. Because coupling the neighbor features transformation with position embedding, Plain-IPE incurs heavy redundant computation. In a summary, EPE is the optimal choice among the three kinds of position embeddings. + +# 4.3. Neighbor Aggregation + +According to whether learnable weights exist in the neighbor aggregation function, we divide the aggregations function into two types: + +Non-learnable aggregation [7, 10, 14-16, 27]. + +
VariantsmIoU (%)Params (M)FLOPs (G)
Point Trans [32]70.5±0.37.85.6
Point Trans (-VSA+Max)70.3±0.25.13.3
+ +Table 2. Ablation of the vector self-attention of Point Transformer [32] on Area 5 of S3DIS [1]. -VSA+Max means replacing VSA with max pooling. + +- Learnable aggregation [6,8,22,29,32]. + +Learnable aggregation. As mentioned in Sec. 3.1, the aggregation function should satisfy permutation invariance because the point cloud is unordered. The permutation invariance is acquired by learning a space transformation [8, 29], learning pointwise weights [32], or pairwise self-attention [6]. In this section, we select the vector self-attention of Point Transformer [32] as the representative implementation of the learnable aggregation function due to its excellent performance. + +Non-learnable aggregation. There are two types of non-learnable aggregation in the literatures: position pooling [10, 15] and max pooling [7, 14, 16, 27]. As mentioned in Sec. 4.2, position pooling embed the position in a nondata-driven way and may compete with the learned neighbor features. As shown in Tab. 1, Plain-EPE-Max surpasses Plain-EPE-PP significantly, which supports our assumption. Therefore, we select the max pooling as the representative non-learnable aggregation function. + +Empirical analysis. We conduct an ablation experiment on Point Transformer [32] to compare the aggregation power of both max pooling and vector self-attention. Specifically, we replace the VSA with a max pooling operation (denoted as -VSA+Max). As a result, the aggregation of the ablated Point Transformer is formulated as: + +$$ +e ^ {j} = \operatorname {M L P} _ {2} \left(p _ {j} - p _ {i}\right), +$$ + +$$ +f _ {i} ^ {(1)} = \operatorname {M a x} \left(\boldsymbol {f} _ {\mathcal {N} (i)} + \boldsymbol {e} _ {\mathcal {N} (i)}\right), j \in \mathcal {N} (i), \tag {4} +$$ + +where $e_{\mathcal{N}(i)} = \{e_j | j \in \mathcal{N}(i)\}$ . As shown in Tab. 2, compared with VSA, max pooling achieves comparable performance with much fewer parameters and FLOPs. Here comes a question, why max pooling has a comparable power with the learnable aggregation? We explore the answer within our PointMeta framework and find that the max pooling is a special case of VSA but with binary and sparse attention. Specifically, the aggregation function Eq. 4 can be reformulated as follows: + +$$ +M _ {j} = \operatorname {S o f t m a x} _ {T \rightarrow 0} \left(\boldsymbol {f} _ {\mathcal {N} (i)}\right), +$$ + +$$ +f _ {i} ^ {(1)} = \sum_ {j \in \mathcal {N} (i)} \left(M _ {j} \odot \left(f _ {j} + e _ {j}\right)\right), +$$ + +
VariantsmIoU (%)Params (M)FLOPs (G)
N1P169.0±0.32.01.7
N2P068.7±0.32.01.7
N0P267.8±0.42.01.7
N1P269.5±0.32.72.0
N1P2-Inv69.7±0.37.13.8
N2P169.4±0.32.72.0
N2P1-Inv69.6±0.47.13.8
N1P369.5±0.33.42.3
N3P169.6±0.53.42.3
+ +Table 3. Semantic segmentation of the Plain-EPE with different MLP configuration on Area 5 of S3DIS [1]. "N1P2" means that the MLP in the neighbor update is 1-layer and 2-layer. "Inv" means the inverted bottleneck. + +where $\mathrm{Softmax}_{T\to 0}$ denotes a softmax function whose temperature approximates 0, which makes the softmax output $M_{j}$ approximates the binary masks. This formulation reveals the collaboration between max pooling and the neighbor update function, i.e., max pooling is an aggregation function with binary and sparse attention which is generated by applying softmax on the output of the neighbor update + +# 4.4. Point Update + +The point update function is used to enhance the point features after the neighbor aggregation. Several existing studies do not implement this function in their building blocks [7, 27], and some other studies use a heavy point update function, such as a 2-layer MLP with an inverted bottleneck [6, 16]. A natural question arises: how much computation complexity should we allocate for the point update function? We answer this question via extensive ablated experiments. + +Empirical analysis. According to the exploration above, we adopt the Plain-EPE-Max as the baseline due to its excellent performance in the ablated experiments (Tab. 1). Since the computation are concentrated on the MLP of the neighbor update function and point update function, we can allocate computation between the two update functions by changing the configurations of the two MLP. We ablate the MLP from two aspects, i.e., the layer number and whether to use an inverted bottleneck that adopted by PointNeXt [16] and StratifiedFormer [6]. As shown in Tab. 3, the configuration "N1P2" can achieve an optimal balance between complexity and performance. More layers or inverted bottlenecks can only promote the model marginally but lead to much more computation. + +# 4.5. Best Practices + +Here we summarize the best practices for the building block as follows: + +- **Position Embedding.** Explicit position embedding can endow the models with better position awareness with a low computation cost comparing with other PE manners. + +- Point Update. Point update is important and the complexity allocation between the neighbor update and the point update should be properly set. + +- Neighbor Aggregation. Max pooling is not a pure non-learnable aggregation function, its aggregation power is comparable with the learnable aggregation function but with much less computation complexity. Any exploration for a new aggregation function should set max pooling as the baseline. + +- Neighbor Update. Applying the MLP before neighbors grouping will bring a $K$ times FLOPs reduction comparing with using the contrary order. However, the position embedding function must be carefully chosen for the position awareness. + +The first and the second practice are usually ignored by researchers but are vital for the effectiveness and efficiency of the building block, which is demonstrated in Sec. 4.2 and Sec. 4.4. The third shares a similar viewpoint with PosPool [10] which claims that different local aggregation operators perform similarly. However, from the perspective of PointMeta, we provide a deeper insight that max pooling can be reformulated as a special case of vector self-attention [32], which makes it have comparable aggregation power comparing with the learnable aggregation function. The fourth practice about neighbor update has been emphasized by ASSANet [15], we restate it here because it is important to notice its collaboration with position embedding function. We believe that these practices can provide a guideline for the future research. + +# 4.6. PointMetaBase + +We apply the best practices summarized in Sec. 4.5 to design the building block and implement the four meta functions as simply as possible. We name this block as PointMetaBase, which we hope to become a solid baseline for future research. Fig. 2 (g) shows the architecture of PointMetaBase. The layer number of the first MLP is set to 1 and that of the second MLP is set to 2. The neighbor aggregation function is a simple max pooling. We use the EPE adopted by Point Transformer [32] as the position embedding function. Applying EPE and the MLP-before-Group order, we tweak the set abstraction module [14, 16] as the reduction block, termed PointMetaSA. We adopt the same scaling strategies and decoder with PointNeXt [16] to construct our PointMetaBase family: + +- PointMetaBase-S: $C = 32$ , $B = 0$ + +
MethodS3DIS 6-FoldS3DIS Area-5ScanNet mIoUParams. MFLOPs GThroughput (ins./sec.)
mIoU (%)OA (%)mIoU (%)OA (%)Val (%)Test (%)
PointNet++ [14]54.581.053.583.053.555.71.07.2237
PointCNN [8]65.488.157.385.9-45.80.6--
DeepGCN [7]60.085.952.5---3.6--
KPConv [22]70.6-67.1-69.268.615.0--
RandLA-Net [5]70.088.0---64.51.35.8-
BAAF-Net [17]72.288.965.488.9--5.0--
Point Transformer [32]73.590.270.490.870.6-7.85.6-
CBL [20]73.189.669.490.6-70.518.6--
ASSANet [15]--65.888.9--2.42.5300
ASSANet-L [15]--68.089.7--115.636.2136
PointNeXt-S [16]68.087.463.4±0.887.9±0.364.5-0.83.6276
PointNeXt-B [16]71.588.867.3±0.289.4±0.168.4-3.88.9161
PointNeXt-L [16]73.989.869.0±0.590.0±0.169.4-7.115.2109
PointNeXt-XL [16]74.990.370.5±0.390.6±0.171.571.241.684.843
PointMetaBase-L75.690.669.5±0.390.5±0.171.0-2.72.0187
PointMetaBase-XL76.391.071.1±0.490.9±0.171.8-15.39.2104
PointMetaBase-XXL77.091.371.3±0.790.8±0.672.871.419.711.090
+ +Table 4. Semantic segmentation on S3DIS [1] (6-Fold and Area 5) and ScanNet V2 [3]. Our PointMetaBase family surpass the state-of-the-art method PointNeXt [16] significantly with large FLOPs reduction. + +- PointMetaBase-L: $C = 32$ , $B = (2,4,2,2)$ +- PointMetaBase-XL: $C = 64$ , $B = (3,6,3,3)$ +- PointMetaBase-XXL: $C = 64$ , $B = (4,8,4,4)$ + +$C$ denotes the channel size of the stem MLP and $B$ denotes the number of the PointMetaBase block in each stage. Note that when $B = 0$ , only one PointMetaSA block but no PointMetaBase blocks are used at each stage. Due to the excellent efficiency of PointMetaBase, we do not need to construct the network at the level "B" as done in Point-NeXt [16]. In contrast, we choose to scale the model to a larger level "XXL". The full details of network architecture are available in the supplementary materials. + +# 5. Experiments + +We studied the performance of PointMetaBase on S3DIS [1], ScanNet V2 [3], ScanObjectNN [24] and ShapeNetPart [30]. To enable a fair comparison, the same data processing and evaluation protocols adopted by the state-of-the-art method PointNeXt [16] were used in our experiments. + +Experimental Setups. We optimized all our models using CrossEntropy loss with label smoothing [19], AdamW optimizer, an initial learning rate 0.001, and weight decay $10^{-4}$ , with Cosine Decay, with a 32G V100 GPU. Model parameters, FLOPs and inference throughput are provided for comparison. For a fair comparison, we use the open + +codes provided by PointNext [16] to conduct the measurements. We also use MindSpore to validate the generalization on different deep learning frameworks. All the measurements are conducted using an NVIDIA Tesla A100 40GB GPU and a 12-core Intel Xeon @ 2.40GHz CPU. + +# 5.1.3D Scene Segmentation + +S3DIS [1] is a challenging benchmark composed of 6 large-scale indoor areas, 271 rooms, and 13 semantic categories in total. The standard 6-fold cross-validation results and the Area-5 results on S3DIS are reported in Tab. 4. Since we adopt the same scaling strategies, PointMetaBase-L and PointMetaBase-XL have the same configuration on width and depth as PointNeXt-L and PointNeXt-XL. With only $13\%$ FLOPs, our PointMetaBase-L outperforms PointNext-L by $1.7\%$ and $0.8\%$ in terms of mean IoU (mIoU) and overall accuracy (OA), respectively, and is much faster in terms of throughput. Compared with PointNext-XL, our PointMetaBase-XL uses only $11\%$ of the FLOPs, but surpasses with $1.4\%$ mIoU and $0.7\%$ OA. The inference speed is $2.4\times$ faster. The excellent efficiency and performance of PointMetaBase can be attributed to the the removal of the computation redundancy in the two update functions and the improved position awareness introduced by explicit position embedding. ScanNet [3] contains 3D indoor scenes of various rooms with 20 semantic categories. The dataset is divided into 3 part: training, + +
Methodins. mIoU (%)cls. mIoU (%)Params (M)FLOPs (G)Throughput (ins./sec)
PointNet++ [14]85.181.91.04.9-
KPConv [22]86.485.1---
ASSANet [15]86.1----
Point Transformer [32]86.683.77.8--
PointMLP [12]86.184.6---
StratifiedFormer [6]86.685.1---
PointNeXt-S [16]86.7±0.084.4±0.21.04.5873
PointNeXt-S (C=64) [16]86.9±0.184.8±0.53.717.8391
PointNeXt-S (C=160) [16]87.0±0.185.2±0.122.5110.2118
PointMetaBase-S (ours)86.7±0.084.3±0.11.01.391194
PointMetaBase-S (C=64) (ours)86.9±0.184.9±0.23.83.85706
PointMetaBase-S (C=160) (ours)87.1±0.085.1±0.322.718.45271
+ +Table 5. Part segmentation on ShapeNetPart [30]. Compaerd with the state-of-the-art method PointNeXt [16], our PointMetaBase family achieve comparable performance with large FLOPs reduction and faster inference speed. + +
VariantsOA (%)FLOPs (G)TP (ins./sec)
PointNet++ [14]77.91.72655
PointCNN [8]78.5--
DGCNN [27]78.14.8735
DRNet [13]80.3--
PointMLP [12]85.4±1.331.3405
PointNeXt-S [16]87.7±0.41.72230
PointMetaBase-S87.9±0.20.62674
+ +Table 6. Classification on ScanObjectNN [24]. Our model surpasses PointNeXt-S [16] both with much less computation cost. + +validation, and test splits, with 1201, 312 and 100 scans, respectively. Following PointNeXt [16], we do not use any voting strategies. As shown in Tab. 4, our PointMetaBase-L surpasses PointNeXt-L $1.6\%$ mIoU with only $14\%$ FLOPs and $1.7\times$ faster inference speed. PointMetaBase-XL surpasses PointNeXt-XL $0.3\%$ mIoU with only $11.3\%$ FLOPs and $2.4\times$ faster inference speed. + +# 5.2. 3D Object Classification + +ScanObjectNN [24] contains about 15000 real scanning objects, which are grouped into 15 classes and have 2902 unique instances. Following [12, 16], we experiment on the hardest perturbed variant (PB_T50_RS). To make a fair comparison with PointNeXt [16], we did not use upscaled variants of PointMetaBase on this benchmark. As reported in Tab. 6, our model surpasses PointNeXt-S, while using much fewer model FLOPs and running much faster. + +# 5.3. 3D Part Segmentation + +ShapeNetPart [30] is a dataset for part segmentation of objects. It consists of 16880 models with 16 differ + +ent shape categories. Following the state-of-the-art method PointNeXt [16], we used voting by averaging the results of 10 randomly scaled input point clouds, with scaling factors equal to [0.8,1.2]. As shown in Tab. 5, compaerd with PointNeXt, our PointMetaBase family achieve comparable performance with large FLOPs reduction and faster inference speed. Specially, for $C = 32 / C = 64 / C = 160$ , the FLOPs of our PointMetaBase are only $30\% 21\% / 17\%$ that of PointNeXt, and the inference speed of PointMetaBase is $1.4 / 1.8 / 2.3$ times faster than that of PointNeXt. + +# 6. Conclusion + +In this paper, we reformulate existing models for point cloud analysis into a general framework, PointMeta. Within this framework, we provide an in-depth analysis on the building blocks of existing models and summarize a few best practices. Furthermore, we follow these best practices and propose a simple building block, PointMetaBase, which enjoys excellent performance and efficiency. Our work can inspire new insights into building block design and future architecture exploration. Next, building on this work, we plan to explore neural architecture search methods [33-37] that are specifically tailored for point cloud analysis. + +Acknowledgement. This work was supported by National Key R&D Program of China (No. 2022ZD0118202), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J01002, No. 2022J06001), Guangdong Basic and Applied Basic Research Foundation (No. 2019B1515120049), and CAAI-Huawei MindSpore Open Fund. + +# References + +[1] Iro Armeni, Sasha Sax, Amir R. Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. CoRR, abs/1702.01105, 2017. 1, 2, 4, 5, 6, 7 +[2] Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Flores Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gulçehre, H. Francis Song, Andrew J. Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, Kelsey R. Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matthew M. Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational inductive biases, deep learning, and graph networks. CoRR, abs/1806.01261, 2018. 2 +[3] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 2432–2443. IEEE Computer Society, 2017. 7 +[4] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1 +[5] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11105-11114, 2020. 2, 3, 7 +[6] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8490-8499. IEEE, 2022. 2, 4, 5, 6, 8 +[7] Guohao Li, Matthias Müller, Ali K. Thabet, and Bernard Ghanem. Deep GCs: Can GCs go as deep as cnns? In ICCV, pages 9266-9275, 2019. 2, 4, 5, 6, 7 +[8] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, pages 828-838, 2018. 2, 3, 4, 5, 7, 8 +[9] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In CVPR, pages 8895-8904, 2019. 2 +[10] Ze Liu, Han Hu, Yue Cao, Zheng Zhang, and Xin Tong. A closer look at local aggregation operators in point cloud analysis. In ECCV, pages 326-342, 2020. 2, 4, 5, 6 +[11] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002, 2021. 3, 5 +[12] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 4, 8 +[13] Charles Ruizhongtai Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 77-85, 2017. 1, 2, 8 +[14] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on + +point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1, 2, 3, 4, 5, 6, 7, 8 +[15] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali K. Thabet, and Bernard Ghanem. Assanet: An anisotropic separable set abstraction for efficient point cloud representation learning. In NeurIPS, pages 28119-28130, 2021. 2, 3, 4, 5, 6, 7, 8 +[16] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Abed Al Kader Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[17] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 7 +[18] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, ICLR, 2015. 1 +[19] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7 +[20] Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. In CVPR, pages 8479-8489, 2022. 7 +[21] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, pages 3887-3896, 2018. 2 +[22] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6410–6419, 2019. 2, 3, 5, 7, 8 +[23] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 3, 5 +[24] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, pages 1588–1597, 2019. 7, 8 +[25] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, pages 10296-10305, 2019. 2 +[26] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In CVPR, pages 2589-2597, 2018. 2 +[27] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. ACM Trans. Graph., 38(5):146:1-146:12, 2019. 2, 3, 4, 5, 6, 8 +[28] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In ICCV, pages 10013-10021, 2021. 3, 5 + +[29] Wenxuan Wu, Zhongang Qi, and Fuxin Li. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, pages 9621-9630, 2019. 3, 5 +[30] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Trans. Graph., 35(6):210:1-210:12, 2016. 7, 8 +[31] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, pages 5565-5573, 2019. 4 +[32] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16239-16248, 2021. 2, 3, 4, 5, 6, 7, 8 +[33] Xiawu Zheng, Xiang Fei, Lei Zhang, Chenglin Wu, Fei Chao, Jianzhuang Liu, Wei Zeng, Yonghong Tian, and Rongrong Ji. Neural architecture search with representation mutual information. In CVPR, pages 11912-11921, 2022. 8 +[34] Xiawu Zheng, Rongrong Ji, Yuhang Chen, Qiang Wang, Baochang Zhang, Jie Chen, Qixiang Ye, Feiyue Huang, and Yonghong Tian. Migo-nas: Towards fast and generalizable neural architecture search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9):2936-2952, 2021. 8 +[35] Xiawu Zheng, Rongrong Ji, Lang Tang, Baochang Zhang, Jianzhuang Liu, and Qi Tian. Multinomial distribution learning for effective neural architecture search. In ICCV, 2019. 8 +[36] Xiawu Zheng, Rongrong Ji, Qiang Wang, Qixiang Ye, Zhenguo Li, Yonghong Tian, and Qi Tian. Rethinking performance estimation in neural architecture search. In CVPR, 2020. 8 +[37] Xiawu Zheng, Yang Zhang, Sirui Hong, Huixia Li, Lang Tang, Youcheng Xiong, Jin Zhou, Yan Wang, Xiaoshuai Sun, Pengfei Zhu, Chenglin Wu, and Rongrong Ji. Evolving fully automated machine learning via life-long knowledge anchors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9):3091-3107, 2021. 8 \ No newline at end of file diff --git a/2023/Meta Architecture for Point Cloud Analysis/images.zip b/2023/Meta Architecture for Point Cloud Analysis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..82dacee49de6050206778280e1db3897edc748dd --- /dev/null +++ b/2023/Meta Architecture for Point Cloud Analysis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2b9bb9728e81e1fc3ed007d339cd708eec2dc7ad70ca317e1ef19d45de5e37d +size 503737 diff --git a/2023/Meta Architecture for Point Cloud Analysis/layout.json b/2023/Meta Architecture for Point Cloud Analysis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3e8de55529961886c86afa9527a62f6f7e99e951 --- /dev/null +++ b/2023/Meta Architecture for Point Cloud Analysis/layout.json @@ -0,0 +1,9097 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 161, + 103, + 433, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 103, + 433, + 121 + ], + "spans": [ + { + "bbox": [ + 161, + 103, + 433, + 121 + ], + "type": "text", + "content": "Meta Architecture for Point Cloud Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "spans": [ + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "content": "Haojia Lin" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Xiawu Zheng" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Lijiang Li" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Fei Chao" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Shanshan Wang" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Yan Wang" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Yonghong Tian" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2,5}" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Rongrong Ji" + }, + { + "bbox": [ + 123, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,6*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 171, + 526, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 171, + 526, + 198 + ], + "spans": [ + { + "bbox": [ + 66, + 171, + 526, + 198 + ], + "type": "text", + "content": "1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, School of Informatics, Xiamen University, 361005, P.R. China." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 199, + 471, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 199, + 471, + 213 + ], + "spans": [ + { + "bbox": [ + 121, + 199, + 471, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 199, + 471, + 213 + ], + "type": "text", + "content": "Peng Cheng Laboratory. " + }, + { + "bbox": [ + 121, + 199, + 471, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 199, + 471, + 213 + ], + "type": "text", + "content": "Chinese Academy of Sciences. " + }, + { + "bbox": [ + 121, + 199, + 471, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 199, + 471, + 213 + ], + "type": "text", + "content": "Samsara Inc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 213, + 493, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 213, + 493, + 241 + ], + "spans": [ + { + "bbox": [ + 99, + 213, + 493, + 241 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 99, + 213, + 493, + 241 + ], + "type": "text", + "content": "National Engineering Research Center for Visual Technology, Peking University. " + }, + { + "bbox": [ + 99, + 213, + 493, + 241 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 99, + 213, + 493, + 241 + ], + "type": "text", + "content": "Shenzhen Research Institute of Xiamen University." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 243, + 516, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 243, + 516, + 255 + ], + "spans": [ + { + "bbox": [ + 75, + 243, + 516, + 255 + ], + "type": "text", + "content": "{linhj, lilijiang}@stu.xmu.edu.cn, zhengxw01@pcl.ac.cn, {rrji, feichao}@xmu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 258, + 455, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 258, + 455, + 269 + ], + "spans": [ + { + "bbox": [ + 135, + 258, + 455, + 269 + ], + "type": "text", + "content": "ss.wang@siat.ac.cn, yan.wang@samsara.com, yhtian@pku.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 297, + 192, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 297, + 192, + 309 + ], + "spans": [ + { + "bbox": [ + 143, + 297, + 192, + 309 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 329, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 289, + 653 + ], + "type": "text", + "content": "Recent advances in 3D point cloud analysis bring a diverse set of network architectures to the field. However, the lack of a unified framework to interpret those networks makes any systematic comparison, contrast, or analysis challenging, and practically limits healthy development of the field. In this paper, we take the initiative to explore and propose a unified framework called PointMeta, to which the popular 3D point cloud analysis approaches could fit. This brings three benefits. First, it allows us to compare different approaches in a fair manner, and use quick experiments to verify any empirical observations or assumptions summarized from the comparison. Second, the big picture brought by PointMeta enables us to think across different components, and revisit common beliefs and key design decisions made by the popular approaches. Third, based on the learnings from the previous two analyses, by doing simple tweaks on the existing approaches, we are able to derive a basic building block, termed PointMetaBase. It shows very strong performance in efficiency and effectiveness through extensive experiments on challenging benchmarks, and thus verifies the necessity and benefits of high-level interpretation, contrast, and comparison like PointMeta. In particular, PointMetaBase surpasses the previous state-of-the-art method by " + }, + { + "bbox": [ + 47, + 329, + 289, + 653 + ], + "type": "inline_equation", + "content": "0.7\\% /1.4\\% 2.1\\%" + }, + { + "bbox": [ + 47, + 329, + 289, + 653 + ], + "type": "text", + "content": " mIoU with only " + }, + { + "bbox": [ + 47, + 329, + 289, + 653 + ], + "type": "inline_equation", + "content": "2\\% /11\\% /13\\%" + }, + { + "bbox": [ + 47, + 329, + 289, + 653 + ], + "type": "text", + "content": " of the computation cost on the S3DIS datasets. The code and models are available at https://github.com/linhaojia13/PointMetaBase." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 298, + 543, + 476 + ], + "blocks": [ + { + "bbox": [ + 309, + 298, + 543, + 476 + ], + "lines": [ + { + "bbox": [ + 309, + 298, + 543, + 476 + ], + "spans": [ + { + "bbox": [ + 309, + 298, + 543, + 476 + ], + "type": "image", + "image_path": "2770bbcbf68355d70490c1ac7cd4a1976b518ec991a73ee3e7a8829f01a3532b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 479, + 545, + 513 + ], + "lines": [ + { + "bbox": [ + 305, + 479, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 545, + 513 + ], + "type": "text", + "content": "Figure 1. Segmentation performance of PointMetaBase on S3DIS [1]. PointMetaBase surpasses the state-of-the-art method PointNeXt [16] significantly with a large FLOPs reduction." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 535, + 386, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 535, + 386, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 535, + 386, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "type": "text", + "content": "In the past two decades, the popularity of 3D data acquisition technology has led to great interest in point cloud analysis. Unlike images, point clouds are inherently sparse, unordered, and irregular. These characteristics make it challenging for the powerful convolutional neural networks (CNN) [4, 18] to extract useful information from point clouds. Early studies attempt to convert the point cloud into grids by either voxelization or 2D projections, such that the standard CNN can be applied, but the applications are limited by the extra computation and information loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "With the introduction of permutation equivariance and invariance by PointNet [13] and PointNet++ [14], it is possible to use CNNs to process point clouds in their unstruc" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "type": "text", + "content": "*Corresponding author: rrji@xmu.edu.cn" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17682" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "tured format. Following PointNet++ [14], many point-based networks have been introduced, most of which focus on the development of complex building blocks to extract local features, such as the " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " - Conv convolution in PointCNN [22] and the self-attention layer in Point Transformer [32]. Although brought significant performance gains, these models are very complicated. On one hand, this makes the approaches computational demanding and thus limited in applications. On the other hand, it also gives extra challenges in systematic comparison and analysis among the models, and thus impacts efficient development of the field, which ideally would benefit from theoretic and empirical guidance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 430 + ], + "type": "text", + "content": "The community is well aware of this issue, and research efforts have been put on looking for a unified perspective to compare these methods and identify the most crucial implementation details. PosPool [10] and PointNeXt [16] take a step toward this goal. Adopting a deep residual architecture as the base network, PosPool [10] evaluates the representative building blocks [8, 14, 22, 26] and finds that they perform similarly well. PointNeXt [16] revisits the improved training and scaling strategies adopted by previous SotA methods [7, 15, 22, 32] and tweaks the early model PointNet++ [14] with the learnings, which achieves state-of-the-art performance. Both of them give insights and inspirations to the community via dedicated experiment designs and extensive empirical analysis. However, neither of them provides a framework that is general enough to fit a broad range of point cloud analysis approaches, such that large scale comparison and contrast could be done." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 432, + 289, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 289, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 289, + 562 + ], + "type": "text", + "content": "In this paper, we argue that if viewed from the perspective of basic building blocks, the majority of existing approaches could be fit into a single meta architecture (Sec. 3). We call it PointMeta. PointMeta abstracts the computation pipeline of the building blocks for point cloud analysis into four meta functions: a neighbor update function, a neighbor aggregation function, a point update function and a position embedding function (implicit or explicit). As will be discussed in more details in the next sections, this framework brings three benefits, which are also our core technical contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 573, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 573, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 573, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 573, + 289, + 693 + ], + "type": "text", + "content": "- In the dimension of models, it allows us to compare and contrast different models in a fair manner. So it becomes practical to observe and summarize learnings and assumptions, whose correctness could be further verified through experiments with variables controlled under the same framework. For example, among the systematic analysis on all the components across popular models, we found for the positional embedding function, explicit positional embedding is empirically the best choice (Sec. 4.2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "content": "- In the dimension of components, it allows us to have" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 156 + ], + "type": "text", + "content": "a higher level view across components, and thus to revisit the common beliefs and design decisions of the existing approaches. For example, despite the common perception, we find that the neighbor aggregation blocks may collaborate or compete with the learned neighbor features (Sec. 4.3), and thus should be designed carefully." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 165, + 547, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 547, + 273 + ], + "type": "text", + "content": "- Based on the learnings from the previous two dimensions, we are then able to do simple tweaks on the building blocks to apply the best practices (Sec. 4.5). The result building block, PointMetaBase, achieves very strong performance, surpassing the previous state-of-the-art method [16] by " + }, + { + "bbox": [ + 317, + 165, + 547, + 273 + ], + "type": "inline_equation", + "content": "0.7\\% / 1.4\\% / 2.1\\%" + }, + { + "bbox": [ + 317, + 165, + 547, + 273 + ], + "type": "text", + "content": " mIoU with only " + }, + { + "bbox": [ + 317, + 165, + 547, + 273 + ], + "type": "inline_equation", + "content": "2\\% / 11\\% / 13\\%" + }, + { + "bbox": [ + 317, + 165, + 547, + 273 + ], + "type": "text", + "content": " of the computation cost on S3DIS [1] (Fig. 1), and can act as a baseline for further research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 285, + 392, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 392, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 392, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 305, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 547, + 545 + ], + "type": "text", + "content": "Neural Networks for Point Cloud Analysis. The unstructured data format of point cloud makes it difficult to apply CNN directly. To address this challenge, PointNet [13] introduces permutation equivariance and invariance via point-wise MLP and max pooling, which makes it possible to process point cloud directly. To better capture locality, PointNet++ [14] presents a Set Abstraction module to aggregate features from the neighbor points. Most works after PointNet++ focus on the design of build blocks. Graph-based methods [7, 25, 27] utilize graph neural networks to extract local features. Pseudo-grid-based methods [8,21,22] transform the local features onto pseudo grids which allow for regular convolutions. Adaptive weight-based methods [5,9] aggregate the neighbor features via weights adaptively learned from the locality. Recently, Transformer-based block [6, 32] are introduced to capture local information through self-attention. These proposed building blocks are so complicated that make challenges in systematic comparison and analysis, which suggests an urgent need of a common framework to unify existing models in this field." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 556, + 375, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 556, + 375, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 556, + 375, + 569 + ], + "type": "text", + "content": "3. PointMeta" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 576, + 416, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 576, + 416, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 576, + 416, + 588 + ], + "type": "text", + "content": "3.1. Meta Architecture" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "content": "Inspired by the graph network framework [2], we reformulate existing building blocks for point cloud analysis into a single meta architecture, named PointMeta. As shown in Fig. 2 (a), a PointMeta block is composed of four meta functions: a neighbor update function " + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\phi^n" + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "content": ", a neighbor aggregation function " + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\phi^a" + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "content": ", a point update function " + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\phi^p" + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "content": " and a position embedding function " + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\phi^e" + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "content": " (implicit or explicit). The forward pass of a PointMeta block is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 701, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 701, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 375, + 701, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\mathcal {N} (i)} = \\phi^ {n} \\circ \\phi^ {e} \\left(f _ {i}, p _ {i}\\right), \\tag {1}", + "image_path": "8ed13a369a93d1f9317e049e7527bb3d8ec3764e64f3996896d75898c8a34c4d.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17683" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 127, + 201 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 127, + 201 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 127, + 201 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 127, + 201 + ], + "type": "image", + "image_path": "41d674684abf894d1a8083199d7c3e93452c04390476ce8670060651356bd6d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 203, + 102, + 212 + ], + "lines": [ + { + "bbox": [ + 59, + 203, + 102, + 212 + ], + "spans": [ + { + "bbox": [ + 59, + 203, + 102, + 212 + ], + "type": "text", + "content": "(a) PointMeta" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 72, + 181, + 202 + ], + "blocks": [ + { + "bbox": [ + 130, + 72, + 181, + 202 + ], + "lines": [ + { + "bbox": [ + 130, + 72, + 181, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 72, + 181, + 202 + ], + "type": "image", + "image_path": "963b7cc4839de276bd2d8d73ed9e1838cbff50ba68b1e4b21dd87b497897bf39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 203, + 179, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 203, + 179, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 179, + 211 + ], + "type": "text", + "content": "(b) PointNet++" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 72, + 241, + 202 + ], + "blocks": [ + { + "bbox": [ + 186, + 72, + 241, + 202 + ], + "lines": [ + { + "bbox": [ + 186, + 72, + 241, + 202 + ], + "spans": [ + { + "bbox": [ + 186, + 72, + 241, + 202 + ], + "type": "image", + "image_path": "88b61868aae50f225cc781fa976ffb589a02a22bf569464011831c6616a699eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 203, + 236, + 212 + ], + "lines": [ + { + "bbox": [ + 199, + 203, + 236, + 212 + ], + "spans": [ + { + "bbox": [ + 199, + 203, + 236, + 212 + ], + "type": "text", + "content": "(c) DGCNN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 248, + 72, + 316, + 203 + ], + "blocks": [ + { + "bbox": [ + 248, + 72, + 316, + 203 + ], + "lines": [ + { + "bbox": [ + 248, + 72, + 316, + 203 + ], + "spans": [ + { + "bbox": [ + 248, + 72, + 316, + 203 + ], + "type": "image", + "image_path": "b6ebcad888433fced510b30228249046d70ff071a2ce60cdb5b18b29f5c5cd56.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 204, + 297, + 212 + ], + "lines": [ + { + "bbox": [ + 255, + 204, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 255, + 204, + 297, + 212 + ], + "type": "text", + "content": "(d) PointCNN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 216, + 547, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 547, + 250 + ], + "type": "text", + "content": "Figure 2. PointMeta and its instantiation examples. (a) In PointMeta, the position embedding function is usually combined with the neighbor update function or the aggregation function implicitly or explicitly. (b) " + }, + { + "bbox": [ + 46, + 216, + 547, + 250 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 46, + 216, + 547, + 250 + ], + "type": "text", + "content": " (f) Representative building blocks can naturally fit into PointMeta. (g) Applying the summarized best practices, we do simple tweaks on the building blocks and propose PointMetaBase." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 400, + 202 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 400, + 202 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 400, + 202 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 400, + 202 + ], + "type": "image", + "image_path": "902577f248e47bcfc2d08ae786f41f008e2e50dfb0cd7bb4c05ca083eab86704.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 203, + 391, + 212 + ], + "lines": [ + { + "bbox": [ + 326, + 203, + 391, + 212 + ], + "spans": [ + { + "bbox": [ + 326, + 203, + 391, + 212 + ], + "type": "text", + "content": "(e) Point Transformer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 401, + 72, + 467, + 202 + ], + "blocks": [ + { + "bbox": [ + 401, + 72, + 467, + 202 + ], + "lines": [ + { + "bbox": [ + 401, + 72, + 467, + 202 + ], + "spans": [ + { + "bbox": [ + 401, + 72, + 467, + 202 + ], + "type": "image", + "image_path": "2e85920323a7b4a9fac7d04d08f53de16d208aa099da8c7a952a389be601048a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 203, + 451, + 212 + ], + "lines": [ + { + "bbox": [ + 413, + 203, + 451, + 212 + ], + "spans": [ + { + "bbox": [ + 413, + 203, + 451, + 212 + ], + "type": "text", + "content": "(f) ASSANet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 473, + 72, + 547, + 203 + ], + "blocks": [ + { + "bbox": [ + 473, + 72, + 547, + 203 + ], + "lines": [ + { + "bbox": [ + 473, + 72, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 473, + 72, + 547, + 203 + ], + "type": "image", + "image_path": "46995296ac8f64a4b113e7208b8f058ce2d77058ec41277de09fb2158561b9d2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 204, + 534, + 212 + ], + "lines": [ + { + "bbox": [ + 478, + 204, + 534, + 212 + ], + "spans": [ + { + "bbox": [ + 478, + 204, + 534, + 212 + ], + "type": "text", + "content": "(g) PointMetaBase" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 93, + 281, + 287, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 281, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 93, + 281, + 287, + 298 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {(1)} = \\phi^ {a} \\circ \\phi^ {e} \\left(f _ {i}, p _ {i}, \\mathbf {f} _ {\\mathcal {N} (i)}, \\mathbf {p} _ {\\mathcal {N} (i)}\\right) \\tag {2}", + "image_path": "f032bf4663fadfe2552d1d2908685e978b7fcfc356c311ffc7d7c765bf5dd580.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 314, + 287, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 314, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 125, + 314, + 287, + 331 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {(2)} = \\phi^ {p} \\left(f _ {i} ^ {(1)}, p _ {i}\\right), \\tag {3}", + "image_path": "2f157d6c3054ec53ef2e363d5b0ca0d304265a6f0f3e105172f6c62acd9fa14a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "f_{i} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " represent the feature and the coordinate of point " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " respectively, and " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " denotes the feature length. Let " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " denote the neighbors of point " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " be the feature and coordinate of each neighbor of point " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathcal{N}(i)} = \\{f_j | j \\in \\mathcal{N}(i)\\}" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " represent all neighbor features of point " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{\\mathcal{N}(i)} = \\{p_j | j \\in \\mathcal{N}(i)\\}" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " represent all neighbor coordinates of point " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 337, + 287, + 432 + ], + "type": "text", + "content": " denote the neighbor number." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "type": "text", + "content": "Neighbor update. The first step is to group and update neighbors for each point. Usually, the neighbor features " + }, + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{N}(i)}" + }, + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "type": "text", + "content": " is produced by K-Nearest query as [27, 32] or ball query as [14-16]. The update on neighbor features is usually conducted by a multi-layer perceptions (MLP) to make sure the permutation equivariance of the neighbor update function " + }, + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\phi^n" + }, + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "type": "text", + "content": ". Specifically, a permutation transformation on " + }, + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{N}(i)}" + }, + { + "bbox": [ + 46, + 434, + 287, + 541 + ], + "type": "text", + "content": " results in the same permutation change on the output of the neighbor update." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "text", + "content": "Neighbor aggregation. Literally, the neighbor aggregation function is used to aggregate the neighbor features for a center point. Different from " + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\phi^n" + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "text", + "content": ", the aggregation function " + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\phi^a" + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "text", + "content": " must satisfy permutation invariance, i.e., a permutation transformation on the input " + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{N}(i)}" + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "text", + "content": " will have no influence on the output " + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "inline_equation", + "content": "f_i^{(1)}" + }, + { + "bbox": [ + 46, + 542, + 287, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 617, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 641 + ], + "type": "text", + "content": "Point update. The point update usually contains an MLP, which is used to further enhance the point features." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Position embedding. Incorporating representations of position information via position embedding is a common practice for Vision Transformer (ViT) [11, 23, 28], since ViT models must capture spatial relationships of tokens. We thus introduce the concept of position embedding into PointMeta. Within a point cloud, points are organized in" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "text", + "content": "an unstructured format, however the points are highly correlated in the 3D Euclidean space. Therefore, the building block of point cloud analysis should possess the position awareness, which is provided by the position embedding function " + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\phi^e" + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\phi^e" + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "text", + "content": " is usually combined with the neighbor update function " + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\phi^n" + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "text", + "content": " or the aggregation function " + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "text", + "content": " implicitly or explicitly, thus we present them with an bundled notation, " + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\phi^n\\circ \\phi^e" + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\phi^a\\circ \\phi^e" + }, + { + "bbox": [ + 304, + 270, + 545, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 373, + 437, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 437, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 437, + 385 + ], + "type": "text", + "content": "3.2. Instantiation Examples" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "type": "text", + "content": "In this section, we select three representative models from the literature as the instantiation examples and explain how these models implement the meta functions within PointMeta. We illustrate more instantiation examples [5, 8, 16, 22, 27, 29] in the supplementary due to the limited space. Sometimes these meta functions are implemented sophisticatedly, we will decompose them into several subfunctions for the sake of brevity and readability." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 487, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 487, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 487, + 545, + 510 + ], + "type": "text", + "content": "PointNet++ [14]. The neighbor update is described as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 365, + 519, + 487, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 519, + 487, + 534 + ], + "spans": [ + { + "bbox": [ + 365, + 519, + 487, + 534 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\mathcal {N} (i)} ^ {\\prime}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right),", + "image_path": "f673dd36f9aad44d89485b53c0ef5ce12b9b2b6875f1664d18f4ffa692e7f2b4.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 357, + 536, + 495, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 536, + 495, + 551 + ], + "spans": [ + { + "bbox": [ + 357, + 536, + 495, + 551 + ], + "type": "interline_equation", + "content": "f _ {j} = \\operatorname {M L P} _ {1} \\left(\\operatorname {C o n c a t} \\left(f _ {j} ^ {\\prime}, p _ {j} - p _ {i}\\right)\\right),", + "image_path": "553196469fcede740bc388c143dbdf61838391dde62e01917d02f7d58ace687d.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 557, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 545, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 557, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_1" + }, + { + "bbox": [ + 306, + 557, + 545, + 594 + ], + "type": "text", + "content": " denotes the multi-layer perceptions applied on the concatenated neighbor features. The neighbor aggregation is implemented using a simple max pooling operation," + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 383, + 601, + 468, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 601, + 468, + 617 + ], + "spans": [ + { + "bbox": [ + 383, + 601, + 468, + 617 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {(1)} = \\operatorname {M a x} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)}\\right).", + "image_path": "5fac4c85e4375c8ad20abeb02004145113ba3003de2f6922630d1d3ca8bf2e91.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 624, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 646 + ], + "type": "text", + "content": "The point update function is another MLP layer, thus we have" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 389, + 645, + 462, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 645, + 462, + 662 + ], + "spans": [ + { + "bbox": [ + 389, + 645, + 462, + 662 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {(2)} = \\mathrm {M L P} _ {2} (f _ {i})", + "image_path": "733edc6dd51abfce1cfee68a99bb562fa5ab0b4d4a0e0bbb187350d232eb3a86.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "Note that the position embedding function " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi_{e}" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " is implicitly implemented by " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_1" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": ", which incorporates the concatenation of the relative positions " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "p_j - p_i" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " and grouping features " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_j'" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " as input and returns position-aware neighbor features " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_j" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17684" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "Point Transformer [32]. The neighbor update is described as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 102, + 201, + 116 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 102, + 201, + 116 + ], + "spans": [ + { + "bbox": [ + 132, + 102, + 201, + 116 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\prime} = \\operatorname {M L P} _ {1} (f _ {i}),", + "image_path": "6ec371c4d33a592d8cd28eac3991cb677c10a046f50adc735656b0aeb6c47dfc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 228, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 118, + 228, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 228, + 132 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\mathcal {N} (i)}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right).", + "image_path": "b6b46223beb86cf6b48552f1849f2cb56272ba54947597ab66a6ad37b9e87844.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 137, + 287, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 137, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 287, + 159 + ], + "type": "text", + "content": "The neighbor aggregation is implemented using vector self-attention, which is described as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 166, + 279, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 166, + 279, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 166, + 279, + 224 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} e ^ {j} = \\operatorname {M L P} _ {2} \\left(p _ {j} - p _ {i}\\right), \\\\ M _ {j} = \\operatorname {S o f t m a x} \\left(\\mathrm {M L P} _ {5} \\left(\\mathrm {M L P} _ {3} \\left(f _ {i}\\right) - \\mathrm {M L P} _ {4} \\left(f _ {j}\\right) + e _ {j}\\right)\\right), \\\\ f _ {i} ^ {(1)} = \\sum_ {j \\in \\mathcal {N} (i)} \\left(M _ {j} \\odot \\left(f _ {j} + e _ {j}\\right)\\right), \\\\ \\end{array}", + "image_path": "5f2293d1e010011429863ba282d907aa7facf955d08475dd2fdac81a42648ff1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 231, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 231, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 287, + 278 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 47, + 231, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 231, + 287, + 278 + ], + "type": "text", + "content": " denotes the Hadamard product and " + }, + { + "bbox": [ + 47, + 231, + 287, + 278 + ], + "type": "inline_equation", + "content": "M\\in \\mathbb{R}^{k\\times d}" + }, + { + "bbox": [ + 47, + 231, + 287, + 278 + ], + "type": "text", + "content": " denotes the vector self-attention (VSA) mask. Finally, the point feature is updated with an MLP layer and a shortcut, then we have" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 283, + 218, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 283, + 218, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 283, + 218, + 300 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {(2)} = \\mathrm {M L P} _ {6} \\left(f _ {i} ^ {(1)}\\right) + f _ {i}", + "image_path": "ae01c7152a101e54c26b75e9ef2868b4d15560bab8f816de3d6532b7f05f1964.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 304, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 304, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 59, + 304, + 287, + 317 + ], + "type": "text", + "content": "ASSANet [15]. The neighbor update is described as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 335, + 201, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 201, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 201, + 348 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\prime} = \\operatorname {M L P} _ {1} (f _ {i}),", + "image_path": "e3b4ab6fa3283f942b206610db9e1b65e59664e32b8868a3c7d910eb7c6688ff.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 350, + 228, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 350, + 228, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 228, + 365 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\mathcal {N} (i)} ^ {\\prime}, \\boldsymbol {p} _ {\\mathcal {N} (i)} = \\operatorname {G r o u p} \\left(f _ {i}, p _ {i}\\right),", + "image_path": "c6379da627372344835d43585267614a7503e4f0036cf05015e2b577781daabb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 367, + 205, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 367, + 205, + 381 + ], + "spans": [ + { + "bbox": [ + 129, + 367, + 205, + 381 + ], + "type": "interline_equation", + "content": "f _ {j} = \\operatorname {R e p e a t} _ {1} \\left(f _ {j} ^ {\\prime}\\right),", + "image_path": "796a1bdb9fefc6db15894d6c02018bc008dae85782506dac885bf24fd36fb810.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 386, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 433 + ], + "type": "text", + "content": "where the operation " + }, + { + "bbox": [ + 47, + 386, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{Repeat}_1" + }, + { + "bbox": [ + 47, + 386, + 287, + 433 + ], + "type": "text", + "content": " means repeating the features three times to produce " + }, + { + "bbox": [ + 47, + 386, + 287, + 433 + ], + "type": "inline_equation", + "content": "f_{j}\\in \\mathbb{R}^{3d\\times 1}" + }, + { + "bbox": [ + 47, + 386, + 287, + 433 + ], + "type": "text", + "content": ". The neighbor aggregation is implemented using an anistropic function, which is described as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 439, + 215, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 439, + 215, + 453 + ], + "spans": [ + { + "bbox": [ + 117, + 439, + 215, + 453 + ], + "type": "interline_equation", + "content": "e ^ {j} = \\operatorname {R e p e a t} _ {2} \\left(p _ {j} - p _ {i}\\right),", + "image_path": "34f776bd0077875650a8b129e8b53229fb2cc235292cc46fc032488eeba9da33.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 455, + 221, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 455, + 221, + 485 + ], + "spans": [ + { + "bbox": [ + 113, + 455, + 221, + 485 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {(1)} = \\frac {1}{k} \\sum_ {j \\in \\mathcal {N} (i)} \\left(f _ {j} \\odot e _ {j}\\right),", + "image_path": "ca77b25b3d912ebebf8a973edc64831625e9fc4752b63a6da6bf33583d140ffa.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": "where the position embedding " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "e^j \\in \\mathbb{R}^{3d \\times 1}" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": " is produced by simply repeating the relative coordinates " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": " times. Note that, a similar position embedding and aggregation method are also adopted by PosPool [10], which is named position pooling. Similar with Point Transformer [32], the point feature is updated with an MLP layer and a shortcut." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 574, + 177, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 177, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 177, + 586 + ], + "type": "text", + "content": "4. Explore Best Practices" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Given how many instantiations of PointMeta have appeared in the literature, we should push this general framework as far as possible in the architecture exploration to determine the most crucial implementation details. With the help of PointMeta, it is convenient to use quick experiments with variables controlled to verify any empirical observations or assumptions summarized from the comparison. In this section, we first compare and analyze the building blocks of existing models from the implementation details of these meta functions (Sec. 4.1, 4.2, 4.3, 4.4). Then," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "we summarize the best practices from the systematic analysis (Sec. 4.5). Finally, applying these practices, we do simple tweaks on the building blocks and propose a simple yet well-performing block, PointMetaBase (Sec. 4.6)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 128, + 409, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 128, + 409, + 140 + ], + "spans": [ + { + "bbox": [ + 305, + 128, + 409, + 140 + ], + "type": "text", + "content": "4.1. Neighbor Update" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 146, + 545, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 146, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 545, + 170 + ], + "type": "text", + "content": "According to the order of Group and MLP operation, the neighbor update function is categorized into two classes:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 178, + 471, + 209 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 317, + 178, + 463, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 178, + 463, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 178, + 463, + 190 + ], + "type": "text", + "content": "- Group before MLP [7, 14, 16, 27]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 198, + 471, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 471, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 471, + 209 + ], + "type": "text", + "content": "MLP before Group [6,8,10,15,32]." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "text", + "content": "As pointed out by ASSANet [15], compared with the first order, adopting the second order reduces the FLOPs by " + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\frac{d\\times d\\times N\\times K\\times L}{d\\times d\\times N\\times L} = K" + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "text", + "content": " times, where " + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "text", + "content": " denotes the point number, " + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "text", + "content": " denotes the neighbor number, " + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 217, + 545, + 336 + ], + "type": "text", + "content": " denotes the layer number of the MLP. Although applying MLP before Group brings a large efficiency gain for the neighbor update, a side effect is that the relative coordinates cannot be used as input for the MLP. However, this problem can be easily solved by other position embedding methods, which is specified in Sec. 4.2." + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 322, + 345, + 529, + 445 + ], + "blocks": [ + { + "bbox": [ + 322, + 345, + 529, + 445 + ], + "lines": [ + { + "bbox": [ + 322, + 345, + 529, + 445 + ], + "spans": [ + { + "bbox": [ + 322, + 345, + 529, + 445 + ], + "type": "table", + "html": "
VariantmIoU (%)Params (M)FLOPs (G)
Plain-Max47.3±0.72.01.4
Plain-PP65.1±0.22.01.5
Plain-PP-Max58.4±0.62.01.5
Plain-IPE-Max68.0±0.32.013.8
Plain-EPE-Max69.0±0.32.01.8
Plain-EPE-PP65.4±0.12.01.8
", + "image_path": "4af9eac1eaf69dcdbb63ad9793f161cc82b165b0336fc3298737331a2d9f376c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 453, + 545, + 497 + ], + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 497 + ], + "type": "text", + "content": "Table 1. Performance of the blocks with different position embedding functions and neighbor aggregation function on Area 5 of S3DIS [1]. Plain-PP-Max means multiplying the neighbor features with relative coordinates as PP but aggregating via max pooling." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 520, + 423, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 423, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 423, + 533 + ], + "type": "text", + "content": "4.2. Position Embedding" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 538, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 538, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 538, + 545, + 562 + ], + "type": "text", + "content": "The manners of implementing the position embedding can be categorized into three classes:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 570, + 523, + 621 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 317, + 570, + 523, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 523, + 582 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 523, + 582 + ], + "type": "text", + "content": "- Implicit position embedding (IPE) [7, 14, 16, 27]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 590, + 507, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 507, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 507, + 602 + ], + "type": "text", + "content": "- Explicit position embedding (EPE) [6,8,32]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 609, + 450, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 609, + 450, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 609, + 450, + 621 + ], + "type": "text", + "content": "- Position pooling (PP) [10, 15]." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Implicit position embedding (IPE). Following PointNet++ [14], many methods [12, 16, 31] concatenate the relative coordinates with the neighbor features as inputs of the MLP layer in the neighbor update function. As going deeper through the networks, the dimension of the neighbor features become higher while the coordinates dimension still remains 3. This may cause the loss of position" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17685" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "information in the high-level features. In addition, IPE requires applying MLP before Group in the neighbor update function, which will increase the computation " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "text", + "content": "Explicit position embedding (EPE). Incorporating explicit representations of position information is a common practice for ViT [11, 23, 28]. As mentioned in Sec. 3.2, Point Transformer [32] learn a position embedding " + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "inline_equation", + "content": "e^j" + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_2" + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "text", + "content": " and add it to the neighbor features " + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "inline_equation", + "content": "f_j" + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "text", + "content": ". Compare with IPE, EPE avoid the dimension imbalance between the position representations and the neighbor features. Because the input dimension of " + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_2" + }, + { + "bbox": [ + 46, + 109, + 287, + 300 + ], + "type": "text", + "content": " is only 3, the computation of EPE becomes very small by controlling the dimension of the hidden layers. PointCNN [8] also uses an MLP to implement the EPE function, but it merges the EPE by concatenation instead of addition, which incurs a computation increase. StratifiedFormer [6] maintains a learnable quantized lookup table as EPE. In this section, we select the EPE of Point Transformer as the representative EPE method for the later experiments due to its simplicity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 301, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 396 + ], + "type": "text", + "content": "Position pooling (PP). As mentioned in Sec. 3.2, AS-SANet [15] and PosPool [10] use relative coordinates as weights to pool neighbor features. This way can embed the position information into the neighbor features and make an aggregation function without learnable weights. The efficiency of position pooling is worth affirming, but embedding the position in a non-data-driven way may harm the learned neighbor features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 397, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 625 + ], + "type": "text", + "content": "Empirical analysis. We conduct empirical experiments to evaluate the different PE functions described above. The first step is to construct a plain block as the baseline. Following the suggestion given in Sec. 4.1, the neighbor update function adopts the order of MLP-before-Group. The layer number of the MLP here is set to 1. The neighbor aggregation function is a simple max pooling. The point update function is a 1-layer MLP. We denote this block as PlainMax. We equip Plain-Max with the three kinds of position embedding described above and compare their performance on Area 5 of S3DIS [1]. The MLP used respectively in EPE and IPE are both 1-layer. The macro-architecture is in level \"L\" (described in Sec. 4.6). As shown in Tab. 1, PlainEPE-Max achieves the best performance with a tiny computation increase against Plain-PP, Plain-PP-Max and Plain-IPE-Max. Because coupling the neighbor features transformation with position embedding, Plain-IPE incurs heavy redundant computation. In a summary, EPE is the optimal choice among the three kinds of position embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 174, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 174, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 174, + 648 + ], + "type": "text", + "content": "4.3. Neighbor Aggregation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 691 + ], + "type": "text", + "content": "According to whether learnable weights exist in the neighbor aggregation function, we divide the aggregations function into two types:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 701, + 249, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 249, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 249, + 713 + ], + "type": "text", + "content": "Non-learnable aggregation [7, 10, 14-16, 27]." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 553, + 122 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 553, + 122 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 553, + 122 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 553, + 122 + ], + "type": "table", + "html": "
VariantsmIoU (%)Params (M)FLOPs (G)
Point Trans [32]70.5±0.37.85.6
Point Trans (-VSA+Max)70.3±0.25.13.3
", + "image_path": "654de023b9701b49ffddef88e810d1fa38cf2158d7f925859b409147ca4771f1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "type": "text", + "content": "Table 2. Ablation of the vector self-attention of Point Transformer [32] on Area 5 of S3DIS [1]. -VSA+Max means replacing VSA with max pooling." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 317, + 183, + 484, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 183, + 484, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 183, + 484, + 196 + ], + "type": "text", + "content": "- Learnable aggregation [6,8,22,29,32]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 203, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 310 + ], + "type": "text", + "content": "Learnable aggregation. As mentioned in Sec. 3.1, the aggregation function should satisfy permutation invariance because the point cloud is unordered. The permutation invariance is acquired by learning a space transformation [8, 29], learning pointwise weights [32], or pairwise self-attention [6]. In this section, we select the vector self-attention of Point Transformer [32] as the representative implementation of the learnable aggregation function due to its excellent performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 311, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 418 + ], + "type": "text", + "content": "Non-learnable aggregation. There are two types of non-learnable aggregation in the literatures: position pooling [10, 15] and max pooling [7, 14, 16, 27]. As mentioned in Sec. 4.2, position pooling embed the position in a nondata-driven way and may compete with the learned neighbor features. As shown in Tab. 1, Plain-EPE-Max surpasses Plain-EPE-PP significantly, which supports our assumption. Therefore, we select the max pooling as the representative non-learnable aggregation function." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "content": "Empirical analysis. We conduct an ablation experiment on Point Transformer [32] to compare the aggregation power of both max pooling and vector self-attention. Specifically, we replace the VSA with a max pooling operation (denoted as -VSA+Max). As a result, the aggregation of the ablated Point Transformer is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 380, + 498, + 471, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 498, + 471, + 512 + ], + "spans": [ + { + "bbox": [ + 380, + 498, + 471, + 512 + ], + "type": "interline_equation", + "content": "e ^ {j} = \\operatorname {M L P} _ {2} \\left(p _ {j} - p _ {i}\\right),", + "image_path": "a652213732e91638fc4fd13ad2124ebbd556c98ed239af539e0211a7f9ce5f65.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 531, + 545, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 531, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 345, + 531, + 545, + 546 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {(1)} = \\operatorname {M a x} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)} + \\boldsymbol {e} _ {\\mathcal {N} (i)}\\right), j \\in \\mathcal {N} (i), \\tag {4}", + "image_path": "377a398dea2b2226736a9c050c9fdb00407c996a564ac0d8b593938edd0aa280.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 550, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 550, + 545, + 658 + ], + "type": "inline_equation", + "content": "e_{\\mathcal{N}(i)} = \\{e_j | j \\in \\mathcal{N}(i)\\}" + }, + { + "bbox": [ + 304, + 550, + 545, + 658 + ], + "type": "text", + "content": ". As shown in Tab. 2, compared with VSA, max pooling achieves comparable performance with much fewer parameters and FLOPs. Here comes a question, why max pooling has a comparable power with the learnable aggregation? We explore the answer within our PointMeta framework and find that the max pooling is a special case of VSA but with binary and sparse attention. Specifically, the aggregation function Eq. 4 can be reformulated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 369, + 668, + 482, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 668, + 482, + 681 + ], + "spans": [ + { + "bbox": [ + 369, + 668, + 482, + 681 + ], + "type": "interline_equation", + "content": "M _ {j} = \\operatorname {S o f t m a x} _ {T \\rightarrow 0} \\left(\\boldsymbol {f} _ {\\mathcal {N} (i)}\\right),", + "image_path": "78b5ff442637edfe3ff831e9aa0121524519f0d701fc9a03da10d2656ac9e06e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 360, + 684, + 492, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 684, + 492, + 710 + ], + "spans": [ + { + "bbox": [ + 360, + 684, + 492, + 710 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {(1)} = \\sum_ {j \\in \\mathcal {N} (i)} \\left(M _ {j} \\odot \\left(f _ {j} + e _ {j}\\right)\\right),", + "image_path": "addc5e6496c1c54508b7edb7cdec2fb8470bb5ca37c86d8df71c571b2225218c.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17686" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 258, + 204 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 258, + 204 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 258, + 204 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 258, + 204 + ], + "type": "table", + "html": "
VariantsmIoU (%)Params (M)FLOPs (G)
N1P169.0±0.32.01.7
N2P068.7±0.32.01.7
N0P267.8±0.42.01.7
N1P269.5±0.32.72.0
N1P2-Inv69.7±0.37.13.8
N2P169.4±0.32.72.0
N2P1-Inv69.6±0.47.13.8
N1P369.5±0.33.42.3
N3P169.6±0.53.42.3
", + "image_path": "323155c65b0e41b65c644c3c9eceaac8ada571f6efde3a1ea6de33dce4e4bc6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 213, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 257 + ], + "type": "text", + "content": "Table 3. Semantic segmentation of the Plain-EPE with different MLP configuration on Area 5 of S3DIS [1]. \"N1P2\" means that the MLP in the neighbor update is 1-layer and 2-layer. \"Inv\" means the inverted bottleneck." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 278, + 286, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 286, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 286, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 278, + 286, + 362 + ], + "type": "inline_equation", + "content": "\\mathrm{Softmax}_{T\\to 0}" + }, + { + "bbox": [ + 46, + 278, + 286, + 362 + ], + "type": "text", + "content": " denotes a softmax function whose temperature approximates 0, which makes the softmax output " + }, + { + "bbox": [ + 46, + 278, + 286, + 362 + ], + "type": "inline_equation", + "content": "M_{j}" + }, + { + "bbox": [ + 46, + 278, + 286, + 362 + ], + "type": "text", + "content": " approximates the binary masks. This formulation reveals the collaboration between max pooling and the neighbor update function, i.e., max pooling is an aggregation function with binary and sparse attention which is generated by applying softmax on the output of the neighbor update" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 370, + 132, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 132, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 132, + 383 + ], + "type": "text", + "content": "4.4. Point Update" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 286, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 495 + ], + "type": "text", + "content": "The point update function is used to enhance the point features after the neighbor aggregation. Several existing studies do not implement this function in their building blocks [7, 27], and some other studies use a heavy point update function, such as a 2-layer MLP with an inverted bottleneck [6, 16]. A natural question arises: how much computation complexity should we allocate for the point update function? We answer this question via extensive ablated experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 286, + 662 + ], + "type": "text", + "content": "Empirical analysis. According to the exploration above, we adopt the Plain-EPE-Max as the baseline due to its excellent performance in the ablated experiments (Tab. 1). Since the computation are concentrated on the MLP of the neighbor update function and point update function, we can allocate computation between the two update functions by changing the configurations of the two MLP. We ablate the MLP from two aspects, i.e., the layer number and whether to use an inverted bottleneck that adopted by PointNeXt [16] and StratifiedFormer [6]. As shown in Tab. 3, the configuration \"N1P2\" can achieve an optimal balance between complexity and performance. More layers or inverted bottlenecks can only promote the model marginally but lead to much more computation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "type": "text", + "content": "4.5. Best Practices" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "content": "Here we summarize the best practices for the building block as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 72, + 545, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 118 + ], + "type": "text", + "content": "- **Position Embedding.** Explicit position embedding can endow the models with better position awareness with a low computation cost comparing with other PE manners." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 128, + 545, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 128, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 317, + 128, + 545, + 164 + ], + "type": "text", + "content": "- Point Update. Point update is important and the complexity allocation between the neighbor update and the point update should be properly set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 172, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 545, + 243 + ], + "type": "text", + "content": "- Neighbor Aggregation. Max pooling is not a pure non-learnable aggregation function, its aggregation power is comparable with the learnable aggregation function but with much less computation complexity. Any exploration for a new aggregation function should set max pooling as the baseline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 252, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 252, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 317, + 252, + 545, + 312 + ], + "type": "text", + "content": "- Neighbor Update. Applying the MLP before neighbors grouping will bring a " + }, + { + "bbox": [ + 317, + 252, + 545, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 317, + 252, + 545, + 312 + ], + "type": "text", + "content": " times FLOPs reduction comparing with using the contrary order. However, the position embedding function must be carefully chosen for the position awareness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 320, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 499 + ], + "type": "text", + "content": "The first and the second practice are usually ignored by researchers but are vital for the effectiveness and efficiency of the building block, which is demonstrated in Sec. 4.2 and Sec. 4.4. The third shares a similar viewpoint with PosPool [10] which claims that different local aggregation operators perform similarly. However, from the perspective of PointMeta, we provide a deeper insight that max pooling can be reformulated as a special case of vector self-attention [32], which makes it have comparable aggregation power comparing with the learnable aggregation function. The fourth practice about neighbor update has been emphasized by ASSANet [15], we restate it here because it is important to notice its collaboration with position embedding function. We believe that these practices can provide a guideline for the future research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 507, + 400, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 507, + 400, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 507, + 400, + 519 + ], + "type": "text", + "content": "4.6. PointMetaBase" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 525, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 693 + ], + "type": "text", + "content": "We apply the best practices summarized in Sec. 4.5 to design the building block and implement the four meta functions as simply as possible. We name this block as PointMetaBase, which we hope to become a solid baseline for future research. Fig. 2 (g) shows the architecture of PointMetaBase. The layer number of the first MLP is set to 1 and that of the second MLP is set to 2. The neighbor aggregation function is a simple max pooling. We use the EPE adopted by Point Transformer [32] as the position embedding function. Applying EPE and the MLP-before-Group order, we tweak the set abstraction module [14, 16] as the reduction block, termed PointMetaSA. We adopt the same scaling strategies and decoder with PointNeXt [16] to construct our PointMetaBase family:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 701, + 468, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 468, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 468, + 712 + ], + "type": "text", + "content": "- PointMetaBase-S: " + }, + { + "bbox": [ + 317, + 701, + 468, + 712 + ], + "type": "inline_equation", + "content": "C = 32" + }, + { + "bbox": [ + 317, + 701, + 468, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 317, + 701, + 468, + 712 + ], + "type": "inline_equation", + "content": "B = 0" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17687" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 541, + 319 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 319 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 319 + ], + "type": "table", + "html": "
MethodS3DIS 6-FoldS3DIS Area-5ScanNet mIoUParams. MFLOPs GThroughput (ins./sec.)
mIoU (%)OA (%)mIoU (%)OA (%)Val (%)Test (%)
PointNet++ [14]54.581.053.583.053.555.71.07.2237
PointCNN [8]65.488.157.385.9-45.80.6--
DeepGCN [7]60.085.952.5---3.6--
KPConv [22]70.6-67.1-69.268.615.0--
RandLA-Net [5]70.088.0---64.51.35.8-
BAAF-Net [17]72.288.965.488.9--5.0--
Point Transformer [32]73.590.270.490.870.6-7.85.6-
CBL [20]73.189.669.490.6-70.518.6--
ASSANet [15]--65.888.9--2.42.5300
ASSANet-L [15]--68.089.7--115.636.2136
PointNeXt-S [16]68.087.463.4±0.887.9±0.364.5-0.83.6276
PointNeXt-B [16]71.588.867.3±0.289.4±0.168.4-3.88.9161
PointNeXt-L [16]73.989.869.0±0.590.0±0.169.4-7.115.2109
PointNeXt-XL [16]74.990.370.5±0.390.6±0.171.571.241.684.843
PointMetaBase-L75.690.669.5±0.390.5±0.171.0-2.72.0187
PointMetaBase-XL76.391.071.1±0.490.9±0.171.8-15.39.2104
PointMetaBase-XXL77.091.371.3±0.790.8±0.672.871.419.711.090
", + "image_path": "140a4fe15b827e27b326c5862fbabbcb1a883b84730102789e9211d4f52eddda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 546, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 546, + 351 + ], + "type": "text", + "content": "Table 4. Semantic segmentation on S3DIS [1] (6-Fold and Area 5) and ScanNet V2 [3]. Our PointMetaBase family surpass the state-of-the-art method PointNeXt [16] significantly with large FLOPs reduction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 58, + 370, + 261, + 429 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 370, + 247, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 370, + 247, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 370, + 247, + 383 + ], + "type": "text", + "content": "- PointMetaBase-L: " + }, + { + "bbox": [ + 58, + 370, + 247, + 383 + ], + "type": "inline_equation", + "content": "C = 32" + }, + { + "bbox": [ + 58, + 370, + 247, + 383 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 58, + 370, + 247, + 383 + ], + "type": "inline_equation", + "content": "B = (2,4,2,2)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 393, + 253, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 393, + 253, + 406 + ], + "spans": [ + { + "bbox": [ + 58, + 393, + 253, + 406 + ], + "type": "text", + "content": "- PointMetaBase-XL: " + }, + { + "bbox": [ + 58, + 393, + 253, + 406 + ], + "type": "inline_equation", + "content": "C = 64" + }, + { + "bbox": [ + 58, + 393, + 253, + 406 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 58, + 393, + 253, + 406 + ], + "type": "inline_equation", + "content": "B = (3,6,3,3)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 416, + 261, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 416, + 261, + 429 + ], + "spans": [ + { + "bbox": [ + 58, + 416, + 261, + 429 + ], + "type": "text", + "content": "- PointMetaBase-XXL: " + }, + { + "bbox": [ + 58, + 416, + 261, + 429 + ], + "type": "inline_equation", + "content": "C = 64" + }, + { + "bbox": [ + 58, + 416, + 261, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 58, + 416, + 261, + 429 + ], + "type": "inline_equation", + "content": "B = (4,8,4,4)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 438, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 548 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 438, + 289, + 548 + ], + "type": "text", + "content": " denotes the channel size of the stem MLP and " + }, + { + "bbox": [ + 46, + 438, + 289, + 548 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 438, + 289, + 548 + ], + "type": "text", + "content": " denotes the number of the PointMetaBase block in each stage. Note that when " + }, + { + "bbox": [ + 46, + 438, + 289, + 548 + ], + "type": "inline_equation", + "content": "B = 0" + }, + { + "bbox": [ + 46, + 438, + 289, + 548 + ], + "type": "text", + "content": ", only one PointMetaSA block but no PointMetaBase blocks are used at each stage. Due to the excellent efficiency of PointMetaBase, we do not need to construct the network at the level \"B\" as done in Point-NeXt [16]. In contrast, we choose to scale the model to a larger level \"XXL\". The full details of network architecture are available in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 560, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 128, + 574 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "content": "We studied the performance of PointMetaBase on S3DIS [1], ScanNet V2 [3], ScanObjectNN [24] and ShapeNetPart [30]. To enable a fair comparison, the same data processing and evaluation protocols adopted by the state-of-the-art method PointNeXt [16] were used in our experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Experimental Setups. We optimized all our models using CrossEntropy loss with label smoothing [19], AdamW optimizer, an initial learning rate 0.001, and weight decay " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ", with Cosine Decay, with a 32G V100 GPU. Model parameters, FLOPs and inference throughput are provided for comparison. For a fair comparison, we use the open" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 370, + 546, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 546, + 431 + ], + "type": "text", + "content": "codes provided by PointNext [16] to conduct the measurements. We also use MindSpore to validate the generalization on different deep learning frameworks. All the measurements are conducted using an NVIDIA Tesla A100 40GB GPU and a 12-core Intel Xeon @ 2.40GHz CPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 437, + 440, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 440, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 440, + 451 + ], + "type": "text", + "content": "5.1.3D Scene Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "content": "S3DIS [1] is a challenging benchmark composed of 6 large-scale indoor areas, 271 rooms, and 13 semantic categories in total. The standard 6-fold cross-validation results and the Area-5 results on S3DIS are reported in Tab. 4. Since we adopt the same scaling strategies, PointMetaBase-L and PointMetaBase-XL have the same configuration on width and depth as PointNeXt-L and PointNeXt-XL. With only " + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "content": " FLOPs, our PointMetaBase-L outperforms PointNext-L by " + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "content": " in terms of mean IoU (mIoU) and overall accuracy (OA), respectively, and is much faster in terms of throughput. Compared with PointNext-XL, our PointMetaBase-XL uses only " + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "content": " of the FLOPs, but surpasses with " + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "content": " OA. The inference speed is " + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "inline_equation", + "content": "2.4\\times" + }, + { + "bbox": [ + 304, + 456, + 547, + 696 + ], + "type": "text", + "content": " faster. The excellent efficiency and performance of PointMetaBase can be attributed to the the removal of the computation redundancy in the two update functions and the improved position awareness introduced by explicit position embedding. ScanNet [3] contains 3D indoor scenes of various rooms with 20 semantic categories. The dataset is divided into 3 part: training," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 465, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 465, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 465, + 714 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 317, + 702, + 465, + 714 + ], + "type": "text", + "content": "https://github.com/guochengqian/PointNeXt" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17688" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 497, + 242 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 497, + 242 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 497, + 242 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 497, + 242 + ], + "type": "table", + "html": "
Methodins. mIoU (%)cls. mIoU (%)Params (M)FLOPs (G)Throughput (ins./sec)
PointNet++ [14]85.181.91.04.9-
KPConv [22]86.485.1---
ASSANet [15]86.1----
Point Transformer [32]86.683.77.8--
PointMLP [12]86.184.6---
StratifiedFormer [6]86.685.1---
PointNeXt-S [16]86.7±0.084.4±0.21.04.5873
PointNeXt-S (C=64) [16]86.9±0.184.8±0.53.717.8391
PointNeXt-S (C=160) [16]87.0±0.185.2±0.122.5110.2118
PointMetaBase-S (ours)86.7±0.084.3±0.11.01.391194
PointMetaBase-S (C=64) (ours)86.9±0.184.9±0.23.83.85706
PointMetaBase-S (C=160) (ours)87.1±0.085.1±0.322.718.45271
", + "image_path": "edadc472c09b94023cc0a7769a5805f4f822370bbe2a5b9415736ce0b6673545.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 291, + 280, + 408 + ], + "blocks": [ + { + "bbox": [ + 46, + 250, + 546, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 546, + 273 + ], + "type": "text", + "content": "Table 5. Part segmentation on ShapeNetPart [30]. Compaerd with the state-of-the-art method PointNeXt [16], our PointMetaBase family achieve comparable performance with large FLOPs reduction and faster inference speed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 291, + 280, + 408 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 280, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 280, + 408 + ], + "type": "table", + "html": "
VariantsOA (%)FLOPs (G)TP (ins./sec)
PointNet++ [14]77.91.72655
PointCNN [8]78.5--
DGCNN [27]78.14.8735
DRNet [13]80.3--
PointMLP [12]85.4±1.331.3405
PointNeXt-S [16]87.7±0.41.72230
PointMetaBase-S87.9±0.20.62674
", + "image_path": "357e88776e7e3329831d6c65160e6ca6b7451f5e6ab604c51fe5c40f13600968.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "type": "text", + "content": "Table 6. Classification on ScanObjectNN [24]. Our model surpasses PointNeXt-S [16] both with much less computation cost." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": "validation, and test splits, with 1201, 312 and 100 scans, respectively. Following PointNeXt [16], we do not use any voting strategies. As shown in Tab. 4, our PointMetaBase-L surpasses PointNeXt-L " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " mIoU with only " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " FLOPs and " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "1.7\\times" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " faster inference speed. PointMetaBase-XL surpasses PointNeXt-XL " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " mIoU with only " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "11.3\\%" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " FLOPs and " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "2.4\\times" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " faster inference speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 549, + 185, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 185, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 185, + 562 + ], + "type": "text", + "content": "5.2. 3D Object Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 665 + ], + "type": "text", + "content": "ScanObjectNN [24] contains about 15000 real scanning objects, which are grouped into 15 classes and have 2902 unique instances. Following [12, 16], we experiment on the hardest perturbed variant (PB_T50_RS). To make a fair comparison with PointNeXt [16], we did not use upscaled variants of PointMetaBase on this benchmark. As reported in Tab. 6, our model surpasses PointNeXt-S, while using much fewer model FLOPs and running much faster." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 671, + 174, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 174, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 174, + 684 + ], + "type": "text", + "content": "5.3. 3D Part Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "ShapeNetPart [30] is a dataset for part segmentation of objects. It consists of 16880 models with 16 differ" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "text", + "content": "ent shape categories. Following the state-of-the-art method PointNeXt [16], we used voting by averaging the results of 10 randomly scaled input point clouds, with scaling factors equal to [0.8,1.2]. As shown in Tab. 5, compaerd with PointNeXt, our PointMetaBase family achieve comparable performance with large FLOPs reduction and faster inference speed. Specially, for " + }, + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "inline_equation", + "content": "C = 32 / C = 64 / C = 160" + }, + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "text", + "content": ", the FLOPs of our PointMetaBase are only " + }, + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "inline_equation", + "content": "30\\% 21\\% / 17\\%" + }, + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "text", + "content": " that of PointNeXt, and the inference speed of PointMetaBase is " + }, + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "inline_equation", + "content": "1.4 / 1.8 / 2.3" + }, + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "text", + "content": " times faster than that of PointNeXt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 428, + 379, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 379, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 379, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 581 + ], + "type": "text", + "content": "In this paper, we reformulate existing models for point cloud analysis into a general framework, PointMeta. Within this framework, we provide an in-depth analysis on the building blocks of existing models and summarize a few best practices. Furthermore, we follow these best practices and propose a simple building block, PointMetaBase, which enjoys excellent performance and efficiency. Our work can inspire new insights into building block design and future architecture exploration. Next, building on this work, we plan to explore neural architecture search methods [33-37] that are specifically tailored for point cloud analysis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 581, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by National Key R&D Program of China (No. 2022ZD0118202), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J01002, No. 2022J06001), Guangdong Basic and Applied Basic Research Foundation (No. 2019B1515120049), and CAAI-Huawei MindSpore Open Fund." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17689" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 94, + 106, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 94, + 106, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 94, + 106, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 114, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "type": "text", + "content": "[1] Iro Armeni, Sasha Sax, Amir R. Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. CoRR, abs/1702.01105, 2017. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 286, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 256 + ], + "type": "text", + "content": "[2] Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Flores Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gulçehre, H. Francis Song, Andrew J. Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, Kelsey R. Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matthew M. Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational inductive biases, deep learning, and graph networks. CoRR, abs/1806.01261, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 258, + 286, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 311 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 311 + ], + "type": "text", + "content": "[3] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 2432–2443. IEEE Computer Society, 2017. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 313, + 286, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 286, + 346 + ], + "type": "text", + "content": "[4] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 346, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 286, + 390 + ], + "type": "text", + "content": "[5] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11105-11114, 2020. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 391, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 286, + 434 + ], + "type": "text", + "content": "[6] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8490-8499. IEEE, 2022. 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 435, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 286, + 468 + ], + "type": "text", + "content": "[7] Guohao Li, Matthias Müller, Ali K. Thabet, and Bernard Ghanem. Deep GCs: Can GCs go as deep as cnns? In ICCV, pages 9266-9275, 2019. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 286, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 286, + 502 + ], + "type": "text", + "content": "[8] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, pages 828-838, 2018. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 286, + 534 + ], + "type": "text", + "content": "[9] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In CVPR, pages 8895-8904, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 568 + ], + "type": "text", + "content": "[10] Ze Liu, Han Hu, Yue Cao, Zheng Zhang, and Xin Tong. A closer look at local aggregation operators in point cloud analysis. In ECCV, pages 326-342, 2020. 2, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 286, + 612 + ], + "type": "text", + "content": "[11] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002, 2021. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "type": "text", + "content": "[12] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 4, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[13] Charles Ruizhongtai Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 77-85, 2017. 1, 2, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[14] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 151 + ], + "type": "text", + "content": "[15] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali K. Thabet, and Bernard Ghanem. Assanet: An anisotropic separable set abstraction for efficient point cloud representation learning. In NeurIPS, pages 28119-28130, 2021. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 153, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 208 + ], + "type": "text", + "content": "[16] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Abed Al Kader Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 210, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 210, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 210, + 545, + 243 + ], + "type": "text", + "content": "[17] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "type": "text", + "content": "[18] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, ICLR, 2015. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 280, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 280, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 280, + 545, + 322 + ], + "type": "text", + "content": "[19] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 325, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 325, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 325, + 545, + 358 + ], + "type": "text", + "content": "[20] Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. In CVPR, pages 8479-8489, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "type": "text", + "content": "[21] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, pages 3887-3896, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 395, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 545, + 438 + ], + "type": "text", + "content": "[22] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6410–6419, 2019. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 441, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 545, + 484 + ], + "type": "text", + "content": "[23] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 487, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 540 + ], + "type": "text", + "content": "[24] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, pages 1588–1597, 2019. 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 543, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 576 + ], + "type": "text", + "content": "[25] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, pages 10296-10305, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "type": "text", + "content": "[26] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In CVPR, pages 2589-2597, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[27] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. ACM Trans. Graph., 38(5):146:1-146:12, 2019. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[28] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In ICCV, pages 10013-10021, 2021. 3, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17690" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 497 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[29] Wenxuan Wu, Zhongang Qi, and Fuxin Li. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, pages 9621-9630, 2019. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "text", + "content": "[30] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Trans. Graph., 35(6):210:1-210:12, 2016. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 196 + ], + "type": "text", + "content": "[31] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, pages 5565-5573, 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 229 + ], + "type": "text", + "content": "[32] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16239-16248, 2021. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "text", + "content": "[33] Xiawu Zheng, Xiang Fei, Lei Zhang, Chenglin Wu, Fei Chao, Jianzhuang Liu, Wei Zeng, Yonghong Tian, and Rongrong Ji. Neural architecture search with representation mutual information. In CVPR, pages 11912-11921, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 286, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 339 + ], + "type": "text", + "content": "[34] Xiawu Zheng, Rongrong Ji, Yuhang Chen, Qiang Wang, Baochang Zhang, Jie Chen, Qixiang Ye, Feiyue Huang, and Yonghong Tian. Migo-nas: Towards fast and generalizable neural architecture search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9):2936-2952, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 342, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 286, + 384 + ], + "type": "text", + "content": "[35] Xiawu Zheng, Rongrong Ji, Lang Tang, Baochang Zhang, Jianzhuang Liu, and Qi Tian. Multinomial distribution learning for effective neural architecture search. In ICCV, 2019. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "type": "text", + "content": "[36] Xiawu Zheng, Rongrong Ji, Qiang Wang, Qixiang Ye, Zhenguo Li, Yonghong Tian, and Qi Tian. Rethinking performance estimation in neural architecture search. In CVPR, 2020. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 432, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 432, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 49, + 432, + 286, + 497 + ], + "type": "text", + "content": "[37] Xiawu Zheng, Yang Zhang, Sirui Hong, Huixia Li, Lang Tang, Youcheng Xiong, Jin Zhou, Yan Wang, Xiaoshuai Sun, Pengfei Zhu, Chenglin Wu, and Rongrong Ji. Evolving fully automated machine learning via life-long knowledge anchors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(9):3091-3107, 2021. 8" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17691" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_content_list.json b/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a336b69dbbe1beb7f0fd7f0f3db486899980e58 --- /dev/null +++ b/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_content_list.json @@ -0,0 +1,1415 @@ +[ + { + "type": "text", + "text": "Meta Compositional Referring Expression Segmentation", + "text_level": 1, + "bbox": [ + 197, + 130, + 772, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Li Xu $^{1}$ , Mark He Huang $^{1,3}$ , Xindi Shang $^{2}$ , Zehuan Yuan $^{2}$ , Ying Sun $^{3}$ , Jun Liu $^{1*}$ $^{1}$ Singapore University of Technology and Design, Singapore \n $^{2}$ ByteDance", + "bbox": [ + 186, + 179, + 816, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Institute for Infocomm Research (I $^2$ R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore", + "bbox": [ + 88, + 233, + 911, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{li_xu, he_huang}@mymail.sutd.edu.sg shangxindi, yuanzehuan}@bytedance.com @i2r.a-star.edu.sg, jun_liu@sutd.edu.sg", + "bbox": [ + 344, + 253, + 684, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring expression segmentation aims to segment an object described by a language expression from an image. Despite the recent progress on this task, existing models tackling this task may not be able to fully capture semantics and visual representations of individual concepts, which limits their generalization capability, especially when handling novel compositions of learned concepts. In this work, through the lens of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework to enhance model compositional generalization performance. Specifically, to handle various levels of novel compositions, our framework first uses training data to construct a virtual training set and multiple virtual testing sets, where data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set. Then, following a novel meta optimization scheme to optimize the model to obtain good testing performance on the virtual testing sets after training on the virtual training set, our framework can effectively drive the model to better capture semantics and visual representations of individual concepts, and thus obtain robust generalization performance even when handling novel compositions. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our framework.", + "bbox": [ + 75, + 371, + 473, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 207, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring expression segmentation (RES) [12, 38, 40] aims to segment a visual entity in an image given a linguistic expression. This task has been receiving increasing attention in recent years [5, 18, 34, 37], as it can play an important role in various applications, such as language-based human-robot interaction and interactive image edit", + "bbox": [ + 75, + 786, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/275a82f17cd616dbd8a9ebcec3ad7080ee08243791991194e43b43d145fd3385.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 532, + 339, + 859, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/385f94c0a861a9b6c4e80d8271248fa0bd5b0956d5cb67b60b7d1b777ee9a8db.jpg", + "image_caption": [ + "(b)", + "Figure 1. Illustration of novel compositions and various levels of compositions. (a) An example of the testing sample containing the novel composition of \"dark coffee\" in RefCOCO dataset [41]. Such a novel composition itself does not exist in the training data, but its individual components (i.e., \"dark\" and \"coffee\") exist in the training data. (b) An example of various levels of compositions in an expression. We perform constituency parsing of the expression using AllenNLP [9]. Based on the obtained parsing tree as shown above, we can then obtain various levels of compositions (e.g., word-word level, word-phrase level) in this expression." + ], + "image_footnote": [], + "bbox": [ + 537, + 469, + 857, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing. However, despite the recent progress on tackling this task [5, 34, 42], existing methods may struggle with handling the testing samples, of which the expressions contain novel compositions of learned concepts. Here a novel composition means that the composition itself does not exist in the training data, but its individual components (e.g., words, phrases) exist in the training data, as shown in Fig. 1a.", + "bbox": [ + 496, + 702, + 890, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We observe that testing samples containing such novel compositions of learned concepts widely exist in RES datasets [26, 28, 41]. However, existing RES models may not be able to well handle novel compositions during testing. Here we test the generalization capability of multiple state-of-the-art models [25, 37, 42] in terms of handling", + "bbox": [ + 496, + 810, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19478", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "novel compositions, as shown in Table 2. Specifically, we first split each testing set of the RefCOCO dataset. In each testing set, one split subset includes the data samples, in which all the contained compositions are seen in the RefCOCO training set. While another subset includes the data samples containing novel compositions, of which the individual components (e.g., words, phrases) exist in the RefCOCO training set but the composition itself is unseen in the training set, i.e., containing novel compositions of learned concepts. Then we evaluate the models [25, 37, 42] on the two subsets in each testing set, and find that for each model, its testing performance on the subset containing novel compositions drops obviously compared to the performance on the other subset. For these models, the performance gap between the two subsets can reach $14\\% - 17\\%$ measured by the metric of overall IoU. Such a clear performance gap indicates that existing models struggle with generalizing to novel compositions of learned concepts. This might due to that the model does not effectively capture the semantics and visual representations of individual concepts (e.g., \"dark\", \"coffee\" in Fig. 1a) during training. Then the trained model may fail to recognize a novel composition (e.g., \"dark coffee\") at testing time, which though is composed of learned concepts.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus to handle this issue, we aim to train the model to effectively capture the semantics and visual representations of individual concepts during training. Despite the conceptual simplicity, how to guide the model's learning behavior towards this goal is a challenging problem. Here from the perspective of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework, to effectively handle such a challenging problem by only changing the model training scheme.", + "bbox": [ + 76, + 455, + 472, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meta learning proposes to perform virtual testing during model training for better performance [7, 29]. Inspired by this, to improve the generalization capability of RES models, our MCRES framework incorporates a meta optimization scheme that consists of three steps: virtual training, virtual testing and meta update. Specifically, we first split the training set to construct a virtual training set for virtual training, and a virtual testing set for virtual testing. The data samples in the virtual testing set contain novel compositions w.r.t. the virtual training set. For example, if the expressions of data samples in the virtual training set contain both words \"dark\" and \"coffee\" but do not contain their composition (i.e., \"dark coffee\"), the virtual testing set can include this novel composition correspondingly.", + "bbox": [ + 76, + 595, + 472, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the constructed virtual training set and virtual testing set, we first train the model using the virtual training set, and then evaluate the trained model on the virtual testing set. During virtual training, the model may learn the compositions of individual concepts as a whole without truly understanding the semantics and visual representa", + "bbox": [ + 76, + 809, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions of individual concepts, which though can still improve model training performance. For example, if there are many training samples containing the composition of \"yellow banana\" in the virtual training set, the model can superficially correlate \"banana\" with \"yellow\" and learn this composition as a whole, since using such spurious correlations can facilitate the model learning [1, 10, 39]. However, learning the compositions as a whole over the virtual training set may not improve model performance much on the virtual testing set in virtual testing, since the virtual testing set contains novel compositions w.r.t. the virtual training set. Thus to achieve good testing performance on such a virtual testing set, the model needs to effectively capture semantics and visual representations of individual concepts during virtual training. In this way, the model testing performance on the virtual testing set serves as a generalization feedback to the model virtual training process.", + "bbox": [ + 496, + 90, + 893, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus after the virtual training and virtual testing, we can further update the model to obtain better testing performance on the virtual testing set (i.e., meta update), so as to drive the model training on the virtual training set towards the direction of learning to capture semantics and visual representations of individual concepts, i.e., learning to learn. In this manner, our framework is able to optimize the model for robust generalization performance, even tackling the challenging testing samples with novel compositions.", + "bbox": [ + 496, + 349, + 893, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, given that expressions can often be hierarchically decomposed, there can exist various levels of novel compositions. Specifically, to identify meaningful compositions in an expression, we can parse an expression into a tree structure based on the constituency parsing tool [9] as shown in Fig.1b. In such a parsing tree, under the same parent node, each pair of child nodes (e.g., \"white\" and \"bird\") are closely semantically related, and thus can form a meaningful composition. Since each child node can be a word or a phrase as in Fig. 1b, there can naturally exist the following three levels of novel compositions: word-word level (e.g., \"white\" and \"bird\"), word-phrase level (e.g., \"standing\" and \"in water\") and phrase-phrase level (e.g., \"white bird\" and \"standing in water\"), which correspond to different levels of comprehension complexity. To better handle such a range of novel compositions, we construct multiple virtual testing sets in our framework, where each virtual testing set is constructed to handle one level of novel compositions.", + "bbox": [ + 496, + 488, + 895, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework only changes the model training scheme without the need to change the model structure. Thus our framework is general, and can be conveniently applied on various RES models. We test our framework on multiple models, and obtain consistent performance improvement.", + "bbox": [ + 496, + 762, + 893, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our work are threefold: 1) We propose a novel framework (MCRES) to effectively improve generalization performance of RES models, especially when handing novel compositions of learned con", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19479", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cepts. 2) Via constructing a virtual training set and multiple virtual testing sets w.r.t. various levels of novel compositions, our framework can train the model to well handle various levels of novel compositions. 3) When applied on various models on three RES benchmarks [28, 41], our framework achieves consistent performance improvement.", + "bbox": [ + 75, + 90, + 470, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 196, + 227, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Referring Expression Segmentation (RES). RES [12] aims to segment a target object from an image based on an expression. Early works [20, 22, 28, 41] employed convolutional and recurrent networks to extract visual and linguistic features respectively, and then fused the extracted features to predict the segmentation mask. Recently, a series of works [5, 18, 37] employed vision and language transformers to boost performance. Besides, with the development of large-scale pretrained models, Wang et al. [34] proposed to leverage CLIP [30] to improve cross-modal matching.", + "bbox": [ + 75, + 223, + 468, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some methods have been explored to help model better understand learned concepts in RES. Yu et al. [40] proposed a modular network that uses different modules to process different types of information in the given expression. Yang et al. [36] designed a reasoning module to help align the language concepts with visual regions. Different from all the above-mentioned works, we propose an MCRES framework to drive RES models to better capture semantics and visual representations of individual concepts. Such a framework only changes the model training scheme, and thus can be flexibly applied on various models to improve their generalization performance.", + "bbox": [ + 75, + 375, + 470, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta Learning. Meta learning, i.e., the paradigm of learning to learn, has emerged to mainly tackle the few-shot learning problem [7, 29, 31, 32]. MAML [7] and its following works [29, 31] aim to learn a good initialization of network parameters, to achieve fast test-time update to adapt to new few-shot learning tasks. More recently, meta learning has also been explored in other areas [8, 11, 14, 19, 35] to enhance model generalization performance without the need of test-time update. Inspired by these works, we leverage a meta learning-based framework to improve generalization performance of RES models especially when handling novel compositions.", + "bbox": [ + 75, + 556, + 470, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 753, + 168, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing RES models may fail to generalize to data samples containing novel compositions of learned concepts. To handle this issue, we aim to encourage the model to better capture the semantics of individual concepts as well as recognize their visual correspondences during training, which however is a non-trivial problem. In this paper, from the perspective of meta learning, we introduce MCRES framework optimizing RES models via a meta optimization", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scheme to improve their generalization performance.", + "bbox": [ + 498, + 90, + 849, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 2, in our framework, we first split the original training set $(\\mathcal{D}_{train})$ to build a virtual training set $(\\mathcal{D}_{v\\_ tr})$ for virtual training, and a group of $K$ virtual testing sets $\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)$ for virtual testing. Each virtual testing set consists of data samples containing one level of novel compositions w.r.t. the virtual training set. We first use the virtual training set to train the model (virtual training), and then perform model testing on the virtual testing sets (virtual testing). Since the virtual testing sets contain novel compositions of individual concepts from the virtual training set, if the model can still achieve robust testing performance on the virtual testing sets after training on the virtual training set, we posit that the trained model has learned to capture more semantics and visual representations of individual concepts during the virtual training. Guided by this principle, we can optimize the model virtual testing performance, to guide the model virtual training process towards learning more semantics and representations of individual concepts. Below, we first introduce the pipeline of our framework, and then detail the virtual training set and virtual testing sets construction process.", + "bbox": [ + 496, + 107, + 890, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework", + "text_level": 1, + "bbox": [ + 500, + 436, + 627, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first train the RES model using the virtual training set, i.e., virtual training. Specifically, we denote the model parameters as $\\theta$ , the loss function (e.g., cross-entropy loss) for training the RES model as $\\mathcal{L}$ . Thus we can calculate the model virtual training loss $(\\mathcal{L}_{v\\_ tr})$ over the virtual training set $(\\mathcal{D}_{v\\_ tr})$ as:", + "bbox": [ + 496, + 460, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v \\_ t r} (\\theta) = \\mathcal {L} (\\theta ; \\mathcal {D} _ {v \\_ t r}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 559, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this loss, we then update the model parameters $(\\theta)$ as follows:", + "bbox": [ + 498, + 582, + 890, + 611 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {\\prime} = \\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v _ {-} t r} (\\theta) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 614, + 890, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ denotes the learning rate. It is worth noting that the model update at this step is virtual, and the updated parameters $\\theta'$ only serve as intermediate parameters for the model evaluation at the following virtual testing step.", + "bbox": [ + 496, + 642, + 890, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After the virtual training, we evaluate the generalization performance of the trained model on the virtual testing sets $\\left(\\{\\mathcal{D}_{v\\_te}^k\\}_{k = 1}^K\\right)$ , i.e., virtual testing. Specifically, on each virtual testing set $\\mathcal{D}_{v\\_te}^k$ , we compute the model loss $\\mathcal{L}_{v\\_te}^k$ as:", + "bbox": [ + 498, + 703, + 890, + 765 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v. t e} ^ {k} \\left(\\theta^ {\\prime}\\right) = \\mathcal {L} \\left(\\theta^ {\\prime}; \\mathcal {D} _ {v. t e} ^ {k}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 768, + 890, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Such a loss measures how well the model generalizes to the virtual testing set after training on the virtual training set.", + "bbox": [ + 496, + 792, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then we perform the meta update step. At this step, we wish to actually update the model parameters $(\\theta)$ , so that after the model training on the virtual training set, the trained model can generalize well to the virtual testing sets, i.e., generalizing well to novel compositions. To this end, we", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19480", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a4b218637b9ae77833b8d3d5ab604e43a0412f521608cc2beeb21c414e829b77.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 467, + 329 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e801d30b4da5eed6ad8ff7092a0717d0e9fd40ac124d0bfb9c67d3a8aaf55b78.jpg", + "image_caption": [ + "(b)", + "Figure 2. Framework overview. Fig. (a) illustrates our framework pipeline. First, we train the model using the virtual training set $(\\mathcal{D}_{v\\_ tr})$ and then obtain the updated model. We then test the model with updated parameters $(\\theta^{\\prime})$ on multiple virtual testing sets $\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)$ . According to the virtual testing losses, we perform meta update to optimize the model for better generalization capability. Fig. (b) shows that we construct a virtual training set and multiple virtual testing sets to handle various levels of novel compositions. The expressions of data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set." + ], + "image_footnote": [], + "bbox": [ + 526, + 92, + 856, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "formulate the optimization objective as:", + "bbox": [ + 76, + 426, + 341, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} _ {v. t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v. t e} ^ {k} \\left(\\theta^ {\\prime}\\right) \\tag {4} \\\\ = \\min _ {\\theta} \\mathcal {L} _ {v - t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v - t e} ^ {k} \\left(\\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v - t r} (\\theta)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 446, + 468, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the first term indicates the model training performance on the virtual training set, and the second term represents the model testing performance on the virtual testing sets after training on the virtual training set. Based on this objective, we then update the model parameters $(\\theta)$ as:", + "bbox": [ + 75, + 539, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\left(\\mathcal {L} _ {v _ {-} t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v _ {-} t e} ^ {k} \\left(\\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v _ {-} t r} (\\theta)\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 621, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is the learning rate for meta update. Through the above optimization process, the model is pushed to learn more semantics and visual representations of individual concepts during training.", + "bbox": [ + 75, + 675, + 468, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the above process, the model is first trained (updated) on the virtual training set. At this step, the model may learn the compositions of individual concepts as a whole without effectively capturing the semantics and visual representations of individual concepts, which though can still improve its training performance on the virtual training set. However, to achieve good testing performance on the virtual testing data which contains novel compositions of individual concepts from the virtual training data, the model is expected to avoid learning compositions as a", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whole and instead capture more semantics and representations of individual concepts. In this way, the second term of Eqn. 5, which includes the second-order gradients of $\\theta$ : $\\nabla_{\\theta}\\mathcal{L}_{v\\_te}^{k}\\big(\\theta -\\alpha \\nabla_{\\theta}\\mathcal{L}_{v\\_tr}(\\theta)\\big)$ , can be regarded as a generalization feedback that can guide the model to capture more semantics and representations of individual concepts.", + "bbox": [ + 496, + 426, + 890, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our framework, the above three steps (i.e., virtual training, virtual testing and meta update) are performed iteratively until the model training converges.", + "bbox": [ + 496, + 517, + 890, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sets Construction", + "text_level": 1, + "bbox": [ + 500, + 574, + 671, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our framework, to handle various levels of novel compositions, we split the original training set to construct a virtual training set and various virtual testing sets. Each virtual testing set is expected to contain one level of novel compositions w.r.t. the virtual training set. Below we first introduce the details of these sets, and then discuss the strategy for constructing the virtual testing sets.", + "bbox": [ + 496, + 597, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We randomly sample a subset of the training set $(\\mathcal{D}_{train})$ as the virtual training set $(\\mathcal{D}_{v\\_ tr})$ , and the remaining training data will be used as the candidate samples to construct a group of $(K)$ virtual testing sets $\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)$ . Each virtual testing set is constructed to handle one level of novel compositions. Here considering the hierarchical semantic structure of each expression, we target the following levels of novel compositions: word-word level, word-phrase level and phrase-phrase level. A phrase means a group of words that serve as a grammatical unit in the expression (e.g., \"white bird\" in Fig. 1b), which can be conveniently identified using off-the-shelf tools (e.g., AllenNLP [9]).", + "bbox": [ + 496, + 704, + 892, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, to identify various levels of novel compo", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19481", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sitions in an expression, we first use the constituency parsing tool [9] to parse an expression into a tree structure as shown in Fig. 1b. In this tree, under the same parent node, each pair of child nodes are closely related, and thus can form a meaningful composition. If a pair of nodes having the same parent node are both words, they form a word-word level composition (e.g., \"white\" and \"bird\"). Similarly, if the nodes with the same parent node are a word and a phrase, they form a word-phrase level composition (e.g., \"standing\" and \"in water\"). In a similar way, we can obtain phrase-phrase level compositions (e.g., \"white bird\" and \"standing in water\"). Thus there can exist three levels of novel compositions, which correspond to different levels of comprehension complexity. Corresponding to these three levels of novel compositions, we will construct a total of three virtual testing sets $(K = 3)$ .", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Virtual testing sets construction. To construct each virtual testing set, we need to select data samples containing the corresponding level of novel compositions w.r.t. the virtual training set, from the candidate samples. A data sample means an expression paired with the corresponding image. Here we design an efficient strategy, which can be leveraged to construct each virtual testing set. Below, we take the process of constructing the virtual testing set for handling word-word level novel compositions as an example, to introduce such a strategy. Specifically, this virtual testing set should include all of those candidate samples that contain word-word level novel compositions w.r.t. the virtual training set. To construct such a virtual testing set, our strategy proceeds as follows.", + "bbox": [ + 75, + 337, + 472, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(i) To find the data samples, of which the expressions contain word-word level novel compositions w.r.t. the virtual training set, from the candidate samples, we need to first obtain all the word-word level compositions in the virtual training set and in the candidate samples respectively. Thus by parsing each expression into a parsing tree as shown in Fig. 1b, we can obtain all word-word level compositions (e.g., \"white\" and \"bird\", \"in\" and \"water\" in Fig. 1b) in each expression. (ii) To identify word-word level novel compositions, we first select the word-word level compositions that exist in the candidate samples but are unseen in the virtual training set. Then for each selected composition, we further check if its individual words exist in the virtual training set. If so, such a composition will be identified as a word-word level novel composition w.r.t. the virtual training set. For example, if the composition \"white bird\" is unseen in the virtual training set, while its individual words \"white\" and \"bird\" both exist in the virtual training set, this composition is a word-word level novel composition w.r.t. the virtual training set. (iii) Finally, we select the candidate samples, of which the expressions contain the identified word-word level novel compositions w.r.t. the virtual training set, to construct the virtual testing set.", + "bbox": [ + 75, + 553, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To efficiently perform the above steps, we can employ parallel matrix operations. Specifically, to record the word-word level compositions in the virtual training set, we build a matrix $\\mathcal{M}_{v,tr}$ with the shape of $|\\mathcal{V}| \\times |\\mathcal{V}|$ , where $|\\mathcal{V}|$ is the size of the vocabulary set $\\mathcal{V}$ of the original training set. In this matrix, each element at the $i-th$ row and the $j-th$ column $(i,j \\in \\{1,\\dots,|\\mathcal{V}|\\})$ is a binary value, and the value 1 indicates that the $i-th$ word and $j-th$ word in the vocabulary set $\\mathcal{V}$ form a word-word level composition in the virtual training set, while 0 means such a word-word level composition does not exist in the virtual training set. Similarly, we can also build a matrix $\\mathcal{M}_{candi}$ to record the word-word level compositions in the candidate samples. Then to identify word-word level novel compositions w.r.t. the virtual training set $(\\mathcal{M}_{v,tr})$ from the candidate samples $(\\mathcal{M}_{candi})$ , we can efficiently compute a difference matrix: $\\mathcal{M}_{diff} = \\mathcal{M}_{candi} - \\mathcal{M}_{v,tr}$ . In $\\mathcal{M}_{diff}$ , any element with the value 1 indicates that the corresponding composition exists in the candidate samples, but is unseen in the virtual training set. Then by further checking whether the individual words in such a composition exist in the virtual training set, we can determine if it is a word-word level novel composition w.r.t. the virtual training set.", + "bbox": [ + 496, + 90, + 893, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we can efficiently construct the other two virtual testing sets. Note that to help the model to learn to handle a wide range of possible novel compositions, at the beginning of each training epoch, we randomly sample a subset of the training data to re-construct the virtual training set, and leverage the remaining data to re-construct the virtual testing sets using the above strategy.", + "bbox": [ + 496, + 439, + 893, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed above, to handle various levels of novel compositions, we construct multiple virtual testing sets. A simpler alternative is to construct only one virtual testing set, which includes all the data samples containing any level of novel compositions w.r.t. the virtual training set. However, compared to this alternative, by constructing multiple virtual testing sets, each level of novel compositions in the corresponding virtual testing set can offer an explicit generalization feedback during model training. Thus the model is explicitly encouraged to well handle each level of novel compositions. Moreover, constructing multiple virtual testing sets can facilitate the use of curriculum learning strategy for model training as discussed in Sec. 3.3.", + "bbox": [ + 496, + 546, + 895, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training and Testing", + "text_level": 1, + "bbox": [ + 500, + 755, + 697, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To help models generalize to novel compositions in RES, our framework only changes the model training scheme. Thus our framework is general, and can be flexibly applied to train various RES models. During training, at each epoch, we first split the training set to construct a virtual training set and multiple virtual testing sets. Then we iteratively optimize the model over the virtual training set and virtual testing sets via meta optimization. Specifically, for each meta", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19482", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/45f90d9dadc0ff43646abb0cbad84a9083ffd30c6dad386f2d193981c145471e.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-arts on three benchmark datasets using the metric of overall IoU. Moreover, we apply our framework on various models [25, 37, 42], and obtain consistent performance improvement. “-” indicates that the corresponding result is not provided in the original paper. U means the UMD split of RefCOCOg dataset, and G means the Google split." + ], + "table_footnote": [], + "table_body": "
MethodRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBval(U)test(U)val(G)
DMN [27]49.7854.8345.1338.8844.2232.29--36.76
RRN [20]55.3357.2653.9339.7542.1536.11--36.45
MAAttNet [40]56.5162.3751.7046.6752.3940.0847.6448.61-
CAC [4]58.9061.7753.81---46.3746.9544.32
CMSA [38]58.3260.6155.0943.7647.6037.89--39.98
STEP [3]60.0463.4657.9748.1952.33404.1--46.40
BRINet [13]60.9862.9959.2148.1752.3242.11--48.04
CMPC [15]61.3664.5359.6449.5653.4443.23--49.05
LSCM [16]61.4764.9959.5549.3453.1243.50--48.05
CMPC+ [23]62.4765.0860.8250.2554.0443.47--49.89
EFN [6]62.7665.6959.6751.5055.2443.01--51.93
BUSNet [36]63.2766.4161.3951.7656.8744.13--50.56
CGAN [24]64.8668.0462.0751.0355.5144.0651.0151.6946.54
LTS [17]65.4367.7663.0854.2158.3248.0254.4054.25-
VLT [5]65.6568.2962.7355.5059.2049.3652.9956.6549.76
ReSTR [18]67.2269.3064.4555.7860.4448.27--54.48
CRIS [34]70.4773.1866.1062.2768.0853.6859.8760.36-
MCN [25]62.4464.2059.7150.6254.9944.6949.2249.40-
MCN + Ours64.5166.4862.8452.2956.7146.7251.4951.63-
SeqTR [42]71.7073.3169.8263.0466.7358.9764.6965.74-
SeqTR + Ours73.2375.0171.9564.7167.8560.8566.7767.48-
LAVT [37]72.7375.8268.7962.1468.3855.1061.2462.0960.50
LAVT + Ours74.9276.9870.8464.3269.6856.6463.5164.9061.63
", + "bbox": [ + 80, + 172, + 468, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimization iteration, we use two batches of data samples: one batch of data for virtual training and the other batch of data for virtual testing. Thus we need to ensure that the batch of data for virtual testing consists of the data samples containing novel compositions w.r.t. the batch of data for virtual training. To this end, for each iteration, we randomly sample some data from each virtual testing set, to form the batch of data for virtual testing. Then we select the virtual training set samples that contain all the individual components of the novel compositions in the batch of data for virtual testing, to form the batch of data for virtual training. Note that the above data preparation procedures can be done before the model training. For model testing, we test the trained model in the conventional way.", + "bbox": [ + 75, + 476, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, given that the various levels of novel compositions correspond to different levels of comprehension complexity, we adopt a curriculum learning strategy [2, 33] for model training, so that the model can progressively learn to handle various levels of novel compositions, i.e., from lower level (word-word) to middle level (word-phrase) to higher level (phrase-phrase), and thus can learn all these levels of novel compositions better. Specifically, in the first $1/3$ of training epochs, we only use the virtual testing set for handling word-word level novel compositions for meta optimization. Then in the middle $1/3 - 2/3$ of training epochs, we add the virtual testing set for handling word-phrase level novel compositions. Finally, in the remaining training epochs, we use all the three virtual testing sets.", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/07ddd9d10ba740c332bba98b1ae7f2d039680e657b33e8978693c8363ce33795.jpg", + "table_caption": [ + "Table 2. We evaluate our framework on different testing subsets w.r.t. novel compositions to validate its effectiveness on optimizing the model to better generalize to novel compositions. We use the metric of overall IoU here. The performance gain of our framework compared to the corresponding baseline model is shown in parentheses." + ], + "table_footnote": [], + "table_body": "
MethodRefCOCO-valRefCOCO-testARefCOCO-testB
NovelNon-novelNovelNon-novelNovelNon-novel
MCN [23]53.1767.4155.4370.0850.6466.47
MCN + Ours57.38 (↑4.21)68.09 (↑0.68)59.66 (↑4.23)70.67 (↑0.59)55.87 (↑5.23)67.04 (↑0.57)
SeqTR [40]64.2478.5965.7179.4660.1477.21
SeqTR + Ours67.31 (↑3.07)79.06 (↑0.47)69.48 (↑3.77)79.84 (↑0.38)64.89 (↑4.75)77.68 (↑0.47)
LAVT [35]63.5278.5567.1780.7158.4976.23
LAVT + Ours67.42 (↑3.90)79.29 (↑0.74)70.05 (↑2.88)81.37 (↑0.66)62.63 (↑4.14)76.92 (↑0.69)
", + "bbox": [ + 503, + 170, + 890, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 261, + 633, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on three commonly used RES benchmarks: RefCOCO [41], RefCOCO+ [41] and RefCOCOg [26, 28]. Images in these three benchmarks are all from the COCO dataset [21]. RefCOCO [41] contains 19994 images, 50000 annotated objects with 142209 referring expressions. RefCOCO+ [41] consists of 19992 images with 49856 annotated objects, and 141564 expressions. Different from RefCOCO, words describing absolute spatial locations (e.g., left, front) are not allowed to be used in the expressions in RefCOCO+. In RefCOCO and RefCOCO+ datasets, following the original split in [41], the visual entities to be segmented in testA subset are people, while the ones in testB subset are objects (i.e., not people). RefCOCOg [26, 28] includes 26711 images, 54822 annotated objects and 104560 expressions. Compared with RefCOCO and RefCOCO+, the expressions in RefCOCOg are more complex, which have an average length of 8.4 words. There exist two different partitions for RefCOCOg dataset: UMD split [28] and Google split [26].", + "bbox": [ + 496, + 287, + 890, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Following [3,37,38], we report our results using two kinds of metrics: overall IoU (oIoU) and Precision@X (P@X). The overall IoU measures the ratio of total intersection regions over total union regions of predicted masks and ground truths of all testing samples. Precision@X calculates the percentage of testing samples, of which the model prediction has an IoU score higher than the threshold value X, and $\\mathrm{X} \\in \\{0.5, 0.6, 0.7, 0.8, 0.9\\}$ .", + "bbox": [ + 496, + 575, + 890, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We conduct our experiments on 8 Tesla V100 GPUs. We applied our framework on various RES models [25,37,42]. On each dataset, we randomly sample $60\\%$ of the training data as the virtual training set and use the remaining training data to construct virtual testing sets at each training epoch. The learning rate $(\\alpha)$ for virtual training is 5e-5, and the learning rate $(\\beta)$ for meta update is 2e-5.", + "bbox": [ + 496, + 696, + 890, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 830, + 702, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, our framework achieves state-of-the-art performance across all three datasets, demonstrating the superiority of our framework. Moreover, we applied our", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19483", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ee113a85514dfbeddeed6f1aa28d2a7479372ad5ed839ff529b843c72e158f2e.jpg", + "table_caption": [ + "Table 3. We test several variants to investigate the impact of each level of novel compositions on model performance." + ], + "table_footnote": [], + "table_body": "
MethodIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline (LAVT)72.7384.4681.2475.2864.7134.30
w/o word-word novel compositions73.4384.7181.6976.1265.2334.47
w/o word-phrase novel compositions73.6884.8281.7676.2865.3734.56
w/o phrase-phrase novel compositions73.5984.7381.7376.1665.3534.45
Ours74.9286.2383.4577.2566.5635.61
", + "bbox": [ + 80, + 116, + 467, + 183 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/92f8d15d714845d4792841d8dbdf81136cfcce7bbf15f23080508ba1aba4cdcf.jpg", + "table_caption": [ + "Table 4. We evaluate a variant to test the effectiveness of meta optimization scheme in our framework. For this variant, its optimization objective is to minimize $\\mathcal{L}_{v\\_ tr}(\\theta) + \\sum_{k=1}^{K} \\mathcal{L}_{v\\_ te}^{k}(\\theta)$ (i.e., replacing $\\theta'$ with $\\theta$ in Eqn. 4)." + ], + "table_footnote": [], + "table_body": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Training w/o meta72.7684.5081.2775.3364.7234.31
Ours74.9286.2383.4577.2566.5635.61
", + "bbox": [ + 80, + 241, + 467, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "framework on two transformer-based SOTA models [37,42] and a CNN-based model [25]. Our framework brings consistent performance improvement on all these models and datasets. This shows that our framework can serve as a general approach to enhance model performance", + "bbox": [ + 76, + 287, + 467, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further validate the effectiveness of our framework for handling novel compositions, we perform the following analysis. Specifically, we split each testing set of RefCOCO dataset to construct two subsets. One subset (Non-novel) includes the data samples, in which all the contained compositions are seen in the RefCOCO training set. While another subset (Novel) includes the data samples that contain any level of novel compositions w.r.t. the training set. As shown in Table 2, on each testing set, we can see a clear performance gap between the two subsets for each of the baseline models [25, 37, 42], showing that existing models struggle with handling novel compositions. Then by applying our framework on each baseline model, we obtain significant performance improvement on the subset of Novel. This validates the general effectiveness of our framework to optimize the model to well generalize to novel compositions. Moreover, our framework also slightly improves the model performance on the subset of Non-novel. This can be attributed to that by training the model to better capture semantics and visual representations of individual concepts, our framework can help the model to better understand the given expression and find its visual correspondence, which thus can generally improve model performance. We also show some qualitative results in Fig. 3. As shown, when handling the testing samples containing novel compositions, our framework achieves better performance than the baseline model [37].", + "bbox": [ + 76, + 364, + 467, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 784, + 238, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [3, 37, 38], we conduct ablation experiments to evaluate our framework on RefCOCO validation set.", + "bbox": [ + 76, + 809, + 467, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of various levels of novel compositions. To investigate the impact of each level of novel compositions on model performance, we test multiple variants. One variant (w/o word-word novel compositions) ignores the word-", + "bbox": [ + 76, + 840, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1558cade3b9dbb7c371f8acd70512fe2ac0e69155f1bf1e60dbf25e697642e99.jpg", + "image_caption": [ + "Figure 3. Qualitative results of our method and the baseline model [37]. The above testing examples contain a word-word novel composition, a word-phrase novel composition, and a phrase-phrase novel composition respectively. As shown, by applying our framework on the baseline model, our method performs better when handling novel compositions of the learned concepts." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 879, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "word level novel compositions. This means that in this variant, we omit the virtual testing set for handling word-word level novel compositions during meta optimization. Similarly, the other variants (w/o word-phrase novel compositions and w/o phrase-phrase novel compositions) ignore the corresponding level of novel compositions respectively. As shown in Table 3, ignoring any level of novel compositions leads to performance drop compared to our framework, demonstrating that each level of novel compositions can affect model generalization performance.", + "bbox": [ + 496, + 318, + 890, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of meta optimization. To investigate the effectiveness of the meta optimization scheme in our framework, we evaluate a variant (training w/o meta). In this variant, we train the model in the conventional manner on the constructed virtual training set and virtual testing sets, i.e., without meta optimization. Note that for this variant, we still construct the virtual training set and virtual testing sets in the same way as in our framework. As shown in Table 4, our framework outperforms this variant obviously, demonstrating the effectiveness of our meta optimization scheme.", + "bbox": [ + 496, + 472, + 890, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of multiple virtual testing sets and curriculum learning. In our framework, we construct multiple virtual testing sets to handle various levels of novel compositions. Moreover, since the various levels of novel compositions correspond to different levels of comprehension complexity, we adopt a curriculum learning strategy to facilitate model training. To investigate the effectiveness of such design, we evaluate two variants. One variant (one virtual testing set) constructs only one virtual testing set to handle all levels of novel compositions. Such a virtual testing set includes all the data samples containing any level of novel compositions w.r.t. the virtual training set. Another variant (multiple virtual testing sets w/o curriculum learning) constructs multiple virtual testing sets as in our framework, but does not adopt the curriculum learning strategy. We compare these two variants to our original framework (multiple virtual testing sets w/ curriculum learning).", + "bbox": [ + 496, + 626, + 890, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 5, compared to the variant construct", + "bbox": [ + 519, + 885, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19484", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing one virtual testing set, the other variant obtains better performance, showing the superiority of using multiple virtual testing sets to handle various levels of novel compositions. Furthermore, by employing the curriculum learning strategy, our framework performs better than the variant constructing multiple virtual testing sets without curriculum learning, demonstrating the effectiveness of our curriculum learning strategy.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the size of virtual training set. In our framework, we randomly sample $60\\%$ of the training data as the virtual training set, and use the remaining $40\\%$ of the training data to construct virtual testing sets $(60\\%:40\\%)$ . Here we test two variants. One variant $(50\\%:50\\%)$ uses $50\\%$ of the training data to construct the virtual training set, and the remaining $50\\%$ to construct virtual testing sets. While another variant $(70\\%:30\\%)$ uses $70\\%$ of the training data for the virtual training set, and the remaining $30\\%$ part for virtual testing sets. As shown in Table 6, our method and these two variants all perform better than the baseline model (i.e., LAVT [37]), showing the robustness of our framework in terms of the varying size of virtual training set.", + "bbox": [ + 75, + 214, + 472, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of virtual testing sets construction strategy. In our framework, after sampling a subset of the training set as the virtual training set, we use the remaining training data to construct multiple virtual testing sets. Each virtual testing set consists of the data samples containing one level of novel compositions w.r.t. the virtual training set. To explore the efficacy of such a strategy, we evaluate a variant (random virtual testing sets), in which we totally randomly select training data to construct each virtual testing set. As shown in Table 7, our framework obviously outperforms this variant. This shows that our virtual testing sets construct strategy can effectively help our framework to improve model generalization performance.", + "bbox": [ + 75, + 412, + 472, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training time. As shown in Table 8, we test the training time of our framework that trains the baseline network [37] with meta optimization, and compare it to the training time of the baseline that trains the same network in the conventional manner without meta optimization, on RefCOCO dataset. Though our framework performs better, it brings only relatively little increase (18.18%) in training time.", + "bbox": [ + 75, + 611, + 472, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of additional gradient updates. As discussed above, compared to the baseline model, our framework trains the model for longer time and involves more gradient updates. To explore whether the performance improvement of our framework comes from the additional gradient updates, we test a variant (baseline w/ additional gradient updates) in which we train the baseline model (following the original training strategy) for as many iterations as in our framework. As shown in Table 9, the performance of this variant is very close to the baseline model [37], and is obviously worse than our framework. This might be because that the originally trained baseline models have al", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dfe5b51fa0acfa0ad4451844afaf75d27cf1ffe83c9273b00de71669e55fa044.jpg", + "table_caption": [ + "Table 5. We evaluate two variants to test the impact of using multiple virtual testing sets and the curriculum learning strategy." + ], + "table_footnote": [], + "table_body": "
MethodoloUP@0.5P@0.6P@0.7P@0.8P@0.9
Ours (one virtual testing set)73.6484.8282.0876.2465.4734.64
Ours (multiple virtual testing sets w/o curriculum learning)74.0285.3482.6376.8766.0934.95
Ours (multiple virtual testing sets w/ curriculum learning)74.9286.2383.4577.2566.5635.61
", + "bbox": [ + 501, + 117, + 890, + 157 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d2d0dff27a28ca93720674d32a282f066866b6a2f7c0eaf8f93c215f94a82e85.jpg", + "table_caption": [ + "Table 6. We test different variants that utilize different proportions of the training data to construct the virtual training set and virtual testing sets." + ], + "table_footnote": [], + "table_body": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline (LAVT)72.7384.4681.2475.2864.7134.30
Ours (50%:50%)74.7886.1183.3277.0866.3835.54
Ours (60%:40%)74.9286.2383.4577.2566.5635.61
Ours (70%:30%)74.7486.0483.2677.0166.2935.52
", + "bbox": [ + 503, + 207, + 890, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/709c18e54624b6f344074a76e3e04e44185d3a18abcb992f168aa30deac65cc1.jpg", + "table_caption": [ + "Table 7. We evaluate a variant to investigate the efficacy of virtual testing sets construction strategy in our framework." + ], + "table_footnote": [], + "table_body": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Random virtual testing sets72.8184.5681.2875.3464.7434.32
Ours74.9286.2383.4577.2566.5635.61
", + "bbox": [ + 503, + 311, + 890, + 349 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6b1ff69a316a900d7420d8bf8857a781ec0b0f4759daeee3999bfa4ea4383872.jpg", + "table_caption": [ + "Table 8. Comparison of the training time. Note that our method achieves much better performance than the baseline." + ], + "table_footnote": [], + "table_body": "
MethodTraining timeoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline33 hours72.7384.4681.2475.2864.7134.30
Ours39 hours74.9286.2383.4577.2566.5635.61
", + "bbox": [ + 503, + 387, + 890, + 421 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0540f0ca43e29bb9812010816ebe714536f83578059c4ac55791f3e3d13c1f9f.jpg", + "table_caption": [ + "Table 9. We test a variant to investigate the impact of additional gradient updates on model performance." + ], + "table_footnote": [], + "table_body": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline72.7384.4681.2475.2864.7134.30
Baseline w/ additional gradient updates72.7484.4381.2575.2564.7334.28
Baseline w/ ours74.9286.2383.4577.2566.5635.61
", + "bbox": [ + 503, + 458, + 890, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ready reached convergence under the original training strategy, and thus additional gradient updates would not bring obvious benefits. Such results further validate the effectiveness of our framework.", + "bbox": [ + 498, + 525, + 890, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 611, + 619, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed a meta learning-based framework (MCRES) to improve the generalization performance of RES models, especially when handling novel compositions of learned concepts. By constructing a virtual training set and multiple virtual testing sets w.r.t. various levels of novel compositions and then optimizing the model via meta optimization, our framework can effectively improve model generalization performance. Extensive experiments show that our framework achieves superior performance on widely used benchmarks. Moreover, our framework is flexible, and can be seamlessly applied on various models with different architectures to enhance their performance.", + "bbox": [ + 496, + 638, + 890, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by MOE AcRF Tier 2 (Proposal ID: T2EP20222-0035), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2020-065), and SUTD SKI Project (SKI 2021_02_06).", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19485", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2", + "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 6", + "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 6, 7", + "[4] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 6", + "[5] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 1, 3, 6", + "[6] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 6", + "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 3", + "[8] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Era: Expert retrieval and assembly for early action prediction. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 670-688. Springer, 2022. 3", + "[9] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. Allennlp: A deep semantic natural language processing platform. arXiv preprint arXiv:1803.07640, 2018. 1, 2, 4, 5", + "[10] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2", + "[11] Jianzhu Guo, Xiangyu Zhu, Chenxu Zhao, Dong Cao, Zhen Lei, and Stan Z Li. Learning meta face recognition in unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163-6172, 2020. 3", + "[12] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 1, 3", + "[13] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4424-4433, 2020. 6", + "[14] Chao Huang, Zhangjie Cao, Yunbo Wang, Jianmin Wang, and Mingsheng Long. Metasets: Meta-learning on point sets for generalizable representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8863-8872, 2021. 3", + "[15] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10488-10497, 2020. 6", + "[16] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In European Conference on Computer Vision, pages 59-75. Springer, 2020. 6", + "[17] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 6", + "[18] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 1, 3, 6", + "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 3", + "[20] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2018. 3, 6", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6", + "[22] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1271-1280, 2017. 3", + "[23] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6", + "[24] Gen Luo, Yiyi Zhou, Rongrong Ji, Xiaoshuai Sun, Jinsong Su, Chia-Wen Lin, and Qi Tian. Cascade grouped attention network for referring expression segmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1274-1282, 2020. 6", + "[25] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collabora" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19486", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tive network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 1, 2, 6, 7", + "[26] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 1, 6", + "[27] Edgar Margffoy-Tuay, Juan C Pérez, Emilio Botero, and Pablo Arbeláez. Dynamic multimodal instance segmentation guided by natural language queries. In Proceedings of the European Conference on Computer Vision (ECCV), pages 630-645, 2018. 6", + "[28] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In European Conference on Computer Vision, pages 792-807. Springer, 2016. 1, 3, 6", + "[29] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2, 3", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3", + "[31] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 3", + "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3", + "[33] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6", + "[34] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11686-11695, 2022. 1, 3, 6", + "[35] Li Xu, Haoxuan Qu, Jason Kuen, Juixiang Gu, and Jun Liu. Meta spatio-temporal debiasing for video scene graph generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 374-390. Springer, 2022. 3", + "[36] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 3, 6", + "[37] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 1, 2, 3, 6, 7, 8" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1, 6, 7", + "[39] Nanyang Ye, Jingxuan Tang, Huayu Deng, Xiao-Yun Zhou, Qianxiao Li, Zhenguo Li, Guang-Zhong Yang, and Zhanxing Zhu. Adversarial invariant learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12441-12449. IEEE, 2021. 2", + "[40] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 1, 3, 6", + "[41] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 1, 3, 6", + "[42] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision. Springer, 2022. 1, 2, 6, 7" + ], + "bbox": [ + 501, + 92, + 892, + 429 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19487", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_model.json b/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9721c0ef6a0b78ff026a81e3cc9b88d227e784bd --- /dev/null +++ b/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.131, + 0.773, + 0.154 + ], + "angle": 0, + "content": "Meta Compositional Referring Expression Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.18, + 0.818, + 0.233 + ], + "angle": 0, + "content": "Li Xu\\(^{1}\\), Mark He Huang\\(^{1,3}\\), Xindi Shang\\(^{2}\\), Zehuan Yuan\\(^{2}\\), Ying Sun\\(^{3}\\), Jun Liu\\(^{1*}\\) \n\\(^{1}\\)Singapore University of Technology and Design, Singapore \n\\(^{2}\\)ByteDance" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.234, + 0.913, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)Institute for Infocomm Research (I\\(^2\\)R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.255, + 0.686, + 0.305 + ], + "angle": 0, + "content": "{li_xu, he_huang}@mymail.sutd.edu.sg shangxindi, yuanzehuan}@bytedance.com @i2r.a-star.edu.sg, jun_liu@sutd.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.354 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.474, + 0.735 + ], + "angle": 0, + "content": "Referring expression segmentation aims to segment an object described by a language expression from an image. Despite the recent progress on this task, existing models tackling this task may not be able to fully capture semantics and visual representations of individual concepts, which limits their generalization capability, especially when handling novel compositions of learned concepts. In this work, through the lens of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework to enhance model compositional generalization performance. Specifically, to handle various levels of novel compositions, our framework first uses training data to construct a virtual training set and multiple virtual testing sets, where data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set. Then, following a novel meta optimization scheme to optimize the model to obtain good testing performance on the virtual testing sets after training on the virtual training set, our framework can effectively drive the model to better capture semantics and visual representations of individual concepts, and thus obtain robust generalization performance even when handling novel compositions. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our framework." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.208, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Referring expression segmentation (RES) [12, 38, 40] aims to segment a visual entity in an image given a linguistic expression. This task has been receiving increasing attention in recent years [5, 18, 34, 37], as it can play an important role in various applications, such as language-based human-robot interaction and interactive image edit" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.34, + 0.86, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.45, + 0.706, + 0.461 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.47, + 0.858, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.545, + 0.706, + 0.556 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Figure 1. Illustration of novel compositions and various levels of compositions. (a) An example of the testing sample containing the novel composition of \"dark coffee\" in RefCOCO dataset [41]. Such a novel composition itself does not exist in the training data, but its individual components (i.e., \"dark\" and \"coffee\") exist in the training data. (b) An example of various levels of compositions in an expression. We perform constituency parsing of the expression using AllenNLP [9]. Based on the obtained parsing tree as shown above, we can then obtain various levels of compositions (e.g., word-word level, word-phrase level) in this expression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.809 + ], + "angle": 0, + "content": "ing. However, despite the recent progress on tackling this task [5, 34, 42], existing methods may struggle with handling the testing samples, of which the expressions contain novel compositions of learned concepts. Here a novel composition means that the composition itself does not exist in the training data, but its individual components (e.g., words, phrases) exist in the training data, as shown in Fig. 1a." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We observe that testing samples containing such novel compositions of learned concepts widely exist in RES datasets [26, 28, 41]. However, existing RES models may not be able to well handle novel compositions during testing. Here we test the generalization capability of multiple state-of-the-art models [25, 37, 42] in terms of handling" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19478" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "novel compositions, as shown in Table 2. Specifically, we first split each testing set of the RefCOCO dataset. In each testing set, one split subset includes the data samples, in which all the contained compositions are seen in the RefCOCO training set. While another subset includes the data samples containing novel compositions, of which the individual components (e.g., words, phrases) exist in the RefCOCO training set but the composition itself is unseen in the training set, i.e., containing novel compositions of learned concepts. Then we evaluate the models [25, 37, 42] on the two subsets in each testing set, and find that for each model, its testing performance on the subset containing novel compositions drops obviously compared to the performance on the other subset. For these models, the performance gap between the two subsets can reach \\(14\\% - 17\\%\\) measured by the metric of overall IoU. Such a clear performance gap indicates that existing models struggle with generalizing to novel compositions of learned concepts. This might due to that the model does not effectively capture the semantics and visual representations of individual concepts (e.g., \"dark\", \"coffee\" in Fig. 1a) during training. Then the trained model may fail to recognize a novel composition (e.g., \"dark coffee\") at testing time, which though is composed of learned concepts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.457, + 0.473, + 0.595 + ], + "angle": 0, + "content": "Thus to handle this issue, we aim to train the model to effectively capture the semantics and visual representations of individual concepts during training. Despite the conceptual simplicity, how to guide the model's learning behavior towards this goal is a challenging problem. Here from the perspective of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework, to effectively handle such a challenging problem by only changing the model training scheme." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.596, + 0.473, + 0.809 + ], + "angle": 0, + "content": "Meta learning proposes to perform virtual testing during model training for better performance [7, 29]. Inspired by this, to improve the generalization capability of RES models, our MCRES framework incorporates a meta optimization scheme that consists of three steps: virtual training, virtual testing and meta update. Specifically, we first split the training set to construct a virtual training set for virtual training, and a virtual testing set for virtual testing. The data samples in the virtual testing set contain novel compositions w.r.t. the virtual training set. For example, if the expressions of data samples in the virtual training set contain both words \"dark\" and \"coffee\" but do not contain their composition (i.e., \"dark coffee\"), the virtual testing set can include this novel composition correspondingly." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Based on the constructed virtual training set and virtual testing set, we first train the model using the virtual training set, and then evaluate the trained model on the virtual testing set. During virtual training, the model may learn the compositions of individual concepts as a whole without truly understanding the semantics and visual representa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.35 + ], + "angle": 0, + "content": "tions of individual concepts, which though can still improve model training performance. For example, if there are many training samples containing the composition of \"yellow banana\" in the virtual training set, the model can superficially correlate \"banana\" with \"yellow\" and learn this composition as a whole, since using such spurious correlations can facilitate the model learning [1, 10, 39]. However, learning the compositions as a whole over the virtual training set may not improve model performance much on the virtual testing set in virtual testing, since the virtual testing set contains novel compositions w.r.t. the virtual training set. Thus to achieve good testing performance on such a virtual testing set, the model needs to effectively capture semantics and visual representations of individual concepts during virtual training. In this way, the model testing performance on the virtual testing set serves as a generalization feedback to the model virtual training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.895, + 0.487 + ], + "angle": 0, + "content": "Thus after the virtual training and virtual testing, we can further update the model to obtain better testing performance on the virtual testing set (i.e., meta update), so as to drive the model training on the virtual training set towards the direction of learning to capture semantics and visual representations of individual concepts, i.e., learning to learn. In this manner, our framework is able to optimize the model for robust generalization performance, even tackling the challenging testing samples with novel compositions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.896, + 0.761 + ], + "angle": 0, + "content": "Moreover, given that expressions can often be hierarchically decomposed, there can exist various levels of novel compositions. Specifically, to identify meaningful compositions in an expression, we can parse an expression into a tree structure based on the constituency parsing tool [9] as shown in Fig.1b. In such a parsing tree, under the same parent node, each pair of child nodes (e.g., \"white\" and \"bird\") are closely semantically related, and thus can form a meaningful composition. Since each child node can be a word or a phrase as in Fig. 1b, there can naturally exist the following three levels of novel compositions: word-word level (e.g., \"white\" and \"bird\"), word-phrase level (e.g., \"standing\" and \"in water\") and phrase-phrase level (e.g., \"white bird\" and \"standing in water\"), which correspond to different levels of comprehension complexity. To better handle such a range of novel compositions, we construct multiple virtual testing sets in our framework, where each virtual testing set is constructed to handle one level of novel compositions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.895, + 0.839 + ], + "angle": 0, + "content": "Our framework only changes the model training scheme without the need to change the model structure. Thus our framework is general, and can be conveniently applied on various RES models. We test our framework on multiple models, and obtain consistent performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The contributions of our work are threefold: 1) We propose a novel framework (MCRES) to effectively improve generalization performance of RES models, especially when handing novel compositions of learned con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19479" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.184 + ], + "angle": 0, + "content": "cepts. 2) Via constructing a virtual training set and multiple virtual testing sets w.r.t. various levels of novel compositions, our framework can train the model to well handle various levels of novel compositions. 3) When applied on various models on three RES benchmarks [28, 41], our framework achieves consistent performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.198, + 0.228, + 0.213 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Referring Expression Segmentation (RES). RES [12] aims to segment a target object from an image based on an expression. Early works [20, 22, 28, 41] employed convolutional and recurrent networks to extract visual and linguistic features respectively, and then fused the extracted features to predict the segmentation mask. Recently, a series of works [5, 18, 37] employed vision and language transformers to boost performance. Besides, with the development of large-scale pretrained models, Wang et al. [34] proposed to leverage CLIP [30] to improve cross-modal matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.471, + 0.557 + ], + "angle": 0, + "content": "Some methods have been explored to help model better understand learned concepts in RES. Yu et al. [40] proposed a modular network that uses different modules to process different types of information in the given expression. Yang et al. [36] designed a reasoning module to help align the language concepts with visual regions. Different from all the above-mentioned works, we propose an MCRES framework to drive RES models to better capture semantics and visual representations of individual concepts. Such a framework only changes the model training scheme, and thus can be flexibly applied on various models to improve their generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.471, + 0.739 + ], + "angle": 0, + "content": "Meta Learning. Meta learning, i.e., the paradigm of learning to learn, has emerged to mainly tackle the few-shot learning problem [7, 29, 31, 32]. MAML [7] and its following works [29, 31] aim to learn a good initialization of network parameters, to achieve fast test-time update to adapt to new few-shot learning tasks. More recently, meta learning has also been explored in other areas [8, 11, 14, 19, 35] to enhance model generalization performance without the need of test-time update. Inspired by these works, we leverage a meta learning-based framework to improve generalization performance of RES models especially when handling novel compositions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.754, + 0.169, + 0.769 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Existing RES models may fail to generalize to data samples containing novel compositions of learned concepts. To handle this issue, we aim to encourage the model to better capture the semantics of individual concepts as well as recognize their visual correspondences during training, which however is a non-trivial problem. In this paper, from the perspective of meta learning, we introduce MCRES framework optimizing RES models via a meta optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.85, + 0.107 + ], + "angle": 0, + "content": "scheme to improve their generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 2, in our framework, we first split the original training set \\((\\mathcal{D}_{train})\\) to build a virtual training set \\((\\mathcal{D}_{v\\_ tr})\\) for virtual training, and a group of \\(K\\) virtual testing sets \\(\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)\\) for virtual testing. Each virtual testing set consists of data samples containing one level of novel compositions w.r.t. the virtual training set. We first use the virtual training set to train the model (virtual training), and then perform model testing on the virtual testing sets (virtual testing). Since the virtual testing sets contain novel compositions of individual concepts from the virtual training set, if the model can still achieve robust testing performance on the virtual testing sets after training on the virtual training set, we posit that the trained model has learned to capture more semantics and visual representations of individual concepts during the virtual training. Guided by this principle, we can optimize the model virtual testing performance, to guide the model virtual training process towards learning more semantics and representations of individual concepts. Below, we first introduce the pipeline of our framework, and then detail the virtual training set and virtual testing sets construction process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.628, + 0.453 + ], + "angle": 0, + "content": "3.1. Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.553 + ], + "angle": 0, + "content": "We first train the RES model using the virtual training set, i.e., virtual training. Specifically, we denote the model parameters as \\(\\theta\\), the loss function (e.g., cross-entropy loss) for training the RES model as \\(\\mathcal{L}\\). Thus we can calculate the model virtual training loss \\((\\mathcal{L}_{v\\_ tr})\\) over the virtual training set \\((\\mathcal{D}_{v\\_ tr})\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.56, + 0.891, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v \\_ t r} (\\theta) = \\mathcal {L} (\\theta ; \\mathcal {D} _ {v \\_ t r}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.583, + 0.891, + 0.612 + ], + "angle": 0, + "content": "Based on this loss, we then update the model parameters \\((\\theta)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.616, + 0.891, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\theta^ {\\prime} = \\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v _ {-} t r} (\\theta) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.891, + 0.703 + ], + "angle": 0, + "content": "where \\(\\alpha\\) denotes the learning rate. It is worth noting that the model update at this step is virtual, and the updated parameters \\(\\theta'\\) only serve as intermediate parameters for the model evaluation at the following virtual testing step." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.766 + ], + "angle": 0, + "content": "After the virtual training, we evaluate the generalization performance of the trained model on the virtual testing sets \\(\\left(\\{\\mathcal{D}_{v\\_te}^k\\}_{k = 1}^K\\right)\\), i.e., virtual testing. Specifically, on each virtual testing set \\(\\mathcal{D}_{v\\_te}^k\\), we compute the model loss \\(\\mathcal{L}_{v\\_te}^k\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.77, + 0.891, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v. t e} ^ {k} \\left(\\theta^ {\\prime}\\right) = \\mathcal {L} \\left(\\theta^ {\\prime}; \\mathcal {D} _ {v. t e} ^ {k}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.891, + 0.825 + ], + "angle": 0, + "content": "Such a loss measures how well the model generalizes to the virtual testing set after training on the virtual training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Then we perform the meta update step. At this step, we wish to actually update the model parameters \\((\\theta)\\), so that after the model training on the virtual training set, the trained model can generalize well to the virtual testing sets, i.e., generalizing well to novel compositions. To this end, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19480" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.468, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.332, + 0.294, + 0.343 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.093, + 0.857, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.332, + 0.7, + 0.343 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Framework overview. Fig. (a) illustrates our framework pipeline. First, we train the model using the virtual training set \\((\\mathcal{D}_{v\\_ tr})\\) and then obtain the updated model. We then test the model with updated parameters \\((\\theta^{\\prime})\\) on multiple virtual testing sets \\(\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)\\). According to the virtual testing losses, we perform meta update to optimize the model for better generalization capability. Fig. (b) shows that we construct a virtual training set and multiple virtual testing sets to handle various levels of novel compositions. The expressions of data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.428, + 0.343, + 0.442 + ], + "angle": 0, + "content": "formulate the optimization objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.448, + 0.469, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} _ {v. t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v. t e} ^ {k} \\left(\\theta^ {\\prime}\\right) \\tag {4} \\\\ = \\min _ {\\theta} \\mathcal {L} _ {v - t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v - t e} ^ {k} \\left(\\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v - t r} (\\theta)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.616 + ], + "angle": 0, + "content": "where the first term indicates the model training performance on the virtual training set, and the second term represents the model testing performance on the virtual testing sets after training on the virtual training set. Based on this objective, we then update the model parameters \\((\\theta)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.622, + 0.469, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\left(\\mathcal {L} _ {v _ {-} t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v _ {-} t e} ^ {k} \\left(\\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v _ {-} t r} (\\theta)\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.676, + 0.47, + 0.737 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the learning rate for meta update. Through the above optimization process, the model is pushed to learn more semantics and visual representations of individual concepts during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "During the above process, the model is first trained (updated) on the virtual training set. At this step, the model may learn the compositions of individual concepts as a whole without effectively capturing the semantics and visual representations of individual concepts, which though can still improve its training performance on the virtual training set. However, to achieve good testing performance on the virtual testing data which contains novel compositions of individual concepts from the virtual training data, the model is expected to avoid learning compositions as a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.517 + ], + "angle": 0, + "content": "whole and instead capture more semantics and representations of individual concepts. In this way, the second term of Eqn. 5, which includes the second-order gradients of \\(\\theta\\): \\(\\nabla_{\\theta}\\mathcal{L}_{v\\_te}^{k}\\big(\\theta -\\alpha \\nabla_{\\theta}\\mathcal{L}_{v\\_tr}(\\theta)\\big)\\), can be regarded as a generalization feedback that can guide the model to capture more semantics and representations of individual concepts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.564 + ], + "angle": 0, + "content": "In our framework, the above three steps (i.e., virtual training, virtual testing and meta update) are performed iteratively until the model training converges." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.673, + 0.589 + ], + "angle": 0, + "content": "3.2. Sets Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.704 + ], + "angle": 0, + "content": "In our framework, to handle various levels of novel compositions, we split the original training set to construct a virtual training set and various virtual testing sets. Each virtual testing set is expected to contain one level of novel compositions w.r.t. the virtual training set. Below we first introduce the details of these sets, and then discuss the strategy for constructing the virtual testing sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.885 + ], + "angle": 0, + "content": "We randomly sample a subset of the training set \\((\\mathcal{D}_{train})\\) as the virtual training set \\((\\mathcal{D}_{v\\_ tr})\\), and the remaining training data will be used as the candidate samples to construct a group of \\((K)\\) virtual testing sets \\(\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)\\). Each virtual testing set is constructed to handle one level of novel compositions. Here considering the hierarchical semantic structure of each expression, we target the following levels of novel compositions: word-word level, word-phrase level and phrase-phrase level. A phrase means a group of words that serve as a grammatical unit in the expression (e.g., \"white bird\" in Fig. 1b), which can be conveniently identified using off-the-shelf tools (e.g., AllenNLP [9])." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Specifically, to identify various levels of novel compo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19481" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "sitions in an expression, we first use the constituency parsing tool [9] to parse an expression into a tree structure as shown in Fig. 1b. In this tree, under the same parent node, each pair of child nodes are closely related, and thus can form a meaningful composition. If a pair of nodes having the same parent node are both words, they form a word-word level composition (e.g., \"white\" and \"bird\"). Similarly, if the nodes with the same parent node are a word and a phrase, they form a word-phrase level composition (e.g., \"standing\" and \"in water\"). In a similar way, we can obtain phrase-phrase level compositions (e.g., \"white bird\" and \"standing in water\"). Thus there can exist three levels of novel compositions, which correspond to different levels of comprehension complexity. Corresponding to these three levels of novel compositions, we will construct a total of three virtual testing sets \\((K = 3)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.473, + 0.549 + ], + "angle": 0, + "content": "Virtual testing sets construction. To construct each virtual testing set, we need to select data samples containing the corresponding level of novel compositions w.r.t. the virtual training set, from the candidate samples. A data sample means an expression paired with the corresponding image. Here we design an efficient strategy, which can be leveraged to construct each virtual testing set. Below, we take the process of constructing the virtual testing set for handling word-word level novel compositions as an example, to introduce such a strategy. Specifically, this virtual testing set should include all of those candidate samples that contain word-word level novel compositions w.r.t. the virtual training set. To construct such a virtual testing set, our strategy proceeds as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.903 + ], + "angle": 0, + "content": "(i) To find the data samples, of which the expressions contain word-word level novel compositions w.r.t. the virtual training set, from the candidate samples, we need to first obtain all the word-word level compositions in the virtual training set and in the candidate samples respectively. Thus by parsing each expression into a parsing tree as shown in Fig. 1b, we can obtain all word-word level compositions (e.g., \"white\" and \"bird\", \"in\" and \"water\" in Fig. 1b) in each expression. (ii) To identify word-word level novel compositions, we first select the word-word level compositions that exist in the candidate samples but are unseen in the virtual training set. Then for each selected composition, we further check if its individual words exist in the virtual training set. If so, such a composition will be identified as a word-word level novel composition w.r.t. the virtual training set. For example, if the composition \"white bird\" is unseen in the virtual training set, while its individual words \"white\" and \"bird\" both exist in the virtual training set, this composition is a word-word level novel composition w.r.t. the virtual training set. (iii) Finally, we select the candidate samples, of which the expressions contain the identified word-word level novel compositions w.r.t. the virtual training set, to construct the virtual testing set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.439 + ], + "angle": 0, + "content": "To efficiently perform the above steps, we can employ parallel matrix operations. Specifically, to record the word-word level compositions in the virtual training set, we build a matrix \\(\\mathcal{M}_{v,tr}\\) with the shape of \\(|\\mathcal{V}| \\times |\\mathcal{V}|\\), where \\(|\\mathcal{V}|\\) is the size of the vocabulary set \\(\\mathcal{V}\\) of the original training set. In this matrix, each element at the \\(i-th\\) row and the \\(j-th\\) column \\((i,j \\in \\{1,\\dots,|\\mathcal{V}|\\})\\) is a binary value, and the value 1 indicates that the \\(i-th\\) word and \\(j-th\\) word in the vocabulary set \\(\\mathcal{V}\\) form a word-word level composition in the virtual training set, while 0 means such a word-word level composition does not exist in the virtual training set. Similarly, we can also build a matrix \\(\\mathcal{M}_{candi}\\) to record the word-word level compositions in the candidate samples. Then to identify word-word level novel compositions w.r.t. the virtual training set \\((\\mathcal{M}_{v,tr})\\) from the candidate samples \\((\\mathcal{M}_{candi})\\), we can efficiently compute a difference matrix: \\(\\mathcal{M}_{diff} = \\mathcal{M}_{candi} - \\mathcal{M}_{v,tr}\\). In \\(\\mathcal{M}_{diff}\\), any element with the value 1 indicates that the corresponding composition exists in the candidate samples, but is unseen in the virtual training set. Then by further checking whether the individual words in such a composition exist in the virtual training set, we can determine if it is a word-word level novel composition w.r.t. the virtual training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.895, + 0.547 + ], + "angle": 0, + "content": "Similarly, we can efficiently construct the other two virtual testing sets. Note that to help the model to learn to handle a wide range of possible novel compositions, at the beginning of each training epoch, we randomly sample a subset of the training data to re-construct the virtual training set, and leverage the remaining data to re-construct the virtual testing sets using the above strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.897, + 0.744 + ], + "angle": 0, + "content": "As discussed above, to handle various levels of novel compositions, we construct multiple virtual testing sets. A simpler alternative is to construct only one virtual testing set, which includes all the data samples containing any level of novel compositions w.r.t. the virtual training set. However, compared to this alternative, by constructing multiple virtual testing sets, each level of novel compositions in the corresponding virtual testing set can offer an explicit generalization feedback during model training. Thus the model is explicitly encouraged to well handle each level of novel compositions. Moreover, constructing multiple virtual testing sets can facilitate the use of curriculum learning strategy for model training as discussed in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.698, + 0.772 + ], + "angle": 0, + "content": "3.3. Training and Testing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To help models generalize to novel compositions in RES, our framework only changes the model training scheme. Thus our framework is general, and can be flexibly applied to train various RES models. During training, at each epoch, we first split the training set to construct a virtual training set and multiple virtual testing sets. Then we iteratively optimize the model over the virtual training set and virtual testing sets via meta optimization. Specifically, for each meta" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19482" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.173 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-arts on three benchmark datasets using the metric of overall IoU. Moreover, we apply our framework on various models [25, 37, 42], and obtain consistent performance improvement. “-” indicates that the corresponding result is not provided in the original paper. U means the UMD split of RefCOCOg dataset, and G means the Google split." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.173, + 0.47, + 0.464 + ], + "angle": 0, + "content": "
MethodRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBval(U)test(U)val(G)
DMN [27]49.7854.8345.1338.8844.2232.29--36.76
RRN [20]55.3357.2653.9339.7542.1536.11--36.45
MAAttNet [40]56.5162.3751.7046.6752.3940.0847.6448.61-
CAC [4]58.9061.7753.81---46.3746.9544.32
CMSA [38]58.3260.6155.0943.7647.6037.89--39.98
STEP [3]60.0463.4657.9748.1952.33404.1--46.40
BRINet [13]60.9862.9959.2148.1752.3242.11--48.04
CMPC [15]61.3664.5359.6449.5653.4443.23--49.05
LSCM [16]61.4764.9959.5549.3453.1243.50--48.05
CMPC+ [23]62.4765.0860.8250.2554.0443.47--49.89
EFN [6]62.7665.6959.6751.5055.2443.01--51.93
BUSNet [36]63.2766.4161.3951.7656.8744.13--50.56
CGAN [24]64.8668.0462.0751.0355.5144.0651.0151.6946.54
LTS [17]65.4367.7663.0854.2158.3248.0254.4054.25-
VLT [5]65.6568.2962.7355.5059.2049.3652.9956.6549.76
ReSTR [18]67.2269.3064.4555.7860.4448.27--54.48
CRIS [34]70.4773.1866.1062.2768.0853.6859.8760.36-
MCN [25]62.4464.2059.7150.6254.9944.6949.2249.40-
MCN + Ours64.5166.4862.8452.2956.7146.7251.4951.63-
SeqTR [42]71.7073.3169.8263.0466.7358.9764.6965.74-
SeqTR + Ours73.2375.0171.9564.7167.8560.8566.7767.48-
LAVT [37]72.7375.8268.7962.1468.3855.1061.2462.0960.50
LAVT + Ours74.9276.9870.8464.3269.6856.6463.5164.9061.63
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.688 + ], + "angle": 0, + "content": "optimization iteration, we use two batches of data samples: one batch of data for virtual training and the other batch of data for virtual testing. Thus we need to ensure that the batch of data for virtual testing consists of the data samples containing novel compositions w.r.t. the batch of data for virtual training. To this end, for each iteration, we randomly sample some data from each virtual testing set, to form the batch of data for virtual testing. Then we select the virtual training set samples that contain all the individual components of the novel compositions in the batch of data for virtual testing, to form the batch of data for virtual training. Note that the above data preparation procedures can be done before the model training. For model testing, we test the trained model in the conventional way." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Besides, given that the various levels of novel compositions correspond to different levels of comprehension complexity, we adopt a curriculum learning strategy [2, 33] for model training, so that the model can progressively learn to handle various levels of novel compositions, i.e., from lower level (word-word) to middle level (word-phrase) to higher level (phrase-phrase), and thus can learn all these levels of novel compositions better. Specifically, in the first \\(1/3\\) of training epochs, we only use the virtual testing set for handling word-word level novel compositions for meta optimization. Then in the middle \\(1/3 - 2/3\\) of training epochs, we add the virtual testing set for handling word-phrase level novel compositions. Finally, in the remaining training epochs, we use all the three virtual testing sets." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.171 + ], + "angle": 0, + "content": "Table 2. We evaluate our framework on different testing subsets w.r.t. novel compositions to validate its effectiveness on optimizing the model to better generalize to novel compositions. We use the metric of overall IoU here. The performance gain of our framework compared to the corresponding baseline model is shown in parentheses." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.171, + 0.892, + 0.242 + ], + "angle": 0, + "content": "
MethodRefCOCO-valRefCOCO-testARefCOCO-testB
NovelNon-novelNovelNon-novelNovelNon-novel
MCN [23]53.1767.4155.4370.0850.6466.47
MCN + Ours57.38 (↑4.21)68.09 (↑0.68)59.66 (↑4.23)70.67 (↑0.59)55.87 (↑5.23)67.04 (↑0.57)
SeqTR [40]64.2478.5965.7179.4660.1477.21
SeqTR + Ours67.31 (↑3.07)79.06 (↑0.47)69.48 (↑3.77)79.84 (↑0.38)64.89 (↑4.75)77.68 (↑0.47)
LAVT [35]63.5278.5567.1780.7158.4976.23
LAVT + Ours67.42 (↑3.90)79.29 (↑0.74)70.05 (↑2.88)81.37 (↑0.66)62.63 (↑4.14)76.92 (↑0.69)
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.262, + 0.634, + 0.279 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.575 + ], + "angle": 0, + "content": "We evaluate our method on three commonly used RES benchmarks: RefCOCO [41], RefCOCO+ [41] and RefCOCOg [26, 28]. Images in these three benchmarks are all from the COCO dataset [21]. RefCOCO [41] contains 19994 images, 50000 annotated objects with 142209 referring expressions. RefCOCO+ [41] consists of 19992 images with 49856 annotated objects, and 141564 expressions. Different from RefCOCO, words describing absolute spatial locations (e.g., left, front) are not allowed to be used in the expressions in RefCOCO+. In RefCOCO and RefCOCO+ datasets, following the original split in [41], the visual entities to be segmented in testA subset are people, while the ones in testB subset are objects (i.e., not people). RefCOCOg [26, 28] includes 26711 images, 54822 annotated objects and 104560 expressions. Compared with RefCOCO and RefCOCO+, the expressions in RefCOCOg are more complex, which have an average length of 8.4 words. There exist two different partitions for RefCOCOg dataset: UMD split [28] and Google split [26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Evaluation metrics. Following [3,37,38], we report our results using two kinds of metrics: overall IoU (oIoU) and Precision@X (P@X). The overall IoU measures the ratio of total intersection regions over total union regions of predicted masks and ground truths of all testing samples. Precision@X calculates the percentage of testing samples, of which the model prediction has an IoU score higher than the threshold value X, and \\( \\mathrm{X} \\in \\{0.5, 0.6, 0.7, 0.8, 0.9\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.819 + ], + "angle": 0, + "content": "Implementation details. We conduct our experiments on 8 Tesla V100 GPUs. We applied our framework on various RES models [25,37,42]. On each dataset, we randomly sample \\(60\\%\\) of the training data as the virtual training set and use the remaining training data to construct virtual testing sets at each training epoch. The learning rate \\((\\alpha)\\) for virtual training is 5e-5, and the learning rate \\((\\beta)\\) for meta update is 2e-5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.704, + 0.848 + ], + "angle": 0, + "content": "4.1. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, our framework achieves state-of-the-art performance across all three datasets, demonstrating the superiority of our framework. Moreover, we applied our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19483" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.468, + 0.117 + ], + "angle": 0, + "content": "Table 3. We test several variants to investigate the impact of each level of novel compositions on model performance." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.117, + 0.468, + 0.184 + ], + "angle": 0, + "content": "
MethodIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline (LAVT)72.7384.4681.2475.2864.7134.30
w/o word-word novel compositions73.4384.7181.6976.1265.2334.47
w/o word-phrase novel compositions73.6884.8281.7676.2865.3734.56
w/o phrase-phrase novel compositions73.5984.7381.7376.1665.3534.45
Ours74.9286.2383.4577.2566.5635.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.187, + 0.468, + 0.24 + ], + "angle": 0, + "content": "Table 4. We evaluate a variant to test the effectiveness of meta optimization scheme in our framework. For this variant, its optimization objective is to minimize \\(\\mathcal{L}_{v\\_ tr}(\\theta) + \\sum_{k=1}^{K} \\mathcal{L}_{v\\_ te}^{k}(\\theta)\\) (i.e., replacing \\(\\theta'\\) with \\(\\theta\\) in Eqn. 4)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.242, + 0.468, + 0.282 + ], + "angle": 0, + "content": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Training w/o meta72.7684.5081.2775.3364.7234.31
Ours74.9286.2383.4577.2566.5635.61
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.289, + 0.468, + 0.364 + ], + "angle": 0, + "content": "framework on two transformer-based SOTA models [37,42] and a CNN-based model [25]. Our framework brings consistent performance improvement on all these models and datasets. This shows that our framework can serve as a general approach to enhance model performance" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.468, + 0.773 + ], + "angle": 0, + "content": "To further validate the effectiveness of our framework for handling novel compositions, we perform the following analysis. Specifically, we split each testing set of RefCOCO dataset to construct two subsets. One subset (Non-novel) includes the data samples, in which all the contained compositions are seen in the RefCOCO training set. While another subset (Novel) includes the data samples that contain any level of novel compositions w.r.t. the training set. As shown in Table 2, on each testing set, we can see a clear performance gap between the two subsets for each of the baseline models [25, 37, 42], showing that existing models struggle with handling novel compositions. Then by applying our framework on each baseline model, we obtain significant performance improvement on the subset of Novel. This validates the general effectiveness of our framework to optimize the model to well generalize to novel compositions. Moreover, our framework also slightly improves the model performance on the subset of Non-novel. This can be attributed to that by training the model to better capture semantics and visual representations of individual concepts, our framework can help the model to better understand the given expression and find its visual correspondence, which thus can generally improve model performance. We also show some qualitative results in Fig. 3. As shown, when handling the testing samples containing novel compositions, our framework achieves better performance than the baseline model [37]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.24, + 0.8 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Following [3, 37, 38], we conduct ablation experiments to evaluate our framework on RefCOCO validation set." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Impact of various levels of novel compositions. To investigate the impact of each level of novel compositions on model performance, we test multiple variants. One variant (w/o word-word novel compositions) ignores the word-" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.09, + 0.88, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.225, + 0.892, + 0.308 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of our method and the baseline model [37]. The above testing examples contain a word-word novel composition, a word-phrase novel composition, and a phrase-phrase novel composition respectively. As shown, by applying our framework on the baseline model, our method performs better when handling novel compositions of the learned concepts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.47 + ], + "angle": 0, + "content": "word level novel compositions. This means that in this variant, we omit the virtual testing set for handling word-word level novel compositions during meta optimization. Similarly, the other variants (w/o word-phrase novel compositions and w/o phrase-phrase novel compositions) ignore the corresponding level of novel compositions respectively. As shown in Table 3, ignoring any level of novel compositions leads to performance drop compared to our framework, demonstrating that each level of novel compositions can affect model generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.624 + ], + "angle": 0, + "content": "Impact of meta optimization. To investigate the effectiveness of the meta optimization scheme in our framework, we evaluate a variant (training w/o meta). In this variant, we train the model in the conventional manner on the constructed virtual training set and virtual testing sets, i.e., without meta optimization. Note that for this variant, we still construct the virtual training set and virtual testing sets in the same way as in our framework. As shown in Table 4, our framework outperforms this variant obviously, demonstrating the effectiveness of our meta optimization scheme." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.883 + ], + "angle": 0, + "content": "Impact of multiple virtual testing sets and curriculum learning. In our framework, we construct multiple virtual testing sets to handle various levels of novel compositions. Moreover, since the various levels of novel compositions correspond to different levels of comprehension complexity, we adopt a curriculum learning strategy to facilitate model training. To investigate the effectiveness of such design, we evaluate two variants. One variant (one virtual testing set) constructs only one virtual testing set to handle all levels of novel compositions. Such a virtual testing set includes all the data samples containing any level of novel compositions w.r.t. the virtual training set. Another variant (multiple virtual testing sets w/o curriculum learning) constructs multiple virtual testing sets as in our framework, but does not adopt the curriculum learning strategy. We compare these two variants to our original framework (multiple virtual testing sets w/ curriculum learning)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.891, + 0.9 + ], + "angle": 0, + "content": "As shown in Table 5, compared to the variant construct" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19484" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "ing one virtual testing set, the other variant obtains better performance, showing the superiority of using multiple virtual testing sets to handle various levels of novel compositions. Furthermore, by employing the curriculum learning strategy, our framework performs better than the variant constructing multiple virtual testing sets without curriculum learning, demonstrating the effectiveness of our curriculum learning strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.473, + 0.412 + ], + "angle": 0, + "content": "Impact of the size of virtual training set. In our framework, we randomly sample \\(60\\%\\) of the training data as the virtual training set, and use the remaining \\(40\\%\\) of the training data to construct virtual testing sets \\((60\\%:40\\%)\\). Here we test two variants. One variant \\((50\\%:50\\%)\\) uses \\(50\\%\\) of the training data to construct the virtual training set, and the remaining \\(50\\%\\) to construct virtual testing sets. While another variant \\((70\\%:30\\%)\\) uses \\(70\\%\\) of the training data for the virtual training set, and the remaining \\(30\\%\\) part for virtual testing sets. As shown in Table 6, our method and these two variants all perform better than the baseline model (i.e., LAVT [37]), showing the robustness of our framework in terms of the varying size of virtual training set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.61 + ], + "angle": 0, + "content": "Impact of virtual testing sets construction strategy. In our framework, after sampling a subset of the training set as the virtual training set, we use the remaining training data to construct multiple virtual testing sets. Each virtual testing set consists of the data samples containing one level of novel compositions w.r.t. the virtual training set. To explore the efficacy of such a strategy, we evaluate a variant (random virtual testing sets), in which we totally randomly select training data to construct each virtual testing set. As shown in Table 7, our framework obviously outperforms this variant. This shows that our virtual testing sets construct strategy can effectively help our framework to improve model generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.473, + 0.718 + ], + "angle": 0, + "content": "Training time. As shown in Table 8, we test the training time of our framework that trains the baseline network [37] with meta optimization, and compare it to the training time of the baseline that trains the same network in the conventional manner without meta optimization, on RefCOCO dataset. Though our framework performs better, it brings only relatively little increase (18.18%) in training time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Impact of additional gradient updates. As discussed above, compared to the baseline model, our framework trains the model for longer time and involves more gradient updates. To explore whether the performance improvement of our framework comes from the additional gradient updates, we test a variant (baseline w/ additional gradient updates) in which we train the baseline model (following the original training strategy) for as many iterations as in our framework. As shown in Table 9, the performance of this variant is very close to the baseline model [37], and is obviously worse than our framework. This might be because that the originally trained baseline models have al" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 5. We evaluate two variants to test the impact of using multiple virtual testing sets and the curriculum learning strategy." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.118, + 0.892, + 0.158 + ], + "angle": 0, + "content": "
MethodoloUP@0.5P@0.6P@0.7P@0.8P@0.9
Ours (one virtual testing set)73.6484.8282.0876.2465.4734.64
Ours (multiple virtual testing sets w/o curriculum learning)74.0285.3482.6376.8766.0934.95
Ours (multiple virtual testing sets w/ curriculum learning)74.9286.2383.4577.2566.5635.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Table 6. We test different variants that utilize different proportions of the training data to construct the virtual training set and virtual testing sets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.276 + ], + "angle": 0, + "content": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline (LAVT)72.7384.4681.2475.2864.7134.30
Ours (50%:50%)74.7886.1183.3277.0866.3835.54
Ours (60%:40%)74.9286.2383.4577.2566.5635.61
Ours (70%:30%)74.7486.0483.2677.0166.2935.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.284, + 0.892, + 0.312 + ], + "angle": 0, + "content": "Table 7. We evaluate a variant to investigate the efficacy of virtual testing sets construction strategy in our framework." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.313, + 0.892, + 0.351 + ], + "angle": 0, + "content": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Random virtual testing sets72.8184.5681.2875.3464.7434.32
Ours74.9286.2383.4577.2566.5635.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.36, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Table 8. Comparison of the training time. Note that our method achieves much better performance than the baseline." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.388, + 0.892, + 0.422 + ], + "angle": 0, + "content": "
MethodTraining timeoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline33 hours72.7384.4681.2475.2864.7134.30
Ours39 hours74.9286.2383.4577.2566.5635.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.43, + 0.892, + 0.458 + ], + "angle": 0, + "content": "Table 9. We test a variant to investigate the impact of additional gradient updates on model performance." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.459, + 0.892, + 0.505 + ], + "angle": 0, + "content": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline72.7384.4681.2475.2864.7134.30
Baseline w/ additional gradient updates72.7484.4381.2575.2564.7334.28
Baseline w/ ours74.9286.2383.4577.2566.5635.61
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.586 + ], + "angle": 0, + "content": "ready reached convergence under the original training strategy, and thus additional gradient updates would not bring obvious benefits. Such results further validate the effectiveness of our framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.612, + 0.62, + 0.628 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.822 + ], + "angle": 0, + "content": "In this work, we proposed a meta learning-based framework (MCRES) to improve the generalization performance of RES models, especially when handling novel compositions of learned concepts. By constructing a virtual training set and multiple virtual testing sets w.r.t. various levels of novel compositions and then optimizing the model via meta optimization, our framework can effectively improve model generalization performance. Extensive experiments show that our framework achieves superior performance on widely used benchmarks. Moreover, our framework is flexible, and can be seamlessly applied on various models with different architectures to enhance their performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by MOE AcRF Tier 2 (Proposal ID: T2EP20222-0035), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2020-065), and SUTD SKI Project (SKI 2021_02_06)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19485" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.471, + 0.212 + ], + "angle": 0, + "content": "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.214, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.409 + ], + "angle": 0, + "content": "[5] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.48 + ], + "angle": 0, + "content": "[6] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[8] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Era: Expert retrieval and assembly for early action prediction. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 670-688. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. Allennlp: A deep semantic natural language processing platform. arXiv preprint arXiv:1803.07640, 2018. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[10] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[11] Jianzhu Guo, Xiangyu Zhu, Chenxu Zhao, Dong Cao, Zhen Lei, and Stan Z Li. Learning meta face recognition in unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163-6172, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[12] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4424-4433, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[14] Chao Huang, Zhangjie Cao, Yunbo Wang, Jianmin Wang, and Mingsheng Long. Metasets: Meta-learning on point sets for generalizable representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8863-8872, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[15] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10488-10497, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[16] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In European Conference on Computer Vision, pages 59-75. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[17] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[18] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[20] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[22] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1271-1280, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[23] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[24] Gen Luo, Yiyi Zhou, Rongrong Ji, Xiaoshuai Sun, Jinsong Su, Chia-Wen Lin, and Qi Tian. Cascade grouped attention network for referring expression segmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1274-1282, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collabora" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19486" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "tive network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[26] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[27] Edgar Margffoy-Tuay, Juan C Pérez, Emilio Botero, and Pablo Arbeláez. Dynamic multimodal instance segmentation guided by natural language queries. In Proceedings of the European Conference on Computer Vision (ECCV), pages 630-645, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[28] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In European Conference on Computer Vision, pages 792-807. Springer, 2016. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[29] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[31] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.608 + ], + "angle": 0, + "content": "[33] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[34] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11686-11695, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.748 + ], + "angle": 0, + "content": "[35] Li Xu, Haoxuan Qu, Jason Kuen, Juixiang Gu, and Jun Liu. Meta spatio-temporal debiasing for video scene graph generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 374-390. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[36] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[37] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[38] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[39] Nanyang Ye, Jingxuan Tang, Huayu Deng, Xiao-Yun Zhou, Qianxiao Li, Zhenguo Li, Guang-Zhong Yang, and Zhanxing Zhu. Adversarial invariant learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12441-12449. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.302 + ], + "angle": 0, + "content": "[40] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.359 + ], + "angle": 0, + "content": "[41] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.893, + 0.43 + ], + "angle": 0, + "content": "[42] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision. Springer, 2022. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19487" + } + ] +] \ No newline at end of file diff --git a/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_origin.pdf b/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3618f6c9fe8bd8a7fa7d85c6e1dbd3d473a4becd --- /dev/null +++ b/2023/Meta Compositional Referring Expression Segmentation/9f238d16-af3f-474e-8b36-772f72bf116f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b441b8537c63a5c8fd88c9c924f4896b0b86b7c3ed59c14a1cdd0904203cc25e +size 735457 diff --git a/2023/Meta Compositional Referring Expression Segmentation/full.md b/2023/Meta Compositional Referring Expression Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9c7e6af6f3c0256970b3eff0b46cb9e3b2ca390a --- /dev/null +++ b/2023/Meta Compositional Referring Expression Segmentation/full.md @@ -0,0 +1,281 @@ +# Meta Compositional Referring Expression Segmentation + +Li Xu $^{1}$ , Mark He Huang $^{1,3}$ , Xindi Shang $^{2}$ , Zehuan Yuan $^{2}$ , Ying Sun $^{3}$ , Jun Liu $^{1*}$ $^{1}$ Singapore University of Technology and Design, Singapore + $^{2}$ ByteDance + +$^{3}$ Institute for Infocomm Research (I $^2$ R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore + +{li_xu, he_huang}@mymail.sutd.edu.sg shangxindi, yuanzehuan}@bytedance.com @i2r.a-star.edu.sg, jun_liu@sutd.edu.sg + +# Abstract + +Referring expression segmentation aims to segment an object described by a language expression from an image. Despite the recent progress on this task, existing models tackling this task may not be able to fully capture semantics and visual representations of individual concepts, which limits their generalization capability, especially when handling novel compositions of learned concepts. In this work, through the lens of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework to enhance model compositional generalization performance. Specifically, to handle various levels of novel compositions, our framework first uses training data to construct a virtual training set and multiple virtual testing sets, where data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set. Then, following a novel meta optimization scheme to optimize the model to obtain good testing performance on the virtual testing sets after training on the virtual training set, our framework can effectively drive the model to better capture semantics and visual representations of individual concepts, and thus obtain robust generalization performance even when handling novel compositions. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our framework. + +# 1. Introduction + +Referring expression segmentation (RES) [12, 38, 40] aims to segment a visual entity in an image given a linguistic expression. This task has been receiving increasing attention in recent years [5, 18, 34, 37], as it can play an important role in various applications, such as language-based human-robot interaction and interactive image edit + +![](images/275a82f17cd616dbd8a9ebcec3ad7080ee08243791991194e43b43d145fd3385.jpg) +(a) + +![](images/385f94c0a861a9b6c4e80d8271248fa0bd5b0956d5cb67b60b7d1b777ee9a8db.jpg) +(b) +Figure 1. Illustration of novel compositions and various levels of compositions. (a) An example of the testing sample containing the novel composition of "dark coffee" in RefCOCO dataset [41]. Such a novel composition itself does not exist in the training data, but its individual components (i.e., "dark" and "coffee") exist in the training data. (b) An example of various levels of compositions in an expression. We perform constituency parsing of the expression using AllenNLP [9]. Based on the obtained parsing tree as shown above, we can then obtain various levels of compositions (e.g., word-word level, word-phrase level) in this expression. + +ing. However, despite the recent progress on tackling this task [5, 34, 42], existing methods may struggle with handling the testing samples, of which the expressions contain novel compositions of learned concepts. Here a novel composition means that the composition itself does not exist in the training data, but its individual components (e.g., words, phrases) exist in the training data, as shown in Fig. 1a. + +We observe that testing samples containing such novel compositions of learned concepts widely exist in RES datasets [26, 28, 41]. However, existing RES models may not be able to well handle novel compositions during testing. Here we test the generalization capability of multiple state-of-the-art models [25, 37, 42] in terms of handling + +novel compositions, as shown in Table 2. Specifically, we first split each testing set of the RefCOCO dataset. In each testing set, one split subset includes the data samples, in which all the contained compositions are seen in the RefCOCO training set. While another subset includes the data samples containing novel compositions, of which the individual components (e.g., words, phrases) exist in the RefCOCO training set but the composition itself is unseen in the training set, i.e., containing novel compositions of learned concepts. Then we evaluate the models [25, 37, 42] on the two subsets in each testing set, and find that for each model, its testing performance on the subset containing novel compositions drops obviously compared to the performance on the other subset. For these models, the performance gap between the two subsets can reach $14\% - 17\%$ measured by the metric of overall IoU. Such a clear performance gap indicates that existing models struggle with generalizing to novel compositions of learned concepts. This might due to that the model does not effectively capture the semantics and visual representations of individual concepts (e.g., "dark", "coffee" in Fig. 1a) during training. Then the trained model may fail to recognize a novel composition (e.g., "dark coffee") at testing time, which though is composed of learned concepts. + +Thus to handle this issue, we aim to train the model to effectively capture the semantics and visual representations of individual concepts during training. Despite the conceptual simplicity, how to guide the model's learning behavior towards this goal is a challenging problem. Here from the perspective of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework, to effectively handle such a challenging problem by only changing the model training scheme. + +Meta learning proposes to perform virtual testing during model training for better performance [7, 29]. Inspired by this, to improve the generalization capability of RES models, our MCRES framework incorporates a meta optimization scheme that consists of three steps: virtual training, virtual testing and meta update. Specifically, we first split the training set to construct a virtual training set for virtual training, and a virtual testing set for virtual testing. The data samples in the virtual testing set contain novel compositions w.r.t. the virtual training set. For example, if the expressions of data samples in the virtual training set contain both words "dark" and "coffee" but do not contain their composition (i.e., "dark coffee"), the virtual testing set can include this novel composition correspondingly. + +Based on the constructed virtual training set and virtual testing set, we first train the model using the virtual training set, and then evaluate the trained model on the virtual testing set. During virtual training, the model may learn the compositions of individual concepts as a whole without truly understanding the semantics and visual representa + +tions of individual concepts, which though can still improve model training performance. For example, if there are many training samples containing the composition of "yellow banana" in the virtual training set, the model can superficially correlate "banana" with "yellow" and learn this composition as a whole, since using such spurious correlations can facilitate the model learning [1, 10, 39]. However, learning the compositions as a whole over the virtual training set may not improve model performance much on the virtual testing set in virtual testing, since the virtual testing set contains novel compositions w.r.t. the virtual training set. Thus to achieve good testing performance on such a virtual testing set, the model needs to effectively capture semantics and visual representations of individual concepts during virtual training. In this way, the model testing performance on the virtual testing set serves as a generalization feedback to the model virtual training process. + +Thus after the virtual training and virtual testing, we can further update the model to obtain better testing performance on the virtual testing set (i.e., meta update), so as to drive the model training on the virtual training set towards the direction of learning to capture semantics and visual representations of individual concepts, i.e., learning to learn. In this manner, our framework is able to optimize the model for robust generalization performance, even tackling the challenging testing samples with novel compositions. + +Moreover, given that expressions can often be hierarchically decomposed, there can exist various levels of novel compositions. Specifically, to identify meaningful compositions in an expression, we can parse an expression into a tree structure based on the constituency parsing tool [9] as shown in Fig.1b. In such a parsing tree, under the same parent node, each pair of child nodes (e.g., "white" and "bird") are closely semantically related, and thus can form a meaningful composition. Since each child node can be a word or a phrase as in Fig. 1b, there can naturally exist the following three levels of novel compositions: word-word level (e.g., "white" and "bird"), word-phrase level (e.g., "standing" and "in water") and phrase-phrase level (e.g., "white bird" and "standing in water"), which correspond to different levels of comprehension complexity. To better handle such a range of novel compositions, we construct multiple virtual testing sets in our framework, where each virtual testing set is constructed to handle one level of novel compositions. + +Our framework only changes the model training scheme without the need to change the model structure. Thus our framework is general, and can be conveniently applied on various RES models. We test our framework on multiple models, and obtain consistent performance improvement. + +The contributions of our work are threefold: 1) We propose a novel framework (MCRES) to effectively improve generalization performance of RES models, especially when handing novel compositions of learned con + +cepts. 2) Via constructing a virtual training set and multiple virtual testing sets w.r.t. various levels of novel compositions, our framework can train the model to well handle various levels of novel compositions. 3) When applied on various models on three RES benchmarks [28, 41], our framework achieves consistent performance improvement. + +# 2. Related Works + +Referring Expression Segmentation (RES). RES [12] aims to segment a target object from an image based on an expression. Early works [20, 22, 28, 41] employed convolutional and recurrent networks to extract visual and linguistic features respectively, and then fused the extracted features to predict the segmentation mask. Recently, a series of works [5, 18, 37] employed vision and language transformers to boost performance. Besides, with the development of large-scale pretrained models, Wang et al. [34] proposed to leverage CLIP [30] to improve cross-modal matching. + +Some methods have been explored to help model better understand learned concepts in RES. Yu et al. [40] proposed a modular network that uses different modules to process different types of information in the given expression. Yang et al. [36] designed a reasoning module to help align the language concepts with visual regions. Different from all the above-mentioned works, we propose an MCRES framework to drive RES models to better capture semantics and visual representations of individual concepts. Such a framework only changes the model training scheme, and thus can be flexibly applied on various models to improve their generalization performance. + +Meta Learning. Meta learning, i.e., the paradigm of learning to learn, has emerged to mainly tackle the few-shot learning problem [7, 29, 31, 32]. MAML [7] and its following works [29, 31] aim to learn a good initialization of network parameters, to achieve fast test-time update to adapt to new few-shot learning tasks. More recently, meta learning has also been explored in other areas [8, 11, 14, 19, 35] to enhance model generalization performance without the need of test-time update. Inspired by these works, we leverage a meta learning-based framework to improve generalization performance of RES models especially when handling novel compositions. + +# 3. Method + +Existing RES models may fail to generalize to data samples containing novel compositions of learned concepts. To handle this issue, we aim to encourage the model to better capture the semantics of individual concepts as well as recognize their visual correspondences during training, which however is a non-trivial problem. In this paper, from the perspective of meta learning, we introduce MCRES framework optimizing RES models via a meta optimization + +scheme to improve their generalization performance. + +Specifically, as shown in Fig. 2, in our framework, we first split the original training set $(\mathcal{D}_{train})$ to build a virtual training set $(\mathcal{D}_{v\_ tr})$ for virtual training, and a group of $K$ virtual testing sets $\left(\{\mathcal{D}_{v\_ te}^{k}\}_{k=1}^{K}\right)$ for virtual testing. Each virtual testing set consists of data samples containing one level of novel compositions w.r.t. the virtual training set. We first use the virtual training set to train the model (virtual training), and then perform model testing on the virtual testing sets (virtual testing). Since the virtual testing sets contain novel compositions of individual concepts from the virtual training set, if the model can still achieve robust testing performance on the virtual testing sets after training on the virtual training set, we posit that the trained model has learned to capture more semantics and visual representations of individual concepts during the virtual training. Guided by this principle, we can optimize the model virtual testing performance, to guide the model virtual training process towards learning more semantics and representations of individual concepts. Below, we first introduce the pipeline of our framework, and then detail the virtual training set and virtual testing sets construction process. + +# 3.1. Framework + +We first train the RES model using the virtual training set, i.e., virtual training. Specifically, we denote the model parameters as $\theta$ , the loss function (e.g., cross-entropy loss) for training the RES model as $\mathcal{L}$ . Thus we can calculate the model virtual training loss $(\mathcal{L}_{v\_ tr})$ over the virtual training set $(\mathcal{D}_{v\_ tr})$ as: + +$$ +\mathcal {L} _ {v \_ t r} (\theta) = \mathcal {L} (\theta ; \mathcal {D} _ {v \_ t r}) \tag {1} +$$ + +Based on this loss, we then update the model parameters $(\theta)$ as follows: + +$$ +\theta^ {\prime} = \theta - \alpha \nabla_ {\theta} \mathcal {L} _ {v _ {-} t r} (\theta) \tag {2} +$$ + +where $\alpha$ denotes the learning rate. It is worth noting that the model update at this step is virtual, and the updated parameters $\theta'$ only serve as intermediate parameters for the model evaluation at the following virtual testing step. + +After the virtual training, we evaluate the generalization performance of the trained model on the virtual testing sets $\left(\{\mathcal{D}_{v\_te}^k\}_{k = 1}^K\right)$ , i.e., virtual testing. Specifically, on each virtual testing set $\mathcal{D}_{v\_te}^k$ , we compute the model loss $\mathcal{L}_{v\_te}^k$ as: + +$$ +\mathcal {L} _ {v. t e} ^ {k} \left(\theta^ {\prime}\right) = \mathcal {L} \left(\theta^ {\prime}; \mathcal {D} _ {v. t e} ^ {k}\right) \tag {3} +$$ + +Such a loss measures how well the model generalizes to the virtual testing set after training on the virtual training set. + +Then we perform the meta update step. At this step, we wish to actually update the model parameters $(\theta)$ , so that after the model training on the virtual training set, the trained model can generalize well to the virtual testing sets, i.e., generalizing well to novel compositions. To this end, we + +![](images/a4b218637b9ae77833b8d3d5ab604e43a0412f521608cc2beeb21c414e829b77.jpg) +(a) + +![](images/e801d30b4da5eed6ad8ff7092a0717d0e9fd40ac124d0bfb9c67d3a8aaf55b78.jpg) +(b) +Figure 2. Framework overview. Fig. (a) illustrates our framework pipeline. First, we train the model using the virtual training set $(\mathcal{D}_{v\_ tr})$ and then obtain the updated model. We then test the model with updated parameters $(\theta^{\prime})$ on multiple virtual testing sets $\left(\{\mathcal{D}_{v\_ te}^{k}\}_{k=1}^{K}\right)$ . According to the virtual testing losses, we perform meta update to optimize the model for better generalization capability. Fig. (b) shows that we construct a virtual training set and multiple virtual testing sets to handle various levels of novel compositions. The expressions of data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set. + +formulate the optimization objective as: + +$$ +\begin{array}{l} \min _ {\theta} \mathcal {L} _ {v. t r} (\theta) + \sum_ {k = 1} ^ {K} \mathcal {L} _ {v. t e} ^ {k} \left(\theta^ {\prime}\right) \tag {4} \\ = \min _ {\theta} \mathcal {L} _ {v - t r} (\theta) + \sum_ {k = 1} ^ {K} \mathcal {L} _ {v - t e} ^ {k} \left(\theta - \alpha \nabla_ {\theta} \mathcal {L} _ {v - t r} (\theta)\right) \\ \end{array} +$$ + +where the first term indicates the model training performance on the virtual training set, and the second term represents the model testing performance on the virtual testing sets after training on the virtual training set. Based on this objective, we then update the model parameters $(\theta)$ as: + +$$ +\theta \leftarrow \theta - \beta \nabla_ {\theta} \left(\mathcal {L} _ {v _ {-} t r} (\theta) + \sum_ {k = 1} ^ {K} \mathcal {L} _ {v _ {-} t e} ^ {k} \left(\theta - \alpha \nabla_ {\theta} \mathcal {L} _ {v _ {-} t r} (\theta)\right)\right) \tag {5} +$$ + +where $\beta$ is the learning rate for meta update. Through the above optimization process, the model is pushed to learn more semantics and visual representations of individual concepts during training. + +During the above process, the model is first trained (updated) on the virtual training set. At this step, the model may learn the compositions of individual concepts as a whole without effectively capturing the semantics and visual representations of individual concepts, which though can still improve its training performance on the virtual training set. However, to achieve good testing performance on the virtual testing data which contains novel compositions of individual concepts from the virtual training data, the model is expected to avoid learning compositions as a + +whole and instead capture more semantics and representations of individual concepts. In this way, the second term of Eqn. 5, which includes the second-order gradients of $\theta$ : $\nabla_{\theta}\mathcal{L}_{v\_te}^{k}\big(\theta -\alpha \nabla_{\theta}\mathcal{L}_{v\_tr}(\theta)\big)$ , can be regarded as a generalization feedback that can guide the model to capture more semantics and representations of individual concepts. + +In our framework, the above three steps (i.e., virtual training, virtual testing and meta update) are performed iteratively until the model training converges. + +# 3.2. Sets Construction + +In our framework, to handle various levels of novel compositions, we split the original training set to construct a virtual training set and various virtual testing sets. Each virtual testing set is expected to contain one level of novel compositions w.r.t. the virtual training set. Below we first introduce the details of these sets, and then discuss the strategy for constructing the virtual testing sets. + +We randomly sample a subset of the training set $(\mathcal{D}_{train})$ as the virtual training set $(\mathcal{D}_{v\_ tr})$ , and the remaining training data will be used as the candidate samples to construct a group of $(K)$ virtual testing sets $\left(\{\mathcal{D}_{v\_ te}^{k}\}_{k=1}^{K}\right)$ . Each virtual testing set is constructed to handle one level of novel compositions. Here considering the hierarchical semantic structure of each expression, we target the following levels of novel compositions: word-word level, word-phrase level and phrase-phrase level. A phrase means a group of words that serve as a grammatical unit in the expression (e.g., "white bird" in Fig. 1b), which can be conveniently identified using off-the-shelf tools (e.g., AllenNLP [9]). + +Specifically, to identify various levels of novel compo + +sitions in an expression, we first use the constituency parsing tool [9] to parse an expression into a tree structure as shown in Fig. 1b. In this tree, under the same parent node, each pair of child nodes are closely related, and thus can form a meaningful composition. If a pair of nodes having the same parent node are both words, they form a word-word level composition (e.g., "white" and "bird"). Similarly, if the nodes with the same parent node are a word and a phrase, they form a word-phrase level composition (e.g., "standing" and "in water"). In a similar way, we can obtain phrase-phrase level compositions (e.g., "white bird" and "standing in water"). Thus there can exist three levels of novel compositions, which correspond to different levels of comprehension complexity. Corresponding to these three levels of novel compositions, we will construct a total of three virtual testing sets $(K = 3)$ . + +Virtual testing sets construction. To construct each virtual testing set, we need to select data samples containing the corresponding level of novel compositions w.r.t. the virtual training set, from the candidate samples. A data sample means an expression paired with the corresponding image. Here we design an efficient strategy, which can be leveraged to construct each virtual testing set. Below, we take the process of constructing the virtual testing set for handling word-word level novel compositions as an example, to introduce such a strategy. Specifically, this virtual testing set should include all of those candidate samples that contain word-word level novel compositions w.r.t. the virtual training set. To construct such a virtual testing set, our strategy proceeds as follows. + +(i) To find the data samples, of which the expressions contain word-word level novel compositions w.r.t. the virtual training set, from the candidate samples, we need to first obtain all the word-word level compositions in the virtual training set and in the candidate samples respectively. Thus by parsing each expression into a parsing tree as shown in Fig. 1b, we can obtain all word-word level compositions (e.g., "white" and "bird", "in" and "water" in Fig. 1b) in each expression. (ii) To identify word-word level novel compositions, we first select the word-word level compositions that exist in the candidate samples but are unseen in the virtual training set. Then for each selected composition, we further check if its individual words exist in the virtual training set. If so, such a composition will be identified as a word-word level novel composition w.r.t. the virtual training set. For example, if the composition "white bird" is unseen in the virtual training set, while its individual words "white" and "bird" both exist in the virtual training set, this composition is a word-word level novel composition w.r.t. the virtual training set. (iii) Finally, we select the candidate samples, of which the expressions contain the identified word-word level novel compositions w.r.t. the virtual training set, to construct the virtual testing set. + +To efficiently perform the above steps, we can employ parallel matrix operations. Specifically, to record the word-word level compositions in the virtual training set, we build a matrix $\mathcal{M}_{v,tr}$ with the shape of $|\mathcal{V}| \times |\mathcal{V}|$ , where $|\mathcal{V}|$ is the size of the vocabulary set $\mathcal{V}$ of the original training set. In this matrix, each element at the $i-th$ row and the $j-th$ column $(i,j \in \{1,\dots,|\mathcal{V}|\})$ is a binary value, and the value 1 indicates that the $i-th$ word and $j-th$ word in the vocabulary set $\mathcal{V}$ form a word-word level composition in the virtual training set, while 0 means such a word-word level composition does not exist in the virtual training set. Similarly, we can also build a matrix $\mathcal{M}_{candi}$ to record the word-word level compositions in the candidate samples. Then to identify word-word level novel compositions w.r.t. the virtual training set $(\mathcal{M}_{v,tr})$ from the candidate samples $(\mathcal{M}_{candi})$ , we can efficiently compute a difference matrix: $\mathcal{M}_{diff} = \mathcal{M}_{candi} - \mathcal{M}_{v,tr}$ . In $\mathcal{M}_{diff}$ , any element with the value 1 indicates that the corresponding composition exists in the candidate samples, but is unseen in the virtual training set. Then by further checking whether the individual words in such a composition exist in the virtual training set, we can determine if it is a word-word level novel composition w.r.t. the virtual training set. + +Similarly, we can efficiently construct the other two virtual testing sets. Note that to help the model to learn to handle a wide range of possible novel compositions, at the beginning of each training epoch, we randomly sample a subset of the training data to re-construct the virtual training set, and leverage the remaining data to re-construct the virtual testing sets using the above strategy. + +As discussed above, to handle various levels of novel compositions, we construct multiple virtual testing sets. A simpler alternative is to construct only one virtual testing set, which includes all the data samples containing any level of novel compositions w.r.t. the virtual training set. However, compared to this alternative, by constructing multiple virtual testing sets, each level of novel compositions in the corresponding virtual testing set can offer an explicit generalization feedback during model training. Thus the model is explicitly encouraged to well handle each level of novel compositions. Moreover, constructing multiple virtual testing sets can facilitate the use of curriculum learning strategy for model training as discussed in Sec. 3.3. + +# 3.3. Training and Testing + +To help models generalize to novel compositions in RES, our framework only changes the model training scheme. Thus our framework is general, and can be flexibly applied to train various RES models. During training, at each epoch, we first split the training set to construct a virtual training set and multiple virtual testing sets. Then we iteratively optimize the model over the virtual training set and virtual testing sets via meta optimization. Specifically, for each meta + +Table 1. Comparison with the state-of-the-arts on three benchmark datasets using the metric of overall IoU. Moreover, we apply our framework on various models [25, 37, 42], and obtain consistent performance improvement. “-” indicates that the corresponding result is not provided in the original paper. U means the UMD split of RefCOCOg dataset, and G means the Google split. + +
MethodRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBval(U)test(U)val(G)
DMN [27]49.7854.8345.1338.8844.2232.29--36.76
RRN [20]55.3357.2653.9339.7542.1536.11--36.45
MAAttNet [40]56.5162.3751.7046.6752.3940.0847.6448.61-
CAC [4]58.9061.7753.81---46.3746.9544.32
CMSA [38]58.3260.6155.0943.7647.6037.89--39.98
STEP [3]60.0463.4657.9748.1952.33404.1--46.40
BRINet [13]60.9862.9959.2148.1752.3242.11--48.04
CMPC [15]61.3664.5359.6449.5653.4443.23--49.05
LSCM [16]61.4764.9959.5549.3453.1243.50--48.05
CMPC+ [23]62.4765.0860.8250.2554.0443.47--49.89
EFN [6]62.7665.6959.6751.5055.2443.01--51.93
BUSNet [36]63.2766.4161.3951.7656.8744.13--50.56
CGAN [24]64.8668.0462.0751.0355.5144.0651.0151.6946.54
LTS [17]65.4367.7663.0854.2158.3248.0254.4054.25-
VLT [5]65.6568.2962.7355.5059.2049.3652.9956.6549.76
ReSTR [18]67.2269.3064.4555.7860.4448.27--54.48
CRIS [34]70.4773.1866.1062.2768.0853.6859.8760.36-
MCN [25]62.4464.2059.7150.6254.9944.6949.2249.40-
MCN + Ours64.5166.4862.8452.2956.7146.7251.4951.63-
SeqTR [42]71.7073.3169.8263.0466.7358.9764.6965.74-
SeqTR + Ours73.2375.0171.9564.7167.8560.8566.7767.48-
LAVT [37]72.7375.8268.7962.1468.3855.1061.2462.0960.50
LAVT + Ours74.9276.9870.8464.3269.6856.6463.5164.9061.63
+ +optimization iteration, we use two batches of data samples: one batch of data for virtual training and the other batch of data for virtual testing. Thus we need to ensure that the batch of data for virtual testing consists of the data samples containing novel compositions w.r.t. the batch of data for virtual training. To this end, for each iteration, we randomly sample some data from each virtual testing set, to form the batch of data for virtual testing. Then we select the virtual training set samples that contain all the individual components of the novel compositions in the batch of data for virtual testing, to form the batch of data for virtual training. Note that the above data preparation procedures can be done before the model training. For model testing, we test the trained model in the conventional way. + +Besides, given that the various levels of novel compositions correspond to different levels of comprehension complexity, we adopt a curriculum learning strategy [2, 33] for model training, so that the model can progressively learn to handle various levels of novel compositions, i.e., from lower level (word-word) to middle level (word-phrase) to higher level (phrase-phrase), and thus can learn all these levels of novel compositions better. Specifically, in the first $1/3$ of training epochs, we only use the virtual testing set for handling word-word level novel compositions for meta optimization. Then in the middle $1/3 - 2/3$ of training epochs, we add the virtual testing set for handling word-phrase level novel compositions. Finally, in the remaining training epochs, we use all the three virtual testing sets. + +Table 2. We evaluate our framework on different testing subsets w.r.t. novel compositions to validate its effectiveness on optimizing the model to better generalize to novel compositions. We use the metric of overall IoU here. The performance gain of our framework compared to the corresponding baseline model is shown in parentheses. + +
MethodRefCOCO-valRefCOCO-testARefCOCO-testB
NovelNon-novelNovelNon-novelNovelNon-novel
MCN [23]53.1767.4155.4370.0850.6466.47
MCN + Ours57.38 (↑4.21)68.09 (↑0.68)59.66 (↑4.23)70.67 (↑0.59)55.87 (↑5.23)67.04 (↑0.57)
SeqTR [40]64.2478.5965.7179.4660.1477.21
SeqTR + Ours67.31 (↑3.07)79.06 (↑0.47)69.48 (↑3.77)79.84 (↑0.38)64.89 (↑4.75)77.68 (↑0.47)
LAVT [35]63.5278.5567.1780.7158.4976.23
LAVT + Ours67.42 (↑3.90)79.29 (↑0.74)70.05 (↑2.88)81.37 (↑0.66)62.63 (↑4.14)76.92 (↑0.69)
+ +# 4. Experiments + +We evaluate our method on three commonly used RES benchmarks: RefCOCO [41], RefCOCO+ [41] and RefCOCOg [26, 28]. Images in these three benchmarks are all from the COCO dataset [21]. RefCOCO [41] contains 19994 images, 50000 annotated objects with 142209 referring expressions. RefCOCO+ [41] consists of 19992 images with 49856 annotated objects, and 141564 expressions. Different from RefCOCO, words describing absolute spatial locations (e.g., left, front) are not allowed to be used in the expressions in RefCOCO+. In RefCOCO and RefCOCO+ datasets, following the original split in [41], the visual entities to be segmented in testA subset are people, while the ones in testB subset are objects (i.e., not people). RefCOCOg [26, 28] includes 26711 images, 54822 annotated objects and 104560 expressions. Compared with RefCOCO and RefCOCO+, the expressions in RefCOCOg are more complex, which have an average length of 8.4 words. There exist two different partitions for RefCOCOg dataset: UMD split [28] and Google split [26]. + +Evaluation metrics. Following [3,37,38], we report our results using two kinds of metrics: overall IoU (oIoU) and Precision@X (P@X). The overall IoU measures the ratio of total intersection regions over total union regions of predicted masks and ground truths of all testing samples. Precision@X calculates the percentage of testing samples, of which the model prediction has an IoU score higher than the threshold value X, and $\mathrm{X} \in \{0.5, 0.6, 0.7, 0.8, 0.9\}$ . + +Implementation details. We conduct our experiments on 8 Tesla V100 GPUs. We applied our framework on various RES models [25,37,42]. On each dataset, we randomly sample $60\%$ of the training data as the virtual training set and use the remaining training data to construct virtual testing sets at each training epoch. The learning rate $(\alpha)$ for virtual training is 5e-5, and the learning rate $(\beta)$ for meta update is 2e-5. + +# 4.1. Experimental Results + +As shown in Table 1, our framework achieves state-of-the-art performance across all three datasets, demonstrating the superiority of our framework. Moreover, we applied our + +Table 3. We test several variants to investigate the impact of each level of novel compositions on model performance. + +
MethodIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline (LAVT)72.7384.4681.2475.2864.7134.30
w/o word-word novel compositions73.4384.7181.6976.1265.2334.47
w/o word-phrase novel compositions73.6884.8281.7676.2865.3734.56
w/o phrase-phrase novel compositions73.5984.7381.7376.1665.3534.45
Ours74.9286.2383.4577.2566.5635.61
+ +Table 4. We evaluate a variant to test the effectiveness of meta optimization scheme in our framework. For this variant, its optimization objective is to minimize $\mathcal{L}_{v\_ tr}(\theta) + \sum_{k=1}^{K} \mathcal{L}_{v\_ te}^{k}(\theta)$ (i.e., replacing $\theta'$ with $\theta$ in Eqn. 4). + +
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Training w/o meta72.7684.5081.2775.3364.7234.31
Ours74.9286.2383.4577.2566.5635.61
+ +framework on two transformer-based SOTA models [37,42] and a CNN-based model [25]. Our framework brings consistent performance improvement on all these models and datasets. This shows that our framework can serve as a general approach to enhance model performance + +To further validate the effectiveness of our framework for handling novel compositions, we perform the following analysis. Specifically, we split each testing set of RefCOCO dataset to construct two subsets. One subset (Non-novel) includes the data samples, in which all the contained compositions are seen in the RefCOCO training set. While another subset (Novel) includes the data samples that contain any level of novel compositions w.r.t. the training set. As shown in Table 2, on each testing set, we can see a clear performance gap between the two subsets for each of the baseline models [25, 37, 42], showing that existing models struggle with handling novel compositions. Then by applying our framework on each baseline model, we obtain significant performance improvement on the subset of Novel. This validates the general effectiveness of our framework to optimize the model to well generalize to novel compositions. Moreover, our framework also slightly improves the model performance on the subset of Non-novel. This can be attributed to that by training the model to better capture semantics and visual representations of individual concepts, our framework can help the model to better understand the given expression and find its visual correspondence, which thus can generally improve model performance. We also show some qualitative results in Fig. 3. As shown, when handling the testing samples containing novel compositions, our framework achieves better performance than the baseline model [37]. + +# 4.2. Ablation Studies + +Following [3, 37, 38], we conduct ablation experiments to evaluate our framework on RefCOCO validation set. + +Impact of various levels of novel compositions. To investigate the impact of each level of novel compositions on model performance, we test multiple variants. One variant (w/o word-word novel compositions) ignores the word- + +![](images/1558cade3b9dbb7c371f8acd70512fe2ac0e69155f1bf1e60dbf25e697642e99.jpg) +Figure 3. Qualitative results of our method and the baseline model [37]. The above testing examples contain a word-word novel composition, a word-phrase novel composition, and a phrase-phrase novel composition respectively. As shown, by applying our framework on the baseline model, our method performs better when handling novel compositions of the learned concepts. + +word level novel compositions. This means that in this variant, we omit the virtual testing set for handling word-word level novel compositions during meta optimization. Similarly, the other variants (w/o word-phrase novel compositions and w/o phrase-phrase novel compositions) ignore the corresponding level of novel compositions respectively. As shown in Table 3, ignoring any level of novel compositions leads to performance drop compared to our framework, demonstrating that each level of novel compositions can affect model generalization performance. + +Impact of meta optimization. To investigate the effectiveness of the meta optimization scheme in our framework, we evaluate a variant (training w/o meta). In this variant, we train the model in the conventional manner on the constructed virtual training set and virtual testing sets, i.e., without meta optimization. Note that for this variant, we still construct the virtual training set and virtual testing sets in the same way as in our framework. As shown in Table 4, our framework outperforms this variant obviously, demonstrating the effectiveness of our meta optimization scheme. + +Impact of multiple virtual testing sets and curriculum learning. In our framework, we construct multiple virtual testing sets to handle various levels of novel compositions. Moreover, since the various levels of novel compositions correspond to different levels of comprehension complexity, we adopt a curriculum learning strategy to facilitate model training. To investigate the effectiveness of such design, we evaluate two variants. One variant (one virtual testing set) constructs only one virtual testing set to handle all levels of novel compositions. Such a virtual testing set includes all the data samples containing any level of novel compositions w.r.t. the virtual training set. Another variant (multiple virtual testing sets w/o curriculum learning) constructs multiple virtual testing sets as in our framework, but does not adopt the curriculum learning strategy. We compare these two variants to our original framework (multiple virtual testing sets w/ curriculum learning). + +As shown in Table 5, compared to the variant construct + +ing one virtual testing set, the other variant obtains better performance, showing the superiority of using multiple virtual testing sets to handle various levels of novel compositions. Furthermore, by employing the curriculum learning strategy, our framework performs better than the variant constructing multiple virtual testing sets without curriculum learning, demonstrating the effectiveness of our curriculum learning strategy. + +Impact of the size of virtual training set. In our framework, we randomly sample $60\%$ of the training data as the virtual training set, and use the remaining $40\%$ of the training data to construct virtual testing sets $(60\%:40\%)$ . Here we test two variants. One variant $(50\%:50\%)$ uses $50\%$ of the training data to construct the virtual training set, and the remaining $50\%$ to construct virtual testing sets. While another variant $(70\%:30\%)$ uses $70\%$ of the training data for the virtual training set, and the remaining $30\%$ part for virtual testing sets. As shown in Table 6, our method and these two variants all perform better than the baseline model (i.e., LAVT [37]), showing the robustness of our framework in terms of the varying size of virtual training set. + +Impact of virtual testing sets construction strategy. In our framework, after sampling a subset of the training set as the virtual training set, we use the remaining training data to construct multiple virtual testing sets. Each virtual testing set consists of the data samples containing one level of novel compositions w.r.t. the virtual training set. To explore the efficacy of such a strategy, we evaluate a variant (random virtual testing sets), in which we totally randomly select training data to construct each virtual testing set. As shown in Table 7, our framework obviously outperforms this variant. This shows that our virtual testing sets construct strategy can effectively help our framework to improve model generalization performance. + +Training time. As shown in Table 8, we test the training time of our framework that trains the baseline network [37] with meta optimization, and compare it to the training time of the baseline that trains the same network in the conventional manner without meta optimization, on RefCOCO dataset. Though our framework performs better, it brings only relatively little increase (18.18%) in training time. + +Impact of additional gradient updates. As discussed above, compared to the baseline model, our framework trains the model for longer time and involves more gradient updates. To explore whether the performance improvement of our framework comes from the additional gradient updates, we test a variant (baseline w/ additional gradient updates) in which we train the baseline model (following the original training strategy) for as many iterations as in our framework. As shown in Table 9, the performance of this variant is very close to the baseline model [37], and is obviously worse than our framework. This might be because that the originally trained baseline models have al + +Table 5. We evaluate two variants to test the impact of using multiple virtual testing sets and the curriculum learning strategy. + +
MethodoloUP@0.5P@0.6P@0.7P@0.8P@0.9
Ours (one virtual testing set)73.6484.8282.0876.2465.4734.64
Ours (multiple virtual testing sets w/o curriculum learning)74.0285.3482.6376.8766.0934.95
Ours (multiple virtual testing sets w/ curriculum learning)74.9286.2383.4577.2566.5635.61
+ +Table 6. We test different variants that utilize different proportions of the training data to construct the virtual training set and virtual testing sets. + +
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline (LAVT)72.7384.4681.2475.2864.7134.30
Ours (50%:50%)74.7886.1183.3277.0866.3835.54
Ours (60%:40%)74.9286.2383.4577.2566.5635.61
Ours (70%:30%)74.7486.0483.2677.0166.2935.52
+ +Table 7. We evaluate a variant to investigate the efficacy of virtual testing sets construction strategy in our framework. + +
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Random virtual testing sets72.8184.5681.2875.3464.7434.32
Ours74.9286.2383.4577.2566.5635.61
+ +Table 8. Comparison of the training time. Note that our method achieves much better performance than the baseline. + +
MethodTraining timeoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline33 hours72.7384.4681.2475.2864.7134.30
Ours39 hours74.9286.2383.4577.2566.5635.61
+ +Table 9. We test a variant to investigate the impact of additional gradient updates on model performance. + +
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline72.7384.4681.2475.2864.7134.30
Baseline w/ additional gradient updates72.7484.4381.2575.2564.7334.28
Baseline w/ ours74.9286.2383.4577.2566.5635.61
+ +ready reached convergence under the original training strategy, and thus additional gradient updates would not bring obvious benefits. Such results further validate the effectiveness of our framework. + +# 5. Conclusion + +In this work, we proposed a meta learning-based framework (MCRES) to improve the generalization performance of RES models, especially when handling novel compositions of learned concepts. By constructing a virtual training set and multiple virtual testing sets w.r.t. various levels of novel compositions and then optimizing the model via meta optimization, our framework can effectively improve model generalization performance. Extensive experiments show that our framework achieves superior performance on widely used benchmarks. Moreover, our framework is flexible, and can be seamlessly applied on various models with different architectures to enhance their performance. + +Acknowledgement. This work is supported by MOE AcRF Tier 2 (Proposal ID: T2EP20222-0035), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2020-065), and SUTD SKI Project (SKI 2021_02_06). + +# References + +[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2 +[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 6 +[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 6, 7 +[4] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 6 +[5] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 1, 3, 6 +[6] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 6 +[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 3 +[8] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Era: Expert retrieval and assembly for early action prediction. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 670-688. Springer, 2022. 3 +[9] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. Allennlp: A deep semantic natural language processing platform. arXiv preprint arXiv:1803.07640, 2018. 1, 2, 4, 5 +[10] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2 +[11] Jianzhu Guo, Xiangyu Zhu, Chenxu Zhao, Dong Cao, Zhen Lei, and Stan Z Li. Learning meta face recognition in unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163-6172, 2020. 3 +[12] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 1, 3 +[13] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4424-4433, 2020. 6 +[14] Chao Huang, Zhangjie Cao, Yunbo Wang, Jianmin Wang, and Mingsheng Long. Metasets: Meta-learning on point sets for generalizable representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8863-8872, 2021. 3 +[15] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10488-10497, 2020. 6 +[16] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In European Conference on Computer Vision, pages 59-75. Springer, 2020. 6 +[17] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 6 +[18] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 1, 3, 6 +[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 3 +[20] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2018. 3, 6 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6 +[22] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1271-1280, 2017. 3 +[23] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6 +[24] Gen Luo, Yiyi Zhou, Rongrong Ji, Xiaoshuai Sun, Jinsong Su, Chia-Wen Lin, and Qi Tian. Cascade grouped attention network for referring expression segmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1274-1282, 2020. 6 +[25] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collabora + +tive network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 1, 2, 6, 7 +[26] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 1, 6 +[27] Edgar Margffoy-Tuay, Juan C Pérez, Emilio Botero, and Pablo Arbeláez. Dynamic multimodal instance segmentation guided by natural language queries. In Proceedings of the European Conference on Computer Vision (ECCV), pages 630-645, 2018. 6 +[28] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In European Conference on Computer Vision, pages 792-807. Springer, 2016. 1, 3, 6 +[29] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2, 3 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3 +[31] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 3 +[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3 +[33] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6 +[34] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11686-11695, 2022. 1, 3, 6 +[35] Li Xu, Haoxuan Qu, Jason Kuen, Juixiang Gu, and Jun Liu. Meta spatio-temporal debiasing for video scene graph generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 374-390. Springer, 2022. 3 +[36] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 3, 6 +[37] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 1, 2, 3, 6, 7, 8 + +[38] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1, 6, 7 +[39] Nanyang Ye, Jingxuan Tang, Huayu Deng, Xiao-Yun Zhou, Qianxiao Li, Zhenguo Li, Guang-Zhong Yang, and Zhanxing Zhu. Adversarial invariant learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12441-12449. IEEE, 2021. 2 +[40] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 1, 3, 6 +[41] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 1, 3, 6 +[42] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision. Springer, 2022. 1, 2, 6, 7 \ No newline at end of file diff --git a/2023/Meta Compositional Referring Expression Segmentation/images.zip b/2023/Meta Compositional Referring Expression Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..902d321246b88c926fccb2dc0c77a44bb60acfd3 --- /dev/null +++ b/2023/Meta Compositional Referring Expression Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f00f574e32d1c27b692e9e8bac0c6e4395b2299de34a97c852c77b517ba8cea +size 449663 diff --git a/2023/Meta Compositional Referring Expression Segmentation/layout.json b/2023/Meta Compositional Referring Expression Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c36ddff922805d6c623edc5d220fadfd62ba0ec8 --- /dev/null +++ b/2023/Meta Compositional Referring Expression Segmentation/layout.json @@ -0,0 +1,7430 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "type": "text", + "content": "Meta Compositional Referring Expression Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "spans": [ + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "content": "Li Xu" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "content": ", Mark He Huang" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "content": ", Xindi Shang" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "content": ", Zehuan Yuan" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "content": ", Ying Sun" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "content": ", Jun Liu" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "content": "Singapore University of Technology and Design, Singapore \n" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 142, + 500, + 184 + ], + "type": "text", + "content": "ByteDance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 185, + 558, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 185, + 558, + 200 + ], + "spans": [ + { + "bbox": [ + 54, + 185, + 558, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 185, + 558, + 200 + ], + "type": "text", + "content": "Institute for Infocomm Research (I" + }, + { + "bbox": [ + 54, + 185, + 558, + 200 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 54, + 185, + 558, + 200 + ], + "type": "text", + "content": "R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 201, + 419, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 201, + 419, + 241 + ], + "spans": [ + { + "bbox": [ + 211, + 201, + 419, + 241 + ], + "type": "text", + "content": "{li_xu, he_huang}@mymail.sutd.edu.sg shangxindi, yuanzehuan}@bytedance.com @i2r.a-star.edu.sg, jun_liu@sutd.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 269, + 192, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 280 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 280 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 294, + 290, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 290, + 582 + ], + "type": "text", + "content": "Referring expression segmentation aims to segment an object described by a language expression from an image. Despite the recent progress on this task, existing models tackling this task may not be able to fully capture semantics and visual representations of individual concepts, which limits their generalization capability, especially when handling novel compositions of learned concepts. In this work, through the lens of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework to enhance model compositional generalization performance. Specifically, to handle various levels of novel compositions, our framework first uses training data to construct a virtual training set and multiple virtual testing sets, where data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set. Then, following a novel meta optimization scheme to optimize the model to obtain good testing performance on the virtual testing sets after training on the virtual training set, our framework can effectively drive the model to better capture semantics and visual representations of individual concepts, and thus obtain robust generalization performance even when handling novel compositions. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 604, + 127, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 127, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 127, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "type": "text", + "content": "Referring expression segmentation (RES) [12, 38, 40] aims to segment a visual entity in an image given a linguistic expression. This task has been receiving increasing attention in recent years [5, 18, 34, 37], as it can play an important role in various applications, such as language-based human-robot interaction and interactive image edit" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 269, + 526, + 353 + ], + "blocks": [ + { + "bbox": [ + 326, + 269, + 526, + 353 + ], + "lines": [ + { + "bbox": [ + 326, + 269, + 526, + 353 + ], + "spans": [ + { + "bbox": [ + 326, + 269, + 526, + 353 + ], + "type": "image", + "image_path": "275a82f17cd616dbd8a9ebcec3ad7080ee08243791991194e43b43d145fd3385.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 356, + 432, + 365 + ], + "lines": [ + { + "bbox": [ + 421, + 356, + 432, + 365 + ], + "spans": [ + { + "bbox": [ + 421, + 356, + 432, + 365 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 329, + 372, + 525, + 430 + ], + "blocks": [ + { + "bbox": [ + 329, + 372, + 525, + 430 + ], + "lines": [ + { + "bbox": [ + 329, + 372, + 525, + 430 + ], + "spans": [ + { + "bbox": [ + 329, + 372, + 525, + 430 + ], + "type": "image", + "image_path": "385f94c0a861a9b6c4e80d8271248fa0bd5b0956d5cb67b60b7d1b777ee9a8db.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 431, + 432, + 440 + ], + "lines": [ + { + "bbox": [ + 420, + 431, + 432, + 440 + ], + "spans": [ + { + "bbox": [ + 420, + 431, + 432, + 440 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 440, + 545, + 550 + ], + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 550 + ], + "type": "text", + "content": "Figure 1. Illustration of novel compositions and various levels of compositions. (a) An example of the testing sample containing the novel composition of \"dark coffee\" in RefCOCO dataset [41]. Such a novel composition itself does not exist in the training data, but its individual components (i.e., \"dark\" and \"coffee\") exist in the training data. (b) An example of various levels of compositions in an expression. We perform constituency parsing of the expression using AllenNLP [9]. Based on the obtained parsing tree as shown above, we can then obtain various levels of compositions (e.g., word-word level, word-phrase level) in this expression." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": "ing. However, despite the recent progress on tackling this task [5, 34, 42], existing methods may struggle with handling the testing samples, of which the expressions contain novel compositions of learned concepts. Here a novel composition means that the composition itself does not exist in the training data, but its individual components (e.g., words, phrases) exist in the training data, as shown in Fig. 1a." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": "We observe that testing samples containing such novel compositions of learned concepts widely exist in RES datasets [26, 28, 41]. However, existing RES models may not be able to well handle novel compositions during testing. Here we test the generalization capability of multiple state-of-the-art models [25, 37, 42] in terms of handling" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19478" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "novel compositions, as shown in Table 2. Specifically, we first split each testing set of the RefCOCO dataset. In each testing set, one split subset includes the data samples, in which all the contained compositions are seen in the RefCOCO training set. While another subset includes the data samples containing novel compositions, of which the individual components (e.g., words, phrases) exist in the RefCOCO training set but the composition itself is unseen in the training set, i.e., containing novel compositions of learned concepts. Then we evaluate the models [25, 37, 42] on the two subsets in each testing set, and find that for each model, its testing performance on the subset containing novel compositions drops obviously compared to the performance on the other subset. For these models, the performance gap between the two subsets can reach " + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "inline_equation", + "content": "14\\% - 17\\%" + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": " measured by the metric of overall IoU. Such a clear performance gap indicates that existing models struggle with generalizing to novel compositions of learned concepts. This might due to that the model does not effectively capture the semantics and visual representations of individual concepts (e.g., \"dark\", \"coffee\" in Fig. 1a) during training. Then the trained model may fail to recognize a novel composition (e.g., \"dark coffee\") at testing time, which though is composed of learned concepts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 361, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 289, + 471 + ], + "type": "text", + "content": "Thus to handle this issue, we aim to train the model to effectively capture the semantics and visual representations of individual concepts during training. Despite the conceptual simplicity, how to guide the model's learning behavior towards this goal is a challenging problem. Here from the perspective of meta learning, we propose a Meta Compositional Referring Expression Segmentation (MCRES) framework, to effectively handle such a challenging problem by only changing the model training scheme." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 472, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 289, + 640 + ], + "type": "text", + "content": "Meta learning proposes to perform virtual testing during model training for better performance [7, 29]. Inspired by this, to improve the generalization capability of RES models, our MCRES framework incorporates a meta optimization scheme that consists of three steps: virtual training, virtual testing and meta update. Specifically, we first split the training set to construct a virtual training set for virtual training, and a virtual testing set for virtual testing. The data samples in the virtual testing set contain novel compositions w.r.t. the virtual training set. For example, if the expressions of data samples in the virtual training set contain both words \"dark\" and \"coffee\" but do not contain their composition (i.e., \"dark coffee\"), the virtual testing set can include this novel composition correspondingly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 289, + 715 + ], + "type": "text", + "content": "Based on the constructed virtual training set and virtual testing set, we first train the model using the virtual training set, and then evaluate the trained model on the virtual testing set. During virtual training, the model may learn the compositions of individual concepts as a whole without truly understanding the semantics and visual representa" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": "tions of individual concepts, which though can still improve model training performance. For example, if there are many training samples containing the composition of \"yellow banana\" in the virtual training set, the model can superficially correlate \"banana\" with \"yellow\" and learn this composition as a whole, since using such spurious correlations can facilitate the model learning [1, 10, 39]. However, learning the compositions as a whole over the virtual training set may not improve model performance much on the virtual testing set in virtual testing, since the virtual testing set contains novel compositions w.r.t. the virtual training set. Thus to achieve good testing performance on such a virtual testing set, the model needs to effectively capture semantics and visual representations of individual concepts during virtual training. In this way, the model testing performance on the virtual testing set serves as a generalization feedback to the model virtual training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 277, + 547, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 385 + ], + "type": "text", + "content": "Thus after the virtual training and virtual testing, we can further update the model to obtain better testing performance on the virtual testing set (i.e., meta update), so as to drive the model training on the virtual training set towards the direction of learning to capture semantics and visual representations of individual concepts, i.e., learning to learn. In this manner, our framework is able to optimize the model for robust generalization performance, even tackling the challenging testing samples with novel compositions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 387, + 548, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 548, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 548, + 602 + ], + "type": "text", + "content": "Moreover, given that expressions can often be hierarchically decomposed, there can exist various levels of novel compositions. Specifically, to identify meaningful compositions in an expression, we can parse an expression into a tree structure based on the constituency parsing tool [9] as shown in Fig.1b. In such a parsing tree, under the same parent node, each pair of child nodes (e.g., \"white\" and \"bird\") are closely semantically related, and thus can form a meaningful composition. Since each child node can be a word or a phrase as in Fig. 1b, there can naturally exist the following three levels of novel compositions: word-word level (e.g., \"white\" and \"bird\"), word-phrase level (e.g., \"standing\" and \"in water\") and phrase-phrase level (e.g., \"white bird\" and \"standing in water\"), which correspond to different levels of comprehension complexity. To better handle such a range of novel compositions, we construct multiple virtual testing sets in our framework, where each virtual testing set is constructed to handle one level of novel compositions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 604, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 547, + 664 + ], + "type": "text", + "content": "Our framework only changes the model training scheme without the need to change the model structure. Thus our framework is general, and can be conveniently applied on various RES models. We test our framework on multiple models, and obtain consistent performance improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "The contributions of our work are threefold: 1) We propose a novel framework (MCRES) to effectively improve generalization performance of RES models, especially when handing novel compositions of learned con" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19479" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": "cepts. 2) Via constructing a virtual training set and multiple virtual testing sets w.r.t. various levels of novel compositions, our framework can train the model to well handle various levels of novel compositions. 3) When applied on various models on three RES benchmarks [28, 41], our framework achieves consistent performance improvement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 139, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 139, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 139, + 168 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 177, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 297 + ], + "type": "text", + "content": "Referring Expression Segmentation (RES). RES [12] aims to segment a target object from an image based on an expression. Early works [20, 22, 28, 41] employed convolutional and recurrent networks to extract visual and linguistic features respectively, and then fused the extracted features to predict the segmentation mask. Recently, a series of works [5, 18, 37] employed vision and language transformers to boost performance. Besides, with the development of large-scale pretrained models, Wang et al. [34] proposed to leverage CLIP [30] to improve cross-modal matching." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 288, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 441 + ], + "type": "text", + "content": "Some methods have been explored to help model better understand learned concepts in RES. Yu et al. [40] proposed a modular network that uses different modules to process different types of information in the given expression. Yang et al. [36] designed a reasoning module to help align the language concepts with visual regions. Different from all the above-mentioned works, we propose an MCRES framework to drive RES models to better capture semantics and visual representations of individual concepts. Such a framework only changes the model training scheme, and thus can be flexibly applied on various models to improve their generalization performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 441, + 288, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 585 + ], + "type": "text", + "content": "Meta Learning. Meta learning, i.e., the paradigm of learning to learn, has emerged to mainly tackle the few-shot learning problem [7, 29, 31, 32]. MAML [7] and its following works [29, 31] aim to learn a good initialization of network parameters, to achieve fast test-time update to adapt to new few-shot learning tasks. More recently, meta learning has also been explored in other areas [8, 11, 14, 19, 35] to enhance model generalization performance without the need of test-time update. Inspired by these works, we leverage a meta learning-based framework to improve generalization performance of RES models especially when handling novel compositions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 103, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 103, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 103, + 609 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Existing RES models may fail to generalize to data samples containing novel compositions of learned concepts. To handle this issue, we aim to encourage the model to better capture the semantics of individual concepts as well as recognize their visual correspondences during training, which however is a non-trivial problem. In this paper, from the perspective of meta learning, we introduce MCRES framework optimizing RES models via a meta optimization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 520, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 520, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 520, + 84 + ], + "type": "text", + "content": "scheme to improve their generalization performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 2, in our framework, we first split the original training set " + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{train})" + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "text", + "content": " to build a virtual training set " + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{v\\_ tr})" + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "text", + "content": " for virtual training, and a group of " + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "text", + "content": " virtual testing sets " + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)" + }, + { + "bbox": [ + 304, + 85, + 545, + 337 + ], + "type": "text", + "content": " for virtual testing. Each virtual testing set consists of data samples containing one level of novel compositions w.r.t. the virtual training set. We first use the virtual training set to train the model (virtual training), and then perform model testing on the virtual testing sets (virtual testing). Since the virtual testing sets contain novel compositions of individual concepts from the virtual training set, if the model can still achieve robust testing performance on the virtual testing sets after training on the virtual training set, we posit that the trained model has learned to capture more semantics and visual representations of individual concepts during the virtual training. Guided by this principle, we can optimize the model virtual testing performance, to guide the model virtual training process towards learning more semantics and representations of individual concepts. Below, we first introduce the pipeline of our framework, and then detail the virtual training set and virtual testing sets construction process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 346, + 384, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 384, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 384, + 358 + ], + "type": "text", + "content": "3.1. Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": "We first train the RES model using the virtual training set, i.e., virtual training. Specifically, we denote the model parameters as " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": ", the loss function (e.g., cross-entropy loss) for training the RES model as " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": ". Thus we can calculate the model virtual training loss " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{v\\_ tr})" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " over the virtual training set " + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{v\\_ tr})" + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 443, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 443, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 377, + 443, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v \\_ t r} (\\theta) = \\mathcal {L} (\\theta ; \\mathcal {D} _ {v \\_ t r}) \\tag {1}", + "image_path": "36e4d2dc46349288764f7518c67dcd4b3abe2caddfd1d870b0d3c0ce321a6937.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 461, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 461, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 461, + 545, + 484 + ], + "type": "text", + "content": "Based on this loss, we then update the model parameters " + }, + { + "bbox": [ + 305, + 461, + 545, + 484 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 305, + 461, + 545, + 484 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 487, + 545, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 487, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 378, + 487, + 545, + 500 + ], + "type": "interline_equation", + "content": "\\theta^ {\\prime} = \\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v _ {-} t r} (\\theta) \\tag {2}", + "image_path": "6e267ee374f378cfcd380a72025c12f35a6692f6b6562616327065b631c0e38b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "content": " denotes the learning rate. It is worth noting that the model update at this step is virtual, and the updated parameters " + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "content": " only serve as intermediate parameters for the model evaluation at the following virtual testing step." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "type": "text", + "content": "After the virtual training, we evaluate the generalization performance of the trained model on the virtual testing sets " + }, + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\left(\\{\\mathcal{D}_{v\\_te}^k\\}_{k = 1}^K\\right)" + }, + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "type": "text", + "content": ", i.e., virtual testing. Specifically, on each virtual testing set " + }, + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{v\\_te}^k" + }, + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "type": "text", + "content": ", we compute the model loss " + }, + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{v\\_te}^k" + }, + { + "bbox": [ + 305, + 557, + 545, + 606 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 609, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 609, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 375, + 609, + 545, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v. t e} ^ {k} \\left(\\theta^ {\\prime}\\right) = \\mathcal {L} \\left(\\theta^ {\\prime}; \\mathcal {D} _ {v. t e} ^ {k}\\right) \\tag {3}", + "image_path": "767639ba4d321348fc070f25466e6d36a668dfef9da5b9abbca52e8f50fc3e5a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 628, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 653 + ], + "type": "text", + "content": "Such a loss measures how well the model generalizes to the virtual testing set after training on the virtual training set." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Then we perform the meta update step. At this step, we wish to actually update the model parameters " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", so that after the model training on the virtual training set, the trained model can generalize well to the virtual testing sets, i.e., generalizing well to novel compositions. To this end, we" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19480" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 286, + 261 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 286, + 261 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 286, + 261 + ], + "type": "image", + "image_path": "a4b218637b9ae77833b8d3d5ab604e43a0412f521608cc2beeb21c414e829b77.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 262, + 179, + 271 + ], + "lines": [ + { + "bbox": [ + 167, + 262, + 179, + 271 + ], + "spans": [ + { + "bbox": [ + 167, + 262, + 179, + 271 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 322, + 73, + 524, + 261 + ], + "blocks": [ + { + "bbox": [ + 322, + 73, + 524, + 261 + ], + "lines": [ + { + "bbox": [ + 322, + 73, + 524, + 261 + ], + "spans": [ + { + "bbox": [ + 322, + 73, + 524, + 261 + ], + "type": "image", + "image_path": "e801d30b4da5eed6ad8ff7092a0717d0e9fd40ac124d0bfb9c67d3a8aaf55b78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 262, + 428, + 271 + ], + "lines": [ + { + "bbox": [ + 417, + 262, + 428, + 271 + ], + "spans": [ + { + "bbox": [ + 417, + 262, + 428, + 271 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "type": "text", + "content": "Figure 2. Framework overview. Fig. (a) illustrates our framework pipeline. First, we train the model using the virtual training set " + }, + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{v\\_ tr})" + }, + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "type": "text", + "content": " and then obtain the updated model. We then test the model with updated parameters " + }, + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "type": "inline_equation", + "content": "(\\theta^{\\prime})" + }, + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "type": "text", + "content": " on multiple virtual testing sets " + }, + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "type": "inline_equation", + "content": "\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)" + }, + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "type": "text", + "content": ". According to the virtual testing losses, we perform meta update to optimize the model for better generalization capability. Fig. (b) shows that we construct a virtual training set and multiple virtual testing sets to handle various levels of novel compositions. The expressions of data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 338, + 209, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 209, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 209, + 350 + ], + "type": "text", + "content": "formulate the optimization objective as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 287, + 423 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} _ {v. t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v. t e} ^ {k} \\left(\\theta^ {\\prime}\\right) \\tag {4} \\\\ = \\min _ {\\theta} \\mathcal {L} _ {v - t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v - t e} ^ {k} \\left(\\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v - t r} (\\theta)\\right) \\\\ \\end{array}", + "image_path": "fc50dbc57751a050450027709be57163bd7e18e6338e150ebcbcebc5ffe7fc60.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 427, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 487 + ], + "type": "text", + "content": "where the first term indicates the model training performance on the virtual training set, and the second term represents the model testing performance on the virtual testing sets after training on the virtual training set. Based on this objective, we then update the model parameters " + }, + { + "bbox": [ + 46, + 427, + 287, + 487 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 46, + 427, + 287, + 487 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 492, + 287, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 492, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 49, + 492, + 287, + 534 + ], + "type": "interline_equation", + "content": "\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\left(\\mathcal {L} _ {v _ {-} t r} (\\theta) + \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {v _ {-} t e} ^ {k} \\left(\\theta - \\alpha \\nabla_ {\\theta} \\mathcal {L} _ {v _ {-} t r} (\\theta)\\right)\\right) \\tag {5}", + "image_path": "d70e3ee5d044702670d33a1aea5aec2da3ba584b54a92fb8143f8a4dd0e44e3a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 535, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 535, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 535, + 287, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 535, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 535, + 287, + 583 + ], + "type": "text", + "content": " is the learning rate for meta update. Through the above optimization process, the model is pushed to learn more semantics and visual representations of individual concepts during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "During the above process, the model is first trained (updated) on the virtual training set. At this step, the model may learn the compositions of individual concepts as a whole without effectively capturing the semantics and visual representations of individual concepts, which though can still improve its training performance on the virtual training set. However, to achieve good testing performance on the virtual testing data which contains novel compositions of individual concepts from the virtual training data, the model is expected to avoid learning compositions as a" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "type": "text", + "content": "whole and instead capture more semantics and representations of individual concepts. In this way, the second term of Eqn. 5, which includes the second-order gradients of " + }, + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}_{v\\_te}^{k}\\big(\\theta -\\alpha \\nabla_{\\theta}\\mathcal{L}_{v\\_tr}(\\theta)\\big)" + }, + { + "bbox": [ + 304, + 338, + 545, + 409 + ], + "type": "text", + "content": ", can be regarded as a generalization feedback that can guide the model to capture more semantics and representations of individual concepts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 410, + 545, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 446 + ], + "type": "text", + "content": "In our framework, the above three steps (i.e., virtual training, virtual testing and meta update) are performed iteratively until the model training converges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 455, + 411, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 411, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 411, + 466 + ], + "type": "text", + "content": "3.2. Sets Construction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 473, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 557 + ], + "type": "text", + "content": "In our framework, to handle various levels of novel compositions, we split the original training set to construct a virtual training set and various virtual testing sets. Each virtual testing set is expected to contain one level of novel compositions w.r.t. the virtual training set. Below we first introduce the details of these sets, and then discuss the strategy for constructing the virtual testing sets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "text", + "content": "We randomly sample a subset of the training set " + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{train})" + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "text", + "content": " as the virtual training set " + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{v\\_ tr})" + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "text", + "content": ", and the remaining training data will be used as the candidate samples to construct a group of " + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "text", + "content": " virtual testing sets " + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "inline_equation", + "content": "\\left(\\{\\mathcal{D}_{v\\_ te}^{k}\\}_{k=1}^{K}\\right)" + }, + { + "bbox": [ + 304, + 558, + 546, + 700 + ], + "type": "text", + "content": ". Each virtual testing set is constructed to handle one level of novel compositions. Here considering the hierarchical semantic structure of each expression, we target the following levels of novel compositions: word-word level, word-phrase level and phrase-phrase level. A phrase means a group of words that serve as a grammatical unit in the expression (e.g., \"white bird\" in Fig. 1b), which can be conveniently identified using off-the-shelf tools (e.g., AllenNLP [9])." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Specifically, to identify various levels of novel compo" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19481" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "sitions in an expression, we first use the constituency parsing tool [9] to parse an expression into a tree structure as shown in Fig. 1b. In this tree, under the same parent node, each pair of child nodes are closely related, and thus can form a meaningful composition. If a pair of nodes having the same parent node are both words, they form a word-word level composition (e.g., \"white\" and \"bird\"). Similarly, if the nodes with the same parent node are a word and a phrase, they form a word-phrase level composition (e.g., \"standing\" and \"in water\"). In a similar way, we can obtain phrase-phrase level compositions (e.g., \"white bird\" and \"standing in water\"). Thus there can exist three levels of novel compositions, which correspond to different levels of comprehension complexity. Corresponding to these three levels of novel compositions, we will construct a total of three virtual testing sets " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "(K = 3)" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 289, + 434 + ], + "type": "text", + "content": "Virtual testing sets construction. To construct each virtual testing set, we need to select data samples containing the corresponding level of novel compositions w.r.t. the virtual training set, from the candidate samples. A data sample means an expression paired with the corresponding image. Here we design an efficient strategy, which can be leveraged to construct each virtual testing set. Below, we take the process of constructing the virtual testing set for handling word-word level novel compositions as an example, to introduce such a strategy. Specifically, this virtual testing set should include all of those candidate samples that contain word-word level novel compositions w.r.t. the virtual training set. To construct such a virtual testing set, our strategy proceeds as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "content": "(i) To find the data samples, of which the expressions contain word-word level novel compositions w.r.t. the virtual training set, from the candidate samples, we need to first obtain all the word-word level compositions in the virtual training set and in the candidate samples respectively. Thus by parsing each expression into a parsing tree as shown in Fig. 1b, we can obtain all word-word level compositions (e.g., \"white\" and \"bird\", \"in\" and \"water\" in Fig. 1b) in each expression. (ii) To identify word-word level novel compositions, we first select the word-word level compositions that exist in the candidate samples but are unseen in the virtual training set. Then for each selected composition, we further check if its individual words exist in the virtual training set. If so, such a composition will be identified as a word-word level novel composition w.r.t. the virtual training set. For example, if the composition \"white bird\" is unseen in the virtual training set, while its individual words \"white\" and \"bird\" both exist in the virtual training set, this composition is a word-word level novel composition w.r.t. the virtual training set. (iii) Finally, we select the candidate samples, of which the expressions contain the identified word-word level novel compositions w.r.t. the virtual training set, to construct the virtual testing set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": "To efficiently perform the above steps, we can employ parallel matrix operations. Specifically, to record the word-word level compositions in the virtual training set, we build a matrix " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{v,tr}" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " with the shape of " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}| \\times |\\mathcal{V}|" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}|" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " is the size of the vocabulary set " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " of the original training set. In this matrix, each element at the " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "i-th" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " row and the " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "j-th" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " column " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "(i,j \\in \\{1,\\dots,|\\mathcal{V}|\\})" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " is a binary value, and the value 1 indicates that the " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "i-th" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " word and " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "j-th" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " word in the vocabulary set " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " form a word-word level composition in the virtual training set, while 0 means such a word-word level composition does not exist in the virtual training set. Similarly, we can also build a matrix " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{candi}" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " to record the word-word level compositions in the candidate samples. Then to identify word-word level novel compositions w.r.t. the virtual training set " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_{v,tr})" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": " from the candidate samples " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_{candi})" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": ", we can efficiently compute a difference matrix: " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{diff} = \\mathcal{M}_{candi} - \\mathcal{M}_{v,tr}" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{diff}" + }, + { + "bbox": [ + 304, + 72, + 547, + 347 + ], + "type": "text", + "content": ", any element with the value 1 indicates that the corresponding composition exists in the candidate samples, but is unseen in the virtual training set. Then by further checking whether the individual words in such a composition exist in the virtual training set, we can determine if it is a word-word level novel composition w.r.t. the virtual training set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 348, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 547, + 433 + ], + "type": "text", + "content": "Similarly, we can efficiently construct the other two virtual testing sets. Note that to help the model to learn to handle a wide range of possible novel compositions, at the beginning of each training epoch, we randomly sample a subset of the training data to re-construct the virtual training set, and leverage the remaining data to re-construct the virtual testing sets using the above strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 433, + 548, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 548, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 548, + 589 + ], + "type": "text", + "content": "As discussed above, to handle various levels of novel compositions, we construct multiple virtual testing sets. A simpler alternative is to construct only one virtual testing set, which includes all the data samples containing any level of novel compositions w.r.t. the virtual training set. However, compared to this alternative, by constructing multiple virtual testing sets, each level of novel compositions in the corresponding virtual testing set can offer an explicit generalization feedback during model training. Thus the model is explicitly encouraged to well handle each level of novel compositions. Moreover, constructing multiple virtual testing sets can facilitate the use of curriculum learning strategy for model training as discussed in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 598, + 427, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 427, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 427, + 611 + ], + "type": "text", + "content": "3.3. Training and Testing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "To help models generalize to novel compositions in RES, our framework only changes the model training scheme. Thus our framework is general, and can be flexibly applied to train various RES models. During training, at each epoch, we first split the training set to construct a virtual training set and multiple virtual testing sets. Then we iteratively optimize the model over the virtual training set and virtual testing sets via meta optimization. Specifically, for each meta" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19482" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 137, + 287, + 367 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 137 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-arts on three benchmark datasets using the metric of overall IoU. Moreover, we apply our framework on various models [25, 37, 42], and obtain consistent performance improvement. “-” indicates that the corresponding result is not provided in the original paper. U means the UMD split of RefCOCOg dataset, and G means the Google split." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 137, + 287, + 367 + ], + "lines": [ + { + "bbox": [ + 49, + 137, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 49, + 137, + 287, + 367 + ], + "type": "table", + "html": "
MethodRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBval(U)test(U)val(G)
DMN [27]49.7854.8345.1338.8844.2232.29--36.76
RRN [20]55.3357.2653.9339.7542.1536.11--36.45
MAAttNet [40]56.5162.3751.7046.6752.3940.0847.6448.61-
CAC [4]58.9061.7753.81---46.3746.9544.32
CMSA [38]58.3260.6155.0943.7647.6037.89--39.98
STEP [3]60.0463.4657.9748.1952.33404.1--46.40
BRINet [13]60.9862.9959.2148.1752.3242.11--48.04
CMPC [15]61.3664.5359.6449.5653.4443.23--49.05
LSCM [16]61.4764.9959.5549.3453.1243.50--48.05
CMPC+ [23]62.4765.0860.8250.2554.0443.47--49.89
EFN [6]62.7665.6959.6751.5055.2443.01--51.93
BUSNet [36]63.2766.4161.3951.7656.8744.13--50.56
CGAN [24]64.8668.0462.0751.0355.5144.0651.0151.6946.54
LTS [17]65.4367.7663.0854.2158.3248.0254.4054.25-
VLT [5]65.6568.2962.7355.5059.2049.3652.9956.6549.76
ReSTR [18]67.2269.3064.4555.7860.4448.27--54.48
CRIS [34]70.4773.1866.1062.2768.0853.6859.8760.36-
MCN [25]62.4464.2059.7150.6254.9944.6949.2249.40-
MCN + Ours64.5166.4862.8452.2956.7146.7251.4951.63-
SeqTR [42]71.7073.3169.8263.0466.7358.9764.6965.74-
SeqTR + Ours73.2375.0171.9564.7167.8560.8566.7767.48-
LAVT [37]72.7375.8268.7962.1468.3855.1061.2462.0960.50
LAVT + Ours74.9276.9870.8464.3269.6856.6463.5164.9061.63
", + "image_path": "45f90d9dadc0ff43646abb0cbad84a9083ffd30c6dad386f2d193981c145471e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 377, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 544 + ], + "type": "text", + "content": "optimization iteration, we use two batches of data samples: one batch of data for virtual training and the other batch of data for virtual testing. Thus we need to ensure that the batch of data for virtual testing consists of the data samples containing novel compositions w.r.t. the batch of data for virtual training. To this end, for each iteration, we randomly sample some data from each virtual testing set, to form the batch of data for virtual testing. Then we select the virtual training set samples that contain all the individual components of the novel compositions in the batch of data for virtual testing, to form the batch of data for virtual training. Note that the above data preparation procedures can be done before the model training. For model testing, we test the trained model in the conventional way." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Besides, given that the various levels of novel compositions correspond to different levels of comprehension complexity, we adopt a curriculum learning strategy [2, 33] for model training, so that the model can progressively learn to handle various levels of novel compositions, i.e., from lower level (word-word) to middle level (word-phrase) to higher level (phrase-phrase), and thus can learn all these levels of novel compositions better. Specifically, in the first " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " of training epochs, we only use the virtual testing set for handling word-word level novel compositions for meta optimization. Then in the middle " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "1/3 - 2/3" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " of training epochs, we add the virtual testing set for handling word-phrase level novel compositions. Finally, in the remaining training epochs, we use all the three virtual testing sets." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 308, + 135, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 135 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 135 + ], + "type": "text", + "content": "Table 2. We evaluate our framework on different testing subsets w.r.t. novel compositions to validate its effectiveness on optimizing the model to better generalize to novel compositions. We use the metric of overall IoU here. The performance gain of our framework compared to the corresponding baseline model is shown in parentheses." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 135, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 308, + 135, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 308, + 135, + 545, + 191 + ], + "type": "table", + "html": "
MethodRefCOCO-valRefCOCO-testARefCOCO-testB
NovelNon-novelNovelNon-novelNovelNon-novel
MCN [23]53.1767.4155.4370.0850.6466.47
MCN + Ours57.38 (↑4.21)68.09 (↑0.68)59.66 (↑4.23)70.67 (↑0.59)55.87 (↑5.23)67.04 (↑0.57)
SeqTR [40]64.2478.5965.7179.4660.1477.21
SeqTR + Ours67.31 (↑3.07)79.06 (↑0.47)69.48 (↑3.77)79.84 (↑0.38)64.89 (↑4.75)77.68 (↑0.47)
LAVT [35]63.5278.5567.1780.7158.4976.23
LAVT + Ours67.42 (↑3.90)79.29 (↑0.74)70.05 (↑2.88)81.37 (↑0.66)62.63 (↑4.14)76.92 (↑0.69)
", + "image_path": "07ddd9d10ba740c332bba98b1ae7f2d039680e657b33e8978693c8363ce33795.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 207, + 388, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 207, + 388, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 388, + 220 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 455 + ], + "type": "text", + "content": "We evaluate our method on three commonly used RES benchmarks: RefCOCO [41], RefCOCO+ [41] and RefCOCOg [26, 28]. Images in these three benchmarks are all from the COCO dataset [21]. RefCOCO [41] contains 19994 images, 50000 annotated objects with 142209 referring expressions. RefCOCO+ [41] consists of 19992 images with 49856 annotated objects, and 141564 expressions. Different from RefCOCO, words describing absolute spatial locations (e.g., left, front) are not allowed to be used in the expressions in RefCOCO+. In RefCOCO and RefCOCO+ datasets, following the original split in [41], the visual entities to be segmented in testA subset are people, while the ones in testB subset are objects (i.e., not people). RefCOCOg [26, 28] includes 26711 images, 54822 annotated objects and 104560 expressions. Compared with RefCOCO and RefCOCO+, the expressions in RefCOCOg are more complex, which have an average length of 8.4 words. There exist two different partitions for RefCOCOg dataset: UMD split [28] and Google split [26]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 456, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 552 + ], + "type": "text", + "content": "Evaluation metrics. Following [3,37,38], we report our results using two kinds of metrics: overall IoU (oIoU) and Precision@X (P@X). The overall IoU measures the ratio of total intersection regions over total union regions of predicted masks and ground truths of all testing samples. Precision@X calculates the percentage of testing samples, of which the model prediction has an IoU score higher than the threshold value X, and " + }, + { + "bbox": [ + 304, + 456, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{X} \\in \\{0.5, 0.6, 0.7, 0.8, 0.9\\}" + }, + { + "bbox": [ + 304, + 456, + 545, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "type": "text", + "content": "Implementation details. We conduct our experiments on 8 Tesla V100 GPUs. We applied our framework on various RES models [25,37,42]. On each dataset, we randomly sample " + }, + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "type": "text", + "content": " of the training data as the virtual training set and use the remaining training data to construct virtual testing sets at each training epoch. The learning rate " + }, + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "type": "text", + "content": " for virtual training is 5e-5, and the learning rate " + }, + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 304, + 552, + 545, + 648 + ], + "type": "text", + "content": " for meta update is 2e-5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 658, + 430, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 430, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 430, + 671 + ], + "type": "text", + "content": "4.1. Experimental Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "As shown in Table 1, our framework achieves state-of-the-art performance across all three datasets, demonstrating the superiority of our framework. Moreover, we applied our" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19483" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 92, + 286, + 145 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 286, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 286, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 286, + 92 + ], + "type": "text", + "content": "Table 3. We test several variants to investigate the impact of each level of novel compositions on model performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 92, + 286, + 145 + ], + "lines": [ + { + "bbox": [ + 49, + 92, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 49, + 92, + 286, + 145 + ], + "type": "table", + "html": "
MethodIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline (LAVT)72.7384.4681.2475.2864.7134.30
w/o word-word novel compositions73.4384.7181.6976.1265.2334.47
w/o word-phrase novel compositions73.6884.8281.7676.2865.3734.56
w/o phrase-phrase novel compositions73.5984.7381.7376.1665.3534.45
Ours74.9286.2383.4577.2566.5635.61
", + "image_path": "ee113a85514dfbeddeed6f1aa28d2a7479372ad5ed839ff529b843c72e158f2e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 191, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "type": "text", + "content": "Table 4. We evaluate a variant to test the effectiveness of meta optimization scheme in our framework. For this variant, its optimization objective is to minimize " + }, + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{v\\_ tr}(\\theta) + \\sum_{k=1}^{K} \\mathcal{L}_{v\\_ te}^{k}(\\theta)" + }, + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "type": "text", + "content": " (i.e., replacing " + }, + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 148, + 286, + 190 + ], + "type": "text", + "content": " in Eqn. 4)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 191, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 49, + 191, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 49, + 191, + 286, + 223 + ], + "type": "table", + "html": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Training w/o meta72.7684.5081.2775.3364.7234.31
Ours74.9286.2383.4577.2566.5635.61
", + "image_path": "92f8d15d714845d4792841d8dbdf81136cfcce7bbf15f23080508ba1aba4cdcf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": "framework on two transformer-based SOTA models [37,42] and a CNN-based model [25]. Our framework brings consistent performance improvement on all these models and datasets. This shows that our framework can serve as a general approach to enhance model performance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 289, + 286, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 286, + 612 + ], + "type": "text", + "content": "To further validate the effectiveness of our framework for handling novel compositions, we perform the following analysis. Specifically, we split each testing set of RefCOCO dataset to construct two subsets. One subset (Non-novel) includes the data samples, in which all the contained compositions are seen in the RefCOCO training set. While another subset (Novel) includes the data samples that contain any level of novel compositions w.r.t. the training set. As shown in Table 2, on each testing set, we can see a clear performance gap between the two subsets for each of the baseline models [25, 37, 42], showing that existing models struggle with handling novel compositions. Then by applying our framework on each baseline model, we obtain significant performance improvement on the subset of Novel. This validates the general effectiveness of our framework to optimize the model to well generalize to novel compositions. Moreover, our framework also slightly improves the model performance on the subset of Non-novel. This can be attributed to that by training the model to better capture semantics and visual representations of individual concepts, our framework can help the model to better understand the given expression and find its visual correspondence, which thus can generally improve model performance. We also show some qualitative results in Fig. 3. As shown, when handling the testing samples containing novel compositions, our framework achieves better performance than the baseline model [37]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 621, + 146, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 146, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 146, + 633 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 641, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 286, + 664 + ], + "type": "text", + "content": "Following [3, 37, 38], we conduct ablation experiments to evaluate our framework on RefCOCO validation set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 666, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 286, + 712 + ], + "type": "text", + "content": "Impact of various levels of novel compositions. To investigate the impact of each level of novel compositions on model performance, we test multiple variants. One variant (w/o word-word novel compositions) ignores the word-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 538, + 177 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 538, + 177 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 538, + 177 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 538, + 177 + ], + "type": "image", + "image_path": "1558cade3b9dbb7c371f8acd70512fe2ac0e69155f1bf1e60dbf25e697642e99.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 243 + ], + "type": "text", + "content": "Figure 3. Qualitative results of our method and the baseline model [37]. The above testing examples contain a word-word novel composition, a word-phrase novel composition, and a phrase-phrase novel composition respectively. As shown, by applying our framework on the baseline model, our method performs better when handling novel compositions of the learned concepts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 252, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 372 + ], + "type": "text", + "content": "word level novel compositions. This means that in this variant, we omit the virtual testing set for handling word-word level novel compositions during meta optimization. Similarly, the other variants (w/o word-phrase novel compositions and w/o phrase-phrase novel compositions) ignore the corresponding level of novel compositions respectively. As shown in Table 3, ignoring any level of novel compositions leads to performance drop compared to our framework, demonstrating that each level of novel compositions can affect model generalization performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 374, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 494 + ], + "type": "text", + "content": "Impact of meta optimization. To investigate the effectiveness of the meta optimization scheme in our framework, we evaluate a variant (training w/o meta). In this variant, we train the model in the conventional manner on the constructed virtual training set and virtual testing sets, i.e., without meta optimization. Note that for this variant, we still construct the virtual training set and virtual testing sets in the same way as in our framework. As shown in Table 4, our framework outperforms this variant obviously, demonstrating the effectiveness of our meta optimization scheme." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 496, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 699 + ], + "type": "text", + "content": "Impact of multiple virtual testing sets and curriculum learning. In our framework, we construct multiple virtual testing sets to handle various levels of novel compositions. Moreover, since the various levels of novel compositions correspond to different levels of comprehension complexity, we adopt a curriculum learning strategy to facilitate model training. To investigate the effectiveness of such design, we evaluate two variants. One variant (one virtual testing set) constructs only one virtual testing set to handle all levels of novel compositions. Such a virtual testing set includes all the data samples containing any level of novel compositions w.r.t. the virtual training set. Another variant (multiple virtual testing sets w/o curriculum learning) constructs multiple virtual testing sets as in our framework, but does not adopt the curriculum learning strategy. We compare these two variants to our original framework (multiple virtual testing sets w/ curriculum learning)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "content": "As shown in Table 5, compared to the variant construct" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19484" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "ing one virtual testing set, the other variant obtains better performance, showing the superiority of using multiple virtual testing sets to handle various levels of novel compositions. Furthermore, by employing the curriculum learning strategy, our framework performs better than the variant constructing multiple virtual testing sets without curriculum learning, demonstrating the effectiveness of our curriculum learning strategy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": "Impact of the size of virtual training set. In our framework, we randomly sample " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": " of the training data as the virtual training set, and use the remaining " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": " of the training data to construct virtual testing sets " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "(60\\%:40\\%)" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": ". Here we test two variants. One variant " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "(50\\%:50\\%)" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": " uses " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": " of the training data to construct the virtual training set, and the remaining " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": " to construct virtual testing sets. While another variant " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "(70\\%:30\\%)" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": " uses " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": " of the training data for the virtual training set, and the remaining " + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": " part for virtual testing sets. As shown in Table 6, our method and these two variants all perform better than the baseline model (i.e., LAVT [37]), showing the robustness of our framework in terms of the varying size of virtual training set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 483 + ], + "type": "text", + "content": "Impact of virtual testing sets construction strategy. In our framework, after sampling a subset of the training set as the virtual training set, we use the remaining training data to construct multiple virtual testing sets. Each virtual testing set consists of the data samples containing one level of novel compositions w.r.t. the virtual training set. To explore the efficacy of such a strategy, we evaluate a variant (random virtual testing sets), in which we totally randomly select training data to construct each virtual testing set. As shown in Table 7, our framework obviously outperforms this variant. This shows that our virtual testing sets construct strategy can effectively help our framework to improve model generalization performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 289, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 289, + 568 + ], + "type": "text", + "content": "Training time. As shown in Table 8, we test the training time of our framework that trains the baseline network [37] with meta optimization, and compare it to the training time of the baseline that trains the same network in the conventional manner without meta optimization, on RefCOCO dataset. Though our framework performs better, it brings only relatively little increase (18.18%) in training time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Impact of additional gradient updates. As discussed above, compared to the baseline model, our framework trains the model for longer time and involves more gradient updates. To explore whether the performance improvement of our framework comes from the additional gradient updates, we test a variant (baseline w/ additional gradient updates) in which we train the baseline model (following the original training strategy) for as many iterations as in our framework. As shown in Table 9, the performance of this variant is very close to the baseline model [37], and is obviously worse than our framework. This might be because that the originally trained baseline models have al" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 93, + 545, + 125 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 92 + ], + "type": "text", + "content": "Table 5. We evaluate two variants to test the impact of using multiple virtual testing sets and the curriculum learning strategy." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 93, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 307, + 93, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 307, + 93, + 545, + 125 + ], + "type": "table", + "html": "
MethodoloUP@0.5P@0.6P@0.7P@0.8P@0.9
Ours (one virtual testing set)73.6484.8282.0876.2465.4734.64
Ours (multiple virtual testing sets w/o curriculum learning)74.0285.3482.6376.8766.0934.95
Ours (multiple virtual testing sets w/ curriculum learning)74.9286.2383.4577.2566.5635.61
", + "image_path": "dfe5b51fa0acfa0ad4451844afaf75d27cf1ffe83c9273b00de71669e55fa044.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 164, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 305, + 130, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 305, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 545, + 163 + ], + "type": "text", + "content": "Table 6. We test different variants that utilize different proportions of the training data to construct the virtual training set and virtual testing sets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 164, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 218 + ], + "type": "table", + "html": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline (LAVT)72.7384.4681.2475.2864.7134.30
Ours (50%:50%)74.7886.1183.3277.0866.3835.54
Ours (60%:40%)74.9286.2383.4577.2566.5635.61
Ours (70%:30%)74.7486.0483.2677.0166.2935.52
", + "image_path": "d2d0dff27a28ca93720674d32a282f066866b6a2f7c0eaf8f93c215f94a82e85.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 247, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 306, + 224, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 306, + 224, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 545, + 247 + ], + "type": "text", + "content": "Table 7. We evaluate a variant to investigate the efficacy of virtual testing sets construction strategy in our framework." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 247, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 308, + 247, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 247, + 545, + 277 + ], + "type": "table", + "html": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Random virtual testing sets72.8184.5681.2875.3464.7434.32
Ours74.9286.2383.4577.2566.5635.61
", + "image_path": "709c18e54624b6f344074a76e3e04e44185d3a18abcb992f168aa30deac65cc1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 307, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 306, + 285, + 545, + 306 + ], + "lines": [ + { + "bbox": [ + 306, + 285, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 545, + 306 + ], + "type": "text", + "content": "Table 8. Comparison of the training time. Note that our method achieves much better performance than the baseline." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 307, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 308, + 307, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 545, + 334 + ], + "type": "table", + "html": "
MethodTraining timeoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline33 hours72.7384.4681.2475.2864.7134.30
Ours39 hours74.9286.2383.4577.2566.5635.61
", + "image_path": "6b1ff69a316a900d7420d8bf8857a781ec0b0f4759daeee3999bfa4ea4383872.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 363, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 306, + 340, + 545, + 362 + ], + "lines": [ + { + "bbox": [ + 306, + 340, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 545, + 362 + ], + "type": "text", + "content": "Table 9. We test a variant to investigate the impact of additional gradient updates on model performance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 363, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 308, + 363, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 363, + 545, + 399 + ], + "type": "table", + "html": "
MethodoIoUP@0.5P@0.6P@0.7P@0.8P@0.9
Baseline72.7384.4681.2475.2864.7134.30
Baseline w/ additional gradient updates72.7484.4381.2575.2564.7334.28
Baseline w/ ours74.9286.2383.4577.2566.5635.61
", + "image_path": "0540f0ca43e29bb9812010816ebe714536f83578059c4ac55791f3e3d13c1f9f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 416, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 464 + ], + "type": "text", + "content": "ready reached convergence under the original training strategy, and thus additional gradient updates would not bring obvious benefits. Such results further validate the effectiveness of our framework." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 484, + 379, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 484, + 379, + 497 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 379, + 497 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 506, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 651 + ], + "type": "text", + "content": "In this work, we proposed a meta learning-based framework (MCRES) to improve the generalization performance of RES models, especially when handling novel compositions of learned concepts. By constructing a virtual training set and multiple virtual testing sets w.r.t. various levels of novel compositions and then optimizing the model via meta optimization, our framework can effectively improve model generalization performance. Extensive experiments show that our framework achieves superior performance on widely used benchmarks. Moreover, our framework is flexible, and can be seamlessly applied on various models with different architectures to enhance their performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by MOE AcRF Tier 2 (Proposal ID: T2EP20222-0035), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2020-065), and SUTD SKI Project (SKI 2021_02_06)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19485" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "text", + "content": "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 169, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 169, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 169, + 288, + 224 + ], + "type": "text", + "content": "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 267 + ], + "type": "text", + "content": "[4] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "type": "text", + "content": "[5] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "type": "text", + "content": "[6] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "text", + "content": "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 480 + ], + "type": "text", + "content": "[8] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Era: Expert retrieval and assembly for early action prediction. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 670-688. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 534 + ], + "type": "text", + "content": "[9] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. Allennlp: A deep semantic natural language processing platform. arXiv preprint arXiv:1803.07640, 2018. 1, 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "text", + "content": "[10] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "text", + "content": "[11] Jianzhu Guo, Xiangyu Zhu, Chenxu Zhao, Dong Cao, Zhen Lei, and Stan Z Li. Learning meta face recognition in unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163-6172, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[12] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[13] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4424-4433, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "type": "text", + "content": "[14] Chao Huang, Zhangjie Cao, Yunbo Wang, Jianmin Wang, and Mingsheng Long. Metasets: Meta-learning on point sets for generalizable representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8863-8872, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "type": "text", + "content": "[15] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10488-10497, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[16] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In European Conference on Computer Vision, pages 59-75. Springer, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "text", + "content": "[17] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[18] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 1, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[20] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2018. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "type": "text", + "content": "[22] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1271-1280, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[23] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[24] Gen Luo, Yiyi Zhou, Rongrong Ji, Xiaoshuai Sun, Jinsong Su, Chia-Wen Lin, and Qi Tian. Cascade grouped attention network for referring expression segmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1274-1282, 2020. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[25] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collabora" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19486" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 116 + ], + "type": "text", + "content": "tive network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[26] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 226 + ], + "type": "text", + "content": "[27] Edgar Margffoy-Tuay, Juan C Pérez, Emilio Botero, and Pablo Arbeláez. Dynamic multimodal instance segmentation guided by natural language queries. In Proceedings of the European Conference on Computer Vision (ECCV), pages 630-645, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[28] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In European Conference on Computer Vision, pages 792-807. Springer, 2016. 1, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[29] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 370 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 413 + ], + "type": "text", + "content": "[31] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "text", + "content": "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 481 + ], + "type": "text", + "content": "[33] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "text", + "content": "[34] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11686-11695, 2022. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 592 + ], + "type": "text", + "content": "[35] Li Xu, Haoxuan Qu, Jason Kuen, Juixiang Gu, and Jun Liu. Meta spatio-temporal debiasing for video scene graph generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 374-390. Springer, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "type": "text", + "content": "[36] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[37] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 340 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[38] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "type": "text", + "content": "[39] Nanyang Ye, Jingxuan Tang, Huayu Deng, Xiao-Yun Zhou, Qianxiao Li, Zhenguo Li, Guang-Zhong Yang, and Zhanxing Zhu. Adversarial invariant learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12441-12449. IEEE, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "text", + "content": "[40] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 240, + 546, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 284 + ], + "type": "text", + "content": "[41] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 285, + 546, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 546, + 340 + ], + "type": "text", + "content": "[42] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision. Springer, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19487" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_content_list.json b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7bfc925c067617e1e57b3a78b43a1fd0d371fecd --- /dev/null +++ b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_content_list.json @@ -0,0 +1,1353 @@ +[ + { + "type": "text", + "text": "Meta Omnium: A Benchmark for General-Purpose Learning-to-Learn", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ondrej Bohdal $^{1,\\ast}$ , Yinbing Tian $^{2,\\ast}$ , Yongshuo Zong $^{1}$ , Ruchika Chavhan $^{1}$ , Da Li $^{3}$ , Henry Gouk $^{1}$ , Li Guo $^{2}$ , Timothy Hospedales $^{1,3}$", + "bbox": [ + 197, + 179, + 771, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Edinburgh 2Beijing University of Posts and Telecommunications 3Samsung AI Center, Cambridge *Joint-First Authors", + "bbox": [ + 205, + 220, + 764, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta-learning and other approaches to few-shot learning are widely studied for image recognition, and are increasingly applied to other vision tasks such as pose estimation and dense prediction. This naturally raises the question of whether there is any few-shot meta-learning algorithm capable of generalizing across these diverse task types? To support the community in answering this question, we introduce Meta Omnium, a dataset-of-datasets spanning multiple vision tasks including recognition, keypoint localization, semantic segmentation and regression. We experiment with popular few-shot meta-learning baselines and analyze their ability to generalize across tasks and to transfer knowledge between them. Meta Omnium enables meta-learning researchers to evaluate model generalization to a much wider array of tasks than previously possible, and provides a single framework for evaluating meta-learners across a wide suite of vision applications in a consistent manner. Code and dataset are available at https://github.com/edi-meta-learning/meta-omnium.", + "bbox": [ + 75, + 325, + 473, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 646, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta-learning is a long-standing research area that aims to replicate the human ability to learn from a few examples by learning-to-learn from a large number of learning problems [61]. This area has become increasingly important recently, as a paradigm with the potential to break the data bottleneck of traditional supervised learning [26, 70]. While the largest body of work is applied to image recognition, few-shot learning algorithms have now been studied in most corners of computer vision, from semantic segmentation [37] to pose estimation [49] and beyond. Nevertheless, most of these applications of few-shot learning are advancing independently, with increasingly divergent application-specific methods and benchmarks. This makes it hard to evaluate whether few-shot meta-learners can solve diverse vision tasks. Importantly it also discourages the de", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1ca96f2b4f105019e810b99833709cf85445be2e6e6c46008e776bb05d4d7f1.jpg", + "image_caption": [ + "Figure 1. Illustration of the diverse visual domains and task types in Meta Omnium. Meta-learners are required to generalize across multiple task types, multiple datasets, and held-out datasets." + ], + "image_footnote": [], + "bbox": [ + 501, + 292, + 893, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "velopment of meta-learners with the ability to learn-to-learn across tasks, transferring knowledge from, e.g., keypoint localization to segmentation – a capability that would be highly valuable for vision systems if achieved.", + "bbox": [ + 496, + 703, + 890, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The overall trend in computer vision [20, 52] and AI [5, 55] more generally is towards more general-purpose models and algorithms that support many tasks and ideally leverage synergies across them. However, it has not yet been possible to explore this trend in meta-learning, due to the lack of few-shot benchmarks spanning multiple tasks. State-of-the-art benchmarks [63, 65] for visual few-shot learning are restricted to image recognition across a handful of visual domains. There is no few-shot benchmark", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7693", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/0b0888bcc8a68974cc66b2d78f6598302bf2a6ae8d0596d58222b9043d42025e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNum TasksNum DomainsNum ImagesCategoriesSizeLightweightMulti-TaskMulti-Domain
Omniglot [30]1132K1623148MB✓✗✗
miniImageNet [66]1160K1001GB✓✗✗
Meta-Dataset [63]17~1053M43~1500210GB✗✗✓
VTAB [80]13~192.2M2~397100GB✗✗✓
FSS1000 [37]11100001000670MB✓✗✗
Meta-Album [65]110~401.5M19~70615GB✓✗✓
Meta Omnium421160K2~7063.1GB✓✓✓
", + "bbox": [ + 81, + 87, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Feature comparison between Meta Omnium and other few-shot meta-learning benchmarks. Meta Omnium uniquely combines a rich set of tasks and visual domains with a lightweight size for accessible use.", + "bbox": [ + 75, + 207, + 892, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that poses the more substantial challenge [57, 77] of generalizing across different tasks. We remark that the term task is used differently in few-shot meta-learning literature [16, 26, 70] (to mean different image recognition problems, such as cat vs dog or car vs bus) and the multi-task literature [20, 57, 74, 77] (to mean different kinds of image understanding problems, such as classification vs segmentation). In this paper, we will use the term task in the multi-task literature sense, and the term episode to refer to tasks in the meta-learning literature sense, corresponding to a support and query set.", + "bbox": [ + 75, + 262, + 470, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Meta Omnium, a dataset-of-datasets spanning multiple vision tasks including recognition, semantic segmentation, keypoint localization/pose estimation, and regression as illustrated in Figure 1. Specifically, Meta Omnium provides the following important contributions: (1) Existing benchmarks only test the ability of meta-learners to learn-to-learn within tasks such as classification [63, 65], or dense prediction [37]. Meta Omnium uniquely tests the ability of meta-learners to learn across multiple task types. (2) Meta Omnium covers multiple visual domains (from natural to medical and industrial images). (3) Meta Omnium provides the ability to thoroughly evaluate both in-distribution and out-of-distribution generalisation. (4) Meta Omnium has a clear hyper-parameter tuning (HPO) and model selection protocol, to facilitate future fair comparison across current and future meta-learning algorithms. (5), Unlike popular predecessors, [63], and despite the diversity of tasks, Meta Omnium has been carefully designed to be of moderate computational cost, making it accessible for research in modestly-resourced universities as well as large institutions. Table 1 compares Meta Omnium to other relevant meta-learning datasets.", + "bbox": [ + 75, + 430, + 468, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We expect Meta Omnium to advance the field by encouraging the development of meta-learning algorithms capable of knowledge transfer across different tasks – as well as across learning episodes within individual tasks as is popularly studied today [16, 70]. In this regard, it provides the next step of the level of a currently topical challenge of dealing with heterogeneity in meta-learning [1, 35, 63, 67]. While existing benchmarks have tested multi-domain heterogeneity (e.g., recognition of written characters and plants", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "within a single network) [63, 65] and shown it to be challenging, Meta Omnium tests multi-task learning (e.g., character recognition vs plant segmentation). This is substantially more ambitious when considered from the perspective of common representation learning. For example, a representation tuned for recognition might benefit from rotation invariance, while one tuned for segmentation might benefit from rotation equivariance [11, 15, 71]. Thus, in contrast to conventional within-task meta-learning benchmarks that have been criticized as relying more on common representation learning than learning-to-learn [53, 62], Meta Omnium better tests the ability of learning-to-learn since the constituent tasks require more diverse representations.", + "bbox": [ + 496, + 262, + 893, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 473, + 640, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Meta-learning Benchmarks", + "text_level": 1, + "bbox": [ + 500, + 498, + 746, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The classic datasets in few-shot meta-learning for computer vision are Omniglot [30] and miniImageNet [66]. Later work criticized these for having insufficient task (episode) diversity and tieredImageNet [56] used the class hierarchy of ImageNet to enforce more diversity between meta-train and meta-test episodes. The main contemporary benchmarks are CD-FSL [22], which challenges few-shot learners to generalize to new visual domains; and Meta-Dataset [63] and Meta-Album [65], which go further in requiring few-shot learners to learn from a mixture of visual domains. Such multi-domain heterogeneous meta-learning turns out to be challenging. A related benchmark to Meta-Dataset is VTAB [80], which similarly provides multiple domains for evaluating data-efficient visual recognition, but their focus is on evaluating representation transfer from large-scale pre-training rather than learning-to-learn and meta-learning. VTAB+MD [13] compare representation transfer and meta-learning approaches on the Meta-Dataset tasks. However, none of these benchmarks address multi-task meta-learning as considered here (Figure 1).", + "bbox": [ + 496, + 522, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Outside of recognition, task-specific few-shot benchmarks have been proposed in vision problems of semantic segmentation [37], regression [19], pose/keypoint estimation [73], etc. These are mostly slightly behind the complexity of the recognition benchmarks with regards to being", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7694", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "single-domain, with the exception of [73]. With regard to multi-task meta-learning as considered here, the only existing benchmark is meta-world [78], which is specific to robotics and reinforcement learning rather than vision.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also mention taskonomy [79] as a popular dataset that has been used for multi-task learning. However, it is not widely used for few-shot meta-learning. This is because, although taskonomy has many tasks, unlike the main meta-learning benchmarks [37, 63], there are not enough visual concepts within each task to provide a large number of concepts for meta-training and a disjoint set of concepts to evaluate few-shot learning for meta-validation and meta-testing.", + "bbox": [ + 75, + 152, + 470, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Heterogeneity in Meta-Learning", + "text_level": 1, + "bbox": [ + 76, + 284, + 362, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are several sophisticated methods in the literature that highlighted the challenge of the task to address heterogeneity in meta-learning. These have gone under various names such as multi-modal meta-learning [1, 38, 67] – in the sense of multi-modal probability distribution (over tasks/episodes). However, with the exception of [38], these have mostly not been shown to scale to the main multimodal benchmarks such as Meta-Dataset [63]. A more common approach to achieving high performance across multiple heterogeneous domains such as those in Meta-Dataset is to train an ensemble of feature extractors across available training domains, and fuse them during metatesting [14, 36, 36]. However, this obviously incurs a substantial additional cost of maintaining a model ensemble. In our evaluation, we focus on the simpler meta-learners that have been shown to work in challenging multi-domain learning scenarios [63, 65], while leaving sophisticated algorithmic and ensemble-based approaches for future researchers to evaluate on the benchmark.", + "bbox": [ + 75, + 308, + 470, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Meta Omnium Benchmark and Datasets", + "text_level": 1, + "bbox": [ + 76, + 611, + 439, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation and Guiding Principles", + "text_level": 1, + "bbox": [ + 76, + 637, + 380, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first explain the motivating goals and guiding principles behind the design of Meta Omnium. The goal is to build a benchmark for multi-task meta-learning that will: (i) Encourage the community to develop meta-learners that are flexible enough to deal with greater task heterogeneity than before, and thus are more likely to be useful in practice with less curated episode distributions. This was identified as a major challenge in the discussion arising in several recent meta-learning and computer vision workshops and challenges1. (ii) Ultimately progress on this benchmark should provide practical improvements in data-efficient learning", + "bbox": [ + 75, + 660, + 468, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for computer vision through the development of methods that can better transfer across different task types.", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In developing this benchmark, we established a few principles that we used to guide design choices. These include: (i) The benchmark should be lightweight in terms of storage and computing, making it accessible to a broad range of researchers and not only large corporations. (ii) The benchmark should cover multiple tasks with heterogeneous output spaces (as opposed to all classification, all regression, or all dense prediction), as well as multiple visual domains. In these regards, Meta Omnium is compared to alternatives in Table 1. (iii) The initial baselines should have only minimal task-specific decoders. This is in contrast to the state of the art within various sub-disciplines of FSL such as segmentation [25, 42], keypoint [39, 72], and classification [2, 75] where specially designed decoders are often used. This is to evaluate and encourage future research on learning-to-learn across tasks, rather than primarily benchmarking how well we can manually engineer prior knowledge of optimal task-specific decoders. While we are not opposed to future competitors on this benchmark developing task-specific decoders, these should be evaluated separately against the minimal-decoder competitors. (iv) The benchmark should provide distinct datasets for in-distribution (ID) training and out-of-distribution (OOD) evaluation, to evaluate the robustness of the distribution-shift. This is already provided by [63, 65] for classification, and we extend such an ID and OOD dataset ensemble to multiple tasks. Figure 2 illustrates our dataset and task-split. (v) The benchmark should provide a clear hyper-parameter tuning protocol. With a number of recent studies showing that hyper-parameter tuning can dominate other effects of interest in computer vision [21, 32, 45], this is important for a future-proof meta-learning benchmark. This is also related to the first cost point (i) above: Only for a benchmark with a modest cost can most institutions realistically expect to conduct hyperparameter tuning. We provide the hyper-parameter tuning protocol. (vi) Finally, following the debate in [35, 53, 62] as to the value of meta-learning vs conventional transfer learning, the dataset should support both episodic meta-learning and conventional transfer learning approaches.", + "bbox": [ + 496, + 123, + 893, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Splits and Tasks", + "text_level": 1, + "bbox": [ + 500, + 724, + 702, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each main task (classification, segmentation, key-point localization), we split the datasets into seen datasets available for meta-training, and unseen datasets that are completely held out for out-of-domain meta-validation and meta-testing. Similarly to [63, 65], for the seen datasets, we construct category-wise splits into meta-train/val/test. While for the unseen datasets, there is no category-wise split as episodes from all categories from the whole dataset will be used for validation and testing respectively. The overall split organization is illustrated in Figure 2. We", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}$ ICLR https://sites.google.com/view/learning-2-learn/, ECML https://janvanrijn.github.io/metalearning/workshop2022, NeurIPS https://metalearning.chalearn.org/metadlneurips2021, and ECCV http://www.ood-cv.org/", + "bbox": [ + 75, + 838, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7695", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a47e27c77a1999e1448109dd61b535451034f1f2c62b83d9a0d91c8bc3dbbded.jpg", + "image_caption": [ + "Figure 2. Schematic of benchmark and dataset splits. For each task, there are multiple datasets, which are divided into seen (solid border) and unseen (dash border) datasets. The seen (ID) datasets are divided class-wise into meta-train/meta-val/meta-test splits. The unseen datasets are held out for out-of-distribution (OD) evaluation. Meta-training is conducted on the ID-meta-train split of the seen datasets (blue). Models are validated on ID validation class splits, or OOD validation datasets (green). Results are reported on the ID test class splits and OOD test datasets (orange). We also hold out an entire task family (regression) for evaluating novel task generalisation." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 472, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "additionally have a completely held-out task: regression. Datasets from this task are not used during meta-training.", + "bbox": [ + 76, + 481, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our multi-task setup enables us to define and compare two training protocols: Single-task meta-learning which evaluates how well meta-learning performs when trained and tested within a particular task family (within each plate in Figure 2); and Multi-task meta-learning which evaluates how well meta-learning performs when trained across all available task families (across all plates in Figure 2).", + "bbox": [ + 75, + 513, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this organization we can separately evaluate: Within-distribution generalization (ID): How well do meta-learners generalize to novel test concepts within the seen datasets?; and Out-distribution generalization (OOD): How well do meta-learners generalize to novel concepts in unseen datasets?", + "bbox": [ + 75, + 619, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide two sources of validation data: ID and OOD, and our models are selected based on the combined performance across both. OOD validation is not supposed by the most popular Meta-Dataset benchmark [63] as despite its larger size it does not provide OOD validation datasets.", + "bbox": [ + 75, + 710, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 800, + 276, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the considerations in Section 3.1, our benchmark consists of three main tasks (classification, segmentation, keypoints/pose) and one held-out task (regression).", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classification For classification we take the 10 datasets from the initial public release of Meta-Album [65]. These", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "images are all $128 \\times 128$ and contain 19-706 classes per dataset, with 40 images per class. Three of these datasets are reserved for out-of-distribution meta-validation, and four for out-of-distribution meta-test.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Segmentation For segmentation, we take FSS1000 [37] for in-distribution (10,000 images, 1,000 classes), and combine it with VizWiz [64] for OOD meta-validation (862 images, 22 classes), and modified Pascal5i [58] (7,242 images, 6 classes) and the very distinct medical imaging dataset PH2 [41] (200 images, 3 classes) for meta-testing. The segmentation images originally were of diverse sizes. We resize them all to $224 \\times 224$ for Meta Omnium. Note that VizWiz and Pascal datasets originally contain more classes and images. We exclude the classes that overlap with that in the FSS1000 dataset for few-shot learning, and thus there are no classes overlapping among all the datasets.", + "bbox": [ + 496, + 152, + 892, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Keypoints For keypoints/pose, we take animal-pose [10] for in-distribution, synthetic animal-pose [44] for OOD meta-validation, and MPII human-pose [4] for OOD meta-testing. All images are resized to $128 \\times 128$ . MPII includes about $40\\mathrm{k}$ people in over $25\\mathrm{k}$ images with annotated body keypoints. Animal Pose includes 5 animal categories for 6K instances in over $4\\mathrm{k}$ images. Each animal is cropped from the original image. We keep cats and dogs for training, horses and sheep for in-domain validation, and cow for indomain testing. Synthetic animal pose generates synthetic images using animal CAD models rendered from various viewpoints and lightings on a random background. We keep only the horse and tiger categories in our final datasets.", + "bbox": [ + 496, + 334, + 892, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regression For evaluating regression as a held-out task, we use four datasets corresponding to the test splits of [19]: ShapeNet1D, ShapeNet2D, Distractor and Pascal1D [76]. All images are resized to $128 \\times 128$ . ShapeNet1D aims to predict azimuth angles. It contains 30 categories in total and we keep the 3 categories from the test set. ShapeNet2D further includes 2D rotation with azimuth angles and elevation. The test set of ShapeNet2D contains 300 categories in total with 30 images per category. Distractor aims to predict the position of a target object in the presence of a distractor. It contains 12 categories in total and the test set has 2 categories. Each category contains 1000 objects with 36 images for each. Pascal1D aims to predict azimuth angle. The whole Pascal1D contains 65 objects from 10 categories. The test set contains 15 objects with 100 images for each object. The supplementary material provides full details of all datasets and splits.", + "bbox": [ + 496, + 532, + 893, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training API", + "text_level": 1, + "bbox": [ + 500, + 800, + 638, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For episodic learning, we proceed by (i) sampling a task, (ii) sampling a dataset, (iii) sampling an episode. Under our main protocol we consider variable 1 to 5-shot evaluation, but also evaluate separate 1 and 5 shot settings (training is always done with a variable number of shots - sup", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7696", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "port examples). For classification tasks, we follow [65] in generating 5-way episodes. For segmentation tasks, we follow [37] in considering each episode to be a binary foreground/background classification problem for a novel class and generate 2-way episodes. For keypoint, we form episodes by randomly selecting a class (e.g., animal category) and then randomly selecting a subset of 5 keypoints to localize for each episode. For regression tasks, we generate variable 5 to 25-shot episodes because it is a common practice to use more shots for regression tasks [19].", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For non-episodic/transfer learning, we provide access to the meta-train portions of the seen datasets in conventional mini-batches for conventional single task and multi-task supervised learning.", + "bbox": [ + 75, + 243, + 470, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics For classification tasks, we use standard top-1 accuracy; for segmentation tasks, we use standard mean intersection-over-union (mIoU) that averages over IoU values of all object classes; for keypoint prediction, we report the Percentage of Correct Keypoints (PCK). In detail, a detected joint is considered correct if the distance between the predicted and the true joint is within a certain threshold. In our experiments, the threshold is 0.01 for normalized value, which stands for about 12.8 pixel of input image resolution. For regression tasks, we follow [19] and use the same metrics.", + "bbox": [ + 75, + 305, + 472, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Architecture and Baseline Competitors", + "text_level": 1, + "bbox": [ + 76, + 483, + 413, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed in Section 3.1, we aim to establish baselines that can be adapted to tasks with heterogeneous outputs, with minimal reliance on task-specific decoders. We follow [63, 65] in using a ResNet-18 CNN [23] as a feature extractor architecture. For recognition tasks, we perform multi-class classification immediately after ResNet's GAP. For regression tasks, we perform linear regression directly after the ResNet's GAP. For keypoint tasks, we consider them to be a regression problem from the feature map to the keypoint location. For segmentation tasks, we use a simplified PSPNet-like [81] strategy. We concatenate the extracted feature maps from ResNet's feature pyramid, with upsampling where appropriate, to generate a feature map of size $w \\times h$ , and then do pixel-wise classification with $1 \\times 1$ convolutional layer to obtain the final segmentation map. All tasks thus have only one learnable weight as a minimal classifier/decoder after the common ResNet feature encoder. Based on this common encoder and minimal decoder architecture, we describe our meta-learning baselines. Prototypical Network [59] is a classic meta-learner that exploits nearest-centroid metric learning for few-shot classification. ProtoNets were adapted to segmentation tasks in PANet [68], by performing pixel-level feature matching between support prototypes and query pixels. We use the same principle together with the PSPNet-like features described earlier. To generalize ProtoNets to regression tasks such as", + "bbox": [ + 75, + 508, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "keypoint prediction, we must relax the prototype assumption, and use them as simple Gaussian kernel-regression models [7]. Specifically, we generate a feature embedding for each support example, and then for query examples, we calculate the negative exponential distance to the support examples, and use this inverse distance-weighted sum of support set labels as the prediction. Thus for regression tasks with a support set $S = \\{(x_i,y_i)\\}$ and query example $x_{q}$ , ProtoRegression predicts", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (y _ {q} | x _ {q}, S) \\propto \\sum_ {i} y _ {i} \\exp (- \\tau (f _ {\\theta} (x _ {i}) - f _ {\\theta} (x _ {q})) ^ {2}) \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 239, + 893, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "enabling us to learn deep feature $f_{\\theta}$ in the usual episodic meta-learning way. We use cross-entropy loss for classification and segmentation, and MAE loss for regression tasks.", + "bbox": [ + 496, + 281, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DDRR Deep differentiable ridge-regression has been considered for few-shot recognition [6], tracking [82], and other tasks. It is related to ProtoNet in that the feature is not adapted after the meta-train stage, but different in that the decoder/classifier layer is learned by differentiable ridge-regression rather than nearest centroid or kernel regression. An elegant property of DDRR methods is that they naturally address regression tasks, although they have been repurposed for classification [6] by conducting MSE-loss regression to a target 1-hot vector. Thus they are a natural choice for our benchmark. For application to segmentation, we apply DDRR in a $1 \\times 1$ convolution-like way, to perform pixel classification for the output mask with a DDRR classifier at each pixel. Further, we calibrate the prediction for binary cross entropy loss with a learnable scale and bias following [6]. DDRR uses MAE loss for only regression tasks and use MSE loss for other tasks.", + "bbox": [ + 496, + 327, + 892, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MAML The seminal few-shot meta-learner MAML [16] aims to learn an initial condition for per-episode gradient-descent. MAML is straightforward to adapt to different types of tasks. Based on each episode's support set, a new output layer is learned, and the feature extractor is updated, both by a few steps of gradient descent. Similarly to Meta-Dataset [63], we do not learn an initialization for the output layer, since it can change size between episodes drawn from multiple tasks. To alleviate this challenge, we also follow Meta-Dataset in evaluating Proto-MAML – a variant that initializes the MAML output layer based on the linear classifier/regressor suggested by nearest-centroid prior to gradient descent. Going beyond this, to adapt Proto-MAML to regression tasks, we also initialize the output layer based on the ridge-regression solution to the support set.", + "bbox": [ + 496, + 583, + 893, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meta-Curvature Meta-Curvature [48] is an enhancement of MAML that learns a pre-conditioning matrix to improve inner-loop adaptation, as well as an initial condition as in standard MAML. Meta-Curvature outperforms MAML in simpler single-task few-shot benchmarks.", + "bbox": [ + 496, + 810, + 893, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transfer Learning We also consider standard supervised", + "bbox": [ + 500, + 885, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7697", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "learning on the meta-train tasks for transfer to the target tasks, a strategy reported to be competitive with meta-learning [62]. For adaptation, we explore both linear readout [62, 69] and fine-tuning [65]. Besides learning a new output layer from scratch, we also consider a fine-tuning version that initializes the classifier weights using class prototypes (recognition/segmentation) or ridge regression weights (keypoints/regression), inspired by Proto-MAML. Train-from-Scratch (TFS) We lastly consider training each episode from scratch using only the support set [65].", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Hyperparameter Optimization", + "text_level": 1, + "bbox": [ + 76, + 250, + 349, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As part of our benchmark, we perform hyperparameter optimization (HPO) to ensure we select appropriate hyperparameters for the diverse tasks that we consider. Multi-objective HPO under restricted resources is challenging, so we devise the following HPO protocol: estimate the performance of each candidate configuration on a lower fidelity (lower number of iterations) and then identify the configuration that works the best across all validation datasets considered (combination of in-domain and out-domain datasets, across various task families). Note that fast multi-fidelity methods such as Hyperband [34], ASHA [33] or PASHA [8] are not applicable out of the box in our multi-objective setup, so we decided to train each candidate configuration using fixed 5,000 training iterations. Since different tasks/datasets are of different difficulties (and use different metrics), we normalize the score of each configuration for each validation dataset by the best score for that dataset across all candidate configurations. We then select the configuration with the best average normalized score.", + "bbox": [ + 75, + 273, + 472, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that due to resource constraints we are only able to sample a relatively smaller number of candidates (30), so we utilize a sample efficient state-of-the-art Multi-Objective TPE method [47], available from the Optuna library [3]. We perform HPO for multi-task and single-tasks setups separately, so single task classification, segmentation and keypoint estimation have their own set of hyperparameters; and the multi-task case has its own set. The hyperparameters include the meta-learning rate and optimizer, momentum, and various method-specific hyperparameters (full details are in the supplementary). Once the hyperparameters are chosen, we perform standard training of the model for the full number of iterations.", + "bbox": [ + 75, + 561, + 472, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we aim to use our benchmark to answer the following questions: (1) Which meta-learner performs best on average across a heterogenous range of tasks? Existing benchmarks have evaluated meta-learners for one task at a time, we now use our common evaluation platform to find out if any meta-learner can provide general-purpose learning to learn across different task types, or whether", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each task type prefers a different learner. Similarly, we can ask which meta-learner is most robust to out-of-distribution tasks? (2) Having defined the first multi-task meta-learning benchmark, and generalizations of seminal meta-learners to different kinds of output spaces, we ask which meta-learner performs best for multi-task meta-learning? More generally, is there a trend in gradient-based vs metric-based meta-learner success? (3) Does multi-task meta-learning improve or worsen performance compared to single-task? The former obviously provides more meta-training data, which should be advantageous, but the increased heterogeneity across meta-training episodes in the multi-task case also makes it harder to learn [63, 67]. (4) How does meta-learning perform compared to simple transfer learning, or learning from scratch?", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 328, + 705, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train each meta-learner for 30,000 meta-training iterations, with meta-validation after every 2,500 iterations (used for checkpoint selection). For evaluation during meta-testing we use 600 tasks for each corresponding dataset, and for meta-validation we use 1200 tasks together. We use random seeds to ensure that the same tasks are used across all methods that we compare. For transfer learning approaches (fine-tuning, training from scratch, etc.), we use 20 update steps during evaluation. We only retain the meta-learned shared feature extractor across tasks, and for each new evaluation task, we randomly initialize the output layer so that we can support any number of classes as well as novel task families during meta-testing (in line with [65]).", + "bbox": [ + 496, + 352, + 893, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 500, + 559, + 591, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The main experimental results are shown in Table 2, where rows correspond to different few-shot learners, and columns report the average performance of test episodes, aggregated across multiple datasets in each task family, and broken down by \"seen\" datasets (ID) and \"unseen\" datasets (OOD). The table also reports the average rank of each meta-learner across each dataset, both overall and broken down by ID and OOD datasets. More specifically, for each setting (e.g. cls. ID) we calculate the rank of each method (separately for STL and MTL), and then we average those ranks across cls., seg. and keypoints. From the results, we can draw the following main conclusions:", + "bbox": [ + 496, + 583, + 893, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(1) ProtoNet is the most versatile meta-learner, as shown by its highest average rank in the single-task scenario. This validates our novel Kernel Regression extension of ProtoNet for tackling regression-type keypoint localization tasks. Somewhat surprisingly, ProtoNet is also the most robust to out-of-distribution episodes (OOD) which is different from the conclusion of [63] and others who suggested that gradient-based adaptation is crucial to adapt to OOD data. However, it is also in line with the results of [65] and", + "bbox": [ + 496, + 765, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7698", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24f653677f47fc11e3a0f4b80d21fa62f582276c66df6585968c36f6c38f833c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ClassificationSegmentationKeypointsAverage Rank
IDOODIDOODIDOODIDOODAVG
Single-TaskMAML58.761.654.742.125.433.04.33.33.8
Proto-MAML50.549.746.444.123.622.56.06.36.2
Meta-Curvature64.861.465.649.843.516.02.04.33.2
ProtoNet70.459.475.857.227.833.31.31.71.5
DDRR63.158.766.748.020.531.94.73.74.2
Proto-FineTuning50.850.760.043.421.333.15.34.34.8
FineTuning42.348.250.540.025.730.05.76.76.2
Linear-Readout48.653.434.022.722.126.97.36.77.0
TFS31.542.042.837.621.026.08.38.08.2
Multi-TaskMAML59.158.543.337.424.323.92.74.73.7
Proto-MAML58.563.753.043.221.633.33.01.72.3
Meta-Curvature70.466.942.634.518.225.34.34.74.5
ProtoNet65.958.863.349.720.133.02.72.02.3
DDRR52.851.940.437.322.830.15.04.74.8
Proto-FineTuning52.453.244.837.821.230.04.34.04.2
FineTuning44.151.241.336.118.120.57.77.07.3
Linear-Readout46.050.941.532.619.923.56.38.07.2
TFS21.923.838.735.814.111.09.08.38.7
", + "bbox": [ + 158, + 88, + 810, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the strong performance of prototypes more broadly [9].", + "bbox": [ + 76, + 520, + 442, + 536 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) Coming to multi-task meta-learning the situation is similar in that ProtoNet dominates the other competitors, but now sharing the first place with Proto-MAML.", + "(3) To compare single-task and multi-task meta-learning (top and bottom blocks of Table 2) more easily, Figure 3 shows the difference in meta-testing episode performance after STL and MTL meta-training for each method. Overall STL outperforms the MTL condition, showing that the difficulty of learning from heterogeneous tasks [57,77] outweighs the benefit of the extra available multi-task data.", + "(4) Finally, comparing meta-learning methods with simple transfer learning methods as discussed in [13, 62], the best meta-learners are clearly better than transfer learning for both single and multi-task scenarios." + ], + "bbox": [ + 75, + 539, + 470, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also note that Proto-MAML is better than MAML in the multi-task case, likely due to the importance of a good output-layer initialization in the case of heterogeneous episodes, as per [63]. Meta-Curvature outperforms MAML in single-task in-domain scenarios, in line with previous results [48], but it did not achieve stronger performance out-of-domain or in the multi-task case. Finally, while DDRR is perhaps the most elegant baseline in terms of most naturally spanning all task types, its overall performance is middling.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Additional Analysis", + "text_level": 1, + "bbox": [ + 500, + 518, + 687, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How well can multi-task meta-learners generalize to completely new held-out tasks? We take the multi-task meta-learners (trained on classification, segmentation, keypoints) and evaluate them on four regression benchmarks inspired by [19]: ShapeNet1D, ShapeNet2D, Distractor and Pascal1D. Because the metrics differ across datasets, we analyse the rankings and summarise the results in Table 3. We see the basic TFS performs the worst, with MAML, ProtoNets, DDRR being the best. However, in several cases the results were not better than predicting the mean (full results in the appendix), showing that learning-to-learn of completely new task families is an open challenge.", + "bbox": [ + 496, + 542, + 890, + 726 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d42ba26c3248fecfba30e63f677723ada5e2521af80f973b3b6cde52e4b808a.jpg", + "table_caption": [ + "Table 2. Main Results. Results are presented as averages across the datasets within each task type and separately for in-distribution (ID) and out-of-distribution (OOD) datasets. Classification, segmentation, and keypoint results are reported in accuracy $(\\%)$ , mIOU $(\\%)$ , and PCK $(\\%)$ respectively. The upper and lower groups correspond to multi-task and single-task meta-training prior to evaluation on the same set of meta-testing episodes. Upper and lower sub row groups correspond to meta-learners and non-meta learners respectively. See the appendix for a full breakdown over individual datasets." + ], + "table_footnote": [], + "table_body": "
MAMLPMAMLMCPNDDRRPFTFTLRTFS
3.36.54.83.53.53.85.84.38.5
", + "bbox": [ + 500, + 738, + 890, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Average ranking of the different methods across four out-of-task regression datasets.", + "bbox": [ + 498, + 777, + 890, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How much does external pre-training help? Our focus is on assessing the efficacy of meta-learning rather than representation transfer, but we also aim to support researchers investigating the impact of representation learning on external data prior to meta-learning [13, 27, 80]. We therefore", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7699", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d0fefa6ee20d4b2272b0d0732322bc1aed07e4d4dcad11f9ea04e6f87149ece.jpg", + "image_caption": [ + "Figure 3. Analysis of the differences in scores between single-task (STL) and multi-task (MTL) learning for different methods." + ], + "image_footnote": [], + "bbox": [ + 122, + 95, + 848, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c44f212ea733b953c0d72e77dd3fb193da7ee1b88884ecaa20dbe0afb9de33c6.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Analysis of the impact of external data pretraining for selected meta-learners in multi-task learning condition. The results show that ImageNet pretraining does not necessarily help improve performance. Cls., Segm., Keyp. represent classification, segmentation and keypoint respectively." + ], + "table_body": "
MethodPretrainCls.Segm.Keyp.
IDOODIDOODIDOOD
Proto-MAMLX58.563.753.043.221.633.3
ProtoNetX66.058.863.349.720.133.0
Proto-MAML✓63.962.756.245.321.833.3
ProtoNet✓63.558.662.049.020.133.1
", + "bbox": [ + 81, + 295, + 468, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "specify evaluation conditions where external data outside our defined in-distribution meta-training set is allowed.", + "bbox": [ + 75, + 477, + 468, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We take two high-performing approaches in the multi-task scenario, Proto-MAML and ProtoNet, and we investigate to what extent external pre-training helps. We use the standard ImageNet1k pre-trained ResNet18, prior to conducting our meta-learning pipeline as usual. We use the same hyperparameters as selected earlier for these models to ensure consistent evaluation, and ensure that the differences in performance are not due to a better selection of hyperparameters. The results in Table 4 show that pretraining is not necessarily helpful in the considered multi-task setting, in contrast to purely recognition-focused evaluations [13,27,80], which were unambiguously positive about representation transfer from external data.", + "bbox": [ + 75, + 508, + 468, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Runtimes We analyze the times that the different meta-learning approaches spend on meta-training, meta-validation and meta-testing in the multitask learning case of our benchmark. The results in Table 5 show that all experiments are relatively lightweight, despite the ambitious goal of our benchmark to learn a meta-learner that can generalize across various task families. Most notably we observe that ProtoNets are the fastest approach, alongside being the best-performing one. Note that fine-tuning and training from scratch are expensive during the test time as they use backpropagation with a larger number of steps.", + "bbox": [ + 75, + 704, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion and Future Work In future Meta Omnium can be used in a variety of ways beyond benchmarking", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/08dd6804cdd15f09fa1646c176cb024d7b8676b5f24612e34232a9227bf81312.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Analysis of times needed by different algorithms in the multitask setting (using one NVIDIA 1080 Ti GPU and 4 CPUs)." + ], + "table_body": "
MethodTrain TimeVal TimeTest TimeTotal Time
MAML1.8h1.9h0.9h5.0h
Proto-MAML1.9h1.9h0.9h5.1h
Meta-Curvature3.4h2.6h1.3h7.6h
ProtoNet0.8h0.4h0.2h1.8h
DDRR1.4h0.6h0.3h2.7h
Proto-FineTuning1.7h4.5h2.3h8.9h
FineTuning1.5h8.1h4.9h14.9h
Linear-Readout1.2h5.1h2.8h9.6h
TFS0.0h0.8h6.2h7.0h
", + "bbox": [ + 501, + 295, + 890, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "multi-task meta-learning per-se. These including: studying the multi-task optimisation [77] in meta-learning, studying HPO for meta-learning, developing validation strategies in meta-learning (using ID vs OD val sets [32]), and studying the benefit of task-specific decoders and external data.", + "bbox": [ + 496, + 488, + 890, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 580, + 617, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced Meta Omnium, the first multi-task few-shot meta-learning benchmark for computer vision. The benchmark is challenging in multiple highly topical ways including requiring learning on heterogeneous task distributions, evaluating generalization to out-of-distribution datasets, and uniquely challenging meta-learners to learn-to-learn and transfer knowledge across tasks with heterogeneous output spaces. Meta Omnium is nevertheless lightweight enough to be of broad interest and use for driving future research, and even to support future research in hyper-parameter optimization for meta-learning.", + "bbox": [ + 496, + 608, + 890, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by the UKRI and EPSRC Grant numbers EP/S000631/1, EP/L016427/1 and EP/S02431X/1; the MOD University Defence Research Collaboration (UDRC) in Signal Processing; the Royal Academy of Engineering under the Research Fellowship programme; Samsung AI Center, Cambridge; China Grant No.2020KJ010802 and No.BYSYZHKC2021115.", + "bbox": [ + 496, + 779, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7700", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Milad Abdollahzadeh, Touba Malekzadeh, and Ngai-Man Man Cheung. Revisit multimodal meta-learning through the lens of multi-task learning. In NeurIPS, 2021. 2, 3", + "[2] Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, and Christian Gagné. Matching feature sets for few-shot image classification. In CVPR, 2022. 3", + "[3] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In KDD, 2019. 6", + "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, 2014. 4, 12", + "[5] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. 2022. 1", + "[6] Luca Bertinetto, João F. Henriques, Philip H.S. Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In ICLR, 2019. 5", + "[7] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. 5", + "[8] Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, and Giovanni Zappella. PASHA: Efficient HPO and NAS with progressive resource allocation. In ICLR, 2023. 6", + "[9] Ondrej Bohdal, Da Li, and Timothy Hospedales. Feedforward source-free domain adaptation via class prototypes. In ECCV OOD-CV workshop, 2022. 7", + "[10] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In ICCV, 2019. 4, 12", + "[11] Ruchika Chavhan, Jan Stuehmer, Calum Heggan, Mehrdad Yaghoobi, and Timothy Hospedales. Amortised invariance learning for contrastive self-supervision. In ICLR, 2023. 2", + "[12] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 2017. 12", + "[13] Vincent Dumoulin, Neil Houlsby, Utku Evci, Xiaohua Zhai, Ross Goroshin, Sylvain Gelly, and Hugo Larochelle. A unified few-shot classification benchmark to compare transfer and meta learning approaches. In NeurIPS Datasets and Benchmarks, 2021. 2, 7, 8", + "[14] Nikita Dvornik, Cordelia Schmid, and Julien Mairal. Selecting relevant features from a universal representation for few-shot classification. In ECCV, 2020. 3", + "[15] Linus Ericsson, Henry Gouk, and Timothy M. Hospedales. Why do self-supervised models transfer? Investigating the impact of invariance on downstream tasks. BMVC, 2022. 2", + "[16] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2, 5", + "[17] Mario Fritz, E. Hayman, B. Caputo, and J. Eklundh. THE KTH-TIPS database. 2004. 12" + ], + "bbox": [ + 78, + 116, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Geetharamani G. and Arun Pandian J. Identification of plant leaf diseases using a nine-layer deep convolutional neural network. Computers and Electrical Engineering, 76:323-338, 2019. 12", + "[19] Ning Gao, Hanna Ziesche, Ngo Anh Vien, Michael Volpp, and Gerhard Neumann. What matters for meta-learning vision regression tasks? In CVPR, 2022. 2, 4, 5, 7, 12", + "[20] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In ICCV, 2021. 1, 2", + "[21] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In ICLR, 2021. 3", + "[22] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, 2020. 2", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5", + "[24] Emily F. Brownlee Heidi M. Sosik, Emily E. Peacock. Annotated plankton images - data set for developing and evaluating classification methods., 2015. 12", + "[25] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In ECCV, 2022. 3", + "[26] Timothy M Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2021. 1, 2", + "[27] Shell Hu, Da Li, Jan Stuehmer, Minyoung Kim, and Timothy Hospedales. Pushing the limits of simple pipelines for few-shot learning. In CVPR, 2022. 7, 8", + "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE Workshop on 3D Representation and Recognition (3dRR-13), 2013. 12", + "[29] Gustaf Kylberg. The Kylberg texture dataset v. 1.0. Technical report, 2011. 12", + "[30] Brenden M. Lake, Ruslan Salakhutdinov, and Joshua B. Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 2015. 2", + "[31] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. A sparse texture representation using local affine regions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(8):1265-1278, Aug. 2005. 12", + "[32] Da Li, Henry Gouk, and Timothy Hospedales. Finding lost DG: Explaining domain generalization via model complexity. In arXiv, 2022. 3, 8", + "[33] Liam Li, Kevin Jamieson, Afshin Rostamizadeh, Ekaterina Gonina, Moritz Hardt Hardt, Benjamin Recht, and Ameet Talwalkar. A system for massively parallel hyperparameter tuning. In MLSys, 2020. 6", + "[34] Lisha Li, Kevin G. Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter optimization. JMLR, 2018. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7701", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Rui Li, Ondrej Bohdal, Rajesh K. Mishra, Hyeji Kim, Da Li, Nicholas D. Lane, and Timothy M. Hospedales. A channel coding benchmark for meta-learning. 2021. 2, 3", + "[36] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In ICCV, 2021. 3", + "[37] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. 1, 2, 3, 4, 5, 12", + "[38] Yanbin Liu, Juho Lee, Linchao Zhu, Ling Chen, Humphrey Shi, and Yi Yang. A multi-mode modulator for multi-domain few-shot classification. In ICCV, 2021. 3", + "[39] Changsheng Lu and Piotr Koniusz. Few-shot keypoint detection with uncertainty learning for unseen species. In CVPR, 2022. 3", + "[40] P. Mallikarjuna, Alireza Tavakoli Targhi, Mario Fritz, E. Hayman, B. Caputo, and J. Eklundh. The KTH-TIPS 2 database. 2006. 12", + "[41] Teresa Mendonça, Pedro M. Ferreira, Jorge S. Marques, André R. S. Marcal, and Jorge Rozeira. PH2 - a dermoscopic image database for research and benchmarking. In Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2013. 4, 12", + "[42] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In ICCV, 2021. 3", + "[43] Ari Morcos, Maithra Raghu, and Samy Bengio. Insights on representational similarity in neural networks with canonical correlation. Advances in Neural Information Processing Systems, 31, 2018. 12", + "[44] Jiteng Mu, Weichao Qiu, Gregory D. Hager, and Alan L. Yuille. Learning from synthetic animals. In CVPR, 2020. 4, 12", + "[45] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. Unsupervised domain adaptation: A reality check. In arXiv, 2021. 3", + "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 12", + "[47] Yoshihiko Ozaki, Yuki Tanigaki, Shuhei Watanabe, and Masaki Onishi. Multiobjective tree-structured parzen estimator for computationally expensive optimization problems. In Genetic and Evolutionary Computation Conference, 2020. 6", + "[48] Eunbyung Park and Junier B. Oliva. Meta-Curvature. In NeurIPS, 2019. 5, 7", + "[49] Massimiliano Patacchiola, Jack Turner, Elliot J Crowley, Michael O'Boyle, and Amos J Storkey. Bayesian meta-learning for the few-shot setting via deep kernels. In NeurIPS, 2020. 1", + "[50] Gerald Piosenko. 100 sports image classification. 12", + "[51] Gerald Piosenko. Birds 400 - species image classification. 12", + "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1", + "[53] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? Towards understanding the effectiveness of maml. In ICLR, 2020. 2, 3, 12", + "[54] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 12", + "[55] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A generalist agent, 2022. 1", + "[56] Mengye Ren, Eleni Triantafillou, Jake Snell Sachin Ravi, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In ICLR, 2018. 2", + "[57] Sebastian Ruder. An overview of multi-task learning in deep neural networks. In arXiv, 2017. 2, 7", + "[58] Amirreza Shaban, Shray Bansal, Z. Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. In BMVC, 2017. 4, 12", + "[59] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 5", + "[60] Haozhe Sun, Wei-Wei Tu, and Isabelle M Guyon. Omniprint: A configurable printed character synthesizer. In NeurIPS Datasets and Benchmarks Track, 2021. 12", + "[61] Sebastian Thrun and Lorien Pratt, editors. Learning to Learn. Kluwer Academic Publishers, 1998. 1", + "[62] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In ECCV, 2020. 2, 3, 6, 7", + "[63] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, and Hugo Larochelle. Meta-dataset: A dataset of datasets for learning to learn from few examples. In ICLR, 2020. 1, 2, 3, 4, 5, 6, 7", + "[64] Yu-Yun Tseng, Alexander Bell, and Danna Gurari. Vizwiz-fewshot: Locating objects in images taken by people with visual impairments. In ECCV, 2022. 4, 12", + "[65] Ihsan Ullah, Dustin Carrion, Sergio Escalera, Isabelle Guyon, Mike Huisman, Felix Mohr, Jan N. van Rijn, Haozhe Sun, Joaquin Vanschoren, and Phan Anh Vu. Meta-album: Multi-domain meta-dataset for few-shot image classification. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2022. 1, 2, 3, 4, 5, 6, 12, 14", + "[66] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NIPS, 2016. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7702", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Risto Vuorio, Shao-Hua Sun, Hexiang Hu, and Joseph J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. In NeurIPS, 2019. 2, 3, 6", + "[68] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In ICCV, 2019. 5", + "[69] Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning, 2019. 6", + "[70] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 1, 2", + "[71] Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darryll. What should not be contrastive in contrastive learning. In ICLR, 2021. 2", + "[72] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV, 2020. 3", + "[73] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In ECCV, 2022. 2, 3", + "[74] Yongxin Yang and Timothy Hospedales. A unified perspective on multi-domain and multi-task learning. In ICLR, 2015. 2", + "[75] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In CVPR, 2020. 3", + "[76] Mingzhang Yin, George Tucker, Mingyuan Zhou, Sergey Levine, and Chelsea Finn. Meta-learning without memorization. 2020. 4, 12", + "[77] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In NeurIPS, 2020. 2, 7, 8", + "[78] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In CORL, 2019. 3", + "[79] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 3", + "[80] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. In arXiv, 2019. 2, 7, 8", + "[81] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 5", + "[82] Linyu Zheng, Ming Tang, Yingying Chen, Jinqiao Wang, and Hanqing Lu. Learning feature embeddings for discriminant model based tracking. In ECCV, 2020. 5", + "[83] Bartosz Zielinski, Anna Plichta, Krzysztof Misztal, Przemysław Spurek, Monika Brzychczy-Wloch, and Dorota Ochońska. Deep learning approach to bacterial colony classification. PLOS ONE, 12(9):1-14, 09 2017. 12" + ], + "bbox": [ + 78, + 90, + 468, + 896 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7703", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_model.json b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3dd31e299aad3eca6c71cf3f590c87ae13809ae4 --- /dev/null +++ b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.842, + 0.154 + ], + "angle": 0, + "content": "Meta Omnium: A Benchmark for General-Purpose Learning-to-Learn" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.18, + 0.772, + 0.218 + ], + "angle": 0, + "content": "Ondrej Bohdal\\(^{1,\\ast}\\), Yinbing Tian\\(^{2,\\ast}\\), Yongshuo Zong\\(^{1}\\), Ruchika Chavhan\\(^{1}\\), Da Li\\(^{3}\\), Henry Gouk\\(^{1}\\), Li Guo\\(^{2}\\), Timothy Hospedales\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.221, + 0.765, + 0.257 + ], + "angle": 0, + "content": "1The University of Edinburgh 2Beijing University of Posts and Telecommunications 3Samsung AI Center, Cambridge *Joint-First Authors" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.613 + ], + "angle": 0, + "content": "Meta-learning and other approaches to few-shot learning are widely studied for image recognition, and are increasingly applied to other vision tasks such as pose estimation and dense prediction. This naturally raises the question of whether there is any few-shot meta-learning algorithm capable of generalizing across these diverse task types? To support the community in answering this question, we introduce Meta Omnium, a dataset-of-datasets spanning multiple vision tasks including recognition, keypoint localization, semantic segmentation and regression. We experiment with popular few-shot meta-learning baselines and analyze their ability to generalize across tasks and to transfer knowledge between them. Meta Omnium enables meta-learning researchers to evaluate model generalization to a much wider array of tasks than previously possible, and provides a single framework for evaluating meta-learners across a wide suite of vision applications in a consistent manner. Code and dataset are available at https://github.com/edi-meta-learning/meta-omnium." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.21, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Meta-learning is a long-standing research area that aims to replicate the human ability to learn from a few examples by learning-to-learn from a large number of learning problems [61]. This area has become increasingly important recently, as a paradigm with the potential to break the data bottleneck of traditional supervised learning [26, 70]. While the largest body of work is applied to image recognition, few-shot learning algorithms have now been studied in most corners of computer vision, from semantic segmentation [37] to pose estimation [49] and beyond. Nevertheless, most of these applications of few-shot learning are advancing independently, with increasingly divergent application-specific methods and benchmarks. This makes it hard to evaluate whether few-shot meta-learners can solve diverse vision tasks. Importantly it also discourages the de" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.293, + 0.895, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.632, + 0.894, + 0.675 + ], + "angle": 0, + "content": "Figure 1. Illustration of the diverse visual domains and task types in Meta Omnium. Meta-learners are required to generalize across multiple task types, multiple datasets, and held-out datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.763 + ], + "angle": 0, + "content": "velopment of meta-learners with the ability to learn-to-learn across tasks, transferring knowledge from, e.g., keypoint localization to segmentation – a capability that would be highly valuable for vision systems if achieved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The overall trend in computer vision [20, 52] and AI [5, 55] more generally is towards more general-purpose models and algorithms that support many tasks and ideally leverage synergies across them. However, it has not yet been possible to explore this trend in meta-learning, due to the lack of few-shot benchmarks spanning multiple tasks. State-of-the-art benchmarks [63, 65] for visual few-shot learning are restricted to image recognition across a handful of visual domains. There is no few-shot benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7693" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.894, + 0.198 + ], + "angle": 0, + "content": "
DatasetNum TasksNum DomainsNum ImagesCategoriesSizeLightweightMulti-TaskMulti-Domain
Omniglot [30]1132K1623148MB✓✗✗
miniImageNet [66]1160K1001GB✓✗✗
Meta-Dataset [63]17~1053M43~1500210GB✗✗✓
VTAB [80]13~192.2M2~397100GB✗✗✓
FSS1000 [37]11100001000670MB✓✗✗
Meta-Album [65]110~401.5M19~70615GB✓✗✓
Meta Omnium421160K2~7063.1GB✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.208, + 0.893, + 0.237 + ], + "angle": 0, + "content": "Table 1. Feature comparison between Meta Omnium and other few-shot meta-learning benchmarks. Meta Omnium uniquely combines a rich set of tasks and visual domains with a lightweight size for accessible use." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.471, + 0.429 + ], + "angle": 0, + "content": "that poses the more substantial challenge [57, 77] of generalizing across different tasks. We remark that the term task is used differently in few-shot meta-learning literature [16, 26, 70] (to mean different image recognition problems, such as cat vs dog or car vs bus) and the multi-task literature [20, 57, 74, 77] (to mean different kinds of image understanding problems, such as classification vs segmentation). In this paper, we will use the term task in the multi-task literature sense, and the term episode to refer to tasks in the meta-learning literature sense, corresponding to a support and query set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.763 + ], + "angle": 0, + "content": "We introduce Meta Omnium, a dataset-of-datasets spanning multiple vision tasks including recognition, semantic segmentation, keypoint localization/pose estimation, and regression as illustrated in Figure 1. Specifically, Meta Omnium provides the following important contributions: (1) Existing benchmarks only test the ability of meta-learners to learn-to-learn within tasks such as classification [63, 65], or dense prediction [37]. Meta Omnium uniquely tests the ability of meta-learners to learn across multiple task types. (2) Meta Omnium covers multiple visual domains (from natural to medical and industrial images). (3) Meta Omnium provides the ability to thoroughly evaluate both in-distribution and out-of-distribution generalisation. (4) Meta Omnium has a clear hyper-parameter tuning (HPO) and model selection protocol, to facilitate future fair comparison across current and future meta-learning algorithms. (5), Unlike popular predecessors, [63], and despite the diversity of tasks, Meta Omnium has been carefully designed to be of moderate computational cost, making it accessible for research in modestly-resourced universities as well as large institutions. Table 1 compares Meta Omnium to other relevant meta-learning datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We expect Meta Omnium to advance the field by encouraging the development of meta-learning algorithms capable of knowledge transfer across different tasks – as well as across learning episodes within individual tasks as is popularly studied today [16, 70]. In this regard, it provides the next step of the level of a currently topical challenge of dealing with heterogeneity in meta-learning [1, 35, 63, 67]. While existing benchmarks have tested multi-domain heterogeneity (e.g., recognition of written characters and plants" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.895, + 0.46 + ], + "angle": 0, + "content": "within a single network) [63, 65] and shown it to be challenging, Meta Omnium tests multi-task learning (e.g., character recognition vs plant segmentation). This is substantially more ambitious when considered from the perspective of common representation learning. For example, a representation tuned for recognition might benefit from rotation invariance, while one tuned for segmentation might benefit from rotation equivariance [11, 15, 71]. Thus, in contrast to conventional within-task meta-learning benchmarks that have been criticized as relying more on common representation learning than learning-to-learn [53, 62], Meta Omnium better tests the ability of learning-to-learn since the constituent tasks require more diverse representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.642, + 0.489 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.747, + 0.515 + ], + "angle": 0, + "content": "2.1. Meta-learning Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.825 + ], + "angle": 0, + "content": "The classic datasets in few-shot meta-learning for computer vision are Omniglot [30] and miniImageNet [66]. Later work criticized these for having insufficient task (episode) diversity and tieredImageNet [56] used the class hierarchy of ImageNet to enforce more diversity between meta-train and meta-test episodes. The main contemporary benchmarks are CD-FSL [22], which challenges few-shot learners to generalize to new visual domains; and Meta-Dataset [63] and Meta-Album [65], which go further in requiring few-shot learners to learn from a mixture of visual domains. Such multi-domain heterogeneous meta-learning turns out to be challenging. A related benchmark to Meta-Dataset is VTAB [80], which similarly provides multiple domains for evaluating data-efficient visual recognition, but their focus is on evaluating representation transfer from large-scale pre-training rather than learning-to-learn and meta-learning. VTAB+MD [13] compare representation transfer and meta-learning approaches on the Meta-Dataset tasks. However, none of these benchmarks address multi-task meta-learning as considered here (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Outside of recognition, task-specific few-shot benchmarks have been proposed in vision problems of semantic segmentation [37], regression [19], pose/keypoint estimation [73], etc. These are mostly slightly behind the complexity of the recognition benchmarks with regards to being" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7694" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "single-domain, with the exception of [73]. With regard to multi-task meta-learning as considered here, the only existing benchmark is meta-world [78], which is specific to robotics and reinforcement learning rather than vision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.274 + ], + "angle": 0, + "content": "We also mention taskonomy [79] as a popular dataset that has been used for multi-task learning. However, it is not widely used for few-shot meta-learning. This is because, although taskonomy has many tasks, unlike the main meta-learning benchmarks [37, 63], there are not enough visual concepts within each task to provide a large number of concepts for meta-training and a disjoint set of concepts to evaluate few-shot learning for meta-validation and meta-testing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.285, + 0.363, + 0.302 + ], + "angle": 0, + "content": "2.2. Heterogeneity in Meta-Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.471, + 0.596 + ], + "angle": 0, + "content": "There are several sophisticated methods in the literature that highlighted the challenge of the task to address heterogeneity in meta-learning. These have gone under various names such as multi-modal meta-learning [1, 38, 67] – in the sense of multi-modal probability distribution (over tasks/episodes). However, with the exception of [38], these have mostly not been shown to scale to the main multimodal benchmarks such as Meta-Dataset [63]. A more common approach to achieving high performance across multiple heterogeneous domains such as those in Meta-Dataset is to train an ensemble of feature extractors across available training domains, and fuse them during metatesting [14, 36, 36]. However, this obviously incurs a substantial additional cost of maintaining a model ensemble. In our evaluation, we focus on the simpler meta-learners that have been shown to work in challenging multi-domain learning scenarios [63, 65], while leaving sophisticated algorithmic and ensemble-based approaches for future researchers to evaluate on the benchmark." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.441, + 0.628 + ], + "angle": 0, + "content": "3. Meta Omnium Benchmark and Datasets" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.638, + 0.381, + 0.654 + ], + "angle": 0, + "content": "3.1. Motivation and Guiding Principles" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.47, + 0.829 + ], + "angle": 0, + "content": "We first explain the motivating goals and guiding principles behind the design of Meta Omnium. The goal is to build a benchmark for multi-task meta-learning that will: (i) Encourage the community to develop meta-learners that are flexible enough to deal with greater task heterogeneity than before, and thus are more likely to be useful in practice with less curated episode distributions. This was identified as a major challenge in the discussion arising in several recent meta-learning and computer vision workshops and challenges1. (ii) Ultimately progress on this benchmark should provide practical improvements in data-efficient learning" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "for computer vision through the development of methods that can better transfer across different task types." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.895, + 0.715 + ], + "angle": 0, + "content": "In developing this benchmark, we established a few principles that we used to guide design choices. These include: (i) The benchmark should be lightweight in terms of storage and computing, making it accessible to a broad range of researchers and not only large corporations. (ii) The benchmark should cover multiple tasks with heterogeneous output spaces (as opposed to all classification, all regression, or all dense prediction), as well as multiple visual domains. In these regards, Meta Omnium is compared to alternatives in Table 1. (iii) The initial baselines should have only minimal task-specific decoders. This is in contrast to the state of the art within various sub-disciplines of FSL such as segmentation [25, 42], keypoint [39, 72], and classification [2, 75] where specially designed decoders are often used. This is to evaluate and encourage future research on learning-to-learn across tasks, rather than primarily benchmarking how well we can manually engineer prior knowledge of optimal task-specific decoders. While we are not opposed to future competitors on this benchmark developing task-specific decoders, these should be evaluated separately against the minimal-decoder competitors. (iv) The benchmark should provide distinct datasets for in-distribution (ID) training and out-of-distribution (OOD) evaluation, to evaluate the robustness of the distribution-shift. This is already provided by [63, 65] for classification, and we extend such an ID and OOD dataset ensemble to multiple tasks. Figure 2 illustrates our dataset and task-split. (v) The benchmark should provide a clear hyper-parameter tuning protocol. With a number of recent studies showing that hyper-parameter tuning can dominate other effects of interest in computer vision [21, 32, 45], this is important for a future-proof meta-learning benchmark. This is also related to the first cost point (i) above: Only for a benchmark with a modest cost can most institutions realistically expect to conduct hyperparameter tuning. We provide the hyper-parameter tuning protocol. (vi) Finally, following the debate in [35, 53, 62] as to the value of meta-learning vs conventional transfer learning, the dataset should support both episodic meta-learning and conventional transfer learning approaches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.704, + 0.741 + ], + "angle": 0, + "content": "3.2. Data Splits and Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.903 + ], + "angle": 0, + "content": "For each main task (classification, segmentation, key-point localization), we split the datasets into seen datasets available for meta-training, and unseen datasets that are completely held out for out-of-domain meta-validation and meta-testing. Similarly to [63, 65], for the seen datasets, we construct category-wise splits into meta-train/val/test. While for the unseen datasets, there is no category-wise split as episodes from all categories from the whole dataset will be used for validation and testing respectively. The overall split organization is illustrated in Figure 2. We" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\(^{1}\\)ICLR https://sites.google.com/view/learning-2-learn/, ECML https://janvanrijn.github.io/metalearning/workshop2022, NeurIPS https://metalearning.chalearn.org/metadlneurips2021, and ECCV http://www.ood-cv.org/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7695" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.088, + 0.473, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.455 + ], + "angle": 0, + "content": "Figure 2. Schematic of benchmark and dataset splits. For each task, there are multiple datasets, which are divided into seen (solid border) and unseen (dash border) datasets. The seen (ID) datasets are divided class-wise into meta-train/meta-val/meta-test splits. The unseen datasets are held out for out-of-distribution (OD) evaluation. Meta-training is conducted on the ID-meta-train split of the seen datasets (blue). Models are validated on ID validation class splits, or OOD validation datasets (green). Results are reported on the ID test class splits and OOD test datasets (orange). We also hold out an entire task family (regression) for evaluating novel task generalisation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.482, + 0.47, + 0.514 + ], + "angle": 0, + "content": "additionally have a completely held-out task: regression. Datasets from this task are not used during meta-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.47, + 0.62 + ], + "angle": 0, + "content": "Our multi-task setup enables us to define and compare two training protocols: Single-task meta-learning which evaluates how well meta-learning performs when trained and tested within a particular task family (within each plate in Figure 2); and Multi-task meta-learning which evaluates how well meta-learning performs when trained across all available task families (across all plates in Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.711 + ], + "angle": 0, + "content": "With this organization we can separately evaluate: Within-distribution generalization (ID): How well do meta-learners generalize to novel test concepts within the seen datasets?; and Out-distribution generalization (OOD): How well do meta-learners generalize to novel concepts in unseen datasets?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.47, + 0.788 + ], + "angle": 0, + "content": "We provide two sources of validation data: ID and OOD, and our models are selected based on the combined performance across both. OOD validation is not supposed by the most popular Meta-Dataset benchmark [63] as despite its larger size it does not provide OOD validation datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.277, + 0.815 + ], + "angle": 0, + "content": "3.3. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Given the considerations in Section 3.1, our benchmark consists of three main tasks (classification, segmentation, keypoints/pose) and one held-out task (regression)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Classification For classification we take the 10 datasets from the initial public release of Meta-Album [65]. These" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "images are all \\(128 \\times 128\\) and contain 19-706 classes per dataset, with 40 images per class. Three of these datasets are reserved for out-of-distribution meta-validation, and four for out-of-distribution meta-test." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.334 + ], + "angle": 0, + "content": "Segmentation For segmentation, we take FSS1000 [37] for in-distribution (10,000 images, 1,000 classes), and combine it with VizWiz [64] for OOD meta-validation (862 images, 22 classes), and modified Pascal5i [58] (7,242 images, 6 classes) and the very distinct medical imaging dataset PH2 [41] (200 images, 3 classes) for meta-testing. The segmentation images originally were of diverse sizes. We resize them all to \\(224 \\times 224\\) for Meta Omnium. Note that VizWiz and Pascal datasets originally contain more classes and images. We exclude the classes that overlap with that in the FSS1000 dataset for few-shot learning, and thus there are no classes overlapping among all the datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.893, + 0.532 + ], + "angle": 0, + "content": "Keypoints For keypoints/pose, we take animal-pose [10] for in-distribution, synthetic animal-pose [44] for OOD meta-validation, and MPII human-pose [4] for OOD meta-testing. All images are resized to \\(128 \\times 128\\). MPII includes about \\(40\\mathrm{k}\\) people in over \\(25\\mathrm{k}\\) images with annotated body keypoints. Animal Pose includes 5 animal categories for 6K instances in over \\(4\\mathrm{k}\\) images. Each animal is cropped from the original image. We keep cats and dogs for training, horses and sheep for in-domain validation, and cow for indomain testing. Synthetic animal pose generates synthetic images using animal CAD models rendered from various viewpoints and lightings on a random background. We keep only the horse and tiger categories in our final datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.789 + ], + "angle": 0, + "content": "Regression For evaluating regression as a held-out task, we use four datasets corresponding to the test splits of [19]: ShapeNet1D, ShapeNet2D, Distractor and Pascal1D [76]. All images are resized to \\(128 \\times 128\\). ShapeNet1D aims to predict azimuth angles. It contains 30 categories in total and we keep the 3 categories from the test set. ShapeNet2D further includes 2D rotation with azimuth angles and elevation. The test set of ShapeNet2D contains 300 categories in total with 30 images per category. Distractor aims to predict the position of a target object in the presence of a distractor. It contains 12 categories in total and the test set has 2 categories. Each category contains 1000 objects with 36 images for each. Pascal1D aims to predict azimuth angle. The whole Pascal1D contains 65 objects from 10 categories. The test set contains 15 objects with 100 images for each object. The supplementary material provides full details of all datasets and splits." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.64, + 0.817 + ], + "angle": 0, + "content": "3.4. Training API" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "For episodic learning, we proceed by (i) sampling a task, (ii) sampling a dataset, (iii) sampling an episode. Under our main protocol we consider variable 1 to 5-shot evaluation, but also evaluate separate 1 and 5 shot settings (training is always done with a variable number of shots - sup" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7696" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "port examples). For classification tasks, we follow [65] in generating 5-way episodes. For segmentation tasks, we follow [37] in considering each episode to be a binary foreground/background classification problem for a novel class and generate 2-way episodes. For keypoint, we form episodes by randomly selecting a class (e.g., animal category) and then randomly selecting a subset of 5 keypoints to localize for each episode. For regression tasks, we generate variable 5 to 25-shot episodes because it is a common practice to use more shots for regression tasks [19]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.472, + 0.305 + ], + "angle": 0, + "content": "For non-episodic/transfer learning, we provide access to the meta-train portions of the seen datasets in conventional mini-batches for conventional single task and multi-task supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.472 + ], + "angle": 0, + "content": "Evaluation Metrics For classification tasks, we use standard top-1 accuracy; for segmentation tasks, we use standard mean intersection-over-union (mIoU) that averages over IoU values of all object classes; for keypoint prediction, we report the Percentage of Correct Keypoints (PCK). In detail, a detected joint is considered correct if the distance between the predicted and the true joint is within a certain threshold. In our experiments, the threshold is 0.01 for normalized value, which stands for about 12.8 pixel of input image resolution. For regression tasks, we follow [19] and use the same metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.415, + 0.501 + ], + "angle": 0, + "content": "3.5. Architecture and Baseline Competitors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.473, + 0.903 + ], + "angle": 0, + "content": "As discussed in Section 3.1, we aim to establish baselines that can be adapted to tasks with heterogeneous outputs, with minimal reliance on task-specific decoders. We follow [63, 65] in using a ResNet-18 CNN [23] as a feature extractor architecture. For recognition tasks, we perform multi-class classification immediately after ResNet's GAP. For regression tasks, we perform linear regression directly after the ResNet's GAP. For keypoint tasks, we consider them to be a regression problem from the feature map to the keypoint location. For segmentation tasks, we use a simplified PSPNet-like [81] strategy. We concatenate the extracted feature maps from ResNet's feature pyramid, with upsampling where appropriate, to generate a feature map of size \\( w \\times h \\), and then do pixel-wise classification with \\( 1 \\times 1 \\) convolutional layer to obtain the final segmentation map. All tasks thus have only one learnable weight as a minimal classifier/decoder after the common ResNet feature encoder. Based on this common encoder and minimal decoder architecture, we describe our meta-learning baselines. Prototypical Network [59] is a classic meta-learner that exploits nearest-centroid metric learning for few-shot classification. ProtoNets were adapted to segmentation tasks in PANet [68], by performing pixel-level feature matching between support prototypes and query pixels. We use the same principle together with the PSPNet-like features described earlier. To generalize ProtoNets to regression tasks such as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "keypoint prediction, we must relax the prototype assumption, and use them as simple Gaussian kernel-regression models [7]. Specifically, we generate a feature embedding for each support example, and then for query examples, we calculate the negative exponential distance to the support examples, and use this inverse distance-weighted sum of support set labels as the prediction. Thus for regression tasks with a support set \\( S = \\{(x_i,y_i)\\} \\) and query example \\( x_{q} \\), ProtoRegression predicts" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.24, + 0.895, + 0.272 + ], + "angle": 0, + "content": "\\[\nf (y _ {q} | x _ {q}, S) \\propto \\sum_ {i} y _ {i} \\exp (- \\tau (f _ {\\theta} (x _ {i}) - f _ {\\theta} (x _ {q})) ^ {2}) \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.327 + ], + "angle": 0, + "content": "enabling us to learn deep feature \\( f_{\\theta} \\) in the usual episodic meta-learning way. We use cross-entropy loss for classification and segmentation, and MAE loss for regression tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.893, + 0.583 + ], + "angle": 0, + "content": "DDRR Deep differentiable ridge-regression has been considered for few-shot recognition [6], tracking [82], and other tasks. It is related to ProtoNet in that the feature is not adapted after the meta-train stage, but different in that the decoder/classifier layer is learned by differentiable ridge-regression rather than nearest centroid or kernel regression. An elegant property of DDRR methods is that they naturally address regression tasks, although they have been repurposed for classification [6] by conducting MSE-loss regression to a target 1-hot vector. Thus they are a natural choice for our benchmark. For application to segmentation, we apply DDRR in a \\(1 \\times 1\\) convolution-like way, to perform pixel classification for the output mask with a DDRR classifier at each pixel. Further, we calibrate the prediction for binary cross entropy loss with a learnable scale and bias following [6]. DDRR uses MAE loss for only regression tasks and use MSE loss for other tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.811 + ], + "angle": 0, + "content": "MAML The seminal few-shot meta-learner MAML [16] aims to learn an initial condition for per-episode gradient-descent. MAML is straightforward to adapt to different types of tasks. Based on each episode's support set, a new output layer is learned, and the feature extractor is updated, both by a few steps of gradient descent. Similarly to Meta-Dataset [63], we do not learn an initialization for the output layer, since it can change size between episodes drawn from multiple tasks. To alleviate this challenge, we also follow Meta-Dataset in evaluating Proto-MAML – a variant that initializes the MAML output layer based on the linear classifier/regressor suggested by nearest-centroid prior to gradient descent. Going beyond this, to adapt Proto-MAML to regression tasks, we also initialize the output layer based on the ridge-regression solution to the support set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.885 + ], + "angle": 0, + "content": "Meta-Curvature Meta-Curvature [48] is an enhancement of MAML that learns a pre-conditioning matrix to improve inner-loop adaptation, as well as an initial condition as in standard MAML. Meta-Curvature outperforms MAML in simpler single-task few-shot benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.886, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Transfer Learning We also consider standard supervised" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7697" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "learning on the meta-train tasks for transfer to the target tasks, a strategy reported to be competitive with meta-learning [62]. For adaptation, we explore both linear readout [62, 69] and fine-tuning [65]. Besides learning a new output layer from scratch, we also consider a fine-tuning version that initializes the classifier weights using class prototypes (recognition/segmentation) or ridge regression weights (keypoints/regression), inspired by Proto-MAML. Train-from-Scratch (TFS) We lastly consider training each episode from scratch using only the support set [65]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.251, + 0.351, + 0.267 + ], + "angle": 0, + "content": "3.6. Hyperparameter Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.56 + ], + "angle": 0, + "content": "As part of our benchmark, we perform hyperparameter optimization (HPO) to ensure we select appropriate hyperparameters for the diverse tasks that we consider. Multi-objective HPO under restricted resources is challenging, so we devise the following HPO protocol: estimate the performance of each candidate configuration on a lower fidelity (lower number of iterations) and then identify the configuration that works the best across all validation datasets considered (combination of in-domain and out-domain datasets, across various task families). Note that fast multi-fidelity methods such as Hyperband [34], ASHA [33] or PASHA [8] are not applicable out of the box in our multi-objective setup, so we decided to train each candidate configuration using fixed 5,000 training iterations. Since different tasks/datasets are of different difficulties (and use different metrics), we normalize the score of each configuration for each validation dataset by the best score for that dataset across all candidate configurations. We then select the configuration with the best average normalized score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.473, + 0.756 + ], + "angle": 0, + "content": "Note that due to resource constraints we are only able to sample a relatively smaller number of candidates (30), so we utilize a sample efficient state-of-the-art Multi-Objective TPE method [47], available from the Optuna library [3]. We perform HPO for multi-task and single-tasks setups separately, so single task classification, segmentation and keypoint estimation have their own set of hyperparameters; and the multi-task case has its own set. The hyperparameters include the meta-learning rate and optimizer, momentum, and various method-specific hyperparameters (full details are in the supplementary). Once the hyperparameters are chosen, we perform standard training of the model for the full number of iterations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.21, + 0.787 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this section, we aim to use our benchmark to answer the following questions: (1) Which meta-learner performs best on average across a heterogenous range of tasks? Existing benchmarks have evaluated meta-learners for one task at a time, we now use our common evaluation platform to find out if any meta-learner can provide general-purpose learning to learn across different task types, or whether" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "each task type prefers a different learner. Similarly, we can ask which meta-learner is most robust to out-of-distribution tasks? (2) Having defined the first multi-task meta-learning benchmark, and generalizations of seminal meta-learners to different kinds of output spaces, we ask which meta-learner performs best for multi-task meta-learning? More generally, is there a trend in gradient-based vs metric-based meta-learner success? (3) Does multi-task meta-learning improve or worsen performance compared to single-task? The former obviously provides more meta-training data, which should be advantageous, but the increased heterogeneity across meta-training episodes in the multi-task case also makes it harder to learn [63, 67]. (4) How does meta-learning perform compared to simple transfer learning, or learning from scratch?" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.329, + 0.707, + 0.346 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.895, + 0.55 + ], + "angle": 0, + "content": "We train each meta-learner for 30,000 meta-training iterations, with meta-validation after every 2,500 iterations (used for checkpoint selection). For evaluation during meta-testing we use 600 tasks for each corresponding dataset, and for meta-validation we use 1200 tasks together. We use random seeds to ensure that the same tasks are used across all methods that we compare. For transfer learning approaches (fine-tuning, training from scratch, etc.), we use 20 update steps during evaluation. We only retain the meta-learned shared feature extractor across tasks, and for each new evaluation task, we randomly initialize the output layer so that we can support any number of classes as well as novel task families during meta-testing (in line with [65])." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.593, + 0.575 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.765 + ], + "angle": 0, + "content": "The main experimental results are shown in Table 2, where rows correspond to different few-shot learners, and columns report the average performance of test episodes, aggregated across multiple datasets in each task family, and broken down by \"seen\" datasets (ID) and \"unseen\" datasets (OOD). The table also reports the average rank of each meta-learner across each dataset, both overall and broken down by ID and OOD datasets. More specifically, for each setting (e.g. cls. ID) we calculate the rank of each method (separately for STL and MTL), and then we average those ranks across cls., seg. and keypoints. From the results, we can draw the following main conclusions:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.901 + ], + "angle": 0, + "content": "(1) ProtoNet is the most versatile meta-learner, as shown by its highest average rank in the single-task scenario. This validates our novel Kernel Regression extension of ProtoNet for tackling regression-type keypoint localization tasks. Somewhat surprisingly, ProtoNet is also the most robust to out-of-distribution episodes (OOD) which is different from the conclusion of [63] and others who suggested that gradient-based adaptation is crucial to adapt to OOD data. However, it is also in line with the results of [65] and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7698" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.812, + 0.413 + ], + "angle": 0, + "content": "
ClassificationSegmentationKeypointsAverage Rank
IDOODIDOODIDOODIDOODAVG
Single-TaskMAML58.761.654.742.125.433.04.33.33.8
Proto-MAML50.549.746.444.123.622.56.06.36.2
Meta-Curvature64.861.465.649.843.516.02.04.33.2
ProtoNet70.459.475.857.227.833.31.31.71.5
DDRR63.158.766.748.020.531.94.73.74.2
Proto-FineTuning50.850.760.043.421.333.15.34.34.8
FineTuning42.348.250.540.025.730.05.76.76.2
Linear-Readout48.653.434.022.722.126.97.36.77.0
TFS31.542.042.837.621.026.08.38.08.2
Multi-TaskMAML59.158.543.337.424.323.92.74.73.7
Proto-MAML58.563.753.043.221.633.33.01.72.3
Meta-Curvature70.466.942.634.518.225.34.34.74.5
ProtoNet65.958.863.349.720.133.02.72.02.3
DDRR52.851.940.437.322.830.15.04.74.8
Proto-FineTuning52.453.244.837.821.230.04.34.04.2
FineTuning44.151.241.336.118.120.57.77.07.3
Linear-Readout46.050.941.532.619.923.56.38.07.2
TFS21.923.838.735.814.111.09.08.38.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.424, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Table 2. Main Results. Results are presented as averages across the datasets within each task type and separately for in-distribution (ID) and out-of-distribution (OOD) datasets. Classification, segmentation, and keypoint results are reported in accuracy \\((\\%)\\), mIOU \\((\\%)\\), and PCK \\((\\%)\\) respectively. The upper and lower groups correspond to multi-task and single-task meta-training prior to evaluation on the same set of meta-testing episodes. Upper and lower sub row groups correspond to meta-learners and non-meta learners respectively. See the appendix for a full breakdown over individual datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.521, + 0.443, + 0.537 + ], + "angle": 0, + "content": "the strong performance of prototypes more broadly [9]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.54, + 0.47, + 0.587 + ], + "angle": 0, + "content": "(2) Coming to multi-task meta-learning the situation is similar in that ProtoNet dominates the other competitors, but now sharing the first place with Proto-MAML." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.697 + ], + "angle": 0, + "content": "(3) To compare single-task and multi-task meta-learning (top and bottom blocks of Table 2) more easily, Figure 3 shows the difference in meta-testing episode performance after STL and MTL meta-training for each method. Overall STL outperforms the MTL condition, showing that the difficulty of learning from heterogeneous tasks [57,77] outweighs the benefit of the extra available multi-task data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.471, + 0.761 + ], + "angle": 0, + "content": "(4) Finally, comparing meta-learning methods with simple transfer learning methods as discussed in [13, 62], the best meta-learners are clearly better than transfer learning for both single and multi-task scenarios." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.54, + 0.471, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We also note that Proto-MAML is better than MAML in the multi-task case, likely due to the importance of a good output-layer initialization in the case of heterogeneous episodes, as per [63]. Meta-Curvature outperforms MAML in single-task in-domain scenarios, in line with previous results [48], but it did not achieve stronger performance out-of-domain or in the multi-task case. Finally, while DDRR is perhaps the most elegant baseline in terms of most naturally spanning all task types, its overall performance is middling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.52, + 0.688, + 0.537 + ], + "angle": 0, + "content": "4.3. Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.727 + ], + "angle": 0, + "content": "How well can multi-task meta-learners generalize to completely new held-out tasks? We take the multi-task meta-learners (trained on classification, segmentation, keypoints) and evaluate them on four regression benchmarks inspired by [19]: ShapeNet1D, ShapeNet2D, Distractor and Pascal1D. Because the metrics differ across datasets, we analyse the rankings and summarise the results in Table 3. We see the basic TFS performs the worst, with MAML, ProtoNets, DDRR being the best. However, in several cases the results were not better than predicting the mean (full results in the appendix), showing that learning-to-learn of completely new task families is an open challenge." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.739, + 0.892, + 0.768 + ], + "angle": 0, + "content": "
MAMLPMAMLMCPNDDRRPFTFTLRTFS
3.36.54.83.53.53.85.84.38.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.779, + 0.891, + 0.808 + ], + "angle": 0, + "content": "Table 3. Average ranking of the different methods across four out-of-task regression datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "How much does external pre-training help? Our focus is on assessing the efficacy of meta-learning rather than representation transfer, but we also aim to support researchers investigating the impact of representation learning on external data prior to meta-learning [13, 27, 80]. We therefore" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7699" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.096, + 0.849, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.258, + 0.859, + 0.273 + ], + "angle": 0, + "content": "Figure 3. Analysis of the differences in scores between single-task (STL) and multi-task (MTL) learning for different methods." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.296, + 0.47, + 0.372 + ], + "angle": 0, + "content": "
MethodPretrainCls.Segm.Keyp.
IDOODIDOODIDOOD
Proto-MAMLX58.563.753.043.221.633.3
ProtoNetX66.058.863.349.720.133.0
Proto-MAML✓63.962.756.245.321.833.3
ProtoNet✓63.558.662.049.020.133.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.452 + ], + "angle": 0, + "content": "Table 4. Analysis of the impact of external data pretraining for selected meta-learners in multi-task learning condition. The results show that ImageNet pretraining does not necessarily help improve performance. Cls., Segm., Keyp. represent classification, segmentation and keypoint respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.469, + 0.508 + ], + "angle": 0, + "content": "specify evaluation conditions where external data outside our defined in-distribution meta-training set is allowed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.469, + 0.704 + ], + "angle": 0, + "content": "We take two high-performing approaches in the multi-task scenario, Proto-MAML and ProtoNet, and we investigate to what extent external pre-training helps. We use the standard ImageNet1k pre-trained ResNet18, prior to conducting our meta-learning pipeline as usual. We use the same hyperparameters as selected earlier for these models to ensure consistent evaluation, and ensure that the differences in performance are not due to a better selection of hyperparameters. The results in Table 4 show that pretraining is not necessarily helpful in the considered multi-task setting, in contrast to purely recognition-focused evaluations [13,27,80], which were unambiguously positive about representation transfer from external data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Analysis of Runtimes We analyze the times that the different meta-learning approaches spend on meta-training, meta-validation and meta-testing in the multitask learning case of our benchmark. The results in Table 5 show that all experiments are relatively lightweight, despite the ambitious goal of our benchmark to learn a meta-learner that can generalize across various task families. Most notably we observe that ProtoNets are the fastest approach, alongside being the best-performing one. Note that fine-tuning and training from scratch are expensive during the test time as they use backpropagation with a larger number of steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Discussion and Future Work In future Meta Omnium can be used in a variety of ways beyond benchmarking" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.422 + ], + "angle": 0, + "content": "
MethodTrain TimeVal TimeTest TimeTotal Time
MAML1.8h1.9h0.9h5.0h
Proto-MAML1.9h1.9h0.9h5.1h
Meta-Curvature3.4h2.6h1.3h7.6h
ProtoNet0.8h0.4h0.2h1.8h
DDRR1.4h0.6h0.3h2.7h
Proto-FineTuning1.7h4.5h2.3h8.9h
FineTuning1.5h8.1h4.9h14.9h
Linear-Readout1.2h5.1h2.8h9.6h
TFS0.0h0.8h6.2h7.0h
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.432, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Table 5. Analysis of times needed by different algorithms in the multitask setting (using one NVIDIA 1080 Ti GPU and 4 CPUs)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.565 + ], + "angle": 0, + "content": "multi-task meta-learning per-se. These including: studying the multi-task optimisation [77] in meta-learning, studying HPO for meta-learning, developing validation strategies in meta-learning (using ID vs OD val sets [32]), and studying the benefit of task-specific decoders and external data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.582, + 0.619, + 0.599 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.776 + ], + "angle": 0, + "content": "We have introduced Meta Omnium, the first multi-task few-shot meta-learning benchmark for computer vision. The benchmark is challenging in multiple highly topical ways including requiring learning on heterogeneous task distributions, evaluating generalization to out-of-distribution datasets, and uniquely challenging meta-learners to learn-to-learn and transfer knowledge across tasks with heterogeneous output spaces. Meta Omnium is nevertheless lightweight enough to be of broad interest and use for driving future research, and even to support future research in hyper-parameter optimization for meta-learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by the UKRI and EPSRC Grant numbers EP/S000631/1, EP/L016427/1 and EP/S02431X/1; the MOD University Defence Research Collaboration (UDRC) in Signal Processing; the Royal Academy of Engineering under the Research Fellowship programme; Samsung AI Center, Cambridge; China Grant No.2020KJ010802 and No.BYSYZHKC2021115." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7700" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Milad Abdollahzadeh, Touba Malekzadeh, and Ngai-Man Man Cheung. Revisit multimodal meta-learning through the lens of multi-task learning. In NeurIPS, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.47, + 0.216 + ], + "angle": 0, + "content": "[2] Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, and Christian Gagné. Matching feature sets for few-shot image classification. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[3] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In KDD, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.262, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, 2014. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.47, + 0.359 + ], + "angle": 0, + "content": "[5] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.363, + 0.47, + 0.405 + ], + "angle": 0, + "content": "[6] Luca Bertinetto, João F. Henriques, Philip H.S. Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In ICLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.407, + 0.47, + 0.435 + ], + "angle": 0, + "content": "[7] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.437, + 0.47, + 0.491 + ], + "angle": 0, + "content": "[8] Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, and Giovanni Zappella. PASHA: Efficient HPO and NAS with progressive resource allocation. In ICLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[9] Ondrej Bohdal, Da Li, and Timothy Hospedales. Feedforward source-free domain adaptation via class prototypes. In ECCV OOD-CV workshop, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[10] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In ICCV, 2019. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.582, + 0.469, + 0.624 + ], + "angle": 0, + "content": "[11] Ruchika Chavhan, Jan Stuehmer, Calum Heggan, Mehrdad Yaghoobi, and Timothy Hospedales. Amortised invariance learning for contrastive self-supervision. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.626, + 0.469, + 0.667 + ], + "angle": 0, + "content": "[12] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.67, + 0.469, + 0.738 + ], + "angle": 0, + "content": "[13] Vincent Dumoulin, Neil Houlsby, Utku Evci, Xiaohua Zhai, Ross Goroshin, Sylvain Gelly, and Hugo Larochelle. A unified few-shot classification benchmark to compare transfer and meta learning approaches. In NeurIPS Datasets and Benchmarks, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.741, + 0.469, + 0.782 + ], + "angle": 0, + "content": "[14] Nikita Dvornik, Cordelia Schmid, and Julien Mairal. Selecting relevant features from a universal representation for few-shot classification. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.785, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[15] Linus Ericsson, Henry Gouk, and Timothy M. Hospedales. Why do self-supervised models transfer? Investigating the impact of invariance on downstream tasks. BMVC, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[16] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[17] Mario Fritz, E. Hayman, B. Caputo, and J. Eklundh. THE KTH-TIPS database. 2004. 12" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[18] Geetharamani G. and Arun Pandian J. Identification of plant leaf diseases using a nine-layer deep convolutional neural network. Computers and Electrical Engineering, 76:323-338, 2019. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[19] Ning Gao, Hanna Ziesche, Ngo Anh Vien, Michael Volpp, and Gerhard Neumann. What matters for meta-learning vision regression tasks? In CVPR, 2022. 2, 4, 5, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[20] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[21] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[22] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[24] Emily F. Brownlee Heidi M. Sosik, Emily E. Peacock. Annotated plankton images - data set for developing and evaluating classification methods., 2015. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[25] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[26] Timothy M Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[27] Shell Hu, Da Li, Jan Stuehmer, Minyoung Kim, and Timothy Hospedales. Pushing the limits of simple pipelines for few-shot learning. In CVPR, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE Workshop on 3D Representation and Recognition (3dRR-13), 2013. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[29] Gustaf Kylberg. The Kylberg texture dataset v. 1.0. Technical report, 2011. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[30] Brenden M. Lake, Ruslan Salakhutdinov, and Joshua B. Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[31] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. A sparse texture representation using local affine regions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(8):1265-1278, Aug. 2005. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[32] Da Li, Henry Gouk, and Timothy Hospedales. Finding lost DG: Explaining domain generalization via model complexity. In arXiv, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[33] Liam Li, Kevin Jamieson, Afshin Rostamizadeh, Ekaterina Gonina, Moritz Hardt Hardt, Benjamin Recht, and Ameet Talwalkar. A system for massively parallel hyperparameter tuning. In MLSys, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] Lisha Li, Kevin G. Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter optimization. JMLR, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7701" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "[35] Rui Li, Ondrej Bohdal, Rajesh K. Mishra, Hyeji Kim, Da Li, Nicholas D. Lane, and Timothy M. Hospedales. A channel coding benchmark for meta-learning. 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[36] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.47, + 0.221 + ], + "angle": 0, + "content": "[37] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. 1, 2, 3, 4, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.47, + 0.264 + ], + "angle": 0, + "content": "[38] Yanbin Liu, Juho Lee, Linchao Zhu, Ling Chen, Humphrey Shi, and Yi Yang. A multi-mode modulator for multi-domain few-shot classification. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.47, + 0.307 + ], + "angle": 0, + "content": "[39] Changsheng Lu and Piotr Koniusz. Few-shot keypoint detection with uncertainty learning for unseen species. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.47, + 0.351 + ], + "angle": 0, + "content": "[40] P. Mallikarjuna, Alireza Tavakoli Targhi, Mario Fritz, E. Hayman, B. Caputo, and J. Eklundh. The KTH-TIPS 2 database. 2006. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[41] Teresa Mendonça, Pedro M. Ferreira, Jorge S. Marques, André R. S. Marcal, and Jorge Rozeira. PH2 - a dermoscopic image database for research and benchmarking. In Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2013. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.424, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[42] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[43] Ari Morcos, Maithra Raghu, and Samy Bengio. Insights on representational similarity in neural networks with canonical correlation. Advances in Neural Information Processing Systems, 31, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.47, + 0.565 + ], + "angle": 0, + "content": "[44] Jiteng Mu, Weichao Qiu, Gregory D. Hager, and Alan L. Yuille. Learning from synthetic animals. In CVPR, 2020. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.47, + 0.609 + ], + "angle": 0, + "content": "[45] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. Unsupervised domain adaptation: A reality check. In arXiv, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.612, + 0.47, + 0.667 + ], + "angle": 0, + "content": "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.67, + 0.47, + 0.737 + ], + "angle": 0, + "content": "[47] Yoshihiko Ozaki, Yuki Tanigaki, Shuhei Watanabe, and Masaki Onishi. Multiobjective tree-structured parzen estimator for computationally expensive optimization problems. In Genetic and Evolutionary Computation Conference, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.741, + 0.47, + 0.768 + ], + "angle": 0, + "content": "[48] Eunbyung Park and Junier B. Oliva. Meta-Curvature. In NeurIPS, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.47, + 0.825 + ], + "angle": 0, + "content": "[49] Massimiliano Patacchiola, Jack Turner, Elliot J Crowley, Michael O'Boyle, and Amos J Storkey. Bayesian meta-learning for the few-shot setting via deep kernels. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.828, + 0.425, + 0.842 + ], + "angle": 0, + "content": "[50] Gerald Piosenko. 100 sports image classification. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.844, + 0.47, + 0.869 + ], + "angle": 0, + "content": "[51] Gerald Piosenko. Birds 400 - species image classification. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[53] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? Towards understanding the effectiveness of maml. In ICLR, 2020. 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[54] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[55] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A generalist agent, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[56] Mengye Ren, Eleni Triantafillou, Jake Snell Sachin Ravi, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[57] Sebastian Ruder. An overview of multi-task learning in deep neural networks. In arXiv, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[58] Amirreza Shaban, Shray Bansal, Z. Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. In BMVC, 2017. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[59] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[60] Haozhe Sun, Wei-Wei Tu, and Isabelle M Guyon. Omniprint: A configurable printed character synthesizer. In NeurIPS Datasets and Benchmarks Track, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[61] Sebastian Thrun and Lorien Pratt, editors. Learning to Learn. Kluwer Academic Publishers, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[62] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In ECCV, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[63] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, and Hugo Larochelle. Meta-dataset: A dataset of datasets for learning to learn from few examples. In ICLR, 2020. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[64] Yu-Yun Tseng, Alexander Bell, and Danna Gurari. Vizwiz-fewshot: Locating objects in images taken by people with visual impairments. In ECCV, 2022. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[65] Ihsan Ullah, Dustin Carrion, Sergio Escalera, Isabelle Guyon, Mike Huisman, Felix Mohr, Jan N. van Rijn, Haozhe Sun, Joaquin Vanschoren, and Phan Anh Vu. Meta-album: Multi-domain meta-dataset for few-shot image classification. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2022. 1, 2, 3, 4, 5, 6, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[66] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NIPS, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7702" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[67] Risto Vuorio, Shao-Hua Sun, Hexiang Hu, and Joseph J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. In NeurIPS, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[68] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[69] Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.218, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[70] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.273, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[71] Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darryll. What should not be contrastive in contrastive learning. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.315, + 0.469, + 0.355 + ], + "angle": 0, + "content": "[72] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.356, + 0.469, + 0.411 + ], + "angle": 0, + "content": "[73] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.412, + 0.469, + 0.452 + ], + "angle": 0, + "content": "[74] Yongxin Yang and Timothy Hospedales. A unified perspective on multi-domain and multi-task learning. In ICLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.453, + 0.469, + 0.494 + ], + "angle": 0, + "content": "[75] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.495, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[76] Mingzhang Yin, George Tucker, Mingyuan Zhou, Sergey Levine, and Chelsea Finn. Meta-learning without memorization. 2020. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.536, + 0.469, + 0.578 + ], + "angle": 0, + "content": "[77] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In NeurIPS, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.579, + 0.469, + 0.633 + ], + "angle": 0, + "content": "[78] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In CORL, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.634, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[79] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.676, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[80] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. In arXiv, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[81] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.8, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[82] Linyu Zheng, Ming Tang, Yingying Chen, Jinqiao Wang, and Hanqing Lu. Learning feature embeddings for discriminant model based tracking. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.842, + 0.469, + 0.897 + ], + "angle": 0, + "content": "[83] Bartosz Zielinski, Anna Plichta, Krzysztof Misztal, Przemysław Spurek, Monika Brzychczy-Wloch, and Dorota Ochońska. Deep learning approach to bacterial colony classification. PLOS ONE, 12(9):1-14, 09 2017. 12" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7703" + } + ] +] \ No newline at end of file diff --git a/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_origin.pdf b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec00a855e78404467afc414d0b00ba5352decf87 --- /dev/null +++ b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/6a3d596b-7e86-4a45-b2b0-ce1ada0b267e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cad5e962f8b0cf8a98b692c4c052b34556a194e4a47027a4b36c8473fd45af20 +size 461827 diff --git a/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/full.md b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/full.md new file mode 100644 index 0000000000000000000000000000000000000000..20ddf754e3c09f5bb934b0642b4cfa251dc34b5d --- /dev/null +++ b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/full.md @@ -0,0 +1,281 @@ +# Meta Omnium: A Benchmark for General-Purpose Learning-to-Learn + +Ondrej Bohdal $^{1,\ast}$ , Yinbing Tian $^{2,\ast}$ , Yongshuo Zong $^{1}$ , Ruchika Chavhan $^{1}$ , Da Li $^{3}$ , Henry Gouk $^{1}$ , Li Guo $^{2}$ , Timothy Hospedales $^{1,3}$ + +1The University of Edinburgh 2Beijing University of Posts and Telecommunications 3Samsung AI Center, Cambridge *Joint-First Authors + +# Abstract + +Meta-learning and other approaches to few-shot learning are widely studied for image recognition, and are increasingly applied to other vision tasks such as pose estimation and dense prediction. This naturally raises the question of whether there is any few-shot meta-learning algorithm capable of generalizing across these diverse task types? To support the community in answering this question, we introduce Meta Omnium, a dataset-of-datasets spanning multiple vision tasks including recognition, keypoint localization, semantic segmentation and regression. We experiment with popular few-shot meta-learning baselines and analyze their ability to generalize across tasks and to transfer knowledge between them. Meta Omnium enables meta-learning researchers to evaluate model generalization to a much wider array of tasks than previously possible, and provides a single framework for evaluating meta-learners across a wide suite of vision applications in a consistent manner. Code and dataset are available at https://github.com/edi-meta-learning/meta-omnium. + +# 1. Introduction + +Meta-learning is a long-standing research area that aims to replicate the human ability to learn from a few examples by learning-to-learn from a large number of learning problems [61]. This area has become increasingly important recently, as a paradigm with the potential to break the data bottleneck of traditional supervised learning [26, 70]. While the largest body of work is applied to image recognition, few-shot learning algorithms have now been studied in most corners of computer vision, from semantic segmentation [37] to pose estimation [49] and beyond. Nevertheless, most of these applications of few-shot learning are advancing independently, with increasingly divergent application-specific methods and benchmarks. This makes it hard to evaluate whether few-shot meta-learners can solve diverse vision tasks. Importantly it also discourages the de + +![](images/a1ca96f2b4f105019e810b99833709cf85445be2e6e6c46008e776bb05d4d7f1.jpg) +Figure 1. Illustration of the diverse visual domains and task types in Meta Omnium. Meta-learners are required to generalize across multiple task types, multiple datasets, and held-out datasets. + +velopment of meta-learners with the ability to learn-to-learn across tasks, transferring knowledge from, e.g., keypoint localization to segmentation – a capability that would be highly valuable for vision systems if achieved. + +The overall trend in computer vision [20, 52] and AI [5, 55] more generally is towards more general-purpose models and algorithms that support many tasks and ideally leverage synergies across them. However, it has not yet been possible to explore this trend in meta-learning, due to the lack of few-shot benchmarks spanning multiple tasks. State-of-the-art benchmarks [63, 65] for visual few-shot learning are restricted to image recognition across a handful of visual domains. There is no few-shot benchmark + +
DatasetNum TasksNum DomainsNum ImagesCategoriesSizeLightweightMulti-TaskMulti-Domain
Omniglot [30]1132K1623148MB✓✗✗
miniImageNet [66]1160K1001GB✓✗✗
Meta-Dataset [63]17~1053M43~1500210GB✗✗✓
VTAB [80]13~192.2M2~397100GB✗✗✓
FSS1000 [37]11100001000670MB✓✗✗
Meta-Album [65]110~401.5M19~70615GB✓✗✓
Meta Omnium421160K2~7063.1GB✓✓✓
+ +Table 1. Feature comparison between Meta Omnium and other few-shot meta-learning benchmarks. Meta Omnium uniquely combines a rich set of tasks and visual domains with a lightweight size for accessible use. + +that poses the more substantial challenge [57, 77] of generalizing across different tasks. We remark that the term task is used differently in few-shot meta-learning literature [16, 26, 70] (to mean different image recognition problems, such as cat vs dog or car vs bus) and the multi-task literature [20, 57, 74, 77] (to mean different kinds of image understanding problems, such as classification vs segmentation). In this paper, we will use the term task in the multi-task literature sense, and the term episode to refer to tasks in the meta-learning literature sense, corresponding to a support and query set. + +We introduce Meta Omnium, a dataset-of-datasets spanning multiple vision tasks including recognition, semantic segmentation, keypoint localization/pose estimation, and regression as illustrated in Figure 1. Specifically, Meta Omnium provides the following important contributions: (1) Existing benchmarks only test the ability of meta-learners to learn-to-learn within tasks such as classification [63, 65], or dense prediction [37]. Meta Omnium uniquely tests the ability of meta-learners to learn across multiple task types. (2) Meta Omnium covers multiple visual domains (from natural to medical and industrial images). (3) Meta Omnium provides the ability to thoroughly evaluate both in-distribution and out-of-distribution generalisation. (4) Meta Omnium has a clear hyper-parameter tuning (HPO) and model selection protocol, to facilitate future fair comparison across current and future meta-learning algorithms. (5), Unlike popular predecessors, [63], and despite the diversity of tasks, Meta Omnium has been carefully designed to be of moderate computational cost, making it accessible for research in modestly-resourced universities as well as large institutions. Table 1 compares Meta Omnium to other relevant meta-learning datasets. + +We expect Meta Omnium to advance the field by encouraging the development of meta-learning algorithms capable of knowledge transfer across different tasks – as well as across learning episodes within individual tasks as is popularly studied today [16, 70]. In this regard, it provides the next step of the level of a currently topical challenge of dealing with heterogeneity in meta-learning [1, 35, 63, 67]. While existing benchmarks have tested multi-domain heterogeneity (e.g., recognition of written characters and plants + +within a single network) [63, 65] and shown it to be challenging, Meta Omnium tests multi-task learning (e.g., character recognition vs plant segmentation). This is substantially more ambitious when considered from the perspective of common representation learning. For example, a representation tuned for recognition might benefit from rotation invariance, while one tuned for segmentation might benefit from rotation equivariance [11, 15, 71]. Thus, in contrast to conventional within-task meta-learning benchmarks that have been criticized as relying more on common representation learning than learning-to-learn [53, 62], Meta Omnium better tests the ability of learning-to-learn since the constituent tasks require more diverse representations. + +# 2. Related Work + +# 2.1. Meta-learning Benchmarks + +The classic datasets in few-shot meta-learning for computer vision are Omniglot [30] and miniImageNet [66]. Later work criticized these for having insufficient task (episode) diversity and tieredImageNet [56] used the class hierarchy of ImageNet to enforce more diversity between meta-train and meta-test episodes. The main contemporary benchmarks are CD-FSL [22], which challenges few-shot learners to generalize to new visual domains; and Meta-Dataset [63] and Meta-Album [65], which go further in requiring few-shot learners to learn from a mixture of visual domains. Such multi-domain heterogeneous meta-learning turns out to be challenging. A related benchmark to Meta-Dataset is VTAB [80], which similarly provides multiple domains for evaluating data-efficient visual recognition, but their focus is on evaluating representation transfer from large-scale pre-training rather than learning-to-learn and meta-learning. VTAB+MD [13] compare representation transfer and meta-learning approaches on the Meta-Dataset tasks. However, none of these benchmarks address multi-task meta-learning as considered here (Figure 1). + +Outside of recognition, task-specific few-shot benchmarks have been proposed in vision problems of semantic segmentation [37], regression [19], pose/keypoint estimation [73], etc. These are mostly slightly behind the complexity of the recognition benchmarks with regards to being + +single-domain, with the exception of [73]. With regard to multi-task meta-learning as considered here, the only existing benchmark is meta-world [78], which is specific to robotics and reinforcement learning rather than vision. + +We also mention taskonomy [79] as a popular dataset that has been used for multi-task learning. However, it is not widely used for few-shot meta-learning. This is because, although taskonomy has many tasks, unlike the main meta-learning benchmarks [37, 63], there are not enough visual concepts within each task to provide a large number of concepts for meta-training and a disjoint set of concepts to evaluate few-shot learning for meta-validation and meta-testing. + +# 2.2. Heterogeneity in Meta-Learning + +There are several sophisticated methods in the literature that highlighted the challenge of the task to address heterogeneity in meta-learning. These have gone under various names such as multi-modal meta-learning [1, 38, 67] – in the sense of multi-modal probability distribution (over tasks/episodes). However, with the exception of [38], these have mostly not been shown to scale to the main multimodal benchmarks such as Meta-Dataset [63]. A more common approach to achieving high performance across multiple heterogeneous domains such as those in Meta-Dataset is to train an ensemble of feature extractors across available training domains, and fuse them during metatesting [14, 36, 36]. However, this obviously incurs a substantial additional cost of maintaining a model ensemble. In our evaluation, we focus on the simpler meta-learners that have been shown to work in challenging multi-domain learning scenarios [63, 65], while leaving sophisticated algorithmic and ensemble-based approaches for future researchers to evaluate on the benchmark. + +# 3. Meta Omnium Benchmark and Datasets + +# 3.1. Motivation and Guiding Principles + +We first explain the motivating goals and guiding principles behind the design of Meta Omnium. The goal is to build a benchmark for multi-task meta-learning that will: (i) Encourage the community to develop meta-learners that are flexible enough to deal with greater task heterogeneity than before, and thus are more likely to be useful in practice with less curated episode distributions. This was identified as a major challenge in the discussion arising in several recent meta-learning and computer vision workshops and challenges1. (ii) Ultimately progress on this benchmark should provide practical improvements in data-efficient learning + +for computer vision through the development of methods that can better transfer across different task types. + +In developing this benchmark, we established a few principles that we used to guide design choices. These include: (i) The benchmark should be lightweight in terms of storage and computing, making it accessible to a broad range of researchers and not only large corporations. (ii) The benchmark should cover multiple tasks with heterogeneous output spaces (as opposed to all classification, all regression, or all dense prediction), as well as multiple visual domains. In these regards, Meta Omnium is compared to alternatives in Table 1. (iii) The initial baselines should have only minimal task-specific decoders. This is in contrast to the state of the art within various sub-disciplines of FSL such as segmentation [25, 42], keypoint [39, 72], and classification [2, 75] where specially designed decoders are often used. This is to evaluate and encourage future research on learning-to-learn across tasks, rather than primarily benchmarking how well we can manually engineer prior knowledge of optimal task-specific decoders. While we are not opposed to future competitors on this benchmark developing task-specific decoders, these should be evaluated separately against the minimal-decoder competitors. (iv) The benchmark should provide distinct datasets for in-distribution (ID) training and out-of-distribution (OOD) evaluation, to evaluate the robustness of the distribution-shift. This is already provided by [63, 65] for classification, and we extend such an ID and OOD dataset ensemble to multiple tasks. Figure 2 illustrates our dataset and task-split. (v) The benchmark should provide a clear hyper-parameter tuning protocol. With a number of recent studies showing that hyper-parameter tuning can dominate other effects of interest in computer vision [21, 32, 45], this is important for a future-proof meta-learning benchmark. This is also related to the first cost point (i) above: Only for a benchmark with a modest cost can most institutions realistically expect to conduct hyperparameter tuning. We provide the hyper-parameter tuning protocol. (vi) Finally, following the debate in [35, 53, 62] as to the value of meta-learning vs conventional transfer learning, the dataset should support both episodic meta-learning and conventional transfer learning approaches. + +# 3.2. Data Splits and Tasks + +For each main task (classification, segmentation, key-point localization), we split the datasets into seen datasets available for meta-training, and unseen datasets that are completely held out for out-of-domain meta-validation and meta-testing. Similarly to [63, 65], for the seen datasets, we construct category-wise splits into meta-train/val/test. While for the unseen datasets, there is no category-wise split as episodes from all categories from the whole dataset will be used for validation and testing respectively. The overall split organization is illustrated in Figure 2. We + +![](images/a47e27c77a1999e1448109dd61b535451034f1f2c62b83d9a0d91c8bc3dbbded.jpg) +Figure 2. Schematic of benchmark and dataset splits. For each task, there are multiple datasets, which are divided into seen (solid border) and unseen (dash border) datasets. The seen (ID) datasets are divided class-wise into meta-train/meta-val/meta-test splits. The unseen datasets are held out for out-of-distribution (OD) evaluation. Meta-training is conducted on the ID-meta-train split of the seen datasets (blue). Models are validated on ID validation class splits, or OOD validation datasets (green). Results are reported on the ID test class splits and OOD test datasets (orange). We also hold out an entire task family (regression) for evaluating novel task generalisation. + +additionally have a completely held-out task: regression. Datasets from this task are not used during meta-training. + +Our multi-task setup enables us to define and compare two training protocols: Single-task meta-learning which evaluates how well meta-learning performs when trained and tested within a particular task family (within each plate in Figure 2); and Multi-task meta-learning which evaluates how well meta-learning performs when trained across all available task families (across all plates in Figure 2). + +With this organization we can separately evaluate: Within-distribution generalization (ID): How well do meta-learners generalize to novel test concepts within the seen datasets?; and Out-distribution generalization (OOD): How well do meta-learners generalize to novel concepts in unseen datasets? + +We provide two sources of validation data: ID and OOD, and our models are selected based on the combined performance across both. OOD validation is not supposed by the most popular Meta-Dataset benchmark [63] as despite its larger size it does not provide OOD validation datasets. + +# 3.3. Datasets and Metrics + +Given the considerations in Section 3.1, our benchmark consists of three main tasks (classification, segmentation, keypoints/pose) and one held-out task (regression). + +Classification For classification we take the 10 datasets from the initial public release of Meta-Album [65]. These + +images are all $128 \times 128$ and contain 19-706 classes per dataset, with 40 images per class. Three of these datasets are reserved for out-of-distribution meta-validation, and four for out-of-distribution meta-test. + +Segmentation For segmentation, we take FSS1000 [37] for in-distribution (10,000 images, 1,000 classes), and combine it with VizWiz [64] for OOD meta-validation (862 images, 22 classes), and modified Pascal5i [58] (7,242 images, 6 classes) and the very distinct medical imaging dataset PH2 [41] (200 images, 3 classes) for meta-testing. The segmentation images originally were of diverse sizes. We resize them all to $224 \times 224$ for Meta Omnium. Note that VizWiz and Pascal datasets originally contain more classes and images. We exclude the classes that overlap with that in the FSS1000 dataset for few-shot learning, and thus there are no classes overlapping among all the datasets. + +Keypoints For keypoints/pose, we take animal-pose [10] for in-distribution, synthetic animal-pose [44] for OOD meta-validation, and MPII human-pose [4] for OOD meta-testing. All images are resized to $128 \times 128$ . MPII includes about $40\mathrm{k}$ people in over $25\mathrm{k}$ images with annotated body keypoints. Animal Pose includes 5 animal categories for 6K instances in over $4\mathrm{k}$ images. Each animal is cropped from the original image. We keep cats and dogs for training, horses and sheep for in-domain validation, and cow for indomain testing. Synthetic animal pose generates synthetic images using animal CAD models rendered from various viewpoints and lightings on a random background. We keep only the horse and tiger categories in our final datasets. + +Regression For evaluating regression as a held-out task, we use four datasets corresponding to the test splits of [19]: ShapeNet1D, ShapeNet2D, Distractor and Pascal1D [76]. All images are resized to $128 \times 128$ . ShapeNet1D aims to predict azimuth angles. It contains 30 categories in total and we keep the 3 categories from the test set. ShapeNet2D further includes 2D rotation with azimuth angles and elevation. The test set of ShapeNet2D contains 300 categories in total with 30 images per category. Distractor aims to predict the position of a target object in the presence of a distractor. It contains 12 categories in total and the test set has 2 categories. Each category contains 1000 objects with 36 images for each. Pascal1D aims to predict azimuth angle. The whole Pascal1D contains 65 objects from 10 categories. The test set contains 15 objects with 100 images for each object. The supplementary material provides full details of all datasets and splits. + +# 3.4. Training API + +For episodic learning, we proceed by (i) sampling a task, (ii) sampling a dataset, (iii) sampling an episode. Under our main protocol we consider variable 1 to 5-shot evaluation, but also evaluate separate 1 and 5 shot settings (training is always done with a variable number of shots - sup + +port examples). For classification tasks, we follow [65] in generating 5-way episodes. For segmentation tasks, we follow [37] in considering each episode to be a binary foreground/background classification problem for a novel class and generate 2-way episodes. For keypoint, we form episodes by randomly selecting a class (e.g., animal category) and then randomly selecting a subset of 5 keypoints to localize for each episode. For regression tasks, we generate variable 5 to 25-shot episodes because it is a common practice to use more shots for regression tasks [19]. + +For non-episodic/transfer learning, we provide access to the meta-train portions of the seen datasets in conventional mini-batches for conventional single task and multi-task supervised learning. + +Evaluation Metrics For classification tasks, we use standard top-1 accuracy; for segmentation tasks, we use standard mean intersection-over-union (mIoU) that averages over IoU values of all object classes; for keypoint prediction, we report the Percentage of Correct Keypoints (PCK). In detail, a detected joint is considered correct if the distance between the predicted and the true joint is within a certain threshold. In our experiments, the threshold is 0.01 for normalized value, which stands for about 12.8 pixel of input image resolution. For regression tasks, we follow [19] and use the same metrics. + +# 3.5. Architecture and Baseline Competitors + +As discussed in Section 3.1, we aim to establish baselines that can be adapted to tasks with heterogeneous outputs, with minimal reliance on task-specific decoders. We follow [63, 65] in using a ResNet-18 CNN [23] as a feature extractor architecture. For recognition tasks, we perform multi-class classification immediately after ResNet's GAP. For regression tasks, we perform linear regression directly after the ResNet's GAP. For keypoint tasks, we consider them to be a regression problem from the feature map to the keypoint location. For segmentation tasks, we use a simplified PSPNet-like [81] strategy. We concatenate the extracted feature maps from ResNet's feature pyramid, with upsampling where appropriate, to generate a feature map of size $w \times h$ , and then do pixel-wise classification with $1 \times 1$ convolutional layer to obtain the final segmentation map. All tasks thus have only one learnable weight as a minimal classifier/decoder after the common ResNet feature encoder. Based on this common encoder and minimal decoder architecture, we describe our meta-learning baselines. Prototypical Network [59] is a classic meta-learner that exploits nearest-centroid metric learning for few-shot classification. ProtoNets were adapted to segmentation tasks in PANet [68], by performing pixel-level feature matching between support prototypes and query pixels. We use the same principle together with the PSPNet-like features described earlier. To generalize ProtoNets to regression tasks such as + +keypoint prediction, we must relax the prototype assumption, and use them as simple Gaussian kernel-regression models [7]. Specifically, we generate a feature embedding for each support example, and then for query examples, we calculate the negative exponential distance to the support examples, and use this inverse distance-weighted sum of support set labels as the prediction. Thus for regression tasks with a support set $S = \{(x_i,y_i)\}$ and query example $x_{q}$ , ProtoRegression predicts + +$$ +f (y _ {q} | x _ {q}, S) \propto \sum_ {i} y _ {i} \exp (- \tau (f _ {\theta} (x _ {i}) - f _ {\theta} (x _ {q})) ^ {2}) \quad (1) +$$ + +enabling us to learn deep feature $f_{\theta}$ in the usual episodic meta-learning way. We use cross-entropy loss for classification and segmentation, and MAE loss for regression tasks. + +DDRR Deep differentiable ridge-regression has been considered for few-shot recognition [6], tracking [82], and other tasks. It is related to ProtoNet in that the feature is not adapted after the meta-train stage, but different in that the decoder/classifier layer is learned by differentiable ridge-regression rather than nearest centroid or kernel regression. An elegant property of DDRR methods is that they naturally address regression tasks, although they have been repurposed for classification [6] by conducting MSE-loss regression to a target 1-hot vector. Thus they are a natural choice for our benchmark. For application to segmentation, we apply DDRR in a $1 \times 1$ convolution-like way, to perform pixel classification for the output mask with a DDRR classifier at each pixel. Further, we calibrate the prediction for binary cross entropy loss with a learnable scale and bias following [6]. DDRR uses MAE loss for only regression tasks and use MSE loss for other tasks. + +MAML The seminal few-shot meta-learner MAML [16] aims to learn an initial condition for per-episode gradient-descent. MAML is straightforward to adapt to different types of tasks. Based on each episode's support set, a new output layer is learned, and the feature extractor is updated, both by a few steps of gradient descent. Similarly to Meta-Dataset [63], we do not learn an initialization for the output layer, since it can change size between episodes drawn from multiple tasks. To alleviate this challenge, we also follow Meta-Dataset in evaluating Proto-MAML – a variant that initializes the MAML output layer based on the linear classifier/regressor suggested by nearest-centroid prior to gradient descent. Going beyond this, to adapt Proto-MAML to regression tasks, we also initialize the output layer based on the ridge-regression solution to the support set. + +Meta-Curvature Meta-Curvature [48] is an enhancement of MAML that learns a pre-conditioning matrix to improve inner-loop adaptation, as well as an initial condition as in standard MAML. Meta-Curvature outperforms MAML in simpler single-task few-shot benchmarks. + +Transfer Learning We also consider standard supervised + +learning on the meta-train tasks for transfer to the target tasks, a strategy reported to be competitive with meta-learning [62]. For adaptation, we explore both linear readout [62, 69] and fine-tuning [65]. Besides learning a new output layer from scratch, we also consider a fine-tuning version that initializes the classifier weights using class prototypes (recognition/segmentation) or ridge regression weights (keypoints/regression), inspired by Proto-MAML. Train-from-Scratch (TFS) We lastly consider training each episode from scratch using only the support set [65]. + +# 3.6. Hyperparameter Optimization + +As part of our benchmark, we perform hyperparameter optimization (HPO) to ensure we select appropriate hyperparameters for the diverse tasks that we consider. Multi-objective HPO under restricted resources is challenging, so we devise the following HPO protocol: estimate the performance of each candidate configuration on a lower fidelity (lower number of iterations) and then identify the configuration that works the best across all validation datasets considered (combination of in-domain and out-domain datasets, across various task families). Note that fast multi-fidelity methods such as Hyperband [34], ASHA [33] or PASHA [8] are not applicable out of the box in our multi-objective setup, so we decided to train each candidate configuration using fixed 5,000 training iterations. Since different tasks/datasets are of different difficulties (and use different metrics), we normalize the score of each configuration for each validation dataset by the best score for that dataset across all candidate configurations. We then select the configuration with the best average normalized score. + +Note that due to resource constraints we are only able to sample a relatively smaller number of candidates (30), so we utilize a sample efficient state-of-the-art Multi-Objective TPE method [47], available from the Optuna library [3]. We perform HPO for multi-task and single-tasks setups separately, so single task classification, segmentation and keypoint estimation have their own set of hyperparameters; and the multi-task case has its own set. The hyperparameters include the meta-learning rate and optimizer, momentum, and various method-specific hyperparameters (full details are in the supplementary). Once the hyperparameters are chosen, we perform standard training of the model for the full number of iterations. + +# 4. Experiments + +In this section, we aim to use our benchmark to answer the following questions: (1) Which meta-learner performs best on average across a heterogenous range of tasks? Existing benchmarks have evaluated meta-learners for one task at a time, we now use our common evaluation platform to find out if any meta-learner can provide general-purpose learning to learn across different task types, or whether + +each task type prefers a different learner. Similarly, we can ask which meta-learner is most robust to out-of-distribution tasks? (2) Having defined the first multi-task meta-learning benchmark, and generalizations of seminal meta-learners to different kinds of output spaces, we ask which meta-learner performs best for multi-task meta-learning? More generally, is there a trend in gradient-based vs metric-based meta-learner success? (3) Does multi-task meta-learning improve or worsen performance compared to single-task? The former obviously provides more meta-training data, which should be advantageous, but the increased heterogeneity across meta-training episodes in the multi-task case also makes it harder to learn [63, 67]. (4) How does meta-learning perform compared to simple transfer learning, or learning from scratch? + +# 4.1. Experimental Settings + +We train each meta-learner for 30,000 meta-training iterations, with meta-validation after every 2,500 iterations (used for checkpoint selection). For evaluation during meta-testing we use 600 tasks for each corresponding dataset, and for meta-validation we use 1200 tasks together. We use random seeds to ensure that the same tasks are used across all methods that we compare. For transfer learning approaches (fine-tuning, training from scratch, etc.), we use 20 update steps during evaluation. We only retain the meta-learned shared feature extractor across tasks, and for each new evaluation task, we randomly initialize the output layer so that we can support any number of classes as well as novel task families during meta-testing (in line with [65]). + +# 4.2. Results + +The main experimental results are shown in Table 2, where rows correspond to different few-shot learners, and columns report the average performance of test episodes, aggregated across multiple datasets in each task family, and broken down by "seen" datasets (ID) and "unseen" datasets (OOD). The table also reports the average rank of each meta-learner across each dataset, both overall and broken down by ID and OOD datasets. More specifically, for each setting (e.g. cls. ID) we calculate the rank of each method (separately for STL and MTL), and then we average those ranks across cls., seg. and keypoints. From the results, we can draw the following main conclusions: + +(1) ProtoNet is the most versatile meta-learner, as shown by its highest average rank in the single-task scenario. This validates our novel Kernel Regression extension of ProtoNet for tackling regression-type keypoint localization tasks. Somewhat surprisingly, ProtoNet is also the most robust to out-of-distribution episodes (OOD) which is different from the conclusion of [63] and others who suggested that gradient-based adaptation is crucial to adapt to OOD data. However, it is also in line with the results of [65] and + +
ClassificationSegmentationKeypointsAverage Rank
IDOODIDOODIDOODIDOODAVG
Single-TaskMAML58.761.654.742.125.433.04.33.33.8
Proto-MAML50.549.746.444.123.622.56.06.36.2
Meta-Curvature64.861.465.649.843.516.02.04.33.2
ProtoNet70.459.475.857.227.833.31.31.71.5
DDRR63.158.766.748.020.531.94.73.74.2
Proto-FineTuning50.850.760.043.421.333.15.34.34.8
FineTuning42.348.250.540.025.730.05.76.76.2
Linear-Readout48.653.434.022.722.126.97.36.77.0
TFS31.542.042.837.621.026.08.38.08.2
Multi-TaskMAML59.158.543.337.424.323.92.74.73.7
Proto-MAML58.563.753.043.221.633.33.01.72.3
Meta-Curvature70.466.942.634.518.225.34.34.74.5
ProtoNet65.958.863.349.720.133.02.72.02.3
DDRR52.851.940.437.322.830.15.04.74.8
Proto-FineTuning52.453.244.837.821.230.04.34.04.2
FineTuning44.151.241.336.118.120.57.77.07.3
Linear-Readout46.050.941.532.619.923.56.38.07.2
TFS21.923.838.735.814.111.09.08.38.7
+ +the strong performance of prototypes more broadly [9]. + +(2) Coming to multi-task meta-learning the situation is similar in that ProtoNet dominates the other competitors, but now sharing the first place with Proto-MAML. +(3) To compare single-task and multi-task meta-learning (top and bottom blocks of Table 2) more easily, Figure 3 shows the difference in meta-testing episode performance after STL and MTL meta-training for each method. Overall STL outperforms the MTL condition, showing that the difficulty of learning from heterogeneous tasks [57,77] outweighs the benefit of the extra available multi-task data. +(4) Finally, comparing meta-learning methods with simple transfer learning methods as discussed in [13, 62], the best meta-learners are clearly better than transfer learning for both single and multi-task scenarios. + +We also note that Proto-MAML is better than MAML in the multi-task case, likely due to the importance of a good output-layer initialization in the case of heterogeneous episodes, as per [63]. Meta-Curvature outperforms MAML in single-task in-domain scenarios, in line with previous results [48], but it did not achieve stronger performance out-of-domain or in the multi-task case. Finally, while DDRR is perhaps the most elegant baseline in terms of most naturally spanning all task types, its overall performance is middling. + +# 4.3. Additional Analysis + +How well can multi-task meta-learners generalize to completely new held-out tasks? We take the multi-task meta-learners (trained on classification, segmentation, keypoints) and evaluate them on four regression benchmarks inspired by [19]: ShapeNet1D, ShapeNet2D, Distractor and Pascal1D. Because the metrics differ across datasets, we analyse the rankings and summarise the results in Table 3. We see the basic TFS performs the worst, with MAML, ProtoNets, DDRR being the best. However, in several cases the results were not better than predicting the mean (full results in the appendix), showing that learning-to-learn of completely new task families is an open challenge. + +Table 2. Main Results. Results are presented as averages across the datasets within each task type and separately for in-distribution (ID) and out-of-distribution (OOD) datasets. Classification, segmentation, and keypoint results are reported in accuracy $(\%)$ , mIOU $(\%)$ , and PCK $(\%)$ respectively. The upper and lower groups correspond to multi-task and single-task meta-training prior to evaluation on the same set of meta-testing episodes. Upper and lower sub row groups correspond to meta-learners and non-meta learners respectively. See the appendix for a full breakdown over individual datasets. + +
MAMLPMAMLMCPNDDRRPFTFTLRTFS
3.36.54.83.53.53.85.84.38.5
+ +Table 3. Average ranking of the different methods across four out-of-task regression datasets. + +How much does external pre-training help? Our focus is on assessing the efficacy of meta-learning rather than representation transfer, but we also aim to support researchers investigating the impact of representation learning on external data prior to meta-learning [13, 27, 80]. We therefore + +![](images/6d0fefa6ee20d4b2272b0d0732322bc1aed07e4d4dcad11f9ea04e6f87149ece.jpg) +Figure 3. Analysis of the differences in scores between single-task (STL) and multi-task (MTL) learning for different methods. + +
MethodPretrainCls.Segm.Keyp.
IDOODIDOODIDOOD
Proto-MAMLX58.563.753.043.221.633.3
ProtoNetX66.058.863.349.720.133.0
Proto-MAML✓63.962.756.245.321.833.3
ProtoNet✓63.558.662.049.020.133.1
+ +Table 4. Analysis of the impact of external data pretraining for selected meta-learners in multi-task learning condition. The results show that ImageNet pretraining does not necessarily help improve performance. Cls., Segm., Keyp. represent classification, segmentation and keypoint respectively. + +specify evaluation conditions where external data outside our defined in-distribution meta-training set is allowed. + +We take two high-performing approaches in the multi-task scenario, Proto-MAML and ProtoNet, and we investigate to what extent external pre-training helps. We use the standard ImageNet1k pre-trained ResNet18, prior to conducting our meta-learning pipeline as usual. We use the same hyperparameters as selected earlier for these models to ensure consistent evaluation, and ensure that the differences in performance are not due to a better selection of hyperparameters. The results in Table 4 show that pretraining is not necessarily helpful in the considered multi-task setting, in contrast to purely recognition-focused evaluations [13,27,80], which were unambiguously positive about representation transfer from external data. + +Analysis of Runtimes We analyze the times that the different meta-learning approaches spend on meta-training, meta-validation and meta-testing in the multitask learning case of our benchmark. The results in Table 5 show that all experiments are relatively lightweight, despite the ambitious goal of our benchmark to learn a meta-learner that can generalize across various task families. Most notably we observe that ProtoNets are the fastest approach, alongside being the best-performing one. Note that fine-tuning and training from scratch are expensive during the test time as they use backpropagation with a larger number of steps. + +Discussion and Future Work In future Meta Omnium can be used in a variety of ways beyond benchmarking + +
MethodTrain TimeVal TimeTest TimeTotal Time
MAML1.8h1.9h0.9h5.0h
Proto-MAML1.9h1.9h0.9h5.1h
Meta-Curvature3.4h2.6h1.3h7.6h
ProtoNet0.8h0.4h0.2h1.8h
DDRR1.4h0.6h0.3h2.7h
Proto-FineTuning1.7h4.5h2.3h8.9h
FineTuning1.5h8.1h4.9h14.9h
Linear-Readout1.2h5.1h2.8h9.6h
TFS0.0h0.8h6.2h7.0h
+ +Table 5. Analysis of times needed by different algorithms in the multitask setting (using one NVIDIA 1080 Ti GPU and 4 CPUs). + +multi-task meta-learning per-se. These including: studying the multi-task optimisation [77] in meta-learning, studying HPO for meta-learning, developing validation strategies in meta-learning (using ID vs OD val sets [32]), and studying the benefit of task-specific decoders and external data. + +# 5. Conclusion + +We have introduced Meta Omnium, the first multi-task few-shot meta-learning benchmark for computer vision. The benchmark is challenging in multiple highly topical ways including requiring learning on heterogeneous task distributions, evaluating generalization to out-of-distribution datasets, and uniquely challenging meta-learners to learn-to-learn and transfer knowledge across tasks with heterogeneous output spaces. Meta Omnium is nevertheless lightweight enough to be of broad interest and use for driving future research, and even to support future research in hyper-parameter optimization for meta-learning. + +Acknowledgements This work was supported by the UKRI and EPSRC Grant numbers EP/S000631/1, EP/L016427/1 and EP/S02431X/1; the MOD University Defence Research Collaboration (UDRC) in Signal Processing; the Royal Academy of Engineering under the Research Fellowship programme; Samsung AI Center, Cambridge; China Grant No.2020KJ010802 and No.BYSYZHKC2021115. + +# References + +[1] Milad Abdollahzadeh, Touba Malekzadeh, and Ngai-Man Man Cheung. Revisit multimodal meta-learning through the lens of multi-task learning. In NeurIPS, 2021. 2, 3 +[2] Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, and Christian Gagné. Matching feature sets for few-shot image classification. In CVPR, 2022. 3 +[3] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In KDD, 2019. 6 +[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, 2014. 4, 12 +[5] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. 2022. 1 +[6] Luca Bertinetto, João F. Henriques, Philip H.S. Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In ICLR, 2019. 5 +[7] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. 5 +[8] Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, and Giovanni Zappella. PASHA: Efficient HPO and NAS with progressive resource allocation. In ICLR, 2023. 6 +[9] Ondrej Bohdal, Da Li, and Timothy Hospedales. Feedforward source-free domain adaptation via class prototypes. In ECCV OOD-CV workshop, 2022. 7 +[10] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In ICCV, 2019. 4, 12 +[11] Ruchika Chavhan, Jan Stuehmer, Calum Heggan, Mehrdad Yaghoobi, and Timothy Hospedales. Amortised invariance learning for contrastive self-supervision. In ICLR, 2023. 2 +[12] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 2017. 12 +[13] Vincent Dumoulin, Neil Houlsby, Utku Evci, Xiaohua Zhai, Ross Goroshin, Sylvain Gelly, and Hugo Larochelle. A unified few-shot classification benchmark to compare transfer and meta learning approaches. In NeurIPS Datasets and Benchmarks, 2021. 2, 7, 8 +[14] Nikita Dvornik, Cordelia Schmid, and Julien Mairal. Selecting relevant features from a universal representation for few-shot classification. In ECCV, 2020. 3 +[15] Linus Ericsson, Henry Gouk, and Timothy M. Hospedales. Why do self-supervised models transfer? Investigating the impact of invariance on downstream tasks. BMVC, 2022. 2 +[16] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2, 5 +[17] Mario Fritz, E. Hayman, B. Caputo, and J. Eklundh. THE KTH-TIPS database. 2004. 12 + +[18] Geetharamani G. and Arun Pandian J. Identification of plant leaf diseases using a nine-layer deep convolutional neural network. Computers and Electrical Engineering, 76:323-338, 2019. 12 +[19] Ning Gao, Hanna Ziesche, Ngo Anh Vien, Michael Volpp, and Gerhard Neumann. What matters for meta-learning vision regression tasks? In CVPR, 2022. 2, 4, 5, 7, 12 +[20] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In ICCV, 2021. 1, 2 +[21] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In ICLR, 2021. 3 +[22] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, 2020. 2 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 +[24] Emily F. Brownlee Heidi M. Sosik, Emily E. Peacock. Annotated plankton images - data set for developing and evaluating classification methods., 2015. 12 +[25] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In ECCV, 2022. 3 +[26] Timothy M Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2021. 1, 2 +[27] Shell Hu, Da Li, Jan Stuehmer, Minyoung Kim, and Timothy Hospedales. Pushing the limits of simple pipelines for few-shot learning. In CVPR, 2022. 7, 8 +[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE Workshop on 3D Representation and Recognition (3dRR-13), 2013. 12 +[29] Gustaf Kylberg. The Kylberg texture dataset v. 1.0. Technical report, 2011. 12 +[30] Brenden M. Lake, Ruslan Salakhutdinov, and Joshua B. Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 2015. 2 +[31] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. A sparse texture representation using local affine regions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(8):1265-1278, Aug. 2005. 12 +[32] Da Li, Henry Gouk, and Timothy Hospedales. Finding lost DG: Explaining domain generalization via model complexity. In arXiv, 2022. 3, 8 +[33] Liam Li, Kevin Jamieson, Afshin Rostamizadeh, Ekaterina Gonina, Moritz Hardt Hardt, Benjamin Recht, and Ameet Talwalkar. A system for massively parallel hyperparameter tuning. In MLSys, 2020. 6 +[34] Lisha Li, Kevin G. Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter optimization. JMLR, 2018. 6 + +[35] Rui Li, Ondrej Bohdal, Rajesh K. Mishra, Hyeji Kim, Da Li, Nicholas D. Lane, and Timothy M. Hospedales. A channel coding benchmark for meta-learning. 2021. 2, 3 +[36] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In ICCV, 2021. 3 +[37] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. 1, 2, 3, 4, 5, 12 +[38] Yanbin Liu, Juho Lee, Linchao Zhu, Ling Chen, Humphrey Shi, and Yi Yang. A multi-mode modulator for multi-domain few-shot classification. In ICCV, 2021. 3 +[39] Changsheng Lu and Piotr Koniusz. Few-shot keypoint detection with uncertainty learning for unseen species. In CVPR, 2022. 3 +[40] P. Mallikarjuna, Alireza Tavakoli Targhi, Mario Fritz, E. Hayman, B. Caputo, and J. Eklundh. The KTH-TIPS 2 database. 2006. 12 +[41] Teresa Mendonça, Pedro M. Ferreira, Jorge S. Marques, André R. S. Marcal, and Jorge Rozeira. PH2 - a dermoscopic image database for research and benchmarking. In Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2013. 4, 12 +[42] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In ICCV, 2021. 3 +[43] Ari Morcos, Maithra Raghu, and Samy Bengio. Insights on representational similarity in neural networks with canonical correlation. Advances in Neural Information Processing Systems, 31, 2018. 12 +[44] Jiteng Mu, Weichao Qiu, Gregory D. Hager, and Alan L. Yuille. Learning from synthetic animals. In CVPR, 2020. 4, 12 +[45] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. Unsupervised domain adaptation: A reality check. In arXiv, 2021. 3 +[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 12 +[47] Yoshihiko Ozaki, Yuki Tanigaki, Shuhei Watanabe, and Masaki Onishi. Multiobjective tree-structured parzen estimator for computationally expensive optimization problems. In Genetic and Evolutionary Computation Conference, 2020. 6 +[48] Eunbyung Park and Junier B. Oliva. Meta-Curvature. In NeurIPS, 2019. 5, 7 +[49] Massimiliano Patacchiola, Jack Turner, Elliot J Crowley, Michael O'Boyle, and Amos J Storkey. Bayesian meta-learning for the few-shot setting via deep kernels. In NeurIPS, 2020. 1 +[50] Gerald Piosenko. 100 sports image classification. 12 +[51] Gerald Piosenko. Birds 400 - species image classification. 12 +[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1 +[53] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? Towards understanding the effectiveness of maml. In ICLR, 2020. 2, 3, 12 +[54] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 12 +[55] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A generalist agent, 2022. 1 +[56] Mengye Ren, Eleni Triantafillou, Jake Snell Sachin Ravi, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In ICLR, 2018. 2 +[57] Sebastian Ruder. An overview of multi-task learning in deep neural networks. In arXiv, 2017. 2, 7 +[58] Amirreza Shaban, Shray Bansal, Z. Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. In BMVC, 2017. 4, 12 +[59] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 5 +[60] Haozhe Sun, Wei-Wei Tu, and Isabelle M Guyon. Omniprint: A configurable printed character synthesizer. In NeurIPS Datasets and Benchmarks Track, 2021. 12 +[61] Sebastian Thrun and Lorien Pratt, editors. Learning to Learn. Kluwer Academic Publishers, 1998. 1 +[62] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In ECCV, 2020. 2, 3, 6, 7 +[63] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, and Hugo Larochelle. Meta-dataset: A dataset of datasets for learning to learn from few examples. In ICLR, 2020. 1, 2, 3, 4, 5, 6, 7 +[64] Yu-Yun Tseng, Alexander Bell, and Danna Gurari. Vizwiz-fewshot: Locating objects in images taken by people with visual impairments. In ECCV, 2022. 4, 12 +[65] Ihsan Ullah, Dustin Carrion, Sergio Escalera, Isabelle Guyon, Mike Huisman, Felix Mohr, Jan N. van Rijn, Haozhe Sun, Joaquin Vanschoren, and Phan Anh Vu. Meta-album: Multi-domain meta-dataset for few-shot image classification. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2022. 1, 2, 3, 4, 5, 6, 12, 14 +[66] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NIPS, 2016. 2 + +[67] Risto Vuorio, Shao-Hua Sun, Hexiang Hu, and Joseph J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. In NeurIPS, 2019. 2, 3, 6 +[68] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In ICCV, 2019. 5 +[69] Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning, 2019. 6 +[70] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 1, 2 +[71] Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darryll. What should not be contrastive in contrastive learning. In ICLR, 2021. 2 +[72] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV, 2020. 3 +[73] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In ECCV, 2022. 2, 3 +[74] Yongxin Yang and Timothy Hospedales. A unified perspective on multi-domain and multi-task learning. In ICLR, 2015. 2 +[75] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In CVPR, 2020. 3 +[76] Mingzhang Yin, George Tucker, Mingyuan Zhou, Sergey Levine, and Chelsea Finn. Meta-learning without memorization. 2020. 4, 12 +[77] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In NeurIPS, 2020. 2, 7, 8 +[78] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In CORL, 2019. 3 +[79] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 3 +[80] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. In arXiv, 2019. 2, 7, 8 +[81] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 5 +[82] Linyu Zheng, Ming Tang, Yingying Chen, Jinqiao Wang, and Hanqing Lu. Learning feature embeddings for discriminant model based tracking. In ECCV, 2020. 5 +[83] Bartosz Zielinski, Anna Plichta, Krzysztof Misztal, Przemysław Spurek, Monika Brzychczy-Wloch, and Dorota Ochońska. Deep learning approach to bacterial colony classification. PLOS ONE, 12(9):1-14, 09 2017. 12 \ No newline at end of file diff --git a/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/images.zip b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b0dda2008f05944f8f55eabc9a41c4ee85ccf4ff --- /dev/null +++ b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb4049e50453c279b225e8849753442d3fd6c3a92c2f88062923694ba80aba94 +size 341424 diff --git a/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/layout.json b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5b8cb824def4871e1fcfb1aae5e0ab445b5d51db --- /dev/null +++ b/2023/Meta Omnium_ A Benchmark for General-Purpose Learning-To-Learn/layout.json @@ -0,0 +1,7261 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "text", + "content": "Meta Omnium: A Benchmark for General-Purpose Learning-to-Learn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "spans": [ + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "content": "Ondrej Bohdal" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "content": ", Yinbing Tian" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "inline_equation", + "content": "^{2,\\ast}" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "content": ", Yongshuo Zong" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "content": ", Ruchika Chavhan" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "content": ", Da Li" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "content": ", Henry Gouk" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "content": ", Li Guo" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "text", + "content": ", Timothy Hospedales" + }, + { + "bbox": [ + 121, + 142, + 472, + 172 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 175, + 468, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 175, + 468, + 203 + ], + "spans": [ + { + "bbox": [ + 126, + 175, + 468, + 203 + ], + "type": "text", + "content": "1The University of Edinburgh 2Beijing University of Posts and Telecommunications 3Samsung AI Center, Cambridge *Joint-First Authors" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 290, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 485 + ], + "type": "text", + "content": "Meta-learning and other approaches to few-shot learning are widely studied for image recognition, and are increasingly applied to other vision tasks such as pose estimation and dense prediction. This naturally raises the question of whether there is any few-shot meta-learning algorithm capable of generalizing across these diverse task types? To support the community in answering this question, we introduce Meta Omnium, a dataset-of-datasets spanning multiple vision tasks including recognition, keypoint localization, semantic segmentation and regression. We experiment with popular few-shot meta-learning baselines and analyze their ability to generalize across tasks and to transfer knowledge between them. Meta Omnium enables meta-learning researchers to evaluate model generalization to a much wider array of tasks than previously possible, and provides a single framework for evaluating meta-learners across a wide suite of vision applications in a consistent manner. Code and dataset are available at https://github.com/edi-meta-learning/meta-omnium." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "Meta-learning is a long-standing research area that aims to replicate the human ability to learn from a few examples by learning-to-learn from a large number of learning problems [61]. This area has become increasingly important recently, as a paradigm with the potential to break the data bottleneck of traditional supervised learning [26, 70]. While the largest body of work is applied to image recognition, few-shot learning algorithms have now been studied in most corners of computer vision, from semantic segmentation [37] to pose estimation [49] and beyond. Nevertheless, most of these applications of few-shot learning are advancing independently, with increasingly divergent application-specific methods and benchmarks. This makes it hard to evaluate whether few-shot meta-learners can solve diverse vision tasks. Importantly it also discourages the de" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 232, + 547, + 494 + ], + "blocks": [ + { + "bbox": [ + 307, + 232, + 547, + 494 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 547, + 494 + ], + "type": "image", + "image_path": "a1ca96f2b4f105019e810b99833709cf85445be2e6e6c46008e776bb05d4d7f1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 500, + 547, + 534 + ], + "lines": [ + { + "bbox": [ + 305, + 500, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 305, + 500, + 547, + 534 + ], + "type": "text", + "content": "Figure 1. Illustration of the diverse visual domains and task types in Meta Omnium. Meta-learners are required to generalize across multiple task types, multiple datasets, and held-out datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "text", + "content": "velopment of meta-learners with the ability to learn-to-learn across tasks, transferring knowledge from, e.g., keypoint localization to segmentation – a capability that would be highly valuable for vision systems if achieved." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "The overall trend in computer vision [20, 52] and AI [5, 55] more generally is towards more general-purpose models and algorithms that support many tasks and ideally leverage synergies across them. However, it has not yet been possible to explore this trend in meta-learning, due to the lack of few-shot benchmarks spanning multiple tasks. State-of-the-art benchmarks [63, 65] for visual few-shot learning are restricted to image recognition across a handful of visual domains. There is no few-shot benchmark" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7693" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 547, + 156 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 547, + 156 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 547, + 156 + ], + "type": "table", + "html": "
DatasetNum TasksNum DomainsNum ImagesCategoriesSizeLightweightMulti-TaskMulti-Domain
Omniglot [30]1132K1623148MB✓✗✗
miniImageNet [66]1160K1001GB✓✗✗
Meta-Dataset [63]17~1053M43~1500210GB✗✗✓
VTAB [80]13~192.2M2~397100GB✗✗✓
FSS1000 [37]11100001000670MB✓✗✗
Meta-Album [65]110~401.5M19~70615GB✓✗✓
Meta Omnium421160K2~7063.1GB✓✓✓
", + "image_path": "0b0888bcc8a68974cc66b2d78f6598302bf2a6ae8d0596d58222b9043d42025e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 164, + 546, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 164, + 546, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 546, + 187 + ], + "type": "text", + "content": "Table 1. Feature comparison between Meta Omnium and other few-shot meta-learning benchmarks. Meta Omnium uniquely combines a rich set of tasks and visual domains with a lightweight size for accessible use." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "content": "that poses the more substantial challenge [57, 77] of generalizing across different tasks. We remark that the term task is used differently in few-shot meta-learning literature [16, 26, 70] (to mean different image recognition problems, such as cat vs dog or car vs bus) and the multi-task literature [20, 57, 74, 77] (to mean different kinds of image understanding problems, such as classification vs segmentation). In this paper, we will use the term task in the multi-task literature sense, and the term episode to refer to tasks in the meta-learning literature sense, corresponding to a support and query set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 341, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 604 + ], + "type": "text", + "content": "We introduce Meta Omnium, a dataset-of-datasets spanning multiple vision tasks including recognition, semantic segmentation, keypoint localization/pose estimation, and regression as illustrated in Figure 1. Specifically, Meta Omnium provides the following important contributions: (1) Existing benchmarks only test the ability of meta-learners to learn-to-learn within tasks such as classification [63, 65], or dense prediction [37]. Meta Omnium uniquely tests the ability of meta-learners to learn across multiple task types. (2) Meta Omnium covers multiple visual domains (from natural to medical and industrial images). (3) Meta Omnium provides the ability to thoroughly evaluate both in-distribution and out-of-distribution generalisation. (4) Meta Omnium has a clear hyper-parameter tuning (HPO) and model selection protocol, to facilitate future fair comparison across current and future meta-learning algorithms. (5), Unlike popular predecessors, [63], and despite the diversity of tasks, Meta Omnium has been carefully designed to be of moderate computational cost, making it accessible for research in modestly-resourced universities as well as large institutions. Table 1 compares Meta Omnium to other relevant meta-learning datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We expect Meta Omnium to advance the field by encouraging the development of meta-learning algorithms capable of knowledge transfer across different tasks – as well as across learning episodes within individual tasks as is popularly studied today [16, 70]. In this regard, it provides the next step of the level of a currently topical challenge of dealing with heterogeneity in meta-learning [1, 35, 63, 67]. While existing benchmarks have tested multi-domain heterogeneity (e.g., recognition of written characters and plants" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 208, + 547, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 547, + 364 + ], + "type": "text", + "content": "within a single network) [63, 65] and shown it to be challenging, Meta Omnium tests multi-task learning (e.g., character recognition vs plant segmentation). This is substantially more ambitious when considered from the perspective of common representation learning. For example, a representation tuned for recognition might benefit from rotation invariance, while one tuned for segmentation might benefit from rotation equivariance [11, 15, 71]. Thus, in contrast to conventional within-task meta-learning benchmarks that have been criticized as relying more on common representation learning than learning-to-learn [53, 62], Meta Omnium better tests the ability of learning-to-learn since the constituent tasks require more diverse representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 375, + 392, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 392, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 392, + 387 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 395, + 457, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 457, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 457, + 407 + ], + "type": "text", + "content": "2.1. Meta-learning Benchmarks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 414, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 653 + ], + "type": "text", + "content": "The classic datasets in few-shot meta-learning for computer vision are Omniglot [30] and miniImageNet [66]. Later work criticized these for having insufficient task (episode) diversity and tieredImageNet [56] used the class hierarchy of ImageNet to enforce more diversity between meta-train and meta-test episodes. The main contemporary benchmarks are CD-FSL [22], which challenges few-shot learners to generalize to new visual domains; and Meta-Dataset [63] and Meta-Album [65], which go further in requiring few-shot learners to learn from a mixture of visual domains. Such multi-domain heterogeneous meta-learning turns out to be challenging. A related benchmark to Meta-Dataset is VTAB [80], which similarly provides multiple domains for evaluating data-efficient visual recognition, but their focus is on evaluating representation transfer from large-scale pre-training rather than learning-to-learn and meta-learning. VTAB+MD [13] compare representation transfer and meta-learning approaches on the Meta-Dataset tasks. However, none of these benchmarks address multi-task meta-learning as considered here (Figure 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Outside of recognition, task-specific few-shot benchmarks have been proposed in vision problems of semantic segmentation [37], regression [19], pose/keypoint estimation [73], etc. These are mostly slightly behind the complexity of the recognition benchmarks with regards to being" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7694" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "single-domain, with the exception of [73]. With regard to multi-task meta-learning as considered here, the only existing benchmark is meta-world [78], which is specific to robotics and reinforcement learning rather than vision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 217 + ], + "type": "text", + "content": "We also mention taskonomy [79] as a popular dataset that has been used for multi-task learning. However, it is not widely used for few-shot meta-learning. This is because, although taskonomy has many tasks, unlike the main meta-learning benchmarks [37, 63], there are not enough visual concepts within each task to provide a large number of concepts for meta-training and a disjoint set of concepts to evaluate few-shot learning for meta-validation and meta-testing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 225, + 222, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 222, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 222, + 239 + ], + "type": "text", + "content": "2.2. Heterogeneity in Meta-Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 244, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 288, + 472 + ], + "type": "text", + "content": "There are several sophisticated methods in the literature that highlighted the challenge of the task to address heterogeneity in meta-learning. These have gone under various names such as multi-modal meta-learning [1, 38, 67] – in the sense of multi-modal probability distribution (over tasks/episodes). However, with the exception of [38], these have mostly not been shown to scale to the main multimodal benchmarks such as Meta-Dataset [63]. A more common approach to achieving high performance across multiple heterogeneous domains such as those in Meta-Dataset is to train an ensemble of feature extractors across available training domains, and fuse them during metatesting [14, 36, 36]. However, this obviously incurs a substantial additional cost of maintaining a model ensemble. In our evaluation, we focus on the simpler meta-learners that have been shown to work in challenging multi-domain learning scenarios [63, 65], while leaving sophisticated algorithmic and ensemble-based approaches for future researchers to evaluate on the benchmark." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 484, + 269, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 269, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 269, + 497 + ], + "type": "text", + "content": "3. Meta Omnium Benchmark and Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 505, + 233, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 233, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 233, + 517 + ], + "type": "text", + "content": "3.1. Motivation and Guiding Principles" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 523, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 656 + ], + "type": "text", + "content": "We first explain the motivating goals and guiding principles behind the design of Meta Omnium. The goal is to build a benchmark for multi-task meta-learning that will: (i) Encourage the community to develop meta-learners that are flexible enough to deal with greater task heterogeneity than before, and thus are more likely to be useful in practice with less curated episode distributions. This was identified as a major challenge in the discussion arising in several recent meta-learning and computer vision workshops and challenges1. (ii) Ultimately progress on this benchmark should provide practical improvements in data-efficient learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "for computer vision through the development of methods that can better transfer across different task types." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 98, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 547, + 566 + ], + "type": "text", + "content": "In developing this benchmark, we established a few principles that we used to guide design choices. These include: (i) The benchmark should be lightweight in terms of storage and computing, making it accessible to a broad range of researchers and not only large corporations. (ii) The benchmark should cover multiple tasks with heterogeneous output spaces (as opposed to all classification, all regression, or all dense prediction), as well as multiple visual domains. In these regards, Meta Omnium is compared to alternatives in Table 1. (iii) The initial baselines should have only minimal task-specific decoders. This is in contrast to the state of the art within various sub-disciplines of FSL such as segmentation [25, 42], keypoint [39, 72], and classification [2, 75] where specially designed decoders are often used. This is to evaluate and encourage future research on learning-to-learn across tasks, rather than primarily benchmarking how well we can manually engineer prior knowledge of optimal task-specific decoders. While we are not opposed to future competitors on this benchmark developing task-specific decoders, these should be evaluated separately against the minimal-decoder competitors. (iv) The benchmark should provide distinct datasets for in-distribution (ID) training and out-of-distribution (OOD) evaluation, to evaluate the robustness of the distribution-shift. This is already provided by [63, 65] for classification, and we extend such an ID and OOD dataset ensemble to multiple tasks. Figure 2 illustrates our dataset and task-split. (v) The benchmark should provide a clear hyper-parameter tuning protocol. With a number of recent studies showing that hyper-parameter tuning can dominate other effects of interest in computer vision [21, 32, 45], this is important for a future-proof meta-learning benchmark. This is also related to the first cost point (i) above: Only for a benchmark with a modest cost can most institutions realistically expect to conduct hyperparameter tuning. We provide the hyper-parameter tuning protocol. (vi) Finally, following the debate in [35, 53, 62] as to the value of meta-learning vs conventional transfer learning, the dataset should support both episodic meta-learning and conventional transfer learning approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 574, + 430, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 430, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 430, + 586 + ], + "type": "text", + "content": "3.2. Data Splits and Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "For each main task (classification, segmentation, key-point localization), we split the datasets into seen datasets available for meta-training, and unseen datasets that are completely held out for out-of-domain meta-validation and meta-testing. Similarly to [63, 65], for the seen datasets, we construct category-wise splits into meta-train/val/test. While for the unseen datasets, there is no category-wise split as episodes from all categories from the whole dataset will be used for validation and testing respectively. The overall split organization is illustrated in Figure 2. We" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 46, + 664, + 287, + 715 + ], + "type": "text", + "content": "ICLR https://sites.google.com/view/learning-2-learn/, ECML https://janvanrijn.github.io/metalearning/workshop2022, NeurIPS https://metalearning.chalearn.org/metadlneurips2021, and ECCV http://www.ood-cv.org/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7695" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 289, + 228 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 289, + 228 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 289, + 228 + ], + "type": "image", + "image_path": "a47e27c77a1999e1448109dd61b535451034f1f2c62b83d9a0d91c8bc3dbbded.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 290, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 360 + ], + "type": "text", + "content": "Figure 2. Schematic of benchmark and dataset splits. For each task, there are multiple datasets, which are divided into seen (solid border) and unseen (dash border) datasets. The seen (ID) datasets are divided class-wise into meta-train/meta-val/meta-test splits. The unseen datasets are held out for out-of-distribution (OD) evaluation. Meta-training is conducted on the ID-meta-train split of the seen datasets (blue). Models are validated on ID validation class splits, or OOD validation datasets (green). Results are reported on the ID test class splits and OOD test datasets (orange). We also hold out an entire task family (regression) for evaluating novel task generalisation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 381, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 287, + 407 + ], + "type": "text", + "content": "additionally have a completely held-out task: regression. Datasets from this task are not used during meta-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": "Our multi-task setup enables us to define and compare two training protocols: Single-task meta-learning which evaluates how well meta-learning performs when trained and tested within a particular task family (within each plate in Figure 2); and Multi-task meta-learning which evaluates how well meta-learning performs when trained across all available task families (across all plates in Figure 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "content": "With this organization we can separately evaluate: Within-distribution generalization (ID): How well do meta-learners generalize to novel test concepts within the seen datasets?; and Out-distribution generalization (OOD): How well do meta-learners generalize to novel concepts in unseen datasets?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 624 + ], + "type": "text", + "content": "We provide two sources of validation data: ID and OOD, and our models are selected based on the combined performance across both. OOD validation is not supposed by the most popular Meta-Dataset benchmark [63] as despite its larger size it does not provide OOD validation datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 169, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 169, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 169, + 645 + ], + "type": "text", + "content": "3.3. Datasets and Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "Given the considerations in Section 3.1, our benchmark consists of three main tasks (classification, segmentation, keypoints/pose) and one held-out task (regression)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Classification For classification we take the 10 datasets from the initial public release of Meta-Album [65]. These" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "images are all " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": " and contain 19-706 classes per dataset, with 40 images per class. Three of these datasets are reserved for out-of-distribution meta-validation, and four for out-of-distribution meta-test." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 121, + 546, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 264 + ], + "type": "text", + "content": "Segmentation For segmentation, we take FSS1000 [37] for in-distribution (10,000 images, 1,000 classes), and combine it with VizWiz [64] for OOD meta-validation (862 images, 22 classes), and modified Pascal5i [58] (7,242 images, 6 classes) and the very distinct medical imaging dataset PH2 [41] (200 images, 3 classes) for meta-testing. The segmentation images originally were of diverse sizes. We resize them all to " + }, + { + "bbox": [ + 304, + 121, + 546, + 264 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 121, + 546, + 264 + ], + "type": "text", + "content": " for Meta Omnium. Note that VizWiz and Pascal datasets originally contain more classes and images. We exclude the classes that overlap with that in the FSS1000 dataset for few-shot learning, and thus there are no classes overlapping among all the datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "text", + "content": "Keypoints For keypoints/pose, we take animal-pose [10] for in-distribution, synthetic animal-pose [44] for OOD meta-validation, and MPII human-pose [4] for OOD meta-testing. All images are resized to " + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "text", + "content": ". MPII includes about " + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "inline_equation", + "content": "40\\mathrm{k}" + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "text", + "content": " people in over " + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "inline_equation", + "content": "25\\mathrm{k}" + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "text", + "content": " images with annotated body keypoints. Animal Pose includes 5 animal categories for 6K instances in over " + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "inline_equation", + "content": "4\\mathrm{k}" + }, + { + "bbox": [ + 304, + 265, + 546, + 421 + ], + "type": "text", + "content": " images. Each animal is cropped from the original image. We keep cats and dogs for training, horses and sheep for in-domain validation, and cow for indomain testing. Synthetic animal pose generates synthetic images using animal CAD models rendered from various viewpoints and lightings on a random background. We keep only the horse and tiger categories in our final datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 624 + ], + "type": "text", + "content": "Regression For evaluating regression as a held-out task, we use four datasets corresponding to the test splits of [19]: ShapeNet1D, ShapeNet2D, Distractor and Pascal1D [76]. All images are resized to " + }, + { + "bbox": [ + 304, + 422, + 547, + 624 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 422, + 547, + 624 + ], + "type": "text", + "content": ". ShapeNet1D aims to predict azimuth angles. It contains 30 categories in total and we keep the 3 categories from the test set. ShapeNet2D further includes 2D rotation with azimuth angles and elevation. The test set of ShapeNet2D contains 300 categories in total with 30 images per category. Distractor aims to predict the position of a target object in the presence of a distractor. It contains 12 categories in total and the test set has 2 categories. Each category contains 1000 objects with 36 images for each. Pascal1D aims to predict azimuth angle. The whole Pascal1D contains 65 objects from 10 categories. The test set contains 15 objects with 100 images for each object. The supplementary material provides full details of all datasets and splits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 634, + 391, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 391, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 391, + 647 + ], + "type": "text", + "content": "3.4. Training API" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "For episodic learning, we proceed by (i) sampling a task, (ii) sampling a dataset, (iii) sampling an episode. Under our main protocol we consider variable 1 to 5-shot evaluation, but also evaluate separate 1 and 5 shot settings (training is always done with a variable number of shots - sup" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7696" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "port examples). For classification tasks, we follow [65] in generating 5-way episodes. For segmentation tasks, we follow [37] in considering each episode to be a binary foreground/background classification problem for a novel class and generate 2-way episodes. For keypoint, we form episodes by randomly selecting a class (e.g., animal category) and then randomly selecting a subset of 5 keypoints to localize for each episode. For regression tasks, we generate variable 5 to 25-shot episodes because it is a common practice to use more shots for regression tasks [19]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 288, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 288, + 241 + ], + "type": "text", + "content": "For non-episodic/transfer learning, we provide access to the meta-train portions of the seen datasets in conventional mini-batches for conventional single task and multi-task supervised learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 242, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 373 + ], + "type": "text", + "content": "Evaluation Metrics For classification tasks, we use standard top-1 accuracy; for segmentation tasks, we use standard mean intersection-over-union (mIoU) that averages over IoU values of all object classes; for keypoint prediction, we report the Percentage of Correct Keypoints (PCK). In detail, a detected joint is considered correct if the distance between the predicted and the true joint is within a certain threshold. In our experiments, the threshold is 0.01 for normalized value, which stands for about 12.8 pixel of input image resolution. For regression tasks, we follow [19] and use the same metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 383, + 253, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 253, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 253, + 396 + ], + "type": "text", + "content": "3.5. Architecture and Baseline Competitors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "text", + "content": "As discussed in Section 3.1, we aim to establish baselines that can be adapted to tasks with heterogeneous outputs, with minimal reliance on task-specific decoders. We follow [63, 65] in using a ResNet-18 CNN [23] as a feature extractor architecture. For recognition tasks, we perform multi-class classification immediately after ResNet's GAP. For regression tasks, we perform linear regression directly after the ResNet's GAP. For keypoint tasks, we consider them to be a regression problem from the feature map to the keypoint location. For segmentation tasks, we use a simplified PSPNet-like [81] strategy. We concatenate the extracted feature maps from ResNet's feature pyramid, with upsampling where appropriate, to generate a feature map of size " + }, + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "inline_equation", + "content": "w \\times h" + }, + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "text", + "content": ", and then do pixel-wise classification with " + }, + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "text", + "content": " convolutional layer to obtain the final segmentation map. All tasks thus have only one learnable weight as a minimal classifier/decoder after the common ResNet feature encoder. Based on this common encoder and minimal decoder architecture, we describe our meta-learning baselines. Prototypical Network [59] is a classic meta-learner that exploits nearest-centroid metric learning for few-shot classification. ProtoNets were adapted to segmentation tasks in PANet [68], by performing pixel-level feature matching between support prototypes and query pixels. We use the same principle together with the PSPNet-like features described earlier. To generalize ProtoNets to regression tasks such as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "keypoint prediction, we must relax the prototype assumption, and use them as simple Gaussian kernel-regression models [7]. Specifically, we generate a feature embedding for each support example, and then for query examples, we calculate the negative exponential distance to the support examples, and use this inverse distance-weighted sum of support set labels as the prediction. Thus for regression tasks with a support set " + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "inline_equation", + "content": "S = \\{(x_i,y_i)\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": " and query example " + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": ", ProtoRegression predicts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 190, + 547, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 190, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 318, + 190, + 547, + 215 + ], + "type": "interline_equation", + "content": "f (y _ {q} | x _ {q}, S) \\propto \\sum_ {i} y _ {i} \\exp (- \\tau (f _ {\\theta} (x _ {i}) - f _ {\\theta} (x _ {q})) ^ {2}) \\quad (1)", + "image_path": "b531cc91ad72353d6607bc707acb01f4e82d4d9537535fd9d9753ae48d93712f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 223, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 258 + ], + "type": "text", + "content": "enabling us to learn deep feature " + }, + { + "bbox": [ + 304, + 223, + 545, + 258 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 223, + 545, + 258 + ], + "type": "text", + "content": " in the usual episodic meta-learning way. We use cross-entropy loss for classification and segmentation, and MAE loss for regression tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 259, + 546, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 546, + 461 + ], + "type": "text", + "content": "DDRR Deep differentiable ridge-regression has been considered for few-shot recognition [6], tracking [82], and other tasks. It is related to ProtoNet in that the feature is not adapted after the meta-train stage, but different in that the decoder/classifier layer is learned by differentiable ridge-regression rather than nearest centroid or kernel regression. An elegant property of DDRR methods is that they naturally address regression tasks, although they have been repurposed for classification [6] by conducting MSE-loss regression to a target 1-hot vector. Thus they are a natural choice for our benchmark. For application to segmentation, we apply DDRR in a " + }, + { + "bbox": [ + 304, + 259, + 546, + 461 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 259, + 546, + 461 + ], + "type": "text", + "content": " convolution-like way, to perform pixel classification for the output mask with a DDRR classifier at each pixel. Further, we calibrate the prediction for binary cross entropy loss with a learnable scale and bias following [6]. DDRR uses MAE loss for only regression tasks and use MSE loss for other tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 547, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 642 + ], + "type": "text", + "content": "MAML The seminal few-shot meta-learner MAML [16] aims to learn an initial condition for per-episode gradient-descent. MAML is straightforward to adapt to different types of tasks. Based on each episode's support set, a new output layer is learned, and the feature extractor is updated, both by a few steps of gradient descent. Similarly to Meta-Dataset [63], we do not learn an initialization for the output layer, since it can change size between episodes drawn from multiple tasks. To alleviate this challenge, we also follow Meta-Dataset in evaluating Proto-MAML – a variant that initializes the MAML output layer based on the linear classifier/regressor suggested by nearest-centroid prior to gradient descent. Going beyond this, to adapt Proto-MAML to regression tasks, we also initialize the output layer based on the ridge-regression solution to the support set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 700 + ], + "type": "text", + "content": "Meta-Curvature Meta-Curvature [48] is an enhancement of MAML that learns a pre-conditioning matrix to improve inner-loop adaptation, as well as an initial condition as in standard MAML. Meta-Curvature outperforms MAML in simpler single-task few-shot benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "content": "Transfer Learning We also consider standard supervised" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7697" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "learning on the meta-train tasks for transfer to the target tasks, a strategy reported to be competitive with meta-learning [62]. For adaptation, we explore both linear readout [62, 69] and fine-tuning [65]. Besides learning a new output layer from scratch, we also consider a fine-tuning version that initializes the classifier weights using class prototypes (recognition/segmentation) or ridge regression weights (keypoints/regression), inspired by Proto-MAML. Train-from-Scratch (TFS) We lastly consider training each episode from scratch using only the support set [65]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 198, + 214, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 214, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 214, + 211 + ], + "type": "text", + "content": "3.6. Hyperparameter Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": "As part of our benchmark, we perform hyperparameter optimization (HPO) to ensure we select appropriate hyperparameters for the diverse tasks that we consider. Multi-objective HPO under restricted resources is challenging, so we devise the following HPO protocol: estimate the performance of each candidate configuration on a lower fidelity (lower number of iterations) and then identify the configuration that works the best across all validation datasets considered (combination of in-domain and out-domain datasets, across various task families). Note that fast multi-fidelity methods such as Hyperband [34], ASHA [33] or PASHA [8] are not applicable out of the box in our multi-objective setup, so we decided to train each candidate configuration using fixed 5,000 training iterations. Since different tasks/datasets are of different difficulties (and use different metrics), we normalize the score of each configuration for each validation dataset by the best score for that dataset across all candidate configurations. We then select the configuration with the best average normalized score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 289, + 598 + ], + "type": "text", + "content": "Note that due to resource constraints we are only able to sample a relatively smaller number of candidates (30), so we utilize a sample efficient state-of-the-art Multi-Objective TPE method [47], available from the Optuna library [3]. We perform HPO for multi-task and single-tasks setups separately, so single task classification, segmentation and keypoint estimation have their own set of hyperparameters; and the multi-task case has its own set. The hyperparameters include the meta-learning rate and optimizer, momentum, and various method-specific hyperparameters (full details are in the supplementary). Once the hyperparameters are chosen, we perform standard training of the model for the full number of iterations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "In this section, we aim to use our benchmark to answer the following questions: (1) Which meta-learner performs best on average across a heterogenous range of tasks? Existing benchmarks have evaluated meta-learners for one task at a time, we now use our common evaluation platform to find out if any meta-learner can provide general-purpose learning to learn across different task types, or whether" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "each task type prefers a different learner. Similarly, we can ask which meta-learner is most robust to out-of-distribution tasks? (2) Having defined the first multi-task meta-learning benchmark, and generalizations of seminal meta-learners to different kinds of output spaces, we ask which meta-learner performs best for multi-task meta-learning? More generally, is there a trend in gradient-based vs metric-based meta-learner success? (3) Does multi-task meta-learning improve or worsen performance compared to single-task? The former obviously provides more meta-training data, which should be advantageous, but the increased heterogeneity across meta-training episodes in the multi-task case also makes it harder to learn [63, 67]. (4) How does meta-learning perform compared to simple transfer learning, or learning from scratch?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 260, + 432, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 260, + 432, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 432, + 274 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 279, + 547, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 547, + 435 + ], + "type": "text", + "content": "We train each meta-learner for 30,000 meta-training iterations, with meta-validation after every 2,500 iterations (used for checkpoint selection). For evaluation during meta-testing we use 600 tasks for each corresponding dataset, and for meta-validation we use 1200 tasks together. We use random seeds to ensure that the same tasks are used across all methods that we compare. For transfer learning approaches (fine-tuning, training from scratch, etc.), we use 20 update steps during evaluation. We only retain the meta-learned shared feature extractor across tasks, and for each new evaluation task, we randomly initialize the output layer so that we can support any number of classes as well as novel task families during meta-testing (in line with [65])." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 443, + 362, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 362, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 362, + 455 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 462, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 605 + ], + "type": "text", + "content": "The main experimental results are shown in Table 2, where rows correspond to different few-shot learners, and columns report the average performance of test episodes, aggregated across multiple datasets in each task family, and broken down by \"seen\" datasets (ID) and \"unseen\" datasets (OOD). The table also reports the average rank of each meta-learner across each dataset, both overall and broken down by ID and OOD datasets. More specifically, for each setting (e.g. cls. ID) we calculate the rank of each method (separately for STL and MTL), and then we average those ranks across cls., seg. and keypoints. From the results, we can draw the following main conclusions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "type": "text", + "content": "(1) ProtoNet is the most versatile meta-learner, as shown by its highest average rank in the single-task scenario. This validates our novel Kernel Regression extension of ProtoNet for tackling regression-type keypoint localization tasks. Somewhat surprisingly, ProtoNet is also the most robust to out-of-distribution episodes (OOD) which is different from the conclusion of [63] and others who suggested that gradient-based adaptation is crucial to adapt to OOD data. However, it is also in line with the results of [65] and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7698" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 496, + 327 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 496, + 327 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 496, + 327 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 496, + 327 + ], + "type": "table", + "html": "
ClassificationSegmentationKeypointsAverage Rank
IDOODIDOODIDOODIDOODAVG
Single-TaskMAML58.761.654.742.125.433.04.33.33.8
Proto-MAML50.549.746.444.123.622.56.06.36.2
Meta-Curvature64.861.465.649.843.516.02.04.33.2
ProtoNet70.459.475.857.227.833.31.31.71.5
DDRR63.158.766.748.020.531.94.73.74.2
Proto-FineTuning50.850.760.043.421.333.15.34.34.8
FineTuning42.348.250.540.025.730.05.76.76.2
Linear-Readout48.653.434.022.722.126.97.36.77.0
TFS31.542.042.837.621.026.08.38.08.2
Multi-TaskMAML59.158.543.337.424.323.92.74.73.7
Proto-MAML58.563.753.043.221.633.33.01.72.3
Meta-Curvature70.466.942.634.518.225.34.34.74.5
ProtoNet65.958.863.349.720.133.02.72.02.3
DDRR52.851.940.437.322.830.15.04.74.8
Proto-FineTuning52.453.244.837.821.230.04.34.04.2
FineTuning44.151.241.336.118.120.57.77.07.3
Linear-Readout46.050.941.532.619.923.56.38.07.2
TFS21.923.838.735.814.111.09.08.38.7
", + "image_path": "24f653677f47fc11e3a0f4b80d21fa62f582276c66df6585968c36f6c38f833c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 412, + 271, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 271, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 271, + 425 + ], + "type": "text", + "content": "the strong performance of prototypes more broadly [9]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 427, + 288, + 602 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 427, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 287, + 464 + ], + "type": "text", + "content": "(2) Coming to multi-task meta-learning the situation is similar in that ProtoNet dominates the other competitors, but now sharing the first place with Proto-MAML." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 552 + ], + "type": "text", + "content": "(3) To compare single-task and multi-task meta-learning (top and bottom blocks of Table 2) more easily, Figure 3 shows the difference in meta-testing episode performance after STL and MTL meta-training for each method. Overall STL outperforms the MTL condition, showing that the difficulty of learning from heterogeneous tasks [57,77] outweighs the benefit of the extra available multi-task data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 554, + 288, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 288, + 602 + ], + "type": "text", + "content": "(4) Finally, comparing meta-learning methods with simple transfer learning methods as discussed in [13, 62], the best meta-learners are clearly better than transfer learning for both single and multi-task scenarios." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We also note that Proto-MAML is better than MAML in the multi-task case, likely due to the importance of a good output-layer initialization in the case of heterogeneous episodes, as per [63]. Meta-Curvature outperforms MAML in single-task in-domain scenarios, in line with previous results [48], but it did not achieve stronger performance out-of-domain or in the multi-task case. Finally, while DDRR is perhaps the most elegant baseline in terms of most naturally spanning all task types, its overall performance is middling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 411, + 421, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 421, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 421, + 425 + ], + "type": "text", + "content": "4.3. Additional Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 430, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 575 + ], + "type": "text", + "content": "How well can multi-task meta-learners generalize to completely new held-out tasks? We take the multi-task meta-learners (trained on classification, segmentation, keypoints) and evaluate them on four regression benchmarks inspired by [19]: ShapeNet1D, ShapeNet2D, Distractor and Pascal1D. Because the metrics differ across datasets, we analyse the rankings and summarise the results in Table 3. We see the basic TFS performs the worst, with MAML, ProtoNets, DDRR being the best. However, in several cases the results were not better than predicting the mean (full results in the appendix), showing that learning-to-learn of completely new task families is an open challenge." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 585, + 545, + 608 + ], + "blocks": [ + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "type": "text", + "content": "Table 2. Main Results. Results are presented as averages across the datasets within each task type and separately for in-distribution (ID) and out-of-distribution (OOD) datasets. Classification, segmentation, and keypoint results are reported in accuracy " + }, + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "type": "text", + "content": ", mIOU " + }, + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "type": "text", + "content": ", and PCK " + }, + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 335, + 545, + 392 + ], + "type": "text", + "content": " respectively. The upper and lower groups correspond to multi-task and single-task meta-training prior to evaluation on the same set of meta-testing episodes. Upper and lower sub row groups correspond to meta-learners and non-meta learners respectively. See the appendix for a full breakdown over individual datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 585, + 545, + 608 + ], + "lines": [ + { + "bbox": [ + 306, + 585, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 545, + 608 + ], + "type": "table", + "html": "
MAMLPMAMLMCPNDDRRPFTFTLRTFS
3.36.54.83.53.53.85.84.38.5
", + "image_path": "3d42ba26c3248fecfba30e63f677723ada5e2521af80f973b3b6cde52e4b808a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 616, + 545, + 639 + ], + "lines": [ + { + "bbox": [ + 305, + 616, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 616, + 545, + 639 + ], + "type": "text", + "content": "Table 3. Average ranking of the different methods across four out-of-task regression datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "How much does external pre-training help? Our focus is on assessing the efficacy of meta-learning rather than representation transfer, but we also aim to support researchers investigating the impact of representation learning on external data prior to meta-learning [13, 27, 80]. We therefore" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7699" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 76, + 519, + 194 + ], + "blocks": [ + { + "bbox": [ + 75, + 76, + 519, + 194 + ], + "lines": [ + { + "bbox": [ + 75, + 76, + 519, + 194 + ], + "spans": [ + { + "bbox": [ + 75, + 76, + 519, + 194 + ], + "type": "image", + "image_path": "6d0fefa6ee20d4b2272b0d0732322bc1aed07e4d4dcad11f9ea04e6f87149ece.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 204, + 525, + 216 + ], + "lines": [ + { + "bbox": [ + 66, + 204, + 525, + 216 + ], + "spans": [ + { + "bbox": [ + 66, + 204, + 525, + 216 + ], + "type": "text", + "content": "Figure 3. Analysis of the differences in scores between single-task (STL) and multi-task (MTL) learning for different methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 234, + 287, + 294 + ], + "blocks": [ + { + "bbox": [ + 50, + 234, + 287, + 294 + ], + "lines": [ + { + "bbox": [ + 50, + 234, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 50, + 234, + 287, + 294 + ], + "type": "table", + "html": "
MethodPretrainCls.Segm.Keyp.
IDOODIDOODIDOOD
Proto-MAMLX58.563.753.043.221.633.3
ProtoNetX66.058.863.349.720.133.0
Proto-MAML✓63.962.756.245.321.833.3
ProtoNet✓63.558.662.049.020.133.1
", + "image_path": "c44f212ea733b953c0d72e77dd3fb193da7ee1b88884ecaa20dbe0afb9de33c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 301, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 357 + ], + "type": "text", + "content": "Table 4. Analysis of the impact of external data pretraining for selected meta-learners in multi-task learning condition. The results show that ImageNet pretraining does not necessarily help improve performance. Cls., Segm., Keyp. represent classification, segmentation and keypoint respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 378, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 402 + ], + "type": "text", + "content": "specify evaluation conditions where external data outside our defined in-distribution meta-training set is allowed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 557 + ], + "type": "text", + "content": "We take two high-performing approaches in the multi-task scenario, Proto-MAML and ProtoNet, and we investigate to what extent external pre-training helps. We use the standard ImageNet1k pre-trained ResNet18, prior to conducting our meta-learning pipeline as usual. We use the same hyperparameters as selected earlier for these models to ensure consistent evaluation, and ensure that the differences in performance are not due to a better selection of hyperparameters. The results in Table 4 show that pretraining is not necessarily helpful in the considered multi-task setting, in contrast to purely recognition-focused evaluations [13,27,80], which were unambiguously positive about representation transfer from external data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "content": "Analysis of Runtimes We analyze the times that the different meta-learning approaches spend on meta-training, meta-validation and meta-testing in the multitask learning case of our benchmark. The results in Table 5 show that all experiments are relatively lightweight, despite the ambitious goal of our benchmark to learn a meta-learner that can generalize across various task families. Most notably we observe that ProtoNets are the fastest approach, alongside being the best-performing one. Note that fine-tuning and training from scratch are expensive during the test time as they use backpropagation with a larger number of steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Discussion and Future Work In future Meta Omnium can be used in a variety of ways beyond benchmarking" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 234, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 307, + 234, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 334 + ], + "type": "table", + "html": "
MethodTrain TimeVal TimeTest TimeTotal Time
MAML1.8h1.9h0.9h5.0h
Proto-MAML1.9h1.9h0.9h5.1h
Meta-Curvature3.4h2.6h1.3h7.6h
ProtoNet0.8h0.4h0.2h1.8h
DDRR1.4h0.6h0.3h2.7h
Proto-FineTuning1.7h4.5h2.3h8.9h
FineTuning1.5h8.1h4.9h14.9h
Linear-Readout1.2h5.1h2.8h9.6h
TFS0.0h0.8h6.2h7.0h
", + "image_path": "08dd6804cdd15f09fa1646c176cb024d7b8676b5f24612e34232a9227bf81312.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 342, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 306, + 342, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 545, + 364 + ], + "type": "text", + "content": "Table 5. Analysis of times needed by different algorithms in the multitask setting (using one NVIDIA 1080 Ti GPU and 4 CPUs)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 387, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 447 + ], + "type": "text", + "content": "multi-task meta-learning per-se. These including: studying the multi-task optimisation [77] in meta-learning, studying HPO for meta-learning, developing validation strategies in meta-learning (using ID vs OD val sets [32]), and studying the benefit of task-specific decoders and external data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 460, + 378, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 378, + 474 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 378, + 474 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 482, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 614 + ], + "type": "text", + "content": "We have introduced Meta Omnium, the first multi-task few-shot meta-learning benchmark for computer vision. The benchmark is challenging in multiple highly topical ways including requiring learning on heterogeneous task distributions, evaluating generalization to out-of-distribution datasets, and uniquely challenging meta-learners to learn-to-learn and transfer knowledge across tasks with heterogeneous output spaces. Meta Omnium is nevertheless lightweight enough to be of broad interest and use for driving future research, and even to support future research in hyper-parameter optimization for meta-learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements This work was supported by the UKRI and EPSRC Grant numbers EP/S000631/1, EP/L016427/1 and EP/S02431X/1; the MOD University Defence Research Collaboration (UDRC) in Signal Processing; the Royal Academy of Engineering under the Research Fellowship programme; Samsung AI Center, Cambridge; China Grant No.2020KJ010802 and No.BYSYZHKC2021115." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7700" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "text", + "content": "[1] Milad Abdollahzadeh, Touba Malekzadeh, and Ngai-Man Man Cheung. Revisit multimodal meta-learning through the lens of multi-task learning. In NeurIPS, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 171 + ], + "type": "text", + "content": "[2] Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, and Christian Gagné. Matching feature sets for few-shot image classification. In CVPR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 172, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 287, + 205 + ], + "type": "text", + "content": "[3] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In KDD, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 207, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 207, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 287, + 239 + ], + "type": "text", + "content": "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, 2014. 4, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 241, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 287, + 284 + ], + "type": "text", + "content": "[5] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 287, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 287, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 53, + 287, + 287, + 320 + ], + "type": "text", + "content": "[6] Luca Bertinetto, João F. Henriques, Philip H.S. Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In ICLR, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 322, + 287, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 322, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 53, + 322, + 287, + 344 + ], + "type": "text", + "content": "[7] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 388 + ], + "type": "text", + "content": "[8] Ondrej Bohdal, Lukas Balles, Martin Wistuba, Beyza Ermis, Cedric Archambeau, and Giovanni Zappella. PASHA: Efficient HPO and NAS with progressive resource allocation. In ICLR, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 391, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 424 + ], + "type": "text", + "content": "[9] Ondrej Bohdal, Da Li, and Timothy Hospedales. Feedforward source-free domain adaptation via class prototypes. In ECCV OOD-CV workshop, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[10] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In ICCV, 2019. 4, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 460, + 287, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 494 + ], + "type": "text", + "content": "[11] Ruchika Chavhan, Jan Stuehmer, Calum Heggan, Mehrdad Yaghoobi, and Timothy Hospedales. Amortised invariance learning for contrastive self-supervision. In ICLR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "type": "text", + "content": "[12] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 2017. 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 530, + 287, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 530, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 530, + 287, + 584 + ], + "type": "text", + "content": "[13] Vincent Dumoulin, Neil Houlsby, Utku Evci, Xiaohua Zhai, Ross Goroshin, Sylvain Gelly, and Hugo Larochelle. A unified few-shot classification benchmark to compare transfer and meta learning approaches. In NeurIPS Datasets and Benchmarks, 2021. 2, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 586, + 287, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 619 + ], + "type": "text", + "content": "[14] Nikita Dvornik, Cordelia Schmid, and Julien Mairal. Selecting relevant features from a universal representation for few-shot classification. In ECCV, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 621, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 621, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 621, + 287, + 654 + ], + "type": "text", + "content": "[15] Linus Ericsson, Henry Gouk, and Timothy M. Hospedales. Why do self-supervised models transfer? Investigating the impact of invariance on downstream tasks. BMVC, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "type": "text", + "content": "[16] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 690, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 712 + ], + "type": "text", + "content": "[17] Mario Fritz, E. Hayman, B. Caputo, and J. Eklundh. THE KTH-TIPS database. 2004. 12" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[18] Geetharamani G. and Arun Pandian J. Identification of plant leaf diseases using a nine-layer deep convolutional neural network. Computers and Electrical Engineering, 76:323-338, 2019. 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[19] Ning Gao, Hanna Ziesche, Ngo Anh Vien, Michael Volpp, and Gerhard Neumann. What matters for meta-learning vision regression tasks? In CVPR, 2022. 2, 4, 5, 7, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[20] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "text", + "content": "[21] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In ICLR, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[22] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "type": "text", + "content": "[24] Emily F. Brownlee Heidi M. Sosik, Emily E. Peacock. Annotated plankton images - data set for developing and evaluating classification methods., 2015. 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "type": "text", + "content": "[25] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In ECCV, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[26] Timothy M Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2021. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "text", + "content": "[27] Shell Hu, Da Li, Jan Stuehmer, Minyoung Kim, and Timothy Hospedales. Pushing the limits of simple pipelines for few-shot learning. In CVPR, 2022. 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "text", + "content": "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE Workshop on 3D Representation and Recognition (3dRR-13), 2013. 12" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 489, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 510 + ], + "type": "text", + "content": "[29] Gustaf Kylberg. The Kylberg texture dataset v. 1.0. Technical report, 2011. 12" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[30] Brenden M. Lake, Ruslan Salakhutdinov, and Joshua B. Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 2015. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "text", + "content": "[31] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. A sparse texture representation using local affine regions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(8):1265-1278, Aug. 2005. 12" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 591, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 623 + ], + "type": "text", + "content": "[32] Da Li, Henry Gouk, and Timothy Hospedales. Finding lost DG: Explaining domain generalization via model complexity. In arXiv, 2022. 3, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[33] Liam Li, Kevin Jamieson, Afshin Rostamizadeh, Ekaterina Gonina, Moritz Hardt Hardt, Benjamin Recht, and Ameet Talwalkar. A system for massively parallel hyperparameter tuning. In MLSys, 2020. 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[34] Lisha Li, Kevin G. Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: A novel bandit-based approach to hyperparameter optimization. JMLR, 2018. 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7701" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[35] Rui Li, Ondrej Bohdal, Rajesh K. Mishra, Hyeji Kim, Da Li, Nicholas D. Lane, and Timothy M. Hospedales. A channel coding benchmark for meta-learning. 2021. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[36] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In ICCV, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 142, + 287, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 287, + 175 + ], + "type": "text", + "content": "[37] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. 1, 2, 3, 4, 5, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "type": "text", + "content": "[38] Yanbin Liu, Juho Lee, Linchao Zhu, Ling Chen, Humphrey Shi, and Yi Yang. A multi-mode modulator for multi-domain few-shot classification. In ICCV, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 211, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 243 + ], + "type": "text", + "content": "[39] Changsheng Lu and Piotr Koniusz. Few-shot keypoint detection with uncertainty learning for unseen species. In CVPR, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "type": "text", + "content": "[40] P. Mallikarjuna, Alireza Tavakoli Targhi, Mario Fritz, E. Hayman, B. Caputo, and J. Eklundh. The KTH-TIPS 2 database. 2006. 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 280, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 334 + ], + "type": "text", + "content": "[41] Teresa Mendonça, Pedro M. Ferreira, Jorge S. Marques, André R. S. Marcal, and Jorge Rozeira. PH2 - a dermoscopic image database for research and benchmarking. In Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2013. 4, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 335, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 287, + 368 + ], + "type": "text", + "content": "[42] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In ICCV, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "type": "text", + "content": "[43] Ari Morcos, Maithra Raghu, and Samy Bengio. Insights on representational similarity in neural networks with canonical correlation. Advances in Neural Information Processing Systems, 31, 2018. 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 416, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 447 + ], + "type": "text", + "content": "[44] Jiteng Mu, Weichao Qiu, Gregory D. Hager, and Alan L. Yuille. Learning from synthetic animals. In CVPR, 2020. 4, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 450, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 482 + ], + "type": "text", + "content": "[45] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. Unsupervised domain adaptation: A reality check. In arXiv, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 484, + 287, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 484, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 484, + 287, + 528 + ], + "type": "text", + "content": "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 530, + 287, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 530, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 48, + 530, + 287, + 583 + ], + "type": "text", + "content": "[47] Yoshihiko Ozaki, Yuki Tanigaki, Shuhei Watanabe, and Masaki Onishi. Multiobjective tree-structured parzen estimator for computationally expensive optimization problems. In Genetic and Evolutionary Computation Conference, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 586, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 608 + ], + "type": "text", + "content": "[48] Eunbyung Park and Junier B. Oliva. Meta-Curvature. In NeurIPS, 2019. 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 610, + 287, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 287, + 653 + ], + "type": "text", + "content": "[49] Massimiliano Patacchiola, Jack Turner, Elliot J Crowley, Michael O'Boyle, and Amos J Storkey. Bayesian meta-learning for the few-shot setting via deep kernels. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 655, + 260, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 655, + 260, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 655, + 260, + 666 + ], + "type": "text", + "content": "[50] Gerald Piosenko. 100 sports image classification. 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 668, + 287, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 668, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 668, + 287, + 688 + ], + "type": "text", + "content": "[51] Gerald Piosenko. Birds 400 - species image classification. 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[53] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? Towards understanding the effectiveness of maml. In ICLR, 2020. 2, 3, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "text", + "content": "[54] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 208, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 284 + ], + "type": "text", + "content": "[55] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A generalist agent, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "type": "text", + "content": "[56] Mengye Ren, Eleni Triantafillou, Jake Snell Sachin Ravi, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In ICLR, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 331, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 352 + ], + "type": "text", + "content": "[57] Sebastian Ruder. An overview of multi-task learning in deep neural networks. In arXiv, 2017. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 386 + ], + "type": "text", + "content": "[58] Amirreza Shaban, Shray Bansal, Z. Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. In BMVC, 2017. 4, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "type": "text", + "content": "[59] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "text", + "content": "[60] Haozhe Sun, Wei-Wei Tu, and Isabelle M Guyon. Omniprint: A configurable printed character synthesizer. In NeurIPS Datasets and Benchmarks Track, 2021. 12" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 445, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 466 + ], + "type": "text", + "content": "[61] Sebastian Thrun and Lorien Pratt, editors. Learning to Learn. Kluwer Academic Publishers, 1998. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "text", + "content": "[62] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In ECCV, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "type": "text", + "content": "[63] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, and Hugo Larochelle. Meta-dataset: A dataset of datasets for learning to learn from few examples. In ICLR, 2020. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 569, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 601 + ], + "type": "text", + "content": "[64] Yu-Yun Tseng, Alexander Bell, and Danna Gurari. Vizwiz-fewshot: Locating objects in images taken by people with visual impairments. In ECCV, 2022. 4, 12" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 602, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 678 + ], + "type": "text", + "content": "[65] Ihsan Ullah, Dustin Carrion, Sergio Escalera, Isabelle Guyon, Mike Huisman, Felix Mohr, Jan N. van Rijn, Haozhe Sun, Joaquin Vanschoren, and Phan Anh Vu. Meta-album: Multi-domain meta-dataset for few-shot image classification. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2022. 1, 2, 3, 4, 5, 6, 12, 14" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[66] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NIPS, 2016. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7702" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 710 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[67] Risto Vuorio, Shao-Hua Sun, Hexiang Hu, and Joseph J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. In NeurIPS, 2019. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 287, + 139 + ], + "type": "text", + "content": "[68] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In ICCV, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 172 + ], + "type": "text", + "content": "[69] Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 172, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 172, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 172, + 287, + 215 + ], + "type": "text", + "content": "[70] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 216, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 216, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 49, + 216, + 287, + 248 + ], + "type": "text", + "content": "[71] Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darryll. What should not be contrastive in contrastive learning. In ICLR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 249, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 249, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 49, + 249, + 287, + 281 + ], + "type": "text", + "content": "[72] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 281, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 281, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 49, + 281, + 287, + 325 + ], + "type": "text", + "content": "[73] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 326, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 326, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 49, + 326, + 287, + 357 + ], + "type": "text", + "content": "[74] Yongxin Yang and Timothy Hospedales. A unified perspective on multi-domain and multi-task learning. In ICLR, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 358, + 287, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 358, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 49, + 358, + 287, + 391 + ], + "type": "text", + "content": "[75] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In CVPR, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 392, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 392, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 49, + 392, + 287, + 423 + ], + "type": "text", + "content": "[76] Mingzhang Yin, George Tucker, Mingyuan Zhou, Sergey Levine, and Chelsea Finn. Meta-learning without memorization. 2020. 4, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 424, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 424, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 49, + 424, + 287, + 457 + ], + "type": "text", + "content": "[77] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In NeurIPS, 2020. 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 458, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 458, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 49, + 458, + 287, + 501 + ], + "type": "text", + "content": "[78] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In CORL, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 502, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 502, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 49, + 502, + 287, + 534 + ], + "type": "text", + "content": "[79] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 535, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 535, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 535, + 287, + 599 + ], + "type": "text", + "content": "[80] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. In arXiv, 2019. 2, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "type": "text", + "content": "[81] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 633, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 633, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 633, + 287, + 666 + ], + "type": "text", + "content": "[82] Linyu Zheng, Ming Tang, Yingying Chen, Jinqiao Wang, and Hanqing Lu. Learning feature embeddings for discriminant model based tracking. In ECCV, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 666, + 287, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 666, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 49, + 666, + 287, + 710 + ], + "type": "text", + "content": "[83] Bartosz Zielinski, Anna Plichta, Krzysztof Misztal, Przemysław Spurek, Monika Brzychczy-Wloch, and Dorota Ochońska. Deep learning approach to bacterial colony classification. PLOS ONE, 12(9):1-14, 09 2017. 12" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7703" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_content_list.json b/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..df3375b9a4c8abd87585c18d3e97a42ee09dc2a0 --- /dev/null +++ b/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_content_list.json @@ -0,0 +1,1555 @@ +[ + { + "type": "text", + "text": "Meta-causal Learning for Single Domain Generalization", + "text_level": 1, + "bbox": [ + 200, + 130, + 767, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jin Chen $^{1*}$ , Zhi Gao $^{1*}$ , Xinxiao Wu $^{1,2\\dagger}$ , Jiebo Luo $^{3}$", + "bbox": [ + 282, + 179, + 683, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science & Technology, Beijing Institute of Technology, China $^{2}$ Guangdong Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, China", + "bbox": [ + 156, + 199, + 833, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Computer Science, University of Rochester, Rochester NY 14627, USA", + "bbox": [ + 138, + 268, + 828, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chen_jin, gaozhi_2017, wuxinxiao}@bit.edu.cn, jluo@cs.rochester.edu", + "bbox": [ + 197, + 289, + 767, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single domain generalization aims to learn a model from a single training domain (source domain) and apply it to multiple unseen test domains (target domains). Existing methods focus on expanding the distribution of the training domain to cover the target domains, but without estimating the domain shift between the source and target domains. In this paper, we propose a new learning paradigm, namely simulate-analyze-reduce, which first simulates the domain shift by building an auxiliary domain as the target domain, then learns to analyze the causes of domain shift, and finally learns to reduce the domain shift for model adaptation. Under this paradigm, we propose a meta-causal learning method to learn meta-knowledge, that is, how to infer the causes of domain shift between the auxiliary and source domains during training. We use the meta-knowledge to analyze the shift between the target and source domains during testing. Specifically, we perform multiple transformations on source data to generate the auxiliary domain, perform counterfactual inference to learn to discover the causal factors of the shift between the auxiliary and source domains, and incorporate the inferred causality into factor-aware domain alignments. Extensive experiments on several benchmarks of image classification show the effectiveness of our method.", + "bbox": [ + 76, + 371, + 473, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single domain generalization [28] aims to generalize a model trained using one training domain (source domain) into multiple unseen test domains (target domains). Since only one source domain is given and the target domains are out-of-distribution and unavailable during training, single", + "bbox": [ + 75, + 787, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3fce0e146fb971161d9a04df4e696f4ab35f465b42c320aebb5423b5996db95.jpg", + "image_caption": [ + "Figure 1. Illustration of the simulate-analyze-reduce paradigm. In this paradigm, we first simulate the domain shift by constructing an auxiliary domain as the unseen target domain, then learn to analyze the domain shift, and finally learn to reduce the domain shift based on inferred causes." + ], + "image_footnote": [], + "bbox": [ + 514, + 338, + 880, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "domain generalization is a challenging task and attracts increasing interests. Existing works have made considerable successes through expanding the distribution of the source domain by data augmentation [19,28,34] or learning adaptive data normalization [8] typically. However, such successes have been achieved without explicitly considering the domain shift between the source and target domains, which limits the generalization performance of model in real-world scenarios.", + "bbox": [ + 496, + 611, + 890, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a new learning paradigm, namely simulate-analyze-reduce, to address single domain generalization by enabling the model to analyze the real domain shift between the source domain and unseen target domain. This new paradigm is shown in Figure 1. We first build an auxiliary domain as the target domain to simulate the real domain shift between the source and target domains, since the target data is unavailable during training. We then learn to analyze the intrinsic causal factors of the domain shift to facilitate the subsequent model adapta", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Jin Chen and Zhi Gao are co-first authors.", + "bbox": [ + 99, + 875, + 334, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: Xinxiao Wu.", + "bbox": [ + 99, + 887, + 303, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7683", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. Finally, we learn to reduce the domain shift with its inferred causes.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Under this paradigm, we propose a meta-causal learning method to learn the meta-knowledge about how to infer the causes of the simulated domain shift between the auxiliary and source domains via causal inference in training, and then apply the meta-knowledge to analyze the real domain shift between the target and source domains during testing, through which the source and given target domains are adaptively aligned. Specifically, we perform multiple transformations on source data to generate an auxiliary domain with great diversity. Then we build a causal graph to represent the dependency among data, variant factors, semantic concepts and category labels, and conduct counterfactual inference over the causal graph to exploit the intrinsic causality of the simulated domain shift between the auxiliary and source domains. For each sample in the auxiliary domain, we construct counterfactual scenes by intervening variant factors to infer their causal effects on the category prediction, and these inferred causal effects of variant factors can be regarded as the causes of domain shift. To reduce the domain shift, we propose a factor-aware domain alignment by learning and integrating multiple feature mappings, where an effect-to-weight network is designed to convert the causal effects of variant factors into the weights of feature mappings.", + "bbox": [ + 76, + 121, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During testing, the distribution discrepancy between the input target sample and the source domain is analyzed and reduced by applying the learnt meta-knowledge, i.e., inferring the causal effects of variant factors and incorporating them into the factor-aware domain alignment. In summary, the main contributions of this paper are as follows:", + "bbox": [ + 76, + 484, + 472, + 575 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel learning paradigm, simulate-analyze-reduce, for single domain generalization. This paradigm empowers the model with the ability to estimate the domain shift between the source domain and unseen target domains, thus boosting the model adaptation across different domains.", + "- We propose a meta-causal learning method based on counterfactual inference to learn the meta-knowledge about analyzing the intrinsic causality of domain shift, thus facilitating the reduction of domain shift.", + "- Our method achieves the state-of-the-art results on several benchmarks of image classification, especially on the more challenging tasks with a large domain shift, clearly demonstrating the effectiveness of our method." + ], + "bbox": [ + 94, + 580, + 468, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 821, + 218, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Domain Generalization", + "text_level": 1, + "bbox": [ + 76, + 847, + 292, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain generalization focuses on generalizing a model learned from multiple source domains to the unseen target", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "domain. The key difference between domain adaptation and domain generalization is that during training, domain adaptation leverages unlabelled target data while domain generalization has no access to the target domain. Existing domain generalization methods can be roughly divided into two categories: learning domain-invariant feature representation from multiple source domains [7, 10, 24, 25, 32] and generating diverse more samples via data augmentation [2, 29, 31, 39].", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, single domain generalization [28] has attracted growing attention, where only one source domain is available during training and the model is evaluated on multiple unseen target domains. A rich line of works employ data augmentation for generating out-of-domain samples to expand the distribution of the source domain [19, 28, 34]. Qiao et al. [28] propose meta-learning based adversarial domain augmentation to generate samples. Li et al. [19] propose a progressive domain expansion network to generate multiple domains progressively via simulating various photometric and geometric transforms by style transfer based generators. Wang et al. [34] propose a style-complement module to generate diverse images with different styles. Fan et al. [8] use data normalization for single domain generalization, where an adaptive normalization scheme is learned to be incorporated with adversarial domain augmentation to enhance the generalization of the model.", + "bbox": [ + 496, + 228, + 892, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The aforementioned methods aim to generate the source data distribution as diverse as possible to cover unseen target domains. When the expanded source distribution does not approximate the target distribution, the performance may significantly degrade, since there still exists a domain gap between the source and target domains. To address this problem, our method learns to analyze and reduce the domain shift by building an auxiliary domain during training.", + "bbox": [ + 496, + 500, + 890, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Causality for Domain Generalization", + "text_level": 1, + "bbox": [ + 500, + 633, + 818, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several recent methods exploit causality to learn domain-invariant semantic representation for domain generalization [20-22]. Considering the cross-domain invariance of the causality between semantic factors and predictions, Liu et al. [20] propose a causal semantic generative model to remove the domain-specific correlation between semantic factors and variant factors, and thus make the prediction affected only by the semantic factors. Assuming that images of the same object across domains should have the same representation, Mahajan et al. [22] use the cross-domain invariance of the causality between objects and feature representations to capture the within-class variation for domain generalization. Lv et al. [21] introduce causal inference to extract causal factors that are invariant across domains in order to learn invariant feature representation.", + "bbox": [ + 496, + 657, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From a different perspective, our method focuses on uti-", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7684", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/05f48d46008de69a045f19d16947f7fbf8f250b695297fdf44e5dd76ffbcc0fa.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed meta-causal learning method. (1) A data transformation module simulates the domain shift via generating an auxiliary domain $\\mathcal{D}_a$ . (2) A counterfactual inference module analyzes the domain shift by inferring the causal effects of $V$ on $Y$ . “-” denotes comparing the values of $Y$ before and after do-operation. The edge $X \\rightarrow Y$ is a dashed arrow to represent the classification model $P(Y|X)$ . (3) A factor-aware domain alignment module reduces the domain shift via multiple feature mappings according to weights learned by an effect-to-weight network. (4) A base module is used for feature extraction and classification." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 861, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lizing causal inference to discover the intrinsic causes of domain shift by constructing counterfactual scenes over the learnt causal graph, which facilitates the reduction of domain shift and in turn benefits cross-domain adaptation.", + "bbox": [ + 75, + 469, + 470, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 545, + 168, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 571, + 259, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For single domain generalization, during training, we are given a labeled source domain $\\mathcal{D}_s = \\{(x_i^s,y_i^s)|_{i = 1}^{N_s}\\}$ drawn from the distribution $P_{s}$ , where $\\pmb{x}_i^s$ is the $i$ -th source sample with its category label $y_{i}^{s}\\in \\mathcal{V}$ . During testing, the learnt model is applied to multiple unseen target domains $\\mathcal{D}_t = \\{\\mathcal{D}_t^j\\}_{j = 1}^J$ , where $\\mathcal{D}_t^j$ is the $j$ -th target domain drawn from the distribution $P_{t}^{j}$ , and $P_{t}^{j}\\neq P_{s}$ .", + "bbox": [ + 76, + 595, + 468, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To bridge the domain gap between the source domain and multiple unseen target domains, we propose a new learning paradigm, called simulate-analyze-reduce, which starts with simulating the domain shift between the source and target domains, then learns to analyze the domain shift, and finally learns to reduce the domain shift. Under this paradigm, we propose a meta-causal learning method that has four components: a data transformation module to generate auxiliary domains as target domains, a counterfactual inference module to discover the causes of the domain shift, a factor-aware domain alignment module to reduce the domain shift, and a base module for feature extraction and classification, as illustrated in Figure 2.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Transformation for Domain Shift Simulation", + "text_level": 1, + "bbox": [ + 500, + 468, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To simulate the real domain shift between the source and target domains, we generate an auxiliary domain $\\mathcal{D}_a$ as the unseen target domain by performing transformations on source data. The domain shift is actually the data distribution discrepancy between the source and auxiliary domains, and is usually caused by the variations of extrinsic attributes of data, independent of intrinsic semantics. Taking the image data for example, the domain shift is mainly caused by the variations of visual attributes, such as the variations of brightness, viewpoint, and color. Therefore, to make the simulated domain shift as realistic as possible, we formulate the extrinsic attributes as variant factors, and design data transformations according to the variant factors.", + "bbox": [ + 496, + 508, + 892, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define a set of variant factors, denoted as $\\mathcal{V} = \\{v_{1}, v_{2}, \\dots, v_{K}\\}$ , where $v_{k}$ denotes the $k$ -th variant factor. Each variant factor corresponds to a data transformation function that aims to generate new data by making changes on the corresponding extrinsic attribute, denoted as $G_{v_{k}}(\\boldsymbol{x}; \\theta_{v_{k}})$ , where $\\boldsymbol{x}$ is an input sample, and $\\theta_{v_{k}} \\in [g_{min}^{k}, g_{max}^{k}]$ represents the degree parameter to control the magnitude of transformation with the scale range $[g_{min}^{k}, g_{max}^{k}]$ .", + "bbox": [ + 496, + 704, + 890, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In real-world scenarios, the complex domain shift is often caused by combinatorial multiple variant factors and accordingly we design a sampling strategy to enable the combinatorial data transformation. Given a source sample $\\boldsymbol{x}_i^s$ ,", + "bbox": [ + 496, + 840, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7685", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we randomly sample $N_{v}^{i}$ variant factors from $\\mathcal{V}$ to form a factor subset $\\mathcal{V}_i = \\{v_1^i,v_2^i,\\dots ,v_{N_v^i}^i\\}$ . Then, a corresponding auxiliary sample $\\pmb{x}_i^a$ is generated by", + "bbox": [ + 76, + 90, + 468, + 138 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {i} ^ {a} = G _ {v _ {N _ {v} ^ {i}} ^ {i}} \\left(\\dots G _ {v _ {2} ^ {i}} \\left(G _ {v _ {1} ^ {i}} \\left(\\boldsymbol {x} _ {i} ^ {s}; \\theta_ {v _ {1} ^ {i}}\\right); \\theta_ {v _ {2} ^ {i}}\\right) \\dots ; \\theta_ {v _ {N _ {v} ^ {i}} ^ {i}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 148, + 468, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the transformation degree parameter $\\theta_{v_k^i}$ is randomly selected from its range scale. In this way, the domain shift is generated as diverse as possible to approximate the real domain shift.", + "bbox": [ + 76, + 188, + 468, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Counterfactual Inference for Domain Shift Analysis", + "text_level": 1, + "bbox": [ + 76, + 257, + 468, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After simulating the domain shift, we introduce counterfactual inference to analyze the domain shift. During training, we learn the meta-knowledge about how to infer the causes of data discrepancy between one auxiliary sample and the source domain, and during testing, we apply the learnt meta-knowledge to unseen target samples.", + "bbox": [ + 76, + 296, + 468, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We build a causal graph to model the causal dependency among the input sample (node $X$ ), the variant factors (node $V$ ), the semantic concepts (node $S$ ), and the output category (node $Y$ ), as shown in Figure 2 (2). The semantic concepts denote the intrinsic attributes of data that are related to the category, and the variant factors denote the extrinsic attributes of data that are domain specific, independent of the intrinsic semantics. For example, when the input sample is an image of \"zebra\", the semantic concepts are like \"four legs\" and \"black-white stripes\", and the variant factors include brightness, viewpoint, and so on. The edge $S \\rightarrow Y$ represents that the category of the input data is determined by the semantic concepts. The edges $S \\rightarrow X$ and $V \\rightarrow X$ represent that the semantic concepts and the variant factors together determine what the input sample looks like. Since the semantic concepts represent the intrinsic semantics and are invariant across different domains, ideally the cross-domain classification is only determined by the semantic concepts, which is implemented by learning the edge $S \\rightarrow Y$ , i.e., estimating the conditional probability $P(Y|S)$ . However, in reality, the semantic concepts are unobserved from the input samples, so the classification should be implemented by learning the edge $X \\rightarrow Y$ , i.e., estimating the conditional probability $P(Y|X)$ . The edge $X \\rightarrow Y$ is a dashed arrow to represent the classification model $P(Y|X)$ . Since the input sample node $X$ is affected by the semantic concept node $S$ and the variant factor node $V$ together, the category node $Y$ is also affected by the variant factor node $V$ through edges $V \\rightarrow X \\rightarrow Y$ . As the variant factors are domain-specific, their causal effects on the category lead to the domain shift. Hence, we infer the causal effects of the variant factors on the category prediction to discover the causes of the domain shift. That is to say, we infer the causal effects of the node $V$ on the node", + "bbox": [ + 76, + 387, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$Y$ as causes of domain shift, and then learn to reduce the domain shift based on its causes.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To infer the causal effects of the node $V$ on the node $Y$ , we first learn the edge $X \\to Y$ using the source data $\\mathcal{D}_s$ by a classification loss:", + "bbox": [ + 498, + 119, + 890, + 165 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\sim P _ {s}} \\left[ - \\sum_ {u} \\mathbb {I} _ {u = y _ {i} ^ {s}} \\log C \\left(F \\left(\\boldsymbol {x} _ {i} ^ {s}\\right)\\right) _ {u} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 172, + 890, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(\\pmb{x}_i^s, y_i^s) \\in \\mathcal{D}_s$ , $F$ is a feature extractor, $C$ is the classifier to output $|\\mathcal{V}|$ category probabilities, $C(\\cdot)_u$ is the $u$ -th element of $|\\mathcal{V}|$ category probabilities, and $\\mathbb{I}_{u=y_i^s}$ is an indicator function, meaning that if $u = y_i^s$ , the value of $\\mathbb{I}_{u=y_i^s}$ is 1 and 0 otherwise. Then, for each sample from the auxiliary domain, we construct a factual scene and multiple counterfactual scenes over the causal graph, so as to infer the causal effects of variant factors on the category prediction. Given an auxiliary sample $\\pmb{x}_i^a \\in \\mathcal{D}_a$ , its factual category is predicted by", + "bbox": [ + 498, + 224, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {i} ^ {a} = P (Y | X) = C \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 387, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{y}_i^a \\in \\mathbb{R}^{|\\mathcal{Y}|}$ is a category probability vector, and represents the value of node $Y$ . For each variant factor, we construct a counterfactual scene to infer its causal effect by doing intervention on the variant factor node $V$ . Let $do(V = v_k)$ denote the intervention on the node $V$ , which is implemented by changing extrinsic attributes of data through the transformation $G_{v_k}$ with multiple degree parameters. Accordingly, the counterfactual category of $\\pmb{x}_i^a$ is predicted by", + "bbox": [ + 498, + 412, + 890, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {y} _ {i, v _ {k}} ^ {a} = P \\left(Y | X, d o (V = v _ {k})\\right) \\\\ = \\frac {1}{| \\mathcal {M} |} \\sum_ {\\theta_ {v _ {k}} \\in \\mathcal {M}} C \\left(F \\left(G _ {v _ {k}} \\left(\\boldsymbol {x} _ {i} ^ {a}; \\theta_ {v _ {k}}\\right)\\right)\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 560, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is a set of degree parameters, obtained by uniformly sampling magnitudes of transformation $G_{v_k}$ from the scale range. By comparing the factual and counterfactual category probabilities of $\\pmb{x}_i^a$ , the causal effect of the $k$ -th variant factor is calculated by", + "bbox": [ + 498, + 628, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {e} _ {k} ^ {i} = P (Y | X) - P \\left(Y | X, d o (V = v _ {k})\\right) = \\boldsymbol {y} _ {i} ^ {a} - \\boldsymbol {y} _ {i, v _ {k}} ^ {a}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 715, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The inferred causal effect represents how much contribution the corresponding variant factor makes to the domain shift, and the larger the causal effect is, the more seriously the domain shift is caused by the variant factor.", + "bbox": [ + 498, + 747, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Factor-aware Domain Alignment for Domain Shift Reduction", + "text_level": 1, + "bbox": [ + 500, + 816, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After analyzing the causes of domain shift, we propose a factor-aware domain alignment to reduce the domain shift by learning multiple feature mappings, with guidance of the", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7686", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8b18c8e4206dffb7d06e3e53f5b175ba03594c1f3d3fe28e73191aa432b087a8.jpg", + "image_caption": [ + "Figure 3. Inference process of meta-causal learning for a given target sample." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 872, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "inferred causal effects of variant factors. Each feature mapping addresses a specific domain shift caused by one variant factor. We construct $K$ feature mappings for $K$ variant factors, and the $k$ -th feature mapping aims to address the domain shift caused by the $k$ -th variant factor. In order to incorporate the causal effects of variant factors into the learning of mappings, we build an effect-to-weight network that converts the causal effect of each variant factor into the weight of the corresponding feature mapping.", + "bbox": [ + 75, + 297, + 468, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the auxiliary sample $\\pmb{x}_i^a$ and its inferred causal effects of all variant factors $\\{e_1^i, e_2^i, \\dots, e_K^i\\}$ , the weights of feature mappings are calculated by", + "bbox": [ + 76, + 431, + 470, + 478 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\omega} ^ {i} = \\operatorname {s o f t m a x} \\big (W (\\boldsymbol {e} _ {1} ^ {i}), W (\\boldsymbol {e} _ {2} ^ {i}), \\dots , W (\\boldsymbol {e} _ {K} ^ {i}) \\big), \\quad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 484, + 468, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega^i\\in \\mathbb{R}^K$ and its $k$ -th element $\\omega_{k}^{i}$ denotes the weight of the $k$ -th factor-aware feature mapping. $W(\\cdot)$ denotes the effect-to-weight network.", + "bbox": [ + 75, + 510, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to the weights of feature mappings, we integrate the $K$ feature mappings to project the auxiliary samples into the source feature space. The alignment of the source and auxiliary domains is implemented by minimizing the feature distance between the source and auxiliary samples in the source feature space. Given a source sample $\\pmb{x}_i^s$ with the corresponding category label $y_i^s$ , generated auxiliary sample $\\pmb{x}_i^a$ and inferred mapping weights $\\omega^i$ , the alignment loss incorporated with inferred causal effects is defined as", + "bbox": [ + 75, + 556, + 470, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {a} ^ {c} = \\frac {1}{N _ {s}} \\sum_ {i} | | F \\left(\\boldsymbol {x} _ {i} ^ {s}\\right) - \\sum_ {k} \\omega_ {k} ^ {i} M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right) | | _ {2} \\tag {7} \\\\ + \\frac {1}{N _ {s}} \\sum_ {i} \\mathcal {H} \\left(C \\left(\\sum_ {k} \\omega_ {k} ^ {i} M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right), y _ {i} ^ {s}\\right)\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 712, + 468, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F$ is the feature extractor, $M_{k}$ is the $k$ -th feature mapping, $C$ is the classifier, $\\omega_{k}^{i}$ is the $k$ -th element of $\\omega^i$ , $\\sum_{k}\\omega_{k}^{i}M_{k}\\big(F(\\pmb{x}_{i}^{a})\\big)$ represents the projected feature of $\\pmb{x}_i^a$ , and $N_{s}$ is the number of source samples. $||\\cdot ||_2$ is the L2-loss, and $\\mathcal{H}(\\cdot)$ is the cross-entropy loss. The first term measures the feature distance between the source and auxiliary samples after projection, i.e., the data distribution discrepancy", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "between the source and auxiliary domains. The second term encourages the auxiliary samples to belong to the same categories as the source samples.", + "bbox": [ + 496, + 297, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, to enable each feature mapping to address the specific domain shift caused by the corresponding variant factor, we introduce another alignment loss $\\mathcal{L}_a^m$ :", + "bbox": [ + 496, + 342, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {a} ^ {m} = \\frac {1}{N _ {s}} \\frac {1}{K} \\sum_ {i} \\sum_ {k} | | F (\\boldsymbol {x} _ {i} ^ {s}) - M _ {k} (F (\\boldsymbol {x} _ {i} ^ {k})) | | _ {2} \\\\ + \\frac {1}{N _ {s}} \\frac {1}{K} \\sum_ {i} \\sum_ {k} \\mathcal {H} \\left(C \\left(M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {k}\\right)\\right), y _ {i} ^ {s}\\right)\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 397, + 890, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}_i^k$ is an sample generated by conducting the data transformation $G_{v_k}$ on the source sample $\\pmb{x}_i^s$ , denoted by $\\pmb{x}_i^k = G_{v_k}(\\pmb{x}_i^s;\\theta_{v_k})$ . The transformation degree parameter $\\theta_{v_k}$ is randomly selected from its range scale.", + "bbox": [ + 496, + 484, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the overall loss function is defined as", + "bbox": [ + 517, + 547, + 803, + 560 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {a} ^ {c} + \\mathcal {L} _ {a} ^ {m}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 573, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The whole training process is summarized in Algorithm 1.", + "bbox": [ + 500, + 601, + 883, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inference", + "text_level": 1, + "bbox": [ + 500, + 626, + 609, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During testing, given a target sample $\\boldsymbol{x}_i^t$ , firstly, the causal effects of variant factors are inferred by counterfactual inference in Eq. (5), and the weights $\\omega^i$ of feature mappings are calculated by the effect-to-weight network in Eq. (6). Then the target feature $\\boldsymbol{f}_i^t = F(\\boldsymbol{x}_i^t)$ is projected into the source feature space by integrating $K$ feature mappings according to their weights $\\omega^i$ to obtain the projected feature $\\hat{\\boldsymbol{f}}_i^t = \\sum_k \\omega_k^i M_k(\\boldsymbol{f}_i^t)$ . Finally, the category of $\\boldsymbol{x}_i^t$ is predicted by $C(\\hat{\\boldsymbol{f}}_i^t)$ . The whole inference process is shown in Figure 3.", + "bbox": [ + 496, + 648, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 821, + 630, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 847, + 601, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Digits. The Digits dataset consists of five datasets: MNIST [16], MNIST-M [9], SVHN [26], USPS [13], and", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7687", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Training process", + "text_level": 1, + "bbox": [ + 80, + 90, + 279, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: The source domain $\\mathcal{D}_s$ , the variant factor set $\\nu$ .", + "bbox": [ + 80, + 109, + 442, + 125 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output: The feature extractor $F$ , the classifier $C$ , $K$ feature mappings $\\{M_i\\}_{i=1}^K$ , the effect-to-weight network $W$ .", + "bbox": [ + 80, + 126, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $F, C, \\{M_i\\}_{i=1}^K, W$ ;", + "2: while not converge do", + "3: Sample an image $\\pmb{x}_j^s$ from $\\mathcal{D}_s$ ;", + "4: Construct a factor subset $\\mathcal{V}_i$ by sampling $N_v^i$ variant factors from $\\mathcal{V}$ ;", + "5: Generate an auxiliary sample $\\pmb{x}_i^a$ by Eq.(1);", + "6: for $v_{k}$ in $\\mathcal{V}$ do", + "7: Calculate the factual category $\\mathbf{y}_i^a$ of auxiliary sample $\\mathbf{x}_i^a$ by Eq.(3);", + "8: Calculate the counterfactual category $\\pmb{y}_{i,v_k}^a$ of auxiliary sample $\\pmb{x}_i^a$ by Eq.(4);", + "9: Infer the causal effect $e_k^i$ of variant factor $v_k$ via comparing $y_i^a$ and $y_{i,v_k}^a$ by Eq.(5);" + ], + "bbox": [ + 88, + 170, + 468, + 368 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "10: end for", + "11: Calculate the weights of $K$ feature mappings $\\{M_i\\}_{i=1}^K$ by Eq.(6);", + "12: Update $F, C, \\{M_i\\}_{i=1}^K$ , $W$ by Eq.(9).", + "13: end while" + ], + "bbox": [ + 83, + 369, + 468, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SYN [9], with 10 categories. We use MNIST as the source domain, and the other four datasets as the target domains. The first 10,000 images in the training set of MNIST are used for training.", + "bbox": [ + 75, + 470, + 468, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR10-C. The CIFAR10-C dataset [11] is proposed to evaluate the robustness of classification model. The images are corrupted from the test set of the CIFAR10 dataset [15] by 19 corruption types with five levels of severity. A higher level means the more serious corruption. There are 10 categories. We use CIFAR10 as the source domain, and CIFAR10-C as the target domains where images of one severity level form one target domain.", + "bbox": [ + 75, + 532, + 468, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PACS. The PACS dataset [17] is a benchmark for domain generalization, and consists of four domains: art painting, cartoon, photo, and sketch. There are 9,991 images of seven categories. We use one domain as the source domain, and the rest three domains as the target domains. So there are four tasks with different domains as the source domains.", + "bbox": [ + 75, + 654, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 755, + 292, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Auxiliary Domain. We define 16 variant factors to generate the images in the auxiliary domain, including 12 photometric factors (Brightness, Contrast, Color, Sharpness, Auto-Contrast, Invert, Equalize, Solarize, SolarizeAdd, Posterize, NoiseSalt, NoiseGaussian) and 4 geometric factors (Shear-X, Shear-Y, Rotate, Flip). Since the Rotate and Flip variant factors will affect the semantic information of digit images, we use the other 14 variant factors for the Digits dataset. For", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c83627f752f99db53bae3dc99f6b0875998600831aec3b49bebdbe930782e82f.jpg", + "table_caption": [ + "Table 1. Single domain generalization results (%) on Digits with ConvNet as backbone. The model is trained on MNIST, and evaluated on SVHN, SYN, MNIST-M, and USPS." + ], + "table_footnote": [], + "table_body": "
MethodSVHNSYNMNIST-MUSPSAvg
ERM [14]27.8339.6552.7276.9449.29
CCSA [24]25.8937.3149.2983.7249.05
d-SNE [36]26.2237.8350.9893.1652.05
JiGen [2]33.8043.7957.8077.1553.14
GUD [31]35.5145.3260.4177.2654.62
M-ADA [28]42.5548.9567.9478.5359.49
ME-ADA [37]42.5650.3963.2781.0459.32
PDEN [19]62.2169.3982.2085.2674.77
L2D [34]62.8663.7287.3083.9774.46
AA [4]45.2364.5260.5380.6262.72
RA [5]54.7759.6074.0577.3366.44
RSDA [30]47.4062.0081.5083.1068.50
RSDA+ASR [8]52.8064.5080.8082.4070.10
Ours69.9478.4778.3488.5478.82
", + "bbox": [ + 501, + 141, + 898, + 345 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/efad143171c7e70ca496d5743e8b99903c56e24b194f945b116f2ba27eb1fe0e.jpg", + "table_caption": [ + "Table 2. Single domain generalization results (\\%) on CIFAR10-C with WRN as backbone. Each level is viewed as a target domain, a higher level denotes the more serious corruption and the domain discrepancy between the source and target domains is larger." + ], + "table_footnote": [], + "table_body": "
Methodlevel1level2level3level4level5Avg
ERM [14]87.8081.5075.5068.2056.1073.82
GUD [31]88.3083.5077.6070.6058.3075.66
M-ADA [28]90.5086.8082.5076.4065.6080.36
PDEN [19]90.6288.9187.0383.7177.4785.55
AA [4]91.4287.8884.1078.4671.1382.60
RA [5]91.7488.8985.8281.0374.9384.48
Ours92.3891.2289.8887.7384.5289.15
", + "bbox": [ + 501, + 426, + 898, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the CIFAR10-C and PACS datasets, all 16 variant factors are used. To make the auxiliary domain as diverse as possible, for each source image at each iteration, we randomly sample several variant factors and use them to generate a new auxiliary image.", + "bbox": [ + 496, + 563, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on Single Domain Generalization", + "text_level": 1, + "bbox": [ + 498, + 648, + 849, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with several state-of-the-art methods, including the baseline method (ERM [14]), the methods of learning domain-invariant features (CCSA [24], d-SNE [36], JiGen [2]), and the methods of making data augmentation (GUD [31], M-ADA [28], ME-ADA [37], PDEN [19], L2D [34], AA [4], RA [5], RSDA [30], RSC [12], ASR [8]).", + "bbox": [ + 496, + 672, + 890, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1, Table 2, and Table 3 show the comparison results on Digits, CIFAR10-C, and PACS, respectively. From the results, there are several interesting observations as follows. First, our method generally outperforms the compared methods on all datasets, which clearly shows the effectiveness of the proposed simulate-analyze-reduce learning paradigm for single domain generalization. Second, compared with the data augmentation methods (e.g.,", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7688", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/02733cade7e4371e1a4a7ca6b936e0596636556edc641ac7d6c643eb4c068b1d.jpg", + "table_caption": [ + "Table 3. Single domain generalization results (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as the target domains." + ], + "table_footnote": [], + "table_body": "
MethodArtpaintCartoonSketchPhotoAvg
ERM [14]70.9076.5053.1042.2060.70
RSC [12]73.4075.9056.2041.6061.80
RSC+ASR [8]76.7079.3061.6054.6068.10
Ours77.1380.1462.5559.6069.86
", + "bbox": [ + 78, + 155, + 470, + 228 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/26d50dcfce57670383e1b8e2156f3e17e8ea3aef3562d0c6c31fd2a773a257f1.jpg", + "table_caption": [ + "Table 4. Leave-one-domain-out results (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the target domain and the other three domains are used as source domains." + ], + "table_footnote": [], + "table_body": "
MethodArtpaintCartoonPhotoSketchAvg
MetaReg [1]83.7077.2095.5070.3081.70
GUD [31]78.3277.6595.6174.2181.44
Epi-FCR [18]82.1077.0093.9073.0081.50
MASF [6]80.2977.1794.9971.6881.03
JiGen [2]79.4275.2596.0371.3580.51
DMG [3]76.9080.3893.5575.2181.46
DDAIG [38]84.2078.1095.3074.7083.10
CSD [27]78.9075.8094.1076.7081.40
L2A-OT [39]83.3078.2096.2073.6082.80
EISNet [33]81.8976.4495.9374.3382.15
RSC [12]83.4380.3195.9980.8585.15
ME-ADA [37]78.6178.6595.5775.5982.10
MMLD [23]81.2877.1696.0972.2981.83
L2D [34]81.4479.5695.5180.5884.27
FACT [35]85.3778.3895.1579.1584.51
MatchDG [22]81.3280.7096.5379.7284.57
CIRL [21]86.0880.5995.9382.6786.32
Ours85.3080.9396.5385.2487.00
", + "bbox": [ + 81, + 310, + 475, + 570 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8abe1ca85421b46cbd10d92b95c93df0c5d1397d88fdf90b5bd05a4b35f8c03e.jpg", + "table_caption": [ + "Table 5. Ablation study (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as target domains. “T”, “A”, “C” denote Domain Transformation, Domain Alignment, and Counterfactual Inference, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTACArtpaintCartoonSketchPhotoAvg
Base71.2667.6443.9736.9954.97
DT✓75.2878.4659.4556.0967.32
DTA✓✓71.6472.7857.1152.0263.39
Ours✓✓✓77.1380.1462.5559.6069.86
", + "bbox": [ + 81, + 667, + 475, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AA [4], RA [5] and RSDA [30]) that are more related to our method, our method achieves much better results, and especially yields a $10.32\\%$ gain over RSDA on Digits, strongly suggesting that it is beneficial to empower the model with the ability of analyzing the causes of domain shift by counterfactual inference. Third, on more difficult tasks with larger domain shift (e.g., SVHN on Digits, level5 on CIFAR10-C, and Photo on PACS), our method significantly improves the performance, further demonstrating the", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "superiority of our method on handling more challenging situations. Forth, our method performs little worse on MNIST-M and USPS of Digits. The possible reason is that other compared methods use more well-designed data augmentation and network regularization, such as AdaIN based generators in PDEN [19] and stochastic neighborhood embedding techniques in d-SNE [36].", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results on Multiple Domain Generalization", + "text_level": 1, + "bbox": [ + 498, + 207, + 870, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We extend the proposed method to multi-source domain setting by regarding the multiple source domains as one source domain without using domain labels. We employ the leave-one-domain-out protocol following existing multi-source domain generalization [21, 35]. We compare our method with most related methods that introduces causal inference into domain generalization (MatchDG [22], CIRL [21]), and existing popular domain generalization methods (MetaReg [1], GUD [31], EpiFCR [18], MASF [6], JiGen [2], DMG [3], DDAIG [38], CSD [27], L2A-OT [39], EISNet [33], RSC [12], ME-ADA [37], MMLD [23], L2D [34], FACT [35]).", + "bbox": [ + 496, + 231, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 shows the leave-one-domain-out results on the PACS dataset with ResNet-18 as backbone. From the results, we make several observations. First, it is noteworthy that our method achieves the state-of-the-art overall performance (\"Avg\") although our method is not designed for multi-source domain generalization, which further demonstrates that the proposed simulate-analyze-reduce learning paradigm not only benefits single domain generalization but also boosts multi-source domain generalization. Second, compared with the methods of introducing causal inference to learn domain-invariant features (MatchDG [22], CIRL [21]), our method achieves better results on the overall metric \"Avg\", especially making $5.52\\%$ and $2.57\\%$ gains over MatchDG [22] and CIRL [21], respectively, on the more challenging task (Sketch $\\rightarrow$ Others). Such improvements are attributed to the ability of analyzing and reducing the domain shift, further demonstrating the advantages of causal inference in analyzing the causes of the domain shift.", + "bbox": [ + 496, + 412, + 890, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 710, + 663, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate each component of our method, we conduct ablation experiments on the PACS dataset. We design several degraded variants of our method for comparison: (1) \"Base\", where only the base module is utilized and optimized by Eq. (2) using the source domain; (2) \"DT\", where the data transformation module is added into the base module and the model is trained using both source and auxiliary domains; (3) \"DTA\", where the simulated domain shift between auxiliary domain and source domain is directly reduced without analyzing the causes of the domain shift via the proposed counterfactual module.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa0dca8fe5dba2c2c953509f91624b91296fab9d7e5aef196994c4972cabb1ec.jpg", + "image_caption": [ + "Figure 4. Examples of the inferred causal effects (represented as weights) of variant factors. The left part shows a source image from the CIFAR10 dataset and three target images from the CIFAR10-C dataset with Gaussian noise corruption. As the corruption severity increases from level 1 to level 5, the inferred weights of the NoiseGaussian variant factor become larger accordingly." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 849, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8ab6985fa502123931c98deec2109422f8262bb9c32c363713f179215b55f0ba.jpg", + "table_caption": [ + "Table 6. Factor analysis (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as target domains." + ], + "table_footnote": [], + "table_body": "
MethodArtpaintCartoonSketchPhotoAvg
Ours_GT74.6269.7852.7443.3460.12
Ours_PT74.4773.4850.3551.6762.49
Ours77.1380.1462.5559.6069.86
", + "bbox": [ + 81, + 362, + 475, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are shown in Table 5. From the results, we make several observations. First, our method achieves better performance than \"DT\", which validates the superiority of our method on analyzing and reducing the domain shift, rather than directly enlarging the source data distribution. Second, \"DTA\" performs worse than \"DT\", probably because brute-force domain alignment without analyzing causes of the domain shift leads to negative transfer and thus hurts the performance. Third, our method achieves the best performance thanks to the proposed simulate-analyze-reduce paradigm.", + "bbox": [ + 75, + 455, + 470, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Factor Analysis", + "text_level": 1, + "bbox": [ + 76, + 633, + 235, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to further analyze the effect of variant factors, we design several variants of our method by using different variant factors, including only using geometric factors (\"Ours_GT\"), and only using photometric factors (\"Ours_PT\"). Since the factor number of the two type factors are different, we randomly select 4 photometric factors to keep the same number as geometric factors and repeat experiments 10 times to avoid the effect of sampling. The results are shown in Table 6. From the results, it is noteworthy that \"Ours_PT\" performs better than \"Ours_GT\", especially with the $8.33\\%$ gain when using Photo as the source domain. The reason may be that the domain shift between Photo and the other target domains is mainly caused by photometric factors. Moreover, our method outperforms both \"Ours_PT\" and \"Ours_GT\", showing that both photometric factors and geometric factors are required to simulate the", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "domain shift as diverse as possible.", + "bbox": [ + 500, + 311, + 733, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Causality Visualization", + "text_level": 1, + "bbox": [ + 500, + 342, + 714, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 4, we visualize the inferred causal effects (represented by the weights) of variant factors for the unseen target images on the CIFAR10-C dataset during testing. The target images are actually corrupted from the source images of CIFAR10 by Gaussian noise of five levels. It is interesting to observe that the inferred weights of the NoiseGaussian variant factor are larger than that of the other variant factors, indicating that the counterfactual inference succeeds in discovering the real cause of the domain shift. We can also observe that when the corruption severity of Gaussian noise increases from level 1 to level 5 (i.e., the domain shift is more serious), the inferred weights of the NoiseGaussian variant factor become larger, which further demonstrates that our method measures the magnitude of the domain shift correctly.", + "bbox": [ + 496, + 367, + 893, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 614, + 619, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a new paradigm, simulate-analyze-reduce, for single domain generalization. Our paradigm empowers the model with the ability to analyze the domain shift, instead of directly expanding the distribution of the source domain to cover unseen target domains. Under this paradigm, we have presented a meta-causal learning method that can learn meta-knowledge about inferring the causes of domain shift during training, and apply such meta-knowledge to reduce the domain shift for boosting adaptation during testing. Extensive experiments on several benchmark datasets have validated the effectiveness of the new learning paradigm and the advantage of meta-causal learning on analyzing the domain shift for domain generalization.", + "bbox": [ + 496, + 641, + 893, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the Natural Science Foundation of China (NSFC) under Grant No 62072041.", + "bbox": [ + 498, + 854, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7690", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. volume 31, pages 1006-1016, 2018. 7", + "[2] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2229-2238, 2019. 2, 6, 7", + "[3] Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In European Conference on Computer Vision (ECCV), pages 301-318. Springer, 2020. 7", + "[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 113-123, 2019. 6, 7", + "[5] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 702-703, 2020. 6, 7", + "[6] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. volume 32, pages 6447–6458, 2019. 7", + "[7] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In European Conference on Computer Vision (ECCV), pages 200-216. Springer, 2020. 2", + "[8] Xinjie Fan, Qifei Wang, Junjie Ke, Feng Yang, Boqing Gong, and Mingyuan Zhou. Adversarily adaptive normalization for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8208-8217, 2021. 1, 2, 6, 7", + "[9] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), pages 1180-1189. PMLR, 2015. 5, 6", + "[10] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2551-2559, 2015. 2", + "[11] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations (ICLR), 2019. 6", + "[12] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV), pages 124-140. Springer, 2020. 6, 7", + "[13] Jonathan J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 16(5):550-554, 1994. 5" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems: École D'Été de Probabilités de Saint-Flour XXXVIII-2008, volume 2033. Springer Science & Business Media, 2011. 6, 7", + "[15] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[16] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5", + "[17] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5542-5550, 2017. 6", + "[18] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 1446-1455, 2019. 7", + "[19] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 1, 2, 6, 7", + "[20] Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. volume 34, pages 6155-6170, 2021. 2", + "[21] Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8046-8056, 2022. 2, 7", + "[22] Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain generalization using causal matching. In International Conference on Machine Learning (ICML), pages 7313-7324. PMLR, 2021. 2, 7", + "[23] Toshihiko Matsuura and Tatsuya Harada. Domain generalization using a mixture of multiple latent domains. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 11749-11756, 2020. 7", + "[24] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5715-5725, 2017, 2, 6", + "[25] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning (ICML), pages 10-18. PMLR, 2013. 2", + "[26] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 5", + "[27] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In International Conference on Machine Learning (ICML), pages 7728-7738. PMLR, 2020. 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7691", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 1, 2, 6", + "[29] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. In International Conference on Learning Representations (ICLR), 2018. 2", + "[30] Riccardo Volpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transformation sets. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7980-7989, 2019. 6, 7", + "[31] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. volume 31, pages 5339-5349, 2018. 2, 6, 7", + "[32] Haohan Wang, Zexue He, Zachary C. Lipton, and Eric P. Xing. Learning robust representations by projecting superficial statistics out. In International Conference on Learning Representations (ICLR), 2019. 2", + "[33] Shujun Wang, Lequan Yu, Caizi Li, Chi-Wing Fu, and Pheng-Ann Heng. Learning from extrinsic and intrinsic supervisions for domain generalization. In European Conference on Computer Vision (ECCV), pages 159–176. Springer, 2020. 7", + "[34] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 834-843, 2021. 1, 2, 6, 7", + "[35] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383-14392, 2021. 7", + "[36] Xiang Xu, Xiong Zhou, Ragav Venkatesan, Gurumurthy Swaminathan, and Orchid Majumder. d-sne: Domain adaptation using stochastic neighborhood embedding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2497-2506, 2019. 6, 7", + "[37] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 14435–14447, 2020. 6, 7", + "[38] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Deep domain-adversarial image generation for domain generalisation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 13025-13032, 2020. 7", + "[39] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In European Conference on Computer Vision (ECCV), pages 561-578. Springer, 2020. 2, 7" + ], + "bbox": [ + 78, + 90, + 468, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7692", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_model.json b/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cb892a54eb9728a72f7cf21c9ba764a9bac8be42 --- /dev/null +++ b/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.768, + 0.152 + ], + "angle": 0, + "content": "Meta-causal Learning for Single Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.18, + 0.684, + 0.199 + ], + "angle": 0, + "content": "Jin Chen\\(^{1*}\\), Zhi Gao\\(^{1*}\\), Xinxiao Wu\\(^{1,2\\dagger}\\), Jiebo Luo\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.2, + 0.834, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\)Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science & Technology, Beijing Institute of Technology, China \\(^{2}\\)Guangdong Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, China" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.27, + 0.83, + 0.288 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Computer Science, University of Rochester, Rochester NY 14627, USA" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.29, + 0.768, + 0.305 + ], + "angle": 0, + "content": "{chen_jin, gaozhi_2017, wuxinxiao}@bit.edu.cn, jluo@cs.rochester.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.372, + 0.474, + 0.735 + ], + "angle": 0, + "content": "Single domain generalization aims to learn a model from a single training domain (source domain) and apply it to multiple unseen test domains (target domains). Existing methods focus on expanding the distribution of the training domain to cover the target domains, but without estimating the domain shift between the source and target domains. In this paper, we propose a new learning paradigm, namely simulate-analyze-reduce, which first simulates the domain shift by building an auxiliary domain as the target domain, then learns to analyze the causes of domain shift, and finally learns to reduce the domain shift for model adaptation. Under this paradigm, we propose a meta-causal learning method to learn meta-knowledge, that is, how to infer the causes of domain shift between the auxiliary and source domains during training. We use the meta-knowledge to analyze the shift between the target and source domains during testing. Specifically, we perform multiple transformations on source data to generate the auxiliary domain, perform counterfactual inference to learn to discover the causal factors of the shift between the auxiliary and source domains, and incorporate the inferred causality into factor-aware domain alignments. Extensive experiments on several benchmarks of image classification show the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.21, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.789, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Single domain generalization [28] aims to generalize a model trained using one training domain (source domain) into multiple unseen test domains (target domains). Since only one source domain is given and the target domains are out-of-distribution and unavailable during training, single" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.339, + 0.882, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.51, + 0.895, + 0.581 + ], + "angle": 0, + "content": "Figure 1. Illustration of the simulate-analyze-reduce paradigm. In this paradigm, we first simulate the domain shift by constructing an auxiliary domain as the unseen target domain, then learn to analyze the domain shift, and finally learn to reduce the domain shift based on inferred causes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.747 + ], + "angle": 0, + "content": "domain generalization is a challenging task and attracts increasing interests. Existing works have made considerable successes through expanding the distribution of the source domain by data augmentation [19,28,34] or learning adaptive data normalization [8] typically. However, such successes have been achieved without explicitly considering the domain shift between the source and target domains, which limits the generalization performance of model in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a new learning paradigm, namely simulate-analyze-reduce, to address single domain generalization by enabling the model to analyze the real domain shift between the source domain and unseen target domain. This new paradigm is shown in Figure 1. We first build an auxiliary domain as the target domain to simulate the real domain shift between the source and target domains, since the target data is unavailable during training. We then learn to analyze the intrinsic causal factors of the domain shift to facilitate the subsequent model adapta" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.335, + 0.887 + ], + "angle": 0, + "content": "* Jin Chen and Zhi Gao are co-first authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.304, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author: Xinxiao Wu." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.335, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "tion. Finally, we learn to reduce the domain shift with its inferred causes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.122, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Under this paradigm, we propose a meta-causal learning method to learn the meta-knowledge about how to infer the causes of the simulated domain shift between the auxiliary and source domains via causal inference in training, and then apply the meta-knowledge to analyze the real domain shift between the target and source domains during testing, through which the source and given target domains are adaptively aligned. Specifically, we perform multiple transformations on source data to generate an auxiliary domain with great diversity. Then we build a causal graph to represent the dependency among data, variant factors, semantic concepts and category labels, and conduct counterfactual inference over the causal graph to exploit the intrinsic causality of the simulated domain shift between the auxiliary and source domains. For each sample in the auxiliary domain, we construct counterfactual scenes by intervening variant factors to infer their causal effects on the category prediction, and these inferred causal effects of variant factors can be regarded as the causes of domain shift. To reduce the domain shift, we propose a factor-aware domain alignment by learning and integrating multiple feature mappings, where an effect-to-weight network is designed to convert the causal effects of variant factors into the weights of feature mappings." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.485, + 0.473, + 0.576 + ], + "angle": 0, + "content": "During testing, the distribution discrepancy between the input target sample and the source domain is analyzed and reduced by applying the learnt meta-knowledge, i.e., inferring the causal effects of variant factors and incorporating them into the factor-aware domain alignment. In summary, the main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.582, + 0.47, + 0.673 + ], + "angle": 0, + "content": "- We propose a novel learning paradigm, simulate-analyze-reduce, for single domain generalization. This paradigm empowers the model with the ability to estimate the domain shift between the source domain and unseen target domains, thus boosting the model adaptation across different domains." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.682, + 0.47, + 0.742 + ], + "angle": 0, + "content": "- We propose a meta-causal learning method based on counterfactual inference to learn the meta-knowledge about analyzing the intrinsic causality of domain shift, thus facilitating the reduction of domain shift." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.75, + 0.468, + 0.811 + ], + "angle": 0, + "content": "- Our method achieves the state-of-the-art results on several benchmarks of image classification, especially on the more challenging tasks with a large domain shift, clearly demonstrating the effectiveness of our method." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.582, + 0.47, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.22, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.293, + 0.862 + ], + "angle": 0, + "content": "2.1. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Domain generalization focuses on generalizing a model learned from multiple source domains to the unseen target" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "domain. The key difference between domain adaptation and domain generalization is that during training, domain adaptation leverages unlabelled target data while domain generalization has no access to the target domain. Existing domain generalization methods can be roughly divided into two categories: learning domain-invariant feature representation from multiple source domains [7, 10, 24, 25, 32] and generating diverse more samples via data augmentation [2, 29, 31, 39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.893, + 0.499 + ], + "angle": 0, + "content": "Recently, single domain generalization [28] has attracted growing attention, where only one source domain is available during training and the model is evaluated on multiple unseen target domains. A rich line of works employ data augmentation for generating out-of-domain samples to expand the distribution of the source domain [19, 28, 34]. Qiao et al. [28] propose meta-learning based adversarial domain augmentation to generate samples. Li et al. [19] propose a progressive domain expansion network to generate multiple domains progressively via simulating various photometric and geometric transforms by style transfer based generators. Wang et al. [34] propose a style-complement module to generate diverse images with different styles. Fan et al. [8] use data normalization for single domain generalization, where an adaptive normalization scheme is learned to be incorporated with adversarial domain augmentation to enhance the generalization of the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.623 + ], + "angle": 0, + "content": "The aforementioned methods aim to generate the source data distribution as diverse as possible to cover unseen target domains. When the expanded source distribution does not approximate the target distribution, the performance may significantly degrade, since there still exists a domain gap between the source and target domains. To address this problem, our method learns to analyze and reduce the domain shift by building an auxiliary domain during training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.82, + 0.65 + ], + "angle": 0, + "content": "2.2. Causality for Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Several recent methods exploit causality to learn domain-invariant semantic representation for domain generalization [20-22]. Considering the cross-domain invariance of the causality between semantic factors and predictions, Liu et al. [20] propose a causal semantic generative model to remove the domain-specific correlation between semantic factors and variant factors, and thus make the prediction affected only by the semantic factors. Assuming that images of the same object across domains should have the same representation, Mahajan et al. [22] use the cross-domain invariance of the causality between objects and feature representations to capture the within-class variation for domain generalization. Lv et al. [21] introduce causal inference to extract causal factors that are invariant across domains in order to learn invariant feature representation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "From a different perspective, our method focuses on uti-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7684" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.862, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.374, + 0.895, + 0.445 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed meta-causal learning method. (1) A data transformation module simulates the domain shift via generating an auxiliary domain \\(\\mathcal{D}_a\\). (2) A counterfactual inference module analyzes the domain shift by inferring the causal effects of \\(V\\) on \\(Y\\). “-” denotes comparing the values of \\(Y\\) before and after do-operation. The edge \\(X \\rightarrow Y\\) is a dashed arrow to represent the classification model \\(P(Y|X)\\). (3) A factor-aware domain alignment module reduces the domain shift via multiple feature mappings according to weights learned by an effect-to-weight network. (4) A base module is used for feature extraction and classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.471, + 0.532 + ], + "angle": 0, + "content": "lizing causal inference to discover the intrinsic causes of domain shift by constructing counterfactual scenes over the learnt causal graph, which facilitates the reduction of domain shift and in turn benefits cross-domain adaptation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.169, + 0.561 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.26, + 0.587 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.596, + 0.47, + 0.704 + ], + "angle": 0, + "content": "For single domain generalization, during training, we are given a labeled source domain \\(\\mathcal{D}_s = \\{(x_i^s,y_i^s)|_{i = 1}^{N_s}\\}\\) drawn from the distribution \\(P_{s}\\), where \\(\\pmb{x}_i^s\\) is the \\(i\\)-th source sample with its category label \\(y_{i}^{s}\\in \\mathcal{V}\\). During testing, the learnt model is applied to multiple unseen target domains \\(\\mathcal{D}_t = \\{\\mathcal{D}_t^j\\}_{j = 1}^J\\), where \\(\\mathcal{D}_t^j\\) is the \\(j\\)-th target domain drawn from the distribution \\(P_{t}^{j}\\), and \\(P_{t}^{j}\\neq P_{s}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To bridge the domain gap between the source domain and multiple unseen target domains, we propose a new learning paradigm, called simulate-analyze-reduce, which starts with simulating the domain shift between the source and target domains, then learns to analyze the domain shift, and finally learns to reduce the domain shift. Under this paradigm, we propose a meta-causal learning method that has four components: a data transformation module to generate auxiliary domains as target domains, a counterfactual inference module to discover the causes of the domain shift, a factor-aware domain alignment module to reduce the domain shift, and a base module for feature extraction and classification, as illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.892, + 0.501 + ], + "angle": 0, + "content": "3.2. Data Transformation for Domain Shift Simulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.704 + ], + "angle": 0, + "content": "To simulate the real domain shift between the source and target domains, we generate an auxiliary domain \\(\\mathcal{D}_a\\) as the unseen target domain by performing transformations on source data. The domain shift is actually the data distribution discrepancy between the source and auxiliary domains, and is usually caused by the variations of extrinsic attributes of data, independent of intrinsic semantics. Taking the image data for example, the domain shift is mainly caused by the variations of visual attributes, such as the variations of brightness, viewpoint, and color. Therefore, to make the simulated domain shift as realistic as possible, we formulate the extrinsic attributes as variant factors, and design data transformations according to the variant factors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.841 + ], + "angle": 0, + "content": "We define a set of variant factors, denoted as \\(\\mathcal{V} = \\{v_{1}, v_{2}, \\dots, v_{K}\\}\\), where \\(v_{k}\\) denotes the \\(k\\)-th variant factor. Each variant factor corresponds to a data transformation function that aims to generate new data by making changes on the corresponding extrinsic attribute, denoted as \\(G_{v_{k}}(\\boldsymbol{x}; \\theta_{v_{k}})\\), where \\(\\boldsymbol{x}\\) is an input sample, and \\(\\theta_{v_{k}} \\in [g_{min}^{k}, g_{max}^{k}]\\) represents the degree parameter to control the magnitude of transformation with the scale range \\([g_{min}^{k}, g_{max}^{k}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In real-world scenarios, the complex domain shift is often caused by combinatorial multiple variant factors and accordingly we design a sampling strategy to enable the combinatorial data transformation. Given a source sample \\( \\boldsymbol{x}_i^s \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7685" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.14 + ], + "angle": 0, + "content": "we randomly sample \\(N_{v}^{i}\\) variant factors from \\(\\mathcal{V}\\) to form a factor subset \\(\\mathcal{V}_i = \\{v_1^i,v_2^i,\\dots ,v_{N_v^i}^i\\}\\). Then, a corresponding auxiliary sample \\(\\pmb{x}_i^a\\) is generated by" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.149, + 0.469, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {i} ^ {a} = G _ {v _ {N _ {v} ^ {i}} ^ {i}} \\left(\\dots G _ {v _ {2} ^ {i}} \\left(G _ {v _ {1} ^ {i}} \\left(\\boldsymbol {x} _ {i} ^ {s}; \\theta_ {v _ {1} ^ {i}}\\right); \\theta_ {v _ {2} ^ {i}}\\right) \\dots ; \\theta_ {v _ {N _ {v} ^ {i}} ^ {i}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.189, + 0.469, + 0.249 + ], + "angle": 0, + "content": "where the transformation degree parameter \\(\\theta_{v_k^i}\\) is randomly selected from its range scale. In this way, the domain shift is generated as diverse as possible to approximate the real domain shift." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.258, + 0.469, + 0.29 + ], + "angle": 0, + "content": "3.3. Counterfactual Inference for Domain Shift Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.297, + 0.469, + 0.387 + ], + "angle": 0, + "content": "After simulating the domain shift, we introduce counterfactual inference to analyze the domain shift. During training, we learn the meta-knowledge about how to infer the causes of data discrepancy between one auxiliary sample and the source domain, and during testing, we apply the learnt meta-knowledge to unseen target samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.388, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We build a causal graph to model the causal dependency among the input sample (node \\( X \\)), the variant factors (node \\( V \\)), the semantic concepts (node \\( S \\)), and the output category (node \\( Y \\)), as shown in Figure 2 (2). The semantic concepts denote the intrinsic attributes of data that are related to the category, and the variant factors denote the extrinsic attributes of data that are domain specific, independent of the intrinsic semantics. For example, when the input sample is an image of \"zebra\", the semantic concepts are like \"four legs\" and \"black-white stripes\", and the variant factors include brightness, viewpoint, and so on. The edge \\( S \\rightarrow Y \\) represents that the category of the input data is determined by the semantic concepts. The edges \\( S \\rightarrow X \\) and \\( V \\rightarrow X \\) represent that the semantic concepts and the variant factors together determine what the input sample looks like. Since the semantic concepts represent the intrinsic semantics and are invariant across different domains, ideally the cross-domain classification is only determined by the semantic concepts, which is implemented by learning the edge \\( S \\rightarrow Y \\), i.e., estimating the conditional probability \\( P(Y|S) \\). However, in reality, the semantic concepts are unobserved from the input samples, so the classification should be implemented by learning the edge \\( X \\rightarrow Y \\), i.e., estimating the conditional probability \\( P(Y|X) \\). The edge \\( X \\rightarrow Y \\) is a dashed arrow to represent the classification model \\( P(Y|X) \\). Since the input sample node \\( X \\) is affected by the semantic concept node \\( S \\) and the variant factor node \\( V \\) together, the category node \\( Y \\) is also affected by the variant factor node \\( V \\) through edges \\( V \\rightarrow X \\rightarrow Y \\). As the variant factors are domain-specific, their causal effects on the category lead to the domain shift. Hence, we infer the causal effects of the variant factors on the category prediction to discover the causes of the domain shift. That is to say, we infer the causal effects of the node \\( V \\) on the node" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "\\(Y\\) as causes of domain shift, and then learn to reduce the domain shift based on its causes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.121, + 0.892, + 0.166 + ], + "angle": 0, + "content": "To infer the causal effects of the node \\(V\\) on the node \\(Y\\), we first learn the edge \\(X \\to Y\\) using the source data \\(\\mathcal{D}_s\\) by a classification loss:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.174, + 0.892, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\sim P _ {s}} \\left[ - \\sum_ {u} \\mathbb {I} _ {u = y _ {i} ^ {s}} \\log C \\left(F \\left(\\boldsymbol {x} _ {i} ^ {s}\\right)\\right) _ {u} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.225, + 0.892, + 0.376 + ], + "angle": 0, + "content": "where \\((\\pmb{x}_i^s, y_i^s) \\in \\mathcal{D}_s\\), \\(F\\) is a feature extractor, \\(C\\) is the classifier to output \\(|\\mathcal{V}|\\) category probabilities, \\(C(\\cdot)_u\\) is the \\(u\\)-th element of \\(|\\mathcal{V}|\\) category probabilities, and \\(\\mathbb{I}_{u=y_i^s}\\) is an indicator function, meaning that if \\(u = y_i^s\\), the value of \\(\\mathbb{I}_{u=y_i^s}\\) is 1 and 0 otherwise. Then, for each sample from the auxiliary domain, we construct a factual scene and multiple counterfactual scenes over the causal graph, so as to infer the causal effects of variant factors on the category prediction. Given an auxiliary sample \\(\\pmb{x}_i^a \\in \\mathcal{D}_a\\), its factual category is predicted by" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.388, + 0.892, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {i} ^ {a} = P (Y | X) = C \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.892, + 0.551 + ], + "angle": 0, + "content": "where \\( \\pmb{y}_i^a \\in \\mathbb{R}^{|\\mathcal{Y}|} \\) is a category probability vector, and represents the value of node \\( Y \\). For each variant factor, we construct a counterfactual scene to infer its causal effect by doing intervention on the variant factor node \\( V \\). Let \\( do(V = v_k) \\) denote the intervention on the node \\( V \\), which is implemented by changing extrinsic attributes of data through the transformation \\( G_{v_k} \\) with multiple degree parameters. Accordingly, the counterfactual category of \\( \\pmb{x}_i^a \\) is predicted by" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.561, + 0.891, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {y} _ {i, v _ {k}} ^ {a} = P \\left(Y | X, d o (V = v _ {k})\\right) \\\\ = \\frac {1}{| \\mathcal {M} |} \\sum_ {\\theta_ {v _ {k}} \\in \\mathcal {M}} C \\left(F \\left(G _ {v _ {k}} \\left(\\boldsymbol {x} _ {i} ^ {a}; \\theta_ {v _ {k}}\\right)\\right)\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.63, + 0.892, + 0.706 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is a set of degree parameters, obtained by uniformly sampling magnitudes of transformation \\(G_{v_k}\\) from the scale range. By comparing the factual and counterfactual category probabilities of \\(\\pmb{x}_i^a\\), the causal effect of the \\(k\\)-th variant factor is calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.716, + 0.892, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {e} _ {k} ^ {i} = P (Y | X) - P \\left(Y | X, d o (V = v _ {k})\\right) = \\boldsymbol {y} _ {i} ^ {a} - \\boldsymbol {y} _ {i, v _ {k}} ^ {a}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.748, + 0.892, + 0.808 + ], + "angle": 0, + "content": "The inferred causal effect represents how much contribution the corresponding variant factor makes to the domain shift, and the larger the causal effect is, the more seriously the domain shift is caused by the variant factor." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.892, + 0.847 + ], + "angle": 0, + "content": "3.4. Factor-aware Domain Alignment for Domain Shift Reduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "After analyzing the causes of domain shift, we propose a factor-aware domain alignment to reduce the domain shift by learning multiple feature mappings, with guidance of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.088, + 0.873, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.256, + 0.715, + 0.272 + ], + "angle": 0, + "content": "Figure 3. Inference process of meta-causal learning for a given target sample." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.47, + 0.433 + ], + "angle": 0, + "content": "inferred causal effects of variant factors. Each feature mapping addresses a specific domain shift caused by one variant factor. We construct \\( K \\) feature mappings for \\( K \\) variant factors, and the \\( k \\)-th feature mapping aims to address the domain shift caused by the \\( k \\)-th variant factor. In order to incorporate the causal effects of variant factors into the learning of mappings, we build an effect-to-weight network that converts the causal effect of each variant factor into the weight of the corresponding feature mapping." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.433, + 0.471, + 0.479 + ], + "angle": 0, + "content": "For the auxiliary sample \\( \\pmb{x}_i^a \\) and its inferred causal effects of all variant factors \\( \\{e_1^i, e_2^i, \\dots, e_K^i\\} \\), the weights of feature mappings are calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.486, + 0.47, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\omega} ^ {i} = \\operatorname {s o f t m a x} \\big (W (\\boldsymbol {e} _ {1} ^ {i}), W (\\boldsymbol {e} _ {2} ^ {i}), \\dots , W (\\boldsymbol {e} _ {K} ^ {i}) \\big), \\quad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.557 + ], + "angle": 0, + "content": "where \\(\\omega^i\\in \\mathbb{R}^K\\) and its \\(k\\) -th element \\(\\omega_{k}^{i}\\) denotes the weight of the \\(k\\) -th factor-aware feature mapping. \\(W(\\cdot)\\) denotes the effect-to-weight network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.471, + 0.707 + ], + "angle": 0, + "content": "According to the weights of feature mappings, we integrate the \\(K\\) feature mappings to project the auxiliary samples into the source feature space. The alignment of the source and auxiliary domains is implemented by minimizing the feature distance between the source and auxiliary samples in the source feature space. Given a source sample \\(\\pmb{x}_i^s\\) with the corresponding category label \\(y_i^s\\), generated auxiliary sample \\(\\pmb{x}_i^a\\) and inferred mapping weights \\(\\omega^i\\), the alignment loss incorporated with inferred causal effects is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.713, + 0.47, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {a} ^ {c} = \\frac {1}{N _ {s}} \\sum_ {i} | | F \\left(\\boldsymbol {x} _ {i} ^ {s}\\right) - \\sum_ {k} \\omega_ {k} ^ {i} M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right) | | _ {2} \\tag {7} \\\\ + \\frac {1}{N _ {s}} \\sum_ {i} \\mathcal {H} \\left(C \\left(\\sum_ {k} \\omega_ {k} ^ {i} M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right), y _ {i} ^ {s}\\right)\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(F\\) is the feature extractor, \\(M_{k}\\) is the \\(k\\)-th feature mapping, \\(C\\) is the classifier, \\(\\omega_{k}^{i}\\) is the \\(k\\)-th element of \\(\\omega^i\\), \\(\\sum_{k}\\omega_{k}^{i}M_{k}\\big(F(\\pmb{x}_{i}^{a})\\big)\\) represents the projected feature of \\(\\pmb{x}_i^a\\), and \\(N_{s}\\) is the number of source samples. \\(||\\cdot ||_2\\) is the L2-loss, and \\(\\mathcal{H}(\\cdot)\\) is the cross-entropy loss. The first term measures the feature distance between the source and auxiliary samples after projection, i.e., the data distribution discrepancy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.892, + 0.342 + ], + "angle": 0, + "content": "between the source and auxiliary domains. The second term encourages the auxiliary samples to belong to the same categories as the source samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.389 + ], + "angle": 0, + "content": "Moreover, to enable each feature mapping to address the specific domain shift caused by the corresponding variant factor, we introduce another alignment loss \\(\\mathcal{L}_a^m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.398, + 0.891, + 0.474 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {a} ^ {m} = \\frac {1}{N _ {s}} \\frac {1}{K} \\sum_ {i} \\sum_ {k} | | F (\\boldsymbol {x} _ {i} ^ {s}) - M _ {k} (F (\\boldsymbol {x} _ {i} ^ {k})) | | _ {2} \\\\ + \\frac {1}{N _ {s}} \\frac {1}{K} \\sum_ {i} \\sum_ {k} \\mathcal {H} \\left(C \\left(M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {k}\\right)\\right), y _ {i} ^ {s}\\right)\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.548 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_i^k\\) is an sample generated by conducting the data transformation \\(G_{v_k}\\) on the source sample \\(\\pmb{x}_i^s\\), denoted by \\(\\pmb{x}_i^k = G_{v_k}(\\pmb{x}_i^s;\\theta_{v_k})\\). The transformation degree parameter \\(\\theta_{v_k}\\) is randomly selected from its range scale." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.548, + 0.804, + 0.561 + ], + "angle": 0, + "content": "Then the overall loss function is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.574, + 0.892, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {a} ^ {c} + \\mathcal {L} _ {a} ^ {m}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.602, + 0.885, + 0.618 + ], + "angle": 0, + "content": "The whole training process is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.61, + 0.641 + ], + "angle": 0, + "content": "3.5. Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.809 + ], + "angle": 0, + "content": "During testing, given a target sample \\( \\boldsymbol{x}_i^t \\), firstly, the causal effects of variant factors are inferred by counterfactual inference in Eq. (5), and the weights \\( \\omega^i \\) of feature mappings are calculated by the effect-to-weight network in Eq. (6). Then the target feature \\( \\boldsymbol{f}_i^t = F(\\boldsymbol{x}_i^t) \\) is projected into the source feature space by integrating \\( K \\) feature mappings according to their weights \\( \\omega^i \\) to obtain the projected feature \\( \\hat{\\boldsymbol{f}}_i^t = \\sum_k \\omega_k^i M_k(\\boldsymbol{f}_i^t) \\). Finally, the category of \\( \\boldsymbol{x}_i^t \\) is predicted by \\( C(\\hat{\\boldsymbol{f}}_i^t) \\). The whole inference process is shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.823, + 0.632, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.602, + 0.862 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Digits. The Digits dataset consists of five datasets: MNIST [16], MNIST-M [9], SVHN [26], USPS [13], and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7687" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.092, + 0.28, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Training process" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.111, + 0.444, + 0.125 + ], + "angle": 0, + "content": "Input: The source domain \\(\\mathcal{D}_s\\), the variant factor set \\(\\nu\\)." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.127, + 0.469, + 0.169 + ], + "angle": 0, + "content": "Output: The feature extractor \\(F\\), the classifier \\(C\\), \\(K\\) feature mappings \\(\\{M_i\\}_{i=1}^K\\), the effect-to-weight network \\(W\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.171, + 0.304, + 0.187 + ], + "angle": 0, + "content": "1: Initialize \\(F, C, \\{M_i\\}_{i=1}^K, W\\);" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.187, + 0.256, + 0.201 + ], + "angle": 0, + "content": "2: while not converge do" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.202, + 0.324, + 0.216 + ], + "angle": 0, + "content": "3: Sample an image \\( \\pmb{x}_j^s \\) from \\( \\mathcal{D}_s \\);" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.217, + 0.469, + 0.245 + ], + "angle": 0, + "content": "4: Construct a factor subset \\(\\mathcal{V}_i\\) by sampling \\(N_v^i\\) variant factors from \\(\\mathcal{V}\\);" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.247, + 0.41, + 0.262 + ], + "angle": 0, + "content": "5: Generate an auxiliary sample \\( \\pmb{x}_i^a \\) by Eq.(1);" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.263, + 0.221, + 0.276 + ], + "angle": 0, + "content": "6: for \\(v_{k}\\) in \\(\\mathcal{V}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.277, + 0.469, + 0.307 + ], + "angle": 0, + "content": "7: Calculate the factual category \\( \\mathbf{y}_i^a \\) of auxiliary sample \\( \\mathbf{x}_i^a \\) by Eq.(3);" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.308, + 0.468, + 0.337 + ], + "angle": 0, + "content": "8: Calculate the counterfactual category \\( \\pmb{y}_{i,v_k}^a \\) of auxiliary sample \\( \\pmb{x}_i^a \\) by Eq.(4);" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.468, + 0.369 + ], + "angle": 0, + "content": "9: Infer the causal effect \\( e_k^i \\) of variant factor \\( v_k \\) via comparing \\( y_i^a \\) and \\( y_{i,v_k}^a \\) by Eq.(5);" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.171, + 0.469, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.37, + 0.176, + 0.381 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.383, + 0.469, + 0.413 + ], + "angle": 0, + "content": "11: Calculate the weights of \\( K \\) feature mappings \\( \\{M_i\\}_{i=1}^K \\) by Eq.(6);" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.413, + 0.373, + 0.429 + ], + "angle": 0, + "content": "12: Update \\(F, C, \\{M_i\\}_{i=1}^K\\), \\(W\\) by Eq.(9)." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.429, + 0.176, + 0.441 + ], + "angle": 0, + "content": "13: end while" + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.37, + 0.469, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.469, + 0.533 + ], + "angle": 0, + "content": "SYN [9], with 10 categories. We use MNIST as the source domain, and the other four datasets as the target domains. The first 10,000 images in the training set of MNIST are used for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.469, + 0.653 + ], + "angle": 0, + "content": "CIFAR10-C. The CIFAR10-C dataset [11] is proposed to evaluate the robustness of classification model. The images are corrupted from the test set of the CIFAR10 dataset [15] by 19 corruption types with five levels of severity. A higher level means the more serious corruption. There are 10 categories. We use CIFAR10 as the source domain, and CIFAR10-C as the target domains where images of one severity level form one target domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.469, + 0.745 + ], + "angle": 0, + "content": "PACS. The PACS dataset [17] is a benchmark for domain generalization, and consists of four domains: art painting, cartoon, photo, and sketch. There are 9,991 images of seven categories. We use one domain as the source domain, and the rest three domains as the target domains. So there are four tasks with different domains as the source domains." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.294, + 0.772 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Auxiliary Domain. We define 16 variant factors to generate the images in the auxiliary domain, including 12 photometric factors (Brightness, Contrast, Color, Sharpness, Auto-Contrast, Invert, Equalize, Solarize, SolarizeAdd, Posterize, NoiseSalt, NoiseGaussian) and 4 geometric factors (Shear-X, Shear-Y, Rotate, Flip). Since the Rotate and Flip variant factors will affect the semantic information of digit images, we use the other 14 variant factors for the Digits dataset. For" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 1. Single domain generalization results (%) on Digits with ConvNet as backbone. The model is trained on MNIST, and evaluated on SVHN, SYN, MNIST-M, and USPS." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.142, + 0.899, + 0.347 + ], + "angle": 0, + "content": "
MethodSVHNSYNMNIST-MUSPSAvg
ERM [14]27.8339.6552.7276.9449.29
CCSA [24]25.8937.3149.2983.7249.05
d-SNE [36]26.2237.8350.9893.1652.05
JiGen [2]33.8043.7957.8077.1553.14
GUD [31]35.5145.3260.4177.2654.62
M-ADA [28]42.5548.9567.9478.5359.49
ME-ADA [37]42.5650.3963.2781.0459.32
PDEN [19]62.2169.3982.2085.2674.77
L2D [34]62.8663.7287.3083.9774.46
AA [4]45.2364.5260.5380.6262.72
RA [5]54.7759.6074.0577.3366.44
RSDA [30]47.4062.0081.5083.1068.50
RSDA+ASR [8]52.8064.5080.8082.4070.10
Ours69.9478.4778.3488.5478.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Table 2. Single domain generalization results (\\%) on CIFAR10-C with WRN as backbone. Each level is viewed as a target domain, a higher level denotes the more serious corruption and the domain discrepancy between the source and target domains is larger." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.428, + 0.899, + 0.538 + ], + "angle": 0, + "content": "
Methodlevel1level2level3level4level5Avg
ERM [14]87.8081.5075.5068.2056.1073.82
GUD [31]88.3083.5077.6070.6058.3075.66
M-ADA [28]90.5086.8082.5076.4065.6080.36
PDEN [19]90.6288.9187.0383.7177.4785.55
AA [4]91.4287.8884.1078.4671.1382.60
RA [5]91.7488.8985.8281.0374.9384.48
Ours92.3891.2289.8887.7384.5289.15
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.64 + ], + "angle": 0, + "content": "the CIFAR10-C and PACS datasets, all 16 variant factors are used. To make the auxiliary domain as diverse as possible, for each source image at each iteration, we randomly sample several variant factors and use them to generate a new auxiliary image." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.65, + 0.851, + 0.667 + ], + "angle": 0, + "content": "4.3. Results on Single Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We compare our method with several state-of-the-art methods, including the baseline method (ERM [14]), the methods of learning domain-invariant features (CCSA [24], d-SNE [36], JiGen [2]), and the methods of making data augmentation (GUD [31], M-ADA [28], ME-ADA [37], PDEN [19], L2D [34], AA [4], RA [5], RSDA [30], RSC [12], ASR [8])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Table 1, Table 2, and Table 3 show the comparison results on Digits, CIFAR10-C, and PACS, respectively. From the results, there are several interesting observations as follows. First, our method generally outperforms the compared methods on all datasets, which clearly shows the effectiveness of the proposed simulate-analyze-reduce learning paradigm for single domain generalization. Second, compared with the data augmentation methods (e.g.," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7688" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 3. Single domain generalization results (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as the target domains." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.156, + 0.472, + 0.229 + ], + "angle": 0, + "content": "
MethodArtpaintCartoonSketchPhotoAvg
ERM [14]70.9076.5053.1042.2060.70
RSC [12]73.4075.9056.2041.6061.80
RSC+ASR [8]76.7079.3061.6054.6068.10
Ours77.1380.1462.5559.6069.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.244, + 0.47, + 0.299 + ], + "angle": 0, + "content": "Table 4. Leave-one-domain-out results (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the target domain and the other three domains are used as source domains." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.311, + 0.477, + 0.571 + ], + "angle": 0, + "content": "
MethodArtpaintCartoonPhotoSketchAvg
MetaReg [1]83.7077.2095.5070.3081.70
GUD [31]78.3277.6595.6174.2181.44
Epi-FCR [18]82.1077.0093.9073.0081.50
MASF [6]80.2977.1794.9971.6881.03
JiGen [2]79.4275.2596.0371.3580.51
DMG [3]76.9080.3893.5575.2181.46
DDAIG [38]84.2078.1095.3074.7083.10
CSD [27]78.9075.8094.1076.7081.40
L2A-OT [39]83.3078.2096.2073.6082.80
EISNet [33]81.8976.4495.9374.3382.15
RSC [12]83.4380.3195.9980.8585.15
ME-ADA [37]78.6178.6595.5775.5982.10
MMLD [23]81.2877.1696.0972.2981.83
L2D [34]81.4479.5695.5180.5884.27
FACT [35]85.3778.3895.1579.1584.51
MatchDG [22]81.3280.7096.5379.7284.57
CIRL [21]86.0880.5995.9382.6786.32
Ours85.3080.9396.5385.2487.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.587, + 0.471, + 0.658 + ], + "angle": 0, + "content": "Table 5. Ablation study (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as target domains. “T”, “A”, “C” denote Domain Transformation, Domain Alignment, and Counterfactual Inference, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.668, + 0.477, + 0.737 + ], + "angle": 0, + "content": "
MethodTACArtpaintCartoonSketchPhotoAvg
Base71.2667.6443.9736.9954.97
DT✓75.2878.4659.4556.0967.32
DTA✓✓71.6472.7857.1152.0263.39
Ours✓✓✓77.1380.1462.5559.6069.86
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "AA [4], RA [5] and RSDA [30]) that are more related to our method, our method achieves much better results, and especially yields a \\(10.32\\%\\) gain over RSDA on Digits, strongly suggesting that it is beneficial to empower the model with the ability of analyzing the causes of domain shift by counterfactual inference. Third, on more difficult tasks with larger domain shift (e.g., SVHN on Digits, level5 on CIFAR10-C, and Photo on PACS), our method significantly improves the performance, further demonstrating the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "superiority of our method on handling more challenging situations. Forth, our method performs little worse on MNIST-M and USPS of Digits. The possible reason is that other compared methods use more well-designed data augmentation and network regularization, such as AdaIN based generators in PDEN [19] and stochastic neighborhood embedding techniques in d-SNE [36]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.208, + 0.871, + 0.224 + ], + "angle": 0, + "content": "4.4. Results on Multiple Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.413 + ], + "angle": 0, + "content": "We extend the proposed method to multi-source domain setting by regarding the multiple source domains as one source domain without using domain labels. We employ the leave-one-domain-out protocol following existing multi-source domain generalization [21, 35]. We compare our method with most related methods that introduces causal inference into domain generalization (MatchDG [22], CIRL [21]), and existing popular domain generalization methods (MetaReg [1], GUD [31], EpiFCR [18], MASF [6], JiGen [2], DMG [3], DDAIG [38], CSD [27], L2A-OT [39], EISNet [33], RSC [12], ME-ADA [37], MMLD [23], L2D [34], FACT [35])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Table 4 shows the leave-one-domain-out results on the PACS dataset with ResNet-18 as backbone. From the results, we make several observations. First, it is noteworthy that our method achieves the state-of-the-art overall performance (\"Avg\") although our method is not designed for multi-source domain generalization, which further demonstrates that the proposed simulate-analyze-reduce learning paradigm not only benefits single domain generalization but also boosts multi-source domain generalization. Second, compared with the methods of introducing causal inference to learn domain-invariant features (MatchDG [22], CIRL [21]), our method achieves better results on the overall metric \"Avg\", especially making \\(5.52\\%\\) and \\(2.57\\%\\) gains over MatchDG [22] and CIRL [21], respectively, on the more challenging task (Sketch \\(\\rightarrow\\) Others). Such improvements are attributed to the ability of analyzing and reducing the domain shift, further demonstrating the advantages of causal inference in analyzing the causes of the domain shift." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.665, + 0.726 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate each component of our method, we conduct ablation experiments on the PACS dataset. We design several degraded variants of our method for comparison: (1) \"Base\", where only the base module is utilized and optimized by Eq. (2) using the source domain; (2) \"DT\", where the data transformation module is added into the base module and the model is trained using both source and auxiliary domains; (3) \"DTA\", where the simulated domain shift between auxiliary domain and source domain is directly reduced without analyzing the causes of the domain shift via the proposed counterfactual module." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7689" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.85, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Figure 4. Examples of the inferred causal effects (represented as weights) of variant factors. The left part shows a source image from the CIFAR10 dataset and three target images from the CIFAR10-C dataset with Gaussian noise corruption. As the corruption severity increases from level 1 to level 5, the inferred weights of the NoiseGaussian variant factor become larger accordingly." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.31, + 0.471, + 0.354 + ], + "angle": 0, + "content": "Table 6. Factor analysis (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as target domains." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.363, + 0.477, + 0.428 + ], + "angle": 0, + "content": "
MethodArtpaintCartoonSketchPhotoAvg
Ours_GT74.6269.7852.7443.3460.12
Ours_PT74.4773.4850.3551.6762.49
Ours77.1380.1462.5559.6069.86
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.471, + 0.623 + ], + "angle": 0, + "content": "The results are shown in Table 5. From the results, we make several observations. First, our method achieves better performance than \"DT\", which validates the superiority of our method on analyzing and reducing the domain shift, rather than directly enlarging the source data distribution. Second, \"DTA\" performs worse than \"DT\", probably because brute-force domain alignment without analyzing causes of the domain shift leads to negative transfer and thus hurts the performance. Third, our method achieves the best performance thanks to the proposed simulate-analyze-reduce paradigm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.236, + 0.65 + ], + "angle": 0, + "content": "4.6. Factor Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In order to further analyze the effect of variant factors, we design several variants of our method by using different variant factors, including only using geometric factors (\"Ours_GT\"), and only using photometric factors (\"Ours_PT\"). Since the factor number of the two type factors are different, we randomly select 4 photometric factors to keep the same number as geometric factors and repeat experiments 10 times to avoid the effect of sampling. The results are shown in Table 6. From the results, it is noteworthy that \"Ours_PT\" performs better than \"Ours_GT\", especially with the \\(8.33\\%\\) gain when using Photo as the source domain. The reason may be that the domain shift between Photo and the other target domains is mainly caused by photometric factors. Moreover, our method outperforms both \"Ours_PT\" and \"Ours_GT\", showing that both photometric factors and geometric factors are required to simulate the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.313, + 0.734, + 0.328 + ], + "angle": 0, + "content": "domain shift as diverse as possible." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.343, + 0.715, + 0.359 + ], + "angle": 0, + "content": "4.7. Causality Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.895, + 0.596 + ], + "angle": 0, + "content": "In Figure 4, we visualize the inferred causal effects (represented by the weights) of variant factors for the unseen target images on the CIFAR10-C dataset during testing. The target images are actually corrupted from the source images of CIFAR10 by Gaussian noise of five levels. It is interesting to observe that the inferred weights of the NoiseGaussian variant factor are larger than that of the other variant factors, indicating that the counterfactual inference succeeds in discovering the real cause of the domain shift. We can also observe that when the corruption severity of Gaussian noise increases from level 1 to level 5 (i.e., the domain shift is more serious), the inferred weights of the NoiseGaussian variant factor become larger, which further demonstrates that our method measures the magnitude of the domain shift correctly." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.615, + 0.62, + 0.631 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.894, + 0.852 + ], + "angle": 0, + "content": "We have presented a new paradigm, simulate-analyze-reduce, for single domain generalization. Our paradigm empowers the model with the ability to analyze the domain shift, instead of directly expanding the distribution of the source domain to cover unseen target domains. Under this paradigm, we have presented a meta-causal learning method that can learn meta-knowledge about inferring the causes of domain shift during training, and apply such meta-knowledge to reduce the domain shift for boosting adaptation during testing. Extensive experiments on several benchmark datasets have validated the effectiveness of the new learning paradigm and the advantage of meta-causal learning on analyzing the domain shift for domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the Natural Science Foundation of China (NSFC) under Grant No 62072041." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7690" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.157 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. volume 31, pages 1006-1016, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.469, + 0.226 + ], + "angle": 0, + "content": "[2] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2229-2238, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.469, + 0.284 + ], + "angle": 0, + "content": "[3] Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In European Conference on Computer Vision (ECCV), pages 301-318. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 113-123, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[5] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 702-703, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.469, + 0.478 + ], + "angle": 0, + "content": "[6] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. volume 32, pages 6447–6458, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[7] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In European Conference on Computer Vision (ECCV), pages 200-216. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[8] Xinjie Fan, Qifei Wang, Junjie Ke, Feng Yang, Boqing Gong, and Mingyuan Zhou. Adversarily adaptive normalization for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8208-8217, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[9] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), pages 1180-1189. PMLR, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[10] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2551-2559, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[11] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations (ICLR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[12] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV), pages 124-140. Springer, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Jonathan J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 16(5):550-554, 1994. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems: École D'Été de Probabilités de Saint-Flour XXXVIII-2008, volume 2033. Springer Science & Business Media, 2011. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[15] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[17] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5542-5550, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[18] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 1446-1455, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[19] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[20] Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. volume 34, pages 6155-6170, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[21] Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8046-8056, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[22] Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain generalization using causal matching. In International Conference on Machine Learning (ICML), pages 7313-7324. PMLR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[23] Toshihiko Matsuura and Tatsuya Harada. Domain generalization using a mixture of multiple latent domains. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 11749-11756, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[24] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5715-5725, 2017, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[25] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning (ICML), pages 10-18. PMLR, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In International Conference on Machine Learning (ICML), pages 7728-7738. PMLR, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7691" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.218 + ], + "angle": 0, + "content": "[29] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. In International Conference on Learning Representations (ICLR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.22, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[30] Riccardo Volpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transformation sets. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7980-7989, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[31] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. volume 31, pages 5339-5349, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[32] Haohan Wang, Zexue He, Zachary C. Lipton, and Eric P. Xing. Learning robust representations by projecting superficial statistics out. In International Conference on Learning Representations (ICLR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.468, + 0.457 + ], + "angle": 0, + "content": "[33] Shujun Wang, Lequan Yu, Caizi Li, Chi-Wing Fu, and Pheng-Ann Heng. Learning from extrinsic and intrinsic supervisions for domain generalization. In European Conference on Computer Vision (ECCV), pages 159–176. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.46, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[34] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 834-843, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.517, + 0.468, + 0.584 + ], + "angle": 0, + "content": "[35] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383-14392, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.587, + 0.468, + 0.655 + ], + "angle": 0, + "content": "[36] Xiang Xu, Xiong Zhou, Ragav Venkatesan, Gurumurthy Swaminathan, and Orchid Majumder. d-sne: Domain adaptation using stochastic neighborhood embedding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2497-2506, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.657, + 0.468, + 0.739 + ], + "angle": 0, + "content": "[37] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 14435–14447, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.742, + 0.468, + 0.809 + ], + "angle": 0, + "content": "[38] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Deep domain-adversarial image generation for domain generalisation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 13025-13032, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.812, + 0.468, + 0.867 + ], + "angle": 0, + "content": "[39] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In European Conference on Computer Vision (ECCV), pages 561-578. Springer, 2020. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7692" + } + ] +] \ No newline at end of file diff --git a/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_origin.pdf b/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3fb7b9416fbc98678c7cec2025f1166eeafa8d2c --- /dev/null +++ b/2023/Meta-Causal Learning for Single Domain Generalization/c68d967a-e5fc-4d94-a06a-72b59f93a733_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1da02135c48c8b900644f40c3f3edbdc47844865061646361356731088326789 +size 719597 diff --git a/2023/Meta-Causal Learning for Single Domain Generalization/full.md b/2023/Meta-Causal Learning for Single Domain Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1e5df9e55ad007a1a2dbdbe16ba45a00a1eff212 --- /dev/null +++ b/2023/Meta-Causal Learning for Single Domain Generalization/full.md @@ -0,0 +1,310 @@ +# Meta-causal Learning for Single Domain Generalization + +Jin Chen $^{1*}$ , Zhi Gao $^{1*}$ , Xinxiao Wu $^{1,2\dagger}$ , Jiebo Luo $^{3}$ + +$^{1}$ Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science & Technology, Beijing Institute of Technology, China $^{2}$ Guangdong Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, China + +$^{3}$ Department of Computer Science, University of Rochester, Rochester NY 14627, USA + +{chen_jin, gaozhi_2017, wuxinxiao}@bit.edu.cn, jluo@cs.rochester.edu + +# Abstract + +Single domain generalization aims to learn a model from a single training domain (source domain) and apply it to multiple unseen test domains (target domains). Existing methods focus on expanding the distribution of the training domain to cover the target domains, but without estimating the domain shift between the source and target domains. In this paper, we propose a new learning paradigm, namely simulate-analyze-reduce, which first simulates the domain shift by building an auxiliary domain as the target domain, then learns to analyze the causes of domain shift, and finally learns to reduce the domain shift for model adaptation. Under this paradigm, we propose a meta-causal learning method to learn meta-knowledge, that is, how to infer the causes of domain shift between the auxiliary and source domains during training. We use the meta-knowledge to analyze the shift between the target and source domains during testing. Specifically, we perform multiple transformations on source data to generate the auxiliary domain, perform counterfactual inference to learn to discover the causal factors of the shift between the auxiliary and source domains, and incorporate the inferred causality into factor-aware domain alignments. Extensive experiments on several benchmarks of image classification show the effectiveness of our method. + +# 1. Introduction + +Single domain generalization [28] aims to generalize a model trained using one training domain (source domain) into multiple unseen test domains (target domains). Since only one source domain is given and the target domains are out-of-distribution and unavailable during training, single + +![](images/d3fce0e146fb971161d9a04df4e696f4ab35f465b42c320aebb5423b5996db95.jpg) +Figure 1. Illustration of the simulate-analyze-reduce paradigm. In this paradigm, we first simulate the domain shift by constructing an auxiliary domain as the unseen target domain, then learn to analyze the domain shift, and finally learn to reduce the domain shift based on inferred causes. + +domain generalization is a challenging task and attracts increasing interests. Existing works have made considerable successes through expanding the distribution of the source domain by data augmentation [19,28,34] or learning adaptive data normalization [8] typically. However, such successes have been achieved without explicitly considering the domain shift between the source and target domains, which limits the generalization performance of model in real-world scenarios. + +In this paper, we propose a new learning paradigm, namely simulate-analyze-reduce, to address single domain generalization by enabling the model to analyze the real domain shift between the source domain and unseen target domain. This new paradigm is shown in Figure 1. We first build an auxiliary domain as the target domain to simulate the real domain shift between the source and target domains, since the target data is unavailable during training. We then learn to analyze the intrinsic causal factors of the domain shift to facilitate the subsequent model adapta + +tion. Finally, we learn to reduce the domain shift with its inferred causes. + +Under this paradigm, we propose a meta-causal learning method to learn the meta-knowledge about how to infer the causes of the simulated domain shift between the auxiliary and source domains via causal inference in training, and then apply the meta-knowledge to analyze the real domain shift between the target and source domains during testing, through which the source and given target domains are adaptively aligned. Specifically, we perform multiple transformations on source data to generate an auxiliary domain with great diversity. Then we build a causal graph to represent the dependency among data, variant factors, semantic concepts and category labels, and conduct counterfactual inference over the causal graph to exploit the intrinsic causality of the simulated domain shift between the auxiliary and source domains. For each sample in the auxiliary domain, we construct counterfactual scenes by intervening variant factors to infer their causal effects on the category prediction, and these inferred causal effects of variant factors can be regarded as the causes of domain shift. To reduce the domain shift, we propose a factor-aware domain alignment by learning and integrating multiple feature mappings, where an effect-to-weight network is designed to convert the causal effects of variant factors into the weights of feature mappings. + +During testing, the distribution discrepancy between the input target sample and the source domain is analyzed and reduced by applying the learnt meta-knowledge, i.e., inferring the causal effects of variant factors and incorporating them into the factor-aware domain alignment. In summary, the main contributions of this paper are as follows: + +- We propose a novel learning paradigm, simulate-analyze-reduce, for single domain generalization. This paradigm empowers the model with the ability to estimate the domain shift between the source domain and unseen target domains, thus boosting the model adaptation across different domains. +- We propose a meta-causal learning method based on counterfactual inference to learn the meta-knowledge about analyzing the intrinsic causality of domain shift, thus facilitating the reduction of domain shift. +- Our method achieves the state-of-the-art results on several benchmarks of image classification, especially on the more challenging tasks with a large domain shift, clearly demonstrating the effectiveness of our method. + +# 2. Related Work + +# 2.1. Domain Generalization + +Domain generalization focuses on generalizing a model learned from multiple source domains to the unseen target + +domain. The key difference between domain adaptation and domain generalization is that during training, domain adaptation leverages unlabelled target data while domain generalization has no access to the target domain. Existing domain generalization methods can be roughly divided into two categories: learning domain-invariant feature representation from multiple source domains [7, 10, 24, 25, 32] and generating diverse more samples via data augmentation [2, 29, 31, 39]. + +Recently, single domain generalization [28] has attracted growing attention, where only one source domain is available during training and the model is evaluated on multiple unseen target domains. A rich line of works employ data augmentation for generating out-of-domain samples to expand the distribution of the source domain [19, 28, 34]. Qiao et al. [28] propose meta-learning based adversarial domain augmentation to generate samples. Li et al. [19] propose a progressive domain expansion network to generate multiple domains progressively via simulating various photometric and geometric transforms by style transfer based generators. Wang et al. [34] propose a style-complement module to generate diverse images with different styles. Fan et al. [8] use data normalization for single domain generalization, where an adaptive normalization scheme is learned to be incorporated with adversarial domain augmentation to enhance the generalization of the model. + +The aforementioned methods aim to generate the source data distribution as diverse as possible to cover unseen target domains. When the expanded source distribution does not approximate the target distribution, the performance may significantly degrade, since there still exists a domain gap between the source and target domains. To address this problem, our method learns to analyze and reduce the domain shift by building an auxiliary domain during training. + +# 2.2. Causality for Domain Generalization + +Several recent methods exploit causality to learn domain-invariant semantic representation for domain generalization [20-22]. Considering the cross-domain invariance of the causality between semantic factors and predictions, Liu et al. [20] propose a causal semantic generative model to remove the domain-specific correlation between semantic factors and variant factors, and thus make the prediction affected only by the semantic factors. Assuming that images of the same object across domains should have the same representation, Mahajan et al. [22] use the cross-domain invariance of the causality between objects and feature representations to capture the within-class variation for domain generalization. Lv et al. [21] introduce causal inference to extract causal factors that are invariant across domains in order to learn invariant feature representation. + +From a different perspective, our method focuses on uti- + +![](images/05f48d46008de69a045f19d16947f7fbf8f250b695297fdf44e5dd76ffbcc0fa.jpg) +Figure 2. Overview of the proposed meta-causal learning method. (1) A data transformation module simulates the domain shift via generating an auxiliary domain $\mathcal{D}_a$ . (2) A counterfactual inference module analyzes the domain shift by inferring the causal effects of $V$ on $Y$ . “-” denotes comparing the values of $Y$ before and after do-operation. The edge $X \rightarrow Y$ is a dashed arrow to represent the classification model $P(Y|X)$ . (3) A factor-aware domain alignment module reduces the domain shift via multiple feature mappings according to weights learned by an effect-to-weight network. (4) A base module is used for feature extraction and classification. + +lizing causal inference to discover the intrinsic causes of domain shift by constructing counterfactual scenes over the learnt causal graph, which facilitates the reduction of domain shift and in turn benefits cross-domain adaptation. + +# 3. Method + +# 3.1. Problem Definition + +For single domain generalization, during training, we are given a labeled source domain $\mathcal{D}_s = \{(x_i^s,y_i^s)|_{i = 1}^{N_s}\}$ drawn from the distribution $P_{s}$ , where $\pmb{x}_i^s$ is the $i$ -th source sample with its category label $y_{i}^{s}\in \mathcal{V}$ . During testing, the learnt model is applied to multiple unseen target domains $\mathcal{D}_t = \{\mathcal{D}_t^j\}_{j = 1}^J$ , where $\mathcal{D}_t^j$ is the $j$ -th target domain drawn from the distribution $P_{t}^{j}$ , and $P_{t}^{j}\neq P_{s}$ . + +To bridge the domain gap between the source domain and multiple unseen target domains, we propose a new learning paradigm, called simulate-analyze-reduce, which starts with simulating the domain shift between the source and target domains, then learns to analyze the domain shift, and finally learns to reduce the domain shift. Under this paradigm, we propose a meta-causal learning method that has four components: a data transformation module to generate auxiliary domains as target domains, a counterfactual inference module to discover the causes of the domain shift, a factor-aware domain alignment module to reduce the domain shift, and a base module for feature extraction and classification, as illustrated in Figure 2. + +# 3.2. Data Transformation for Domain Shift Simulation + +To simulate the real domain shift between the source and target domains, we generate an auxiliary domain $\mathcal{D}_a$ as the unseen target domain by performing transformations on source data. The domain shift is actually the data distribution discrepancy between the source and auxiliary domains, and is usually caused by the variations of extrinsic attributes of data, independent of intrinsic semantics. Taking the image data for example, the domain shift is mainly caused by the variations of visual attributes, such as the variations of brightness, viewpoint, and color. Therefore, to make the simulated domain shift as realistic as possible, we formulate the extrinsic attributes as variant factors, and design data transformations according to the variant factors. + +We define a set of variant factors, denoted as $\mathcal{V} = \{v_{1}, v_{2}, \dots, v_{K}\}$ , where $v_{k}$ denotes the $k$ -th variant factor. Each variant factor corresponds to a data transformation function that aims to generate new data by making changes on the corresponding extrinsic attribute, denoted as $G_{v_{k}}(\boldsymbol{x}; \theta_{v_{k}})$ , where $\boldsymbol{x}$ is an input sample, and $\theta_{v_{k}} \in [g_{min}^{k}, g_{max}^{k}]$ represents the degree parameter to control the magnitude of transformation with the scale range $[g_{min}^{k}, g_{max}^{k}]$ . + +In real-world scenarios, the complex domain shift is often caused by combinatorial multiple variant factors and accordingly we design a sampling strategy to enable the combinatorial data transformation. Given a source sample $\boldsymbol{x}_i^s$ , + +we randomly sample $N_{v}^{i}$ variant factors from $\mathcal{V}$ to form a factor subset $\mathcal{V}_i = \{v_1^i,v_2^i,\dots ,v_{N_v^i}^i\}$ . Then, a corresponding auxiliary sample $\pmb{x}_i^a$ is generated by + +$$ +\boldsymbol {x} _ {i} ^ {a} = G _ {v _ {N _ {v} ^ {i}} ^ {i}} \left(\dots G _ {v _ {2} ^ {i}} \left(G _ {v _ {1} ^ {i}} \left(\boldsymbol {x} _ {i} ^ {s}; \theta_ {v _ {1} ^ {i}}\right); \theta_ {v _ {2} ^ {i}}\right) \dots ; \theta_ {v _ {N _ {v} ^ {i}} ^ {i}}\right), \tag {1} +$$ + +where the transformation degree parameter $\theta_{v_k^i}$ is randomly selected from its range scale. In this way, the domain shift is generated as diverse as possible to approximate the real domain shift. + +# 3.3. Counterfactual Inference for Domain Shift Analysis + +After simulating the domain shift, we introduce counterfactual inference to analyze the domain shift. During training, we learn the meta-knowledge about how to infer the causes of data discrepancy between one auxiliary sample and the source domain, and during testing, we apply the learnt meta-knowledge to unseen target samples. + +We build a causal graph to model the causal dependency among the input sample (node $X$ ), the variant factors (node $V$ ), the semantic concepts (node $S$ ), and the output category (node $Y$ ), as shown in Figure 2 (2). The semantic concepts denote the intrinsic attributes of data that are related to the category, and the variant factors denote the extrinsic attributes of data that are domain specific, independent of the intrinsic semantics. For example, when the input sample is an image of "zebra", the semantic concepts are like "four legs" and "black-white stripes", and the variant factors include brightness, viewpoint, and so on. The edge $S \rightarrow Y$ represents that the category of the input data is determined by the semantic concepts. The edges $S \rightarrow X$ and $V \rightarrow X$ represent that the semantic concepts and the variant factors together determine what the input sample looks like. Since the semantic concepts represent the intrinsic semantics and are invariant across different domains, ideally the cross-domain classification is only determined by the semantic concepts, which is implemented by learning the edge $S \rightarrow Y$ , i.e., estimating the conditional probability $P(Y|S)$ . However, in reality, the semantic concepts are unobserved from the input samples, so the classification should be implemented by learning the edge $X \rightarrow Y$ , i.e., estimating the conditional probability $P(Y|X)$ . The edge $X \rightarrow Y$ is a dashed arrow to represent the classification model $P(Y|X)$ . Since the input sample node $X$ is affected by the semantic concept node $S$ and the variant factor node $V$ together, the category node $Y$ is also affected by the variant factor node $V$ through edges $V \rightarrow X \rightarrow Y$ . As the variant factors are domain-specific, their causal effects on the category lead to the domain shift. Hence, we infer the causal effects of the variant factors on the category prediction to discover the causes of the domain shift. That is to say, we infer the causal effects of the node $V$ on the node + +$Y$ as causes of domain shift, and then learn to reduce the domain shift based on its causes. + +To infer the causal effects of the node $V$ on the node $Y$ , we first learn the edge $X \to Y$ using the source data $\mathcal{D}_s$ by a classification loss: + +$$ +\mathcal {L} _ {c} = \mathbb {E} _ {\left(\boldsymbol {x} _ {i} ^ {s}, y _ {i} ^ {s}\right) \sim P _ {s}} \left[ - \sum_ {u} \mathbb {I} _ {u = y _ {i} ^ {s}} \log C \left(F \left(\boldsymbol {x} _ {i} ^ {s}\right)\right) _ {u} \right], \tag {2} +$$ + +where $(\pmb{x}_i^s, y_i^s) \in \mathcal{D}_s$ , $F$ is a feature extractor, $C$ is the classifier to output $|\mathcal{V}|$ category probabilities, $C(\cdot)_u$ is the $u$ -th element of $|\mathcal{V}|$ category probabilities, and $\mathbb{I}_{u=y_i^s}$ is an indicator function, meaning that if $u = y_i^s$ , the value of $\mathbb{I}_{u=y_i^s}$ is 1 and 0 otherwise. Then, for each sample from the auxiliary domain, we construct a factual scene and multiple counterfactual scenes over the causal graph, so as to infer the causal effects of variant factors on the category prediction. Given an auxiliary sample $\pmb{x}_i^a \in \mathcal{D}_a$ , its factual category is predicted by + +$$ +\boldsymbol {y} _ {i} ^ {a} = P (Y | X) = C \left(F \left(\boldsymbol {x} _ {i} ^ {a}\right)\right), \tag {3} +$$ + +where $\pmb{y}_i^a \in \mathbb{R}^{|\mathcal{Y}|}$ is a category probability vector, and represents the value of node $Y$ . For each variant factor, we construct a counterfactual scene to infer its causal effect by doing intervention on the variant factor node $V$ . Let $do(V = v_k)$ denote the intervention on the node $V$ , which is implemented by changing extrinsic attributes of data through the transformation $G_{v_k}$ with multiple degree parameters. Accordingly, the counterfactual category of $\pmb{x}_i^a$ is predicted by + +$$ +\begin{array}{l} \boldsymbol {y} _ {i, v _ {k}} ^ {a} = P \left(Y | X, d o (V = v _ {k})\right) \\ = \frac {1}{| \mathcal {M} |} \sum_ {\theta_ {v _ {k}} \in \mathcal {M}} C \left(F \left(G _ {v _ {k}} \left(\boldsymbol {x} _ {i} ^ {a}; \theta_ {v _ {k}}\right)\right)\right), \tag {4} \\ \end{array} +$$ + +where $\mathcal{M}$ is a set of degree parameters, obtained by uniformly sampling magnitudes of transformation $G_{v_k}$ from the scale range. By comparing the factual and counterfactual category probabilities of $\pmb{x}_i^a$ , the causal effect of the $k$ -th variant factor is calculated by + +$$ +\boldsymbol {e} _ {k} ^ {i} = P (Y | X) - P \left(Y | X, d o (V = v _ {k})\right) = \boldsymbol {y} _ {i} ^ {a} - \boldsymbol {y} _ {i, v _ {k}} ^ {a}. \tag {5} +$$ + +The inferred causal effect represents how much contribution the corresponding variant factor makes to the domain shift, and the larger the causal effect is, the more seriously the domain shift is caused by the variant factor. + +# 3.4. Factor-aware Domain Alignment for Domain Shift Reduction + +After analyzing the causes of domain shift, we propose a factor-aware domain alignment to reduce the domain shift by learning multiple feature mappings, with guidance of the + +![](images/8b18c8e4206dffb7d06e3e53f5b175ba03594c1f3d3fe28e73191aa432b087a8.jpg) +Figure 3. Inference process of meta-causal learning for a given target sample. + +inferred causal effects of variant factors. Each feature mapping addresses a specific domain shift caused by one variant factor. We construct $K$ feature mappings for $K$ variant factors, and the $k$ -th feature mapping aims to address the domain shift caused by the $k$ -th variant factor. In order to incorporate the causal effects of variant factors into the learning of mappings, we build an effect-to-weight network that converts the causal effect of each variant factor into the weight of the corresponding feature mapping. + +For the auxiliary sample $\pmb{x}_i^a$ and its inferred causal effects of all variant factors $\{e_1^i, e_2^i, \dots, e_K^i\}$ , the weights of feature mappings are calculated by + +$$ +\boldsymbol {\omega} ^ {i} = \operatorname {s o f t m a x} \big (W (\boldsymbol {e} _ {1} ^ {i}), W (\boldsymbol {e} _ {2} ^ {i}), \dots , W (\boldsymbol {e} _ {K} ^ {i}) \big), \quad (6) +$$ + +where $\omega^i\in \mathbb{R}^K$ and its $k$ -th element $\omega_{k}^{i}$ denotes the weight of the $k$ -th factor-aware feature mapping. $W(\cdot)$ denotes the effect-to-weight network. + +According to the weights of feature mappings, we integrate the $K$ feature mappings to project the auxiliary samples into the source feature space. The alignment of the source and auxiliary domains is implemented by minimizing the feature distance between the source and auxiliary samples in the source feature space. Given a source sample $\pmb{x}_i^s$ with the corresponding category label $y_i^s$ , generated auxiliary sample $\pmb{x}_i^a$ and inferred mapping weights $\omega^i$ , the alignment loss incorporated with inferred causal effects is defined as + +$$ +\begin{array}{l} \mathcal {L} _ {a} ^ {c} = \frac {1}{N _ {s}} \sum_ {i} | | F \left(\boldsymbol {x} _ {i} ^ {s}\right) - \sum_ {k} \omega_ {k} ^ {i} M _ {k} \left(F \left(\boldsymbol {x} _ {i} ^ {a}\right)\right) | | _ {2} \tag {7} \\ + \frac {1}{N _ {s}} \sum_ {i} \mathcal {H} \left(C \left(\sum_ {k} \omega_ {k} ^ {i} M _ {k} \left(F \left(\boldsymbol {x} _ {i} ^ {a}\right)\right), y _ {i} ^ {s}\right)\right). \\ \end{array} +$$ + +where $F$ is the feature extractor, $M_{k}$ is the $k$ -th feature mapping, $C$ is the classifier, $\omega_{k}^{i}$ is the $k$ -th element of $\omega^i$ , $\sum_{k}\omega_{k}^{i}M_{k}\big(F(\pmb{x}_{i}^{a})\big)$ represents the projected feature of $\pmb{x}_i^a$ , and $N_{s}$ is the number of source samples. $||\cdot ||_2$ is the L2-loss, and $\mathcal{H}(\cdot)$ is the cross-entropy loss. The first term measures the feature distance between the source and auxiliary samples after projection, i.e., the data distribution discrepancy + +between the source and auxiliary domains. The second term encourages the auxiliary samples to belong to the same categories as the source samples. + +Moreover, to enable each feature mapping to address the specific domain shift caused by the corresponding variant factor, we introduce another alignment loss $\mathcal{L}_a^m$ : + +$$ +\begin{array}{l} \mathcal {L} _ {a} ^ {m} = \frac {1}{N _ {s}} \frac {1}{K} \sum_ {i} \sum_ {k} | | F (\boldsymbol {x} _ {i} ^ {s}) - M _ {k} (F (\boldsymbol {x} _ {i} ^ {k})) | | _ {2} \\ + \frac {1}{N _ {s}} \frac {1}{K} \sum_ {i} \sum_ {k} \mathcal {H} \left(C \left(M _ {k} \left(F \left(\boldsymbol {x} _ {i} ^ {k}\right)\right), y _ {i} ^ {s}\right)\right), \tag {8} \\ \end{array} +$$ + +where $\pmb{x}_i^k$ is an sample generated by conducting the data transformation $G_{v_k}$ on the source sample $\pmb{x}_i^s$ , denoted by $\pmb{x}_i^k = G_{v_k}(\pmb{x}_i^s;\theta_{v_k})$ . The transformation degree parameter $\theta_{v_k}$ is randomly selected from its range scale. + +Then the overall loss function is defined as + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \mathcal {L} _ {a} ^ {c} + \mathcal {L} _ {a} ^ {m}. \tag {9} +$$ + +The whole training process is summarized in Algorithm 1. + +# 3.5. Inference + +During testing, given a target sample $\boldsymbol{x}_i^t$ , firstly, the causal effects of variant factors are inferred by counterfactual inference in Eq. (5), and the weights $\omega^i$ of feature mappings are calculated by the effect-to-weight network in Eq. (6). Then the target feature $\boldsymbol{f}_i^t = F(\boldsymbol{x}_i^t)$ is projected into the source feature space by integrating $K$ feature mappings according to their weights $\omega^i$ to obtain the projected feature $\hat{\boldsymbol{f}}_i^t = \sum_k \omega_k^i M_k(\boldsymbol{f}_i^t)$ . Finally, the category of $\boldsymbol{x}_i^t$ is predicted by $C(\hat{\boldsymbol{f}}_i^t)$ . The whole inference process is shown in Figure 3. + +# 4. Experiments + +# 4.1. Datasets + +Digits. The Digits dataset consists of five datasets: MNIST [16], MNIST-M [9], SVHN [26], USPS [13], and + +# Algorithm 1 Training process + +Input: The source domain $\mathcal{D}_s$ , the variant factor set $\nu$ . + +Output: The feature extractor $F$ , the classifier $C$ , $K$ feature mappings $\{M_i\}_{i=1}^K$ , the effect-to-weight network $W$ . + +1: Initialize $F, C, \{M_i\}_{i=1}^K, W$ ; +2: while not converge do +3: Sample an image $\pmb{x}_j^s$ from $\mathcal{D}_s$ ; +4: Construct a factor subset $\mathcal{V}_i$ by sampling $N_v^i$ variant factors from $\mathcal{V}$ ; +5: Generate an auxiliary sample $\pmb{x}_i^a$ by Eq.(1); +6: for $v_{k}$ in $\mathcal{V}$ do +7: Calculate the factual category $\mathbf{y}_i^a$ of auxiliary sample $\mathbf{x}_i^a$ by Eq.(3); +8: Calculate the counterfactual category $\pmb{y}_{i,v_k}^a$ of auxiliary sample $\pmb{x}_i^a$ by Eq.(4); +9: Infer the causal effect $e_k^i$ of variant factor $v_k$ via comparing $y_i^a$ and $y_{i,v_k}^a$ by Eq.(5); + +10: end for +11: Calculate the weights of $K$ feature mappings $\{M_i\}_{i=1}^K$ by Eq.(6); +12: Update $F, C, \{M_i\}_{i=1}^K$ , $W$ by Eq.(9). +13: end while + +SYN [9], with 10 categories. We use MNIST as the source domain, and the other four datasets as the target domains. The first 10,000 images in the training set of MNIST are used for training. + +CIFAR10-C. The CIFAR10-C dataset [11] is proposed to evaluate the robustness of classification model. The images are corrupted from the test set of the CIFAR10 dataset [15] by 19 corruption types with five levels of severity. A higher level means the more serious corruption. There are 10 categories. We use CIFAR10 as the source domain, and CIFAR10-C as the target domains where images of one severity level form one target domain. + +PACS. The PACS dataset [17] is a benchmark for domain generalization, and consists of four domains: art painting, cartoon, photo, and sketch. There are 9,991 images of seven categories. We use one domain as the source domain, and the rest three domains as the target domains. So there are four tasks with different domains as the source domains. + +# 4.2. Implementation Details + +Auxiliary Domain. We define 16 variant factors to generate the images in the auxiliary domain, including 12 photometric factors (Brightness, Contrast, Color, Sharpness, Auto-Contrast, Invert, Equalize, Solarize, SolarizeAdd, Posterize, NoiseSalt, NoiseGaussian) and 4 geometric factors (Shear-X, Shear-Y, Rotate, Flip). Since the Rotate and Flip variant factors will affect the semantic information of digit images, we use the other 14 variant factors for the Digits dataset. For + +Table 1. Single domain generalization results (%) on Digits with ConvNet as backbone. The model is trained on MNIST, and evaluated on SVHN, SYN, MNIST-M, and USPS. + +
MethodSVHNSYNMNIST-MUSPSAvg
ERM [14]27.8339.6552.7276.9449.29
CCSA [24]25.8937.3149.2983.7249.05
d-SNE [36]26.2237.8350.9893.1652.05
JiGen [2]33.8043.7957.8077.1553.14
GUD [31]35.5145.3260.4177.2654.62
M-ADA [28]42.5548.9567.9478.5359.49
ME-ADA [37]42.5650.3963.2781.0459.32
PDEN [19]62.2169.3982.2085.2674.77
L2D [34]62.8663.7287.3083.9774.46
AA [4]45.2364.5260.5380.6262.72
RA [5]54.7759.6074.0577.3366.44
RSDA [30]47.4062.0081.5083.1068.50
RSDA+ASR [8]52.8064.5080.8082.4070.10
Ours69.9478.4778.3488.5478.82
+ +Table 2. Single domain generalization results (\%) on CIFAR10-C with WRN as backbone. Each level is viewed as a target domain, a higher level denotes the more serious corruption and the domain discrepancy between the source and target domains is larger. + +
Methodlevel1level2level3level4level5Avg
ERM [14]87.8081.5075.5068.2056.1073.82
GUD [31]88.3083.5077.6070.6058.3075.66
M-ADA [28]90.5086.8082.5076.4065.6080.36
PDEN [19]90.6288.9187.0383.7177.4785.55
AA [4]91.4287.8884.1078.4671.1382.60
RA [5]91.7488.8985.8281.0374.9384.48
Ours92.3891.2289.8887.7384.5289.15
+ +the CIFAR10-C and PACS datasets, all 16 variant factors are used. To make the auxiliary domain as diverse as possible, for each source image at each iteration, we randomly sample several variant factors and use them to generate a new auxiliary image. + +# 4.3. Results on Single Domain Generalization + +We compare our method with several state-of-the-art methods, including the baseline method (ERM [14]), the methods of learning domain-invariant features (CCSA [24], d-SNE [36], JiGen [2]), and the methods of making data augmentation (GUD [31], M-ADA [28], ME-ADA [37], PDEN [19], L2D [34], AA [4], RA [5], RSDA [30], RSC [12], ASR [8]). + +Table 1, Table 2, and Table 3 show the comparison results on Digits, CIFAR10-C, and PACS, respectively. From the results, there are several interesting observations as follows. First, our method generally outperforms the compared methods on all datasets, which clearly shows the effectiveness of the proposed simulate-analyze-reduce learning paradigm for single domain generalization. Second, compared with the data augmentation methods (e.g., + +Table 3. Single domain generalization results (\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as the target domains. + +
MethodArtpaintCartoonSketchPhotoAvg
ERM [14]70.9076.5053.1042.2060.70
RSC [12]73.4075.9056.2041.6061.80
RSC+ASR [8]76.7079.3061.6054.6068.10
Ours77.1380.1462.5559.6069.86
+ +Table 4. Leave-one-domain-out results (\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the target domain and the other three domains are used as source domains. + +
MethodArtpaintCartoonPhotoSketchAvg
MetaReg [1]83.7077.2095.5070.3081.70
GUD [31]78.3277.6595.6174.2181.44
Epi-FCR [18]82.1077.0093.9073.0081.50
MASF [6]80.2977.1794.9971.6881.03
JiGen [2]79.4275.2596.0371.3580.51
DMG [3]76.9080.3893.5575.2181.46
DDAIG [38]84.2078.1095.3074.7083.10
CSD [27]78.9075.8094.1076.7081.40
L2A-OT [39]83.3078.2096.2073.6082.80
EISNet [33]81.8976.4495.9374.3382.15
RSC [12]83.4380.3195.9980.8585.15
ME-ADA [37]78.6178.6595.5775.5982.10
MMLD [23]81.2877.1696.0972.2981.83
L2D [34]81.4479.5695.5180.5884.27
FACT [35]85.3778.3895.1579.1584.51
MatchDG [22]81.3280.7096.5379.7284.57
CIRL [21]86.0880.5995.9382.6786.32
Ours85.3080.9396.5385.2487.00
+ +Table 5. Ablation study (\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as target domains. “T”, “A”, “C” denote Domain Transformation, Domain Alignment, and Counterfactual Inference, respectively. + +
MethodTACArtpaintCartoonSketchPhotoAvg
Base71.2667.6443.9736.9954.97
DT✓75.2878.4659.4556.0967.32
DTA✓✓71.6472.7857.1152.0263.39
Ours✓✓✓77.1380.1462.5559.6069.86
+ +AA [4], RA [5] and RSDA [30]) that are more related to our method, our method achieves much better results, and especially yields a $10.32\%$ gain over RSDA on Digits, strongly suggesting that it is beneficial to empower the model with the ability of analyzing the causes of domain shift by counterfactual inference. Third, on more difficult tasks with larger domain shift (e.g., SVHN on Digits, level5 on CIFAR10-C, and Photo on PACS), our method significantly improves the performance, further demonstrating the + +superiority of our method on handling more challenging situations. Forth, our method performs little worse on MNIST-M and USPS of Digits. The possible reason is that other compared methods use more well-designed data augmentation and network regularization, such as AdaIN based generators in PDEN [19] and stochastic neighborhood embedding techniques in d-SNE [36]. + +# 4.4. Results on Multiple Domain Generalization + +We extend the proposed method to multi-source domain setting by regarding the multiple source domains as one source domain without using domain labels. We employ the leave-one-domain-out protocol following existing multi-source domain generalization [21, 35]. We compare our method with most related methods that introduces causal inference into domain generalization (MatchDG [22], CIRL [21]), and existing popular domain generalization methods (MetaReg [1], GUD [31], EpiFCR [18], MASF [6], JiGen [2], DMG [3], DDAIG [38], CSD [27], L2A-OT [39], EISNet [33], RSC [12], ME-ADA [37], MMLD [23], L2D [34], FACT [35]). + +Table 4 shows the leave-one-domain-out results on the PACS dataset with ResNet-18 as backbone. From the results, we make several observations. First, it is noteworthy that our method achieves the state-of-the-art overall performance ("Avg") although our method is not designed for multi-source domain generalization, which further demonstrates that the proposed simulate-analyze-reduce learning paradigm not only benefits single domain generalization but also boosts multi-source domain generalization. Second, compared with the methods of introducing causal inference to learn domain-invariant features (MatchDG [22], CIRL [21]), our method achieves better results on the overall metric "Avg", especially making $5.52\%$ and $2.57\%$ gains over MatchDG [22] and CIRL [21], respectively, on the more challenging task (Sketch $\rightarrow$ Others). Such improvements are attributed to the ability of analyzing and reducing the domain shift, further demonstrating the advantages of causal inference in analyzing the causes of the domain shift. + +# 4.5. Ablation Studies + +To evaluate each component of our method, we conduct ablation experiments on the PACS dataset. We design several degraded variants of our method for comparison: (1) "Base", where only the base module is utilized and optimized by Eq. (2) using the source domain; (2) "DT", where the data transformation module is added into the base module and the model is trained using both source and auxiliary domains; (3) "DTA", where the simulated domain shift between auxiliary domain and source domain is directly reduced without analyzing the causes of the domain shift via the proposed counterfactual module. + +![](images/fa0dca8fe5dba2c2c953509f91624b91296fab9d7e5aef196994c4972cabb1ec.jpg) +Figure 4. Examples of the inferred causal effects (represented as weights) of variant factors. The left part shows a source image from the CIFAR10 dataset and three target images from the CIFAR10-C dataset with Gaussian noise corruption. As the corruption severity increases from level 1 to level 5, the inferred weights of the NoiseGaussian variant factor become larger accordingly. + +Table 6. Factor analysis (\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as target domains. + +
MethodArtpaintCartoonSketchPhotoAvg
Ours_GT74.6269.7852.7443.3460.12
Ours_PT74.4773.4850.3551.6762.49
Ours77.1380.1462.5559.6069.86
+ +The results are shown in Table 5. From the results, we make several observations. First, our method achieves better performance than "DT", which validates the superiority of our method on analyzing and reducing the domain shift, rather than directly enlarging the source data distribution. Second, "DTA" performs worse than "DT", probably because brute-force domain alignment without analyzing causes of the domain shift leads to negative transfer and thus hurts the performance. Third, our method achieves the best performance thanks to the proposed simulate-analyze-reduce paradigm. + +# 4.6. Factor Analysis + +In order to further analyze the effect of variant factors, we design several variants of our method by using different variant factors, including only using geometric factors ("Ours_GT"), and only using photometric factors ("Ours_PT"). Since the factor number of the two type factors are different, we randomly select 4 photometric factors to keep the same number as geometric factors and repeat experiments 10 times to avoid the effect of sampling. The results are shown in Table 6. From the results, it is noteworthy that "Ours_PT" performs better than "Ours_GT", especially with the $8.33\%$ gain when using Photo as the source domain. The reason may be that the domain shift between Photo and the other target domains is mainly caused by photometric factors. Moreover, our method outperforms both "Ours_PT" and "Ours_GT", showing that both photometric factors and geometric factors are required to simulate the + +domain shift as diverse as possible. + +# 4.7. Causality Visualization + +In Figure 4, we visualize the inferred causal effects (represented by the weights) of variant factors for the unseen target images on the CIFAR10-C dataset during testing. The target images are actually corrupted from the source images of CIFAR10 by Gaussian noise of five levels. It is interesting to observe that the inferred weights of the NoiseGaussian variant factor are larger than that of the other variant factors, indicating that the counterfactual inference succeeds in discovering the real cause of the domain shift. We can also observe that when the corruption severity of Gaussian noise increases from level 1 to level 5 (i.e., the domain shift is more serious), the inferred weights of the NoiseGaussian variant factor become larger, which further demonstrates that our method measures the magnitude of the domain shift correctly. + +# 5. Conclusion + +We have presented a new paradigm, simulate-analyze-reduce, for single domain generalization. Our paradigm empowers the model with the ability to analyze the domain shift, instead of directly expanding the distribution of the source domain to cover unseen target domains. Under this paradigm, we have presented a meta-causal learning method that can learn meta-knowledge about inferring the causes of domain shift during training, and apply such meta-knowledge to reduce the domain shift for boosting adaptation during testing. Extensive experiments on several benchmark datasets have validated the effectiveness of the new learning paradigm and the advantage of meta-causal learning on analyzing the domain shift for domain generalization. + +Acknowledgments. This work was supported in part by the Natural Science Foundation of China (NSFC) under Grant No 62072041. + +# References + +[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. volume 31, pages 1006-1016, 2018. 7 +[2] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2229-2238, 2019. 2, 6, 7 +[3] Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In European Conference on Computer Vision (ECCV), pages 301-318. Springer, 2020. 7 +[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 113-123, 2019. 6, 7 +[5] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 702-703, 2020. 6, 7 +[6] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. volume 32, pages 6447–6458, 2019. 7 +[7] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In European Conference on Computer Vision (ECCV), pages 200-216. Springer, 2020. 2 +[8] Xinjie Fan, Qifei Wang, Junjie Ke, Feng Yang, Boqing Gong, and Mingyuan Zhou. Adversarily adaptive normalization for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8208-8217, 2021. 1, 2, 6, 7 +[9] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), pages 1180-1189. PMLR, 2015. 5, 6 +[10] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2551-2559, 2015. 2 +[11] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations (ICLR), 2019. 6 +[12] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV), pages 124-140. Springer, 2020. 6, 7 +[13] Jonathan J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 16(5):550-554, 1994. 5 + +[14] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems: École D'Été de Probabilités de Saint-Flour XXXVIII-2008, volume 2033. Springer Science & Business Media, 2011. 6, 7 +[15] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[16] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5 +[17] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5542-5550, 2017. 6 +[18] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 1446-1455, 2019. 7 +[19] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 1, 2, 6, 7 +[20] Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. volume 34, pages 6155-6170, 2021. 2 +[21] Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8046-8056, 2022. 2, 7 +[22] Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain generalization using causal matching. In International Conference on Machine Learning (ICML), pages 7313-7324. PMLR, 2021. 2, 7 +[23] Toshihiko Matsuura and Tatsuya Harada. Domain generalization using a mixture of multiple latent domains. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 11749-11756, 2020. 7 +[24] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5715-5725, 2017, 2, 6 +[25] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning (ICML), pages 10-18. PMLR, 2013. 2 +[26] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 5 +[27] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In International Conference on Machine Learning (ICML), pages 7728-7738. PMLR, 2020. 7 + +[28] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 1, 2, 6 +[29] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. In International Conference on Learning Representations (ICLR), 2018. 2 +[30] Riccardo Volpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transformation sets. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7980-7989, 2019. 6, 7 +[31] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. volume 31, pages 5339-5349, 2018. 2, 6, 7 +[32] Haohan Wang, Zexue He, Zachary C. Lipton, and Eric P. Xing. Learning robust representations by projecting superficial statistics out. In International Conference on Learning Representations (ICLR), 2019. 2 +[33] Shujun Wang, Lequan Yu, Caizi Li, Chi-Wing Fu, and Pheng-Ann Heng. Learning from extrinsic and intrinsic supervisions for domain generalization. In European Conference on Computer Vision (ECCV), pages 159–176. Springer, 2020. 7 +[34] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 834-843, 2021. 1, 2, 6, 7 +[35] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383-14392, 2021. 7 +[36] Xiang Xu, Xiong Zhou, Ragav Venkatesan, Gurumurthy Swaminathan, and Orchid Majumder. d-sne: Domain adaptation using stochastic neighborhood embedding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2497-2506, 2019. 6, 7 +[37] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 14435–14447, 2020. 6, 7 +[38] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Deep domain-adversarial image generation for domain generalisation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 13025-13032, 2020. 7 +[39] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In European Conference on Computer Vision (ECCV), pages 561-578. Springer, 2020. 2, 7 \ No newline at end of file diff --git a/2023/Meta-Causal Learning for Single Domain Generalization/images.zip b/2023/Meta-Causal Learning for Single Domain Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e94bfde8db0768d1003148865ec78c30ccb76c9 --- /dev/null +++ b/2023/Meta-Causal Learning for Single Domain Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cad3fab13ceb2677c78324e18f14908decdaa955e2827c025458fc1f45d1a0f +size 592322 diff --git a/2023/Meta-Causal Learning for Single Domain Generalization/layout.json b/2023/Meta-Causal Learning for Single Domain Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..afb09f8b1f4f34ef182a0aa4c1d856513f447289 --- /dev/null +++ b/2023/Meta-Causal Learning for Single Domain Generalization/layout.json @@ -0,0 +1,9646 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 470, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 470, + 120 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 470, + 120 + ], + "type": "text", + "content": "Meta-causal Learning for Single Domain Generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "spans": [ + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "text", + "content": "Jin Chen" + }, + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "text", + "content": ", Zhi Gao" + }, + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "text", + "content": ", Xinxiao Wu" + }, + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "text", + "content": ", Jiebo Luo" + }, + { + "bbox": [ + 173, + 142, + 418, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 158, + 510, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 158, + 510, + 213 + ], + "spans": [ + { + "bbox": [ + 96, + 158, + 510, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 158, + 510, + 213 + ], + "type": "text", + "content": "Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science & Technology, Beijing Institute of Technology, China " + }, + { + "bbox": [ + 96, + 158, + 510, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 158, + 510, + 213 + ], + "type": "text", + "content": "Guangdong Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 213, + 507, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 213, + 507, + 228 + ], + "spans": [ + { + "bbox": [ + 85, + 213, + 507, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 213, + 507, + 228 + ], + "type": "text", + "content": "Department of Computer Science, University of Rochester, Rochester NY 14627, USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 229, + 470, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 229, + 470, + 241 + ], + "spans": [ + { + "bbox": [ + 121, + 229, + 470, + 241 + ], + "type": "text", + "content": "{chen_jin, gaozhi_2017, wuxinxiao}@bit.edu.cn, jluo@cs.rochester.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 294, + 290, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 290, + 582 + ], + "type": "text", + "content": "Single domain generalization aims to learn a model from a single training domain (source domain) and apply it to multiple unseen test domains (target domains). Existing methods focus on expanding the distribution of the training domain to cover the target domains, but without estimating the domain shift between the source and target domains. In this paper, we propose a new learning paradigm, namely simulate-analyze-reduce, which first simulates the domain shift by building an auxiliary domain as the target domain, then learns to analyze the causes of domain shift, and finally learns to reduce the domain shift for model adaptation. Under this paradigm, we propose a meta-causal learning method to learn meta-knowledge, that is, how to infer the causes of domain shift between the auxiliary and source domains during training. We use the meta-knowledge to analyze the shift between the target and source domains during testing. Specifically, we perform multiple transformations on source data to generate the auxiliary domain, perform counterfactual inference to learn to discover the causal factors of the shift between the auxiliary and source domains, and incorporate the inferred causality into factor-aware domain alignments. Extensive experiments on several benchmarks of image classification show the effectiveness of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 624, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 685 + ], + "type": "text", + "content": "Single domain generalization [28] aims to generalize a model trained using one training domain (source domain) into multiple unseen test domains (target domains). Since only one source domain is given and the target domains are out-of-distribution and unavailable during training, single" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 268, + 539, + 396 + ], + "blocks": [ + { + "bbox": [ + 315, + 268, + 539, + 396 + ], + "lines": [ + { + "bbox": [ + 315, + 268, + 539, + 396 + ], + "spans": [ + { + "bbox": [ + 315, + 268, + 539, + 396 + ], + "type": "image", + "image_path": "d3fce0e146fb971161d9a04df4e696f4ab35f465b42c320aebb5423b5996db95.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 403, + 547, + 460 + ], + "lines": [ + { + "bbox": [ + 305, + 403, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 547, + 460 + ], + "type": "text", + "content": "Figure 1. Illustration of the simulate-analyze-reduce paradigm. In this paradigm, we first simulate the domain shift by constructing an auxiliary domain as the unseen target domain, then learn to analyze the domain shift, and finally learn to reduce the domain shift based on inferred causes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 484, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 591 + ], + "type": "text", + "content": "domain generalization is a challenging task and attracts increasing interests. Existing works have made considerable successes through expanding the distribution of the source domain by data augmentation [19,28,34] or learning adaptive data normalization [8] typically. However, such successes have been achieved without explicitly considering the domain shift between the source and target domains, which limits the generalization performance of model in real-world scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we propose a new learning paradigm, namely simulate-analyze-reduce, to address single domain generalization by enabling the model to analyze the real domain shift between the source domain and unseen target domain. This new paradigm is shown in Figure 1. We first build an auxiliary domain as the target domain to simulate the real domain shift between the source and target domains, since the target data is unavailable during training. We then learn to analyze the intrinsic causal factors of the domain shift to facilitate the subsequent model adapta" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 205, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 205, + 702 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 205, + 702 + ], + "type": "text", + "content": "* Jin Chen and Zhi Gao are co-first authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 61, + 703, + 186, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 186, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 186, + 712 + ], + "type": "text", + "content": "† Corresponding author: Xinxiao Wu." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7683" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "tion. Finally, we learn to reduce the domain shift with its inferred causes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 96, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 96, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 289, + 384 + ], + "type": "text", + "content": "Under this paradigm, we propose a meta-causal learning method to learn the meta-knowledge about how to infer the causes of the simulated domain shift between the auxiliary and source domains via causal inference in training, and then apply the meta-knowledge to analyze the real domain shift between the target and source domains during testing, through which the source and given target domains are adaptively aligned. Specifically, we perform multiple transformations on source data to generate an auxiliary domain with great diversity. Then we build a causal graph to represent the dependency among data, variant factors, semantic concepts and category labels, and conduct counterfactual inference over the causal graph to exploit the intrinsic causality of the simulated domain shift between the auxiliary and source domains. For each sample in the auxiliary domain, we construct counterfactual scenes by intervening variant factors to infer their causal effects on the category prediction, and these inferred causal effects of variant factors can be regarded as the causes of domain shift. To reduce the domain shift, we propose a factor-aware domain alignment by learning and integrating multiple feature mappings, where an effect-to-weight network is designed to convert the causal effects of variant factors into the weights of feature mappings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 384, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 289, + 456 + ], + "type": "text", + "content": "During testing, the distribution discrepancy between the input target sample and the source domain is analyzed and reduced by applying the learnt meta-knowledge, i.e., inferring the causal effects of variant factors and incorporating them into the factor-aware domain alignment. In summary, the main contributions of this paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 460, + 287, + 642 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 460, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 460, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 58, + 460, + 287, + 533 + ], + "type": "text", + "content": "- We propose a novel learning paradigm, simulate-analyze-reduce, for single domain generalization. This paradigm empowers the model with the ability to estimate the domain shift between the source domain and unseen target domains, thus boosting the model adaptation across different domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 540, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 540, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 58, + 540, + 287, + 587 + ], + "type": "text", + "content": "- We propose a meta-causal learning method based on counterfactual inference to learn the meta-knowledge about analyzing the intrinsic causality of domain shift, thus facilitating the reduction of domain shift." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 594, + 286, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 594, + 286, + 642 + ], + "spans": [ + { + "bbox": [ + 58, + 594, + 286, + 642 + ], + "type": "text", + "content": "- Our method achieves the state-of-the-art results on several benchmarks of image classification, especially on the more challenging tasks with a large domain shift, clearly demonstrating the effectiveness of our method." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 179, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 179, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 179, + 682 + ], + "type": "text", + "content": "2.1. Domain Generalization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Domain generalization focuses on generalizing a model learned from multiple source domains to the unseen target" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "domain. The key difference between domain adaptation and domain generalization is that during training, domain adaptation leverages unlabelled target data while domain generalization has no access to the target domain. Existing domain generalization methods can be roughly divided into two categories: learning domain-invariant feature representation from multiple source domains [7, 10, 24, 25, 32] and generating diverse more samples via data augmentation [2, 29, 31, 39]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 181, + 546, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 546, + 395 + ], + "type": "text", + "content": "Recently, single domain generalization [28] has attracted growing attention, where only one source domain is available during training and the model is evaluated on multiple unseen target domains. A rich line of works employ data augmentation for generating out-of-domain samples to expand the distribution of the source domain [19, 28, 34]. Qiao et al. [28] propose meta-learning based adversarial domain augmentation to generate samples. Li et al. [19] propose a progressive domain expansion network to generate multiple domains progressively via simulating various photometric and geometric transforms by style transfer based generators. Wang et al. [34] propose a style-complement module to generate diverse images with different styles. Fan et al. [8] use data normalization for single domain generalization, where an adaptive normalization scheme is learned to be incorporated with adversarial domain augmentation to enhance the generalization of the model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 396, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 493 + ], + "type": "text", + "content": "The aforementioned methods aim to generate the source data distribution as diverse as possible to cover unseen target domains. When the expanded source distribution does not approximate the target distribution, the performance may significantly degrade, since there still exists a domain gap between the source and target domains. To address this problem, our method learns to analyze and reduce the domain shift by building an auxiliary domain during training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 502, + 501, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 501, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 501, + 514 + ], + "type": "text", + "content": "2.2. Causality for Domain Generalization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": "Several recent methods exploit causality to learn domain-invariant semantic representation for domain generalization [20-22]. Considering the cross-domain invariance of the causality between semantic factors and predictions, Liu et al. [20] propose a causal semantic generative model to remove the domain-specific correlation between semantic factors and variant factors, and thus make the prediction affected only by the semantic factors. Assuming that images of the same object across domains should have the same representation, Mahajan et al. [22] use the cross-domain invariance of the causality between objects and feature representations to capture the within-class variation for domain generalization. Lv et al. [21] introduce causal inference to extract causal factors that are invariant across domains in order to learn invariant feature representation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "From a different perspective, our method focuses on uti-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7684" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 527, + 289 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 527, + 289 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 527, + 289 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 527, + 289 + ], + "type": "image", + "image_path": "05f48d46008de69a045f19d16947f7fbf8f250b695297fdf44e5dd76ffbcc0fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed meta-causal learning method. (1) A data transformation module simulates the domain shift via generating an auxiliary domain " + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_a" + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "text", + "content": ". (2) A counterfactual inference module analyzes the domain shift by inferring the causal effects of " + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "text", + "content": ". “-” denotes comparing the values of " + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "text", + "content": " before and after do-operation. The edge " + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "inline_equation", + "content": "X \\rightarrow Y" + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "text", + "content": " is a dashed arrow to represent the classification model " + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "inline_equation", + "content": "P(Y|X)" + }, + { + "bbox": [ + 46, + 296, + 547, + 352 + ], + "type": "text", + "content": ". (3) A factor-aware domain alignment module reduces the domain shift via multiple feature mappings according to weights learned by an effect-to-weight network. (4) A base module is used for feature extraction and classification." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 372, + 288, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 288, + 421 + ], + "type": "text", + "content": "lizing causal inference to discover the intrinsic causes of domain shift by constructing counterfactual scenes over the learnt causal graph, which facilitates the reduction of domain shift and in turn benefits cross-domain adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 432, + 103, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 103, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 103, + 444 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 453, + 159, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 159, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 159, + 464 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": "For single domain generalization, during training, we are given a labeled source domain " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = \\{(x_i^s,y_i^s)|_{i = 1}^{N_s}\\}" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": " drawn from the distribution " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^s" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": "-th source sample with its category label " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "y_{i}^{s}\\in \\mathcal{V}" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": ". During testing, the learnt model is applied to multiple unseen target domains " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{\\mathcal{D}_t^j\\}_{j = 1}^J" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t^j" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": "-th target domain drawn from the distribution " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "P_{t}^{j}" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "inline_equation", + "content": "P_{t}^{j}\\neq P_{s}" + }, + { + "bbox": [ + 47, + 472, + 287, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "To bridge the domain gap between the source domain and multiple unseen target domains, we propose a new learning paradigm, called simulate-analyze-reduce, which starts with simulating the domain shift between the source and target domains, then learns to analyze the domain shift, and finally learns to reduce the domain shift. Under this paradigm, we propose a meta-causal learning method that has four components: a data transformation module to generate auxiliary domains as target domains, a counterfactual inference module to discover the causes of the domain shift, a factor-aware domain alignment module to reduce the domain shift, and a base module for feature extraction and classification, as illustrated in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 371, + 545, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 545, + 396 + ], + "type": "text", + "content": "3.2. Data Transformation for Domain Shift Simulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 546, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 557 + ], + "type": "text", + "content": "To simulate the real domain shift between the source and target domains, we generate an auxiliary domain " + }, + { + "bbox": [ + 304, + 403, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_a" + }, + { + "bbox": [ + 304, + 403, + 546, + 557 + ], + "type": "text", + "content": " as the unseen target domain by performing transformations on source data. The domain shift is actually the data distribution discrepancy between the source and auxiliary domains, and is usually caused by the variations of extrinsic attributes of data, independent of intrinsic semantics. Taking the image data for example, the domain shift is mainly caused by the variations of visual attributes, such as the variations of brightness, viewpoint, and color. Therefore, to make the simulated domain shift as realistic as possible, we formulate the extrinsic attributes as variant factors, and design data transformations according to the variant factors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": "We define a set of variant factors, denoted as " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_{1}, v_{2}, \\dots, v_{K}\\}" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": "-th variant factor. Each variant factor corresponds to a data transformation function that aims to generate new data by making changes on the corresponding extrinsic attribute, denoted as " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "G_{v_{k}}(\\boldsymbol{x}; \\theta_{v_{k}})" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": " is an input sample, and " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\theta_{v_{k}} \\in [g_{min}^{k}, g_{max}^{k}]" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": " represents the degree parameter to control the magnitude of transformation with the scale range " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "[g_{min}^{k}, g_{max}^{k}]" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": "In real-world scenarios, the complex domain shift is often caused by combinatorial multiple variant factors and accordingly we design a sampling strategy to enable the combinatorial data transformation. Given a source sample " + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^s" + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7685" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": "we randomly sample " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "N_{v}^{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": " variant factors from " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": " to form a factor subset " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i = \\{v_1^i,v_2^i,\\dots ,v_{N_v^i}^i\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": ". Then, a corresponding auxiliary sample " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": " is generated by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 148 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {i} ^ {a} = G _ {v _ {N _ {v} ^ {i}} ^ {i}} \\left(\\dots G _ {v _ {2} ^ {i}} \\left(G _ {v _ {1} ^ {i}} \\left(\\boldsymbol {x} _ {i} ^ {s}; \\theta_ {v _ {1} ^ {i}}\\right); \\theta_ {v _ {2} ^ {i}}\\right) \\dots ; \\theta_ {v _ {N _ {v} ^ {i}} ^ {i}}\\right), \\tag {1}", + "image_path": "ef6f17b70ab5b594205bc952e734570619f77031edf69a2aa75f2621ca36133c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 149, + 287, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 149, + 287, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 287, + 197 + ], + "type": "text", + "content": "where the transformation degree parameter " + }, + { + "bbox": [ + 47, + 149, + 287, + 197 + ], + "type": "inline_equation", + "content": "\\theta_{v_k^i}" + }, + { + "bbox": [ + 47, + 149, + 287, + 197 + ], + "type": "text", + "content": " is randomly selected from its range scale. In this way, the domain shift is generated as diverse as possible to approximate the real domain shift." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 204, + 287, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 287, + 229 + ], + "type": "text", + "content": "3.3. Counterfactual Inference for Domain Shift Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 235, + 287, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 287, + 306 + ], + "type": "text", + "content": "After simulating the domain shift, we introduce counterfactual inference to analyze the domain shift. During training, we learn the meta-knowledge about how to infer the causes of data discrepancy between one auxiliary sample and the source domain, and during testing, we apply the learnt meta-knowledge to unseen target samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": "We build a causal graph to model the causal dependency among the input sample (node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": "), the variant factors (node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": "), the semantic concepts (node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": "), and the output category (node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": "), as shown in Figure 2 (2). The semantic concepts denote the intrinsic attributes of data that are related to the category, and the variant factors denote the extrinsic attributes of data that are domain specific, independent of the intrinsic semantics. For example, when the input sample is an image of \"zebra\", the semantic concepts are like \"four legs\" and \"black-white stripes\", and the variant factors include brightness, viewpoint, and so on. The edge " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "S \\rightarrow Y" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " represents that the category of the input data is determined by the semantic concepts. The edges " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "S \\rightarrow X" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "V \\rightarrow X" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " represent that the semantic concepts and the variant factors together determine what the input sample looks like. Since the semantic concepts represent the intrinsic semantics and are invariant across different domains, ideally the cross-domain classification is only determined by the semantic concepts, which is implemented by learning the edge " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "S \\rightarrow Y" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": ", i.e., estimating the conditional probability " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "P(Y|S)" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": ". However, in reality, the semantic concepts are unobserved from the input samples, so the classification should be implemented by learning the edge " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "X \\rightarrow Y" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": ", i.e., estimating the conditional probability " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "P(Y|X)" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": ". The edge " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "X \\rightarrow Y" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " is a dashed arrow to represent the classification model " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "P(Y|X)" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": ". Since the input sample node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " is affected by the semantic concept node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " and the variant factor node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " together, the category node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " is also affected by the variant factor node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " through edges " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "V \\rightarrow X \\rightarrow Y" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": ". As the variant factors are domain-specific, their causal effects on the category lead to the domain shift. Hence, we infer the causal effects of the variant factors on the category prediction to discover the causes of the domain shift. That is to say, we infer the causal effects of the node " + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 307, + 287, + 713 + ], + "type": "text", + "content": " on the node" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " as causes of domain shift, and then learn to reduce the domain shift based on its causes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "text", + "content": "To infer the causal effects of the node " + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "text", + "content": " on the node " + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "text", + "content": ", we first learn the edge " + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "inline_equation", + "content": "X \\to Y" + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "text", + "content": " using the source data " + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 305, + 95, + 545, + 131 + ], + "type": "text", + "content": " by a classification loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 137, + 545, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 137, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 137, + 545, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\sim P _ {s}} \\left[ - \\sum_ {u} \\mathbb {I} _ {u = y _ {i} ^ {s}} \\log C \\left(F \\left(\\boldsymbol {x} _ {i} ^ {s}\\right)\\right) _ {u} \\right], \\tag {2}", + "image_path": "1922ec960be195ca96b1f1dc250c4fd97afe8ba52604260fb621c4f8b56a8768.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_i^s, y_i^s) \\in \\mathcal{D}_s" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": " is a feature extractor, " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": " is the classifier to output " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}|" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": " category probabilities, " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "C(\\cdot)_u" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}|" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": " category probabilities, and " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{u=y_i^s}" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": " is an indicator function, meaning that if " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "u = y_i^s" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": ", the value of " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{u=y_i^s}" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": " is 1 and 0 otherwise. Then, for each sample from the auxiliary domain, we construct a factual scene and multiple counterfactual scenes over the causal graph, so as to infer the causal effects of variant factors on the category prediction. Given an auxiliary sample " + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a \\in \\mathcal{D}_a" + }, + { + "bbox": [ + 305, + 178, + 545, + 297 + ], + "type": "text", + "content": ", its factual category is predicted by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 307, + 545, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 307, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 365, + 307, + 545, + 320 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {i} ^ {a} = P (Y | X) = C \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right), \\tag {3}", + "image_path": "41453699ee98972ba373273560d86c500da4c53cb42c534e38b9bb68c715db2a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^a \\in \\mathbb{R}^{|\\mathcal{Y}|}" + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "content": " is a category probability vector, and represents the value of node " + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "content": ". For each variant factor, we construct a counterfactual scene to infer its causal effect by doing intervention on the variant factor node " + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "inline_equation", + "content": "do(V = v_k)" + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "content": " denote the intervention on the node " + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "content": ", which is implemented by changing extrinsic attributes of data through the transformation " + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "inline_equation", + "content": "G_{v_k}" + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "content": " with multiple degree parameters. Accordingly, the counterfactual category of " + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a" + }, + { + "bbox": [ + 305, + 327, + 545, + 436 + ], + "type": "text", + "content": " is predicted by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 444, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 444, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 335, + 444, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {y} _ {i, v _ {k}} ^ {a} = P \\left(Y | X, d o (V = v _ {k})\\right) \\\\ = \\frac {1}{| \\mathcal {M} |} \\sum_ {\\theta_ {v _ {k}} \\in \\mathcal {M}} C \\left(F \\left(G _ {v _ {k}} \\left(\\boldsymbol {x} _ {i} ^ {a}; \\theta_ {v _ {k}}\\right)\\right)\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "d9351e6ed5306f2fcfb47fcf68b4a3aa410d675f741cf7bda5fd79c8d3d883f1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "text", + "content": " is a set of degree parameters, obtained by uniformly sampling magnitudes of transformation " + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "inline_equation", + "content": "G_{v_k}" + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "text", + "content": " from the scale range. By comparing the factual and counterfactual category probabilities of " + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a" + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "text", + "content": ", the causal effect of the " + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 498, + 545, + 559 + ], + "type": "text", + "content": "-th variant factor is calculated by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 567, + 545, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 545, + 591 + ], + "type": "interline_equation", + "content": "\\boldsymbol {e} _ {k} ^ {i} = P (Y | X) - P \\left(Y | X, d o (V = v _ {k})\\right) = \\boldsymbol {y} _ {i} ^ {a} - \\boldsymbol {y} _ {i, v _ {k}} ^ {a}. \\tag {5}", + "image_path": "8253e6c3a955e3f537dc83b5f8a461f066b3730767239d3ff742c09e8a31f3bc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 592, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 592, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 592, + 545, + 639 + ], + "type": "text", + "content": "The inferred causal effect represents how much contribution the corresponding variant factor makes to the domain shift, and the larger the causal effect is, the more seriously the domain shift is caused by the variant factor." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 647, + 545, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 545, + 670 + ], + "type": "text", + "content": "3.4. Factor-aware Domain Alignment for Domain Shift Reduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "After analyzing the causes of domain shift, we propose a factor-aware domain alignment to reduce the domain shift by learning multiple feature mappings, with guidance of the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7686" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 534, + 196 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 534, + 196 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 534, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 534, + 196 + ], + "type": "image", + "image_path": "8b18c8e4206dffb7d06e3e53f5b175ba03594c1f3d3fe28e73191aa432b087a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 202, + 437, + 215 + ], + "lines": [ + { + "bbox": [ + 155, + 202, + 437, + 215 + ], + "spans": [ + { + "bbox": [ + 155, + 202, + 437, + 215 + ], + "type": "text", + "content": "Figure 3. Inference process of meta-causal learning for a given target sample." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "text", + "content": "inferred causal effects of variant factors. Each feature mapping addresses a specific domain shift caused by one variant factor. We construct " + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "text", + "content": " feature mappings for " + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "text", + "content": " variant factors, and the " + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "text", + "content": "-th feature mapping aims to address the domain shift caused by the " + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 236, + 287, + 342 + ], + "type": "text", + "content": "-th variant factor. In order to incorporate the causal effects of variant factors into the learning of mappings, we build an effect-to-weight network that converts the causal effect of each variant factor into the weight of the corresponding feature mapping." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 342, + 288, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 288, + 379 + ], + "type": "text", + "content": "For the auxiliary sample " + }, + { + "bbox": [ + 47, + 342, + 288, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a" + }, + { + "bbox": [ + 47, + 342, + 288, + 379 + ], + "type": "text", + "content": " and its inferred causal effects of all variant factors " + }, + { + "bbox": [ + 47, + 342, + 288, + 379 + ], + "type": "inline_equation", + "content": "\\{e_1^i, e_2^i, \\dots, e_K^i\\}" + }, + { + "bbox": [ + 47, + 342, + 288, + 379 + ], + "type": "text", + "content": ", the weights of feature mappings are calculated by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 384, + 287, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 384, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 71, + 384, + 287, + 400 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\omega} ^ {i} = \\operatorname {s o f t m a x} \\big (W (\\boldsymbol {e} _ {1} ^ {i}), W (\\boldsymbol {e} _ {2} ^ {i}), \\dots , W (\\boldsymbol {e} _ {K} ^ {i}) \\big), \\quad (6)", + "image_path": "b8663a707beb8e5280569b625cd4e77746ab0601fd46c4715d4ee5c9f5a23b4c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\omega^i\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "text", + "content": " and its " + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "text", + "content": " -th element " + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\omega_{k}^{i}" + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "text", + "content": " denotes the weight of the " + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "text", + "content": " -th factor-aware feature mapping. " + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "inline_equation", + "content": "W(\\cdot)" + }, + { + "bbox": [ + 46, + 404, + 287, + 441 + ], + "type": "text", + "content": " denotes the effect-to-weight network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "content": "According to the weights of feature mappings, we integrate the " + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "content": " feature mappings to project the auxiliary samples into the source feature space. The alignment of the source and auxiliary domains is implemented by minimizing the feature distance between the source and auxiliary samples in the source feature space. Given a source sample " + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^s" + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "content": " with the corresponding category label " + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "inline_equation", + "content": "y_i^s" + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "content": ", generated auxiliary sample " + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a" + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "content": " and inferred mapping weights " + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\omega^i" + }, + { + "bbox": [ + 46, + 441, + 288, + 559 + ], + "type": "text", + "content": ", the alignment loss incorporated with inferred causal effects is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 564, + 287, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 564, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 61, + 564, + 287, + 623 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {a} ^ {c} = \\frac {1}{N _ {s}} \\sum_ {i} | | F \\left(\\boldsymbol {x} _ {i} ^ {s}\\right) - \\sum_ {k} \\omega_ {k} ^ {i} M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right) | | _ {2} \\tag {7} \\\\ + \\frac {1}{N _ {s}} \\sum_ {i} \\mathcal {H} \\left(C \\left(\\sum_ {k} \\omega_ {k} ^ {i} M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {a}\\right)\\right), y _ {i} ^ {s}\\right)\\right). \\\\ \\end{array}", + "image_path": "df86bd1cba52725b68a30d0c7ac444ef61579e6cc0c7349717c2d0c29631e733.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is the feature extractor, " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "M_{k}" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "-th feature mapping, " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is the classifier, " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\omega_{k}^{i}" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\omega^i" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sum_{k}\\omega_{k}^{i}M_{k}\\big(F(\\pmb{x}_{i}^{a})\\big)" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " represents the projected feature of " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is the number of source samples. " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "||\\cdot ||_2" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is the L2-loss, and " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is the cross-entropy loss. The first term measures the feature distance between the source and auxiliary samples after projection, i.e., the data distribution discrepancy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 236, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 270 + ], + "type": "text", + "content": "between the source and auxiliary domains. The second term encourages the auxiliary samples to belong to the same categories as the source samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "text", + "content": "Moreover, to enable each feature mapping to address the specific domain shift caused by the corresponding variant factor, we introduce another alignment loss " + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_a^m" + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 315, + 545, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 315, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 319, + 315, + 545, + 375 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {a} ^ {m} = \\frac {1}{N _ {s}} \\frac {1}{K} \\sum_ {i} \\sum_ {k} | | F (\\boldsymbol {x} _ {i} ^ {s}) - M _ {k} (F (\\boldsymbol {x} _ {i} ^ {k})) | | _ {2} \\\\ + \\frac {1}{N _ {s}} \\frac {1}{K} \\sum_ {i} \\sum_ {k} \\mathcal {H} \\left(C \\left(M _ {k} \\left(F \\left(\\boldsymbol {x} _ {i} ^ {k}\\right)\\right), y _ {i} ^ {s}\\right)\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "440ebacab7051d389eed9bf03a4fcec8d88c4d645087c09679b216b22623da4a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^k" + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "text", + "content": " is an sample generated by conducting the data transformation " + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "inline_equation", + "content": "G_{v_k}" + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "text", + "content": " on the source sample " + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^s" + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^k = G_{v_k}(\\pmb{x}_i^s;\\theta_{v_k})" + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "text", + "content": ". The transformation degree parameter " + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\theta_{v_k}" + }, + { + "bbox": [ + 304, + 384, + 545, + 434 + ], + "type": "text", + "content": " is randomly selected from its range scale." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 434, + 492, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 492, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 492, + 444 + ], + "type": "text", + "content": "Then the overall loss function is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 382, + 454, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 454, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 382, + 454, + 545, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {a} ^ {c} + \\mathcal {L} _ {a} ^ {m}. \\tag {9}", + "image_path": "45d9c67d198310f842e5f5f1858001921ad536a167e6fdfb82acc74e9c39f791.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 476, + 541, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 541, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 541, + 489 + ], + "type": "text", + "content": "The whole training process is summarized in Algorithm 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 496, + 373, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 373, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 373, + 507 + ], + "type": "text", + "content": "3.5. Inference" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": "During testing, given a target sample " + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^t" + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": ", firstly, the causal effects of variant factors are inferred by counterfactual inference in Eq. (5), and the weights " + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\omega^i" + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": " of feature mappings are calculated by the effect-to-weight network in Eq. (6). Then the target feature " + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f}_i^t = F(\\boldsymbol{x}_i^t)" + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": " is projected into the source feature space by integrating " + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": " feature mappings according to their weights " + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\omega^i" + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": " to obtain the projected feature " + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{f}}_i^t = \\sum_k \\omega_k^i M_k(\\boldsymbol{f}_i^t)" + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": ". Finally, the category of " + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^t" + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": " is predicted by " + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "inline_equation", + "content": "C(\\hat{\\boldsymbol{f}}_i^t)" + }, + { + "bbox": [ + 304, + 514, + 545, + 640 + ], + "type": "text", + "content": ". The whole inference process is shown in Figure 3." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 651, + 386, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 386, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 386, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 671, + 368, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 368, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 368, + 682 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Digits. The Digits dataset consists of five datasets: MNIST [16], MNIST-M [9], SVHN [26], USPS [13], and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7687" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 171, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 171, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 171, + 83 + ], + "type": "text", + "content": "Algorithm 1 Training process" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 87, + 271, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 87, + 271, + 99 + ], + "spans": [ + { + "bbox": [ + 49, + 87, + 271, + 99 + ], + "type": "text", + "content": "Input: The source domain " + }, + { + "bbox": [ + 49, + 87, + 271, + 99 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 49, + 87, + 271, + 99 + ], + "type": "text", + "content": ", the variant factor set " + }, + { + "bbox": [ + 49, + 87, + 271, + 99 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 49, + 87, + 271, + 99 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "text", + "content": "Output: The feature extractor " + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "text", + "content": ", the classifier " + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "text", + "content": " feature mappings " + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\{M_i\\}_{i=1}^K" + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "text", + "content": ", the effect-to-weight network " + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 49, + 100, + 287, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 135, + 287, + 292 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 54, + 135, + 186, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 135, + 186, + 148 + ], + "spans": [ + { + "bbox": [ + 54, + 135, + 186, + 148 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 54, + 135, + 186, + 148 + ], + "type": "inline_equation", + "content": "F, C, \\{M_i\\}_{i=1}^K, W" + }, + { + "bbox": [ + 54, + 135, + 186, + 148 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 148, + 156, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 148, + 156, + 159 + ], + "spans": [ + { + "bbox": [ + 54, + 148, + 156, + 159 + ], + "type": "text", + "content": "2: while not converge do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 159, + 198, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 198, + 171 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 198, + 171 + ], + "type": "text", + "content": "3: Sample an image " + }, + { + "bbox": [ + 54, + 159, + 198, + 171 + ], + "type": "inline_equation", + "content": "\\pmb{x}_j^s" + }, + { + "bbox": [ + 54, + 159, + 198, + 171 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 159, + 198, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 54, + 159, + 198, + 171 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "type": "text", + "content": "4: Construct a factor subset " + }, + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "type": "text", + "content": " by sampling " + }, + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "type": "inline_equation", + "content": "N_v^i" + }, + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "type": "text", + "content": " variant factors from " + }, + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 171, + 287, + 194 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 195, + 250, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 195, + 250, + 207 + ], + "spans": [ + { + "bbox": [ + 54, + 195, + 250, + 207 + ], + "type": "text", + "content": "5: Generate an auxiliary sample " + }, + { + "bbox": [ + 54, + 195, + 250, + 207 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a" + }, + { + "bbox": [ + 54, + 195, + 250, + 207 + ], + "type": "text", + "content": " by Eq.(1);" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 208, + 135, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 208, + 135, + 218 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 135, + 218 + ], + "type": "text", + "content": "6: for " + }, + { + "bbox": [ + 54, + 208, + 135, + 218 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 54, + 208, + 135, + 218 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 208, + 135, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 208, + 135, + 218 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 219, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 219, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 219, + 287, + 243 + ], + "type": "text", + "content": "7: Calculate the factual category " + }, + { + "bbox": [ + 54, + 219, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^a" + }, + { + "bbox": [ + 54, + 219, + 287, + 243 + ], + "type": "text", + "content": " of auxiliary sample " + }, + { + "bbox": [ + 54, + 219, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^a" + }, + { + "bbox": [ + 54, + 219, + 287, + 243 + ], + "type": "text", + "content": " by Eq.(3);" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 243, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 286, + 266 + ], + "type": "text", + "content": "8: Calculate the counterfactual category " + }, + { + "bbox": [ + 54, + 243, + 286, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{i,v_k}^a" + }, + { + "bbox": [ + 54, + 243, + 286, + 266 + ], + "type": "text", + "content": " of auxiliary sample " + }, + { + "bbox": [ + 54, + 243, + 286, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^a" + }, + { + "bbox": [ + 54, + 243, + 286, + 266 + ], + "type": "text", + "content": " by Eq.(4);" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "text", + "content": "9: Infer the causal effect " + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "inline_equation", + "content": "e_k^i" + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "text", + "content": " of variant factor " + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "inline_equation", + "content": "v_k" + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "text", + "content": " via comparing " + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "inline_equation", + "content": "y_i^a" + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "inline_equation", + "content": "y_{i,v_k}^a" + }, + { + "bbox": [ + 54, + 267, + 286, + 292 + ], + "type": "text", + "content": " by Eq.(5);" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 51, + 293, + 287, + 349 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 51, + 293, + 107, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 293, + 107, + 301 + ], + "spans": [ + { + "bbox": [ + 51, + 293, + 107, + 301 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 51, + 303, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 303, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 51, + 303, + 287, + 327 + ], + "type": "text", + "content": "11: Calculate the weights of " + }, + { + "bbox": [ + 51, + 303, + 287, + 327 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 51, + 303, + 287, + 327 + ], + "type": "text", + "content": " feature mappings " + }, + { + "bbox": [ + 51, + 303, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\{M_i\\}_{i=1}^K" + }, + { + "bbox": [ + 51, + 303, + 287, + 327 + ], + "type": "text", + "content": " by Eq.(6);" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 51, + 327, + 228, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 327, + 228, + 339 + ], + "spans": [ + { + "bbox": [ + 51, + 327, + 228, + 339 + ], + "type": "text", + "content": "12: Update " + }, + { + "bbox": [ + 51, + 327, + 228, + 339 + ], + "type": "inline_equation", + "content": "F, C, \\{M_i\\}_{i=1}^K" + }, + { + "bbox": [ + 51, + 327, + 228, + 339 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 51, + 327, + 228, + 339 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 51, + 327, + 228, + 339 + ], + "type": "text", + "content": " by Eq.(9)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 51, + 339, + 107, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 339, + 107, + 349 + ], + "spans": [ + { + "bbox": [ + 51, + 339, + 107, + 349 + ], + "type": "text", + "content": "13: end while" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 373, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 422 + ], + "type": "text", + "content": "SYN [9], with 10 categories. We use MNIST as the source domain, and the other four datasets as the target domains. The first 10,000 images in the training set of MNIST are used for training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 422, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 517 + ], + "type": "text", + "content": "CIFAR10-C. The CIFAR10-C dataset [11] is proposed to evaluate the robustness of classification model. The images are corrupted from the test set of the CIFAR10 dataset [15] by 19 corruption types with five levels of severity. A higher level means the more serious corruption. There are 10 categories. We use CIFAR10 as the source domain, and CIFAR10-C as the target domains where images of one severity level form one target domain." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 518, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 590 + ], + "type": "text", + "content": "PACS. The PACS dataset [17] is a benchmark for domain generalization, and consists of four domains: art painting, cartoon, photo, and sketch. There are 9,991 images of seven categories. We use one domain as the source domain, and the rest three domains as the target domains. So there are four tasks with different domains as the source domains." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 598, + 179, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 179, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 179, + 611 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Auxiliary Domain. We define 16 variant factors to generate the images in the auxiliary domain, including 12 photometric factors (Brightness, Contrast, Color, Sharpness, Auto-Contrast, Invert, Equalize, Solarize, SolarizeAdd, Posterize, NoiseSalt, NoiseGaussian) and 4 geometric factors (Shear-X, Shear-Y, Rotate, Flip). Since the Rotate and Flip variant factors will affect the semantic information of digit images, we use the other 14 variant factors for the Digits dataset. For" + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 307, + 112, + 550, + 274 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 103 + ], + "type": "text", + "content": "Table 1. Single domain generalization results (%) on Digits with ConvNet as backbone. The model is trained on MNIST, and evaluated on SVHN, SYN, MNIST-M, and USPS." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 112, + 550, + 274 + ], + "lines": [ + { + "bbox": [ + 307, + 112, + 550, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 112, + 550, + 274 + ], + "type": "table", + "html": "
MethodSVHNSYNMNIST-MUSPSAvg
ERM [14]27.8339.6552.7276.9449.29
CCSA [24]25.8937.3149.2983.7249.05
d-SNE [36]26.2237.8350.9893.1652.05
JiGen [2]33.8043.7957.8077.1553.14
GUD [31]35.5145.3260.4177.2654.62
M-ADA [28]42.5548.9567.9478.5359.49
ME-ADA [37]42.5650.3963.2781.0459.32
PDEN [19]62.2169.3982.2085.2674.77
L2D [34]62.8663.7287.3083.9774.46
AA [4]45.2364.5260.5380.6262.72
RA [5]54.7759.6074.0577.3366.44
RSDA [30]47.4062.0081.5083.1068.50
RSDA+ASR [8]52.8064.5080.8082.4070.10
Ours69.9478.4778.3488.5478.82
", + "image_path": "c83627f752f99db53bae3dc99f6b0875998600831aec3b49bebdbe930782e82f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 307, + 338, + 550, + 426 + ], + "blocks": [ + { + "bbox": [ + 305, + 285, + 545, + 330 + ], + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 330 + ], + "type": "text", + "content": "Table 2. Single domain generalization results (\\%) on CIFAR10-C with WRN as backbone. Each level is viewed as a target domain, a higher level denotes the more serious corruption and the domain discrepancy between the source and target domains is larger." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 338, + 550, + 426 + ], + "lines": [ + { + "bbox": [ + 307, + 338, + 550, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 550, + 426 + ], + "type": "table", + "html": "
Methodlevel1level2level3level4level5Avg
ERM [14]87.8081.5075.5068.2056.1073.82
GUD [31]88.3083.5077.6070.6058.3075.66
M-ADA [28]90.5086.8082.5076.4065.6080.36
PDEN [19]90.6288.9187.0383.7177.4785.55
AA [4]91.4287.8884.1078.4671.1382.60
RA [5]91.7488.8985.8281.0374.9384.48
Ours92.3891.2289.8887.7384.5289.15
", + "image_path": "efad143171c7e70ca496d5743e8b99903c56e24b194f945b116f2ba27eb1fe0e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "type": "text", + "content": "the CIFAR10-C and PACS datasets, all 16 variant factors are used. To make the auxiliary domain as diverse as possible, for each source image at each iteration, we randomly sample several variant factors and use them to generate a new auxiliary image." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 514, + 520, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 520, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 520, + 528 + ], + "type": "text", + "content": "4.3. Results on Single Domain Generalization" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": "We compare our method with several state-of-the-art methods, including the baseline method (ERM [14]), the methods of learning domain-invariant features (CCSA [24], d-SNE [36], JiGen [2]), and the methods of making data augmentation (GUD [31], M-ADA [28], ME-ADA [37], PDEN [19], L2D [34], AA [4], RA [5], RSDA [30], RSC [12], ASR [8])." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "Table 1, Table 2, and Table 3 show the comparison results on Digits, CIFAR10-C, and PACS, respectively. From the results, there are several interesting observations as follows. First, our method generally outperforms the compared methods on all datasets, which clearly shows the effectiveness of the proposed simulate-analyze-reduce learning paradigm for single domain generalization. Second, compared with the data augmentation methods (e.g.," + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7688" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 123, + 288, + 181 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 3. Single domain generalization results (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as the target domains." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 123, + 288, + 181 + ], + "lines": [ + { + "bbox": [ + 48, + 123, + 288, + 181 + ], + "spans": [ + { + "bbox": [ + 48, + 123, + 288, + 181 + ], + "type": "table", + "html": "
MethodArtpaintCartoonSketchPhotoAvg
ERM [14]70.9076.5053.1042.2060.70
RSC [12]73.4075.9056.2041.6061.80
RSC+ASR [8]76.7079.3061.6054.6068.10
Ours77.1380.1462.5559.6069.86
", + "image_path": "02733cade7e4371e1a4a7ca6b936e0596636556edc641ac7d6c643eb4c068b1d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 246, + 291, + 452 + ], + "blocks": [ + { + "bbox": [ + 47, + 193, + 287, + 236 + ], + "lines": [ + { + "bbox": [ + 47, + 193, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 287, + 236 + ], + "type": "text", + "content": "Table 4. Leave-one-domain-out results (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the target domain and the other three domains are used as source domains." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 246, + 291, + 452 + ], + "lines": [ + { + "bbox": [ + 50, + 246, + 291, + 452 + ], + "spans": [ + { + "bbox": [ + 50, + 246, + 291, + 452 + ], + "type": "table", + "html": "
MethodArtpaintCartoonPhotoSketchAvg
MetaReg [1]83.7077.2095.5070.3081.70
GUD [31]78.3277.6595.6174.2181.44
Epi-FCR [18]82.1077.0093.9073.0081.50
MASF [6]80.2977.1794.9971.6881.03
JiGen [2]79.4275.2596.0371.3580.51
DMG [3]76.9080.3893.5575.2181.46
DDAIG [38]84.2078.1095.3074.7083.10
CSD [27]78.9075.8094.1076.7081.40
L2A-OT [39]83.3078.2096.2073.6082.80
EISNet [33]81.8976.4495.9374.3382.15
RSC [12]83.4380.3195.9980.8585.15
ME-ADA [37]78.6178.6595.5775.5982.10
MMLD [23]81.2877.1696.0972.2981.83
L2D [34]81.4479.5695.5180.5884.27
FACT [35]85.3778.3895.1579.1584.51
MatchDG [22]81.3280.7096.5379.7284.57
CIRL [21]86.0880.5995.9382.6786.32
Ours85.3080.9396.5385.2487.00
", + "image_path": "26d50dcfce57670383e1b8e2156f3e17e8ea3aef3562d0c6c31fd2a773a257f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 529, + 291, + 583 + ], + "blocks": [ + { + "bbox": [ + 46, + 464, + 288, + 521 + ], + "lines": [ + { + "bbox": [ + 46, + 464, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 288, + 521 + ], + "type": "text", + "content": "Table 5. Ablation study (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as target domains. “T”, “A”, “C” denote Domain Transformation, Domain Alignment, and Counterfactual Inference, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 529, + 291, + 583 + ], + "lines": [ + { + "bbox": [ + 50, + 529, + 291, + 583 + ], + "spans": [ + { + "bbox": [ + 50, + 529, + 291, + 583 + ], + "type": "table", + "html": "
MethodTACArtpaintCartoonSketchPhotoAvg
Base71.2667.6443.9736.9954.97
DT✓75.2878.4659.4556.0967.32
DTA✓✓71.6472.7857.1152.0263.39
Ours✓✓✓77.1380.1462.5559.6069.86
", + "image_path": "8abe1ca85421b46cbd10d92b95c93df0c5d1397d88fdf90b5bd05a4b35f8c03e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "AA [4], RA [5] and RSDA [30]) that are more related to our method, our method achieves much better results, and especially yields a " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "10.32\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " gain over RSDA on Digits, strongly suggesting that it is beneficial to empower the model with the ability of analyzing the causes of domain shift by counterfactual inference. Third, on more difficult tasks with larger domain shift (e.g., SVHN on Digits, level5 on CIFAR10-C, and Photo on PACS), our method significantly improves the performance, further demonstrating the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "superiority of our method on handling more challenging situations. Forth, our method performs little worse on MNIST-M and USPS of Digits. The possible reason is that other compared methods use more well-designed data augmentation and network regularization, such as AdaIN based generators in PDEN [19] and stochastic neighborhood embedding techniques in d-SNE [36]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 164, + 533, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 164, + 533, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 533, + 177 + ], + "type": "text", + "content": "4.4. Results on Multiple Domain Generalization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 183, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 327 + ], + "type": "text", + "content": "We extend the proposed method to multi-source domain setting by regarding the multiple source domains as one source domain without using domain labels. We employ the leave-one-domain-out protocol following existing multi-source domain generalization [21, 35]. We compare our method with most related methods that introduces causal inference into domain generalization (MatchDG [22], CIRL [21]), and existing popular domain generalization methods (MetaReg [1], GUD [31], EpiFCR [18], MASF [6], JiGen [2], DMG [3], DDAIG [38], CSD [27], L2A-OT [39], EISNet [33], RSC [12], ME-ADA [37], MMLD [23], L2D [34], FACT [35])." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "text", + "content": "Table 4 shows the leave-one-domain-out results on the PACS dataset with ResNet-18 as backbone. From the results, we make several observations. First, it is noteworthy that our method achieves the state-of-the-art overall performance (\"Avg\") although our method is not designed for multi-source domain generalization, which further demonstrates that the proposed simulate-analyze-reduce learning paradigm not only benefits single domain generalization but also boosts multi-source domain generalization. Second, compared with the methods of introducing causal inference to learn domain-invariant features (MatchDG [22], CIRL [21]), our method achieves better results on the overall metric \"Avg\", especially making " + }, + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "inline_equation", + "content": "5.52\\%" + }, + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "text", + "content": " gains over MatchDG [22] and CIRL [21], respectively, on the more challenging task (Sketch " + }, + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "text", + "content": " Others). Such improvements are attributed to the ability of analyzing and reducing the domain shift, further demonstrating the advantages of causal inference in analyzing the causes of the domain shift." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "To evaluate each component of our method, we conduct ablation experiments on the PACS dataset. We design several degraded variants of our method for comparison: (1) \"Base\", where only the base module is utilized and optimized by Eq. (2) using the source domain; (2) \"DT\", where the data transformation module is added into the base module and the model is trained using both source and auxiliary domains; (3) \"DTA\", where the simulated domain shift between auxiliary domain and source domain is directly reduced without analyzing the causes of the domain shift via the proposed counterfactual module." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7689" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 520, + 182 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 520, + 182 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 520, + 182 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 520, + 182 + ], + "type": "image", + "image_path": "fa0dca8fe5dba2c2c953509f91624b91296fab9d7e5aef196994c4972cabb1ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "type": "text", + "content": "Figure 4. Examples of the inferred causal effects (represented as weights) of variant factors. The left part shows a source image from the CIFAR10 dataset and three target images from the CIFAR10-C dataset with Gaussian noise corruption. As the corruption severity increases from level 1 to level 5, the inferred weights of the NoiseGaussian variant factor become larger accordingly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 287, + 291, + 338 + ], + "blocks": [ + { + "bbox": [ + 46, + 245, + 288, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 245, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 288, + 280 + ], + "type": "text", + "content": "Table 6. Factor analysis (\\%) on PACS with ResNet-18 as backbone. One domain (name in column) is used as the source domain and the other three domains are used as target domains." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 287, + 291, + 338 + ], + "lines": [ + { + "bbox": [ + 50, + 287, + 291, + 338 + ], + "spans": [ + { + "bbox": [ + 50, + 287, + 291, + 338 + ], + "type": "table", + "html": "
MethodArtpaintCartoonSketchPhotoAvg
Ours_GT74.6269.7852.7443.3460.12
Ours_PT74.4773.4850.3551.6762.49
Ours77.1380.1462.5559.6069.86
", + "image_path": "8ab6985fa502123931c98deec2109422f8262bb9c32c363713f179215b55f0ba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 288, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 493 + ], + "type": "text", + "content": "The results are shown in Table 5. From the results, we make several observations. First, our method achieves better performance than \"DT\", which validates the superiority of our method on analyzing and reducing the domain shift, rather than directly enlarging the source data distribution. Second, \"DTA\" performs worse than \"DT\", probably because brute-force domain alignment without analyzing causes of the domain shift leads to negative transfer and thus hurts the performance. Third, our method achieves the best performance thanks to the proposed simulate-analyze-reduce paradigm." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 502, + 144, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 144, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 144, + 514 + ], + "type": "text", + "content": "4.6. Factor Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "In order to further analyze the effect of variant factors, we design several variants of our method by using different variant factors, including only using geometric factors (\"Ours_GT\"), and only using photometric factors (\"Ours_PT\"). Since the factor number of the two type factors are different, we randomly select 4 photometric factors to keep the same number as geometric factors and repeat experiments 10 times to avoid the effect of sampling. The results are shown in Table 6. From the results, it is noteworthy that \"Ours_PT\" performs better than \"Ours_GT\", especially with the " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "8.33\\%" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " gain when using Photo as the source domain. The reason may be that the domain shift between Photo and the other target domains is mainly caused by photometric factors. Moreover, our method outperforms both \"Ours_PT\" and \"Ours_GT\", showing that both photometric factors and geometric factors are required to simulate the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 247, + 449, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 449, + 259 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 449, + 259 + ], + "type": "text", + "content": "domain shift as diverse as possible." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 271, + 437, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 271, + 437, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 437, + 284 + ], + "type": "text", + "content": "4.7. Causality Visualization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 291, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 547, + 472 + ], + "type": "text", + "content": "In Figure 4, we visualize the inferred causal effects (represented by the weights) of variant factors for the unseen target images on the CIFAR10-C dataset during testing. The target images are actually corrupted from the source images of CIFAR10 by Gaussian noise of five levels. It is interesting to observe that the inferred weights of the NoiseGaussian variant factor are larger than that of the other variant factors, indicating that the counterfactual inference succeeds in discovering the real cause of the domain shift. We can also observe that when the corruption severity of Gaussian noise increases from level 1 to level 5 (i.e., the domain shift is more serious), the inferred weights of the NoiseGaussian variant factor become larger, which further demonstrates that our method measures the magnitude of the domain shift correctly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 487, + 379, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 379, + 499 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 379, + 499 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 508, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 674 + ], + "type": "text", + "content": "We have presented a new paradigm, simulate-analyze-reduce, for single domain generalization. Our paradigm empowers the model with the ability to analyze the domain shift, instead of directly expanding the distribution of the source domain to cover unseen target domains. Under this paradigm, we have presented a meta-causal learning method that can learn meta-knowledge about inferring the causes of domain shift during training, and apply such meta-knowledge to reduce the domain shift for boosting adaptation during testing. Extensive experiments on several benchmark datasets have validated the effectiveness of the new learning paradigm and the advantage of meta-causal learning on analyzing the domain shift for domain generalization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the Natural Science Foundation of China (NSFC) under Grant No 62072041." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7690" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. volume 31, pages 1006-1016, 2018. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 178 + ], + "type": "text", + "content": "[2] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2229-2238, 2019. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 224 + ], + "type": "text", + "content": "[3] Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In European Conference on Computer Vision (ECCV), pages 301-318. Springer, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 287, + 278 + ], + "type": "text", + "content": "[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 113-123, 2019. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 335 + ], + "type": "text", + "content": "[5] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 702-703, 2020. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 378 + ], + "type": "text", + "content": "[6] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. volume 32, pages 6447–6458, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "type": "text", + "content": "[7] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In European Conference on Computer Vision (ECCV), pages 200-216. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "type": "text", + "content": "[8] Xinjie Fan, Qifei Wang, Junjie Ke, Feng Yang, Boqing Gong, and Mingyuan Zhou. Adversarily adaptive normalization for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8208-8217, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "type": "text", + "content": "[9] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), pages 1180-1189. PMLR, 2015. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[10] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2551-2559, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[11] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations (ICLR), 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "text", + "content": "[12] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV), pages 124-140. Springer, 2020. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[13] Jonathan J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 16(5):550-554, 1994. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems: École D'Été de Probabilités de Saint-Flour XXXVIII-2008, volume 2033. Springer Science & Business Media, 2011. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "type": "text", + "content": "[15] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[16] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "text", + "content": "[17] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5542-5550, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 275 + ], + "type": "text", + "content": "[18] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 1446-1455, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "type": "text", + "content": "[19] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "text", + "content": "[20] Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. volume 34, pages 6155-6170, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "type": "text", + "content": "[21] Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8046-8056, 2022. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "type": "text", + "content": "[22] Divyat Mahajan, Shruti Tople, and Amit Sharma. Domain generalization using causal matching. In International Conference on Machine Learning (ICML), pages 7313-7324. PMLR, 2021. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "text", + "content": "[23] Toshihiko Matsuura and Tatsuya Harada. Domain generalization using a mixture of multiple latent domains. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 11749-11756, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "text", + "content": "[24] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5715-5725, 2017, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[25] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning (ICML), pages 10-18. PMLR, 2013. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "type": "text", + "content": "[26] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In International Conference on Machine Learning (ICML), pages 7728-7738. PMLR, 2020. 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7691" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 686 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 172 + ], + "type": "text", + "content": "[29] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. In International Conference on Learning Representations (ICLR), 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 174, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 174, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 286, + 217 + ], + "type": "text", + "content": "[30] Riccardo Volpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transformation sets. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7980-7989, 2019. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 286, + 261 + ], + "type": "text", + "content": "[31] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. volume 31, pages 5339-5349, 2018. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 286, + 306 + ], + "type": "text", + "content": "[32] Haohan Wang, Zexue He, Zachary C. Lipton, and Eric P. Xing. Learning robust representations by projecting superficial statistics out. In International Conference on Learning Representations (ICLR), 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 308, + 286, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 286, + 361 + ], + "type": "text", + "content": "[33] Shujun Wang, Lequan Yu, Caizi Li, Chi-Wing Fu, and Pheng-Ann Heng. Learning from extrinsic and intrinsic supervisions for domain generalization. In European Conference on Computer Vision (ECCV), pages 159–176. Springer, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 364, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 364, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 49, + 364, + 286, + 407 + ], + "type": "text", + "content": "[34] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 834-843, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 409, + 286, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 409, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 49, + 409, + 286, + 462 + ], + "type": "text", + "content": "[35] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383-14392, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 464, + 286, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 464, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 49, + 464, + 286, + 518 + ], + "type": "text", + "content": "[36] Xiang Xu, Xiong Zhou, Ragav Venkatesan, Gurumurthy Swaminathan, and Orchid Majumder. d-sne: Domain adaptation using stochastic neighborhood embedding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2497-2506, 2019. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 520, + 286, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 520, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 49, + 520, + 286, + 585 + ], + "type": "text", + "content": "[37] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 14435–14447, 2020. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 587, + 286, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 587, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 49, + 587, + 286, + 640 + ], + "type": "text", + "content": "[38] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Deep domain-adversarial image generation for domain generalisation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 34, pages 13025-13032, 2020. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 643, + 286, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 643, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 49, + 643, + 286, + 686 + ], + "type": "text", + "content": "[39] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In European Conference on Computer Vision (ECCV), pages 561-578. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "7692" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_content_list.json b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f11aa4ea9c5244d034b37009a8b1ed83249c6363 --- /dev/null +++ b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_content_list.json @@ -0,0 +1,1413 @@ +[ + { + "type": "text", + "text": "Meta-Explore: Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding", + "text_level": 1, + "bbox": [ + 110, + 161, + 858, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minyoung Hwang $^{1}$ , Jaeyeon Jeong $^{1}$ , Minsoo Kim $^{3}$ , Yoonseon Oh $^{2}$ , Songhwai Oh $^{1}$ $^{1}$ Electrical and Computer Engineering and ASRI, Seoul National University \n $^{2}$ Department of Electronic Engineering, Hanyang University \n $^{3}$ Interdisciplinary Major in Artificial Intelligence, Seoul National University", + "bbox": [ + 161, + 226, + 811, + 293 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{minyoung.hwang, jaeyeon.jeong}@rlllab.snu.ac.kr, goldbird5@snu.ac.kr, yoh21@hanyang.ac.kr, songhwai@snu.ac.kr", + "bbox": [ + 87, + 296, + 884, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 326, + 312, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main challenge in vision-and-language navigation (VLN) is how to understand natural-language instructions in an unseen environment. The main limitation of conventional VLN algorithms is that if an action is mistaken, the agent fails to follow the instructions or explores unnecessary regions, leading the agent to an irrecoverable path. To tackle this problem, we propose Meta-Explore, a hierarchical navigation method deploying an exploitation policy to correct misled recent actions. We show that an exploitation policy, which moves the agent toward a well-chosen local goal among unvisited but observable states, outperforms a method which moves the agent to a previously visited state. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. The key to our approach is understanding the object placements around the agent in spectral-domain. Specifically, we present a novel visual representation, called scene object spectrum (SOS), which performs category-wise 2D Fourier transform of detected objects. Combining exploitation policy and SOS features, the agent can correct its path by choosing a promising local goal. We evaluate our method in three VLN benchmarks: R2R, SOON, and REVERIE. Meta-Explore outperforms other baselines and shows significant generalization performance. In addition, local goal search using the proposed spectral-domain SOS features significantly improves the success rate by $17.1\\%$ and SPL by $20.6\\%$ against the state-of-the-art method of the SOON benchmark. Project page: https://rllab-snu.github.io/projects/Meta-Explore/doc.html", + "bbox": [ + 75, + 353, + 473, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 749, + 208, + 764 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual navigation in indoor environments has been studied widely and shown that an agent can navigate in unex-", + "bbox": [ + 75, + 769, + 468, + 798 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3edce37aa747505b4e8eaf55c9d230dc101641e3984131ef27e1d082eb0b25a1.jpg", + "image_caption": [ + "Figure 1. Hierarchical Exploration. At each episode, a natural-language instruction is given to the agent to navigate to a goal location. The agent explores the environment and constructs a topological map by recording visited nodes $\\odot$ and next step reachable nodes $\\clubsuit$ . Each node consists of the position of the agent and visual features. $o_t$ denotes the observation at time $t$ . The agent chooses an unvisited local goal to solve the regretful exploration problem." + ], + "image_footnote": [], + "bbox": [ + 522, + 326, + 870, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plored environments [1]. By recognizing the visual context and constructing a map, an agent can explore the environment and solve tasks such as moving towards a goal or following a desired trajectory. With the increasing development in human language understanding, vision-and-language navigation (VLN) [2] has enabled robots to communicate with humans using natural languages. The high degree of freedom in natural language instructions allows VLN to expand to various tasks, including (1) following fine-grained step-by-step instructions [2-13] and (2) reaching a target location described by goal-oriented language instructions [14-20].", + "bbox": [ + 497, + 546, + 892, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A challenging issue in VLN is the case when an action is mistaken with respect to the given language instruction [21-26]. For instance, if the agent is asked to turn right at the end of the hallway but turns left, the agent may end up in irrecoverable paths. Several existing studies solve this issue via hierarchical exploration, where the high-level planner decides when to explore and the low-level planner chooses what actions to take. If the high-level planner chooses to explore, the agent searches unexplored regions, and if it chooses to exploit, the agent executes the best action based on the previous exploration. Prior work [21-23] returns the agent to the last successful state and resumes exploration.", + "bbox": [ + 497, + 700, + 894, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 86 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was in part supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-01190, [SW Star Lab] Robot Learning: Efficient, Safe, and Socially-Acceptable Machine Learning, $80\\%$ , and No.2022-0-00907, Development of AI Bots Collaboration Platform and Self-Organizing AI, $20\\%$ ). (Corresponding authors: Yoonseon Oh and Songhwai Oh.)", + "bbox": [ + 75, + 807, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6683", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, such methods take a heuristic approach because the agent only backtracks to a recently visited location. The agent does not take advantage of the constructed map and instead naively uses its recent trajectory for backtracking. Another recent work [26] suggests graph-based exploitation, which uses a topological map to expand the action space in global planning. Still, this method assumes that the agent can directly jump to a previously visited node. Since this method can perform a jump action at every timestep, there is no trigger that explicitly decides when to explore and when to exploit. Therefore, we address the importance of time scheduling for exploration-exploitation and efficient global planning using a topological map to avoid reexploring visited regions.", + "bbox": [ + 75, + 125, + 468, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We expand the notion of hierarchical exploration by proposing Meta-Explore, which not only allows the high-level planner to choose when to correct misled local movements but also finds an unvisited state inferred to be close to the global goal. We illustrate the overview of hierarchical exploration in Figure 1. Instead of backtracking, we present an exploitation method called local goal search. We show that it is more efficient to plan a path to a local goal, which is the most promising node from the unvisited but reachable nodes. We illustrate the difference between conventional backtracking and local goal search in Figure 2. Based on our method, we show that exploration and exploitation are not independent and can complement each other: (1) to overtake regretful explorations, the agent can perform exploitation and (2) the agent can utilize the constructed topological map for local goal search. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. Most VLN tasks require a level of understanding objects nearby the agent, but previous studies simply encode observed panoramic or object images [2, 3, 16-18, 21-35]. In this paper, we present a novel semantic representation of the scene called scene object spectrum (SOS), which is a matrix containing the arrangements and frequencies of objects from the visual observation at each location. Using SOS features, we can sufficiently estimate the context of the environment. We show that the proposed spectral-domain SOS features manifest better linguistic interpretability than conventional spatial-domain visual features. Combining exploitation policy and SOS features, we design a navigation score that measures the alignment between a given language instruction and a corrected trajectory toward a local goal. The agent compares local goal candidates and selects a near-optimal candidate with the highest navigation score from corrected trajectories. This involves high-level reasoning related to the landmarks (e.g., bedroom and kitchen) and objects (e.g., table and window) that appear in the instructions.", + "bbox": [ + 75, + 305, + 470, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows:", + "bbox": [ + 95, + 815, + 436, + 828 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a hierarchical navigation method called Meta-Explore, deploying an exploitation policy to cor", + "bbox": [ + 95, + 838, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cae102e992b908d5afb9a1f840f5b5aa8a60047e5e8a448b980f2d3f4eab3042.jpg", + "image_caption": [ + "Figure 2. Local Goal Search for Exploitation. The local goal is likely to be chosen as the closest node to the global goal. Existing methods only backtrack to a visited node (left). We expand the searchable area by including unvisited but reachable nodes (right)." + ], + "image_footnote": [], + "bbox": [ + 500, + 122, + 875, + 283 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rect misled recent actions. The agent searches for an appropriate local goal instead of reversing the recent action sequence.", + "bbox": [ + 529, + 347, + 890, + 390 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In the exploitation mode, the agent uses a novel scene representation called scene object spectrum (SOS), which contains the spectral information of the object placements in the scene. SOS features provide semantically meaningful clues to choose a near-optimal local goal and help the agent to solve the regretful exploration problem.", + "- We evaluate our method on three VLN benchmarks: R2R [2], SOON [16], and REVERIE [17]. The experimental results show that the proposed method, MetaExplore, improves the success rate and SPL in test splits of R2R, SOON and val split of REVERIE. The proposed method shows better generalization results compared to all baselines." + ], + "bbox": [ + 517, + 393, + 890, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 595, + 640, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-and-Language Navigation", + "text_level": 1, + "bbox": [ + 499, + 612, + 786, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In VLN, an agent encodes the natural language instructions and follows the instructions, which can be either (1) a fine-grained step-by-step instruction the agent can follow [2-4], (2) a description of the target object and location [16, 17], or (3) additional guidance given to the agent [18, 27]. These tasks require the agent to recognize its current location using some words in the natural-language instructions. Prior work [2, 28-31] show that an agent can align visual features to language instructions via neural networks and use the multimodal output embeddings to generate a suitable action at each timestep. Most VLN methods utilize cross-modal attention, either with recurrent neural networks [2, 28] or with transformer-based architectures [29-31]. For sequential action prediction, Hong et al. [32] further use recurrent units inside transformer architectures, while Pashevich et al. [33] and Chen et al. [34] use additional transformers to embed past observations and actions.", + "bbox": [ + 497, + 631, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6684", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Exploration-Exploitation", + "text_level": 1, + "bbox": [ + 77, + 112, + 307, + 127 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In an unseen environment, the agent must maximize the return without knowing the true value functions. One of the solutions to this problem is to switch back and forth between exploration and exploitation [36]. In the exploration mode, the agent gathers more information about the environment. On the other hand, the agent uses information collected during exploration and chooses the best action for exploitation. Ecoffet et al. [37] reduced the exploration step by archiving the states and exploring again from the successful states. Pislar et al. [38] addressed the various scheduling policies and demonstrated their method on Atari games. Recent work [39, 40] successfully demonstrates the effectiveness of hierarchical exploration in image-goal navigation.", + "bbox": [ + 75, + 129, + 468, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Like commonly used greedy navigation policies, VLN tasks also deal with the problem of maximizing the chance to reach the goal without knowing the ground truth map. Several VLN methods employ the concept of exploitation to tackle this problem. Ke et al. [35] look forward to several possible future trajectories and decide whether to backtrack or not and where to backtrack. Others [21-23] estimate the progress to tell whether the agent becomes lost and make the agent backtrack to a previously visited location to restart exploration. However, previous studies do not take into account what should be done in the exploitation mode. In order to handle this problem, we propose a hierarchical navigation method which determines the scheduling between exploration and exploitation.", + "bbox": [ + 75, + 309, + 468, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Visual Representations", + "text_level": 1, + "bbox": [ + 77, + 505, + 290, + 519 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Popular visual encoding methods via ResNet [41] and ViT [42] can be trained to learn rotation-invariant visual features. Both methods learn to extract visual features with high information gain for global and local spatial information. The high complexity of the features leads to low interpretability of the scene and therefore requires the agent to use additional neural networks or complex processing to utilize them. On the other hand, traditional visual representation methods such as Fourier transform use spectral analysis, which is highly interpretable and computationally efficient. One drawback of the traditional methods is that they fail to maximize the information gain. Nonetheless, an appropriate use of essential information can be helpful for high-level decision making and enables more straightforward interpretation and prediction of the visual features. One traditional navigation method, Sturz et al. [43] used Fourier transform to generate rotation-invariant visual features. However, no research has transformed the spectral information of the detected objects to represent high-level semantics from visual observations. Focusing on the fact that 2D Fourier transform can extract morphological properties of images [44], we can find out the shape or structure of detected objects through 2D Fourier transform. In this paper, we decompose the object mask into binary masks by object categories and perform a 2D Fourier transform on each binary mask.", + "bbox": [ + 75, + 521, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 499, + 124, + 589, + 138 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 499, + 143, + 700, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We deal with VLN in discrete environments, where the environment is given as an undirected graph $G_{e} = \\{V,E\\}$ . $V$ denotes a set of $N$ navigable nodes, $\\{V_{i}\\}_{i = 1}^{N}$ , and $E$ is the adjacency matrix describing connectivity among the nodes in $V$ . We denote the observation at node $V_{i}$ as $O_{i}$ . The agent uses a panoramic RGB image observation $o_{t}$ and current node $v_{t}$ , which are collected at time $t$ . The agent either moves to a neighboring node or executes a stop action. $a_{t}$ denotes the action at time $t$ . The objectives of VLN are categorized as follows: (1) to follow language instructions [2] and (2) to find a target object described by language instructions in a fixed time $T$ [16, 17]. We present a general hierarchical exploration method that can be applied to both tasks. We also enhance the navigation policy by extracting cross-domain visual representations from the environments, i.e., spatial-domain and spectral-domain representations. To balance the information loss and interpretability of the visual feature, we adopt multi-channel fast Fourier transform (FFT) to encode semantic masks of the detected objects into category-wise spectral-domain features.", + "bbox": [ + 497, + 161, + 890, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Meta-Explore", + "text_level": 1, + "bbox": [ + 499, + 442, + 643, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design a learnable hierarchical exploration method for VLN called Meta-Explore, which decides (1) when to explore or exploit and (2) a new imagined local goal to seek during exploitation. The overall network architecture of the proposed Meta-Explore is shown in Figure 3. Given a language instruction $L$ , the agent navigates in the environment until it finds the target described in $L$ . Meta-Explore consists of a mode selector and two navigation modules corresponding to two modes: exploration and exploitation. At each timestep, the mode selector chooses to explore or exploit. At $t = 0$ , the mode is initialized to exploration. In the exploration mode, the agent outputs an action toward a neighboring node to move the agent toward the goal. When the mode selector recognizes that the agent is not following the instruction successfully, the mode is switched to exploitation. In the exploitation mode, the agent seeks a new local goal with the highest correspondence against the language instructions from the previously unvisited candidate nodes using spectral-domain visual features. The agent moves toward the local goal by planning a path. After the agent arrives at the local goal, the mode is reset to exploration. The explore-exploit switching decision occurs through the mode selector by estimating the probability to explore. The agent repeats this explore-exploit behavior until it determines that the target is found and decides to stop.", + "bbox": [ + 497, + 460, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Mode Selector", + "text_level": 1, + "bbox": [ + 499, + 809, + 653, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At time $t$ , the agent observes visual features about the current node $v_{t}$ and several reachable nodes. We call the nodes reachable at the current timestep as candidate nodes. $n_{cand}$", + "bbox": [ + 497, + 824, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6685", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/beec1b93785ed13d1fb664fb0d718a97f7174f0600af642877fcb3add3493e74.jpg", + "image_caption": [ + "Figure 3. Network Architecture. Three types of visual features: panoramic (yellow), object image (aquamarine), and object spectrum (red) are encoded. The color in each parenthesis denotes the color describing the corresponding feature. The cross-modal transformer encodes language and spatial visual features as hidden state $H_{t}$ . A mode selector gives explore or exploit command to the agent by predicting the explore probability $P_{\\text{explore}}$ . The selected navigation module outputs an action $a_{t}$ from the possible $n_{\\text{cand}}$ candidate nodes." + ], + "image_footnote": [], + "bbox": [ + 163, + 124, + 811, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "denotes the number of candidate nodes. We use a cross-modal transformer with $n_{L}$ layers to relate visual observations to language instructions. The cross-modal transformer takes the visual features of nodes in the constructed topological map at time $t$ , $G_{t}$ , and outputs cross-modal embedding $H_{t}$ to encode visual observations with $L$ . We concatenate location encoding and history encoding [24] to the visual features as node features to consider the relative pose from $v_{t}$ and the last visited timestep of each node, respectively. Each word is encoded via a pretrained language encoder [45], which is used for general vision-language tasks.", + "bbox": [ + 75, + 408, + 470, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cross-modal transformer consists of cross-attention layer L2V_Atn(\\hat{W},\\hat{V}) = Softmax(\\hat{W}\\Theta_q^\\dagger (\\hat{V}\\Theta_k^\\dagger)^T /\\sqrt{d})\\hat{V}\\Theta_v^\\dagger and self-attention layer SelfAttn(X) = Softmax $(X\\Theta_{q}(X\\Theta_{k})^{T}+$ $A\\Theta_e + b_e) / \\sqrt{d})X\\Theta_v$ , where $\\hat{W},\\hat{V},X$ , and $A$ denote word, visual, node representations and adjacency matrix of $G_{t}$ respectively. The (query, key, value) weight matrices of self-attention and cross-attention layers are denoted as $(\\Theta_q,\\Theta_k,\\Theta_v)$ and $(\\Theta_q^\\dagger ,\\Theta_k^\\dagger ,\\Theta_v^\\dagger)$ , respectively. The final crossmodal embedding at time $t$ after passing through $n_L$ transformer layers is denoted as $H_{t}$ . To encourage the monotonic increasing relationship between language and visual attentions at each timestep, we define a correlation loss $L_{corr} = \\sum_{t = 1}^{T}\\| \\mathrm{L2V\\_Attn} - I_{n_x}\\| _1$ for training the crossmodal transformer, where $n_x$ denotes the dimension of the $H_{t}$ and $I_{n_x}$ denotes an identity matrix of size $n_x\\times n_x$", + "bbox": [ + 75, + 561, + 468, + 769 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 4, the mode selector estimates the probability to explore $P_{\\text{explore}}$ given the cross-modal hidden state $H_t$ . We denote the mode selector as $S_{\\text{mode}}$ and use a two-layer feed-forward neural network. Given $H_t$ , $S_{\\text{mode}}$ outputs the exploration probability as $P_{\\text{explore}} = 1 - S_{\\text{mode}}(H_t)$ . If $P_{\\text{explore}} \\geq 0.5$ , the exploration policy outputs a probability distribution for reachable nodes at the", + "bbox": [ + 75, + 769, + 470, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "next step. At time $t + 1$ , the agent moves to the node with the highest probability. If $P_{\\text{explore}} < 0.5$ , the agent determines that the current trajectory is regretful, so the agent should traverse to find a local goal, which is the most likely to be the closest node to the global goal. The exploitation policy mainly utilizes object-level features to search for the local goal with high-level reasoning. After the local goal is chosen, the path planning module outputs an action following the shortest path to the local goal.", + "bbox": [ + 497, + 408, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train the mode selector, we require additional demonstration data other than the ground truth trajectory, such that it switches between exploration and exploitation. We generate the demonstration data from the ground truth trajectories, with additional detours. For the detours, we stochastically select candidate nodes other than the ground truth paths and add the trajectory that returns to the current viewpoint. The imitation learning loss for training the mode selector is defined as $L_{mode} = \\sum_{t=1}^{T} \\mathbb{1}(m_t = \\mathrm{gt}_t)$ , where $m_t$ is the mode of the agent, 0 for exploitation and 1 for exploration. $\\mathrm{gt}_t$ is 1 if the current node is in the shortest ground truth trajectory and $\\mathrm{gt}_t = 0$ , otherwise.", + "bbox": [ + 497, + 533, + 892, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Exploration Module", + "text_level": 1, + "bbox": [ + 499, + 707, + 694, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the exploration mode, the agent follows the following sequential operations: topological map construction, self-monitoring, and an exploration policy. To improve the exploration, we adopt self-monitoring [21] to predict the current progress of exploration to enhance the exploration policy itself. Prior work [21, 22] has shown that auxiliary loss using self-monitoring can regularize the exploration policy.", + "bbox": [ + 497, + 723, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Topological Map Construction. The agent constructs graph $G_{t}$ by classifying nodes into two types: (1) visited nodes and (2) unvisited but observable nodes. At current", + "bbox": [ + 497, + 825, + 892, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6686", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a84bd4402484b8d5c99c18b4e3a47c29d64252e158398d83efef5e5d34ea4b21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 122, + 257, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/49a1a66bfa018bbdb7f02df07e6ecca1068f310984d4e327817c647b63e6fe41.jpg", + "image_caption": [ + "Exploration Module", + "Figure 4. Navigation Modules. Mode selector estimates $P_{\\text{explore}}$ , i.e., the probability to explore, and chooses between exploration and exploitation modules. The selected navigation module outputs the next action $a_t$ ." + ], + "image_footnote": [], + "bbox": [ + 99, + 200, + 253, + 296 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c4a582aac44a3dbcdcd77c052679afa3b2a36df0d9cf7919d8eaa275f55f1a0.jpg", + "image_caption": [ + "Exploitation Module" + ], + "image_footnote": [], + "bbox": [ + 258, + 122, + 448, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "time $t$ , the agent at node $v_{t} \\in \\{V_{i}\\}_{i=1}^{N}$ observes $N(v_{t})$ neighbor nodes as next step candidates at time $t + 1$ . The visited nodes consist of visual features of their own and the neighboring nodes from panoramic RGB observations. The unvisited nodes can be observed only if they are connected to at least one visited node. The topological map records the positions and visual features of observed nodes at each timestep. By knowing the positions of nodes in $G_{t}$ , the agent can plan the shortest path trajectory between two nodes.", + "bbox": [ + 75, + 374, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-Monitoring. We use a progress monitor to estimate the current navigation progress at each episode. Self-monitoring via estimating current progress helps the agent choose the next action that can increase the progress. The estimated progress $\\hat{p}_t = F_{\\text{progress}}(H_t)$ is the output of a feed-forward neural network, given $H_t$ as input. We measure the ground truth progress $p_t$ as the ratio between the current distance to the goal and the shortest path length of the episode subtracted from 1, described as $1 - \\frac{d_{geo}(v_t, v_{goal})}{d_{geo}(v_0, v_{goal})}$ , where $d_{geo}(a, b)$ is the geodesic distance between $a$ and $b$ . $v_0, v_t$ , and $v_{goal}$ denote initial, current, and goal positions, respectively. We add progress loss $L_{\\text{progress}} = \\sum_{t=1}^{T} (\\hat{p}_t - p_t)^2$ to train the progress monitor while training the exploration policy.", + "bbox": [ + 75, + 505, + 468, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exploration Policy. The exploration policy $F_{explore}$ estimates the probability of moving to the candidate nodes at the next step. The agent chooses the action $a_t$ at time $t$ based on the estimated probability distribution among candidate nodes, described as $a_t = \\arg \\max_{V_i} (F_{explore}([H_t]_i))$ . $F_{explore}$ is implemented via a two-layer feed-forward network with the cross-modal hidden state $H_t$ given as input. The output of $F_{explore}$ becomes a probability distribution over possible actions. To only consider unvisited nodes, we mask out the output for visited nodes. For training, we sample the next action from the probability distribution instead of choosing a node with the highest probability. We describe the training details in Section 3.3.", + "bbox": [ + 75, + 695, + 470, + 876 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f87060c14ebff4c2fc1cd3db0294d7b820c3aa365ea34d85d467c07f4f0fc055.jpg", + "image_caption": [ + "Figure 5. Scene Object Spectrum (SOS). The agent calculates scene object spectrum (SOS) features for efficient exploitation. SOS features incorporate semantic information observed in a single panoramic image by performing category-wise 2D FFT." + ], + "image_footnote": [], + "bbox": [ + 512, + 122, + 880, + 264 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Exploitation Module", + "text_level": 1, + "bbox": [ + 499, + 332, + 697, + 346 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the exploitation mode, the agent requires high-level reasoning with identifiable environmental clues to imagine regretful exploration cases. To find clues in an object-level manner, we present a novel visual representation by capturing object information in the spectral-domain. The novel representation is more easily predictable than spatial features such as RGB image embeddings. The agent can take advantage of the predictability by expanding the searchable area to find a local goal. We choose the local goal as the closest node to the global goal in the feature space.", + "bbox": [ + 497, + 349, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spectral-Domain Visual Representations. Common navigation policies can lead the agent toward the node with the highest similarity to the target. However, even with a good learned policy, the agent can act in a novice manner in unseen environments. In this paper, we seek extra information from the environment for generalizable high-level reasoning to resolve the issue. As illustrated in Figure 5, scene object spectrum (SOS) incorporates semantic information observed in a single panoramic image by generating a semantic mask for each object category and applying Fourier transform to each semantic mask. The semantic mask for object class $k$ at time $t$ is calculated as a binary mask $[m_t^k]_{ij}$ that detects the object at pixel $(i,j)$ . Suppose there are a total of $K$ object categories. When multiple objects are detected for one object category, the binary mask appears as a union of the bounding boxes of the detected objects. We define FFT as a channel-wise 2D fast Fourier transform that receives $K$ binary semantic masks and outputs $K$ spectral-domain features, where $K$ is the number of object classes. Then, SOS feature $\\vec{S}_t = [s_t^1,\\dots,s_t^K ]^T$ can be defined as $s_t^k = \\log |\\mathbf{FFT}(m_t^k)|$ . For simplicity, we perform mean pooling on the vertical spectral axis and normalize the output. The final SOS feature has shape $(K,\\eta)$ , where $\\eta$ is the maximum horizontal frequency.", + "bbox": [ + 497, + 492, + 892, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local Goal Search Using Semantic Clues. We argue that returning to a previously visited node does not guarantee the agent escapes from the local optima. Instead of backtracking", + "bbox": [ + 497, + 824, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6687", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to a previously visited node, the agent searches for a local goal to move towards. If the agent plans a path and moves towards the local goal, the agent does not need to repeat unnecessary actions in visited regions after the exploitation ends. Additionally, searching for a local goal takes full advantage of the topological map by utilizing the connections among the observed nodes. To expand the searchable area further, we let the agent choose the local goal from previously unvisited and unchosen candidate nodes.", + "bbox": [ + 75, + 125, + 468, + 249 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To choose a local goal, we first score the corrected trajectories to measure the alignment with the language instruction $L$ . We use SOS features as semantic environmental clues to estimate the navigation score $S_{\\text{nav}}$ of the corrected trajectory, which is the shortest path trajectory from the initial node to the local goal in the constructed topological map. To simplify, we convert the language instruction into a list of objects $W^{o} = [w_{1}^{o}, \\dots, w_{B}^{o}]$ consisting of $B[\\leq K]$ object categories (e.g., desk, cabinet, and microwave). We approximate the corresponding reference SOS features as $[\\hat{\\delta}(w_{1}^{o}), \\dots, \\hat{\\delta}(w_{B}^{o})]$ where the $i^{th}$ row of $\\hat{\\delta}(w_{k}^{o})$ is defined as $[\\hat{\\delta}(w_{k}^{o})]_{i} = \\mathbb{1}(k = i)\\lambda(\\delta(w_{k}^{o}))\\mathrm{sinc}\\left(\\frac{j}{2} - \\frac{\\eta}{4}\\right)$ . $\\lambda(\\delta(w_{k}^{o}))$ denotes the average width of detected bounding boxes of object $w_{k}^{o}$ in the environment. A detailed approximation process is explained in the supplementary material. To simulate a corrected trajectory $\\mathcal{T}' = (v_{1}', \\dots, v_{t'})$ , we calculate the SOS features $[\\vec{S}_1', \\dots, \\vec{S}_{t'}']$ corresponding to the nodes in $\\mathcal{T}'$ . We measure the similarity between two object spectrum features via the cosine similarity of the flattened vectors. Finally, the navigation score $S_{\\text{nav}}$ of $\\mathcal{T}'$ is computed as:", + "bbox": [ + 75, + 250, + 470, + 519 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS _ {n a v} \\left(\\mathcal {T} ^ {\\prime}\\right) = \\frac {\\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {t ^ {\\prime}} \\left(\\frac {\\hat {\\delta} \\left(w _ {i} ^ {o}\\right)}{\\left| \\hat {\\delta} \\left(w _ {i} ^ {o}\\right) \\right|} \\cdot \\frac {\\vec {S} _ {j} ^ {\\prime}}{\\left| \\vec {S} _ {j} ^ {\\prime} \\right|}\\right) \\left(\\left(\\hat {\\delta} \\left(w _ {i} ^ {o}\\right) - \\hat {\\delta} (\\bar {w} ^ {o})\\right) \\cdot \\left(\\vec {S} _ {j} ^ {\\prime} - \\bar {\\vec {S}} ^ {\\prime}\\right)\\right)}{\\sqrt {\\frac {t ^ {\\prime}}{B} \\cdot \\sum_ {i = 1} ^ {B} \\left(\\hat {\\delta} \\left(w _ {i} ^ {o}\\right) - \\hat {\\delta} (\\bar {w} ^ {o})\\right) ^ {2} \\sum_ {j = 1} ^ {t ^ {\\prime}} \\left(\\vec {S} _ {j} ^ {\\prime} - \\bar {\\vec {S}} ^ {\\prime}\\right) ^ {2}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 77, + 524, + 468, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\delta} (\\overline{w}^o)$ and $\\vec{S}'$ denote the average values of SOS features $\\hat{\\delta} (w_i^o)$ and $\\vec{S}_j^\\prime$ , respectively. This equation can also be interpreted as a pseudo correlation-coefficient function between object list $W^{o}$ and trajectory $\\mathcal{T}'$ . The exploitation policy selects the node with the highest navigation score as the local goal from the previously unvisited candidates.", + "bbox": [ + 75, + 590, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6 illustrates a simple scenario of entering a room. Suppose $W^{o} =$ [sculpture, door, bed] and the agent has to compare two trajectories $\\mathcal{T}_1 = (v_1,v_2,A)$ and $\\mathcal{T}_1 = (v_1,v_2,B)$ . Each similarity matrix in Figure 6 has the $(t,j)$ element as the similarity between the SOS feature of $V_{t}$ and $\\hat{\\delta}(w_j^o)$ , which is calculated as $\\hat{\\delta}(w_j^o) \\cdot \\vec{S}_t'$ . Notably, the similarity matrix shows monotonic alignment and the navigation score is higher when the next action is chosen correctly.", + "bbox": [ + 75, + 678, + 468, + 788 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Training Details", + "text_level": 1, + "bbox": [ + 75, + 794, + 236, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use [24] for pretraining the visual encoder with panoramic RGB observations. We use the DAgger algorithm [46] to pretrain the navigation policy and the mode selector. To prevent overfitting, we iteratively perform", + "bbox": [ + 75, + 811, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/953ad6452a75b7dd3dfd02078929ac7b6d9100d714fc9faa876085f058d646c8.jpg", + "image_caption": [ + "Figure 6. Toy Example. Monotonic alignment between language instruction and visual observation is desirable. Yellow dots $\\bullet$ in the nodes describe the ground truth trajectory. Based on the node at $t = 3$ , the similarity matrix can show either monotonic or non-monotonic alignment between object tokens and SOS features. The green circles $\\clubsuit$ describe the possible candidates $A, B$ for next action." + ], + "image_footnote": [], + "bbox": [ + 500, + 122, + 769, + 271 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/836f12f84d172a7363a704ee5a7c22f441b83c33d87b40a05c1d4ad46e4f1703.jpg", + "table_caption": [ + "similarity matrix" + ], + "table_footnote": [], + "table_body": "
w01w2w3
v1=10.90.00.0
v1=20.31.00.0
v1=30.00.01.0
v1=4, Snav = 0.296
w01w2w3
v1=10.90.00.0
v1=20.31.00.0
v1=30.50.40.0
v1=4, Snav = 0.296
", + "bbox": [ + 784, + 136, + 890, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "teacher forcing and student forcing to choose the action from the exploration policy. Imitation learning loss is calculated as $L_{IL} = \\sum_{t=1}^{T} -\\log p(a_t^* | a_t)$ and object grounding loss is calculated as $L_{OG} = -\\log p(\\mathrm{obj}^* | \\mathrm{obj}_{pred})$ , where $\\mathrm{obj}^*$ denotes the ground truth and $\\mathrm{obj}_{pred}$ denotes the predicted object location. The total loss function is defined as $L_{total} = L_{mode} + L_{progress} + L_{corr} + L_{IL} + L_{OG}$ . We further finetune the agent via A2C [47]. The exploration policy selects the action $a_t$ with probability $p_t^a$ . Reinforcement learning loss is defined as $L_{RL} = -\\sum_t a_t^s \\log(p_t^a) A_t - \\lambda \\sum_t a_t^* \\log(p_t^a)$ . To train the mode selector, progress monitor, and exploration policy in an end-to-end manner, we use the total loss function as $L_{fine} = L_{mode} + L_{progress} + L_{RL}$ . The exploitation policy searches the path toward the local goal from the constructed navigation graph. Thus, the exploitation policy is not learned.", + "bbox": [ + 497, + 364, + 890, + 586 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Navigation Experiments", + "text_level": 1, + "bbox": [ + 497, + 593, + 726, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 499, + 612, + 692, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on three VLN benchmarks, Room-to-Room(R2R) [2], SOON [16], and REVERIE [17]. R2R evaluates the visually-grounded natural navigation performance of the agent. The agent must navigate to the predefined goal point given image observations and language instructions in an unseen environment.", + "bbox": [ + 497, + 629, + 890, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SOON is also a goal-oriented VLN benchmark. Natural language instructions in SOON have an average length of 47 words. The agent should locate the target location and detect the location of an object to find the target object.", + "bbox": [ + 497, + 713, + 890, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "REVERIE is a goal-oriented VLN benchmark that provides natural language instruction about target locations and objects. In REVERIE, the agent is given an instruction referring to a remote object with an average length of 21 words. With this instruction and a panoramic observation from the environment, the agent should navigate to the location the instruction describes and find the correct object", + "bbox": [ + 497, + 769, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6688", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bounding box among the predefined object bounding boxes.", + "bbox": [ + 75, + 125, + 467, + 138 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 77, + 144, + 260, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Navigation performance", + "text_level": 1, + "bbox": [ + 77, + 162, + 299, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate algorithms using the trajectory length (TL), success rate (SR), and success weighted by inverse path length (SPL) [48], and oracle success rate (OSR) for the navigation performance comparison. An episode is recorded as a success if the agent takes a stop action within $3\\mathrm{m}$ of the target location. TL is the average path length in meters. SR is denoted as the number of successes divided by the total number of episodes, $M$ . SPL is calculated as $\\frac{1}{M}\\sum_{i = 1}^{M}S_{i}\\frac{l_{i}}{\\max(p_{i},l_{i})}$ , where $S_{i}$ denotes the success as a binary value. $p_i$ and $l_{i}$ denote the shortest path and actual path lengths for the $i^{th}$ episode. OSR uses the oracle stop policy instead of the stop policy of the agent.", + "bbox": [ + 75, + 178, + 468, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Object grounding performance", + "text_level": 1, + "bbox": [ + 77, + 349, + 347, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate the object grounding performance of the agent by the success rate of finding the target object (FSR) and the target finding success weighted by inverse path length (FSPL)1 [16, 17]. FSPL is calculated as FSPL = $\\frac{1}{N}\\sum_{i=1}^{N}S_{i}^{nav}S_{i}^{loc} \\cdot l_{i}^{nav}/\\max(l_{i}^{nav},l_{i}^{gt})$ , where $S_{i}^{nav}$ is whether the agent navigates to the target, $S_{i}^{loc}$ is whether the agent finds a target object bounding box, and $l_{i}^{nav}$ and $l_{i}^{gt}$ are the navigation trajectory length and ground truth trajectory length, respectively.", + "bbox": [ + 75, + 365, + 468, + 490 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Baselines and Implementation Details", + "text_level": 1, + "bbox": [ + 77, + 494, + 401, + 509 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with several other baselines as follows. For each task, we compare our method with a number of baselines that use various types of memory (recurrent, sequential, and topological map). For methods implemented with a hierarchical navigation framework, we compare the specific exploitation methods: homing, jump, and local goal search. Homing makes the agent backtrack, and jump makes the agent jump to a previously visited node. The hyperparameters and detailed model architecture of Meta-Explore are described in the supplementary material.", + "bbox": [ + 75, + 511, + 468, + 649 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with Navigation Baselines", + "text_level": 1, + "bbox": [ + 77, + 652, + 408, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with navigation baselines. We focus on the success rate and SPL. Rendered results and detailed analyses with other evaluation metrics are provided in the supplementary material.", + "bbox": [ + 75, + 668, + 468, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "R2R. Table 1 compares the proposed Meta-Explore with baselines for the R2R navigation task. We categorize the baseline methods based on the type of constructed memory and the type of exploitation. Our method outperforms other exploration-only baselines over all types of validation and test splits in success rate and SPL. Compared with hierarchical baselines SMNA [21], Regretful-Agent [22], FAST [35], and SSM [26], Meta-Explore improves success", + "bbox": [ + 75, + 724, + 467, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rate and SPL by at least $16.4\\%$ and $8.9\\%$ , respectively. The main difference is that Meta-Explore constructs a topological map during exploration and uses the map for local goal search in exploitation. On the contrary, homing exploitation policies in SMNA, Regretful-Agent, and FAST only rely on the current trajectory, instead of taking advantage of the constructed memory. Jump exploitation in SSM uses a topological map to search a successful previous node, but it makes an unrealistic assumption that the agent can directly jump to a previously visited distant node and unfairly saves time. In our approach, we plan a path to the local goal based on the topological map. The experiment results reveal that even if we design a hierarchical navigation framework, exploration and exploitation are not entirely separate but they can complement each other.", + "bbox": [ + 499, + 125, + 890, + 331 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SOON, REVERIE. Table 2 compares Meta-Explore with baselines in the SOON navigation task. While the proposed method does not improve performance in val seen split, Meta-Explore outperforms other baselines in the test unseen split of SOON for success rate by $17.1\\%$ and SPL by $20.6\\%$ . The result implies that for the goal-oriented VLN task, high performance in train or val seen splits can be the overfitted result. Because the agent can be easily overfitted to the training data, making a generalizable model or providing a deterministic error-correction module for inference is essential. Meta-Explore chooses the latter approach by correcting the trajectory via exploitation in regretful cases. The evaluation results in the REVERIE navigation task are described in the supplementary material. Meta-Explore shows improvement in the val split of REVERIE for success rate and SPL, but the improvement in the test split is lower than the results in R2R and SOON. We found 252 meaningless object categories (e.g., verbs, adjectives, and prepositions) and 418 replaceable object categories (e.g., typographical errors and synonyms) in the REVERIE3 dataset. Because our exploitation method utilizes object-based parsing of the given instruction to match with the detected object categories, the effectiveness of the proposed method is lessened due to inaccuracies and inconsistencies in the dataset. We expect to have higher performance if the mistakes in the dataset are fixed.", + "bbox": [ + 497, + 333, + 890, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Local Goal Search using SOS Features", + "text_level": 1, + "bbox": [ + 499, + 695, + 830, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To discuss the significance of modeling exploitation policy, we conduct specific experiments about choosing the local goal for R2R and SOON. We evaluate our method using different types of local goal search, as shown in Table 3 and 4. Oracle denotes a method which selects a local goal using the ground truth trajectory. The performance of the oracle provides the achievable performance for each dataset. The results imply that local goal search using either", + "bbox": [ + 497, + 711, + 890, + 821 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1Identical with its original term, Remote Grounding Success (RGS).", + "bbox": [ + 92, + 840, + 453, + 853 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$2^{\\dagger}$ indicates reproduced results.", + "bbox": [ + 95, + 853, + 263, + 865 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$310.7\\%$ and $41.2\\%$ of a total of 46,476 words in the bounding box dataset correspond to meaningless and replaceable object categories, respectively.", + "bbox": [ + 499, + 831, + 890, + 865 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6689", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/39c78cef6cb759ef02796f4670567674882b0af678015557444c28ca87363894.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMemoryExploitVal SeenVal UnseenTest Unseen
SR↑SPL↑TL↓NE↓SR↑SPL↑TL↓NE↓SR↑SPL↑TL↓NE↓
Random--16-9.589.4516-9.779.2313129.899.79
Human----------11.851.618676
Seq2Seq [2]RecX6.03911.33-22-8.397.8420188.137.85
VLN\\(\\circ\\)BERT [32]RecX726811.132.90635712.013.93635712.354.09
SMNA†[21]Rechoming696311.693.31474112.615.486156-4.48
Regretful-Agent [22]Rechoming6963-3.235041-5.324840-5.69
FAST (short) [35]Rechoming----564321.174.97544122.085.14
FAST (long) [35]Rechoming7004188.063.136302224.424.036103196.534.29
HAMT-e2e [34]SeqX767211.152.51666111.462.29656012.273.93
DUET [24]Top. MapX797312.322.28726013.943.31695914.733.65
SSM [26]Top. Mapjump716214.73.10624520.74.32614620.44.57
Meta-Explore (Ours)Top. Maplocal goal817511.952.11726213.093.22716114.253.57
", + "bbox": [ + 78, + 124, + 892, + 271 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/07feae75b547e47b2a2080788085b4c9b79b5b21f5a979c612a2ae4bf20dbed3.jpg", + "table_caption": [ + "Table 1. Comparison and evaluation results of the baselines and our model in the R2R Navigation Task. Gray shaded rows describe hierarchical navigation baselines. Three memory types: Rec(recurrent), Seq(sequential), and Top. Map(topological map)" + ], + "table_footnote": [], + "table_body": "
MethodsMemoryExploitVal Seen InstructionVal Seen HouseTest Unseen House
SR↑SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑
Human----------90.459.291.451.1
RandomRecX0.01.50.11.40.10.00.40.92.10.42.70.0
Speaker-Follower [28]RecX97.997.797.824.561.260.469.49.17.06.19.80.6
RCM [49]RecX84.082.689.110.962.460.972.77.87.46.212.40.7
AuxRN [23]RecX98.497.498.713.768.867.378.58.38.16.711.00.5
GBE w/o GETop. MapX89.588.391.824.262.560.873.06.711.48.718.80.8
GBE [16]Top. MapX98.497.998.644.276.362.564.17.311.910.219.51.4
\\( GBE^† \\)Top. MapX----19.513.328.51.212.99.221.50.5
DUET [24]Top. MapX94.091.690.031.136.322.650.93.833.421.443.04.2
Meta-Explore (Ours)Top. Maplocal goal100.099.196.033.944.734.852.78.939.125.848.74.0
", + "bbox": [ + 78, + 304, + 892, + 432 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "spatial or spectral visual representations is more effective than random local goal search. The results show that local goal search using spectral visual representations, i.e., SOS features, lead the agent to desirable nodes the most. We also compare local goal search with homing and the difference between the performance of the two methods is most noticeable in the test split of the SOON navigation task. As shown in Table 4, choosing the local goal with only spatial-domain features, the navigation performance does not improve compared to homing. On the contrary, spectral-domain local goal search shows significant improvement against homing by $10.4\\%$ in success rate, $34.5\\%$ on SPL, and $27.4\\%$ on FSPL. The results imply that using spectral-domain SOS features helps high-level decision making, thereby enhancing the navigation performance. To further show the effectiveness of SOS features, we provide sample local goal search scenarios in the supplementary material.", + "bbox": [ + 75, + 458, + 468, + 694 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a7461b174b2dc80f1684741d73048eebc0e3203a5eebe3b6fd52ddc6b153b40a.jpg", + "table_caption": [ + "Table 2. Comparison and evaluation results of the baselines and our model in the SOON Navigation Task." + ], + "table_footnote": [], + "table_body": "
Local GoalVal SeenVal Unseen
SR↑SPL↑OSR↑TL↓NE↓SR↑SPL↑OSR↑TL↓NE↓
Oracle81.8874.1287.4613.061.9375.9562.5384.1614.002.71
Random79.3372.6785.3113.192.2270.9759.4580.1614.923.34
Homing80.2273.6385.6012.512.1471.6560.6080.3313.913.26
Spatial79.6373.1485.6012.992.2271.5660.0180.3314.903.27
Spectral80.6175.1585.8011.952.1171.7861.6880.7613.093.22
", + "bbox": [ + 78, + 697, + 468, + 767 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ab44f70c9422182776a4fe87432d10326335e52d22ff454561812d763660587.jpg", + "table_caption": [ + "Table 3. Comparison of Exploitation Policies. (R2R)" + ], + "table_footnote": [], + "table_body": "
Local GoalSR↑Val Seen HouseTest Unseen House
SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑
Oracle54.4237.9663.7211.0148.3828.4562.984.74
Random24.7811.9734.963.0824.197.4135.841.29
Homing42.0427.7248.2310.1835.4019.1851.623.14
Spatial32.3011.6039.381.9026.1110.5839.231.43
Spectral44.6934.8452.658.8939.0925.8048.674.01
", + "bbox": [ + 78, + 789, + 468, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison of Exploitation Policies. (SOON)", + "bbox": [ + 112, + 862, + 433, + 875 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 499, + 458, + 653, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study to compare the proposed method against language-triggered hierarchical exploration. Results in the supplementary material show that among the three representation domains, spatial, spectral, and language, the spectral-domain features enhance navigation performance the most. Additionally, to implicate further applications of Meta-Explore in continuous environments, we evaluate our method on the photo-realistic Habitat [50] simulator to solve image-goal navigation and vision-and-language navigation tasks. Implementation details and results are included in the supplementary material. Results show that our method outperforms baselines in both tasks.", + "bbox": [ + 497, + 473, + 890, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 642, + 618, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed Meta-Explore, a hierarchical navigation method for VLN, by correcting mistaken short-term actions via efficient exploitation. In the exploitation mode, the agent is directed to a local goal which is inferred to be the closest to the target. A topological map constructed during exploration helps the agent to search and plan the shortest path toward the local goal. To further search beyond the frontier of the map, we present a novel visual representation called scene object spectrum (SOS), which compactly encodes the arrangements and frequencies of nearby objects. Meta-Explore achieves the highest generalization performance for test splits of R2R, SOON, and val split of REVERIE navigation tasks by showing less overfitting and high success rates. We plan to apply Meta-Explore for VLN tasks in continuous environments in our future work.", + "bbox": [ + 497, + 657, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6690", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 173, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fengda Zhu, Yi Zhu, Vincent Lee, Xiaodan Liang, and Xiao-jun Chang. Deep learning for embodied vision navigation: A survey. arXiv preprint arXiv:2108.04097, 2021. 1", + "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2018. 1, 2, 3, 6, 8", + "[3] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. Touchdown: Natural language navigation and spatial reasoning in visual street environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12538-12547, 2019. 1, 2", + "[4] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Conference on Empirical Methods for Natural Language Processing, 2020. 1, 2", + "[5] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In *Transactions of the Association for Computational Linguistics*, pages 1862–1872, Florence, Italy, July 2019. Association for Computational Linguistics. 1", + "[6] An Yan, Xin Eric Wang, Jiangtao Feng, Lei Li, and William Yang Wang. Cross-lingual vision-language navigation. arXiv preprint arXiv:1910.11301, 2019. 1", + "[7] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. In Proceedings of the International Conference on Neural Information Processing Systems, volume 34, pages 652-663. Curran Associates, Inc., 2021. 1", + "[8] Jacob Krantz, Erik Wijmans, Arjun Majundar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision and language navigation in continuous environments. In Proceedings of the European Conference on Computer Vision, 2020. 1", + "[9] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teplyashin, Karl Moritz Hermann, Mateusz Malinowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 1", + "[10] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11773–11781, Apr. 2020. 1", + "[11] Arun Balajee Vasudevan, Dengxin Dai, and Luc Van Gool. Talk2nav: Long-range vision-and-language navigation with dual attention and spatial memory. In Proceedings of the" + ], + "bbox": [ + 78, + 147, + 470, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, volume 129, pages 246-266. Springer, 2021. 1", + "[12] Dipendra Misra, Andrew Bennett, Valts Blukis, Eyvind Niklasson, Max Shatkhin, and Yoav Artzi. Mapping instructions to actions in 3D environments with visual goal prediction. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 2667-2678, Brussels, Belgium, October-November 2018. Association for Computational Linguistics. 1", + "[13] Ta-Chung Chi, Minmin Shen, Mihail Eric, Seokhwan Kim, and Dilek Hakkani-tur. Justask: An interactive learning framework for vision and language navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2459-2466, 2020. 1", + "[14] Yi Wu, Yuxin Wu, Georgia Gkioxari, and Yuandong Tian. Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209, 2018. 1", + "[15] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2018. 1", + "[16] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12689-12699, 2021. 1, 2, 3, 6, 7, 8", + "[17] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 1, 2, 3, 6, 7", + "[18] Khanh Nguyen and Hal Daume III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, November 2019. 1, 2", + "[19] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2019. 1", + "[20] Alane Suhr, Claudia Yan, Jack Schluger, Stanley Yu, Hadi Khader, Marwa Mouallem, Iris Zhang, and Yoav Artzi. Executing instructions in situated collaborative interactions. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 2119-2130, 2019. 1", + "[21] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In Proceedings of the International Conference on Learning Representations, 2019. 1, 2, 3, 4, 7, 8", + "[22] Chih-Yao Ma, Zuxuan Wu, Ghassan AlRegib, Caiming Xiong, and Zsolt Kira. The regretful agent: Heuristic-aided navigation" + ], + "bbox": [ + 502, + 125, + 892, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6691", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "through progress estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6732-6740, 2019. 1, 2, 3, 4, 7, 8", + "[23] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10012-10022, 2020. 1, 2, 3, 8", + "[24] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537-16547, 2022. 1, 2, 4, 6, 8", + "[25] Zhiwei Deng, Karthik Narasimhan, and Olga Russakovsky. Evolving graphical planner: Contextual global planning for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, volume 33, pages 20660-20672, 2020. 1, 2", + "[26] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8455-8464, June 2021. 1, 2, 7, 8", + "[27] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Proceedings of the Conference on Robot Learning, pages 394-406, 2020. 2", + "[28] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, 2018. 2, 8", + "[29] Xiujun Li, Chunyuan Li, Qiaolin Xia, Yonatan Bisk, Asli Celikyilmaz, Jianfeng Gao, Noah A Smith, and Yejin Choi. Robust navigation with language pretraining and stochastic sampling. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 1494-1499, 2019. 2", + "[30] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Proceedings of the International Conference on Neural Information Processing Systems, 2019. 2", + "[31] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1634-1643, 2021. 2", + "[32] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. VIn $\\text{©}$ bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2, 8" + ], + "bbox": [ + 78, + 125, + 468, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15942-15952, 2021. 2", + "[34] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, pages 5834-5847, 2021. 2, 8", + "[35] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6741-6749, 2019. 2, 3, 7, 8", + "[36] James G March. Exploration and exploitation in organizational learning. Organization science, 2(1):71-87, 1991. 3", + "[37] Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O Stanley, and Jeff Clune. First return, then explore. Nature, 590(7847):580-586, 2021. 3", + "[38] Miruna Pislar, David Szepesvari, Georg Ostrovski, Diana L Borsa, and Tom Schaul. When should agents explore? In Proceedings of the International Conference on Learning Representations, 2022. 3", + "[39] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In Proceedings of the International Conference on Learning Representations, 2020. 3", + "[40] Meera Hahn, Devendra Singh Chaplot, Shubham Tulsiani, Mustafa Mukadam, James M Rehg, and Abhinav Gupta. No rl, no simulation: Learning to navigate without navigating. In Proceedings of the International Conference on Neural Information Processing Systems, volume 34, pages 26661-26673. Curran Associates, Inc., 2021. 3", + "[41] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3", + "[42] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 3", + "[43] Wolfgang Stürzl and Hanspeter A Mallot. Efficient visual homing based on fourier transformed panoramic images. Robotics and Autonomous Systems, 54(4):300-313, 2006. 3", + "[44] Jean Serra. Mathematical Morphology, pages 1-16. Springer International Publishing, Cham, 2020. 3", + "[45] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 5100-5111, 2019. 4" + ], + "bbox": [ + 502, + 127, + 890, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6692", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Stéphane Ross, Geoffrey J Gordon, and J Andrew Bagnell. No-regret reductions for imitation learning and structured prediction. arXiv preprint arXiv:1011.0686, 2010. 6", + "[47] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proceedings of the International Conference on Machine Learning, pages 1928-1937, 2016. 6", + "[48] Peter Anderson, Angel Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, et al. On evaluation of embodied navigation agents. arXiv preprint arXiv:1807.06757, 2018. 7", + "[49] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6629-6638, 2019. 8", + "[50] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Proceedings of the International Conference on Neural Information Processing Systems, 2021. 8" + ], + "bbox": [ + 78, + 125, + 468, + 498 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6693", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_model.json b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e160a52d103e4793ca24a4e45558501712026fec --- /dev/null +++ b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_model.json @@ -0,0 +1,2026 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.792, + 0.087 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.162, + 0.86, + 0.205 + ], + "angle": 0, + "content": "Meta-Explore: Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.228, + 0.813, + 0.294 + ], + "angle": 0, + "content": "Minyoung Hwang\\(^{1}\\), Jaeyeon Jeong\\(^{1}\\), Minsoo Kim\\(^{3}\\), Yoonseon Oh\\(^{2}\\), Songhwai Oh\\(^{1}\\) \n\\(^{1}\\)Electrical and Computer Engineering and ASRI, Seoul National University \n\\(^{2}\\)Department of Electronic Engineering, Hanyang University \n\\(^{3}\\)Interdisciplinary Major in Artificial Intelligence, Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.297, + 0.885, + 0.31 + ], + "angle": 0, + "content": "{minyoung.hwang, jaeyeon.jeong}@rlllab.snu.ac.kr, goldbird5@snu.ac.kr, yoh21@hanyang.ac.kr, songhwai@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.341 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.474, + 0.742 + ], + "angle": 0, + "content": "The main challenge in vision-and-language navigation (VLN) is how to understand natural-language instructions in an unseen environment. The main limitation of conventional VLN algorithms is that if an action is mistaken, the agent fails to follow the instructions or explores unnecessary regions, leading the agent to an irrecoverable path. To tackle this problem, we propose Meta-Explore, a hierarchical navigation method deploying an exploitation policy to correct misled recent actions. We show that an exploitation policy, which moves the agent toward a well-chosen local goal among unvisited but observable states, outperforms a method which moves the agent to a previously visited state. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. The key to our approach is understanding the object placements around the agent in spectral-domain. Specifically, we present a novel visual representation, called scene object spectrum (SOS), which performs category-wise 2D Fourier transform of detected objects. Combining exploitation policy and SOS features, the agent can correct its path by choosing a promising local goal. We evaluate our method in three VLN benchmarks: R2R, SOON, and REVERIE. Meta-Explore outperforms other baselines and shows significant generalization performance. In addition, local goal search using the proposed spectral-domain SOS features significantly improves the success rate by \\(17.1\\%\\) and SPL by \\(20.6\\%\\) against the state-of-the-art method of the SOON benchmark. Project page: https://rllab-snu.github.io/projects/Meta-Explore/doc.html" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.75, + 0.21, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.77, + 0.47, + 0.799 + ], + "angle": 0, + "content": "Visual navigation in indoor environments has been studied widely and shown that an agent can navigate in unex-" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.327, + 0.871, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.541 + ], + "angle": 0, + "content": "Figure 1. Hierarchical Exploration. At each episode, a natural-language instruction is given to the agent to navigate to a goal location. The agent explores the environment and constructs a topological map by recording visited nodes \\(\\odot\\) and next step reachable nodes \\(\\clubsuit\\). Each node consists of the position of the agent and visual features. \\(o_t\\) denotes the observation at time \\(t\\). The agent chooses an unvisited local goal to solve the regretful exploration problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.893, + 0.7 + ], + "angle": 0, + "content": "plored environments [1]. By recognizing the visual context and constructing a map, an agent can explore the environment and solve tasks such as moving towards a goal or following a desired trajectory. With the increasing development in human language understanding, vision-and-language navigation (VLN) [2] has enabled robots to communicate with humans using natural languages. The high degree of freedom in natural language instructions allows VLN to expand to various tasks, including (1) following fine-grained step-by-step instructions [2-13] and (2) reaching a target location described by goal-oriented language instructions [14-20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.895, + 0.868 + ], + "angle": 0, + "content": "A challenging issue in VLN is the case when an action is mistaken with respect to the given language instruction [21-26]. For instance, if the agent is asked to turn right at the end of the hallway but turns left, the agent may end up in irrecoverable paths. Several existing studies solve this issue via hierarchical exploration, where the high-level planner decides when to explore and the low-level planner chooses what actions to take. If the high-level planner chooses to explore, the agent searches unexplored regions, and if it chooses to exploit, the agent executes the best action based on the previous exploration. Prior work [21-23] returns the agent to the last successful state and resumes exploration." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.808, + 0.47, + 0.868 + ], + "angle": 0, + "content": "This work was in part supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-01190, [SW Star Lab] Robot Learning: Efficient, Safe, and Socially-Acceptable Machine Learning, \\(80\\%\\), and No.2022-0-00907, Development of AI Bots Collaboration Platform and Self-Organizing AI, \\(20\\%\\)). (Corresponding authors: Yoonseon Oh and Songhwai Oh.)" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.899 + ], + "angle": 0, + "content": "6683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.469, + 0.306 + ], + "angle": 0, + "content": "However, such methods take a heuristic approach because the agent only backtracks to a recently visited location. The agent does not take advantage of the constructed map and instead naively uses its recent trajectory for backtracking. Another recent work [26] suggests graph-based exploitation, which uses a topological map to expand the action space in global planning. Still, this method assumes that the agent can directly jump to a previously visited node. Since this method can perform a jump action at every timestep, there is no trigger that explicitly decides when to explore and when to exploit. Therefore, we address the importance of time scheduling for exploration-exploitation and efficient global planning using a topological map to avoid reexploring visited regions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.306, + 0.471, + 0.816 + ], + "angle": 0, + "content": "We expand the notion of hierarchical exploration by proposing Meta-Explore, which not only allows the high-level planner to choose when to correct misled local movements but also finds an unvisited state inferred to be close to the global goal. We illustrate the overview of hierarchical exploration in Figure 1. Instead of backtracking, we present an exploitation method called local goal search. We show that it is more efficient to plan a path to a local goal, which is the most promising node from the unvisited but reachable nodes. We illustrate the difference between conventional backtracking and local goal search in Figure 2. Based on our method, we show that exploration and exploitation are not independent and can complement each other: (1) to overtake regretful explorations, the agent can perform exploitation and (2) the agent can utilize the constructed topological map for local goal search. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. Most VLN tasks require a level of understanding objects nearby the agent, but previous studies simply encode observed panoramic or object images [2, 3, 16-18, 21-35]. In this paper, we present a novel semantic representation of the scene called scene object spectrum (SOS), which is a matrix containing the arrangements and frequencies of objects from the visual observation at each location. Using SOS features, we can sufficiently estimate the context of the environment. We show that the proposed spectral-domain SOS features manifest better linguistic interpretability than conventional spatial-domain visual features. Combining exploitation policy and SOS features, we design a navigation score that measures the alignment between a given language instruction and a corrected trajectory toward a local goal. The agent compares local goal candidates and selects a near-optimal candidate with the highest navigation score from corrected trajectories. This involves high-level reasoning related to the landmarks (e.g., bedroom and kitchen) and objects (e.g., table and window) that appear in the instructions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.817, + 0.438, + 0.83 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.839, + 0.47, + 0.867 + ], + "angle": 0, + "content": "- We propose a hierarchical navigation method called Meta-Explore, deploying an exploitation policy to cor" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.123, + 0.877, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.294, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Local Goal Search for Exploitation. The local goal is likely to be chosen as the closest node to the global goal. Existing methods only backtrack to a visited node (left). We expand the searchable area by including unvisited but reachable nodes (right)." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.349, + 0.892, + 0.391 + ], + "angle": 0, + "content": "rect misled recent actions. The agent searches for an appropriate local goal instead of reversing the recent action sequence." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.394, + 0.892, + 0.489 + ], + "angle": 0, + "content": "- In the exploitation mode, the agent uses a novel scene representation called scene object spectrum (SOS), which contains the spectral information of the object placements in the scene. SOS features provide semantically meaningful clues to choose a near-optimal local goal and help the agent to solve the regretful exploration problem." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.493, + 0.892, + 0.589 + ], + "angle": 0, + "content": "- We evaluate our method on three VLN benchmarks: R2R [2], SOON [16], and REVERIE [17]. The experimental results show that the proposed method, MetaExplore, improves the success rate and SPL in test splits of R2R, SOON and val split of REVERIE. The proposed method shows better generalization results compared to all baselines." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.394, + 0.892, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.642, + 0.61 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.614, + 0.788, + 0.629 + ], + "angle": 0, + "content": "2.1. Vision-and-Language Navigation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.867 + ], + "angle": 0, + "content": "In VLN, an agent encodes the natural language instructions and follows the instructions, which can be either (1) a fine-grained step-by-step instruction the agent can follow [2-4], (2) a description of the target object and location [16, 17], or (3) additional guidance given to the agent [18, 27]. These tasks require the agent to recognize its current location using some words in the natural-language instructions. Prior work [2, 28-31] show that an agent can align visual features to language instructions via neural networks and use the multimodal output embeddings to generate a suitable action at each timestep. Most VLN methods utilize cross-modal attention, either with recurrent neural networks [2, 28] or with transformer-based architectures [29-31]. For sequential action prediction, Hong et al. [32] further use recurrent units inside transformer architectures, while Pashevich et al. [33] and Chen et al. [34] use additional transformers to embed past observations and actions." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "6684" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.113, + 0.308, + 0.128 + ], + "angle": 0, + "content": "2.2. Exploration-Exploitation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.13, + 0.47, + 0.309 + ], + "angle": 0, + "content": "In an unseen environment, the agent must maximize the return without knowing the true value functions. One of the solutions to this problem is to switch back and forth between exploration and exploitation [36]. In the exploration mode, the agent gathers more information about the environment. On the other hand, the agent uses information collected during exploration and chooses the best action for exploitation. Ecoffet et al. [37] reduced the exploration step by archiving the states and exploring again from the successful states. Pislar et al. [38] addressed the various scheduling policies and demonstrated their method on Atari games. Recent work [39, 40] successfully demonstrates the effectiveness of hierarchical exploration in image-goal navigation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Like commonly used greedy navigation policies, VLN tasks also deal with the problem of maximizing the chance to reach the goal without knowing the ground truth map. Several VLN methods employ the concept of exploitation to tackle this problem. Ke et al. [35] look forward to several possible future trajectories and decide whether to backtrack or not and where to backtrack. Others [21-23] estimate the progress to tell whether the agent becomes lost and make the agent backtrack to a previously visited location to restart exploration. However, previous studies do not take into account what should be done in the exploitation mode. In order to handle this problem, we propose a hierarchical navigation method which determines the scheduling between exploration and exploitation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.506, + 0.291, + 0.52 + ], + "angle": 0, + "content": "2.3. Visual Representations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Popular visual encoding methods via ResNet [41] and ViT [42] can be trained to learn rotation-invariant visual features. Both methods learn to extract visual features with high information gain for global and local spatial information. The high complexity of the features leads to low interpretability of the scene and therefore requires the agent to use additional neural networks or complex processing to utilize them. On the other hand, traditional visual representation methods such as Fourier transform use spectral analysis, which is highly interpretable and computationally efficient. One drawback of the traditional methods is that they fail to maximize the information gain. Nonetheless, an appropriate use of essential information can be helpful for high-level decision making and enables more straightforward interpretation and prediction of the visual features. One traditional navigation method, Sturz et al. [43] used Fourier transform to generate rotation-invariant visual features. However, no research has transformed the spectral information of the detected objects to represent high-level semantics from visual observations. Focusing on the fact that 2D Fourier transform can extract morphological properties of images [44], we can find out the shape or structure of detected objects through 2D Fourier transform. In this paper, we decompose the object mask into binary masks by object categories and perform a 2D Fourier transform on each binary mask." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.125, + 0.591, + 0.139 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.144, + 0.702, + 0.158 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.162, + 0.892, + 0.44 + ], + "angle": 0, + "content": "We deal with VLN in discrete environments, where the environment is given as an undirected graph \\( G_{e} = \\{V,E\\} \\). \\( V \\) denotes a set of \\( N \\) navigable nodes, \\( \\{V_{i}\\}_{i = 1}^{N} \\), and \\( E \\) is the adjacency matrix describing connectivity among the nodes in \\( V \\). We denote the observation at node \\( V_{i} \\) as \\( O_{i} \\). The agent uses a panoramic RGB image observation \\( o_{t} \\) and current node \\( v_{t} \\), which are collected at time \\( t \\). The agent either moves to a neighboring node or executes a stop action. \\( a_{t} \\) denotes the action at time \\( t \\). The objectives of VLN are categorized as follows: (1) to follow language instructions [2] and (2) to find a target object described by language instructions in a fixed time \\( T \\) [16, 17]. We present a general hierarchical exploration method that can be applied to both tasks. We also enhance the navigation policy by extracting cross-domain visual representations from the environments, i.e., spatial-domain and spectral-domain representations. To balance the information loss and interpretability of the visual feature, we adopt multi-channel fast Fourier transform (FFT) to encode semantic masks of the detected objects into category-wise spectral-domain features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.644, + 0.459 + ], + "angle": 0, + "content": "3.2. Meta-Explore" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We design a learnable hierarchical exploration method for VLN called Meta-Explore, which decides (1) when to explore or exploit and (2) a new imagined local goal to seek during exploitation. The overall network architecture of the proposed Meta-Explore is shown in Figure 3. Given a language instruction \\( L \\), the agent navigates in the environment until it finds the target described in \\( L \\). Meta-Explore consists of a mode selector and two navigation modules corresponding to two modes: exploration and exploitation. At each timestep, the mode selector chooses to explore or exploit. At \\( t = 0 \\), the mode is initialized to exploration. In the exploration mode, the agent outputs an action toward a neighboring node to move the agent toward the goal. When the mode selector recognizes that the agent is not following the instruction successfully, the mode is switched to exploitation. In the exploitation mode, the agent seeks a new local goal with the highest correspondence against the language instructions from the previously unvisited candidate nodes using spectral-domain visual features. The agent moves toward the local goal by planning a path. After the agent arrives at the local goal, the mode is reset to exploration. The explore-exploit switching decision occurs through the mode selector by estimating the probability to explore. The agent repeats this explore-exploit behavior until it determines that the target is found and decides to stop." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.81, + 0.655, + 0.823 + ], + "angle": 0, + "content": "3.2.1 Mode Selector" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.868 + ], + "angle": 0, + "content": "At time \\( t \\), the agent observes visual features about the current node \\( v_{t} \\) and several reachable nodes. We call the nodes reachable at the current timestep as candidate nodes. \\( n_{cand} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.899 + ], + "angle": 0, + "content": "6685" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.125, + 0.812, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.354, + 0.895, + 0.403 + ], + "angle": 0, + "content": "Figure 3. Network Architecture. Three types of visual features: panoramic (yellow), object image (aquamarine), and object spectrum (red) are encoded. The color in each parenthesis denotes the color describing the corresponding feature. The cross-modal transformer encodes language and spatial visual features as hidden state \\( H_{t} \\). A mode selector gives explore or exploit command to the agent by predicting the explore probability \\( P_{\\text{explore}} \\). The selected navigation module outputs an action \\( a_{t} \\) from the possible \\( n_{\\text{cand}} \\) candidate nodes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.471, + 0.56 + ], + "angle": 0, + "content": "denotes the number of candidate nodes. We use a cross-modal transformer with \\( n_{L} \\) layers to relate visual observations to language instructions. The cross-modal transformer takes the visual features of nodes in the constructed topological map at time \\( t \\), \\( G_{t} \\), and outputs cross-modal embedding \\( H_{t} \\) to encode visual observations with \\( L \\). We concatenate location encoding and history encoding [24] to the visual features as node features to consider the relative pose from \\( v_{t} \\) and the last visited timestep of each node, respectively. Each word is encoded via a pretrained language encoder [45], which is used for general vision-language tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.562, + 0.47, + 0.77 + ], + "angle": 0, + "content": "The cross-modal transformer consists of cross-attention layer L2V_Atn(\\hat{W},\\hat{V}) = Softmax(\\hat{W}\\Theta_q^\\dagger (\\hat{V}\\Theta_k^\\dagger)^T /\\sqrt{d})\\hat{V}\\Theta_v^\\dagger and self-attention layer SelfAttn(X) = Softmax \\((X\\Theta_{q}(X\\Theta_{k})^{T}+\\) \\(A\\Theta_e + b_e) / \\sqrt{d})X\\Theta_v\\) , where \\(\\hat{W},\\hat{V},X\\) , and \\(A\\) denote word, visual, node representations and adjacency matrix of \\(G_{t}\\) respectively. The (query, key, value) weight matrices of self-attention and cross-attention layers are denoted as \\((\\Theta_q,\\Theta_k,\\Theta_v)\\) and \\((\\Theta_q^\\dagger ,\\Theta_k^\\dagger ,\\Theta_v^\\dagger)\\) , respectively. The final crossmodal embedding at time \\(t\\) after passing through \\(n_L\\) transformer layers is denoted as \\(H_{t}\\) . To encourage the monotonic increasing relationship between language and visual attentions at each timestep, we define a correlation loss \\(L_{corr} = \\sum_{t = 1}^{T}\\| \\mathrm{L2V\\_Attn} - I_{n_x}\\| _1\\) for training the crossmodal transformer, where \\(n_x\\) denotes the dimension of the \\(H_{t}\\) and \\(I_{n_x}\\) denotes an identity matrix of size \\(n_x\\times n_x\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.77, + 0.471, + 0.868 + ], + "angle": 0, + "content": "As illustrated in Figure 4, the mode selector estimates the probability to explore \\( P_{\\text{explore}} \\) given the cross-modal hidden state \\( H_t \\). We denote the mode selector as \\( S_{\\text{mode}} \\) and use a two-layer feed-forward neural network. Given \\( H_t \\), \\( S_{\\text{mode}} \\) outputs the exploration probability as \\( P_{\\text{explore}} = 1 - S_{\\text{mode}}(H_t) \\). If \\( P_{\\text{explore}} \\geq 0.5 \\), the exploration policy outputs a probability distribution for reachable nodes at the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.533 + ], + "angle": 0, + "content": "next step. At time \\( t + 1 \\), the agent moves to the node with the highest probability. If \\( P_{\\text{explore}} < 0.5 \\), the agent determines that the current trajectory is regretful, so the agent should traverse to find a local goal, which is the most likely to be the closest node to the global goal. The exploitation policy mainly utilizes object-level features to search for the local goal with high-level reasoning. After the local goal is chosen, the path planning module outputs an action following the shortest path to the local goal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.893, + 0.7 + ], + "angle": 0, + "content": "To train the mode selector, we require additional demonstration data other than the ground truth trajectory, such that it switches between exploration and exploitation. We generate the demonstration data from the ground truth trajectories, with additional detours. For the detours, we stochastically select candidate nodes other than the ground truth paths and add the trajectory that returns to the current viewpoint. The imitation learning loss for training the mode selector is defined as \\( L_{mode} = \\sum_{t=1}^{T} \\mathbb{1}(m_t = \\mathrm{gt}_t) \\), where \\( m_t \\) is the mode of the agent, 0 for exploitation and 1 for exploration. \\( \\mathrm{gt}_t \\) is 1 if the current node is in the shortest ground truth trajectory and \\( \\mathrm{gt}_t = 0 \\), otherwise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.695, + 0.722 + ], + "angle": 0, + "content": "3.2.2 Exploration Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.892, + 0.822 + ], + "angle": 0, + "content": "In the exploration mode, the agent follows the following sequential operations: topological map construction, self-monitoring, and an exploration policy. To improve the exploration, we adopt self-monitoring [21] to predict the current progress of exploration to enhance the exploration policy itself. Prior work [21, 22] has shown that auxiliary loss using self-monitoring can regularize the exploration policy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.867 + ], + "angle": 0, + "content": "Topological Map Construction. The agent constructs graph \\( G_{t} \\) by classifying nodes into two types: (1) visited nodes and (2) unvisited but observable nodes. At current" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.502, + 0.898 + ], + "angle": 0, + "content": "6686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.123, + 0.258, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.201, + 0.255, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.3, + 0.234, + 0.31 + ], + "angle": 0, + "content": "Exploration Module" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.123, + 0.45, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.3, + 0.406, + 0.31 + ], + "angle": 0, + "content": "Exploitation Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.47, + 0.358 + ], + "angle": 0, + "content": "Figure 4. Navigation Modules. Mode selector estimates \\( P_{\\text{explore}} \\), i.e., the probability to explore, and chooses between exploration and exploitation modules. The selected navigation module outputs the next action \\( a_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.469, + 0.501 + ], + "angle": 0, + "content": "time \\( t \\), the agent at node \\( v_{t} \\in \\{V_{i}\\}_{i=1}^{N} \\) observes \\( N(v_{t}) \\) neighbor nodes as next step candidates at time \\( t + 1 \\). The visited nodes consist of visual features of their own and the neighboring nodes from panoramic RGB observations. The unvisited nodes can be observed only if they are connected to at least one visited node. The topological map records the positions and visual features of observed nodes at each timestep. By knowing the positions of nodes in \\( G_{t} \\), the agent can plan the shortest path trajectory between two nodes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.691 + ], + "angle": 0, + "content": "Self-Monitoring. We use a progress monitor to estimate the current navigation progress at each episode. Self-monitoring via estimating current progress helps the agent choose the next action that can increase the progress. The estimated progress \\(\\hat{p}_t = F_{\\text{progress}}(H_t)\\) is the output of a feed-forward neural network, given \\(H_t\\) as input. We measure the ground truth progress \\(p_t\\) as the ratio between the current distance to the goal and the shortest path length of the episode subtracted from 1, described as \\(1 - \\frac{d_{geo}(v_t, v_{goal})}{d_{geo}(v_0, v_{goal})}\\), where \\(d_{geo}(a, b)\\) is the geodesic distance between \\(a\\) and \\(b\\). \\(v_0, v_t\\), and \\(v_{goal}\\) denote initial, current, and goal positions, respectively. We add progress loss \\(L_{\\text{progress}} = \\sum_{t=1}^{T} (\\hat{p}_t - p_t)^2\\) to train the progress monitor while training the exploration policy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.471, + 0.877 + ], + "angle": 0, + "content": "Exploration Policy. The exploration policy \\( F_{explore} \\) estimates the probability of moving to the candidate nodes at the next step. The agent chooses the action \\( a_t \\) at time \\( t \\) based on the estimated probability distribution among candidate nodes, described as \\( a_t = \\arg \\max_{V_i} (F_{explore}([H_t]_i)) \\). \\( F_{explore} \\) is implemented via a two-layer feed-forward network with the cross-modal hidden state \\( H_t \\) given as input. The output of \\( F_{explore} \\) becomes a probability distribution over possible actions. To only consider unvisited nodes, we mask out the output for visited nodes. For training, we sample the next action from the probability distribution instead of choosing a node with the highest probability. We describe the training details in Section 3.3." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.123, + 0.882, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.322 + ], + "angle": 0, + "content": "Figure 5. Scene Object Spectrum (SOS). The agent calculates scene object spectrum (SOS) features for efficient exploitation. SOS features incorporate semantic information observed in a single panoramic image by performing category-wise 2D FFT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.333, + 0.698, + 0.347 + ], + "angle": 0, + "content": "3.2.3 Exploitation Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.489 + ], + "angle": 0, + "content": "In the exploitation mode, the agent requires high-level reasoning with identifiable environmental clues to imagine regretful exploration cases. To find clues in an object-level manner, we present a novel visual representation by capturing object information in the spectral-domain. The novel representation is more easily predictable than spatial features such as RGB image embeddings. The agent can take advantage of the predictability by expanding the searchable area to find a local goal. We choose the local goal as the closest node to the global goal in the feature space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Spectral-Domain Visual Representations. Common navigation policies can lead the agent toward the node with the highest similarity to the target. However, even with a good learned policy, the agent can act in a novice manner in unseen environments. In this paper, we seek extra information from the environment for generalizable high-level reasoning to resolve the issue. As illustrated in Figure 5, scene object spectrum (SOS) incorporates semantic information observed in a single panoramic image by generating a semantic mask for each object category and applying Fourier transform to each semantic mask. The semantic mask for object class \\( k \\) at time \\( t \\) is calculated as a binary mask \\( [m_t^k]_{ij} \\) that detects the object at pixel \\( (i,j) \\). Suppose there are a total of \\( K \\) object categories. When multiple objects are detected for one object category, the binary mask appears as a union of the bounding boxes of the detected objects. We define FFT as a channel-wise 2D fast Fourier transform that receives \\( K \\) binary semantic masks and outputs \\( K \\) spectral-domain features, where \\( K \\) is the number of object classes. Then, SOS feature \\( \\vec{S}_t = [s_t^1,\\dots,s_t^K ]^T \\) can be defined as \\( s_t^k = \\log |\\mathbf{FFT}(m_t^k)| \\). For simplicity, we perform mean pooling on the vertical spectral axis and normalize the output. The final SOS feature has shape \\( (K,\\eta) \\), where \\( \\eta \\) is the maximum horizontal frequency." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Local Goal Search Using Semantic Clues. We argue that returning to a previously visited node does not guarantee the agent escapes from the local optima. Instead of backtracking" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "6687" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.47, + 0.25 + ], + "angle": 0, + "content": "to a previously visited node, the agent searches for a local goal to move towards. If the agent plans a path and moves towards the local goal, the agent does not need to repeat unnecessary actions in visited regions after the exploitation ends. Additionally, searching for a local goal takes full advantage of the topological map by utilizing the connections among the observed nodes. To expand the searchable area further, we let the agent choose the local goal from previously unvisited and unchosen candidate nodes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.471, + 0.52 + ], + "angle": 0, + "content": "To choose a local goal, we first score the corrected trajectories to measure the alignment with the language instruction \\( L \\). We use SOS features as semantic environmental clues to estimate the navigation score \\( S_{\\text{nav}} \\) of the corrected trajectory, which is the shortest path trajectory from the initial node to the local goal in the constructed topological map. To simplify, we convert the language instruction into a list of objects \\( W^{o} = [w_{1}^{o}, \\dots, w_{B}^{o}] \\) consisting of \\( B[\\leq K] \\) object categories (e.g., desk, cabinet, and microwave). We approximate the corresponding reference SOS features as \\( [\\hat{\\delta}(w_{1}^{o}), \\dots, \\hat{\\delta}(w_{B}^{o})] \\) where the \\( i^{th} \\) row of \\( \\hat{\\delta}(w_{k}^{o}) \\) is defined as \\( [\\hat{\\delta}(w_{k}^{o})]_{i} = \\mathbb{1}(k = i)\\lambda(\\delta(w_{k}^{o}))\\mathrm{sinc}\\left(\\frac{j}{2} - \\frac{\\eta}{4}\\right) \\). \\( \\lambda(\\delta(w_{k}^{o})) \\) denotes the average width of detected bounding boxes of object \\( w_{k}^{o} \\) in the environment. A detailed approximation process is explained in the supplementary material. To simulate a corrected trajectory \\( \\mathcal{T}' = (v_{1}', \\dots, v_{t'}) \\), we calculate the SOS features \\( [\\vec{S}_1', \\dots, \\vec{S}_{t'}'] \\) corresponding to the nodes in \\( \\mathcal{T}' \\). We measure the similarity between two object spectrum features via the cosine similarity of the flattened vectors. Finally, the navigation score \\( S_{\\text{nav}} \\) of \\( \\mathcal{T}' \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.525, + 0.469, + 0.595 + ], + "angle": 0, + "content": "\\[\nS _ {n a v} \\left(\\mathcal {T} ^ {\\prime}\\right) = \\frac {\\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {t ^ {\\prime}} \\left(\\frac {\\hat {\\delta} \\left(w _ {i} ^ {o}\\right)}{\\left| \\hat {\\delta} \\left(w _ {i} ^ {o}\\right) \\right|} \\cdot \\frac {\\vec {S} _ {j} ^ {\\prime}}{\\left| \\vec {S} _ {j} ^ {\\prime} \\right|}\\right) \\left(\\left(\\hat {\\delta} \\left(w _ {i} ^ {o}\\right) - \\hat {\\delta} (\\bar {w} ^ {o})\\right) \\cdot \\left(\\vec {S} _ {j} ^ {\\prime} - \\bar {\\vec {S}} ^ {\\prime}\\right)\\right)}{\\sqrt {\\frac {t ^ {\\prime}}{B} \\cdot \\sum_ {i = 1} ^ {B} \\left(\\hat {\\delta} \\left(w _ {i} ^ {o}\\right) - \\hat {\\delta} (\\bar {w} ^ {o})\\right) ^ {2} \\sum_ {j = 1} ^ {t ^ {\\prime}} \\left(\\vec {S} _ {j} ^ {\\prime} - \\bar {\\vec {S}} ^ {\\prime}\\right) ^ {2}}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.469, + 0.678 + ], + "angle": 0, + "content": "where \\(\\hat{\\delta} (\\overline{w}^o)\\) and \\(\\vec{S}'\\) denote the average values of SOS features \\(\\hat{\\delta} (w_i^o)\\) and \\(\\vec{S}_j^\\prime\\), respectively. This equation can also be interpreted as a pseudo correlation-coefficient function between object list \\(W^{o}\\) and trajectory \\(\\mathcal{T}'\\). The exploitation policy selects the node with the highest navigation score as the local goal from the previously unvisited candidates." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.679, + 0.469, + 0.789 + ], + "angle": 0, + "content": "Figure 6 illustrates a simple scenario of entering a room. Suppose \\( W^{o} = \\) [sculpture, door, bed] and the agent has to compare two trajectories \\( \\mathcal{T}_1 = (v_1,v_2,A) \\) and \\( \\mathcal{T}_1 = (v_1,v_2,B) \\). Each similarity matrix in Figure 6 has the \\( (t,j) \\) element as the similarity between the SOS feature of \\( V_{t} \\) and \\( \\hat{\\delta}(w_j^o) \\), which is calculated as \\( \\hat{\\delta}(w_j^o) \\cdot \\vec{S}_t' \\). Notably, the similarity matrix shows monotonic alignment and the navigation score is higher when the next action is chosen correctly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.795, + 0.238, + 0.809 + ], + "angle": 0, + "content": "3.3. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.469, + 0.867 + ], + "angle": 0, + "content": "We use [24] for pretraining the visual encoder with panoramic RGB observations. We use the DAgger algorithm [46] to pretrain the navigation policy and the mode selector. To prevent overfitting, we iteratively perform" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.123, + 0.77, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.785, + 0.124, + 0.889, + 0.137 + ], + "angle": 0, + "content": "similarity matrix" + }, + { + "type": "table", + "bbox": [ + 0.785, + 0.137, + 0.892, + 0.281 + ], + "angle": 0, + "content": "
w01w2w3
v1=10.90.00.0
v1=20.31.00.0
v1=30.00.01.0
v1=4, Snav = 0.296
w01w2w3
v1=10.90.00.0
v1=20.31.00.0
v1=30.50.40.0
v1=4, Snav = 0.296
" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Figure 6. Toy Example. Monotonic alignment between language instruction and visual observation is desirable. Yellow dots \\(\\bullet\\) in the nodes describe the ground truth trajectory. Based on the node at \\(t = 3\\), the similarity matrix can show either monotonic or non-monotonic alignment between object tokens and SOS features. The green circles \\(\\clubsuit\\) describe the possible candidates \\(A, B\\) for next action." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.587 + ], + "angle": 0, + "content": "teacher forcing and student forcing to choose the action from the exploration policy. Imitation learning loss is calculated as \\( L_{IL} = \\sum_{t=1}^{T} -\\log p(a_t^* | a_t) \\) and object grounding loss is calculated as \\( L_{OG} = -\\log p(\\mathrm{obj}^* | \\mathrm{obj}_{pred}) \\), where \\( \\mathrm{obj}^* \\) denotes the ground truth and \\( \\mathrm{obj}_{pred} \\) denotes the predicted object location. The total loss function is defined as \\( L_{total} = L_{mode} + L_{progress} + L_{corr} + L_{IL} + L_{OG} \\). We further finetune the agent via A2C [47]. The exploration policy selects the action \\( a_t \\) with probability \\( p_t^a \\). Reinforcement learning loss is defined as \\( L_{RL} = -\\sum_t a_t^s \\log(p_t^a) A_t - \\lambda \\sum_t a_t^* \\log(p_t^a) \\). To train the mode selector, progress monitor, and exploration policy in an end-to-end manner, we use the total loss function as \\( L_{fine} = L_{mode} + L_{progress} + L_{RL} \\). The exploitation policy searches the path toward the local goal from the constructed navigation graph. Thus, the exploitation policy is not learned." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.595, + 0.727, + 0.61 + ], + "angle": 0, + "content": "4. Navigation Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.693, + 0.627 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.891, + 0.713 + ], + "angle": 0, + "content": "We evaluate our method on three VLN benchmarks, Room-to-Room(R2R) [2], SOON [16], and REVERIE [17]. R2R evaluates the visually-grounded natural navigation performance of the agent. The agent must navigate to the predefined goal point given image observations and language instructions in an unseen environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.769 + ], + "angle": 0, + "content": "SOON is also a goal-oriented VLN benchmark. Natural language instructions in SOON have an average length of 47 words. The agent should locate the target location and detect the location of an object to find the target object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.892, + 0.867 + ], + "angle": 0, + "content": "REVERIE is a goal-oriented VLN benchmark that provides natural language instruction about target locations and objects. In REVERIE, the agent is given an instruction referring to a remote object with an average length of 21 words. With this instruction and a panoramic observation from the environment, the agent should navigate to the location the instruction describes and find the correct object" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "6688" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.127, + 0.468, + 0.14 + ], + "angle": 0, + "content": "bounding box among the predefined object bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.146, + 0.262, + 0.159 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.163, + 0.3, + 0.176 + ], + "angle": 0, + "content": "4.2.1 Navigation performance" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.179, + 0.469, + 0.346 + ], + "angle": 0, + "content": "We evaluate algorithms using the trajectory length (TL), success rate (SR), and success weighted by inverse path length (SPL) [48], and oracle success rate (OSR) for the navigation performance comparison. An episode is recorded as a success if the agent takes a stop action within \\(3\\mathrm{m}\\) of the target location. TL is the average path length in meters. SR is denoted as the number of successes divided by the total number of episodes, \\(M\\). SPL is calculated as \\(\\frac{1}{M}\\sum_{i = 1}^{M}S_{i}\\frac{l_{i}}{\\max(p_{i},l_{i})}\\), where \\(S_{i}\\) denotes the success as a binary value. \\(p_i\\) and \\(l_{i}\\) denote the shortest path and actual path lengths for the \\(i^{th}\\) episode. OSR uses the oracle stop policy instead of the stop policy of the agent." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.35, + 0.348, + 0.364 + ], + "angle": 0, + "content": "4.2.2 Object grounding performance" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.469, + 0.491 + ], + "angle": 0, + "content": "We also evaluate the object grounding performance of the agent by the success rate of finding the target object (FSR) and the target finding success weighted by inverse path length (FSPL)1 [16, 17]. FSPL is calculated as FSPL = \\(\\frac{1}{N}\\sum_{i=1}^{N}S_{i}^{nav}S_{i}^{loc} \\cdot l_{i}^{nav}/\\max(l_{i}^{nav},l_{i}^{gt})\\), where \\(S_{i}^{nav}\\) is whether the agent navigates to the target, \\(S_{i}^{loc}\\) is whether the agent finds a target object bounding box, and \\(l_{i}^{nav}\\) and \\(l_{i}^{gt}\\) are the navigation trajectory length and ground truth trajectory length, respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.495, + 0.402, + 0.51 + ], + "angle": 0, + "content": "4.3. Baselines and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.512, + 0.469, + 0.65 + ], + "angle": 0, + "content": "We compare our method with several other baselines as follows. For each task, we compare our method with a number of baselines that use various types of memory (recurrent, sequential, and topological map). For methods implemented with a hierarchical navigation framework, we compare the specific exploitation methods: homing, jump, and local goal search. Homing makes the agent backtrack, and jump makes the agent jump to a previously visited node. The hyperparameters and detailed model architecture of Meta-Explore are described in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.653, + 0.409, + 0.667 + ], + "angle": 0, + "content": "4.4. Comparison with Navigation Baselines" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.669, + 0.469, + 0.724 + ], + "angle": 0, + "content": "We compare our method with navigation baselines. We focus on the success rate and SPL. Rendered results and detailed analyses with other evaluation metrics are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.725, + 0.468, + 0.835 + ], + "angle": 0, + "content": "R2R. Table 1 compares the proposed Meta-Explore with baselines for the R2R navigation task. We categorize the baseline methods based on the type of constructed memory and the type of exploitation. Our method outperforms other exploration-only baselines over all types of validation and test splits in success rate and SPL. Compared with hierarchical baselines SMNA [21], Regretful-Agent [22], FAST [35], and SSM [26], Meta-Explore improves success" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.127, + 0.892, + 0.332 + ], + "angle": 0, + "content": "rate and SPL by at least \\(16.4\\%\\) and \\(8.9\\%\\), respectively. The main difference is that Meta-Explore constructs a topological map during exploration and uses the map for local goal search in exploitation. On the contrary, homing exploitation policies in SMNA, Regretful-Agent, and FAST only rely on the current trajectory, instead of taking advantage of the constructed memory. Jump exploitation in SSM uses a topological map to search a successful previous node, but it makes an unrealistic assumption that the agent can directly jump to a previously visited distant node and unfairly saves time. In our approach, we plan a path to the local goal based on the topological map. The experiment results reveal that even if we design a hierarchical navigation framework, exploration and exploitation are not entirely separate but they can complement each other." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.334, + 0.892, + 0.692 + ], + "angle": 0, + "content": "SOON, REVERIE. Table 2 compares Meta-Explore with baselines in the SOON navigation task. While the proposed method does not improve performance in val seen split, Meta-Explore outperforms other baselines in the test unseen split of SOON for success rate by \\(17.1\\%\\) and SPL by \\(20.6\\%\\). The result implies that for the goal-oriented VLN task, high performance in train or val seen splits can be the overfitted result. Because the agent can be easily overfitted to the training data, making a generalizable model or providing a deterministic error-correction module for inference is essential. Meta-Explore chooses the latter approach by correcting the trajectory via exploitation in regretful cases. The evaluation results in the REVERIE navigation task are described in the supplementary material. Meta-Explore shows improvement in the val split of REVERIE for success rate and SPL, but the improvement in the test split is lower than the results in R2R and SOON. We found 252 meaningless object categories (e.g., verbs, adjectives, and prepositions) and 418 replaceable object categories (e.g., typographical errors and synonyms) in the REVERIE3 dataset. Because our exploitation method utilizes object-based parsing of the given instruction to match with the detected object categories, the effectiveness of the proposed method is lessened due to inaccuracies and inconsistencies in the dataset. We expect to have higher performance if the mistakes in the dataset are fixed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.831, + 0.71 + ], + "angle": 0, + "content": "4.5. Local Goal Search using SOS Features" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.712, + 0.891, + 0.823 + ], + "angle": 0, + "content": "To discuss the significance of modeling exploitation policy, we conduct specific experiments about choosing the local goal for R2R and SOON. We evaluate our method using different types of local goal search, as shown in Table 3 and 4. Oracle denotes a method which selects a local goal using the ground truth trajectory. The performance of the oracle provides the achievable performance for each dataset. The results imply that local goal search using either" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.842, + 0.455, + 0.854 + ], + "angle": 0, + "content": "1Identical with its original term, Remote Grounding Success (RGS)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.855, + 0.264, + 0.866 + ], + "angle": 0, + "content": "\\(2^{\\dagger}\\) indicates reproduced results." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.842, + 0.455, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.832, + 0.891, + 0.866 + ], + "angle": 0, + "content": "\\(310.7\\%\\) and \\(41.2\\%\\) of a total of 46,476 words in the bounding box dataset correspond to meaningless and replaceable object categories, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "6689" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.125, + 0.894, + 0.272 + ], + "angle": 0, + "content": "
MethodsMemoryExploitVal SeenVal UnseenTest Unseen
SR↑SPL↑TL↓NE↓SR↑SPL↑TL↓NE↓SR↑SPL↑TL↓NE↓
Random--16-9.589.4516-9.779.2313129.899.79
Human----------11.851.618676
Seq2Seq [2]RecX6.03911.33-22-8.397.8420188.137.85
VLN\\(\\circ\\)BERT [32]RecX726811.132.90635712.013.93635712.354.09
SMNA†[21]Rechoming696311.693.31474112.615.486156-4.48
Regretful-Agent [22]Rechoming6963-3.235041-5.324840-5.69
FAST (short) [35]Rechoming----564321.174.97544122.085.14
FAST (long) [35]Rechoming7004188.063.136302224.424.036103196.534.29
HAMT-e2e [34]SeqX767211.152.51666111.462.29656012.273.93
DUET [24]Top. MapX797312.322.28726013.943.31695914.733.65
SSM [26]Top. Mapjump716214.73.10624520.74.32614620.44.57
Meta-Explore (Ours)Top. Maplocal goal817511.952.11726213.093.22716114.253.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.277, + 0.877, + 0.3 + ], + "angle": 0, + "content": "Table 1. Comparison and evaluation results of the baselines and our model in the R2R Navigation Task. Gray shaded rows describe hierarchical navigation baselines. Three memory types: Rec(recurrent), Seq(sequential), and Top. Map(topological map)" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.305, + 0.893, + 0.433 + ], + "angle": 0, + "content": "
MethodsMemoryExploitVal Seen InstructionVal Seen HouseTest Unseen House
SR↑SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑
Human----------90.459.291.451.1
RandomRecX0.01.50.11.40.10.00.40.92.10.42.70.0
Speaker-Follower [28]RecX97.997.797.824.561.260.469.49.17.06.19.80.6
RCM [49]RecX84.082.689.110.962.460.972.77.87.46.212.40.7
AuxRN [23]RecX98.497.498.713.768.867.378.58.38.16.711.00.5
GBE w/o GETop. MapX89.588.391.824.262.560.873.06.711.48.718.80.8
GBE [16]Top. MapX98.497.998.644.276.362.564.17.311.910.219.51.4
\\( GBE^† \\)Top. MapX----19.513.328.51.212.99.221.50.5
DUET [24]Top. MapX94.091.690.031.136.322.650.93.833.421.443.04.2
Meta-Explore (Ours)Top. Maplocal goal100.099.196.033.944.734.852.78.939.125.848.74.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.182, + 0.438, + 0.785, + 0.451 + ], + "angle": 0, + "content": "Table 2. Comparison and evaluation results of the baselines and our model in the SOON Navigation Task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.695 + ], + "angle": 0, + "content": "spatial or spectral visual representations is more effective than random local goal search. The results show that local goal search using spectral visual representations, i.e., SOS features, lead the agent to desirable nodes the most. We also compare local goal search with homing and the difference between the performance of the two methods is most noticeable in the test split of the SOON navigation task. As shown in Table 4, choosing the local goal with only spatial-domain features, the navigation performance does not improve compared to homing. On the contrary, spectral-domain local goal search shows significant improvement against homing by \\(10.4\\%\\) in success rate, \\(34.5\\%\\) on SPL, and \\(27.4\\%\\) on FSPL. The results imply that using spectral-domain SOS features helps high-level decision making, thereby enhancing the navigation performance. To further show the effectiveness of SOS features, we provide sample local goal search scenarios in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.698, + 0.47, + 0.768 + ], + "angle": 0, + "content": "
Local GoalVal SeenVal Unseen
SR↑SPL↑OSR↑TL↓NE↓SR↑SPL↑OSR↑TL↓NE↓
Oracle81.8874.1287.4613.061.9375.9562.5384.1614.002.71
Random79.3372.6785.3113.192.2270.9759.4580.1614.923.34
Homing80.2273.6385.6012.512.1471.6560.6080.3313.913.26
Spatial79.6373.1485.6012.992.2271.5660.0180.3314.903.27
Spectral80.6175.1585.8011.952.1171.7861.6880.7613.093.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.772, + 0.427, + 0.784 + ], + "angle": 0, + "content": "Table 3. Comparison of Exploitation Policies. (R2R)" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.79, + 0.469, + 0.861 + ], + "angle": 0, + "content": "
Local GoalSR↑Val Seen HouseTest Unseen House
SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑
Oracle54.4237.9663.7211.0148.3828.4562.984.74
Random24.7811.9734.963.0824.197.4135.841.29
Homing42.0427.7248.2310.1835.4019.1851.623.14
Spatial32.3011.6039.381.9026.1110.5839.231.43
Spectral44.6934.8452.658.8939.0925.8048.674.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.863, + 0.434, + 0.876 + ], + "angle": 0, + "content": "Table 4. Comparison of Exploitation Policies. (SOON)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.459, + 0.655, + 0.473 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.639 + ], + "angle": 0, + "content": "We conduct an ablation study to compare the proposed method against language-triggered hierarchical exploration. Results in the supplementary material show that among the three representation domains, spatial, spectral, and language, the spectral-domain features enhance navigation performance the most. Additionally, to implicate further applications of Meta-Explore in continuous environments, we evaluate our method on the photo-realistic Habitat [50] simulator to solve image-goal navigation and vision-and-language navigation tasks. Implementation details and results are included in the supplementary material. Results show that our method outperforms baselines in both tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.619, + 0.657 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.867 + ], + "angle": 0, + "content": "We have proposed Meta-Explore, a hierarchical navigation method for VLN, by correcting mistaken short-term actions via efficient exploitation. In the exploitation mode, the agent is directed to a local goal which is inferred to be the closest to the target. A topological map constructed during exploration helps the agent to search and plan the shortest path toward the local goal. To further search beyond the frontier of the map, we present a novel visual representation called scene object spectrum (SOS), which compactly encodes the arrangements and frequencies of nearby objects. Meta-Explore achieves the highest generalization performance for test splits of R2R, SOON, and val split of REVERIE navigation tasks by showing less overfitting and high success rates. We plan to apply Meta-Explore for VLN tasks in continuous environments in our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.502, + 0.898 + ], + "angle": 0, + "content": "6690" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Fengda Zhu, Yi Zhu, Vincent Lee, Xiaodan Liang, and Xiao-jun Chang. Deep learning for embodied vision navigation: A survey. arXiv preprint arXiv:2108.04097, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.472, + 0.279 + ], + "angle": 0, + "content": "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2018. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[3] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. Touchdown: Natural language navigation and spatial reasoning in visual street environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12538-12547, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.352, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[4] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Conference on Empirical Methods for Natural Language Processing, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.42, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[5] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In *Transactions of the Association for Computational Linguistics*, pages 1862–1872, Florence, Italy, July 2019. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.539 + ], + "angle": 0, + "content": "[6] An Yan, Xin Eric Wang, Jiangtao Feng, Lei Li, and William Yang Wang. Cross-lingual vision-language navigation. arXiv preprint arXiv:1910.11301, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[7] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. In Proceedings of the International Conference on Neural Information Processing Systems, volume 34, pages 652-663. Curran Associates, Inc., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.625, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[8] Jacob Krantz, Erik Wijmans, Arjun Majundar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision and language navigation in continuous environments. In Proceedings of the European Conference on Computer Vision, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.68, + 0.47, + 0.755 + ], + "angle": 0, + "content": "[9] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teplyashin, Karl Moritz Hermann, Mateusz Malinowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.824 + ], + "angle": 0, + "content": "[10] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11773–11781, Apr. 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.828, + 0.469, + 0.867 + ], + "angle": 0, + "content": "[11] Arun Balajee Vasudevan, Dengxin Dai, and Luc Van Gool. Talk2nav: Long-range vision-and-language navigation with dual attention and spatial memory. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.127, + 0.892, + 0.153 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, volume 129, pages 246-266. Springer, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.157, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[12] Dipendra Misra, Andrew Bennett, Valts Blukis, Eyvind Niklasson, Max Shatkhin, and Yoav Artzi. Mapping instructions to actions in 3D environments with visual goal prediction. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 2667-2678, Brussels, Belgium, October-November 2018. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.893, + 0.314 + ], + "angle": 0, + "content": "[13] Ta-Chung Chi, Minmin Shen, Mihail Eric, Seokhwan Kim, and Dilek Hakkani-tur. Justask: An interactive learning framework for vision and language navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2459-2466, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.893, + 0.358 + ], + "angle": 0, + "content": "[14] Yi Wu, Yuxin Wu, Georgia Gkioxari, and Yuandong Tian. Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.893, + 0.413 + ], + "angle": 0, + "content": "[15] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.894, + 0.482 + ], + "angle": 0, + "content": "[16] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12689-12699, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.487, + 0.893, + 0.562 + ], + "angle": 0, + "content": "[17] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.894, + 0.631 + ], + "angle": 0, + "content": "[18] Khanh Nguyen and Hal Daume III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, November 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.893, + 0.699 + ], + "angle": 0, + "content": "[19] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.893, + 0.768 + ], + "angle": 0, + "content": "[20] Alane Suhr, Claudia Yan, Jack Schluger, Stanley Yu, Hadi Khader, Marwa Mouallem, Iris Zhang, and Yoav Artzi. Executing instructions in situated collaborative interactions. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 2119-2130, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.893, + 0.836 + ], + "angle": 0, + "content": "[21] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In Proceedings of the International Conference on Learning Representations, 2019. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.893, + 0.867 + ], + "angle": 0, + "content": "[22] Chih-Yao Ma, Zuxuan Wu, Ghassan AlRegib, Caiming Xiong, and Zsolt Kira. The regretful agent: Heuristic-aided navigation" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.894, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.899 + ], + "angle": 0, + "content": "6691" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.127, + 0.47, + 0.166 + ], + "angle": 0, + "content": "through progress estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6732-6740, 2019. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.17, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[23] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10012-10022, 2020. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.237, + 0.469, + 0.313 + ], + "angle": 0, + "content": "[24] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537-16547, 2022. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.318, + 0.469, + 0.381 + ], + "angle": 0, + "content": "[25] Zhiwei Deng, Karthik Narasimhan, and Olga Russakovsky. Evolving graphical planner: Contextual global planning for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, volume 33, pages 20660-20672, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.385, + 0.469, + 0.448 + ], + "angle": 0, + "content": "[26] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8455-8464, June 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.453, + 0.468, + 0.502 + ], + "angle": 0, + "content": "[27] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Proceedings of the Conference on Robot Learning, pages 394-406, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.508, + 0.468, + 0.584 + ], + "angle": 0, + "content": "[28] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.588, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[29] Xiujun Li, Chunyuan Li, Qiaolin Xia, Yonatan Bisk, Asli Celikyilmaz, Jianfeng Gao, Noah A Smith, and Yejin Choi. Robust navigation with language pretraining and stochastic sampling. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 1494-1499, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.668, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[30] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Proceedings of the International Conference on Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.736, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[31] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1634-1643, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.867 + ], + "angle": 0, + "content": "[32] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. VIn\\(\\text{©}\\)bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[33] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15942-15952, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.183, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[34] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, pages 5834-5847, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.892, + 0.326 + ], + "angle": 0, + "content": "[35] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6741-6749, 2019. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.891, + 0.356 + ], + "angle": 0, + "content": "[36] James G March. Exploration and exploitation in organizational learning. Organization science, 2(1):71-87, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.891, + 0.398 + ], + "angle": 0, + "content": "[37] Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O Stanley, and Jeff Clune. First return, then explore. Nature, 590(7847):580-586, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.891, + 0.453 + ], + "angle": 0, + "content": "[38] Miruna Pislar, David Szepesvari, Georg Ostrovski, Diana L Borsa, and Tom Schaul. When should agents explore? In Proceedings of the International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.891, + 0.509 + ], + "angle": 0, + "content": "[39] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In Proceedings of the International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.515, + 0.892, + 0.59 + ], + "angle": 0, + "content": "[40] Meera Hahn, Devendra Singh Chaplot, Shubham Tulsiani, Mustafa Mukadam, James M Rehg, and Abhinav Gupta. No rl, no simulation: Learning to navigate without navigating. In Proceedings of the International Conference on Neural Information Processing Systems, volume 34, pages 26661-26673. Curran Associates, Inc., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.891, + 0.645 + ], + "angle": 0, + "content": "[41] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.738 + ], + "angle": 0, + "content": "[42] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.891, + 0.781 + ], + "angle": 0, + "content": "[43] Wolfgang Stürzl and Hanspeter A Mallot. Efficient visual homing based on fourier transformed panoramic images. Robotics and Autonomous Systems, 54(4):300-313, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.891, + 0.811 + ], + "angle": 0, + "content": "[44] Jean Serra. Mathematical Morphology, pages 1-16. Springer International Publishing, Cham, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[45] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 5100-5111, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "6692" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.165 + ], + "angle": 0, + "content": "[46] Stéphane Ross, Geoffrey J Gordon, and J Andrew Bagnell. No-regret reductions for imitation learning and structured prediction. arXiv preprint arXiv:1011.0686, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.17, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[47] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proceedings of the International Conference on Machine Learning, pages 1928-1937, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.469, + 0.301 + ], + "angle": 0, + "content": "[48] Peter Anderson, Angel Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, et al. On evaluation of embodied navigation agents. arXiv preprint arXiv:1807.06757, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.469, + 0.381 + ], + "angle": 0, + "content": "[49] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6629-6638, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[50] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Proceedings of the International Conference on Neural Information Processing Systems, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.47, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "6693" + } + ] +] \ No newline at end of file diff --git a/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_origin.pdf b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f0cb25a57a68194bb1dc46a26b3cc8eebe2338d4 --- /dev/null +++ b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/50db118f-41a0-4ca2-be57-c2ebbcb2698f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9c0213013776cb71b3a3c073b56c7415dfc7c15a9084ce664c0d4cc34712b88 +size 1088286 diff --git a/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/full.md b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c8d4695144aeb6ceaee10c61266326b335fce8e2 --- /dev/null +++ b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/full.md @@ -0,0 +1,263 @@ +# Meta-Explore: Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding + +Minyoung Hwang $^{1}$ , Jaeyeon Jeong $^{1}$ , Minsoo Kim $^{3}$ , Yoonseon Oh $^{2}$ , Songhwai Oh $^{1}$ $^{1}$ Electrical and Computer Engineering and ASRI, Seoul National University + $^{2}$ Department of Electronic Engineering, Hanyang University + $^{3}$ Interdisciplinary Major in Artificial Intelligence, Seoul National University + +{minyoung.hwang, jaeyeon.jeong}@rlllab.snu.ac.kr, goldbird5@snu.ac.kr, yoh21@hanyang.ac.kr, songhwai@snu.ac.kr + +# Abstract + +The main challenge in vision-and-language navigation (VLN) is how to understand natural-language instructions in an unseen environment. The main limitation of conventional VLN algorithms is that if an action is mistaken, the agent fails to follow the instructions or explores unnecessary regions, leading the agent to an irrecoverable path. To tackle this problem, we propose Meta-Explore, a hierarchical navigation method deploying an exploitation policy to correct misled recent actions. We show that an exploitation policy, which moves the agent toward a well-chosen local goal among unvisited but observable states, outperforms a method which moves the agent to a previously visited state. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. The key to our approach is understanding the object placements around the agent in spectral-domain. Specifically, we present a novel visual representation, called scene object spectrum (SOS), which performs category-wise 2D Fourier transform of detected objects. Combining exploitation policy and SOS features, the agent can correct its path by choosing a promising local goal. We evaluate our method in three VLN benchmarks: R2R, SOON, and REVERIE. Meta-Explore outperforms other baselines and shows significant generalization performance. In addition, local goal search using the proposed spectral-domain SOS features significantly improves the success rate by $17.1\%$ and SPL by $20.6\%$ against the state-of-the-art method of the SOON benchmark. Project page: https://rllab-snu.github.io/projects/Meta-Explore/doc.html + +# 1. Introduction + +Visual navigation in indoor environments has been studied widely and shown that an agent can navigate in unex- + +![](images/3edce37aa747505b4e8eaf55c9d230dc101641e3984131ef27e1d082eb0b25a1.jpg) +Figure 1. Hierarchical Exploration. At each episode, a natural-language instruction is given to the agent to navigate to a goal location. The agent explores the environment and constructs a topological map by recording visited nodes $\odot$ and next step reachable nodes $\clubsuit$ . Each node consists of the position of the agent and visual features. $o_t$ denotes the observation at time $t$ . The agent chooses an unvisited local goal to solve the regretful exploration problem. + +plored environments [1]. By recognizing the visual context and constructing a map, an agent can explore the environment and solve tasks such as moving towards a goal or following a desired trajectory. With the increasing development in human language understanding, vision-and-language navigation (VLN) [2] has enabled robots to communicate with humans using natural languages. The high degree of freedom in natural language instructions allows VLN to expand to various tasks, including (1) following fine-grained step-by-step instructions [2-13] and (2) reaching a target location described by goal-oriented language instructions [14-20]. + +A challenging issue in VLN is the case when an action is mistaken with respect to the given language instruction [21-26]. For instance, if the agent is asked to turn right at the end of the hallway but turns left, the agent may end up in irrecoverable paths. Several existing studies solve this issue via hierarchical exploration, where the high-level planner decides when to explore and the low-level planner chooses what actions to take. If the high-level planner chooses to explore, the agent searches unexplored regions, and if it chooses to exploit, the agent executes the best action based on the previous exploration. Prior work [21-23] returns the agent to the last successful state and resumes exploration. + +However, such methods take a heuristic approach because the agent only backtracks to a recently visited location. The agent does not take advantage of the constructed map and instead naively uses its recent trajectory for backtracking. Another recent work [26] suggests graph-based exploitation, which uses a topological map to expand the action space in global planning. Still, this method assumes that the agent can directly jump to a previously visited node. Since this method can perform a jump action at every timestep, there is no trigger that explicitly decides when to explore and when to exploit. Therefore, we address the importance of time scheduling for exploration-exploitation and efficient global planning using a topological map to avoid reexploring visited regions. + +We expand the notion of hierarchical exploration by proposing Meta-Explore, which not only allows the high-level planner to choose when to correct misled local movements but also finds an unvisited state inferred to be close to the global goal. We illustrate the overview of hierarchical exploration in Figure 1. Instead of backtracking, we present an exploitation method called local goal search. We show that it is more efficient to plan a path to a local goal, which is the most promising node from the unvisited but reachable nodes. We illustrate the difference between conventional backtracking and local goal search in Figure 2. Based on our method, we show that exploration and exploitation are not independent and can complement each other: (1) to overtake regretful explorations, the agent can perform exploitation and (2) the agent can utilize the constructed topological map for local goal search. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. Most VLN tasks require a level of understanding objects nearby the agent, but previous studies simply encode observed panoramic or object images [2, 3, 16-18, 21-35]. In this paper, we present a novel semantic representation of the scene called scene object spectrum (SOS), which is a matrix containing the arrangements and frequencies of objects from the visual observation at each location. Using SOS features, we can sufficiently estimate the context of the environment. We show that the proposed spectral-domain SOS features manifest better linguistic interpretability than conventional spatial-domain visual features. Combining exploitation policy and SOS features, we design a navigation score that measures the alignment between a given language instruction and a corrected trajectory toward a local goal. The agent compares local goal candidates and selects a near-optimal candidate with the highest navigation score from corrected trajectories. This involves high-level reasoning related to the landmarks (e.g., bedroom and kitchen) and objects (e.g., table and window) that appear in the instructions. + +The main contributions of this paper are as follows: + +- We propose a hierarchical navigation method called Meta-Explore, deploying an exploitation policy to cor + +![](images/cae102e992b908d5afb9a1f840f5b5aa8a60047e5e8a448b980f2d3f4eab3042.jpg) +Figure 2. Local Goal Search for Exploitation. The local goal is likely to be chosen as the closest node to the global goal. Existing methods only backtrack to a visited node (left). We expand the searchable area by including unvisited but reachable nodes (right). + +rect misled recent actions. The agent searches for an appropriate local goal instead of reversing the recent action sequence. + +- In the exploitation mode, the agent uses a novel scene representation called scene object spectrum (SOS), which contains the spectral information of the object placements in the scene. SOS features provide semantically meaningful clues to choose a near-optimal local goal and help the agent to solve the regretful exploration problem. +- We evaluate our method on three VLN benchmarks: R2R [2], SOON [16], and REVERIE [17]. The experimental results show that the proposed method, MetaExplore, improves the success rate and SPL in test splits of R2R, SOON and val split of REVERIE. The proposed method shows better generalization results compared to all baselines. + +# 2. Related Work + +# 2.1. Vision-and-Language Navigation + +In VLN, an agent encodes the natural language instructions and follows the instructions, which can be either (1) a fine-grained step-by-step instruction the agent can follow [2-4], (2) a description of the target object and location [16, 17], or (3) additional guidance given to the agent [18, 27]. These tasks require the agent to recognize its current location using some words in the natural-language instructions. Prior work [2, 28-31] show that an agent can align visual features to language instructions via neural networks and use the multimodal output embeddings to generate a suitable action at each timestep. Most VLN methods utilize cross-modal attention, either with recurrent neural networks [2, 28] or with transformer-based architectures [29-31]. For sequential action prediction, Hong et al. [32] further use recurrent units inside transformer architectures, while Pashevich et al. [33] and Chen et al. [34] use additional transformers to embed past observations and actions. + +# 2.2. Exploration-Exploitation + +In an unseen environment, the agent must maximize the return without knowing the true value functions. One of the solutions to this problem is to switch back and forth between exploration and exploitation [36]. In the exploration mode, the agent gathers more information about the environment. On the other hand, the agent uses information collected during exploration and chooses the best action for exploitation. Ecoffet et al. [37] reduced the exploration step by archiving the states and exploring again from the successful states. Pislar et al. [38] addressed the various scheduling policies and demonstrated their method on Atari games. Recent work [39, 40] successfully demonstrates the effectiveness of hierarchical exploration in image-goal navigation. + +Like commonly used greedy navigation policies, VLN tasks also deal with the problem of maximizing the chance to reach the goal without knowing the ground truth map. Several VLN methods employ the concept of exploitation to tackle this problem. Ke et al. [35] look forward to several possible future trajectories and decide whether to backtrack or not and where to backtrack. Others [21-23] estimate the progress to tell whether the agent becomes lost and make the agent backtrack to a previously visited location to restart exploration. However, previous studies do not take into account what should be done in the exploitation mode. In order to handle this problem, we propose a hierarchical navigation method which determines the scheduling between exploration and exploitation. + +# 2.3. Visual Representations + +Popular visual encoding methods via ResNet [41] and ViT [42] can be trained to learn rotation-invariant visual features. Both methods learn to extract visual features with high information gain for global and local spatial information. The high complexity of the features leads to low interpretability of the scene and therefore requires the agent to use additional neural networks or complex processing to utilize them. On the other hand, traditional visual representation methods such as Fourier transform use spectral analysis, which is highly interpretable and computationally efficient. One drawback of the traditional methods is that they fail to maximize the information gain. Nonetheless, an appropriate use of essential information can be helpful for high-level decision making and enables more straightforward interpretation and prediction of the visual features. One traditional navigation method, Sturz et al. [43] used Fourier transform to generate rotation-invariant visual features. However, no research has transformed the spectral information of the detected objects to represent high-level semantics from visual observations. Focusing on the fact that 2D Fourier transform can extract morphological properties of images [44], we can find out the shape or structure of detected objects through 2D Fourier transform. In this paper, we decompose the object mask into binary masks by object categories and perform a 2D Fourier transform on each binary mask. + +# 3. Method + +# 3.1. Problem Formulation + +We deal with VLN in discrete environments, where the environment is given as an undirected graph $G_{e} = \{V,E\}$ . $V$ denotes a set of $N$ navigable nodes, $\{V_{i}\}_{i = 1}^{N}$ , and $E$ is the adjacency matrix describing connectivity among the nodes in $V$ . We denote the observation at node $V_{i}$ as $O_{i}$ . The agent uses a panoramic RGB image observation $o_{t}$ and current node $v_{t}$ , which are collected at time $t$ . The agent either moves to a neighboring node or executes a stop action. $a_{t}$ denotes the action at time $t$ . The objectives of VLN are categorized as follows: (1) to follow language instructions [2] and (2) to find a target object described by language instructions in a fixed time $T$ [16, 17]. We present a general hierarchical exploration method that can be applied to both tasks. We also enhance the navigation policy by extracting cross-domain visual representations from the environments, i.e., spatial-domain and spectral-domain representations. To balance the information loss and interpretability of the visual feature, we adopt multi-channel fast Fourier transform (FFT) to encode semantic masks of the detected objects into category-wise spectral-domain features. + +# 3.2. Meta-Explore + +We design a learnable hierarchical exploration method for VLN called Meta-Explore, which decides (1) when to explore or exploit and (2) a new imagined local goal to seek during exploitation. The overall network architecture of the proposed Meta-Explore is shown in Figure 3. Given a language instruction $L$ , the agent navigates in the environment until it finds the target described in $L$ . Meta-Explore consists of a mode selector and two navigation modules corresponding to two modes: exploration and exploitation. At each timestep, the mode selector chooses to explore or exploit. At $t = 0$ , the mode is initialized to exploration. In the exploration mode, the agent outputs an action toward a neighboring node to move the agent toward the goal. When the mode selector recognizes that the agent is not following the instruction successfully, the mode is switched to exploitation. In the exploitation mode, the agent seeks a new local goal with the highest correspondence against the language instructions from the previously unvisited candidate nodes using spectral-domain visual features. The agent moves toward the local goal by planning a path. After the agent arrives at the local goal, the mode is reset to exploration. The explore-exploit switching decision occurs through the mode selector by estimating the probability to explore. The agent repeats this explore-exploit behavior until it determines that the target is found and decides to stop. + +# 3.2.1 Mode Selector + +At time $t$ , the agent observes visual features about the current node $v_{t}$ and several reachable nodes. We call the nodes reachable at the current timestep as candidate nodes. $n_{cand}$ + +![](images/beec1b93785ed13d1fb664fb0d718a97f7174f0600af642877fcb3add3493e74.jpg) +Figure 3. Network Architecture. Three types of visual features: panoramic (yellow), object image (aquamarine), and object spectrum (red) are encoded. The color in each parenthesis denotes the color describing the corresponding feature. The cross-modal transformer encodes language and spatial visual features as hidden state $H_{t}$ . A mode selector gives explore or exploit command to the agent by predicting the explore probability $P_{\text{explore}}$ . The selected navigation module outputs an action $a_{t}$ from the possible $n_{\text{cand}}$ candidate nodes. + +denotes the number of candidate nodes. We use a cross-modal transformer with $n_{L}$ layers to relate visual observations to language instructions. The cross-modal transformer takes the visual features of nodes in the constructed topological map at time $t$ , $G_{t}$ , and outputs cross-modal embedding $H_{t}$ to encode visual observations with $L$ . We concatenate location encoding and history encoding [24] to the visual features as node features to consider the relative pose from $v_{t}$ and the last visited timestep of each node, respectively. Each word is encoded via a pretrained language encoder [45], which is used for general vision-language tasks. + +The cross-modal transformer consists of cross-attention layer L2V_Atn(\hat{W},\hat{V}) = Softmax(\hat{W}\Theta_q^\dagger (\hat{V}\Theta_k^\dagger)^T /\sqrt{d})\hat{V}\Theta_v^\dagger and self-attention layer SelfAttn(X) = Softmax $(X\Theta_{q}(X\Theta_{k})^{T}+$ $A\Theta_e + b_e) / \sqrt{d})X\Theta_v$ , where $\hat{W},\hat{V},X$ , and $A$ denote word, visual, node representations and adjacency matrix of $G_{t}$ respectively. The (query, key, value) weight matrices of self-attention and cross-attention layers are denoted as $(\Theta_q,\Theta_k,\Theta_v)$ and $(\Theta_q^\dagger ,\Theta_k^\dagger ,\Theta_v^\dagger)$ , respectively. The final crossmodal embedding at time $t$ after passing through $n_L$ transformer layers is denoted as $H_{t}$ . To encourage the monotonic increasing relationship between language and visual attentions at each timestep, we define a correlation loss $L_{corr} = \sum_{t = 1}^{T}\| \mathrm{L2V\_Attn} - I_{n_x}\| _1$ for training the crossmodal transformer, where $n_x$ denotes the dimension of the $H_{t}$ and $I_{n_x}$ denotes an identity matrix of size $n_x\times n_x$ + +As illustrated in Figure 4, the mode selector estimates the probability to explore $P_{\text{explore}}$ given the cross-modal hidden state $H_t$ . We denote the mode selector as $S_{\text{mode}}$ and use a two-layer feed-forward neural network. Given $H_t$ , $S_{\text{mode}}$ outputs the exploration probability as $P_{\text{explore}} = 1 - S_{\text{mode}}(H_t)$ . If $P_{\text{explore}} \geq 0.5$ , the exploration policy outputs a probability distribution for reachable nodes at the + +next step. At time $t + 1$ , the agent moves to the node with the highest probability. If $P_{\text{explore}} < 0.5$ , the agent determines that the current trajectory is regretful, so the agent should traverse to find a local goal, which is the most likely to be the closest node to the global goal. The exploitation policy mainly utilizes object-level features to search for the local goal with high-level reasoning. After the local goal is chosen, the path planning module outputs an action following the shortest path to the local goal. + +To train the mode selector, we require additional demonstration data other than the ground truth trajectory, such that it switches between exploration and exploitation. We generate the demonstration data from the ground truth trajectories, with additional detours. For the detours, we stochastically select candidate nodes other than the ground truth paths and add the trajectory that returns to the current viewpoint. The imitation learning loss for training the mode selector is defined as $L_{mode} = \sum_{t=1}^{T} \mathbb{1}(m_t = \mathrm{gt}_t)$ , where $m_t$ is the mode of the agent, 0 for exploitation and 1 for exploration. $\mathrm{gt}_t$ is 1 if the current node is in the shortest ground truth trajectory and $\mathrm{gt}_t = 0$ , otherwise. + +# 3.2.2 Exploration Module + +In the exploration mode, the agent follows the following sequential operations: topological map construction, self-monitoring, and an exploration policy. To improve the exploration, we adopt self-monitoring [21] to predict the current progress of exploration to enhance the exploration policy itself. Prior work [21, 22] has shown that auxiliary loss using self-monitoring can regularize the exploration policy. + +Topological Map Construction. The agent constructs graph $G_{t}$ by classifying nodes into two types: (1) visited nodes and (2) unvisited but observable nodes. At current + +![](images/a84bd4402484b8d5c99c18b4e3a47c29d64252e158398d83efef5e5d34ea4b21.jpg) + +![](images/49a1a66bfa018bbdb7f02df07e6ecca1068f310984d4e327817c647b63e6fe41.jpg) +Exploration Module +Figure 4. Navigation Modules. Mode selector estimates $P_{\text{explore}}$ , i.e., the probability to explore, and chooses between exploration and exploitation modules. The selected navigation module outputs the next action $a_t$ . + +![](images/5c4a582aac44a3dbcdcd77c052679afa3b2a36df0d9cf7919d8eaa275f55f1a0.jpg) +Exploitation Module + +time $t$ , the agent at node $v_{t} \in \{V_{i}\}_{i=1}^{N}$ observes $N(v_{t})$ neighbor nodes as next step candidates at time $t + 1$ . The visited nodes consist of visual features of their own and the neighboring nodes from panoramic RGB observations. The unvisited nodes can be observed only if they are connected to at least one visited node. The topological map records the positions and visual features of observed nodes at each timestep. By knowing the positions of nodes in $G_{t}$ , the agent can plan the shortest path trajectory between two nodes. + +Self-Monitoring. We use a progress monitor to estimate the current navigation progress at each episode. Self-monitoring via estimating current progress helps the agent choose the next action that can increase the progress. The estimated progress $\hat{p}_t = F_{\text{progress}}(H_t)$ is the output of a feed-forward neural network, given $H_t$ as input. We measure the ground truth progress $p_t$ as the ratio between the current distance to the goal and the shortest path length of the episode subtracted from 1, described as $1 - \frac{d_{geo}(v_t, v_{goal})}{d_{geo}(v_0, v_{goal})}$ , where $d_{geo}(a, b)$ is the geodesic distance between $a$ and $b$ . $v_0, v_t$ , and $v_{goal}$ denote initial, current, and goal positions, respectively. We add progress loss $L_{\text{progress}} = \sum_{t=1}^{T} (\hat{p}_t - p_t)^2$ to train the progress monitor while training the exploration policy. + +Exploration Policy. The exploration policy $F_{explore}$ estimates the probability of moving to the candidate nodes at the next step. The agent chooses the action $a_t$ at time $t$ based on the estimated probability distribution among candidate nodes, described as $a_t = \arg \max_{V_i} (F_{explore}([H_t]_i))$ . $F_{explore}$ is implemented via a two-layer feed-forward network with the cross-modal hidden state $H_t$ given as input. The output of $F_{explore}$ becomes a probability distribution over possible actions. To only consider unvisited nodes, we mask out the output for visited nodes. For training, we sample the next action from the probability distribution instead of choosing a node with the highest probability. We describe the training details in Section 3.3. + +![](images/f87060c14ebff4c2fc1cd3db0294d7b820c3aa365ea34d85d467c07f4f0fc055.jpg) +Figure 5. Scene Object Spectrum (SOS). The agent calculates scene object spectrum (SOS) features for efficient exploitation. SOS features incorporate semantic information observed in a single panoramic image by performing category-wise 2D FFT. + +# 3.2.3 Exploitation Module + +In the exploitation mode, the agent requires high-level reasoning with identifiable environmental clues to imagine regretful exploration cases. To find clues in an object-level manner, we present a novel visual representation by capturing object information in the spectral-domain. The novel representation is more easily predictable than spatial features such as RGB image embeddings. The agent can take advantage of the predictability by expanding the searchable area to find a local goal. We choose the local goal as the closest node to the global goal in the feature space. + +Spectral-Domain Visual Representations. Common navigation policies can lead the agent toward the node with the highest similarity to the target. However, even with a good learned policy, the agent can act in a novice manner in unseen environments. In this paper, we seek extra information from the environment for generalizable high-level reasoning to resolve the issue. As illustrated in Figure 5, scene object spectrum (SOS) incorporates semantic information observed in a single panoramic image by generating a semantic mask for each object category and applying Fourier transform to each semantic mask. The semantic mask for object class $k$ at time $t$ is calculated as a binary mask $[m_t^k]_{ij}$ that detects the object at pixel $(i,j)$ . Suppose there are a total of $K$ object categories. When multiple objects are detected for one object category, the binary mask appears as a union of the bounding boxes of the detected objects. We define FFT as a channel-wise 2D fast Fourier transform that receives $K$ binary semantic masks and outputs $K$ spectral-domain features, where $K$ is the number of object classes. Then, SOS feature $\vec{S}_t = [s_t^1,\dots,s_t^K ]^T$ can be defined as $s_t^k = \log |\mathbf{FFT}(m_t^k)|$ . For simplicity, we perform mean pooling on the vertical spectral axis and normalize the output. The final SOS feature has shape $(K,\eta)$ , where $\eta$ is the maximum horizontal frequency. + +Local Goal Search Using Semantic Clues. We argue that returning to a previously visited node does not guarantee the agent escapes from the local optima. Instead of backtracking + +to a previously visited node, the agent searches for a local goal to move towards. If the agent plans a path and moves towards the local goal, the agent does not need to repeat unnecessary actions in visited regions after the exploitation ends. Additionally, searching for a local goal takes full advantage of the topological map by utilizing the connections among the observed nodes. To expand the searchable area further, we let the agent choose the local goal from previously unvisited and unchosen candidate nodes. + +To choose a local goal, we first score the corrected trajectories to measure the alignment with the language instruction $L$ . We use SOS features as semantic environmental clues to estimate the navigation score $S_{\text{nav}}$ of the corrected trajectory, which is the shortest path trajectory from the initial node to the local goal in the constructed topological map. To simplify, we convert the language instruction into a list of objects $W^{o} = [w_{1}^{o}, \dots, w_{B}^{o}]$ consisting of $B[\leq K]$ object categories (e.g., desk, cabinet, and microwave). We approximate the corresponding reference SOS features as $[\hat{\delta}(w_{1}^{o}), \dots, \hat{\delta}(w_{B}^{o})]$ where the $i^{th}$ row of $\hat{\delta}(w_{k}^{o})$ is defined as $[\hat{\delta}(w_{k}^{o})]_{i} = \mathbb{1}(k = i)\lambda(\delta(w_{k}^{o}))\mathrm{sinc}\left(\frac{j}{2} - \frac{\eta}{4}\right)$ . $\lambda(\delta(w_{k}^{o}))$ denotes the average width of detected bounding boxes of object $w_{k}^{o}$ in the environment. A detailed approximation process is explained in the supplementary material. To simulate a corrected trajectory $\mathcal{T}' = (v_{1}', \dots, v_{t'})$ , we calculate the SOS features $[\vec{S}_1', \dots, \vec{S}_{t'}']$ corresponding to the nodes in $\mathcal{T}'$ . We measure the similarity between two object spectrum features via the cosine similarity of the flattened vectors. Finally, the navigation score $S_{\text{nav}}$ of $\mathcal{T}'$ is computed as: + +$$ +S _ {n a v} \left(\mathcal {T} ^ {\prime}\right) = \frac {\sum_ {i = 1} ^ {B} \sum_ {j = 1} ^ {t ^ {\prime}} \left(\frac {\hat {\delta} \left(w _ {i} ^ {o}\right)}{\left| \hat {\delta} \left(w _ {i} ^ {o}\right) \right|} \cdot \frac {\vec {S} _ {j} ^ {\prime}}{\left| \vec {S} _ {j} ^ {\prime} \right|}\right) \left(\left(\hat {\delta} \left(w _ {i} ^ {o}\right) - \hat {\delta} (\bar {w} ^ {o})\right) \cdot \left(\vec {S} _ {j} ^ {\prime} - \bar {\vec {S}} ^ {\prime}\right)\right)}{\sqrt {\frac {t ^ {\prime}}{B} \cdot \sum_ {i = 1} ^ {B} \left(\hat {\delta} \left(w _ {i} ^ {o}\right) - \hat {\delta} (\bar {w} ^ {o})\right) ^ {2} \sum_ {j = 1} ^ {t ^ {\prime}} \left(\vec {S} _ {j} ^ {\prime} - \bar {\vec {S}} ^ {\prime}\right) ^ {2}}}, \tag {1} +$$ + +where $\hat{\delta} (\overline{w}^o)$ and $\vec{S}'$ denote the average values of SOS features $\hat{\delta} (w_i^o)$ and $\vec{S}_j^\prime$ , respectively. This equation can also be interpreted as a pseudo correlation-coefficient function between object list $W^{o}$ and trajectory $\mathcal{T}'$ . The exploitation policy selects the node with the highest navigation score as the local goal from the previously unvisited candidates. + +Figure 6 illustrates a simple scenario of entering a room. Suppose $W^{o} =$ [sculpture, door, bed] and the agent has to compare two trajectories $\mathcal{T}_1 = (v_1,v_2,A)$ and $\mathcal{T}_1 = (v_1,v_2,B)$ . Each similarity matrix in Figure 6 has the $(t,j)$ element as the similarity between the SOS feature of $V_{t}$ and $\hat{\delta}(w_j^o)$ , which is calculated as $\hat{\delta}(w_j^o) \cdot \vec{S}_t'$ . Notably, the similarity matrix shows monotonic alignment and the navigation score is higher when the next action is chosen correctly. + +# 3.3. Training Details + +We use [24] for pretraining the visual encoder with panoramic RGB observations. We use the DAgger algorithm [46] to pretrain the navigation policy and the mode selector. To prevent overfitting, we iteratively perform + +![](images/953ad6452a75b7dd3dfd02078929ac7b6d9100d714fc9faa876085f058d646c8.jpg) +Figure 6. Toy Example. Monotonic alignment between language instruction and visual observation is desirable. Yellow dots $\bullet$ in the nodes describe the ground truth trajectory. Based on the node at $t = 3$ , the similarity matrix can show either monotonic or non-monotonic alignment between object tokens and SOS features. The green circles $\clubsuit$ describe the possible candidates $A, B$ for next action. + +similarity matrix + +
w01w2w3
v1=10.90.00.0
v1=20.31.00.0
v1=30.00.01.0
v1=4, Snav = 0.296
w01w2w3
v1=10.90.00.0
v1=20.31.00.0
v1=30.50.40.0
v1=4, Snav = 0.296
+ +teacher forcing and student forcing to choose the action from the exploration policy. Imitation learning loss is calculated as $L_{IL} = \sum_{t=1}^{T} -\log p(a_t^* | a_t)$ and object grounding loss is calculated as $L_{OG} = -\log p(\mathrm{obj}^* | \mathrm{obj}_{pred})$ , where $\mathrm{obj}^*$ denotes the ground truth and $\mathrm{obj}_{pred}$ denotes the predicted object location. The total loss function is defined as $L_{total} = L_{mode} + L_{progress} + L_{corr} + L_{IL} + L_{OG}$ . We further finetune the agent via A2C [47]. The exploration policy selects the action $a_t$ with probability $p_t^a$ . Reinforcement learning loss is defined as $L_{RL} = -\sum_t a_t^s \log(p_t^a) A_t - \lambda \sum_t a_t^* \log(p_t^a)$ . To train the mode selector, progress monitor, and exploration policy in an end-to-end manner, we use the total loss function as $L_{fine} = L_{mode} + L_{progress} + L_{RL}$ . The exploitation policy searches the path toward the local goal from the constructed navigation graph. Thus, the exploitation policy is not learned. + +# 4. Navigation Experiments + +# 4.1. Experiment Settings + +We evaluate our method on three VLN benchmarks, Room-to-Room(R2R) [2], SOON [16], and REVERIE [17]. R2R evaluates the visually-grounded natural navigation performance of the agent. The agent must navigate to the predefined goal point given image observations and language instructions in an unseen environment. + +SOON is also a goal-oriented VLN benchmark. Natural language instructions in SOON have an average length of 47 words. The agent should locate the target location and detect the location of an object to find the target object. + +REVERIE is a goal-oriented VLN benchmark that provides natural language instruction about target locations and objects. In REVERIE, the agent is given an instruction referring to a remote object with an average length of 21 words. With this instruction and a panoramic observation from the environment, the agent should navigate to the location the instruction describes and find the correct object + +bounding box among the predefined object bounding boxes. + +# 4.2. Evaluation Metrics + +# 4.2.1 Navigation performance + +We evaluate algorithms using the trajectory length (TL), success rate (SR), and success weighted by inverse path length (SPL) [48], and oracle success rate (OSR) for the navigation performance comparison. An episode is recorded as a success if the agent takes a stop action within $3\mathrm{m}$ of the target location. TL is the average path length in meters. SR is denoted as the number of successes divided by the total number of episodes, $M$ . SPL is calculated as $\frac{1}{M}\sum_{i = 1}^{M}S_{i}\frac{l_{i}}{\max(p_{i},l_{i})}$ , where $S_{i}$ denotes the success as a binary value. $p_i$ and $l_{i}$ denote the shortest path and actual path lengths for the $i^{th}$ episode. OSR uses the oracle stop policy instead of the stop policy of the agent. + +# 4.2.2 Object grounding performance + +We also evaluate the object grounding performance of the agent by the success rate of finding the target object (FSR) and the target finding success weighted by inverse path length (FSPL)1 [16, 17]. FSPL is calculated as FSPL = $\frac{1}{N}\sum_{i=1}^{N}S_{i}^{nav}S_{i}^{loc} \cdot l_{i}^{nav}/\max(l_{i}^{nav},l_{i}^{gt})$ , where $S_{i}^{nav}$ is whether the agent navigates to the target, $S_{i}^{loc}$ is whether the agent finds a target object bounding box, and $l_{i}^{nav}$ and $l_{i}^{gt}$ are the navigation trajectory length and ground truth trajectory length, respectively. + +# 4.3. Baselines and Implementation Details + +We compare our method with several other baselines as follows. For each task, we compare our method with a number of baselines that use various types of memory (recurrent, sequential, and topological map). For methods implemented with a hierarchical navigation framework, we compare the specific exploitation methods: homing, jump, and local goal search. Homing makes the agent backtrack, and jump makes the agent jump to a previously visited node. The hyperparameters and detailed model architecture of Meta-Explore are described in the supplementary material. + +# 4.4. Comparison with Navigation Baselines + +We compare our method with navigation baselines. We focus on the success rate and SPL. Rendered results and detailed analyses with other evaluation metrics are provided in the supplementary material. + +R2R. Table 1 compares the proposed Meta-Explore with baselines for the R2R navigation task. We categorize the baseline methods based on the type of constructed memory and the type of exploitation. Our method outperforms other exploration-only baselines over all types of validation and test splits in success rate and SPL. Compared with hierarchical baselines SMNA [21], Regretful-Agent [22], FAST [35], and SSM [26], Meta-Explore improves success + +rate and SPL by at least $16.4\%$ and $8.9\%$ , respectively. The main difference is that Meta-Explore constructs a topological map during exploration and uses the map for local goal search in exploitation. On the contrary, homing exploitation policies in SMNA, Regretful-Agent, and FAST only rely on the current trajectory, instead of taking advantage of the constructed memory. Jump exploitation in SSM uses a topological map to search a successful previous node, but it makes an unrealistic assumption that the agent can directly jump to a previously visited distant node and unfairly saves time. In our approach, we plan a path to the local goal based on the topological map. The experiment results reveal that even if we design a hierarchical navigation framework, exploration and exploitation are not entirely separate but they can complement each other. + +SOON, REVERIE. Table 2 compares Meta-Explore with baselines in the SOON navigation task. While the proposed method does not improve performance in val seen split, Meta-Explore outperforms other baselines in the test unseen split of SOON for success rate by $17.1\%$ and SPL by $20.6\%$ . The result implies that for the goal-oriented VLN task, high performance in train or val seen splits can be the overfitted result. Because the agent can be easily overfitted to the training data, making a generalizable model or providing a deterministic error-correction module for inference is essential. Meta-Explore chooses the latter approach by correcting the trajectory via exploitation in regretful cases. The evaluation results in the REVERIE navigation task are described in the supplementary material. Meta-Explore shows improvement in the val split of REVERIE for success rate and SPL, but the improvement in the test split is lower than the results in R2R and SOON. We found 252 meaningless object categories (e.g., verbs, adjectives, and prepositions) and 418 replaceable object categories (e.g., typographical errors and synonyms) in the REVERIE3 dataset. Because our exploitation method utilizes object-based parsing of the given instruction to match with the detected object categories, the effectiveness of the proposed method is lessened due to inaccuracies and inconsistencies in the dataset. We expect to have higher performance if the mistakes in the dataset are fixed. + +# 4.5. Local Goal Search using SOS Features + +To discuss the significance of modeling exploitation policy, we conduct specific experiments about choosing the local goal for R2R and SOON. We evaluate our method using different types of local goal search, as shown in Table 3 and 4. Oracle denotes a method which selects a local goal using the ground truth trajectory. The performance of the oracle provides the achievable performance for each dataset. The results imply that local goal search using either + +
MethodsMemoryExploitVal SeenVal UnseenTest Unseen
SR↑SPL↑TL↓NE↓SR↑SPL↑TL↓NE↓SR↑SPL↑TL↓NE↓
Random--16-9.589.4516-9.779.2313129.899.79
Human----------11.851.618676
Seq2Seq [2]RecX6.03911.33-22-8.397.8420188.137.85
VLN\(\circ\)BERT [32]RecX726811.132.90635712.013.93635712.354.09
SMNA†[21]Rechoming696311.693.31474112.615.486156-4.48
Regretful-Agent [22]Rechoming6963-3.235041-5.324840-5.69
FAST (short) [35]Rechoming----564321.174.97544122.085.14
FAST (long) [35]Rechoming7004188.063.136302224.424.036103196.534.29
HAMT-e2e [34]SeqX767211.152.51666111.462.29656012.273.93
DUET [24]Top. MapX797312.322.28726013.943.31695914.733.65
SSM [26]Top. Mapjump716214.73.10624520.74.32614620.44.57
Meta-Explore (Ours)Top. Maplocal goal817511.952.11726213.093.22716114.253.57
+ +Table 1. Comparison and evaluation results of the baselines and our model in the R2R Navigation Task. Gray shaded rows describe hierarchical navigation baselines. Three memory types: Rec(recurrent), Seq(sequential), and Top. Map(topological map) + +
MethodsMemoryExploitVal Seen InstructionVal Seen HouseTest Unseen House
SR↑SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑
Human----------90.459.291.451.1
RandomRecX0.01.50.11.40.10.00.40.92.10.42.70.0
Speaker-Follower [28]RecX97.997.797.824.561.260.469.49.17.06.19.80.6
RCM [49]RecX84.082.689.110.962.460.972.77.87.46.212.40.7
AuxRN [23]RecX98.497.498.713.768.867.378.58.38.16.711.00.5
GBE w/o GETop. MapX89.588.391.824.262.560.873.06.711.48.718.80.8
GBE [16]Top. MapX98.497.998.644.276.362.564.17.311.910.219.51.4
\( GBE^† \)Top. MapX----19.513.328.51.212.99.221.50.5
DUET [24]Top. MapX94.091.690.031.136.322.650.93.833.421.443.04.2
Meta-Explore (Ours)Top. Maplocal goal100.099.196.033.944.734.852.78.939.125.848.74.0
+ +spatial or spectral visual representations is more effective than random local goal search. The results show that local goal search using spectral visual representations, i.e., SOS features, lead the agent to desirable nodes the most. We also compare local goal search with homing and the difference between the performance of the two methods is most noticeable in the test split of the SOON navigation task. As shown in Table 4, choosing the local goal with only spatial-domain features, the navigation performance does not improve compared to homing. On the contrary, spectral-domain local goal search shows significant improvement against homing by $10.4\%$ in success rate, $34.5\%$ on SPL, and $27.4\%$ on FSPL. The results imply that using spectral-domain SOS features helps high-level decision making, thereby enhancing the navigation performance. To further show the effectiveness of SOS features, we provide sample local goal search scenarios in the supplementary material. + +Table 2. Comparison and evaluation results of the baselines and our model in the SOON Navigation Task. + +
Local GoalVal SeenVal Unseen
SR↑SPL↑OSR↑TL↓NE↓SR↑SPL↑OSR↑TL↓NE↓
Oracle81.8874.1287.4613.061.9375.9562.5384.1614.002.71
Random79.3372.6785.3113.192.2270.9759.4580.1614.923.34
Homing80.2273.6385.6012.512.1471.6560.6080.3313.913.26
Spatial79.6373.1485.6012.992.2271.5660.0180.3314.903.27
Spectral80.6175.1585.8011.952.1171.7861.6880.7613.093.22
+ +Table 3. Comparison of Exploitation Policies. (R2R) + +
Local GoalSR↑Val Seen HouseTest Unseen House
SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑
Oracle54.4237.9663.7211.0148.3828.4562.984.74
Random24.7811.9734.963.0824.197.4135.841.29
Homing42.0427.7248.2310.1835.4019.1851.623.14
Spatial32.3011.6039.381.9026.1110.5839.231.43
Spectral44.6934.8452.658.8939.0925.8048.674.01
+ +Table 4. Comparison of Exploitation Policies. (SOON) + +# 4.6. Ablation Study + +We conduct an ablation study to compare the proposed method against language-triggered hierarchical exploration. Results in the supplementary material show that among the three representation domains, spatial, spectral, and language, the spectral-domain features enhance navigation performance the most. Additionally, to implicate further applications of Meta-Explore in continuous environments, we evaluate our method on the photo-realistic Habitat [50] simulator to solve image-goal navigation and vision-and-language navigation tasks. Implementation details and results are included in the supplementary material. Results show that our method outperforms baselines in both tasks. + +# 5. Conclusion + +We have proposed Meta-Explore, a hierarchical navigation method for VLN, by correcting mistaken short-term actions via efficient exploitation. In the exploitation mode, the agent is directed to a local goal which is inferred to be the closest to the target. A topological map constructed during exploration helps the agent to search and plan the shortest path toward the local goal. To further search beyond the frontier of the map, we present a novel visual representation called scene object spectrum (SOS), which compactly encodes the arrangements and frequencies of nearby objects. Meta-Explore achieves the highest generalization performance for test splits of R2R, SOON, and val split of REVERIE navigation tasks by showing less overfitting and high success rates. We plan to apply Meta-Explore for VLN tasks in continuous environments in our future work. + +# References + +[1] Fengda Zhu, Yi Zhu, Vincent Lee, Xiaodan Liang, and Xiao-jun Chang. Deep learning for embodied vision navigation: A survey. arXiv preprint arXiv:2108.04097, 2021. 1 +[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2018. 1, 2, 3, 6, 8 +[3] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. Touchdown: Natural language navigation and spatial reasoning in visual street environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12538-12547, 2019. 1, 2 +[4] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Conference on Empirical Methods for Natural Language Processing, 2020. 1, 2 +[5] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In *Transactions of the Association for Computational Linguistics*, pages 1862–1872, Florence, Italy, July 2019. Association for Computational Linguistics. 1 +[6] An Yan, Xin Eric Wang, Jiangtao Feng, Lei Li, and William Yang Wang. Cross-lingual vision-language navigation. arXiv preprint arXiv:1910.11301, 2019. 1 +[7] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. In Proceedings of the International Conference on Neural Information Processing Systems, volume 34, pages 652-663. Curran Associates, Inc., 2021. 1 +[8] Jacob Krantz, Erik Wijmans, Arjun Majundar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision and language navigation in continuous environments. In Proceedings of the European Conference on Computer Vision, 2020. 1 +[9] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teplyashin, Karl Moritz Hermann, Mateusz Malinowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 1 +[10] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11773–11781, Apr. 2020. 1 +[11] Arun Balajee Vasudevan, Dengxin Dai, and Luc Van Gool. Talk2nav: Long-range vision-and-language navigation with dual attention and spatial memory. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, volume 129, pages 246-266. Springer, 2021. 1 +[12] Dipendra Misra, Andrew Bennett, Valts Blukis, Eyvind Niklasson, Max Shatkhin, and Yoav Artzi. Mapping instructions to actions in 3D environments with visual goal prediction. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 2667-2678, Brussels, Belgium, October-November 2018. Association for Computational Linguistics. 1 +[13] Ta-Chung Chi, Minmin Shen, Mihail Eric, Seokhwan Kim, and Dilek Hakkani-tur. Justask: An interactive learning framework for vision and language navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2459-2466, 2020. 1 +[14] Yi Wu, Yuxin Wu, Georgia Gkioxari, and Yuandong Tian. Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209, 2018. 1 +[15] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2018. 1 +[16] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12689-12699, 2021. 1, 2, 3, 6, 7, 8 +[17] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 1, 2, 3, 6, 7 +[18] Khanh Nguyen and Hal Daume III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, November 2019. 1, 2 +[19] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2019. 1 +[20] Alane Suhr, Claudia Yan, Jack Schluger, Stanley Yu, Hadi Khader, Marwa Mouallem, Iris Zhang, and Yoav Artzi. Executing instructions in situated collaborative interactions. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 2119-2130, 2019. 1 +[21] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In Proceedings of the International Conference on Learning Representations, 2019. 1, 2, 3, 4, 7, 8 +[22] Chih-Yao Ma, Zuxuan Wu, Ghassan AlRegib, Caiming Xiong, and Zsolt Kira. The regretful agent: Heuristic-aided navigation + +through progress estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6732-6740, 2019. 1, 2, 3, 4, 7, 8 +[23] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10012-10022, 2020. 1, 2, 3, 8 +[24] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537-16547, 2022. 1, 2, 4, 6, 8 +[25] Zhiwei Deng, Karthik Narasimhan, and Olga Russakovsky. Evolving graphical planner: Contextual global planning for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, volume 33, pages 20660-20672, 2020. 1, 2 +[26] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8455-8464, June 2021. 1, 2, 7, 8 +[27] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Proceedings of the Conference on Robot Learning, pages 394-406, 2020. 2 +[28] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, 2018. 2, 8 +[29] Xiujun Li, Chunyuan Li, Qiaolin Xia, Yonatan Bisk, Asli Celikyilmaz, Jianfeng Gao, Noah A Smith, and Yejin Choi. Robust navigation with language pretraining and stochastic sampling. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 1494-1499, 2019. 2 +[30] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Proceedings of the International Conference on Neural Information Processing Systems, 2019. 2 +[31] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1634-1643, 2021. 2 +[32] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. VIn $\text{©}$ bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2, 8 + +[33] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15942-15952, 2021. 2 +[34] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, pages 5834-5847, 2021. 2, 8 +[35] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6741-6749, 2019. 2, 3, 7, 8 +[36] James G March. Exploration and exploitation in organizational learning. Organization science, 2(1):71-87, 1991. 3 +[37] Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O Stanley, and Jeff Clune. First return, then explore. Nature, 590(7847):580-586, 2021. 3 +[38] Miruna Pislar, David Szepesvari, Georg Ostrovski, Diana L Borsa, and Tom Schaul. When should agents explore? In Proceedings of the International Conference on Learning Representations, 2022. 3 +[39] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In Proceedings of the International Conference on Learning Representations, 2020. 3 +[40] Meera Hahn, Devendra Singh Chaplot, Shubham Tulsiani, Mustafa Mukadam, James M Rehg, and Abhinav Gupta. No rl, no simulation: Learning to navigate without navigating. In Proceedings of the International Conference on Neural Information Processing Systems, volume 34, pages 26661-26673. Curran Associates, Inc., 2021. 3 +[41] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3 +[42] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 3 +[43] Wolfgang Stürzl and Hanspeter A Mallot. Efficient visual homing based on fourier transformed panoramic images. Robotics and Autonomous Systems, 54(4):300-313, 2006. 3 +[44] Jean Serra. Mathematical Morphology, pages 1-16. Springer International Publishing, Cham, 2020. 3 +[45] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 5100-5111, 2019. 4 + +[46] Stéphane Ross, Geoffrey J Gordon, and J Andrew Bagnell. No-regret reductions for imitation learning and structured prediction. arXiv preprint arXiv:1011.0686, 2010. 6 +[47] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proceedings of the International Conference on Machine Learning, pages 1928-1937, 2016. 6 +[48] Peter Anderson, Angel Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, et al. On evaluation of embodied navigation agents. arXiv preprint arXiv:1807.06757, 2018. 7 +[49] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6629-6638, 2019. 8 +[50] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Proceedings of the International Conference on Neural Information Processing Systems, 2021. 8 \ No newline at end of file diff --git a/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/images.zip b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7e5e61916fe78901bc7ae8459fc724ad5e5202ea --- /dev/null +++ b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:864221064fc5d7b578f38869bcdb1fde2d7c2acd4de5ba3a198ef6734b9f26ea +size 483067 diff --git a/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/layout.json b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4be535fed7f6f0bc143c9b10dc0266f1ac527e6e --- /dev/null +++ b/2023/Meta-Explore_ Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding/layout.json @@ -0,0 +1,9588 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 136, + 511, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 136, + 511, + 172 + ], + "spans": [ + { + "bbox": [ + 66, + 136, + 511, + 172 + ], + "type": "text", + "content": "Meta-Explore: Exploratory Hierarchical Vision-and-Language Navigation Using Scene Object Spectrum Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "spans": [ + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "content": "Minyoung Hwang" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "content": ", Jaeyeon Jeong" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "content": ", Minsoo Kim" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "content": ", Yoonseon Oh" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "content": ", Songhwai Oh" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "content": "Electrical and Computer Engineering and ASRI, Seoul National University \n" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "content": "Department of Electronic Engineering, Hanyang University \n" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 191, + 483, + 247 + ], + "type": "text", + "content": "Interdisciplinary Major in Artificial Intelligence, Seoul National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 250, + 526, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 250, + 526, + 261 + ], + "spans": [ + { + "bbox": [ + 52, + 250, + 526, + 261 + ], + "type": "text", + "content": "{minyoung.hwang, jaeyeon.jeong}@rlllab.snu.ac.kr, goldbird5@snu.ac.kr, yoh21@hanyang.ac.kr, songhwai@snu.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 275, + 186, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 275, + 186, + 287 + ], + "spans": [ + { + "bbox": [ + 139, + 275, + 186, + 287 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 298, + 282, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 298, + 282, + 624 + ], + "spans": [ + { + "bbox": [ + 45, + 298, + 282, + 624 + ], + "type": "text", + "content": "The main challenge in vision-and-language navigation (VLN) is how to understand natural-language instructions in an unseen environment. The main limitation of conventional VLN algorithms is that if an action is mistaken, the agent fails to follow the instructions or explores unnecessary regions, leading the agent to an irrecoverable path. To tackle this problem, we propose Meta-Explore, a hierarchical navigation method deploying an exploitation policy to correct misled recent actions. We show that an exploitation policy, which moves the agent toward a well-chosen local goal among unvisited but observable states, outperforms a method which moves the agent to a previously visited state. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. The key to our approach is understanding the object placements around the agent in spectral-domain. Specifically, we present a novel visual representation, called scene object spectrum (SOS), which performs category-wise 2D Fourier transform of detected objects. Combining exploitation policy and SOS features, the agent can correct its path by choosing a promising local goal. We evaluate our method in three VLN benchmarks: R2R, SOON, and REVERIE. Meta-Explore outperforms other baselines and shows significant generalization performance. In addition, local goal search using the proposed spectral-domain SOS features significantly improves the success rate by " + }, + { + "bbox": [ + 45, + 298, + 282, + 624 + ], + "type": "inline_equation", + "content": "17.1\\%" + }, + { + "bbox": [ + 45, + 298, + 282, + 624 + ], + "type": "text", + "content": " and SPL by " + }, + { + "bbox": [ + 45, + 298, + 282, + 624 + ], + "type": "inline_equation", + "content": "20.6\\%" + }, + { + "bbox": [ + 45, + 298, + 282, + 624 + ], + "type": "text", + "content": " against the state-of-the-art method of the SOON benchmark. Project page: https://rllab-snu.github.io/projects/Meta-Explore/doc.html" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 631, + 124, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 631, + 124, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 631, + 124, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 648, + 279, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 279, + 672 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 279, + 672 + ], + "type": "text", + "content": "Visual navigation in indoor environments has been studied widely and shown that an agent can navigate in unex-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 311, + 275, + 518, + 383 + ], + "blocks": [ + { + "bbox": [ + 311, + 275, + 518, + 383 + ], + "lines": [ + { + "bbox": [ + 311, + 275, + 518, + 383 + ], + "spans": [ + { + "bbox": [ + 311, + 275, + 518, + 383 + ], + "type": "image", + "image_path": "3edce37aa747505b4e8eaf55c9d230dc101641e3984131ef27e1d082eb0b25a1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "lines": [ + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "spans": [ + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "text", + "content": "Figure 1. Hierarchical Exploration. At each episode, a natural-language instruction is given to the agent to navigate to a goal location. The agent explores the environment and constructs a topological map by recording visited nodes " + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "text", + "content": " and next step reachable nodes " + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "text", + "content": ". Each node consists of the position of the agent and visual features. " + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "text", + "content": " denotes the observation at time " + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 386, + 530, + 455 + ], + "type": "text", + "content": ". The agent chooses an unvisited local goal to solve the regretful exploration problem." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 460, + 531, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 460, + 531, + 589 + ], + "spans": [ + { + "bbox": [ + 296, + 460, + 531, + 589 + ], + "type": "text", + "content": "plored environments [1]. By recognizing the visual context and constructing a map, an agent can explore the environment and solve tasks such as moving towards a goal or following a desired trajectory. With the increasing development in human language understanding, vision-and-language navigation (VLN) [2] has enabled robots to communicate with humans using natural languages. The high degree of freedom in natural language instructions allows VLN to expand to various tasks, including (1) following fine-grained step-by-step instructions [2-13] and (2) reaching a target location described by goal-oriented language instructions [14-20]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 590, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 590, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 590, + 532, + 730 + ], + "type": "text", + "content": "A challenging issue in VLN is the case when an action is mistaken with respect to the given language instruction [21-26]. For instance, if the agent is asked to turn right at the end of the hallway but turns left, the agent may end up in irrecoverable paths. Several existing studies solve this issue via hierarchical exploration, where the high-level planner decides when to explore and the low-level planner chooses what actions to take. If the high-level planner chooses to explore, the agent searches unexplored regions, and if it chooses to exploit, the agent executes the best action based on the previous exploration. Prior work [21-23] returns the agent to the last successful state and resumes exploration." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 680, + 279, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 680, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 680, + 279, + 730 + ], + "type": "text", + "content": "This work was in part supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-01190, [SW Star Lab] Robot Learning: Efficient, Safe, and Socially-Acceptable Machine Learning, " + }, + { + "bbox": [ + 45, + 680, + 279, + 730 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 45, + 680, + 279, + 730 + ], + "type": "text", + "content": ", and No.2022-0-00907, Development of AI Bots Collaboration Platform and Self-Organizing AI, " + }, + { + "bbox": [ + 45, + 680, + 279, + 730 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 45, + 680, + 279, + 730 + ], + "type": "text", + "content": "). (Corresponding authors: Yoonseon Oh and Songhwai Oh.)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6683" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 257 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 257 + ], + "type": "text", + "content": "However, such methods take a heuristic approach because the agent only backtracks to a recently visited location. The agent does not take advantage of the constructed map and instead naively uses its recent trajectory for backtracking. Another recent work [26] suggests graph-based exploitation, which uses a topological map to expand the action space in global planning. Still, this method assumes that the agent can directly jump to a previously visited node. Since this method can perform a jump action at every timestep, there is no trigger that explicitly decides when to explore and when to exploit. Therefore, we address the importance of time scheduling for exploration-exploitation and efficient global planning using a topological map to avoid reexploring visited regions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 257, + 280, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 257, + 280, + 687 + ], + "spans": [ + { + "bbox": [ + 45, + 257, + 280, + 687 + ], + "type": "text", + "content": "We expand the notion of hierarchical exploration by proposing Meta-Explore, which not only allows the high-level planner to choose when to correct misled local movements but also finds an unvisited state inferred to be close to the global goal. We illustrate the overview of hierarchical exploration in Figure 1. Instead of backtracking, we present an exploitation method called local goal search. We show that it is more efficient to plan a path to a local goal, which is the most promising node from the unvisited but reachable nodes. We illustrate the difference between conventional backtracking and local goal search in Figure 2. Based on our method, we show that exploration and exploitation are not independent and can complement each other: (1) to overtake regretful explorations, the agent can perform exploitation and (2) the agent can utilize the constructed topological map for local goal search. We also highlight the demand for imagining regretful explorations with semantically meaningful clues. Most VLN tasks require a level of understanding objects nearby the agent, but previous studies simply encode observed panoramic or object images [2, 3, 16-18, 21-35]. In this paper, we present a novel semantic representation of the scene called scene object spectrum (SOS), which is a matrix containing the arrangements and frequencies of objects from the visual observation at each location. Using SOS features, we can sufficiently estimate the context of the environment. We show that the proposed spectral-domain SOS features manifest better linguistic interpretability than conventional spatial-domain visual features. Combining exploitation policy and SOS features, we design a navigation score that measures the alignment between a given language instruction and a corrected trajectory toward a local goal. The agent compares local goal candidates and selects a near-optimal candidate with the highest navigation score from corrected trajectories. This involves high-level reasoning related to the landmarks (e.g., bedroom and kitchen) and objects (e.g., table and window) that appear in the instructions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 687, + 260, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 687, + 260, + 698 + ], + "spans": [ + { + "bbox": [ + 57, + 687, + 260, + 698 + ], + "type": "text", + "content": "The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 57, + 706, + 279, + 730 + ], + "type": "text", + "content": "- We propose a hierarchical navigation method called Meta-Explore, deploying an exploitation policy to cor" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 298, + 103, + 521, + 239 + ], + "blocks": [ + { + "bbox": [ + 298, + 103, + 521, + 239 + ], + "lines": [ + { + "bbox": [ + 298, + 103, + 521, + 239 + ], + "spans": [ + { + "bbox": [ + 298, + 103, + 521, + 239 + ], + "type": "image", + "image_path": "cae102e992b908d5afb9a1f840f5b5aa8a60047e5e8a448b980f2d3f4eab3042.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 247, + 530, + 287 + ], + "lines": [ + { + "bbox": [ + 296, + 247, + 530, + 287 + ], + "spans": [ + { + "bbox": [ + 296, + 247, + 530, + 287 + ], + "type": "text", + "content": "Figure 2. Local Goal Search for Exploitation. The local goal is likely to be chosen as the closest node to the global goal. Existing methods only backtrack to a visited node (left). We expand the searchable area by including unvisited but reachable nodes (right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 293, + 530, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 293, + 530, + 329 + ], + "spans": [ + { + "bbox": [ + 315, + 293, + 530, + 329 + ], + "type": "text", + "content": "rect misled recent actions. The agent searches for an appropriate local goal instead of reversing the recent action sequence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 331, + 530, + 495 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 308, + 331, + 530, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 530, + 411 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 530, + 411 + ], + "type": "text", + "content": "- In the exploitation mode, the agent uses a novel scene representation called scene object spectrum (SOS), which contains the spectral information of the object placements in the scene. SOS features provide semantically meaningful clues to choose a near-optimal local goal and help the agent to solve the regretful exploration problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 415, + 530, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 530, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 530, + 495 + ], + "type": "text", + "content": "- We evaluate our method on three VLN benchmarks: R2R [2], SOON [16], and REVERIE [17]. The experimental results show that the proposed method, MetaExplore, improves the success rate and SPL in test splits of R2R, SOON and val split of REVERIE. The proposed method shows better generalization results compared to all baselines." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 501, + 381, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 501, + 381, + 513 + ], + "spans": [ + { + "bbox": [ + 297, + 501, + 381, + 513 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 516, + 468, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 516, + 468, + 529 + ], + "spans": [ + { + "bbox": [ + 297, + 516, + 468, + 529 + ], + "type": "text", + "content": "2.1. Vision-and-Language Navigation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 532, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 532, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 532, + 530, + 730 + ], + "type": "text", + "content": "In VLN, an agent encodes the natural language instructions and follows the instructions, which can be either (1) a fine-grained step-by-step instruction the agent can follow [2-4], (2) a description of the target object and location [16, 17], or (3) additional guidance given to the agent [18, 27]. These tasks require the agent to recognize its current location using some words in the natural-language instructions. Prior work [2, 28-31] show that an agent can align visual features to language instructions via neural networks and use the multimodal output embeddings to generate a suitable action at each timestep. Most VLN methods utilize cross-modal attention, either with recurrent neural networks [2, 28] or with transformer-based architectures [29-31]. For sequential action prediction, Hong et al. [32] further use recurrent units inside transformer architectures, while Pashevich et al. [33] and Chen et al. [34] use additional transformers to embed past observations and actions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6684" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 95, + 183, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 95, + 183, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 95, + 183, + 107 + ], + "type": "text", + "content": "2.2. Exploration-Exploitation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 109, + 279, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 109, + 279, + 260 + ], + "spans": [ + { + "bbox": [ + 45, + 109, + 279, + 260 + ], + "type": "text", + "content": "In an unseen environment, the agent must maximize the return without knowing the true value functions. One of the solutions to this problem is to switch back and forth between exploration and exploitation [36]. In the exploration mode, the agent gathers more information about the environment. On the other hand, the agent uses information collected during exploration and chooses the best action for exploitation. Ecoffet et al. [37] reduced the exploration step by archiving the states and exploring again from the successful states. Pislar et al. [38] addressed the various scheduling policies and demonstrated their method on Atari games. Recent work [39, 40] successfully demonstrates the effectiveness of hierarchical exploration in image-goal navigation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 261, + 279, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 261, + 279, + 424 + ], + "spans": [ + { + "bbox": [ + 45, + 261, + 279, + 424 + ], + "type": "text", + "content": "Like commonly used greedy navigation policies, VLN tasks also deal with the problem of maximizing the chance to reach the goal without knowing the ground truth map. Several VLN methods employ the concept of exploitation to tackle this problem. Ke et al. [35] look forward to several possible future trajectories and decide whether to backtrack or not and where to backtrack. Others [21-23] estimate the progress to tell whether the agent becomes lost and make the agent backtrack to a previously visited location to restart exploration. However, previous studies do not take into account what should be done in the exploitation mode. In order to handle this problem, we propose a hierarchical navigation method which determines the scheduling between exploration and exploitation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 173, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 173, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 173, + 437 + ], + "type": "text", + "content": "2.3. Visual Representations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 439, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 439, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 439, + 279, + 730 + ], + "type": "text", + "content": "Popular visual encoding methods via ResNet [41] and ViT [42] can be trained to learn rotation-invariant visual features. Both methods learn to extract visual features with high information gain for global and local spatial information. The high complexity of the features leads to low interpretability of the scene and therefore requires the agent to use additional neural networks or complex processing to utilize them. On the other hand, traditional visual representation methods such as Fourier transform use spectral analysis, which is highly interpretable and computationally efficient. One drawback of the traditional methods is that they fail to maximize the information gain. Nonetheless, an appropriate use of essential information can be helpful for high-level decision making and enables more straightforward interpretation and prediction of the visual features. One traditional navigation method, Sturz et al. [43] used Fourier transform to generate rotation-invariant visual features. However, no research has transformed the spectral information of the detected objects to represent high-level semantics from visual observations. Focusing on the fact that 2D Fourier transform can extract morphological properties of images [44], we can find out the shape or structure of detected objects through 2D Fourier transform. In this paper, we decompose the object mask into binary masks by object categories and perform a 2D Fourier transform on each binary mask." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 297, + 105, + 351, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 105, + 351, + 117 + ], + "spans": [ + { + "bbox": [ + 297, + 105, + 351, + 117 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 121, + 417, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 121, + 417, + 133 + ], + "spans": [ + { + "bbox": [ + 297, + 121, + 417, + 133 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "spans": [ + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": "We deal with VLN in discrete environments, where the environment is given as an undirected graph " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "G_{e} = \\{V,E\\}" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": " denotes a set of " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": " navigable nodes, " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "\\{V_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": " is the adjacency matrix describing connectivity among the nodes in " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": ". We denote the observation at node " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": ". The agent uses a panoramic RGB image observation " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "o_{t}" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": " and current node " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": ", which are collected at time " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": ". The agent either moves to a neighboring node or executes a stop action. " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": " denotes the action at time " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": ". The objectives of VLN are categorized as follows: (1) to follow language instructions [2] and (2) to find a target object described by language instructions in a fixed time " + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 296, + 136, + 530, + 370 + ], + "type": "text", + "content": " [16, 17]. We present a general hierarchical exploration method that can be applied to both tasks. We also enhance the navigation policy by extracting cross-domain visual representations from the environments, i.e., spatial-domain and spectral-domain representations. To balance the information loss and interpretability of the visual feature, we adopt multi-channel fast Fourier transform (FFT) to encode semantic masks of the detected objects into category-wise spectral-domain features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 373, + 383, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 373, + 383, + 386 + ], + "spans": [ + { + "bbox": [ + 297, + 373, + 383, + 386 + ], + "type": "text", + "content": "3.2. Meta-Explore" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "spans": [ + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "type": "text", + "content": "We design a learnable hierarchical exploration method for VLN called Meta-Explore, which decides (1) when to explore or exploit and (2) a new imagined local goal to seek during exploitation. The overall network architecture of the proposed Meta-Explore is shown in Figure 3. Given a language instruction " + }, + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "type": "text", + "content": ", the agent navigates in the environment until it finds the target described in " + }, + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "type": "text", + "content": ". Meta-Explore consists of a mode selector and two navigation modules corresponding to two modes: exploration and exploitation. At each timestep, the mode selector chooses to explore or exploit. At " + }, + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 296, + 388, + 530, + 679 + ], + "type": "text", + "content": ", the mode is initialized to exploration. In the exploration mode, the agent outputs an action toward a neighboring node to move the agent toward the goal. When the mode selector recognizes that the agent is not following the instruction successfully, the mode is switched to exploitation. In the exploitation mode, the agent seeks a new local goal with the highest correspondence against the language instructions from the previously unvisited candidate nodes using spectral-domain visual features. The agent moves toward the local goal by planning a path. After the agent arrives at the local goal, the mode is reset to exploration. The explore-exploit switching decision occurs through the mode selector by estimating the probability to explore. The agent repeats this explore-exploit behavior until it determines that the target is found and decides to stop." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 682, + 389, + 692 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 682, + 389, + 692 + ], + "spans": [ + { + "bbox": [ + 297, + 682, + 389, + 692 + ], + "type": "text", + "content": "3.2.1 Mode Selector" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": "At time " + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": ", the agent observes visual features about the current node " + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": " and several reachable nodes. We call the nodes reachable at the current timestep as candidate nodes. " + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "inline_equation", + "content": "n_{cand}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6685" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 105, + 483, + 290 + ], + "blocks": [ + { + "bbox": [ + 97, + 105, + 483, + 290 + ], + "lines": [ + { + "bbox": [ + 97, + 105, + 483, + 290 + ], + "spans": [ + { + "bbox": [ + 97, + 105, + 483, + 290 + ], + "type": "image", + "image_path": "beec1b93785ed13d1fb664fb0d718a97f7174f0600af642877fcb3add3493e74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "lines": [ + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "spans": [ + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "text", + "content": "Figure 3. Network Architecture. Three types of visual features: panoramic (yellow), object image (aquamarine), and object spectrum (red) are encoded. The color in each parenthesis denotes the color describing the corresponding feature. The cross-modal transformer encodes language and spatial visual features as hidden state " + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "text", + "content": ". A mode selector gives explore or exploit command to the agent by predicting the explore probability " + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "inline_equation", + "content": "P_{\\text{explore}}" + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "text", + "content": ". The selected navigation module outputs an action " + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "text", + "content": " from the possible " + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "inline_equation", + "content": "n_{\\text{cand}}" + }, + { + "bbox": [ + 45, + 298, + 532, + 339 + ], + "type": "text", + "content": " candidate nodes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "spans": [ + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "text", + "content": "denotes the number of candidate nodes. We use a cross-modal transformer with " + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "inline_equation", + "content": "n_{L}" + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "text", + "content": " layers to relate visual observations to language instructions. The cross-modal transformer takes the visual features of nodes in the constructed topological map at time " + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "text", + "content": ", and outputs cross-modal embedding " + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "text", + "content": " to encode visual observations with " + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "text", + "content": ". We concatenate location encoding and history encoding [24] to the visual features as node features to consider the relative pose from " + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 45, + 344, + 280, + 471 + ], + "type": "text", + "content": " and the last visited timestep of each node, respectively. Each word is encoded via a pretrained language encoder [45], which is used for general vision-language tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "spans": [ + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": "The cross-modal transformer consists of cross-attention layer L2V_Atn(\\hat{W},\\hat{V}) = Softmax(\\hat{W}\\Theta_q^\\dagger (\\hat{V}\\Theta_k^\\dagger)^T /\\sqrt{d})\\hat{V}\\Theta_v^\\dagger and self-attention layer SelfAttn(X) = Softmax " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "(X\\Theta_{q}(X\\Theta_{k})^{T}+" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "A\\Theta_e + b_e) / \\sqrt{d})X\\Theta_v" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "\\hat{W},\\hat{V},X" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " denote word, visual, node representations and adjacency matrix of " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " respectively. The (query, key, value) weight matrices of self-attention and cross-attention layers are denoted as " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "(\\Theta_q,\\Theta_k,\\Theta_v)" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "(\\Theta_q^\\dagger ,\\Theta_k^\\dagger ,\\Theta_v^\\dagger)" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " , respectively. The final crossmodal embedding at time " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " after passing through " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "n_L" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " transformer layers is denoted as " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " . To encourage the monotonic increasing relationship between language and visual attentions at each timestep, we define a correlation loss " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "L_{corr} = \\sum_{t = 1}^{T}\\| \\mathrm{L2V\\_Attn} - I_{n_x}\\| _1" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " for training the crossmodal transformer, where " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "n_x" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " denotes the dimension of the " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "I_{n_x}" + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "text", + "content": " denotes an identity matrix of size " + }, + { + "bbox": [ + 45, + 473, + 279, + 648 + ], + "type": "inline_equation", + "content": "n_x\\times n_x" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": "As illustrated in Figure 4, the mode selector estimates the probability to explore " + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "inline_equation", + "content": "P_{\\text{explore}}" + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": " given the cross-modal hidden state " + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "inline_equation", + "content": "H_t" + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": ". We denote the mode selector as " + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "inline_equation", + "content": "S_{\\text{mode}}" + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": " and use a two-layer feed-forward neural network. Given " + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "inline_equation", + "content": "H_t" + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "inline_equation", + "content": "S_{\\text{mode}}" + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": " outputs the exploration probability as " + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "inline_equation", + "content": "P_{\\text{explore}} = 1 - S_{\\text{mode}}(H_t)" + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "inline_equation", + "content": "P_{\\text{explore}} \\geq 0.5" + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": ", the exploration policy outputs a probability distribution for reachable nodes at the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 344, + 530, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 344, + 530, + 448 + ], + "spans": [ + { + "bbox": [ + 296, + 344, + 530, + 448 + ], + "type": "text", + "content": "next step. At time " + }, + { + "bbox": [ + 296, + 344, + 530, + 448 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 296, + 344, + 530, + 448 + ], + "type": "text", + "content": ", the agent moves to the node with the highest probability. If " + }, + { + "bbox": [ + 296, + 344, + 530, + 448 + ], + "type": "inline_equation", + "content": "P_{\\text{explore}} < 0.5" + }, + { + "bbox": [ + 296, + 344, + 530, + 448 + ], + "type": "text", + "content": ", the agent determines that the current trajectory is regretful, so the agent should traverse to find a local goal, which is the most likely to be the closest node to the global goal. The exploitation policy mainly utilizes object-level features to search for the local goal with high-level reasoning. After the local goal is chosen, the path planning module outputs an action following the shortest path to the local goal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "spans": [ + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "text", + "content": "To train the mode selector, we require additional demonstration data other than the ground truth trajectory, such that it switches between exploration and exploitation. We generate the demonstration data from the ground truth trajectories, with additional detours. For the detours, we stochastically select candidate nodes other than the ground truth paths and add the trajectory that returns to the current viewpoint. The imitation learning loss for training the mode selector is defined as " + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "inline_equation", + "content": "L_{mode} = \\sum_{t=1}^{T} \\mathbb{1}(m_t = \\mathrm{gt}_t)" + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "text", + "content": " is the mode of the agent, 0 for exploitation and 1 for exploration. " + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{gt}_t" + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "text", + "content": " is 1 if the current node is in the shortest ground truth trajectory and " + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{gt}_t = 0" + }, + { + "bbox": [ + 296, + 449, + 531, + 589 + ], + "type": "text", + "content": ", otherwise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 596, + 413, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 596, + 413, + 607 + ], + "spans": [ + { + "bbox": [ + 297, + 596, + 413, + 607 + ], + "type": "text", + "content": "3.2.2 Exploration Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 609, + 530, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 609, + 530, + 692 + ], + "spans": [ + { + "bbox": [ + 296, + 609, + 530, + 692 + ], + "type": "text", + "content": "In the exploration mode, the agent follows the following sequential operations: topological map construction, self-monitoring, and an exploration policy. To improve the exploration, we adopt self-monitoring [21] to predict the current progress of exploration to enhance the exploration policy itself. Prior work [21, 22] has shown that auxiliary loss using self-monitoring can regularize the exploration policy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 695, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 695, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 695, + 531, + 730 + ], + "type": "text", + "content": "Topological Map Construction. The agent constructs graph " + }, + { + "bbox": [ + 296, + 695, + 531, + 730 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 296, + 695, + 531, + 730 + ], + "type": "text", + "content": " by classifying nodes into two types: (1) visited nodes and (2) unvisited but observable nodes. At current" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6686" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 103, + 153, + 169 + ], + "blocks": [ + { + "bbox": [ + 59, + 103, + 153, + 169 + ], + "lines": [ + { + "bbox": [ + 59, + 103, + 153, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 153, + 169 + ], + "type": "image", + "image_path": "a84bd4402484b8d5c99c18b4e3a47c29d64252e158398d83efef5e5d34ea4b21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 169, + 151, + 250 + ], + "blocks": [ + { + "bbox": [ + 59, + 169, + 151, + 250 + ], + "lines": [ + { + "bbox": [ + 59, + 169, + 151, + 250 + ], + "spans": [ + { + "bbox": [ + 59, + 169, + 151, + 250 + ], + "type": "image", + "image_path": "49a1a66bfa018bbdb7f02df07e6ecca1068f310984d4e327817c647b63e6fe41.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 252, + 139, + 261 + ], + "lines": [ + { + "bbox": [ + 76, + 252, + 139, + 261 + ], + "spans": [ + { + "bbox": [ + 76, + 252, + 139, + 261 + ], + "type": "text", + "content": "Exploration Module" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 262, + 279, + 301 + ], + "lines": [ + { + "bbox": [ + 45, + 262, + 279, + 301 + ], + "spans": [ + { + "bbox": [ + 45, + 262, + 279, + 301 + ], + "type": "text", + "content": "Figure 4. Navigation Modules. Mode selector estimates " + }, + { + "bbox": [ + 45, + 262, + 279, + 301 + ], + "type": "inline_equation", + "content": "P_{\\text{explore}}" + }, + { + "bbox": [ + 45, + 262, + 279, + 301 + ], + "type": "text", + "content": ", i.e., the probability to explore, and chooses between exploration and exploitation modules. The selected navigation module outputs the next action " + }, + { + "bbox": [ + 45, + 262, + 279, + 301 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 45, + 262, + 279, + 301 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 154, + 103, + 267, + 250 + ], + "blocks": [ + { + "bbox": [ + 154, + 103, + 267, + 250 + ], + "lines": [ + { + "bbox": [ + 154, + 103, + 267, + 250 + ], + "spans": [ + { + "bbox": [ + 154, + 103, + 267, + 250 + ], + "type": "image", + "image_path": "5c4a582aac44a3dbcdcd77c052679afa3b2a36df0d9cf7919d8eaa275f55f1a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 252, + 241, + 261 + ], + "lines": [ + { + "bbox": [ + 177, + 252, + 241, + 261 + ], + "spans": [ + { + "bbox": [ + 177, + 252, + 241, + 261 + ], + "type": "text", + "content": "Exploitation Module" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "spans": [ + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "text", + "content": "time " + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "text", + "content": ", the agent at node " + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "inline_equation", + "content": "v_{t} \\in \\{V_{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "text", + "content": " observes " + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "inline_equation", + "content": "N(v_{t})" + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "text", + "content": " neighbor nodes as next step candidates at time " + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "text", + "content": ". The visited nodes consist of visual features of their own and the neighboring nodes from panoramic RGB observations. The unvisited nodes can be observed only if they are connected to at least one visited node. The topological map records the positions and visual features of observed nodes at each timestep. By knowing the positions of nodes in " + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 45, + 315, + 279, + 421 + ], + "type": "text", + "content": ", the agent can plan the shortest path trajectory between two nodes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "spans": [ + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": "Self-Monitoring. We use a progress monitor to estimate the current navigation progress at each episode. Self-monitoring via estimating current progress helps the agent choose the next action that can increase the progress. The estimated progress " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "\\hat{p}_t = F_{\\text{progress}}(H_t)" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": " is the output of a feed-forward neural network, given " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "H_t" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": " as input. We measure the ground truth progress " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": " as the ratio between the current distance to the goal and the shortest path length of the episode subtracted from 1, described as " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "1 - \\frac{d_{geo}(v_t, v_{goal})}{d_{geo}(v_0, v_{goal})}" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "d_{geo}(a, b)" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": " is the geodesic distance between " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "v_0, v_t" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": " denote initial, current, and goal positions, respectively. We add progress loss " + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "inline_equation", + "content": "L_{\\text{progress}} = \\sum_{t=1}^{T} (\\hat{p}_t - p_t)^2" + }, + { + "bbox": [ + 45, + 426, + 279, + 581 + ], + "type": "text", + "content": " to train the progress monitor while training the exploration policy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "spans": [ + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "content": "Exploration Policy. The exploration policy " + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "inline_equation", + "content": "F_{explore}" + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "content": " estimates the probability of moving to the candidate nodes at the next step. The agent chooses the action " + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "content": " based on the estimated probability distribution among candidate nodes, described as " + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "inline_equation", + "content": "a_t = \\arg \\max_{V_i} (F_{explore}([H_t]_i))" + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "inline_equation", + "content": "F_{explore}" + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "content": " is implemented via a two-layer feed-forward network with the cross-modal hidden state " + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "inline_equation", + "content": "H_t" + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "content": " given as input. The output of " + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "inline_equation", + "content": "F_{explore}" + }, + { + "bbox": [ + 45, + 586, + 280, + 738 + ], + "type": "text", + "content": " becomes a probability distribution over possible actions. To only consider unvisited nodes, we mask out the output for visited nodes. For training, we sample the next action from the probability distribution instead of choosing a node with the highest probability. We describe the training details in Section 3.3." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 103, + 524, + 223 + ], + "blocks": [ + { + "bbox": [ + 305, + 103, + 524, + 223 + ], + "lines": [ + { + "bbox": [ + 305, + 103, + 524, + 223 + ], + "spans": [ + { + "bbox": [ + 305, + 103, + 524, + 223 + ], + "type": "image", + "image_path": "f87060c14ebff4c2fc1cd3db0294d7b820c3aa365ea34d85d467c07f4f0fc055.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 229, + 530, + 271 + ], + "lines": [ + { + "bbox": [ + 296, + 229, + 530, + 271 + ], + "spans": [ + { + "bbox": [ + 296, + 229, + 530, + 271 + ], + "type": "text", + "content": "Figure 5. Scene Object Spectrum (SOS). The agent calculates scene object spectrum (SOS) features for efficient exploitation. SOS features incorporate semantic information observed in a single panoramic image by performing category-wise 2D FFT." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 280, + 415, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 280, + 415, + 292 + ], + "spans": [ + { + "bbox": [ + 297, + 280, + 415, + 292 + ], + "type": "text", + "content": "3.2.3 Exploitation Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 294, + 530, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 294, + 530, + 411 + ], + "spans": [ + { + "bbox": [ + 296, + 294, + 530, + 411 + ], + "type": "text", + "content": "In the exploitation mode, the agent requires high-level reasoning with identifiable environmental clues to imagine regretful exploration cases. To find clues in an object-level manner, we present a novel visual representation by capturing object information in the spectral-domain. The novel representation is more easily predictable than spatial features such as RGB image embeddings. The agent can take advantage of the predictability by expanding the searchable area to find a local goal. We choose the local goal as the closest node to the global goal in the feature space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "spans": [ + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": "Spectral-Domain Visual Representations. Common navigation policies can lead the agent toward the node with the highest similarity to the target. However, even with a good learned policy, the agent can act in a novice manner in unseen environments. In this paper, we seek extra information from the environment for generalizable high-level reasoning to resolve the issue. As illustrated in Figure 5, scene object spectrum (SOS) incorporates semantic information observed in a single panoramic image by generating a semantic mask for each object category and applying Fourier transform to each semantic mask. The semantic mask for object class " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " is calculated as a binary mask " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "[m_t^k]_{ij}" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " that detects the object at pixel " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": ". Suppose there are a total of " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " object categories. When multiple objects are detected for one object category, the binary mask appears as a union of the bounding boxes of the detected objects. We define FFT as a channel-wise 2D fast Fourier transform that receives " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " binary semantic masks and outputs " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " spectral-domain features, where " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " is the number of object classes. Then, SOS feature " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "\\vec{S}_t = [s_t^1,\\dots,s_t^K ]^T" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " can be defined as " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "s_t^k = \\log |\\mathbf{FFT}(m_t^k)|" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": ". For simplicity, we perform mean pooling on the vertical spectral axis and normalize the output. The final SOS feature has shape " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "(K,\\eta)" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 296, + 415, + 531, + 693 + ], + "type": "text", + "content": " is the maximum horizontal frequency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": "Local Goal Search Using Semantic Clues. We argue that returning to a previously visited node does not guarantee the agent escapes from the local optima. Instead of backtracking" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6687" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 210 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 210 + ], + "type": "text", + "content": "to a previously visited node, the agent searches for a local goal to move towards. If the agent plans a path and moves towards the local goal, the agent does not need to repeat unnecessary actions in visited regions after the exploitation ends. Additionally, searching for a local goal takes full advantage of the topological map by utilizing the connections among the observed nodes. To expand the searchable area further, we let the agent choose the local goal from previously unvisited and unchosen candidate nodes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "spans": [ + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": "To choose a local goal, we first score the corrected trajectories to measure the alignment with the language instruction " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": ". We use SOS features as semantic environmental clues to estimate the navigation score " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "S_{\\text{nav}}" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " of the corrected trajectory, which is the shortest path trajectory from the initial node to the local goal in the constructed topological map. To simplify, we convert the language instruction into a list of objects " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "W^{o} = [w_{1}^{o}, \\dots, w_{B}^{o}]" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "B[\\leq K]" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " object categories (e.g., desk, cabinet, and microwave). We approximate the corresponding reference SOS features as " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "[\\hat{\\delta}(w_{1}^{o}), \\dots, \\hat{\\delta}(w_{B}^{o})]" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " where the " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " row of " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}(w_{k}^{o})" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "[\\hat{\\delta}(w_{k}^{o})]_{i} = \\mathbb{1}(k = i)\\lambda(\\delta(w_{k}^{o}))\\mathrm{sinc}\\left(\\frac{j}{2} - \\frac{\\eta}{4}\\right)" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "\\lambda(\\delta(w_{k}^{o}))" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " denotes the average width of detected bounding boxes of object " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "w_{k}^{o}" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " in the environment. A detailed approximation process is explained in the supplementary material. To simulate a corrected trajectory " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{T}' = (v_{1}', \\dots, v_{t'})" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": ", we calculate the SOS features " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "[\\vec{S}_1', \\dots, \\vec{S}_{t'}']" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " corresponding to the nodes in " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": ". We measure the similarity between two object spectrum features via the cosine similarity of the flattened vectors. Finally, the navigation score " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "S_{\\text{nav}}" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 45, + 211, + 280, + 437 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 442, + 279, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 279, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 279, + 500 + ], + "type": "interline_equation", + "content": "S _ {n a v} \\left(\\mathcal {T} ^ {\\prime}\\right) = \\frac {\\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {t ^ {\\prime}} \\left(\\frac {\\hat {\\delta} \\left(w _ {i} ^ {o}\\right)}{\\left| \\hat {\\delta} \\left(w _ {i} ^ {o}\\right) \\right|} \\cdot \\frac {\\vec {S} _ {j} ^ {\\prime}}{\\left| \\vec {S} _ {j} ^ {\\prime} \\right|}\\right) \\left(\\left(\\hat {\\delta} \\left(w _ {i} ^ {o}\\right) - \\hat {\\delta} (\\bar {w} ^ {o})\\right) \\cdot \\left(\\vec {S} _ {j} ^ {\\prime} - \\bar {\\vec {S}} ^ {\\prime}\\right)\\right)}{\\sqrt {\\frac {t ^ {\\prime}}{B} \\cdot \\sum_ {i = 1} ^ {B} \\left(\\hat {\\delta} \\left(w _ {i} ^ {o}\\right) - \\hat {\\delta} (\\bar {w} ^ {o})\\right) ^ {2} \\sum_ {j = 1} ^ {t ^ {\\prime}} \\left(\\vec {S} _ {j} ^ {\\prime} - \\bar {\\vec {S}} ^ {\\prime}\\right) ^ {2}}}, \\tag {1}", + "image_path": "fac47ca30d53af19f277e02efdb62a55a491f7baed194975a45e676845fac599.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "spans": [ + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\delta} (\\overline{w}^o)" + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "inline_equation", + "content": "\\vec{S}'" + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "text", + "content": " denote the average values of SOS features " + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\delta} (w_i^o)" + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "inline_equation", + "content": "\\vec{S}_j^\\prime" + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "text", + "content": ", respectively. This equation can also be interpreted as a pseudo correlation-coefficient function between object list " + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "inline_equation", + "content": "W^{o}" + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "text", + "content": " and trajectory " + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 45, + 497, + 279, + 570 + ], + "type": "text", + "content": ". The exploitation policy selects the node with the highest navigation score as the local goal from the previously unvisited candidates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "spans": [ + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "content": "Figure 6 illustrates a simple scenario of entering a room. Suppose " + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "inline_equation", + "content": "W^{o} =" + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "content": " [sculpture, door, bed] and the agent has to compare two trajectories " + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1 = (v_1,v_2,A)" + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1 = (v_1,v_2,B)" + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "content": ". Each similarity matrix in Figure 6 has the " + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "inline_equation", + "content": "(t,j)" + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "content": " element as the similarity between the SOS feature of " + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}(w_j^o)" + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "content": ", which is calculated as " + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}(w_j^o) \\cdot \\vec{S}_t'" + }, + { + "bbox": [ + 45, + 571, + 279, + 664 + ], + "type": "text", + "content": ". Notably, the similarity matrix shows monotonic alignment and the navigation score is higher when the next action is chosen correctly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 669, + 141, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 669, + 141, + 681 + ], + "spans": [ + { + "bbox": [ + 45, + 669, + 141, + 681 + ], + "type": "text", + "content": "3.3. Training Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "text", + "content": "We use [24] for pretraining the visual encoder with panoramic RGB observations. We use the DAgger algorithm [46] to pretrain the navigation policy and the mode selector. To prevent overfitting, we iteratively perform" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 103, + 458, + 229 + ], + "blocks": [ + { + "bbox": [ + 298, + 103, + 458, + 229 + ], + "lines": [ + { + "bbox": [ + 298, + 103, + 458, + 229 + ], + "spans": [ + { + "bbox": [ + 298, + 103, + 458, + 229 + ], + "type": "image", + "image_path": "953ad6452a75b7dd3dfd02078929ac7b6d9100d714fc9faa876085f058d646c8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "lines": [ + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "spans": [ + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "text", + "content": "Figure 6. Toy Example. Monotonic alignment between language instruction and visual observation is desirable. Yellow dots " + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "text", + "content": " in the nodes describe the ground truth trajectory. Based on the node at " + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "inline_equation", + "content": "t = 3" + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "text", + "content": ", the similarity matrix can show either monotonic or non-monotonic alignment between object tokens and SOS features. The green circles " + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "text", + "content": " describe the possible candidates " + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "inline_equation", + "content": "A, B" + }, + { + "bbox": [ + 296, + 239, + 530, + 298 + ], + "type": "text", + "content": " for next action." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 467, + 115, + 530, + 236 + ], + "blocks": [ + { + "bbox": [ + 467, + 104, + 528, + 115 + ], + "lines": [ + { + "bbox": [ + 467, + 104, + 528, + 115 + ], + "spans": [ + { + "bbox": [ + 467, + 104, + 528, + 115 + ], + "type": "text", + "content": "similarity matrix" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 467, + 115, + 530, + 236 + ], + "lines": [ + { + "bbox": [ + 467, + 115, + 530, + 236 + ], + "spans": [ + { + "bbox": [ + 467, + 115, + 530, + 236 + ], + "type": "table", + "html": "
w01w2w3
v1=10.90.00.0
v1=20.31.00.0
v1=30.00.01.0
v1=4, Snav = 0.296
w01w2w3
v1=10.90.00.0
v1=20.31.00.0
v1=30.50.40.0
v1=4, Snav = 0.296
", + "image_path": "836f12f84d172a7363a704ee5a7c22f441b83c33d87b40a05c1d4ad46e4f1703.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "spans": [ + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": "teacher forcing and student forcing to choose the action from the exploration policy. Imitation learning loss is calculated as " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "L_{IL} = \\sum_{t=1}^{T} -\\log p(a_t^* | a_t)" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": " and object grounding loss is calculated as " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "L_{OG} = -\\log p(\\mathrm{obj}^* | \\mathrm{obj}_{pred})" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "\\mathrm{obj}^*" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": " denotes the ground truth and " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "\\mathrm{obj}_{pred}" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": " denotes the predicted object location. The total loss function is defined as " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "L_{total} = L_{mode} + L_{progress} + L_{corr} + L_{IL} + L_{OG}" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": ". We further finetune the agent via A2C [47]. The exploration policy selects the action " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": " with probability " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "p_t^a" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": ". Reinforcement learning loss is defined as " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "L_{RL} = -\\sum_t a_t^s \\log(p_t^a) A_t - \\lambda \\sum_t a_t^* \\log(p_t^a)" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": ". To train the mode selector, progress monitor, and exploration policy in an end-to-end manner, we use the total loss function as " + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "inline_equation", + "content": "L_{fine} = L_{mode} + L_{progress} + L_{RL}" + }, + { + "bbox": [ + 296, + 307, + 530, + 494 + ], + "type": "text", + "content": ". The exploitation policy searches the path toward the local goal from the constructed navigation graph. Thus, the exploitation policy is not learned." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 500, + 432, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 500, + 432, + 513 + ], + "spans": [ + { + "bbox": [ + 296, + 500, + 432, + 513 + ], + "type": "text", + "content": "4. Navigation Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 516, + 412, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 516, + 412, + 527 + ], + "spans": [ + { + "bbox": [ + 297, + 516, + 412, + 527 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 530, + 530, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 530, + 530, + 600 + ], + "spans": [ + { + "bbox": [ + 296, + 530, + 530, + 600 + ], + "type": "text", + "content": "We evaluate our method on three VLN benchmarks, Room-to-Room(R2R) [2], SOON [16], and REVERIE [17]. R2R evaluates the visually-grounded natural navigation performance of the agent. The agent must navigate to the predefined goal point given image observations and language instructions in an unseen environment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 601, + 530, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 601, + 530, + 647 + ], + "spans": [ + { + "bbox": [ + 296, + 601, + 530, + 647 + ], + "type": "text", + "content": "SOON is also a goal-oriented VLN benchmark. Natural language instructions in SOON have an average length of 47 words. The agent should locate the target location and detect the location of an object to find the target object." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 648, + 530, + 730 + ], + "type": "text", + "content": "REVERIE is a goal-oriented VLN benchmark that provides natural language instruction about target locations and objects. In REVERIE, the agent is given an instruction referring to a remote object with an average length of 21 words. With this instruction and a panoramic observation from the environment, the agent should navigate to the location the instruction describes and find the correct object" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6688" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 278, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 278, + 117 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 278, + 117 + ], + "type": "text", + "content": "bounding box among the predefined object bounding boxes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 155, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 155, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 155, + 133 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 137, + 178, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 137, + 178, + 148 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 178, + 148 + ], + "type": "text", + "content": "4.2.1 Navigation performance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "spans": [ + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "content": "We evaluate algorithms using the trajectory length (TL), success rate (SR), and success weighted by inverse path length (SPL) [48], and oracle success rate (OSR) for the navigation performance comparison. An episode is recorded as a success if the agent takes a stop action within " + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "inline_equation", + "content": "3\\mathrm{m}" + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "content": " of the target location. TL is the average path length in meters. SR is denoted as the number of successes divided by the total number of episodes, " + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "content": ". SPL is calculated as " + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "inline_equation", + "content": "\\frac{1}{M}\\sum_{i = 1}^{M}S_{i}\\frac{l_{i}}{\\max(p_{i},l_{i})}" + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "content": " denotes the success as a binary value. " + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "content": " denote the shortest path and actual path lengths for the " + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 45, + 150, + 279, + 291 + ], + "type": "text", + "content": " episode. OSR uses the oracle stop policy instead of the stop policy of the agent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 294, + 207, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 207, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 207, + 306 + ], + "type": "text", + "content": "4.2.2 Object grounding performance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "spans": [ + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "text", + "content": "We also evaluate the object grounding performance of the agent by the success rate of finding the target object (FSR) and the target finding success weighted by inverse path length (FSPL)1 [16, 17]. FSPL is calculated as FSPL = " + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{i=1}^{N}S_{i}^{nav}S_{i}^{loc} \\cdot l_{i}^{nav}/\\max(l_{i}^{nav},l_{i}^{gt})" + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "inline_equation", + "content": "S_{i}^{nav}" + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "text", + "content": " is whether the agent navigates to the target, " + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "inline_equation", + "content": "S_{i}^{loc}" + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "text", + "content": " is whether the agent finds a target object bounding box, and " + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "inline_equation", + "content": "l_{i}^{nav}" + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "inline_equation", + "content": "l_{i}^{gt}" + }, + { + "bbox": [ + 45, + 308, + 279, + 413 + ], + "type": "text", + "content": " are the navigation trajectory length and ground truth trajectory length, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 416, + 239, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 239, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 239, + 429 + ], + "type": "text", + "content": "4.3. Baselines and Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 431, + 279, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 431, + 279, + 547 + ], + "spans": [ + { + "bbox": [ + 45, + 431, + 279, + 547 + ], + "type": "text", + "content": "We compare our method with several other baselines as follows. For each task, we compare our method with a number of baselines that use various types of memory (recurrent, sequential, and topological map). For methods implemented with a hierarchical navigation framework, we compare the specific exploitation methods: homing, jump, and local goal search. Homing makes the agent backtrack, and jump makes the agent jump to a previously visited node. The hyperparameters and detailed model architecture of Meta-Explore are described in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 549, + 243, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 243, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 243, + 561 + ], + "type": "text", + "content": "4.4. Comparison with Navigation Baselines" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 563, + 279, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 563, + 279, + 609 + ], + "spans": [ + { + "bbox": [ + 45, + 563, + 279, + 609 + ], + "type": "text", + "content": "We compare our method with navigation baselines. We focus on the success rate and SPL. Rendered results and detailed analyses with other evaluation metrics are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 610, + 278, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 610, + 278, + 703 + ], + "spans": [ + { + "bbox": [ + 45, + 610, + 278, + 703 + ], + "type": "text", + "content": "R2R. Table 1 compares the proposed Meta-Explore with baselines for the R2R navigation task. We categorize the baseline methods based on the type of constructed memory and the type of exploitation. Our method outperforms other exploration-only baselines over all types of validation and test splits in success rate and SPL. Compared with hierarchical baselines SMNA [21], Regretful-Agent [22], FAST [35], and SSM [26], Meta-Explore improves success" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 106, + 530, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 106, + 530, + 279 + ], + "spans": [ + { + "bbox": [ + 297, + 106, + 530, + 279 + ], + "type": "text", + "content": "rate and SPL by at least " + }, + { + "bbox": [ + 297, + 106, + 530, + 279 + ], + "type": "inline_equation", + "content": "16.4\\%" + }, + { + "bbox": [ + 297, + 106, + 530, + 279 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 297, + 106, + 530, + 279 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 297, + 106, + 530, + 279 + ], + "type": "text", + "content": ", respectively. The main difference is that Meta-Explore constructs a topological map during exploration and uses the map for local goal search in exploitation. On the contrary, homing exploitation policies in SMNA, Regretful-Agent, and FAST only rely on the current trajectory, instead of taking advantage of the constructed memory. Jump exploitation in SSM uses a topological map to search a successful previous node, but it makes an unrealistic assumption that the agent can directly jump to a previously visited distant node and unfairly saves time. In our approach, we plan a path to the local goal based on the topological map. The experiment results reveal that even if we design a hierarchical navigation framework, exploration and exploitation are not entirely separate but they can complement each other." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 281, + 530, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 281, + 530, + 582 + ], + "spans": [ + { + "bbox": [ + 296, + 281, + 530, + 582 + ], + "type": "text", + "content": "SOON, REVERIE. Table 2 compares Meta-Explore with baselines in the SOON navigation task. While the proposed method does not improve performance in val seen split, Meta-Explore outperforms other baselines in the test unseen split of SOON for success rate by " + }, + { + "bbox": [ + 296, + 281, + 530, + 582 + ], + "type": "inline_equation", + "content": "17.1\\%" + }, + { + "bbox": [ + 296, + 281, + 530, + 582 + ], + "type": "text", + "content": " and SPL by " + }, + { + "bbox": [ + 296, + 281, + 530, + 582 + ], + "type": "inline_equation", + "content": "20.6\\%" + }, + { + "bbox": [ + 296, + 281, + 530, + 582 + ], + "type": "text", + "content": ". The result implies that for the goal-oriented VLN task, high performance in train or val seen splits can be the overfitted result. Because the agent can be easily overfitted to the training data, making a generalizable model or providing a deterministic error-correction module for inference is essential. Meta-Explore chooses the latter approach by correcting the trajectory via exploitation in regretful cases. The evaluation results in the REVERIE navigation task are described in the supplementary material. Meta-Explore shows improvement in the val split of REVERIE for success rate and SPL, but the improvement in the test split is lower than the results in R2R and SOON. We found 252 meaningless object categories (e.g., verbs, adjectives, and prepositions) and 418 replaceable object categories (e.g., typographical errors and synonyms) in the REVERIE3 dataset. Because our exploitation method utilizes object-based parsing of the given instruction to match with the detected object categories, the effectiveness of the proposed method is lessened due to inaccuracies and inconsistencies in the dataset. We expect to have higher performance if the mistakes in the dataset are fixed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 586, + 494, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 586, + 494, + 597 + ], + "spans": [ + { + "bbox": [ + 297, + 586, + 494, + 597 + ], + "type": "text", + "content": "4.5. Local Goal Search using SOS Features" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 599, + 530, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 599, + 530, + 692 + ], + "spans": [ + { + "bbox": [ + 296, + 599, + 530, + 692 + ], + "type": "text", + "content": "To discuss the significance of modeling exploitation policy, we conduct specific experiments about choosing the local goal for R2R and SOON. We evaluate our method using different types of local goal search, as shown in Table 3 and 4. Oracle denotes a method which selects a local goal using the ground truth trajectory. The performance of the oracle provides the achievable performance for each dataset. The results imply that local goal search using either" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 708, + 270, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 708, + 270, + 719 + ], + "spans": [ + { + "bbox": [ + 55, + 708, + 270, + 719 + ], + "type": "text", + "content": "1Identical with its original term, Remote Grounding Success (RGS)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 719, + 157, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 719, + 157, + 729 + ], + "spans": [ + { + "bbox": [ + 57, + 719, + 157, + 729 + ], + "type": "inline_equation", + "content": "2^{\\dagger}" + }, + { + "bbox": [ + 57, + 719, + 157, + 729 + ], + "type": "text", + "content": " indicates reproduced results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 700, + 530, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 700, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 297, + 700, + 530, + 729 + ], + "type": "inline_equation", + "content": "310.7\\%" + }, + { + "bbox": [ + 297, + 700, + 530, + 729 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 297, + 700, + 530, + 729 + ], + "type": "inline_equation", + "content": "41.2\\%" + }, + { + "bbox": [ + 297, + 700, + 530, + 729 + ], + "type": "text", + "content": " of a total of 46,476 words in the bounding box dataset correspond to meaningless and replaceable object categories, respectively." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6689" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 105, + 531, + 229 + ], + "blocks": [ + { + "bbox": [ + 47, + 105, + 531, + 229 + ], + "lines": [ + { + "bbox": [ + 47, + 105, + 531, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 531, + 229 + ], + "type": "table", + "html": "
MethodsMemoryExploitVal SeenVal UnseenTest Unseen
SR↑SPL↑TL↓NE↓SR↑SPL↑TL↓NE↓SR↑SPL↑TL↓NE↓
Random--16-9.589.4516-9.779.2313129.899.79
Human----------11.851.618676
Seq2Seq [2]RecX6.03911.33-22-8.397.8420188.137.85
VLN\\(\\circ\\)BERT [32]RecX726811.132.90635712.013.93635712.354.09
SMNA†[21]Rechoming696311.693.31474112.615.486156-4.48
Regretful-Agent [22]Rechoming6963-3.235041-5.324840-5.69
FAST (short) [35]Rechoming----564321.174.97544122.085.14
FAST (long) [35]Rechoming7004188.063.136302224.424.036103196.534.29
HAMT-e2e [34]SeqX767211.152.51666111.462.29656012.273.93
DUET [24]Top. MapX797312.322.28726013.943.31695914.733.65
SSM [26]Top. Mapjump716214.73.10624520.74.32614620.44.57
Meta-Explore (Ours)Top. Maplocal goal817511.952.11726213.093.22716114.253.57
", + "image_path": "39c78cef6cb759ef02796f4670567674882b0af678015557444c28ca87363894.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 256, + 531, + 364 + ], + "blocks": [ + { + "bbox": [ + 55, + 233, + 521, + 252 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 521, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 521, + 252 + ], + "type": "text", + "content": "Table 1. Comparison and evaluation results of the baselines and our model in the R2R Navigation Task. Gray shaded rows describe hierarchical navigation baselines. Three memory types: Rec(recurrent), Seq(sequential), and Top. Map(topological map)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 256, + 531, + 364 + ], + "lines": [ + { + "bbox": [ + 47, + 256, + 531, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 531, + 364 + ], + "type": "table", + "html": "
MethodsMemoryExploitVal Seen InstructionVal Seen HouseTest Unseen House
SR↑SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑
Human----------90.459.291.451.1
RandomRecX0.01.50.11.40.10.00.40.92.10.42.70.0
Speaker-Follower [28]RecX97.997.797.824.561.260.469.49.17.06.19.80.6
RCM [49]RecX84.082.689.110.962.460.972.77.87.46.212.40.7
AuxRN [23]RecX98.497.498.713.768.867.378.58.38.16.711.00.5
GBE w/o GETop. MapX89.588.391.824.262.560.873.06.711.48.718.80.8
GBE [16]Top. MapX98.497.998.644.276.362.564.17.311.910.219.51.4
\\( GBE^† \\)Top. MapX----19.513.328.51.212.99.221.50.5
DUET [24]Top. MapX94.091.690.031.136.322.650.93.833.421.443.04.2
Meta-Explore (Ours)Top. Maplocal goal100.099.196.033.944.734.852.78.939.125.848.74.0
", + "image_path": "07feae75b547e47b2a2080788085b4c9b79b5b21f5a979c612a2ae4bf20dbed3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "spans": [ + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "type": "text", + "content": "spatial or spectral visual representations is more effective than random local goal search. The results show that local goal search using spectral visual representations, i.e., SOS features, lead the agent to desirable nodes the most. We also compare local goal search with homing and the difference between the performance of the two methods is most noticeable in the test split of the SOON navigation task. As shown in Table 4, choosing the local goal with only spatial-domain features, the navigation performance does not improve compared to homing. On the contrary, spectral-domain local goal search shows significant improvement against homing by " + }, + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "type": "text", + "content": " in success rate, " + }, + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "type": "inline_equation", + "content": "34.5\\%" + }, + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "type": "text", + "content": " on SPL, and " + }, + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "type": "inline_equation", + "content": "27.4\\%" + }, + { + "bbox": [ + 45, + 386, + 279, + 585 + ], + "type": "text", + "content": " on FSPL. The results imply that using spectral-domain SOS features helps high-level decision making, thereby enhancing the navigation performance. To further show the effectiveness of SOS features, we provide sample local goal search scenarios in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 47, + 587, + 279, + 646 + ], + "blocks": [ + { + "bbox": [ + 108, + 368, + 467, + 379 + ], + "lines": [ + { + "bbox": [ + 108, + 368, + 467, + 379 + ], + "spans": [ + { + "bbox": [ + 108, + 368, + 467, + 379 + ], + "type": "text", + "content": "Table 2. Comparison and evaluation results of the baselines and our model in the SOON Navigation Task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 587, + 279, + 646 + ], + "lines": [ + { + "bbox": [ + 47, + 587, + 279, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 279, + 646 + ], + "type": "table", + "html": "
Local GoalVal SeenVal Unseen
SR↑SPL↑OSR↑TL↓NE↓SR↑SPL↑OSR↑TL↓NE↓
Oracle81.8874.1287.4613.061.9375.9562.5384.1614.002.71
Random79.3372.6785.3113.192.2270.9759.4580.1614.923.34
Homing80.2273.6385.6012.512.1471.6560.6080.3313.913.26
Spatial79.6373.1485.6012.992.2271.5660.0180.3314.903.27
Spectral80.6175.1585.8011.952.1171.7861.6880.7613.093.22
", + "image_path": "a7461b174b2dc80f1684741d73048eebc0e3203a5eebe3b6fd52ddc6b153b40a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 47, + 665, + 279, + 724 + ], + "blocks": [ + { + "bbox": [ + 71, + 650, + 254, + 660 + ], + "lines": [ + { + "bbox": [ + 71, + 650, + 254, + 660 + ], + "spans": [ + { + "bbox": [ + 71, + 650, + 254, + 660 + ], + "type": "text", + "content": "Table 3. Comparison of Exploitation Policies. (R2R)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 665, + 279, + 724 + ], + "lines": [ + { + "bbox": [ + 47, + 665, + 279, + 724 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 279, + 724 + ], + "type": "table", + "html": "
Local GoalSR↑Val Seen HouseTest Unseen House
SPL↑OSR↑FSPL↑SR↑SPL↑OSR↑FSPL↑
Oracle54.4237.9663.7211.0148.3828.4562.984.74
Random24.7811.9734.963.0824.197.4135.841.29
Homing42.0427.7248.2310.1835.4019.1851.623.14
Spatial32.3011.6039.381.9026.1110.5839.231.43
Spectral44.6934.8452.658.8939.0925.8048.674.01
", + "image_path": "9ab44f70c9422182776a4fe87432d10326335e52d22ff454561812d763660587.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 726, + 258, + 737 + ], + "lines": [ + { + "bbox": [ + 67, + 726, + 258, + 737 + ], + "spans": [ + { + "bbox": [ + 67, + 726, + 258, + 737 + ], + "type": "text", + "content": "Table 4. Comparison of Exploitation Policies. (SOON)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 297, + 386, + 389, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 386, + 389, + 398 + ], + "spans": [ + { + "bbox": [ + 297, + 386, + 389, + 398 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 399, + 530, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 399, + 530, + 538 + ], + "spans": [ + { + "bbox": [ + 296, + 399, + 530, + 538 + ], + "type": "text", + "content": "We conduct an ablation study to compare the proposed method against language-triggered hierarchical exploration. Results in the supplementary material show that among the three representation domains, spatial, spectral, and language, the spectral-domain features enhance navigation performance the most. Additionally, to implicate further applications of Meta-Explore in continuous environments, we evaluate our method on the photo-realistic Habitat [50] simulator to solve image-goal navigation and vision-and-language navigation tasks. Implementation details and results are included in the supplementary material. Results show that our method outperforms baselines in both tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 541, + 368, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 541, + 368, + 553 + ], + "spans": [ + { + "bbox": [ + 297, + 541, + 368, + 553 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 554, + 530, + 730 + ], + "type": "text", + "content": "We have proposed Meta-Explore, a hierarchical navigation method for VLN, by correcting mistaken short-term actions via efficient exploitation. In the exploitation mode, the agent is directed to a local goal which is inferred to be the closest to the target. A topological map constructed during exploration helps the agent to search and plan the shortest path toward the local goal. To further search beyond the frontier of the map, we present a novel visual representation called scene object spectrum (SOS), which compactly encodes the arrangements and frequencies of nearby objects. Meta-Explore achieves the highest generalization performance for test splits of R2R, SOON, and val split of REVERIE navigation tasks by showing less overfitting and high success rates. We plan to apply Meta-Explore for VLN tasks in continuous environments in our future work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6690" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 280, + 730 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 51, + 124, + 279, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 279, + 156 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 279, + 156 + ], + "type": "text", + "content": "[1] Fengda Zhu, Yi Zhu, Vincent Lee, Xiaodan Liang, and Xiao-jun Chang. Deep learning for embodied vision navigation: A survey. arXiv preprint arXiv:2108.04097, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 159, + 280, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 159, + 280, + 234 + ], + "spans": [ + { + "bbox": [ + 51, + 159, + 280, + 234 + ], + "type": "text", + "content": "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2018. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 239, + 280, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 239, + 280, + 293 + ], + "spans": [ + { + "bbox": [ + 51, + 239, + 280, + 293 + ], + "type": "text", + "content": "[3] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. Touchdown: Natural language navigation and spatial reasoning in visual street environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12538-12547, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 296, + 279, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 296, + 279, + 349 + ], + "spans": [ + { + "bbox": [ + 51, + 296, + 279, + 349 + ], + "type": "text", + "content": "[4] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-Across-Room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Conference on Empirical Methods for Natural Language Processing, 2020. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 353, + 279, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 353, + 279, + 417 + ], + "spans": [ + { + "bbox": [ + 51, + 353, + 279, + 417 + ], + "type": "text", + "content": "[5] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In *Transactions of the Association for Computational Linguistics*, pages 1862–1872, Florence, Italy, July 2019. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 421, + 279, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 421, + 279, + 453 + ], + "spans": [ + { + "bbox": [ + 51, + 421, + 279, + 453 + ], + "type": "text", + "content": "[6] An Yan, Xin Eric Wang, Jiangtao Feng, Lei Li, and William Yang Wang. Cross-lingual vision-language navigation. arXiv preprint arXiv:1910.11301, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 458, + 279, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 458, + 279, + 522 + ], + "spans": [ + { + "bbox": [ + 51, + 458, + 279, + 522 + ], + "type": "text", + "content": "[7] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. In Proceedings of the International Conference on Neural Information Processing Systems, volume 34, pages 652-663. Curran Associates, Inc., 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 526, + 279, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 526, + 279, + 568 + ], + "spans": [ + { + "bbox": [ + 51, + 526, + 279, + 568 + ], + "type": "text", + "content": "[8] Jacob Krantz, Erik Wijmans, Arjun Majundar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision and language navigation in continuous environments. In Proceedings of the European Conference on Computer Vision, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 572, + 279, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 572, + 279, + 635 + ], + "spans": [ + { + "bbox": [ + 51, + 572, + 279, + 635 + ], + "type": "text", + "content": "[9] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teplyashin, Karl Moritz Hermann, Mateusz Malinowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 639, + 279, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 279, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 279, + 693 + ], + "type": "text", + "content": "[10] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11773–11781, Apr. 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "type": "text", + "content": "[11] Arun Balajee Vasudevan, Dengxin Dai, and Luc Van Gool. Talk2nav: Long-range vision-and-language navigation with dual attention and spatial memory. In Proceedings of the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 531, + 730 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 317, + 106, + 530, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 530, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 530, + 128 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, volume 129, pages 246-266. Springer, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 299, + 132, + 531, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 132, + 531, + 207 + ], + "spans": [ + { + "bbox": [ + 299, + 132, + 531, + 207 + ], + "type": "text", + "content": "[12] Dipendra Misra, Andrew Bennett, Valts Blukis, Eyvind Niklasson, Max Shatkhin, and Yoav Artzi. Mapping instructions to actions in 3D environments with visual goal prediction. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 2667-2678, Brussels, Belgium, October-November 2018. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 211, + 531, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 211, + 531, + 264 + ], + "spans": [ + { + "bbox": [ + 299, + 211, + 531, + 264 + ], + "type": "text", + "content": "[13] Ta-Chung Chi, Minmin Shen, Mihail Eric, Seokhwan Kim, and Dilek Hakkani-tur. Justask: An interactive learning framework for vision and language navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2459-2466, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 268, + 531, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 268, + 531, + 301 + ], + "spans": [ + { + "bbox": [ + 299, + 268, + 531, + 301 + ], + "type": "text", + "content": "[14] Yi Wu, Yuxin Wu, Georgia Gkioxari, and Yuandong Tian. Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 305, + 531, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 305, + 531, + 347 + ], + "spans": [ + { + "bbox": [ + 299, + 305, + 531, + 347 + ], + "type": "text", + "content": "[15] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 351, + 531, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 351, + 531, + 405 + ], + "spans": [ + { + "bbox": [ + 299, + 351, + 531, + 405 + ], + "type": "text", + "content": "[16] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12689-12699, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 410, + 531, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 410, + 531, + 473 + ], + "spans": [ + { + "bbox": [ + 299, + 410, + 531, + 473 + ], + "type": "text", + "content": "[17] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 478, + 531, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 478, + 531, + 531 + ], + "spans": [ + { + "bbox": [ + 299, + 478, + 531, + 531 + ], + "type": "text", + "content": "[18] Khanh Nguyen and Hal Daume III. Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, November 2019. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 535, + 531, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 535, + 531, + 588 + ], + "spans": [ + { + "bbox": [ + 299, + 535, + 531, + 588 + ], + "type": "text", + "content": "[19] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 592, + 531, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 592, + 531, + 646 + ], + "spans": [ + { + "bbox": [ + 299, + 592, + 531, + 646 + ], + "type": "text", + "content": "[20] Alane Suhr, Claudia Yan, Jack Schluger, Stanley Yu, Hadi Khader, Marwa Mouallem, Iris Zhang, and Yoav Artzi. Executing instructions in situated collaborative interactions. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 2119-2130, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 650, + 531, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 650, + 531, + 703 + ], + "spans": [ + { + "bbox": [ + 299, + 650, + 531, + 703 + ], + "type": "text", + "content": "[21] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In Proceedings of the International Conference on Learning Representations, 2019. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 708, + 531, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 708, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 708, + 531, + 730 + ], + "type": "text", + "content": "[22] Chih-Yao Ma, Zuxuan Wu, Ghassan AlRegib, Caiming Xiong, and Zsolt Kira. The regretful agent: Heuristic-aided navigation" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6691" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 730 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 64, + 106, + 279, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 106, + 279, + 139 + ], + "spans": [ + { + "bbox": [ + 64, + 106, + 279, + 139 + ], + "type": "text", + "content": "through progress estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6732-6740, 2019. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 143, + 279, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 279, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 279, + 196 + ], + "type": "text", + "content": "[23] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10012-10022, 2020. 1, 2, 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 199, + 279, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 279, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 279, + 263 + ], + "type": "text", + "content": "[24] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16537-16547, 2022. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 267, + 279, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 279, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 279, + 320 + ], + "type": "text", + "content": "[25] Zhiwei Deng, Karthik Narasimhan, and Olga Russakovsky. Evolving graphical planner: Contextual global planning for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, volume 33, pages 20660-20672, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 324, + 279, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 279, + 377 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 279, + 377 + ], + "type": "text", + "content": "[26] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8455-8464, June 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 381, + 278, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 278, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 278, + 422 + ], + "type": "text", + "content": "[27] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Proceedings of the Conference on Robot Learning, pages 394-406, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 427, + 278, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 278, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 278, + 491 + ], + "type": "text", + "content": "[28] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, 2018. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 495, + 279, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 279, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 279, + 557 + ], + "type": "text", + "content": "[29] Xiujun Li, Chunyuan Li, Qiaolin Xia, Yonatan Bisk, Asli Celikyilmaz, Jianfeng Gao, Noah A Smith, and Yejin Choi. Robust navigation with language pretraining and stochastic sampling. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 1494-1499, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 562, + 279, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 562, + 279, + 615 + ], + "spans": [ + { + "bbox": [ + 48, + 562, + 279, + 615 + ], + "type": "text", + "content": "[30] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Proceedings of the International Conference on Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 619, + 279, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 619, + 279, + 672 + ], + "spans": [ + { + "bbox": [ + 48, + 619, + 279, + 672 + ], + "type": "text", + "content": "[31] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1634-1643, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 676, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 676, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 48, + 676, + 279, + 730 + ], + "type": "text", + "content": "[32] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. VIn" + }, + { + "bbox": [ + 48, + 676, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\text{©}" + }, + { + "bbox": [ + 48, + 676, + 279, + 730 + ], + "type": "text", + "content": "bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021. 2, 8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 730 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "text", + "content": "[33] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15942-15952, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 299, + 154, + 530, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 154, + 530, + 207 + ], + "spans": [ + { + "bbox": [ + 299, + 154, + 530, + 207 + ], + "type": "text", + "content": "[34] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In Proceedings of the International Conference on Neural Information Processing Systems, pages 5834-5847, 2021. 2, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 299, + 211, + 530, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 211, + 530, + 274 + ], + "spans": [ + { + "bbox": [ + 299, + 211, + 530, + 274 + ], + "type": "text", + "content": "[35] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6741-6749, 2019. 2, 3, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 279, + 530, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 279, + 530, + 299 + ], + "spans": [ + { + "bbox": [ + 299, + 279, + 530, + 299 + ], + "type": "text", + "content": "[36] James G March. Exploration and exploitation in organizational learning. Organization science, 2(1):71-87, 1991. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 304, + 530, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 304, + 530, + 335 + ], + "spans": [ + { + "bbox": [ + 299, + 304, + 530, + 335 + ], + "type": "text", + "content": "[37] Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O Stanley, and Jeff Clune. First return, then explore. Nature, 590(7847):580-586, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 340, + 530, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 340, + 530, + 381 + ], + "spans": [ + { + "bbox": [ + 299, + 340, + 530, + 381 + ], + "type": "text", + "content": "[38] Miruna Pislar, David Szepesvari, Georg Ostrovski, Diana L Borsa, and Tom Schaul. When should agents explore? In Proceedings of the International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 386, + 530, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 386, + 530, + 428 + ], + "spans": [ + { + "bbox": [ + 299, + 386, + 530, + 428 + ], + "type": "text", + "content": "[39] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In Proceedings of the International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 433, + 530, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 433, + 530, + 496 + ], + "spans": [ + { + "bbox": [ + 299, + 433, + 530, + 496 + ], + "type": "text", + "content": "[40] Meera Hahn, Devendra Singh Chaplot, Shubham Tulsiani, Mustafa Mukadam, James M Rehg, and Abhinav Gupta. No rl, no simulation: Learning to navigate without navigating. In Proceedings of the International Conference on Neural Information Processing Systems, volume 34, pages 26661-26673. Curran Associates, Inc., 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 500, + 530, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 500, + 530, + 543 + ], + "spans": [ + { + "bbox": [ + 299, + 500, + 530, + 543 + ], + "type": "text", + "content": "[41] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 547, + 530, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 547, + 530, + 621 + ], + "spans": [ + { + "bbox": [ + 299, + 547, + 530, + 621 + ], + "type": "text", + "content": "[42] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 626, + 530, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 626, + 530, + 657 + ], + "spans": [ + { + "bbox": [ + 299, + 626, + 530, + 657 + ], + "type": "text", + "content": "[43] Wolfgang Stürzl and Hanspeter A Mallot. Efficient visual homing based on fourier transformed panoramic images. Robotics and Autonomous Systems, 54(4):300-313, 2006. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 661, + 530, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 661, + 530, + 682 + ], + "spans": [ + { + "bbox": [ + 299, + 661, + 530, + 682 + ], + "type": "text", + "content": "[44] Jean Serra. Mathematical Morphology, pages 1-16. Springer International Publishing, Cham, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "type": "text", + "content": "[45] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the Conference on Empirical Methods for Natural Language Processing, pages 5100-5111, 2019. 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6692" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 420 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[46] Stéphane Ross, Geoffrey J Gordon, and J Andrew Bagnell. No-regret reductions for imitation learning and structured prediction. arXiv preprint arXiv:1011.0686, 2010. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 143, + 279, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 279, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 279, + 196 + ], + "type": "text", + "content": "[47] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proceedings of the International Conference on Machine Learning, pages 1928-1937, 2016. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 200, + 279, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 279, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 279, + 253 + ], + "type": "text", + "content": "[48] Peter Anderson, Angel Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, et al. On evaluation of embodied navigation agents. arXiv preprint arXiv:1807.06757, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 279, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 279, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 279, + 320 + ], + "type": "text", + "content": "[49] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6629-6638, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 325, + 279, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 279, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 279, + 420 + ], + "type": "text", + "content": "[50] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Proceedings of the International Conference on Neural Information Processing Systems, 2021. 8" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "6693" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_content_list.json b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d233885a5489feab346a8a06e4eded297d46bdfd --- /dev/null +++ b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_content_list.json @@ -0,0 +1,1796 @@ +[ + { + "type": "text", + "text": "Meta-Learning with a Geometry-Adaptive Preconditioner", + "text_level": 1, + "bbox": [ + 192, + 130, + 777, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suhyun Kang $^{1*}$ , Duhun Hwang $^{1*}$ , Moonjung Eo $^{1}$ , Taesup Kim $^{2}$ , Wonjong Rhee $^{1,3\\dagger}$", + "bbox": [ + 166, + 179, + 807, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Department of Intelligence and Information, Seoul National University", + "bbox": [ + 200, + 198, + 769, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Graduate School of Data Science, Seoul National University", + "bbox": [ + 240, + 215, + 728, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ IPAI and AIIS, Seoul National University", + "bbox": [ + 313, + 233, + 655, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{su_hyun, yelobean, eod87, taesup.kim, wrhee}@snu.ac.kr", + "bbox": [ + 238, + 253, + 725, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model-agnostic meta-learning (MAML) is one of the most successful meta-learning algorithms. It has a bi-level optimization structure where the outer-loop process learns a shared initialization and the inner-loop process optimizes task-specific weights. Although MAML relies on the standard gradient descent in the inner-loop, recent studies have shown that controlling the inner-loop's gradient descent with a meta-learning preconditioner can be beneficial. Existing preconditioners, however, cannot simultaneously adapt in a task-specific and path-dependent way. Additionally, they do not satisfy the Riemannian metric condition, which can enable the steepest descent learning with preconditioned gradient. In this study, we propose Geometry-Adaptive Preconditioned gradient descent (GAP) that can overcome the limitations in MAML; GAP can efficiently meta-learn a preconditioner that is dependent on task-specific parameters, and its preconditioner can be shown to be a Riemannian metric. Thanks to the two properties, the geometry-adaptive preconditioner is effective for improving the inner-loop optimization. Experiment results show that GAP outperforms the state-of-the-art MAML family and preconditioned gradient descent-MAML (PGD-MAML) family in a variety of few-shot learning tasks. Code is available at: https://github.com/Suhyun777/CVPR23-GAP.", + "bbox": [ + 76, + 335, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 728, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta-learning, or learning to learn, enables algorithms to quickly learn new concepts with only a small number of samples by extracting prior-knowledge known as meta-knowledge from a variety of tasks and by improving the generalization capability over the new tasks. Among the meta-learning algorithms, the category of optimization-based meta-learning [8, 17, 20, 21, 48] has been gaining popularity", + "bbox": [ + 75, + 755, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "due to its flexible applicability over diverse fields including robotics [55, 59], medical image analysis [40, 54], language modeling [37, 42], and object detection [46, 61]. In particular, Model-Agnostic Meta-Learning (MAML) [20] is one of the most prevalent gradient-based meta-learning algorithms.", + "bbox": [ + 500, + 304, + 890, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many recent studies have improved MAML by adopting a Preconditioned Gradient Descent (PGD $^1$ ) for inner-loop optimization [34, 36, 44, 49, 53, 57, 66]. In this paper, we collectively address PGD-based MAML algorithms as the PGD-MAML family. A PGD is different from the ordinary gradient descent because it performs a preconditioning on the gradient using a preconditioning matrix $\\mathbf{P}$ , also called a preconditioner. A PGD-MAML algorithm meta-learns not only the initialization parameter $\\theta_0$ of the network but also the meta-parameter $\\phi$ of the preconditioner $\\mathbf{P}$ .", + "bbox": [ + 496, + 381, + 892, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For the inner-loop optimization, $\\mathbf{P}$ was kept static in most of the previous works (Figure 1(b)) [34, 36, 44, 57, 66]. Some of the previous works considered adapting the preconditioner $\\mathbf{P}$ with the inner-step $k$ (Figure 1(c)) [49] and some others with the individual task (Figure 1(d)) [53]. They achieved performance improvement by considering individual tasks and inner-step, respectively, and showed that both factors were valuable. However, both factors have not been considered simultaneously in the existing studies.", + "bbox": [ + 496, + 532, + 893, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When a parameter space has a certain underlying structure, there exists a Riemannian metric corresponding the parameter space [3,4]. If the preconditioning matrix is the Riemannian metric, the preconditioned gradient is known to become the steepest descent on the parameter space [2-4,6,27]. An illustration of a toy example is shown in Figure 2. The optimization path of an ordinary gradient descent is shown in Figure 2(a). Compared to the ordinary gradient descent, a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can actually harm the optimization. For the example in Figure 2(b), the preconditioner affects the optimization into an undesirable direction and negatively affects the gradient descent. On the", + "bbox": [ + 496, + 667, + 893, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1PGD in our work should not be confused with Projected Gradient Descent [13].", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 205, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16080", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22f94a546d5669996f508fd7a8de533e9ecf86714cfd0f4f740d739e3e8c590d.jpg", + "image_caption": [ + "(a) MAML" + ], + "image_footnote": [], + "bbox": [ + 116, + 90, + 243, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fc08048b30068de72750ee3c66ee519927bf6ab0c8cf79c359d78bb633857f94.jpg", + "image_caption": [ + "(b) Non-adaptive $\\mathbf{P}(\\phi)$", + "Figure 1. Diagram of MAML and PGD-MAML family. For the inner-loop adaptation in each diagram, the dotted lines of the same color indicate that they use a common preconditioning matrix (preconditioner). (a) MAML adaptation: no preconditioner is used (i.e., $\\mathbf{P} = \\mathbf{I}$ ). (b) $\\mathbf{P}(\\phi)$ : a constant preconditioner is used in the inner-loop where the preconditioner's meta-parameter $\\phi$ is meta-learned. (c) $\\mathbf{P}(k; \\phi)$ : a constant preconditioner is used for each inner-step $k$ . Preconditioner for each step is meta-learned, but $\\mathbf{P}(k, \\phi)$ is not task-specific. (d) $\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}}; \\phi)$ : a constant preconditioner is used for each task. Preconditioner for each task is meta-learned, but $\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}}; \\phi)$ is not dependent on $k$ . (e) GAP adapts $\\mathbf{P}(\\theta_{\\tau, k}; \\phi)$ : a fully adaptive preconditioner is used where it is task-specific and path-dependent. Instead of saying 'dependent on $k$ ', we specifically say it is path-dependent because the exact dependency is on the task-specific parameter set $\\theta_{\\tau, k}$ that is considerably more informative than $k$ ." + ], + "image_footnote": [], + "bbox": [ + 248, + 93, + 395, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/46b74e423ecb4e495bc02b70428a0eecb8433c5a9499a669faeeb13e4bbe74dd.jpg", + "image_caption": [ + "(c) Adaptive $\\mathbf{P}(k;\\phi)$" + ], + "image_footnote": [], + "bbox": [ + 398, + 103, + 549, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/038e361c7d389e45c4e7073230ae9930ebb75bdd33dac471b48952256502f58e.jpg", + "image_caption": [ + "(d) Adaptive $\\mathbf{P}(D_{\\tau}^{tr};\\phi)$" + ], + "image_footnote": [], + "bbox": [ + 552, + 104, + 702, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a5c6cd1ca0fac950bc4b3c6165c35dcab0dfd8582eb6825828d95786f0099240.jpg", + "image_caption": [ + "(e) Adaptive $\\mathbf{P}(\\theta_{\\tau ,k};\\phi)$" + ], + "image_footnote": [], + "bbox": [ + 705, + 103, + 857, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a17d7c689a2c2d9f27ee5d1d4ee36f32ecd5bf0dafc979e08fa11e9a17aa020.jpg", + "image_caption": [ + "(a) Gradient Descent (GD)" + ], + "image_footnote": [], + "bbox": [ + 109, + 338, + 251, + 425 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/23f7288c3df23d2178d36282242e1bced73b5cfd1cb10de0b9bc550a8e2ac384.jpg", + "image_caption": [ + "(b) Preconditioned GD (Non-Riemannian metric)" + ], + "image_footnote": [], + "bbox": [ + 285, + 338, + 426, + 425 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb7ee49f34060de67ab21dc136cd1251f87a3d5dcaae13f93e4a578395eeae29.jpg", + "image_caption": [ + "(c) Preconditioned GD (Riemannian metric)", + "Figure 2. A toy example for illustrating the effect of Riemannian metric (see Supplementary Section A). When the curvature of the parameter space is poorly conditioned, (a) gradient descent can suffer from the difficulty of finding the solution, (b) a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can suffer further, and (c) a preconditioned gradient descent with a preconditioner that is a Riemannian metric can perform better." + ], + "image_footnote": [], + "bbox": [ + 196, + 468, + 336, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "contrary, if the preconditioner is the Riemannian metric corresponding the parameter space, the preconditioned gradient descent can become the steepest descent and can exhibit a better optimization behavior as shown in Figure 2(c). While the Riemannian metric condition (i.e., positive definiteness) is a necessary condition for steepest descent learning, the existing studies on PGD-MAML family did not consider constraining preconditioners to satisfy the condition for Riemannian metric.", + "bbox": [ + 75, + 729, + 470, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we propose a new PGD method named Geometry Aaptive Preconditioned gradient descent (GAP).", + "bbox": [ + 76, + 869, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, GAP satisfies two desirable properties which have not been considered before. First, GAP's preconditioner $\\mathbf{P}_{\\mathrm{GAP}}$ is a fully adaptive preconditioner that can adapt to the individual task (task-specific) and to the optimization-path (path-dependent). The full adaptation is made possible by having the preconditioner depend on the task-specific parameter $\\theta_{\\tau,k}$ (Figure 1(e)). Second, we prove that $\\mathbf{P}_{\\mathrm{GAP}}$ is a Riemannian metric. To this end, we force the metaparameters of $\\mathbf{P}_{\\mathrm{GAP}}$ to be positive definite. Thus, GAP guarantees the steepest descent learning on the parameter space corresponding to $\\mathbf{P}_{\\mathrm{GAP}}$ . Owing to the two properties, GAP enables a geometry-adaptive learning in inner-loop optimization.", + "bbox": [ + 496, + 330, + 893, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the implementation of GAP, we utilize the Singular Value Decomposition (SVD) operation to come up with our preconditioner satisfying the desired properties. For the recently proposed large-scale architectures, computational overhead can be an important design factor and we provide a low-computational approximation, Approximate GAP, that can be proven to asymptotically approximate the operation of GAP.", + "bbox": [ + 496, + 527, + 890, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of GAP, we empirically evaluate our algorithm on popular few-shot learning tasks; few-shot regression, few-shot classification, and few-shot cross-domain classification. The results show that GAP outperforms the state-of-the-art MAML family and PGDMAML family.", + "bbox": [ + 496, + 648, + 893, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our study can be summarized as follows:", + "bbox": [ + 500, + 739, + 890, + 767 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new preconditioned gradient descent method called GAP, where it learns a preconditioner that enables a geometry-adaptive learning in the inner-loop optimization.", + "- We prove that GAP's preconditioner has two desirable properties: (1) It is both task-specific and path-dependent (i.e., dependent on task-specific parameter $\\theta_{\\tau, k}$ ). (2) It is a Riemannian metric." + ], + "bbox": [ + 514, + 770, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16081", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For large-scale architectures, we provide a low-computational approximation that can be theoretically shown to approximate the GAP method.", + "- For popular few-shot learning benchmark tasks, we empirically show that GAP outperforms the state-of-the-art MAML family and PGD-MAML family." + ], + "bbox": [ + 93, + 90, + 470, + 194 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 76, + 209, + 205, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Model-Agnostic Meta-Learning (MAML)", + "text_level": 1, + "bbox": [ + 76, + 234, + 429, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of MAML [20] is to find the best initialization that the model can quickly adapt from, such that the model can perform well for a new task. MAML consists of two levels of main optimization processes: inner-loop and outer-loop optimizations. Consider the model $f_{\\theta}(\\cdot)$ with parameter $\\theta$ . For a task $\\tau = \\{D_{\\tau}^{\\mathrm{tr}}, D_{\\tau}^{\\mathrm{val}}\\}$ sampled from the task distribution $p(\\mathcal{T})$ , the inner-loop optimization is defined as:", + "bbox": [ + 76, + 257, + 470, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\tau , K} = \\theta_ {\\tau , 0} - \\alpha \\sum_ {k = 0} ^ {K - 1} \\nabla_ {\\theta_ {\\tau , k}} \\mathcal {L} _ {\\tau} ^ {\\mathrm {i n}} \\left(\\theta_ {\\tau , k}; D _ {\\tau} ^ {\\mathrm {t r}}\\right) s. t \\theta_ {\\tau , 0} = \\theta , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 376, + 468, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{\\tau ,k}$ is task-specific parameters for task $\\tau$ and $\\alpha$ is the learning rate for inner-loop optimization, $\\mathcal{L}_{\\tau}^{\\mathrm{in}}$ is the inner-loop's loss function, and $K$ is the number of gradient descent steps. With $D_{\\tau_i}^{\\mathrm{val}}$ in each task, we can define outer-loop optimization as", + "bbox": [ + 76, + 422, + 468, + 497 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 506, + 468, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta$ is the learning rate for outer-loop optimization, and $\\mathcal{L}_{\\tau}^{\\mathrm{out}}$ is the outer-loop's loss function.", + "bbox": [ + 76, + 539, + 468, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Preconditioned Gradient Descent (PGD)", + "text_level": 1, + "bbox": [ + 76, + 580, + 421, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PGD is a method that minimizes the empirical risk through a gradient update with a preconditioner that rescales the geometry of the parameter space. Given model parameters $\\theta$ and task $\\tau = \\{D_{\\tau}^{\\mathrm{tr}}, D_{\\tau}^{\\mathrm{val}}\\}$ , we can define the preconditioned gradient update with a preconditioner $\\mathbf{P}$ as follows:", + "bbox": [ + 76, + 604, + 468, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta_ {\\tau , k + 1} = \\theta_ {\\tau , k} - \\alpha \\mathbf {P} \\nabla_ {\\theta_ {\\tau , k}} \\mathcal {L} _ {\\tau} \\left(\\theta_ {\\tau , k}; D _ {\\tau} ^ {\\mathrm {t r}}\\right) \\\\ k = 0, 1, \\dots \\text {a n d} \\theta_ {\\tau , 0} = \\theta \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 703, + 468, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\tau}(\\theta_{\\tau,k}; D_{\\tau}^{\\mathrm{tr}})$ is the empirical loss for task $\\tau$ and parameters $\\theta_{\\tau,k}$ . Setting $\\mathbf{P} = \\mathbf{I}$ recovers Eq. (3) to the basic gradient descent (GD). Choice of $\\mathbf{P}$ for exploiting the second-order information includes the inverse Fisher information matrix $\\mathbf{F}^{-1}$ which leads to the natural gradient descent (NGD) [4]; the inverse Hessian $\\mathbf{H}^{-1}$ which corresponds to the Newton's method [31]; and the diagonal matrix estimation with the past gradients which results in the adaptive gradient methods [18, 28]. They often reduce the effect of pathological curvature and speed up the optimization [5].", + "bbox": [ + 76, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Unfolding: reshaping a tensor into a matrix", + "text_level": 1, + "bbox": [ + 500, + 90, + 867, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this study, the concept of unfolding is used to transform the gradient tensor of convolutional kernels into a matrix form. Tensor unfolding, also known as matricization or flattening, is the process of reshaping the elements of an $N$ -dimensional tensor $\\mathbf{X} \\in \\mathbb{R}^{I_1 \\times \\dots \\times I_N}$ into a matrix [29]. The mode- $n$ unfolding of an $N$ -dimensional tensor $\\mathbf{X} \\in \\mathbb{R}^{I_1 \\times \\dots \\times I_N}$ is defined as:", + "bbox": [ + 496, + 113, + 893, + 219 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} \\xrightarrow [ \\text {m o d e - n u n f o l d i n g} ]{} \\mathbf {X} _ {[ n ]} \\in \\mathbb {R} ^ {I _ {n} \\times I _ {M}}, \\text {w h e r e} I _ {M} = \\prod_ {k \\neq n} I _ {k} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 224, + 890, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For example, the weight tensor of a convolutional layer is represented as a 4-D tensor $(\\mathbf{W} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times C_{\\mathrm{in}} \\times k_h \\times k_w})$ , where it is composed of kernels and it can be unfolded into a matrix as one of the following four forms: (1) $\\mathbf{W}_{[1]} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times (C_{\\mathrm{in}}k_hk_w)}$ , (2) $\\mathbf{W}_{[2]} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times (C_{\\mathrm{out}}k_hk_w)}$ , (3) $\\mathbf{W}_{[3]} \\in \\mathbb{R}^{k_h \\times (C_{\\mathrm{out}}C_{\\mathrm{in}}k_w)}$ , (4) $\\mathbf{W}_{[4]} \\in \\mathbb{R}^{k_w \\times (C_{\\mathrm{out}}C_{\\mathrm{in}}k_h)}$ .", + "bbox": [ + 496, + 262, + 893, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Riemannian manifold", + "text_level": 1, + "bbox": [ + 500, + 364, + 702, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An $n$ -dimensional Riemannian manifold is defined by a manifold $\\mathcal{M}$ and a Riemannian metric $g: \\mathcal{M} \\to \\mathbb{R}^{n \\times n}$ , which is a smooth function from each point $\\pmb{x} \\in \\mathcal{M}$ to a positive definite matrix [32]. The metric $g(\\pmb{x})$ defines the inner product of two tangent vectors for each point of the manifold $\\langle \\cdot, \\cdot \\rangle: \\mathcal{T}_{\\pmb{x}}\\mathcal{M} \\times \\mathcal{T}_{\\pmb{x}}\\mathcal{M} \\to \\mathbb{R}$ , where $\\mathcal{T}_{\\pmb{x}}\\mathcal{M}$ is the tangent space of $\\pmb{x}$ . For $\\pmb{v}, \\pmb{w} \\in \\mathcal{T}_{\\pmb{x}}\\mathcal{M}$ , the inner product can be expressed $\\langle \\pmb{v}, \\pmb{w} \\rangle = \\pmb{v}^T g(\\pmb{x})\\pmb{w}$ . A Riemannian manifold can be characterized by the curvature of the curves defined by a metric. The curvature of a Riemannian manifold can be computed at each point of the curves, while some manifolds have curvatures of a constant value. For example, the unit sphere $\\mathcal{S}$ has constant positive curvature of $+1$ .", + "bbox": [ + 496, + 387, + 890, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 595, + 633, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose a new preconditioned gradient descent method called GAP in the MAML framework. In Section 3.1, we introduce GAP in the inner-loop optimization and describe how to meta-train GAP in the outer-loop optimization. In Section 3.2, we prove that GAP has two desirable properties. In Section 3.3, we provide a low-computational approximation that can be useful for large-scale architectures.", + "bbox": [ + 496, + 622, + 893, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. GAP: Geometry-Adaptive Preconditioner", + "text_level": 1, + "bbox": [ + 500, + 751, + 854, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Inner-loop optimization", + "text_level": 1, + "bbox": [ + 500, + 773, + 718, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider an $L$ -layer neural network $f_{\\theta}(\\cdot)$ with parameters $\\theta = \\{\\mathbf{W}^1, \\dots, \\mathbf{W}^l, \\dots, \\mathbf{W}^L\\}$ . In the standard MAML with task $\\tau \\sim p(\\mathcal{T})$ , each $\\mathbf{W}^l$ is adapted with the gradient update as below:", + "bbox": [ + 496, + 797, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {\\tau , K} ^ {l} \\leftarrow \\mathbf {W} _ {\\tau , 0} ^ {l} - \\alpha \\cdot \\sum_ {k = 0} ^ {K - 1} \\mathbf {G} _ {\\tau , k} ^ {l} s. t \\mathbf {W} _ {\\tau , 0} ^ {l} = \\mathbf {W} ^ {l}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 866, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16082", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Geometry-Adaptive Preconditioned gradient descent (GAP)" + ], + "code_body": "Require: $p(\\mathcal{T})$ : distribution over tasks \nRequire: $\\alpha ,\\beta_{1},\\beta_{2}$ : learning rates 1: Randomly initialize $\\theta = \\{\\mathbf{W}^1,\\dots ,\\mathbf{W}^L\\}$ \n2: Initialize $\\phi = \\{\\mathbf{M}^{1},\\dots ,\\mathbf{M}^{L}\\}$ as identity matrix. \n3: while not converged do \n4: Sample a batch of tasks $\\mathcal{T}_i\\sim p(\\mathcal{T})$ \n5: for all $\\tau \\in \\mathcal{T}_i$ do \n6: for inner-loop step $k = 0$ to $K - 1$ do \n7: for layer $l = 1$ to $L$ do \n8: Compute $\\mathbf{G}_{\\tau ,k}^{l} = \\nabla_{\\mathbf{W}_{\\tau ,k}^{l}}\\mathcal{L}_{\\tau}^{\\mathrm{in}}(\\theta_{\\tau ,k};D_{\\tau}^{\\mathrm{tr}})$ using $D_{\\tau}^{\\mathrm{tr}}$ \n9: Reshape $\\mathbf{G}_{\\tau ,k}^{l}$ to $\\mathbf{G}_{\\tau ,k}^{l}$ via Eq. (6) \n10: Transform $\\mathbf{G}_{\\tau ,k}^{l}$ to $\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}$ using $\\mathbf{M}^l$ via Eq. (8) \n11: Reshape $\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}$ back to the original form of gradient tensor, $\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}$ \n12: Compute $l$ -layer adapted weight: $\\mathbf{W}_{\\tau ,k + 1}^{l} = \\mathbf{W}_{\\tau ,k}^{l} - \\alpha \\cdot \\tilde{\\mathbf{G}}_{\\tau ,k}^{l}$ \n13: end for \n14: end for \n15: Compute $\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})$ by evaluating $\\mathcal{L}_{\\tau}^{\\mathrm{out}}$ w.r.t $D_{\\tau}^{\\mathrm{val}}$ \n16: end for \n17: Update the weights and meta parameters: \n18: $\\theta \\gets \\theta -\\beta_1\\nabla_\\theta \\sum_{\\tau \\in \\mathcal{T}}\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})$ \n19: $\\phi \\leftarrow \\phi -\\beta_{2}\\nabla_{\\phi}\\sum_{\\tau \\in \\mathcal{T}}\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})$ \n20: end while", + "bbox": [ + 76, + 125, + 470, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{G}_{\\tau,k}^{l} = \\nabla_{\\mathbf{W}_{\\tau,k}^{l}} \\mathcal{L}_{\\tau}^{\\mathrm{in}}(\\theta_{\\tau,k}; D_{\\tau}^{\\mathrm{tr}})$ is the gradient with respect to $\\mathbf{W}_{\\tau,k}^{l}$ and $\\alpha$ is the learning rate for inner-loop optimization. In the GAP, we first use the mode-1 unfolding to reshape the gradient tensor into a matrix form (see Section 2.3). For a convolutional layer (i.e., $\\mathbf{G}_{\\tau,k}^{l} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times C_{\\mathrm{in}} \\times k \\times k}$ ), we reshape the gradient tensor as below:", + "bbox": [ + 75, + 405, + 472, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {\\tau , k} ^ {l} \\xrightarrow [ \\text {m o d e - 1} ]{} \\left\\{ \\begin{array}{l l} \\mathbf {G} _ {\\tau , k} ^ {l} \\in \\mathbb {R} ^ {C _ {\\mathrm {o u t}} \\times C _ {\\mathrm {i n}} k ^ {2}} & \\text {i f} C _ {\\mathrm {o u t}} \\leq C _ {\\mathrm {i n}} k ^ {2} \\\\ \\mathbf {G} _ {\\tau , k} ^ {l} \\in \\mathbb {R} ^ {C _ {\\mathrm {i n}} k ^ {2} \\times C _ {\\mathrm {o u t}}} & \\text {i f} C _ {\\mathrm {i n}} k ^ {2} < C _ {\\mathrm {o u t}}, \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 516, + 468, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{G}_{\\tau ,k}^{l}$ denotes $(\\mathbf{G}_{\\tau ,k}^{l})_{[1]}$ for the notational brevity. Note that we chose mode-1 unfolding because it performs best among the four unfolding forms as shown in Table 1. For a linear layer in a matrix form, there is no need for an unfolding. Second, we transform the singular values of the gradient matrix using additional meta parameters $\\phi = \\{\\mathbf{M}^l\\}_{l = 1}^L$ . The meta parameter $\\mathbf{M}^l$ are diagonal matrices with positive elements defined as:", + "bbox": [ + 75, + 569, + 472, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} ^ {l} = \\left\\{ \\begin{array}{l l} \\operatorname {d i a g} \\left(\\operatorname {S p} \\left(m _ {1} ^ {l}\\right), \\dots , \\operatorname {S p} \\left(m _ {C _ {\\text {o u t}}} ^ {l}\\right)\\right) & \\text {i f} C _ {\\text {o u t}} \\leq C _ {\\text {i n}} k ^ {2} \\\\ \\operatorname {d i a g} \\left(\\operatorname {S p} \\left(m _ {1} ^ {l}\\right), \\dots , \\operatorname {S p} \\left(m _ {C _ {\\text {i n}} k ^ {2}} ^ {l}\\right)\\right) & \\text {i f} C _ {\\text {i n}} k ^ {2} < C _ {\\text {o u t}} \\end{array} , \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 702, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m_i^l\\in \\mathbb{R}$ and $\\mathbf{Sp}(x) = \\frac{1}{2}\\cdot \\log (1 + \\exp (2*x))$ . They are applied to the gradient matrix as follows:", + "bbox": [ + 75, + 750, + 468, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {G}} _ {\\tau , k} ^ {l} = \\mathbf {U} _ {\\tau , k} ^ {l} \\left(\\mathbf {M} ^ {l} \\cdot \\boldsymbol {\\Sigma} _ {\\tau , k} ^ {l}\\right) \\mathbf {V} _ {\\tau , k} ^ {l}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 792, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{G}_{\\tau,k}^{l} = \\mathbf{U}_{\\tau,k}^{l}\\pmb{\\Sigma}_{\\tau,k}^{l}\\mathbf{V}_{\\tau,k}^{l}$ is the singular value decomposition (SVD) of $\\mathbf{G}_{\\tau,k}^{l}$ . Finally, we reshape $\\tilde{\\mathbf{G}}_{\\tau,k}^{l}$ back to its original gradient tensor form $\\tilde{\\mathbf{G}}_{\\tau,k}^{l}$ (i.e., inverse unfolding). The resulting preconditioned gradient descent of GAP", + "bbox": [ + 75, + 827, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/08b7499f3b2c145998995c12708492f12709c46f15481919a0d1ed8208dc089b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
mode-1mode-2mode-3mode-4
54.86 ± 0.8553.02 ± 0.8751.23 ± 0.7651.45 ± 0.77
", + "bbox": [ + 513, + 89, + 877, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Performance comparison of unfolding types. We performed the experiment with 5-way 1-shot on mini-[ImageNet] and used the standard Conv-4 backbone.", + "bbox": [ + 498, + 131, + 893, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "becomes the following:", + "bbox": [ + 500, + 189, + 658, + 205 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {\\tau , K} ^ {l} \\leftarrow \\mathbf {W} _ {\\tau , 0} ^ {l} - \\alpha \\cdot \\sum_ {k = 0} ^ {K - 1} \\tilde {\\mathbf {G}} _ {\\tau , k} ^ {l} s. t \\mathbf {W} _ {\\tau , 0} ^ {l} = \\mathbf {W} ^ {l}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 214, + 890, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}$ is the preconditioned gradient based on the meta parameters $\\phi$", + "bbox": [ + 498, + 263, + 890, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Outer-loop optimization", + "text_level": 1, + "bbox": [ + 500, + 316, + 723, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For outer-loop optimization, GAP follows the typical process of MAML. Unlike MAML, however, GAP meta-learns two meta parameter sets $\\theta$ and $\\phi$ as follows:", + "bbox": [ + 498, + 340, + 890, + 385 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta \\leftarrow \\theta - \\beta_ {1} \\nabla_ {\\theta} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 397, + 890, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\leftarrow \\phi - \\beta_ {2} \\nabla_ {\\phi} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 417, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta_{1}$ and $\\beta_{2}$ are the learning rates for the outer-loop optimization. We initialize $\\mathbf{M}^l$ as an identity matrix for all $l$ . The training procedure is provided in Algorithm 1.", + "bbox": [ + 496, + 446, + 893, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Desirable properties of GAP", + "text_level": 1, + "bbox": [ + 500, + 502, + 754, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we prove that GAP's preconditioner $\\mathbf{P}_{\\mathrm{GAP}}$ satisfies two desirable properties.", + "bbox": [ + 498, + 525, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Let $\\tilde{\\mathbf{G}}_{\\tau,k}^{l} \\in \\mathbb{R}^{m \\times n}$ be the 'l-layer $k$ -th inner-step' gradient matrix transformed by meta parameter $\\mathbf{M}^l$ for task $\\tau$ . Then preconditioner $\\mathbf{P}_{GAP}$ induced by $\\tilde{\\mathbf{G}}_{\\tau,k}^{l}$ is a Riemannian metric and depends on the task-specific parameters $\\theta_{\\tau,k}$ .", + "bbox": [ + 498, + 565, + 893, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof and closed form of $\\mathbf{P}_{\\mathrm{GAP}}$ are provided in Supplementary Section B. The following two properties emerge from the theorem.", + "bbox": [ + 498, + 650, + 893, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Property 1. Dependency on task-specific parameters: Theorem 1 formally shows that $\\mathbf{P}_{\\mathrm{GAP}}$ depends on the task-specific parameters $\\theta_{\\tau,k}$ . While the previous studies considered a non-adaptive preconditioner $\\mathbf{P}(\\phi)$ [34, 36, 44, 49, 57, 66] and a partially adaptive preconditioner $\\mathbf{P}(k;\\phi)$ [49] or $\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}};\\phi)$ [53], $\\mathbf{P}_{\\mathrm{GAP}}$ can be considered to be the most advanced adaptive preconditioner because it is fully adaptive (i.e., task-specific and path-dependent) by being dependent on $\\theta_{\\tau,k}$ as shown in Figure 1.", + "bbox": [ + 496, + 707, + 893, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Property 2. Riemannian metric: If the parameter space has a certain underlying structure, the ordinary gradient of a function $\\nabla \\mathcal{L}$ does not represent its steepest direction [4].", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16083", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To define the steepest direction on the parameter space, we need a Riemannian metric $g(\\boldsymbol{w})$ , which is a positive-definite matrix defined for each parameter $\\boldsymbol{w}$ . A Riemannian metric defines the steepest descent direction by $-g(\\boldsymbol{w})^{-1}\\nabla \\mathcal{L}$ [4]. If a preconditioning matrix is a Riemannian metric, it defines the geometry of the underlying structure and enables steepest descent learning. Because we prove that $\\mathbf{P}_{\\mathrm{GAP}}$ is a Riemannian metric for each parameter in Theorem 1, $\\mathbf{P}_{\\mathrm{GAP}}$ is theoretically guaranteed to enable steepest descent learning on its corresponding parameter space. $\\mathbf{P}_{\\mathrm{GAP}}$ consists of two factors, a unitary matrix of the inner-loop gradient $\\mathbf{U}_{\\tau,k}$ and a meta-parameter $\\mathbf{M}$ ; $\\mathbf{M}$ enables us to reflect the shared geometry information across the tasks. Task-specific and path-dependent geometry information can be reflected in the metric through $\\mathbf{U}_{\\tau,k}$ . Two factors allow our Riemannian metric to have higher function complexity than a constant metric. For example, we can consider various structures other than a unit-sphere that corresponds to the constant metric of $+1$ . Even though $\\mathbf{P}_{\\mathrm{GAP}}$ is guaranteed to be a Riemannian metric, it is crucial that the meta-learned $\\mathbf{P}_{\\mathrm{GAP}}$ corresponds to the true parameter space or at least $\\mathbf{P}_{\\mathrm{GAP}}$ is close enough to be useful. We will discuss this issue in Section 5.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Approximate GAP: a low-computational approximation of GAP", + "text_level": 1, + "bbox": [ + 76, + 433, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As presented in Section 3.1, GAP uses an SVD operation. The SVD operation can be burdensome for large-scale networks because it implies that the computational cost can be significantly increased. In recent studies, the use of a large-scale architecture has been emphasized as the key factor for improving the performance of meta-learning [25]. To make use of GAP for large-scale architectures without causing a computational problem, we provide an efficient approximation, named Approximate GAP, under an assumption.", + "bbox": [ + 75, + 470, + 472, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assumption 1. The elements of the gradient matrix follow an i.i.d. normal distribution with zero means.", + "bbox": [ + 75, + 618, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [39, 60], we adopt the assumption to have the gradient matrix become orthogonal as $n$ increases. Although the utilization of the assumption is a limiting factor, we empirically confirmed that the row vectors of the gradient matrix are indeed asymptotically orthogonal as $n$ increases (see Figure 3). For the assumption, our approximation can be established as the following.", + "bbox": [ + 75, + 659, + 468, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. Let $\\mathbf{G} \\in \\mathbb{R}^{m \\times n}$ be a gradient matrix and $\\tilde{\\mathbf{G}}$ be the gradient transformed by a meta parameter $\\mathbf{M}$ . Under the Assumption 1, as $n$ becomes large, $\\tilde{\\mathbf{G}}$ asymptotically becomes equivalent to $\\mathbf{MG}$ as follows:", + "bbox": [ + 75, + 773, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {G}} \\cong \\mathbf {M G} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 847, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we have chosen the larger dimension of the gradient matrix as $n$ when reshaping with Eq. (6). We provide", + "bbox": [ + 76, + 869, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31919d55723fcabb9ab7563c5f8fe11f0bbc340a025fd60acea68c3f1cb5061e.jpg", + "image_caption": [ + "Figure 3. The average of cosine similarity between row vectors of gradient matrix as $n$ increases." + ], + "image_footnote": [], + "bbox": [ + 576, + 90, + 816, + 229 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/60001991cec2eae7bce66791bcda83002be32fb879b0ed2d5622087c54159c82.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Algorithmtrain time (secs)test time (secs)GPU-memory (MB)
MAML35796.6152.6610185
GAP50916.63120.4410279
Approximate GAP36684.7253.6210217
", + "bbox": [ + 514, + 276, + 874, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Comparison of training time, GPU memory, and test time for MAML, GAP, and Approximate GAP. We performed the experiment with 5-way 1-shot on mini-ImageNet and used 600 tasks in the test. We used the standard Conv-4 backbone.", + "bbox": [ + 498, + 330, + 893, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the proof in Supplementary Section B. This approximation has a clear trade-off between scalability and adaptiveness. Approximate GAP efficiently reduces the computational cost as shown in Table 2, whereas the preconditioner $\\mathbf{P}_{\\mathrm{GAP}}$ becomes not adaptive but constant. However, Approximate GAP still guarantees the preconditioner to be a Riemannian metric. As shown in Table 4 & 5, Approximate GAP incurs a slight performance drop but it still achieves a high performance owing to the Riemannian metric property.", + "bbox": [ + 496, + 398, + 893, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 547, + 633, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we show the superiority of GAP by comparing it with the state-of-the-art PGD-MAML family and the MAML family. Hyper-parameter setups used in our experiments can be found in Supplementary Section C.1.", + "bbox": [ + 498, + 573, + 893, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Few-shot regression", + "text_level": 1, + "bbox": [ + 500, + 643, + 689, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and experimental setup. The goal of few-shot regression is to fit an unknown target function for the given $K$ sample points from the function. For the evaluation of few-shot regression, we use the sinusoid regression benchmark [20]. In this benchmark, sinusoid is used as the target function. Each task has a sinusoid $y(x) = A\\sin (\\omega x + b)$ as the target function, where the parameter values are within the following range: amplitude $A\\in [0.1,5.0]$ , frequency $\\omega \\in [0.8,1.2]$ , and phase $b\\in [0,\\pi ]$ . For each task, input data point $x$ is sampled from $[-5.0,5.0]$ . In the experiment, we use a simple Multi-Layer Perceptron (MLP), following the setting in [20]. The details of the architecture are provided in Supplementary Section C.2.1.", + "bbox": [ + 496, + 666, + 893, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. We evaluate GAP and compare it with MAML family and PGD-MAML family on a regression task. As", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16084", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fbd668e3fb1a36a0460ccab608a6ecbfe2a656ff081258005b6e9b0c7e10ea17.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Few-shot regression for the sinusoid regression benchmark with a 2-layer MLP backbone. We report MSE ± 95% confidence intervals(ci) for 600 tasks following the setup in [20].† denotes PGD-MAML family." + ], + "table_body": "
Algorithm5-shot10-shot20-shot
MAML [20]1.13 ± 0.180.77 ± 0.110.48 ± 0.08
Meta-SGD† [36]0.90 ± 0.160.53 ± 0.090.31 ± 0.05
MT-Net [34]0.76 ± 0.090.49 ± 0.050.33 ± 0.04
ALFA [9]0.92 ± 0.190.62 ± 0.160.34 ± 0.07
L2F [10]0.71 ± N/A0.37 ± N/A0.16 ± N/A
PAMELA† [49]0.54 ± 0.060.41 ± 0.040.17 ± 0.03
MeTAL [8]0.74 ± 0.180.44 ± 0.110.21 ± 0.06
GAP†0.16 ± 0.040.04 ± 0.010.01 ± 0.01
", + "bbox": [ + 94, + 88, + 447, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shown in Table 3, GAP consistently achieves the lowest mean squared error (MSE) scores, with the lowest confidence intervals in all three cases. The performance of GAP is improved by $89\\%$ on 10-shot and $94\\%$ on 20-shot compared to the performance of state-of-the-art algorithms.", + "bbox": [ + 75, + 268, + 470, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Few-shot classification", + "text_level": 1, + "bbox": [ + 76, + 354, + 284, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and experimental setup. For the few-shot classification, we evaluate two benchmarks: (1) miniImageNet [56]; this dataset has 100 classes and it is a subset of ImageNet [16], and we use the same split as in [51], with 64, 16 and 20 classes for train, validation and test, respectively. (2) tiered-ImageNet [52]; this is also a subset of ImageNet with 608 classes grouped into 34 high-level categories, and divided into 20, 6 and 8 for train, validation, and test, respectively. For all the experiments, our model follows the standard Conv-4 backbone used in [56]. The details of the architecture are provided in Supplementary Section C.2.2. Following the experimental protocol in [20], we use 15 samples per class in the query-set to compute the meta gradients. In meta training and meta testing, the inner-loop optimization is updated in five and ten steps, respectively.", + "bbox": [ + 75, + 378, + 473, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. Table 4 & 5 present the performance of GAP, the state-of-the-art PGD-MAML family, and the state-of-the-art MAML-family on mini-[ImageNet and tiered-[ImageNet under two typical settings: 5-way 1-shot and 5-way 5-shot. The GAP outperforms all of the previous PGD-MAML family and MAML family. Compared to the state-of-the-art MAML family, GAP improves the performance with a quite significant margin for both mini-[ImageNet and tiered-[ImageNet datasets. Compared to the state-of-the-art PGD-MAML family, GAP shows that the 1- and 5-shot accuracy can be increased by $1.4\\%$ and $1.5\\%$ on mini-[ImageNet dataset, and by $0.7\\%$ and $0.68\\%$ on tiered-[ImageNet dataset, respectively. We also evaluated Approximate GAP that is introduced in Section 3.3. The results show that the approximated version can perform comparably to the original GAP. Although Approximate GAP shows slightly lower accuracies than the original, its performance is superior to most of the existing algorithms because of its Riemannian metric property.", + "bbox": [ + 75, + 613, + 473, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3b1cb1c7b02d9f29066bd155430a911ddd64285053312f4738a51ad0eb989752.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. 5-way few-shot classification accuracy $(\\%)$ on miniImageNet with a Conv-4 backbone. We report mean $\\pm 95\\%$ confidence intervals(ci) for 600 tasks according to [20]. $\\dagger$ denotes PGD-MAML family." + ], + "table_body": "
Algorithm5-way 1-shot5-way 5-shot
MAML [20]47.89 ± 1.2064.59 ± 0.88
Meta-SGD† [36]50.47 ± 1.8764.00 ± 0.90
BMAML [64]53.80 ± 1.4664.23 ± 0.69
ANIL [48]46.70 ± 0.4061.50 ± 0.50
LLAMA [23].49.40 ± 1.83N/A
PLATIPUS [21]50.13 ± 1.86-
T-net [34]50.86 ± 1.82N/A
MT-net [34]51.70 ± 1.84N/A
MAML++ [7]52.15 ± 0.2668.32 ± 0.44
iMAML-HF [50]49.30 ± 1.88N/A
WarpGrad [22]52.30 ± 0.9068.40 ± 0.60
MC1† [44]53.74 ± 0.8468.01 ± 0.73
MC2† [44]54.08 ± 0.8867.99 ± 0.73
MH-C† [66]48.64 ± 0.3364.52 ± 0.51
MH† [66]49.41 ± 0.9667.16 ± 0.42
BOIL [43]49.61 ± 0.1666.46 ± 0.37
ARML [63]50.42 ± 1.7964.12 ± 0.90
ALFA [9]50.58 ± 0.5169.12 ± 0.47
L2F [10]52.10 ± 0.5069.38 ± 0.46
ModGrad† [53]53.20 ± 0.8669.17 ± 0.69
PAMELA† [49]53.50 ± 0.8970.51 ± 0.67
SignMAML [19]42.90 ± 1.5060.70 ± 0.70
CTML [45]50.47 ± 1.8364.15 ± 0.90
MeTAL [8]52.63 ± 0.3770.52 ± 0.29
ECML [24]48.94 ± 0.8065.26 ± 0.67
Sharp-MAML_up [1]49.56 ± N/A63.06 ± N/A
Sharp-MAML_low [1]49.72 ± N/A63.18 ± N/A
Sharp-MAMLboth [1]50.28 ± N/A65.04 ± N/A
FBM [62]50.62 ± 1.7964.78 ± 0.35
CxGrad [33]51.80 ± 0.4669.82 ± 0.42
HyperMAML [47]51.84 ± 0.5766.29 ± 0.43
EEML [35]52.42 ± 1.7568.40 ± 0.95
MH-O† [66]52.50 ± 0.6167.76 ± 0.34
Sparse-MAML† [57]50.35 ± 0.3967.03 ± 0.74
Sparse-ReLU-MAML† [57]50.39 ± 0.8964.84 ± 0.46
Sparse-MAML+† [57]51.04 ± 0.5967.03 ± 0.74
Approximate GAP†53.52 ± 0.8870.75 ± 0.67
GAP†54.86 ± 0.8571.55 ± 0.61
", + "bbox": [ + 517, + 87, + 870, + 534 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f48b7ebe878150bbeb66f33a2582bf6ac729d8894b7955bdef8a5fa9ada7f212.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. 5-way few-shot classification accuracy (%) on tiered-ImageNet dataset with a Conv-4 backbone. We report mean ± 95% confidence intervals(ci) for 600 tasks according to [20].† denotes PGD-MAML family." + ], + "table_body": "
Algorithm5-way 1-shot5-way 5-shot
Meta-SGD†[36]50.92 ± 0.9369.28 ± 0.80
MAML [20]51.70 ± 1.8070.30 ± 1.80
MT-net [34]51.95 ± 1.83N/A
WarpGrad [22]57.20 ± 0.9074.10 ± 0.70
BOIL [43].48.58 ± 0.2769.37 ± 0.12
ALFA [9]53.16 ± 0.4970.54 ± 0.46
L2F [10]54.40 ± 0.5073.34 ± 0.44
ARML [63]52.91 ± 1.83N/A
PAMELA†[49]54.81 ± 0.8874.39 ± 0.71
Sparse-ReLU-MAML†[57]53.18 ± 0.5269.06 ± 0.28
Sparse-MAML†[57]53.47 ± 0.5368.83 ± 0.65
Sparse-MAML+†[57]53.91 ± 0.6769.92 ± 0.21
MeTAL [8]54.34 ± 0.3170.40 ± 0.21
CxGrad [33]55.55 ± 0.4673.55 ± 0.41
ECML [24]47.34 ± 0.8864.77 ± 0.75
Approximate GAP†56.86 ± 0.9174.41 ± 0.72
GAP†57.60 ± 0.9374.90 ± 0.68
", + "bbox": [ + 517, + 595, + 870, + 811 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16085", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bb85ba485c0ae5c303032b7042dc412448487831e4704d755dc6f458b848194e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Algorithmtiered-ImageNetCUBCars
1-shot5-shot1-shot5-shot1-shot5-shot
MAML [20]51.61 ± 0.2065.76 ± 0.2740.51 ± 0.0853.09 ± 0.1633.57 ± 0.1444.56 ± 0.21
ANIL [48]52.82 ± 0.2966.52 ± 0.2841.12 ± 0.1555.82 ± 0.2134.77 ± 0.3146.55 ± 0.29
BOIL [43]53.23 ± 0.4169.37 ± 0.2344.20 ± 0.1560.92 ± 0.1136.12 ± 0.2950.64 ± 0.22
BMAML [64]N/AN/A33.52 ± 0.3651.35 ± 0.16N/AN/A
ALFA [9]N/AN/AN/A58.35 ± 0.25N/AN/A
L2F [10]N/AN/AN/A60.89 ± 0.22N/AN/A
MeTAL [8]N/AN/AN/A58.20 ± 0.24N/AN/A
HyperMAML [47]N/AN/A36.52 ± 0.6149.43 ± 0.14N/AN/A
CxGrad [33]N/AN/AN/A63.92 ± 0.44N/AN/A
Sparse-MAML†[57]53.47 ± 0.5368.83 ± 0.6541.37 ± 0.7360.58 ± 1.1035.90 ± 0.5052.63 ± 0.56
Sparse-ReLU-MAML†[57]53.77 ± 0.9468.12 ± 0.6942.89 ± 0.4557.53 ± 0.9436.04 ± 0.5549.95 ± 0.42
Sparse-MAML+†[57]53.91 ± 0.6769.92 ± 0.2143.43 ± 1.0462.02 ± 0.7837.14 ± 0.7753.18 ± 0.44
Approximate GAP†57.47 ± 0.9971.66 ± 0.7643.77 ± 0.7962.92 ± 0.7337.00 ± 0.7553.28 ± 0.76
GAP†58.56 ± 0.9372.82 ± 0.7744.74 ± 0.7564.88 ± 0.7238.44 ± 0.7755.04 ± 0.77
", + "bbox": [ + 135, + 88, + 831, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. 5-way few-shot cross domain classification accuracy (%) with a Conv-4 backbone, meta training on mini-[ImageNet dataset, and meta-testing on tiered-[ImageNet, CUB, or Cars datasets. We report mean ± 95% confidence intervals(ci) for 600 tasks according to [20].† denotes PGD-MAML family.", + "bbox": [ + 75, + 301, + 893, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Cross-domain few-shot classification", + "text_level": 1, + "bbox": [ + 76, + 371, + 393, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The cross-domain few-shot classification introduced by [14] addresses a more challenging and practical few-shot classification scenario in which meta-train and meta-test tasks are sampled from different task distributions. These scenarios are designed to evaluate meta-level overfitting of meta-learning algorithms by creating a large domain gap between meta-trains and meta-tests. In particular, an algorithm can be said to be meta-overfitting if it relies too much on the prior knowledge of previously seen meta-train tasks instead of focusing on a few given examples to learn a new task. This meta-level overfitting makes the learning system more likely to fail to adapt to new tasks sampled from substantially different task distributions.", + "bbox": [ + 75, + 393, + 473, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and experimental setup. To evaluate the level of meta-overfitting for GAP, we evaluate a cross-domain few-shot classification experiment. The mini-ImageNet is used for the meta-train task, and the tiered-ImageNet [52], CUB-200-2011 [58], Cars [11] datasets are used for the meta-test task. The CUB dataset has 200 fine-grained classes and consists of a total of 11,788 images; it is further divided into 100 meta-train classes, 50 meta-validation classes, and 50 meta-test classes. The Cars [30] dataset consists of 16,185 images of 196 classes of cars; it is split into 8,144 training images and 8,041 testing images, where each class has been split roughly in 50-50. The classes are typically at the level of Make, Model, Year, e.g., 2012 Tesla Model S or 2012 BMW M3 coupe. As with the few-shot classification experiment, we use the standard Conv-4 backbone and follow the same experimental protocol.", + "bbox": [ + 75, + 604, + 472, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Table 6 presents the cross-domain few-shot performance for GAP, MAML family, and PGD-MAML family. GAP significantly outperforms the state-of-the-art al", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gorithms on 5-way 1-shot and 5-way 5-shot cross-domain classification tasks. In particular, for the tiered-ImageNet dataset, the performance was improved by $8.6\\%$ and $4.1\\%$ on 1-shot and 5-shot classification tasks, respectively. Because GAP can simultaneously consider a task's individuality and optimization trajectory in the inner-loop optimization, it can overcome meta-overfitting better than the existing methods. However, Approximate GAP shows more performance degradation in cross-domain few-shot classification than in few-shot classification. In particular, when the domain difference with the meta-train is more significant (i.e., the tiered-ImageNet dataset) than when the domain difference with the meta-train is marginal (i.e., CARS and CUB datasets), it shows a more considerable performance drop. We can see that full adaptation plays an important role in cross-domain few-shot classification.", + "bbox": [ + 496, + 371, + 897, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 498, + 628, + 614, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of meta parameters: Recent MAML family and PGD-MAML family require a large increase in the number of meta-learning parameters as shown in Table 7. One advantage of GAP is that it requires only a very small increase in the number of meta parameters, when compared to the baseline of MAML. This is possible because we transform a gradient tensor into a gradient matrix, perform the SVD of the matrix, and assign only a small number of meta parameters that correspond to the diagonal matrix of the gradient matrix. For the Conv-4 network, GAP requires only $0.2\\%$ increase of the meta parameters. Although the increase in the number of meta parameters is negligible, SVD of the gradient matrix can incur a large computational burden for large networks. This is addressed by Approximate GAP.", + "bbox": [ + 496, + 648, + 893, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Approximate GAP vs. simple constant preconditioners: Approximate GAP is a low-complexity method where SVD", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16086", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c3f61bf0f6eedafdbe3b76cc3dab9dd858577028da8b0594dcaf2519de0b1862.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Algorithm# of params% increase
MAML [20]1.2109 × 105
Meta-SGD [36]2.4218 × 105100.0%
MC [44]2.7106 × 1062140.4%
PAMELA [49]1.6239 × 10534.1%
MH [66]7.2196 × 10759586.7%
Sparse-MAML [57]2.4218 × 105100.0%
GAP1.2131 × 1050.2%
", + "bbox": [ + 99, + 88, + 444, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa8e326b7def17789b540bcf82ec93d1c4420aa5ac0b88aad471a18f41286c3a.jpg", + "table_caption": [ + "Table 7. Comparison of the number of parameters for MAML, existing methods, and GAP." + ], + "table_footnote": [], + "table_body": "
AlgorithmStructureRiemannian metric (i.e., positive definite)Acc. (%)
Meta-SGDdiag(a1, ·, an)X50.47%
Meta-SGD with positive definitenessdiag(a1, ·, an)O52.39%
Approximate GAPblkdiag(Mt, ·, M)O53.52%
", + "bbox": [ + 96, + 262, + 444, + 315 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b6a5ebed9c63bf234c92763df44f976deabf9d547bd64fc74b04ad8232397df7.jpg", + "table_caption": [ + "Table 8. Performance comparison of three constant (i.e., non-adaptive) preconditioners for 5-way 1-shot on mini-ImageNet: Meta-SGD, Meta-SGD modified to satisfy positive definiteness, and our Approximate GAP." + ], + "table_footnote": [], + "table_body": "
Algorithm1-shot5-shot
GAP w/o PGAP48.23 ± 0.8065.80 ± 0.75
GAP w/ PGAP54.86 ± 0.8571.55 ± 0.61
", + "bbox": [ + 98, + 388, + 444, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation study of $\\mathbf{P}_{\\mathrm{GAP}}$ on mini-ImageNet. Performance of the GAP-trained model is significantly affected by not applying $\\mathbf{P}_{\\mathrm{GAP}}$ .", + "bbox": [ + 76, + 450, + 467, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "operation is avoided by approximating GAP with a constant diagonal preconditioner. A natural question to ask is how does Approximate GAP compare with other constant diagonal preconditioners. To answer this question, we have compared Approximate GAP with Meta-SGD and a modified Meta-SGD. Meta-SGD [36] is a well-known constant diagonal preconditioner (i.e., $\\mathrm{diag}(a_1,\\dots ,a_n)$ ) that does not need to be positive definite and we also investigate its modification with a constraint on positive definiteness. The results are shown in Table 8. It can be observed that enforcing positive definiteness can improve Meta-SGD. Furthermore, an additional improvement can be achieved by Approximate GAP. While both modified Meta-SGD and Approximate GAP are positive definite, Approximate GAP is different because it inherits an additional constraint from GAP - a block diagonal structure where a constant diagonal matrix $M$ is repeated (i.e., $\\mathrm{blkdiag}(M,\\dots ,M)$ ). The inherited constraint provides a gain over the modified Meta-SGD.", + "bbox": [ + 75, + 512, + 468, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does GAP learn a useful preconditioner: While a Riemannian metric can be helpful, it does not mean any Riemannian metric will result in an improvement. For the true parameter space with a specific underlying structure, the corresponding Riemannian metric needs to be applied to enable steepest descent [3, 4]. For the special case of a two-layer neural network with a mean squared error (MSE) loss, it", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "was proven that Fisher information matrix is the corresponding Riemannian metric [4]. For a general neural network, however, a proper Riemannian metric is unknown and it needs to be learned. In our work, we have devised a method to guarantee a Riemannian metric and have used the outer-loop optimization to learn the Riemannian metric. In general, the learned Riemannian metric is unlikely to correspond perfectly to the true parameter space. Then, an important question is if the Riemannian metric learned by GAP is close enough to the desired one and if it is useful. To investigate this issue, we performed an ablation study by not applying the preconditioner $\\mathbf{P}_{\\mathrm{GAP}}$ . After training a GAP model, we have evaluated the performance with and without applying $\\mathbf{P}_{\\mathrm{GAP}}$ . The results are shown in Table 9 and clearly the preconditioner learned with the outer-loop optimization plays an essential role for improving the performance.", + "bbox": [ + 496, + 90, + 893, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Why is preconditioner helpful for meta-learning: When the batch size is small, the resulting empirical gradient can be noisy [53, 65]. A typical few-shot learning has only a small number of samples for the inner-loop optimization, and its gradient can be noisy. On the other hand, it was shown in [5] that preconditioned gradient descent with a positive definite preconditioner can achieve a lower risk than gradient descent when the labels are noisy, the model is mis-specified, or the signal is misaligned with the features. Under a misalignment, a properly chosen positive definite preconditioner can generalize better than gradient descent [5]. Considering the noisy gradient of inner loop optimization, it can be surmised that a positive definite preconditioner that is adaptive (i.e., a Riemannian metric) can be helpful for improving MAML. Note that the noisy case is in contrast to the case of supervised learning with a large amount of data [5].", + "bbox": [ + 496, + 356, + 893, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 619, + 617, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have proposed GAP that is a PGDMAML algorithm utilizing SVD operation. Thanks to the fully adaptive property, GAP can handle individual tasks with a wide diversity within the MAML framework. Additionally, GAP can enable steepest descent on the parameter space owing to the Riemannian metric property.", + "bbox": [ + 496, + 648, + 893, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 762, + 666, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by ETRI [23ZR1100, A Study of Hyper-Connected Thinking Internet Technology by autonomous connecting, controlling and evolving ways], NRF (NRF-2020R1A2C2007139), IITP [NO.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University)], and the New Faculty Startup Fund from Seoul National University.", + "bbox": [ + 496, + 787, + 893, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16087", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning. arXiv preprint arXiv:2206.03996, 2022. 6", + "[2] Shunichi Amari. A theory of adaptive pattern classifiers. IEEE Transactions on Electronic Computers, (3):299-307, 1967. 1", + "[3] Shun-ichi Amari. Neural learning in structured parameter spaces-natural riemannian gradient. Advances in neural information processing systems, 9, 1996. 1, 8", + "[4] Shun-Ichi Amari. Natural gradient works efficiently in learning. Neural computation, 10(2):251-276, 1998. 1, 3, 4, 5, 8", + "[5] Shun-ichi Amari, Jimmy Ba, Roger Grosse, Xuechen Li, Atsushi Nitanda, Taiji Suzuki, Denny Wu, and Ji Xu. When does preconditioning help or hurt generalization? arXiv preprint arXiv:2006.10732, 2020. 3, 8", + "[6] Shun-Ichi Amari and Scott C Douglas. Why natural gradient? In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), volume 2, pages 1213-1216. IEEE, 1998. 1", + "[7] Antreas Antoniou, Harrison Edwards, and Amos Storkey. How to train your maml. arXiv preprint arXiv:1810.09502, 2018.6", + "[8] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9465-9474, 2021. 1, 6, 7", + "[9] Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. Meta-learning with adaptive hyperparameters. Advances in Neural Information Processing Systems, 33:20755-20765, 2020. 6, 7", + "[10] Sungyong Baik, Seokil Hong, and Kyoung Mu Lee. Learning to forget for meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2379-2387, 2020. 6, 7", + "[11] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 7", + "[12] Irénée-Jules Bienaymé. Considerations à l'appui de la découverte de Laplace sur la loi de probabilité dans la méthode des moindres carrés. Imprimerie de Mallet-Bachelier, 1853. 13", + "[13] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 1", + "[14] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 7", + "[15] Tristan Deleu, Tobias Würfl, Mandana Samiei, Joseph Paul Cohen, and Yoshua Bengio. Torchmeta: A meta-learning library for pytorch. arXiv preprint arXiv:1909.06576, 2019. 14", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[17] Lin Ding, Peng Liu, Wenfeng Shen, Weijia Lu, and Shengbo Chen. Gradient-based meta-learning using uncertainty to weigh loss for few-shot learning. arXiv preprint arXiv:2208.08135, 2022. 1", + "[18] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011. 3", + "[19] Chen Fan, Parikshit Ram, and Sijia Liu. Sign-maml: Efficient model-agnostic meta-learning by signsgd. arXiv preprint arXiv:2109.07497, 2021. 6", + "[20] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 1, 3, 5, 6, 7, 8, 14", + "[21] Chelsea Finn, Kelvin Xu, and Sergey Levine. Probabilistic model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 1, 6", + "[22] Sebastian Flennerhag, Andrei A Rusu, Razvan Pascanu, Francesco Visin, Hujun Yin, and Raia Hadsell. Meta-learning with warped gradient descent. arXiv preprint arXiv:1909.00025, 2019. 6", + "[23] Erin Grant, Chelsea Finn, Sergey Levine, Trevor Darrell, and Thomas Griffiths. Recasting gradient-based meta-learning as hierarchical bayes. arXiv preprint arXiv:1801.08930, 2018.6", + "[24] Markus Hiller, Mehrtash Harandi, and Tom Drummond. On enforcing better conditioned meta-learning for rapid few-shot adaptation. arXiv preprint arXiv:2206.07260, 2022. 6", + "[25] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9068-9077, 2022. 5", + "[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 14", + "[27] Sham M Kakade. A natural policy gradient. Advances in neural information processing systems, 14, 2001. 1", + "[28] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3,14", + "[29] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 3", + "[30] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 7", + "[31] Yann A LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-48. Springer, 2012. 3", + "[32] John M Lee. Smooth manifolds. Springer, 2012. 3, 12", + "[33] Sanghyuk Lee, Seunghyun Lee, and Byung Cheol Song. Contextual gradient scaling for few-shot learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 834-843, 2022. 6, 7" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16088", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Yoonho Lee and Seungjin Choi. Gradient-based meta-learning with learned layerwise metric and subspace. In International Conference on Machine Learning, pages 2927-2936. PMLR, 2018. 1, 4, 6", + "[35] Geng Li, Boyuan Ren, and Hongzhi Wang. Eeml: Ensemble embedded meta-learning. arXiv preprint arXiv:2206.09195, 2022. 6", + "[36] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1, 4, 6, 8", + "[37] Zequn Liu, Ruiyi Zhang, Yiping Song, and Ming Zhang. When does maml work the best? an empirical study on model-agnostic meta-learning in nlp applications. arXiv preprint arXiv:2005.11700, 2020. 1", + "[38] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 14", + "[39] Ahmed M Abdelmoniem, Ahmed Elzanaty, Mohamed-Slim Alouini, and Marco Canini. An efficient statistical-based gradient compression technique for distributed training systems. Proceedings of Machine Learning and Systems, 3:297-322, 2021. 5", + "[40] Gabriel Maicas, Andrew P Bradley, Jacinto C Nascimento, Ian Reid, and Gustavo Carneiro. Training medical image analysis systems like radiologists. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 546-554. Springer, 2018. 1", + "[41] George Marsaglia. Choosing a point from the surface of a sphere. The Annals of Mathematical Statistics, 43(2):645-646, 1972. 13", + "[42] Fei Mi, Minlie Huang, Jiyong Zhang, and Boi Faltings. Meta-learning for low-resource natural language generation in task-oriented dialogue systems. arXiv preprint arXiv:1905.05644, 2019. 1", + "[43] Jaehoon Oh, Hyungjun Yoo, ChangHwan Kim, and Se-Young Yun. Boil: Towards representation change for few-shot learning. arXiv preprint arXiv:2008.08882, 2020. 6, 7", + "[44] Eunbyung Park and Junier B Oliva. Meta-curvature. Advances in Neural Information Processing Systems, 32, 2019. 1, 4, 6, 8", + "[45] Danni Peng and Sinno Pan. Clustered task-aware meta-learning by learning from learning paths. 2021. 6", + "[46] Juan-Manuel Perez-Rua, Xiatian Zhu, Timothy M Hospedales, and Tao Xiang. Incremental few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13846-13855, 2020. 1", + "[47] M Przewięźlikowski, P Przybysz, J Tabor, M Zieba, and P Spurek. Hypermaml: Few-shot adaptation of deep models with hypernetworks. arXiv preprint arXiv:2205.15745, 2022. 6, 7", + "[48] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 1, 6, 7", + "[49] Jathushan Rajasegaran, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Mubarak Shah. Meta-learning" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the learning trends shared across tasks. arXiv preprint arXiv:2010.09291, 2020. 1, 4, 6, 8", + "[50] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 6", + "[51] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. 2016. 6", + "[52] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 6, 7", + "[53] Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. On modulating the gradient for meta-learning. In European Conference on Computer Vision, pages 556-572. Springer, 2020. 1, 4, 6, 8", + "[54] Rishav Singh, Vandana Bharti, Vishal Purohit, Abhinav Kumar, Amit Kumar Singh, and Sanjay Kumar Singh. Metamed: Few-shot medical image classification using gradient-based meta-learning. Pattern Recognition, 120:108111, 2021. 1", + "[55] Xingyou Song, Yuxiang Yang, Krzysztof Choromanski, Ken Caluwaerts, Wenbo Gao, Chelsea Finn, and Jie Tan. Rapidly adaptable legged robots via evolutionary meta-learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3769-3776. IEEE, 2020. 1", + "[56] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016, 6, 14", + "[57] Johannes Von Oswald, Dominic Zhao, Seijin Kobayashi, Simon Schug, Massimo Caccia, Nicolas Zucchet, and João Sacramento. Learning where to learn: Gradient sparsity in meta and continual learning. Advances in Neural Information Processing Systems, 34:5250-5263, 2021. 1, 4, 6, 7, 8", + "[58] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 7", + "[59] Shuhuan Wen, Zeteng Wen, Di Zhang, Hong Zhang, and Tao Wang. A multi-robot path-planning algorithm for autonomous navigation using meta-reinforcement learning based on transfer learning. Applied Soft Computing, 110:107605, 2021. 1", + "[60] Simon Wiedemann, Temesgen Mehari, Kevin Kepp, and Wojciech Samek. Dithered backprop: A sparse and quantized backpropagation algorithm for more efficient deep neural network training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 720-721, 2020. 5", + "[61] Xiongwei Wu, Doyen Sahoo, and Steven Hoi. Meta-rcnn: Meta learning for few-shot object detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1679–1687, 2020. 1", + "[62] Peng Yang, Shaogang Ren, Yang Zhao, and Ping Li. Calibrating cnns for few-shot meta learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2090–2099, 2022. 6" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16089", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Huaxiu Yao, Xian Wu, Zhiqiang Tao, Yaliang Li, Bolin Ding, Ruirui Li, and Zhenhui Li. Automated relational meta-learning. arXiv preprint arXiv:2001.00745, 2020. 6", + "[64] Jaesik Yoon, Taesup Kim, Ousmane Dia, Sungwoong Kim, Yoshua Bengio, and Sungjin Ahn. Bayesian model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 6, 7", + "[65] Yikai Zhang, Hui Qu, Chao Chen, and Dimitris Metaxas. Taming the noisy gradient: train deep neural networks with small batch sizes. In The Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI), 2019. 8", + "[66] Dominic Zhao, Johannes von Oswald, Seijin Kobayashi, João Sacramento, and Benjamin F Grewe. Meta-learning via hypernetworks. 2020. 1, 4, 6, 8" + ], + "bbox": [ + 78, + 90, + 470, + 289 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16090", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_model.json b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f732601102f1e965365457c3d3035fecc4ff0e36 --- /dev/null +++ b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.193, + 0.131, + 0.778, + 0.154 + ], + "angle": 0, + "content": "Meta-Learning with a Geometry-Adaptive Preconditioner" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.18, + 0.808, + 0.199 + ], + "angle": 0, + "content": "Suhyun Kang\\(^{1*}\\), Duhun Hwang\\(^{1*}\\), Moonjung Eo\\(^{1}\\), Taesup Kim\\(^{2}\\), Wonjong Rhee\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.199, + 0.77, + 0.217 + ], + "angle": 0, + "content": "1 Department of Intelligence and Information, Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.217, + 0.729, + 0.234 + ], + "angle": 0, + "content": "2 Graduate School of Data Science, Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.234, + 0.656, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\) IPAI and AIIS, Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.255, + 0.726, + 0.269 + ], + "angle": 0, + "content": "{su_hyun, yelobean, eod87, taesup.kim, wrhee}@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.474, + 0.701 + ], + "angle": 0, + "content": "Model-agnostic meta-learning (MAML) is one of the most successful meta-learning algorithms. It has a bi-level optimization structure where the outer-loop process learns a shared initialization and the inner-loop process optimizes task-specific weights. Although MAML relies on the standard gradient descent in the inner-loop, recent studies have shown that controlling the inner-loop's gradient descent with a meta-learning preconditioner can be beneficial. Existing preconditioners, however, cannot simultaneously adapt in a task-specific and path-dependent way. Additionally, they do not satisfy the Riemannian metric condition, which can enable the steepest descent learning with preconditioned gradient. In this study, we propose Geometry-Adaptive Preconditioned gradient descent (GAP) that can overcome the limitations in MAML; GAP can efficiently meta-learn a preconditioner that is dependent on task-specific parameters, and its preconditioner can be shown to be a Riemannian metric. Thanks to the two properties, the geometry-adaptive preconditioner is effective for improving the inner-loop optimization. Experiment results show that GAP outperforms the state-of-the-art MAML family and preconditioned gradient descent-MAML (PGD-MAML) family in a variety of few-shot learning tasks. Code is available at: https://github.com/Suhyun777/CVPR23-GAP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.729, + 0.21, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.471, + 0.864 + ], + "angle": 0, + "content": "Meta-learning, or learning to learn, enables algorithms to quickly learn new concepts with only a small number of samples by extracting prior-knowledge known as meta-knowledge from a variety of tasks and by improving the generalization capability over the new tasks. Among the meta-learning algorithms, the category of optimization-based meta-learning [8, 17, 20, 21, 48] has been gaining popularity" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.305, + 0.892, + 0.381 + ], + "angle": 0, + "content": "due to its flexible applicability over diverse fields including robotics [55, 59], medical image analysis [40, 54], language modeling [37, 42], and object detection [46, 61]. In particular, Model-Agnostic Meta-Learning (MAML) [20] is one of the most prevalent gradient-based meta-learning algorithms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.893, + 0.533 + ], + "angle": 0, + "content": "Many recent studies have improved MAML by adopting a Preconditioned Gradient Descent (PGD\\(^1\\)) for inner-loop optimization [34, 36, 44, 49, 53, 57, 66]. In this paper, we collectively address PGD-based MAML algorithms as the PGD-MAML family. A PGD is different from the ordinary gradient descent because it performs a preconditioning on the gradient using a preconditioning matrix \\(\\mathbf{P}\\), also called a preconditioner. A PGD-MAML algorithm meta-learns not only the initialization parameter \\(\\theta_0\\) of the network but also the meta-parameter \\(\\phi\\) of the preconditioner \\(\\mathbf{P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.669 + ], + "angle": 0, + "content": "For the inner-loop optimization, \\(\\mathbf{P}\\) was kept static in most of the previous works (Figure 1(b)) [34, 36, 44, 57, 66]. Some of the previous works considered adapting the preconditioner \\(\\mathbf{P}\\) with the inner-step \\(k\\) (Figure 1(c)) [49] and some others with the individual task (Figure 1(d)) [53]. They achieved performance improvement by considering individual tasks and inner-step, respectively, and showed that both factors were valuable. However, both factors have not been considered simultaneously in the existing studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.895, + 0.867 + ], + "angle": 0, + "content": "When a parameter space has a certain underlying structure, there exists a Riemannian metric corresponding the parameter space [3,4]. If the preconditioning matrix is the Riemannian metric, the preconditioned gradient is known to become the steepest descent on the parameter space [2-4,6,27]. An illustration of a toy example is shown in Figure 2. The optimization path of an ordinary gradient descent is shown in Figure 2(a). Compared to the ordinary gradient descent, a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can actually harm the optimization. For the example in Figure 2(b), the preconditioner affects the optimization into an undesirable direction and negatively affects the gradient descent. On the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1PGD in our work should not be confused with Projected Gradient Descent [13]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.206, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16080" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.092, + 0.245, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.186, + 0.209, + 0.197 + ], + "angle": 0, + "content": "(a) MAML" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.094, + 0.396, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.186, + 0.381, + 0.198 + ], + "angle": 0, + "content": "(b) Non-adaptive \\(\\mathbf{P}(\\phi)\\)" + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.104, + 0.55, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.186, + 0.529, + 0.198 + ], + "angle": 0, + "content": "(c) Adaptive \\(\\mathbf{P}(k;\\phi)\\)" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.106, + 0.703, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.185, + 0.692, + 0.198 + ], + "angle": 0, + "content": "(d) Adaptive \\(\\mathbf{P}(D_{\\tau}^{tr};\\phi)\\)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.104, + 0.858, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.186, + 0.845, + 0.198 + ], + "angle": 0, + "content": "(e) Adaptive \\(\\mathbf{P}(\\theta_{\\tau ,k};\\phi)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.205, + 0.894, + 0.314 + ], + "angle": 0, + "content": "Figure 1. Diagram of MAML and PGD-MAML family. For the inner-loop adaptation in each diagram, the dotted lines of the same color indicate that they use a common preconditioning matrix (preconditioner). (a) MAML adaptation: no preconditioner is used (i.e., \\(\\mathbf{P} = \\mathbf{I}\\)). (b) \\(\\mathbf{P}(\\phi)\\): a constant preconditioner is used in the inner-loop where the preconditioner's meta-parameter \\(\\phi\\) is meta-learned. (c) \\(\\mathbf{P}(k; \\phi)\\): a constant preconditioner is used for each inner-step \\(k\\). Preconditioner for each step is meta-learned, but \\(\\mathbf{P}(k, \\phi)\\) is not task-specific. (d) \\(\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}}; \\phi)\\): a constant preconditioner is used for each task. Preconditioner for each task is meta-learned, but \\(\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}}; \\phi)\\) is not dependent on \\(k\\). (e) GAP adapts \\(\\mathbf{P}(\\theta_{\\tau, k}; \\phi)\\): a fully adaptive preconditioner is used where it is task-specific and path-dependent. Instead of saying 'dependent on \\(k\\)', we specifically say it is path-dependent because the exact dependency is on the task-specific parameter set \\(\\theta_{\\tau, k}\\) that is considerably more informative than \\(k\\)." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.339, + 0.252, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.434, + 0.256, + 0.446 + ], + "angle": 0, + "content": "(a) Gradient Descent (GD)" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.339, + 0.428, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.434, + 0.432, + 0.458 + ], + "angle": 0, + "content": "(b) Preconditioned GD (Non-Riemannian metric)" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.469, + 0.337, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.564, + 0.333, + 0.588 + ], + "angle": 0, + "content": "(c) Preconditioned GD (Riemannian metric)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Figure 2. A toy example for illustrating the effect of Riemannian metric (see Supplementary Section A). When the curvature of the parameter space is poorly conditioned, (a) gradient descent can suffer from the difficulty of finding the solution, (b) a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can suffer further, and (c) a preconditioned gradient descent with a preconditioner that is a Riemannian metric can perform better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.471, + 0.866 + ], + "angle": 0, + "content": "contrary, if the preconditioner is the Riemannian metric corresponding the parameter space, the preconditioned gradient descent can become the steepest descent and can exhibit a better optimization behavior as shown in Figure 2(c). While the Riemannian metric condition (i.e., positive definiteness) is a necessary condition for steepest descent learning, the existing studies on PGD-MAML family did not consider constraining preconditioners to satisfy the condition for Riemannian metric." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In this study, we propose a new PGD method named Geometry Aaptive Preconditioned gradient descent (GAP)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.895, + 0.527 + ], + "angle": 0, + "content": "Specifically, GAP satisfies two desirable properties which have not been considered before. First, GAP's preconditioner \\(\\mathbf{P}_{\\mathrm{GAP}}\\) is a fully adaptive preconditioner that can adapt to the individual task (task-specific) and to the optimization-path (path-dependent). The full adaptation is made possible by having the preconditioner depend on the task-specific parameter \\(\\theta_{\\tau,k}\\) (Figure 1(e)). Second, we prove that \\(\\mathbf{P}_{\\mathrm{GAP}}\\) is a Riemannian metric. To this end, we force the metaparameters of \\(\\mathbf{P}_{\\mathrm{GAP}}\\) to be positive definite. Thus, GAP guarantees the steepest descent learning on the parameter space corresponding to \\(\\mathbf{P}_{\\mathrm{GAP}}\\). Owing to the two properties, GAP enables a geometry-adaptive learning in inner-loop optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.648 + ], + "angle": 0, + "content": "For the implementation of GAP, we utilize the Singular Value Decomposition (SVD) operation to come up with our preconditioner satisfying the desired properties. For the recently proposed large-scale architectures, computational overhead can be an important design factor and we provide a low-computational approximation, Approximate GAP, that can be proven to asymptotically approximate the operation of GAP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.894, + 0.739 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of GAP, we empirically evaluate our algorithm on popular few-shot learning tasks; few-shot regression, few-shot classification, and few-shot cross-domain classification. The results show that GAP outperforms the state-of-the-art MAML family and PGDMAML family." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.74, + 0.892, + 0.768 + ], + "angle": 0, + "content": "The main contributions of our study can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.771, + 0.894, + 0.832 + ], + "angle": 0, + "content": "- We propose a new preconditioned gradient descent method called GAP, where it learns a preconditioner that enables a geometry-adaptive learning in the inner-loop optimization." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "- We prove that GAP's preconditioner has two desirable properties: (1) It is both task-specific and path-dependent (i.e., dependent on task-specific parameter \\(\\theta_{\\tau, k}\\)). (2) It is a Riemannian metric." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.771, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16081" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "- For large-scale architectures, we provide a low-computational approximation that can be theoretically shown to approximate the GAP method." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.149, + 0.471, + 0.195 + ], + "angle": 0, + "content": "- For popular few-shot learning benchmark tasks, we empirically show that GAP outperforms the state-of-the-art MAML family and PGD-MAML family." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.471, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.21, + 0.206, + 0.226 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.235, + 0.431, + 0.251 + ], + "angle": 0, + "content": "2.1. Model-Agnostic Meta-Learning (MAML)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.258, + 0.471, + 0.365 + ], + "angle": 0, + "content": "The goal of MAML [20] is to find the best initialization that the model can quickly adapt from, such that the model can perform well for a new task. MAML consists of two levels of main optimization processes: inner-loop and outer-loop optimizations. Consider the model \\( f_{\\theta}(\\cdot) \\) with parameter \\( \\theta \\). For a task \\( \\tau = \\{D_{\\tau}^{\\mathrm{tr}}, D_{\\tau}^{\\mathrm{val}}\\} \\) sampled from the task distribution \\( p(\\mathcal{T}) \\), the inner-loop optimization is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.377, + 0.47, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\tau , K} = \\theta_ {\\tau , 0} - \\alpha \\sum_ {k = 0} ^ {K - 1} \\nabla_ {\\theta_ {\\tau , k}} \\mathcal {L} _ {\\tau} ^ {\\mathrm {i n}} \\left(\\theta_ {\\tau , k}; D _ {\\tau} ^ {\\mathrm {t r}}\\right) s. t \\theta_ {\\tau , 0} = \\theta , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.424, + 0.47, + 0.498 + ], + "angle": 0, + "content": "where \\(\\theta_{\\tau ,k}\\) is task-specific parameters for task \\(\\tau\\) and \\(\\alpha\\) is the learning rate for inner-loop optimization, \\(\\mathcal{L}_{\\tau}^{\\mathrm{in}}\\) is the inner-loop's loss function, and \\(K\\) is the number of gradient descent steps. With \\(D_{\\tau_i}^{\\mathrm{val}}\\) in each task, we can define outer-loop optimization as" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.507, + 0.469, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.54, + 0.47, + 0.571 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the learning rate for outer-loop optimization, and \\(\\mathcal{L}_{\\tau}^{\\mathrm{out}}\\) is the outer-loop's loss function." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.581, + 0.423, + 0.597 + ], + "angle": 0, + "content": "2.2. Preconditioned Gradient Descent (PGD)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.605, + 0.47, + 0.681 + ], + "angle": 0, + "content": "PGD is a method that minimizes the empirical risk through a gradient update with a preconditioner that rescales the geometry of the parameter space. Given model parameters \\(\\theta\\) and task \\(\\tau = \\{D_{\\tau}^{\\mathrm{tr}}, D_{\\tau}^{\\mathrm{val}}\\}\\), we can define the preconditioned gradient update with a preconditioner \\(\\mathbf{P}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.704, + 0.47, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta_ {\\tau , k + 1} = \\theta_ {\\tau , k} - \\alpha \\mathbf {P} \\nabla_ {\\theta_ {\\tau , k}} \\mathcal {L} _ {\\tau} \\left(\\theta_ {\\tau , k}; D _ {\\tau} ^ {\\mathrm {t r}}\\right) \\\\ k = 0, 1, \\dots \\text {a n d} \\theta_ {\\tau , 0} = \\theta \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\tau}(\\theta_{\\tau,k}; D_{\\tau}^{\\mathrm{tr}})\\) is the empirical loss for task \\(\\tau\\) and parameters \\(\\theta_{\\tau,k}\\). Setting \\(\\mathbf{P} = \\mathbf{I}\\) recovers Eq. (3) to the basic gradient descent (GD). Choice of \\(\\mathbf{P}\\) for exploiting the second-order information includes the inverse Fisher information matrix \\(\\mathbf{F}^{-1}\\) which leads to the natural gradient descent (NGD) [4]; the inverse Hessian \\(\\mathbf{H}^{-1}\\) which corresponds to the Newton's method [31]; and the diagonal matrix estimation with the past gradients which results in the adaptive gradient methods [18, 28]. They often reduce the effect of pathological curvature and speed up the optimization [5]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.869, + 0.107 + ], + "angle": 0, + "content": "2.3. Unfolding: reshaping a tensor into a matrix" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.894, + 0.22 + ], + "angle": 0, + "content": "In this study, the concept of unfolding is used to transform the gradient tensor of convolutional kernels into a matrix form. Tensor unfolding, also known as matricization or flattening, is the process of reshaping the elements of an \\(N\\)-dimensional tensor \\(\\mathbf{X} \\in \\mathbb{R}^{I_1 \\times \\dots \\times I_N}\\) into a matrix [29]. The mode-\\(n\\) unfolding of an \\(N\\)-dimensional tensor \\(\\mathbf{X} \\in \\mathbb{R}^{I_1 \\times \\dots \\times I_N}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.226, + 0.892, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} \\xrightarrow [ \\text {m o d e - n u n f o l d i n g} ]{} \\mathbf {X} _ {[ n ]} \\in \\mathbb {R} ^ {I _ {n} \\times I _ {M}}, \\text {w h e r e} I _ {M} = \\prod_ {k \\neq n} I _ {k} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.894, + 0.358 + ], + "angle": 0, + "content": "For example, the weight tensor of a convolutional layer is represented as a 4-D tensor \\((\\mathbf{W} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times C_{\\mathrm{in}} \\times k_h \\times k_w})\\), where it is composed of kernels and it can be unfolded into a matrix as one of the following four forms: (1) \\(\\mathbf{W}_{[1]} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times (C_{\\mathrm{in}}k_hk_w)}\\), (2) \\(\\mathbf{W}_{[2]} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times (C_{\\mathrm{out}}k_hk_w)}\\), (3) \\(\\mathbf{W}_{[3]} \\in \\mathbb{R}^{k_h \\times (C_{\\mathrm{out}}C_{\\mathrm{in}}k_w)}\\), (4) \\(\\mathbf{W}_{[4]} \\in \\mathbb{R}^{k_w \\times (C_{\\mathrm{out}}C_{\\mathrm{in}}k_h)}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.365, + 0.704, + 0.379 + ], + "angle": 0, + "content": "2.4. Riemannian manifold" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.585 + ], + "angle": 0, + "content": "An \\(n\\)-dimensional Riemannian manifold is defined by a manifold \\(\\mathcal{M}\\) and a Riemannian metric \\(g: \\mathcal{M} \\to \\mathbb{R}^{n \\times n}\\), which is a smooth function from each point \\(\\pmb{x} \\in \\mathcal{M}\\) to a positive definite matrix [32]. The metric \\(g(\\pmb{x})\\) defines the inner product of two tangent vectors for each point of the manifold \\(\\langle \\cdot, \\cdot \\rangle: \\mathcal{T}_{\\pmb{x}}\\mathcal{M} \\times \\mathcal{T}_{\\pmb{x}}\\mathcal{M} \\to \\mathbb{R}\\), where \\(\\mathcal{T}_{\\pmb{x}}\\mathcal{M}\\) is the tangent space of \\(\\pmb{x}\\). For \\(\\pmb{v}, \\pmb{w} \\in \\mathcal{T}_{\\pmb{x}}\\mathcal{M}\\), the inner product can be expressed \\(\\langle \\pmb{v}, \\pmb{w} \\rangle = \\pmb{v}^T g(\\pmb{x})\\pmb{w}\\). A Riemannian manifold can be characterized by the curvature of the curves defined by a metric. The curvature of a Riemannian manifold can be computed at each point of the curves, while some manifolds have curvatures of a constant value. For example, the unit sphere \\(\\mathcal{S}\\) has constant positive curvature of \\(+1\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.597, + 0.634, + 0.615 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.894, + 0.743 + ], + "angle": 0, + "content": "In this section, we propose a new preconditioned gradient descent method called GAP in the MAML framework. In Section 3.1, we introduce GAP in the inner-loop optimization and describe how to meta-train GAP in the outer-loop optimization. In Section 3.2, we prove that GAP has two desirable properties. In Section 3.3, we provide a low-computational approximation that can be useful for large-scale architectures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.855, + 0.768 + ], + "angle": 0, + "content": "3.1. GAP: Geometry-Adaptive Preconditioner" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.775, + 0.72, + 0.79 + ], + "angle": 0, + "content": "3.1.1 Inner-loop optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.892, + 0.861 + ], + "angle": 0, + "content": "We consider an \\(L\\)-layer neural network \\(f_{\\theta}(\\cdot)\\) with parameters \\(\\theta = \\{\\mathbf{W}^1, \\dots, \\mathbf{W}^l, \\dots, \\mathbf{W}^L\\}\\). In the standard MAML with task \\(\\tau \\sim p(\\mathcal{T})\\), each \\(\\mathbf{W}^l\\) is adapted with the gradient update as below:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.867, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {\\tau , K} ^ {l} \\leftarrow \\mathbf {W} _ {\\tau , 0} ^ {l} - \\alpha \\cdot \\sum_ {k = 0} ^ {K - 1} \\mathbf {G} _ {\\tau , k} ^ {l} s. t \\mathbf {W} _ {\\tau , 0} ^ {l} = \\mathbf {W} ^ {l}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16082" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.078, + 0.091, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Algorithm 1 Geometry-Adaptive Preconditioned gradient descent (GAP)" + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.125, + 0.472, + 0.378 + ], + "angle": 0, + "content": "Require: \\(p(\\mathcal{T})\\) : distribution over tasks \nRequire: \\(\\alpha ,\\beta_{1},\\beta_{2}\\) : learning rates 1: Randomly initialize \\(\\theta = \\{\\mathbf{W}^1,\\dots ,\\mathbf{W}^L\\}\\) \n2: Initialize \\(\\phi = \\{\\mathbf{M}^{1},\\dots ,\\mathbf{M}^{L}\\}\\) as identity matrix. \n3: while not converged do \n4: Sample a batch of tasks \\(\\mathcal{T}_i\\sim p(\\mathcal{T})\\) \n5: for all \\(\\tau \\in \\mathcal{T}_i\\) do \n6: for inner-loop step \\(k = 0\\) to \\(K - 1\\) do \n7: for layer \\(l = 1\\) to \\(L\\) do \n8: Compute \\(\\mathbf{G}_{\\tau ,k}^{l} = \\nabla_{\\mathbf{W}_{\\tau ,k}^{l}}\\mathcal{L}_{\\tau}^{\\mathrm{in}}(\\theta_{\\tau ,k};D_{\\tau}^{\\mathrm{tr}})\\) using \\(D_{\\tau}^{\\mathrm{tr}}\\) \n9: Reshape \\(\\mathbf{G}_{\\tau ,k}^{l}\\) to \\(\\mathbf{G}_{\\tau ,k}^{l}\\) via Eq. (6) \n10: Transform \\(\\mathbf{G}_{\\tau ,k}^{l}\\) to \\(\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}\\) using \\(\\mathbf{M}^l\\) via Eq. (8) \n11: Reshape \\(\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}\\) back to the original form of gradient tensor, \\(\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}\\) \n12: Compute \\(l\\) -layer adapted weight: \\(\\mathbf{W}_{\\tau ,k + 1}^{l} = \\mathbf{W}_{\\tau ,k}^{l} - \\alpha \\cdot \\tilde{\\mathbf{G}}_{\\tau ,k}^{l}\\) \n13: end for \n14: end for \n15: Compute \\(\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})\\) by evaluating \\(\\mathcal{L}_{\\tau}^{\\mathrm{out}}\\) w.r.t \\(D_{\\tau}^{\\mathrm{val}}\\) \n16: end for \n17: Update the weights and meta parameters: \n18: \\(\\theta \\gets \\theta -\\beta_1\\nabla_\\theta \\sum_{\\tau \\in \\mathcal{T}}\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})\\) \n19: \\(\\phi \\leftarrow \\phi -\\beta_{2}\\nabla_{\\phi}\\sum_{\\tau \\in \\mathcal{T}}\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})\\) \n20: end while" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.473, + 0.504 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}_{\\tau,k}^{l} = \\nabla_{\\mathbf{W}_{\\tau,k}^{l}} \\mathcal{L}_{\\tau}^{\\mathrm{in}}(\\theta_{\\tau,k}; D_{\\tau}^{\\mathrm{tr}})\\) is the gradient with respect to \\(\\mathbf{W}_{\\tau,k}^{l}\\) and \\(\\alpha\\) is the learning rate for inner-loop optimization. In the GAP, we first use the mode-1 unfolding to reshape the gradient tensor into a matrix form (see Section 2.3). For a convolutional layer (i.e., \\(\\mathbf{G}_{\\tau,k}^{l} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times C_{\\mathrm{in}} \\times k \\times k}\\)), we reshape the gradient tensor as below:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.517, + 0.47, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {\\tau , k} ^ {l} \\xrightarrow [ \\text {m o d e - 1} ]{} \\left\\{ \\begin{array}{l l} \\mathbf {G} _ {\\tau , k} ^ {l} \\in \\mathbb {R} ^ {C _ {\\mathrm {o u t}} \\times C _ {\\mathrm {i n}} k ^ {2}} & \\text {i f} C _ {\\mathrm {o u t}} \\leq C _ {\\mathrm {i n}} k ^ {2} \\\\ \\mathbf {G} _ {\\tau , k} ^ {l} \\in \\mathbb {R} ^ {C _ {\\mathrm {i n}} k ^ {2} \\times C _ {\\mathrm {o u t}}} & \\text {i f} C _ {\\mathrm {i n}} k ^ {2} < C _ {\\mathrm {o u t}}, \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.473, + 0.691 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}_{\\tau ,k}^{l}\\) denotes \\((\\mathbf{G}_{\\tau ,k}^{l})_{[1]}\\) for the notational brevity. Note that we chose mode-1 unfolding because it performs best among the four unfolding forms as shown in Table 1. For a linear layer in a matrix form, there is no need for an unfolding. Second, we transform the singular values of the gradient matrix using additional meta parameters \\(\\phi = \\{\\mathbf{M}^l\\}_{l = 1}^L\\) . The meta parameter \\(\\mathbf{M}^l\\) are diagonal matrices with positive elements defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.703, + 0.469, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} ^ {l} = \\left\\{ \\begin{array}{l l} \\operatorname {d i a g} \\left(\\operatorname {S p} \\left(m _ {1} ^ {l}\\right), \\dots , \\operatorname {S p} \\left(m _ {C _ {\\text {o u t}}} ^ {l}\\right)\\right) & \\text {i f} C _ {\\text {o u t}} \\leq C _ {\\text {i n}} k ^ {2} \\\\ \\operatorname {d i a g} \\left(\\operatorname {S p} \\left(m _ {1} ^ {l}\\right), \\dots , \\operatorname {S p} \\left(m _ {C _ {\\text {i n}} k ^ {2}} ^ {l}\\right)\\right) & \\text {i f} C _ {\\text {i n}} k ^ {2} < C _ {\\text {o u t}} \\end{array} , \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.781 + ], + "angle": 0, + "content": "where \\(m_i^l\\in \\mathbb{R}\\) and \\(\\mathbf{Sp}(x) = \\frac{1}{2}\\cdot \\log (1 + \\exp (2*x))\\) . They are applied to the gradient matrix as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.794, + 0.47, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {G}} _ {\\tau , k} ^ {l} = \\mathbf {U} _ {\\tau , k} ^ {l} \\left(\\mathbf {M} ^ {l} \\cdot \\boldsymbol {\\Sigma} _ {\\tau , k} ^ {l}\\right) \\mathbf {V} _ {\\tau , k} ^ {l}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.828, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}_{\\tau,k}^{l} = \\mathbf{U}_{\\tau,k}^{l}\\pmb{\\Sigma}_{\\tau,k}^{l}\\mathbf{V}_{\\tau,k}^{l}\\) is the singular value decomposition (SVD) of \\(\\mathbf{G}_{\\tau,k}^{l}\\). Finally, we reshape \\(\\tilde{\\mathbf{G}}_{\\tau,k}^{l}\\) back to its original gradient tensor form \\(\\tilde{\\mathbf{G}}_{\\tau,k}^{l}\\) (i.e., inverse unfolding). The resulting preconditioned gradient descent of GAP" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.09, + 0.878, + 0.128 + ], + "angle": 0, + "content": "
mode-1mode-2mode-3mode-4
54.86 ± 0.8553.02 ± 0.8751.23 ± 0.7651.45 ± 0.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.132, + 0.894, + 0.173 + ], + "angle": 0, + "content": "Table 1. Performance comparison of unfolding types. We performed the experiment with 5-way 1-shot on mini-[ImageNet] and used the standard Conv-4 backbone." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.19, + 0.659, + 0.206 + ], + "angle": 0, + "content": "becomes the following:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.215, + 0.892, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {\\tau , K} ^ {l} \\leftarrow \\mathbf {W} _ {\\tau , 0} ^ {l} - \\alpha \\cdot \\sum_ {k = 0} ^ {K - 1} \\tilde {\\mathbf {G}} _ {\\tau , k} ^ {l} s. t \\mathbf {W} _ {\\tau , 0} ^ {l} = \\mathbf {W} ^ {l}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.264, + 0.892, + 0.299 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}\\) is the preconditioned gradient based on the meta parameters \\(\\phi\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.724, + 0.333 + ], + "angle": 0, + "content": "3.1.2 Outer-loop optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.341, + 0.892, + 0.386 + ], + "angle": 0, + "content": "For outer-loop optimization, GAP follows the typical process of MAML. Unlike MAML, however, GAP meta-learns two meta parameter sets \\(\\theta\\) and \\(\\phi\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.398, + 0.891, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\theta \\leftarrow \\theta - \\beta_ {1} \\nabla_ {\\theta} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.419, + 0.891, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\phi \\leftarrow \\phi - \\beta_ {2} \\nabla_ {\\phi} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.894, + 0.494 + ], + "angle": 0, + "content": "where \\(\\beta_{1}\\) and \\(\\beta_{2}\\) are the learning rates for the outer-loop optimization. We initialize \\(\\mathbf{M}^l\\) as an identity matrix for all \\(l\\). The training procedure is provided in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.503, + 0.756, + 0.519 + ], + "angle": 0, + "content": "3.2. Desirable properties of GAP" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.557 + ], + "angle": 0, + "content": "In this section, we prove that GAP's preconditioner \\(\\mathbf{P}_{\\mathrm{GAP}}\\) satisfies two desirable properties." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.566, + 0.894, + 0.645 + ], + "angle": 0, + "content": "Theorem 1. Let \\(\\tilde{\\mathbf{G}}_{\\tau,k}^{l} \\in \\mathbb{R}^{m \\times n}\\) be the 'l-layer \\(k\\)-th inner-step' gradient matrix transformed by meta parameter \\(\\mathbf{M}^l\\) for task \\(\\tau\\). Then preconditioner \\(\\mathbf{P}_{GAP}\\) induced by \\(\\tilde{\\mathbf{G}}_{\\tau,k}^{l}\\) is a Riemannian metric and depends on the task-specific parameters \\(\\theta_{\\tau,k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.651, + 0.894, + 0.696 + ], + "angle": 0, + "content": "The proof and closed form of \\(\\mathbf{P}_{\\mathrm{GAP}}\\) are provided in Supplementary Section B. The following two properties emerge from the theorem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.895, + 0.845 + ], + "angle": 0, + "content": "Property 1. Dependency on task-specific parameters: Theorem 1 formally shows that \\(\\mathbf{P}_{\\mathrm{GAP}}\\) depends on the task-specific parameters \\(\\theta_{\\tau,k}\\). While the previous studies considered a non-adaptive preconditioner \\(\\mathbf{P}(\\phi)\\) [34, 36, 44, 49, 57, 66] and a partially adaptive preconditioner \\(\\mathbf{P}(k;\\phi)\\) [49] or \\(\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}};\\phi)\\) [53], \\(\\mathbf{P}_{\\mathrm{GAP}}\\) can be considered to be the most advanced adaptive preconditioner because it is fully adaptive (i.e., task-specific and path-dependent) by being dependent on \\(\\theta_{\\tau,k}\\) as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Property 2. Riemannian metric: If the parameter space has a certain underlying structure, the ordinary gradient of a function \\(\\nabla \\mathcal{L}\\) does not represent its steepest direction [4]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16083" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "To define the steepest direction on the parameter space, we need a Riemannian metric \\( g(\\boldsymbol{w}) \\), which is a positive-definite matrix defined for each parameter \\( \\boldsymbol{w} \\). A Riemannian metric defines the steepest descent direction by \\( -g(\\boldsymbol{w})^{-1}\\nabla \\mathcal{L} \\) [4]. If a preconditioning matrix is a Riemannian metric, it defines the geometry of the underlying structure and enables steepest descent learning. Because we prove that \\( \\mathbf{P}_{\\mathrm{GAP}} \\) is a Riemannian metric for each parameter in Theorem 1, \\( \\mathbf{P}_{\\mathrm{GAP}} \\) is theoretically guaranteed to enable steepest descent learning on its corresponding parameter space. \\( \\mathbf{P}_{\\mathrm{GAP}} \\) consists of two factors, a unitary matrix of the inner-loop gradient \\( \\mathbf{U}_{\\tau,k} \\) and a meta-parameter \\( \\mathbf{M} \\); \\( \\mathbf{M} \\) enables us to reflect the shared geometry information across the tasks. Task-specific and path-dependent geometry information can be reflected in the metric through \\( \\mathbf{U}_{\\tau,k} \\). Two factors allow our Riemannian metric to have higher function complexity than a constant metric. For example, we can consider various structures other than a unit-sphere that corresponds to the constant metric of \\( +1 \\). Even though \\( \\mathbf{P}_{\\mathrm{GAP}} \\) is guaranteed to be a Riemannian metric, it is crucial that the meta-learned \\( \\mathbf{P}_{\\mathrm{GAP}} \\) corresponds to the true parameter space or at least \\( \\mathbf{P}_{\\mathrm{GAP}} \\) is close enough to be useful. We will discuss this issue in Section 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.47, + 0.466 + ], + "angle": 0, + "content": "3.3. Approximate GAP: a low-computational approximation of GAP" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.473, + 0.61 + ], + "angle": 0, + "content": "As presented in Section 3.1, GAP uses an SVD operation. The SVD operation can be burdensome for large-scale networks because it implies that the computational cost can be significantly increased. In recent studies, the use of a large-scale architecture has been emphasized as the key factor for improving the performance of meta-learning [25]. To make use of GAP for large-scale architectures without causing a computational problem, we provide an efficient approximation, named Approximate GAP, under an assumption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.65 + ], + "angle": 0, + "content": "Assumption 1. The elements of the gradient matrix follow an i.i.d. normal distribution with zero means." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.766 + ], + "angle": 0, + "content": "Following [39, 60], we adopt the assumption to have the gradient matrix become orthogonal as \\( n \\) increases. Although the utilization of the assumption is a limiting factor, we empirically confirmed that the row vectors of the gradient matrix are indeed asymptotically orthogonal as \\( n \\) increases (see Figure 3). For the assumption, our approximation can be established as the following." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.47, + 0.838 + ], + "angle": 0, + "content": "Theorem 2. Let \\(\\mathbf{G} \\in \\mathbb{R}^{m \\times n}\\) be a gradient matrix and \\(\\tilde{\\mathbf{G}}\\) be the gradient transformed by a meta parameter \\(\\mathbf{M}\\). Under the Assumption 1, as \\(n\\) becomes large, \\(\\tilde{\\mathbf{G}}\\) asymptotically becomes equivalent to \\(\\mathbf{MG}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.848, + 0.469, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {G}} \\cong \\mathbf {M G} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Note that we have chosen the larger dimension of the gradient matrix as \\( n \\) when reshaping with Eq. (6). We provide" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.092, + 0.817, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.236, + 0.895, + 0.264 + ], + "angle": 0, + "content": "Figure 3. The average of cosine similarity between row vectors of gradient matrix as \\( n \\) increases." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.277, + 0.875, + 0.329 + ], + "angle": 0, + "content": "
Algorithmtrain time (secs)test time (secs)GPU-memory (MB)
MAML35796.6152.6610185
GAP50916.63120.4410279
Approximate GAP36684.7253.6210217
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.332, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Table 2. Comparison of training time, GPU memory, and test time for MAML, GAP, and Approximate GAP. We performed the experiment with 5-way 1-shot on mini-ImageNet and used 600 tasks in the test. We used the standard Conv-4 backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.537 + ], + "angle": 0, + "content": "the proof in Supplementary Section B. This approximation has a clear trade-off between scalability and adaptiveness. Approximate GAP efficiently reduces the computational cost as shown in Table 2, whereas the preconditioner \\(\\mathbf{P}_{\\mathrm{GAP}}\\) becomes not adaptive but constant. However, Approximate GAP still guarantees the preconditioner to be a Riemannian metric. As shown in Table 4 & 5, Approximate GAP incurs a slight performance drop but it still achieves a high performance owing to the Riemannian metric property." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.549, + 0.634, + 0.566 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.574, + 0.895, + 0.636 + ], + "angle": 0, + "content": "In this section, we show the superiority of GAP by comparing it with the state-of-the-art PGD-MAML family and the MAML family. Hyper-parameter setups used in our experiments can be found in Supplementary Section C.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.644, + 0.691, + 0.66 + ], + "angle": 0, + "content": "4.1. Few-shot regression" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.894, + 0.865 + ], + "angle": 0, + "content": "Datasets and experimental setup. The goal of few-shot regression is to fit an unknown target function for the given \\( K \\) sample points from the function. For the evaluation of few-shot regression, we use the sinusoid regression benchmark [20]. In this benchmark, sinusoid is used as the target function. Each task has a sinusoid \\( y(x) = A\\sin (\\omega x + b) \\) as the target function, where the parameter values are within the following range: amplitude \\( A\\in [0.1,5.0] \\), frequency \\( \\omega \\in [0.8,1.2] \\), and phase \\( b\\in [0,\\pi ] \\). For each task, input data point \\( x \\) is sampled from \\( [-5.0,5.0] \\). In the experiment, we use a simple Multi-Layer Perceptron (MLP), following the setting in [20]. The details of the architecture are provided in Supplementary Section C.2.1." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Results. We evaluate GAP and compare it with MAML family and PGD-MAML family on a regression task. As" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16084" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.449, + 0.197 + ], + "angle": 0, + "content": "
Algorithm5-shot10-shot20-shot
MAML [20]1.13 ± 0.180.77 ± 0.110.48 ± 0.08
Meta-SGD† [36]0.90 ± 0.160.53 ± 0.090.31 ± 0.05
MT-Net [34]0.76 ± 0.090.49 ± 0.050.33 ± 0.04
ALFA [9]0.92 ± 0.190.62 ± 0.160.34 ± 0.07
L2F [10]0.71 ± N/A0.37 ± N/A0.16 ± N/A
PAMELA† [49]0.54 ± 0.060.41 ± 0.040.17 ± 0.03
MeTAL [8]0.74 ± 0.180.44 ± 0.110.21 ± 0.06
GAP†0.16 ± 0.040.04 ± 0.010.01 ± 0.01
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.256 + ], + "angle": 0, + "content": "Table 3. Few-shot regression for the sinusoid regression benchmark with a 2-layer MLP backbone. We report MSE ± 95% confidence intervals(ci) for 600 tasks following the setup in [20].† denotes PGD-MAML family." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.472, + 0.346 + ], + "angle": 0, + "content": "shown in Table 3, GAP consistently achieves the lowest mean squared error (MSE) scores, with the lowest confidence intervals in all three cases. The performance of GAP is improved by \\(89\\%\\) on 10-shot and \\(94\\%\\) on 20-shot compared to the performance of state-of-the-art algorithms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.285, + 0.37 + ], + "angle": 0, + "content": "4.2. Few-shot classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.474, + 0.606 + ], + "angle": 0, + "content": "Datasets and experimental setup. For the few-shot classification, we evaluate two benchmarks: (1) miniImageNet [56]; this dataset has 100 classes and it is a subset of ImageNet [16], and we use the same split as in [51], with 64, 16 and 20 classes for train, validation and test, respectively. (2) tiered-ImageNet [52]; this is also a subset of ImageNet with 608 classes grouped into 34 high-level categories, and divided into 20, 6 and 8 for train, validation, and test, respectively. For all the experiments, our model follows the standard Conv-4 backbone used in [56]. The details of the architecture are provided in Supplementary Section C.2.2. Following the experimental protocol in [20], we use 15 samples per class in the query-set to compute the meta gradients. In meta training and meta testing, the inner-loop optimization is updated in five and ten steps, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.474, + 0.902 + ], + "angle": 0, + "content": "Results. Table 4 & 5 present the performance of GAP, the state-of-the-art PGD-MAML family, and the state-of-the-art MAML-family on mini-[ImageNet and tiered-[ImageNet under two typical settings: 5-way 1-shot and 5-way 5-shot. The GAP outperforms all of the previous PGD-MAML family and MAML family. Compared to the state-of-the-art MAML family, GAP improves the performance with a quite significant margin for both mini-[ImageNet and tiered-[ImageNet datasets. Compared to the state-of-the-art PGD-MAML family, GAP shows that the 1- and 5-shot accuracy can be increased by \\(1.4\\%\\) and \\(1.5\\%\\) on mini-[ImageNet dataset, and by \\(0.7\\%\\) and \\(0.68\\%\\) on tiered-[ImageNet dataset, respectively. We also evaluated Approximate GAP that is introduced in Section 3.3. The results show that the approximated version can perform comparably to the original GAP. Although Approximate GAP shows slightly lower accuracies than the original, its performance is superior to most of the existing algorithms because of its Riemannian metric property." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.088, + 0.872, + 0.535 + ], + "angle": 0, + "content": "
Algorithm5-way 1-shot5-way 5-shot
MAML [20]47.89 ± 1.2064.59 ± 0.88
Meta-SGD† [36]50.47 ± 1.8764.00 ± 0.90
BMAML [64]53.80 ± 1.4664.23 ± 0.69
ANIL [48]46.70 ± 0.4061.50 ± 0.50
LLAMA [23].49.40 ± 1.83N/A
PLATIPUS [21]50.13 ± 1.86-
T-net [34]50.86 ± 1.82N/A
MT-net [34]51.70 ± 1.84N/A
MAML++ [7]52.15 ± 0.2668.32 ± 0.44
iMAML-HF [50]49.30 ± 1.88N/A
WarpGrad [22]52.30 ± 0.9068.40 ± 0.60
MC1† [44]53.74 ± 0.8468.01 ± 0.73
MC2† [44]54.08 ± 0.8867.99 ± 0.73
MH-C† [66]48.64 ± 0.3364.52 ± 0.51
MH† [66]49.41 ± 0.9667.16 ± 0.42
BOIL [43]49.61 ± 0.1666.46 ± 0.37
ARML [63]50.42 ± 1.7964.12 ± 0.90
ALFA [9]50.58 ± 0.5169.12 ± 0.47
L2F [10]52.10 ± 0.5069.38 ± 0.46
ModGrad† [53]53.20 ± 0.8669.17 ± 0.69
PAMELA† [49]53.50 ± 0.8970.51 ± 0.67
SignMAML [19]42.90 ± 1.5060.70 ± 0.70
CTML [45]50.47 ± 1.8364.15 ± 0.90
MeTAL [8]52.63 ± 0.3770.52 ± 0.29
ECML [24]48.94 ± 0.8065.26 ± 0.67
Sharp-MAML_up [1]49.56 ± N/A63.06 ± N/A
Sharp-MAML_low [1]49.72 ± N/A63.18 ± N/A
Sharp-MAMLboth [1]50.28 ± N/A65.04 ± N/A
FBM [62]50.62 ± 1.7964.78 ± 0.35
CxGrad [33]51.80 ± 0.4669.82 ± 0.42
HyperMAML [47]51.84 ± 0.5766.29 ± 0.43
EEML [35]52.42 ± 1.7568.40 ± 0.95
MH-O† [66]52.50 ± 0.6167.76 ± 0.34
Sparse-MAML† [57]50.35 ± 0.3967.03 ± 0.74
Sparse-ReLU-MAML† [57]50.39 ± 0.8964.84 ± 0.46
Sparse-MAML+† [57]51.04 ± 0.5967.03 ± 0.74
Approximate GAP†53.52 ± 0.8870.75 ± 0.67
GAP†54.86 ± 0.8571.55 ± 0.61
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.538, + 0.895, + 0.594 + ], + "angle": 0, + "content": "Table 4. 5-way few-shot classification accuracy \\((\\%)\\) on miniImageNet with a Conv-4 backbone. We report mean \\(\\pm 95\\%\\) confidence intervals(ci) for 600 tasks according to [20]. \\(\\dagger\\) denotes PGD-MAML family." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.597, + 0.872, + 0.813 + ], + "angle": 0, + "content": "
Algorithm5-way 1-shot5-way 5-shot
Meta-SGD†[36]50.92 ± 0.9369.28 ± 0.80
MAML [20]51.70 ± 1.8070.30 ± 1.80
MT-net [34]51.95 ± 1.83N/A
WarpGrad [22]57.20 ± 0.9074.10 ± 0.70
BOIL [43].48.58 ± 0.2769.37 ± 0.12
ALFA [9]53.16 ± 0.4970.54 ± 0.46
L2F [10]54.40 ± 0.5073.34 ± 0.44
ARML [63]52.91 ± 1.83N/A
PAMELA†[49]54.81 ± 0.8874.39 ± 0.71
Sparse-ReLU-MAML†[57]53.18 ± 0.5269.06 ± 0.28
Sparse-MAML†[57]53.47 ± 0.5368.83 ± 0.65
Sparse-MAML+†[57]53.91 ± 0.6769.92 ± 0.21
MeTAL [8]54.34 ± 0.3170.40 ± 0.21
CxGrad [33]55.55 ± 0.4673.55 ± 0.41
ECML [24]47.34 ± 0.8864.77 ± 0.75
Approximate GAP†56.86 ± 0.9174.41 ± 0.72
GAP†57.60 ± 0.9374.90 ± 0.68
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.815, + 0.895, + 0.872 + ], + "angle": 0, + "content": "Table 5. 5-way few-shot classification accuracy (%) on tiered-ImageNet dataset with a Conv-4 backbone. We report mean ± 95% confidence intervals(ci) for 600 tasks according to [20].† denotes PGD-MAML family." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16085" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.833, + 0.3 + ], + "angle": 0, + "content": "
Algorithmtiered-ImageNetCUBCars
1-shot5-shot1-shot5-shot1-shot5-shot
MAML [20]51.61 ± 0.2065.76 ± 0.2740.51 ± 0.0853.09 ± 0.1633.57 ± 0.1444.56 ± 0.21
ANIL [48]52.82 ± 0.2966.52 ± 0.2841.12 ± 0.1555.82 ± 0.2134.77 ± 0.3146.55 ± 0.29
BOIL [43]53.23 ± 0.4169.37 ± 0.2344.20 ± 0.1560.92 ± 0.1136.12 ± 0.2950.64 ± 0.22
BMAML [64]N/AN/A33.52 ± 0.3651.35 ± 0.16N/AN/A
ALFA [9]N/AN/AN/A58.35 ± 0.25N/AN/A
L2F [10]N/AN/AN/A60.89 ± 0.22N/AN/A
MeTAL [8]N/AN/AN/A58.20 ± 0.24N/AN/A
HyperMAML [47]N/AN/A36.52 ± 0.6149.43 ± 0.14N/AN/A
CxGrad [33]N/AN/AN/A63.92 ± 0.44N/AN/A
Sparse-MAML†[57]53.47 ± 0.5368.83 ± 0.6541.37 ± 0.7360.58 ± 1.1035.90 ± 0.5052.63 ± 0.56
Sparse-ReLU-MAML†[57]53.77 ± 0.9468.12 ± 0.6942.89 ± 0.4557.53 ± 0.9436.04 ± 0.5549.95 ± 0.42
Sparse-MAML+†[57]53.91 ± 0.6769.92 ± 0.2143.43 ± 1.0462.02 ± 0.7837.14 ± 0.7753.18 ± 0.44
Approximate GAP†57.47 ± 0.9971.66 ± 0.7643.77 ± 0.7962.92 ± 0.7337.00 ± 0.7553.28 ± 0.76
GAP†58.56 ± 0.9372.82 ± 0.7744.74 ± 0.7564.88 ± 0.7238.44 ± 0.7755.04 ± 0.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.303, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Table 6. 5-way few-shot cross domain classification accuracy (%) with a Conv-4 backbone, meta training on mini-[ImageNet dataset, and meta-testing on tiered-[ImageNet, CUB, or Cars datasets. We report mean ± 95% confidence intervals(ci) for 600 tasks according to [20].† denotes PGD-MAML family." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.372, + 0.395, + 0.387 + ], + "angle": 0, + "content": "4.3. Cross-domain few-shot classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.475, + 0.593 + ], + "angle": 0, + "content": "The cross-domain few-shot classification introduced by [14] addresses a more challenging and practical few-shot classification scenario in which meta-train and meta-test tasks are sampled from different task distributions. These scenarios are designed to evaluate meta-level overfitting of meta-learning algorithms by creating a large domain gap between meta-trains and meta-tests. In particular, an algorithm can be said to be meta-overfitting if it relies too much on the prior knowledge of previously seen meta-train tasks instead of focusing on a few given examples to learn a new task. This meta-level overfitting makes the learning system more likely to fail to adapt to new tasks sampled from substantially different task distributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.473, + 0.848 + ], + "angle": 0, + "content": "Datasets and experimental setup. To evaluate the level of meta-overfitting for GAP, we evaluate a cross-domain few-shot classification experiment. The mini-ImageNet is used for the meta-train task, and the tiered-ImageNet [52], CUB-200-2011 [58], Cars [11] datasets are used for the meta-test task. The CUB dataset has 200 fine-grained classes and consists of a total of 11,788 images; it is further divided into 100 meta-train classes, 50 meta-validation classes, and 50 meta-test classes. The Cars [30] dataset consists of 16,185 images of 196 classes of cars; it is split into 8,144 training images and 8,041 testing images, where each class has been split roughly in 50-50. The classes are typically at the level of Make, Model, Year, e.g., 2012 Tesla Model S or 2012 BMW M3 coupe. As with the few-shot classification experiment, we use the standard Conv-4 backbone and follow the same experimental protocol." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Results. Table 6 presents the cross-domain few-shot performance for GAP, MAML family, and PGD-MAML family. GAP significantly outperforms the state-of-the-art al" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.898, + 0.614 + ], + "angle": 0, + "content": "gorithms on 5-way 1-shot and 5-way 5-shot cross-domain classification tasks. In particular, for the tiered-ImageNet dataset, the performance was improved by \\(8.6\\%\\) and \\(4.1\\%\\) on 1-shot and 5-shot classification tasks, respectively. Because GAP can simultaneously consider a task's individuality and optimization trajectory in the inner-loop optimization, it can overcome meta-overfitting better than the existing methods. However, Approximate GAP shows more performance degradation in cross-domain few-shot classification than in few-shot classification. In particular, when the domain difference with the meta-train is more significant (i.e., the tiered-ImageNet dataset) than when the domain difference with the meta-train is marginal (i.e., CARS and CUB datasets), it shows a more considerable performance drop. We can see that full adaptation plays an important role in cross-domain few-shot classification." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.63, + 0.615, + 0.645 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.895, + 0.86 + ], + "angle": 0, + "content": "Number of meta parameters: Recent MAML family and PGD-MAML family require a large increase in the number of meta-learning parameters as shown in Table 7. One advantage of GAP is that it requires only a very small increase in the number of meta parameters, when compared to the baseline of MAML. This is possible because we transform a gradient tensor into a gradient matrix, perform the SVD of the matrix, and assign only a small number of meta parameters that correspond to the diagonal matrix of the gradient matrix. For the Conv-4 network, GAP requires only \\(0.2\\%\\) increase of the meta parameters. Although the increase in the number of meta parameters is negligible, SVD of the gradient matrix can incur a large computational burden for large networks. This is addressed by Approximate GAP." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Approximate GAP vs. simple constant preconditioners: Approximate GAP is a low-complexity method where SVD" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "16086" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.445, + 0.216 + ], + "angle": 0, + "content": "
Algorithm# of params% increase
MAML [20]1.2109 × 105
Meta-SGD [36]2.4218 × 105100.0%
MC [44]2.7106 × 1062140.4%
PAMELA [49]1.6239 × 10534.1%
MH [66]7.2196 × 10759586.7%
Sparse-MAML [57]2.4218 × 105100.0%
GAP1.2131 × 1050.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.221, + 0.47, + 0.248 + ], + "angle": 0, + "content": "Table 7. Comparison of the number of parameters for MAML, existing methods, and GAP." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.263, + 0.446, + 0.316 + ], + "angle": 0, + "content": "
AlgorithmStructureRiemannian metric (i.e., positive definite)Acc. (%)
Meta-SGDdiag(a1, ·, an)X50.47%
Meta-SGD with positive definitenessdiag(a1, ·, an)O52.39%
Approximate GAPblkdiag(Mt, ·, M)O53.52%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.321, + 0.471, + 0.376 + ], + "angle": 0, + "content": "Table 8. Performance comparison of three constant (i.e., non-adaptive) preconditioners for 5-way 1-shot on mini-ImageNet: Meta-SGD, Meta-SGD modified to satisfy positive definiteness, and our Approximate GAP." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.39, + 0.445, + 0.447 + ], + "angle": 0, + "content": "
Algorithm1-shot5-shot
GAP w/o PGAP48.23 ± 0.8065.80 ± 0.75
GAP w/ PGAP54.86 ± 0.8571.55 ± 0.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.451, + 0.468, + 0.492 + ], + "angle": 0, + "content": "Table 9. Ablation study of \\(\\mathbf{P}_{\\mathrm{GAP}}\\) on mini-ImageNet. Performance of the GAP-trained model is significantly affected by not applying \\(\\mathbf{P}_{\\mathrm{GAP}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.786 + ], + "angle": 0, + "content": "operation is avoided by approximating GAP with a constant diagonal preconditioner. A natural question to ask is how does Approximate GAP compare with other constant diagonal preconditioners. To answer this question, we have compared Approximate GAP with Meta-SGD and a modified Meta-SGD. Meta-SGD [36] is a well-known constant diagonal preconditioner (i.e., \\(\\mathrm{diag}(a_1,\\dots ,a_n)\\)) that does not need to be positive definite and we also investigate its modification with a constraint on positive definiteness. The results are shown in Table 8. It can be observed that enforcing positive definiteness can improve Meta-SGD. Furthermore, an additional improvement can be achieved by Approximate GAP. While both modified Meta-SGD and Approximate GAP are positive definite, Approximate GAP is different because it inherits an additional constraint from GAP - a block diagonal structure where a constant diagonal matrix \\(M\\) is repeated (i.e., \\(\\mathrm{blkdiag}(M,\\dots ,M)\\)). The inherited constraint provides a gain over the modified Meta-SGD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Does GAP learn a useful preconditioner: While a Riemannian metric can be helpful, it does not mean any Riemannian metric will result in an improvement. For the true parameter space with a specific underlying structure, the corresponding Riemannian metric needs to be applied to enable steepest descent [3, 4]. For the special case of a two-layer neural network with a mean squared error (MSE) loss, it" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.333 + ], + "angle": 0, + "content": "was proven that Fisher information matrix is the corresponding Riemannian metric [4]. For a general neural network, however, a proper Riemannian metric is unknown and it needs to be learned. In our work, we have devised a method to guarantee a Riemannian metric and have used the outer-loop optimization to learn the Riemannian metric. In general, the learned Riemannian metric is unlikely to correspond perfectly to the true parameter space. Then, an important question is if the Riemannian metric learned by GAP is close enough to the desired one and if it is useful. To investigate this issue, we performed an ablation study by not applying the preconditioner \\(\\mathbf{P}_{\\mathrm{GAP}}\\). After training a GAP model, we have evaluated the performance with and without applying \\(\\mathbf{P}_{\\mathrm{GAP}}\\). The results are shown in Table 9 and clearly the preconditioner learned with the outer-loop optimization plays an essential role for improving the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.895, + 0.599 + ], + "angle": 0, + "content": "Why is preconditioner helpful for meta-learning: When the batch size is small, the resulting empirical gradient can be noisy [53, 65]. A typical few-shot learning has only a small number of samples for the inner-loop optimization, and its gradient can be noisy. On the other hand, it was shown in [5] that preconditioned gradient descent with a positive definite preconditioner can achieve a lower risk than gradient descent when the labels are noisy, the model is mis-specified, or the signal is misaligned with the features. Under a misalignment, a properly chosen positive definite preconditioner can generalize better than gradient descent [5]. Considering the noisy gradient of inner loop optimization, it can be surmised that a positive definite preconditioner that is adaptive (i.e., a Riemannian metric) can be helpful for improving MAML. Note that the noisy case is in contrast to the case of supervised learning with a large amount of data [5]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.618, + 0.637 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.894, + 0.74 + ], + "angle": 0, + "content": "In this work, we have proposed GAP that is a PGDMAML algorithm utilizing SVD operation. Thanks to the fully adaptive property, GAP can handle individual tasks with a wide diversity within the MAML framework. Additionally, GAP can enable steepest descent on the parameter space owing to the Riemannian metric property." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.667, + 0.779 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.894, + 0.888 + ], + "angle": 0, + "content": "This work was supported by ETRI [23ZR1100, A Study of Hyper-Connected Thinking Internet Technology by autonomous connecting, controlling and evolving ways], NRF (NRF-2020R1A2C2007139), IITP [NO.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University)], and the New Faculty Startup Fund from Seoul National University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16087" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning. arXiv preprint arXiv:2206.03996, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] Shunichi Amari. A theory of adaptive pattern classifiers. IEEE Transactions on Electronic Computers, (3):299-307, 1967. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[3] Shun-ichi Amari. Neural learning in structured parameter spaces-natural riemannian gradient. Advances in neural information processing systems, 9, 1996. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.472, + 0.285 + ], + "angle": 0, + "content": "[4] Shun-Ichi Amari. Natural gradient works efficiently in learning. Neural computation, 10(2):251-276, 1998. 1, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[5] Shun-ichi Amari, Jimmy Ba, Roger Grosse, Xuechen Li, Atsushi Nitanda, Taiji Suzuki, Denny Wu, and Ji Xu. When does preconditioning help or hurt generalization? arXiv preprint arXiv:2006.10732, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[6] Shun-Ichi Amari and Scott C Douglas. Why natural gradient? In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), volume 2, pages 1213-1216. IEEE, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[7] Antreas Antoniou, Harrison Edwards, and Amos Storkey. How to train your maml. arXiv preprint arXiv:1810.09502, 2018.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[8] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9465-9474, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.472, + 0.585 + ], + "angle": 0, + "content": "[9] Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. Meta-learning with adaptive hyperparameters. Advances in Neural Information Processing Systems, 33:20755-20765, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.472, + 0.642 + ], + "angle": 0, + "content": "[10] Sungyong Baik, Seokil Hong, and Kyoung Mu Lee. Learning to forget for meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2379-2387, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[11] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.472, + 0.741 + ], + "angle": 0, + "content": "[12] Irénée-Jules Bienaymé. Considerations à l'appui de la découverte de Laplace sur la loi de probabilité dans la méthode des moindres carrés. Imprimerie de Mallet-Bachelier, 1853. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.472, + 0.772 + ], + "angle": 0, + "content": "[13] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[14] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[15] Tristan Deleu, Tobias Würfl, Mandana Samiei, Joseph Paul Cohen, and Yoshua Bengio. Torchmeta: A meta-learning library for pytorch. arXiv preprint arXiv:1909.06576, 2019. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[17] Lin Ding, Peng Liu, Wenfeng Shen, Weijia Lu, and Shengbo Chen. Gradient-based meta-learning using uncertainty to weigh loss for few-shot learning. arXiv preprint arXiv:2208.08135, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[18] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[19] Chen Fan, Parikshit Ram, and Sijia Liu. Sign-maml: Efficient model-agnostic meta-learning by signsgd. arXiv preprint arXiv:2109.07497, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.894, + 0.315 + ], + "angle": 0, + "content": "[20] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 1, 3, 5, 6, 7, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[21] Chelsea Finn, Kelvin Xu, and Sergey Levine. Probabilistic model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[22] Sebastian Flennerhag, Andrei A Rusu, Razvan Pascanu, Francesco Visin, Hujun Yin, and Raia Hadsell. Meta-learning with warped gradient descent. arXiv preprint arXiv:1909.00025, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[23] Erin Grant, Chelsea Finn, Sergey Levine, Trevor Darrell, and Thomas Griffiths. Recasting gradient-based meta-learning as hierarchical bayes. arXiv preprint arXiv:1801.08930, 2018.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[24] Markus Hiller, Mehrtash Harandi, and Tom Drummond. On enforcing better conditioned meta-learning for rapid few-shot adaptation. arXiv preprint arXiv:2206.07260, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[25] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9068-9077, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[27] Sham M Kakade. A natural policy gradient. Advances in neural information processing systems, 14, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[28] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3,14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[29] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[30] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[31] Yann A LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-48. Springer, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.856, + 0.845 + ], + "angle": 0, + "content": "[32] John M Lee. Smooth manifolds. Springer, 2012. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[33] Sanghyuk Lee, Seunghyun Lee, and Byung Cheol Song. Contextual gradient scaling for few-shot learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 834-843, 2022. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16088" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[34] Yoonho Lee and Seungjin Choi. Gradient-based meta-learning with learned layerwise metric and subspace. In International Conference on Machine Learning, pages 2927-2936. PMLR, 2018. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.189 + ], + "angle": 0, + "content": "[35] Geng Li, Boyuan Ren, and Hongzhi Wang. Eeml: Ensemble embedded meta-learning. arXiv preprint arXiv:2206.09195, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[36] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[37] Zequn Liu, Ruiyi Zhang, Yiping Song, and Ming Zhang. When does maml work the best? an empirical study on model-agnostic meta-learning in nlp applications. arXiv preprint arXiv:2005.11700, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.472, + 0.402 + ], + "angle": 0, + "content": "[39] Ahmed M Abdelmoniem, Ahmed Elzanaty, Mohamed-Slim Alouini, and Marco Canini. An efficient statistical-based gradient compression technique for distributed training systems. Proceedings of Machine Learning and Systems, 3:297-322, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.472, + 0.474 + ], + "angle": 0, + "content": "[40] Gabriel Maicas, Andrew P Bradley, Jacinto C Nascimento, Ian Reid, and Gustavo Carneiro. Training medical image analysis systems like radiologists. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 546-554. Springer, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[41] George Marsaglia. Choosing a point from the surface of a sphere. The Annals of Mathematical Statistics, 43(2):645-646, 1972. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[42] Fei Mi, Minlie Huang, Jiyong Zhang, and Boi Faltings. Meta-learning for low-resource natural language generation in task-oriented dialogue systems. arXiv preprint arXiv:1905.05644, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.575, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[43] Jaehoon Oh, Hyungjun Yoo, ChangHwan Kim, and Se-Young Yun. Boil: Towards representation change for few-shot learning. arXiv preprint arXiv:2008.08882, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[44] Eunbyung Park and Junier B Oliva. Meta-curvature. Advances in Neural Information Processing Systems, 32, 2019. 1, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.472, + 0.688 + ], + "angle": 0, + "content": "[45] Danni Peng and Sinno Pan. Clustered task-aware meta-learning by learning from learning paths. 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[46] Juan-Manuel Perez-Rua, Xiatian Zhu, Timothy M Hospedales, and Tao Xiang. Incremental few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13846-13855, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[47] M Przewięźlikowski, P Przybysz, J Tabor, M Zieba, and P Spurek. Hypermaml: Few-shot adaptation of deep models with hypernetworks. arXiv preprint arXiv:2205.15745, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[48] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[49] Jathushan Rajasegaran, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Mubarak Shah. Meta-learning" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the learning trends shared across tasks. arXiv preprint arXiv:2010.09291, 2020. 1, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[50] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[51] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.209, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[52] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[53] Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. On modulating the gradient for meta-learning. In European Conference on Computer Vision, pages 556-572. Springer, 2020. 1, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.336, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[54] Rishav Singh, Vandana Bharti, Vishal Purohit, Abhinav Kumar, Amit Kumar Singh, and Sanjay Kumar Singh. Metamed: Few-shot medical image classification using gradient-based meta-learning. Pattern Recognition, 120:108111, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.393, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[55] Xingyou Song, Yuxiang Yang, Krzysztof Choromanski, Ken Caluwaerts, Wenbo Gao, Chelsea Finn, and Jie Tan. Rapidly adaptable legged robots via evolutionary meta-learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3769-3776. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[56] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[57] Johannes Von Oswald, Dominic Zhao, Seijin Kobayashi, Simon Schug, Massimo Caccia, Nicolas Zucchet, and João Sacramento. Learning where to learn: Gradient sparsity in meta and continual learning. Advances in Neural Information Processing Systems, 34:5250-5263, 2021. 1, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[58] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[59] Shuhuan Wen, Zeteng Wen, Di Zhang, Hong Zhang, and Tao Wang. A multi-robot path-planning algorithm for autonomous navigation using meta-reinforcement learning based on transfer learning. Applied Soft Computing, 110:107605, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.705, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[60] Simon Wiedemann, Temesgen Mehari, Kevin Kepp, and Wojciech Samek. Dithered backprop: A sparse and quantized backpropagation algorithm for more efficient deep neural network training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 720-721, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[61] Xiongwei Wu, Doyen Sahoo, and Steven Hoi. Meta-rcnn: Meta learning for few-shot object detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1679–1687, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[62] Peng Yang, Shaogang Ren, Yang Zhao, and Ping Li. Calibrating cnns for few-shot meta learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2090–2099, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16089" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "[63] Huaxiu Yao, Xian Wu, Zhiqiang Tao, Yaliang Li, Bolin Ding, Ruirui Li, and Zhenhui Li. Automated relational meta-learning. arXiv preprint arXiv:2001.00745, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[64] Jaesik Yoon, Taesup Kim, Ousmane Dia, Sungwoong Kim, Yoshua Bengio, and Sungjin Ahn. Bayesian model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[65] Yikai Zhang, Hui Qu, Chao Chen, and Dimitris Metaxas. Taming the noisy gradient: train deep neural networks with small batch sizes. In The Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI), 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[66] Dominic Zhao, Johannes von Oswald, Seijin Kobayashi, João Sacramento, and Benjamin F Grewe. Meta-learning via hypernetworks. 2020. 1, 4, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16090" + } + ] +] \ No newline at end of file diff --git a/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_origin.pdf b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1105bef208e6fc881ae5dbebce26009d7edb2ffa --- /dev/null +++ b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/79e00a9d-10a7-492d-9f6c-4be9854d55e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57eed10efdb9a26a5921c9a88a502810dd56b6a9fee0da50fddc0defab1e5fa0 +size 1735960 diff --git a/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/full.md b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0ba8720dc72ec5c9edbcbabf5a81287dd6a1207a --- /dev/null +++ b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/full.md @@ -0,0 +1,388 @@ +# Meta-Learning with a Geometry-Adaptive Preconditioner + +Suhyun Kang $^{1*}$ , Duhun Hwang $^{1*}$ , Moonjung Eo $^{1}$ , Taesup Kim $^{2}$ , Wonjong Rhee $^{1,3\dagger}$ + +1 Department of Intelligence and Information, Seoul National University + +2 Graduate School of Data Science, Seoul National University + +$^{3}$ IPAI and AIIS, Seoul National University + +{su_hyun, yelobean, eod87, taesup.kim, wrhee}@snu.ac.kr + +# Abstract + +Model-agnostic meta-learning (MAML) is one of the most successful meta-learning algorithms. It has a bi-level optimization structure where the outer-loop process learns a shared initialization and the inner-loop process optimizes task-specific weights. Although MAML relies on the standard gradient descent in the inner-loop, recent studies have shown that controlling the inner-loop's gradient descent with a meta-learning preconditioner can be beneficial. Existing preconditioners, however, cannot simultaneously adapt in a task-specific and path-dependent way. Additionally, they do not satisfy the Riemannian metric condition, which can enable the steepest descent learning with preconditioned gradient. In this study, we propose Geometry-Adaptive Preconditioned gradient descent (GAP) that can overcome the limitations in MAML; GAP can efficiently meta-learn a preconditioner that is dependent on task-specific parameters, and its preconditioner can be shown to be a Riemannian metric. Thanks to the two properties, the geometry-adaptive preconditioner is effective for improving the inner-loop optimization. Experiment results show that GAP outperforms the state-of-the-art MAML family and preconditioned gradient descent-MAML (PGD-MAML) family in a variety of few-shot learning tasks. Code is available at: https://github.com/Suhyun777/CVPR23-GAP. + +# 1. Introduction + +Meta-learning, or learning to learn, enables algorithms to quickly learn new concepts with only a small number of samples by extracting prior-knowledge known as meta-knowledge from a variety of tasks and by improving the generalization capability over the new tasks. Among the meta-learning algorithms, the category of optimization-based meta-learning [8, 17, 20, 21, 48] has been gaining popularity + +due to its flexible applicability over diverse fields including robotics [55, 59], medical image analysis [40, 54], language modeling [37, 42], and object detection [46, 61]. In particular, Model-Agnostic Meta-Learning (MAML) [20] is one of the most prevalent gradient-based meta-learning algorithms. + +Many recent studies have improved MAML by adopting a Preconditioned Gradient Descent (PGD $^1$ ) for inner-loop optimization [34, 36, 44, 49, 53, 57, 66]. In this paper, we collectively address PGD-based MAML algorithms as the PGD-MAML family. A PGD is different from the ordinary gradient descent because it performs a preconditioning on the gradient using a preconditioning matrix $\mathbf{P}$ , also called a preconditioner. A PGD-MAML algorithm meta-learns not only the initialization parameter $\theta_0$ of the network but also the meta-parameter $\phi$ of the preconditioner $\mathbf{P}$ . + +For the inner-loop optimization, $\mathbf{P}$ was kept static in most of the previous works (Figure 1(b)) [34, 36, 44, 57, 66]. Some of the previous works considered adapting the preconditioner $\mathbf{P}$ with the inner-step $k$ (Figure 1(c)) [49] and some others with the individual task (Figure 1(d)) [53]. They achieved performance improvement by considering individual tasks and inner-step, respectively, and showed that both factors were valuable. However, both factors have not been considered simultaneously in the existing studies. + +When a parameter space has a certain underlying structure, there exists a Riemannian metric corresponding the parameter space [3,4]. If the preconditioning matrix is the Riemannian metric, the preconditioned gradient is known to become the steepest descent on the parameter space [2-4,6,27]. An illustration of a toy example is shown in Figure 2. The optimization path of an ordinary gradient descent is shown in Figure 2(a). Compared to the ordinary gradient descent, a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can actually harm the optimization. For the example in Figure 2(b), the preconditioner affects the optimization into an undesirable direction and negatively affects the gradient descent. On the + +![](images/22f94a546d5669996f508fd7a8de533e9ecf86714cfd0f4f740d739e3e8c590d.jpg) +(a) MAML + +![](images/fc08048b30068de72750ee3c66ee519927bf6ab0c8cf79c359d78bb633857f94.jpg) +(b) Non-adaptive $\mathbf{P}(\phi)$ +Figure 1. Diagram of MAML and PGD-MAML family. For the inner-loop adaptation in each diagram, the dotted lines of the same color indicate that they use a common preconditioning matrix (preconditioner). (a) MAML adaptation: no preconditioner is used (i.e., $\mathbf{P} = \mathbf{I}$ ). (b) $\mathbf{P}(\phi)$ : a constant preconditioner is used in the inner-loop where the preconditioner's meta-parameter $\phi$ is meta-learned. (c) $\mathbf{P}(k; \phi)$ : a constant preconditioner is used for each inner-step $k$ . Preconditioner for each step is meta-learned, but $\mathbf{P}(k, \phi)$ is not task-specific. (d) $\mathbf{P}(D_{\tau}^{\mathrm{tr}}; \phi)$ : a constant preconditioner is used for each task. Preconditioner for each task is meta-learned, but $\mathbf{P}(D_{\tau}^{\mathrm{tr}}; \phi)$ is not dependent on $k$ . (e) GAP adapts $\mathbf{P}(\theta_{\tau, k}; \phi)$ : a fully adaptive preconditioner is used where it is task-specific and path-dependent. Instead of saying 'dependent on $k$ ', we specifically say it is path-dependent because the exact dependency is on the task-specific parameter set $\theta_{\tau, k}$ that is considerably more informative than $k$ . + +![](images/46b74e423ecb4e495bc02b70428a0eecb8433c5a9499a669faeeb13e4bbe74dd.jpg) +(c) Adaptive $\mathbf{P}(k;\phi)$ + +![](images/038e361c7d389e45c4e7073230ae9930ebb75bdd33dac471b48952256502f58e.jpg) +(d) Adaptive $\mathbf{P}(D_{\tau}^{tr};\phi)$ + +![](images/a5c6cd1ca0fac950bc4b3c6165c35dcab0dfd8582eb6825828d95786f0099240.jpg) +(e) Adaptive $\mathbf{P}(\theta_{\tau ,k};\phi)$ + +![](images/4a17d7c689a2c2d9f27ee5d1d4ee36f32ecd5bf0dafc979e08fa11e9a17aa020.jpg) +(a) Gradient Descent (GD) + +![](images/23f7288c3df23d2178d36282242e1bced73b5cfd1cb10de0b9bc550a8e2ac384.jpg) +(b) Preconditioned GD (Non-Riemannian metric) + +![](images/bb7ee49f34060de67ab21dc136cd1251f87a3d5dcaae13f93e4a578395eeae29.jpg) +(c) Preconditioned GD (Riemannian metric) +Figure 2. A toy example for illustrating the effect of Riemannian metric (see Supplementary Section A). When the curvature of the parameter space is poorly conditioned, (a) gradient descent can suffer from the difficulty of finding the solution, (b) a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can suffer further, and (c) a preconditioned gradient descent with a preconditioner that is a Riemannian metric can perform better. + +contrary, if the preconditioner is the Riemannian metric corresponding the parameter space, the preconditioned gradient descent can become the steepest descent and can exhibit a better optimization behavior as shown in Figure 2(c). While the Riemannian metric condition (i.e., positive definiteness) is a necessary condition for steepest descent learning, the existing studies on PGD-MAML family did not consider constraining preconditioners to satisfy the condition for Riemannian metric. + +In this study, we propose a new PGD method named Geometry Aaptive Preconditioned gradient descent (GAP). + +Specifically, GAP satisfies two desirable properties which have not been considered before. First, GAP's preconditioner $\mathbf{P}_{\mathrm{GAP}}$ is a fully adaptive preconditioner that can adapt to the individual task (task-specific) and to the optimization-path (path-dependent). The full adaptation is made possible by having the preconditioner depend on the task-specific parameter $\theta_{\tau,k}$ (Figure 1(e)). Second, we prove that $\mathbf{P}_{\mathrm{GAP}}$ is a Riemannian metric. To this end, we force the metaparameters of $\mathbf{P}_{\mathrm{GAP}}$ to be positive definite. Thus, GAP guarantees the steepest descent learning on the parameter space corresponding to $\mathbf{P}_{\mathrm{GAP}}$ . Owing to the two properties, GAP enables a geometry-adaptive learning in inner-loop optimization. + +For the implementation of GAP, we utilize the Singular Value Decomposition (SVD) operation to come up with our preconditioner satisfying the desired properties. For the recently proposed large-scale architectures, computational overhead can be an important design factor and we provide a low-computational approximation, Approximate GAP, that can be proven to asymptotically approximate the operation of GAP. + +To demonstrate the effectiveness of GAP, we empirically evaluate our algorithm on popular few-shot learning tasks; few-shot regression, few-shot classification, and few-shot cross-domain classification. The results show that GAP outperforms the state-of-the-art MAML family and PGDMAML family. + +The main contributions of our study can be summarized as follows: + +- We propose a new preconditioned gradient descent method called GAP, where it learns a preconditioner that enables a geometry-adaptive learning in the inner-loop optimization. +- We prove that GAP's preconditioner has two desirable properties: (1) It is both task-specific and path-dependent (i.e., dependent on task-specific parameter $\theta_{\tau, k}$ ). (2) It is a Riemannian metric. + +- For large-scale architectures, we provide a low-computational approximation that can be theoretically shown to approximate the GAP method. +- For popular few-shot learning benchmark tasks, we empirically show that GAP outperforms the state-of-the-art MAML family and PGD-MAML family. + +# 2. Background + +# 2.1. Model-Agnostic Meta-Learning (MAML) + +The goal of MAML [20] is to find the best initialization that the model can quickly adapt from, such that the model can perform well for a new task. MAML consists of two levels of main optimization processes: inner-loop and outer-loop optimizations. Consider the model $f_{\theta}(\cdot)$ with parameter $\theta$ . For a task $\tau = \{D_{\tau}^{\mathrm{tr}}, D_{\tau}^{\mathrm{val}}\}$ sampled from the task distribution $p(\mathcal{T})$ , the inner-loop optimization is defined as: + +$$ +\theta_ {\tau , K} = \theta_ {\tau , 0} - \alpha \sum_ {k = 0} ^ {K - 1} \nabla_ {\theta_ {\tau , k}} \mathcal {L} _ {\tau} ^ {\mathrm {i n}} \left(\theta_ {\tau , k}; D _ {\tau} ^ {\mathrm {t r}}\right) s. t \theta_ {\tau , 0} = \theta , \tag {1} +$$ + +where $\theta_{\tau ,k}$ is task-specific parameters for task $\tau$ and $\alpha$ is the learning rate for inner-loop optimization, $\mathcal{L}_{\tau}^{\mathrm{in}}$ is the inner-loop's loss function, and $K$ is the number of gradient descent steps. With $D_{\tau_i}^{\mathrm{val}}$ in each task, we can define outer-loop optimization as + +$$ +\theta \leftarrow \theta - \beta \nabla_ {\theta} \mathbb {E} _ {\tau} \left[ \mathcal {L} _ {\tau} ^ {\text {o u t}} \left(\theta_ {\tau , K}; D _ {\tau} ^ {\text {v a l}}\right) \right], \tag {2} +$$ + +where $\beta$ is the learning rate for outer-loop optimization, and $\mathcal{L}_{\tau}^{\mathrm{out}}$ is the outer-loop's loss function. + +# 2.2. Preconditioned Gradient Descent (PGD) + +PGD is a method that minimizes the empirical risk through a gradient update with a preconditioner that rescales the geometry of the parameter space. Given model parameters $\theta$ and task $\tau = \{D_{\tau}^{\mathrm{tr}}, D_{\tau}^{\mathrm{val}}\}$ , we can define the preconditioned gradient update with a preconditioner $\mathbf{P}$ as follows: + +$$ +\begin{array}{l} \theta_ {\tau , k + 1} = \theta_ {\tau , k} - \alpha \mathbf {P} \nabla_ {\theta_ {\tau , k}} \mathcal {L} _ {\tau} \left(\theta_ {\tau , k}; D _ {\tau} ^ {\mathrm {t r}}\right) \\ k = 0, 1, \dots \text {a n d} \theta_ {\tau , 0} = \theta \\ \end{array} +$$ + +where $\mathcal{L}_{\tau}(\theta_{\tau,k}; D_{\tau}^{\mathrm{tr}})$ is the empirical loss for task $\tau$ and parameters $\theta_{\tau,k}$ . Setting $\mathbf{P} = \mathbf{I}$ recovers Eq. (3) to the basic gradient descent (GD). Choice of $\mathbf{P}$ for exploiting the second-order information includes the inverse Fisher information matrix $\mathbf{F}^{-1}$ which leads to the natural gradient descent (NGD) [4]; the inverse Hessian $\mathbf{H}^{-1}$ which corresponds to the Newton's method [31]; and the diagonal matrix estimation with the past gradients which results in the adaptive gradient methods [18, 28]. They often reduce the effect of pathological curvature and speed up the optimization [5]. + +# 2.3. Unfolding: reshaping a tensor into a matrix + +In this study, the concept of unfolding is used to transform the gradient tensor of convolutional kernels into a matrix form. Tensor unfolding, also known as matricization or flattening, is the process of reshaping the elements of an $N$ -dimensional tensor $\mathbf{X} \in \mathbb{R}^{I_1 \times \dots \times I_N}$ into a matrix [29]. The mode- $n$ unfolding of an $N$ -dimensional tensor $\mathbf{X} \in \mathbb{R}^{I_1 \times \dots \times I_N}$ is defined as: + +$$ +\mathbf {X} \xrightarrow [ \text {m o d e - n u n f o l d i n g} ]{} \mathbf {X} _ {[ n ]} \in \mathbb {R} ^ {I _ {n} \times I _ {M}}, \text {w h e r e} I _ {M} = \prod_ {k \neq n} I _ {k} \tag {4} +$$ + +For example, the weight tensor of a convolutional layer is represented as a 4-D tensor $(\mathbf{W} \in \mathbb{R}^{C_{\mathrm{out}} \times C_{\mathrm{in}} \times k_h \times k_w})$ , where it is composed of kernels and it can be unfolded into a matrix as one of the following four forms: (1) $\mathbf{W}_{[1]} \in \mathbb{R}^{C_{\mathrm{out}} \times (C_{\mathrm{in}}k_hk_w)}$ , (2) $\mathbf{W}_{[2]} \in \mathbb{R}^{C_{\mathrm{in}} \times (C_{\mathrm{out}}k_hk_w)}$ , (3) $\mathbf{W}_{[3]} \in \mathbb{R}^{k_h \times (C_{\mathrm{out}}C_{\mathrm{in}}k_w)}$ , (4) $\mathbf{W}_{[4]} \in \mathbb{R}^{k_w \times (C_{\mathrm{out}}C_{\mathrm{in}}k_h)}$ . + +# 2.4. Riemannian manifold + +An $n$ -dimensional Riemannian manifold is defined by a manifold $\mathcal{M}$ and a Riemannian metric $g: \mathcal{M} \to \mathbb{R}^{n \times n}$ , which is a smooth function from each point $\pmb{x} \in \mathcal{M}$ to a positive definite matrix [32]. The metric $g(\pmb{x})$ defines the inner product of two tangent vectors for each point of the manifold $\langle \cdot, \cdot \rangle: \mathcal{T}_{\pmb{x}}\mathcal{M} \times \mathcal{T}_{\pmb{x}}\mathcal{M} \to \mathbb{R}$ , where $\mathcal{T}_{\pmb{x}}\mathcal{M}$ is the tangent space of $\pmb{x}$ . For $\pmb{v}, \pmb{w} \in \mathcal{T}_{\pmb{x}}\mathcal{M}$ , the inner product can be expressed $\langle \pmb{v}, \pmb{w} \rangle = \pmb{v}^T g(\pmb{x})\pmb{w}$ . A Riemannian manifold can be characterized by the curvature of the curves defined by a metric. The curvature of a Riemannian manifold can be computed at each point of the curves, while some manifolds have curvatures of a constant value. For example, the unit sphere $\mathcal{S}$ has constant positive curvature of $+1$ . + +# 3. Methodology + +In this section, we propose a new preconditioned gradient descent method called GAP in the MAML framework. In Section 3.1, we introduce GAP in the inner-loop optimization and describe how to meta-train GAP in the outer-loop optimization. In Section 3.2, we prove that GAP has two desirable properties. In Section 3.3, we provide a low-computational approximation that can be useful for large-scale architectures. + +# 3.1. GAP: Geometry-Adaptive Preconditioner + +# 3.1.1 Inner-loop optimization + +We consider an $L$ -layer neural network $f_{\theta}(\cdot)$ with parameters $\theta = \{\mathbf{W}^1, \dots, \mathbf{W}^l, \dots, \mathbf{W}^L\}$ . In the standard MAML with task $\tau \sim p(\mathcal{T})$ , each $\mathbf{W}^l$ is adapted with the gradient update as below: + +$$ +\mathbf {W} _ {\tau , K} ^ {l} \leftarrow \mathbf {W} _ {\tau , 0} ^ {l} - \alpha \cdot \sum_ {k = 0} ^ {K - 1} \mathbf {G} _ {\tau , k} ^ {l} s. t \mathbf {W} _ {\tau , 0} ^ {l} = \mathbf {W} ^ {l}, \tag {5} +$$ + +Algorithm 1 Geometry-Adaptive Preconditioned gradient descent (GAP) +Require: $p(\mathcal{T})$ : distribution over tasks +Require: $\alpha ,\beta_{1},\beta_{2}$ : learning rates 1: Randomly initialize $\theta = \{\mathbf{W}^1,\dots ,\mathbf{W}^L\}$ +2: Initialize $\phi = \{\mathbf{M}^{1},\dots ,\mathbf{M}^{L}\}$ as identity matrix. +3: while not converged do +4: Sample a batch of tasks $\mathcal{T}_i\sim p(\mathcal{T})$ +5: for all $\tau \in \mathcal{T}_i$ do +6: for inner-loop step $k = 0$ to $K - 1$ do +7: for layer $l = 1$ to $L$ do +8: Compute $\mathbf{G}_{\tau ,k}^{l} = \nabla_{\mathbf{W}_{\tau ,k}^{l}}\mathcal{L}_{\tau}^{\mathrm{in}}(\theta_{\tau ,k};D_{\tau}^{\mathrm{tr}})$ using $D_{\tau}^{\mathrm{tr}}$ +9: Reshape $\mathbf{G}_{\tau ,k}^{l}$ to $\mathbf{G}_{\tau ,k}^{l}$ via Eq. (6) +10: Transform $\mathbf{G}_{\tau ,k}^{l}$ to $\tilde{\mathbf{G}}_{\tau ,k}^{l}$ using $\mathbf{M}^l$ via Eq. (8) +11: Reshape $\tilde{\mathbf{G}}_{\tau ,k}^{l}$ back to the original form of gradient tensor, $\tilde{\mathbf{G}}_{\tau ,k}^{l}$ +12: Compute $l$ -layer adapted weight: $\mathbf{W}_{\tau ,k + 1}^{l} = \mathbf{W}_{\tau ,k}^{l} - \alpha \cdot \tilde{\mathbf{G}}_{\tau ,k}^{l}$ +13: end for +14: end for +15: Compute $\mathcal{L}_{\tau}^{\mathrm{out}}(\theta_{\tau ,K};D_{\tau}^{\mathrm{val}})$ by evaluating $\mathcal{L}_{\tau}^{\mathrm{out}}$ w.r.t $D_{\tau}^{\mathrm{val}}$ +16: end for +17: Update the weights and meta parameters: +18: $\theta \gets \theta -\beta_1\nabla_\theta \sum_{\tau \in \mathcal{T}}\mathcal{L}_{\tau}^{\mathrm{out}}(\theta_{\tau ,K};D_{\tau}^{\mathrm{val}})$ +19: $\phi \leftarrow \phi -\beta_{2}\nabla_{\phi}\sum_{\tau \in \mathcal{T}}\mathcal{L}_{\tau}^{\mathrm{out}}(\theta_{\tau ,K};D_{\tau}^{\mathrm{val}})$ +20: end while + +where $\mathbf{G}_{\tau,k}^{l} = \nabla_{\mathbf{W}_{\tau,k}^{l}} \mathcal{L}_{\tau}^{\mathrm{in}}(\theta_{\tau,k}; D_{\tau}^{\mathrm{tr}})$ is the gradient with respect to $\mathbf{W}_{\tau,k}^{l}$ and $\alpha$ is the learning rate for inner-loop optimization. In the GAP, we first use the mode-1 unfolding to reshape the gradient tensor into a matrix form (see Section 2.3). For a convolutional layer (i.e., $\mathbf{G}_{\tau,k}^{l} \in \mathbb{R}^{C_{\mathrm{out}} \times C_{\mathrm{in}} \times k \times k}$ ), we reshape the gradient tensor as below: + +$$ +\mathbf {G} _ {\tau , k} ^ {l} \xrightarrow [ \text {m o d e - 1} ]{} \left\{ \begin{array}{l l} \mathbf {G} _ {\tau , k} ^ {l} \in \mathbb {R} ^ {C _ {\mathrm {o u t}} \times C _ {\mathrm {i n}} k ^ {2}} & \text {i f} C _ {\mathrm {o u t}} \leq C _ {\mathrm {i n}} k ^ {2} \\ \mathbf {G} _ {\tau , k} ^ {l} \in \mathbb {R} ^ {C _ {\mathrm {i n}} k ^ {2} \times C _ {\mathrm {o u t}}} & \text {i f} C _ {\mathrm {i n}} k ^ {2} < C _ {\mathrm {o u t}}, \end{array} \right. \tag {6} +$$ + +where $\mathbf{G}_{\tau ,k}^{l}$ denotes $(\mathbf{G}_{\tau ,k}^{l})_{[1]}$ for the notational brevity. Note that we chose mode-1 unfolding because it performs best among the four unfolding forms as shown in Table 1. For a linear layer in a matrix form, there is no need for an unfolding. Second, we transform the singular values of the gradient matrix using additional meta parameters $\phi = \{\mathbf{M}^l\}_{l = 1}^L$ . The meta parameter $\mathbf{M}^l$ are diagonal matrices with positive elements defined as: + +$$ +\mathbf {M} ^ {l} = \left\{ \begin{array}{l l} \operatorname {d i a g} \left(\operatorname {S p} \left(m _ {1} ^ {l}\right), \dots , \operatorname {S p} \left(m _ {C _ {\text {o u t}}} ^ {l}\right)\right) & \text {i f} C _ {\text {o u t}} \leq C _ {\text {i n}} k ^ {2} \\ \operatorname {d i a g} \left(\operatorname {S p} \left(m _ {1} ^ {l}\right), \dots , \operatorname {S p} \left(m _ {C _ {\text {i n}} k ^ {2}} ^ {l}\right)\right) & \text {i f} C _ {\text {i n}} k ^ {2} < C _ {\text {o u t}} \end{array} , \right. \tag {7} +$$ + +where $m_i^l\in \mathbb{R}$ and $\mathbf{Sp}(x) = \frac{1}{2}\cdot \log (1 + \exp (2*x))$ . They are applied to the gradient matrix as follows: + +$$ +\tilde {\mathbf {G}} _ {\tau , k} ^ {l} = \mathbf {U} _ {\tau , k} ^ {l} \left(\mathbf {M} ^ {l} \cdot \boldsymbol {\Sigma} _ {\tau , k} ^ {l}\right) \mathbf {V} _ {\tau , k} ^ {l}, \tag {8} +$$ + +where $\mathbf{G}_{\tau,k}^{l} = \mathbf{U}_{\tau,k}^{l}\pmb{\Sigma}_{\tau,k}^{l}\mathbf{V}_{\tau,k}^{l}$ is the singular value decomposition (SVD) of $\mathbf{G}_{\tau,k}^{l}$ . Finally, we reshape $\tilde{\mathbf{G}}_{\tau,k}^{l}$ back to its original gradient tensor form $\tilde{\mathbf{G}}_{\tau,k}^{l}$ (i.e., inverse unfolding). The resulting preconditioned gradient descent of GAP + +
mode-1mode-2mode-3mode-4
54.86 ± 0.8553.02 ± 0.8751.23 ± 0.7651.45 ± 0.77
+ +Table 1. Performance comparison of unfolding types. We performed the experiment with 5-way 1-shot on mini-[ImageNet] and used the standard Conv-4 backbone. + +becomes the following: + +$$ +\mathbf {W} _ {\tau , K} ^ {l} \leftarrow \mathbf {W} _ {\tau , 0} ^ {l} - \alpha \cdot \sum_ {k = 0} ^ {K - 1} \tilde {\mathbf {G}} _ {\tau , k} ^ {l} s. t \mathbf {W} _ {\tau , 0} ^ {l} = \mathbf {W} ^ {l}, \tag {9} +$$ + +where $\tilde{\mathbf{G}}_{\tau ,k}^{l}$ is the preconditioned gradient based on the meta parameters $\phi$ + +# 3.1.2 Outer-loop optimization + +For outer-loop optimization, GAP follows the typical process of MAML. Unlike MAML, however, GAP meta-learns two meta parameter sets $\theta$ and $\phi$ as follows: + +$$ +\theta \leftarrow \theta - \beta_ {1} \nabla_ {\theta} \mathbb {E} _ {\tau} \left[ \mathcal {L} _ {\tau} ^ {\text {o u t}} \left(\theta_ {\tau , K}; D _ {\tau} ^ {\text {v a l}}\right) \right], \tag {10} +$$ + +$$ +\phi \leftarrow \phi - \beta_ {2} \nabla_ {\phi} \mathbb {E} _ {\tau} \left[ \mathcal {L} _ {\tau} ^ {\text {o u t}} \left(\theta_ {\tau , K}; D _ {\tau} ^ {\text {v a l}}\right) \right], \tag {11} +$$ + +where $\beta_{1}$ and $\beta_{2}$ are the learning rates for the outer-loop optimization. We initialize $\mathbf{M}^l$ as an identity matrix for all $l$ . The training procedure is provided in Algorithm 1. + +# 3.2. Desirable properties of GAP + +In this section, we prove that GAP's preconditioner $\mathbf{P}_{\mathrm{GAP}}$ satisfies two desirable properties. + +Theorem 1. Let $\tilde{\mathbf{G}}_{\tau,k}^{l} \in \mathbb{R}^{m \times n}$ be the 'l-layer $k$ -th inner-step' gradient matrix transformed by meta parameter $\mathbf{M}^l$ for task $\tau$ . Then preconditioner $\mathbf{P}_{GAP}$ induced by $\tilde{\mathbf{G}}_{\tau,k}^{l}$ is a Riemannian metric and depends on the task-specific parameters $\theta_{\tau,k}$ . + +The proof and closed form of $\mathbf{P}_{\mathrm{GAP}}$ are provided in Supplementary Section B. The following two properties emerge from the theorem. + +Property 1. Dependency on task-specific parameters: Theorem 1 formally shows that $\mathbf{P}_{\mathrm{GAP}}$ depends on the task-specific parameters $\theta_{\tau,k}$ . While the previous studies considered a non-adaptive preconditioner $\mathbf{P}(\phi)$ [34, 36, 44, 49, 57, 66] and a partially adaptive preconditioner $\mathbf{P}(k;\phi)$ [49] or $\mathbf{P}(D_{\tau}^{\mathrm{tr}};\phi)$ [53], $\mathbf{P}_{\mathrm{GAP}}$ can be considered to be the most advanced adaptive preconditioner because it is fully adaptive (i.e., task-specific and path-dependent) by being dependent on $\theta_{\tau,k}$ as shown in Figure 1. + +Property 2. Riemannian metric: If the parameter space has a certain underlying structure, the ordinary gradient of a function $\nabla \mathcal{L}$ does not represent its steepest direction [4]. + +To define the steepest direction on the parameter space, we need a Riemannian metric $g(\boldsymbol{w})$ , which is a positive-definite matrix defined for each parameter $\boldsymbol{w}$ . A Riemannian metric defines the steepest descent direction by $-g(\boldsymbol{w})^{-1}\nabla \mathcal{L}$ [4]. If a preconditioning matrix is a Riemannian metric, it defines the geometry of the underlying structure and enables steepest descent learning. Because we prove that $\mathbf{P}_{\mathrm{GAP}}$ is a Riemannian metric for each parameter in Theorem 1, $\mathbf{P}_{\mathrm{GAP}}$ is theoretically guaranteed to enable steepest descent learning on its corresponding parameter space. $\mathbf{P}_{\mathrm{GAP}}$ consists of two factors, a unitary matrix of the inner-loop gradient $\mathbf{U}_{\tau,k}$ and a meta-parameter $\mathbf{M}$ ; $\mathbf{M}$ enables us to reflect the shared geometry information across the tasks. Task-specific and path-dependent geometry information can be reflected in the metric through $\mathbf{U}_{\tau,k}$ . Two factors allow our Riemannian metric to have higher function complexity than a constant metric. For example, we can consider various structures other than a unit-sphere that corresponds to the constant metric of $+1$ . Even though $\mathbf{P}_{\mathrm{GAP}}$ is guaranteed to be a Riemannian metric, it is crucial that the meta-learned $\mathbf{P}_{\mathrm{GAP}}$ corresponds to the true parameter space or at least $\mathbf{P}_{\mathrm{GAP}}$ is close enough to be useful. We will discuss this issue in Section 5. + +# 3.3. Approximate GAP: a low-computational approximation of GAP + +As presented in Section 3.1, GAP uses an SVD operation. The SVD operation can be burdensome for large-scale networks because it implies that the computational cost can be significantly increased. In recent studies, the use of a large-scale architecture has been emphasized as the key factor for improving the performance of meta-learning [25]. To make use of GAP for large-scale architectures without causing a computational problem, we provide an efficient approximation, named Approximate GAP, under an assumption. + +Assumption 1. The elements of the gradient matrix follow an i.i.d. normal distribution with zero means. + +Following [39, 60], we adopt the assumption to have the gradient matrix become orthogonal as $n$ increases. Although the utilization of the assumption is a limiting factor, we empirically confirmed that the row vectors of the gradient matrix are indeed asymptotically orthogonal as $n$ increases (see Figure 3). For the assumption, our approximation can be established as the following. + +Theorem 2. Let $\mathbf{G} \in \mathbb{R}^{m \times n}$ be a gradient matrix and $\tilde{\mathbf{G}}$ be the gradient transformed by a meta parameter $\mathbf{M}$ . Under the Assumption 1, as $n$ becomes large, $\tilde{\mathbf{G}}$ asymptotically becomes equivalent to $\mathbf{MG}$ as follows: + +$$ +\tilde {\mathbf {G}} \cong \mathbf {M G} \tag {12} +$$ + +Note that we have chosen the larger dimension of the gradient matrix as $n$ when reshaping with Eq. (6). We provide + +![](images/31919d55723fcabb9ab7563c5f8fe11f0bbc340a025fd60acea68c3f1cb5061e.jpg) +Figure 3. The average of cosine similarity between row vectors of gradient matrix as $n$ increases. + +
Algorithmtrain time (secs)test time (secs)GPU-memory (MB)
MAML35796.6152.6610185
GAP50916.63120.4410279
Approximate GAP36684.7253.6210217
+ +Table 2. Comparison of training time, GPU memory, and test time for MAML, GAP, and Approximate GAP. We performed the experiment with 5-way 1-shot on mini-ImageNet and used 600 tasks in the test. We used the standard Conv-4 backbone. + +the proof in Supplementary Section B. This approximation has a clear trade-off between scalability and adaptiveness. Approximate GAP efficiently reduces the computational cost as shown in Table 2, whereas the preconditioner $\mathbf{P}_{\mathrm{GAP}}$ becomes not adaptive but constant. However, Approximate GAP still guarantees the preconditioner to be a Riemannian metric. As shown in Table 4 & 5, Approximate GAP incurs a slight performance drop but it still achieves a high performance owing to the Riemannian metric property. + +# 4. Experiments + +In this section, we show the superiority of GAP by comparing it with the state-of-the-art PGD-MAML family and the MAML family. Hyper-parameter setups used in our experiments can be found in Supplementary Section C.1. + +# 4.1. Few-shot regression + +Datasets and experimental setup. The goal of few-shot regression is to fit an unknown target function for the given $K$ sample points from the function. For the evaluation of few-shot regression, we use the sinusoid regression benchmark [20]. In this benchmark, sinusoid is used as the target function. Each task has a sinusoid $y(x) = A\sin (\omega x + b)$ as the target function, where the parameter values are within the following range: amplitude $A\in [0.1,5.0]$ , frequency $\omega \in [0.8,1.2]$ , and phase $b\in [0,\pi ]$ . For each task, input data point $x$ is sampled from $[-5.0,5.0]$ . In the experiment, we use a simple Multi-Layer Perceptron (MLP), following the setting in [20]. The details of the architecture are provided in Supplementary Section C.2.1. + +Results. We evaluate GAP and compare it with MAML family and PGD-MAML family on a regression task. As + +
Algorithm5-shot10-shot20-shot
MAML [20]1.13 ± 0.180.77 ± 0.110.48 ± 0.08
Meta-SGD† [36]0.90 ± 0.160.53 ± 0.090.31 ± 0.05
MT-Net [34]0.76 ± 0.090.49 ± 0.050.33 ± 0.04
ALFA [9]0.92 ± 0.190.62 ± 0.160.34 ± 0.07
L2F [10]0.71 ± N/A0.37 ± N/A0.16 ± N/A
PAMELA† [49]0.54 ± 0.060.41 ± 0.040.17 ± 0.03
MeTAL [8]0.74 ± 0.180.44 ± 0.110.21 ± 0.06
GAP†0.16 ± 0.040.04 ± 0.010.01 ± 0.01
+ +Table 3. Few-shot regression for the sinusoid regression benchmark with a 2-layer MLP backbone. We report MSE ± 95% confidence intervals(ci) for 600 tasks following the setup in [20].† denotes PGD-MAML family. + +shown in Table 3, GAP consistently achieves the lowest mean squared error (MSE) scores, with the lowest confidence intervals in all three cases. The performance of GAP is improved by $89\%$ on 10-shot and $94\%$ on 20-shot compared to the performance of state-of-the-art algorithms. + +# 4.2. Few-shot classification + +Datasets and experimental setup. For the few-shot classification, we evaluate two benchmarks: (1) miniImageNet [56]; this dataset has 100 classes and it is a subset of ImageNet [16], and we use the same split as in [51], with 64, 16 and 20 classes for train, validation and test, respectively. (2) tiered-ImageNet [52]; this is also a subset of ImageNet with 608 classes grouped into 34 high-level categories, and divided into 20, 6 and 8 for train, validation, and test, respectively. For all the experiments, our model follows the standard Conv-4 backbone used in [56]. The details of the architecture are provided in Supplementary Section C.2.2. Following the experimental protocol in [20], we use 15 samples per class in the query-set to compute the meta gradients. In meta training and meta testing, the inner-loop optimization is updated in five and ten steps, respectively. + +Results. Table 4 & 5 present the performance of GAP, the state-of-the-art PGD-MAML family, and the state-of-the-art MAML-family on mini-[ImageNet and tiered-[ImageNet under two typical settings: 5-way 1-shot and 5-way 5-shot. The GAP outperforms all of the previous PGD-MAML family and MAML family. Compared to the state-of-the-art MAML family, GAP improves the performance with a quite significant margin for both mini-[ImageNet and tiered-[ImageNet datasets. Compared to the state-of-the-art PGD-MAML family, GAP shows that the 1- and 5-shot accuracy can be increased by $1.4\%$ and $1.5\%$ on mini-[ImageNet dataset, and by $0.7\%$ and $0.68\%$ on tiered-[ImageNet dataset, respectively. We also evaluated Approximate GAP that is introduced in Section 3.3. The results show that the approximated version can perform comparably to the original GAP. Although Approximate GAP shows slightly lower accuracies than the original, its performance is superior to most of the existing algorithms because of its Riemannian metric property. + +
Algorithm5-way 1-shot5-way 5-shot
MAML [20]47.89 ± 1.2064.59 ± 0.88
Meta-SGD† [36]50.47 ± 1.8764.00 ± 0.90
BMAML [64]53.80 ± 1.4664.23 ± 0.69
ANIL [48]46.70 ± 0.4061.50 ± 0.50
LLAMA [23].49.40 ± 1.83N/A
PLATIPUS [21]50.13 ± 1.86-
T-net [34]50.86 ± 1.82N/A
MT-net [34]51.70 ± 1.84N/A
MAML++ [7]52.15 ± 0.2668.32 ± 0.44
iMAML-HF [50]49.30 ± 1.88N/A
WarpGrad [22]52.30 ± 0.9068.40 ± 0.60
MC1† [44]53.74 ± 0.8468.01 ± 0.73
MC2† [44]54.08 ± 0.8867.99 ± 0.73
MH-C† [66]48.64 ± 0.3364.52 ± 0.51
MH† [66]49.41 ± 0.9667.16 ± 0.42
BOIL [43]49.61 ± 0.1666.46 ± 0.37
ARML [63]50.42 ± 1.7964.12 ± 0.90
ALFA [9]50.58 ± 0.5169.12 ± 0.47
L2F [10]52.10 ± 0.5069.38 ± 0.46
ModGrad† [53]53.20 ± 0.8669.17 ± 0.69
PAMELA† [49]53.50 ± 0.8970.51 ± 0.67
SignMAML [19]42.90 ± 1.5060.70 ± 0.70
CTML [45]50.47 ± 1.8364.15 ± 0.90
MeTAL [8]52.63 ± 0.3770.52 ± 0.29
ECML [24]48.94 ± 0.8065.26 ± 0.67
Sharp-MAML_up [1]49.56 ± N/A63.06 ± N/A
Sharp-MAML_low [1]49.72 ± N/A63.18 ± N/A
Sharp-MAMLboth [1]50.28 ± N/A65.04 ± N/A
FBM [62]50.62 ± 1.7964.78 ± 0.35
CxGrad [33]51.80 ± 0.4669.82 ± 0.42
HyperMAML [47]51.84 ± 0.5766.29 ± 0.43
EEML [35]52.42 ± 1.7568.40 ± 0.95
MH-O† [66]52.50 ± 0.6167.76 ± 0.34
Sparse-MAML† [57]50.35 ± 0.3967.03 ± 0.74
Sparse-ReLU-MAML† [57]50.39 ± 0.8964.84 ± 0.46
Sparse-MAML+† [57]51.04 ± 0.5967.03 ± 0.74
Approximate GAP†53.52 ± 0.8870.75 ± 0.67
GAP†54.86 ± 0.8571.55 ± 0.61
+ +Table 4. 5-way few-shot classification accuracy $(\%)$ on miniImageNet with a Conv-4 backbone. We report mean $\pm 95\%$ confidence intervals(ci) for 600 tasks according to [20]. $\dagger$ denotes PGD-MAML family. + +
Algorithm5-way 1-shot5-way 5-shot
Meta-SGD†[36]50.92 ± 0.9369.28 ± 0.80
MAML [20]51.70 ± 1.8070.30 ± 1.80
MT-net [34]51.95 ± 1.83N/A
WarpGrad [22]57.20 ± 0.9074.10 ± 0.70
BOIL [43].48.58 ± 0.2769.37 ± 0.12
ALFA [9]53.16 ± 0.4970.54 ± 0.46
L2F [10]54.40 ± 0.5073.34 ± 0.44
ARML [63]52.91 ± 1.83N/A
PAMELA†[49]54.81 ± 0.8874.39 ± 0.71
Sparse-ReLU-MAML†[57]53.18 ± 0.5269.06 ± 0.28
Sparse-MAML†[57]53.47 ± 0.5368.83 ± 0.65
Sparse-MAML+†[57]53.91 ± 0.6769.92 ± 0.21
MeTAL [8]54.34 ± 0.3170.40 ± 0.21
CxGrad [33]55.55 ± 0.4673.55 ± 0.41
ECML [24]47.34 ± 0.8864.77 ± 0.75
Approximate GAP†56.86 ± 0.9174.41 ± 0.72
GAP†57.60 ± 0.9374.90 ± 0.68
+ +Table 5. 5-way few-shot classification accuracy (%) on tiered-ImageNet dataset with a Conv-4 backbone. We report mean ± 95% confidence intervals(ci) for 600 tasks according to [20].† denotes PGD-MAML family. + +
Algorithmtiered-ImageNetCUBCars
1-shot5-shot1-shot5-shot1-shot5-shot
MAML [20]51.61 ± 0.2065.76 ± 0.2740.51 ± 0.0853.09 ± 0.1633.57 ± 0.1444.56 ± 0.21
ANIL [48]52.82 ± 0.2966.52 ± 0.2841.12 ± 0.1555.82 ± 0.2134.77 ± 0.3146.55 ± 0.29
BOIL [43]53.23 ± 0.4169.37 ± 0.2344.20 ± 0.1560.92 ± 0.1136.12 ± 0.2950.64 ± 0.22
BMAML [64]N/AN/A33.52 ± 0.3651.35 ± 0.16N/AN/A
ALFA [9]N/AN/AN/A58.35 ± 0.25N/AN/A
L2F [10]N/AN/AN/A60.89 ± 0.22N/AN/A
MeTAL [8]N/AN/AN/A58.20 ± 0.24N/AN/A
HyperMAML [47]N/AN/A36.52 ± 0.6149.43 ± 0.14N/AN/A
CxGrad [33]N/AN/AN/A63.92 ± 0.44N/AN/A
Sparse-MAML†[57]53.47 ± 0.5368.83 ± 0.6541.37 ± 0.7360.58 ± 1.1035.90 ± 0.5052.63 ± 0.56
Sparse-ReLU-MAML†[57]53.77 ± 0.9468.12 ± 0.6942.89 ± 0.4557.53 ± 0.9436.04 ± 0.5549.95 ± 0.42
Sparse-MAML+†[57]53.91 ± 0.6769.92 ± 0.2143.43 ± 1.0462.02 ± 0.7837.14 ± 0.7753.18 ± 0.44
Approximate GAP†57.47 ± 0.9971.66 ± 0.7643.77 ± 0.7962.92 ± 0.7337.00 ± 0.7553.28 ± 0.76
GAP†58.56 ± 0.9372.82 ± 0.7744.74 ± 0.7564.88 ± 0.7238.44 ± 0.7755.04 ± 0.77
+ +Table 6. 5-way few-shot cross domain classification accuracy (%) with a Conv-4 backbone, meta training on mini-[ImageNet dataset, and meta-testing on tiered-[ImageNet, CUB, or Cars datasets. We report mean ± 95% confidence intervals(ci) for 600 tasks according to [20].† denotes PGD-MAML family. + +# 4.3. Cross-domain few-shot classification + +The cross-domain few-shot classification introduced by [14] addresses a more challenging and practical few-shot classification scenario in which meta-train and meta-test tasks are sampled from different task distributions. These scenarios are designed to evaluate meta-level overfitting of meta-learning algorithms by creating a large domain gap between meta-trains and meta-tests. In particular, an algorithm can be said to be meta-overfitting if it relies too much on the prior knowledge of previously seen meta-train tasks instead of focusing on a few given examples to learn a new task. This meta-level overfitting makes the learning system more likely to fail to adapt to new tasks sampled from substantially different task distributions. + +Datasets and experimental setup. To evaluate the level of meta-overfitting for GAP, we evaluate a cross-domain few-shot classification experiment. The mini-ImageNet is used for the meta-train task, and the tiered-ImageNet [52], CUB-200-2011 [58], Cars [11] datasets are used for the meta-test task. The CUB dataset has 200 fine-grained classes and consists of a total of 11,788 images; it is further divided into 100 meta-train classes, 50 meta-validation classes, and 50 meta-test classes. The Cars [30] dataset consists of 16,185 images of 196 classes of cars; it is split into 8,144 training images and 8,041 testing images, where each class has been split roughly in 50-50. The classes are typically at the level of Make, Model, Year, e.g., 2012 Tesla Model S or 2012 BMW M3 coupe. As with the few-shot classification experiment, we use the standard Conv-4 backbone and follow the same experimental protocol. + +Results. Table 6 presents the cross-domain few-shot performance for GAP, MAML family, and PGD-MAML family. GAP significantly outperforms the state-of-the-art al + +gorithms on 5-way 1-shot and 5-way 5-shot cross-domain classification tasks. In particular, for the tiered-ImageNet dataset, the performance was improved by $8.6\%$ and $4.1\%$ on 1-shot and 5-shot classification tasks, respectively. Because GAP can simultaneously consider a task's individuality and optimization trajectory in the inner-loop optimization, it can overcome meta-overfitting better than the existing methods. However, Approximate GAP shows more performance degradation in cross-domain few-shot classification than in few-shot classification. In particular, when the domain difference with the meta-train is more significant (i.e., the tiered-ImageNet dataset) than when the domain difference with the meta-train is marginal (i.e., CARS and CUB datasets), it shows a more considerable performance drop. We can see that full adaptation plays an important role in cross-domain few-shot classification. + +# 5. Discussion + +Number of meta parameters: Recent MAML family and PGD-MAML family require a large increase in the number of meta-learning parameters as shown in Table 7. One advantage of GAP is that it requires only a very small increase in the number of meta parameters, when compared to the baseline of MAML. This is possible because we transform a gradient tensor into a gradient matrix, perform the SVD of the matrix, and assign only a small number of meta parameters that correspond to the diagonal matrix of the gradient matrix. For the Conv-4 network, GAP requires only $0.2\%$ increase of the meta parameters. Although the increase in the number of meta parameters is negligible, SVD of the gradient matrix can incur a large computational burden for large networks. This is addressed by Approximate GAP. + +Approximate GAP vs. simple constant preconditioners: Approximate GAP is a low-complexity method where SVD + +
Algorithm# of params% increase
MAML [20]1.2109 × 105
Meta-SGD [36]2.4218 × 105100.0%
MC [44]2.7106 × 1062140.4%
PAMELA [49]1.6239 × 10534.1%
MH [66]7.2196 × 10759586.7%
Sparse-MAML [57]2.4218 × 105100.0%
GAP1.2131 × 1050.2%
+ +Table 7. Comparison of the number of parameters for MAML, existing methods, and GAP. + +
AlgorithmStructureRiemannian metric (i.e., positive definite)Acc. (%)
Meta-SGDdiag(a1, ·, an)X50.47%
Meta-SGD with positive definitenessdiag(a1, ·, an)O52.39%
Approximate GAPblkdiag(Mt, ·, M)O53.52%
+ +Table 8. Performance comparison of three constant (i.e., non-adaptive) preconditioners for 5-way 1-shot on mini-ImageNet: Meta-SGD, Meta-SGD modified to satisfy positive definiteness, and our Approximate GAP. + +
Algorithm1-shot5-shot
GAP w/o PGAP48.23 ± 0.8065.80 ± 0.75
GAP w/ PGAP54.86 ± 0.8571.55 ± 0.61
+ +Table 9. Ablation study of $\mathbf{P}_{\mathrm{GAP}}$ on mini-ImageNet. Performance of the GAP-trained model is significantly affected by not applying $\mathbf{P}_{\mathrm{GAP}}$ . + +operation is avoided by approximating GAP with a constant diagonal preconditioner. A natural question to ask is how does Approximate GAP compare with other constant diagonal preconditioners. To answer this question, we have compared Approximate GAP with Meta-SGD and a modified Meta-SGD. Meta-SGD [36] is a well-known constant diagonal preconditioner (i.e., $\mathrm{diag}(a_1,\dots ,a_n)$ ) that does not need to be positive definite and we also investigate its modification with a constraint on positive definiteness. The results are shown in Table 8. It can be observed that enforcing positive definiteness can improve Meta-SGD. Furthermore, an additional improvement can be achieved by Approximate GAP. While both modified Meta-SGD and Approximate GAP are positive definite, Approximate GAP is different because it inherits an additional constraint from GAP - a block diagonal structure where a constant diagonal matrix $M$ is repeated (i.e., $\mathrm{blkdiag}(M,\dots ,M)$ ). The inherited constraint provides a gain over the modified Meta-SGD. + +Does GAP learn a useful preconditioner: While a Riemannian metric can be helpful, it does not mean any Riemannian metric will result in an improvement. For the true parameter space with a specific underlying structure, the corresponding Riemannian metric needs to be applied to enable steepest descent [3, 4]. For the special case of a two-layer neural network with a mean squared error (MSE) loss, it + +was proven that Fisher information matrix is the corresponding Riemannian metric [4]. For a general neural network, however, a proper Riemannian metric is unknown and it needs to be learned. In our work, we have devised a method to guarantee a Riemannian metric and have used the outer-loop optimization to learn the Riemannian metric. In general, the learned Riemannian metric is unlikely to correspond perfectly to the true parameter space. Then, an important question is if the Riemannian metric learned by GAP is close enough to the desired one and if it is useful. To investigate this issue, we performed an ablation study by not applying the preconditioner $\mathbf{P}_{\mathrm{GAP}}$ . After training a GAP model, we have evaluated the performance with and without applying $\mathbf{P}_{\mathrm{GAP}}$ . The results are shown in Table 9 and clearly the preconditioner learned with the outer-loop optimization plays an essential role for improving the performance. + +Why is preconditioner helpful for meta-learning: When the batch size is small, the resulting empirical gradient can be noisy [53, 65]. A typical few-shot learning has only a small number of samples for the inner-loop optimization, and its gradient can be noisy. On the other hand, it was shown in [5] that preconditioned gradient descent with a positive definite preconditioner can achieve a lower risk than gradient descent when the labels are noisy, the model is mis-specified, or the signal is misaligned with the features. Under a misalignment, a properly chosen positive definite preconditioner can generalize better than gradient descent [5]. Considering the noisy gradient of inner loop optimization, it can be surmised that a positive definite preconditioner that is adaptive (i.e., a Riemannian metric) can be helpful for improving MAML. Note that the noisy case is in contrast to the case of supervised learning with a large amount of data [5]. + +# 6. Conclusion + +In this work, we have proposed GAP that is a PGDMAML algorithm utilizing SVD operation. Thanks to the fully adaptive property, GAP can handle individual tasks with a wide diversity within the MAML framework. Additionally, GAP can enable steepest descent on the parameter space owing to the Riemannian metric property. + +# Acknowledgements + +This work was supported by ETRI [23ZR1100, A Study of Hyper-Connected Thinking Internet Technology by autonomous connecting, controlling and evolving ways], NRF (NRF-2020R1A2C2007139), IITP [NO.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University)], and the New Faculty Startup Fund from Seoul National University. + +# References + +[1] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning. arXiv preprint arXiv:2206.03996, 2022. 6 +[2] Shunichi Amari. A theory of adaptive pattern classifiers. IEEE Transactions on Electronic Computers, (3):299-307, 1967. 1 +[3] Shun-ichi Amari. Neural learning in structured parameter spaces-natural riemannian gradient. Advances in neural information processing systems, 9, 1996. 1, 8 +[4] Shun-Ichi Amari. Natural gradient works efficiently in learning. Neural computation, 10(2):251-276, 1998. 1, 3, 4, 5, 8 +[5] Shun-ichi Amari, Jimmy Ba, Roger Grosse, Xuechen Li, Atsushi Nitanda, Taiji Suzuki, Denny Wu, and Ji Xu. When does preconditioning help or hurt generalization? arXiv preprint arXiv:2006.10732, 2020. 3, 8 +[6] Shun-Ichi Amari and Scott C Douglas. Why natural gradient? In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), volume 2, pages 1213-1216. IEEE, 1998. 1 +[7] Antreas Antoniou, Harrison Edwards, and Amos Storkey. How to train your maml. arXiv preprint arXiv:1810.09502, 2018.6 +[8] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9465-9474, 2021. 1, 6, 7 +[9] Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. Meta-learning with adaptive hyperparameters. Advances in Neural Information Processing Systems, 33:20755-20765, 2020. 6, 7 +[10] Sungyong Baik, Seokil Hong, and Kyoung Mu Lee. Learning to forget for meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2379-2387, 2020. 6, 7 +[11] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 7 +[12] Irénée-Jules Bienaymé. Considerations à l'appui de la découverte de Laplace sur la loi de probabilité dans la méthode des moindres carrés. Imprimerie de Mallet-Bachelier, 1853. 13 +[13] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 1 +[14] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 7 +[15] Tristan Deleu, Tobias Würfl, Mandana Samiei, Joseph Paul Cohen, and Yoshua Bengio. Torchmeta: A meta-learning library for pytorch. arXiv preprint arXiv:1909.06576, 2019. 14 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. + +In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[17] Lin Ding, Peng Liu, Wenfeng Shen, Weijia Lu, and Shengbo Chen. Gradient-based meta-learning using uncertainty to weigh loss for few-shot learning. arXiv preprint arXiv:2208.08135, 2022. 1 +[18] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011. 3 +[19] Chen Fan, Parikshit Ram, and Sijia Liu. Sign-maml: Efficient model-agnostic meta-learning by signsgd. arXiv preprint arXiv:2109.07497, 2021. 6 +[20] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 1, 3, 5, 6, 7, 8, 14 +[21] Chelsea Finn, Kelvin Xu, and Sergey Levine. Probabilistic model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 1, 6 +[22] Sebastian Flennerhag, Andrei A Rusu, Razvan Pascanu, Francesco Visin, Hujun Yin, and Raia Hadsell. Meta-learning with warped gradient descent. arXiv preprint arXiv:1909.00025, 2019. 6 +[23] Erin Grant, Chelsea Finn, Sergey Levine, Trevor Darrell, and Thomas Griffiths. Recasting gradient-based meta-learning as hierarchical bayes. arXiv preprint arXiv:1801.08930, 2018.6 +[24] Markus Hiller, Mehrtash Harandi, and Tom Drummond. On enforcing better conditioned meta-learning for rapid few-shot adaptation. arXiv preprint arXiv:2206.07260, 2022. 6 +[25] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9068-9077, 2022. 5 +[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 14 +[27] Sham M Kakade. A natural policy gradient. Advances in neural information processing systems, 14, 2001. 1 +[28] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3,14 +[29] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 3 +[30] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 7 +[31] Yann A LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-48. Springer, 2012. 3 +[32] John M Lee. Smooth manifolds. Springer, 2012. 3, 12 +[33] Sanghyuk Lee, Seunghyun Lee, and Byung Cheol Song. Contextual gradient scaling for few-shot learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 834-843, 2022. 6, 7 + +[34] Yoonho Lee and Seungjin Choi. Gradient-based meta-learning with learned layerwise metric and subspace. In International Conference on Machine Learning, pages 2927-2936. PMLR, 2018. 1, 4, 6 +[35] Geng Li, Boyuan Ren, and Hongzhi Wang. Eeml: Ensemble embedded meta-learning. arXiv preprint arXiv:2206.09195, 2022. 6 +[36] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1, 4, 6, 8 +[37] Zequn Liu, Ruiyi Zhang, Yiping Song, and Ming Zhang. When does maml work the best? an empirical study on model-agnostic meta-learning in nlp applications. arXiv preprint arXiv:2005.11700, 2020. 1 +[38] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 14 +[39] Ahmed M Abdelmoniem, Ahmed Elzanaty, Mohamed-Slim Alouini, and Marco Canini. An efficient statistical-based gradient compression technique for distributed training systems. Proceedings of Machine Learning and Systems, 3:297-322, 2021. 5 +[40] Gabriel Maicas, Andrew P Bradley, Jacinto C Nascimento, Ian Reid, and Gustavo Carneiro. Training medical image analysis systems like radiologists. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 546-554. Springer, 2018. 1 +[41] George Marsaglia. Choosing a point from the surface of a sphere. The Annals of Mathematical Statistics, 43(2):645-646, 1972. 13 +[42] Fei Mi, Minlie Huang, Jiyong Zhang, and Boi Faltings. Meta-learning for low-resource natural language generation in task-oriented dialogue systems. arXiv preprint arXiv:1905.05644, 2019. 1 +[43] Jaehoon Oh, Hyungjun Yoo, ChangHwan Kim, and Se-Young Yun. Boil: Towards representation change for few-shot learning. arXiv preprint arXiv:2008.08882, 2020. 6, 7 +[44] Eunbyung Park and Junier B Oliva. Meta-curvature. Advances in Neural Information Processing Systems, 32, 2019. 1, 4, 6, 8 +[45] Danni Peng and Sinno Pan. Clustered task-aware meta-learning by learning from learning paths. 2021. 6 +[46] Juan-Manuel Perez-Rua, Xiatian Zhu, Timothy M Hospedales, and Tao Xiang. Incremental few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13846-13855, 2020. 1 +[47] M Przewięźlikowski, P Przybysz, J Tabor, M Zieba, and P Spurek. Hypermaml: Few-shot adaptation of deep models with hypernetworks. arXiv preprint arXiv:2205.15745, 2022. 6, 7 +[48] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 1, 6, 7 +[49] Jathushan Rajasegaran, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Mubarak Shah. Meta-learning + +the learning trends shared across tasks. arXiv preprint arXiv:2010.09291, 2020. 1, 4, 6, 8 +[50] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 6 +[51] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. 2016. 6 +[52] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 6, 7 +[53] Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. On modulating the gradient for meta-learning. In European Conference on Computer Vision, pages 556-572. Springer, 2020. 1, 4, 6, 8 +[54] Rishav Singh, Vandana Bharti, Vishal Purohit, Abhinav Kumar, Amit Kumar Singh, and Sanjay Kumar Singh. Metamed: Few-shot medical image classification using gradient-based meta-learning. Pattern Recognition, 120:108111, 2021. 1 +[55] Xingyou Song, Yuxiang Yang, Krzysztof Choromanski, Ken Caluwaerts, Wenbo Gao, Chelsea Finn, and Jie Tan. Rapidly adaptable legged robots via evolutionary meta-learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3769-3776. IEEE, 2020. 1 +[56] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016, 6, 14 +[57] Johannes Von Oswald, Dominic Zhao, Seijin Kobayashi, Simon Schug, Massimo Caccia, Nicolas Zucchet, and João Sacramento. Learning where to learn: Gradient sparsity in meta and continual learning. Advances in Neural Information Processing Systems, 34:5250-5263, 2021. 1, 4, 6, 7, 8 +[58] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 7 +[59] Shuhuan Wen, Zeteng Wen, Di Zhang, Hong Zhang, and Tao Wang. A multi-robot path-planning algorithm for autonomous navigation using meta-reinforcement learning based on transfer learning. Applied Soft Computing, 110:107605, 2021. 1 +[60] Simon Wiedemann, Temesgen Mehari, Kevin Kepp, and Wojciech Samek. Dithered backprop: A sparse and quantized backpropagation algorithm for more efficient deep neural network training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 720-721, 2020. 5 +[61] Xiongwei Wu, Doyen Sahoo, and Steven Hoi. Meta-rcnn: Meta learning for few-shot object detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1679–1687, 2020. 1 +[62] Peng Yang, Shaogang Ren, Yang Zhao, and Ping Li. Calibrating cnns for few-shot meta learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2090–2099, 2022. 6 + +[63] Huaxiu Yao, Xian Wu, Zhiqiang Tao, Yaliang Li, Bolin Ding, Ruirui Li, and Zhenhui Li. Automated relational meta-learning. arXiv preprint arXiv:2001.00745, 2020. 6 +[64] Jaesik Yoon, Taesup Kim, Ousmane Dia, Sungwoong Kim, Yoshua Bengio, and Sungjin Ahn. Bayesian model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 6, 7 +[65] Yikai Zhang, Hui Qu, Chao Chen, and Dimitris Metaxas. Taming the noisy gradient: train deep neural networks with small batch sizes. In The Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI), 2019. 8 +[66] Dominic Zhao, Johannes von Oswald, Seijin Kobayashi, João Sacramento, and Benjamin F Grewe. Meta-learning via hypernetworks. 2020. 1, 4, 6, 8 \ No newline at end of file diff --git a/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/images.zip b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d45ca3f39eb446bb7126863566d6176d8a2926e --- /dev/null +++ b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4a7bcb12d0dc038f1fb816aa531cea02754f0cf50a6759a812a57d6961dfccd +size 504402 diff --git a/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/layout.json b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..361f4de81f38abf9ffccfb324f71cd787cd7f5e7 --- /dev/null +++ b/2023/Meta-Learning With a Geometry-Adaptive Preconditioner/layout.json @@ -0,0 +1,11936 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "type": "text", + "content": "Meta-Learning with a Geometry-Adaptive Preconditioner" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "text", + "content": "Suhyun Kang" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "text", + "content": ", Duhun Hwang" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "text", + "content": ", Moonjung Eo" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "text", + "content": ", Taesup Kim" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "text", + "content": ", Wonjong Rhee" + }, + { + "bbox": [ + 102, + 142, + 494, + 157 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 157, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 157, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 123, + 157, + 471, + 171 + ], + "type": "text", + "content": "1 Department of Intelligence and Information, Seoul National University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 171, + 446, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 171, + 446, + 185 + ], + "spans": [ + { + "bbox": [ + 147, + 171, + 446, + 185 + ], + "type": "text", + "content": "2 Graduate School of Data Science, Seoul National University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 185, + 401, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 185, + 401, + 200 + ], + "spans": [ + { + "bbox": [ + 192, + 185, + 401, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 192, + 185, + 401, + 200 + ], + "type": "text", + "content": " IPAI and AIIS, Seoul National University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 201, + 444, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 201, + 444, + 213 + ], + "spans": [ + { + "bbox": [ + 146, + 201, + 444, + 213 + ], + "type": "text", + "content": "{su_hyun, yelobean, eod87, taesup.kim, wrhee}@snu.ac.kr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 266, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 555 + ], + "type": "text", + "content": "Model-agnostic meta-learning (MAML) is one of the most successful meta-learning algorithms. It has a bi-level optimization structure where the outer-loop process learns a shared initialization and the inner-loop process optimizes task-specific weights. Although MAML relies on the standard gradient descent in the inner-loop, recent studies have shown that controlling the inner-loop's gradient descent with a meta-learning preconditioner can be beneficial. Existing preconditioners, however, cannot simultaneously adapt in a task-specific and path-dependent way. Additionally, they do not satisfy the Riemannian metric condition, which can enable the steepest descent learning with preconditioned gradient. In this study, we propose Geometry-Adaptive Preconditioned gradient descent (GAP) that can overcome the limitations in MAML; GAP can efficiently meta-learn a preconditioner that is dependent on task-specific parameters, and its preconditioner can be shown to be a Riemannian metric. Thanks to the two properties, the geometry-adaptive preconditioner is effective for improving the inner-loop optimization. Experiment results show that GAP outperforms the state-of-the-art MAML family and preconditioned gradient descent-MAML (PGD-MAML) family in a variety of few-shot learning tasks. Code is available at: https://github.com/Suhyun777/CVPR23-GAP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 598, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 288, + 684 + ], + "type": "text", + "content": "Meta-learning, or learning to learn, enables algorithms to quickly learn new concepts with only a small number of samples by extracting prior-knowledge known as meta-knowledge from a variety of tasks and by improving the generalization capability over the new tasks. Among the meta-learning algorithms, the category of optimization-based meta-learning [8, 17, 20, 21, 48] has been gaining popularity" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 241, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 545, + 301 + ], + "type": "text", + "content": "due to its flexible applicability over diverse fields including robotics [55, 59], medical image analysis [40, 54], language modeling [37, 42], and object detection [46, 61]. In particular, Model-Agnostic Meta-Learning (MAML) [20] is one of the most prevalent gradient-based meta-learning algorithms." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "text", + "content": "Many recent studies have improved MAML by adopting a Preconditioned Gradient Descent (PGD" + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "text", + "content": ") for inner-loop optimization [34, 36, 44, 49, 53, 57, 66]. In this paper, we collectively address PGD-based MAML algorithms as the PGD-MAML family. A PGD is different from the ordinary gradient descent because it performs a preconditioning on the gradient using a preconditioning matrix " + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "text", + "content": ", also called a preconditioner. A PGD-MAML algorithm meta-learns not only the initialization parameter " + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "text", + "content": " of the network but also the meta-parameter " + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "text", + "content": " of the preconditioner " + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 302, + 546, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "type": "text", + "content": "For the inner-loop optimization, " + }, + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "type": "text", + "content": " was kept static in most of the previous works (Figure 1(b)) [34, 36, 44, 57, 66]. Some of the previous works considered adapting the preconditioner " + }, + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "type": "text", + "content": " with the inner-step " + }, + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 422, + 547, + 529 + ], + "type": "text", + "content": " (Figure 1(c)) [49] and some others with the individual task (Figure 1(d)) [53]. They achieved performance improvement by considering individual tasks and inner-step, respectively, and showed that both factors were valuable. However, both factors have not been considered simultaneously in the existing studies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 529, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 686 + ], + "type": "text", + "content": "When a parameter space has a certain underlying structure, there exists a Riemannian metric corresponding the parameter space [3,4]. If the preconditioning matrix is the Riemannian metric, the preconditioned gradient is known to become the steepest descent on the parameter space [2-4,6,27]. An illustration of a toy example is shown in Figure 2. The optimization path of an ordinary gradient descent is shown in Figure 2(a). Compared to the ordinary gradient descent, a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can actually harm the optimization. For the example in Figure 2(b), the preconditioner affects the optimization into an undesirable direction and negatively affects the gradient descent. On the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1PGD in our work should not be confused with Projected Gradient Descent [13]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16080" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 72, + 149, + 144 + ], + "blocks": [ + { + "bbox": [ + 71, + 72, + 149, + 144 + ], + "lines": [ + { + "bbox": [ + 71, + 72, + 149, + 144 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 149, + 144 + ], + "type": "image", + "image_path": "22f94a546d5669996f508fd7a8de533e9ecf86714cfd0f4f740d739e3e8c590d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 147, + 127, + 156 + ], + "lines": [ + { + "bbox": [ + 89, + 147, + 127, + 156 + ], + "spans": [ + { + "bbox": [ + 89, + 147, + 127, + 156 + ], + "type": "text", + "content": "(a) MAML" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 152, + 74, + 242, + 144 + ], + "blocks": [ + { + "bbox": [ + 152, + 74, + 242, + 144 + ], + "lines": [ + { + "bbox": [ + 152, + 74, + 242, + 144 + ], + "spans": [ + { + "bbox": [ + 152, + 74, + 242, + 144 + ], + "type": "image", + "image_path": "fc08048b30068de72750ee3c66ee519927bf6ab0c8cf79c359d78bb633857f94.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 147, + 233, + 156 + ], + "lines": [ + { + "bbox": [ + 157, + 147, + 233, + 156 + ], + "spans": [ + { + "bbox": [ + 157, + 147, + 233, + 156 + ], + "type": "text", + "content": "(b) Non-adaptive " + }, + { + "bbox": [ + 157, + 147, + 233, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(\\phi)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": "Figure 1. Diagram of MAML and PGD-MAML family. For the inner-loop adaptation in each diagram, the dotted lines of the same color indicate that they use a common preconditioning matrix (preconditioner). (a) MAML adaptation: no preconditioner is used (i.e., " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\mathbf{I}" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": "). (b) " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(\\phi)" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": ": a constant preconditioner is used in the inner-loop where the preconditioner's meta-parameter " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": " is meta-learned. (c) " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(k; \\phi)" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": ": a constant preconditioner is used for each inner-step " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": ". Preconditioner for each step is meta-learned, but " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(k, \\phi)" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": " is not task-specific. (d) " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}}; \\phi)" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": ": a constant preconditioner is used for each task. Preconditioner for each task is meta-learned, but " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}}; \\phi)" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": " is not dependent on " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": ". (e) GAP adapts " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(\\theta_{\\tau, k}; \\phi)" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": ": a fully adaptive preconditioner is used where it is task-specific and path-dependent. Instead of saying 'dependent on " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": "', we specifically say it is path-dependent because the exact dependency is on the task-specific parameter set " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau, k}" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": " that is considerably more informative than " + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 162, + 547, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 82, + 336, + 144 + ], + "blocks": [ + { + "bbox": [ + 244, + 82, + 336, + 144 + ], + "lines": [ + { + "bbox": [ + 244, + 82, + 336, + 144 + ], + "spans": [ + { + "bbox": [ + 244, + 82, + 336, + 144 + ], + "type": "image", + "image_path": "46b74e423ecb4e495bc02b70428a0eecb8433c5a9499a669faeeb13e4bbe74dd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 147, + 323, + 156 + ], + "lines": [ + { + "bbox": [ + 254, + 147, + 323, + 156 + ], + "spans": [ + { + "bbox": [ + 254, + 147, + 323, + 156 + ], + "type": "text", + "content": "(c) Adaptive " + }, + { + "bbox": [ + 254, + 147, + 323, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(k;\\phi)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 338, + 83, + 430, + 144 + ], + "blocks": [ + { + "bbox": [ + 338, + 83, + 430, + 144 + ], + "lines": [ + { + "bbox": [ + 338, + 83, + 430, + 144 + ], + "spans": [ + { + "bbox": [ + 338, + 83, + 430, + 144 + ], + "type": "image", + "image_path": "038e361c7d389e45c4e7073230ae9930ebb75bdd33dac471b48952256502f58e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 146, + 423, + 156 + ], + "lines": [ + { + "bbox": [ + 343, + 146, + 423, + 156 + ], + "spans": [ + { + "bbox": [ + 343, + 146, + 423, + 156 + ], + "type": "text", + "content": "(d) Adaptive " + }, + { + "bbox": [ + 343, + 146, + 423, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(D_{\\tau}^{tr};\\phi)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 432, + 82, + 525, + 144 + ], + "blocks": [ + { + "bbox": [ + 432, + 82, + 525, + 144 + ], + "lines": [ + { + "bbox": [ + 432, + 82, + 525, + 144 + ], + "spans": [ + { + "bbox": [ + 432, + 82, + 525, + 144 + ], + "type": "image", + "image_path": "a5c6cd1ca0fac950bc4b3c6165c35dcab0dfd8582eb6825828d95786f0099240.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 147, + 517, + 156 + ], + "lines": [ + { + "bbox": [ + 437, + 147, + 517, + 156 + ], + "spans": [ + { + "bbox": [ + 437, + 147, + 517, + 156 + ], + "type": "text", + "content": "(e) Adaptive " + }, + { + "bbox": [ + 437, + 147, + 517, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(\\theta_{\\tau ,k};\\phi)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 67, + 268, + 154, + 337 + ], + "blocks": [ + { + "bbox": [ + 67, + 268, + 154, + 337 + ], + "lines": [ + { + "bbox": [ + 67, + 268, + 154, + 337 + ], + "spans": [ + { + "bbox": [ + 67, + 268, + 154, + 337 + ], + "type": "image", + "image_path": "4a17d7c689a2c2d9f27ee5d1d4ee36f32ecd5bf0dafc979e08fa11e9a17aa020.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 343, + 156, + 353 + ], + "lines": [ + { + "bbox": [ + 69, + 343, + 156, + 353 + ], + "spans": [ + { + "bbox": [ + 69, + 343, + 156, + 353 + ], + "type": "text", + "content": "(a) Gradient Descent (GD)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 175, + 268, + 261, + 337 + ], + "blocks": [ + { + "bbox": [ + 175, + 268, + 261, + 337 + ], + "lines": [ + { + "bbox": [ + 175, + 268, + 261, + 337 + ], + "spans": [ + { + "bbox": [ + 175, + 268, + 261, + 337 + ], + "type": "image", + "image_path": "23f7288c3df23d2178d36282242e1bced73b5cfd1cb10de0b9bc550a8e2ac384.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 343, + 264, + 362 + ], + "lines": [ + { + "bbox": [ + 180, + 343, + 264, + 362 + ], + "spans": [ + { + "bbox": [ + 180, + 343, + 264, + 362 + ], + "type": "text", + "content": "(b) Preconditioned GD (Non-Riemannian metric)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 120, + 371, + 206, + 440 + ], + "blocks": [ + { + "bbox": [ + 120, + 371, + 206, + 440 + ], + "lines": [ + { + "bbox": [ + 120, + 371, + 206, + 440 + ], + "spans": [ + { + "bbox": [ + 120, + 371, + 206, + 440 + ], + "type": "image", + "image_path": "bb7ee49f34060de67ab21dc136cd1251f87a3d5dcaae13f93e4a578395eeae29.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 446, + 203, + 465 + ], + "lines": [ + { + "bbox": [ + 129, + 446, + 203, + 465 + ], + "spans": [ + { + "bbox": [ + 129, + 446, + 203, + 465 + ], + "type": "text", + "content": "(c) Preconditioned GD (Riemannian metric)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 469, + 287, + 557 + ], + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 557 + ], + "type": "text", + "content": "Figure 2. A toy example for illustrating the effect of Riemannian metric (see Supplementary Section A). When the curvature of the parameter space is poorly conditioned, (a) gradient descent can suffer from the difficulty of finding the solution, (b) a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can suffer further, and (c) a preconditioned gradient descent with a preconditioner that is a Riemannian metric can perform better." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 578, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 685 + ], + "type": "text", + "content": "contrary, if the preconditioner is the Riemannian metric corresponding the parameter space, the preconditioned gradient descent can become the steepest descent and can exhibit a better optimization behavior as shown in Figure 2(c). While the Riemannian metric condition (i.e., positive definiteness) is a necessary condition for steepest descent learning, the existing studies on PGD-MAML family did not consider constraining preconditioners to satisfy the condition for Riemannian metric." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": "In this study, we propose a new PGD method named Geometry Aaptive Preconditioned gradient descent (GAP)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "text", + "content": "Specifically, GAP satisfies two desirable properties which have not been considered before. First, GAP's preconditioner " + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "text", + "content": " is a fully adaptive preconditioner that can adapt to the individual task (task-specific) and to the optimization-path (path-dependent). The full adaptation is made possible by having the preconditioner depend on the task-specific parameter " + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau,k}" + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "text", + "content": " (Figure 1(e)). Second, we prove that " + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "text", + "content": " is a Riemannian metric. To this end, we force the metaparameters of " + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "text", + "content": " to be positive definite. Thus, GAP guarantees the steepest descent learning on the parameter space corresponding to " + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 262, + 547, + 417 + ], + "type": "text", + "content": ". Owing to the two properties, GAP enables a geometry-adaptive learning in inner-loop optimization." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 418, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 513 + ], + "type": "text", + "content": "For the implementation of GAP, we utilize the Singular Value Decomposition (SVD) operation to come up with our preconditioner satisfying the desired properties. For the recently proposed large-scale architectures, computational overhead can be an important design factor and we provide a low-computational approximation, Approximate GAP, that can be proven to asymptotically approximate the operation of GAP." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 514, + 547, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 585 + ], + "type": "text", + "content": "To demonstrate the effectiveness of GAP, we empirically evaluate our algorithm on popular few-shot learning tasks; few-shot regression, few-shot classification, and few-shot cross-domain classification. The results show that GAP outperforms the state-of-the-art MAML family and PGDMAML family." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 586, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 545, + 608 + ], + "type": "text", + "content": "The main contributions of our study can be summarized as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 610, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 315, + 610, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 610, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 315, + 610, + 547, + 658 + ], + "type": "text", + "content": "- We propose a new preconditioned gradient descent method called GAP, where it learns a preconditioner that enables a geometry-adaptive learning in the inner-loop optimization." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 665, + 547, + 713 + ], + "type": "text", + "content": "- We prove that GAP's preconditioner has two desirable properties: (1) It is both task-specific and path-dependent (i.e., dependent on task-specific parameter " + }, + { + "bbox": [ + 315, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau, k}" + }, + { + "bbox": [ + 315, + 665, + 547, + 713 + ], + "type": "text", + "content": "). (2) It is a Riemannian metric." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16081" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 288, + 154 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 57, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 287, + 108 + ], + "type": "text", + "content": "- For large-scale architectures, we provide a low-computational approximation that can be theoretically shown to approximate the GAP method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 118, + 288, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 118, + 288, + 154 + ], + "spans": [ + { + "bbox": [ + 57, + 118, + 288, + 154 + ], + "type": "text", + "content": "- For popular few-shot learning benchmark tasks, we empirically show that GAP outperforms the state-of-the-art MAML family and PGD-MAML family." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 166, + 126, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 166, + 126, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 126, + 178 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 186, + 263, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 263, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 263, + 198 + ], + "type": "text", + "content": "2.1. Model-Agnostic Meta-Learning (MAML)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "text", + "content": "The goal of MAML [20] is to find the best initialization that the model can quickly adapt from, such that the model can perform well for a new task. MAML consists of two levels of main optimization processes: inner-loop and outer-loop optimizations. Consider the model " + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "text", + "content": ". For a task " + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\tau = \\{D_{\\tau}^{\\mathrm{tr}}, D_{\\tau}^{\\mathrm{val}}\\}" + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "text", + "content": " sampled from the task distribution " + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "inline_equation", + "content": "p(\\mathcal{T})" + }, + { + "bbox": [ + 47, + 204, + 288, + 289 + ], + "type": "text", + "content": ", the inner-loop optimization is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 287, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 287, + 327 + ], + "type": "interline_equation", + "content": "\\theta_ {\\tau , K} = \\theta_ {\\tau , 0} - \\alpha \\sum_ {k = 0} ^ {K - 1} \\nabla_ {\\theta_ {\\tau , k}} \\mathcal {L} _ {\\tau} ^ {\\mathrm {i n}} \\left(\\theta_ {\\tau , k}; D _ {\\tau} ^ {\\mathrm {t r}}\\right) s. t \\theta_ {\\tau , 0} = \\theta , \\tag {1}", + "image_path": "cbfef5b8e63ddcd59cf7a09a37257448e8a44169ca91a2f2fda753cbb25f37c4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau ,k}" + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "text", + "content": " is task-specific parameters for task " + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "text", + "content": " is the learning rate for inner-loop optimization, " + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\tau}^{\\mathrm{in}}" + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "text", + "content": " is the inner-loop's loss function, and " + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "text", + "content": " is the number of gradient descent steps. With " + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "inline_equation", + "content": "D_{\\tau_i}^{\\mathrm{val}}" + }, + { + "bbox": [ + 47, + 335, + 287, + 394 + ], + "type": "text", + "content": " in each task, we can define outer-loop optimization as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 95, + 401, + 287, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 401, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 95, + 401, + 287, + 419 + ], + "type": "interline_equation", + "content": "\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {2}", + "image_path": "3a8c88f584c84c49952be801c987bb5798e840c40bc848081b3a80be12d290cf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 427, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 287, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 427, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 427, + 287, + 452 + ], + "type": "text", + "content": " is the learning rate for outer-loop optimization, and " + }, + { + "bbox": [ + 47, + 427, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\tau}^{\\mathrm{out}}" + }, + { + "bbox": [ + 47, + 427, + 287, + 452 + ], + "type": "text", + "content": " is the outer-loop's loss function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 460, + 258, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 258, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 258, + 472 + ], + "type": "text", + "content": "2.2. Preconditioned Gradient Descent (PGD)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "type": "text", + "content": "PGD is a method that minimizes the empirical risk through a gradient update with a preconditioner that rescales the geometry of the parameter space. Given model parameters " + }, + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "type": "text", + "content": " and task " + }, + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\tau = \\{D_{\\tau}^{\\mathrm{tr}}, D_{\\tau}^{\\mathrm{val}}\\}" + }, + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "type": "text", + "content": ", we can define the preconditioned gradient update with a preconditioner " + }, + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 47, + 479, + 287, + 539 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 557, + 287, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 557, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 89, + 557, + 287, + 583 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta_ {\\tau , k + 1} = \\theta_ {\\tau , k} - \\alpha \\mathbf {P} \\nabla_ {\\theta_ {\\tau , k}} \\mathcal {L} _ {\\tau} \\left(\\theta_ {\\tau , k}; D _ {\\tau} ^ {\\mathrm {t r}}\\right) \\\\ k = 0, 1, \\dots \\text {a n d} \\theta_ {\\tau , 0} = \\theta \\\\ \\end{array}", + "image_path": "3bea9e06535b6d315a54fc80cef378d18b9c3ac3905d0f55cda69364021b5530.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\tau}(\\theta_{\\tau,k}; D_{\\tau}^{\\mathrm{tr}})" + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "content": " is the empirical loss for task " + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "content": " and parameters " + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau,k}" + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "content": ". Setting " + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\mathbf{I}" + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "content": " recovers Eq. (3) to the basic gradient descent (GD). Choice of " + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "content": " for exploiting the second-order information includes the inverse Fisher information matrix " + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{-1}" + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "content": " which leads to the natural gradient descent (NGD) [4]; the inverse Hessian " + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{-1}" + }, + { + "bbox": [ + 47, + 594, + 288, + 714 + ], + "type": "text", + "content": " which corresponds to the Newton's method [31]; and the diagonal matrix estimation with the past gradients which results in the adaptive gradient methods [18, 28]. They often reduce the effect of pathological curvature and speed up the optimization [5]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 531, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 531, + 84 + ], + "type": "text", + "content": "2.3. Unfolding: reshaping a tensor into a matrix" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "text", + "content": "In this study, the concept of unfolding is used to transform the gradient tensor of convolutional kernels into a matrix form. Tensor unfolding, also known as matricization or flattening, is the process of reshaping the elements of an " + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "text", + "content": "-dimensional tensor " + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{I_1 \\times \\dots \\times I_N}" + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "text", + "content": " into a matrix [29]. The mode-" + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "text", + "content": " unfolding of an " + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "text", + "content": "-dimensional tensor " + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{I_1 \\times \\dots \\times I_N}" + }, + { + "bbox": [ + 304, + 90, + 547, + 174 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 178, + 545, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 178, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 319, + 178, + 545, + 201 + ], + "type": "interline_equation", + "content": "\\mathbf {X} \\xrightarrow [ \\text {m o d e - n u n f o l d i n g} ]{} \\mathbf {X} _ {[ n ]} \\in \\mathbb {R} ^ {I _ {n} \\times I _ {M}}, \\text {w h e r e} I _ {M} = \\prod_ {k \\neq n} I _ {k} \\tag {4}", + "image_path": "d6c564a24f8969f115debe5a1c0492b63913b7d759251253b0963084e8cfd173.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "text", + "content": "For example, the weight tensor of a convolutional layer is represented as a 4-D tensor " + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "inline_equation", + "content": "(\\mathbf{W} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times C_{\\mathrm{in}} \\times k_h \\times k_w})" + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "text", + "content": ", where it is composed of kernels and it can be unfolded into a matrix as one of the following four forms: (1) " + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{[1]} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times (C_{\\mathrm{in}}k_hk_w)}" + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "text", + "content": ", (2) " + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{[2]} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times (C_{\\mathrm{out}}k_hk_w)}" + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "text", + "content": ", (3) " + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{[3]} \\in \\mathbb{R}^{k_h \\times (C_{\\mathrm{out}}C_{\\mathrm{in}}k_w)}" + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "text", + "content": ", (4) " + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{[4]} \\in \\mathbb{R}^{k_w \\times (C_{\\mathrm{out}}C_{\\mathrm{in}}k_h)}" + }, + { + "bbox": [ + 304, + 208, + 547, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 289, + 430, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 430, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 430, + 300 + ], + "type": "text", + "content": "2.4. Riemannian manifold" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": "An " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": "-dimensional Riemannian manifold is defined by a manifold " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": " and a Riemannian metric " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "g: \\mathcal{M} \\to \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": ", which is a smooth function from each point " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{M}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": " to a positive definite matrix [32]. The metric " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "g(\\pmb{x})" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": " defines the inner product of two tangent vectors for each point of the manifold " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle: \\mathcal{T}_{\\pmb{x}}\\mathcal{M} \\times \\mathcal{T}_{\\pmb{x}}\\mathcal{M} \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\pmb{x}}\\mathcal{M}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": " is the tangent space of " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\pmb{v}, \\pmb{w} \\in \\mathcal{T}_{\\pmb{x}}\\mathcal{M}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": ", the inner product can be expressed " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\langle \\pmb{v}, \\pmb{w} \\rangle = \\pmb{v}^T g(\\pmb{x})\\pmb{w}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": ". A Riemannian manifold can be characterized by the curvature of the curves defined by a metric. The curvature of a Riemannian manifold can be computed at each point of the curves, while some manifolds have curvatures of a constant value. For example, the unit sphere " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": " has constant positive curvature of " + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "inline_equation", + "content": "+1" + }, + { + "bbox": [ + 304, + 307, + 545, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 472, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 388, + 487 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 493, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 547, + 588 + ], + "type": "text", + "content": "In this section, we propose a new preconditioned gradient descent method called GAP in the MAML framework. In Section 3.1, we introduce GAP in the inner-loop optimization and describe how to meta-train GAP in the outer-loop optimization. In Section 3.2, we prove that GAP has two desirable properties. In Section 3.3, we provide a low-computational approximation that can be useful for large-scale architectures." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 595, + 523, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 523, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 523, + 608 + ], + "type": "text", + "content": "3.1. GAP: Geometry-Adaptive Preconditioner" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 613, + 440, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 613, + 440, + 625 + ], + "spans": [ + { + "bbox": [ + 306, + 613, + 440, + 625 + ], + "type": "text", + "content": "3.1.1 Inner-loop optimization" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": "We consider an " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": "-layer neural network " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\mathbf{W}^1, \\dots, \\mathbf{W}^l, \\dots, \\mathbf{W}^L\\}" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": ". In the standard MAML with task " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\tau \\sim p(\\mathcal{T})" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": ", each " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^l" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": " is adapted with the gradient update as below:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 334, + 686, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 686, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 334, + 686, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {\\tau , K} ^ {l} \\leftarrow \\mathbf {W} _ {\\tau , 0} ^ {l} - \\alpha \\cdot \\sum_ {k = 0} ^ {K - 1} \\mathbf {G} _ {\\tau , k} ^ {l} s. t \\mathbf {W} _ {\\tau , 0} ^ {l} = \\mathbf {W} ^ {l}, \\tag {5}", + "image_path": "87d0d426f87959325e30c04f72bef82370bf85a6a3469c4cc83b7d05eb89a877.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16082" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 99, + 288, + 299 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "Algorithm 1 Geometry-Adaptive Preconditioned gradient descent (GAP)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "lines": [ + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "p(\\mathcal{T})" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " : distribution over tasks \nRequire: " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta_{1},\\beta_{2}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " : learning rates 1: Randomly initialize " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\mathbf{W}^1,\\dots ,\\mathbf{W}^L\\}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " \n2: Initialize " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\mathbf{M}^{1},\\dots ,\\mathbf{M}^{L}\\}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " as identity matrix. \n3: while not converged do \n4: Sample a batch of tasks " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i\\sim p(\\mathcal{T})" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " \n5: for all " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathcal{T}_i" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " do \n6: for inner-loop step " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " do \n7: for layer " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " do \n8: Compute " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau ,k}^{l} = \\nabla_{\\mathbf{W}_{\\tau ,k}^{l}}\\mathcal{L}_{\\tau}^{\\mathrm{in}}(\\theta_{\\tau ,k};D_{\\tau}^{\\mathrm{tr}})" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "D_{\\tau}^{\\mathrm{tr}}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " \n9: Reshape " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " via Eq. (6) \n10: Transform " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^l" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " via Eq. (8) \n11: Reshape " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " back to the original form of gradient tensor, " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " \n12: Compute " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " -layer adapted weight: " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\tau ,k + 1}^{l} = \\mathbf{W}_{\\tau ,k}^{l} - \\alpha \\cdot \\tilde{\\mathbf{G}}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " \n13: end for \n14: end for \n15: Compute " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " by evaluating " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\tau}^{\\mathrm{out}}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " w.r.t " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "D_{\\tau}^{\\mathrm{val}}" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " \n16: end for \n17: Update the weights and meta parameters: \n18: " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta -\\beta_1\\nabla_\\theta \\sum_{\\tau \\in \\mathcal{T}}\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " \n19: " + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\phi \\leftarrow \\phi -\\beta_{2}\\nabla_{\\phi}\\sum_{\\tau \\in \\mathcal{T}}\\mathcal{L}_{\\tau}^{\\mathrm{out}}(\\theta_{\\tau ,K};D_{\\tau}^{\\mathrm{val}})" + }, + { + "bbox": [ + 47, + 99, + 288, + 299 + ], + "type": "text", + "content": " \n20: end while" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau,k}^{l} = \\nabla_{\\mathbf{W}_{\\tau,k}^{l}} \\mathcal{L}_{\\tau}^{\\mathrm{in}}(\\theta_{\\tau,k}; D_{\\tau}^{\\mathrm{tr}})" + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "text", + "content": " is the gradient with respect to " + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\tau,k}^{l}" + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "text", + "content": " is the learning rate for inner-loop optimization. In the GAP, we first use the mode-1 unfolding to reshape the gradient tensor into a matrix form (see Section 2.3). For a convolutional layer (i.e., " + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau,k}^{l} \\in \\mathbb{R}^{C_{\\mathrm{out}} \\times C_{\\mathrm{in}} \\times k \\times k}" + }, + { + "bbox": [ + 46, + 321, + 289, + 399 + ], + "type": "text", + "content": "), we reshape the gradient tensor as below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 409, + 287, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 409, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 59, + 409, + 287, + 440 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {\\tau , k} ^ {l} \\xrightarrow [ \\text {m o d e - 1} ]{} \\left\\{ \\begin{array}{l l} \\mathbf {G} _ {\\tau , k} ^ {l} \\in \\mathbb {R} ^ {C _ {\\mathrm {o u t}} \\times C _ {\\mathrm {i n}} k ^ {2}} & \\text {i f} C _ {\\mathrm {o u t}} \\leq C _ {\\mathrm {i n}} k ^ {2} \\\\ \\mathbf {G} _ {\\tau , k} ^ {l} \\in \\mathbb {R} ^ {C _ {\\mathrm {i n}} k ^ {2} \\times C _ {\\mathrm {o u t}}} & \\text {i f} C _ {\\mathrm {i n}} k ^ {2} < C _ {\\mathrm {o u t}}, \\end{array} \\right. \\tag {6}", + "image_path": "7cf4a1855079e37556d54d6296b70a077b03689920d454aa5041e290b894e712.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "inline_equation", + "content": "(\\mathbf{G}_{\\tau ,k}^{l})_{[1]}" + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "text", + "content": " for the notational brevity. Note that we chose mode-1 unfolding because it performs best among the four unfolding forms as shown in Table 1. For a linear layer in a matrix form, there is no need for an unfolding. Second, we transform the singular values of the gradient matrix using additional meta parameters " + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\mathbf{M}^l\\}_{l = 1}^L" + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "text", + "content": " . The meta parameter " + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^l" + }, + { + "bbox": [ + 46, + 451, + 289, + 547 + ], + "type": "text", + "content": " are diagonal matrices with positive elements defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\mathbf {M} ^ {l} = \\left\\{ \\begin{array}{l l} \\operatorname {d i a g} \\left(\\operatorname {S p} \\left(m _ {1} ^ {l}\\right), \\dots , \\operatorname {S p} \\left(m _ {C _ {\\text {o u t}}} ^ {l}\\right)\\right) & \\text {i f} C _ {\\text {o u t}} \\leq C _ {\\text {i n}} k ^ {2} \\\\ \\operatorname {d i a g} \\left(\\operatorname {S p} \\left(m _ {1} ^ {l}\\right), \\dots , \\operatorname {S p} \\left(m _ {C _ {\\text {i n}} k ^ {2}} ^ {l}\\right)\\right) & \\text {i f} C _ {\\text {i n}} k ^ {2} < C _ {\\text {o u t}} \\end{array} , \\right. \\tag {7}", + "image_path": "0057e608b85a6c791492934e3353215671f1c5417b0de7e8483c15097f3e54bc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "inline_equation", + "content": "m_i^l\\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{Sp}(x) = \\frac{1}{2}\\cdot \\log (1 + \\exp (2*x))" + }, + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "text", + "content": " . They are applied to the gradient matrix as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 628, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 628, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {G}} _ {\\tau , k} ^ {l} = \\mathbf {U} _ {\\tau , k} ^ {l} \\left(\\mathbf {M} ^ {l} \\cdot \\boldsymbol {\\Sigma} _ {\\tau , k} ^ {l}\\right) \\mathbf {V} _ {\\tau , k} ^ {l}, \\tag {8}", + "image_path": "7b1d6fb4a3c6ed6432fe309104f1f29298be36b21d25faf9116edb87f95036d4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau,k}^{l} = \\mathbf{U}_{\\tau,k}^{l}\\pmb{\\Sigma}_{\\tau,k}^{l}\\mathbf{V}_{\\tau,k}^{l}" + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "text", + "content": " is the singular value decomposition (SVD) of " + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\tau,k}^{l}" + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "text", + "content": ". Finally, we reshape " + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_{\\tau,k}^{l}" + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "text", + "content": " back to its original gradient tensor form " + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_{\\tau,k}^{l}" + }, + { + "bbox": [ + 46, + 655, + 289, + 714 + ], + "type": "text", + "content": " (i.e., inverse unfolding). The resulting preconditioned gradient descent of GAP" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 314, + 71, + 537, + 101 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 537, + 101 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 537, + 101 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 537, + 101 + ], + "type": "table", + "html": "
mode-1mode-2mode-3mode-4
54.86 ± 0.8553.02 ± 0.8751.23 ± 0.7651.45 ± 0.77
", + "image_path": "08b7499f3b2c145998995c12708492f12709c46f15481919a0d1ed8208dc089b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 104, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 305, + 104, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 305, + 104, + 547, + 137 + ], + "type": "text", + "content": "Table 1. Performance comparison of unfolding types. We performed the experiment with 5-way 1-shot on mini-[ImageNet] and used the standard Conv-4 backbone." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 150, + 403, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 150, + 403, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 150, + 403, + 163 + ], + "type": "text", + "content": "becomes the following:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 334, + 170, + 545, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 170, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 334, + 170, + 545, + 199 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {\\tau , K} ^ {l} \\leftarrow \\mathbf {W} _ {\\tau , 0} ^ {l} - \\alpha \\cdot \\sum_ {k = 0} ^ {K - 1} \\tilde {\\mathbf {G}} _ {\\tau , k} ^ {l} s. t \\mathbf {W} _ {\\tau , 0} ^ {l} = \\mathbf {W} ^ {l}, \\tag {9}", + "image_path": "0344c317352f26fa15fb37a7255634c703e212a35898df1b1f8735dd376368fe.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 209, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 209, + 545, + 236 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_{\\tau ,k}^{l}" + }, + { + "bbox": [ + 305, + 209, + 545, + 236 + ], + "type": "text", + "content": " is the preconditioned gradient based on the meta parameters " + }, + { + "bbox": [ + 305, + 209, + 545, + 236 + ], + "type": "inline_equation", + "content": "\\phi" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 251, + 443, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 443, + 263 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 443, + 263 + ], + "type": "text", + "content": "3.1.2 Outer-loop optimization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "type": "text", + "content": "For outer-loop optimization, GAP follows the typical process of MAML. Unlike MAML, however, GAP meta-learns two meta parameter sets " + }, + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 315, + 545, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 315, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 347, + 315, + 545, + 329 + ], + "type": "interline_equation", + "content": "\\theta \\leftarrow \\theta - \\beta_ {1} \\nabla_ {\\theta} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {10}", + "image_path": "da93080f169a7803b26573e3c79bc60897a53af1c924ad992184157319bfafe4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 347, + 331, + 545, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 331, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 347, + 331, + 545, + 346 + ], + "type": "interline_equation", + "content": "\\phi \\leftarrow \\phi - \\beta_ {2} \\nabla_ {\\phi} \\mathbb {E} _ {\\tau} \\left[ \\mathcal {L} _ {\\tau} ^ {\\text {o u t}} \\left(\\theta_ {\\tau , K}; D _ {\\tau} ^ {\\text {v a l}}\\right) \\right], \\tag {11}", + "image_path": "e24e7d43ab364470412aaf1d633dae17ce3932f7d0ef4235da2b0a893bf536ca.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "text", + "content": " are the learning rates for the outer-loop optimization. We initialize " + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^l" + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "text", + "content": " as an identity matrix for all " + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 354, + 547, + 391 + ], + "type": "text", + "content": ". The training procedure is provided in Algorithm 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 398, + 462, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 462, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 462, + 411 + ], + "type": "text", + "content": "3.2. Desirable properties of GAP" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 416, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 441 + ], + "type": "text", + "content": "In this section, we prove that GAP's preconditioner " + }, + { + "bbox": [ + 305, + 416, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 305, + 416, + 545, + 441 + ], + "type": "text", + "content": " satisfies two desirable properties." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "content": "Theorem 1. Let " + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_{\\tau,k}^{l} \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "content": " be the 'l-layer " + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "content": "-th inner-step' gradient matrix transformed by meta parameter " + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^l" + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "content": " for task " + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "content": ". Then preconditioner " + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{GAP}" + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "content": " induced by " + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}_{\\tau,k}^{l}" + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "content": " is a Riemannian metric and depends on the task-specific parameters " + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau,k}" + }, + { + "bbox": [ + 305, + 448, + 547, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 515, + 547, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 547, + 551 + ], + "type": "text", + "content": "The proof and closed form of " + }, + { + "bbox": [ + 305, + 515, + 547, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 305, + 515, + 547, + 551 + ], + "type": "text", + "content": " are provided in Supplementary Section B. The following two properties emerge from the theorem." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "content": "Property 1. Dependency on task-specific parameters: Theorem 1 formally shows that " + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "content": " depends on the task-specific parameters " + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau,k}" + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "content": ". While the previous studies considered a non-adaptive preconditioner " + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(\\phi)" + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "content": " [34, 36, 44, 49, 57, 66] and a partially adaptive preconditioner " + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(k;\\phi)" + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "content": " [49] or " + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(D_{\\tau}^{\\mathrm{tr}};\\phi)" + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "content": " [53], " + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "content": " can be considered to be the most advanced adaptive preconditioner because it is fully adaptive (i.e., task-specific and path-dependent) by being dependent on " + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau,k}" + }, + { + "bbox": [ + 304, + 560, + 547, + 669 + ], + "type": "text", + "content": " as shown in Figure 1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Property 2. Riemannian metric: If the parameter space has a certain underlying structure, the ordinary gradient of a function " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": " does not represent its steepest direction [4]." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16083" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "To define the steepest direction on the parameter space, we need a Riemannian metric " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "g(\\boldsymbol{w})" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ", which is a positive-definite matrix defined for each parameter " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ". A Riemannian metric defines the steepest descent direction by " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "-g(\\boldsymbol{w})^{-1}\\nabla \\mathcal{L}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " [4]. If a preconditioning matrix is a Riemannian metric, it defines the geometry of the underlying structure and enables steepest descent learning. Because we prove that " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " is a Riemannian metric for each parameter in Theorem 1, " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " is theoretically guaranteed to enable steepest descent learning on its corresponding parameter space. " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " consists of two factors, a unitary matrix of the inner-loop gradient " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\tau,k}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " and a meta-parameter " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " enables us to reflect the shared geometry information across the tasks. Task-specific and path-dependent geometry information can be reflected in the metric through " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\tau,k}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ". Two factors allow our Riemannian metric to have higher function complexity than a constant metric. For example, we can consider various structures other than a unit-sphere that corresponds to the constant metric of " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "+1" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": ". Even though " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " is guaranteed to be a Riemannian metric, it is crucial that the meta-learned " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " corresponds to the true parameter space or at least " + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": " is close enough to be useful. We will discuss this issue in Section 5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 343, + 287, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 287, + 369 + ], + "type": "text", + "content": "3.3. Approximate GAP: a low-computational approximation of GAP" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 373, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 483 + ], + "type": "text", + "content": "As presented in Section 3.1, GAP uses an SVD operation. The SVD operation can be burdensome for large-scale networks because it implies that the computational cost can be significantly increased. In recent studies, the use of a large-scale architecture has been emphasized as the key factor for improving the performance of meta-learning [25]. To make use of GAP for large-scale architectures without causing a computational problem, we provide an efficient approximation, named Approximate GAP, under an assumption." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 514 + ], + "type": "text", + "content": "Assumption 1. The elements of the gradient matrix follow an i.i.d. normal distribution with zero means." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 606 + ], + "type": "text", + "content": "Following [39, 60], we adopt the assumption to have the gradient matrix become orthogonal as " + }, + { + "bbox": [ + 46, + 522, + 287, + 606 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 522, + 287, + 606 + ], + "type": "text", + "content": " increases. Although the utilization of the assumption is a limiting factor, we empirically confirmed that the row vectors of the gradient matrix are indeed asymptotically orthogonal as " + }, + { + "bbox": [ + 46, + 522, + 287, + 606 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 522, + 287, + 606 + ], + "type": "text", + "content": " increases (see Figure 3). For the assumption, our approximation can be established as the following." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "text", + "content": "Theorem 2. Let " + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "text", + "content": " be a gradient matrix and " + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}" + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "text", + "content": " be the gradient transformed by a meta parameter " + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "text", + "content": ". Under the Assumption 1, as " + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "text", + "content": " becomes large, " + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{G}}" + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "text", + "content": " asymptotically becomes equivalent to " + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{MG}" + }, + { + "bbox": [ + 46, + 613, + 287, + 663 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 144, + 671, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 671, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 144, + 671, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {G}} \\cong \\mathbf {M G} \\tag {12}", + "image_path": "9010cb425779d35d3b1a21908497aea5152b24b7b574aeb8b4db05e37bc7b4fb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": "Note that we have chosen the larger dimension of the gradient matrix as " + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": " when reshaping with Eq. (6). We provide" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 353, + 72, + 500, + 182 + ], + "blocks": [ + { + "bbox": [ + 353, + 72, + 500, + 182 + ], + "lines": [ + { + "bbox": [ + 353, + 72, + 500, + 182 + ], + "spans": [ + { + "bbox": [ + 353, + 72, + 500, + 182 + ], + "type": "image", + "image_path": "31919d55723fcabb9ab7563c5f8fe11f0bbc340a025fd60acea68c3f1cb5061e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 186, + 547, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 186, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 547, + 209 + ], + "type": "text", + "content": "Figure 3. The average of cosine similarity between row vectors of gradient matrix as " + }, + { + "bbox": [ + 305, + 186, + 547, + 209 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 186, + 547, + 209 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 219, + 535, + 260 + ], + "blocks": [ + { + "bbox": [ + 315, + 219, + 535, + 260 + ], + "lines": [ + { + "bbox": [ + 315, + 219, + 535, + 260 + ], + "spans": [ + { + "bbox": [ + 315, + 219, + 535, + 260 + ], + "type": "table", + "html": "
Algorithmtrain time (secs)test time (secs)GPU-memory (MB)
MAML35796.6152.6610185
GAP50916.63120.4410279
Approximate GAP36684.7253.6210217
", + "image_path": "60001991cec2eae7bce66791bcda83002be32fb879b0ed2d5622087c54159c82.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 262, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 547, + 307 + ], + "type": "text", + "content": "Table 2. Comparison of training time, GPU memory, and test time for MAML, GAP, and Approximate GAP. We performed the experiment with 5-way 1-shot on mini-ImageNet and used 600 tasks in the test. We used the standard Conv-4 backbone." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": "the proof in Supplementary Section B. This approximation has a clear trade-off between scalability and adaptiveness. Approximate GAP efficiently reduces the computational cost as shown in Table 2, whereas the preconditioner " + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 316, + 547, + 425 + ], + "type": "text", + "content": " becomes not adaptive but constant. However, Approximate GAP still guarantees the preconditioner to be a Riemannian metric. As shown in Table 4 & 5, Approximate GAP incurs a slight performance drop but it still achieves a high performance owing to the Riemannian metric property." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 434, + 388, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 388, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 388, + 448 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 454, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 454, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 454, + 547, + 503 + ], + "type": "text", + "content": "In this section, we show the superiority of GAP by comparing it with the state-of-the-art PGD-MAML family and the MAML family. Hyper-parameter setups used in our experiments can be found in Supplementary Section C.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 510, + 422, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 422, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 422, + 522 + ], + "type": "text", + "content": "4.1. Few-shot regression" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "content": "Datasets and experimental setup. The goal of few-shot regression is to fit an unknown target function for the given " + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "content": " sample points from the function. For the evaluation of few-shot regression, we use the sinusoid regression benchmark [20]. In this benchmark, sinusoid is used as the target function. Each task has a sinusoid " + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "inline_equation", + "content": "y(x) = A\\sin (\\omega x + b)" + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "content": " as the target function, where the parameter values are within the following range: amplitude " + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "inline_equation", + "content": "A\\in [0.1,5.0]" + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "content": ", frequency " + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "inline_equation", + "content": "\\omega \\in [0.8,1.2]" + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "content": ", and phase " + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "inline_equation", + "content": "b\\in [0,\\pi ]" + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "content": ". For each task, input data point " + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "content": " is sampled from " + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "inline_equation", + "content": "[-5.0,5.0]" + }, + { + "bbox": [ + 304, + 528, + 547, + 685 + ], + "type": "text", + "content": ". In the experiment, we use a simple Multi-Layer Perceptron (MLP), following the setting in [20]. The details of the architecture are provided in Supplementary Section C.2.1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Results. We evaluate GAP and compare it with MAML family and PGD-MAML family on a regression task. As" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16084" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 274, + 156 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 274, + 156 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 274, + 156 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 274, + 156 + ], + "type": "table", + "html": "
Algorithm5-shot10-shot20-shot
MAML [20]1.13 ± 0.180.77 ± 0.110.48 ± 0.08
Meta-SGD† [36]0.90 ± 0.160.53 ± 0.090.31 ± 0.05
MT-Net [34]0.76 ± 0.090.49 ± 0.050.33 ± 0.04
ALFA [9]0.92 ± 0.190.62 ± 0.160.34 ± 0.07
L2F [10]0.71 ± N/A0.37 ± N/A0.16 ± N/A
PAMELA† [49]0.54 ± 0.060.41 ± 0.040.17 ± 0.03
MeTAL [8]0.74 ± 0.180.44 ± 0.110.21 ± 0.06
GAP†0.16 ± 0.040.04 ± 0.010.01 ± 0.01
", + "image_path": "fbd668e3fb1a36a0460ccab608a6ecbfe2a656ff081258005b6e9b0c7e10ea17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 158, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 202 + ], + "type": "text", + "content": "Table 3. Few-shot regression for the sinusoid regression benchmark with a 2-layer MLP backbone. We report MSE ± 95% confidence intervals(ci) for 600 tasks following the setup in [20].† denotes PGD-MAML family." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 213, + 288, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 288, + 274 + ], + "type": "text", + "content": "shown in Table 3, GAP consistently achieves the lowest mean squared error (MSE) scores, with the lowest confidence intervals in all three cases. The performance of GAP is improved by " + }, + { + "bbox": [ + 46, + 213, + 288, + 274 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 46, + 213, + 288, + 274 + ], + "type": "text", + "content": " on 10-shot and " + }, + { + "bbox": [ + 46, + 213, + 288, + 274 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 46, + 213, + 288, + 274 + ], + "type": "text", + "content": " on 20-shot compared to the performance of state-of-the-art algorithms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 281, + 174, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 174, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 174, + 293 + ], + "type": "text", + "content": "4.2. Few-shot classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 300, + 290, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 290, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 290, + 479 + ], + "type": "text", + "content": "Datasets and experimental setup. For the few-shot classification, we evaluate two benchmarks: (1) miniImageNet [56]; this dataset has 100 classes and it is a subset of ImageNet [16], and we use the same split as in [51], with 64, 16 and 20 classes for train, validation and test, respectively. (2) tiered-ImageNet [52]; this is also a subset of ImageNet with 608 classes grouped into 34 high-level categories, and divided into 20, 6 and 8 for train, validation, and test, respectively. For all the experiments, our model follows the standard Conv-4 backbone used in [56]. The details of the architecture are provided in Supplementary Section C.2.2. Following the experimental protocol in [20], we use 15 samples per class in the query-set to compute the meta gradients. In meta training and meta testing, the inner-loop optimization is updated in five and ten steps, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": "Results. Table 4 & 5 present the performance of GAP, the state-of-the-art PGD-MAML family, and the state-of-the-art MAML-family on mini-[ImageNet and tiered-[ImageNet under two typical settings: 5-way 1-shot and 5-way 5-shot. The GAP outperforms all of the previous PGD-MAML family and MAML family. Compared to the state-of-the-art MAML family, GAP improves the performance with a quite significant margin for both mini-[ImageNet and tiered-[ImageNet datasets. Compared to the state-of-the-art PGD-MAML family, GAP shows that the 1- and 5-shot accuracy can be increased by " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": " on mini-[ImageNet dataset, and by " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "0.68\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": " on tiered-[ImageNet dataset, respectively. We also evaluated Approximate GAP that is introduced in Section 3.3. The results show that the approximated version can perform comparably to the original GAP. Although Approximate GAP shows slightly lower accuracies than the original, its performance is superior to most of the existing algorithms because of its Riemannian metric property." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 69, + 533, + 423 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 533, + 423 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 533, + 423 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 533, + 423 + ], + "type": "table", + "html": "
Algorithm5-way 1-shot5-way 5-shot
MAML [20]47.89 ± 1.2064.59 ± 0.88
Meta-SGD† [36]50.47 ± 1.8764.00 ± 0.90
BMAML [64]53.80 ± 1.4664.23 ± 0.69
ANIL [48]46.70 ± 0.4061.50 ± 0.50
LLAMA [23].49.40 ± 1.83N/A
PLATIPUS [21]50.13 ± 1.86-
T-net [34]50.86 ± 1.82N/A
MT-net [34]51.70 ± 1.84N/A
MAML++ [7]52.15 ± 0.2668.32 ± 0.44
iMAML-HF [50]49.30 ± 1.88N/A
WarpGrad [22]52.30 ± 0.9068.40 ± 0.60
MC1† [44]53.74 ± 0.8468.01 ± 0.73
MC2† [44]54.08 ± 0.8867.99 ± 0.73
MH-C† [66]48.64 ± 0.3364.52 ± 0.51
MH† [66]49.41 ± 0.9667.16 ± 0.42
BOIL [43]49.61 ± 0.1666.46 ± 0.37
ARML [63]50.42 ± 1.7964.12 ± 0.90
ALFA [9]50.58 ± 0.5169.12 ± 0.47
L2F [10]52.10 ± 0.5069.38 ± 0.46
ModGrad† [53]53.20 ± 0.8669.17 ± 0.69
PAMELA† [49]53.50 ± 0.8970.51 ± 0.67
SignMAML [19]42.90 ± 1.5060.70 ± 0.70
CTML [45]50.47 ± 1.8364.15 ± 0.90
MeTAL [8]52.63 ± 0.3770.52 ± 0.29
ECML [24]48.94 ± 0.8065.26 ± 0.67
Sharp-MAML_up [1]49.56 ± N/A63.06 ± N/A
Sharp-MAML_low [1]49.72 ± N/A63.18 ± N/A
Sharp-MAMLboth [1]50.28 ± N/A65.04 ± N/A
FBM [62]50.62 ± 1.7964.78 ± 0.35
CxGrad [33]51.80 ± 0.4669.82 ± 0.42
HyperMAML [47]51.84 ± 0.5766.29 ± 0.43
EEML [35]52.42 ± 1.7568.40 ± 0.95
MH-O† [66]52.50 ± 0.6167.76 ± 0.34
Sparse-MAML† [57]50.35 ± 0.3967.03 ± 0.74
Sparse-ReLU-MAML† [57]50.39 ± 0.8964.84 ± 0.46
Sparse-MAML+† [57]51.04 ± 0.5967.03 ± 0.74
Approximate GAP†53.52 ± 0.8870.75 ± 0.67
GAP†54.86 ± 0.8571.55 ± 0.61
", + "image_path": "3b1cb1c7b02d9f29066bd155430a911ddd64285053312f4738a51ad0eb989752.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "type": "text", + "content": "Table 4. 5-way few-shot classification accuracy " + }, + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "type": "text", + "content": " on miniImageNet with a Conv-4 backbone. We report mean " + }, + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\pm 95\\%" + }, + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "type": "text", + "content": " confidence intervals(ci) for 600 tasks according to [20]. " + }, + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 426, + 547, + 470 + ], + "type": "text", + "content": " denotes PGD-MAML family." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 472, + 533, + 643 + ], + "blocks": [ + { + "bbox": [ + 317, + 472, + 533, + 643 + ], + "lines": [ + { + "bbox": [ + 317, + 472, + 533, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 533, + 643 + ], + "type": "table", + "html": "
Algorithm5-way 1-shot5-way 5-shot
Meta-SGD†[36]50.92 ± 0.9369.28 ± 0.80
MAML [20]51.70 ± 1.8070.30 ± 1.80
MT-net [34]51.95 ± 1.83N/A
WarpGrad [22]57.20 ± 0.9074.10 ± 0.70
BOIL [43].48.58 ± 0.2769.37 ± 0.12
ALFA [9]53.16 ± 0.4970.54 ± 0.46
L2F [10]54.40 ± 0.5073.34 ± 0.44
ARML [63]52.91 ± 1.83N/A
PAMELA†[49]54.81 ± 0.8874.39 ± 0.71
Sparse-ReLU-MAML†[57]53.18 ± 0.5269.06 ± 0.28
Sparse-MAML†[57]53.47 ± 0.5368.83 ± 0.65
Sparse-MAML+†[57]53.91 ± 0.6769.92 ± 0.21
MeTAL [8]54.34 ± 0.3170.40 ± 0.21
CxGrad [33]55.55 ± 0.4673.55 ± 0.41
ECML [24]47.34 ± 0.8864.77 ± 0.75
Approximate GAP†56.86 ± 0.9174.41 ± 0.72
GAP†57.60 ± 0.9374.90 ± 0.68
", + "image_path": "f48b7ebe878150bbeb66f33a2582bf6ac729d8894b7955bdef8a5fa9ada7f212.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 645, + 547, + 690 + ], + "lines": [ + { + "bbox": [ + 305, + 645, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 305, + 645, + 547, + 690 + ], + "type": "text", + "content": "Table 5. 5-way few-shot classification accuracy (%) on tiered-ImageNet dataset with a Conv-4 backbone. We report mean ± 95% confidence intervals(ci) for 600 tasks according to [20].† denotes PGD-MAML family." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16085" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 509, + 237 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 509, + 237 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 509, + 237 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 509, + 237 + ], + "type": "table", + "html": "
Algorithmtiered-ImageNetCUBCars
1-shot5-shot1-shot5-shot1-shot5-shot
MAML [20]51.61 ± 0.2065.76 ± 0.2740.51 ± 0.0853.09 ± 0.1633.57 ± 0.1444.56 ± 0.21
ANIL [48]52.82 ± 0.2966.52 ± 0.2841.12 ± 0.1555.82 ± 0.2134.77 ± 0.3146.55 ± 0.29
BOIL [43]53.23 ± 0.4169.37 ± 0.2344.20 ± 0.1560.92 ± 0.1136.12 ± 0.2950.64 ± 0.22
BMAML [64]N/AN/A33.52 ± 0.3651.35 ± 0.16N/AN/A
ALFA [9]N/AN/AN/A58.35 ± 0.25N/AN/A
L2F [10]N/AN/AN/A60.89 ± 0.22N/AN/A
MeTAL [8]N/AN/AN/A58.20 ± 0.24N/AN/A
HyperMAML [47]N/AN/A36.52 ± 0.6149.43 ± 0.14N/AN/A
CxGrad [33]N/AN/AN/A63.92 ± 0.44N/AN/A
Sparse-MAML†[57]53.47 ± 0.5368.83 ± 0.6541.37 ± 0.7360.58 ± 1.1035.90 ± 0.5052.63 ± 0.56
Sparse-ReLU-MAML†[57]53.77 ± 0.9468.12 ± 0.6942.89 ± 0.4557.53 ± 0.9436.04 ± 0.5549.95 ± 0.42
Sparse-MAML+†[57]53.91 ± 0.6769.92 ± 0.2143.43 ± 1.0462.02 ± 0.7837.14 ± 0.7753.18 ± 0.44
Approximate GAP†57.47 ± 0.9971.66 ± 0.7643.77 ± 0.7962.92 ± 0.7337.00 ± 0.7553.28 ± 0.76
GAP†58.56 ± 0.9372.82 ± 0.7744.74 ± 0.7564.88 ± 0.7238.44 ± 0.7755.04 ± 0.77
", + "image_path": "bb85ba485c0ae5c303032b7042dc412448487831e4704d755dc6f458b848194e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "type": "text", + "content": "Table 6. 5-way few-shot cross domain classification accuracy (%) with a Conv-4 backbone, meta training on mini-[ImageNet dataset, and meta-testing on tiered-[ImageNet, CUB, or Cars datasets. We report mean ± 95% confidence intervals(ci) for 600 tasks according to [20].† denotes PGD-MAML family." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 294, + 241, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 241, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 241, + 306 + ], + "type": "text", + "content": "4.3. Cross-domain few-shot classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 290, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 290, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 290, + 469 + ], + "type": "text", + "content": "The cross-domain few-shot classification introduced by [14] addresses a more challenging and practical few-shot classification scenario in which meta-train and meta-test tasks are sampled from different task distributions. These scenarios are designed to evaluate meta-level overfitting of meta-learning algorithms by creating a large domain gap between meta-trains and meta-tests. In particular, an algorithm can be said to be meta-overfitting if it relies too much on the prior knowledge of previously seen meta-train tasks instead of focusing on a few given examples to learn a new task. This meta-level overfitting makes the learning system more likely to fail to adapt to new tasks sampled from substantially different task distributions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 289, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 289, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 289, + 671 + ], + "type": "text", + "content": "Datasets and experimental setup. To evaluate the level of meta-overfitting for GAP, we evaluate a cross-domain few-shot classification experiment. The mini-ImageNet is used for the meta-train task, and the tiered-ImageNet [52], CUB-200-2011 [58], Cars [11] datasets are used for the meta-test task. The CUB dataset has 200 fine-grained classes and consists of a total of 11,788 images; it is further divided into 100 meta-train classes, 50 meta-validation classes, and 50 meta-test classes. The Cars [30] dataset consists of 16,185 images of 196 classes of cars; it is split into 8,144 training images and 8,041 testing images, where each class has been split roughly in 50-50. The classes are typically at the level of Make, Model, Year, e.g., 2012 Tesla Model S or 2012 BMW M3 coupe. As with the few-shot classification experiment, we use the standard Conv-4 backbone and follow the same experimental protocol." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "Results. Table 6 presents the cross-domain few-shot performance for GAP, MAML family, and PGD-MAML family. GAP significantly outperforms the state-of-the-art al" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 294, + 549, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 549, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 549, + 486 + ], + "type": "text", + "content": "gorithms on 5-way 1-shot and 5-way 5-shot cross-domain classification tasks. In particular, for the tiered-ImageNet dataset, the performance was improved by " + }, + { + "bbox": [ + 304, + 294, + 549, + 486 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 304, + 294, + 549, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 294, + 549, + 486 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 304, + 294, + 549, + 486 + ], + "type": "text", + "content": " on 1-shot and 5-shot classification tasks, respectively. Because GAP can simultaneously consider a task's individuality and optimization trajectory in the inner-loop optimization, it can overcome meta-overfitting better than the existing methods. However, Approximate GAP shows more performance degradation in cross-domain few-shot classification than in few-shot classification. In particular, when the domain difference with the meta-train is more significant (i.e., the tiered-ImageNet dataset) than when the domain difference with the meta-train is marginal (i.e., CARS and CUB datasets), it shows a more considerable performance drop. We can see that full adaptation plays an important role in cross-domain few-shot classification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 498, + 376, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 376, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 376, + 510 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 514, + 547, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 681 + ], + "type": "text", + "content": "Number of meta parameters: Recent MAML family and PGD-MAML family require a large increase in the number of meta-learning parameters as shown in Table 7. One advantage of GAP is that it requires only a very small increase in the number of meta parameters, when compared to the baseline of MAML. This is possible because we transform a gradient tensor into a gradient matrix, perform the SVD of the matrix, and assign only a small number of meta parameters that correspond to the diagonal matrix of the gradient matrix. For the Conv-4 network, GAP requires only " + }, + { + "bbox": [ + 304, + 514, + 547, + 681 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 514, + 547, + 681 + ], + "type": "text", + "content": " increase of the meta parameters. Although the increase in the number of meta parameters is negligible, SVD of the gradient matrix can incur a large computational burden for large networks. This is addressed by Approximate GAP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Approximate GAP vs. simple constant preconditioners: Approximate GAP is a low-complexity method where SVD" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16086" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 272, + 171 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 272, + 171 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 272, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 272, + 171 + ], + "type": "table", + "html": "
Algorithm# of params% increase
MAML [20]1.2109 × 105
Meta-SGD [36]2.4218 × 105100.0%
MC [44]2.7106 × 1062140.4%
PAMELA [49]1.6239 × 10534.1%
MH [66]7.2196 × 10759586.7%
Sparse-MAML [57]2.4218 × 105100.0%
GAP1.2131 × 1050.2%
", + "image_path": "c3f61bf0f6eedafdbe3b76cc3dab9dd858577028da8b0594dcaf2519de0b1862.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 208, + 272, + 250 + ], + "blocks": [ + { + "bbox": [ + 47, + 175, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 196 + ], + "type": "text", + "content": "Table 7. Comparison of the number of parameters for MAML, existing methods, and GAP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 208, + 272, + 250 + ], + "lines": [ + { + "bbox": [ + 59, + 208, + 272, + 250 + ], + "spans": [ + { + "bbox": [ + 59, + 208, + 272, + 250 + ], + "type": "table", + "html": "
AlgorithmStructureRiemannian metric (i.e., positive definite)Acc. (%)
Meta-SGDdiag(a1, ·, an)X50.47%
Meta-SGD with positive definitenessdiag(a1, ·, an)O52.39%
Approximate GAPblkdiag(Mt, ·, M)O53.52%
", + "image_path": "fa8e326b7def17789b540bcf82ec93d1c4420aa5ac0b88aad471a18f41286c3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 60, + 308, + 272, + 354 + ], + "blocks": [ + { + "bbox": [ + 47, + 254, + 288, + 297 + ], + "lines": [ + { + "bbox": [ + 47, + 254, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 288, + 297 + ], + "type": "text", + "content": "Table 8. Performance comparison of three constant (i.e., non-adaptive) preconditioners for 5-way 1-shot on mini-ImageNet: Meta-SGD, Meta-SGD modified to satisfy positive definiteness, and our Approximate GAP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 308, + 272, + 354 + ], + "lines": [ + { + "bbox": [ + 60, + 308, + 272, + 354 + ], + "spans": [ + { + "bbox": [ + 60, + 308, + 272, + 354 + ], + "type": "table", + "html": "
Algorithm1-shot5-shot
GAP w/o PGAP48.23 ± 0.8065.80 ± 0.75
GAP w/ PGAP54.86 ± 0.8571.55 ± 0.61
", + "image_path": "b6a5ebed9c63bf234c92763df44f976deabf9d547bd64fc74b04ad8232397df7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 357, + 286, + 389 + ], + "lines": [ + { + "bbox": [ + 47, + 357, + 286, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 286, + 389 + ], + "type": "text", + "content": "Table 9. Ablation study of " + }, + { + "bbox": [ + 47, + 357, + 286, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 47, + 357, + 286, + 389 + ], + "type": "text", + "content": " on mini-ImageNet. Performance of the GAP-trained model is significantly affected by not applying " + }, + { + "bbox": [ + 47, + 357, + 286, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 47, + 357, + 286, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "type": "text", + "content": "operation is avoided by approximating GAP with a constant diagonal preconditioner. A natural question to ask is how does Approximate GAP compare with other constant diagonal preconditioners. To answer this question, we have compared Approximate GAP with Meta-SGD and a modified Meta-SGD. Meta-SGD [36] is a well-known constant diagonal preconditioner (i.e., " + }, + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(a_1,\\dots ,a_n)" + }, + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "type": "text", + "content": ") that does not need to be positive definite and we also investigate its modification with a constraint on positive definiteness. The results are shown in Table 8. It can be observed that enforcing positive definiteness can improve Meta-SGD. Furthermore, an additional improvement can be achieved by Approximate GAP. While both modified Meta-SGD and Approximate GAP are positive definite, Approximate GAP is different because it inherits an additional constraint from GAP - a block diagonal structure where a constant diagonal matrix " + }, + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "type": "text", + "content": " is repeated (i.e., " + }, + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\mathrm{blkdiag}(M,\\dots ,M)" + }, + { + "bbox": [ + 46, + 406, + 287, + 622 + ], + "type": "text", + "content": "). The inherited constraint provides a gain over the modified Meta-SGD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Does GAP learn a useful preconditioner: While a Riemannian metric can be helpful, it does not mean any Riemannian metric will result in an improvement. For the true parameter space with a specific underlying structure, the corresponding Riemannian metric needs to be applied to enable steepest descent [3, 4]. For the special case of a two-layer neural network with a mean squared error (MSE) loss, it" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "content": "was proven that Fisher information matrix is the corresponding Riemannian metric [4]. For a general neural network, however, a proper Riemannian metric is unknown and it needs to be learned. In our work, we have devised a method to guarantee a Riemannian metric and have used the outer-loop optimization to learn the Riemannian metric. In general, the learned Riemannian metric is unlikely to correspond perfectly to the true parameter space. Then, an important question is if the Riemannian metric learned by GAP is close enough to the desired one and if it is useful. To investigate this issue, we performed an ablation study by not applying the preconditioner " + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "content": ". After training a GAP model, we have evaluated the performance with and without applying " + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{GAP}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "content": ". The results are shown in Table 9 and clearly the preconditioner learned with the outer-loop optimization plays an essential role for improving the performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 282, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 547, + 474 + ], + "type": "text", + "content": "Why is preconditioner helpful for meta-learning: When the batch size is small, the resulting empirical gradient can be noisy [53, 65]. A typical few-shot learning has only a small number of samples for the inner-loop optimization, and its gradient can be noisy. On the other hand, it was shown in [5] that preconditioned gradient descent with a positive definite preconditioner can achieve a lower risk than gradient descent when the labels are noisy, the model is mis-specified, or the signal is misaligned with the features. Under a misalignment, a properly chosen positive definite preconditioner can generalize better than gradient descent [5]. Considering the noisy gradient of inner loop optimization, it can be surmised that a positive definite preconditioner that is adaptive (i.e., a Riemannian metric) can be helpful for improving MAML. Note that the noisy case is in contrast to the case of supervised learning with a large amount of data [5]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 491, + 378, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 378, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 378, + 504 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 514, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 586 + ], + "type": "text", + "content": "In this work, we have proposed GAP that is a PGDMAML algorithm utilizing SVD operation. Thanks to the fully adaptive property, GAP can handle individual tasks with a wide diversity within the MAML framework. Additionally, GAP can enable steepest descent on the parameter space owing to the Riemannian metric property." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 604, + 408, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 408, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 408, + 616 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 624, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 547, + 703 + ], + "type": "text", + "content": "This work was supported by ETRI [23ZR1100, A Study of Hyper-Connected Thinking Internet Technology by autonomous connecting, controlling and evolving ways], NRF (NRF-2020R1A2C2007139), IITP [NO.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University)], and the New Faculty Startup Fund from Seoul National University." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16087" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Momin Abbas, Quan Xiao, Lisha Chen, Pin-Yu Chen, and Tianyi Chen. Sharp-maml: Sharpness-aware model-agnostic meta learning. arXiv preprint arXiv:2206.03996, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Shunichi Amari. A theory of adaptive pattern classifiers. IEEE Transactions on Electronic Computers, (3):299-307, 1967. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "type": "text", + "content": "[3] Shun-ichi Amari. Neural learning in structured parameter spaces-natural riemannian gradient. Advances in neural information processing systems, 9, 1996. 1, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "type": "text", + "content": "[4] Shun-Ichi Amari. Natural gradient works efficiently in learning. Neural computation, 10(2):251-276, 1998. 1, 3, 4, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[5] Shun-ichi Amari, Jimmy Ba, Roger Grosse, Xuechen Li, Atsushi Nitanda, Taiji Suzuki, Denny Wu, and Ji Xu. When does preconditioning help or hurt generalization? arXiv preprint arXiv:2006.10732, 2020. 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 327 + ], + "type": "text", + "content": "[6] Shun-Ichi Amari and Scott C Douglas. Why natural gradient? In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), volume 2, pages 1213-1216. IEEE, 1998. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 329, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 361 + ], + "type": "text", + "content": "[7] Antreas Antoniou, Harrison Edwards, and Amos Storkey. How to train your maml. arXiv preprint arXiv:1810.09502, 2018.6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "type": "text", + "content": "[8] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9465-9474, 2021. 1, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "type": "text", + "content": "[9] Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. Meta-learning with adaptive hyperparameters. Advances in Neural Information Processing Systems, 33:20755-20765, 2020. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 464, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 288, + 508 + ], + "type": "text", + "content": "[10] Sungyong Baik, Seokil Hong, and Kyoung Mu Lee. Learning to forget for meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2379-2387, 2020. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 542 + ], + "type": "text", + "content": "[11] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "type": "text", + "content": "[12] Irénée-Jules Bienaymé. Considerations à l'appui de la découverte de Laplace sur la loi de probabilité dans la méthode des moindres carrés. Imprimerie de Mallet-Bachelier, 1853. 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 288, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 288, + 611 + ], + "type": "text", + "content": "[13] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 644 + ], + "type": "text", + "content": "[14] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[15] Tristan Deleu, Tobias Würfl, Mandana Samiei, Joseph Paul Cohen, and Yoshua Bengio. Torchmeta: A meta-learning library for pytorch. arXiv preprint arXiv:1909.06576, 2019. 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 95, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 138 + ], + "type": "text", + "content": "[17] Lin Ding, Peng Liu, Wenfeng Shen, Weijia Lu, and Shengbo Chen. Gradient-based meta-learning using uncertainty to weigh loss for few-shot learning. arXiv preprint arXiv:2208.08135, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 547, + 172 + ], + "type": "text", + "content": "[18] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "text", + "content": "[19] Chen Fan, Parikshit Ram, and Sijia Liu. Sign-maml: Efficient model-agnostic meta-learning by signsgd. arXiv preprint arXiv:2109.07497, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 206, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 547, + 249 + ], + "type": "text", + "content": "[20] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 1, 3, 5, 6, 7, 8, 14" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "text", + "content": "[21] Chelsea Finn, Kelvin Xu, and Sergey Levine. Probabilistic model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "text", + "content": "[22] Sebastian Flennerhag, Andrei A Rusu, Razvan Pascanu, Francesco Visin, Hujun Yin, and Raia Hadsell. Meta-learning with warped gradient descent. arXiv preprint arXiv:1909.00025, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[23] Erin Grant, Chelsea Finn, Sergey Levine, Trevor Darrell, and Thomas Griffiths. Recasting gradient-based meta-learning as hierarchical bayes. arXiv preprint arXiv:1801.08930, 2018.6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "text", + "content": "[24] Markus Hiller, Mehrtash Harandi, and Tom Drummond. On enforcing better conditioned meta-learning for rapid few-shot adaptation. arXiv preprint arXiv:2206.07260, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "type": "text", + "content": "[25] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9068-9077, 2022. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 14" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "type": "text", + "content": "[27] Sham M Kakade. A natural policy gradient. Advances in neural information processing systems, 14, 2001. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 526, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 557 + ], + "type": "text", + "content": "[28] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3,14" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 559, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 580 + ], + "type": "text", + "content": "[29] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[30] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "type": "text", + "content": "[31] Yann A LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-48. Springer, 2012. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 658, + 523, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 523, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 523, + 669 + ], + "type": "text", + "content": "[32] John M Lee. Smooth manifolds. Springer, 2012. 3, 12" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[33] Sanghyuk Lee, Seunghyun Lee, and Byung Cheol Song. Contextual gradient scaling for few-shot learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 834-843, 2022. 6, 7" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16088" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[34] Yoonho Lee and Seungjin Choi. Gradient-based meta-learning with learned layerwise metric and subspace. In International Conference on Machine Learning, pages 2927-2936. PMLR, 2018. 1, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 149 + ], + "type": "text", + "content": "[35] Geng Li, Boyuan Ren, and Hongzhi Wang. Eeml: Ensemble embedded meta-learning. arXiv preprint arXiv:2206.09195, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 288, + 185 + ], + "type": "text", + "content": "[36] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1, 4, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 288, + 228 + ], + "type": "text", + "content": "[37] Zequn Liu, Ruiyi Zhang, Yiping Song, and Ming Zhang. When does maml work the best? an empirical study on model-agnostic meta-learning in nlp applications. arXiv preprint arXiv:2005.11700, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 288, + 262 + ], + "type": "text", + "content": "[38] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 288, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 288, + 318 + ], + "type": "text", + "content": "[39] Ahmed M Abdelmoniem, Ahmed Elzanaty, Mohamed-Slim Alouini, and Marco Canini. An efficient statistical-based gradient compression technique for distributed training systems. Proceedings of Machine Learning and Systems, 3:297-322, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 288, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 288, + 375 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 288, + 375 + ], + "type": "text", + "content": "[40] Gabriel Maicas, Andrew P Bradley, Jacinto C Nascimento, Ian Reid, and Gustavo Carneiro. Training medical image analysis systems like radiologists. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 546-554. Springer, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 376, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 288, + 408 + ], + "type": "text", + "content": "[41] George Marsaglia. Choosing a point from the surface of a sphere. The Annals of Mathematical Statistics, 43(2):645-646, 1972. 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 411, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 288, + 453 + ], + "type": "text", + "content": "[42] Fei Mi, Minlie Huang, Jiyong Zhang, and Boi Faltings. Meta-learning for low-resource natural language generation in task-oriented dialogue systems. arXiv preprint arXiv:1905.05644, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 455, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 455, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 455, + 288, + 487 + ], + "type": "text", + "content": "[43] Jaehoon Oh, Hyungjun Yoo, ChangHwan Kim, and Se-Young Yun. Boil: Towards representation change for few-shot learning. arXiv preprint arXiv:2008.08882, 2020. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 489, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 288, + 520 + ], + "type": "text", + "content": "[44] Eunbyung Park and Junier B Oliva. Meta-curvature. Advances in Neural Information Processing Systems, 32, 2019. 1, 4, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 523, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 288, + 544 + ], + "type": "text", + "content": "[45] Danni Peng and Sinno Pan. Clustered task-aware meta-learning by learning from learning paths. 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 545, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 288, + 599 + ], + "type": "text", + "content": "[46] Juan-Manuel Perez-Rua, Xiatian Zhu, Timothy M Hospedales, and Tao Xiang. Incremental few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13846-13855, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 601, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 288, + 644 + ], + "type": "text", + "content": "[47] M Przewięźlikowski, P Przybysz, J Tabor, M Zieba, and P Spurek. Hypermaml: Few-shot adaptation of deep models with hypernetworks. arXiv preprint arXiv:2205.15745, 2022. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "type": "text", + "content": "[48] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 1, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 288, + 713 + ], + "type": "text", + "content": "[49] Jathushan Rajasegaran, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Mubarak Shah. Meta-learning" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "the learning trends shared across tasks. arXiv preprint arXiv:2010.09291, 2020. 1, 4, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "type": "text", + "content": "[50] Aravind Rajeswaran, Chelsea Finn, Sham M Kakade, and Sergey Levine. Meta-learning with implicit gradients. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 163 + ], + "type": "text", + "content": "[51] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 165, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 165, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 547, + 217 + ], + "type": "text", + "content": "[52] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 263 + ], + "type": "text", + "content": "[53] Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. On modulating the gradient for meta-learning. In European Conference on Computer Vision, pages 556-572. Springer, 2020. 1, 4, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 266, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 545, + 308 + ], + "type": "text", + "content": "[54] Rishav Singh, Vandana Bharti, Vishal Purohit, Abhinav Kumar, Amit Kumar Singh, and Sanjay Kumar Singh. Metamed: Few-shot medical image classification using gradient-based meta-learning. Pattern Recognition, 120:108111, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 311, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 311, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 311, + 545, + 365 + ], + "type": "text", + "content": "[55] Xingyou Song, Yuxiang Yang, Krzysztof Choromanski, Ken Caluwaerts, Wenbo Gao, Chelsea Finn, and Jie Tan. Rapidly adaptable legged robots via evolutionary meta-learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3769-3776. IEEE, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "type": "text", + "content": "[56] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016, 6, 14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "text", + "content": "[57] Johannes Von Oswald, Dominic Zhao, Seijin Kobayashi, Simon Schug, Massimo Caccia, Nicolas Zucchet, and João Sacramento. Learning where to learn: Gradient sparsity in meta and continual learning. Advances in Neural Information Processing Systems, 34:5250-5263, 2021. 1, 4, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 468, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 498 + ], + "type": "text", + "content": "[58] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 502, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 545, + 555 + ], + "type": "text", + "content": "[59] Shuhuan Wen, Zeteng Wen, Di Zhang, Hong Zhang, and Tao Wang. A multi-robot path-planning algorithm for autonomous navigation using meta-reinforcement learning based on transfer learning. Applied Soft Computing, 110:107605, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 558, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 558, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 558, + 545, + 623 + ], + "type": "text", + "content": "[60] Simon Wiedemann, Temesgen Mehari, Kevin Kepp, and Wojciech Samek. Dithered backprop: A sparse and quantized backpropagation algorithm for more efficient deep neural network training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 720-721, 2020. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[61] Xiongwei Wu, Doyen Sahoo, and Steven Hoi. Meta-rcnn: Meta learning for few-shot object detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1679–1687, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[62] Peng Yang, Shaogang Ren, Yang Zhao, and Ping Li. Calibrating cnns for few-shot meta learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2090–2099, 2022. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16089" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 229 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[63] Huaxiu Yao, Xian Wu, Zhiqiang Tao, Yaliang Li, Bolin Ding, Ruirui Li, and Zhenhui Li. Automated relational meta-learning. arXiv preprint arXiv:2001.00745, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "text", + "content": "[64] Jaesik Yoon, Taesup Kim, Ousmane Dia, Sungwoong Kim, Yoshua Bengio, and Sungjin Ahn. Bayesian model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[65] Yikai Zhang, Hui Qu, Chao Chen, and Dimitris Metaxas. Taming the noisy gradient: train deep neural networks with small batch sizes. In The Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI), 2019. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "type": "text", + "content": "[66] Dominic Zhao, Johannes von Oswald, Seijin Kobayashi, João Sacramento, and Benjamin F Grewe. Meta-learning via hypernetworks. 2020. 1, 4, 6, 8" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16090" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_content_list.json b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..473d3ff14354182c9de9b52c637f1bf52ee4fe62 --- /dev/null +++ b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "Meta-Personalizing Vision-Language Models to Find Named Instances in Video", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chun-Hsiao Yeh $^{1,3*}$ Bryan Russell $^{3}$ Josef Sivic $^{2,3}$ Fabian Caba Heilbron $^{3\\dagger}$ Simon Jenni $^{3\\dagger}$", + "bbox": [ + 94, + 178, + 869, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of California, Berkeley 2CIIRC CTU 3Adobe Research", + "bbox": [ + 181, + 202, + 789, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "daniel.yeh@berkeley.edu {brussell, inr03127, caba, jenni} @adobe.com", + "bbox": [ + 194, + 223, + 772, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 318, + 313, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale vision-language models (VLM) have shown impressive results for language-guided search applications. While these models allow category-level queries, they currently struggle with personalized searches for moments in a video where a specific object instance such as \"My dog Biscuit\" appears. We present the following three contributions to address this problem. First, we describe a method to meta-personalize a pre-trained VLM, i.e., learning how to learn to personalize a VLM at test time to search in video. Our method extends the VLM's token vocabulary by learning novel word embeddings specific to each instance. To capture only instance-specific features, we represent each instance embedding as a combination of shared and learned global category features. Second, we propose to learn such personalization without explicit human supervision. Our approach automatically identifies moments of named visual instances in video using transcripts and vision-language similarity in the VLM's embedding space. Finally, we introduce This-Is-My, a personal video instance retrieval benchmark. We evaluate our approach on This-Is-My and DeepFashion2 and show that we obtain a $15\\%$ relative improvement over the state of the art on the latter dataset.", + "bbox": [ + 75, + 351, + 473, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 710, + 209, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent introduction of large-scale pre-trained vision-language models (VLMs) has enabled many new vision tasks, including zero-shot classification and retrieval [14, 18, 22], image/video generation [12, 24, 25, 27, 28, 32], or language-guided question answering [1, 19, 42]. It is now possible to search not only for specific object categories (e.g., dogs) but also for more specific descriptions of both", + "bbox": [ + 75, + 734, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48552c1d4940bd8f73574710c34c6c90dea7d045f9fbe7695919b460665f9156.jpg", + "image_caption": [ + "Figure 1. Meta-Personalized Vision-Language Model (VLM) to Retrieve Named Instances in Video. Given a video where a user-specific instance, e.g., \"My dog Biscuit\" is mentioned, our method automatically learns a representation for the user-specific instance in the VLM's text input space. The personalized VLM can then be used to retrieve the learned instance in other contexts through natural language queries, e.g., grabbing a pink frisbee. This result is enabled by metapersonalizing the VLM on a large-scale dataset of narrated videos by pre-learning shared global category tokens (in this example for the category of 'dogs'), which are then easily personalized to user-specific instances from only a few user-given training examples." + ], + "image_footnote": [ + "...create recipes to them...This is my dog Biscuit...If you teach them a trick..." + ], + "bbox": [ + 511, + 276, + 875, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Input Video with Transcript", + "text_level": 1, + "bbox": [ + 584, + 571, + 799, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the object and scene attributes (e.g., \"A small white dog playing at the dog park\"). However, we often do not want to search for just any example of a generic category but instead to find a specific instance. For example, a user might want to search their personal video library for all the scenes that show their dog \"Biscuit grabbing a pink frisbee\", as illustrated in Figure 1. Since VLMs do not have a representation of \"Biscuit,\" such queries are beyond the capabilities", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal advising.", + "bbox": [ + 101, + 851, + 192, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during CHY's summer internship at Adobe Research.", + "bbox": [ + 102, + 864, + 450, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Czech Institute of Informatics, Robotics and Cybernetics at the Czech Technical University in Prague.", + "bbox": [ + 78, + 876, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19123", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of off-the-shelf VLMs.", + "bbox": [ + 76, + 90, + 230, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent work [5] proposed a method to extend the language encoder's vocabulary with a newly learned token that represents a specific personal instance to address this issue. While this approach enables language-guided search for personal instances by placing the learned tokens in the query prompt, their solution assumes a collection of manually annotated images showing the individual instance in various contexts for successful token learning. For this approach to work in practice, a user must manually annotate all their important personal instances in various contexts, such that the instance representation does not capture nuisance features, e.g., the background. We thus identify two key challenges: 1) collecting personal instance examples without explicit human labeling and 2) learning a generalizable object-centric representation of personal instances from very few examples.", + "bbox": [ + 75, + 109, + 472, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are three-fold. As our first contribution, we propose a method to automatically identify important personal instances in videos for personalizing a vision-language model without explicit human annotations. Indeed, people often record and refer to personal items or relationships in videos found online. Our approach is thus to identify mentions of personal instances in a video automatically and leverage these moments to build a set of personal instances for training. To this end, we extract the transcripts of videos using speech-to-text models and find candidate moments by looking for occurrences of \"this is my *\" or similar possessive adjective patterns. The symbol * in this example could represent a single word or sequence of words describing the instance (e.g., *= \"dog Biscuit\"). We then use vision-language similarity to filter non-visual examples and to find additional occurrences in the video for training. For example, we found more than six thousand named instances in 50K videos randomly sampled from the Merlot Reserve dataset [44]. We call the resulting collection of named instances in videos the This-Is-My dataset.", + "bbox": [ + 75, + 353, + 472, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As our second contribution, we propose a novel model and training procedure to learn text tokens representing the named instances in video from possibly very few and noisy training examples. Our method represents each instance with learned tokens and models each token as a linear combination of a set of pre-learned category-specific features shared across different instances. This set of shared category-specific features (similar to object attributes) improves the generalization of our method by preventing the instance representations from capturing nuisance features (e.g., the scene background). Furthermore, we show how to pre-train and adapt the shared category features using a large set of automatically collected This-Is-My examples, further improving our model's few-shot personalization performance at test-time. We call this pre-training of shared category features meta-personalization. In contrast to prior", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work [5], our method does not require training additional neural network models and requires only the optimization of a contrastive learning objective.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As our final contribution, we demonstrate and evaluate our model on an existing fashion item retrieval benchmark, DeepFashion2, and our new challenging This-Is-My video instance retrieval dataset4 depicting specific object instances across different videos and contexts. Our experiments demonstrate that our method outperforms several baselines and prior approaches on these challenging language-guided instance retrieval tasks.", + "bbox": [ + 496, + 137, + 893, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 273, + 640, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Models for Video Retrieval. Vision-language foundational models [14, 18, 22] have been successful for zero-shot and other diverse video tasks, such as video question answering [19, 42], language-video grounding [13, 37, 41], and text-to-video retrieval [11, 16, 17, 20, 21, 39, 40]. These models have a powerful representation that transfers well to the video domain to achieve competitive performance on video-language tasks. Our approach builds on these powerful representations to retrieve specific named instances in video.", + "bbox": [ + 496, + 299, + 893, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Personalized Concept Learning. Adapting a model to learn a user-specific representation has been a significant topic in machine learning research, including recommendation systems [2, 3] and federated learning [15]. Relevant to us are recent approaches for adapting vision-language models to object instances. PALAVRA [5] proposes a learning scheme that appends a learnable token for a new personalized concept to the token embedding of the input text prompt. This learned representation helps to preserve the personalized concept. DualPrompt [36] introduces a framework that learns a small set of prompts to emphasize more specific concepts without forgetting the learned concepts in the pre-trained foundational model. There have also been works that have extended this personalized concept learning to image generation [4,8,26]. While these approaches adapt a vision-language model to personal instances, they perform the adaptation independently for each instance and do not \"meta-train\" for the personalization task for improved few-shot learning as we do in this work.", + "bbox": [ + 496, + 450, + 893, + 737 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fine-tuning and Test-time Adaptation. Fine-tuning is a common strategy to adapt a pretrained model for a specific downstream task by transferring the source model to a target domain. Recent works on vision-language model tuning include CLIP-Adapter [9] that proposes to conduct fine-tuning with an extra bottleneck layer while freezing the pre-trained CLIP model. WiSE-FT [38] resolves the distribution shift caused by fine-tuning and ensembles the weights of the original and fine-tuned model to increase ro", + "bbox": [ + 496, + 739, + 893, + 876 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4Available at https://danielchyeh.github.io/metaper/", + "bbox": [ + 514, + 886, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19124", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ee01b6b95138c0199ecf6e9c38be9b8baf03bdd861b048b8471a721a4d48fc4.jpg", + "image_caption": [ + "Figure 2. Overview of our Personalized Vision-Language Model. Our model augments a frozen VLM (blue) with novel personal instance tokens $w = Cz$ (orange) that are a combination of global category features $C$ (green) with instance-specific weights $z \\in \\mathcal{Z}$ (yellow). Our approach for personalized instance retrieval has three stages. First, we pre-learn global category features $C_{\\mathcal{D}}$ on a large set of automatically mined named personal instances in videos. We call this process Meta-Personalization (left). In the second step (middle), we adapt the meta-personalized category features $C_{\\mathcal{D}}$ at test-time and learn novel instance weights $z \\in \\mathcal{Z}_{\\mathcal{P}}$ to represent a user's personal instances via $w = C_{\\mathcal{P}}z$ . Finally (right), we leverage the (frozen) personalized instance tokens $w$ in natural language queries at query time." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 343, + 318 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/948389672c9dc1d540297882241a97776f06808533f67166442f166d11e9a93f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 93, + 596, + 319 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9646bc818d22e4bfaf9d795838662cfb13340f332f3b72d0ca3613d84f3b7a0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 93, + 880, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bustness. Prior test-time adaption works [31, 34, 35] fine-tune on target data without information from the source data. Our approach leverages test-time adaptation for updating our meta-personalized model to user-specific data.", + "bbox": [ + 75, + 428, + 468, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta-Learning. We draw inspiration from meta-learning (\"learning to learn\") [6, 7, 43], which enables models to quickly adapt to new tasks by learning on a diverse set of tasks. Hence, given only a handful of novel training examples, the model can be adapted to novel tasks. Our approach borrows the idea of meta-learning to meta-personalize a model by learning global category features from a large video database. We then adapt the global features during test-time training from only few examples of user-specific instances to enable query-time retrieval.", + "bbox": [ + 75, + 489, + 468, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. (Meta-)Personalization of Named Instances", + "text_level": 1, + "bbox": [ + 76, + 660, + 464, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to learn representations of personal items in video that enable retrieval through natural language queries. To achieve this goal, we have to address the key challenge of adapting a model with one or few examples of a named instance. To address this challenge, we propose a metapersonalization approach that learns to personalize given a large corpus of named instances mined from videos with transcriptions. We illustrate our approach in Figure 2.", + "bbox": [ + 75, + 686, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first step (Figure 2 (left)), we mine automatically a collection $\\mathcal{D}$ of named instances from videos with transcriptions for meta-personalization. We use this collection to train a proposed model $\\mathcal{M}_{C,z}$ that includes global category features $C$ and instance-specific parameters $z$ . The global category features $C$ are lightweight and shared across all", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "instances. Given a natural language query $u$ and video $v$ , the model returns a score $\\mathcal{M}_{C,z}(u,v)$ . During metapersonalization, given a training loss $\\mathcal{L}$ , we jointly optimize the loss over the global category features $C$ and instance parameters $\\mathcal{Z}$ for each named instance in the collection $\\mathcal{D}$ ,", + "bbox": [ + 498, + 428, + 890, + 503 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(C _ {\\mathcal {D}}, \\mathcal {Z} _ {\\mathcal {D}}\\right) \\in \\underset {(C, \\mathcal {Z})} {\\arg \\min } \\sum_ {z \\in \\mathcal {Z}} \\mathcal {L} (C, z). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 515, + 890, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that here the instance-specific parameters $\\mathcal{Z}_{\\mathcal{D}}$ learnt via (1) are discarded while the global category features $C_{\\mathcal{D}}$ are kept as the meta-personalized part of the model. The global category features $C_{\\mathcal{D}}$ capture information shared across instances relevant to the personalization task.", + "bbox": [ + 498, + 558, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the second step (Figure 2 (middle)), we are given a set of named video instances $\\mathcal{P}$ (e.g., automatically mined from someone's personal video library) and wish to perform test-time personalization of the model to this person's instances. Here each instance is represented by only one or few examples. In this step, we optimize the training loss $\\mathcal{L}$ over the global category features $C$ and the set of instance parameters $\\mathcal{Z}$ for all instances in the personal set $\\mathcal{P}$ starting from the pre-trained global category features $C_{\\mathcal{D}}$ and random $\\mathcal{Z}$ . We obtain the personalized model parameters $(C_{\\mathcal{P}},\\mathcal{Z}_{\\mathcal{P}})$ as", + "bbox": [ + 496, + 633, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(C _ {\\mathcal {P}}, \\mathcal {Z} _ {\\mathcal {P}}\\right) \\in \\underset {(C, \\mathcal {Z})} {\\arg \\min } \\sum_ {z \\in \\mathcal {Z}} \\mathcal {L} (C, z). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 796, + 890, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now keep both $C_{\\mathcal{P}}$ and $\\mathcal{Z}_{\\mathcal{P}}$", + "bbox": [ + 500, + 840, + 710, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the final step (Figure 2 (right)), we perform retrieval over a potentially large dataset using the test-time personalized model $\\mathcal{M}_{C_{\\mathcal{P}},z}$ (where $z \\in \\mathcal{Z}_{\\mathcal{P}}$ ). We next describe", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19125", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b74e414fc093a921d5a9a1dab87f549dd0b73bcd07430e8ac1ab6448e74f7a6.jpg", + "image_caption": [ + "visual similarity for every pair" + ], + "image_footnote": [], + "bbox": [ + 83, + 102, + 166, + 143 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/788d6595b4af62523175399336e012db3ab4b6061e6bca806ce0bc5fb42f9c12.jpg", + "image_caption": [ + "0.2" + ], + "image_footnote": [], + "bbox": [ + 241, + 97, + 310, + 146 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/70e7dfee7a79237d126b138c718afcd2bf12d62f224a0755dc23350ec8400c55.jpg", + "image_caption": [ + "set of candidates" + ], + "image_footnote": [], + "bbox": [ + 392, + 97, + 462, + 146 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/53a57fea915bc7aa2aff6185f49d5b588c64b32e4ea128c6bb174f4939f10fd3.jpg", + "image_caption": [ + "Finding Additional Instance Examples (Step 3)" + ], + "image_footnote": [], + "bbox": [ + 81, + 172, + 104, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6426d3a23890796359acef37c8a38cabc8f06c3bf07c44323719e5560fc3bdc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 193, + 174, + 238 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3b20ae614e3ac354e5e83ebb7e136378a311f5f2ff1fb9cb2ecedb23e6aa7e6.jpg", + "image_caption": [ + "instance name: Time to talk about" + ], + "image_footnote": [], + "bbox": [ + 174, + 193, + 264, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2606f2a888dd299fe0f41072ec5e46716a45e88bb51f6d1b07e7f732ba7f4e68.jpg", + "image_caption": [ + "visual relevance: 0.1", + "instance name: fender guitar" + ], + "image_footnote": [], + "bbox": [ + 279, + 193, + 364, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/46fd0c6eaaa5ee6a60bca5e18a3fc897a75e17f33f1f70dd467a51c2cff0308b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 193, + 452, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/61adb1288277139d8ca03dc56ef128da6e74185220e114f3bce1717d1ccef932.jpg", + "image_caption": [ + "Filtering Non-visual Instances (Step 2)" + ], + "image_footnote": [], + "bbox": [ + 83, + 267, + 101, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/67903db5e0f75d42b56b19ad4f5a062b5cb2246b2123e95a3342eb21b46de532.jpg", + "image_caption": [ + "Video:", + "Transcript: This is our time to talk about ... This is my fender guitar", + "Spotting Named Instances (Step 1)", + "Figure 3. Automatic mining of named instances in video for meta-personalization. Our automatic mining pipeline includes three steps (from bottom to top). Step 1 finds named instances via string-matching of possessive patterns in video transcripts. Step 2 filters non-visual instances using text-to-visual relevance between the instance name and the video shots neighboring the named instance. Finally, Step 3 retrieves additional shots with high visual similarity to the instance reference shot." + ], + "image_footnote": [], + "bbox": [ + 138, + 277, + 455, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "how we automatically mine the named instances in video $\\mathcal{D}$ for meta-personalization (Sec. 3.1), and our full model $\\mathcal{M}$ for query-time retrieval and loss $\\mathcal{L}$ for meta-personalization and test-time personalization (Sec. 3.2).", + "bbox": [ + 75, + 484, + 468, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Automatic Mining of Named Instances in Video", + "text_level": 1, + "bbox": [ + 76, + 556, + 468, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To identify personal instances without explicit supervision, we leverage a collection of videos from the web along with their corresponding time-aligned transcripts. These transcripts can be automatically generated through speech-to-text models [23, 33]. We now describe a procedure to mine these data for a set of moments (i.e., a collection of video shots) depicting a referred personal instance in the transcript without the need for manual annotation. We will use these moments for training a meta-personalization model (Sec. 3.2).", + "bbox": [ + 75, + 580, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spotting Named Instances. Here our goal is to find moments where candidate personal instances are mentioned in videos. We do so by searching for possessive text patterns5 such as \"This is my *\" in a corpus of time-aligned video transcripts. This string-matching process outputs a list of candidate instance names * associated with video timestamps $t^*$ . In prac", + "bbox": [ + 75, + 733, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tice, we keep up to four words after a possessive text pattern is matched based on text-visual similarity (see supp. for details). That way, we can retrieve simple named instances such as This is my dog ( $=$ dog) but also complex ones like This is my favorite CHANEL classic handbag ( $=$ CHANEL classic handbag). Note also that a single video might include multiple string matches; for instance, the example video illustrated in Figure 3 (Step 1) includes two matches: \"This is our time to talk about\" at time 1:30 and \"This is my fender guitar\" at time 3:25.", + "bbox": [ + 500, + 90, + 890, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Filtering Non-visual Instances. The previous spotting step only searches for potential instances using the transcript, yielding many string matches that are non-visual, i.e., the strings do not describe the visible content in the video. Here we aim to filter out these non-visual instances. We do so by computing the text-to-visual relevance between the instance name (e.g., fender guitar) and the neighboring shots around the time when the instance is mentioned. We add neighboring shots to cover cases where the named instances are shown just before or after they are mentioned. Concretely, given a sequence of $m$ video shots $S = [s_1, \\ldots, s_m]$ automatically extracted with [30], we find the shot $s_{t^*}$ that overlaps with $t^*$ (the time when the instance was mentioned). Next, we form a set of candidate visual references $S_{t^*} = [s_{t^* - 1}, s_{t^*}, s_{t^* + 1}]$ comprising a window of shots that are previous and subsequent to $s_{t^*}$ . We then compute text-to-visual relevance scores using CLIP [22] encoders. This encoding process yields $L_2$ -normalized embeddings $f_l(*)$ for the named instance and $f_v(s_i)$ for each shot $s_i \\in S_{t^*}$ . We compute $f_v(s_i)$ by averaging the visual embeddings of all frames in the corresponding shot. Finally, we compute the cosine similarity between every $(f_l(*), f_v(s_i))$ pair and retain a visual reference shot $s^*$ if the highest cosine similarity is greater than 0.3. This filtering step outputs a cleaned set of named instances with a corresponding visual reference. Figure 3 (Step 2) illustrates how we prune out non-visual matches such as time to talk about. In contrast, visual instances such as fender guitar are kept and matched with a visual reference.", + "bbox": [ + 496, + 261, + 892, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finding Additional Instance Shots. Since frames from a single video shot provide only limited variability in the instance appearance and could thus limit the learning, we aim to recover other shots from the video where that instance appears. We leverage CLIP's visual encoder to compute the visual similarity between the instance's reference shot $s^*$ and every shot $s_i \\in S$ . We extract an embedding $f_v(s^*)$ for the reference shot $s^*$ and an embedding $f_v(s_i)$ for each shot $s_i$ . Similar to the non-visual filtering step, we average the CLIP embeddings of each frame belonging to a shot. Then, we compute the cosine similarity between the embeddings for the reference shot and every candidate shot in the video. We keep the shots whose cosine similarity with the refer", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "5List of possessive text patterns: , , , , , , , , , ", + "bbox": [ + 76, + 849, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19126", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e8379900235a8874a3fddede6276fb59ec0913edffd3118dabffc64be6c38eb3.jpg", + "image_caption": [ + "$\\mathcal{M}_{\\mathcal{C},z}(\\pmb {u},\\pmb {v})$", + "Figure 4. Model Overview. Our model $\\mathcal{M}_{C,z}$ extends CLIP's language input vocabulary with $n_w$ novel instance-specific tokens $w_i^y = C_l z_i^y$ , which we model as a linear combination of metapersonalized category features $C_l$ with weights $z_i^y$ . Note that the vision and language encoders are frozen during this process." + ], + "image_footnote": [], + "bbox": [ + 76, + 104, + 472, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ence is greater than 0.9. Figure 3 (Step 3) illustrates the output of this final step in the automatic mining pipeline. Our mining algorithm allows us to retrieve additional shots for the instance fender guitar. Now the instance examples not only include a clean close-up of the guitar, but also shots where the guitar is being played or held by its owner.", + "bbox": [ + 75, + 431, + 468, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Learning Personal Instance Representations", + "text_level": 1, + "bbox": [ + 76, + 554, + 452, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The result of the mining procedure described above is a dataset consisting of a set of video shots $\\mathcal{D} = \\{s_1, \\ldots, s_n\\}$ and corresponding instance IDs $Y = \\{y_1, \\ldots, y_n\\}$ , where $y_i = y_j$ if $s_i$ and $s_j$ are video shots that are assumed to contain the same instance. We now describe how we use these data to learn representations of the collected instances.", + "bbox": [ + 75, + 579, + 468, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach is to leverage a large-scale pre-trained vision-language model (CLIP [22]) and augment its language encoder with a set of novel personal instance tokens. Let $f_{v}(s)$ be the output of the visual encoder for shot $s$ (computed as the average over the frame embeddings) and let $f_{l}(u)$ be the output of the language encoder for a natural language input $u = [v_{1},\\dots ,v_{m}]$ of $m$ token embeddings, where $v_{i}\\in \\mathbb{R}^{d}$ denote learned token embeddings of the pretraining vocabulary (positional embeddings are included but omitted from the notation). We propose to extend this vocabulary with novel personal instance tokens. Concretely, our approach introduces a set $w^{y} = \\{w_{i}^{y}\\}_{i = 1}^{n_{w}}$ of $n_w$ new tokens that represent a personal instance $y\\in Y$ .", + "bbox": [ + 75, + 671, + 468, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To learn these tokens and to perform personalized retrieval at test time, we construct natural language person", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "alized queries as", + "bbox": [ + 500, + 90, + 612, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {u} ^ {p} = \\left[ p _ {1}, \\dots , p _ {k - 1}, w _ {1} ^ {y}, \\dots , w _ {n _ {w}} ^ {y}, p _ {k + 1}, \\dots , p _ {m} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 116, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i$ are token embeddings of a chosen prompt $p = [p_1, \\ldots, p_m]$ . During training, the prompt $p$ corresponds to a random template of the form [An image of *, [* can be seen in this photo], [There is * in this image], etc., and $k$ denotes the position of the * placeholder for the instance tokens.", + "bbox": [ + 498, + 143, + 893, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instances as Combinations of Category Features. Since learning personal instance tokens from possibly very few examples runs the danger of overfitting (e.g., to nuisance features in the background), we propose to parameterize them as", + "bbox": [ + 498, + 234, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {y} = C _ {l} z _ {i} ^ {y} \\in \\mathbb {R} ^ {d \\times 1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 308, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{i}^{y} \\in \\mathbb{R}^{q \\times 1}$ is a vector of learnable weights specific to each instance and $C_{l} \\in \\mathbb{R}^{d \\times q}$ is a matrix of learnable global category features, which are shared for all instances belonging to the same object category $l \\in \\mathcal{V}$ (e.g., $\\mathcal{V} = \\{\\text{car}, \\text{person}, \\text{dog}, \\ldots\\}$ ) and constitute the set $C_{\\mathcal{D}} = \\{C_{l}\\}_{l \\in \\mathcal{V}}$ . We illustrate the model in Figure 4.", + "bbox": [ + 498, + 330, + 890, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can think of the columns of $C_l$ as representing a set of shared category features and the final instance token $w_i^y$ as a linear combination of these features with weights $z_i^y$ . Our aim is that only category-specific features are captured during training and irrelevant features (e.g., about the background) are discarded. Intuitively, the columns of $C_l$ could correspond to attributes of an object category, e.g., if we were to learn \"car\" features, these could capture their color, brand, type, or age, to name a few. To identify which category matrix $C_l$ to use for an instance $y$ , we rely on 0-shot classification using vision-language similarity between instance shots and a generic prompt for the category $l \\in \\mathcal{V}$ .", + "bbox": [ + 496, + 422, + 892, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive Personal Token Learning. We propose a contrastive learning objective to learn the personal instance tokens $w^{y}$ using a set of video shots containing the instance $y$ . To this end, let $\\psi_{i} := f_{v}(s_{i})$ be an encoding of a video shot $s_{i}$ , i.e., the average frame encoding of all frames in the shot, and let $\\phi_{i} := f_{l}(\\hat{u}_{i}^{p})$ denote the language encoding of a corresponding personalized query. We learn the novel tokens $w^{y}$ and shared category features $C_{l}$ by optimizing two contrastive objectives: a language-language contrastive objective $\\mathcal{L}_{l}$ and a vision-language contrastive objective $\\mathcal{L}_{vl}$ . The language-language objective is given by", + "bbox": [ + 496, + 603, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l} = \\sum_ {i \\in \\mathcal {B}} \\sum_ {j \\neq i \\in \\mathcal {B}} - 1 \\left\\{y _ {i} = y _ {j} \\right\\} \\log \\left(\\frac {d \\left(\\phi_ {i} , \\phi_ {j}\\right)}{\\sum_ {k \\neq i \\in \\mathcal {B}} d \\left(\\phi_ {i} , \\phi_ {k}\\right)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 777, + 890, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(a,b)\\coloneqq \\exp \\left(\\frac{1}{\\lambda}\\frac{a^{\\intercal}b}{\\|a\\|_{2}\\|b\\|_{2}}\\right)$ measures the similarity between the feature vectors $a$ and $b$ , $\\lambda = 0.1$ is a temperature parameter, and $\\mathcal{B}$ is a randomly sampled training minibatch. The vision-language objective is similarly defined as", + "bbox": [ + 496, + 834, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19127", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/60425bf28057a6954da847bfede9fb95f80f36c3ab27741bf639b951d1c9c3dd.jpg", + "image_caption": [ + "Figure 5. Examples from This-Is-My { Meta-Personalization $\\mathcal{D}$ (top) vs Test-time personalization $\\mathcal{P}$ (bottom-left) vs Query-time $\\mathcal{Q}$ (bottom-right)} datasets. In the Query-time dataset (bottom-right), we design a challenging video instance retrieval task. For example, in (a) the named instance (i.e., Alex's piano) is in the background and is barely visible and in (b) the background scenes in the query-time dataset (bottom-right) are completely different from the test-time personalization dataset (bottom-left) depicting the same named instance." + ], + "image_footnote": [], + "bbox": [ + 145, + 92, + 826, + 315 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v l} = \\sum_ {i, j \\in \\mathcal {B}} - \\mathbb {1} \\left\\{y _ {i} = y _ {j} \\right\\} \\log \\left(\\frac {d \\left(\\phi_ {i} , \\psi_ {j}\\right)}{\\sum_ {k \\in \\mathcal {N}} d \\left(\\phi_ {i} , \\psi_ {k}\\right)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 402, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with the set of negative examples $\\mathcal{N}$ comprising both other examples in the batch $\\mathcal{B}$ and non-instance shots from the videos containing the named instances, i.e., shots that have low vision-language similarity. The loss is low when the encodings for video shots and personalized queries with the same instance ID are more similar than to other queries. Including non-instance segments as negatives can help discard non-instance features such as scene background.", + "bbox": [ + 75, + 449, + 468, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further constrain the learning toward category-specific attributes, we include a loss that maximizes the similarity between a personal instance query and a generic category query. Concretely, let $c_{l}$ be a category query embedding for category $l$ (e.g., \"An image of a [car]\") to which instance $y$ belongs. We then include the following category-anchoring loss", + "bbox": [ + 75, + 570, + 468, + 675 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = - \\sum_ {i \\in \\mathcal {B}} \\frac {c _ {l} ^ {\\top} \\phi_ {i}}{\\| c _ {l} \\| _ {2} \\| \\phi_ {i} \\| _ {2}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 681, + 468, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To summarize, our training loss $\\mathcal{L}$ (see Equations 1 and 2) for meta- and test-time personalization is given by", + "bbox": [ + 76, + 724, + 468, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {l} + \\mathcal {L} _ {v l} + \\lambda_ {c} \\mathcal {L} _ {c}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 763, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{c} = 0.5$ , controls the amount of category anchoring.", + "bbox": [ + 76, + 787, + 467, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. This-Is-My Dataset", + "text_level": 1, + "bbox": [ + 76, + 829, + 259, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our This-Is-My dataset comprises three subsets for metapersonalization, test-time personalization, and querying. We describe each subset next.", + "bbox": [ + 76, + 854, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meta-Personalization Dataset $\\mathcal{D}$ . We gather this subset with the automatic mining algorithm introduced in Section 3.1. We leverage the Merlot Reserve dataset [44], which contains more than 20 million videos from YouTube and their corresponding time-aligned transcripts. In practice, we start from a subset of $50K$ randomly sampled Merlot Reserve videos. We spot 6058 named instances using various text possessive templates. Our visual filtering step removes $52\\%$ of instances, generating a total of 2908 named instances with a visual reference. Finally, we mine additional samples for each instance, yielding a total of 49256 instance samples. This subset includes a wide variety of visual concepts, ranging from common objects such as bikes to rare concepts such as toaster. While we design our mining pipeline to minimize noise, it is limited by CLIP's capability to distinguish between similar object instances. For instance, while we find several samples of the fender guitar instance, it includes other shots that do not correspond to the aforementioned guitar (Figure 5 (top)). Nevertheless, we will show empirically that this subset is still useful for meta-personalization purposes (Section 5).", + "bbox": [ + 496, + 383, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test-time Personalization Dataset $\\mathcal{P}$ . Our goal is to create a test dataset that recreates the scenario where a person wants to find when their personal instances appear in their video collection. We want to make this task close to the real scenario where a person records the same visual instance, e.g., their dog, across multiple places and situations. Our strategy is to emulate such a scenario by finding YouTube channels that frequently mention the same instance across multiple videos. While Merlot Reserve is large and diverse, only few channels are represented with more than one video in the dataset. Instead, we download all videos and automatic transcripts from the channels of 15 popular YouTube bloggers. We then run our mining algorithm, but manu", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19128", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ally supervise the steps for visual filtering and finding additional instance samples. We first manually verify that each instance name and its visual reference are good matches. Then, we find additional sample shots across all videos in the channel by ranking them according to their visual similarity to the instance reference shot. Finally, we review and label the top 1000-scored shots for each named instance. The supplementary material includes a screenshot of the annotation tool. In total, this subset includes 15 named instances with more than 686 labeled samples. Figure 5 (bottom left) shows two example instances in our dataset.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query-time Dataset $\\mathcal{Q}$ . Our end goal is to retrieve named instances via natural language queries. We would like to be able to find videos when is grabbing a pink Frisbee. To this end, we manually caption 30 instance samples with descriptive queries. Thus, the Query-time dataset includes (manually captioned) video-caption pairs containing instances from the Test-time Personalization dataset. Figure 5 (bottom right) shows manually curated captions for two instances in our dataset.", + "bbox": [ + 75, + 257, + 472, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 407, + 209, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments first ablate our model's contributions and loss design (Section 5.1), and evaluate our final model in personalized instance retrieval benchmarks (Section 5.2). Evaluation Datasets and Metrics. We evaluate our approach on two datasets: (i) our newly introduced This-IsMy personal video instance retrieval benchmark, and (ii) the personalized image retrieval benchmark built on DeepFashion2 [10] proposed by [5]. We evaluate retrieval performance in two settings:", + "bbox": [ + 75, + 431, + 468, + 568 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Generic Instance Retrieval: In this case, we build queries for learned instances using a generic prompt, i.e., \"An image of *\", and measure retrieval performance using mean Average Precision (mAP) and Mean Reciprocal Rank (MRR). In this setting, there are multiple correct matches.", + "2) Contextualized Instance Retrieval: In this case, we use natural language queries that describe the personal instance in a specific context, e.g., using a scene description such as \"A photo of * lying on the beach\". In this case, we assume there is only a single correct match, and we measure performance with MRR and Recall-at-5 (R@5)." + ], + "bbox": [ + 75, + 568, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We use the ViT-B/16 version of CLIP in most of our experiments if not otherwise indicated. All the learnable parameters $z_{i}$ and $C_l$ are randomly initialized from $\\mathcal{N}(0,0.1)$ . We set the number of category features to $q = 512$ and the number of instance tokens to $n_w = 1$ by default. No data augmentation is used during training (the visual embeddings thus have to be computed just once). For meta-personalization of category features $C_l$ , we randomly select 32 named instances from $\\mathcal{D}$ that are 0-shot classified to each category $l$ and train for 20 epochs. This process is repeated 10 times, re-initializing $z_{i}$ each", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cebc7e374bdf02282a6538cdf81deb70bb37ffa5e5b2e34be3745885a1144e32.jpg", + "table_caption": [ + "Table 1. Ablation Experiments. We verify our model and training objective design through ablations on This-Is-My and DeepFashion2. We report personal instance retrieval performance in terms of mAP and MRR (higher is better)." + ], + "table_footnote": [], + "table_body": "
AblationThis-Is-MyDeepFashion2
mAPMRRmAPMRR
a) w/o meta-pers.54.1±1.383.1±2.135.2±0.555.2±1.9
b) single C55.5±0.486.9±0.544.1±0.564.7±1.5
c) w/o Ll53.9±1.086.4±2.247.3±0.868.3±1.2
d) w/o Lc48.9±1.078.6±2.447.1±0.568.4±0.8
e) N = B47.4±0.473.7±1.8--
f) w/o pre-trained C53.0±0.785.1±3.044.3±0.867.1±1.0
Ours56.4±0.687.4±1.247.3±0.769.9±0.8
", + "bbox": [ + 509, + 146, + 879, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "time while retaining $C_l$ from the previous run. Although updating the parameters of $C_l$ to a small set of instances $\\mathcal{P}$ at test time is beneficial, there is a risk of overfitting when only a few or just a single instance of each category is provided. To mitigate this, we include additional instances from $\\mathcal{D}$ with the same categories as contained in $\\mathcal{P}$ during test time personalization. All our results are averaged over five runs, each with different random seeds, and we report the standard error for each experiment.", + "bbox": [ + 496, + 291, + 890, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Ablations", + "text_level": 1, + "bbox": [ + 500, + 438, + 609, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1 we report results in the generic instance retrieval setting for the following ablations to illustrate the effect of our model and training objective design:", + "bbox": [ + 496, + 462, + 890, + 508 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "a) w/o meta-personalization: In this case, we do not learn a meta-personalized $C$ and instead directly learn instance embeddings $w_{i}$ . This ablation demonstrates that learning global category attributes through meta-personalization improves generalization at test-time personalization.", + "b) single $C$ (shared for all categories): In this experiment, we learn only a single category matrix $C$ , which is shared among all the categories. The results demonstrate that learning separate attributes per category is better than sharing global attributes among all categories.", + "c) $\\mathbf{w} / \\mathbf{o}\\mathcal{L}_l$ : We explore the influence of the language-language contrastive loss $\\mathcal{L}_l$ . The benefits of $\\mathcal{L}_l$ are more pronounced in This-Is-My, where instances belong to different categories.", + "d) $\\mathbf{w} / \\mathbf{o}\\mathcal{L}_c$ : We analyze the contribution of the category-anchoring loss $\\mathcal{L}_c$ . Including $\\mathcal{L}_c$ is important on This-Is-My, likely because it keeps the learning focused on the category and prevents it from capturing other scene features. This is less the case on DeepFashion2, where images are very object-centric, and there is less background variety.", + "e) $\\mathcal{N} = \\mathcal{B}$ (w/o non-instance segments): We do not use non-instance segments $\\mathcal{N}$ (i.e., segments of the same video but low vision-language similarity) as additional negatives in this ablation. Including $\\mathcal{N}$ is important, as it can help prevent the instance embedding from capturing non-instance nuisance features from the scene." + ], + "bbox": [ + 496, + 508, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19129", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/325c0220828f2c93b9e3e989a7151786af47645a95998c682cbf6350a1f1be53.jpg", + "image_caption": [ + "Language Queries", + "Top-5 Personalized Retrievals" + ], + "image_footnote": [ + " is eating food with a white plate" + ], + "bbox": [ + 189, + 109, + 467, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38ee14909a757f87ca57d54923524830cd3430f5ecaac3a00cf26b4ba2a8f4ee.jpg", + "image_caption": [], + "image_footnote": [ + "man is leading $<$ Zak's dog Coffee> with a dog leash" + ], + "bbox": [ + 189, + 152, + 467, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/204fd41855bab52074f33c7febc1597354df3685b5f3a3fb4bf85dc78d34b5f4.jpg", + "image_caption": [], + "image_footnote": [ + " is with a computer monitor showing 1:12 or the screen" + ], + "bbox": [ + 189, + 195, + 467, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d3a9141a56783e276ad130c492e950e39ed0293ae554e978d35976fc4cd470b.jpg", + "image_caption": [ + "Figure 6. Contextualized This-Is-My Retrievals. We show personalized query-time retrievals for four This-Is-My instances. Search prompts are shown on the left and correct retrievals are highlighted in green." + ], + "image_footnote": [ + "a man is putting into a black case near red cases" + ], + "bbox": [ + 189, + 241, + 467, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "f) w/o pre-trained $C$ : In this case, we randomly initialize $C$ instead of using a meta-personalized $C$ . Compared to a) the sharing of category attributes among fashion items in DeepFashion2 shows clear benefits. This is less the case for This-Is-My, where instances belong to different categories.", + "bbox": [ + 75, + 337, + 467, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Personal Instance Retrieval Benchmarks", + "text_level": 1, + "bbox": [ + 76, + 422, + 424, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. To demonstrate the effectiveness of our approach, we compare it to prior work [5] and the following strong baselines: CLIP (visual), which represents each instance as the average visual embedding of the corresponding training examples, CLIP (language), which represents each instance with a 0-shot prompt embedding of the corresponding category name, and CLIP $(\\mathbf{V} + \\mathbf{L})$ , which uses the average of the visual and language embedding.", + "bbox": [ + 75, + 446, + 467, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Named Video Instance Retrieval on This-Is-My. We evaluate our approach on the 15 named instances in our This-Is-My test-time personalization dataset $\\mathcal{P}$ . To learn personal instance tokens, we train on video shots that occur in the video where the instance was named. All other video shots (belonging to other videos) are in the retrieval corpus for evaluation. Furthermore, we also include all other shots that do not contain the instance in the retrieval corpus as distractors. For contextualized retrieval we use the manually collected text captions from the query-time dataset $\\mathcal{Q}$ as queries. We compare against the baselines and report contextualized retrieval performance on the left and generic instance retrieval performance on the right of Table 2. Our model clearly outperforms the baselines in both settings. Interestingly, the visual and language baselines have opposing advantages in generic vs. contextualized retrieval, while our model performs well in both settings. Qualitative retrieval results are shown in Figures 6.", + "bbox": [ + 75, + 568, + 467, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fashion Item Retrieval on DeepFashion2. We follow the setup of [5] and consider DeepFashion2 [10] for personalized fashion item retrieval. The dataset in this setting consists of 653 training and 221 evaluation images across 50", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/147261c9f630c44c669e510376fc154e8db36166a1be1b34e785be709099ea4e.jpg", + "table_caption": [ + "Table 2. This-Is-My Video Instance Retrieval Task. We report personal instance retrieval performance in language-specified contexts (e.g., \"* catching a pink frisbee\") on the left and generic instance retrieval (e.g., \"An image of *\") on the right." + ], + "table_footnote": [], + "table_body": "
MethodContext. Retr.Generic Retr.
MRRR@5mAPMRR
Random0.0±0.00.0±0.01.1±0.22.2±1.3
CLIP (language)30.8±0.036.7±0.016.6±0.044.2±0.0
CLIP (visual)10.3±1.012.0±1.648.0±0.675.0±3.2
CLIP (V+L)20.9±1.623.3±2.151.7±0.481.9±0.0
Ours42.0±1.350.7±1.856.4±0.687.4±1.2
", + "bbox": [ + 521, + 146, + 867, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bb719e136c6c5b74199099ac08379dd87ae8db6acefce70a16354afc7e1be294.jpg", + "table_caption": [ + "Table 3. Fashion Item Retrieval on DeepFashion2. We evaluate our approach on the personalized instance retrieval task with contextualized queries from [5] (left) and generic instance retrieval (right). Results with * use ViT-B/32 instead of ViT-B/16." + ], + "table_footnote": [], + "table_body": "
MethodContext. Retr.Generic Retr.
MRRR@5mAPMRR
Random2.9±0.21.8±0.54.7±0.59.5±2.0
CLIP (language)21.2±0.025.3±0.08.3±0.016.9±0.0
CLIP (visual)14.2±0.317.3±0.320.6±0.443.7±1.1
CLIP (V+L)20.8±0.825.4±1.720.5±0.542.8±1.2
Adapter* [5]5.9±0.7---
COLLIE* [5,29]7.9±0.7---
PALAVRA* [5]28.4±0.739.2±1.3--
Ours*34.4±0.745.2±1.140.0±1.069.3±1.8
Ours38.4±0.451.4±0.453.4±0.477.7±0.6
", + "bbox": [ + 519, + 315, + 867, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "instances (i.e., unique fashion items). While our focus is on video retrieval, our model and training objective can be used for image retrieval with minimal adjustments. We perform meta-personalization by pre-training on other non-instance-labeled images of DeepFashion2 (using only a single training image per instance). For a fair comparison with [5], the instance tokens are learned by randomly sampling $k = 5$ images per instance for training. We compare to results from [5] and the baselines in Table 3. Results using contextualized queries provided by [5] are on the left, and generic instance retrieval performance on the right. Note that [5] does not provide results for generic instance retrieval so we leave these two columns empty (-). Our approach outperforms the baselines and prior work by a large margin.", + "bbox": [ + 496, + 483, + 890, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 709, + 617, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced a meta-personalization approach for learning to retrieve named instances in video given a natural language query. We demonstrated the effectiveness of our approach on two datasets and showed that it outperforms strong baselines. Our effort is a step towards vision-language models trained over a large number of general and personal concepts. Our approach opens up the possibility of personalized language-based video editing, summarization, and generation, e.g., identify the key instances in a collection of footage and create an edited story about the instances using natural language.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19130", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 1", + "[2] Fernando Amat, Ashok Chandrashekar, Tony Jebara, and Justin Basilico. Artwork personalization at netflix. In Proceedings of the 12th ACM conference on recommender systems, pages 487-488, 2018. 2", + "[3] Soulef Benhamdi, Abdesselam Babouri, and Raja Chiky. Personalized recommender system for e-learning environment. Education and Information Technologies, 22(4):1455-1477, 2017. 2", + "[4] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 2", + "[5] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. “This is my unicorn, Fluffy”: Personalizing frozen vision-language representations. In ECCV, 2022. 2, 7, 8", + "[6] Chelsea Finn, P. Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 3", + "[7] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In International Conference on Machine Learning, pages 1920-1930. PMLR, 2019. 3", + "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. ArXiv, abs/2208.01618, 2022. 2", + "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2", + "[10] Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaou Tang, and Ping Luo. Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 7, 8", + "[11] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5006-5015, 2022. 2", + "[12] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1", + "[13] De-An Huang, S. Buch, Lucio M. Dery, Animesh Garg, Li Fei-Fei, and Juan Carlos Niebles. Finding \"it\": Weakly-supervised reference-aware visual grounding in instructional videos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5948–5957, 2018. 2", + "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2", + "[15] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv preprint arXiv:1909.12488, 2019.2", + "[16] Alexander Kunitsyn, Maksim Kalashnikov, Maksim Dzabraev, and Andrei Ivaniuta. Mdmmt-2: Multidomain multimodal transformer for video retrieval, one more step towards generalization. arXiv preprint arXiv:2203.07086, 2022. 2", + "[17] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2", + "[18] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1, 2", + "[19] Shuang Li, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, and Igor Mordatch. Composing ensembles of pretrained models via iterative consensus. arXiv preprint arXiv:2210.11522, 2022. 1, 2", + "[20] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2", + "[21] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 2", + "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5", + "[23] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. Technical report, Tech. Rep., Technical report, OpenAI, 2022. 4", + "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19131", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1", + "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2", + "[27] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1", + "[28] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 1", + "[29] Gabriel Skantze and Bram Willemsen. Collie: Continual learning of language grounding from language-image embeddings. Journal of Artificial Intelligence Research, 74:1201-1223, 2022. 8", + "[30] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 4", + "[31] Yu Sun, X. Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 3", + "[32] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and D. Erhan. Phenaki: Variable length video generation from open domain textual description. ArXiv, abs/2210.02399, 2022. 1", + "[33] Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, and Juan Miguel Pino. Fairoq s2t: Fast speech-to-text modeling with faireq. In AACL, 2020. 4", + "[34] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. ArXiv, abs/2109.01087, 2021. 3", + "[35] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 3", + "[36] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, 2022. 2", + "[37] Sangmin Woo, Jinyoung Park, Inyong Koo, Sumin Lee, Minki Jeong, and Changick Kim. Explore and match: End-to-end video grounding with transformer. arXiv preprint arXiv:2201.10168, 2022. 2", + "[38] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2", + "[39] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2", + "[40] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 2", + "[41] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 2", + "[42] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 1, 2", + "[43] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 3", + "[44] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 2, 6" + ], + "bbox": [ + 501, + 92, + 890, + 568 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19132", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_model.json b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2eb2b639465858f13fcf94990421371b00a97566 --- /dev/null +++ b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "Meta-Personalizing Vision-Language Models to Find Named Instances in Video" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.179, + 0.87, + 0.199 + ], + "angle": 0, + "content": "Chun-Hsiao Yeh\\(^{1,3*}\\) Bryan Russell\\(^{3}\\) Josef Sivic\\(^{2,3}\\) Fabian Caba Heilbron\\(^{3\\dagger}\\) Simon Jenni\\(^{3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.203, + 0.79, + 0.222 + ], + "angle": 0, + "content": "1University of California, Berkeley 2CIIRC CTU 3Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.224, + 0.773, + 0.24 + ], + "angle": 0, + "content": "daniel.yeh@berkeley.edu {brussell, inr03127, caba, jenni} @adobe.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.319, + 0.314, + 0.335 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.474, + 0.685 + ], + "angle": 0, + "content": "Large-scale vision-language models (VLM) have shown impressive results for language-guided search applications. While these models allow category-level queries, they currently struggle with personalized searches for moments in a video where a specific object instance such as \"My dog Biscuit\" appears. We present the following three contributions to address this problem. First, we describe a method to meta-personalize a pre-trained VLM, i.e., learning how to learn to personalize a VLM at test time to search in video. Our method extends the VLM's token vocabulary by learning novel word embeddings specific to each instance. To capture only instance-specific features, we represent each instance embedding as a combination of shared and learned global category features. Second, we propose to learn such personalization without explicit human supervision. Our approach automatically identifies moments of named visual instances in video using transcripts and vision-language similarity in the VLM's embedding space. Finally, we introduce This-Is-My, a personal video instance retrieval benchmark. We evaluate our approach on This-Is-My and DeepFashion2 and show that we obtain a \\(15\\%\\) relative improvement over the state of the art on the latter dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.21, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.47, + 0.844 + ], + "angle": 0, + "content": "The recent introduction of large-scale pre-trained vision-language models (VLMs) has enabled many new vision tasks, including zero-shot classification and retrieval [14, 18, 22], image/video generation [12, 24, 25, 27, 28, 32], or language-guided question answering [1, 19, 42]. It is now possible to search not only for specific object categories (e.g., dogs) but also for more specific descriptions of both" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.277, + 0.877, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.516, + 0.556, + 0.878, + 0.568 + ], + "angle": 0, + "content": "...create recipes to them...This is my dog Biscuit...If you teach them a trick..." + }, + { + "type": "title", + "bbox": [ + 0.586, + 0.572, + 0.8, + 0.588 + ], + "angle": 0, + "content": "Input Video with Transcript" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Figure 1. Meta-Personalized Vision-Language Model (VLM) to Retrieve Named Instances in Video. Given a video where a user-specific instance, e.g., \"My dog Biscuit\" is mentioned, our method automatically learns a representation for the user-specific instance in the VLM's text input space. The personalized VLM can then be used to retrieve the learned instance in other contexts through natural language queries, e.g., grabbing a pink frisbee. This result is enabled by metapersonalizing the VLM on a large-scale dataset of narrated videos by pre-learning shared global category tokens (in this example for the category of 'dogs'), which are then easily personalized to user-specific instances from only a few user-given training examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "the object and scene attributes (e.g., \"A small white dog playing at the dog park\"). However, we often do not want to search for just any example of a generic category but instead to find a specific instance. For example, a user might want to search their personal video library for all the scenes that show their dog \"Biscuit grabbing a pink frisbee\", as illustrated in Figure 1. Since VLMs do not have a representation of \"Biscuit,\" such queries are beyond the capabilities" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.852, + 0.194, + 0.864 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal advising." + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.865, + 0.452, + 0.876 + ], + "angle": 0, + "content": "*Work done during CHY's summer internship at Adobe Research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.877, + 0.469, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)Czech Institute of Informatics, Robotics and Cybernetics at the Czech Technical University in Prague." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.852, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19123" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.232, + 0.106 + ], + "angle": 0, + "content": "of off-the-shelf VLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.11, + 0.473, + 0.352 + ], + "angle": 0, + "content": "Recent work [5] proposed a method to extend the language encoder's vocabulary with a newly learned token that represents a specific personal instance to address this issue. While this approach enables language-guided search for personal instances by placing the learned tokens in the query prompt, their solution assumes a collection of manually annotated images showing the individual instance in various contexts for successful token learning. For this approach to work in practice, a user must manually annotate all their important personal instances in various contexts, such that the instance representation does not capture nuisance features, e.g., the background. We thus identify two key challenges: 1) collecting personal instance examples without explicit human labeling and 2) learning a generalizable object-centric representation of personal instances from very few examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.473, + 0.656 + ], + "angle": 0, + "content": "The contributions of this work are three-fold. As our first contribution, we propose a method to automatically identify important personal instances in videos for personalizing a vision-language model without explicit human annotations. Indeed, people often record and refer to personal items or relationships in videos found online. Our approach is thus to identify mentions of personal instances in a video automatically and leverage these moments to build a set of personal instances for training. To this end, we extract the transcripts of videos using speech-to-text models and find candidate moments by looking for occurrences of \"this is my *\" or similar possessive adjective patterns. The symbol * in this example could represent a single word or sequence of words describing the instance (e.g., *= \"dog Biscuit\"). We then use vision-language similarity to filter non-visual examples and to find additional occurrences in the video for training. For example, we found more than six thousand named instances in 50K videos randomly sampled from the Merlot Reserve dataset [44]. We call the resulting collection of named instances in videos the This-Is-My dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "As our second contribution, we propose a novel model and training procedure to learn text tokens representing the named instances in video from possibly very few and noisy training examples. Our method represents each instance with learned tokens and models each token as a linear combination of a set of pre-learned category-specific features shared across different instances. This set of shared category-specific features (similar to object attributes) improves the generalization of our method by preventing the instance representations from capturing nuisance features (e.g., the scene background). Furthermore, we show how to pre-train and adapt the shared category features using a large set of automatically collected This-Is-My examples, further improving our model's few-shot personalization performance at test-time. We call this pre-training of shared category features meta-personalization. In contrast to prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "work [5], our method does not require training additional neural network models and requires only the optimization of a contrastive learning objective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.895, + 0.259 + ], + "angle": 0, + "content": "As our final contribution, we demonstrate and evaluate our model on an existing fashion item retrieval benchmark, DeepFashion2, and our new challenging This-Is-My video instance retrieval dataset4 depicting specific object instances across different videos and contexts. Our experiments demonstrate that our method outperforms several baselines and prior approaches on these challenging language-guided instance retrieval tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.274, + 0.642, + 0.29 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.45 + ], + "angle": 0, + "content": "Vision-Language Models for Video Retrieval. Vision-language foundational models [14, 18, 22] have been successful for zero-shot and other diverse video tasks, such as video question answering [19, 42], language-video grounding [13, 37, 41], and text-to-video retrieval [11, 16, 17, 20, 21, 39, 40]. These models have a powerful representation that transfers well to the video domain to achieve competitive performance on video-language tasks. Our approach builds on these powerful representations to retrieve specific named instances in video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.895, + 0.738 + ], + "angle": 0, + "content": "Personalized Concept Learning. Adapting a model to learn a user-specific representation has been a significant topic in machine learning research, including recommendation systems [2, 3] and federated learning [15]. Relevant to us are recent approaches for adapting vision-language models to object instances. PALAVRA [5] proposes a learning scheme that appends a learnable token for a new personalized concept to the token embedding of the input text prompt. This learned representation helps to preserve the personalized concept. DualPrompt [36] introduces a framework that learns a small set of prompts to emphasize more specific concepts without forgetting the learned concepts in the pre-trained foundational model. There have also been works that have extended this personalized concept learning to image generation [4,8,26]. While these approaches adapt a vision-language model to personal instances, they perform the adaptation independently for each instance and do not \"meta-train\" for the personalization task for improved few-shot learning as we do in this work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.895, + 0.877 + ], + "angle": 0, + "content": "Fine-tuning and Test-time Adaptation. Fine-tuning is a common strategy to adapt a pretrained model for a specific downstream task by transferring the source model to a target domain. Recent works on vision-language model tuning include CLIP-Adapter [9] that proposes to conduct fine-tuning with an extra bottleneck layer while freezing the pre-trained CLIP model. WiSE-FT [38] resolves the distribution shift caused by fine-tuning and ensembles the weights of the original and fine-tuned model to increase ro" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.892, + 0.9 + ], + "angle": 0, + "content": "4Available at https://danielchyeh.github.io/metaper/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19124" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.344, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.094, + 0.597, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.094, + 0.882, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.894, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Overview of our Personalized Vision-Language Model. Our model augments a frozen VLM (blue) with novel personal instance tokens \\( w = Cz \\) (orange) that are a combination of global category features \\( C \\) (green) with instance-specific weights \\( z \\in \\mathcal{Z} \\) (yellow). Our approach for personalized instance retrieval has three stages. First, we pre-learn global category features \\( C_{\\mathcal{D}} \\) on a large set of automatically mined named personal instances in videos. We call this process Meta-Personalization (left). In the second step (middle), we adapt the meta-personalized category features \\( C_{\\mathcal{D}} \\) at test-time and learn novel instance weights \\( z \\in \\mathcal{Z}_{\\mathcal{P}} \\) to represent a user's personal instances via \\( w = C_{\\mathcal{P}}z \\). Finally (right), we leverage the (frozen) personalized instance tokens \\( w \\) in natural language queries at query time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.469, + 0.489 + ], + "angle": 0, + "content": "bustness. Prior test-time adaption works [31, 34, 35] fine-tune on target data without information from the source data. Our approach leverages test-time adaptation for updating our meta-personalized model to user-specific data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.469, + 0.642 + ], + "angle": 0, + "content": "Meta-Learning. We draw inspiration from meta-learning (\"learning to learn\") [6, 7, 43], which enables models to quickly adapt to new tasks by learning on a diverse set of tasks. Hence, given only a handful of novel training examples, the model can be adapted to novel tasks. Our approach borrows the idea of meta-learning to meta-personalize a model by learning global category features from a large video database. We then adapt the global features during test-time training from only few examples of user-specific instances to enable query-time retrieval." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.661, + 0.465, + 0.679 + ], + "angle": 0, + "content": "3. (Meta-)Personalization of Named Instances" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Our goal is to learn representations of personal items in video that enable retrieval through natural language queries. To achieve this goal, we have to address the key challenge of adapting a model with one or few examples of a named instance. To address this challenge, we propose a metapersonalization approach that learns to personalize given a large corpus of named instances mined from videos with transcriptions. We illustrate our approach in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In the first step (Figure 2 (left)), we mine automatically a collection \\(\\mathcal{D}\\) of named instances from videos with transcriptions for meta-personalization. We use this collection to train a proposed model \\(\\mathcal{M}_{C,z}\\) that includes global category features \\(C\\) and instance-specific parameters \\(z\\). The global category features \\(C\\) are lightweight and shared across all" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.892, + 0.504 + ], + "angle": 0, + "content": "instances. Given a natural language query \\( u \\) and video \\( v \\), the model returns a score \\( \\mathcal{M}_{C,z}(u,v) \\). During metapersonalization, given a training loss \\( \\mathcal{L} \\), we jointly optimize the loss over the global category features \\( C \\) and instance parameters \\( \\mathcal{Z} \\) for each named instance in the collection \\( \\mathcal{D} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.516, + 0.892, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\left(C _ {\\mathcal {D}}, \\mathcal {Z} _ {\\mathcal {D}}\\right) \\in \\underset {(C, \\mathcal {Z})} {\\arg \\min } \\sum_ {z \\in \\mathcal {Z}} \\mathcal {L} (C, z). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.559, + 0.892, + 0.634 + ], + "angle": 0, + "content": "Note that here the instance-specific parameters \\(\\mathcal{Z}_{\\mathcal{D}}\\) learnt via (1) are discarded while the global category features \\(C_{\\mathcal{D}}\\) are kept as the meta-personalized part of the model. The global category features \\(C_{\\mathcal{D}}\\) capture information shared across instances relevant to the personalization task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.786 + ], + "angle": 0, + "content": "In the second step (Figure 2 (middle)), we are given a set of named video instances \\(\\mathcal{P}\\) (e.g., automatically mined from someone's personal video library) and wish to perform test-time personalization of the model to this person's instances. Here each instance is represented by only one or few examples. In this step, we optimize the training loss \\(\\mathcal{L}\\) over the global category features \\(C\\) and the set of instance parameters \\(\\mathcal{Z}\\) for all instances in the personal set \\(\\mathcal{P}\\) starting from the pre-trained global category features \\(C_{\\mathcal{D}}\\) and random \\(\\mathcal{Z}\\). We obtain the personalized model parameters \\((C_{\\mathcal{P}},\\mathcal{Z}_{\\mathcal{P}})\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.797, + 0.891, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\left(C _ {\\mathcal {P}}, \\mathcal {Z} _ {\\mathcal {P}}\\right) \\in \\underset {(C, \\mathcal {Z})} {\\arg \\min } \\sum_ {z \\in \\mathcal {Z}} \\mathcal {L} (C, z). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.841, + 0.712, + 0.856 + ], + "angle": 0, + "content": "We now keep both \\(C_{\\mathcal{P}}\\) and \\(\\mathcal{Z}_{\\mathcal{P}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In the final step (Figure 2 (right)), we perform retrieval over a potentially large dataset using the test-time personalized model \\(\\mathcal{M}_{C_{\\mathcal{P}},z}\\) (where \\(z \\in \\mathcal{Z}_{\\mathcal{P}}\\)). We next describe" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19125" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.103, + 0.167, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.101, + 0.24, + 0.133 + ], + "angle": 0, + "content": "visual similarity for every pair" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.098, + 0.311, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.094, + 0.382, + 0.147 + ], + "angle": 0, + "content": "0.2" + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.098, + 0.464, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.149, + 0.404, + 0.158 + ], + "angle": 0, + "content": "set of candidates" + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.162, + 0.389, + 0.175 + ], + "angle": 0, + "content": "Finding Additional Instance Examples (Step 3)" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.174, + 0.105, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.184, + 0.308, + 0.192 + ], + "angle": 0, + "content": "visual relevance: 0.1" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.194, + 0.175, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.194, + 0.265, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.194, + 0.366, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.194, + 0.454, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.24, + 0.256, + 0.248 + ], + "angle": 0, + "content": "instance name: Time to talk about" + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.24, + 0.437, + 0.248 + ], + "angle": 0, + "content": "instance name: fender guitar" + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.255, + 0.342, + 0.268 + ], + "angle": 0, + "content": "Filtering Non-visual Instances (Step 2)" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.268, + 0.102, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.302, + 0.13, + 0.312 + ], + "angle": 0, + "content": "Video:" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.278, + 0.457, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.327, + 0.456, + 0.338 + ], + "angle": 0, + "content": "Transcript: This is our time to talk about ... This is my fender guitar" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.342, + 0.323, + 0.355 + ], + "angle": 0, + "content": "Spotting Named Instances (Step 1)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.469 + ], + "angle": 0, + "content": "Figure 3. Automatic mining of named instances in video for meta-personalization. Our automatic mining pipeline includes three steps (from bottom to top). Step 1 finds named instances via string-matching of possessive patterns in video transcripts. Step 2 filters non-visual instances using text-to-visual relevance between the instance name and the video shots neighboring the named instance. Finally, Step 3 retrieves additional shots with high visual similarity to the instance reference shot." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.469, + 0.547 + ], + "angle": 0, + "content": "how we automatically mine the named instances in video \\(\\mathcal{D}\\) for meta-personalization (Sec. 3.1), and our full model \\(\\mathcal{M}\\) for query-time retrieval and loss \\(\\mathcal{L}\\) for meta-personalization and test-time personalization (Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.469, + 0.574 + ], + "angle": 0, + "content": "3.1. Automatic Mining of Named Instances in Video" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.732 + ], + "angle": 0, + "content": "To identify personal instances without explicit supervision, we leverage a collection of videos from the web along with their corresponding time-aligned transcripts. These transcripts can be automatically generated through speech-to-text models [23, 33]. We now describe a procedure to mine these data for a set of moments (i.e., a collection of video shots) depicting a referred personal instance in the transcript without the need for manual annotation. We will use these moments for training a meta-personalization model (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Spotting Named Instances. Here our goal is to find moments where candidate personal instances are mentioned in videos. We do so by searching for possessive text patterns5 such as \"This is my *\" in a corpus of time-aligned video transcripts. This string-matching process outputs a list of candidate instance names * associated with video timestamps \\( t^* \\). In prac" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "tice, we keep up to four words after a possessive text pattern is matched based on text-visual similarity (see supp. for details). That way, we can retrieve simple named instances such as This is my dog (\\(=\\)dog) but also complex ones like This is my favorite CHANEL classic handbag (\\(=\\)CHANEL classic handbag). Note also that a single video might include multiple string matches; for instance, the example video illustrated in Figure 3 (Step 1) includes two matches: \"This is our time to talk about\" at time 1:30 and \"This is my fender guitar\" at time 3:25." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.893, + 0.699 + ], + "angle": 0, + "content": "Filtering Non-visual Instances. The previous spotting step only searches for potential instances using the transcript, yielding many string matches that are non-visual, i.e., the strings do not describe the visible content in the video. Here we aim to filter out these non-visual instances. We do so by computing the text-to-visual relevance between the instance name (e.g., fender guitar) and the neighboring shots around the time when the instance is mentioned. We add neighboring shots to cover cases where the named instances are shown just before or after they are mentioned. Concretely, given a sequence of \\( m \\) video shots \\( S = [s_1, \\ldots, s_m] \\) automatically extracted with [30], we find the shot \\( s_{t^*} \\) that overlaps with \\( t^* \\) (the time when the instance was mentioned). Next, we form a set of candidate visual references \\( S_{t^*} = [s_{t^* - 1}, s_{t^*}, s_{t^* + 1}] \\) comprising a window of shots that are previous and subsequent to \\( s_{t^*} \\). We then compute text-to-visual relevance scores using CLIP [22] encoders. This encoding process yields \\( L_2 \\)-normalized embeddings \\( f_l(*) \\) for the named instance and \\( f_v(s_i) \\) for each shot \\( s_i \\in S_{t^*} \\). We compute \\( f_v(s_i) \\) by averaging the visual embeddings of all frames in the corresponding shot. Finally, we compute the cosine similarity between every \\( (f_l(*), f_v(s_i)) \\) pair and retain a visual reference shot \\( s^* \\) if the highest cosine similarity is greater than 0.3. This filtering step outputs a cleaned set of named instances with a corresponding visual reference. Figure 3 (Step 2) illustrates how we prune out non-visual matches such as time to talk about. In contrast, visual instances such as fender guitar are kept and matched with a visual reference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Finding Additional Instance Shots. Since frames from a single video shot provide only limited variability in the instance appearance and could thus limit the learning, we aim to recover other shots from the video where that instance appears. We leverage CLIP's visual encoder to compute the visual similarity between the instance's reference shot \\( s^* \\) and every shot \\( s_i \\in S \\). We extract an embedding \\( f_v(s^*) \\) for the reference shot \\( s^* \\) and an embedding \\( f_v(s_i) \\) for each shot \\( s_i \\). Similar to the non-visual filtering step, we average the CLIP embeddings of each frame belonging to a shot. Then, we compute the cosine similarity between the embeddings for the reference shot and every candidate shot in the video. We keep the shots whose cosine similarity with the refer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.851, + 0.469, + 0.9 + ], + "angle": 0, + "content": "5List of possessive text patterns: , , , , , , , , , " + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19126" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.092, + 0.159, + 0.106 + ], + "angle": 0, + "content": "\\(\\mathcal{M}_{\\mathcal{C},z}(\\pmb {u},\\pmb {v})\\)" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.106, + 0.473, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.47, + 0.414 + ], + "angle": 0, + "content": "Figure 4. Model Overview. Our model \\(\\mathcal{M}_{C,z}\\) extends CLIP's language input vocabulary with \\(n_w\\) novel instance-specific tokens \\(w_i^y = C_l z_i^y\\), which we model as a linear combination of metapersonalized category features \\(C_l\\) with weights \\(z_i^y\\). Note that the vision and language encoders are frozen during this process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.538 + ], + "angle": 0, + "content": "ence is greater than 0.9. Figure 3 (Step 3) illustrates the output of this final step in the automatic mining pipeline. Our mining algorithm allows us to retrieve additional shots for the instance fender guitar. Now the instance examples not only include a clean close-up of the guitar, but also shots where the guitar is being played or held by its owner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.453, + 0.571 + ], + "angle": 0, + "content": "3.2. Learning Personal Instance Representations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.67 + ], + "angle": 0, + "content": "The result of the mining procedure described above is a dataset consisting of a set of video shots \\(\\mathcal{D} = \\{s_1, \\ldots, s_n\\}\\) and corresponding instance IDs \\(Y = \\{y_1, \\ldots, y_n\\}\\), where \\(y_i = y_j\\) if \\(s_i\\) and \\(s_j\\) are video shots that are assumed to contain the same instance. We now describe how we use these data to learn representations of the collected instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.868 + ], + "angle": 0, + "content": "Our approach is to leverage a large-scale pre-trained vision-language model (CLIP [22]) and augment its language encoder with a set of novel personal instance tokens. Let \\( f_{v}(s) \\) be the output of the visual encoder for shot \\( s \\) (computed as the average over the frame embeddings) and let \\( f_{l}(u) \\) be the output of the language encoder for a natural language input \\( u = [v_{1},\\dots ,v_{m}] \\) of \\( m \\) token embeddings, where \\( v_{i}\\in \\mathbb{R}^{d} \\) denote learned token embeddings of the pretraining vocabulary (positional embeddings are included but omitted from the notation). We propose to extend this vocabulary with novel personal instance tokens. Concretely, our approach introduces a set \\( w^{y} = \\{w_{i}^{y}\\}_{i = 1}^{n_{w}} \\) of \\( n_w \\) new tokens that represent a personal instance \\( y\\in Y \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To learn these tokens and to perform personalized retrieval at test time, we construct natural language person" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.614, + 0.107 + ], + "angle": 0, + "content": "alized queries as" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.117, + 0.892, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\hat {u} ^ {p} = \\left[ p _ {1}, \\dots , p _ {k - 1}, w _ {1} ^ {y}, \\dots , w _ {n _ {w}} ^ {y}, p _ {k + 1}, \\dots , p _ {m} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.144, + 0.894, + 0.234 + ], + "angle": 0, + "content": "where \\( p_i \\) are token embeddings of a chosen prompt \\( p = [p_1, \\ldots, p_m] \\). During training, the prompt \\( p \\) corresponds to a random template of the form [An image of *, [* can be seen in this photo], [There is * in this image], etc., and \\( k \\) denotes the position of the * placeholder for the instance tokens." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.235, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Instances as Combinations of Category Features. Since learning personal instance tokens from possibly very few examples runs the danger of overfitting (e.g., to nuisance features in the background), we propose to parameterize them as" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.309, + 0.891, + 0.327 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {y} = C _ {l} z _ {i} ^ {y} \\in \\mathbb {R} ^ {d \\times 1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.331, + 0.892, + 0.422 + ], + "angle": 0, + "content": "where \\( z_{i}^{y} \\in \\mathbb{R}^{q \\times 1} \\) is a vector of learnable weights specific to each instance and \\( C_{l} \\in \\mathbb{R}^{d \\times q} \\) is a matrix of learnable global category features, which are shared for all instances belonging to the same object category \\( l \\in \\mathcal{V} \\) (e.g., \\( \\mathcal{V} = \\{\\text{car}, \\text{person}, \\text{dog}, \\ldots\\} \\)) and constitute the set \\( C_{\\mathcal{D}} = \\{C_{l}\\}_{l \\in \\mathcal{V}} \\). We illustrate the model in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.893, + 0.603 + ], + "angle": 0, + "content": "We can think of the columns of \\( C_l \\) as representing a set of shared category features and the final instance token \\( w_i^y \\) as a linear combination of these features with weights \\( z_i^y \\). Our aim is that only category-specific features are captured during training and irrelevant features (e.g., about the background) are discarded. Intuitively, the columns of \\( C_l \\) could correspond to attributes of an object category, e.g., if we were to learn \"car\" features, these could capture their color, brand, type, or age, to name a few. To identify which category matrix \\( C_l \\) to use for an instance \\( y \\), we rely on 0-shot classification using vision-language similarity between instance shots and a generic prompt for the category \\( l \\in \\mathcal{V} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Contrastive Personal Token Learning. We propose a contrastive learning objective to learn the personal instance tokens \\( w^{y} \\) using a set of video shots containing the instance \\( y \\). To this end, let \\( \\psi_{i} := f_{v}(s_{i}) \\) be an encoding of a video shot \\( s_{i} \\), i.e., the average frame encoding of all frames in the shot, and let \\( \\phi_{i} := f_{l}(\\hat{u}_{i}^{p}) \\) denote the language encoding of a corresponding personalized query. We learn the novel tokens \\( w^{y} \\) and shared category features \\( C_{l} \\) by optimizing two contrastive objectives: a language-language contrastive objective \\( \\mathcal{L}_{l} \\) and a vision-language contrastive objective \\( \\mathcal{L}_{vl} \\). The language-language objective is given by" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.779, + 0.891, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l} = \\sum_ {i \\in \\mathcal {B}} \\sum_ {j \\neq i \\in \\mathcal {B}} - 1 \\left\\{y _ {i} = y _ {j} \\right\\} \\log \\left(\\frac {d \\left(\\phi_ {i} , \\phi_ {j}\\right)}{\\sum_ {k \\neq i \\in \\mathcal {B}} d \\left(\\phi_ {i} , \\phi_ {k}\\right)}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.835, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(d(a,b)\\coloneqq \\exp \\left(\\frac{1}{\\lambda}\\frac{a^{\\intercal}b}{\\|a\\|_{2}\\|b\\|_{2}}\\right)\\) measures the similarity between the feature vectors \\(a\\) and \\(b\\), \\(\\lambda = 0.1\\) is a temperature parameter, and \\(\\mathcal{B}\\) is a randomly sampled training minibatch. The vision-language objective is similarly defined as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19127" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.146, + 0.093, + 0.827, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.892, + 0.375 + ], + "angle": 0, + "content": "Figure 5. Examples from This-Is-My { Meta-Personalization \\(\\mathcal{D}\\) (top) vs Test-time personalization \\(\\mathcal{P}\\) (bottom-left) vs Query-time \\(\\mathcal{Q}\\) (bottom-right)} datasets. In the Query-time dataset (bottom-right), we design a challenging video instance retrieval task. For example, in (a) the named instance (i.e., Alex's piano) is in the background and is barely visible and in (b) the background scenes in the query-time dataset (bottom-right) are completely different from the test-time personalization dataset (bottom-left) depicting the same named instance." + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.404, + 0.469, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v l} = \\sum_ {i, j \\in \\mathcal {B}} - \\mathbb {1} \\left\\{y _ {i} = y _ {j} \\right\\} \\log \\left(\\frac {d \\left(\\phi_ {i} , \\psi_ {j}\\right)}{\\sum_ {k \\in \\mathcal {N}} d \\left(\\phi_ {i} , \\psi_ {k}\\right)}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.469, + 0.57 + ], + "angle": 0, + "content": "with the set of negative examples \\(\\mathcal{N}\\) comprising both other examples in the batch \\(\\mathcal{B}\\) and non-instance shots from the videos containing the named instances, i.e., shots that have low vision-language similarity. The loss is low when the encodings for video shots and personalized queries with the same instance ID are more similar than to other queries. Including non-instance segments as negatives can help discard non-instance features such as scene background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.469, + 0.676 + ], + "angle": 0, + "content": "To further constrain the learning toward category-specific attributes, we include a loss that maximizes the similarity between a personal instance query and a generic category query. Concretely, let \\( c_{l} \\) be a category query embedding for category \\( l \\) (e.g., \"An image of a [car]\") to which instance \\( y \\) belongs. We then include the following category-anchoring loss" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.683, + 0.469, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = - \\sum_ {i \\in \\mathcal {B}} \\frac {c _ {l} ^ {\\top} \\phi_ {i}}{\\| c _ {l} \\| _ {2} \\| \\phi_ {i} \\| _ {2}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.726, + 0.469, + 0.756 + ], + "angle": 0, + "content": "To summarize, our training loss \\(\\mathcal{L}\\) (see Equations 1 and 2) for meta- and test-time personalization is given by" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.765, + 0.469, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {l} + \\mathcal {L} _ {v l} + \\lambda_ {c} \\mathcal {L} _ {c}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.788, + 0.468, + 0.804 + ], + "angle": 0, + "content": "where \\(\\lambda_{c} = 0.5\\), controls the amount of category anchoring." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.26, + 0.848 + ], + "angle": 0, + "content": "4. This-Is-My Dataset" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Our This-Is-My dataset comprises three subsets for metapersonalization, test-time personalization, and querying. We describe each subset next." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Meta-Personalization Dataset \\(\\mathcal{D}\\). We gather this subset with the automatic mining algorithm introduced in Section 3.1. We leverage the Merlot Reserve dataset [44], which contains more than 20 million videos from YouTube and their corresponding time-aligned transcripts. In practice, we start from a subset of \\(50K\\) randomly sampled Merlot Reserve videos. We spot 6058 named instances using various text possessive templates. Our visual filtering step removes \\(52\\%\\) of instances, generating a total of 2908 named instances with a visual reference. Finally, we mine additional samples for each instance, yielding a total of 49256 instance samples. This subset includes a wide variety of visual concepts, ranging from common objects such as bikes to rare concepts such as toaster. While we design our mining pipeline to minimize noise, it is limited by CLIP's capability to distinguish between similar object instances. For instance, while we find several samples of the fender guitar instance, it includes other shots that do not correspond to the aforementioned guitar (Figure 5 (top)). Nevertheless, we will show empirically that this subset is still useful for meta-personalization purposes (Section 5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Test-time Personalization Dataset \\(\\mathcal{P}\\). Our goal is to create a test dataset that recreates the scenario where a person wants to find when their personal instances appear in their video collection. We want to make this task close to the real scenario where a person records the same visual instance, e.g., their dog, across multiple places and situations. Our strategy is to emulate such a scenario by finding YouTube channels that frequently mention the same instance across multiple videos. While Merlot Reserve is large and diverse, only few channels are represented with more than one video in the dataset. Instead, we download all videos and automatic transcripts from the channels of 15 popular YouTube bloggers. We then run our mining algorithm, but manu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19128" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.257 + ], + "angle": 0, + "content": "ally supervise the steps for visual filtering and finding additional instance samples. We first manually verify that each instance name and its visual reference are good matches. Then, we find additional sample shots across all videos in the channel by ranking them according to their visual similarity to the instance reference shot. Finally, we review and label the top 1000-scored shots for each named instance. The supplementary material includes a screenshot of the annotation tool. In total, this subset includes 15 named instances with more than 686 labeled samples. Figure 5 (bottom left) shows two example instances in our dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.395 + ], + "angle": 0, + "content": "Query-time Dataset \\(\\mathcal{Q}\\). Our end goal is to retrieve named instances via natural language queries. We would like to be able to find videos when is grabbing a pink Frisbee. To this end, we manually caption 30 instance samples with descriptive queries. Thus, the Query-time dataset includes (manually captioned) video-caption pairs containing instances from the Test-time Personalization dataset. Figure 5 (bottom right) shows manually curated captions for two instances in our dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.21, + 0.425 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Our experiments first ablate our model's contributions and loss design (Section 5.1), and evaluate our final model in personalized instance retrieval benchmarks (Section 5.2). Evaluation Datasets and Metrics. We evaluate our approach on two datasets: (i) our newly introduced This-IsMy personal video instance retrieval benchmark, and (ii) the personalized image retrieval benchmark built on DeepFashion2 [10] proposed by [5]. We evaluate retrieval performance in two settings:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.645 + ], + "angle": 0, + "content": "1) Generic Instance Retrieval: In this case, we build queries for learned instances using a generic prompt, i.e., \"An image of *\", and measure retrieval performance using mean Average Precision (mAP) and Mean Reciprocal Rank (MRR). In this setting, there are multiple correct matches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.734 + ], + "angle": 0, + "content": "2) Contextualized Instance Retrieval: In this case, we use natural language queries that describe the personal instance in a specific context, e.g., using a scene description such as \"A photo of * lying on the beach\". In this case, we assume there is only a single correct match, and we measure performance with MRR and Recall-at-5 (R@5)." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We use the ViT-B/16 version of CLIP in most of our experiments if not otherwise indicated. All the learnable parameters \\( z_{i} \\) and \\( C_l \\) are randomly initialized from \\( \\mathcal{N}(0,0.1) \\). We set the number of category features to \\( q = 512 \\) and the number of instance tokens to \\( n_w = 1 \\) by default. No data augmentation is used during training (the visual embeddings thus have to be computed just once). For meta-personalization of category features \\( C_l \\), we randomly select 32 named instances from \\( \\mathcal{D} \\) that are 0-shot classified to each category \\( l \\) and train for 20 epochs. This process is repeated 10 times, re-initializing \\( z_{i} \\) each" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Table 1. Ablation Experiments. We verify our model and training objective design through ablations on This-Is-My and DeepFashion2. We report personal instance retrieval performance in terms of mAP and MRR (higher is better)." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.147, + 0.88, + 0.277 + ], + "angle": 0, + "content": "
AblationThis-Is-MyDeepFashion2
mAPMRRmAPMRR
a) w/o meta-pers.54.1±1.383.1±2.135.2±0.555.2±1.9
b) single C55.5±0.486.9±0.544.1±0.564.7±1.5
c) w/o Ll53.9±1.086.4±2.247.3±0.868.3±1.2
d) w/o Lc48.9±1.078.6±2.447.1±0.568.4±0.8
e) N = B47.4±0.473.7±1.8--
f) w/o pre-trained C53.0±0.785.1±3.044.3±0.867.1±1.0
Ours56.4±0.687.4±1.247.3±0.769.9±0.8
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.429 + ], + "angle": 0, + "content": "time while retaining \\( C_l \\) from the previous run. Although updating the parameters of \\( C_l \\) to a small set of instances \\( \\mathcal{P} \\) at test time is beneficial, there is a risk of overfitting when only a few or just a single instance of each category is provided. To mitigate this, we include additional instances from \\( \\mathcal{D} \\) with the same categories as contained in \\( \\mathcal{P} \\) during test time personalization. All our results are averaged over five runs, each with different random seeds, and we report the standard error for each experiment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.439, + 0.611, + 0.455 + ], + "angle": 0, + "content": "5.1. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.509 + ], + "angle": 0, + "content": "In Table 1 we report results in the generic instance retrieval setting for the following ablations to illustrate the effect of our model and training objective design:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.584 + ], + "angle": 0, + "content": "a) w/o meta-personalization: In this case, we do not learn a meta-personalized \\( C \\) and instead directly learn instance embeddings \\( w_{i} \\). This ablation demonstrates that learning global category attributes through meta-personalization improves generalization at test-time personalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.659 + ], + "angle": 0, + "content": "b) single \\( C \\) (shared for all categories): In this experiment, we learn only a single category matrix \\( C \\), which is shared among all the categories. The results demonstrate that learning separate attributes per category is better than sharing global attributes among all categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.719 + ], + "angle": 0, + "content": "c) \\(\\mathbf{w} / \\mathbf{o}\\mathcal{L}_l\\): We explore the influence of the language-language contrastive loss \\(\\mathcal{L}_l\\). The benefits of \\(\\mathcal{L}_l\\) are more pronounced in This-Is-My, where instances belong to different categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.81 + ], + "angle": 0, + "content": "d) \\(\\mathbf{w} / \\mathbf{o}\\mathcal{L}_c\\): We analyze the contribution of the category-anchoring loss \\(\\mathcal{L}_c\\). Including \\(\\mathcal{L}_c\\) is important on This-Is-My, likely because it keeps the learning focused on the category and prevents it from capturing other scene features. This is less the case on DeepFashion2, where images are very object-centric, and there is less background variety." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "e) \\(\\mathcal{N} = \\mathcal{B}\\) (w/o non-instance segments): We do not use non-instance segments \\(\\mathcal{N}\\) (i.e., segments of the same video but low vision-language similarity) as additional negatives in this ablation. Including \\(\\mathcal{N}\\) is important, as it can help prevent the instance embedding from capturing non-instance nuisance features from the scene." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19129" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.093, + 0.192, + 0.105 + ], + "angle": 0, + "content": "Language Queries" + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.093, + 0.414, + 0.105 + ], + "angle": 0, + "content": "Top-5 Personalized Retrievals" + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.118, + 0.176, + 0.144 + ], + "angle": 0, + "content": " is eating food with a white plate" + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.158, + 0.176, + 0.183 + ], + "angle": 0, + "content": "man is leading \\(< \\)Zak's dog Coffee> with a dog leash" + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.197, + 0.189, + 0.23 + ], + "angle": 0, + "content": " is with a computer monitor showing 1:12 or the screen" + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.242, + 0.174, + 0.275 + ], + "angle": 0, + "content": "a man is putting into a black case near red cases" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.118, + 0.189, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.111, + 0.468, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.154, + 0.468, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.196, + 0.468, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.242, + 0.468, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.468, + 0.336 + ], + "angle": 0, + "content": "Figure 6. Contextualized This-Is-My Retrievals. We show personalized query-time retrievals for four This-Is-My instances. Search prompts are shown on the left and correct retrievals are highlighted in green." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.468, + 0.414 + ], + "angle": 0, + "content": "f) w/o pre-trained \\( C \\): In this case, we randomly initialize \\( C \\) instead of using a meta-personalized \\( C \\). Compared to a) the sharing of category attributes among fashion items in DeepFashion2 shows clear benefits. This is less the case for This-Is-My, where instances belong to different categories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.425, + 0.438 + ], + "angle": 0, + "content": "5.2. Personal Instance Retrieval Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.468, + 0.568 + ], + "angle": 0, + "content": "Baselines. To demonstrate the effectiveness of our approach, we compare it to prior work [5] and the following strong baselines: CLIP (visual), which represents each instance as the average visual embedding of the corresponding training examples, CLIP (language), which represents each instance with a 0-shot prompt embedding of the corresponding category name, and CLIP \\((\\mathbf{V} + \\mathbf{L})\\) , which uses the average of the visual and language embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Named Video Instance Retrieval on This-Is-My. We evaluate our approach on the 15 named instances in our This-Is-My test-time personalization dataset \\(\\mathcal{P}\\). To learn personal instance tokens, we train on video shots that occur in the video where the instance was named. All other video shots (belonging to other videos) are in the retrieval corpus for evaluation. Furthermore, we also include all other shots that do not contain the instance in the retrieval corpus as distractors. For contextualized retrieval we use the manually collected text captions from the query-time dataset \\(\\mathcal{Q}\\) as queries. We compare against the baselines and report contextualized retrieval performance on the left and generic instance retrieval performance on the right of Table 2. Our model clearly outperforms the baselines in both settings. Interestingly, the visual and language baselines have opposing advantages in generic vs. contextualized retrieval, while our model performs well in both settings. Qualitative retrieval results are shown in Figures 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Fashion Item Retrieval on DeepFashion2. We follow the setup of [5] and consider DeepFashion2 [10] for personalized fashion item retrieval. The dataset in this setting consists of 653 training and 221 evaluation images across 50" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Table 2. This-Is-My Video Instance Retrieval Task. We report personal instance retrieval performance in language-specified contexts (e.g., \"* catching a pink frisbee\") on the left and generic instance retrieval (e.g., \"An image of *\") on the right." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.147, + 0.868, + 0.251 + ], + "angle": 0, + "content": "
MethodContext. Retr.Generic Retr.
MRRR@5mAPMRR
Random0.0±0.00.0±0.01.1±0.22.2±1.3
CLIP (language)30.8±0.036.7±0.016.6±0.044.2±0.0
CLIP (visual)10.3±1.012.0±1.648.0±0.675.0±3.2
CLIP (V+L)20.9±1.623.3±2.151.7±0.481.9±0.0
Ours42.0±1.350.7±1.856.4±0.687.4±1.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.255, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Table 3. Fashion Item Retrieval on DeepFashion2. We evaluate our approach on the personalized instance retrieval task with contextualized queries from [5] (left) and generic instance retrieval (right). Results with * use ViT-B/32 instead of ViT-B/16." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.316, + 0.869, + 0.469 + ], + "angle": 0, + "content": "
MethodContext. Retr.Generic Retr.
MRRR@5mAPMRR
Random2.9±0.21.8±0.54.7±0.59.5±2.0
CLIP (language)21.2±0.025.3±0.08.3±0.016.9±0.0
CLIP (visual)14.2±0.317.3±0.320.6±0.443.7±1.1
CLIP (V+L)20.8±0.825.4±1.720.5±0.542.8±1.2
Adapter* [5]5.9±0.7---
COLLIE* [5,29]7.9±0.7---
PALAVRA* [5]28.4±0.739.2±1.3--
Ours*34.4±0.745.2±1.140.0±1.069.3±1.8
Ours38.4±0.451.4±0.453.4±0.477.7±0.6
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.696 + ], + "angle": 0, + "content": "instances (i.e., unique fashion items). While our focus is on video retrieval, our model and training objective can be used for image retrieval with minimal adjustments. We perform meta-personalization by pre-training on other non-instance-labeled images of DeepFashion2 (using only a single training image per instance). For a fair comparison with [5], the instance tokens are learned by randomly sampling \\( k = 5 \\) images per instance for training. We compare to results from [5] and the baselines in Table 3. Results using contextualized queries provided by [5] are on the left, and generic instance retrieval performance on the right. Note that [5] does not provide results for generic instance retrieval so we leave these two columns empty (-). Our approach outperforms the baselines and prior work by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.619, + 0.725 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We have introduced a meta-personalization approach for learning to retrieve named instances in video given a natural language query. We demonstrated the effectiveness of our approach on two datasets and showed that it outperforms strong baselines. Our effort is a step towards vision-language models trained over a large number of general and personal concepts. Our approach opens up the possibility of personalized language-based video editing, summarization, and generation, e.g., identify the key instances in a collection of footage and create an edited story about the instances using natural language." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19130" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.253 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[2] Fernando Amat, Ashok Chandrashekar, Tony Jebara, and Justin Basilico. Artwork personalization at netflix. In Proceedings of the 12th ACM conference on recommender systems, pages 487-488, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.366 + ], + "angle": 0, + "content": "[3] Soulef Benhamdi, Abdesselam Babouri, and Raja Chiky. Personalized recommender system for e-learning environment. Education and Information Technologies, 22(4):1455-1477, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[4] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[5] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. “This is my unicorn, Fluffy”: Personalizing frozen vision-language representations. In ECCV, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.52 + ], + "angle": 0, + "content": "[6] Chelsea Finn, P. Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[7] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In International Conference on Machine Learning, pages 1920-1930. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. ArXiv, abs/2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.705, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[10] Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaou Tang, and Ping Luo. Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.79, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[11] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5006-5015, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[12] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[13] De-An Huang, S. Buch, Lucio M. Dery, Animesh Garg, Li Fei-Fei, and Juan Carlos Niebles. Finding \"it\": Weakly-supervised reference-aware visual grounding in instructional videos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5948–5957, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.894, + 0.286 + ], + "angle": 0, + "content": "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[15] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv preprint arXiv:1909.12488, 2019.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[16] Alexander Kunitsyn, Maksim Kalashnikov, Maksim Dzabraev, and Andrei Ivaniuta. Mdmmt-2: Multidomain multimodal transformer for video retrieval, one more step towards generalization. arXiv preprint arXiv:2203.07086, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[17] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[18] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[19] Shuang Li, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, and Igor Mordatch. Composing ensembles of pretrained models via iterative consensus. arXiv preprint arXiv:2210.11522, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[20] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[21] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[23] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. Technical report, Tech. Rep., Technical report, OpenAI, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19131" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[25] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[27] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[28] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[29] Gabriel Skantze and Bram Willemsen. Collie: Continual learning of language grounding from language-image embeddings. Journal of Artificial Intelligence Research, 74:1201-1223, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[30] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[31] Yu Sun, X. Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[32] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and D. Erhan. Phenaki: Variable length video generation from open domain textual description. ArXiv, abs/2210.02399, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[33] Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, and Juan Miguel Pino. Fairoq s2t: Fast speech-to-text modeling with faireq. In AACL, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[34] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. ArXiv, abs/2109.01087, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[35] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[36] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[37] Sangmin Woo, Jinyoung Park, Inyong Koo, Sumin Lee, Minki Jeong, and Changick Kim. Explore and match: End-to-end video grounding with transformer. arXiv preprint arXiv:2201.10168, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[39] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[40] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[41] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[42] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[43] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[44] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19132" + } + ] +] \ No newline at end of file diff --git a/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_origin.pdf b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..603e03b14d07c5246b7f37fd4b58d2e83a7e84c6 --- /dev/null +++ b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/7dfac407-7dc4-48cc-9bb7-f23ca54c68cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9e2d3492aee23e6e0afb9eb2426fd2d62096d2cf4c6f7a27e639471bf8ae580 +size 5440634 diff --git a/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/full.md b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e94b232a1ff79cf92ab1f4d727bee11619588fa0 --- /dev/null +++ b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/full.md @@ -0,0 +1,315 @@ +# Meta-Personalizing Vision-Language Models to Find Named Instances in Video + +Chun-Hsiao Yeh $^{1,3*}$ Bryan Russell $^{3}$ Josef Sivic $^{2,3}$ Fabian Caba Heilbron $^{3\dagger}$ Simon Jenni $^{3\dagger}$ + +1University of California, Berkeley 2CIIRC CTU 3Adobe Research + +daniel.yeh@berkeley.edu {brussell, inr03127, caba, jenni} @adobe.com + +# Abstract + +Large-scale vision-language models (VLM) have shown impressive results for language-guided search applications. While these models allow category-level queries, they currently struggle with personalized searches for moments in a video where a specific object instance such as "My dog Biscuit" appears. We present the following three contributions to address this problem. First, we describe a method to meta-personalize a pre-trained VLM, i.e., learning how to learn to personalize a VLM at test time to search in video. Our method extends the VLM's token vocabulary by learning novel word embeddings specific to each instance. To capture only instance-specific features, we represent each instance embedding as a combination of shared and learned global category features. Second, we propose to learn such personalization without explicit human supervision. Our approach automatically identifies moments of named visual instances in video using transcripts and vision-language similarity in the VLM's embedding space. Finally, we introduce This-Is-My, a personal video instance retrieval benchmark. We evaluate our approach on This-Is-My and DeepFashion2 and show that we obtain a $15\%$ relative improvement over the state of the art on the latter dataset. + +# 1. Introduction + +The recent introduction of large-scale pre-trained vision-language models (VLMs) has enabled many new vision tasks, including zero-shot classification and retrieval [14, 18, 22], image/video generation [12, 24, 25, 27, 28, 32], or language-guided question answering [1, 19, 42]. It is now possible to search not only for specific object categories (e.g., dogs) but also for more specific descriptions of both + +Figure 1. Meta-Personalized Vision-Language Model (VLM) to Retrieve Named Instances in Video. Given a video where a user-specific instance, e.g., "My dog Biscuit" is mentioned, our method automatically learns a representation for the user-specific instance in the VLM's text input space. The personalized VLM can then be used to retrieve the learned instance in other contexts through natural language queries, e.g., grabbing a pink frisbee. This result is enabled by metapersonalizing the VLM on a large-scale dataset of narrated videos by pre-learning shared global category tokens (in this example for the category of 'dogs'), which are then easily personalized to user-specific instances from only a few user-given training examples. +![](images/48552c1d4940bd8f73574710c34c6c90dea7d045f9fbe7695919b460665f9156.jpg) +...create recipes to them...This is my dog Biscuit...If you teach them a trick... + +# Input Video with Transcript + +the object and scene attributes (e.g., "A small white dog playing at the dog park"). However, we often do not want to search for just any example of a generic category but instead to find a specific instance. For example, a user might want to search their personal video library for all the scenes that show their dog "Biscuit grabbing a pink frisbee", as illustrated in Figure 1. Since VLMs do not have a representation of "Biscuit," such queries are beyond the capabilities + +of off-the-shelf VLMs. + +Recent work [5] proposed a method to extend the language encoder's vocabulary with a newly learned token that represents a specific personal instance to address this issue. While this approach enables language-guided search for personal instances by placing the learned tokens in the query prompt, their solution assumes a collection of manually annotated images showing the individual instance in various contexts for successful token learning. For this approach to work in practice, a user must manually annotate all their important personal instances in various contexts, such that the instance representation does not capture nuisance features, e.g., the background. We thus identify two key challenges: 1) collecting personal instance examples without explicit human labeling and 2) learning a generalizable object-centric representation of personal instances from very few examples. + +The contributions of this work are three-fold. As our first contribution, we propose a method to automatically identify important personal instances in videos for personalizing a vision-language model without explicit human annotations. Indeed, people often record and refer to personal items or relationships in videos found online. Our approach is thus to identify mentions of personal instances in a video automatically and leverage these moments to build a set of personal instances for training. To this end, we extract the transcripts of videos using speech-to-text models and find candidate moments by looking for occurrences of "this is my *" or similar possessive adjective patterns. The symbol * in this example could represent a single word or sequence of words describing the instance (e.g., *= "dog Biscuit"). We then use vision-language similarity to filter non-visual examples and to find additional occurrences in the video for training. For example, we found more than six thousand named instances in 50K videos randomly sampled from the Merlot Reserve dataset [44]. We call the resulting collection of named instances in videos the This-Is-My dataset. + +As our second contribution, we propose a novel model and training procedure to learn text tokens representing the named instances in video from possibly very few and noisy training examples. Our method represents each instance with learned tokens and models each token as a linear combination of a set of pre-learned category-specific features shared across different instances. This set of shared category-specific features (similar to object attributes) improves the generalization of our method by preventing the instance representations from capturing nuisance features (e.g., the scene background). Furthermore, we show how to pre-train and adapt the shared category features using a large set of automatically collected This-Is-My examples, further improving our model's few-shot personalization performance at test-time. We call this pre-training of shared category features meta-personalization. In contrast to prior + +work [5], our method does not require training additional neural network models and requires only the optimization of a contrastive learning objective. + +As our final contribution, we demonstrate and evaluate our model on an existing fashion item retrieval benchmark, DeepFashion2, and our new challenging This-Is-My video instance retrieval dataset4 depicting specific object instances across different videos and contexts. Our experiments demonstrate that our method outperforms several baselines and prior approaches on these challenging language-guided instance retrieval tasks. + +# 2. Related Work + +Vision-Language Models for Video Retrieval. Vision-language foundational models [14, 18, 22] have been successful for zero-shot and other diverse video tasks, such as video question answering [19, 42], language-video grounding [13, 37, 41], and text-to-video retrieval [11, 16, 17, 20, 21, 39, 40]. These models have a powerful representation that transfers well to the video domain to achieve competitive performance on video-language tasks. Our approach builds on these powerful representations to retrieve specific named instances in video. + +Personalized Concept Learning. Adapting a model to learn a user-specific representation has been a significant topic in machine learning research, including recommendation systems [2, 3] and federated learning [15]. Relevant to us are recent approaches for adapting vision-language models to object instances. PALAVRA [5] proposes a learning scheme that appends a learnable token for a new personalized concept to the token embedding of the input text prompt. This learned representation helps to preserve the personalized concept. DualPrompt [36] introduces a framework that learns a small set of prompts to emphasize more specific concepts without forgetting the learned concepts in the pre-trained foundational model. There have also been works that have extended this personalized concept learning to image generation [4,8,26]. While these approaches adapt a vision-language model to personal instances, they perform the adaptation independently for each instance and do not "meta-train" for the personalization task for improved few-shot learning as we do in this work. + +Fine-tuning and Test-time Adaptation. Fine-tuning is a common strategy to adapt a pretrained model for a specific downstream task by transferring the source model to a target domain. Recent works on vision-language model tuning include CLIP-Adapter [9] that proposes to conduct fine-tuning with an extra bottleneck layer while freezing the pre-trained CLIP model. WiSE-FT [38] resolves the distribution shift caused by fine-tuning and ensembles the weights of the original and fine-tuned model to increase ro + +![](images/4ee01b6b95138c0199ecf6e9c38be9b8baf03bdd861b048b8471a721a4d48fc4.jpg) +Figure 2. Overview of our Personalized Vision-Language Model. Our model augments a frozen VLM (blue) with novel personal instance tokens $w = Cz$ (orange) that are a combination of global category features $C$ (green) with instance-specific weights $z \in \mathcal{Z}$ (yellow). Our approach for personalized instance retrieval has three stages. First, we pre-learn global category features $C_{\mathcal{D}}$ on a large set of automatically mined named personal instances in videos. We call this process Meta-Personalization (left). In the second step (middle), we adapt the meta-personalized category features $C_{\mathcal{D}}$ at test-time and learn novel instance weights $z \in \mathcal{Z}_{\mathcal{P}}$ to represent a user's personal instances via $w = C_{\mathcal{P}}z$ . Finally (right), we leverage the (frozen) personalized instance tokens $w$ in natural language queries at query time. + +![](images/948389672c9dc1d540297882241a97776f06808533f67166442f166d11e9a93f.jpg) + +![](images/9646bc818d22e4bfaf9d795838662cfb13340f332f3b72d0ca3613d84f3b7a0b.jpg) + +bustness. Prior test-time adaption works [31, 34, 35] fine-tune on target data without information from the source data. Our approach leverages test-time adaptation for updating our meta-personalized model to user-specific data. + +Meta-Learning. We draw inspiration from meta-learning ("learning to learn") [6, 7, 43], which enables models to quickly adapt to new tasks by learning on a diverse set of tasks. Hence, given only a handful of novel training examples, the model can be adapted to novel tasks. Our approach borrows the idea of meta-learning to meta-personalize a model by learning global category features from a large video database. We then adapt the global features during test-time training from only few examples of user-specific instances to enable query-time retrieval. + +# 3. (Meta-)Personalization of Named Instances + +Our goal is to learn representations of personal items in video that enable retrieval through natural language queries. To achieve this goal, we have to address the key challenge of adapting a model with one or few examples of a named instance. To address this challenge, we propose a metapersonalization approach that learns to personalize given a large corpus of named instances mined from videos with transcriptions. We illustrate our approach in Figure 2. + +In the first step (Figure 2 (left)), we mine automatically a collection $\mathcal{D}$ of named instances from videos with transcriptions for meta-personalization. We use this collection to train a proposed model $\mathcal{M}_{C,z}$ that includes global category features $C$ and instance-specific parameters $z$ . The global category features $C$ are lightweight and shared across all + +instances. Given a natural language query $u$ and video $v$ , the model returns a score $\mathcal{M}_{C,z}(u,v)$ . During metapersonalization, given a training loss $\mathcal{L}$ , we jointly optimize the loss over the global category features $C$ and instance parameters $\mathcal{Z}$ for each named instance in the collection $\mathcal{D}$ , + +$$ +\left(C _ {\mathcal {D}}, \mathcal {Z} _ {\mathcal {D}}\right) \in \underset {(C, \mathcal {Z})} {\arg \min } \sum_ {z \in \mathcal {Z}} \mathcal {L} (C, z). \tag {1} +$$ + +Note that here the instance-specific parameters $\mathcal{Z}_{\mathcal{D}}$ learnt via (1) are discarded while the global category features $C_{\mathcal{D}}$ are kept as the meta-personalized part of the model. The global category features $C_{\mathcal{D}}$ capture information shared across instances relevant to the personalization task. + +In the second step (Figure 2 (middle)), we are given a set of named video instances $\mathcal{P}$ (e.g., automatically mined from someone's personal video library) and wish to perform test-time personalization of the model to this person's instances. Here each instance is represented by only one or few examples. In this step, we optimize the training loss $\mathcal{L}$ over the global category features $C$ and the set of instance parameters $\mathcal{Z}$ for all instances in the personal set $\mathcal{P}$ starting from the pre-trained global category features $C_{\mathcal{D}}$ and random $\mathcal{Z}$ . We obtain the personalized model parameters $(C_{\mathcal{P}},\mathcal{Z}_{\mathcal{P}})$ as + +$$ +\left(C _ {\mathcal {P}}, \mathcal {Z} _ {\mathcal {P}}\right) \in \underset {(C, \mathcal {Z})} {\arg \min } \sum_ {z \in \mathcal {Z}} \mathcal {L} (C, z). \tag {2} +$$ + +We now keep both $C_{\mathcal{P}}$ and $\mathcal{Z}_{\mathcal{P}}$ + +In the final step (Figure 2 (right)), we perform retrieval over a potentially large dataset using the test-time personalized model $\mathcal{M}_{C_{\mathcal{P}},z}$ (where $z \in \mathcal{Z}_{\mathcal{P}}$ ). We next describe + +![](images/5b74e414fc093a921d5a9a1dab87f549dd0b73bcd07430e8ac1ab6448e74f7a6.jpg) +visual similarity for every pair + +![](images/788d6595b4af62523175399336e012db3ab4b6061e6bca806ce0bc5fb42f9c12.jpg) +0.2 + +![](images/70e7dfee7a79237d126b138c718afcd2bf12d62f224a0755dc23350ec8400c55.jpg) +set of candidates + +![](images/53a57fea915bc7aa2aff6185f49d5b588c64b32e4ea128c6bb174f4939f10fd3.jpg) +Finding Additional Instance Examples (Step 3) + +![](images/6426d3a23890796359acef37c8a38cabc8f06c3bf07c44323719e5560fc3bdc2.jpg) + +![](images/d3b20ae614e3ac354e5e83ebb7e136378a311f5f2ff1fb9cb2ecedb23e6aa7e6.jpg) +instance name: Time to talk about + +![](images/2606f2a888dd299fe0f41072ec5e46716a45e88bb51f6d1b07e7f732ba7f4e68.jpg) +visual relevance: 0.1 +instance name: fender guitar + +![](images/46fd0c6eaaa5ee6a60bca5e18a3fc897a75e17f33f1f70dd467a51c2cff0308b.jpg) + +![](images/61adb1288277139d8ca03dc56ef128da6e74185220e114f3bce1717d1ccef932.jpg) +Filtering Non-visual Instances (Step 2) + +![](images/67903db5e0f75d42b56b19ad4f5a062b5cb2246b2123e95a3342eb21b46de532.jpg) +Video: +Transcript: This is our time to talk about ... This is my fender guitar +Spotting Named Instances (Step 1) +Figure 3. Automatic mining of named instances in video for meta-personalization. Our automatic mining pipeline includes three steps (from bottom to top). Step 1 finds named instances via string-matching of possessive patterns in video transcripts. Step 2 filters non-visual instances using text-to-visual relevance between the instance name and the video shots neighboring the named instance. Finally, Step 3 retrieves additional shots with high visual similarity to the instance reference shot. + +how we automatically mine the named instances in video $\mathcal{D}$ for meta-personalization (Sec. 3.1), and our full model $\mathcal{M}$ for query-time retrieval and loss $\mathcal{L}$ for meta-personalization and test-time personalization (Sec. 3.2). + +# 3.1. Automatic Mining of Named Instances in Video + +To identify personal instances without explicit supervision, we leverage a collection of videos from the web along with their corresponding time-aligned transcripts. These transcripts can be automatically generated through speech-to-text models [23, 33]. We now describe a procedure to mine these data for a set of moments (i.e., a collection of video shots) depicting a referred personal instance in the transcript without the need for manual annotation. We will use these moments for training a meta-personalization model (Sec. 3.2). + +Spotting Named Instances. Here our goal is to find moments where candidate personal instances are mentioned in videos. We do so by searching for possessive text patterns5 such as "This is my *" in a corpus of time-aligned video transcripts. This string-matching process outputs a list of candidate instance names * associated with video timestamps $t^*$ . In prac + +tice, we keep up to four words after a possessive text pattern is matched based on text-visual similarity (see supp. for details). That way, we can retrieve simple named instances such as This is my dog ( $=$ dog) but also complex ones like This is my favorite CHANEL classic handbag ( $=$ CHANEL classic handbag). Note also that a single video might include multiple string matches; for instance, the example video illustrated in Figure 3 (Step 1) includes two matches: "This is our time to talk about" at time 1:30 and "This is my fender guitar" at time 3:25. + +Filtering Non-visual Instances. The previous spotting step only searches for potential instances using the transcript, yielding many string matches that are non-visual, i.e., the strings do not describe the visible content in the video. Here we aim to filter out these non-visual instances. We do so by computing the text-to-visual relevance between the instance name (e.g., fender guitar) and the neighboring shots around the time when the instance is mentioned. We add neighboring shots to cover cases where the named instances are shown just before or after they are mentioned. Concretely, given a sequence of $m$ video shots $S = [s_1, \ldots, s_m]$ automatically extracted with [30], we find the shot $s_{t^*}$ that overlaps with $t^*$ (the time when the instance was mentioned). Next, we form a set of candidate visual references $S_{t^*} = [s_{t^* - 1}, s_{t^*}, s_{t^* + 1}]$ comprising a window of shots that are previous and subsequent to $s_{t^*}$ . We then compute text-to-visual relevance scores using CLIP [22] encoders. This encoding process yields $L_2$ -normalized embeddings $f_l(*)$ for the named instance and $f_v(s_i)$ for each shot $s_i \in S_{t^*}$ . We compute $f_v(s_i)$ by averaging the visual embeddings of all frames in the corresponding shot. Finally, we compute the cosine similarity between every $(f_l(*), f_v(s_i))$ pair and retain a visual reference shot $s^*$ if the highest cosine similarity is greater than 0.3. This filtering step outputs a cleaned set of named instances with a corresponding visual reference. Figure 3 (Step 2) illustrates how we prune out non-visual matches such as time to talk about. In contrast, visual instances such as fender guitar are kept and matched with a visual reference. + +Finding Additional Instance Shots. Since frames from a single video shot provide only limited variability in the instance appearance and could thus limit the learning, we aim to recover other shots from the video where that instance appears. We leverage CLIP's visual encoder to compute the visual similarity between the instance's reference shot $s^*$ and every shot $s_i \in S$ . We extract an embedding $f_v(s^*)$ for the reference shot $s^*$ and an embedding $f_v(s_i)$ for each shot $s_i$ . Similar to the non-visual filtering step, we average the CLIP embeddings of each frame belonging to a shot. Then, we compute the cosine similarity between the embeddings for the reference shot and every candidate shot in the video. We keep the shots whose cosine similarity with the refer + +![](images/e8379900235a8874a3fddede6276fb59ec0913edffd3118dabffc64be6c38eb3.jpg) +$\mathcal{M}_{\mathcal{C},z}(\pmb {u},\pmb {v})$ +Figure 4. Model Overview. Our model $\mathcal{M}_{C,z}$ extends CLIP's language input vocabulary with $n_w$ novel instance-specific tokens $w_i^y = C_l z_i^y$ , which we model as a linear combination of metapersonalized category features $C_l$ with weights $z_i^y$ . Note that the vision and language encoders are frozen during this process. + +ence is greater than 0.9. Figure 3 (Step 3) illustrates the output of this final step in the automatic mining pipeline. Our mining algorithm allows us to retrieve additional shots for the instance fender guitar. Now the instance examples not only include a clean close-up of the guitar, but also shots where the guitar is being played or held by its owner. + +# 3.2. Learning Personal Instance Representations + +The result of the mining procedure described above is a dataset consisting of a set of video shots $\mathcal{D} = \{s_1, \ldots, s_n\}$ and corresponding instance IDs $Y = \{y_1, \ldots, y_n\}$ , where $y_i = y_j$ if $s_i$ and $s_j$ are video shots that are assumed to contain the same instance. We now describe how we use these data to learn representations of the collected instances. + +Our approach is to leverage a large-scale pre-trained vision-language model (CLIP [22]) and augment its language encoder with a set of novel personal instance tokens. Let $f_{v}(s)$ be the output of the visual encoder for shot $s$ (computed as the average over the frame embeddings) and let $f_{l}(u)$ be the output of the language encoder for a natural language input $u = [v_{1},\dots ,v_{m}]$ of $m$ token embeddings, where $v_{i}\in \mathbb{R}^{d}$ denote learned token embeddings of the pretraining vocabulary (positional embeddings are included but omitted from the notation). We propose to extend this vocabulary with novel personal instance tokens. Concretely, our approach introduces a set $w^{y} = \{w_{i}^{y}\}_{i = 1}^{n_{w}}$ of $n_w$ new tokens that represent a personal instance $y\in Y$ . + +To learn these tokens and to perform personalized retrieval at test time, we construct natural language person + +alized queries as + +$$ +\hat {u} ^ {p} = \left[ p _ {1}, \dots , p _ {k - 1}, w _ {1} ^ {y}, \dots , w _ {n _ {w}} ^ {y}, p _ {k + 1}, \dots , p _ {m} \right], \tag {3} +$$ + +where $p_i$ are token embeddings of a chosen prompt $p = [p_1, \ldots, p_m]$ . During training, the prompt $p$ corresponds to a random template of the form [An image of *, [* can be seen in this photo], [There is * in this image], etc., and $k$ denotes the position of the * placeholder for the instance tokens. + +Instances as Combinations of Category Features. Since learning personal instance tokens from possibly very few examples runs the danger of overfitting (e.g., to nuisance features in the background), we propose to parameterize them as + +$$ +w _ {i} ^ {y} = C _ {l} z _ {i} ^ {y} \in \mathbb {R} ^ {d \times 1}, \tag {4} +$$ + +where $z_{i}^{y} \in \mathbb{R}^{q \times 1}$ is a vector of learnable weights specific to each instance and $C_{l} \in \mathbb{R}^{d \times q}$ is a matrix of learnable global category features, which are shared for all instances belonging to the same object category $l \in \mathcal{V}$ (e.g., $\mathcal{V} = \{\text{car}, \text{person}, \text{dog}, \ldots\}$ ) and constitute the set $C_{\mathcal{D}} = \{C_{l}\}_{l \in \mathcal{V}}$ . We illustrate the model in Figure 4. + +We can think of the columns of $C_l$ as representing a set of shared category features and the final instance token $w_i^y$ as a linear combination of these features with weights $z_i^y$ . Our aim is that only category-specific features are captured during training and irrelevant features (e.g., about the background) are discarded. Intuitively, the columns of $C_l$ could correspond to attributes of an object category, e.g., if we were to learn "car" features, these could capture their color, brand, type, or age, to name a few. To identify which category matrix $C_l$ to use for an instance $y$ , we rely on 0-shot classification using vision-language similarity between instance shots and a generic prompt for the category $l \in \mathcal{V}$ . + +Contrastive Personal Token Learning. We propose a contrastive learning objective to learn the personal instance tokens $w^{y}$ using a set of video shots containing the instance $y$ . To this end, let $\psi_{i} := f_{v}(s_{i})$ be an encoding of a video shot $s_{i}$ , i.e., the average frame encoding of all frames in the shot, and let $\phi_{i} := f_{l}(\hat{u}_{i}^{p})$ denote the language encoding of a corresponding personalized query. We learn the novel tokens $w^{y}$ and shared category features $C_{l}$ by optimizing two contrastive objectives: a language-language contrastive objective $\mathcal{L}_{l}$ and a vision-language contrastive objective $\mathcal{L}_{vl}$ . The language-language objective is given by + +$$ +\mathcal {L} _ {l} = \sum_ {i \in \mathcal {B}} \sum_ {j \neq i \in \mathcal {B}} - 1 \left\{y _ {i} = y _ {j} \right\} \log \left(\frac {d \left(\phi_ {i} , \phi_ {j}\right)}{\sum_ {k \neq i \in \mathcal {B}} d \left(\phi_ {i} , \phi_ {k}\right)}\right), \tag {5} +$$ + +where $d(a,b)\coloneqq \exp \left(\frac{1}{\lambda}\frac{a^{\intercal}b}{\|a\|_{2}\|b\|_{2}}\right)$ measures the similarity between the feature vectors $a$ and $b$ , $\lambda = 0.1$ is a temperature parameter, and $\mathcal{B}$ is a randomly sampled training minibatch. The vision-language objective is similarly defined as + +![](images/60425bf28057a6954da847bfede9fb95f80f36c3ab27741bf639b951d1c9c3dd.jpg) +Figure 5. Examples from This-Is-My { Meta-Personalization $\mathcal{D}$ (top) vs Test-time personalization $\mathcal{P}$ (bottom-left) vs Query-time $\mathcal{Q}$ (bottom-right)} datasets. In the Query-time dataset (bottom-right), we design a challenging video instance retrieval task. For example, in (a) the named instance (i.e., Alex's piano) is in the background and is barely visible and in (b) the background scenes in the query-time dataset (bottom-right) are completely different from the test-time personalization dataset (bottom-left) depicting the same named instance. + +$$ +\mathcal {L} _ {v l} = \sum_ {i, j \in \mathcal {B}} - \mathbb {1} \left\{y _ {i} = y _ {j} \right\} \log \left(\frac {d \left(\phi_ {i} , \psi_ {j}\right)}{\sum_ {k \in \mathcal {N}} d \left(\phi_ {i} , \psi_ {k}\right)}\right), \tag {6} +$$ + +with the set of negative examples $\mathcal{N}$ comprising both other examples in the batch $\mathcal{B}$ and non-instance shots from the videos containing the named instances, i.e., shots that have low vision-language similarity. The loss is low when the encodings for video shots and personalized queries with the same instance ID are more similar than to other queries. Including non-instance segments as negatives can help discard non-instance features such as scene background. + +To further constrain the learning toward category-specific attributes, we include a loss that maximizes the similarity between a personal instance query and a generic category query. Concretely, let $c_{l}$ be a category query embedding for category $l$ (e.g., "An image of a [car]") to which instance $y$ belongs. We then include the following category-anchoring loss + +$$ +\mathcal {L} _ {c} = - \sum_ {i \in \mathcal {B}} \frac {c _ {l} ^ {\top} \phi_ {i}}{\| c _ {l} \| _ {2} \| \phi_ {i} \| _ {2}}. \tag {7} +$$ + +To summarize, our training loss $\mathcal{L}$ (see Equations 1 and 2) for meta- and test-time personalization is given by + +$$ +\mathcal {L} = \mathcal {L} _ {l} + \mathcal {L} _ {v l} + \lambda_ {c} \mathcal {L} _ {c}, \tag {8} +$$ + +where $\lambda_{c} = 0.5$ , controls the amount of category anchoring. + +# 4. This-Is-My Dataset + +Our This-Is-My dataset comprises three subsets for metapersonalization, test-time personalization, and querying. We describe each subset next. + +Meta-Personalization Dataset $\mathcal{D}$ . We gather this subset with the automatic mining algorithm introduced in Section 3.1. We leverage the Merlot Reserve dataset [44], which contains more than 20 million videos from YouTube and their corresponding time-aligned transcripts. In practice, we start from a subset of $50K$ randomly sampled Merlot Reserve videos. We spot 6058 named instances using various text possessive templates. Our visual filtering step removes $52\%$ of instances, generating a total of 2908 named instances with a visual reference. Finally, we mine additional samples for each instance, yielding a total of 49256 instance samples. This subset includes a wide variety of visual concepts, ranging from common objects such as bikes to rare concepts such as toaster. While we design our mining pipeline to minimize noise, it is limited by CLIP's capability to distinguish between similar object instances. For instance, while we find several samples of the fender guitar instance, it includes other shots that do not correspond to the aforementioned guitar (Figure 5 (top)). Nevertheless, we will show empirically that this subset is still useful for meta-personalization purposes (Section 5). + +Test-time Personalization Dataset $\mathcal{P}$ . Our goal is to create a test dataset that recreates the scenario where a person wants to find when their personal instances appear in their video collection. We want to make this task close to the real scenario where a person records the same visual instance, e.g., their dog, across multiple places and situations. Our strategy is to emulate such a scenario by finding YouTube channels that frequently mention the same instance across multiple videos. While Merlot Reserve is large and diverse, only few channels are represented with more than one video in the dataset. Instead, we download all videos and automatic transcripts from the channels of 15 popular YouTube bloggers. We then run our mining algorithm, but manu + +ally supervise the steps for visual filtering and finding additional instance samples. We first manually verify that each instance name and its visual reference are good matches. Then, we find additional sample shots across all videos in the channel by ranking them according to their visual similarity to the instance reference shot. Finally, we review and label the top 1000-scored shots for each named instance. The supplementary material includes a screenshot of the annotation tool. In total, this subset includes 15 named instances with more than 686 labeled samples. Figure 5 (bottom left) shows two example instances in our dataset. + +Query-time Dataset $\mathcal{Q}$ . Our end goal is to retrieve named instances via natural language queries. We would like to be able to find videos when is grabbing a pink Frisbee. To this end, we manually caption 30 instance samples with descriptive queries. Thus, the Query-time dataset includes (manually captioned) video-caption pairs containing instances from the Test-time Personalization dataset. Figure 5 (bottom right) shows manually curated captions for two instances in our dataset. + +# 5. Experiments + +Our experiments first ablate our model's contributions and loss design (Section 5.1), and evaluate our final model in personalized instance retrieval benchmarks (Section 5.2). Evaluation Datasets and Metrics. We evaluate our approach on two datasets: (i) our newly introduced This-IsMy personal video instance retrieval benchmark, and (ii) the personalized image retrieval benchmark built on DeepFashion2 [10] proposed by [5]. We evaluate retrieval performance in two settings: + +1) Generic Instance Retrieval: In this case, we build queries for learned instances using a generic prompt, i.e., "An image of *", and measure retrieval performance using mean Average Precision (mAP) and Mean Reciprocal Rank (MRR). In this setting, there are multiple correct matches. +2) Contextualized Instance Retrieval: In this case, we use natural language queries that describe the personal instance in a specific context, e.g., using a scene description such as "A photo of * lying on the beach". In this case, we assume there is only a single correct match, and we measure performance with MRR and Recall-at-5 (R@5). + +Implementation Details. We use the ViT-B/16 version of CLIP in most of our experiments if not otherwise indicated. All the learnable parameters $z_{i}$ and $C_l$ are randomly initialized from $\mathcal{N}(0,0.1)$ . We set the number of category features to $q = 512$ and the number of instance tokens to $n_w = 1$ by default. No data augmentation is used during training (the visual embeddings thus have to be computed just once). For meta-personalization of category features $C_l$ , we randomly select 32 named instances from $\mathcal{D}$ that are 0-shot classified to each category $l$ and train for 20 epochs. This process is repeated 10 times, re-initializing $z_{i}$ each + +Table 1. Ablation Experiments. We verify our model and training objective design through ablations on This-Is-My and DeepFashion2. We report personal instance retrieval performance in terms of mAP and MRR (higher is better). + +
AblationThis-Is-MyDeepFashion2
mAPMRRmAPMRR
a) w/o meta-pers.54.1±1.383.1±2.135.2±0.555.2±1.9
b) single C55.5±0.486.9±0.544.1±0.564.7±1.5
c) w/o Ll53.9±1.086.4±2.247.3±0.868.3±1.2
d) w/o Lc48.9±1.078.6±2.447.1±0.568.4±0.8
e) N = B47.4±0.473.7±1.8--
f) w/o pre-trained C53.0±0.785.1±3.044.3±0.867.1±1.0
Ours56.4±0.687.4±1.247.3±0.769.9±0.8
+ +time while retaining $C_l$ from the previous run. Although updating the parameters of $C_l$ to a small set of instances $\mathcal{P}$ at test time is beneficial, there is a risk of overfitting when only a few or just a single instance of each category is provided. To mitigate this, we include additional instances from $\mathcal{D}$ with the same categories as contained in $\mathcal{P}$ during test time personalization. All our results are averaged over five runs, each with different random seeds, and we report the standard error for each experiment. + +# 5.1. Ablations + +In Table 1 we report results in the generic instance retrieval setting for the following ablations to illustrate the effect of our model and training objective design: + +a) w/o meta-personalization: In this case, we do not learn a meta-personalized $C$ and instead directly learn instance embeddings $w_{i}$ . This ablation demonstrates that learning global category attributes through meta-personalization improves generalization at test-time personalization. +b) single $C$ (shared for all categories): In this experiment, we learn only a single category matrix $C$ , which is shared among all the categories. The results demonstrate that learning separate attributes per category is better than sharing global attributes among all categories. +c) $\mathbf{w} / \mathbf{o}\mathcal{L}_l$ : We explore the influence of the language-language contrastive loss $\mathcal{L}_l$ . The benefits of $\mathcal{L}_l$ are more pronounced in This-Is-My, where instances belong to different categories. +d) $\mathbf{w} / \mathbf{o}\mathcal{L}_c$ : We analyze the contribution of the category-anchoring loss $\mathcal{L}_c$ . Including $\mathcal{L}_c$ is important on This-Is-My, likely because it keeps the learning focused on the category and prevents it from capturing other scene features. This is less the case on DeepFashion2, where images are very object-centric, and there is less background variety. +e) $\mathcal{N} = \mathcal{B}$ (w/o non-instance segments): We do not use non-instance segments $\mathcal{N}$ (i.e., segments of the same video but low vision-language similarity) as additional negatives in this ablation. Including $\mathcal{N}$ is important, as it can help prevent the instance embedding from capturing non-instance nuisance features from the scene. + +Language Queries +Top-5 Personalized Retrievals +![](images/325c0220828f2c93b9e3e989a7151786af47645a95998c682cbf6350a1f1be53.jpg) + is eating food with a white plate + +![](images/38ee14909a757f87ca57d54923524830cd3430f5ecaac3a00cf26b4ba2a8f4ee.jpg) +man is leading $<$ Zak's dog Coffee> with a dog leash + +![](images/204fd41855bab52074f33c7febc1597354df3685b5f3a3fb4bf85dc78d34b5f4.jpg) + is with a computer monitor showing 1:12 or the screen + +Figure 6. Contextualized This-Is-My Retrievals. We show personalized query-time retrievals for four This-Is-My instances. Search prompts are shown on the left and correct retrievals are highlighted in green. +![](images/6d3a9141a56783e276ad130c492e950e39ed0293ae554e978d35976fc4cd470b.jpg) +a man is putting into a black case near red cases + +f) w/o pre-trained $C$ : In this case, we randomly initialize $C$ instead of using a meta-personalized $C$ . Compared to a) the sharing of category attributes among fashion items in DeepFashion2 shows clear benefits. This is less the case for This-Is-My, where instances belong to different categories. + +# 5.2. Personal Instance Retrieval Benchmarks + +Baselines. To demonstrate the effectiveness of our approach, we compare it to prior work [5] and the following strong baselines: CLIP (visual), which represents each instance as the average visual embedding of the corresponding training examples, CLIP (language), which represents each instance with a 0-shot prompt embedding of the corresponding category name, and CLIP $(\mathbf{V} + \mathbf{L})$ , which uses the average of the visual and language embedding. + +Named Video Instance Retrieval on This-Is-My. We evaluate our approach on the 15 named instances in our This-Is-My test-time personalization dataset $\mathcal{P}$ . To learn personal instance tokens, we train on video shots that occur in the video where the instance was named. All other video shots (belonging to other videos) are in the retrieval corpus for evaluation. Furthermore, we also include all other shots that do not contain the instance in the retrieval corpus as distractors. For contextualized retrieval we use the manually collected text captions from the query-time dataset $\mathcal{Q}$ as queries. We compare against the baselines and report contextualized retrieval performance on the left and generic instance retrieval performance on the right of Table 2. Our model clearly outperforms the baselines in both settings. Interestingly, the visual and language baselines have opposing advantages in generic vs. contextualized retrieval, while our model performs well in both settings. Qualitative retrieval results are shown in Figures 6. + +Fashion Item Retrieval on DeepFashion2. We follow the setup of [5] and consider DeepFashion2 [10] for personalized fashion item retrieval. The dataset in this setting consists of 653 training and 221 evaluation images across 50 + +Table 2. This-Is-My Video Instance Retrieval Task. We report personal instance retrieval performance in language-specified contexts (e.g., "* catching a pink frisbee") on the left and generic instance retrieval (e.g., "An image of *") on the right. + +
MethodContext. Retr.Generic Retr.
MRRR@5mAPMRR
Random0.0±0.00.0±0.01.1±0.22.2±1.3
CLIP (language)30.8±0.036.7±0.016.6±0.044.2±0.0
CLIP (visual)10.3±1.012.0±1.648.0±0.675.0±3.2
CLIP (V+L)20.9±1.623.3±2.151.7±0.481.9±0.0
Ours42.0±1.350.7±1.856.4±0.687.4±1.2
+ +Table 3. Fashion Item Retrieval on DeepFashion2. We evaluate our approach on the personalized instance retrieval task with contextualized queries from [5] (left) and generic instance retrieval (right). Results with * use ViT-B/32 instead of ViT-B/16. + +
MethodContext. Retr.Generic Retr.
MRRR@5mAPMRR
Random2.9±0.21.8±0.54.7±0.59.5±2.0
CLIP (language)21.2±0.025.3±0.08.3±0.016.9±0.0
CLIP (visual)14.2±0.317.3±0.320.6±0.443.7±1.1
CLIP (V+L)20.8±0.825.4±1.720.5±0.542.8±1.2
Adapter* [5]5.9±0.7---
COLLIE* [5,29]7.9±0.7---
PALAVRA* [5]28.4±0.739.2±1.3--
Ours*34.4±0.745.2±1.140.0±1.069.3±1.8
Ours38.4±0.451.4±0.453.4±0.477.7±0.6
+ +instances (i.e., unique fashion items). While our focus is on video retrieval, our model and training objective can be used for image retrieval with minimal adjustments. We perform meta-personalization by pre-training on other non-instance-labeled images of DeepFashion2 (using only a single training image per instance). For a fair comparison with [5], the instance tokens are learned by randomly sampling $k = 5$ images per instance for training. We compare to results from [5] and the baselines in Table 3. Results using contextualized queries provided by [5] are on the left, and generic instance retrieval performance on the right. Note that [5] does not provide results for generic instance retrieval so we leave these two columns empty (-). Our approach outperforms the baselines and prior work by a large margin. + +# 6. Conclusion + +We have introduced a meta-personalization approach for learning to retrieve named instances in video given a natural language query. We demonstrated the effectiveness of our approach on two datasets and showed that it outperforms strong baselines. Our effort is a step towards vision-language models trained over a large number of general and personal concepts. Our approach opens up the possibility of personalized language-based video editing, summarization, and generation, e.g., identify the key instances in a collection of footage and create an edited story about the instances using natural language. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 1 +[2] Fernando Amat, Ashok Chandrashekar, Tony Jebara, and Justin Basilico. Artwork personalization at netflix. In Proceedings of the 12th ACM conference on recommender systems, pages 487-488, 2018. 2 +[3] Soulef Benhamdi, Abdesselam Babouri, and Raja Chiky. Personalized recommender system for e-learning environment. Education and Information Technologies, 22(4):1455-1477, 2017. 2 +[4] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 2 +[5] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. “This is my unicorn, Fluffy”: Personalizing frozen vision-language representations. In ECCV, 2022. 2, 7, 8 +[6] Chelsea Finn, P. Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 3 +[7] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In International Conference on Machine Learning, pages 1920-1930. PMLR, 2019. 3 +[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. ArXiv, abs/2208.01618, 2022. 2 +[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2 +[10] Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaou Tang, and Ping Luo. Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 7, 8 +[11] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5006-5015, 2022. 2 +[12] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben + +Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1 +[13] De-An Huang, S. Buch, Lucio M. Dery, Animesh Garg, Li Fei-Fei, and Juan Carlos Niebles. Finding "it": Weakly-supervised reference-aware visual grounding in instructional videos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5948–5957, 2018. 2 +[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2 +[15] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv preprint arXiv:1909.12488, 2019.2 +[16] Alexander Kunitsyn, Maksim Kalashnikov, Maksim Dzabraev, and Andrei Ivaniuta. Mdmmt-2: Multidomain multimodal transformer for video retrieval, one more step towards generalization. arXiv preprint arXiv:2203.07086, 2022. 2 +[17] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2 +[18] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1, 2 +[19] Shuang Li, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, and Igor Mordatch. Composing ensembles of pretrained models via iterative consensus. arXiv preprint arXiv:2210.11522, 2022. 1, 2 +[20] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2 +[21] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 2 +[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5 +[23] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. Technical report, Tech. Rep., Technical report, OpenAI, 2022. 4 +[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1 + +[25] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1 +[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2 +[27] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1 +[28] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 1 +[29] Gabriel Skantze and Bram Willemsen. Collie: Continual learning of language grounding from language-image embeddings. Journal of Artificial Intelligence Research, 74:1201-1223, 2022. 8 +[30] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 4 +[31] Yu Sun, X. Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 3 +[32] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and D. Erhan. Phenaki: Variable length video generation from open domain textual description. ArXiv, abs/2210.02399, 2022. 1 +[33] Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, and Juan Miguel Pino. Fairoq s2t: Fast speech-to-text modeling with faireq. In AACL, 2020. 4 +[34] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. ArXiv, abs/2109.01087, 2021. 3 +[35] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 3 +[36] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, 2022. 2 +[37] Sangmin Woo, Jinyoung Park, Inyong Koo, Sumin Lee, Minki Jeong, and Changick Kim. Explore and match: End-to-end video grounding with transformer. arXiv preprint arXiv:2201.10168, 2022. 2 +[38] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok + +Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2 +[39] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2 +[40] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 2 +[41] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 2 +[42] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 1, 2 +[43] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 3 +[44] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 2, 6 \ No newline at end of file diff --git a/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/images.zip b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dfa3c078aa1b455567e9b28c47cc2039bcf4bd70 --- /dev/null +++ b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25140ac179a8c9e76f7ec619d8743138488ffb8f3e3558d9d63ad57696de2276 +size 567885 diff --git a/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/layout.json b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f9d493fa5dc1577fe9f76ba9a75aebe20ba3d68e --- /dev/null +++ b/2023/Meta-Personalizing Vision-Language Models To Find Named Instances in Video/layout.json @@ -0,0 +1,10486 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "text", + "content": "Meta-Personalizing Vision-Language Models to Find Named Instances in Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "spans": [ + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "text", + "content": "Chun-Hsiao Yeh" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "text", + "content": " Bryan Russell" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "text", + "content": " Josef Sivic" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "text", + "content": " Fabian Caba Heilbron" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "text", + "content": " Simon Jenni" + }, + { + "bbox": [ + 58, + 141, + 532, + 157 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 160, + 483, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 160, + 483, + 175 + ], + "spans": [ + { + "bbox": [ + 111, + 160, + 483, + 175 + ], + "type": "text", + "content": "1University of California, Berkeley 2CIIRC CTU 3Adobe Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 177, + 473, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 177, + 473, + 190 + ], + "spans": [ + { + "bbox": [ + 119, + 177, + 473, + 190 + ], + "type": "text", + "content": "daniel.yeh@berkeley.edu {brussell, inr03127, caba, jenni} @adobe.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "spans": [ + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 278, + 290, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 290, + 542 + ], + "type": "text", + "content": "Large-scale vision-language models (VLM) have shown impressive results for language-guided search applications. While these models allow category-level queries, they currently struggle with personalized searches for moments in a video where a specific object instance such as \"My dog Biscuit\" appears. We present the following three contributions to address this problem. First, we describe a method to meta-personalize a pre-trained VLM, i.e., learning how to learn to personalize a VLM at test time to search in video. Our method extends the VLM's token vocabulary by learning novel word embeddings specific to each instance. To capture only instance-specific features, we represent each instance embedding as a combination of shared and learned global category features. Second, we propose to learn such personalization without explicit human supervision. Our approach automatically identifies moments of named visual instances in video using transcripts and vision-language similarity in the VLM's embedding space. Finally, we introduce This-Is-My, a personal video instance retrieval benchmark. We evaluate our approach on This-Is-My and DeepFashion2 and show that we obtain a " + }, + { + "bbox": [ + 46, + 278, + 290, + 542 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 278, + 290, + 542 + ], + "type": "text", + "content": " relative improvement over the state of the art on the latter dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 668 + ], + "type": "text", + "content": "The recent introduction of large-scale pre-trained vision-language models (VLMs) has enabled many new vision tasks, including zero-shot classification and retrieval [14, 18, 22], image/video generation [12, 24, 25, 27, 28, 32], or language-guided question answering [1, 19, 42]. It is now possible to search not only for specific object categories (e.g., dogs) but also for more specific descriptions of both" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 219, + 536, + 438 + ], + "blocks": [ + { + "bbox": [ + 313, + 219, + 536, + 438 + ], + "lines": [ + { + "bbox": [ + 313, + 219, + 536, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 536, + 438 + ], + "type": "image", + "image_path": "48552c1d4940bd8f73574710c34c6c90dea7d045f9fbe7695919b460665f9156.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 440, + 537, + 449 + ], + "lines": [ + { + "bbox": [ + 315, + 440, + 537, + 449 + ], + "spans": [ + { + "bbox": [ + 315, + 440, + 537, + 449 + ], + "type": "text", + "content": "...create recipes to them...This is my dog Biscuit...If you teach them a trick..." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 304, + 471, + 547, + 605 + ], + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 605 + ], + "type": "text", + "content": "Figure 1. Meta-Personalized Vision-Language Model (VLM) to Retrieve Named Instances in Video. Given a video where a user-specific instance, e.g., \"My dog Biscuit\" is mentioned, our method automatically learns a representation for the user-specific instance in the VLM's text input space. The personalized VLM can then be used to retrieve the learned instance in other contexts through natural language queries, e.g., grabbing a pink frisbee. This result is enabled by metapersonalizing the VLM on a large-scale dataset of narrated videos by pre-learning shared global category tokens (in this example for the category of 'dogs'), which are then easily personalized to user-specific instances from only a few user-given training examples." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 453, + 489, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 453, + 489, + 465 + ], + "spans": [ + { + "bbox": [ + 358, + 453, + 489, + 465 + ], + "type": "text", + "content": "Input Video with Transcript" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "the object and scene attributes (e.g., \"A small white dog playing at the dog park\"). However, we often do not want to search for just any example of a generic category but instead to find a specific instance. For example, a user might want to search their personal video library for all the scenes that show their dog \"Biscuit grabbing a pink frisbee\", as illustrated in Figure 1. Since VLMs do not have a representation of \"Biscuit,\" such queries are beyond the capabilities" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 674, + 118, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 674, + 118, + 684 + ], + "spans": [ + { + "bbox": [ + 62, + 674, + 118, + 684 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 62, + 674, + 118, + 684 + ], + "type": "text", + "content": " Equal advising." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 685, + 276, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 685, + 276, + 693 + ], + "spans": [ + { + "bbox": [ + 63, + 685, + 276, + 693 + ], + "type": "text", + "content": "*Work done during CHY's summer internship at Adobe Research." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "text", + "content": "Czech Institute of Informatics, Robotics and Cybernetics at the Czech Technical University in Prague." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19123" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 141, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 141, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 141, + 83 + ], + "type": "text", + "content": "of off-the-shelf VLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 87, + 289, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 87, + 289, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 87, + 289, + 278 + ], + "type": "text", + "content": "Recent work [5] proposed a method to extend the language encoder's vocabulary with a newly learned token that represents a specific personal instance to address this issue. While this approach enables language-guided search for personal instances by placing the learned tokens in the query prompt, their solution assumes a collection of manually annotated images showing the individual instance in various contexts for successful token learning. For this approach to work in practice, a user must manually annotate all their important personal instances in various contexts, such that the instance representation does not capture nuisance features, e.g., the background. We thus identify two key challenges: 1) collecting personal instance examples without explicit human labeling and 2) learning a generalizable object-centric representation of personal instances from very few examples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 280, + 289, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 289, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 289, + 519 + ], + "type": "text", + "content": "The contributions of this work are three-fold. As our first contribution, we propose a method to automatically identify important personal instances in videos for personalizing a vision-language model without explicit human annotations. Indeed, people often record and refer to personal items or relationships in videos found online. Our approach is thus to identify mentions of personal instances in a video automatically and leverage these moments to build a set of personal instances for training. To this end, we extract the transcripts of videos using speech-to-text models and find candidate moments by looking for occurrences of \"this is my *\" or similar possessive adjective patterns. The symbol * in this example could represent a single word or sequence of words describing the instance (e.g., *= \"dog Biscuit\"). We then use vision-language similarity to filter non-visual examples and to find additional occurrences in the video for training. For example, we found more than six thousand named instances in 50K videos randomly sampled from the Merlot Reserve dataset [44]. We call the resulting collection of named instances in videos the This-Is-My dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "As our second contribution, we propose a novel model and training procedure to learn text tokens representing the named instances in video from possibly very few and noisy training examples. Our method represents each instance with learned tokens and models each token as a linear combination of a set of pre-learned category-specific features shared across different instances. This set of shared category-specific features (similar to object attributes) improves the generalization of our method by preventing the instance representations from capturing nuisance features (e.g., the scene background). Furthermore, we show how to pre-train and adapt the shared category features using a large set of automatically collected This-Is-My examples, further improving our model's few-shot personalization performance at test-time. We call this pre-training of shared category features meta-personalization. In contrast to prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "work [5], our method does not require training additional neural network models and requires only the optimization of a contrastive learning objective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 109, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 205 + ], + "type": "text", + "content": "As our final contribution, we demonstrate and evaluate our model on an existing fashion item retrieval benchmark, DeepFashion2, and our new challenging This-Is-My video instance retrieval dataset4 depicting specific object instances across different videos and contexts. Our experiments demonstrate that our method outperforms several baselines and prior approaches on these challenging language-guided instance retrieval tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 237, + 547, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 356 + ], + "type": "text", + "content": "Vision-Language Models for Video Retrieval. Vision-language foundational models [14, 18, 22] have been successful for zero-shot and other diverse video tasks, such as video question answering [19, 42], language-video grounding [13, 37, 41], and text-to-video retrieval [11, 16, 17, 20, 21, 39, 40]. These models have a powerful representation that transfers well to the video domain to achieve competitive performance on video-language tasks. Our approach builds on these powerful representations to retrieve specific named instances in video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 357, + 547, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 584 + ], + "type": "text", + "content": "Personalized Concept Learning. Adapting a model to learn a user-specific representation has been a significant topic in machine learning research, including recommendation systems [2, 3] and federated learning [15]. Relevant to us are recent approaches for adapting vision-language models to object instances. PALAVRA [5] proposes a learning scheme that appends a learnable token for a new personalized concept to the token embedding of the input text prompt. This learned representation helps to preserve the personalized concept. DualPrompt [36] introduces a framework that learns a small set of prompts to emphasize more specific concepts without forgetting the learned concepts in the pre-trained foundational model. There have also been works that have extended this personalized concept learning to image generation [4,8,26]. While these approaches adapt a vision-language model to personal instances, they perform the adaptation independently for each instance and do not \"meta-train\" for the personalization task for improved few-shot learning as we do in this work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 586, + 547, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 547, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 547, + 694 + ], + "type": "text", + "content": "Fine-tuning and Test-time Adaptation. Fine-tuning is a common strategy to adapt a pretrained model for a specific downstream task by transferring the source model to a target domain. Recent works on vision-language model tuning include CLIP-Adapter [9] that proposes to conduct fine-tuning with an extra bottleneck layer while freezing the pre-trained CLIP model. WiSE-FT [38] resolves the distribution shift caused by fine-tuning and ensembles the weights of the original and fine-tuned model to increase ro" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 545, + 712 + ], + "type": "text", + "content": "4Available at https://danielchyeh.github.io/metaper/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19124" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 210, + 252 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 210, + 252 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 210, + 252 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 210, + 252 + ], + "type": "image", + "image_path": "4ee01b6b95138c0199ecf6e9c38be9b8baf03bdd861b048b8471a721a4d48fc4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": "Figure 2. Overview of our Personalized Vision-Language Model. Our model augments a frozen VLM (blue) with novel personal instance tokens " + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "inline_equation", + "content": "w = Cz" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": " (orange) that are a combination of global category features " + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": " (green) with instance-specific weights " + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z}" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": " (yellow). Our approach for personalized instance retrieval has three stages. First, we pre-learn global category features " + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "inline_equation", + "content": "C_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": " on a large set of automatically mined named personal instances in videos. We call this process Meta-Personalization (left). In the second step (middle), we adapt the meta-personalized category features " + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "inline_equation", + "content": "C_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": " at test-time and learn novel instance weights " + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z}_{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": " to represent a user's personal instances via " + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "inline_equation", + "content": "w = C_{\\mathcal{P}}z" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": ". Finally (right), we leverage the (frozen) personalized instance tokens " + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 255, + 547, + 323 + ], + "type": "text", + "content": " in natural language queries at query time." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 74, + 365, + 253 + ], + "blocks": [ + { + "bbox": [ + 215, + 74, + 365, + 253 + ], + "lines": [ + { + "bbox": [ + 215, + 74, + 365, + 253 + ], + "spans": [ + { + "bbox": [ + 215, + 74, + 365, + 253 + ], + "type": "image", + "image_path": "948389672c9dc1d540297882241a97776f06808533f67166442f166d11e9a93f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 372, + 74, + 539, + 251 + ], + "blocks": [ + { + "bbox": [ + 372, + 74, + 539, + 251 + ], + "lines": [ + { + "bbox": [ + 372, + 74, + 539, + 251 + ], + "spans": [ + { + "bbox": [ + 372, + 74, + 539, + 251 + ], + "type": "image", + "image_path": "9646bc818d22e4bfaf9d795838662cfb13340f332f3b72d0ca3613d84f3b7a0b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 387 + ], + "type": "text", + "content": "bustness. Prior test-time adaption works [31, 34, 35] fine-tune on target data without information from the source data. Our approach leverages test-time adaptation for updating our meta-personalized model to user-specific data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 508 + ], + "type": "text", + "content": "Meta-Learning. We draw inspiration from meta-learning (\"learning to learn\") [6, 7, 43], which enables models to quickly adapt to new tasks by learning on a diverse set of tasks. Hence, given only a handful of novel training examples, the model can be adapted to novel tasks. Our approach borrows the idea of meta-learning to meta-personalize a model by learning global category features from a large video database. We then adapt the global features during test-time training from only few examples of user-specific instances to enable query-time retrieval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 523, + 284, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 284, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 284, + 537 + ], + "type": "text", + "content": "3. (Meta-)Personalization of Named Instances" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 640 + ], + "type": "text", + "content": "Our goal is to learn representations of personal items in video that enable retrieval through natural language queries. To achieve this goal, we have to address the key challenge of adapting a model with one or few examples of a named instance. To address this challenge, we propose a metapersonalization approach that learns to personalize given a large corpus of named instances mined from videos with transcriptions. We illustrate our approach in Figure 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "In the first step (Figure 2 (left)), we mine automatically a collection " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " of named instances from videos with transcriptions for meta-personalization. We use this collection to train a proposed model " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{C,z}" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " that includes global category features " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " and instance-specific parameters " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ". The global category features " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " are lightweight and shared across all" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": "instances. Given a natural language query " + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": " and video " + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": ", the model returns a score " + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{C,z}(u,v)" + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": ". During metapersonalization, given a training loss " + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": ", we jointly optimize the loss over the global category features " + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": " and instance parameters " + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": " for each named instance in the collection " + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 355, + 408, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 408, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 355, + 408, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\left(C _ {\\mathcal {D}}, \\mathcal {Z} _ {\\mathcal {D}}\\right) \\in \\underset {(C, \\mathcal {Z})} {\\arg \\min } \\sum_ {z \\in \\mathcal {Z}} \\mathcal {L} (C, z). \\tag {1}", + "image_path": "0c705a2088cb76c37d632e0c136069eb05ce9485626e02cb8026a5beaa212754.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "type": "text", + "content": "Note that here the instance-specific parameters " + }, + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathcal{D}}" + }, + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "type": "text", + "content": " learnt via (1) are discarded while the global category features " + }, + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "type": "inline_equation", + "content": "C_{\\mathcal{D}}" + }, + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "type": "text", + "content": " are kept as the meta-personalized part of the model. The global category features " + }, + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "type": "inline_equation", + "content": "C_{\\mathcal{D}}" + }, + { + "bbox": [ + 305, + 442, + 545, + 502 + ], + "type": "text", + "content": " capture information shared across instances relevant to the personalization task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": "In the second step (Figure 2 (middle)), we are given a set of named video instances " + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": " (e.g., automatically mined from someone's personal video library) and wish to perform test-time personalization of the model to this person's instances. Here each instance is represented by only one or few examples. In this step, we optimize the training loss " + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": " over the global category features " + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": " and the set of instance parameters " + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": " for all instances in the personal set " + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": " starting from the pre-trained global category features " + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "inline_equation", + "content": "C_{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": " and random " + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": ". We obtain the personalized model parameters " + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "inline_equation", + "content": "(C_{\\mathcal{P}},\\mathcal{Z}_{\\mathcal{P}})" + }, + { + "bbox": [ + 304, + 502, + 545, + 622 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 355, + 631, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 631, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 355, + 631, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\left(C _ {\\mathcal {P}}, \\mathcal {Z} _ {\\mathcal {P}}\\right) \\in \\underset {(C, \\mathcal {Z})} {\\arg \\min } \\sum_ {z \\in \\mathcal {Z}} \\mathcal {L} (C, z). \\tag {2}", + "image_path": "9edf249e170d00b6eb59e7a55e4d53b3add082f8f859b0b7f3ffcee57c9adee5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 666, + 435, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 435, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 435, + 677 + ], + "type": "text", + "content": "We now keep both " + }, + { + "bbox": [ + 306, + 666, + 435, + 677 + ], + "type": "inline_equation", + "content": "C_{\\mathcal{P}}" + }, + { + "bbox": [ + 306, + 666, + 435, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 666, + 435, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathcal{P}}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "In the final step (Figure 2 (right)), we perform retrieval over a potentially large dataset using the test-time personalized model " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{C_{\\mathcal{P}},z}" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z}_{\\mathcal{P}}" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "). We next describe" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19125" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 81, + 102, + 114 + ], + "blocks": [ + { + "bbox": [ + 51, + 81, + 102, + 114 + ], + "lines": [ + { + "bbox": [ + 51, + 81, + 102, + 114 + ], + "spans": [ + { + "bbox": [ + 51, + 81, + 102, + 114 + ], + "type": "image", + "image_path": "5b74e414fc093a921d5a9a1dab87f549dd0b73bcd07430e8ac1ab6448e74f7a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 79, + 146, + 105 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 146, + 105 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 146, + 105 + ], + "type": "text", + "content": "visual similarity for every pair" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 148, + 77, + 190, + 116 + ], + "blocks": [ + { + "bbox": [ + 148, + 77, + 190, + 116 + ], + "lines": [ + { + "bbox": [ + 148, + 77, + 190, + 116 + ], + "spans": [ + { + "bbox": [ + 148, + 77, + 190, + 116 + ], + "type": "image", + "image_path": "788d6595b4af62523175399336e012db3ab4b6061e6bca806ce0bc5fb42f9c12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 74, + 233, + 116 + ], + "lines": [ + { + "bbox": [ + 190, + 74, + 233, + 116 + ], + "spans": [ + { + "bbox": [ + 190, + 74, + 233, + 116 + ], + "type": "text", + "content": "0.2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 240, + 77, + 283, + 116 + ], + "blocks": [ + { + "bbox": [ + 240, + 77, + 283, + 116 + ], + "lines": [ + { + "bbox": [ + 240, + 77, + 283, + 116 + ], + "spans": [ + { + "bbox": [ + 240, + 77, + 283, + 116 + ], + "type": "image", + "image_path": "70e7dfee7a79237d126b138c718afcd2bf12d62f224a0755dc23350ec8400c55.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 118, + 247, + 125 + ], + "lines": [ + { + "bbox": [ + 192, + 118, + 247, + 125 + ], + "spans": [ + { + "bbox": [ + 192, + 118, + 247, + 125 + ], + "type": "text", + "content": "set of candidates" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 137, + 64, + 150 + ], + "blocks": [ + { + "bbox": [ + 66, + 128, + 238, + 138 + ], + "lines": [ + { + "bbox": [ + 66, + 128, + 238, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 128, + 238, + 138 + ], + "type": "text", + "content": "Finding Additional Instance Examples (Step 3)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 137, + 64, + 150 + ], + "lines": [ + { + "bbox": [ + 50, + 137, + 64, + 150 + ], + "spans": [ + { + "bbox": [ + 50, + 137, + 64, + 150 + ], + "type": "image", + "image_path": "53a57fea915bc7aa2aff6185f49d5b588c64b32e4ea128c6bb174f4939f10fd3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 51, + 153, + 107, + 189 + ], + "blocks": [ + { + "bbox": [ + 51, + 153, + 107, + 189 + ], + "lines": [ + { + "bbox": [ + 51, + 153, + 107, + 189 + ], + "spans": [ + { + "bbox": [ + 51, + 153, + 107, + 189 + ], + "type": "image", + "image_path": "6426d3a23890796359acef37c8a38cabc8f06c3bf07c44323719e5560fc3bdc2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 107, + 153, + 162, + 190 + ], + "blocks": [ + { + "bbox": [ + 107, + 153, + 162, + 190 + ], + "lines": [ + { + "bbox": [ + 107, + 153, + 162, + 190 + ], + "spans": [ + { + "bbox": [ + 107, + 153, + 162, + 190 + ], + "type": "image", + "image_path": "d3b20ae614e3ac354e5e83ebb7e136378a311f5f2ff1fb9cb2ecedb23e6aa7e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 190, + 156, + 196 + ], + "lines": [ + { + "bbox": [ + 52, + 190, + 156, + 196 + ], + "spans": [ + { + "bbox": [ + 52, + 190, + 156, + 196 + ], + "type": "text", + "content": "instance name: Time to talk about" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 171, + 153, + 223, + 190 + ], + "blocks": [ + { + "bbox": [ + 127, + 145, + 188, + 152 + ], + "lines": [ + { + "bbox": [ + 127, + 145, + 188, + 152 + ], + "spans": [ + { + "bbox": [ + 127, + 145, + 188, + 152 + ], + "type": "text", + "content": "visual relevance: 0.1" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 153, + 223, + 190 + ], + "lines": [ + { + "bbox": [ + 171, + 153, + 223, + 190 + ], + "spans": [ + { + "bbox": [ + 171, + 153, + 223, + 190 + ], + "type": "image", + "image_path": "2606f2a888dd299fe0f41072ec5e46716a45e88bb51f6d1b07e7f732ba7f4e68.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 190, + 267, + 196 + ], + "lines": [ + { + "bbox": [ + 180, + 190, + 267, + 196 + ], + "spans": [ + { + "bbox": [ + 180, + 190, + 267, + 196 + ], + "type": "text", + "content": "instance name: fender guitar" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 224, + 153, + 277, + 190 + ], + "blocks": [ + { + "bbox": [ + 224, + 153, + 277, + 190 + ], + "lines": [ + { + "bbox": [ + 224, + 153, + 277, + 190 + ], + "spans": [ + { + "bbox": [ + 224, + 153, + 277, + 190 + ], + "type": "image", + "image_path": "46fd0c6eaaa5ee6a60bca5e18a3fc897a75e17f33f1f70dd467a51c2cff0308b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 51, + 212, + 62, + 224 + ], + "blocks": [ + { + "bbox": [ + 66, + 201, + 209, + 212 + ], + "lines": [ + { + "bbox": [ + 66, + 201, + 209, + 212 + ], + "spans": [ + { + "bbox": [ + 66, + 201, + 209, + 212 + ], + "type": "text", + "content": "Filtering Non-visual Instances (Step 2)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 212, + 62, + 224 + ], + "lines": [ + { + "bbox": [ + 51, + 212, + 62, + 224 + ], + "spans": [ + { + "bbox": [ + 51, + 212, + 62, + 224 + ], + "type": "image", + "image_path": "61adb1288277139d8ca03dc56ef128da6e74185220e114f3bce1717d1ccef932.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 85, + 220, + 279, + 258 + ], + "blocks": [ + { + "bbox": [ + 58, + 239, + 79, + 247 + ], + "lines": [ + { + "bbox": [ + 58, + 239, + 79, + 247 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 79, + 247 + ], + "type": "text", + "content": "Video:" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 220, + 279, + 258 + ], + "lines": [ + { + "bbox": [ + 85, + 220, + 279, + 258 + ], + "spans": [ + { + "bbox": [ + 85, + 220, + 279, + 258 + ], + "type": "image", + "image_path": "67903db5e0f75d42b56b19ad4f5a062b5cb2246b2123e95a3342eb21b46de532.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 258, + 279, + 267 + ], + "lines": [ + { + "bbox": [ + 52, + 258, + 279, + 267 + ], + "spans": [ + { + "bbox": [ + 52, + 258, + 279, + 267 + ], + "type": "text", + "content": "Transcript: This is our time to talk about ... This is my fender guitar" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 270, + 197, + 281 + ], + "lines": [ + { + "bbox": [ + 67, + 270, + 197, + 281 + ], + "spans": [ + { + "bbox": [ + 67, + 270, + 197, + 281 + ], + "type": "text", + "content": "Spotting Named Instances (Step 1)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 283, + 287, + 371 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 371 + ], + "type": "text", + "content": "Figure 3. Automatic mining of named instances in video for meta-personalization. Our automatic mining pipeline includes three steps (from bottom to top). Step 1 finds named instances via string-matching of possessive patterns in video transcripts. Step 2 filters non-visual instances using text-to-visual relevance between the instance name and the video shots neighboring the named instance. Finally, Step 3 retrieves additional shots with high visual similarity to the instance reference shot." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "type": "text", + "content": "how we automatically mine the named instances in video " + }, + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "type": "text", + "content": " for meta-personalization (Sec. 3.1), and our full model " + }, + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "type": "text", + "content": " for query-time retrieval and loss " + }, + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 384, + 287, + 433 + ], + "type": "text", + "content": " for meta-personalization and test-time personalization (Sec. 3.2)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 441, + 287, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 454 + ], + "type": "text", + "content": "3.1. Automatic Mining of Named Instances in Video" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "type": "text", + "content": "To identify personal instances without explicit supervision, we leverage a collection of videos from the web along with their corresponding time-aligned transcripts. These transcripts can be automatically generated through speech-to-text models [23, 33]. We now describe a procedure to mine these data for a set of moments (i.e., a collection of video shots) depicting a referred personal instance in the transcript without the need for manual annotation. We will use these moments for training a meta-personalization model (Sec. 3.2)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "content": "Spotting Named Instances. Here our goal is to find moments where candidate personal instances are mentioned in videos. We do so by searching for possessive text patterns5 such as \"This is my *\" in a corpus of time-aligned video transcripts. This string-matching process outputs a list of candidate instance names * associated with video timestamps " + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "content": ". In prac" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "content": "tice, we keep up to four words after a possessive text pattern is matched based on text-visual similarity (see supp. for details). That way, we can retrieve simple named instances such as This is my dog (" + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "content": "dog) but also complex ones like This is my favorite CHANEL classic handbag (" + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "content": "CHANEL classic handbag). Note also that a single video might include multiple string matches; for instance, the example video illustrated in Figure 3 (Step 1) includes two matches: \"This is our time to talk about\" at time 1:30 and \"This is my fender guitar\" at time 3:25." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": "Filtering Non-visual Instances. The previous spotting step only searches for potential instances using the transcript, yielding many string matches that are non-visual, i.e., the strings do not describe the visible content in the video. Here we aim to filter out these non-visual instances. We do so by computing the text-to-visual relevance between the instance name (e.g., fender guitar) and the neighboring shots around the time when the instance is mentioned. We add neighboring shots to cover cases where the named instances are shown just before or after they are mentioned. Concretely, given a sequence of " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " video shots " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "S = [s_1, \\ldots, s_m]" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " automatically extracted with [30], we find the shot " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "s_{t^*}" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " that overlaps with " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " (the time when the instance was mentioned). Next, we form a set of candidate visual references " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "S_{t^*} = [s_{t^* - 1}, s_{t^*}, s_{t^* + 1}]" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " comprising a window of shots that are previous and subsequent to " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "s_{t^*}" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": ". We then compute text-to-visual relevance scores using CLIP [22] encoders. This encoding process yields " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": "-normalized embeddings " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "f_l(*)" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " for the named instance and " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "f_v(s_i)" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " for each shot " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "s_i \\in S_{t^*}" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": ". We compute " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "f_v(s_i)" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " by averaging the visual embeddings of all frames in the corresponding shot. Finally, we compute the cosine similarity between every " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "(f_l(*), f_v(s_i))" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " pair and retain a visual reference shot " + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "inline_equation", + "content": "s^*" + }, + { + "bbox": [ + 304, + 207, + 546, + 553 + ], + "type": "text", + "content": " if the highest cosine similarity is greater than 0.3. This filtering step outputs a cleaned set of named instances with a corresponding visual reference. Figure 3 (Step 2) illustrates how we prune out non-visual matches such as time to talk about. In contrast, visual instances such as fender guitar are kept and matched with a visual reference." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "Finding Additional Instance Shots. Since frames from a single video shot provide only limited variability in the instance appearance and could thus limit the learning, we aim to recover other shots from the video where that instance appears. We leverage CLIP's visual encoder to compute the visual similarity between the instance's reference shot " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "s^*" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " and every shot " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "s_i \\in S" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ". We extract an embedding " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_v(s^*)" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " for the reference shot " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "s^*" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " and an embedding " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_v(s_i)" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " for each shot " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ". Similar to the non-visual filtering step, we average the CLIP embeddings of each frame belonging to a shot. Then, we compute the cosine similarity between the embeddings for the reference shot and every candidate shot in the video. We keep the shots whose cosine similarity with the refer" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 673, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 287, + 712 + ], + "type": "text", + "content": "5List of possessive text patterns: , , , , , , , , , " + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19126" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 83, + 289, + 270 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 97, + 83 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 97, + 83 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 97, + 83 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C},z}(\\pmb {u},\\pmb {v})" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 83, + 289, + 270 + ], + "lines": [ + { + "bbox": [ + 47, + 83, + 289, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 83, + 289, + 270 + ], + "type": "image", + "image_path": "e8379900235a8874a3fddede6276fb59ec0913edffd3118dabffc64be6c38eb3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "text", + "content": "Figure 4. Model Overview. Our model " + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{C,z}" + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "text", + "content": " extends CLIP's language input vocabulary with " + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "inline_equation", + "content": "n_w" + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "text", + "content": " novel instance-specific tokens " + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "inline_equation", + "content": "w_i^y = C_l z_i^y" + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "text", + "content": ", which we model as a linear combination of metapersonalized category features " + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "inline_equation", + "content": "z_i^y" + }, + { + "bbox": [ + 46, + 271, + 287, + 327 + ], + "type": "text", + "content": ". Note that the vision and language encoders are frozen during this process." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": "ence is greater than 0.9. Figure 3 (Step 3) illustrates the output of this final step in the automatic mining pipeline. Our mining algorithm allows us to retrieve additional shots for the instance fender guitar. Now the instance examples not only include a clean close-up of the guitar, but also shots where the guitar is being played or held by its owner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 439, + 277, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 277, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 277, + 452 + ], + "type": "text", + "content": "3.2. Learning Personal Instance Representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "text", + "content": "The result of the mining procedure described above is a dataset consisting of a set of video shots " + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{s_1, \\ldots, s_n\\}" + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "text", + "content": " and corresponding instance IDs " + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "inline_equation", + "content": "Y = \\{y_1, \\ldots, y_n\\}" + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "inline_equation", + "content": "y_i = y_j" + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 46, + 459, + 287, + 530 + ], + "type": "text", + "content": " are video shots that are assumed to contain the same instance. We now describe how we use these data to learn representations of the collected instances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": "Our approach is to leverage a large-scale pre-trained vision-language model (CLIP [22]) and augment its language encoder with a set of novel personal instance tokens. Let " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "f_{v}(s)" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": " be the output of the visual encoder for shot " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": " (computed as the average over the frame embeddings) and let " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "f_{l}(u)" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": " be the output of the language encoder for a natural language input " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "u = [v_{1},\\dots ,v_{m}]" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": " token embeddings, where " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "v_{i}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": " denote learned token embeddings of the pretraining vocabulary (positional embeddings are included but omitted from the notation). We propose to extend this vocabulary with novel personal instance tokens. Concretely, our approach introduces a set " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "w^{y} = \\{w_{i}^{y}\\}_{i = 1}^{n_{w}}" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "n_w" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": " new tokens that represent a personal instance " + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "inline_equation", + "content": "y\\in Y" + }, + { + "bbox": [ + 46, + 532, + 287, + 687 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "To learn these tokens and to perform personalized retrieval at test time, we construct natural language person" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 375, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 375, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 375, + 84 + ], + "type": "text", + "content": "alized queries as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 92, + 545, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 92, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 318, + 92, + 545, + 107 + ], + "type": "interline_equation", + "content": "\\hat {u} ^ {p} = \\left[ p _ {1}, \\dots , p _ {k - 1}, w _ {1} ^ {y}, \\dots , w _ {n _ {w}} ^ {y}, p _ {k + 1}, \\dots , p _ {m} \\right], \\tag {3}", + "image_path": "f15c43603a31167fb0f03c5032639b3ce8614b88b47b5418927398b7eee8a8e6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "content": " are token embeddings of a chosen prompt " + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "inline_equation", + "content": "p = [p_1, \\ldots, p_m]" + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "content": ". During training, the prompt " + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "content": " corresponds to a random template of the form [An image of *, [* can be seen in this photo], [There is * in this image], etc., and " + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 114, + 547, + 185 + ], + "type": "text", + "content": " denotes the position of the * placeholder for the instance tokens." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 186, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 545, + 243 + ], + "type": "text", + "content": "Instances as Combinations of Category Features. Since learning personal instance tokens from possibly very few examples runs the danger of overfitting (e.g., to nuisance features in the background), we propose to parameterize them as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 382, + 244, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 244, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 382, + 244, + 545, + 258 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {y} = C _ {l} z _ {i} ^ {y} \\in \\mathbb {R} ^ {d \\times 1}, \\tag {4}", + "image_path": "96c7ec320a7772ddf4e34beea2db98ff1f4c51eec189a1c829a4a9fad1c52b6a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "z_{i}^{y} \\in \\mathbb{R}^{q \\times 1}" + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "content": " is a vector of learnable weights specific to each instance and " + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "C_{l} \\in \\mathbb{R}^{d \\times q}" + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "content": " is a matrix of learnable global category features, which are shared for all instances belonging to the same object category " + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "l \\in \\mathcal{V}" + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\text{car}, \\text{person}, \\text{dog}, \\ldots\\}" + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "content": ") and constitute the set " + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "C_{\\mathcal{D}} = \\{C_{l}\\}_{l \\in \\mathcal{V}}" + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "content": ". We illustrate the model in Figure 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "content": "We can think of the columns of " + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "content": " as representing a set of shared category features and the final instance token " + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "inline_equation", + "content": "w_i^y" + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "content": " as a linear combination of these features with weights " + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "inline_equation", + "content": "z_i^y" + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "content": ". Our aim is that only category-specific features are captured during training and irrelevant features (e.g., about the background) are discarded. Intuitively, the columns of " + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "content": " could correspond to attributes of an object category, e.g., if we were to learn \"car\" features, these could capture their color, brand, type, or age, to name a few. To identify which category matrix " + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "content": " to use for an instance " + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "content": ", we rely on 0-shot classification using vision-language similarity between instance shots and a generic prompt for the category " + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "inline_equation", + "content": "l \\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 335, + 546, + 477 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": "Contrastive Personal Token Learning. We propose a contrastive learning objective to learn the personal instance tokens " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "w^{y}" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": " using a set of video shots containing the instance " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": ". To this end, let " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\psi_{i} := f_{v}(s_{i})" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": " be an encoding of a video shot " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": ", i.e., the average frame encoding of all frames in the shot, and let " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\phi_{i} := f_{l}(\\hat{u}_{i}^{p})" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": " denote the language encoding of a corresponding personalized query. We learn the novel tokens " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "w^{y}" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": " and shared category features " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "C_{l}" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": " by optimizing two contrastive objectives: a language-language contrastive objective " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l}" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": " and a vision-language contrastive objective " + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vl}" + }, + { + "bbox": [ + 304, + 478, + 545, + 609 + ], + "type": "text", + "content": ". The language-language objective is given by" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 616, + 545, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 545, + 660 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l} = \\sum_ {i \\in \\mathcal {B}} \\sum_ {j \\neq i \\in \\mathcal {B}} - 1 \\left\\{y _ {i} = y _ {j} \\right\\} \\log \\left(\\frac {d \\left(\\phi_ {i} , \\phi_ {j}\\right)}{\\sum_ {k \\neq i \\in \\mathcal {B}} d \\left(\\phi_ {i} , \\phi_ {k}\\right)}\\right), \\tag {5}", + "image_path": "e7e19e107ccf5f6d00033f76e3d29a55dbece5e55db106a818daf783243ee881.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "inline_equation", + "content": "d(a,b)\\coloneqq \\exp \\left(\\frac{1}{\\lambda}\\frac{a^{\\intercal}b}{\\|a\\|_{2}\\|b\\|_{2}}\\right)" + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "text", + "content": " measures the similarity between the feature vectors " + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "text", + "content": " is a temperature parameter, and " + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 661, + 545, + 713 + ], + "type": "text", + "content": " is a randomly sampled training minibatch. The vision-language objective is similarly defined as" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19127" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 73, + 506, + 250 + ], + "blocks": [ + { + "bbox": [ + 89, + 73, + 506, + 250 + ], + "lines": [ + { + "bbox": [ + 89, + 73, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 89, + 73, + 506, + 250 + ], + "type": "image", + "image_path": "60425bf28057a6954da847bfede9fb95f80f36c3ab27741bf639b951d1c9c3dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "type": "text", + "content": "Figure 5. Examples from This-Is-My { Meta-Personalization " + }, + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "type": "text", + "content": " (top) vs Test-time personalization " + }, + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "type": "text", + "content": " (bottom-left) vs Query-time " + }, + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 252, + 545, + 297 + ], + "type": "text", + "content": " (bottom-right)} datasets. In the Query-time dataset (bottom-right), we design a challenging video instance retrieval task. For example, in (a) the named instance (i.e., Alex's piano) is in the background and is barely visible and in (b) the background scenes in the query-time dataset (bottom-right) are completely different from the test-time personalization dataset (bottom-left) depicting the same named instance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 319, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v l} = \\sum_ {i, j \\in \\mathcal {B}} - \\mathbb {1} \\left\\{y _ {i} = y _ {j} \\right\\} \\log \\left(\\frac {d \\left(\\phi_ {i} , \\psi_ {j}\\right)}{\\sum_ {k \\in \\mathcal {N}} d \\left(\\phi_ {i} , \\psi_ {k}\\right)}\\right), \\tag {6}", + "image_path": "c7cd39e2fc310169ec3b7f2d9f484a6ace266c04c7c07f3f2ba20bae4a5efe55.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 356, + 287, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 451 + ], + "type": "text", + "content": "with the set of negative examples " + }, + { + "bbox": [ + 46, + 356, + 287, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 356, + 287, + 451 + ], + "type": "text", + "content": " comprising both other examples in the batch " + }, + { + "bbox": [ + 46, + 356, + 287, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 356, + 287, + 451 + ], + "type": "text", + "content": " and non-instance shots from the videos containing the named instances, i.e., shots that have low vision-language similarity. The loss is low when the encodings for video shots and personalized queries with the same instance ID are more similar than to other queries. Including non-instance segments as negatives can help discard non-instance features such as scene background." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "type": "text", + "content": "To further constrain the learning toward category-specific attributes, we include a loss that maximizes the similarity between a personal instance query and a generic category query. Concretely, let " + }, + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "type": "inline_equation", + "content": "c_{l}" + }, + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "type": "text", + "content": " be a category query embedding for category " + }, + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "type": "text", + "content": " (e.g., \"An image of a [car]\") to which instance " + }, + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 452, + 287, + 535 + ], + "type": "text", + "content": " belongs. We then include the following category-anchoring loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 540, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 540, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 540, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = - \\sum_ {i \\in \\mathcal {B}} \\frac {c _ {l} ^ {\\top} \\phi_ {i}}{\\| c _ {l} \\| _ {2} \\| \\phi_ {i} \\| _ {2}}. \\tag {7}", + "image_path": "59330434253899fc8799dba862c7131decef842cde575c7006f23e75f4b13738.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 598 + ], + "type": "text", + "content": "To summarize, our training loss " + }, + { + "bbox": [ + 47, + 574, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 574, + 287, + 598 + ], + "type": "text", + "content": " (see Equations 1 and 2) for meta- and test-time personalization is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 605, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 605, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 119, + 605, + 287, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {l} + \\mathcal {L} _ {v l} + \\lambda_ {c} \\mathcal {L} _ {c}, \\tag {8}", + "image_path": "bffeca73b744efcb285927e1703a999de68c0e5603f35a3c064d6e1f2b5e9f31.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 624, + 286, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 286, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 286, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 624, + 286, + 636 + ], + "type": "inline_equation", + "content": "\\lambda_{c} = 0.5" + }, + { + "bbox": [ + 47, + 624, + 286, + 636 + ], + "type": "text", + "content": ", controls the amount of category anchoring." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 657, + 159, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 159, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 159, + 671 + ], + "type": "text", + "content": "4. This-Is-My Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": "Our This-Is-My dataset comprises three subsets for metapersonalization, test-time personalization, and querying. We describe each subset next." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "type": "text", + "content": "Meta-Personalization Dataset " + }, + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "type": "text", + "content": ". We gather this subset with the automatic mining algorithm introduced in Section 3.1. We leverage the Merlot Reserve dataset [44], which contains more than 20 million videos from YouTube and their corresponding time-aligned transcripts. In practice, we start from a subset of " + }, + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "type": "inline_equation", + "content": "50K" + }, + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "type": "text", + "content": " randomly sampled Merlot Reserve videos. We spot 6058 named instances using various text possessive templates. Our visual filtering step removes " + }, + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "type": "inline_equation", + "content": "52\\%" + }, + { + "bbox": [ + 304, + 304, + 545, + 555 + ], + "type": "text", + "content": " of instances, generating a total of 2908 named instances with a visual reference. Finally, we mine additional samples for each instance, yielding a total of 49256 instance samples. This subset includes a wide variety of visual concepts, ranging from common objects such as bikes to rare concepts such as toaster. While we design our mining pipeline to minimize noise, it is limited by CLIP's capability to distinguish between similar object instances. For instance, while we find several samples of the fender guitar instance, it includes other shots that do not correspond to the aforementioned guitar (Figure 5 (top)). Nevertheless, we will show empirically that this subset is still useful for meta-personalization purposes (Section 5)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Test-time Personalization Dataset " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": ". Our goal is to create a test dataset that recreates the scenario where a person wants to find when their personal instances appear in their video collection. We want to make this task close to the real scenario where a person records the same visual instance, e.g., their dog, across multiple places and situations. Our strategy is to emulate such a scenario by finding YouTube channels that frequently mention the same instance across multiple videos. While Merlot Reserve is large and diverse, only few channels are represented with more than one video in the dataset. Instead, we download all videos and automatic transcripts from the channels of 15 popular YouTube bloggers. We then run our mining algorithm, but manu" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19128" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "ally supervise the steps for visual filtering and finding additional instance samples. We first manually verify that each instance name and its visual reference are good matches. Then, we find additional sample shots across all videos in the channel by ranking them according to their visual similarity to the instance reference shot. Finally, we review and label the top 1000-scored shots for each named instance. The supplementary material includes a screenshot of the annotation tool. In total, this subset includes 15 named instances with more than 686 labeled samples. Figure 5 (bottom left) shows two example instances in our dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": "Query-time Dataset " + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": ". Our end goal is to retrieve named instances via natural language queries. We would like to be able to find videos when is grabbing a pink Frisbee. To this end, we manually caption 30 instance samples with descriptive queries. Thus, the Query-time dataset includes (manually captioned) video-caption pairs containing instances from the Test-time Personalization dataset. Figure 5 (bottom right) shows manually curated captions for two instances in our dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 323, + 128, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 128, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 128, + 336 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "text", + "content": "Our experiments first ablate our model's contributions and loss design (Section 5.1), and evaluate our final model in personalized instance retrieval benchmarks (Section 5.2). Evaluation Datasets and Metrics. We evaluate our approach on two datasets: (i) our newly introduced This-IsMy personal video instance retrieval benchmark, and (ii) the personalized image retrieval benchmark built on DeepFashion2 [10] proposed by [5]. We evaluate retrieval performance in two settings:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": "1) Generic Instance Retrieval: In this case, we build queries for learned instances using a generic prompt, i.e., \"An image of *\", and measure retrieval performance using mean Average Precision (mAP) and Mean Reciprocal Rank (MRR). In this setting, there are multiple correct matches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "content": "2) Contextualized Instance Retrieval: In this case, we use natural language queries that describe the personal instance in a specific context, e.g., using a scene description such as \"A photo of * lying on the beach\". In this case, we assume there is only a single correct match, and we measure performance with MRR and Recall-at-5 (R@5)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Implementation Details. We use the ViT-B/16 version of CLIP in most of our experiments if not otherwise indicated. All the learnable parameters " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " are randomly initialized from " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,0.1)" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". We set the number of category features to " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "q = 512" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and the number of instance tokens to " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "n_w = 1" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " by default. No data augmentation is used during training (the visual embeddings thus have to be computed just once). For meta-personalization of category features " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ", we randomly select 32 named instances from " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " that are 0-shot classified to each category " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and train for 20 epochs. This process is repeated 10 times, re-initializing " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " each" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 312, + 116, + 538, + 219 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 114 + ], + "type": "text", + "content": "Table 1. Ablation Experiments. We verify our model and training objective design through ablations on This-Is-My and DeepFashion2. We report personal instance retrieval performance in terms of mAP and MRR (higher is better)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 116, + 538, + 219 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 538, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 538, + 219 + ], + "type": "table", + "html": "
AblationThis-Is-MyDeepFashion2
mAPMRRmAPMRR
a) w/o meta-pers.54.1±1.383.1±2.135.2±0.555.2±1.9
b) single C55.5±0.486.9±0.544.1±0.564.7±1.5
c) w/o Ll53.9±1.086.4±2.247.3±0.868.3±1.2
d) w/o Lc48.9±1.078.6±2.447.1±0.568.4±0.8
e) N = B47.4±0.473.7±1.8--
f) w/o pre-trained C53.0±0.785.1±3.044.3±0.867.1±1.0
Ours56.4±0.687.4±1.247.3±0.769.9±0.8
", + "image_path": "cebc7e374bdf02282a6538cdf81deb70bb37ffa5e5b2e34be3745885a1144e32.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "text", + "content": "time while retaining " + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "text", + "content": " from the previous run. Although updating the parameters of " + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "text", + "content": " to a small set of instances " + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "text", + "content": " at test time is beneficial, there is a risk of overfitting when only a few or just a single instance of each category is provided. To mitigate this, we include additional instances from " + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "text", + "content": " with the same categories as contained in " + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 231, + 545, + 339 + ], + "type": "text", + "content": " during test time personalization. All our results are averaged over five runs, each with different random seeds, and we report the standard error for each experiment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 347, + 373, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 373, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 373, + 360 + ], + "type": "text", + "content": "5.1. Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 366, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 403 + ], + "type": "text", + "content": "In Table 1 we report results in the generic instance retrieval setting for the following ablations to illustrate the effect of our model and training objective design:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 403, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "text", + "content": "a) w/o meta-personalization: In this case, we do not learn a meta-personalized " + }, + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "text", + "content": " and instead directly learn instance embeddings " + }, + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "text", + "content": ". This ablation demonstrates that learning global category attributes through meta-personalization improves generalization at test-time personalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 462, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 521 + ], + "type": "text", + "content": "b) single " + }, + { + "bbox": [ + 304, + 462, + 545, + 521 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 462, + 545, + 521 + ], + "type": "text", + "content": " (shared for all categories): In this experiment, we learn only a single category matrix " + }, + { + "bbox": [ + 304, + 462, + 545, + 521 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 462, + 545, + 521 + ], + "type": "text", + "content": ", which is shared among all the categories. The results demonstrate that learning separate attributes per category is better than sharing global attributes among all categories." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "content": "c) " + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{w} / \\mathbf{o}\\mathcal{L}_l" + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "content": ": We explore the influence of the language-language contrastive loss " + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "content": ". The benefits of " + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "content": " are more pronounced in This-Is-My, where instances belong to different categories." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": "d) " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{w} / \\mathbf{o}\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": ": We analyze the contribution of the category-anchoring loss " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": ". Including " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " is important on This-Is-My, likely because it keeps the learning focused on the category and prevents it from capturing other scene features. This is less the case on DeepFashion2, where images are very object-centric, and there is less background variety." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "e) " + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\mathcal{B}" + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": " (w/o non-instance segments): We do not use non-instance segments " + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": " (i.e., segments of the same video but low vision-language similarity) as additional negatives in this ablation. Including " + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": " is important, as it can help prevent the instance embedding from capturing non-instance nuisance features from the scene." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19129" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 87, + 286, + 117 + ], + "blocks": [ + { + "bbox": [ + 53, + 73, + 117, + 83 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 117, + 83 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 117, + 83 + ], + "type": "text", + "content": "Language Queries" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 73, + 253, + 83 + ], + "lines": [ + { + "bbox": [ + 149, + 73, + 253, + 83 + ], + "spans": [ + { + "bbox": [ + 149, + 73, + 253, + 83 + ], + "type": "text", + "content": "Top-5 Personalized Retrievals" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 93, + 107, + 114 + ], + "lines": [ + { + "bbox": [ + 56, + 93, + 107, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 93, + 107, + 114 + ], + "type": "text", + "content": " is eating food with a white plate" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 116, + 87, + 286, + 117 + ], + "lines": [ + { + "bbox": [ + 116, + 87, + 286, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 87, + 286, + 117 + ], + "type": "image", + "image_path": "325c0220828f2c93b9e3e989a7151786af47645a95998c682cbf6350a1f1be53.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 116, + 121, + 286, + 150 + ], + "blocks": [ + { + "bbox": [ + 56, + 125, + 107, + 144 + ], + "lines": [ + { + "bbox": [ + 56, + 125, + 107, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 107, + 144 + ], + "type": "text", + "content": "man is leading " + }, + { + "bbox": [ + 56, + 125, + 107, + 144 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 56, + 125, + 107, + 144 + ], + "type": "text", + "content": "Zak's dog Coffee> with a dog leash" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 116, + 121, + 286, + 150 + ], + "lines": [ + { + "bbox": [ + 116, + 121, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 116, + 121, + 286, + 150 + ], + "type": "image", + "image_path": "38ee14909a757f87ca57d54923524830cd3430f5ecaac3a00cf26b4ba2a8f4ee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 116, + 155, + 286, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 156, + 115, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 182 + ], + "type": "text", + "content": " is with a computer monitor showing 1:12 or the screen" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 116, + 155, + 286, + 182 + ], + "lines": [ + { + "bbox": [ + 116, + 155, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 116, + 155, + 286, + 182 + ], + "type": "image", + "image_path": "204fd41855bab52074f33c7febc1597354df3685b5f3a3fb4bf85dc78d34b5f4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 116, + 191, + 286, + 218 + ], + "blocks": [ + { + "bbox": [ + 56, + 191, + 106, + 217 + ], + "lines": [ + { + "bbox": [ + 56, + 191, + 106, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 191, + 106, + 217 + ], + "type": "text", + "content": "a man is putting into a black case near red cases" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 116, + 191, + 286, + 218 + ], + "lines": [ + { + "bbox": [ + 116, + 191, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 116, + 191, + 286, + 218 + ], + "type": "image", + "image_path": "6d3a9141a56783e276ad130c492e950e39ed0293ae554e978d35976fc4cd470b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 286, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 286, + 266 + ], + "type": "text", + "content": "Figure 6. Contextualized This-Is-My Retrievals. We show personalized query-time retrievals for four This-Is-My instances. Search prompts are shown on the left and correct retrievals are highlighted in green." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "content": "f) w/o pre-trained " + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "content": ": In this case, we randomly initialize " + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "content": " instead of using a meta-personalized " + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 267, + 286, + 327 + ], + "type": "text", + "content": ". Compared to a) the sharing of category attributes among fashion items in DeepFashion2 shows clear benefits. This is less the case for This-Is-My, where instances belong to different categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 335, + 260, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 260, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 260, + 346 + ], + "type": "text", + "content": "5.2. Personal Instance Retrieval Benchmarks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 354, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 286, + 449 + ], + "type": "text", + "content": "Baselines. To demonstrate the effectiveness of our approach, we compare it to prior work [5] and the following strong baselines: CLIP (visual), which represents each instance as the average visual embedding of the corresponding training examples, CLIP (language), which represents each instance with a 0-shot prompt embedding of the corresponding category name, and CLIP " + }, + { + "bbox": [ + 46, + 354, + 286, + 449 + ], + "type": "inline_equation", + "content": "(\\mathbf{V} + \\mathbf{L})" + }, + { + "bbox": [ + 46, + 354, + 286, + 449 + ], + "type": "text", + "content": " , which uses the average of the visual and language embedding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 450, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 665 + ], + "type": "text", + "content": "Named Video Instance Retrieval on This-Is-My. We evaluate our approach on the 15 named instances in our This-Is-My test-time personalization dataset " + }, + { + "bbox": [ + 46, + 450, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 450, + 286, + 665 + ], + "type": "text", + "content": ". To learn personal instance tokens, we train on video shots that occur in the video where the instance was named. All other video shots (belonging to other videos) are in the retrieval corpus for evaluation. Furthermore, we also include all other shots that do not contain the instance in the retrieval corpus as distractors. For contextualized retrieval we use the manually collected text captions from the query-time dataset " + }, + { + "bbox": [ + 46, + 450, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 450, + 286, + 665 + ], + "type": "text", + "content": " as queries. We compare against the baselines and report contextualized retrieval performance on the left and generic instance retrieval performance on the right of Table 2. Our model clearly outperforms the baselines in both settings. Interestingly, the visual and language baselines have opposing advantages in generic vs. contextualized retrieval, while our model performs well in both settings. Qualitative retrieval results are shown in Figures 6." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Fashion Item Retrieval on DeepFashion2. We follow the setup of [5] and consider DeepFashion2 [10] for personalized fashion item retrieval. The dataset in this setting consists of 653 training and 221 evaluation images across 50" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 319, + 116, + 531, + 198 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": "Table 2. This-Is-My Video Instance Retrieval Task. We report personal instance retrieval performance in language-specified contexts (e.g., \"* catching a pink frisbee\") on the left and generic instance retrieval (e.g., \"An image of *\") on the right." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 116, + 531, + 198 + ], + "lines": [ + { + "bbox": [ + 319, + 116, + 531, + 198 + ], + "spans": [ + { + "bbox": [ + 319, + 116, + 531, + 198 + ], + "type": "table", + "html": "
MethodContext. Retr.Generic Retr.
MRRR@5mAPMRR
Random0.0±0.00.0±0.01.1±0.22.2±1.3
CLIP (language)30.8±0.036.7±0.016.6±0.044.2±0.0
CLIP (visual)10.3±1.012.0±1.648.0±0.675.0±3.2
CLIP (V+L)20.9±1.623.3±2.151.7±0.481.9±0.0
Ours42.0±1.350.7±1.856.4±0.687.4±1.2
", + "image_path": "147261c9f630c44c669e510376fc154e8db36166a1be1b34e785be709099ea4e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 318, + 250, + 531, + 371 + ], + "blocks": [ + { + "bbox": [ + 305, + 201, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 245 + ], + "type": "text", + "content": "Table 3. Fashion Item Retrieval on DeepFashion2. We evaluate our approach on the personalized instance retrieval task with contextualized queries from [5] (left) and generic instance retrieval (right). Results with * use ViT-B/32 instead of ViT-B/16." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 250, + 531, + 371 + ], + "lines": [ + { + "bbox": [ + 318, + 250, + 531, + 371 + ], + "spans": [ + { + "bbox": [ + 318, + 250, + 531, + 371 + ], + "type": "table", + "html": "
MethodContext. Retr.Generic Retr.
MRRR@5mAPMRR
Random2.9±0.21.8±0.54.7±0.59.5±2.0
CLIP (language)21.2±0.025.3±0.08.3±0.016.9±0.0
CLIP (visual)14.2±0.317.3±0.320.6±0.443.7±1.1
CLIP (V+L)20.8±0.825.4±1.720.5±0.542.8±1.2
Adapter* [5]5.9±0.7---
COLLIE* [5,29]7.9±0.7---
PALAVRA* [5]28.4±0.739.2±1.3--
Ours*34.4±0.745.2±1.140.0±1.069.3±1.8
Ours38.4±0.451.4±0.453.4±0.477.7±0.6
", + "image_path": "bb719e136c6c5b74199099ac08379dd87ae8db6acefce70a16354afc7e1be294.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 383, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 551 + ], + "type": "text", + "content": "instances (i.e., unique fashion items). While our focus is on video retrieval, our model and training objective can be used for image retrieval with minimal adjustments. We perform meta-personalization by pre-training on other non-instance-labeled images of DeepFashion2 (using only a single training image per instance). For a fair comparison with [5], the instance tokens are learned by randomly sampling " + }, + { + "bbox": [ + 304, + 383, + 545, + 551 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 304, + 383, + 545, + 551 + ], + "type": "text", + "content": " images per instance for training. We compare to results from [5] and the baselines in Table 3. Results using contextualized queries provided by [5] are on the left, and generic instance retrieval performance on the right. Note that [5] does not provide results for generic instance retrieval so we leave these two columns empty (-). Our approach outperforms the baselines and prior work by a large margin." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We have introduced a meta-personalization approach for learning to retrieve named instances in video given a natural language query. We demonstrated the effectiveness of our approach on two datasets and showed that it outperforms strong baselines. Our effort is a step towards vision-language models trained over a large number of general and personal concepts. Our approach opens up the possibility of personalized language-based video editing, summarization, and generation, e.g., identify the key instances in a collection of footage and create an edited story about the instances using natural language." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19130" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "text", + "content": "[2] Fernando Amat, Ashok Chandrashekar, Tony Jebara, and Justin Basilico. Artwork personalization at netflix. In Proceedings of the 12th ACM conference on recommender systems, pages 487-488, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 246, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 289 + ], + "type": "text", + "content": "[3] Soulef Benhamdi, Abdesselam Babouri, and Raja Chiky. Personalized recommender system for e-learning environment. Education and Information Technologies, 22(4):1455-1477, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "text", + "content": "[4] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "text", + "content": "[5] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. “This is my unicorn, Fluffy”: Personalizing frozen vision-language representations. In ECCV, 2022. 2, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "type": "text", + "content": "[6] Chelsea Finn, P. Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "text", + "content": "[7] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In International Conference on Machine Learning, pages 1920-1930. PMLR, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "type": "text", + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. ArXiv, abs/2208.01618, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 557 + ], + "type": "text", + "content": "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 624 + ], + "type": "text", + "content": "[10] Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaou Tang, and Ping Luo. Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 689 + ], + "type": "text", + "content": "[11] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5006-5015, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[12] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 547, + 161 + ], + "type": "text", + "content": "[13] De-An Huang, S. Buch, Lucio M. Dery, Animesh Garg, Li Fei-Fei, and Juan Carlos Niebles. Finding \"it\": Weakly-supervised reference-aware visual grounding in instructional videos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5948–5957, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 547, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 547, + 226 + ], + "type": "text", + "content": "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "text", + "content": "[15] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv preprint arXiv:1909.12488, 2019.2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "text", + "content": "[16] Alexander Kunitsyn, Maksim Kalashnikov, Maksim Dzabraev, and Andrei Ivaniuta. Mdmmt-2: Multidomain multimodal transformer for video retrieval, one more step towards generalization. arXiv preprint arXiv:2203.07086, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "type": "text", + "content": "[17] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "type": "text", + "content": "[18] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[19] Shuang Li, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, and Igor Mordatch. Composing ensembles of pretrained models via iterative consensus. arXiv preprint arXiv:2210.11522, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[20] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "text", + "content": "[21] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "type": "text", + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[23] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. Technical report, Tech. Rep., Technical report, OpenAI, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19131" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[25] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 239 + ], + "type": "text", + "content": "[27] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 295 + ], + "type": "text", + "content": "[28] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[29] Gabriel Skantze and Bram Willemsen. Collie: Continual learning of language grounding from language-image embeddings. Journal of Artificial Intelligence Research, 74:1201-1223, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 374 + ], + "type": "text", + "content": "[30] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "text", + "content": "[31] Yu Sun, X. Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 475 + ], + "type": "text", + "content": "[32] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and D. Erhan. Phenaki: Variable length video generation from open domain textual description. ArXiv, abs/2210.02399, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "text", + "content": "[33] Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, and Juan Miguel Pino. Fairoq s2t: Fast speech-to-text modeling with faireq. In AACL, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 543 + ], + "type": "text", + "content": "[34] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. ArXiv, abs/2109.01087, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[35] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "text", + "content": "[36] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[37] Sangmin Woo, Jinyoung Park, Inyong Koo, Sumin Lee, Minki Jeong, and Changick Kim. Explore and match: End-to-end video grounding with transformer. arXiv preprint arXiv:2201.10168, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[38] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 450 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[39] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[40] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "type": "text", + "content": "[41] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "type": "text", + "content": "[42] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "text", + "content": "[43] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 375, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 450 + ], + "type": "text", + "content": "[44] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 2, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19132" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_content_list.json b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..df5a07ca85c1a424718b8a9d1ded3750bb877761 --- /dev/null +++ b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "Meta-tuning Loss Functions and Data Augmentation for Few-shot Object Detection", + "text_level": 1, + "bbox": [ + 116, + 128, + 854, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Berkan Demirel1,2 Orhun Bugra Baran1 Ramazan Gokberk Cinbis1 \n1Middle East Technical University 2HAVELSAN Inc.", + "bbox": [ + 192, + 202, + 772, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bdemirel@havelsan.com.tr bugra@ceng.metu.edu.tr gcinbis@ceng.metu.edu.tr", + "bbox": [ + 147, + 242, + 815, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot object detection, the problem of modelling novel object detection categories with few training instances, is an emerging topic in the area of few-shot learning and object detection. Contemporary techniques can be divided into two groups: fine-tuning based and meta-learning based approaches. While meta-learning approaches aim to learn dedicated meta-models for mapping samples to novel class models, fine-tuning approaches tackle few-shot detection in a simpler manner, by adapting the detection model to novel classes through gradient based optimization. Despite their simplicity, fine-tuning based approaches typically yield competitive detection results. Based on this observation, we focus on the role of loss functions and augmentations as the force driving the fine-tuning process, and propose to tune their dynamics through meta-learning principles. The proposed training scheme, therefore, allows learning inductive biases that can boost few-shot detection, while keeping the advantages of fine-tuning based approaches. In addition, the proposed approach yields interpretable loss functions, as opposed to highly parametric and complex few-shot meta-models. The experimental results highlight the merits of the proposed scheme, with significant improvements over the strong fine-tuning based few-shot detection baselines on benchmark Pascal VOC and MS-COCO datasets, in terms of both standard and generalized few-shot performance metrics.", + "bbox": [ + 75, + 325, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 752, + 209, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection is one of the computer vision problems that has greatly benefited from the advances in supervised deep learning approaches. However, similar to the case in many other problems, state-of-the-art in object detection relies on the availability of large-scale fully-annotated datasets, which is particularly problematic due to the difficulty of collecting accurate bounding box annotations [18,46]. This practical burden has lead to a great interest in the approaches", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69104d70ca31758b63e72fb1d45a40fc0573a5278fa7cf9a8cbd4dfd31242d74.jpg", + "image_caption": [ + "Figure 1. The overall architecture of the meta-tuning approach." + ], + "image_footnote": [], + "bbox": [ + 503, + 272, + 890, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that can potentially reduce the annotation cost, such as weakly-supervised learning [29, 57], learning from point annotations [7], and mixed supervised learning [45]. A more recently emerging paradigm in this direction is few-shot object detection (FSOD). In the FSOD problem, the goal is to build detection models for the novel classes with few labeled training images by transferring knowledge from the base classes with a large set of training images. In the closely related Generalized-FSOD (G-FSOD) problem, the goal is to build few-shot detection models that perform well on both base and novel classes.", + "bbox": [ + 496, + 444, + 893, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "FSOD methods can be categorized into meta-learning and fine-tuning approaches. Although meta-learning based methods are predominantly used in the literature in FSOD research [8,22,31,36,52,75,76,79,81,83], several fine-tuning based works have recently reported competitive results [6, 15,32,53,61,65,72,84]. The main premise of meta-learning approaches is to design and train dedicated meta-models that map given few train samples to novel class detection models, e.g. [73] or learn easy-to-adapt models [30] in a MAML [16] fashion. In contrast, however, fine-tuning based methods tackle the problem as a typical transfer learning problem and apply the general purpose supervised training techniques, i.e. regularized loss minimization via gradient-based optimization, to adapt a pre-trained model to few-shot classes. It is also worth noting that the recent results on fine-tuning based FSOD are aligned with related observations on few-shot classification [9, 12, 63] and segmentation [4].", + "bbox": [ + 496, + 611, + 895, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While some of the FSOD meta-learning approaches are attractive for being able to learn dedicated parametric training", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7339", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mechanisms, they also come with two important shortcomings: (i) the risk of overfitting to the base classes used for training the meta-model due to model complexity, and (ii) the difficulty of interpreting what is actually learned; both of which can be crucially important for real-world, in-the-wild utilization of a meta-learned model. From this point of view, the simplicity and generality of a fine-tuning based FSOD approach can be seen as major advantages. In fact, one can find a large machine learning literature on the components (optimization techniques, loss functions, data augmentation, and architectures) of an FT approach, as opposed to the unique and typically unknown nature of a meta-learning inference model, especially when the model aims to replace standard training procedures for modeling the novel few-shot classes. While MAML [16] like meta-learning for quick adaptation is closer in nature to fine-tuning based approaches, the vanishing gradient problems and the overall complexity of the meta-learning task practically limits the approach to target only one or few model update steps, whereas an FT approach has no such computational difficulty.", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Perhaps the biggest advantage of a fine-tuning based FSOD approach, however, can also be its biggest disadvantage: its generality may lack the inductive biases needed for effective learning with few novel class samples while preserving the knowledge of base classes. To this end, such approaches focus on the design of fine-tuning details, e.g. whether to freeze the representation parameters [65], use contrastive fine-tuning losses [61], increase the novel class variances [84], introduce the using additional detection heads and branches [15, 72]. However, optimizing such details specifically for few-shot classes in a hand-crafted manner is clearly difficult, and likely to be sub-optimal.", + "bbox": [ + 76, + 396, + 472, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, we focus on applying meta-learning principles to tune the loss functions and augmentations to be used in the fine-tuning stage for FSOD, which we call meta-tuning (Figure 1). More specifically, much like the meta-learning of a meta-model, we define an episodic training procedure that aims to progressively discover the optimal loss function and augmentation details for FSOD purposes in a data-driven manner. Using reinforcement learning (RL) techniques, we aim to tune the loss function and augmentation details such that they maximize the expected detection quality of an FSOD model obtained by fine-tuning to a set of novel classes. By defining meta-tuning over well-designed loss terms and an augmentation list, we restrict the search process to effective function families, reducing the computational costs compared to AutoML methods that aim to discover loss terms from scratch for fully-supervised learning [20, 42]. The resulting meta-tuned loss functions and augmentations, therefore, inject the learned FSOD-specific inductive biases into a fine-tuning based approach.", + "bbox": [ + 76, + 580, + 472, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To explore the potential of the meta-tuning scheme for FSOD, we focus on the details of classification loss func", + "bbox": [ + 76, + 869, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions, based on the observations that FSOD prediction mistakes tend to be in classification rather than localization details [61]. In particular, we first focus on the softmax temperature parameter, for which we define two versions: (i) a simple constant temperature, and (ii) time (fine-tuning iteration index) varying dynamic temperature, parameterized as an exponentiated polynomial. In all cases, the parameters learned via meta-tuning yield an interpretable loss function that has a negligible risk of over-fitting to the base classes, in contrast to a complex meta-model. We also model augmentation magnitudes during meta-tuning for improving the data loading pipeline for few-shot learning purposes. Additionally, we incorporate a score scaling coefficient for learning to balance base versus novel class scores.", + "bbox": [ + 496, + 90, + 893, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide an experimental analysis on the Pascal VOC [13] and MS-COCO [40] benchmarks for FSOD, using the state-of-the-art fine-tuning based baselines MPSR [72] and DeFRCN [53]. Our experimental results show that the proposed meta-tuning approach provides significant performance gains in both FSOD and Generalized FSOD settings, suggesting that meta-tuning loss functions and data augmentation can be a promising direction in FSOD research.", + "bbox": [ + 496, + 301, + 895, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 439, + 640, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section provides an overview of recent developments on few-shot image classification, few-shot object detection, automated loss function and data augmentation discovery.", + "bbox": [ + 496, + 465, + 893, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot classification. Most of the meta-learning approaches for few-shot learning (FSL) of classification models can be grouped as adaptation-based and mapping-based approaches. Adaptation-based (also called gradient-based) approaches aim to learn model parameters that can easily be adapted to new unseen few-shot tasks within a few model update steps, e.g. [17, 38, 47, 48, 51, 54, 58]. Mapping-based approaches (also called metric-based) aim to bypass a gradient-descent based adaptation step, and instead learn a data-to-classifier mapping, e.g. [5, 44, 49, 59, 60, 62, 64, 77, 78, 80].", + "bbox": [ + 496, + 516, + 895, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some of the other notable approaches include learning to generate synthetic data for novel classes [23, 33, 68], using better feature representations [1, 2, 19, 28, 41, 63, 67] or utilizing differentiable convex solvers [3, 34]. Importantly, several works highlight that a carefully trained representation combined with simple fine-tuning or even just shallow classifiers can yield competitive or better performance than meta-learning based approaches, e.g. [9, 12, 63].", + "bbox": [ + 496, + 667, + 895, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot object detection. The FSOD approaches can be summarized as meta-learning and fine-tuning (also called transfer-learning) based ones. Most meta-learning based FSOD approaches embrace formulations similar to those used in mapping-based meta-learning approaches for FSL, e.g. [8, 22, 31, 36, 52, 75, 76, 79, 81, 83]. Support feature aggregation is one of the main aspects where meta-learning-", + "bbox": [ + 496, + 795, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7340", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based methods differ from each other. Xiao and Marlet [75] use both the differences and the channel-wise multiplication of the features in addition to the combination of the features directly for support-query aggregation. Fan et al. [14] use attention blocks to make support and query features more distinguishable for base and novel object classes. Zhang et al. [81] use inter-class correlations to highlight important support features. Li et al. [36] propose to use specialized support and query features for classification and localization.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent efforts towards improving meta-learning based FSOD include complimentary techniques, mainly to improve loss functions, feature matching, and novel class sample usage efficiency. [36] uses class margin loss, [26] uses margin-based ranking loss, [82] uses hybrid loss which consist of focal loss, adaptive margin loss and contrastive loss. Hu et al. [27] perform feature matching between query and support images to use the information from the support images more effectively. Similarly, Han et al. [21] construct a matching network between query and support instances using heterogeneous graph convolutional networks. Li and Li [35] augment novel class samples via adding Gaussian noise. Yin et al. [79] decouple classification task from localization by using the proposed class-conditional architecture.", + "bbox": [ + 75, + 229, + 472, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning-based methods typically freeze parts of a pre-trained detection network, add auxiliary detection heads, increase the novel class variances and then apply gradient descent based model update steps, unlike meta-learning-based methods that use complex episodic learning [15, 32, 53, 61, 65, 71, 72]. Wang et al. [65] propose a Faster-RCNN [56] based approach, where the class-agnostic region proposal network (RPN) component is kept frozen during fine-tuning. Sun et al. [61] use a similar approach and differently include FPN and RPN layers to the learnable parameter set in the same architecture. These learnable layers allow using contrastive proposal encodings that facilitate the more accurate classification of novel objects. Wu et al. [72] show that the scale distribution of support set tends to be imbalanced, and proposes a multi-scale positive sample refinement (MPSR) branch as an addition to the main model. Fan et al. [15] propose Retentive R-CNN architecture to prevent forgetting during fine-tuning for base classes. The obtained object proposals are fed into two ROI detectors responsible for base class and novel class instances. Qiao et al. [53] focus on decoupling network modules, and introduce a gradient decoupling layer and prototypical calibration block. Kaul et al. [32] extend the novel class annotations in the training set. In this context, the proposed method obtains object candidates from the base detector, and applies the box refinement step.", + "bbox": [ + 75, + 444, + 472, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While our approach is based on fine-tuning based FSOD, we embrace meta-learn principles to optimize the loss function and augmentations to improve the fine-tuning process for FSOD, without learning a complex and over-fitting-prone", + "bbox": [ + 76, + 839, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "meta-model. The resulting loss function and data augmentations are then utilized within the fine-tuning steps.", + "bbox": [ + 498, + 90, + 893, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Automated loss function discovery. Loss function discovery is an emerging AutoML topic towards improving the learning systems in a data-driven manner. Existing methods are mainly based on either (i) constructing the loss function directly from the basic operators [20, 42, 55] or (ii) optimizing parameterized loss functions [37, 66]. For loss construction, [42] proposes a genetic algorithm that consists of loss function verification and quality filtering modules. In this approach, the predefined proxy task eliminates divergent and poor candidate loss functions and survives the promising loss functions for other steps. [20] uses a genetic algorithm to select candidate loss functions from a tree of simple mathematical operations, and the successful loss functions pass to other stages to mutate. [55] suggests a method to learn not only the loss function but also the whole machine learning algorithm from scratch. For loss optimization, [37] re-analyzes the existing loss functions and presents them in a combined formula. [66] observes that the search space used in [37] can be too complex, and propose to simplify the search space via heuristics. In contrast to these works targeting supervised training scenarios, we aim to adapt loss function learning principles to the FSOD problem.", + "bbox": [ + 496, + 126, + 893, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AutoML for data augmentation. A variety of automated data augmentation techniques have recently been proposed [10, 11, 25, 39]. Cubuk et al. [10] generate augmentation policies using reinforcement learning and a controller RNN. Ho et al. [25] propose a method that reduces the computational costs compared to [10] by using a population-based framework. Similarly, Lim et al. [39] propose a direct Bayesian method to reduce costs. Cubuk et al. [11] show that the optimal augmentation magnitudes tend to be similar across transformations, and the search process can greatly be simplified by using a shared value. We follow this suggestion and use a shared magnitude across the transforms in our formulation. In contrast to these works on supervised learning, however, we focus on learning detectors with few-samples.", + "bbox": [ + 496, + 464, + 895, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, while loss function and augmentation discovery topics increasingly attract attention towards improving supervised training pipelines, ours is the first work on learning few-sample specific inductive biases for fine-tuning based few-shot object detection based on meta-learning and AutoML principles, to the best of our knowledge.", + "bbox": [ + 496, + 675, + 893, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 779, + 589, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides a brief summary of the FSOD problem definition and the baseline model we utilize. We then present our definition and instantiation of meta-tuning.", + "bbox": [ + 498, + 804, + 893, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem definition. We follow the FSOD setup of [31], where a relatively large set of training images for the set $C_b$ of base classes is made available. Each training im", + "bbox": [ + 498, + 854, + 895, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7341", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "age corresponds to a tuple $(x,y)$ consisting of image $x$ and annotations $y = \\{y_0,\\dots,y_M\\}$ . Each object annotation $y_{i} = \\{c_{i},b_{i}\\}$ contains a category label $(c_{i})$ and a bounding box $(b_{i} = \\{x_{i},y_{i},w_{i},h_{i}\\})$ . Once the FSOD model training is complete, the evaluation is carried out based on a limited number $(k)$ of training images made available for the set $C_n$ of distinct novel (i.e. few-shot) classes.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Base model. We use the MPSR FSOD method [72] as the infrastructure for our loss function and data augmentation search methods. MPSR adapts the Faster-RCNN to be suitable for fine-tuning-based FSOD and uses an auxiliary multi-scale positive sample refinement (MPSR) branch to handle the scale scarcity problems. This branch expands the scale space of positive samples without increasing improper negative instances, unlike feature pyramid networks and image pyramids that do not change data distribution, hence the scale sparsity problem. In this context, objects in the images are cropped and resized in multiple sizes to create scale pyramids. The MPSR uses two groups of loss functions for the region proposal network (RPN) and detection heads, and feeds differently scaled positive samples to these loss functions together with the main detection branch. Finally, we note that the proposed approach can in principle be applied to virtually any fine-tuning based FSOD model.", + "bbox": [ + 75, + 200, + 472, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Meta-tuning loss functions", + "text_level": 1, + "bbox": [ + 76, + 467, + 316, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our main goal is to improve few-shot detector fine-tuning based on meta-learning principles. For meta-tuning the FSOD loss, we specifically focus on the classification loss term, as the FSOD errors tend to be primarily caused by misclassifications [61]. The MPSR classification loss term can be expressed as follows:", + "bbox": [ + 75, + 489, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c l s} (x, y) = - \\frac {1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {f \\left(x _ {i} , y _ {i}\\right)}}{\\sum_ {y} e ^ {f \\left(x _ {i} , y\\right)}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 590, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{ROI}$ is the number of ROIs (i.e. candidate regions) in an image, $y_{i}$ is the groundtruth class label for the $i$ -th ROI, and $f(x_{i},y)$ is the corresponding class $y$ prediction score. To add more flexibility into the loss function, we re-define it as a parametric function $\\ell_{cls}(x,y;\\rho)$ , where $\\rho$ represents the loss function parameters. First, we introduce a temperature scalar $\\rho_{\\tau}$ , i.e. $\\rho = (\\rho_{\\tau})$ :", + "bbox": [ + 75, + 640, + 468, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c l s} (x, y; \\rho) = - \\frac {1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {f \\left(x _ {i} , y _ {i}\\right) / \\rho_ {\\tau}}}{\\sum_ {y ^ {\\prime}} e ^ {f \\left(x _ {i} , y ^ {\\prime}\\right) / \\rho_ {\\tau}}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 755, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our motivation comes from the observations on the importance of temperature scaling in log loss on various other problems, such as knowledge distillation [24], few-shot classification [49, 78], and zero-shot learning [43]. While temperature is typically tuned in a manual manner, here we aim to meta-learn it specifically for fine-tuning based FSOD", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "purposes, giving a chance to observe the behavior of metatuning in a simple case. We also define a more sophisticated variant of the loss function by defining the dynamic temperature function $f_{\\rho}$ and novel class scaling $\\alpha$ :", + "bbox": [ + 496, + 90, + 893, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c l s} (x, y; \\rho) = \\frac {- 1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {\\alpha \\left(y _ {i}\\right) f \\left(x _ {i} , y _ {i}\\right) / f _ {\\rho} (t)}}{\\sum_ {y ^ {\\prime}} e ^ {\\alpha \\left(y ^ {\\prime}\\right) f \\left(x _ {i} , y ^ {\\prime}\\right) / f _ {\\rho} (t)}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 162, + 890, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\rho}(t) = \\exp (\\rho_{a}t^{2} + \\rho_{b}t + \\rho_{c})$ . Here, $\\rho = (\\rho_{a},\\rho_{b},\\rho_{c})$ is a 3-tuple of polynomial coefficients, and $t\\in [0,1]$ is the normalized fine-tuning iteration index. The temperature can increase or decrease over time, making the predicted class distributions smoother or sharper. $\\alpha (y)$ is set to 1 for $y\\in C_b$ , and otherwise the novel class score scaling coefficient $\\rho_{\\alpha}$ , as a way to learn base and novel score balancing.", + "bbox": [ + 496, + 215, + 893, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Meta-tuning augmentations", + "text_level": 1, + "bbox": [ + 500, + 330, + 750, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For meta-tuning augmentations, we focus on the photometric augmentations that are likely to be transferable from base to novel classes. In this context, we model the brightness, saturation, contrast, and hue transforms, with a shared magnitude parameter $(\\rho_{aug})$ , which is known to be effective for supervised training [11].", + "bbox": [ + 496, + 354, + 893, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Meta-tuning procedure", + "text_level": 1, + "bbox": [ + 500, + 454, + 714, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our work, we utilize a REINFORCE [70] based reinforcement learning (RL) approach to search for the optimal loss function and augmentations, where we use the AutoML approach of Wang et al. [66] on loss function search for fully-supervised face recognition as our starting point.", + "bbox": [ + 496, + 477, + 893, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to meta-tune the loss function and augmentations to maximize FSOD generalization abilities, we generate proxy tasks over base class training data to imitate real FSOD tasks over the novel classes. For this purpose, we divide base classes into two subsets, proxy-base $C_{\\mathrm{p - base}}$ and proxy-novel $C_{\\mathrm{p - novel}}$ . We then construct three non-overlapping data set splits using the base class training set: (i) $D_{\\mathrm{p - pretrain}}$ containing $C_{\\mathrm{p - base}}$ -only samples, used for training a temporary object detection model for meta-tuning purposes; (ii) $D_{\\mathrm{p - support}}$ containing samples of $C_{\\mathrm{p - base}} \\cup C_{\\mathrm{p - novel}}$ classes to be used as fine-tuning images during meta-tuning; (iii) $D_{\\mathrm{p - query}}$ containing samples of $C_{\\mathrm{p - base}} \\cup C_{\\mathrm{p - novel}}$ classes to be used for evaluating the generalized FSOD performance of a fine-tuned model during meta-tuning.", + "bbox": [ + 496, + 554, + 893, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We generate a series of FSOD proxy tasks for metatuning, similar to episodic meta-learning: at each proxy task $T$ , we sample a few-shot training set from $D_{\\mathrm{p - support}}$ . We also sample a loss function/augmentation magnitude parameter combination $\\rho$ , where each $\\rho_j \\in \\rho$ is modeled in terms of a Gaussian distribution: $\\rho_j \\sim \\mathcal{N}(\\mu_j, \\sigma^2)$ . Using the loss function or augmentations corresponding to the sampled $\\rho$ , we fine-tune the initial model on the support images using gradient-based optimization, and compute the mean", + "bbox": [ + 496, + 765, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7342", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff3bc558467eb304c8dc5cfa596f6af2101a5be28269c400feb2da6860cc9a5c.jpg", + "image_caption": [ + "Figure 2. The meta-tuning approach. At each RL iteration over a proxy task, the distribution parameters modeling the loss function and augmentations are updated as a function of the obtained mAP scores, towards improved training with few-samples." + ], + "image_footnote": [], + "bbox": [ + 81, + 108, + 874, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "average precision (mAP) scores on $D_{\\mathrm{p - query}}$ . We get multiple mAP scores by repeating this process multiple times over multiple proxy support samples. Meta-tuning is then carried over by updating $\\mu$ values via the REINFORCE rule after each episode, towards finding $\\mu$ values centered around well-performing $\\rho$ combinations.", + "bbox": [ + 75, + 342, + 472, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {j} ^ {\\prime} \\leftarrow \\mu_ {j} + \\eta R (\\rho) \\nabla_ {\\mu} \\log (p \\left(\\rho_ {j}; \\mu_ {j}, \\sigma\\right)) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 445, + 468, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(\\rho ;\\mu ,\\sigma)$ is the Gaussian probability density function, $\\eta$ is the RL learning rate.", + "bbox": [ + 75, + 473, + 470, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply the REINFORCE update rule using the $\\rho$ with the highest reward per episode. $R(\\rho)$ is the normalized reward function obtained by whitening the mAP scores. We empirically observe that normalization improves the results (Section 4) since without reward normalization, the RL updates are scaled with respect to the inherent difficulty of the proxy task, which greatly varies depending on the sampled support examples. Reward normalization approximately removes the average reward, enabling better performing $\\rho$ samples to influence based on their relative success.", + "bbox": [ + 75, + 503, + 470, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, similar to [50], starting with $\\sigma = 0.1$ , we diminish $\\sigma$ over the RL iterations to progressively reduce explorations by sampling more conservatively, which improves converge. The final scheme is illustrated in Figure 2.", + "bbox": [ + 75, + 654, + 470, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 728, + 209, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We use mAP to evaluate the base and novel class detection results separately. To evaluate the generalized FSOD performance, we use the Harmonic Mean (HM) metric to compute a balanced aggregation of base and novel class performance scores. Adapted from generalized zero-shot learning [74], HM is defined as the harmonic mean of $\\mathrm{mAP}_{\\mathrm{base}}$ and $\\mathrm{mAP}_{\\mathrm{novel}}$ scores.", + "bbox": [ + 75, + 758, + 470, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use Pascal VOC [13] and MS COCO [40] with the same splits defined in FSOD benchmarks [65, 72].", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On Pascal VOC, three separate base/novel class splits exist, where each one consists of 15 base and 5 novel classes. In each split, we select 5 base classes to mimic novel classes during meta-tuning. On MS-COCO, we select 15 base classes to mimic novel classes in each proxy task, and evaluate the models for the 10-shot and 30-shot settings.", + "bbox": [ + 496, + 342, + 893, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We primarily use the MPSR [72] and DeFRCN [53] as our baselines, which are among the best performing fine-tuning based FSOD methods on Pascal VOC. For the DeFRCN experiments, we transfer the meta-tuned loss functions and augmentation magnitudes from MPSR to the DeFRCN method, which are both based on Faster-RCNN. We take the results for FRCN [76], Ret. R-CNN [15], Meta-RCNN [76], FSRW [31], MetaDet [69], FsDetView [75] and ONCE [52] from [15] for a fair comparison. For the MPSR, DeFRCN (seed is set to 0) and FSCE [61], we report the results we obtain experimentally. We take the results for TFA+Hal [84], CME [36], TIP [35], DCNet [27], QA-FewDet [21] FADI [6], LVC [32], KFSOD [83] and FCT [22] from the original papers. Finally, while it is difficult to fairly compare fine-tuning versus meta-learning based approaches, we provide a discussion in the supplementary material.", + "bbox": [ + 496, + 438, + 893, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We use 200 RL episodes for loss function meta-tuning, with REINFORCE learning rate set to 0.0005. The meta-tuning for augmentation parameter is carried out using the trained and frozen the loss function parameters. We keep the fine-tuning implementation details of MPSR unchanged, which uses 4000 and 8000 gradient descent iterations for 10-shot and 30-shot experiments on MS-COCO, and 2000 iterations on Pascal VOC. We will publish the full source code upon publication; a preliminary version is provided as supplementary material.", + "bbox": [ + 496, + 685, + 893, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Main results", + "text_level": 1, + "bbox": [ + 500, + 847, + 633, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first compare the meta-tuning results against the corresponding MPSR baseline in Table 1. In the table, Meta", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7343", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/74f2b3f2e42e398d215c6c4b03511810495f0e31a358a2ebba72301b018b1c7f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method/ShotPascal VOCMS-COCO
Novel ClassesAll Classes (HM)Novel ClassesAll Classes (HM)
12351012351010301030
MPSR [72]33.137.244.347.152.143.147.454.557.260.89.113.711.515.0
MPSR+Meta-Static33.439.445.147.352.643.750.455.457.561.410.114.812.716.4
MPSR+Meta-Dynamic34.539.845.048.252.545.051.055.558.361.611.914.914.316.6
MPSR+Meta-ScaledDynamic35.240.345.848.452.945.651.255.958.361.812.315.014.416.7
MPSR+Aug34.638.646.048.352.745.149.556.258.462.09.914.912.516.3
MPSR+Meta-Static+Aug35.339.146.148.452.745.949.956.258.361.810.215.212.816.7
MPSR+Meta-Dynamic+Aug35.439.646.548.953.346.050.556.858.962.512.115.314.516.8
MPSR+Meta-ScaledDynamic+Aug35.840.646.849.253.746.351.557.059.262.712.515.414.716.9
", + "bbox": [ + 102, + 88, + 870, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. FSOD (mAP) and G-FSOD (HM of the base and novel class mAPs) results on Pascal VOC and MS-COCO datasets for MPSR baseline method. HM stands for harmonic mean.", + "bbox": [ + 75, + 233, + 893, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Static, Meta-Dynamic, Meta-ScaledDynamic refer to meta-tuning a single temperature, dynamic temperature, and novel class scaled dynamic temperature functions, respectively. Similarly, Aug, Meta-Static $^+$ Aug, Meta-Dynamic $^+$ Aug, and Meta-ScaledDynamic $^+$ Aug refer to meta-tuning only augmentation, single temperature and augmentation, dynamic temperature and augmentation, and novel class scaled dynamic temperature and augmentation functions, respectively. We observe that meta-tuning consistently improves the FSOD and G-FSOD results of the MPSR model. We also observe steady improvements gradually from the baseline to Meta-Static, to Meta-Dynamic, and finally to Meta-ScaledDynamic. In addition, the meta-tuned augmentation magnitude parameter also contributes positively to the few-shot object detection performance. The overall consistency of the improvements provides positive evidence for the value of loss and augmentation meta-tuning.", + "bbox": [ + 75, + 287, + 473, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pascal VOC results. In Table 2, we report the Pascal VOC results for our MPSR and DeFRCN based Meta-ScaledDynamic+Aug approach and compare them against the state-of-the-art fine-tuning based FSOD methods. While we present the scores averaged over the three splits in this table, additional per-split FSOD and G-FSOD results can be found in the supplementary material. The left side of Table 2 presents the FSOD results for the varying number of support images. We observe that DeFRCN combined with Meta-ScaledDynamic+Aug, i.e. meta-tuning of the score coefficient, dynamic temperature and the augmentation parameter, yields the best mAP scores in all $k$ -shot settings among all methods.", + "bbox": [ + 75, + 550, + 470, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The right side of Table 2 presents the G-FSOD results on Pascal VOC. We observe that the best-performing Meta-ScaledDynamic+Aug method improves the HM scores further above the state-of-the-art in all $k$ -shot settings. Overall, these results suggest that the proposed framework is an effective way for meta-learning inductive biases to be used in fine-tuning-based FSOD.", + "bbox": [ + 75, + 748, + 470, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 presents visual detection examples without and with meta-tuned scaled dynamic temperature and augmentations in the first and second rows, respectively. We observe", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "various improvements, such as reductions in false positives, improved recall, and more precise boxes, most likely due to the improved model fitting in the low-data regime.", + "bbox": [ + 496, + 287, + 893, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MS-COCO results. In Table 3, we compare the MPSR and DeFRCN based Meta-ScaleDynamic+Aug results against other fine-tuning based FSOD methods that report 10-shot and 30-shot results on the MS-COCO dataset. We observe that with meta-tuning, the FSOD scores of MPSR improve from 9.1 to 12.5 (10-shot mAP), and from 13.7 to 15.4 (30-shot mAP). We also observe that the scores of DeFRCN improve from 18.5 to 18.8 (10-shot mAP), and from 21.9 to 23.4 (30-shot mAP), obtaining the best and second best results against all other models. Similarly, in the case of GFSD, with meta-tuning, the 10-shot HM score of DeFRCN improves from 24.0 to 24.4, outperforming all other models. In addition, the 30-shot HM score of DeFRCN improves from 26.8 to 28.0, which is slightly below the 28.1 score of LVC-PL [32].", + "bbox": [ + 496, + 340, + 895, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 580, + 661, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meta-tuning details. The proposed meta-tuning approach involves three important technical details: Proxy-novel imitation, model re-initialization, and reward normalization. Proxy-novel imitation refers to reinforcement learning over the sampled proxy-novel tasks, instead of the whole training set, to mimic the test-time FSOD challenges. Model re-initialization is the re-initialization of the base model for each task. Without re-initialization, not only the sampled loss/augmentation parameters and tasks but also the accumulated model updates undesirably affect the rewards. Reward normalization further reduces the effect of task difficulty variance by normalizing the rewards obtained within a single episode, allowing a more isolated assessment of the sampled loss functions and augmentations.", + "bbox": [ + 496, + 612, + 895, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the contributions of these three important details in terms of G-FSOD HM scores using the 5-shot setting of Pascal VOC Split-1 with MPSR+Meta-Dynamic. The results averaged over 5 runs are given in Table 4. We observe that each component progressively improves the", + "bbox": [ + 496, + 825, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7344", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0b1b24a6a210f0c6fb18443734c165d5ae6d4be3a6ea534e1cacf72623255f4a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method/ShotNovel ClassesAll Classes (HM)
123510123510
FRCN [76] (ICCV'19)16.120.628.833.436.525.931.740.044.346.7
TFA-fc [65] (ICML'20)27.630.639.846.648.740.544.152.958.359.9
TFA-cos [65] (ICML'20)31.432.640.546.848.344.646.053.558.459.6
FSCE [61] (CVPR'21)29.236.342.547.152.241.848.854.257.761.0
Ret. R-CNN [15] (CVPR'21)31.437.141.446.848.844.750.554.759.160.8
TFA+Hal [84] (CVPR'21)32.935.540.446.348.1-----
FADI [6] (NeurIPS'21)42.246.547.952.456.9-----
LVC [32] (CVPR'22)30.935.443.651.154.1-----
LVC-PL [32] (CVPR'22)45.245.054.857.558.6-----
MPSR [72] (ECCV'20)33.137.244.347.152.143.147.454.557.260.8
DeFRCN [53] (ICCV'21)46.552.655.960.060.857.662.564.767.667.8
MPSR+Meta-ScaledDynamic+Aug35.840.646.849.253.746.351.557.059.262.7
DeFRCN+Meta-ScaledDynamic+Aug49.254.057.261.361.859.863.765.968.668.7
", + "bbox": [ + 187, + 88, + 782, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c84e8ccae23d30cfef10a3a7b73b22f5d6316a4b99065d13a673d414046cfd03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 316, + 217, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07d07e33e01afe655237ea462d0df27346a0d4cdfbb053f7cfaf3c4d86cc597d.jpg", + "image_caption": [ + "Split-1/5-shot" + ], + "image_footnote": [], + "bbox": [ + 84, + 407, + 218, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9be8da8655e2408b293b01a02e818e701c1b1257b9293d94df3792dcd139c9bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 316, + 349, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33840c689ca2bf073231d0a4381805805015ddb9a7011954efdb97c8067e658a.jpg", + "image_caption": [ + "Split-1/10-shot" + ], + "image_footnote": [], + "bbox": [ + 220, + 407, + 349, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00cd92d0f5889b6ce2ac6aff11354bb9c79e68d040a21a5b7854fbf4d8bd6304.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 316, + 482, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ae390992bf7542c471325f10dc035b0f094f06d6f53708c39e863bcfa10e780.jpg", + "image_caption": [ + "Split-2/5-shot", + "Figure 3. Qualitative results using MPSR without (first row) and with (second row) meta-tuning, over multiple Pascal VOC splits. Base and novel class detections are shown with green and red boxes, respectively. (Best viewed in color.)" + ], + "image_footnote": [], + "bbox": [ + 352, + 409, + 482, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ba54553e65bf95e2dd0604a7fffd07cf776df0ecce98278d36adc2cc70c8628.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 316, + 614, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/129654a3e58852f10898f12a210faa73be234f73690e28553587a21c7f725414.jpg", + "image_caption": [ + "Split-2/10-shot" + ], + "image_footnote": [], + "bbox": [ + 483, + 409, + 614, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f6f3319f594f4eda440e44e5e8e8252502128783c0ae46b0187cd0b99fe0c49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 316, + 746, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d4fedf116c18219edb06b9359f08470a8b32ba8b790502595be9e0fb0385995.jpg", + "image_caption": [ + "Split-3/5-shot" + ], + "image_footnote": [], + "bbox": [ + 616, + 409, + 746, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/199345f0ee8602ff84564ce16d0aee3eaea222704f924c2543e8fbfb22f742d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 316, + 880, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b61ecf822828a1a2fd411c35799eb9412ec39cd0b45e31200dfc08dc0534076.jpg", + "image_caption": [ + "Split-3/10-shot" + ], + "image_footnote": [], + "bbox": [ + 750, + 409, + 880, + 500 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/308522c7f40d4e6d811db0c0f68c3aaf087f0f0ccbc358c9ed4764085a97d2f6.jpg", + "table_caption": [ + "Table 2. FSOD (mAP) and G-FSOD (HM of the base and novel class mAPs) results on Pascal VOC. The best and the second-best results are marked with red and blue. HM stands for harmonic mean." + ], + "table_footnote": [], + "table_body": "
Method/ShotsNovel ClassesAll Classes (HM)
10-shot30-shot10-shot30-shot
FRCN [76] (ICCV'19)9.212.512.815.6
FRCN-BCE [65] (ICML'20)6.410.310.916.1
TFA-fc [65] (ICML'20)10.013.415.419.4
TFA-cos [65] (ICML'20)10.013.715.619.8
MPSR [72] (ECCV'20)9.113.711.515.0
FSCE [61] (CVPR'21)10.514.416.020.2
Ret. R-CNN [15] (CVPR'21)10.513.816.620.4
FADI [6] (NeurIPS'21)12.216.1--
DeFRCN [53] (ICCV'21)18.521.924.026.8
LVC [32] (CVPR'22)12.117.817.822.8
LVC-PL [32] (CVPR'22)17.824.522.828.1
MPSR+Meta-ScaledDynamic+Aug12.515.414.716.9
DeFRCN+Meta-ScaledDynamic+Aug18.823.424.428.0
", + "bbox": [ + 78, + 577, + 493, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HM scores, and the most significant contribution is made by reward normalization, which improves from 62.1 to 63.3. We also observe that reward normalization considerably improves the overall experimental stability. To quantify this", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f0aae38acf665faaed19c4b22ded6d178baaa1ac95cd2623ab15ff49efd4d697.jpg", + "table_caption": [ + "Table 3. Comparison of Meta-ScaledDynamic results to the fine-tuning based (G-)FSOD methods on the MS-COCO dataset. The best and the second-best results are marked with red and blue." + ], + "table_footnote": [], + "table_body": "
Proxy-novel imit.Model re-init.Reward norm.HM
XXX61.5
✓XX61.8
✓✓X62.1
✓✓✓63.3
", + "bbox": [ + 540, + 577, + 852, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Evaluation of meta-tuning details. Proxy-novel imitation is the imitation of novel classes using a subset of base classes. Model re-initialization is the re-initialization of the base model at each task. Reward normalization is within-episode normalization of the mAP scores during meta-tuning.", + "bbox": [ + 496, + 656, + 892, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "observation, we estimate the $95\\%$ confidence interval over the runs using $CI = 1.96\\frac{s}{\\sqrt{n}}$ , where $s$ , $n$ , and $1.96$ are the standard deviation, number of runs, and $Z$ -value, respectively [65]. According to this estimator, the normalization step narrows the confidence interval from $\\pm 0.75$ to $\\pm 0.13$ , providing a clear improvement in reliability.", + "bbox": [ + 496, + 755, + 893, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Learned loss functions. In Figure 4, we plot the learned loss functions according to the $\\mu$ values obtained at the end of the RL process. The upper plot shows the dynamic", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7345", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/398273ef69347b6ecae7cbc7d987dc77d5389ebf546fa8f522214dff745fcfc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 108, + 413, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d30b7a0b618214bfbeeb4284573888ed1e96615f5a9d492058a0a79c92609f25.jpg", + "image_caption": [ + "Figure 4. The dynamic temperature functions and score scaling coefficients learned by the meta-tuning process, using Meta-Dynamic (upper) and Meta-ScaledDynamic (lower) formulations. Results for each Pascal VOC split is shown with a separate curve." + ], + "image_footnote": [], + "bbox": [ + 102, + 286, + 413, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "temperature functions learned over three different splits. We observe that temporally attenuated temperature values are preferred consistently, sharpening the predictions towards the end of the fine-tuning process. The lower plot shows the learned dynamic temperature functions with novel class score scaling. The learned scaling coefficients, i.e. $\\mu_{\\alpha}$ of the learned $\\rho_{\\alpha}$ distribution, are shown as horizontal lines. We observe that similar dynamic temperature functions are learned, and $\\mu_{\\alpha}$ values vary between 1.09 to 1.2, suggesting that the meta-tuning process learns to boost the novel class scores. The interpretability of these outcomes, we believe, highlights a significant advantage of loss meta-tuning. In the context of interpretability, we observe that as the fine-tuning process continues on the few-shot training set, the predictions are progressively made sharper, i.e. the loss becomes more sensitive to classification errors and enforces towards making more confident correct predictions. This is in alignment with one of our original motivations for reducing the dominating classification errors in G-FSOD, as the meta-tuning process automatically learns to enforce more accurate classifications, where the curve steepness and the numerical ranges are learned via RL.", + "bbox": [ + 75, + 546, + 470, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learned augmentations. The learned photometric augmen", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0b37a4a9f7e84081188f5cb9e35582c075198e5d0623cef1732678aead7d2648.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
S/MTFA [65]TFA+Hal [84]TFA+Meta-ScaledDynamic+Aug
13.43.84.7
24.65.05.8
36.66.97.1
", + "bbox": [ + 514, + 88, + 879, + 145 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Low-shot (1-shot, 2-shot and 3-shot) experiments on MS-COCO dataset with novel classes.", + "bbox": [ + 500, + 156, + 893, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tation magnitude values learned are 0.29, 0.24, 0.13, and 0.36 for Pascal VOC split-1, split-2, split-3, and MS-COCO datasets, respectively. We observe that the learned augmentation magnitudes positively contribute to the performance. According to the results in Table 1, the average Pascal VOC split-1/1-shot score increases from 33.1 to 34.6 with only augmentation steps.", + "bbox": [ + 498, + 215, + 893, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Very low-shot experiments. Finally, we evaluate the meta-tuning approach in low-shot many-class settings. [84] proposes TFA+Hal method that uses the TFA baseline and conducts 1-shot, 2-shot, and 3-shot FSOD on the MS-COCO dataset. As we already observe the positive effects of the loss terms and augmentation magnitudes obtained from the MPSR on the DeFRCN, we similarly apply the learned parameters to the TFA baseline. The results are presented in Table 5. We observe that results are consistently improved using the meta-tuned functions on the TFA baseline.", + "bbox": [ + 498, + 329, + 895, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 500, + 617, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning based frameworks offer simple and reliable approaches to building detection models from few samples. However, a major limitation of the existing fine-tuning-based FSOD models is their focus on the hand-crafting the design of fine-tuning details for few-shot training, which is inherently difficult and likely to be sub-optimal. Towards addressing this limitation, we propose to meta-learn the fine-tuning based learning dynamics as a way of introducing learned inductive biases for few-shot learning. The proposed tuning scheme uses meta-learning principles with reinforcement learning, and obtains interpretable loss functions and augmentation magnitudes for few-shot training. Our comprehensive experimental results on Pascal VOC and MS COCO datasets show that the proposed meta-tuning approach consistently provides significant performance improvements over the strong fine-tuning based few-shot detection baselines in both FSOD and G-FSOD settings.", + "bbox": [ + 496, + 527, + 893, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While we restrict our experiments to loss and augmentation functions, meta-tuning other learning components, e.g. initial model, and applications to other few-shot learning problems can be interesting future work directions.", + "bbox": [ + 498, + 787, + 893, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the TUBITAK Grant 119E597 and a Google Faculty Research Award.", + "bbox": [ + 498, + 854, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7346", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, and Frank Wood. Enhancing few-shot image classification with unlabelled examples. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2796-2805, January 2022. 2", + "[2] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Sigal. Improved Few-Shot Visual Classification. arXiv e-prints, page arXiv:1912.03432, Dec. 2019. 2", + "[3] Luca Bertinetto, Joao F. Henriques, Philip Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Proc. Int. Conf. Learn. Represent., 2019. 2", + "[4] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-Shot Segmentation Without Meta-Learning: A Good Transductive Inference Is All You Need? In arXiv:2012.06166 [cs], 2021. 1", + "[5] Kaidi Cao, Maria Brbić, and Jure Leskovec. Concept learners for few-shot learning. In Proc. Int. Conf. Learn. Represent., 2021. 2", + "[6] Yuhang Cao, Jiaqi Wang, Ying Jin, Tong Wu, Kai Chen, Ziwei Liu, and Dahua Lin. Few-shot object detection via association and discrimination. Proc. Adv. Neural Inf. Process. Syst., 34:16570-16581, 2021. 1, 5, 7", + "[7] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 8819-8828, 2021. 1", + "[8] Tung-I Chen, Yueh-Cheng Liu, Hung-Ting Su, Yu-Cheng Chang, Yu-Hsiang Lin, Jia-Fong Yeh, Wen-Chin Chen, and Winston Hsu. Dual-awareness attention for few-shot object detection. IEEE Transactions on Multimedia, 2021. 1, 2", + "[9] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A Closer Look at Few-shot Classification. In ICLR 2019, 2019. 1, 2", + "[10] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 3", + "[11] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proc. IEEE Conf. Comput. Vis. Pattern Recog. Workshops, pages 702-703, 2020. 3, 4", + "[12] Guneet S Dhillon, Pratik Chaudhari, Avinash Ravichandran, and Stefano Soatto. A Baseline for Few-Shot Image Classification. In ICLR, page 20, 2020. 1, 2", + "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 5", + "[14] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4013–4022, 2020. 3", + "[15] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4527-4536, 2021. 1, 2, 3, 5, 7" + ], + "bbox": [ + 78, + 114, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 1, 2", + "[17] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proc. Int. Conf. Mach. Learn., volume 70, pages 1126–1135, 2017. 2", + "[18] Golnaz Ghiasi, Yin Cui, A. Srinivas, Rui Qian, Tsung-Yi Lin, Ekin Dogus Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 2917-2927, 2021. 1", + "[19] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proc. IEEE Int. Conf. on Computer Vision, 2019. 2", + "[20] Santiago Gonzalez and Risto Miikkulainen. Improved training speed, accuracy, and data utilization through loss function optimization. In 2020 IEEE Congress on Evolutionary Computation (CEC), pages 1-8. IEEE, 2020. 2, 3", + "[21] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proc. IEEE Int. Conf. on Computer Vision, pages 3263-3272, 2021. 3, 5", + "[22] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 5321-5330, 2022. 1, 2, 5", + "[23] Bharath Hariharan and Ross B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. Proc. IEEE Int. Conf. on Computer Vision, pages 3037-3046, 2017. 2", + "[24] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 4", + "[25] Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. In Proc. Int. Conf. Mach. Learn., pages 2731-2741. PMLR, 2019. 3", + "[26] Ting-I Hsieh, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. One-shot object detection with co-attention and co-excitation. arXiv preprint arXiv:1911.12529, 2019. 3", + "[27] Hanzhe Hu, Shuai Bai, Aoxue Li, Jinshi Cui, and Liwei Wang. Dense relation distillation with context-aware aggregation for few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 10185-10194, 2021. 3, 5", + "[28] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference. arXiv e-prints, page arXiv:2204.07305, Apr. 2022. 2", + "[29] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Proc. Adv. Neural Inf. Process. Syst., 33, 2020. 1" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7347", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Taewon Jeong and Heeyoung Kim. Ood-maml: Meta-learning for few-shot out-of-distribution detection and classification. In Advances in Neural Information Processing Systems, volume 33, pages 3907-3916, 2020. 1", + "[31] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In Proc. IEEE Int. Conf. on Computer Vision, pages 8420-8429, 2019. 1, 2, 3, 5", + "[32] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Label, verify, correct: A simple few shot object detection method. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 14237-14247, 2022. 1, 3, 5, 6, 7", + "[33] Michalis Lazarou, Yannis Avrithis, and Tania Stathaki. Tensor feature hallucination for few-shot learning. ArXiv, abs/2106.05321, 2021. 2", + "[34] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 10649-10657, 2019. 2", + "[35] Aoxue Li and Zhenguo Li. Transformation invariant few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 3094-3102, 2021. 3, 5", + "[36] Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. Beyond max-margin: Class margin equilibrium for few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 7363-7372, 2021. 1, 2, 3, 5", + "[37] Chuming Li, Xin Yuan, Chen Lin, Minghao Guo, Wei Wu, Junjie Yan, and Wanli Ouyang. Am-lfs: Automl for loss function search. In Proc. IEEE Int. Conf. on Computer Vision, pages 8410-8419, 2019. 3", + "[38] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. MetaSGD: Learning to Learn Quickly for Few-Shot Learning. arXiv e-prints, page arXiv:1707.09835, July 2017. 2", + "[39] Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment. Proc. Adv. Neural Inf. Process. Syst., 32, 2019. 3", + "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conf. on Computer Vision, pages 740-755. Springer, 2014. 2, 5", + "[41] Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters: Understanding margin in few-shot classification. arXiv preprint arXiv:2003.12060, 2020. 2", + "[42] Peidong Liu, Gengwei Zhang, Bochao Wang, Hang Xu, Xiaodan Liang, Yong Jiang, and Zhenguo Li. Loss function discovery for object detection via convergence-simulation driven search. arXiv preprint arXiv:2102.04700, 2021. 2, 3", + "[43] Shichen Liu, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Generalized Zero-Shot Learning with Deep Calibration Network. In NeurIPS, pages 2005–2015. 2018. 4", + "[44] Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sungju Hwang, and Yi Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In Proc. Int. Conf. Learn. Represent., 2019. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Yan Liu, Zhijie Zhang, Li Niu, Junjie Chen, and Liqing Zhang. Mixed supervised object detection by transferring mask prior and semantic similarity. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Proc. Adv. Neural Inf. Process. Syst., 2021. 1", + "[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. Proc. IEEE Int. Conf. on Computer Vision, 2021. 1", + "[47] Tsendsumen Munkhdalai, Xingdi Yuan, Soroush Mehri, and Adam Trischler. Rapid adaptation with conditionally shifted neurons. In ICML, 2018. 2", + "[48] Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv: Learning, 2018. 2", + "[49] Boris N. Oreshkin, Pau Rodriguez Lopez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. In NeurIPS, 2018. 2, 4", + "[50] Matteo Papini, Andrea Battistello, and Marcello Restelli. Balancing learning speed and stability in policy gradient via adaptive exploration. In Proc. Int. Conf. on Artif. Intellig. and Stat., pages 1188-1199, 2020. 5", + "[51] Eunbyung Park and Junier B. Oliva. Meta-curvature. In NeurIPS, 2019. 2", + "[52] Juan-Manuel Perez-Rua, Xiatian Zhu, Timothy M Hospedales, and Tao Xiang. Incremental few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 13846-13855, 2020. 1, 2, 5", + "[53] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In Proc. IEEE Int. Conf. on Computer Vision, pages 8681-8690, 2021. 1, 2, 3, 5, 7", + "[54] Aravind Rajeswaran, Chelsea Finn, Sham M. Kakade, and Sergey Levine. Meta-learning with implicit gradients. In NeurIPS, 2019. 2", + "[55] Esteban Real, Chen Liang, David So, and Quoc Le. Automlzero: Evolving machine learning algorithms from scratch. In Proc. Int. Conf. Mach. Learn., pages 8007-8019. PMLR, 2020. 3", + "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Proc. Adv. Neural Inf. Process. Syst., 28:91-99, 2015. 3", + "[57] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, MingYu Liu, Yong Jae Lee, Alexander G. Schwing, and Jan Kautz. Instance-aware, context-focused, and memoryefficient weakly supervised object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2020. 1", + "[58] Andrei A. Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In Proc. Int. Conf. Learn. Represent., 2019. 2", + "[59] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In Proc. Int. Conf. Mach. Learn., volume 48, pages 1842-1850, 2016. 2", + "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Proc. Adv. Neural Inf. Process. Syst., volume 30, 2017. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7348", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 7352-7362, 2021. 1, 2, 3, 4, 5, 7", + "[62] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 1199-1208, 2018. 2", + "[63] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B. Tenenbaum, and Phillip Isola. Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need? In Proc. European Conf. on Computer Vision, 2020. 1, 2", + "[64] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proc. Adv. Neural Inf. Process. Syst., volume 29, 2016. 2", + "[65] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. arXiv preprint arXiv:2003.06957, 2020. 1, 2, 3, 5, 7, 8", + "[66] Xiaobo Wang, Shuo Wang, Cheng Chi, Shifeng Zhang, and Tao Mei. Loss function search for face recognition. In Proc. Int. Conf. Mach. Learn., pages 10029-10038. PMLR, 2020. 3, 4", + "[67] Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning. arXiv preprint arXiv:1911.04623, 2019. 2", + "[68] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-Shot Learning from Imaginary Data. arXiv:1801.05401 [cs], Jan. 2018. 2", + "[69] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Meta-learning to detect rare objects. In Proc. IEEE Int. Conf. on Computer Vision, pages 9925-9934, 2019. 5", + "[70] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 4", + "[71] Among Wu, Suqi Zhao, Cheng Deng, and Wei Liu. Generalized and discriminative few-shot object detection via svddictionary enhancement. Proc. Adv. Neural Inf. Process. Syst., 34:6353-6364, 2021. 3", + "[72] Jiaxi Wu, Songtao Liu, Di Huang, and Yunhong Wang. Multiscale positive sample refinement for few-shot object detection. In Proc. European Conf. on Computer Vision, pages 456-472. Springer, 2020. 1, 2, 3, 4, 5, 6, 7", + "[73] Xiongwei Wu, Doyen Sahoo, and Steven Hoi. Meta-rcnn: Meta learning for few-shot object detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1679–1687, 2020. 1", + "[74] Yongqin Xian, Bernt Schiele, and Zeynep Akata. Zero-shot learning-the good, the bad and the ugly. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4582-4591, 2017. 5", + "[75] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In Proc. European Conf. on Computer Vision, pages 192-210. Springer, 2020. 1, 2, 3, 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proc. IEEE Int. Conf. on Computer Vision, pages 9577-9586, 2019. 1, 2, 5, 7", + "[77] Huaxiu Yao, Linjun Zhang, and Chelsea Finn. Meta-learning with fewer tasks through task interpolation. In Proceeding of the 10th International Conference on Learning Representations, 2022. 2", + "[78] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 8805-8814, 2020. 2, 4", + "[79] Li Yin, Juan M Perez-Rua, and Kevin J Liang. Sylph: A hypernetwork framework for incremental few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 9035-9045, 2022. 1, 2, 3", + "[80] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., June 2020. 2", + "[81] Gongjie Zhang, Zhipeng Luo, Kaiwen Cui, Shijian Lu, and Eric P Xing. Meta-detr: Image-level few-shot detection with inter-class correlation exploitation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3", + "[82] Lu Zhang, Shuigeng Zhou, Jihong Guan, and Ji Zhang. Accurate few-shot object detection with support-query mutual guidance and hybrid loss. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 14424-14432, 2021. 3", + "[83] Shan Zhang, Lei Wang, Naila Murray, and Piotr Koniusz. Kernelized few-shot object detection with efficient integral aggregation. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 19207-19216, 2022. 1, 2, 5", + "[84] Weilin Zhang and Yu-Xiong Wang. Hallucination improves few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 13008-13017, 2021. 1, 2, 5, 7, 8" + ], + "bbox": [ + 501, + 92, + 893, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7349", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_model.json b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f39956cacaf9d451b48671c004c2e7dfbda59222 --- /dev/null +++ b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.13, + 0.856, + 0.175 + ], + "angle": 0, + "content": "Meta-tuning Loss Functions and Data Augmentation for Few-shot Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.203, + 0.773, + 0.24 + ], + "angle": 0, + "content": "Berkan Demirel1,2 Orhun Bugra Baran1 Ramazan Gokberk Cinbis1 \n1Middle East Technical University 2HAVELSAN Inc." + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.243, + 0.816, + 0.257 + ], + "angle": 0, + "content": "bdemirel@havelsan.com.tr bugra@ceng.metu.edu.tr gcinbis@ceng.metu.edu.tr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Few-shot object detection, the problem of modelling novel object detection categories with few training instances, is an emerging topic in the area of few-shot learning and object detection. Contemporary techniques can be divided into two groups: fine-tuning based and meta-learning based approaches. While meta-learning approaches aim to learn dedicated meta-models for mapping samples to novel class models, fine-tuning approaches tackle few-shot detection in a simpler manner, by adapting the detection model to novel classes through gradient based optimization. Despite their simplicity, fine-tuning based approaches typically yield competitive detection results. Based on this observation, we focus on the role of loss functions and augmentations as the force driving the fine-tuning process, and propose to tune their dynamics through meta-learning principles. The proposed training scheme, therefore, allows learning inductive biases that can boost few-shot detection, while keeping the advantages of fine-tuning based approaches. In addition, the proposed approach yields interpretable loss functions, as opposed to highly parametric and complex few-shot meta-models. The experimental results highlight the merits of the proposed scheme, with significant improvements over the strong fine-tuning based few-shot detection baselines on benchmark Pascal VOC and MS-COCO datasets, in terms of both standard and generalized few-shot performance metrics." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.753, + 0.21, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Object detection is one of the computer vision problems that has greatly benefited from the advances in supervised deep learning approaches. However, similar to the case in many other problems, state-of-the-art in object detection relies on the availability of large-scale fully-annotated datasets, which is particularly problematic due to the difficulty of collecting accurate bounding box annotations [18,46]. This practical burden has lead to a great interest in the approaches" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.273, + 0.891, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.418, + 0.884, + 0.433 + ], + "angle": 0, + "content": "Figure 1. The overall architecture of the meta-tuning approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.61 + ], + "angle": 0, + "content": "that can potentially reduce the annotation cost, such as weakly-supervised learning [29, 57], learning from point annotations [7], and mixed supervised learning [45]. A more recently emerging paradigm in this direction is few-shot object detection (FSOD). In the FSOD problem, the goal is to build detection models for the novel classes with few labeled training images by transferring knowledge from the base classes with a large set of training images. In the closely related Generalized-FSOD (G-FSOD) problem, the goal is to build few-shot detection models that perform well on both base and novel classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.896, + 0.869 + ], + "angle": 0, + "content": "FSOD methods can be categorized into meta-learning and fine-tuning approaches. Although meta-learning based methods are predominantly used in the literature in FSOD research [8,22,31,36,52,75,76,79,81,83], several fine-tuning based works have recently reported competitive results [6, 15,32,53,61,65,72,84]. The main premise of meta-learning approaches is to design and train dedicated meta-models that map given few train samples to novel class detection models, e.g. [73] or learn easy-to-adapt models [30] in a MAML [16] fashion. In contrast, however, fine-tuning based methods tackle the problem as a typical transfer learning problem and apply the general purpose supervised training techniques, i.e. regularized loss minimization via gradient-based optimization, to adapt a pre-trained model to few-shot classes. It is also worth noting that the recent results on fine-tuning based FSOD are aligned with related observations on few-shot classification [9, 12, 63] and segmentation [4]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "While some of the FSOD meta-learning approaches are attractive for being able to learn dedicated parametric training" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7339" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.394 + ], + "angle": 0, + "content": "mechanisms, they also come with two important shortcomings: (i) the risk of overfitting to the base classes used for training the meta-model due to model complexity, and (ii) the difficulty of interpreting what is actually learned; both of which can be crucially important for real-world, in-the-wild utilization of a meta-learned model. From this point of view, the simplicity and generality of a fine-tuning based FSOD approach can be seen as major advantages. In fact, one can find a large machine learning literature on the components (optimization techniques, loss functions, data augmentation, and architectures) of an FT approach, as opposed to the unique and typically unknown nature of a meta-learning inference model, especially when the model aims to replace standard training procedures for modeling the novel few-shot classes. While MAML [16] like meta-learning for quick adaptation is closer in nature to fine-tuning based approaches, the vanishing gradient problems and the overall complexity of the meta-learning task practically limits the approach to target only one or few model update steps, whereas an FT approach has no such computational difficulty." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.397, + 0.473, + 0.578 + ], + "angle": 0, + "content": "Perhaps the biggest advantage of a fine-tuning based FSOD approach, however, can also be its biggest disadvantage: its generality may lack the inductive biases needed for effective learning with few novel class samples while preserving the knowledge of base classes. To this end, such approaches focus on the design of fine-tuning details, e.g. whether to freeze the representation parameters [65], use contrastive fine-tuning losses [61], increase the novel class variances [84], introduce the using additional detection heads and branches [15, 72]. However, optimizing such details specifically for few-shot classes in a hand-crafted manner is clearly difficult, and likely to be sub-optimal." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.581, + 0.473, + 0.868 + ], + "angle": 0, + "content": "To address this problem, we focus on applying meta-learning principles to tune the loss functions and augmentations to be used in the fine-tuning stage for FSOD, which we call meta-tuning (Figure 1). More specifically, much like the meta-learning of a meta-model, we define an episodic training procedure that aims to progressively discover the optimal loss function and augmentation details for FSOD purposes in a data-driven manner. Using reinforcement learning (RL) techniques, we aim to tune the loss function and augmentation details such that they maximize the expected detection quality of an FSOD model obtained by fine-tuning to a set of novel classes. By defining meta-tuning over well-designed loss terms and an augmentation list, we restrict the search process to effective function families, reducing the computational costs compared to AutoML methods that aim to discover loss terms from scratch for fully-supervised learning [20, 42]. The resulting meta-tuned loss functions and augmentations, therefore, inject the learned FSOD-specific inductive biases into a fine-tuning based approach." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.473, + 0.901 + ], + "angle": 0, + "content": "To explore the potential of the meta-tuning scheme for FSOD, we focus on the details of classification loss func" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.302 + ], + "angle": 0, + "content": "tions, based on the observations that FSOD prediction mistakes tend to be in classification rather than localization details [61]. In particular, we first focus on the softmax temperature parameter, for which we define two versions: (i) a simple constant temperature, and (ii) time (fine-tuning iteration index) varying dynamic temperature, parameterized as an exponentiated polynomial. In all cases, the parameters learned via meta-tuning yield an interpretable loss function that has a negligible risk of over-fitting to the base classes, in contrast to a complex meta-model. We also model augmentation magnitudes during meta-tuning for improving the data loading pipeline for few-shot learning purposes. Additionally, we incorporate a score scaling coefficient for learning to balance base versus novel class scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.897, + 0.426 + ], + "angle": 0, + "content": "We provide an experimental analysis on the Pascal VOC [13] and MS-COCO [40] benchmarks for FSOD, using the state-of-the-art fine-tuning based baselines MPSR [72] and DeFRCN [53]. Our experimental results show that the proposed meta-tuning approach provides significant performance gains in both FSOD and Generalized FSOD settings, suggesting that meta-tuning loss functions and data augmentation can be a promising direction in FSOD research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.642, + 0.456 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.895, + 0.512 + ], + "angle": 0, + "content": "This section provides an overview of recent developments on few-shot image classification, few-shot object detection, automated loss function and data augmentation discovery." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.897, + 0.668 + ], + "angle": 0, + "content": "Few-shot classification. Most of the meta-learning approaches for few-shot learning (FSL) of classification models can be grouped as adaptation-based and mapping-based approaches. Adaptation-based (also called gradient-based) approaches aim to learn model parameters that can easily be adapted to new unseen few-shot tasks within a few model update steps, e.g. [17, 38, 47, 48, 51, 54, 58]. Mapping-based approaches (also called metric-based) aim to bypass a gradient-descent based adaptation step, and instead learn a data-to-classifier mapping, e.g. [5, 44, 49, 59, 60, 62, 64, 77, 78, 80]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.897, + 0.79 + ], + "angle": 0, + "content": "Some of the other notable approaches include learning to generate synthetic data for novel classes [23, 33, 68], using better feature representations [1, 2, 19, 28, 41, 63, 67] or utilizing differentiable convex solvers [3, 34]. Importantly, several works highlight that a carefully trained representation combined with simple fine-tuning or even just shallow classifiers can yield competitive or better performance than meta-learning based approaches, e.g. [9, 12, 63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Few-shot object detection. The FSOD approaches can be summarized as meta-learning and fine-tuning (also called transfer-learning) based ones. Most meta-learning based FSOD approaches embrace formulations similar to those used in mapping-based meta-learning approaches for FSL, e.g. [8, 22, 31, 36, 52, 75, 76, 79, 81, 83]. Support feature aggregation is one of the main aspects where meta-learning-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7340" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "based methods differ from each other. Xiao and Marlet [75] use both the differences and the channel-wise multiplication of the features in addition to the combination of the features directly for support-query aggregation. Fan et al. [14] use attention blocks to make support and query features more distinguishable for base and novel object classes. Zhang et al. [81] use inter-class correlations to highlight important support features. Li et al. [36] propose to use specialized support and query features for classification and localization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.473, + 0.442 + ], + "angle": 0, + "content": "Recent efforts towards improving meta-learning based FSOD include complimentary techniques, mainly to improve loss functions, feature matching, and novel class sample usage efficiency. [36] uses class margin loss, [26] uses margin-based ranking loss, [82] uses hybrid loss which consist of focal loss, adaptive margin loss and contrastive loss. Hu et al. [27] perform feature matching between query and support images to use the information from the support images more effectively. Similarly, Han et al. [21] construct a matching network between query and support instances using heterogeneous graph convolutional networks. Li and Li [35] augment novel class samples via adding Gaussian noise. Yin et al. [79] decouple classification task from localization by using the proposed class-conditional architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.838 + ], + "angle": 0, + "content": "Fine-tuning-based methods typically freeze parts of a pre-trained detection network, add auxiliary detection heads, increase the novel class variances and then apply gradient descent based model update steps, unlike meta-learning-based methods that use complex episodic learning [15, 32, 53, 61, 65, 71, 72]. Wang et al. [65] propose a Faster-RCNN [56] based approach, where the class-agnostic region proposal network (RPN) component is kept frozen during fine-tuning. Sun et al. [61] use a similar approach and differently include FPN and RPN layers to the learnable parameter set in the same architecture. These learnable layers allow using contrastive proposal encodings that facilitate the more accurate classification of novel objects. Wu et al. [72] show that the scale distribution of support set tends to be imbalanced, and proposes a multi-scale positive sample refinement (MPSR) branch as an addition to the main model. Fan et al. [15] propose Retentive R-CNN architecture to prevent forgetting during fine-tuning for base classes. The obtained object proposals are fed into two ROI detectors responsible for base class and novel class instances. Qiao et al. [53] focus on decoupling network modules, and introduce a gradient decoupling layer and prototypical calibration block. Kaul et al. [32] extend the novel class annotations in the training set. In this context, the proposed method obtains object candidates from the base detector, and applies the box refinement step." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "While our approach is based on fine-tuning based FSOD, we embrace meta-learn principles to optimize the loss function and augmentations to improve the fine-tuning process for FSOD, without learning a complex and over-fitting-prone" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.895, + 0.122 + ], + "angle": 0, + "content": "meta-model. The resulting loss function and data augmentations are then utilized within the fine-tuning steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.895, + 0.46 + ], + "angle": 0, + "content": "Automated loss function discovery. Loss function discovery is an emerging AutoML topic towards improving the learning systems in a data-driven manner. Existing methods are mainly based on either (i) constructing the loss function directly from the basic operators [20, 42, 55] or (ii) optimizing parameterized loss functions [37, 66]. For loss construction, [42] proposes a genetic algorithm that consists of loss function verification and quality filtering modules. In this approach, the predefined proxy task eliminates divergent and poor candidate loss functions and survives the promising loss functions for other steps. [20] uses a genetic algorithm to select candidate loss functions from a tree of simple mathematical operations, and the successful loss functions pass to other stages to mutate. [55] suggests a method to learn not only the loss function but also the whole machine learning algorithm from scratch. For loss optimization, [37] re-analyzes the existing loss functions and presents them in a combined formula. [66] observes that the search space used in [37] can be too complex, and propose to simplify the search space via heuristics. In contrast to these works targeting supervised training scenarios, we aim to adapt loss function learning principles to the FSOD problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.897, + 0.675 + ], + "angle": 0, + "content": "AutoML for data augmentation. A variety of automated data augmentation techniques have recently been proposed [10, 11, 25, 39]. Cubuk et al. [10] generate augmentation policies using reinforcement learning and a controller RNN. Ho et al. [25] propose a method that reduces the computational costs compared to [10] by using a population-based framework. Similarly, Lim et al. [39] propose a direct Bayesian method to reduce costs. Cubuk et al. [11] show that the optimal augmentation magnitudes tend to be similar across transformations, and the search process can greatly be simplified by using a shared value. We follow this suggestion and use a shared magnitude across the transforms in our formulation. In contrast to these works on supervised learning, however, we focus on learning detectors with few-samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.895, + 0.767 + ], + "angle": 0, + "content": "In summary, while loss function and augmentation discovery topics increasingly attract attention towards improving supervised training pipelines, ours is the first work on learning few-sample specific inductive biases for fine-tuning based few-shot object detection based on meta-learning and AutoML principles, to the best of our knowledge." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.78, + 0.591, + 0.795 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.805, + 0.895, + 0.851 + ], + "angle": 0, + "content": "This section provides a brief summary of the FSOD problem definition and the baseline model we utilize. We then present our definition and instantiation of meta-tuning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Problem definition. We follow the FSOD setup of [31], where a relatively large set of training images for the set \\( C_b \\) of base classes is made available. Each training im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7341" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "age corresponds to a tuple \\((x,y)\\) consisting of image \\(x\\) and annotations \\(y = \\{y_0,\\dots,y_M\\}\\). Each object annotation \\(y_{i} = \\{c_{i},b_{i}\\}\\) contains a category label \\((c_{i})\\) and a bounding box \\((b_{i} = \\{x_{i},y_{i},w_{i},h_{i}\\})\\). Once the FSOD model training is complete, the evaluation is carried out based on a limited number \\((k)\\) of training images made available for the set \\(C_n\\) of distinct novel (i.e. few-shot) classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.473, + 0.46 + ], + "angle": 0, + "content": "Base model. We use the MPSR FSOD method [72] as the infrastructure for our loss function and data augmentation search methods. MPSR adapts the Faster-RCNN to be suitable for fine-tuning-based FSOD and uses an auxiliary multi-scale positive sample refinement (MPSR) branch to handle the scale scarcity problems. This branch expands the scale space of positive samples without increasing improper negative instances, unlike feature pyramid networks and image pyramids that do not change data distribution, hence the scale sparsity problem. In this context, objects in the images are cropped and resized in multiple sizes to create scale pyramids. The MPSR uses two groups of loss functions for the region proposal network (RPN) and detection heads, and feeds differently scaled positive samples to these loss functions together with the main detection branch. Finally, we note that the proposed approach can in principle be applied to virtually any fine-tuning based FSOD model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.318, + 0.484 + ], + "angle": 0, + "content": "3.1. Meta-tuning loss functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Our main goal is to improve few-shot detector fine-tuning based on meta-learning principles. For meta-tuning the FSOD loss, we specifically focus on the classification loss term, as the FSOD errors tend to be primarily caused by misclassifications [61]. The MPSR classification loss term can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.591, + 0.47, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c l s} (x, y) = - \\frac {1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {f \\left(x _ {i} , y _ {i}\\right)}}{\\sum_ {y} e ^ {f \\left(x _ {i} , y\\right)}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.748 + ], + "angle": 0, + "content": "where \\(N_{ROI}\\) is the number of ROIs (i.e. candidate regions) in an image, \\(y_{i}\\) is the groundtruth class label for the \\(i\\)-th ROI, and \\(f(x_{i},y)\\) is the corresponding class \\(y\\) prediction score. To add more flexibility into the loss function, we re-define it as a parametric function \\(\\ell_{cls}(x,y;\\rho)\\), where \\(\\rho\\) represents the loss function parameters. First, we introduce a temperature scalar \\(\\rho_{\\tau}\\), i.e. \\(\\rho = (\\rho_{\\tau})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.756, + 0.47, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c l s} (x, y; \\rho) = - \\frac {1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {f \\left(x _ {i} , y _ {i}\\right) / \\rho_ {\\tau}}}{\\sum_ {y ^ {\\prime}} e ^ {f \\left(x _ {i} , y ^ {\\prime}\\right) / \\rho_ {\\tau}}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our motivation comes from the observations on the importance of temperature scaling in log loss on various other problems, such as knowledge distillation [24], few-shot classification [49, 78], and zero-shot learning [43]. While temperature is typically tuned in a manual manner, here we aim to meta-learn it specifically for fine-tuning based FSOD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.153 + ], + "angle": 0, + "content": "purposes, giving a chance to observe the behavior of metatuning in a simple case. We also define a more sophisticated variant of the loss function by defining the dynamic temperature function \\( f_{\\rho} \\) and novel class scaling \\( \\alpha \\):" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.163, + 0.892, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c l s} (x, y; \\rho) = \\frac {- 1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {\\alpha \\left(y _ {i}\\right) f \\left(x _ {i} , y _ {i}\\right) / f _ {\\rho} (t)}}{\\sum_ {y ^ {\\prime}} e ^ {\\alpha \\left(y ^ {\\prime}\\right) f \\left(x _ {i} , y ^ {\\prime}\\right) / f _ {\\rho} (t)}}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.894, + 0.324 + ], + "angle": 0, + "content": "where \\( f_{\\rho}(t) = \\exp (\\rho_{a}t^{2} + \\rho_{b}t + \\rho_{c}) \\). Here, \\( \\rho = (\\rho_{a},\\rho_{b},\\rho_{c}) \\) is a 3-tuple of polynomial coefficients, and \\( t\\in [0,1] \\) is the normalized fine-tuning iteration index. The temperature can increase or decrease over time, making the predicted class distributions smoother or sharper. \\( \\alpha (y) \\) is set to 1 for \\( y\\in C_b \\), and otherwise the novel class score scaling coefficient \\( \\rho_{\\alpha} \\), as a way to learn base and novel score balancing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.75, + 0.348 + ], + "angle": 0, + "content": "3.2. Meta-tuning augmentations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.894, + 0.447 + ], + "angle": 0, + "content": "For meta-tuning augmentations, we focus on the photometric augmentations that are likely to be transferable from base to novel classes. In this context, we model the brightness, saturation, contrast, and hue transforms, with a shared magnitude parameter \\((\\rho_{aug})\\), which is known to be effective for supervised training [11]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.715, + 0.472 + ], + "angle": 0, + "content": "3.3. Meta-tuning procedure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.894, + 0.554 + ], + "angle": 0, + "content": "In our work, we utilize a REINFORCE [70] based reinforcement learning (RL) approach to search for the optimal loss function and augmentations, where we use the AutoML approach of Wang et al. [66] on loss function search for fully-supervised face recognition as our starting point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.895, + 0.765 + ], + "angle": 0, + "content": "In order to meta-tune the loss function and augmentations to maximize FSOD generalization abilities, we generate proxy tasks over base class training data to imitate real FSOD tasks over the novel classes. For this purpose, we divide base classes into two subsets, proxy-base \\( C_{\\mathrm{p - base}} \\) and proxy-novel \\( C_{\\mathrm{p - novel}} \\). We then construct three non-overlapping data set splits using the base class training set: (i) \\( D_{\\mathrm{p - pretrain}} \\) containing \\( C_{\\mathrm{p - base}} \\)-only samples, used for training a temporary object detection model for meta-tuning purposes; (ii) \\( D_{\\mathrm{p - support}} \\) containing samples of \\( C_{\\mathrm{p - base}} \\cup C_{\\mathrm{p - novel}} \\) classes to be used as fine-tuning images during meta-tuning; (iii) \\( D_{\\mathrm{p - query}} \\) containing samples of \\( C_{\\mathrm{p - base}} \\cup C_{\\mathrm{p - novel}} \\) classes to be used for evaluating the generalized FSOD performance of a fine-tuned model during meta-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We generate a series of FSOD proxy tasks for metatuning, similar to episodic meta-learning: at each proxy task \\( T \\), we sample a few-shot training set from \\( D_{\\mathrm{p - support}} \\). We also sample a loss function/augmentation magnitude parameter combination \\( \\rho \\), where each \\( \\rho_j \\in \\rho \\) is modeled in terms of a Gaussian distribution: \\( \\rho_j \\sim \\mathcal{N}(\\mu_j, \\sigma^2) \\). Using the loss function or augmentations corresponding to the sampled \\( \\rho \\), we fine-tune the initial model on the support images using gradient-based optimization, and compute the mean" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7342" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.109, + 0.875, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Figure 2. The meta-tuning approach. At each RL iteration over a proxy task, the distribution parameters modeling the loss function and augmentations are updated as a function of the obtained mAP scores, towards improved training with few-samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.473, + 0.435 + ], + "angle": 0, + "content": "average precision (mAP) scores on \\( D_{\\mathrm{p - query}} \\). We get multiple mAP scores by repeating this process multiple times over multiple proxy support samples. Meta-tuning is then carried over by updating \\( \\mu \\) values via the REINFORCE rule after each episode, towards finding \\( \\mu \\) values centered around well-performing \\( \\rho \\) combinations." + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.446, + 0.47, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mu_ {j} ^ {\\prime} \\leftarrow \\mu_ {j} + \\eta R (\\rho) \\nabla_ {\\mu} \\log (p \\left(\\rho_ {j}; \\mu_ {j}, \\sigma\\right)) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.471, + 0.504 + ], + "angle": 0, + "content": "where \\(p(\\rho ;\\mu ,\\sigma)\\) is the Gaussian probability density function, \\(\\eta\\) is the RL learning rate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.472, + 0.653 + ], + "angle": 0, + "content": "We apply the REINFORCE update rule using the \\(\\rho\\) with the highest reward per episode. \\(R(\\rho)\\) is the normalized reward function obtained by whitening the mAP scores. We empirically observe that normalization improves the results (Section 4) since without reward normalization, the RL updates are scaled with respect to the inherent difficulty of the proxy task, which greatly varies depending on the sampled support examples. Reward normalization approximately removes the average reward, enabling better performing \\(\\rho\\) samples to influence based on their relative success." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.472, + 0.716 + ], + "angle": 0, + "content": "Finally, similar to [50], starting with \\(\\sigma = 0.1\\), we diminish \\(\\sigma\\) over the RL iterations to progressively reduce explorations by sampling more conservatively, which improves converge. The final scheme is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.729, + 0.21, + 0.745 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.472, + 0.866 + ], + "angle": 0, + "content": "Metrics. We use mAP to evaluate the base and novel class detection results separately. To evaluate the generalized FSOD performance, we use the Harmonic Mean (HM) metric to compute a balanced aggregation of base and novel class performance scores. Adapted from generalized zero-shot learning [74], HM is defined as the harmonic mean of \\(\\mathrm{mAP}_{\\mathrm{base}}\\) and \\(\\mathrm{mAP}_{\\mathrm{novel}}\\) scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Datasets. We use Pascal VOC [13] and MS COCO [40] with the same splits defined in FSOD benchmarks [65, 72]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.895, + 0.435 + ], + "angle": 0, + "content": "On Pascal VOC, three separate base/novel class splits exist, where each one consists of 15 base and 5 novel classes. In each split, we select 5 base classes to mimic novel classes during meta-tuning. On MS-COCO, we select 15 base classes to mimic novel classes in each proxy task, and evaluate the models for the 10-shot and 30-shot settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.895, + 0.681 + ], + "angle": 0, + "content": "Baselines. We primarily use the MPSR [72] and DeFRCN [53] as our baselines, which are among the best performing fine-tuning based FSOD methods on Pascal VOC. For the DeFRCN experiments, we transfer the meta-tuned loss functions and augmentation magnitudes from MPSR to the DeFRCN method, which are both based on Faster-RCNN. We take the results for FRCN [76], Ret. R-CNN [15], Meta-RCNN [76], FSRW [31], MetaDet [69], FsDetView [75] and ONCE [52] from [15] for a fair comparison. For the MPSR, DeFRCN (seed is set to 0) and FSCE [61], we report the results we obtain experimentally. We take the results for TFA+Hal [84], CME [36], TIP [35], DCNet [27], QA-FewDet [21] FADI [6], LVC [32], KFSOD [83] and FCT [22] from the original papers. Finally, while it is difficult to fairly compare fine-tuning versus meta-learning based approaches, we provide a discussion in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.894, + 0.838 + ], + "angle": 0, + "content": "Implementation details. We use 200 RL episodes for loss function meta-tuning, with REINFORCE learning rate set to 0.0005. The meta-tuning for augmentation parameter is carried out using the trained and frozen the loss function parameters. We keep the fine-tuning implementation details of MPSR unchanged, which uses 4000 and 8000 gradient descent iterations for 10-shot and 30-shot experiments on MS-COCO, and 2000 iterations on Pascal VOC. We will publish the full source code upon publication; a preliminary version is provided as supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.634, + 0.862 + ], + "angle": 0, + "content": "4.1. Main results" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We first compare the meta-tuning results against the corresponding MPSR baseline in Table 1. In the table, Meta" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7343" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.872, + 0.226 + ], + "angle": 0, + "content": "
Method/ShotPascal VOCMS-COCO
Novel ClassesAll Classes (HM)Novel ClassesAll Classes (HM)
12351012351010301030
MPSR [72]33.137.244.347.152.143.147.454.557.260.89.113.711.515.0
MPSR+Meta-Static33.439.445.147.352.643.750.455.457.561.410.114.812.716.4
MPSR+Meta-Dynamic34.539.845.048.252.545.051.055.558.361.611.914.914.316.6
MPSR+Meta-ScaledDynamic35.240.345.848.452.945.651.255.958.361.812.315.014.416.7
MPSR+Aug34.638.646.048.352.745.149.556.258.462.09.914.912.516.3
MPSR+Meta-Static+Aug35.339.146.148.452.745.949.956.258.361.810.215.212.816.7
MPSR+Meta-Dynamic+Aug35.439.646.548.953.346.050.556.858.962.512.115.314.516.8
MPSR+Meta-ScaledDynamic+Aug35.840.646.849.253.746.351.557.059.262.712.515.414.716.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.234, + 0.895, + 0.263 + ], + "angle": 0, + "content": "Table 1. FSOD (mAP) and G-FSOD (HM of the base and novel class mAPs) results on Pascal VOC and MS-COCO datasets for MPSR baseline method. HM stands for harmonic mean." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.475, + 0.546 + ], + "angle": 0, + "content": "Static, Meta-Dynamic, Meta-ScaledDynamic refer to meta-tuning a single temperature, dynamic temperature, and novel class scaled dynamic temperature functions, respectively. Similarly, Aug, Meta-Static \\(^+\\) Aug, Meta-Dynamic \\(^+\\) Aug, and Meta-ScaledDynamic \\(^+\\) Aug refer to meta-tuning only augmentation, single temperature and augmentation, dynamic temperature and augmentation, and novel class scaled dynamic temperature and augmentation functions, respectively. We observe that meta-tuning consistently improves the FSOD and G-FSOD results of the MPSR model. We also observe steady improvements gradually from the baseline to Meta-Static, to Meta-Dynamic, and finally to Meta-ScaledDynamic. In addition, the meta-tuned augmentation magnitude parameter also contributes positively to the few-shot object detection performance. The overall consistency of the improvements provides positive evidence for the value of loss and augmentation meta-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.472, + 0.748 + ], + "angle": 0, + "content": "Pascal VOC results. In Table 2, we report the Pascal VOC results for our MPSR and DeFRCN based Meta-ScaledDynamic+Aug approach and compare them against the state-of-the-art fine-tuning based FSOD methods. While we present the scores averaged over the three splits in this table, additional per-split FSOD and G-FSOD results can be found in the supplementary material. The left side of Table 2 presents the FSOD results for the varying number of support images. We observe that DeFRCN combined with Meta-ScaledDynamic+Aug, i.e. meta-tuning of the score coefficient, dynamic temperature and the augmentation parameter, yields the best mAP scores in all \\(k\\)-shot settings among all methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.472, + 0.854 + ], + "angle": 0, + "content": "The right side of Table 2 presents the G-FSOD results on Pascal VOC. We observe that the best-performing Meta-ScaledDynamic+Aug method improves the HM scores further above the state-of-the-art in all \\(k\\)-shot settings. Overall, these results suggest that the proposed framework is an effective way for meta-learning inductive biases to be used in fine-tuning-based FSOD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Figure 3 presents visual detection examples without and with meta-tuned scaled dynamic temperature and augmentations in the first and second rows, respectively. We observe" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.335 + ], + "angle": 0, + "content": "various improvements, such as reductions in false positives, improved recall, and more precise boxes, most likely due to the improved model fitting in the low-data regime." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.897, + 0.568 + ], + "angle": 0, + "content": "MS-COCO results. In Table 3, we compare the MPSR and DeFRCN based Meta-ScaleDynamic+Aug results against other fine-tuning based FSOD methods that report 10-shot and 30-shot results on the MS-COCO dataset. We observe that with meta-tuning, the FSOD scores of MPSR improve from 9.1 to 12.5 (10-shot mAP), and from 13.7 to 15.4 (30-shot mAP). We also observe that the scores of DeFRCN improve from 18.5 to 18.8 (10-shot mAP), and from 21.9 to 23.4 (30-shot mAP), obtaining the best and second best results against all other models. Similarly, in the case of GFSD, with meta-tuning, the 10-shot HM score of DeFRCN improves from 24.0 to 24.4, outperforming all other models. In addition, the 30-shot HM score of DeFRCN improves from 26.8 to 28.0, which is slightly below the 28.1 score of LVC-PL [32]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.582, + 0.662, + 0.598 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.896, + 0.824 + ], + "angle": 0, + "content": "Meta-tuning details. The proposed meta-tuning approach involves three important technical details: Proxy-novel imitation, model re-initialization, and reward normalization. Proxy-novel imitation refers to reinforcement learning over the sampled proxy-novel tasks, instead of the whole training set, to mimic the test-time FSOD challenges. Model re-initialization is the re-initialization of the base model for each task. Without re-initialization, not only the sampled loss/augmentation parameters and tasks but also the accumulated model updates undesirably affect the rewards. Reward normalization further reduces the effect of task difficulty variance by normalizing the rewards obtained within a single episode, allowing a more isolated assessment of the sampled loss functions and augmentations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.897, + 0.903 + ], + "angle": 0, + "content": "We evaluate the contributions of these three important details in terms of G-FSOD HM scores using the 5-shot setting of Pascal VOC Split-1 with MPSR+Meta-Dynamic. The results averaged over 5 runs are given in Table 4. We observe that each component progressively improves the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7344" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.189, + 0.089, + 0.784, + 0.264 + ], + "angle": 0, + "content": "
Method/ShotNovel ClassesAll Classes (HM)
123510123510
FRCN [76] (ICCV'19)16.120.628.833.436.525.931.740.044.346.7
TFA-fc [65] (ICML'20)27.630.639.846.648.740.544.152.958.359.9
TFA-cos [65] (ICML'20)31.432.640.546.848.344.646.053.558.459.6
FSCE [61] (CVPR'21)29.236.342.547.152.241.848.854.257.761.0
Ret. R-CNN [15] (CVPR'21)31.437.141.446.848.844.750.554.759.160.8
TFA+Hal [84] (CVPR'21)32.935.540.446.348.1-----
FADI [6] (NeurIPS'21)42.246.547.952.456.9-----
LVC [32] (CVPR'22)30.935.443.651.154.1-----
LVC-PL [32] (CVPR'22)45.245.054.857.558.6-----
MPSR [72] (ECCV'20)33.137.244.347.152.143.147.454.557.260.8
DeFRCN [53] (ICCV'21)46.552.655.960.060.857.662.564.767.667.8
MPSR+Meta-ScaledDynamic+Aug35.840.646.849.253.746.351.557.059.262.7
DeFRCN+Meta-ScaledDynamic+Aug49.254.057.261.361.859.863.765.968.668.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.274, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Table 2. FSOD (mAP) and G-FSOD (HM of the base and novel class mAPs) results on Pascal VOC. The best and the second-best results are marked with red and blue. HM stands for harmonic mean." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.317, + 0.218, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.409, + 0.219, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.503, + 0.187, + 0.514 + ], + "angle": 0, + "content": "Split-1/5-shot" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.317, + 0.351, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.409, + 0.351, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.503, + 0.322, + 0.514 + ], + "angle": 0, + "content": "Split-1/10-shot" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.317, + 0.483, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.41, + 0.483, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.503, + 0.45, + 0.514 + ], + "angle": 0, + "content": "Split-2/5-shot" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.317, + 0.615, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.41, + 0.615, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.503, + 0.585, + 0.514 + ], + "angle": 0, + "content": "Split-2/10-shot" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.317, + 0.747, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.41, + 0.747, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.503, + 0.714, + 0.514 + ], + "angle": 0, + "content": "Split-3/5-shot" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.317, + 0.881, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.41, + 0.881, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.503, + 0.849, + 0.514 + ], + "angle": 0, + "content": "Split-3/10-shot" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.526, + 0.893, + 0.555 + ], + "angle": 0, + "content": "Figure 3. Qualitative results using MPSR without (first row) and with (second row) meta-tuning, over multiple Pascal VOC splits. Base and novel class detections are shown with green and red boxes, respectively. (Best viewed in color.)" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.578, + 0.495, + 0.753 + ], + "angle": 0, + "content": "
Method/ShotsNovel ClassesAll Classes (HM)
10-shot30-shot10-shot30-shot
FRCN [76] (ICCV'19)9.212.512.815.6
FRCN-BCE [65] (ICML'20)6.410.310.916.1
TFA-fc [65] (ICML'20)10.013.415.419.4
TFA-cos [65] (ICML'20)10.013.715.619.8
MPSR [72] (ECCV'20)9.113.711.515.0
FSCE [61] (CVPR'21)10.514.416.020.2
Ret. R-CNN [15] (CVPR'21)10.513.816.620.4
FADI [6] (NeurIPS'21)12.216.1--
DeFRCN [53] (ICCV'21)18.521.924.026.8
LVC [32] (CVPR'22)12.117.817.822.8
LVC-PL [32] (CVPR'22)17.824.522.828.1
MPSR+Meta-ScaledDynamic+Aug12.515.414.716.9
DeFRCN+Meta-ScaledDynamic+Aug18.823.424.428.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.764, + 0.471, + 0.806 + ], + "angle": 0, + "content": "Table 3. Comparison of Meta-ScaledDynamic results to the fine-tuning based (G-)FSOD methods on the MS-COCO dataset. The best and the second-best results are marked with red and blue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "HM scores, and the most significant contribution is made by reward normalization, which improves from 62.1 to 63.3. We also observe that reward normalization considerably improves the overall experimental stability. To quantify this" + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.578, + 0.853, + 0.646 + ], + "angle": 0, + "content": "
Proxy-novel imit.Model re-init.Reward norm.HM
XXX61.5
✓XX61.8
✓✓X62.1
✓✓✓63.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.657, + 0.893, + 0.727 + ], + "angle": 0, + "content": "Table 4. Evaluation of meta-tuning details. Proxy-novel imitation is the imitation of novel classes using a subset of base classes. Model re-initialization is the re-initialization of the base model at each task. Reward normalization is within-episode normalization of the mAP scores during meta-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.894, + 0.85 + ], + "angle": 0, + "content": "observation, we estimate the \\(95\\%\\) confidence interval over the runs using \\(CI = 1.96\\frac{s}{\\sqrt{n}}\\), where \\(s\\), \\(n\\), and \\(1.96\\) are the standard deviation, number of runs, and \\(Z\\)-value, respectively [65]. According to this estimator, the normalization step narrows the confidence interval from \\(\\pm 0.75\\) to \\(\\pm 0.13\\), providing a clear improvement in reliability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Learned loss functions. In Figure 4, we plot the learned loss functions according to the \\(\\mu\\) values obtained at the end of the RL process. The upper plot shows the dynamic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7345" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.109, + 0.415, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.287, + 0.415, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.461, + 0.472, + 0.518 + ], + "angle": 0, + "content": "Figure 4. The dynamic temperature functions and score scaling coefficients learned by the meta-tuning process, using Meta-Dynamic (upper) and Meta-ScaledDynamic (lower) formulations. Results for each Pascal VOC split is shown with a separate curve." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.472, + 0.88 + ], + "angle": 0, + "content": "temperature functions learned over three different splits. We observe that temporally attenuated temperature values are preferred consistently, sharpening the predictions towards the end of the fine-tuning process. The lower plot shows the learned dynamic temperature functions with novel class score scaling. The learned scaling coefficients, i.e. \\(\\mu_{\\alpha}\\) of the learned \\(\\rho_{\\alpha}\\) distribution, are shown as horizontal lines. We observe that similar dynamic temperature functions are learned, and \\(\\mu_{\\alpha}\\) values vary between 1.09 to 1.2, suggesting that the meta-tuning process learns to boost the novel class scores. The interpretability of these outcomes, we believe, highlights a significant advantage of loss meta-tuning. In the context of interpretability, we observe that as the fine-tuning process continues on the few-shot training set, the predictions are progressively made sharper, i.e. the loss becomes more sensitive to classification errors and enforces towards making more confident correct predictions. This is in alignment with one of our original motivations for reducing the dominating classification errors in G-FSOD, as the meta-tuning process automatically learns to enforce more accurate classifications, where the curve steepness and the numerical ranges are learned via RL." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Learned augmentations. The learned photometric augmen" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.88, + 0.146 + ], + "angle": 0, + "content": "
S/MTFA [65]TFA+Hal [84]TFA+Meta-ScaledDynamic+Aug
13.43.84.7
24.65.05.8
36.66.97.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.157, + 0.894, + 0.184 + ], + "angle": 0, + "content": "Table 5. Low-shot (1-shot, 2-shot and 3-shot) experiments on MS-COCO dataset with novel classes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.217, + 0.895, + 0.323 + ], + "angle": 0, + "content": "tation magnitude values learned are 0.29, 0.24, 0.13, and 0.36 for Pascal VOC split-1, split-2, split-3, and MS-COCO datasets, respectively. We observe that the learned augmentation magnitudes positively contribute to the performance. According to the results in Table 1, the average Pascal VOC split-1/1-shot score increases from 33.1 to 34.6 with only augmentation steps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.33, + 0.897, + 0.481 + ], + "angle": 0, + "content": "Very low-shot experiments. Finally, we evaluate the meta-tuning approach in low-shot many-class settings. [84] proposes TFA+Hal method that uses the TFA baseline and conducts 1-shot, 2-shot, and 3-shot FSOD on the MS-COCO dataset. As we already observe the positive effects of the loss terms and augmentation magnitudes obtained from the MPSR on the DeFRCN, we similarly apply the learned parameters to the TFA baseline. The results are presented in Table 5. We observe that results are consistently improved using the meta-tuned functions on the TFA baseline." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.501, + 0.619, + 0.517 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.895, + 0.786 + ], + "angle": 0, + "content": "Fine-tuning based frameworks offer simple and reliable approaches to building detection models from few samples. However, a major limitation of the existing fine-tuning-based FSOD models is their focus on the hand-crafting the design of fine-tuning details for few-shot training, which is inherently difficult and likely to be sub-optimal. Towards addressing this limitation, we propose to meta-learn the fine-tuning based learning dynamics as a way of introducing learned inductive biases for few-shot learning. The proposed tuning scheme uses meta-learning principles with reinforcement learning, and obtains interpretable loss functions and augmentation magnitudes for few-shot training. Our comprehensive experimental results on Pascal VOC and MS COCO datasets show that the proposed meta-tuning approach consistently provides significant performance improvements over the strong fine-tuning based few-shot detection baselines in both FSOD and G-FSOD settings." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.788, + 0.895, + 0.849 + ], + "angle": 0, + "content": "While we restrict our experiments to loss and augmentation functions, meta-tuning other learning components, e.g. initial model, and applications to other few-shot learning problems can be interesting future work directions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the TUBITAK Grant 119E597 and a Google Faculty Research Award." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7346" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, and Frank Wood. Enhancing few-shot image classification with unlabelled examples. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2796-2805, January 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Sigal. Improved Few-Shot Visual Classification. arXiv e-prints, page arXiv:1912.03432, Dec. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Luca Bertinetto, Joao F. Henriques, Philip Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Proc. Int. Conf. Learn. Represent., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-Shot Segmentation Without Meta-Learning: A Good Transductive Inference Is All You Need? In arXiv:2012.06166 [cs], 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[5] Kaidi Cao, Maria Brbić, and Jure Leskovec. Concept learners for few-shot learning. In Proc. Int. Conf. Learn. Represent., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[6] Yuhang Cao, Jiaqi Wang, Ying Jin, Tong Wu, Kai Chen, Ziwei Liu, and Dahua Lin. Few-shot object detection via association and discrimination. Proc. Adv. Neural Inf. Process. Syst., 34:16570-16581, 2021. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[7] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 8819-8828, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.549 + ], + "angle": 0, + "content": "[8] Tung-I Chen, Yueh-Cheng Liu, Hung-Ting Su, Yu-Cheng Chang, Yu-Hsiang Lin, Jia-Fong Yeh, Wen-Chin Chen, and Winston Hsu. Dual-awareness attention for few-shot object detection. IEEE Transactions on Multimedia, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[9] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A Closer Look at Few-shot Classification. In ICLR 2019, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[10] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[11] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proc. IEEE Conf. Comput. Vis. Pattern Recog. Workshops, pages 702-703, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[12] Guneet S Dhillon, Pratik Chaudhari, Avinash Ravichandran, and Stefano Soatto. A Baseline for Few-Shot Image Classification. In ICLR, page 20, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[14] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4013–4022, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[15] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4527-4536, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[16] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[17] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proc. Int. Conf. Mach. Learn., volume 70, pages 1126–1135, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[18] Golnaz Ghiasi, Yin Cui, A. Srinivas, Rui Qian, Tsung-Yi Lin, Ekin Dogus Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 2917-2927, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.332 + ], + "angle": 0, + "content": "[19] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proc. IEEE Int. Conf. on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[20] Santiago Gonzalez and Risto Miikkulainen. Improved training speed, accuracy, and data utilization through loss function optimization. In 2020 IEEE Congress on Evolutionary Computation (CEC), pages 1-8. IEEE, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[21] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proc. IEEE Int. Conf. on Computer Vision, pages 3263-3272, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[22] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 5321-5330, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.894, + 0.572 + ], + "angle": 0, + "content": "[23] Bharath Hariharan and Ross B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. Proc. IEEE Int. Conf. on Computer Vision, pages 3037-3046, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[24] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.894, + 0.674 + ], + "angle": 0, + "content": "[25] Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. In Proc. Int. Conf. Mach. Learn., pages 2731-2741. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[26] Ting-I Hsieh, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. One-shot object detection with co-attention and co-excitation. arXiv preprint arXiv:1911.12529, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.894, + 0.773 + ], + "angle": 0, + "content": "[27] Hanzhe Hu, Shuai Bai, Aoxue Li, Jinshi Cui, and Liwei Wang. Dense relation distillation with context-aware aggregation for few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 10185-10194, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.894, + 0.842 + ], + "angle": 0, + "content": "[28] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference. arXiv e-prints, page arXiv:2204.07305, Apr. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[29] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Proc. Adv. Neural Inf. Process. Syst., 33, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7347" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[30] Taewon Jeong and Heeyoung Kim. Ood-maml: Meta-learning for few-shot out-of-distribution detection and classification. In Advances in Neural Information Processing Systems, volume 33, pages 3907-3916, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[31] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In Proc. IEEE Int. Conf. on Computer Vision, pages 8420-8429, 2019. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.472, + 0.263 + ], + "angle": 0, + "content": "[32] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Label, verify, correct: A simple few shot object detection method. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 14237-14247, 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.472, + 0.307 + ], + "angle": 0, + "content": "[33] Michalis Lazarou, Yannis Avrithis, and Tania Stathaki. Tensor feature hallucination for few-shot learning. ArXiv, abs/2106.05321, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.472, + 0.365 + ], + "angle": 0, + "content": "[34] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 10649-10657, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.472, + 0.409 + ], + "angle": 0, + "content": "[35] Aoxue Li and Zhenguo Li. Transformation invariant few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 3094-3102, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.472, + 0.467 + ], + "angle": 0, + "content": "[36] Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. Beyond max-margin: Class margin equilibrium for few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 7363-7372, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.525 + ], + "angle": 0, + "content": "[37] Chuming Li, Xin Yuan, Chen Lin, Minghao Guo, Wei Wu, Junjie Yan, and Wanli Ouyang. Am-lfs: Automl for loss function search. In Proc. IEEE Int. Conf. on Computer Vision, pages 8410-8419, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.472, + 0.569 + ], + "angle": 0, + "content": "[38] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. MetaSGD: Learning to Learn Quickly for Few-Shot Learning. arXiv e-prints, page arXiv:1707.09835, July 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.472, + 0.612 + ], + "angle": 0, + "content": "[39] Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment. Proc. Adv. Neural Inf. Process. Syst., 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.472, + 0.684 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conf. on Computer Vision, pages 740-755. Springer, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.472, + 0.741 + ], + "angle": 0, + "content": "[41] Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters: Understanding margin in few-shot classification. arXiv preprint arXiv:2003.12060, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.472, + 0.799 + ], + "angle": 0, + "content": "[42] Peidong Liu, Gengwei Zhang, Bochao Wang, Hang Xu, Xiaodan Liang, Yong Jiang, and Zhenguo Li. Loss function discovery for object detection via convergence-simulation driven search. arXiv preprint arXiv:2102.04700, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[43] Shichen Liu, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Generalized Zero-Shot Learning with Deep Calibration Network. In NeurIPS, pages 2005–2015. 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[44] Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sungju Hwang, and Yi Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In Proc. Int. Conf. Learn. Represent., 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[45] Yan Liu, Zhijie Zhang, Li Niu, Junjie Chen, and Liqing Zhang. Mixed supervised object detection by transferring mask prior and semantic similarity. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Proc. Adv. Neural Inf. Process. Syst., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. Proc. IEEE Int. Conf. on Computer Vision, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[47] Tsendsumen Munkhdalai, Xingdi Yuan, Soroush Mehri, and Adam Trischler. Rapid adaptation with conditionally shifted neurons. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[48] Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv: Learning, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.894, + 0.329 + ], + "angle": 0, + "content": "[49] Boris N. Oreshkin, Pau Rodriguez Lopez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. In NeurIPS, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[50] Matteo Papini, Andrea Battistello, and Marcello Restelli. Balancing learning speed and stability in policy gradient via adaptive exploration. In Proc. Int. Conf. on Artif. Intellig. and Stat., pages 1188-1199, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.387, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[51] Eunbyung Park and Junier B. Oliva. Meta-curvature. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[52] Juan-Manuel Perez-Rua, Xiatian Zhu, Timothy M Hospedales, and Tao Xiang. Incremental few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 13846-13855, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[53] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In Proc. IEEE Int. Conf. on Computer Vision, pages 8681-8690, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[54] Aravind Rajeswaran, Chelsea Finn, Sham M. Kakade, and Sergey Levine. Meta-learning with implicit gradients. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.894, + 0.621 + ], + "angle": 0, + "content": "[55] Esteban Real, Chen Liang, David So, and Quoc Le. Automlzero: Evolving machine learning algorithms from scratch. In Proc. Int. Conf. Mach. Learn., pages 8007-8019. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.894, + 0.677 + ], + "angle": 0, + "content": "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Proc. Adv. Neural Inf. Process. Syst., 28:91-99, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[57] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, MingYu Liu, Yong Jae Lee, Alexander G. Schwing, and Jan Kautz. Instance-aware, context-focused, and memoryefficient weakly supervised object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[58] Andrei A. Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In Proc. Int. Conf. Learn. Represent., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.805, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[59] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In Proc. Int. Conf. Mach. Learn., volume 48, pages 1842-1850, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Proc. Adv. Neural Inf. Process. Syst., volume 30, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7348" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.148 + ], + "angle": 0, + "content": "[61] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 7352-7362, 2021. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[62] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 1199-1208, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.472, + 0.263 + ], + "angle": 0, + "content": "[63] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B. Tenenbaum, and Phillip Isola. Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need? In Proc. European Conf. on Computer Vision, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.472, + 0.32 + ], + "angle": 0, + "content": "[64] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proc. Adv. Neural Inf. Process. Syst., volume 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.472, + 0.377 + ], + "angle": 0, + "content": "[65] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. arXiv preprint arXiv:2003.06957, 2020. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.381, + 0.472, + 0.435 + ], + "angle": 0, + "content": "[66] Xiaobo Wang, Shuo Wang, Cheng Chi, Shifeng Zhang, and Tao Mei. Loss function search for face recognition. In Proc. Int. Conf. Mach. Learn., pages 10029-10038. PMLR, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.472, + 0.493 + ], + "angle": 0, + "content": "[67] Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning. arXiv preprint arXiv:1911.04623, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.472, + 0.538 + ], + "angle": 0, + "content": "[68] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-Shot Learning from Imaginary Data. arXiv:1801.05401 [cs], Jan. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.472, + 0.582 + ], + "angle": 0, + "content": "[69] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Meta-learning to detect rare objects. In Proc. IEEE Int. Conf. on Computer Vision, pages 9925-9934, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.472, + 0.626 + ], + "angle": 0, + "content": "[70] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.472, + 0.683 + ], + "angle": 0, + "content": "[71] Among Wu, Suqi Zhao, Cheng Deng, and Wei Liu. Generalized and discriminative few-shot object detection via svddictionary enhancement. Proc. Adv. Neural Inf. Process. Syst., 34:6353-6364, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.472, + 0.741 + ], + "angle": 0, + "content": "[72] Jiaxi Wu, Songtao Liu, Di Huang, and Yunhong Wang. Multiscale positive sample refinement for few-shot object detection. In Proc. European Conf. on Computer Vision, pages 456-472. Springer, 2020. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.472, + 0.799 + ], + "angle": 0, + "content": "[73] Xiongwei Wu, Doyen Sahoo, and Steven Hoi. Meta-rcnn: Meta learning for few-shot object detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1679–1687, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[74] Yongqin Xian, Bernt Schiele, and Zeynep Akata. Zero-shot learning-the good, the bad and the ugly. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4582-4591, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[75] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In Proc. European Conf. on Computer Vision, pages 192-210. Springer, 2020. 1, 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[76] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proc. IEEE Int. Conf. on Computer Vision, pages 9577-9586, 2019. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[77] Huaxiu Yao, Linjun Zhang, and Chelsea Finn. Meta-learning with fewer tasks through task interpolation. In Proceeding of the 10th International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[78] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 8805-8814, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.894, + 0.318 + ], + "angle": 0, + "content": "[79] Li Yin, Juan M Perez-Rua, and Kevin J Liang. Sylph: A hypernetwork framework for incremental few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 9035-9045, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[80] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[81] Gongjie Zhang, Zhipeng Luo, Kaiwen Cui, Shijian Lu, and Eric P Xing. Meta-detr: Image-level few-shot detection with inter-class correlation exploitation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[82] Lu Zhang, Shuigeng Zhou, Jihong Guan, and Ji Zhang. Accurate few-shot object detection with support-query mutual guidance and hybrid loss. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 14424-14432, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.894, + 0.544 + ], + "angle": 0, + "content": "[83] Shan Zhang, Lei Wang, Naila Murray, and Piotr Koniusz. Kernelized few-shot object detection with efficient integral aggregation. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 19207-19216, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[84] Weilin Zhang and Yu-Xiong Wang. Hallucination improves few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 13008-13017, 2021. 1, 2, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7349" + } + ] +] \ No newline at end of file diff --git a/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_origin.pdf b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0c1582bcd08325367605bcdf33e98ddf799989b6 --- /dev/null +++ b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/985b2823-7f93-4fba-8afc-b7a84ab316d8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b22c2dfd5472270ccdb2ce0ca5b87085add5dcfc4cc3bc16c7bc94e66dc6dbf +size 5226749 diff --git a/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/full.md b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e42f6abda77108e7a578992fdb1cbbd5173af27 --- /dev/null +++ b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/full.md @@ -0,0 +1,326 @@ +# Meta-tuning Loss Functions and Data Augmentation for Few-shot Object Detection + +Berkan Demirel1,2 Orhun Bugra Baran1 Ramazan Gokberk Cinbis1 +1Middle East Technical University 2HAVELSAN Inc. + +bdemirel@havelsan.com.tr bugra@ceng.metu.edu.tr gcinbis@ceng.metu.edu.tr + +# Abstract + +Few-shot object detection, the problem of modelling novel object detection categories with few training instances, is an emerging topic in the area of few-shot learning and object detection. Contemporary techniques can be divided into two groups: fine-tuning based and meta-learning based approaches. While meta-learning approaches aim to learn dedicated meta-models for mapping samples to novel class models, fine-tuning approaches tackle few-shot detection in a simpler manner, by adapting the detection model to novel classes through gradient based optimization. Despite their simplicity, fine-tuning based approaches typically yield competitive detection results. Based on this observation, we focus on the role of loss functions and augmentations as the force driving the fine-tuning process, and propose to tune their dynamics through meta-learning principles. The proposed training scheme, therefore, allows learning inductive biases that can boost few-shot detection, while keeping the advantages of fine-tuning based approaches. In addition, the proposed approach yields interpretable loss functions, as opposed to highly parametric and complex few-shot meta-models. The experimental results highlight the merits of the proposed scheme, with significant improvements over the strong fine-tuning based few-shot detection baselines on benchmark Pascal VOC and MS-COCO datasets, in terms of both standard and generalized few-shot performance metrics. + +# 1. Introduction + +Object detection is one of the computer vision problems that has greatly benefited from the advances in supervised deep learning approaches. However, similar to the case in many other problems, state-of-the-art in object detection relies on the availability of large-scale fully-annotated datasets, which is particularly problematic due to the difficulty of collecting accurate bounding box annotations [18,46]. This practical burden has lead to a great interest in the approaches + +![](images/69104d70ca31758b63e72fb1d45a40fc0573a5278fa7cf9a8cbd4dfd31242d74.jpg) +Figure 1. The overall architecture of the meta-tuning approach. + +that can potentially reduce the annotation cost, such as weakly-supervised learning [29, 57], learning from point annotations [7], and mixed supervised learning [45]. A more recently emerging paradigm in this direction is few-shot object detection (FSOD). In the FSOD problem, the goal is to build detection models for the novel classes with few labeled training images by transferring knowledge from the base classes with a large set of training images. In the closely related Generalized-FSOD (G-FSOD) problem, the goal is to build few-shot detection models that perform well on both base and novel classes. + +FSOD methods can be categorized into meta-learning and fine-tuning approaches. Although meta-learning based methods are predominantly used in the literature in FSOD research [8,22,31,36,52,75,76,79,81,83], several fine-tuning based works have recently reported competitive results [6, 15,32,53,61,65,72,84]. The main premise of meta-learning approaches is to design and train dedicated meta-models that map given few train samples to novel class detection models, e.g. [73] or learn easy-to-adapt models [30] in a MAML [16] fashion. In contrast, however, fine-tuning based methods tackle the problem as a typical transfer learning problem and apply the general purpose supervised training techniques, i.e. regularized loss minimization via gradient-based optimization, to adapt a pre-trained model to few-shot classes. It is also worth noting that the recent results on fine-tuning based FSOD are aligned with related observations on few-shot classification [9, 12, 63] and segmentation [4]. + +While some of the FSOD meta-learning approaches are attractive for being able to learn dedicated parametric training + +mechanisms, they also come with two important shortcomings: (i) the risk of overfitting to the base classes used for training the meta-model due to model complexity, and (ii) the difficulty of interpreting what is actually learned; both of which can be crucially important for real-world, in-the-wild utilization of a meta-learned model. From this point of view, the simplicity and generality of a fine-tuning based FSOD approach can be seen as major advantages. In fact, one can find a large machine learning literature on the components (optimization techniques, loss functions, data augmentation, and architectures) of an FT approach, as opposed to the unique and typically unknown nature of a meta-learning inference model, especially when the model aims to replace standard training procedures for modeling the novel few-shot classes. While MAML [16] like meta-learning for quick adaptation is closer in nature to fine-tuning based approaches, the vanishing gradient problems and the overall complexity of the meta-learning task practically limits the approach to target only one or few model update steps, whereas an FT approach has no such computational difficulty. + +Perhaps the biggest advantage of a fine-tuning based FSOD approach, however, can also be its biggest disadvantage: its generality may lack the inductive biases needed for effective learning with few novel class samples while preserving the knowledge of base classes. To this end, such approaches focus on the design of fine-tuning details, e.g. whether to freeze the representation parameters [65], use contrastive fine-tuning losses [61], increase the novel class variances [84], introduce the using additional detection heads and branches [15, 72]. However, optimizing such details specifically for few-shot classes in a hand-crafted manner is clearly difficult, and likely to be sub-optimal. + +To address this problem, we focus on applying meta-learning principles to tune the loss functions and augmentations to be used in the fine-tuning stage for FSOD, which we call meta-tuning (Figure 1). More specifically, much like the meta-learning of a meta-model, we define an episodic training procedure that aims to progressively discover the optimal loss function and augmentation details for FSOD purposes in a data-driven manner. Using reinforcement learning (RL) techniques, we aim to tune the loss function and augmentation details such that they maximize the expected detection quality of an FSOD model obtained by fine-tuning to a set of novel classes. By defining meta-tuning over well-designed loss terms and an augmentation list, we restrict the search process to effective function families, reducing the computational costs compared to AutoML methods that aim to discover loss terms from scratch for fully-supervised learning [20, 42]. The resulting meta-tuned loss functions and augmentations, therefore, inject the learned FSOD-specific inductive biases into a fine-tuning based approach. + +To explore the potential of the meta-tuning scheme for FSOD, we focus on the details of classification loss func + +tions, based on the observations that FSOD prediction mistakes tend to be in classification rather than localization details [61]. In particular, we first focus on the softmax temperature parameter, for which we define two versions: (i) a simple constant temperature, and (ii) time (fine-tuning iteration index) varying dynamic temperature, parameterized as an exponentiated polynomial. In all cases, the parameters learned via meta-tuning yield an interpretable loss function that has a negligible risk of over-fitting to the base classes, in contrast to a complex meta-model. We also model augmentation magnitudes during meta-tuning for improving the data loading pipeline for few-shot learning purposes. Additionally, we incorporate a score scaling coefficient for learning to balance base versus novel class scores. + +We provide an experimental analysis on the Pascal VOC [13] and MS-COCO [40] benchmarks for FSOD, using the state-of-the-art fine-tuning based baselines MPSR [72] and DeFRCN [53]. Our experimental results show that the proposed meta-tuning approach provides significant performance gains in both FSOD and Generalized FSOD settings, suggesting that meta-tuning loss functions and data augmentation can be a promising direction in FSOD research. + +# 2. Related Work + +This section provides an overview of recent developments on few-shot image classification, few-shot object detection, automated loss function and data augmentation discovery. + +Few-shot classification. Most of the meta-learning approaches for few-shot learning (FSL) of classification models can be grouped as adaptation-based and mapping-based approaches. Adaptation-based (also called gradient-based) approaches aim to learn model parameters that can easily be adapted to new unseen few-shot tasks within a few model update steps, e.g. [17, 38, 47, 48, 51, 54, 58]. Mapping-based approaches (also called metric-based) aim to bypass a gradient-descent based adaptation step, and instead learn a data-to-classifier mapping, e.g. [5, 44, 49, 59, 60, 62, 64, 77, 78, 80]. + +Some of the other notable approaches include learning to generate synthetic data for novel classes [23, 33, 68], using better feature representations [1, 2, 19, 28, 41, 63, 67] or utilizing differentiable convex solvers [3, 34]. Importantly, several works highlight that a carefully trained representation combined with simple fine-tuning or even just shallow classifiers can yield competitive or better performance than meta-learning based approaches, e.g. [9, 12, 63]. + +Few-shot object detection. The FSOD approaches can be summarized as meta-learning and fine-tuning (also called transfer-learning) based ones. Most meta-learning based FSOD approaches embrace formulations similar to those used in mapping-based meta-learning approaches for FSL, e.g. [8, 22, 31, 36, 52, 75, 76, 79, 81, 83]. Support feature aggregation is one of the main aspects where meta-learning- + +based methods differ from each other. Xiao and Marlet [75] use both the differences and the channel-wise multiplication of the features in addition to the combination of the features directly for support-query aggregation. Fan et al. [14] use attention blocks to make support and query features more distinguishable for base and novel object classes. Zhang et al. [81] use inter-class correlations to highlight important support features. Li et al. [36] propose to use specialized support and query features for classification and localization. + +Recent efforts towards improving meta-learning based FSOD include complimentary techniques, mainly to improve loss functions, feature matching, and novel class sample usage efficiency. [36] uses class margin loss, [26] uses margin-based ranking loss, [82] uses hybrid loss which consist of focal loss, adaptive margin loss and contrastive loss. Hu et al. [27] perform feature matching between query and support images to use the information from the support images more effectively. Similarly, Han et al. [21] construct a matching network between query and support instances using heterogeneous graph convolutional networks. Li and Li [35] augment novel class samples via adding Gaussian noise. Yin et al. [79] decouple classification task from localization by using the proposed class-conditional architecture. + +Fine-tuning-based methods typically freeze parts of a pre-trained detection network, add auxiliary detection heads, increase the novel class variances and then apply gradient descent based model update steps, unlike meta-learning-based methods that use complex episodic learning [15, 32, 53, 61, 65, 71, 72]. Wang et al. [65] propose a Faster-RCNN [56] based approach, where the class-agnostic region proposal network (RPN) component is kept frozen during fine-tuning. Sun et al. [61] use a similar approach and differently include FPN and RPN layers to the learnable parameter set in the same architecture. These learnable layers allow using contrastive proposal encodings that facilitate the more accurate classification of novel objects. Wu et al. [72] show that the scale distribution of support set tends to be imbalanced, and proposes a multi-scale positive sample refinement (MPSR) branch as an addition to the main model. Fan et al. [15] propose Retentive R-CNN architecture to prevent forgetting during fine-tuning for base classes. The obtained object proposals are fed into two ROI detectors responsible for base class and novel class instances. Qiao et al. [53] focus on decoupling network modules, and introduce a gradient decoupling layer and prototypical calibration block. Kaul et al. [32] extend the novel class annotations in the training set. In this context, the proposed method obtains object candidates from the base detector, and applies the box refinement step. + +While our approach is based on fine-tuning based FSOD, we embrace meta-learn principles to optimize the loss function and augmentations to improve the fine-tuning process for FSOD, without learning a complex and over-fitting-prone + +meta-model. The resulting loss function and data augmentations are then utilized within the fine-tuning steps. + +Automated loss function discovery. Loss function discovery is an emerging AutoML topic towards improving the learning systems in a data-driven manner. Existing methods are mainly based on either (i) constructing the loss function directly from the basic operators [20, 42, 55] or (ii) optimizing parameterized loss functions [37, 66]. For loss construction, [42] proposes a genetic algorithm that consists of loss function verification and quality filtering modules. In this approach, the predefined proxy task eliminates divergent and poor candidate loss functions and survives the promising loss functions for other steps. [20] uses a genetic algorithm to select candidate loss functions from a tree of simple mathematical operations, and the successful loss functions pass to other stages to mutate. [55] suggests a method to learn not only the loss function but also the whole machine learning algorithm from scratch. For loss optimization, [37] re-analyzes the existing loss functions and presents them in a combined formula. [66] observes that the search space used in [37] can be too complex, and propose to simplify the search space via heuristics. In contrast to these works targeting supervised training scenarios, we aim to adapt loss function learning principles to the FSOD problem. + +AutoML for data augmentation. A variety of automated data augmentation techniques have recently been proposed [10, 11, 25, 39]. Cubuk et al. [10] generate augmentation policies using reinforcement learning and a controller RNN. Ho et al. [25] propose a method that reduces the computational costs compared to [10] by using a population-based framework. Similarly, Lim et al. [39] propose a direct Bayesian method to reduce costs. Cubuk et al. [11] show that the optimal augmentation magnitudes tend to be similar across transformations, and the search process can greatly be simplified by using a shared value. We follow this suggestion and use a shared magnitude across the transforms in our formulation. In contrast to these works on supervised learning, however, we focus on learning detectors with few-samples. + +In summary, while loss function and augmentation discovery topics increasingly attract attention towards improving supervised training pipelines, ours is the first work on learning few-sample specific inductive biases for fine-tuning based few-shot object detection based on meta-learning and AutoML principles, to the best of our knowledge. + +# 3. Method + +This section provides a brief summary of the FSOD problem definition and the baseline model we utilize. We then present our definition and instantiation of meta-tuning. + +Problem definition. We follow the FSOD setup of [31], where a relatively large set of training images for the set $C_b$ of base classes is made available. Each training im + +age corresponds to a tuple $(x,y)$ consisting of image $x$ and annotations $y = \{y_0,\dots,y_M\}$ . Each object annotation $y_{i} = \{c_{i},b_{i}\}$ contains a category label $(c_{i})$ and a bounding box $(b_{i} = \{x_{i},y_{i},w_{i},h_{i}\})$ . Once the FSOD model training is complete, the evaluation is carried out based on a limited number $(k)$ of training images made available for the set $C_n$ of distinct novel (i.e. few-shot) classes. + +Base model. We use the MPSR FSOD method [72] as the infrastructure for our loss function and data augmentation search methods. MPSR adapts the Faster-RCNN to be suitable for fine-tuning-based FSOD and uses an auxiliary multi-scale positive sample refinement (MPSR) branch to handle the scale scarcity problems. This branch expands the scale space of positive samples without increasing improper negative instances, unlike feature pyramid networks and image pyramids that do not change data distribution, hence the scale sparsity problem. In this context, objects in the images are cropped and resized in multiple sizes to create scale pyramids. The MPSR uses two groups of loss functions for the region proposal network (RPN) and detection heads, and feeds differently scaled positive samples to these loss functions together with the main detection branch. Finally, we note that the proposed approach can in principle be applied to virtually any fine-tuning based FSOD model. + +# 3.1. Meta-tuning loss functions + +Our main goal is to improve few-shot detector fine-tuning based on meta-learning principles. For meta-tuning the FSOD loss, we specifically focus on the classification loss term, as the FSOD errors tend to be primarily caused by misclassifications [61]. The MPSR classification loss term can be expressed as follows: + +$$ +\ell_ {c l s} (x, y) = - \frac {1}{N _ {R O I}} \sum_ {i} ^ {N _ {R O I}} \log \left(\frac {e ^ {f \left(x _ {i} , y _ {i}\right)}}{\sum_ {y} e ^ {f \left(x _ {i} , y\right)}}\right) \tag {1} +$$ + +where $N_{ROI}$ is the number of ROIs (i.e. candidate regions) in an image, $y_{i}$ is the groundtruth class label for the $i$ -th ROI, and $f(x_{i},y)$ is the corresponding class $y$ prediction score. To add more flexibility into the loss function, we re-define it as a parametric function $\ell_{cls}(x,y;\rho)$ , where $\rho$ represents the loss function parameters. First, we introduce a temperature scalar $\rho_{\tau}$ , i.e. $\rho = (\rho_{\tau})$ : + +$$ +\ell_ {c l s} (x, y; \rho) = - \frac {1}{N _ {R O I}} \sum_ {i} ^ {N _ {R O I}} \log \left(\frac {e ^ {f \left(x _ {i} , y _ {i}\right) / \rho_ {\tau}}}{\sum_ {y ^ {\prime}} e ^ {f \left(x _ {i} , y ^ {\prime}\right) / \rho_ {\tau}}}\right) \tag {2} +$$ + +Our motivation comes from the observations on the importance of temperature scaling in log loss on various other problems, such as knowledge distillation [24], few-shot classification [49, 78], and zero-shot learning [43]. While temperature is typically tuned in a manual manner, here we aim to meta-learn it specifically for fine-tuning based FSOD + +purposes, giving a chance to observe the behavior of metatuning in a simple case. We also define a more sophisticated variant of the loss function by defining the dynamic temperature function $f_{\rho}$ and novel class scaling $\alpha$ : + +$$ +\ell_ {c l s} (x, y; \rho) = \frac {- 1}{N _ {R O I}} \sum_ {i} ^ {N _ {R O I}} \log \left(\frac {e ^ {\alpha \left(y _ {i}\right) f \left(x _ {i} , y _ {i}\right) / f _ {\rho} (t)}}{\sum_ {y ^ {\prime}} e ^ {\alpha \left(y ^ {\prime}\right) f \left(x _ {i} , y ^ {\prime}\right) / f _ {\rho} (t)}}\right) \tag {3} +$$ + +where $f_{\rho}(t) = \exp (\rho_{a}t^{2} + \rho_{b}t + \rho_{c})$ . Here, $\rho = (\rho_{a},\rho_{b},\rho_{c})$ is a 3-tuple of polynomial coefficients, and $t\in [0,1]$ is the normalized fine-tuning iteration index. The temperature can increase or decrease over time, making the predicted class distributions smoother or sharper. $\alpha (y)$ is set to 1 for $y\in C_b$ , and otherwise the novel class score scaling coefficient $\rho_{\alpha}$ , as a way to learn base and novel score balancing. + +# 3.2. Meta-tuning augmentations + +For meta-tuning augmentations, we focus on the photometric augmentations that are likely to be transferable from base to novel classes. In this context, we model the brightness, saturation, contrast, and hue transforms, with a shared magnitude parameter $(\rho_{aug})$ , which is known to be effective for supervised training [11]. + +# 3.3. Meta-tuning procedure + +In our work, we utilize a REINFORCE [70] based reinforcement learning (RL) approach to search for the optimal loss function and augmentations, where we use the AutoML approach of Wang et al. [66] on loss function search for fully-supervised face recognition as our starting point. + +In order to meta-tune the loss function and augmentations to maximize FSOD generalization abilities, we generate proxy tasks over base class training data to imitate real FSOD tasks over the novel classes. For this purpose, we divide base classes into two subsets, proxy-base $C_{\mathrm{p - base}}$ and proxy-novel $C_{\mathrm{p - novel}}$ . We then construct three non-overlapping data set splits using the base class training set: (i) $D_{\mathrm{p - pretrain}}$ containing $C_{\mathrm{p - base}}$ -only samples, used for training a temporary object detection model for meta-tuning purposes; (ii) $D_{\mathrm{p - support}}$ containing samples of $C_{\mathrm{p - base}} \cup C_{\mathrm{p - novel}}$ classes to be used as fine-tuning images during meta-tuning; (iii) $D_{\mathrm{p - query}}$ containing samples of $C_{\mathrm{p - base}} \cup C_{\mathrm{p - novel}}$ classes to be used for evaluating the generalized FSOD performance of a fine-tuned model during meta-tuning. + +We generate a series of FSOD proxy tasks for metatuning, similar to episodic meta-learning: at each proxy task $T$ , we sample a few-shot training set from $D_{\mathrm{p - support}}$ . We also sample a loss function/augmentation magnitude parameter combination $\rho$ , where each $\rho_j \in \rho$ is modeled in terms of a Gaussian distribution: $\rho_j \sim \mathcal{N}(\mu_j, \sigma^2)$ . Using the loss function or augmentations corresponding to the sampled $\rho$ , we fine-tune the initial model on the support images using gradient-based optimization, and compute the mean + +![](images/ff3bc558467eb304c8dc5cfa596f6af2101a5be28269c400feb2da6860cc9a5c.jpg) +Figure 2. The meta-tuning approach. At each RL iteration over a proxy task, the distribution parameters modeling the loss function and augmentations are updated as a function of the obtained mAP scores, towards improved training with few-samples. + +average precision (mAP) scores on $D_{\mathrm{p - query}}$ . We get multiple mAP scores by repeating this process multiple times over multiple proxy support samples. Meta-tuning is then carried over by updating $\mu$ values via the REINFORCE rule after each episode, towards finding $\mu$ values centered around well-performing $\rho$ combinations. + +$$ +\mu_ {j} ^ {\prime} \leftarrow \mu_ {j} + \eta R (\rho) \nabla_ {\mu} \log (p \left(\rho_ {j}; \mu_ {j}, \sigma\right)) \tag {4} +$$ + +where $p(\rho ;\mu ,\sigma)$ is the Gaussian probability density function, $\eta$ is the RL learning rate. + +We apply the REINFORCE update rule using the $\rho$ with the highest reward per episode. $R(\rho)$ is the normalized reward function obtained by whitening the mAP scores. We empirically observe that normalization improves the results (Section 4) since without reward normalization, the RL updates are scaled with respect to the inherent difficulty of the proxy task, which greatly varies depending on the sampled support examples. Reward normalization approximately removes the average reward, enabling better performing $\rho$ samples to influence based on their relative success. + +Finally, similar to [50], starting with $\sigma = 0.1$ , we diminish $\sigma$ over the RL iterations to progressively reduce explorations by sampling more conservatively, which improves converge. The final scheme is illustrated in Figure 2. + +# 4. Experiments + +Metrics. We use mAP to evaluate the base and novel class detection results separately. To evaluate the generalized FSOD performance, we use the Harmonic Mean (HM) metric to compute a balanced aggregation of base and novel class performance scores. Adapted from generalized zero-shot learning [74], HM is defined as the harmonic mean of $\mathrm{mAP}_{\mathrm{base}}$ and $\mathrm{mAP}_{\mathrm{novel}}$ scores. + +Datasets. We use Pascal VOC [13] and MS COCO [40] with the same splits defined in FSOD benchmarks [65, 72]. + +On Pascal VOC, three separate base/novel class splits exist, where each one consists of 15 base and 5 novel classes. In each split, we select 5 base classes to mimic novel classes during meta-tuning. On MS-COCO, we select 15 base classes to mimic novel classes in each proxy task, and evaluate the models for the 10-shot and 30-shot settings. + +Baselines. We primarily use the MPSR [72] and DeFRCN [53] as our baselines, which are among the best performing fine-tuning based FSOD methods on Pascal VOC. For the DeFRCN experiments, we transfer the meta-tuned loss functions and augmentation magnitudes from MPSR to the DeFRCN method, which are both based on Faster-RCNN. We take the results for FRCN [76], Ret. R-CNN [15], Meta-RCNN [76], FSRW [31], MetaDet [69], FsDetView [75] and ONCE [52] from [15] for a fair comparison. For the MPSR, DeFRCN (seed is set to 0) and FSCE [61], we report the results we obtain experimentally. We take the results for TFA+Hal [84], CME [36], TIP [35], DCNet [27], QA-FewDet [21] FADI [6], LVC [32], KFSOD [83] and FCT [22] from the original papers. Finally, while it is difficult to fairly compare fine-tuning versus meta-learning based approaches, we provide a discussion in the supplementary material. + +Implementation details. We use 200 RL episodes for loss function meta-tuning, with REINFORCE learning rate set to 0.0005. The meta-tuning for augmentation parameter is carried out using the trained and frozen the loss function parameters. We keep the fine-tuning implementation details of MPSR unchanged, which uses 4000 and 8000 gradient descent iterations for 10-shot and 30-shot experiments on MS-COCO, and 2000 iterations on Pascal VOC. We will publish the full source code upon publication; a preliminary version is provided as supplementary material. + +# 4.1. Main results + +We first compare the meta-tuning results against the corresponding MPSR baseline in Table 1. In the table, Meta + +
Method/ShotPascal VOCMS-COCO
Novel ClassesAll Classes (HM)Novel ClassesAll Classes (HM)
12351012351010301030
MPSR [72]33.137.244.347.152.143.147.454.557.260.89.113.711.515.0
MPSR+Meta-Static33.439.445.147.352.643.750.455.457.561.410.114.812.716.4
MPSR+Meta-Dynamic34.539.845.048.252.545.051.055.558.361.611.914.914.316.6
MPSR+Meta-ScaledDynamic35.240.345.848.452.945.651.255.958.361.812.315.014.416.7
MPSR+Aug34.638.646.048.352.745.149.556.258.462.09.914.912.516.3
MPSR+Meta-Static+Aug35.339.146.148.452.745.949.956.258.361.810.215.212.816.7
MPSR+Meta-Dynamic+Aug35.439.646.548.953.346.050.556.858.962.512.115.314.516.8
MPSR+Meta-ScaledDynamic+Aug35.840.646.849.253.746.351.557.059.262.712.515.414.716.9
+ +Table 1. FSOD (mAP) and G-FSOD (HM of the base and novel class mAPs) results on Pascal VOC and MS-COCO datasets for MPSR baseline method. HM stands for harmonic mean. + +Static, Meta-Dynamic, Meta-ScaledDynamic refer to meta-tuning a single temperature, dynamic temperature, and novel class scaled dynamic temperature functions, respectively. Similarly, Aug, Meta-Static $^+$ Aug, Meta-Dynamic $^+$ Aug, and Meta-ScaledDynamic $^+$ Aug refer to meta-tuning only augmentation, single temperature and augmentation, dynamic temperature and augmentation, and novel class scaled dynamic temperature and augmentation functions, respectively. We observe that meta-tuning consistently improves the FSOD and G-FSOD results of the MPSR model. We also observe steady improvements gradually from the baseline to Meta-Static, to Meta-Dynamic, and finally to Meta-ScaledDynamic. In addition, the meta-tuned augmentation magnitude parameter also contributes positively to the few-shot object detection performance. The overall consistency of the improvements provides positive evidence for the value of loss and augmentation meta-tuning. + +Pascal VOC results. In Table 2, we report the Pascal VOC results for our MPSR and DeFRCN based Meta-ScaledDynamic+Aug approach and compare them against the state-of-the-art fine-tuning based FSOD methods. While we present the scores averaged over the three splits in this table, additional per-split FSOD and G-FSOD results can be found in the supplementary material. The left side of Table 2 presents the FSOD results for the varying number of support images. We observe that DeFRCN combined with Meta-ScaledDynamic+Aug, i.e. meta-tuning of the score coefficient, dynamic temperature and the augmentation parameter, yields the best mAP scores in all $k$ -shot settings among all methods. + +The right side of Table 2 presents the G-FSOD results on Pascal VOC. We observe that the best-performing Meta-ScaledDynamic+Aug method improves the HM scores further above the state-of-the-art in all $k$ -shot settings. Overall, these results suggest that the proposed framework is an effective way for meta-learning inductive biases to be used in fine-tuning-based FSOD. + +Figure 3 presents visual detection examples without and with meta-tuned scaled dynamic temperature and augmentations in the first and second rows, respectively. We observe + +various improvements, such as reductions in false positives, improved recall, and more precise boxes, most likely due to the improved model fitting in the low-data regime. + +MS-COCO results. In Table 3, we compare the MPSR and DeFRCN based Meta-ScaleDynamic+Aug results against other fine-tuning based FSOD methods that report 10-shot and 30-shot results on the MS-COCO dataset. We observe that with meta-tuning, the FSOD scores of MPSR improve from 9.1 to 12.5 (10-shot mAP), and from 13.7 to 15.4 (30-shot mAP). We also observe that the scores of DeFRCN improve from 18.5 to 18.8 (10-shot mAP), and from 21.9 to 23.4 (30-shot mAP), obtaining the best and second best results against all other models. Similarly, in the case of GFSD, with meta-tuning, the 10-shot HM score of DeFRCN improves from 24.0 to 24.4, outperforming all other models. In addition, the 30-shot HM score of DeFRCN improves from 26.8 to 28.0, which is slightly below the 28.1 score of LVC-PL [32]. + +# 4.2. Ablation studies + +Meta-tuning details. The proposed meta-tuning approach involves three important technical details: Proxy-novel imitation, model re-initialization, and reward normalization. Proxy-novel imitation refers to reinforcement learning over the sampled proxy-novel tasks, instead of the whole training set, to mimic the test-time FSOD challenges. Model re-initialization is the re-initialization of the base model for each task. Without re-initialization, not only the sampled loss/augmentation parameters and tasks but also the accumulated model updates undesirably affect the rewards. Reward normalization further reduces the effect of task difficulty variance by normalizing the rewards obtained within a single episode, allowing a more isolated assessment of the sampled loss functions and augmentations. + +We evaluate the contributions of these three important details in terms of G-FSOD HM scores using the 5-shot setting of Pascal VOC Split-1 with MPSR+Meta-Dynamic. The results averaged over 5 runs are given in Table 4. We observe that each component progressively improves the + +
Method/ShotNovel ClassesAll Classes (HM)
123510123510
FRCN [76] (ICCV'19)16.120.628.833.436.525.931.740.044.346.7
TFA-fc [65] (ICML'20)27.630.639.846.648.740.544.152.958.359.9
TFA-cos [65] (ICML'20)31.432.640.546.848.344.646.053.558.459.6
FSCE [61] (CVPR'21)29.236.342.547.152.241.848.854.257.761.0
Ret. R-CNN [15] (CVPR'21)31.437.141.446.848.844.750.554.759.160.8
TFA+Hal [84] (CVPR'21)32.935.540.446.348.1-----
FADI [6] (NeurIPS'21)42.246.547.952.456.9-----
LVC [32] (CVPR'22)30.935.443.651.154.1-----
LVC-PL [32] (CVPR'22)45.245.054.857.558.6-----
MPSR [72] (ECCV'20)33.137.244.347.152.143.147.454.557.260.8
DeFRCN [53] (ICCV'21)46.552.655.960.060.857.662.564.767.667.8
MPSR+Meta-ScaledDynamic+Aug35.840.646.849.253.746.351.557.059.262.7
DeFRCN+Meta-ScaledDynamic+Aug49.254.057.261.361.859.863.765.968.668.7
+ +![](images/c84e8ccae23d30cfef10a3a7b73b22f5d6316a4b99065d13a673d414046cfd03.jpg) + +![](images/07d07e33e01afe655237ea462d0df27346a0d4cdfbb053f7cfaf3c4d86cc597d.jpg) +Split-1/5-shot + +![](images/9be8da8655e2408b293b01a02e818e701c1b1257b9293d94df3792dcd139c9bb.jpg) + +![](images/33840c689ca2bf073231d0a4381805805015ddb9a7011954efdb97c8067e658a.jpg) +Split-1/10-shot + +![](images/00cd92d0f5889b6ce2ac6aff11354bb9c79e68d040a21a5b7854fbf4d8bd6304.jpg) + +![](images/0ae390992bf7542c471325f10dc035b0f094f06d6f53708c39e863bcfa10e780.jpg) +Split-2/5-shot +Figure 3. Qualitative results using MPSR without (first row) and with (second row) meta-tuning, over multiple Pascal VOC splits. Base and novel class detections are shown with green and red boxes, respectively. (Best viewed in color.) + +![](images/0ba54553e65bf95e2dd0604a7fffd07cf776df0ecce98278d36adc2cc70c8628.jpg) + +![](images/129654a3e58852f10898f12a210faa73be234f73690e28553587a21c7f725414.jpg) +Split-2/10-shot + +![](images/7f6f3319f594f4eda440e44e5e8e8252502128783c0ae46b0187cd0b99fe0c49.jpg) + +![](images/0d4fedf116c18219edb06b9359f08470a8b32ba8b790502595be9e0fb0385995.jpg) +Split-3/5-shot + +![](images/199345f0ee8602ff84564ce16d0aee3eaea222704f924c2543e8fbfb22f742d3.jpg) + +![](images/0b61ecf822828a1a2fd411c35799eb9412ec39cd0b45e31200dfc08dc0534076.jpg) +Split-3/10-shot + +Table 2. FSOD (mAP) and G-FSOD (HM of the base and novel class mAPs) results on Pascal VOC. The best and the second-best results are marked with red and blue. HM stands for harmonic mean. + +
Method/ShotsNovel ClassesAll Classes (HM)
10-shot30-shot10-shot30-shot
FRCN [76] (ICCV'19)9.212.512.815.6
FRCN-BCE [65] (ICML'20)6.410.310.916.1
TFA-fc [65] (ICML'20)10.013.415.419.4
TFA-cos [65] (ICML'20)10.013.715.619.8
MPSR [72] (ECCV'20)9.113.711.515.0
FSCE [61] (CVPR'21)10.514.416.020.2
Ret. R-CNN [15] (CVPR'21)10.513.816.620.4
FADI [6] (NeurIPS'21)12.216.1--
DeFRCN [53] (ICCV'21)18.521.924.026.8
LVC [32] (CVPR'22)12.117.817.822.8
LVC-PL [32] (CVPR'22)17.824.522.828.1
MPSR+Meta-ScaledDynamic+Aug12.515.414.716.9
DeFRCN+Meta-ScaledDynamic+Aug18.823.424.428.0
+ +HM scores, and the most significant contribution is made by reward normalization, which improves from 62.1 to 63.3. We also observe that reward normalization considerably improves the overall experimental stability. To quantify this + +Table 3. Comparison of Meta-ScaledDynamic results to the fine-tuning based (G-)FSOD methods on the MS-COCO dataset. The best and the second-best results are marked with red and blue. + +
Proxy-novel imit.Model re-init.Reward norm.HM
XXX61.5
✓XX61.8
✓✓X62.1
✓✓✓63.3
+ +Table 4. Evaluation of meta-tuning details. Proxy-novel imitation is the imitation of novel classes using a subset of base classes. Model re-initialization is the re-initialization of the base model at each task. Reward normalization is within-episode normalization of the mAP scores during meta-tuning. + +observation, we estimate the $95\%$ confidence interval over the runs using $CI = 1.96\frac{s}{\sqrt{n}}$ , where $s$ , $n$ , and $1.96$ are the standard deviation, number of runs, and $Z$ -value, respectively [65]. According to this estimator, the normalization step narrows the confidence interval from $\pm 0.75$ to $\pm 0.13$ , providing a clear improvement in reliability. + +Learned loss functions. In Figure 4, we plot the learned loss functions according to the $\mu$ values obtained at the end of the RL process. The upper plot shows the dynamic + +![](images/398273ef69347b6ecae7cbc7d987dc77d5389ebf546fa8f522214dff745fcfc8.jpg) + +![](images/d30b7a0b618214bfbeeb4284573888ed1e96615f5a9d492058a0a79c92609f25.jpg) +Figure 4. The dynamic temperature functions and score scaling coefficients learned by the meta-tuning process, using Meta-Dynamic (upper) and Meta-ScaledDynamic (lower) formulations. Results for each Pascal VOC split is shown with a separate curve. + +temperature functions learned over three different splits. We observe that temporally attenuated temperature values are preferred consistently, sharpening the predictions towards the end of the fine-tuning process. The lower plot shows the learned dynamic temperature functions with novel class score scaling. The learned scaling coefficients, i.e. $\mu_{\alpha}$ of the learned $\rho_{\alpha}$ distribution, are shown as horizontal lines. We observe that similar dynamic temperature functions are learned, and $\mu_{\alpha}$ values vary between 1.09 to 1.2, suggesting that the meta-tuning process learns to boost the novel class scores. The interpretability of these outcomes, we believe, highlights a significant advantage of loss meta-tuning. In the context of interpretability, we observe that as the fine-tuning process continues on the few-shot training set, the predictions are progressively made sharper, i.e. the loss becomes more sensitive to classification errors and enforces towards making more confident correct predictions. This is in alignment with one of our original motivations for reducing the dominating classification errors in G-FSOD, as the meta-tuning process automatically learns to enforce more accurate classifications, where the curve steepness and the numerical ranges are learned via RL. + +Learned augmentations. The learned photometric augmen + +
S/MTFA [65]TFA+Hal [84]TFA+Meta-ScaledDynamic+Aug
13.43.84.7
24.65.05.8
36.66.97.1
+ +Table 5. Low-shot (1-shot, 2-shot and 3-shot) experiments on MS-COCO dataset with novel classes. + +tation magnitude values learned are 0.29, 0.24, 0.13, and 0.36 for Pascal VOC split-1, split-2, split-3, and MS-COCO datasets, respectively. We observe that the learned augmentation magnitudes positively contribute to the performance. According to the results in Table 1, the average Pascal VOC split-1/1-shot score increases from 33.1 to 34.6 with only augmentation steps. + +Very low-shot experiments. Finally, we evaluate the meta-tuning approach in low-shot many-class settings. [84] proposes TFA+Hal method that uses the TFA baseline and conducts 1-shot, 2-shot, and 3-shot FSOD on the MS-COCO dataset. As we already observe the positive effects of the loss terms and augmentation magnitudes obtained from the MPSR on the DeFRCN, we similarly apply the learned parameters to the TFA baseline. The results are presented in Table 5. We observe that results are consistently improved using the meta-tuned functions on the TFA baseline. + +# 5. Conclusion + +Fine-tuning based frameworks offer simple and reliable approaches to building detection models from few samples. However, a major limitation of the existing fine-tuning-based FSOD models is their focus on the hand-crafting the design of fine-tuning details for few-shot training, which is inherently difficult and likely to be sub-optimal. Towards addressing this limitation, we propose to meta-learn the fine-tuning based learning dynamics as a way of introducing learned inductive biases for few-shot learning. The proposed tuning scheme uses meta-learning principles with reinforcement learning, and obtains interpretable loss functions and augmentation magnitudes for few-shot training. Our comprehensive experimental results on Pascal VOC and MS COCO datasets show that the proposed meta-tuning approach consistently provides significant performance improvements over the strong fine-tuning based few-shot detection baselines in both FSOD and G-FSOD settings. + +While we restrict our experiments to loss and augmentation functions, meta-tuning other learning components, e.g. initial model, and applications to other few-shot learning problems can be interesting future work directions. + +Acknowledgements. This work was supported in part by the TUBITAK Grant 119E597 and a Google Faculty Research Award. + +# References + +[1] Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, and Frank Wood. Enhancing few-shot image classification with unlabelled examples. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2796-2805, January 2022. 2 +[2] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Sigal. Improved Few-Shot Visual Classification. arXiv e-prints, page arXiv:1912.03432, Dec. 2019. 2 +[3] Luca Bertinetto, Joao F. Henriques, Philip Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Proc. Int. Conf. Learn. Represent., 2019. 2 +[4] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-Shot Segmentation Without Meta-Learning: A Good Transductive Inference Is All You Need? In arXiv:2012.06166 [cs], 2021. 1 +[5] Kaidi Cao, Maria Brbić, and Jure Leskovec. Concept learners for few-shot learning. In Proc. Int. Conf. Learn. Represent., 2021. 2 +[6] Yuhang Cao, Jiaqi Wang, Ying Jin, Tong Wu, Kai Chen, Ziwei Liu, and Dahua Lin. Few-shot object detection via association and discrimination. Proc. Adv. Neural Inf. Process. Syst., 34:16570-16581, 2021. 1, 5, 7 +[7] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 8819-8828, 2021. 1 +[8] Tung-I Chen, Yueh-Cheng Liu, Hung-Ting Su, Yu-Cheng Chang, Yu-Hsiang Lin, Jia-Fong Yeh, Wen-Chin Chen, and Winston Hsu. Dual-awareness attention for few-shot object detection. IEEE Transactions on Multimedia, 2021. 1, 2 +[9] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A Closer Look at Few-shot Classification. In ICLR 2019, 2019. 1, 2 +[10] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 3 +[11] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proc. IEEE Conf. Comput. Vis. Pattern Recog. Workshops, pages 702-703, 2020. 3, 4 +[12] Guneet S Dhillon, Pratik Chaudhari, Avinash Ravichandran, and Stefano Soatto. A Baseline for Few-Shot Image Classification. In ICLR, page 20, 2020. 1, 2 +[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 5 +[14] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4013–4022, 2020. 3 +[15] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4527-4536, 2021. 1, 2, 3, 5, 7 + +[16] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 1, 2 +[17] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proc. Int. Conf. Mach. Learn., volume 70, pages 1126–1135, 2017. 2 +[18] Golnaz Ghiasi, Yin Cui, A. Srinivas, Rui Qian, Tsung-Yi Lin, Ekin Dogus Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 2917-2927, 2021. 1 +[19] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proc. IEEE Int. Conf. on Computer Vision, 2019. 2 +[20] Santiago Gonzalez and Risto Miikkulainen. Improved training speed, accuracy, and data utilization through loss function optimization. In 2020 IEEE Congress on Evolutionary Computation (CEC), pages 1-8. IEEE, 2020. 2, 3 +[21] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proc. IEEE Int. Conf. on Computer Vision, pages 3263-3272, 2021. 3, 5 +[22] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 5321-5330, 2022. 1, 2, 5 +[23] Bharath Hariharan and Ross B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. Proc. IEEE Int. Conf. on Computer Vision, pages 3037-3046, 2017. 2 +[24] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 4 +[25] Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. In Proc. Int. Conf. Mach. Learn., pages 2731-2741. PMLR, 2019. 3 +[26] Ting-I Hsieh, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. One-shot object detection with co-attention and co-excitation. arXiv preprint arXiv:1911.12529, 2019. 3 +[27] Hanzhe Hu, Shuai Bai, Aoxue Li, Jinshi Cui, and Liwei Wang. Dense relation distillation with context-aware aggregation for few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 10185-10194, 2021. 3, 5 +[28] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference. arXiv e-prints, page arXiv:2204.07305, Apr. 2022. 2 +[29] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Proc. Adv. Neural Inf. Process. Syst., 33, 2020. 1 + +[30] Taewon Jeong and Heeyoung Kim. Ood-maml: Meta-learning for few-shot out-of-distribution detection and classification. In Advances in Neural Information Processing Systems, volume 33, pages 3907-3916, 2020. 1 +[31] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In Proc. IEEE Int. Conf. on Computer Vision, pages 8420-8429, 2019. 1, 2, 3, 5 +[32] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Label, verify, correct: A simple few shot object detection method. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 14237-14247, 2022. 1, 3, 5, 6, 7 +[33] Michalis Lazarou, Yannis Avrithis, and Tania Stathaki. Tensor feature hallucination for few-shot learning. ArXiv, abs/2106.05321, 2021. 2 +[34] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 10649-10657, 2019. 2 +[35] Aoxue Li and Zhenguo Li. Transformation invariant few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 3094-3102, 2021. 3, 5 +[36] Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. Beyond max-margin: Class margin equilibrium for few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 7363-7372, 2021. 1, 2, 3, 5 +[37] Chuming Li, Xin Yuan, Chen Lin, Minghao Guo, Wei Wu, Junjie Yan, and Wanli Ouyang. Am-lfs: Automl for loss function search. In Proc. IEEE Int. Conf. on Computer Vision, pages 8410-8419, 2019. 3 +[38] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. MetaSGD: Learning to Learn Quickly for Few-Shot Learning. arXiv e-prints, page arXiv:1707.09835, July 2017. 2 +[39] Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment. Proc. Adv. Neural Inf. Process. Syst., 32, 2019. 3 +[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conf. on Computer Vision, pages 740-755. Springer, 2014. 2, 5 +[41] Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters: Understanding margin in few-shot classification. arXiv preprint arXiv:2003.12060, 2020. 2 +[42] Peidong Liu, Gengwei Zhang, Bochao Wang, Hang Xu, Xiaodan Liang, Yong Jiang, and Zhenguo Li. Loss function discovery for object detection via convergence-simulation driven search. arXiv preprint arXiv:2102.04700, 2021. 2, 3 +[43] Shichen Liu, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Generalized Zero-Shot Learning with Deep Calibration Network. In NeurIPS, pages 2005–2015. 2018. 4 +[44] Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sungju Hwang, and Yi Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In Proc. Int. Conf. Learn. Represent., 2019. 2 + +[45] Yan Liu, Zhijie Zhang, Li Niu, Junjie Chen, and Liqing Zhang. Mixed supervised object detection by transferring mask prior and semantic similarity. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Proc. Adv. Neural Inf. Process. Syst., 2021. 1 +[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. Proc. IEEE Int. Conf. on Computer Vision, 2021. 1 +[47] Tsendsumen Munkhdalai, Xingdi Yuan, Soroush Mehri, and Adam Trischler. Rapid adaptation with conditionally shifted neurons. In ICML, 2018. 2 +[48] Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv: Learning, 2018. 2 +[49] Boris N. Oreshkin, Pau Rodriguez Lopez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. In NeurIPS, 2018. 2, 4 +[50] Matteo Papini, Andrea Battistello, and Marcello Restelli. Balancing learning speed and stability in policy gradient via adaptive exploration. In Proc. Int. Conf. on Artif. Intellig. and Stat., pages 1188-1199, 2020. 5 +[51] Eunbyung Park and Junier B. Oliva. Meta-curvature. In NeurIPS, 2019. 2 +[52] Juan-Manuel Perez-Rua, Xiatian Zhu, Timothy M Hospedales, and Tao Xiang. Incremental few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 13846-13855, 2020. 1, 2, 5 +[53] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In Proc. IEEE Int. Conf. on Computer Vision, pages 8681-8690, 2021. 1, 2, 3, 5, 7 +[54] Aravind Rajeswaran, Chelsea Finn, Sham M. Kakade, and Sergey Levine. Meta-learning with implicit gradients. In NeurIPS, 2019. 2 +[55] Esteban Real, Chen Liang, David So, and Quoc Le. Automlzero: Evolving machine learning algorithms from scratch. In Proc. Int. Conf. Mach. Learn., pages 8007-8019. PMLR, 2020. 3 +[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Proc. Adv. Neural Inf. Process. Syst., 28:91-99, 2015. 3 +[57] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, MingYu Liu, Yong Jae Lee, Alexander G. Schwing, and Jan Kautz. Instance-aware, context-focused, and memoryefficient weakly supervised object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2020. 1 +[58] Andrei A. Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In Proc. Int. Conf. Learn. Represent., 2019. 2 +[59] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In Proc. Int. Conf. Mach. Learn., volume 48, pages 1842-1850, 2016. 2 +[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Proc. Adv. Neural Inf. Process. Syst., volume 30, 2017. 2 + +[61] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 7352-7362, 2021. 1, 2, 3, 4, 5, 7 +[62] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 1199-1208, 2018. 2 +[63] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B. Tenenbaum, and Phillip Isola. Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need? In Proc. European Conf. on Computer Vision, 2020. 1, 2 +[64] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proc. Adv. Neural Inf. Process. Syst., volume 29, 2016. 2 +[65] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. arXiv preprint arXiv:2003.06957, 2020. 1, 2, 3, 5, 7, 8 +[66] Xiaobo Wang, Shuo Wang, Cheng Chi, Shifeng Zhang, and Tao Mei. Loss function search for face recognition. In Proc. Int. Conf. Mach. Learn., pages 10029-10038. PMLR, 2020. 3, 4 +[67] Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning. arXiv preprint arXiv:1911.04623, 2019. 2 +[68] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-Shot Learning from Imaginary Data. arXiv:1801.05401 [cs], Jan. 2018. 2 +[69] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Meta-learning to detect rare objects. In Proc. IEEE Int. Conf. on Computer Vision, pages 9925-9934, 2019. 5 +[70] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 4 +[71] Among Wu, Suqi Zhao, Cheng Deng, and Wei Liu. Generalized and discriminative few-shot object detection via svddictionary enhancement. Proc. Adv. Neural Inf. Process. Syst., 34:6353-6364, 2021. 3 +[72] Jiaxi Wu, Songtao Liu, Di Huang, and Yunhong Wang. Multiscale positive sample refinement for few-shot object detection. In Proc. European Conf. on Computer Vision, pages 456-472. Springer, 2020. 1, 2, 3, 4, 5, 6, 7 +[73] Xiongwei Wu, Doyen Sahoo, and Steven Hoi. Meta-rcnn: Meta learning for few-shot object detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1679–1687, 2020. 1 +[74] Yongqin Xian, Bernt Schiele, and Zeynep Akata. Zero-shot learning-the good, the bad and the ugly. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4582-4591, 2017. 5 +[75] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In Proc. European Conf. on Computer Vision, pages 192-210. Springer, 2020. 1, 2, 3, 5 + +[76] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proc. IEEE Int. Conf. on Computer Vision, pages 9577-9586, 2019. 1, 2, 5, 7 +[77] Huaxiu Yao, Linjun Zhang, and Chelsea Finn. Meta-learning with fewer tasks through task interpolation. In Proceeding of the 10th International Conference on Learning Representations, 2022. 2 +[78] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 8805-8814, 2020. 2, 4 +[79] Li Yin, Juan M Perez-Rua, and Kevin J Liang. Sylph: A hypernetwork framework for incremental few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 9035-9045, 2022. 1, 2, 3 +[80] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., June 2020. 2 +[81] Gongjie Zhang, Zhipeng Luo, Kaiwen Cui, Shijian Lu, and Eric P Xing. Meta-detr: Image-level few-shot detection with inter-class correlation exploitation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3 +[82] Lu Zhang, Shuigeng Zhou, Jihong Guan, and Ji Zhang. Accurate few-shot object detection with support-query mutual guidance and hybrid loss. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 14424-14432, 2021. 3 +[83] Shan Zhang, Lei Wang, Naila Murray, and Piotr Koniusz. Kernelized few-shot object detection with efficient integral aggregation. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 19207-19216, 2022. 1, 2, 5 +[84] Weilin Zhang and Yu-Xiong Wang. Hallucination improves few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 13008-13017, 2021. 1, 2, 5, 7, 8 \ No newline at end of file diff --git a/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/images.zip b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fba6afe05710e84aeab336f22821f665d3c45b88 --- /dev/null +++ b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67a4c7da2568097d6ade4d45872d19f370d9d0478430738b10709c368725d208 +size 502109 diff --git a/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/layout.json b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..819ed1ceb151e613a10faef8135a2efd9c648d17 --- /dev/null +++ b/2023/Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection/layout.json @@ -0,0 +1,9020 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 102, + 523, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 102, + 523, + 138 + ], + "spans": [ + { + "bbox": [ + 71, + 102, + 523, + 138 + ], + "type": "text", + "content": "Meta-tuning Loss Functions and Data Augmentation for Few-shot Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 160, + 473, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 160, + 473, + 190 + ], + "spans": [ + { + "bbox": [ + 118, + 160, + 473, + 190 + ], + "type": "text", + "content": "Berkan Demirel1,2 Orhun Bugra Baran1 Ramazan Gokberk Cinbis1 \n1Middle East Technical University 2HAVELSAN Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 192, + 499, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 192, + 499, + 203 + ], + "spans": [ + { + "bbox": [ + 90, + 192, + 499, + 203 + ], + "type": "text", + "content": "bdemirel@havelsan.com.tr bugra@ceng.metu.edu.tr gcinbis@ceng.metu.edu.tr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "type": "text", + "content": "Few-shot object detection, the problem of modelling novel object detection categories with few training instances, is an emerging topic in the area of few-shot learning and object detection. Contemporary techniques can be divided into two groups: fine-tuning based and meta-learning based approaches. While meta-learning approaches aim to learn dedicated meta-models for mapping samples to novel class models, fine-tuning approaches tackle few-shot detection in a simpler manner, by adapting the detection model to novel classes through gradient based optimization. Despite their simplicity, fine-tuning based approaches typically yield competitive detection results. Based on this observation, we focus on the role of loss functions and augmentations as the force driving the fine-tuning process, and propose to tune their dynamics through meta-learning principles. The proposed training scheme, therefore, allows learning inductive biases that can boost few-shot detection, while keeping the advantages of fine-tuning based approaches. In addition, the proposed approach yields interpretable loss functions, as opposed to highly parametric and complex few-shot meta-models. The experimental results highlight the merits of the proposed scheme, with significant improvements over the strong fine-tuning based few-shot detection baselines on benchmark Pascal VOC and MS-COCO datasets, in terms of both standard and generalized few-shot performance metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Object detection is one of the computer vision problems that has greatly benefited from the advances in supervised deep learning approaches. However, similar to the case in many other problems, state-of-the-art in object detection relies on the availability of large-scale fully-annotated datasets, which is particularly problematic due to the difficulty of collecting accurate bounding box annotations [18,46]. This practical burden has lead to a great interest in the approaches" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 216, + 545, + 322 + ], + "blocks": [ + { + "bbox": [ + 308, + 216, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 308, + 216, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 216, + 545, + 322 + ], + "type": "image", + "image_path": "69104d70ca31758b63e72fb1d45a40fc0573a5278fa7cf9a8cbd4dfd31242d74.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 331, + 541, + 342 + ], + "lines": [ + { + "bbox": [ + 310, + 331, + 541, + 342 + ], + "spans": [ + { + "bbox": [ + 310, + 331, + 541, + 342 + ], + "type": "text", + "content": "Figure 1. The overall architecture of the meta-tuning approach." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 352, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 483 + ], + "type": "text", + "content": "that can potentially reduce the annotation cost, such as weakly-supervised learning [29, 57], learning from point annotations [7], and mixed supervised learning [45]. A more recently emerging paradigm in this direction is few-shot object detection (FSOD). In the FSOD problem, the goal is to build detection models for the novel classes with few labeled training images by transferring knowledge from the base classes with a large set of training images. In the closely related Generalized-FSOD (G-FSOD) problem, the goal is to build few-shot detection models that perform well on both base and novel classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 484, + 548, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 548, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 548, + 688 + ], + "type": "text", + "content": "FSOD methods can be categorized into meta-learning and fine-tuning approaches. Although meta-learning based methods are predominantly used in the literature in FSOD research [8,22,31,36,52,75,76,79,81,83], several fine-tuning based works have recently reported competitive results [6, 15,32,53,61,65,72,84]. The main premise of meta-learning approaches is to design and train dedicated meta-models that map given few train samples to novel class detection models, e.g. [73] or learn easy-to-adapt models [30] in a MAML [16] fashion. In contrast, however, fine-tuning based methods tackle the problem as a typical transfer learning problem and apply the general purpose supervised training techniques, i.e. regularized loss minimization via gradient-based optimization, to adapt a pre-trained model to few-shot classes. It is also worth noting that the recent results on fine-tuning based FSOD are aligned with related observations on few-shot classification [9, 12, 63] and segmentation [4]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "While some of the FSOD meta-learning approaches are attractive for being able to learn dedicated parametric training" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7339" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "mechanisms, they also come with two important shortcomings: (i) the risk of overfitting to the base classes used for training the meta-model due to model complexity, and (ii) the difficulty of interpreting what is actually learned; both of which can be crucially important for real-world, in-the-wild utilization of a meta-learned model. From this point of view, the simplicity and generality of a fine-tuning based FSOD approach can be seen as major advantages. In fact, one can find a large machine learning literature on the components (optimization techniques, loss functions, data augmentation, and architectures) of an FT approach, as opposed to the unique and typically unknown nature of a meta-learning inference model, especially when the model aims to replace standard training procedures for modeling the novel few-shot classes. While MAML [16] like meta-learning for quick adaptation is closer in nature to fine-tuning based approaches, the vanishing gradient problems and the overall complexity of the meta-learning task practically limits the approach to target only one or few model update steps, whereas an FT approach has no such computational difficulty." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 314, + 289, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 289, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 289, + 457 + ], + "type": "text", + "content": "Perhaps the biggest advantage of a fine-tuning based FSOD approach, however, can also be its biggest disadvantage: its generality may lack the inductive biases needed for effective learning with few novel class samples while preserving the knowledge of base classes. To this end, such approaches focus on the design of fine-tuning details, e.g. whether to freeze the representation parameters [65], use contrastive fine-tuning losses [61], increase the novel class variances [84], introduce the using additional detection heads and branches [15, 72]. However, optimizing such details specifically for few-shot classes in a hand-crafted manner is clearly difficult, and likely to be sub-optimal." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 460, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 289, + 687 + ], + "type": "text", + "content": "To address this problem, we focus on applying meta-learning principles to tune the loss functions and augmentations to be used in the fine-tuning stage for FSOD, which we call meta-tuning (Figure 1). More specifically, much like the meta-learning of a meta-model, we define an episodic training procedure that aims to progressively discover the optimal loss function and augmentation details for FSOD purposes in a data-driven manner. Using reinforcement learning (RL) techniques, we aim to tune the loss function and augmentation details such that they maximize the expected detection quality of an FSOD model obtained by fine-tuning to a set of novel classes. By defining meta-tuning over well-designed loss terms and an augmentation list, we restrict the search process to effective function families, reducing the computational costs compared to AutoML methods that aim to discover loss terms from scratch for fully-supervised learning [20, 42]. The resulting meta-tuned loss functions and augmentations, therefore, inject the learned FSOD-specific inductive biases into a fine-tuning based approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 713 + ], + "type": "text", + "content": "To explore the potential of the meta-tuning scheme for FSOD, we focus on the details of classification loss func" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": "tions, based on the observations that FSOD prediction mistakes tend to be in classification rather than localization details [61]. In particular, we first focus on the softmax temperature parameter, for which we define two versions: (i) a simple constant temperature, and (ii) time (fine-tuning iteration index) varying dynamic temperature, parameterized as an exponentiated polynomial. In all cases, the parameters learned via meta-tuning yield an interpretable loss function that has a negligible risk of over-fitting to the base classes, in contrast to a complex meta-model. We also model augmentation magnitudes during meta-tuning for improving the data loading pipeline for few-shot learning purposes. Additionally, we incorporate a score scaling coefficient for learning to balance base versus novel class scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 239, + 548, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 548, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 548, + 337 + ], + "type": "text", + "content": "We provide an experimental analysis on the Pascal VOC [13] and MS-COCO [40] benchmarks for FSOD, using the state-of-the-art fine-tuning based baselines MPSR [72] and DeFRCN [53]. Our experimental results show that the proposed meta-tuning approach provides significant performance gains in both FSOD and Generalized FSOD settings, suggesting that meta-tuning loss functions and data augmentation can be a promising direction in FSOD research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 348, + 392, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 392, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 392, + 361 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 369, + 547, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 405 + ], + "type": "text", + "content": "This section provides an overview of recent developments on few-shot image classification, few-shot object detection, automated loss function and data augmentation discovery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 409, + 548, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 548, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 548, + 529 + ], + "type": "text", + "content": "Few-shot classification. Most of the meta-learning approaches for few-shot learning (FSL) of classification models can be grouped as adaptation-based and mapping-based approaches. Adaptation-based (also called gradient-based) approaches aim to learn model parameters that can easily be adapted to new unseen few-shot tasks within a few model update steps, e.g. [17, 38, 47, 48, 51, 54, 58]. Mapping-based approaches (also called metric-based) aim to bypass a gradient-descent based adaptation step, and instead learn a data-to-classifier mapping, e.g. [5, 44, 49, 59, 60, 62, 64, 77, 78, 80]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 529, + 548, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 548, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 548, + 625 + ], + "type": "text", + "content": "Some of the other notable approaches include learning to generate synthetic data for novel classes [23, 33, 68], using better feature representations [1, 2, 19, 28, 41, 63, 67] or utilizing differentiable convex solvers [3, 34]. Importantly, several works highlight that a carefully trained representation combined with simple fine-tuning or even just shallow classifiers can yield competitive or better performance than meta-learning based approaches, e.g. [9, 12, 63]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "content": "Few-shot object detection. The FSOD approaches can be summarized as meta-learning and fine-tuning (also called transfer-learning) based ones. Most meta-learning based FSOD approaches embrace formulations similar to those used in mapping-based meta-learning approaches for FSL, e.g. [8, 22, 31, 36, 52, 75, 76, 79, 81, 83]. Support feature aggregation is one of the main aspects where meta-learning-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7340" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "based methods differ from each other. Xiao and Marlet [75] use both the differences and the channel-wise multiplication of the features in addition to the combination of the features directly for support-query aggregation. Fan et al. [14] use attention blocks to make support and query features more distinguishable for base and novel object classes. Zhang et al. [81] use inter-class correlations to highlight important support features. Li et al. [36] propose to use specialized support and query features for classification and localization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "content": "Recent efforts towards improving meta-learning based FSOD include complimentary techniques, mainly to improve loss functions, feature matching, and novel class sample usage efficiency. [36] uses class margin loss, [26] uses margin-based ranking loss, [82] uses hybrid loss which consist of focal loss, adaptive margin loss and contrastive loss. Hu et al. [27] perform feature matching between query and support images to use the information from the support images more effectively. Similarly, Han et al. [21] construct a matching network between query and support instances using heterogeneous graph convolutional networks. Li and Li [35] augment novel class samples via adding Gaussian noise. Yin et al. [79] decouple classification task from localization by using the proposed class-conditional architecture." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 289, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 663 + ], + "type": "text", + "content": "Fine-tuning-based methods typically freeze parts of a pre-trained detection network, add auxiliary detection heads, increase the novel class variances and then apply gradient descent based model update steps, unlike meta-learning-based methods that use complex episodic learning [15, 32, 53, 61, 65, 71, 72]. Wang et al. [65] propose a Faster-RCNN [56] based approach, where the class-agnostic region proposal network (RPN) component is kept frozen during fine-tuning. Sun et al. [61] use a similar approach and differently include FPN and RPN layers to the learnable parameter set in the same architecture. These learnable layers allow using contrastive proposal encodings that facilitate the more accurate classification of novel objects. Wu et al. [72] show that the scale distribution of support set tends to be imbalanced, and proposes a multi-scale positive sample refinement (MPSR) branch as an addition to the main model. Fan et al. [15] propose Retentive R-CNN architecture to prevent forgetting during fine-tuning for base classes. The obtained object proposals are fed into two ROI detectors responsible for base class and novel class instances. Qiao et al. [53] focus on decoupling network modules, and introduce a gradient decoupling layer and prototypical calibration block. Kaul et al. [32] extend the novel class annotations in the training set. In this context, the proposed method obtains object candidates from the base detector, and applies the box refinement step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "content": "While our approach is based on fine-tuning based FSOD, we embrace meta-learn principles to optimize the loss function and augmentations to improve the fine-tuning process for FSOD, without learning a complex and over-fitting-prone" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "content": "meta-model. The resulting loss function and data augmentations are then utilized within the fine-tuning steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 100, + 547, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 547, + 364 + ], + "type": "text", + "content": "Automated loss function discovery. Loss function discovery is an emerging AutoML topic towards improving the learning systems in a data-driven manner. Existing methods are mainly based on either (i) constructing the loss function directly from the basic operators [20, 42, 55] or (ii) optimizing parameterized loss functions [37, 66]. For loss construction, [42] proposes a genetic algorithm that consists of loss function verification and quality filtering modules. In this approach, the predefined proxy task eliminates divergent and poor candidate loss functions and survives the promising loss functions for other steps. [20] uses a genetic algorithm to select candidate loss functions from a tree of simple mathematical operations, and the successful loss functions pass to other stages to mutate. [55] suggests a method to learn not only the loss function but also the whole machine learning algorithm from scratch. For loss optimization, [37] re-analyzes the existing loss functions and presents them in a combined formula. [66] observes that the search space used in [37] can be too complex, and propose to simplify the search space via heuristics. In contrast to these works targeting supervised training scenarios, we aim to adapt loss function learning principles to the FSOD problem." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 368, + 548, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 548, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 548, + 534 + ], + "type": "text", + "content": "AutoML for data augmentation. A variety of automated data augmentation techniques have recently been proposed [10, 11, 25, 39]. Cubuk et al. [10] generate augmentation policies using reinforcement learning and a controller RNN. Ho et al. [25] propose a method that reduces the computational costs compared to [10] by using a population-based framework. Similarly, Lim et al. [39] propose a direct Bayesian method to reduce costs. Cubuk et al. [11] show that the optimal augmentation magnitudes tend to be similar across transformations, and the search process can greatly be simplified by using a shared value. We follow this suggestion and use a shared magnitude across the transforms in our formulation. In contrast to these works on supervised learning, however, we focus on learning detectors with few-samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 535, + 547, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 547, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 547, + 607 + ], + "type": "text", + "content": "In summary, while loss function and augmentation discovery topics increasingly attract attention towards improving supervised training pipelines, ours is the first work on learning few-sample specific inductive biases for fine-tuning based few-shot object detection based on meta-learning and AutoML principles, to the best of our knowledge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 617, + 361, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 361, + 629 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 361, + 629 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 637, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 637, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 305, + 637, + 547, + 673 + ], + "type": "text", + "content": "This section provides a brief summary of the FSOD problem definition and the baseline model we utilize. We then present our definition and instantiation of meta-tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 548, + 713 + ], + "type": "text", + "content": "Problem definition. We follow the FSOD setup of [31], where a relatively large set of training images for the set " + }, + { + "bbox": [ + 305, + 677, + 548, + 713 + ], + "type": "inline_equation", + "content": "C_b" + }, + { + "bbox": [ + 305, + 677, + 548, + 713 + ], + "type": "text", + "content": " of base classes is made available. Each training im" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7341" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "age corresponds to a tuple " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " consisting of image " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " and annotations " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "y = \\{y_0,\\dots,y_M\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". Each object annotation " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "y_{i} = \\{c_{i},b_{i}\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " contains a category label " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "(c_{i})" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " and a bounding box " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "(b_{i} = \\{x_{i},y_{i},w_{i},h_{i}\\})" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". Once the FSOD model training is complete, the evaluation is carried out based on a limited number " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "(k)" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " of training images made available for the set " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " of distinct novel (i.e. few-shot) classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 159, + 289, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 289, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 289, + 364 + ], + "type": "text", + "content": "Base model. We use the MPSR FSOD method [72] as the infrastructure for our loss function and data augmentation search methods. MPSR adapts the Faster-RCNN to be suitable for fine-tuning-based FSOD and uses an auxiliary multi-scale positive sample refinement (MPSR) branch to handle the scale scarcity problems. This branch expands the scale space of positive samples without increasing improper negative instances, unlike feature pyramid networks and image pyramids that do not change data distribution, hence the scale sparsity problem. In this context, objects in the images are cropped and resized in multiple sizes to create scale pyramids. The MPSR uses two groups of loss functions for the region proposal network (RPN) and detection heads, and feeds differently scaled positive samples to these loss functions together with the main detection branch. Finally, we note that the proposed approach can in principle be applied to virtually any fine-tuning based FSOD model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 370, + 194, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 194, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 194, + 383 + ], + "type": "text", + "content": "3.1. Meta-tuning loss functions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "type": "text", + "content": "Our main goal is to improve few-shot detector fine-tuning based on meta-learning principles. For meta-tuning the FSOD loss, we specifically focus on the classification loss term, as the FSOD errors tend to be primarily caused by misclassifications [61]. The MPSR classification loss term can be expressed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 468, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 468, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 468, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\ell_ {c l s} (x, y) = - \\frac {1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {f \\left(x _ {i} , y _ {i}\\right)}}{\\sum_ {y} e ^ {f \\left(x _ {i} , y\\right)}}\\right) \\tag {1}", + "image_path": "a8ac892a7410e71caf9877fd191f831309fba115676588fff8df1e52b8fa78de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "N_{ROI}" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": " is the number of ROIs (i.e. candidate regions) in an image, " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": " is the groundtruth class label for the " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": "-th ROI, and " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "f(x_{i},y)" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": " is the corresponding class " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": " prediction score. To add more flexibility into the loss function, we re-define it as a parametric function " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\ell_{cls}(x,y;\\rho)" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": " represents the loss function parameters. First, we introduce a temperature scalar " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\rho_{\\tau}" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\rho = (\\rho_{\\tau})" + }, + { + "bbox": [ + 46, + 507, + 287, + 592 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 598, + 287, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 598, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 56, + 598, + 287, + 641 + ], + "type": "interline_equation", + "content": "\\ell_ {c l s} (x, y; \\rho) = - \\frac {1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {f \\left(x _ {i} , y _ {i}\\right) / \\rho_ {\\tau}}}{\\sum_ {y ^ {\\prime}} e ^ {f \\left(x _ {i} , y ^ {\\prime}\\right) / \\rho_ {\\tau}}}\\right) \\tag {2}", + "image_path": "f70ff344c0bd8e29b4deba958788ec2a504c2ce36ed8fc17bb48bb053b77a4e2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Our motivation comes from the observations on the importance of temperature scaling in log loss on various other problems, such as knowledge distillation [24], few-shot classification [49, 78], and zero-shot learning [43]. While temperature is typically tuned in a manual manner, here we aim to meta-learn it specifically for fine-tuning based FSOD" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": "purposes, giving a chance to observe the behavior of metatuning in a simple case. We also define a more sophisticated variant of the loss function by defining the dynamic temperature function " + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "f_{\\rho}" + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": " and novel class scaling " + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 129, + 545, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 545, + 171 + ], + "type": "interline_equation", + "content": "\\ell_ {c l s} (x, y; \\rho) = \\frac {- 1}{N _ {R O I}} \\sum_ {i} ^ {N _ {R O I}} \\log \\left(\\frac {e ^ {\\alpha \\left(y _ {i}\\right) f \\left(x _ {i} , y _ {i}\\right) / f _ {\\rho} (t)}}{\\sum_ {y ^ {\\prime}} e ^ {\\alpha \\left(y ^ {\\prime}\\right) f \\left(x _ {i} , y ^ {\\prime}\\right) / f _ {\\rho} (t)}}\\right) \\tag {3}", + "image_path": "736ca05c14949387f080467368e925ff940eeb3e10ca2b4eefcb2aa9855ccfb6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "inline_equation", + "content": "f_{\\rho}(t) = \\exp (\\rho_{a}t^{2} + \\rho_{b}t + \\rho_{c})" + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "inline_equation", + "content": "\\rho = (\\rho_{a},\\rho_{b},\\rho_{c})" + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": " is a 3-tuple of polynomial coefficients, and " + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": " is the normalized fine-tuning iteration index. The temperature can increase or decrease over time, making the predicted class distributions smoother or sharper. " + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "inline_equation", + "content": "\\alpha (y)" + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": " is set to 1 for " + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "inline_equation", + "content": "y\\in C_b" + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": ", and otherwise the novel class score scaling coefficient " + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "inline_equation", + "content": "\\rho_{\\alpha}" + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": ", as a way to learn base and novel score balancing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 262, + 459, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 459, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 459, + 275 + ], + "type": "text", + "content": "3.2. Meta-tuning augmentations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 281, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 354 + ], + "type": "text", + "content": "For meta-tuning augmentations, we focus on the photometric augmentations that are likely to be transferable from base to novel classes. In this context, we model the brightness, saturation, contrast, and hue transforms, with a shared magnitude parameter " + }, + { + "bbox": [ + 304, + 281, + 547, + 354 + ], + "type": "inline_equation", + "content": "(\\rho_{aug})" + }, + { + "bbox": [ + 304, + 281, + 547, + 354 + ], + "type": "text", + "content": ", which is known to be effective for supervised training [11]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 360, + 437, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 437, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 437, + 373 + ], + "type": "text", + "content": "3.3. Meta-tuning procedure" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 378, + 547, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 438 + ], + "type": "text", + "content": "In our work, we utilize a REINFORCE [70] based reinforcement learning (RL) approach to search for the optimal loss function and augmentations, where we use the AutoML approach of Wang et al. [66] on loss function search for fully-supervised face recognition as our starting point." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": "In order to meta-tune the loss function and augmentations to maximize FSOD generalization abilities, we generate proxy tasks over base class training data to imitate real FSOD tasks over the novel classes. For this purpose, we divide base classes into two subsets, proxy-base " + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{p - base}}" + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": " and proxy-novel " + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{p - novel}}" + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": ". We then construct three non-overlapping data set splits using the base class training set: (i) " + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{p - pretrain}}" + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{p - base}}" + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": "-only samples, used for training a temporary object detection model for meta-tuning purposes; (ii) " + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{p - support}}" + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": " containing samples of " + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{p - base}} \\cup C_{\\mathrm{p - novel}}" + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": " classes to be used as fine-tuning images during meta-tuning; (iii) " + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{p - query}}" + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": " containing samples of " + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{p - base}} \\cup C_{\\mathrm{p - novel}}" + }, + { + "bbox": [ + 304, + 439, + 547, + 605 + ], + "type": "text", + "content": " classes to be used for evaluating the generalized FSOD performance of a fine-tuned model during meta-tuning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": "We generate a series of FSOD proxy tasks for metatuning, similar to episodic meta-learning: at each proxy task " + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": ", we sample a few-shot training set from " + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{p - support}}" + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": ". We also sample a loss function/augmentation magnitude parameter combination " + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\rho_j \\in \\rho" + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": " is modeled in terms of a Gaussian distribution: " + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\rho_j \\sim \\mathcal{N}(\\mu_j, \\sigma^2)" + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": ". Using the loss function or augmentations corresponding to the sampled " + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": ", we fine-tune the initial model on the support images using gradient-based optimization, and compute the mean" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7342" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 86, + 535, + 235 + ], + "blocks": [ + { + "bbox": [ + 50, + 86, + 535, + 235 + ], + "lines": [ + { + "bbox": [ + 50, + 86, + 535, + 235 + ], + "spans": [ + { + "bbox": [ + 50, + 86, + 535, + 235 + ], + "type": "image", + "image_path": "ff3bc558467eb304c8dc5cfa596f6af2101a5be28269c400feb2da6860cc9a5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 546, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 546, + 266 + ], + "type": "text", + "content": "Figure 2. The meta-tuning approach. At each RL iteration over a proxy task, the distribution parameters modeling the loss function and augmentations are updated as a function of the obtained mAP scores, towards improved training with few-samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "text", + "content": "average precision (mAP) scores on " + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{p - query}}" + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "text", + "content": ". We get multiple mAP scores by repeating this process multiple times over multiple proxy support samples. Meta-tuning is then carried over by updating " + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "text", + "content": " values via the REINFORCE rule after each episode, towards finding " + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "text", + "content": " values centered around well-performing " + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 271, + 289, + 344 + ], + "type": "text", + "content": " combinations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 353, + 287, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 353, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 85, + 353, + 287, + 368 + ], + "type": "interline_equation", + "content": "\\mu_ {j} ^ {\\prime} \\leftarrow \\mu_ {j} + \\eta R (\\rho) \\nabla_ {\\mu} \\log (p \\left(\\rho_ {j}; \\mu_ {j}, \\sigma\\right)) \\tag {4}", + "image_path": "de80d47a9124903a6901c209ef0d75ec50eb28cd64692d051b7444cc2c515bdd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 375, + 288, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 288, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 375, + 288, + 399 + ], + "type": "inline_equation", + "content": "p(\\rho ;\\mu ,\\sigma)" + }, + { + "bbox": [ + 46, + 375, + 288, + 399 + ], + "type": "text", + "content": " is the Gaussian probability density function, " + }, + { + "bbox": [ + 46, + 375, + 288, + 399 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 375, + 288, + 399 + ], + "type": "text", + "content": " is the RL learning rate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "text", + "content": "We apply the REINFORCE update rule using the " + }, + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "text", + "content": " with the highest reward per episode. " + }, + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "inline_equation", + "content": "R(\\rho)" + }, + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "text", + "content": " is the normalized reward function obtained by whitening the mAP scores. We empirically observe that normalization improves the results (Section 4) since without reward normalization, the RL updates are scaled with respect to the inherent difficulty of the proxy task, which greatly varies depending on the sampled support examples. Reward normalization approximately removes the average reward, enabling better performing " + }, + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "text", + "content": " samples to influence based on their relative success." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 518, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 288, + 567 + ], + "type": "text", + "content": "Finally, similar to [50], starting with " + }, + { + "bbox": [ + 46, + 518, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\sigma = 0.1" + }, + { + "bbox": [ + 46, + 518, + 288, + 567 + ], + "type": "text", + "content": ", we diminish " + }, + { + "bbox": [ + 46, + 518, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 518, + 288, + 567 + ], + "type": "text", + "content": " over the RL iterations to progressively reduce explorations by sampling more conservatively, which improves converge. The final scheme is illustrated in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "text", + "content": "Metrics. We use mAP to evaluate the base and novel class detection results separately. To evaluate the generalized FSOD performance, we use the Harmonic Mean (HM) metric to compute a balanced aggregation of base and novel class performance scores. Adapted from generalized zero-shot learning [74], HM is defined as the harmonic mean of " + }, + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{\\mathrm{base}}" + }, + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{\\mathrm{novel}}" + }, + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "text", + "content": " scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "Datasets. We use Pascal VOC [13] and MS COCO [40] with the same splits defined in FSOD benchmarks [65, 72]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 271, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 547, + 344 + ], + "type": "text", + "content": "On Pascal VOC, three separate base/novel class splits exist, where each one consists of 15 base and 5 novel classes. In each split, we select 5 base classes to mimic novel classes during meta-tuning. On MS-COCO, we select 15 base classes to mimic novel classes in each proxy task, and evaluate the models for the 10-shot and 30-shot settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 347, + 547, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 539 + ], + "type": "text", + "content": "Baselines. We primarily use the MPSR [72] and DeFRCN [53] as our baselines, which are among the best performing fine-tuning based FSOD methods on Pascal VOC. For the DeFRCN experiments, we transfer the meta-tuned loss functions and augmentation magnitudes from MPSR to the DeFRCN method, which are both based on Faster-RCNN. We take the results for FRCN [76], Ret. R-CNN [15], Meta-RCNN [76], FSRW [31], MetaDet [69], FsDetView [75] and ONCE [52] from [15] for a fair comparison. For the MPSR, DeFRCN (seed is set to 0) and FSCE [61], we report the results we obtain experimentally. We take the results for TFA+Hal [84], CME [36], TIP [35], DCNet [27], QA-FewDet [21] FADI [6], LVC [32], KFSOD [83] and FCT [22] from the original papers. Finally, while it is difficult to fairly compare fine-tuning versus meta-learning based approaches, we provide a discussion in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 543, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 547, + 663 + ], + "type": "text", + "content": "Implementation details. We use 200 RL episodes for loss function meta-tuning, with REINFORCE learning rate set to 0.0005. The meta-tuning for augmentation parameter is carried out using the trained and frozen the loss function parameters. We keep the fine-tuning implementation details of MPSR unchanged, which uses 4000 and 8000 gradient descent iterations for 10-shot and 30-shot experiments on MS-COCO, and 2000 iterations on Pascal VOC. We will publish the full source code upon publication; a preliminary version is provided as supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 671, + 388, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 388, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 388, + 682 + ], + "type": "text", + "content": "4.1. Main results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "We first compare the meta-tuning results against the corresponding MPSR baseline in Table 1. In the table, Meta" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7343" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 533, + 178 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 533, + 178 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 533, + 178 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 533, + 178 + ], + "type": "table", + "html": "
Method/ShotPascal VOCMS-COCO
Novel ClassesAll Classes (HM)Novel ClassesAll Classes (HM)
12351012351010301030
MPSR [72]33.137.244.347.152.143.147.454.557.260.89.113.711.515.0
MPSR+Meta-Static33.439.445.147.352.643.750.455.457.561.410.114.812.716.4
MPSR+Meta-Dynamic34.539.845.048.252.545.051.055.558.361.611.914.914.316.6
MPSR+Meta-ScaledDynamic35.240.345.848.452.945.651.255.958.361.812.315.014.416.7
MPSR+Aug34.638.646.048.352.745.149.556.258.462.09.914.912.516.3
MPSR+Meta-Static+Aug35.339.146.148.452.745.949.956.258.361.810.215.212.816.7
MPSR+Meta-Dynamic+Aug35.439.646.548.953.346.050.556.858.962.512.115.314.516.8
MPSR+Meta-ScaledDynamic+Aug35.840.646.849.253.746.351.557.059.262.712.515.414.716.9
", + "image_path": "74f2b3f2e42e398d215c6c4b03511810495f0e31a358a2ebba72301b018b1c7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 185, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 547, + 208 + ], + "type": "text", + "content": "Table 1. FSOD (mAP) and G-FSOD (HM of the base and novel class mAPs) results on Pascal VOC and MS-COCO datasets for MPSR baseline method. HM stands for harmonic mean." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "content": "Static, Meta-Dynamic, Meta-ScaledDynamic refer to meta-tuning a single temperature, dynamic temperature, and novel class scaled dynamic temperature functions, respectively. Similarly, Aug, Meta-Static " + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "content": " Aug, Meta-Dynamic " + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "content": " Aug, and Meta-ScaledDynamic " + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "content": " Aug refer to meta-tuning only augmentation, single temperature and augmentation, dynamic temperature and augmentation, and novel class scaled dynamic temperature and augmentation functions, respectively. We observe that meta-tuning consistently improves the FSOD and G-FSOD results of the MPSR model. We also observe steady improvements gradually from the baseline to Meta-Static, to Meta-Dynamic, and finally to Meta-ScaledDynamic. In addition, the meta-tuned augmentation magnitude parameter also contributes positively to the few-shot object detection performance. The overall consistency of the improvements provides positive evidence for the value of loss and augmentation meta-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 436, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 288, + 592 + ], + "type": "text", + "content": "Pascal VOC results. In Table 2, we report the Pascal VOC results for our MPSR and DeFRCN based Meta-ScaledDynamic+Aug approach and compare them against the state-of-the-art fine-tuning based FSOD methods. While we present the scores averaged over the three splits in this table, additional per-split FSOD and G-FSOD results can be found in the supplementary material. The left side of Table 2 presents the FSOD results for the varying number of support images. We observe that DeFRCN combined with Meta-ScaledDynamic+Aug, i.e. meta-tuning of the score coefficient, dynamic temperature and the augmentation parameter, yields the best mAP scores in all " + }, + { + "bbox": [ + 46, + 436, + 288, + 592 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 436, + 288, + 592 + ], + "type": "text", + "content": "-shot settings among all methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "text", + "content": "The right side of Table 2 presents the G-FSOD results on Pascal VOC. We observe that the best-performing Meta-ScaledDynamic+Aug method improves the HM scores further above the state-of-the-art in all " + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 593, + 288, + 676 + ], + "type": "text", + "content": "-shot settings. Overall, these results suggest that the proposed framework is an effective way for meta-learning inductive biases to be used in fine-tuning-based FSOD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Figure 3 presents visual detection examples without and with meta-tuned scaled dynamic temperature and augmentations in the first and second rows, respectively. We observe" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 228, + 547, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 265 + ], + "type": "text", + "content": "various improvements, such as reductions in false positives, improved recall, and more precise boxes, most likely due to the improved model fitting in the low-data regime." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 270, + 548, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 548, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 548, + 449 + ], + "type": "text", + "content": "MS-COCO results. In Table 3, we compare the MPSR and DeFRCN based Meta-ScaleDynamic+Aug results against other fine-tuning based FSOD methods that report 10-shot and 30-shot results on the MS-COCO dataset. We observe that with meta-tuning, the FSOD scores of MPSR improve from 9.1 to 12.5 (10-shot mAP), and from 13.7 to 15.4 (30-shot mAP). We also observe that the scores of DeFRCN improve from 18.5 to 18.8 (10-shot mAP), and from 21.9 to 23.4 (30-shot mAP), obtaining the best and second best results against all other models. Similarly, in the case of GFSD, with meta-tuning, the 10-shot HM score of DeFRCN improves from 24.0 to 24.4, outperforming all other models. In addition, the 30-shot HM score of DeFRCN improves from 26.8 to 28.0, which is slightly below the 28.1 score of LVC-PL [32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 460, + 405, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 405, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 405, + 473 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 485, + 548, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 548, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 548, + 652 + ], + "type": "text", + "content": "Meta-tuning details. The proposed meta-tuning approach involves three important technical details: Proxy-novel imitation, model re-initialization, and reward normalization. Proxy-novel imitation refers to reinforcement learning over the sampled proxy-novel tasks, instead of the whole training set, to mimic the test-time FSOD challenges. Model re-initialization is the re-initialization of the base model for each task. Without re-initialization, not only the sampled loss/augmentation parameters and tasks but also the accumulated model updates undesirably affect the rewards. Reward normalization further reduces the effect of task difficulty variance by normalizing the rewards obtained within a single episode, allowing a more isolated assessment of the sampled loss functions and augmentations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": "We evaluate the contributions of these three important details in terms of G-FSOD HM scores using the 5-shot setting of Pascal VOC Split-1 with MPSR+Meta-Dynamic. The results averaged over 5 runs are given in Table 4. We observe that each component progressively improves the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7344" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 70, + 479, + 209 + ], + "blocks": [ + { + "bbox": [ + 115, + 70, + 479, + 209 + ], + "lines": [ + { + "bbox": [ + 115, + 70, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 70, + 479, + 209 + ], + "type": "table", + "html": "
Method/ShotNovel ClassesAll Classes (HM)
123510123510
FRCN [76] (ICCV'19)16.120.628.833.436.525.931.740.044.346.7
TFA-fc [65] (ICML'20)27.630.639.846.648.740.544.152.958.359.9
TFA-cos [65] (ICML'20)31.432.640.546.848.344.646.053.558.459.6
FSCE [61] (CVPR'21)29.236.342.547.152.241.848.854.257.761.0
Ret. R-CNN [15] (CVPR'21)31.437.141.446.848.844.750.554.759.160.8
TFA+Hal [84] (CVPR'21)32.935.540.446.348.1-----
FADI [6] (NeurIPS'21)42.246.547.952.456.9-----
LVC [32] (CVPR'22)30.935.443.651.154.1-----
LVC-PL [32] (CVPR'22)45.245.054.857.558.6-----
MPSR [72] (ECCV'20)33.137.244.347.152.143.147.454.557.260.8
DeFRCN [53] (ICCV'21)46.552.655.960.060.857.662.564.767.667.8
MPSR+Meta-ScaledDynamic+Aug35.840.646.849.253.746.351.557.059.262.7
DeFRCN+Meta-ScaledDynamic+Aug49.254.057.261.361.859.863.765.968.668.7
", + "image_path": "0b1b24a6a210f0c6fb18443734c165d5ae6d4be3a6ea534e1cacf72623255f4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 251, + 133, + 322 + ], + "blocks": [ + { + "bbox": [ + 52, + 251, + 133, + 322 + ], + "lines": [ + { + "bbox": [ + 52, + 251, + 133, + 322 + ], + "spans": [ + { + "bbox": [ + 52, + 251, + 133, + 322 + ], + "type": "image", + "image_path": "c84e8ccae23d30cfef10a3a7b73b22f5d6316a4b99065d13a673d414046cfd03.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 323, + 134, + 396 + ], + "blocks": [ + { + "bbox": [ + 52, + 323, + 134, + 396 + ], + "lines": [ + { + "bbox": [ + 52, + 323, + 134, + 396 + ], + "spans": [ + { + "bbox": [ + 52, + 323, + 134, + 396 + ], + "type": "image", + "image_path": "07d07e33e01afe655237ea462d0df27346a0d4cdfbb053f7cfaf3c4d86cc597d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 398, + 114, + 407 + ], + "lines": [ + { + "bbox": [ + 72, + 398, + 114, + 407 + ], + "spans": [ + { + "bbox": [ + 72, + 398, + 114, + 407 + ], + "type": "text", + "content": "Split-1/5-shot" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 251, + 214, + 323 + ], + "blocks": [ + { + "bbox": [ + 134, + 251, + 214, + 323 + ], + "lines": [ + { + "bbox": [ + 134, + 251, + 214, + 323 + ], + "spans": [ + { + "bbox": [ + 134, + 251, + 214, + 323 + ], + "type": "image", + "image_path": "9be8da8655e2408b293b01a02e818e701c1b1257b9293d94df3792dcd139c9bb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 135, + 323, + 214, + 396 + ], + "blocks": [ + { + "bbox": [ + 135, + 323, + 214, + 396 + ], + "lines": [ + { + "bbox": [ + 135, + 323, + 214, + 396 + ], + "spans": [ + { + "bbox": [ + 135, + 323, + 214, + 396 + ], + "type": "image", + "image_path": "33840c689ca2bf073231d0a4381805805015ddb9a7011954efdb97c8067e658a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 398, + 197, + 407 + ], + "lines": [ + { + "bbox": [ + 151, + 398, + 197, + 407 + ], + "spans": [ + { + "bbox": [ + 151, + 398, + 197, + 407 + ], + "type": "text", + "content": "Split-1/10-shot" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 216, + 251, + 295, + 323 + ], + "blocks": [ + { + "bbox": [ + 216, + 251, + 295, + 323 + ], + "lines": [ + { + "bbox": [ + 216, + 251, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 216, + 251, + 295, + 323 + ], + "type": "image", + "image_path": "00cd92d0f5889b6ce2ac6aff11354bb9c79e68d040a21a5b7854fbf4d8bd6304.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 216, + 324, + 295, + 396 + ], + "blocks": [ + { + "bbox": [ + 216, + 324, + 295, + 396 + ], + "lines": [ + { + "bbox": [ + 216, + 324, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 216, + 324, + 295, + 396 + ], + "type": "image", + "image_path": "0ae390992bf7542c471325f10dc035b0f094f06d6f53708c39e863bcfa10e780.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 398, + 275, + 407 + ], + "lines": [ + { + "bbox": [ + 234, + 398, + 275, + 407 + ], + "spans": [ + { + "bbox": [ + 234, + 398, + 275, + 407 + ], + "type": "text", + "content": "Split-2/5-shot" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 416, + 546, + 439 + ], + "lines": [ + { + "bbox": [ + 46, + 416, + 546, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 546, + 439 + ], + "type": "text", + "content": "Figure 3. Qualitative results using MPSR without (first row) and with (second row) meta-tuning, over multiple Pascal VOC splits. Base and novel class detections are shown with green and red boxes, respectively. (Best viewed in color.)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 296, + 251, + 376, + 323 + ], + "blocks": [ + { + "bbox": [ + 296, + 251, + 376, + 323 + ], + "lines": [ + { + "bbox": [ + 296, + 251, + 376, + 323 + ], + "spans": [ + { + "bbox": [ + 296, + 251, + 376, + 323 + ], + "type": "image", + "image_path": "0ba54553e65bf95e2dd0604a7fffd07cf776df0ecce98278d36adc2cc70c8628.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 296, + 324, + 376, + 396 + ], + "blocks": [ + { + "bbox": [ + 296, + 324, + 376, + 396 + ], + "lines": [ + { + "bbox": [ + 296, + 324, + 376, + 396 + ], + "spans": [ + { + "bbox": [ + 296, + 324, + 376, + 396 + ], + "type": "image", + "image_path": "129654a3e58852f10898f12a210faa73be234f73690e28553587a21c7f725414.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 398, + 358, + 407 + ], + "lines": [ + { + "bbox": [ + 312, + 398, + 358, + 407 + ], + "spans": [ + { + "bbox": [ + 312, + 398, + 358, + 407 + ], + "type": "text", + "content": "Split-2/10-shot" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 377, + 251, + 457, + 323 + ], + "blocks": [ + { + "bbox": [ + 377, + 251, + 457, + 323 + ], + "lines": [ + { + "bbox": [ + 377, + 251, + 457, + 323 + ], + "spans": [ + { + "bbox": [ + 377, + 251, + 457, + 323 + ], + "type": "image", + "image_path": "7f6f3319f594f4eda440e44e5e8e8252502128783c0ae46b0187cd0b99fe0c49.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 377, + 324, + 457, + 396 + ], + "blocks": [ + { + "bbox": [ + 377, + 324, + 457, + 396 + ], + "lines": [ + { + "bbox": [ + 377, + 324, + 457, + 396 + ], + "spans": [ + { + "bbox": [ + 377, + 324, + 457, + 396 + ], + "type": "image", + "image_path": "0d4fedf116c18219edb06b9359f08470a8b32ba8b790502595be9e0fb0385995.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 398, + 436, + 407 + ], + "lines": [ + { + "bbox": [ + 395, + 398, + 436, + 407 + ], + "spans": [ + { + "bbox": [ + 395, + 398, + 436, + 407 + ], + "type": "text", + "content": "Split-3/5-shot" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 459, + 251, + 539, + 323 + ], + "blocks": [ + { + "bbox": [ + 459, + 251, + 539, + 323 + ], + "lines": [ + { + "bbox": [ + 459, + 251, + 539, + 323 + ], + "spans": [ + { + "bbox": [ + 459, + 251, + 539, + 323 + ], + "type": "image", + "image_path": "199345f0ee8602ff84564ce16d0aee3eaea222704f924c2543e8fbfb22f742d3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 459, + 324, + 539, + 396 + ], + "blocks": [ + { + "bbox": [ + 459, + 324, + 539, + 396 + ], + "lines": [ + { + "bbox": [ + 459, + 324, + 539, + 396 + ], + "spans": [ + { + "bbox": [ + 459, + 324, + 539, + 396 + ], + "type": "image", + "image_path": "0b61ecf822828a1a2fd411c35799eb9412ec39cd0b45e31200dfc08dc0534076.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 398, + 519, + 407 + ], + "lines": [ + { + "bbox": [ + 474, + 398, + 519, + 407 + ], + "spans": [ + { + "bbox": [ + 474, + 398, + 519, + 407 + ], + "type": "text", + "content": "Split-3/10-shot" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 48, + 457, + 302, + 596 + ], + "blocks": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "type": "text", + "content": "Table 2. FSOD (mAP) and G-FSOD (HM of the base and novel class mAPs) results on Pascal VOC. The best and the second-best results are marked with red and blue. HM stands for harmonic mean." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 457, + 302, + 596 + ], + "lines": [ + { + "bbox": [ + 48, + 457, + 302, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 302, + 596 + ], + "type": "table", + "html": "
Method/ShotsNovel ClassesAll Classes (HM)
10-shot30-shot10-shot30-shot
FRCN [76] (ICCV'19)9.212.512.815.6
FRCN-BCE [65] (ICML'20)6.410.310.916.1
TFA-fc [65] (ICML'20)10.013.415.419.4
TFA-cos [65] (ICML'20)10.013.715.619.8
MPSR [72] (ECCV'20)9.113.711.515.0
FSCE [61] (CVPR'21)10.514.416.020.2
Ret. R-CNN [15] (CVPR'21)10.513.816.620.4
FADI [6] (NeurIPS'21)12.216.1--
DeFRCN [53] (ICCV'21)18.521.924.026.8
LVC [32] (CVPR'22)12.117.817.822.8
LVC-PL [32] (CVPR'22)17.824.522.828.1
MPSR+Meta-ScaledDynamic+Aug12.515.414.716.9
DeFRCN+Meta-ScaledDynamic+Aug18.823.424.428.0
", + "image_path": "308522c7f40d4e6d811db0c0f68c3aaf087f0f0ccbc358c9ed4764085a97d2f6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "HM scores, and the most significant contribution is made by reward normalization, which improves from 62.1 to 63.3. We also observe that reward normalization considerably improves the overall experimental stability. To quantify this" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 331, + 457, + 522, + 511 + ], + "blocks": [ + { + "bbox": [ + 46, + 605, + 288, + 638 + ], + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 638 + ], + "type": "text", + "content": "Table 3. Comparison of Meta-ScaledDynamic results to the fine-tuning based (G-)FSOD methods on the MS-COCO dataset. The best and the second-best results are marked with red and blue." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 457, + 522, + 511 + ], + "lines": [ + { + "bbox": [ + 331, + 457, + 522, + 511 + ], + "spans": [ + { + "bbox": [ + 331, + 457, + 522, + 511 + ], + "type": "table", + "html": "
Proxy-novel imit.Model re-init.Reward norm.HM
XXX61.5
✓XX61.8
✓✓X62.1
✓✓✓63.3
", + "image_path": "f0aae38acf665faaed19c4b22ded6d178baaa1ac95cd2623ab15ff49efd4d697.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 520, + 546, + 575 + ], + "lines": [ + { + "bbox": [ + 304, + 520, + 546, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 546, + 575 + ], + "type": "text", + "content": "Table 4. Evaluation of meta-tuning details. Proxy-novel imitation is the imitation of novel classes using a subset of base classes. Model re-initialization is the re-initialization of the base model at each task. Reward normalization is within-episode normalization of the mAP scores during meta-tuning." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": "observation, we estimate the " + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": " confidence interval over the runs using " + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "inline_equation", + "content": "CI = 1.96\\frac{s}{\\sqrt{n}}" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "inline_equation", + "content": "1.96" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": " are the standard deviation, number of runs, and " + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": "-value, respectively [65]. According to this estimator, the normalization step narrows the confidence interval from " + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "inline_equation", + "content": "\\pm 0.75" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "inline_equation", + "content": "\\pm 0.13" + }, + { + "bbox": [ + 304, + 598, + 547, + 673 + ], + "type": "text", + "content": ", providing a clear improvement in reliability." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Learned loss functions. In Figure 4, we plot the learned loss functions according to the " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " values obtained at the end of the RL process. The upper plot shows the dynamic" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7345" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 86, + 253, + 213 + ], + "blocks": [ + { + "bbox": [ + 62, + 86, + 253, + 213 + ], + "lines": [ + { + "bbox": [ + 62, + 86, + 253, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 86, + 253, + 213 + ], + "type": "image", + "image_path": "398273ef69347b6ecae7cbc7d987dc77d5389ebf546fa8f522214dff745fcfc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 227, + 253, + 356 + ], + "blocks": [ + { + "bbox": [ + 63, + 227, + 253, + 356 + ], + "lines": [ + { + "bbox": [ + 63, + 227, + 253, + 356 + ], + "spans": [ + { + "bbox": [ + 63, + 227, + 253, + 356 + ], + "type": "image", + "image_path": "d30b7a0b618214bfbeeb4284573888ed1e96615f5a9d492058a0a79c92609f25.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 288, + 410 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 288, + 410 + ], + "type": "text", + "content": "Figure 4. The dynamic temperature functions and score scaling coefficients learned by the meta-tuning process, using Meta-Dynamic (upper) and Meta-ScaledDynamic (lower) formulations. Results for each Pascal VOC split is shown with a separate curve." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "type": "text", + "content": "temperature functions learned over three different splits. We observe that temporally attenuated temperature values are preferred consistently, sharpening the predictions towards the end of the fine-tuning process. The lower plot shows the learned dynamic temperature functions with novel class score scaling. The learned scaling coefficients, i.e. " + }, + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "type": "inline_equation", + "content": "\\mu_{\\alpha}" + }, + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "type": "text", + "content": " of the learned " + }, + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "type": "inline_equation", + "content": "\\rho_{\\alpha}" + }, + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "type": "text", + "content": " distribution, are shown as horizontal lines. We observe that similar dynamic temperature functions are learned, and " + }, + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "type": "inline_equation", + "content": "\\mu_{\\alpha}" + }, + { + "bbox": [ + 46, + 433, + 288, + 696 + ], + "type": "text", + "content": " values vary between 1.09 to 1.2, suggesting that the meta-tuning process learns to boost the novel class scores. The interpretability of these outcomes, we believe, highlights a significant advantage of loss meta-tuning. In the context of interpretability, we observe that as the fine-tuning process continues on the few-shot training set, the predictions are progressively made sharper, i.e. the loss becomes more sensitive to classification errors and enforces towards making more confident correct predictions. This is in alignment with one of our original motivations for reducing the dominating classification errors in G-FSOD, as the meta-tuning process automatically learns to enforce more accurate classifications, where the curve steepness and the numerical ranges are learned via RL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "Learned augmentations. The learned photometric augmen" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 538, + 115 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 538, + 115 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 538, + 115 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 538, + 115 + ], + "type": "table", + "html": "
S/MTFA [65]TFA+Hal [84]TFA+Meta-ScaledDynamic+Aug
13.43.84.7
24.65.05.8
36.66.97.1
", + "image_path": "0b37a4a9f7e84081188f5cb9e35582c075198e5d0623cef1732678aead7d2648.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 124, + 547, + 145 + ], + "lines": [ + { + "bbox": [ + 306, + 124, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 124, + 547, + 145 + ], + "type": "text", + "content": "Table 5. Low-shot (1-shot, 2-shot and 3-shot) experiments on MS-COCO dataset with novel classes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 171, + 547, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 171, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 547, + 255 + ], + "type": "text", + "content": "tation magnitude values learned are 0.29, 0.24, 0.13, and 0.36 for Pascal VOC split-1, split-2, split-3, and MS-COCO datasets, respectively. We observe that the learned augmentation magnitudes positively contribute to the performance. According to the results in Table 1, the average Pascal VOC split-1/1-shot score increases from 33.1 to 34.6 with only augmentation steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 261, + 548, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 261, + 548, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 548, + 380 + ], + "type": "text", + "content": "Very low-shot experiments. Finally, we evaluate the meta-tuning approach in low-shot many-class settings. [84] proposes TFA+Hal method that uses the TFA baseline and conducts 1-shot, 2-shot, and 3-shot FSOD on the MS-COCO dataset. As we already observe the positive effects of the loss terms and augmentation magnitudes obtained from the MPSR on the DeFRCN, we similarly apply the learned parameters to the TFA baseline. The results are presented in Table 5. We observe that results are consistently improved using the meta-tuned functions on the TFA baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 396, + 378, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 378, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 378, + 409 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 622 + ], + "type": "text", + "content": "Fine-tuning based frameworks offer simple and reliable approaches to building detection models from few samples. However, a major limitation of the existing fine-tuning-based FSOD models is their focus on the hand-crafting the design of fine-tuning details for few-shot training, which is inherently difficult and likely to be sub-optimal. Towards addressing this limitation, we propose to meta-learn the fine-tuning based learning dynamics as a way of introducing learned inductive biases for few-shot learning. The proposed tuning scheme uses meta-learning principles with reinforcement learning, and obtains interpretable loss functions and augmentation magnitudes for few-shot training. Our comprehensive experimental results on Pascal VOC and MS COCO datasets show that the proposed meta-tuning approach consistently provides significant performance improvements over the strong fine-tuning based few-shot detection baselines in both FSOD and G-FSOD settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 624, + 547, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 624, + 547, + 672 + ], + "spans": [ + { + "bbox": [ + 305, + 624, + 547, + 672 + ], + "type": "text", + "content": "While we restrict our experiments to loss and augmentation functions, meta-tuning other learning components, e.g. initial model, and applications to other few-shot learning problems can be interesting future work directions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the TUBITAK Grant 119E597 and a Google Faculty Research Award." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7346" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, and Frank Wood. Enhancing few-shot image classification with unlabelled examples. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2796-2805, January 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 178 + ], + "type": "text", + "content": "[2] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Sigal. Improved Few-Shot Visual Classification. arXiv e-prints, page arXiv:1912.03432, Dec. 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 213 + ], + "type": "text", + "content": "[3] Luca Bertinetto, Joao F. Henriques, Philip Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Proc. Int. Conf. Learn. Represent., 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 267 + ], + "type": "text", + "content": "[4] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-Shot Segmentation Without Meta-Learning: A Good Transductive Inference Is All You Need? In arXiv:2012.06166 [cs], 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 300 + ], + "type": "text", + "content": "[5] Kaidi Cao, Maria Brbić, and Jure Leskovec. Concept learners for few-shot learning. In Proc. Int. Conf. Learn. Represent., 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 345 + ], + "type": "text", + "content": "[6] Yuhang Cao, Jiaqi Wang, Ying Jin, Tong Wu, Kai Chen, Ziwei Liu, and Dahua Lin. Few-shot object detection via association and discrimination. Proc. Adv. Neural Inf. Process. Syst., 34:16570-16581, 2021. 1, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "text", + "content": "[7] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 8819-8828, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "type": "text", + "content": "[8] Tung-I Chen, Yueh-Cheng Liu, Hung-Ting Su, Yu-Cheng Chang, Yu-Hsiang Lin, Jia-Fong Yeh, Wen-Chin Chen, and Winston Hsu. Dual-awareness attention for few-shot object detection. IEEE Transactions on Multimedia, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 468 + ], + "type": "text", + "content": "[9] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A Closer Look at Few-shot Classification. In ICLR 2019, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 469, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 288, + 502 + ], + "type": "text", + "content": "[10] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 288, + 546 + ], + "type": "text", + "content": "[11] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proc. IEEE Conf. Comput. Vis. Pattern Recog. Workshops, pages 702-703, 2020. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "type": "text", + "content": "[12] Guneet S Dhillon, Pratik Chaudhari, Avinash Ravichandran, and Stefano Soatto. A Baseline for Few-Shot Image Classification. In ICLR, page 20, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "text", + "content": "[14] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4013–4022, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[15] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4527-4536, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[16] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[17] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proc. Int. Conf. Mach. Learn., volume 70, pages 1126–1135, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[18] Golnaz Ghiasi, Yin Cui, A. Srinivas, Rui Qian, Tsung-Yi Lin, Ekin Dogus Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 2917-2927, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "type": "text", + "content": "[19] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proc. IEEE Int. Conf. on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 547, + 308 + ], + "type": "text", + "content": "[20] Santiago Gonzalez and Risto Miikkulainen. Improved training speed, accuracy, and data utilization through loss function optimization. In 2020 IEEE Congress on Evolutionary Computation (CEC), pages 1-8. IEEE, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 547, + 363 + ], + "type": "text", + "content": "[21] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proc. IEEE Int. Conf. on Computer Vision, pages 3263-3272, 2021. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 409 + ], + "type": "text", + "content": "[22] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 5321-5330, 2022. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 410, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 547, + 453 + ], + "type": "text", + "content": "[23] Bharath Hariharan and Ross B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. Proc. IEEE Int. Conf. on Computer Vision, pages 3037-3046, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 455, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 547, + 487 + ], + "type": "text", + "content": "[24] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "type": "text", + "content": "[25] Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. In Proc. Int. Conf. Mach. Learn., pages 2731-2741. PMLR, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 534, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 547, + 567 + ], + "type": "text", + "content": "[26] Ting-I Hsieh, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. One-shot object detection with co-attention and co-excitation. arXiv preprint arXiv:1911.12529, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 568, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 547, + 612 + ], + "type": "text", + "content": "[27] Hanzhe Hu, Shuai Bai, Aoxue Li, Jinshi Cui, and Liwei Wang. Dense relation distillation with context-aware aggregation for few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 10185-10194, 2021. 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 547, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 666 + ], + "type": "text", + "content": "[28] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference. arXiv e-prints, page arXiv:2204.07305, Apr. 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "type": "text", + "content": "[29] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Proc. Adv. Neural Inf. Process. Syst., 33, 2020. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7347" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[30] Taewon Jeong and Heeyoung Kim. Ood-maml: Meta-learning for few-shot out-of-distribution detection and classification. In Advances in Neural Information Processing Systems, volume 33, pages 3907-3916, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "text", + "content": "[31] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In Proc. IEEE Int. Conf. on Computer Vision, pages 8420-8429, 2019. 1, 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "type": "text", + "content": "[32] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Label, verify, correct: A simple few shot object detection method. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 14237-14247, 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 288, + 243 + ], + "type": "text", + "content": "[33] Michalis Lazarou, Yannis Avrithis, and Tania Stathaki. Tensor feature hallucination for few-shot learning. ArXiv, abs/2106.05321, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 245, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 288, + 289 + ], + "type": "text", + "content": "[34] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 10649-10657, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 291, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 288, + 323 + ], + "type": "text", + "content": "[35] Aoxue Li and Zhenguo Li. Transformation invariant few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 3094-3102, 2021. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 326, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 288, + 369 + ], + "type": "text", + "content": "[36] Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. Beyond max-margin: Class margin equilibrium for few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 7363-7372, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "text", + "content": "[37] Chuming Li, Xin Yuan, Chen Lin, Minghao Guo, Wei Wu, Junjie Yan, and Wanli Ouyang. Am-lfs: Automl for loss function search. In Proc. IEEE Int. Conf. on Computer Vision, pages 8410-8419, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 417, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 288, + 450 + ], + "type": "text", + "content": "[38] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. MetaSGD: Learning to Learn Quickly for Few-Shot Learning. arXiv e-prints, page arXiv:1707.09835, July 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 452, + 288, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 288, + 484 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 288, + 484 + ], + "type": "text", + "content": "[39] Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment. Proc. Adv. Neural Inf. Process. Syst., 32, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 288, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 541 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conf. on Computer Vision, pages 740-755. Springer, 2014. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "text", + "content": "[41] Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters: Understanding margin in few-shot classification. arXiv preprint arXiv:2003.12060, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "type": "text", + "content": "[42] Peidong Liu, Gengwei Zhang, Bochao Wang, Hang Xu, Xiaodan Liang, Yong Jiang, and Zhenguo Li. Loss function discovery for object detection via convergence-simulation driven search. arXiv preprint arXiv:2102.04700, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "text", + "content": "[43] Shichen Liu, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Generalized Zero-Shot Learning with Deep Calibration Network. In NeurIPS, pages 2005–2015. 2018. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[44] Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sungju Hwang, and Yi Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In Proc. Int. Conf. Learn. Represent., 2019. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[45] Yan Liu, Zhijie Zhang, Li Niu, Junjie Chen, and Liqing Zhang. Mixed supervised object detection by transferring mask prior and semantic similarity. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Proc. Adv. Neural Inf. Process. Syst., 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "type": "text", + "content": "[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. Proc. IEEE Int. Conf. on Computer Vision, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "text", + "content": "[47] Tsendsumen Munkhdalai, Xingdi Yuan, Soroush Mehri, and Adam Trischler. Rapid adaptation with conditionally shifted neurons. In ICML, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 206, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 547, + 228 + ], + "type": "text", + "content": "[48] Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv: Learning, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 547, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 547, + 260 + ], + "type": "text", + "content": "[49] Boris N. Oreshkin, Pau Rodriguez Lopez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. In NeurIPS, 2018. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "type": "text", + "content": "[50] Matteo Papini, Andrea Battistello, and Marcello Restelli. Balancing learning speed and stability in policy gradient via adaptive exploration. In Proc. Int. Conf. on Artif. Intellig. and Stat., pages 1188-1199, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 306, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 306, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 306, + 545, + 326 + ], + "type": "text", + "content": "[51] Eunbyung Park and Junier B. Oliva. Meta-curvature. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "text", + "content": "[52] Juan-Manuel Perez-Rua, Xiatian Zhu, Timothy M Hospedales, and Tao Xiang. Incremental few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 13846-13855, 2020. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "type": "text", + "content": "[53] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In Proc. IEEE Int. Conf. on Computer Vision, pages 8681-8690, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 416, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 448 + ], + "type": "text", + "content": "[54] Aravind Rajeswaran, Chelsea Finn, Sham M. Kakade, and Sergey Levine. Meta-learning with implicit gradients. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "type": "text", + "content": "[55] Esteban Real, Chen Liang, David So, and Quoc Le. Automlzero: Evolving machine learning algorithms from scratch. In Proc. Int. Conf. Mach. Learn., pages 8007-8019. PMLR, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 493, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 547, + 536 + ], + "type": "text", + "content": "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Proc. Adv. Neural Inf. Process. Syst., 28:91-99, 2015. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "type": "text", + "content": "[57] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, MingYu Liu, Yong Jae Lee, Alexander G. Schwing, and Jan Kautz. Instance-aware, context-focused, and memoryefficient weakly supervised object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "type": "text", + "content": "[58] Andrei A. Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In Proc. Int. Conf. Learn. Represent., 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 637, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 637, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 637, + 547, + 680 + ], + "type": "text", + "content": "[59] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In Proc. Int. Conf. Mach. Learn., volume 48, pages 1842-1850, 2016. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "text", + "content": "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Proc. Adv. Neural Inf. Process. Syst., volume 30, 2017. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7348" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "type": "text", + "content": "[61] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 7352-7362, 2021. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "text", + "content": "[62] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 1199-1208, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "type": "text", + "content": "[63] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B. Tenenbaum, and Phillip Isola. Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need? In Proc. European Conf. on Computer Vision, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 288, + 253 + ], + "type": "text", + "content": "[64] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proc. Adv. Neural Inf. Process. Syst., volume 29, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 288, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 288, + 298 + ], + "type": "text", + "content": "[65] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. arXiv preprint arXiv:2003.06957, 2020. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 301, + 288, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 301, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 48, + 301, + 288, + 344 + ], + "type": "text", + "content": "[66] Xiaobo Wang, Shuo Wang, Cheng Chi, Shifeng Zhang, and Tao Mei. Loss function search for face recognition. In Proc. Int. Conf. Mach. Learn., pages 10029-10038. PMLR, 2020. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 347, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 347, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 48, + 347, + 288, + 390 + ], + "type": "text", + "content": "[67] Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning. arXiv preprint arXiv:1911.04623, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "type": "text", + "content": "[68] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-Shot Learning from Imaginary Data. arXiv:1801.05401 [cs], Jan. 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 428, + 288, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 288, + 460 + ], + "type": "text", + "content": "[69] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Meta-learning to detect rare objects. In Proc. IEEE Int. Conf. on Computer Vision, pages 9925-9934, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 463, + 288, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 288, + 495 + ], + "type": "text", + "content": "[70] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 288, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 288, + 540 + ], + "type": "text", + "content": "[71] Among Wu, Suqi Zhao, Cheng Deng, and Wei Liu. Generalized and discriminative few-shot object detection via svddictionary enhancement. Proc. Adv. Neural Inf. Process. Syst., 34:6353-6364, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "text", + "content": "[72] Jiaxi Wu, Songtao Liu, Di Huang, and Yunhong Wang. Multiscale positive sample refinement for few-shot object detection. In Proc. European Conf. on Computer Vision, pages 456-472. Springer, 2020. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "type": "text", + "content": "[73] Xiongwei Wu, Doyen Sahoo, and Steven Hoi. Meta-rcnn: Meta learning for few-shot object detection. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1679–1687, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "text", + "content": "[74] Yongqin Xian, Bernt Schiele, and Zeynep Akata. Zero-shot learning-the good, the bad and the ugly. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 4582-4591, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[75] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In Proc. European Conf. on Computer Vision, pages 192-210. Springer, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 464 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[76] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proc. IEEE Int. Conf. on Computer Vision, pages 9577-9586, 2019. 1, 2, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[77] Huaxiu Yao, Linjun Zhang, and Chelsea Finn. Meta-learning with fewer tasks through task interpolation. In Proceeding of the 10th International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 205 + ], + "type": "text", + "content": "[78] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 8805-8814, 2020. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 547, + 251 + ], + "type": "text", + "content": "[79] Li Yin, Juan M Perez-Rua, and Kevin J Liang. Sylph: A hypernetwork framework for incremental few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 9035-9045, 2022. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 253, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 547, + 296 + ], + "type": "text", + "content": "[80] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., June 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[81] Gongjie Zhang, Zhipeng Luo, Kaiwen Cui, Shijian Lu, and Eric P Xing. Meta-detr: Image-level few-shot detection with inter-class correlation exploitation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 547, + 386 + ], + "type": "text", + "content": "[82] Lu Zhang, Shuigeng Zhou, Jihong Guan, and Ji Zhang. Accurate few-shot object detection with support-query mutual guidance and hybrid loss. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 14424-14432, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "type": "text", + "content": "[83] Shan Zhang, Lei Wang, Naila Murray, and Piotr Koniusz. Kernelized few-shot object detection with efficient integral aggregation. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 19207-19216, 2022. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 432, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 547, + 464 + ], + "type": "text", + "content": "[84] Weilin Zhang and Yu-Xiong Wang. Hallucination improves few-shot object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 13008-13017, 2021. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7349" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_content_list.json b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..070b6887f0b3a1d0baf1aaf51cdb1f6adb3031cd --- /dev/null +++ b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_content_list.json @@ -0,0 +1,1689 @@ +[ + { + "type": "text", + "text": "MetaCLUE: Towards Comprehensive Visual Metaphors Research", + "text_level": 1, + "bbox": [ + 151, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arjun R. Akula*, Brendan Driscoll*, Pradyumna Narayana, Soravit Changpinyo, Zhiwei Jia, Suyash Damle, Garima Pruthi, Sugato Basu, Leonidas Guibas, William T. Freeman, Yuanzhen Li, Varun Jampani* Google", + "bbox": [ + 166, + 180, + 802, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/209a120a3d720efd6732ac818466b1da66b20f7edd8f48d41c73de8e0c48d583.jpg", + "image_caption": [ + "Metaphor: Killing trees is as harmful as killing wildlife." + ], + "image_footnote": [], + "bbox": [ + 84, + 290, + 349, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Classification", + "text_level": 1, + "bbox": [ + 418, + 290, + 549, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Is this a visual metaphor? - YES - NO", + "bbox": [ + 357, + 313, + 599, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Localization", + "text_level": 1, + "bbox": [ + 424, + 354, + 542, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detect image regions that invoke the concepts:", + "bbox": [ + 356, + 375, + 593, + 404 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Killing trees", + "- killing wildlife" + ], + "bbox": [ + 356, + 410, + 459, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a639cb28fbb5f4dd8079fbda6c71370593e24e033a72a38e4a7efb389868de45.jpg", + "image_caption": [ + "Figure 1. With MetaCLUE, we introduce several interesting tasks related to visual metaphors. We collect metaphor annotations (objects, abstract concepts, relationships and object boxes) for evaluating existing models on these tasks. Specifically we perform a comprehensive evaluation of vision and language models on four different tasks (Classification, Localization, Understanding, and gEneration). Comprehensive experiments in this work show that state-of-the-art techniques mostly focus on literal interpretation and perform poorly in understanding and generation of metaphor images." + ], + "image_footnote": [], + "bbox": [ + 464, + 388, + 599, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gEneration", + "text_level": 1, + "bbox": [ + 697, + 292, + 803, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt: \"An advertisement where killing trees is as harmful as killing wildlife.\"", + "bbox": [ + 622, + 314, + 852, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6401876215261697f19124b169d600cb701db81e91126ced2a0c158e7de52e1f.jpg", + "image_caption": [ + "Stable Diffusion" + ], + "image_footnote": [], + "bbox": [ + 630, + 364, + 743, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ccc5ac901c561ad3aca7d3018ac7e221c233c7aa31b53e61fded73e4225eae9.jpg", + "image_caption": [ + "Imagen" + ], + "image_footnote": [], + "bbox": [ + 758, + 364, + 870, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding", + "text_level": 1, + "bbox": [ + 411, + 479, + 555, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Retrieval", + "bbox": [ + 91, + 500, + 155, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pick the right one:", + "bbox": [ + 91, + 516, + 212, + 529 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Killing the forest is as deadly as killing the animals too.", + "(b) Birds is as much a part of our world as used cans." + ], + "bbox": [ + 91, + 532, + 323, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Captioning", + "bbox": [ + 334, + 500, + 413, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sample predictions:", + "bbox": [ + 334, + 516, + 460, + 530 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Deforestation is as damaging as killing wildlife.", + "2. Deforestation is as bad as ending the death penalty." + ], + "bbox": [ + 334, + 532, + 593, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Question Answering Sample Questions:", + "bbox": [ + 620, + 500, + 807, + 530 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q. What is as harmful as killing wildlife?", + "Q. What is compared to killing wildlife?" + ], + "bbox": [ + 619, + 532, + 879, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 685, + 313, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creativity is an indispensable part of human cognition and also an inherent part of how we make sense of the world. Metaphorical abstraction is fundamental in communicating creative ideas through nuanced relationships between abstract concepts such as feelings. While computer vision benchmarks and approaches predominantly focus on understanding and generating literal interpretations of images, metaphorical comprehension of images remains relatively unexplored. Towards this goal, we introduce MetaCLUE, a set of vision tasks on visual metaphor. We also collect high-quality and rich metaphor annotations (ab", + "bbox": [ + 75, + 707, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "struct objects, concepts, relationships along with their corresponding object boxes) as there do not exist any datasets that facilitate the evaluation of these tasks. We perform a comprehensive analysis of state-of-the-art models in vision and language based on our annotations, highlighting strengths and weaknesses of current approaches in visual metaphor classification, localization, understanding (retrieval, question answering, captioning) and generation (text-to-image synthesis) tasks. We hope this work provides a concrete step towards developing AI systems with human-like creative capabilities. Project page: https://metaclue.github.io", + "bbox": [ + 496, + 686, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23201", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 209, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"Metaphor is pervasive in everyday life ... Our ordinary conceptual system, in terms of which we both think and act, is fundamentally metaphorical in nature.\" — Lakoff & Johnson [25]", + "bbox": [ + 109, + 114, + 437, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Creativity is a process of generating a new perspective on a problem or a situation. Metaphorical thinking has been recognized as a key and powerful mechanism of creativity [24,29,50]. Humans engage metaphors in their creative thinking process as strategies to link or blend concepts, or to view a concept from a target domain in terms of another, apparently dissimilar concept from a source domain [25]. metaphors also provide a sophisticated tool for nuanced human communication. Let us take a closer look at the structure of metaphors – and especially visual metaphors.", + "bbox": [ + 75, + 190, + 468, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Metaphors1 are a cognitive construct in which a concept is compared to a seemingly unrelated concept via some shared attribute. Take as an example 'This car is a cheetah', where 'This car' is compared to 'a cheetah' in terms of speed. Metaphors have a simple syntactic structure of 'A is B' where A is referred to as the primary concept and B as the secondary concept. The implied analogy in a metaphor is of the form: ' (primary concept) is as (relationship)2 as (secondary concept)' and often involves an attribute transfer from the secondary to the primary concept. Some examples include 'This phone is as fast as a rocket', 'Cigarettes are as harmful as bullets' etc. The primary and secondary concepts are usually unrelated at a glance, resulting in an element of surprise and creativity in metaphorical expressions. Despite following such simple structure, metaphors are quite powerful in conveying creative ideas. Metaphors are pervasive in all forms of communication, such as speech, text, visual etc.", + "bbox": [ + 75, + 340, + 470, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Metaphors are images where the primary and secondary concepts are visually depicted in an image conveying the metaphorical message to the viewers. Visual metaphors are widely used in mass media communications like advertising and journalism [15, 43, 45]. In this work, we work with Ad images, as metaphors tend to be prevalent in ads. There are numerous ways a metaphor can be represented visually. Following the classification in [15], there are at least 4 different types of visual metaphors. Fig. 2 shows sample images that belong to these types along with our annotations of primary, secondary concepts and their relationship. In contextual metaphors, either the primary or secondary concept is not explicitly visible, but is inferred from the context (e.g., apple in the left-most image). In hybrid metaphors, the primary and secondary concepts are", + "bbox": [ + 75, + 613, + 468, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visually conflated. Juxtaposition forms one of the simplest visual metaphor types, where the two concepts are just presented next to each other. Multimodal metaphors represent one of the concepts with another modality, such as text or logo. In practice, visual metaphors use multiple of these strategies to convey a metaphor in an effective manner. In many cases, the implied metaphorical meaning is somewhat open-ended. Interpretation of visual metaphors depends on several external factors, such as familiarity with the brands and cultural context.", + "bbox": [ + 500, + 90, + 890, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These visual variations and nuances make automatic cognition or generation of visual metaphors highly challenging. While the last decade has seen rapid progress in many areas of understanding and generation tasks, prior works in computer vision focus heavily on literal interpretation of images and overlook the importance of metaphorical reasoning in understanding the image message [1, 48]. We believe that developing AI systems with metaphorical comprehension and generation capabilities can greatly assist humans in creative endeavors involving conveying concepts in new and exciting ways. Such systems provide an important step towards conferring human-like creativity to AI models.", + "bbox": [ + 496, + 242, + 890, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce multiple interesting tasks and construct metaphor annotations that enable comprehensive research on visual metaphors. As metaphors are more common in visual Ads, we start with the Pitt's Ads dataset images [20] and then perform a rigorous multi-stage annotation process with expert annotators to filter metaphorical images, add metaphor annotations, and perform additional validation steps to clean the annotations3. While there is recent work making advances in understanding non-literal interpretations in natural language research [7, 12], this work proposes the first step towards metaphor analysis in images.", + "bbox": [ + 496, + 424, + 890, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, we perform comprehensive evaluations with state-of-the-art techniques on four sets of tasks, which we call MetaCLUE: 1. Classification: This is binary classification task of estimating whether a given image contains a metaphor or not. In other words, Are visual features indicative of whether there exists a metaphor in a given image or not?. 2. Localization: Here, the task is to localize the image regions that invoke the primary and secondary concepts in the viewer. This is similar to a standard object detection task, but is more complicated in the case of visual metaphors as the primary/secondary concepts may not be explicitly present in an image. 3. Understanding: Can our models understand the metaphorical message in a given image? We pose this understanding problem as 3 tasks where we can quantitatively measure the performance: Retrieval, Captioning and Visual question answering. 4. gEneration: Can we generate an image that conveys the metaphor, given the metaphorical message as a text prompt?", + "bbox": [ + 496, + 590, + 890, + 863 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3The Ads, while useful for the purposes of our paper, some images may perpetuate harmful stereotypes according to characteristics such as gender.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1. Grammarians distinguish a metaphor \"A is B\" from a simile \"A is like B\". In our work we use \"metaphor\" to encompass both variants.", + "bbox": [ + 76, + 849, + 467, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2We use the word 'relationship' to denote the shared property of primary and secondary concepts, usually adjectives or adjectival phrases.", + "bbox": [ + 76, + 875, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23202", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fbd965311b0c74e4ca2dd41001132f6b7d4a01a2b7a6234d78a30650f44e892d.jpg", + "image_caption": [ + "Contextual", + "This beer is as tasty as a real apple." + ], + "image_footnote": [], + "bbox": [ + 96, + 104, + 282, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aa12fc950b3973b9820cec40425dcc1988ea0c7bdb1e4c4afaa801b13b84dfd5.jpg", + "image_caption": [ + "This car is as adventurous as a space ship." + ], + "image_footnote": [], + "bbox": [ + 295, + 108, + 473, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0861d88e2a3d360853e8c299b9877fc03857a240952a8effc8420dd5c0929bd4.jpg", + "image_caption": [ + "Hybrid", + "Driving this SUV is as smooth as birds flying in the sky." + ], + "image_footnote": [], + "bbox": [ + 491, + 106, + 676, + 202 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/32d41a06a69b86d55b19d50bccf4b6b508bbadc17eee7aeca30e1db324fb7627.jpg", + "image_caption": [ + "This pencil is as red as a fire truck." + ], + "image_footnote": [], + "bbox": [ + 683, + 106, + 870, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fba8eda3cf31a78df732cc054f9bdd677dfe2575d1766b800aa5f980c5547ec6.jpg", + "image_caption": [ + "This chocolate bar is as rich as gold.", + "This car is as made for the beach as a crab.", + "Figure 2. Sample Visual Metaphors with their Annotations. There are different types of visual metaphors. The type depends on how the primary and secondary concepts are visually depicted. Here are sample Ad images from [20] where we annotated the primary concept, secondary concept and their relationship." + ], + "image_footnote": [], + "bbox": [ + 98, + 247, + 282, + 347 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fc19b684838be1c1a648bc9ae41f9b69669726d22fad4e4302152277af5a54b5.jpg", + "image_caption": [ + "Juxtaposition" + ], + "image_footnote": [], + "bbox": [ + 290, + 246, + 477, + 339 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/360d096004498d375b4996e1c2b82bf050d23ce8ce98d5ff58237b8e06e194b7.jpg", + "image_caption": [ + "These donuts are as unique as as talking people." + ], + "image_footnote": [], + "bbox": [ + 491, + 248, + 671, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/57232ccf52b7920f5087f9ef19437801071cf4ba62848cd162eb574cd3373a43.jpg", + "image_caption": [ + "Multimodal", + "The car is as rugged as this muddy trailer." + ], + "image_footnote": [], + "bbox": [ + 683, + 247, + 867, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We comprehensively evaluate existing state-of-the-art techniques for each of these tasks on our collected metaphor annotations. We evaluate the models both in a zero-shot manner as well as with finetuning on our annotations. Even though finetuning resulted in some improvements, most models struggle to produce satisfactory results in many cases, demonstrating the difficulty of these tasks. Our experiments highlight several strengths and weaknesses of the existing techniques on comprehending and generating visual metaphors, providing a concrete first step towards further AI research on this fascinating topic.", + "bbox": [ + 75, + 452, + 472, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 633, + 218, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Creativity and Metaphorical Abstraction. Creativity often involves an innovative fusion of objects, attributes, or relationships from previous knowledge to generate new concepts [5, 52]. Metaphors can serve as an invaluable tool for expressing creative insights and also to stimulate new ones [14, 21]. The cognitive research community has made initial attempts in understanding different realizations of metaphors such as language metaphors and visual metaphors [15, 16, 32]. Studies of visual persuasion show that visual metaphors may be more effective than language metaphors in terms of producing a greater degree of cognitive creativity [34]. In addition to improving creativity, metaphors are also known to elicit pleasure since the initial ambiguity in (re-)conceptualizing a target entity in terms of a source stimulus generates interest and motivation, and the subsequent resolution is rewarding – explaining the impor", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tance of using creative metaphorical processes in art, advertising, and marketing [6,30,38].", + "bbox": [ + 498, + 452, + 890, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Metaphors in Language Research. Computational Linguistic studies show that metaphors are ubiquitous in language, occurring once per three sentences on average [40, 44, 49]. Recently, an increasing number of research efforts have explored the limitations and challenges in detecting and decoding the meaning of language metaphors [7, 48]. While there exist several computational models [9, 12, 17, 33, 35, 45, 47, 56] for metaphor identification, interpretation and generation in language, there exists very little work on computational modeling of visual metaphors.", + "bbox": [ + 496, + 487, + 892, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Metaphors in Computer Vision. Much of computer vision literature is focused on understanding and generating literal images. Automatic metaphorical interpretations of images is highly challenging and requires multi-faceted cognitive reasoning that involves visual reasoning coupled with the use of external knowledge. Recent work on affective captioning using the ArtEmis data set [1] includes some captions that evoke metaphors to explain emotions, but that dataset is focused on visual art and does not specifically consider metaphors. There is no explicit prior work for comprehensive evaluation and development of models that can automatically comprehend or generate visual metaphors. At the same time, there exist several studies that demonstrate the potential of visual metaphors. For instance, some studies [4,37] suggest that advertisements containing metaphorical images will be more persuasive compared to ads with language metaphors, non-metaphorical", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cd66b94d5c2c2b6dead4a5348fa6d5db897abec5c905b7e5548545e2f80e0212.jpg", + "image_caption": [ + "Figure 3. Distribution of Topics in our annotated metaphorical images from Pitt's Ad dataset." + ], + "image_footnote": [], + "bbox": [ + 127, + 88, + 419, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ads, or literal images. There have been some prior computational models [11, 22] for metaphor generation, but they are not thoroughly validated against any benchmark datasets or user studies. Some works [20, 54] propose datasets and techniques for general Ad image understanding with a focus on the challenging aspects of non-literal interpretations in Ad images. However they do not explicitly collect any metaphorical annotations, nor do they provide the corresponding analysis. In this work, we start with an existing Ad image dataset [20] and perform extensive human studies to filter metaphorical images and collect detailed annotations accounting various aspects of metaphoric interpretation. Prior works such as Multi-MET [57] and MET-Meme [53], propose metaphorical annotations but does not annotate the relationship between primary and secondary. In MetaCLUE, in addition to providing relationship annotations, we also collect detailed bounding box annotations that help localize the image regions invoking the primary, secondary concepts in the viewer. We further provide VQA style question and answers.", + "bbox": [ + 75, + 314, + 472, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. MetaCLUE", + "text_level": 1, + "bbox": [ + 76, + 630, + 200, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce four different high-level tasks in MetaCLUE that enable comprehensive evaluation and development of visual metaphor research: Classification, Localization, Understanding and gEneration. In the rest of this section, we first describe our annotation collection process for the tasks and next provide the analysis of using existing state-of-the-art techniques for each of these tasks.", + "bbox": [ + 75, + 655, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Metaphor Classification", + "text_level": 1, + "bbox": [ + 76, + 770, + 297, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the fundamental vision task of image classification, we first ask: Can we develop models that can classify whether or not a given image contains a visual metaphor? In other words, is it possible to just use visual cues to estimate whether or not there exists a metaphorical interpretation of an image?", + "bbox": [ + 75, + 794, + 468, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Annotations. For this task, we need to label whether or", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "not each image contains a metaphor. Since we tend to see more metaphorical images in Ads, we start with images from an existing Ads dataset published by the University of Pittsburgh [20] and manually annotate whether a given Ad image contains a metaphor or not. Pitt's Ads dataset contains images of both product ads (e.g. phone ads) as well as public service announcements (e.g. forest conservation ads). Concretely, to make annotations more efficient, we use a subset of 8.5K Ad images from this dataset that are annotated to have 'symbolic' (fun, adventurous, etc.) references. We find this subset to contain a considerable portion of metaphors. Specifically, we collect 5 Yes/No annotations for each of these 8K images from different annotators. To obtain high-quality annotations, we explained what a visual metaphor is to the annotators and also conducted qualifying exams to pick final annotators for this task. We consider the images with 3 or more 'Yes' annotations as visual metaphors and the remaining as non-metaphorical. At the end, we identify 5061 images containing visual metaphors out of 8480 images. Pitt's Ads dataset [20] also comes with topic annotations (e.g., restaurant, car, animal rights etc.). Fig. 3 shows the word cloud plot of different topics in these 5061 metaphorical images. We split the metaphorical images into 3730 train and 1331 test images by maintaining the same distribution of topics in both the splits. We have two types of negative sets (non-metaphorical images) for classification experiments. One is formed by the remaining 3419 non-metaphorical images in the symbolic set, and another one is created separately by annotating an additional 3000 literal (and non-symbolic) images from Pitt's Ads dataset [20]. We add 2000 of 3419 symbolic negative images to our train split, and add the remaining 1419 images to our test split. Similarly 2000 of 3000 literal negatives are used in training, and the remaining 1000 are used for testing.", + "bbox": [ + 496, + 90, + 893, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation and Results. Using our collected metaphor images, we evaluate the performance of the following state-of-the-art models in classifying an input image as metaphor: EfficientNet [46] and Vision Transformer (ViT) [13]. We fine-tune these models to classify metaphor vs. symbolic/non-metaphors and metaphor vs. literal, and use $20\\%$ of corresponding train splits for validation. Test results are reported in Table 1. Although the performance of ViT is significantly better than EfficientNet and random baselines, there is still ample room for improvement. In particular, models find it easier to distinguish metaphor images from literal images, and struggle to identify metaphors within the symbolic image pool.", + "bbox": [ + 496, + 621, + 893, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Metaphor Understanding", + "text_level": 1, + "bbox": [ + 500, + 830, + 732, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now describe how we collect annotations that help in capturing the metaphorical message from the images collected in previous section.", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23204", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0dbaeae6207fe8bcc0a13f5e333b7b23c047d2e05b2a720db3cb1cd31d2a93d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model#ParamsSymbolic Neg.Literal Neg.
ValTestValTest
RandomN/A63.1051.6060.6657.12
EfficientNet-B05.3M60.7649.6770.9450.30
EfficientNet-B766M61.4448.5469.8449.82
ViT-B/1686M69.3166.9884.0481.24
ViT-L/16307M65.8360.6581.4580.52
", + "bbox": [ + 94, + 88, + 452, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Accuracy of Metaphor Classification (binary classification accuracy) using state-of-the-art classification architectures of EfficientNet [46] and ViT [13].", + "bbox": [ + 76, + 204, + 470, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotations. We provide detailed instructions and several examples to the annotators to help them annotate primary and secondary concepts in the metaphor and also the characteristic/relationship that is transferred from secondary to the primary. We conduct multiple pilot studies to reduce the noise and to improve inter-annotator agreement. Enforcing the annotators to make sure that their annotations are linguistically readable in the following syntactic structure helped us in improving quality and consistency of the annotations: “__ is as __ as __,” where the first blank is the primary concept, the second blank is the relationship, and the third blank is the secondary concept. Figure 2 shows some examples of these annotations. We collect 5 metaphor annotations for each image. As interpretations of metaphorical images can be highly subjective, there can exist more than one interpretation for each image, which makes it difficult to automatically remove noisy annotations. Therefore we conduct an additional human study where we show each of the annotation to five annotators and ask them to verify the correctness along three dimensions: (a) Is the grammar correct?; (b) Are primary and secondary concepts correct?; and (c) Is the relationship correct?. We remove annotations with a low number of votes out of 5 along each of the three dimensions, resulting in a total of over 26k clean annotations. We evaluate state-of-the-art models in understanding metaphorical message from the input images using 3 tasks namely, Retrieval, Captioning and Visual question answering.", + "bbox": [ + 75, + 286, + 468, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Retrieval. The goal of this task is to retrieve the correct metaphor interpretation/statement from a candidate set given an image. In our candidate set, we choose exactly one positive (correct) metaphorical statement from its ground truth messages and uniformly sample $K - 1$ random negative statements from other images. Table 2 shows the results obtained with CLIP [39] and ALBEF [26] for $K = 50$ , reporting retrieval precision@1 and rank4. Although CLIP ViT-L/14 shows good zero-shot performance on random negatives with more than $76\\%$ accuracy, we observe a large drop", + "bbox": [ + 75, + 712, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ca14d96a149751ef91b3f5bcb6263b959bd8dd54231fe77feafd9042acb5e98.jpg", + "image_caption": [ + "Metaphor: Smoking cigarettes is as life-shortening as sharpening a pencil" + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 694, + 193 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Captioning Result Prediction:Smoking is as dangerous as burning a pencil.", + "Visual Question Answering Result Q1: What is a smoker's life compared to a sharpened pencil?", + "Pred: Sharp GT: Shortened Q2: What is used as a visual metaphor for a sharpened pencil?", + "Pred: Not smoking GT:Smoker's life" + ], + "bbox": [ + 699, + 92, + 883, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18d9bca6907411565d11d98da02daafa75f25b1aa588b96755fde662c2f44e7d.jpg", + "image_caption": [ + "Metaphor: Smoking is as deadly as injecting drugs.", + "Figure 4. Results of PaLI [10] for captioning and visual question answering on sample images in our test split." + ], + "image_footnote": [], + "bbox": [ + 506, + 224, + 696, + 330 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Captioning Result Prediction: Drinking and driving is as bad as injecting drugs.", + "Visual Question Answering Result Q1: What is a cigarette as compared to injecting drugs?", + "Pred: Scary GT: Deadly Q2: What is used as a visual metaphor for injecting drugs?", + "Pred:Drinking GT:Cigarette" + ], + "bbox": [ + 699, + 228, + 879, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in performance as we increase $K$ to $\\{100, 500, 1000\\}$ . We further fine-tune CLIP models using $70\\%$ of metaphor annotations as the train set and see gains by up to $+7$ absolute points in $p@1$ . In summary, the performance of models is impressive with less than 50 negative candidates whereas the performance drops greatly by increasing $K$ .", + "bbox": [ + 496, + 424, + 890, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We hypothesize that models might simply be looking at salient objects rather than comprehending the underlying semantics of metaphor in finding the correct candidate. To test this, we mine hard negative (HN) statements and use them as our candidate set. Specifically, we construct the following five types of HNs: (a) Neg Prim: candidates obtained by replacing the primary concept in the metaphor statement with the primary concept from another image; (b) Neg Sec: replacing secondary concept likewise; (c) Neg Prim+Rel: replacing primary and relationship; (d) Neg Sec+Rel: replacing secondary and relationship; (e) Swap Prim&Neg: swapping primary and secondary from the same image. Table 2 shows the results. We see a significant drop of up to $30\\%$ by using HNs as negative statements for $K = 50$ , indicating the difficulty in comprehending and distinguishing metaphorical abstraction of concepts. We find performance of Neg Prim is significantly higher than Neg Sec, suggesting that the models tend to rely more on primary object than secondary object in identifying correct interpretation. Overall, there is ample room for improvement indicated by the steep drop in model performance with HNs.", + "bbox": [ + 496, + 517, + 892, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Captioning. Here, we propose metaphor image cap", + "bbox": [ + 500, + 837, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Rank measures the averaged ranking value of the highest-ranked ground-truth statement with 1 being the highest possible rank.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "5We swap objects with images having different topic to make sure that the generated HNs are actually negatives.", + "bbox": [ + 500, + 862, + 890, + 887 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "With swapping $K$ is always 2.", + "bbox": [ + 517, + 887, + 687, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23205", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2326bfcc24d2af8ca28cf2546fd5be9da9166cc881f724e9117598dbec0422bc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelRandom NegNeg PrimNeg SecNeg Prim+RelNeg Sec+RelSwap accuracy↑
p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓
CLIP (ViT-B/16)70.973.4946.6710.1138.1413.9749.368.8042.2511.6940.61
CLIP (ViT-B/32)61.784.1938.7411.6033.2014.8439.7910.2936.2712.7141.28
CLIP (ViT-L/14)76.663.1751.759.2239.8613.1954.747.7045.9911.0243.08
ALBEF39.798.5727.0016.5329.3115.7926.7715.5828.4214.5946.67
ALBEF (MSCOCO)44.877.5531.7815.1331.4114.0932.5314.0433.2812.8548.24
ALBEF (Flickr30k)47.498.7735.6014.5635.2213.6636.3513.5836.1212.7949.81
FT CLIP (ViT-B/16)76.812.2549.819.6545.4710.2453.408.0750.638.1844.65
FT CLIP (ViT-B/32)68.062.8243.0010.8239.6411.2344.729.3943.459.3545.69
FT CLIP (ViT-L/14)81.751.7857.667.4849.069.4061.255.9957.067.3343.75
", + "bbox": [ + 99, + 88, + 870, + 250 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/62632c3d908e6212262e70a23088be00e58edcd04dc3df7aa03e97c4d0e925c4.jpg", + "table_caption": [ + "Table 2. Performance of retrieval models on $K$ random (column 2) and hard negative candidates (columns 3-7) ( $K = 50$ for columns 2-6, $K = 2$ for last column)." + ], + "table_footnote": [], + "table_body": "
CaptioningAccBLEU4ROUGE-LMETEORSPICECIDErVQAAccBLEU4ROUGE-LMETEORSPICECIDEr
Whole caption1.1%0.2540.5360.2200.1861.076All questions19.9%0.3290.2860.2490.1850.851
Primary29.9%0.3270.4070.3380.2440.931Primary21.5%0.2910.3480.2770.2901.099
Secondary13.7%0.2490.3070.1930.1550.550Secondary12.8%0.2380.2320.1810.2340.735
Relationship23.6%0.4850.2030.2260.0280.296Relationship25.6%0.4490.2750.2850.0380.706
", + "bbox": [ + 86, + 314, + 883, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Metaphorical image captioning (left) and visual question answering (right) performance of PaLI [10]. We report different metrics (higher the better) using both the exact match accuracy (Acc) and standard text generation metrics. For image captioning, we evaluate the whole predicted caption as well as parsed primary object, secondary object, and relationship. For VQA, we evaluate on all predicted answers as well as provide the breakdown for each question type.", + "bbox": [ + 75, + 401, + 892, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tioning task, where the input is an image and the target is the metaphorical message in the syntactic structure is as as . We fine-tune and evaluate the state-of-the-art literal image caption model PaLI-17B [10] based on the exact match accuracy (maximum over all references) and standard metrics for image captioning BLEU4 [36], ROUGE-L [28], METEOR [3], SPICE [2], and CIDEr [51]. Since the target captions follow the fixed syntactic structure, we parse each predicted caption into the primary concept, the secondary concept, and their relationship, and use the same set of metrics for the whole caption to evaluate. Table 3 (left) summarizes the results and Figure 4 provides sample qualitative results. We observe that PaLI generally struggles on this task. For instance, it achieves a CIDEr score of 1.076, compared to 1.491 for state-of-the-art literal image captioning on the popular COCO-Captions [8]. Further, the model struggles with predicting the target relationship when the metrics favor recall (e.g., ROUGE-L) and with predicting the target secondary concept when the metrics favor precision (e.g., BLEU).", + "bbox": [ + 75, + 497, + 472, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Question-Answering (VQA). We propose metaphorical open-ended (i.e., not vocab-based) VQA task, where the input is an image and a given question, and", + "bbox": [ + 75, + 832, + 470, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the target is the answer. We use fixed templates to generate 2 VQA questions whose answer is the primary concept, 2 for secondary concept, and 2 for relationship. Again, we fine-tune and evaluate the state-of-the-art literal VQA model PaLI-17B [10], using the same set of metrics as in image captioning. Table 3 (right) summarizes the results and Figure 4 provides sample qualitative results. Overall, we find that PaLI performs poorly, only achieving the average accuracy score of $19.9\\%$ , while the state-of-the-art literal VQA on the popular VQAv2 [18] benchmark is $77.6\\%$ on \"other\" questions. Additionally, the model struggles the most with answering questions that ask for the secondary concept.", + "bbox": [ + 500, + 497, + 892, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Metaphor Localization", + "text_level": 1, + "bbox": [ + 500, + 707, + 714, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, the task is to localize the image regions that invoke either the primary or secondary concept in the viewer. This is similar to the phrase grounding task of localizing objects using free-form natural language phrases [23, 31, 55], but with some key differences due to the peculiarities of visual metaphors in comparison to literal images used in standard vision datasets.", + "bbox": [ + 496, + 732, + 893, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Annotations. As discussed earlier, there are diverse types of metaphors based on how the primary and secondary concepts are visually depicted in an image (see Fig. 2). There are at least two key differences in metaphor localization", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "The score of 0 is given if parsing fails.", + "bbox": [ + 94, + 886, + 305, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23206", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b4b8ddebaf432558714448de80b905fa83aefe24b6fb1a6d2244719f4e3b589.jpg", + "image_caption": [ + "The shoe is as light as feathers." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 274, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecdf30748a846733fd43bca6f7f9046d1f337aa12a79b070a67ffe786f7a7ec8.jpg", + "image_caption": [ + "Ketchup is as hot as a volcano." + ], + "image_footnote": [], + "bbox": [ + 287, + 90, + 472, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf759da64092325daa75d573d8f7b0baa80180683bae3f615f2f08dd477c86f5.jpg", + "image_caption": [ + "Getting two ice creams is as exciting as getting two movie tickets for one." + ], + "image_footnote": [], + "bbox": [ + 485, + 90, + 679, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d385c119769f7f879f00b0beb73fb05361ac6890a80b1d96969615f633fd4aa.jpg", + "image_caption": [ + "This car is as out of this world as three dimensional." + ], + "image_footnote": [], + "bbox": [ + 687, + 90, + 879, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2af22fca7d1313656213e48edb1222d4085026bb710ff8103b6bf6cf3ce082f5.jpg", + "image_caption": [ + "Figure 5. Sample Localization annotations showing annotated bounding boxes around primary and secondary concepts. Notice the diversity of types in bounding boxes: explicit, contextual, logo and texts. This makes metaphor localization more challenging compared to standard object detection.", + "Parking assist is as safe as parking far from objects.", + "Figure 6. Sample Localizations with the phrase-grounding model from [27], where the secondary concepts are contextual. GT boxes are shown in green, whereas the predictions are shown in blue." + ], + "image_footnote": [], + "bbox": [ + 86, + 301, + 264, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc838d87c41af418b4aba5ec7b745dbc9daf79a7b8e8e8a784f338f7b9b27c44.jpg", + "image_caption": [ + "This chocolate is as fun as going to a concert." + ], + "image_footnote": [], + "bbox": [ + 277, + 301, + 452, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared to standard localization in literal images: 1. A given concept can be present in the image either explicitly or in a contextual manner (for e.g., contextual visual metaphors in Fig. 2). 2. Visual metaphor Ads are inherently multimodal and a concept can be invoked by other modalities such as text or logo. See the multimodal visual metaphors in Fig. 2. As a result, we not only annotate the bounding boxes that invoke the primary/secondary concept in the viewer, but we also annotate the type of that bounding box. A bounding box can be of one of the 4 types: Explicitly present, Contextually present, Logo or a Text. Specifically, for each of the 5061 metaphorical images, we pick the best metaphor annotation (primary, secondary concepts and their relationship) according to their validation scores (see previous section) and collect bounding box annotations for both the primary and secondary concepts. We collect all the bounding boxes that invoke both the primary and secondary concepts and also their type (explicit, contextual, logo or text) for each of the images. We use 5 annotators for each annotation and choose the bounding boxes with the best inter-annotator agreement. We use the same train and test splits as used in understanding tasks. Fig. 5 shows sample localization annotations. We collected over $30\\mathrm{k}$ bounding box annotations for this task on 5061 metaphor images.", + "bbox": [ + 75, + 505, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Detection. Recent detection and localization models [27, 58] pre-trained on image and caption pairs are shown to", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67c5583be3961a464bfc52e45f28027b26f8446ca5354ff2d10472007146ede5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
mAP50mAP70
Primary concept33.2214.25
Secondary concept43.5431.23
", + "bbox": [ + 573, + 297, + 820, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Localization results with CLIP based phrase localization model [27] on our test split.", + "bbox": [ + 500, + 366, + 890, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieve remarkable localization performance on discriminating fine-grained objects and unseen concepts. Specifically, we evaluate [27] which leverages the effective image representations in CLIP by extracting spatial features from it. Using these spatial features, for each pixel location, the model computes the inner product between the spatial feature and the phrase embedding extracted from CLIP to predict the bounding box. In our case, we pass the primary or secondary concept as input phrase to [27] the estimate the corresponding bounding box. Table 4 summarizes the detection results (using Mean Average Precision) on our test split. We find relatively better performance in localizing secondary objects compared to primary objects. We show few qualitative results in Figure 6. It is worth noting that our collected annotations allow for more comprehensive analysis on localization tasks due to the availability of different types of bounding boxes (explicit, contextual, logos and texts).", + "bbox": [ + 496, + 410, + 890, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Metaphor Generation", + "text_level": 1, + "bbox": [ + 500, + 694, + 705, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recent large-scale text-to-image (T2I) generative models show remarkable success in generating highly realistic images from text prompts. Can these models also work well in metaphorical image generation? We evaluate two state-of-the-art generative models (Imagen [42], Stable diffusion [41] (SD)) using 300 samples from MetaCLUE test set where we use the text prompts: \"An advertisement where primary-concept is as relationship as secondary concept.\" In addition, we finetune the stable diffusion model on our train split (same split as in Sec. 3.2).", + "bbox": [ + 496, + 718, + 890, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Fig. 7 shows sample visual results from different T2I models, with the metaphor annotation shown on the", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23207", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/78887cce51915947bcea4c90f5cd4f5a7ccb0f3acee0e19ea3b4fde41a05d0d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFID ↓CLIP Similarity ↑
Imagen [42]153.132.1
Stable Diffusion [41]161.630.8
Stable Diffusion - FT154.332.0
", + "bbox": [ + 130, + 88, + 416, + 154 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Analysis of Image Generation results with standard metrics of FID [42] and CLIP similarity [39] scores.", + "bbox": [ + 76, + 162, + 468, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "top. The generated images capture different aspects of the metaphor (tablet, waterproof), but not the entire metaphorical message. We compute standard metrics to automatically evaluate the quality of the generations. Tab. 5 shows the standard FID [19] and CLIP-Similarity [39] scores of different models. FID score evaluates the image distribution similarity between the generation images with the corresponding real image distribution. FID scores in Tab. 5 shows that Imagen performs slightly better than SD. And, there is a slight improvement in FID with finetuning (SDFT vs. SD). In general, high FID scores in Tab. 5 indicate the large distribution gap between the generated and real images. CLIP similarity score, on the other hand, measures the prompt fidelity - similarity between the generated image and the corresponding input text prompt according to the CLIP model [39]. CLIP similarity scores in Tab. 5 follow the similar trend as FID scores: Imagen performs better than SD in terms of the prompt fidelity, and finetuning SD model improved its prompt fidelity.", + "bbox": [ + 75, + 223, + 473, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given that both FID and CLIP scores are not tailored towards metaphorical images, these metrics are not reliable in assessing whether the generated images capture the essence of visual metaphors. To analyze this, we perform human studies comparing two different models at a time. Specifically, we show a metaphorical message and the corresponding generations from two models and ask the users to pick an image that best conveys that metaphorical message. We obtain 7 user ratings for each image pair and consider 3-out-of-7 or 4-out-of-7 to be ties. Fig. 8 shows the percentage of user preferences across different pairs of results. User studies also indicate that Imagen performed better than SD. An interesting finding is that finetuning SD resulted in slightly worse user preference compared to base SD model (SDFT vs. SD). This is in contrast to standard FID and CLIP metrics that improved with finetuning. We hypothesize that finetuning SD resulted in more realistic Ad images, but the resulting model may have forgotten some of the metaphorical priors. This calls for more effective finetuning strategies with the limited training datasets, which forms an important future work. In addition, both Imagen and SD performed quite poorly compared to real images in conveying metaphorical messages. Real images are preferred around $88\\%$ of time over Imagen results. This illustrates the big scope of improvements in generating visual metaphors.", + "bbox": [ + 75, + 523, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metaphor: This android tablet is as waterproof as someone in a swimtube.", + "bbox": [ + 514, + 85, + 852, + 108 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/057c8a7d03abade2c4c6a9902380c8451e89473675b0d84921203e9a77d6f0a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 113, + 666, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/610235dcb4fdee2ad461930d71f923cb8d5ce027804e68ca032a25f03d8fe1b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 108, + 875, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9c9a9e32bff7517f5c17a00f640199c65dc936bebc7651716b99588707cf02c.jpg", + "image_caption": [ + "Figure 7. Sample Image Generations for a given metaphorical message (shown on top) with Imagen [42], Stable Diffusion [41] and fine-tuned (FT) version of Stable Diffusion." + ], + "image_footnote": [], + "bbox": [ + 506, + 263, + 689, + 415 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e9a00b90199accce31f522c019995823bcd16b8e9949fb7bcf904d536de09cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 265, + 875, + 415 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/317d1e758b9dbbcecdd3e33759232a5e45f0b0ad5239d742ba643c80836fdac5.jpg", + "image_caption": [ + "Figure 8. User Study on Image Generation Results. Percentage of results users preferred across real, Imagen [42], Stable Diffusion (SD) [41] and its fine-tuned version (SD-FT) results. Users are asked to choose the image that better depicts a given metaphor." + ], + "image_footnote": [], + "bbox": [ + 504, + 491, + 888, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 676, + 619, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a step towards comprehensive evaluation of progress on visual metaphor research. Specifically, we propose a collection of tasks related to comprehending and generating visual metaphors using AI techniques. Our MetaCLUE tasks include Classification, Understanding (Retrieval, Captioning, VQA), Localization and Generation. For comprehensive evaluations, we collected high quality and rich annotations that facilitate the measurable progress. Existing methods demonstrate poor results in many cases with our experimental analysis shedding light on strengths and drawbacks of different approaches paving a path for future research in this fascinating field.", + "bbox": [ + 496, + 704, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23208", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569-11579, 2021. 2, 3", + "[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. SPICE: semantic propositional image caption evaluation. In ECCV, 2016. 6", + "[3] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In ACL Workshops, 2005. 6", + "[4] Roland Barthes. Image-music-text. Macmillan, 1977. 3", + "[5] Nathalie Bonnardel and Evelyne Marmèche. Towards supporting evocation processes in creative design: A cognitive approach. International journal of human-computer studies, 63(4-5):422-435, 2005. 3", + "[6] Robert W Boozer, David C Wyld, and James Grant. Using metaphor to create more effective sales messages. Journal of services marketing, 1990. 3", + "[7] Tuhin Chakrabarty, Xurui Zhang, Smaranda Muresan, and Nanyun Peng. Mermaid: Metaphor generation with symbolism and discriminative decoding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4250-4261, 2021. 2, 3", + "[8] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO Captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6", + "[9] Xianyang Chen, Chee Wee Leong, Michael Flor, and Beata Beigman Klebanov. Go figure! multi-task transformer-based architecture for metaphor detection using idioms: Ets team in 2020 metaphor shared task. In Proceedings of the second workshop on figurative language processing, pages 235-243, 2020. 3", + "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 5, 6", + "[11] Lydia B Chilton, Savvas Petridis, and Maneesh Agrawala. Visiblends: A flexible workflow for visual blends. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2019. 4", + "[12] Minjin Choi, Sunkyung Lee, Eunseong Choi, Heesoo Park, Junhyuk Lee, Dongwon Lee, and Jongwuk Lee. Melbert: Metaphor detection via contextualized late interaction using metaphorical identification theories. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1763-1773, 2021. 2, 3" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 5", + "[14] Gilles Fauconnier and Mark Turner. The way we think: Conceptual blending and the mind's hidden complexities. Basic books, 2008. 3", + "[15] Charles Forceville. Pictorial metaphor in advertising. Routledge, 1996. 2, 3", + "[16] Susan R Fussell and Mallie M Moss. Figurative language in emotional communication. Social and cognitive approaches to interpersonal communication, pages 113-141, 1998. 3", + "[17] Hongyu Gong, Kshitij Gupta, Akriti Jain, and Suma Bhat. Illinimet: Illinois system for metaphor detection with contextual and linguistic information. In Proceedings of the Second Workshop on Figurative Language Processing, pages 146-153, 2020. 3", + "[18] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6", + "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8", + "[20] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic understanding of image and video advertisements. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1705-1715, 2017. 2, 3, 4", + "[21] Bipin Indurkhya. On the role of metaphor in creative cognition. In ICCC, pages 51-59, 2010. 3", + "[22] Bipin Indurkhya and Amitash Ojha. An empirical study on the role of perceptual similarity in visual metaphors and creativity. Metaphor and Symbol, 28(4):233-253, 2013. 4", + "[23] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 6", + "[24] George Lakoff. The contemporary theory of metaphor. 1993. 2", + "[25] George Lakoff and Mark Johnson. Metaphors we live by. University of Chicago press, 2008. 2", + "[26] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 5", + "[27] Jiahao Li, Greg Shakhnarovich, and Raymond A Yeh. Adapting clip for phrase localization without further training. arXiv preprint arXiv:2204.03647, 2022. 7", + "[28] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23209", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Todd I Lubart and Isaac Getz. Emotion, metaphor, and the creative process. Creativity research journal, 10(4):285-301, 1997. 2", + "[30] Carita Lundmark. Metaphor and creativity in British magazine advertising. PhD thesis, Luleå teknska universitet, 2005. 3", + "[31] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6", + "[32] James H Martin. A corpus-based analysis of context effects on metaphor comprehension. Trends in Linguistics Studies and Monographs, 171:214, 2006. 3", + "[33] Zachary J Mason. Cormet: a computational, corpus-based conventional metaphor extraction system. Computational linguistics, 30(1):23-44, 2004. 3", + "[34] Edward F McQuarrie and David Glen Mick. Visual rhetoric in advertising: Text-interpretive, experimental, and reader-response analyses. Journal of consumer research, 26(1):37-54, 1999. 3", + "[35] Ekaterina Ovchinnikova, Vladimir Zaytsev, Suzanne Wertheim, and Ross Israel. Generating conceptual metaphors from proposition stores. arXiv preprint arXiv:1409.7619, 2014. 3", + "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. 2002. 6", + "[37] Barbara J Phillips. The impact of verbal anchoring on consumer response to image ads. Journal of advertising, 29(1):15-24, 2000. 3", + "[38] Barbara J Phillips and Edward F McQuarrie. Beyond visual metaphor: A new typology of visual rhetoric in advertising. Marketing theory, 4(1-2):113-136, 2004. 3", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5, 8", + "[40] Sunny Rai and Shampa Chakraverty. A survey on computational metaphor processing. ACM Computing Surveys (CSUR), 53(2):1-37, 2020. 3", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 7, 8", + "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 7, 8", + "[43] Linda M Scott. Images in advertising: The need for a theory of visual rhetoric. Journal of consumer research, 21(2):252-273, 1994. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Gerard Steen. A method for linguistic metaphor identification: From MIP to MIPVU, volume 14. John Benjamins Publishing, 2010. 3", + "[45] Kevin Stowe, Tuhin Chakrabarty, Nanyun Peng, Smaranda Muresan, and Iryna Gurevych. Metaphor generation with conceptual mappings. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6724-6736, 2021. 2, 3", + "[46] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 4, 5", + "[47] Asuka Terai and Masanori Nakagawa. A computational system of metaphor generation with evaluation mechanism. In International Conference on Artificial Neural Networks, pages 142-147. Springer, 2010. 3", + "[48] Xiaoyu Tong, Ekaterina Shutova, and Martha Lewis. Recent advances in neural metaphor processing: A linguistic, cognitive and social perspective. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4673-4686, 2021. 2, 3", + "[49] Peter Turney, Yair Neuman, Dan Assaf, and Yohai Cohen. *Literal and metaphorical sense identification through concrete and abstract context*. In *Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing*, pages 680–690, 2011. 3", + "[50] Tony Veale, Ekaterina Shutova, and Beata Beigman Klebanov. Metaphor: A computational perspective. Synthesis Lectures on Human Language Technologies, 9(1):1-160, 2016. 2", + "[51] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015. 6", + "[52] Merryl J Wilkenfeld and Thomas B Ward. Similarity and emergence in conceptual combination. Journal of Memory and Language, 45(1):21-38, 2001. 3", + "[53] Bo Xu, Tingting Li, Junzhe Zheng, Mehdi Naseriparsa, Zhehuan Zhao, Hongfei Lin, and Feng Xia. Met-meme: A multimodal meme dataset rich in metaphors. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2887-2899, 2022. 4", + "[54] Keren Ye and Adriana Kovashka. Advise: Symbolism and external knowledge for decoding advertisements. In Proceedings of the European Conference on Computer Vision (ECCV), pages 837-855, 2018. 4", + "[55] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 6", + "[56] Zhiwei Yu and Xiaojun Wan. How to avoid sentences spelling boring? towards a neural approach to unsupervised metaphor generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23210", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 861-871, 2019. 3", + "[57] Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, and Hongfei Lin. Multimet: A multimodal dataset for metaphor understanding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3214-3225, 2021. 4", + "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 7" + ], + "bbox": [ + 78, + 90, + 468, + 258 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23211", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_model.json b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9fb888d615c4ffad40b157d5a0cd262ef41bfe3d --- /dev/null +++ b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_model.json @@ -0,0 +1,2675 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.153, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "MetaCLUE: Towards Comprehensive Visual Metaphors Research" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.181, + 0.803, + 0.253 + ], + "angle": 0, + "content": "Arjun R. Akula*, Brendan Driscoll*, Pradyumna Narayana, Soravit Changpinyo, Zhiwei Jia, Suyash Damle, Garima Pruthi, Sugato Basu, Leonidas Guibas, William T. Freeman, Yuanzhen Li, Varun Jampani* Google" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.291, + 0.351, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.439, + 0.345, + 0.47 + ], + "angle": 0, + "content": "Metaphor: Killing trees is as harmful as killing wildlife." + }, + { + "type": "title", + "bbox": [ + 0.419, + 0.291, + 0.55, + 0.309 + ], + "angle": 0, + "content": "Classification" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.314, + 0.6, + 0.34 + ], + "angle": 0, + "content": "Is this a visual metaphor? - YES - NO" + }, + { + "type": "title", + "bbox": [ + 0.425, + 0.356, + 0.544, + 0.373 + ], + "angle": 0, + "content": "Localization" + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.376, + 0.594, + 0.405 + ], + "angle": 0, + "content": "Detect image regions that invoke the concepts:" + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.411, + 0.45, + 0.424 + ], + "angle": 0, + "content": "- Killing trees" + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.428, + 0.46, + 0.44 + ], + "angle": 0, + "content": "- killing wildlife" + }, + { + "type": "list", + "bbox": [ + 0.357, + 0.411, + 0.46, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.389, + 0.601, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.698, + 0.293, + 0.805, + 0.311 + ], + "angle": 0, + "content": "gEneration" + }, + { + "type": "text", + "bbox": [ + 0.623, + 0.315, + 0.853, + 0.36 + ], + "angle": 0, + "content": "Prompt: \"An advertisement where killing trees is as harmful as killing wildlife.\"" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.366, + 0.745, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.456, + 0.739, + 0.467 + ], + "angle": 0, + "content": "Stable Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.366, + 0.871, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.455, + 0.839, + 0.468 + ], + "angle": 0, + "content": "Imagen" + }, + { + "type": "title", + "bbox": [ + 0.413, + 0.48, + 0.557, + 0.499 + ], + "angle": 0, + "content": "Understanding" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.501, + 0.156, + 0.514 + ], + "angle": 0, + "content": "Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.517, + 0.213, + 0.53 + ], + "angle": 0, + "content": "Pick the right one:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.533, + 0.325, + 0.562 + ], + "angle": 0, + "content": "(a) Killing the forest is as deadly as killing the animals too." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.564, + 0.312, + 0.593 + ], + "angle": 0, + "content": "(b) Birds is as much a part of our world as used cans." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.533, + 0.325, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.5, + 0.414, + 0.515 + ], + "angle": 0, + "content": "Captioning" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.517, + 0.461, + 0.531 + ], + "angle": 0, + "content": "Sample predictions:" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.533, + 0.559, + 0.562 + ], + "angle": 0, + "content": "1. Deforestation is as damaging as killing wildlife." + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.564, + 0.594, + 0.593 + ], + "angle": 0, + "content": "2. Deforestation is as bad as ending the death penalty." + }, + { + "type": "list", + "bbox": [ + 0.336, + 0.533, + 0.594, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.5, + 0.808, + 0.531 + ], + "angle": 0, + "content": "Visual Question Answering Sample Questions:" + }, + { + "type": "text", + "bbox": [ + 0.62, + 0.533, + 0.83, + 0.562 + ], + "angle": 0, + "content": "Q. What is as harmful as killing wildlife?" + }, + { + "type": "text", + "bbox": [ + 0.62, + 0.57, + 0.88, + 0.585 + ], + "angle": 0, + "content": "Q. What is compared to killing wildlife?" + }, + { + "type": "list", + "bbox": [ + 0.62, + 0.533, + 0.88, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.607, + 0.894, + 0.678 + ], + "angle": 0, + "content": "Figure 1. With MetaCLUE, we introduce several interesting tasks related to visual metaphors. We collect metaphor annotations (objects, abstract concepts, relationships and object boxes) for evaluating existing models on these tasks. Specifically we perform a comprehensive evaluation of vision and language models on four different tasks (Classification, Localization, Understanding, and gEneration). Comprehensive experiments in this work show that state-of-the-art techniques mostly focus on literal interpretation and perform poorly in understanding and generation of metaphor images." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.686, + 0.314, + 0.701 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Creativity is an indispensable part of human cognition and also an inherent part of how we make sense of the world. Metaphorical abstraction is fundamental in communicating creative ideas through nuanced relationships between abstract concepts such as feelings. While computer vision benchmarks and approaches predominantly focus on understanding and generating literal interpretations of images, metaphorical comprehension of images remains relatively unexplored. Towards this goal, we introduce MetaCLUE, a set of vision tasks on visual metaphor. We also collect high-quality and rich metaphor annotations (ab" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.87 + ], + "angle": 0, + "content": "struct objects, concepts, relationships along with their corresponding object boxes) as there do not exist any datasets that facilitate the evaluation of these tasks. We perform a comprehensive analysis of state-of-the-art models in vision and language based on our annotations, highlighting strengths and weaknesses of current approaches in visual metaphor classification, localization, understanding (retrieval, question answering, captioning) and generation (text-to-image synthesis) tasks. We hope this work provides a concrete step towards developing AI systems with human-like creative capabilities. Project page: https://metaclue.github.io" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23201" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.21, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.115, + 0.438, + 0.177 + ], + "angle": 0, + "content": "\"Metaphor is pervasive in everyday life ... Our ordinary conceptual system, in terms of which we both think and act, is fundamentally metaphorical in nature.\" — Lakoff & Johnson [25]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.191, + 0.47, + 0.341 + ], + "angle": 0, + "content": "Creativity is a process of generating a new perspective on a problem or a situation. Metaphorical thinking has been recognized as a key and powerful mechanism of creativity [24,29,50]. Humans engage metaphors in their creative thinking process as strategies to link or blend concepts, or to view a concept from a target domain in terms of another, apparently dissimilar concept from a source domain [25]. metaphors also provide a sophisticated tool for nuanced human communication. Let us take a closer look at the structure of metaphors – and especially visual metaphors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.471, + 0.612 + ], + "angle": 0, + "content": "Metaphors1 are a cognitive construct in which a concept is compared to a seemingly unrelated concept via some shared attribute. Take as an example 'This car is a cheetah', where 'This car' is compared to 'a cheetah' in terms of speed. Metaphors have a simple syntactic structure of 'A is B' where A is referred to as the primary concept and B as the secondary concept. The implied analogy in a metaphor is of the form: ' (primary concept) is as (relationship)2 as (secondary concept)' and often involves an attribute transfer from the secondary to the primary concept. Some examples include 'This phone is as fast as a rocket', 'Cigarettes are as harmful as bullets' etc. The primary and secondary concepts are usually unrelated at a glance, resulting in an element of surprise and creativity in metaphorical expressions. Despite following such simple structure, metaphors are quite powerful in conveying creative ideas. Metaphors are pervasive in all forms of communication, such as speech, text, visual etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.841 + ], + "angle": 0, + "content": "Visual Metaphors are images where the primary and secondary concepts are visually depicted in an image conveying the metaphorical message to the viewers. Visual metaphors are widely used in mass media communications like advertising and journalism [15, 43, 45]. In this work, we work with Ad images, as metaphors tend to be prevalent in ads. There are numerous ways a metaphor can be represented visually. Following the classification in [15], there are at least 4 different types of visual metaphors. Fig. 2 shows sample images that belong to these types along with our annotations of primary, secondary concepts and their relationship. In contextual metaphors, either the primary or secondary concept is not explicitly visible, but is inferred from the context (e.g., apple in the left-most image). In hybrid metaphors, the primary and secondary concepts are" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.241 + ], + "angle": 0, + "content": "visually conflated. Juxtaposition forms one of the simplest visual metaphor types, where the two concepts are just presented next to each other. Multimodal metaphors represent one of the concepts with another modality, such as text or logo. In practice, visual metaphors use multiple of these strategies to convey a metaphor in an effective manner. In many cases, the implied metaphorical meaning is somewhat open-ended. Interpretation of visual metaphors depends on several external factors, such as familiarity with the brands and cultural context." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.424 + ], + "angle": 0, + "content": "These visual variations and nuances make automatic cognition or generation of visual metaphors highly challenging. While the last decade has seen rapid progress in many areas of understanding and generation tasks, prior works in computer vision focus heavily on literal interpretation of images and overlook the importance of metaphorical reasoning in understanding the image message [1, 48]. We believe that developing AI systems with metaphorical comprehension and generation capabilities can greatly assist humans in creative endeavors involving conveying concepts in new and exciting ways. Such systems provide an important step towards conferring human-like creativity to AI models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.591 + ], + "angle": 0, + "content": "To this end, we introduce multiple interesting tasks and construct metaphor annotations that enable comprehensive research on visual metaphors. As metaphors are more common in visual Ads, we start with the Pitt's Ads dataset images [20] and then perform a rigorous multi-stage annotation process with expert annotators to filter metaphorical images, add metaphor annotations, and perform additional validation steps to clean the annotations3. While there is recent work making advances in understanding non-literal interpretations in natural language research [7, 12], this work proposes the first step towards metaphor analysis in images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.864 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, we perform comprehensive evaluations with state-of-the-art techniques on four sets of tasks, which we call MetaCLUE: 1. Classification: This is binary classification task of estimating whether a given image contains a metaphor or not. In other words, Are visual features indicative of whether there exists a metaphor in a given image or not?. 2. Localization: Here, the task is to localize the image regions that invoke the primary and secondary concepts in the viewer. This is similar to a standard object detection task, but is more complicated in the case of visual metaphors as the primary/secondary concepts may not be explicitly present in an image. 3. Understanding: Can our models understand the metaphorical message in a given image? We pose this understanding problem as 3 tasks where we can quantitatively measure the performance: Retrieval, Captioning and Visual question answering. 4. gEneration: Can we generate an image that conveys the metaphor, given the metaphorical message as a text prompt?" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3The Ads, while useful for the purposes of our paper, some images may perpetuate harmful stereotypes according to characteristics such as gender." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.85, + 0.468, + 0.875 + ], + "angle": 0, + "content": "1. Grammarians distinguish a metaphor \"A is B\" from a simile \"A is like B\". In our work we use \"metaphor\" to encompass both variants." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.468, + 0.901 + ], + "angle": 0, + "content": "2We use the word 'relationship' to denote the shared property of primary and secondary concepts, usually adjectives or adjectival phrases." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.85, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23202" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.093, + 0.325, + 0.104 + ], + "angle": 0, + "content": "Contextual" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.106, + 0.284, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.211, + 0.274, + 0.222 + ], + "angle": 0, + "content": "This beer is as tasty as a real apple." + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.109, + 0.474, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.205, + 0.472, + 0.227 + ], + "angle": 0, + "content": "This car is as adventurous as a space ship." + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.092, + 0.704, + 0.105 + ], + "angle": 0, + "content": "Hybrid" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.107, + 0.677, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.205, + 0.679, + 0.227 + ], + "angle": 0, + "content": "Driving this SUV is as smooth as birds flying in the sky." + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.107, + 0.871, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.211, + 0.856, + 0.222 + ], + "angle": 0, + "content": "This pencil is as red as a fire truck." + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.234, + 0.334, + 0.246 + ], + "angle": 0, + "content": "Juxtaposition" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.248, + 0.284, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.352, + 0.275, + 0.363 + ], + "angle": 0, + "content": "This chocolate bar is as rich as gold." + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.247, + 0.478, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.348, + 0.477, + 0.369 + ], + "angle": 0, + "content": "This car is as made for the beach as a crab." + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.235, + 0.72, + 0.247 + ], + "angle": 0, + "content": "Multimodal" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.249, + 0.673, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.348, + 0.675, + 0.369 + ], + "angle": 0, + "content": "These donuts are as unique as as talking people." + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.248, + 0.869, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.348, + 0.854, + 0.369 + ], + "angle": 0, + "content": "The car is as rugged as this muddy trailer." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.893, + 0.426 + ], + "angle": 0, + "content": "Figure 2. Sample Visual Metaphors with their Annotations. There are different types of visual metaphors. The type depends on how the primary and secondary concepts are visually depicted. Here are sample Ad images from [20] where we annotated the primary concept, secondary concept and their relationship." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.473, + 0.62 + ], + "angle": 0, + "content": "We comprehensively evaluate existing state-of-the-art techniques for each of these tasks on our collected metaphor annotations. We evaluate the models both in a zero-shot manner as well as with finetuning on our annotations. Even though finetuning resulted in some improvements, most models struggle to produce satisfactory results in many cases, demonstrating the difficulty of these tasks. Our experiments highlight several strengths and weaknesses of the existing techniques on comprehending and generating visual metaphors, providing a concrete first step towards further AI research on this fascinating topic." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.634, + 0.219, + 0.649 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Creativity and Metaphorical Abstraction. Creativity often involves an innovative fusion of objects, attributes, or relationships from previous knowledge to generate new concepts [5, 52]. Metaphors can serve as an invaluable tool for expressing creative insights and also to stimulate new ones [14, 21]. The cognitive research community has made initial attempts in understanding different realizations of metaphors such as language metaphors and visual metaphors [15, 16, 32]. Studies of visual persuasion show that visual metaphors may be more effective than language metaphors in terms of producing a greater degree of cognitive creativity [34]. In addition to improving creativity, metaphors are also known to elicit pleasure since the initial ambiguity in (re-)conceptualizing a target entity in terms of a source stimulus generates interest and motivation, and the subsequent resolution is rewarding – explaining the impor" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.453, + 0.892, + 0.484 + ], + "angle": 0, + "content": "tance of using creative metaphorical processes in art, advertising, and marketing [6,30,38]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.893, + 0.64 + ], + "angle": 0, + "content": "Metaphors in Language Research. Computational Linguistic studies show that metaphors are ubiquitous in language, occurring once per three sentences on average [40, 44, 49]. Recently, an increasing number of research efforts have explored the limitations and challenges in detecting and decoding the meaning of language metaphors [7, 48]. While there exist several computational models [9, 12, 17, 33, 35, 45, 47, 56] for metaphor identification, interpretation and generation in language, there exists very little work on computational modeling of visual metaphors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Metaphors in Computer Vision. Much of computer vision literature is focused on understanding and generating literal images. Automatic metaphorical interpretations of images is highly challenging and requires multi-faceted cognitive reasoning that involves visual reasoning coupled with the use of external knowledge. Recent work on affective captioning using the ArtEmis data set [1] includes some captions that evoke metaphors to explain emotions, but that dataset is focused on visual art and does not specifically consider metaphors. There is no explicit prior work for comprehensive evaluation and development of models that can automatically comprehend or generate visual metaphors. At the same time, there exist several studies that demonstrate the potential of visual metaphors. For instance, some studies [4,37] suggest that advertisements containing metaphorical images will be more persuasive compared to ads with language metaphors, non-metaphorical" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23203" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.089, + 0.42, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.26, + 0.47, + 0.289 + ], + "angle": 0, + "content": "Figure 3. Distribution of Topics in our annotated metaphorical images from Pitt's Ad dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.473, + 0.618 + ], + "angle": 0, + "content": "ads, or literal images. There have been some prior computational models [11, 22] for metaphor generation, but they are not thoroughly validated against any benchmark datasets or user studies. Some works [20, 54] propose datasets and techniques for general Ad image understanding with a focus on the challenging aspects of non-literal interpretations in Ad images. However they do not explicitly collect any metaphorical annotations, nor do they provide the corresponding analysis. In this work, we start with an existing Ad image dataset [20] and perform extensive human studies to filter metaphorical images and collect detailed annotations accounting various aspects of metaphoric interpretation. Prior works such as Multi-MET [57] and MET-Meme [53], propose metaphorical annotations but does not annotate the relationship between primary and secondary. In MetaCLUE, in addition to providing relationship annotations, we also collect detailed bounding box annotations that help localize the image regions invoking the primary, secondary concepts in the viewer. We further provide VQA style question and answers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.201, + 0.646 + ], + "angle": 0, + "content": "3. MetaCLUE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.763 + ], + "angle": 0, + "content": "We introduce four different high-level tasks in MetaCLUE that enable comprehensive evaluation and development of visual metaphor research: Classification, Localization, Understanding and gEneration. In the rest of this section, we first describe our annotation collection process for the tasks and next provide the analysis of using existing state-of-the-art techniques for each of these tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.298, + 0.788 + ], + "angle": 0, + "content": "3.1. Metaphor Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Following the fundamental vision task of image classification, we first ask: Can we develop models that can classify whether or not a given image contains a visual metaphor? In other words, is it possible to just use visual cues to estimate whether or not there exists a metaphorical interpretation of an image?" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Annotations. For this task, we need to label whether or" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.621 + ], + "angle": 0, + "content": "not each image contains a metaphor. Since we tend to see more metaphorical images in Ads, we start with images from an existing Ads dataset published by the University of Pittsburgh [20] and manually annotate whether a given Ad image contains a metaphor or not. Pitt's Ads dataset contains images of both product ads (e.g. phone ads) as well as public service announcements (e.g. forest conservation ads). Concretely, to make annotations more efficient, we use a subset of 8.5K Ad images from this dataset that are annotated to have 'symbolic' (fun, adventurous, etc.) references. We find this subset to contain a considerable portion of metaphors. Specifically, we collect 5 Yes/No annotations for each of these 8K images from different annotators. To obtain high-quality annotations, we explained what a visual metaphor is to the annotators and also conducted qualifying exams to pick final annotators for this task. We consider the images with 3 or more 'Yes' annotations as visual metaphors and the remaining as non-metaphorical. At the end, we identify 5061 images containing visual metaphors out of 8480 images. Pitt's Ads dataset [20] also comes with topic annotations (e.g., restaurant, car, animal rights etc.). Fig. 3 shows the word cloud plot of different topics in these 5061 metaphorical images. We split the metaphorical images into 3730 train and 1331 test images by maintaining the same distribution of topics in both the splits. We have two types of negative sets (non-metaphorical images) for classification experiments. One is formed by the remaining 3419 non-metaphorical images in the symbolic set, and another one is created separately by annotating an additional 3000 literal (and non-symbolic) images from Pitt's Ads dataset [20]. We add 2000 of 3419 symbolic negative images to our train split, and add the remaining 1419 images to our test split. Similarly 2000 of 3000 literal negatives are used in training, and the remaining 1000 are used for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.895, + 0.819 + ], + "angle": 0, + "content": "Evaluation and Results. Using our collected metaphor images, we evaluate the performance of the following state-of-the-art models in classifying an input image as metaphor: EfficientNet [46] and Vision Transformer (ViT) [13]. We fine-tune these models to classify metaphor vs. symbolic/non-metaphors and metaphor vs. literal, and use \\(20\\%\\) of corresponding train splits for validation. Test results are reported in Table 1. Although the performance of ViT is significantly better than EfficientNet and random baselines, there is still ample room for improvement. In particular, models find it easier to distinguish metaphor images from literal images, and struggle to identify metaphors within the symbolic image pool." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.733, + 0.848 + ], + "angle": 0, + "content": "3.2. Metaphor Understanding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We now describe how we collect annotations that help in capturing the metaphorical message from the images collected in previous section." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23204" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.454, + 0.194 + ], + "angle": 0, + "content": "
Model#ParamsSymbolic Neg.Literal Neg.
ValTestValTest
RandomN/A63.1051.6060.6657.12
EfficientNet-B05.3M60.7649.6770.9450.30
EfficientNet-B766M61.4448.5469.8449.82
ViT-B/1686M69.3166.9884.0481.24
ViT-L/16307M65.8360.6581.4580.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.205, + 0.471, + 0.247 + ], + "angle": 0, + "content": "Table 1. Accuracy of Metaphor Classification (binary classification accuracy) using state-of-the-art classification architectures of EfficientNet [46] and ViT [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.47, + 0.71 + ], + "angle": 0, + "content": "Annotations. We provide detailed instructions and several examples to the annotators to help them annotate primary and secondary concepts in the metaphor and also the characteristic/relationship that is transferred from secondary to the primary. We conduct multiple pilot studies to reduce the noise and to improve inter-annotator agreement. Enforcing the annotators to make sure that their annotations are linguistically readable in the following syntactic structure helped us in improving quality and consistency of the annotations: “__ is as __ as __,” where the first blank is the primary concept, the second blank is the relationship, and the third blank is the secondary concept. Figure 2 shows some examples of these annotations. We collect 5 metaphor annotations for each image. As interpretations of metaphorical images can be highly subjective, there can exist more than one interpretation for each image, which makes it difficult to automatically remove noisy annotations. Therefore we conduct an additional human study where we show each of the annotation to five annotators and ask them to verify the correctness along three dimensions: (a) Is the grammar correct?; (b) Are primary and secondary concepts correct?; and (c) Is the relationship correct?. We remove annotations with a low number of votes out of 5 along each of the three dimensions, resulting in a total of over 26k clean annotations. We evaluate state-of-the-art models in understanding metaphorical message from the input images using 3 tasks namely, Retrieval, Captioning and Visual question answering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Retrieval. The goal of this task is to retrieve the correct metaphor interpretation/statement from a candidate set given an image. In our candidate set, we choose exactly one positive (correct) metaphorical statement from its ground truth messages and uniformly sample \\( K - 1 \\) random negative statements from other images. Table 2 shows the results obtained with CLIP [39] and ALBEF [26] for \\( K = 50 \\), reporting retrieval precision@1 and rank4. Although CLIP ViT-L/14 shows good zero-shot performance on random negatives with more than \\( 76\\% \\) accuracy, we observe a large drop" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.695, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.198, + 0.688, + 0.221 + ], + "angle": 0, + "content": "Metaphor: Smoking cigarettes is as life-shortening as sharpening a pencil" + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.093, + 0.884, + 0.131 + ], + "angle": 0, + "content": "Captioning Result Prediction:Smoking is as dangerous as burning a pencil." + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.139, + 0.88, + 0.174 + ], + "angle": 0, + "content": "Visual Question Answering Result Q1: What is a smoker's life compared to a sharpened pencil?" + }, + { + "type": "text", + "bbox": [ + 0.701, + 0.175, + 0.875, + 0.21 + ], + "angle": 0, + "content": "Pred: Sharp GT: Shortened Q2: What is used as a visual metaphor for a sharpened pencil?" + }, + { + "type": "text", + "bbox": [ + 0.701, + 0.212, + 0.873, + 0.222 + ], + "angle": 0, + "content": "Pred: Not smoking GT:Smoker's life" + }, + { + "type": "list", + "bbox": [ + 0.7, + 0.093, + 0.884, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.226, + 0.697, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.333, + 0.683, + 0.356 + ], + "angle": 0, + "content": "Metaphor: Smoking is as deadly as injecting drugs." + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.229, + 0.88, + 0.266 + ], + "angle": 0, + "content": "Captioning Result Prediction: Drinking and driving is as bad as injecting drugs." + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.275, + 0.878, + 0.308 + ], + "angle": 0, + "content": "Visual Question Answering Result Q1: What is a cigarette as compared to injecting drugs?" + }, + { + "type": "text", + "bbox": [ + 0.701, + 0.31, + 0.872, + 0.345 + ], + "angle": 0, + "content": "Pred: Scary GT: Deadly Q2: What is used as a visual metaphor for injecting drugs?" + }, + { + "type": "text", + "bbox": [ + 0.701, + 0.347, + 0.857, + 0.357 + ], + "angle": 0, + "content": "Pred:Drinking GT:Cigarette" + }, + { + "type": "list", + "bbox": [ + 0.7, + 0.229, + 0.88, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.37, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Figure 4. Results of PaLI [10] for captioning and visual question answering on sample images in our test split." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.517 + ], + "angle": 0, + "content": "in performance as we increase \\( K \\) to \\( \\{100, 500, 1000\\} \\). We further fine-tune CLIP models using \\( 70\\% \\) of metaphor annotations as the train set and see gains by up to \\( +7 \\) absolute points in \\( p@1 \\). In summary, the performance of models is impressive with less than 50 negative candidates whereas the performance drops greatly by increasing \\( K \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.893, + 0.833 + ], + "angle": 0, + "content": "We hypothesize that models might simply be looking at salient objects rather than comprehending the underlying semantics of metaphor in finding the correct candidate. To test this, we mine hard negative (HN) statements and use them as our candidate set. Specifically, we construct the following five types of HNs: (a) Neg Prim: candidates obtained by replacing the primary concept in the metaphor statement with the primary concept from another image; (b) Neg Sec: replacing secondary concept likewise; (c) Neg Prim+Rel: replacing primary and relationship; (d) Neg Sec+Rel: replacing secondary and relationship; (e) Swap Prim&Neg: swapping primary and secondary from the same image. Table 2 shows the results. We see a significant drop of up to \\(30\\%\\) by using HNs as negative statements for \\(K = 50\\), indicating the difficulty in comprehending and distinguishing metaphorical abstraction of concepts. We find performance of Neg Prim is significantly higher than Neg Sec, suggesting that the models tend to rely more on primary object than secondary object in identifying correct interpretation. Overall, there is ample room for improvement indicated by the steep drop in model performance with HNs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.838, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Captioning. Here, we propose metaphor image cap" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Rank measures the averaged ranking value of the highest-ranked ground-truth statement with 1 being the highest possible rank." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.888 + ], + "angle": 0, + "content": "5We swap objects with images having different topic to make sure that the generated HNs are actually negatives." + }, + { + "type": "page_footnote", + "bbox": [ + 0.519, + 0.888, + 0.688, + 0.901 + ], + "angle": 0, + "content": "With swapping \\(K\\) is always 2." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23205" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.872, + 0.251 + ], + "angle": 0, + "content": "
ModelRandom NegNeg PrimNeg SecNeg Prim+RelNeg Sec+RelSwap accuracy↑
p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓
CLIP (ViT-B/16)70.973.4946.6710.1138.1413.9749.368.8042.2511.6940.61
CLIP (ViT-B/32)61.784.1938.7411.6033.2014.8439.7910.2936.2712.7141.28
CLIP (ViT-L/14)76.663.1751.759.2239.8613.1954.747.7045.9911.0243.08
ALBEF39.798.5727.0016.5329.3115.7926.7715.5828.4214.5946.67
ALBEF (MSCOCO)44.877.5531.7815.1331.4114.0932.5314.0433.2812.8548.24
ALBEF (Flickr30k)47.498.7735.6014.5635.2213.6636.3513.5836.1212.7949.81
FT CLIP (ViT-B/16)76.812.2549.819.6545.4710.2453.408.0750.638.1844.65
FT CLIP (ViT-B/32)68.062.8243.0010.8239.6411.2344.729.3943.459.3545.69
FT CLIP (ViT-L/14)81.751.7857.667.4849.069.4061.255.9957.067.3343.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.261, + 0.893, + 0.29 + ], + "angle": 0, + "content": "Table 2. Performance of retrieval models on \\( K \\) random (column 2) and hard negative candidates (columns 3-7) (\\( K = 50 \\) for columns 2-6, \\( K = 2 \\) for last column)." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.315, + 0.885, + 0.392 + ], + "angle": 0, + "content": "
CaptioningAccBLEU4ROUGE-LMETEORSPICECIDErVQAAccBLEU4ROUGE-LMETEORSPICECIDEr
Whole caption1.1%0.2540.5360.2200.1861.076All questions19.9%0.3290.2860.2490.1850.851
Primary29.9%0.3270.4070.3380.2440.931Primary21.5%0.2910.3480.2770.2901.099
Secondary13.7%0.2490.3070.1930.1550.550Secondary12.8%0.2380.2320.1810.2340.735
Relationship23.6%0.4850.2030.2260.0280.296Relationship25.6%0.4490.2750.2850.0380.706
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.402, + 0.893, + 0.46 + ], + "angle": 0, + "content": "Table 3. Metaphorical image captioning (left) and visual question answering (right) performance of PaLI [10]. We report different metrics (higher the better) using both the exact match accuracy (Acc) and standard text generation metrics. For image captioning, we evaluate the whole predicted caption as well as parsed primary object, secondary object, and relationship. For VQA, we evaluate on all predicted answers as well as provide the breakdown for each question type." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.473, + 0.829 + ], + "angle": 0, + "content": "tioning task, where the input is an image and the target is the metaphorical message in the syntactic structure is as as . We fine-tune and evaluate the state-of-the-art literal image caption model PaLI-17B [10] based on the exact match accuracy (maximum over all references) and standard metrics for image captioning BLEU4 [36], ROUGE-L [28], METEOR [3], SPICE [2], and CIDEr [51]. Since the target captions follow the fixed syntactic structure, we parse each predicted caption into the primary concept, the secondary concept, and their relationship, and use the same set of metrics for the whole caption to evaluate. Table 3 (left) summarizes the results and Figure 4 provides sample qualitative results. We observe that PaLI generally struggles on this task. For instance, it achieves a CIDEr score of 1.076, compared to 1.491 for state-of-the-art literal image captioning on the popular COCO-Captions [8]. Further, the model struggles with predicting the target relationship when the metrics favor recall (e.g., ROUGE-L) and with predicting the target secondary concept when the metrics favor precision (e.g., BLEU)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.833, + 0.471, + 0.88 + ], + "angle": 0, + "content": "Visual Question-Answering (VQA). We propose metaphorical open-ended (i.e., not vocab-based) VQA task, where the input is an image and a given question, and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.498, + 0.893, + 0.695 + ], + "angle": 0, + "content": "the target is the answer. We use fixed templates to generate 2 VQA questions whose answer is the primary concept, 2 for secondary concept, and 2 for relationship. Again, we fine-tune and evaluate the state-of-the-art literal VQA model PaLI-17B [10], using the same set of metrics as in image captioning. Table 3 (right) summarizes the results and Figure 4 provides sample qualitative results. Overall, we find that PaLI performs poorly, only achieving the average accuracy score of \\(19.9\\%\\), while the state-of-the-art literal VQA on the popular VQAv2 [18] benchmark is \\(77.6\\%\\) on \"other\" questions. Additionally, the model struggles the most with answering questions that ask for the secondary concept." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.715, + 0.725 + ], + "angle": 0, + "content": "3.3. Metaphor Localization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.894, + 0.838 + ], + "angle": 0, + "content": "Here, the task is to localize the image regions that invoke either the primary or secondary concept in the viewer. This is similar to the phrase grounding task of localizing objects using free-form natural language phrases [23, 31, 55], but with some key differences due to the peculiarities of visual metaphors in comparison to literal images used in standard vision datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Annotations. As discussed earlier, there are diverse types of metaphors based on how the primary and secondary concepts are visually depicted in an image (see Fig. 2). There are at least two key differences in metaphor localization" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.306, + 0.901 + ], + "angle": 0, + "content": "The score of 0 is given if parsing fails." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23206" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.275, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.202, + 0.241, + 0.212 + ], + "angle": 0, + "content": "The shoe is as light as feathers." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.092, + 0.473, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.202, + 0.441, + 0.213 + ], + "angle": 0, + "content": "Ketchup is as hot as a volcano." + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.092, + 0.68, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.197, + 0.684, + 0.219 + ], + "angle": 0, + "content": "Getting two ice creams is as exciting as getting two movie tickets for one." + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.092, + 0.88, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.198, + 0.857, + 0.219 + ], + "angle": 0, + "content": "This car is as out of this world as three dimensional." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.894, + 0.274 + ], + "angle": 0, + "content": "Figure 5. Sample Localization annotations showing annotated bounding boxes around primary and secondary concepts. Notice the diversity of types in bounding boxes: explicit, contextual, logo and texts. This makes metaphor localization more challenging compared to standard object detection." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.302, + 0.266, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.4, + 0.266, + 0.419 + ], + "angle": 0, + "content": "Parking assist is as safe as parking far from objects." + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.303, + 0.454, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.4, + 0.449, + 0.419 + ], + "angle": 0, + "content": "This chocolate is as fun as going to a concert." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.433, + 0.47, + 0.475 + ], + "angle": 0, + "content": "Figure 6. Sample Localizations with the phrase-grounding model from [27], where the secondary concepts are contextual. GT boxes are shown in green, whereas the predictions are shown in blue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.868 + ], + "angle": 0, + "content": "compared to standard localization in literal images: 1. A given concept can be present in the image either explicitly or in a contextual manner (for e.g., contextual visual metaphors in Fig. 2). 2. Visual metaphor Ads are inherently multimodal and a concept can be invoked by other modalities such as text or logo. See the multimodal visual metaphors in Fig. 2. As a result, we not only annotate the bounding boxes that invoke the primary/secondary concept in the viewer, but we also annotate the type of that bounding box. A bounding box can be of one of the 4 types: Explicitly present, Contextually present, Logo or a Text. Specifically, for each of the 5061 metaphorical images, we pick the best metaphor annotation (primary, secondary concepts and their relationship) according to their validation scores (see previous section) and collect bounding box annotations for both the primary and secondary concepts. We collect all the bounding boxes that invoke both the primary and secondary concepts and also their type (explicit, contextual, logo or text) for each of the images. We use 5 annotators for each annotation and choose the bounding boxes with the best inter-annotator agreement. We use the same train and test splits as used in understanding tasks. Fig. 5 shows sample localization annotations. We collected over \\(30\\mathrm{k}\\) bounding box annotations for this task on 5061 metaphor images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Detection. Recent detection and localization models [27, 58] pre-trained on image and caption pairs are shown to" + }, + { + "type": "table", + "bbox": [ + 0.574, + 0.298, + 0.821, + 0.356 + ], + "angle": 0, + "content": "
mAP50mAP70
Primary concept33.2214.25
Secondary concept43.5431.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.367, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Table 4. Localization results with CLIP based phrase localization model [27] on our test split." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.682 + ], + "angle": 0, + "content": "achieve remarkable localization performance on discriminating fine-grained objects and unseen concepts. Specifically, we evaluate [27] which leverages the effective image representations in CLIP by extracting spatial features from it. Using these spatial features, for each pixel location, the model computes the inner product between the spatial feature and the phrase embedding extracted from CLIP to predict the bounding box. In our case, we pass the primary or secondary concept as input phrase to [27] the estimate the corresponding bounding box. Table 4 summarizes the detection results (using Mean Average Precision) on our test split. We find relatively better performance in localizing secondary objects compared to primary objects. We show few qualitative results in Figure 6. It is worth noting that our collected annotations allow for more comprehensive analysis on localization tasks due to the availability of different types of bounding boxes (explicit, contextual, logos and texts)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.706, + 0.71 + ], + "angle": 0, + "content": "3.4. Metaphor Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Recent large-scale text-to-image (T2I) generative models show remarkable success in generating highly realistic images from text prompts. Can these models also work well in metaphorical image generation? We evaluate two state-of-the-art generative models (Imagen [42], Stable diffusion [41] (SD)) using 300 samples from MetaCLUE test set where we use the text prompts: \"An advertisement where primary-concept is as relationship as secondary concept.\" In addition, we finetune the stable diffusion model on our train split (same split as in Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Results. Fig. 7 shows sample visual results from different T2I models, with the metaphor annotation shown on the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23207" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.418, + 0.155 + ], + "angle": 0, + "content": "
ModelFID ↓CLIP Similarity ↑
Imagen [42]153.132.1
Stable Diffusion [41]161.630.8
Stable Diffusion - FT154.332.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.164, + 0.47, + 0.193 + ], + "angle": 0, + "content": "Table 5. Analysis of Image Generation results with standard metrics of FID [42] and CLIP similarity [39] scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.474, + 0.513 + ], + "angle": 0, + "content": "top. The generated images capture different aspects of the metaphor (tablet, waterproof), but not the entire metaphorical message. We compute standard metrics to automatically evaluate the quality of the generations. Tab. 5 shows the standard FID [19] and CLIP-Similarity [39] scores of different models. FID score evaluates the image distribution similarity between the generation images with the corresponding real image distribution. FID scores in Tab. 5 shows that Imagen performs slightly better than SD. And, there is a slight improvement in FID with finetuning (SDFT vs. SD). In general, high FID scores in Tab. 5 indicate the large distribution gap between the generated and real images. CLIP similarity score, on the other hand, measures the prompt fidelity - similarity between the generated image and the corresponding input text prompt according to the CLIP model [39]. CLIP similarity scores in Tab. 5 follow the similar trend as FID scores: Imagen performs better than SD in terms of the prompt fidelity, and finetuning SD model improved its prompt fidelity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given that both FID and CLIP scores are not tailored towards metaphorical images, these metrics are not reliable in assessing whether the generated images capture the essence of visual metaphors. To analyze this, we perform human studies comparing two different models at a time. Specifically, we show a metaphorical message and the corresponding generations from two models and ask the users to pick an image that best conveys that metaphorical message. We obtain 7 user ratings for each image pair and consider 3-out-of-7 or 4-out-of-7 to be ties. Fig. 8 shows the percentage of user preferences across different pairs of results. User studies also indicate that Imagen performed better than SD. An interesting finding is that finetuning SD resulted in slightly worse user preference compared to base SD model (SDFT vs. SD). This is in contrast to standard FID and CLIP metrics that improved with finetuning. We hypothesize that finetuning SD resulted in more realistic Ad images, but the resulting model may have forgotten some of the metaphorical priors. This calls for more effective finetuning strategies with the limited training datasets, which forms an important future work. In addition, both Imagen and SD performed quite poorly compared to real images in conveying metaphorical messages. Real images are preferred around \\(88\\%\\) of time over Imagen results. This illustrates the big scope of improvements in generating visual metaphors." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.086, + 0.854, + 0.109 + ], + "angle": 0, + "content": "Metaphor: This android tablet is as waterproof as someone in a swimtube." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.114, + 0.667, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.109, + 0.876, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.265, + 0.691, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.266, + 0.876, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.431, + 0.895, + 0.473 + ], + "angle": 0, + "content": "Figure 7. Sample Image Generations for a given metaphorical message (shown on top) with Imagen [42], Stable Diffusion [41] and fine-tuned (FT) version of Stable Diffusion." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.492, + 0.889, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.597, + 0.893, + 0.655 + ], + "angle": 0, + "content": "Figure 8. User Study on Image Generation Results. Percentage of results users preferred across real, Imagen [42], Stable Diffusion (SD) [41] and its fine-tuned version (SD-FT) results. Users are asked to choose the image that better depicts a given metaphor." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.62, + 0.693 + ], + "angle": 0, + "content": "4. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.9 + ], + "angle": 0, + "content": "In this paper, we present a step towards comprehensive evaluation of progress on visual metaphor research. Specifically, we propose a collection of tasks related to comprehending and generating visual metaphors using AI techniques. Our MetaCLUE tasks include Classification, Understanding (Retrieval, Captioning, VQA), Localization and Generation. For comprehensive evaluations, we collected high quality and rich annotations that facilitate the measurable progress. Existing methods demonstrate poor results in many cases with our experimental analysis shedding light on strengths and drawbacks of different approaches paving a path for future research in this fascinating field." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23208" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569-11579, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. SPICE: semantic propositional image caption evaluation. In ECCV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.469, + 0.269 + ], + "angle": 0, + "content": "[3] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In ACL Workshops, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.442, + 0.284 + ], + "angle": 0, + "content": "[4] Roland Barthes. Image-music-text. Macmillan, 1977. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.286, + 0.469, + 0.34 + ], + "angle": 0, + "content": "[5] Nathalie Bonnardel and Evelyne Marmèche. Towards supporting evocation processes in creative design: A cognitive approach. International journal of human-computer studies, 63(4-5):422-435, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.342, + 0.47, + 0.383 + ], + "angle": 0, + "content": "[6] Robert W Boozer, David C Wyld, and James Grant. Using metaphor to create more effective sales messages. Journal of services marketing, 1990. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[7] Tuhin Chakrabarty, Xurui Zhang, Smaranda Muresan, and Nanyun Peng. Mermaid: Metaphor generation with symbolism and discriminative decoding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4250-4261, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO Captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[9] Xianyang Chen, Chee Wee Leong, Michael Flor, and Beata Beigman Klebanov. Go figure! multi-task transformer-based architecture for metaphor detection using idioms: Ets team in 2020 metaphor shared task. In Proceedings of the second workshop on figurative language processing, pages 235-243, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[11] Lydia B Chilton, Savvas Petridis, and Maneesh Agrawala. Visiblends: A flexible workflow for visual blends. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[12] Minjin Choi, Sunkyung Lee, Eunseong Choi, Heesoo Park, Junhyuk Lee, Dongwon Lee, and Jongwuk Lee. Melbert: Metaphor detection via contextualized late interaction using metaphorical identification theories. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1763-1773, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[14] Gilles Fauconnier and Mark Turner. The way we think: Conceptual blending and the mind's hidden complexities. Basic books, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] Charles Forceville. Pictorial metaphor in advertising. Routledge, 1996. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Susan R Fussell and Mallie M Moss. Figurative language in emotional communication. Social and cognitive approaches to interpersonal communication, pages 113-141, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[17] Hongyu Gong, Kshitij Gupta, Akriti Jain, and Suma Bhat. Illinimet: Illinois system for metaphor detection with contextual and linguistic information. In Proceedings of the Second Workshop on Figurative Language Processing, pages 146-153, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[18] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[20] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic understanding of image and video advertisements. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1705-1715, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[21] Bipin Indurkhya. On the role of metaphor in creative cognition. In ICCC, pages 51-59, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[22] Bipin Indurkhya and Amitash Ojha. An empirical study on the role of perceptual similarity in visual metaphors and creativity. Metaphor and Symbol, 28(4):233-253, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[23] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[24] George Lakoff. The contemporary theory of metaphor. 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[25] George Lakoff and Mark Johnson. Metaphors we live by. University of Chicago press, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[27] Jiahao Li, Greg Shakhnarovich, and Raymond A Yeh. Adapting clip for phrase localization without further training. arXiv preprint arXiv:2204.03647, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23209" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[29] Todd I Lubart and Isaac Getz. Emotion, metaphor, and the creative process. Creativity research journal, 10(4):285-301, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[30] Carita Lundmark. Metaphor and creativity in British magazine advertising. PhD thesis, Luleå teknska universitet, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[31] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.291 + ], + "angle": 0, + "content": "[32] James H Martin. A corpus-based analysis of context effects on metaphor comprehension. Trends in Linguistics Studies and Monographs, 171:214, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.47, + 0.333 + ], + "angle": 0, + "content": "[33] Zachary J Mason. Cormet: a computational, corpus-based conventional metaphor extraction system. Computational linguistics, 30(1):23-44, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[34] Edward F McQuarrie and David Glen Mick. Visual rhetoric in advertising: Text-interpretive, experimental, and reader-response analyses. Journal of consumer research, 26(1):37-54, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[35] Ekaterina Ovchinnikova, Vladimir Zaytsev, Suzanne Wertheim, and Ross Israel. Generating conceptual metaphors from proposition stores. arXiv preprint arXiv:1409.7619, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[37] Barbara J Phillips. The impact of verbal anchoring on consumer response to image ads. Journal of advertising, 29(1):15-24, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[38] Barbara J Phillips and Edward F McQuarrie. Beyond visual metaphor: A new typology of visual rhetoric in advertising. Marketing theory, 4(1-2):113-136, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[40] Sunny Rai and Shampa Chakraverty. A survey on computational metaphor processing. ACM Computing Surveys (CSUR), 53(2):1-37, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[43] Linda M Scott. Images in advertising: The need for a theory of visual rhetoric. Journal of consumer research, 21(2):252-273, 1994. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[44] Gerard Steen. A method for linguistic metaphor identification: From MIP to MIPVU, volume 14. John Benjamins Publishing, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[45] Kevin Stowe, Tuhin Chakrabarty, Nanyun Peng, Smaranda Muresan, and Iryna Gurevych. Metaphor generation with conceptual mappings. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6724-6736, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[46] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[47] Asuka Terai and Masanori Nakagawa. A computational system of metaphor generation with evaluation mechanism. In International Conference on Artificial Neural Networks, pages 142-147. Springer, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[48] Xiaoyu Tong, Ekaterina Shutova, and Martha Lewis. Recent advances in neural metaphor processing: A linguistic, cognitive and social perspective. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4673-4686, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[49] Peter Turney, Yair Neuman, Dan Assaf, and Yohai Cohen. *Literal and metaphorical sense identification through concrete and abstract context*. In *Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing*, pages 680–690, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[50] Tony Veale, Ekaterina Shutova, and Beata Beigman Klebanov. Metaphor: A computational perspective. Synthesis Lectures on Human Language Technologies, 9(1):1-160, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[51] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[52] Merryl J Wilkenfeld and Thomas B Ward. Similarity and emergence in conceptual combination. Journal of Memory and Language, 45(1):21-38, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[53] Bo Xu, Tingting Li, Junzhe Zheng, Mehdi Naseriparsa, Zhehuan Zhao, Hongfei Lin, and Feng Xia. Met-meme: A multimodal meme dataset rich in metaphors. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2887-2899, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[54] Keren Ye and Adriana Kovashka. Advise: Symbolism and external knowledge for decoding advertisements. In Proceedings of the European Conference on Computer Vision (ECCV), pages 837-855, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[55] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[56] Zhiwei Yu and Xiaojun Wan. How to avoid sentences spelling boring? towards a neural approach to unsupervised metaphor generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23210" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 861-871, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[57] Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, and Hongfei Lin. Multimet: A multimodal dataset for metaphor understanding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3214-3225, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23211" + } + ] +] \ No newline at end of file diff --git a/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_origin.pdf b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d264ebdcf97dce271c821e78cabb1181f9c86f48 --- /dev/null +++ b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/5a097123-23c7-455f-934d-5ab8f12c4f80_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ada23a384ff1454d2047f793a7b3bf9f367b30d9d875d5a7f06fe6d4fa57de0 +size 2686293 diff --git a/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/full.md b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/full.md new file mode 100644 index 0000000000000000000000000000000000000000..487c657a1820b70cbaae0cf1f8164b52d4393eb8 --- /dev/null +++ b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/full.md @@ -0,0 +1,318 @@ +# MetaCLUE: Towards Comprehensive Visual Metaphors Research + +Arjun R. Akula*, Brendan Driscoll*, Pradyumna Narayana, Soravit Changpinyo, Zhiwei Jia, Suyash Damle, Garima Pruthi, Sugato Basu, Leonidas Guibas, William T. Freeman, Yuanzhen Li, Varun Jampani* Google + +![](images/209a120a3d720efd6732ac818466b1da66b20f7edd8f48d41c73de8e0c48d583.jpg) +Metaphor: Killing trees is as harmful as killing wildlife. + +# Classification + +Is this a visual metaphor? - YES - NO + +# Localization + +Detect image regions that invoke the concepts: + +- Killing trees +- killing wildlife + +![](images/a639cb28fbb5f4dd8079fbda6c71370593e24e033a72a38e4a7efb389868de45.jpg) +Figure 1. With MetaCLUE, we introduce several interesting tasks related to visual metaphors. We collect metaphor annotations (objects, abstract concepts, relationships and object boxes) for evaluating existing models on these tasks. Specifically we perform a comprehensive evaluation of vision and language models on four different tasks (Classification, Localization, Understanding, and gEneration). Comprehensive experiments in this work show that state-of-the-art techniques mostly focus on literal interpretation and perform poorly in understanding and generation of metaphor images. + +# gEneration + +Prompt: "An advertisement where killing trees is as harmful as killing wildlife." + +![](images/6401876215261697f19124b169d600cb701db81e91126ced2a0c158e7de52e1f.jpg) +Stable Diffusion + +![](images/9ccc5ac901c561ad3aca7d3018ac7e221c233c7aa31b53e61fded73e4225eae9.jpg) +Imagen + +# Understanding + +Retrieval + +Pick the right one: + +(a) Killing the forest is as deadly as killing the animals too. +(b) Birds is as much a part of our world as used cans. + +Captioning + +Sample predictions: + +1. Deforestation is as damaging as killing wildlife. +2. Deforestation is as bad as ending the death penalty. + +Visual Question Answering Sample Questions: + +Q. What is as harmful as killing wildlife? +Q. What is compared to killing wildlife? + +# Abstract + +Creativity is an indispensable part of human cognition and also an inherent part of how we make sense of the world. Metaphorical abstraction is fundamental in communicating creative ideas through nuanced relationships between abstract concepts such as feelings. While computer vision benchmarks and approaches predominantly focus on understanding and generating literal interpretations of images, metaphorical comprehension of images remains relatively unexplored. Towards this goal, we introduce MetaCLUE, a set of vision tasks on visual metaphor. We also collect high-quality and rich metaphor annotations (ab + +struct objects, concepts, relationships along with their corresponding object boxes) as there do not exist any datasets that facilitate the evaluation of these tasks. We perform a comprehensive analysis of state-of-the-art models in vision and language based on our annotations, highlighting strengths and weaknesses of current approaches in visual metaphor classification, localization, understanding (retrieval, question answering, captioning) and generation (text-to-image synthesis) tasks. We hope this work provides a concrete step towards developing AI systems with human-like creative capabilities. Project page: https://metaclue.github.io + +# 1. Introduction + +"Metaphor is pervasive in everyday life ... Our ordinary conceptual system, in terms of which we both think and act, is fundamentally metaphorical in nature." — Lakoff & Johnson [25] + +Creativity is a process of generating a new perspective on a problem or a situation. Metaphorical thinking has been recognized as a key and powerful mechanism of creativity [24,29,50]. Humans engage metaphors in their creative thinking process as strategies to link or blend concepts, or to view a concept from a target domain in terms of another, apparently dissimilar concept from a source domain [25]. metaphors also provide a sophisticated tool for nuanced human communication. Let us take a closer look at the structure of metaphors – and especially visual metaphors. + +Metaphors1 are a cognitive construct in which a concept is compared to a seemingly unrelated concept via some shared attribute. Take as an example 'This car is a cheetah', where 'This car' is compared to 'a cheetah' in terms of speed. Metaphors have a simple syntactic structure of 'A is B' where A is referred to as the primary concept and B as the secondary concept. The implied analogy in a metaphor is of the form: ' (primary concept) is as (relationship)2 as (secondary concept)' and often involves an attribute transfer from the secondary to the primary concept. Some examples include 'This phone is as fast as a rocket', 'Cigarettes are as harmful as bullets' etc. The primary and secondary concepts are usually unrelated at a glance, resulting in an element of surprise and creativity in metaphorical expressions. Despite following such simple structure, metaphors are quite powerful in conveying creative ideas. Metaphors are pervasive in all forms of communication, such as speech, text, visual etc. + +Visual Metaphors are images where the primary and secondary concepts are visually depicted in an image conveying the metaphorical message to the viewers. Visual metaphors are widely used in mass media communications like advertising and journalism [15, 43, 45]. In this work, we work with Ad images, as metaphors tend to be prevalent in ads. There are numerous ways a metaphor can be represented visually. Following the classification in [15], there are at least 4 different types of visual metaphors. Fig. 2 shows sample images that belong to these types along with our annotations of primary, secondary concepts and their relationship. In contextual metaphors, either the primary or secondary concept is not explicitly visible, but is inferred from the context (e.g., apple in the left-most image). In hybrid metaphors, the primary and secondary concepts are + +visually conflated. Juxtaposition forms one of the simplest visual metaphor types, where the two concepts are just presented next to each other. Multimodal metaphors represent one of the concepts with another modality, such as text or logo. In practice, visual metaphors use multiple of these strategies to convey a metaphor in an effective manner. In many cases, the implied metaphorical meaning is somewhat open-ended. Interpretation of visual metaphors depends on several external factors, such as familiarity with the brands and cultural context. + +These visual variations and nuances make automatic cognition or generation of visual metaphors highly challenging. While the last decade has seen rapid progress in many areas of understanding and generation tasks, prior works in computer vision focus heavily on literal interpretation of images and overlook the importance of metaphorical reasoning in understanding the image message [1, 48]. We believe that developing AI systems with metaphorical comprehension and generation capabilities can greatly assist humans in creative endeavors involving conveying concepts in new and exciting ways. Such systems provide an important step towards conferring human-like creativity to AI models. + +To this end, we introduce multiple interesting tasks and construct metaphor annotations that enable comprehensive research on visual metaphors. As metaphors are more common in visual Ads, we start with the Pitt's Ads dataset images [20] and then perform a rigorous multi-stage annotation process with expert annotators to filter metaphorical images, add metaphor annotations, and perform additional validation steps to clean the annotations3. While there is recent work making advances in understanding non-literal interpretations in natural language research [7, 12], this work proposes the first step towards metaphor analysis in images. + +As illustrated in Fig. 1, we perform comprehensive evaluations with state-of-the-art techniques on four sets of tasks, which we call MetaCLUE: 1. Classification: This is binary classification task of estimating whether a given image contains a metaphor or not. In other words, Are visual features indicative of whether there exists a metaphor in a given image or not?. 2. Localization: Here, the task is to localize the image regions that invoke the primary and secondary concepts in the viewer. This is similar to a standard object detection task, but is more complicated in the case of visual metaphors as the primary/secondary concepts may not be explicitly present in an image. 3. Understanding: Can our models understand the metaphorical message in a given image? We pose this understanding problem as 3 tasks where we can quantitatively measure the performance: Retrieval, Captioning and Visual question answering. 4. gEneration: Can we generate an image that conveys the metaphor, given the metaphorical message as a text prompt? + +![](images/fbd965311b0c74e4ca2dd41001132f6b7d4a01a2b7a6234d78a30650f44e892d.jpg) +Contextual +This beer is as tasty as a real apple. + +![](images/aa12fc950b3973b9820cec40425dcc1988ea0c7bdb1e4c4afaa801b13b84dfd5.jpg) +This car is as adventurous as a space ship. + +![](images/0861d88e2a3d360853e8c299b9877fc03857a240952a8effc8420dd5c0929bd4.jpg) +Hybrid +Driving this SUV is as smooth as birds flying in the sky. + +![](images/32d41a06a69b86d55b19d50bccf4b6b508bbadc17eee7aeca30e1db324fb7627.jpg) +This pencil is as red as a fire truck. + +![](images/fba8eda3cf31a78df732cc054f9bdd677dfe2575d1766b800aa5f980c5547ec6.jpg) +This chocolate bar is as rich as gold. +This car is as made for the beach as a crab. +Figure 2. Sample Visual Metaphors with their Annotations. There are different types of visual metaphors. The type depends on how the primary and secondary concepts are visually depicted. Here are sample Ad images from [20] where we annotated the primary concept, secondary concept and their relationship. + +![](images/fc19b684838be1c1a648bc9ae41f9b69669726d22fad4e4302152277af5a54b5.jpg) +Juxtaposition + +![](images/360d096004498d375b4996e1c2b82bf050d23ce8ce98d5ff58237b8e06e194b7.jpg) +These donuts are as unique as as talking people. + +![](images/57232ccf52b7920f5087f9ef19437801071cf4ba62848cd162eb574cd3373a43.jpg) +Multimodal +The car is as rugged as this muddy trailer. + +We comprehensively evaluate existing state-of-the-art techniques for each of these tasks on our collected metaphor annotations. We evaluate the models both in a zero-shot manner as well as with finetuning on our annotations. Even though finetuning resulted in some improvements, most models struggle to produce satisfactory results in many cases, demonstrating the difficulty of these tasks. Our experiments highlight several strengths and weaknesses of the existing techniques on comprehending and generating visual metaphors, providing a concrete first step towards further AI research on this fascinating topic. + +# 2. Related Work + +Creativity and Metaphorical Abstraction. Creativity often involves an innovative fusion of objects, attributes, or relationships from previous knowledge to generate new concepts [5, 52]. Metaphors can serve as an invaluable tool for expressing creative insights and also to stimulate new ones [14, 21]. The cognitive research community has made initial attempts in understanding different realizations of metaphors such as language metaphors and visual metaphors [15, 16, 32]. Studies of visual persuasion show that visual metaphors may be more effective than language metaphors in terms of producing a greater degree of cognitive creativity [34]. In addition to improving creativity, metaphors are also known to elicit pleasure since the initial ambiguity in (re-)conceptualizing a target entity in terms of a source stimulus generates interest and motivation, and the subsequent resolution is rewarding – explaining the impor + +tance of using creative metaphorical processes in art, advertising, and marketing [6,30,38]. + +Metaphors in Language Research. Computational Linguistic studies show that metaphors are ubiquitous in language, occurring once per three sentences on average [40, 44, 49]. Recently, an increasing number of research efforts have explored the limitations and challenges in detecting and decoding the meaning of language metaphors [7, 48]. While there exist several computational models [9, 12, 17, 33, 35, 45, 47, 56] for metaphor identification, interpretation and generation in language, there exists very little work on computational modeling of visual metaphors. + +Metaphors in Computer Vision. Much of computer vision literature is focused on understanding and generating literal images. Automatic metaphorical interpretations of images is highly challenging and requires multi-faceted cognitive reasoning that involves visual reasoning coupled with the use of external knowledge. Recent work on affective captioning using the ArtEmis data set [1] includes some captions that evoke metaphors to explain emotions, but that dataset is focused on visual art and does not specifically consider metaphors. There is no explicit prior work for comprehensive evaluation and development of models that can automatically comprehend or generate visual metaphors. At the same time, there exist several studies that demonstrate the potential of visual metaphors. For instance, some studies [4,37] suggest that advertisements containing metaphorical images will be more persuasive compared to ads with language metaphors, non-metaphorical + +![](images/cd66b94d5c2c2b6dead4a5348fa6d5db897abec5c905b7e5548545e2f80e0212.jpg) +Figure 3. Distribution of Topics in our annotated metaphorical images from Pitt's Ad dataset. + +ads, or literal images. There have been some prior computational models [11, 22] for metaphor generation, but they are not thoroughly validated against any benchmark datasets or user studies. Some works [20, 54] propose datasets and techniques for general Ad image understanding with a focus on the challenging aspects of non-literal interpretations in Ad images. However they do not explicitly collect any metaphorical annotations, nor do they provide the corresponding analysis. In this work, we start with an existing Ad image dataset [20] and perform extensive human studies to filter metaphorical images and collect detailed annotations accounting various aspects of metaphoric interpretation. Prior works such as Multi-MET [57] and MET-Meme [53], propose metaphorical annotations but does not annotate the relationship between primary and secondary. In MetaCLUE, in addition to providing relationship annotations, we also collect detailed bounding box annotations that help localize the image regions invoking the primary, secondary concepts in the viewer. We further provide VQA style question and answers. + +# 3. MetaCLUE + +We introduce four different high-level tasks in MetaCLUE that enable comprehensive evaluation and development of visual metaphor research: Classification, Localization, Understanding and gEneration. In the rest of this section, we first describe our annotation collection process for the tasks and next provide the analysis of using existing state-of-the-art techniques for each of these tasks. + +# 3.1. Metaphor Classification + +Following the fundamental vision task of image classification, we first ask: Can we develop models that can classify whether or not a given image contains a visual metaphor? In other words, is it possible to just use visual cues to estimate whether or not there exists a metaphorical interpretation of an image? + +Annotations. For this task, we need to label whether or + +not each image contains a metaphor. Since we tend to see more metaphorical images in Ads, we start with images from an existing Ads dataset published by the University of Pittsburgh [20] and manually annotate whether a given Ad image contains a metaphor or not. Pitt's Ads dataset contains images of both product ads (e.g. phone ads) as well as public service announcements (e.g. forest conservation ads). Concretely, to make annotations more efficient, we use a subset of 8.5K Ad images from this dataset that are annotated to have 'symbolic' (fun, adventurous, etc.) references. We find this subset to contain a considerable portion of metaphors. Specifically, we collect 5 Yes/No annotations for each of these 8K images from different annotators. To obtain high-quality annotations, we explained what a visual metaphor is to the annotators and also conducted qualifying exams to pick final annotators for this task. We consider the images with 3 or more 'Yes' annotations as visual metaphors and the remaining as non-metaphorical. At the end, we identify 5061 images containing visual metaphors out of 8480 images. Pitt's Ads dataset [20] also comes with topic annotations (e.g., restaurant, car, animal rights etc.). Fig. 3 shows the word cloud plot of different topics in these 5061 metaphorical images. We split the metaphorical images into 3730 train and 1331 test images by maintaining the same distribution of topics in both the splits. We have two types of negative sets (non-metaphorical images) for classification experiments. One is formed by the remaining 3419 non-metaphorical images in the symbolic set, and another one is created separately by annotating an additional 3000 literal (and non-symbolic) images from Pitt's Ads dataset [20]. We add 2000 of 3419 symbolic negative images to our train split, and add the remaining 1419 images to our test split. Similarly 2000 of 3000 literal negatives are used in training, and the remaining 1000 are used for testing. + +Evaluation and Results. Using our collected metaphor images, we evaluate the performance of the following state-of-the-art models in classifying an input image as metaphor: EfficientNet [46] and Vision Transformer (ViT) [13]. We fine-tune these models to classify metaphor vs. symbolic/non-metaphors and metaphor vs. literal, and use $20\%$ of corresponding train splits for validation. Test results are reported in Table 1. Although the performance of ViT is significantly better than EfficientNet and random baselines, there is still ample room for improvement. In particular, models find it easier to distinguish metaphor images from literal images, and struggle to identify metaphors within the symbolic image pool. + +# 3.2. Metaphor Understanding + +We now describe how we collect annotations that help in capturing the metaphorical message from the images collected in previous section. + +
Model#ParamsSymbolic Neg.Literal Neg.
ValTestValTest
RandomN/A63.1051.6060.6657.12
EfficientNet-B05.3M60.7649.6770.9450.30
EfficientNet-B766M61.4448.5469.8449.82
ViT-B/1686M69.3166.9884.0481.24
ViT-L/16307M65.8360.6581.4580.52
+ +Table 1. Accuracy of Metaphor Classification (binary classification accuracy) using state-of-the-art classification architectures of EfficientNet [46] and ViT [13]. + +Annotations. We provide detailed instructions and several examples to the annotators to help them annotate primary and secondary concepts in the metaphor and also the characteristic/relationship that is transferred from secondary to the primary. We conduct multiple pilot studies to reduce the noise and to improve inter-annotator agreement. Enforcing the annotators to make sure that their annotations are linguistically readable in the following syntactic structure helped us in improving quality and consistency of the annotations: “__ is as __ as __,” where the first blank is the primary concept, the second blank is the relationship, and the third blank is the secondary concept. Figure 2 shows some examples of these annotations. We collect 5 metaphor annotations for each image. As interpretations of metaphorical images can be highly subjective, there can exist more than one interpretation for each image, which makes it difficult to automatically remove noisy annotations. Therefore we conduct an additional human study where we show each of the annotation to five annotators and ask them to verify the correctness along three dimensions: (a) Is the grammar correct?; (b) Are primary and secondary concepts correct?; and (c) Is the relationship correct?. We remove annotations with a low number of votes out of 5 along each of the three dimensions, resulting in a total of over 26k clean annotations. We evaluate state-of-the-art models in understanding metaphorical message from the input images using 3 tasks namely, Retrieval, Captioning and Visual question answering. + +Retrieval. The goal of this task is to retrieve the correct metaphor interpretation/statement from a candidate set given an image. In our candidate set, we choose exactly one positive (correct) metaphorical statement from its ground truth messages and uniformly sample $K - 1$ random negative statements from other images. Table 2 shows the results obtained with CLIP [39] and ALBEF [26] for $K = 50$ , reporting retrieval precision@1 and rank4. Although CLIP ViT-L/14 shows good zero-shot performance on random negatives with more than $76\%$ accuracy, we observe a large drop + +![](images/8ca14d96a149751ef91b3f5bcb6263b959bd8dd54231fe77feafd9042acb5e98.jpg) +Metaphor: Smoking cigarettes is as life-shortening as sharpening a pencil + +Captioning Result Prediction:Smoking is as dangerous as burning a pencil. +Visual Question Answering Result Q1: What is a smoker's life compared to a sharpened pencil? +Pred: Sharp GT: Shortened Q2: What is used as a visual metaphor for a sharpened pencil? +Pred: Not smoking GT:Smoker's life + +![](images/18d9bca6907411565d11d98da02daafa75f25b1aa588b96755fde662c2f44e7d.jpg) +Metaphor: Smoking is as deadly as injecting drugs. +Figure 4. Results of PaLI [10] for captioning and visual question answering on sample images in our test split. + +Captioning Result Prediction: Drinking and driving is as bad as injecting drugs. +Visual Question Answering Result Q1: What is a cigarette as compared to injecting drugs? +Pred: Scary GT: Deadly Q2: What is used as a visual metaphor for injecting drugs? +Pred:Drinking GT:Cigarette + +in performance as we increase $K$ to $\{100, 500, 1000\}$ . We further fine-tune CLIP models using $70\%$ of metaphor annotations as the train set and see gains by up to $+7$ absolute points in $p@1$ . In summary, the performance of models is impressive with less than 50 negative candidates whereas the performance drops greatly by increasing $K$ . + +We hypothesize that models might simply be looking at salient objects rather than comprehending the underlying semantics of metaphor in finding the correct candidate. To test this, we mine hard negative (HN) statements and use them as our candidate set. Specifically, we construct the following five types of HNs: (a) Neg Prim: candidates obtained by replacing the primary concept in the metaphor statement with the primary concept from another image; (b) Neg Sec: replacing secondary concept likewise; (c) Neg Prim+Rel: replacing primary and relationship; (d) Neg Sec+Rel: replacing secondary and relationship; (e) Swap Prim&Neg: swapping primary and secondary from the same image. Table 2 shows the results. We see a significant drop of up to $30\%$ by using HNs as negative statements for $K = 50$ , indicating the difficulty in comprehending and distinguishing metaphorical abstraction of concepts. We find performance of Neg Prim is significantly higher than Neg Sec, suggesting that the models tend to rely more on primary object than secondary object in identifying correct interpretation. Overall, there is ample room for improvement indicated by the steep drop in model performance with HNs. + +Captioning. Here, we propose metaphor image cap + +
ModelRandom NegNeg PrimNeg SecNeg Prim+RelNeg Sec+RelSwap accuracy↑
p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓
CLIP (ViT-B/16)70.973.4946.6710.1138.1413.9749.368.8042.2511.6940.61
CLIP (ViT-B/32)61.784.1938.7411.6033.2014.8439.7910.2936.2712.7141.28
CLIP (ViT-L/14)76.663.1751.759.2239.8613.1954.747.7045.9911.0243.08
ALBEF39.798.5727.0016.5329.3115.7926.7715.5828.4214.5946.67
ALBEF (MSCOCO)44.877.5531.7815.1331.4114.0932.5314.0433.2812.8548.24
ALBEF (Flickr30k)47.498.7735.6014.5635.2213.6636.3513.5836.1212.7949.81
FT CLIP (ViT-B/16)76.812.2549.819.6545.4710.2453.408.0750.638.1844.65
FT CLIP (ViT-B/32)68.062.8243.0010.8239.6411.2344.729.3943.459.3545.69
FT CLIP (ViT-L/14)81.751.7857.667.4849.069.4061.255.9957.067.3343.75
+ +Table 2. Performance of retrieval models on $K$ random (column 2) and hard negative candidates (columns 3-7) ( $K = 50$ for columns 2-6, $K = 2$ for last column). + +
CaptioningAccBLEU4ROUGE-LMETEORSPICECIDErVQAAccBLEU4ROUGE-LMETEORSPICECIDEr
Whole caption1.1%0.2540.5360.2200.1861.076All questions19.9%0.3290.2860.2490.1850.851
Primary29.9%0.3270.4070.3380.2440.931Primary21.5%0.2910.3480.2770.2901.099
Secondary13.7%0.2490.3070.1930.1550.550Secondary12.8%0.2380.2320.1810.2340.735
Relationship23.6%0.4850.2030.2260.0280.296Relationship25.6%0.4490.2750.2850.0380.706
+ +Table 3. Metaphorical image captioning (left) and visual question answering (right) performance of PaLI [10]. We report different metrics (higher the better) using both the exact match accuracy (Acc) and standard text generation metrics. For image captioning, we evaluate the whole predicted caption as well as parsed primary object, secondary object, and relationship. For VQA, we evaluate on all predicted answers as well as provide the breakdown for each question type. + +tioning task, where the input is an image and the target is the metaphorical message in the syntactic structure is as as . We fine-tune and evaluate the state-of-the-art literal image caption model PaLI-17B [10] based on the exact match accuracy (maximum over all references) and standard metrics for image captioning BLEU4 [36], ROUGE-L [28], METEOR [3], SPICE [2], and CIDEr [51]. Since the target captions follow the fixed syntactic structure, we parse each predicted caption into the primary concept, the secondary concept, and their relationship, and use the same set of metrics for the whole caption to evaluate. Table 3 (left) summarizes the results and Figure 4 provides sample qualitative results. We observe that PaLI generally struggles on this task. For instance, it achieves a CIDEr score of 1.076, compared to 1.491 for state-of-the-art literal image captioning on the popular COCO-Captions [8]. Further, the model struggles with predicting the target relationship when the metrics favor recall (e.g., ROUGE-L) and with predicting the target secondary concept when the metrics favor precision (e.g., BLEU). + +Visual Question-Answering (VQA). We propose metaphorical open-ended (i.e., not vocab-based) VQA task, where the input is an image and a given question, and + +the target is the answer. We use fixed templates to generate 2 VQA questions whose answer is the primary concept, 2 for secondary concept, and 2 for relationship. Again, we fine-tune and evaluate the state-of-the-art literal VQA model PaLI-17B [10], using the same set of metrics as in image captioning. Table 3 (right) summarizes the results and Figure 4 provides sample qualitative results. Overall, we find that PaLI performs poorly, only achieving the average accuracy score of $19.9\%$ , while the state-of-the-art literal VQA on the popular VQAv2 [18] benchmark is $77.6\%$ on "other" questions. Additionally, the model struggles the most with answering questions that ask for the secondary concept. + +# 3.3. Metaphor Localization + +Here, the task is to localize the image regions that invoke either the primary or secondary concept in the viewer. This is similar to the phrase grounding task of localizing objects using free-form natural language phrases [23, 31, 55], but with some key differences due to the peculiarities of visual metaphors in comparison to literal images used in standard vision datasets. + +Annotations. As discussed earlier, there are diverse types of metaphors based on how the primary and secondary concepts are visually depicted in an image (see Fig. 2). There are at least two key differences in metaphor localization + +![](images/0b4b8ddebaf432558714448de80b905fa83aefe24b6fb1a6d2244719f4e3b589.jpg) +The shoe is as light as feathers. + +![](images/ecdf30748a846733fd43bca6f7f9046d1f337aa12a79b070a67ffe786f7a7ec8.jpg) +Ketchup is as hot as a volcano. + +![](images/bf759da64092325daa75d573d8f7b0baa80180683bae3f615f2f08dd477c86f5.jpg) +Getting two ice creams is as exciting as getting two movie tickets for one. + +![](images/9d385c119769f7f879f00b0beb73fb05361ac6890a80b1d96969615f633fd4aa.jpg) +This car is as out of this world as three dimensional. + +![](images/2af22fca7d1313656213e48edb1222d4085026bb710ff8103b6bf6cf3ce082f5.jpg) +Figure 5. Sample Localization annotations showing annotated bounding boxes around primary and secondary concepts. Notice the diversity of types in bounding boxes: explicit, contextual, logo and texts. This makes metaphor localization more challenging compared to standard object detection. +Parking assist is as safe as parking far from objects. +Figure 6. Sample Localizations with the phrase-grounding model from [27], where the secondary concepts are contextual. GT boxes are shown in green, whereas the predictions are shown in blue. + +![](images/dc838d87c41af418b4aba5ec7b745dbc9daf79a7b8e8e8a784f338f7b9b27c44.jpg) +This chocolate is as fun as going to a concert. + +compared to standard localization in literal images: 1. A given concept can be present in the image either explicitly or in a contextual manner (for e.g., contextual visual metaphors in Fig. 2). 2. Visual metaphor Ads are inherently multimodal and a concept can be invoked by other modalities such as text or logo. See the multimodal visual metaphors in Fig. 2. As a result, we not only annotate the bounding boxes that invoke the primary/secondary concept in the viewer, but we also annotate the type of that bounding box. A bounding box can be of one of the 4 types: Explicitly present, Contextually present, Logo or a Text. Specifically, for each of the 5061 metaphorical images, we pick the best metaphor annotation (primary, secondary concepts and their relationship) according to their validation scores (see previous section) and collect bounding box annotations for both the primary and secondary concepts. We collect all the bounding boxes that invoke both the primary and secondary concepts and also their type (explicit, contextual, logo or text) for each of the images. We use 5 annotators for each annotation and choose the bounding boxes with the best inter-annotator agreement. We use the same train and test splits as used in understanding tasks. Fig. 5 shows sample localization annotations. We collected over $30\mathrm{k}$ bounding box annotations for this task on 5061 metaphor images. + +Detection. Recent detection and localization models [27, 58] pre-trained on image and caption pairs are shown to + +
mAP50mAP70
Primary concept33.2214.25
Secondary concept43.5431.23
+ +Table 4. Localization results with CLIP based phrase localization model [27] on our test split. + +achieve remarkable localization performance on discriminating fine-grained objects and unseen concepts. Specifically, we evaluate [27] which leverages the effective image representations in CLIP by extracting spatial features from it. Using these spatial features, for each pixel location, the model computes the inner product between the spatial feature and the phrase embedding extracted from CLIP to predict the bounding box. In our case, we pass the primary or secondary concept as input phrase to [27] the estimate the corresponding bounding box. Table 4 summarizes the detection results (using Mean Average Precision) on our test split. We find relatively better performance in localizing secondary objects compared to primary objects. We show few qualitative results in Figure 6. It is worth noting that our collected annotations allow for more comprehensive analysis on localization tasks due to the availability of different types of bounding boxes (explicit, contextual, logos and texts). + +# 3.4. Metaphor Generation + +Recent large-scale text-to-image (T2I) generative models show remarkable success in generating highly realistic images from text prompts. Can these models also work well in metaphorical image generation? We evaluate two state-of-the-art generative models (Imagen [42], Stable diffusion [41] (SD)) using 300 samples from MetaCLUE test set where we use the text prompts: "An advertisement where primary-concept is as relationship as secondary concept." In addition, we finetune the stable diffusion model on our train split (same split as in Sec. 3.2). + +Results. Fig. 7 shows sample visual results from different T2I models, with the metaphor annotation shown on the + +
ModelFID ↓CLIP Similarity ↑
Imagen [42]153.132.1
Stable Diffusion [41]161.630.8
Stable Diffusion - FT154.332.0
+ +Table 5. Analysis of Image Generation results with standard metrics of FID [42] and CLIP similarity [39] scores. + +top. The generated images capture different aspects of the metaphor (tablet, waterproof), but not the entire metaphorical message. We compute standard metrics to automatically evaluate the quality of the generations. Tab. 5 shows the standard FID [19] and CLIP-Similarity [39] scores of different models. FID score evaluates the image distribution similarity between the generation images with the corresponding real image distribution. FID scores in Tab. 5 shows that Imagen performs slightly better than SD. And, there is a slight improvement in FID with finetuning (SDFT vs. SD). In general, high FID scores in Tab. 5 indicate the large distribution gap between the generated and real images. CLIP similarity score, on the other hand, measures the prompt fidelity - similarity between the generated image and the corresponding input text prompt according to the CLIP model [39]. CLIP similarity scores in Tab. 5 follow the similar trend as FID scores: Imagen performs better than SD in terms of the prompt fidelity, and finetuning SD model improved its prompt fidelity. + +Given that both FID and CLIP scores are not tailored towards metaphorical images, these metrics are not reliable in assessing whether the generated images capture the essence of visual metaphors. To analyze this, we perform human studies comparing two different models at a time. Specifically, we show a metaphorical message and the corresponding generations from two models and ask the users to pick an image that best conveys that metaphorical message. We obtain 7 user ratings for each image pair and consider 3-out-of-7 or 4-out-of-7 to be ties. Fig. 8 shows the percentage of user preferences across different pairs of results. User studies also indicate that Imagen performed better than SD. An interesting finding is that finetuning SD resulted in slightly worse user preference compared to base SD model (SDFT vs. SD). This is in contrast to standard FID and CLIP metrics that improved with finetuning. We hypothesize that finetuning SD resulted in more realistic Ad images, but the resulting model may have forgotten some of the metaphorical priors. This calls for more effective finetuning strategies with the limited training datasets, which forms an important future work. In addition, both Imagen and SD performed quite poorly compared to real images in conveying metaphorical messages. Real images are preferred around $88\%$ of time over Imagen results. This illustrates the big scope of improvements in generating visual metaphors. + +Metaphor: This android tablet is as waterproof as someone in a swimtube. + +![](images/057c8a7d03abade2c4c6a9902380c8451e89473675b0d84921203e9a77d6f0a1.jpg) + +![](images/610235dcb4fdee2ad461930d71f923cb8d5ce027804e68ca032a25f03d8fe1b0.jpg) + +![](images/c9c9a9e32bff7517f5c17a00f640199c65dc936bebc7651716b99588707cf02c.jpg) +Figure 7. Sample Image Generations for a given metaphorical message (shown on top) with Imagen [42], Stable Diffusion [41] and fine-tuned (FT) version of Stable Diffusion. + +![](images/6e9a00b90199accce31f522c019995823bcd16b8e9949fb7bcf904d536de09cf.jpg) + +![](images/317d1e758b9dbbcecdd3e33759232a5e45f0b0ad5239d742ba643c80836fdac5.jpg) +Figure 8. User Study on Image Generation Results. Percentage of results users preferred across real, Imagen [42], Stable Diffusion (SD) [41] and its fine-tuned version (SD-FT) results. Users are asked to choose the image that better depicts a given metaphor. + +# 4. Conclusion + +In this paper, we present a step towards comprehensive evaluation of progress on visual metaphor research. Specifically, we propose a collection of tasks related to comprehending and generating visual metaphors using AI techniques. Our MetaCLUE tasks include Classification, Understanding (Retrieval, Captioning, VQA), Localization and Generation. For comprehensive evaluations, we collected high quality and rich annotations that facilitate the measurable progress. Existing methods demonstrate poor results in many cases with our experimental analysis shedding light on strengths and drawbacks of different approaches paving a path for future research in this fascinating field. + +# References + +[1] Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569-11579, 2021. 2, 3 +[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. SPICE: semantic propositional image caption evaluation. In ECCV, 2016. 6 +[3] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In ACL Workshops, 2005. 6 +[4] Roland Barthes. Image-music-text. Macmillan, 1977. 3 +[5] Nathalie Bonnardel and Evelyne Marmèche. Towards supporting evocation processes in creative design: A cognitive approach. International journal of human-computer studies, 63(4-5):422-435, 2005. 3 +[6] Robert W Boozer, David C Wyld, and James Grant. Using metaphor to create more effective sales messages. Journal of services marketing, 1990. 3 +[7] Tuhin Chakrabarty, Xurui Zhang, Smaranda Muresan, and Nanyun Peng. Mermaid: Metaphor generation with symbolism and discriminative decoding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4250-4261, 2021. 2, 3 +[8] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO Captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6 +[9] Xianyang Chen, Chee Wee Leong, Michael Flor, and Beata Beigman Klebanov. Go figure! multi-task transformer-based architecture for metaphor detection using idioms: Ets team in 2020 metaphor shared task. In Proceedings of the second workshop on figurative language processing, pages 235-243, 2020. 3 +[10] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 5, 6 +[11] Lydia B Chilton, Savvas Petridis, and Maneesh Agrawala. Visiblends: A flexible workflow for visual blends. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2019. 4 +[12] Minjin Choi, Sunkyung Lee, Eunseong Choi, Heesoo Park, Junhyuk Lee, Dongwon Lee, and Jongwuk Lee. Melbert: Metaphor detection via contextualized late interaction using metaphorical identification theories. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1763-1773, 2021. 2, 3 + +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 5 +[14] Gilles Fauconnier and Mark Turner. The way we think: Conceptual blending and the mind's hidden complexities. Basic books, 2008. 3 +[15] Charles Forceville. Pictorial metaphor in advertising. Routledge, 1996. 2, 3 +[16] Susan R Fussell and Mallie M Moss. Figurative language in emotional communication. Social and cognitive approaches to interpersonal communication, pages 113-141, 1998. 3 +[17] Hongyu Gong, Kshitij Gupta, Akriti Jain, and Suma Bhat. Illinimet: Illinois system for metaphor detection with contextual and linguistic information. In Proceedings of the Second Workshop on Figurative Language Processing, pages 146-153, 2020. 3 +[18] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6 +[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8 +[20] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic understanding of image and video advertisements. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1705-1715, 2017. 2, 3, 4 +[21] Bipin Indurkhya. On the role of metaphor in creative cognition. In ICCC, pages 51-59, 2010. 3 +[22] Bipin Indurkhya and Amitash Ojha. An empirical study on the role of perceptual similarity in visual metaphors and creativity. Metaphor and Symbol, 28(4):233-253, 2013. 4 +[23] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 6 +[24] George Lakoff. The contemporary theory of metaphor. 1993. 2 +[25] George Lakoff and Mark Johnson. Metaphors we live by. University of Chicago press, 2008. 2 +[26] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 5 +[27] Jiahao Li, Greg Shakhnarovich, and Raymond A Yeh. Adapting clip for phrase localization without further training. arXiv preprint arXiv:2204.03647, 2022. 7 +[28] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 6 + +[29] Todd I Lubart and Isaac Getz. Emotion, metaphor, and the creative process. Creativity research journal, 10(4):285-301, 1997. 2 +[30] Carita Lundmark. Metaphor and creativity in British magazine advertising. PhD thesis, Luleå teknska universitet, 2005. 3 +[31] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6 +[32] James H Martin. A corpus-based analysis of context effects on metaphor comprehension. Trends in Linguistics Studies and Monographs, 171:214, 2006. 3 +[33] Zachary J Mason. Cormet: a computational, corpus-based conventional metaphor extraction system. Computational linguistics, 30(1):23-44, 2004. 3 +[34] Edward F McQuarrie and David Glen Mick. Visual rhetoric in advertising: Text-interpretive, experimental, and reader-response analyses. Journal of consumer research, 26(1):37-54, 1999. 3 +[35] Ekaterina Ovchinnikova, Vladimir Zaytsev, Suzanne Wertheim, and Ross Israel. Generating conceptual metaphors from proposition stores. arXiv preprint arXiv:1409.7619, 2014. 3 +[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. 2002. 6 +[37] Barbara J Phillips. The impact of verbal anchoring on consumer response to image ads. Journal of advertising, 29(1):15-24, 2000. 3 +[38] Barbara J Phillips and Edward F McQuarrie. Beyond visual metaphor: A new typology of visual rhetoric in advertising. Marketing theory, 4(1-2):113-136, 2004. 3 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5, 8 +[40] Sunny Rai and Shampa Chakraverty. A survey on computational metaphor processing. ACM Computing Surveys (CSUR), 53(2):1-37, 2020. 3 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 7, 8 +[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 7, 8 +[43] Linda M Scott. Images in advertising: The need for a theory of visual rhetoric. Journal of consumer research, 21(2):252-273, 1994. 2 + +[44] Gerard Steen. A method for linguistic metaphor identification: From MIP to MIPVU, volume 14. John Benjamins Publishing, 2010. 3 +[45] Kevin Stowe, Tuhin Chakrabarty, Nanyun Peng, Smaranda Muresan, and Iryna Gurevych. Metaphor generation with conceptual mappings. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6724-6736, 2021. 2, 3 +[46] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 4, 5 +[47] Asuka Terai and Masanori Nakagawa. A computational system of metaphor generation with evaluation mechanism. In International Conference on Artificial Neural Networks, pages 142-147. Springer, 2010. 3 +[48] Xiaoyu Tong, Ekaterina Shutova, and Martha Lewis. Recent advances in neural metaphor processing: A linguistic, cognitive and social perspective. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4673-4686, 2021. 2, 3 +[49] Peter Turney, Yair Neuman, Dan Assaf, and Yohai Cohen. *Literal and metaphorical sense identification through concrete and abstract context*. In *Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing*, pages 680–690, 2011. 3 +[50] Tony Veale, Ekaterina Shutova, and Beata Beigman Klebanov. Metaphor: A computational perspective. Synthesis Lectures on Human Language Technologies, 9(1):1-160, 2016. 2 +[51] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015. 6 +[52] Merryl J Wilkenfeld and Thomas B Ward. Similarity and emergence in conceptual combination. Journal of Memory and Language, 45(1):21-38, 2001. 3 +[53] Bo Xu, Tingting Li, Junzhe Zheng, Mehdi Naseriparsa, Zhehuan Zhao, Hongfei Lin, and Feng Xia. Met-meme: A multimodal meme dataset rich in metaphors. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2887-2899, 2022. 4 +[54] Keren Ye and Adriana Kovashka. Advise: Symbolism and external knowledge for decoding advertisements. In Proceedings of the European Conference on Computer Vision (ECCV), pages 837-855, 2018. 4 +[55] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 6 +[56] Zhiwei Yu and Xiaojun Wan. How to avoid sentences spelling boring? towards a neural approach to unsupervised metaphor generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for + +Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 861-871, 2019. 3 +[57] Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, and Hongfei Lin. Multimet: A multimodal dataset for metaphor understanding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3214-3225, 2021. 4 +[58] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 7 \ No newline at end of file diff --git a/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/images.zip b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9294b6206941cb1417e5ca2a6c3aade312d2e344 --- /dev/null +++ b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8661d8b42c409923c72a9110c60b7220f4ff401c050dc98f0b1e359b940d50c +size 531989 diff --git a/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/layout.json b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..34b95a6e8b1a7b7b0c7cf1f57c05c9280cc19597 --- /dev/null +++ b/2023/MetaCLUE_ Towards Comprehensive Visual Metaphors Research/layout.json @@ -0,0 +1,8582 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "type": "text", + "content": "MetaCLUE: Towards Comprehensive Visual Metaphors Research" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 143, + 491, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 143, + 491, + 200 + ], + "spans": [ + { + "bbox": [ + 102, + 143, + 491, + 200 + ], + "type": "text", + "content": "Arjun R. Akula*, Brendan Driscoll*, Pradyumna Narayana, Soravit Changpinyo, Zhiwei Jia, Suyash Damle, Garima Pruthi, Sugato Basu, Leonidas Guibas, William T. Freeman, Yuanzhen Li, Varun Jampani* Google" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 230, + 214, + 345 + ], + "blocks": [ + { + "bbox": [ + 52, + 230, + 214, + 345 + ], + "lines": [ + { + "bbox": [ + 52, + 230, + 214, + 345 + ], + "spans": [ + { + "bbox": [ + 52, + 230, + 214, + 345 + ], + "type": "image", + "image_path": "209a120a3d720efd6732ac818466b1da66b20f7edd8f48d41c73de8e0c48d583.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 347, + 211, + 372 + ], + "lines": [ + { + "bbox": [ + 52, + 347, + 211, + 372 + ], + "spans": [ + { + "bbox": [ + 52, + 347, + 211, + 372 + ], + "type": "text", + "content": "Metaphor: Killing trees is as harmful as killing wildlife." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 230, + 336, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 230, + 336, + 244 + ], + "spans": [ + { + "bbox": [ + 256, + 230, + 336, + 244 + ], + "type": "text", + "content": "Classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 248, + 367, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 248, + 367, + 269 + ], + "spans": [ + { + "bbox": [ + 219, + 248, + 367, + 269 + ], + "type": "text", + "content": "Is this a visual metaphor? - YES - NO" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 260, + 281, + 332, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 281, + 332, + 295 + ], + "spans": [ + { + "bbox": [ + 260, + 281, + 332, + 295 + ], + "type": "text", + "content": "Localization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 218, + 297, + 363, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 297, + 363, + 320 + ], + "spans": [ + { + "bbox": [ + 218, + 297, + 363, + 320 + ], + "type": "text", + "content": "Detect image regions that invoke the concepts:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 218, + 325, + 281, + 348 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 218, + 325, + 275, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 325, + 275, + 335 + ], + "spans": [ + { + "bbox": [ + 218, + 325, + 275, + 335 + ], + "type": "text", + "content": "- Killing trees" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 218, + 338, + 281, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 338, + 281, + 348 + ], + "spans": [ + { + "bbox": [ + 218, + 338, + 281, + 348 + ], + "type": "text", + "content": "- killing wildlife" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 284, + 308, + 367, + 369 + ], + "blocks": [ + { + "bbox": [ + 284, + 308, + 367, + 369 + ], + "lines": [ + { + "bbox": [ + 284, + 308, + 367, + 369 + ], + "spans": [ + { + "bbox": [ + 284, + 308, + 367, + 369 + ], + "type": "image", + "image_path": "a639cb28fbb5f4dd8079fbda6c71370593e24e033a72a38e4a7efb389868de45.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 480, + 547, + 536 + ], + "lines": [ + { + "bbox": [ + 46, + 480, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 547, + 536 + ], + "type": "text", + "content": "Figure 1. With MetaCLUE, we introduce several interesting tasks related to visual metaphors. We collect metaphor annotations (objects, abstract concepts, relationships and object boxes) for evaluating existing models on these tasks. Specifically we perform a comprehensive evaluation of vision and language models on four different tasks (Classification, Localization, Understanding, and gEneration). Comprehensive experiments in this work show that state-of-the-art techniques mostly focus on literal interpretation and perform poorly in understanding and generation of metaphor images." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 427, + 232, + 492, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 232, + 492, + 246 + ], + "spans": [ + { + "bbox": [ + 427, + 232, + 492, + 246 + ], + "type": "text", + "content": "gEneration" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 381, + 249, + 522, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 249, + 522, + 285 + ], + "spans": [ + { + "bbox": [ + 381, + 249, + 522, + 285 + ], + "type": "text", + "content": "Prompt: \"An advertisement where killing trees is as harmful as killing wildlife.\"" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 386, + 289, + 455, + 358 + ], + "blocks": [ + { + "bbox": [ + 386, + 289, + 455, + 358 + ], + "lines": [ + { + "bbox": [ + 386, + 289, + 455, + 358 + ], + "spans": [ + { + "bbox": [ + 386, + 289, + 455, + 358 + ], + "type": "image", + "image_path": "6401876215261697f19124b169d600cb701db81e91126ced2a0c158e7de52e1f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 361, + 452, + 369 + ], + "lines": [ + { + "bbox": [ + 389, + 361, + 452, + 369 + ], + "spans": [ + { + "bbox": [ + 389, + 361, + 452, + 369 + ], + "type": "text", + "content": "Stable Diffusion" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 464, + 289, + 533, + 357 + ], + "blocks": [ + { + "bbox": [ + 464, + 289, + 533, + 357 + ], + "lines": [ + { + "bbox": [ + 464, + 289, + 533, + 357 + ], + "spans": [ + { + "bbox": [ + 464, + 289, + 533, + 357 + ], + "type": "image", + "image_path": "9ccc5ac901c561ad3aca7d3018ac7e221c233c7aa31b53e61fded73e4225eae9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 483, + 360, + 513, + 370 + ], + "lines": [ + { + "bbox": [ + 483, + 360, + 513, + 370 + ], + "spans": [ + { + "bbox": [ + 483, + 360, + 513, + 370 + ], + "type": "text", + "content": "Imagen" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 252, + 380, + 340, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 380, + 340, + 395 + ], + "spans": [ + { + "bbox": [ + 252, + 380, + 340, + 395 + ], + "type": "text", + "content": "Understanding" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 56, + 396, + 95, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 95, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 95, + 407 + ], + "type": "text", + "content": "Retrieval" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 56, + 409, + 130, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 130, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 130, + 419 + ], + "type": "text", + "content": "Pick the right one:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 56, + 422, + 198, + 469 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 56, + 422, + 198, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 198, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 198, + 445 + ], + "type": "text", + "content": "(a) Killing the forest is as deadly as killing the animals too." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 56, + 446, + 190, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 446, + 190, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 446, + 190, + 469 + ], + "type": "text", + "content": "(b) Birds is as much a part of our world as used cans." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 205, + 396, + 253, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 396, + 253, + 407 + ], + "spans": [ + { + "bbox": [ + 205, + 396, + 253, + 407 + ], + "type": "text", + "content": "Captioning" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 205, + 409, + 282, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 409, + 282, + 420 + ], + "spans": [ + { + "bbox": [ + 205, + 409, + 282, + 420 + ], + "type": "text", + "content": "Sample predictions:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 205, + 422, + 363, + 469 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 205, + 422, + 342, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 422, + 342, + 445 + ], + "spans": [ + { + "bbox": [ + 205, + 422, + 342, + 445 + ], + "type": "text", + "content": "1. Deforestation is as damaging as killing wildlife." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 205, + 446, + 363, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 446, + 363, + 469 + ], + "spans": [ + { + "bbox": [ + 205, + 446, + 363, + 469 + ], + "type": "text", + "content": "2. Deforestation is as bad as ending the death penalty." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 380, + 396, + 494, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 396, + 494, + 420 + ], + "spans": [ + { + "bbox": [ + 380, + 396, + 494, + 420 + ], + "type": "text", + "content": "Visual Question Answering Sample Questions:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 379, + 422, + 538, + 463 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 379, + 422, + 507, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 422, + 507, + 445 + ], + "spans": [ + { + "bbox": [ + 379, + 422, + 507, + 445 + ], + "type": "text", + "content": "Q. What is as harmful as killing wildlife?" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 379, + 451, + 538, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 451, + 538, + 463 + ], + "spans": [ + { + "bbox": [ + 379, + 451, + 538, + 463 + ], + "type": "text", + "content": "Q. What is compared to killing wildlife?" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 143, + 543, + 192, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 543, + 192, + 555 + ], + "spans": [ + { + "bbox": [ + 143, + 543, + 192, + 555 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "type": "text", + "content": "Creativity is an indispensable part of human cognition and also an inherent part of how we make sense of the world. Metaphorical abstraction is fundamental in communicating creative ideas through nuanced relationships between abstract concepts such as feelings. While computer vision benchmarks and approaches predominantly focus on understanding and generating literal interpretations of images, metaphorical comprehension of images remains relatively unexplored. Towards this goal, we introduce MetaCLUE, a set of vision tasks on visual metaphor. We also collect high-quality and rich metaphor annotations (ab" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 544, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 689 + ], + "type": "text", + "content": "struct objects, concepts, relationships along with their corresponding object boxes) as there do not exist any datasets that facilitate the evaluation of these tasks. We perform a comprehensive analysis of state-of-the-art models in vision and language based on our annotations, highlighting strengths and weaknesses of current approaches in visual metaphor classification, localization, understanding (retrieval, question answering, captioning) and generation (text-to-image synthesis) tasks. We hope this work provides a concrete step towards developing AI systems with human-like creative capabilities. Project page: https://metaclue.github.io" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23201" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 128, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 128, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 128, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 91, + 268, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 91, + 268, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 91, + 268, + 140 + ], + "type": "text", + "content": "\"Metaphor is pervasive in everyday life ... Our ordinary conceptual system, in terms of which we both think and act, is fundamentally metaphorical in nature.\" — Lakoff & Johnson [25]" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 151, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 151, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 151, + 287, + 270 + ], + "type": "text", + "content": "Creativity is a process of generating a new perspective on a problem or a situation. Metaphorical thinking has been recognized as a key and powerful mechanism of creativity [24,29,50]. Humans engage metaphors in their creative thinking process as strategies to link or blend concepts, or to view a concept from a target domain in terms of another, apparently dissimilar concept from a source domain [25]. metaphors also provide a sophisticated tool for nuanced human communication. Let us take a closer look at the structure of metaphors – and especially visual metaphors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 288, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 288, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 288, + 484 + ], + "type": "text", + "content": "Metaphors1 are a cognitive construct in which a concept is compared to a seemingly unrelated concept via some shared attribute. Take as an example 'This car is a cheetah', where 'This car' is compared to 'a cheetah' in terms of speed. Metaphors have a simple syntactic structure of 'A is B' where A is referred to as the primary concept and B as the secondary concept. The implied analogy in a metaphor is of the form: ' (primary concept) is as (relationship)2 as (secondary concept)' and often involves an attribute transfer from the secondary to the primary concept. Some examples include 'This phone is as fast as a rocket', 'Cigarettes are as harmful as bullets' etc. The primary and secondary concepts are usually unrelated at a glance, resulting in an element of surprise and creativity in metaphorical expressions. Despite following such simple structure, metaphors are quite powerful in conveying creative ideas. Metaphors are pervasive in all forms of communication, such as speech, text, visual etc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 666 + ], + "type": "text", + "content": "Visual Metaphors are images where the primary and secondary concepts are visually depicted in an image conveying the metaphorical message to the viewers. Visual metaphors are widely used in mass media communications like advertising and journalism [15, 43, 45]. In this work, we work with Ad images, as metaphors tend to be prevalent in ads. There are numerous ways a metaphor can be represented visually. Following the classification in [15], there are at least 4 different types of visual metaphors. Fig. 2 shows sample images that belong to these types along with our annotations of primary, secondary concepts and their relationship. In contextual metaphors, either the primary or secondary concept is not explicitly visible, but is inferred from the context (e.g., apple in the left-most image). In hybrid metaphors, the primary and secondary concepts are" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 190 + ], + "type": "text", + "content": "visually conflated. Juxtaposition forms one of the simplest visual metaphor types, where the two concepts are just presented next to each other. Multimodal metaphors represent one of the concepts with another modality, such as text or logo. In practice, visual metaphors use multiple of these strategies to convey a metaphor in an effective manner. In many cases, the implied metaphorical meaning is somewhat open-ended. Interpretation of visual metaphors depends on several external factors, such as familiarity with the brands and cultural context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 192, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 335 + ], + "type": "text", + "content": "These visual variations and nuances make automatic cognition or generation of visual metaphors highly challenging. While the last decade has seen rapid progress in many areas of understanding and generation tasks, prior works in computer vision focus heavily on literal interpretation of images and overlook the importance of metaphorical reasoning in understanding the image message [1, 48]. We believe that developing AI systems with metaphorical comprehension and generation capabilities can greatly assist humans in creative endeavors involving conveying concepts in new and exciting ways. Such systems provide an important step towards conferring human-like creativity to AI models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 336, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 468 + ], + "type": "text", + "content": "To this end, we introduce multiple interesting tasks and construct metaphor annotations that enable comprehensive research on visual metaphors. As metaphors are more common in visual Ads, we start with the Pitt's Ads dataset images [20] and then perform a rigorous multi-stage annotation process with expert annotators to filter metaphorical images, add metaphor annotations, and perform additional validation steps to clean the annotations3. While there is recent work making advances in understanding non-literal interpretations in natural language research [7, 12], this work proposes the first step towards metaphor analysis in images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 468, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 684 + ], + "type": "text", + "content": "As illustrated in Fig. 1, we perform comprehensive evaluations with state-of-the-art techniques on four sets of tasks, which we call MetaCLUE: 1. Classification: This is binary classification task of estimating whether a given image contains a metaphor or not. In other words, Are visual features indicative of whether there exists a metaphor in a given image or not?. 2. Localization: Here, the task is to localize the image regions that invoke the primary and secondary concepts in the viewer. This is similar to a standard object detection task, but is more complicated in the case of visual metaphors as the primary/secondary concepts may not be explicitly present in an image. 3. Understanding: Can our models understand the metaphorical message in a given image? We pose this understanding problem as 3 tasks where we can quantitatively measure the performance: Retrieval, Captioning and Visual question answering. 4. gEneration: Can we generate an image that conveys the metaphor, given the metaphorical message as a text prompt?" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "3The Ads, while useful for the purposes of our paper, some images may perpetuate harmful stereotypes according to characteristics such as gender." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 673, + 286, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 286, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 286, + 693 + ], + "type": "text", + "content": "1. Grammarians distinguish a metaphor \"A is B\" from a simile \"A is like B\". In our work we use \"metaphor\" to encompass both variants." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "text", + "content": "2We use the word 'relationship' to denote the shared property of primary and secondary concepts, usually adjectives or adjectival phrases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23202" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 83, + 173, + 163 + ], + "blocks": [ + { + "bbox": [ + 153, + 73, + 198, + 82 + ], + "lines": [ + { + "bbox": [ + 153, + 73, + 198, + 82 + ], + "spans": [ + { + "bbox": [ + 153, + 73, + 198, + 82 + ], + "type": "text", + "content": "Contextual" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 83, + 173, + 163 + ], + "lines": [ + { + "bbox": [ + 59, + 83, + 173, + 163 + ], + "spans": [ + { + "bbox": [ + 59, + 83, + 173, + 163 + ], + "type": "image", + "image_path": "fbd965311b0c74e4ca2dd41001132f6b7d4a01a2b7a6234d78a30650f44e892d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 167, + 167, + 175 + ], + "lines": [ + { + "bbox": [ + 62, + 167, + 167, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 167, + 167, + 175 + ], + "type": "text", + "content": "This beer is as tasty as a real apple." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 181, + 86, + 290, + 158 + ], + "blocks": [ + { + "bbox": [ + 181, + 86, + 290, + 158 + ], + "lines": [ + { + "bbox": [ + 181, + 86, + 290, + 158 + ], + "spans": [ + { + "bbox": [ + 181, + 86, + 290, + 158 + ], + "type": "image", + "image_path": "aa12fc950b3973b9820cec40425dcc1988ea0c7bdb1e4c4afaa801b13b84dfd5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 162, + 288, + 179 + ], + "lines": [ + { + "bbox": [ + 179, + 162, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 179, + 162, + 288, + 179 + ], + "type": "text", + "content": "This car is as adventurous as a space ship." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 301, + 84, + 414, + 160 + ], + "blocks": [ + { + "bbox": [ + 401, + 72, + 430, + 83 + ], + "lines": [ + { + "bbox": [ + 401, + 72, + 430, + 83 + ], + "spans": [ + { + "bbox": [ + 401, + 72, + 430, + 83 + ], + "type": "text", + "content": "Hybrid" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 84, + 414, + 160 + ], + "lines": [ + { + "bbox": [ + 301, + 84, + 414, + 160 + ], + "spans": [ + { + "bbox": [ + 301, + 84, + 414, + 160 + ], + "type": "image", + "image_path": "0861d88e2a3d360853e8c299b9877fc03857a240952a8effc8420dd5c0929bd4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 162, + 415, + 179 + ], + "lines": [ + { + "bbox": [ + 301, + 162, + 415, + 179 + ], + "spans": [ + { + "bbox": [ + 301, + 162, + 415, + 179 + ], + "type": "text", + "content": "Driving this SUV is as smooth as birds flying in the sky." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 418, + 84, + 533, + 163 + ], + "blocks": [ + { + "bbox": [ + 418, + 84, + 533, + 163 + ], + "lines": [ + { + "bbox": [ + 418, + 84, + 533, + 163 + ], + "spans": [ + { + "bbox": [ + 418, + 84, + 533, + 163 + ], + "type": "image", + "image_path": "32d41a06a69b86d55b19d50bccf4b6b508bbadc17eee7aeca30e1db324fb7627.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 167, + 523, + 175 + ], + "lines": [ + { + "bbox": [ + 419, + 167, + 523, + 175 + ], + "spans": [ + { + "bbox": [ + 419, + 167, + 523, + 175 + ], + "type": "text", + "content": "This pencil is as red as a fire truck." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 196, + 173, + 275 + ], + "blocks": [ + { + "bbox": [ + 60, + 196, + 173, + 275 + ], + "lines": [ + { + "bbox": [ + 60, + 196, + 173, + 275 + ], + "spans": [ + { + "bbox": [ + 60, + 196, + 173, + 275 + ], + "type": "image", + "image_path": "fba8eda3cf31a78df732cc054f9bdd677dfe2575d1766b800aa5f980c5547ec6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 278, + 168, + 287 + ], + "lines": [ + { + "bbox": [ + 61, + 278, + 168, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 278, + 168, + 287 + ], + "type": "text", + "content": "This chocolate bar is as rich as gold." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 275, + 291, + 292 + ], + "lines": [ + { + "bbox": [ + 179, + 275, + 291, + 292 + ], + "spans": [ + { + "bbox": [ + 179, + 275, + 291, + 292 + ], + "type": "text", + "content": "This car is as made for the beach as a crab." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 304, + 546, + 337 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 546, + 337 + ], + "type": "text", + "content": "Figure 2. Sample Visual Metaphors with their Annotations. There are different types of visual metaphors. The type depends on how the primary and secondary concepts are visually depicted. Here are sample Ad images from [20] where we annotated the primary concept, secondary concept and their relationship." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 178, + 195, + 292, + 269 + ], + "blocks": [ + { + "bbox": [ + 148, + 185, + 204, + 194 + ], + "lines": [ + { + "bbox": [ + 148, + 185, + 204, + 194 + ], + "spans": [ + { + "bbox": [ + 148, + 185, + 204, + 194 + ], + "type": "text", + "content": "Juxtaposition" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 195, + 292, + 269 + ], + "lines": [ + { + "bbox": [ + 178, + 195, + 292, + 269 + ], + "spans": [ + { + "bbox": [ + 178, + 195, + 292, + 269 + ], + "type": "image", + "image_path": "fc19b684838be1c1a648bc9ae41f9b69669726d22fad4e4302152277af5a54b5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 301, + 197, + 411, + 274 + ], + "blocks": [ + { + "bbox": [ + 301, + 197, + 411, + 274 + ], + "lines": [ + { + "bbox": [ + 301, + 197, + 411, + 274 + ], + "spans": [ + { + "bbox": [ + 301, + 197, + 411, + 274 + ], + "type": "image", + "image_path": "360d096004498d375b4996e1c2b82bf050d23ce8ce98d5ff58237b8e06e194b7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 275, + 413, + 292 + ], + "lines": [ + { + "bbox": [ + 301, + 275, + 413, + 292 + ], + "spans": [ + { + "bbox": [ + 301, + 275, + 413, + 292 + ], + "type": "text", + "content": "These donuts are as unique as as talking people." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 418, + 196, + 531, + 274 + ], + "blocks": [ + { + "bbox": [ + 391, + 186, + 440, + 195 + ], + "lines": [ + { + "bbox": [ + 391, + 186, + 440, + 195 + ], + "spans": [ + { + "bbox": [ + 391, + 186, + 440, + 195 + ], + "type": "text", + "content": "Multimodal" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 418, + 196, + 531, + 274 + ], + "lines": [ + { + "bbox": [ + 418, + 196, + 531, + 274 + ], + "spans": [ + { + "bbox": [ + 418, + 196, + 531, + 274 + ], + "type": "image", + "image_path": "57232ccf52b7920f5087f9ef19437801071cf4ba62848cd162eb574cd3373a43.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 275, + 522, + 292 + ], + "lines": [ + { + "bbox": [ + 419, + 275, + 522, + 292 + ], + "spans": [ + { + "bbox": [ + 419, + 275, + 522, + 292 + ], + "type": "text", + "content": "The car is as rugged as this muddy trailer." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 358, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 289, + 491 + ], + "type": "text", + "content": "We comprehensively evaluate existing state-of-the-art techniques for each of these tasks on our collected metaphor annotations. We evaluate the models both in a zero-shot manner as well as with finetuning on our annotations. Even though finetuning resulted in some improvements, most models struggle to produce satisfactory results in many cases, demonstrating the difficulty of these tasks. Our experiments highlight several strengths and weaknesses of the existing techniques on comprehending and generating visual metaphors, providing a concrete first step towards further AI research on this fascinating topic." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 502, + 134, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 134, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 134, + 514 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "Creativity and Metaphorical Abstraction. Creativity often involves an innovative fusion of objects, attributes, or relationships from previous knowledge to generate new concepts [5, 52]. Metaphors can serve as an invaluable tool for expressing creative insights and also to stimulate new ones [14, 21]. The cognitive research community has made initial attempts in understanding different realizations of metaphors such as language metaphors and visual metaphors [15, 16, 32]. Studies of visual persuasion show that visual metaphors may be more effective than language metaphors in terms of producing a greater degree of cognitive creativity [34]. In addition to improving creativity, metaphors are also known to elicit pleasure since the initial ambiguity in (re-)conceptualizing a target entity in terms of a source stimulus generates interest and motivation, and the subsequent resolution is rewarding – explaining the impor" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 358, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 545, + 383 + ], + "type": "text", + "content": "tance of using creative metaphorical processes in art, advertising, and marketing [6,30,38]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 386, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 546, + 506 + ], + "type": "text", + "content": "Metaphors in Language Research. Computational Linguistic studies show that metaphors are ubiquitous in language, occurring once per three sentences on average [40, 44, 49]. Recently, an increasing number of research efforts have explored the limitations and challenges in detecting and decoding the meaning of language metaphors [7, 48]. While there exist several computational models [9, 12, 17, 33, 35, 45, 47, 56] for metaphor identification, interpretation and generation in language, there exists very little work on computational modeling of visual metaphors." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "Metaphors in Computer Vision. Much of computer vision literature is focused on understanding and generating literal images. Automatic metaphorical interpretations of images is highly challenging and requires multi-faceted cognitive reasoning that involves visual reasoning coupled with the use of external knowledge. Recent work on affective captioning using the ArtEmis data set [1] includes some captions that evoke metaphors to explain emotions, but that dataset is focused on visual art and does not specifically consider metaphors. There is no explicit prior work for comprehensive evaluation and development of models that can automatically comprehend or generate visual metaphors. At the same time, there exist several studies that demonstrate the potential of visual metaphors. For instance, some studies [4,37] suggest that advertisements containing metaphorical images will be more persuasive compared to ads with language metaphors, non-metaphorical" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23203" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 70, + 257, + 196 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 257, + 196 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 257, + 196 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 257, + 196 + ], + "type": "image", + "image_path": "cd66b94d5c2c2b6dead4a5348fa6d5db897abec5c905b7e5548545e2f80e0212.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "type": "text", + "content": "Figure 3. Distribution of Topics in our annotated metaphorical images from Pitt's Ad dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 249, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 289, + 489 + ], + "type": "text", + "content": "ads, or literal images. There have been some prior computational models [11, 22] for metaphor generation, but they are not thoroughly validated against any benchmark datasets or user studies. Some works [20, 54] propose datasets and techniques for general Ad image understanding with a focus on the challenging aspects of non-literal interpretations in Ad images. However they do not explicitly collect any metaphorical annotations, nor do they provide the corresponding analysis. In this work, we start with an existing Ad image dataset [20] and perform extensive human studies to filter metaphorical images and collect detailed annotations accounting various aspects of metaphoric interpretation. Prior works such as Multi-MET [57] and MET-Meme [53], propose metaphorical annotations but does not annotate the relationship between primary and secondary. In MetaCLUE, in addition to providing relationship annotations, we also collect detailed bounding box annotations that help localize the image regions invoking the primary, secondary concepts in the viewer. We further provide VQA style question and answers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 499, + 123, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 123, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 123, + 511 + ], + "type": "text", + "content": "3. MetaCLUE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 519, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 604 + ], + "type": "text", + "content": "We introduce four different high-level tasks in MetaCLUE that enable comprehensive evaluation and development of visual metaphor research: Classification, Localization, Understanding and gEneration. In the rest of this section, we first describe our annotation collection process for the tasks and next provide the analysis of using existing state-of-the-art techniques for each of these tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 610, + 182, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 182, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 182, + 624 + ], + "type": "text", + "content": "3.1. Metaphor Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 701 + ], + "type": "text", + "content": "Following the fundamental vision task of image classification, we first ask: Can we develop models that can classify whether or not a given image contains a visual metaphor? In other words, is it possible to just use visual cues to estimate whether or not there exists a metaphorical interpretation of an image?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Annotations. For this task, we need to label whether or" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 491 + ], + "type": "text", + "content": "not each image contains a metaphor. Since we tend to see more metaphorical images in Ads, we start with images from an existing Ads dataset published by the University of Pittsburgh [20] and manually annotate whether a given Ad image contains a metaphor or not. Pitt's Ads dataset contains images of both product ads (e.g. phone ads) as well as public service announcements (e.g. forest conservation ads). Concretely, to make annotations more efficient, we use a subset of 8.5K Ad images from this dataset that are annotated to have 'symbolic' (fun, adventurous, etc.) references. We find this subset to contain a considerable portion of metaphors. Specifically, we collect 5 Yes/No annotations for each of these 8K images from different annotators. To obtain high-quality annotations, we explained what a visual metaphor is to the annotators and also conducted qualifying exams to pick final annotators for this task. We consider the images with 3 or more 'Yes' annotations as visual metaphors and the remaining as non-metaphorical. At the end, we identify 5061 images containing visual metaphors out of 8480 images. Pitt's Ads dataset [20] also comes with topic annotations (e.g., restaurant, car, animal rights etc.). Fig. 3 shows the word cloud plot of different topics in these 5061 metaphorical images. We split the metaphorical images into 3730 train and 1331 test images by maintaining the same distribution of topics in both the splits. We have two types of negative sets (non-metaphorical images) for classification experiments. One is formed by the remaining 3419 non-metaphorical images in the symbolic set, and another one is created separately by annotating an additional 3000 literal (and non-symbolic) images from Pitt's Ads dataset [20]. We add 2000 of 3419 symbolic negative images to our train split, and add the remaining 1419 images to our test split. Similarly 2000 of 3000 literal negatives are used in training, and the remaining 1000 are used for testing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 492, + 547, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 547, + 648 + ], + "type": "text", + "content": "Evaluation and Results. Using our collected metaphor images, we evaluate the performance of the following state-of-the-art models in classifying an input image as metaphor: EfficientNet [46] and Vision Transformer (ViT) [13]. We fine-tune these models to classify metaphor vs. symbolic/non-metaphors and metaphor vs. literal, and use " + }, + { + "bbox": [ + 304, + 492, + 547, + 648 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 492, + 547, + 648 + ], + "type": "text", + "content": " of corresponding train splits for validation. Test results are reported in Table 1. Although the performance of ViT is significantly better than EfficientNet and random baselines, there is still ample room for improvement. In particular, models find it easier to distinguish metaphor images from literal images, and struggle to identify metaphors within the symbolic image pool." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 658, + 448, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 448, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 448, + 671 + ], + "type": "text", + "content": "3.2. Metaphor Understanding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "We now describe how we collect annotations that help in capturing the metaphorical message from the images collected in previous section." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23204" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 277, + 153 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 277, + 153 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 277, + 153 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 277, + 153 + ], + "type": "table", + "html": "
Model#ParamsSymbolic Neg.Literal Neg.
ValTestValTest
RandomN/A63.1051.6060.6657.12
EfficientNet-B05.3M60.7649.6770.9450.30
EfficientNet-B766M61.4448.5469.8449.82
ViT-B/1686M69.3166.9884.0481.24
ViT-L/16307M65.8360.6581.4580.52
", + "image_path": "0dbaeae6207fe8bcc0a13f5e333b7b23c047d2e05b2a720db3cb1cd31d2a93d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 162, + 288, + 195 + ], + "lines": [ + { + "bbox": [ + 47, + 162, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 47, + 162, + 288, + 195 + ], + "type": "text", + "content": "Table 1. Accuracy of Metaphor Classification (binary classification accuracy) using state-of-the-art classification architectures of EfficientNet [46] and ViT [13]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 227, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 562 + ], + "type": "text", + "content": "Annotations. We provide detailed instructions and several examples to the annotators to help them annotate primary and secondary concepts in the metaphor and also the characteristic/relationship that is transferred from secondary to the primary. We conduct multiple pilot studies to reduce the noise and to improve inter-annotator agreement. Enforcing the annotators to make sure that their annotations are linguistically readable in the following syntactic structure helped us in improving quality and consistency of the annotations: “__ is as __ as __,” where the first blank is the primary concept, the second blank is the relationship, and the third blank is the secondary concept. Figure 2 shows some examples of these annotations. We collect 5 metaphor annotations for each image. As interpretations of metaphorical images can be highly subjective, there can exist more than one interpretation for each image, which makes it difficult to automatically remove noisy annotations. Therefore we conduct an additional human study where we show each of the annotation to five annotators and ask them to verify the correctness along three dimensions: (a) Is the grammar correct?; (b) Are primary and secondary concepts correct?; and (c) Is the relationship correct?. We remove annotations with a low number of votes out of 5 along each of the three dimensions, resulting in a total of over 26k clean annotations. We evaluate state-of-the-art models in understanding metaphorical message from the input images using 3 tasks namely, Retrieval, Captioning and Visual question answering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "text", + "content": "Retrieval. The goal of this task is to retrieve the correct metaphor interpretation/statement from a candidate set given an image. In our candidate set, we choose exactly one positive (correct) metaphorical statement from its ground truth messages and uniformly sample " + }, + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "text", + "content": " random negative statements from other images. Table 2 shows the results obtained with CLIP [39] and ALBEF [26] for " + }, + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "inline_equation", + "content": "K = 50" + }, + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "text", + "content": ", reporting retrieval precision@1 and rank4. Although CLIP ViT-L/14 shows good zero-shot performance on random negatives with more than " + }, + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "text", + "content": " accuracy, we observe a large drop" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 425, + 153 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 425, + 153 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 425, + 153 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 425, + 153 + ], + "type": "image", + "image_path": "8ca14d96a149751ef91b3f5bcb6263b959bd8dd54231fe77feafd9042acb5e98.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 156, + 421, + 175 + ], + "lines": [ + { + "bbox": [ + 310, + 156, + 421, + 175 + ], + "spans": [ + { + "bbox": [ + 310, + 156, + 421, + 175 + ], + "type": "text", + "content": "Metaphor: Smoking cigarettes is as life-shortening as sharpening a pencil" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 428, + 73, + 541, + 175 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 428, + 73, + 541, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 73, + 541, + 103 + ], + "spans": [ + { + "bbox": [ + 428, + 73, + 541, + 103 + ], + "type": "text", + "content": "Captioning Result Prediction:Smoking is as dangerous as burning a pencil." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 428, + 110, + 538, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 110, + 538, + 137 + ], + "spans": [ + { + "bbox": [ + 428, + 110, + 538, + 137 + ], + "type": "text", + "content": "Visual Question Answering Result Q1: What is a smoker's life compared to a sharpened pencil?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 429, + 138, + 535, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 138, + 535, + 166 + ], + "spans": [ + { + "bbox": [ + 429, + 138, + 535, + 166 + ], + "type": "text", + "content": "Pred: Sharp GT: Shortened Q2: What is used as a visual metaphor for a sharpened pencil?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 429, + 167, + 534, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 167, + 534, + 175 + ], + "spans": [ + { + "bbox": [ + 429, + 167, + 534, + 175 + ], + "type": "text", + "content": "Pred: Not smoking GT:Smoker's life" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 178, + 426, + 262 + ], + "blocks": [ + { + "bbox": [ + 310, + 178, + 426, + 262 + ], + "lines": [ + { + "bbox": [ + 310, + 178, + 426, + 262 + ], + "spans": [ + { + "bbox": [ + 310, + 178, + 426, + 262 + ], + "type": "image", + "image_path": "18d9bca6907411565d11d98da02daafa75f25b1aa588b96755fde662c2f44e7d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 263, + 417, + 281 + ], + "lines": [ + { + "bbox": [ + 310, + 263, + 417, + 281 + ], + "spans": [ + { + "bbox": [ + 310, + 263, + 417, + 281 + ], + "type": "text", + "content": "Metaphor: Smoking is as deadly as injecting drugs." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "text", + "content": "Figure 4. Results of PaLI [10] for captioning and visual question answering on sample images in our test split." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 428, + 181, + 538, + 282 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 428, + 181, + 538, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 181, + 538, + 210 + ], + "spans": [ + { + "bbox": [ + 428, + 181, + 538, + 210 + ], + "type": "text", + "content": "Captioning Result Prediction: Drinking and driving is as bad as injecting drugs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 428, + 217, + 537, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 217, + 537, + 243 + ], + "spans": [ + { + "bbox": [ + 428, + 217, + 537, + 243 + ], + "type": "text", + "content": "Visual Question Answering Result Q1: What is a cigarette as compared to injecting drugs?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 429, + 245, + 533, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 245, + 533, + 273 + ], + "spans": [ + { + "bbox": [ + 429, + 245, + 533, + 273 + ], + "type": "text", + "content": "Pred: Scary GT: Deadly Q2: What is used as a visual metaphor for injecting drugs?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 429, + 274, + 524, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 274, + 524, + 282 + ], + "spans": [ + { + "bbox": [ + 429, + 274, + 524, + 282 + ], + "type": "text", + "content": "Pred:Drinking GT:Cigarette" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "text", + "content": "in performance as we increase " + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\{100, 500, 1000\\}" + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "text", + "content": ". We further fine-tune CLIP models using " + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "text", + "content": " of metaphor annotations as the train set and see gains by up to " + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "inline_equation", + "content": "+7" + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "text", + "content": " absolute points in " + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "inline_equation", + "content": "p@1" + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "text", + "content": ". In summary, the performance of models is impressive with less than 50 negative candidates whereas the performance drops greatly by increasing " + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 336, + 545, + 409 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 410, + 546, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 546, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 546, + 659 + ], + "type": "text", + "content": "We hypothesize that models might simply be looking at salient objects rather than comprehending the underlying semantics of metaphor in finding the correct candidate. To test this, we mine hard negative (HN) statements and use them as our candidate set. Specifically, we construct the following five types of HNs: (a) Neg Prim: candidates obtained by replacing the primary concept in the metaphor statement with the primary concept from another image; (b) Neg Sec: replacing secondary concept likewise; (c) Neg Prim+Rel: replacing primary and relationship; (d) Neg Sec+Rel: replacing secondary and relationship; (e) Swap Prim&Neg: swapping primary and secondary from the same image. Table 2 shows the results. We see a significant drop of up to " + }, + { + "bbox": [ + 304, + 410, + 546, + 659 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 410, + 546, + 659 + ], + "type": "text", + "content": " by using HNs as negative statements for " + }, + { + "bbox": [ + 304, + 410, + 546, + 659 + ], + "type": "inline_equation", + "content": "K = 50" + }, + { + "bbox": [ + 304, + 410, + 546, + 659 + ], + "type": "text", + "content": ", indicating the difficulty in comprehending and distinguishing metaphorical abstraction of concepts. We find performance of Neg Prim is significantly higher than Neg Sec, suggesting that the models tend to rely more on primary object than secondary object in identifying correct interpretation. Overall, there is ample room for improvement indicated by the steep drop in model performance with HNs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 663, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 663, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 306, + 663, + 545, + 676 + ], + "type": "text", + "content": "Captioning. Here, we propose metaphor image cap" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "Rank measures the averaged ranking value of the highest-ranked ground-truth statement with 1 being the highest possible rank." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 683, + 545, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 545, + 703 + ], + "type": "text", + "content": "5We swap objects with images having different topic to make sure that the generated HNs are actually negatives." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 703, + 421, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 421, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 421, + 713 + ], + "type": "text", + "content": "With swapping " + }, + { + "bbox": [ + 317, + 703, + 421, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 317, + 703, + 421, + 713 + ], + "type": "text", + "content": " is always 2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23205" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 533, + 198 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 533, + 198 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 533, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 533, + 198 + ], + "type": "table", + "html": "
ModelRandom NegNeg PrimNeg SecNeg Prim+RelNeg Sec+RelSwap accuracy↑
p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓p@1↑rank ↓
CLIP (ViT-B/16)70.973.4946.6710.1138.1413.9749.368.8042.2511.6940.61
CLIP (ViT-B/32)61.784.1938.7411.6033.2014.8439.7910.2936.2712.7141.28
CLIP (ViT-L/14)76.663.1751.759.2239.8613.1954.747.7045.9911.0243.08
ALBEF39.798.5727.0016.5329.3115.7926.7715.5828.4214.5946.67
ALBEF (MSCOCO)44.877.5531.7815.1331.4114.0932.5314.0433.2812.8548.24
ALBEF (Flickr30k)47.498.7735.6014.5635.2213.6636.3513.5836.1212.7949.81
FT CLIP (ViT-B/16)76.812.2549.819.6545.4710.2453.408.0750.638.1844.65
FT CLIP (ViT-B/32)68.062.8243.0010.8239.6411.2344.729.3943.459.3545.69
FT CLIP (ViT-L/14)81.751.7857.667.4849.069.4061.255.9957.067.3343.75
", + "image_path": "2326bfcc24d2af8ca28cf2546fd5be9da9166cc881f724e9117598dbec0422bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 249, + 541, + 310 + ], + "blocks": [ + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "type": "text", + "content": "Table 2. Performance of retrieval models on " + }, + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "type": "text", + "content": " random (column 2) and hard negative candidates (columns 3-7) (" + }, + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "type": "inline_equation", + "content": "K = 50" + }, + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "type": "text", + "content": " for columns 2-6, " + }, + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 206, + 546, + 229 + ], + "type": "text", + "content": " for last column)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 249, + 541, + 310 + ], + "lines": [ + { + "bbox": [ + 53, + 249, + 541, + 310 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 541, + 310 + ], + "type": "table", + "html": "
CaptioningAccBLEU4ROUGE-LMETEORSPICECIDErVQAAccBLEU4ROUGE-LMETEORSPICECIDEr
Whole caption1.1%0.2540.5360.2200.1861.076All questions19.9%0.3290.2860.2490.1850.851
Primary29.9%0.3270.4070.3380.2440.931Primary21.5%0.2910.3480.2770.2901.099
Secondary13.7%0.2490.3070.1930.1550.550Secondary12.8%0.2380.2320.1810.2340.735
Relationship23.6%0.4850.2030.2260.0280.296Relationship25.6%0.4490.2750.2850.0380.706
", + "image_path": "62632c3d908e6212262e70a23088be00e58edcd04dc3df7aa03e97c4d0e925c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 318, + 546, + 364 + ], + "lines": [ + { + "bbox": [ + 46, + 318, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 546, + 364 + ], + "type": "text", + "content": "Table 3. Metaphorical image captioning (left) and visual question answering (right) performance of PaLI [10]. We report different metrics (higher the better) using both the exact match accuracy (Acc) and standard text generation metrics. For image captioning, we evaluate the whole predicted caption as well as parsed primary object, secondary object, and relationship. For VQA, we evaluate on all predicted answers as well as provide the breakdown for each question type." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 394, + 289, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 289, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 289, + 656 + ], + "type": "text", + "content": "tioning task, where the input is an image and the target is the metaphorical message in the syntactic structure is as as . We fine-tune and evaluate the state-of-the-art literal image caption model PaLI-17B [10] based on the exact match accuracy (maximum over all references) and standard metrics for image captioning BLEU4 [36], ROUGE-L [28], METEOR [3], SPICE [2], and CIDEr [51]. Since the target captions follow the fixed syntactic structure, we parse each predicted caption into the primary concept, the secondary concept, and their relationship, and use the same set of metrics for the whole caption to evaluate. Table 3 (left) summarizes the results and Figure 4 provides sample qualitative results. We observe that PaLI generally struggles on this task. For instance, it achieves a CIDEr score of 1.076, compared to 1.491 for state-of-the-art literal image captioning on the popular COCO-Captions [8]. Further, the model struggles with predicting the target relationship when the metrics favor recall (e.g., ROUGE-L) and with predicting the target secondary concept when the metrics favor precision (e.g., BLEU)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "type": "text", + "content": "Visual Question-Answering (VQA). We propose metaphorical open-ended (i.e., not vocab-based) VQA task, where the input is an image and a given question, and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 394, + 546, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 546, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 546, + 550 + ], + "type": "text", + "content": "the target is the answer. We use fixed templates to generate 2 VQA questions whose answer is the primary concept, 2 for secondary concept, and 2 for relationship. Again, we fine-tune and evaluate the state-of-the-art literal VQA model PaLI-17B [10], using the same set of metrics as in image captioning. Table 3 (right) summarizes the results and Figure 4 provides sample qualitative results. Overall, we find that PaLI performs poorly, only achieving the average accuracy score of " + }, + { + "bbox": [ + 306, + 394, + 546, + 550 + ], + "type": "inline_equation", + "content": "19.9\\%" + }, + { + "bbox": [ + 306, + 394, + 546, + 550 + ], + "type": "text", + "content": ", while the state-of-the-art literal VQA on the popular VQAv2 [18] benchmark is " + }, + { + "bbox": [ + 306, + 394, + 546, + 550 + ], + "type": "inline_equation", + "content": "77.6\\%" + }, + { + "bbox": [ + 306, + 394, + 546, + 550 + ], + "type": "text", + "content": " on \"other\" questions. Additionally, the model struggles the most with answering questions that ask for the secondary concept." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 560, + 437, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 437, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 437, + 574 + ], + "type": "text", + "content": "3.3. Metaphor Localization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 580, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 547, + 663 + ], + "type": "text", + "content": "Here, the task is to localize the image regions that invoke either the primary or secondary concept in the viewer. This is similar to the phrase grounding task of localizing objects using free-form natural language phrases [23, 31, 55], but with some key differences due to the peculiarities of visual metaphors in comparison to literal images used in standard vision datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Annotations. As discussed earlier, there are diverse types of metaphors based on how the primary and secondary concepts are visually depicted in an image (see Fig. 2). There are at least two key differences in metaphor localization" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 187, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 187, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 187, + 713 + ], + "type": "text", + "content": "The score of 0 is given if parsing fails." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23206" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 168, + 156 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 168, + 156 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 168, + 156 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 168, + 156 + ], + "type": "image", + "image_path": "0b4b8ddebaf432558714448de80b905fa83aefe24b6fb1a6d2244719f4e3b589.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 159, + 147, + 167 + ], + "lines": [ + { + "bbox": [ + 52, + 159, + 147, + 167 + ], + "spans": [ + { + "bbox": [ + 52, + 159, + 147, + 167 + ], + "type": "text", + "content": "The shoe is as light as feathers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 72, + 289, + 153 + ], + "blocks": [ + { + "bbox": [ + 176, + 72, + 289, + 153 + ], + "lines": [ + { + "bbox": [ + 176, + 72, + 289, + 153 + ], + "spans": [ + { + "bbox": [ + 176, + 72, + 289, + 153 + ], + "type": "image", + "image_path": "ecdf30748a846733fd43bca6f7f9046d1f337aa12a79b070a67ffe786f7a7ec8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 159, + 269, + 168 + ], + "lines": [ + { + "bbox": [ + 175, + 159, + 269, + 168 + ], + "spans": [ + { + "bbox": [ + 175, + 159, + 269, + 168 + ], + "type": "text", + "content": "Ketchup is as hot as a volcano." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 297, + 72, + 416, + 155 + ], + "blocks": [ + { + "bbox": [ + 297, + 72, + 416, + 155 + ], + "lines": [ + { + "bbox": [ + 297, + 72, + 416, + 155 + ], + "spans": [ + { + "bbox": [ + 297, + 72, + 416, + 155 + ], + "type": "image", + "image_path": "bf759da64092325daa75d573d8f7b0baa80180683bae3f615f2f08dd477c86f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 156, + 418, + 173 + ], + "lines": [ + { + "bbox": [ + 298, + 156, + 418, + 173 + ], + "spans": [ + { + "bbox": [ + 298, + 156, + 418, + 173 + ], + "type": "text", + "content": "Getting two ice creams is as exciting as getting two movie tickets for one." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 538, + 156 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 538, + 156 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 538, + 156 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 538, + 156 + ], + "type": "image", + "image_path": "9d385c119769f7f879f00b0beb73fb05361ac6890a80b1d96969615f633fd4aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 156, + 524, + 173 + ], + "lines": [ + { + "bbox": [ + 422, + 156, + 524, + 173 + ], + "spans": [ + { + "bbox": [ + 422, + 156, + 524, + 173 + ], + "type": "text", + "content": "This car is as out of this world as three dimensional." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 239, + 162, + 316 + ], + "blocks": [ + { + "bbox": [ + 46, + 183, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 547, + 217 + ], + "type": "text", + "content": "Figure 5. Sample Localization annotations showing annotated bounding boxes around primary and secondary concepts. Notice the diversity of types in bounding boxes: explicit, contextual, logo and texts. This makes metaphor localization more challenging compared to standard object detection." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 239, + 162, + 316 + ], + "lines": [ + { + "bbox": [ + 53, + 239, + 162, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 162, + 316 + ], + "type": "image", + "image_path": "2af22fca7d1313656213e48edb1222d4085026bb710ff8103b6bf6cf3ce082f5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 316, + 162, + 331 + ], + "lines": [ + { + "bbox": [ + 53, + 316, + 162, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 162, + 331 + ], + "type": "text", + "content": "Parking assist is as safe as parking far from objects." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 342, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 376 + ], + "type": "text", + "content": "Figure 6. Sample Localizations with the phrase-grounding model from [27], where the secondary concepts are contextual. GT boxes are shown in green, whereas the predictions are shown in blue." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 170, + 239, + 277, + 314 + ], + "blocks": [ + { + "bbox": [ + 170, + 239, + 277, + 314 + ], + "lines": [ + { + "bbox": [ + 170, + 239, + 277, + 314 + ], + "spans": [ + { + "bbox": [ + 170, + 239, + 277, + 314 + ], + "type": "image", + "image_path": "dc838d87c41af418b4aba5ec7b745dbc9daf79a7b8e8e8a784f338f7b9b27c44.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 316, + 274, + 331 + ], + "lines": [ + { + "bbox": [ + 168, + 316, + 274, + 331 + ], + "spans": [ + { + "bbox": [ + 168, + 316, + 274, + 331 + ], + "type": "text", + "content": "This chocolate is as fun as going to a concert." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 400, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 687 + ], + "type": "text", + "content": "compared to standard localization in literal images: 1. A given concept can be present in the image either explicitly or in a contextual manner (for e.g., contextual visual metaphors in Fig. 2). 2. Visual metaphor Ads are inherently multimodal and a concept can be invoked by other modalities such as text or logo. See the multimodal visual metaphors in Fig. 2. As a result, we not only annotate the bounding boxes that invoke the primary/secondary concept in the viewer, but we also annotate the type of that bounding box. A bounding box can be of one of the 4 types: Explicitly present, Contextually present, Logo or a Text. Specifically, for each of the 5061 metaphorical images, we pick the best metaphor annotation (primary, secondary concepts and their relationship) according to their validation scores (see previous section) and collect bounding box annotations for both the primary and secondary concepts. We collect all the bounding boxes that invoke both the primary and secondary concepts and also their type (explicit, contextual, logo or text) for each of the images. We use 5 annotators for each annotation and choose the bounding boxes with the best inter-annotator agreement. We use the same train and test splits as used in understanding tasks. Fig. 5 shows sample localization annotations. We collected over " + }, + { + "bbox": [ + 46, + 400, + 287, + 687 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}" + }, + { + "bbox": [ + 46, + 400, + 287, + 687 + ], + "type": "text", + "content": " bounding box annotations for this task on 5061 metaphor images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Detection. Recent detection and localization models [27, 58] pre-trained on image and caption pairs are shown to" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 351, + 236, + 502, + 281 + ], + "blocks": [ + { + "bbox": [ + 351, + 236, + 502, + 281 + ], + "lines": [ + { + "bbox": [ + 351, + 236, + 502, + 281 + ], + "spans": [ + { + "bbox": [ + 351, + 236, + 502, + 281 + ], + "type": "table", + "html": "
mAP50mAP70
Primary concept33.2214.25
Secondary concept43.5431.23
", + "image_path": "67c5583be3961a464bfc52e45f28027b26f8446ca5354ff2d10472007146ede5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 290, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 306, + 290, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 545, + 312 + ], + "type": "text", + "content": "Table 4. Localization results with CLIP based phrase localization model [27] on our test split." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 325, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 540 + ], + "type": "text", + "content": "achieve remarkable localization performance on discriminating fine-grained objects and unseen concepts. Specifically, we evaluate [27] which leverages the effective image representations in CLIP by extracting spatial features from it. Using these spatial features, for each pixel location, the model computes the inner product between the spatial feature and the phrase embedding extracted from CLIP to predict the bounding box. In our case, we pass the primary or secondary concept as input phrase to [27] the estimate the corresponding bounding box. Table 4 summarizes the detection results (using Mean Average Precision) on our test split. We find relatively better performance in localizing secondary objects compared to primary objects. We show few qualitative results in Figure 6. It is worth noting that our collected annotations allow for more comprehensive analysis on localization tasks due to the availability of different types of bounding boxes (explicit, contextual, logos and texts)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 550, + 432, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 432, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 432, + 562 + ], + "type": "text", + "content": "3.4. Metaphor Generation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 569, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 688 + ], + "type": "text", + "content": "Recent large-scale text-to-image (T2I) generative models show remarkable success in generating highly realistic images from text prompts. Can these models also work well in metaphorical image generation? We evaluate two state-of-the-art generative models (Imagen [42], Stable diffusion [41] (SD)) using 300 samples from MetaCLUE test set where we use the text prompts: \"An advertisement where primary-concept is as relationship as secondary concept.\" In addition, we finetune the stable diffusion model on our train split (same split as in Sec. 3.2)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Results. Fig. 7 shows sample visual results from different T2I models, with the metaphor annotation shown on the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23207" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 255, + 122 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 255, + 122 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 255, + 122 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 255, + 122 + ], + "type": "table", + "html": "
ModelFID ↓CLIP Similarity ↑
Imagen [42]153.132.1
Stable Diffusion [41]161.630.8
Stable Diffusion - FT154.332.0
", + "image_path": "78887cce51915947bcea4c90f5cd4f5a7ccb0f3acee0e19ea3b4fde41a05d0d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 287, + 152 + ], + "lines": [ + { + "bbox": [ + 47, + 129, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 287, + 152 + ], + "type": "text", + "content": "Table 5. Analysis of Image Generation results with standard metrics of FID [42] and CLIP similarity [39] scores." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 177, + 290, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 290, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 290, + 406 + ], + "type": "text", + "content": "top. The generated images capture different aspects of the metaphor (tablet, waterproof), but not the entire metaphorical message. We compute standard metrics to automatically evaluate the quality of the generations. Tab. 5 shows the standard FID [19] and CLIP-Similarity [39] scores of different models. FID score evaluates the image distribution similarity between the generation images with the corresponding real image distribution. FID scores in Tab. 5 shows that Imagen performs slightly better than SD. And, there is a slight improvement in FID with finetuning (SDFT vs. SD). In general, high FID scores in Tab. 5 indicate the large distribution gap between the generated and real images. CLIP similarity score, on the other hand, measures the prompt fidelity - similarity between the generated image and the corresponding input text prompt according to the CLIP model [39]. CLIP similarity scores in Tab. 5 follow the similar trend as FID scores: Imagen performs better than SD in terms of the prompt fidelity, and finetuning SD model improved its prompt fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "text", + "content": "Given that both FID and CLIP scores are not tailored towards metaphorical images, these metrics are not reliable in assessing whether the generated images capture the essence of visual metaphors. To analyze this, we perform human studies comparing two different models at a time. Specifically, we show a metaphorical message and the corresponding generations from two models and ask the users to pick an image that best conveys that metaphorical message. We obtain 7 user ratings for each image pair and consider 3-out-of-7 or 4-out-of-7 to be ties. Fig. 8 shows the percentage of user preferences across different pairs of results. User studies also indicate that Imagen performed better than SD. An interesting finding is that finetuning SD resulted in slightly worse user preference compared to base SD model (SDFT vs. SD). This is in contrast to standard FID and CLIP metrics that improved with finetuning. We hypothesize that finetuning SD resulted in more realistic Ad images, but the resulting model may have forgotten some of the metaphorical priors. This calls for more effective finetuning strategies with the limited training datasets, which forms an important future work. In addition, both Imagen and SD performed quite poorly compared to real images in conveying metaphorical messages. Real images are preferred around " + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "text", + "content": " of time over Imagen results. This illustrates the big scope of improvements in generating visual metaphors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 315, + 68, + 522, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 68, + 522, + 86 + ], + "spans": [ + { + "bbox": [ + 315, + 68, + 522, + 86 + ], + "type": "text", + "content": "Metaphor: This android tablet is as waterproof as someone in a swimtube." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 325, + 90, + 408, + 203 + ], + "blocks": [ + { + "bbox": [ + 325, + 90, + 408, + 203 + ], + "lines": [ + { + "bbox": [ + 325, + 90, + 408, + 203 + ], + "spans": [ + { + "bbox": [ + 325, + 90, + 408, + 203 + ], + "type": "image", + "image_path": "057c8a7d03abade2c4c6a9902380c8451e89473675b0d84921203e9a77d6f0a1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 426, + 86, + 536, + 204 + ], + "blocks": [ + { + "bbox": [ + 426, + 86, + 536, + 204 + ], + "lines": [ + { + "bbox": [ + 426, + 86, + 536, + 204 + ], + "spans": [ + { + "bbox": [ + 426, + 86, + 536, + 204 + ], + "type": "image", + "image_path": "610235dcb4fdee2ad461930d71f923cb8d5ce027804e68ca032a25f03d8fe1b0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 209, + 422, + 329 + ], + "blocks": [ + { + "bbox": [ + 310, + 209, + 422, + 329 + ], + "lines": [ + { + "bbox": [ + 310, + 209, + 422, + 329 + ], + "spans": [ + { + "bbox": [ + 310, + 209, + 422, + 329 + ], + "type": "image", + "image_path": "c9c9a9e32bff7517f5c17a00f640199c65dc936bebc7651716b99588707cf02c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 341, + 547, + 374 + ], + "lines": [ + { + "bbox": [ + 305, + 341, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 547, + 374 + ], + "type": "text", + "content": "Figure 7. Sample Image Generations for a given metaphorical message (shown on top) with Imagen [42], Stable Diffusion [41] and fine-tuned (FT) version of Stable Diffusion." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 210, + 536, + 329 + ], + "blocks": [ + { + "bbox": [ + 426, + 210, + 536, + 329 + ], + "lines": [ + { + "bbox": [ + 426, + 210, + 536, + 329 + ], + "spans": [ + { + "bbox": [ + 426, + 210, + 536, + 329 + ], + "type": "image", + "image_path": "6e9a00b90199accce31f522c019995823bcd16b8e9949fb7bcf904d536de09cf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 389, + 544, + 463 + ], + "blocks": [ + { + "bbox": [ + 309, + 389, + 544, + 463 + ], + "lines": [ + { + "bbox": [ + 309, + 389, + 544, + 463 + ], + "spans": [ + { + "bbox": [ + 309, + 389, + 544, + 463 + ], + "type": "image", + "image_path": "317d1e758b9dbbcecdd3e33759232a5e45f0b0ad5239d742ba643c80836fdac5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 472, + 546, + 518 + ], + "lines": [ + { + "bbox": [ + 305, + 472, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 546, + 518 + ], + "type": "text", + "content": "Figure 8. User Study on Image Generation Results. Percentage of results users preferred across real, Imagen [42], Stable Diffusion (SD) [41] and its fine-tuned version (SD-FT) results. Users are asked to choose the image that better depicts a given metaphor." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 536, + 379, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 379, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 379, + 548 + ], + "type": "text", + "content": "4. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 712 + ], + "type": "text", + "content": "In this paper, we present a step towards comprehensive evaluation of progress on visual metaphor research. Specifically, we propose a collection of tasks related to comprehending and generating visual metaphors using AI techniques. Our MetaCLUE tasks include Classification, Understanding (Retrieval, Captioning, VQA), Localization and Generation. For comprehensive evaluations, we collected high quality and rich annotations that facilitate the measurable progress. Existing methods demonstrate poor results in many cases with our experimental analysis shedding light on strengths and drawbacks of different approaches paving a path for future research in this fascinating field." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23208" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569-11579, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 287, + 178 + ], + "type": "text", + "content": "[2] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. SPICE: semantic propositional image caption evaluation. In ECCV, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "type": "text", + "content": "[3] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In ACL Workshops, 2005. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 270, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 270, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 270, + 224 + ], + "type": "text", + "content": "[4] Roland Barthes. Image-music-text. Macmillan, 1977. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 226, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 226, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 226, + 287, + 269 + ], + "type": "text", + "content": "[5] Nathalie Bonnardel and Evelyne Marmèche. Towards supporting evocation processes in creative design: A cognitive approach. International journal of human-computer studies, 63(4-5):422-435, 2005. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 303 + ], + "type": "text", + "content": "[6] Robert W Boozer, David C Wyld, and James Grant. Using metaphor to create more effective sales messages. Journal of services marketing, 1990. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 304, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 369 + ], + "type": "text", + "content": "[7] Tuhin Chakrabarty, Xurui Zhang, Smaranda Muresan, and Nanyun Peng. Mermaid: Metaphor generation with symbolism and discriminative decoding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4250-4261, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 287, + 414 + ], + "type": "text", + "content": "[8] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO Captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 415, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 287, + 480 + ], + "type": "text", + "content": "[9] Xianyang Chen, Chee Wee Leong, Michael Flor, and Beata Beigman Klebanov. Go figure! multi-task transformer-based architecture for metaphor detection using idioms: Ets team in 2020 metaphor shared task. In Proceedings of the second workshop on figurative language processing, pages 235-243, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 590 + ], + "type": "text", + "content": "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[11] Lydia B Chilton, Savvas Petridis, and Maneesh Agrawala. Visiblends: A flexible workflow for visual blends. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 637, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 287, + 712 + ], + "type": "text", + "content": "[12] Minjin Choi, Sunkyung Lee, Eunseong Choi, Heesoo Park, Junhyuk Lee, Dongwon Lee, and Jongwuk Lee. Melbert: Metaphor detection via contextualized late interaction using metaphorical identification theories. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1763-1773, 2021. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "text", + "content": "[14] Gilles Fauconnier and Mark Turner. The way we think: Conceptual blending and the mind's hidden complexities. Basic books, 2008. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "type": "text", + "content": "[15] Charles Forceville. Pictorial metaphor in advertising. Routledge, 1996. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "text", + "content": "[16] Susan R Fussell and Mallie M Moss. Figurative language in emotional communication. Social and cognitive approaches to interpersonal communication, pages 113-141, 1998. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "text", + "content": "[17] Hongyu Gong, Kshitij Gupta, Akriti Jain, and Suma Bhat. Illinimet: Illinois system for metaphor detection with contextual and linguistic information. In Proceedings of the Second Workshop on Figurative Language Processing, pages 146-153, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "text", + "content": "[18] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "text", + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "type": "text", + "content": "[20] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic understanding of image and video advertisements. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1705-1715, 2017. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 469 + ], + "type": "text", + "content": "[21] Bipin Indurkhya. On the role of metaphor in creative cognition. In ICCC, pages 51-59, 2010. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "type": "text", + "content": "[22] Bipin Indurkhya and Amitash Ojha. An empirical study on the role of perceptual similarity in visual metaphors and creativity. Metaphor and Symbol, 28(4):233-253, 2013. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 558 + ], + "type": "text", + "content": "[23] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 578 + ], + "type": "text", + "content": "[24] George Lakoff. The contemporary theory of metaphor. 1993. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 581, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 602 + ], + "type": "text", + "content": "[25] George Lakoff and Mark Johnson. Metaphors we live by. University of Chicago press, 2008. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "type": "text", + "content": "[26] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "text", + "content": "[27] Jiahao Li, Greg Shakhnarovich, and Raymond A Yeh. Adapting clip for phrase localization without further training. arXiv preprint arXiv:2204.03647, 2022. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "type": "text", + "content": "[28] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23209" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[29] Todd I Lubart and Isaac Getz. Emotion, metaphor, and the creative process. Creativity research journal, 10(4):285-301, 1997. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "text", + "content": "[30] Carita Lundmark. Metaphor and creativity in British magazine advertising. PhD thesis, Luleå teknska universitet, 2005. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "type": "text", + "content": "[31] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "text", + "content": "[32] James H Martin. A corpus-based analysis of context effects on metaphor comprehension. Trends in Linguistics Studies and Monographs, 171:214, 2006. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "type": "text", + "content": "[33] Zachary J Mason. Cormet: a computational, corpus-based conventional metaphor extraction system. Computational linguistics, 30(1):23-44, 2004. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "text", + "content": "[34] Edward F McQuarrie and David Glen Mick. Visual rhetoric in advertising: Text-interpretive, experimental, and reader-response analyses. Journal of consumer research, 26(1):37-54, 1999. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "text", + "content": "[35] Ekaterina Ovchinnikova, Vladimir Zaytsev, Suzanne Wertheim, and Ross Israel. Generating conceptual metaphors from proposition stores. arXiv preprint arXiv:1409.7619, 2014. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 387 + ], + "type": "text", + "content": "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. 2002. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "text", + "content": "[37] Barbara J Phillips. The impact of verbal anchoring on consumer response to image ads. Journal of advertising, 29(1):15-24, 2000. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 423, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 455 + ], + "type": "text", + "content": "[38] Barbara J Phillips and Edward F McQuarrie. Beyond visual metaphor: A new typology of visual rhetoric in advertising. Marketing theory, 4(1-2):113-136, 2004. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 457, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 287, + 521 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 524, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 555 + ], + "type": "text", + "content": "[40] Sunny Rai and Shampa Chakraverty. A survey on computational metaphor processing. ACM Computing Surveys (CSUR), 53(2):1-37, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 679 + ], + "type": "text", + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[43] Linda M Scott. Images in advertising: The need for a theory of visual rhetoric. Journal of consumer research, 21(2):252-273, 1994. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[44] Gerard Steen. A method for linguistic metaphor identification: From MIP to MIPVU, volume 14. John Benjamins Publishing, 2010. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 182 + ], + "type": "text", + "content": "[45] Kevin Stowe, Tuhin Chakrabarty, Nanyun Peng, Smaranda Muresan, and Iryna Gurevych. Metaphor generation with conceptual mappings. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6724-6736, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "type": "text", + "content": "[46] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[47] Asuka Terai and Masanori Nakagawa. A computational system of metaphor generation with evaluation mechanism. In International Conference on Artificial Neural Networks, pages 142-147. Springer, 2010. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 273, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 338 + ], + "type": "text", + "content": "[48] Xiaoyu Tong, Ekaterina Shutova, and Martha Lewis. Recent advances in neural metaphor processing: A linguistic, cognitive and social perspective. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4673-4686, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[49] Peter Turney, Yair Neuman, Dan Assaf, and Yohai Cohen. *Literal and metaphorical sense identification through concrete and abstract context*. In *Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing*, pages 680–690, 2011. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 435 + ], + "type": "text", + "content": "[50] Tony Veale, Ekaterina Shutova, and Beata Beigman Klebanov. Metaphor: A computational perspective. Synthesis Lectures on Human Language Technologies, 9(1):1-160, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "text", + "content": "[51] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "text", + "content": "[52] Merryl J Wilkenfeld and Thomas B Ward. Similarity and emergence in conceptual combination. Journal of Memory and Language, 45(1):21-38, 2001. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 505, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 568 + ], + "type": "text", + "content": "[53] Bo Xu, Tingting Li, Junzhe Zheng, Mehdi Naseriparsa, Zhehuan Zhao, Hongfei Lin, and Feng Xia. Met-meme: A multimodal meme dataset rich in metaphors. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2887-2899, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[54] Keren Ye and Adriana Kovashka. Advise: Symbolism and external knowledge for decoding advertisements. In Proceedings of the European Conference on Computer Vision (ECCV), pages 837-855, 2018. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[55] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[56] Zhiwei Yu and Xiaojun Wan. How to avoid sentences spelling boring? towards a neural approach to unsupervised metaphor generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23210" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 205 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 861-871, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 171 + ], + "type": "text", + "content": "[57] Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, and Hongfei Lin. Multimet: A multimodal dataset for metaphor understanding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3214-3225, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "type": "text", + "content": "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 7" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23211" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_content_list.json b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3ace6443608fb04ee5c1e5ce46fa305ea06835a6 --- /dev/null +++ b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_content_list.json @@ -0,0 +1,2114 @@ +[ + { + "type": "text", + "text": "MetaFusion: Infrared and Visible Image Fusion via Meta-Feature Embedding from Object Detection", + "text_level": 1, + "bbox": [ + 91, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenda Zhao $^{1}$ , Shigeng Xie $^{1}$ , Fan Zhao $^{2*}$ , You He $^{3}$ , Huchuan Lu $^{1}$", + "bbox": [ + 220, + 202, + 746, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dalian University of Technology, China", + "bbox": [ + 321, + 220, + 648, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Liaoning Normal University, China", + "bbox": [ + 338, + 238, + 630, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Tsinghua University, China", + "bbox": [ + 367, + 256, + 599, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhaowenda@dlut.edu.cn; xieshigeng@mail.dlut.edu.cn", + "bbox": [ + 263, + 277, + 705, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fan-Zhao20@163.com; youhe_nau@163.com; lhchuan@dlut.edu.cn", + "bbox": [ + 228, + 295, + 735, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fusing infrared and visible images can provide more texture details for subsequent object detection task. Conversely, detection task furnishes object semantic information to improve the infrared and visible image fusion. Thus, a joint fusion and detection learning to use their mutual promotion is attracting more attention. However, the feature gap between these two different-level tasks hinders the progress. Addressing this issue, this paper proposes an infrared and visible image fusion via meta-feature embedding from object detection. The core idea is that meta-feature embedding model is designed to generate object semantic features according to fusion network ability, and thus the semantic features are naturally compatible with fusion features. It is optimized by simulating a meta learning. Moreover, we further implement a mutual promotion learning between fusion and detection tasks to improve their performances. Comprehensive experiments on three public datasets demonstrate the effectiveness of our method. Code and model are available at: https://github.com/wdzhao123/MetaFusion.", + "bbox": [ + 73, + 376, + 473, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modality sensor technology has promoted the application of multi-modality images in different areas. Among them, infrared images and visible images have been utilized commonly, as the information contained in these two modalities is complementary. Specifically, infrared images can supply object thermal structures without being affected by illumination. But they are short of texture details. On the contrary, visible images can catch the texture information for the scene. But they are severely affected by light. Thus, many methods [15,25,35,43-45,47] focus on s", + "bbox": [ + 75, + 722, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a829f059b0b06fa7ccf0112ee4701d81542d6560ff3a0acd00647ef62525dea4.jpg", + "image_caption": [ + "(a) Separate optimization method" + ], + "image_footnote": [], + "bbox": [ + 516, + 344, + 885, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e2b9f5f41446252228b71c0564d6a3d37cf305c7addcac4d9a0d9e12f1c36f1.jpg", + "image_caption": [ + "(b) Cascaded optimization method" + ], + "image_footnote": [], + "bbox": [ + 514, + 434, + 887, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d75aaed38dfc8eca509b018fd0184e39f1718fff02a6b437df2ce7e876fb386.jpg", + "image_caption": [ + "(c) Meta-feature embedding method", + "Figure 1. Different joint learning methods of infrared and visible image fusion (IVIF) and object detection (OD). (a) Separate optimization method: IVIF network $\\Psi$ is firstly optimized by fusion loss $L_{f}$ . Then, fusion result $z$ is generated by $\\Psi$ from the input infrared and visible image pair $x,y$ . Finally, OD network $\\Phi$ is optimized by detection loss $L_{d}$ using $z$ . (b) Cascaded optimization method: OD network $\\Phi$ is treated as a constraint to optimize IVIF network $\\Psi$ by the loss $L_{f}$ and $L_{d}$ . (c) Meta-feature embedding method: Meta-feature embedding $\\Gamma$ is optimized to learn how to guide $\\Psi$ to have a low $L_{f}$ . Then, $\\Gamma$ generates meta feature from detection feature $F_{i}^{d}$ and fusion feature $F_{i}^{d}$ . Finally, the meta feature is used to guide $\\Psi$ by the loss $L_{d}$ ." + ], + "image_footnote": [], + "bbox": [ + 511, + 518, + 885, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tudying pixel-level infrared and visible image fusion (IVIF), thereby helping high-level tasks improve performance, e.g., object detection (OD) [20, 34].", + "bbox": [ + 496, + 835, + 892, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IVIF and OD can greatly benefit form each other. IVIF", + "bbox": [ + 517, + 885, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13955", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "generates the fused image that contains more information than any single modality image to improve OD. However, IVIF mainly focuses on the pixel relationship between an image pair, and there is little consideration for object semantic. In contrast, OD can provide rich object semantic information to IVIF, as its aim is locating the objects. Therefore, this paper studies a joint learning framework between IVIF and OD to improve their performances.", + "bbox": [ + 75, + 90, + 470, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing joint learning methods of IVIF and OD can be divided into two categories: separate optimization and cascaded optimization. Separate optimization firstly trains IVIF network, and then trains OD network using IVIF results, as shown in Figure 1(a). Thus, most methods focus on improving the fusion effect, e.g., designing networks [13, 23, 35, 36] and introducing constraints [10, 26, 45]. Obviously, separate optimization neglects the help of OD. Cascaded optimization adopts OD network as a constraint to train IVIF network, and thus forces the IVIF network to generate fusion images with easily detected objects [20], as shown in Figure 1(b). However, directly utilizing the high-level OD constraint to guide the pixel-level IVIF will result in limited effect. Therefore, we leverage OD feature maps that guide IVIF feature maps to obtain more semantic information. Unfortunately, OD features are mismatched with IVIF features due to their task-level difference. Addressing this issue, we propose a meta-feature embedding network (MFE), as shown in Figure 1(c). The idea is that if MFE generates OD features according to the IVIF network ability, the OD features are naturally compatible with the IVIF network, and the optimization can be achieved by simulating a meta learning.", + "bbox": [ + 75, + 215, + 470, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, an infrared and visible image fusion via meta-feature embedding from object detection is proposed, which is named as MetaFusion. MetaFusion includes IVIF network $(F)$ , OD network $(D)$ , and $MFE$ . In particular, $MFE$ is expected to generate meta features to bridge the gap between $F$ and $D$ , which is optimized by two alternate steps: inner update and outer update. In the inner update process, we firstly optimize $F$ using meta training set $S_{mtr}$ to obtain its updated network $F'$ . Then, $F'$ calculates the fusion loss on meta testing set $S_{mts}$ to optimize $MFE$ . The motivation is that if $MFE$ successfully generates meta features which are compatible with $F$ , $F'$ will produce better fused images, i.e., the fusion loss should be lower. In the outer update process, $F$ is optimized with the guide of the meta features generated by the fixed $MFE$ on $S_{mtr}$ and $S_{mts}$ . In this way, $F$ can learn how to extract semantic information to improve fusion quality. In the above two alternate steps, $D$ is fixed to offer detection semantic information. Thus, we further implement a mutual promotion learning, where we use the improved $F$ to generate fusion results to finetune $D$ , and then the improved $D$ offers better semantic information to optimize $F$ .", + "bbox": [ + 75, + 568, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows. (1) We explore the joint learning framework of IVIF and OD, and propose MetaFusion to obtain superior performance on these two tasks. (2) Meta-feature embedding network is designed to generate meta features that bridge the gap between $F$ and $D$ . (3) Sequentially, mutual promotion learning between $F$ and $D$ is introduced to improve their performances. (4) Extensive experiments on image fusion and object detection validate the effectiveness of the proposed method.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 239, + 640, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Fusion. Traditional image fusion methods use hand-crafted features, such as sparse representation [22], spectral variation [53] and low-rank representation [14], which can not handle the complex scenes well. Nowadays, deep learning-based image fusion methods [16, 24, 27, 39, 49, 50, 54, 56] are proposed. Ma et al. [23] propose a swim transformer based image fusion method with cross-domain long-range learning. Xu et al. [44] use feature extraction and measurement to estimate the degree of information preservation in image fusion. Zhao et al. [51] design a self-supervised feature extraction model. Besides, feature diversity enhancement and fusion is received attention [52, 55]. However, most of them neglect the help of high-level tasks. Recently, Liu et al. [20] propose a joint learning of IVIF and OD. They treat OD network as an additional constraint to help IVIF network generate fusion result with more clear objects.", + "bbox": [ + 496, + 266, + 890, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, directly using the high-level OD network to guide pixel-level IVIF may not obtain an effective constraint. Moreover, they use the fusion results as the connection between IVIF network and OD network, ignoring the pixel-level semantic information in OD features. In contrast, we adopt semantic information in OD features to help IVIF, and conduct a meta-feature embedding to generate meta features from OD features. Then, those meta features are used to guide IVIF network to learn pixel-level semantic information.", + "bbox": [ + 496, + 523, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object Detection. Deep learning-based object detection methods have made great progress, e.g., network structure-designed methods [2, 8, 30, 31] and loss constraint methods [18, 32, 42]. Carion et al. [3] introduce a transformer encoder-decoder architecture for object detection. Xu et al. [42] rank positive and negative sample pairs and improve the ranking effect by a clustering algorithm to built object detection loss. On the other hand, many weakly supervised methods [40, 46, 48] are proposed. Wu et al. [40] design a model to suppress the background activation and obtain object region, where image-level labels are used. Zhang et al. [48] use instance proposal grouping to learn object detection from a single point annotation.", + "bbox": [ + 496, + 674, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of those object detection methods leverage single modality image, e.g., visible images. However, the visible", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13956", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/caef070d7ffe13a45f7345ec4df249e9870fe0669478a6b797e9d39d2eb76a5e.jpg", + "image_caption": [ + "Figure 2. Framework illustration of the proposed MetaFusion. MetaFusion includes three parts: infrared and visible image fusion (IVIF), object detection (OD), and meta-feature embedding (MFE). IFIV network includes four blocks: three feature fusion blocks $FFB_{1}$ , $FFB_{2}$ , $FFB_{3}$ used to extract and fuse features and one image reconstruction module $IRM$ used to reconstruct fusion result. MFE contains three meta-feature generators $MFG_{1}$ , $MFG_{2}$ , $MFG_{3}$ conducted to generate multi-level meta features from OD features $F_{ej}$ , and three feature transform networks $FT_{1}$ , $FT_{2}$ , $FT_{3}$ used to transfer the meta features to fusion features. OD network is divided as three feature extraction blocks $DFB_{1}$ , $DFB_{2}$ , $DFB_{3}$ and one OD detection head $DH$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 869, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "images are affected by light conditions, thereby influencing the performance of object detection. On the contrary, we leverage IVIF results to provide infrared information without being influenced by illumination, therefore can offer more stable detection performance.", + "bbox": [ + 75, + 420, + 468, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta Learning. Meta learning aims to use the learned knowledge to quickly adapt to new tasks, which has been used to many fields: model pre-training [5, 11, 29], image classification [17, 19], few-shot learning [1, 4, 6, 41], etc. For example, Raghu et al. [29] propose an implicit differentiation and backpropagation meta-learning method and improve the efficiency of learning pre-training hyperparameters. Xu et al. [41] design a dynamic alignment in meta learning, which can emphasize information in query according to the support. Li et al. [12] use the meta learning to handle different input resolutions, thereby generating fusion results with arbitrary resolutions.", + "bbox": [ + 75, + 496, + 468, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the above meta learning methods, we design a meta-feature embedding network to generate the meta features from object detection features, thereby helping IVIF network fuse more object semantic information.", + "bbox": [ + 75, + 678, + 468, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 753, + 250, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our MetaFusion framework is shown in Figure 2, which includes three subnetworks: IVIF network $(F)$ generates fusion results, OD network $(D)$ offers semantic features, and meta-feature embedding network $(MFE)$ uses OD features to guide IVIF to fuse more object semantic information. The challenge is that OD features can hardly be directly used to guide IVIF due to their task gap. Thus, our focus is to design $MFE$ and optimize it. Details are as follows.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/136eafc183a7d3af98a5a0e498950ee091625d9583a47f04d88607982d561a7c.jpg", + "image_caption": [ + "Figure 3. Illustration of the meta-feature embedding learning. There are two stages in the meta-feature embedding: (a) inner update and (b) outer update. In the inner update, $MFG$ and $F^T$ are trained according to current $F'$ semantic extracting ability to generate meta features. In the outer update, $F$ is trained which is guided by the meta features, thereby learning to extract object semantic. The two stages alternate every N epochs." + ], + "image_footnote": [], + "bbox": [ + 509, + 420, + 897, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Meta-feature Embedding", + "text_level": 1, + "bbox": [ + 500, + 693, + 732, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta-feature embedding includes meta-feature generator (MFG) and feature transform (FT). MFG generates meta feature $F_{mj}$ according to IVIF feature $F_{uj}$ from OD feature $F_{ej}$ , i.e., $F_{mj} = MFG(F_{uj}, F_{ej})$ , where $j$ is the feature level index. FT transfers the meta feature $F_{mj}$ to fusion feature $F_{uj}$ by producing feature bridge $F_{tj}$ . Specifically, meta-feature embedding learning is divided two stages: inner update stage and outer update stage, as shown in Figure 3.", + "bbox": [ + 496, + 717, + 893, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the inner update stage, we optimize $MFG$ , $FT$ and $F$ (see Figure 3(a)). Given meta training set $S_{mtr}$ and meta testing set $S_{mts}$ , we firstly update the parameter $\\theta_F$ of $F$ on", + "bbox": [ + 500, + 854, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13957", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$S_{mtr}$ through meta feature guidance:", + "bbox": [ + 76, + 90, + 323, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {F ^ {\\prime}} = \\theta_ {F} - \\beta_ {F ^ {\\prime}} \\nabla_ {\\theta_ {F}} L _ {g} \\left(F _ {m j}, F _ {t j}\\right) = \\theta_ {F} - \\beta_ {F ^ {\\prime}} \\frac {\\partial L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 116, + 468, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the guide loss $L_{g}$ is $L_{2}$ distance, $\\beta_{F^{\\prime}}$ is the learning rate of $F^{\\prime}$ , $\\theta_{F^{\\prime}}$ is the parameter of $F^{\\prime}$ that is updated from $F$ . Then, we use $F^{\\prime}$ with parameter $\\theta_{F^{\\prime}}$ to calculate the fusion loss $L_{f}$ on meta testing set $S_{mts}$ . Here, $L_{f}$ can measure the effect of using the meta features to guide $F$ , i.e., when $L_{f}$ is lower, the semantic extracting ability of $F$ will be improved. Therefore, we use $L_{f}$ to update parameter $\\theta_{MFG}$ of $MFG$ and parameter $\\theta_{FT}$ of $FT$ by", + "bbox": [ + 76, + 161, + 472, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {M F G} = \\theta_ {M F G} - \\beta_ {M F G} \\nabla_ {\\theta_ {M F G}} L _ {f} \\left(I _ {f} ^ {m t s}, I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 292, + 468, + 311 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {F T} = \\theta_ {F T} - \\beta_ {F T} \\nabla_ {\\theta_ {F T}} L _ {f} \\left(I _ {f} ^ {m t s}, I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 321, + 468, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the fusion loss $L_{f}$ is SSIM loss [38], $I_{i}^{mts}$ and $I_{v}^{mts}$ are infrared and visible images from $S_{mts}$ , $I_{f}^{mts} = F'(I_{i}^{mts}, I_{v}^{mts})$ is the meta testing fusion result, and $\\beta_{MFG}$ and $\\beta_{FT}$ are the learning rate of MFG and $FT$ . $\\nabla_{\\theta_{MFG}} L_{f}$ can be calculated by", + "bbox": [ + 76, + 345, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta_ {M F G}} L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right) \\\\ = \\frac {\\partial L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right) , I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right)}{\\partial \\theta_ {F ^ {\\prime}}} * \\left(- \\frac {\\partial^ {2} L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F} \\partial \\theta_ {M F G}}\\right). \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 433, + 468, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, $\\nabla_{\\theta_{FT}}L_f$ can be calculated by", + "bbox": [ + 76, + 500, + 341, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta_ {F T}} L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right) \\\\ = \\frac {\\partial L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right) , I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right)}{\\partial \\theta_ {F ^ {\\prime}}} * \\left(- \\frac {\\partial^ {2} L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F} \\partial \\theta_ {F T}}\\right). \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 525, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, $MFG$ can learn how to generate meta features according to current $F^{\\prime}$ semantic extracting ability, i.e., makes meta features be compatible with $F$ .", + "bbox": [ + 76, + 590, + 468, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the outer update stage, $F$ is trained with the initial parameter $\\theta_F$ , as shown in Figure 3(b). Given the training data $S = \\{S_{mtr}, S_{mts}\\}$ , $F$ is optimized using the fusion loss $L_f$ and guide loss $L_g$ by", + "bbox": [ + 76, + 636, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta_ {F} = \\theta_ {F} - \\beta_ {F} \\nabla_ {\\theta_ {F}} \\left(L _ {f} \\left(I _ {f}, I _ {i}, I _ {v}\\right) + \\lambda_ {g} \\sum_ {j = 1} ^ {3} L _ {g} \\left(F _ {m j}, F _ {t j}\\right)\\right) \\\\ = \\theta_ {F} - \\beta_ {F} \\left(\\frac {\\partial L _ {f} \\left(I _ {f} , I _ {i} , I _ {v}\\right)}{\\partial \\theta_ {F}} + \\lambda_ {g} \\sum_ {j = 1} ^ {3} \\frac {\\partial L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F}}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 708, + 470, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{i}$ and $I_{v}$ are infrared and visible images from $S$ , $\\lambda_{g}$ is a hyper-parameter to balance $L_{f}$ and $L_{g}$ , and $\\beta_{F}$ is the learning rate of $F$ .", + "bbox": [ + 76, + 810, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the inner update stage and outer update stage are carried out alternately for every $N$ epochs, which improves the optimization of $MFG$ and $F$ .", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ce69d4d5025548242fcc713ea43b1a5aa3a141ca0f701fe1e91fc1559f44905.jpg", + "image_caption": [ + "Figure 4. Illustration of the mutual promotion learning between $F$ and $D$ . Firstly, $F$ is used to generate fusion results to optimize $D$ . Thus, a fine-tuned $D$ with higher semantic extracting ability is obtained. Then, $D$ is adopted to optimize $F$ . This process is alternated." + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 864, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Mutual Promotion Learning", + "text_level": 1, + "bbox": [ + 500, + 441, + 754, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Sec. 3.1, we fix the object detection network $D$ to offer stable object semantic features. In this way, $MFG$ does not need to handle variable object semantic features, thereby reducing the difficulty of training $MFG$ . However, the fixed $D$ will restrict the object semantic extracting ability of $F$ . Addressing this problem, we propose the mutual promotion learning that promotes not only $F$ but also $D$ , as shown in Figure 4. Specifically, we firstly train $F$ individually, and then use the fusion results to train $D$ . Sequentially, $D$ offers object semantic features to the meta-feature embedding, thereby improving $F$ . After that, the improved fusion results generated by $F$ are used to fine-tune $D$ . In this way, both $F$ and $D$ are improved.", + "bbox": [ + 498, + 467, + 890, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Architecture", + "text_level": 1, + "bbox": [ + 500, + 679, + 633, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IVIF network $F$ aims to generate fusion results according to the input infrared and visible images, which is flexible. Here, we adopt the similar structure with [50]. $F$ consists of three feature fusion blocks $(FFB_{j}, j = 1,2,3)$ and one image reconstruction module (IRM), as shown in Figure 2. $FFB_{j}$ is used to extract and fuse features from input infrared image $I_{i}$ and visible image $I_{v}$ . It is designed based on the idea of [50] that if $FFB_{j}$ can well fuse the features of $I_{i}$ and $I_{v}$ , then the two kinds of features can be recovered from the fused feature. Thus, $FFB_{j}$ is described as", + "bbox": [ + 498, + 703, + 890, + 856 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF F B _ {j} = \\left\\{ \\begin{array}{c} C _ {2} T _ {1} \\biguplus C _ {2} T _ {1}, j = 1 \\\\ C _ {2} T _ {j - 1} \\biguplus C _ {2} T _ {j - 1}, j > 1 \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 869, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13958", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j$ is the block index, $C_k$ represents the number of \"Convolution with $3 \\times 3$ kernel +ReLU\" layers is $k$ , $F_{u(j-1)}$ is the output feature of the $j-1$ th block, $T_1 = C_2(I_i, I_v)$ and $T_{j-1} = C_2(F_{u(j-1)})$ represent feature integration operations, and $\\boxplus$ means feature concatenation. IRM reconstructs the fusion result, which contains six \"Convolution with $3 \\times 3$ kernel +ReLU\" layers.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object detection network $D$ offers object semantic features. In our framework, we choose $\\mathrm{yolov5s}^1$ as our $D$ . According to the size of the feature maps, the backbone of $D$ is divided into three feature extraction blocks $(\\{DFB_{i}, i = 1,2,3\\})$ . In addition, we represent the neck and detection head in yolov5s into one block $(DH)$ to simplify the expression.", + "bbox": [ + 75, + 196, + 468, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meta-feature embedding network $MFE$ includes metafeature generator $MFG_{j}$ and feature transform network $FT_{j}$ . Since OD feature $F_{ej}$ generated by $DFB_{j}$ offers object semantic information and $F_{uj}$ provides scene details, they are fed into the $MFG_{j}$ . Then, $MFG_{j}$ is built as", + "bbox": [ + 75, + 301, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM F G _ {j} = C _ {6} \\left(C _ {4} \\left(F _ {u j}\\right) \\biguplus C _ {2} \\left(U p \\left(F _ {e j}\\right)\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 390, + 468, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Up()$ is the feature up-sampling. $FT_{j}$ transfers the meta feature to fusion feature, which includes three \"Convolution with $3 \\times 3$ kernel +ReLU\" layers.", + "bbox": [ + 75, + 421, + 468, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 76, + 476, + 181, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training details of our framework is shown in Algorithm 1, which contains four steps. Generally, we firstly pretrain $F$ , and then use the fusion results to finetune $D$ . Sequentially, we train $FT$ and $MFG$ , where $MFG$ and $FT$ are updated $n = 200$ times every $N = 8$ epochs. Lastly, the mutual promotion between $F$ and $D$ is trained for $R = 2$ rounds, thereby improving their performances.", + "bbox": [ + 75, + 500, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 619, + 207, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 643, + 156, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct experiment on three widely-used datasets: $\\mathbf{M}^3\\mathrm{FD}$ [20], RoadScene [44] and TNO [37]. $\\mathbf{M}^3\\mathrm{FD}$ is divided as training set (2940 image pairs) and testing set (1260 image pairs). RoadScene with 221 image pairs and TNO with 40 image pairs are only used for testing. Besides, $\\mathbf{M}^3\\mathrm{FD}$ is adopted to evaluate OD performance.", + "bbox": [ + 75, + 666, + 468, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation. Our framework is implemented with PyTorch on a NVIDIA GeForce RTX 3090 GPU. $F$ and MFE are trained using optimizer Adam with the learning rate $\\beta_{F},\\beta_{F^{\\prime}},\\beta_{FT}$ and $\\beta_{MFG}$ of $1\\times 10^{-3}$ , respectively. $D$ is trained followed by yolov5s that adopts the optimizer SGD using the learning rate of $1\\times 10^{-2}$ with decaying rate of 0.1 every round. We firstly train $F$ for 100 epochs and $D$ for 150 epochs. Then, we conduct MFE for 50 epochs. After", + "bbox": [ + 75, + 757, + 468, + 877 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 MetaFusion Training Algorithm" + ], + "code_body": "Input: Training dataset $S = \\{S_{mtr}, S_{mts}\\}$ , IVIF model $F(\\theta_F)$ , meta-feature generator $MFG(\\theta_F)$ , feature transform network $FT(\\theta_{FT})$ , OD network $D(\\theta_D)$ \nOutput: IVIF model $F(\\theta_F^*)$ \n1: Initialize $F(\\theta_F)$ , $MFG(\\theta_F)$ , $FT(\\theta_{FT})$ , $D(\\theta_D)$ \n2: /* Pretrain $F^{*}/$ \n3: while not converged do \n4: Sample image pair $(I_i, I_v)$ from $S$ \n5: Forward pass on $F$ to generate fusion result $I_f$ \n6: Optimize $F$ using fusion loss $L_f$ \n7: end while \n8: /* Pretrain $D^{*}/$ \n9: while not converged do \n10: Sample image $I_f$ from the fused images generated by $F$ \n11: Forward pass on $D$ to get detection result \n12: Optimize $D$ using detection loss $L_d$ \n13: end while \n14: /* Meta-feature embedding outer update */ \n15: while not converged do \n16: Sample a batch image pair $(I_i, I_v)$ from $S$ \n17: Forward pass on $F$ to generate fusion result $I_f$ \n18: Forward pass on $MFG$ to generate meta feature \n19: Optimize $F$ using fusion loss $L_f$ and guide loss $L_g$ by Eq. 6 \n20: /* Meta-feature embedding inner update */ \n21: if epoch%N==0 then \n22: for t=1 to n do \n23: Sample an image pair $(I_i^{mtr}, I_v^{mtr})$ from $S_{mtr}$ \n24: Compute $\\theta_{F'}$ by Eq. 1 \n25: Sample an image pair $(I_i^{mts}, I_v^{mts})$ from $S_{mts}$ \n26: Calculate $L_f$ using $\\theta_{F'}$ \n27: Update $\\theta_{MFG}$ and $\\theta_{FT}$ using the gradient of $L_f$ by Eqs. 2 and 3 \n28: end for \n29: end if \n30: end while \n31: /* Mutual promotion*/ \n32: if Number of round ≤ R then \n33: Go back to line 9 \n34: end if", + "bbox": [ + 500, + 109, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that, $D$ is finetuned for 150 epochs. Following the same strategy, the mutual promotion is carried out $R$ rounds. The images are resized to $512 \\times 384$ with batchsize of 1. The hyperparameter $\\lambda_{g}$ is set to 0.1.", + "bbox": [ + 496, + 762, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metric. Three metrics are used for IVIF evaluation: entropy (EN) [33], mutual information (MI) [28] and visual information fidelity (VIF) [7]. EN evaluates the information richness in an image, and the higher EN means more information. MI evaluates the information similarity between", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/ultralytics/YOLOv5", + "bbox": [ + 94, + 886, + 308, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13959", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad293010675822b131f424cffc076bde8fd54c2a123d2075e4dbce42638ad3ab.jpg", + "image_caption": [ + "(a) Visible-infrared images" + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 184, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d84997b06a4f246fb6b57071c6fe16f6d4755bcd016e6e8d08defa9513f78a5.jpg", + "image_caption": [ + "(b) SoFusion and feature map" + ], + "image_footnote": [], + "bbox": [ + 186, + 90, + 281, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9391a83037ed5dc068fd2a84eaff4e7bc2f11318622a7a96427f068c312ef89c.jpg", + "image_caption": [ + "(c) CoFusion and feature map", + "Figure 5. Visual comparison of different optimization methods." + ], + "image_footnote": [], + "bbox": [ + 282, + 90, + 375, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc40a81a98b73fc408062d9cc005fd1b877e9acc7ee712b0e4d50c63f03f89f2.jpg", + "image_caption": [ + "(d) MetaFusion and feature map" + ], + "image_footnote": [], + "bbox": [ + 377, + 90, + 470, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Effect study of meta-feature embedding by comparing with different optimization methods on $\\mathrm{M}^3\\mathrm{FD}$ . The best result is in red.", + "bbox": [ + 76, + 226, + 467, + 266 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1a132f492a4064ce525aa1dd2d7804c837e88a24fdad39bfebb5c46d4cd7d5fd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMetric
MIENVIF
SoFusion14.1647.0781.190
CoFusion14.1877.0891.345
MetaFusion14.5117.2491.515
", + "bbox": [ + 93, + 273, + 450, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the input images and fused image. The higher MI illustrates more information of the input images is fused. VIF measures the ability to extract visible information from the input image, and a larger VIF represents less visible distortion in the fused result. Here, we use the V channel in HSV space of the fusion results to calculate these metrics. Moreover, we use $\\mathrm{mAP}_{50\\rightarrow 95}$ [9] to comprehensively evaluate OD performance, where the average of mAPs sampling every 5 from $\\mathrm{AP}_{50}$ to $\\mathrm{AP}_{95}$ is calculated. A higher $\\mathrm{mAP}_{50\\rightarrow 95}$ means better OD effect.", + "bbox": [ + 75, + 397, + 467, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 559, + 228, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of meta-feature embedding. In Sec. 3.1, we introduce the meta-feature embedding that generates the meta features from object detection features to help IVIF network learn more object semantic information. To verify its effect, separate optimization method (SoFusion) and cascaded optimization method (CoFusion) are compared, as shown in Figure 1(a)-(b). Specifically, SoFusion firstly optimizes $F$ and then uses the generated fusion results to optimize $D$ . CoFusion treats $D$ as a constraint to optimize $F$ with the loss $L_{f}$ and $L_{d}$ . The results are shown in Table 1. Our MetaFusion achieves the best results on IVIF. The reason is that SoFusion neglects the help of $D$ , and directly utilizing the high-level $D$ to guide the pixel-level $F$ has feature mismatch in CoFusion. In contrast, we implement the meta-feature embedding that generates meta features from $D$ according to the ability of $F$ . Thus, the meta features are naturally compatible with $F$ , thereby providing effective object semantic features to $F$ . Visual comparison is shown in Figure 5. MetaFusion highlights the object feature and generates the clearest result.", + "bbox": [ + 75, + 582, + 468, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Influence of mutual promotion learning. As described", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d4ef4ee68911d1950f5d8435ee47b9769027ec586dc409647b4d42a5aa127a63.jpg", + "table_caption": [ + "Table 2. Influence study of mutual promotion learning by evaluating the performances of round $R$ on ${\\mathrm{M}}^{3}\\mathrm{{FD}}$ . The best result is in red." + ], + "table_footnote": [], + "table_body": "
MethodMetric
MIENVIFmAP50→95 (%)
R = 014.1647.0781.19055.6
R = 114.4647.2261.47455.8
R = 214.5117.2491.51556.5
", + "bbox": [ + 519, + 135, + 869, + 239 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e44d7b2893c890b4475788958f1b68e8005bb9368f25cce4b82165c1b8c6bba.jpg", + "table_caption": [ + "Table 3. Study of multi-level meta-feature embedding by comparing different configurations on $\\mathbf{M}^3\\mathrm{FD}$ . The best result is in red." + ], + "table_footnote": [], + "table_body": "
MethodMetric
MIENVIF
MetaFusion-L114.4507.2201.259
MetaFusion-L214.4527.2201.377
MetaFusion-L314.4647.2261.474
", + "bbox": [ + 514, + 284, + 874, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Sec. 3.2, we propose the mutual promotion learning promote $F$ and $D$ alternately. Here, we study the influence of mutual promotion learning by evaluating round $R = 0,1,2$ . In particular, $R = 0$ represents $F$ and $D$ are trained respectively. As shown in Table 2, the results of $F$ are better with the $R$ increases. Comprehensively considering training efficiency and performance, we take $R = 2$ .", + "bbox": [ + 496, + 409, + 890, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Study of multi-level meta-feature embedding. We implement multi-level meta-feature embedding to build MetaFusion framework, as introduced in Sec. 3.3. Here, we study the number of meta-feature embedding levels with the following configurations. One-level meta-feature embedding (MetaFusion-L1) with $MFE = \\{MFG_1, FT_1\\}$ , $F = \\{FFB_1, IRM\\}$ and $D = \\{DFB_1, DH\\}$ . Two-level meta-feature embedding (MetaFusion-L2) with $MFE = \\{MFG_1, MFG_2, FT_1, FT_2\\}$ , $F = \\{FFB_1, FFB_2, IRM\\}$ and $D = \\{DFB_1, DFB_2, DH\\}$ . Three-level meta-feature embedding (MetaFusion-L3) with $MFE = \\{MFG_1, MFG_2, MFG_3, FT_1, FT_2, FT_3\\}$ , $F = \\{FFB_1, FFB_2, FFB_3, IRM\\}$ and $D = \\{DFB_1, DFB_2, DFB_3, DH\\}$ . The results are shown in Table 3. With the number of multi-level meta-feature embedding increases, $F$ achieves higher image fusion quality. Since meta-feature embedding provide multi-level object semantic features to $F$ .", + "bbox": [ + 496, + 516, + 892, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 500, + 800, + 862, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the proposed MetaFusion with eight SOTA fusion methods to verify the superiority: FusionGAN [25], GANMcC [26], MFEIF [21], YDTR [36], PIAFusion [35], SwinFusion [23], Tardal [20], and U2Fusion [44]. Their available codes and recommended parameter settings are", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13960", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6948d69ed6adca343da4e9b96dfa341d6031457013efd9491b1519cbed6cd5f.jpg", + "image_caption": [ + "Infrared-visible image" + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 264, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88508f5965299148b31b56b5d6d3f8c58eccf04a84d59cd27c19d62f09d6a498.jpg", + "image_caption": [ + "FusionGAN" + ], + "image_footnote": [], + "bbox": [ + 267, + 92, + 410, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a54a70e631bac6372eb2f5827c768b359dd40a4dbfaef7a3813ca965493cef89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 92, + 555, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2e7f06cb59fbf4dbb9bac8f4ba5f94c633b094f1c20b5b14d5c0d20c8a33f35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 92, + 700, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/852d3dd4c719b3eb05074b4b0c4cc075b771b2b9bf705c2505aa8b8684eb302c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 92, + 844, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8e91f56e3c7b9b817435a0c052e9a751ce44fb42e8826395cbe8dda7b5db68a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 196, + 264, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa2151fce9011d3a1ee0802be7a594fc55d8654e144fa0861e7d1213147c3688.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 196, + 408, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20db97aad59377cb1f63ca37a2fdc03c755daf3a6af2006e0c88a40878e21656.jpg", + "image_caption": [ + "GANMcC" + ], + "image_footnote": [], + "bbox": [ + 413, + 196, + 553, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e1cb7d4873a1eb56a6e1195c24bcd75a85a7eba53fa26479e755d62dc7b202e.jpg", + "image_caption": [ + "MFEIF" + ], + "image_footnote": [], + "bbox": [ + 557, + 196, + 699, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03ba215c1803ec5764514a99aa83c4471b5ca4e85616aeab5171f5cf599a4f0c.jpg", + "image_caption": [ + "U2Fusion" + ], + "image_footnote": [], + "bbox": [ + 702, + 196, + 844, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb4d87835cb6a47d9194550c743d2137e6724038f9993d1310cc9a2a162b2fe4.jpg", + "image_caption": [ + "YDTR" + ], + "image_footnote": [], + "bbox": [ + 122, + 301, + 267, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5656eb01f22b5676b841337b6986b965402f9f920406bf6f622c74d46f004dd3.jpg", + "image_caption": [ + "PIAFusion", + "FusionGAN" + ], + "image_footnote": [], + "bbox": [ + 267, + 301, + 410, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15b224bf6ae870c6265ec3f4032a5103d04a7cd5a83a5673af802f6bce405637.jpg", + "image_caption": [ + "SwinFusion" + ], + "image_footnote": [], + "bbox": [ + 413, + 301, + 553, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fbb470afbc421d16b4a4621902e8d60da426399a06bdc68223be9a3c6acd569.jpg", + "image_caption": [ + "Tardal", + "MFEIF" + ], + "image_footnote": [], + "bbox": [ + 557, + 301, + 700, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/515fcbf6d7e2e13c469ab0e3c6b4617ffea677ca8b2a02d3b62959ea945f9458.jpg", + "image_caption": [ + "MetaFusion" + ], + "image_footnote": [], + "bbox": [ + 702, + 301, + 844, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d36d1aae1d2ef1013486fc2f1c9d43f9d4ae16d6356d7f4fa2b7f52773e420fa.jpg", + "image_caption": [ + "Infrared-visible image", + "YDTR" + ], + "image_footnote": [], + "bbox": [ + 122, + 412, + 264, + 503 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ad257a316b584678fb08417affd4db817082088d0b0249cf87cccd64845a3ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 412, + 410, + 502 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8285b498c10e22dfba2c6eedc0e914b8620db201cca92225e008313cbb3740e8.jpg", + "image_caption": [ + "GANMcC" + ], + "image_footnote": [], + "bbox": [ + 413, + 412, + 553, + 502 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/658322fc0e3150ddd744a4e7a8209033695ffad91c723aacab6174198aafcadb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 412, + 700, + 502 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4745ce2a8ff41e1922328a7877ce1df3248ff8f67452eb1dcd6bcdb29bbbbc4d.jpg", + "image_caption": [ + "U2Fusion" + ], + "image_footnote": [], + "bbox": [ + 702, + 412, + 844, + 502 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d76b022c349acfc86c866944a2e04bafd3fb7c3e23db1ea35750c871a1a3ee53.jpg", + "image_caption": [ + "Infrared-visible image" + ], + "image_footnote": [], + "bbox": [ + 122, + 516, + 264, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/995e45e504a43da9738de81ddd817dbb0d2a9253221afe2754e25e0d31ee27b0.jpg", + "image_caption": [ + "PIAFusion", + "FusionGAN" + ], + "image_footnote": [], + "bbox": [ + 267, + 516, + 410, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c564e8d7527f75731c5f671c6be12c2194ab566ce1646802ac0c70841819cae5.jpg", + "image_caption": [ + "SwinFusion", + "GANMcC" + ], + "image_footnote": [], + "bbox": [ + 413, + 516, + 553, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72cc7c78d62de00a53c2e4651e1ca4f4285985d1f1739a464ec971318f4033d5.jpg", + "image_caption": [ + "Tardal", + "MFEIF" + ], + "image_footnote": [], + "bbox": [ + 557, + 516, + 700, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/341595a9f2a5131a9d527e2c5eede76f72a6fa7647935fa5594162b178e2c52c.jpg", + "image_caption": [ + "MetaFusion", + "U2Fusion" + ], + "image_footnote": [], + "bbox": [ + 702, + 516, + 844, + 613 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6cb12c5541ad10eafda0fc1f0cff1d005ec662d214ed311e581e5489272e1dd.jpg", + "image_caption": [ + "YDTR", + "Figure 6. Qualitative results of different fusion methods on $\\mathbf{M}^3\\mathrm{FD}$ (first group), RoadScene (second group) and TNO (third group)." + ], + "image_footnote": [], + "bbox": [ + 124, + 628, + 264, + 720 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/924e4f0a226b7a4935fb85fedb606e19bb232cd9ee7c74cb84ebf060146d90be.jpg", + "image_caption": [ + "PIAFusion" + ], + "image_footnote": [], + "bbox": [ + 267, + 628, + 410, + 720 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/691b89cf755bf1359f60d8cefd5c3d54770854d45b616b9767bb33f07867267e.jpg", + "image_caption": [ + "SwinFusion" + ], + "image_footnote": [], + "bbox": [ + 413, + 628, + 553, + 720 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6bc5db959b22a9872bcb6f684e4368999ba3a587ec0f7313aeeb397b62f5964b.jpg", + "image_caption": [ + "Tardal" + ], + "image_footnote": [], + "bbox": [ + 557, + 628, + 699, + 720 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/685a7fef49377996e8219b74c7cfedb1be499bbbcb1a73d9b491d15de8cc7c2b.jpg", + "image_caption": [ + "MetaFusion" + ], + "image_footnote": [], + "bbox": [ + 702, + 628, + 844, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "adopted to generate fusion results for the fair comparison.", + "bbox": [ + 76, + 770, + 459, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results of different fusion methods are shown in Figure 6. All the fusion methods can fuse the main features of the infrared and visible images to some extent. However, FusionGAN, GANMcC, MFEIF and YDTR produce blurred edge details, and U2Fusion, PIAFusion, Swinfusion and Tardal generate low contrast objects, as shown in the red rectangular boxes. In contrast, the images gener", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ated by the proposed MetaFusion contain clear edge details and high contrast objects.", + "bbox": [ + 498, + 770, + 890, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sequentially, we provide quantitative results of different fusion methods in Table 4. Our MetaFusion generally achieves the largest or the second-largest metric values. In detail, the high EN and MI illustrate the fusion images by our MetaFusion contain high contrast object and clear edge details. The large VIF shows our fusion results have high-", + "bbox": [ + 496, + 809, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13961", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e36083457f362deb290572f174b9eae2df103aaf34e80f70ce6cc64bfdfc7f6c.jpg", + "table_caption": [ + "Table 4. Quantitative results of different fusion methods on TNO, RoadScene and $\\mathbf{M}^3\\mathrm{FD}$ datasets. The model inference time is counted on a NVIDIA GeForce RTX 2080 Ti. The best result is in red and the second best one is in violet." + ], + "table_footnote": [], + "table_body": "
MethodM3FDRoadSceneTNO40Time (s)
MIEntVIFMIEntVIFMIEntVIF
FusionGAN [25]13.4456.7220.30314.2037.1010.25113.0686.5340.2520.040
GANMcC [26]13.7316.8650.45314.0177.0080.42213.4856.7420.4240.081
MFEIF [21]12.9576.4780.40113.4896.7420.26013.3606.6800.3950.029
U2Fusion [44]13.8166.9080.54513.8886.9440.45613.8896.9440.6360.043
YDTR [36]12.6946.3470.36113.1666.5830.23612.8626.4310.2800.086
PIAFusion [35]13.3266.6630.43713.2626.5330.20513.9316.9600.5060.055
SwinFusion [23]13.0516.5250.46013.2676.6330.31313.8796.9330.4791.081
Tardal [20]13.6366.8180.65015.0097.5040.48313.0306.5150.9100.030
MetaFusion14.5117.2491.51514.2187.0220.96914.6577.3231.4620.015
", + "bbox": [ + 93, + 125, + 874, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdcd42e5b93865b1136dd887f9ab32722b6d75e9c7be68926411453fa17f5d3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 354, + 186, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c22264dbdd5ab373084f0a61ed0876b143d758176fdc5aad3b9e879640f97567.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 187, + 354, + 281, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb61026d2b5b36f7323ce85006915eae8ed2e73cea3ebe1e1e5353ddc7c62074.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 354, + 375, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b1f7750091301e96c5ff96b59ba8243280f6a92a8868adf0783f3bbe6057ca1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 354, + 473, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7dbcd590492a89d29c05ed3e2ae000c8d379d9e62ccaafd73fa44c3eed1c48c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 410, + 186, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa46622c4c836260e003e102d1cbe4cb775ce2450a09639e2d66e7107a073779.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 187, + 410, + 281, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/feccfbfe8c5ddbf158ec44996b51a076155566fae41d1f2f167bf02f8769a664.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 410, + 375, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a70e72e5c666c9755bb36dd2b788b5c04598f690e5b5c79a850ab0eedb68e27c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 410, + 473, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9454166dc21ce7a57c98c07515a73a40df63984fc7542c60c469158e2dc28f3.jpg", + "image_caption": [ + "SwinFusion" + ], + "image_footnote": [], + "bbox": [ + 88, + 474, + 186, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d5c4172d94e9654594d1425b91a4acddc3c2237a424f62aabcf6eb63e01bce3.jpg", + "image_caption": [ + "Tardal" + ], + "image_footnote": [], + "bbox": [ + 187, + 469, + 281, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a14122c348be79ee8a9a3f7a31e946ec3789bb650e339f0ec8cff17e3f314c82.jpg", + "image_caption": [ + "MetaFusion", + "Figure 7. Visual results of object detection based on different fusion methods on $\\mathbf{M}^3\\mathrm{FD}$" + ], + "image_footnote": [], + "bbox": [ + 282, + 474, + 375, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a31329ecfc5325b98c81c63f5f364718a68cb1bbffd94430a92dc9cc429a3fbe.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 377, + 474, + 473, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "quality visual effect and small distortion compared with the source images. Moreover, our method achieves the fastest inference time that only needs 0.015s to generate one fusion result.", + "bbox": [ + 75, + 580, + 468, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Evaluation on Infrared-visible Object Detection", + "text_level": 1, + "bbox": [ + 76, + 650, + 468, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A better fusion image can provide more information to improve object detection. Here, we verify the effectiveness of our method by evaluating object detection accuracy with the infrared and visible fusion images. To make a fair comparison, we firstly generate fusion results using our method and SOTA methods, and then use the results to retrain the object detection baseline YOLOv5s, respectively.", + "bbox": [ + 75, + 689, + 468, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 shows the object detection results. In general, the SOTA methods improve the performance of object detection. In contrast, our MetaFusion achieves a better performance, e.g., the person is accurately detected (see the blue ellipse). Table 5 shows the quantitative results. Our fusion results help the detection network achieve the highest object detection accuracy. This further proves that our", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/083a47be8efac396a40d841af152634be033863291b3d8932a9c5fd9ba49f0d9.jpg", + "table_caption": [ + "Table 5. Quantitative results of object detection based on different fusion methods on $\\mathbf{M}^3\\mathrm{FD}$ dataset. The best result is in red and the second best one is in violet." + ], + "table_footnote": [], + "table_body": "
Method\\( \\mathrm{mAP}_{50\\rightarrow 95}(\\%) \\)
FusionGAN [25]54.2
GANMcC [26]55.2
MFEIF [21]55.4
U2Fusion [44]55.7
YDTR [36]55.4
PIAFusion [35]55.6
SwinFusion [23]55.4
Tardal [20]54.4
MetaFusion56.5
", + "bbox": [ + 553, + 401, + 836, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method can generate high-quality fusion results especially for the objects.", + "bbox": [ + 498, + 612, + 890, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 660, + 619, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a joint fusion and detection learning framework through introducing the meta-feature embedding model. Based on the meta learning idea, the meta-feature embedding model can generate object semantic features according to the fusion network ability, thereby bridging the feature gap between these two different-level tasks. Moreover, a mutual promotion learning between fusion and detection tasks is further implemented to improve their performances. Both quantitative and qualitative results demonstrate the superior performance of our method compared with the state-of-the-art methods.", + "bbox": [ + 496, + 686, + 890, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by National Natural Science Foundation of China under Grant Nos. 62176038, 62001450 and U1903215.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13962", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9465-9474, October 2021. 3", + "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6154-6162, 2018. 2", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. 2", + "[4] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-sclil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14166-14175, June 2022. 3", + "[5] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 3", + "[6] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster R-CNN: towards accurate few-shot object detection with attentive feature alignment. In AAAI, pages 780-789. AAAI Press, 2022. 3", + "[7] Yu Han, Yunze Cai, Yin Cao, and Xiaoming Xu. A new image fusion performance metric based on visual information fidelity. Information Fusion, 14:-, 04 2013. 5", + "[8] Kaiming He, Georgia Gkioxari, Piotr Dollr, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 2", + "[9] Liqiang He and Sinisa Todorovic. Destr: Object detection with split transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9367-9376, 2022. 6", + "[10] Ruichao Hou, Dongming Zhou, Rencan Nie, Dong Liu, Lei Xiong, Yanbu Guo, and Chuanbo Yu. Vif-net: An unsupervised framework for infrared and visible image fusion. IEEE Transactions on Computational Imaging, 6:640-651, 2020. 2", + "[11] Muhammad Abdullah Jamal, Liqiang Wang, and Boqing Gong. A lazy approach to long-horizon gradient-based meta-learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6577-6586, October 2021. 3", + "[12] Huafeng Li, Yueliang Cen, Yu Liu, Xun Chen, and Zhengtao Yu. Different input resolutions and arbitrary output resolution: A meta learning-based deep framework for infrared and visible image fusion. IEEE Transactions on Image Processing, 30:4070-4083, 2021. 3", + "[13] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Hui Li, Xiao-Jun Wu, and Josef Kittler. Mdlatrr: A novel decomposition method for infrared and visible image fusion. IEEE Transactions on Image Processing, 29:4733-4746, 2020. 2", + "[15] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 1", + "[16] Jing Li, Jianming Zhu, Chang Li, Xun Chen, and Bin Yang. Cgtf: Convolution-guided transformer for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 71:1-14, 2022. 2", + "[17] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 3", + "[18] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP:1-1, 07 2018. 2", + "[19] Benlin Liu, Yongming Rao, Jiwen Lu, Jie Zhou, and Chojui Hsieh. Metadistiller: Network self-boosting via metalearned top-down distillation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16, pages 694–709. Springer, 2020. 3", + "[20] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 1, 2, 5, 6, 8", + "[21] J. Liu, X. Fan, J. Jiang, R. Liu, and Z. Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1, 2021. 6, 8", + "[22] Yu Liu, Shuping Liu, and Zengfu Wang. A general framework for image fusion based on multi-scale transform and s-parse representation. Information Fusion, 24:147-164, 2015. 2", + "[23] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 2, 6, 8", + "[24] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping Zhang. Ddcgan: A dual-discriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29:4980-4995, 2020. 2", + "[25] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusion: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11-26, 08 2019. 1, 6, 8", + "[26] Jiayi Ma, Hao Zhang, Zhenfeng Shao, Pengwei Liang, and Han Xu. Ganmcc: A generative adversarial network with" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13963", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multiclassification constraints for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1-14, 2021. 2, 6, 8", + "[27] Yan Mo, Xudong Kang, Puhong Duan, Bin Sun, and Shutao Li. Attribute filter based infrared and visible image fusion. Information Fusion, 75:41-54, 2021. 2", + "[28] Guihong Qu, Dali Zhang, and Pingfan Yan. Information measure for performance of image fusion. *Electronics Letters*, 38:313-315, 04 2002. 5", + "[29] Aniruddh Raghu, Jonathan Lorraine, Simon Kornblith, Matthew McDermott, and David K Duvenaud. Meta-learning to improve pre-training. Advances in Neural Information Processing Systems, 34:23231-23244, 2021. 3", + "[30] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 2", + "[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 2", + "[32] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 658-666, 2019. 2", + "[33] Wesley Roberts, Jan van Aardt, and Fethi Ahmed. Assessment of image fusion procedures using entropy, image quality, and multispectral classification. Journal of Applied Remote Sensing, 2:1-28, 05 2008. 5", + "[34] Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshitaka Ushiku, and Tatsuya Harada. Multispectral object detection for autonomous vehicles. In Proceedings of the on Thematic Workshops of ACM Multimedia 2017, Thematic Workshops '17, page 3543, New York, NY, USA, 2017. Association for Computing Machinery. 1", + "[35] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83-84:79-92, 2022. 1, 2, 6, 8", + "[36] Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer. IEEE Transactions on Multimedia, pages 1-16, 2022. 2, 6, 8", + "[37] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249-251, 2017. 5", + "[38] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 4", + "[39] Zhishe Wang, Yuanyuan Wu, Junyao Wang, Jiawei Xu, and Wenyu Shao. Res2fusion: Infrared and visible image fusion based on dense res2net and double nonlocal attention models. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 2", + "[40] Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. In" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14248-14257, June 2022. 2", + "[41] Chengming Xu, Yanwei Fu, Chen Liu, Chengjie Wang, Jilin Li, Feiyue Huang, Li Zhang, and Xiangyang Xue. Learning dynamic alignment via meta-filter for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5182-5191, June 2021. 3", + "[42] Dongli Xu, Jinghong Deng, and Wen Li. Revisiting ap loss for dense object detection: Adaptive ranking pair selection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 14187-14196, 2022. 2", + "[43] Han Xu, Meiqi Gong, Xin Tian, Jun Huang, and Jiayi Ma. Cufd: An encoderdecoder network for visible and infrared image fusion based on common and unique feature decomposition. Computer Vision and Image Understanding, 218:103407, 03 2022. 1", + "[44] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 1, 2, 5, 6, 8", + "[45] Han Xu, Xinya Wang, and Jiayi Ma. Drf: Disentangled representation for visible and infrared image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1-13, 2021. 1, 2", + "[46] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Rui-Wei Zhao, Tao Zhang, Xuequan Lu, and Shang Gao. Cream: Weakly supervised object localization via class re-activation mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9437–9446, June 2022. 2", + "[47] Meilong Xu, Linfeng Tang, Hao Zhang, and Jiayi Ma. Infrared and visible image fusion via parallel scene and texture learning. Pattern Recognition, 132:108929, 2022. 1", + "[48] Shilong Zhang, Zhuoran Yu, Liyang Liu, Xinjiang Wang, Aojun Zhou, and Kai Chen. Group r-cnn for weakly semi-supervised object detection with points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9417-9426, June 2022. 2", + "[49] Fan Zhao and Wenda Zhao. Learning specific and general realm feature representations for image fusion. IEEE Transactions on Multimedia, 23:2745-2756, 2020. 2", + "[50] Fan Zhao, Wenda Zhao, and Hutchuan Lu. Interactive feature embedding for infrared and visible image fusion. arXiv preprint arXiv:2211.04877, 2022. 2, 4", + "[51] Fan Zhao, Wenda Zhao, Libo Yao, and Yu Liu. Self-supervised feature adaption for infrared and visible image fusion. Information Fusion, 76:189-203, 2021. 2", + "[52] Wenda Zhao, Xueqing Hou, You He, and Huchuan Lu. Defocus blur detection via boosting diversity of deep ensemble networks. IEEE Transactions on Image Processing, 30:5426-5438, 2021. 2", + "[53] Wenda Zhao, Huimin Lu, and Dong Wang. Multisensor image fusion and enhancement in spectral total variation domain. IEEE Transactions on Multimedia, 20(4):866-879, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13964", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Wenda Zhao, Dong Wang, and Huchuan Lu. Multi-focus image fusion with a natural enhancement via a joint multi-level deeply supervised convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 29(4):1102-1115, 2018. 2", + "[55] Wenda Zhao, Bowen Zheng, Qiuhua Lin, and Hutchuan Lu. Enhancing diversity of defocus blur detectors via cross-ensemble network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8905–8913, 2019. 2", + "[56] Zhengjie Zhu, Xiaogang Yang, Ruitao Lu, Tong Shen, Xueli Xie, and Tao Zhang. Clf-net: Contrastive learning for infrared and visible image fusion network. IEEE Transactions on Instrumentation and Measurement, 71:1-15, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13965", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_model.json b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f200e07e5898b6034273b842c656a769ec13f0f --- /dev/null +++ b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_model.json @@ -0,0 +1,3049 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.88, + 0.178 + ], + "angle": 0, + "content": "MetaFusion: Infrared and Visible Image Fusion via Meta-Feature Embedding from Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.203, + 0.747, + 0.221 + ], + "angle": 0, + "content": "Wenda Zhao\\(^{1}\\), Shigeng Xie\\(^{1}\\), Fan Zhao\\(^{2*}\\), You He\\(^{3}\\), Huchuan Lu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.221, + 0.65, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\) Dalian University of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.239, + 0.632, + 0.256 + ], + "angle": 0, + "content": "\\(^{2}\\) Liaoning Normal University, China" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.257, + 0.601, + 0.275 + ], + "angle": 0, + "content": "3 Tsinghua University, China" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.278, + 0.707, + 0.292 + ], + "angle": 0, + "content": "zhaowenda@dlut.edu.cn; xieshigeng@mail.dlut.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.296, + 0.736, + 0.309 + ], + "angle": 0, + "content": "Fan-Zhao20@163.com; youhe_nau@163.com; lhchuan@dlut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.377, + 0.474, + 0.665 + ], + "angle": 0, + "content": "Fusing infrared and visible images can provide more texture details for subsequent object detection task. Conversely, detection task furnishes object semantic information to improve the infrared and visible image fusion. Thus, a joint fusion and detection learning to use their mutual promotion is attracting more attention. However, the feature gap between these two different-level tasks hinders the progress. Addressing this issue, this paper proposes an infrared and visible image fusion via meta-feature embedding from object detection. The core idea is that meta-feature embedding model is designed to generate object semantic features according to fusion network ability, and thus the semantic features are naturally compatible with fusion features. It is optimized by simulating a meta learning. Moreover, we further implement a mutual promotion learning between fusion and detection tasks to improve their performances. Comprehensive experiments on three public datasets demonstrate the effectiveness of our method. Code and model are available at: https://github.com/wdzhao123/MetaFusion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.21, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Multi-modality sensor technology has promoted the application of multi-modality images in different areas. Among them, infrared images and visible images have been utilized commonly, as the information contained in these two modalities is complementary. Specifically, infrared images can supply object thermal structures without being affected by illumination. But they are short of texture details. On the contrary, visible images can catch the texture information for the scene. But they are severely affected by light. Thus, many methods [15,25,35,43-45,47] focus on s" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.345, + 0.887, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.418, + 0.802, + 0.43 + ], + "angle": 0, + "content": "(a) Separate optimization method" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.435, + 0.888, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.504, + 0.808, + 0.515 + ], + "angle": 0, + "content": "(b) Cascaded optimization method" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.519, + 0.886, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.628, + 0.807, + 0.64 + ], + "angle": 0, + "content": "(c) Meta-feature embedding method" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.649, + 0.894, + 0.816 + ], + "angle": 0, + "content": "Figure 1. Different joint learning methods of infrared and visible image fusion (IVIF) and object detection (OD). (a) Separate optimization method: IVIF network \\(\\Psi\\) is firstly optimized by fusion loss \\(L_{f}\\). Then, fusion result \\(z\\) is generated by \\(\\Psi\\) from the input infrared and visible image pair \\(x,y\\). Finally, OD network \\(\\Phi\\) is optimized by detection loss \\(L_{d}\\) using \\(z\\). (b) Cascaded optimization method: OD network \\(\\Phi\\) is treated as a constraint to optimize IVIF network \\(\\Psi\\) by the loss \\(L_{f}\\) and \\(L_{d}\\). (c) Meta-feature embedding method: Meta-feature embedding \\(\\Gamma\\) is optimized to learn how to guide \\(\\Psi\\) to have a low \\(L_{f}\\). Then, \\(\\Gamma\\) generates meta feature from detection feature \\(F_{i}^{d}\\) and fusion feature \\(F_{i}^{d}\\). Finally, the meta feature is used to guide \\(\\Psi\\) by the loss \\(L_{d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.836, + 0.893, + 0.882 + ], + "angle": 0, + "content": "tudying pixel-level infrared and visible image fusion (IVIF), thereby helping high-level tasks improve performance, e.g., object detection (OD) [20, 34]." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.893, + 0.901 + ], + "angle": 0, + "content": "IVIF and OD can greatly benefit form each other. IVIF" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13955" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.213 + ], + "angle": 0, + "content": "generates the fused image that contains more information than any single modality image to improve OD. However, IVIF mainly focuses on the pixel relationship between an image pair, and there is little consideration for object semantic. In contrast, OD can provide rich object semantic information to IVIF, as its aim is locating the objects. Therefore, this paper studies a joint learning framework between IVIF and OD to improve their performances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.471, + 0.565 + ], + "angle": 0, + "content": "Existing joint learning methods of IVIF and OD can be divided into two categories: separate optimization and cascaded optimization. Separate optimization firstly trains IVIF network, and then trains OD network using IVIF results, as shown in Figure 1(a). Thus, most methods focus on improving the fusion effect, e.g., designing networks [13, 23, 35, 36] and introducing constraints [10, 26, 45]. Obviously, separate optimization neglects the help of OD. Cascaded optimization adopts OD network as a constraint to train IVIF network, and thus forces the IVIF network to generate fusion images with easily detected objects [20], as shown in Figure 1(b). However, directly utilizing the high-level OD constraint to guide the pixel-level IVIF will result in limited effect. Therefore, we leverage OD feature maps that guide IVIF feature maps to obtain more semantic information. Unfortunately, OD features are mismatched with IVIF features due to their task-level difference. Addressing this issue, we propose a meta-feature embedding network (MFE), as shown in Figure 1(c). The idea is that if MFE generates OD features according to the IVIF network ability, the OD features are naturally compatible with the IVIF network, and the optimization can be achieved by simulating a meta learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Specifically, an infrared and visible image fusion via meta-feature embedding from object detection is proposed, which is named as MetaFusion. MetaFusion includes IVIF network \\((F)\\), OD network \\((D)\\), and \\(MFE\\). In particular, \\(MFE\\) is expected to generate meta features to bridge the gap between \\(F\\) and \\(D\\), which is optimized by two alternate steps: inner update and outer update. In the inner update process, we firstly optimize \\(F\\) using meta training set \\(S_{mtr}\\) to obtain its updated network \\(F'\\). Then, \\(F'\\) calculates the fusion loss on meta testing set \\(S_{mts}\\) to optimize \\(MFE\\). The motivation is that if \\(MFE\\) successfully generates meta features which are compatible with \\(F\\), \\(F'\\) will produce better fused images, i.e., the fusion loss should be lower. In the outer update process, \\(F\\) is optimized with the guide of the meta features generated by the fixed \\(MFE\\) on \\(S_{mtr}\\) and \\(S_{mts}\\). In this way, \\(F\\) can learn how to extract semantic information to improve fusion quality. In the above two alternate steps, \\(D\\) is fixed to offer detection semantic information. Thus, we further implement a mutual promotion learning, where we use the improved \\(F\\) to generate fusion results to finetune \\(D\\), and then the improved \\(D\\) offers better semantic information to optimize \\(F\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "In summary, our contributions are as follows. (1) We explore the joint learning framework of IVIF and OD, and propose MetaFusion to obtain superior performance on these two tasks. (2) Meta-feature embedding network is designed to generate meta features that bridge the gap between \\( F \\) and \\( D \\). (3) Sequentially, mutual promotion learning between \\( F \\) and \\( D \\) is introduced to improve their performances. (4) Extensive experiments on image fusion and object detection validate the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.241, + 0.642, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Image Fusion. Traditional image fusion methods use hand-crafted features, such as sparse representation [22], spectral variation [53] and low-rank representation [14], which can not handle the complex scenes well. Nowadays, deep learning-based image fusion methods [16, 24, 27, 39, 49, 50, 54, 56] are proposed. Ma et al. [23] propose a swim transformer based image fusion method with cross-domain long-range learning. Xu et al. [44] use feature extraction and measurement to estimate the degree of information preservation in image fusion. Zhao et al. [51] design a self-supervised feature extraction model. Besides, feature diversity enhancement and fusion is received attention [52, 55]. However, most of them neglect the help of high-level tasks. Recently, Liu et al. [20] propose a joint learning of IVIF and OD. They treat OD network as an additional constraint to help IVIF network generate fusion result with more clear objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.674 + ], + "angle": 0, + "content": "However, directly using the high-level OD network to guide pixel-level IVIF may not obtain an effective constraint. Moreover, they use the fusion results as the connection between IVIF network and OD network, ignoring the pixel-level semantic information in OD features. In contrast, we adopt semantic information in OD features to help IVIF, and conduct a meta-feature embedding to generate meta features from OD features. Then, those meta features are used to guide IVIF network to learn pixel-level semantic information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Object Detection. Deep learning-based object detection methods have made great progress, e.g., network structure-designed methods [2, 8, 30, 31] and loss constraint methods [18, 32, 42]. Carion et al. [3] introduce a transformer encoder-decoder architecture for object detection. Xu et al. [42] rank positive and negative sample pairs and improve the ranking effect by a clustering algorithm to built object detection loss. On the other hand, many weakly supervised methods [40, 46, 48] are proposed. Wu et al. [40] design a model to suppress the background activation and obtain object region, where image-level labels are used. Zhang et al. [48] use instance proposal grouping to learn object detection from a single point annotation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Most of those object detection methods leverage single modality image, e.g., visible images. However, the visible" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.87, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.896, + 0.399 + ], + "angle": 0, + "content": "Figure 2. Framework illustration of the proposed MetaFusion. MetaFusion includes three parts: infrared and visible image fusion (IVIF), object detection (OD), and meta-feature embedding (MFE). IFIV network includes four blocks: three feature fusion blocks \\( FFB_{1} \\), \\( FFB_{2} \\), \\( FFB_{3} \\) used to extract and fuse features and one image reconstruction module \\( IRM \\) used to reconstruct fusion result. MFE contains three meta-feature generators \\( MFG_{1} \\), \\( MFG_{2} \\), \\( MFG_{3} \\) conducted to generate multi-level meta features from OD features \\( F_{ej} \\), and three feature transform networks \\( FT_{1} \\), \\( FT_{2} \\), \\( FT_{3} \\) used to transfer the meta features to fusion features. OD network is divided as three feature extraction blocks \\( DFB_{1} \\), \\( DFB_{2} \\), \\( DFB_{3} \\) and one OD detection head \\( DH \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.496 + ], + "angle": 0, + "content": "images are affected by light conditions, thereby influencing the performance of object detection. On the contrary, we leverage IVIF results to provide infrared information without being influenced by illumination, therefore can offer more stable detection performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Meta Learning. Meta learning aims to use the learned knowledge to quickly adapt to new tasks, which has been used to many fields: model pre-training [5, 11, 29], image classification [17, 19], few-shot learning [1, 4, 6, 41], etc. For example, Raghu et al. [29] propose an implicit differentiation and backpropagation meta-learning method and improve the efficiency of learning pre-training hyperparameters. Xu et al. [41] design a dynamic alignment in meta learning, which can emphasize information in query according to the support. Li et al. [12] use the meta learning to handle different input resolutions, thereby generating fusion results with arbitrary resolutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Inspired by the above meta learning methods, we design a meta-feature embedding network to generate the meta features from object detection features, thereby helping IVIF network fuse more object semantic information." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.754, + 0.251, + 0.772 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our MetaFusion framework is shown in Figure 2, which includes three subnetworks: IVIF network \\((F)\\) generates fusion results, OD network \\((D)\\) offers semantic features, and meta-feature embedding network \\((MFE)\\) uses OD features to guide IVIF to fuse more object semantic information. The challenge is that OD features can hardly be directly used to guide IVIF due to their task gap. Thus, our focus is to design \\(MFE\\) and optimize it. Details are as follows." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.421, + 0.898, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.581, + 0.894, + 0.68 + ], + "angle": 0, + "content": "Figure 3. Illustration of the meta-feature embedding learning. There are two stages in the meta-feature embedding: (a) inner update and (b) outer update. In the inner update, \\( MFG \\) and \\( F^T \\) are trained according to current \\( F' \\) semantic extracting ability to generate meta features. In the outer update, \\( F \\) is trained which is guided by the meta features, thereby learning to extract object semantic. The two stages alternate every N epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.733, + 0.71 + ], + "angle": 0, + "content": "3.1. Meta-feature Embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.854 + ], + "angle": 0, + "content": "Meta-feature embedding includes meta-feature generator (MFG) and feature transform (FT). MFG generates meta feature \\( F_{mj} \\) according to IVIF feature \\( F_{uj} \\) from OD feature \\( F_{ej} \\), i.e., \\( F_{mj} = MFG(F_{uj}, F_{ej}) \\), where \\( j \\) is the feature level index. FT transfers the meta feature \\( F_{mj} \\) to fusion feature \\( F_{uj} \\) by producing feature bridge \\( F_{tj} \\). Specifically, meta-feature embedding learning is divided two stages: inner update stage and outer update stage, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In the inner update stage, we optimize \\(MFG\\), \\(FT\\) and \\(F\\) (see Figure 3(a)). Given meta training set \\(S_{mtr}\\) and meta testing set \\(S_{mts}\\), we firstly update the parameter \\(\\theta_F\\) of \\(F\\) on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13957" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.325, + 0.107 + ], + "angle": 0, + "content": "\\(S_{mtr}\\) through meta feature guidance:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.117, + 0.47, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\theta_ {F ^ {\\prime}} = \\theta_ {F} - \\beta_ {F ^ {\\prime}} \\nabla_ {\\theta_ {F}} L _ {g} \\left(F _ {m j}, F _ {t j}\\right) = \\theta_ {F} - \\beta_ {F ^ {\\prime}} \\frac {\\partial L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.162, + 0.473, + 0.284 + ], + "angle": 0, + "content": "where the guide loss \\(L_{g}\\) is \\(L_{2}\\) distance, \\(\\beta_{F^{\\prime}}\\) is the learning rate of \\(F^{\\prime}\\), \\(\\theta_{F^{\\prime}}\\) is the parameter of \\(F^{\\prime}\\) that is updated from \\(F\\). Then, we use \\(F^{\\prime}\\) with parameter \\(\\theta_{F^{\\prime}}\\) to calculate the fusion loss \\(L_{f}\\) on meta testing set \\(S_{mts}\\). Here, \\(L_{f}\\) can measure the effect of using the meta features to guide \\(F\\), i.e., when \\(L_{f}\\) is lower, the semantic extracting ability of \\(F\\) will be improved. Therefore, we use \\(L_{f}\\) to update parameter \\(\\theta_{MFG}\\) of \\(MFG\\) and parameter \\(\\theta_{FT}\\) of \\(FT\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.293, + 0.47, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\theta_ {M F G} = \\theta_ {M F G} - \\beta_ {M F G} \\nabla_ {\\theta_ {M F G}} L _ {f} \\left(I _ {f} ^ {m t s}, I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), (2)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.322, + 0.469, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\theta_ {F T} = \\theta_ {F T} - \\beta_ {F T} \\nabla_ {\\theta_ {F T}} L _ {f} \\left(I _ {f} ^ {m t s}, I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.347, + 0.47, + 0.424 + ], + "angle": 0, + "content": "where the fusion loss \\(L_{f}\\) is SSIM loss [38], \\(I_{i}^{mts}\\) and \\(I_{v}^{mts}\\) are infrared and visible images from \\(S_{mts}\\), \\(I_{f}^{mts} = F'(I_{i}^{mts}, I_{v}^{mts})\\) is the meta testing fusion result, and \\(\\beta_{MFG}\\) and \\(\\beta_{FT}\\) are the learning rate of MFG and \\(FT\\). \\(\\nabla_{\\theta_{MFG}} L_{f}\\) can be calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.434, + 0.469, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta_ {M F G}} L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right) \\\\ = \\frac {\\partial L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right) , I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right)}{\\partial \\theta_ {F ^ {\\prime}}} * \\left(- \\frac {\\partial^ {2} L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F} \\partial \\theta_ {M F G}}\\right). \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.5, + 0.343, + 0.516 + ], + "angle": 0, + "content": "Similarly, \\(\\nabla_{\\theta_{FT}}L_f\\) can be calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.526, + 0.469, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta_ {F T}} L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right) \\\\ = \\frac {\\partial L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right) , I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right)}{\\partial \\theta_ {F ^ {\\prime}}} * \\left(- \\frac {\\partial^ {2} L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F} \\partial \\theta_ {F T}}\\right). \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.592, + 0.469, + 0.636 + ], + "angle": 0, + "content": "Thus, \\(MFG\\) can learn how to generate meta features according to current \\(F^{\\prime}\\) semantic extracting ability, i.e., makes meta features be compatible with \\(F\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.637, + 0.469, + 0.699 + ], + "angle": 0, + "content": "In the outer update stage, \\( F \\) is trained with the initial parameter \\( \\theta_F \\), as shown in Figure 3(b). Given the training data \\( S = \\{S_{mtr}, S_{mts}\\} \\), \\( F \\) is optimized using the fusion loss \\( L_f \\) and guide loss \\( L_g \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.709, + 0.472, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta_ {F} = \\theta_ {F} - \\beta_ {F} \\nabla_ {\\theta_ {F}} \\left(L _ {f} \\left(I _ {f}, I _ {i}, I _ {v}\\right) + \\lambda_ {g} \\sum_ {j = 1} ^ {3} L _ {g} \\left(F _ {m j}, F _ {t j}\\right)\\right) \\\\ = \\theta_ {F} - \\beta_ {F} \\left(\\frac {\\partial L _ {f} \\left(I _ {f} , I _ {i} , I _ {v}\\right)}{\\partial \\theta_ {F}} + \\lambda_ {g} \\sum_ {j = 1} ^ {3} \\frac {\\partial L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F}}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\(I_{i}\\) and \\(I_{v}\\) are infrared and visible images from \\(S\\), \\(\\lambda_{g}\\) is a hyper-parameter to balance \\(L_{f}\\) and \\(L_{g}\\), and \\(\\beta_{F}\\) is the learning rate of \\(F\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Finally, the inner update stage and outer update stage are carried out alternately for every \\(N\\) epochs, which improves the optimization of \\(MFG\\) and \\(F\\)." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.865, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.357, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Figure 4. Illustration of the mutual promotion learning between \\( F \\) and \\( D \\). Firstly, \\( F \\) is used to generate fusion results to optimize \\( D \\). Thus, a fine-tuned \\( D \\) with higher semantic extracting ability is obtained. Then, \\( D \\) is adopted to optimize \\( F \\). This process is alternated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.442, + 0.756, + 0.459 + ], + "angle": 0, + "content": "3.2. Mutual Promotion Learning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.892, + 0.665 + ], + "angle": 0, + "content": "In Sec. 3.1, we fix the object detection network \\( D \\) to offer stable object semantic features. In this way, \\( MFG \\) does not need to handle variable object semantic features, thereby reducing the difficulty of training \\( MFG \\). However, the fixed \\( D \\) will restrict the object semantic extracting ability of \\( F \\). Addressing this problem, we propose the mutual promotion learning that promotes not only \\( F \\) but also \\( D \\), as shown in Figure 4. Specifically, we firstly train \\( F \\) individually, and then use the fusion results to train \\( D \\). Sequentially, \\( D \\) offers object semantic features to the meta-feature embedding, thereby improving \\( F \\). After that, the improved fusion results generated by \\( F \\) are used to fine-tune \\( D \\). In this way, both \\( F \\) and \\( D \\) are improved." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "3.3. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.857 + ], + "angle": 0, + "content": "IVIF network \\(F\\) aims to generate fusion results according to the input infrared and visible images, which is flexible. Here, we adopt the similar structure with [50]. \\(F\\) consists of three feature fusion blocks \\((FFB_{j}, j = 1,2,3)\\) and one image reconstruction module (IRM), as shown in Figure 2. \\(FFB_{j}\\) is used to extract and fuse features from input infrared image \\(I_{i}\\) and visible image \\(I_{v}\\). It is designed based on the idea of [50] that if \\(FFB_{j}\\) can well fuse the features of \\(I_{i}\\) and \\(I_{v}\\), then the two kinds of features can be recovered from the fused feature. Thus, \\(FFB_{j}\\) is described as" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.871, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nF F B _ {j} = \\left\\{ \\begin{array}{c} C _ {2} T _ {1} \\biguplus C _ {2} T _ {1}, j = 1 \\\\ C _ {2} T _ {j - 1} \\biguplus C _ {2} T _ {j - 1}, j > 1 \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13958" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.197 + ], + "angle": 0, + "content": "where \\( j \\) is the block index, \\( C_k \\) represents the number of \"Convolution with \\( 3 \\times 3 \\) kernel +ReLU\" layers is \\( k \\), \\( F_{u(j-1)} \\) is the output feature of the \\( j-1 \\)th block, \\( T_1 = C_2(I_i, I_v) \\) and \\( T_{j-1} = C_2(F_{u(j-1)}) \\) represent feature integration operations, and \\( \\boxplus \\) means feature concatenation. IRM reconstructs the fusion result, which contains six \"Convolution with \\( 3 \\times 3 \\) kernel +ReLU\" layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.47, + 0.302 + ], + "angle": 0, + "content": "Object detection network \\(D\\) offers object semantic features. In our framework, we choose \\(\\mathrm{yolov5s}^1\\) as our \\(D\\). According to the size of the feature maps, the backbone of \\(D\\) is divided into three feature extraction blocks \\((\\{DFB_{i}, i = 1,2,3\\})\\). In addition, we represent the neck and detection head in yolov5s into one block \\((DH)\\) to simplify the expression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.469, + 0.38 + ], + "angle": 0, + "content": "Meta-feature embedding network \\( MFE \\) includes metafeature generator \\( MFG_{j} \\) and feature transform network \\( FT_{j} \\). Since OD feature \\( F_{ej} \\) generated by \\( DFB_{j} \\) offers object semantic information and \\( F_{uj} \\) provides scene details, they are fed into the \\( MFG_{j} \\). Then, \\( MFG_{j} \\) is built as" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.391, + 0.469, + 0.412 + ], + "angle": 0, + "content": "\\[\nM F G _ {j} = C _ {6} \\left(C _ {4} \\left(F _ {u j}\\right) \\biguplus C _ {2} \\left(U p \\left(F _ {e j}\\right)\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.468 + ], + "angle": 0, + "content": "where \\( Up() \\) is the feature up-sampling. \\( FT_{j} \\) transfers the meta feature to fusion feature, which includes three \"Convolution with \\( 3 \\times 3 \\) kernel +ReLU\" layers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.477, + 0.182, + 0.494 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.606 + ], + "angle": 0, + "content": "The training details of our framework is shown in Algorithm 1, which contains four steps. Generally, we firstly pretrain \\( F \\), and then use the fusion results to finetune \\( D \\). Sequentially, we train \\( FT \\) and \\( MFG \\), where \\( MFG \\) and \\( FT \\) are updated \\( n = 200 \\) times every \\( N = 8 \\) epochs. Lastly, the mutual promotion between \\( F \\) and \\( D \\) is trained for \\( R = 2 \\) rounds, thereby improving their performances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.209, + 0.637 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.644, + 0.158, + 0.66 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.469, + 0.757 + ], + "angle": 0, + "content": "Dataset. We conduct experiment on three widely-used datasets: \\(\\mathbf{M}^3\\mathrm{FD}\\) [20], RoadScene [44] and TNO [37]. \\(\\mathbf{M}^3\\mathrm{FD}\\) is divided as training set (2940 image pairs) and testing set (1260 image pairs). RoadScene with 221 image pairs and TNO with 40 image pairs are only used for testing. Besides, \\(\\mathbf{M}^3\\mathrm{FD}\\) is adopted to evaluate OD performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.469, + 0.878 + ], + "angle": 0, + "content": "Implementation. Our framework is implemented with PyTorch on a NVIDIA GeForce RTX 3090 GPU. \\(F\\) and MFE are trained using optimizer Adam with the learning rate \\(\\beta_{F},\\beta_{F^{\\prime}},\\beta_{FT}\\) and \\(\\beta_{MFG}\\) of \\(1\\times 10^{-3}\\), respectively. \\(D\\) is trained followed by yolov5s that adopts the optimizer SGD using the learning rate of \\(1\\times 10^{-2}\\) with decaying rate of 0.1 every round. We firstly train \\(F\\) for 100 epochs and \\(D\\) for 150 epochs. Then, we conduct MFE for 50 epochs. After" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.804, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 MetaFusion Training Algorithm" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.11, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Input: Training dataset \\(S = \\{S_{mtr}, S_{mts}\\}\\), IVIF model \\(F(\\theta_F)\\), meta-feature generator \\(MFG(\\theta_F)\\), feature transform network \\(FT(\\theta_{FT})\\), OD network \\(D(\\theta_D)\\) \nOutput: IVIF model \\(F(\\theta_F^*)\\) \n1: Initialize \\(F(\\theta_F)\\), \\(MFG(\\theta_F)\\), \\(FT(\\theta_{FT})\\), \\(D(\\theta_D)\\) \n2: /* Pretrain \\(F^{*}/\\) \n3: while not converged do \n4: Sample image pair \\((I_i, I_v)\\) from \\(S\\) \n5: Forward pass on \\(F\\) to generate fusion result \\(I_f\\) \n6: Optimize \\(F\\) using fusion loss \\(L_f\\) \n7: end while \n8: /* Pretrain \\(D^{*}/\\) \n9: while not converged do \n10: Sample image \\(I_f\\) from the fused images generated by \\(F\\) \n11: Forward pass on \\(D\\) to get detection result \n12: Optimize \\(D\\) using detection loss \\(L_d\\) \n13: end while \n14: /* Meta-feature embedding outer update */ \n15: while not converged do \n16: Sample a batch image pair \\((I_i, I_v)\\) from \\(S\\) \n17: Forward pass on \\(F\\) to generate fusion result \\(I_f\\) \n18: Forward pass on \\(MFG\\) to generate meta feature \n19: Optimize \\(F\\) using fusion loss \\(L_f\\) and guide loss \\(L_g\\) by Eq. 6 \n20: /* Meta-feature embedding inner update */ \n21: if epoch%N==0 then \n22: for t=1 to n do \n23: Sample an image pair \\((I_i^{mtr}, I_v^{mtr})\\) from \\(S_{mtr}\\) \n24: Compute \\(\\theta_{F'}\\) by Eq. 1 \n25: Sample an image pair \\((I_i^{mts}, I_v^{mts})\\) from \\(S_{mts}\\) \n26: Calculate \\(L_f\\) using \\(\\theta_{F'}\\) \n27: Update \\(\\theta_{MFG}\\) and \\(\\theta_{FT}\\) using the gradient of \\(L_f\\) by Eqs. 2 and 3 \n28: end for \n29: end if \n30: end while \n31: /* Mutual promotion*/ \n32: if Number of round ≤ R then \n33: Go back to line 9 \n34: end if" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.824 + ], + "angle": 0, + "content": "that, \\(D\\) is finetuned for 150 epochs. Following the same strategy, the mutual promotion is carried out \\(R\\) rounds. The images are resized to \\(512 \\times 384\\) with batchsize of 1. The hyperparameter \\(\\lambda_{g}\\) is set to 0.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Metric. Three metrics are used for IVIF evaluation: entropy (EN) [33], mutual information (MI) [28] and visual information fidelity (VIF) [7]. EN evaluates the information richness in an image, and the higher EN means more information. MI evaluates the information similarity between" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.309, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/ultralytics/YOLOv5" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13959" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.09, + 0.186, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.191, + 0.174, + 0.197 + ], + "angle": 0, + "content": "(a) Visible-infrared images" + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.091, + 0.282, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.191, + 0.273, + 0.197 + ], + "angle": 0, + "content": "(b) SoFusion and feature map" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.091, + 0.376, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.191, + 0.371, + 0.197 + ], + "angle": 0, + "content": "(c) CoFusion and feature map" + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.091, + 0.472, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.191, + 0.468, + 0.197 + ], + "angle": 0, + "content": "(d) MetaFusion and feature map" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.208, + 0.461, + 0.223 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of different optimization methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.227, + 0.468, + 0.267 + ], + "angle": 0, + "content": "Table 1. Effect study of meta-feature embedding by comparing with different optimization methods on \\( \\mathrm{M}^3\\mathrm{FD} \\). The best result is in red." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.274, + 0.451, + 0.378 + ], + "angle": 0, + "content": "
MethodMetric
MIENVIF
SoFusion14.1647.0781.190
CoFusion14.1877.0891.345
MetaFusion14.5117.2491.515
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.468, + 0.548 + ], + "angle": 0, + "content": "the input images and fused image. The higher MI illustrates more information of the input images is fused. VIF measures the ability to extract visible information from the input image, and a larger VIF represents less visible distortion in the fused result. Here, we use the V channel in HSV space of the fusion results to calculate these metrics. Moreover, we use \\(\\mathrm{mAP}_{50\\rightarrow 95}\\) [9] to comprehensively evaluate OD performance, where the average of mAPs sampling every 5 from \\(\\mathrm{AP}_{50}\\) to \\(\\mathrm{AP}_{95}\\) is calculated. A higher \\(\\mathrm{mAP}_{50\\rightarrow 95}\\) means better OD effect." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.23, + 0.575 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Effect of meta-feature embedding. In Sec. 3.1, we introduce the meta-feature embedding that generates the meta features from object detection features to help IVIF network learn more object semantic information. To verify its effect, separate optimization method (SoFusion) and cascaded optimization method (CoFusion) are compared, as shown in Figure 1(a)-(b). Specifically, SoFusion firstly optimizes \\( F \\) and then uses the generated fusion results to optimize \\( D \\). CoFusion treats \\( D \\) as a constraint to optimize \\( F \\) with the loss \\( L_{f} \\) and \\( L_{d} \\). The results are shown in Table 1. Our MetaFusion achieves the best results on IVIF. The reason is that SoFusion neglects the help of \\( D \\), and directly utilizing the high-level \\( D \\) to guide the pixel-level \\( F \\) has feature mismatch in CoFusion. In contrast, we implement the meta-feature embedding that generates meta features from \\( D \\) according to the ability of \\( F \\). Thus, the meta features are naturally compatible with \\( F \\), thereby providing effective object semantic features to \\( F \\). Visual comparison is shown in Figure 5. MetaFusion highlights the object feature and generates the clearest result." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Influence of mutual promotion learning. As described" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.13 + ], + "angle": 0, + "content": "Table 2. Influence study of mutual promotion learning by evaluating the performances of round \\( R \\) on \\( {\\mathrm{M}}^{3}\\mathrm{{FD}} \\) . The best result is in red." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.136, + 0.87, + 0.24 + ], + "angle": 0, + "content": "
MethodMetric
MIENVIFmAP50→95 (%)
R = 014.1647.0781.19055.6
R = 114.4647.2261.47455.8
R = 214.5117.2491.51556.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.253, + 0.891, + 0.28 + ], + "angle": 0, + "content": "Table 3. Study of multi-level meta-feature embedding by comparing different configurations on \\(\\mathbf{M}^3\\mathrm{FD}\\). The best result is in red." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.285, + 0.875, + 0.389 + ], + "angle": 0, + "content": "
MethodMetric
MIENVIF
MetaFusion-L114.4507.2201.259
MetaFusion-L214.4527.2201.377
MetaFusion-L314.4647.2261.474
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.516 + ], + "angle": 0, + "content": "in Sec. 3.2, we propose the mutual promotion learning promote \\( F \\) and \\( D \\) alternately. Here, we study the influence of mutual promotion learning by evaluating round \\( R = 0,1,2 \\). In particular, \\( R = 0 \\) represents \\( F \\) and \\( D \\) are trained respectively. As shown in Table 2, the results of \\( F \\) are better with the \\( R \\) increases. Comprehensively considering training efficiency and performance, we take \\( R = 2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.893, + 0.789 + ], + "angle": 0, + "content": "Study of multi-level meta-feature embedding. We implement multi-level meta-feature embedding to build MetaFusion framework, as introduced in Sec. 3.3. Here, we study the number of meta-feature embedding levels with the following configurations. One-level meta-feature embedding (MetaFusion-L1) with \\( MFE = \\{MFG_1, FT_1\\} \\), \\( F = \\{FFB_1, IRM\\} \\) and \\( D = \\{DFB_1, DH\\} \\). Two-level meta-feature embedding (MetaFusion-L2) with \\( MFE = \\{MFG_1, MFG_2, FT_1, FT_2\\} \\), \\( F = \\{FFB_1, FFB_2, IRM\\} \\) and \\( D = \\{DFB_1, DFB_2, DH\\} \\). Three-level meta-feature embedding (MetaFusion-L3) with \\( MFE = \\{MFG_1, MFG_2, MFG_3, FT_1, FT_2, FT_3\\} \\), \\( F = \\{FFB_1, FFB_2, FFB_3, IRM\\} \\) and \\( D = \\{DFB_1, DFB_2, DFB_3, DH\\} \\). The results are shown in Table 3. With the number of multi-level meta-feature embedding increases, \\( F \\) achieves higher image fusion quality. Since meta-feature embedding provide multi-level object semantic features to \\( F \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.864, + 0.817 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare the proposed MetaFusion with eight SOTA fusion methods to verify the superiority: FusionGAN [25], GANMcC [26], MFEIF [21], YDTR [36], PIAFusion [35], SwinFusion [23], Tardal [20], and U2Fusion [44]. Their available codes and recommended parameter settings are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13960" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.092, + 0.266, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.185, + 0.255, + 0.196 + ], + "angle": 0, + "content": "Infrared-visible image" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.093, + 0.411, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.185, + 0.374, + 0.195 + ], + "angle": 0, + "content": "FusionGAN" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.093, + 0.556, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.186, + 0.518, + 0.196 + ], + "angle": 0, + "content": "GANMcC" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.093, + 0.701, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.186, + 0.657, + 0.196 + ], + "angle": 0, + "content": "MFEIF" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.093, + 0.846, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.186, + 0.806, + 0.196 + ], + "angle": 0, + "content": "U2Fusion" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.197, + 0.266, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.291, + 0.213, + 0.301 + ], + "angle": 0, + "content": "YDTR" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.197, + 0.41, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.291, + 0.374, + 0.301 + ], + "angle": 0, + "content": "PIAFusion" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.197, + 0.555, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.291, + 0.522, + 0.301 + ], + "angle": 0, + "content": "SwinFusion" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.197, + 0.7, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.291, + 0.652, + 0.301 + ], + "angle": 0, + "content": "Tardal" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.197, + 0.846, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.291, + 0.813, + 0.301 + ], + "angle": 0, + "content": "MetaFusion" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.302, + 0.268, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.401, + 0.254, + 0.413 + ], + "angle": 0, + "content": "Infrared-visible image" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.302, + 0.411, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.401, + 0.371, + 0.411 + ], + "angle": 0, + "content": "FusionGAN" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.302, + 0.555, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.402, + 0.515, + 0.412 + ], + "angle": 0, + "content": "GANMcC" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.302, + 0.701, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.401, + 0.654, + 0.411 + ], + "angle": 0, + "content": "MFEIF" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.302, + 0.846, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.402, + 0.803, + 0.412 + ], + "angle": 0, + "content": "U2Fusion" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.413, + 0.266, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.506, + 0.214, + 0.516 + ], + "angle": 0, + "content": "YDTR" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.413, + 0.411, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.506, + 0.374, + 0.516 + ], + "angle": 0, + "content": "PIAFusion" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.413, + 0.555, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.506, + 0.523, + 0.516 + ], + "angle": 0, + "content": "SwinFusion" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.413, + 0.701, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.506, + 0.653, + 0.516 + ], + "angle": 0, + "content": "Tardal" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.413, + 0.846, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.506, + 0.813, + 0.516 + ], + "angle": 0, + "content": "MetaFusion" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.517, + 0.266, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.616, + 0.253, + 0.627 + ], + "angle": 0, + "content": "Infrared-visible image" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.517, + 0.411, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.616, + 0.37, + 0.626 + ], + "angle": 0, + "content": "FusionGAN" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.517, + 0.555, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.616, + 0.513, + 0.626 + ], + "angle": 0, + "content": "GANMcC" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.517, + 0.701, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.616, + 0.653, + 0.626 + ], + "angle": 0, + "content": "MFEIF" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.517, + 0.846, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.616, + 0.802, + 0.626 + ], + "angle": 0, + "content": "U2Fusion" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.63, + 0.266, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.724, + 0.212, + 0.734 + ], + "angle": 0, + "content": "YDTR" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.63, + 0.411, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.724, + 0.373, + 0.734 + ], + "angle": 0, + "content": "PIAFusion" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.63, + 0.555, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.724, + 0.522, + 0.734 + ], + "angle": 0, + "content": "SwinFusion" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.63, + 0.7, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.724, + 0.652, + 0.734 + ], + "angle": 0, + "content": "Tardal" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.63, + 0.846, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.724, + 0.812, + 0.734 + ], + "angle": 0, + "content": "MetaFusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.744, + 0.871, + 0.76 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of different fusion methods on \\(\\mathbf{M}^3\\mathrm{FD}\\) (first group), RoadScene (second group) and TNO (third group)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.771, + 0.46, + 0.787 + ], + "angle": 0, + "content": "adopted to generate fusion results for the fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Qualitative results of different fusion methods are shown in Figure 6. All the fusion methods can fuse the main features of the infrared and visible images to some extent. However, FusionGAN, GANMcC, MFEIF and YDTR produce blurred edge details, and U2Fusion, PIAFusion, Swinfusion and Tardal generate low contrast objects, as shown in the red rectangular boxes. In contrast, the images gener" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.771, + 0.892, + 0.801 + ], + "angle": 0, + "content": "ated by the proposed MetaFusion contain clear edge details and high contrast objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Sequentially, we provide quantitative results of different fusion methods in Table 4. Our MetaFusion generally achieves the largest or the second-largest metric values. In detail, the high EN and MI illustrate the fusion images by our MetaFusion contain high contrast object and clear edge details. The large VIF shows our fusion results have high-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13961" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 4. Quantitative results of different fusion methods on TNO, RoadScene and \\(\\mathbf{M}^3\\mathrm{FD}\\) datasets. The model inference time is counted on a NVIDIA GeForce RTX 2080 Ti. The best result is in red and the second best one is in violet." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.125, + 0.875, + 0.338 + ], + "angle": 0, + "content": "
MethodM3FDRoadSceneTNO40Time (s)
MIEntVIFMIEntVIFMIEntVIF
FusionGAN [25]13.4456.7220.30314.2037.1010.25113.0686.5340.2520.040
GANMcC [26]13.7316.8650.45314.0177.0080.42213.4856.7420.4240.081
MFEIF [21]12.9576.4780.40113.4896.7420.26013.3606.6800.3950.029
U2Fusion [44]13.8166.9080.54513.8886.9440.45613.8896.9440.6360.043
YDTR [36]12.6946.3470.36113.1666.5830.23612.8626.4310.2800.086
PIAFusion [35]13.3266.6630.43713.2626.5330.20513.9316.9600.5060.055
SwinFusion [23]13.0516.5250.46013.2676.6330.31313.8796.9330.4791.081
Tardal [20]13.6366.8180.65015.0097.5040.48313.0306.5150.9100.030
MetaFusion14.5117.2491.51514.2187.0220.96914.6577.3231.4620.015
" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.356, + 0.187, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.356, + 0.282, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.356, + 0.376, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.356, + 0.474, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.411, + 0.187, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.411, + 0.282, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.411, + 0.377, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.411, + 0.474, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.475, + 0.187, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.47, + 0.282, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.475, + 0.377, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.475, + 0.474, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.53, + 0.157, + 0.535 + ], + "angle": 0, + "content": "SwinFusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.529, + 0.246, + 0.535 + ], + "angle": 0, + "content": "Tardal" + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.529, + 0.349, + 0.535 + ], + "angle": 0, + "content": "MetaFusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.529, + 0.445, + 0.535 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Figure 7. Visual results of object detection based on different fusion methods on \\(\\mathbf{M}^3\\mathrm{FD}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.643 + ], + "angle": 0, + "content": "quality visual effect and small distortion compared with the source images. Moreover, our method achieves the fastest inference time that only needs 0.015s to generate one fusion result." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.469, + 0.681 + ], + "angle": 0, + "content": "4.4. Evaluation on Infrared-visible Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.469, + 0.796 + ], + "angle": 0, + "content": "A better fusion image can provide more information to improve object detection. Here, we verify the effectiveness of our method by evaluating object detection accuracy with the infrared and visible fusion images. To make a fair comparison, we firstly generate fusion results using our method and SOTA methods, and then use the results to retrain the object detection baseline YOLOv5s, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Figure 7 shows the object detection results. In general, the SOTA methods improve the performance of object detection. In contrast, our MetaFusion achieves a better performance, e.g., the person is accurately detected (see the blue ellipse). Table 5 shows the quantitative results. Our fusion results help the detection network achieve the highest object detection accuracy. This further proves that our" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Table 5. Quantitative results of object detection based on different fusion methods on \\(\\mathbf{M}^3\\mathrm{FD}\\) dataset. The best result is in red and the second best one is in violet." + }, + { + "type": "table", + "bbox": [ + 0.555, + 0.402, + 0.837, + 0.599 + ], + "angle": 0, + "content": "
Method\\( \\mathrm{mAP}_{50\\rightarrow 95}(\\%) \\)
FusionGAN [25]54.2
GANMcC [26]55.2
MFEIF [21]55.4
U2Fusion [44]55.7
YDTR [36]55.4
PIAFusion [35]55.6
SwinFusion [23]55.4
Tardal [20]54.4
MetaFusion56.5
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.613, + 0.892, + 0.643 + ], + "angle": 0, + "content": "method can generate high-quality fusion results especially for the objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.661, + 0.62, + 0.677 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.853 + ], + "angle": 0, + "content": "This paper presents a joint fusion and detection learning framework through introducing the meta-feature embedding model. Based on the meta learning idea, the meta-feature embedding model can generate object semantic features according to the fusion network ability, thereby bridging the feature gap between these two different-level tasks. Moreover, a mutual promotion learning between fusion and detection tasks is further implemented to improve their performances. Both quantitative and qualitative results demonstrate the superior performance of our method compared with the state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by National Natural Science Foundation of China under Grant Nos. 62176038, 62001450 and U1903215." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13962" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9465-9474, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6154-6162, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[4] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-sclil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14166-14175, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[5] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster R-CNN: towards accurate few-shot object detection with attentive feature alignment. In AAAI, pages 780-789. AAAI Press, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[7] Yu Han, Yunze Cai, Yin Cao, and Xiaoming Xu. A new image fusion performance metric based on visual information fidelity. Information Fusion, 14:-, 04 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Kaiming He, Georgia Gkioxari, Piotr Dollr, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[9] Liqiang He and Sinisa Todorovic. Destr: Object detection with split transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9367-9376, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[10] Ruichao Hou, Dongming Zhou, Rencan Nie, Dong Liu, Lei Xiong, Yanbu Guo, and Chuanbo Yu. Vif-net: An unsupervised framework for infrared and visible image fusion. IEEE Transactions on Computational Imaging, 6:640-651, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[11] Muhammad Abdullah Jamal, Liqiang Wang, and Boqing Gong. A lazy approach to long-horizon gradient-based meta-learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6577-6586, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[12] Huafeng Li, Yueliang Cen, Yu Liu, Xun Chen, and Zhengtao Yu. Different input resolutions and arbitrary output resolution: A meta learning-based deep framework for infrared and visible image fusion. IEEE Transactions on Image Processing, 30:4070-4083, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Hui Li, Xiao-Jun Wu, and Josef Kittler. Mdlatrr: A novel decomposition method for infrared and visible image fusion. IEEE Transactions on Image Processing, 29:4733-4746, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Jing Li, Jianming Zhu, Chang Li, Xun Chen, and Bin Yang. Cgtf: Convolution-guided transformer for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 71:1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[17] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP:1-1, 07 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[19] Benlin Liu, Yongming Rao, Jiwen Lu, Jie Zhou, and Chojui Hsieh. Metadistiller: Network self-boosting via metalearned top-down distillation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16, pages 694–709. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[20] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[21] J. Liu, X. Fan, J. Jiang, R. Liu, and Z. Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[22] Yu Liu, Shuping Liu, and Zengfu Wang. A general framework for image fusion based on multi-scale transform and s-parse representation. Information Fusion, 24:147-164, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[23] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[24] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping Zhang. Ddcgan: A dual-discriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29:4980-4995, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[25] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusion: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11-26, 08 2019. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Jiayi Ma, Hao Zhang, Zhenfeng Shao, Pengwei Liang, and Han Xu. Ganmcc: A generative adversarial network with" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13963" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "multiclassification constraints for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1-14, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Yan Mo, Xudong Kang, Puhong Duan, Bin Sun, and Shutao Li. Attribute filter based infrared and visible image fusion. Information Fusion, 75:41-54, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[28] Guihong Qu, Dali Zhang, and Pingfan Yan. Information measure for performance of image fusion. *Electronics Letters*, 38:313-315, 04 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[29] Aniruddh Raghu, Jonathan Lorraine, Simon Kornblith, Matthew McDermott, and David K Duvenaud. Meta-learning to improve pre-training. Advances in Neural Information Processing Systems, 34:23231-23244, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[30] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[32] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 658-666, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[33] Wesley Roberts, Jan van Aardt, and Fethi Ahmed. Assessment of image fusion procedures using entropy, image quality, and multispectral classification. Journal of Applied Remote Sensing, 2:1-28, 05 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[34] Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshitaka Ushiku, and Tatsuya Harada. Multispectral object detection for autonomous vehicles. In Proceedings of the on Thematic Workshops of ACM Multimedia 2017, Thematic Workshops '17, page 3543, New York, NY, USA, 2017. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[35] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83-84:79-92, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[36] Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer. IEEE Transactions on Multimedia, pages 1-16, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[37] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249-251, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[38] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[39] Zhishe Wang, Yuanyuan Wu, Junyao Wang, Jiawei Xu, and Wenyu Shao. Res2fusion: Infrared and visible image fusion based on dense res2net and double nonlocal attention models. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14248-14257, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[41] Chengming Xu, Yanwei Fu, Chen Liu, Chengjie Wang, Jilin Li, Feiyue Huang, Li Zhang, and Xiangyang Xue. Learning dynamic alignment via meta-filter for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5182-5191, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[42] Dongli Xu, Jinghong Deng, and Wen Li. Revisiting ap loss for dense object detection: Adaptive ranking pair selection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 14187-14196, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[43] Han Xu, Meiqi Gong, Xin Tian, Jun Huang, and Jiayi Ma. Cufd: An encoderdecoder network for visible and infrared image fusion based on common and unique feature decomposition. Computer Vision and Image Understanding, 218:103407, 03 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[44] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[45] Han Xu, Xinya Wang, and Jiayi Ma. Drf: Disentangled representation for visible and infrared image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1-13, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[46] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Rui-Wei Zhao, Tao Zhang, Xuequan Lu, and Shang Gao. Cream: Weakly supervised object localization via class re-activation mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9437–9446, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[47] Meilong Xu, Linfeng Tang, Hao Zhang, and Jiayi Ma. Infrared and visible image fusion via parallel scene and texture learning. Pattern Recognition, 132:108929, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[48] Shilong Zhang, Zhuoran Yu, Liyang Liu, Xinjiang Wang, Aojun Zhou, and Kai Chen. Group r-cnn for weakly semi-supervised object detection with points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9417-9426, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[49] Fan Zhao and Wenda Zhao. Learning specific and general realm feature representations for image fusion. IEEE Transactions on Multimedia, 23:2745-2756, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[50] Fan Zhao, Wenda Zhao, and Hutchuan Lu. Interactive feature embedding for infrared and visible image fusion. arXiv preprint arXiv:2211.04877, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[51] Fan Zhao, Wenda Zhao, Libo Yao, and Yu Liu. Self-supervised feature adaption for infrared and visible image fusion. Information Fusion, 76:189-203, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[52] Wenda Zhao, Xueqing Hou, You He, and Huchuan Lu. Defocus blur detection via boosting diversity of deep ensemble networks. IEEE Transactions on Image Processing, 30:5426-5438, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[53] Wenda Zhao, Huimin Lu, and Dong Wang. Multisensor image fusion and enhancement in spectral total variation domain. IEEE Transactions on Multimedia, 20(4):866-879, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13964" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[54] Wenda Zhao, Dong Wang, and Huchuan Lu. Multi-focus image fusion with a natural enhancement via a joint multi-level deeply supervised convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 29(4):1102-1115, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[55] Wenda Zhao, Bowen Zheng, Qiuhua Lin, and Hutchuan Lu. Enhancing diversity of defocus blur detectors via cross-ensemble network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8905–8913, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[56] Zhengjie Zhu, Xiaogang Yang, Ruitao Lu, Tong Shen, Xueli Xie, and Tao Zhang. Clf-net: Contrastive learning for infrared and visible image fusion network. IEEE Transactions on Instrumentation and Measurement, 71:1-15, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13965" + } + ] +] \ No newline at end of file diff --git a/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_origin.pdf b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d03e78c5a705f5f5fc31cdf5a0e456334b82e21f --- /dev/null +++ b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/70dfb2c3-4648-4819-a764-7cd79a227474_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ec328e526f6247931c86126126ff2acdc78c1f38810ffd1a25724e73f379b17 +size 4755870 diff --git a/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/full.md b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..72b93c0342da98a82f3eaa70132c80e0fa6c79fd --- /dev/null +++ b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/full.md @@ -0,0 +1,454 @@ +# MetaFusion: Infrared and Visible Image Fusion via Meta-Feature Embedding from Object Detection + +Wenda Zhao $^{1}$ , Shigeng Xie $^{1}$ , Fan Zhao $^{2*}$ , You He $^{3}$ , Huchuan Lu $^{1}$ + +$^{1}$ Dalian University of Technology, China + +$^{2}$ Liaoning Normal University, China + +3 Tsinghua University, China + +zhaowenda@dlut.edu.cn; xieshigeng@mail.dlut.edu.cn + +Fan-Zhao20@163.com; youhe_nau@163.com; lhchuan@dlut.edu.cn + +# Abstract + +Fusing infrared and visible images can provide more texture details for subsequent object detection task. Conversely, detection task furnishes object semantic information to improve the infrared and visible image fusion. Thus, a joint fusion and detection learning to use their mutual promotion is attracting more attention. However, the feature gap between these two different-level tasks hinders the progress. Addressing this issue, this paper proposes an infrared and visible image fusion via meta-feature embedding from object detection. The core idea is that meta-feature embedding model is designed to generate object semantic features according to fusion network ability, and thus the semantic features are naturally compatible with fusion features. It is optimized by simulating a meta learning. Moreover, we further implement a mutual promotion learning between fusion and detection tasks to improve their performances. Comprehensive experiments on three public datasets demonstrate the effectiveness of our method. Code and model are available at: https://github.com/wdzhao123/MetaFusion. + +# 1. Introduction + +Multi-modality sensor technology has promoted the application of multi-modality images in different areas. Among them, infrared images and visible images have been utilized commonly, as the information contained in these two modalities is complementary. Specifically, infrared images can supply object thermal structures without being affected by illumination. But they are short of texture details. On the contrary, visible images can catch the texture information for the scene. But they are severely affected by light. Thus, many methods [15,25,35,43-45,47] focus on s + +![](images/a829f059b0b06fa7ccf0112ee4701d81542d6560ff3a0acd00647ef62525dea4.jpg) +(a) Separate optimization method + +![](images/8e2b9f5f41446252228b71c0564d6a3d37cf305c7addcac4d9a0d9e12f1c36f1.jpg) +(b) Cascaded optimization method + +![](images/7d75aaed38dfc8eca509b018fd0184e39f1718fff02a6b437df2ce7e876fb386.jpg) +(c) Meta-feature embedding method +Figure 1. Different joint learning methods of infrared and visible image fusion (IVIF) and object detection (OD). (a) Separate optimization method: IVIF network $\Psi$ is firstly optimized by fusion loss $L_{f}$ . Then, fusion result $z$ is generated by $\Psi$ from the input infrared and visible image pair $x,y$ . Finally, OD network $\Phi$ is optimized by detection loss $L_{d}$ using $z$ . (b) Cascaded optimization method: OD network $\Phi$ is treated as a constraint to optimize IVIF network $\Psi$ by the loss $L_{f}$ and $L_{d}$ . (c) Meta-feature embedding method: Meta-feature embedding $\Gamma$ is optimized to learn how to guide $\Psi$ to have a low $L_{f}$ . Then, $\Gamma$ generates meta feature from detection feature $F_{i}^{d}$ and fusion feature $F_{i}^{d}$ . Finally, the meta feature is used to guide $\Psi$ by the loss $L_{d}$ . + +tudying pixel-level infrared and visible image fusion (IVIF), thereby helping high-level tasks improve performance, e.g., object detection (OD) [20, 34]. + +IVIF and OD can greatly benefit form each other. IVIF + +generates the fused image that contains more information than any single modality image to improve OD. However, IVIF mainly focuses on the pixel relationship between an image pair, and there is little consideration for object semantic. In contrast, OD can provide rich object semantic information to IVIF, as its aim is locating the objects. Therefore, this paper studies a joint learning framework between IVIF and OD to improve their performances. + +Existing joint learning methods of IVIF and OD can be divided into two categories: separate optimization and cascaded optimization. Separate optimization firstly trains IVIF network, and then trains OD network using IVIF results, as shown in Figure 1(a). Thus, most methods focus on improving the fusion effect, e.g., designing networks [13, 23, 35, 36] and introducing constraints [10, 26, 45]. Obviously, separate optimization neglects the help of OD. Cascaded optimization adopts OD network as a constraint to train IVIF network, and thus forces the IVIF network to generate fusion images with easily detected objects [20], as shown in Figure 1(b). However, directly utilizing the high-level OD constraint to guide the pixel-level IVIF will result in limited effect. Therefore, we leverage OD feature maps that guide IVIF feature maps to obtain more semantic information. Unfortunately, OD features are mismatched with IVIF features due to their task-level difference. Addressing this issue, we propose a meta-feature embedding network (MFE), as shown in Figure 1(c). The idea is that if MFE generates OD features according to the IVIF network ability, the OD features are naturally compatible with the IVIF network, and the optimization can be achieved by simulating a meta learning. + +Specifically, an infrared and visible image fusion via meta-feature embedding from object detection is proposed, which is named as MetaFusion. MetaFusion includes IVIF network $(F)$ , OD network $(D)$ , and $MFE$ . In particular, $MFE$ is expected to generate meta features to bridge the gap between $F$ and $D$ , which is optimized by two alternate steps: inner update and outer update. In the inner update process, we firstly optimize $F$ using meta training set $S_{mtr}$ to obtain its updated network $F'$ . Then, $F'$ calculates the fusion loss on meta testing set $S_{mts}$ to optimize $MFE$ . The motivation is that if $MFE$ successfully generates meta features which are compatible with $F$ , $F'$ will produce better fused images, i.e., the fusion loss should be lower. In the outer update process, $F$ is optimized with the guide of the meta features generated by the fixed $MFE$ on $S_{mtr}$ and $S_{mts}$ . In this way, $F$ can learn how to extract semantic information to improve fusion quality. In the above two alternate steps, $D$ is fixed to offer detection semantic information. Thus, we further implement a mutual promotion learning, where we use the improved $F$ to generate fusion results to finetune $D$ , and then the improved $D$ offers better semantic information to optimize $F$ . + +In summary, our contributions are as follows. (1) We explore the joint learning framework of IVIF and OD, and propose MetaFusion to obtain superior performance on these two tasks. (2) Meta-feature embedding network is designed to generate meta features that bridge the gap between $F$ and $D$ . (3) Sequentially, mutual promotion learning between $F$ and $D$ is introduced to improve their performances. (4) Extensive experiments on image fusion and object detection validate the effectiveness of the proposed method. + +# 2. Related Work + +Image Fusion. Traditional image fusion methods use hand-crafted features, such as sparse representation [22], spectral variation [53] and low-rank representation [14], which can not handle the complex scenes well. Nowadays, deep learning-based image fusion methods [16, 24, 27, 39, 49, 50, 54, 56] are proposed. Ma et al. [23] propose a swim transformer based image fusion method with cross-domain long-range learning. Xu et al. [44] use feature extraction and measurement to estimate the degree of information preservation in image fusion. Zhao et al. [51] design a self-supervised feature extraction model. Besides, feature diversity enhancement and fusion is received attention [52, 55]. However, most of them neglect the help of high-level tasks. Recently, Liu et al. [20] propose a joint learning of IVIF and OD. They treat OD network as an additional constraint to help IVIF network generate fusion result with more clear objects. + +However, directly using the high-level OD network to guide pixel-level IVIF may not obtain an effective constraint. Moreover, they use the fusion results as the connection between IVIF network and OD network, ignoring the pixel-level semantic information in OD features. In contrast, we adopt semantic information in OD features to help IVIF, and conduct a meta-feature embedding to generate meta features from OD features. Then, those meta features are used to guide IVIF network to learn pixel-level semantic information. + +Object Detection. Deep learning-based object detection methods have made great progress, e.g., network structure-designed methods [2, 8, 30, 31] and loss constraint methods [18, 32, 42]. Carion et al. [3] introduce a transformer encoder-decoder architecture for object detection. Xu et al. [42] rank positive and negative sample pairs and improve the ranking effect by a clustering algorithm to built object detection loss. On the other hand, many weakly supervised methods [40, 46, 48] are proposed. Wu et al. [40] design a model to suppress the background activation and obtain object region, where image-level labels are used. Zhang et al. [48] use instance proposal grouping to learn object detection from a single point annotation. + +Most of those object detection methods leverage single modality image, e.g., visible images. However, the visible + +![](images/caef070d7ffe13a45f7345ec4df249e9870fe0669478a6b797e9d39d2eb76a5e.jpg) +Figure 2. Framework illustration of the proposed MetaFusion. MetaFusion includes three parts: infrared and visible image fusion (IVIF), object detection (OD), and meta-feature embedding (MFE). IFIV network includes four blocks: three feature fusion blocks $FFB_{1}$ , $FFB_{2}$ , $FFB_{3}$ used to extract and fuse features and one image reconstruction module $IRM$ used to reconstruct fusion result. MFE contains three meta-feature generators $MFG_{1}$ , $MFG_{2}$ , $MFG_{3}$ conducted to generate multi-level meta features from OD features $F_{ej}$ , and three feature transform networks $FT_{1}$ , $FT_{2}$ , $FT_{3}$ used to transfer the meta features to fusion features. OD network is divided as three feature extraction blocks $DFB_{1}$ , $DFB_{2}$ , $DFB_{3}$ and one OD detection head $DH$ . + +images are affected by light conditions, thereby influencing the performance of object detection. On the contrary, we leverage IVIF results to provide infrared information without being influenced by illumination, therefore can offer more stable detection performance. + +Meta Learning. Meta learning aims to use the learned knowledge to quickly adapt to new tasks, which has been used to many fields: model pre-training [5, 11, 29], image classification [17, 19], few-shot learning [1, 4, 6, 41], etc. For example, Raghu et al. [29] propose an implicit differentiation and backpropagation meta-learning method and improve the efficiency of learning pre-training hyperparameters. Xu et al. [41] design a dynamic alignment in meta learning, which can emphasize information in query according to the support. Li et al. [12] use the meta learning to handle different input resolutions, thereby generating fusion results with arbitrary resolutions. + +Inspired by the above meta learning methods, we design a meta-feature embedding network to generate the meta features from object detection features, thereby helping IVIF network fuse more object semantic information. + +# 3. Proposed Method + +Our MetaFusion framework is shown in Figure 2, which includes three subnetworks: IVIF network $(F)$ generates fusion results, OD network $(D)$ offers semantic features, and meta-feature embedding network $(MFE)$ uses OD features to guide IVIF to fuse more object semantic information. The challenge is that OD features can hardly be directly used to guide IVIF due to their task gap. Thus, our focus is to design $MFE$ and optimize it. Details are as follows. + +![](images/136eafc183a7d3af98a5a0e498950ee091625d9583a47f04d88607982d561a7c.jpg) +Figure 3. Illustration of the meta-feature embedding learning. There are two stages in the meta-feature embedding: (a) inner update and (b) outer update. In the inner update, $MFG$ and $F^T$ are trained according to current $F'$ semantic extracting ability to generate meta features. In the outer update, $F$ is trained which is guided by the meta features, thereby learning to extract object semantic. The two stages alternate every N epochs. + +# 3.1. Meta-feature Embedding + +Meta-feature embedding includes meta-feature generator (MFG) and feature transform (FT). MFG generates meta feature $F_{mj}$ according to IVIF feature $F_{uj}$ from OD feature $F_{ej}$ , i.e., $F_{mj} = MFG(F_{uj}, F_{ej})$ , where $j$ is the feature level index. FT transfers the meta feature $F_{mj}$ to fusion feature $F_{uj}$ by producing feature bridge $F_{tj}$ . Specifically, meta-feature embedding learning is divided two stages: inner update stage and outer update stage, as shown in Figure 3. + +In the inner update stage, we optimize $MFG$ , $FT$ and $F$ (see Figure 3(a)). Given meta training set $S_{mtr}$ and meta testing set $S_{mts}$ , we firstly update the parameter $\theta_F$ of $F$ on + +$S_{mtr}$ through meta feature guidance: + +$$ +\theta_ {F ^ {\prime}} = \theta_ {F} - \beta_ {F ^ {\prime}} \nabla_ {\theta_ {F}} L _ {g} \left(F _ {m j}, F _ {t j}\right) = \theta_ {F} - \beta_ {F ^ {\prime}} \frac {\partial L _ {g} \left(F _ {m j} , F _ {t j}\right)}{\partial \theta_ {F}}, \tag {1} +$$ + +where the guide loss $L_{g}$ is $L_{2}$ distance, $\beta_{F^{\prime}}$ is the learning rate of $F^{\prime}$ , $\theta_{F^{\prime}}$ is the parameter of $F^{\prime}$ that is updated from $F$ . Then, we use $F^{\prime}$ with parameter $\theta_{F^{\prime}}$ to calculate the fusion loss $L_{f}$ on meta testing set $S_{mts}$ . Here, $L_{f}$ can measure the effect of using the meta features to guide $F$ , i.e., when $L_{f}$ is lower, the semantic extracting ability of $F$ will be improved. Therefore, we use $L_{f}$ to update parameter $\theta_{MFG}$ of $MFG$ and parameter $\theta_{FT}$ of $FT$ by + +$$ +\theta_ {M F G} = \theta_ {M F G} - \beta_ {M F G} \nabla_ {\theta_ {M F G}} L _ {f} \left(I _ {f} ^ {m t s}, I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\right), (2) +$$ + +$$ +\theta_ {F T} = \theta_ {F T} - \beta_ {F T} \nabla_ {\theta_ {F T}} L _ {f} \left(I _ {f} ^ {m t s}, I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\right), \tag {3} +$$ + +where the fusion loss $L_{f}$ is SSIM loss [38], $I_{i}^{mts}$ and $I_{v}^{mts}$ are infrared and visible images from $S_{mts}$ , $I_{f}^{mts} = F'(I_{i}^{mts}, I_{v}^{mts})$ is the meta testing fusion result, and $\beta_{MFG}$ and $\beta_{FT}$ are the learning rate of MFG and $FT$ . $\nabla_{\theta_{MFG}} L_{f}$ can be calculated by + +$$ +\begin{array}{l} \nabla_ {\theta_ {M F G}} L _ {f} \left(F ^ {\prime} \left(I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\right), I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\right) \\ = \frac {\partial L _ {f} \left(F ^ {\prime} \left(I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\right) , I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\right)}{\partial \theta_ {F ^ {\prime}}} * \left(- \frac {\partial^ {2} L _ {g} \left(F _ {m j} , F _ {t j}\right)}{\partial \theta_ {F} \partial \theta_ {M F G}}\right). \tag {4} \\ \end{array} +$$ + +Similarly, $\nabla_{\theta_{FT}}L_f$ can be calculated by + +$$ +\begin{array}{l} \nabla_ {\theta_ {F T}} L _ {f} \left(F ^ {\prime} \left(I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\right), I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\right) \\ = \frac {\partial L _ {f} \left(F ^ {\prime} \left(I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\right) , I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\right)}{\partial \theta_ {F ^ {\prime}}} * \left(- \frac {\partial^ {2} L _ {g} \left(F _ {m j} , F _ {t j}\right)}{\partial \theta_ {F} \partial \theta_ {F T}}\right). \tag {5} \\ \end{array} +$$ + +Thus, $MFG$ can learn how to generate meta features according to current $F^{\prime}$ semantic extracting ability, i.e., makes meta features be compatible with $F$ . + +In the outer update stage, $F$ is trained with the initial parameter $\theta_F$ , as shown in Figure 3(b). Given the training data $S = \{S_{mtr}, S_{mts}\}$ , $F$ is optimized using the fusion loss $L_f$ and guide loss $L_g$ by + +$$ +\begin{array}{l} \theta_ {F} = \theta_ {F} - \beta_ {F} \nabla_ {\theta_ {F}} \left(L _ {f} \left(I _ {f}, I _ {i}, I _ {v}\right) + \lambda_ {g} \sum_ {j = 1} ^ {3} L _ {g} \left(F _ {m j}, F _ {t j}\right)\right) \\ = \theta_ {F} - \beta_ {F} \left(\frac {\partial L _ {f} \left(I _ {f} , I _ {i} , I _ {v}\right)}{\partial \theta_ {F}} + \lambda_ {g} \sum_ {j = 1} ^ {3} \frac {\partial L _ {g} \left(F _ {m j} , F _ {t j}\right)}{\partial \theta_ {F}}\right), \tag {6} \\ \end{array} +$$ + +where $I_{i}$ and $I_{v}$ are infrared and visible images from $S$ , $\lambda_{g}$ is a hyper-parameter to balance $L_{f}$ and $L_{g}$ , and $\beta_{F}$ is the learning rate of $F$ . + +Finally, the inner update stage and outer update stage are carried out alternately for every $N$ epochs, which improves the optimization of $MFG$ and $F$ . + +![](images/0ce69d4d5025548242fcc713ea43b1a5aa3a141ca0f701fe1e91fc1559f44905.jpg) +Figure 4. Illustration of the mutual promotion learning between $F$ and $D$ . Firstly, $F$ is used to generate fusion results to optimize $D$ . Thus, a fine-tuned $D$ with higher semantic extracting ability is obtained. Then, $D$ is adopted to optimize $F$ . This process is alternated. + +# 3.2. Mutual Promotion Learning + +In Sec. 3.1, we fix the object detection network $D$ to offer stable object semantic features. In this way, $MFG$ does not need to handle variable object semantic features, thereby reducing the difficulty of training $MFG$ . However, the fixed $D$ will restrict the object semantic extracting ability of $F$ . Addressing this problem, we propose the mutual promotion learning that promotes not only $F$ but also $D$ , as shown in Figure 4. Specifically, we firstly train $F$ individually, and then use the fusion results to train $D$ . Sequentially, $D$ offers object semantic features to the meta-feature embedding, thereby improving $F$ . After that, the improved fusion results generated by $F$ are used to fine-tune $D$ . In this way, both $F$ and $D$ are improved. + +# 3.3. Architecture + +IVIF network $F$ aims to generate fusion results according to the input infrared and visible images, which is flexible. Here, we adopt the similar structure with [50]. $F$ consists of three feature fusion blocks $(FFB_{j}, j = 1,2,3)$ and one image reconstruction module (IRM), as shown in Figure 2. $FFB_{j}$ is used to extract and fuse features from input infrared image $I_{i}$ and visible image $I_{v}$ . It is designed based on the idea of [50] that if $FFB_{j}$ can well fuse the features of $I_{i}$ and $I_{v}$ , then the two kinds of features can be recovered from the fused feature. Thus, $FFB_{j}$ is described as + +$$ +F F B _ {j} = \left\{ \begin{array}{c} C _ {2} T _ {1} \biguplus C _ {2} T _ {1}, j = 1 \\ C _ {2} T _ {j - 1} \biguplus C _ {2} T _ {j - 1}, j > 1 \end{array} \right. \tag {7} +$$ + +where $j$ is the block index, $C_k$ represents the number of "Convolution with $3 \times 3$ kernel +ReLU" layers is $k$ , $F_{u(j-1)}$ is the output feature of the $j-1$ th block, $T_1 = C_2(I_i, I_v)$ and $T_{j-1} = C_2(F_{u(j-1)})$ represent feature integration operations, and $\boxplus$ means feature concatenation. IRM reconstructs the fusion result, which contains six "Convolution with $3 \times 3$ kernel +ReLU" layers. + +Object detection network $D$ offers object semantic features. In our framework, we choose $\mathrm{yolov5s}^1$ as our $D$ . According to the size of the feature maps, the backbone of $D$ is divided into three feature extraction blocks $(\{DFB_{i}, i = 1,2,3\})$ . In addition, we represent the neck and detection head in yolov5s into one block $(DH)$ to simplify the expression. + +Meta-feature embedding network $MFE$ includes metafeature generator $MFG_{j}$ and feature transform network $FT_{j}$ . Since OD feature $F_{ej}$ generated by $DFB_{j}$ offers object semantic information and $F_{uj}$ provides scene details, they are fed into the $MFG_{j}$ . Then, $MFG_{j}$ is built as + +$$ +M F G _ {j} = C _ {6} \left(C _ {4} \left(F _ {u j}\right) \biguplus C _ {2} \left(U p \left(F _ {e j}\right)\right)\right), \tag {8} +$$ + +where $Up()$ is the feature up-sampling. $FT_{j}$ transfers the meta feature to fusion feature, which includes three "Convolution with $3 \times 3$ kernel +ReLU" layers. + +# 3.4. Training + +The training details of our framework is shown in Algorithm 1, which contains four steps. Generally, we firstly pretrain $F$ , and then use the fusion results to finetune $D$ . Sequentially, we train $FT$ and $MFG$ , where $MFG$ and $FT$ are updated $n = 200$ times every $N = 8$ epochs. Lastly, the mutual promotion between $F$ and $D$ is trained for $R = 2$ rounds, thereby improving their performances. + +# 4. Experiments + +# 4.1. Setup + +Dataset. We conduct experiment on three widely-used datasets: $\mathbf{M}^3\mathrm{FD}$ [20], RoadScene [44] and TNO [37]. $\mathbf{M}^3\mathrm{FD}$ is divided as training set (2940 image pairs) and testing set (1260 image pairs). RoadScene with 221 image pairs and TNO with 40 image pairs are only used for testing. Besides, $\mathbf{M}^3\mathrm{FD}$ is adopted to evaluate OD performance. + +Implementation. Our framework is implemented with PyTorch on a NVIDIA GeForce RTX 3090 GPU. $F$ and MFE are trained using optimizer Adam with the learning rate $\beta_{F},\beta_{F^{\prime}},\beta_{FT}$ and $\beta_{MFG}$ of $1\times 10^{-3}$ , respectively. $D$ is trained followed by yolov5s that adopts the optimizer SGD using the learning rate of $1\times 10^{-2}$ with decaying rate of 0.1 every round. We firstly train $F$ for 100 epochs and $D$ for 150 epochs. Then, we conduct MFE for 50 epochs. After + +Algorithm 1 MetaFusion Training Algorithm +Input: Training dataset $S = \{S_{mtr}, S_{mts}\}$ , IVIF model $F(\theta_F)$ , meta-feature generator $MFG(\theta_F)$ , feature transform network $FT(\theta_{FT})$ , OD network $D(\theta_D)$ +Output: IVIF model $F(\theta_F^*)$ +1: Initialize $F(\theta_F)$ , $MFG(\theta_F)$ , $FT(\theta_{FT})$ , $D(\theta_D)$ +2: /* Pretrain $F^{*}/$ +3: while not converged do +4: Sample image pair $(I_i, I_v)$ from $S$ +5: Forward pass on $F$ to generate fusion result $I_f$ +6: Optimize $F$ using fusion loss $L_f$ +7: end while +8: /* Pretrain $D^{*}/$ +9: while not converged do +10: Sample image $I_f$ from the fused images generated by $F$ +11: Forward pass on $D$ to get detection result +12: Optimize $D$ using detection loss $L_d$ +13: end while +14: /* Meta-feature embedding outer update */ +15: while not converged do +16: Sample a batch image pair $(I_i, I_v)$ from $S$ +17: Forward pass on $F$ to generate fusion result $I_f$ +18: Forward pass on $MFG$ to generate meta feature +19: Optimize $F$ using fusion loss $L_f$ and guide loss $L_g$ by Eq. 6 +20: /* Meta-feature embedding inner update */ +21: if epoch%N==0 then +22: for t=1 to n do +23: Sample an image pair $(I_i^{mtr}, I_v^{mtr})$ from $S_{mtr}$ +24: Compute $\theta_{F'}$ by Eq. 1 +25: Sample an image pair $(I_i^{mts}, I_v^{mts})$ from $S_{mts}$ +26: Calculate $L_f$ using $\theta_{F'}$ +27: Update $\theta_{MFG}$ and $\theta_{FT}$ using the gradient of $L_f$ by Eqs. 2 and 3 +28: end for +29: end if +30: end while +31: /* Mutual promotion*/ +32: if Number of round ≤ R then +33: Go back to line 9 +34: end if + +that, $D$ is finetuned for 150 epochs. Following the same strategy, the mutual promotion is carried out $R$ rounds. The images are resized to $512 \times 384$ with batchsize of 1. The hyperparameter $\lambda_{g}$ is set to 0.1. + +Metric. Three metrics are used for IVIF evaluation: entropy (EN) [33], mutual information (MI) [28] and visual information fidelity (VIF) [7]. EN evaluates the information richness in an image, and the higher EN means more information. MI evaluates the information similarity between + +![](images/ad293010675822b131f424cffc076bde8fd54c2a123d2075e4dbce42638ad3ab.jpg) +(a) Visible-infrared images + +![](images/5d84997b06a4f246fb6b57071c6fe16f6d4755bcd016e6e8d08defa9513f78a5.jpg) +(b) SoFusion and feature map + +![](images/9391a83037ed5dc068fd2a84eaff4e7bc2f11318622a7a96427f068c312ef89c.jpg) +(c) CoFusion and feature map +Figure 5. Visual comparison of different optimization methods. + +![](images/fc40a81a98b73fc408062d9cc005fd1b877e9acc7ee712b0e4d50c63f03f89f2.jpg) +(d) MetaFusion and feature map + +Table 1. Effect study of meta-feature embedding by comparing with different optimization methods on $\mathrm{M}^3\mathrm{FD}$ . The best result is in red. + +
MethodMetric
MIENVIF
SoFusion14.1647.0781.190
CoFusion14.1877.0891.345
MetaFusion14.5117.2491.515
+ +the input images and fused image. The higher MI illustrates more information of the input images is fused. VIF measures the ability to extract visible information from the input image, and a larger VIF represents less visible distortion in the fused result. Here, we use the V channel in HSV space of the fusion results to calculate these metrics. Moreover, we use $\mathrm{mAP}_{50\rightarrow 95}$ [9] to comprehensively evaluate OD performance, where the average of mAPs sampling every 5 from $\mathrm{AP}_{50}$ to $\mathrm{AP}_{95}$ is calculated. A higher $\mathrm{mAP}_{50\rightarrow 95}$ means better OD effect. + +# 4.2. Ablation Study + +Effect of meta-feature embedding. In Sec. 3.1, we introduce the meta-feature embedding that generates the meta features from object detection features to help IVIF network learn more object semantic information. To verify its effect, separate optimization method (SoFusion) and cascaded optimization method (CoFusion) are compared, as shown in Figure 1(a)-(b). Specifically, SoFusion firstly optimizes $F$ and then uses the generated fusion results to optimize $D$ . CoFusion treats $D$ as a constraint to optimize $F$ with the loss $L_{f}$ and $L_{d}$ . The results are shown in Table 1. Our MetaFusion achieves the best results on IVIF. The reason is that SoFusion neglects the help of $D$ , and directly utilizing the high-level $D$ to guide the pixel-level $F$ has feature mismatch in CoFusion. In contrast, we implement the meta-feature embedding that generates meta features from $D$ according to the ability of $F$ . Thus, the meta features are naturally compatible with $F$ , thereby providing effective object semantic features to $F$ . Visual comparison is shown in Figure 5. MetaFusion highlights the object feature and generates the clearest result. + +Influence of mutual promotion learning. As described + +Table 2. Influence study of mutual promotion learning by evaluating the performances of round $R$ on ${\mathrm{M}}^{3}\mathrm{{FD}}$ . The best result is in red. + +
MethodMetric
MIENVIFmAP50→95 (%)
R = 014.1647.0781.19055.6
R = 114.4647.2261.47455.8
R = 214.5117.2491.51556.5
+ +Table 3. Study of multi-level meta-feature embedding by comparing different configurations on $\mathbf{M}^3\mathrm{FD}$ . The best result is in red. + +
MethodMetric
MIENVIF
MetaFusion-L114.4507.2201.259
MetaFusion-L214.4527.2201.377
MetaFusion-L314.4647.2261.474
+ +in Sec. 3.2, we propose the mutual promotion learning promote $F$ and $D$ alternately. Here, we study the influence of mutual promotion learning by evaluating round $R = 0,1,2$ . In particular, $R = 0$ represents $F$ and $D$ are trained respectively. As shown in Table 2, the results of $F$ are better with the $R$ increases. Comprehensively considering training efficiency and performance, we take $R = 2$ . + +Study of multi-level meta-feature embedding. We implement multi-level meta-feature embedding to build MetaFusion framework, as introduced in Sec. 3.3. Here, we study the number of meta-feature embedding levels with the following configurations. One-level meta-feature embedding (MetaFusion-L1) with $MFE = \{MFG_1, FT_1\}$ , $F = \{FFB_1, IRM\}$ and $D = \{DFB_1, DH\}$ . Two-level meta-feature embedding (MetaFusion-L2) with $MFE = \{MFG_1, MFG_2, FT_1, FT_2\}$ , $F = \{FFB_1, FFB_2, IRM\}$ and $D = \{DFB_1, DFB_2, DH\}$ . Three-level meta-feature embedding (MetaFusion-L3) with $MFE = \{MFG_1, MFG_2, MFG_3, FT_1, FT_2, FT_3\}$ , $F = \{FFB_1, FFB_2, FFB_3, IRM\}$ and $D = \{DFB_1, DFB_2, DFB_3, DH\}$ . The results are shown in Table 3. With the number of multi-level meta-feature embedding increases, $F$ achieves higher image fusion quality. Since meta-feature embedding provide multi-level object semantic features to $F$ . + +# 4.3. Comparison with State-of-the-art Methods + +We compare the proposed MetaFusion with eight SOTA fusion methods to verify the superiority: FusionGAN [25], GANMcC [26], MFEIF [21], YDTR [36], PIAFusion [35], SwinFusion [23], Tardal [20], and U2Fusion [44]. Their available codes and recommended parameter settings are + +![](images/a6948d69ed6adca343da4e9b96dfa341d6031457013efd9491b1519cbed6cd5f.jpg) +Infrared-visible image + +![](images/88508f5965299148b31b56b5d6d3f8c58eccf04a84d59cd27c19d62f09d6a498.jpg) +FusionGAN + +![](images/a54a70e631bac6372eb2f5827c768b359dd40a4dbfaef7a3813ca965493cef89.jpg) + +![](images/f2e7f06cb59fbf4dbb9bac8f4ba5f94c633b094f1c20b5b14d5c0d20c8a33f35.jpg) + +![](images/852d3dd4c719b3eb05074b4b0c4cc075b771b2b9bf705c2505aa8b8684eb302c.jpg) + +![](images/d8e91f56e3c7b9b817435a0c052e9a751ce44fb42e8826395cbe8dda7b5db68a.jpg) + +![](images/aa2151fce9011d3a1ee0802be7a594fc55d8654e144fa0861e7d1213147c3688.jpg) + +![](images/20db97aad59377cb1f63ca37a2fdc03c755daf3a6af2006e0c88a40878e21656.jpg) +GANMcC + +![](images/0e1cb7d4873a1eb56a6e1195c24bcd75a85a7eba53fa26479e755d62dc7b202e.jpg) +MFEIF + +![](images/03ba215c1803ec5764514a99aa83c4471b5ca4e85616aeab5171f5cf599a4f0c.jpg) +U2Fusion + +![](images/cb4d87835cb6a47d9194550c743d2137e6724038f9993d1310cc9a2a162b2fe4.jpg) +YDTR + +![](images/5656eb01f22b5676b841337b6986b965402f9f920406bf6f622c74d46f004dd3.jpg) +PIAFusion +FusionGAN + +![](images/15b224bf6ae870c6265ec3f4032a5103d04a7cd5a83a5673af802f6bce405637.jpg) +SwinFusion + +![](images/9fbb470afbc421d16b4a4621902e8d60da426399a06bdc68223be9a3c6acd569.jpg) +Tardal +MFEIF + +![](images/515fcbf6d7e2e13c469ab0e3c6b4617ffea677ca8b2a02d3b62959ea945f9458.jpg) +MetaFusion + +![](images/d36d1aae1d2ef1013486fc2f1c9d43f9d4ae16d6356d7f4fa2b7f52773e420fa.jpg) +Infrared-visible image +YDTR + +![](images/1ad257a316b584678fb08417affd4db817082088d0b0249cf87cccd64845a3ab.jpg) + +![](images/8285b498c10e22dfba2c6eedc0e914b8620db201cca92225e008313cbb3740e8.jpg) +GANMcC + +![](images/658322fc0e3150ddd744a4e7a8209033695ffad91c723aacab6174198aafcadb.jpg) + +![](images/4745ce2a8ff41e1922328a7877ce1df3248ff8f67452eb1dcd6bcdb29bbbbc4d.jpg) +U2Fusion + +![](images/d76b022c349acfc86c866944a2e04bafd3fb7c3e23db1ea35750c871a1a3ee53.jpg) +Infrared-visible image + +![](images/995e45e504a43da9738de81ddd817dbb0d2a9253221afe2754e25e0d31ee27b0.jpg) +PIAFusion +FusionGAN + +![](images/c564e8d7527f75731c5f671c6be12c2194ab566ce1646802ac0c70841819cae5.jpg) +SwinFusion +GANMcC + +![](images/72cc7c78d62de00a53c2e4651e1ca4f4285985d1f1739a464ec971318f4033d5.jpg) +Tardal +MFEIF + +![](images/341595a9f2a5131a9d527e2c5eede76f72a6fa7647935fa5594162b178e2c52c.jpg) +MetaFusion +U2Fusion + +![](images/d6cb12c5541ad10eafda0fc1f0cff1d005ec662d214ed311e581e5489272e1dd.jpg) +YDTR +Figure 6. Qualitative results of different fusion methods on $\mathbf{M}^3\mathrm{FD}$ (first group), RoadScene (second group) and TNO (third group). + +![](images/924e4f0a226b7a4935fb85fedb606e19bb232cd9ee7c74cb84ebf060146d90be.jpg) +PIAFusion + +![](images/691b89cf755bf1359f60d8cefd5c3d54770854d45b616b9767bb33f07867267e.jpg) +SwinFusion + +![](images/6bc5db959b22a9872bcb6f684e4368999ba3a587ec0f7313aeeb397b62f5964b.jpg) +Tardal + +![](images/685a7fef49377996e8219b74c7cfedb1be499bbbcb1a73d9b491d15de8cc7c2b.jpg) +MetaFusion + +adopted to generate fusion results for the fair comparison. + +Qualitative results of different fusion methods are shown in Figure 6. All the fusion methods can fuse the main features of the infrared and visible images to some extent. However, FusionGAN, GANMcC, MFEIF and YDTR produce blurred edge details, and U2Fusion, PIAFusion, Swinfusion and Tardal generate low contrast objects, as shown in the red rectangular boxes. In contrast, the images gener + +ated by the proposed MetaFusion contain clear edge details and high contrast objects. + +Sequentially, we provide quantitative results of different fusion methods in Table 4. Our MetaFusion generally achieves the largest or the second-largest metric values. In detail, the high EN and MI illustrate the fusion images by our MetaFusion contain high contrast object and clear edge details. The large VIF shows our fusion results have high- + +Table 4. Quantitative results of different fusion methods on TNO, RoadScene and $\mathbf{M}^3\mathrm{FD}$ datasets. The model inference time is counted on a NVIDIA GeForce RTX 2080 Ti. The best result is in red and the second best one is in violet. + +
MethodM3FDRoadSceneTNO40Time (s)
MIEntVIFMIEntVIFMIEntVIF
FusionGAN [25]13.4456.7220.30314.2037.1010.25113.0686.5340.2520.040
GANMcC [26]13.7316.8650.45314.0177.0080.42213.4856.7420.4240.081
MFEIF [21]12.9576.4780.40113.4896.7420.26013.3606.6800.3950.029
U2Fusion [44]13.8166.9080.54513.8886.9440.45613.8896.9440.6360.043
YDTR [36]12.6946.3470.36113.1666.5830.23612.8626.4310.2800.086
PIAFusion [35]13.3266.6630.43713.2626.5330.20513.9316.9600.5060.055
SwinFusion [23]13.0516.5250.46013.2676.6330.31313.8796.9330.4791.081
Tardal [20]13.6366.8180.65015.0097.5040.48313.0306.5150.9100.030
MetaFusion14.5117.2491.51514.2187.0220.96914.6577.3231.4620.015
+ +![](images/fdcd42e5b93865b1136dd887f9ab32722b6d75e9c7be68926411453fa17f5d3a.jpg) + +![](images/c22264dbdd5ab373084f0a61ed0876b143d758176fdc5aad3b9e879640f97567.jpg) + +![](images/eb61026d2b5b36f7323ce85006915eae8ed2e73cea3ebe1e1e5353ddc7c62074.jpg) + +![](images/9b1f7750091301e96c5ff96b59ba8243280f6a92a8868adf0783f3bbe6057ca1.jpg) + +![](images/7dbcd590492a89d29c05ed3e2ae000c8d379d9e62ccaafd73fa44c3eed1c48c1.jpg) + +![](images/fa46622c4c836260e003e102d1cbe4cb775ce2450a09639e2d66e7107a073779.jpg) + +![](images/feccfbfe8c5ddbf158ec44996b51a076155566fae41d1f2f167bf02f8769a664.jpg) + +![](images/a70e72e5c666c9755bb36dd2b788b5c04598f690e5b5c79a850ab0eedb68e27c.jpg) + +![](images/e9454166dc21ce7a57c98c07515a73a40df63984fc7542c60c469158e2dc28f3.jpg) +SwinFusion + +![](images/9d5c4172d94e9654594d1425b91a4acddc3c2237a424f62aabcf6eb63e01bce3.jpg) +Tardal + +![](images/a14122c348be79ee8a9a3f7a31e946ec3789bb650e339f0ec8cff17e3f314c82.jpg) +MetaFusion +Figure 7. Visual results of object detection based on different fusion methods on $\mathbf{M}^3\mathrm{FD}$ + +![](images/a31329ecfc5325b98c81c63f5f364718a68cb1bbffd94430a92dc9cc429a3fbe.jpg) +Ground truth + +quality visual effect and small distortion compared with the source images. Moreover, our method achieves the fastest inference time that only needs 0.015s to generate one fusion result. + +# 4.4. Evaluation on Infrared-visible Object Detection + +A better fusion image can provide more information to improve object detection. Here, we verify the effectiveness of our method by evaluating object detection accuracy with the infrared and visible fusion images. To make a fair comparison, we firstly generate fusion results using our method and SOTA methods, and then use the results to retrain the object detection baseline YOLOv5s, respectively. + +Figure 7 shows the object detection results. In general, the SOTA methods improve the performance of object detection. In contrast, our MetaFusion achieves a better performance, e.g., the person is accurately detected (see the blue ellipse). Table 5 shows the quantitative results. Our fusion results help the detection network achieve the highest object detection accuracy. This further proves that our + +Table 5. Quantitative results of object detection based on different fusion methods on $\mathbf{M}^3\mathrm{FD}$ dataset. The best result is in red and the second best one is in violet. + +
Method\( \mathrm{mAP}_{50\rightarrow 95}(\%) \)
FusionGAN [25]54.2
GANMcC [26]55.2
MFEIF [21]55.4
U2Fusion [44]55.7
YDTR [36]55.4
PIAFusion [35]55.6
SwinFusion [23]55.4
Tardal [20]54.4
MetaFusion56.5
+ +method can generate high-quality fusion results especially for the objects. + +# 5. Conclusion + +This paper presents a joint fusion and detection learning framework through introducing the meta-feature embedding model. Based on the meta learning idea, the meta-feature embedding model can generate object semantic features according to the fusion network ability, thereby bridging the feature gap between these two different-level tasks. Moreover, a mutual promotion learning between fusion and detection tasks is further implemented to improve their performances. Both quantitative and qualitative results demonstrate the superior performance of our method compared with the state-of-the-art methods. + +Acknowledgement. This work is supported by National Natural Science Foundation of China under Grant Nos. 62176038, 62001450 and U1903215. + +# References + +[1] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9465-9474, October 2021. 3 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6154-6162, 2018. 2 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. 2 +[4] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-sclil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14166-14175, June 2022. 3 +[5] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 3 +[6] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster R-CNN: towards accurate few-shot object detection with attentive feature alignment. In AAAI, pages 780-789. AAAI Press, 2022. 3 +[7] Yu Han, Yunze Cai, Yin Cao, and Xiaoming Xu. A new image fusion performance metric based on visual information fidelity. Information Fusion, 14:-, 04 2013. 5 +[8] Kaiming He, Georgia Gkioxari, Piotr Dollr, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 2 +[9] Liqiang He and Sinisa Todorovic. Destr: Object detection with split transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9367-9376, 2022. 6 +[10] Ruichao Hou, Dongming Zhou, Rencan Nie, Dong Liu, Lei Xiong, Yanbu Guo, and Chuanbo Yu. Vif-net: An unsupervised framework for infrared and visible image fusion. IEEE Transactions on Computational Imaging, 6:640-651, 2020. 2 +[11] Muhammad Abdullah Jamal, Liqiang Wang, and Boqing Gong. A lazy approach to long-horizon gradient-based meta-learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6577-6586, October 2021. 3 +[12] Huafeng Li, Yueliang Cen, Yu Liu, Xun Chen, and Zhengtao Yu. Different input resolutions and arbitrary output resolution: A meta learning-based deep framework for infrared and visible image fusion. IEEE Transactions on Image Processing, 30:4070-4083, 2021. 3 +[13] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 2 + +[14] Hui Li, Xiao-Jun Wu, and Josef Kittler. Mdlatrr: A novel decomposition method for infrared and visible image fusion. IEEE Transactions on Image Processing, 29:4733-4746, 2020. 2 +[15] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 1 +[16] Jing Li, Jianming Zhu, Chang Li, Xun Chen, and Bin Yang. Cgtf: Convolution-guided transformer for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 71:1-14, 2022. 2 +[17] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 3 +[18] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP:1-1, 07 2018. 2 +[19] Benlin Liu, Yongming Rao, Jiwen Lu, Jie Zhou, and Chojui Hsieh. Metadistiller: Network self-boosting via metalearned top-down distillation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16, pages 694–709. Springer, 2020. 3 +[20] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 1, 2, 5, 6, 8 +[21] J. Liu, X. Fan, J. Jiang, R. Liu, and Z. Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1, 2021. 6, 8 +[22] Yu Liu, Shuping Liu, and Zengfu Wang. A general framework for image fusion based on multi-scale transform and s-parse representation. Information Fusion, 24:147-164, 2015. 2 +[23] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 2, 6, 8 +[24] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping Zhang. Ddcgan: A dual-discriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29:4980-4995, 2020. 2 +[25] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusion: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11-26, 08 2019. 1, 6, 8 +[26] Jiayi Ma, Hao Zhang, Zhenfeng Shao, Pengwei Liang, and Han Xu. Ganmcc: A generative adversarial network with + +multiclassification constraints for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1-14, 2021. 2, 6, 8 +[27] Yan Mo, Xudong Kang, Puhong Duan, Bin Sun, and Shutao Li. Attribute filter based infrared and visible image fusion. Information Fusion, 75:41-54, 2021. 2 +[28] Guihong Qu, Dali Zhang, and Pingfan Yan. Information measure for performance of image fusion. *Electronics Letters*, 38:313-315, 04 2002. 5 +[29] Aniruddh Raghu, Jonathan Lorraine, Simon Kornblith, Matthew McDermott, and David K Duvenaud. Meta-learning to improve pre-training. Advances in Neural Information Processing Systems, 34:23231-23244, 2021. 3 +[30] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 2 +[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 2 +[32] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 658-666, 2019. 2 +[33] Wesley Roberts, Jan van Aardt, and Fethi Ahmed. Assessment of image fusion procedures using entropy, image quality, and multispectral classification. Journal of Applied Remote Sensing, 2:1-28, 05 2008. 5 +[34] Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshitaka Ushiku, and Tatsuya Harada. Multispectral object detection for autonomous vehicles. In Proceedings of the on Thematic Workshops of ACM Multimedia 2017, Thematic Workshops '17, page 3543, New York, NY, USA, 2017. Association for Computing Machinery. 1 +[35] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83-84:79-92, 2022. 1, 2, 6, 8 +[36] Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer. IEEE Transactions on Multimedia, pages 1-16, 2022. 2, 6, 8 +[37] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249-251, 2017. 5 +[38] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 4 +[39] Zhishe Wang, Yuanyuan Wu, Junyao Wang, Jiawei Xu, and Wenyu Shao. Res2fusion: Infrared and visible image fusion based on dense res2net and double nonlocal attention models. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 2 +[40] Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14248-14257, June 2022. 2 +[41] Chengming Xu, Yanwei Fu, Chen Liu, Chengjie Wang, Jilin Li, Feiyue Huang, Li Zhang, and Xiangyang Xue. Learning dynamic alignment via meta-filter for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5182-5191, June 2021. 3 +[42] Dongli Xu, Jinghong Deng, and Wen Li. Revisiting ap loss for dense object detection: Adaptive ranking pair selection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 14187-14196, 2022. 2 +[43] Han Xu, Meiqi Gong, Xin Tian, Jun Huang, and Jiayi Ma. Cufd: An encoderdecoder network for visible and infrared image fusion based on common and unique feature decomposition. Computer Vision and Image Understanding, 218:103407, 03 2022. 1 +[44] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 1, 2, 5, 6, 8 +[45] Han Xu, Xinya Wang, and Jiayi Ma. Drf: Disentangled representation for visible and infrared image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1-13, 2021. 1, 2 +[46] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Rui-Wei Zhao, Tao Zhang, Xuequan Lu, and Shang Gao. Cream: Weakly supervised object localization via class re-activation mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9437–9446, June 2022. 2 +[47] Meilong Xu, Linfeng Tang, Hao Zhang, and Jiayi Ma. Infrared and visible image fusion via parallel scene and texture learning. Pattern Recognition, 132:108929, 2022. 1 +[48] Shilong Zhang, Zhuoran Yu, Liyang Liu, Xinjiang Wang, Aojun Zhou, and Kai Chen. Group r-cnn for weakly semi-supervised object detection with points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9417-9426, June 2022. 2 +[49] Fan Zhao and Wenda Zhao. Learning specific and general realm feature representations for image fusion. IEEE Transactions on Multimedia, 23:2745-2756, 2020. 2 +[50] Fan Zhao, Wenda Zhao, and Hutchuan Lu. Interactive feature embedding for infrared and visible image fusion. arXiv preprint arXiv:2211.04877, 2022. 2, 4 +[51] Fan Zhao, Wenda Zhao, Libo Yao, and Yu Liu. Self-supervised feature adaption for infrared and visible image fusion. Information Fusion, 76:189-203, 2021. 2 +[52] Wenda Zhao, Xueqing Hou, You He, and Huchuan Lu. Defocus blur detection via boosting diversity of deep ensemble networks. IEEE Transactions on Image Processing, 30:5426-5438, 2021. 2 +[53] Wenda Zhao, Huimin Lu, and Dong Wang. Multisensor image fusion and enhancement in spectral total variation domain. IEEE Transactions on Multimedia, 20(4):866-879, 2018. 2 + +[54] Wenda Zhao, Dong Wang, and Huchuan Lu. Multi-focus image fusion with a natural enhancement via a joint multi-level deeply supervised convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 29(4):1102-1115, 2018. 2 +[55] Wenda Zhao, Bowen Zheng, Qiuhua Lin, and Hutchuan Lu. Enhancing diversity of defocus blur detectors via cross-ensemble network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8905–8913, 2019. 2 +[56] Zhengjie Zhu, Xiaogang Yang, Ruitao Lu, Tong Shen, Xueli Xie, and Tao Zhang. Clf-net: Contrastive learning for infrared and visible image fusion network. IEEE Transactions on Instrumentation and Measurement, 71:1-15, 2022. 2 \ No newline at end of file diff --git a/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/images.zip b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b2efdfbdb779567a12e90a5c5382760f9c4d556 --- /dev/null +++ b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3628b9d7e8943327184f56cda2578b7fb4b910d07d5838a2560bad1c33fdf3d +size 773215 diff --git a/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/layout.json b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d66aa1acd6c3a2e52e00f69643c706e91958dd34 --- /dev/null +++ b/2023/MetaFusion_ Infrared and Visible Image Fusion via Meta-Feature Embedding From Object Detection/layout.json @@ -0,0 +1,16097 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "type": "text", + "content": "MetaFusion: Infrared and Visible Image Fusion via Meta-Feature Embedding from Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "text", + "content": "Wenda Zhao" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "text", + "content": ", Shigeng Xie" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "text", + "content": ", Fan Zhao" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "text", + "content": ", You He" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "text", + "content": ", Huchuan Lu" + }, + { + "bbox": [ + 135, + 160, + 457, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 197, + 175, + 397, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 175, + 397, + 189 + ], + "spans": [ + { + "bbox": [ + 197, + 175, + 397, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 175, + 397, + 189 + ], + "type": "text", + "content": " Dalian University of Technology, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 189, + 386, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 189, + 386, + 202 + ], + "spans": [ + { + "bbox": [ + 207, + 189, + 386, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 207, + 189, + 386, + 202 + ], + "type": "text", + "content": " Liaoning Normal University, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 203, + 367, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 203, + 367, + 217 + ], + "spans": [ + { + "bbox": [ + 225, + 203, + 367, + 217 + ], + "type": "text", + "content": "3 Tsinghua University, China" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 161, + 220, + 432, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 220, + 432, + 231 + ], + "spans": [ + { + "bbox": [ + 161, + 220, + 432, + 231 + ], + "type": "text", + "content": "zhaowenda@dlut.edu.cn; xieshigeng@mail.dlut.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 140, + 234, + 450, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 234, + 450, + 244 + ], + "spans": [ + { + "bbox": [ + 140, + 234, + 450, + 244 + ], + "type": "text", + "content": "Fan-Zhao20@163.com; youhe_nau@163.com; lhchuan@dlut.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 298, + 290, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 298, + 290, + 526 + ], + "spans": [ + { + "bbox": [ + 45, + 298, + 290, + 526 + ], + "type": "text", + "content": "Fusing infrared and visible images can provide more texture details for subsequent object detection task. Conversely, detection task furnishes object semantic information to improve the infrared and visible image fusion. Thus, a joint fusion and detection learning to use their mutual promotion is attracting more attention. However, the feature gap between these two different-level tasks hinders the progress. Addressing this issue, this paper proposes an infrared and visible image fusion via meta-feature embedding from object detection. The core idea is that meta-feature embedding model is designed to generate object semantic features according to fusion network ability, and thus the semantic features are naturally compatible with fusion features. It is optimized by simulating a meta learning. Moreover, we further implement a mutual promotion learning between fusion and detection tasks to improve their performances. Comprehensive experiments on three public datasets demonstrate the effectiveness of our method. Code and model are available at: https://github.com/wdzhao123/MetaFusion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "content": "Multi-modality sensor technology has promoted the application of multi-modality images in different areas. Among them, infrared images and visible images have been utilized commonly, as the information contained in these two modalities is complementary. Specifically, infrared images can supply object thermal structures without being affected by illumination. But they are short of texture details. On the contrary, visible images can catch the texture information for the scene. But they are severely affected by light. Thus, many methods [15,25,35,43-45,47] focus on s" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 273, + 542, + 327 + ], + "blocks": [ + { + "bbox": [ + 316, + 273, + 542, + 327 + ], + "lines": [ + { + "bbox": [ + 316, + 273, + 542, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 542, + 327 + ], + "type": "image", + "image_path": "a829f059b0b06fa7ccf0112ee4701d81542d6560ff3a0acd00647ef62525dea4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 331, + 490, + 340 + ], + "lines": [ + { + "bbox": [ + 384, + 331, + 490, + 340 + ], + "spans": [ + { + "bbox": [ + 384, + 331, + 490, + 340 + ], + "type": "text", + "content": "(a) Separate optimization method" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 344, + 543, + 397 + ], + "blocks": [ + { + "bbox": [ + 315, + 344, + 543, + 397 + ], + "lines": [ + { + "bbox": [ + 315, + 344, + 543, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 344, + 543, + 397 + ], + "type": "image", + "image_path": "8e2b9f5f41446252228b71c0564d6a3d37cf305c7addcac4d9a0d9e12f1c36f1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 399, + 494, + 407 + ], + "lines": [ + { + "bbox": [ + 387, + 399, + 494, + 407 + ], + "spans": [ + { + "bbox": [ + 387, + 399, + 494, + 407 + ], + "type": "text", + "content": "(b) Cascaded optimization method" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 313, + 411, + 542, + 495 + ], + "blocks": [ + { + "bbox": [ + 313, + 411, + 542, + 495 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 542, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 542, + 495 + ], + "type": "image", + "image_path": "7d75aaed38dfc8eca509b018fd0184e39f1718fff02a6b437df2ce7e876fb386.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 497, + 493, + 506 + ], + "lines": [ + { + "bbox": [ + 381, + 497, + 493, + 506 + ], + "spans": [ + { + "bbox": [ + 381, + 497, + 493, + 506 + ], + "type": "text", + "content": "(c) Meta-feature embedding method" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": "Figure 1. Different joint learning methods of infrared and visible image fusion (IVIF) and object detection (OD). (a) Separate optimization method: IVIF network " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " is firstly optimized by fusion loss " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": ". Then, fusion result " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " is generated by " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " from the input infrared and visible image pair " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "x,y" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": ". Finally, OD network " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " is optimized by detection loss " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": ". (b) Cascaded optimization method: OD network " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " is treated as a constraint to optimize IVIF network " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " by the loss " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": ". (c) Meta-feature embedding method: Meta-feature embedding " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " is optimized to learn how to guide " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " to have a low " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " generates meta feature from detection feature " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "F_{i}^{d}" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " and fusion feature " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "F_{i}^{d}" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": ". Finally, the meta feature is used to guide " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": " by the loss " + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 304, + 514, + 547, + 646 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 662, + 546, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 546, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 546, + 698 + ], + "type": "text", + "content": "tudying pixel-level infrared and visible image fusion (IVIF), thereby helping high-level tasks improve performance, e.g., object detection (OD) [20, 34]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 701, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 546, + 713 + ], + "type": "text", + "content": "IVIF and OD can greatly benefit form each other. IVIF" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13955" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": "generates the fused image that contains more information than any single modality image to improve OD. However, IVIF mainly focuses on the pixel relationship between an image pair, and there is little consideration for object semantic. In contrast, OD can provide rich object semantic information to IVIF, as its aim is locating the objects. Therefore, this paper studies a joint learning framework between IVIF and OD to improve their performances." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 288, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 288, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 288, + 447 + ], + "type": "text", + "content": "Existing joint learning methods of IVIF and OD can be divided into two categories: separate optimization and cascaded optimization. Separate optimization firstly trains IVIF network, and then trains OD network using IVIF results, as shown in Figure 1(a). Thus, most methods focus on improving the fusion effect, e.g., designing networks [13, 23, 35, 36] and introducing constraints [10, 26, 45]. Obviously, separate optimization neglects the help of OD. Cascaded optimization adopts OD network as a constraint to train IVIF network, and thus forces the IVIF network to generate fusion images with easily detected objects [20], as shown in Figure 1(b). However, directly utilizing the high-level OD constraint to guide the pixel-level IVIF will result in limited effect. Therefore, we leverage OD feature maps that guide IVIF feature maps to obtain more semantic information. Unfortunately, OD features are mismatched with IVIF features due to their task-level difference. Addressing this issue, we propose a meta-feature embedding network (MFE), as shown in Figure 1(c). The idea is that if MFE generates OD features according to the IVIF network ability, the OD features are naturally compatible with the IVIF network, and the optimization can be achieved by simulating a meta learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": "Specifically, an infrared and visible image fusion via meta-feature embedding from object detection is proposed, which is named as MetaFusion. MetaFusion includes IVIF network " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ", OD network " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "MFE" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ". In particular, " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "MFE" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " is expected to generate meta features to bridge the gap between " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ", which is optimized by two alternate steps: inner update and outer update. In the inner update process, we firstly optimize " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " using meta training set " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "S_{mtr}" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " to obtain its updated network " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " calculates the fusion loss on meta testing set " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "S_{mts}" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " to optimize " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "MFE" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ". The motivation is that if " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "MFE" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " successfully generates meta features which are compatible with " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " will produce better fused images, i.e., the fusion loss should be lower. In the outer update process, " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " is optimized with the guide of the meta features generated by the fixed " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "MFE" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "S_{mtr}" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "S_{mts}" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ". In this way, " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " can learn how to extract semantic information to improve fusion quality. In the above two alternate steps, " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " is fixed to offer detection semantic information. Thus, we further implement a mutual promotion learning, where we use the improved " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " to generate fusion results to finetune " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": ", and then the improved " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": " offers better semantic information to optimize " + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 450, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "In summary, our contributions are as follows. (1) We explore the joint learning framework of IVIF and OD, and propose MetaFusion to obtain superior performance on these two tasks. (2) Meta-feature embedding network is designed to generate meta features that bridge the gap between " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". (3) Sequentially, mutual promotion learning between " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " is introduced to improve their performances. (4) Extensive experiments on image fusion and object detection validate the effectiveness of the proposed method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 190, + 392, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 190, + 392, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 392, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 211, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 414 + ], + "type": "text", + "content": "Image Fusion. Traditional image fusion methods use hand-crafted features, such as sparse representation [22], spectral variation [53] and low-rank representation [14], which can not handle the complex scenes well. Nowadays, deep learning-based image fusion methods [16, 24, 27, 39, 49, 50, 54, 56] are proposed. Ma et al. [23] propose a swim transformer based image fusion method with cross-domain long-range learning. Xu et al. [44] use feature extraction and measurement to estimate the degree of information preservation in image fusion. Zhao et al. [51] design a self-supervised feature extraction model. Besides, feature diversity enhancement and fusion is received attention [52, 55]. However, most of them neglect the help of high-level tasks. Recently, Liu et al. [20] propose a joint learning of IVIF and OD. They treat OD network as an additional constraint to help IVIF network generate fusion result with more clear objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 415, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 533 + ], + "type": "text", + "content": "However, directly using the high-level OD network to guide pixel-level IVIF may not obtain an effective constraint. Moreover, they use the fusion results as the connection between IVIF network and OD network, ignoring the pixel-level semantic information in OD features. In contrast, we adopt semantic information in OD features to help IVIF, and conduct a meta-feature embedding to generate meta features from OD features. Then, those meta features are used to guide IVIF network to learn pixel-level semantic information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "content": "Object Detection. Deep learning-based object detection methods have made great progress, e.g., network structure-designed methods [2, 8, 30, 31] and loss constraint methods [18, 32, 42]. Carion et al. [3] introduce a transformer encoder-decoder architecture for object detection. Xu et al. [42] rank positive and negative sample pairs and improve the ranking effect by a clustering algorithm to built object detection loss. On the other hand, many weakly supervised methods [40, 46, 48] are proposed. Wu et al. [40] design a model to suppress the background activation and obtain object region, where image-level labels are used. Zhang et al. [48] use instance proposal grouping to learn object detection from a single point annotation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Most of those object detection methods leverage single modality image, e.g., visible images. However, the visible" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13956" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 532, + 242 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 532, + 242 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 532, + 242 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 532, + 242 + ], + "type": "image", + "image_path": "caef070d7ffe13a45f7345ec4df249e9870fe0669478a6b797e9d39d2eb76a5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": "Figure 2. Framework illustration of the proposed MetaFusion. MetaFusion includes three parts: infrared and visible image fusion (IVIF), object detection (OD), and meta-feature embedding (MFE). IFIV network includes four blocks: three feature fusion blocks " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "FFB_{1}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "FFB_{2}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "FFB_{3}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": " used to extract and fuse features and one image reconstruction module " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "IRM" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": " used to reconstruct fusion result. MFE contains three meta-feature generators " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "MFG_{1}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "MFG_{2}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "MFG_{3}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": " conducted to generate multi-level meta features from OD features " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "F_{ej}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", and three feature transform networks " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "FT_{1}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "FT_{2}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "FT_{3}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": " used to transfer the meta features to fusion features. OD network is divided as three feature extraction blocks " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "DFB_{1}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "DFB_{2}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "DFB_{3}" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": " and one OD detection head " + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "inline_equation", + "content": "DH" + }, + { + "bbox": [ + 46, + 248, + 548, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 333, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 392 + ], + "type": "text", + "content": "images are affected by light conditions, thereby influencing the performance of object detection. On the contrary, we leverage IVIF results to provide infrared information without being influenced by illumination, therefore can offer more stable detection performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 393, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 537 + ], + "type": "text", + "content": "Meta Learning. Meta learning aims to use the learned knowledge to quickly adapt to new tasks, which has been used to many fields: model pre-training [5, 11, 29], image classification [17, 19], few-shot learning [1, 4, 6, 41], etc. For example, Raghu et al. [29] propose an implicit differentiation and backpropagation meta-learning method and improve the efficiency of learning pre-training hyperparameters. Xu et al. [41] design a dynamic alignment in meta learning, which can emphasize information in query according to the support. Li et al. [12] use the meta learning to handle different input resolutions, thereby generating fusion results with arbitrary resolutions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 586 + ], + "type": "text", + "content": "Inspired by the above meta learning methods, we design a meta-feature embedding network to generate the meta features from object detection features, thereby helping IVIF network fuse more object semantic information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 153, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 153, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 153, + 611 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Our MetaFusion framework is shown in Figure 2, which includes three subnetworks: IVIF network " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " generates fusion results, OD network " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " offers semantic features, and meta-feature embedding network " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "(MFE)" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " uses OD features to guide IVIF to fuse more object semantic information. The challenge is that OD features can hardly be directly used to guide IVIF due to their task gap. Thus, our focus is to design " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "MFE" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " and optimize it. Details are as follows." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 333, + 549, + 453 + ], + "blocks": [ + { + "bbox": [ + 312, + 333, + 549, + 453 + ], + "lines": [ + { + "bbox": [ + 312, + 333, + 549, + 453 + ], + "spans": [ + { + "bbox": [ + 312, + 333, + 549, + 453 + ], + "type": "image", + "image_path": "136eafc183a7d3af98a5a0e498950ee091625d9583a47f04d88607982d561a7c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "text", + "content": "Figure 3. Illustration of the meta-feature embedding learning. There are two stages in the meta-feature embedding: (a) inner update and (b) outer update. In the inner update, " + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "inline_equation", + "content": "F^T" + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "text", + "content": " are trained according to current " + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "text", + "content": " semantic extracting ability to generate meta features. In the outer update, " + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 460, + 547, + 538 + ], + "type": "text", + "content": " is trained which is guided by the meta features, thereby learning to extract object semantic. The two stages alternate every N epochs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 549, + 448, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 448, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 448, + 562 + ], + "type": "text", + "content": "3.1. Meta-feature Embedding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": "Meta-feature embedding includes meta-feature generator (MFG) and feature transform (FT). MFG generates meta feature " + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "inline_equation", + "content": "F_{mj}" + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": " according to IVIF feature " + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "inline_equation", + "content": "F_{uj}" + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": " from OD feature " + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "inline_equation", + "content": "F_{ej}" + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "inline_equation", + "content": "F_{mj} = MFG(F_{uj}, F_{ej})" + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": " is the feature level index. FT transfers the meta feature " + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "inline_equation", + "content": "F_{mj}" + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": " to fusion feature " + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "inline_equation", + "content": "F_{uj}" + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": " by producing feature bridge " + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "inline_equation", + "content": "F_{tj}" + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": ". Specifically, meta-feature embedding learning is divided two stages: inner update stage and outer update stage, as shown in Figure 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": "In the inner update stage, we optimize " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": " (see Figure 3(a)). Given meta training set " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "S_{mtr}" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": " and meta testing set " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "S_{mts}" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": ", we firstly update the parameter " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta_F" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": " on" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13957" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 198, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 198, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 198, + 84 + ], + "type": "inline_equation", + "content": "S_{mtr}" + }, + { + "bbox": [ + 47, + 72, + 198, + 84 + ], + "type": "text", + "content": " through meta feature guidance:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 92, + 287, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 92, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 287, + 127 + ], + "type": "interline_equation", + "content": "\\theta_ {F ^ {\\prime}} = \\theta_ {F} - \\beta_ {F ^ {\\prime}} \\nabla_ {\\theta_ {F}} L _ {g} \\left(F _ {m j}, F _ {t j}\\right) = \\theta_ {F} - \\beta_ {F ^ {\\prime}} \\frac {\\partial L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F}}, \\tag {1}", + "image_path": "fbaae0ea883b24fb48c340b9e8f4f67fcd75f425159abec23eb0bba24551fee2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": "where the guide loss " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "L_{g}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " distance, " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "\\beta_{F^{\\prime}}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " is the learning rate of " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "F^{\\prime}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "\\theta_{F^{\\prime}}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " is the parameter of " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "F^{\\prime}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " that is updated from " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": ". Then, we use " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "F^{\\prime}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "\\theta_{F^{\\prime}}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " to calculate the fusion loss " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " on meta testing set " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "S_{mts}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " can measure the effect of using the meta features to guide " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": ", i.e., when " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " is lower, the semantic extracting ability of " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " will be improved. Therefore, we use " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " to update parameter " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "\\theta_{MFG}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " and parameter " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "\\theta_{FT}" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 47, + 128, + 289, + 224 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 232, + 287, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 232, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 232, + 287, + 247 + ], + "type": "interline_equation", + "content": "\\theta_ {M F G} = \\theta_ {M F G} - \\beta_ {M F G} \\nabla_ {\\theta_ {M F G}} L _ {f} \\left(I _ {f} ^ {m t s}, I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), (2)", + "image_path": "650a80752d0b63d58c3a849cfced3a15e48a04fe986f9bc7dd77a552c548011d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 255, + 287, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 255, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 70, + 255, + 287, + 270 + ], + "type": "interline_equation", + "content": "\\theta_ {F T} = \\theta_ {F T} - \\beta_ {F T} \\nabla_ {\\theta_ {F T}} L _ {f} \\left(I _ {f} ^ {m t s}, I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), \\tag {3}", + "image_path": "d16e75ff721e4fb5b9d96f2257af7bada52196e9ed0a304b587943ea223750ba.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": "where the fusion loss " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": " is SSIM loss [38], " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "I_{i}^{mts}" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "I_{v}^{mts}" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": " are infrared and visible images from " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "S_{mts}" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "I_{f}^{mts} = F'(I_{i}^{mts}, I_{v}^{mts})" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": " is the meta testing fusion result, and " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\beta_{MFG}" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\beta_{FT}" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": " are the learning rate of MFG and " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta_{MFG}} L_{f}" + }, + { + "bbox": [ + 47, + 274, + 287, + 335 + ], + "type": "text", + "content": " can be calculated by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 343, + 287, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 343, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 51, + 343, + 287, + 395 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta_ {M F G}} L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right) \\\\ = \\frac {\\partial L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right) , I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right)}{\\partial \\theta_ {F ^ {\\prime}}} * \\left(- \\frac {\\partial^ {2} L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F} \\partial \\theta_ {M F G}}\\right). \\tag {4} \\\\ \\end{array}", + "image_path": "a57af2255cb04afdbe50a013ffacb32681ae35cf718728b604a528c44e837dde.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 396, + 209, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 209, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 209, + 408 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 47, + 396, + 209, + 408 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta_{FT}}L_f" + }, + { + "bbox": [ + 47, + 396, + 209, + 408 + ], + "type": "text", + "content": " can be calculated by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 416, + 287, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 416, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 52, + 416, + 287, + 467 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta_ {F T}} L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right), I _ {i} ^ {m t s}, I _ {v} ^ {m t s}\\right) \\\\ = \\frac {\\partial L _ {f} \\left(F ^ {\\prime} \\left(I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right) , I _ {i} ^ {m t s} , I _ {v} ^ {m t s}\\right)}{\\partial \\theta_ {F ^ {\\prime}}} * \\left(- \\frac {\\partial^ {2} L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F} \\partial \\theta_ {F T}}\\right). \\tag {5} \\\\ \\end{array}", + "image_path": "c4bb950829d6e4274ff224c7fa0d2f443f8000d9d690b82812660ef6aadaf6bb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "type": "text", + "content": "Thus, " + }, + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "type": "text", + "content": " can learn how to generate meta features according to current " + }, + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "type": "inline_equation", + "content": "F^{\\prime}" + }, + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "type": "text", + "content": " semantic extracting ability, i.e., makes meta features be compatible with " + }, + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 468, + 287, + 503 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "text", + "content": "In the outer update stage, " + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "text", + "content": " is trained with the initial parameter " + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\theta_F" + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "text", + "content": ", as shown in Figure 3(b). Given the training data " + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "S = \\{S_{mtr}, S_{mts}\\}" + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "text", + "content": " is optimized using the fusion loss " + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "L_f" + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "text", + "content": " and guide loss " + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "L_g" + }, + { + "bbox": [ + 47, + 504, + 287, + 553 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 561, + 288, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 561, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 48, + 561, + 288, + 640 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta_ {F} = \\theta_ {F} - \\beta_ {F} \\nabla_ {\\theta_ {F}} \\left(L _ {f} \\left(I _ {f}, I _ {i}, I _ {v}\\right) + \\lambda_ {g} \\sum_ {j = 1} ^ {3} L _ {g} \\left(F _ {m j}, F _ {t j}\\right)\\right) \\\\ = \\theta_ {F} - \\beta_ {F} \\left(\\frac {\\partial L _ {f} \\left(I _ {f} , I _ {i} , I _ {v}\\right)}{\\partial \\theta_ {F}} + \\lambda_ {g} \\sum_ {j = 1} ^ {3} \\frac {\\partial L _ {g} \\left(F _ {m j} , F _ {t j}\\right)}{\\partial \\theta_ {F}}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "9ec84466b98a83dcd465977cd254029e94b7a336466d33bf790d39d563e7d160.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "I_{v}" + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": " are infrared and visible images from " + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\lambda_{g}" + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": " is a hyper-parameter to balance " + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "L_{g}" + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\beta_{F}" + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": " is the learning rate of " + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 642, + 287, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Finally, the inner update stage and outer update stage are carried out alternately for every " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " epochs, which improves the optimization of " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 529, + 277 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 529, + 277 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 529, + 277 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 529, + 277 + ], + "type": "image", + "image_path": "0ce69d4d5025548242fcc713ea43b1a5aa3a141ca0f701fe1e91fc1559f44905.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "text", + "content": "Figure 4. Illustration of the mutual promotion learning between " + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "text", + "content": ". Firstly, " + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "text", + "content": " is used to generate fusion results to optimize " + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "text", + "content": ". Thus, a fine-tuned " + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "text", + "content": " with higher semantic extracting ability is obtained. Then, " + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "text", + "content": " is adopted to optimize " + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 282, + 545, + 337 + ], + "type": "text", + "content": ". This process is alternated." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 350, + 462, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 462, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 462, + 363 + ], + "type": "text", + "content": "3.2. Mutual Promotion Learning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": "In Sec. 3.1, we fix the object detection network " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " to offer stable object semantic features. In this way, " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " does not need to handle variable object semantic features, thereby reducing the difficulty of training " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": ". However, the fixed " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " will restrict the object semantic extracting ability of " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": ". Addressing this problem, we propose the mutual promotion learning that promotes not only " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " but also " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": ", as shown in Figure 4. Specifically, we firstly train " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " individually, and then use the fusion results to train " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": ". Sequentially, " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " offers object semantic features to the meta-feature embedding, thereby improving " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": ". After that, the improved fusion results generated by " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " are used to fine-tune " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": ". In this way, both " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 370, + 545, + 526 + ], + "type": "text", + "content": " are improved." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 388, + 550 + ], + "type": "text", + "content": "3.3. Architecture" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": "IVIF network " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": " aims to generate fusion results according to the input infrared and visible images, which is flexible. Here, we adopt the similar structure with [50]. " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": " consists of three feature fusion blocks " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "(FFB_{j}, j = 1,2,3)" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": " and one image reconstruction module (IRM), as shown in Figure 2. " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "FFB_{j}" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": " is used to extract and fuse features from input infrared image " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": " and visible image " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "I_{v}" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": ". It is designed based on the idea of [50] that if " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "FFB_{j}" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": " can well fuse the features of " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "I_{v}" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": ", then the two kinds of features can be recovered from the fused feature. Thus, " + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "inline_equation", + "content": "FFB_{j}" + }, + { + "bbox": [ + 305, + 557, + 545, + 678 + ], + "type": "text", + "content": " is described as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 350, + 689, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 689, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 350, + 689, + 545, + 717 + ], + "type": "interline_equation", + "content": "F F B _ {j} = \\left\\{ \\begin{array}{c} C _ {2} T _ {1} \\biguplus C _ {2} T _ {1}, j = 1 \\\\ C _ {2} T _ {j - 1} \\biguplus C _ {2} T _ {j - 1}, j > 1 \\end{array} \\right. \\tag {7}", + "image_path": "ee038526bb222cf956a6fe77ddfdd863d4e389391075d53b13f8ee43baccb888.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13958" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " is the block index, " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " represents the number of \"Convolution with " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " kernel +ReLU\" layers is " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "F_{u(j-1)}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " is the output feature of the " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "j-1" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "th block, " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "T_1 = C_2(I_i, I_v)" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "T_{j-1} = C_2(F_{u(j-1)})" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " represent feature integration operations, and " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\boxplus" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " means feature concatenation. IRM reconstructs the fusion result, which contains six \"Convolution with " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " kernel +ReLU\" layers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "content": "Object detection network " + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "content": " offers object semantic features. In our framework, we choose " + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{yolov5s}^1" + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "content": " as our " + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "content": ". According to the size of the feature maps, the backbone of " + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "content": " is divided into three feature extraction blocks " + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "inline_equation", + "content": "(\\{DFB_{i}, i = 1,2,3\\})" + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "content": ". In addition, we represent the neck and detection head in yolov5s into one block " + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "inline_equation", + "content": "(DH)" + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "content": " to simplify the expression." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": "Meta-feature embedding network " + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "MFE" + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": " includes metafeature generator " + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "MFG_{j}" + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": " and feature transform network " + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "FT_{j}" + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": ". Since OD feature " + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "F_{ej}" + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": " generated by " + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "DFB_{j}" + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": " offers object semantic information and " + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "F_{uj}" + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": " provides scene details, they are fed into the " + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "MFG_{j}" + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "MFG_{j}" + }, + { + "bbox": [ + 46, + 239, + 287, + 300 + ], + "type": "text", + "content": " is built as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 309, + 287, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 309, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 81, + 309, + 287, + 326 + ], + "type": "interline_equation", + "content": "M F G _ {j} = C _ {6} \\left(C _ {4} \\left(F _ {u j}\\right) \\biguplus C _ {2} \\left(U p \\left(F _ {e j}\\right)\\right)\\right), \\tag {8}", + "image_path": "73c32d0dee718cbada4d1504315773d1e3e1d6ef82987a7377c3e45a34e74ea6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "type": "inline_equation", + "content": "Up()" + }, + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "type": "text", + "content": " is the feature up-sampling. " + }, + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "type": "inline_equation", + "content": "FT_{j}" + }, + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "type": "text", + "content": " transfers the meta feature to fusion feature, which includes three \"Convolution with " + }, + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 334, + 287, + 370 + ], + "type": "text", + "content": " kernel +ReLU\" layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 377, + 111, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 111, + 391 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 111, + 391 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": "The training details of our framework is shown in Algorithm 1, which contains four steps. Generally, we firstly pretrain " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": ", and then use the fusion results to finetune " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": ". Sequentially, we train " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " are updated " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "n = 200" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " times every " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " epochs. Lastly, the mutual promotion between " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " is trained for " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "R = 2" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " rounds, thereby improving their performances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 127, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 127, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 127, + 504 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 510, + 96, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 96, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 96, + 522 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "content": "Dataset. We conduct experiment on three widely-used datasets: " + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "content": " [20], RoadScene [44] and TNO [37]. " + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "content": " is divided as training set (2940 image pairs) and testing set (1260 image pairs). RoadScene with 221 image pairs and TNO with 40 image pairs are only used for testing. Besides, " + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "content": " is adopted to evaluate OD performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": "Implementation. Our framework is implemented with PyTorch on a NVIDIA GeForce RTX 3090 GPU. " + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": " and MFE are trained using optimizer Adam with the learning rate " + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\beta_{F},\\beta_{F^{\\prime}},\\beta_{FT}" + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\beta_{MFG}" + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": " is trained followed by yolov5s that adopts the optimizer SGD using the learning rate of " + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "inline_equation", + "content": "1\\times 10^{-2}" + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": " with decaying rate of 0.1 every round. We firstly train " + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": " for 100 epochs and " + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 600, + 287, + 695 + ], + "type": "text", + "content": " for 150 epochs. Then, we conduct MFE for 50 epochs. After" + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 306, + 87, + 545, + 578 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 492, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 492, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 492, + 84 + ], + "type": "text", + "content": "Algorithm 1 MetaFusion Training Algorithm" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": "Input: Training dataset " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "S = \\{S_{mtr}, S_{mts}\\}" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": ", IVIF model " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F(\\theta_F)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": ", meta-feature generator " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "MFG(\\theta_F)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": ", feature transform network " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "FT(\\theta_{FT})" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": ", OD network " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "D(\\theta_D)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \nOutput: IVIF model " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F(\\theta_F^*)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n1: Initialize " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F(\\theta_F)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "MFG(\\theta_F)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "FT(\\theta_{FT})" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "D(\\theta_D)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n2: /* Pretrain " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F^{*}/" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n3: while not converged do \n4: Sample image pair " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "(I_i, I_v)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n5: Forward pass on " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " to generate fusion result " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n6: Optimize " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " using fusion loss " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "L_f" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n7: end while \n8: /* Pretrain " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "D^{*}/" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n9: while not converged do \n10: Sample image " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " from the fused images generated by " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n11: Forward pass on " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " to get detection result \n12: Optimize " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " using detection loss " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "L_d" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n13: end while \n14: /* Meta-feature embedding outer update */ \n15: while not converged do \n16: Sample a batch image pair " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "(I_i, I_v)" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n17: Forward pass on " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " to generate fusion result " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n18: Forward pass on " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "MFG" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " to generate meta feature \n19: Optimize " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " using fusion loss " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "L_f" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " and guide loss " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "L_g" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " by Eq. 6 \n20: /* Meta-feature embedding inner update */ \n21: if epoch%N==0 then \n22: for t=1 to n do \n23: Sample an image pair " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "(I_i^{mtr}, I_v^{mtr})" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "S_{mtr}" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n24: Compute " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\theta_{F'}" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " by Eq. 1 \n25: Sample an image pair " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "(I_i^{mts}, I_v^{mts})" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "S_{mts}" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n26: Calculate " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "L_f" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\theta_{F'}" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " \n27: Update " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\theta_{MFG}" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\theta_{FT}" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " using the gradient of " + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "inline_equation", + "content": "L_f" + }, + { + "bbox": [ + 306, + 87, + 545, + 578 + ], + "type": "text", + "content": " by Eqs. 2 and 3 \n28: end for \n29: end if \n30: end while \n31: /* Mutual promotion*/ \n32: if Number of round ≤ R then \n33: Go back to line 9 \n34: end if" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": "that, " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " is finetuned for 150 epochs. Following the same strategy, the mutual promotion is carried out " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " rounds. The images are resized to " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "512 \\times 384" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " with batchsize of 1. The hyperparameter " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\lambda_{g}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " is set to 0.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Metric. Three metrics are used for IVIF evaluation: entropy (EN) [33], mutual information (MI) [28] and visual information fidelity (VIF) [7]. EN evaluates the information richness in an image, and the higher EN means more information. MI evaluates the information similarity between" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 189, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 189, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 189, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 702, + 189, + 713 + ], + "type": "text", + "content": "https://github.com/ultralytics/YOLOv5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13959" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 113, + 150 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 113, + 150 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 113, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 113, + 150 + ], + "type": "image", + "image_path": "ad293010675822b131f424cffc076bde8fd54c2a123d2075e4dbce42638ad3ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 151, + 106, + 156 + ], + "lines": [ + { + "bbox": [ + 61, + 151, + 106, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 151, + 106, + 156 + ], + "type": "text", + "content": "(a) Visible-infrared images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 114, + 72, + 172, + 150 + ], + "blocks": [ + { + "bbox": [ + 114, + 72, + 172, + 150 + ], + "lines": [ + { + "bbox": [ + 114, + 72, + 172, + 150 + ], + "spans": [ + { + "bbox": [ + 114, + 72, + 172, + 150 + ], + "type": "image", + "image_path": "5d84997b06a4f246fb6b57071c6fe16f6d4755bcd016e6e8d08defa9513f78a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 151, + 167, + 156 + ], + "lines": [ + { + "bbox": [ + 117, + 151, + 167, + 156 + ], + "spans": [ + { + "bbox": [ + 117, + 151, + 167, + 156 + ], + "type": "text", + "content": "(b) SoFusion and feature map" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 230, + 150 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 230, + 150 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 230, + 150 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 230, + 150 + ], + "type": "image", + "image_path": "9391a83037ed5dc068fd2a84eaff4e7bc2f11318622a7a96427f068c312ef89c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 151, + 227, + 156 + ], + "lines": [ + { + "bbox": [ + 178, + 151, + 227, + 156 + ], + "spans": [ + { + "bbox": [ + 178, + 151, + 227, + 156 + ], + "type": "text", + "content": "(c) CoFusion and feature map" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 164, + 282, + 176 + ], + "lines": [ + { + "bbox": [ + 51, + 164, + 282, + 176 + ], + "spans": [ + { + "bbox": [ + 51, + 164, + 282, + 176 + ], + "type": "text", + "content": "Figure 5. Visual comparison of different optimization methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 231, + 72, + 288, + 150 + ], + "blocks": [ + { + "bbox": [ + 231, + 72, + 288, + 150 + ], + "lines": [ + { + "bbox": [ + 231, + 72, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 231, + 72, + 288, + 150 + ], + "type": "image", + "image_path": "fc40a81a98b73fc408062d9cc005fd1b877e9acc7ee712b0e4d50c63f03f89f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 151, + 286, + 156 + ], + "lines": [ + { + "bbox": [ + 233, + 151, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 233, + 151, + 286, + 156 + ], + "type": "text", + "content": "(d) MetaFusion and feature map" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 179, + 286, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 179, + 286, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 179, + 286, + 211 + ], + "type": "text", + "content": "Table 1. Effect study of meta-feature embedding by comparing with different optimization methods on " + }, + { + "bbox": [ + 47, + 179, + 286, + 211 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 47, + 179, + 286, + 211 + ], + "type": "text", + "content": ". The best result is in red." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 57, + 217, + 276, + 299 + ], + "blocks": [ + { + "bbox": [ + 57, + 217, + 276, + 299 + ], + "lines": [ + { + "bbox": [ + 57, + 217, + 276, + 299 + ], + "spans": [ + { + "bbox": [ + 57, + 217, + 276, + 299 + ], + "type": "table", + "html": "
MethodMetric
MIENVIF
SoFusion14.1647.0781.190
CoFusion14.1877.0891.345
MetaFusion14.5117.2491.515
", + "image_path": "1a132f492a4064ce525aa1dd2d7804c837e88a24fdad39bfebb5c46d4cd7d5fd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "text", + "content": "the input images and fused image. The higher MI illustrates more information of the input images is fused. VIF measures the ability to extract visible information from the input image, and a larger VIF represents less visible distortion in the fused result. Here, we use the V channel in HSV space of the fusion results to calculate these metrics. Moreover, we use " + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50\\rightarrow 95}" + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "text", + "content": " [9] to comprehensively evaluate OD performance, where the average of mAPs sampling every 5 from " + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{95}" + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "text", + "content": " is calculated. A higher " + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50\\rightarrow 95}" + }, + { + "bbox": [ + 46, + 315, + 286, + 434 + ], + "type": "text", + "content": " means better OD effect." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 443, + 140, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 140, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 140, + 455 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": "Effect of meta-feature embedding. In Sec. 3.1, we introduce the meta-feature embedding that generates the meta features from object detection features to help IVIF network learn more object semantic information. To verify its effect, separate optimization method (SoFusion) and cascaded optimization method (CoFusion) are compared, as shown in Figure 1(a)-(b). Specifically, SoFusion firstly optimizes " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": " and then uses the generated fusion results to optimize " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": ". CoFusion treats " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": " as a constraint to optimize " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": " with the loss " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": ". The results are shown in Table 1. Our MetaFusion achieves the best results on IVIF. The reason is that SoFusion neglects the help of " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": ", and directly utilizing the high-level " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": " to guide the pixel-level " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": " has feature mismatch in CoFusion. In contrast, we implement the meta-feature embedding that generates meta features from " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": " according to the ability of " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": ". Thus, the meta features are naturally compatible with " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": ", thereby providing effective object semantic features to " + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 461, + 287, + 700 + ], + "type": "text", + "content": ". Visual comparison is shown in Figure 5. MetaFusion highlights the object feature and generates the clearest result." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Influence of mutual promotion learning. As described" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 107, + 532, + 190 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "type": "text", + "content": "Table 2. Influence study of mutual promotion learning by evaluating the performances of round " + }, + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "type": "inline_equation", + "content": "{\\mathrm{M}}^{3}\\mathrm{{FD}}" + }, + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "type": "text", + "content": " . The best result is in red." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 107, + 532, + 190 + ], + "lines": [ + { + "bbox": [ + 318, + 107, + 532, + 190 + ], + "spans": [ + { + "bbox": [ + 318, + 107, + 532, + 190 + ], + "type": "table", + "html": "
MethodMetric
MIENVIFmAP50→95 (%)
R = 014.1647.0781.19055.6
R = 114.4647.2261.47455.8
R = 214.5117.2491.51556.5
", + "image_path": "d4ef4ee68911d1950f5d8435ee47b9769027ec586dc409647b4d42a5aa127a63.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 315, + 225, + 535, + 308 + ], + "blocks": [ + { + "bbox": [ + 306, + 200, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 306, + 200, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 545, + 221 + ], + "type": "text", + "content": "Table 3. Study of multi-level meta-feature embedding by comparing different configurations on " + }, + { + "bbox": [ + 306, + 200, + 545, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 306, + 200, + 545, + 221 + ], + "type": "text", + "content": ". The best result is in red." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 225, + 535, + 308 + ], + "lines": [ + { + "bbox": [ + 315, + 225, + 535, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 225, + 535, + 308 + ], + "type": "table", + "html": "
MethodMetric
MIENVIF
MetaFusion-L114.4507.2201.259
MetaFusion-L214.4527.2201.377
MetaFusion-L314.4647.2261.474
", + "image_path": "6e44d7b2893c890b4475788958f1b68e8005bb9368f25cce4b82165c1b8c6bba.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": "in Sec. 3.2, we propose the mutual promotion learning promote " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " alternately. Here, we study the influence of mutual promotion learning by evaluating round " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "R = 0,1,2" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": ". In particular, " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "R = 0" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " are trained respectively. As shown in Table 2, the results of " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " are better with the " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " increases. Comprehensively considering training efficiency and performance, we take " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "R = 2" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": "Study of multi-level meta-feature embedding. We implement multi-level meta-feature embedding to build MetaFusion framework, as introduced in Sec. 3.3. Here, we study the number of meta-feature embedding levels with the following configurations. One-level meta-feature embedding (MetaFusion-L1) with " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "MFE = \\{MFG_1, FT_1\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "F = \\{FFB_1, IRM\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "D = \\{DFB_1, DH\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": ". Two-level meta-feature embedding (MetaFusion-L2) with " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "MFE = \\{MFG_1, MFG_2, FT_1, FT_2\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "F = \\{FFB_1, FFB_2, IRM\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "D = \\{DFB_1, DFB_2, DH\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": ". Three-level meta-feature embedding (MetaFusion-L3) with " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "MFE = \\{MFG_1, MFG_2, MFG_3, FT_1, FT_2, FT_3\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "F = \\{FFB_1, FFB_2, FFB_3, IRM\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "D = \\{DFB_1, DFB_2, DFB_3, DH\\}" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": ". The results are shown in Table 3. With the number of multi-level meta-feature embedding increases, " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": " achieves higher image fusion quality. Since meta-feature embedding provide multi-level object semantic features to " + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 409, + 546, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 634, + 528, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 528, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 528, + 647 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "We compare the proposed MetaFusion with eight SOTA fusion methods to verify the superiority: FusionGAN [25], GANMcC [26], MFEIF [21], YDTR [36], PIAFusion [35], SwinFusion [23], Tardal [20], and U2Fusion [44]. Their available codes and recommended parameter settings are" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13960" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 162, + 144 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 162, + 144 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 162, + 144 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 162, + 144 + ], + "type": "image", + "image_path": "a6948d69ed6adca343da4e9b96dfa341d6031457013efd9491b1519cbed6cd5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 146, + 156, + 155 + ], + "lines": [ + { + "bbox": [ + 83, + 146, + 156, + 155 + ], + "spans": [ + { + "bbox": [ + 83, + 146, + 156, + 155 + ], + "type": "text", + "content": "Infrared-visible image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 73, + 251, + 144 + ], + "blocks": [ + { + "bbox": [ + 164, + 73, + 251, + 144 + ], + "lines": [ + { + "bbox": [ + 164, + 73, + 251, + 144 + ], + "spans": [ + { + "bbox": [ + 164, + 73, + 251, + 144 + ], + "type": "image", + "image_path": "88508f5965299148b31b56b5d6d3f8c58eccf04a84d59cd27c19d62f09d6a498.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 146, + 228, + 154 + ], + "lines": [ + { + "bbox": [ + 187, + 146, + 228, + 154 + ], + "spans": [ + { + "bbox": [ + 187, + 146, + 228, + 154 + ], + "type": "text", + "content": "FusionGAN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 73, + 340, + 144 + ], + "blocks": [ + { + "bbox": [ + 253, + 73, + 340, + 144 + ], + "lines": [ + { + "bbox": [ + 253, + 73, + 340, + 144 + ], + "spans": [ + { + "bbox": [ + 253, + 73, + 340, + 144 + ], + "type": "image", + "image_path": "a54a70e631bac6372eb2f5827c768b359dd40a4dbfaef7a3813ca965493cef89.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 73, + 429, + 144 + ], + "blocks": [ + { + "bbox": [ + 342, + 73, + 429, + 144 + ], + "lines": [ + { + "bbox": [ + 342, + 73, + 429, + 144 + ], + "spans": [ + { + "bbox": [ + 342, + 73, + 429, + 144 + ], + "type": "image", + "image_path": "f2e7f06cb59fbf4dbb9bac8f4ba5f94c633b094f1c20b5b14d5c0d20c8a33f35.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 430, + 73, + 517, + 144 + ], + "blocks": [ + { + "bbox": [ + 430, + 73, + 517, + 144 + ], + "lines": [ + { + "bbox": [ + 430, + 73, + 517, + 144 + ], + "spans": [ + { + "bbox": [ + 430, + 73, + 517, + 144 + ], + "type": "image", + "image_path": "852d3dd4c719b3eb05074b4b0c4cc075b771b2b9bf705c2505aa8b8684eb302c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 75, + 156, + 162, + 227 + ], + "blocks": [ + { + "bbox": [ + 75, + 156, + 162, + 227 + ], + "lines": [ + { + "bbox": [ + 75, + 156, + 162, + 227 + ], + "spans": [ + { + "bbox": [ + 75, + 156, + 162, + 227 + ], + "type": "image", + "image_path": "d8e91f56e3c7b9b817435a0c052e9a751ce44fb42e8826395cbe8dda7b5db68a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 164, + 156, + 250, + 227 + ], + "blocks": [ + { + "bbox": [ + 164, + 156, + 250, + 227 + ], + "lines": [ + { + "bbox": [ + 164, + 156, + 250, + 227 + ], + "spans": [ + { + "bbox": [ + 164, + 156, + 250, + 227 + ], + "type": "image", + "image_path": "aa2151fce9011d3a1ee0802be7a594fc55d8654e144fa0861e7d1213147c3688.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 253, + 156, + 339, + 227 + ], + "blocks": [ + { + "bbox": [ + 282, + 147, + 317, + 155 + ], + "lines": [ + { + "bbox": [ + 282, + 147, + 317, + 155 + ], + "spans": [ + { + "bbox": [ + 282, + 147, + 317, + 155 + ], + "type": "text", + "content": "GANMcC" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 156, + 339, + 227 + ], + "lines": [ + { + "bbox": [ + 253, + 156, + 339, + 227 + ], + "spans": [ + { + "bbox": [ + 253, + 156, + 339, + 227 + ], + "type": "image", + "image_path": "20db97aad59377cb1f63ca37a2fdc03c755daf3a6af2006e0c88a40878e21656.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 341, + 156, + 428, + 227 + ], + "blocks": [ + { + "bbox": [ + 376, + 147, + 402, + 155 + ], + "lines": [ + { + "bbox": [ + 376, + 147, + 402, + 155 + ], + "spans": [ + { + "bbox": [ + 376, + 147, + 402, + 155 + ], + "type": "text", + "content": "MFEIF" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 156, + 428, + 227 + ], + "lines": [ + { + "bbox": [ + 341, + 156, + 428, + 227 + ], + "spans": [ + { + "bbox": [ + 341, + 156, + 428, + 227 + ], + "type": "image", + "image_path": "0e1cb7d4873a1eb56a6e1195c24bcd75a85a7eba53fa26479e755d62dc7b202e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 430, + 156, + 517, + 227 + ], + "blocks": [ + { + "bbox": [ + 460, + 147, + 493, + 155 + ], + "lines": [ + { + "bbox": [ + 460, + 147, + 493, + 155 + ], + "spans": [ + { + "bbox": [ + 460, + 147, + 493, + 155 + ], + "type": "text", + "content": "U2Fusion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 156, + 517, + 227 + ], + "lines": [ + { + "bbox": [ + 430, + 156, + 517, + 227 + ], + "spans": [ + { + "bbox": [ + 430, + 156, + 517, + 227 + ], + "type": "image", + "image_path": "03ba215c1803ec5764514a99aa83c4471b5ca4e85616aeab5171f5cf599a4f0c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 75, + 239, + 164, + 316 + ], + "blocks": [ + { + "bbox": [ + 107, + 230, + 130, + 238 + ], + "lines": [ + { + "bbox": [ + 107, + 230, + 130, + 238 + ], + "spans": [ + { + "bbox": [ + 107, + 230, + 130, + 238 + ], + "type": "text", + "content": "YDTR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 239, + 164, + 316 + ], + "lines": [ + { + "bbox": [ + 75, + 239, + 164, + 316 + ], + "spans": [ + { + "bbox": [ + 75, + 239, + 164, + 316 + ], + "type": "image", + "image_path": "cb4d87835cb6a47d9194550c743d2137e6724038f9993d1310cc9a2a162b2fe4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 164, + 239, + 251, + 316 + ], + "blocks": [ + { + "bbox": [ + 192, + 230, + 228, + 238 + ], + "lines": [ + { + "bbox": [ + 192, + 230, + 228, + 238 + ], + "spans": [ + { + "bbox": [ + 192, + 230, + 228, + 238 + ], + "type": "text", + "content": "PIAFusion" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 239, + 251, + 316 + ], + "lines": [ + { + "bbox": [ + 164, + 239, + 251, + 316 + ], + "spans": [ + { + "bbox": [ + 164, + 239, + 251, + 316 + ], + "type": "image", + "image_path": "5656eb01f22b5676b841337b6986b965402f9f920406bf6f622c74d46f004dd3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 317, + 227, + 325 + ], + "lines": [ + { + "bbox": [ + 186, + 317, + 227, + 325 + ], + "spans": [ + { + "bbox": [ + 186, + 317, + 227, + 325 + ], + "type": "text", + "content": "FusionGAN" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 253, + 239, + 339, + 316 + ], + "blocks": [ + { + "bbox": [ + 279, + 230, + 319, + 238 + ], + "lines": [ + { + "bbox": [ + 279, + 230, + 319, + 238 + ], + "spans": [ + { + "bbox": [ + 279, + 230, + 319, + 238 + ], + "type": "text", + "content": "SwinFusion" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 239, + 339, + 316 + ], + "lines": [ + { + "bbox": [ + 253, + 239, + 339, + 316 + ], + "spans": [ + { + "bbox": [ + 253, + 239, + 339, + 316 + ], + "type": "image", + "image_path": "15b224bf6ae870c6265ec3f4032a5103d04a7cd5a83a5673af802f6bce405637.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 341, + 239, + 429, + 316 + ], + "blocks": [ + { + "bbox": [ + 376, + 230, + 399, + 238 + ], + "lines": [ + { + "bbox": [ + 376, + 230, + 399, + 238 + ], + "spans": [ + { + "bbox": [ + 376, + 230, + 399, + 238 + ], + "type": "text", + "content": "Tardal" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 239, + 429, + 316 + ], + "lines": [ + { + "bbox": [ + 341, + 239, + 429, + 316 + ], + "spans": [ + { + "bbox": [ + 341, + 239, + 429, + 316 + ], + "type": "image", + "image_path": "9fbb470afbc421d16b4a4621902e8d60da426399a06bdc68223be9a3c6acd569.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 317, + 400, + 325 + ], + "lines": [ + { + "bbox": [ + 375, + 317, + 400, + 325 + ], + "spans": [ + { + "bbox": [ + 375, + 317, + 400, + 325 + ], + "type": "text", + "content": "MFEIF" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 430, + 239, + 517, + 316 + ], + "blocks": [ + { + "bbox": [ + 457, + 230, + 497, + 238 + ], + "lines": [ + { + "bbox": [ + 457, + 230, + 497, + 238 + ], + "spans": [ + { + "bbox": [ + 457, + 230, + 497, + 238 + ], + "type": "text", + "content": "MetaFusion" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 239, + 517, + 316 + ], + "lines": [ + { + "bbox": [ + 430, + 239, + 517, + 316 + ], + "spans": [ + { + "bbox": [ + 430, + 239, + 517, + 316 + ], + "type": "image", + "image_path": "515fcbf6d7e2e13c469ab0e3c6b4617ffea677ca8b2a02d3b62959ea945f9458.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 75, + 327, + 162, + 399 + ], + "blocks": [ + { + "bbox": [ + 82, + 317, + 155, + 327 + ], + "lines": [ + { + "bbox": [ + 82, + 317, + 155, + 327 + ], + "spans": [ + { + "bbox": [ + 82, + 317, + 155, + 327 + ], + "type": "text", + "content": "Infrared-visible image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 327, + 162, + 399 + ], + "lines": [ + { + "bbox": [ + 75, + 327, + 162, + 399 + ], + "spans": [ + { + "bbox": [ + 75, + 327, + 162, + 399 + ], + "type": "image", + "image_path": "d36d1aae1d2ef1013486fc2f1c9d43f9d4ae16d6356d7f4fa2b7f52773e420fa.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 400, + 130, + 408 + ], + "lines": [ + { + "bbox": [ + 107, + 400, + 130, + 408 + ], + "spans": [ + { + "bbox": [ + 107, + 400, + 130, + 408 + ], + "type": "text", + "content": "YDTR" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 164, + 327, + 251, + 398 + ], + "blocks": [ + { + "bbox": [ + 164, + 327, + 251, + 398 + ], + "lines": [ + { + "bbox": [ + 164, + 327, + 251, + 398 + ], + "spans": [ + { + "bbox": [ + 164, + 327, + 251, + 398 + ], + "type": "image", + "image_path": "1ad257a316b584678fb08417affd4db817082088d0b0249cf87cccd64845a3ab.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 253, + 327, + 339, + 398 + ], + "blocks": [ + { + "bbox": [ + 280, + 318, + 315, + 326 + ], + "lines": [ + { + "bbox": [ + 280, + 318, + 315, + 326 + ], + "spans": [ + { + "bbox": [ + 280, + 318, + 315, + 326 + ], + "type": "text", + "content": "GANMcC" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 327, + 339, + 398 + ], + "lines": [ + { + "bbox": [ + 253, + 327, + 339, + 398 + ], + "spans": [ + { + "bbox": [ + 253, + 327, + 339, + 398 + ], + "type": "image", + "image_path": "8285b498c10e22dfba2c6eedc0e914b8620db201cca92225e008313cbb3740e8.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 341, + 327, + 429, + 398 + ], + "blocks": [ + { + "bbox": [ + 341, + 327, + 429, + 398 + ], + "lines": [ + { + "bbox": [ + 341, + 327, + 429, + 398 + ], + "spans": [ + { + "bbox": [ + 341, + 327, + 429, + 398 + ], + "type": "image", + "image_path": "658322fc0e3150ddd744a4e7a8209033695ffad91c723aacab6174198aafcadb.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 430, + 327, + 517, + 398 + ], + "blocks": [ + { + "bbox": [ + 459, + 318, + 491, + 326 + ], + "lines": [ + { + "bbox": [ + 459, + 318, + 491, + 326 + ], + "spans": [ + { + "bbox": [ + 459, + 318, + 491, + 326 + ], + "type": "text", + "content": "U2Fusion" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 327, + 517, + 398 + ], + "lines": [ + { + "bbox": [ + 430, + 327, + 517, + 398 + ], + "spans": [ + { + "bbox": [ + 430, + 327, + 517, + 398 + ], + "type": "image", + "image_path": "4745ce2a8ff41e1922328a7877ce1df3248ff8f67452eb1dcd6bcdb29bbbbc4d.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 75, + 409, + 162, + 486 + ], + "blocks": [ + { + "bbox": [ + 75, + 409, + 162, + 486 + ], + "lines": [ + { + "bbox": [ + 75, + 409, + 162, + 486 + ], + "spans": [ + { + "bbox": [ + 75, + 409, + 162, + 486 + ], + "type": "image", + "image_path": "d76b022c349acfc86c866944a2e04bafd3fb7c3e23db1ea35750c871a1a3ee53.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 487, + 154, + 496 + ], + "lines": [ + { + "bbox": [ + 82, + 487, + 154, + 496 + ], + "spans": [ + { + "bbox": [ + 82, + 487, + 154, + 496 + ], + "type": "text", + "content": "Infrared-visible image" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 164, + 409, + 251, + 486 + ], + "blocks": [ + { + "bbox": [ + 193, + 400, + 228, + 408 + ], + "lines": [ + { + "bbox": [ + 193, + 400, + 228, + 408 + ], + "spans": [ + { + "bbox": [ + 193, + 400, + 228, + 408 + ], + "type": "text", + "content": "PIAFusion" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 409, + 251, + 486 + ], + "lines": [ + { + "bbox": [ + 164, + 409, + 251, + 486 + ], + "spans": [ + { + "bbox": [ + 164, + 409, + 251, + 486 + ], + "type": "image", + "image_path": "995e45e504a43da9738de81ddd817dbb0d2a9253221afe2754e25e0d31ee27b0.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 487, + 226, + 495 + ], + "lines": [ + { + "bbox": [ + 186, + 487, + 226, + 495 + ], + "spans": [ + { + "bbox": [ + 186, + 487, + 226, + 495 + ], + "type": "text", + "content": "FusionGAN" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 253, + 409, + 339, + 486 + ], + "blocks": [ + { + "bbox": [ + 280, + 400, + 320, + 408 + ], + "lines": [ + { + "bbox": [ + 280, + 400, + 320, + 408 + ], + "spans": [ + { + "bbox": [ + 280, + 400, + 320, + 408 + ], + "type": "text", + "content": "SwinFusion" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 409, + 339, + 486 + ], + "lines": [ + { + "bbox": [ + 253, + 409, + 339, + 486 + ], + "spans": [ + { + "bbox": [ + 253, + 409, + 339, + 486 + ], + "type": "image", + "image_path": "c564e8d7527f75731c5f671c6be12c2194ab566ce1646802ac0c70841819cae5.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 487, + 313, + 495 + ], + "lines": [ + { + "bbox": [ + 279, + 487, + 313, + 495 + ], + "spans": [ + { + "bbox": [ + 279, + 487, + 313, + 495 + ], + "type": "text", + "content": "GANMcC" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 341, + 409, + 429, + 486 + ], + "blocks": [ + { + "bbox": [ + 376, + 400, + 399, + 408 + ], + "lines": [ + { + "bbox": [ + 376, + 400, + 399, + 408 + ], + "spans": [ + { + "bbox": [ + 376, + 400, + 399, + 408 + ], + "type": "text", + "content": "Tardal" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 409, + 429, + 486 + ], + "lines": [ + { + "bbox": [ + 341, + 409, + 429, + 486 + ], + "spans": [ + { + "bbox": [ + 341, + 409, + 429, + 486 + ], + "type": "image", + "image_path": "72cc7c78d62de00a53c2e4651e1ca4f4285985d1f1739a464ec971318f4033d5.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 487, + 399, + 495 + ], + "lines": [ + { + "bbox": [ + 374, + 487, + 399, + 495 + ], + "spans": [ + { + "bbox": [ + 374, + 487, + 399, + 495 + ], + "type": "text", + "content": "MFEIF" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 430, + 409, + 517, + 486 + ], + "blocks": [ + { + "bbox": [ + 459, + 400, + 497, + 408 + ], + "lines": [ + { + "bbox": [ + 459, + 400, + 497, + 408 + ], + "spans": [ + { + "bbox": [ + 459, + 400, + 497, + 408 + ], + "type": "text", + "content": "MetaFusion" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 409, + 517, + 486 + ], + "lines": [ + { + "bbox": [ + 430, + 409, + 517, + 486 + ], + "spans": [ + { + "bbox": [ + 430, + 409, + 517, + 486 + ], + "type": "image", + "image_path": "341595a9f2a5131a9d527e2c5eede76f72a6fa7647935fa5594162b178e2c52c.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 487, + 490, + 495 + ], + "lines": [ + { + "bbox": [ + 458, + 487, + 490, + 495 + ], + "spans": [ + { + "bbox": [ + 458, + 487, + 490, + 495 + ], + "type": "text", + "content": "U2Fusion" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 76, + 498, + 162, + 571 + ], + "blocks": [ + { + "bbox": [ + 76, + 498, + 162, + 571 + ], + "lines": [ + { + "bbox": [ + 76, + 498, + 162, + 571 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 162, + 571 + ], + "type": "image", + "image_path": "d6cb12c5541ad10eafda0fc1f0cff1d005ec662d214ed311e581e5489272e1dd.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 573, + 129, + 581 + ], + "lines": [ + { + "bbox": [ + 107, + 573, + 129, + 581 + ], + "spans": [ + { + "bbox": [ + 107, + 573, + 129, + 581 + ], + "type": "text", + "content": "YDTR" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 589, + 533, + 601 + ], + "lines": [ + { + "bbox": [ + 59, + 589, + 533, + 601 + ], + "spans": [ + { + "bbox": [ + 59, + 589, + 533, + 601 + ], + "type": "text", + "content": "Figure 6. Qualitative results of different fusion methods on " + }, + { + "bbox": [ + 59, + 589, + 533, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 59, + 589, + 533, + 601 + ], + "type": "text", + "content": " (first group), RoadScene (second group) and TNO (third group)." + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 164, + 498, + 251, + 571 + ], + "blocks": [ + { + "bbox": [ + 164, + 498, + 251, + 571 + ], + "lines": [ + { + "bbox": [ + 164, + 498, + 251, + 571 + ], + "spans": [ + { + "bbox": [ + 164, + 498, + 251, + 571 + ], + "type": "image", + "image_path": "924e4f0a226b7a4935fb85fedb606e19bb232cd9ee7c74cb84ebf060146d90be.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 573, + 228, + 581 + ], + "lines": [ + { + "bbox": [ + 192, + 573, + 228, + 581 + ], + "spans": [ + { + "bbox": [ + 192, + 573, + 228, + 581 + ], + "type": "text", + "content": "PIAFusion" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 253, + 498, + 339, + 571 + ], + "blocks": [ + { + "bbox": [ + 253, + 498, + 339, + 571 + ], + "lines": [ + { + "bbox": [ + 253, + 498, + 339, + 571 + ], + "spans": [ + { + "bbox": [ + 253, + 498, + 339, + 571 + ], + "type": "image", + "image_path": "691b89cf755bf1359f60d8cefd5c3d54770854d45b616b9767bb33f07867267e.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 573, + 319, + 581 + ], + "lines": [ + { + "bbox": [ + 279, + 573, + 319, + 581 + ], + "spans": [ + { + "bbox": [ + 279, + 573, + 319, + 581 + ], + "type": "text", + "content": "SwinFusion" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 341, + 498, + 428, + 571 + ], + "blocks": [ + { + "bbox": [ + 341, + 498, + 428, + 571 + ], + "lines": [ + { + "bbox": [ + 341, + 498, + 428, + 571 + ], + "spans": [ + { + "bbox": [ + 341, + 498, + 428, + 571 + ], + "type": "image", + "image_path": "6bc5db959b22a9872bcb6f684e4368999ba3a587ec0f7313aeeb397b62f5964b.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 573, + 399, + 581 + ], + "lines": [ + { + "bbox": [ + 376, + 573, + 399, + 581 + ], + "spans": [ + { + "bbox": [ + 376, + 573, + 399, + 581 + ], + "type": "text", + "content": "Tardal" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 430, + 498, + 517, + 571 + ], + "blocks": [ + { + "bbox": [ + 430, + 498, + 517, + 571 + ], + "lines": [ + { + "bbox": [ + 430, + 498, + 517, + 571 + ], + "spans": [ + { + "bbox": [ + 430, + 498, + 517, + 571 + ], + "type": "image", + "image_path": "685a7fef49377996e8219b74c7cfedb1be499bbbcb1a73d9b491d15de8cc7c2b.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 573, + 496, + 581 + ], + "lines": [ + { + "bbox": [ + 457, + 573, + 496, + 581 + ], + "spans": [ + { + "bbox": [ + 457, + 573, + 496, + 581 + ], + "type": "text", + "content": "MetaFusion" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "bbox": [ + 47, + 610, + 281, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 281, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 281, + 623 + ], + "type": "text", + "content": "adopted to generate fusion results for the fair comparison." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "Qualitative results of different fusion methods are shown in Figure 6. All the fusion methods can fuse the main features of the infrared and visible images to some extent. However, FusionGAN, GANMcC, MFEIF and YDTR produce blurred edge details, and U2Fusion, PIAFusion, Swinfusion and Tardal generate low contrast objects, as shown in the red rectangular boxes. In contrast, the images gener" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 305, + 610, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 545, + 634 + ], + "type": "text", + "content": "ated by the proposed MetaFusion contain clear edge details and high contrast objects." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": "Sequentially, we provide quantitative results of different fusion methods in Table 4. Our MetaFusion generally achieves the largest or the second-largest metric values. In detail, the high EN and MI illustrate the fusion images by our MetaFusion contain high contrast object and clear edge details. The large VIF shows our fusion results have high-" + } + ] + } + ], + "index": 64 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13961" + } + ] + } + ], + "index": 65 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 99, + 535, + 267 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 4. Quantitative results of different fusion methods on TNO, RoadScene and " + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": " datasets. The model inference time is counted on a NVIDIA GeForce RTX 2080 Ti. The best result is in red and the second best one is in violet." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 99, + 535, + 267 + ], + "lines": [ + { + "bbox": [ + 57, + 99, + 535, + 267 + ], + "spans": [ + { + "bbox": [ + 57, + 99, + 535, + 267 + ], + "type": "table", + "html": "
MethodM3FDRoadSceneTNO40Time (s)
MIEntVIFMIEntVIFMIEntVIF
FusionGAN [25]13.4456.7220.30314.2037.1010.25113.0686.5340.2520.040
GANMcC [26]13.7316.8650.45314.0177.0080.42213.4856.7420.4240.081
MFEIF [21]12.9576.4780.40113.4896.7420.26013.3606.6800.3950.029
U2Fusion [44]13.8166.9080.54513.8886.9440.45613.8896.9440.6360.043
YDTR [36]12.6946.3470.36113.1666.5830.23612.8626.4310.2800.086
PIAFusion [35]13.3266.6630.43713.2626.5330.20513.9316.9600.5060.055
SwinFusion [23]13.0516.5250.46013.2676.6330.31313.8796.9330.4791.081
Tardal [20]13.6366.8180.65015.0097.5040.48313.0306.5150.9100.030
MetaFusion14.5117.2491.51514.2187.0220.96914.6577.3231.4620.015
", + "image_path": "e36083457f362deb290572f174b9eae2df103aaf34e80f70ce6cc64bfdfc7f6c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 55, + 281, + 114, + 323 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 114, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 114, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 114, + 323 + ], + "type": "image", + "image_path": "fdcd42e5b93865b1136dd887f9ab32722b6d75e9c7be68926411453fa17f5d3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 115, + 281, + 172, + 323 + ], + "blocks": [ + { + "bbox": [ + 115, + 281, + 172, + 323 + ], + "lines": [ + { + "bbox": [ + 115, + 281, + 172, + 323 + ], + "spans": [ + { + "bbox": [ + 115, + 281, + 172, + 323 + ], + "type": "image", + "image_path": "c22264dbdd5ab373084f0a61ed0876b143d758176fdc5aad3b9e879640f97567.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 173, + 281, + 230, + 323 + ], + "blocks": [ + { + "bbox": [ + 173, + 281, + 230, + 323 + ], + "lines": [ + { + "bbox": [ + 173, + 281, + 230, + 323 + ], + "spans": [ + { + "bbox": [ + 173, + 281, + 230, + 323 + ], + "type": "image", + "image_path": "eb61026d2b5b36f7323ce85006915eae8ed2e73cea3ebe1e1e5353ddc7c62074.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 230, + 281, + 290, + 323 + ], + "blocks": [ + { + "bbox": [ + 230, + 281, + 290, + 323 + ], + "lines": [ + { + "bbox": [ + 230, + 281, + 290, + 323 + ], + "spans": [ + { + "bbox": [ + 230, + 281, + 290, + 323 + ], + "type": "image", + "image_path": "9b1f7750091301e96c5ff96b59ba8243280f6a92a8868adf0783f3bbe6057ca1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 54, + 325, + 114, + 370 + ], + "blocks": [ + { + "bbox": [ + 54, + 325, + 114, + 370 + ], + "lines": [ + { + "bbox": [ + 54, + 325, + 114, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 325, + 114, + 370 + ], + "type": "image", + "image_path": "7dbcd590492a89d29c05ed3e2ae000c8d379d9e62ccaafd73fa44c3eed1c48c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 115, + 325, + 172, + 370 + ], + "blocks": [ + { + "bbox": [ + 115, + 325, + 172, + 370 + ], + "lines": [ + { + "bbox": [ + 115, + 325, + 172, + 370 + ], + "spans": [ + { + "bbox": [ + 115, + 325, + 172, + 370 + ], + "type": "image", + "image_path": "fa46622c4c836260e003e102d1cbe4cb775ce2450a09639e2d66e7107a073779.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 173, + 325, + 230, + 370 + ], + "blocks": [ + { + "bbox": [ + 173, + 325, + 230, + 370 + ], + "lines": [ + { + "bbox": [ + 173, + 325, + 230, + 370 + ], + "spans": [ + { + "bbox": [ + 173, + 325, + 230, + 370 + ], + "type": "image", + "image_path": "feccfbfe8c5ddbf158ec44996b51a076155566fae41d1f2f167bf02f8769a664.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 231, + 325, + 290, + 370 + ], + "blocks": [ + { + "bbox": [ + 231, + 325, + 290, + 370 + ], + "lines": [ + { + "bbox": [ + 231, + 325, + 290, + 370 + ], + "spans": [ + { + "bbox": [ + 231, + 325, + 290, + 370 + ], + "type": "image", + "image_path": "a70e72e5c666c9755bb36dd2b788b5c04598f690e5b5c79a850ab0eedb68e27c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 54, + 376, + 114, + 418 + ], + "blocks": [ + { + "bbox": [ + 54, + 376, + 114, + 418 + ], + "lines": [ + { + "bbox": [ + 54, + 376, + 114, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 114, + 418 + ], + "type": "image", + "image_path": "e9454166dc21ce7a57c98c07515a73a40df63984fc7542c60c469158e2dc28f3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 419, + 96, + 423 + ], + "lines": [ + { + "bbox": [ + 73, + 419, + 96, + 423 + ], + "spans": [ + { + "bbox": [ + 73, + 419, + 96, + 423 + ], + "type": "text", + "content": "SwinFusion" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 115, + 372, + 172, + 418 + ], + "blocks": [ + { + "bbox": [ + 115, + 372, + 172, + 418 + ], + "lines": [ + { + "bbox": [ + 115, + 372, + 172, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 172, + 418 + ], + "type": "image", + "image_path": "9d5c4172d94e9654594d1425b91a4acddc3c2237a424f62aabcf6eb63e01bce3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 418, + 150, + 423 + ], + "lines": [ + { + "bbox": [ + 137, + 418, + 150, + 423 + ], + "spans": [ + { + "bbox": [ + 137, + 418, + 150, + 423 + ], + "type": "text", + "content": "Tardal" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 173, + 376, + 230, + 418 + ], + "blocks": [ + { + "bbox": [ + 173, + 376, + 230, + 418 + ], + "lines": [ + { + "bbox": [ + 173, + 376, + 230, + 418 + ], + "spans": [ + { + "bbox": [ + 173, + 376, + 230, + 418 + ], + "type": "image", + "image_path": "a14122c348be79ee8a9a3f7a31e946ec3789bb650e339f0ec8cff17e3f314c82.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 418, + 213, + 423 + ], + "lines": [ + { + "bbox": [ + 191, + 418, + 213, + 423 + ], + "spans": [ + { + "bbox": [ + 191, + 418, + 213, + 423 + ], + "type": "text", + "content": "MetaFusion" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 432, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 453 + ], + "type": "text", + "content": "Figure 7. Visual results of object detection based on different fusion methods on " + }, + { + "bbox": [ + 46, + 432, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 231, + 376, + 290, + 418 + ], + "blocks": [ + { + "bbox": [ + 231, + 376, + 290, + 418 + ], + "lines": [ + { + "bbox": [ + 231, + 376, + 290, + 418 + ], + "spans": [ + { + "bbox": [ + 231, + 376, + 290, + 418 + ], + "type": "image", + "image_path": "a31329ecfc5325b98c81c63f5f364718a68cb1bbffd94430a92dc9cc429a3fbe.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 418, + 272, + 423 + ], + "lines": [ + { + "bbox": [ + 248, + 418, + 272, + 423 + ], + "spans": [ + { + "bbox": [ + 248, + 418, + 272, + 423 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "type": "text", + "content": "quality visual effect and small distortion compared with the source images. Moreover, our method achieves the fastest inference time that only needs 0.015s to generate one fusion result." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 515, + 287, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 287, + 539 + ], + "type": "text", + "content": "4.4. Evaluation on Infrared-visible Object Detection" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 546, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 630 + ], + "type": "text", + "content": "A better fusion image can provide more information to improve object detection. Here, we verify the effectiveness of our method by evaluating object detection accuracy with the infrared and visible fusion images. To make a fair comparison, we firstly generate fusion results using our method and SOTA methods, and then use the results to retrain the object detection baseline YOLOv5s, respectively." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Figure 7 shows the object detection results. In general, the SOTA methods improve the performance of object detection. In contrast, our MetaFusion achieves a better performance, e.g., the person is accurately detected (see the blue ellipse). Table 5 shows the quantitative results. Our fusion results help the detection network achieve the highest object detection accuracy. This further proves that our" + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 339, + 318, + 512, + 474 + ], + "blocks": [ + { + "bbox": [ + 305, + 281, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 314 + ], + "type": "text", + "content": "Table 5. Quantitative results of object detection based on different fusion methods on " + }, + { + "bbox": [ + 305, + 281, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 305, + 281, + 545, + 314 + ], + "type": "text", + "content": " dataset. The best result is in red and the second best one is in violet." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 318, + 512, + 474 + ], + "lines": [ + { + "bbox": [ + 339, + 318, + 512, + 474 + ], + "spans": [ + { + "bbox": [ + 339, + 318, + 512, + 474 + ], + "type": "table", + "html": "
Method\\( \\mathrm{mAP}_{50\\rightarrow 95}(\\%) \\)
FusionGAN [25]54.2
GANMcC [26]55.2
MFEIF [21]55.4
U2Fusion [44]55.7
YDTR [36]55.4
PIAFusion [35]55.6
SwinFusion [23]55.4
Tardal [20]54.4
MetaFusion56.5
", + "image_path": "083a47be8efac396a40d841af152634be033863291b3d8932a9c5fd9ba49f0d9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 485, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 545, + 509 + ], + "type": "text", + "content": "method can generate high-quality fusion results especially for the objects." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 523, + 379, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 379, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 379, + 536 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 544, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 675 + ], + "type": "text", + "content": "This paper presents a joint fusion and detection learning framework through introducing the meta-feature embedding model. Based on the meta learning idea, the meta-feature embedding model can generate object semantic features according to the fusion network ability, thereby bridging the feature gap between these two different-level tasks. Moreover, a mutual promotion learning between fusion and detection tasks is further implemented to improve their performances. Both quantitative and qualitative results demonstrate the superior performance of our method compared with the state-of-the-art methods." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by National Natural Science Foundation of China under Grant Nos. 62176038, 62001450 and U1903215." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13962" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9465-9474, October 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "text", + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6154-6162, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[4] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-sclil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14166-14175, June 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "type": "text", + "content": "[5] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "text", + "content": "[6] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster R-CNN: towards accurate few-shot object detection with attentive feature alignment. In AAAI, pages 780-789. AAAI Press, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "text", + "content": "[7] Yu Han, Yunze Cai, Yin Cao, and Xiaoming Xu. A new image fusion performance metric based on visual information fidelity. Information Fusion, 14:-, 04 2013. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 468 + ], + "type": "text", + "content": "[8] Kaiming He, Georgia Gkioxari, Piotr Dollr, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 512 + ], + "type": "text", + "content": "[9] Liqiang He and Sinisa Todorovic. Destr: Object detection with split transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9367-9376, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "text", + "content": "[10] Ruichao Hou, Dongming Zhou, Rencan Nie, Dong Liu, Lei Xiong, Yanbu Guo, and Chuanbo Yu. Vif-net: An unsupervised framework for infrared and visible image fusion. IEEE Transactions on Computational Imaging, 6:640-651, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "text", + "content": "[11] Muhammad Abdullah Jamal, Liqiang Wang, and Boqing Gong. A lazy approach to long-horizon gradient-based meta-learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6577-6586, October 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "text", + "content": "[12] Huafeng Li, Yueliang Cen, Yu Liu, Xun Chen, and Zhengtao Yu. Different input resolutions and arbitrary output resolution: A meta learning-based deep framework for infrared and visible image fusion. IEEE Transactions on Image Processing, 30:4070-4083, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Hui Li, Xiao-Jun Wu, and Josef Kittler. Mdlatrr: A novel decomposition method for infrared and visible image fusion. IEEE Transactions on Image Processing, 29:4733-4746, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[15] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[16] Jing Li, Jianming Zhu, Chang Li, Xun Chen, and Bin Yang. Cgtf: Convolution-guided transformer for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 71:1-14, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 252 + ], + "type": "text", + "content": "[17] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP:1-1, 07 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 299, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 364 + ], + "type": "text", + "content": "[19] Benlin Liu, Yongming Rao, Jiwen Lu, Jie Zhou, and Chojui Hsieh. Metadistiller: Network self-boosting via metalearned top-down distillation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16, pages 694–709. Springer, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 441 + ], + "type": "text", + "content": "[20] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[21] J. Liu, X. Fan, J. Jiang, R. Liu, and Z. Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1, 2021. 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "text", + "content": "[22] Yu Liu, Shuping Liu, and Zengfu Wang. A general framework for image fusion based on multi-scale transform and s-parse representation. Information Fusion, 24:147-164, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 534, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 588 + ], + "type": "text", + "content": "[23] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 2, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "type": "text", + "content": "[24] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping Zhang. Ddcgan: A dual-discriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29:4980-4995, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "text", + "content": "[25] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusion: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11-26, 08 2019. 1, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[26] Jiayi Ma, Hao Zhang, Zhenfeng Shao, Pengwei Liang, and Han Xu. Ganmcc: A generative adversarial network with" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13963" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "multiclassification constraints for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1-14, 2021. 2, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[27] Yan Mo, Xudong Kang, Puhong Duan, Bin Sun, and Shutao Li. Attribute filter based infrared and visible image fusion. Information Fusion, 75:41-54, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "text", + "content": "[28] Guihong Qu, Dali Zhang, and Pingfan Yan. Information measure for performance of image fusion. *Electronics Letters*, 38:313-315, 04 2002. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[29] Aniruddh Raghu, Jonathan Lorraine, Simon Kornblith, Matthew McDermott, and David K Duvenaud. Meta-learning to improve pre-training. Advances in Neural Information Processing Systems, 34:23231-23244, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 261 + ], + "type": "text", + "content": "[30] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "type": "text", + "content": "[32] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 658-666, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "text", + "content": "[33] Wesley Roberts, Jan van Aardt, and Fethi Ahmed. Assessment of image fusion procedures using entropy, image quality, and multispectral classification. Journal of Applied Remote Sensing, 2:1-28, 05 2008. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 491 + ], + "type": "text", + "content": "[34] Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshitaka Ushiku, and Tatsuya Harada. Multispectral object detection for autonomous vehicles. In Proceedings of the on Thematic Workshops of ACM Multimedia 2017, Thematic Workshops '17, page 3543, New York, NY, USA, 2017. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[35] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83-84:79-92, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "text", + "content": "[36] Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer. IEEE Transactions on Multimedia, pages 1-16, 2022. 2, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 591 + ], + "type": "text", + "content": "[37] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249-251, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[38] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[39] Zhishe Wang, Yuanyuan Wu, Junyao Wang, Jiawei Xu, and Wenyu Shao. Res2fusion: Infrared and visible image fusion based on dense res2net and double nonlocal attention models. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[40] Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. In" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14248-14257, June 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 106, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 172 + ], + "type": "text", + "content": "[41] Chengming Xu, Yanwei Fu, Chen Liu, Chengjie Wang, Jilin Li, Feiyue Huang, Li Zhang, and Xiangyang Xue. Learning dynamic alignment via meta-filter for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5182-5191, June 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 218 + ], + "type": "text", + "content": "[42] Dongli Xu, Jinghong Deng, and Wen Li. Revisiting ap loss for dense object detection: Adaptive ranking pair selection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 14187-14196, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "text", + "content": "[43] Han Xu, Meiqi Gong, Xin Tian, Jun Huang, and Jiayi Ma. Cufd: An encoderdecoder network for visible and infrared image fusion based on common and unique feature decomposition. Computer Vision and Image Understanding, 218:103407, 03 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "type": "text", + "content": "[44] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "text", + "content": "[45] Han Xu, Xinya Wang, and Jiayi Ma. Drf: Disentangled representation for visible and infrared image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1-13, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "type": "text", + "content": "[46] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Rui-Wei Zhao, Tao Zhang, Xuequan Lu, and Shang Gao. Cream: Weakly supervised object localization via class re-activation mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9437–9446, June 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 431, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 464 + ], + "type": "text", + "content": "[47] Meilong Xu, Linfeng Tang, Hao Zhang, and Jiayi Ma. Infrared and visible image fusion via parallel scene and texture learning. Pattern Recognition, 132:108929, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 466, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 521 + ], + "type": "text", + "content": "[48] Shilong Zhang, Zhuoran Yu, Liyang Liu, Xinjiang Wang, Aojun Zhou, and Kai Chen. Group r-cnn for weakly semi-supervised object detection with points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9417-9426, June 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[49] Fan Zhao and Wenda Zhao. Learning specific and general realm feature representations for image fusion. IEEE Transactions on Multimedia, 23:2745-2756, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "type": "text", + "content": "[50] Fan Zhao, Wenda Zhao, and Hutchuan Lu. Interactive feature embedding for infrared and visible image fusion. arXiv preprint arXiv:2211.04877, 2022. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[51] Fan Zhao, Wenda Zhao, Libo Yao, and Yu Liu. Self-supervised feature adaption for infrared and visible image fusion. Information Fusion, 76:189-203, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[52] Wenda Zhao, Xueqing Hou, You He, and Huchuan Lu. Defocus blur detection via boosting diversity of deep ensemble networks. IEEE Transactions on Image Processing, 30:5426-5438, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[53] Wenda Zhao, Huimin Lu, and Dong Wang. Multisensor image fusion and enhancement in spectral total variation domain. IEEE Transactions on Multimedia, 20(4):866-879, 2018. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13964" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[54] Wenda Zhao, Dong Wang, and Huchuan Lu. Multi-focus image fusion with a natural enhancement via a joint multi-level deeply supervised convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 29(4):1102-1115, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[55] Wenda Zhao, Bowen Zheng, Qiuhua Lin, and Hutchuan Lu. Enhancing diversity of defocus blur detectors via cross-ensemble network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8905–8913, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "text", + "content": "[56] Zhengjie Zhu, Xiaogang Yang, Ruitao Lu, Tong Shen, Xueli Xie, and Tao Zhang. Clf-net: Contrastive learning for infrared and visible image fusion network. IEEE Transactions on Instrumentation and Measurement, 71:1-15, 2022. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13965" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_content_list.json b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f494f9383d1e3c1000524bfee7a8182e3438caa3 --- /dev/null +++ b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_content_list.json @@ -0,0 +1,1507 @@ +[ + { + "type": "text", + "text": "MetaMix: Towards Corruption-Robust Continual Learning with Temporally Self-Adaptive Data Transformation", + "text_level": 1, + "bbox": [ + 158, + 128, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenyi Wang $^{1}$ Li Shen $^{2,*}$ Donglin Zhan $^{3}$ Qiuling Suo $^{1}$", + "bbox": [ + 235, + 202, + 730, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanjun Zhu $^{1}$ Tiehang Duan $^{4}$ Mingchen Gao $^{1}$", + "bbox": [ + 279, + 220, + 687, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State University of New York at Buffalo, USA; $^{2}$ JD Explore Academy, China", + "bbox": [ + 178, + 239, + 792, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Columbia University, USA; 4 Meta, USA", + "bbox": [ + 315, + 256, + 653, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhenyiwa, qiulings, yjzhu, mgao8}@buffalo.edu; dz2478@columbia.edu", + "bbox": [ + 189, + 277, + 781, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mathshenli, tiehang.duan}@gmail.com", + "bbox": [ + 320, + 295, + 638, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Continual Learning (CL) has achieved rapid progress in recent years. However, it is still largely unknown how to determine whether a CL model is trustworthy and how to foster its trustworthiness. This work focuses on evaluating and improving the robustness to corruptions of existing CL models. Our empirical evaluation results show that existing state-of-the-art (SOTA) CL models are particularly vulnerable to various data corruptions during testing. To make them trustworthy and robust to corruptions deployed in safety-critical scenarios, we propose a meta-learning framework of self-adaptive data augmentation to tackle the corruption robustness in CL. The proposed framework, MetaMix, learns to augment and mix data, automatically transforming the new task data or memory data. It directly optimizes the generalization performance against data corruptions during training. To evaluate the corruption robustness of our proposed approach, we construct several CL corruption datasets with different levels of severity. We perform comprehensive experiments on both task- and class-continual learning. Extensive experiments demonstrate the effectiveness of our proposed method compared to SOTA baselines.", + "bbox": [ + 75, + 376, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 726, + 209, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans constantly acquire new information throughout their lifespan and easily recognize information shifts such as structure and style variations in images. Continual learning (CL) aims at imitating human's ability to learn from non-stationary data distributions without forgetting the previously learned knowledge. The past few years have witnessed rapid progress in CL research [1, 30, 31, 36, 45]. Despite the success, existing CL systems overlook the robustness", + "bbox": [ + 75, + 752, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "against unforeseen data shifts during testing. They assume that training and test images for each task follow the same distribution. However, as data distributions evolve and new scenarios occur, test images often encounter various corruptions such as snow, blur, pixelation, and combinations, resulting in a shifted distribution from the training set. For example, Figure 1 shows various corruptions applied on one image from Split-miniImageNet.", + "bbox": [ + 500, + 345, + 893, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data corruption can drastically impair the performance of existing image recognition systems. A recent study [21] shows that classification accuracy of various architectures has dropped significantly on the ImageNet test set with some simple and natural corruptions. Our empirical evaluation results show that state-of-the-art CL models are even more vulnerable to these corruptions during testing. For example, the accuracy of $\\mathrm{DER}++$ [5] for task-continual learning decreases from $93.9\\%$ to $50.5\\%$ on split-CIFAR10; accuracy drops to $10.6\\%$ from $75.6\\%$ on split-CIFAR100; accuracy decreases to $9.8\\%$ from $61.3\\%$ on split-miniImageNet by applying those common corruptions on test data of each CL task. This severe issue makes existing CL models highly unreliable in safety-critical applications. Thus, improving the robustness of CL models to foster their trustworthiness when deployed in real-world scenarios is essential.", + "bbox": [ + 496, + 472, + 893, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training a CL model robust to various corruptions is difficult due to the following challenges. 1) Unseen corruptions could perturb the test set far beyond those encountered during training. A model that naively augments training images with seen corruptions cannot generalize to the new ones during testing. Also, it is unrealistic to enumerate all possible corruptions during training since there are infinite types of corruptions and their combinations. 2) With the ever-evolving data distributions in CL, an effective augmentation strategy learned on previous tasks may gradually become less effective because the optimal augmentation strategies are task-dependent and dynamically change over tasks [11].", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24521", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7018865a0779f5e2d6097ceeab51aaa398a576f5adcca44e96478509a63043cb.jpg", + "image_caption": [ + "(a) Original data" + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 248, + 180 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bd248e56f61825c5094a381447e681e063e85f8458e33af3352119463e6883e4.jpg", + "image_caption": [ + "(b) Brightness" + ], + "image_footnote": [], + "bbox": [ + 250, + 88, + 364, + 180 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/37e78b46c73ed97a17a6757dc2f0e68afbceb916b50f86b9c6442501111f50c9.jpg", + "image_caption": [ + "(c) Contrast", + "Figure 1. Visualization of five types of different corruption operations on the testing images of Split-miniImageNet." + ], + "image_footnote": [], + "bbox": [ + 367, + 88, + 478, + 180 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8a957671ca9433a669081fd58d90e4e56156bdfb908584d1921f399c859f7997.jpg", + "image_caption": [ + "(d) Defocus Blur" + ], + "image_footnote": [], + "bbox": [ + 478, + 88, + 591, + 180 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe0ece7366514591855722284a7840932d204682a7eadf66ea104892bafff090.jpg", + "image_caption": [ + "(e) Elastic" + ], + "image_footnote": [], + "bbox": [ + 591, + 89, + 720, + 180 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6fb2639880539c52c8ee1db674f88b42d5a135973c64dafd693bb4e2062e627d.jpg", + "image_caption": [ + "(f) Fog" + ], + "image_footnote": [], + "bbox": [ + 722, + 89, + 852, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although we can adopt a memory buffer to store data from previous tasks, augmenting and replaying them at later training iterations, the augmented memory may gradually become less effective as the model could memorize the stored information after replay runs. Recent approaches, such as Augmix [22] composes and combines multiple pre-defined augmentation operations with different depths and widths, show efficacy in improving robustness under traditional supervised classification tasks. However, these approaches are not directly applicable to corruption-robust CL since they only use a fixed random augmentation strategy that is often not optimal for non-stationary data distributions in CL.", + "bbox": [ + 75, + 226, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these unique challenges of corruption-robustness in CL, we propose a temporally self-adaptive Augmix within a meta-learning framework, named MetaMix. It adaptively augments the memory buffer data and the currently received new data by learning to mix the augmentation operations tailored to the evolving data distributions. In particular, our automatic self-adaptive MetaMix is a bi-level optimization, simulating the evaluation process on unseen corruptions. We randomly divide the training augmentation operations into pseudo-seen and pseudo-unseen operations at each CL step. The lower-level optimization is to optimize the model performance on the pseudo-seen operations; the upper-level optimization is to optimize the generalization of the pseudo-unseen operations. The augmentation strategy is governed by an LSTM, which inputs context information and outputs the corresponding mixing parameters for the augmentations. The proposed MetaMix ensures the augmentation strategy automatically adapts to non-stationary data distribution. Furthermore, the objective is to optimize the performance of the pseudo-unseen corruption operations, which aligns with our goal during testing and encourages the generalization to unseen corruptions.", + "bbox": [ + 75, + 412, + 472, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the corruption robustness of existing and the proposed methods, we propose a new challenging benchmark where various corruptions perturb the testing data of each CL task. To facilitate future research, we construct several new datasets, including split-CIFAR-10-C, split-CIFAR-100-C, and split-miniImageNet-C. Extensive experiments on the constructed benchmarks demonstrate the effectiveness of our proposed MetaMix approach compared with several SOTA data-augmentation approaches adapted for CL. We summarize our contributions as follows:", + "bbox": [ + 75, + 750, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our best knowledge, we are the first to study the corruption-robustness of CL methods. Accordingly, we propose the first set of novel benchmarks for evaluating the corruption-robustness of existing CL methods and moving towards trustworthy CL.", + "- We propose a self-adaptive augmentation method, MetaMix, by learning to mix and augment the training data of each CL task to achieve corruption-robustness on unseen corruptions for each CL task.", + "- Our method is versatile and can be seamlessly integrated with existing CL methods. Extensive experiments with both task/class continual learning demonstrate the effectiveness of MetaMix." + ], + "bbox": [ + 517, + 226, + 893, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 459, + 640, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Continual Learning", + "text_level": 1, + "bbox": [ + 500, + 484, + 687, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CL focuses on learning non-stationary data distribution without forgetting previous knowledge. These methods can be categorized into: 1) retaining memory for future replay [1,4,8,9,17,31,35,42,46,48,49]; 2) designing tailored network architectures [14,24,37,41,54]; 3) performing proper regularization during parameter updates [7,26,40,56]; 4) introducing Bayesian methods for model parameter inference [13,23,34,53]; 5) subspace-based methods [27,30,38] and 6) prompt-tuning [29] based methods [50,51]; 7) transformer-based methods [47,52]. However, all these works focus on the simplified setting in which training and testing data of each CL task follow the same distribution. In practice, the testing data may come from another different distribution due to the variations in weather, light, snow, etc. These distortions pose significant robustness challenges for existing CL methods. Existing literature largely ignored these trustworthy properties of CL. According to the terminology of [43], CL settings can be categorized into three scenarios based on whether task identities of the testing data are available to the CL model during test time. Among the three scenarios, task-continual learning studies the scenario that the task identities are available during testing and is the easiest learning setting; domain incremental learning focuses on the case that all the tasks during CL solve the same classification task with different data distribution and the task identities are unknown during testing; class-continual learning focuses", + "bbox": [ + 496, + 508, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24522", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on the scenario that task identities are unavailable and is the most difficult learning setting. We focus on task/class CL scenarios.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, robust CL began to attract researchers' attention. Self-Purified Replay [25] and PuriDivER [3] consider CL on noisy data stream with label noise. Their work is orthogonal to ours since they consider noisy data during training; we instead consider the robustness to data corruptions during testing. A closely related work to our is Continual Active Adaptation (CAA) [32], which is different from and orthogonal to ours since they focus on the setting that for each data distribution, the label space for each learning task is the same. Each task only encounters a single data corruption. During testing, CAA focuses on generalization to previously seen corruption operations. In contrast, we focus on the setting that each task solves a different task. During testing, the CL model should be robust to unseen data corruptions.", + "bbox": [ + 75, + 138, + 470, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Data Augmentation", + "text_level": 1, + "bbox": [ + 76, + 367, + 264, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We categorize existing data augmentation techniques into two classes: 1) improving performance on clean data; 2) improving performance and being robust to data distribution shift and corruptions.", + "bbox": [ + 75, + 393, + 468, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Augmentations for Improving Clean Data Performance. Standard data augmentation aims to improve the generalization performance on clean data. To improve the performance of neural network, standard supervised learning adopts common augmentations, such as random flips, rotations, and crops. Recently, Cutout [12], Mixup [57], CutMix [55] and MaxUp [16] show excellent performance on clean datasets. Other techniques include automatically learning augmentation policy from data, e.g., AutoAugment [10] and RandAugment [11]. However, AutoAugment is computationally impractical for CL problems since it needs a huge computation cost and separate proxy tasks to train the policy network.", + "bbox": [ + 75, + 455, + 470, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Augmentations for Being Robust to Corruptions. Augmix [22] is one of the SOTA methods on the data corruption benchmarks, such as ImageNet-C, CIFAR100-C, etc. DeepAugment [20] and AdversarialAugment [6] transform and augment clean images by perturbing the representations/features of neural networks to generate new augmented images. However, this increases a lot of computation costs. Thus, DeepAugment [20] and AdversarialAugment [6] are impractical for CL. We thus choose the more efficient and effective AugMix and propose a new self-adaptive AugMix for CL.", + "bbox": [ + 75, + 640, + 470, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 811, + 210, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first define the corruption robustness of CL and then describe the proposed MetaMix approach. We describe the notation in the main text. We also provide a notation table in Table 4 in Appendix for ease of reading.", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 500, + 90, + 651, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Corruption robustness of CL. We consider the problem of learning a sequence of tasks denoted as $\\mathcal{D}^{tr} = \\{\\mathcal{D}_1^{tr},\\mathcal{D}_2^{tr},\\dots ,\\mathcal{D}_N^{tr}\\}$ , where $N$ is the number of training tasks. The $k$ -th task training data $\\mathcal{D}_k^{tr}$ consists of many labeled examples $(\\pmb{x}^k,y^k,\\mathcal{T}_k)$ , where $\\pmb{x}^k$ is the data example in the task, $y^{k}$ is the corresponding data label, and $\\mathcal{T}_k$ is the task identifier. The goal is to learn a model $f_{\\theta}$ on the training task sequence $\\mathcal{D}^{tr}$ so that it performs well on the testing set of all the learned tasks $\\mathcal{D}^{te} = \\{\\mathcal{D}_1^{te},\\mathcal{D}_2^{te},\\dots ,\\mathcal{D}_N^{te}\\}$ without forgetting previously learned knowledge. In particular, we focus on improving the model robustness on the testing set for each task $\\mathcal{T}_k$ under a collection of corruption operations $\\mathcal{C}$ . Below, we define the CL corruption robustness.", + "bbox": [ + 498, + 114, + 893, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (CL corruption robustness). Given a sequence of $N$ tasks with test data $\\mathcal{D}^{te} = \\{\\mathcal{D}_1^{te},\\mathcal{D}_2^{te},\\dots ,\\mathcal{D}_N^{te}\\}$ , and a collection of corruption operations $\\mathcal{C}$ , the corruption robustness is the expectation accuracy after applying corruptions $\\mathcal{C}$ on the test data of each task $k$ , i.e.,", + "bbox": [ + 498, + 323, + 893, + 398 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA C C = \\frac {1}{N} \\sum_ {k = 1} ^ {k = N} \\mathbb {E} _ {c \\sim \\mathcal {C}} \\left[ \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {k}, y ^ {k}, \\mathcal {T} _ {k}\\right) \\sim D _ {k} ^ {t e}} \\mathbb {I} \\left(f \\left(c \\left(\\boldsymbol {x} ^ {k}\\right)\\right) = y ^ {k}\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 411, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c(\\pmb{x}^k)$ is the resulting testing data after applying the corruption operation $c$ on $\\pmb{x}^k$ . $\\mathbb{I}(.)$ is the indicator function. ACC measures the average robust accuracy after applying a collection of corruption operations $\\mathcal{C}$ .", + "bbox": [ + 498, + 464, + 893, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary of AugMix. AugMix [22] is a recently proposed SOTA augmentation technique that significantly improves model robustness in standard supervised learning. Considering that CL requires the algorithm to be both efficient and effective, we thus choose the efficient and effective AugMix to adapt to CL. AugMix maintains augmented data diversity by mixing several augmentation chains. Suppose we have an augmentation operation set $\\mathcal{O} = \\{op_1, op_2, \\dots, op_M\\}$ during training. The standard AugMix implementation samples three operations $op_1$ , $op_2$ and $op_3$ from $\\mathcal{O}$ , and composes them with varying depths to form several operation chains; where each chain consists of one to three randomly selected augmentation operations. The augmented image is a convex mix of $A$ chains ( $A$ denotes the number of augmentation chains), i.e., $\\boldsymbol{x}_{aug} = w_1 \\cdot \\text{chain}_1(\\boldsymbol{x}) + w_2 \\cdot \\text{chain}_2(\\boldsymbol{x}) + \\dots + w_A \\cdot \\text{chain}_A(\\boldsymbol{x})$ , where mixing weights $\\boldsymbol{w} = (w_1, w_2, \\dots, w_A)$ is sampled from the distribution, Drichelet $(\\alpha, \\alpha, \\dots, \\alpha)$ with $\\alpha = 1$ . With the mixed image $\\boldsymbol{x}_{aug}$ , AugMix combines it with the original image through another random convex combination to obtain the final image $\\boldsymbol{x}_{augmix}$ , i.e., $\\boldsymbol{x}_{augmix} = m\\boldsymbol{x} + (1 - m)\\boldsymbol{x}_{aug}$ , where $m$ is sampled from a $\\operatorname{Beta}(\\alpha, \\alpha)$ distribution. AugMix adopts a fixed random augmentation policy, i.e., Dirichlet and Beta distributions, during the entire training process.", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24523", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9bc3e58034b6a6c6710135d88b192c538cf7e7fdb13fec050b39bee734bfd5b4.jpg", + "image_caption": [ + "(a) Continual Learning Process" + ], + "image_footnote": [], + "bbox": [ + 89, + 93, + 455, + 151 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b8b5da81454f5d2fcae324074954ddf84fe5a702baf00360719987d5635bd247.jpg", + "image_caption": [ + "(b) AugMix" + ], + "image_footnote": [], + "bbox": [ + 86, + 171, + 459, + 297 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/18f1487a47f0f0d0d0d9c1ba79eaa3d74579d1f72e97b6f124ba36706d274f47.jpg", + "image_caption": [ + "(c) MetaMix (Ours)", + "Figure 2. Comparisons between AugMix [22] and our proposed MetaMix. (a) Tasks sequentially arrive during CL. (b) At each CL step, AugMix adopts a fixed and random augmentation policy during the entire learning process, thus lacking adaptation to the non-stationary data in CL and may not be the optimal strategy for each task. (c) MetaMix automatically adapts to the non-stationary data distribution during CL by meta-learning the mixing parameters of augmentation. At time $t$ , we sample a batch of data $(\\pmb{x}_r, y_r)$ from the memory buffer and concatenate them with the current received data $(\\pmb{x}_t, y_t)$ as $(\\pmb{x}_b, y_b)$ . We apply MetaMix to augment each data example. Specifically, we feed the context information $I_t$ into the LSTM cell with parameters $\\phi_t$ , and obtain the mixing weights $w_t$ and $m_t$ . The CL model $f_{\\theta_t}$ calculates the losses for the augmented data. We adopt bi-level optimization to improve model generalization on unseen corruptions by splitting the augmentation operations into a pseudo-seen set $S_t$ and pseudo-unseen set $\\mathcal{U}_t$ . The loss $\\mathcal{L}_{\\phi_t}$ is calculated on data augmented with $\\mathcal{U}_t$ to update $\\phi_t$ , and loss $\\mathcal{L}_{\\theta_t}$ is calculated on data augmented with $S_t$ to update $\\theta_t$ . We repeat this process for all tasks in the sequence." + ], + "image_footnote": [], + "bbox": [ + 86, + 315, + 460, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The augmentation procedure of AugMix is illustrated in Figure 2 (b). The optimization objective of AugMix is shown in the following:", + "bbox": [ + 76, + 747, + 470, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {*} = \\arg \\min _ {\\boldsymbol {\\theta}} \\underset {o p \\in \\mathcal {O}} {\\mathbb {E}} \\mathcal {L} (\\boldsymbol {x}, y, \\boldsymbol {\\theta}) + \\lambda J S (\\boldsymbol {x}, \\boldsymbol {x} _ {a u g m i x 1}, \\boldsymbol {x} _ {a u g m i x 2}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 804, + 472, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ is the original data and $y$ is the data label; $\\mathcal{L}(\\pmb{x},\\pmb{y},\\pmb{\\theta})$ is the loss function on the labeled data (cross-entropy for classification); $\\pmb{x}_{augmix1}$ and $\\pmb{x}_{augmix2}$ are two augmenta", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions of original data $\\mathbf{x}$ by AugMix; $JS$ denotes the Jensen Shannon divergence for regularizing the output consistency between the original and augmented data. We will define it in the following section.", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. MetaMix Approach", + "text_level": 1, + "bbox": [ + 500, + 164, + 687, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Direct extension of AugMix [22] for CL is to augment data examples sampled from the memory buffer, and the received new data at each time point for memory-based method across the sequentially arrived tasks. In the following, we present our methods for memory-based CL methods, but our methods can be easily extended to other families of CL methods as well, also demonstrated in Section 4. However, since the memory buffer is usually small for CL models, using a fixed random augmentation strategy (like AugMix) may make the CL model gradually memorize the training data corruption operations, but fail to generalize to unseen corruptions during testing. Moreover, with the non-stationary nature of the data, the optimal augmentation strategy for image mixing is data-dependent, and the mixing weights should be specifically optimized for each task. However, since we cannot repeatedly tune the mixing weights for each task during CL learning process, an automatic self-adaptive process is necessary to achieve good performance. We thus introduce the details of our self-adaptive data augmentation approach, namely MetaMix, for improving robustness in CL. Figure 2 (c) illustrates the training process of MetaMix at time $t$ . Specifically, we adopt meta-learning to learn a set of data-dependent parameters $w_{t}$ and $m_{t}$ , which can dynamically control the mixing weights of augmentation operations and automatically adapt to the non-stationary data distribution.", + "bbox": [ + 496, + 188, + 893, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, to memorize previous knowledge, we allocate a small memory buffer $\\mathcal{M}$ to store data from previous tasks by reservoir sampling, similar to [35]. The data in the memory buffer will be replayed later with a new task together. Suppose we have $M$ augmentation operations $\\mathcal{O} = \\{op_1, op_2, \\dots, op_M\\}$ during training, CL model parameters at time $t$ are $\\theta_t$ , and the function represented by the network with $\\theta_t$ is $f_{\\theta_t}(x)$ .", + "bbox": [ + 496, + 582, + 893, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To capture the sequential dependencies among previous tasks, we design a self-adaptive LSTM mixing strategy. The LSTM inputs the current context information and outputs the adaptive mixing parameters. We use an LSTM (dynamic) model to dynamically change the mixing weights instead of an MLP (static) model since the tasks in CL sequentially arrive. The LSTM can encode the information of previous tasks into an additional hidden state while the MLP can only capture the data information of the current task. It does not capture the information of previous tasks without the hidden state. Formally, the LSTM adaptive process is defined as:", + "bbox": [ + 496, + 704, + 893, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {o} _ {t + 1}, \\boldsymbol {h} _ {t + 1}, \\boldsymbol {g} _ {t + 1} = \\operatorname {L S T M} _ {\\phi_ {t}} \\left(\\boldsymbol {I} _ {t}, \\boldsymbol {h} _ {t}, \\boldsymbol {g} _ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 885, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24524", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_t$ is the LSTM MetaMixer parameters at time $t$ , $h_t$ is the hidden state of LSTM at time $t$ , $g_t$ is the cell state of LSTM at time $t$ , $I_t$ is the context information encoding as input to LSTM at time $t$ ; $o_{t+1}$ is the output of LSTM at time $t+1$ for the mixing parameters, i.e., $o_{t+1} = \\{w_t', m_t'\\}$ , where $m_t'$ is the mixing weight for the augmented data and original raw data and $w_t'$ are the mixing weights for different chains in the augmentation. We apply softmax function to $w_t'$ and sigmoid function to $m_t'$ to ensure they are within [0,1] interval, i.e.,", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {t} = \\operatorname {S i g m o i d} \\left(m _ {t} ^ {\\prime}\\right), \\boldsymbol {w} _ {t} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {w} _ {t} ^ {\\prime}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 250, + 470, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The context $I_{t}$ contains essential information to characterize data in current and previous tasks. It is shown as:", + "bbox": [ + 76, + 271, + 472, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {t} = \\left\\{\\nabla \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {y} _ {r}\\right) \\cdot \\nabla \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {t}\\right), \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {y} _ {r}\\right), \\boldsymbol {e} _ {r}, \\boldsymbol {e} _ {t} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 323, + 468, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(\\pmb{x}_r, y_r)$ is the labeled mini-batch data sampled from the memory buffer and $(\\pmb{x}_t, y_t)$ is the labeled data received at time $t$ . We adopt this information for several reasons: (1) $\\nabla \\mathcal{L}_{\\pmb{\\theta}_t}(\\pmb{x}_r, y_r) \\cdot \\nabla \\mathcal{L}_{\\pmb{\\theta}_t}(\\pmb{x}_t, y_t)$ is the gradient dot product between memory data gradient and current mini-batch data gradient. It determines the degree of interference and transfer between memory data and current task mini-batch data. It is commonly used in existing CL as context information [18, 35]; (2) we also input the memory mini-batch data loss, as it indicates the degree of forgetting of previous tasks; (3) $\\pmb{e}_r$ and $\\pmb{e}_t$ are data encoding for mini-batch data, $\\pmb{x}_r$ , sampled from memory buffer and currently received new mini-batch data, $\\pmb{x}_t$ , respectively. That is, $\\pmb{e}_r$ and $\\pmb{e}_t$ are the last layer features outputted by ResNet-18. This determines the data-dependent features for calculating the data-dependent mixing weights.", + "bbox": [ + 75, + 354, + 472, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since corruption operations on testing data are unseen during training, we cannot directly optimize the augmentation strategy for these corruptions. To make the CL model generalize better to unseen corruptions, at each training step $t$ , we randomly split the augmentation operations $\\mathcal{O}$ into non-overlapping pseudo-seen $S_{t}$ and pseudo-unseen $\\mathcal{U}_t$ operations, i.e., $S_{t} \\cup \\mathcal{U}_{t} = \\mathcal{O}$ and $S_{t} \\cap \\mathcal{U}_{t} = \\emptyset$ . We use the pseudo operation sets to simulate the scenario in which training and testing corruptions have no overlap. We formulate MetaMix as a bi-level optimization problem to make the CL model learn to augment and mix. This ensures that the update of model parameters $\\theta$ on the pseudo-seen set can improve the optimization of MetaMixer parameters $\\phi$ on the pseudo-unseen set, which could, in turn, be beneficial to improve the CL model generalization. We formulate the MetaMix as the following bi-level optimization:", + "bbox": [ + 75, + 597, + 472, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\phi_ {*} = \\arg \\min _ {\\phi} \\mathbb {E} _ {o p \\in \\mathcal {U} _ {t}} \\mathcal {L} \\left(\\boldsymbol {x} _ {b}, y _ {b}, \\boldsymbol {\\theta} _ {*}, \\phi\\right) + \\lambda J S \\left(\\boldsymbol {x} _ {b}, \\widehat {\\boldsymbol {x}} _ {b 1}, \\widehat {\\boldsymbol {x}} _ {b 2}\\right), \\tag {6} \\\\ \\boldsymbol{\\theta}_{*} = \\arg \\min_{\\boldsymbol{\\theta}}\\underset {op\\in \\mathcal{S}_{t}}{\\mathbb{E}}\\mathcal{L}(\\boldsymbol{x}_{b},y_{b},\\boldsymbol {\\theta},\\boldsymbol {\\phi}) + \\lambda J S(\\boldsymbol{x}_{b},\\boldsymbol{x}^{\\prime}_{b1},\\boldsymbol{x}^{\\prime}_{b2}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 844, + 468, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we concatenate the mini-batch data $(\\pmb{x}_r, y_r)$ from the memory buffer and the current received mini-batch data $(\\pmb{x}_t, y_t)$ together as $(\\pmb{x}_b, y_b)$ ; $\\widehat{\\pmb{x}}_{b1}$ and $\\widehat{\\pmb{x}}_{b2}$ are the two minibatch data augmented by applying the pseudo-unseen augmentation operations $\\mathcal{U}_t$ on $(\\pmb{x}_b, y_b)$ ; $\\pmb{x}_{b1}'$ and $\\pmb{x}_{b2}'$ are the two mini-batch data augmented by applying the pseudo-seen augmentation operations $\\mathcal{S}_t$ on $(\\pmb{x}_b, y_b)$ . $\\lambda$ is the regularization constant for the consistency loss. The lower-level optimization is to optimize the performance on the pseudoseen augmentation operations $\\mathcal{S}_t$ at the training step $t$ ; and the upper-level optimization is to optimize the generalization on the pseudo-unseen augmentation operations $\\mathcal{U}_t$ . This bi-level optimization for learning to augment and mix can be efficiently solved by first-order gradient descent method, similar to the first-order MAML [15].", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The advantage of MetaMix (Eq. (6)) over AugMix (Eq. (2)) are two folds. First, MetaMix improves the generalization to unseen corruptions since we split the training data augmentation operations into pseudo-seen and unseen operations. This is similar to the train-validation data split in standard meta-learning, which improves the generalization compared to the training without data split [39]; It could be understood as introducing an implicit regularization so that it can benefit from low sample complexity [39]. Second, MetaMix can automatically adapt to the non-stationary data distribution in CL, while AugMix can only work well in the stationary data distribution in a single task.", + "bbox": [ + 496, + 323, + 893, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consistency Regularization To ensure the network output consistency between the original raw image data and the augmented data, we use a similar consistency loss [22] to regularize the network output to ensure smoother neural network responses. We hope the model to have similar responses to $\\pmb{x}_b, \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2}$ . We minimize the Jensen-Shannon divergence among the posterior distributions of the original sample $\\pmb{x}_b$ and its augmented variants $\\widehat{\\pmb{x}}_{b1}$ and $\\widehat{\\pmb{x}}_{b2}$ . The consistency loss is defined as below:", + "bbox": [ + 496, + 511, + 893, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {m e a n} = \\left(p _ {\\boldsymbol {x} _ {b}} + p _ {\\hat {\\boldsymbol {x}} _ {b 1}} + p _ {\\hat {\\boldsymbol {x}} _ {b 2}}\\right) / 3 (7) \\\\ J S \\left(\\boldsymbol {x} _ {b}, \\widehat {\\boldsymbol {x}} _ {b 1}, \\widehat {\\boldsymbol {x}} _ {b 2}\\right) = \\left(\\mathbb {K L} \\left(p _ {\\boldsymbol {x} _ {b}} \\mid p _ {\\text {m e a n}}\\right) + \\mathbb {K L} \\left(p _ {\\widehat {\\boldsymbol {x}} _ {b 1}} \\mid p _ {\\text {m e a n}}\\right) \\right. \\\\ + \\mathbb {K L} \\left(p _ {\\widehat {x} _ {b 2}} \\mid p _ {\\text {m e a n}}\\right)) / 3 (8) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 493, + 672, + 890, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $\\mathbb{KL}$ denotes the KL divergence between two distributions. $p_{\\boldsymbol{x}_b} = f_{\\boldsymbol{\\theta}_t}(\\boldsymbol{x}_b)$ is the network output probabilities of each class for original raw data $\\boldsymbol{x}_b$ . Similarly, we can define $p_{\\widehat{\\boldsymbol{x}}_{b1}} = f_{\\boldsymbol{\\theta}_t}(\\widehat{\\boldsymbol{x}}_{b1})$ and $p_{\\widehat{\\boldsymbol{x}}_{b2}} = f_{\\boldsymbol{\\theta}_t}(\\widehat{\\boldsymbol{x}}_{b2})$ . We summarize our proposed algorithm in Algorithm 3. Line 5-8 takes the context information as input and computes desired augmentation mixing parameters. Line 9-11 augments the original data with the pseudo-seen augmentation operations $\\mathcal{S}_t$ and pseudo-unseen operations $\\mathcal{U}_t$ by Algorithm 1. Line 12 solves the bi-level optimization problem (Eq. (6)) by Algorithm 2.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24525", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 DataMix" + ], + "code_body": "1: Function DataMix $(\\pmb{x}_b, \\pmb{w}_t, m_t, \\mathcal{V})$ \n2: MetaMix augmented images initialized as $\\pmb{x}_b' = 0$ \n3: $\\pmb{x}_{orig} = \\pmb{x}_b$ \n4: for $i = 1$ to $A$ do { A is the number of augmentation chains} \n5: $\\pmb{x}_b = \\pmb{x}_{orig}$ \n6: for $l = 1$ to $d$ do { $d$ is the augmentation depth} \n7: randomly sample operation $op \\in \\mathcal{V}$ ( $\\mathcal{V}$ is the collection of augmentation operations) \n8: $\\pmb{x}_b = op(\\pmb{x}_b)$ \n9: end for \n10: $\\pmb{x}_b' + = \\pmb{w}_t^i \\cdot \\pmb{x}_b$ \n11: end for \n12: $\\pmb{x}_b' = m_t \\pmb{x}_{orig} + (1 - m_t) \\pmb{x}_b'$ ; \n13: similar to the above procedure, we can get two different minibatch $\\pmb{x}_{b1}', \\pmb{x}_{b2}'$ \n14: return $\\pmb{x}_{b1}', \\pmb{x}_{b2}'$ \n15: EndFunction", + "bbox": [ + 83, + 109, + 472, + 345 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Bi-level Solver" + ], + "code_body": "1: Function Bi-level Solver $(\\pmb{\\theta}_t, \\pmb{\\phi}_t, \\pmb{x}_b, y_b, \\pmb{x}_{b1}', \\pmb{x}_{b2}', \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2})$ \n2: optimize the lower-level optimization problem in Eq. (6) by $\\pmb{\\theta}_{t+1} = \\pmb{\\theta}_t - \\gamma \\nabla_\\theta \\mathbb{E}_{op \\in S_t}[\\mathcal{L}(\\pmb{x}_b, y_b, \\pmb{\\theta}_t, \\phi) + \\lambda JS(\\pmb{x}_b, \\pmb{x}_{b1}', \\pmb{x}_{b2}')]$ \n3: calculate the upper-level optimization loss function in Eq. (6) by $\\mathcal{E}(\\pmb{\\theta}_{t+1}, \\phi) = \\mathbb{E}_{op \\in U_t}[\\mathcal{L}(\\pmb{x}_b, y_b, \\pmb{\\theta}_{t+1}, \\phi) + \\lambda JS(\\pmb{x}_b, \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2})]$ \n4: $\\phi_1' = \\phi_t$ \n5: for $s = 1$ to $J$ do \n6: $\\phi_{s+1}' = \\phi_s' - \\beta \\nabla_{\\phi'} \\mathcal{E}(\\pmb{\\theta}_{t+1}, \\phi_s')$ \n7: end for \n8: $\\phi_{t+1} = \\phi_{J+1}'$ \n9: return $\\pmb{\\theta}_{t+1}, \\phi_{t+1}$ \n10: EndFunction", + "bbox": [ + 81, + 383, + 472, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 609, + 200, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the model performance for both task- and class-continual learning. The CL models are evaluated on the proposed benchmark for each dataset. We then describe experiment setup, results, and ablation study.", + "bbox": [ + 75, + 635, + 468, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 709, + 253, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Below, we first construct the corruption CL benchmark and then describe the evaluation metrics, baselines, and implementation details.", + "bbox": [ + 75, + 732, + 470, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Corruption CL benchmark. Existing CL benchmarks, such as Split-CIFAR10, Split-CIFAR100, and Split-miniImageNet, are commonly used in literature. However, each task's training and testing data follow the same distribution for these datasets. Thus, these benchmarks are not suitable for evaluating the corruption robustness of CL methods. To properly evaluate the corruption robustness, we follow the protocols [21] and create several new CL datasets.", + "bbox": [ + 75, + 779, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 3 MetaMix" + ], + "code_body": "1: REQUIRE: Augmentation Operations $\\mathcal{O}$ $=$ {sterize, rotate, ..., solarize}; augmentation width $A$ ; augmentation depth $d$ ; the number of CL tasks N; the number of training iterations for each task $\\mathcal{T}_k$ is $N_{k}$ ; memory buffer $\\mathcal{M}$ ; CL model parameters $\\theta$ ; LSTM MetaMix with parameters $\\phi$ ; MetaMixer LSTM update steps $J$ ; CL model parameter learning rate $\\gamma$ ; regularization weight $\\lambda$ ; original image $\\pmb{x}_{orig}$ ; initial hidden state and cell state $\\pmb{h}_t,\\pmb{g}_t$ are initialized as random noise $\\pmb{h}_1,\\pmb{g}_1\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})$ \n2: for $k = 1$ to $N$ do \n3: for $t = 1$ to $N_{k}$ do \n4: randomly sample mini-batch data $(x_{r},y_{r})$ from memory buffer $\\mathcal{M}$ and concatenate it with the current received mini-batch data $(x_{t},y_{t})$ to obtain $(x_{b},y_{b})$ \n5: compute context vector $\\pmb{I}_t = \\{\\nabla \\mathcal{L}_{\\theta_t}(\\pmb{x}_r,\\pmb{y}_r)$ $\\nabla \\mathcal{L}_{\\theta_t}(\\pmb{x}_t,\\pmb{y}_t),\\mathcal{L}_{\\theta_t}(\\pmb{x}_r,\\pmb{y}_r),\\pmb{e}_r,\\pmb{e}_t\\}$ \n6: $o_{t + 1},h_{t + 1},g_{t + 1} = \\mathrm{LSTM}_{\\phi_t}(I_t,h_t,g_t)$ \n7: generate mixing parameters: $\\pmb{w}_t^{\\prime},m_t^{\\prime} = o_{t + 1}$ \n8: $m_t = \\mathrm{Sigmoid}(m_t^{\\prime}),w_t = \\mathrm{Softmax}(w_t^{\\prime})$ \n9: randomly split augmentation operations $\\mathcal{O}$ into nonoverlapping pseudo-seen $S_{t}$ and unseen $\\mathcal{U}_t$ \n10: $\\pmb{x}_{b1}^{\\prime},\\pmb{x}_{b2}^{\\prime} = \\mathrm{DataMix}(\\pmb{x}_b,\\pmb{w}_t,m_t,\\mathcal{S}_t)$ (Algorithm 1) \n11: $\\widehat{\\pmb{x}}_{b1},\\widehat{\\pmb{x}}_{b2} = \\mathrm{DataMix}(\\pmb{x}_b,\\pmb{w}_t,m_t,\\mathcal{U}_t)$ (Algorithm 1) \n12: $\\theta_{t + 1},\\phi_{t + 1} = \\mathrm{Bi - level~Solver~}(\\theta_t,\\phi_t,\\pmb{x}_b,\\pmb{y}_b,\\pmb{x}_{b1}'',\\pmb{x}_{b2}'',\\widehat{\\pmb{x}}_{b1},\\widehat{\\pmb{x}}_{b2})$ (Algorithm 2) \n13: update memory buffer $\\mathcal{M}$ through reservoir sampling [35] to determine whether to store $(x_{t},y_{t})$ \n14: end for \n15: end for", + "bbox": [ + 506, + 108, + 893, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply 15 common corruption operations to the testing data of the three CL benchmarks, such as snow, frost, fog, etc. We provide a brief description for each corruption operation in Table 5 in Appendix. We perform each corruption at five different severity levels. Note that these corruptions only appear on the testing dataset of each continually learned task and are not seen in the training data of each task. This is to ensure that we can measure the robustness to unseen corruptions during testing for each compared method. To our best knowledge, we build the first CL corruption benchmark for measuring the generalization of CL models to unseen corruptions during testing. We name these new datasets Split-CIFAR10-C, Split-CIFAR100-C, and Split-miniImageNet-C. Specifically, for Split-CIFAR-10-C, we split the CIFAR-10 dataset [28] into 5 disjoint tasks, where each task has 2 classes; for Split-CIFAR-100-C, we split CIFAR-100 [28] consisting of 100 classes into 10 disjoint tasks, where each task has 10 classes; for Split-miniImageNet-C, we split miniImageNet [44] that consists of 100 classes, into 10 disjoint tasks, where each task has 10 classes.", + "bbox": [ + 496, + 547, + 893, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the performance of different methods with average accuracy and backward transfer at the end of CL training to measure the final performance", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24526", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the extent of forgetting for different methods. For each task, the accuracy for a specific corruption $c$ is the average accuracy over the five severity level. Suppose for task $k$ , $ACC_{c,s}^{k}$ measures the robust accuracy after applying the corruption operation $c$ with severity level $s$ :", + "bbox": [ + 76, + 90, + 470, + 167 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA C C _ {c, s} ^ {k} = \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {k}, y ^ {k}, \\mathcal {T} _ {k}\\right) \\sim D _ {t e} ^ {t e}} \\mathbb {I} \\left(f \\left(c _ {s} \\left(\\boldsymbol {x} ^ {k}\\right)\\right) = y ^ {k}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 175, + 431, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $c_{s}(\\pmb{x}^{k})$ is the corruption operation $c$ with severity level $s$ applied on $\\pmb{x}^{k}$ . The accuracy for corruption $c$ denoted as $ACC_{c}^{k}$ is the average robust accuracy across all the five severity levels. The overall accuracy for task $k$ , $ACC^{k}$ , is the average accuracy across all corruption types, i.e.,", + "bbox": [ + 76, + 205, + 468, + 282 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA C C _ {c} ^ {k} = \\frac {1}{5} \\sum_ {s = 1} ^ {s = 5} A C C _ {c, s} ^ {k}, \\text {a n d} A C C ^ {k} = \\frac {1}{| \\mathcal {C} |} \\sum_ {c \\in \\mathcal {C}} A C C _ {c} ^ {k}.\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 291, + 462, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To measure catastrophic forgetting, we also evaluate the backward transfer (BWT), which measures the extent of forgetting on previous tasks after learning new ones. $BWT < 0$ reveals the occurrence of catastrophic forgetting on previous tasks, and $BWT > 0$ indicates that learning new tasks is helpful for previous tasks. Formally, the backward transfer (BWT) is defined as: $BWT = \\frac{1}{N - 1}\\sum_{k = 1}^{k = N - 1}(R_{N,k} - R_{k,k})$ ; where $R_{N,k}$ is the testing accuracy on task $k$ after learning on task $N$ and $R_{N,k}$ is defined the same as the above mentioned $ACC^k$ . For BWT, the higher, the better.", + "bbox": [ + 76, + 340, + 468, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Continual Learning Setting. We apply our proposed MetaMix to task-continual learning (Task-CL) and class-continual learning (Class-CL). The former provides task identities for the CL learner to select the relevant classifier for each example during testing, whereas the latter does not.", + "bbox": [ + 76, + 492, + 470, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline. Our method can be seamlessly integrated with exiting CL methods. For illustration, we integrate the proposed methods with SOTA memory-based methods: $DER++$ [5] and CLS-ER [2]. To evaluate the effectiveness of the proposed MetaMix, we compare it with several SOTA data augmentation approaches originally designed for traditional supervised learning: Adversarial Training (AT) [33], RandAugment (RA) [11], Maxup [16], DeepAugment [20], Augmix [22]. We adapt these augmentation methods for CL problems by applying them on the new task data and the memory buffer data. We provide more detailed descriptions of baselines in Appendix. Due to space limitations, we put the experiments that integrate our proposed method with CLS-ER in Appendix.", + "bbox": [ + 75, + 568, + 470, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. We use ResNet18 [19] as classifier for all datasets. All the other hyperparameter settings follow from [5]. The types of corruption during testing are not seen during training for all the compared augmentation methods to evaluate the robustness against unseen corruptions. Following [22], the augmentation operations performed during training are autocontrast, equalize, posterize, rotate, solarize, shear-x, shear-y, translate-x, translate-y.", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All the above-compared data augmentation methods apply these augmentation operations to CL baseline method to improve robustness. The memory buffer has capacity of 500 data points by default. We average the result for 5 runs for each experiment. We provide both mean and standard deviation across different runs. Due to space limitations, we provide standard deviation results in Appendix. We provide more implementation details in Appendix.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LSTM Architecture. It has one recurrent layer with 5 hidden units. A linear layer (fully connected layer) is appended next to the LSTM output to generate the mixing parameters $\\boldsymbol{w}_t$ and $m_t$ . The LSTM only has about 20K parameters and is negligible compared to ResNet18, which has more than 11.22 million parameters.", + "bbox": [ + 496, + 218, + 890, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Robustness of Task-CL", + "text_level": 1, + "bbox": [ + 498, + 335, + 714, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task-Continual Learning In this section, we evaluate the corruption robustness under task-continual learning, where the task identifiers are provided to the CL learner during testing. We present the results that integrate baselines and the proposed method with DER++ [5]. Due to space limitations, we put the experiments that integrate baselines and MetaMix with CLS-ER [2] in Appendix.", + "bbox": [ + 496, + 366, + 893, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We show the results across different corruptions and compared methods in Tables 1-2; where $Avg$ is the average robust accuracy across all the corruptions. Due to the space limitations, we provide the standard deviation across different runs in Table 8 and 9 in Appendix. We also put results on split-CIFAR10-C in Table 6 and Table 7 in Appendix. We can observe that without data augmentation, the performance of $\\mathrm{DER}++$ is close to random guessing, indicating that the CL model does not generalize to common corruptions. The proposed self-adaptive MetaMix brings significant improvement toward corruption robustness. It outperforms the average corruption robustness accuracy of the other models by a large margin of $2.4\\%$ , $2.6\\%$ on Split-CIFAR100-C, and Split-miniImageNet-C, respectively. The improvement shows the effectiveness of proposed MetaMix, which can automatically adapt to the non-stationary data distribution. The compared methods only use fixed static/random augmentation methods which lack adaptation to the non-stationary data in CL, and thus do not perform well. Random augmentation achieves second-best results. Deep-augment would distort the image such that they are very different from original images. Adversarial training and Maxup do not help much corruption robustness among the compared methods. Adversarial training can generate similar images compared to original images. Thus, there is a large gap between training and testing corrupted data for those methods adapted for CL. Maxup aims for improving clean data performance, not designed for robust accuracy.", + "bbox": [ + 496, + 478, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24527", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f0bb9b21acdf49e9d9d018c950a7543eaaf15d11dd9caa879cc59636713dcb2.jpg", + "table_caption": [ + "Table 1. Robust accuracy of Task-CL on Split-CIFAR100-C with DER++" + ], + "table_footnote": [], + "table_body": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++10.4610.6910.2610.7610.5110.7110.6610.4710.5610.5110.7910.5410.7510.5510.7210.60
AT10.3610.4510.2510.2310.2410.2710.2510.129.889.9810.169.8310.1910.310.3110.19
RA26.7727.1525.4627.8423.4624.0525.9923.2626.3430.4527.9528.825.5429.1229.0926.75
DeepAugment11.0011.1610.8611.3511.1511.4511.3310.7610.4610.5911.1010.5711.2011.3611.2511.04
Maxup11.111.1711.1311.0911.0611.1311.110.8810.7610.8110.8610.3711.0611.1311.1210.98
Augmix59.5862.6859.9267.9558.8865.9367.0363.3963.9363.2468.6263.2765.8266.0664.0764.02
Ours61.5365.1461.1870.3561.2868.0168.7666.0266.4966.1471.4966.1668.2468.9166.7566.43
", + "bbox": [ + 107, + 107, + 859, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d3ea6fc99a9946344001d94e4d408641fcc408a084e968d7ccf94ca75c0b0926.jpg", + "table_caption": [ + "Table 2. Robust accuracy of Task-CL on Split-MiniImageNet-C with DER++" + ], + "table_footnote": [], + "table_body": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++10.2710.110.079.789.739.679.729.619.719.859.6910.159.559.459.579.79
AT9.079.249.279.659.629.69.489.179.109.019.128.389.549.569.359.28
RA17.2318.2216.719.8116.618.3219.0718.9417.1418.2617.2614.5621.1420.8423.018.47
DeepAugment10.5410.5310.6210.6810.7010.6310.7410.5610.3210.3210.529.9210.5510.4110.4910.50
Maxup10.410.910.6810.659.7510.2110.259.6210.1210.2110.3410.910.210.1610.0710.30
Augmix43.4548.9142.2053.7849.6859.8851.8353.2254.0255.6359.5943.4661.3551.3462.7552.74
Ours48.853.3343.8159.753.6462.7856.9556.1957.3156.0862.0741.7163.1749.7864.5555.32
", + "bbox": [ + 107, + 232, + 857, + 339 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8193a5865e8b6ae064f1e154e54344c2474dddcfe2427dbfecdd8ee2a38a7637.jpg", + "table_caption": [ + "Table 3. Robust accuracy of Class-CL on Split-CIFAR100-C with DER++" + ], + "table_footnote": [], + "table_body": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++0.750.750.880.750.860.770.680.880.960.920.750.960.780.780.790.82
AT0.990.971.040.910.950.970.920.831.050.940.921.00.910.950.930.95
RA6.226.35.676.254.424.545.595.256.247.726.426.935.166.817.116.04
DeepAugment1.291.291.251.261.261.231.241.261.241.241.141.261.221.281.331.25
Maxup1.171.151.161.151.081.141.121.081.081.091.141.061.131.151.111.12
Augmix19.5721.2120.3526.1819.0825.1725.4622.0922.4322.8226.4422.5423.7224.6423.0122.98
Ours22.3224.4423.1529.1922.1328.1028.3525.4725.226.0829.5625.7827.0128.1625.8226.05
", + "bbox": [ + 107, + 354, + 857, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Robustness of Class-CL", + "text_level": 1, + "bbox": [ + 76, + 481, + 295, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class-Continual Learning. In this section, we evaluate the corruption robustness under the more challenging class-continual learning setting, where the task identifiers are unavailable to the CL learner during testing.", + "bbox": [ + 75, + 505, + 470, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. We present the results across different corruptions and compared methods on Split-CIFAR100-C in Table 3, we present the results on Split-miniImageNet-C in Table 10 in Appendix. Similar to the Task-CL, MetaMix significantly improves over the baseline augmentation methods in Class-CL setting due to its self-adaptivity property. It outperforms the average corruption robustness accuracy of the other models by a large margin of $3.1\\%$ , $2.2\\%$ on Split-CIFAR100-C, and Split-miniImageNet-C, respectively. This is because our proposed method can automatically adapt to the non-stationary distributions. Most compared methods do not perform well on this challenging setup.", + "bbox": [ + 75, + 566, + 470, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to space limitations, we put BWT results in Table 15 in Appendix. In the corruption-robustness scenario, BWT is no longer a meaningful metric with such extreme accuracy differences, as the significantly lower accuracy of comparison methods results in much less space for further performance variations during backward transfer.", + "bbox": [ + 75, + 747, + 470, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study and Hyperparameter Analysis", + "text_level": 1, + "bbox": [ + 76, + 847, + 467, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the limited space, we provide (1) ablation studies; (2) hyperparameter analyses, including $\\lambda$ , $J$ , $\\beta$ , etc.; (3)", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of using LSTM vs MLP and the benefit of using additional hidden state information from previous tasks; (4) computation cost in Appendix.", + "bbox": [ + 498, + 481, + 893, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Memory Size. We evaluate the effect of memory size with 500 and 3000, respectively. The memory size of 500 is the default setting in the above tables. We provide experiment results on Split-CIFAR100-C, Split-miniImageNet-C with memory size 3000 on task-CL and class-CL respectively in Table 11, 12, 13, 14 in Appendix. In these cases, our method significantly outperforms baselines.", + "bbox": [ + 496, + 532, + 893, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 667, + 617, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper tackles a more challenging problem of corruption robustness in CL with non-stationary data distribution, where the testing data distribution may significantly differ from training data distribution with various forms of common corruptions. We propose a meta-learning framework, MetaMix, for self-adaptive data augmentation specialized for CL. Our proposed MetaMix can substantially improve the model's robustness. Comprehensive experiments on both task- and class-CL settings demonstrate the effectiveness of the proposed method.", + "bbox": [ + 496, + 699, + 893, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement We thank all the anonymous reviewers for their insightful and thoughtful comments. This research was supported in part by grant NSF III-1910492.", + "bbox": [ + 498, + 854, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24528", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. Advances in Neural Information Processing Systems 32, pages 11849-11860, 2019. 1, 2", + "[2] Elahe Arani, Fahad Sarfraz, and Bahram Zonooz. Learning fast, learning slow: A general continual learning method based on complementary learning system. In International Conference on Learning Representations, 2022. 7, 12, 17", + "[3] Jihwan Bang, Hyunseo Koh, Seulki Park, Hwanjun Song, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on a contaminated data stream with blurry task boundaries. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9265-9274, 2022. 3", + "[4] E. Belouadah and A. Popescu. II2m: Class incremental learning with dual memory. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 583-592, 2019. 2", + "[5] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. 34th Conference on Neural Information Processing Systems, 2020. 1, 7, 12", + "[6] Dan A. Calian, Florian Stimberg, Olivia Wiles, Sylveste-Alvise Rebuffi, Andras Gyorgy, Timothy Mann, and Sven Gowal. Defending against image corruptions through adversarial augmentations. https://arxiv.org/abs/2104.01086, 2021.3", + "[7] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio Calmon, and Taesup Moon. {CPR}: Classifier-projection regularization for continual learning. In International Conference on Learning Representations, 2021. 2", + "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. Proceedings of the International Conference on Learning Representations, 2019. 2", + "[9] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. https://arxiv.org/abs/1902.10486, 2019.", + "[10] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation policies from data. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[11] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. Advances in Neural Information Processing Systems, 2020. 1, 3, 7, 12", + "[12] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. https://arxiv.org/abs/1708.04552, 2017.3", + "[13] Sayna Ebrahimi, Mohamed Elhoseiny, Trevor Darrell, and Marcus Rohrbach. Uncertainty-guided continual learning with bayesian neural networks. Proceedings of the International Conference on Learning Representations, 2020. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. https://arxiv.org/abs/1701.08734, 2017.2", + "[15] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. International Conference on Machine Learning, 2017. 5", + "[16] Chengyue Gong, Tongzheng Ren, Mao Ye, and Qiang Liu. Maxup: Lightweight adversarial training with data augmentation improves neural network training. Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 7, 12", + "[17] Yanan Gu, Xu Yang, Kun Wei, and Cheng Deng. Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[18] Gunshi Gupta, Karmesh Yadav, and Liam Paull. La-maml: Look-ahead meta learning for continual learning. In Advances in Neural Information Processing Systems, 2020. 5", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2016. 7", + "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. International Conference on Computer Vision, 2021. 3, 7, 12", + "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 1, 6", + "[22] Dan Hendrycks, Norman Mu, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple method to improve robustness and uncertainty under data shift. 2020. 2, 3, 4, 5, 7, 12", + "[23] Christian Henning, Maria Cervera, Francesco D'Angelo, Johannes Von Oswald, Regina Traber, Benjamin Ehret, Seijin Kobayashi, Benjamin F Grewe, and Joao Sacramento. Posterior meta-replay for continual learning. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2", + "[24] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Intertask association in continual learning. In European Conference on Computer Vision, 2022. 2", + "[25] Chris Dongjoo Kim, Jinseo Jeong, Sangwoo Moon, and Gunhee Kim. Continual learning on noisy data streams via self-purified replay. International Conference on Computer Vision, 2021. 3", + "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24529", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "networks. Proceedings of the national academy of sciences, 2017. 2", + "[27] Yajing Kong, Liu Liu, Zhen Wang, and Dacheng Tao. Balancing stability and plasticity through advanced null space in continual learning. In European Conference on Computer Vision, 2022. 2", + "[28] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6", + "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Conference on Empirical Methods in Natural Language Processing, 2021. 2", + "[30] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. TRGP: Trust region gradient projection for continual learning. In International Conference on Learning Representations, 2022. 1, 2", + "[31] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in Neural Information Processing Systems, 2017. 1, 2", + "[32] Amrutha Machireddy, Ranganath Krishnan, Nilesh Ahuja, and Omesh Tickoo. Continual active adaptation to evolving distributional shifts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3443-3449, 2022. 3", + "[33] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. Proceedings of the International Conference on Learning Representations, 2019. 7, 12", + "[34] Cuong V. Nguyen, Yingzhen Li, Thang D. Bui, and Richard E. Turner. Variational continual learning. Proceedings of the International Conference on Learning Representations, 2018. 2", + "[35] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. International Conference on Learning Representations, 2019. 2, 4, 5, 6", + "[36] Tim G. J. Rudner, Freddie Bickford Smith, Qixuan Feng, Yee Whye Teh, and Yarin Gal. Continual learning via sequential function-space variational inference. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 18871-18887, 17-23 Jul 2022. 1", + "[37] Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. https://arxiv.org/abs/1606.04671, 2016. 2", + "[38] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. In International Conference on Learning Representations, 2021. 2", + "[39] Nikunj Saunshi, Arushi Gupta, and Wei Hu. A representation learning perspective on the importance of train-validation splitting in meta-learning. In Proceedings of the 38th International Conference on Machine Learning, pages 9333–9343, 2021. 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Jonathan Schwarz, Jelena Luketina, Wojciech M. Czarnecki, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress and compress: A scalable framework for continual learning. In Proceedings of the International Conference on Machine Learning, 2018. 2", + "[41] Joan Serrá, Dídac Surís, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In Proceedings of the International Conference on Machine Learning, 2018. 2", + "[42] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2", + "[43] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning. 2019. 2", + "[44] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. Advances in neural information processing systems, 2017. 6", + "[45] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing HONG, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In International Conference on Learning Representations, 2022. 1", + "[46] Zhenyi Wang, Tiehang Duan, Le Fang, Qiuling Suo, and Mingchen Gao. Meta learning on a sequence of imbalanced domains with difficulty awareness. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8947-8957, 2021. 2", + "[47] Zhen Wang, Liu Liu, Yajing Kong, Jiaxian Guo, and Dacheng Tao. Online continual learning with contrastive vision transformer. In European Conference on Computer Vision, 2022. 2", + "[48] Zhenyi Wang, Li Shen, Tiehang Duan, Donglin Zhan, Le Fang, and Mingchen Gao. Learning to learn and remember super long multi-domain task sequence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7992, 2022. 2", + "[49] Zhenyi Wang, Li Shen, Le Fang, Qiuling Suo, Tiehang Duan, and Mingchen Gao. Improving task-free continual learning by distributionally robust memory evolution. In International Conference on Machine Learning, pages 22985-22998, 2022. 2", + "[50] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. European Conference on Computer Vision, 2022. 2", + "[51] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2", + "[52] Mengqi Xue, Haofei Zhang, Jie Song, and Mingli Song. Meta-attention for vit-backed continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 150–159, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24530", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Qingsen Yan, Dong Gong, Yuhang Liu, Anton van den Hengel, and Javen Qinfeng Shi. Learning bayesian sparse networks with full experience replay for continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 109-118, 2022. 2", + "[54] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. International Conference on Learning Representations, 2018. 2", + "[55] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. IEEE International Conference on Computer Vision, 2019. 3", + "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. https://arxiv.org/abs/1703.04200, 2017.2", + "[57] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. International Conference on Learning Representations, 2018. 3, 12" + ], + "bbox": [ + 78, + 90, + 470, + 386 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24531", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_model.json b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94479eab3399c4466d7b38c4d9c785b79992cb42 --- /dev/null +++ b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.13, + 0.817, + 0.178 + ], + "angle": 0, + "content": "MetaMix: Towards Corruption-Robust Continual Learning with Temporally Self-Adaptive Data Transformation" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.203, + 0.732, + 0.222 + ], + "angle": 0, + "content": "Zhenyi Wang\\(^{1}\\) Li Shen\\(^{2,*}\\) Donglin Zhan\\(^{3}\\) Qiuling Suo\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.222, + 0.688, + 0.239 + ], + "angle": 0, + "content": "Yanjun Zhu \\(^{1}\\) Tiehang Duan \\(^{4}\\) Mingchen Gao \\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.24, + 0.793, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)State University of New York at Buffalo, USA; \\(^{2}\\)JD Explore Academy, China" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.257, + 0.654, + 0.274 + ], + "angle": 0, + "content": "3 Columbia University, USA; 4 Meta, USA" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.278, + 0.782, + 0.292 + ], + "angle": 0, + "content": "{zhenyiwa, qiulings, yjzhu, mgao8}@buffalo.edu; dz2478@columbia.edu" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.296, + 0.64, + 0.309 + ], + "angle": 0, + "content": "{mathshenli, tiehang.duan}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.474, + 0.695 + ], + "angle": 0, + "content": "Continual Learning (CL) has achieved rapid progress in recent years. However, it is still largely unknown how to determine whether a CL model is trustworthy and how to foster its trustworthiness. This work focuses on evaluating and improving the robustness to corruptions of existing CL models. Our empirical evaluation results show that existing state-of-the-art (SOTA) CL models are particularly vulnerable to various data corruptions during testing. To make them trustworthy and robust to corruptions deployed in safety-critical scenarios, we propose a meta-learning framework of self-adaptive data augmentation to tackle the corruption robustness in CL. The proposed framework, MetaMix, learns to augment and mix data, automatically transforming the new task data or memory data. It directly optimizes the generalization performance against data corruptions during training. To evaluate the corruption robustness of our proposed approach, we construct several CL corruption datasets with different levels of severity. We perform comprehensive experiments on both task- and class-continual learning. Extensive experiments demonstrate the effectiveness of our proposed method compared to SOTA baselines." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.21, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Humans constantly acquire new information throughout their lifespan and easily recognize information shifts such as structure and style variations in images. Continual learning (CL) aims at imitating human's ability to learn from non-stationary data distributions without forgetting the previously learned knowledge. The past few years have witnessed rapid progress in CL research [1, 30, 31, 36, 45]. Despite the success, existing CL systems overlook the robustness" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.346, + 0.895, + 0.468 + ], + "angle": 0, + "content": "against unforeseen data shifts during testing. They assume that training and test images for each task follow the same distribution. However, as data distributions evolve and new scenarios occur, test images often encounter various corruptions such as snow, blur, pixelation, and combinations, resulting in a shifted distribution from the training set. For example, Figure 1 shows various corruptions applied on one image from Split-miniImageNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.715 + ], + "angle": 0, + "content": "Data corruption can drastically impair the performance of existing image recognition systems. A recent study [21] shows that classification accuracy of various architectures has dropped significantly on the ImageNet test set with some simple and natural corruptions. Our empirical evaluation results show that state-of-the-art CL models are even more vulnerable to these corruptions during testing. For example, the accuracy of \\(\\mathrm{DER}++\\) [5] for task-continual learning decreases from \\(93.9\\%\\) to \\(50.5\\%\\) on split-CIFAR10; accuracy drops to \\(10.6\\%\\) from \\(75.6\\%\\) on split-CIFAR100; accuracy decreases to \\(9.8\\%\\) from \\(61.3\\%\\) on split-miniImageNet by applying those common corruptions on test data of each CL task. This severe issue makes existing CL models highly unreliable in safety-critical applications. Thus, improving the robustness of CL models to foster their trustworthiness when deployed in real-world scenarios is essential." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Training a CL model robust to various corruptions is difficult due to the following challenges. 1) Unseen corruptions could perturb the test set far beyond those encountered during training. A model that naively augments training images with seen corruptions cannot generalize to the new ones during testing. Also, it is unrealistic to enumerate all possible corruptions during training since there are infinite types of corruptions and their combinations. 2) With the ever-evolving data distributions in CL, an effective augmentation strategy learned on previous tasks may gradually become less effective because the optimal augmentation strategies are task-dependent and dynamically change over tasks [11]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "24521" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.249, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.183, + 0.228, + 0.195 + ], + "angle": 0, + "content": "(a) Original data" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.089, + 0.366, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.183, + 0.352, + 0.195 + ], + "angle": 0, + "content": "(b) Brightness" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.089, + 0.479, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.184, + 0.463, + 0.195 + ], + "angle": 0, + "content": "(c) Contrast" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.089, + 0.593, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.184, + 0.589, + 0.195 + ], + "angle": 0, + "content": "(d) Defocus Blur" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.09, + 0.721, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.184, + 0.685, + 0.195 + ], + "angle": 0, + "content": "(e) Elastic" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.09, + 0.853, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.184, + 0.806, + 0.196 + ], + "angle": 0, + "content": "(f) Fog" + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.2, + 0.825, + 0.215 + ], + "angle": 0, + "content": "Figure 1. Visualization of five types of different corruption operations on the testing images of Split-miniImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.473, + 0.409 + ], + "angle": 0, + "content": "Although we can adopt a memory buffer to store data from previous tasks, augmenting and replaying them at later training iterations, the augmented memory may gradually become less effective as the model could memorize the stored information after replay runs. Recent approaches, such as Augmix [22] composes and combines multiple pre-defined augmentation operations with different depths and widths, show efficacy in improving robustness under traditional supervised classification tasks. However, these approaches are not directly applicable to corruption-robust CL since they only use a fixed random augmentation strategy that is often not optimal for non-stationary data distributions in CL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.746 + ], + "angle": 0, + "content": "To address these unique challenges of corruption-robustness in CL, we propose a temporally self-adaptive Augmix within a meta-learning framework, named MetaMix. It adaptively augments the memory buffer data and the currently received new data by learning to mix the augmentation operations tailored to the evolving data distributions. In particular, our automatic self-adaptive MetaMix is a bi-level optimization, simulating the evaluation process on unseen corruptions. We randomly divide the training augmentation operations into pseudo-seen and pseudo-unseen operations at each CL step. The lower-level optimization is to optimize the model performance on the pseudo-seen operations; the upper-level optimization is to optimize the generalization of the pseudo-unseen operations. The augmentation strategy is governed by an LSTM, which inputs context information and outputs the corresponding mixing parameters for the augmentations. The proposed MetaMix ensures the augmentation strategy automatically adapts to non-stationary data distribution. Furthermore, the objective is to optimize the performance of the pseudo-unseen corruption operations, which aligns with our goal during testing and encourages the generalization to unseen corruptions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.901 + ], + "angle": 0, + "content": "To evaluate the corruption robustness of existing and the proposed methods, we propose a new challenging benchmark where various corruptions perturb the testing data of each CL task. To facilitate future research, we construct several new datasets, including split-CIFAR-10-C, split-CIFAR-100-C, and split-miniImageNet-C. Extensive experiments on the constructed benchmarks demonstrate the effectiveness of our proposed MetaMix approach compared with several SOTA data-augmentation approaches adapted for CL. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.227, + 0.892, + 0.303 + ], + "angle": 0, + "content": "- To our best knowledge, we are the first to study the corruption-robustness of CL methods. Accordingly, we propose the first set of novel benchmarks for evaluating the corruption-robustness of existing CL methods and moving towards trustworthy CL." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.313, + 0.892, + 0.373 + ], + "angle": 0, + "content": "- We propose a self-adaptive augmentation method, MetaMix, by learning to mix and augment the training data of each CL task to achieve corruption-robustness on unseen corruptions for each CL task." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.385, + 0.894, + 0.445 + ], + "angle": 0, + "content": "- Our method is versatile and can be seamlessly integrated with existing CL methods. Extensive experiments with both task/class continual learning demonstrate the effectiveness of MetaMix." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.227, + 0.894, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.642, + 0.475 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.689, + 0.501 + ], + "angle": 0, + "content": "2.1. Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.895, + 0.903 + ], + "angle": 0, + "content": "CL focuses on learning non-stationary data distribution without forgetting previous knowledge. These methods can be categorized into: 1) retaining memory for future replay [1,4,8,9,17,31,35,42,46,48,49]; 2) designing tailored network architectures [14,24,37,41,54]; 3) performing proper regularization during parameter updates [7,26,40,56]; 4) introducing Bayesian methods for model parameter inference [13,23,34,53]; 5) subspace-based methods [27,30,38] and 6) prompt-tuning [29] based methods [50,51]; 7) transformer-based methods [47,52]. However, all these works focus on the simplified setting in which training and testing data of each CL task follow the same distribution. In practice, the testing data may come from another different distribution due to the variations in weather, light, snow, etc. These distortions pose significant robustness challenges for existing CL methods. Existing literature largely ignored these trustworthy properties of CL. According to the terminology of [43], CL settings can be categorized into three scenarios based on whether task identities of the testing data are available to the CL model during test time. Among the three scenarios, task-continual learning studies the scenario that the task identities are available during testing and is the easiest learning setting; domain incremental learning focuses on the case that all the tasks during CL solve the same classification task with different data distribution and the task identities are unknown during testing; class-continual learning focuses" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24522" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "on the scenario that task identities are unavailable and is the most difficult learning setting. We focus on task/class CL scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.471, + 0.352 + ], + "angle": 0, + "content": "Recently, robust CL began to attract researchers' attention. Self-Purified Replay [25] and PuriDivER [3] consider CL on noisy data stream with label noise. Their work is orthogonal to ours since they consider noisy data during training; we instead consider the robustness to data corruptions during testing. A closely related work to our is Continual Active Adaptation (CAA) [32], which is different from and orthogonal to ours since they focus on the setting that for each data distribution, the label space for each learning task is the same. Each task only encounters a single data corruption. During testing, CAA focuses on generalization to previously seen corruption operations. In contrast, we focus on the setting that each task solves a different task. During testing, the CL model should be robust to unseen data corruptions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.265, + 0.384 + ], + "angle": 0, + "content": "2.2. Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.454 + ], + "angle": 0, + "content": "We categorize existing data augmentation techniques into two classes: 1) improving performance on clean data; 2) improving performance and being robust to data distribution shift and corruptions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.471, + 0.638 + ], + "angle": 0, + "content": "Augmentations for Improving Clean Data Performance. Standard data augmentation aims to improve the generalization performance on clean data. To improve the performance of neural network, standard supervised learning adopts common augmentations, such as random flips, rotations, and crops. Recently, Cutout [12], Mixup [57], CutMix [55] and MaxUp [16] show excellent performance on clean datasets. Other techniques include automatically learning augmentation policy from data, e.g., AutoAugment [10] and RandAugment [11]. However, AutoAugment is computationally impractical for CL problems since it needs a huge computation cost and separate proxy tasks to train the policy network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.471, + 0.792 + ], + "angle": 0, + "content": "Augmentations for Being Robust to Corruptions. Augmix [22] is one of the SOTA methods on the data corruption benchmarks, such as ImageNet-C, CIFAR100-C, etc. DeepAugment [20] and AdversarialAugment [6] transform and augment clean images by perturbing the representations/features of neural networks to generate new augmented images. However, this increases a lot of computation costs. Thus, DeepAugment [20] and AdversarialAugment [6] are impractical for CL. We thus choose the more efficient and effective AugMix and propose a new self-adaptive AugMix for CL." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.812, + 0.212, + 0.83 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this section, we first define the corruption robustness of CL and then describe the proposed MetaMix approach. We describe the notation in the main text. We also provide a notation table in Table 4 in Appendix for ease of reading." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.652, + 0.108 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.894, + 0.312 + ], + "angle": 0, + "content": "Corruption robustness of CL. We consider the problem of learning a sequence of tasks denoted as \\(\\mathcal{D}^{tr} = \\{\\mathcal{D}_1^{tr},\\mathcal{D}_2^{tr},\\dots ,\\mathcal{D}_N^{tr}\\}\\), where \\(N\\) is the number of training tasks. The \\(k\\)-th task training data \\(\\mathcal{D}_k^{tr}\\) consists of many labeled examples \\((\\pmb{x}^k,y^k,\\mathcal{T}_k)\\), where \\(\\pmb{x}^k\\) is the data example in the task, \\(y^{k}\\) is the corresponding data label, and \\(\\mathcal{T}_k\\) is the task identifier. The goal is to learn a model \\(f_{\\theta}\\) on the training task sequence \\(\\mathcal{D}^{tr}\\) so that it performs well on the testing set of all the learned tasks \\(\\mathcal{D}^{te} = \\{\\mathcal{D}_1^{te},\\mathcal{D}_2^{te},\\dots ,\\mathcal{D}_N^{te}\\}\\) without forgetting previously learned knowledge. In particular, we focus on improving the model robustness on the testing set for each task \\(\\mathcal{T}_k\\) under a collection of corruption operations \\(\\mathcal{C}\\). Below, we define the CL corruption robustness." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.324, + 0.894, + 0.4 + ], + "angle": 0, + "content": "Definition 1 (CL corruption robustness). Given a sequence of \\(N\\) tasks with test data \\(\\mathcal{D}^{te} = \\{\\mathcal{D}_1^{te},\\mathcal{D}_2^{te},\\dots ,\\mathcal{D}_N^{te}\\}\\), and a collection of corruption operations \\(\\mathcal{C}\\), the corruption robustness is the expectation accuracy after applying corruptions \\(\\mathcal{C}\\) on the test data of each task \\(k\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.412, + 0.892, + 0.466 + ], + "angle": 0, + "content": "\\[\nA C C = \\frac {1}{N} \\sum_ {k = 1} ^ {k = N} \\mathbb {E} _ {c \\sim \\mathcal {C}} \\left[ \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {k}, y ^ {k}, \\mathcal {T} _ {k}\\right) \\sim D _ {k} ^ {t e}} \\mathbb {I} \\left(f \\left(c \\left(\\boldsymbol {x} ^ {k}\\right)\\right) = y ^ {k}\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.465, + 0.894, + 0.527 + ], + "angle": 0, + "content": "where \\( c(\\pmb{x}^k) \\) is the resulting testing data after applying the corruption operation \\( c \\) on \\( \\pmb{x}^k \\). \\( \\mathbb{I}(.) \\) is the indicator function. ACC measures the average robust accuracy after applying a collection of corruption operations \\( \\mathcal{C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Preliminary of AugMix. AugMix [22] is a recently proposed SOTA augmentation technique that significantly improves model robustness in standard supervised learning. Considering that CL requires the algorithm to be both efficient and effective, we thus choose the efficient and effective AugMix to adapt to CL. AugMix maintains augmented data diversity by mixing several augmentation chains. Suppose we have an augmentation operation set \\(\\mathcal{O} = \\{op_1, op_2, \\dots, op_M\\}\\) during training. The standard AugMix implementation samples three operations \\(op_1\\), \\(op_2\\) and \\(op_3\\) from \\(\\mathcal{O}\\), and composes them with varying depths to form several operation chains; where each chain consists of one to three randomly selected augmentation operations. The augmented image is a convex mix of \\(A\\) chains (\\(A\\) denotes the number of augmentation chains), i.e., \\(\\boldsymbol{x}_{aug} = w_1 \\cdot \\text{chain}_1(\\boldsymbol{x}) + w_2 \\cdot \\text{chain}_2(\\boldsymbol{x}) + \\dots + w_A \\cdot \\text{chain}_A(\\boldsymbol{x})\\), where mixing weights \\(\\boldsymbol{w} = (w_1, w_2, \\dots, w_A)\\) is sampled from the distribution, Drichelet \\((\\alpha, \\alpha, \\dots, \\alpha)\\) with \\(\\alpha = 1\\). With the mixed image \\(\\boldsymbol{x}_{aug}\\), AugMix combines it with the original image through another random convex combination to obtain the final image \\(\\boldsymbol{x}_{augmix}\\), i.e., \\(\\boldsymbol{x}_{augmix} = m\\boldsymbol{x} + (1 - m)\\boldsymbol{x}_{aug}\\), where \\(m\\) is sampled from a \\(\\operatorname{Beta}(\\alpha, \\alpha)\\) distribution. AugMix adopts a fixed random augmentation policy, i.e., Dirichlet and Beta distributions, during the entire training process." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24523" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.094, + 0.457, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.157, + 0.357, + 0.17 + ], + "angle": 0, + "content": "(a) Continual Learning Process" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.172, + 0.46, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.302, + 0.307, + 0.314 + ], + "angle": 0, + "content": "(b) AugMix" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.316, + 0.462, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.456, + 0.327, + 0.468 + ], + "angle": 0, + "content": "(c) MetaMix (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.473, + 0.473, + 0.737 + ], + "angle": 0, + "content": "Figure 2. Comparisons between AugMix [22] and our proposed MetaMix. (a) Tasks sequentially arrive during CL. (b) At each CL step, AugMix adopts a fixed and random augmentation policy during the entire learning process, thus lacking adaptation to the non-stationary data in CL and may not be the optimal strategy for each task. (c) MetaMix automatically adapts to the non-stationary data distribution during CL by meta-learning the mixing parameters of augmentation. At time \\( t \\), we sample a batch of data \\( (\\pmb{x}_r, y_r) \\) from the memory buffer and concatenate them with the current received data \\( (\\pmb{x}_t, y_t) \\) as \\( (\\pmb{x}_b, y_b) \\). We apply MetaMix to augment each data example. Specifically, we feed the context information \\( I_t \\) into the LSTM cell with parameters \\( \\phi_t \\), and obtain the mixing weights \\( w_t \\) and \\( m_t \\). The CL model \\( f_{\\theta_t} \\) calculates the losses for the augmented data. We adopt bi-level optimization to improve model generalization on unseen corruptions by splitting the augmentation operations into a pseudo-seen set \\( S_t \\) and pseudo-unseen set \\( \\mathcal{U}_t \\). The loss \\( \\mathcal{L}_{\\phi_t} \\) is calculated on data augmented with \\( \\mathcal{U}_t \\) to update \\( \\phi_t \\), and loss \\( \\mathcal{L}_{\\theta_t} \\) is calculated on data augmented with \\( S_t \\) to update \\( \\theta_t \\). We repeat this process for all tasks in the sequence." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.748, + 0.472, + 0.793 + ], + "angle": 0, + "content": "The augmentation procedure of AugMix is illustrated in Figure 2 (b). The optimization objective of AugMix is shown in the following:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.805, + 0.473, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {*} = \\arg \\min _ {\\boldsymbol {\\theta}} \\underset {o p \\in \\mathcal {O}} {\\mathbb {E}} \\mathcal {L} (\\boldsymbol {x}, y, \\boldsymbol {\\theta}) + \\lambda J S (\\boldsymbol {x}, \\boldsymbol {x} _ {a u g m i x 1}, \\boldsymbol {x} _ {a u g m i x 2}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(\\pmb{x}\\) is the original data and \\(y\\) is the data label; \\(\\mathcal{L}(\\pmb{x},\\pmb{y},\\pmb{\\theta})\\) is the loss function on the labeled data (cross-entropy for classification); \\(\\pmb{x}_{augmix1}\\) and \\(\\pmb{x}_{augmix2}\\) are two augmenta" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "tions of original data \\( \\mathbf{x} \\) by AugMix; \\( JS \\) denotes the Jensen Shannon divergence for regularizing the output consistency between the original and augmented data. We will define it in the following section." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.165, + 0.688, + 0.181 + ], + "angle": 0, + "content": "3.2. MetaMix Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.189, + 0.895, + 0.582 + ], + "angle": 0, + "content": "Direct extension of AugMix [22] for CL is to augment data examples sampled from the memory buffer, and the received new data at each time point for memory-based method across the sequentially arrived tasks. In the following, we present our methods for memory-based CL methods, but our methods can be easily extended to other families of CL methods as well, also demonstrated in Section 4. However, since the memory buffer is usually small for CL models, using a fixed random augmentation strategy (like AugMix) may make the CL model gradually memorize the training data corruption operations, but fail to generalize to unseen corruptions during testing. Moreover, with the non-stationary nature of the data, the optimal augmentation strategy for image mixing is data-dependent, and the mixing weights should be specifically optimized for each task. However, since we cannot repeatedly tune the mixing weights for each task during CL learning process, an automatic self-adaptive process is necessary to achieve good performance. We thus introduce the details of our self-adaptive data augmentation approach, namely MetaMix, for improving robustness in CL. Figure 2 (c) illustrates the training process of MetaMix at time \\( t \\). Specifically, we adopt meta-learning to learn a set of data-dependent parameters \\( w_{t} \\) and \\( m_{t} \\), which can dynamically control the mixing weights of augmentation operations and automatically adapt to the non-stationary data distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.894, + 0.704 + ], + "angle": 0, + "content": "First, to memorize previous knowledge, we allocate a small memory buffer \\(\\mathcal{M}\\) to store data from previous tasks by reservoir sampling, similar to [35]. The data in the memory buffer will be replayed later with a new task together. Suppose we have \\(M\\) augmentation operations \\(\\mathcal{O} = \\{op_1, op_2, \\dots, op_M\\}\\) during training, CL model parameters at time \\(t\\) are \\(\\theta_t\\), and the function represented by the network with \\(\\theta_t\\) is \\(f_{\\theta_t}(x)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.872 + ], + "angle": 0, + "content": "To capture the sequential dependencies among previous tasks, we design a self-adaptive LSTM mixing strategy. The LSTM inputs the current context information and outputs the adaptive mixing parameters. We use an LSTM (dynamic) model to dynamically change the mixing weights instead of an MLP (static) model since the tasks in CL sequentially arrive. The LSTM can encode the information of previous tasks into an additional hidden state while the MLP can only capture the data information of the current task. It does not capture the information of previous tasks without the hidden state. Formally, the LSTM adaptive process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.886, + 0.893, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {o} _ {t + 1}, \\boldsymbol {h} _ {t + 1}, \\boldsymbol {g} _ {t + 1} = \\operatorname {L S T M} _ {\\phi_ {t}} \\left(\\boldsymbol {I} _ {t}, \\boldsymbol {h} _ {t}, \\boldsymbol {g} _ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24524" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "where \\(\\phi_t\\) is the LSTM MetaMixer parameters at time \\(t\\), \\(h_t\\) is the hidden state of LSTM at time \\(t\\), \\(g_t\\) is the cell state of LSTM at time \\(t\\), \\(I_t\\) is the context information encoding as input to LSTM at time \\(t\\); \\(o_{t+1}\\) is the output of LSTM at time \\(t+1\\) for the mixing parameters, i.e., \\(o_{t+1} = \\{w_t', m_t'\\}\\), where \\(m_t'\\) is the mixing weight for the augmented data and original raw data and \\(w_t'\\) are the mixing weights for different chains in the augmentation. We apply softmax function to \\(w_t'\\) and sigmoid function to \\(m_t'\\) to ensure they are within [0,1] interval, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.25, + 0.471, + 0.267 + ], + "angle": 0, + "content": "\\[\nm _ {t} = \\operatorname {S i g m o i d} \\left(m _ {t} ^ {\\prime}\\right), \\boldsymbol {w} _ {t} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {w} _ {t} ^ {\\prime}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.272, + 0.473, + 0.304 + ], + "angle": 0, + "content": "The context \\(I_{t}\\) contains essential information to characterize data in current and previous tasks. It is shown as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.324, + 0.47, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {t} = \\left\\{\\nabla \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {y} _ {r}\\right) \\cdot \\nabla \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {t}\\right), \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {y} _ {r}\\right), \\boldsymbol {e} _ {r}, \\boldsymbol {e} _ {t} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.473, + 0.596 + ], + "angle": 0, + "content": "where \\((\\pmb{x}_r, y_r)\\) is the labeled mini-batch data sampled from the memory buffer and \\((\\pmb{x}_t, y_t)\\) is the labeled data received at time \\(t\\). We adopt this information for several reasons: (1) \\(\\nabla \\mathcal{L}_{\\pmb{\\theta}_t}(\\pmb{x}_r, y_r) \\cdot \\nabla \\mathcal{L}_{\\pmb{\\theta}_t}(\\pmb{x}_t, y_t)\\) is the gradient dot product between memory data gradient and current mini-batch data gradient. It determines the degree of interference and transfer between memory data and current task mini-batch data. It is commonly used in existing CL as context information [18, 35]; (2) we also input the memory mini-batch data loss, as it indicates the degree of forgetting of previous tasks; (3) \\(\\pmb{e}_r\\) and \\(\\pmb{e}_t\\) are data encoding for mini-batch data, \\(\\pmb{x}_r\\), sampled from memory buffer and currently received new mini-batch data, \\(\\pmb{x}_t\\), respectively. That is, \\(\\pmb{e}_r\\) and \\(\\pmb{e}_t\\) are the last layer features outputted by ResNet-18. This determines the data-dependent features for calculating the data-dependent mixing weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.473, + 0.839 + ], + "angle": 0, + "content": "Since corruption operations on testing data are unseen during training, we cannot directly optimize the augmentation strategy for these corruptions. To make the CL model generalize better to unseen corruptions, at each training step \\( t \\), we randomly split the augmentation operations \\( \\mathcal{O} \\) into non-overlapping pseudo-seen \\( S_{t} \\) and pseudo-unseen \\( \\mathcal{U}_t \\) operations, i.e., \\( S_{t} \\cup \\mathcal{U}_{t} = \\mathcal{O} \\) and \\( S_{t} \\cap \\mathcal{U}_{t} = \\emptyset \\). We use the pseudo operation sets to simulate the scenario in which training and testing corruptions have no overlap. We formulate MetaMix as a bi-level optimization problem to make the CL model learn to augment and mix. This ensures that the update of model parameters \\( \\theta \\) on the pseudo-seen set can improve the optimization of MetaMixer parameters \\( \\phi \\) on the pseudo-unseen set, which could, in turn, be beneficial to improve the CL model generalization. We formulate the MetaMix as the following bi-level optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.845, + 0.47, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\phi_ {*} = \\arg \\min _ {\\phi} \\mathbb {E} _ {o p \\in \\mathcal {U} _ {t}} \\mathcal {L} \\left(\\boldsymbol {x} _ {b}, y _ {b}, \\boldsymbol {\\theta} _ {*}, \\phi\\right) + \\lambda J S \\left(\\boldsymbol {x} _ {b}, \\widehat {\\boldsymbol {x}} _ {b 1}, \\widehat {\\boldsymbol {x}} _ {b 2}\\right), \\tag {6} \\\\ \\boldsymbol{\\theta}_{*} = \\arg \\min_{\\boldsymbol{\\theta}}\\underset {op\\in \\mathcal{S}_{t}}{\\mathbb{E}}\\mathcal{L}(\\boldsymbol{x}_{b},y_{b},\\boldsymbol {\\theta},\\boldsymbol {\\phi}) + \\lambda J S(\\boldsymbol{x}_{b},\\boldsymbol{x}^{\\prime}_{b1},\\boldsymbol{x}^{\\prime}_{b2}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "where we concatenate the mini-batch data \\((\\pmb{x}_r, y_r)\\) from the memory buffer and the current received mini-batch data \\((\\pmb{x}_t, y_t)\\) together as \\((\\pmb{x}_b, y_b)\\); \\(\\widehat{\\pmb{x}}_{b1}\\) and \\(\\widehat{\\pmb{x}}_{b2}\\) are the two minibatch data augmented by applying the pseudo-unseen augmentation operations \\(\\mathcal{U}_t\\) on \\((\\pmb{x}_b, y_b)\\); \\(\\pmb{x}_{b1}'\\) and \\(\\pmb{x}_{b2}'\\) are the two mini-batch data augmented by applying the pseudo-seen augmentation operations \\(\\mathcal{S}_t\\) on \\((\\pmb{x}_b, y_b)\\). \\(\\lambda\\) is the regularization constant for the consistency loss. The lower-level optimization is to optimize the performance on the pseudoseen augmentation operations \\(\\mathcal{S}_t\\) at the training step \\(t\\); and the upper-level optimization is to optimize the generalization on the pseudo-unseen augmentation operations \\(\\mathcal{U}_t\\). This bi-level optimization for learning to augment and mix can be efficiently solved by first-order gradient descent method, similar to the first-order MAML [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.895, + 0.506 + ], + "angle": 0, + "content": "The advantage of MetaMix (Eq. (6)) over AugMix (Eq. (2)) are two folds. First, MetaMix improves the generalization to unseen corruptions since we split the training data augmentation operations into pseudo-seen and unseen operations. This is similar to the train-validation data split in standard meta-learning, which improves the generalization compared to the training without data split [39]; It could be understood as introducing an implicit regularization so that it can benefit from low sample complexity [39]. Second, MetaMix can automatically adapt to the non-stationary data distribution in CL, while AugMix can only work well in the stationary data distribution in a single task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.895, + 0.648 + ], + "angle": 0, + "content": "Consistency Regularization To ensure the network output consistency between the original raw image data and the augmented data, we use a similar consistency loss [22] to regularize the network output to ensure smoother neural network responses. We hope the model to have similar responses to \\( \\pmb{x}_b, \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2} \\). We minimize the Jensen-Shannon divergence among the posterior distributions of the original sample \\( \\pmb{x}_b \\) and its augmented variants \\( \\widehat{\\pmb{x}}_{b1} \\) and \\( \\widehat{\\pmb{x}}_{b2} \\). The consistency loss is defined as below:" + }, + { + "type": "equation", + "bbox": [ + 0.495, + 0.673, + 0.892, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {m e a n} = \\left(p _ {\\boldsymbol {x} _ {b}} + p _ {\\hat {\\boldsymbol {x}} _ {b 1}} + p _ {\\hat {\\boldsymbol {x}} _ {b 2}}\\right) / 3 (7) \\\\ J S \\left(\\boldsymbol {x} _ {b}, \\widehat {\\boldsymbol {x}} _ {b 1}, \\widehat {\\boldsymbol {x}} _ {b 2}\\right) = \\left(\\mathbb {K L} \\left(p _ {\\boldsymbol {x} _ {b}} \\mid p _ {\\text {m e a n}}\\right) + \\mathbb {K L} \\left(p _ {\\widehat {\\boldsymbol {x}} _ {b 1}} \\mid p _ {\\text {m e a n}}\\right) \\right. \\\\ + \\mathbb {K L} \\left(p _ {\\widehat {x} _ {b 2}} \\mid p _ {\\text {m e a n}}\\right)) / 3 (8) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Where \\(\\mathbb{KL}\\) denotes the KL divergence between two distributions. \\(p_{\\boldsymbol{x}_b} = f_{\\boldsymbol{\\theta}_t}(\\boldsymbol{x}_b)\\) is the network output probabilities of each class for original raw data \\(\\boldsymbol{x}_b\\). Similarly, we can define \\(p_{\\widehat{\\boldsymbol{x}}_{b1}} = f_{\\boldsymbol{\\theta}_t}(\\widehat{\\boldsymbol{x}}_{b1})\\) and \\(p_{\\widehat{\\boldsymbol{x}}_{b2}} = f_{\\boldsymbol{\\theta}_t}(\\widehat{\\boldsymbol{x}}_{b2})\\). We summarize our proposed algorithm in Algorithm 3. Line 5-8 takes the context information as input and computes desired augmentation mixing parameters. Line 9-11 augments the original data with the pseudo-seen augmentation operations \\(\\mathcal{S}_t\\) and pseudo-unseen operations \\(\\mathcal{U}_t\\) by Algorithm 1. Line 12 solves the bi-level optimization problem (Eq. (6)) by Algorithm 2." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24525" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.091, + 0.229, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 DataMix" + }, + { + "type": "algorithm", + "bbox": [ + 0.084, + 0.11, + 0.473, + 0.347 + ], + "angle": 0, + "content": "1: Function DataMix \\((\\pmb{x}_b, \\pmb{w}_t, m_t, \\mathcal{V})\\) \n2: MetaMix augmented images initialized as \\(\\pmb{x}_b' = 0\\) \n3: \\(\\pmb{x}_{orig} = \\pmb{x}_b\\) \n4: for \\(i = 1\\) to \\(A\\) do { A is the number of augmentation chains} \n5: \\(\\pmb{x}_b = \\pmb{x}_{orig}\\) \n6: for \\(l = 1\\) to \\(d\\) do { \\(d\\) is the augmentation depth} \n7: randomly sample operation \\(op \\in \\mathcal{V}\\) (\\(\\mathcal{V}\\) is the collection of augmentation operations) \n8: \\(\\pmb{x}_b = op(\\pmb{x}_b)\\) \n9: end for \n10: \\(\\pmb{x}_b' + = \\pmb{w}_t^i \\cdot \\pmb{x}_b\\) \n11: end for \n12: \\(\\pmb{x}_b' = m_t \\pmb{x}_{orig} + (1 - m_t) \\pmb{x}_b'\\); \n13: similar to the above procedure, we can get two different minibatch \\(\\pmb{x}_{b1}', \\pmb{x}_{b2}'\\) \n14: return \\(\\pmb{x}_{b1}', \\pmb{x}_{b2}'\\) \n15: EndFunction" + }, + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.366, + 0.27, + 0.381 + ], + "angle": 0, + "content": "Algorithm 2 Bi-level Solver" + }, + { + "type": "algorithm", + "bbox": [ + 0.083, + 0.385, + 0.473, + 0.58 + ], + "angle": 0, + "content": "1: Function Bi-level Solver \\((\\pmb{\\theta}_t, \\pmb{\\phi}_t, \\pmb{x}_b, y_b, \\pmb{x}_{b1}', \\pmb{x}_{b2}', \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2})\\) \n2: optimize the lower-level optimization problem in Eq. (6) by \\(\\pmb{\\theta}_{t+1} = \\pmb{\\theta}_t - \\gamma \\nabla_\\theta \\mathbb{E}_{op \\in S_t}[\\mathcal{L}(\\pmb{x}_b, y_b, \\pmb{\\theta}_t, \\phi) + \\lambda JS(\\pmb{x}_b, \\pmb{x}_{b1}', \\pmb{x}_{b2}')]\\) \n3: calculate the upper-level optimization loss function in Eq. (6) by \\(\\mathcal{E}(\\pmb{\\theta}_{t+1}, \\phi) = \\mathbb{E}_{op \\in U_t}[\\mathcal{L}(\\pmb{x}_b, y_b, \\pmb{\\theta}_{t+1}, \\phi) + \\lambda JS(\\pmb{x}_b, \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2})]\\) \n4: \\(\\phi_1' = \\phi_t\\) \n5: for \\(s = 1\\) to \\(J\\) do \n6: \\(\\phi_{s+1}' = \\phi_s' - \\beta \\nabla_{\\phi'} \\mathcal{E}(\\pmb{\\theta}_{t+1}, \\phi_s')\\) \n7: end for \n8: \\(\\phi_{t+1} = \\phi_{J+1}'\\) \n9: return \\(\\pmb{\\theta}_{t+1}, \\phi_{t+1}\\) \n10: EndFunction" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.202, + 0.627 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.697 + ], + "angle": 0, + "content": "We evaluate the model performance for both task- and class-continual learning. The CL models are evaluated on the proposed benchmark for each dataset. We then describe experiment setup, results, and ablation study." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.254, + 0.726 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Below, we first construct the corruption CL benchmark and then describe the evaluation metrics, baselines, and implementation details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Corruption CL benchmark. Existing CL benchmarks, such as Split-CIFAR10, Split-CIFAR100, and Split-miniImageNet, are commonly used in literature. However, each task's training and testing data follow the same distribution for these datasets. Thus, these benchmarks are not suitable for evaluating the corruption robustness of CL methods. To properly evaluate the corruption robustness, we follow the protocols [21] and create several new CL datasets." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.655, + 0.106 + ], + "angle": 0, + "content": "Algorithm 3 MetaMix" + }, + { + "type": "algorithm", + "bbox": [ + 0.507, + 0.109, + 0.895, + 0.513 + ], + "angle": 0, + "content": "1: REQUIRE: Augmentation Operations \\(\\mathcal{O}\\) \\(=\\) {sterize, rotate, ..., solarize}; augmentation width \\(A\\) ; augmentation depth \\(d\\) ; the number of CL tasks N; the number of training iterations for each task \\(\\mathcal{T}_k\\) is \\(N_{k}\\) ; memory buffer \\(\\mathcal{M}\\) ; CL model parameters \\(\\theta\\) ; LSTM MetaMix with parameters \\(\\phi\\) ; MetaMixer LSTM update steps \\(J\\) ; CL model parameter learning rate \\(\\gamma\\) ; regularization weight \\(\\lambda\\) ; original image \\(\\pmb{x}_{orig}\\) ; initial hidden state and cell state \\(\\pmb{h}_t,\\pmb{g}_t\\) are initialized as random noise \\(\\pmb{h}_1,\\pmb{g}_1\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})\\) \n2: for \\(k = 1\\) to \\(N\\) do \n3: for \\(t = 1\\) to \\(N_{k}\\) do \n4: randomly sample mini-batch data \\((x_{r},y_{r})\\) from memory buffer \\(\\mathcal{M}\\) and concatenate it with the current received mini-batch data \\((x_{t},y_{t})\\) to obtain \\((x_{b},y_{b})\\) \n5: compute context vector \\(\\pmb{I}_t = \\{\\nabla \\mathcal{L}_{\\theta_t}(\\pmb{x}_r,\\pmb{y}_r)\\) \\(\\nabla \\mathcal{L}_{\\theta_t}(\\pmb{x}_t,\\pmb{y}_t),\\mathcal{L}_{\\theta_t}(\\pmb{x}_r,\\pmb{y}_r),\\pmb{e}_r,\\pmb{e}_t\\}\\) \n6: \\(o_{t + 1},h_{t + 1},g_{t + 1} = \\mathrm{LSTM}_{\\phi_t}(I_t,h_t,g_t)\\) \n7: generate mixing parameters: \\(\\pmb{w}_t^{\\prime},m_t^{\\prime} = o_{t + 1}\\) \n8: \\(m_t = \\mathrm{Sigmoid}(m_t^{\\prime}),w_t = \\mathrm{Softmax}(w_t^{\\prime})\\) \n9: randomly split augmentation operations \\(\\mathcal{O}\\) into nonoverlapping pseudo-seen \\(S_{t}\\) and unseen \\(\\mathcal{U}_t\\) \n10: \\(\\pmb{x}_{b1}^{\\prime},\\pmb{x}_{b2}^{\\prime} = \\mathrm{DataMix}(\\pmb{x}_b,\\pmb{w}_t,m_t,\\mathcal{S}_t)\\) (Algorithm 1) \n11: \\(\\widehat{\\pmb{x}}_{b1},\\widehat{\\pmb{x}}_{b2} = \\mathrm{DataMix}(\\pmb{x}_b,\\pmb{w}_t,m_t,\\mathcal{U}_t)\\) (Algorithm 1) \n12: \\(\\theta_{t + 1},\\phi_{t + 1} = \\mathrm{Bi - level~Solver~}(\\theta_t,\\phi_t,\\pmb{x}_b,\\pmb{y}_b,\\pmb{x}_{b1}'',\\pmb{x}_{b2}'',\\widehat{\\pmb{x}}_{b1},\\widehat{\\pmb{x}}_{b2})\\) (Algorithm 2) \n13: update memory buffer \\(\\mathcal{M}\\) through reservoir sampling [35] to determine whether to store \\((x_{t},y_{t})\\) \n14: end for \n15: end for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.852 + ], + "angle": 0, + "content": "We apply 15 common corruption operations to the testing data of the three CL benchmarks, such as snow, frost, fog, etc. We provide a brief description for each corruption operation in Table 5 in Appendix. We perform each corruption at five different severity levels. Note that these corruptions only appear on the testing dataset of each continually learned task and are not seen in the training data of each task. This is to ensure that we can measure the robustness to unseen corruptions during testing for each compared method. To our best knowledge, we build the first CL corruption benchmark for measuring the generalization of CL models to unseen corruptions during testing. We name these new datasets Split-CIFAR10-C, Split-CIFAR100-C, and Split-miniImageNet-C. Specifically, for Split-CIFAR-10-C, we split the CIFAR-10 dataset [28] into 5 disjoint tasks, where each task has 2 classes; for Split-CIFAR-100-C, we split CIFAR-100 [28] consisting of 100 classes into 10 disjoint tasks, where each task has 10 classes; for Split-miniImageNet-C, we split miniImageNet [44] that consists of 100 classes, into 10 disjoint tasks, where each task has 10 classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the performance of different methods with average accuracy and backward transfer at the end of CL training to measure the final performance" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24526" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "and the extent of forgetting for different methods. For each task, the accuracy for a specific corruption \\( c \\) is the average accuracy over the five severity level. Suppose for task \\( k \\), \\( ACC_{c,s}^{k} \\) measures the robust accuracy after applying the corruption operation \\( c \\) with severity level \\( s \\):" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.176, + 0.432, + 0.197 + ], + "angle": 0, + "content": "\\[\nA C C _ {c, s} ^ {k} = \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {k}, y ^ {k}, \\mathcal {T} _ {k}\\right) \\sim D _ {t e} ^ {t e}} \\mathbb {I} \\left(f \\left(c _ {s} \\left(\\boldsymbol {x} ^ {k}\\right)\\right) = y ^ {k}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.206, + 0.47, + 0.283 + ], + "angle": 0, + "content": "where \\( c_{s}(\\pmb{x}^{k}) \\) is the corruption operation \\( c \\) with severity level \\( s \\) applied on \\( \\pmb{x}^{k} \\). The accuracy for corruption \\( c \\) denoted as \\( ACC_{c}^{k} \\) is the average robust accuracy across all the five severity levels. The overall accuracy for task \\( k \\), \\( ACC^{k} \\), is the average accuracy across all corruption types, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.292, + 0.463, + 0.332 + ], + "angle": 0, + "content": "\\[\nA C C _ {c} ^ {k} = \\frac {1}{5} \\sum_ {s = 1} ^ {s = 5} A C C _ {c, s} ^ {k}, \\text {a n d} A C C ^ {k} = \\frac {1}{| \\mathcal {C} |} \\sum_ {c \\in \\mathcal {C}} A C C _ {c} ^ {k}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.47, + 0.493 + ], + "angle": 0, + "content": "To measure catastrophic forgetting, we also evaluate the backward transfer (BWT), which measures the extent of forgetting on previous tasks after learning new ones. \\( BWT < 0 \\) reveals the occurrence of catastrophic forgetting on previous tasks, and \\( BWT > 0 \\) indicates that learning new tasks is helpful for previous tasks. Formally, the backward transfer (BWT) is defined as: \\( BWT = \\frac{1}{N - 1}\\sum_{k = 1}^{k = N - 1}(R_{N,k} - R_{k,k}) \\); where \\( R_{N,k} \\) is the testing accuracy on task \\( k \\) after learning on task \\( N \\) and \\( R_{N,k} \\) is defined the same as the above mentioned \\( ACC^k \\). For BWT, the higher, the better." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.493, + 0.472, + 0.569 + ], + "angle": 0, + "content": "Continual Learning Setting. We apply our proposed MetaMix to task-continual learning (Task-CL) and class-continual learning (Class-CL). The former provides task identities for the CL learner to select the relevant classifier for each example during testing, whereas the latter does not." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.78 + ], + "angle": 0, + "content": "Baseline. Our method can be seamlessly integrated with exiting CL methods. For illustration, we integrate the proposed methods with SOTA memory-based methods: \\( DER++ \\) [5] and CLS-ER [2]. To evaluate the effectiveness of the proposed MetaMix, we compare it with several SOTA data augmentation approaches originally designed for traditional supervised learning: Adversarial Training (AT) [33], RandAugment (RA) [11], Maxup [16], DeepAugment [20], Augmix [22]. We adapt these augmentation methods for CL problems by applying them on the new task data and the memory buffer data. We provide more detailed descriptions of baselines in Appendix. Due to space limitations, we put the experiments that integrate our proposed method with CLS-ER in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Implementation details. We use ResNet18 [19] as classifier for all datasets. All the other hyperparameter settings follow from [5]. The types of corruption during testing are not seen during training for all the compared augmentation methods to evaluate the robustness against unseen corruptions. Following [22], the augmentation operations performed during training are autocontrast, equalize, posterize, rotate, solarize, shear-x, shear-y, translate-x, translate-y." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "All the above-compared data augmentation methods apply these augmentation operations to CL baseline method to improve robustness. The memory buffer has capacity of 500 data points by default. We average the result for 5 runs for each experiment. We provide both mean and standard deviation across different runs. Due to space limitations, we provide standard deviation results in Appendix. We provide more implementation details in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.31 + ], + "angle": 0, + "content": "LSTM Architecture. It has one recurrent layer with 5 hidden units. A linear layer (fully connected layer) is appended next to the LSTM output to generate the mixing parameters \\( \\boldsymbol{w}_t \\) and \\( m_t \\). The LSTM only has about 20K parameters and is negligible compared to ResNet18, which has more than 11.22 million parameters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.337, + 0.715, + 0.352 + ], + "angle": 0, + "content": "4.2. Robustness of Task-CL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.895, + 0.473 + ], + "angle": 0, + "content": "Task-Continual Learning In this section, we evaluate the corruption robustness under task-continual learning, where the task identifiers are provided to the CL learner during testing. We present the results that integrate baselines and the proposed method with DER++ [5]. Due to space limitations, we put the experiments that integrate baselines and MetaMix with CLS-ER [2] in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Results. We show the results across different corruptions and compared methods in Tables 1-2; where \\(Avg\\) is the average robust accuracy across all the corruptions. Due to the space limitations, we provide the standard deviation across different runs in Table 8 and 9 in Appendix. We also put results on split-CIFAR10-C in Table 6 and Table 7 in Appendix. We can observe that without data augmentation, the performance of \\(\\mathrm{DER}++\\) is close to random guessing, indicating that the CL model does not generalize to common corruptions. The proposed self-adaptive MetaMix brings significant improvement toward corruption robustness. It outperforms the average corruption robustness accuracy of the other models by a large margin of \\(2.4\\%\\), \\(2.6\\%\\) on Split-CIFAR100-C, and Split-miniImageNet-C, respectively. The improvement shows the effectiveness of proposed MetaMix, which can automatically adapt to the non-stationary data distribution. The compared methods only use fixed static/random augmentation methods which lack adaptation to the non-stationary data in CL, and thus do not perform well. Random augmentation achieves second-best results. Deep-augment would distort the image such that they are very different from original images. Adversarial training and Maxup do not help much corruption robustness among the compared methods. Adversarial training can generate similar images compared to original images. Thus, there is a large gap between training and testing corrupted data for those methods adapted for CL. Maxup aims for improving clean data performance, not designed for robust accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24527" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.261, + 0.089, + 0.709, + 0.104 + ], + "angle": 0, + "content": "Table 1. Robust accuracy of Task-CL on Split-CIFAR100-C with DER++" + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.108, + 0.86, + 0.216 + ], + "angle": 0, + "content": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++10.4610.6910.2610.7610.5110.7110.6610.4710.5610.5110.7910.5410.7510.5510.7210.60
AT10.3610.4510.2510.2310.2410.2710.2510.129.889.9810.169.8310.1910.310.3110.19
RA26.7727.1525.4627.8423.4624.0525.9923.2626.3430.4527.9528.825.5429.1229.0926.75
DeepAugment11.0011.1610.8611.3511.1511.4511.3310.7610.4610.5911.1010.5711.2011.3611.2511.04
Maxup11.111.1711.1311.0911.0611.1311.110.8810.7610.8110.8610.3711.0611.1311.1210.98
Augmix59.5862.6859.9267.9558.8865.9367.0363.3963.9363.2468.6263.2765.8266.0664.0764.02
Ours61.5365.1461.1870.3561.2868.0168.7666.0266.4966.1471.4966.1668.2468.9166.7566.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.249, + 0.219, + 0.72, + 0.232 + ], + "angle": 0, + "content": "Table 2. Robust accuracy of Task-CL on Split-MiniImageNet-C with DER++" + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.233, + 0.859, + 0.34 + ], + "angle": 0, + "content": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++10.2710.110.079.789.739.679.729.619.719.859.6910.159.559.459.579.79
AT9.079.249.279.659.629.69.489.179.109.019.128.389.549.569.359.28
RA17.2318.2216.719.8116.618.3219.0718.9417.1418.2617.2614.5621.1420.8423.018.47
DeepAugment10.5410.5310.6210.6810.7010.6310.7410.5610.3210.3210.529.9210.5510.4110.4910.50
Maxup10.410.910.6810.659.7510.2110.259.6210.1210.2110.3410.910.210.1610.0710.30
Augmix43.4548.9142.2053.7849.6859.8851.8353.2254.0255.6359.5943.4661.3551.3462.7552.74
Ours48.853.3343.8159.753.6462.7856.9556.1957.3156.0862.0741.7163.1749.7864.5555.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.259, + 0.342, + 0.71, + 0.355 + ], + "angle": 0, + "content": "Table 3. Robust accuracy of Class-CL on Split-CIFAR100-C with DER++" + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.356, + 0.859, + 0.465 + ], + "angle": 0, + "content": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++0.750.750.880.750.860.770.680.880.960.920.750.960.780.780.790.82
AT0.990.971.040.910.950.970.920.831.050.940.921.00.910.950.930.95
RA6.226.35.676.254.424.545.595.256.247.726.426.935.166.817.116.04
DeepAugment1.291.291.251.261.261.231.241.261.241.241.141.261.221.281.331.25
Maxup1.171.151.161.151.081.141.121.081.081.091.141.061.131.151.111.12
Augmix19.5721.2120.3526.1819.0825.1725.4622.0922.4322.8226.4422.5423.7224.6423.0122.98
Ours22.3224.4423.1529.1922.1328.1028.3525.4725.226.0829.5625.7827.0128.1625.8226.05
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.296, + 0.497 + ], + "angle": 0, + "content": "4.3. Robustness of Class-CL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.471, + 0.566 + ], + "angle": 0, + "content": "Class-Continual Learning. In this section, we evaluate the corruption robustness under the more challenging class-continual learning setting, where the task identifiers are unavailable to the CL learner during testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.471, + 0.747 + ], + "angle": 0, + "content": "Results. We present the results across different corruptions and compared methods on Split-CIFAR100-C in Table 3, we present the results on Split-miniImageNet-C in Table 10 in Appendix. Similar to the Task-CL, MetaMix significantly improves over the baseline augmentation methods in Class-CL setting due to its self-adaptivity property. It outperforms the average corruption robustness accuracy of the other models by a large margin of \\(3.1\\%\\), \\(2.2\\%\\) on Split-CIFAR100-C, and Split-miniImageNet-C, respectively. This is because our proposed method can automatically adapt to the non-stationary distributions. Most compared methods do not perform well on this challenging setup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Due to space limitations, we put BWT results in Table 15 in Appendix. In the corruption-robustness scenario, BWT is no longer a meaningful metric with such extreme accuracy differences, as the significantly lower accuracy of comparison methods results in much less space for further performance variations during backward transfer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.468, + 0.865 + ], + "angle": 0, + "content": "4.4. Ablation Study and Hyperparameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Due to the limited space, we provide (1) ablation studies; (2) hyperparameter analyses, including \\(\\lambda\\), \\(J\\), \\(\\beta\\), etc.; (3)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.482, + 0.894, + 0.529 + ], + "angle": 0, + "content": "The effect of using LSTM vs MLP and the benefit of using additional hidden state information from previous tasks; (4) computation cost in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.895, + 0.64 + ], + "angle": 0, + "content": "Effect of Memory Size. We evaluate the effect of memory size with 500 and 3000, respectively. The memory size of 500 is the default setting in the above tables. We provide experiment results on Split-CIFAR100-C, Split-miniImageNet-C with memory size 3000 on task-CL and class-CL respectively in Table 11, 12, 13, 14 in Appendix. In these cases, our method significantly outperforms baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.619, + 0.685 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.895, + 0.851 + ], + "angle": 0, + "content": "This paper tackles a more challenging problem of corruption robustness in CL with non-stationary data distribution, where the testing data distribution may significantly differ from training data distribution with various forms of common corruptions. We propose a meta-learning framework, MetaMix, for self-adaptive data augmentation specialized for CL. Our proposed MetaMix can substantially improve the model's robustness. Comprehensive experiments on both task- and class-CL settings demonstrate the effectiveness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement We thank all the anonymous reviewers for their insightful and thoughtful comments. This research was supported in part by grant NSF III-1910492." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24528" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.472, + 0.185 + ], + "angle": 0, + "content": "[1] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. Advances in Neural Information Processing Systems 32, pages 11849-11860, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Elahe Arani, Fahad Sarfraz, and Bahram Zonooz. Learning fast, learning slow: A general continual learning method based on complementary learning system. In International Conference on Learning Representations, 2022. 7, 12, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[3] Jihwan Bang, Hyunseo Koh, Seulki Park, Hwanjun Song, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on a contaminated data stream with blurry task boundaries. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9265-9274, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[4] E. Belouadah and A. Popescu. II2m: Class incremental learning with dual memory. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 583-592, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.429 + ], + "angle": 0, + "content": "[5] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. 34th Conference on Neural Information Processing Systems, 2020. 1, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.472, + 0.497 + ], + "angle": 0, + "content": "[6] Dan A. Calian, Florian Stimberg, Olivia Wiles, Sylveste-Alvise Rebuffi, Andras Gyorgy, Timothy Mann, and Sven Gowal. Defending against image corruptions through adversarial augmentations. https://arxiv.org/abs/2104.01086, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.472, + 0.557 + ], + "angle": 0, + "content": "[7] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio Calmon, and Taesup Moon. {CPR}: Classifier-projection regularization for continual learning. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. Proceedings of the International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.472, + 0.683 + ], + "angle": 0, + "content": "[9] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. https://arxiv.org/abs/1902.10486, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.472, + 0.743 + ], + "angle": 0, + "content": "[10] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation policies from data. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[11] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. Advances in Neural Information Processing Systems, 2020. 1, 3, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[12] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. https://arxiv.org/abs/1708.04552, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] Sayna Ebrahimi, Mohamed Elhoseiny, Trevor Darrell, and Marcus Rohrbach. Uncertainty-guided continual learning with bayesian neural networks. Proceedings of the International Conference on Learning Representations, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[14] Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. https://arxiv.org/abs/1701.08734, 2017.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[15] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. International Conference on Machine Learning, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[16] Chengyue Gong, Tongzheng Ren, Mao Ye, and Qiang Liu. Maxup: Lightweight adversarial training with data augmentation improves neural network training. Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.894, + 0.345 + ], + "angle": 0, + "content": "[17] Yanan Gu, Xu Yang, Kun Wei, and Cheng Deng. Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[18] Gunshi Gupta, Karmesh Yadav, and Liam Paull. La-maml: Look-ahead meta learning for continual learning. In Advances in Neural Information Processing Systems, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.894, + 0.446 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. International Conference on Computer Vision, 2021. 3, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.894, + 0.588 + ], + "angle": 0, + "content": "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[22] Dan Hendrycks, Norman Mu, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple method to improve robustness and uncertainty under data shift. 2020. 2, 3, 4, 5, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[23] Christian Henning, Maria Cervera, Francesco D'Angelo, Johannes Von Oswald, Regina Traber, Benjamin Ehret, Seijin Kobayashi, Benjamin F Grewe, and Joao Sacramento. Posterior meta-replay for continual learning. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.894, + 0.773 + ], + "angle": 0, + "content": "[24] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Intertask association in continual learning. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[25] Chris Dongjoo Kim, Jinseo Jeong, Sangwoo Moon, and Gunhee Kim. Continual learning on noisy data streams via self-purified replay. International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24529" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.12 + ], + "angle": 0, + "content": "networks. Proceedings of the national academy of sciences, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.472, + 0.178 + ], + "angle": 0, + "content": "[27] Yajing Kong, Liu Liu, Zhen Wang, and Dacheng Tao. Balancing stability and plasticity through advanced null space in continual learning. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.472, + 0.209 + ], + "angle": 0, + "content": "[28] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.211, + 0.472, + 0.266 + ], + "angle": 0, + "content": "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Conference on Empirical Methods in Natural Language Processing, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.472, + 0.323 + ], + "angle": 0, + "content": "[30] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. TRGP: Trust region gradient projection for continual learning. In International Conference on Learning Representations, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.327, + 0.472, + 0.369 + ], + "angle": 0, + "content": "[31] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in Neural Information Processing Systems, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.472, + 0.44 + ], + "angle": 0, + "content": "[32] Amrutha Machireddy, Ranganath Krishnan, Nilesh Ahuja, and Omesh Tickoo. Continual active adaptation to evolving distributional shifts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3443-3449, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.472, + 0.511 + ], + "angle": 0, + "content": "[33] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. Proceedings of the International Conference on Learning Representations, 2019. 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.514, + 0.472, + 0.569 + ], + "angle": 0, + "content": "[34] Cuong V. Nguyen, Yingzhen Li, Thang D. Bui, and Richard E. Turner. Variational continual learning. Proceedings of the International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.472, + 0.641 + ], + "angle": 0, + "content": "[35] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. International Conference on Learning Representations, 2019. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.472, + 0.728 + ], + "angle": 0, + "content": "[36] Tim G. J. Rudner, Freddie Bickford Smith, Qixuan Feng, Yee Whye Teh, and Yarin Gal. Continual learning via sequential function-space variational inference. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 18871-18887, 17-23 Jul 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.472, + 0.785 + ], + "angle": 0, + "content": "[37] Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. https://arxiv.org/abs/1606.04671, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[38] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[39] Nikunj Saunshi, Arushi Gupta, and Wei Hu. A representation learning perspective on the importance of train-validation splitting in meta-learning. In Proceedings of the 38th International Conference on Machine Learning, pages 9333–9343, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[40] Jonathan Schwarz, Jelena Luketina, Wojciech M. Czarnecki, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress and compress: A scalable framework for continual learning. In Proceedings of the International Conference on Machine Learning, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[41] Joan Serrá, Dídac Surís, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In Proceedings of the International Conference on Machine Learning, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.894, + 0.265 + ], + "angle": 0, + "content": "[42] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[43] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.298, + 0.892, + 0.353 + ], + "angle": 0, + "content": "[44] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. Advances in neural information processing systems, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.356, + 0.894, + 0.424 + ], + "angle": 0, + "content": "[45] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing HONG, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.894, + 0.497 + ], + "angle": 0, + "content": "[46] Zhenyi Wang, Tiehang Duan, Le Fang, Qiuling Suo, and Mingchen Gao. Meta learning on a sequence of imbalanced domains with difficulty awareness. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8947-8957, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.5, + 0.894, + 0.554 + ], + "angle": 0, + "content": "[47] Zhen Wang, Liu Liu, Yajing Kong, Jiaxian Guo, and Dacheng Tao. Online continual learning with contrastive vision transformer. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.894, + 0.627 + ], + "angle": 0, + "content": "[48] Zhenyi Wang, Li Shen, Tiehang Duan, Donglin Zhan, Le Fang, and Mingchen Gao. Learning to learn and remember super long multi-domain task sequence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7992, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.894, + 0.697 + ], + "angle": 0, + "content": "[49] Zhenyi Wang, Li Shen, Le Fang, Qiuling Suo, Tiehang Duan, and Mingchen Gao. Improving task-free continual learning by distributionally robust memory evolution. In International Conference on Machine Learning, pages 22985-22998, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.894, + 0.77 + ], + "angle": 0, + "content": "[50] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[51] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[52] Mengqi Xue, Haofei Zhang, Jie Song, and Mingli Song. Meta-attention for vit-backed continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 150–159, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24530" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[53] Qingsen Yan, Dong Gong, Yuhang Liu, Anton van den Hengel, and Javen Qinfeng Shi. Learning bayesian sparse networks with full experience replay for continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 109-118, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[54] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[55] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. IEEE International Conference on Computer Vision, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. https://arxiv.org/abs/1703.04200, 2017.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[57] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. International Conference on Learning Representations, 2018. 3, 12" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24531" + } + ] +] \ No newline at end of file diff --git a/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_origin.pdf b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e3d9c6f56aa79ab4f01eb51fd424cda909bec7dc --- /dev/null +++ b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/350ee3f5-281e-480f-bce3-b1c6caa8838a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0007e3d319457e47e9094ed7f1a2230b7adecf7d0c331a55a3f881fae3ff5fb2 +size 679633 diff --git a/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/full.md b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..afa5349917dfdd939916737b0ed5a8e2463b813b --- /dev/null +++ b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/full.md @@ -0,0 +1,345 @@ +# MetaMix: Towards Corruption-Robust Continual Learning with Temporally Self-Adaptive Data Transformation + +Zhenyi Wang $^{1}$ Li Shen $^{2,*}$ Donglin Zhan $^{3}$ Qiuling Suo $^{1}$ + +Yanjun Zhu $^{1}$ Tiehang Duan $^{4}$ Mingchen Gao $^{1}$ + +$^{1}$ State University of New York at Buffalo, USA; $^{2}$ JD Explore Academy, China + +3 Columbia University, USA; 4 Meta, USA + +{zhenyiwa, qiulings, yjzhu, mgao8}@buffalo.edu; dz2478@columbia.edu + +{mathshenli, tiehang.duan}@gmail.com + +# Abstract + +Continual Learning (CL) has achieved rapid progress in recent years. However, it is still largely unknown how to determine whether a CL model is trustworthy and how to foster its trustworthiness. This work focuses on evaluating and improving the robustness to corruptions of existing CL models. Our empirical evaluation results show that existing state-of-the-art (SOTA) CL models are particularly vulnerable to various data corruptions during testing. To make them trustworthy and robust to corruptions deployed in safety-critical scenarios, we propose a meta-learning framework of self-adaptive data augmentation to tackle the corruption robustness in CL. The proposed framework, MetaMix, learns to augment and mix data, automatically transforming the new task data or memory data. It directly optimizes the generalization performance against data corruptions during training. To evaluate the corruption robustness of our proposed approach, we construct several CL corruption datasets with different levels of severity. We perform comprehensive experiments on both task- and class-continual learning. Extensive experiments demonstrate the effectiveness of our proposed method compared to SOTA baselines. + +# 1. Introduction + +Humans constantly acquire new information throughout their lifespan and easily recognize information shifts such as structure and style variations in images. Continual learning (CL) aims at imitating human's ability to learn from non-stationary data distributions without forgetting the previously learned knowledge. The past few years have witnessed rapid progress in CL research [1, 30, 31, 36, 45]. Despite the success, existing CL systems overlook the robustness + +against unforeseen data shifts during testing. They assume that training and test images for each task follow the same distribution. However, as data distributions evolve and new scenarios occur, test images often encounter various corruptions such as snow, blur, pixelation, and combinations, resulting in a shifted distribution from the training set. For example, Figure 1 shows various corruptions applied on one image from Split-miniImageNet. + +Data corruption can drastically impair the performance of existing image recognition systems. A recent study [21] shows that classification accuracy of various architectures has dropped significantly on the ImageNet test set with some simple and natural corruptions. Our empirical evaluation results show that state-of-the-art CL models are even more vulnerable to these corruptions during testing. For example, the accuracy of $\mathrm{DER}++$ [5] for task-continual learning decreases from $93.9\%$ to $50.5\%$ on split-CIFAR10; accuracy drops to $10.6\%$ from $75.6\%$ on split-CIFAR100; accuracy decreases to $9.8\%$ from $61.3\%$ on split-miniImageNet by applying those common corruptions on test data of each CL task. This severe issue makes existing CL models highly unreliable in safety-critical applications. Thus, improving the robustness of CL models to foster their trustworthiness when deployed in real-world scenarios is essential. + +Training a CL model robust to various corruptions is difficult due to the following challenges. 1) Unseen corruptions could perturb the test set far beyond those encountered during training. A model that naively augments training images with seen corruptions cannot generalize to the new ones during testing. Also, it is unrealistic to enumerate all possible corruptions during training since there are infinite types of corruptions and their combinations. 2) With the ever-evolving data distributions in CL, an effective augmentation strategy learned on previous tasks may gradually become less effective because the optimal augmentation strategies are task-dependent and dynamically change over tasks [11]. + +![](images/7018865a0779f5e2d6097ceeab51aaa398a576f5adcca44e96478509a63043cb.jpg) +(a) Original data + +![](images/bd248e56f61825c5094a381447e681e063e85f8458e33af3352119463e6883e4.jpg) +(b) Brightness + +![](images/37e78b46c73ed97a17a6757dc2f0e68afbceb916b50f86b9c6442501111f50c9.jpg) +(c) Contrast +Figure 1. Visualization of five types of different corruption operations on the testing images of Split-miniImageNet. + +![](images/8a957671ca9433a669081fd58d90e4e56156bdfb908584d1921f399c859f7997.jpg) +(d) Defocus Blur + +![](images/fe0ece7366514591855722284a7840932d204682a7eadf66ea104892bafff090.jpg) +(e) Elastic + +![](images/6fb2639880539c52c8ee1db674f88b42d5a135973c64dafd693bb4e2062e627d.jpg) +(f) Fog + +Although we can adopt a memory buffer to store data from previous tasks, augmenting and replaying them at later training iterations, the augmented memory may gradually become less effective as the model could memorize the stored information after replay runs. Recent approaches, such as Augmix [22] composes and combines multiple pre-defined augmentation operations with different depths and widths, show efficacy in improving robustness under traditional supervised classification tasks. However, these approaches are not directly applicable to corruption-robust CL since they only use a fixed random augmentation strategy that is often not optimal for non-stationary data distributions in CL. + +To address these unique challenges of corruption-robustness in CL, we propose a temporally self-adaptive Augmix within a meta-learning framework, named MetaMix. It adaptively augments the memory buffer data and the currently received new data by learning to mix the augmentation operations tailored to the evolving data distributions. In particular, our automatic self-adaptive MetaMix is a bi-level optimization, simulating the evaluation process on unseen corruptions. We randomly divide the training augmentation operations into pseudo-seen and pseudo-unseen operations at each CL step. The lower-level optimization is to optimize the model performance on the pseudo-seen operations; the upper-level optimization is to optimize the generalization of the pseudo-unseen operations. The augmentation strategy is governed by an LSTM, which inputs context information and outputs the corresponding mixing parameters for the augmentations. The proposed MetaMix ensures the augmentation strategy automatically adapts to non-stationary data distribution. Furthermore, the objective is to optimize the performance of the pseudo-unseen corruption operations, which aligns with our goal during testing and encourages the generalization to unseen corruptions. + +To evaluate the corruption robustness of existing and the proposed methods, we propose a new challenging benchmark where various corruptions perturb the testing data of each CL task. To facilitate future research, we construct several new datasets, including split-CIFAR-10-C, split-CIFAR-100-C, and split-miniImageNet-C. Extensive experiments on the constructed benchmarks demonstrate the effectiveness of our proposed MetaMix approach compared with several SOTA data-augmentation approaches adapted for CL. We summarize our contributions as follows: + +- To our best knowledge, we are the first to study the corruption-robustness of CL methods. Accordingly, we propose the first set of novel benchmarks for evaluating the corruption-robustness of existing CL methods and moving towards trustworthy CL. +- We propose a self-adaptive augmentation method, MetaMix, by learning to mix and augment the training data of each CL task to achieve corruption-robustness on unseen corruptions for each CL task. +- Our method is versatile and can be seamlessly integrated with existing CL methods. Extensive experiments with both task/class continual learning demonstrate the effectiveness of MetaMix. + +# 2. Related Work + +# 2.1. Continual Learning + +CL focuses on learning non-stationary data distribution without forgetting previous knowledge. These methods can be categorized into: 1) retaining memory for future replay [1,4,8,9,17,31,35,42,46,48,49]; 2) designing tailored network architectures [14,24,37,41,54]; 3) performing proper regularization during parameter updates [7,26,40,56]; 4) introducing Bayesian methods for model parameter inference [13,23,34,53]; 5) subspace-based methods [27,30,38] and 6) prompt-tuning [29] based methods [50,51]; 7) transformer-based methods [47,52]. However, all these works focus on the simplified setting in which training and testing data of each CL task follow the same distribution. In practice, the testing data may come from another different distribution due to the variations in weather, light, snow, etc. These distortions pose significant robustness challenges for existing CL methods. Existing literature largely ignored these trustworthy properties of CL. According to the terminology of [43], CL settings can be categorized into three scenarios based on whether task identities of the testing data are available to the CL model during test time. Among the three scenarios, task-continual learning studies the scenario that the task identities are available during testing and is the easiest learning setting; domain incremental learning focuses on the case that all the tasks during CL solve the same classification task with different data distribution and the task identities are unknown during testing; class-continual learning focuses + +on the scenario that task identities are unavailable and is the most difficult learning setting. We focus on task/class CL scenarios. + +Recently, robust CL began to attract researchers' attention. Self-Purified Replay [25] and PuriDivER [3] consider CL on noisy data stream with label noise. Their work is orthogonal to ours since they consider noisy data during training; we instead consider the robustness to data corruptions during testing. A closely related work to our is Continual Active Adaptation (CAA) [32], which is different from and orthogonal to ours since they focus on the setting that for each data distribution, the label space for each learning task is the same. Each task only encounters a single data corruption. During testing, CAA focuses on generalization to previously seen corruption operations. In contrast, we focus on the setting that each task solves a different task. During testing, the CL model should be robust to unseen data corruptions. + +# 2.2. Data Augmentation + +We categorize existing data augmentation techniques into two classes: 1) improving performance on clean data; 2) improving performance and being robust to data distribution shift and corruptions. + +Augmentations for Improving Clean Data Performance. Standard data augmentation aims to improve the generalization performance on clean data. To improve the performance of neural network, standard supervised learning adopts common augmentations, such as random flips, rotations, and crops. Recently, Cutout [12], Mixup [57], CutMix [55] and MaxUp [16] show excellent performance on clean datasets. Other techniques include automatically learning augmentation policy from data, e.g., AutoAugment [10] and RandAugment [11]. However, AutoAugment is computationally impractical for CL problems since it needs a huge computation cost and separate proxy tasks to train the policy network. + +Augmentations for Being Robust to Corruptions. Augmix [22] is one of the SOTA methods on the data corruption benchmarks, such as ImageNet-C, CIFAR100-C, etc. DeepAugment [20] and AdversarialAugment [6] transform and augment clean images by perturbing the representations/features of neural networks to generate new augmented images. However, this increases a lot of computation costs. Thus, DeepAugment [20] and AdversarialAugment [6] are impractical for CL. We thus choose the more efficient and effective AugMix and propose a new self-adaptive AugMix for CL. + +# 3. Methodology + +In this section, we first define the corruption robustness of CL and then describe the proposed MetaMix approach. We describe the notation in the main text. We also provide a notation table in Table 4 in Appendix for ease of reading. + +# 3.1. Problem Setup + +Corruption robustness of CL. We consider the problem of learning a sequence of tasks denoted as $\mathcal{D}^{tr} = \{\mathcal{D}_1^{tr},\mathcal{D}_2^{tr},\dots ,\mathcal{D}_N^{tr}\}$ , where $N$ is the number of training tasks. The $k$ -th task training data $\mathcal{D}_k^{tr}$ consists of many labeled examples $(\pmb{x}^k,y^k,\mathcal{T}_k)$ , where $\pmb{x}^k$ is the data example in the task, $y^{k}$ is the corresponding data label, and $\mathcal{T}_k$ is the task identifier. The goal is to learn a model $f_{\theta}$ on the training task sequence $\mathcal{D}^{tr}$ so that it performs well on the testing set of all the learned tasks $\mathcal{D}^{te} = \{\mathcal{D}_1^{te},\mathcal{D}_2^{te},\dots ,\mathcal{D}_N^{te}\}$ without forgetting previously learned knowledge. In particular, we focus on improving the model robustness on the testing set for each task $\mathcal{T}_k$ under a collection of corruption operations $\mathcal{C}$ . Below, we define the CL corruption robustness. + +Definition 1 (CL corruption robustness). Given a sequence of $N$ tasks with test data $\mathcal{D}^{te} = \{\mathcal{D}_1^{te},\mathcal{D}_2^{te},\dots ,\mathcal{D}_N^{te}\}$ , and a collection of corruption operations $\mathcal{C}$ , the corruption robustness is the expectation accuracy after applying corruptions $\mathcal{C}$ on the test data of each task $k$ , i.e., + +$$ +A C C = \frac {1}{N} \sum_ {k = 1} ^ {k = N} \mathbb {E} _ {c \sim \mathcal {C}} \left[ \mathbb {E} _ {\left(\boldsymbol {x} ^ {k}, y ^ {k}, \mathcal {T} _ {k}\right) \sim D _ {k} ^ {t e}} \mathbb {I} \left(f \left(c \left(\boldsymbol {x} ^ {k}\right)\right) = y ^ {k}\right) \right] \tag {1} +$$ + +where $c(\pmb{x}^k)$ is the resulting testing data after applying the corruption operation $c$ on $\pmb{x}^k$ . $\mathbb{I}(.)$ is the indicator function. ACC measures the average robust accuracy after applying a collection of corruption operations $\mathcal{C}$ . + +Preliminary of AugMix. AugMix [22] is a recently proposed SOTA augmentation technique that significantly improves model robustness in standard supervised learning. Considering that CL requires the algorithm to be both efficient and effective, we thus choose the efficient and effective AugMix to adapt to CL. AugMix maintains augmented data diversity by mixing several augmentation chains. Suppose we have an augmentation operation set $\mathcal{O} = \{op_1, op_2, \dots, op_M\}$ during training. The standard AugMix implementation samples three operations $op_1$ , $op_2$ and $op_3$ from $\mathcal{O}$ , and composes them with varying depths to form several operation chains; where each chain consists of one to three randomly selected augmentation operations. The augmented image is a convex mix of $A$ chains ( $A$ denotes the number of augmentation chains), i.e., $\boldsymbol{x}_{aug} = w_1 \cdot \text{chain}_1(\boldsymbol{x}) + w_2 \cdot \text{chain}_2(\boldsymbol{x}) + \dots + w_A \cdot \text{chain}_A(\boldsymbol{x})$ , where mixing weights $\boldsymbol{w} = (w_1, w_2, \dots, w_A)$ is sampled from the distribution, Drichelet $(\alpha, \alpha, \dots, \alpha)$ with $\alpha = 1$ . With the mixed image $\boldsymbol{x}_{aug}$ , AugMix combines it with the original image through another random convex combination to obtain the final image $\boldsymbol{x}_{augmix}$ , i.e., $\boldsymbol{x}_{augmix} = m\boldsymbol{x} + (1 - m)\boldsymbol{x}_{aug}$ , where $m$ is sampled from a $\operatorname{Beta}(\alpha, \alpha)$ distribution. AugMix adopts a fixed random augmentation policy, i.e., Dirichlet and Beta distributions, during the entire training process. + +![](images/9bc3e58034b6a6c6710135d88b192c538cf7e7fdb13fec050b39bee734bfd5b4.jpg) +(a) Continual Learning Process + +![](images/b8b5da81454f5d2fcae324074954ddf84fe5a702baf00360719987d5635bd247.jpg) +(b) AugMix + +![](images/18f1487a47f0f0d0d0d9c1ba79eaa3d74579d1f72e97b6f124ba36706d274f47.jpg) +(c) MetaMix (Ours) +Figure 2. Comparisons between AugMix [22] and our proposed MetaMix. (a) Tasks sequentially arrive during CL. (b) At each CL step, AugMix adopts a fixed and random augmentation policy during the entire learning process, thus lacking adaptation to the non-stationary data in CL and may not be the optimal strategy for each task. (c) MetaMix automatically adapts to the non-stationary data distribution during CL by meta-learning the mixing parameters of augmentation. At time $t$ , we sample a batch of data $(\pmb{x}_r, y_r)$ from the memory buffer and concatenate them with the current received data $(\pmb{x}_t, y_t)$ as $(\pmb{x}_b, y_b)$ . We apply MetaMix to augment each data example. Specifically, we feed the context information $I_t$ into the LSTM cell with parameters $\phi_t$ , and obtain the mixing weights $w_t$ and $m_t$ . The CL model $f_{\theta_t}$ calculates the losses for the augmented data. We adopt bi-level optimization to improve model generalization on unseen corruptions by splitting the augmentation operations into a pseudo-seen set $S_t$ and pseudo-unseen set $\mathcal{U}_t$ . The loss $\mathcal{L}_{\phi_t}$ is calculated on data augmented with $\mathcal{U}_t$ to update $\phi_t$ , and loss $\mathcal{L}_{\theta_t}$ is calculated on data augmented with $S_t$ to update $\theta_t$ . We repeat this process for all tasks in the sequence. + +The augmentation procedure of AugMix is illustrated in Figure 2 (b). The optimization objective of AugMix is shown in the following: + +$$ +\boldsymbol {\theta} _ {*} = \arg \min _ {\boldsymbol {\theta}} \underset {o p \in \mathcal {O}} {\mathbb {E}} \mathcal {L} (\boldsymbol {x}, y, \boldsymbol {\theta}) + \lambda J S (\boldsymbol {x}, \boldsymbol {x} _ {a u g m i x 1}, \boldsymbol {x} _ {a u g m i x 2}), \tag {2} +$$ + +where $\pmb{x}$ is the original data and $y$ is the data label; $\mathcal{L}(\pmb{x},\pmb{y},\pmb{\theta})$ is the loss function on the labeled data (cross-entropy for classification); $\pmb{x}_{augmix1}$ and $\pmb{x}_{augmix2}$ are two augmenta + +tions of original data $\mathbf{x}$ by AugMix; $JS$ denotes the Jensen Shannon divergence for regularizing the output consistency between the original and augmented data. We will define it in the following section. + +# 3.2. MetaMix Approach + +Direct extension of AugMix [22] for CL is to augment data examples sampled from the memory buffer, and the received new data at each time point for memory-based method across the sequentially arrived tasks. In the following, we present our methods for memory-based CL methods, but our methods can be easily extended to other families of CL methods as well, also demonstrated in Section 4. However, since the memory buffer is usually small for CL models, using a fixed random augmentation strategy (like AugMix) may make the CL model gradually memorize the training data corruption operations, but fail to generalize to unseen corruptions during testing. Moreover, with the non-stationary nature of the data, the optimal augmentation strategy for image mixing is data-dependent, and the mixing weights should be specifically optimized for each task. However, since we cannot repeatedly tune the mixing weights for each task during CL learning process, an automatic self-adaptive process is necessary to achieve good performance. We thus introduce the details of our self-adaptive data augmentation approach, namely MetaMix, for improving robustness in CL. Figure 2 (c) illustrates the training process of MetaMix at time $t$ . Specifically, we adopt meta-learning to learn a set of data-dependent parameters $w_{t}$ and $m_{t}$ , which can dynamically control the mixing weights of augmentation operations and automatically adapt to the non-stationary data distribution. + +First, to memorize previous knowledge, we allocate a small memory buffer $\mathcal{M}$ to store data from previous tasks by reservoir sampling, similar to [35]. The data in the memory buffer will be replayed later with a new task together. Suppose we have $M$ augmentation operations $\mathcal{O} = \{op_1, op_2, \dots, op_M\}$ during training, CL model parameters at time $t$ are $\theta_t$ , and the function represented by the network with $\theta_t$ is $f_{\theta_t}(x)$ . + +To capture the sequential dependencies among previous tasks, we design a self-adaptive LSTM mixing strategy. The LSTM inputs the current context information and outputs the adaptive mixing parameters. We use an LSTM (dynamic) model to dynamically change the mixing weights instead of an MLP (static) model since the tasks in CL sequentially arrive. The LSTM can encode the information of previous tasks into an additional hidden state while the MLP can only capture the data information of the current task. It does not capture the information of previous tasks without the hidden state. Formally, the LSTM adaptive process is defined as: + +$$ +\boldsymbol {o} _ {t + 1}, \boldsymbol {h} _ {t + 1}, \boldsymbol {g} _ {t + 1} = \operatorname {L S T M} _ {\phi_ {t}} \left(\boldsymbol {I} _ {t}, \boldsymbol {h} _ {t}, \boldsymbol {g} _ {t}\right), \tag {3} +$$ + +where $\phi_t$ is the LSTM MetaMixer parameters at time $t$ , $h_t$ is the hidden state of LSTM at time $t$ , $g_t$ is the cell state of LSTM at time $t$ , $I_t$ is the context information encoding as input to LSTM at time $t$ ; $o_{t+1}$ is the output of LSTM at time $t+1$ for the mixing parameters, i.e., $o_{t+1} = \{w_t', m_t'\}$ , where $m_t'$ is the mixing weight for the augmented data and original raw data and $w_t'$ are the mixing weights for different chains in the augmentation. We apply softmax function to $w_t'$ and sigmoid function to $m_t'$ to ensure they are within [0,1] interval, i.e., + +$$ +m _ {t} = \operatorname {S i g m o i d} \left(m _ {t} ^ {\prime}\right), \boldsymbol {w} _ {t} = \operatorname {S o f t m a x} \left(\boldsymbol {w} _ {t} ^ {\prime}\right). \tag {4} +$$ + +The context $I_{t}$ contains essential information to characterize data in current and previous tasks. It is shown as: + +$$ +\boldsymbol {I} _ {t} = \left\{\nabla \mathcal {L} _ {\boldsymbol {\theta} _ {t}} \left(\boldsymbol {x} _ {r}, \boldsymbol {y} _ {r}\right) \cdot \nabla \mathcal {L} _ {\boldsymbol {\theta} _ {t}} \left(\boldsymbol {x} _ {t}, \boldsymbol {y} _ {t}\right), \mathcal {L} _ {\boldsymbol {\theta} _ {t}} \left(\boldsymbol {x} _ {r}, \boldsymbol {y} _ {r}\right), \boldsymbol {e} _ {r}, \boldsymbol {e} _ {t} \right\}, \tag {5} +$$ + +where $(\pmb{x}_r, y_r)$ is the labeled mini-batch data sampled from the memory buffer and $(\pmb{x}_t, y_t)$ is the labeled data received at time $t$ . We adopt this information for several reasons: (1) $\nabla \mathcal{L}_{\pmb{\theta}_t}(\pmb{x}_r, y_r) \cdot \nabla \mathcal{L}_{\pmb{\theta}_t}(\pmb{x}_t, y_t)$ is the gradient dot product between memory data gradient and current mini-batch data gradient. It determines the degree of interference and transfer between memory data and current task mini-batch data. It is commonly used in existing CL as context information [18, 35]; (2) we also input the memory mini-batch data loss, as it indicates the degree of forgetting of previous tasks; (3) $\pmb{e}_r$ and $\pmb{e}_t$ are data encoding for mini-batch data, $\pmb{x}_r$ , sampled from memory buffer and currently received new mini-batch data, $\pmb{x}_t$ , respectively. That is, $\pmb{e}_r$ and $\pmb{e}_t$ are the last layer features outputted by ResNet-18. This determines the data-dependent features for calculating the data-dependent mixing weights. + +Since corruption operations on testing data are unseen during training, we cannot directly optimize the augmentation strategy for these corruptions. To make the CL model generalize better to unseen corruptions, at each training step $t$ , we randomly split the augmentation operations $\mathcal{O}$ into non-overlapping pseudo-seen $S_{t}$ and pseudo-unseen $\mathcal{U}_t$ operations, i.e., $S_{t} \cup \mathcal{U}_{t} = \mathcal{O}$ and $S_{t} \cap \mathcal{U}_{t} = \emptyset$ . We use the pseudo operation sets to simulate the scenario in which training and testing corruptions have no overlap. We formulate MetaMix as a bi-level optimization problem to make the CL model learn to augment and mix. This ensures that the update of model parameters $\theta$ on the pseudo-seen set can improve the optimization of MetaMixer parameters $\phi$ on the pseudo-unseen set, which could, in turn, be beneficial to improve the CL model generalization. We formulate the MetaMix as the following bi-level optimization: + +$$ +\begin{array}{l} \phi_ {*} = \arg \min _ {\phi} \mathbb {E} _ {o p \in \mathcal {U} _ {t}} \mathcal {L} \left(\boldsymbol {x} _ {b}, y _ {b}, \boldsymbol {\theta} _ {*}, \phi\right) + \lambda J S \left(\boldsymbol {x} _ {b}, \widehat {\boldsymbol {x}} _ {b 1}, \widehat {\boldsymbol {x}} _ {b 2}\right), \tag {6} \\ \boldsymbol{\theta}_{*} = \arg \min_{\boldsymbol{\theta}}\underset {op\in \mathcal{S}_{t}}{\mathbb{E}}\mathcal{L}(\boldsymbol{x}_{b},y_{b},\boldsymbol {\theta},\boldsymbol {\phi}) + \lambda J S(\boldsymbol{x}_{b},\boldsymbol{x}^{\prime}_{b1},\boldsymbol{x}^{\prime}_{b2}), \\ \end{array} +$$ + +where we concatenate the mini-batch data $(\pmb{x}_r, y_r)$ from the memory buffer and the current received mini-batch data $(\pmb{x}_t, y_t)$ together as $(\pmb{x}_b, y_b)$ ; $\widehat{\pmb{x}}_{b1}$ and $\widehat{\pmb{x}}_{b2}$ are the two minibatch data augmented by applying the pseudo-unseen augmentation operations $\mathcal{U}_t$ on $(\pmb{x}_b, y_b)$ ; $\pmb{x}_{b1}'$ and $\pmb{x}_{b2}'$ are the two mini-batch data augmented by applying the pseudo-seen augmentation operations $\mathcal{S}_t$ on $(\pmb{x}_b, y_b)$ . $\lambda$ is the regularization constant for the consistency loss. The lower-level optimization is to optimize the performance on the pseudoseen augmentation operations $\mathcal{S}_t$ at the training step $t$ ; and the upper-level optimization is to optimize the generalization on the pseudo-unseen augmentation operations $\mathcal{U}_t$ . This bi-level optimization for learning to augment and mix can be efficiently solved by first-order gradient descent method, similar to the first-order MAML [15]. + +The advantage of MetaMix (Eq. (6)) over AugMix (Eq. (2)) are two folds. First, MetaMix improves the generalization to unseen corruptions since we split the training data augmentation operations into pseudo-seen and unseen operations. This is similar to the train-validation data split in standard meta-learning, which improves the generalization compared to the training without data split [39]; It could be understood as introducing an implicit regularization so that it can benefit from low sample complexity [39]. Second, MetaMix can automatically adapt to the non-stationary data distribution in CL, while AugMix can only work well in the stationary data distribution in a single task. + +Consistency Regularization To ensure the network output consistency between the original raw image data and the augmented data, we use a similar consistency loss [22] to regularize the network output to ensure smoother neural network responses. We hope the model to have similar responses to $\pmb{x}_b, \widehat{\pmb{x}}_{b1}, \widehat{\pmb{x}}_{b2}$ . We minimize the Jensen-Shannon divergence among the posterior distributions of the original sample $\pmb{x}_b$ and its augmented variants $\widehat{\pmb{x}}_{b1}$ and $\widehat{\pmb{x}}_{b2}$ . The consistency loss is defined as below: + +$$ +\begin{array}{l} p _ {m e a n} = \left(p _ {\boldsymbol {x} _ {b}} + p _ {\hat {\boldsymbol {x}} _ {b 1}} + p _ {\hat {\boldsymbol {x}} _ {b 2}}\right) / 3 (7) \\ J S \left(\boldsymbol {x} _ {b}, \widehat {\boldsymbol {x}} _ {b 1}, \widehat {\boldsymbol {x}} _ {b 2}\right) = \left(\mathbb {K L} \left(p _ {\boldsymbol {x} _ {b}} \mid p _ {\text {m e a n}}\right) + \mathbb {K L} \left(p _ {\widehat {\boldsymbol {x}} _ {b 1}} \mid p _ {\text {m e a n}}\right) \right. \\ + \mathbb {K L} \left(p _ {\widehat {x} _ {b 2}} \mid p _ {\text {m e a n}}\right)) / 3 (8) \\ \end{array} +$$ + +Where $\mathbb{KL}$ denotes the KL divergence between two distributions. $p_{\boldsymbol{x}_b} = f_{\boldsymbol{\theta}_t}(\boldsymbol{x}_b)$ is the network output probabilities of each class for original raw data $\boldsymbol{x}_b$ . Similarly, we can define $p_{\widehat{\boldsymbol{x}}_{b1}} = f_{\boldsymbol{\theta}_t}(\widehat{\boldsymbol{x}}_{b1})$ and $p_{\widehat{\boldsymbol{x}}_{b2}} = f_{\boldsymbol{\theta}_t}(\widehat{\boldsymbol{x}}_{b2})$ . We summarize our proposed algorithm in Algorithm 3. Line 5-8 takes the context information as input and computes desired augmentation mixing parameters. Line 9-11 augments the original data with the pseudo-seen augmentation operations $\mathcal{S}_t$ and pseudo-unseen operations $\mathcal{U}_t$ by Algorithm 1. Line 12 solves the bi-level optimization problem (Eq. (6)) by Algorithm 2. + +Algorithm 1 DataMix +1: Function DataMix $(\pmb{x}_b, \pmb{w}_t, m_t, \mathcal{V})$ +2: MetaMix augmented images initialized as $\pmb{x}_b' = 0$ +3: $\pmb{x}_{orig} = \pmb{x}_b$ +4: for $i = 1$ to $A$ do { A is the number of augmentation chains} +5: $\pmb{x}_b = \pmb{x}_{orig}$ +6: for $l = 1$ to $d$ do { $d$ is the augmentation depth} +7: randomly sample operation $op \in \mathcal{V}$ ( $\mathcal{V}$ is the collection of augmentation operations) +8: $\pmb{x}_b = op(\pmb{x}_b)$ +9: end for +10: $\pmb{x}_b' + = \pmb{w}_t^i \cdot \pmb{x}_b$ +11: end for +12: $\pmb{x}_b' = m_t \pmb{x}_{orig} + (1 - m_t) \pmb{x}_b'$ ; +13: similar to the above procedure, we can get two different minibatch $\pmb{x}_{b1}', \pmb{x}_{b2}'$ +14: return $\pmb{x}_{b1}', \pmb{x}_{b2}'$ +15: EndFunction + +Algorithm 2 Bi-level Solver +1: Function Bi-level Solver $(\pmb{\theta}_t, \pmb{\phi}_t, \pmb{x}_b, y_b, \pmb{x}_{b1}', \pmb{x}_{b2}', \widehat{\pmb{x}}_{b1}, \widehat{\pmb{x}}_{b2})$ +2: optimize the lower-level optimization problem in Eq. (6) by $\pmb{\theta}_{t+1} = \pmb{\theta}_t - \gamma \nabla_\theta \mathbb{E}_{op \in S_t}[\mathcal{L}(\pmb{x}_b, y_b, \pmb{\theta}_t, \phi) + \lambda JS(\pmb{x}_b, \pmb{x}_{b1}', \pmb{x}_{b2}')]$ +3: calculate the upper-level optimization loss function in Eq. (6) by $\mathcal{E}(\pmb{\theta}_{t+1}, \phi) = \mathbb{E}_{op \in U_t}[\mathcal{L}(\pmb{x}_b, y_b, \pmb{\theta}_{t+1}, \phi) + \lambda JS(\pmb{x}_b, \widehat{\pmb{x}}_{b1}, \widehat{\pmb{x}}_{b2})]$ +4: $\phi_1' = \phi_t$ +5: for $s = 1$ to $J$ do +6: $\phi_{s+1}' = \phi_s' - \beta \nabla_{\phi'} \mathcal{E}(\pmb{\theta}_{t+1}, \phi_s')$ +7: end for +8: $\phi_{t+1} = \phi_{J+1}'$ +9: return $\pmb{\theta}_{t+1}, \phi_{t+1}$ +10: EndFunction + +# 4. Experiment + +We evaluate the model performance for both task- and class-continual learning. The CL models are evaluated on the proposed benchmark for each dataset. We then describe experiment setup, results, and ablation study. + +# 4.1. Experiment Setup + +Below, we first construct the corruption CL benchmark and then describe the evaluation metrics, baselines, and implementation details. + +Corruption CL benchmark. Existing CL benchmarks, such as Split-CIFAR10, Split-CIFAR100, and Split-miniImageNet, are commonly used in literature. However, each task's training and testing data follow the same distribution for these datasets. Thus, these benchmarks are not suitable for evaluating the corruption robustness of CL methods. To properly evaluate the corruption robustness, we follow the protocols [21] and create several new CL datasets. + +Algorithm 3 MetaMix +1: REQUIRE: Augmentation Operations $\mathcal{O}$ $=$ {sterize, rotate, ..., solarize}; augmentation width $A$ ; augmentation depth $d$ ; the number of CL tasks N; the number of training iterations for each task $\mathcal{T}_k$ is $N_{k}$ ; memory buffer $\mathcal{M}$ ; CL model parameters $\theta$ ; LSTM MetaMix with parameters $\phi$ ; MetaMixer LSTM update steps $J$ ; CL model parameter learning rate $\gamma$ ; regularization weight $\lambda$ ; original image $\pmb{x}_{orig}$ ; initial hidden state and cell state $\pmb{h}_t,\pmb{g}_t$ are initialized as random noise $\pmb{h}_1,\pmb{g}_1\sim \mathcal{N}(\pmb {0},\pmb {I})$ +2: for $k = 1$ to $N$ do +3: for $t = 1$ to $N_{k}$ do +4: randomly sample mini-batch data $(x_{r},y_{r})$ from memory buffer $\mathcal{M}$ and concatenate it with the current received mini-batch data $(x_{t},y_{t})$ to obtain $(x_{b},y_{b})$ +5: compute context vector $\pmb{I}_t = \{\nabla \mathcal{L}_{\theta_t}(\pmb{x}_r,\pmb{y}_r)$ $\nabla \mathcal{L}_{\theta_t}(\pmb{x}_t,\pmb{y}_t),\mathcal{L}_{\theta_t}(\pmb{x}_r,\pmb{y}_r),\pmb{e}_r,\pmb{e}_t\}$ +6: $o_{t + 1},h_{t + 1},g_{t + 1} = \mathrm{LSTM}_{\phi_t}(I_t,h_t,g_t)$ +7: generate mixing parameters: $\pmb{w}_t^{\prime},m_t^{\prime} = o_{t + 1}$ +8: $m_t = \mathrm{Sigmoid}(m_t^{\prime}),w_t = \mathrm{Softmax}(w_t^{\prime})$ +9: randomly split augmentation operations $\mathcal{O}$ into nonoverlapping pseudo-seen $S_{t}$ and unseen $\mathcal{U}_t$ +10: $\pmb{x}_{b1}^{\prime},\pmb{x}_{b2}^{\prime} = \mathrm{DataMix}(\pmb{x}_b,\pmb{w}_t,m_t,\mathcal{S}_t)$ (Algorithm 1) +11: $\widehat{\pmb{x}}_{b1},\widehat{\pmb{x}}_{b2} = \mathrm{DataMix}(\pmb{x}_b,\pmb{w}_t,m_t,\mathcal{U}_t)$ (Algorithm 1) +12: $\theta_{t + 1},\phi_{t + 1} = \mathrm{Bi - level~Solver~}(\theta_t,\phi_t,\pmb{x}_b,\pmb{y}_b,\pmb{x}_{b1}'',\pmb{x}_{b2}'',\widehat{\pmb{x}}_{b1},\widehat{\pmb{x}}_{b2})$ (Algorithm 2) +13: update memory buffer $\mathcal{M}$ through reservoir sampling [35] to determine whether to store $(x_{t},y_{t})$ +14: end for +15: end for + +We apply 15 common corruption operations to the testing data of the three CL benchmarks, such as snow, frost, fog, etc. We provide a brief description for each corruption operation in Table 5 in Appendix. We perform each corruption at five different severity levels. Note that these corruptions only appear on the testing dataset of each continually learned task and are not seen in the training data of each task. This is to ensure that we can measure the robustness to unseen corruptions during testing for each compared method. To our best knowledge, we build the first CL corruption benchmark for measuring the generalization of CL models to unseen corruptions during testing. We name these new datasets Split-CIFAR10-C, Split-CIFAR100-C, and Split-miniImageNet-C. Specifically, for Split-CIFAR-10-C, we split the CIFAR-10 dataset [28] into 5 disjoint tasks, where each task has 2 classes; for Split-CIFAR-100-C, we split CIFAR-100 [28] consisting of 100 classes into 10 disjoint tasks, where each task has 10 classes; for Split-miniImageNet-C, we split miniImageNet [44] that consists of 100 classes, into 10 disjoint tasks, where each task has 10 classes. + +Evaluation Metrics. We evaluate the performance of different methods with average accuracy and backward transfer at the end of CL training to measure the final performance + +and the extent of forgetting for different methods. For each task, the accuracy for a specific corruption $c$ is the average accuracy over the five severity level. Suppose for task $k$ , $ACC_{c,s}^{k}$ measures the robust accuracy after applying the corruption operation $c$ with severity level $s$ : + +$$ +A C C _ {c, s} ^ {k} = \mathbb {E} _ {\left(\boldsymbol {x} ^ {k}, y ^ {k}, \mathcal {T} _ {k}\right) \sim D _ {t e} ^ {t e}} \mathbb {I} \left(f \left(c _ {s} \left(\boldsymbol {x} ^ {k}\right)\right) = y ^ {k}\right), +$$ + +where $c_{s}(\pmb{x}^{k})$ is the corruption operation $c$ with severity level $s$ applied on $\pmb{x}^{k}$ . The accuracy for corruption $c$ denoted as $ACC_{c}^{k}$ is the average robust accuracy across all the five severity levels. The overall accuracy for task $k$ , $ACC^{k}$ , is the average accuracy across all corruption types, i.e., + +$$ +A C C _ {c} ^ {k} = \frac {1}{5} \sum_ {s = 1} ^ {s = 5} A C C _ {c, s} ^ {k}, \text {a n d} A C C ^ {k} = \frac {1}{| \mathcal {C} |} \sum_ {c \in \mathcal {C}} A C C _ {c} ^ {k}. +$$ + +To measure catastrophic forgetting, we also evaluate the backward transfer (BWT), which measures the extent of forgetting on previous tasks after learning new ones. $BWT < 0$ reveals the occurrence of catastrophic forgetting on previous tasks, and $BWT > 0$ indicates that learning new tasks is helpful for previous tasks. Formally, the backward transfer (BWT) is defined as: $BWT = \frac{1}{N - 1}\sum_{k = 1}^{k = N - 1}(R_{N,k} - R_{k,k})$ ; where $R_{N,k}$ is the testing accuracy on task $k$ after learning on task $N$ and $R_{N,k}$ is defined the same as the above mentioned $ACC^k$ . For BWT, the higher, the better. + +Continual Learning Setting. We apply our proposed MetaMix to task-continual learning (Task-CL) and class-continual learning (Class-CL). The former provides task identities for the CL learner to select the relevant classifier for each example during testing, whereas the latter does not. + +Baseline. Our method can be seamlessly integrated with exiting CL methods. For illustration, we integrate the proposed methods with SOTA memory-based methods: $DER++$ [5] and CLS-ER [2]. To evaluate the effectiveness of the proposed MetaMix, we compare it with several SOTA data augmentation approaches originally designed for traditional supervised learning: Adversarial Training (AT) [33], RandAugment (RA) [11], Maxup [16], DeepAugment [20], Augmix [22]. We adapt these augmentation methods for CL problems by applying them on the new task data and the memory buffer data. We provide more detailed descriptions of baselines in Appendix. Due to space limitations, we put the experiments that integrate our proposed method with CLS-ER in Appendix. + +Implementation details. We use ResNet18 [19] as classifier for all datasets. All the other hyperparameter settings follow from [5]. The types of corruption during testing are not seen during training for all the compared augmentation methods to evaluate the robustness against unseen corruptions. Following [22], the augmentation operations performed during training are autocontrast, equalize, posterize, rotate, solarize, shear-x, shear-y, translate-x, translate-y. + +All the above-compared data augmentation methods apply these augmentation operations to CL baseline method to improve robustness. The memory buffer has capacity of 500 data points by default. We average the result for 5 runs for each experiment. We provide both mean and standard deviation across different runs. Due to space limitations, we provide standard deviation results in Appendix. We provide more implementation details in Appendix. + +LSTM Architecture. It has one recurrent layer with 5 hidden units. A linear layer (fully connected layer) is appended next to the LSTM output to generate the mixing parameters $\boldsymbol{w}_t$ and $m_t$ . The LSTM only has about 20K parameters and is negligible compared to ResNet18, which has more than 11.22 million parameters. + +# 4.2. Robustness of Task-CL + +Task-Continual Learning In this section, we evaluate the corruption robustness under task-continual learning, where the task identifiers are provided to the CL learner during testing. We present the results that integrate baselines and the proposed method with DER++ [5]. Due to space limitations, we put the experiments that integrate baselines and MetaMix with CLS-ER [2] in Appendix. + +Results. We show the results across different corruptions and compared methods in Tables 1-2; where $Avg$ is the average robust accuracy across all the corruptions. Due to the space limitations, we provide the standard deviation across different runs in Table 8 and 9 in Appendix. We also put results on split-CIFAR10-C in Table 6 and Table 7 in Appendix. We can observe that without data augmentation, the performance of $\mathrm{DER}++$ is close to random guessing, indicating that the CL model does not generalize to common corruptions. The proposed self-adaptive MetaMix brings significant improvement toward corruption robustness. It outperforms the average corruption robustness accuracy of the other models by a large margin of $2.4\%$ , $2.6\%$ on Split-CIFAR100-C, and Split-miniImageNet-C, respectively. The improvement shows the effectiveness of proposed MetaMix, which can automatically adapt to the non-stationary data distribution. The compared methods only use fixed static/random augmentation methods which lack adaptation to the non-stationary data in CL, and thus do not perform well. Random augmentation achieves second-best results. Deep-augment would distort the image such that they are very different from original images. Adversarial training and Maxup do not help much corruption robustness among the compared methods. Adversarial training can generate similar images compared to original images. Thus, there is a large gap between training and testing corrupted data for those methods adapted for CL. Maxup aims for improving clean data performance, not designed for robust accuracy. + +Table 1. Robust accuracy of Task-CL on Split-CIFAR100-C with DER++ + +
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++10.4610.6910.2610.7610.5110.7110.6610.4710.5610.5110.7910.5410.7510.5510.7210.60
AT10.3610.4510.2510.2310.2410.2710.2510.129.889.9810.169.8310.1910.310.3110.19
RA26.7727.1525.4627.8423.4624.0525.9923.2626.3430.4527.9528.825.5429.1229.0926.75
DeepAugment11.0011.1610.8611.3511.1511.4511.3310.7610.4610.5911.1010.5711.2011.3611.2511.04
Maxup11.111.1711.1311.0911.0611.1311.110.8810.7610.8110.8610.3711.0611.1311.1210.98
Augmix59.5862.6859.9267.9558.8865.9367.0363.3963.9363.2468.6263.2765.8266.0664.0764.02
Ours61.5365.1461.1870.3561.2868.0168.7666.0266.4966.1471.4966.1668.2468.9166.7566.43
+ +Table 2. Robust accuracy of Task-CL on Split-MiniImageNet-C with DER++ + +
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++10.2710.110.079.789.739.679.729.619.719.859.6910.159.559.459.579.79
AT9.079.249.279.659.629.69.489.179.109.019.128.389.549.569.359.28
RA17.2318.2216.719.8116.618.3219.0718.9417.1418.2617.2614.5621.1420.8423.018.47
DeepAugment10.5410.5310.6210.6810.7010.6310.7410.5610.3210.3210.529.9210.5510.4110.4910.50
Maxup10.410.910.6810.659.7510.2110.259.6210.1210.2110.3410.910.210.1610.0710.30
Augmix43.4548.9142.2053.7849.6859.8851.8353.2254.0255.6359.5943.4661.3551.3462.7552.74
Ours48.853.3343.8159.753.6462.7856.9556.1957.3156.0862.0741.7163.1749.7864.5555.32
+ +Table 3. Robust accuracy of Class-CL on Split-CIFAR100-C with DER++ + +
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++0.750.750.880.750.860.770.680.880.960.920.750.960.780.780.790.82
AT0.990.971.040.910.950.970.920.831.050.940.921.00.910.950.930.95
RA6.226.35.676.254.424.545.595.256.247.726.426.935.166.817.116.04
DeepAugment1.291.291.251.261.261.231.241.261.241.241.141.261.221.281.331.25
Maxup1.171.151.161.151.081.141.121.081.081.091.141.061.131.151.111.12
Augmix19.5721.2120.3526.1819.0825.1725.4622.0922.4322.8226.4422.5423.7224.6423.0122.98
Ours22.3224.4423.1529.1922.1328.1028.3525.4725.226.0829.5625.7827.0128.1625.8226.05
+ +# 4.3. Robustness of Class-CL + +Class-Continual Learning. In this section, we evaluate the corruption robustness under the more challenging class-continual learning setting, where the task identifiers are unavailable to the CL learner during testing. + +Results. We present the results across different corruptions and compared methods on Split-CIFAR100-C in Table 3, we present the results on Split-miniImageNet-C in Table 10 in Appendix. Similar to the Task-CL, MetaMix significantly improves over the baseline augmentation methods in Class-CL setting due to its self-adaptivity property. It outperforms the average corruption robustness accuracy of the other models by a large margin of $3.1\%$ , $2.2\%$ on Split-CIFAR100-C, and Split-miniImageNet-C, respectively. This is because our proposed method can automatically adapt to the non-stationary distributions. Most compared methods do not perform well on this challenging setup. + +Due to space limitations, we put BWT results in Table 15 in Appendix. In the corruption-robustness scenario, BWT is no longer a meaningful metric with such extreme accuracy differences, as the significantly lower accuracy of comparison methods results in much less space for further performance variations during backward transfer. + +# 4.4. Ablation Study and Hyperparameter Analysis + +Due to the limited space, we provide (1) ablation studies; (2) hyperparameter analyses, including $\lambda$ , $J$ , $\beta$ , etc.; (3) + +The effect of using LSTM vs MLP and the benefit of using additional hidden state information from previous tasks; (4) computation cost in Appendix. + +Effect of Memory Size. We evaluate the effect of memory size with 500 and 3000, respectively. The memory size of 500 is the default setting in the above tables. We provide experiment results on Split-CIFAR100-C, Split-miniImageNet-C with memory size 3000 on task-CL and class-CL respectively in Table 11, 12, 13, 14 in Appendix. In these cases, our method significantly outperforms baselines. + +# 5. Conclusion + +This paper tackles a more challenging problem of corruption robustness in CL with non-stationary data distribution, where the testing data distribution may significantly differ from training data distribution with various forms of common corruptions. We propose a meta-learning framework, MetaMix, for self-adaptive data augmentation specialized for CL. Our proposed MetaMix can substantially improve the model's robustness. Comprehensive experiments on both task- and class-CL settings demonstrate the effectiveness of the proposed method. + +Acknowledgement We thank all the anonymous reviewers for their insightful and thoughtful comments. This research was supported in part by grant NSF III-1910492. + +# References + +[1] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. Advances in Neural Information Processing Systems 32, pages 11849-11860, 2019. 1, 2 +[2] Elahe Arani, Fahad Sarfraz, and Bahram Zonooz. Learning fast, learning slow: A general continual learning method based on complementary learning system. In International Conference on Learning Representations, 2022. 7, 12, 17 +[3] Jihwan Bang, Hyunseo Koh, Seulki Park, Hwanjun Song, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on a contaminated data stream with blurry task boundaries. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9265-9274, 2022. 3 +[4] E. Belouadah and A. Popescu. II2m: Class incremental learning with dual memory. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 583-592, 2019. 2 +[5] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. 34th Conference on Neural Information Processing Systems, 2020. 1, 7, 12 +[6] Dan A. Calian, Florian Stimberg, Olivia Wiles, Sylveste-Alvise Rebuffi, Andras Gyorgy, Timothy Mann, and Sven Gowal. Defending against image corruptions through adversarial augmentations. https://arxiv.org/abs/2104.01086, 2021.3 +[7] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio Calmon, and Taesup Moon. {CPR}: Classifier-projection regularization for continual learning. In International Conference on Learning Representations, 2021. 2 +[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. Proceedings of the International Conference on Learning Representations, 2019. 2 +[9] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. https://arxiv.org/abs/1902.10486, 2019. +[10] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation policies from data. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[11] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. Advances in Neural Information Processing Systems, 2020. 1, 3, 7, 12 +[12] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. https://arxiv.org/abs/1708.04552, 2017.3 +[13] Sayna Ebrahimi, Mohamed Elhoseiny, Trevor Darrell, and Marcus Rohrbach. Uncertainty-guided continual learning with bayesian neural networks. Proceedings of the International Conference on Learning Representations, 2020. 2 + +[14] Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. https://arxiv.org/abs/1701.08734, 2017.2 +[15] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. International Conference on Machine Learning, 2017. 5 +[16] Chengyue Gong, Tongzheng Ren, Mao Ye, and Qiang Liu. Maxup: Lightweight adversarial training with data augmentation improves neural network training. Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 7, 12 +[17] Yanan Gu, Xu Yang, Kun Wei, and Cheng Deng. Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[18] Gunshi Gupta, Karmesh Yadav, and Liam Paull. La-maml: Look-ahead meta learning for continual learning. In Advances in Neural Information Processing Systems, 2020. 5 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2016. 7 +[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. International Conference on Computer Vision, 2021. 3, 7, 12 +[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 1, 6 +[22] Dan Hendrycks, Norman Mu, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple method to improve robustness and uncertainty under data shift. 2020. 2, 3, 4, 5, 7, 12 +[23] Christian Henning, Maria Cervera, Francesco D'Angelo, Johannes Von Oswald, Regina Traber, Benjamin Ehret, Seijin Kobayashi, Benjamin F Grewe, and Joao Sacramento. Posterior meta-replay for continual learning. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2 +[24] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Intertask association in continual learning. In European Conference on Computer Vision, 2022. 2 +[25] Chris Dongjoo Kim, Jinseo Jeong, Sangwoo Moon, and Gunhee Kim. Continual learning on noisy data streams via self-purified replay. International Conference on Computer Vision, 2021. 3 +[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural + +networks. Proceedings of the national academy of sciences, 2017. 2 +[27] Yajing Kong, Liu Liu, Zhen Wang, and Dacheng Tao. Balancing stability and plasticity through advanced null space in continual learning. In European Conference on Computer Vision, 2022. 2 +[28] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6 +[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Conference on Empirical Methods in Natural Language Processing, 2021. 2 +[30] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. TRGP: Trust region gradient projection for continual learning. In International Conference on Learning Representations, 2022. 1, 2 +[31] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in Neural Information Processing Systems, 2017. 1, 2 +[32] Amrutha Machireddy, Ranganath Krishnan, Nilesh Ahuja, and Omesh Tickoo. Continual active adaptation to evolving distributional shifts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3443-3449, 2022. 3 +[33] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. Proceedings of the International Conference on Learning Representations, 2019. 7, 12 +[34] Cuong V. Nguyen, Yingzhen Li, Thang D. Bui, and Richard E. Turner. Variational continual learning. Proceedings of the International Conference on Learning Representations, 2018. 2 +[35] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. International Conference on Learning Representations, 2019. 2, 4, 5, 6 +[36] Tim G. J. Rudner, Freddie Bickford Smith, Qixuan Feng, Yee Whye Teh, and Yarin Gal. Continual learning via sequential function-space variational inference. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 18871-18887, 17-23 Jul 2022. 1 +[37] Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. https://arxiv.org/abs/1606.04671, 2016. 2 +[38] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. In International Conference on Learning Representations, 2021. 2 +[39] Nikunj Saunshi, Arushi Gupta, and Wei Hu. A representation learning perspective on the importance of train-validation splitting in meta-learning. In Proceedings of the 38th International Conference on Machine Learning, pages 9333–9343, 2021. 5 + +[40] Jonathan Schwarz, Jelena Luketina, Wojciech M. Czarnecki, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress and compress: A scalable framework for continual learning. In Proceedings of the International Conference on Machine Learning, 2018. 2 +[41] Joan Serrá, Dídac Surís, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In Proceedings of the International Conference on Machine Learning, 2018. 2 +[42] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2 +[43] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning. 2019. 2 +[44] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. Advances in neural information processing systems, 2017. 6 +[45] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing HONG, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In International Conference on Learning Representations, 2022. 1 +[46] Zhenyi Wang, Tiehang Duan, Le Fang, Qiuling Suo, and Mingchen Gao. Meta learning on a sequence of imbalanced domains with difficulty awareness. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8947-8957, 2021. 2 +[47] Zhen Wang, Liu Liu, Yajing Kong, Jiaxian Guo, and Dacheng Tao. Online continual learning with contrastive vision transformer. In European Conference on Computer Vision, 2022. 2 +[48] Zhenyi Wang, Li Shen, Tiehang Duan, Donglin Zhan, Le Fang, and Mingchen Gao. Learning to learn and remember super long multi-domain task sequence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7992, 2022. 2 +[49] Zhenyi Wang, Li Shen, Le Fang, Qiuling Suo, Tiehang Duan, and Mingchen Gao. Improving task-free continual learning by distributionally robust memory evolution. In International Conference on Machine Learning, pages 22985-22998, 2022. 2 +[50] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. European Conference on Computer Vision, 2022. 2 +[51] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2 +[52] Mengqi Xue, Haofei Zhang, Jie Song, and Mingli Song. Meta-attention for vit-backed continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 150–159, 2022. 2 + +[53] Qingsen Yan, Dong Gong, Yuhang Liu, Anton van den Hengel, and Javen Qinfeng Shi. Learning bayesian sparse networks with full experience replay for continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 109-118, 2022. 2 +[54] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. International Conference on Learning Representations, 2018. 2 +[55] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. IEEE International Conference on Computer Vision, 2019. 3 +[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. https://arxiv.org/abs/1703.04200, 2017.2 +[57] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. International Conference on Learning Representations, 2018. 3, 12 \ No newline at end of file diff --git a/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/images.zip b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..00527c07b6bdadd47e9560d1d99c84c7d6d77827 --- /dev/null +++ b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e5398b8d6a27c619d2e9eee616bbb7a552393e38ace37704b3559bfd82a5507 +size 349868 diff --git a/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/layout.json b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4fbcf4c5824e6463dab97d06bd3a52dcc0c1365b --- /dev/null +++ b/2023/MetaMix_ Towards Corruption-Robust Continual Learning With Temporally Self-Adaptive Data Transformation/layout.json @@ -0,0 +1,11618 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 102, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 102, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 97, + 102, + 500, + 140 + ], + "type": "text", + "content": "MetaMix: Towards Corruption-Robust Continual Learning with Temporally Self-Adaptive Data Transformation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "spans": [ + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "text", + "content": "Zhenyi Wang" + }, + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "text", + "content": " Li Shen" + }, + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "text", + "content": " Donglin Zhan" + }, + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "text", + "content": " Qiuling Suo" + }, + { + "bbox": [ + 144, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 175, + 421, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 175, + 421, + 189 + ], + "spans": [ + { + "bbox": [ + 171, + 175, + 421, + 189 + ], + "type": "text", + "content": "Yanjun Zhu " + }, + { + "bbox": [ + 171, + 175, + 421, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 175, + 421, + 189 + ], + "type": "text", + "content": " Tiehang Duan " + }, + { + "bbox": [ + 171, + 175, + 421, + 189 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 171, + 175, + 421, + 189 + ], + "type": "text", + "content": " Mingchen Gao " + }, + { + "bbox": [ + 171, + 175, + 421, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 190, + 485, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 190, + 485, + 203 + ], + "spans": [ + { + "bbox": [ + 109, + 190, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 190, + 485, + 203 + ], + "type": "text", + "content": "State University of New York at Buffalo, USA; " + }, + { + "bbox": [ + 109, + 190, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 190, + 485, + 203 + ], + "type": "text", + "content": "JD Explore Academy, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 193, + 203, + 400, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 203, + 400, + 217 + ], + "spans": [ + { + "bbox": [ + 193, + 203, + 400, + 217 + ], + "type": "text", + "content": "3 Columbia University, USA; 4 Meta, USA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 220, + 478, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 220, + 478, + 231 + ], + "spans": [ + { + "bbox": [ + 116, + 220, + 478, + 231 + ], + "type": "text", + "content": "{zhenyiwa, qiulings, yjzhu, mgao8}@buffalo.edu; dz2478@columbia.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 196, + 234, + 391, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 234, + 391, + 244 + ], + "spans": [ + { + "bbox": [ + 196, + 234, + 391, + 244 + ], + "type": "text", + "content": "{mathshenli, tiehang.duan}@gmail.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 298, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 290, + 550 + ], + "type": "text", + "content": "Continual Learning (CL) has achieved rapid progress in recent years. However, it is still largely unknown how to determine whether a CL model is trustworthy and how to foster its trustworthiness. This work focuses on evaluating and improving the robustness to corruptions of existing CL models. Our empirical evaluation results show that existing state-of-the-art (SOTA) CL models are particularly vulnerable to various data corruptions during testing. To make them trustworthy and robust to corruptions deployed in safety-critical scenarios, we propose a meta-learning framework of self-adaptive data augmentation to tackle the corruption robustness in CL. The proposed framework, MetaMix, learns to augment and mix data, automatically transforming the new task data or memory data. It directly optimizes the generalization performance against data corruptions during training. To evaluate the corruption robustness of our proposed approach, we construct several CL corruption datasets with different levels of severity. We perform comprehensive experiments on both task- and class-continual learning. Extensive experiments demonstrate the effectiveness of our proposed method compared to SOTA baselines." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "type": "text", + "content": "Humans constantly acquire new information throughout their lifespan and easily recognize information shifts such as structure and style variations in images. Continual learning (CL) aims at imitating human's ability to learn from non-stationary data distributions without forgetting the previously learned knowledge. The past few years have witnessed rapid progress in CL research [1, 30, 31, 36, 45]. Despite the success, existing CL systems overlook the robustness" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 274, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 547, + 370 + ], + "type": "text", + "content": "against unforeseen data shifts during testing. They assume that training and test images for each task follow the same distribution. However, as data distributions evolve and new scenarios occur, test images often encounter various corruptions such as snow, blur, pixelation, and combinations, resulting in a shifted distribution from the training set. For example, Figure 1 shows various corruptions applied on one image from Split-miniImageNet." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "content": "Data corruption can drastically impair the performance of existing image recognition systems. A recent study [21] shows that classification accuracy of various architectures has dropped significantly on the ImageNet test set with some simple and natural corruptions. Our empirical evaluation results show that state-of-the-art CL models are even more vulnerable to these corruptions during testing. For example, the accuracy of " + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{DER}++" + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "content": " [5] for task-continual learning decreases from " + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "inline_equation", + "content": "93.9\\%" + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "inline_equation", + "content": "50.5\\%" + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "content": " on split-CIFAR10; accuracy drops to " + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "inline_equation", + "content": "75.6\\%" + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "content": " on split-CIFAR100; accuracy decreases to " + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "inline_equation", + "content": "61.3\\%" + }, + { + "bbox": [ + 304, + 374, + 547, + 566 + ], + "type": "text", + "content": " on split-miniImageNet by applying those common corruptions on test data of each CL task. This severe issue makes existing CL models highly unreliable in safety-critical applications. Thus, improving the robustness of CL models to foster their trustworthiness when deployed in real-world scenarios is essential." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Training a CL model robust to various corruptions is difficult due to the following challenges. 1) Unseen corruptions could perturb the test set far beyond those encountered during training. A model that naively augments training images with seen corruptions cannot generalize to the new ones during testing. Also, it is unrealistic to enumerate all possible corruptions during training since there are infinite types of corruptions and their combinations. 2) With the ever-evolving data distributions in CL, an effective augmentation strategy learned on previous tasks may gradually become less effective because the optimal augmentation strategies are task-dependent and dynamically change over tasks [11]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "24521" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 152, + 143 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 152, + 143 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 152, + 143 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 152, + 143 + ], + "type": "image", + "image_path": "7018865a0779f5e2d6097ceeab51aaa398a576f5adcca44e96478509a63043cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 144, + 139, + 154 + ], + "lines": [ + { + "bbox": [ + 85, + 144, + 139, + 154 + ], + "spans": [ + { + "bbox": [ + 85, + 144, + 139, + 154 + ], + "type": "text", + "content": "(a) Original data" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 153, + 70, + 223, + 143 + ], + "blocks": [ + { + "bbox": [ + 153, + 70, + 223, + 143 + ], + "lines": [ + { + "bbox": [ + 153, + 70, + 223, + 143 + ], + "spans": [ + { + "bbox": [ + 153, + 70, + 223, + 143 + ], + "type": "image", + "image_path": "bd248e56f61825c5094a381447e681e063e85f8458e33af3352119463e6883e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 144, + 215, + 154 + ], + "lines": [ + { + "bbox": [ + 168, + 144, + 215, + 154 + ], + "spans": [ + { + "bbox": [ + 168, + 144, + 215, + 154 + ], + "type": "text", + "content": "(b) Brightness" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 293, + 143 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 293, + 143 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 293, + 143 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 293, + 143 + ], + "type": "image", + "image_path": "37e78b46c73ed97a17a6757dc2f0e68afbceb916b50f86b9c6442501111f50c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 145, + 283, + 154 + ], + "lines": [ + { + "bbox": [ + 244, + 145, + 283, + 154 + ], + "spans": [ + { + "bbox": [ + 244, + 145, + 283, + 154 + ], + "type": "text", + "content": "(c) Contrast" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 158, + 504, + 170 + ], + "lines": [ + { + "bbox": [ + 87, + 158, + 504, + 170 + ], + "spans": [ + { + "bbox": [ + 87, + 158, + 504, + 170 + ], + "type": "text", + "content": "Figure 1. Visualization of five types of different corruption operations on the testing images of Split-miniImageNet." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 293, + 70, + 362, + 143 + ], + "blocks": [ + { + "bbox": [ + 293, + 70, + 362, + 143 + ], + "lines": [ + { + "bbox": [ + 293, + 70, + 362, + 143 + ], + "spans": [ + { + "bbox": [ + 293, + 70, + 362, + 143 + ], + "type": "image", + "image_path": "8a957671ca9433a669081fd58d90e4e56156bdfb908584d1921f399c859f7997.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 145, + 360, + 154 + ], + "lines": [ + { + "bbox": [ + 304, + 145, + 360, + 154 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 360, + 154 + ], + "type": "text", + "content": "(d) Defocus Blur" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 362, + 71, + 441, + 143 + ], + "blocks": [ + { + "bbox": [ + 362, + 71, + 441, + 143 + ], + "lines": [ + { + "bbox": [ + 362, + 71, + 441, + 143 + ], + "spans": [ + { + "bbox": [ + 362, + 71, + 441, + 143 + ], + "type": "image", + "image_path": "fe0ece7366514591855722284a7840932d204682a7eadf66ea104892bafff090.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 145, + 419, + 154 + ], + "lines": [ + { + "bbox": [ + 385, + 145, + 419, + 154 + ], + "spans": [ + { + "bbox": [ + 385, + 145, + 419, + 154 + ], + "type": "text", + "content": "(e) Elastic" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 442, + 71, + 522, + 143 + ], + "blocks": [ + { + "bbox": [ + 442, + 71, + 522, + 143 + ], + "lines": [ + { + "bbox": [ + 442, + 71, + 522, + 143 + ], + "spans": [ + { + "bbox": [ + 442, + 71, + 522, + 143 + ], + "type": "image", + "image_path": "6fb2639880539c52c8ee1db674f88b42d5a135973c64dafd693bb4e2062e627d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 145, + 493, + 155 + ], + "lines": [ + { + "bbox": [ + 469, + 145, + 493, + 155 + ], + "spans": [ + { + "bbox": [ + 469, + 145, + 493, + 155 + ], + "type": "text", + "content": "(f) Fog" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 179, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 289, + 323 + ], + "type": "text", + "content": "Although we can adopt a memory buffer to store data from previous tasks, augmenting and replaying them at later training iterations, the augmented memory may gradually become less effective as the model could memorize the stored information after replay runs. Recent approaches, such as Augmix [22] composes and combines multiple pre-defined augmentation operations with different depths and widths, show efficacy in improving robustness under traditional supervised classification tasks. However, these approaches are not directly applicable to corruption-robust CL since they only use a fixed random augmentation strategy that is often not optimal for non-stationary data distributions in CL." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 327, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 590 + ], + "type": "text", + "content": "To address these unique challenges of corruption-robustness in CL, we propose a temporally self-adaptive Augmix within a meta-learning framework, named MetaMix. It adaptively augments the memory buffer data and the currently received new data by learning to mix the augmentation operations tailored to the evolving data distributions. In particular, our automatic self-adaptive MetaMix is a bi-level optimization, simulating the evaluation process on unseen corruptions. We randomly divide the training augmentation operations into pseudo-seen and pseudo-unseen operations at each CL step. The lower-level optimization is to optimize the model performance on the pseudo-seen operations; the upper-level optimization is to optimize the generalization of the pseudo-unseen operations. The augmentation strategy is governed by an LSTM, which inputs context information and outputs the corresponding mixing parameters for the augmentations. The proposed MetaMix ensures the augmentation strategy automatically adapts to non-stationary data distribution. Furthermore, the objective is to optimize the performance of the pseudo-unseen corruption operations, which aligns with our goal during testing and encourages the generalization to unseen corruptions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": "To evaluate the corruption robustness of existing and the proposed methods, we propose a new challenging benchmark where various corruptions perturb the testing data of each CL task. To facilitate future research, we construct several new datasets, including split-CIFAR-10-C, split-CIFAR-100-C, and split-miniImageNet-C. Extensive experiments on the constructed benchmarks demonstrate the effectiveness of our proposed MetaMix approach compared with several SOTA data-augmentation approaches adapted for CL. We summarize our contributions as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 179, + 547, + 352 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 317, + 179, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 179, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 179, + 545, + 239 + ], + "type": "text", + "content": "- To our best knowledge, we are the first to study the corruption-robustness of CL methods. Accordingly, we propose the first set of novel benchmarks for evaluating the corruption-robustness of existing CL methods and moving towards trustworthy CL." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 247, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 247, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 247, + 545, + 295 + ], + "type": "text", + "content": "- We propose a self-adaptive augmentation method, MetaMix, by learning to mix and augment the training data of each CL task to achieve corruption-robustness on unseen corruptions for each CL task." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 304, + 547, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 304, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 547, + 352 + ], + "type": "text", + "content": "- Our method is versatile and can be seamlessly integrated with existing CL methods. Extensive experiments with both task/class continual learning demonstrate the effectiveness of MetaMix." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 364, + 392, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 392, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 392, + 376 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 384, + 421, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 421, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 421, + 396 + ], + "type": "text", + "content": "2.1. Continual Learning" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "type": "text", + "content": "CL focuses on learning non-stationary data distribution without forgetting previous knowledge. These methods can be categorized into: 1) retaining memory for future replay [1,4,8,9,17,31,35,42,46,48,49]; 2) designing tailored network architectures [14,24,37,41,54]; 3) performing proper regularization during parameter updates [7,26,40,56]; 4) introducing Bayesian methods for model parameter inference [13,23,34,53]; 5) subspace-based methods [27,30,38] and 6) prompt-tuning [29] based methods [50,51]; 7) transformer-based methods [47,52]. However, all these works focus on the simplified setting in which training and testing data of each CL task follow the same distribution. In practice, the testing data may come from another different distribution due to the variations in weather, light, snow, etc. These distortions pose significant robustness challenges for existing CL methods. Existing literature largely ignored these trustworthy properties of CL. According to the terminology of [43], CL settings can be categorized into three scenarios based on whether task identities of the testing data are available to the CL model during test time. Among the three scenarios, task-continual learning studies the scenario that the task identities are available during testing and is the easiest learning setting; domain incremental learning focuses on the case that all the tasks during CL solve the same classification task with different data distribution and the task identities are unknown during testing; class-continual learning focuses" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24522" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "on the scenario that task identities are unavailable and is the most difficult learning setting. We focus on task/class CL scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 288, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 278 + ], + "type": "text", + "content": "Recently, robust CL began to attract researchers' attention. Self-Purified Replay [25] and PuriDivER [3] consider CL on noisy data stream with label noise. Their work is orthogonal to ours since they consider noisy data during training; we instead consider the robustness to data corruptions during testing. A closely related work to our is Continual Active Adaptation (CAA) [32], which is different from and orthogonal to ours since they focus on the setting that for each data distribution, the label space for each learning task is the same. Each task only encounters a single data corruption. During testing, CAA focuses on generalization to previously seen corruption operations. In contrast, we focus on the setting that each task solves a different task. During testing, the CL model should be robust to unseen data corruptions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 291, + 162, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 162, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 162, + 304 + ], + "type": "text", + "content": "2.2. Data Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 287, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 359 + ], + "type": "text", + "content": "We categorize existing data augmentation techniques into two classes: 1) improving performance on clean data; 2) improving performance and being robust to data distribution shift and corruptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 288, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 505 + ], + "type": "text", + "content": "Augmentations for Improving Clean Data Performance. Standard data augmentation aims to improve the generalization performance on clean data. To improve the performance of neural network, standard supervised learning adopts common augmentations, such as random flips, rotations, and crops. Recently, Cutout [12], Mixup [57], CutMix [55] and MaxUp [16] show excellent performance on clean datasets. Other techniques include automatically learning augmentation policy from data, e.g., AutoAugment [10] and RandAugment [11]. However, AutoAugment is computationally impractical for CL problems since it needs a huge computation cost and separate proxy tasks to train the policy network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 507, + 288, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 288, + 627 + ], + "type": "text", + "content": "Augmentations for Being Robust to Corruptions. Augmix [22] is one of the SOTA methods on the data corruption benchmarks, such as ImageNet-C, CIFAR100-C, etc. DeepAugment [20] and AdversarialAugment [6] transform and augment clean images by perturbing the representations/features of neural networks to generate new augmented images. However, this increases a lot of computation costs. Thus, DeepAugment [20] and AdversarialAugment [6] are impractical for CL. We thus choose the more efficient and effective AugMix and propose a new self-adaptive AugMix for CL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 643, + 129, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 129, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 129, + 657 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "In this section, we first define the corruption robustness of CL and then describe the proposed MetaMix approach. We describe the notation in the main text. We also provide a notation table in Table 4 in Appendix for ease of reading." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 399, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 399, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 399, + 85 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": "Corruption robustness of CL. We consider the problem of learning a sequence of tasks denoted as " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{tr} = \\{\\mathcal{D}_1^{tr},\\mathcal{D}_2^{tr},\\dots ,\\mathcal{D}_N^{tr}\\}" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " is the number of training tasks. The " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": "-th task training data " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k^{tr}" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " consists of many labeled examples " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "(\\pmb{x}^k,y^k,\\mathcal{T}_k)" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\pmb{x}^k" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " is the data example in the task, " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "y^{k}" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " is the corresponding data label, and " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " is the task identifier. The goal is to learn a model " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " on the training task sequence " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{tr}" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " so that it performs well on the testing set of all the learned tasks " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{te} = \\{\\mathcal{D}_1^{te},\\mathcal{D}_2^{te},\\dots ,\\mathcal{D}_N^{te}\\}" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " without forgetting previously learned knowledge. In particular, we focus on improving the model robustness on the testing set for each task " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": " under a collection of corruption operations " + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 91, + 547, + 247 + ], + "type": "text", + "content": ". Below, we define the CL corruption robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "text", + "content": "Definition 1 (CL corruption robustness). Given a sequence of " + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "text", + "content": " tasks with test data " + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{te} = \\{\\mathcal{D}_1^{te},\\mathcal{D}_2^{te},\\dots ,\\mathcal{D}_N^{te}\\}" + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "text", + "content": ", and a collection of corruption operations " + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "text", + "content": ", the corruption robustness is the expectation accuracy after applying corruptions " + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "text", + "content": " on the test data of each task " + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 256, + 547, + 316 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 326, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 326, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 311, + 326, + 545, + 369 + ], + "type": "interline_equation", + "content": "A C C = \\frac {1}{N} \\sum_ {k = 1} ^ {k = N} \\mathbb {E} _ {c \\sim \\mathcal {C}} \\left[ \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {k}, y ^ {k}, \\mathcal {T} _ {k}\\right) \\sim D _ {k} ^ {t e}} \\mathbb {I} \\left(f \\left(c \\left(\\boldsymbol {x} ^ {k}\\right)\\right) = y ^ {k}\\right) \\right] \\tag {1}", + "image_path": "3b06f8ed0bdcfc120d94c4169827d6cdca25f1eb549a09d552d65eb16835b0bf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "inline_equation", + "content": "c(\\pmb{x}^k)" + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "text", + "content": " is the resulting testing data after applying the corruption operation " + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\pmb{x}^k" + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(.)" + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "text", + "content": " is the indicator function. ACC measures the average robust accuracy after applying a collection of corruption operations " + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 368, + 547, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "Preliminary of AugMix. AugMix [22] is a recently proposed SOTA augmentation technique that significantly improves model robustness in standard supervised learning. Considering that CL requires the algorithm to be both efficient and effective, we thus choose the efficient and effective AugMix to adapt to CL. AugMix maintains augmented data diversity by mixing several augmentation chains. Suppose we have an augmentation operation set " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{op_1, op_2, \\dots, op_M\\}" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " during training. The standard AugMix implementation samples three operations " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "op_1" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "op_2" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "op_3" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ", and composes them with varying depths to form several operation chains; where each chain consists of one to three randomly selected augmentation operations. The augmented image is a convex mix of " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " chains (" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " denotes the number of augmentation chains), i.e., " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{aug} = w_1 \\cdot \\text{chain}_1(\\boldsymbol{x}) + w_2 \\cdot \\text{chain}_2(\\boldsymbol{x}) + \\dots + w_A \\cdot \\text{chain}_A(\\boldsymbol{x})" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ", where mixing weights " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w} = (w_1, w_2, \\dots, w_A)" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " is sampled from the distribution, Drichelet " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\alpha, \\alpha, \\dots, \\alpha)" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ". With the mixed image " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{aug}" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ", AugMix combines it with the original image through another random convex combination to obtain the final image " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{augmix}" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{augmix} = m\\boldsymbol{x} + (1 - m)\\boldsymbol{x}_{aug}" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " is sampled from a " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\operatorname{Beta}(\\alpha, \\alpha)" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " distribution. AugMix adopts a fixed random augmentation policy, i.e., Dirichlet and Beta distributions, during the entire training process." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24523" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 74, + 279, + 120 + ], + "blocks": [ + { + "bbox": [ + 55, + 74, + 279, + 120 + ], + "lines": [ + { + "bbox": [ + 55, + 74, + 279, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 74, + 279, + 120 + ], + "type": "image", + "image_path": "9bc3e58034b6a6c6710135d88b192c538cf7e7fdb13fec050b39bee734bfd5b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 124, + 218, + 134 + ], + "lines": [ + { + "bbox": [ + 116, + 124, + 218, + 134 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 218, + 134 + ], + "type": "text", + "content": "(a) Continual Learning Process" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 136, + 281, + 236 + ], + "blocks": [ + { + "bbox": [ + 53, + 136, + 281, + 236 + ], + "lines": [ + { + "bbox": [ + 53, + 136, + 281, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 281, + 236 + ], + "type": "image", + "image_path": "b8b5da81454f5d2fcae324074954ddf84fe5a702baf00360719987d5635bd247.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 239, + 187, + 248 + ], + "lines": [ + { + "bbox": [ + 147, + 239, + 187, + 248 + ], + "spans": [ + { + "bbox": [ + 147, + 239, + 187, + 248 + ], + "type": "text", + "content": "(b) AugMix" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 250, + 282, + 358 + ], + "blocks": [ + { + "bbox": [ + 53, + 250, + 282, + 358 + ], + "lines": [ + { + "bbox": [ + 53, + 250, + 282, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 282, + 358 + ], + "type": "image", + "image_path": "18f1487a47f0f0d0d0d9c1ba79eaa3d74579d1f72e97b6f124ba36706d274f47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 361, + 200, + 370 + ], + "lines": [ + { + "bbox": [ + 135, + 361, + 200, + 370 + ], + "spans": [ + { + "bbox": [ + 135, + 361, + 200, + 370 + ], + "type": "text", + "content": "(c) MetaMix (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "lines": [ + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": "Figure 2. Comparisons between AugMix [22] and our proposed MetaMix. (a) Tasks sequentially arrive during CL. (b) At each CL step, AugMix adopts a fixed and random augmentation policy during the entire learning process, thus lacking adaptation to the non-stationary data in CL and may not be the optimal strategy for each task. (c) MetaMix automatically adapts to the non-stationary data distribution during CL by meta-learning the mixing parameters of augmentation. At time " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": ", we sample a batch of data " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_r, y_r)" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " from the memory buffer and concatenate them with the current received data " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_t, y_t)" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_b, y_b)" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": ". We apply MetaMix to augment each data example. Specifically, we feed the context information " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " into the LSTM cell with parameters " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": ", and obtain the mixing weights " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "w_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": ". The CL model " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " calculates the losses for the augmented data. We adopt bi-level optimization to improve model generalization on unseen corruptions by splitting the augmentation operations into a pseudo-seen set " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " and pseudo-unseen set " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": ". The loss " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\phi_t}" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " is calculated on data augmented with " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " to update " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": ", and loss " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta_t}" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " is calculated on data augmented with " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": " to update " + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 47, + 374, + 289, + 583 + ], + "type": "text", + "content": ". We repeat this process for all tasks in the sequence." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 592, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 592, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 592, + 288, + 628 + ], + "type": "text", + "content": "The augmentation procedure of AugMix is illustrated in Figure 2 (b). The optimization objective of AugMix is shown in the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 637, + 289, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 289, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 289, + 669 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {*} = \\arg \\min _ {\\boldsymbol {\\theta}} \\underset {o p \\in \\mathcal {O}} {\\mathbb {E}} \\mathcal {L} (\\boldsymbol {x}, y, \\boldsymbol {\\theta}) + \\lambda J S (\\boldsymbol {x}, \\boldsymbol {x} _ {a u g m i x 1}, \\boldsymbol {x} _ {a u g m i x 2}), \\tag {2}", + "image_path": "f36eb0defd0b6aa7e67b2c953a4fa0f868ccfde0a7e3111abac3512cf42d0c26.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " is the original data and " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " is the data label; " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\pmb{x},\\pmb{y},\\pmb{\\theta})" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " is the loss function on the labeled data (cross-entropy for classification); " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{augmix1}" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{augmix2}" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " are two augmenta" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "tions of original data " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": " by AugMix; " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "JS" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": " denotes the Jensen Shannon divergence for regularizing the output consistency between the original and augmented data. We will define it in the following section." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 130, + 421, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 130, + 421, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 130, + 421, + 143 + ], + "type": "text", + "content": "3.2. MetaMix Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "type": "text", + "content": "Direct extension of AugMix [22] for CL is to augment data examples sampled from the memory buffer, and the received new data at each time point for memory-based method across the sequentially arrived tasks. In the following, we present our methods for memory-based CL methods, but our methods can be easily extended to other families of CL methods as well, also demonstrated in Section 4. However, since the memory buffer is usually small for CL models, using a fixed random augmentation strategy (like AugMix) may make the CL model gradually memorize the training data corruption operations, but fail to generalize to unseen corruptions during testing. Moreover, with the non-stationary nature of the data, the optimal augmentation strategy for image mixing is data-dependent, and the mixing weights should be specifically optimized for each task. However, since we cannot repeatedly tune the mixing weights for each task during CL learning process, an automatic self-adaptive process is necessary to achieve good performance. We thus introduce the details of our self-adaptive data augmentation approach, namely MetaMix, for improving robustness in CL. Figure 2 (c) illustrates the training process of MetaMix at time " + }, + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "type": "text", + "content": ". Specifically, we adopt meta-learning to learn a set of data-dependent parameters " + }, + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 304, + 149, + 547, + 460 + ], + "type": "text", + "content": ", which can dynamically control the mixing weights of augmentation operations and automatically adapt to the non-stationary data distribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "content": "First, to memorize previous knowledge, we allocate a small memory buffer " + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "content": " to store data from previous tasks by reservoir sampling, similar to [35]. The data in the memory buffer will be replayed later with a new task together. Suppose we have " + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "content": " augmentation operations " + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{op_1, op_2, \\dots, op_M\\}" + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "content": " during training, CL model parameters at time " + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "content": ", and the function represented by the network with " + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}(x)" + }, + { + "bbox": [ + 304, + 461, + 547, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 547, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 690 + ], + "type": "text", + "content": "To capture the sequential dependencies among previous tasks, we design a self-adaptive LSTM mixing strategy. The LSTM inputs the current context information and outputs the adaptive mixing parameters. We use an LSTM (dynamic) model to dynamically change the mixing weights instead of an MLP (static) model since the tasks in CL sequentially arrive. The LSTM can encode the information of previous tasks into an additional hidden state while the MLP can only capture the data information of the current task. It does not capture the information of previous tasks without the hidden state. Formally, the LSTM adaptive process is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 701, + 546, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 701, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 341, + 701, + 546, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {o} _ {t + 1}, \\boldsymbol {h} _ {t + 1}, \\boldsymbol {g} _ {t + 1} = \\operatorname {L S T M} _ {\\phi_ {t}} \\left(\\boldsymbol {I} _ {t}, \\boldsymbol {h} _ {t}, \\boldsymbol {g} _ {t}\\right), \\tag {3}", + "image_path": "37bbef6aeda33d4f28c71f35829f1a60caa1a433ce8bcf913b7c3421732b3806.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24524" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " is the LSTM MetaMixer parameters at time " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " is the hidden state of LSTM at time " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " is the cell state of LSTM at time " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " is the context information encoding as input to LSTM at time " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "o_{t+1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " is the output of LSTM at time " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "t+1" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " for the mixing parameters, i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "o_{t+1} = \\{w_t', m_t'\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "m_t'" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " is the mixing weight for the augmented data and original raw data and " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "w_t'" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " are the mixing weights for different chains in the augmentation. We apply softmax function to " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "w_t'" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " and sigmoid function to " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "m_t'" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " to ensure they are within [0,1] interval, i.e.," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 81, + 198, + 288, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 198, + 288, + 211 + ], + "spans": [ + { + "bbox": [ + 81, + 198, + 288, + 211 + ], + "type": "interline_equation", + "content": "m _ {t} = \\operatorname {S i g m o i d} \\left(m _ {t} ^ {\\prime}\\right), \\boldsymbol {w} _ {t} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {w} _ {t} ^ {\\prime}\\right). \\tag {4}", + "image_path": "379612a3ffc1eee251164cbec30025c41bee03fc73fd37e95efc5ca30269b389.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 215, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 289, + 240 + ], + "type": "text", + "content": "The context " + }, + { + "bbox": [ + 47, + 215, + 289, + 240 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 47, + 215, + 289, + 240 + ], + "type": "text", + "content": " contains essential information to characterize data in current and previous tasks. It is shown as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 256, + 287, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 256, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 51, + 256, + 287, + 280 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {t} = \\left\\{\\nabla \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {y} _ {r}\\right) \\cdot \\nabla \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {t}\\right), \\mathcal {L} _ {\\boldsymbol {\\theta} _ {t}} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {y} _ {r}\\right), \\boldsymbol {e} _ {r}, \\boldsymbol {e} _ {t} \\right\\}, \\tag {5}", + "image_path": "bad13fc55f123aca09972d8617c6101e8029b5564087577a20cccf6b33bf3fad.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_r, y_r)" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": " is the labeled mini-batch data sampled from the memory buffer and " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_t, y_t)" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": " is the labeled data received at time " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": ". We adopt this information for several reasons: (1) " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}_{\\pmb{\\theta}_t}(\\pmb{x}_r, y_r) \\cdot \\nabla \\mathcal{L}_{\\pmb{\\theta}_t}(\\pmb{x}_t, y_t)" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": " is the gradient dot product between memory data gradient and current mini-batch data gradient. It determines the degree of interference and transfer between memory data and current task mini-batch data. It is commonly used in existing CL as context information [18, 35]; (2) we also input the memory mini-batch data loss, as it indicates the degree of forgetting of previous tasks; (3) " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{e}_r" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": " are data encoding for mini-batch data, " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": ", sampled from memory buffer and currently received new mini-batch data, " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": ", respectively. That is, " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{e}_r" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t" + }, + { + "bbox": [ + 46, + 281, + 289, + 472 + ], + "type": "text", + "content": " are the last layer features outputted by ResNet-18. This determines the data-dependent features for calculating the data-dependent mixing weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": "Since corruption operations on testing data are unseen during training, we cannot directly optimize the augmentation strategy for these corruptions. To make the CL model generalize better to unseen corruptions, at each training step " + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": ", we randomly split the augmentation operations " + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": " into non-overlapping pseudo-seen " + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": " and pseudo-unseen " + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_t" + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": " operations, i.e., " + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "inline_equation", + "content": "S_{t} \\cup \\mathcal{U}_{t} = \\mathcal{O}" + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "inline_equation", + "content": "S_{t} \\cap \\mathcal{U}_{t} = \\emptyset" + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": ". We use the pseudo operation sets to simulate the scenario in which training and testing corruptions have no overlap. We formulate MetaMix as a bi-level optimization problem to make the CL model learn to augment and mix. This ensures that the update of model parameters " + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": " on the pseudo-seen set can improve the optimization of MetaMixer parameters " + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 473, + 289, + 664 + ], + "type": "text", + "content": " on the pseudo-unseen set, which could, in turn, be beneficial to improve the CL model generalization. We formulate the MetaMix as the following bi-level optimization:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 669, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 669, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 52, + 669, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\phi_ {*} = \\arg \\min _ {\\phi} \\mathbb {E} _ {o p \\in \\mathcal {U} _ {t}} \\mathcal {L} \\left(\\boldsymbol {x} _ {b}, y _ {b}, \\boldsymbol {\\theta} _ {*}, \\phi\\right) + \\lambda J S \\left(\\boldsymbol {x} _ {b}, \\widehat {\\boldsymbol {x}} _ {b 1}, \\widehat {\\boldsymbol {x}} _ {b 2}\\right), \\tag {6} \\\\ \\boldsymbol{\\theta}_{*} = \\arg \\min_{\\boldsymbol{\\theta}}\\underset {op\\in \\mathcal{S}_{t}}{\\mathbb{E}}\\mathcal{L}(\\boldsymbol{x}_{b},y_{b},\\boldsymbol {\\theta},\\boldsymbol {\\phi}) + \\lambda J S(\\boldsymbol{x}_{b},\\boldsymbol{x}^{\\prime}_{b1},\\boldsymbol{x}^{\\prime}_{b2}), \\\\ \\end{array}", + "image_path": "5b4ab678f9ca8c7569c862e17fe4df212ec8656c8aa42c5e89e197f046363b56.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "where we concatenate the mini-batch data " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_r, y_r)" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " from the memory buffer and the current received mini-batch data " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_t, y_t)" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " together as " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_b, y_b)" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}_{b1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}_{b2}" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " are the two minibatch data augmented by applying the pseudo-unseen augmentation operations " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_t" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_b, y_b)" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{b1}'" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{b2}'" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " are the two mini-batch data augmented by applying the pseudo-seen augmentation operations " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_t" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_b, y_b)" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " is the regularization constant for the consistency loss. The lower-level optimization is to optimize the performance on the pseudoseen augmentation operations " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_t" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": " at the training step " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "; and the upper-level optimization is to optimize the generalization on the pseudo-unseen augmentation operations " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_t" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": ". This bi-level optimization for learning to augment and mix can be efficiently solved by first-order gradient descent method, similar to the first-order MAML [15]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 256, + 547, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 547, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 547, + 400 + ], + "type": "text", + "content": "The advantage of MetaMix (Eq. (6)) over AugMix (Eq. (2)) are two folds. First, MetaMix improves the generalization to unseen corruptions since we split the training data augmentation operations into pseudo-seen and unseen operations. This is similar to the train-validation data split in standard meta-learning, which improves the generalization compared to the training without data split [39]; It could be understood as introducing an implicit regularization so that it can benefit from low sample complexity [39]. Second, MetaMix can automatically adapt to the non-stationary data distribution in CL, while AugMix can only work well in the stationary data distribution in a single task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "text", + "content": "Consistency Regularization To ensure the network output consistency between the original raw image data and the augmented data, we use a similar consistency loss [22] to regularize the network output to ensure smoother neural network responses. We hope the model to have similar responses to " + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b, \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2}" + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "text", + "content": ". We minimize the Jensen-Shannon divergence among the posterior distributions of the original sample " + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b" + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "text", + "content": " and its augmented variants " + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}_{b1}" + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}_{b2}" + }, + { + "bbox": [ + 304, + 405, + 547, + 513 + ], + "type": "text", + "content": ". The consistency loss is defined as below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 533, + 545, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 533, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 302, + 533, + 545, + 575 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {m e a n} = \\left(p _ {\\boldsymbol {x} _ {b}} + p _ {\\hat {\\boldsymbol {x}} _ {b 1}} + p _ {\\hat {\\boldsymbol {x}} _ {b 2}}\\right) / 3 (7) \\\\ J S \\left(\\boldsymbol {x} _ {b}, \\widehat {\\boldsymbol {x}} _ {b 1}, \\widehat {\\boldsymbol {x}} _ {b 2}\\right) = \\left(\\mathbb {K L} \\left(p _ {\\boldsymbol {x} _ {b}} \\mid p _ {\\text {m e a n}}\\right) + \\mathbb {K L} \\left(p _ {\\widehat {\\boldsymbol {x}} _ {b 1}} \\mid p _ {\\text {m e a n}}\\right) \\right. \\\\ + \\mathbb {K L} \\left(p _ {\\widehat {x} _ {b 2}} \\mid p _ {\\text {m e a n}}\\right)) / 3 (8) \\\\ \\end{array}", + "image_path": "545a258582d290b9914363c80486fb0329f705b931589ea3fc5c14024041e982.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{KL}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " denotes the KL divergence between two distributions. " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "p_{\\boldsymbol{x}_b} = f_{\\boldsymbol{\\theta}_t}(\\boldsymbol{x}_b)" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " is the network output probabilities of each class for original raw data " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_b" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ". Similarly, we can define " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "p_{\\widehat{\\boldsymbol{x}}_{b1}} = f_{\\boldsymbol{\\theta}_t}(\\widehat{\\boldsymbol{x}}_{b1})" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "p_{\\widehat{\\boldsymbol{x}}_{b2}} = f_{\\boldsymbol{\\theta}_t}(\\widehat{\\boldsymbol{x}}_{b2})" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ". We summarize our proposed algorithm in Algorithm 3. Line 5-8 takes the context information as input and computes desired augmentation mixing parameters. Line 9-11 augments the original data with the pseudo-seen augmentation operations " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_t" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " and pseudo-unseen operations " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_t" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " by Algorithm 1. Line 12 solves the bi-level optimization problem (Eq. (6)) by Algorithm 2." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24525" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 51, + 87, + 289, + 274 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 140, + 83 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 140, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 140, + 83 + ], + "type": "text", + "content": "Algorithm 1 DataMix" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "lines": [ + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "spans": [ + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": "1: Function DataMix " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_b, \\pmb{w}_t, m_t, \\mathcal{V})" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " \n2: MetaMix augmented images initialized as " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b' = 0" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{orig} = \\pmb{x}_b" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " \n4: for " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " do { A is the number of augmentation chains} \n5: " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b = \\pmb{x}_{orig}" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " \n6: for " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " do { " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " is the augmentation depth} \n7: randomly sample operation " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "op \\in \\mathcal{V}" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " is the collection of augmentation operations) \n8: " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b = op(\\pmb{x}_b)" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " \n9: end for \n10: " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b' + = \\pmb{w}_t^i \\cdot \\pmb{x}_b" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " \n11: end for \n12: " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b' = m_t \\pmb{x}_{orig} + (1 - m_t) \\pmb{x}_b'" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": "; \n13: similar to the above procedure, we can get two different minibatch " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{b1}', \\pmb{x}_{b2}'" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " \n14: return " + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{b1}', \\pmb{x}_{b2}'" + }, + { + "bbox": [ + 51, + 87, + 289, + 274 + ], + "type": "text", + "content": " \n15: EndFunction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 50, + 304, + 289, + 459 + ], + "blocks": [ + { + "bbox": [ + 48, + 289, + 165, + 301 + ], + "lines": [ + { + "bbox": [ + 48, + 289, + 165, + 301 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 165, + 301 + ], + "type": "text", + "content": "Algorithm 2 Bi-level Solver" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "lines": [ + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "spans": [ + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": "1: Function Bi-level Solver " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "(\\pmb{\\theta}_t, \\pmb{\\phi}_t, \\pmb{x}_b, y_b, \\pmb{x}_{b1}', \\pmb{x}_{b2}', \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2})" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " \n2: optimize the lower-level optimization problem in Eq. (6) by " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{t+1} = \\pmb{\\theta}_t - \\gamma \\nabla_\\theta \\mathbb{E}_{op \\in S_t}[\\mathcal{L}(\\pmb{x}_b, y_b, \\pmb{\\theta}_t, \\phi) + \\lambda JS(\\pmb{x}_b, \\pmb{x}_{b1}', \\pmb{x}_{b2}')]" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " \n3: calculate the upper-level optimization loss function in Eq. (6) by " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\pmb{\\theta}_{t+1}, \\phi) = \\mathbb{E}_{op \\in U_t}[\\mathcal{L}(\\pmb{x}_b, y_b, \\pmb{\\theta}_{t+1}, \\phi) + \\lambda JS(\\pmb{x}_b, \\widehat{\\pmb{x}}_{b1}, \\widehat{\\pmb{x}}_{b2})]" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\phi_1' = \\phi_t" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " \n5: for " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " do \n6: " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\phi_{s+1}' = \\phi_s' - \\beta \\nabla_{\\phi'} \\mathcal{E}(\\pmb{\\theta}_{t+1}, \\phi_s')" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " \n7: end for \n8: " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\phi_{t+1} = \\phi_{J+1}'" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " \n9: return " + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{t+1}, \\phi_{t+1}" + }, + { + "bbox": [ + 50, + 304, + 289, + 459 + ], + "type": "text", + "content": " \n10: EndFunction" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 483, + 123, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 123, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 123, + 496 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 503, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 552 + ], + "type": "text", + "content": "We evaluate the model performance for both task- and class-continual learning. The CL models are evaluated on the proposed benchmark for each dataset. We then describe experiment setup, results, and ablation study." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 155, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 155, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 155, + 574 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "text", + "content": "Below, we first construct the corruption CL benchmark and then describe the evaluation metrics, baselines, and implementation details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 714 + ], + "type": "text", + "content": "Corruption CL benchmark. Existing CL benchmarks, such as Split-CIFAR10, Split-CIFAR100, and Split-miniImageNet, are commonly used in literature. However, each task's training and testing data follow the same distribution for these datasets. Thus, these benchmarks are not suitable for evaluating the corruption robustness of CL methods. To properly evaluate the corruption robustness, we follow the protocols [21] and create several new CL datasets." + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 310, + 86, + 547, + 406 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 400, + 83 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 400, + 83 + ], + "type": "text", + "content": "Algorithm 3 MetaMix" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "lines": [ + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": "1: REQUIRE: Augmentation Operations " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " {sterize, rotate, ..., solarize}; augmentation width " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; augmentation depth " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; the number of CL tasks N; the number of training iterations for each task " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; memory buffer " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; CL model parameters " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; LSTM MetaMix with parameters " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; MetaMixer LSTM update steps " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; CL model parameter learning rate " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; regularization weight " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; original image " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{orig}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " ; initial hidden state and cell state " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\pmb{h}_t,\\pmb{g}_t" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " are initialized as random noise " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\pmb{h}_1,\\pmb{g}_1\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " do \n3: for " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " do \n4: randomly sample mini-batch data " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "(x_{r},y_{r})" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " from memory buffer " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " and concatenate it with the current received mini-batch data " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "(x_{t},y_{t})" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "(x_{b},y_{b})" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " \n5: compute context vector " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\pmb{I}_t = \\{\\nabla \\mathcal{L}_{\\theta_t}(\\pmb{x}_r,\\pmb{y}_r)" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}_{\\theta_t}(\\pmb{x}_t,\\pmb{y}_t),\\mathcal{L}_{\\theta_t}(\\pmb{x}_r,\\pmb{y}_r),\\pmb{e}_r,\\pmb{e}_t\\}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "o_{t + 1},h_{t + 1},g_{t + 1} = \\mathrm{LSTM}_{\\phi_t}(I_t,h_t,g_t)" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " \n7: generate mixing parameters: " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\pmb{w}_t^{\\prime},m_t^{\\prime} = o_{t + 1}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "m_t = \\mathrm{Sigmoid}(m_t^{\\prime}),w_t = \\mathrm{Softmax}(w_t^{\\prime})" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " \n9: randomly split augmentation operations " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " into nonoverlapping pseudo-seen " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " and unseen " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_t" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{b1}^{\\prime},\\pmb{x}_{b2}^{\\prime} = \\mathrm{DataMix}(\\pmb{x}_b,\\pmb{w}_t,m_t,\\mathcal{S}_t)" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " (Algorithm 1) \n11: " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}_{b1},\\widehat{\\pmb{x}}_{b2} = \\mathrm{DataMix}(\\pmb{x}_b,\\pmb{w}_t,m_t,\\mathcal{U}_t)" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " (Algorithm 1) \n12: " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\theta_{t + 1},\\phi_{t + 1} = \\mathrm{Bi - level~Solver~}(\\theta_t,\\phi_t,\\pmb{x}_b,\\pmb{y}_b,\\pmb{x}_{b1}'',\\pmb{x}_{b2}'',\\widehat{\\pmb{x}}_{b1},\\widehat{\\pmb{x}}_{b2})" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " (Algorithm 2) \n13: update memory buffer " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " through reservoir sampling [35] to determine whether to store " + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "inline_equation", + "content": "(x_{t},y_{t})" + }, + { + "bbox": [ + 310, + 86, + 547, + 406 + ], + "type": "text", + "content": " \n14: end for \n15: end for" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 434, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 674 + ], + "type": "text", + "content": "We apply 15 common corruption operations to the testing data of the three CL benchmarks, such as snow, frost, fog, etc. We provide a brief description for each corruption operation in Table 5 in Appendix. We perform each corruption at five different severity levels. Note that these corruptions only appear on the testing dataset of each continually learned task and are not seen in the training data of each task. This is to ensure that we can measure the robustness to unseen corruptions during testing for each compared method. To our best knowledge, we build the first CL corruption benchmark for measuring the generalization of CL models to unseen corruptions during testing. We name these new datasets Split-CIFAR10-C, Split-CIFAR100-C, and Split-miniImageNet-C. Specifically, for Split-CIFAR-10-C, we split the CIFAR-10 dataset [28] into 5 disjoint tasks, where each task has 2 classes; for Split-CIFAR-100-C, we split CIFAR-100 [28] consisting of 100 classes into 10 disjoint tasks, where each task has 10 classes; for Split-miniImageNet-C, we split miniImageNet [44] that consists of 100 classes, into 10 disjoint tasks, where each task has 10 classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the performance of different methods with average accuracy and backward transfer at the end of CL training to measure the final performance" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24526" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "content": "and the extent of forgetting for different methods. For each task, the accuracy for a specific corruption " + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "content": " is the average accuracy over the five severity level. Suppose for task " + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "ACC_{c,s}^{k}" + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "content": " measures the robust accuracy after applying the corruption operation " + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "content": " with severity level " + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 139, + 264, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 139, + 264, + 156 + ], + "spans": [ + { + "bbox": [ + 70, + 139, + 264, + 156 + ], + "type": "interline_equation", + "content": "A C C _ {c, s} ^ {k} = \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {k}, y ^ {k}, \\mathcal {T} _ {k}\\right) \\sim D _ {t e} ^ {t e}} \\mathbb {I} \\left(f \\left(c _ {s} \\left(\\boldsymbol {x} ^ {k}\\right)\\right) = y ^ {k}\\right),", + "image_path": "8e4ccf737ff8f88ee36759ae3160a0eacf8fa793accde6b9bae3e3ea24c599b8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "inline_equation", + "content": "c_{s}(\\pmb{x}^{k})" + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": " is the corruption operation " + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": " with severity level " + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": " applied on " + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{k}" + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": ". The accuracy for corruption " + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "inline_equation", + "content": "ACC_{c}^{k}" + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": " is the average robust accuracy across all the five severity levels. The overall accuracy for task " + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "inline_equation", + "content": "ACC^{k}" + }, + { + "bbox": [ + 47, + 163, + 287, + 224 + ], + "type": "text", + "content": ", is the average accuracy across all corruption types, i.e.," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 231, + 283, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 231, + 283, + 262 + ], + "spans": [ + { + "bbox": [ + 51, + 231, + 283, + 262 + ], + "type": "interline_equation", + "content": "A C C _ {c} ^ {k} = \\frac {1}{5} \\sum_ {s = 1} ^ {s = 5} A C C _ {c, s} ^ {k}, \\text {a n d} A C C ^ {k} = \\frac {1}{| \\mathcal {C} |} \\sum_ {c \\in \\mathcal {C}} A C C _ {c} ^ {k}.", + "image_path": "ca7a164e80e87d0642a0a261f387bbd23e80997aac45bf9729e2860ad531cdfe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": "To measure catastrophic forgetting, we also evaluate the backward transfer (BWT), which measures the extent of forgetting on previous tasks after learning new ones. " + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "inline_equation", + "content": "BWT < 0" + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": " reveals the occurrence of catastrophic forgetting on previous tasks, and " + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "inline_equation", + "content": "BWT > 0" + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": " indicates that learning new tasks is helpful for previous tasks. Formally, the backward transfer (BWT) is defined as: " + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "inline_equation", + "content": "BWT = \\frac{1}{N - 1}\\sum_{k = 1}^{k = N - 1}(R_{N,k} - R_{k,k})" + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": "; where " + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "inline_equation", + "content": "R_{N,k}" + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": " is the testing accuracy on task " + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": " after learning on task " + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "inline_equation", + "content": "R_{N,k}" + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": " is defined the same as the above mentioned " + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "inline_equation", + "content": "ACC^k" + }, + { + "bbox": [ + 47, + 270, + 287, + 390 + ], + "type": "text", + "content": ". For BWT, the higher, the better." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 390, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 288, + 450 + ], + "type": "text", + "content": "Continual Learning Setting. We apply our proposed MetaMix to task-continual learning (Task-CL) and class-continual learning (Class-CL). The former provides task identities for the CL learner to select the relevant classifier for each example during testing, whereas the latter does not." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 450, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 617 + ], + "type": "text", + "content": "Baseline. Our method can be seamlessly integrated with exiting CL methods. For illustration, we integrate the proposed methods with SOTA memory-based methods: " + }, + { + "bbox": [ + 46, + 450, + 288, + 617 + ], + "type": "inline_equation", + "content": "DER++" + }, + { + "bbox": [ + 46, + 450, + 288, + 617 + ], + "type": "text", + "content": " [5] and CLS-ER [2]. To evaluate the effectiveness of the proposed MetaMix, we compare it with several SOTA data augmentation approaches originally designed for traditional supervised learning: Adversarial Training (AT) [33], RandAugment (RA) [11], Maxup [16], DeepAugment [20], Augmix [22]. We adapt these augmentation methods for CL problems by applying them on the new task data and the memory buffer data. We provide more detailed descriptions of baselines in Appendix. Due to space limitations, we put the experiments that integrate our proposed method with CLS-ER in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Implementation details. We use ResNet18 [19] as classifier for all datasets. All the other hyperparameter settings follow from [5]. The types of corruption during testing are not seen during training for all the compared augmentation methods to evaluate the robustness against unseen corruptions. Following [22], the augmentation operations performed during training are autocontrast, equalize, posterize, rotate, solarize, shear-x, shear-y, translate-x, translate-y." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "All the above-compared data augmentation methods apply these augmentation operations to CL baseline method to improve robustness. The memory buffer has capacity of 500 data points by default. We average the result for 5 runs for each experiment. We provide both mean and standard deviation across different runs. Due to space limitations, we provide standard deviation results in Appendix. We provide more implementation details in Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 173, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 245 + ], + "type": "text", + "content": "LSTM Architecture. It has one recurrent layer with 5 hidden units. A linear layer (fully connected layer) is appended next to the LSTM output to generate the mixing parameters " + }, + { + "bbox": [ + 304, + 173, + 545, + 245 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_t" + }, + { + "bbox": [ + 304, + 173, + 545, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 173, + 545, + 245 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 304, + 173, + 545, + 245 + ], + "type": "text", + "content": ". The LSTM only has about 20K parameters and is negligible compared to ResNet18, which has more than 11.22 million parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 266, + 437, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 437, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 437, + 278 + ], + "type": "text", + "content": "4.2. Robustness of Task-CL" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 290, + 547, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 547, + 374 + ], + "type": "text", + "content": "Task-Continual Learning In this section, we evaluate the corruption robustness under task-continual learning, where the task identifiers are provided to the CL learner during testing. We present the results that integrate baselines and the proposed method with DER++ [5]. Due to space limitations, we put the experiments that integrate baselines and MetaMix with CLS-ER [2] in Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "text", + "content": "Results. We show the results across different corruptions and compared methods in Tables 1-2; where " + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "inline_equation", + "content": "Avg" + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "text", + "content": " is the average robust accuracy across all the corruptions. Due to the space limitations, we provide the standard deviation across different runs in Table 8 and 9 in Appendix. We also put results on split-CIFAR10-C in Table 6 and Table 7 in Appendix. We can observe that without data augmentation, the performance of " + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{DER}++" + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "text", + "content": " is close to random guessing, indicating that the CL model does not generalize to common corruptions. The proposed self-adaptive MetaMix brings significant improvement toward corruption robustness. It outperforms the average corruption robustness accuracy of the other models by a large margin of " + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 379, + 547, + 715 + ], + "type": "text", + "content": " on Split-CIFAR100-C, and Split-miniImageNet-C, respectively. The improvement shows the effectiveness of proposed MetaMix, which can automatically adapt to the non-stationary data distribution. The compared methods only use fixed static/random augmentation methods which lack adaptation to the non-stationary data in CL, and thus do not perform well. Random augmentation achieves second-best results. Deep-augment would distort the image such that they are very different from original images. Adversarial training and Maxup do not help much corruption robustness among the compared methods. Adversarial training can generate similar images compared to original images. Thus, there is a large gap between training and testing corrupted data for those methods adapted for CL. Maxup aims for improving clean data performance, not designed for robust accuracy." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24527" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 85, + 526, + 171 + ], + "blocks": [ + { + "bbox": [ + 159, + 70, + 433, + 82 + ], + "lines": [ + { + "bbox": [ + 159, + 70, + 433, + 82 + ], + "spans": [ + { + "bbox": [ + 159, + 70, + 433, + 82 + ], + "type": "text", + "content": "Table 1. Robust accuracy of Task-CL on Split-CIFAR100-C with DER++" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 85, + 526, + 171 + ], + "lines": [ + { + "bbox": [ + 66, + 85, + 526, + 171 + ], + "spans": [ + { + "bbox": [ + 66, + 85, + 526, + 171 + ], + "type": "table", + "html": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++10.4610.6910.2610.7610.5110.7110.6610.4710.5610.5110.7910.5410.7510.5510.7210.60
AT10.3610.4510.2510.2310.2410.2710.2510.129.889.9810.169.8310.1910.310.3110.19
RA26.7727.1525.4627.8423.4624.0525.9923.2626.3430.4527.9528.825.5429.1229.0926.75
DeepAugment11.0011.1610.8611.3511.1511.4511.3310.7610.4610.5911.1010.5711.2011.3611.2511.04
Maxup11.111.1711.1311.0911.0611.1311.110.8810.7610.8110.8610.3711.0611.1311.1210.98
Augmix59.5862.6859.9267.9558.8865.9367.0363.3963.9363.2468.6263.2765.8266.0664.0764.02
Ours61.5365.1461.1870.3561.2868.0168.7666.0266.4966.1471.4966.1668.2468.9166.7566.43
", + "image_path": "2f0bb9b21acdf49e9d9d018c950a7543eaaf15d11dd9caa879cc59636713dcb2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 184, + 525, + 269 + ], + "blocks": [ + { + "bbox": [ + 152, + 173, + 440, + 183 + ], + "lines": [ + { + "bbox": [ + 152, + 173, + 440, + 183 + ], + "spans": [ + { + "bbox": [ + 152, + 173, + 440, + 183 + ], + "type": "text", + "content": "Table 2. Robust accuracy of Task-CL on Split-MiniImageNet-C with DER++" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 184, + 525, + 269 + ], + "lines": [ + { + "bbox": [ + 66, + 184, + 525, + 269 + ], + "spans": [ + { + "bbox": [ + 66, + 184, + 525, + 269 + ], + "type": "table", + "html": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++10.2710.110.079.789.739.679.729.619.719.859.6910.159.559.459.579.79
AT9.079.249.279.659.629.69.489.179.109.019.128.389.549.569.359.28
RA17.2318.2216.719.8116.618.3219.0718.9417.1418.2617.2614.5621.1420.8423.018.47
DeepAugment10.5410.5310.6210.6810.7010.6310.7410.5610.3210.3210.529.9210.5510.4110.4910.50
Maxup10.410.910.6810.659.7510.2110.259.6210.1210.2110.3410.910.210.1610.0710.30
Augmix43.4548.9142.2053.7849.6859.8851.8353.2254.0255.6359.5943.4661.3551.3462.7552.74
Ours48.853.3343.8159.753.6462.7856.9556.1957.3156.0862.0741.7163.1749.7864.5555.32
", + "image_path": "d3ea6fc99a9946344001d94e4d408641fcc408a084e968d7ccf94ca75c0b0926.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 66, + 281, + 525, + 368 + ], + "blocks": [ + { + "bbox": [ + 158, + 270, + 434, + 281 + ], + "lines": [ + { + "bbox": [ + 158, + 270, + 434, + 281 + ], + "spans": [ + { + "bbox": [ + 158, + 270, + 434, + 281 + ], + "type": "text", + "content": "Table 3. Robust accuracy of Class-CL on Split-CIFAR100-C with DER++" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 281, + 525, + 368 + ], + "lines": [ + { + "bbox": [ + 66, + 281, + 525, + 368 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 525, + 368 + ], + "type": "table", + "html": "
MethodNoiseBlurWeatherDigitalAvg
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG
DER++0.750.750.880.750.860.770.680.880.960.920.750.960.780.780.790.82
AT0.990.971.040.910.950.970.920.831.050.940.921.00.910.950.930.95
RA6.226.35.676.254.424.545.595.256.247.726.426.935.166.817.116.04
DeepAugment1.291.291.251.261.261.231.241.261.241.241.141.261.221.281.331.25
Maxup1.171.151.161.151.081.141.121.081.081.091.141.061.131.151.111.12
Augmix19.5721.2120.3526.1819.0825.1725.4622.0922.4322.8226.4422.5423.7224.6423.0122.98
Ours22.3224.4423.1529.1922.1328.1028.3525.4725.226.0829.5625.7827.0128.1625.8226.05
", + "image_path": "8193a5865e8b6ae064f1e154e54344c2474dddcfe2427dbfecdd8ee2a38a7637.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 381, + 181, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 181, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 181, + 393 + ], + "type": "text", + "content": "4.3. Robustness of Class-CL" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 400, + 288, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 448 + ], + "type": "text", + "content": "Class-Continual Learning. In this section, we evaluate the corruption robustness under the more challenging class-continual learning setting, where the task identifiers are unavailable to the CL learner during testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 449, + 288, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 591 + ], + "type": "text", + "content": "Results. We present the results across different corruptions and compared methods on Split-CIFAR100-C in Table 3, we present the results on Split-miniImageNet-C in Table 10 in Appendix. Similar to the Task-CL, MetaMix significantly improves over the baseline augmentation methods in Class-CL setting due to its self-adaptivity property. It outperforms the average corruption robustness accuracy of the other models by a large margin of " + }, + { + "bbox": [ + 46, + 449, + 288, + 591 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 449, + 288, + 591 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 449, + 288, + 591 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 449, + 288, + 591 + ], + "type": "text", + "content": " on Split-CIFAR100-C, and Split-miniImageNet-C, respectively. This is because our proposed method can automatically adapt to the non-stationary distributions. Most compared methods do not perform well on this challenging setup." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 592, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 288, + 664 + ], + "type": "text", + "content": "Due to space limitations, we put BWT results in Table 15 in Appendix. In the corruption-robustness scenario, BWT is no longer a meaningful metric with such extreme accuracy differences, as the significantly lower accuracy of comparison methods results in much less space for further performance variations during backward transfer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 671, + 286, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 286, + 685 + ], + "type": "text", + "content": "4.4. Ablation Study and Hyperparameter Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Due to the limited space, we provide (1) ablation studies; (2) hyperparameter analyses, including " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": ", etc.; (3)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 381, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 381, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 547, + 418 + ], + "type": "text", + "content": "The effect of using LSTM vs MLP and the benefit of using additional hidden state information from previous tasks; (4) computation cost in Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "content": "Effect of Memory Size. We evaluate the effect of memory size with 500 and 3000, respectively. The memory size of 500 is the default setting in the above tables. We provide experiment results on Split-CIFAR100-C, Split-miniImageNet-C with memory size 3000 on task-CL and class-CL respectively in Table 11, 12, 13, 14 in Appendix. In these cases, our method significantly outperforms baselines." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 529, + 378, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 378, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 378, + 542 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 554, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 673 + ], + "type": "text", + "content": "This paper tackles a more challenging problem of corruption robustness in CL with non-stationary data distribution, where the testing data distribution may significantly differ from training data distribution with various forms of common corruptions. We propose a meta-learning framework, MetaMix, for self-adaptive data augmentation specialized for CL. Our proposed MetaMix can substantially improve the model's robustness. Comprehensive experiments on both task- and class-CL settings demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgement We thank all the anonymous reviewers for their insightful and thoughtful comments. This research was supported in part by grant NSF III-1910492." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24528" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 146 + ], + "type": "text", + "content": "[1] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. Advances in Neural Information Processing Systems 32, pages 11849-11860, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Elahe Arani, Fahad Sarfraz, and Bahram Zonooz. Learning fast, learning slow: A general continual learning method based on complementary learning system. In International Conference on Learning Representations, 2022. 7, 12, 17" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "text", + "content": "[3] Jihwan Bang, Hyunseo Koh, Seulki Park, Hwanjun Song, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on a contaminated data stream with blurry task boundaries. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9265-9274, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 293 + ], + "type": "text", + "content": "[4] E. Belouadah and A. Popescu. II2m: Class incremental learning with dual memory. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 583-592, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 339 + ], + "type": "text", + "content": "[5] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. 34th Conference on Neural Information Processing Systems, 2020. 1, 7, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 393 + ], + "type": "text", + "content": "[6] Dan A. Calian, Florian Stimberg, Olivia Wiles, Sylveste-Alvise Rebuffi, Andras Gyorgy, Timothy Mann, and Sven Gowal. Defending against image corruptions through adversarial augmentations. https://arxiv.org/abs/2104.01086, 2021.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "text", + "content": "[7] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio Calmon, and Taesup Moon. {CPR}: Classifier-projection regularization for continual learning. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 485 + ], + "type": "text", + "content": "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. Proceedings of the International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 540 + ], + "type": "text", + "content": "[9] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. https://arxiv.org/abs/1902.10486, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "type": "text", + "content": "[10] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation policies from data. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "text", + "content": "[11] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. Advances in Neural Information Processing Systems, 2020. 1, 3, 7, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "text", + "content": "[12] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. https://arxiv.org/abs/1708.04552, 2017.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[13] Sayna Ebrahimi, Mohamed Elhoseiny, Trevor Darrell, and Marcus Rohrbach. Uncertainty-guided continual learning with bayesian neural networks. Proceedings of the International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "text", + "content": "[14] Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. https://arxiv.org/abs/1701.08734, 2017.2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "type": "text", + "content": "[15] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. International Conference on Machine Learning, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[16] Chengyue Gong, Tongzheng Ren, Mao Ye, and Qiang Liu. Maxup: Lightweight adversarial training with data augmentation improves neural network training. Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 7, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 547, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 273 + ], + "type": "text", + "content": "[17] Yanan Gu, Xu Yang, Kun Wei, and Cheng Deng. Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 547, + 308 + ], + "type": "text", + "content": "[18] Gunshi Gupta, Karmesh Yadav, and Liam Paull. La-maml: Look-ahead meta learning for continual learning. In Advances in Neural Information Processing Systems, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 310, + 547, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 547, + 353 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2016. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 355, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 547, + 420 + ], + "type": "text", + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. International Conference on Computer Vision, 2021. 3, 7, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 547, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 547, + 465 + ], + "type": "text", + "content": "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 468, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 547, + 510 + ], + "type": "text", + "content": "[22] Dan Hendrycks, Norman Mu, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple method to improve robustness and uncertainty under data shift. 2020. 2, 3, 4, 5, 7, 12" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 513, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 547, + 578 + ], + "type": "text", + "content": "[23] Christian Henning, Maria Cervera, Francesco D'Angelo, Johannes Von Oswald, Regina Traber, Benjamin Ehret, Seijin Kobayashi, Benjamin F Grewe, and Joao Sacramento. Posterior meta-replay for continual learning. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 580, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 547, + 612 + ], + "type": "text", + "content": "[24] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Intertask association in continual learning. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 547, + 656 + ], + "type": "text", + "content": "[25] Chris Dongjoo Kim, Jinseo Jeong, Sangwoo Moon, and Gunhee Kim. Continual learning on noisy data streams via self-purified replay. International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "text", + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24529" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "text", + "content": "networks. Proceedings of the national academy of sciences, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 288, + 140 + ], + "type": "text", + "content": "[27] Yajing Kong, Liu Liu, Zhen Wang, and Dacheng Tao. Balancing stability and plasticity through advanced null space in continual learning. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 288, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 288, + 165 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 288, + 165 + ], + "type": "text", + "content": "[28] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 167, + 288, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 167, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 167, + 288, + 210 + ], + "type": "text", + "content": "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Conference on Empirical Methods in Natural Language Processing, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "type": "text", + "content": "[30] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. TRGP: Trust region gradient projection for continual learning. In International Conference on Learning Representations, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 258, + 288, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 258, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 288, + 292 + ], + "type": "text", + "content": "[31] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in Neural Information Processing Systems, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 293, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 348 + ], + "type": "text", + "content": "[32] Amrutha Machireddy, Ranganath Krishnan, Nilesh Ahuja, and Omesh Tickoo. Continual active adaptation to evolving distributional shifts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3443-3449, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 404 + ], + "type": "text", + "content": "[33] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. Proceedings of the International Conference on Learning Representations, 2019. 7, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 407, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 288, + 450 + ], + "type": "text", + "content": "[34] Cuong V. Nguyen, Yingzhen Li, Thang D. Bui, and Richard E. Turner. Variational continual learning. Proceedings of the International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 453, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 288, + 507 + ], + "type": "text", + "content": "[35] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. International Conference on Learning Representations, 2019. 2, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 576 + ], + "type": "text", + "content": "[36] Tim G. J. Rudner, Freddie Bickford Smith, Qixuan Feng, Yee Whye Teh, and Yarin Gal. Continual learning via sequential function-space variational inference. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 18871-18887, 17-23 Jul 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "type": "text", + "content": "[37] Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. https://arxiv.org/abs/1606.04671, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 623, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 288, + 656 + ], + "type": "text", + "content": "[38] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[39] Nikunj Saunshi, Arushi Gupta, and Wei Hu. A representation learning perspective on the importance of train-validation splitting in meta-learning. In Proceedings of the 38th International Conference on Machine Learning, pages 9333–9343, 2021. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[40] Jonathan Schwarz, Jelena Luketina, Wojciech M. Czarnecki, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress and compress: A scalable framework for continual learning. In Proceedings of the International Conference on Machine Learning, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 130, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 547, + 174 + ], + "type": "text", + "content": "[41] Joan Serrá, Dídac Surís, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In Proceedings of the International Conference on Machine Learning, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 176, + 547, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 547, + 209 + ], + "type": "text", + "content": "[42] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 212, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 233 + ], + "type": "text", + "content": "[43] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning. 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 236, + 545, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 236, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 545, + 279 + ], + "type": "text", + "content": "[44] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. Advances in neural information processing systems, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 281, + 547, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 281, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 547, + 335 + ], + "type": "text", + "content": "[45] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing HONG, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 547, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 547, + 393 + ], + "type": "text", + "content": "[46] Zhenyi Wang, Tiehang Duan, Le Fang, Qiuling Suo, and Mingchen Gao. Meta learning on a sequence of imbalanced domains with difficulty awareness. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8947-8957, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 396, + 547, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 547, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 547, + 438 + ], + "type": "text", + "content": "[47] Zhen Wang, Liu Liu, Yajing Kong, Jiaxian Guo, and Dacheng Tao. Online continual learning with contrastive vision transformer. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 441, + 547, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 547, + 496 + ], + "type": "text", + "content": "[48] Zhenyi Wang, Li Shen, Tiehang Duan, Donglin Zhan, Le Fang, and Mingchen Gao. Learning to learn and remember super long multi-domain task sequence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7992, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 498, + 547, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 547, + 552 + ], + "type": "text", + "content": "[49] Zhenyi Wang, Li Shen, Le Fang, Qiuling Suo, Tiehang Duan, and Mingchen Gao. Improving task-free continual learning by distributionally robust memory evolution. In International Conference on Machine Learning, pages 22985-22998, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 555, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 547, + 609 + ], + "type": "text", + "content": "[50] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 612, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 547, + 667 + ], + "type": "text", + "content": "[51] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[52] Mengqi Xue, Haofei Zhang, Jie Song, and Mingli Song. Meta-attention for vit-backed continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 150–159, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24530" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 306 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[53] Qingsen Yan, Dong Gong, Yuhang Liu, Anton van den Hengel, and Javen Qinfeng Shi. Learning bayesian sparse networks with full experience replay for continual learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 109-118, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[54] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "text", + "content": "[55] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. IEEE International Conference on Computer Vision, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 262 + ], + "type": "text", + "content": "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. https://arxiv.org/abs/1703.04200, 2017.2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "type": "text", + "content": "[57] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. International Conference on Learning Representations, 2018. 3, 12" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24531" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_content_list.json b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f7165e24d0e8a540e60369964b7865d61f5ac13d --- /dev/null +++ b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_content_list.json @@ -0,0 +1,1827 @@ +[ + { + "type": "text", + "text": "MetaPortrait: Identity-Preserving Talking Head Generation with Fast Personalized Adaptation", + "text_level": 1, + "bbox": [ + 181, + 161, + 789, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bowen Zhang $^{1*}$", + "bbox": [ + 122, + 226, + 250, + 245 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenyang Qi $^{2*}$", + "bbox": [ + 292, + 228, + 411, + 245 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pan Zhang1", + "bbox": [ + 453, + 228, + 547, + 245 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo Zhang $^{3\\dagger}$", + "bbox": [ + 589, + 228, + 685, + 245 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "HsiangTao $\\mathbf{W}\\mathbf{u}^{3}$", + "bbox": [ + 717, + 228, + 845, + 245 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dong Chen", + "bbox": [ + 238, + 247, + 339, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qifeng Chen $^{2\\dagger}$", + "bbox": [ + 381, + 245, + 497, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yong Wang1", + "bbox": [ + 534, + 245, + 633, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fang Wen3", + "bbox": [ + 673, + 245, + 764, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1USTC", + "bbox": [ + 329, + 266, + 391, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ HKUST", + "bbox": [ + 430, + 266, + 507, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Microsoft", + "bbox": [ + 547, + 266, + 636, + 281 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b75b5dbbd8923e19d6f83ac11e6b60850c10440cb43ed6e5b098a9c4ff78437.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 97, + 289, + 250, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b51786e4d3264e6ac6dc6dd885488b1fe601d8b032ff89400658c9754bb5092a.jpg", + "image_caption": [ + "Source", + "Figure 1. Our method yields identity-preserving talking head generation. See the webpage for video demos." + ], + "image_footnote": [], + "bbox": [ + 97, + 397, + 248, + 505 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/967a942b302c45d756cde2491f8b48e7a90c6e2478801b6c8f31abc9059cba7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 252, + 289, + 403, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40be958db55379b27dcd67e9adb7be7974fe1148722f4e68302d185d21e57c0f.jpg", + "image_caption": [ + "Driving" + ], + "image_footnote": [], + "bbox": [ + 252, + 397, + 403, + 505 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51516888e33fd5eeb1f6d40d56507332f88176cf6f5f5369e8611f5f67437184.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 289, + 559, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/714b603906da0ed8f43d498389d4ae21c1a6eb46d3db74291ef245fef4fd4c98.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 406, + 397, + 559, + 505 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec661b1f362f159a77b94ce3fa3f5b04f8f5b5565f109a2cc33334ef9fcc7020.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 561, + 289, + 712, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61df584a2845a9df2d0819114f19bcf60ae94b68c82e7ef1e9f86942d5fbae7f.jpg", + "image_caption": [ + "Driving" + ], + "image_footnote": [], + "bbox": [ + 561, + 397, + 712, + 505 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37c874edcae1660da2469f966120204f2864f0e12cc9d25feaac669c1c0a91a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 289, + 868, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5450721974ec78c7e90723e0ee463eeed0a38ee2bb3343ef3b0fcf7416feac88.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 717, + 397, + 868, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 552, + 312, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose an ID-preserving talking head generation framework, which advances previous methods in two aspects. First, as opposed to interpolating from sparse flow, we claim that dense landmarks are crucial to achieving accurate geometry-aware flow fields. Second, inspired by face-swapping methods, we adaptively fuse the source identity during synthesis, so that the network better preserves the key characteristics of the image portrait. Although the proposed model surpasses prior generation fidelity on established benchmarks, personalized fine-tuning is still needed to further make the talking head generation qualified for real usage. However, this process is rather computationally demanding that is unaffordable to standard users. To alleviate this, we propose a fast adaptation model using a meta-learning approach. The learned model can be adapted to a high-quality personalized model as fast as 30 seconds. Last but not least, a spatial-temporal enhancement module is proposed to improve the fine details while ensuring temporal coherency. Extensive experiments", + "bbox": [ + 75, + 570, + 472, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prove the significant superiority of our approach over the state of the arts in both one-shot and personalized settings.", + "bbox": [ + 499, + 554, + 890, + 581 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 604, + 630, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Talking head generation [1,6,7,24,27,31,33,39,41,47] has found extensive applications in face-to-face live chat, virtual reality and virtual avatars in games and videos. In this paper, we aim to synthesize a realistic talking head with a single source image (one-shot) that provides the appearance of a given person while being animatable according to the motion of the driving person. Recently, considerable progress has been made with neural rendering techniques, bypassing the sophisticated 3D human modeling process and expensive driving sensors. While these works attain increasing fidelity and higher rendering resolution, identity preserving remains a challenging issue since the human vision system is particularly sensitive to any nuanced deviation from the person's facial geometry.", + "bbox": [ + 497, + 629, + 890, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior arts mainly focus on learning a geometry-aware warping field, either by interpolating from sparse 2D/3D landmarks or leveraging 3D face prior, e.g., 3D morphable", + "bbox": [ + 497, + 825, + 890, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 84 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 86 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, interns at Microsoft Research.", + "bbox": [ + 95, + 842, + 364, + 853 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Joint corresponding authors.", + "bbox": [ + 95, + 853, + 252, + 865 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22096", + "bbox": [ + 463, + 885, + 505, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "face model (3DMM) [2, 3]. However, fine-grained facial geometry may not be well described by a set of sparse landmarks or inaccurate face reconstruction. Indeed, the warping field, trained in a self-supervised manner rather than using accurate flow ground truth, can only model coarse geometry deformation, lacking the expressivity that captures the subtle semantic characteristics of the portrait.", + "bbox": [ + 75, + 125, + 468, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose to better preserve the portrait identity in two ways. First, we claim that dense facial landmarks are sufficient for an accurate warping field prediction without the need for local affine transformation. Specifically, we adopt a landmark prediction model [43] trained on synthetic data [42], yielding 669 head landmarks that offer significantly richer information on facial geometry. In addition, we build upon the face-swapping approach [23] and propose to enhance the perceptual identity by attentionally fusing the identity feature of the source portrait while retaining the pose and expression of the intermediate warping. Equipped with these two improvements, our one-shot model demonstrates a significant advantage over prior works in terms of both image quality and perceptual identity preservation when animating in-the-wild portraits.", + "bbox": [ + 75, + 224, + 468, + 432 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While our one-shot talking head model has achieved state-of-the-art quality, it is still infeasible to guarantee satisfactory synthesis results because such a one-shot setting is inherently ill-posed—one may never hallucinate the person-specific facial shape and occluded content from a single photo. Hence, ultimately we encounter the uncanny valley [32] that a user becomes uncomfortable as the synthesis results approach to realism. To circumvent this, one workaround is to finetune the model using several minutes of a personal video. Such personalized training has been widely adopted in industry to ensure product-level quality, yet this process is computationally expensive, which greatly limits its use scenarios. Thus, speeding up this personalized training, a task previously under-explored, is of great significance to the application of talking head synthesis.", + "bbox": [ + 75, + 433, + 468, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose to achieve fast personalization with meta-learning. The key idea is to find an initialization model that can be easily adapted to a given identity with limited training iterations. To this end, we resort to a meta-learning approach [9, 26] that finds success in quickly learning discriminative tasks, yet is rarely explored in generative tasks. Specifically, we optimize the model for specific personal data with a few iterations. In this way, we get a slightly fine-tuned personal model towards which we move the initialization model weight a little bit. Such meta-learned initialization allows us to train a personal model within 30 seconds, which is 3 times faster than a vanilla pretrained model while requiring less amount of personal data.", + "bbox": [ + 75, + 643, + 468, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we propose a novel temporal super-resolution network to enhance the resolution of the generated talking head video. To do this, we leverage the generative prior", + "bbox": [ + 75, + 824, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to boost the high-frequency details for portraits and meanwhile take into account adjacent frames that are helpful to reduce temporal flickering. Finally, we reach temporally coherent video results of $512 \\times 512$ resolution with compelling facial details. In summary, this work innovates in the following aspects:", + "bbox": [ + 497, + 125, + 890, + 207 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a carefully designed framework to significantly improve the identity-preserving capability when animating a one-shot in-the-wild portrait.", + "- To the best of our knowledge, we are the first to explore meta-learning to accelerate personalized training, thus obtaining ultra-high-quality results at affordable cost.", + "- Our novel video super-resolution model effectively enhances details without introducing temporal flickering." + ], + "bbox": [ + 499, + 209, + 890, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 331, + 640, + 346 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D-based talking head synthesis. Methods along this line [24, 28, 33, 34, 39] predict explicit warping flow by interpolating the sparse flow defined by 2D landmarks. FOMM [33] assumes local affine transformation for flow interpolation. Recently, Mallya et al. [24] computes landmarks from multiple source images using an attention mechanism. However, the landmarks learned in an unsupervised manner are too sparse (e.g., 20 in FaceVid2Vid [39]) to be interpolated into dense flows. Using predefined facial landmarks [12] is also a straightforward approach [13, 36, 48, 53] to represent the motion of driving images. For example, PFLD [12] uses 98 facial landmarks. However, they could not generate an accurate warping flow since the landmarks are not dense enough.", + "bbox": [ + 497, + 353, + 890, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Talking head synthesis with 3D face prior. 3D Morphable Models [2,3] represent a face image as PCA coefficients relating to identity, expression, and pose, which provides an easy tool to edit and render portrait images [10, 11, 19, 20, 35, 44]. Some attempts [19, 20] render the animated faces by combining the identity coefficients of the source image and the motion coefficients of the driving video. Recent works [6, 31] predict a warping flow field for talking head synthesis using 3DMM. Although 3DMM-based methods allow explicit face control, using these coefficients to represent detailed face geometry and expression remains challenging.", + "bbox": [ + 497, + 548, + 890, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards higher-resolution talking head. The video resolution of most talking head methods [1, 31, 33, 41] is $256 \\times 256$ , bounded by the available video datasets [4, 25]. To enhance the visual quality of output videos, previous literature trains an additional single-image super-resolution network [7, 40, 45] or utilizes pretrained 2D StyleGAN [40, 47]. However, these methods upsample the original video in a frame-by-frame manner and ignore the temporal consistency of adjacent video frames. In this work, we propose a novel temporal super-resolution module to boost temporal consistency while preserving per-frame quality.", + "bbox": [ + 497, + 714, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22097", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/98c9f11cfa12e4f1c0d56579a6f358518e3722052592fbf2d612250bcc4d9161.jpg", + "image_caption": [ + "Figure 2. Overview of our one-shot framework. (a) Given a source image $\\mathbf{I}_s$ and a driving video $\\{\\mathbf{I}_d^1, \\mathbf{I}_d^2, \\dots, \\mathbf{I}_d^t\\}$ , we first extract their dense landmarks $(\\mathbf{I}_s^{\\mathrm{ldmk}}, \\mathbf{I}_d^{\\mathrm{ldmk}})$ using a pretrained landmark detector. (b) Then, we estimate warping flows $\\boldsymbol{w}$ between the source image and each driving frame according to concatenated input $\\boldsymbol{x}_{\\mathrm{in}}$ . (c) We further refine the warped source input $\\tilde{\\mathbf{I}}_s$ using an ID-preserving network. (d) Finally, we enhance and upsample the $256 \\times 256$ results $\\{\\tilde{I}_y^1, \\tilde{I}_y^2, \\dots, \\tilde{I}_y^t\\}$ to high-fidelity output $\\{\\mathbf{I}'_y^1, \\mathbf{I}'_y^2, \\dots, \\mathbf{I}'_y^t\\}$ in $512 \\times 512$ ." + ], + "image_footnote": [], + "bbox": [ + 97, + 125, + 868, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta learning and fast adaptation. The goal of meta-learning methods [9, 26, 38, 49] is to achieve good performance with a limited amount of training data. In this paper, we aim to quickly build a personalized model given a new identity, which correlates with the goal of meta-learning. Zakharov et al. [49] propose to improve GAN inversion using a hyper-network. We leverage the idea of model-agnostic meta-learning (MAML) [9] to obtain the best initialization that can be easily adapted to different persons.", + "bbox": [ + 75, + 446, + 472, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 75, + 584, + 169, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 illustrates an overview of our synthesis framework. Given an image $I_{s}$ of a person which we refer as source image and a sequence of $t$ driving video frames $\\{I_d^1, I_d^2, \\dots, I_d^t\\}$ , we aim to generate an output video $\\{I_y^1, I_y^2, \\dots, I_y^t\\}$ with the motions derived from the driving video while maintaining the identity of the source image. Section 3.1 introduces our one-shot model (Figure 2(a, b, c)) for identity-preserving talking head generation $\\tilde{I}_y$ at $256 \\times 256$ resolution. We describe our meta-learning scheme in Section 3.2, which allows fast adaption using a few images. In Section 3.3, we propose a spatial-temporal enhancement network that generally improves the perceptual quality for both the one-shot and personalized models, yielding video frames with $512 \\times 512$ resolution, as shown in Figure 2(d).", + "bbox": [ + 75, + 608, + 468, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. ID-preserving One-shot Base Model", + "text_level": 1, + "bbox": [ + 75, + 815, + 389, + 831 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will introduce our warping network using dense facial landmarks and an identity-aware refine-", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ment network for one-shot talking head synthesis.", + "bbox": [ + 497, + 446, + 830, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Warping prediction with dense landmarks. To predict an accurate geometry-aware warping field, we claim that dense landmark prediction [43] is the key to a geometry-aware warping field estimation. While dense facial landmarks are tedious to annotate, our dense landmark prediction is trained on synthetic faces [42], which reliably produces 669 points covering the entire head, including the ears, eyeballs, and teeth, for in-the-wild faces. These dense landmarks capture rich information about the person's facial geometry and considerably ease the flow field prediction.", + "bbox": [ + 497, + 461, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, it is non-trivial to fully make use of these dense landmarks. A naive approach is to channel-wise concatenate the landmarks before feeding into the network, as previous works [33, 34, 39]. However, processing such input is computationally demanding due to the inordinate number of input channels. Hence, we propose an efficient way to digest these landmarks. Specifically, We draw the neighboring connected landmark points, with each connection encoded in different colors as shown in Figure 2(a).", + "bbox": [ + 497, + 600, + 890, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One can thus take the landmark images of the source and driving, along with the source image, i.e., $\\boldsymbol{x}_{\\mathrm{in}} = \\mathrm{Concat}(I_s, I_s^{\\mathrm{ldmk}}, I_d^{\\mathrm{ldmk}})$ , for warping field prediction. To ensure a globally coherent prediction, we strengthen the warping capability using the condition of a latent motion code $\\boldsymbol{z}_w$ which is derived from the input, i.e., $\\boldsymbol{z}_w = E_w(\\boldsymbol{x}_{\\mathrm{in}})$ , where $E_w$ is a CNN encoder. The motion code $\\boldsymbol{z}_w$ is injected into the flow estimation network $F_w$ through AdaIN [15]. By modulating the mean and variance of the normalized feature map, the network could be effectively", + "bbox": [ + 497, + 726, + 892, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22098", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guided by the motion vector which induces a globally coherent flow prediction. Formally, we obtain the prediction of a dense flow field through $\\boldsymbol{w} = F_{w}(\\boldsymbol{x}_{\\mathrm{in}},\\boldsymbol{z}_{w})$", + "bbox": [ + 75, + 125, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ID-preserving refinement network. Directly warping the source image with the predicted flow field inevitably introduces artifacts and the loss of subtle perceived identity. Therefore, an ID-preserving refinement network is needed to produce a photo-realistic result while maintaining the identity of the source image. Prior works primarily focus on the geometry-aware flow field prediction, whereas such deformation may not well characterize the fine-grained facial geometry. In this work, we resolve the identity loss via a well-designed identity-preserving refinement network.", + "bbox": [ + 75, + 167, + 468, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose to attentionally incorporate the semantic identity vector with the intermediate warping results. Let $h_{\\mathrm{in}}^l$ be the $l$ -th layer feature map of the refinement network. We obtain the identity-aware feature output $h_{\\mathrm{id}}^l$ by modulating $h_{\\mathrm{in}}^l$ with the identity embedding $z_{\\mathrm{id}}$ through AdaIN, where $z_{\\mathrm{id}}$ is extracted using a pre-trained face recognition model $E_{\\mathrm{id}}$ [5]. Meanwhile, we obtain a motion-aware feature $h_{\\mathrm{motion}}^l$ , which keeps the head motion and expression of the driving video. Specifically, $h_{\\mathrm{motion}}^l$ is obtained via a Feature-wise Linear Modulate (FiLM) [8, 29, 30] according to the warped image $\\tilde{I}_s = w(I_s)$ , i.e., $h_{\\mathrm{motion}}^l = \\operatorname{Conv}(\\tilde{I}_s) \\times h_{\\mathrm{id}}^l + \\operatorname{Conv}(\\tilde{I}_s)$ .", + "bbox": [ + 75, + 306, + 468, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With both the identity-aware and motion-aware features, we adaptively fuse the features through an attention-based fusion block. Inspired by recent face-swapping approaches [23], we suppose that the driving motions and source identity should be fused in a spatially-adaptive manner, in which the facial parts that mostly characterize the key facial features express the identity should rely more on the identity-aware feature, whereas other parts (e.g. hair and clothes) should make greater use of the motion-aware feature. A learnable fusion mask $M^l$ is used for the fusion of these two parts, which is predicted by,", + "bbox": [ + 75, + 473, + 468, + 624 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} ^ {l} = \\sigma \\left(\\operatorname {C o n v} \\left(\\boldsymbol {h} _ {\\text {i n}} ^ {l}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 631, + 468, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma$ indicates the sigmoid activation function. In this way, the model learns to properly inject the identity-aware features into identity-related regions. The output of layer $l$ can be derived by fusing features according to the mask $M^l$ , which is,", + "bbox": [ + 75, + 655, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} _ {\\text {o u t}} ^ {l} = \\boldsymbol {M} ^ {l} \\otimes \\boldsymbol {h} _ {\\text {m o t i o n}} ^ {l} + (1 - \\boldsymbol {M} ^ {l}) \\otimes \\boldsymbol {h} _ {\\text {i d}} ^ {l}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 134, + 730, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes the Hadamard product. Through a cascade of such blocks, we obtain the final output image $\\tilde{I}_y$ , which well preserves the source identity while accurately following the head motion and expression as the driving person.", + "bbox": [ + 75, + 755, + 468, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training objective. Perceptual loss [16] is computed between the warped source image $\\tilde{I}_s$ and ground-truth driving image $I_d$ for accurate warping prediction. The same loss is also applied to enforce the refinement output $\\tilde{I}_y$ .", + "bbox": [ + 75, + 811, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3b11bbe1f78cb753bc77a4b297b679fbe6ab85449c6118d7a0cbd3c93aaef216.jpg", + "image_caption": [ + "Figure 3. Visualization of meta-learning. Compared with the one-shot pre-trained model $\\phi$ , the meta-learned model $\\hat{\\phi}$ can be rapidly adapted to a unique personal model $\\hat{\\phi}_j^K$ in $K$ steps." + ], + "image_footnote": [], + "bbox": [ + 541, + 124, + 852, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also extract the feature using face recognition model $E_{\\mathrm{id}}$ [5], and penalize the dissimilarity between the ID vectors of the output image $\\tilde{I}_y$ and source $I_s$ , using", + "bbox": [ + 499, + 372, + 890, + 414 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i d}} = 1 - \\cos \\left(E _ {\\mathrm {i d}} \\left(\\tilde {I} _ {y}\\right), E _ {\\mathrm {i d}} \\left(I _ {s}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 427, + 890, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A multi-scale patch discriminator $\\mathcal{L}_{\\mathrm{adv}}$ [29] is adopted to enhance the photo-realism of the outputs. To further improve the generation quality on the hard eye and mouth areas, we add additional $\\mathcal{L}_1$ reconstruction losses, i.e., $\\mathcal{L}_{\\mathrm{eye}}, \\mathcal{L}_{\\mathrm{mouth}}$ for these parts.", + "bbox": [ + 497, + 454, + 890, + 524 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall training loss can be formulated as", + "bbox": [ + 517, + 526, + 821, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {w}} ^ {\\mathrm {V G G}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}} ^ {\\mathrm {V G G}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}} + \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 559, + 549, + 890, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {\\text {e y e}} \\mathcal {L} _ {\\text {e y e}} + \\lambda_ {\\text {m o u t h}} \\mathcal {L} _ {\\text {m o u t h}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}},\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 571, + 826, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{r}}, \\lambda_{\\mathrm{id}}, \\lambda_{\\mathrm{eye}}, \\lambda_{\\mathrm{mouth}}$ and $\\lambda_{\\mathrm{adv}}$ are the loss weights.", + "bbox": [ + 499, + 597, + 865, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Meta-learning based Faster Personalization", + "text_level": 1, + "bbox": [ + 499, + 622, + 868, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While achieving state-of-the-art generation quality using our one-shot model, there always exists challenging cases that are intractable since the one-shot setting is inherently ill-posed. Person-specific characteristics and occlusion would never be faithfully recovered using a one-shot general model. Thus, personalized fine-tuning is necessary to achieve robust and authentic results that are truly usable. Nonetheless, fine-tuning the one-shot pre-trained model on a long video is computationally prohibitive for common users. To solve this, we propose a meta-learned model whose initialization weights can be easily adapted according to low-shot personal data within a few training steps, as illustrated in Figure 3.", + "bbox": [ + 497, + 644, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given a person $j$ , the personalized adaptation starts from the pre-trained model weight $\\phi$ and aims to reach the optimal personal model weight $\\hat{\\phi}_j$ by minimiz-", + "bbox": [ + 497, + 825, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22099", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing the error against the personal images $\\hat{X}_j$ :", + "bbox": [ + 77, + 125, + 379, + 140 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\phi} _ {j} = \\min _ {\\phi_ {j}} \\mathcal {L} \\left(G _ {\\phi_ {j}} \\left(\\hat {\\boldsymbol {X}} _ {j}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 144, + 468, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G_{\\phi_j}$ denotes the whole generator with weight $\\phi_j$ . Usually we perform $K$ steps of stochastic gradient descent (SGD) from initialization $\\phi$ to approach $\\hat{\\phi}_j$ , so the weight updating process can be formulated as:", + "bbox": [ + 75, + 173, + 468, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {j} ^ {k} = \\operatorname {S G D} _ {K} (\\phi , \\hat {\\boldsymbol {X}} _ {j}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 233, + 468, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to find an optimal initialization $\\hat{\\phi}^K$ which could approach any personal model after $K$ steps of SGD update, even for a small $K$ , i.e.,", + "bbox": [ + 75, + 256, + 468, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\phi} ^ {K} = \\min _ {\\phi} \\sum_ {j = 1} ^ {M} \\| \\hat {\\phi} _ {j} - \\phi_ {j} ^ {k} \\|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 302, + 468, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Indeed, the general one-shot pretrained model $\\phi$ is a special case in the above formulation, which essentially learns the model weight in Equation 7 when $K = 0$ . When we are allowed to perform a few adaption steps ( $K > 0$ ), there is a gap between $\\phi$ and desired $\\hat{\\phi}^K$ , since the optimization target of the standard pre-training is to minimize the overall error across all training data, it does not necessarily find the best weight suitable for personalization.", + "bbox": [ + 75, + 347, + 468, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with general one-shot models, we leverage the idea of Model-Agnostic Meta-Learning (MAML) to bridge this gap and enable surprisingly fast personalized training. The goal of MAML-based methods is to optimize the initialized weights such that they could be fast adapted to a new identity within a few steps of gradient descent, which directly matches our goal. Directly optimizing the initialization weight using Equation (7) involves the computation of second-order derivatives, which is computationally expensive on large-scale training. Therefore, we utilize Reptile [26], a first-order MAML-based approach to obtain suitable initialization for fast personalization.", + "bbox": [ + 75, + 457, + 468, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To be more specific, our meta-learning model explicitly considers the personalized adaptation during training. For each person $j$ , we start from the $\\phi_j^0 = \\phi$ , which is the initialization to be optimized. Formally, we sample a batch of personal training data $\\hat{X}_j$ , the $K$ steps of personalized training yield the finetuned model weights as:", + "bbox": [ + 75, + 622, + 468, + 706 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {j} ^ {k} = \\operatorname {S G D} \\left(\\phi_ {j} ^ {k - 1}, \\hat {\\boldsymbol {X}} _ {j}\\right), k = 1, \\dots , K. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 711, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the personal update, i.e., the difference of $\\phi_j^K$ and $\\phi_j^0$ , is used as the gradient to update our initialization $\\phi$ :", + "bbox": [ + 75, + 738, + 468, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\leftarrow \\phi - \\beta \\left(\\phi_ {j} ^ {K} - \\phi\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 775, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is the meta-learning rate. The full algorithm is shown in Algorithm 1. Our model progressively learns a more suitable initialization through meta-training as visualized in Figure 3, which could fast adapt to personalized models after limited steps of adaptation.", + "bbox": [ + 75, + 796, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Optimization of Initial Weights with Reptile", + "text_level": 1, + "bbox": [ + 500, + 125, + 882, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: weights $\\phi$ of generation network $G$ , inner loop learning rate $\\alpha$ , meta-learning rate $\\beta$ , number of training iterations $N$ , number of training persons $M$ , number of inner loop iterations $K$", + "bbox": [ + 522, + 142, + 890, + 192 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $i = 1, \\dots, N$ do", + "2: for $j = 1,\\dots ,M$ do", + "3: $\\phi_j^0 = \\mathbf{Clone}(\\phi)$", + "4: Sample a training batch $\\hat{X}_j$ of person $j$", + "5: for $k = 1,\\dots ,K$ do", + "6: $\\phi_j^k = \\phi_j^{k - 1} - \\alpha \\nabla_\\phi \\mathcal{L}(G_\\phi (\\hat{\\mathbf{X}}_j))$", + "7: end for", + "8: end for", + "9: $\\phi \\gets \\phi -\\frac{\\beta}{M}\\sum_{j = 1}^{M}(\\phi_j^K -\\phi)$", + "10: end for" + ], + "bbox": [ + 502, + 194, + 805, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Temporal-consistent Super-resolution Network", + "text_level": 1, + "bbox": [ + 499, + 350, + 890, + 365 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further enhance the generation resolution and improve the high-fidelity details of our output, a video super-resolution network is needed as the final stage of the generation framework. Previous talking head synthesis works [7,47] utilize single-frame super-resolution as the last stage and ignore the quality of temporal consistency and stability. Performing super-resolution in frame-by-frame manner tends to produce texture flickering which severely hampers the visual quality. In this work, we consider multiple adjacent frames to ensure temporal coherency.", + "bbox": [ + 497, + 372, + 890, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by previous 2D face restoration works [40, 46], the pre-trained generative models [17, 18, 50] like StyleGAN contain rich face prior and could significantly help to enhance the high-frequency details. Moreover, the disentangled $\\mathcal{W}$ space in StyleGAN provides desirable temporal consistency during manipulation [37], which also benefits our framework. Therefore, we propose a temporally consistent super-resolution module by leveraging pretrained StyleGAN and 3D convolution, where the latter brings quality enhancement in spatio-temporal domain. As shown in Fig. 2, we feed the concatenated sequence of $t$ video frames $\\{\\tilde{I}_y^1,\\tilde{I}_y^2,\\dots,\\tilde{I}_y^t\\}$ into a U-Net composed of 3D convolution with reflection padding on the temporal dimension. To ensure pre-trained per-frame quality while improving temporal consistency, we initialize the 3D convolution weight in U-Net with a pretrained 2D face restoration network [40].", + "bbox": [ + 497, + 511, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These spatio-temporally enhanced features from the U-Net decoder further modulate the pretrained StyleGAN features through FiLM. Thus, the super-resolution frames $\\{I_y^{1}, I_y^{2}, \\dots, I_y^{t}\\}$ are obtained as:", + "bbox": [ + 497, + 735, + 890, + 793 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {I} _ {y} ^ {\\prime 1}, \\boldsymbol {I} _ {y} ^ {\\prime 2}, \\dots , \\boldsymbol {I} _ {y} ^ {\\prime t} \\right\\} = F _ {\\text {S t y l e G A N}} \\left(F _ {\\text {3 D}} \\left(\\left\\{\\tilde {\\boldsymbol {I}} _ {y} ^ {1}, \\tilde {\\boldsymbol {I}} _ {y} ^ {2}, \\dots , \\tilde {\\boldsymbol {I}} _ {y} ^ {t} \\right\\}\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 515, + 805, + 890, + 836 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we optimize the output $I_{y}^{\\prime}$ towards the $512 \\times 512$ ground truth $I_{d}$ using $\\ell_{1}$ and perceptual loss.", + "bbox": [ + 499, + 838, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22100", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/94cc2c767e91afe306b209cc76e33a36b8d607f85e27bcdedf820a6cdcdf1e5b.jpg", + "image_caption": [ + "Figure 4. Qualitative results of our self-reconstruction (top row) and cross-identity reenactment (bottom two rows) at $256 \\times 256$ resolution. Our method synthesizes more faithful expression and motion, while better preserving the identity of the source portrait." + ], + "image_footnote": [], + "bbox": [ + 92, + 122, + 877, + 377 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5540f3c19d16e01b241b0346ed41d2accd1944429397f622ab27bd9fd5e87cb6.jpg", + "image_caption": [ + "Figure 5. Qualitative results of self reconstruction and cross-identity reenactment at $512 \\times 512$ resolution. Our spatial-temporal superresolution module further enhances more high-frequency details on teeth, eyes, and hair." + ], + "image_footnote": [], + "bbox": [ + 87, + 416, + 878, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 75, + 710, + 208, + 725 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 75, + 733, + 253, + 749 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Following [7], we train our warping and refinement networks on cropped VoxCeleb2 dataset [4] at $256^2$ resolution. We randomly select 500 videos from the test set for evaluation. For our meta-learning-based fast personalization, we finetune our base model on HDTF dataset [52], which is composed of 410 videos from 300 different identities. We downsample cropped faces to $256^2$ resolution and split the original HDTF dataset into 400 training videos and", + "bbox": [ + 75, + 755, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "10 test videos. After the convergence of our meta-training, we further evaluate the personalization speed of our model on the HDTF test set. Our temporal super-resolution module $F_{3d}$ is trained on the HDTF dataset [52] which has 300 frames per video with $512^2$ resolution. We feed downsampled $256^2$ frames into fixed warping and refinement network and use the outputs of the refinement network as inputs for the next temporal super-resolution module. More training details are provided in the supplementary.", + "bbox": [ + 497, + 711, + 890, + 836 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics We evaluate the fidelity of self-reconstruction using FID [14] and LPIPS [51]. Our motion transfer qual", + "bbox": [ + 497, + 838, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22101", + "bbox": [ + 463, + 885, + 502, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d275d5bc1da0f88561e454ea5f9fc7d5ca56b9d60453ee8c358c4cfd72eed30d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSelf Reconstruction (256 × 256)Cross Reenactment (256 × 256)
FID↓LPIPS↓ID Loss↓AED↓APD↓FID↓ID Loss↓AED↓APD↓
X2Face [41]45.29080.68060.96320.21470.100791.14850.64960.31120.1210
Bi-layer [48]100.91960.58810.52800.12580.0139127.78230.63360.23300.0208
FOMM [33]12.19790.23380.20960.09640.010080.16370.57600.23400.0239
PIRender [31]14.40650.26390.30240.10800.016278.84300.54400.21130.0214
Ours11.95280.22620.12960.09420.012477.50480.29440.25240.0258
MethodsSelf Reconstruction (512 × 512)Cross Reenactment (512 × 512)
StyleHEAT [47]44.52070.28400.41120.11550.0131111.34500.47200.25050.0218
Ours21.49740.20790.08320.09040.012149.60200.19520.27370.0242
", + "bbox": [ + 149, + 123, + 823, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results for self-reconstruction and cross-reenactment. We evaluate both $256 \\times 256$ results and $512 \\times 512$ results. Our method outperforms all baselines on both resolutions across image fidelity metrics with comparable motion transfer results.", + "bbox": [ + 75, + 277, + 892, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab90219bc083f4c5b1c8491790e86148f1237ab2322c7b00d45a794b519c3f84.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 78, + 315, + 238, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f2433757bdb30aafc52f44dbf9cc74d9336e69f8405b94732f681aaa23e413b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 242, + 315, + 401, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9a23ba10ea9dbf8347ee038be83ad616905e906646aacb308d65c56707282b2.jpg", + "image_caption": [ + "Driving" + ], + "image_footnote": [], + "bbox": [ + 242, + 432, + 401, + 543 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7da9b6081c1637b0b43d8e2463e80b3bc9331bce43c22bcffcf59377273974e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 315, + 566, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e35a835034b52e24561264ffb79341570e3cd5f52dd063440e8fb5021a30db9.jpg", + "image_caption": [ + "FOMM", + "Figure 6. A comparison of different personalized models at the same epochs. Our personalized base model without reptile adaptation reconstructs more accurate skin color and source identity than the personalized FOMM method. For the model personalized from a reptile learning stage, finer detail on teeth and eye colors could be further generated." + ], + "image_footnote": [], + "bbox": [ + 406, + 432, + 566, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a648b25a52e7f060dd462def80a6abe88931461ca07db5507de4a2cd059c511.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 315, + 729, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a359a42399421bb22e3dbe8135384ce6ac15e41ea08414e9e31aa4f50298d507.jpg", + "image_caption": [ + "Ours w/o Reptile" + ], + "image_footnote": [], + "bbox": [ + 568, + 432, + 729, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c35b06cb76ccf71e9e40a3aab9b90d3fc2f87fc0735b80b9e1b780062237e55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 315, + 892, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28e1446c407015a50c511bd3a9560809daa94b73ae5c70356e3cf95cb9e5128f.jpg", + "image_caption": [ + "Ours w/ Reptile" + ], + "image_footnote": [], + "bbox": [ + 732, + 432, + 892, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ity is measured using average expression distance (AED) and average pose distance (APD) with driving videos. For our meta-learning-based fast personalization, we illustrate the LPIPS at different personalization epochs for each fine-tuning approach. Following previous works [21, 22], we evaluate our temporal consistency using warping error $E_{\\mathrm{warp}}$ . For each frame $\\tilde{I}_y^t$ , we calculate the warping error with the previous frame $\\tilde{I}_y^{t - 1}$ warped by the estimated optical flow in the occlusion map.", + "bbox": [ + 75, + 625, + 468, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 75, + 761, + 440, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our warping and refinement models at $256 \\times 256$ resolution against several state-of-the-art face reenactment works: X2Face [41], Bi-Layer [1], First-Order Motion Model [33], and PIRender [31]. The top row of Figure 4 presents our qualitative results of self-reconstruction. Different from sparse landmarks in unsupervised learning [33]", + "bbox": [ + 75, + 783, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "or 1D 3DMM [31], our dense landmarks provide strong guidance to accurate and fine-detailed synthesis on gaze, mouth, and expressions. The last two rows show the results of our cross-identity reenactment. Since our landmarks have a better decomposition of identity and motion and our refinement network is identity-aware, our method is the only one that well preserves the identity of source image. In contrast, previous methods suffer from appearance data leakage directly from the driver and generate faces with a similar identity to the driving image. Note that the 3DMM coefficients of AED and APD evaluation are not fully decomposed. Thus, other baselines with identity leakage may have better quantitative metrics. We compare our full framework with the proposed temporal super-resolution module at $512 \\times 512$ resolution against StyleHEAT [47], which is the only open-sourced method that generates high-resolution talking heads. StyleHEAT [47] fails to synthesize", + "bbox": [ + 497, + 625, + 890, + 862 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22102", + "bbox": [ + 463, + 885, + 505, + 897 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c660bb192bb6d895f0f779e58464f663b5ef9a7cc30632aa6d6d0445af0f80e4.jpg", + "image_caption": [ + "Figure 7. Comparison of test LPIPS at fine-tuning epochs for different approaches. Our reptile-based model achieves more than $3 \\times$ speedup compared with the base model and previous baseline." + ], + "image_footnote": [], + "bbox": [ + 121, + 122, + 426, + 298 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sharp and accurate teeth in our experiment, and the identity of the output image is quite different from the source portrait due to the limitations of GAN inversion. In contrast, our refinement network is identity-aware and we leverage pretrained StyleGAN in our temporal super-resolution module to fully exploit its face prior knowledge. Figure 5 shows that our method produces sharp teeth and hairs, while bicubic-upsampled results are blurry with artifacts.", + "bbox": [ + 75, + 365, + 467, + 475 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 demonstrates that our method achieves the best quantitative fidelity and comparable motion transfer quality on both self-reconstruction and cross-identity reenactment.", + "bbox": [ + 75, + 477, + 467, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Evaluation of Fast Personalization", + "text_level": 1, + "bbox": [ + 75, + 524, + 376, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our base model achieves state-of-the-art quality as a general model, there still exists ill-posedness in some cases. For example, it is very difficult to synthesize teeth according to a closed source mouth. Thus, industry products typically conduct personalization using fine-tuning strategy, which can be computationally expensive. To achieve faster convergence, we finetune our base model using a meta-learning strategy to provide a better weight initialization for the following personalization. In Fig 7, we evaluate the personalization speed of our metalearned model against our base model and previous baseline FOMM [33]. It takes our meta-learning model 0.5 epoch to decrease LPIPS to 0.14, which is $3 \\times$ speedup against our base model, and $4 \\times$ against FOMM. Figure 6 compares the personalization of our method and FOMM [33] at the same epoch, which illustrates our fast adaptation speed on ambiguous areas (e.g., teeth, eyes, and wrinkle details).", + "bbox": [ + 75, + 547, + 467, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Evaluation of Temporal Super-resolution", + "text_level": 1, + "bbox": [ + 75, + 789, + 426, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 2, we also evaluate the performance of our temporal super-resolution using 2D image fidelity and warping error $E_{\\text{warp}}$ . We train a 2D super-resolution baseline using GFPGAN [40]. The quantitative result in Table 2 shows", + "bbox": [ + 75, + 809, + 467, + 866 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/48868c82f0208d8377c7b8a342433cb05c0eca439004424b97ac0f7ce162fba0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsFID↓LPIPS↓E_warp ↓
Ground Truth--0.0182
Base w/ bicubic25.57620.22850.0184
Base w/ GFPGAN [40]22.63510.21780.0242
Ours21.49740.20790.0213
", + "bbox": [ + 526, + 123, + 867, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that although naive 2D super-resolution improves per-frame fidelity, it also brings more flickering and larger warping error (0.0242) than simple bicubic upsampling (0.0184). To achieve temporally coherent results, we combine a U-Net composed of 3D convolution with facial prior [18], which significantly reduces $E_{\\text{warp}}$ of our final videos from 0.0242 to 0.0213, and preserves compelling 2D facial details.", + "bbox": [ + 497, + 256, + 890, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study of Base Model", + "text_level": 1, + "bbox": [ + 499, + 362, + 764, + 377 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/473363f226a2139d7b95e0e99c0820a3b3b446b81d96b21a2fc2becc17548230.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of our temporal super-resolution on self-reconstruction at $512 \\times 512$ resolution." + ], + "table_footnote": [], + "table_body": "
MethodsFID↓LPIPS↓ID Loss↓
Ours Sparse Landmark14.31900.24850.1424
Ours w/o ID12.27360.22560.2144
Ours11.95280.22620.1296
", + "bbox": [ + 526, + 396, + 867, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Quantitative ablation study of landmarks and ID coefficients in our base model.", + "bbox": [ + 499, + 461, + 889, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies to validate the effectiveness of our driving motion and source identity representation in our base model. If we replace our 669 dense landmarks with sparse landmarks, the LPIPS of warped source images degrades by 0.2. To evaluate our identity-aware refinement, the removal of the identity input causes significant increase of identity loss from 0.1296 to 0.2144.", + "bbox": [ + 497, + 502, + 890, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 612, + 616, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel framework for identity-preserving one-shot talking head generation. To faithfully maintain the source ID, we propose to leverage accurate dense landmarks in the warping network and explicit source identity during refinement. Further, we significantly advance the applicability of personalized model by reducing its training to 30 seconds with meta-learning. Last but not least, we enhance the final resolution and temporal consistency with 3D convolution and generative prior. Comprehensive experiments demonstrate the state-of-the-art performance of our system.", + "bbox": [ + 497, + 635, + 890, + 774 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 786, + 596, + 800 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 1, 2, 7", + "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. international conference on computer graphics and interactive techniques, 1999. 2" + ], + "bbox": [ + 505, + 808, + 892, + 865 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22103", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponni-ahy, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In CVPR, 2016. 2", + "[4] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. conference of the international speech communication association, 2018. 2, 6", + "[5] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In CVPR, 2020. 4", + "[6] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In ICCV, 2021. 1, 2", + "[7] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In ACM Multimedia, 2022. 1, 2, 5, 6", + "[8] Vincent Dumoulin, Ethan Perez, Nathan Schucher, Florian Strub, Harm de Vries, Aaron Courville, and Yoshua Bengio. Feature-wise transformations. Distill, 2018. https://distill.pub/2018/features-wise-transformations.4", + "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2, 3", + "[10] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Trans. Graph., 2019. 2", + "[11] Jiahao Geng, Tianjia Shao, Youyi Zheng, Yanlin Weng, and Kun Zhou. Warp-guided gans for single-photo facial animation. ACM Transactions on Graphics (TOG), 2018. 2", + "[12] Xiaojie Guo, Siyuan Li, Jiawan Zhang, Jiayi Ma, Lin Ma, Wei Liu, and Haibin Ling. Pfld: A practical facial landmark detector. arXiv: Computer Vision and Pattern Recognition, 2019. 2", + "[13] Sungjoo Ha, Martin Kersner, Beomsu Kim, Seokjun Seo, and Dongyoung Kim. Marionette: Few-shot face reenactment preserving identity of unseen targets. In AAAI, 2020. 2", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6", + "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 3", + "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 4", + "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 5", + "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 5, 8", + "[19] Hyeongwoo Kim, Mohamed Elgharib, Hans-Peter Zollofer, Michael Seidel, Thabo Beeler, Christian Richardt, and Christian Theobalt. Neural style-preserving visual dubbing. ACM Transactions on Graphics (TOG), 2019. 2", + "[20] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Nießner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. Association for Computing Machinery, 2018. 2", + "[21] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 7", + "[22] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. In NeurIPS, 2020. 7", + "[23] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv: Computer Vision and Pattern Recognition, 2019. 2, 4" + ], + "bbox": [ + 78, + 127, + 468, + 864 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit Warping for Animation with Image Sets. In NeurIPS, 2022. 1, 2", + "[25] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: A large-scale speaker identification dataset. conference of the international speech communication association, 2017. 2", + "[26] Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv: Learning, 2018. 2, 3, 5", + "[27] Hao Ouyang, Bo Zhang, Pan Zhang, Hao Yang, Jiaolong Yang, Dong Chen, Qifeng Chen, and Fang Wen. Real-time neural character rendering with pose-guided multiplane images. ECCV, 2022. 1", + "[28] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. arXiv preprint arXiv:2301.06281, 2023. 2", + "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 4", + "[30] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In AAAI, 2018. 4", + "[31] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In ICCV, 2021. 1, 2, 7", + "[32] Mincheol Shin, Se Jung Kim, and Frank Biocca. The uncanny valley: No need for any further judgments when an avatar looks eerie. Computers in Human Behavior, 94:100-109, 2019. 2", + "[33] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In NeurIPS, 2019. 1, 2, 3, 7, 8", + "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In CVPR, 2019. 2, 3", + "[35] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 2", + "[36] Soumya Tripathy, Juho Kannala, and Esa Rahtu. Facegan: Facial attribute controllable reenactment gan. In CVPR, 2021. 2", + "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit H. Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial editing of real videos*, 2022. 5", + "[38] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaevator: Learning animatable clothed human models from few depth images. In NeurIPS, 2021. 3", + "[39] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, 2020. 1, 2, 3", + "[40] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In CVPR, 2021. 2, 5, 8", + "[41] Olivia Wiles, A Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018. 1, 2, 7", + "[42] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Matthew Johnson, Virginia Estellers, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone. In ICCV, 2021. 2, 3", + "[43] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Chirag Raman, Jamie Shotton, Toby Sharp, Ivan Stojiljkovic, Tom Cashman, and Julien Valentin. 3d face reconstruction with dense landmarks. In ECCV, 2022. 2, 3", + "[44] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. arXiv preprint arXiv:2301.02379, 2023. 2" + ], + "bbox": [ + 500, + 127, + 890, + 864 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22104", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Lingbo Yang, Chang Liu, Pan Wang, Shanshe Wang, Peiran Ren, Siwei Ma, and Wen Gao. Hifacegan: Face renovation via collaborative suppression and replenishment. ACM Multimedia, 2020. 2", + "[46] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 5", + "[47] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. In ECCV, 2022. 1, 2, 5, 7", + "[48] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In ECCV, 2020. 2, 7", + "[49] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In ICCV, 2019. 3", + "[50] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In CVPR, 2022. 5", + "[51] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[52] Zhimeng Zhang, Lincheng Li, and Yu Ding. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In CVPR, 2021. 6", + "[53] Ruiqi Zhao, Tianyi Wu, and Guodong Guo. Sparse to dense motion transfer for face image animation. In ICCV, 2021. 2" + ], + "bbox": [ + 78, + 127, + 468, + 432 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22105", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_model.json b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3895e0ba209992ddd96b5264d97dc41ba880ff0a --- /dev/null +++ b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_model.json @@ -0,0 +1,2585 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.085 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.792, + 0.087 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.162, + 0.79, + 0.204 + ], + "angle": 0, + "content": "MetaPortrait: Identity-Preserving Talking Head Generation with Fast Personalized Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.228, + 0.252, + 0.246 + ], + "angle": 0, + "content": "Bowen Zhang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.229, + 0.412, + 0.246 + ], + "angle": 0, + "content": "Chenyang Qi\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.454, + 0.229, + 0.548, + 0.246 + ], + "angle": 0, + "content": "Pan Zhang1" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.229, + 0.686, + 0.246 + ], + "angle": 0, + "content": "Bo Zhang\\(^{3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.719, + 0.229, + 0.846, + 0.246 + ], + "angle": 0, + "content": "HsiangTao \\(\\mathbf{W}\\mathbf{u}^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.248, + 0.34, + 0.264 + ], + "angle": 0, + "content": "Dong Chen" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.247, + 0.498, + 0.264 + ], + "angle": 0, + "content": "Qifeng Chen\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.247, + 0.634, + 0.264 + ], + "angle": 0, + "content": "Yong Wang1" + }, + { + "type": "text", + "bbox": [ + 0.675, + 0.247, + 0.766, + 0.264 + ], + "angle": 0, + "content": "Fang Wen3" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.267, + 0.392, + 0.282 + ], + "angle": 0, + "content": "1USTC" + }, + { + "type": "text", + "bbox": [ + 0.431, + 0.267, + 0.508, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\)HKUST" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.267, + 0.637, + 0.282 + ], + "angle": 0, + "content": "3Microsoft" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.29, + 0.251, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.399, + 0.25, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.507, + 0.195, + 0.518 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.29, + 0.405, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.398, + 0.405, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.507, + 0.352, + 0.518 + ], + "angle": 0, + "content": "Driving" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.29, + 0.561, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.399, + 0.56, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.507, + 0.5, + 0.518 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.29, + 0.714, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.399, + 0.714, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.507, + 0.665, + 0.518 + ], + "angle": 0, + "content": "Driving" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.29, + 0.869, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.399, + 0.869, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.507, + 0.807, + 0.517 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.523, + 0.802, + 0.537 + ], + "angle": 0, + "content": "Figure 1. Our method yields identity-preserving talking head generation. See the webpage for video demos." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.553, + 0.314, + 0.567 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.473, + 0.834 + ], + "angle": 0, + "content": "In this work, we propose an ID-preserving talking head generation framework, which advances previous methods in two aspects. First, as opposed to interpolating from sparse flow, we claim that dense landmarks are crucial to achieving accurate geometry-aware flow fields. Second, inspired by face-swapping methods, we adaptively fuse the source identity during synthesis, so that the network better preserves the key characteristics of the image portrait. Although the proposed model surpasses prior generation fidelity on established benchmarks, personalized fine-tuning is still needed to further make the talking head generation qualified for real usage. However, this process is rather computationally demanding that is unaffordable to standard users. To alleviate this, we propose a fast adaptation model using a meta-learning approach. The learned model can be adapted to a high-quality personalized model as fast as 30 seconds. Last but not least, a spatial-temporal enhancement module is proposed to improve the fine details while ensuring temporal coherency. Extensive experiments" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.555, + 0.892, + 0.583 + ], + "angle": 0, + "content": "prove the significant superiority of our approach over the state of the arts in both one-shot and personalized settings." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.605, + 0.631, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Talking head generation [1,6,7,24,27,31,33,39,41,47] has found extensive applications in face-to-face live chat, virtual reality and virtual avatars in games and videos. In this paper, we aim to synthesize a realistic talking head with a single source image (one-shot) that provides the appearance of a given person while being animatable according to the motion of the driving person. Recently, considerable progress has been made with neural rendering techniques, bypassing the sophisticated 3D human modeling process and expensive driving sensors. While these works attain increasing fidelity and higher rendering resolution, identity preserving remains a challenging issue since the human vision system is particularly sensitive to any nuanced deviation from the person's facial geometry." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Prior arts mainly focus on learning a geometry-aware warping field, either by interpolating from sparse 2D/3D landmarks or leveraging 3D face prior, e.g., 3D morphable" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.843, + 0.366, + 0.855 + ], + "angle": 0, + "content": "*Equal contribution, interns at Microsoft Research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.855, + 0.253, + 0.866 + ], + "angle": 0, + "content": "† Joint corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.843, + 0.366, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "22096" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.47, + 0.224 + ], + "angle": 0, + "content": "face model (3DMM) [2, 3]. However, fine-grained facial geometry may not be well described by a set of sparse landmarks or inaccurate face reconstruction. Indeed, the warping field, trained in a self-supervised manner rather than using accurate flow ground truth, can only model coarse geometry deformation, lacking the expressivity that captures the subtle semantic characteristics of the portrait." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.225, + 0.47, + 0.433 + ], + "angle": 0, + "content": "In this paper, we propose to better preserve the portrait identity in two ways. First, we claim that dense facial landmarks are sufficient for an accurate warping field prediction without the need for local affine transformation. Specifically, we adopt a landmark prediction model [43] trained on synthetic data [42], yielding 669 head landmarks that offer significantly richer information on facial geometry. In addition, we build upon the face-swapping approach [23] and propose to enhance the perceptual identity by attentionally fusing the identity feature of the source portrait while retaining the pose and expression of the intermediate warping. Equipped with these two improvements, our one-shot model demonstrates a significant advantage over prior works in terms of both image quality and perceptual identity preservation when animating in-the-wild portraits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.47, + 0.642 + ], + "angle": 0, + "content": "While our one-shot talking head model has achieved state-of-the-art quality, it is still infeasible to guarantee satisfactory synthesis results because such a one-shot setting is inherently ill-posed—one may never hallucinate the person-specific facial shape and occluded content from a single photo. Hence, ultimately we encounter the uncanny valley [32] that a user becomes uncomfortable as the synthesis results approach to realism. To circumvent this, one workaround is to finetune the model using several minutes of a personal video. Such personalized training has been widely adopted in industry to ensure product-level quality, yet this process is computationally expensive, which greatly limits its use scenarios. Thus, speeding up this personalized training, a task previously under-explored, is of great significance to the application of talking head synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.824 + ], + "angle": 0, + "content": "We propose to achieve fast personalization with meta-learning. The key idea is to find an initialization model that can be easily adapted to a given identity with limited training iterations. To this end, we resort to a meta-learning approach [9, 26] that finds success in quickly learning discriminative tasks, yet is rarely explored in generative tasks. Specifically, we optimize the model for specific personal data with a few iterations. In this way, we get a slightly fine-tuned personal model towards which we move the initialization model weight a little bit. Such meta-learned initialization allows us to train a personal model within 30 seconds, which is 3 times faster than a vanilla pretrained model while requiring less amount of personal data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Moreover, we propose a novel temporal super-resolution network to enhance the resolution of the generated talking head video. To do this, we leverage the generative prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.892, + 0.209 + ], + "angle": 0, + "content": "to boost the high-frequency details for portraits and meanwhile take into account adjacent frames that are helpful to reduce temporal flickering. Finally, we reach temporally coherent video results of \\(512 \\times 512\\) resolution with compelling facial details. In summary, this work innovates in the following aspects:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.21, + 0.891, + 0.251 + ], + "angle": 0, + "content": "- We propose a carefully designed framework to significantly improve the identity-preserving capability when animating a one-shot in-the-wild portrait." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.251, + 0.891, + 0.292 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to explore meta-learning to accelerate personalized training, thus obtaining ultra-high-quality results at affordable cost." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.292, + 0.891, + 0.32 + ], + "angle": 0, + "content": "- Our novel video super-resolution model effectively enhances details without introducing temporal flickering." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.21, + 0.891, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.642, + 0.347 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.548 + ], + "angle": 0, + "content": "2D-based talking head synthesis. Methods along this line [24, 28, 33, 34, 39] predict explicit warping flow by interpolating the sparse flow defined by 2D landmarks. FOMM [33] assumes local affine transformation for flow interpolation. Recently, Mallya et al. [24] computes landmarks from multiple source images using an attention mechanism. However, the landmarks learned in an unsupervised manner are too sparse (e.g., 20 in FaceVid2Vid [39]) to be interpolated into dense flows. Using predefined facial landmarks [12] is also a straightforward approach [13, 36, 48, 53] to represent the motion of driving images. For example, PFLD [12] uses 98 facial landmarks. However, they could not generate an accurate warping flow since the landmarks are not dense enough." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Talking head synthesis with 3D face prior. 3D Morphable Models [2,3] represent a face image as PCA coefficients relating to identity, expression, and pose, which provides an easy tool to edit and render portrait images [10, 11, 19, 20, 35, 44]. Some attempts [19, 20] render the animated faces by combining the identity coefficients of the source image and the motion coefficients of the driving video. Recent works [6, 31] predict a warping flow field for talking head synthesis using 3DMM. Although 3DMM-based methods allow explicit face control, using these coefficients to represent detailed face geometry and expression remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Towards higher-resolution talking head. The video resolution of most talking head methods [1, 31, 33, 41] is \\(256 \\times 256\\), bounded by the available video datasets [4, 25]. To enhance the visual quality of output videos, previous literature trains an additional single-image super-resolution network [7, 40, 45] or utilizes pretrained 2D StyleGAN [40, 47]. However, these methods upsample the original video in a frame-by-frame manner and ignore the temporal consistency of adjacent video frames. In this work, we propose a novel temporal super-resolution module to boost temporal consistency while preserving per-frame quality." + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "22097" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.126, + 0.87, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.387, + 0.893, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Overview of our one-shot framework. (a) Given a source image \\( \\mathbf{I}_s \\) and a driving video \\( \\{\\mathbf{I}_d^1, \\mathbf{I}_d^2, \\dots, \\mathbf{I}_d^t\\} \\), we first extract their dense landmarks \\( (\\mathbf{I}_s^{\\mathrm{ldmk}}, \\mathbf{I}_d^{\\mathrm{ldmk}}) \\) using a pretrained landmark detector. (b) Then, we estimate warping flows \\( \\boldsymbol{w} \\) between the source image and each driving frame according to concatenated input \\( \\boldsymbol{x}_{\\mathrm{in}} \\). (c) We further refine the warped source input \\( \\tilde{\\mathbf{I}}_s \\) using an ID-preserving network. (d) Finally, we enhance and upsample the \\( 256 \\times 256 \\) results \\( \\{\\tilde{I}_y^1, \\tilde{I}_y^2, \\dots, \\tilde{I}_y^t\\} \\) to high-fidelity output \\( \\{\\mathbf{I}'_y^1, \\mathbf{I}'_y^2, \\dots, \\mathbf{I}'_y^t\\} \\) in \\( 512 \\times 512 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.473, + 0.571 + ], + "angle": 0, + "content": "Meta learning and fast adaptation. The goal of meta-learning methods [9, 26, 38, 49] is to achieve good performance with a limited amount of training data. In this paper, we aim to quickly build a personalized model given a new identity, which correlates with the goal of meta-learning. Zakharov et al. [49] propose to improve GAN inversion using a hyper-network. We leverage the idea of model-agnostic meta-learning (MAML) [9] to obtain the best initialization that can be easily adapted to different persons." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.585, + 0.17, + 0.6 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Figure 2 illustrates an overview of our synthesis framework. Given an image \\(I_{s}\\) of a person which we refer as source image and a sequence of \\(t\\) driving video frames \\(\\{I_d^1, I_d^2, \\dots, I_d^t\\}\\), we aim to generate an output video \\(\\{I_y^1, I_y^2, \\dots, I_y^t\\}\\) with the motions derived from the driving video while maintaining the identity of the source image. Section 3.1 introduces our one-shot model (Figure 2(a, b, c)) for identity-preserving talking head generation \\(\\tilde{I}_y\\) at \\(256 \\times 256\\) resolution. We describe our meta-learning scheme in Section 3.2, which allows fast adaption using a few images. In Section 3.3, we propose a spatial-temporal enhancement network that generally improves the perceptual quality for both the one-shot and personalized models, yielding video frames with \\(512 \\times 512\\) resolution, as shown in Figure 2(d)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.391, + 0.832 + ], + "angle": 0, + "content": "3.1. ID-preserving One-shot Base Model" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.867 + ], + "angle": 0, + "content": "In this section, we will introduce our warping network using dense facial landmarks and an identity-aware refine-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.831, + 0.461 + ], + "angle": 0, + "content": "ment network for one-shot talking head synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.6 + ], + "angle": 0, + "content": "Warping prediction with dense landmarks. To predict an accurate geometry-aware warping field, we claim that dense landmark prediction [43] is the key to a geometry-aware warping field estimation. While dense facial landmarks are tedious to annotate, our dense landmark prediction is trained on synthetic faces [42], which reliably produces 669 points covering the entire head, including the ears, eyeballs, and teeth, for in-the-wild faces. These dense landmarks capture rich information about the person's facial geometry and considerably ease the flow field prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.727 + ], + "angle": 0, + "content": "However, it is non-trivial to fully make use of these dense landmarks. A naive approach is to channel-wise concatenate the landmarks before feeding into the network, as previous works [33, 34, 39]. However, processing such input is computationally demanding due to the inordinate number of input channels. Hence, we propose an efficient way to digest these landmarks. Specifically, We draw the neighboring connected landmark points, with each connection encoded in different colors as shown in Figure 2(a)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.893, + 0.867 + ], + "angle": 0, + "content": "One can thus take the landmark images of the source and driving, along with the source image, i.e., \\( \\boldsymbol{x}_{\\mathrm{in}} = \\mathrm{Concat}(I_s, I_s^{\\mathrm{ldmk}}, I_d^{\\mathrm{ldmk}}) \\), for warping field prediction. To ensure a globally coherent prediction, we strengthen the warping capability using the condition of a latent motion code \\( \\boldsymbol{z}_w \\) which is derived from the input, i.e., \\( \\boldsymbol{z}_w = E_w(\\boldsymbol{x}_{\\mathrm{in}}) \\), where \\( E_w \\) is a CNN encoder. The motion code \\( \\boldsymbol{z}_w \\) is injected into the flow estimation network \\( F_w \\) through AdaIN [15]. By modulating the mean and variance of the normalized feature map, the network could be effectively" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "22098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.47, + 0.168 + ], + "angle": 0, + "content": "guided by the motion vector which induces a globally coherent flow prediction. Formally, we obtain the prediction of a dense flow field through \\(\\boldsymbol{w} = F_{w}(\\boldsymbol{x}_{\\mathrm{in}},\\boldsymbol{z}_{w})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.469, + 0.306 + ], + "angle": 0, + "content": "ID-preserving refinement network. Directly warping the source image with the predicted flow field inevitably introduces artifacts and the loss of subtle perceived identity. Therefore, an ID-preserving refinement network is needed to produce a photo-realistic result while maintaining the identity of the source image. Prior works primarily focus on the geometry-aware flow field prediction, whereas such deformation may not well characterize the fine-grained facial geometry. In this work, we resolve the identity loss via a well-designed identity-preserving refinement network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.469, + 0.473 + ], + "angle": 0, + "content": "We propose to attentionally incorporate the semantic identity vector with the intermediate warping results. Let \\( h_{\\mathrm{in}}^l \\) be the \\( l \\)-th layer feature map of the refinement network. We obtain the identity-aware feature output \\( h_{\\mathrm{id}}^l \\) by modulating \\( h_{\\mathrm{in}}^l \\) with the identity embedding \\( z_{\\mathrm{id}} \\) through AdaIN, where \\( z_{\\mathrm{id}} \\) is extracted using a pre-trained face recognition model \\( E_{\\mathrm{id}} \\) [5]. Meanwhile, we obtain a motion-aware feature \\( h_{\\mathrm{motion}}^l \\), which keeps the head motion and expression of the driving video. Specifically, \\( h_{\\mathrm{motion}}^l \\) is obtained via a Feature-wise Linear Modulate (FiLM) [8, 29, 30] according to the warped image \\( \\tilde{I}_s = w(I_s) \\), i.e., \\( h_{\\mathrm{motion}}^l = \\operatorname{Conv}(\\tilde{I}_s) \\times h_{\\mathrm{id}}^l + \\operatorname{Conv}(\\tilde{I}_s) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.625 + ], + "angle": 0, + "content": "With both the identity-aware and motion-aware features, we adaptively fuse the features through an attention-based fusion block. Inspired by recent face-swapping approaches [23], we suppose that the driving motions and source identity should be fused in a spatially-adaptive manner, in which the facial parts that mostly characterize the key facial features express the identity should rely more on the identity-aware feature, whereas other parts (e.g. hair and clothes) should make greater use of the motion-aware feature. A learnable fusion mask \\( M^l \\) is used for the fusion of these two parts, which is predicted by," + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.632, + 0.469, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} ^ {l} = \\sigma \\left(\\operatorname {C o n v} \\left(\\boldsymbol {h} _ {\\text {i n}} ^ {l}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.656, + 0.469, + 0.724 + ], + "angle": 0, + "content": "where \\(\\sigma\\) indicates the sigmoid activation function. In this way, the model learns to properly inject the identity-aware features into identity-related regions. The output of layer \\(l\\) can be derived by fusing features according to the mask \\(M^l\\), which is," + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.731, + 0.469, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} _ {\\text {o u t}} ^ {l} = \\boldsymbol {M} ^ {l} \\otimes \\boldsymbol {h} _ {\\text {m o t i o n}} ^ {l} + (1 - \\boldsymbol {M} ^ {l}) \\otimes \\boldsymbol {h} _ {\\text {i d}} ^ {l}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.469, + 0.812 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes the Hadamard product. Through a cascade of such blocks, we obtain the final output image \\(\\tilde{I}_y\\), which well preserves the source identity while accurately following the head motion and expression as the driving person." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Training objective. Perceptual loss [16] is computed between the warped source image \\(\\tilde{I}_s\\) and ground-truth driving image \\(I_d\\) for accurate warping prediction. The same loss is also applied to enforce the refinement output \\(\\tilde{I}_y\\)." + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.125, + 0.853, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.325, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Figure 3. Visualization of meta-learning. Compared with the one-shot pre-trained model \\(\\phi\\), the meta-learned model \\(\\hat{\\phi}\\) can be rapidly adapted to a unique personal model \\(\\hat{\\phi}_j^K\\) in \\(K\\) steps." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.373, + 0.892, + 0.415 + ], + "angle": 0, + "content": "We also extract the feature using face recognition model \\( E_{\\mathrm{id}} \\) [5], and penalize the dissimilarity between the ID vectors of the output image \\( \\tilde{I}_y \\) and source \\( I_s \\), using" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.428, + 0.891, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {i d}} = 1 - \\cos \\left(E _ {\\mathrm {i d}} \\left(\\tilde {I} _ {y}\\right), E _ {\\mathrm {i d}} \\left(I _ {s}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.456, + 0.891, + 0.525 + ], + "angle": 0, + "content": "A multi-scale patch discriminator \\(\\mathcal{L}_{\\mathrm{adv}}\\) [29] is adopted to enhance the photo-realism of the outputs. To further improve the generation quality on the hard eye and mouth areas, we add additional \\(\\mathcal{L}_1\\) reconstruction losses, i.e., \\(\\mathcal{L}_{\\mathrm{eye}}, \\mathcal{L}_{\\mathrm{mouth}}\\) for these parts." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.527, + 0.822, + 0.54 + ], + "angle": 0, + "content": "The overall training loss can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.551, + 0.891, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {w}} ^ {\\mathrm {V G G}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}} ^ {\\mathrm {V G G}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}} + \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.572, + 0.828, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {\\text {e y e}} \\mathcal {L} _ {\\text {e y e}} + \\lambda_ {\\text {m o u t h}} \\mathcal {L} _ {\\text {m o u t h}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.598, + 0.867, + 0.612 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{r}}, \\lambda_{\\mathrm{id}}, \\lambda_{\\mathrm{eye}}, \\lambda_{\\mathrm{mouth}}\\) and \\(\\lambda_{\\mathrm{adv}}\\) are the loss weights." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.623, + 0.869, + 0.637 + ], + "angle": 0, + "content": "3.2. Meta-learning based Faster Personalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.824 + ], + "angle": 0, + "content": "While achieving state-of-the-art generation quality using our one-shot model, there always exists challenging cases that are intractable since the one-shot setting is inherently ill-posed. Person-specific characteristics and occlusion would never be faithfully recovered using a one-shot general model. Thus, personalized fine-tuning is necessary to achieve robust and authentic results that are truly usable. Nonetheless, fine-tuning the one-shot pre-trained model on a long video is computationally prohibitive for common users. To solve this, we propose a meta-learned model whose initialization weights can be easily adapted according to low-shot personal data within a few training steps, as illustrated in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Formally, given a person \\(j\\), the personalized adaptation starts from the pre-trained model weight \\(\\phi\\) and aims to reach the optimal personal model weight \\(\\hat{\\phi}_j\\) by minimiz-" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "22099" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.126, + 0.38, + 0.141 + ], + "angle": 0, + "content": "ing the error against the personal images \\(\\hat{X}_j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.146, + 0.469, + 0.17 + ], + "angle": 0, + "content": "\\[\n\\hat {\\phi} _ {j} = \\min _ {\\phi_ {j}} \\mathcal {L} \\left(G _ {\\phi_ {j}} \\left(\\hat {\\boldsymbol {X}} _ {j}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.174, + 0.469, + 0.23 + ], + "angle": 0, + "content": "where \\(G_{\\phi_j}\\) denotes the whole generator with weight \\(\\phi_j\\). Usually we perform \\(K\\) steps of stochastic gradient descent (SGD) from initialization \\(\\phi\\) to approach \\(\\hat{\\phi}_j\\), so the weight updating process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.234, + 0.469, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\phi_ {j} ^ {k} = \\operatorname {S G D} _ {K} (\\phi , \\hat {\\boldsymbol {X}} _ {j}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.257, + 0.469, + 0.3 + ], + "angle": 0, + "content": "Our goal is to find an optimal initialization \\(\\hat{\\phi}^K\\) which could approach any personal model after \\(K\\) steps of SGD update, even for a small \\(K\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.304, + 0.469, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\hat {\\phi} ^ {K} = \\min _ {\\phi} \\sum_ {j = 1} ^ {M} \\| \\hat {\\phi} _ {j} - \\phi_ {j} ^ {k} \\|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.469, + 0.458 + ], + "angle": 0, + "content": "Indeed, the general one-shot pretrained model \\(\\phi\\) is a special case in the above formulation, which essentially learns the model weight in Equation 7 when \\(K = 0\\). When we are allowed to perform a few adaption steps (\\(K > 0\\)), there is a gap between \\(\\phi\\) and desired \\(\\hat{\\phi}^K\\), since the optimization target of the standard pre-training is to minimize the overall error across all training data, it does not necessarily find the best weight suitable for personalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.469, + 0.623 + ], + "angle": 0, + "content": "Compared with general one-shot models, we leverage the idea of Model-Agnostic Meta-Learning (MAML) to bridge this gap and enable surprisingly fast personalized training. The goal of MAML-based methods is to optimize the initialized weights such that they could be fast adapted to a new identity within a few steps of gradient descent, which directly matches our goal. Directly optimizing the initialization weight using Equation (7) involves the computation of second-order derivatives, which is computationally expensive on large-scale training. Therefore, we utilize Reptile [26], a first-order MAML-based approach to obtain suitable initialization for fast personalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.469, + 0.707 + ], + "angle": 0, + "content": "To be more specific, our meta-learning model explicitly considers the personalized adaptation during training. For each person \\(j\\), we start from the \\(\\phi_j^0 = \\phi\\), which is the initialization to be optimized. Formally, we sample a batch of personal training data \\(\\hat{X}_j\\), the \\(K\\) steps of personalized training yield the finetuned model weights as:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.712, + 0.469, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\phi_ {j} ^ {k} = \\operatorname {S G D} \\left(\\phi_ {j} ^ {k - 1}, \\hat {\\boldsymbol {X}} _ {j}\\right), k = 1, \\dots , K. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.739, + 0.469, + 0.771 + ], + "angle": 0, + "content": "Finally, the personal update, i.e., the difference of \\(\\phi_j^K\\) and \\(\\phi_j^0\\), is used as the gradient to update our initialization \\(\\phi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.776, + 0.469, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\phi \\leftarrow \\phi - \\beta \\left(\\phi_ {j} ^ {K} - \\phi\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.797, + 0.469, + 0.867 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the meta-learning rate. The full algorithm is shown in Algorithm 1. Our model progressively learns a more suitable initialization through meta-training as visualized in Figure 3, which could fast adapt to personalized models after limited steps of adaptation." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.126, + 0.884, + 0.14 + ], + "angle": 0, + "content": "Algorithm 1 Optimization of Initial Weights with Reptile" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.143, + 0.892, + 0.193 + ], + "angle": 0, + "content": "Input: weights \\(\\phi\\) of generation network \\(G\\), inner loop learning rate \\(\\alpha\\), meta-learning rate \\(\\beta\\), number of training iterations \\(N\\), number of training persons \\(M\\), number of inner loop iterations \\(K\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.195, + 0.654, + 0.206 + ], + "angle": 0, + "content": "1: for \\( i = 1, \\dots, N \\) do" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.208, + 0.68, + 0.219 + ], + "angle": 0, + "content": "2: for \\(j = 1,\\dots ,M\\) do" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.22, + 0.669, + 0.233 + ], + "angle": 0, + "content": "3: \\(\\phi_j^0 = \\mathbf{Clone}(\\phi)\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.234, + 0.806, + 0.246 + ], + "angle": 0, + "content": "4: Sample a training batch \\(\\hat{X}_j\\) of person \\(j\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.247, + 0.702, + 0.258 + ], + "angle": 0, + "content": "5: for \\(k = 1,\\dots ,K\\) do" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.258, + 0.792, + 0.273 + ], + "angle": 0, + "content": "6: \\(\\phi_j^k = \\phi_j^{k - 1} - \\alpha \\nabla_\\phi \\mathcal{L}(G_\\phi (\\hat{\\mathbf{X}}_j))\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.273, + 0.619, + 0.283 + ], + "angle": 0, + "content": "7: end for" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.285, + 0.597, + 0.295 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.296, + 0.733, + 0.312 + ], + "angle": 0, + "content": "9: \\(\\phi \\gets \\phi -\\frac{\\beta}{M}\\sum_{j = 1}^{M}(\\phi_j^K -\\phi)\\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.312, + 0.575, + 0.321 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.195, + 0.806, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.351, + 0.891, + 0.366 + ], + "angle": 0, + "content": "3.3. Temporal-consistent Super-resolution Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.512 + ], + "angle": 0, + "content": "To further enhance the generation resolution and improve the high-fidelity details of our output, a video super-resolution network is needed as the final stage of the generation framework. Previous talking head synthesis works [7,47] utilize single-frame super-resolution as the last stage and ignore the quality of temporal consistency and stability. Performing super-resolution in frame-by-frame manner tends to produce texture flickering which severely hampers the visual quality. In this work, we consider multiple adjacent frames to ensure temporal coherency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Inspired by previous 2D face restoration works [40, 46], the pre-trained generative models [17, 18, 50] like StyleGAN contain rich face prior and could significantly help to enhance the high-frequency details. Moreover, the disentangled \\(\\mathcal{W}\\) space in StyleGAN provides desirable temporal consistency during manipulation [37], which also benefits our framework. Therefore, we propose a temporally consistent super-resolution module by leveraging pretrained StyleGAN and 3D convolution, where the latter brings quality enhancement in spatio-temporal domain. As shown in Fig. 2, we feed the concatenated sequence of \\(t\\) video frames \\(\\{\\tilde{I}_y^1,\\tilde{I}_y^2,\\dots,\\tilde{I}_y^t\\}\\) into a U-Net composed of 3D convolution with reflection padding on the temporal dimension. To ensure pre-trained per-frame quality while improving temporal consistency, we initialize the 3D convolution weight in U-Net with a pretrained 2D face restoration network [40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.794 + ], + "angle": 0, + "content": "These spatio-temporally enhanced features from the U-Net decoder further modulate the pretrained StyleGAN features through FiLM. Thus, the super-resolution frames \\(\\{I_y^{1}, I_y^{2}, \\dots, I_y^{t}\\}\\) are obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.806, + 0.891, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {I} _ {y} ^ {\\prime 1}, \\boldsymbol {I} _ {y} ^ {\\prime 2}, \\dots , \\boldsymbol {I} _ {y} ^ {\\prime t} \\right\\} = F _ {\\text {S t y l e G A N}} \\left(F _ {\\text {3 D}} \\left(\\left\\{\\tilde {\\boldsymbol {I}} _ {y} ^ {1}, \\tilde {\\boldsymbol {I}} _ {y} ^ {2}, \\dots , \\tilde {\\boldsymbol {I}} _ {y} ^ {t} \\right\\}\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.867 + ], + "angle": 0, + "content": "During training, we optimize the output \\( I_{y}^{\\prime} \\) towards the \\( 512 \\times 512 \\) ground truth \\( I_{d} \\) using \\( \\ell_{1} \\) and perceptual loss." + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "22100" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.123, + 0.878, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.38, + 0.893, + 0.407 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of our self-reconstruction (top row) and cross-identity reenactment (bottom two rows) at \\(256 \\times 256\\) resolution. Our method synthesizes more faithful expression and motion, while better preserving the identity of the source portrait." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.418, + 0.88, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.662, + 0.893, + 0.689 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of self reconstruction and cross-identity reenactment at \\(512 \\times 512\\) resolution. Our spatial-temporal superresolution module further enhances more high-frequency details on teeth, eyes, and hair." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.21, + 0.726 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.734, + 0.254, + 0.75 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Dataset. Following [7], we train our warping and refinement networks on cropped VoxCeleb2 dataset [4] at \\(256^2\\) resolution. We randomly select 500 videos from the test set for evaluation. For our meta-learning-based fast personalization, we finetune our base model on HDTF dataset [52], which is composed of 410 videos from 300 different identities. We downsample cropped faces to \\(256^2\\) resolution and split the original HDTF dataset into 400 training videos and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.837 + ], + "angle": 0, + "content": "10 test videos. After the convergence of our meta-training, we further evaluate the personalization speed of our model on the HDTF test set. Our temporal super-resolution module \\( F_{3d} \\) is trained on the HDTF dataset [52] which has 300 frames per video with \\( 512^2 \\) resolution. We feed downsampled \\( 256^2 \\) frames into fixed warping and refinement network and use the outputs of the refinement network as inputs for the next temporal super-resolution module. More training details are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Metrics We evaluate the fidelity of self-reconstruction using FID [14] and LPIPS [51]. Our motion transfer qual" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.504, + 0.898 + ], + "angle": 0, + "content": "22101" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.151, + 0.124, + 0.825, + 0.277 + ], + "angle": 0, + "content": "
MethodsSelf Reconstruction (256 × 256)Cross Reenactment (256 × 256)
FID↓LPIPS↓ID Loss↓AED↓APD↓FID↓ID Loss↓AED↓APD↓
X2Face [41]45.29080.68060.96320.21470.100791.14850.64960.31120.1210
Bi-layer [48]100.91960.58810.52800.12580.0139127.78230.63360.23300.0208
FOMM [33]12.19790.23380.20960.09640.010080.16370.57600.23400.0239
PIRender [31]14.40650.26390.30240.10800.016278.84300.54400.21130.0214
Ours11.95280.22620.12960.09420.012477.50480.29440.25240.0258
MethodsSelf Reconstruction (512 × 512)Cross Reenactment (512 × 512)
StyleHEAT [47]44.52070.28400.41120.11550.0131111.34500.47200.25050.0218
Ours21.49740.20790.08320.09040.012149.60200.19520.27370.0242
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.279, + 0.893, + 0.306 + ], + "angle": 0, + "content": "Table 1. Quantitative results for self-reconstruction and cross-reenactment. We evaluate both \\(256 \\times 256\\) results and \\(512 \\times 512\\) results. Our method outperforms all baselines on both resolutions across image fidelity metrics with comparable motion transfer results." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.317, + 0.24, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.549, + 0.179, + 0.559 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.317, + 0.403, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.433, + 0.403, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.548, + 0.338, + 0.559 + ], + "angle": 0, + "content": "Driving" + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.317, + 0.567, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.433, + 0.567, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.549, + 0.5, + 0.559 + ], + "angle": 0, + "content": "FOMM" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.317, + 0.73, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.433, + 0.73, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.549, + 0.693, + 0.56 + ], + "angle": 0, + "content": "Ours w/o Reptile" + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.317, + 0.893, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.433, + 0.893, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.548, + 0.847, + 0.559 + ], + "angle": 0, + "content": "Ours w/ Reptile" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.565, + 0.893, + 0.604 + ], + "angle": 0, + "content": "Figure 6. A comparison of different personalized models at the same epochs. Our personalized base model without reptile adaptation reconstructs more accurate skin color and source identity than the personalized FOMM method. For the model personalized from a reptile learning stage, finer detail on teeth and eye colors could be further generated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.754 + ], + "angle": 0, + "content": "ity is measured using average expression distance (AED) and average pose distance (APD) with driving videos. For our meta-learning-based fast personalization, we illustrate the LPIPS at different personalization epochs for each fine-tuning approach. Following previous works [21, 22], we evaluate our temporal consistency using warping error \\( E_{\\mathrm{warp}} \\). For each frame \\( \\tilde{I}_y^t \\), we calculate the warping error with the previous frame \\( \\tilde{I}_y^{t - 1} \\) warped by the estimated optical flow in the occlusion map." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.442, + 0.777 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.47, + 0.868 + ], + "angle": 0, + "content": "We compare our warping and refinement models at \\(256 \\times 256\\) resolution against several state-of-the-art face reenactment works: X2Face [41], Bi-Layer [1], First-Order Motion Model [33], and PIRender [31]. The top row of Figure 4 presents our qualitative results of self-reconstruction. Different from sparse landmarks in unsupervised learning [33]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.863 + ], + "angle": 0, + "content": "or 1D 3DMM [31], our dense landmarks provide strong guidance to accurate and fine-detailed synthesis on gaze, mouth, and expressions. The last two rows show the results of our cross-identity reenactment. Since our landmarks have a better decomposition of identity and motion and our refinement network is identity-aware, our method is the only one that well preserves the identity of source image. In contrast, previous methods suffer from appearance data leakage directly from the driver and generate faces with a similar identity to the driving image. Note that the 3DMM coefficients of AED and APD evaluation are not fully decomposed. Thus, other baselines with identity leakage may have better quantitative metrics. We compare our full framework with the proposed temporal super-resolution module at \\(512 \\times 512\\) resolution against StyleHEAT [47], which is the only open-sourced method that generates high-resolution talking heads. StyleHEAT [47] fails to synthesize" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "22102" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.123, + 0.428, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.471, + 0.346 + ], + "angle": 0, + "content": "Figure 7. Comparison of test LPIPS at fine-tuning epochs for different approaches. Our reptile-based model achieves more than \\(3 \\times\\) speedup compared with the base model and previous baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.468, + 0.476 + ], + "angle": 0, + "content": "sharp and accurate teeth in our experiment, and the identity of the output image is quite different from the source portrait due to the limitations of GAN inversion. In contrast, our refinement network is identity-aware and we leverage pretrained StyleGAN in our temporal super-resolution module to fully exploit its face prior knowledge. Figure 5 shows that our method produces sharp teeth and hairs, while bicubic-upsampled results are blurry with artifacts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.478, + 0.468, + 0.518 + ], + "angle": 0, + "content": "Table 1 demonstrates that our method achieves the best quantitative fidelity and comparable motion transfer quality on both self-reconstruction and cross-identity reenactment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.377, + 0.54 + ], + "angle": 0, + "content": "4.3. Evaluation of Fast Personalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.468, + 0.782 + ], + "angle": 0, + "content": "Although our base model achieves state-of-the-art quality as a general model, there still exists ill-posedness in some cases. For example, it is very difficult to synthesize teeth according to a closed source mouth. Thus, industry products typically conduct personalization using fine-tuning strategy, which can be computationally expensive. To achieve faster convergence, we finetune our base model using a meta-learning strategy to provide a better weight initialization for the following personalization. In Fig 7, we evaluate the personalization speed of our metalearned model against our base model and previous baseline FOMM [33]. It takes our meta-learning model 0.5 epoch to decrease LPIPS to 0.14, which is \\(3 \\times\\) speedup against our base model, and \\(4 \\times\\) against FOMM. Figure 6 compares the personalization of our method and FOMM [33] at the same epoch, which illustrates our fast adaptation speed on ambiguous areas (e.g., teeth, eyes, and wrinkle details)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.427, + 0.805 + ], + "angle": 0, + "content": "4.4. Evaluation of Temporal Super-resolution" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.468, + 0.867 + ], + "angle": 0, + "content": "In Table 2, we also evaluate the performance of our temporal super-resolution using 2D image fidelity and warping error \\( E_{\\text{warp}} \\). We train a 2D super-resolution baseline using GFPGAN [40]. The quantitative result in Table 2 shows" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.124, + 0.868, + 0.205 + ], + "angle": 0, + "content": "
MethodsFID↓LPIPS↓E_warp ↓
Ground Truth--0.0182
Base w/ bicubic25.57620.22850.0184
Base w/ GFPGAN [40]22.63510.21780.0242
Ours21.49740.20790.0213
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.207, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of our temporal super-resolution on self-reconstruction at \\(512 \\times 512\\) resolution." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.257, + 0.892, + 0.355 + ], + "angle": 0, + "content": "that although naive 2D super-resolution improves per-frame fidelity, it also brings more flickering and larger warping error (0.0242) than simple bicubic upsampling (0.0184). To achieve temporally coherent results, we combine a U-Net composed of 3D convolution with facial prior [18], which significantly reduces \\( E_{\\text{warp}} \\) of our final videos from 0.0242 to 0.0213, and preserves compelling 2D facial details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.363, + 0.766, + 0.378 + ], + "angle": 0, + "content": "4.5. Ablation Study of Base Model" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.397, + 0.868, + 0.459 + ], + "angle": 0, + "content": "
MethodsFID↓LPIPS↓ID Loss↓
Ours Sparse Landmark14.31900.24850.1424
Ours w/o ID12.27360.22560.2144
Ours11.95280.22620.1296
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.462, + 0.89, + 0.487 + ], + "angle": 0, + "content": "Table 3. Quantitative ablation study of landmarks and ID coefficients in our base model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.503, + 0.892, + 0.6 + ], + "angle": 0, + "content": "We conduct ablation studies to validate the effectiveness of our driving motion and source identity representation in our base model. If we replace our 669 dense landmarks with sparse landmarks, the LPIPS of warped source images degrades by 0.2. To evaluate our identity-aware refinement, the removal of the identity input causes significant increase of identity loss from 0.1296 to 0.2144." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.618, + 0.627 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.636, + 0.892, + 0.775 + ], + "angle": 0, + "content": "We present a novel framework for identity-preserving one-shot talking head generation. To faithfully maintain the source ID, we propose to leverage accurate dense landmarks in the warping network and explicit source identity during refinement. Further, we significantly advance the applicability of personalized model by reducing its training to 30 seconds with meta-learning. Last but not least, we enhance the final resolution and temporal consistency with 3D convolution and generative prior. Comprehensive experiments demonstrate the state-of-the-art performance of our system." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.787, + 0.597, + 0.801 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.506, + 0.809, + 0.892, + 0.832 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.833, + 0.894, + 0.866 + ], + "angle": 0, + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. international conference on computer graphics and interactive techniques, 1999. 2" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.809, + 0.894, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "22103" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.128, + 0.469, + 0.161 + ], + "angle": 0, + "content": "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponni-ahy, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.162, + 0.469, + 0.195 + ], + "angle": 0, + "content": "[4] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. conference of the international speech communication association, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.196, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[5] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.229, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[6] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.263, + 0.469, + 0.306 + ], + "angle": 0, + "content": "[7] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In ACM Multimedia, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.307, + 0.469, + 0.351 + ], + "angle": 0, + "content": "[8] Vincent Dumoulin, Ethan Perez, Nathan Schucher, Florian Strub, Harm de Vries, Aaron Courville, and Yoshua Bengio. Feature-wise transformations. Distill, 2018. https://distill.pub/2018/features-wise-transformations.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.352, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[10] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Trans. Graph., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.43, + 0.469, + 0.463 + ], + "angle": 0, + "content": "[11] Jiahao Geng, Tianjia Shao, Youyi Zheng, Yanlin Weng, and Kun Zhou. Warp-guided gans for single-photo facial animation. ACM Transactions on Graphics (TOG), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.469, + 0.497 + ], + "angle": 0, + "content": "[12] Xiaojie Guo, Siyuan Li, Jiawan Zhang, Jiayi Ma, Lin Ma, Wei Liu, and Haibin Ling. Pfld: A practical facial landmark detector. arXiv: Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[13] Sungjoo Ha, Martin Kersner, Beomsu Kim, Seokjun Seo, and Dongyoung Kim. Marionette: Few-shot face reenactment preserving identity of unseen targets. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.72 + ], + "angle": 0, + "content": "[19] Hyeongwoo Kim, Mohamed Elgharib, Hans-Peter Zollofer, Michael Seidel, Thabo Beeler, Christian Richardt, and Christian Theobalt. Neural style-preserving visual dubbing. ACM Transactions on Graphics (TOG), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.765 + ], + "angle": 0, + "content": "[20] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Nießner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. Association for Computing Machinery, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.766, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[21] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.8, + 0.469, + 0.822 + ], + "angle": 0, + "content": "[22] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. In NeurIPS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.822, + 0.469, + 0.865 + ], + "angle": 0, + "content": "[23] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv: Computer Vision and Pattern Recognition, 2019. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.128, + 0.469, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.501, + 0.128, + 0.891, + 0.151 + ], + "angle": 0, + "content": "[24] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit Warping for Animation with Image Sets. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.152, + 0.891, + 0.185 + ], + "angle": 0, + "content": "[25] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: A large-scale speaker identification dataset. conference of the international speech communication association, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.186, + 0.891, + 0.208 + ], + "angle": 0, + "content": "[26] Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv: Learning, 2018. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.891, + 0.242 + ], + "angle": 0, + "content": "[27] Hao Ouyang, Bo Zhang, Pan Zhang, Hao Yang, Jiaolong Yang, Dong Chen, Qifeng Chen, and Fang Wen. Real-time neural character rendering with pose-guided multiplane images. ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.243, + 0.891, + 0.287 + ], + "angle": 0, + "content": "[28] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. arXiv preprint arXiv:2301.06281, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.891, + 0.321 + ], + "angle": 0, + "content": "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.891, + 0.355 + ], + "angle": 0, + "content": "[30] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In AAAI, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.356, + 0.891, + 0.39 + ], + "angle": 0, + "content": "[31] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In ICCV, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.891, + 0.424 + ], + "angle": 0, + "content": "[32] Mincheol Shin, Se Jung Kim, and Frank Biocca. The uncanny valley: No need for any further judgments when an avatar looks eerie. Computers in Human Behavior, 94:100-109, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.425, + 0.891, + 0.458 + ], + "angle": 0, + "content": "[33] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In NeurIPS, 2019. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.891, + 0.492 + ], + "angle": 0, + "content": "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In CVPR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.493, + 0.891, + 0.527 + ], + "angle": 0, + "content": "[35] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.891, + 0.549 + ], + "angle": 0, + "content": "[36] Soumya Tripathy, Juho Kannala, and Esa Rahtu. Facegan: Facial attribute controllable reenactment gan. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.55, + 0.891, + 0.583 + ], + "angle": 0, + "content": "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit H. Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial editing of real videos*, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.891, + 0.618 + ], + "angle": 0, + "content": "[38] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaevator: Learning animatable clothed human models from few depth images. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.891, + 0.651 + ], + "angle": 0, + "content": "[39] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.652, + 0.891, + 0.685 + ], + "angle": 0, + "content": "[40] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In CVPR, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.891, + 0.719 + ], + "angle": 0, + "content": "[41] Olivia Wiles, A Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.891, + 0.765 + ], + "angle": 0, + "content": "[42] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Matthew Johnson, Virginia Estellers, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.766, + 0.891, + 0.82 + ], + "angle": 0, + "content": "[43] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Chirag Raman, Jamie Shotton, Toby Sharp, Ivan Stojiljkovic, Tom Cashman, and Julien Valentin. 3d face reconstruction with dense landmarks. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.822, + 0.891, + 0.865 + ], + "angle": 0, + "content": "[44] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. arXiv preprint arXiv:2301.02379, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.128, + 0.891, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "22104" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.128, + 0.469, + 0.161 + ], + "angle": 0, + "content": "[45] Lingbo Yang, Chang Liu, Pan Wang, Shanshe Wang, Peiran Ren, Siwei Ma, and Wen Gao. Hifacegan: Face renovation via collaborative suppression and replenishment. ACM Multimedia, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.469, + 0.195 + ], + "angle": 0, + "content": "[46] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.196, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[47] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. In ECCV, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.242, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[48] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In ECCV, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.469, + 0.308 + ], + "angle": 0, + "content": "[49] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[50] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[51] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.41 + ], + "angle": 0, + "content": "[52] Zhimeng Zhang, Lincheng Li, and Yu Ding. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.41, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[53] Ruiqi Zhao, Tianyi Wu, and Guodong Guo. Sparse to dense motion transfer for face image animation. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.128, + 0.469, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "22105" + } + ] +] \ No newline at end of file diff --git a/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_origin.pdf b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c9e017530c3dd41ff8dd5372153220e401e97bd --- /dev/null +++ b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/23766ee0-1d0b-472f-8959-8f7e09df80cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8f86b432b31ae86c1eea8238f964c32c4f2041352bfe36ac940d9055a76fa12 +size 9207658 diff --git a/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/full.md b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..23337c7861a58eb7dcb62451318f98da2f0cabc2 --- /dev/null +++ b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/full.md @@ -0,0 +1,368 @@ +# MetaPortrait: Identity-Preserving Talking Head Generation with Fast Personalized Adaptation + +Bowen Zhang $^{1*}$ + +Chenyang Qi $^{2*}$ + +Pan Zhang1 + +Bo Zhang $^{3\dagger}$ + +HsiangTao $\mathbf{W}\mathbf{u}^{3}$ + +Dong Chen + +Qifeng Chen $^{2\dagger}$ + +Yong Wang1 + +Fang Wen3 + +1USTC + +$^{2}$ HKUST + +3Microsoft + +![](images/5b75b5dbbd8923e19d6f83ac11e6b60850c10440cb43ed6e5b098a9c4ff78437.jpg) + +![](images/b51786e4d3264e6ac6dc6dd885488b1fe601d8b032ff89400658c9754bb5092a.jpg) +Source +Figure 1. Our method yields identity-preserving talking head generation. See the webpage for video demos. + +![](images/967a942b302c45d756cde2491f8b48e7a90c6e2478801b6c8f31abc9059cba7a.jpg) + +![](images/40be958db55379b27dcd67e9adb7be7974fe1148722f4e68302d185d21e57c0f.jpg) +Driving + +![](images/51516888e33fd5eeb1f6d40d56507332f88176cf6f5f5369e8611f5f67437184.jpg) + +![](images/714b603906da0ed8f43d498389d4ae21c1a6eb46d3db74291ef245fef4fd4c98.jpg) +Ours + +![](images/ec661b1f362f159a77b94ce3fa3f5b04f8f5b5565f109a2cc33334ef9fcc7020.jpg) + +![](images/61df584a2845a9df2d0819114f19bcf60ae94b68c82e7ef1e9f86942d5fbae7f.jpg) +Driving + +![](images/37c874edcae1660da2469f966120204f2864f0e12cc9d25feaac669c1c0a91a6.jpg) + +![](images/5450721974ec78c7e90723e0ee463eeed0a38ee2bb3343ef3b0fcf7416feac88.jpg) +Ours + +# Abstract + +In this work, we propose an ID-preserving talking head generation framework, which advances previous methods in two aspects. First, as opposed to interpolating from sparse flow, we claim that dense landmarks are crucial to achieving accurate geometry-aware flow fields. Second, inspired by face-swapping methods, we adaptively fuse the source identity during synthesis, so that the network better preserves the key characteristics of the image portrait. Although the proposed model surpasses prior generation fidelity on established benchmarks, personalized fine-tuning is still needed to further make the talking head generation qualified for real usage. However, this process is rather computationally demanding that is unaffordable to standard users. To alleviate this, we propose a fast adaptation model using a meta-learning approach. The learned model can be adapted to a high-quality personalized model as fast as 30 seconds. Last but not least, a spatial-temporal enhancement module is proposed to improve the fine details while ensuring temporal coherency. Extensive experiments + +prove the significant superiority of our approach over the state of the arts in both one-shot and personalized settings. + +# 1. Introduction + +Talking head generation [1,6,7,24,27,31,33,39,41,47] has found extensive applications in face-to-face live chat, virtual reality and virtual avatars in games and videos. In this paper, we aim to synthesize a realistic talking head with a single source image (one-shot) that provides the appearance of a given person while being animatable according to the motion of the driving person. Recently, considerable progress has been made with neural rendering techniques, bypassing the sophisticated 3D human modeling process and expensive driving sensors. While these works attain increasing fidelity and higher rendering resolution, identity preserving remains a challenging issue since the human vision system is particularly sensitive to any nuanced deviation from the person's facial geometry. + +Prior arts mainly focus on learning a geometry-aware warping field, either by interpolating from sparse 2D/3D landmarks or leveraging 3D face prior, e.g., 3D morphable + +face model (3DMM) [2, 3]. However, fine-grained facial geometry may not be well described by a set of sparse landmarks or inaccurate face reconstruction. Indeed, the warping field, trained in a self-supervised manner rather than using accurate flow ground truth, can only model coarse geometry deformation, lacking the expressivity that captures the subtle semantic characteristics of the portrait. + +In this paper, we propose to better preserve the portrait identity in two ways. First, we claim that dense facial landmarks are sufficient for an accurate warping field prediction without the need for local affine transformation. Specifically, we adopt a landmark prediction model [43] trained on synthetic data [42], yielding 669 head landmarks that offer significantly richer information on facial geometry. In addition, we build upon the face-swapping approach [23] and propose to enhance the perceptual identity by attentionally fusing the identity feature of the source portrait while retaining the pose and expression of the intermediate warping. Equipped with these two improvements, our one-shot model demonstrates a significant advantage over prior works in terms of both image quality and perceptual identity preservation when animating in-the-wild portraits. + +While our one-shot talking head model has achieved state-of-the-art quality, it is still infeasible to guarantee satisfactory synthesis results because such a one-shot setting is inherently ill-posed—one may never hallucinate the person-specific facial shape and occluded content from a single photo. Hence, ultimately we encounter the uncanny valley [32] that a user becomes uncomfortable as the synthesis results approach to realism. To circumvent this, one workaround is to finetune the model using several minutes of a personal video. Such personalized training has been widely adopted in industry to ensure product-level quality, yet this process is computationally expensive, which greatly limits its use scenarios. Thus, speeding up this personalized training, a task previously under-explored, is of great significance to the application of talking head synthesis. + +We propose to achieve fast personalization with meta-learning. The key idea is to find an initialization model that can be easily adapted to a given identity with limited training iterations. To this end, we resort to a meta-learning approach [9, 26] that finds success in quickly learning discriminative tasks, yet is rarely explored in generative tasks. Specifically, we optimize the model for specific personal data with a few iterations. In this way, we get a slightly fine-tuned personal model towards which we move the initialization model weight a little bit. Such meta-learned initialization allows us to train a personal model within 30 seconds, which is 3 times faster than a vanilla pretrained model while requiring less amount of personal data. + +Moreover, we propose a novel temporal super-resolution network to enhance the resolution of the generated talking head video. To do this, we leverage the generative prior + +to boost the high-frequency details for portraits and meanwhile take into account adjacent frames that are helpful to reduce temporal flickering. Finally, we reach temporally coherent video results of $512 \times 512$ resolution with compelling facial details. In summary, this work innovates in the following aspects: + +- We propose a carefully designed framework to significantly improve the identity-preserving capability when animating a one-shot in-the-wild portrait. +- To the best of our knowledge, we are the first to explore meta-learning to accelerate personalized training, thus obtaining ultra-high-quality results at affordable cost. +- Our novel video super-resolution model effectively enhances details without introducing temporal flickering. + +# 2. Related Work + +2D-based talking head synthesis. Methods along this line [24, 28, 33, 34, 39] predict explicit warping flow by interpolating the sparse flow defined by 2D landmarks. FOMM [33] assumes local affine transformation for flow interpolation. Recently, Mallya et al. [24] computes landmarks from multiple source images using an attention mechanism. However, the landmarks learned in an unsupervised manner are too sparse (e.g., 20 in FaceVid2Vid [39]) to be interpolated into dense flows. Using predefined facial landmarks [12] is also a straightforward approach [13, 36, 48, 53] to represent the motion of driving images. For example, PFLD [12] uses 98 facial landmarks. However, they could not generate an accurate warping flow since the landmarks are not dense enough. + +Talking head synthesis with 3D face prior. 3D Morphable Models [2,3] represent a face image as PCA coefficients relating to identity, expression, and pose, which provides an easy tool to edit and render portrait images [10, 11, 19, 20, 35, 44]. Some attempts [19, 20] render the animated faces by combining the identity coefficients of the source image and the motion coefficients of the driving video. Recent works [6, 31] predict a warping flow field for talking head synthesis using 3DMM. Although 3DMM-based methods allow explicit face control, using these coefficients to represent detailed face geometry and expression remains challenging. + +Towards higher-resolution talking head. The video resolution of most talking head methods [1, 31, 33, 41] is $256 \times 256$ , bounded by the available video datasets [4, 25]. To enhance the visual quality of output videos, previous literature trains an additional single-image super-resolution network [7, 40, 45] or utilizes pretrained 2D StyleGAN [40, 47]. However, these methods upsample the original video in a frame-by-frame manner and ignore the temporal consistency of adjacent video frames. In this work, we propose a novel temporal super-resolution module to boost temporal consistency while preserving per-frame quality. + +![](images/98c9f11cfa12e4f1c0d56579a6f358518e3722052592fbf2d612250bcc4d9161.jpg) +Figure 2. Overview of our one-shot framework. (a) Given a source image $\mathbf{I}_s$ and a driving video $\{\mathbf{I}_d^1, \mathbf{I}_d^2, \dots, \mathbf{I}_d^t\}$ , we first extract their dense landmarks $(\mathbf{I}_s^{\mathrm{ldmk}}, \mathbf{I}_d^{\mathrm{ldmk}})$ using a pretrained landmark detector. (b) Then, we estimate warping flows $\boldsymbol{w}$ between the source image and each driving frame according to concatenated input $\boldsymbol{x}_{\mathrm{in}}$ . (c) We further refine the warped source input $\tilde{\mathbf{I}}_s$ using an ID-preserving network. (d) Finally, we enhance and upsample the $256 \times 256$ results $\{\tilde{I}_y^1, \tilde{I}_y^2, \dots, \tilde{I}_y^t\}$ to high-fidelity output $\{\mathbf{I}'_y^1, \mathbf{I}'_y^2, \dots, \mathbf{I}'_y^t\}$ in $512 \times 512$ . + +Meta learning and fast adaptation. The goal of meta-learning methods [9, 26, 38, 49] is to achieve good performance with a limited amount of training data. In this paper, we aim to quickly build a personalized model given a new identity, which correlates with the goal of meta-learning. Zakharov et al. [49] propose to improve GAN inversion using a hyper-network. We leverage the idea of model-agnostic meta-learning (MAML) [9] to obtain the best initialization that can be easily adapted to different persons. + +# 3. Method + +Figure 2 illustrates an overview of our synthesis framework. Given an image $I_{s}$ of a person which we refer as source image and a sequence of $t$ driving video frames $\{I_d^1, I_d^2, \dots, I_d^t\}$ , we aim to generate an output video $\{I_y^1, I_y^2, \dots, I_y^t\}$ with the motions derived from the driving video while maintaining the identity of the source image. Section 3.1 introduces our one-shot model (Figure 2(a, b, c)) for identity-preserving talking head generation $\tilde{I}_y$ at $256 \times 256$ resolution. We describe our meta-learning scheme in Section 3.2, which allows fast adaption using a few images. In Section 3.3, we propose a spatial-temporal enhancement network that generally improves the perceptual quality for both the one-shot and personalized models, yielding video frames with $512 \times 512$ resolution, as shown in Figure 2(d). + +# 3.1. ID-preserving One-shot Base Model + +In this section, we will introduce our warping network using dense facial landmarks and an identity-aware refine- + +ment network for one-shot talking head synthesis. + +Warping prediction with dense landmarks. To predict an accurate geometry-aware warping field, we claim that dense landmark prediction [43] is the key to a geometry-aware warping field estimation. While dense facial landmarks are tedious to annotate, our dense landmark prediction is trained on synthetic faces [42], which reliably produces 669 points covering the entire head, including the ears, eyeballs, and teeth, for in-the-wild faces. These dense landmarks capture rich information about the person's facial geometry and considerably ease the flow field prediction. + +However, it is non-trivial to fully make use of these dense landmarks. A naive approach is to channel-wise concatenate the landmarks before feeding into the network, as previous works [33, 34, 39]. However, processing such input is computationally demanding due to the inordinate number of input channels. Hence, we propose an efficient way to digest these landmarks. Specifically, We draw the neighboring connected landmark points, with each connection encoded in different colors as shown in Figure 2(a). + +One can thus take the landmark images of the source and driving, along with the source image, i.e., $\boldsymbol{x}_{\mathrm{in}} = \mathrm{Concat}(I_s, I_s^{\mathrm{ldmk}}, I_d^{\mathrm{ldmk}})$ , for warping field prediction. To ensure a globally coherent prediction, we strengthen the warping capability using the condition of a latent motion code $\boldsymbol{z}_w$ which is derived from the input, i.e., $\boldsymbol{z}_w = E_w(\boldsymbol{x}_{\mathrm{in}})$ , where $E_w$ is a CNN encoder. The motion code $\boldsymbol{z}_w$ is injected into the flow estimation network $F_w$ through AdaIN [15]. By modulating the mean and variance of the normalized feature map, the network could be effectively + +guided by the motion vector which induces a globally coherent flow prediction. Formally, we obtain the prediction of a dense flow field through $\boldsymbol{w} = F_{w}(\boldsymbol{x}_{\mathrm{in}},\boldsymbol{z}_{w})$ + +ID-preserving refinement network. Directly warping the source image with the predicted flow field inevitably introduces artifacts and the loss of subtle perceived identity. Therefore, an ID-preserving refinement network is needed to produce a photo-realistic result while maintaining the identity of the source image. Prior works primarily focus on the geometry-aware flow field prediction, whereas such deformation may not well characterize the fine-grained facial geometry. In this work, we resolve the identity loss via a well-designed identity-preserving refinement network. + +We propose to attentionally incorporate the semantic identity vector with the intermediate warping results. Let $h_{\mathrm{in}}^l$ be the $l$ -th layer feature map of the refinement network. We obtain the identity-aware feature output $h_{\mathrm{id}}^l$ by modulating $h_{\mathrm{in}}^l$ with the identity embedding $z_{\mathrm{id}}$ through AdaIN, where $z_{\mathrm{id}}$ is extracted using a pre-trained face recognition model $E_{\mathrm{id}}$ [5]. Meanwhile, we obtain a motion-aware feature $h_{\mathrm{motion}}^l$ , which keeps the head motion and expression of the driving video. Specifically, $h_{\mathrm{motion}}^l$ is obtained via a Feature-wise Linear Modulate (FiLM) [8, 29, 30] according to the warped image $\tilde{I}_s = w(I_s)$ , i.e., $h_{\mathrm{motion}}^l = \operatorname{Conv}(\tilde{I}_s) \times h_{\mathrm{id}}^l + \operatorname{Conv}(\tilde{I}_s)$ . + +With both the identity-aware and motion-aware features, we adaptively fuse the features through an attention-based fusion block. Inspired by recent face-swapping approaches [23], we suppose that the driving motions and source identity should be fused in a spatially-adaptive manner, in which the facial parts that mostly characterize the key facial features express the identity should rely more on the identity-aware feature, whereas other parts (e.g. hair and clothes) should make greater use of the motion-aware feature. A learnable fusion mask $M^l$ is used for the fusion of these two parts, which is predicted by, + +$$ +\boldsymbol {M} ^ {l} = \sigma \left(\operatorname {C o n v} \left(\boldsymbol {h} _ {\text {i n}} ^ {l}\right)\right), \tag {1} +$$ + +where $\sigma$ indicates the sigmoid activation function. In this way, the model learns to properly inject the identity-aware features into identity-related regions. The output of layer $l$ can be derived by fusing features according to the mask $M^l$ , which is, + +$$ +\boldsymbol {h} _ {\text {o u t}} ^ {l} = \boldsymbol {M} ^ {l} \otimes \boldsymbol {h} _ {\text {m o t i o n}} ^ {l} + (1 - \boldsymbol {M} ^ {l}) \otimes \boldsymbol {h} _ {\text {i d}} ^ {l}, \tag {2} +$$ + +where $\otimes$ denotes the Hadamard product. Through a cascade of such blocks, we obtain the final output image $\tilde{I}_y$ , which well preserves the source identity while accurately following the head motion and expression as the driving person. + +Training objective. Perceptual loss [16] is computed between the warped source image $\tilde{I}_s$ and ground-truth driving image $I_d$ for accurate warping prediction. The same loss is also applied to enforce the refinement output $\tilde{I}_y$ . + +![](images/3b11bbe1f78cb753bc77a4b297b679fbe6ab85449c6118d7a0cbd3c93aaef216.jpg) +Figure 3. Visualization of meta-learning. Compared with the one-shot pre-trained model $\phi$ , the meta-learned model $\hat{\phi}$ can be rapidly adapted to a unique personal model $\hat{\phi}_j^K$ in $K$ steps. + +We also extract the feature using face recognition model $E_{\mathrm{id}}$ [5], and penalize the dissimilarity between the ID vectors of the output image $\tilde{I}_y$ and source $I_s$ , using + +$$ +\mathcal {L} _ {\mathrm {i d}} = 1 - \cos \left(E _ {\mathrm {i d}} \left(\tilde {I} _ {y}\right), E _ {\mathrm {i d}} \left(I _ {s}\right)\right). \tag {3} +$$ + +A multi-scale patch discriminator $\mathcal{L}_{\mathrm{adv}}$ [29] is adopted to enhance the photo-realism of the outputs. To further improve the generation quality on the hard eye and mouth areas, we add additional $\mathcal{L}_1$ reconstruction losses, i.e., $\mathcal{L}_{\mathrm{eye}}, \mathcal{L}_{\mathrm{mouth}}$ for these parts. + +The overall training loss can be formulated as + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {w}} ^ {\mathrm {V G G}} + \lambda_ {\mathrm {r}} \mathcal {L} _ {\mathrm {r}} ^ {\mathrm {V G G}} + \lambda_ {\mathrm {i d}} \mathcal {L} _ {\mathrm {i d}} + \tag {4} +$$ + +$$ +\lambda_ {\text {e y e}} \mathcal {L} _ {\text {e y e}} + \lambda_ {\text {m o u t h}} \mathcal {L} _ {\text {m o u t h}} + \lambda_ {\text {a d v}} \mathcal {L} _ {\text {a d v}}, +$$ + +where $\lambda_{\mathrm{r}}, \lambda_{\mathrm{id}}, \lambda_{\mathrm{eye}}, \lambda_{\mathrm{mouth}}$ and $\lambda_{\mathrm{adv}}$ are the loss weights. + +# 3.2. Meta-learning based Faster Personalization + +While achieving state-of-the-art generation quality using our one-shot model, there always exists challenging cases that are intractable since the one-shot setting is inherently ill-posed. Person-specific characteristics and occlusion would never be faithfully recovered using a one-shot general model. Thus, personalized fine-tuning is necessary to achieve robust and authentic results that are truly usable. Nonetheless, fine-tuning the one-shot pre-trained model on a long video is computationally prohibitive for common users. To solve this, we propose a meta-learned model whose initialization weights can be easily adapted according to low-shot personal data within a few training steps, as illustrated in Figure 3. + +Formally, given a person $j$ , the personalized adaptation starts from the pre-trained model weight $\phi$ and aims to reach the optimal personal model weight $\hat{\phi}_j$ by minimiz- + +ing the error against the personal images $\hat{X}_j$ : + +$$ +\hat {\phi} _ {j} = \min _ {\phi_ {j}} \mathcal {L} \left(G _ {\phi_ {j}} \left(\hat {\boldsymbol {X}} _ {j}\right)\right), \tag {5} +$$ + +where $G_{\phi_j}$ denotes the whole generator with weight $\phi_j$ . Usually we perform $K$ steps of stochastic gradient descent (SGD) from initialization $\phi$ to approach $\hat{\phi}_j$ , so the weight updating process can be formulated as: + +$$ +\phi_ {j} ^ {k} = \operatorname {S G D} _ {K} (\phi , \hat {\boldsymbol {X}} _ {j}). \tag {6} +$$ + +Our goal is to find an optimal initialization $\hat{\phi}^K$ which could approach any personal model after $K$ steps of SGD update, even for a small $K$ , i.e., + +$$ +\hat {\phi} ^ {K} = \min _ {\phi} \sum_ {j = 1} ^ {M} \| \hat {\phi} _ {j} - \phi_ {j} ^ {k} \|. \tag {7} +$$ + +Indeed, the general one-shot pretrained model $\phi$ is a special case in the above formulation, which essentially learns the model weight in Equation 7 when $K = 0$ . When we are allowed to perform a few adaption steps ( $K > 0$ ), there is a gap between $\phi$ and desired $\hat{\phi}^K$ , since the optimization target of the standard pre-training is to minimize the overall error across all training data, it does not necessarily find the best weight suitable for personalization. + +Compared with general one-shot models, we leverage the idea of Model-Agnostic Meta-Learning (MAML) to bridge this gap and enable surprisingly fast personalized training. The goal of MAML-based methods is to optimize the initialized weights such that they could be fast adapted to a new identity within a few steps of gradient descent, which directly matches our goal. Directly optimizing the initialization weight using Equation (7) involves the computation of second-order derivatives, which is computationally expensive on large-scale training. Therefore, we utilize Reptile [26], a first-order MAML-based approach to obtain suitable initialization for fast personalization. + +To be more specific, our meta-learning model explicitly considers the personalized adaptation during training. For each person $j$ , we start from the $\phi_j^0 = \phi$ , which is the initialization to be optimized. Formally, we sample a batch of personal training data $\hat{X}_j$ , the $K$ steps of personalized training yield the finetuned model weights as: + +$$ +\phi_ {j} ^ {k} = \operatorname {S G D} \left(\phi_ {j} ^ {k - 1}, \hat {\boldsymbol {X}} _ {j}\right), k = 1, \dots , K. \tag {8} +$$ + +Finally, the personal update, i.e., the difference of $\phi_j^K$ and $\phi_j^0$ , is used as the gradient to update our initialization $\phi$ : + +$$ +\phi \leftarrow \phi - \beta \left(\phi_ {j} ^ {K} - \phi\right), \tag {9} +$$ + +where $\beta$ is the meta-learning rate. The full algorithm is shown in Algorithm 1. Our model progressively learns a more suitable initialization through meta-training as visualized in Figure 3, which could fast adapt to personalized models after limited steps of adaptation. + +# Algorithm 1 Optimization of Initial Weights with Reptile + +Input: weights $\phi$ of generation network $G$ , inner loop learning rate $\alpha$ , meta-learning rate $\beta$ , number of training iterations $N$ , number of training persons $M$ , number of inner loop iterations $K$ + +1: for $i = 1, \dots, N$ do +2: for $j = 1,\dots ,M$ do +3: $\phi_j^0 = \mathbf{Clone}(\phi)$ +4: Sample a training batch $\hat{X}_j$ of person $j$ +5: for $k = 1,\dots ,K$ do +6: $\phi_j^k = \phi_j^{k - 1} - \alpha \nabla_\phi \mathcal{L}(G_\phi (\hat{\mathbf{X}}_j))$ +7: end for +8: end for +9: $\phi \gets \phi -\frac{\beta}{M}\sum_{j = 1}^{M}(\phi_j^K -\phi)$ +10: end for + +# 3.3. Temporal-consistent Super-resolution Network + +To further enhance the generation resolution and improve the high-fidelity details of our output, a video super-resolution network is needed as the final stage of the generation framework. Previous talking head synthesis works [7,47] utilize single-frame super-resolution as the last stage and ignore the quality of temporal consistency and stability. Performing super-resolution in frame-by-frame manner tends to produce texture flickering which severely hampers the visual quality. In this work, we consider multiple adjacent frames to ensure temporal coherency. + +Inspired by previous 2D face restoration works [40, 46], the pre-trained generative models [17, 18, 50] like StyleGAN contain rich face prior and could significantly help to enhance the high-frequency details. Moreover, the disentangled $\mathcal{W}$ space in StyleGAN provides desirable temporal consistency during manipulation [37], which also benefits our framework. Therefore, we propose a temporally consistent super-resolution module by leveraging pretrained StyleGAN and 3D convolution, where the latter brings quality enhancement in spatio-temporal domain. As shown in Fig. 2, we feed the concatenated sequence of $t$ video frames $\{\tilde{I}_y^1,\tilde{I}_y^2,\dots,\tilde{I}_y^t\}$ into a U-Net composed of 3D convolution with reflection padding on the temporal dimension. To ensure pre-trained per-frame quality while improving temporal consistency, we initialize the 3D convolution weight in U-Net with a pretrained 2D face restoration network [40]. + +These spatio-temporally enhanced features from the U-Net decoder further modulate the pretrained StyleGAN features through FiLM. Thus, the super-resolution frames $\{I_y^{1}, I_y^{2}, \dots, I_y^{t}\}$ are obtained as: + +$$ +\left\{\boldsymbol {I} _ {y} ^ {\prime 1}, \boldsymbol {I} _ {y} ^ {\prime 2}, \dots , \boldsymbol {I} _ {y} ^ {\prime t} \right\} = F _ {\text {S t y l e G A N}} \left(F _ {\text {3 D}} \left(\left\{\tilde {\boldsymbol {I}} _ {y} ^ {1}, \tilde {\boldsymbol {I}} _ {y} ^ {2}, \dots , \tilde {\boldsymbol {I}} _ {y} ^ {t} \right\}\right)\right). \tag {10} +$$ + +During training, we optimize the output $I_{y}^{\prime}$ towards the $512 \times 512$ ground truth $I_{d}$ using $\ell_{1}$ and perceptual loss. + +![](images/94cc2c767e91afe306b209cc76e33a36b8d607f85e27bcdedf820a6cdcdf1e5b.jpg) +Figure 4. Qualitative results of our self-reconstruction (top row) and cross-identity reenactment (bottom two rows) at $256 \times 256$ resolution. Our method synthesizes more faithful expression and motion, while better preserving the identity of the source portrait. + +![](images/5540f3c19d16e01b241b0346ed41d2accd1944429397f622ab27bd9fd5e87cb6.jpg) +Figure 5. Qualitative results of self reconstruction and cross-identity reenactment at $512 \times 512$ resolution. Our spatial-temporal superresolution module further enhances more high-frequency details on teeth, eyes, and hair. + +# 4. Experiments + +# 4.1. Experiment Setup + +Dataset. Following [7], we train our warping and refinement networks on cropped VoxCeleb2 dataset [4] at $256^2$ resolution. We randomly select 500 videos from the test set for evaluation. For our meta-learning-based fast personalization, we finetune our base model on HDTF dataset [52], which is composed of 410 videos from 300 different identities. We downsample cropped faces to $256^2$ resolution and split the original HDTF dataset into 400 training videos and + +10 test videos. After the convergence of our meta-training, we further evaluate the personalization speed of our model on the HDTF test set. Our temporal super-resolution module $F_{3d}$ is trained on the HDTF dataset [52] which has 300 frames per video with $512^2$ resolution. We feed downsampled $256^2$ frames into fixed warping and refinement network and use the outputs of the refinement network as inputs for the next temporal super-resolution module. More training details are provided in the supplementary. + +Metrics We evaluate the fidelity of self-reconstruction using FID [14] and LPIPS [51]. Our motion transfer qual + +
MethodsSelf Reconstruction (256 × 256)Cross Reenactment (256 × 256)
FID↓LPIPS↓ID Loss↓AED↓APD↓FID↓ID Loss↓AED↓APD↓
X2Face [41]45.29080.68060.96320.21470.100791.14850.64960.31120.1210
Bi-layer [48]100.91960.58810.52800.12580.0139127.78230.63360.23300.0208
FOMM [33]12.19790.23380.20960.09640.010080.16370.57600.23400.0239
PIRender [31]14.40650.26390.30240.10800.016278.84300.54400.21130.0214
Ours11.95280.22620.12960.09420.012477.50480.29440.25240.0258
MethodsSelf Reconstruction (512 × 512)Cross Reenactment (512 × 512)
StyleHEAT [47]44.52070.28400.41120.11550.0131111.34500.47200.25050.0218
Ours21.49740.20790.08320.09040.012149.60200.19520.27370.0242
+ +Table 1. Quantitative results for self-reconstruction and cross-reenactment. We evaluate both $256 \times 256$ results and $512 \times 512$ results. Our method outperforms all baselines on both resolutions across image fidelity metrics with comparable motion transfer results. + +![](images/ab90219bc083f4c5b1c8491790e86148f1237ab2322c7b00d45a794b519c3f84.jpg) +Source + +![](images/5f2433757bdb30aafc52f44dbf9cc74d9336e69f8405b94732f681aaa23e413b.jpg) + +![](images/d9a23ba10ea9dbf8347ee038be83ad616905e906646aacb308d65c56707282b2.jpg) +Driving + +![](images/b7da9b6081c1637b0b43d8e2463e80b3bc9331bce43c22bcffcf59377273974e.jpg) + +![](images/3e35a835034b52e24561264ffb79341570e3cd5f52dd063440e8fb5021a30db9.jpg) +FOMM +Figure 6. A comparison of different personalized models at the same epochs. Our personalized base model without reptile adaptation reconstructs more accurate skin color and source identity than the personalized FOMM method. For the model personalized from a reptile learning stage, finer detail on teeth and eye colors could be further generated. + +![](images/0a648b25a52e7f060dd462def80a6abe88931461ca07db5507de4a2cd059c511.jpg) + +![](images/a359a42399421bb22e3dbe8135384ce6ac15e41ea08414e9e31aa4f50298d507.jpg) +Ours w/o Reptile + +![](images/7c35b06cb76ccf71e9e40a3aab9b90d3fc2f87fc0735b80b9e1b780062237e55.jpg) + +![](images/28e1446c407015a50c511bd3a9560809daa94b73ae5c70356e3cf95cb9e5128f.jpg) +Ours w/ Reptile + +ity is measured using average expression distance (AED) and average pose distance (APD) with driving videos. For our meta-learning-based fast personalization, we illustrate the LPIPS at different personalization epochs for each fine-tuning approach. Following previous works [21, 22], we evaluate our temporal consistency using warping error $E_{\mathrm{warp}}$ . For each frame $\tilde{I}_y^t$ , we calculate the warping error with the previous frame $\tilde{I}_y^{t - 1}$ warped by the estimated optical flow in the occlusion map. + +# 4.2. Comparison with State-of-the-art Methods + +We compare our warping and refinement models at $256 \times 256$ resolution against several state-of-the-art face reenactment works: X2Face [41], Bi-Layer [1], First-Order Motion Model [33], and PIRender [31]. The top row of Figure 4 presents our qualitative results of self-reconstruction. Different from sparse landmarks in unsupervised learning [33] + +or 1D 3DMM [31], our dense landmarks provide strong guidance to accurate and fine-detailed synthesis on gaze, mouth, and expressions. The last two rows show the results of our cross-identity reenactment. Since our landmarks have a better decomposition of identity and motion and our refinement network is identity-aware, our method is the only one that well preserves the identity of source image. In contrast, previous methods suffer from appearance data leakage directly from the driver and generate faces with a similar identity to the driving image. Note that the 3DMM coefficients of AED and APD evaluation are not fully decomposed. Thus, other baselines with identity leakage may have better quantitative metrics. We compare our full framework with the proposed temporal super-resolution module at $512 \times 512$ resolution against StyleHEAT [47], which is the only open-sourced method that generates high-resolution talking heads. StyleHEAT [47] fails to synthesize + +![](images/c660bb192bb6d895f0f779e58464f663b5ef9a7cc30632aa6d6d0445af0f80e4.jpg) +Figure 7. Comparison of test LPIPS at fine-tuning epochs for different approaches. Our reptile-based model achieves more than $3 \times$ speedup compared with the base model and previous baseline. + +sharp and accurate teeth in our experiment, and the identity of the output image is quite different from the source portrait due to the limitations of GAN inversion. In contrast, our refinement network is identity-aware and we leverage pretrained StyleGAN in our temporal super-resolution module to fully exploit its face prior knowledge. Figure 5 shows that our method produces sharp teeth and hairs, while bicubic-upsampled results are blurry with artifacts. + +Table 1 demonstrates that our method achieves the best quantitative fidelity and comparable motion transfer quality on both self-reconstruction and cross-identity reenactment. + +# 4.3. Evaluation of Fast Personalization + +Although our base model achieves state-of-the-art quality as a general model, there still exists ill-posedness in some cases. For example, it is very difficult to synthesize teeth according to a closed source mouth. Thus, industry products typically conduct personalization using fine-tuning strategy, which can be computationally expensive. To achieve faster convergence, we finetune our base model using a meta-learning strategy to provide a better weight initialization for the following personalization. In Fig 7, we evaluate the personalization speed of our metalearned model against our base model and previous baseline FOMM [33]. It takes our meta-learning model 0.5 epoch to decrease LPIPS to 0.14, which is $3 \times$ speedup against our base model, and $4 \times$ against FOMM. Figure 6 compares the personalization of our method and FOMM [33] at the same epoch, which illustrates our fast adaptation speed on ambiguous areas (e.g., teeth, eyes, and wrinkle details). + +# 4.4. Evaluation of Temporal Super-resolution + +In Table 2, we also evaluate the performance of our temporal super-resolution using 2D image fidelity and warping error $E_{\text{warp}}$ . We train a 2D super-resolution baseline using GFPGAN [40]. The quantitative result in Table 2 shows + +
MethodsFID↓LPIPS↓E_warp ↓
Ground Truth--0.0182
Base w/ bicubic25.57620.22850.0184
Base w/ GFPGAN [40]22.63510.21780.0242
Ours21.49740.20790.0213
+ +that although naive 2D super-resolution improves per-frame fidelity, it also brings more flickering and larger warping error (0.0242) than simple bicubic upsampling (0.0184). To achieve temporally coherent results, we combine a U-Net composed of 3D convolution with facial prior [18], which significantly reduces $E_{\text{warp}}$ of our final videos from 0.0242 to 0.0213, and preserves compelling 2D facial details. + +# 4.5. Ablation Study of Base Model + +Table 2. Quantitative evaluation of our temporal super-resolution on self-reconstruction at $512 \times 512$ resolution. + +
MethodsFID↓LPIPS↓ID Loss↓
Ours Sparse Landmark14.31900.24850.1424
Ours w/o ID12.27360.22560.2144
Ours11.95280.22620.1296
+ +Table 3. Quantitative ablation study of landmarks and ID coefficients in our base model. + +We conduct ablation studies to validate the effectiveness of our driving motion and source identity representation in our base model. If we replace our 669 dense landmarks with sparse landmarks, the LPIPS of warped source images degrades by 0.2. To evaluate our identity-aware refinement, the removal of the identity input causes significant increase of identity loss from 0.1296 to 0.2144. + +# 5. Conclusion + +We present a novel framework for identity-preserving one-shot talking head generation. To faithfully maintain the source ID, we propose to leverage accurate dense landmarks in the warping network and explicit source identity during refinement. Further, we significantly advance the applicability of personalized model by reducing its training to 30 seconds with meta-learning. Last but not least, we enhance the final resolution and temporal consistency with 3D convolution and generative prior. Comprehensive experiments demonstrate the state-of-the-art performance of our system. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 1, 2, 7 +[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. international conference on computer graphics and interactive techniques, 1999. 2 + +[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponni-ahy, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In CVPR, 2016. 2 +[4] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. conference of the international speech communication association, 2018. 2, 6 +[5] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In CVPR, 2020. 4 +[6] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In ICCV, 2021. 1, 2 +[7] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In ACM Multimedia, 2022. 1, 2, 5, 6 +[8] Vincent Dumoulin, Ethan Perez, Nathan Schucher, Florian Strub, Harm de Vries, Aaron Courville, and Yoshua Bengio. Feature-wise transformations. Distill, 2018. https://distill.pub/2018/features-wise-transformations.4 +[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2, 3 +[10] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Trans. Graph., 2019. 2 +[11] Jiahao Geng, Tianjia Shao, Youyi Zheng, Yanlin Weng, and Kun Zhou. Warp-guided gans for single-photo facial animation. ACM Transactions on Graphics (TOG), 2018. 2 +[12] Xiaojie Guo, Siyuan Li, Jiawan Zhang, Jiayi Ma, Lin Ma, Wei Liu, and Haibin Ling. Pfld: A practical facial landmark detector. arXiv: Computer Vision and Pattern Recognition, 2019. 2 +[13] Sungjoo Ha, Martin Kersner, Beomsu Kim, Seokjun Seo, and Dongyoung Kim. Marionette: Few-shot face reenactment preserving identity of unseen targets. In AAAI, 2020. 2 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6 +[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 3 +[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 4 +[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 5 +[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 5, 8 +[19] Hyeongwoo Kim, Mohamed Elgharib, Hans-Peter Zollofer, Michael Seidel, Thabo Beeler, Christian Richardt, and Christian Theobalt. Neural style-preserving visual dubbing. ACM Transactions on Graphics (TOG), 2019. 2 +[20] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Nießner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. Association for Computing Machinery, 2018. 2 +[21] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 7 +[22] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. In NeurIPS, 2020. 7 +[23] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv: Computer Vision and Pattern Recognition, 2019. 2, 4 + +[24] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit Warping for Animation with Image Sets. In NeurIPS, 2022. 1, 2 +[25] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: A large-scale speaker identification dataset. conference of the international speech communication association, 2017. 2 +[26] Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv: Learning, 2018. 2, 3, 5 +[27] Hao Ouyang, Bo Zhang, Pan Zhang, Hao Yang, Jiaolong Yang, Dong Chen, Qifeng Chen, and Fang Wen. Real-time neural character rendering with pose-guided multiplane images. ECCV, 2022. 1 +[28] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. arXiv preprint arXiv:2301.06281, 2023. 2 +[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 4 +[30] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In AAAI, 2018. 4 +[31] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In ICCV, 2021. 1, 2, 7 +[32] Mincheol Shin, Se Jung Kim, and Frank Biocca. The uncanny valley: No need for any further judgments when an avatar looks eerie. Computers in Human Behavior, 94:100-109, 2019. 2 +[33] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In NeurIPS, 2019. 1, 2, 3, 7, 8 +[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In CVPR, 2019. 2, 3 +[35] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 2 +[36] Soumya Tripathy, Juho Kannala, and Esa Rahtu. Facegan: Facial attribute controllable reenactment gan. In CVPR, 2021. 2 +[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit H. Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial editing of real videos*, 2022. 5 +[38] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaevator: Learning animatable clothed human models from few depth images. In NeurIPS, 2021. 3 +[39] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, 2020. 1, 2, 3 +[40] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In CVPR, 2021. 2, 5, 8 +[41] Olivia Wiles, A Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018. 1, 2, 7 +[42] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Matthew Johnson, Virginia Estellers, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone. In ICCV, 2021. 2, 3 +[43] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Chirag Raman, Jamie Shotton, Toby Sharp, Ivan Stojiljkovic, Tom Cashman, and Julien Valentin. 3d face reconstruction with dense landmarks. In ECCV, 2022. 2, 3 +[44] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. arXiv preprint arXiv:2301.02379, 2023. 2 + +[45] Lingbo Yang, Chang Liu, Pan Wang, Shanshe Wang, Peiran Ren, Siwei Ma, and Wen Gao. Hifacegan: Face renovation via collaborative suppression and replenishment. ACM Multimedia, 2020. 2 +[46] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 5 +[47] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. In ECCV, 2022. 1, 2, 5, 7 +[48] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In ECCV, 2020. 2, 7 +[49] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In ICCV, 2019. 3 +[50] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In CVPR, 2022. 5 +[51] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[52] Zhimeng Zhang, Lincheng Li, and Yu Ding. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In CVPR, 2021. 6 +[53] Ruiqi Zhao, Tianyi Wu, and Guodong Guo. Sparse to dense motion transfer for face image animation. In ICCV, 2021. 2 \ No newline at end of file diff --git a/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/images.zip b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..02c702c31d5bb7af6cef01f3df4f432830f1001a --- /dev/null +++ b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e801f793f86199880df388dbfd8858cdabd0deb1cd108c2b03d5cf6e1afcd6a9 +size 669537 diff --git a/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/layout.json b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cd098048e05d5d05ff61f25e266743ca842c0f22 --- /dev/null +++ b/2023/MetaPortrait_ Identity-Preserving Talking Head Generation With Fast Personalized Adaptation/layout.json @@ -0,0 +1,10516 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 136, + 470, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 136, + 470, + 171 + ], + "spans": [ + { + "bbox": [ + 108, + 136, + 470, + 171 + ], + "type": "text", + "content": "MetaPortrait: Identity-Preserving Talking Head Generation with Fast Personalized Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 191, + 149, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 191, + 149, + 207 + ], + "spans": [ + { + "bbox": [ + 73, + 191, + 149, + 207 + ], + "type": "text", + "content": "Bowen Zhang" + }, + { + "bbox": [ + 73, + 191, + 149, + 207 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 174, + 192, + 245, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 192, + 245, + 207 + ], + "spans": [ + { + "bbox": [ + 174, + 192, + 245, + 207 + ], + "type": "text", + "content": "Chenyang Qi" + }, + { + "bbox": [ + 174, + 192, + 245, + 207 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 270, + 192, + 326, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 192, + 326, + 207 + ], + "spans": [ + { + "bbox": [ + 270, + 192, + 326, + 207 + ], + "type": "text", + "content": "Pan Zhang1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 351, + 192, + 408, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 192, + 408, + 207 + ], + "spans": [ + { + "bbox": [ + 351, + 192, + 408, + 207 + ], + "type": "text", + "content": "Bo Zhang" + }, + { + "bbox": [ + 351, + 192, + 408, + 207 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 427, + 192, + 503, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 192, + 503, + 207 + ], + "spans": [ + { + "bbox": [ + 427, + 192, + 503, + 207 + ], + "type": "text", + "content": "HsiangTao " + }, + { + "bbox": [ + 427, + 192, + 503, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{3}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 208, + 202, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 208, + 202, + 222 + ], + "spans": [ + { + "bbox": [ + 142, + 208, + 202, + 222 + ], + "type": "text", + "content": "Dong Chen" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 227, + 207, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 207, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 227, + 207, + 296, + 222 + ], + "type": "text", + "content": "Qifeng Chen" + }, + { + "bbox": [ + 227, + 207, + 296, + 222 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 207, + 377, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 207, + 377, + 222 + ], + "spans": [ + { + "bbox": [ + 318, + 207, + 377, + 222 + ], + "type": "text", + "content": "Yong Wang1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 401, + 207, + 455, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 207, + 455, + 222 + ], + "spans": [ + { + "bbox": [ + 401, + 207, + 455, + 222 + ], + "type": "text", + "content": "Fang Wen3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 196, + 224, + 233, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 224, + 233, + 237 + ], + "spans": [ + { + "bbox": [ + 196, + 224, + 233, + 237 + ], + "type": "text", + "content": "1USTC" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 256, + 224, + 302, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 224, + 302, + 237 + ], + "spans": [ + { + "bbox": [ + 256, + 224, + 302, + 237 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 256, + 224, + 302, + 237 + ], + "type": "text", + "content": "HKUST" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 224, + 379, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 224, + 379, + 237 + ], + "spans": [ + { + "bbox": [ + 326, + 224, + 379, + 237 + ], + "type": "text", + "content": "3Microsoft" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 244, + 149, + 335 + ], + "blocks": [ + { + "bbox": [ + 58, + 244, + 149, + 335 + ], + "lines": [ + { + "bbox": [ + 58, + 244, + 149, + 335 + ], + "spans": [ + { + "bbox": [ + 58, + 244, + 149, + 335 + ], + "type": "image", + "image_path": "5b75b5dbbd8923e19d6f83ac11e6b60850c10440cb43ed6e5b098a9c4ff78437.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 58, + 335, + 148, + 426 + ], + "blocks": [ + { + "bbox": [ + 58, + 335, + 148, + 426 + ], + "lines": [ + { + "bbox": [ + 58, + 335, + 148, + 426 + ], + "spans": [ + { + "bbox": [ + 58, + 335, + 148, + 426 + ], + "type": "image", + "image_path": "b51786e4d3264e6ac6dc6dd885488b1fe601d8b032ff89400658c9754bb5092a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 426, + 116, + 436 + ], + "lines": [ + { + "bbox": [ + 91, + 426, + 116, + 436 + ], + "spans": [ + { + "bbox": [ + 91, + 426, + 116, + 436 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 98, + 440, + 477, + 452 + ], + "lines": [ + { + "bbox": [ + 98, + 440, + 477, + 452 + ], + "spans": [ + { + "bbox": [ + 98, + 440, + 477, + 452 + ], + "type": "text", + "content": "Figure 1. Our method yields identity-preserving talking head generation. See the webpage for video demos." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 150, + 244, + 240, + 335 + ], + "blocks": [ + { + "bbox": [ + 150, + 244, + 240, + 335 + ], + "lines": [ + { + "bbox": [ + 150, + 244, + 240, + 335 + ], + "spans": [ + { + "bbox": [ + 150, + 244, + 240, + 335 + ], + "type": "image", + "image_path": "967a942b302c45d756cde2491f8b48e7a90c6e2478801b6c8f31abc9059cba7a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 150, + 335, + 240, + 426 + ], + "blocks": [ + { + "bbox": [ + 150, + 335, + 240, + 426 + ], + "lines": [ + { + "bbox": [ + 150, + 335, + 240, + 426 + ], + "spans": [ + { + "bbox": [ + 150, + 335, + 240, + 426 + ], + "type": "image", + "image_path": "40be958db55379b27dcd67e9adb7be7974fe1148722f4e68302d185d21e57c0f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 426, + 209, + 436 + ], + "lines": [ + { + "bbox": [ + 182, + 426, + 209, + 436 + ], + "spans": [ + { + "bbox": [ + 182, + 426, + 209, + 436 + ], + "type": "text", + "content": "Driving" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 242, + 244, + 333, + 335 + ], + "blocks": [ + { + "bbox": [ + 242, + 244, + 333, + 335 + ], + "lines": [ + { + "bbox": [ + 242, + 244, + 333, + 335 + ], + "spans": [ + { + "bbox": [ + 242, + 244, + 333, + 335 + ], + "type": "image", + "image_path": "51516888e33fd5eeb1f6d40d56507332f88176cf6f5f5369e8611f5f67437184.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 242, + 335, + 333, + 426 + ], + "blocks": [ + { + "bbox": [ + 242, + 335, + 333, + 426 + ], + "lines": [ + { + "bbox": [ + 242, + 335, + 333, + 426 + ], + "spans": [ + { + "bbox": [ + 242, + 335, + 333, + 426 + ], + "type": "image", + "image_path": "714b603906da0ed8f43d498389d4ae21c1a6eb46d3db74291ef245fef4fd4c98.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 426, + 297, + 436 + ], + "lines": [ + { + "bbox": [ + 279, + 426, + 297, + 436 + ], + "spans": [ + { + "bbox": [ + 279, + 426, + 297, + 436 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 334, + 244, + 424, + 335 + ], + "blocks": [ + { + "bbox": [ + 334, + 244, + 424, + 335 + ], + "lines": [ + { + "bbox": [ + 334, + 244, + 424, + 335 + ], + "spans": [ + { + "bbox": [ + 334, + 244, + 424, + 335 + ], + "type": "image", + "image_path": "ec661b1f362f159a77b94ce3fa3f5b04f8f5b5565f109a2cc33334ef9fcc7020.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 334, + 335, + 424, + 426 + ], + "blocks": [ + { + "bbox": [ + 334, + 335, + 424, + 426 + ], + "lines": [ + { + "bbox": [ + 334, + 335, + 424, + 426 + ], + "spans": [ + { + "bbox": [ + 334, + 335, + 424, + 426 + ], + "type": "image", + "image_path": "61df584a2845a9df2d0819114f19bcf60ae94b68c82e7ef1e9f86942d5fbae7f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 426, + 395, + 436 + ], + "lines": [ + { + "bbox": [ + 367, + 426, + 395, + 436 + ], + "spans": [ + { + "bbox": [ + 367, + 426, + 395, + 436 + ], + "type": "text", + "content": "Driving" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 427, + 244, + 517, + 335 + ], + "blocks": [ + { + "bbox": [ + 427, + 244, + 517, + 335 + ], + "lines": [ + { + "bbox": [ + 427, + 244, + 517, + 335 + ], + "spans": [ + { + "bbox": [ + 427, + 244, + 517, + 335 + ], + "type": "image", + "image_path": "37c874edcae1660da2469f966120204f2864f0e12cc9d25feaac669c1c0a91a6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 427, + 335, + 517, + 426 + ], + "blocks": [ + { + "bbox": [ + 427, + 335, + 517, + 426 + ], + "lines": [ + { + "bbox": [ + 427, + 335, + 517, + 426 + ], + "spans": [ + { + "bbox": [ + 427, + 335, + 517, + 426 + ], + "type": "image", + "image_path": "5450721974ec78c7e90723e0ee463eeed0a38ee2bb3343ef3b0fcf7416feac88.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 426, + 480, + 435 + ], + "lines": [ + { + "bbox": [ + 463, + 426, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 463, + 426, + 480, + 435 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 139, + 465, + 186, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 465, + 186, + 477 + ], + "spans": [ + { + "bbox": [ + 139, + 465, + 186, + 477 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 45, + 480, + 281, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 480, + 281, + 702 + ], + "spans": [ + { + "bbox": [ + 45, + 480, + 281, + 702 + ], + "type": "text", + "content": "In this work, we propose an ID-preserving talking head generation framework, which advances previous methods in two aspects. First, as opposed to interpolating from sparse flow, we claim that dense landmarks are crucial to achieving accurate geometry-aware flow fields. Second, inspired by face-swapping methods, we adaptively fuse the source identity during synthesis, so that the network better preserves the key characteristics of the image portrait. Although the proposed model surpasses prior generation fidelity on established benchmarks, personalized fine-tuning is still needed to further make the talking head generation qualified for real usage. However, this process is rather computationally demanding that is unaffordable to standard users. To alleviate this, we propose a fast adaptation model using a meta-learning approach. The learned model can be adapted to a high-quality personalized model as fast as 30 seconds. Last but not least, a spatial-temporal enhancement module is proposed to improve the fine details while ensuring temporal coherency. Extensive experiments" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 297, + 467, + 530, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 467, + 530, + 490 + ], + "spans": [ + { + "bbox": [ + 297, + 467, + 530, + 490 + ], + "type": "text", + "content": "prove the significant superiority of our approach over the state of the arts in both one-shot and personalized settings." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 298, + 509, + 375, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 509, + 375, + 521 + ], + "spans": [ + { + "bbox": [ + 298, + 509, + 375, + 521 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 296, + 530, + 530, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 530, + 530, + 693 + ], + "spans": [ + { + "bbox": [ + 296, + 530, + 530, + 693 + ], + "type": "text", + "content": "Talking head generation [1,6,7,24,27,31,33,39,41,47] has found extensive applications in face-to-face live chat, virtual reality and virtual avatars in games and videos. In this paper, we aim to synthesize a realistic talking head with a single source image (one-shot) that provides the appearance of a given person while being animatable according to the motion of the driving person. Recently, considerable progress has been made with neural rendering techniques, bypassing the sophisticated 3D human modeling process and expensive driving sensors. While these works attain increasing fidelity and higher rendering resolution, identity preserving remains a challenging issue since the human vision system is particularly sensitive to any nuanced deviation from the person's facial geometry." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "content": "Prior arts mainly focus on learning a geometry-aware warping field, either by interpolating from sparse 2D/3D landmarks or leveraging 3D face prior, e.g., 3D morphable" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 73 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 709, + 217, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 709, + 217, + 719 + ], + "spans": [ + { + "bbox": [ + 57, + 709, + 217, + 719 + ], + "type": "text", + "content": "*Equal contribution, interns at Microsoft Research." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 57, + 719, + 150, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 719, + 150, + 729 + ], + "spans": [ + { + "bbox": [ + 57, + 719, + 150, + 729 + ], + "type": "text", + "content": "† Joint corresponding authors." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "type": "text", + "content": "22096" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 188 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 188 + ], + "type": "text", + "content": "face model (3DMM) [2, 3]. However, fine-grained facial geometry may not be well described by a set of sparse landmarks or inaccurate face reconstruction. Indeed, the warping field, trained in a self-supervised manner rather than using accurate flow ground truth, can only model coarse geometry deformation, lacking the expressivity that captures the subtle semantic characteristics of the portrait." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 189, + 279, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 189, + 279, + 364 + ], + "spans": [ + { + "bbox": [ + 45, + 189, + 279, + 364 + ], + "type": "text", + "content": "In this paper, we propose to better preserve the portrait identity in two ways. First, we claim that dense facial landmarks are sufficient for an accurate warping field prediction without the need for local affine transformation. Specifically, we adopt a landmark prediction model [43] trained on synthetic data [42], yielding 669 head landmarks that offer significantly richer information on facial geometry. In addition, we build upon the face-swapping approach [23] and propose to enhance the perceptual identity by attentionally fusing the identity feature of the source portrait while retaining the pose and expression of the intermediate warping. Equipped with these two improvements, our one-shot model demonstrates a significant advantage over prior works in terms of both image quality and perceptual identity preservation when animating in-the-wild portraits." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 365, + 279, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 365, + 279, + 540 + ], + "spans": [ + { + "bbox": [ + 45, + 365, + 279, + 540 + ], + "type": "text", + "content": "While our one-shot talking head model has achieved state-of-the-art quality, it is still infeasible to guarantee satisfactory synthesis results because such a one-shot setting is inherently ill-posed—one may never hallucinate the person-specific facial shape and occluded content from a single photo. Hence, ultimately we encounter the uncanny valley [32] that a user becomes uncomfortable as the synthesis results approach to realism. To circumvent this, one workaround is to finetune the model using several minutes of a personal video. Such personalized training has been widely adopted in industry to ensure product-level quality, yet this process is computationally expensive, which greatly limits its use scenarios. Thus, speeding up this personalized training, a task previously under-explored, is of great significance to the application of talking head synthesis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 542, + 279, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 542, + 279, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 542, + 279, + 693 + ], + "type": "text", + "content": "We propose to achieve fast personalization with meta-learning. The key idea is to find an initialization model that can be easily adapted to a given identity with limited training iterations. To this end, we resort to a meta-learning approach [9, 26] that finds success in quickly learning discriminative tasks, yet is rarely explored in generative tasks. Specifically, we optimize the model for specific personal data with a few iterations. In this way, we get a slightly fine-tuned personal model towards which we move the initialization model weight a little bit. Such meta-learned initialization allows us to train a personal model within 30 seconds, which is 3 times faster than a vanilla pretrained model while requiring less amount of personal data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": "Moreover, we propose a novel temporal super-resolution network to enhance the resolution of the generated talking head video. To do this, we leverage the generative prior" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "content": "to boost the high-frequency details for portraits and meanwhile take into account adjacent frames that are helpful to reduce temporal flickering. Finally, we reach temporally coherent video results of " + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "content": " resolution with compelling facial details. In summary, this work innovates in the following aspects:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 176, + 530, + 269 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 297, + 176, + 530, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 176, + 530, + 211 + ], + "spans": [ + { + "bbox": [ + 297, + 176, + 530, + 211 + ], + "type": "text", + "content": "- We propose a carefully designed framework to significantly improve the identity-preserving capability when animating a one-shot in-the-wild portrait." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 211, + 530, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 211, + 530, + 245 + ], + "spans": [ + { + "bbox": [ + 297, + 211, + 530, + 245 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to explore meta-learning to accelerate personalized training, thus obtaining ultra-high-quality results at affordable cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 245, + 530, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 245, + 530, + 269 + ], + "spans": [ + { + "bbox": [ + 297, + 245, + 530, + 269 + ], + "type": "text", + "content": "- Our novel video super-resolution model effectively enhances details without introducing temporal flickering." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 279, + 381, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 279, + 381, + 292 + ], + "spans": [ + { + "bbox": [ + 297, + 279, + 381, + 292 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 298, + 530, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 298, + 530, + 461 + ], + "spans": [ + { + "bbox": [ + 296, + 298, + 530, + 461 + ], + "type": "text", + "content": "2D-based talking head synthesis. Methods along this line [24, 28, 33, 34, 39] predict explicit warping flow by interpolating the sparse flow defined by 2D landmarks. FOMM [33] assumes local affine transformation for flow interpolation. Recently, Mallya et al. [24] computes landmarks from multiple source images using an attention mechanism. However, the landmarks learned in an unsupervised manner are too sparse (e.g., 20 in FaceVid2Vid [39]) to be interpolated into dense flows. Using predefined facial landmarks [12] is also a straightforward approach [13, 36, 48, 53] to represent the motion of driving images. For example, PFLD [12] uses 98 facial landmarks. However, they could not generate an accurate warping flow since the landmarks are not dense enough." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 462, + 530, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 462, + 530, + 602 + ], + "spans": [ + { + "bbox": [ + 296, + 462, + 530, + 602 + ], + "type": "text", + "content": "Talking head synthesis with 3D face prior. 3D Morphable Models [2,3] represent a face image as PCA coefficients relating to identity, expression, and pose, which provides an easy tool to edit and render portrait images [10, 11, 19, 20, 35, 44]. Some attempts [19, 20] render the animated faces by combining the identity coefficients of the source image and the motion coefficients of the driving video. Recent works [6, 31] predict a warping flow field for talking head synthesis using 3DMM. Although 3DMM-based methods allow explicit face control, using these coefficients to represent detailed face geometry and expression remains challenging." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 602, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 602, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 602, + 530, + 730 + ], + "type": "text", + "content": "Towards higher-resolution talking head. The video resolution of most talking head methods [1, 31, 33, 41] is " + }, + { + "bbox": [ + 296, + 602, + 530, + 730 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 296, + 602, + 530, + 730 + ], + "type": "text", + "content": ", bounded by the available video datasets [4, 25]. To enhance the visual quality of output videos, previous literature trains an additional single-image super-resolution network [7, 40, 45] or utilizes pretrained 2D StyleGAN [40, 47]. However, these methods upsample the original video in a frame-by-frame manner and ignore the temporal consistency of adjacent video frames. In this work, we propose a novel temporal super-resolution module to boost temporal consistency while preserving per-frame quality." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "22097" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 106, + 517, + 323 + ], + "blocks": [ + { + "bbox": [ + 58, + 106, + 517, + 323 + ], + "lines": [ + { + "bbox": [ + 58, + 106, + 517, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 106, + 517, + 323 + ], + "type": "image", + "image_path": "98c9f11cfa12e4f1c0d56579a6f358518e3722052592fbf2d612250bcc4d9161.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "lines": [ + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "spans": [ + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": "Figure 2. Overview of our one-shot framework. (a) Given a source image " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_s" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": " and a driving video " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_d^1, \\mathbf{I}_d^2, \\dots, \\mathbf{I}_d^t\\}" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": ", we first extract their dense landmarks " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_s^{\\mathrm{ldmk}}, \\mathbf{I}_d^{\\mathrm{ldmk}})" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": " using a pretrained landmark detector. (b) Then, we estimate warping flows " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": " between the source image and each driving frame according to concatenated input " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{in}}" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": ". (c) We further refine the warped source input " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}_s" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": " using an ID-preserving network. (d) Finally, we enhance and upsample the " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": " results " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "\\{\\tilde{I}_y^1, \\tilde{I}_y^2, \\dots, \\tilde{I}_y^t\\}" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": " to high-fidelity output " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}'_y^1, \\mathbf{I}'_y^2, \\dots, \\mathbf{I}'_y^t\\}" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 45, + 325, + 531, + 372 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 376, + 281, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 376, + 281, + 480 + ], + "spans": [ + { + "bbox": [ + 45, + 376, + 281, + 480 + ], + "type": "text", + "content": "Meta learning and fast adaptation. The goal of meta-learning methods [9, 26, 38, 49] is to achieve good performance with a limited amount of training data. In this paper, we aim to quickly build a personalized model given a new identity, which correlates with the goal of meta-learning. Zakharov et al. [49] propose to improve GAN inversion using a hyper-network. We leverage the idea of model-agnostic meta-learning (MAML) [9] to obtain the best initialization that can be easily adapted to different persons." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 492, + 101, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 492, + 101, + 505 + ], + "spans": [ + { + "bbox": [ + 45, + 492, + 101, + 505 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "spans": [ + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "content": "Figure 2 illustrates an overview of our synthesis framework. Given an image " + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "content": " of a person which we refer as source image and a sequence of " + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "content": " driving video frames " + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "inline_equation", + "content": "\\{I_d^1, I_d^2, \\dots, I_d^t\\}" + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "content": ", we aim to generate an output video " + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "inline_equation", + "content": "\\{I_y^1, I_y^2, \\dots, I_y^t\\}" + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "content": " with the motions derived from the driving video while maintaining the identity of the source image. Section 3.1 introduces our one-shot model (Figure 2(a, b, c)) for identity-preserving talking head generation " + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "inline_equation", + "content": "\\tilde{I}_y" + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "content": " resolution. We describe our meta-learning scheme in Section 3.2, which allows fast adaption using a few images. In Section 3.3, we propose a spatial-temporal enhancement network that generally improves the perceptual quality for both the one-shot and personalized models, yielding video frames with " + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 45, + 512, + 279, + 679 + ], + "type": "text", + "content": " resolution, as shown in Figure 2(d)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 687, + 232, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 687, + 232, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 687, + 232, + 700 + ], + "type": "text", + "content": "3.1. ID-preserving One-shot Base Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "In this section, we will introduce our warping network using dense facial landmarks and an identity-aware refine-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 376, + 494, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 376, + 494, + 388 + ], + "spans": [ + { + "bbox": [ + 296, + 376, + 494, + 388 + ], + "type": "text", + "content": "ment network for one-shot talking head synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 389, + 530, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 389, + 530, + 505 + ], + "spans": [ + { + "bbox": [ + 296, + 389, + 530, + 505 + ], + "type": "text", + "content": "Warping prediction with dense landmarks. To predict an accurate geometry-aware warping field, we claim that dense landmark prediction [43] is the key to a geometry-aware warping field estimation. While dense facial landmarks are tedious to annotate, our dense landmark prediction is trained on synthetic faces [42], which reliably produces 669 points covering the entire head, including the ears, eyeballs, and teeth, for in-the-wild faces. These dense landmarks capture rich information about the person's facial geometry and considerably ease the flow field prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 506, + 530, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 506, + 530, + 612 + ], + "spans": [ + { + "bbox": [ + 296, + 506, + 530, + 612 + ], + "type": "text", + "content": "However, it is non-trivial to fully make use of these dense landmarks. A naive approach is to channel-wise concatenate the landmarks before feeding into the network, as previous works [33, 34, 39]. However, processing such input is computationally demanding due to the inordinate number of input channels. Hence, we propose an efficient way to digest these landmarks. Specifically, We draw the neighboring connected landmark points, with each connection encoded in different colors as shown in Figure 2(a)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "text", + "content": "One can thus take the landmark images of the source and driving, along with the source image, i.e., " + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{in}} = \\mathrm{Concat}(I_s, I_s^{\\mathrm{ldmk}}, I_d^{\\mathrm{ldmk}})" + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "text", + "content": ", for warping field prediction. To ensure a globally coherent prediction, we strengthen the warping capability using the condition of a latent motion code " + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_w" + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "text", + "content": " which is derived from the input, i.e., " + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_w = E_w(\\boldsymbol{x}_{\\mathrm{in}})" + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "inline_equation", + "content": "E_w" + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "text", + "content": " is a CNN encoder. The motion code " + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_w" + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "text", + "content": " is injected into the flow estimation network " + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "inline_equation", + "content": "F_w" + }, + { + "bbox": [ + 296, + 612, + 531, + 730 + ], + "type": "text", + "content": " through AdaIN [15]. By modulating the mean and variance of the normalized feature map, the network could be effectively" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "22098" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "text", + "content": "guided by the motion vector which induces a globally coherent flow prediction. Formally, we obtain the prediction of a dense flow field through " + }, + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w} = F_{w}(\\boldsymbol{x}_{\\mathrm{in}},\\boldsymbol{z}_{w})" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 141, + 279, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 141, + 279, + 257 + ], + "spans": [ + { + "bbox": [ + 45, + 141, + 279, + 257 + ], + "type": "text", + "content": "ID-preserving refinement network. Directly warping the source image with the predicted flow field inevitably introduces artifacts and the loss of subtle perceived identity. Therefore, an ID-preserving refinement network is needed to produce a photo-realistic result while maintaining the identity of the source image. Prior works primarily focus on the geometry-aware flow field prediction, whereas such deformation may not well characterize the fine-grained facial geometry. In this work, we resolve the identity loss via a well-designed identity-preserving refinement network." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "spans": [ + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": "We propose to attentionally incorporate the semantic identity vector with the intermediate warping results. Let " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{in}}^l" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": "-th layer feature map of the refinement network. We obtain the identity-aware feature output " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{id}}^l" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": " by modulating " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{in}}^l" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": " with the identity embedding " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{id}}" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": " through AdaIN, where " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{id}}" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": " is extracted using a pre-trained face recognition model " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{id}}" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": " [5]. Meanwhile, we obtain a motion-aware feature " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{motion}}^l" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": ", which keeps the head motion and expression of the driving video. Specifically, " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{motion}}^l" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": " is obtained via a Feature-wise Linear Modulate (FiLM) [8, 29, 30] according to the warped image " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "\\tilde{I}_s = w(I_s)" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{motion}}^l = \\operatorname{Conv}(\\tilde{I}_s) \\times h_{\\mathrm{id}}^l + \\operatorname{Conv}(\\tilde{I}_s)" + }, + { + "bbox": [ + 45, + 258, + 279, + 398 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 399, + 279, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 399, + 279, + 526 + ], + "spans": [ + { + "bbox": [ + 45, + 399, + 279, + 526 + ], + "type": "text", + "content": "With both the identity-aware and motion-aware features, we adaptively fuse the features through an attention-based fusion block. Inspired by recent face-swapping approaches [23], we suppose that the driving motions and source identity should be fused in a spatially-adaptive manner, in which the facial parts that mostly characterize the key facial features express the identity should rely more on the identity-aware feature, whereas other parts (e.g. hair and clothes) should make greater use of the motion-aware feature. A learnable fusion mask " + }, + { + "bbox": [ + 45, + 399, + 279, + 526 + ], + "type": "inline_equation", + "content": "M^l" + }, + { + "bbox": [ + 45, + 399, + 279, + 526 + ], + "type": "text", + "content": " is used for the fusion of these two parts, which is predicted by," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 532, + 279, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 532, + 279, + 545 + ], + "spans": [ + { + "bbox": [ + 119, + 532, + 279, + 545 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} ^ {l} = \\sigma \\left(\\operatorname {C o n v} \\left(\\boldsymbol {h} _ {\\text {i n}} ^ {l}\\right)\\right), \\tag {1}", + "image_path": "3986132ca99fa9832c9f6b83d465cfa6f923e094a21037c32b2fad4145e9d9f4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "spans": [ + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "text", + "content": " indicates the sigmoid activation function. In this way, the model learns to properly inject the identity-aware features into identity-related regions. The output of layer " + }, + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "text", + "content": " can be derived by fusing features according to the mask " + }, + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "inline_equation", + "content": "M^l" + }, + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "text", + "content": ", which is," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 615, + 279, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 615, + 279, + 630 + ], + "spans": [ + { + "bbox": [ + 80, + 615, + 279, + 630 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} _ {\\text {o u t}} ^ {l} = \\boldsymbol {M} ^ {l} \\otimes \\boldsymbol {h} _ {\\text {m o t i o n}} ^ {l} + (1 - \\boldsymbol {M} ^ {l}) \\otimes \\boldsymbol {h} _ {\\text {i d}} ^ {l}, \\tag {2}", + "image_path": "f4ffbdf23fc3c02e90a08c05a42fa3741b62c9419745915cbf61b10aaf3122b3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 636, + 279, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 636, + 279, + 683 + ], + "spans": [ + { + "bbox": [ + 45, + 636, + 279, + 683 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 636, + 279, + 683 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 45, + 636, + 279, + 683 + ], + "type": "text", + "content": " denotes the Hadamard product. Through a cascade of such blocks, we obtain the final output image " + }, + { + "bbox": [ + 45, + 636, + 279, + 683 + ], + "type": "inline_equation", + "content": "\\tilde{I}_y" + }, + { + "bbox": [ + 45, + 636, + 279, + 683 + ], + "type": "text", + "content": ", which well preserves the source identity while accurately following the head motion and expression as the driving person." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "text", + "content": "Training objective. Perceptual loss [16] is computed between the warped source image " + }, + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\tilde{I}_s" + }, + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "text", + "content": " and ground-truth driving image " + }, + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "text", + "content": " for accurate warping prediction. The same loss is also applied to enforce the refinement output " + }, + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\tilde{I}_y" + }, + { + "bbox": [ + 45, + 683, + 279, + 730 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 105, + 507, + 269 + ], + "blocks": [ + { + "bbox": [ + 322, + 105, + 507, + 269 + ], + "lines": [ + { + "bbox": [ + 322, + 105, + 507, + 269 + ], + "spans": [ + { + "bbox": [ + 322, + 105, + 507, + 269 + ], + "type": "image", + "image_path": "3b11bbe1f78cb753bc77a4b297b679fbe6ab85449c6118d7a0cbd3c93aaef216.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "lines": [ + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "spans": [ + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "text", + "content": "Figure 3. Visualization of meta-learning. Compared with the one-shot pre-trained model " + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "text", + "content": ", the meta-learned model " + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}" + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "text", + "content": " can be rapidly adapted to a unique personal model " + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_j^K" + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 296, + 273, + 530, + 306 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "spans": [ + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "type": "text", + "content": "We also extract the feature using face recognition model " + }, + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{id}}" + }, + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "type": "text", + "content": " [5], and penalize the dissimilarity between the ID vectors of the output image " + }, + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "type": "inline_equation", + "content": "\\tilde{I}_y" + }, + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "type": "text", + "content": " and source " + }, + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 297, + 314, + 530, + 349 + ], + "type": "text", + "content": ", using" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 346, + 360, + 530, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 360, + 530, + 374 + ], + "spans": [ + { + "bbox": [ + 346, + 360, + 530, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {i d}} = 1 - \\cos \\left(E _ {\\mathrm {i d}} \\left(\\tilde {I} _ {y}\\right), E _ {\\mathrm {i d}} \\left(I _ {s}\\right)\\right). \\tag {3}", + "image_path": "60b3630f320e848a1920114c56510c9310b1f60e4cc287fd64fd5abaaf90f700.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "spans": [ + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "type": "text", + "content": "A multi-scale patch discriminator " + }, + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "type": "text", + "content": " [29] is adopted to enhance the photo-realism of the outputs. To further improve the generation quality on the hard eye and mouth areas, we add additional " + }, + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "type": "text", + "content": " reconstruction losses, i.e., " + }, + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eye}}, \\mathcal{L}_{\\mathrm{mouth}}" + }, + { + "bbox": [ + 296, + 383, + 530, + 442 + ], + "type": "text", + "content": " for these parts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 443, + 489, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 489, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 489, + 454 + ], + "type": "text", + "content": "The overall training loss can be formulated as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 333, + 463, + 530, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 463, + 530, + 483 + ], + "spans": [ + { + "bbox": [ + 333, + 463, + 530, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {w}} ^ {\\mathrm {V G G}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}} ^ {\\mathrm {V G G}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}} + \\tag {4}", + "image_path": "733140958fa950b62ede8d795726b0ad2040db2972cf94d41f8e48f9e0f1f5df.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 481, + 492, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 481, + 492, + 492 + ], + "spans": [ + { + "bbox": [ + 351, + 481, + 492, + 492 + ], + "type": "interline_equation", + "content": "\\lambda_ {\\text {e y e}} \\mathcal {L} _ {\\text {e y e}} + \\lambda_ {\\text {m o u t h}} \\mathcal {L} _ {\\text {m o u t h}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}},", + "image_path": "47ecf5d8c991c3da65e11e8793ccef546fca67b8267c810b2541a1b7300eaeee.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 503, + 515, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 503, + 515, + 515 + ], + "spans": [ + { + "bbox": [ + 297, + 503, + 515, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 297, + 503, + 515, + 515 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{r}}, \\lambda_{\\mathrm{id}}, \\lambda_{\\mathrm{eye}}, \\lambda_{\\mathrm{mouth}}" + }, + { + "bbox": [ + 297, + 503, + 515, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 297, + 503, + 515, + 515 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{adv}}" + }, + { + "bbox": [ + 297, + 503, + 515, + 515 + ], + "type": "text", + "content": " are the loss weights." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 524, + 517, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 524, + 517, + 536 + ], + "spans": [ + { + "bbox": [ + 297, + 524, + 517, + 536 + ], + "type": "text", + "content": "3.2. Meta-learning based Faster Personalization" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 543, + 530, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 543, + 530, + 693 + ], + "spans": [ + { + "bbox": [ + 296, + 543, + 530, + 693 + ], + "type": "text", + "content": "While achieving state-of-the-art generation quality using our one-shot model, there always exists challenging cases that are intractable since the one-shot setting is inherently ill-posed. Person-specific characteristics and occlusion would never be faithfully recovered using a one-shot general model. Thus, personalized fine-tuning is necessary to achieve robust and authentic results that are truly usable. Nonetheless, fine-tuning the one-shot pre-trained model on a long video is computationally prohibitive for common users. To solve this, we propose a meta-learned model whose initialization weights can be easily adapted according to low-shot personal data within a few training steps, as illustrated in Figure 3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "content": "Formally, given a person " + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "content": ", the personalized adaptation starts from the pre-trained model weight " + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "content": " and aims to reach the optimal personal model weight " + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_j" + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "content": " by minimiz-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "22099" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 106, + 226, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 106, + 226, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 106, + 226, + 118 + ], + "type": "text", + "content": "ing the error against the personal images " + }, + { + "bbox": [ + 46, + 106, + 226, + 118 + ], + "type": "inline_equation", + "content": "\\hat{X}_j" + }, + { + "bbox": [ + 46, + 106, + 226, + 118 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 122, + 279, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 122, + 279, + 143 + ], + "spans": [ + { + "bbox": [ + 113, + 122, + 279, + 143 + ], + "type": "interline_equation", + "content": "\\hat {\\phi} _ {j} = \\min _ {\\phi_ {j}} \\mathcal {L} \\left(G _ {\\phi_ {j}} \\left(\\hat {\\boldsymbol {X}} _ {j}\\right)\\right), \\tag {5}", + "image_path": "078f04aa81e1db6e834b0470259b37962075dea3f4793302cbfa37fd23c583fe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "spans": [ + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "inline_equation", + "content": "G_{\\phi_j}" + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "text", + "content": " denotes the whole generator with weight " + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "text", + "content": ". Usually we perform " + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "text", + "content": " steps of stochastic gradient descent (SGD) from initialization " + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "text", + "content": " to approach " + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_j" + }, + { + "bbox": [ + 45, + 146, + 279, + 193 + ], + "type": "text", + "content": ", so the weight updating process can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 197, + 279, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 197, + 279, + 212 + ], + "spans": [ + { + "bbox": [ + 117, + 197, + 279, + 212 + ], + "type": "interline_equation", + "content": "\\phi_ {j} ^ {k} = \\operatorname {S G D} _ {K} (\\phi , \\hat {\\boldsymbol {X}} _ {j}). \\tag {6}", + "image_path": "dda643b2cafd52a8b795d2b9e65e4fe97be155f4140de98db16cb79ea0ab1a32.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "spans": [ + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "type": "text", + "content": "Our goal is to find an optimal initialization " + }, + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}^K" + }, + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "type": "text", + "content": " which could approach any personal model after " + }, + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "type": "text", + "content": " steps of SGD update, even for a small " + }, + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 216, + 279, + 252 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 255, + 279, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 255, + 279, + 288 + ], + "spans": [ + { + "bbox": [ + 107, + 255, + 279, + 288 + ], + "type": "interline_equation", + "content": "\\hat {\\phi} ^ {K} = \\min _ {\\phi} \\sum_ {j = 1} ^ {M} \\| \\hat {\\phi} _ {j} - \\phi_ {j} ^ {k} \\|. \\tag {7}", + "image_path": "da66fbebc1a580f4849de8c15b8bcd689c1dcfb737809d674f6e1b49d4ba07f1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "spans": [ + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "text", + "content": "Indeed, the general one-shot pretrained model " + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "text", + "content": " is a special case in the above formulation, which essentially learns the model weight in Equation 7 when " + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "text", + "content": ". When we are allowed to perform a few adaption steps (" + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "inline_equation", + "content": "K > 0" + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "text", + "content": "), there is a gap between " + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "text", + "content": " and desired " + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}^K" + }, + { + "bbox": [ + 45, + 293, + 279, + 385 + ], + "type": "text", + "content": ", since the optimization target of the standard pre-training is to minimize the overall error across all training data, it does not necessarily find the best weight suitable for personalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 385, + 279, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 385, + 279, + 524 + ], + "spans": [ + { + "bbox": [ + 45, + 385, + 279, + 524 + ], + "type": "text", + "content": "Compared with general one-shot models, we leverage the idea of Model-Agnostic Meta-Learning (MAML) to bridge this gap and enable surprisingly fast personalized training. The goal of MAML-based methods is to optimize the initialized weights such that they could be fast adapted to a new identity within a few steps of gradient descent, which directly matches our goal. Directly optimizing the initialization weight using Equation (7) involves the computation of second-order derivatives, which is computationally expensive on large-scale training. Therefore, we utilize Reptile [26], a first-order MAML-based approach to obtain suitable initialization for fast personalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "spans": [ + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "text", + "content": "To be more specific, our meta-learning model explicitly considers the personalized adaptation during training. For each person " + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "text", + "content": ", we start from the " + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "inline_equation", + "content": "\\phi_j^0 = \\phi" + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "text", + "content": ", which is the initialization to be optimized. Formally, we sample a batch of personal training data " + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "inline_equation", + "content": "\\hat{X}_j" + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 524, + 279, + 595 + ], + "type": "text", + "content": " steps of personalized training yield the finetuned model weights as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 79, + 599, + 279, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 599, + 279, + 618 + ], + "spans": [ + { + "bbox": [ + 79, + 599, + 279, + 618 + ], + "type": "interline_equation", + "content": "\\phi_ {j} ^ {k} = \\operatorname {S G D} \\left(\\phi_ {j} ^ {k - 1}, \\hat {\\boldsymbol {X}} _ {j}\\right), k = 1, \\dots , K. \\tag {8}", + "image_path": "8ec73274b2865e5ebc97f4f258d2e07da493ce21dfbf7a5125b7a6ff5e209a1e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "spans": [ + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "type": "text", + "content": "Finally, the personal update, i.e., the difference of " + }, + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "type": "inline_equation", + "content": "\\phi_j^K" + }, + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "type": "inline_equation", + "content": "\\phi_j^0" + }, + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "type": "text", + "content": ", is used as the gradient to update our initialization " + }, + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 45, + 622, + 279, + 649 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 653, + 279, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 653, + 279, + 667 + ], + "spans": [ + { + "bbox": [ + 113, + 653, + 279, + 667 + ], + "type": "interline_equation", + "content": "\\phi \\leftarrow \\phi - \\beta \\left(\\phi_ {j} ^ {K} - \\phi\\right), \\tag {9}", + "image_path": "12764bf84e38b89d02ff68fd804f389a7daf8d978a3d287d48b9863d3f5347a1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": " is the meta-learning rate. The full algorithm is shown in Algorithm 1. Our model progressively learns a more suitable initialization through meta-training as visualized in Figure 3, which could fast adapt to personalized models after limited steps of adaptation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 298, + 106, + 525, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 106, + 525, + 117 + ], + "spans": [ + { + "bbox": [ + 298, + 106, + 525, + 117 + ], + "type": "text", + "content": "Algorithm 1 Optimization of Initial Weights with Reptile" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "spans": [ + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "text", + "content": "Input: weights " + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "text", + "content": " of generation network " + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "text", + "content": ", inner loop learning rate " + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "text", + "content": ", meta-learning rate " + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "text", + "content": ", number of training iterations " + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "text", + "content": ", number of training persons " + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "text", + "content": ", number of inner loop iterations " + }, + { + "bbox": [ + 311, + 120, + 530, + 162 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 164, + 479, + 270 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 302, + 164, + 389, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 164, + 389, + 173 + ], + "spans": [ + { + "bbox": [ + 302, + 164, + 389, + 173 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 302, + 164, + 389, + 173 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, N" + }, + { + "bbox": [ + 302, + 164, + 389, + 173 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 175, + 404, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 175, + 404, + 184 + ], + "spans": [ + { + "bbox": [ + 303, + 175, + 404, + 184 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 303, + 175, + 404, + 184 + ], + "type": "inline_equation", + "content": "j = 1,\\dots ,M" + }, + { + "bbox": [ + 303, + 175, + 404, + 184 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 303, + 185, + 398, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 185, + 398, + 196 + ], + "spans": [ + { + "bbox": [ + 303, + 185, + 398, + 196 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 303, + 185, + 398, + 196 + ], + "type": "inline_equation", + "content": "\\phi_j^0 = \\mathbf{Clone}(\\phi)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 303, + 197, + 479, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 197, + 479, + 207 + ], + "spans": [ + { + "bbox": [ + 303, + 197, + 479, + 207 + ], + "type": "text", + "content": "4: Sample a training batch " + }, + { + "bbox": [ + 303, + 197, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\hat{X}_j" + }, + { + "bbox": [ + 303, + 197, + 479, + 207 + ], + "type": "text", + "content": " of person " + }, + { + "bbox": [ + 303, + 197, + 479, + 207 + ], + "type": "inline_equation", + "content": "j" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 303, + 207, + 417, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 207, + 417, + 217 + ], + "spans": [ + { + "bbox": [ + 303, + 207, + 417, + 217 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 303, + 207, + 417, + 217 + ], + "type": "inline_equation", + "content": "k = 1,\\dots ,K" + }, + { + "bbox": [ + 303, + 207, + 417, + 217 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 303, + 217, + 471, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 217, + 471, + 229 + ], + "spans": [ + { + "bbox": [ + 303, + 217, + 471, + 229 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 303, + 217, + 471, + 229 + ], + "type": "inline_equation", + "content": "\\phi_j^k = \\phi_j^{k - 1} - \\alpha \\nabla_\\phi \\mathcal{L}(G_\\phi (\\hat{\\mathbf{X}}_j))" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 303, + 229, + 368, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 229, + 368, + 238 + ], + "spans": [ + { + "bbox": [ + 303, + 229, + 368, + 238 + ], + "type": "text", + "content": "7: end for" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 303, + 239, + 355, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 239, + 355, + 248 + ], + "spans": [ + { + "bbox": [ + 303, + 239, + 355, + 248 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 303, + 249, + 436, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 249, + 436, + 262 + ], + "spans": [ + { + "bbox": [ + 303, + 249, + 436, + 262 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 303, + 249, + 436, + 262 + ], + "type": "inline_equation", + "content": "\\phi \\gets \\phi -\\frac{\\beta}{M}\\sum_{j = 1}^{M}(\\phi_j^K -\\phi)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 262, + 342, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 262, + 342, + 270 + ], + "spans": [ + { + "bbox": [ + 299, + 262, + 342, + 270 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 295, + 530, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 295, + 530, + 308 + ], + "spans": [ + { + "bbox": [ + 297, + 295, + 530, + 308 + ], + "type": "text", + "content": "3.3. Temporal-consistent Super-resolution Network" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 296, + 314, + 530, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 314, + 530, + 431 + ], + "spans": [ + { + "bbox": [ + 296, + 314, + 530, + 431 + ], + "type": "text", + "content": "To further enhance the generation resolution and improve the high-fidelity details of our output, a video super-resolution network is needed as the final stage of the generation framework. Previous talking head synthesis works [7,47] utilize single-frame super-resolution as the last stage and ignore the quality of temporal consistency and stability. Performing super-resolution in frame-by-frame manner tends to produce texture flickering which severely hampers the visual quality. In this work, we consider multiple adjacent frames to ensure temporal coherency." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "spans": [ + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "type": "text", + "content": "Inspired by previous 2D face restoration works [40, 46], the pre-trained generative models [17, 18, 50] like StyleGAN contain rich face prior and could significantly help to enhance the high-frequency details. Moreover, the disentangled " + }, + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "type": "text", + "content": " space in StyleGAN provides desirable temporal consistency during manipulation [37], which also benefits our framework. Therefore, we propose a temporally consistent super-resolution module by leveraging pretrained StyleGAN and 3D convolution, where the latter brings quality enhancement in spatio-temporal domain. As shown in Fig. 2, we feed the concatenated sequence of " + }, + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "type": "text", + "content": " video frames " + }, + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "type": "inline_equation", + "content": "\\{\\tilde{I}_y^1,\\tilde{I}_y^2,\\dots,\\tilde{I}_y^t\\}" + }, + { + "bbox": [ + 296, + 431, + 530, + 618 + ], + "type": "text", + "content": " into a U-Net composed of 3D convolution with reflection padding on the temporal dimension. To ensure pre-trained per-frame quality while improving temporal consistency, we initialize the 3D convolution weight in U-Net with a pretrained 2D face restoration network [40]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 296, + 619, + 530, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 619, + 530, + 668 + ], + "spans": [ + { + "bbox": [ + 296, + 619, + 530, + 668 + ], + "type": "text", + "content": "These spatio-temporally enhanced features from the U-Net decoder further modulate the pretrained StyleGAN features through FiLM. Thus, the super-resolution frames " + }, + { + "bbox": [ + 296, + 619, + 530, + 668 + ], + "type": "inline_equation", + "content": "\\{I_y^{1}, I_y^{2}, \\dots, I_y^{t}\\}" + }, + { + "bbox": [ + 296, + 619, + 530, + 668 + ], + "type": "text", + "content": " are obtained as:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 678, + 530, + 704 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 678, + 530, + 704 + ], + "spans": [ + { + "bbox": [ + 307, + 678, + 530, + 704 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {I} _ {y} ^ {\\prime 1}, \\boldsymbol {I} _ {y} ^ {\\prime 2}, \\dots , \\boldsymbol {I} _ {y} ^ {\\prime t} \\right\\} = F _ {\\text {S t y l e G A N}} \\left(F _ {\\text {3 D}} \\left(\\left\\{\\tilde {\\boldsymbol {I}} _ {y} ^ {1}, \\tilde {\\boldsymbol {I}} _ {y} ^ {2}, \\dots , \\tilde {\\boldsymbol {I}} _ {y} ^ {t} \\right\\}\\right)\\right). \\tag {10}", + "image_path": "5ff91f3b516d42881c023481ce0e0cf2bcdc2fd1a66b51da4cadb1fde10abf3d.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": "During training, we optimize the output " + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "inline_equation", + "content": "I_{y}^{\\prime}" + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": " towards the " + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": " ground truth " + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\ell_{1}" + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": " and perceptual loss." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "22100" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 103, + 522, + 318 + ], + "blocks": [ + { + "bbox": [ + 55, + 103, + 522, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 103, + 522, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 522, + 318 + ], + "type": "image", + "image_path": "94cc2c767e91afe306b209cc76e33a36b8d607f85e27bcdedf820a6cdcdf1e5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 319, + 531, + 342 + ], + "lines": [ + { + "bbox": [ + 45, + 319, + 531, + 342 + ], + "spans": [ + { + "bbox": [ + 45, + 319, + 531, + 342 + ], + "type": "text", + "content": "Figure 4. Qualitative results of our self-reconstruction (top row) and cross-identity reenactment (bottom two rows) at " + }, + { + "bbox": [ + 45, + 319, + 531, + 342 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 45, + 319, + 531, + 342 + ], + "type": "text", + "content": " resolution. Our method synthesizes more faithful expression and motion, while better preserving the identity of the source portrait." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 351, + 523, + 554 + ], + "blocks": [ + { + "bbox": [ + 52, + 351, + 523, + 554 + ], + "lines": [ + { + "bbox": [ + 52, + 351, + 523, + 554 + ], + "spans": [ + { + "bbox": [ + 52, + 351, + 523, + 554 + ], + "type": "image", + "image_path": "5540f3c19d16e01b241b0346ed41d2accd1944429397f622ab27bd9fd5e87cb6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 557, + 531, + 580 + ], + "lines": [ + { + "bbox": [ + 45, + 557, + 531, + 580 + ], + "spans": [ + { + "bbox": [ + 45, + 557, + 531, + 580 + ], + "type": "text", + "content": "Figure 5. Qualitative results of self reconstruction and cross-identity reenactment at " + }, + { + "bbox": [ + 45, + 557, + 531, + 580 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 45, + 557, + 531, + 580 + ], + "type": "text", + "content": " resolution. Our spatial-temporal superresolution module further enhances more high-frequency details on teeth, eyes, and hair." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 598, + 124, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 598, + 124, + 611 + ], + "spans": [ + { + "bbox": [ + 45, + 598, + 124, + 611 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 618, + 151, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 618, + 151, + 631 + ], + "spans": [ + { + "bbox": [ + 45, + 618, + 151, + 631 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": "Dataset. Following [7], we train our warping and refinement networks on cropped VoxCeleb2 dataset [4] at " + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": " resolution. We randomly select 500 videos from the test set for evaluation. For our meta-learning-based fast personalization, we finetune our base model on HDTF dataset [52], which is composed of 410 videos from 300 different identities. We downsample cropped faces to " + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": " resolution and split the original HDTF dataset into 400 training videos and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "spans": [ + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "type": "text", + "content": "10 test videos. After the convergence of our meta-training, we further evaluate the personalization speed of our model on the HDTF test set. Our temporal super-resolution module " + }, + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "type": "inline_equation", + "content": "F_{3d}" + }, + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "type": "text", + "content": " is trained on the HDTF dataset [52] which has 300 frames per video with " + }, + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "type": "text", + "content": " resolution. We feed downsampled " + }, + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 296, + 599, + 530, + 704 + ], + "type": "text", + "content": " frames into fixed warping and refinement network and use the outputs of the refinement network as inputs for the next temporal super-resolution module. More training details are provided in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "type": "text", + "content": "Metrics We evaluate the fidelity of self-reconstruction using FID [14] and LPIPS [51]. Our motion transfer qual" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 299, + 756 + ], + "type": "text", + "content": "22101" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 104, + 490, + 233 + ], + "blocks": [ + { + "bbox": [ + 89, + 104, + 490, + 233 + ], + "lines": [ + { + "bbox": [ + 89, + 104, + 490, + 233 + ], + "spans": [ + { + "bbox": [ + 89, + 104, + 490, + 233 + ], + "type": "table", + "html": "
MethodsSelf Reconstruction (256 × 256)Cross Reenactment (256 × 256)
FID↓LPIPS↓ID Loss↓AED↓APD↓FID↓ID Loss↓AED↓APD↓
X2Face [41]45.29080.68060.96320.21470.100791.14850.64960.31120.1210
Bi-layer [48]100.91960.58810.52800.12580.0139127.78230.63360.23300.0208
FOMM [33]12.19790.23380.20960.09640.010080.16370.57600.23400.0239
PIRender [31]14.40650.26390.30240.10800.016278.84300.54400.21130.0214
Ours11.95280.22620.12960.09420.012477.50480.29440.25240.0258
MethodsSelf Reconstruction (512 × 512)Cross Reenactment (512 × 512)
StyleHEAT [47]44.52070.28400.41120.11550.0131111.34500.47200.25050.0218
Ours21.49740.20790.08320.09040.012149.60200.19520.27370.0242
", + "image_path": "d275d5bc1da0f88561e454ea5f9fc7d5ca56b9d60453ee8c358c4cfd72eed30d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 234, + 531, + 257 + ], + "lines": [ + { + "bbox": [ + 45, + 234, + 531, + 257 + ], + "spans": [ + { + "bbox": [ + 45, + 234, + 531, + 257 + ], + "type": "text", + "content": "Table 1. Quantitative results for self-reconstruction and cross-reenactment. We evaluate both " + }, + { + "bbox": [ + 45, + 234, + 531, + 257 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 45, + 234, + 531, + 257 + ], + "type": "text", + "content": " results and " + }, + { + "bbox": [ + 45, + 234, + 531, + 257 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 45, + 234, + 531, + 257 + ], + "type": "text", + "content": " results. Our method outperforms all baselines on both resolutions across image fidelity metrics with comparable motion transfer results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 47, + 266, + 142, + 459 + ], + "blocks": [ + { + "bbox": [ + 47, + 266, + 142, + 459 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 142, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 142, + 459 + ], + "type": "image", + "image_path": "ab90219bc083f4c5b1c8491790e86148f1237ab2322c7b00d45a794b519c3f84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 462, + 106, + 470 + ], + "lines": [ + { + "bbox": [ + 82, + 462, + 106, + 470 + ], + "spans": [ + { + "bbox": [ + 82, + 462, + 106, + 470 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 144, + 266, + 239, + 362 + ], + "blocks": [ + { + "bbox": [ + 144, + 266, + 239, + 362 + ], + "lines": [ + { + "bbox": [ + 144, + 266, + 239, + 362 + ], + "spans": [ + { + "bbox": [ + 144, + 266, + 239, + 362 + ], + "type": "image", + "image_path": "5f2433757bdb30aafc52f44dbf9cc74d9336e69f8405b94732f681aaa23e413b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 144, + 364, + 239, + 458 + ], + "blocks": [ + { + "bbox": [ + 144, + 364, + 239, + 458 + ], + "lines": [ + { + "bbox": [ + 144, + 364, + 239, + 458 + ], + "spans": [ + { + "bbox": [ + 144, + 364, + 239, + 458 + ], + "type": "image", + "image_path": "d9a23ba10ea9dbf8347ee038be83ad616905e906646aacb308d65c56707282b2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 461, + 201, + 470 + ], + "lines": [ + { + "bbox": [ + 174, + 461, + 201, + 470 + ], + "spans": [ + { + "bbox": [ + 174, + 461, + 201, + 470 + ], + "type": "text", + "content": "Driving" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 242, + 266, + 337, + 362 + ], + "blocks": [ + { + "bbox": [ + 242, + 266, + 337, + 362 + ], + "lines": [ + { + "bbox": [ + 242, + 266, + 337, + 362 + ], + "spans": [ + { + "bbox": [ + 242, + 266, + 337, + 362 + ], + "type": "image", + "image_path": "b7da9b6081c1637b0b43d8e2463e80b3bc9331bce43c22bcffcf59377273974e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 242, + 364, + 337, + 459 + ], + "blocks": [ + { + "bbox": [ + 242, + 364, + 337, + 459 + ], + "lines": [ + { + "bbox": [ + 242, + 364, + 337, + 459 + ], + "spans": [ + { + "bbox": [ + 242, + 364, + 337, + 459 + ], + "type": "image", + "image_path": "3e35a835034b52e24561264ffb79341570e3cd5f52dd063440e8fb5021a30db9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 462, + 297, + 470 + ], + "lines": [ + { + "bbox": [ + 271, + 462, + 297, + 470 + ], + "spans": [ + { + "bbox": [ + 271, + 462, + 297, + 470 + ], + "type": "text", + "content": "FOMM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 475, + 531, + 508 + ], + "lines": [ + { + "bbox": [ + 45, + 475, + 531, + 508 + ], + "spans": [ + { + "bbox": [ + 45, + 475, + 531, + 508 + ], + "type": "text", + "content": "Figure 6. A comparison of different personalized models at the same epochs. Our personalized base model without reptile adaptation reconstructs more accurate skin color and source identity than the personalized FOMM method. For the model personalized from a reptile learning stage, finer detail on teeth and eye colors could be further generated." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 338, + 266, + 434, + 362 + ], + "blocks": [ + { + "bbox": [ + 338, + 266, + 434, + 362 + ], + "lines": [ + { + "bbox": [ + 338, + 266, + 434, + 362 + ], + "spans": [ + { + "bbox": [ + 338, + 266, + 434, + 362 + ], + "type": "image", + "image_path": "0a648b25a52e7f060dd462def80a6abe88931461ca07db5507de4a2cd059c511.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 338, + 364, + 434, + 459 + ], + "blocks": [ + { + "bbox": [ + 338, + 364, + 434, + 459 + ], + "lines": [ + { + "bbox": [ + 338, + 364, + 434, + 459 + ], + "spans": [ + { + "bbox": [ + 338, + 364, + 434, + 459 + ], + "type": "image", + "image_path": "a359a42399421bb22e3dbe8135384ce6ac15e41ea08414e9e31aa4f50298d507.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 462, + 412, + 471 + ], + "lines": [ + { + "bbox": [ + 356, + 462, + 412, + 471 + ], + "spans": [ + { + "bbox": [ + 356, + 462, + 412, + 471 + ], + "type": "text", + "content": "Ours w/o Reptile" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 266, + 531, + 362 + ], + "blocks": [ + { + "bbox": [ + 436, + 266, + 531, + 362 + ], + "lines": [ + { + "bbox": [ + 436, + 266, + 531, + 362 + ], + "spans": [ + { + "bbox": [ + 436, + 266, + 531, + 362 + ], + "type": "image", + "image_path": "7c35b06cb76ccf71e9e40a3aab9b90d3fc2f87fc0735b80b9e1b780062237e55.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 436, + 364, + 531, + 459 + ], + "blocks": [ + { + "bbox": [ + 436, + 364, + 531, + 459 + ], + "lines": [ + { + "bbox": [ + 436, + 364, + 531, + 459 + ], + "spans": [ + { + "bbox": [ + 436, + 364, + 531, + 459 + ], + "type": "image", + "image_path": "28e1446c407015a50c511bd3a9560809daa94b73ae5c70356e3cf95cb9e5128f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 461, + 503, + 470 + ], + "lines": [ + { + "bbox": [ + 452, + 461, + 503, + 470 + ], + "spans": [ + { + "bbox": [ + 452, + 461, + 503, + 470 + ], + "type": "text", + "content": "Ours w/ Reptile" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "spans": [ + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "type": "text", + "content": "ity is measured using average expression distance (AED) and average pose distance (APD) with driving videos. For our meta-learning-based fast personalization, we illustrate the LPIPS at different personalization epochs for each fine-tuning approach. Following previous works [21, 22], we evaluate our temporal consistency using warping error " + }, + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{warp}}" + }, + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "type": "text", + "content": ". For each frame " + }, + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "type": "inline_equation", + "content": "\\tilde{I}_y^t" + }, + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "type": "text", + "content": ", we calculate the warping error with the previous frame " + }, + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "type": "inline_equation", + "content": "\\tilde{I}_y^{t - 1}" + }, + { + "bbox": [ + 45, + 527, + 279, + 634 + ], + "type": "text", + "content": " warped by the estimated optical flow in the occlusion map." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 45, + 641, + 262, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 641, + 262, + 654 + ], + "spans": [ + { + "bbox": [ + 45, + 641, + 262, + 654 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "We compare our warping and refinement models at " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " resolution against several state-of-the-art face reenactment works: X2Face [41], Bi-Layer [1], First-Order Motion Model [33], and PIRender [31]. The top row of Figure 4 presents our qualitative results of self-reconstruction. Different from sparse landmarks in unsupervised learning [33]" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 527, + 530, + 726 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 527, + 530, + 726 + ], + "spans": [ + { + "bbox": [ + 296, + 527, + 530, + 726 + ], + "type": "text", + "content": "or 1D 3DMM [31], our dense landmarks provide strong guidance to accurate and fine-detailed synthesis on gaze, mouth, and expressions. The last two rows show the results of our cross-identity reenactment. Since our landmarks have a better decomposition of identity and motion and our refinement network is identity-aware, our method is the only one that well preserves the identity of source image. In contrast, previous methods suffer from appearance data leakage directly from the driver and generate faces with a similar identity to the driving image. Note that the 3DMM coefficients of AED and APD evaluation are not fully decomposed. Thus, other baselines with identity leakage may have better quantitative metrics. We compare our full framework with the proposed temporal super-resolution module at " + }, + { + "bbox": [ + 296, + 527, + 530, + 726 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 296, + 527, + 530, + 726 + ], + "type": "text", + "content": " resolution against StyleHEAT [47], which is the only open-sourced method that generates high-resolution talking heads. StyleHEAT [47] fails to synthesize" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "type": "text", + "content": "22102" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 103, + 254, + 251 + ], + "blocks": [ + { + "bbox": [ + 72, + 103, + 254, + 251 + ], + "lines": [ + { + "bbox": [ + 72, + 103, + 254, + 251 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 254, + 251 + ], + "type": "image", + "image_path": "c660bb192bb6d895f0f779e58464f663b5ef9a7cc30632aa6d6d0445af0f80e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 256, + 280, + 291 + ], + "lines": [ + { + "bbox": [ + 45, + 256, + 280, + 291 + ], + "spans": [ + { + "bbox": [ + 45, + 256, + 280, + 291 + ], + "type": "text", + "content": "Figure 7. Comparison of test LPIPS at fine-tuning epochs for different approaches. Our reptile-based model achieves more than " + }, + { + "bbox": [ + 45, + 256, + 280, + 291 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 45, + 256, + 280, + 291 + ], + "type": "text", + "content": " speedup compared with the base model and previous baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 308, + 278, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 308, + 278, + 400 + ], + "spans": [ + { + "bbox": [ + 45, + 308, + 278, + 400 + ], + "type": "text", + "content": "sharp and accurate teeth in our experiment, and the identity of the output image is quite different from the source portrait due to the limitations of GAN inversion. In contrast, our refinement network is identity-aware and we leverage pretrained StyleGAN in our temporal super-resolution module to fully exploit its face prior knowledge. Figure 5 shows that our method produces sharp teeth and hairs, while bicubic-upsampled results are blurry with artifacts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 402, + 278, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 402, + 278, + 436 + ], + "spans": [ + { + "bbox": [ + 45, + 402, + 278, + 436 + ], + "type": "text", + "content": "Table 1 demonstrates that our method achieves the best quantitative fidelity and comparable motion transfer quality on both self-reconstruction and cross-identity reenactment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 442, + 224, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 442, + 224, + 454 + ], + "spans": [ + { + "bbox": [ + 45, + 442, + 224, + 454 + ], + "type": "text", + "content": "4.3. Evaluation of Fast Personalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 461, + 278, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 461, + 278, + 658 + ], + "spans": [ + { + "bbox": [ + 45, + 461, + 278, + 658 + ], + "type": "text", + "content": "Although our base model achieves state-of-the-art quality as a general model, there still exists ill-posedness in some cases. For example, it is very difficult to synthesize teeth according to a closed source mouth. Thus, industry products typically conduct personalization using fine-tuning strategy, which can be computationally expensive. To achieve faster convergence, we finetune our base model using a meta-learning strategy to provide a better weight initialization for the following personalization. In Fig 7, we evaluate the personalization speed of our metalearned model against our base model and previous baseline FOMM [33]. It takes our meta-learning model 0.5 epoch to decrease LPIPS to 0.14, which is " + }, + { + "bbox": [ + 45, + 461, + 278, + 658 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 45, + 461, + 278, + 658 + ], + "type": "text", + "content": " speedup against our base model, and " + }, + { + "bbox": [ + 45, + 461, + 278, + 658 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 45, + 461, + 278, + 658 + ], + "type": "text", + "content": " against FOMM. Figure 6 compares the personalization of our method and FOMM [33] at the same epoch, which illustrates our fast adaptation speed on ambiguous areas (e.g., teeth, eyes, and wrinkle details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 665, + 254, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 665, + 254, + 677 + ], + "spans": [ + { + "bbox": [ + 45, + 665, + 254, + 677 + ], + "type": "text", + "content": "4.4. Evaluation of Temporal Super-resolution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 682, + 278, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 682, + 278, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 682, + 278, + 730 + ], + "type": "text", + "content": "In Table 2, we also evaluate the performance of our temporal super-resolution using 2D image fidelity and warping error " + }, + { + "bbox": [ + 45, + 682, + 278, + 730 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 45, + 682, + 278, + 730 + ], + "type": "text", + "content": ". We train a 2D super-resolution baseline using GFPGAN [40]. The quantitative result in Table 2 shows" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 104, + 516, + 172 + ], + "blocks": [ + { + "bbox": [ + 313, + 104, + 516, + 172 + ], + "lines": [ + { + "bbox": [ + 313, + 104, + 516, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 516, + 172 + ], + "type": "table", + "html": "
MethodsFID↓LPIPS↓E_warp ↓
Ground Truth--0.0182
Base w/ bicubic25.57620.22850.0184
Base w/ GFPGAN [40]22.63510.21780.0242
Ours21.49740.20790.0213
", + "image_path": "48868c82f0208d8377c7b8a342433cb05c0eca439004424b97ac0f7ce162fba0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 216, + 530, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 216, + 530, + 298 + ], + "spans": [ + { + "bbox": [ + 296, + 216, + 530, + 298 + ], + "type": "text", + "content": "that although naive 2D super-resolution improves per-frame fidelity, it also brings more flickering and larger warping error (0.0242) than simple bicubic upsampling (0.0184). To achieve temporally coherent results, we combine a U-Net composed of 3D convolution with facial prior [18], which significantly reduces " + }, + { + "bbox": [ + 296, + 216, + 530, + 298 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 296, + 216, + 530, + 298 + ], + "type": "text", + "content": " of our final videos from 0.0242 to 0.0213, and preserves compelling 2D facial details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 305, + 455, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 305, + 455, + 318 + ], + "spans": [ + { + "bbox": [ + 297, + 305, + 455, + 318 + ], + "type": "text", + "content": "4.5. Ablation Study of Base Model" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 313, + 334, + 516, + 386 + ], + "blocks": [ + { + "bbox": [ + 297, + 174, + 530, + 195 + ], + "lines": [ + { + "bbox": [ + 297, + 174, + 530, + 195 + ], + "spans": [ + { + "bbox": [ + 297, + 174, + 530, + 195 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of our temporal super-resolution on self-reconstruction at " + }, + { + "bbox": [ + 297, + 174, + 530, + 195 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 297, + 174, + 530, + 195 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 334, + 516, + 386 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 516, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 516, + 386 + ], + "type": "table", + "html": "
MethodsFID↓LPIPS↓ID Loss↓
Ours Sparse Landmark14.31900.24850.1424
Ours w/o ID12.27360.22560.2144
Ours11.95280.22620.1296
", + "image_path": "473363f226a2139d7b95e0e99c0820a3b3b446b81d96b21a2fc2becc17548230.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 389, + 529, + 410 + ], + "lines": [ + { + "bbox": [ + 297, + 389, + 529, + 410 + ], + "spans": [ + { + "bbox": [ + 297, + 389, + 529, + 410 + ], + "type": "text", + "content": "Table 3. Quantitative ablation study of landmarks and ID coefficients in our base model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 423, + 530, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 423, + 530, + 505 + ], + "spans": [ + { + "bbox": [ + 296, + 423, + 530, + 505 + ], + "type": "text", + "content": "We conduct ablation studies to validate the effectiveness of our driving motion and source identity representation in our base model. If we replace our 669 dense landmarks with sparse landmarks, the LPIPS of warped source images degrades by 0.2. To evaluate our identity-aware refinement, the removal of the identity input causes significant increase of identity loss from 0.1296 to 0.2144." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 516, + 367, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 516, + 367, + 527 + ], + "spans": [ + { + "bbox": [ + 297, + 516, + 367, + 527 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 535, + 530, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 535, + 530, + 652 + ], + "spans": [ + { + "bbox": [ + 296, + 535, + 530, + 652 + ], + "type": "text", + "content": "We present a novel framework for identity-preserving one-shot talking head generation. To faithfully maintain the source ID, we propose to leverage accurate dense landmarks in the warping network and explicit source identity during refinement. Further, we significantly advance the applicability of personalized model by reducing its training to 30 seconds with meta-learning. Last but not least, we enhance the final resolution and temporal consistency with 3D convolution and generative prior. Comprehensive experiments demonstrate the state-of-the-art performance of our system." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 298, + 662, + 355, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 662, + 355, + 674 + ], + "spans": [ + { + "bbox": [ + 298, + 662, + 355, + 674 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 301, + 681, + 531, + 729 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 301, + 681, + 530, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 681, + 530, + 700 + ], + "spans": [ + { + "bbox": [ + 301, + 681, + 530, + 700 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 701, + 531, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 701, + 531, + 729 + ], + "spans": [ + { + "bbox": [ + 302, + 701, + 531, + 729 + ], + "type": "text", + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. international conference on computer graphics and interactive techniques, 1999. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "22103" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 107, + 279, + 728 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 51, + 107, + 279, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 107, + 279, + 135 + ], + "spans": [ + { + "bbox": [ + 51, + 107, + 279, + 135 + ], + "type": "text", + "content": "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponni-ahy, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In CVPR, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 51, + 136, + 279, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 136, + 279, + 164 + ], + "spans": [ + { + "bbox": [ + 51, + 136, + 279, + 164 + ], + "type": "text", + "content": "[4] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. conference of the international speech communication association, 2018. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 165, + 279, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 165, + 279, + 191 + ], + "spans": [ + { + "bbox": [ + 51, + 165, + 279, + 191 + ], + "type": "text", + "content": "[5] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In CVPR, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 192, + 279, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 192, + 279, + 220 + ], + "spans": [ + { + "bbox": [ + 51, + 192, + 279, + 220 + ], + "type": "text", + "content": "[6] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 221, + 279, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 221, + 279, + 257 + ], + "spans": [ + { + "bbox": [ + 51, + 221, + 279, + 257 + ], + "type": "text", + "content": "[7] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In ACM Multimedia, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 258, + 279, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 258, + 279, + 295 + ], + "spans": [ + { + "bbox": [ + 51, + 258, + 279, + 295 + ], + "type": "text", + "content": "[8] Vincent Dumoulin, Ethan Perez, Nathan Schucher, Florian Strub, Harm de Vries, Aaron Courville, and Yoshua Bengio. Feature-wise transformations. Distill, 2018. https://distill.pub/2018/features-wise-transformations.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 296, + 279, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 296, + 279, + 323 + ], + "spans": [ + { + "bbox": [ + 51, + 296, + 279, + 323 + ], + "type": "text", + "content": "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 324, + 279, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 279, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 279, + 361 + ], + "type": "text", + "content": "[10] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Trans. Graph., 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 362, + 279, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 279, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 279, + 389 + ], + "type": "text", + "content": "[11] Jiahao Geng, Tianjia Shao, Youyi Zheng, Yanlin Weng, and Kun Zhou. Warp-guided gans for single-photo facial animation. ACM Transactions on Graphics (TOG), 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 390, + 279, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 279, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 279, + 418 + ], + "type": "text", + "content": "[12] Xiaojie Guo, Siyuan Li, Jiawan Zhang, Jiayi Ma, Lin Ma, Wei Liu, and Haibin Ling. Pfld: A practical facial landmark detector. arXiv: Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 418, + 279, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 279, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 279, + 446 + ], + "type": "text", + "content": "[13] Sungjoo Ha, Martin Kersner, Beomsu Kim, Seokjun Seo, and Dongyoung Kim. Marionette: Few-shot face reenactment preserving identity of unseen targets. In AAAI, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 447, + 279, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 279, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 279, + 483 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 484, + 279, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 279, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 279, + 503 + ], + "type": "text", + "content": "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 504, + 279, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 279, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 279, + 522 + ], + "type": "text", + "content": "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 522, + 279, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 279, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 279, + 541 + ], + "type": "text", + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 541, + 279, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 279, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 279, + 569 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 570, + 279, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 279, + 606 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 279, + 606 + ], + "type": "text", + "content": "[19] Hyeongwoo Kim, Mohamed Elgharib, Hans-Peter Zollofer, Michael Seidel, Thabo Beeler, Christian Richardt, and Christian Theobalt. Neural style-preserving visual dubbing. ACM Transactions on Graphics (TOG), 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 607, + 279, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 279, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 279, + 644 + ], + "type": "text", + "content": "[20] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Nießner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. Association for Computing Machinery, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 644, + 279, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 279, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 279, + 672 + ], + "type": "text", + "content": "[21] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 673, + 279, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 279, + 692 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 279, + 692 + ], + "type": "text", + "content": "[22] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. In NeurIPS, 2020. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 692, + 279, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 692, + 279, + 728 + ], + "spans": [ + { + "bbox": [ + 47, + 692, + 279, + 728 + ], + "type": "text", + "content": "[23] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv: Computer Vision and Pattern Recognition, 2019. 2, 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 298, + 107, + 530, + 728 + ], + "type": "list", + "angle": 0, + "index": 43, + "blocks": [ + { + "bbox": [ + 298, + 107, + 530, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 107, + 530, + 127 + ], + "spans": [ + { + "bbox": [ + 298, + 107, + 530, + 127 + ], + "type": "text", + "content": "[24] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit Warping for Animation with Image Sets. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 298, + 127, + 530, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 127, + 530, + 155 + ], + "spans": [ + { + "bbox": [ + 298, + 127, + 530, + 155 + ], + "type": "text", + "content": "[25] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: A large-scale speaker identification dataset. conference of the international speech communication association, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 298, + 156, + 530, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 156, + 530, + 175 + ], + "spans": [ + { + "bbox": [ + 298, + 156, + 530, + 175 + ], + "type": "text", + "content": "[26] Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv: Learning, 2018. 2, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 175, + 530, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 175, + 530, + 203 + ], + "spans": [ + { + "bbox": [ + 299, + 175, + 530, + 203 + ], + "type": "text", + "content": "[27] Hao Ouyang, Bo Zhang, Pan Zhang, Hao Yang, Jiaolong Yang, Dong Chen, Qifeng Chen, and Fang Wen. Real-time neural character rendering with pose-guided multiplane images. ECCV, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 204, + 530, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 204, + 530, + 241 + ], + "spans": [ + { + "bbox": [ + 299, + 204, + 530, + 241 + ], + "type": "text", + "content": "[28] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. arXiv preprint arXiv:2301.06281, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 242, + 530, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 242, + 530, + 270 + ], + "spans": [ + { + "bbox": [ + 299, + 242, + 530, + 270 + ], + "type": "text", + "content": "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 271, + 530, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 271, + 530, + 298 + ], + "spans": [ + { + "bbox": [ + 299, + 271, + 530, + 298 + ], + "type": "text", + "content": "[30] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In AAAI, 2018. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 299, + 530, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 299, + 530, + 328 + ], + "spans": [ + { + "bbox": [ + 299, + 299, + 530, + 328 + ], + "type": "text", + "content": "[31] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In ICCV, 2021. 1, 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 329, + 530, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 329, + 530, + 357 + ], + "spans": [ + { + "bbox": [ + 299, + 329, + 530, + 357 + ], + "type": "text", + "content": "[32] Mincheol Shin, Se Jung Kim, and Frank Biocca. The uncanny valley: No need for any further judgments when an avatar looks eerie. Computers in Human Behavior, 94:100-109, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 357, + 530, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 357, + 530, + 385 + ], + "spans": [ + { + "bbox": [ + 299, + 357, + 530, + 385 + ], + "type": "text", + "content": "[33] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In NeurIPS, 2019. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 386, + 530, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 386, + 530, + 414 + ], + "spans": [ + { + "bbox": [ + 299, + 386, + 530, + 414 + ], + "type": "text", + "content": "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In CVPR, 2019. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 299, + 415, + 530, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 415, + 530, + 443 + ], + "spans": [ + { + "bbox": [ + 299, + 415, + 530, + 443 + ], + "type": "text", + "content": "[35] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 299, + 444, + 530, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 444, + 530, + 462 + ], + "spans": [ + { + "bbox": [ + 299, + 444, + 530, + 462 + ], + "type": "text", + "content": "[36] Soumya Tripathy, Juho Kannala, and Esa Rahtu. Facegan: Facial attribute controllable reenactment gan. In CVPR, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 299, + 463, + 530, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 463, + 530, + 490 + ], + "spans": [ + { + "bbox": [ + 299, + 463, + 530, + 490 + ], + "type": "text", + "content": "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit H. Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial editing of real videos*, 2022. 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 299, + 491, + 530, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 491, + 530, + 520 + ], + "spans": [ + { + "bbox": [ + 299, + 491, + 530, + 520 + ], + "type": "text", + "content": "[38] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaevator: Learning animatable clothed human models from few depth images. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 299, + 521, + 530, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 521, + 530, + 548 + ], + "spans": [ + { + "bbox": [ + 299, + 521, + 530, + 548 + ], + "type": "text", + "content": "[39] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, 2020. 1, 2, 3" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 299, + 548, + 530, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 548, + 530, + 576 + ], + "spans": [ + { + "bbox": [ + 299, + 548, + 530, + 576 + ], + "type": "text", + "content": "[40] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In CVPR, 2021. 2, 5, 8" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 299, + 578, + 530, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 578, + 530, + 605 + ], + "spans": [ + { + "bbox": [ + 299, + 578, + 530, + 605 + ], + "type": "text", + "content": "[41] Olivia Wiles, A Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018. 1, 2, 7" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 299, + 607, + 530, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 607, + 530, + 644 + ], + "spans": [ + { + "bbox": [ + 299, + 607, + 530, + 644 + ], + "type": "text", + "content": "[42] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Matthew Johnson, Virginia Estellers, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 299, + 644, + 530, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 644, + 530, + 690 + ], + "spans": [ + { + "bbox": [ + 299, + 644, + 530, + 690 + ], + "type": "text", + "content": "[43] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Chirag Raman, Jamie Shotton, Toby Sharp, Ivan Stojiljkovic, Tom Cashman, and Julien Valentin. 3d face reconstruction with dense landmarks. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 299, + 692, + 530, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 692, + 530, + 728 + ], + "spans": [ + { + "bbox": [ + 299, + 692, + 530, + 728 + ], + "type": "text", + "content": "[44] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. arXiv preprint arXiv:2301.02379, 2023. 2" + } + ] + } + ], + "index": 42 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "22104" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 107, + 279, + 364 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 107, + 279, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 107, + 279, + 135 + ], + "spans": [ + { + "bbox": [ + 47, + 107, + 279, + 135 + ], + "type": "text", + "content": "[45] Lingbo Yang, Chang Liu, Pan Wang, Shanshe Wang, Peiran Ren, Siwei Ma, and Wen Gao. Hifacegan: Face renovation via collaborative suppression and replenishment. ACM Multimedia, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 136, + 279, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 136, + 279, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 136, + 279, + 164 + ], + "type": "text", + "content": "[46] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 165, + 279, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 165, + 279, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 279, + 202 + ], + "type": "text", + "content": "[47] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. In ECCV, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 203, + 279, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 279, + 230 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 279, + 230 + ], + "type": "text", + "content": "[48] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In ECCV, 2020. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 231, + 279, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 231, + 279, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 279, + 259 + ], + "type": "text", + "content": "[49] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In ICCV, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 260, + 279, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 279, + 287 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 279, + 287 + ], + "type": "text", + "content": "[50] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In CVPR, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 288, + 279, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 279, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 279, + 316 + ], + "type": "text", + "content": "[51] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 317, + 279, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 279, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 279, + 345 + ], + "type": "text", + "content": "[52] Zhimeng Zhang, Lincheng Li, and Yu Ding. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In CVPR, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 345, + 279, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 345, + 279, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 279, + 364 + ], + "type": "text", + "content": "[53] Ruiqi Zhao, Tianyi Wu, and Guodong Guo. Sparse to dense motion transfer for face image animation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "22105" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_content_list.json b/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b47692dc43329e795ff801f25ce642be331c9ddc --- /dev/null +++ b/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_content_list.json @@ -0,0 +1,1853 @@ +[ + { + "type": "text", + "text": "MetaViewer: Towards A Unified Multi-View Representation", + "text_level": 1, + "bbox": [ + 179, + 130, + 790, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ren Wang \nShandong University", + "bbox": [ + 147, + 181, + 320, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xxlifelover@gmail.com", + "bbox": [ + 140, + 219, + 328, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoliang Sun* Shandong University", + "bbox": [ + 375, + 181, + 547, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "haolsun.cn@gmail.com", + "bbox": [ + 372, + 219, + 553, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuling Ma \nShandong Jianzhu University", + "bbox": [ + 594, + 181, + 830, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mayuling20@sdjzu.edu.cn", + "bbox": [ + 609, + 219, + 815, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoming Xi", + "bbox": [ + 315, + 258, + 423, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shandong Jianzhu University", + "bbox": [ + 253, + 277, + 485, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fyzq10@126.com", + "bbox": [ + 305, + 297, + 433, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yilong Yin* Shandong University", + "bbox": [ + 547, + 258, + 717, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ylyin@sdu.edu.cn", + "bbox": [ + 558, + 297, + 702, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 345, + 313, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing multi-view representation learning methods typically follow a specific-to-uniform pipeline, extracting latent features from each view and then fusing or aligning them to obtain the unified object representation. However, the manually pre-specified fusion functions and aligning criteria could potentially degrade the quality of the derived representation. To overcome them, we propose a novel uniform-to-specific multi-view learning framework from a meta-learning perspective, where the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. Specifically, we formulated the extraction and fusion of view-specific latent features as a nested optimization problem and solved it by using a bi-level optimization scheme. In this way, MetaViewer automatically fuses view-specific features into a unified one and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views. Extensive experimental results in downstream classification and clustering tasks demonstrate the efficiency and effectiveness of the proposed method.", + "bbox": [ + 73, + 378, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 712, + 209, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-view representation learning aims to learn a unified representation of the entity from its multiple observable views for the benefit of downstream tasks [35, 43, 58]. Each view acquired by different sensors or sources contains both view-shared consistency information and view-specific information [8]. The view-specific part further consists of complementary and redundant components, where the former can be considered as a supplement to the consistency information, while the latter is view-private and may be", + "bbox": [ + 75, + 738, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3523d0ccd48cabdf9ce9d9e848b184eeb4c48aef543a7882449f9307d0b67dd5.jpg", + "image_caption": [ + "(a) Joint Representation" + ], + "image_footnote": [], + "bbox": [ + 527, + 348, + 689, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5e2d22c9a68dfb1b4b4f9806e0e57c8341dbd12976b32c9cd65f98f17e8a774.jpg", + "image_caption": [ + "(b) Alignment Representation" + ], + "image_footnote": [], + "bbox": [ + 705, + 349, + 859, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/624596e2010e1985f387e1e8121a354d9d4bd77a4e31f525a9c56a9b2a812a6d.jpg", + "image_caption": [ + "(c) S&S Representation" + ], + "image_footnote": [], + "bbox": [ + 535, + 481, + 684, + 588 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e233cccfa9bdf64eb6bf3e9200f085f56138aa8241f42b95706b8700637dd40.jpg", + "image_caption": [ + "(d) Meta Representation", + "Figure 1. (a), (b) and (c) show three multi-view learning frameworks following the specific-to-uniform pipeline, where the unified representation is obtained by fusing or concatenating view-specific features. (d) illustrates our uniform-to-specific manner, where a meta-learner learns to fusion by observing reconstruction from unified representation to specific views." + ], + "image_footnote": [], + "bbox": [ + 700, + 481, + 864, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adverse for the unified representation [16]. Therefore, a high-quality representation is required to retain the consistency and complementary information, as well as filter out the view-private redundant ones [51].", + "bbox": [ + 496, + 727, + 893, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given the data containing two views, $x_{1}$ and $x_{2}$ , prevailing multi-view learning methods typically follow a specific-to-uniform pipeline and can be roughly characterized as:", + "bbox": [ + 496, + 787, + 893, + 834 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\nH := f \\left(x _ {1}; W _ {f}\\right) \\circ g \\left(x _ {2}; W _ {g}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 848, + 892, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $f$ and $g$ are encoding (or embedding [27]) functions that map the original view data into the corresponding latent", + "bbox": [ + 496, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 20 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 31 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 101, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11590", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "features with the trainable parameters $W_{f}$ and $W_{g}$ . These latent features are subsequently aggregated into the unified representation $H$ using the designed aggregation operator o. With different aggregation strategies, existing approaches can be further subdivided into the joint, alignment, and a combined share-specific (S&S) representation [22, 26].", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1 (a) - (c) show the above three branches of the specific-to-uniform framework. Joint representation focuses on the integration of complementary information by directly fusing latent features, where the $\\circ$ is represented as fusion strategies, such as graph-based modules [33], neural networks [10], or other elaborate functions [7, 34]. While alignment representation seeks alignment between view-specific features to retain consistency information and specifies $\\circ$ as the alignment operator, measured by distance [11], correlation [24, 47], or similarity [25, 38, 39]. The aligned features can be concatenated as the unified representation for downstream tasks. As a trade-off strategy, S&S representation explicitly distinguishes latent features into shared and specific representations and only aligns the shared part [20, 30, 53].", + "bbox": [ + 75, + 185, + 473, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite demonstrating promising results, the specific-to-uniform framework inherently suffers from potential risks in the following two aspects: (1) Compared with the data-oriented fusion rules, manually pre-specified rules are designed to compute the unified representation by concatenating or fusing latent features, restricting its extensibility in complex real-world applications. (2) Even if we can find a well-performing fusion scheme, the sequential training manner limits the model to constrain the various component information separately. For example, it is difficult to automatically separate out view-private redundant information from the feature level [22, 37]. Recent researches have attempted to address the second issue by decomposing latent features using matrix factorization [61] or hierarchical feature modeling [51], but still cannot avoid manually pre-specified fusion or even decomposition strategies.", + "bbox": [ + 75, + 398, + 473, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we provide a new meta-learning perspective for multi-view representation learning and propose a novel uniform-to-specific framework to address these potential risks. Fig. 1 (d) shows the overall schematic. In contrast to the specific-to-uniform pipeline, the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. To train the MetaViewer, we first decouple the learning of view-specific latent features and unified meta representation. These two learning processes can then be formulated as a nested optimization problem and eventually solved by a bi-level optimization scheme. In detail, MetaViewer fuses view-specific features into a unified one at the outer level and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views at the inner level. In addition, our uniform-to-specific framework is compatible with most existing objective functions and pre-text", + "bbox": [ + 75, + 643, + 473, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks to cope with complex real-world scenarios. Extensive experiments validate that MetaViewer achieves comparable performance to the state-of-the-art methods in downstream clustering and classification tasks. The core contributions of this work are as follows.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We provide a new meta-learning perspective and develop a novel uniform-to-specific framework to learn a unified multi-view representation. To the best of our knowledge, it could be the first meta-learning-based work in multi-view representation learning community.", + "2. We propose MetaViewer, a meta-learner that formulates the modeling of view-specific features and unified representation as a nested bi-level optimization and ultimately meta-learns a data-driven optimal fusion.", + "3. Extensive experiments on multiple benchmarks validate that our MetaViewer achieves comparable performance to the existing methods in two downstream tasks." + ], + "bbox": [ + 511, + 175, + 893, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 383, + 640, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-view representation learning", + "text_level": 1, + "bbox": [ + 500, + 407, + 805, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Representation learning is not a new topic and plays a critical role in numerous downstream tasks [29, 36, 49]. This work focuses on multi-view representation in unsupervised deep learning scope [56]. Related studies can be roughly summarized into two categories [54]. One is the deep extension of traditional methods, where representative ones include deep canonical correlation analysis (DCCA) and its variants [44, 57]. DCCA [2] intends to discover the nonlinear mapping for two views to a common space in which their correlations are maximally preserved. These methods benefit from a sound theoretical foundation, but also usually have strict restrictions on the number and form of views.", + "bbox": [ + 496, + 431, + 893, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another alternative is deep multi-view learning [5, 32]. Early deep-based approaches attempted to design effective architectures for multiple views, such as CNN-based [12, 42] and GAN-based models [18, 52]. Recent works resort to leveraging mutual or comparative information to derive the constraint for parameters [3, 27, 55]. Most of them follow the pipeline from view-specific features to unified representation. In contrast, our MetaViewer learns by observing the learning from the uniform to the specific. The most related work is $\\mathrm{AE^2}$ -Nets [60], which treats the unified representation as a set of trainable parameters that are optimized during the degradation process of each view. The essential difference is that we learn the data-oriented fusion strategy rather than optimizing the unified representation itself.", + "bbox": [ + 496, + 613, + 895, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Meta-learning", + "text_level": 1, + "bbox": [ + 500, + 832, + 648, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimization-based meta-learning is a classic application of bi-level optimization designed to learn task-level knowledge to quickly handle new tasks [19, 21]. A typical work,", + "bbox": [ + 496, + 854, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9c05b632a0d125a190a8ccae793b904072902733fb371f9901dc903f56baa0a.jpg", + "image_caption": [ + "Figure 2. The overall framework of MetaViewer, contains (a) the meta-split of multi-view data and three modules: (b) embedding module, (c) representation module, and (d) self-supervised module. These modules are trained with a bi-level optimization. Inner-level (dark gray arrows) learns the view-specific reconstruction on the support set, and outer-level (red arrows) updates the entire model to learn the fusion scheme and the unified representation $H$ for downstream tasks by validating on the query set." + ], + "image_footnote": [], + "bbox": [ + 91, + 92, + 883, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAML [13], learns a set of initialization parameters to solve different tasks with a few steps of updates. Similar meta paradigm has been used to learn other manually designed parts, such as the network structure [28], optimizer [62], and even sample weight [40, 41, 48]. Similarly, we try to metalearn the fusion rule of multi-view features for a unified representation. There also exist some works that consider both meta-learning paradigms and multi-view data [14, 31, 45]. However, they are dedicated to exploiting the rich information contained in multiple views to improve the performance of the meta-learner in a few-shot or self-supervised scenario. Instead, we train a meta-learner to derive high-quality shared representations from multi-view data. To the best of our knowledge, this could be the first work to learn multi-view representation with a meta-learning paradigm.", + "bbox": [ + 75, + 436, + 473, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. MetaViewer", + "text_level": 1, + "bbox": [ + 76, + 678, + 205, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of unlabeled multi-view data, $D = \\{x^{i} \\in \\mathbb{R}^{d_{x}}\\}_{i=1}^{N}$ , where $N$ and superscript $i$ are the number and index of entity samples, respectively. Each sample $x^{i} = \\{x_{1}^{i}, x_{2}^{i}, \\ldots, x_{v}^{i}\\}_{v=1}^{V}$ contains $V$ views, where subscript $v$ is the index of views. Our goal is to learn a unified, high-quality representation $H$ for each entity by observing these views and filtering view-private information as much as possible. The overall framework of our MetaViewer is shown in Fig. 2, including three main modules and a bi-level optimization process. The outer-level trains a meta-learner to learn an optimal fusion function and derive the unified representation $H$ , and the inner-level reconstructs original views from $H$ in a few update steps, which explicitly models and separates", + "bbox": [ + 76, + 703, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "view-private information and ensure the quality of the representation. In the following subsections, we first introduce the entire structure of the MetaViewer and then elaborate on the bi-level optimization process.", + "bbox": [ + 496, + 436, + 893, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The entire structure", + "text_level": 1, + "bbox": [ + 500, + 507, + 691, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Embedding module** aims to transform heterogeneous views into the latent feature space, where transformed view embeddings have the same dimension as each other. To this end, we conduct a view-specific embedding function $f_{v}$ for each view, where $v = 1,2,\\dots ,V$ . Given the $v$ -th view data $x_{v}$ of the entity $x$ , the corresponding embedding $z_{v}\\in \\mathbb{R}^{d}$ can be computed by", + "bbox": [ + 496, + 530, + 893, + 636 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {v} = f _ {v} \\left(x _ {v}, \\theta_ {f _ {v}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 647, + 892, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{v}$ is typically instantiated as a multi-layer neural network with learnable parameters $\\theta_{f_v}$ .", + "bbox": [ + 496, + 672, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representation learning module maps the obtained embedding to the view-specific or unified representation by constructing a meta-learner $m$ (i.e., MetaViewer). Note that the two types of representations are both derived from the same set of parameters, but at different learning stages (see 3.2). The view-specific representation is obtained at the inner level, where MetaViewer receives the embedding of the $v$ -th view and outputs the corresponding representation $h_v$ . In contrast, the unified representation $H$ is learned by MetaViewer from all view embeddings at the outer level and is ultimately used for downstream tasks. Therefore, the structure of MetaViewer should be flexible enough to meet the two requirements above simultaneously. Specifically,", + "bbox": [ + 496, + 704, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11592", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MetaViewer is implemented as a channel-oriented 1-d convolutional layer (C-Conv) with a non-linear function (e.g., ReLU [17]), as shown in Fig. 2 (c).", + "bbox": [ + 76, + 90, + 470, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the one hand, at the outer level, we first concatenate all view embeddings at the channel level, i.e., the number of the channel in the concatenated embedding is equal to the number of views, and then train the MetaViewer to learn the fusion of cross-view information $H \\in \\mathbb{R}^{d_h}$ by", + "bbox": [ + 76, + 137, + 468, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH = m \\left(z _ {c a t}, \\omega\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 224, + 468, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{cat} \\in \\mathbb{R}^{d \\times V}$ indicates the concatenated embedding and $\\omega$ is the parameter of MetaViewer. On the other hand, the inner-level initializes and trains MetaViewer to learning the representation $h_v \\in \\mathbb{R}^{d_h}$ of the $v$ -th view via", + "bbox": [ + 76, + 252, + 468, + 313 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {v} = m _ {v} \\left(z _ {v}, \\omega_ {v}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 325, + 468, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m_v$ denotes the learner with the learnable parameter $\\omega_v$ for handling the $v$ -th view in the inner level, also known as the base learner to avoid confusion. It is worth noting that $\\omega_v$ is initialized by a subset of the MetaViewer parameters.", + "bbox": [ + 76, + 354, + 468, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-supervised module provides effectively supervised information for model training by constructing a series of pretext tasks1. These tasks serve as the heads of our framework. Given the input $z$ , the output $s$ can be formalized as", + "bbox": [ + 76, + 416, + 468, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns = t \\left(z, \\theta_ {t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 488, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ and $\\theta_t$ are the network and its parameters involved in pre-text tasks, respectively. Taking the reconstruction head as an example, it achieves the reconstruction object by re-mapping the representation back to the original view space. In this case, the $s := \\hat{x}_v$ indicates the reconstruction result for input $x_v$ , and $t(\\cdot, \\theta_t) := t_v(\\cdot, \\theta_{t_v})$ is a multi-layer neural network with the opposite structure to the embedding module. More similar pre-text tasks have been studied in previous works [55, 57], and this is not the focus of this work.", + "bbox": [ + 76, + 517, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training via bi-level optimization", + "text_level": 1, + "bbox": [ + 76, + 662, + 369, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now have the entire structure, which can be end-to-end trained following the proposed uniform-to-specific pipeline to derive the unified multi-view representation. The training flowchart and step are shown in the top row in Fig. 3, where the learning of view-private features and unified representation are formulated as a nested bi-level optimization problem. Inner-level focus on independent training on a specific view, which receives a given representation and reconstructs it to each view. Outer-level updates the meta-learner to find the optimal fusion rule through observing the training processes across all views. This pipeline decouples the unified entity", + "bbox": [ + 76, + 686, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "representations from view-specific features and enables data-driven fusion. Before the detailed description, we introduce a split strategy for multi-view data in meta-learning style to adapt the bi-level training.", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meta-split of multi-view data. Consider a batch multi-view entity samples $\\{D_v^{batch}\\}_{v=1}^V$ from set $D$ . We randomly and proportionally divide it into two disjoint subsets, marked as support set $S = \\{S_v\\}_{v=1}^V$ and query set $Q = \\{Q_v\\}_{v=1}^V$ , respectively. As shown in 2 (a), support set is used in inner-level for leaning view-specific information, so the sample attributes in it could be ignored. In contrast, query set retains both view and sample attributes for outer-level optimization.", + "bbox": [ + 496, + 152, + 893, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inner-level optimization. Without loss of generality, take the inner-level update after the $o$ -th outer-level optimization as an example. Let $\\omega^o$ be the lasted parameters of the MetaViewer, and $\\theta_v^o = \\{\\theta_{f_v}^o,\\theta_{t_v}^o\\}$ denotes the lasted parameters in embedding and self-supervised modules for brevity. We first initial base learner from meta-learner, i.e., $\\omega_v^0 = \\omega^o$ , and make a copy of $\\theta_v^o$ for $\\tilde{\\theta}_v$ . Note that the copy means gradients with respect to the $\\theta_v^o$ will not be backpropagated to $\\tilde{\\theta}_v$ and vice versa. Thus, $\\omega_v^0$ and $\\tilde{\\theta}_v$ form the initial states of parameters in the inner-level optimization. Suppose $\\mathcal{L}_v^{in}$ is the loss function of the inner-level with respect to the $v$ -th view, the corresponding update goal is", + "bbox": [ + 496, + 272, + 893, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {v} ^ {*} \\left(\\omega^ {o}\\right) = \\arg \\min \\mathcal {L} _ {v} ^ {i n} \\left(\\omega_ {v} \\left(\\omega^ {o}\\right), \\tilde {\\phi} _ {v}; S _ {v}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 463, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider a gradient descent strategy (e.g., SGD [4]), we can further write the update process of $\\omega_{v}$ :", + "bbox": [ + 496, + 493, + 890, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {v} ^ {i} = \\omega_ {v} ^ {i - 1} - \\beta \\frac {\\partial \\mathcal {L} _ {v} ^ {i n}}{\\partial \\omega_ {v} ^ {i - 1}}, \\dots , \\omega_ {v} ^ {0} = \\omega^ {o}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 532, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ and $i$ denote the learning rate and iterative step of inner-level optimization, respectively.", + "bbox": [ + 496, + 574, + 890, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Outer-level optimization. After several inner-level updates, we obtain a set of optimal view-specific parameters for the support set. Outer-level then updates the meta-learner, embedding, and self-supervised modules by training on the query set. With the loss function $\\mathcal{L}^{out}$ , the outer-level optimization goal is", + "bbox": [ + 496, + 604, + 893, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega^ {*}, \\left\\{\\theta_ {v} ^ {*} \\right\\} _ {v = 1} ^ {V} = \\arg \\min \\mathcal {L} ^ {\\text {o u t}} \\left(\\left\\{\\omega_ {v} ^ {*} (\\omega), \\phi_ {v} \\right\\} _ {v = 1} ^ {V}; Q\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 704, + 890, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By alternately optimizing Eq. 6 and Eq. 8, we end up with the optimal meta-parameters $\\omega^{*}$ and a set of view-specific parameters $\\{\\theta_v^*\\}_{v = 1}^V$ . Given a test sample $x_{test}$ , the corresponding unified representation is derived by feeding it sequentially into the embedding module and the meta-learner. The overall framework of MetaViewer is shown in Alg. 1.", + "bbox": [ + 496, + 732, + 893, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Specific-to-uniform versus uniform-to-specific", + "text_level": 1, + "bbox": [ + 500, + 832, + 885, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We discuss the difference between specific-to-uniform and our uniform-to-specific pipeline through three aspects shown in Fig. 3, including the overall flowchart, update steps,", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Constructing pre-text tasks is a classic strategy in self-supervised learning [1], which emphasizes the use of intrinsic properties within or between data as ground-truth to provide supervised information.", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11593", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 The framework of our MetaViewer.", + "bbox": [ + 78, + 90, + 398, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: Training dataset $D$ , meta parameters $\\omega$ , base parameters $\\{\\omega_v\\}_{v=1}^V$ , view-specific parameters $\\{\\theta_v\\}_{v=1}^V$ , the number of views $V$ , the iteration step in inner-level optimization $T$ .", + "bbox": [ + 76, + 108, + 470, + 166 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $\\omega, \\{\\theta_v\\}_{v=1}^V$ ;", + "2: while not done do", + "3: # Outer-level", + "4: Sample and meta-split a batch set from $D$ :", + "5: $\\{D_v^{batch}\\}_{v = 1}^V = \\{\\tilde{S}_v\\}_{v = 1}^V +\\{Q_v\\}_{v = 1}^V.$", + "6: for $t = 1,\\dots ,T$ do", + "7: for $v = 1, \\ldots, V$ do", + "8: #Inner-level", + "9: Initialize $\\omega_{v} = \\omega ,\\tilde{\\theta}_{v} = \\theta_{v};$", + "10: Optimize $\\omega_v(\\omega)$ and $\\theta_v$ via Eq. 6.", + "11: end for", + "12: end for", + "13: Optimize $\\omega$ and $\\{\\phi_v\\}_{v=1}^V$ via Eq. 8.", + "14: end while" + ], + "bbox": [ + 81, + 167, + 401, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and key gradients. For clarity, we here slightly confuse the notation and let $\\theta_v$ and $\\mathcal{L}_v^{in}$ correspond to the view-specific parameters and loss functions in specific-to-uniform pipeline, respectively. Similarly, $\\omega$ and $\\mathcal{L}^{out}$ denote the view-shared parameters and loss functions, respectively. Let us start by reviewing two branches in specific-to-uniform framework.", + "bbox": [ + 75, + 409, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One-level methods is a classical branch in specific-to-uniform framework, where parameters $\\{\\theta_v\\}_{v=1}^V$ and $\\omega$ are simultaneously optimized by combining the functions $\\mathcal{L}_v^{in}$ and $\\mathcal{L}^{out}$ .", + "bbox": [ + 75, + 500, + 470, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-level methods also follow the specific-to-uniform framework, the difference is that they hierarchically learn specific and shared features to avoid interference. That is, parameters $\\theta_v$ are updated at the low level by minimizing the $\\mathcal{L}_v^{in}$ , and parameters $\\omega$ are updated at the high level by minimizing the $\\mathcal{L}^{out}$ , respectively.", + "bbox": [ + 75, + 560, + 468, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bi-level MetaViewer further decouples view-private information, except for view-specific and view-sharing components. Consider the simplest case, where the base learner updates only one step on each view in the inner loop, i.e., $\\omega_{v}^{*} = \\omega -\\beta \\nabla_{\\omega}\\mathcal{L}_{v}^{in}(\\omega)$ . Thus, the final optimization of $\\omega$ in the outer level in Eq. 8 can be rewritten as", + "bbox": [ + 75, + 651, + 470, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\omega} \\sum_ {v} \\mathcal {L} ^ {o u t} (\\omega_ {v} ^ {*}) = \\sum_ {v} \\mathcal {L} ^ {o u t} \\left(\\omega - \\beta \\nabla_ {\\omega} \\mathcal {L} _ {v} ^ {i n} (\\omega)\\right). \\quad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 753, + 470, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the optimal $\\omega$ is expected to be close enough to each view that the view-specific representation can be produced after once updating. In other words, our uniform-to-specific framework requires explicit modeling the gap caused by view-private information when mapping from a unified representation to a particular view, which is achieved by inner-level optimization in MetaViewer.", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/73fa9fb5a546aa25fb4461c13bf5b574097d7ab2e5e3f546531e43917913f6d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
our bi-level MetaViewer uniform-to-specific
Q→θv→ωv(ω)→Lvinωv(ω)→ωv*(ω)∂Lvin/∂ωv(ω)
DInner-Level
S→θv→Σv→ω→Loutθv→θv*, ω→ω*∂Lout/∂θv, ∂Lout/∂2ω
", + "bbox": [ + 504, + 90, + 888, + 166 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b425a2c23c424eb6ef6f011ab586bd3ba2683436f5f35711aaa551fd3b95e5ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
one-level methodsspecific-to-uniform
D→θv→Σv→ωLvnθv→θ*, ω→ω*∂(Lout+Lvn)/∂θv, ∂(Lout+∑v) / ∂ω
", + "bbox": [ + 506, + 167, + 888, + 226 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4fec3400b26f4460afd80131bbad9eee6a1501caf186b4683a2d31a47e1316dd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
two-/multi-level methods
D→θvLinvθv→θ*Low-Level
ω→ω*High-Level
∂Lu/∂θv
", + "bbox": [ + 506, + 227, + 888, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(a) Flowcharts", + "bbox": [ + 549, + 282, + 632, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(b) Steps", + "bbox": [ + 697, + 282, + 750, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(c) Gradients", + "bbox": [ + 795, + 282, + 870, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3. Comparison of two frameworks in terms of flowchart, update steps, and key gradients. The top row is our bi-level MetaViewer with uniform-to-specific framework, where red and black arrows indicate the flow at outer-level and inner-level, respectively. The bottom two rows show two main branches in uniform-to-specific framework, i.e., one-level methods [47] and two- or multi-level methods [51].", + "bbox": [ + 496, + 308, + 893, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. The instances of the objective function", + "text_level": 1, + "bbox": [ + 500, + 431, + 828, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our uniform-to-specific framework emphasizes learning from the training process in the inner loop. In unsupervised multi-view learning, this process typically refers to the reconstruction on views, thus the $\\mathcal{L}_v^{in}$ is specified as the reconstruction loss [27, 60]", + "bbox": [ + 496, + 455, + 893, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v} ^ {i n} = \\mathcal {L} _ {v} ^ {r e c} \\left(S _ {v}, S _ {v} ^ {r e c}\\right) = \\| S _ {v} - S _ {v} ^ {r e c} \\| _ {F} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 542, + 890, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While parameters updated at the outer level can be constrained by richer self-supervised feedbacks, as maintained in Sec. 3.1. Here we provide two instances of the outer-level loss function to demonstrate how MetaViewer can be extended with different learning objectives.", + "bbox": [ + 496, + 571, + 893, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MVer-R adopts the same reconstruction loss as the inner-level, and $\\mathcal{L}^{out} = \\sum_{v}\\mathcal{L}_{v}^{rec}(Q_{v},Q_{v}^{rec})$ , which is the purest implementation of MetaViewer.", + "bbox": [ + 496, + 647, + 893, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MVer-C additionally utilizes a contrastive objective, where the similarities of views belonging to the same entity (i.e., positive pairs) should be maximized and those of different entities (i.e., negative pairs) should be minimized, i.e., $\\mathcal{L}^{out} = \\sum_{v}\\left(\\mathcal{L}_{v}^{rec} + \\sum_{v^{\\prime},v^{\\prime}\\neq v}\\mathcal{L}_{v,v^{\\prime}}^{con}\\right)$ . Following previous work [18, 51, 55], the contrastive loss $\\mathcal{L}_{v,v^{\\prime}}^{con}$ is formed as", + "bbox": [ + 496, + 691, + 893, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v, v ^ {\\prime}} ^ {c o n} = - \\frac {1}{N _ {Q}} \\sum_ {i = 1} ^ {N _ {Q}} \\log \\frac {e ^ {d \\left(q _ {v} ^ {i} , q _ {v ^ {\\prime}} ^ {i}\\right) / \\tau}}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {Q}} e ^ {d \\left(q _ {v} ^ {i} , q _ {v} ^ {j}\\right) / \\tau} + \\sum_ {j = 1} ^ {N _ {Q}} e ^ {d \\left(q _ {v} ^ {i} , q _ {v ^ {\\prime}} ^ {j}\\right) / \\tau}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 801, + 905, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q_v^i$ is the $v$ -th view of the $i$ -th query sample and $d$ is the similarity metric (e.g., cosine similarity [6]). The $N_Q$ and $\\tau$ denote the number of query set samples and the", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11594", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b526b25e2bc8bd1f831a449229665b862a84f816421eb92fd916b751fb875700.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Datasets#Views#Classes#Samples (train, val, test)View Dimensions
Handwirtten2102000 (1200, 400, 400)240; 216
RGB-D250500 (300, 100, 100)12288; 4096
Animal25010158 (6074, 2011, 2073)4096; 4096
Fashion-MV31010000 (6000, 2000, 2000)784; 784; 784
Caltech101-206202386 (1425, 469, 492)48; 40; 254; 1984; 512; 928
", + "bbox": [ + 158, + 88, + 812, + 196 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0348a81c98b12687185c2b641c2fc40a771ea983139cf1267f716ffa88d7b6bf.jpg", + "table_caption": [ + "Table 1. The attributes for all datasets used in our experiments." + ], + "table_footnote": [], + "table_body": "
MethodsHandwrittenRGB-DAnimalFashion-MVCaltech101-20
ACCNMIARIACCNMIARIACCNMIARIACCNMIARIACCNMIARI
Baseline43.5552.1925.6848.3080.9704.9545.7966.2233.8750.9145.9832.0936.6348.9623.54
DCCA [2] (2013)57.2569.8052.1551.0082.9952.0260.3468.5146.1070.7080.4261.8038.6250.8822.73
DCCAE [47] (2015)63.0075.0459.2948.0081.5848.3465.1766.8248.8371.0581.1262.3436.5952.2425.25
MIB [8] (2020)63.2567.5852.1650.0081.1351.2760.8961.9852.4957.2073.8347.6235.9847.0022.18
WTNNM [15] (2020)75.6977.6261.5152.5481.9552.4671.9271.4361.0683.3386.3077.9939.7455.5826.92
TLRR [23] (2021)78.6077.9966.4052.8782.3853.3769.7973.0161.2379.5484.0178.5440.1852.3827.60
MFLVC [51] (2022)64.0064.5348.8553.0083.3154.0774.1076.0864.2883.2088.7578.9336.5958.3626.87
DCP [27] (2022)66.2570.5656.1052.0082.0452.6472.7774.0265.4862.6068.3854.3036.7944.3723.50
MVer-R (ours)75.0078.5367.2153.0082.4153.0476.4978.2565.2880.8088.1375.0541.8758.5229.19
MVer-C (ours)86.2578.9672.2557.0084.9757.0775.9278.0166.0785.4088.7680.0745.1260.8635.00
", + "bbox": [ + 81, + 234, + 888, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Clustering results of all methods on six datasets. Bold and underline denote the best and second-best results, respectively.", + "bbox": [ + 98, + 416, + 867, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "temperature parameter, respectively. Note that, the derived meta representation can also be used in contrastive learning as an additional novel view.", + "bbox": [ + 75, + 457, + 470, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 520, + 209, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present extensive experimental results to validate the quality of the unified representation derived from our MetaViewer. The remainder of the experiments are organized as follows: Subsection 4.1 lists datasets, compared methods, and implementation details. Subsection 4.2 compares the performance of our method with classical and state-of-the-art methods on two common downstream scenarios, clustering and classification tasks. Comparisons with manually designed fusion and ablation studies are shown in Subsections 4.3 and 4.4, respectively.", + "bbox": [ + 75, + 546, + 470, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 710, + 266, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To comprehensively evaluate the effectiveness of our MetaViewer, we conduct five multi-view benchmarks in experiments. All datasets are scaled to $[0,1]$ and split into training, validation, and test sets in the ratio of $6:2:2$ , as shown in Tab. 1. Handwritten contains 2,000 handwritten digital images from 0 to 9, where two types of descriptors, i.e., 240-D pixel average in $2\\times 3$ windows and 216-D profile correlations, are selected as two views [60]. RGB-D dataset contains visual and depth images of 300 distinct objects across 50 categories [46, 63]. Two views are obtained by flattening the $64\\times 64\\times 3$ color images and $64\\times 64$ depth", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images. Animal consists of 10158 images from 50 classes with two views. Unlike other datasets, these two views are features extracted by deep networks [59]. Fashion-MV is an image dataset that contains 10 categories with a total of 30,000 fashion products. It has three views and each of which consists of 10,000 gray images sampled from the same category [50]. Caltech101-20 is a subset of the Caltech101 image set [9], which consists of 2,386 images of 20 subjects. Six features are used, including Gabor, Wavelet Moments, CENTRIST, HOG, GIST, and LBP.", + "bbox": [ + 496, + 457, + 893, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare the performance of MetaViewer with seven representative multi-view learning methods, including two classical methods (DCCA [2] and DCCAE [47]) and five state-of-the-art methods (WTNNM [15], TLRR [23], MIB [8], MFLVC [51] and DCP [27]). Among them, DCCA and DCCAE are the deep extensions of traditional correlation strategies. WTNNM and TLRR are representative tensor-based methods. MIB is a typical generative method with mutual information constraints. DCP learns unified representation in both unsupervised and supervised scenarios, and we report the unsupervised version for a fair comparison. In particular, MFLVC also notices the view-private redundant information and designs a multilevel feature network for clustering tasks. In addition, we also compare the model without MetaViewer as a baseline, where the unified representation is obtained by concatenating view-specific features output by the embedding module.", + "bbox": [ + 496, + 609, + 895, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. For a fair comparison, all methods are trained from scratch, and all deep-network-based", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11595", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/128e9fe9b96af3afffe7992a100375c652893922ba197b92202bae4fa59fa96c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsHandwrittenRGB-DAnimalFashion-MVCaltech101-20
ACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-score
Baseline87.0087.3287.0414.0006.4507.7071.7865.0263.1087.5087.6387.5475.4151.9546.21
DCCA [2] (2013)88.2589.2088.0530.0021.1022.0462.8461.6160.6584.9085.2283.5471.5445.2739.81
DCCAE [47] (2015)90.0090.4889.9224.0016.0016.9166.8262.5362.8385.3585.9783.8471.5460.5743.25
MIB [8] (2020)79.0083.9078.5233.0028.5027.3759.3263.7862.1386.8086.8086.5572.7261.6452.47
WTNNM [15] (2020)96.7796.2096.3645.0047.1843.9069.9067.4065.4894.5094.5094.5883.2780.4974.78
TLRR [23] (2021)97.0097.0597.1749.0053.3247.6671.9069.5268.3596.3596.2896.4085.5577.3075.24
MFLVC [51] (2022)94.0094.2094.0144.0046.0941.8175.6272.1770.8196.5096.5296.4985.3771.8369.07
DCP [27] (2022)97.2597.3097.2437.0028.8730.7877.9573.4370.0189.2582.0682.9092.4889.4184.58
MVer-R (ours)97.0097.0897.0051.0053.6548.7377.6973.9171.2296.8596.3796.4892.2889.4684.21
MVer-C (ours)97.7597.9097.7556.0055.2052.7878.0374.5671.5597.7096.7897.0792.1690.6885.72
", + "bbox": [ + 81, + 88, + 890, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods share the same embedding backbone, following the previous work [51]. For two-view-based methods, such as DCCA, DCCAE, and MIB, we report the results of the two best-performing views. The final unified representation $H$ in alignment-based methods is a concatenation of the processed features, and the representation dimension is specified as $d_h = 256$ . The results on clustering and classification tasks are obtained by subsequent K-means and SVM classifier, respectively. For MetaViewer, we train 2,000 epochs for all benchmarks, and set the batch size is 32 for RGBD and 128 for others. The learning rates in outer- and inner-level are set to $10^{-4}$ and $10^{-3}$ , respectively. We use a single convolutional layer with 32 kernels of size 5 as the meta-learner, and we set the proportion of support set and the update step in the inner level to 0.2 and 2, respectively. All experiments have been verified using the PyTorch library on a single RTX3090. Code is available at https://github.com/xxLifeLover/MetaViewer.", + "bbox": [ + 76, + 300, + 473, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Performance on downstream tasks", + "text_level": 1, + "bbox": [ + 76, + 587, + 377, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Clustering results. Tab. 2 lists the results of the clustering task, where the performance is measured by three standard evaluation metrics, i.e., Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI). A higher value of these metrics indicates better clustering performance. It can be observed that (1) our MVer-C variant significantly outperforms other compared methods on all benchmarks; (2) the second-best results appear between MVer-R and MFLVC, both of which explicitly separate the view-private information; (3) a larger number of categories and views is the main reason for the degradation of clustering performance, and our Metaviewer improves most significantly in such a scenario (e.g., Fashion-MV and Caltech101-20).", + "bbox": [ + 75, + 612, + 470, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classification results. Tab. 3 lists the results of the classification task, where three common metrics are used, including Accuracy, Precision, and F-score. A higher value indicates better classification performance. Similar to the clustering results, two variants of MetaViewer significantly", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a060463772f5583fc92a76d3b84f49fa29ac6ad78702191595d161c3e3e95c03.jpg", + "table_caption": [ + "Table 3. Classification results of all methods on six datasets. Bold and underline denote the best and second-best results, respectively." + ], + "table_footnote": [], + "table_body": "
StrategiesRulesACC↑NMI↑ARI↑MSE↓
Sumzx + zy69.2571.8959.021.84
Maxmax(zx, zy)80.7573.9363.75-
Concat.cat[zx, zy]78.7572.0261.521.77
Linearl(zx, zy, θl)85.0077.4069.714.74
C-Convm(zx, zy, ω)69.7565.2151.332.37
MetaViewermeta-learning86.2578.9672.252.45
", + "bbox": [ + 504, + 297, + 888, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Clustering resulting on the Handwritten dataset.", + "bbox": [ + 526, + 412, + 864, + 426 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6baaf3e6a29beef395c5ab07655a3d9c308950faeffa017db965f0ecdf1e1b55.jpg", + "image_caption": [ + "Figure 4. Class activate maps on RGB-D samples for different fusion strategies, including single views, Sum, and our MetaViewer. View 1 and view 2 are the RGB and depth view, respectively." + ], + "image_footnote": [], + "bbox": [ + 522, + 444, + 869, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "outperform the comparison methods. It is worth noting that (1) DCP learns unified representation and therefore achieves the second-best result instead of MFLVC. (2) The number of categories is the main factor affecting the classification performance, and our method obtains the most significant improvement in the RGB-D dataset with up to 50 classes.", + "bbox": [ + 496, + 669, + 893, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with manually designed fusion.", + "text_level": 1, + "bbox": [ + 498, + 770, + 872, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in 3.3, MetaViewer essentially learns an optimal fusion rule that preserves as much consistency and complementary information as possible while filtering out view-private information. To verify this, we compare it with commonly used fusion strategies [26], including sum, maxima, concatenation, linear layer and $C$ -Conv. The former three are the specified fusion rules without trainable param-", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11596", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f48bdb6a99fc4b704fcbbf5744ca4dddb96009c29daa933fb69f0bc450bae03.jpg", + "image_caption": [ + "(a) Effect of network width" + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 272, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e8e719885cd11547462682ff6d9cbac33b9b3b85e7688e22ad2c2f240f29672d.jpg", + "image_caption": [ + "(b) Effect of kernel size" + ], + "image_footnote": [], + "bbox": [ + 272, + 89, + 464, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a6e0a0af9705b163b1b6c0f5014a349a1123fbfe531cbf97ecfb7ee29a7720a.jpg", + "image_caption": [ + "Figure 5. Effect of meta-learner architectures with different depth, width, and kernel size on classification accuracy.", + "(a) Proportion of support set", + "Figure 6. Effect of (a) different meta-division ratios and (b) the number of inner-loop iterations on classification accuracy." + ], + "image_footnote": [], + "bbox": [ + 86, + 273, + 274, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c60e460c8994000eddf195e84b26d7a58a01f158a284f0a03f1a4466edc0fc4.jpg", + "image_caption": [ + "(b) Steps of inner-level update" + ], + "image_footnote": [], + "bbox": [ + 277, + 273, + 465, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eters, and the remaining two are the trainable fusion layer trained via the specific-to-uniform manner. Tab. 4 lists the clustering results and an additional MSE score on the Hand-written dataset with the same embedding and reconstruction network. We can observe that (1) trainable fusion layers outperform the hand-designed rules, and our MetaViewer yields the best performance; (2) the MSE scores listed in the last column indicate that the quality of the unified representation cannot be measured and guaranteed only with the reconstruction constraint, due to the view-private redundant information mixed in view-specific latent features.", + "bbox": [ + 75, + 468, + 468, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further explore the fusion preferences learned by MetaViewer, we visualize the class activation map for the RGB-D samples in Fig. 4. MetaViewer balances information from all views instead of just the salient one (view 1), ensuring a high-quality unified multi-view representation.", + "bbox": [ + 75, + 636, + 470, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 724, + 241, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Meta-learner structures. We implement the meta-learner as a channel-level convolution structure in this work. Albeit simple, this layer can be considered as a universal approximator for almost any continuous function [40], and thus can fit a wide range of conventional fusion functions. To investigate the effect of network depth, width, and convolution kernel size on the performance of the representation, we alternately fix the 32 kernels and the $1 \\times 3$ kernel size and show the classification results on Handwritten data in Fig. 5. It is clear that (1) the meta-learner works well with just a", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shallow structure, as shown in Fig. 5 (a), instead of gradually overfitting to the training data as the network deepens or widens, and (2) our MetaViewer is stable and insensitive to the hyper-parameters within reasonable ranges.", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Meta-split ratios. Fig. 6 (a) shows the impact of the meta-split mentioned in Sec. 3.2 on the classification performance, where the proportion of support set is set from 0.1 to 0.9 in steps of 0.1, and the rest is query set. In addition to the single view, we also compare the sum and concat. fusion as baselines. MetaViewer consistently surpasses all baselines over the experimental proportion. In addition, fusion baselines are more dependent on the better-performing view at lower proportions, instead becoming unstable as the available query sample decreases.", + "bbox": [ + 496, + 154, + 893, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inner-level update steps. Another hyper-parameter is the number of iteration steps in inner-level optimization. More iterations mean a larger gap from the learned meta representation to the specific view space, i.e., coarser modeling of view-private information. Fig. 6 (b) shows the classification results with various steps, where $n$ steps mean that the inner-level optimization is updated $n$ times throughout the training. MetaViewer achieves the best results when using 1 steps, and remains stable within 15 steps.", + "bbox": [ + 496, + 310, + 893, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 470, + 617, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduced a novel meta-learning perspective for multi-view learning, and proposed a meta-learner, namely MetaViewer, to derive a high-quality unified representation for downstream tasks. In contrast to the prevailing specific-to-uniform pipeline, MetaViewer observes the reconstruction process from unified representation to specific views and essentially learns an optimal data-driven fusion that separates and filters out meaningless view-private information. The proposed framework is compatible with most excellent objective functions and pre-text tasks to cope with complex real-world scenarios. Extensive experimental results on clustering and classification tasks demonstrate the performance of meta-learned unified representation. In addition, we believe that our bi-level training pipeline is also promising for handling incomplete views, and leave it to future work.", + "bbox": [ + 496, + 500, + 893, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 751, + 666, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported in part by the Natural Science Foundation of China (No. 62106129, 62176139, and 62177031), the Natural Science Foundation of Shandong Province (No. ZR2021QF053, ZR2021ZD15), and the China Postdoctoral Science Foundation (No. 2021TQ0195, 2021M701984). The author acknowledges Prof. Xiushan Nie of Shandong Jianzhu University (China) for the financial support and valuable discussions.", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11597", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sara Atito Ali Ahmed, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. CoRR, abs/2104.03602, 2021. 4", + "[2] Galen Andrew, Raman Arora, Jeff A. Bilmes, and Karen Livescu. Deep canonical correlation analysis. In ICML, volume 28, pages 1247-1255, 2013. 2, 6, 7", + "[3] Philip Bachman, R. Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In NeurIPS, pages 15509-15519, 2019. 2", + "[4] Léon Bottou. Large-scale machine learning with stochastic gradient descent. In COMPSTAT, pages 177-186, 2010. 4", + "[5] Abhra Chaudhuri, Massimiliano Mancini, Yanbei Chen, Zeynep Akata, and Anjan Dutta. Cross-modal fusion distillation for fine-grained sketch-based image retrieval. In BMVC, page 499, 2022. 2", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119, pages 1597-1607, 2020. 5", + "[7] Yanhua Cheng, Xin Zhao, Rui Cai, Zhiwei Li, Kaiqi Huang, and Yong Rui. Semi-supervised multimodal deep learning for RGB-D object recognition. In IJCAI, pages 3345-3351, 2016. 2", + "[8] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushner, and Zeynep Akata. Learning robust representations via multiview information bottleneck. In ICLR, 2020. 1, 6, 7", + "[9] Li Fei-Fei, Robert Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Comput. Vis. Image Underst., 106(1):59-70, 2007. 6", + "[10] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, pages 1933-1941, 2016. 2", + "[11] Fangxiang Feng, Xiaojie Wang, and Ruifan Li. Cross-modal retrieval with correspondence autoencoder. In ACM MM, pages 7-16, 2014. 2", + "[12] Yifan Feng, Zizhao Zhang, Xibin Zhao, Rongrong Ji, and Yue Gao. GVCNN: group-view convolutional neural networks for 3d shape recognition. In CVPR, pages 264-272, 2018. 2", + "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, volume 70, pages 1126–1135, 2017. 3", + "[14] Kuiliang Gao, Bing Liu, Xuchu Yu, and Anzhu Yu. Unsupervised meta learning with multiview constraints for hyperspectral image small sample set classification. IEEE TIP, 31:3449-3462, 2022. 3", + "[15] Quanxue Gao, Wei Xia, Zhizhen Wan, De-Yan Xie, and Pu Zhang. Tensor-svd based graph learning for multi-view subspace clustering. In AAAI, pages 3930-3937, 2020. 6, 7", + "[16] Yu Geng, Zongbo Han, Changqing Zhang, and Qinghua Hu. Uncertainty-aware multi-view representation learning. In AAAI, pages 7545-7553, 2021. 1", + "[17] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In AISTATS, volume 15, pages 315-323, 2011. 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Kaveh Hassani and Amir Hosein Khas Ahmadi. Contrastive multi-view representation learning on graphs. In ICML, volume 119, pages 4116-4126, 2020. 2, 5", + "[19] Timothy M. Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE TPAMI, 44(9):5149-5169, 2022. 2", + "[20] Junlin Hu, Jiwen Lu, and Yap-Peng Tan. Sharable and individual multi-view metric learning. IEEE TPAMI, 40(9):2281-2288, 2018. 2", + "[21] Mike Huisman, Jan N. van Rijn, and Aske Plaat. A survey of deep meta-learning. Artif. Intell. Rev., 54(6):4483-4541, 2021. 2", + "[22] Xiaodong Jia, Xiao-Yuan Jing, Xiaoke Zhu, Songcan Chen, Bo Du, Ziyun Cai, Zhenyu He, and Dong Yue. Semi-supervised multi-view deep discriminant representation learning. IEEE TPAMI, 43(7):2496-2509, 2021. 2", + "[23] Yuheng Jia, Hui Liu, Junhui Hou, Sam Kwong, and Qingfu Zhang. Multi-view spectral clustering tailored tensor low-rank representation. IEEE TCSVT, 31(12):4784-4797, 2021. 6, 7", + "[24] Xiao-Yuan Jing, Fei Wu, Xiwei Dong, Shiguang Shan, and Songcan Chen. Semi-supervised multi-view correlation feature learning with application to webpage classification. In AAAI, pages 1374-1381, 2017. 2", + "[25] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. IEEE TPAMI, 39(4):664-676, 2017. 2", + "[26] Yingming Li, Ming Yang, and Zhongfei Zhang. A survey of multi-view representation learning. IEEE TKDE, 31(10):1863-1883, 2019. 2, 7", + "[27] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE TPAMI, 2022. 1, 2, 5, 6, 7", + "[28] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In ICCV, pages 3295-3304, 2019. 3", + "[29] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In CVPR, pages 3623–3632, 2019. 2", + "[30] Shirui Luo, Changqing Zhang, Wei Zhang, and Xiaochun Cao. Consistent and specific multi-view subspace clustering. In AAAI, pages 3730-3737, 2018. 2", + "[31] Yao Ma, Shilin Zhao, Weixiao Wang, Yaoman Li, and Irwin King. Multimodality in meta-learning: A comprehensive survey. KBS, 250:108976, 2022. 3", + "[32] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In CVPR, pages 9826–9836, 2021. 2", + "[33] Feiping Nie, Guohao Cai, Jing Li, and Xuelong Li. Autoweighted multi-view learning for image clustering and semi-supervised classification. IEEE TIP, 27(3):1501-1511, 2018. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11598", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Xiushan Nie, Weizhen Jing, Chaoran Cui, Chen Jason Zhang, Lei Zhu, and Yilong Yin. Joint multi-view hashing for large-scale near-duplicate video retrieval. IEEE TKDE, 32(10):1951-1965, 2020. 2", + "[35] Xiushan Nie, Xingbo Liu, Jie Guo, Letian Wang, and Yilong Yin. Supervised discrete multiple-length hashing for image retrieval. IEEE TBD, 9(1):312-327, 2023. 1", + "[36] Zheyun Qin, Xiankai Lu, Xiushan Nie, Dongfang Liu, Yi-long Yin, and Wenguan Wang. Coarse-to-fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica, 10(5):1-17, 2023. 2", + "[37] Mathieu Salzmann, Carl Henrik Ek, Raquel Urtasun, and Trevor Darrell. Factorized orthogonal latent spaces. In AISTATS, volume 9, pages 701-708, 2010. 2", + "[38] Yang Shi, Xiushan Nie, Xingbo Liu, Lu Yang, and Yilong Yin. Zero-shot hashing via asymmetric ratio similarity matrix. IEEE TKDE, 2022. 2", + "[39] Yang Shi, Xiushan Nie, Xingbo Liu, Li Zou, and Yilong Yin. Supervised adaptive similarity matrix hashing. IEEE TIP, 31:2755-2766, 2022. 2", + "[40] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pages 1917-1928, 2019. 3, 8", + "[41] Haoliang Sun, Chenhui Guo, Qi Wei, Zhongyi Han, and Yilong Yin. Learning to rectify for robust learning with noisy labels. PR, 124:108467, 2022. 3", + "[42] Kai Sun, Jiangshe Zhang, Junmin Liu, Ruixuan Yu, and Zengjie Song. DRCNN: dynamic routing convolutional neural network for multi-view 3d object recognition. IEEE TIP, 30:868-877, 2021. 2", + "[43] Shiliang Sun, Wenbo Dong, and Qiuyang Liu. Multi-view representation learning with deep gaussian processes. IEEE TPAMI, 43(12):4453-4468, 2021. 1", + "[44] Zhongkai Sun, Prathusha Kameswara Sarma, William A. Sethares, and Yingyu Liang. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In AAAI, pages 8992-8999, 2020. 2", + "[45] Risto Vuorio, Shao-Hua Sun, Hexiang Hu, and Joseph J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. In NeurIPS, pages 1–12, 2019. 3", + "[46] Shiye Wang, Changsheng Li, Yanming Li, Ye Yuan, and Guoren Wang. Self-supervised information bottleneck for deep multi-view subspace clustering. CoRR, abs/2204.12496, 2022. 6", + "[47] Weiran Wang, Raman Arora, Karen Livescu, and Jeff A. Bilmes. On deep multi-view representation learning. In ICML, volume 37, pages 1083-1092, 2015. 2, 5, 6, 7", + "[48] Qi Wei, Lei Feng, Haoliang Sun, Ren Wang, Chenhui Guo, and Yilong Yin. Fine-grained classification with noisy labels. arXiv preprint arXiv:2303.02404, 2023. 3", + "[49] Qi Wei, Haoliang Sun, Xiankai Lu, and Yilong Yin. Self-filtering: A noise-aware sample selection for label noise with confidence penalization. In ECCV, pages 516-532, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Jie Xu, Yazhou Ren, Huayi Tang, Zhimeng Yang, Lili Pan, Yang Yang, and Xiaorong Pu. Self-supervised discriminative feature learning for multi-view clustering. CoRR, abs/2103.15069, 2021. 6", + "[51] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In CVPR, pages 16030-16039, 2022. 1, 2, 5, 6, 7", + "[52] Fei Xue, Xin Wu, Shaojun Cai, and Junqiu Wang. Learning multi-view camera relocalization with graph neural networks. In CVPR, pages 11372-11381, 2020. 2", + "[53] Xiaowei Xue, Feiping Nie, Sen Wang, Xiaojun Chang, Bela Stantic, and Min Yao. Multi-view correlated feature learning by uncovering shared component. In AAAI, pages 2810-2816, 2017. 2", + "[54] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 2", + "[55] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, pages 8824-8833, 2022. 2, 4, 5", + "[56] Xiao Yu, Hui Liu, Yuxiu Lin, Yan Wu, and Caiming Zhang. Auto-weighted sample-level fusion with anchors for incomplete multi-view clustering. PR, 130:108772, 2022. 2", + "[57] Yun-Hao Yuan, Jin Li, Yun Li, Jipeng Qiang, Yi Zhu, Xiaobo Shen, and Jianping Gou. Learning canonical f-correlation projection for compact multiview representation. In CVPR, pages 19238-19247, 2022. 2, 4", + "[58] Changqing Zhang, Huazhu Fu, Qinghua Hu, Xiaochun Cao, Yuan Xie, Dacheng Tao, and Dong Xu. Generalized latent multi-view subspace clustering. IEEE TPAMI, 42(1):86-99, 2020. 1", + "[59] Changqing Zhang, Zongbo Han, Yajie Cui, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Cpm-nets: Cross partial multi-view networks. In NeurIPS, pages 557-567, 2019. 6", + "[60] Changqing Zhang, Yeqing Liu, and Huazhu Fu. Ae2-nets: Autoencoder in autoencoder networks. In CVPR, pages 2577-2585, 2019. 2, 5, 6", + "[61] Liang Zhao, Tao Yang, Jie Zhang, Zhikui Chen, Yi Yang, and Z. Jane Wang. Co-learning non-negative correlated and uncorrelated features for multi-view data. IEEE TNNLS, 32(4):1486-1496, 2021. 2", + "[62] Wenqing Zheng, Tianlong Chen, Ting-Kuei Hu, and Zhangyang Wang. Symbolic learning to optimize: Towards interpretability and scalability. In ICLR, 2022. 3", + "[63] Pengfei Zhu, Binyuan Hui, Changqing Zhang, Dawei Du, Longyin Wen, and Qinghua Hu. Multi-view deep subspace clustering networks. CoRR, abs/1908.01978, 2019. 6" + ], + "bbox": [ + 501, + 92, + 893, + 786 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11599", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_model.json b/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8ccd442bc7bd7e152a39c306cd272ed8e58afdc2 --- /dev/null +++ b/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_model.json @@ -0,0 +1,2728 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.021 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.032 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.791, + 0.154 + ], + "angle": 0, + "content": "MetaViewer: Towards A Unified Multi-View Representation" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.182, + 0.321, + 0.217 + ], + "angle": 0, + "content": "Ren Wang \nShandong University" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.22, + 0.33, + 0.234 + ], + "angle": 0, + "content": "xxlifelover@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.182, + 0.548, + 0.217 + ], + "angle": 0, + "content": "Haoliang Sun* Shandong University" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.22, + 0.554, + 0.234 + ], + "angle": 0, + "content": "haolsun.cn@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.596, + 0.182, + 0.831, + 0.218 + ], + "angle": 0, + "content": "Yuling Ma \nShandong Jianzhu University" + }, + { + "type": "text", + "bbox": [ + 0.61, + 0.22, + 0.816, + 0.234 + ], + "angle": 0, + "content": "mayuling20@sdjzu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.26, + 0.424, + 0.277 + ], + "angle": 0, + "content": "Xiaoming Xi" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.278, + 0.486, + 0.295 + ], + "angle": 0, + "content": "Shandong Jianzhu University" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.299, + 0.434, + 0.312 + ], + "angle": 0, + "content": "fyzq10@126.com" + }, + { + "type": "text", + "bbox": [ + 0.549, + 0.26, + 0.718, + 0.296 + ], + "angle": 0, + "content": "Yilong Yin* Shandong University" + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.299, + 0.704, + 0.312 + ], + "angle": 0, + "content": "ylyin@sdu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.347, + 0.314, + 0.363 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.38, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Existing multi-view representation learning methods typically follow a specific-to-uniform pipeline, extracting latent features from each view and then fusing or aligning them to obtain the unified object representation. However, the manually pre-specified fusion functions and aligning criteria could potentially degrade the quality of the derived representation. To overcome them, we propose a novel uniform-to-specific multi-view learning framework from a meta-learning perspective, where the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. Specifically, we formulated the extraction and fusion of view-specific latent features as a nested optimization problem and solved it by using a bi-level optimization scheme. In this way, MetaViewer automatically fuses view-specific features into a unified one and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views. Extensive experimental results in downstream classification and clustering tasks demonstrate the efficiency and effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.713, + 0.21, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.472, + 0.876 + ], + "angle": 0, + "content": "Multi-view representation learning aims to learn a unified representation of the entity from its multiple observable views for the benefit of downstream tasks [35, 43, 58]. Each view acquired by different sensors or sources contains both view-shared consistency information and view-specific information [8]. The view-specific part further consists of complementary and redundant components, where the former can be considered as a supplement to the consistency information, while the latter is view-private and may be" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.349, + 0.691, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.459, + 0.675, + 0.473 + ], + "angle": 0, + "content": "(a) Joint Representation" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.35, + 0.861, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.459, + 0.863, + 0.473 + ], + "angle": 0, + "content": "(b) Alignment Representation" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.482, + 0.685, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.593, + 0.678, + 0.605 + ], + "angle": 0, + "content": "(c) S&S Representation" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.482, + 0.865, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.593, + 0.848, + 0.605 + ], + "angle": 0, + "content": "(d) Meta Representation" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.619, + 0.896, + 0.704 + ], + "angle": 0, + "content": "Figure 1. (a), (b) and (c) show three multi-view learning frameworks following the specific-to-uniform pipeline, where the unified representation is obtained by fusing or concatenating view-specific features. (d) illustrates our uniform-to-specific manner, where a meta-learner learns to fusion by observing reconstruction from unified representation to specific views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.895, + 0.789 + ], + "angle": 0, + "content": "adverse for the unified representation [16]. Therefore, a high-quality representation is required to retain the consistency and complementary information, as well as filter out the view-private redundant ones [51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.895, + 0.835 + ], + "angle": 0, + "content": "Given the data containing two views, \\( x_{1} \\) and \\( x_{2} \\), prevailing multi-view learning methods typically follow a specific-to-uniform pipeline and can be roughly characterized as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.849, + 0.893, + 0.866 + ], + "angle": 0, + "content": "\\[\nH := f \\left(x _ {1}; W _ {f}\\right) \\circ g \\left(x _ {2}; W _ {g}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "where \\( f \\) and \\( g \\) are encoding (or embedding [27]) functions that map the original view data into the corresponding latent" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.234, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11590" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.183 + ], + "angle": 0, + "content": "features with the trainable parameters \\( W_{f} \\) and \\( W_{g} \\). These latent features are subsequently aggregated into the unified representation \\( H \\) using the designed aggregation operator o. With different aggregation strategies, existing approaches can be further subdivided into the joint, alignment, and a combined share-specific (S&S) representation [22, 26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.474, + 0.398 + ], + "angle": 0, + "content": "Fig. 1 (a) - (c) show the above three branches of the specific-to-uniform framework. Joint representation focuses on the integration of complementary information by directly fusing latent features, where the \\(\\circ\\) is represented as fusion strategies, such as graph-based modules [33], neural networks [10], or other elaborate functions [7, 34]. While alignment representation seeks alignment between view-specific features to retain consistency information and specifies \\(\\circ\\) as the alignment operator, measured by distance [11], correlation [24, 47], or similarity [25, 38, 39]. The aligned features can be concatenated as the unified representation for downstream tasks. As a trade-off strategy, S&S representation explicitly distinguishes latent features into shared and specific representations and only aligns the shared part [20, 30, 53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.474, + 0.642 + ], + "angle": 0, + "content": "Despite demonstrating promising results, the specific-to-uniform framework inherently suffers from potential risks in the following two aspects: (1) Compared with the data-oriented fusion rules, manually pre-specified rules are designed to compute the unified representation by concatenating or fusing latent features, restricting its extensibility in complex real-world applications. (2) Even if we can find a well-performing fusion scheme, the sequential training manner limits the model to constrain the various component information separately. For example, it is difficult to automatically separate out view-private redundant information from the feature level [22, 37]. Recent researches have attempted to address the second issue by decomposing latent features using matrix factorization [61] or hierarchical feature modeling [51], but still cannot avoid manually pre-specified fusion or even decomposition strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.474, + 0.903 + ], + "angle": 0, + "content": "In this work, we provide a new meta-learning perspective for multi-view representation learning and propose a novel uniform-to-specific framework to address these potential risks. Fig. 1 (d) shows the overall schematic. In contrast to the specific-to-uniform pipeline, the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. To train the MetaViewer, we first decouple the learning of view-specific latent features and unified meta representation. These two learning processes can then be formulated as a nested optimization problem and eventually solved by a bi-level optimization scheme. In detail, MetaViewer fuses view-specific features into a unified one at the outer level and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views at the inner level. In addition, our uniform-to-specific framework is compatible with most existing objective functions and pre-text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "tasks to cope with complex real-world scenarios. Extensive experiments validate that MetaViewer achieves comparable performance to the state-of-the-art methods in downstream clustering and classification tasks. The core contributions of this work are as follows." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.176, + 0.895, + 0.253 + ], + "angle": 0, + "content": "1. We provide a new meta-learning perspective and develop a novel uniform-to-specific framework to learn a unified multi-view representation. To the best of our knowledge, it could be the first meta-learning-based work in multi-view representation learning community." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.26, + 0.895, + 0.322 + ], + "angle": 0, + "content": "2. We propose MetaViewer, a meta-learner that formulates the modeling of view-specific features and unified representation as a nested bi-level optimization and ultimately meta-learns a data-driven optimal fusion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.329, + 0.895, + 0.375 + ], + "angle": 0, + "content": "3. Extensive experiments on multiple benchmarks validate that our MetaViewer achieves comparable performance to the existing methods in two downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.176, + 0.895, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.384, + 0.642, + 0.4 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.409, + 0.806, + 0.426 + ], + "angle": 0, + "content": "2.1. Multi-view representation learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.612 + ], + "angle": 0, + "content": "Representation learning is not a new topic and plays a critical role in numerous downstream tasks [29, 36, 49]. This work focuses on multi-view representation in unsupervised deep learning scope [56]. Related studies can be roughly summarized into two categories [54]. One is the deep extension of traditional methods, where representative ones include deep canonical correlation analysis (DCCA) and its variants [44, 57]. DCCA [2] intends to discover the nonlinear mapping for two views to a common space in which their correlations are maximally preserved. These methods benefit from a sound theoretical foundation, but also usually have strict restrictions on the number and form of views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.897, + 0.826 + ], + "angle": 0, + "content": "Another alternative is deep multi-view learning [5, 32]. Early deep-based approaches attempted to design effective architectures for multiple views, such as CNN-based [12, 42] and GAN-based models [18, 52]. Recent works resort to leveraging mutual or comparative information to derive the constraint for parameters [3, 27, 55]. Most of them follow the pipeline from view-specific features to unified representation. In contrast, our MetaViewer learns by observing the learning from the uniform to the specific. The most related work is \\(\\mathrm{AE^2}\\)-Nets [60], which treats the unified representation as a set of trainable parameters that are optimized during the degradation process of each view. The essential difference is that we learn the data-oriented fusion strategy rather than optimizing the unified representation itself." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.649, + 0.849 + ], + "angle": 0, + "content": "2.2. Meta-learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Optimization-based meta-learning is a classic application of bi-level optimization designed to learn task-level knowledge to quickly handle new tasks [19, 21]. A typical work," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11591" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.093, + 0.885, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.356, + 0.893, + 0.412 + ], + "angle": 0, + "content": "Figure 2. The overall framework of MetaViewer, contains (a) the meta-split of multi-view data and three modules: (b) embedding module, (c) representation module, and (d) self-supervised module. These modules are trained with a bi-level optimization. Inner-level (dark gray arrows) learns the view-specific reconstruction on the support set, and outer-level (red arrows) updates the entire model to learn the fusion scheme and the unified representation \\( H \\) for downstream tasks by validating on the query set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.474, + 0.667 + ], + "angle": 0, + "content": "MAML [13], learns a set of initialization parameters to solve different tasks with a few steps of updates. Similar meta paradigm has been used to learn other manually designed parts, such as the network structure [28], optimizer [62], and even sample weight [40, 41, 48]. Similarly, we try to metalearn the fusion rule of multi-view features for a unified representation. There also exist some works that consider both meta-learning paradigms and multi-view data [14, 31, 45]. However, they are dedicated to exploiting the rich information contained in multiple views to improve the performance of the meta-learner in a few-shot or self-supervised scenario. Instead, we train a meta-learner to derive high-quality shared representations from multi-view data. To the best of our knowledge, this could be the first work to learn multi-view representation with a meta-learning paradigm." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.679, + 0.206, + 0.695 + ], + "angle": 0, + "content": "3. MetaViewer" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.704, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Given a set of unlabeled multi-view data, \\( D = \\{x^{i} \\in \\mathbb{R}^{d_{x}}\\}_{i=1}^{N} \\), where \\( N \\) and superscript \\( i \\) are the number and index of entity samples, respectively. Each sample \\( x^{i} = \\{x_{1}^{i}, x_{2}^{i}, \\ldots, x_{v}^{i}\\}_{v=1}^{V} \\) contains \\( V \\) views, where subscript \\( v \\) is the index of views. Our goal is to learn a unified, high-quality representation \\( H \\) for each entity by observing these views and filtering view-private information as much as possible. The overall framework of our MetaViewer is shown in Fig. 2, including three main modules and a bi-level optimization process. The outer-level trains a meta-learner to learn an optimal fusion function and derive the unified representation \\( H \\), and the inner-level reconstructs original views from \\( H \\) in a few update steps, which explicitly models and separates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.895, + 0.5 + ], + "angle": 0, + "content": "view-private information and ensure the quality of the representation. In the following subsections, we first introduce the entire structure of the MetaViewer and then elaborate on the bi-level optimization process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.508, + 0.692, + 0.523 + ], + "angle": 0, + "content": "3.1. The entire structure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.894, + 0.637 + ], + "angle": 0, + "content": "**Embedding module** aims to transform heterogeneous views into the latent feature space, where transformed view embeddings have the same dimension as each other. To this end, we conduct a view-specific embedding function \\( f_{v} \\) for each view, where \\( v = 1,2,\\dots ,V \\). Given the \\( v \\)-th view data \\( x_{v} \\) of the entity \\( x \\), the corresponding embedding \\( z_{v}\\in \\mathbb{R}^{d} \\) can be computed by" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.648, + 0.893, + 0.665 + ], + "angle": 0, + "content": "\\[\nz _ {v} = f _ {v} \\left(x _ {v}, \\theta_ {f _ {v}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.704 + ], + "angle": 0, + "content": "where \\( f_{v} \\) is typically instantiated as a multi-layer neural network with learnable parameters \\( \\theta_{f_v} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Representation learning module maps the obtained embedding to the view-specific or unified representation by constructing a meta-learner \\( m \\) (i.e., MetaViewer). Note that the two types of representations are both derived from the same set of parameters, but at different learning stages (see 3.2). The view-specific representation is obtained at the inner level, where MetaViewer receives the embedding of the \\( v \\)-th view and outputs the corresponding representation \\( h_v \\). In contrast, the unified representation \\( H \\) is learned by MetaViewer from all view embeddings at the outer level and is ultimately used for downstream tasks. Therefore, the structure of MetaViewer should be flexible enough to meet the two requirements above simultaneously. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "MetaViewer is implemented as a channel-oriented 1-d convolutional layer (C-Conv) with a non-linear function (e.g., ReLU [17]), as shown in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.138, + 0.47, + 0.213 + ], + "angle": 0, + "content": "On the one hand, at the outer level, we first concatenate all view embeddings at the channel level, i.e., the number of the channel in the concatenated embedding is equal to the number of views, and then train the MetaViewer to learn the fusion of cross-view information \\( H \\in \\mathbb{R}^{d_h} \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.226, + 0.47, + 0.243 + ], + "angle": 0, + "content": "\\[\nH = m \\left(z _ {c a t}, \\omega\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.253, + 0.469, + 0.314 + ], + "angle": 0, + "content": "where \\(z_{cat} \\in \\mathbb{R}^{d \\times V}\\) indicates the concatenated embedding and \\(\\omega\\) is the parameter of MetaViewer. On the other hand, the inner-level initializes and trains MetaViewer to learning the representation \\(h_v \\in \\mathbb{R}^{d_h}\\) of the \\(v\\)-th view via" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.327, + 0.47, + 0.344 + ], + "angle": 0, + "content": "\\[\nh _ {v} = m _ {v} \\left(z _ {v}, \\omega_ {v}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.356, + 0.469, + 0.416 + ], + "angle": 0, + "content": "where \\( m_v \\) denotes the learner with the learnable parameter \\( \\omega_v \\) for handling the \\( v \\)-th view in the inner level, also known as the base learner to avoid confusion. It is worth noting that \\( \\omega_v \\) is initialized by a subset of the MetaViewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Self-supervised module provides effectively supervised information for model training by constructing a series of pretext tasks1. These tasks serve as the heads of our framework. Given the input \\( z \\), the output \\( s \\) can be formalized as" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.489, + 0.47, + 0.506 + ], + "angle": 0, + "content": "\\[\ns = t \\left(z, \\theta_ {t}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.518, + 0.47, + 0.653 + ], + "angle": 0, + "content": "where \\( t \\) and \\( \\theta_t \\) are the network and its parameters involved in pre-text tasks, respectively. Taking the reconstruction head as an example, it achieves the reconstruction object by re-mapping the representation back to the original view space. In this case, the \\( s := \\hat{x}_v \\) indicates the reconstruction result for input \\( x_v \\), and \\( t(\\cdot, \\theta_t) := t_v(\\cdot, \\theta_{t_v}) \\) is a multi-layer neural network with the opposite structure to the embedding module. More similar pre-text tasks have been studied in previous works [55, 57], and this is not the focus of this work." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.37, + 0.68 + ], + "angle": 0, + "content": "3.2. Training via bi-level optimization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.47, + 0.854 + ], + "angle": 0, + "content": "We now have the entire structure, which can be end-to-end trained following the proposed uniform-to-specific pipeline to derive the unified multi-view representation. The training flowchart and step are shown in the top row in Fig. 3, where the learning of view-private features and unified representation are formulated as a nested bi-level optimization problem. Inner-level focus on independent training on a specific view, which receives a given representation and reconstructs it to each view. Outer-level updates the meta-learner to find the optimal fusion rule through observing the training processes across all views. This pipeline decouples the unified entity" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Constructing pre-text tasks is a classic strategy in self-supervised learning [1], which emphasizes the use of intrinsic properties within or between data as ground-truth to provide supervised information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "representations from view-specific features and enables data-driven fusion. Before the detailed description, we introduce a split strategy for multi-view data in meta-learning style to adapt the bi-level training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.894, + 0.272 + ], + "angle": 0, + "content": "Meta-split of multi-view data. Consider a batch multi-view entity samples \\(\\{D_v^{batch}\\}_{v=1}^V\\) from set \\(D\\). We randomly and proportionally divide it into two disjoint subsets, marked as support set \\(S = \\{S_v\\}_{v=1}^V\\) and query set \\(Q = \\{Q_v\\}_{v=1}^V\\), respectively. As shown in 2 (a), support set is used in inner-level for leaning view-specific information, so the sample attributes in it could be ignored. In contrast, query set retains both view and sample attributes for outer-level optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.894, + 0.455 + ], + "angle": 0, + "content": "Inner-level optimization. Without loss of generality, take the inner-level update after the \\(o\\)-th outer-level optimization as an example. Let \\(\\omega^o\\) be the lasted parameters of the MetaViewer, and \\(\\theta_v^o = \\{\\theta_{f_v}^o,\\theta_{t_v}^o\\}\\) denotes the lasted parameters in embedding and self-supervised modules for brevity. We first initial base learner from meta-learner, i.e., \\(\\omega_v^0 = \\omega^o\\), and make a copy of \\(\\theta_v^o\\) for \\(\\tilde{\\theta}_v\\). Note that the copy means gradients with respect to the \\(\\theta_v^o\\) will not be backpropagated to \\(\\tilde{\\theta}_v\\) and vice versa. Thus, \\(\\omega_v^0\\) and \\(\\tilde{\\theta}_v\\) form the initial states of parameters in the inner-level optimization. Suppose \\(\\mathcal{L}_v^{in}\\) is the loss function of the inner-level with respect to the \\(v\\)-th view, the corresponding update goal is" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.464, + 0.892, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\omega_ {v} ^ {*} \\left(\\omega^ {o}\\right) = \\arg \\min \\mathcal {L} _ {v} ^ {i n} \\left(\\omega_ {v} \\left(\\omega^ {o}\\right), \\tilde {\\phi} _ {v}; S _ {v}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.524 + ], + "angle": 0, + "content": "Consider a gradient descent strategy (e.g., SGD [4]), we can further write the update process of \\(\\omega_{v}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.533, + 0.892, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\omega_ {v} ^ {i} = \\omega_ {v} ^ {i - 1} - \\beta \\frac {\\partial \\mathcal {L} _ {v} ^ {i n}}{\\partial \\omega_ {v} ^ {i - 1}}, \\dots , \\omega_ {v} ^ {0} = \\omega^ {o}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.605 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(i\\) denote the learning rate and iterative step of inner-level optimization, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.894, + 0.696 + ], + "angle": 0, + "content": "Outer-level optimization. After several inner-level updates, we obtain a set of optimal view-specific parameters for the support set. Outer-level then updates the meta-learner, embedding, and self-supervised modules by training on the query set. With the loss function \\(\\mathcal{L}^{out}\\), the outer-level optimization goal is" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.705, + 0.892, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\omega^ {*}, \\left\\{\\theta_ {v} ^ {*} \\right\\} _ {v = 1} ^ {V} = \\arg \\min \\mathcal {L} ^ {\\text {o u t}} \\left(\\left\\{\\omega_ {v} ^ {*} (\\omega), \\phi_ {v} \\right\\} _ {v = 1} ^ {V}; Q\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.894, + 0.824 + ], + "angle": 0, + "content": "By alternately optimizing Eq. 6 and Eq. 8, we end up with the optimal meta-parameters \\(\\omega^{*}\\) and a set of view-specific parameters \\(\\{\\theta_v^*\\}_{v = 1}^V\\). Given a test sample \\(x_{test}\\), the corresponding unified representation is derived by feeding it sequentially into the embedding module and the meta-learner. The overall framework of MetaViewer is shown in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.887, + 0.849 + ], + "angle": 0, + "content": "3.3. Specific-to-uniform versus uniform-to-specific" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We discuss the difference between specific-to-uniform and our uniform-to-specific pipeline through three aspects shown in Fig. 3, including the overall flowchart, update steps," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11593" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.092, + 0.4, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 The framework of our MetaViewer." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.109, + 0.471, + 0.167 + ], + "angle": 0, + "content": "Require: Training dataset \\(D\\), meta parameters \\(\\omega\\), base parameters \\(\\{\\omega_v\\}_{v=1}^V\\), view-specific parameters \\(\\{\\theta_v\\}_{v=1}^V\\), the number of views \\(V\\), the iteration step in inner-level optimization \\(T\\)." + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.169, + 0.252, + 0.187 + ], + "angle": 0, + "content": "1: Initialize \\(\\omega, \\{\\theta_v\\}_{v=1}^V\\);" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.187, + 0.23, + 0.2 + ], + "angle": 0, + "content": "2: while not done do" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.201, + 0.222, + 0.214 + ], + "angle": 0, + "content": "3: # Outer-level" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.216, + 0.403, + 0.231 + ], + "angle": 0, + "content": "4: Sample and meta-split a batch set from \\(D\\):" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.23, + 0.377, + 0.248 + ], + "angle": 0, + "content": "5: \\(\\{D_v^{batch}\\}_{v = 1}^V = \\{\\tilde{S}_v\\}_{v = 1}^V +\\{Q_v\\}_{v = 1}^V.\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.248, + 0.258, + 0.261 + ], + "angle": 0, + "content": "6: for \\(t = 1,\\dots ,T\\) do" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.262, + 0.277, + 0.276 + ], + "angle": 0, + "content": "7: for \\( v = 1, \\ldots, V \\) do" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.277, + 0.255, + 0.29 + ], + "angle": 0, + "content": "8: #Inner-level" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.291, + 0.337, + 0.307 + ], + "angle": 0, + "content": "9: Initialize \\(\\omega_{v} = \\omega ,\\tilde{\\theta}_{v} = \\theta_{v};\\)" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.308, + 0.379, + 0.323 + ], + "angle": 0, + "content": "10: Optimize \\(\\omega_v(\\omega)\\) and \\(\\theta_v\\) via Eq. 6." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.323, + 0.193, + 0.335 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.338, + 0.177, + 0.35 + ], + "angle": 0, + "content": "12: end for" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.352, + 0.358, + 0.369 + ], + "angle": 0, + "content": "13: Optimize \\(\\omega\\) and \\(\\{\\phi_v\\}_{v=1}^V\\) via Eq. 8." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.369, + 0.177, + 0.381 + ], + "angle": 0, + "content": "14: end while" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.169, + 0.403, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.501 + ], + "angle": 0, + "content": "and key gradients. For clarity, we here slightly confuse the notation and let \\(\\theta_v\\) and \\(\\mathcal{L}_v^{in}\\) correspond to the view-specific parameters and loss functions in specific-to-uniform pipeline, respectively. Similarly, \\(\\omega\\) and \\(\\mathcal{L}^{out}\\) denote the view-shared parameters and loss functions, respectively. Let us start by reviewing two branches in specific-to-uniform framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.471, + 0.56 + ], + "angle": 0, + "content": "One-level methods is a classical branch in specific-to-uniform framework, where parameters \\(\\{\\theta_v\\}_{v=1}^V\\) and \\(\\omega\\) are simultaneously optimized by combining the functions \\(\\mathcal{L}_v^{in}\\) and \\(\\mathcal{L}^{out}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.652 + ], + "angle": 0, + "content": "Multi-level methods also follow the specific-to-uniform framework, the difference is that they hierarchically learn specific and shared features to avoid interference. That is, parameters \\(\\theta_v\\) are updated at the low level by minimizing the \\(\\mathcal{L}_v^{in}\\), and parameters \\(\\omega\\) are updated at the high level by minimizing the \\(\\mathcal{L}^{out}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.471, + 0.744 + ], + "angle": 0, + "content": "Bi-level MetaViewer further decouples view-private information, except for view-specific and view-sharing components. Consider the simplest case, where the base learner updates only one step on each view in the inner loop, i.e., \\(\\omega_{v}^{*} = \\omega -\\beta \\nabla_{\\omega}\\mathcal{L}_{v}^{in}(\\omega)\\). Thus, the final optimization of \\(\\omega\\) in the outer level in Eq. 8 can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.754, + 0.471, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\omega} \\sum_ {v} \\mathcal {L} ^ {o u t} (\\omega_ {v} ^ {*}) = \\sum_ {v} \\mathcal {L} ^ {o u t} \\left(\\omega - \\beta \\nabla_ {\\omega} \\mathcal {L} _ {v} ^ {i n} (\\omega)\\right). \\quad (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Intuitively, the optimal \\(\\omega\\) is expected to be close enough to each view that the view-specific representation can be produced after once updating. In other words, our uniform-to-specific framework requires explicit modeling the gap caused by view-private information when mapping from a unified representation to a particular view, which is achieved by inner-level optimization in MetaViewer." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.092, + 0.89, + 0.167 + ], + "angle": 0, + "content": "
our bi-level MetaViewer uniform-to-specific
Q→θv→ωv(ω)→Lvinωv(ω)→ωv*(ω)∂Lvin/∂ωv(ω)
DInner-Level
S→θv→Σv→ω→Loutθv→θv*, ω→ω*∂Lout/∂θv, ∂Lout/∂2ω
" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.168, + 0.89, + 0.227 + ], + "angle": 0, + "content": "
one-level methodsspecific-to-uniform
D→θv→Σv→ωLvnθv→θ*, ω→ω*∂(Lout+Lvn)/∂θv, ∂(Lout+∑v) / ∂ω
" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.228, + 0.889, + 0.282 + ], + "angle": 0, + "content": "
two-/multi-level methods
D→θvLinvθv→θ*Low-Level
ω→ω*High-Level
∂Lu/∂θv
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.284, + 0.633, + 0.297 + ], + "angle": 0, + "content": "(a) Flowcharts" + }, + { + "type": "table_caption", + "bbox": [ + 0.699, + 0.284, + 0.751, + 0.297 + ], + "angle": 0, + "content": "(b) Steps" + }, + { + "type": "table_caption", + "bbox": [ + 0.797, + 0.284, + 0.871, + 0.297 + ], + "angle": 0, + "content": "(c) Gradients" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.309, + 0.895, + 0.409 + ], + "angle": 0, + "content": "Figure 3. Comparison of two frameworks in terms of flowchart, update steps, and key gradients. The top row is our bi-level MetaViewer with uniform-to-specific framework, where red and black arrows indicate the flow at outer-level and inner-level, respectively. The bottom two rows show two main branches in uniform-to-specific framework, i.e., one-level methods [47] and two- or multi-level methods [51]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.433, + 0.83, + 0.451 + ], + "angle": 0, + "content": "3.4. The instances of the objective function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.894, + 0.533 + ], + "angle": 0, + "content": "Our uniform-to-specific framework emphasizes learning from the training process in the inner loop. In unsupervised multi-view learning, this process typically refers to the reconstruction on views, thus the \\(\\mathcal{L}_v^{in}\\) is specified as the reconstruction loss [27, 60]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.543, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v} ^ {i n} = \\mathcal {L} _ {v} ^ {r e c} \\left(S _ {v}, S _ {v} ^ {r e c}\\right) = \\| S _ {v} - S _ {v} ^ {r e c} \\| _ {F} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.894, + 0.648 + ], + "angle": 0, + "content": "While parameters updated at the outer level can be constrained by richer self-supervised feedbacks, as maintained in Sec. 3.1. Here we provide two instances of the outer-level loss function to demonstrate how MetaViewer can be extended with different learning objectives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.894, + 0.692 + ], + "angle": 0, + "content": "MVer-R adopts the same reconstruction loss as the inner-level, and \\(\\mathcal{L}^{out} = \\sum_{v}\\mathcal{L}_{v}^{rec}(Q_{v},Q_{v}^{rec})\\), which is the purest implementation of MetaViewer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.894, + 0.793 + ], + "angle": 0, + "content": "MVer-C additionally utilizes a contrastive objective, where the similarities of views belonging to the same entity (i.e., positive pairs) should be maximized and those of different entities (i.e., negative pairs) should be minimized, i.e., \\(\\mathcal{L}^{out} = \\sum_{v}\\left(\\mathcal{L}_{v}^{rec} + \\sum_{v^{\\prime},v^{\\prime}\\neq v}\\mathcal{L}_{v,v^{\\prime}}^{con}\\right)\\). Following previous work [18, 51, 55], the contrastive loss \\(\\mathcal{L}_{v,v^{\\prime}}^{con}\\) is formed as" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.803, + 0.906, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v, v ^ {\\prime}} ^ {c o n} = - \\frac {1}{N _ {Q}} \\sum_ {i = 1} ^ {N _ {Q}} \\log \\frac {e ^ {d \\left(q _ {v} ^ {i} , q _ {v ^ {\\prime}} ^ {i}\\right) / \\tau}}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {Q}} e ^ {d \\left(q _ {v} ^ {i} , q _ {v} ^ {j}\\right) / \\tau} + \\sum_ {j = 1} ^ {N _ {Q}} e ^ {d \\left(q _ {v} ^ {i} , q _ {v ^ {\\prime}} ^ {j}\\right) / \\tau}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "where \\( q_v^i \\) is the \\( v \\)-th view of the \\( i \\)-th query sample and \\( d \\) is the similarity metric (e.g., cosine similarity [6]). The \\( N_Q \\) and \\( \\tau \\) denote the number of query set samples and the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11594" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.089, + 0.813, + 0.197 + ], + "angle": 0, + "content": "
Datasets#Views#Classes#Samples (train, val, test)View Dimensions
Handwirtten2102000 (1200, 400, 400)240; 216
RGB-D250500 (300, 100, 100)12288; 4096
Animal25010158 (6074, 2011, 2073)4096; 4096
Fashion-MV31010000 (6000, 2000, 2000)784; 784; 784
Caltech101-206202386 (1425, 469, 492)48; 40; 254; 1984; 512; 928
" + }, + { + "type": "table_caption", + "bbox": [ + 0.297, + 0.207, + 0.672, + 0.221 + ], + "angle": 0, + "content": "Table 1. The attributes for all datasets used in our experiments." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.235, + 0.89, + 0.407 + ], + "angle": 0, + "content": "
MethodsHandwrittenRGB-DAnimalFashion-MVCaltech101-20
ACCNMIARIACCNMIARIACCNMIARIACCNMIARIACCNMIARI
Baseline43.5552.1925.6848.3080.9704.9545.7966.2233.8750.9145.9832.0936.6348.9623.54
DCCA [2] (2013)57.2569.8052.1551.0082.9952.0260.3468.5146.1070.7080.4261.8038.6250.8822.73
DCCAE [47] (2015)63.0075.0459.2948.0081.5848.3465.1766.8248.8371.0581.1262.3436.5952.2425.25
MIB [8] (2020)63.2567.5852.1650.0081.1351.2760.8961.9852.4957.2073.8347.6235.9847.0022.18
WTNNM [15] (2020)75.6977.6261.5152.5481.9552.4671.9271.4361.0683.3386.3077.9939.7455.5826.92
TLRR [23] (2021)78.6077.9966.4052.8782.3853.3769.7973.0161.2379.5484.0178.5440.1852.3827.60
MFLVC [51] (2022)64.0064.5348.8553.0083.3154.0774.1076.0864.2883.2088.7578.9336.5958.3626.87
DCP [27] (2022)66.2570.5656.1052.0082.0452.6472.7774.0265.4862.6068.3854.3036.7944.3723.50
MVer-R (ours)75.0078.5367.2153.0082.4153.0476.4978.2565.2880.8088.1375.0541.8758.5229.19
MVer-C (ours)86.2578.9672.2557.0084.9757.0775.9278.0166.0785.4088.7680.0745.1260.8635.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.417, + 0.868, + 0.431 + ], + "angle": 0, + "content": "Table 2. Clustering results of all methods on six datasets. Bold and underline denote the best and second-best results, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.471, + 0.503 + ], + "angle": 0, + "content": "temperature parameter, respectively. Note that, the derived meta representation can also be used in contrastive learning as an additional novel view." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.521, + 0.21, + 0.538 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.471, + 0.699 + ], + "angle": 0, + "content": "In this section, we present extensive experimental results to validate the quality of the unified representation derived from our MetaViewer. The remainder of the experiments are organized as follows: Subsection 4.1 lists datasets, compared methods, and implementation details. Subsection 4.2 compares the performance of our method with classical and state-of-the-art methods on two common downstream scenarios, clustering and classification tasks. Comparisons with manually designed fusion and ablation studies are shown in Subsections 4.3 and 4.4, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.267, + 0.728 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Datasets. To comprehensively evaluate the effectiveness of our MetaViewer, we conduct five multi-view benchmarks in experiments. All datasets are scaled to \\([0,1]\\) and split into training, validation, and test sets in the ratio of \\(6:2:2\\), as shown in Tab. 1. Handwritten contains 2,000 handwritten digital images from 0 to 9, where two types of descriptors, i.e., 240-D pixel average in \\(2\\times 3\\) windows and 216-D profile correlations, are selected as two views [60]. RGB-D dataset contains visual and depth images of 300 distinct objects across 50 categories [46, 63]. Two views are obtained by flattening the \\(64\\times 64\\times 3\\) color images and \\(64\\times 64\\) depth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.895, + 0.608 + ], + "angle": 0, + "content": "images. Animal consists of 10158 images from 50 classes with two views. Unlike other datasets, these two views are features extracted by deep networks [59]. Fashion-MV is an image dataset that contains 10 categories with a total of 30,000 fashion products. It has three views and each of which consists of 10,000 gray images sampled from the same category [50]. Caltech101-20 is a subset of the Caltech101 image set [9], which consists of 2,386 images of 20 subjects. Six features are used, including Gabor, Wavelet Moments, CENTRIST, HOG, GIST, and LBP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.896, + 0.869 + ], + "angle": 0, + "content": "Compared methods. We compare the performance of MetaViewer with seven representative multi-view learning methods, including two classical methods (DCCA [2] and DCCAE [47]) and five state-of-the-art methods (WTNNM [15], TLRR [23], MIB [8], MFLVC [51] and DCP [27]). Among them, DCCA and DCCAE are the deep extensions of traditional correlation strategies. WTNNM and TLRR are representative tensor-based methods. MIB is a typical generative method with mutual information constraints. DCP learns unified representation in both unsupervised and supervised scenarios, and we report the unsupervised version for a fair comparison. In particular, MFLVC also notices the view-private redundant information and designs a multilevel feature network for clustering tasks. In addition, we also compare the model without MetaViewer as a baseline, where the unified representation is obtained by concatenating view-specific features output by the embedding module." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Implementation details. For a fair comparison, all methods are trained from scratch, and all deep-network-based" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11595" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.251 + ], + "angle": 0, + "content": "
MethodsHandwrittenRGB-DAnimalFashion-MVCaltech101-20
ACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-score
Baseline87.0087.3287.0414.0006.4507.7071.7865.0263.1087.5087.6387.5475.4151.9546.21
DCCA [2] (2013)88.2589.2088.0530.0021.1022.0462.8461.6160.6584.9085.2283.5471.5445.2739.81
DCCAE [47] (2015)90.0090.4889.9224.0016.0016.9166.8262.5362.8385.3585.9783.8471.5460.5743.25
MIB [8] (2020)79.0083.9078.5233.0028.5027.3759.3263.7862.1386.8086.8086.5572.7261.6452.47
WTNNM [15] (2020)96.7796.2096.3645.0047.1843.9069.9067.4065.4894.5094.5094.5883.2780.4974.78
TLRR [23] (2021)97.0097.0597.1749.0053.3247.6671.9069.5268.3596.3596.2896.4085.5577.3075.24
MFLVC [51] (2022)94.0094.2094.0144.0046.0941.8175.6272.1770.8196.5096.5296.4985.3771.8369.07
DCP [27] (2022)97.2597.3097.2437.0028.8730.7877.9573.4370.0189.2582.0682.9092.4889.4184.58
MVer-R (ours)97.0097.0897.0051.0053.6548.7377.6973.9171.2296.8596.3796.4892.2889.4684.21
MVer-C (ours)97.7597.9097.7556.0055.2052.7878.0374.5671.5597.7096.7897.0792.1690.6885.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.26, + 0.878, + 0.275 + ], + "angle": 0, + "content": "Table 3. Classification results of all methods on six datasets. Bold and underline denote the best and second-best results, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.474, + 0.574 + ], + "angle": 0, + "content": "methods share the same embedding backbone, following the previous work [51]. For two-view-based methods, such as DCCA, DCCAE, and MIB, we report the results of the two best-performing views. The final unified representation \\( H \\) in alignment-based methods is a concatenation of the processed features, and the representation dimension is specified as \\( d_h = 256 \\). The results on clustering and classification tasks are obtained by subsequent K-means and SVM classifier, respectively. For MetaViewer, we train 2,000 epochs for all benchmarks, and set the batch size is 32 for RGBD and 128 for others. The learning rates in outer- and inner-level are set to \\( 10^{-4} \\) and \\( 10^{-3} \\), respectively. We use a single convolutional layer with 32 kernels of size 5 as the meta-learner, and we set the proportion of support set and the update step in the inner level to 0.2 and 2, respectively. All experiments have been verified using the PyTorch library on a single RTX3090. Code is available at https://github.com/xxLifeLover/MetaViewer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.379, + 0.603 + ], + "angle": 0, + "content": "4.2. Performance on downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.822 + ], + "angle": 0, + "content": "Clustering results. Tab. 2 lists the results of the clustering task, where the performance is measured by three standard evaluation metrics, i.e., Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI). A higher value of these metrics indicates better clustering performance. It can be observed that (1) our MVer-C variant significantly outperforms other compared methods on all benchmarks; (2) the second-best results appear between MVer-R and MFLVC, both of which explicitly separate the view-private information; (3) a larger number of categories and views is the main reason for the degradation of clustering performance, and our Metaviewer improves most significantly in such a scenario (e.g., Fashion-MV and Caltech101-20)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Classification results. Tab. 3 lists the results of the classification task, where three common metrics are used, including Accuracy, Precision, and F-score. A higher value indicates better classification performance. Similar to the clustering results, two variants of MetaViewer significantly" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.299, + 0.89, + 0.404 + ], + "angle": 0, + "content": "
StrategiesRulesACC↑NMI↑ARI↑MSE↓
Sumzx + zy69.2571.8959.021.84
Maxmax(zx, zy)80.7573.9363.75-
Concat.cat[zx, zy]78.7572.0261.521.77
Linearl(zx, zy, θl)85.0077.4069.714.74
C-Convm(zx, zy, ω)69.7565.2151.332.37
MetaViewermeta-learning86.2578.9672.252.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.413, + 0.866, + 0.428 + ], + "angle": 0, + "content": "Table 4. Clustering resulting on the Handwritten dataset." + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.445, + 0.87, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.599, + 0.895, + 0.643 + ], + "angle": 0, + "content": "Figure 4. Class activate maps on RGB-D samples for different fusion strategies, including single views, Sum, and our MetaViewer. View 1 and view 2 are the RGB and depth view, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.894, + 0.76 + ], + "angle": 0, + "content": "outperform the comparison methods. It is worth noting that (1) DCP learns unified representation and therefore achieves the second-best result instead of MFLVC. (2) The number of categories is the main factor affecting the classification performance, and our method obtains the most significant improvement in the RGB-D dataset with up to 50 classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.874, + 0.789 + ], + "angle": 0, + "content": "4.3. Comparison with manually designed fusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "As mentioned in 3.3, MetaViewer essentially learns an optimal fusion rule that preserves as much consistency and complementary information as possible while filtering out view-private information. To verify this, we compare it with commonly used fusion strategies [26], including sum, maxima, concatenation, linear layer and \\( C \\)-Conv. The former three are the specified fusion rules without trainable param-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11596" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.273, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.204, + 0.248, + 0.217 + ], + "angle": 0, + "content": "(a) Effect of network width" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.09, + 0.465, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.204, + 0.426, + 0.217 + ], + "angle": 0, + "content": "(b) Effect of kernel size" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.228, + 0.471, + 0.256 + ], + "angle": 0, + "content": "Figure 5. Effect of meta-learner architectures with different depth, width, and kernel size on classification accuracy." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.274, + 0.275, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.387, + 0.26, + 0.399 + ], + "angle": 0, + "content": "(a) Proportion of support set" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.274, + 0.466, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.386, + 0.458, + 0.399 + ], + "angle": 0, + "content": "(b) Steps of inner-level update" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.412, + 0.47, + 0.44 + ], + "angle": 0, + "content": "Figure 6. Effect of (a) different meta-division ratios and (b) the number of inner-loop iterations on classification accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.636 + ], + "angle": 0, + "content": "eters, and the remaining two are the trainable fusion layer trained via the specific-to-uniform manner. Tab. 4 lists the clustering results and an additional MSE score on the Hand-written dataset with the same embedding and reconstruction network. We can observe that (1) trainable fusion layers outperform the hand-designed rules, and our MetaViewer yields the best performance; (2) the MSE scores listed in the last column indicate that the quality of the unified representation cannot be measured and guaranteed only with the reconstruction constraint, due to the view-private redundant information mixed in view-specific latent features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.471, + 0.713 + ], + "angle": 0, + "content": "To further explore the fusion preferences learned by MetaViewer, we visualize the class activation map for the RGB-D samples in Fig. 4. MetaViewer balances information from all views instead of just the salient one (view 1), ensuring a high-quality unified multi-view representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.242, + 0.74 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Meta-learner structures. We implement the meta-learner as a channel-level convolution structure in this work. Albeit simple, this layer can be considered as a universal approximator for almost any continuous function [40], and thus can fit a wide range of conventional fusion functions. To investigate the effect of network depth, width, and convolution kernel size on the performance of the representation, we alternately fix the 32 kernels and the \\(1 \\times 3\\) kernel size and show the classification results on Handwritten data in Fig. 5. It is clear that (1) the meta-learner works well with just a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "shallow structure, as shown in Fig. 5 (a), instead of gradually overfitting to the training data as the network deepens or widens, and (2) our MetaViewer is stable and insensitive to the hyper-parameters within reasonable ranges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.895, + 0.307 + ], + "angle": 0, + "content": "Meta-split ratios. Fig. 6 (a) shows the impact of the meta-split mentioned in Sec. 3.2 on the classification performance, where the proportion of support set is set from 0.1 to 0.9 in steps of 0.1, and the rest is query set. In addition to the single view, we also compare the sum and concat. fusion as baselines. MetaViewer consistently surpasses all baselines over the experimental proportion. In addition, fusion baselines are more dependent on the better-performing view at lower proportions, instead becoming unstable as the available query sample decreases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.894, + 0.448 + ], + "angle": 0, + "content": "Inner-level update steps. Another hyper-parameter is the number of iteration steps in inner-level optimization. More iterations mean a larger gap from the learned meta representation to the specific view space, i.e., coarser modeling of view-private information. Fig. 6 (b) shows the classification results with various steps, where \\( n \\) steps mean that the inner-level optimization is updated \\( n \\) times throughout the training. MetaViewer achieves the best results when using 1 steps, and remains stable within 15 steps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.619, + 0.485 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.727 + ], + "angle": 0, + "content": "This work introduced a novel meta-learning perspective for multi-view learning, and proposed a meta-learner, namely MetaViewer, to derive a high-quality unified representation for downstream tasks. In contrast to the prevailing specific-to-uniform pipeline, MetaViewer observes the reconstruction process from unified representation to specific views and essentially learns an optimal data-driven fusion that separates and filters out meaningless view-private information. The proposed framework is compatible with most excellent objective functions and pre-text tasks to cope with complex real-world scenarios. Extensive experimental results on clustering and classification tasks demonstrate the performance of meta-learned unified representation. In addition, we believe that our bi-level training pipeline is also promising for handling incomplete views, and leave it to future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.667, + 0.768 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This research was supported in part by the Natural Science Foundation of China (No. 62106129, 62176139, and 62177031), the Natural Science Foundation of Shandong Province (No. ZR2021QF053, ZR2021ZD15), and the China Postdoctoral Science Foundation (No. 2021TQ0195, 2021M701984). The author acknowledges Prof. Xiushan Nie of Shandong Jianzhu University (China) for the financial support and valuable discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11597" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.471, + 0.156 + ], + "angle": 0, + "content": "[1] Sara Atito Ali Ahmed, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. CoRR, abs/2104.03602, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] Galen Andrew, Raman Arora, Jeff A. Bilmes, and Karen Livescu. Deep canonical correlation analysis. In ICML, volume 28, pages 1247-1255, 2013. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.471, + 0.243 + ], + "angle": 0, + "content": "[3] Philip Bachman, R. Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In NeurIPS, pages 15509-15519, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[4] Léon Bottou. Large-scale machine learning with stochastic gradient descent. In COMPSTAT, pages 177-186, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.329 + ], + "angle": 0, + "content": "[5] Abhra Chaudhuri, Massimiliano Mancini, Yanbei Chen, Zeynep Akata, and Anjan Dutta. Cross-modal fusion distillation for fine-grained sketch-based image retrieval. In BMVC, page 499, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119, pages 1597-1607, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[7] Yanhua Cheng, Xin Zhao, Rui Cai, Zhiwei Li, Kaiqi Huang, and Yong Rui. Semi-supervised multimodal deep learning for RGB-D object recognition. In IJCAI, pages 3345-3351, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.471, + 0.486 + ], + "angle": 0, + "content": "[8] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushner, and Zeynep Akata. Learning robust representations via multiview information bottleneck. In ICLR, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[9] Li Fei-Fei, Robert Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Comput. Vis. Image Underst., 106(1):59-70, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.47, + 0.585 + ], + "angle": 0, + "content": "[10] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, pages 1933-1941, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[11] Fangxiang Feng, Xiaojie Wang, and Ruifan Li. Cross-modal retrieval with correspondence autoencoder. In ACM MM, pages 7-16, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[12] Yifan Feng, Zizhao Zhang, Xibin Zhao, Rongrong Ji, and Yue Gao. GVCNN: group-view convolutional neural networks for 3d shape recognition. In CVPR, pages 264-272, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.471, + 0.715 + ], + "angle": 0, + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, volume 70, pages 1126–1135, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[14] Kuiliang Gao, Bing Liu, Xuchu Yu, and Anzhu Yu. Unsupervised meta learning with multiview constraints for hyperspectral image small sample set classification. IEEE TIP, 31:3449-3462, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[15] Quanxue Gao, Wei Xia, Zhizhen Wan, De-Yan Xie, and Pu Zhang. Tensor-svd based graph learning for multi-view subspace clustering. In AAAI, pages 3930-3937, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[16] Yu Geng, Zongbo Han, Changqing Zhang, and Qinghua Hu. Uncertainty-aware multi-view representation learning. In AAAI, pages 7545-7553, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[17] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In AISTATS, volume 15, pages 315-323, 2011. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[18] Kaveh Hassani and Amir Hosein Khas Ahmadi. Contrastive multi-view representation learning on graphs. In ICML, volume 119, pages 4116-4126, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[19] Timothy M. Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE TPAMI, 44(9):5149-5169, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[20] Junlin Hu, Jiwen Lu, and Yap-Peng Tan. Sharable and individual multi-view metric learning. IEEE TPAMI, 40(9):2281-2288, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.894, + 0.264 + ], + "angle": 0, + "content": "[21] Mike Huisman, Jan N. van Rijn, and Aske Plaat. A survey of deep meta-learning. Artif. Intell. Rev., 54(6):4483-4541, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.894, + 0.323 + ], + "angle": 0, + "content": "[22] Xiaodong Jia, Xiao-Yuan Jing, Xiaoke Zhu, Songcan Chen, Bo Du, Ziyun Cai, Zhenyu He, and Dong Yue. Semi-supervised multi-view deep discriminant representation learning. IEEE TPAMI, 43(7):2496-2509, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.326, + 0.894, + 0.379 + ], + "angle": 0, + "content": "[23] Yuheng Jia, Hui Liu, Junhui Hou, Sam Kwong, and Qingfu Zhang. Multi-view spectral clustering tailored tensor low-rank representation. IEEE TCSVT, 31(12):4784-4797, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.894, + 0.438 + ], + "angle": 0, + "content": "[24] Xiao-Yuan Jing, Fei Wu, Xiwei Dong, Shiguang Shan, and Songcan Chen. Semi-supervised multi-view correlation feature learning with application to webpage classification. In AAAI, pages 1374-1381, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.894, + 0.481 + ], + "angle": 0, + "content": "[25] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. IEEE TPAMI, 39(4):664-676, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[26] Yingming Li, Ming Yang, and Zhongfei Zhang. A survey of multi-view representation learning. IEEE TKDE, 31(10):1863-1883, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.894, + 0.582 + ], + "angle": 0, + "content": "[27] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE TPAMI, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.894, + 0.641 + ], + "angle": 0, + "content": "[28] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In ICCV, pages 3295-3304, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.894, + 0.698 + ], + "angle": 0, + "content": "[29] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In CVPR, pages 3623–3632, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.894, + 0.742 + ], + "angle": 0, + "content": "[30] Shirui Luo, Changqing Zhang, Wei Zhang, and Xiaochun Cao. Consistent and specific multi-view subspace clustering. In AAAI, pages 3730-3737, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.894, + 0.786 + ], + "angle": 0, + "content": "[31] Yao Ma, Shilin Zhao, Weixiao Wang, Yaoman Li, and Irwin King. Multimodality in meta-learning: A comprehensive survey. KBS, 250:108976, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[32] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In CVPR, pages 9826–9836, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[33] Feiping Nie, Guohao Cai, Jing Li, and Xuelong Li. Autoweighted multi-view learning for image clustering and semi-supervised classification. IEEE TIP, 27(3):1501-1511, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11598" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[34] Xiushan Nie, Weizhen Jing, Chaoran Cui, Chen Jason Zhang, Lei Zhu, and Yilong Yin. Joint multi-view hashing for large-scale near-duplicate video retrieval. IEEE TKDE, 32(10):1951-1965, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[35] Xiushan Nie, Xingbo Liu, Jie Guo, Letian Wang, and Yilong Yin. Supervised discrete multiple-length hashing for image retrieval. IEEE TBD, 9(1):312-327, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[36] Zheyun Qin, Xiankai Lu, Xiushan Nie, Dongfang Liu, Yi-long Yin, and Wenguan Wang. Coarse-to-fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica, 10(5):1-17, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.472, + 0.307 + ], + "angle": 0, + "content": "[37] Mathieu Salzmann, Carl Henrik Ek, Raquel Urtasun, and Trevor Darrell. Factorized orthogonal latent spaces. In AISTATS, volume 9, pages 701-708, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.472, + 0.35 + ], + "angle": 0, + "content": "[38] Yang Shi, Xiushan Nie, Xingbo Liu, Lu Yang, and Yilong Yin. Zero-shot hashing via asymmetric ratio similarity matrix. IEEE TKDE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.472, + 0.394 + ], + "angle": 0, + "content": "[39] Yang Shi, Xiushan Nie, Xingbo Liu, Li Zou, and Yilong Yin. Supervised adaptive similarity matrix hashing. IEEE TIP, 31:2755-2766, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.472, + 0.451 + ], + "angle": 0, + "content": "[40] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pages 1917-1928, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.472, + 0.495 + ], + "angle": 0, + "content": "[41] Haoliang Sun, Chenhui Guo, Qi Wei, Zhongyi Han, and Yilong Yin. Learning to rectify for robust learning with noisy labels. PR, 124:108467, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[42] Kai Sun, Jiangshe Zhang, Junmin Liu, Ruixuan Yu, and Zengjie Song. DRCNN: dynamic routing convolutional neural network for multi-view 3d object recognition. IEEE TIP, 30:868-877, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.472, + 0.597 + ], + "angle": 0, + "content": "[43] Shiliang Sun, Wenbo Dong, and Qiuyang Liu. Multi-view representation learning with deep gaussian processes. IEEE TPAMI, 43(12):4453-4468, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.472, + 0.667 + ], + "angle": 0, + "content": "[44] Zhongkai Sun, Prathusha Kameswara Sarma, William A. Sethares, and Yingyu Liang. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In AAAI, pages 8992-8999, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.472, + 0.712 + ], + "angle": 0, + "content": "[45] Risto Vuorio, Shao-Hua Sun, Hexiang Hu, and Joseph J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. In NeurIPS, pages 1–12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.472, + 0.768 + ], + "angle": 0, + "content": "[46] Shiye Wang, Changsheng Li, Yanming Li, Ye Yuan, and Guoren Wang. Self-supervised information bottleneck for deep multi-view subspace clustering. CoRR, abs/2204.12496, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.472, + 0.813 + ], + "angle": 0, + "content": "[47] Weiran Wang, Raman Arora, Karen Livescu, and Jeff A. Bilmes. On deep multi-view representation learning. In ICML, volume 37, pages 1083-1092, 2015. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[48] Qi Wei, Lei Feng, Haoliang Sun, Ren Wang, Chenhui Guo, and Yilong Yin. Fine-grained classification with noisy labels. arXiv preprint arXiv:2303.02404, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[49] Qi Wei, Haoliang Sun, Xiankai Lu, and Yilong Yin. Self-filtering: A noise-aware sample selection for label noise with confidence penalization. In ECCV, pages 516-532, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[50] Jie Xu, Yazhou Ren, Huayi Tang, Zhimeng Yang, Lili Pan, Yang Yang, and Xiaorong Pu. Self-supervised discriminative feature learning for multi-view clustering. CoRR, abs/2103.15069, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[51] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In CVPR, pages 16030-16039, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[52] Fei Xue, Xin Wu, Shaojun Cai, and Junqiu Wang. Learning multi-view camera relocalization with graph neural networks. In CVPR, pages 11372-11381, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[53] Xiaowei Xue, Feiping Nie, Sen Wang, Xiaojun Chang, Bela Stantic, and Min Yao. Multi-view correlated feature learning by uncovering shared component. In AAAI, pages 2810-2816, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[54] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[55] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, pages 8824-8833, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.894, + 0.447 + ], + "angle": 0, + "content": "[56] Xiao Yu, Hui Liu, Yuxiu Lin, Yan Wu, and Caiming Zhang. Auto-weighted sample-level fusion with anchors for incomplete multi-view clustering. PR, 130:108772, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.894, + 0.503 + ], + "angle": 0, + "content": "[57] Yun-Hao Yuan, Jin Li, Yun Li, Jipeng Qiang, Yi Zhu, Xiaobo Shen, and Jianping Gou. Learning canonical f-correlation projection for compact multiview representation. In CVPR, pages 19238-19247, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.894, + 0.558 + ], + "angle": 0, + "content": "[58] Changqing Zhang, Huazhu Fu, Qinghua Hu, Xiaochun Cao, Yuan Xie, Dacheng Tao, and Dong Xu. Generalized latent multi-view subspace clustering. IEEE TPAMI, 42(1):86-99, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.894, + 0.603 + ], + "angle": 0, + "content": "[59] Changqing Zhang, Zongbo Han, Yajie Cui, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Cpm-nets: Cross partial multi-view networks. In NeurIPS, pages 557-567, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.894, + 0.644 + ], + "angle": 0, + "content": "[60] Changqing Zhang, Yeqing Liu, and Huazhu Fu. Ae2-nets: Autoencoder in autoencoder networks. In CVPR, pages 2577-2585, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[61] Liang Zhao, Tao Yang, Jie Zhang, Zhikui Chen, Yi Yang, and Z. Jane Wang. Co-learning non-negative correlated and uncorrelated features for multi-view data. IEEE TNNLS, 32(4):1486-1496, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.894, + 0.745 + ], + "angle": 0, + "content": "[62] Wenqing Zheng, Tianlong Chen, Ting-Kuei Hu, and Zhangyang Wang. Symbolic learning to optimize: Towards interpretability and scalability. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.894, + 0.787 + ], + "angle": 0, + "content": "[63] Pengfei Zhu, Binyuan Hui, Changqing Zhang, Dawei Du, Longyin Wen, and Qinghua Hu. Multi-view deep subspace clustering networks. CoRR, abs/1908.01978, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11599" + } + ] +] \ No newline at end of file diff --git a/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_origin.pdf b/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e62073c426dc241d567626c41a649fc7f56acf31 --- /dev/null +++ b/2023/MetaViewer_ Towards a Unified Multi-View Representation/cfe9d4c1-69d9-4f8e-8e6b-a1ece2428418_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:100c9f0c5d1e3196c0330e077446ff836210570514518ed1b979748aac33cc79 +size 1261831 diff --git a/2023/MetaViewer_ Towards a Unified Multi-View Representation/full.md b/2023/MetaViewer_ Towards a Unified Multi-View Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0faaa5a3b5ebaec80c0de529679f47c30ba84997 --- /dev/null +++ b/2023/MetaViewer_ Towards a Unified Multi-View Representation/full.md @@ -0,0 +1,384 @@ +# MetaViewer: Towards A Unified Multi-View Representation + +Ren Wang +Shandong University + +xxlifelover@gmail.com + +Haoliang Sun* Shandong University + +haolsun.cn@gmail.com + +Yuling Ma +Shandong Jianzhu University + +mayuling20@sdjzu.edu.cn + +Xiaoming Xi + +Shandong Jianzhu University + +fyzq10@126.com + +Yilong Yin* Shandong University + +ylyin@sdu.edu.cn + +# Abstract + +Existing multi-view representation learning methods typically follow a specific-to-uniform pipeline, extracting latent features from each view and then fusing or aligning them to obtain the unified object representation. However, the manually pre-specified fusion functions and aligning criteria could potentially degrade the quality of the derived representation. To overcome them, we propose a novel uniform-to-specific multi-view learning framework from a meta-learning perspective, where the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. Specifically, we formulated the extraction and fusion of view-specific latent features as a nested optimization problem and solved it by using a bi-level optimization scheme. In this way, MetaViewer automatically fuses view-specific features into a unified one and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views. Extensive experimental results in downstream classification and clustering tasks demonstrate the efficiency and effectiveness of the proposed method. + +# 1. Introduction + +Multi-view representation learning aims to learn a unified representation of the entity from its multiple observable views for the benefit of downstream tasks [35, 43, 58]. Each view acquired by different sensors or sources contains both view-shared consistency information and view-specific information [8]. The view-specific part further consists of complementary and redundant components, where the former can be considered as a supplement to the consistency information, while the latter is view-private and may be + +![](images/3523d0ccd48cabdf9ce9d9e848b184eeb4c48aef543a7882449f9307d0b67dd5.jpg) +(a) Joint Representation + +![](images/b5e2d22c9a68dfb1b4b4f9806e0e57c8341dbd12976b32c9cd65f98f17e8a774.jpg) +(b) Alignment Representation + +![](images/624596e2010e1985f387e1e8121a354d9d4bd77a4e31f525a9c56a9b2a812a6d.jpg) +(c) S&S Representation + +![](images/0e233cccfa9bdf64eb6bf3e9200f085f56138aa8241f42b95706b8700637dd40.jpg) +(d) Meta Representation +Figure 1. (a), (b) and (c) show three multi-view learning frameworks following the specific-to-uniform pipeline, where the unified representation is obtained by fusing or concatenating view-specific features. (d) illustrates our uniform-to-specific manner, where a meta-learner learns to fusion by observing reconstruction from unified representation to specific views. + +adverse for the unified representation [16]. Therefore, a high-quality representation is required to retain the consistency and complementary information, as well as filter out the view-private redundant ones [51]. + +Given the data containing two views, $x_{1}$ and $x_{2}$ , prevailing multi-view learning methods typically follow a specific-to-uniform pipeline and can be roughly characterized as: + +$$ +H := f \left(x _ {1}; W _ {f}\right) \circ g \left(x _ {2}; W _ {g}\right), \tag {1} +$$ + +where $f$ and $g$ are encoding (or embedding [27]) functions that map the original view data into the corresponding latent + +features with the trainable parameters $W_{f}$ and $W_{g}$ . These latent features are subsequently aggregated into the unified representation $H$ using the designed aggregation operator o. With different aggregation strategies, existing approaches can be further subdivided into the joint, alignment, and a combined share-specific (S&S) representation [22, 26]. + +Fig. 1 (a) - (c) show the above three branches of the specific-to-uniform framework. Joint representation focuses on the integration of complementary information by directly fusing latent features, where the $\circ$ is represented as fusion strategies, such as graph-based modules [33], neural networks [10], or other elaborate functions [7, 34]. While alignment representation seeks alignment between view-specific features to retain consistency information and specifies $\circ$ as the alignment operator, measured by distance [11], correlation [24, 47], or similarity [25, 38, 39]. The aligned features can be concatenated as the unified representation for downstream tasks. As a trade-off strategy, S&S representation explicitly distinguishes latent features into shared and specific representations and only aligns the shared part [20, 30, 53]. + +Despite demonstrating promising results, the specific-to-uniform framework inherently suffers from potential risks in the following two aspects: (1) Compared with the data-oriented fusion rules, manually pre-specified rules are designed to compute the unified representation by concatenating or fusing latent features, restricting its extensibility in complex real-world applications. (2) Even if we can find a well-performing fusion scheme, the sequential training manner limits the model to constrain the various component information separately. For example, it is difficult to automatically separate out view-private redundant information from the feature level [22, 37]. Recent researches have attempted to address the second issue by decomposing latent features using matrix factorization [61] or hierarchical feature modeling [51], but still cannot avoid manually pre-specified fusion or even decomposition strategies. + +In this work, we provide a new meta-learning perspective for multi-view representation learning and propose a novel uniform-to-specific framework to address these potential risks. Fig. 1 (d) shows the overall schematic. In contrast to the specific-to-uniform pipeline, the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. To train the MetaViewer, we first decouple the learning of view-specific latent features and unified meta representation. These two learning processes can then be formulated as a nested optimization problem and eventually solved by a bi-level optimization scheme. In detail, MetaViewer fuses view-specific features into a unified one at the outer level and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views at the inner level. In addition, our uniform-to-specific framework is compatible with most existing objective functions and pre-text + +tasks to cope with complex real-world scenarios. Extensive experiments validate that MetaViewer achieves comparable performance to the state-of-the-art methods in downstream clustering and classification tasks. The core contributions of this work are as follows. + +1. We provide a new meta-learning perspective and develop a novel uniform-to-specific framework to learn a unified multi-view representation. To the best of our knowledge, it could be the first meta-learning-based work in multi-view representation learning community. +2. We propose MetaViewer, a meta-learner that formulates the modeling of view-specific features and unified representation as a nested bi-level optimization and ultimately meta-learns a data-driven optimal fusion. +3. Extensive experiments on multiple benchmarks validate that our MetaViewer achieves comparable performance to the existing methods in two downstream tasks. + +# 2. Related Work + +# 2.1. Multi-view representation learning + +Representation learning is not a new topic and plays a critical role in numerous downstream tasks [29, 36, 49]. This work focuses on multi-view representation in unsupervised deep learning scope [56]. Related studies can be roughly summarized into two categories [54]. One is the deep extension of traditional methods, where representative ones include deep canonical correlation analysis (DCCA) and its variants [44, 57]. DCCA [2] intends to discover the nonlinear mapping for two views to a common space in which their correlations are maximally preserved. These methods benefit from a sound theoretical foundation, but also usually have strict restrictions on the number and form of views. + +Another alternative is deep multi-view learning [5, 32]. Early deep-based approaches attempted to design effective architectures for multiple views, such as CNN-based [12, 42] and GAN-based models [18, 52]. Recent works resort to leveraging mutual or comparative information to derive the constraint for parameters [3, 27, 55]. Most of them follow the pipeline from view-specific features to unified representation. In contrast, our MetaViewer learns by observing the learning from the uniform to the specific. The most related work is $\mathrm{AE^2}$ -Nets [60], which treats the unified representation as a set of trainable parameters that are optimized during the degradation process of each view. The essential difference is that we learn the data-oriented fusion strategy rather than optimizing the unified representation itself. + +# 2.2. Meta-learning + +Optimization-based meta-learning is a classic application of bi-level optimization designed to learn task-level knowledge to quickly handle new tasks [19, 21]. A typical work, + +![](images/a9c05b632a0d125a190a8ccae793b904072902733fb371f9901dc903f56baa0a.jpg) +Figure 2. The overall framework of MetaViewer, contains (a) the meta-split of multi-view data and three modules: (b) embedding module, (c) representation module, and (d) self-supervised module. These modules are trained with a bi-level optimization. Inner-level (dark gray arrows) learns the view-specific reconstruction on the support set, and outer-level (red arrows) updates the entire model to learn the fusion scheme and the unified representation $H$ for downstream tasks by validating on the query set. + +MAML [13], learns a set of initialization parameters to solve different tasks with a few steps of updates. Similar meta paradigm has been used to learn other manually designed parts, such as the network structure [28], optimizer [62], and even sample weight [40, 41, 48]. Similarly, we try to metalearn the fusion rule of multi-view features for a unified representation. There also exist some works that consider both meta-learning paradigms and multi-view data [14, 31, 45]. However, they are dedicated to exploiting the rich information contained in multiple views to improve the performance of the meta-learner in a few-shot or self-supervised scenario. Instead, we train a meta-learner to derive high-quality shared representations from multi-view data. To the best of our knowledge, this could be the first work to learn multi-view representation with a meta-learning paradigm. + +# 3. MetaViewer + +Given a set of unlabeled multi-view data, $D = \{x^{i} \in \mathbb{R}^{d_{x}}\}_{i=1}^{N}$ , where $N$ and superscript $i$ are the number and index of entity samples, respectively. Each sample $x^{i} = \{x_{1}^{i}, x_{2}^{i}, \ldots, x_{v}^{i}\}_{v=1}^{V}$ contains $V$ views, where subscript $v$ is the index of views. Our goal is to learn a unified, high-quality representation $H$ for each entity by observing these views and filtering view-private information as much as possible. The overall framework of our MetaViewer is shown in Fig. 2, including three main modules and a bi-level optimization process. The outer-level trains a meta-learner to learn an optimal fusion function and derive the unified representation $H$ , and the inner-level reconstructs original views from $H$ in a few update steps, which explicitly models and separates + +view-private information and ensure the quality of the representation. In the following subsections, we first introduce the entire structure of the MetaViewer and then elaborate on the bi-level optimization process. + +# 3.1. The entire structure + +**Embedding module** aims to transform heterogeneous views into the latent feature space, where transformed view embeddings have the same dimension as each other. To this end, we conduct a view-specific embedding function $f_{v}$ for each view, where $v = 1,2,\dots ,V$ . Given the $v$ -th view data $x_{v}$ of the entity $x$ , the corresponding embedding $z_{v}\in \mathbb{R}^{d}$ can be computed by + +$$ +z _ {v} = f _ {v} \left(x _ {v}, \theta_ {f _ {v}}\right), \tag {2} +$$ + +where $f_{v}$ is typically instantiated as a multi-layer neural network with learnable parameters $\theta_{f_v}$ . + +Representation learning module maps the obtained embedding to the view-specific or unified representation by constructing a meta-learner $m$ (i.e., MetaViewer). Note that the two types of representations are both derived from the same set of parameters, but at different learning stages (see 3.2). The view-specific representation is obtained at the inner level, where MetaViewer receives the embedding of the $v$ -th view and outputs the corresponding representation $h_v$ . In contrast, the unified representation $H$ is learned by MetaViewer from all view embeddings at the outer level and is ultimately used for downstream tasks. Therefore, the structure of MetaViewer should be flexible enough to meet the two requirements above simultaneously. Specifically, + +MetaViewer is implemented as a channel-oriented 1-d convolutional layer (C-Conv) with a non-linear function (e.g., ReLU [17]), as shown in Fig. 2 (c). + +On the one hand, at the outer level, we first concatenate all view embeddings at the channel level, i.e., the number of the channel in the concatenated embedding is equal to the number of views, and then train the MetaViewer to learn the fusion of cross-view information $H \in \mathbb{R}^{d_h}$ by + +$$ +H = m \left(z _ {c a t}, \omega\right), \tag {3} +$$ + +where $z_{cat} \in \mathbb{R}^{d \times V}$ indicates the concatenated embedding and $\omega$ is the parameter of MetaViewer. On the other hand, the inner-level initializes and trains MetaViewer to learning the representation $h_v \in \mathbb{R}^{d_h}$ of the $v$ -th view via + +$$ +h _ {v} = m _ {v} \left(z _ {v}, \omega_ {v}\right), \tag {4} +$$ + +where $m_v$ denotes the learner with the learnable parameter $\omega_v$ for handling the $v$ -th view in the inner level, also known as the base learner to avoid confusion. It is worth noting that $\omega_v$ is initialized by a subset of the MetaViewer parameters. + +Self-supervised module provides effectively supervised information for model training by constructing a series of pretext tasks1. These tasks serve as the heads of our framework. Given the input $z$ , the output $s$ can be formalized as + +$$ +s = t \left(z, \theta_ {t}\right), \tag {5} +$$ + +where $t$ and $\theta_t$ are the network and its parameters involved in pre-text tasks, respectively. Taking the reconstruction head as an example, it achieves the reconstruction object by re-mapping the representation back to the original view space. In this case, the $s := \hat{x}_v$ indicates the reconstruction result for input $x_v$ , and $t(\cdot, \theta_t) := t_v(\cdot, \theta_{t_v})$ is a multi-layer neural network with the opposite structure to the embedding module. More similar pre-text tasks have been studied in previous works [55, 57], and this is not the focus of this work. + +# 3.2. Training via bi-level optimization + +We now have the entire structure, which can be end-to-end trained following the proposed uniform-to-specific pipeline to derive the unified multi-view representation. The training flowchart and step are shown in the top row in Fig. 3, where the learning of view-private features and unified representation are formulated as a nested bi-level optimization problem. Inner-level focus on independent training on a specific view, which receives a given representation and reconstructs it to each view. Outer-level updates the meta-learner to find the optimal fusion rule through observing the training processes across all views. This pipeline decouples the unified entity + +representations from view-specific features and enables data-driven fusion. Before the detailed description, we introduce a split strategy for multi-view data in meta-learning style to adapt the bi-level training. + +Meta-split of multi-view data. Consider a batch multi-view entity samples $\{D_v^{batch}\}_{v=1}^V$ from set $D$ . We randomly and proportionally divide it into two disjoint subsets, marked as support set $S = \{S_v\}_{v=1}^V$ and query set $Q = \{Q_v\}_{v=1}^V$ , respectively. As shown in 2 (a), support set is used in inner-level for leaning view-specific information, so the sample attributes in it could be ignored. In contrast, query set retains both view and sample attributes for outer-level optimization. + +Inner-level optimization. Without loss of generality, take the inner-level update after the $o$ -th outer-level optimization as an example. Let $\omega^o$ be the lasted parameters of the MetaViewer, and $\theta_v^o = \{\theta_{f_v}^o,\theta_{t_v}^o\}$ denotes the lasted parameters in embedding and self-supervised modules for brevity. We first initial base learner from meta-learner, i.e., $\omega_v^0 = \omega^o$ , and make a copy of $\theta_v^o$ for $\tilde{\theta}_v$ . Note that the copy means gradients with respect to the $\theta_v^o$ will not be backpropagated to $\tilde{\theta}_v$ and vice versa. Thus, $\omega_v^0$ and $\tilde{\theta}_v$ form the initial states of parameters in the inner-level optimization. Suppose $\mathcal{L}_v^{in}$ is the loss function of the inner-level with respect to the $v$ -th view, the corresponding update goal is + +$$ +\omega_ {v} ^ {*} \left(\omega^ {o}\right) = \arg \min \mathcal {L} _ {v} ^ {i n} \left(\omega_ {v} \left(\omega^ {o}\right), \tilde {\phi} _ {v}; S _ {v}\right). \tag {6} +$$ + +Consider a gradient descent strategy (e.g., SGD [4]), we can further write the update process of $\omega_{v}$ : + +$$ +\omega_ {v} ^ {i} = \omega_ {v} ^ {i - 1} - \beta \frac {\partial \mathcal {L} _ {v} ^ {i n}}{\partial \omega_ {v} ^ {i - 1}}, \dots , \omega_ {v} ^ {0} = \omega^ {o}, \tag {7} +$$ + +where $\beta$ and $i$ denote the learning rate and iterative step of inner-level optimization, respectively. + +Outer-level optimization. After several inner-level updates, we obtain a set of optimal view-specific parameters for the support set. Outer-level then updates the meta-learner, embedding, and self-supervised modules by training on the query set. With the loss function $\mathcal{L}^{out}$ , the outer-level optimization goal is + +$$ +\omega^ {*}, \left\{\theta_ {v} ^ {*} \right\} _ {v = 1} ^ {V} = \arg \min \mathcal {L} ^ {\text {o u t}} \left(\left\{\omega_ {v} ^ {*} (\omega), \phi_ {v} \right\} _ {v = 1} ^ {V}; Q\right). \tag {8} +$$ + +By alternately optimizing Eq. 6 and Eq. 8, we end up with the optimal meta-parameters $\omega^{*}$ and a set of view-specific parameters $\{\theta_v^*\}_{v = 1}^V$ . Given a test sample $x_{test}$ , the corresponding unified representation is derived by feeding it sequentially into the embedding module and the meta-learner. The overall framework of MetaViewer is shown in Alg. 1. + +# 3.3. Specific-to-uniform versus uniform-to-specific + +We discuss the difference between specific-to-uniform and our uniform-to-specific pipeline through three aspects shown in Fig. 3, including the overall flowchart, update steps, + +Algorithm 1 The framework of our MetaViewer. + +Require: Training dataset $D$ , meta parameters $\omega$ , base parameters $\{\omega_v\}_{v=1}^V$ , view-specific parameters $\{\theta_v\}_{v=1}^V$ , the number of views $V$ , the iteration step in inner-level optimization $T$ . + +1: Initialize $\omega, \{\theta_v\}_{v=1}^V$ ; +2: while not done do +3: # Outer-level +4: Sample and meta-split a batch set from $D$ : +5: $\{D_v^{batch}\}_{v = 1}^V = \{\tilde{S}_v\}_{v = 1}^V +\{Q_v\}_{v = 1}^V.$ +6: for $t = 1,\dots ,T$ do +7: for $v = 1, \ldots, V$ do +8: #Inner-level +9: Initialize $\omega_{v} = \omega ,\tilde{\theta}_{v} = \theta_{v};$ +10: Optimize $\omega_v(\omega)$ and $\theta_v$ via Eq. 6. +11: end for +12: end for +13: Optimize $\omega$ and $\{\phi_v\}_{v=1}^V$ via Eq. 8. +14: end while + +and key gradients. For clarity, we here slightly confuse the notation and let $\theta_v$ and $\mathcal{L}_v^{in}$ correspond to the view-specific parameters and loss functions in specific-to-uniform pipeline, respectively. Similarly, $\omega$ and $\mathcal{L}^{out}$ denote the view-shared parameters and loss functions, respectively. Let us start by reviewing two branches in specific-to-uniform framework. + +One-level methods is a classical branch in specific-to-uniform framework, where parameters $\{\theta_v\}_{v=1}^V$ and $\omega$ are simultaneously optimized by combining the functions $\mathcal{L}_v^{in}$ and $\mathcal{L}^{out}$ . + +Multi-level methods also follow the specific-to-uniform framework, the difference is that they hierarchically learn specific and shared features to avoid interference. That is, parameters $\theta_v$ are updated at the low level by minimizing the $\mathcal{L}_v^{in}$ , and parameters $\omega$ are updated at the high level by minimizing the $\mathcal{L}^{out}$ , respectively. + +Bi-level MetaViewer further decouples view-private information, except for view-specific and view-sharing components. Consider the simplest case, where the base learner updates only one step on each view in the inner loop, i.e., $\omega_{v}^{*} = \omega -\beta \nabla_{\omega}\mathcal{L}_{v}^{in}(\omega)$ . Thus, the final optimization of $\omega$ in the outer level in Eq. 8 can be rewritten as + +$$ +\min _ {\omega} \sum_ {v} \mathcal {L} ^ {o u t} (\omega_ {v} ^ {*}) = \sum_ {v} \mathcal {L} ^ {o u t} \left(\omega - \beta \nabla_ {\omega} \mathcal {L} _ {v} ^ {i n} (\omega)\right). \quad (9) +$$ + +Intuitively, the optimal $\omega$ is expected to be close enough to each view that the view-specific representation can be produced after once updating. In other words, our uniform-to-specific framework requires explicit modeling the gap caused by view-private information when mapping from a unified representation to a particular view, which is achieved by inner-level optimization in MetaViewer. + +
our bi-level MetaViewer uniform-to-specific
Q→θv→ωv(ω)→Lvinωv(ω)→ωv*(ω)∂Lvin/∂ωv(ω)
DInner-Level
S→θv→Σv→ω→Loutθv→θv*, ω→ω*∂Lout/∂θv, ∂Lout/∂2ω
+ +
one-level methodsspecific-to-uniform
D→θv→Σv→ωLvnθv→θ*, ω→ω*∂(Lout+Lvn)/∂θv, ∂(Lout+∑v) / ∂ω
+ +
two-/multi-level methods
D→θvLinvθv→θ*Low-Level
ω→ω*High-Level
∂Lu/∂θv
+ +(a) Flowcharts + +(b) Steps + +(c) Gradients + +Figure 3. Comparison of two frameworks in terms of flowchart, update steps, and key gradients. The top row is our bi-level MetaViewer with uniform-to-specific framework, where red and black arrows indicate the flow at outer-level and inner-level, respectively. The bottom two rows show two main branches in uniform-to-specific framework, i.e., one-level methods [47] and two- or multi-level methods [51]. + +# 3.4. The instances of the objective function + +Our uniform-to-specific framework emphasizes learning from the training process in the inner loop. In unsupervised multi-view learning, this process typically refers to the reconstruction on views, thus the $\mathcal{L}_v^{in}$ is specified as the reconstruction loss [27, 60] + +$$ +\mathcal {L} _ {v} ^ {i n} = \mathcal {L} _ {v} ^ {r e c} \left(S _ {v}, S _ {v} ^ {r e c}\right) = \| S _ {v} - S _ {v} ^ {r e c} \| _ {F} ^ {2}. \tag {10} +$$ + +While parameters updated at the outer level can be constrained by richer self-supervised feedbacks, as maintained in Sec. 3.1. Here we provide two instances of the outer-level loss function to demonstrate how MetaViewer can be extended with different learning objectives. + +MVer-R adopts the same reconstruction loss as the inner-level, and $\mathcal{L}^{out} = \sum_{v}\mathcal{L}_{v}^{rec}(Q_{v},Q_{v}^{rec})$ , which is the purest implementation of MetaViewer. + +MVer-C additionally utilizes a contrastive objective, where the similarities of views belonging to the same entity (i.e., positive pairs) should be maximized and those of different entities (i.e., negative pairs) should be minimized, i.e., $\mathcal{L}^{out} = \sum_{v}\left(\mathcal{L}_{v}^{rec} + \sum_{v^{\prime},v^{\prime}\neq v}\mathcal{L}_{v,v^{\prime}}^{con}\right)$ . Following previous work [18, 51, 55], the contrastive loss $\mathcal{L}_{v,v^{\prime}}^{con}$ is formed as + +$$ +\mathcal {L} _ {v, v ^ {\prime}} ^ {c o n} = - \frac {1}{N _ {Q}} \sum_ {i = 1} ^ {N _ {Q}} \log \frac {e ^ {d \left(q _ {v} ^ {i} , q _ {v ^ {\prime}} ^ {i}\right) / \tau}}{\sum_ {j = 1 , j \neq i} ^ {N _ {Q}} e ^ {d \left(q _ {v} ^ {i} , q _ {v} ^ {j}\right) / \tau} + \sum_ {j = 1} ^ {N _ {Q}} e ^ {d \left(q _ {v} ^ {i} , q _ {v ^ {\prime}} ^ {j}\right) / \tau}}, \tag {11} +$$ + +where $q_v^i$ is the $v$ -th view of the $i$ -th query sample and $d$ is the similarity metric (e.g., cosine similarity [6]). The $N_Q$ and $\tau$ denote the number of query set samples and the + +
Datasets#Views#Classes#Samples (train, val, test)View Dimensions
Handwirtten2102000 (1200, 400, 400)240; 216
RGB-D250500 (300, 100, 100)12288; 4096
Animal25010158 (6074, 2011, 2073)4096; 4096
Fashion-MV31010000 (6000, 2000, 2000)784; 784; 784
Caltech101-206202386 (1425, 469, 492)48; 40; 254; 1984; 512; 928
+ +Table 1. The attributes for all datasets used in our experiments. + +
MethodsHandwrittenRGB-DAnimalFashion-MVCaltech101-20
ACCNMIARIACCNMIARIACCNMIARIACCNMIARIACCNMIARI
Baseline43.5552.1925.6848.3080.9704.9545.7966.2233.8750.9145.9832.0936.6348.9623.54
DCCA [2] (2013)57.2569.8052.1551.0082.9952.0260.3468.5146.1070.7080.4261.8038.6250.8822.73
DCCAE [47] (2015)63.0075.0459.2948.0081.5848.3465.1766.8248.8371.0581.1262.3436.5952.2425.25
MIB [8] (2020)63.2567.5852.1650.0081.1351.2760.8961.9852.4957.2073.8347.6235.9847.0022.18
WTNNM [15] (2020)75.6977.6261.5152.5481.9552.4671.9271.4361.0683.3386.3077.9939.7455.5826.92
TLRR [23] (2021)78.6077.9966.4052.8782.3853.3769.7973.0161.2379.5484.0178.5440.1852.3827.60
MFLVC [51] (2022)64.0064.5348.8553.0083.3154.0774.1076.0864.2883.2088.7578.9336.5958.3626.87
DCP [27] (2022)66.2570.5656.1052.0082.0452.6472.7774.0265.4862.6068.3854.3036.7944.3723.50
MVer-R (ours)75.0078.5367.2153.0082.4153.0476.4978.2565.2880.8088.1375.0541.8758.5229.19
MVer-C (ours)86.2578.9672.2557.0084.9757.0775.9278.0166.0785.4088.7680.0745.1260.8635.00
+ +Table 2. Clustering results of all methods on six datasets. Bold and underline denote the best and second-best results, respectively. + +temperature parameter, respectively. Note that, the derived meta representation can also be used in contrastive learning as an additional novel view. + +# 4. Experiments + +In this section, we present extensive experimental results to validate the quality of the unified representation derived from our MetaViewer. The remainder of the experiments are organized as follows: Subsection 4.1 lists datasets, compared methods, and implementation details. Subsection 4.2 compares the performance of our method with classical and state-of-the-art methods on two common downstream scenarios, clustering and classification tasks. Comparisons with manually designed fusion and ablation studies are shown in Subsections 4.3 and 4.4, respectively. + +# 4.1. Experimental Setup + +Datasets. To comprehensively evaluate the effectiveness of our MetaViewer, we conduct five multi-view benchmarks in experiments. All datasets are scaled to $[0,1]$ and split into training, validation, and test sets in the ratio of $6:2:2$ , as shown in Tab. 1. Handwritten contains 2,000 handwritten digital images from 0 to 9, where two types of descriptors, i.e., 240-D pixel average in $2\times 3$ windows and 216-D profile correlations, are selected as two views [60]. RGB-D dataset contains visual and depth images of 300 distinct objects across 50 categories [46, 63]. Two views are obtained by flattening the $64\times 64\times 3$ color images and $64\times 64$ depth + +images. Animal consists of 10158 images from 50 classes with two views. Unlike other datasets, these two views are features extracted by deep networks [59]. Fashion-MV is an image dataset that contains 10 categories with a total of 30,000 fashion products. It has three views and each of which consists of 10,000 gray images sampled from the same category [50]. Caltech101-20 is a subset of the Caltech101 image set [9], which consists of 2,386 images of 20 subjects. Six features are used, including Gabor, Wavelet Moments, CENTRIST, HOG, GIST, and LBP. + +Compared methods. We compare the performance of MetaViewer with seven representative multi-view learning methods, including two classical methods (DCCA [2] and DCCAE [47]) and five state-of-the-art methods (WTNNM [15], TLRR [23], MIB [8], MFLVC [51] and DCP [27]). Among them, DCCA and DCCAE are the deep extensions of traditional correlation strategies. WTNNM and TLRR are representative tensor-based methods. MIB is a typical generative method with mutual information constraints. DCP learns unified representation in both unsupervised and supervised scenarios, and we report the unsupervised version for a fair comparison. In particular, MFLVC also notices the view-private redundant information and designs a multilevel feature network for clustering tasks. In addition, we also compare the model without MetaViewer as a baseline, where the unified representation is obtained by concatenating view-specific features output by the embedding module. + +Implementation details. For a fair comparison, all methods are trained from scratch, and all deep-network-based + +
MethodsHandwrittenRGB-DAnimalFashion-MVCaltech101-20
ACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-score
Baseline87.0087.3287.0414.0006.4507.7071.7865.0263.1087.5087.6387.5475.4151.9546.21
DCCA [2] (2013)88.2589.2088.0530.0021.1022.0462.8461.6160.6584.9085.2283.5471.5445.2739.81
DCCAE [47] (2015)90.0090.4889.9224.0016.0016.9166.8262.5362.8385.3585.9783.8471.5460.5743.25
MIB [8] (2020)79.0083.9078.5233.0028.5027.3759.3263.7862.1386.8086.8086.5572.7261.6452.47
WTNNM [15] (2020)96.7796.2096.3645.0047.1843.9069.9067.4065.4894.5094.5094.5883.2780.4974.78
TLRR [23] (2021)97.0097.0597.1749.0053.3247.6671.9069.5268.3596.3596.2896.4085.5577.3075.24
MFLVC [51] (2022)94.0094.2094.0144.0046.0941.8175.6272.1770.8196.5096.5296.4985.3771.8369.07
DCP [27] (2022)97.2597.3097.2437.0028.8730.7877.9573.4370.0189.2582.0682.9092.4889.4184.58
MVer-R (ours)97.0097.0897.0051.0053.6548.7377.6973.9171.2296.8596.3796.4892.2889.4684.21
MVer-C (ours)97.7597.9097.7556.0055.2052.7878.0374.5671.5597.7096.7897.0792.1690.6885.72
+ +methods share the same embedding backbone, following the previous work [51]. For two-view-based methods, such as DCCA, DCCAE, and MIB, we report the results of the two best-performing views. The final unified representation $H$ in alignment-based methods is a concatenation of the processed features, and the representation dimension is specified as $d_h = 256$ . The results on clustering and classification tasks are obtained by subsequent K-means and SVM classifier, respectively. For MetaViewer, we train 2,000 epochs for all benchmarks, and set the batch size is 32 for RGBD and 128 for others. The learning rates in outer- and inner-level are set to $10^{-4}$ and $10^{-3}$ , respectively. We use a single convolutional layer with 32 kernels of size 5 as the meta-learner, and we set the proportion of support set and the update step in the inner level to 0.2 and 2, respectively. All experiments have been verified using the PyTorch library on a single RTX3090. Code is available at https://github.com/xxLifeLover/MetaViewer. + +# 4.2. Performance on downstream tasks + +Clustering results. Tab. 2 lists the results of the clustering task, where the performance is measured by three standard evaluation metrics, i.e., Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI). A higher value of these metrics indicates better clustering performance. It can be observed that (1) our MVer-C variant significantly outperforms other compared methods on all benchmarks; (2) the second-best results appear between MVer-R and MFLVC, both of which explicitly separate the view-private information; (3) a larger number of categories and views is the main reason for the degradation of clustering performance, and our Metaviewer improves most significantly in such a scenario (e.g., Fashion-MV and Caltech101-20). + +Classification results. Tab. 3 lists the results of the classification task, where three common metrics are used, including Accuracy, Precision, and F-score. A higher value indicates better classification performance. Similar to the clustering results, two variants of MetaViewer significantly + +Table 3. Classification results of all methods on six datasets. Bold and underline denote the best and second-best results, respectively. + +
StrategiesRulesACC↑NMI↑ARI↑MSE↓
Sumzx + zy69.2571.8959.021.84
Maxmax(zx, zy)80.7573.9363.75-
Concat.cat[zx, zy]78.7572.0261.521.77
Linearl(zx, zy, θl)85.0077.4069.714.74
C-Convm(zx, zy, ω)69.7565.2151.332.37
MetaViewermeta-learning86.2578.9672.252.45
+ +Table 4. Clustering resulting on the Handwritten dataset. + +![](images/6baaf3e6a29beef395c5ab07655a3d9c308950faeffa017db965f0ecdf1e1b55.jpg) +Figure 4. Class activate maps on RGB-D samples for different fusion strategies, including single views, Sum, and our MetaViewer. View 1 and view 2 are the RGB and depth view, respectively. + +outperform the comparison methods. It is worth noting that (1) DCP learns unified representation and therefore achieves the second-best result instead of MFLVC. (2) The number of categories is the main factor affecting the classification performance, and our method obtains the most significant improvement in the RGB-D dataset with up to 50 classes. + +# 4.3. Comparison with manually designed fusion. + +As mentioned in 3.3, MetaViewer essentially learns an optimal fusion rule that preserves as much consistency and complementary information as possible while filtering out view-private information. To verify this, we compare it with commonly used fusion strategies [26], including sum, maxima, concatenation, linear layer and $C$ -Conv. The former three are the specified fusion rules without trainable param- + +![](images/5f48bdb6a99fc4b704fcbbf5744ca4dddb96009c29daa933fb69f0bc450bae03.jpg) +(a) Effect of network width + +![](images/e8e719885cd11547462682ff6d9cbac33b9b3b85e7688e22ad2c2f240f29672d.jpg) +(b) Effect of kernel size + +![](images/3a6e0a0af9705b163b1b6c0f5014a349a1123fbfe531cbf97ecfb7ee29a7720a.jpg) +Figure 5. Effect of meta-learner architectures with different depth, width, and kernel size on classification accuracy. +(a) Proportion of support set +Figure 6. Effect of (a) different meta-division ratios and (b) the number of inner-loop iterations on classification accuracy. + +![](images/0c60e460c8994000eddf195e84b26d7a58a01f158a284f0a03f1a4466edc0fc4.jpg) +(b) Steps of inner-level update + +eters, and the remaining two are the trainable fusion layer trained via the specific-to-uniform manner. Tab. 4 lists the clustering results and an additional MSE score on the Hand-written dataset with the same embedding and reconstruction network. We can observe that (1) trainable fusion layers outperform the hand-designed rules, and our MetaViewer yields the best performance; (2) the MSE scores listed in the last column indicate that the quality of the unified representation cannot be measured and guaranteed only with the reconstruction constraint, due to the view-private redundant information mixed in view-specific latent features. + +To further explore the fusion preferences learned by MetaViewer, we visualize the class activation map for the RGB-D samples in Fig. 4. MetaViewer balances information from all views instead of just the salient one (view 1), ensuring a high-quality unified multi-view representation. + +# 4.4. Ablation Studies + +Meta-learner structures. We implement the meta-learner as a channel-level convolution structure in this work. Albeit simple, this layer can be considered as a universal approximator for almost any continuous function [40], and thus can fit a wide range of conventional fusion functions. To investigate the effect of network depth, width, and convolution kernel size on the performance of the representation, we alternately fix the 32 kernels and the $1 \times 3$ kernel size and show the classification results on Handwritten data in Fig. 5. It is clear that (1) the meta-learner works well with just a + +shallow structure, as shown in Fig. 5 (a), instead of gradually overfitting to the training data as the network deepens or widens, and (2) our MetaViewer is stable and insensitive to the hyper-parameters within reasonable ranges. + +Meta-split ratios. Fig. 6 (a) shows the impact of the meta-split mentioned in Sec. 3.2 on the classification performance, where the proportion of support set is set from 0.1 to 0.9 in steps of 0.1, and the rest is query set. In addition to the single view, we also compare the sum and concat. fusion as baselines. MetaViewer consistently surpasses all baselines over the experimental proportion. In addition, fusion baselines are more dependent on the better-performing view at lower proportions, instead becoming unstable as the available query sample decreases. + +Inner-level update steps. Another hyper-parameter is the number of iteration steps in inner-level optimization. More iterations mean a larger gap from the learned meta representation to the specific view space, i.e., coarser modeling of view-private information. Fig. 6 (b) shows the classification results with various steps, where $n$ steps mean that the inner-level optimization is updated $n$ times throughout the training. MetaViewer achieves the best results when using 1 steps, and remains stable within 15 steps. + +# 5. Conclusion + +This work introduced a novel meta-learning perspective for multi-view learning, and proposed a meta-learner, namely MetaViewer, to derive a high-quality unified representation for downstream tasks. In contrast to the prevailing specific-to-uniform pipeline, MetaViewer observes the reconstruction process from unified representation to specific views and essentially learns an optimal data-driven fusion that separates and filters out meaningless view-private information. The proposed framework is compatible with most excellent objective functions and pre-text tasks to cope with complex real-world scenarios. Extensive experimental results on clustering and classification tasks demonstrate the performance of meta-learned unified representation. In addition, we believe that our bi-level training pipeline is also promising for handling incomplete views, and leave it to future work. + +# Acknowledgements + +This research was supported in part by the Natural Science Foundation of China (No. 62106129, 62176139, and 62177031), the Natural Science Foundation of Shandong Province (No. ZR2021QF053, ZR2021ZD15), and the China Postdoctoral Science Foundation (No. 2021TQ0195, 2021M701984). The author acknowledges Prof. Xiushan Nie of Shandong Jianzhu University (China) for the financial support and valuable discussions. + +# References + +[1] Sara Atito Ali Ahmed, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. CoRR, abs/2104.03602, 2021. 4 +[2] Galen Andrew, Raman Arora, Jeff A. Bilmes, and Karen Livescu. Deep canonical correlation analysis. In ICML, volume 28, pages 1247-1255, 2013. 2, 6, 7 +[3] Philip Bachman, R. Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In NeurIPS, pages 15509-15519, 2019. 2 +[4] Léon Bottou. Large-scale machine learning with stochastic gradient descent. In COMPSTAT, pages 177-186, 2010. 4 +[5] Abhra Chaudhuri, Massimiliano Mancini, Yanbei Chen, Zeynep Akata, and Anjan Dutta. Cross-modal fusion distillation for fine-grained sketch-based image retrieval. In BMVC, page 499, 2022. 2 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119, pages 1597-1607, 2020. 5 +[7] Yanhua Cheng, Xin Zhao, Rui Cai, Zhiwei Li, Kaiqi Huang, and Yong Rui. Semi-supervised multimodal deep learning for RGB-D object recognition. In IJCAI, pages 3345-3351, 2016. 2 +[8] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushner, and Zeynep Akata. Learning robust representations via multiview information bottleneck. In ICLR, 2020. 1, 6, 7 +[9] Li Fei-Fei, Robert Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Comput. Vis. Image Underst., 106(1):59-70, 2007. 6 +[10] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, pages 1933-1941, 2016. 2 +[11] Fangxiang Feng, Xiaojie Wang, and Ruifan Li. Cross-modal retrieval with correspondence autoencoder. In ACM MM, pages 7-16, 2014. 2 +[12] Yifan Feng, Zizhao Zhang, Xibin Zhao, Rongrong Ji, and Yue Gao. GVCNN: group-view convolutional neural networks for 3d shape recognition. In CVPR, pages 264-272, 2018. 2 +[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, volume 70, pages 1126–1135, 2017. 3 +[14] Kuiliang Gao, Bing Liu, Xuchu Yu, and Anzhu Yu. Unsupervised meta learning with multiview constraints for hyperspectral image small sample set classification. IEEE TIP, 31:3449-3462, 2022. 3 +[15] Quanxue Gao, Wei Xia, Zhizhen Wan, De-Yan Xie, and Pu Zhang. Tensor-svd based graph learning for multi-view subspace clustering. In AAAI, pages 3930-3937, 2020. 6, 7 +[16] Yu Geng, Zongbo Han, Changqing Zhang, and Qinghua Hu. Uncertainty-aware multi-view representation learning. In AAAI, pages 7545-7553, 2021. 1 +[17] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In AISTATS, volume 15, pages 315-323, 2011. 4 + +[18] Kaveh Hassani and Amir Hosein Khas Ahmadi. Contrastive multi-view representation learning on graphs. In ICML, volume 119, pages 4116-4126, 2020. 2, 5 +[19] Timothy M. Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE TPAMI, 44(9):5149-5169, 2022. 2 +[20] Junlin Hu, Jiwen Lu, and Yap-Peng Tan. Sharable and individual multi-view metric learning. IEEE TPAMI, 40(9):2281-2288, 2018. 2 +[21] Mike Huisman, Jan N. van Rijn, and Aske Plaat. A survey of deep meta-learning. Artif. Intell. Rev., 54(6):4483-4541, 2021. 2 +[22] Xiaodong Jia, Xiao-Yuan Jing, Xiaoke Zhu, Songcan Chen, Bo Du, Ziyun Cai, Zhenyu He, and Dong Yue. Semi-supervised multi-view deep discriminant representation learning. IEEE TPAMI, 43(7):2496-2509, 2021. 2 +[23] Yuheng Jia, Hui Liu, Junhui Hou, Sam Kwong, and Qingfu Zhang. Multi-view spectral clustering tailored tensor low-rank representation. IEEE TCSVT, 31(12):4784-4797, 2021. 6, 7 +[24] Xiao-Yuan Jing, Fei Wu, Xiwei Dong, Shiguang Shan, and Songcan Chen. Semi-supervised multi-view correlation feature learning with application to webpage classification. In AAAI, pages 1374-1381, 2017. 2 +[25] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. IEEE TPAMI, 39(4):664-676, 2017. 2 +[26] Yingming Li, Ming Yang, and Zhongfei Zhang. A survey of multi-view representation learning. IEEE TKDE, 31(10):1863-1883, 2019. 2, 7 +[27] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE TPAMI, 2022. 1, 2, 5, 6, 7 +[28] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In ICCV, pages 3295-3304, 2019. 3 +[29] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In CVPR, pages 3623–3632, 2019. 2 +[30] Shirui Luo, Changqing Zhang, Wei Zhang, and Xiaochun Cao. Consistent and specific multi-view subspace clustering. In AAAI, pages 3730-3737, 2018. 2 +[31] Yao Ma, Shilin Zhao, Weixiao Wang, Yaoman Li, and Irwin King. Multimodality in meta-learning: A comprehensive survey. KBS, 250:108976, 2022. 3 +[32] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In CVPR, pages 9826–9836, 2021. 2 +[33] Feiping Nie, Guohao Cai, Jing Li, and Xuelong Li. Autoweighted multi-view learning for image clustering and semi-supervised classification. IEEE TIP, 27(3):1501-1511, 2018. 2 + +[34] Xiushan Nie, Weizhen Jing, Chaoran Cui, Chen Jason Zhang, Lei Zhu, and Yilong Yin. Joint multi-view hashing for large-scale near-duplicate video retrieval. IEEE TKDE, 32(10):1951-1965, 2020. 2 +[35] Xiushan Nie, Xingbo Liu, Jie Guo, Letian Wang, and Yilong Yin. Supervised discrete multiple-length hashing for image retrieval. IEEE TBD, 9(1):312-327, 2023. 1 +[36] Zheyun Qin, Xiankai Lu, Xiushan Nie, Dongfang Liu, Yi-long Yin, and Wenguan Wang. Coarse-to-fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica, 10(5):1-17, 2023. 2 +[37] Mathieu Salzmann, Carl Henrik Ek, Raquel Urtasun, and Trevor Darrell. Factorized orthogonal latent spaces. In AISTATS, volume 9, pages 701-708, 2010. 2 +[38] Yang Shi, Xiushan Nie, Xingbo Liu, Lu Yang, and Yilong Yin. Zero-shot hashing via asymmetric ratio similarity matrix. IEEE TKDE, 2022. 2 +[39] Yang Shi, Xiushan Nie, Xingbo Liu, Li Zou, and Yilong Yin. Supervised adaptive similarity matrix hashing. IEEE TIP, 31:2755-2766, 2022. 2 +[40] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pages 1917-1928, 2019. 3, 8 +[41] Haoliang Sun, Chenhui Guo, Qi Wei, Zhongyi Han, and Yilong Yin. Learning to rectify for robust learning with noisy labels. PR, 124:108467, 2022. 3 +[42] Kai Sun, Jiangshe Zhang, Junmin Liu, Ruixuan Yu, and Zengjie Song. DRCNN: dynamic routing convolutional neural network for multi-view 3d object recognition. IEEE TIP, 30:868-877, 2021. 2 +[43] Shiliang Sun, Wenbo Dong, and Qiuyang Liu. Multi-view representation learning with deep gaussian processes. IEEE TPAMI, 43(12):4453-4468, 2021. 1 +[44] Zhongkai Sun, Prathusha Kameswara Sarma, William A. Sethares, and Yingyu Liang. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In AAAI, pages 8992-8999, 2020. 2 +[45] Risto Vuorio, Shao-Hua Sun, Hexiang Hu, and Joseph J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. In NeurIPS, pages 1–12, 2019. 3 +[46] Shiye Wang, Changsheng Li, Yanming Li, Ye Yuan, and Guoren Wang. Self-supervised information bottleneck for deep multi-view subspace clustering. CoRR, abs/2204.12496, 2022. 6 +[47] Weiran Wang, Raman Arora, Karen Livescu, and Jeff A. Bilmes. On deep multi-view representation learning. In ICML, volume 37, pages 1083-1092, 2015. 2, 5, 6, 7 +[48] Qi Wei, Lei Feng, Haoliang Sun, Ren Wang, Chenhui Guo, and Yilong Yin. Fine-grained classification with noisy labels. arXiv preprint arXiv:2303.02404, 2023. 3 +[49] Qi Wei, Haoliang Sun, Xiankai Lu, and Yilong Yin. Self-filtering: A noise-aware sample selection for label noise with confidence penalization. In ECCV, pages 516-532, 2022. 2 + +[50] Jie Xu, Yazhou Ren, Huayi Tang, Zhimeng Yang, Lili Pan, Yang Yang, and Xiaorong Pu. Self-supervised discriminative feature learning for multi-view clustering. CoRR, abs/2103.15069, 2021. 6 +[51] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In CVPR, pages 16030-16039, 2022. 1, 2, 5, 6, 7 +[52] Fei Xue, Xin Wu, Shaojun Cai, and Junqiu Wang. Learning multi-view camera relocalization with graph neural networks. In CVPR, pages 11372-11381, 2020. 2 +[53] Xiaowei Xue, Feiping Nie, Sen Wang, Xiaojun Chang, Bela Stantic, and Min Yao. Multi-view correlated feature learning by uncovering shared component. In AAAI, pages 2810-2816, 2017. 2 +[54] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 2 +[55] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, pages 8824-8833, 2022. 2, 4, 5 +[56] Xiao Yu, Hui Liu, Yuxiu Lin, Yan Wu, and Caiming Zhang. Auto-weighted sample-level fusion with anchors for incomplete multi-view clustering. PR, 130:108772, 2022. 2 +[57] Yun-Hao Yuan, Jin Li, Yun Li, Jipeng Qiang, Yi Zhu, Xiaobo Shen, and Jianping Gou. Learning canonical f-correlation projection for compact multiview representation. In CVPR, pages 19238-19247, 2022. 2, 4 +[58] Changqing Zhang, Huazhu Fu, Qinghua Hu, Xiaochun Cao, Yuan Xie, Dacheng Tao, and Dong Xu. Generalized latent multi-view subspace clustering. IEEE TPAMI, 42(1):86-99, 2020. 1 +[59] Changqing Zhang, Zongbo Han, Yajie Cui, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Cpm-nets: Cross partial multi-view networks. In NeurIPS, pages 557-567, 2019. 6 +[60] Changqing Zhang, Yeqing Liu, and Huazhu Fu. Ae2-nets: Autoencoder in autoencoder networks. In CVPR, pages 2577-2585, 2019. 2, 5, 6 +[61] Liang Zhao, Tao Yang, Jie Zhang, Zhikui Chen, Yi Yang, and Z. Jane Wang. Co-learning non-negative correlated and uncorrelated features for multi-view data. IEEE TNNLS, 32(4):1486-1496, 2021. 2 +[62] Wenqing Zheng, Tianlong Chen, Ting-Kuei Hu, and Zhangyang Wang. Symbolic learning to optimize: Towards interpretability and scalability. In ICLR, 2022. 3 +[63] Pengfei Zhu, Binyuan Hui, Changqing Zhang, Dawei Du, Longyin Wen, and Qinghua Hu. Multi-view deep subspace clustering networks. CoRR, abs/1908.01978, 2019. 6 \ No newline at end of file diff --git a/2023/MetaViewer_ Towards a Unified Multi-View Representation/images.zip b/2023/MetaViewer_ Towards a Unified Multi-View Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..029275e8b0d63eb53c42407c4ee55fb0ea743772 --- /dev/null +++ b/2023/MetaViewer_ Towards a Unified Multi-View Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fa1228f207f707e8d36836e235569cd728529eecc2d354cf09301087bb1a4b8 +size 636592 diff --git a/2023/MetaViewer_ Towards a Unified Multi-View Representation/layout.json b/2023/MetaViewer_ Towards a Unified Multi-View Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..51901b575efdeffb8740c17caad8df2cc0e8af4b --- /dev/null +++ b/2023/MetaViewer_ Towards a Unified Multi-View Representation/layout.json @@ -0,0 +1,10822 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "type": "text", + "content": "MetaViewer: Towards A Unified Multi-View Representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 144, + 196, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 144, + 196, + 171 + ], + "spans": [ + { + "bbox": [ + 90, + 144, + 196, + 171 + ], + "type": "text", + "content": "Ren Wang \nShandong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 174, + 201, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 174, + 201, + 185 + ], + "spans": [ + { + "bbox": [ + 86, + 174, + 201, + 185 + ], + "type": "text", + "content": "xxlifelover@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 144, + 335, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 144, + 335, + 171 + ], + "spans": [ + { + "bbox": [ + 230, + 144, + 335, + 171 + ], + "type": "text", + "content": "Haoliang Sun* Shandong University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 174, + 339, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 174, + 339, + 185 + ], + "spans": [ + { + "bbox": [ + 228, + 174, + 339, + 185 + ], + "type": "text", + "content": "haolsun.cn@gmail.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 364, + 144, + 508, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 144, + 508, + 172 + ], + "spans": [ + { + "bbox": [ + 364, + 144, + 508, + 172 + ], + "type": "text", + "content": "Yuling Ma \nShandong Jianzhu University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 174, + 499, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 174, + 499, + 185 + ], + "spans": [ + { + "bbox": [ + 373, + 174, + 499, + 185 + ], + "type": "text", + "content": "mayuling20@sdjzu.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 193, + 205, + 259, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 205, + 259, + 219 + ], + "spans": [ + { + "bbox": [ + 193, + 205, + 259, + 219 + ], + "type": "text", + "content": "Xiaoming Xi" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 155, + 220, + 297, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 220, + 297, + 233 + ], + "spans": [ + { + "bbox": [ + 155, + 220, + 297, + 233 + ], + "type": "text", + "content": "Shandong Jianzhu University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 187, + 236, + 265, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 236, + 265, + 247 + ], + "spans": [ + { + "bbox": [ + 187, + 236, + 265, + 247 + ], + "type": "text", + "content": "fyzq10@126.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 205, + 439, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 205, + 439, + 234 + ], + "spans": [ + { + "bbox": [ + 335, + 205, + 439, + 234 + ], + "type": "text", + "content": "Yilong Yin* Shandong University" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 342, + 236, + 430, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 236, + 430, + 247 + ], + "spans": [ + { + "bbox": [ + 342, + 236, + 430, + 247 + ], + "type": "text", + "content": "ylyin@sdu.edu.cn" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 143, + 274, + 192, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 274, + 192, + 287 + ], + "spans": [ + { + "bbox": [ + 143, + 274, + 192, + 287 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 45, + 300, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 300, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 45, + 300, + 290, + 540 + ], + "type": "text", + "content": "Existing multi-view representation learning methods typically follow a specific-to-uniform pipeline, extracting latent features from each view and then fusing or aligning them to obtain the unified object representation. However, the manually pre-specified fusion functions and aligning criteria could potentially degrade the quality of the derived representation. To overcome them, we propose a novel uniform-to-specific multi-view learning framework from a meta-learning perspective, where the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. Specifically, we formulated the extraction and fusion of view-specific latent features as a nested optimization problem and solved it by using a bi-level optimization scheme. In this way, MetaViewer automatically fuses view-specific features into a unified one and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views. Extensive experimental results in downstream classification and clustering tasks demonstrate the efficiency and effectiveness of the proposed method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 288, + 693 + ], + "type": "text", + "content": "Multi-view representation learning aims to learn a unified representation of the entity from its multiple observable views for the benefit of downstream tasks [35, 43, 58]. Each view acquired by different sensors or sources contains both view-shared consistency information and view-specific information [8]. The view-specific part further consists of complementary and redundant components, where the former can be considered as a supplement to the consistency information, while the latter is view-private and may be" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 323, + 276, + 422, + 360 + ], + "blocks": [ + { + "bbox": [ + 323, + 276, + 422, + 360 + ], + "lines": [ + { + "bbox": [ + 323, + 276, + 422, + 360 + ], + "spans": [ + { + "bbox": [ + 323, + 276, + 422, + 360 + ], + "type": "image", + "image_path": "3523d0ccd48cabdf9ce9d9e848b184eeb4c48aef543a7882449f9307d0b67dd5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 363, + 413, + 374 + ], + "lines": [ + { + "bbox": [ + 333, + 363, + 413, + 374 + ], + "spans": [ + { + "bbox": [ + 333, + 363, + 413, + 374 + ], + "type": "text", + "content": "(a) Joint Representation" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 432, + 277, + 526, + 360 + ], + "blocks": [ + { + "bbox": [ + 432, + 277, + 526, + 360 + ], + "lines": [ + { + "bbox": [ + 432, + 277, + 526, + 360 + ], + "spans": [ + { + "bbox": [ + 432, + 277, + 526, + 360 + ], + "type": "image", + "image_path": "b5e2d22c9a68dfb1b4b4f9806e0e57c8341dbd12976b32c9cd65f98f17e8a774.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 363, + 528, + 374 + ], + "lines": [ + { + "bbox": [ + 429, + 363, + 528, + 374 + ], + "spans": [ + { + "bbox": [ + 429, + 363, + 528, + 374 + ], + "type": "text", + "content": "(b) Alignment Representation" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 328, + 381, + 419, + 466 + ], + "blocks": [ + { + "bbox": [ + 328, + 381, + 419, + 466 + ], + "lines": [ + { + "bbox": [ + 328, + 381, + 419, + 466 + ], + "spans": [ + { + "bbox": [ + 328, + 381, + 419, + 466 + ], + "type": "image", + "image_path": "624596e2010e1985f387e1e8121a354d9d4bd77a4e31f525a9c56a9b2a812a6d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 469, + 414, + 479 + ], + "lines": [ + { + "bbox": [ + 334, + 469, + 414, + 479 + ], + "spans": [ + { + "bbox": [ + 334, + 469, + 414, + 479 + ], + "type": "text", + "content": "(c) S&S Representation" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 429, + 381, + 529, + 466 + ], + "blocks": [ + { + "bbox": [ + 429, + 381, + 529, + 466 + ], + "lines": [ + { + "bbox": [ + 429, + 381, + 529, + 466 + ], + "spans": [ + { + "bbox": [ + 429, + 381, + 529, + 466 + ], + "type": "image", + "image_path": "0e233cccfa9bdf64eb6bf3e9200f085f56138aa8241f42b95706b8700637dd40.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 469, + 518, + 479 + ], + "lines": [ + { + "bbox": [ + 436, + 469, + 518, + 479 + ], + "spans": [ + { + "bbox": [ + 436, + 469, + 518, + 479 + ], + "type": "text", + "content": "(d) Meta Representation" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 490, + 548, + 557 + ], + "lines": [ + { + "bbox": [ + 304, + 490, + 548, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 548, + 557 + ], + "type": "text", + "content": "Figure 1. (a), (b) and (c) show three multi-view learning frameworks following the specific-to-uniform pipeline, where the unified representation is obtained by fusing or concatenating view-specific features. (d) illustrates our uniform-to-specific manner, where a meta-learner learns to fusion by observing reconstruction from unified representation to specific views." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 576, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 547, + 624 + ], + "type": "text", + "content": "adverse for the unified representation [16]. Therefore, a high-quality representation is required to retain the consistency and complementary information, as well as filter out the view-private redundant ones [51]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 624, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 547, + 661 + ], + "type": "text", + "content": "Given the data containing two views, " + }, + { + "bbox": [ + 304, + 624, + 547, + 661 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 304, + 624, + 547, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 624, + 547, + 661 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 304, + 624, + 547, + 661 + ], + "type": "text", + "content": ", prevailing multi-view learning methods typically follow a specific-to-uniform pipeline and can be roughly characterized as:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 363, + 672, + 546, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 672, + 546, + 685 + ], + "spans": [ + { + "bbox": [ + 363, + 672, + 546, + 685 + ], + "type": "interline_equation", + "content": "H := f \\left(x _ {1}; W _ {f}\\right) \\circ g \\left(x _ {2}; W _ {g}\\right), \\tag {1}", + "image_path": "54d5630db345ecf935b5c455241c3d18ee39440835a8fc71c593a7ab9fe86dfb.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": " are encoding (or embedding [27]) functions that map the original view data into the corresponding latent" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11590" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "features with the trainable parameters " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "W_{f}" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "W_{g}" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": ". These latent features are subsequently aggregated into the unified representation " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " using the designed aggregation operator o. With different aggregation strategies, existing approaches can be further subdivided into the joint, alignment, and a combined share-specific (S&S) representation [22, 26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 290, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 290, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 290, + 315 + ], + "type": "text", + "content": "Fig. 1 (a) - (c) show the above three branches of the specific-to-uniform framework. Joint representation focuses on the integration of complementary information by directly fusing latent features, where the " + }, + { + "bbox": [ + 46, + 147, + 290, + 315 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 147, + 290, + 315 + ], + "type": "text", + "content": " is represented as fusion strategies, such as graph-based modules [33], neural networks [10], or other elaborate functions [7, 34]. While alignment representation seeks alignment between view-specific features to retain consistency information and specifies " + }, + { + "bbox": [ + 46, + 147, + 290, + 315 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 147, + 290, + 315 + ], + "type": "text", + "content": " as the alignment operator, measured by distance [11], correlation [24, 47], or similarity [25, 38, 39]. The aligned features can be concatenated as the unified representation for downstream tasks. As a trade-off strategy, S&S representation explicitly distinguishes latent features into shared and specific representations and only aligns the shared part [20, 30, 53]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 290, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 290, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 290, + 508 + ], + "type": "text", + "content": "Despite demonstrating promising results, the specific-to-uniform framework inherently suffers from potential risks in the following two aspects: (1) Compared with the data-oriented fusion rules, manually pre-specified rules are designed to compute the unified representation by concatenating or fusing latent features, restricting its extensibility in complex real-world applications. (2) Even if we can find a well-performing fusion scheme, the sequential training manner limits the model to constrain the various component information separately. For example, it is difficult to automatically separate out view-private redundant information from the feature level [22, 37]. Recent researches have attempted to address the second issue by decomposing latent features using matrix factorization [61] or hierarchical feature modeling [51], but still cannot avoid manually pre-specified fusion or even decomposition strategies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": "In this work, we provide a new meta-learning perspective for multi-view representation learning and propose a novel uniform-to-specific framework to address these potential risks. Fig. 1 (d) shows the overall schematic. In contrast to the specific-to-uniform pipeline, the unified representation no longer involves manual manipulation but is automatically derived from a meta-learner named MetaViewer. To train the MetaViewer, we first decouple the learning of view-specific latent features and unified meta representation. These two learning processes can then be formulated as a nested optimization problem and eventually solved by a bi-level optimization scheme. In detail, MetaViewer fuses view-specific features into a unified one at the outer level and learns the optimal fusion scheme by observing reconstruction processes from the unified to the specific over all views at the inner level. In addition, our uniform-to-specific framework is compatible with most existing objective functions and pre-text" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "tasks to cope with complex real-world scenarios. Extensive experiments validate that MetaViewer achieves comparable performance to the state-of-the-art methods in downstream clustering and classification tasks. The core contributions of this work are as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 139, + 547, + 297 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 139, + 547, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 547, + 200 + ], + "type": "text", + "content": "1. We provide a new meta-learning perspective and develop a novel uniform-to-specific framework to learn a unified multi-view representation. To the best of our knowledge, it could be the first meta-learning-based work in multi-view representation learning community." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 547, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 547, + 255 + ], + "type": "text", + "content": "2. We propose MetaViewer, a meta-learner that formulates the modeling of view-specific features and unified representation as a nested bi-level optimization and ultimately meta-learns a data-driven optimal fusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 260, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 547, + 297 + ], + "type": "text", + "content": "3. Extensive experiments on multiple benchmarks validate that our MetaViewer achieves comparable performance to the existing methods in two downstream tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 304, + 392, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 392, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 392, + 316 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 323, + 493, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 493, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 493, + 337 + ], + "type": "text", + "content": "2.1. Multi-view representation learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 484 + ], + "type": "text", + "content": "Representation learning is not a new topic and plays a critical role in numerous downstream tasks [29, 36, 49]. This work focuses on multi-view representation in unsupervised deep learning scope [56]. Related studies can be roughly summarized into two categories [54]. One is the deep extension of traditional methods, where representative ones include deep canonical correlation analysis (DCCA) and its variants [44, 57]. DCCA [2] intends to discover the nonlinear mapping for two views to a common space in which their correlations are maximally preserved. These methods benefit from a sound theoretical foundation, but also usually have strict restrictions on the number and form of views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 548, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 548, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 548, + 654 + ], + "type": "text", + "content": "Another alternative is deep multi-view learning [5, 32]. Early deep-based approaches attempted to design effective architectures for multiple views, such as CNN-based [12, 42] and GAN-based models [18, 52]. Recent works resort to leveraging mutual or comparative information to derive the constraint for parameters [3, 27, 55]. Most of them follow the pipeline from view-specific features to unified representation. In contrast, our MetaViewer learns by observing the learning from the uniform to the specific. The most related work is " + }, + { + "bbox": [ + 304, + 486, + 548, + 654 + ], + "type": "inline_equation", + "content": "\\mathrm{AE^2}" + }, + { + "bbox": [ + 304, + 486, + 548, + 654 + ], + "type": "text", + "content": "-Nets [60], which treats the unified representation as a set of trainable parameters that are optimized during the degradation process of each view. The essential difference is that we learn the data-oriented fusion strategy rather than optimizing the unified representation itself." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 659, + 397, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 397, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 397, + 672 + ], + "type": "text", + "content": "2.2. Meta-learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "type": "text", + "content": "Optimization-based meta-learning is a classic application of bi-level optimization designed to learn task-level knowledge to quickly handle new tasks [19, 21]. A typical work," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11591" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 73, + 541, + 270 + ], + "blocks": [ + { + "bbox": [ + 56, + 73, + 541, + 270 + ], + "lines": [ + { + "bbox": [ + 56, + 73, + 541, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 541, + 270 + ], + "type": "image", + "image_path": "a9c05b632a0d125a190a8ccae793b904072902733fb371f9901dc903f56baa0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 546, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 546, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 546, + 326 + ], + "type": "text", + "content": "Figure 2. The overall framework of MetaViewer, contains (a) the meta-split of multi-view data and three modules: (b) embedding module, (c) representation module, and (d) self-supervised module. These modules are trained with a bi-level optimization. Inner-level (dark gray arrows) learns the view-specific reconstruction on the support set, and outer-level (red arrows) updates the entire model to learn the fusion scheme and the unified representation " + }, + { + "bbox": [ + 46, + 281, + 546, + 326 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 281, + 546, + 326 + ], + "type": "text", + "content": " for downstream tasks by validating on the query set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 346, + 290, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 290, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 290, + 528 + ], + "type": "text", + "content": "MAML [13], learns a set of initialization parameters to solve different tasks with a few steps of updates. Similar meta paradigm has been used to learn other manually designed parts, such as the network structure [28], optimizer [62], and even sample weight [40, 41, 48]. Similarly, we try to metalearn the fusion rule of multi-view features for a unified representation. There also exist some works that consider both meta-learning paradigms and multi-view data [14, 31, 45]. However, they are dedicated to exploiting the rich information contained in multiple views to improve the performance of the meta-learner in a few-shot or self-supervised scenario. Instead, we train a meta-learner to derive high-quality shared representations from multi-view data. To the best of our knowledge, this could be the first work to learn multi-view representation with a meta-learning paradigm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 537, + 126, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 126, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 126, + 550 + ], + "type": "text", + "content": "3. MetaViewer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": "Given a set of unlabeled multi-view data, " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "D = \\{x^{i} \\in \\mathbb{R}^{d_{x}}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": " and superscript " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": " are the number and index of entity samples, respectively. Each sample " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "x^{i} = \\{x_{1}^{i}, x_{2}^{i}, \\ldots, x_{v}^{i}\\}_{v=1}^{V}" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": " views, where subscript " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": " is the index of views. Our goal is to learn a unified, high-quality representation " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": " for each entity by observing these views and filtering view-private information as much as possible. The overall framework of our MetaViewer is shown in Fig. 2, including three main modules and a bi-level optimization process. The outer-level trains a meta-learner to learn an optimal fusion function and derive the unified representation " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": ", and the inner-level reconstructs original views from " + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 557, + 288, + 714 + ], + "type": "text", + "content": " in a few update steps, which explicitly models and separates" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 346, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 396 + ], + "type": "text", + "content": "view-private information and ensure the quality of the representation. In the following subsections, we first introduce the entire structure of the MetaViewer and then elaborate on the bi-level optimization process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 402, + 423, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 423, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 423, + 414 + ], + "type": "text", + "content": "3.1. The entire structure" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "text", + "content": "**Embedding module** aims to transform heterogeneous views into the latent feature space, where transformed view embeddings have the same dimension as each other. To this end, we conduct a view-specific embedding function " + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "text", + "content": " for each view, where " + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "inline_equation", + "content": "v = 1,2,\\dots ,V" + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "text", + "content": ". Given the " + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "text", + "content": "-th view data " + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "text", + "content": " of the entity " + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "text", + "content": ", the corresponding embedding " + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "inline_equation", + "content": "z_{v}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 420, + 547, + 504 + ], + "type": "text", + "content": " can be computed by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 388, + 513, + 546, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 513, + 546, + 526 + ], + "spans": [ + { + "bbox": [ + 388, + 513, + 546, + 526 + ], + "type": "interline_equation", + "content": "z _ {v} = f _ {v} \\left(x _ {v}, \\theta_ {f _ {v}}\\right), \\tag {2}", + "image_path": "17ec803ebcd320522c3c5ca22091a6b7f91efee43a9b1e6a976b838e2627e181.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "text", + "content": " is typically instantiated as a multi-layer neural network with learnable parameters " + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\theta_{f_v}" + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": "Representation learning module maps the obtained embedding to the view-specific or unified representation by constructing a meta-learner " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": " (i.e., MetaViewer). Note that the two types of representations are both derived from the same set of parameters, but at different learning stages (see 3.2). The view-specific representation is obtained at the inner level, where MetaViewer receives the embedding of the " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": "-th view and outputs the corresponding representation " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "h_v" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": ". In contrast, the unified representation " + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": " is learned by MetaViewer from all view embeddings at the outer level and is ultimately used for downstream tasks. Therefore, the structure of MetaViewer should be flexible enough to meet the two requirements above simultaneously. Specifically," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11592" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "type": "text", + "content": "MetaViewer is implemented as a channel-oriented 1-d convolutional layer (C-Conv) with a non-linear function (e.g., ReLU [17]), as shown in Fig. 2 (c)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 109, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 109, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 287, + 168 + ], + "type": "text", + "content": "On the one hand, at the outer level, we first concatenate all view embeddings at the channel level, i.e., the number of the channel in the concatenated embedding is equal to the number of views, and then train the MetaViewer to learn the fusion of cross-view information " + }, + { + "bbox": [ + 47, + 109, + 287, + 168 + ], + "type": "inline_equation", + "content": "H \\in \\mathbb{R}^{d_h}" + }, + { + "bbox": [ + 47, + 109, + 287, + 168 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 178, + 287, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 287, + 192 + ], + "type": "interline_equation", + "content": "H = m \\left(z _ {c a t}, \\omega\\right), \\tag {3}", + "image_path": "717342ae9b92aed5b0bf8debd50871b971e0edd56915db43dc29e67f75ae205e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "inline_equation", + "content": "z_{cat} \\in \\mathbb{R}^{d \\times V}" + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "text", + "content": " indicates the concatenated embedding and " + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "text", + "content": " is the parameter of MetaViewer. On the other hand, the inner-level initializes and trains MetaViewer to learning the representation " + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "inline_equation", + "content": "h_v \\in \\mathbb{R}^{d_h}" + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 200, + 287, + 248 + ], + "type": "text", + "content": "-th view via" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 258, + 287, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 258, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 129, + 258, + 287, + 272 + ], + "type": "interline_equation", + "content": "h _ {v} = m _ {v} \\left(z _ {v}, \\omega_ {v}\\right), \\tag {4}", + "image_path": "c47dade43dd7323dfa6b1c6aae3c14afd697b4e26a1a44e1ec8bcf882bc8a842.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "inline_equation", + "content": "m_v" + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "text", + "content": " denotes the learner with the learnable parameter " + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\omega_v" + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "text", + "content": " for handling the " + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "text", + "content": "-th view in the inner level, also known as the base learner to avoid confusion. It is worth noting that " + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\omega_v" + }, + { + "bbox": [ + 47, + 281, + 287, + 329 + ], + "type": "text", + "content": " is initialized by a subset of the MetaViewer parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 330, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 287, + 376 + ], + "type": "text", + "content": "Self-supervised module provides effectively supervised information for model training by constructing a series of pretext tasks1. These tasks serve as the heads of our framework. Given the input " + }, + { + "bbox": [ + 47, + 330, + 287, + 376 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 330, + 287, + 376 + ], + "type": "text", + "content": ", the output " + }, + { + "bbox": [ + 47, + 330, + 287, + 376 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 330, + 287, + 376 + ], + "type": "text", + "content": " can be formalized as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 140, + 387, + 287, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 387, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 140, + 387, + 287, + 400 + ], + "type": "interline_equation", + "content": "s = t \\left(z, \\theta_ {t}\\right), \\tag {5}", + "image_path": "8c6453195d06f5514046d712e37077cfa3156829716271c5c78fc7fb7460ddda.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "text", + "content": " are the network and its parameters involved in pre-text tasks, respectively. Taking the reconstruction head as an example, it achieves the reconstruction object by re-mapping the representation back to the original view space. In this case, the " + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "inline_equation", + "content": "s := \\hat{x}_v" + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "text", + "content": " indicates the reconstruction result for input " + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "inline_equation", + "content": "x_v" + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "inline_equation", + "content": "t(\\cdot, \\theta_t) := t_v(\\cdot, \\theta_{t_v})" + }, + { + "bbox": [ + 47, + 410, + 287, + 517 + ], + "type": "text", + "content": " is a multi-layer neural network with the opposite structure to the embedding module. More similar pre-text tasks have been studied in previous works [55, 57], and this is not the focus of this work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 525, + 226, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 226, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 226, + 538 + ], + "type": "text", + "content": "3.2. Training via bi-level optimization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 544, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 287, + 676 + ], + "type": "text", + "content": "We now have the entire structure, which can be end-to-end trained following the proposed uniform-to-specific pipeline to derive the unified multi-view representation. The training flowchart and step are shown in the top row in Fig. 3, where the learning of view-private features and unified representation are formulated as a nested bi-level optimization problem. Inner-level focus on independent training on a specific view, which receives a given representation and reconstructs it to each view. Outer-level updates the meta-learner to find the optimal fusion rule through observing the training processes across all views. This pipeline decouples the unified entity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "representations from view-specific features and enables data-driven fusion. Before the detailed description, we introduce a split strategy for multi-view data in meta-learning style to adapt the bi-level training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "text", + "content": "Meta-split of multi-view data. Consider a batch multi-view entity samples " + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "inline_equation", + "content": "\\{D_v^{batch}\\}_{v=1}^V" + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "text", + "content": " from set " + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "text", + "content": ". We randomly and proportionally divide it into two disjoint subsets, marked as support set " + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "inline_equation", + "content": "S = \\{S_v\\}_{v=1}^V" + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "text", + "content": " and query set " + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "inline_equation", + "content": "Q = \\{Q_v\\}_{v=1}^V" + }, + { + "bbox": [ + 304, + 121, + 547, + 215 + ], + "type": "text", + "content": ", respectively. As shown in 2 (a), support set is used in inner-level for leaning view-specific information, so the sample attributes in it could be ignored. In contrast, query set retains both view and sample attributes for outer-level optimization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": "Inner-level optimization. Without loss of generality, take the inner-level update after the " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": "-th outer-level optimization as an example. Let " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\omega^o" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": " be the lasted parameters of the MetaViewer, and " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\theta_v^o = \\{\\theta_{f_v}^o,\\theta_{t_v}^o\\}" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": " denotes the lasted parameters in embedding and self-supervised modules for brevity. We first initial base learner from meta-learner, i.e., " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\omega_v^0 = \\omega^o" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": ", and make a copy of " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\theta_v^o" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}_v" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": ". Note that the copy means gradients with respect to the " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\theta_v^o" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": " will not be backpropagated to " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}_v" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": " and vice versa. Thus, " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\omega_v^0" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}_v" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": " form the initial states of parameters in the inner-level optimization. Suppose " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v^{in}" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": " is the loss function of the inner-level with respect to the " + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 216, + 547, + 360 + ], + "type": "text", + "content": "-th view, the corresponding update goal is" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 338, + 367, + 545, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 367, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 338, + 367, + 545, + 387 + ], + "type": "interline_equation", + "content": "\\omega_ {v} ^ {*} \\left(\\omega^ {o}\\right) = \\arg \\min \\mathcal {L} _ {v} ^ {i n} \\left(\\omega_ {v} \\left(\\omega^ {o}\\right), \\tilde {\\phi} _ {v}; S _ {v}\\right). \\tag {6}", + "image_path": "ca8af399f783cf2620a9c17569c183512052368e310be53526c6f85a69ab4f52.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 391, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 415 + ], + "type": "text", + "content": "Consider a gradient descent strategy (e.g., SGD [4]), we can further write the update process of " + }, + { + "bbox": [ + 304, + 391, + 545, + 415 + ], + "type": "inline_equation", + "content": "\\omega_{v}" + }, + { + "bbox": [ + 304, + 391, + 545, + 415 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 347, + 422, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 422, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 347, + 422, + 545, + 449 + ], + "type": "interline_equation", + "content": "\\omega_ {v} ^ {i} = \\omega_ {v} ^ {i - 1} - \\beta \\frac {\\partial \\mathcal {L} _ {v} ^ {i n}}{\\partial \\omega_ {v} ^ {i - 1}}, \\dots , \\omega_ {v} ^ {0} = \\omega^ {o}, \\tag {7}", + "image_path": "7a793d31d728c95dbae1b06ebd63686416d5bcda3069c62b1612f2f71a389f65.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "text", + "content": " denote the learning rate and iterative step of inner-level optimization, respectively." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 479, + 547, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 551 + ], + "type": "text", + "content": "Outer-level optimization. After several inner-level updates, we obtain a set of optimal view-specific parameters for the support set. Outer-level then updates the meta-learner, embedding, and self-supervised modules by training on the query set. With the loss function " + }, + { + "bbox": [ + 304, + 479, + 547, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out}" + }, + { + "bbox": [ + 304, + 479, + 547, + 551 + ], + "type": "text", + "content": ", the outer-level optimization goal is" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 558, + 545, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 545, + 573 + ], + "type": "interline_equation", + "content": "\\omega^ {*}, \\left\\{\\theta_ {v} ^ {*} \\right\\} _ {v = 1} ^ {V} = \\arg \\min \\mathcal {L} ^ {\\text {o u t}} \\left(\\left\\{\\omega_ {v} ^ {*} (\\omega), \\phi_ {v} \\right\\} _ {v = 1} ^ {V}; Q\\right). \\tag {8}", + "image_path": "0b208797cf728a12d8ca7520f6b466a288544713ee32c3b353ed2da803d9e733.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "type": "text", + "content": "By alternately optimizing Eq. 6 and Eq. 8, we end up with the optimal meta-parameters " + }, + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "type": "inline_equation", + "content": "\\omega^{*}" + }, + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "type": "text", + "content": " and a set of view-specific parameters " + }, + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "type": "inline_equation", + "content": "\\{\\theta_v^*\\}_{v = 1}^V" + }, + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "type": "text", + "content": ". Given a test sample " + }, + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "type": "inline_equation", + "content": "x_{test}" + }, + { + "bbox": [ + 304, + 580, + 547, + 652 + ], + "type": "text", + "content": ", the corresponding unified representation is derived by feeding it sequentially into the embedding module and the meta-learner. The overall framework of MetaViewer is shown in Alg. 1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 659, + 542, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 542, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 542, + 672 + ], + "type": "text", + "content": "3.3. Specific-to-uniform versus uniform-to-specific" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "We discuss the difference between specific-to-uniform and our uniform-to-specific pipeline through three aspects shown in Fig. 3, including the overall flowchart, update steps," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "Constructing pre-text tasks is a classic strategy in self-supervised learning [1], which emphasizes the use of intrinsic properties within or between data as ground-truth to provide supervised information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11593" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 244, + 83 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 244, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 244, + 83 + ], + "type": "text", + "content": "Algorithm 1 The framework of our MetaViewer." + } + ] + } + ], + "index": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "text", + "content": "Require: Training dataset " + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "text", + "content": ", meta parameters " + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "text", + "content": ", base parameters " + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "inline_equation", + "content": "\\{\\omega_v\\}_{v=1}^V" + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "text", + "content": ", view-specific parameters " + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "inline_equation", + "content": "\\{\\theta_v\\}_{v=1}^V" + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "text", + "content": ", the number of views " + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "text", + "content": ", the iteration step in inner-level optimization " + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 86, + 288, + 132 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 133, + 246, + 301 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 133, + 154, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 133, + 154, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 133, + 154, + 148 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 53, + 133, + 154, + 148 + ], + "type": "inline_equation", + "content": "\\omega, \\{\\theta_v\\}_{v=1}^V" + }, + { + "bbox": [ + 53, + 133, + 154, + 148 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 140, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 140, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 140, + 158 + ], + "type": "text", + "content": "2: while not done do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 159, + 135, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 135, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 135, + 169 + ], + "type": "text", + "content": "3: # Outer-level" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 171, + 246, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 246, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 246, + 182 + ], + "type": "text", + "content": "4: Sample and meta-split a batch set from " + }, + { + "bbox": [ + 53, + 171, + 246, + 182 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 53, + 171, + 246, + 182 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 182, + 230, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 230, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 230, + 196 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 53, + 182, + 230, + 196 + ], + "type": "inline_equation", + "content": "\\{D_v^{batch}\\}_{v = 1}^V = \\{\\tilde{S}_v\\}_{v = 1}^V +\\{Q_v\\}_{v = 1}^V." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 196, + 157, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 157, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 157, + 206 + ], + "type": "text", + "content": "6: for " + }, + { + "bbox": [ + 53, + 196, + 157, + 206 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + }, + { + "bbox": [ + 53, + 196, + 157, + 206 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 207, + 169, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 207, + 169, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 169, + 218 + ], + "type": "text", + "content": "7: for " + }, + { + "bbox": [ + 53, + 207, + 169, + 218 + ], + "type": "inline_equation", + "content": "v = 1, \\ldots, V" + }, + { + "bbox": [ + 53, + 207, + 169, + 218 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 219, + 156, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 219, + 156, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 156, + 229 + ], + "type": "text", + "content": "8: #Inner-level" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 230, + 206, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 206, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 206, + 243 + ], + "type": "text", + "content": "9: Initialize " + }, + { + "bbox": [ + 53, + 230, + 206, + 243 + ], + "type": "inline_equation", + "content": "\\omega_{v} = \\omega ,\\tilde{\\theta}_{v} = \\theta_{v};" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 243, + 231, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 243, + 231, + 255 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 231, + 255 + ], + "type": "text", + "content": "10: Optimize " + }, + { + "bbox": [ + 50, + 243, + 231, + 255 + ], + "type": "inline_equation", + "content": "\\omega_v(\\omega)" + }, + { + "bbox": [ + 50, + 243, + 231, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 50, + 243, + 231, + 255 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 50, + 243, + 231, + 255 + ], + "type": "text", + "content": " via Eq. 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 255, + 118, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 255, + 118, + 265 + ], + "spans": [ + { + "bbox": [ + 50, + 255, + 118, + 265 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 267, + 108, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 267, + 108, + 277 + ], + "spans": [ + { + "bbox": [ + 50, + 267, + 108, + 277 + ], + "type": "text", + "content": "12: end for" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 278, + 219, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 278, + 219, + 292 + ], + "spans": [ + { + "bbox": [ + 50, + 278, + 219, + 292 + ], + "type": "text", + "content": "13: Optimize " + }, + { + "bbox": [ + 50, + 278, + 219, + 292 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 50, + 278, + 219, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 50, + 278, + 219, + 292 + ], + "type": "inline_equation", + "content": "\\{\\phi_v\\}_{v=1}^V" + }, + { + "bbox": [ + 50, + 278, + 219, + 292 + ], + "type": "text", + "content": " via Eq. 8." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 50, + 292, + 108, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 292, + 108, + 301 + ], + "spans": [ + { + "bbox": [ + 50, + 292, + 108, + 301 + ], + "type": "text", + "content": "14: end while" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "content": "and key gradients. For clarity, we here slightly confuse the notation and let " + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v^{in}" + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "content": " correspond to the view-specific parameters and loss functions in specific-to-uniform pipeline, respectively. Similarly, " + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out}" + }, + { + "bbox": [ + 46, + 324, + 287, + 396 + ], + "type": "text", + "content": " denote the view-shared parameters and loss functions, respectively. Let us start by reviewing two branches in specific-to-uniform framework." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "text", + "content": "One-level methods is a classical branch in specific-to-uniform framework, where parameters " + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\{\\theta_v\\}_{v=1}^V" + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "text", + "content": " are simultaneously optimized by combining the functions " + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v^{in}" + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out}" + }, + { + "bbox": [ + 46, + 396, + 288, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "text", + "content": "Multi-level methods also follow the specific-to-uniform framework, the difference is that they hierarchically learn specific and shared features to avoid interference. That is, parameters " + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "text", + "content": " are updated at the low level by minimizing the " + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v^{in}" + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "text", + "content": ", and parameters " + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "text", + "content": " are updated at the high level by minimizing the " + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out}" + }, + { + "bbox": [ + 46, + 444, + 287, + 516 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 516, + 288, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 288, + 589 + ], + "type": "text", + "content": "Bi-level MetaViewer further decouples view-private information, except for view-specific and view-sharing components. Consider the simplest case, where the base learner updates only one step on each view in the inner loop, i.e., " + }, + { + "bbox": [ + 46, + 516, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\omega_{v}^{*} = \\omega -\\beta \\nabla_{\\omega}\\mathcal{L}_{v}^{in}(\\omega)" + }, + { + "bbox": [ + 46, + 516, + 288, + 589 + ], + "type": "text", + "content": ". Thus, the final optimization of " + }, + { + "bbox": [ + 46, + 516, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 516, + 288, + 589 + ], + "type": "text", + "content": " in the outer level in Eq. 8 can be rewritten as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 58, + 597, + 288, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 597, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 58, + 597, + 288, + 622 + ], + "type": "interline_equation", + "content": "\\min _ {\\omega} \\sum_ {v} \\mathcal {L} ^ {o u t} (\\omega_ {v} ^ {*}) = \\sum_ {v} \\mathcal {L} ^ {o u t} \\left(\\omega - \\beta \\nabla_ {\\omega} \\mathcal {L} _ {v} ^ {i n} (\\omega)\\right). \\quad (9)", + "image_path": "76560b1ebf6f362dba13e2d0e385d875c950c09f4d9c0c939449cc7f3764494a.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "Intuitively, the optimal " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is expected to be close enough to each view that the view-specific representation can be produced after once updating. In other words, our uniform-to-specific framework requires explicit modeling the gap caused by view-private information when mapping from a unified representation to a particular view, which is achieved by inner-level optimization in MetaViewer." + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 309, + 72, + 544, + 132 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 544, + 132 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 544, + 132 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 544, + 132 + ], + "type": "table", + "html": "
our bi-level MetaViewer uniform-to-specific
Q→θv→ωv(ω)→Lvinωv(ω)→ωv*(ω)∂Lvin/∂ωv(ω)
DInner-Level
S→θv→Σv→ω→Loutθv→θv*, ω→ω*∂Lout/∂θv, ∂Lout/∂2ω
", + "image_path": "73fa9fb5a546aa25fb4461c13bf5b574097d7ab2e5e3f546531e43917913f6d0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 310, + 133, + 544, + 179 + ], + "blocks": [ + { + "bbox": [ + 310, + 133, + 544, + 179 + ], + "lines": [ + { + "bbox": [ + 310, + 133, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 310, + 133, + 544, + 179 + ], + "type": "table", + "html": "
one-level methodsspecific-to-uniform
D→θv→Σv→ωLvnθv→θ*, ω→ω*∂(Lout+Lvn)/∂θv, ∂(Lout+∑v) / ∂ω
", + "image_path": "b425a2c23c424eb6ef6f011ab586bd3ba2683436f5f35711aaa551fd3b95e5ed.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 310, + 180, + 544, + 223 + ], + "blocks": [ + { + "bbox": [ + 310, + 180, + 544, + 223 + ], + "lines": [ + { + "bbox": [ + 310, + 180, + 544, + 223 + ], + "spans": [ + { + "bbox": [ + 310, + 180, + 544, + 223 + ], + "type": "table", + "html": "
two-/multi-level methods
D→θvLinvθv→θ*Low-Level
ω→ω*High-Level
∂Lu/∂θv
", + "image_path": "4fec3400b26f4460afd80131bbad9eee6a1501caf186b4683a2d31a47e1316dd.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 336, + 224, + 387, + 235 + ], + "lines": [ + { + "bbox": [ + 336, + 224, + 387, + 235 + ], + "spans": [ + { + "bbox": [ + 336, + 224, + 387, + 235 + ], + "type": "text", + "content": "(a) Flowcharts" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 427, + 224, + 459, + 235 + ], + "lines": [ + { + "bbox": [ + 427, + 224, + 459, + 235 + ], + "spans": [ + { + "bbox": [ + 427, + 224, + 459, + 235 + ], + "type": "text", + "content": "(b) Steps" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 487, + 224, + 533, + 235 + ], + "lines": [ + { + "bbox": [ + 487, + 224, + 533, + 235 + ], + "spans": [ + { + "bbox": [ + 487, + 224, + 533, + 235 + ], + "type": "text", + "content": "(c) Gradients" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 244, + 547, + 323 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 323 + ], + "type": "text", + "content": "Figure 3. Comparison of two frameworks in terms of flowchart, update steps, and key gradients. The top row is our bi-level MetaViewer with uniform-to-specific framework, where red and black arrows indicate the flow at outer-level and inner-level, respectively. The bottom two rows show two main branches in uniform-to-specific framework, i.e., one-level methods [47] and two- or multi-level methods [51]." + } + ] + } + ], + "index": 29, + "type": "text" + }, + { + "bbox": [ + 306, + 342, + 507, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 507, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 507, + 357 + ], + "type": "text", + "content": "3.4. The instances of the objective function" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 361, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 422 + ], + "type": "text", + "content": "Our uniform-to-specific framework emphasizes learning from the training process in the inner loop. In unsupervised multi-view learning, this process typically refers to the reconstruction on views, thus the " + }, + { + "bbox": [ + 304, + 361, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v^{in}" + }, + { + "bbox": [ + 304, + 361, + 547, + 422 + ], + "type": "text", + "content": " is specified as the reconstruction loss [27, 60]" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 342, + 430, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 430, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 342, + 430, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v} ^ {i n} = \\mathcal {L} _ {v} ^ {r e c} \\left(S _ {v}, S _ {v} ^ {r e c}\\right) = \\| S _ {v} - S _ {v} ^ {r e c} \\| _ {F} ^ {2}. \\tag {10}", + "image_path": "9ce8c10c8f82ae91fd7fda949d3ba6412d7eb18472c9b82a98be06048daa17bc.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 453, + 547, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 513 + ], + "type": "text", + "content": "While parameters updated at the outer level can be constrained by richer self-supervised feedbacks, as maintained in Sec. 3.1. Here we provide two instances of the outer-level loss function to demonstrate how MetaViewer can be extended with different learning objectives." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 513, + 547, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 547, + 548 + ], + "type": "text", + "content": "MVer-R adopts the same reconstruction loss as the inner-level, and " + }, + { + "bbox": [ + 304, + 513, + 547, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out} = \\sum_{v}\\mathcal{L}_{v}^{rec}(Q_{v},Q_{v}^{rec})" + }, + { + "bbox": [ + 304, + 513, + 547, + 548 + ], + "type": "text", + "content": ", which is the purest implementation of MetaViewer." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 548, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 628 + ], + "type": "text", + "content": "MVer-C additionally utilizes a contrastive objective, where the similarities of views belonging to the same entity (i.e., positive pairs) should be maximized and those of different entities (i.e., negative pairs) should be minimized, i.e., " + }, + { + "bbox": [ + 304, + 548, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out} = \\sum_{v}\\left(\\mathcal{L}_{v}^{rec} + \\sum_{v^{\\prime},v^{\\prime}\\neq v}\\mathcal{L}_{v,v^{\\prime}}^{con}\\right)" + }, + { + "bbox": [ + 304, + 548, + 547, + 628 + ], + "type": "text", + "content": ". Following previous work [18, 51, 55], the contrastive loss " + }, + { + "bbox": [ + 304, + 548, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{v,v^{\\prime}}^{con}" + }, + { + "bbox": [ + 304, + 548, + 547, + 628 + ], + "type": "text", + "content": " is formed as" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 635, + 554, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 554, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v, v ^ {\\prime}} ^ {c o n} = - \\frac {1}{N _ {Q}} \\sum_ {i = 1} ^ {N _ {Q}} \\log \\frac {e ^ {d \\left(q _ {v} ^ {i} , q _ {v ^ {\\prime}} ^ {i}\\right) / \\tau}}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {Q}} e ^ {d \\left(q _ {v} ^ {i} , q _ {v} ^ {j}\\right) / \\tau} + \\sum_ {j = 1} ^ {N _ {Q}} e ^ {d \\left(q _ {v} ^ {i} , q _ {v ^ {\\prime}} ^ {j}\\right) / \\tau}}, \\tag {11}", + "image_path": "d13a995742237e9724573c758d88e9350a2a7a42a4c648a2bf11b4b445c7c990.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "q_v^i" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "-th view of the " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "-th query sample and " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is the similarity metric (e.g., cosine similarity [6]). The " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "N_Q" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " denote the number of query set samples and the" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11594" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 497, + 156 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 497, + 156 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 497, + 156 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 497, + 156 + ], + "type": "table", + "html": "
Datasets#Views#Classes#Samples (train, val, test)View Dimensions
Handwirtten2102000 (1200, 400, 400)240; 216
RGB-D250500 (300, 100, 100)12288; 4096
Animal25010158 (6074, 2011, 2073)4096; 4096
Fashion-MV31010000 (6000, 2000, 2000)784; 784; 784
Caltech101-206202386 (1425, 469, 492)48; 40; 254; 1984; 512; 928
", + "image_path": "b526b25e2bc8bd1f831a449229665b862a84f816421eb92fd916b751fb875700.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 186, + 544, + 322 + ], + "blocks": [ + { + "bbox": [ + 181, + 163, + 411, + 175 + ], + "lines": [ + { + "bbox": [ + 181, + 163, + 411, + 175 + ], + "spans": [ + { + "bbox": [ + 181, + 163, + 411, + 175 + ], + "type": "text", + "content": "Table 1. The attributes for all datasets used in our experiments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 186, + 544, + 322 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 544, + 322 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 544, + 322 + ], + "type": "table", + "html": "
MethodsHandwrittenRGB-DAnimalFashion-MVCaltech101-20
ACCNMIARIACCNMIARIACCNMIARIACCNMIARIACCNMIARI
Baseline43.5552.1925.6848.3080.9704.9545.7966.2233.8750.9145.9832.0936.6348.9623.54
DCCA [2] (2013)57.2569.8052.1551.0082.9952.0260.3468.5146.1070.7080.4261.8038.6250.8822.73
DCCAE [47] (2015)63.0075.0459.2948.0081.5848.3465.1766.8248.8371.0581.1262.3436.5952.2425.25
MIB [8] (2020)63.2567.5852.1650.0081.1351.2760.8961.9852.4957.2073.8347.6235.9847.0022.18
WTNNM [15] (2020)75.6977.6261.5152.5481.9552.4671.9271.4361.0683.3386.3077.9939.7455.5826.92
TLRR [23] (2021)78.6077.9966.4052.8782.3853.3769.7973.0161.2379.5484.0178.5440.1852.3827.60
MFLVC [51] (2022)64.0064.5348.8553.0083.3154.0774.1076.0864.2883.2088.7578.9336.5958.3626.87
DCP [27] (2022)66.2570.5656.1052.0082.0452.6472.7774.0265.4862.6068.3854.3036.7944.3723.50
MVer-R (ours)75.0078.5367.2153.0082.4153.0476.4978.2565.2880.8088.1375.0541.8758.5229.19
MVer-C (ours)86.2578.9672.2557.0084.9757.0775.9278.0166.0785.4088.7680.0745.1260.8635.00
", + "image_path": "0348a81c98b12687185c2b641c2fc40a771ea983139cf1267f716ffa88d7b6bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 330, + 531, + 341 + ], + "lines": [ + { + "bbox": [ + 60, + 330, + 531, + 341 + ], + "spans": [ + { + "bbox": [ + 60, + 330, + 531, + 341 + ], + "type": "text", + "content": "Table 2. Clustering results of all methods on six datasets. Bold and underline denote the best and second-best results, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 362, + 288, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 288, + 398 + ], + "type": "text", + "content": "temperature parameter, respectively. Note that, the derived meta representation can also be used in contrastive learning as an additional novel view." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 412, + 128, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 128, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 128, + 426 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 433, + 288, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 288, + 553 + ], + "type": "text", + "content": "In this section, we present extensive experimental results to validate the quality of the unified representation derived from our MetaViewer. The remainder of the experiments are organized as follows: Subsection 4.1 lists datasets, compared methods, and implementation details. Subsection 4.2 compares the performance of our method with classical and state-of-the-art methods on two common downstream scenarios, clustering and classification tasks. Comparisons with manually designed fusion and ablation studies are shown in Subsections 4.3 and 4.4, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 163, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 163, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 163, + 576 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Datasets. To comprehensively evaluate the effectiveness of our MetaViewer, we conduct five multi-view benchmarks in experiments. All datasets are scaled to " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " and split into training, validation, and test sets in the ratio of " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "6:2:2" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ", as shown in Tab. 1. Handwritten contains 2,000 handwritten digital images from 0 to 9, where two types of descriptors, i.e., 240-D pixel average in " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "2\\times 3" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " windows and 216-D profile correlations, are selected as two views [60]. RGB-D dataset contains visual and depth images of 300 distinct objects across 50 categories [46, 63]. Two views are obtained by flattening the " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "64\\times 64\\times 3" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " color images and " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " depth" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 362, + 547, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 547, + 481 + ], + "type": "text", + "content": "images. Animal consists of 10158 images from 50 classes with two views. Unlike other datasets, these two views are features extracted by deep networks [59]. Fashion-MV is an image dataset that contains 10 categories with a total of 30,000 fashion products. It has three views and each of which consists of 10,000 gray images sampled from the same category [50]. Caltech101-20 is a subset of the Caltech101 image set [9], which consists of 2,386 images of 20 subjects. Six features are used, including Gabor, Wavelet Moments, CENTRIST, HOG, GIST, and LBP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 483, + 548, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 548, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 548, + 688 + ], + "type": "text", + "content": "Compared methods. We compare the performance of MetaViewer with seven representative multi-view learning methods, including two classical methods (DCCA [2] and DCCAE [47]) and five state-of-the-art methods (WTNNM [15], TLRR [23], MIB [8], MFLVC [51] and DCP [27]). Among them, DCCA and DCCAE are the deep extensions of traditional correlation strategies. WTNNM and TLRR are representative tensor-based methods. MIB is a typical generative method with mutual information constraints. DCP learns unified representation in both unsupervised and supervised scenarios, and we report the unsupervised version for a fair comparison. In particular, MFLVC also notices the view-private redundant information and designs a multilevel feature network for clustering tasks. In addition, we also compare the model without MetaViewer as a baseline, where the unified representation is obtained by concatenating view-specific features output by the embedding module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Implementation details. For a fair comparison, all methods are trained from scratch, and all deep-network-based" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11595" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 198 + ], + "type": "table", + "html": "
MethodsHandwrittenRGB-DAnimalFashion-MVCaltech101-20
ACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-scoreACCPrec.F-score
Baseline87.0087.3287.0414.0006.4507.7071.7865.0263.1087.5087.6387.5475.4151.9546.21
DCCA [2] (2013)88.2589.2088.0530.0021.1022.0462.8461.6160.6584.9085.2283.5471.5445.2739.81
DCCAE [47] (2015)90.0090.4889.9224.0016.0016.9166.8262.5362.8385.3585.9783.8471.5460.5743.25
MIB [8] (2020)79.0083.9078.5233.0028.5027.3759.3263.7862.1386.8086.8086.5572.7261.6452.47
WTNNM [15] (2020)96.7796.2096.3645.0047.1843.9069.9067.4065.4894.5094.5094.5883.2780.4974.78
TLRR [23] (2021)97.0097.0597.1749.0053.3247.6671.9069.5268.3596.3596.2896.4085.5577.3075.24
MFLVC [51] (2022)94.0094.2094.0144.0046.0941.8175.6272.1770.8196.5096.5296.4985.3771.8369.07
DCP [27] (2022)97.2597.3097.2437.0028.8730.7877.9573.4370.0189.2582.0682.9092.4889.4184.58
MVer-R (ours)97.0097.0897.0051.0053.6548.7377.6973.9171.2296.8596.3796.4892.2889.4684.21
MVer-C (ours)97.7597.9097.7556.0055.2052.7878.0374.5671.5597.7096.7897.0792.1690.6885.72
", + "image_path": "128e9fe9b96af3afffe7992a100375c652893922ba197b92202bae4fa59fa96c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "text", + "content": "methods share the same embedding backbone, following the previous work [51]. For two-view-based methods, such as DCCA, DCCAE, and MIB, we report the results of the two best-performing views. The final unified representation " + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "text", + "content": " in alignment-based methods is a concatenation of the processed features, and the representation dimension is specified as " + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "inline_equation", + "content": "d_h = 256" + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "text", + "content": ". The results on clustering and classification tasks are obtained by subsequent K-means and SVM classifier, respectively. For MetaViewer, we train 2,000 epochs for all benchmarks, and set the batch size is 32 for RGBD and 128 for others. The learning rates in outer- and inner-level are set to " + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 47, + 238, + 290, + 454 + ], + "type": "text", + "content": ", respectively. We use a single convolutional layer with 32 kernels of size 5 as the meta-learner, and we set the proportion of support set and the update step in the inner level to 0.2 and 2, respectively. All experiments have been verified using the PyTorch library on a single RTX3090. Code is available at https://github.com/xxLifeLover/MetaViewer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 465, + 231, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 231, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 231, + 477 + ], + "type": "text", + "content": "4.2. Performance on downstream tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 651 + ], + "type": "text", + "content": "Clustering results. Tab. 2 lists the results of the clustering task, where the performance is measured by three standard evaluation metrics, i.e., Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI). A higher value of these metrics indicates better clustering performance. It can be observed that (1) our MVer-C variant significantly outperforms other compared methods on all benchmarks; (2) the second-best results appear between MVer-R and MFLVC, both of which explicitly separate the view-private information; (3) a larger number of categories and views is the main reason for the degradation of clustering performance, and our Metaviewer improves most significantly in such a scenario (e.g., Fashion-MV and Caltech101-20)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Classification results. Tab. 3 lists the results of the classification task, where three common metrics are used, including Accuracy, Precision, and F-score. A higher value indicates better classification performance. Similar to the clustering results, two variants of MetaViewer significantly" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 236, + 544, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 537, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 537, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 537, + 217 + ], + "type": "text", + "content": "Table 3. Classification results of all methods on six datasets. Bold and underline denote the best and second-best results, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 236, + 544, + 319 + ], + "lines": [ + { + "bbox": [ + 309, + 236, + 544, + 319 + ], + "spans": [ + { + "bbox": [ + 309, + 236, + 544, + 319 + ], + "type": "table", + "html": "
StrategiesRulesACC↑NMI↑ARI↑MSE↓
Sumzx + zy69.2571.8959.021.84
Maxmax(zx, zy)80.7573.9363.75-
Concat.cat[zx, zy]78.7572.0261.521.77
Linearl(zx, zy, θl)85.0077.4069.714.74
C-Convm(zx, zy, ω)69.7565.2151.332.37
MetaViewermeta-learning86.2578.9672.252.45
", + "image_path": "a060463772f5583fc92a76d3b84f49fa29ac6ad78702191595d161c3e3e95c03.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 322, + 327, + 529, + 338 + ], + "lines": [ + { + "bbox": [ + 322, + 327, + 529, + 338 + ], + "spans": [ + { + "bbox": [ + 322, + 327, + 529, + 338 + ], + "type": "text", + "content": "Table 4. Clustering resulting on the Handwritten dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 320, + 352, + 532, + 464 + ], + "blocks": [ + { + "bbox": [ + 320, + 352, + 532, + 464 + ], + "lines": [ + { + "bbox": [ + 320, + 352, + 532, + 464 + ], + "spans": [ + { + "bbox": [ + 320, + 352, + 532, + 464 + ], + "type": "image", + "image_path": "6baaf3e6a29beef395c5ab07655a3d9c308950faeffa017db965f0ecdf1e1b55.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 474, + 547, + 509 + ], + "lines": [ + { + "bbox": [ + 305, + 474, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 547, + 509 + ], + "type": "text", + "content": "Figure 4. Class activate maps on RGB-D samples for different fusion strategies, including single views, Sum, and our MetaViewer. View 1 and view 2 are the RGB and depth view, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 530, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 547, + 601 + ], + "type": "text", + "content": "outperform the comparison methods. It is worth noting that (1) DCP learns unified representation and therefore achieves the second-best result instead of MFLVC. (2) The number of categories is the main factor affecting the classification performance, and our method obtains the most significant improvement in the RGB-D dataset with up to 50 classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 610, + 534, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 534, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 534, + 624 + ], + "type": "text", + "content": "4.3. Comparison with manually designed fusion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "As mentioned in 3.3, MetaViewer essentially learns an optimal fusion rule that preserves as much consistency and complementary information as possible while filtering out view-private information. To verify this, we compare it with commonly used fusion strategies [26], including sum, maxima, concatenation, linear layer and " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "-Conv. The former three are the specified fusion rules without trainable param-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11596" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 167, + 156 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 167, + 156 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 167, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 167, + 156 + ], + "type": "image", + "image_path": "5f48bdb6a99fc4b704fcbbf5744ca4dddb96009c29daa933fb69f0bc450bae03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 161, + 151, + 171 + ], + "lines": [ + { + "bbox": [ + 59, + 161, + 151, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 161, + 151, + 171 + ], + "type": "text", + "content": "(a) Effect of network width" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 71, + 284, + 156 + ], + "blocks": [ + { + "bbox": [ + 167, + 71, + 284, + 156 + ], + "lines": [ + { + "bbox": [ + 167, + 71, + 284, + 156 + ], + "spans": [ + { + "bbox": [ + 167, + 71, + 284, + 156 + ], + "type": "image", + "image_path": "e8e719885cd11547462682ff6d9cbac33b9b3b85e7688e22ad2c2f240f29672d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 161, + 260, + 171 + ], + "lines": [ + { + "bbox": [ + 180, + 161, + 260, + 171 + ], + "spans": [ + { + "bbox": [ + 180, + 161, + 260, + 171 + ], + "type": "text", + "content": "(b) Effect of kernel size" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 217, + 168, + 303 + ], + "blocks": [ + { + "bbox": [ + 47, + 180, + 288, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 180, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 288, + 202 + ], + "type": "text", + "content": "Figure 5. Effect of meta-learner architectures with different depth, width, and kernel size on classification accuracy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 217, + 168, + 303 + ], + "lines": [ + { + "bbox": [ + 53, + 217, + 168, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 168, + 303 + ], + "type": "image", + "image_path": "3a6e0a0af9705b163b1b6c0f5014a349a1123fbfe531cbf97ecfb7ee29a7720a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 306, + 159, + 316 + ], + "lines": [ + { + "bbox": [ + 64, + 306, + 159, + 316 + ], + "spans": [ + { + "bbox": [ + 64, + 306, + 159, + 316 + ], + "type": "text", + "content": "(a) Proportion of support set" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 326, + 287, + 348 + ], + "lines": [ + { + "bbox": [ + 47, + 326, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 287, + 348 + ], + "type": "text", + "content": "Figure 6. Effect of (a) different meta-division ratios and (b) the number of inner-loop iterations on classification accuracy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 170, + 217, + 285, + 303 + ], + "blocks": [ + { + "bbox": [ + 170, + 217, + 285, + 303 + ], + "lines": [ + { + "bbox": [ + 170, + 217, + 285, + 303 + ], + "spans": [ + { + "bbox": [ + 170, + 217, + 285, + 303 + ], + "type": "image", + "image_path": "0c60e460c8994000eddf195e84b26d7a58a01f158a284f0a03f1a4466edc0fc4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 305, + 280, + 316 + ], + "lines": [ + { + "bbox": [ + 179, + 305, + 280, + 316 + ], + "spans": [ + { + "bbox": [ + 179, + 305, + 280, + 316 + ], + "type": "text", + "content": "(b) Steps of inner-level update" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 371, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 503 + ], + "type": "text", + "content": "eters, and the remaining two are the trainable fusion layer trained via the specific-to-uniform manner. Tab. 4 lists the clustering results and an additional MSE score on the Hand-written dataset with the same embedding and reconstruction network. We can observe that (1) trainable fusion layers outperform the hand-designed rules, and our MetaViewer yields the best performance; (2) the MSE scores listed in the last column indicate that the quality of the unified representation cannot be measured and guaranteed only with the reconstruction constraint, due to the view-private redundant information mixed in view-specific latent features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 504, + 288, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 288, + 564 + ], + "type": "text", + "content": "To further explore the fusion preferences learned by MetaViewer, we visualize the class activation map for the RGB-D samples in Fig. 4. MetaViewer balances information from all views instead of just the salient one (view 1), ensuring a high-quality unified multi-view representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 574, + 148, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 148, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 148, + 586 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Meta-learner structures. We implement the meta-learner as a channel-level convolution structure in this work. Albeit simple, this layer can be considered as a universal approximator for almost any continuous function [40], and thus can fit a wide range of conventional fusion functions. To investigate the effect of network depth, width, and convolution kernel size on the performance of the representation, we alternately fix the 32 kernels and the " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 3" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " kernel size and show the classification results on Handwritten data in Fig. 5. It is clear that (1) the meta-learner works well with just a" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "shallow structure, as shown in Fig. 5 (a), instead of gradually overfitting to the training data as the network deepens or widens, and (2) our MetaViewer is stable and insensitive to the hyper-parameters within reasonable ranges." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 122, + 547, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 547, + 243 + ], + "type": "text", + "content": "Meta-split ratios. Fig. 6 (a) shows the impact of the meta-split mentioned in Sec. 3.2 on the classification performance, where the proportion of support set is set from 0.1 to 0.9 in steps of 0.1, and the rest is query set. In addition to the single view, we also compare the sum and concat. fusion as baselines. MetaViewer consistently surpasses all baselines over the experimental proportion. In addition, fusion baselines are more dependent on the better-performing view at lower proportions, instead becoming unstable as the available query sample decreases." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "text", + "content": "Inner-level update steps. Another hyper-parameter is the number of iteration steps in inner-level optimization. More iterations mean a larger gap from the learned meta representation to the specific view space, i.e., coarser modeling of view-private information. Fig. 6 (b) shows the classification results with various steps, where " + }, + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "text", + "content": " steps mean that the inner-level optimization is updated " + }, + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "text", + "content": " times throughout the training. MetaViewer achieves the best results when using 1 steps, and remains stable within 15 steps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 373, + 378, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 378, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 378, + 384 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 396, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 575 + ], + "type": "text", + "content": "This work introduced a novel meta-learning perspective for multi-view learning, and proposed a meta-learner, namely MetaViewer, to derive a high-quality unified representation for downstream tasks. In contrast to the prevailing specific-to-uniform pipeline, MetaViewer observes the reconstruction process from unified representation to specific views and essentially learns an optimal data-driven fusion that separates and filters out meaningless view-private information. The proposed framework is compatible with most excellent objective functions and pre-text tasks to cope with complex real-world scenarios. Extensive experimental results on clustering and classification tasks demonstrate the performance of meta-learned unified representation. In addition, we believe that our bi-level training pipeline is also promising for handling incomplete views, and leave it to future work." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 595, + 408, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 408, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 408, + 608 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "This research was supported in part by the Natural Science Foundation of China (No. 62106129, 62176139, and 62177031), the Natural Science Foundation of Shandong Province (No. ZR2021QF053, ZR2021ZD15), and the China Postdoctoral Science Foundation (No. 2021TQ0195, 2021M701984). The author acknowledges Prof. Xiushan Nie of Shandong Jianzhu University (China) for the financial support and valuable discussions." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11597" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] Sara Atito Ali Ahmed, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. CoRR, abs/2104.03602, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Galen Andrew, Raman Arora, Jeff A. Bilmes, and Karen Livescu. Deep canonical correlation analysis. In ICML, volume 28, pages 1247-1255, 2013. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "type": "text", + "content": "[3] Philip Bachman, R. Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In NeurIPS, pages 15509-15519, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 216 + ], + "type": "text", + "content": "[4] Léon Bottou. Large-scale machine learning with stochastic gradient descent. In COMPSTAT, pages 177-186, 2010. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 217, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 260 + ], + "type": "text", + "content": "[5] Abhra Chaudhuri, Massimiliano Mancini, Yanbei Chen, Zeynep Akata, and Anjan Dutta. Cross-modal fusion distillation for fine-grained sketch-based image retrieval. In BMVC, page 499, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 262, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 304 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119, pages 1597-1607, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 307, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 350 + ], + "type": "text", + "content": "[7] Yanhua Cheng, Xin Zhao, Rui Cai, Zhiwei Li, Kaiqi Huang, and Yong Rui. Semi-supervised multimodal deep learning for RGB-D object recognition. In IJCAI, pages 3345-3351, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 352, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 288, + 384 + ], + "type": "text", + "content": "[8] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushner, and Zeynep Akata. Learning robust representations via multiview information bottleneck. In ICLR, 2020. 1, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "text", + "content": "[9] Li Fei-Fei, Robert Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Comput. Vis. Image Underst., 106(1):59-70, 2007. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 431, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 287, + 463 + ], + "type": "text", + "content": "[10] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, pages 1933-1941, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "type": "text", + "content": "[11] Fangxiang Feng, Xiaojie Wang, and Ruifan Li. Cross-modal retrieval with correspondence autoencoder. In ACM MM, pages 7-16, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 499, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 288, + 532 + ], + "type": "text", + "content": "[12] Yifan Feng, Zizhao Zhang, Xibin Zhao, Rongrong Ji, and Yue Gao. GVCNN: group-view convolutional neural networks for 3d shape recognition. In CVPR, pages 264-272, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 533, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 566 + ], + "type": "text", + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, volume 70, pages 1126–1135, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "text", + "content": "[14] Kuiliang Gao, Bing Liu, Xuchu Yu, and Anzhu Yu. Unsupervised meta learning with multiview constraints for hyperspectral image small sample set classification. IEEE TIP, 31:3449-3462, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "type": "text", + "content": "[15] Quanxue Gao, Wei Xia, Zhizhen Wan, De-Yan Xie, and Pu Zhang. Tensor-svd based graph learning for multi-view subspace clustering. In AAAI, pages 3930-3937, 2020. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "text", + "content": "[16] Yu Geng, Zongbo Han, Changqing Zhang, and Qinghua Hu. Uncertainty-aware multi-view representation learning. In AAAI, pages 7545-7553, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[17] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In AISTATS, volume 15, pages 315-323, 2011. 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[18] Kaveh Hassani and Amir Hosein Khas Ahmadi. Contrastive multi-view representation learning on graphs. In ICML, volume 119, pages 4116-4126, 2020. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "text", + "content": "[19] Timothy M. Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE TPAMI, 44(9):5149-5169, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 143, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 547, + 174 + ], + "type": "text", + "content": "[20] Junlin Hu, Jiwen Lu, and Yap-Peng Tan. Sharable and individual multi-view metric learning. IEEE TPAMI, 40(9):2281-2288, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 177, + 547, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 547, + 209 + ], + "type": "text", + "content": "[21] Mike Huisman, Jan N. van Rijn, and Aske Plaat. A survey of deep meta-learning. Artif. Intell. Rev., 54(6):4483-4541, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 212, + 547, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 547, + 255 + ], + "type": "text", + "content": "[22] Xiaodong Jia, Xiao-Yuan Jing, Xiaoke Zhu, Songcan Chen, Bo Du, Ziyun Cai, Zhenyu He, and Dong Yue. Semi-supervised multi-view deep discriminant representation learning. IEEE TPAMI, 43(7):2496-2509, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 258, + 547, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 258, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 547, + 300 + ], + "type": "text", + "content": "[23] Yuheng Jia, Hui Liu, Junhui Hou, Sam Kwong, and Qingfu Zhang. Multi-view spectral clustering tailored tensor low-rank representation. IEEE TCSVT, 31(12):4784-4797, 2021. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 303, + 547, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 547, + 346 + ], + "type": "text", + "content": "[24] Xiao-Yuan Jing, Fei Wu, Xiwei Dong, Shiguang Shan, and Songcan Chen. Semi-supervised multi-view correlation feature learning with application to webpage classification. In AAAI, pages 1374-1381, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 349, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 547, + 380 + ], + "type": "text", + "content": "[25] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. IEEE TPAMI, 39(4):664-676, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 384, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 547, + 415 + ], + "type": "text", + "content": "[26] Yingming Li, Ming Yang, and Zhongfei Zhang. A survey of multi-view representation learning. IEEE TKDE, 31(10):1863-1883, 2019. 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 418, + 547, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 547, + 460 + ], + "type": "text", + "content": "[27] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE TPAMI, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 463, + 547, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 547, + 507 + ], + "type": "text", + "content": "[28] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In ICCV, pages 3295-3304, 2019. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 509, + 547, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 547, + 552 + ], + "type": "text", + "content": "[29] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In CVPR, pages 3623–3632, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 555, + 547, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 547, + 587 + ], + "type": "text", + "content": "[30] Shirui Luo, Changqing Zhang, Wei Zhang, and Xiaochun Cao. Consistent and specific multi-view subspace clustering. In AAAI, pages 3730-3737, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 590, + 547, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 622 + ], + "type": "text", + "content": "[31] Yao Ma, Shilin Zhao, Weixiao Wang, Yaoman Li, and Irwin King. Multimodality in meta-learning: A comprehensive survey. KBS, 250:108976, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 624, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 547, + 667 + ], + "type": "text", + "content": "[32] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In CVPR, pages 9826–9836, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 670, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 712 + ], + "type": "text", + "content": "[33] Feiping Nie, Guohao Cai, Jing Li, and Xuelong Li. Autoweighted multi-view learning for image clustering and semi-supervised classification. IEEE TIP, 27(3):1501-1511, 2018. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11598" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[34] Xiushan Nie, Weizhen Jing, Chaoran Cui, Chen Jason Zhang, Lei Zhu, and Yilong Yin. Joint multi-view hashing for large-scale near-duplicate video retrieval. IEEE TKDE, 32(10):1951-1965, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "text", + "content": "[35] Xiushan Nie, Xingbo Liu, Jie Guo, Letian Wang, and Yilong Yin. Supervised discrete multiple-length hashing for image retrieval. IEEE TBD, 9(1):312-327, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 288, + 206 + ], + "type": "text", + "content": "[36] Zheyun Qin, Xiankai Lu, Xiushan Nie, Dongfang Liu, Yi-long Yin, and Wenguan Wang. Coarse-to-fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica, 10(5):1-17, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 243 + ], + "type": "text", + "content": "[37] Mathieu Salzmann, Carl Henrik Ek, Raquel Urtasun, and Trevor Darrell. Factorized orthogonal latent spaces. In AISTATS, volume 9, pages 701-708, 2010. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "type": "text", + "content": "[38] Yang Shi, Xiushan Nie, Xingbo Liu, Lu Yang, and Yilong Yin. Zero-shot hashing via asymmetric ratio similarity matrix. IEEE TKDE, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 288, + 312 + ], + "type": "text", + "content": "[39] Yang Shi, Xiushan Nie, Xingbo Liu, Li Zou, and Yilong Yin. Supervised adaptive similarity matrix hashing. IEEE TIP, 31:2755-2766, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 314, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 288, + 357 + ], + "type": "text", + "content": "[40] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pages 1917-1928, 2019. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 359, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 288, + 392 + ], + "type": "text", + "content": "[41] Haoliang Sun, Chenhui Guo, Qi Wei, Zhongyi Han, and Yilong Yin. Learning to rectify for robust learning with noisy labels. PR, 124:108467, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 394, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 437 + ], + "type": "text", + "content": "[42] Kai Sun, Jiangshe Zhang, Junmin Liu, Ruixuan Yu, and Zengjie Song. DRCNN: dynamic routing convolutional neural network for multi-view 3d object recognition. IEEE TIP, 30:868-877, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 440, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 288, + 472 + ], + "type": "text", + "content": "[43] Shiliang Sun, Wenbo Dong, and Qiuyang Liu. Multi-view representation learning with deep gaussian processes. IEEE TPAMI, 43(12):4453-4468, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 474, + 288, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 288, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 288, + 528 + ], + "type": "text", + "content": "[44] Zhongkai Sun, Prathusha Kameswara Sarma, William A. Sethares, and Yingyu Liang. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In AAAI, pages 8992-8999, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 531, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 288, + 563 + ], + "type": "text", + "content": "[45] Risto Vuorio, Shao-Hua Sun, Hexiang Hu, and Joseph J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. In NeurIPS, pages 1–12, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 566, + 288, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 288, + 608 + ], + "type": "text", + "content": "[46] Shiye Wang, Changsheng Li, Yanming Li, Ye Yuan, and Guoren Wang. Self-supervised information bottleneck for deep multi-view subspace clustering. CoRR, abs/2204.12496, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "type": "text", + "content": "[47] Weiran Wang, Raman Arora, Karen Livescu, and Jeff A. Bilmes. On deep multi-view representation learning. In ICML, volume 37, pages 1083-1092, 2015. 2, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "type": "text", + "content": "[48] Qi Wei, Lei Feng, Haoliang Sun, Ren Wang, Chenhui Guo, and Yilong Yin. Fine-grained classification with noisy labels. arXiv preprint arXiv:2303.02404, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[49] Qi Wei, Haoliang Sun, Xiankai Lu, and Yilong Yin. Self-filtering: A noise-aware sample selection for label noise with confidence penalization. In ECCV, pages 516-532, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 623 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[50] Jie Xu, Yazhou Ren, Huayi Tang, Zhimeng Yang, Lili Pan, Yang Yang, and Xiaorong Pu. Self-supervised discriminative feature learning for multi-view clustering. CoRR, abs/2103.15069, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[51] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In CVPR, pages 16030-16039, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 196 + ], + "type": "text", + "content": "[52] Fei Xue, Xin Wu, Shaojun Cai, and Junqiu Wang. Learning multi-view camera relocalization with graph neural networks. In CVPR, pages 11372-11381, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 198, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 239 + ], + "type": "text", + "content": "[53] Xiaowei Xue, Feiping Nie, Sen Wang, Xiaojun Chang, Bela Stantic, and Min Yao. Multi-view correlated feature learning by uncovering shared component. In AAAI, pages 2810-2816, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 274 + ], + "type": "text", + "content": "[54] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "type": "text", + "content": "[55] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, pages 8824-8833, 2022. 2, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 320, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 547, + 354 + ], + "type": "text", + "content": "[56] Xiao Yu, Hui Liu, Yuxiu Lin, Yan Wu, and Caiming Zhang. Auto-weighted sample-level fusion with anchors for incomplete multi-view clustering. PR, 130:108772, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 547, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 398 + ], + "type": "text", + "content": "[57] Yun-Hao Yuan, Jin Li, Yun Li, Jipeng Qiang, Yi Zhu, Xiaobo Shen, and Jianping Gou. Learning canonical f-correlation projection for compact multiview representation. In CVPR, pages 19238-19247, 2022. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 547, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 441 + ], + "type": "text", + "content": "[58] Changqing Zhang, Huazhu Fu, Qinghua Hu, Xiaochun Cao, Yuan Xie, Dacheng Tao, and Dong Xu. Generalized latent multi-view subspace clustering. IEEE TPAMI, 42(1):86-99, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 547, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 547, + 477 + ], + "type": "text", + "content": "[59] Changqing Zhang, Zongbo Han, Yajie Cui, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Cpm-nets: Cross partial multi-view networks. In NeurIPS, pages 557-567, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 478, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 547, + 510 + ], + "type": "text", + "content": "[60] Changqing Zhang, Yeqing Liu, and Huazhu Fu. Ae2-nets: Autoencoder in autoencoder networks. In CVPR, pages 2577-2585, 2019. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 511, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 547, + 555 + ], + "type": "text", + "content": "[61] Liang Zhao, Tao Yang, Jie Zhang, Zhikui Chen, Yi Yang, and Z. Jane Wang. Co-learning non-negative correlated and uncorrelated features for multi-view data. IEEE TNNLS, 32(4):1486-1496, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 556, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 547, + 590 + ], + "type": "text", + "content": "[62] Wenqing Zheng, Tianlong Chen, Ting-Kuei Hu, and Zhangyang Wang. Symbolic learning to optimize: Towards interpretability and scalability. In ICLR, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 590, + 547, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 623 + ], + "type": "text", + "content": "[63] Pengfei Zhu, Binyuan Hui, Changqing Zhang, Dawei Du, Longyin Wen, and Qinghua Hu. Multi-view deep subspace clustering networks. CoRR, abs/1908.01978, 2019. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11599" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_content_list.json b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fec8b5703190ac22c0d165591969e8ea5dfc9790 --- /dev/null +++ b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_content_list.json @@ -0,0 +1,1510 @@ +[ + { + "type": "text", + "text": "Metadata-Based RAW Reconstruction via Implicit Neural Functions", + "text_level": 1, + "bbox": [ + 140, + 130, + 828, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leyi Li $^{1}$ Huijie Qiao $^{2}$ Qi Ye $^{1,3}$ Qinmin Yang $^{1}$ $^{1}$ Zhejiang University $^{2}$ Chinese Academy of Sciences $^{3}$ Key Lab of CS&AUS of Zhejiang Province $\\{lileyi, qi.ye, qmyang\\} @zju.edu.cn$ qiaohj@ioz.ac.cn", + "bbox": [ + 81, + 178, + 879, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many low-level computer vision tasks are desirable to utilize the unprocessed RAW image as input, which remains the linear relationship between pixel values and scene radiance. Recent works advocate to embed the RAW image samples into sRGB images at capture time, and reconstruct the RAW from sRGB by these metadata when needed. However, there still exist some limitations in making full use of the metadata. In this paper, instead of following the perspective of sRGB-to-RAW mapping, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values. With this novel formulation, we propose to reconstruct the RAW image with an implicit neural function, which achieves significant performance improvement (more than 10dB average PSNR) only with the uniform sampling. Compared with most deep learning-based approaches, our method is trained in a self-supervised way that requiring no pre-training on different camera ISPs. We perform further experiments to demonstrate the effectiveness of our method, and show that our framework is also suitable for the task of guided super-resolution.", + "bbox": [ + 76, + 301, + 473, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 647, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-level computer vision tasks benefit a lot from the scene-referred RAW images [7, 39, 19, 17, 16], which is rendered to the display-referred standard RGB (sRGB) images via camera image signal processors (ISPs). Compared with sRGB images, typical RAW images has the advantages of linear relationship between pixel values and scene radiance, as well as higher dynamic range. However, RAW images occupy obviously more memory than the sRGB images in common format like JPEG, which is unfavourable for transferring and sharing. Moreover, since most display and printing devices are designed for images stored and shared in sRGB format, it is inconvenient to directly replace sRGB with RAW. Consequently, mapping sRGB images back to their RAW counterparts, which is also called RAW reconstruction, is regarded as the appropriate way to", + "bbox": [ + 76, + 674, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31a015de6f98ea740612b792ad93ecb9cf0bd51955312dbb18e2bbe775b307e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 267, + 625, + 339 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c8d4334f820bffa702019a34bfb31ca4a28364fce885a43e7eee9bb4add11661.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 267, + 750, + 339 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2e9c957293d37b991170c1821173687f873d0b839aac09bb86ac70703554e09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 267, + 874, + 339 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fca0e4de21f1902af753f138645bcb5eaf85b9d422cce49a476af77cd3769bcf.jpg", + "image_caption": [ + "sRGB (input)", + "RAW (GT)", + "Figure 1. As RAW images are beneficial to many low-level computer vision tasks, we aim to reconstruct the RAW image from the corresponding sRGB image with the assistance of extra metadata. In this figure, the reconstructed RAW images are visualized through error maps. As can be seen, our method remarkably outperforms other related methods with the improvement of more than $10\\mathrm{dB}$ PSNR. We owe this performance boost to the effectiveness of our proposed implicit neural function (INF)." + ], + "image_footnote": [], + "bbox": [ + 504, + 357, + 627, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d90cb0ea54e8c45ed93a019b6dc10d41143a66a3c0986585e85c1396252d4f5c.jpg", + "image_caption": [ + "RIR [25]", + "CAM [24]" + ], + "image_footnote": [], + "bbox": [ + 627, + 357, + 750, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4ec6ff7eefc56932a3671d30aafe47cae63c53e79063eb37c652587f8031fd0.jpg", + "image_caption": [ + "SAM [28]", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 751, + 357, + 874, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "utilize the advantage of RAW data [23, 25, 36, 10, 28, 24].", + "bbox": [ + 500, + 597, + 887, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early RAW reconstruction methods focus on building standard models to reverse ISPs, which is parameterized by either explicit functions [4, 18, 14, 5] or neural networks [23, 36, 10]. However, these approaches are faced with the same issue that a parameterized model is only suitable for a specific ISP. Meanwhile, a series of methods [25, 26, 28, 24] propose to overcome this problem by embedding extra metadata into sRGB images at capture time. For such methods, the main challenge is to improve the accuracy with lower metadata generation cost. RIR [25] implements complex optimization algorithm to estimate the global mapping parameters as metadata, but suffers high computational cost. SAM [28] adopts a uniform sampling on RAW images to generate the metadata, which is further replaced with a sampler network by CAM [24].", + "bbox": [ + 496, + 613, + 892, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For the metadata-based methods of SAM [28] and CAM [24], the embedded RAW samples stores partial information of ISPs which helps to reconstruct the RAW images better; also, by conditioning the reconstruction algorithm on", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18196", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these metadata, the recovery of the RAW data turns into a conditional mapping function instead of a function fitted to a specific case, enabling the potential to achieve better generalization. Therefore, we adopt this strategy in the paper.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the progress that SAM [28] and CAM [24] have made, there still exist some limitations for the metadata-based RAW reconstruction methods. SAM utilizes RBF interpolation [3], the main idea of which is to calculate the difference between sampling and target points by a kernel function. However, a fixed kernel function lacks the flexibility to model various sRGB-to-RAW mappings. CAM directly uses a neural network for reconstruction but requires pre-training on pairs of sRGB and raw data from different types of ISPs. Also, we observe that the results of these methods fail to recover the saturated regions [26] (i.e., pixels with any channel value close to the maximum), as is shown in Figure 1.", + "bbox": [ + 75, + 152, + 470, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the limitation, we propose a two-way RAW reconstruction algorithm based on an implicit neural function (INF). Previously, RAW reconstruction is formulated as mapping a sRGB image and the metadata to its RAW image. In this paper, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values, i.e. an implicit function. With this novel formulation, we can also decompose the problem into two aspects: a mapping function from the sRGB values to the corresponding raw values; a super-resolution function to interpolate the RAW image from the sparse samples. We observe that the super-resolution part usually exhibits much higher errors, indicating the latter a more challenging task. Accordingly, two branches are designed for each task inside an implicit neural network and the hyper-parameters for these branches are tuned to accommodate the difficulty of the tasks. Also, notice that with this formulation, the network can be trained in a self-supervised way, without the need of corresponding RAW images.", + "bbox": [ + 75, + 347, + 470, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution can be summarized as follows:", + "bbox": [ + 96, + 635, + 413, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reformulate the RAW reconstruction problem as a RAW image approximation problem that learns the 2D-to-RAW mapping of image coordinates to RAW values conditioned on its sRGB image.", + "- We decompose the reconstruction into two aspects and design the implicit neural network accordingly.", + "- We conduct extensive experiments on different cameras and demonstrate our algorithm outperforms existing work significantly." + ], + "bbox": [ + 96, + 657, + 467, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 825, + 218, + 842 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Blind RAW reconstruction. Early works of RAW reconstruction were blindly taken sRGB images as input without extra metadata. Since the processing stages of modern ISPs", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are designed more complicated, more complex models are proposed [4, 18, 14, 5]. However, these methods suffer the inconvenient calibration procedures that need to be repeated on each camera or even each camera setting. Deep learning-based methods (e.g., [23, 36]), similarly, are faced with the parallel issues of camera-specific models, which required abundant training data captured for each camera. Recent methods that simulate ISP architectures by assuming a classical set of ISP operations, such as [37, 10], cannot handle different camera settings as they take fixed parameters (e.g., gamma correction) and ISP length as a priori knowledge.", + "bbox": [ + 496, + 90, + 890, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RAW reconstruction with metadata. Compared with blind RAW reconstruction, a series of recent methods [25, 26, 28, 24] benefit from additional metadata embedded into sRGB images at capture time. Nguyen and Brown [25, 26] propose to extract and store the necessary parameters for recovering a RAW image from the sRGB counterpart. These parameters model the specific sRGB-to-RAW mapping and are restricted to a small memory (e.g., 128KB). However, their main algorithm deployed on the device has high computational cost and only considers the global tone mapping. Punnappurath and Brown [28], on the contrary, implement a uniform sampling on RAW images to save as metadata. They employ radial basis function (RBF) interpolation using sRGB pixel values and coordinates as input to reconstruct the RAW values. Nevertheless, the fixed type of RBF interpolation does not take full use of the information of sampled data. A further work [24] improves the effectiveness of sampling and recovering by taking a U-Net architecture [29] as both sampler and reconstruction networks, but their method requires to run a deep neural network on the device, which would lead to high computational cost and additional memory cost to save the pixel positions into metadata. We follow the uniform sampling in [28], but improve the reconstruction performance by INF.", + "bbox": [ + 496, + 261, + 890, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit neural representation. Implicit neural representation (INR) has recently been introduced to represent 2D images and 3D objects using coordinate-based multi-layer perceptron (MLP). To overcome the problem that conventional MLPs are incapable of representing high-frequency details of signals, two methods have been proposed. Sitzmann et al. [31] introduce SIREN, which replace ReLU activation with periodic activation (e.g., sine function). They demonstrate that the representation power of SIREN comes from the derivation invariance of sine function, and provide a number of potential applications and future works. A concurrent work [32] leverages random Fourier feature mapping on input coordinates to enable the MLP with ReLU to learn high-frequency details. It indicates that a Fourier feature mapping can beneficially address the spectral bias of a conventional MLP. Based on these breakthroughs, INR has been successfully adopted in various tasks [8, 1, 6, 35]. In this work, the proposed INF is based on [31], which shows", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18197", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "remarkably improvement in reconstruction accuracy.", + "bbox": [ + 76, + 90, + 426, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Guided super-resolution. Another topic similar to RAW reconstruction with metadata is guided image superresolution, which aims to convert a low-resolution (LR) source image to the high-resolution (HR) target with a guided HR image [20, 30, 33]. The difference of these two topics is that RAW reconstruction has obvious mapping of pixel values between two HR images. On this aspect, PixTransform [20] is more related to our method. It uses an MLP with ReLU to learn the pixel-to-pixel mapping of guided HR image to target HR image, where the LR image is treated like the metadata in RAW reconstruction. Our method differs in the MLP layer, where we use SIREN layer [31] to enhance the expressive power [38].", + "bbox": [ + 75, + 109, + 468, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 316, + 168, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation with Implicit Function", + "text_level": 1, + "bbox": [ + 76, + 342, + 457, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $S$ and $R$ denote an sRGB image and the corresponding RAW image, respectively. Since $R$ is converted to $S$ through a series of operations in the camera ISP, classical methods aim to build a model $g$ to map $S$ back to $R$ , i.e., $R = g(S)$ . However, the model $g$ is specific to a camera ISP or even a set of ISP parameters. Metadata-based methods introduce extra data $M$ for RAW reconstruction, i.e., $R = f(S, M)$ . Here $f$ is supposed to be a general function, and $M$ includes the image-specific information generated at capture time, which is commonly implemented by sampling on the RAW images for low computation cost. Hence $M = \\{\\mathbf{p}_i, \\mathbf{r}_i\\}_{i=0}^N$ , where $\\mathbf{p}_i = (x_i, y_i)$ is the coordinate of a sampled RAW pixel and $\\mathbf{r}_i = (r_{R_i}, r_{G_i}, r_{B_i})$ its RAW value. $N$ represents the number of sampled RAW pixels and $i$ refers to the $i^{th}$ sampling one.", + "bbox": [ + 75, + 366, + 468, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to previous work that establishes a mapping from RGB to RGB values, $f:(r,s)\\to r$ , we reformulate the problem as a mapping from the 2D coordinates to the RAW values conditioning on the sRGB value", + "bbox": [ + 75, + 593, + 468, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf: (p; s) \\rightarrow r, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 661, + 468, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $r, s, p$ denote the variables for the RAW value, the RGB value and coordinates. They can either represent the value for a pixel or a patch, or a set of coordinates.", + "bbox": [ + 75, + 685, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, for the sRGB image, we only use the the sRGB value $(s_{R_i}, s_{G_i}, s_{B_i})$ at $\\mathbf{p}_i$ , queried by $S(\\mathbf{p}_i)$ . Therefore we aim to learn an implicit neural function $f_\\theta$ taking a coordinate $\\mathbf{p}_i$ and its sRGB value and output $\\mathbf{r}_i$ .", + "bbox": [ + 75, + 731, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta}: \\left(\\mathbf {p} _ {i}; \\mathbf {s} _ {i}\\right)\\rightarrow \\mathbf {r} _ {i}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 800, + 468, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn the function, we define the loss function below to find the best configuration for the parameters $\\theta$ of $f$", + "bbox": [ + 76, + 824, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\| f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {s} _ {i}, \\mathbf {p} _ {i}\\right) - \\mathbf {r} _ {i} \\|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 862, + 468, + 904 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/044f91c70c87888b6d885fa68ded9dd79a496b72e01e155b78c5b7a8de67c3ba.jpg", + "image_caption": [ + "RAW (GT)" + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 632, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4145960c48f539469765404e92c981b40e8a5afd9c91ec4bbc9ed34f6c3d9a90.jpg", + "image_caption": [ + "(a) $f:\\mathbf{s}_i\\to \\mathbf{r}_i$" + ], + "image_footnote": [], + "bbox": [ + 633, + 89, + 758, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d2e7b2a0d1ae35636404deff0c8b44099a426ada025f67cffd6d0565f79858d.jpg", + "image_caption": [ + "(b) $f:\\mathbf{p}_i\\to \\mathbf{r}_i$" + ], + "image_footnote": [], + "bbox": [ + 761, + 89, + 885, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b81fb7584acee094625535f4ed1a6aebff0de0e8e4f40b5ec85659a747b454a2.jpg", + "image_caption": [ + "Figure 2. Illustration of the two aspects of reconstruction. It can be observed that the reconstruction from (a) pixel value mapping is much more accurate than (b) spatial super-resolution. Our method utilizes the information from both $\\mathbf{s}_i$ and $\\mathbf{p}_i$ , but give a different constraint to them to control their impacts.", + "Figure 3. The network structure of our proposed implicit neural function (INF). The inputs are divided into sRGB values and coordinates, as their branches are regularized with different parameters of weight decay. Each cube represents the output of each linear layer with sine activation proposed in [31]." + ], + "image_footnote": [], + "bbox": [ + 506, + 304, + 883, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of our method is shown in Figure 4. As mentioned above, considering the on-device computational cost, we follow SAM [28] to implement a uniform sampling on the RAW image to get the metadata $M$ . For the reconstruction stage, we implement the same sampling on the sRGB image, which is used to train the implicit neural function (INF) together with $M$ . The sRGB image is then processed by the trained INF to reconstruct the RAW image.", + "bbox": [ + 496, + 607, + 890, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Two-way Raw Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 739, + 766, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We note that Equation (2) contains the reconstruction from two aspects: (a) a pixel values mapping (i.e., $f:\\mathbf{s}_i\\to \\mathbf{r}_i$ ) and (b) a spatial super-resolution (i.e., $f:\\mathbf{p}_i\\rightarrow \\mathbf{r}_i$ ). Since $M$ will cost additional memory, the sampling rate is required to be constrained (e.g., no more than 1.5%), which limits the accuracy of (b). On the contrary, the sRGB-to-RAW mapping is piece-wise smooth [27], hence only such small amount of samples is able to model the whole mapping function theoretically. We provide an example in Fig-", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18198", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/36658c6332cb8f442cdc43c5d757fc25a22f9ff8e998466eb5a247db73c24c1d.jpg", + "image_caption": [ + "Figure 4. Overview of our proposed RAW reconstruction method. At capture stage, we follow [28] to implement an uniform sampling on the RAW image to generate the metadata, which does not require to store the pixel positions. The metadata is then embedded into the sRGB image for storage and transmission. At reconstruction stage, the metadata together with corresponding sRGB samples are used to train the implicit neural functions (INFs), which then convert the sRGB image to the RAW. Compared with the state-of-the-art metadata-based RAW reconstruction approach [24], our method achieves both lower on-device computational cost and much higher reconstruction accuracy." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 885, + 287 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/373b0b37cca2ffd48117d60a8bc3c977db2b78c6462083509d144e255e2f2e67.jpg", + "image_caption": [ + "Figure 5. Illustration of the patch-specific INF. Inspired by [28], we split the image into patches, and train the INF for each patch with its neighbours to make it robust to the local mapping. We further discuss the impact of patch-size in the ablation study." + ], + "image_footnote": [], + "bbox": [ + 84, + 401, + 460, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ure 2 to intuitively illustrate this phenomenon. Therefore (a) is supposed to be more significant than (b) in $f$ .", + "bbox": [ + 75, + 681, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the analysis of the sub-tasks, we design an implicit neural network consisting of two branches as shown in Figure 3. INF is an MLP structure built with SIREN layers [31], where the inputs are separated into two branches and then concatenated into the output branch. Each branch contains 4 layers, with 256 channels for the hidden layer. The output channels of two input branches are reduced to half, which are then concatenated into the output branch. Here $\\theta$ refers to the network weights, and Equation (4) can then be solved by training the INF.", + "bbox": [ + 75, + 715, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Also, considering the difficulty of the tasks, we introduce regularisation parameters to Equation (3) to control", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the complexity of different branches. That is,", + "bbox": [ + 500, + 398, + 799, + 414 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\theta}} = \\underset {\\boldsymbol {\\theta}} {\\arg \\min } \\sum_ {i = 1} ^ {N} \\| f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {s} _ {i}, \\mathbf {p} _ {i}\\right) - \\mathbf {r} _ {i} \\| + \\lambda_ {c} \\| \\boldsymbol {\\theta} _ {c} \\| ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 421, + 890, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c \\in \\{\\mathbf{s}, \\mathbf{p}, \\mathbf{r}\\}$ and $\\lambda_c$ is a hyper-parameter to control the strength of regularisation for different parts of $M$ .", + "bbox": [ + 498, + 470, + 890, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We note that the structure of INF is most similar to the network of [20]. However, instead of merging the two input branches by adding their output, we first reduce the output size of these two branches and then concatenate them as the next input. This is because the output of SIREN Layers $X_{l}$ is distributed in $X_{l} \\sim \\operatorname{Arcsin}(-1,1)$ [31], and adding the outputs would let the output branch treat them equally. Hence we use concatenation to enable the network to learn different weights for the outputs of $\\mathbf{p}_i$ and $\\mathbf{s}_i$ branches.", + "bbox": [ + 498, + 501, + 892, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction conditioned on sRGB patches. Additionally, it is worth noting that local tone mapping is commonly applied as a favourable part of the ISP pipeline, which is adopted on local images to enhance the special objects (e.g., food, flower, sky). In order to make our method robust to local tone mapping, we split the input image into patches, and then train the INF for each patch. Moreover, inspired by [28], considering that the mapping of each patch and its neighbours are supposed to have strong relationship, we use the samples from each patch together with its eight neighbour patches to train the patch-specific INF. Therefore the mapping in Eq. (2) becomes", + "bbox": [ + 496, + 641, + 893, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta}: \\left(\\mathbf {p} _ {i}; \\{s _ {j} \\} _ {j \\in \\mathcal {N} (i)}\\right)\\rightarrow \\mathbf {r} _ {i}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 830, + 890, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(i)$ represents neighbour patches of current pixel $\\mathbf{p}_i$ . Figure 5 illustrates the training strategy of INF for such patch-wise RAW reconstruction.", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18199", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 90, + 209, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 114, + 266, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We take three metadata-based RAW reconstruction methods for comparison: RIR [25], SAM [28] and CAM [24]. The RIR method computes the parameters of global sRGB-to-RAW mapping as metadata. The SAM and CAM methods adopt different sampling approaches on RAW images to generate metadata. We follow the uniform sampling as SAM [28], and show that it is enough to get much higher performance using uniform samples with the proposed method. Note that neither RIR nor SAM make their source codes publicly available, and only CAM reports their results reproduced on its own dataset. Therefore we only give a comparison with RIR and SAM on the dataset of CAM by inheriting its results in Table 1 and Figure 6.", + "bbox": [ + 75, + 142, + 468, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We deploy the same dataset of the previous work [24] to test our proposed method, which contains the preprocessed [13] and downsampled (with a downsampling factor of 4) images of three cameras from the NUS dataset [9]—Samsung NX2000, Olympus E-PL6 and Sony SLT-A57. Note that as our method does not require any pretraining step, the training and validation sets are not used in our method. Besides, since our method is able to directly process full-resolution images, we also compare it with CAM [24] on the original NUS dataset [9]. Additionally, as the NUS dataset [9] includes JPEG-compressed sRGB images rendered by real camera ISPs, we also conduct the experiments on these RAW-JPEG image pairs. These results are reported in the supplementary.", + "bbox": [ + 75, + 343, + 470, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. For capture stage, we use the same sampling rate of $1.5\\%$ like [24] in all experiments for fair comparison with other metadata-based methods. With respect to reconstruction stage, we split the image into different patches—(228, 272) for Samsung NX2000, (216, 288) for Olympus E-PL6 and (204, 304) for Sony SLT-A57—to make the resolution divisible. The proposed INF is trained by Adam optimizer [15] with $L_{2}$ loss in 500 iterations. We set the initial learning rate to 0.0001, and reduce it to half every 200 iterations. The regularisation parameters are set to $\\lambda_{\\mathrm{s}} = 0.0001, \\lambda_{\\mathrm{p}} = 0.1, \\lambda_{\\mathrm{r}} = 0.001$ to produce the final results. It is worth noting that the reconstruction network is only trained for each specific patch, and will be re-initialized for the training of other patches.", + "bbox": [ + 75, + 559, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 76, + 782, + 171, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results. We report the quantitative comparison results against other three baselines in Table 1. It can be observed that our method makes a remarkably improvement in both PSNR and SSIM metrics, especially outperforms the closest competitor CAM [24] by more than $10\\mathrm{dB}$ in average PSNR. This indicates the effectiveness of INF to", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "modeling the sRGB-to-RAW mapping.", + "bbox": [ + 500, + 90, + 759, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative results. We also provide examples of the qualitative comparison in Figure 6, which gives an intuitive explanation of the success of our method. As can be seen from the error maps, our method significantly reduces the reconstructed error on the whole images, especially the saturated regions [26]. We note that recovering saturated regions is the most difficult problem for all RAW reconstruction methods, and is the only case of our method getting large errors $(\\geqslant 1\\%)$ . We attribute this to the expressive power [38] of INF, which is further discussed in Section 4.4.", + "bbox": [ + 496, + 109, + 893, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Ablation study.", + "text_level": 1, + "bbox": [ + 500, + 272, + 651, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RBF vs. INF. We first compare RBF and INF with different input, which is shown in Table 2. As can be seen, the proposed INF outperforms the RBF interpolation [28] whatever the input is. We attribute the main performance boost to the basic structure, i.e. SIREN [31], as INF is able to modelling the sRGB-to-RAW mapping more accurately (59.62dB) only with the pixel value as input. Moreover, compared with RBF, our proposed INF can make further improvement by splitting the input into separate branches and regularize them with different weights.", + "bbox": [ + 496, + 299, + 892, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The structure of INF. We also conduct experiments on the structure of INF, which is reported in Figure 7. We compare the different activation functions and the layer location where two input branches are merged. It is obviously shown that the INF with sine activation remarkably outperforms the INF with ReLU function, which can be attributed to the expressive power of INRs [38]. As to the location of fusion layer, we find there is no significant difference except the final layer for the INF with sine activation, since it directly maps the merged vector to the output by a fully connection layer without activation. Therefore, we merge the two branches at fourth layer to balance the complexity of each branch1. On the contrary, due to lack of such expressive power, the INF with ReLU requires more layers to model the mapping. Hence the performance increases when the input branch goes deep.", + "bbox": [ + 496, + 454, + 893, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regularisation parameters. We further discuss the influence of key setups to INF, which mainly focuses on the regularisation parameters and patch size. As is shown in Table 3, it is obvious that the regularisation parameters play a vital role in the final accuracy. Here we use the grid search with a commonly-used tuner TPE [2] (implemented by [22]) in a grid of $[1, 10^{-1}, \\dots, 10^{-5}]$ to determine each parameter. In our experiments, we find the regulation parameters are mainly related to the complexity of camera ISP rather than data dependent, which means we only need", + "bbox": [ + 496, + 700, + 893, + 852 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1In fact, we find the fusion at third layer would achieve a slightly higher performance, but which can be ignored compared with the instability of random initialization. See more detail in supplementary material.", + "bbox": [ + 500, + 862, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18200", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5a932bb0e86d9d819e33d1bb53fce77df64092e0e33a53a819c49e6d231deee7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodSamsung NX2000Olympus E-PL6Sony SLT-A57
PSNRSSIMPSNRSSIMPSNRSSIM
CAM [24]RIR [25]45.660.993948.420.992451.260.9982
SAM [28]47.030.996249.350.997850.440.9982
CAM [24]49.570.997551.540.998053.110.9985
Ours53.530.998556.710.999158.360.9992
NUS [9]CAM [24]49.510.993752.870.996153.340.9959
Ours61.480.999163.380.999363.800.9995
", + "bbox": [ + 116, + 88, + 854, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with [25, 28, 24]. Here CAM dataset [24] involves downsampling operation (with the factor of 4) on the original NUS dataset [9], which undermines the strength of local mapping. For CAM dataset, we conduct the experiments on their test set to inherit the results of [24]. For NUS dataset, we report the results tested on the whole dataset. The best score for each column is in bold.", + "bbox": [ + 75, + 255, + 890, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0a13b958cb61dd10a4256e3a44811f0b4337e77e3adfe1c74e8b205873a7213.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison with [25, 28, 24]. Each two rows from top to bottom are respectively selected from Olympus E-PL6, Samsung NX2000 and Sony SLT-A57. The GT RAW images are processed with a gamma function for better visualization, and the reconstructed RAW images are visualized through error maps. This figure is best viewed in the electronic version." + ], + "image_footnote": [], + "bbox": [ + 80, + 311, + 885, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to tune it for different ISPs, e.g., digital camera ISP, smart phone ISP, or software ISP. It also means we just need to", + "bbox": [ + 75, + 859, + 468, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "use one or quite a few images for tuning, resulting in a fast tuning process—within 2 hours on a single RTX3090. The", + "bbox": [ + 500, + 859, + 890, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18201", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/79f5a9ea9ee1d76fcabf9a588d5530750d173d2fafc779bbf8818d46a2055830.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInputPSNR
RBF [28]s29.10
RBF [28]p21.71
RBF [28](s, p)49.69
INFs59.62
INFp31.09
INF(s, p)60.91
INF{s; p}62.24
", + "bbox": [ + 125, + 88, + 421, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78a1b604ae63ab7d72b714c8b4adda9fd326c3e2baf78a155478f312dbe90725.jpg", + "image_caption": [ + "Figure 7. Ablation study on the structure of INF. The abscissa value represents the layer where we merge the two input branches, e.g., the branches are merged at fourth layer in Figure 3." + ], + "image_footnote": [], + "bbox": [ + 101, + 358, + 431, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tuned parameters can then be used for all scenes and all other images from the ISPs with similar complexity, hence we use the same parameters for the three reported cameras.", + "bbox": [ + 75, + 593, + 467, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 3, the proposed INF achieves the best performance at $\\lambda_{\\mathbf{s}} = 0.0001$ , $\\lambda_{\\mathbf{p}} = 0.1$ , $\\lambda_{\\mathbf{r}} = 0.001$ , which indicates the spatial branch is not supposed to be more complex. This is because the output $\\mathbf{r}$ has weaker relationship with the coordinates $\\mathbf{p}$ than sRGB values $\\mathbf{s}$ , and regularisation could help to control the contribution of $\\mathbf{s}$ and $\\mathbf{p}$ . Therefore when we set $\\lambda_{\\mathbf{p}} = 0$ , the reconstruction performance dramatically drops down due to inaccurate information from $\\mathbf{p}$ is expanded. We believe such intuitive relationship would be helpful to set the initial values for regularisation parameters. Besides, for different ISPs with similar complexity, the results are insensitive to the parameters in a wide range (for example, where all branches are regularized by $\\lambda \\geqslant 10^{-4}$ ). In this sense, the risk of over-fitting due to these hyper-parameter settings is very small.", + "bbox": [ + 75, + 638, + 467, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of patch size. Moreover, we find that different size of the patch would lead to an obvious accuracy gap. For the", + "bbox": [ + 75, + 869, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9521b458d119f69a609c27eba7004dfeeec32f96a19b6b54e06460af5db51cce.jpg", + "table_caption": [ + "Table 2. Ablation study on different reconstruction methods. The best score is in bold and the second in italic. $(\\mathbf{s},\\mathbf{p})$ refers to combine $\\mathbf{s}$ and $\\mathbf{p}$ into a single input vector, while $\\{\\mathbf{s};\\mathbf{p}\\}$ denotes separating them as different input. For the RBF interpolation, we use the linear RBF kernel as proposed in [28]." + ], + "table_footnote": [], + "table_body": "
Patch size(λs,λp,λr)PSNR
×147.48
×2(0,0,0)49.17
×450.53
×161.13
×2(0.0001,0.001,0.0001)60.38
×457.80
×162.58
×2(0.0001,0.1,0.001)62.24
×459.64
", + "bbox": [ + 514, + 88, + 879, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on the patch size and regularisation parameters. The best score is in bold and the second in italic. For the patch size, $\\times 2$ refers to the sizes we report in our main experiments, and the rescale factor of 1, 2, 4 corresponds to the length of patch sides. We note that the accuracy gap less than $0.5\\mathrm{dB}$ can not represent an improvement, which may be caused by the different calculations of PSNR and the random initialization of INF. This is further discussed in our supplementary material.", + "bbox": [ + 496, + 291, + 890, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "case of all regularisation parameters set to 0, we find small patch would lead to accuracy reduction. When the regularisation is correctly set, reconstruction on a small patch is more accurate than on a large one. This can be explained by the various local mapping functions in different image regions, which means modelling the local mapping for a small patch would be more accurate than a larger one. Note that though reducing the patch size would be helpful to the reconstruction accuracy, it would take much more time to recover an image. Therefore, we compromise to use the patch size around 200-300 in the main experiments.", + "bbox": [ + 496, + 429, + 890, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Discussion", + "text_level": 1, + "bbox": [ + 500, + 604, + 617, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Why INF works in sRGB-to-RAW mapping. We first rethink the RBF interpolation [3] used in SAM [28], by which the sRGB-to-RAW mapping is built in the following form:", + "bbox": [ + 496, + 630, + 890, + 676 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = \\mathbf {M} ^ {\\text {t e s t}} \\mathbf {M} ^ {- 1} y \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 685, + 890, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{M}$ denotes the RBF interpolation matrix with linear kernel function [27], and test refers to the pixels that need to be reconstructed. As discussed in [32], the output of an MLP can be approximated as:", + "bbox": [ + 496, + 712, + 890, + 773 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} ^ {(t)} = \\mathbf {K} ^ {\\text {t e s t}} \\mathbf {K} ^ {- 1} \\left(\\mathbf {I} - e ^ {- \\eta \\mathbf {K} t}\\right) y \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 781, + 890, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{K}$ denotes the NTK matrix [12] and $t$ refers to training iterations. In our experiments, an MLP (typical with ReLU activation) tends to behave like RBF when $t$ is large enough (e.g., $t > 1000$ ), which meets the results of Equation (6) and Equation (7). The reason why RBF interpolation works can be attributed to that sRGB-RAW mapping", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18202", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5733ff64a408211ea0e05a940d066f65243b416966bef20b38336d4fabe723d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method×4×8×16
Bicubic2.624.196.68
PixTransform [20]3.513.944.94
INF2.503.065.23
INF with ReLU2.392.693.57
", + "bbox": [ + 102, + 88, + 444, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative comparison on guided depth superresolution. We use the Middlebury dataset provided by [33] for testing. The results are reported in terms of average RMSE.", + "bbox": [ + 75, + 210, + 468, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is piecewise smooth [27], but there still exists errors when using a linear kernel function, which is also suffered by MLPs with ReLU. On the other hand, Sitzmann et al. [31] demonstrate that MLPs with periodic activation functions (i.e., SIRENs) are more suitable for representing complicated signals and their derivatives compared with the typical MLP architectures. SIRENs have the advantage of representing the derivatives of the signal [31], which is more helpful to reconstruct fine details. Therefore, recovering RAW images by SIRENs can effectively improve the performance, which is also proved by Table 2.", + "bbox": [ + 75, + 281, + 467, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future research. One of the unresolved problem of our method is that the regularisation parameters of weight decay require pre-experiments to determine. These three parameters play a significant role to control the final result by limiting the complexity of network. In our experiments, we find the parameters are strongly related to the correlation of input and output data. For example, compared with the coordinates, sRGB values have a more obvious relationship with the RAW values, where the regularisation parameters are respectively set to $\\lambda_{p} = 0.1$ and $\\lambda_{s} = 0.0001$ . Such correlation is suitable for modelling with another neural network, and would lead to the performance improvement with optimized parameters. Therefore, learning the regularisation parameters from the input and output is supposed to be a valuable topic of future work.", + "bbox": [ + 75, + 452, + 467, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Applications", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Guided depth super-resolution As mentioned in Section 2, our framework is related to guided super-resolution, therefore we also provide an experiment to illustrate the applicability of our framework. Considering that our method is only self-supervised without any priori knowledge, we give a comparison with another self-supervised method PixTransform [20]. As shown in Table 4, INF with ReLU outperforms other methods, which indicates the ReLU activation is more suitable for this task than the sine activation. The reason for ReLU working better can be attributed to the sharp changes of the depth near the object edges, as ReLU is piecewise and discontinuous, which fits this depth disconti", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c4a8c3a74025fc666c754094d2dbf1e8474ebcf49fc335ea6b590a6ea2314ca.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 632, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1219412462c45f3b2453bf8d2b1ff29cb2970be19a8ef7d7a3ef3031c0e52b06.jpg", + "image_caption": [ + "Guide" + ], + "image_footnote": [], + "bbox": [ + 635, + 89, + 756, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57788384b97aebde3c91e7b824c0d49d6b3910a9ea1b979c37252e274f3c47fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 89, + 885, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0945424c2419cee1bcb58fd28cc0a1c6164cda3bd61709a9b1f2c1eef896ecb7.jpg", + "image_caption": [ + "PixTransform [20]" + ], + "image_footnote": [], + "bbox": [ + 506, + 204, + 632, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e61d52c2b98b2ff1e9d0b694935e7e866d83ef0a3e7cca024c36e421cd6e02ca.jpg", + "image_caption": [ + "INF" + ], + "image_footnote": [], + "bbox": [ + 635, + 204, + 756, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d549c9409a9663144e61a9a339cb1deb72c4f2709220ba774a0acbe089abd2b3.jpg", + "image_caption": [ + "Target", + "INF with ReLU", + "Figure 8. Qualitative comparison on guided depth super-resolution." + ], + "image_footnote": [], + "bbox": [ + 759, + 204, + 885, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nuity. In contrast, the discontinuity in RAW-to-sRGB mapping is moderate and the smooth sine activation function is a better option. We also provide qualitative comparison in Figure 8. which shows that the result of INF, though does not achieve higher accuracy, obtains better visual quality compared with the result of INF with ReLU.", + "bbox": [ + 496, + 388, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Low-light image enhancement. As a potential application of RAW reconstruction, we test our method on low-light image enhancement (LLIE) task. We find that the LLIE task can be simplified into a linear degradation problem (e.g., enlarging the pixel values by the same multiple) on the reconstructed RAW images, which achieves equal visual quality compared with the state-of-the-art deep learning-based approaches [11, 34, 21]. The results are reported in our supplementary. Note that such comparison is merely aimed to illustrate the effectiveness of executing low-light enhancement on reconstructed RAW images.", + "bbox": [ + 496, + 484, + 890, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 669, + 617, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a method for recovering the RAW image from the sRGB counterpart with assistance of additional metadata, which is sampled from the RAW image at capture time. We introduce the implicit neural function (INF) to remarkably improve the reconstruction accuracy (10 dB average PSNR) only with uniform sampling. We prove that the structure of INF is beneficial to merging the information from both pixel values and coordinates. Further experiments indicate that our framework is also suitable for the task of guided super-resolution.", + "bbox": [ + 496, + 696, + 890, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was partly supported by National Key R&D Program of China (Grant No. 2021YFD1400200) and National Natural Science Foundation of China (Grant No. 62088101).", + "bbox": [ + 500, + 856, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18203", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19323-19332, 2022. 2", + "[2] James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization. Advances in neural information processing systems, 24, 2011. 5", + "[3] Marco Evangelos Biancolini. Fast radial basis functions for engineering applications. Springer, 2017. 2, 7", + "[4] Ayan Chakrabarti, Daniel Scharstein, and Todd E Zickler. An empirical camera model for internet color vision. In BMVC, volume 1, page 4. CiteSeer, 2009. 1, 2", + "[5] Ayan Chakrabarti, Ying Xiong, Baochen Sun, Trevor Darrell, Daniel Scharstein, Todd Zickler, and Kate Saenko. Modeling radiometric uncertainty for vision with tonemapred color images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11):2185-2198, 2014. 1, 2", + "[6] Bowei Chen, Tiancheng Zhi, Martial Hebert, and Srinivasa G Narasimhan. Learning continuous implicit representation for near-periodic patterns. arXiv preprint arXiv:2208.12278, 2022. 2", + "[7] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 1", + "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2", + "[9] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 5, 6", + "[10] Marcos V Conde, Steven McDonagh, Matteo Maggioni, Ales Leonardis, and Eduardo Pérez-Pellitero. Model-based image signal processors via learnable dictionaries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 481-489, 2022. 1, 2", + "[11] Chunle Guo Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1780-1789, June 2020. 8", + "[12] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 7", + "[13] Hakki Can Karaimer and Michael S Brown. A software platform for manipulating the camera imaging pipeline. In European Conference on Computer Vision, pages 429-444. Springer, 2016. 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Seon Joo Kim, Hai Ting Lin, Zheng Lu, Sabine Süssstrunk, Stephen Lin, and Michael S Brown. A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(12):2289-2302, 2012. 1, 2", + "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, 2015. 5", + "[16] Shayan Kousha, Ali Maleky, Michael S Brown, and Marcus A Brubaker. Modeling srgb camera noise with normalizing flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17463-17471, 2022. 1", + "[17] Zhihao Li, Si Yi, and Zhan Ma. Rendering nighttime image via cascaded color and brightness compensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 897–905, 2022. 1", + "[18] Hai Ting Lin, Zheng Lu, Seon Joo Kim, and Michael S Brown. Nonuniform lattice regression for modeling the camera imaging pipeline. In European Conference on Computer Vision, pages 556-568. Springer, 2012. 1, 2", + "[19] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image hdr reconstruction by learning to reverse the camera pipeline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1651-1660, 2020. 1", + "[20] Riccardo de Lutio, Stefano D'aronco, Jan Dirk Wegner, and Konrad Schindler. Guided super-resolution as pixel-to-pixel transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8829-8837, 2019. 3, 4, 8", + "[21] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 8", + "[22] Microsoft. Neural Network Intelligence, 2021. 5", + "[23] Seonghyeon Nam and Seon Joo Kim. Modelling the scene dependent imaging in cameras with a deep neural network. In Proceedings of the IEEE International Conference on Computer Vision, pages 1717-1725, 2017. 1, 2", + "[24] Seonghyeon Nam, Abhijith Punnappurath, Marcus A Brubaker, and Michael S Brown. Learning srgb-to-raw-rgb de-rendering with content-aware metadata. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17704–17713, 2022. 1, 2, 4, 5, 6", + "[25] Rang MH Nguyen and Michael S Brown. Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1655-1663, 2016. 1, 2, 5, 6", + "[26] Rang MH Nguyen and Michael S Brown. Raw image reconstruction using a self-contained srgb-jpeg image with small memory overhead. International journal of computer vision, 126(6):637-650, 2018. 1, 2, 5", + "[27] Abhijith Punnappurath and Michael S Brown. Learning raw image reconstruction-aware deep image compressors. IEEE" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18204", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transactions on pattern analysis and machine intelligence, 42(4):1013-1019, 2019. 3, 7, 8", + "[28] Abhijith Punnappurath and Michael S Brown. Spatially aware metadata for raw reconstruction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 218-226, 2021. 1, 2, 3, 4, 5, 6, 7", + "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2", + "[30] Guy Shacht, Dov Danon, Sharon Fogel, and Daniel Cohen-Or. Single pair cross-modality super resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6378–6387, 2021. 3", + "[31] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2, 3, 4, 5, 8", + "[32] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 2, 7", + "[33] Jiaxiang Tang, Xiaokang Chen, and Gang Zeng. Joint implicit image function for guided depth super-resolution. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4390-4399, 2021. 3, 8", + "[34] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2604-2612, 2022. 8", + "[35] Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neuralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1526-1535, 2022. 2", + "[36] Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6287-6296, 2021. 1, 2", + "[37] Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248-4257, 2021. 2", + "[38] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 3, 5", + "[39] Xuaner Zhang, Qifeng Chen, Ren Ng, and Vladlen Koltun. Zoom to learn, learn to zoom. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3762-3770, 2019. 1" + ], + "bbox": [ + 78, + 92, + 468, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18205", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_model.json b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7a717f99fa2bfc5daa04e9622076b0783b8999fd --- /dev/null +++ b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.83, + 0.152 + ], + "angle": 0, + "content": "Metadata-Based RAW Reconstruction via Implicit Neural Functions" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.179, + 0.88, + 0.234 + ], + "angle": 0, + "content": "Leyi Li\\(^{1}\\) Huijie Qiao\\(^{2}\\) Qi Ye\\(^{1,3}\\) Qinmin Yang\\(^{1}\\) \n\\(^{1}\\)Zhejiang University \\(^{2}\\)Chinese Academy of Sciences \\(^{3}\\)Key Lab of CS&AUS of Zhejiang Province \\(\\{lileyi, qi.ye, qmyang\\} @zju.edu.cn\\) qiaohj@ioz.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.62 + ], + "angle": 0, + "content": "Many low-level computer vision tasks are desirable to utilize the unprocessed RAW image as input, which remains the linear relationship between pixel values and scene radiance. Recent works advocate to embed the RAW image samples into sRGB images at capture time, and reconstruct the RAW from sRGB by these metadata when needed. However, there still exist some limitations in making full use of the metadata. In this paper, instead of following the perspective of sRGB-to-RAW mapping, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values. With this novel formulation, we propose to reconstruct the RAW image with an implicit neural function, which achieves significant performance improvement (more than 10dB average PSNR) only with the uniform sampling. Compared with most deep learning-based approaches, our method is trained in a self-supervised way that requiring no pre-training on different camera ISPs. We perform further experiments to demonstrate the effectiveness of our method, and show that our framework is also suitable for the task of guided super-resolution." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.648, + 0.21, + 0.664 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Low-level computer vision tasks benefit a lot from the scene-referred RAW images [7, 39, 19, 17, 16], which is rendered to the display-referred standard RGB (sRGB) images via camera image signal processors (ISPs). Compared with sRGB images, typical RAW images has the advantages of linear relationship between pixel values and scene radiance, as well as higher dynamic range. However, RAW images occupy obviously more memory than the sRGB images in common format like JPEG, which is unfavourable for transferring and sharing. Moreover, since most display and printing devices are designed for images stored and shared in sRGB format, it is inconvenient to directly replace sRGB with RAW. Consequently, mapping sRGB images back to their RAW counterparts, which is also called RAW reconstruction, is regarded as the appropriate way to" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.268, + 0.627, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.344, + 0.608, + 0.357 + ], + "angle": 0, + "content": "sRGB (input)" + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.268, + 0.751, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.344, + 0.717, + 0.357 + ], + "angle": 0, + "content": "RIR [25]" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.268, + 0.875, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.344, + 0.843, + 0.357 + ], + "angle": 0, + "content": "SAM [28]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.358, + 0.628, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.433, + 0.601, + 0.446 + ], + "angle": 0, + "content": "RAW (GT)" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.358, + 0.751, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.433, + 0.721, + 0.446 + ], + "angle": 0, + "content": "CAM [24]" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.358, + 0.875, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.434, + 0.828, + 0.445 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Figure 1. As RAW images are beneficial to many low-level computer vision tasks, we aim to reconstruct the RAW image from the corresponding sRGB image with the assistance of extra metadata. In this figure, the reconstructed RAW images are visualized through error maps. As can be seen, our method remarkably outperforms other related methods with the improvement of more than \\(10\\mathrm{dB}\\) PSNR. We owe this performance boost to the effectiveness of our proposed implicit neural function (INF)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.598, + 0.888, + 0.613 + ], + "angle": 0, + "content": "utilize the advantage of RAW data [23, 25, 36, 10, 28, 24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.84 + ], + "angle": 0, + "content": "Early RAW reconstruction methods focus on building standard models to reverse ISPs, which is parameterized by either explicit functions [4, 18, 14, 5] or neural networks [23, 36, 10]. However, these approaches are faced with the same issue that a parameterized model is only suitable for a specific ISP. Meanwhile, a series of methods [25, 26, 28, 24] propose to overcome this problem by embedding extra metadata into sRGB images at capture time. For such methods, the main challenge is to improve the accuracy with lower metadata generation cost. RIR [25] implements complex optimization algorithm to estimate the global mapping parameters as metadata, but suffers high computational cost. SAM [28] adopts a uniform sampling on RAW images to generate the metadata, which is further replaced with a sampler network by CAM [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "For the metadata-based methods of SAM [28] and CAM [24], the embedded RAW samples stores partial information of ISPs which helps to reconstruct the RAW images better; also, by conditioning the reconstruction algorithm on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18196" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "these metadata, the recovery of the RAW data turns into a conditional mapping function instead of a function fitted to a specific case, enabling the potential to achieve better generalization. Therefore, we adopt this strategy in the paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.348 + ], + "angle": 0, + "content": "Despite the progress that SAM [28] and CAM [24] have made, there still exist some limitations for the metadata-based RAW reconstruction methods. SAM utilizes RBF interpolation [3], the main idea of which is to calculate the difference between sampling and target points by a kernel function. However, a fixed kernel function lacks the flexibility to model various sRGB-to-RAW mappings. CAM directly uses a neural network for reconstruction but requires pre-training on pairs of sRGB and raw data from different types of ISPs. Also, we observe that the results of these methods fail to recover the saturated regions [26] (i.e., pixels with any channel value close to the maximum), as is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.471, + 0.635 + ], + "angle": 0, + "content": "To address the limitation, we propose a two-way RAW reconstruction algorithm based on an implicit neural function (INF). Previously, RAW reconstruction is formulated as mapping a sRGB image and the metadata to its RAW image. In this paper, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values, i.e. an implicit function. With this novel formulation, we can also decompose the problem into two aspects: a mapping function from the sRGB values to the corresponding raw values; a super-resolution function to interpolate the RAW image from the sparse samples. We observe that the super-resolution part usually exhibits much higher errors, indicating the latter a more challenging task. Accordingly, two branches are designed for each task inside an implicit neural network and the hyper-parameters for these branches are tuned to accommodate the difficulty of the tasks. Also, notice that with this formulation, the network can be trained in a self-supervised way, without the need of corresponding RAW images." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.636, + 0.415, + 0.649 + ], + "angle": 0, + "content": "Our contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.659, + 0.468, + 0.719 + ], + "angle": 0, + "content": "- We reformulate the RAW reconstruction problem as a RAW image approximation problem that learns the 2D-to-RAW mapping of image coordinates to RAW values conditioned on its sRGB image." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.729, + 0.468, + 0.759 + ], + "angle": 0, + "content": "- We decompose the reconstruction into two aspects and design the implicit neural network accordingly." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.769, + 0.468, + 0.814 + ], + "angle": 0, + "content": "- We conduct extensive experiments on different cameras and demonstrate our algorithm outperforms existing work significantly." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.659, + 0.468, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.827, + 0.219, + 0.843 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Blind RAW reconstruction. Early works of RAW reconstruction were blindly taken sRGB images as input without extra metadata. Since the processing stages of modern ISPs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.259 + ], + "angle": 0, + "content": "are designed more complicated, more complex models are proposed [4, 18, 14, 5]. However, these methods suffer the inconvenient calibration procedures that need to be repeated on each camera or even each camera setting. Deep learning-based methods (e.g., [23, 36]), similarly, are faced with the parallel issues of camera-specific models, which required abundant training data captured for each camera. Recent methods that simulate ISP architectures by assuming a classical set of ISP operations, such as [37, 10], cannot handle different camera settings as they take fixed parameters (e.g., gamma correction) and ISP length as a priori knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.626 + ], + "angle": 0, + "content": "RAW reconstruction with metadata. Compared with blind RAW reconstruction, a series of recent methods [25, 26, 28, 24] benefit from additional metadata embedded into sRGB images at capture time. Nguyen and Brown [25, 26] propose to extract and store the necessary parameters for recovering a RAW image from the sRGB counterpart. These parameters model the specific sRGB-to-RAW mapping and are restricted to a small memory (e.g., 128KB). However, their main algorithm deployed on the device has high computational cost and only considers the global tone mapping. Punnappurath and Brown [28], on the contrary, implement a uniform sampling on RAW images to save as metadata. They employ radial basis function (RBF) interpolation using sRGB pixel values and coordinates as input to reconstruct the RAW values. Nevertheless, the fixed type of RBF interpolation does not take full use of the information of sampled data. A further work [24] improves the effectiveness of sampling and recovering by taking a U-Net architecture [29] as both sampler and reconstruction networks, but their method requires to run a deep neural network on the device, which would lead to high computational cost and additional memory cost to save the pixel positions into metadata. We follow the uniform sampling in [28], but improve the reconstruction performance by INF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implicit neural representation. Implicit neural representation (INR) has recently been introduced to represent 2D images and 3D objects using coordinate-based multi-layer perceptron (MLP). To overcome the problem that conventional MLPs are incapable of representing high-frequency details of signals, two methods have been proposed. Sitzmann et al. [31] introduce SIREN, which replace ReLU activation with periodic activation (e.g., sine function). They demonstrate that the representation power of SIREN comes from the derivation invariance of sine function, and provide a number of potential applications and future works. A concurrent work [32] leverages random Fourier feature mapping on input coordinates to enable the MLP with ReLU to learn high-frequency details. It indicates that a Fourier feature mapping can beneficially address the spectral bias of a conventional MLP. Based on these breakthroughs, INR has been successfully adopted in various tasks [8, 1, 6, 35]. In this work, the proposed INF is based on [31], which shows" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18197" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.427, + 0.107 + ], + "angle": 0, + "content": "remarkably improvement in reconstruction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.111, + 0.47, + 0.307 + ], + "angle": 0, + "content": "Guided super-resolution. Another topic similar to RAW reconstruction with metadata is guided image superresolution, which aims to convert a low-resolution (LR) source image to the high-resolution (HR) target with a guided HR image [20, 30, 33]. The difference of these two topics is that RAW reconstruction has obvious mapping of pixel values between two HR images. On this aspect, PixTransform [20] is more related to our method. It uses an MLP with ReLU to learn the pixel-to-pixel mapping of guided HR image to target HR image, where the LR image is treated like the metadata in RAW reconstruction. Our method differs in the MLP layer, where we use SIREN layer [31] to enhance the expressive power [38]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.318, + 0.169, + 0.333 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.343, + 0.458, + 0.36 + ], + "angle": 0, + "content": "3.1. Problem Formulation with Implicit Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.593 + ], + "angle": 0, + "content": "Let \\( S \\) and \\( R \\) denote an sRGB image and the corresponding RAW image, respectively. Since \\( R \\) is converted to \\( S \\) through a series of operations in the camera ISP, classical methods aim to build a model \\( g \\) to map \\( S \\) back to \\( R \\), i.e., \\( R = g(S) \\). However, the model \\( g \\) is specific to a camera ISP or even a set of ISP parameters. Metadata-based methods introduce extra data \\( M \\) for RAW reconstruction, i.e., \\( R = f(S, M) \\). Here \\( f \\) is supposed to be a general function, and \\( M \\) includes the image-specific information generated at capture time, which is commonly implemented by sampling on the RAW images for low computation cost. Hence \\( M = \\{\\mathbf{p}_i, \\mathbf{r}_i\\}_{i=0}^N \\), where \\( \\mathbf{p}_i = (x_i, y_i) \\) is the coordinate of a sampled RAW pixel and \\( \\mathbf{r}_i = (r_{R_i}, r_{G_i}, r_{B_i}) \\) its RAW value. \\( N \\) represents the number of sampled RAW pixels and \\( i \\) refers to the \\( i^{th} \\) sampling one." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.655 + ], + "angle": 0, + "content": "In contrast to previous work that establishes a mapping from RGB to RGB values, \\( f:(r,s)\\to r \\), we reformulate the problem as a mapping from the 2D coordinates to the RAW values conditioning on the sRGB value" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.662, + 0.469, + 0.679 + ], + "angle": 0, + "content": "\\[\nf: (p; s) \\rightarrow r, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.732 + ], + "angle": 0, + "content": "where \\( r, s, p \\) denote the variables for the RAW value, the RGB value and coordinates. They can either represent the value for a pixel or a patch, or a set of coordinates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.469, + 0.793 + ], + "angle": 0, + "content": "Specifically, for the sRGB image, we only use the the sRGB value \\((s_{R_i}, s_{G_i}, s_{B_i})\\) at \\(\\mathbf{p}_i\\), queried by \\(S(\\mathbf{p}_i)\\). Therefore we aim to learn an implicit neural function \\(f_\\theta\\) taking a coordinate \\(\\mathbf{p}_i\\) and its sRGB value and output \\(\\mathbf{r}_i\\)." + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.801, + 0.469, + 0.817 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta}: \\left(\\mathbf {p} _ {i}; \\mathbf {s} _ {i}\\right)\\rightarrow \\mathbf {r} _ {i}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.856 + ], + "angle": 0, + "content": "To learn the function, we define the loss function below to find the best configuration for the parameters \\(\\theta\\) of \\(f\\)" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.863, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\| f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {s} _ {i}, \\mathbf {p} _ {i}\\right) - \\mathbf {r} _ {i} \\|, \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.633, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.19, + 0.605, + 0.204 + ], + "angle": 0, + "content": "RAW (GT)" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.09, + 0.759, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.191, + 0.742, + 0.205 + ], + "angle": 0, + "content": "(a) \\(f:\\mathbf{s}_i\\to \\mathbf{r}_i\\)" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.09, + 0.887, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.191, + 0.871, + 0.205 + ], + "angle": 0, + "content": "(b) \\(f:\\mathbf{p}_i\\to \\mathbf{r}_i\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Figure 2. Illustration of the two aspects of reconstruction. It can be observed that the reconstruction from (a) pixel value mapping is much more accurate than (b) spatial super-resolution. Our method utilizes the information from both \\(\\mathbf{s}_i\\) and \\(\\mathbf{p}_i\\), but give a different constraint to them to control their impacts." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.305, + 0.885, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Figure 3. The network structure of our proposed implicit neural function (INF). The inputs are divided into sRGB values and coordinates, as their branches are regularized with different parameters of weight decay. Each cube represents the output of each linear layer with sine activation proposed in [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.73 + ], + "angle": 0, + "content": "The overview of our method is shown in Figure 4. As mentioned above, considering the on-device computational cost, we follow SAM [28] to implement a uniform sampling on the RAW image to get the metadata \\( M \\). For the reconstruction stage, we implement the same sampling on the sRGB image, which is used to train the implicit neural function (INF) together with \\( M \\). The sRGB image is then processed by the trained INF to reconstruct the RAW image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.767, + 0.756 + ], + "angle": 0, + "content": "3.2. Two-way Raw Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We note that Equation (2) contains the reconstruction from two aspects: (a) a pixel values mapping (i.e., \\( f:\\mathbf{s}_i\\to \\mathbf{r}_i \\)) and (b) a spatial super-resolution (i.e., \\( f:\\mathbf{p}_i\\rightarrow \\mathbf{r}_i \\)). Since \\( M \\) will cost additional memory, the sampling rate is required to be constrained (e.g., no more than 1.5%), which limits the accuracy of (b). On the contrary, the sRGB-to-RAW mapping is piece-wise smooth [27], hence only such small amount of samples is able to model the whole mapping function theoretically. We provide an example in Fig-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18198" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.887, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.303, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Figure 4. Overview of our proposed RAW reconstruction method. At capture stage, we follow [28] to implement an uniform sampling on the RAW image to generate the metadata, which does not require to store the pixel positions. The metadata is then embedded into the sRGB image for storage and transmission. At reconstruction stage, the metadata together with corresponding sRGB samples are used to train the implicit neural functions (INFs), which then convert the sRGB image to the RAW. Compared with the state-of-the-art metadata-based RAW reconstruction approach [24], our method achieves both lower on-device computational cost and much higher reconstruction accuracy." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.402, + 0.462, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.592, + 0.472, + 0.649 + ], + "angle": 0, + "content": "Figure 5. Illustration of the patch-specific INF. Inspired by [28], we split the image into patches, and train the INF for each patch with its neighbours to make it robust to the local mapping. We further discuss the impact of patch-size in the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.713 + ], + "angle": 0, + "content": "ure 2 to intuitively illustrate this phenomenon. Therefore (a) is supposed to be more significant than (b) in \\( f \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Based on the analysis of the sub-tasks, we design an implicit neural network consisting of two branches as shown in Figure 3. INF is an MLP structure built with SIREN layers [31], where the inputs are separated into two branches and then concatenated into the output branch. Each branch contains 4 layers, with 256 channels for the hidden layer. The output channels of two input branches are reduced to half, which are then concatenated into the output branch. Here \\(\\theta\\) refers to the network weights, and Equation (4) can then be solved by training the INF." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Also, considering the difficulty of the tasks, we introduce regularisation parameters to Equation (3) to control" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.4, + 0.8, + 0.415 + ], + "angle": 0, + "content": "the complexity of different branches. That is," + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.422, + 0.892, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\theta}} = \\underset {\\boldsymbol {\\theta}} {\\arg \\min } \\sum_ {i = 1} ^ {N} \\| f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {s} _ {i}, \\mathbf {p} _ {i}\\right) - \\mathbf {r} _ {i} \\| + \\lambda_ {c} \\| \\boldsymbol {\\theta} _ {c} \\| ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.502 + ], + "angle": 0, + "content": "where \\(c \\in \\{\\mathbf{s}, \\mathbf{p}, \\mathbf{r}\\}\\) and \\(\\lambda_c\\) is a hyper-parameter to control the strength of regularisation for different parts of \\(M\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.502, + 0.893, + 0.638 + ], + "angle": 0, + "content": "We note that the structure of INF is most similar to the network of [20]. However, instead of merging the two input branches by adding their output, we first reduce the output size of these two branches and then concatenate them as the next input. This is because the output of SIREN Layers \\( X_{l} \\) is distributed in \\( X_{l} \\sim \\operatorname{Arcsin}(-1,1) \\) [31], and adding the outputs would let the output branch treat them equally. Hence we use concatenation to enable the network to learn different weights for the outputs of \\( \\mathbf{p}_i \\) and \\( \\mathbf{s}_i \\) branches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.822 + ], + "angle": 0, + "content": "Reconstruction conditioned on sRGB patches. Additionally, it is worth noting that local tone mapping is commonly applied as a favourable part of the ISP pipeline, which is adopted on local images to enhance the special objects (e.g., food, flower, sky). In order to make our method robust to local tone mapping, we split the input image into patches, and then train the INF for each patch. Moreover, inspired by [28], considering that the mapping of each patch and its neighbours are supposed to have strong relationship, we use the samples from each patch together with its eight neighbour patches to train the patch-specific INF. Therefore the mapping in Eq. (2) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.831, + 0.892, + 0.849 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta}: \\left(\\mathbf {p} _ {i}; \\{s _ {j} \\} _ {j \\in \\mathcal {N} (i)}\\right)\\rightarrow \\mathbf {r} _ {i}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(i)\\) represents neighbour patches of current pixel \\(\\mathbf{p}_i\\). Figure 5 illustrates the training strategy of INF for such patch-wise RAW reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18199" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.267, + 0.133 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.143, + 0.47, + 0.34 + ], + "angle": 0, + "content": "Baselines. We take three metadata-based RAW reconstruction methods for comparison: RIR [25], SAM [28] and CAM [24]. The RIR method computes the parameters of global sRGB-to-RAW mapping as metadata. The SAM and CAM methods adopt different sampling approaches on RAW images to generate metadata. We follow the uniform sampling as SAM [28], and show that it is enough to get much higher performance using uniform samples with the proposed method. Note that neither RIR nor SAM make their source codes publicly available, and only CAM reports their results reproduced on its own dataset. Therefore we only give a comparison with RIR and SAM on the dataset of CAM by inheriting its results in Table 1 and Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.471, + 0.556 + ], + "angle": 0, + "content": "Dataset. We deploy the same dataset of the previous work [24] to test our proposed method, which contains the preprocessed [13] and downsampled (with a downsampling factor of 4) images of three cameras from the NUS dataset [9]—Samsung NX2000, Olympus E-PL6 and Sony SLT-A57. Note that as our method does not require any pretraining step, the training and validation sets are not used in our method. Besides, since our method is able to directly process full-resolution images, we also compare it with CAM [24] on the original NUS dataset [9]. Additionally, as the NUS dataset [9] includes JPEG-compressed sRGB images rendered by real camera ISPs, we also conduct the experiments on these RAW-JPEG image pairs. These results are reported in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Implementation details. For capture stage, we use the same sampling rate of \\(1.5\\%\\) like [24] in all experiments for fair comparison with other metadata-based methods. With respect to reconstruction stage, we split the image into different patches—(228, 272) for Samsung NX2000, (216, 288) for Olympus E-PL6 and (204, 304) for Sony SLT-A57—to make the resolution divisible. The proposed INF is trained by Adam optimizer [15] with \\(L_{2}\\) loss in 500 iterations. We set the initial learning rate to 0.0001, and reduce it to half every 200 iterations. The regularisation parameters are set to \\(\\lambda_{\\mathrm{s}} = 0.0001, \\lambda_{\\mathrm{p}} = 0.1, \\lambda_{\\mathrm{r}} = 0.001\\) to produce the final results. It is worth noting that the reconstruction network is only trained for each specific patch, and will be re-initialized for the training of other patches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.783, + 0.172, + 0.798 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Quantitative results. We report the quantitative comparison results against other three baselines in Table 1. It can be observed that our method makes a remarkably improvement in both PSNR and SSIM metrics, especially outperforms the closest competitor CAM [24] by more than \\(10\\mathrm{dB}\\) in average PSNR. This indicates the effectiveness of INF to" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.76, + 0.108 + ], + "angle": 0, + "content": "modeling the sRGB-to-RAW mapping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.111, + 0.894, + 0.262 + ], + "angle": 0, + "content": "Qualitative results. We also provide examples of the qualitative comparison in Figure 6, which gives an intuitive explanation of the success of our method. As can be seen from the error maps, our method significantly reduces the reconstructed error on the whole images, especially the saturated regions [26]. We note that recovering saturated regions is the most difficult problem for all RAW reconstruction methods, and is the only case of our method getting large errors \\((\\geqslant 1\\%)\\). We attribute this to the expressive power [38] of INF, which is further discussed in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.273, + 0.653, + 0.289 + ], + "angle": 0, + "content": "4.3. Ablation study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.893, + 0.452 + ], + "angle": 0, + "content": "RBF vs. INF. We first compare RBF and INF with different input, which is shown in Table 2. As can be seen, the proposed INF outperforms the RBF interpolation [28] whatever the input is. We attribute the main performance boost to the basic structure, i.e. SIREN [31], as INF is able to modelling the sRGB-to-RAW mapping more accurately (59.62dB) only with the pixel value as input. Moreover, compared with RBF, our proposed INF can make further improvement by splitting the input into separate branches and regularize them with different weights." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.698 + ], + "angle": 0, + "content": "The structure of INF. We also conduct experiments on the structure of INF, which is reported in Figure 7. We compare the different activation functions and the layer location where two input branches are merged. It is obviously shown that the INF with sine activation remarkably outperforms the INF with ReLU function, which can be attributed to the expressive power of INRs [38]. As to the location of fusion layer, we find there is no significant difference except the final layer for the INF with sine activation, since it directly maps the merged vector to the output by a fully connection layer without activation. Therefore, we merge the two branches at fourth layer to balance the complexity of each branch1. On the contrary, due to lack of such expressive power, the INF with ReLU requires more layers to model the mapping. Hence the performance increases when the input branch goes deep." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.894, + 0.853 + ], + "angle": 0, + "content": "Regularisation parameters. We further discuss the influence of key setups to INF, which mainly focuses on the regularisation parameters and patch size. As is shown in Table 3, it is obvious that the regularisation parameters play a vital role in the final accuracy. Here we use the grid search with a commonly-used tuner TPE [2] (implemented by [22]) in a grid of \\([1, 10^{-1}, \\dots, 10^{-5}]\\) to determine each parameter. In our experiments, we find the regulation parameters are mainly related to the complexity of camera ISP rather than data dependent, which means we only need" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.894, + 0.901 + ], + "angle": 0, + "content": "1In fact, we find the fusion at third layer would achieve a slightly higher performance, but which can be ignored compared with the instability of random initialization. See more detail in supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18200" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.089, + 0.855, + 0.233 + ], + "angle": 0, + "content": "
DatasetMethodSamsung NX2000Olympus E-PL6Sony SLT-A57
PSNRSSIMPSNRSSIMPSNRSSIM
CAM [24]RIR [25]45.660.993948.420.992451.260.9982
SAM [28]47.030.996249.350.997850.440.9982
CAM [24]49.570.997551.540.998053.110.9985
Ours53.530.998556.710.999158.360.9992
NUS [9]CAM [24]49.510.993752.870.996153.340.9959
Ours61.480.999163.380.999363.800.9995
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.256, + 0.892, + 0.298 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with [25, 28, 24]. Here CAM dataset [24] involves downsampling operation (with the factor of 4) on the original NUS dataset [9], which undermines the strength of local mapping. For CAM dataset, we conduct the experiments on their test set to inherit the results of [24]. For NUS dataset, we report the results tested on the whole dataset. The best score for each column is in bold." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.312, + 0.887, + 0.781 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.791, + 0.89, + 0.834 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison with [25, 28, 24]. Each two rows from top to bottom are respectively selected from Olympus E-PL6, Samsung NX2000 and Sony SLT-A57. The GT RAW images are processed with a gamma function for better visualization, and the reconstructed RAW images are visualized through error maps. This figure is best viewed in the electronic version." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.86, + 0.47, + 0.89 + ], + "angle": 0, + "content": "to tune it for different ISPs, e.g., digital camera ISP, smart phone ISP, or software ISP. It also means we just need to" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.86, + 0.892, + 0.89 + ], + "angle": 0, + "content": "use one or quite a few images for tuning, resulting in a fast tuning process—within 2 hours on a single RTX3090. The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18201" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.422, + 0.233 + ], + "angle": 0, + "content": "
MethodInputPSNR
RBF [28]s29.10
RBF [28]p21.71
RBF [28](s, p)49.69
INFs59.62
INFp31.09
INF(s, p)60.91
INF{s; p}62.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.256, + 0.47, + 0.327 + ], + "angle": 0, + "content": "Table 2. Ablation study on different reconstruction methods. The best score is in bold and the second in italic. \\((\\mathbf{s},\\mathbf{p})\\) refers to combine \\(\\mathbf{s}\\) and \\(\\mathbf{p}\\) into a single input vector, while \\(\\{\\mathbf{s};\\mathbf{p}\\}\\) denotes separating them as different input. For the RBF interpolation, we use the linear RBF kernel as proposed in [28]." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.359, + 0.432, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the structure of INF. The abscissa value represents the layer where we merge the two input branches, e.g., the branches are merged at fourth layer in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.468, + 0.639 + ], + "angle": 0, + "content": "tuned parameters can then be used for all scenes and all other images from the ISPs with similar complexity, hence we use the same parameters for the three reported cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.468, + 0.867 + ], + "angle": 0, + "content": "As shown in Table 3, the proposed INF achieves the best performance at \\(\\lambda_{\\mathbf{s}} = 0.0001\\), \\(\\lambda_{\\mathbf{p}} = 0.1\\), \\(\\lambda_{\\mathbf{r}} = 0.001\\), which indicates the spatial branch is not supposed to be more complex. This is because the output \\(\\mathbf{r}\\) has weaker relationship with the coordinates \\(\\mathbf{p}\\) than sRGB values \\(\\mathbf{s}\\), and regularisation could help to control the contribution of \\(\\mathbf{s}\\) and \\(\\mathbf{p}\\). Therefore when we set \\(\\lambda_{\\mathbf{p}} = 0\\), the reconstruction performance dramatically drops down due to inaccurate information from \\(\\mathbf{p}\\) is expanded. We believe such intuitive relationship would be helpful to set the initial values for regularisation parameters. Besides, for different ISPs with similar complexity, the results are insensitive to the parameters in a wide range (for example, where all branches are regularized by \\(\\lambda \\geqslant 10^{-4}\\)). In this sense, the risk of over-fitting due to these hyper-parameter settings is very small." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Impact of patch size. Moreover, we find that different size of the patch would lead to an obvious accuracy gap. For the" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.88, + 0.27 + ], + "angle": 0, + "content": "
Patch size(λs,λp,λr)PSNR
×147.48
×2(0,0,0)49.17
×450.53
×161.13
×2(0.0001,0.001,0.0001)60.38
×457.80
×162.58
×2(0.0001,0.1,0.001)62.24
×459.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.405 + ], + "angle": 0, + "content": "Table 3. Ablation study on the patch size and regularisation parameters. The best score is in bold and the second in italic. For the patch size, \\(\\times 2\\) refers to the sizes we report in our main experiments, and the rescale factor of 1, 2, 4 corresponds to the length of patch sides. We note that the accuracy gap less than \\(0.5\\mathrm{dB}\\) can not represent an improvement, which may be caused by the different calculations of PSNR and the random initialization of INF. This is further discussed in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.596 + ], + "angle": 0, + "content": "case of all regularisation parameters set to 0, we find small patch would lead to accuracy reduction. When the regularisation is correctly set, reconstruction on a small patch is more accurate than on a large one. This can be explained by the various local mapping functions in different image regions, which means modelling the local mapping for a small patch would be more accurate than a larger one. Note that though reducing the patch size would be helpful to the reconstruction accuracy, it would take much more time to recover an image. Therefore, we compromise to use the patch size around 200-300 in the main experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.618, + 0.619 + ], + "angle": 0, + "content": "4.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.677 + ], + "angle": 0, + "content": "Why INF works in sRGB-to-RAW mapping. We first rethink the RBF interpolation [3] used in SAM [28], by which the sRGB-to-RAW mapping is built in the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.686, + 0.891, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = \\mathbf {M} ^ {\\text {t e s t}} \\mathbf {M} ^ {- 1} y \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.774 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}\\) denotes the RBF interpolation matrix with linear kernel function [27], and test refers to the pixels that need to be reconstructed. As discussed in [32], the output of an MLP can be approximated as:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.782, + 0.891, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\hat {y} ^ {(t)} = \\mathbf {K} ^ {\\text {t e s t}} \\mathbf {K} ^ {- 1} \\left(\\mathbf {I} - e ^ {- \\eta \\mathbf {K} t}\\right) y \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}\\) denotes the NTK matrix [12] and \\(t\\) refers to training iterations. In our experiments, an MLP (typical with ReLU activation) tends to behave like RBF when \\(t\\) is large enough (e.g., \\(t > 1000\\)), which meets the results of Equation (6) and Equation (7). The reason why RBF interpolation works can be attributed to that sRGB-RAW mapping" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18202" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.445, + 0.188 + ], + "angle": 0, + "content": "
Method×4×8×16
Bicubic2.624.196.68
PixTransform [20]3.513.944.94
INF2.503.065.23
INF with ReLU2.392.693.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.211, + 0.47, + 0.253 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on guided depth superresolution. We use the Middlebury dataset provided by [33] for testing. The results are reported in terms of average RMSE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.468, + 0.448 + ], + "angle": 0, + "content": "is piecewise smooth [27], but there still exists errors when using a linear kernel function, which is also suffered by MLPs with ReLU. On the other hand, Sitzmann et al. [31] demonstrate that MLPs with periodic activation functions (i.e., SIRENs) are more suitable for representing complicated signals and their derivatives compared with the typical MLP architectures. SIRENs have the advantage of representing the derivatives of the signal [31], which is more helpful to reconstruct fine details. Therefore, recovering RAW images by SIRENs can effectively improve the performance, which is also proved by Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.468, + 0.68 + ], + "angle": 0, + "content": "Limitations and future research. One of the unresolved problem of our method is that the regularisation parameters of weight decay require pre-experiments to determine. These three parameters play a significant role to control the final result by limiting the complexity of network. In our experiments, we find the parameters are strongly related to the correlation of input and output data. For example, compared with the coordinates, sRGB values have a more obvious relationship with the RAW values, where the regularisation parameters are respectively set to \\(\\lambda_{p} = 0.1\\) and \\(\\lambda_{s} = 0.0001\\). Such correlation is suitable for modelling with another neural network, and would lead to the performance improvement with optimized parameters. Therefore, learning the regularisation parameters from the input and output is supposed to be a valuable topic of future work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.692, + 0.21, + 0.709 + ], + "angle": 0, + "content": "4.5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Guided depth super-resolution As mentioned in Section 2, our framework is related to guided super-resolution, therefore we also provide an experiment to illustrate the applicability of our framework. Considering that our method is only self-supervised without any priori knowledge, we give a comparison with another self-supervised method PixTransform [20]. As shown in Table 4, INF with ReLU outperforms other methods, which indicates the ReLU activation is more suitable for this task than the sine activation. The reason for ReLU working better can be attributed to the sharp changes of the depth near the object edges, as ReLU is piecewise and discontinuous, which fits this depth disconti" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.633, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.191, + 0.591, + 0.203 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.09, + 0.758, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.191, + 0.716, + 0.203 + ], + "angle": 0, + "content": "Guide" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.09, + 0.886, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.191, + 0.843, + 0.204 + ], + "angle": 0, + "content": "Target" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.205, + 0.633, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.306, + 0.625, + 0.319 + ], + "angle": 0, + "content": "PixTransform [20]" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.205, + 0.758, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.306, + 0.71, + 0.317 + ], + "angle": 0, + "content": "INF" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.205, + 0.887, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.306, + 0.871, + 0.318 + ], + "angle": 0, + "content": "INF with ReLU" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.332, + 0.89, + 0.359 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison on guided depth super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.48 + ], + "angle": 0, + "content": "nuity. In contrast, the discontinuity in RAW-to-sRGB mapping is moderate and the smooth sine activation function is a better option. We also provide qualitative comparison in Figure 8. which shows that the result of INF, though does not achieve higher accuracy, obtains better visual quality compared with the result of INF with ReLU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Low-light image enhancement. As a potential application of RAW reconstruction, we test our method on low-light image enhancement (LLIE) task. We find that the LLIE task can be simplified into a linear degradation problem (e.g., enlarging the pixel values by the same multiple) on the reconstructed RAW images, which achieves equal visual quality compared with the state-of-the-art deep learning-based approaches [11, 34, 21]. The results are reported in our supplementary. Note that such comparison is merely aimed to illustrate the effectiveness of executing low-light enhancement on reconstructed RAW images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.67, + 0.618, + 0.685 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.848 + ], + "angle": 0, + "content": "We propose a method for recovering the RAW image from the sRGB counterpart with assistance of additional metadata, which is sampled from the RAW image at capture time. We introduce the implicit neural function (INF) to remarkably improve the reconstruction accuracy (10 dB average PSNR) only with uniform sampling. We prove that the structure of INF is beneficial to merging the information from both pixel values and coordinates. Further experiments indicate that our framework is also suitable for the task of guided super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.857, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was partly supported by National Key R&D Program of China (Grant No. 2021YFD1400200) and National Natural Science Foundation of China (Grant No. 62088101)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18203" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19323-19332, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.239 + ], + "angle": 0, + "content": "[2] James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization. Advances in neural information processing systems, 24, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[3] Marco Evangelos Biancolini. Fast radial basis functions for engineering applications. Springer, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.312 + ], + "angle": 0, + "content": "[4] Ayan Chakrabarti, Daniel Scharstein, and Todd E Zickler. An empirical camera model for internet color vision. In BMVC, volume 1, page 4. CiteSeer, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[5] Ayan Chakrabarti, Ying Xiong, Baochen Sun, Trevor Darrell, Daniel Scharstein, Todd Zickler, and Kate Saenko. Modeling radiometric uncertainty for vision with tonemapred color images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11):2185-2198, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[6] Bowei Chen, Tiancheng Zhi, Martial Hebert, and Srinivasa G Narasimhan. Learning continuous implicit representation for near-periodic patterns. arXiv preprint arXiv:2208.12278, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[7] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.469, + 0.634 + ], + "angle": 0, + "content": "[9] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[10] Marcos V Conde, Steven McDonagh, Matteo Maggioni, Ales Leonardis, and Eduardo Pérez-Pellitero. Model-based image signal processors via learnable dictionaries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 481-489, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[11] Chunle Guo Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1780-1789, June 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[12] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Hakki Can Karaimer and Michael S Brown. A software platform for manipulating the camera imaging pipeline. In European Conference on Computer Vision, pages 429-444. Springer, 2016. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] Seon Joo Kim, Hai Ting Lin, Zheng Lu, Sabine Süssstrunk, Stephen Lin, and Michael S Brown. A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(12):2289-2302, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[16] Shayan Kousha, Ali Maleky, Michael S Brown, and Marcus A Brubaker. Modeling srgb camera noise with normalizing flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17463-17471, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Zhihao Li, Si Yi, and Zhan Ma. Rendering nighttime image via cascaded color and brightness compensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 897–905, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[18] Hai Ting Lin, Zheng Lu, Seon Joo Kim, and Michael S Brown. Nonuniform lattice regression for modeling the camera imaging pipeline. In European Conference on Computer Vision, pages 556-568. Springer, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[19] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image hdr reconstruction by learning to reverse the camera pipeline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1651-1660, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[20] Riccardo de Lutio, Stefano D'aronco, Jan Dirk Wegner, and Konrad Schindler. Guided super-resolution as pixel-to-pixel transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8829-8837, 2019. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.825, + 0.621 + ], + "angle": 0, + "content": "[22] Microsoft. Neural Network Intelligence, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[23] Seonghyeon Nam and Seon Joo Kim. Modelling the scene dependent imaging in cameras with a deep neural network. In Proceedings of the IEEE International Conference on Computer Vision, pages 1717-1725, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[24] Seonghyeon Nam, Abhijith Punnappurath, Marcus A Brubaker, and Michael S Brown. Learning srgb-to-raw-rgb de-rendering with content-aware metadata. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17704–17713, 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[25] Rang MH Nguyen and Michael S Brown. Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1655-1663, 2016. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] Rang MH Nguyen and Michael S Brown. Raw image reconstruction using a self-contained srgb-jpeg image with small memory overhead. International journal of computer vision, 126(6):637-650, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Abhijith Punnappurath and Michael S Brown. Learning raw image reconstruction-aware deep image compressors. IEEE" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18204" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "transactions on pattern analysis and machine intelligence, 42(4):1013-1019, 2019. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[28] Abhijith Punnappurath and Michael S Brown. Spatially aware metadata for raw reconstruction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 218-226, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[30] Guy Shacht, Dov Danon, Sharon Fogel, and Daniel Cohen-Or. Single pair cross-modality super resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6378–6387, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[31] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[32] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[33] Jiaxiang Tang, Xiaokang Chen, and Gang Zeng. Joint implicit image function for guided depth super-resolution. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4390-4399, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.516, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[34] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2604-2612, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.653 + ], + "angle": 0, + "content": "[35] Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neuralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1526-1535, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.469, + 0.71 + ], + "angle": 0, + "content": "[36] Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6287-6296, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.712, + 0.469, + 0.767 + ], + "angle": 0, + "content": "[37] Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248-4257, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.769, + 0.469, + 0.837 + ], + "angle": 0, + "content": "[38] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.839, + 0.469, + 0.894 + ], + "angle": 0, + "content": "[39] Xuaner Zhang, Qifeng Chen, Ren Ng, and Vladlen Koltun. Zoom to learn, learn to zoom. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3762-3770, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18205" + } + ] +] \ No newline at end of file diff --git a/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_origin.pdf b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..91c8c9e8a5641743350ea57039acf36229845405 --- /dev/null +++ b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/c16d5ffd-61d0-4f71-a138-3e75fa3b8833_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:008a885efc3c8b483e24d92ba7197eaf39fe0b0cd8e340c34ffa92a29d245e1a +size 9790552 diff --git a/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/full.md b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..83e5c1825cdf74803f850cc3265d6e3316da30a3 --- /dev/null +++ b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/full.md @@ -0,0 +1,295 @@ +# Metadata-Based RAW Reconstruction via Implicit Neural Functions + +Leyi Li $^{1}$ Huijie Qiao $^{2}$ Qi Ye $^{1,3}$ Qinmin Yang $^{1}$ $^{1}$ Zhejiang University $^{2}$ Chinese Academy of Sciences $^{3}$ Key Lab of CS&AUS of Zhejiang Province $\{lileyi, qi.ye, qmyang\} @zju.edu.cn$ qiaohj@ioz.ac.cn + +# Abstract + +Many low-level computer vision tasks are desirable to utilize the unprocessed RAW image as input, which remains the linear relationship between pixel values and scene radiance. Recent works advocate to embed the RAW image samples into sRGB images at capture time, and reconstruct the RAW from sRGB by these metadata when needed. However, there still exist some limitations in making full use of the metadata. In this paper, instead of following the perspective of sRGB-to-RAW mapping, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values. With this novel formulation, we propose to reconstruct the RAW image with an implicit neural function, which achieves significant performance improvement (more than 10dB average PSNR) only with the uniform sampling. Compared with most deep learning-based approaches, our method is trained in a self-supervised way that requiring no pre-training on different camera ISPs. We perform further experiments to demonstrate the effectiveness of our method, and show that our framework is also suitable for the task of guided super-resolution. + +# 1. Introduction + +Low-level computer vision tasks benefit a lot from the scene-referred RAW images [7, 39, 19, 17, 16], which is rendered to the display-referred standard RGB (sRGB) images via camera image signal processors (ISPs). Compared with sRGB images, typical RAW images has the advantages of linear relationship between pixel values and scene radiance, as well as higher dynamic range. However, RAW images occupy obviously more memory than the sRGB images in common format like JPEG, which is unfavourable for transferring and sharing. Moreover, since most display and printing devices are designed for images stored and shared in sRGB format, it is inconvenient to directly replace sRGB with RAW. Consequently, mapping sRGB images back to their RAW counterparts, which is also called RAW reconstruction, is regarded as the appropriate way to + +![](images/31a015de6f98ea740612b792ad93ecb9cf0bd51955312dbb18e2bbe775b307e1.jpg) + +![](images/c8d4334f820bffa702019a34bfb31ca4a28364fce885a43e7eee9bb4add11661.jpg) + +![](images/a2e9c957293d37b991170c1821173687f873d0b839aac09bb86ac70703554e09.jpg) + +![](images/fca0e4de21f1902af753f138645bcb5eaf85b9d422cce49a476af77cd3769bcf.jpg) +sRGB (input) +RAW (GT) +Figure 1. As RAW images are beneficial to many low-level computer vision tasks, we aim to reconstruct the RAW image from the corresponding sRGB image with the assistance of extra metadata. In this figure, the reconstructed RAW images are visualized through error maps. As can be seen, our method remarkably outperforms other related methods with the improvement of more than $10\mathrm{dB}$ PSNR. We owe this performance boost to the effectiveness of our proposed implicit neural function (INF). + +![](images/d90cb0ea54e8c45ed93a019b6dc10d41143a66a3c0986585e85c1396252d4f5c.jpg) +RIR [25] +CAM [24] + +![](images/e4ec6ff7eefc56932a3671d30aafe47cae63c53e79063eb37c652587f8031fd0.jpg) +SAM [28] +Ours + +utilize the advantage of RAW data [23, 25, 36, 10, 28, 24]. + +Early RAW reconstruction methods focus on building standard models to reverse ISPs, which is parameterized by either explicit functions [4, 18, 14, 5] or neural networks [23, 36, 10]. However, these approaches are faced with the same issue that a parameterized model is only suitable for a specific ISP. Meanwhile, a series of methods [25, 26, 28, 24] propose to overcome this problem by embedding extra metadata into sRGB images at capture time. For such methods, the main challenge is to improve the accuracy with lower metadata generation cost. RIR [25] implements complex optimization algorithm to estimate the global mapping parameters as metadata, but suffers high computational cost. SAM [28] adopts a uniform sampling on RAW images to generate the metadata, which is further replaced with a sampler network by CAM [24]. + +For the metadata-based methods of SAM [28] and CAM [24], the embedded RAW samples stores partial information of ISPs which helps to reconstruct the RAW images better; also, by conditioning the reconstruction algorithm on + +these metadata, the recovery of the RAW data turns into a conditional mapping function instead of a function fitted to a specific case, enabling the potential to achieve better generalization. Therefore, we adopt this strategy in the paper. + +Despite the progress that SAM [28] and CAM [24] have made, there still exist some limitations for the metadata-based RAW reconstruction methods. SAM utilizes RBF interpolation [3], the main idea of which is to calculate the difference between sampling and target points by a kernel function. However, a fixed kernel function lacks the flexibility to model various sRGB-to-RAW mappings. CAM directly uses a neural network for reconstruction but requires pre-training on pairs of sRGB and raw data from different types of ISPs. Also, we observe that the results of these methods fail to recover the saturated regions [26] (i.e., pixels with any channel value close to the maximum), as is shown in Figure 1. + +To address the limitation, we propose a two-way RAW reconstruction algorithm based on an implicit neural function (INF). Previously, RAW reconstruction is formulated as mapping a sRGB image and the metadata to its RAW image. In this paper, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values, i.e. an implicit function. With this novel formulation, we can also decompose the problem into two aspects: a mapping function from the sRGB values to the corresponding raw values; a super-resolution function to interpolate the RAW image from the sparse samples. We observe that the super-resolution part usually exhibits much higher errors, indicating the latter a more challenging task. Accordingly, two branches are designed for each task inside an implicit neural network and the hyper-parameters for these branches are tuned to accommodate the difficulty of the tasks. Also, notice that with this formulation, the network can be trained in a self-supervised way, without the need of corresponding RAW images. + +Our contribution can be summarized as follows: + +- We reformulate the RAW reconstruction problem as a RAW image approximation problem that learns the 2D-to-RAW mapping of image coordinates to RAW values conditioned on its sRGB image. +- We decompose the reconstruction into two aspects and design the implicit neural network accordingly. +- We conduct extensive experiments on different cameras and demonstrate our algorithm outperforms existing work significantly. + +# 2. Related Work + +Blind RAW reconstruction. Early works of RAW reconstruction were blindly taken sRGB images as input without extra metadata. Since the processing stages of modern ISPs + +are designed more complicated, more complex models are proposed [4, 18, 14, 5]. However, these methods suffer the inconvenient calibration procedures that need to be repeated on each camera or even each camera setting. Deep learning-based methods (e.g., [23, 36]), similarly, are faced with the parallel issues of camera-specific models, which required abundant training data captured for each camera. Recent methods that simulate ISP architectures by assuming a classical set of ISP operations, such as [37, 10], cannot handle different camera settings as they take fixed parameters (e.g., gamma correction) and ISP length as a priori knowledge. + +RAW reconstruction with metadata. Compared with blind RAW reconstruction, a series of recent methods [25, 26, 28, 24] benefit from additional metadata embedded into sRGB images at capture time. Nguyen and Brown [25, 26] propose to extract and store the necessary parameters for recovering a RAW image from the sRGB counterpart. These parameters model the specific sRGB-to-RAW mapping and are restricted to a small memory (e.g., 128KB). However, their main algorithm deployed on the device has high computational cost and only considers the global tone mapping. Punnappurath and Brown [28], on the contrary, implement a uniform sampling on RAW images to save as metadata. They employ radial basis function (RBF) interpolation using sRGB pixel values and coordinates as input to reconstruct the RAW values. Nevertheless, the fixed type of RBF interpolation does not take full use of the information of sampled data. A further work [24] improves the effectiveness of sampling and recovering by taking a U-Net architecture [29] as both sampler and reconstruction networks, but their method requires to run a deep neural network on the device, which would lead to high computational cost and additional memory cost to save the pixel positions into metadata. We follow the uniform sampling in [28], but improve the reconstruction performance by INF. + +Implicit neural representation. Implicit neural representation (INR) has recently been introduced to represent 2D images and 3D objects using coordinate-based multi-layer perceptron (MLP). To overcome the problem that conventional MLPs are incapable of representing high-frequency details of signals, two methods have been proposed. Sitzmann et al. [31] introduce SIREN, which replace ReLU activation with periodic activation (e.g., sine function). They demonstrate that the representation power of SIREN comes from the derivation invariance of sine function, and provide a number of potential applications and future works. A concurrent work [32] leverages random Fourier feature mapping on input coordinates to enable the MLP with ReLU to learn high-frequency details. It indicates that a Fourier feature mapping can beneficially address the spectral bias of a conventional MLP. Based on these breakthroughs, INR has been successfully adopted in various tasks [8, 1, 6, 35]. In this work, the proposed INF is based on [31], which shows + +remarkably improvement in reconstruction accuracy. + +Guided super-resolution. Another topic similar to RAW reconstruction with metadata is guided image superresolution, which aims to convert a low-resolution (LR) source image to the high-resolution (HR) target with a guided HR image [20, 30, 33]. The difference of these two topics is that RAW reconstruction has obvious mapping of pixel values between two HR images. On this aspect, PixTransform [20] is more related to our method. It uses an MLP with ReLU to learn the pixel-to-pixel mapping of guided HR image to target HR image, where the LR image is treated like the metadata in RAW reconstruction. Our method differs in the MLP layer, where we use SIREN layer [31] to enhance the expressive power [38]. + +# 3. Method + +# 3.1. Problem Formulation with Implicit Function + +Let $S$ and $R$ denote an sRGB image and the corresponding RAW image, respectively. Since $R$ is converted to $S$ through a series of operations in the camera ISP, classical methods aim to build a model $g$ to map $S$ back to $R$ , i.e., $R = g(S)$ . However, the model $g$ is specific to a camera ISP or even a set of ISP parameters. Metadata-based methods introduce extra data $M$ for RAW reconstruction, i.e., $R = f(S, M)$ . Here $f$ is supposed to be a general function, and $M$ includes the image-specific information generated at capture time, which is commonly implemented by sampling on the RAW images for low computation cost. Hence $M = \{\mathbf{p}_i, \mathbf{r}_i\}_{i=0}^N$ , where $\mathbf{p}_i = (x_i, y_i)$ is the coordinate of a sampled RAW pixel and $\mathbf{r}_i = (r_{R_i}, r_{G_i}, r_{B_i})$ its RAW value. $N$ represents the number of sampled RAW pixels and $i$ refers to the $i^{th}$ sampling one. + +In contrast to previous work that establishes a mapping from RGB to RGB values, $f:(r,s)\to r$ , we reformulate the problem as a mapping from the 2D coordinates to the RAW values conditioning on the sRGB value + +$$ +f: (p; s) \rightarrow r, \tag {1} +$$ + +where $r, s, p$ denote the variables for the RAW value, the RGB value and coordinates. They can either represent the value for a pixel or a patch, or a set of coordinates. + +Specifically, for the sRGB image, we only use the the sRGB value $(s_{R_i}, s_{G_i}, s_{B_i})$ at $\mathbf{p}_i$ , queried by $S(\mathbf{p}_i)$ . Therefore we aim to learn an implicit neural function $f_\theta$ taking a coordinate $\mathbf{p}_i$ and its sRGB value and output $\mathbf{r}_i$ . + +$$ +f _ {\theta}: \left(\mathbf {p} _ {i}; \mathbf {s} _ {i}\right)\rightarrow \mathbf {r} _ {i}. \tag {2} +$$ + +To learn the function, we define the loss function below to find the best configuration for the parameters $\theta$ of $f$ + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {N} \| f _ {\boldsymbol {\theta}} \left(\mathbf {s} _ {i}, \mathbf {p} _ {i}\right) - \mathbf {r} _ {i} \|, \tag {3} +$$ + +![](images/044f91c70c87888b6d885fa68ded9dd79a496b72e01e155b78c5b7a8de67c3ba.jpg) +RAW (GT) + +![](images/4145960c48f539469765404e92c981b40e8a5afd9c91ec4bbc9ed34f6c3d9a90.jpg) +(a) $f:\mathbf{s}_i\to \mathbf{r}_i$ + +![](images/1d2e7b2a0d1ae35636404deff0c8b44099a426ada025f67cffd6d0565f79858d.jpg) +(b) $f:\mathbf{p}_i\to \mathbf{r}_i$ + +![](images/b81fb7584acee094625535f4ed1a6aebff0de0e8e4f40b5ec85659a747b454a2.jpg) +Figure 2. Illustration of the two aspects of reconstruction. It can be observed that the reconstruction from (a) pixel value mapping is much more accurate than (b) spatial super-resolution. Our method utilizes the information from both $\mathbf{s}_i$ and $\mathbf{p}_i$ , but give a different constraint to them to control their impacts. +Figure 3. The network structure of our proposed implicit neural function (INF). The inputs are divided into sRGB values and coordinates, as their branches are regularized with different parameters of weight decay. Each cube represents the output of each linear layer with sine activation proposed in [31]. + +The overview of our method is shown in Figure 4. As mentioned above, considering the on-device computational cost, we follow SAM [28] to implement a uniform sampling on the RAW image to get the metadata $M$ . For the reconstruction stage, we implement the same sampling on the sRGB image, which is used to train the implicit neural function (INF) together with $M$ . The sRGB image is then processed by the trained INF to reconstruct the RAW image. + +# 3.2. Two-way Raw Reconstruction + +We note that Equation (2) contains the reconstruction from two aspects: (a) a pixel values mapping (i.e., $f:\mathbf{s}_i\to \mathbf{r}_i$ ) and (b) a spatial super-resolution (i.e., $f:\mathbf{p}_i\rightarrow \mathbf{r}_i$ ). Since $M$ will cost additional memory, the sampling rate is required to be constrained (e.g., no more than 1.5%), which limits the accuracy of (b). On the contrary, the sRGB-to-RAW mapping is piece-wise smooth [27], hence only such small amount of samples is able to model the whole mapping function theoretically. We provide an example in Fig- + +![](images/36658c6332cb8f442cdc43c5d757fc25a22f9ff8e998466eb5a247db73c24c1d.jpg) +Figure 4. Overview of our proposed RAW reconstruction method. At capture stage, we follow [28] to implement an uniform sampling on the RAW image to generate the metadata, which does not require to store the pixel positions. The metadata is then embedded into the sRGB image for storage and transmission. At reconstruction stage, the metadata together with corresponding sRGB samples are used to train the implicit neural functions (INFs), which then convert the sRGB image to the RAW. Compared with the state-of-the-art metadata-based RAW reconstruction approach [24], our method achieves both lower on-device computational cost and much higher reconstruction accuracy. + +![](images/373b0b37cca2ffd48117d60a8bc3c977db2b78c6462083509d144e255e2f2e67.jpg) +Figure 5. Illustration of the patch-specific INF. Inspired by [28], we split the image into patches, and train the INF for each patch with its neighbours to make it robust to the local mapping. We further discuss the impact of patch-size in the ablation study. + +ure 2 to intuitively illustrate this phenomenon. Therefore (a) is supposed to be more significant than (b) in $f$ . + +Based on the analysis of the sub-tasks, we design an implicit neural network consisting of two branches as shown in Figure 3. INF is an MLP structure built with SIREN layers [31], where the inputs are separated into two branches and then concatenated into the output branch. Each branch contains 4 layers, with 256 channels for the hidden layer. The output channels of two input branches are reduced to half, which are then concatenated into the output branch. Here $\theta$ refers to the network weights, and Equation (4) can then be solved by training the INF. + +Also, considering the difficulty of the tasks, we introduce regularisation parameters to Equation (3) to control + +the complexity of different branches. That is, + +$$ +\hat {\boldsymbol {\theta}} = \underset {\boldsymbol {\theta}} {\arg \min } \sum_ {i = 1} ^ {N} \| f _ {\boldsymbol {\theta}} \left(\mathbf {s} _ {i}, \mathbf {p} _ {i}\right) - \mathbf {r} _ {i} \| + \lambda_ {c} \| \boldsymbol {\theta} _ {c} \| ^ {2}, \tag {4} +$$ + +where $c \in \{\mathbf{s}, \mathbf{p}, \mathbf{r}\}$ and $\lambda_c$ is a hyper-parameter to control the strength of regularisation for different parts of $M$ . + +We note that the structure of INF is most similar to the network of [20]. However, instead of merging the two input branches by adding their output, we first reduce the output size of these two branches and then concatenate them as the next input. This is because the output of SIREN Layers $X_{l}$ is distributed in $X_{l} \sim \operatorname{Arcsin}(-1,1)$ [31], and adding the outputs would let the output branch treat them equally. Hence we use concatenation to enable the network to learn different weights for the outputs of $\mathbf{p}_i$ and $\mathbf{s}_i$ branches. + +Reconstruction conditioned on sRGB patches. Additionally, it is worth noting that local tone mapping is commonly applied as a favourable part of the ISP pipeline, which is adopted on local images to enhance the special objects (e.g., food, flower, sky). In order to make our method robust to local tone mapping, we split the input image into patches, and then train the INF for each patch. Moreover, inspired by [28], considering that the mapping of each patch and its neighbours are supposed to have strong relationship, we use the samples from each patch together with its eight neighbour patches to train the patch-specific INF. Therefore the mapping in Eq. (2) becomes + +$$ +f _ {\theta}: \left(\mathbf {p} _ {i}; \{s _ {j} \} _ {j \in \mathcal {N} (i)}\right)\rightarrow \mathbf {r} _ {i}. \tag {5} +$$ + +where $\mathcal{N}(i)$ represents neighbour patches of current pixel $\mathbf{p}_i$ . Figure 5 illustrates the training strategy of INF for such patch-wise RAW reconstruction. + +# 4. Experiments + +# 4.1. Experimental Setup + +Baselines. We take three metadata-based RAW reconstruction methods for comparison: RIR [25], SAM [28] and CAM [24]. The RIR method computes the parameters of global sRGB-to-RAW mapping as metadata. The SAM and CAM methods adopt different sampling approaches on RAW images to generate metadata. We follow the uniform sampling as SAM [28], and show that it is enough to get much higher performance using uniform samples with the proposed method. Note that neither RIR nor SAM make their source codes publicly available, and only CAM reports their results reproduced on its own dataset. Therefore we only give a comparison with RIR and SAM on the dataset of CAM by inheriting its results in Table 1 and Figure 6. + +Dataset. We deploy the same dataset of the previous work [24] to test our proposed method, which contains the preprocessed [13] and downsampled (with a downsampling factor of 4) images of three cameras from the NUS dataset [9]—Samsung NX2000, Olympus E-PL6 and Sony SLT-A57. Note that as our method does not require any pretraining step, the training and validation sets are not used in our method. Besides, since our method is able to directly process full-resolution images, we also compare it with CAM [24] on the original NUS dataset [9]. Additionally, as the NUS dataset [9] includes JPEG-compressed sRGB images rendered by real camera ISPs, we also conduct the experiments on these RAW-JPEG image pairs. These results are reported in the supplementary. + +Implementation details. For capture stage, we use the same sampling rate of $1.5\%$ like [24] in all experiments for fair comparison with other metadata-based methods. With respect to reconstruction stage, we split the image into different patches—(228, 272) for Samsung NX2000, (216, 288) for Olympus E-PL6 and (204, 304) for Sony SLT-A57—to make the resolution divisible. The proposed INF is trained by Adam optimizer [15] with $L_{2}$ loss in 500 iterations. We set the initial learning rate to 0.0001, and reduce it to half every 200 iterations. The regularisation parameters are set to $\lambda_{\mathrm{s}} = 0.0001, \lambda_{\mathrm{p}} = 0.1, \lambda_{\mathrm{r}} = 0.001$ to produce the final results. It is worth noting that the reconstruction network is only trained for each specific patch, and will be re-initialized for the training of other patches. + +# 4.2. Results + +Quantitative results. We report the quantitative comparison results against other three baselines in Table 1. It can be observed that our method makes a remarkably improvement in both PSNR and SSIM metrics, especially outperforms the closest competitor CAM [24] by more than $10\mathrm{dB}$ in average PSNR. This indicates the effectiveness of INF to + +modeling the sRGB-to-RAW mapping. + +Qualitative results. We also provide examples of the qualitative comparison in Figure 6, which gives an intuitive explanation of the success of our method. As can be seen from the error maps, our method significantly reduces the reconstructed error on the whole images, especially the saturated regions [26]. We note that recovering saturated regions is the most difficult problem for all RAW reconstruction methods, and is the only case of our method getting large errors $(\geqslant 1\%)$ . We attribute this to the expressive power [38] of INF, which is further discussed in Section 4.4. + +# 4.3. Ablation study. + +RBF vs. INF. We first compare RBF and INF with different input, which is shown in Table 2. As can be seen, the proposed INF outperforms the RBF interpolation [28] whatever the input is. We attribute the main performance boost to the basic structure, i.e. SIREN [31], as INF is able to modelling the sRGB-to-RAW mapping more accurately (59.62dB) only with the pixel value as input. Moreover, compared with RBF, our proposed INF can make further improvement by splitting the input into separate branches and regularize them with different weights. + +The structure of INF. We also conduct experiments on the structure of INF, which is reported in Figure 7. We compare the different activation functions and the layer location where two input branches are merged. It is obviously shown that the INF with sine activation remarkably outperforms the INF with ReLU function, which can be attributed to the expressive power of INRs [38]. As to the location of fusion layer, we find there is no significant difference except the final layer for the INF with sine activation, since it directly maps the merged vector to the output by a fully connection layer without activation. Therefore, we merge the two branches at fourth layer to balance the complexity of each branch1. On the contrary, due to lack of such expressive power, the INF with ReLU requires more layers to model the mapping. Hence the performance increases when the input branch goes deep. + +Regularisation parameters. We further discuss the influence of key setups to INF, which mainly focuses on the regularisation parameters and patch size. As is shown in Table 3, it is obvious that the regularisation parameters play a vital role in the final accuracy. Here we use the grid search with a commonly-used tuner TPE [2] (implemented by [22]) in a grid of $[1, 10^{-1}, \dots, 10^{-5}]$ to determine each parameter. In our experiments, we find the regulation parameters are mainly related to the complexity of camera ISP rather than data dependent, which means we only need + +
DatasetMethodSamsung NX2000Olympus E-PL6Sony SLT-A57
PSNRSSIMPSNRSSIMPSNRSSIM
CAM [24]RIR [25]45.660.993948.420.992451.260.9982
SAM [28]47.030.996249.350.997850.440.9982
CAM [24]49.570.997551.540.998053.110.9985
Ours53.530.998556.710.999158.360.9992
NUS [9]CAM [24]49.510.993752.870.996153.340.9959
Ours61.480.999163.380.999363.800.9995
+ +Table 1. Quantitative comparison with [25, 28, 24]. Here CAM dataset [24] involves downsampling operation (with the factor of 4) on the original NUS dataset [9], which undermines the strength of local mapping. For CAM dataset, we conduct the experiments on their test set to inherit the results of [24]. For NUS dataset, we report the results tested on the whole dataset. The best score for each column is in bold. + +![](images/c0a13b958cb61dd10a4256e3a44811f0b4337e77e3adfe1c74e8b205873a7213.jpg) +Figure 6. Qualitative comparison with [25, 28, 24]. Each two rows from top to bottom are respectively selected from Olympus E-PL6, Samsung NX2000 and Sony SLT-A57. The GT RAW images are processed with a gamma function for better visualization, and the reconstructed RAW images are visualized through error maps. This figure is best viewed in the electronic version. + +to tune it for different ISPs, e.g., digital camera ISP, smart phone ISP, or software ISP. It also means we just need to + +use one or quite a few images for tuning, resulting in a fast tuning process—within 2 hours on a single RTX3090. The + +
MethodInputPSNR
RBF [28]s29.10
RBF [28]p21.71
RBF [28](s, p)49.69
INFs59.62
INFp31.09
INF(s, p)60.91
INF{s; p}62.24
+ +![](images/78a1b604ae63ab7d72b714c8b4adda9fd326c3e2baf78a155478f312dbe90725.jpg) +Figure 7. Ablation study on the structure of INF. The abscissa value represents the layer where we merge the two input branches, e.g., the branches are merged at fourth layer in Figure 3. + +tuned parameters can then be used for all scenes and all other images from the ISPs with similar complexity, hence we use the same parameters for the three reported cameras. + +As shown in Table 3, the proposed INF achieves the best performance at $\lambda_{\mathbf{s}} = 0.0001$ , $\lambda_{\mathbf{p}} = 0.1$ , $\lambda_{\mathbf{r}} = 0.001$ , which indicates the spatial branch is not supposed to be more complex. This is because the output $\mathbf{r}$ has weaker relationship with the coordinates $\mathbf{p}$ than sRGB values $\mathbf{s}$ , and regularisation could help to control the contribution of $\mathbf{s}$ and $\mathbf{p}$ . Therefore when we set $\lambda_{\mathbf{p}} = 0$ , the reconstruction performance dramatically drops down due to inaccurate information from $\mathbf{p}$ is expanded. We believe such intuitive relationship would be helpful to set the initial values for regularisation parameters. Besides, for different ISPs with similar complexity, the results are insensitive to the parameters in a wide range (for example, where all branches are regularized by $\lambda \geqslant 10^{-4}$ ). In this sense, the risk of over-fitting due to these hyper-parameter settings is very small. + +Impact of patch size. Moreover, we find that different size of the patch would lead to an obvious accuracy gap. For the + +Table 2. Ablation study on different reconstruction methods. The best score is in bold and the second in italic. $(\mathbf{s},\mathbf{p})$ refers to combine $\mathbf{s}$ and $\mathbf{p}$ into a single input vector, while $\{\mathbf{s};\mathbf{p}\}$ denotes separating them as different input. For the RBF interpolation, we use the linear RBF kernel as proposed in [28]. + +
Patch size(λs,λp,λr)PSNR
×147.48
×2(0,0,0)49.17
×450.53
×161.13
×2(0.0001,0.001,0.0001)60.38
×457.80
×162.58
×2(0.0001,0.1,0.001)62.24
×459.64
+ +Table 3. Ablation study on the patch size and regularisation parameters. The best score is in bold and the second in italic. For the patch size, $\times 2$ refers to the sizes we report in our main experiments, and the rescale factor of 1, 2, 4 corresponds to the length of patch sides. We note that the accuracy gap less than $0.5\mathrm{dB}$ can not represent an improvement, which may be caused by the different calculations of PSNR and the random initialization of INF. This is further discussed in our supplementary material. + +case of all regularisation parameters set to 0, we find small patch would lead to accuracy reduction. When the regularisation is correctly set, reconstruction on a small patch is more accurate than on a large one. This can be explained by the various local mapping functions in different image regions, which means modelling the local mapping for a small patch would be more accurate than a larger one. Note that though reducing the patch size would be helpful to the reconstruction accuracy, it would take much more time to recover an image. Therefore, we compromise to use the patch size around 200-300 in the main experiments. + +# 4.4. Discussion + +Why INF works in sRGB-to-RAW mapping. We first rethink the RBF interpolation [3] used in SAM [28], by which the sRGB-to-RAW mapping is built in the following form: + +$$ +\hat {y} = \mathbf {M} ^ {\text {t e s t}} \mathbf {M} ^ {- 1} y \tag {6} +$$ + +where $\mathbf{M}$ denotes the RBF interpolation matrix with linear kernel function [27], and test refers to the pixels that need to be reconstructed. As discussed in [32], the output of an MLP can be approximated as: + +$$ +\hat {y} ^ {(t)} = \mathbf {K} ^ {\text {t e s t}} \mathbf {K} ^ {- 1} \left(\mathbf {I} - e ^ {- \eta \mathbf {K} t}\right) y \tag {7} +$$ + +where $\mathbf{K}$ denotes the NTK matrix [12] and $t$ refers to training iterations. In our experiments, an MLP (typical with ReLU activation) tends to behave like RBF when $t$ is large enough (e.g., $t > 1000$ ), which meets the results of Equation (6) and Equation (7). The reason why RBF interpolation works can be attributed to that sRGB-RAW mapping + +
Method×4×8×16
Bicubic2.624.196.68
PixTransform [20]3.513.944.94
INF2.503.065.23
INF with ReLU2.392.693.57
+ +Table 4. Quantitative comparison on guided depth superresolution. We use the Middlebury dataset provided by [33] for testing. The results are reported in terms of average RMSE. + +is piecewise smooth [27], but there still exists errors when using a linear kernel function, which is also suffered by MLPs with ReLU. On the other hand, Sitzmann et al. [31] demonstrate that MLPs with periodic activation functions (i.e., SIRENs) are more suitable for representing complicated signals and their derivatives compared with the typical MLP architectures. SIRENs have the advantage of representing the derivatives of the signal [31], which is more helpful to reconstruct fine details. Therefore, recovering RAW images by SIRENs can effectively improve the performance, which is also proved by Table 2. + +Limitations and future research. One of the unresolved problem of our method is that the regularisation parameters of weight decay require pre-experiments to determine. These three parameters play a significant role to control the final result by limiting the complexity of network. In our experiments, we find the parameters are strongly related to the correlation of input and output data. For example, compared with the coordinates, sRGB values have a more obvious relationship with the RAW values, where the regularisation parameters are respectively set to $\lambda_{p} = 0.1$ and $\lambda_{s} = 0.0001$ . Such correlation is suitable for modelling with another neural network, and would lead to the performance improvement with optimized parameters. Therefore, learning the regularisation parameters from the input and output is supposed to be a valuable topic of future work. + +# 4.5. Applications + +Guided depth super-resolution As mentioned in Section 2, our framework is related to guided super-resolution, therefore we also provide an experiment to illustrate the applicability of our framework. Considering that our method is only self-supervised without any priori knowledge, we give a comparison with another self-supervised method PixTransform [20]. As shown in Table 4, INF with ReLU outperforms other methods, which indicates the ReLU activation is more suitable for this task than the sine activation. The reason for ReLU working better can be attributed to the sharp changes of the depth near the object edges, as ReLU is piecewise and discontinuous, which fits this depth disconti + +![](images/1c4a8c3a74025fc666c754094d2dbf1e8474ebcf49fc335ea6b590a6ea2314ca.jpg) +Source + +![](images/1219412462c45f3b2453bf8d2b1ff29cb2970be19a8ef7d7a3ef3031c0e52b06.jpg) +Guide + +![](images/57788384b97aebde3c91e7b824c0d49d6b3910a9ea1b979c37252e274f3c47fe.jpg) + +![](images/0945424c2419cee1bcb58fd28cc0a1c6164cda3bd61709a9b1f2c1eef896ecb7.jpg) +PixTransform [20] + +![](images/e61d52c2b98b2ff1e9d0b694935e7e866d83ef0a3e7cca024c36e421cd6e02ca.jpg) +INF + +![](images/d549c9409a9663144e61a9a339cb1deb72c4f2709220ba774a0acbe089abd2b3.jpg) +Target +INF with ReLU +Figure 8. Qualitative comparison on guided depth super-resolution. + +nuity. In contrast, the discontinuity in RAW-to-sRGB mapping is moderate and the smooth sine activation function is a better option. We also provide qualitative comparison in Figure 8. which shows that the result of INF, though does not achieve higher accuracy, obtains better visual quality compared with the result of INF with ReLU. + +Low-light image enhancement. As a potential application of RAW reconstruction, we test our method on low-light image enhancement (LLIE) task. We find that the LLIE task can be simplified into a linear degradation problem (e.g., enlarging the pixel values by the same multiple) on the reconstructed RAW images, which achieves equal visual quality compared with the state-of-the-art deep learning-based approaches [11, 34, 21]. The results are reported in our supplementary. Note that such comparison is merely aimed to illustrate the effectiveness of executing low-light enhancement on reconstructed RAW images. + +# 5. Conclusion + +We propose a method for recovering the RAW image from the sRGB counterpart with assistance of additional metadata, which is sampled from the RAW image at capture time. We introduce the implicit neural function (INF) to remarkably improve the reconstruction accuracy (10 dB average PSNR) only with uniform sampling. We prove that the structure of INF is beneficial to merging the information from both pixel values and coordinates. Further experiments indicate that our framework is also suitable for the task of guided super-resolution. + +Acknowledgement. This work was partly supported by National Key R&D Program of China (Grant No. 2021YFD1400200) and National Natural Science Foundation of China (Grant No. 62088101). + +# References + +[1] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19323-19332, 2022. 2 +[2] James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization. Advances in neural information processing systems, 24, 2011. 5 +[3] Marco Evangelos Biancolini. Fast radial basis functions for engineering applications. Springer, 2017. 2, 7 +[4] Ayan Chakrabarti, Daniel Scharstein, and Todd E Zickler. An empirical camera model for internet color vision. In BMVC, volume 1, page 4. CiteSeer, 2009. 1, 2 +[5] Ayan Chakrabarti, Ying Xiong, Baochen Sun, Trevor Darrell, Daniel Scharstein, Todd Zickler, and Kate Saenko. Modeling radiometric uncertainty for vision with tonemapred color images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11):2185-2198, 2014. 1, 2 +[6] Bowei Chen, Tiancheng Zhi, Martial Hebert, and Srinivasa G Narasimhan. Learning continuous implicit representation for near-periodic patterns. arXiv preprint arXiv:2208.12278, 2022. 2 +[7] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 1 +[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2 +[9] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 5, 6 +[10] Marcos V Conde, Steven McDonagh, Matteo Maggioni, Ales Leonardis, and Eduardo Pérez-Pellitero. Model-based image signal processors via learnable dictionaries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 481-489, 2022. 1, 2 +[11] Chunle Guo Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1780-1789, June 2020. 8 +[12] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 7 +[13] Hakki Can Karaimer and Michael S Brown. A software platform for manipulating the camera imaging pipeline. In European Conference on Computer Vision, pages 429-444. Springer, 2016. 5 + +[14] Seon Joo Kim, Hai Ting Lin, Zheng Lu, Sabine Süssstrunk, Stephen Lin, and Michael S Brown. A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(12):2289-2302, 2012. 1, 2 +[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, 2015. 5 +[16] Shayan Kousha, Ali Maleky, Michael S Brown, and Marcus A Brubaker. Modeling srgb camera noise with normalizing flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17463-17471, 2022. 1 +[17] Zhihao Li, Si Yi, and Zhan Ma. Rendering nighttime image via cascaded color and brightness compensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 897–905, 2022. 1 +[18] Hai Ting Lin, Zheng Lu, Seon Joo Kim, and Michael S Brown. Nonuniform lattice regression for modeling the camera imaging pipeline. In European Conference on Computer Vision, pages 556-568. Springer, 2012. 1, 2 +[19] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image hdr reconstruction by learning to reverse the camera pipeline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1651-1660, 2020. 1 +[20] Riccardo de Lutio, Stefano D'aronco, Jan Dirk Wegner, and Konrad Schindler. Guided super-resolution as pixel-to-pixel transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8829-8837, 2019. 3, 4, 8 +[21] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 8 +[22] Microsoft. Neural Network Intelligence, 2021. 5 +[23] Seonghyeon Nam and Seon Joo Kim. Modelling the scene dependent imaging in cameras with a deep neural network. In Proceedings of the IEEE International Conference on Computer Vision, pages 1717-1725, 2017. 1, 2 +[24] Seonghyeon Nam, Abhijith Punnappurath, Marcus A Brubaker, and Michael S Brown. Learning srgb-to-raw-rgb de-rendering with content-aware metadata. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17704–17713, 2022. 1, 2, 4, 5, 6 +[25] Rang MH Nguyen and Michael S Brown. Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1655-1663, 2016. 1, 2, 5, 6 +[26] Rang MH Nguyen and Michael S Brown. Raw image reconstruction using a self-contained srgb-jpeg image with small memory overhead. International journal of computer vision, 126(6):637-650, 2018. 1, 2, 5 +[27] Abhijith Punnappurath and Michael S Brown. Learning raw image reconstruction-aware deep image compressors. IEEE + +transactions on pattern analysis and machine intelligence, 42(4):1013-1019, 2019. 3, 7, 8 +[28] Abhijith Punnappurath and Michael S Brown. Spatially aware metadata for raw reconstruction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 218-226, 2021. 1, 2, 3, 4, 5, 6, 7 +[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2 +[30] Guy Shacht, Dov Danon, Sharon Fogel, and Daniel Cohen-Or. Single pair cross-modality super resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6378–6387, 2021. 3 +[31] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2, 3, 4, 5, 8 +[32] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 2, 7 +[33] Jiaxiang Tang, Xiaokang Chen, and Gang Zeng. Joint implicit image function for guided depth super-resolution. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4390-4399, 2021. 3, 8 +[34] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2604-2612, 2022. 8 +[35] Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neuralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1526-1535, 2022. 2 +[36] Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6287-6296, 2021. 1, 2 +[37] Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248-4257, 2021. 2 +[38] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 3, 5 +[39] Xuaner Zhang, Qifeng Chen, Ren Ng, and Vladlen Koltun. Zoom to learn, learn to zoom. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3762-3770, 2019. 1 \ No newline at end of file diff --git a/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/images.zip b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3fa0f05e899c133bc90751b305f54e87586ab90c --- /dev/null +++ b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94ae82296d02c03b6f56608be04e4eb4f033bff0e074be783f7af64e7319236a +size 697572 diff --git a/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/layout.json b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a955f98bd39aef6ff35efaec7e3218524c09a946 --- /dev/null +++ b/2023/Metadata-Based RAW Reconstruction via Implicit Neural Functions/layout.json @@ -0,0 +1,8296 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 507, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 507, + 120 + ], + "type": "text", + "content": "Metadata-Based RAW Reconstruction via Implicit Neural Functions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "spans": [ + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "content": "Leyi Li" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "content": " Huijie Qiao" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "content": " Qi Ye" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "content": " Qinmin Yang" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "content": "Chinese Academy of Sciences " + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "content": "Key Lab of CS&AUS of Zhejiang Province " + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "inline_equation", + "content": "\\{lileyi, qi.ye, qmyang\\} @zju.edu.cn" + }, + { + "bbox": [ + 50, + 141, + 538, + 185 + ], + "type": "text", + "content": " qiaohj@ioz.ac.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 239, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 491 + ], + "type": "text", + "content": "Many low-level computer vision tasks are desirable to utilize the unprocessed RAW image as input, which remains the linear relationship between pixel values and scene radiance. Recent works advocate to embed the RAW image samples into sRGB images at capture time, and reconstruct the RAW from sRGB by these metadata when needed. However, there still exist some limitations in making full use of the metadata. In this paper, instead of following the perspective of sRGB-to-RAW mapping, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values. With this novel formulation, we propose to reconstruct the RAW image with an implicit neural function, which achieves significant performance improvement (more than 10dB average PSNR) only with the uniform sampling. Compared with most deep learning-based approaches, our method is trained in a self-supervised way that requiring no pre-training on different camera ISPs. We perform further experiments to demonstrate the effectiveness of our method, and show that our framework is also suitable for the task of guided super-resolution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 288, + 715 + ], + "type": "text", + "content": "Low-level computer vision tasks benefit a lot from the scene-referred RAW images [7, 39, 19, 17, 16], which is rendered to the display-referred standard RGB (sRGB) images via camera image signal processors (ISPs). Compared with sRGB images, typical RAW images has the advantages of linear relationship between pixel values and scene radiance, as well as higher dynamic range. However, RAW images occupy obviously more memory than the sRGB images in common format like JPEG, which is unfavourable for transferring and sharing. Moreover, since most display and printing devices are designed for images stored and shared in sRGB format, it is inconvenient to directly replace sRGB with RAW. Consequently, mapping sRGB images back to their RAW counterparts, which is also called RAW reconstruction, is regarded as the appropriate way to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 212, + 383, + 269 + ], + "blocks": [ + { + "bbox": [ + 309, + 212, + 383, + 269 + ], + "lines": [ + { + "bbox": [ + 309, + 212, + 383, + 269 + ], + "spans": [ + { + "bbox": [ + 309, + 212, + 383, + 269 + ], + "type": "image", + "image_path": "31a015de6f98ea740612b792ad93ecb9cf0bd51955312dbb18e2bbe775b307e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 384, + 212, + 459, + 269 + ], + "blocks": [ + { + "bbox": [ + 384, + 212, + 459, + 269 + ], + "lines": [ + { + "bbox": [ + 384, + 212, + 459, + 269 + ], + "spans": [ + { + "bbox": [ + 384, + 212, + 459, + 269 + ], + "type": "image", + "image_path": "c8d4334f820bffa702019a34bfb31ca4a28364fce885a43e7eee9bb4add11661.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 460, + 212, + 535, + 269 + ], + "blocks": [ + { + "bbox": [ + 460, + 212, + 535, + 269 + ], + "lines": [ + { + "bbox": [ + 460, + 212, + 535, + 269 + ], + "spans": [ + { + "bbox": [ + 460, + 212, + 535, + 269 + ], + "type": "image", + "image_path": "a2e9c957293d37b991170c1821173687f873d0b839aac09bb86ac70703554e09.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 283, + 384, + 339 + ], + "blocks": [ + { + "bbox": [ + 321, + 272, + 372, + 282 + ], + "lines": [ + { + "bbox": [ + 321, + 272, + 372, + 282 + ], + "spans": [ + { + "bbox": [ + 321, + 272, + 372, + 282 + ], + "type": "text", + "content": "sRGB (input)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 283, + 384, + 339 + ], + "lines": [ + { + "bbox": [ + 309, + 283, + 384, + 339 + ], + "spans": [ + { + "bbox": [ + 309, + 283, + 384, + 339 + ], + "type": "image", + "image_path": "fca0e4de21f1902af753f138645bcb5eaf85b9d422cce49a476af77cd3769bcf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 342, + 367, + 353 + ], + "lines": [ + { + "bbox": [ + 325, + 342, + 367, + 353 + ], + "spans": [ + { + "bbox": [ + 325, + 342, + 367, + 353 + ], + "type": "text", + "content": "RAW (GT)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 363, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 452 + ], + "type": "text", + "content": "Figure 1. As RAW images are beneficial to many low-level computer vision tasks, we aim to reconstruct the RAW image from the corresponding sRGB image with the assistance of extra metadata. In this figure, the reconstructed RAW images are visualized through error maps. As can be seen, our method remarkably outperforms other related methods with the improvement of more than " + }, + { + "bbox": [ + 304, + 363, + 545, + 452 + ], + "type": "inline_equation", + "content": "10\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 363, + 545, + 452 + ], + "type": "text", + "content": " PSNR. We owe this performance boost to the effectiveness of our proposed implicit neural function (INF)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 384, + 283, + 459, + 339 + ], + "blocks": [ + { + "bbox": [ + 405, + 272, + 438, + 282 + ], + "lines": [ + { + "bbox": [ + 405, + 272, + 438, + 282 + ], + "spans": [ + { + "bbox": [ + 405, + 272, + 438, + 282 + ], + "type": "text", + "content": "RIR [25]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 283, + 459, + 339 + ], + "lines": [ + { + "bbox": [ + 384, + 283, + 459, + 339 + ], + "spans": [ + { + "bbox": [ + 384, + 283, + 459, + 339 + ], + "type": "image", + "image_path": "d90cb0ea54e8c45ed93a019b6dc10d41143a66a3c0986585e85c1396252d4f5c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 342, + 441, + 353 + ], + "lines": [ + { + "bbox": [ + 402, + 342, + 441, + 353 + ], + "spans": [ + { + "bbox": [ + 402, + 342, + 441, + 353 + ], + "type": "text", + "content": "CAM [24]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 460, + 283, + 535, + 339 + ], + "blocks": [ + { + "bbox": [ + 477, + 272, + 515, + 282 + ], + "lines": [ + { + "bbox": [ + 477, + 272, + 515, + 282 + ], + "spans": [ + { + "bbox": [ + 477, + 272, + 515, + 282 + ], + "type": "text", + "content": "SAM [28]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 283, + 535, + 339 + ], + "lines": [ + { + "bbox": [ + 460, + 283, + 535, + 339 + ], + "spans": [ + { + "bbox": [ + 460, + 283, + 535, + 339 + ], + "type": "image", + "image_path": "e4ec6ff7eefc56932a3671d30aafe47cae63c53e79063eb37c652587f8031fd0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 343, + 506, + 352 + ], + "lines": [ + { + "bbox": [ + 487, + 343, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 487, + 343, + 506, + 352 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 473, + 543, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 543, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 543, + 485 + ], + "type": "text", + "content": "utilize the advantage of RAW data [23, 25, 36, 10, 28, 24]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 486, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 665 + ], + "type": "text", + "content": "Early RAW reconstruction methods focus on building standard models to reverse ISPs, which is parameterized by either explicit functions [4, 18, 14, 5] or neural networks [23, 36, 10]. However, these approaches are faced with the same issue that a parameterized model is only suitable for a specific ISP. Meanwhile, a series of methods [25, 26, 28, 24] propose to overcome this problem by embedding extra metadata into sRGB images at capture time. For such methods, the main challenge is to improve the accuracy with lower metadata generation cost. RIR [25] implements complex optimization algorithm to estimate the global mapping parameters as metadata, but suffers high computational cost. SAM [28] adopts a uniform sampling on RAW images to generate the metadata, which is further replaced with a sampler network by CAM [24]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "For the metadata-based methods of SAM [28] and CAM [24], the embedded RAW samples stores partial information of ISPs which helps to reconstruct the RAW images better; also, by conditioning the reconstruction algorithm on" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18196" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "these metadata, the recovery of the RAW data turns into a conditional mapping function instead of a function fitted to a specific case, enabling the potential to achieve better generalization. Therefore, we adopt this strategy in the paper." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 275 + ], + "type": "text", + "content": "Despite the progress that SAM [28] and CAM [24] have made, there still exist some limitations for the metadata-based RAW reconstruction methods. SAM utilizes RBF interpolation [3], the main idea of which is to calculate the difference between sampling and target points by a kernel function. However, a fixed kernel function lacks the flexibility to model various sRGB-to-RAW mappings. CAM directly uses a neural network for reconstruction but requires pre-training on pairs of sRGB and raw data from different types of ISPs. Also, we observe that the results of these methods fail to recover the saturated regions [26] (i.e., pixels with any channel value close to the maximum), as is shown in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 275, + 288, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 288, + 502 + ], + "type": "text", + "content": "To address the limitation, we propose a two-way RAW reconstruction algorithm based on an implicit neural function (INF). Previously, RAW reconstruction is formulated as mapping a sRGB image and the metadata to its RAW image. In this paper, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values conditioned on the corresponding sRGB values, i.e. an implicit function. With this novel formulation, we can also decompose the problem into two aspects: a mapping function from the sRGB values to the corresponding raw values; a super-resolution function to interpolate the RAW image from the sparse samples. We observe that the super-resolution part usually exhibits much higher errors, indicating the latter a more challenging task. Accordingly, two branches are designed for each task inside an implicit neural network and the hyper-parameters for these branches are tuned to accommodate the difficulty of the tasks. Also, notice that with this formulation, the network can be trained in a self-supervised way, without the need of corresponding RAW images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 503, + 253, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 503, + 253, + 514 + ], + "spans": [ + { + "bbox": [ + 59, + 503, + 253, + 514 + ], + "type": "text", + "content": "Our contribution can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 521, + 286, + 644 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 521, + 286, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 521, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 59, + 521, + 286, + 569 + ], + "type": "text", + "content": "- We reformulate the RAW reconstruction problem as a RAW image approximation problem that learns the 2D-to-RAW mapping of image coordinates to RAW values conditioned on its sRGB image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 577, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 577, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 59, + 577, + 286, + 601 + ], + "type": "text", + "content": "- We decompose the reconstruction into two aspects and design the implicit neural network accordingly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 609, + 286, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 609, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 59, + 609, + 286, + 644 + ], + "type": "text", + "content": "- We conduct extensive experiments on different cameras and demonstrate our algorithm outperforms existing work significantly." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 654, + 134, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 134, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 134, + 667 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Blind RAW reconstruction. Early works of RAW reconstruction were blindly taken sRGB images as input without extra metadata. Since the processing stages of modern ISPs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": "are designed more complicated, more complex models are proposed [4, 18, 14, 5]. However, these methods suffer the inconvenient calibration procedures that need to be repeated on each camera or even each camera setting. Deep learning-based methods (e.g., [23, 36]), similarly, are faced with the parallel issues of camera-specific models, which required abundant training data captured for each camera. Recent methods that simulate ISP architectures by assuming a classical set of ISP operations, such as [37, 10], cannot handle different camera settings as they take fixed parameters (e.g., gamma correction) and ISP length as a priori knowledge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 207, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 495 + ], + "type": "text", + "content": "RAW reconstruction with metadata. Compared with blind RAW reconstruction, a series of recent methods [25, 26, 28, 24] benefit from additional metadata embedded into sRGB images at capture time. Nguyen and Brown [25, 26] propose to extract and store the necessary parameters for recovering a RAW image from the sRGB counterpart. These parameters model the specific sRGB-to-RAW mapping and are restricted to a small memory (e.g., 128KB). However, their main algorithm deployed on the device has high computational cost and only considers the global tone mapping. Punnappurath and Brown [28], on the contrary, implement a uniform sampling on RAW images to save as metadata. They employ radial basis function (RBF) interpolation using sRGB pixel values and coordinates as input to reconstruct the RAW values. Nevertheless, the fixed type of RBF interpolation does not take full use of the information of sampled data. A further work [24] improves the effectiveness of sampling and recovering by taking a U-Net architecture [29] as both sampler and reconstruction networks, but their method requires to run a deep neural network on the device, which would lead to high computational cost and additional memory cost to save the pixel positions into metadata. We follow the uniform sampling in [28], but improve the reconstruction performance by INF." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Implicit neural representation. Implicit neural representation (INR) has recently been introduced to represent 2D images and 3D objects using coordinate-based multi-layer perceptron (MLP). To overcome the problem that conventional MLPs are incapable of representing high-frequency details of signals, two methods have been proposed. Sitzmann et al. [31] introduce SIREN, which replace ReLU activation with periodic activation (e.g., sine function). They demonstrate that the representation power of SIREN comes from the derivation invariance of sine function, and provide a number of potential applications and future works. A concurrent work [32] leverages random Fourier feature mapping on input coordinates to enable the MLP with ReLU to learn high-frequency details. It indicates that a Fourier feature mapping can beneficially address the spectral bias of a conventional MLP. Based on these breakthroughs, INR has been successfully adopted in various tasks [8, 1, 6, 35]. In this work, the proposed INF is based on [31], which shows" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18197" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 261, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 261, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 261, + 84 + ], + "type": "text", + "content": "remarkably improvement in reconstruction accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 87, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 87, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 87, + 287, + 243 + ], + "type": "text", + "content": "Guided super-resolution. Another topic similar to RAW reconstruction with metadata is guided image superresolution, which aims to convert a low-resolution (LR) source image to the high-resolution (HR) target with a guided HR image [20, 30, 33]. The difference of these two topics is that RAW reconstruction has obvious mapping of pixel values between two HR images. On this aspect, PixTransform [20] is more related to our method. It uses an MLP with ReLU to learn the pixel-to-pixel mapping of guided HR image to target HR image, where the LR image is treated like the metadata in RAW reconstruction. Our method differs in the MLP layer, where we use SIREN layer [31] to enhance the expressive power [38]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 251, + 103, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 103, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 103, + 263 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 271, + 280, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 280, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 280, + 285 + ], + "type": "text", + "content": "3.1. Problem Formulation with Implicit Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " denote an sRGB image and the corresponding RAW image, respectively. Since " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " is converted to " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " through a series of operations in the camera ISP, classical methods aim to build a model " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " to map " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "R = g(S)" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": ". However, the model " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " is specific to a camera ISP or even a set of ISP parameters. Metadata-based methods introduce extra data " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " for RAW reconstruction, i.e., " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "R = f(S, M)" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " is supposed to be a general function, and " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " includes the image-specific information generated at capture time, which is commonly implemented by sampling on the RAW images for low computation cost. Hence " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "M = \\{\\mathbf{p}_i, \\mathbf{r}_i\\}_{i=0}^N" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i = (x_i, y_i)" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " is the coordinate of a sampled RAW pixel and " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i = (r_{R_i}, r_{G_i}, r_{B_i})" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " its RAW value. " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " represents the number of sampled RAW pixels and " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 290, + 287, + 469 + ], + "type": "text", + "content": " sampling one." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": "In contrast to previous work that establishes a mapping from RGB to RGB values, " + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "inline_equation", + "content": "f:(r,s)\\to r" + }, + { + "bbox": [ + 46, + 470, + 287, + 518 + ], + "type": "text", + "content": ", we reformulate the problem as a mapping from the 2D coordinates to the RAW values conditioning on the sRGB value" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 524, + 287, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 524, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 135, + 524, + 287, + 537 + ], + "type": "interline_equation", + "content": "f: (p; s) \\rightarrow r, \\tag {1}", + "image_path": "c418efc44aec3caaa4926c3abcea667e48109c96dafde12f23d8b5c6d74ce83c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 543, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 543, + 287, + 579 + ], + "type": "inline_equation", + "content": "r, s, p" + }, + { + "bbox": [ + 46, + 543, + 287, + 579 + ], + "type": "text", + "content": " denote the variables for the RAW value, the RGB value and coordinates. They can either represent the value for a pixel or a patch, or a set of coordinates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": "Specifically, for the sRGB image, we only use the the sRGB value " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "(s_{R_i}, s_{G_i}, s_{B_i})" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": ", queried by " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "S(\\mathbf{p}_i)" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": ". Therefore we aim to learn an implicit neural function " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": " taking a coordinate " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": " and its sRGB value and output " + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 128, + 634, + 287, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 634, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 128, + 634, + 287, + 647 + ], + "type": "interline_equation", + "content": "f _ {\\theta}: \\left(\\mathbf {p} _ {i}; \\mathbf {s} _ {i}\\right)\\rightarrow \\mathbf {r} _ {i}. \\tag {2}", + "image_path": "845c16ecbac143e11d03ab83779f0fa2bad1af07fb78503e0a997234847757a0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "type": "text", + "content": "To learn the function, we define the loss function below to find the best configuration for the parameters " + }, + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 653, + 287, + 677 + ], + "type": "inline_equation", + "content": "f" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 683, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 683, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 110, + 683, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\| f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {s} _ {i}, \\mathbf {p} _ {i}\\right) - \\mathbf {r} _ {i} \\|, \\tag {3}", + "image_path": "7ae5be118938694f8bf26648a76fe69361476850e345a34be495a9954d369630.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 387, + 148 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 387, + 148 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 387, + 148 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 387, + 148 + ], + "type": "image", + "image_path": "044f91c70c87888b6d885fa68ded9dd79a496b72e01e155b78c5b7a8de67c3ba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 150, + 370, + 161 + ], + "lines": [ + { + "bbox": [ + 328, + 150, + 370, + 161 + ], + "spans": [ + { + "bbox": [ + 328, + 150, + 370, + 161 + ], + "type": "text", + "content": "RAW (GT)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 388, + 71, + 464, + 148 + ], + "blocks": [ + { + "bbox": [ + 388, + 71, + 464, + 148 + ], + "lines": [ + { + "bbox": [ + 388, + 71, + 464, + 148 + ], + "spans": [ + { + "bbox": [ + 388, + 71, + 464, + 148 + ], + "type": "image", + "image_path": "4145960c48f539469765404e92c981b40e8a5afd9c91ec4bbc9ed34f6c3d9a90.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 151, + 454, + 162 + ], + "lines": [ + { + "bbox": [ + 398, + 151, + 454, + 162 + ], + "spans": [ + { + "bbox": [ + 398, + 151, + 454, + 162 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 398, + 151, + 454, + 162 + ], + "type": "inline_equation", + "content": "f:\\mathbf{s}_i\\to \\mathbf{r}_i" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 466, + 71, + 542, + 148 + ], + "blocks": [ + { + "bbox": [ + 466, + 71, + 542, + 148 + ], + "lines": [ + { + "bbox": [ + 466, + 71, + 542, + 148 + ], + "spans": [ + { + "bbox": [ + 466, + 71, + 542, + 148 + ], + "type": "image", + "image_path": "1d2e7b2a0d1ae35636404deff0c8b44099a426ada025f67cffd6d0565f79858d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 151, + 533, + 162 + ], + "lines": [ + { + "bbox": [ + 474, + 151, + 533, + 162 + ], + "spans": [ + { + "bbox": [ + 474, + 151, + 533, + 162 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 474, + 151, + 533, + 162 + ], + "type": "inline_equation", + "content": "f:\\mathbf{p}_i\\to \\mathbf{r}_i" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 310, + 241, + 541, + 393 + ], + "blocks": [ + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "type": "text", + "content": "Figure 2. Illustration of the two aspects of reconstruction. It can be observed that the reconstruction from (a) pixel value mapping is much more accurate than (b) spatial super-resolution. Our method utilizes the information from both " + }, + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "type": "text", + "content": ", but give a different constraint to them to control their impacts." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 241, + 541, + 393 + ], + "lines": [ + { + "bbox": [ + 310, + 241, + 541, + 393 + ], + "spans": [ + { + "bbox": [ + 310, + 241, + 541, + 393 + ], + "type": "image", + "image_path": "b81fb7584acee094625535f4ed1a6aebff0de0e8e4f40b5ec85659a747b454a2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 403, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 459 + ], + "type": "text", + "content": "Figure 3. The network structure of our proposed implicit neural function (INF). The inputs are divided into sRGB values and coordinates, as their branches are regularized with different parameters of weight decay. Each cube represents the output of each linear layer with sine activation proposed in [31]." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 481, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 578 + ], + "type": "text", + "content": "The overview of our method is shown in Figure 4. As mentioned above, considering the on-device computational cost, we follow SAM [28] to implement a uniform sampling on the RAW image to get the metadata " + }, + { + "bbox": [ + 304, + 481, + 545, + 578 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 481, + 545, + 578 + ], + "type": "text", + "content": ". For the reconstruction stage, we implement the same sampling on the sRGB image, which is used to train the implicit neural function (INF) together with " + }, + { + "bbox": [ + 304, + 481, + 545, + 578 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 481, + 545, + 578 + ], + "type": "text", + "content": ". The sRGB image is then processed by the trained INF to reconstruct the RAW image." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 586, + 469, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 469, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 469, + 598 + ], + "type": "text", + "content": "3.2. Two-way Raw Reconstruction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "We note that Equation (2) contains the reconstruction from two aspects: (a) a pixel values mapping (i.e., " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "f:\\mathbf{s}_i\\to \\mathbf{r}_i" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": ") and (b) a spatial super-resolution (i.e., " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "f:\\mathbf{p}_i\\rightarrow \\mathbf{r}_i" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "). Since " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " will cost additional memory, the sampling rate is required to be constrained (e.g., no more than 1.5%), which limits the accuracy of (b). On the contrary, the sRGB-to-RAW mapping is piece-wise smooth [27], hence only such small amount of samples is able to model the whole mapping function theoretically. We provide an example in Fig-" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18198" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 542, + 228 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 542, + 228 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 542, + 228 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 542, + 228 + ], + "type": "image", + "image_path": "36658c6332cb8f442cdc43c5d757fc25a22f9ff8e998466eb5a247db73c24c1d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 298 + ], + "type": "text", + "content": "Figure 4. Overview of our proposed RAW reconstruction method. At capture stage, we follow [28] to implement an uniform sampling on the RAW image to generate the metadata, which does not require to store the pixel positions. The metadata is then embedded into the sRGB image for storage and transmission. At reconstruction stage, the metadata together with corresponding sRGB samples are used to train the implicit neural functions (INFs), which then convert the sRGB image to the RAW. Compared with the state-of-the-art metadata-based RAW reconstruction approach [24], our method achieves both lower on-device computational cost and much higher reconstruction accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 318, + 282, + 456 + ], + "blocks": [ + { + "bbox": [ + 52, + 318, + 282, + 456 + ], + "lines": [ + { + "bbox": [ + 52, + 318, + 282, + 456 + ], + "spans": [ + { + "bbox": [ + 52, + 318, + 282, + 456 + ], + "type": "image", + "image_path": "373b0b37cca2ffd48117d60a8bc3c977db2b78c6462083509d144e255e2f2e67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 468, + 288, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 288, + 514 + ], + "type": "text", + "content": "Figure 5. Illustration of the patch-specific INF. Inspired by [28], we split the image into patches, and train the INF for each patch with its neighbours to make it robust to the local mapping. We further discuss the impact of patch-size in the ablation study." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 540, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 564 + ], + "type": "text", + "content": "ure 2 to intuitively illustrate this phenomenon. Therefore (a) is supposed to be more significant than (b) in " + }, + { + "bbox": [ + 46, + 540, + 287, + 564 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 540, + 287, + 564 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 567, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 686 + ], + "type": "text", + "content": "Based on the analysis of the sub-tasks, we design an implicit neural network consisting of two branches as shown in Figure 3. INF is an MLP structure built with SIREN layers [31], where the inputs are separated into two branches and then concatenated into the output branch. Each branch contains 4 layers, with 256 channels for the hidden layer. The output channels of two input branches are reduced to half, which are then concatenated into the output branch. Here " + }, + { + "bbox": [ + 46, + 567, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 567, + 287, + 686 + ], + "type": "text", + "content": " refers to the network weights, and Equation (4) can then be solved by training the INF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Also, considering the difficulty of the tasks, we introduce regularisation parameters to Equation (3) to control" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 316, + 489, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 489, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 489, + 328 + ], + "type": "text", + "content": "the complexity of different branches. That is," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 334, + 545, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 334, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 321, + 334, + 545, + 367 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\theta}} = \\underset {\\boldsymbol {\\theta}} {\\arg \\min } \\sum_ {i = 1} ^ {N} \\| f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {s} _ {i}, \\mathbf {p} _ {i}\\right) - \\mathbf {r} _ {i} \\| + \\lambda_ {c} \\| \\boldsymbol {\\theta} _ {c} \\| ^ {2}, \\tag {4}", + "image_path": "42ccc0a3e2221dab01d2aedcfebc020273585dc9d208b27cf9ebcd9e3a77d541.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "inline_equation", + "content": "c \\in \\{\\mathbf{s}, \\mathbf{p}, \\mathbf{r}\\}" + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\lambda_c" + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "text", + "content": " is a hyper-parameter to control the strength of regularisation for different parts of " + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "text", + "content": "We note that the structure of INF is most similar to the network of [20]. However, instead of merging the two input branches by adding their output, we first reduce the output size of these two branches and then concatenate them as the next input. This is because the output of SIREN Layers " + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "text", + "content": " is distributed in " + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "inline_equation", + "content": "X_{l} \\sim \\operatorname{Arcsin}(-1,1)" + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "text", + "content": " [31], and adding the outputs would let the output branch treat them equally. Hence we use concatenation to enable the network to learn different weights for the outputs of " + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 305, + 397, + 546, + 505 + ], + "type": "text", + "content": " branches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 508, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 651 + ], + "type": "text", + "content": "Reconstruction conditioned on sRGB patches. Additionally, it is worth noting that local tone mapping is commonly applied as a favourable part of the ISP pipeline, which is adopted on local images to enhance the special objects (e.g., food, flower, sky). In order to make our method robust to local tone mapping, we split the input image into patches, and then train the INF for each patch. Moreover, inspired by [28], considering that the mapping of each patch and its neighbours are supposed to have strong relationship, we use the samples from each patch together with its eight neighbour patches to train the patch-specific INF. Therefore the mapping in Eq. (2) becomes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 368, + 658, + 545, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 658, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 368, + 658, + 545, + 672 + ], + "type": "interline_equation", + "content": "f _ {\\theta}: \\left(\\mathbf {p} _ {i}; \\{s _ {j} \\} _ {j \\in \\mathcal {N} (i)}\\right)\\rightarrow \\mathbf {r} _ {i}. \\tag {5}", + "image_path": "e4e351b50dad94c728f254a6e1a7d3fb4215ccf612df99ce6bae84df63c0c952.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": " represents neighbour patches of current pixel " + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": ". Figure 5 illustrates the training strategy of INF for such patch-wise RAW reconstruction." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18199" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 163, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 163, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 163, + 105 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 113, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 113, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 113, + 287, + 269 + ], + "type": "text", + "content": "Baselines. We take three metadata-based RAW reconstruction methods for comparison: RIR [25], SAM [28] and CAM [24]. The RIR method computes the parameters of global sRGB-to-RAW mapping as metadata. The SAM and CAM methods adopt different sampling approaches on RAW images to generate metadata. We follow the uniform sampling as SAM [28], and show that it is enough to get much higher performance using uniform samples with the proposed method. Note that neither RIR nor SAM make their source codes publicly available, and only CAM reports their results reproduced on its own dataset. Therefore we only give a comparison with RIR and SAM on the dataset of CAM by inheriting its results in Table 1 and Figure 6." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 272, + 288, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 288, + 440 + ], + "type": "text", + "content": "Dataset. We deploy the same dataset of the previous work [24] to test our proposed method, which contains the preprocessed [13] and downsampled (with a downsampling factor of 4) images of three cameras from the NUS dataset [9]—Samsung NX2000, Olympus E-PL6 and Sony SLT-A57. Note that as our method does not require any pretraining step, the training and validation sets are not used in our method. Besides, since our method is able to directly process full-resolution images, we also compare it with CAM [24] on the original NUS dataset [9]. Additionally, as the NUS dataset [9] includes JPEG-compressed sRGB images rendered by real camera ISPs, we also conduct the experiments on these RAW-JPEG image pairs. These results are reported in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "type": "text", + "content": "Implementation details. For capture stage, we use the same sampling rate of " + }, + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "type": "text", + "content": " like [24] in all experiments for fair comparison with other metadata-based methods. With respect to reconstruction stage, we split the image into different patches—(228, 272) for Samsung NX2000, (216, 288) for Olympus E-PL6 and (204, 304) for Sony SLT-A57—to make the resolution divisible. The proposed INF is trained by Adam optimizer [15] with " + }, + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "type": "text", + "content": " loss in 500 iterations. We set the initial learning rate to 0.0001, and reduce it to half every 200 iterations. The regularisation parameters are set to " + }, + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{s}} = 0.0001, \\lambda_{\\mathrm{p}} = 0.1, \\lambda_{\\mathrm{r}} = 0.001" + }, + { + "bbox": [ + 46, + 443, + 287, + 612 + ], + "type": "text", + "content": " to produce the final results. It is worth noting that the reconstruction network is only trained for each specific patch, and will be re-initialized for the training of other patches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 620, + 105, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 105, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 105, + 632 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Quantitative results. We report the quantitative comparison results against other three baselines in Table 1. It can be observed that our method makes a remarkably improvement in both PSNR and SSIM metrics, especially outperforms the closest competitor CAM [24] by more than " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "10\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " in average PSNR. This indicates the effectiveness of INF to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 465, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 465, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 465, + 85 + ], + "type": "text", + "content": "modeling the sRGB-to-RAW mapping." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 87, + 547, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 87, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 87, + 547, + 207 + ], + "type": "text", + "content": "Qualitative results. We also provide examples of the qualitative comparison in Figure 6, which gives an intuitive explanation of the success of our method. As can be seen from the error maps, our method significantly reduces the reconstructed error on the whole images, especially the saturated regions [26]. We note that recovering saturated regions is the most difficult problem for all RAW reconstruction methods, and is the only case of our method getting large errors " + }, + { + "bbox": [ + 304, + 87, + 547, + 207 + ], + "type": "inline_equation", + "content": "(\\geqslant 1\\%)" + }, + { + "bbox": [ + 304, + 87, + 547, + 207 + ], + "type": "text", + "content": ". We attribute this to the expressive power [38] of INF, which is further discussed in Section 4.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 216, + 399, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 216, + 399, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 216, + 399, + 228 + ], + "type": "text", + "content": "4.3. Ablation study." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 237, + 546, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 546, + 357 + ], + "type": "text", + "content": "RBF vs. INF. We first compare RBF and INF with different input, which is shown in Table 2. As can be seen, the proposed INF outperforms the RBF interpolation [28] whatever the input is. We attribute the main performance boost to the basic structure, i.e. SIREN [31], as INF is able to modelling the sRGB-to-RAW mapping more accurately (59.62dB) only with the pixel value as input. Moreover, compared with RBF, our proposed INF can make further improvement by splitting the input into separate branches and regularize them with different weights." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 360, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 552 + ], + "type": "text", + "content": "The structure of INF. We also conduct experiments on the structure of INF, which is reported in Figure 7. We compare the different activation functions and the layer location where two input branches are merged. It is obviously shown that the INF with sine activation remarkably outperforms the INF with ReLU function, which can be attributed to the expressive power of INRs [38]. As to the location of fusion layer, we find there is no significant difference except the final layer for the INF with sine activation, since it directly maps the merged vector to the output by a fully connection layer without activation. Therefore, we merge the two branches at fourth layer to balance the complexity of each branch1. On the contrary, due to lack of such expressive power, the INF with ReLU requires more layers to model the mapping. Hence the performance increases when the input branch goes deep." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "type": "text", + "content": "Regularisation parameters. We further discuss the influence of key setups to INF, which mainly focuses on the regularisation parameters and patch size. As is shown in Table 3, it is obvious that the regularisation parameters play a vital role in the final accuracy. Here we use the grid search with a commonly-used tuner TPE [2] (implemented by [22]) in a grid of " + }, + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "type": "inline_equation", + "content": "[1, 10^{-1}, \\dots, 10^{-5}]" + }, + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "type": "text", + "content": " to determine each parameter. In our experiments, we find the regulation parameters are mainly related to the complexity of camera ISP rather than data dependent, which means we only need" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 683, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 547, + 713 + ], + "type": "text", + "content": "1In fact, we find the fusion at third layer would achieve a slightly higher performance, but which can be ignored compared with the instability of random initialization. See more detail in supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18200" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 523, + 184 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 523, + 184 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 523, + 184 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 523, + 184 + ], + "type": "table", + "html": "
DatasetMethodSamsung NX2000Olympus E-PL6Sony SLT-A57
PSNRSSIMPSNRSSIMPSNRSSIM
CAM [24]RIR [25]45.660.993948.420.992451.260.9982
SAM [28]47.030.996249.350.997850.440.9982
CAM [24]49.570.997551.540.998053.110.9985
Ours53.530.998556.710.999158.360.9992
NUS [9]CAM [24]49.510.993752.870.996153.340.9959
Ours61.480.999163.380.999363.800.9995
", + "image_path": "5a932bb0e86d9d819e33d1bb53fce77df64092e0e33a53a819c49e6d231deee7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 202, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 545, + 236 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with [25, 28, 24]. Here CAM dataset [24] involves downsampling operation (with the factor of 4) on the original NUS dataset [9], which undermines the strength of local mapping. For CAM dataset, we conduct the experiments on their test set to inherit the results of [24]. For NUS dataset, we report the results tested on the whole dataset. The best score for each column is in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 247, + 542, + 618 + ], + "blocks": [ + { + "bbox": [ + 49, + 247, + 542, + 618 + ], + "lines": [ + { + "bbox": [ + 49, + 247, + 542, + 618 + ], + "spans": [ + { + "bbox": [ + 49, + 247, + 542, + 618 + ], + "type": "image", + "image_path": "c0a13b958cb61dd10a4256e3a44811f0b4337e77e3adfe1c74e8b205873a7213.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 626, + 544, + 660 + ], + "lines": [ + { + "bbox": [ + 46, + 626, + 544, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 544, + 660 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison with [25, 28, 24]. Each two rows from top to bottom are respectively selected from Olympus E-PL6, Samsung NX2000 and Sony SLT-A57. The GT RAW images are processed with a gamma function for better visualization, and the reconstructed RAW images are visualized through error maps. This figure is best viewed in the electronic version." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 681, + 287, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 681, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 681, + 287, + 704 + ], + "type": "text", + "content": "to tune it for different ISPs, e.g., digital camera ISP, smart phone ISP, or software ISP. It also means we just need to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 681, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 681, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 306, + 681, + 545, + 704 + ], + "type": "text", + "content": "use one or quite a few images for tuning, resulting in a fast tuning process—within 2 hours on a single RTX3090. The" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18201" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 258, + 184 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 258, + 184 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 258, + 184 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 258, + 184 + ], + "type": "table", + "html": "
MethodInputPSNR
RBF [28]s29.10
RBF [28]p21.71
RBF [28](s, p)49.69
INFs59.62
INFp31.09
INF(s, p)60.91
INF{s; p}62.24
", + "image_path": "79f5a9ea9ee1d76fcabf9a588d5530750d173d2fafc779bbf8818d46a2055830.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 284, + 264, + 404 + ], + "blocks": [ + { + "bbox": [ + 62, + 284, + 264, + 404 + ], + "lines": [ + { + "bbox": [ + 62, + 284, + 264, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 264, + 404 + ], + "type": "image", + "image_path": "78a1b604ae63ab7d72b714c8b4adda9fd326c3e2baf78a155478f312dbe90725.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 414, + 287, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 448 + ], + "type": "text", + "content": "Figure 7. Ablation study on the structure of INF. The abscissa value represents the layer where we merge the two input branches, e.g., the branches are merged at fourth layer in Figure 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 470, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 286, + 506 + ], + "type": "text", + "content": "tuned parameters can then be used for all scenes and all other images from the ISPs with similar complexity, hence we use the same parameters for the three reported cameras." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": "As shown in Table 3, the proposed INF achieves the best performance at " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathbf{s}} = 0.0001" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathbf{p}} = 0.1" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathbf{r}} = 0.001" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": ", which indicates the spatial branch is not supposed to be more complex. This is because the output " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": " has weaker relationship with the coordinates " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": " than sRGB values " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": ", and regularisation could help to control the contribution of " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": ". Therefore when we set " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathbf{p}} = 0" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": ", the reconstruction performance dramatically drops down due to inaccurate information from " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": " is expanded. We believe such intuitive relationship would be helpful to set the initial values for regularisation parameters. Besides, for different ISPs with similar complexity, the results are insensitive to the parameters in a wide range (for example, where all branches are regularized by " + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "inline_equation", + "content": "\\lambda \\geqslant 10^{-4}" + }, + { + "bbox": [ + 46, + 506, + 286, + 686 + ], + "type": "text", + "content": "). In this sense, the risk of over-fitting due to these hyper-parameter settings is very small." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 714 + ], + "type": "text", + "content": "Impact of patch size. Moreover, we find that different size of the patch would lead to an obvious accuracy gap. For the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 538, + 213 + ], + "blocks": [ + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "text", + "content": "Table 2. Ablation study on different reconstruction methods. The best score is in bold and the second in italic. " + }, + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "inline_equation", + "content": "(\\mathbf{s},\\mathbf{p})" + }, + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "text", + "content": " refers to combine " + }, + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "text", + "content": " into a single input vector, while " + }, + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s};\\mathbf{p}\\}" + }, + { + "bbox": [ + 46, + 202, + 287, + 258 + ], + "type": "text", + "content": " denotes separating them as different input. For the RBF interpolation, we use the linear RBF kernel as proposed in [28]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 538, + 213 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 538, + 213 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 538, + 213 + ], + "type": "table", + "html": "
Patch size(λs,λp,λr)PSNR
×147.48
×2(0,0,0)49.17
×450.53
×161.13
×2(0.0001,0.001,0.0001)60.38
×457.80
×162.58
×2(0.0001,0.1,0.001)62.24
×459.64
", + "image_path": "9521b458d119f69a609c27eba7004dfeeec32f96a19b6b54e06460af5db51cce.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 231, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 320 + ], + "type": "text", + "content": "Table 3. Ablation study on the patch size and regularisation parameters. The best score is in bold and the second in italic. For the patch size, " + }, + { + "bbox": [ + 304, + 231, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 304, + 231, + 545, + 320 + ], + "type": "text", + "content": " refers to the sizes we report in our main experiments, and the rescale factor of 1, 2, 4 corresponds to the length of patch sides. We note that the accuracy gap less than " + }, + { + "bbox": [ + 304, + 231, + 545, + 320 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 231, + 545, + 320 + ], + "type": "text", + "content": " can not represent an improvement, which may be caused by the different calculations of PSNR and the random initialization of INF. This is further discussed in our supplementary material." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 340, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 472 + ], + "type": "text", + "content": "case of all regularisation parameters set to 0, we find small patch would lead to accuracy reduction. When the regularisation is correctly set, reconstruction on a small patch is more accurate than on a large one. This can be explained by the various local mapping functions in different image regions, which means modelling the local mapping for a small patch would be more accurate than a larger one. Note that though reducing the patch size would be helpful to the reconstruction accuracy, it would take much more time to recover an image. Therefore, we compromise to use the patch size around 200-300 in the main experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 479, + 378, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 378, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 378, + 490 + ], + "type": "text", + "content": "4.4. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 499, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 536 + ], + "type": "text", + "content": "Why INF works in sRGB-to-RAW mapping. We first rethink the RBF interpolation [3] used in SAM [28], by which the sRGB-to-RAW mapping is built in the following form:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 389, + 543, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 543, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 389, + 543, + 545, + 557 + ], + "type": "interline_equation", + "content": "\\hat {y} = \\mathbf {M} ^ {\\text {t e s t}} \\mathbf {M} ^ {- 1} y \\tag {6}", + "image_path": "db60b5ac51239863c63b7a42628b7d21b6c24e9c67500aabd889839b657db52f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 564, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 564, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 564, + 545, + 613 + ], + "type": "text", + "content": " denotes the RBF interpolation matrix with linear kernel function [27], and test refers to the pixels that need to be reconstructed. As discussed in [32], the output of an MLP can be approximated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 619, + 545, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 619, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 358, + 619, + 545, + 634 + ], + "type": "interline_equation", + "content": "\\hat {y} ^ {(t)} = \\mathbf {K} ^ {\\text {t e s t}} \\mathbf {K} ^ {- 1} \\left(\\mathbf {I} - e ^ {- \\eta \\mathbf {K} t}\\right) y \\tag {7}", + "image_path": "65b8c899377422532edc6ed00e455e9980c6ebff9ef2a0335a00cbdf3eb449a0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " denotes the NTK matrix [12] and " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " refers to training iterations. In our experiments, an MLP (typical with ReLU activation) tends to behave like RBF when " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is large enough (e.g., " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "t > 1000" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "), which meets the results of Equation (6) and Equation (7). The reason why RBF interpolation works can be attributed to that sRGB-RAW mapping" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18202" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 272, + 148 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 272, + 148 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 272, + 148 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 272, + 148 + ], + "type": "table", + "html": "
Method×4×8×16
Bicubic2.624.196.68
PixTransform [20]3.513.944.94
INF2.503.065.23
INF with ReLU2.392.693.57
", + "image_path": "5733ff64a408211ea0e05a940d066f65243b416966bef20b38336d4fabe723d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 167, + 287, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 287, + 200 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on guided depth superresolution. We use the Middlebury dataset provided by [33] for testing. The results are reported in terms of average RMSE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 223, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 286, + 354 + ], + "type": "text", + "content": "is piecewise smooth [27], but there still exists errors when using a linear kernel function, which is also suffered by MLPs with ReLU. On the other hand, Sitzmann et al. [31] demonstrate that MLPs with periodic activation functions (i.e., SIRENs) are more suitable for representing complicated signals and their derivatives compared with the typical MLP architectures. SIRENs have the advantage of representing the derivatives of the signal [31], which is more helpful to reconstruct fine details. Therefore, recovering RAW images by SIRENs can effectively improve the performance, which is also proved by Table 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 358, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 286, + 538 + ], + "type": "text", + "content": "Limitations and future research. One of the unresolved problem of our method is that the regularisation parameters of weight decay require pre-experiments to determine. These three parameters play a significant role to control the final result by limiting the complexity of network. In our experiments, we find the parameters are strongly related to the correlation of input and output data. For example, compared with the coordinates, sRGB values have a more obvious relationship with the RAW values, where the regularisation parameters are respectively set to " + }, + { + "bbox": [ + 46, + 358, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\lambda_{p} = 0.1" + }, + { + "bbox": [ + 46, + 358, + 286, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 358, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\lambda_{s} = 0.0001" + }, + { + "bbox": [ + 46, + 358, + 286, + 538 + ], + "type": "text", + "content": ". Such correlation is suitable for modelling with another neural network, and would lead to the performance improvement with optimized parameters. Therefore, learning the regularisation parameters from the input and output is supposed to be a valuable topic of future work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "type": "text", + "content": "4.5. Applications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Guided depth super-resolution As mentioned in Section 2, our framework is related to guided super-resolution, therefore we also provide an experiment to illustrate the applicability of our framework. Considering that our method is only self-supervised without any priori knowledge, we give a comparison with another self-supervised method PixTransform [20]. As shown in Table 4, INF with ReLU outperforms other methods, which indicates the ReLU activation is more suitable for this task than the sine activation. The reason for ReLU working better can be attributed to the sharp changes of the depth near the object edges, as ReLU is piecewise and discontinuous, which fits this depth disconti" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 387, + 148 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 387, + 148 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 387, + 148 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 387, + 148 + ], + "type": "image", + "image_path": "1c4a8c3a74025fc666c754094d2dbf1e8474ebcf49fc335ea6b590a6ea2314ca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 151, + 361, + 160 + ], + "lines": [ + { + "bbox": [ + 335, + 151, + 361, + 160 + ], + "spans": [ + { + "bbox": [ + 335, + 151, + 361, + 160 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 71, + 463, + 148 + ], + "blocks": [ + { + "bbox": [ + 389, + 71, + 463, + 148 + ], + "lines": [ + { + "bbox": [ + 389, + 71, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 389, + 71, + 463, + 148 + ], + "type": "image", + "image_path": "1219412462c45f3b2453bf8d2b1ff29cb2970be19a8ef7d7a3ef3031c0e52b06.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 151, + 438, + 160 + ], + "lines": [ + { + "bbox": [ + 414, + 151, + 438, + 160 + ], + "spans": [ + { + "bbox": [ + 414, + 151, + 438, + 160 + ], + "type": "text", + "content": "Guide" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 465, + 71, + 542, + 148 + ], + "blocks": [ + { + "bbox": [ + 465, + 71, + 542, + 148 + ], + "lines": [ + { + "bbox": [ + 465, + 71, + 542, + 148 + ], + "spans": [ + { + "bbox": [ + 465, + 71, + 542, + 148 + ], + "type": "image", + "image_path": "57788384b97aebde3c91e7b824c0d49d6b3910a9ea1b979c37252e274f3c47fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 162, + 387, + 239 + ], + "blocks": [ + { + "bbox": [ + 310, + 162, + 387, + 239 + ], + "lines": [ + { + "bbox": [ + 310, + 162, + 387, + 239 + ], + "spans": [ + { + "bbox": [ + 310, + 162, + 387, + 239 + ], + "type": "image", + "image_path": "0945424c2419cee1bcb58fd28cc0a1c6164cda3bd61709a9b1f2c1eef896ecb7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 242, + 382, + 252 + ], + "lines": [ + { + "bbox": [ + 314, + 242, + 382, + 252 + ], + "spans": [ + { + "bbox": [ + 314, + 242, + 382, + 252 + ], + "type": "text", + "content": "PixTransform [20]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 162, + 463, + 239 + ], + "blocks": [ + { + "bbox": [ + 389, + 162, + 463, + 239 + ], + "lines": [ + { + "bbox": [ + 389, + 162, + 463, + 239 + ], + "spans": [ + { + "bbox": [ + 389, + 162, + 463, + 239 + ], + "type": "image", + "image_path": "e61d52c2b98b2ff1e9d0b694935e7e866d83ef0a3e7cca024c36e421cd6e02ca.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 242, + 434, + 251 + ], + "lines": [ + { + "bbox": [ + 418, + 242, + 434, + 251 + ], + "spans": [ + { + "bbox": [ + 418, + 242, + 434, + 251 + ], + "type": "text", + "content": "INF" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 465, + 162, + 542, + 239 + ], + "blocks": [ + { + "bbox": [ + 492, + 151, + 515, + 161 + ], + "lines": [ + { + "bbox": [ + 492, + 151, + 515, + 161 + ], + "spans": [ + { + "bbox": [ + 492, + 151, + 515, + 161 + ], + "type": "text", + "content": "Target" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 162, + 542, + 239 + ], + "lines": [ + { + "bbox": [ + 465, + 162, + 542, + 239 + ], + "spans": [ + { + "bbox": [ + 465, + 162, + 542, + 239 + ], + "type": "image", + "image_path": "d549c9409a9663144e61a9a339cb1deb72c4f2709220ba774a0acbe089abd2b3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 242, + 533, + 251 + ], + "lines": [ + { + "bbox": [ + 474, + 242, + 533, + 251 + ], + "spans": [ + { + "bbox": [ + 474, + 242, + 533, + 251 + ], + "type": "text", + "content": "INF with ReLU" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 262, + 544, + 284 + ], + "lines": [ + { + "bbox": [ + 306, + 262, + 544, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 544, + 284 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison on guided depth super-resolution." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "type": "text", + "content": "nuity. In contrast, the discontinuity in RAW-to-sRGB mapping is moderate and the smooth sine activation function is a better option. We also provide qualitative comparison in Figure 8. which shows that the result of INF, though does not achieve higher accuracy, obtains better visual quality compared with the result of INF with ReLU." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "type": "text", + "content": "Low-light image enhancement. As a potential application of RAW reconstruction, we test our method on low-light image enhancement (LLIE) task. We find that the LLIE task can be simplified into a linear degradation problem (e.g., enlarging the pixel values by the same multiple) on the reconstructed RAW images, which achieves equal visual quality compared with the state-of-the-art deep learning-based approaches [11, 34, 21]. The results are reported in our supplementary. Note that such comparison is merely aimed to illustrate the effectiveness of executing low-light enhancement on reconstructed RAW images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 378, + 542 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 552, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 671 + ], + "type": "text", + "content": "We propose a method for recovering the RAW image from the sRGB counterpart with assistance of additional metadata, which is sampled from the RAW image at capture time. We introduce the implicit neural function (INF) to remarkably improve the reconstruction accuracy (10 dB average PSNR) only with uniform sampling. We prove that the structure of INF is beneficial to merging the information from both pixel values and coordinates. Further experiments indicate that our framework is also suitable for the task of guided super-resolution." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 678, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 678, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 678, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was partly supported by National Key R&D Program of China (Grant No. 2021YFD1400200) and National Natural Science Foundation of China (Grant No. 62088101)." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18203" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19323-19332, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 189 + ], + "type": "text", + "content": "[2] James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization. Advances in neural information processing systems, 24, 2011. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 213 + ], + "type": "text", + "content": "[3] Marco Evangelos Biancolini. Fast radial basis functions for engineering applications. Springer, 2017. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "type": "text", + "content": "[4] Ayan Chakrabarti, Daniel Scharstein, and Todd E Zickler. An empirical camera model for internet color vision. In BMVC, volume 1, page 4. CiteSeer, 2009. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 312 + ], + "type": "text", + "content": "[5] Ayan Chakrabarti, Ying Xiong, Baochen Sun, Trevor Darrell, Daniel Scharstein, Todd Zickler, and Kate Saenko. Modeling radiometric uncertainty for vision with tonemapred color images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11):2185-2198, 2014. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 356 + ], + "type": "text", + "content": "[6] Bowei Chen, Tiancheng Zhi, Martial Hebert, and Srinivasa G Narasimhan. Learning continuous implicit representation for near-periodic patterns. arXiv preprint arXiv:2208.12278, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "type": "text", + "content": "[7] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "type": "text", + "content": "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "type": "text", + "content": "[9] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "text", + "content": "[10] Marcos V Conde, Steven McDonagh, Matteo Maggioni, Ales Leonardis, and Eduardo Pérez-Pellitero. Model-based image signal processors via learnable dictionaries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 481-489, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 624 + ], + "type": "text", + "content": "[11] Chunle Guo Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1780-1789, June 2020. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "text", + "content": "[12] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Hakki Can Karaimer and Michael S Brown. A software platform for manipulating the camera imaging pipeline. In European Conference on Computer Vision, pages 429-444. Springer, 2016. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] Seon Joo Kim, Hai Ting Lin, Zheng Lu, Sabine Süssstrunk, Stephen Lin, and Michael S Brown. A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(12):2289-2302, 2012. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, 2015. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[16] Shayan Kousha, Ali Maleky, Michael S Brown, and Marcus A Brubaker. Modeling srgb camera noise with normalizing flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17463-17471, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[17] Zhihao Li, Si Yi, and Zhan Ma. Rendering nighttime image via cascaded color and brightness compensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 897–905, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[18] Hai Ting Lin, Zheng Lu, Seon Joo Kim, and Michael S Brown. Nonuniform lattice regression for modeling the camera imaging pipeline. In European Conference on Computer Vision, pages 556-568. Springer, 2012. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 305, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 370 + ], + "type": "text", + "content": "[19] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image hdr reconstruction by learning to reverse the camera pipeline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1651-1660, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 371, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 425 + ], + "type": "text", + "content": "[20] Riccardo de Lutio, Stefano D'aronco, Jan Dirk Wegner, and Konrad Schindler. Guided super-resolution as pixel-to-pixel transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8829-8837, 2019. 3, 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[21] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 481, + 504, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 504, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 504, + 491 + ], + "type": "text", + "content": "[22] Microsoft. Neural Network Intelligence, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[23] Seonghyeon Nam and Seon Joo Kim. Modelling the scene dependent imaging in cameras with a deep neural network. In Proceedings of the IEEE International Conference on Computer Vision, pages 1717-1725, 2017. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "text", + "content": "[24] Seonghyeon Nam, Abhijith Punnappurath, Marcus A Brubaker, and Michael S Brown. Learning srgb-to-raw-rgb de-rendering with content-aware metadata. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17704–17713, 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "text", + "content": "[25] Rang MH Nguyen and Michael S Brown. Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1655-1663, 2016. 1, 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[26] Rang MH Nguyen and Michael S Brown. Raw image reconstruction using a self-contained srgb-jpeg image with small memory overhead. International journal of computer vision, 126(6):637-650, 2018. 1, 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[27] Abhijith Punnappurath and Michael S Brown. Learning raw image reconstruction-aware deep image compressors. IEEE" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18204" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "text", + "content": "transactions on pattern analysis and machine intelligence, 42(4):1013-1019, 2019. 3, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[28] Abhijith Punnappurath and Michael S Brown. Spatially aware metadata for raw reconstruction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 218-226, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 195 + ], + "type": "text", + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[30] Guy Shacht, Dov Danon, Sharon Fogel, and Daniel Cohen-Or. Single pair cross-modality super resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6378–6387, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 294 + ], + "type": "text", + "content": "[31] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 361 + ], + "type": "text", + "content": "[32] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 363, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 287, + 407 + ], + "type": "text", + "content": "[33] Jiaxiang Tang, Xiaokang Chen, and Gang Zeng. Joint implicit image function for guided depth super-resolution. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4390-4399, 2021. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 408, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 287, + 452 + ], + "type": "text", + "content": "[34] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2604-2612, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 453, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 517 + ], + "type": "text", + "content": "[35] Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neuralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1526-1535, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 519, + 287, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 562 + ], + "type": "text", + "content": "[36] Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6287-6296, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 563, + 287, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 287, + 607 + ], + "type": "text", + "content": "[37] Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248-4257, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 609, + 287, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 287, + 662 + ], + "type": "text", + "content": "[38] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 664, + 287, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 664, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 664, + 287, + 708 + ], + "type": "text", + "content": "[39] Xuaner Zhang, Qifeng Chen, Ren Ng, and Vladlen Koltun. Zoom to learn, learn to zoom. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3762-3770, 2019. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18205" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_content_list.json b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e77085979f405d2a51adf5a8c91da7e5021f822d --- /dev/null +++ b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_content_list.json @@ -0,0 +1,1606 @@ +[ + { + "type": "text", + "text": "MethaneMapper: Spectral Absorption aware Hyperspectral Transformer for Methane Detection", + "text_level": 1, + "bbox": [ + 96, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Satish Kumar", + "bbox": [ + 171, + 204, + 282, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "satishkumar@ucsb.edu", + "bbox": [ + 137, + 224, + 316, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ivan Arevalo", + "bbox": [ + 362, + 204, + 467, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ifa@ucsb.edu", + "bbox": [ + 361, + 224, + 468, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ASM Iftekhar", + "bbox": [ + 532, + 204, + 647, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "iftekhar@ucsb.edu", + "bbox": [ + 514, + 224, + 666, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "B S Manjunath", + "bbox": [ + 710, + 204, + 831, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "manj@ucsb.edu", + "bbox": [ + 710, + 224, + 830, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Electrical and Computer Engineering \nUniversity of California Santa Barbara", + "bbox": [ + 277, + 246, + 692, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 316, + 313, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Methane $(\\mathsf{CH}_4)$ is the chief contributor to global climate change. Recent Airborne Visible-Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) has been very useful in quantitative mapping of methane emissions. Existing methods for analyzing this data are sensitive to local terrain conditions, often require manual inspection from domain experts, prone to significant error and hence are not scalable. To address these challenges, we propose a novel end-to-end spectral absorption wavelength aware transformer network, MethaneMapper, to detect and quantify the emissions. MethaneMapper introduces two novel modules that help to locate the most relevant methane plume regions in the spectral domain and uses them to localize these accurately. Thorough evaluation shows that MethaneMapper achieves 0.63 mAP in detection and reduces the model size (by $5\\times$ ) compared to the current state of the art. In addition, we also introduce a large-scale dataset of methane plume segmentation mask for over 1200 AVIRIS-NG flight lines from 2015-2022. It contains over 4000 methane plume sites. Our dataset will provide researchers the opportunity to develop and advance new methods for tackling this challenging green-house gas detection problem with significant broader social impact. Dataset and source code link1.", + "bbox": [ + 75, + 348, + 473, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 717, + 209, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider the problem of detecting and localizing methane $\\left(\\mathrm{CH}_4\\right)$ plumes from hyperspectral imaging data. Detecting and localizing potential $\\mathrm{CH}_4$ hot spots is a necessary first step in combating global warming due to greenhouse gas emissions. Methane gas is estimated to contribute $20\\%$ of global warming induced by greenhouse gasses [24] with a Global Warming Potential (GWP) 86 times higher than carbon dioxide $(\\mathrm{CO}_2)$ in a 20 year period [33]. To", + "bbox": [ + 75, + 743, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "put into perspective, the amount of environmental damage that $\\mathrm{CO}_{2}$ can do in 100 years, $\\mathrm{CH}_4$ can do in 1.2 years. Hence it is critical to monitor and curb the $\\mathrm{CH}_4$ emissions. While $\\mathrm{CH}_4$ emission has many sources, of particular interest are those from oil and natural gas industries. According to the United States Environmental Protection Agency report, $\\mathrm{CH}_4$ emissions from these industries accounts to 84 million tons per year [18]. These $\\mathrm{CH}_4$ emissions emanate from specific locations, mainly from pipeline leakages, storage tank leak or leakage from oil extraction point.", + "bbox": [ + 496, + 316, + 893, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current efforts to detect these sources mostly depend on aerial imagery. The Jet Propulsion Laboratory (JPL) has conducted thousands of aerial surveys in the last decade to collect data using an airborne sensor AVIRIS-NG [21]. Several methods have been proposed to detect potential emission sites from such imagery, for example, see [8,9,35,38, 40, 41]. However, these methods are in general very sensitive to background context and land-cover types, resulting in a large number of false positives that often require significant domain expert time to correct the detections. The primary reason is that these pixel-based methods are solely dependent on spectral correlations for detection. Spatial information can be very effective in reducing these false positives as $\\mathrm{CH}_4$ plumes exhibit a plume-like structure morphology. There has been recent efforts in utilizing spatial correlation using deep learning methods [22, 29], however, these works don't leverage spectral properties to filter out confusers. For example, methane has similar spectral properties as white-painted commercial roofs or paved surfaces such as airport asphalts [1]. This paper presents a novel deep-network based solution to minimize the effects of such confusers in accurately localizing methane plumes.", + "bbox": [ + 496, + 473, + 895, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our proposed approach, referred to as the MethaneMapper (MM), adapts the DETR [4], a transformer model that combines the spectral and spatial correlations in the imaging data to generate a map of potential methane $\\mathrm{(CH_4)}$ plume candidates. These candidates reduce the search space for a hyperspectral decoder to detect $\\mathrm{CH}_4$ plumes and re", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/UCSB-VRL/MethaneMapper-Spectral-Absorption-aware-Hyperspectral-Transformer-for-Methane-Detection", + "bbox": [ + 75, + 875, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17609", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "move potential confusers. MM is a light-weight end-to-end single-stage $\\mathrm{CH}_4$ detector and introduces two novel modules: a Spectral Feature Generator and a Query Refiner. The former generates spectral features from a linear filter that maximizes the $\\mathrm{CH}_4$ -to-noise ratio in the presence of additive background noise, while the latter integrates these features for decoding.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A major bottle neck for development of $\\mathrm{CH}_4$ detection methods is the limited availability of public training data. To address this, another significant contribution of this research is the introduction of a new Methane Hot Spots (MHS) dataset, largest of its kind available for computer vision researchers. MHS is curated by systematically collecting information from different publicly available datasets (airborne sensor [6], Non-profits [3,31] and satellites [34]) and generating the annotations as described in Section 4.1.1. This curated dataset contains methane segmentation masks for over 1200 AVIRIS-NG flight lines from years 2015 to 2022. Each flight line contains anywhere from $3 - 4\\mathrm{CH}_4$ plume sites for a total of 4000 in the MHS dataset.", + "bbox": [ + 75, + 196, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 393, + 421, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a novel single-stage end-to-end approach for methane plume detection using a hyperspectral transformer. The two modules, Spectral Feature Generator and Query Refiner, work together to improve upon the traditional transformer design and enable localization of potential methane hot spots in the hyperspectral images using a Spectral-Aware Linear Filter and refine the query representation for better decoding.", + "2. A new Spectral Linear Filter (SLF) improves upon traditional linear filters by strategically picking correlated pixels in spectral domain to better whiten background distribution and amplify methane signal.", + "3. A new benchmark dataset, MHS, provides the largest $(\\sim 35\\times)$ publicly available dataset of annotated AVIRIS-NG flight lines from years 2015-2022." + ], + "bbox": [ + 88, + 415, + 468, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 678, + 225, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work is at the intersection of hyperspectral data for $\\mathrm{CH}_4$ detection, deterministic linear filtering methods for spectral features and encoder-decoder based transformer. A review of the pertinent related works is given below.", + "bbox": [ + 75, + 704, + 468, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are several recent papers on detecting methane plumes from the airborne imaging spectrometer AVIRIS-NG [21]. This includes the Iterative Maximum a Posterior Differential Optical Absorption Spectroscopy algorithm (IMAP-DOAS) [7, 8] and matched filters [9, 35, 38, 40, 41]. IMAP-DOAS requires data from 2 hyperspectral sensors, one airborne and another on ground, hence not very practical for most application scenarios. Matched-filter based methods use background statistics to normalize", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c642682fdf22bacd4de13a996f1a9c7a413a775a8368849b65ce61b639bdef5.jpg", + "image_caption": [ + "Figure 1. Depiction of data collection process. Each flightline is $\\sim 300$ km long. An array of 598 sensors records data at $1.5\\mathrm{m}$ pixel spatial resolution. All flightlines are ortho-corrected. Each data-cube is of dimension $\\sim 25000\\times \\sim 1500\\times 432$" + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 883, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the spectral signals and match with the $\\mathrm{CH}_4$ spectral signature at every spatial location (pixel-wise). This process, however, is sensitive to surface albedo and land cover with spectral absorption similar to $\\mathrm{CH}_4$ , leading to spurious detections. Domain experts must then manually inspect each flight line to identify and delineate real $\\mathrm{CH}_4$ plumes [39]. To suppress the effect of false positives due to variability of elements on ground, Christopher et.al. [10] and Thorbe et. al. [41] introduced cluster-tuned matched filter. It involves clustering the pixels with similar spectral properties using k-means clustering. Both IMAP-DOAS and all versions of matched filters are heavily prone to false positives as the information is processed pixel-wise.", + "bbox": [ + 496, + 362, + 893, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Machine learning approaches have been used for target detection, including $\\mathrm{CH}_4$ identification, in hyperspectral imagery [12,17,28,36,37]. Similar to matched-filtering, these methods do not take into account the influence of confusers on the $\\mathrm{CH}_4$ spectral signature and have similar issues concerning false positives. Recently introduced deep learning based H-mrcnn model [29] focus on capturing spatial correlation. H-mrcnn is an ensemble of mask-rcnncite networks processing blocks of hyperspectral data. This block processing in the spectral domain is inefficient and often results in overall poor performance. Methanet [22] is a more recent work focusing on estimating methane concentration from matched-filter data. In this regard, our proposed MethaneMapper uses both spectral and spatial correlation to accurately delineates $\\mathrm{CH}_4$ plumes.", + "bbox": [ + 496, + 563, + 893, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets: The only dataset publicly available with annotation for $\\mathrm{CH}_4$ plume detection is JPL- $\\mathrm{CH}_4$ -detection2017-V1.0 dataset [39]. It contains only 46 AVIRIS-NG [21] flight lines in the US Four-Corners region. Deep learning architectures require a large number of annotated samples, and for this reason we introduce the new MHS dataset with over 1200 annotated flightlines and $\\sim 4000$ plume sites.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17610", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7da038c7b5b519847a23b51c3895625349fa028d309a2347e15dc08561707b6b.jpg", + "image_caption": [ + "Figure 2. Overview of MethaneMapper (MM) architecture. Given a hyperspectral image, our RGB $(400\\mathrm{nm} - 700\\mathrm{nm})$ and SWIR $(2000\\mathrm{nm} - 2500\\mathrm{nm})$ band-pass filters passes a subset of channels in desired wavelength range and feed them to CNN backbones (ResNet) to extract features. These features are concatenated and fed to Transformer Encoder. Parallelly, our Spectral Feature Generator (SFG) modules take in all channels of input image and generate methane candidates features. Next these candidates are sent to Query Refiner (QR) to refine queries. Then these queries decoded using encoded feature from Transformer Encoder. Finally each decoded query is used to predict a plume mask via Mask Prediction and, bounding box and class via FFNs (Feed Forward Network)." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 911, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. MethaneMapper (MM) Architecture", + "text_level": 1, + "bbox": [ + 75, + 363, + 410, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Overview", + "text_level": 1, + "bbox": [ + 75, + 388, + 230, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AVIRIS-NG hyperspectral imaging sensors capture spectral radiance values from $\\mathrm{N}_0$ ( $\\mathrm{N}_0 = 432$ ) channels corresponding to wavelengths ranging from $400nm - 2500nm$ as shown in Fig. 1. The complete hyperspectral image is represented as $\\mathbf{x} \\in \\mathbb{R}^{H_0 \\times W_0 \\times N_0}$ where $H_0, W_0$ are the height & width, respectively, and $N_0 = 432$ is number of channels. This hyperspectral data includes a very weak signature of $\\mathrm{CH}_4$ around $2100 - 2400nm$ , conflated with radiations from the surrounding land cover and background clutter. A single flight-line could be over a couple miles long (about $25\\mathrm{K}$ pixels in one of the dimensions), with an array of sensors recording the data at $1.5\\mathrm{m/pixel}$ resolution. The images are orthorectified before processing.", + "bbox": [ + 73, + 412, + 473, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Technical Overview", + "text_level": 1, + "bbox": [ + 75, + 619, + 264, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Referring to Fig.2, MM contains the following main components: (i) 2 CNN backbones to extract a compact feature representation of the spectral regions of interest from the hyperspectral image, (ii) a Spectral Feature Generator (SFG), and (iii) a Query Refiner (QR) in between an encoder-decoder pair (inspired by GTNet [20], SSRT [19]). The hyperspectral image is first processed through two separate band-pass filters to select the channels in visible $(400 - 700nm)$ and short-wave infrared (SWIR)(2000 - 2500nm) wavelength regions, and are then passed through CNN backbones. Output of these backbones are concatenated together and then encoded using a transformer encoder.", + "bbox": [ + 73, + 643, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The SFG (Sec. 3.4) takes in all channels of the hyperspectral image and process them through a spectral linear filter. The SFG exploits the spectral correlation to generate methane candidates feature maps and passes them to QR.", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The QR (Sec. 3.5) uses these methane candidates to refine the learnable queries. Our hyperspectral decoder takes the encoded features from the encoder and refined queries from QR to generate the embeddings. The mask-prediction layer processes these embeddings along with the feature pyramid from the backbone layers to generate the final methane-plume segmentation prediction.", + "bbox": [ + 496, + 364, + 890, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These individual blocks are discussed in more detail below.", + "bbox": [ + 498, + 473, + 890, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Bandpass filtering for the Encoder", + "text_level": 1, + "bbox": [ + 498, + 506, + 802, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The HSI is processed by two parallel band-pass filters; a visible wavelength $(400 - 700nm)$ (RGB) and a shortwave infrared wavelength $(2000 - 2500nm)$ (SWIR) bandpass filter. The RGB filter results in a 3 channel output corresponding to the normal red, green, and blue wavelengths. The SWIR generates channels, approximately $5\\mathrm{nm}$ apart. The filtered outputs are $\\mathbf{x}_{rgb} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}$ and $\\mathbf{x}_{swir} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 100}$ . Using $\\mathbf{x}_{rgb}$ and $\\mathbf{x}_{swir}$ , two conventional CNN backbones (e.g. ResNet-50 [16, 25]) generate two feature maps respectively of size $\\in \\mathbb{R}^{H \\times W \\times N}$ . Here $H = \\frac{H_0}{32}$ , $W = \\frac{W_0}{32}$ and $N = 2048$ typically. We concatenate these feature maps along channel dimension and project through a $1 \\times 1$ convolution layer to retain channel dimension of $N$ . The resulting output is $f_{comb} \\in \\mathbb{R}^{H \\times W \\times N}$ .", + "bbox": [ + 496, + 532, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the standard architecture of transformer encoder from previous works [4, 19, 20, 26, 43], we reduce the channel dimension of $f_{comb}$ using $1 \\times 1$ convolution to $f_z \\in \\mathbb{R}^{H \\times W \\times d}$ and supplement position information by adding a fixed positional embedding $p \\in \\mathbb{R}^{H \\times W \\times d}$ . The encoder consists of a stack of multi-head self-attention modules and feed-forward networks (FFN). The encoded feature map is $f_e \\in \\mathbb{R}^{H \\times W \\times d}$ :", + "bbox": [ + 496, + 746, + 893, + 868 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {e} = \\operatorname {E n c o d e r} \\left(f _ {z}, p\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17611", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Spectral Feature Generator (SFG)", + "text_level": 1, + "bbox": [ + 76, + 90, + 375, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In parallel, the input hyperspectral image is processed by the SFG module to generate methane candidates feature map $f_{mc}$ , providing the QR module with spatial information to help the network delineate the methane plumes.", + "bbox": [ + 76, + 114, + 468, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The SFG consist of a spectral linear filter (SLF) and a Feature Extractor (e.g.ResNet-50 [16]). The most common linear filtering approach for detecting $\\mathrm{CH}_4$ is to take each pixel from the input hyperspectral image $\\{\\mathbf{x}_{ij} \\mid \\mathbf{x}_{ij} \\in \\mathbb{R}^{1 \\times 1 \\times N_0}\\}_{i,j=1}^{H_0, W_0}$ and project it onto a $\\mathrm{CH}_4$ spectral absorption signature vector of same size [13]. This is to reduce the interference from ground terrain and amplify the $\\mathrm{CH}_4$ visibility in that pixel. Accurately modeling SLF is critical given that it is designed to reduce ground terrain interference. To model SLF we use the most common approach to matched filtering from information theory [42].", + "bbox": [ + 76, + 176, + 468, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spectral Linear Filter (SLF): The design of SLF is dependent on the spectral absorption pattern of $\\mathrm{CH}_4$ gas [13] and distribution of ground terrain. Since our signal of interest, $\\mathrm{CH}_4$ , is very weak, traditional methods of linear filtering [40, 41] are not effective. The conventional methods to whiten the ground terrain noise includes calculating the covariance $(\\mathbf{Cov} \\in \\mathbb{R}^{N_0 \\times N_0})$ of background by selecting a set of 10-15 adjacent columns $\\{\\mathbf{x}_i \\mid \\mathbf{x}_i \\in \\mathbb{R}^{1 \\times H_0 \\times N_0}\\}_{i=1}^{W_0}$ . However, in a given flight-line, the terrain changes frequently, from water bodies to bare soil, vegetation, buildings and other urban structures. Therefore single approximation of the covariance can not provide correct estimate of $\\mathrm{CH}_4$ and a localized context-based whitening will be more effective. To address this problem, we took a very simple and effective approach of doing land cover classification and segmentation [11, 32, 44], and then compute covariance per class from the land cover. More details in supplementary materials. This improves the quality of methane candidates in presence of confusers (materials with similar spectral absorption patterns as $\\mathrm{CH}_4$ ) and also in cases where $\\mathrm{CH}_4$ concentration is low. The final SLF design with per class covariance is:", + "bbox": [ + 76, + 364, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S L F} \\left(\\mathbf {x} _ {i j}\\right) = \\frac {\\left(\\mathbf {x} _ {i j} - \\mu_ {k}\\right) ^ {T} \\mathbf {C o v} _ {k} ^ {- 1} t}{\\sqrt {t ^ {T} \\mathbf {C o v} _ {k} ^ {- 1}} t} \\forall (i, j) \\in \\text {c l a s s} k \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 708, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ represents the spectral absorption pattern [13] of $\\mathrm{CH}_4$ gas, and $\\mathbf{Cov}_k$ , $\\mu_k$ are the covariance and mean of $k^{th}$ class respectively. $\\mathbf{x}_{ij}$ represents the pixel in input hyperspectral image at $(i,j)$ index in $k^{th}$ class. This operation generates a 2-D spatial $\\mathrm{CH}_4$ candidates map of size $\\mathbb{R}^{H_0\\times W_0}$ . Next this $\\mathrm{CH}_4$ candidates map is fed to a Feature Extractor to generate $\\mathrm{CH}_4$ candidates feature map $f_{mc}$ . Details of the land cover segmentation/classification and complete SLF derivation are in the Supplementary materials.", + "bbox": [ + 76, + 763, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {m c} = \\operatorname {F e a t u r e E x t r a c t o r} (\\mathbf {S L F} (\\mathbf {x} _ {i j}) \\forall i, j) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 104, + 890, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Query Refiner (QR)", + "text_level": 1, + "bbox": [ + 500, + 128, + 689, + 145 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next the methane candidate feature map $f_{mc} \\in \\mathbb{R}^{H \\times W \\times d}$ is fed to the QR module along with a set of 100 learnable queries $Q \\in \\mathbb{R}^{100 \\times d}$ . The $f_{mc}$ refines the learnable queries via cross-attention mechanism. This operation provides a narrow search space for the queries. The QR module follows a transformer decoder-like architecture inspired from [19,20]. The randomly initialized queries $Q \\in \\mathbb{R}^{100 \\times d}$ are first passed through a self-attention layer to attend to themselves. Next, these queries attend to our methane candidates feature map $f_{mc}$ from SFG module through a cross-attention layer. The methane candidates feature map serves as key-values pairs in our attention architecture. The output of QR is $Q_{ref}$ .", + "bbox": [ + 498, + 152, + 890, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {r e f} = \\mathbf {Q R} \\left(f _ {m c}, Q\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 359, + 890, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.6. Hyperspectral Decoder", + "text_level": 1, + "bbox": [ + 500, + 383, + 705, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $Q_{ref}$ is fed to the decoder module along with encoder output $f_{e}$ to generate output embeddings. Our hyperspectral decoder follows the standard architecture with a minor difference. There are no self-attention layers, just stack of multi-headed cross attention layers. The refined queries are transformed into output embeddings $E_{out} \\in \\mathbb{R}^{100 \\times d}$ .", + "bbox": [ + 498, + 407, + 890, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {o u t}} = \\operatorname {D e c o d e r} \\left(f _ {e}, p, Q _ {\\text {r e f}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 508, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.7. Box and Mask Prediction", + "text_level": 1, + "bbox": [ + 500, + 532, + 732, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The decoder output embeddings $(E_{out})$ are fed to two Feed Forward Network (FFNs) and a Mask prediction layer. The outputs of the FFNs are the bounding boxes covering each $\\mathrm{CH}_4$ plume and a confidence score corresponding to each box. The mask-prediction module follows the standard segmentation head of DETR [4]. It computes multi-head attention scores of each embedding over the $f_{e}$ (Eq. 1), generating a low-resolution heatmap for each embedding. To make the final prediction a Feature Pyramid Network [14] like structure is used. Each heatmap is designed to capture one methane plume. A simple thresholding is used to merge the heatmaps as final segmentation mask.", + "bbox": [ + 498, + 555, + 890, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm a s k = \\operatorname {M a s k} _ {-} \\operatorname {p r e d} \\left(E _ {\\text {o u t}}, f _ {e}, f _ {\\text {c o m b}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 747, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.8. Training and Inference", + "text_level": 1, + "bbox": [ + 500, + 771, + 712, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train MethaneMapper in two stages; first we train bounding box detection corresponding to each $\\mathrm{CH}_4$ plume, and second by freezing the box detection network and training only the mask prediction module. We also trained both box and mask prediction modules end-to-end and achieved similar performance. We use a similar two-stage loss strategy for training MethaneMapper as that used in DETR [4]:", + "bbox": [ + 498, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17612", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4656e40c8c344ccdb4e448e5e781ba0cd9fac1831c406ca2c2e680ff5c98466a.jpg", + "image_caption": [ + "Figure 3. Sample images from MHS dataset. The colormap in black circle shows concentration maps corresponding to the plume mask shown in red. We are showing different types of leakage sources and land cover types. For better visualization, we plotted the binary mask on color image created using visible bands of hyperspectral image." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 444, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "first stage is the bipartite matching between the predictions and the ground truths both in bounding box and mask prediction, and then second stage is loss calculation for the matched pairs. The bipartite matching employs the Hungarian algorithm [4] to find the optimal matching between the predictions and the ground truths. After this matching, every prediction is associated with a ground truth. Next, we calculate the $l_{1}$ and $GIoU$ loss on both box and mask predictions and cross entropy loss for class prediction [4].", + "bbox": [ + 75, + 556, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference: The inference pipeline is similar to training pipeline and can be implemented using approximately 50 lines of code. During inference, we first filter the detections with confidence levels below $50\\%$ and a per-pixel max to determine which pixels are predicted to belong to a CH4 plume.", + "bbox": [ + 75, + 707, + 468, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Methane Hot Spots (MHS) dataset", + "text_level": 1, + "bbox": [ + 76, + 799, + 390, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another significant contribution of this work is a large scale curated MHS dataset. It contains the AVIRIS-NG spectral data with wavelength ranging from $380nm$ to $2510nm$ , a $5nm$ sampling [21], and capturing 432 channels per pixel. The images from the flight-line are orthorec", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cb9970bd25ff7d3f5a997f06b12daef31ba86ca955f9ebbe8c96a731e90769d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMHS (Ours) DatasetJPL-CH4 detection-V1.0 [39]
# plume sites3961161
# flightlines118546
# point source3675114
# diffused source28657
Time period2015 - 2022(8 years)2015(1 year)
Segmentation MaskYesYes
Bonding boxYesNo
Concentration mapYesNo
Number of Regions61
", + "bbox": [ + 501, + 88, + 911, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Statistics shows MHS dataset comparison with JPL-CH4-detection-V1.0 [39] dataset. Each flightline have multiple large and small plume sites. Each flightline have at least 4 plume sites. The Point Source represents high concentration $(300kg/hr)$ to leakage from sources like pipeline leak, storage tanks, oil and gas refineries. Diffused Source represent low concentration leakages from sources like biomass degradation in landfills. Our dataset is covers more diverse type of terrain over 6 states.", + "bbox": [ + 496, + 282, + 893, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tified and of size $\\sim 23K\\times \\sim 1.5K\\times 432$ . The only currently publicly-available dataset with methane plume segmentation masks is the JPL-CH4-detection-V1.0 [39] dataset released by JPL-NASA in 2017. The MHS dataset has approximately 4000 plume sites corresponding to approximately 1200 AVIRIS-NG flightlines as shown in Table 1. MHS also has higher diversity data with flight lines spanning from 2015-2022 and covering terrain from 6 states– California, Nevada, New Mexico, Colorado, Midland Texas, and Virginia.", + "bbox": [ + 496, + 422, + 890, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Pruning: We selected AVIRIS-NG flight lines over varying regions as it covers a wide variety of $\\mathrm{CH}_4$ plume sources, such as leaks in oil and gas refineries, oil and gas extraction points, natural seeps, leaking underground storage tank, coal mines, dairy farms, landfill sites, and pipeline leaks. Along with varying emission sources, we selected regions with different types of ground terrains like, bare soil, rocks, mountains, light vegetation, water bodies and dense vegetation as shown with few samples in Fig. 3. Different types of ground terrain exhibit widely varying albedo and thus have a major impact on the quality of $\\mathrm{CH}_4$ detections as shown in Fig. 4. Given this, training models with diverse ground terrain data leads to a more robust model.", + "bbox": [ + 496, + 585, + 890, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Concentration map and Segmentation mask", + "text_level": 1, + "bbox": [ + 498, + 792, + 872, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concentration map is provided in the form of a matrix of spatial dimensions same as the flightline $(\\sim 23k\\times \\sim 1.5k\\times 1)$ . There is one concentration map per flight-line (orthorectified). It shows methane concentration in parts-per-million (ppm) per-pixel on the ground. Pixel-regions with no methane presence are set to zero.", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17613", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segmentation mask provided in the format of a \"png\" image file with three channels and of the same spatial dimension as the corresponding flight line ( $\\sim 23k \\times \\sim 1.5k \\times 3$ ). The segmentation mask is obtained from the concentration mask file by setting all pixel values above zero to represent methane plumes. We manually annotated Point Source and Diffused Source based on the type of ground terrain and concentration of methane gas. Following the benchmark dataset [39], three channels are used to color code Point Source (Red) and Diffused Source (Green). The distinction of Point Source and Diffused Source is derived from the JPL-CH4-detection-V1.0 benchmark dataset [39]. Our annotation style is also consistent with the JPL-CH4-detection-V1.0 benchmark dataset [39], so that both datasets can be merged seamlessly.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Constructing Concentration map", + "text_level": 1, + "bbox": [ + 76, + 342, + 362, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Concentration maps are generated by mapping expert-annotated methane-plume concentration maps to the orthocorrected AVIRIS-NG flightlines. These methane plume annotations are systematically collected from a nonprofit [3] entity. They provide concentration masks of methane emissions in $150 \\times 150$ size patches along with location information from different sources (airborne sensors [6], satellites [34]). In order to map these patches from different sources to the AVIRIS-NG flight-lines, we use the pixel coordinate locations provided for both the annotations and flight-lines. We use this information to create a homography transformation to map each pixel to its corresponding location in the flight-line. Fig. 3 shows a sample of varying types of terrains with $\\mathrm{CH}_4$ segmentation mask in red and concentration mask in black circle. Details about matching the resolution, ortho-correction, and transformation are discussed in supplementary materials. The patch annotations are verified by experts visiting the physical location of emission the same day [2]. Most of the regions in California are verified by physical visits by California Air Resource Board [2,3].", + "bbox": [ + 75, + 367, + 472, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. MHS Statistics", + "text_level": 1, + "bbox": [ + 76, + 696, + 228, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MHS statistics and properties are summarized in Table 1.", + "bbox": [ + 96, + 715, + 467, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Annotations: MHS provides both segmentation masks and concentration maps which enable development of deep learning algorithms than can produce both $\\mathrm{CH}_4$ plume location and concentration predictions.", + "bbox": [ + 76, + 732, + 467, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diversity: MHS dataset includes AVARIS-NG flightlines spanning 8 years (2015 - 2022) from six states in the U.S.: California, Nevada, New Mexico, Colorado, Texas, and Virginia.", + "bbox": [ + 76, + 792, + 467, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Split: We divide MHS dataset into train/test splits of $80 - 20\\%$ with overlapping time periods and locations. Our dataset covers 6 states. Each state has sub-regions/locations", + "bbox": [ + 76, + 854, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(e.g. Permian basin) that are covered by multiple nonoverlapping flightlines ( $25k \\times 1.5k \\times 432$ pixels). These flightlines are split into train and test sets. In each set, we create patches ( $256 \\times 256 \\times 432$ pixels) from the corresponding flightlines. From the patches/tiles, we take all positives patches (methane ( $\\mathrm{CH}_4$ ) and randomly sample equal number of negative ( $\\mathrm{no - CH_4}$ ) patches. This is done for both train and test sets separately to balance the data and we refer to Section 6.2 for detailed ablation studies.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental settings", + "text_level": 1, + "bbox": [ + 498, + 251, + 707, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: Following the evaluation protocol of H-mrcnn [29] we report our performance in mean intersection-over-union (mIOU). Here, mIOU indicates the overlap between the predicted and the ground truth $\\mathrm{CH}_4$ plume masks. ED represents the accuracy in plume core prediction. Additionally, as first stage of our two stage training procedure contains bounding box prediction, we also report our performance in predicting plume bounding boxes in terms of mean Average Precision (mAP) which tells us the effectiveness of MethaneMapper in eliminating the false positives in plume prediction.", + "bbox": [ + 496, + 273, + 890, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Pre-Processing: Each input hyperspectral image is approximately of size $25000 \\times 1500 \\times 432$ taking up memory space of $55 - 60\\mathrm{GB}$ . We create tiles of each image in spatial domain, each tile is of size $256 \\times 256 \\times 432$ [29] with an overlap of 128. The $\\mathrm{CH}_4$ plume is available in very few pixels in the whole image, $90\\%$ of the tiles are negative samples (no methane, just ground terrain). We can not use the whole hyperspectral image because of GPU memory limitations", + "bbox": [ + 496, + 444, + 890, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: The band-selectors module takes 432-channels hyperspectral image as input, the RGB band selector picks 60 channel from $400nm - 700nm$ wavelength range and creates a 3-channel RGB image, the SWIR band selector picks 100 channel from wavelength range $2000nm - 2500nm$ . These input images are passed to two ResNet-50 [16] feature extractor backbones. The backbone networks are initialized with DETR [4] trained on COCO dataset [30] and input layer initialized randomly [15]. The transformer encoder-decoder and our query refiner have 6 layers and 8 heads. We initialized the transformer encoder-decoder with weights extracted and stripped from DETR [4] model. The dimension of transformer architecture is 256 and number of queries is 100. The SFG module takes in all 432-channels hyperspectral image and generates 1-channel output map of same spatial dimension as input. The feature extractor in SFG is ResNet-50 [16] initialized with DETR [4] trained on COCO dataset [30]. The decoder output embeddings are of size 512. The feature pyramid network in mask prediction module has 3 layers. More details are mentioned in supplementary materials.", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17614", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/52149fbf023a2cb80b0a30fe5950268d9cf30a453cca08208897783f9fd917bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBack boneSFG F.Ext.#paramsmAPmIOU
JPL-CH4-detection-v1.0 Dataset
1Hu et. alR-50-75M0.260.48
2H-mrcnnR-50-353M0.530.86
3MMR-50R-5080M0.630.91
MHS (Ours) Dataset
4SpectralFormerR-50-84M0.330.41
5UPSNet (stuff)R-50-69M0.320.38
6UPSNet (stuff + things)R-50-69M0.290.35
7DETRR-18*33M0.370.56
8DETRR-50*59M0.440.59
10R-18Linear Layer39M0.450.60
11MMR-18R-1844M0.520.63
12R-50R-5080M0.590.68
", + "bbox": [ + 81, + 88, + 467, + 305 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f2fda5f90154d35c0c45fb96478c1bea80411dd2222522969ef0a8c75b17042e.jpg", + "table_caption": [ + "Table 2. Comparison with baselines. “-” represent Not Applicable and “*” represent no SFG module and a random query used for transformer decoder. The top section shows performance on JPL-CH $_4$ dataset [39]. MethaneMapper achieves better results than heavily tuned H-mrcnn with $\\sim 5\\times$ fewer parameters. The overall detection accuracy is higher on this dataset because the type of ground terrain is uniform across all flightlines. In MHS dataset, MM outperforms multiple baselines as shown in rows 4-12. MM accuracy is lower in MHS than JPL-CH $_4$ dataset because MHS dataset has more variety of ground terrain spreading over 6 states" + ], + "table_footnote": [], + "table_body": "
MethodsmAPmIOU
LogReg [5]-0.05
SVM [36]-0.29
PCA + LogReg-0.06
PCA + SVM-0.31
MM (R-50)0.630.91
", + "bbox": [ + 158, + 468, + 387, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison with classical machine learning methods. \" - \" represent Not Available. The classical ML methods are not suited for the ${\\mathrm{{CH}}}_{4}$ detection task. MethaneMapper outperforms all methods on JPL dataset [39]", + "bbox": [ + 75, + 570, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Results", + "text_level": 1, + "bbox": [ + 76, + 627, + 163, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we will discuss and validate all the design choices for MethaneMapper (MM) with ablations. We show that MM achieves state-of-the-art results in overall performance compared all other methods shown in Tables 2 & 3.", + "bbox": [ + 75, + 646, + 468, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Performance comparison", + "text_level": 1, + "bbox": [ + 76, + 715, + 307, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deep Learning methods: We trained MM with ResNet-50 [16] backbone on the same dataset that H-mrcnn [29] (JPL-CH4-detection-V1.0 [39]) was trained on for fair comparison. To align with H-mrcnn we used the same split and input image size. The MM model with 80M parameters trained for 250 epochs outperforms by significant margin the H-mrcnn model with 352M parameters. Results are summarized in Table 2 that includes the performance of MM on the new larger MHS dataset. We note that though the code for H-mrcnn is available, many of the modules are deprecated and can not be reproduced. The 'Backbone' col", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "unn represents backbones used for feature extraction from input image,'SFG F.Ext.' represents the feature extractor in SFG module in MethaneMapper. We observed (qualitatively) that H-mrcnn fails to detect small $\\mathrm{CH}_4$ plumes with concentration lower than $100\\mathrm{kg / hr}$ while MM detects those.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We did evaluation by implementing 3 baseline models [4, 17, 45] shown rows 4-8 of Table 2. These methods were not designed for $\\mathrm{CH}_4$ detection task, therefore we needed to modify their input channel size. The poor performance of these methods may be attributed to the weak signal of interest in a high dimensional data, high number of confusers, and limited annotated data. Additionally, the only hyperspectral baseline method SpectralFormer [17] has low efficiency due its pixel-wise training scheme.", + "bbox": [ + 496, + 167, + 890, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classical ML methods: We trained and tested multiple existing machine learning based approaches that are used for methane detection, performance shown in Table 3. Logistic regression (LogReg) [5] and multinomial logistic regression (MLR) [23] failed to produce any meaningful detection with $90\\%$ false positive detections. We also trained a Support Vector Machine (SVM) [36,37] based classifier, it performed slightly better than LR and MLR methods with an IOU of $21\\%$ . SVMs are prone to false positives detections same as Gaussian Mixture Models [36]. We observed that all traditional methods are not suited for the task of $\\mathrm{CH}_4$ detection. We also tested reducing the dimension using principal component analysis (PCA) or just taking bands which shows maximum $\\mathrm{CH}_4$ absorption. In the later case, the traditional methods performed better than using all 432 bands, this backs our idea of just using bands from SWIR region.", + "bbox": [ + 496, + 303, + 892, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 5 shows comparison of MM's mask and bounding box prediction with ground truth mask on different ground terrains. The Leakages are from different type of sources such as, oil refinery, pipeline and storage tank. MM makes correct predictions in varying scenarios.", + "bbox": [ + 496, + 544, + 890, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 628, + 663, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We did the experiments for ablation on MHS dataset with ResNet-50 as backbone and validate the design choices. One parameter is changed for each ablation and others kept at best settings. More ablations in Supplementary.", + "bbox": [ + 496, + 646, + 890, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spectral Feature Generator Module: In Table 2 lower section, we show the effectiveness of our SFG module for the query refiner block. Our baseline is standard implementation of DETR [4] for segmentation task represented by row-1 and row-2 of Tab. 2 lower section. Using $\\mathrm{CH}_4$ candidates feature from SFG improves the bounding box detection performance by $0.14\\mathrm{mAP}$ and mask prediction by $0.09\\mathrm{mIOU}$ . This demonstrate that guiding queries with $\\mathrm{CH}_4$ candidates feature generated by SFG produces better embeddings as compared to random queries.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Along with this, we explored the provision of $\\mathrm{CH}_4$ candidates feature at 2 places, (i) at input level concatenating", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17615", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e3a6acfd3326c9d821617f2496dffeb73040da5c91197ffc6bac4229893a22b.jpg", + "image_caption": [ + "Figure 4. Comparison of SLF with traditional filter in SFG module. White pixels represent methane and black no-methane. Red boundary represents ground-truth plume mask. SLF module generates better $\\mathrm{CH}_4$ candidates" + ], + "image_footnote": [], + "bbox": [ + 80, + 98, + 467, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "it with $f_{comb}$ ; and (ii) as input to query refiner. We see an improvement of 0.09 mAP and 0.08 mIOU when SFG module output is passed to query refiner. We hypothesize that this is because on concatenating with input features, the CH $_4$ candidates feature information gets lost, while as cross-attention with queries reduces the search space for decoder and generate better embeddings.", + "bbox": [ + 75, + 367, + 467, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also experimented with different types of feature extractors for SFG module, and observed that a Resnet18 or Resnet50 [16] is more effective than a 2 linear layer feature extractor as shown in Table 2.", + "bbox": [ + 75, + 474, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spectral Linear Filter: We experimented with SLF for computing covariance $(Cov)$ using different subsets of columns in the input hyperspectral image. We observed that the SLF is most effective when covariance is computed class-wise based on land cover. Class-wise $Cov$ ensures that the radiance absorption by ground terrain is same for all the pixels while computing $\\mathrm{CH}_4$ enhancement. As can be seen in Fig. 4, SLF amplifies $\\mathrm{CH}_4$ candidate detection and reduces false positives. SLF leads to a $0.03\\mathrm{mAP}$ improved in detection compared to traditional filters. The prediction from MM is shown row-1 of Fig. 5.", + "bbox": [ + 75, + 535, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Geographic generalization: To assess the geographical generalization capabilities of MM, we trained it on MHS data from all states except California and tested it on flightlines from California. We observed a slight drop of 0.04 mAP in detections. However, when trained on all data except Virginia, we noticed a significant drop of $0.09\\mathrm{mAP}$ in detections. We attribute this to the fact that the land cover in Virginia is dense and moist vegetation, has a lower solar reflectance compared to the arid regions of California, Texas, and Nevada.", + "bbox": [ + 75, + 702, + 468, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal generalization: Testing MM on 2015 after training on data from 2016-2022 showed no performance drop.", + "bbox": [ + 76, + 854, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unbalanced test set: MM's performance dropped by 0.05", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1a144124c273a7dfdba175c468825c0bd032af4ff156e04861e464c5dd07c01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 88, + 700, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b28eeb09c44832b36a8595c354a7ecc2adc7c44c12b1b03c02745bdfe0b4660.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 202, + 700, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea65ccdbb0634334d83662aa8a54192394c8358967e1abafa58806ad2476bcc0.jpg", + "image_caption": [ + "Figure 5. Sample ground truths and predictions on MHS dataset. We show robustness of MethaneMapper predictions on different kind of ground terrain, rows 1 and 3 shows leakage at a refinery, row 2 shows leakage from pipeline in agricultural land, row 4 shows leakage from storage tank with concrete background." + ], + "image_footnote": [], + "bbox": [ + 557, + 314, + 702, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5eb15fe50361ce53547376d135c4c5a870c1277cfa21d169e86dbdbe62478522.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 88, + 849, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/016440e284b5f226b4af73e29fba63d590340f4cf634ea27a3a6f49c2749ac93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 202, + 849, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/963cd45d6c3f7a64a247ca55e350dcfa086d013c4c8d3b5309aca1d649ee5a53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 314, + 849, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mAP on an unbalanced test set with only $10\\%$ positive samples $\\mathrm{(CH_4)}$ and $90\\%$ negative samples (no- $\\mathrm{CH}_4$ ). This highlights the challenges in $\\mathrm{CH}_4$ detection. Future work will address this issue.", + "bbox": [ + 498, + 517, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 592, + 617, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents MethaneMapper - a hyperspectral Transformer for methane plume detection. It utilize spectral and spatial correlations using a spectral feature generator and a query refiner, to accurately delineate the $\\mathrm{CH}_4$ plumes. Additionally, we curated a large-scale dataset for the task, a first of its kind, which will be made available to all researchers. The proposed MethaneMapper significantly improves upon the current methods in terms of detection and localization accuracy, as our extensive experiments demonstrate. Future work will extend the model to global monitoring [27] using multispectral satellite imaging data.", + "bbox": [ + 496, + 614, + 890, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 801, + 678, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is partially supported by the following grants: NSF award SI2-SSI #1664172 and the US Army Research Laboratory (ARL) under agreement number W911NF2020157. We also thank Aditya Ramakrishnan in helping with the initial data cleaning process.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17616", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alana K. Ayasse, Andrew K. Thorpe, Dar A. Roberts, Christopher C. Funk, Philip E. Dennison, Christian Frankenberg, Andrea Steffke, and Andrew D. Aubrey. Evaluating the effects of surface properties on methane retrievals using a synthetic airborne visible/infrared imaging spectrometer next generation (aviris-ng) image. Remote Sensing of Environment, 215:386-397, 2018. 1", + "[2] CALIFORNIA AIR RESOURCE BOARD. Green house gas inventory by california air resource board, 2022. 6", + "[3] INC CARBON MAPPER. Carbon mapper methane emission exploration, 2022. 2, 6", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3, 4, 5, 6, 7", + "[5] Qi Cheng, Pramod K Varshney, and Manoj K Arora. Logistic regression for feature selection and soft classification of remote sensing data. IEEE Geoscience and Remote Sensing Letters, 3(4):491-494, 2006. 7", + "[6] E. Knapp David, Heckler Joseph, Seely Megs, and P. Asner Gregory. Global airborne observatory visible to infrared imaging spectrometer report, 2020. 2, 6", + "[7] C Frankenberg, U Platt, and T Wagner. Iterative maximum a posteriori (imap)-doas for retrieval of strongly absorbing trace gases: Model studies for ch 4 and co 2 retrieval from near infrared spectra of sciamachy onboard envisat. Atmospheric Chemistry and Physics Discussions, 4(5):6067-6106, 2004. 2", + "[8] Christian Frankenberg, U Platt, and T Wagner. Iterative maximum a posteriori (imap)-doas for retrieval of strongly absorbing trace gases: Model studies for ch 4 and co 2 retrieval from near infrared spectra of sciamachy onboard envisat. Atmospheric Chemistry and Physics, 5(1):9-22, 2005. 1, 2", + "[9] Christian Frankenberg, Andrew K Thorpe, David R Thompson, Glynn Hulley, Eric Adam Kort, Nick Vance, Jakob Borchardt, Thomas Krings, Konstantin Gerilowski, Colm Sweeney, et al. Airborne methane remote measurements reveal heavy-tail flux distribution in four corners region. Proceedings of the national academy of sciences, 113(35):9734-9739, 2016. 1, 2", + "[10] Christopher C Funk, James Theiler, Dar A Roberts, and Christoph C Borel. Clustering to improve matched filter detection of weak gas plumes in hyperspectral thermal imagery. IEEE transactions on geoscience and remote sensing, 39(7):1410-1420, 2001. 2", + "[11] Bo-Cai Gao. Ndwi—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote sensing of environment, 58(3):257-266, 1996. 4", + "[12] Utsav B Gewali, Sildomar T Monteiro, and Eli Saber. Machine learning based hyperspectral image analysis: a survey. arXiv preprint arXiv:1802.08701, 2018. 2", + "[13] IE Gordon, LS Rothman, RJ Hargreaves, R Hashemi, EV Karlovets, FM Skinner, EK Conway, C Hill, RV Kochanov, Y Tan, et al. The hitran2020 molecular spectroscopic" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "database. Journal of quantitative spectroscopy and radiative transfer, 277:107949, 2022. 4", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4, 6, 7, 8", + "[17] Danfeng Hong, Zhu Han, Jing Yao, Lianru Gao, Bing Zhang, Antonio Plaza, and Jocelyn Chanussot. Spectralformer: Rethinking hyperspectral image classification with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-15, 2021. 2, 7", + "[18] Paris IEA. Methane from oil & gas, 2020. 1", + "[19] ASM Iftekhar, Hao Chen, Kaustav Kundu, Xinyu Li, Joseph Tighe, and Davide Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5353-5363, 2022. 3, 4", + "[20] ASM Iftekhar, Satish Kumar, R Austin McEver, Suya You, and BS Manjunath. Gtnet: Guided transformer network for detecting human-object interactions. arXiv preprint arXiv:2108.00596, 2021. 3, 4", + "[21] California Institute of Technology Jet Propulsion Laboratory. Airborne visible infrared imaging spectrometer - next generation (aviris-ng) overview, 2009. 1, 2, 5", + "[22] Siraput Jongaramrungruang, Andrew K Thorpe, Georgios Matheou, and Christian Frankenberg. Methanet—an ai-driven approach to quantifying methane point-source emission from high-resolution 2-d plume imagery. Remote Sensing of Environment, 269:112809, 2022. 1, 2", + "[23] Mahdi Khodadadzadeh, Jun Li, Antonio Plaza, and José M Bioucas-Dias. A subspace-based multinomial logistic regression for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 11(12):2105-2109, 2014. 7", + "[24] Stefanie Kirschke, Philippe Bousquet, Philippe Ciais, Marielle Saunois, Josep G Canadell, Edward J Dlugokencky, Peter Bergamaschi, Daniel Bergmann, Donald R Blake, Lori Bruhwiler, et al. Three decades of global methane sources and sinks. Nature geoscience, 6(10):813-823, 2013. 1", + "[25] Satish Kumar, ASM Iftekhar, Michael Goebel, Tom Bullock, Mary H MacLean, Michael B Miller, Tyler Santander, Barry Giesbrecht, Scott T Grafton, and BS Manjunath. Stressnet: detecting stress in thermal videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 999-1009, 2021. 3", + "[26] Satish Kumar, ASM Iftekhar, Ekta Prashnani, and BS Manjunath. Locl: Learning object-attribute composition using localization. arXiv preprint arXiv:2210.03780, 2022. 3" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17617", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Satish Kumar, William Kingwill, Orbio Earth, Rozanne Mouton, Wojciech Adamczyk, Robert Huppertz, and Evan Sherwin. Guided transformer network for detecting methane emissions in sentinel-2 satellite imagery. 8", + "[28] Satish Kumar, Rui Kou, Henry Hill, Jake Lempges, Eric Qian, and Vikram Jayaram. In-situ water quality monitoring in oil and gas operations. arXiv preprint arXiv:2301.08800, 2023. 2", + "[29] Satish Kumar, Carlos Torres, Oytun Ulutan, Alana Ayasse, Dar Roberts, and BS Manjunath. Deep remote sensing methods for methane detection in overhead hyperspectral imagery. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1776-1785, 2020. 1, 2, 6, 7", + "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6", + "[31] META-STANFORD. Methane emissions technology alliance (meta), stanford natural gas initiative, 2022. 2", + "[32] MGISGeography. Ndvi (normalized difference vegetation index), 1979. 4", + "[33] Gunnar Myhre, Drew Shindell, and Julia Pongratz. Anthropogenic and natural radiative forcing. 2014. 1", + "[34] Darius Phiri, Matamyo Simwanda, Serajis Salekin, Vincent R Nyirenda, Yuji Murayama, and Manjula Ranagalage. Sentinel-2 data for land cover/use mapping: a review. Remote Sensing, 12(14):2291, 2020. 2, 6", + "[35] Dar A Roberts, Eliza S Bradley, Ross Cheung, Ira Leifer, Philip E Dennison, and Jack S Margolis. Mapping methane emissions from a marine geological seep source using imaging spectrometry. Remote Sensing of Environment, 114(3):592-606, 2010. 1, 2", + "[36] CA Shah, PK Varshney, and MK Arora. Ica mixture model algorithm for unsupervised classification of remote sensing imagery. International Journal of Remote Sensing, 28(8):1711-1731, 2007. 2, 7", + "[37] David MJ Tax and Robert PW Duin. Support vector data description. Machine learning, 54(1):45-66, 2004. 2, 7", + "[38] DR Thompson, I Leifer, H Bovensmann, M Eastwood, M Fladeland, C Frankenberg, K Gerilowski, RO Green, S Kratwurst, T Krings, et al. Real-time remote detection and measurement for airborne imaging spectroscopy: a case study with methane. Atmospheric Measurement Techniques, 8(10):4383-4397, 2015. 1, 2", + "[39] David R Thompson, Anuj Karpatne, Imme Ebert-Uphoff, Christian Frankenberg, Andrew K Thorpe, Brian D Bue, and Robert O Green. Isgeo dataset jpl-ch4-detection-2017-v1. 0: A benchmark for methane source detection from imaging spectrometer data. 2017. 2, 5, 6, 7", + "[40] Andrew K Thorpe, Christian Frankenberg, David R Thompson, Riley M Duren, Andrew D Aubrey, Brian D Bue, Robert O Green, Konstantin Gerilowski, Thomas Krings, Jakob Borchardt, et al. Airborne doas retrievals of methane, carbon dioxide, and water vapor concentrations at high spatial resolution: application to aviris-ng. Atmospheric Measurement Techniques, 10(10):3833-3850, 2017. 1, 2, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Andrew K Thorpe, Dar A Roberts, Eliza S Bradley, Christopher C Funk, Philip E Dennison, and Ira Leifer. High resolution mapping of methane emissions from marine and terrestrial sources using a cluster-tuned matched filter technique and imaging spectrometry. Remote Sensing of Environment, 134:305-318, 2013. 1, 2, 4", + "[42] George Turin. An introduction to matched filters. IRE transactions on Information theory, 6(3):311-329, 1960. 4", + "[43] Oytun Ulutan, ASM Iftekhar, and Bangalore S Manjunath. Vsgnet: Spatial attention network for detecting human object interactions using graph convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13617-13626, 2020. 3", + "[44] Foreign Agriculture Service US Dept. of Agriculture. Normalized difference vegetation index (ndvi), 1969. 4", + "[45] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8826, 2019. 7" + ], + "bbox": [ + 501, + 92, + 892, + 373 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17618", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_model.json b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1c5c71a566e8239336685bda9aa4df930cca4b93 --- /dev/null +++ b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "MethaneMapper: Spectral Absorption aware Hyperspectral Transformer for Methane Detection" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.205, + 0.284, + 0.221 + ], + "angle": 0, + "content": "Satish Kumar" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.226, + 0.318, + 0.239 + ], + "angle": 0, + "content": "satishkumar@ucsb.edu" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.205, + 0.468, + 0.221 + ], + "angle": 0, + "content": "Ivan Arevalo" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.226, + 0.47, + 0.238 + ], + "angle": 0, + "content": "ifa@ucsb.edu" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.205, + 0.648, + 0.221 + ], + "angle": 0, + "content": "ASM Iftekhar" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.226, + 0.667, + 0.238 + ], + "angle": 0, + "content": "iftekhar@ucsb.edu" + }, + { + "type": "text", + "bbox": [ + 0.711, + 0.205, + 0.833, + 0.222 + ], + "angle": 0, + "content": "B S Manjunath" + }, + { + "type": "text", + "bbox": [ + 0.712, + 0.226, + 0.831, + 0.239 + ], + "angle": 0, + "content": "manj@ucsb.edu" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.247, + 0.694, + 0.283 + ], + "angle": 0, + "content": "Department of Electrical and Computer Engineering \nUniversity of California Santa Barbara" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.317, + 0.314, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.474, + 0.698 + ], + "angle": 0, + "content": "Methane \\((\\mathsf{CH}_4)\\) is the chief contributor to global climate change. Recent Airborne Visible-Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) has been very useful in quantitative mapping of methane emissions. Existing methods for analyzing this data are sensitive to local terrain conditions, often require manual inspection from domain experts, prone to significant error and hence are not scalable. To address these challenges, we propose a novel end-to-end spectral absorption wavelength aware transformer network, MethaneMapper, to detect and quantify the emissions. MethaneMapper introduces two novel modules that help to locate the most relevant methane plume regions in the spectral domain and uses them to localize these accurately. Thorough evaluation shows that MethaneMapper achieves 0.63 mAP in detection and reduces the model size (by \\(5\\times\\)) compared to the current state of the art. In addition, we also introduce a large-scale dataset of methane plume segmentation mask for over 1200 AVIRIS-NG flight lines from 2015-2022. It contains over 4000 methane plume sites. Our dataset will provide researchers the opportunity to develop and advance new methods for tackling this challenging green-house gas detection problem with significant broader social impact. Dataset and source code link1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.718, + 0.21, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.866 + ], + "angle": 0, + "content": "We consider the problem of detecting and localizing methane \\(\\left(\\mathrm{CH}_4\\right)\\) plumes from hyperspectral imaging data. Detecting and localizing potential \\(\\mathrm{CH}_4\\) hot spots is a necessary first step in combating global warming due to greenhouse gas emissions. Methane gas is estimated to contribute \\(20\\%\\) of global warming induced by greenhouse gasses [24] with a Global Warming Potential (GWP) 86 times higher than carbon dioxide \\((\\mathrm{CO}_2)\\) in a 20 year period [33]. To" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.903 + ], + "angle": 0, + "content": "1https://github.com/UCSB-VRL/MethaneMapper-Spectral-Absorption-aware-Hyperspectral-Transformer-for-Methane-Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.895, + 0.472 + ], + "angle": 0, + "content": "put into perspective, the amount of environmental damage that \\(\\mathrm{CO}_{2}\\) can do in 100 years, \\(\\mathrm{CH}_4\\) can do in 1.2 years. Hence it is critical to monitor and curb the \\(\\mathrm{CH}_4\\) emissions. While \\(\\mathrm{CH}_4\\) emission has many sources, of particular interest are those from oil and natural gas industries. According to the United States Environmental Protection Agency report, \\(\\mathrm{CH}_4\\) emissions from these industries accounts to 84 million tons per year [18]. These \\(\\mathrm{CH}_4\\) emissions emanate from specific locations, mainly from pipeline leakages, storage tank leak or leakage from oil extraction point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.896, + 0.807 + ], + "angle": 0, + "content": "Current efforts to detect these sources mostly depend on aerial imagery. The Jet Propulsion Laboratory (JPL) has conducted thousands of aerial surveys in the last decade to collect data using an airborne sensor AVIRIS-NG [21]. Several methods have been proposed to detect potential emission sites from such imagery, for example, see [8,9,35,38, 40, 41]. However, these methods are in general very sensitive to background context and land-cover types, resulting in a large number of false positives that often require significant domain expert time to correct the detections. The primary reason is that these pixel-based methods are solely dependent on spectral correlations for detection. Spatial information can be very effective in reducing these false positives as \\(\\mathrm{CH}_4\\) plumes exhibit a plume-like structure morphology. There has been recent efforts in utilizing spatial correlation using deep learning methods [22, 29], however, these works don't leverage spectral properties to filter out confusers. For example, methane has similar spectral properties as white-painted commercial roofs or paved surfaces such as airport asphalts [1]. This paper presents a novel deep-network based solution to minimize the effects of such confusers in accurately localizing methane plumes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our proposed approach, referred to as the MethaneMapper (MM), adapts the DETR [4], a transformer model that combines the spectral and spatial correlations in the imaging data to generate a map of potential methane \\(\\mathrm{(CH_4)}\\) plume candidates. These candidates reduce the search space for a hyperspectral decoder to detect \\(\\mathrm{CH}_4\\) plumes and re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17609" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "move potential confusers. MM is a light-weight end-to-end single-stage \\(\\mathrm{CH}_4\\) detector and introduces two novel modules: a Spectral Feature Generator and a Query Refiner. The former generates spectral features from a linear filter that maximizes the \\(\\mathrm{CH}_4\\)-to-noise ratio in the presence of additive background noise, while the latter integrates these features for decoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.473, + 0.394 + ], + "angle": 0, + "content": "A major bottle neck for development of \\(\\mathrm{CH}_4\\) detection methods is the limited availability of public training data. To address this, another significant contribution of this research is the introduction of a new Methane Hot Spots (MHS) dataset, largest of its kind available for computer vision researchers. MHS is curated by systematically collecting information from different publicly available datasets (airborne sensor [6], Non-profits [3,31] and satellites [34]) and generating the annotations as described in Section 4.1.1. This curated dataset contains methane segmentation masks for over 1200 AVIRIS-NG flight lines from years 2015 to 2022. Each flight line contains anywhere from \\(3 - 4\\mathrm{CH}_4\\) plume sites for a total of 4000 in the MHS dataset." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.394, + 0.422, + 0.408 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.416, + 0.47, + 0.538 + ], + "angle": 0, + "content": "1. We introduce a novel single-stage end-to-end approach for methane plume detection using a hyperspectral transformer. The two modules, Spectral Feature Generator and Query Refiner, work together to improve upon the traditional transformer design and enable localization of potential methane hot spots in the hyperspectral images using a Spectral-Aware Linear Filter and refine the query representation for better decoding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.545, + 0.47, + 0.607 + ], + "angle": 0, + "content": "2. A new Spectral Linear Filter (SLF) improves upon traditional linear filters by strategically picking correlated pixels in spectral domain to better whiten background distribution and amplify methane signal." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.615, + 0.47, + 0.66 + ], + "angle": 0, + "content": "3. A new benchmark dataset, MHS, provides the largest \\((\\sim 35\\times)\\) publicly available dataset of annotated AVIRIS-NG flight lines from years 2015-2022." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.416, + 0.47, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.227, + 0.696 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Our work is at the intersection of hyperspectral data for \\(\\mathrm{CH}_4\\) detection, deterministic linear filtering methods for spectral features and encoder-decoder based transformer. A review of the pertinent related works is given below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "There are several recent papers on detecting methane plumes from the airborne imaging spectrometer AVIRIS-NG [21]. This includes the Iterative Maximum a Posterior Differential Optical Absorption Spectroscopy algorithm (IMAP-DOAS) [7, 8] and matched filters [9, 35, 38, 40, 41]. IMAP-DOAS requires data from 2 hyperspectral sensors, one airborne and another on ground, hence not very practical for most application scenarios. Matched-filter based methods use background statistics to normalize" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.09, + 0.885, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.271, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 1. Depiction of data collection process. Each flightline is \\(\\sim 300\\) km long. An array of 598 sensors records data at \\(1.5\\mathrm{m}\\) pixel spatial resolution. All flightlines are ortho-corrected. Each data-cube is of dimension \\(\\sim 25000\\times \\sim 1500\\times 432\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.56 + ], + "angle": 0, + "content": "the spectral signals and match with the \\(\\mathrm{CH}_4\\) spectral signature at every spatial location (pixel-wise). This process, however, is sensitive to surface albedo and land cover with spectral absorption similar to \\(\\mathrm{CH}_4\\), leading to spurious detections. Domain experts must then manually inspect each flight line to identify and delineate real \\(\\mathrm{CH}_4\\) plumes [39]. To suppress the effect of false positives due to variability of elements on ground, Christopher et.al. [10] and Thorbe et. al. [41] introduced cluster-tuned matched filter. It involves clustering the pixels with similar spectral properties using k-means clustering. Both IMAP-DOAS and all versions of matched filters are heavily prone to false positives as the information is processed pixel-wise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.895, + 0.791 + ], + "angle": 0, + "content": "Machine learning approaches have been used for target detection, including \\(\\mathrm{CH}_4\\) identification, in hyperspectral imagery [12,17,28,36,37]. Similar to matched-filtering, these methods do not take into account the influence of confusers on the \\(\\mathrm{CH}_4\\) spectral signature and have similar issues concerning false positives. Recently introduced deep learning based H-mrcnn model [29] focus on capturing spatial correlation. H-mrcnn is an ensemble of mask-rcnncite networks processing blocks of hyperspectral data. This block processing in the spectral domain is inefficient and often results in overall poor performance. Methanet [22] is a more recent work focusing on estimating methane concentration from matched-filter data. In this regard, our proposed MethaneMapper uses both spectral and spatial correlation to accurately delineates \\(\\mathrm{CH}_4\\) plumes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Datasets: The only dataset publicly available with annotation for \\(\\mathrm{CH}_4\\) plume detection is JPL- \\(\\mathrm{CH}_4\\)-detection2017-V1.0 dataset [39]. It contains only 46 AVIRIS-NG [21] flight lines in the US Four-Corners region. Deep learning architectures require a large number of annotated samples, and for this reason we introduce the new MHS dataset with over 1200 annotated flightlines and \\(\\sim 4000\\) plume sites." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17610" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.912, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.262, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Overview of MethaneMapper (MM) architecture. Given a hyperspectral image, our RGB \\((400\\mathrm{nm} - 700\\mathrm{nm})\\) and SWIR \\((2000\\mathrm{nm} - 2500\\mathrm{nm})\\) band-pass filters passes a subset of channels in desired wavelength range and feed them to CNN backbones (ResNet) to extract features. These features are concatenated and fed to Transformer Encoder. Parallelly, our Spectral Feature Generator (SFG) modules take in all channels of input image and generate methane candidates features. Next these candidates are sent to Query Refiner (QR) to refine queries. Then these queries decoded using encoded feature from Transformer Encoder. Finally each decoded query is used to predict a plume mask via Mask Prediction and, bounding box and class via FFNs (Feed Forward Network)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.364, + 0.411, + 0.383 + ], + "angle": 0, + "content": "3. MethaneMapper (MM) Architecture" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.389, + 0.231, + 0.405 + ], + "angle": 0, + "content": "3.1. Data Overview" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.413, + 0.474, + 0.611 + ], + "angle": 0, + "content": "AVIRIS-NG hyperspectral imaging sensors capture spectral radiance values from \\(\\mathrm{N}_0\\) (\\(\\mathrm{N}_0 = 432\\)) channels corresponding to wavelengths ranging from \\(400nm - 2500nm\\) as shown in Fig. 1. The complete hyperspectral image is represented as \\(\\mathbf{x} \\in \\mathbb{R}^{H_0 \\times W_0 \\times N_0}\\) where \\(H_0, W_0\\) are the height & width, respectively, and \\(N_0 = 432\\) is number of channels. This hyperspectral data includes a very weak signature of \\(\\mathrm{CH}_4\\) around \\(2100 - 2400nm\\), conflated with radiations from the surrounding land cover and background clutter. A single flight-line could be over a couple miles long (about \\(25\\mathrm{K}\\) pixels in one of the dimensions), with an array of sensors recording the data at \\(1.5\\mathrm{m/pixel}\\) resolution. The images are orthorectified before processing." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.62, + 0.266, + 0.636 + ], + "angle": 0, + "content": "3.2. Technical Overview" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.644, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Referring to Fig.2, MM contains the following main components: (i) 2 CNN backbones to extract a compact feature representation of the spectral regions of interest from the hyperspectral image, (ii) a Spectral Feature Generator (SFG), and (iii) a Query Refiner (QR) in between an encoder-decoder pair (inspired by GTNet [20], SSRT [19]). The hyperspectral image is first processed through two separate band-pass filters to select the channels in visible \\((400 - 700nm)\\) and short-wave infrared (SWIR)(2000 - 2500nm) wavelength regions, and are then passed through CNN backbones. Output of these backbones are concatenated together and then encoded using a transformer encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The SFG (Sec. 3.4) takes in all channels of the hyperspectral image and process them through a spectral linear filter. The SFG exploits the spectral correlation to generate methane candidates feature maps and passes them to QR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.472 + ], + "angle": 0, + "content": "The QR (Sec. 3.5) uses these methane candidates to refine the learnable queries. Our hyperspectral decoder takes the encoded features from the encoder and refined queries from QR to generate the embeddings. The mask-prediction layer processes these embeddings along with the feature pyramid from the backbone layers to generate the final methane-plume segmentation prediction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.474, + 0.891, + 0.489 + ], + "angle": 0, + "content": "These individual blocks are discussed in more detail below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.507, + 0.803, + 0.524 + ], + "angle": 0, + "content": "3.3. Bandpass filtering for the Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.746 + ], + "angle": 0, + "content": "The HSI is processed by two parallel band-pass filters; a visible wavelength \\((400 - 700nm)\\) (RGB) and a shortwave infrared wavelength \\((2000 - 2500nm)\\) (SWIR) bandpass filter. The RGB filter results in a 3 channel output corresponding to the normal red, green, and blue wavelengths. The SWIR generates channels, approximately \\(5\\mathrm{nm}\\) apart. The filtered outputs are \\(\\mathbf{x}_{rgb} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}\\) and \\(\\mathbf{x}_{swir} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 100}\\). Using \\(\\mathbf{x}_{rgb}\\) and \\(\\mathbf{x}_{swir}\\), two conventional CNN backbones (e.g. ResNet-50 [16, 25]) generate two feature maps respectively of size \\(\\in \\mathbb{R}^{H \\times W \\times N}\\). Here \\(H = \\frac{H_0}{32}\\), \\(W = \\frac{W_0}{32}\\) and \\(N = 2048\\) typically. We concatenate these feature maps along channel dimension and project through a \\(1 \\times 1\\) convolution layer to retain channel dimension of \\(N\\). The resulting output is \\(f_{comb} \\in \\mathbb{R}^{H \\times W \\times N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.895, + 0.869 + ], + "angle": 0, + "content": "Following the standard architecture of transformer encoder from previous works [4, 19, 20, 26, 43], we reduce the channel dimension of \\( f_{comb} \\) using \\( 1 \\times 1 \\) convolution to \\( f_z \\in \\mathbb{R}^{H \\times W \\times d} \\) and supplement position information by adding a fixed positional embedding \\( p \\in \\mathbb{R}^{H \\times W \\times d} \\). The encoder consists of a stack of multi-head self-attention modules and feed-forward networks (FFN). The encoded feature map is \\( f_e \\in \\mathbb{R}^{H \\times W \\times d} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\nf _ {e} = \\operatorname {E n c o d e r} \\left(f _ {z}, p\\right) \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17611" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.377, + 0.108 + ], + "angle": 0, + "content": "3.4. Spectral Feature Generator (SFG)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.469, + 0.176 + ], + "angle": 0, + "content": "In parallel, the input hyperspectral image is processed by the SFG module to generate methane candidates feature map \\( f_{mc} \\), providing the QR module with spatial information to help the network delineate the methane plumes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.177, + 0.469, + 0.344 + ], + "angle": 0, + "content": "The SFG consist of a spectral linear filter (SLF) and a Feature Extractor (e.g.ResNet-50 [16]). The most common linear filtering approach for detecting \\(\\mathrm{CH}_4\\) is to take each pixel from the input hyperspectral image \\(\\{\\mathbf{x}_{ij} \\mid \\mathbf{x}_{ij} \\in \\mathbb{R}^{1 \\times 1 \\times N_0}\\}_{i,j=1}^{H_0, W_0}\\) and project it onto a \\(\\mathrm{CH}_4\\) spectral absorption signature vector of same size [13]. This is to reduce the interference from ground terrain and amplify the \\(\\mathrm{CH}_4\\) visibility in that pixel. Accurately modeling SLF is critical given that it is designed to reduce ground terrain interference. To model SLF we use the most common approach to matched filtering from information theory [42]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Spectral Linear Filter (SLF): The design of SLF is dependent on the spectral absorption pattern of \\(\\mathrm{CH}_4\\) gas [13] and distribution of ground terrain. Since our signal of interest, \\(\\mathrm{CH}_4\\), is very weak, traditional methods of linear filtering [40, 41] are not effective. The conventional methods to whiten the ground terrain noise includes calculating the covariance \\((\\mathbf{Cov} \\in \\mathbb{R}^{N_0 \\times N_0})\\) of background by selecting a set of 10-15 adjacent columns \\(\\{\\mathbf{x}_i \\mid \\mathbf{x}_i \\in \\mathbb{R}^{1 \\times H_0 \\times N_0}\\}_{i=1}^{W_0}\\). However, in a given flight-line, the terrain changes frequently, from water bodies to bare soil, vegetation, buildings and other urban structures. Therefore single approximation of the covariance can not provide correct estimate of \\(\\mathrm{CH}_4\\) and a localized context-based whitening will be more effective. To address this problem, we took a very simple and effective approach of doing land cover classification and segmentation [11, 32, 44], and then compute covariance per class from the land cover. More details in supplementary materials. This improves the quality of methane candidates in presence of confusers (materials with similar spectral absorption patterns as \\(\\mathrm{CH}_4\\)) and also in cases where \\(\\mathrm{CH}_4\\) concentration is low. The final SLF design with per class covariance is:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.709, + 0.469, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S L F} \\left(\\mathbf {x} _ {i j}\\right) = \\frac {\\left(\\mathbf {x} _ {i j} - \\mu_ {k}\\right) ^ {T} \\mathbf {C o v} _ {k} ^ {- 1} t}{\\sqrt {t ^ {T} \\mathbf {C o v} _ {k} ^ {- 1}} t} \\forall (i, j) \\in \\text {c l a s s} k \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(t\\) represents the spectral absorption pattern [13] of \\(\\mathrm{CH}_4\\) gas, and \\(\\mathbf{Cov}_k\\), \\(\\mu_k\\) are the covariance and mean of \\(k^{th}\\) class respectively. \\(\\mathbf{x}_{ij}\\) represents the pixel in input hyperspectral image at \\((i,j)\\) index in \\(k^{th}\\) class. This operation generates a 2-D spatial \\(\\mathrm{CH}_4\\) candidates map of size \\(\\mathbb{R}^{H_0\\times W_0}\\). Next this \\(\\mathrm{CH}_4\\) candidates map is fed to a Feature Extractor to generate \\(\\mathrm{CH}_4\\) candidates feature map \\(f_{mc}\\). Details of the land cover segmentation/classification and complete SLF derivation are in the Supplementary materials." + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.106, + 0.892, + 0.123 + ], + "angle": 0, + "content": "\\[\nf _ {m c} = \\operatorname {F e a t u r e E x t r a c t o r} (\\mathbf {S L F} (\\mathbf {x} _ {i j}) \\forall i, j) \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.13, + 0.691, + 0.146 + ], + "angle": 0, + "content": "3.5. Query Refiner (QR)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.154, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Next the methane candidate feature map \\( f_{mc} \\in \\mathbb{R}^{H \\times W \\times d} \\) is fed to the QR module along with a set of 100 learnable queries \\( Q \\in \\mathbb{R}^{100 \\times d} \\). The \\( f_{mc} \\) refines the learnable queries via cross-attention mechanism. This operation provides a narrow search space for the queries. The QR module follows a transformer decoder-like architecture inspired from [19,20]. The randomly initialized queries \\( Q \\in \\mathbb{R}^{100 \\times d} \\) are first passed through a self-attention layer to attend to themselves. Next, these queries attend to our methane candidates feature map \\( f_{mc} \\) from SFG module through a cross-attention layer. The methane candidates feature map serves as key-values pairs in our attention architecture. The output of QR is \\( Q_{ref} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.36, + 0.892, + 0.377 + ], + "angle": 0, + "content": "\\[\nQ _ {r e f} = \\mathbf {Q R} \\left(f _ {m c}, Q\\right) \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.385, + 0.707, + 0.401 + ], + "angle": 0, + "content": "3.6. Hyperspectral Decoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.408, + 0.892, + 0.499 + ], + "angle": 0, + "content": "The \\( Q_{ref} \\) is fed to the decoder module along with encoder output \\( f_{e} \\) to generate output embeddings. Our hyperspectral decoder follows the standard architecture with a minor difference. There are no self-attention layers, just stack of multi-headed cross attention layers. The refined queries are transformed into output embeddings \\( E_{out} \\in \\mathbb{R}^{100 \\times d} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.509, + 0.892, + 0.526 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {o u t}} = \\operatorname {D e c o d e r} \\left(f _ {e}, p, Q _ {\\text {r e f}}\\right) \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.533, + 0.733, + 0.548 + ], + "angle": 0, + "content": "3.7. Box and Mask Prediction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.556, + 0.892, + 0.738 + ], + "angle": 0, + "content": "The decoder output embeddings \\((E_{out})\\) are fed to two Feed Forward Network (FFNs) and a Mask prediction layer. The outputs of the FFNs are the bounding boxes covering each \\(\\mathrm{CH}_4\\) plume and a confidence score corresponding to each box. The mask-prediction module follows the standard segmentation head of DETR [4]. It computes multi-head attention scores of each embedding over the \\(f_{e}\\) (Eq. 1), generating a low-resolution heatmap for each embedding. To make the final prediction a Feature Pyramid Network [14] like structure is used. Each heatmap is designed to capture one methane plume. A simple thresholding is used to merge the heatmaps as final segmentation mask." + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.748, + 0.891, + 0.765 + ], + "angle": 0, + "content": "\\[\nm a s k = \\operatorname {M a s k} _ {-} \\operatorname {p r e d} \\left(E _ {\\text {o u t}}, f _ {e}, f _ {\\text {c o m b}}\\right) \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.714, + 0.788 + ], + "angle": 0, + "content": "3.8. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We train MethaneMapper in two stages; first we train bounding box detection corresponding to each \\(\\mathrm{CH}_4\\) plume, and second by freezing the box detection network and training only the mask prediction module. We also trained both box and mask prediction modules end-to-end and achieved similar performance. We use a similar two-stage loss strategy for training MethaneMapper as that used in DETR [4]:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17612" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.446, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.458, + 0.472, + 0.542 + ], + "angle": 0, + "content": "Figure 3. Sample images from MHS dataset. The colormap in black circle shows concentration maps corresponding to the plume mask shown in red. We are showing different types of leakage sources and land cover types. For better visualization, we plotted the binary mask on color image created using visible bands of hyperspectral image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.47, + 0.693 + ], + "angle": 0, + "content": "first stage is the bipartite matching between the predictions and the ground truths both in bounding box and mask prediction, and then second stage is loss calculation for the matched pairs. The bipartite matching employs the Hungarian algorithm [4] to find the optimal matching between the predictions and the ground truths. After this matching, every prediction is associated with a ground truth. Next, we calculate the \\( l_{1} \\) and \\( GIoU \\) loss on both box and mask predictions and cross entropy loss for class prediction [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.47, + 0.785 + ], + "angle": 0, + "content": "Inference: The inference pipeline is similar to training pipeline and can be implemented using approximately 50 lines of code. During inference, we first filter the detections with confidence levels below \\(50\\%\\) and a per-pixel max to determine which pixels are predicted to belong to a CH4 plume." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.391, + 0.817 + ], + "angle": 0, + "content": "4. Methane Hot Spots (MHS) dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Another significant contribution of this work is a large scale curated MHS dataset. It contains the AVIRIS-NG spectral data with wavelength ranging from \\(380nm\\) to \\(2510nm\\), a \\(5nm\\) sampling [21], and capturing 432 channels per pixel. The images from the flight-line are orthorec" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.089, + 0.912, + 0.275 + ], + "angle": 0, + "content": "
DatasetMHS (Ours) DatasetJPL-CH4 detection-V1.0 [39]
# plume sites3961161
# flightlines118546
# point source3675114
# diffused source28657
Time period2015 - 2022(8 years)2015(1 year)
Segmentation MaskYesYes
Bonding boxYesNo
Concentration mapYesNo
Number of Regions61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.284, + 0.895, + 0.397 + ], + "angle": 0, + "content": "Table 1. Statistics shows MHS dataset comparison with JPL-CH4-detection-V1.0 [39] dataset. Each flightline have multiple large and small plume sites. Each flightline have at least 4 plume sites. The Point Source represents high concentration \\((300kg/hr)\\) to leakage from sources like pipeline leak, storage tanks, oil and gas refineries. Diffused Source represent low concentration leakages from sources like biomass degradation in landfills. Our dataset is covers more diverse type of terrain over 6 states." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.575 + ], + "angle": 0, + "content": "tified and of size \\(\\sim 23K\\times \\sim 1.5K\\times 432\\). The only currently publicly-available dataset with methane plume segmentation masks is the JPL-CH4-detection-V1.0 [39] dataset released by JPL-NASA in 2017. The MHS dataset has approximately 4000 plume sites corresponding to approximately 1200 AVIRIS-NG flightlines as shown in Table 1. MHS also has higher diversity data with flight lines spanning from 2015-2022 and covering terrain from 6 states– California, Nevada, New Mexico, Colorado, Midland Texas, and Virginia." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.782 + ], + "angle": 0, + "content": "Data Pruning: We selected AVIRIS-NG flight lines over varying regions as it covers a wide variety of \\(\\mathrm{CH}_4\\) plume sources, such as leaks in oil and gas refineries, oil and gas extraction points, natural seeps, leaking underground storage tank, coal mines, dairy farms, landfill sites, and pipeline leaks. Along with varying emission sources, we selected regions with different types of ground terrains like, bare soil, rocks, mountains, light vegetation, water bodies and dense vegetation as shown with few samples in Fig. 3. Different types of ground terrain exhibit widely varying albedo and thus have a major impact on the quality of \\(\\mathrm{CH}_4\\) detections as shown in Fig. 4. Given this, training models with diverse ground terrain data leads to a more robust model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.793, + 0.874, + 0.809 + ], + "angle": 0, + "content": "4.1. Concentration map and Segmentation mask" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Concentration map is provided in the form of a matrix of spatial dimensions same as the flightline \\((\\sim 23k\\times \\sim 1.5k\\times 1)\\). There is one concentration map per flight-line (orthorectified). It shows methane concentration in parts-per-million (ppm) per-pixel on the ground. Pixel-regions with no methane presence are set to zero." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17613" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "Segmentation mask provided in the format of a \"png\" image file with three channels and of the same spatial dimension as the corresponding flight line (\\(\\sim 23k \\times \\sim 1.5k \\times 3\\)). The segmentation mask is obtained from the concentration mask file by setting all pixel values above zero to represent methane plumes. We manually annotated Point Source and Diffused Source based on the type of ground terrain and concentration of methane gas. Following the benchmark dataset [39], three channels are used to color code Point Source (Red) and Diffused Source (Green). The distinction of Point Source and Diffused Source is derived from the JPL-CH4-detection-V1.0 benchmark dataset [39]. Our annotation style is also consistent with the JPL-CH4-detection-V1.0 benchmark dataset [39], so that both datasets can be merged seamlessly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.343, + 0.364, + 0.358 + ], + "angle": 0, + "content": "4.1.1 Constructing Concentration map" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.686 + ], + "angle": 0, + "content": "Concentration maps are generated by mapping expert-annotated methane-plume concentration maps to the orthocorrected AVIRIS-NG flightlines. These methane plume annotations are systematically collected from a nonprofit [3] entity. They provide concentration masks of methane emissions in \\(150 \\times 150\\) size patches along with location information from different sources (airborne sensors [6], satellites [34]). In order to map these patches from different sources to the AVIRIS-NG flight-lines, we use the pixel coordinate locations provided for both the annotations and flight-lines. We use this information to create a homography transformation to map each pixel to its corresponding location in the flight-line. Fig. 3 shows a sample of varying types of terrains with \\(\\mathrm{CH}_4\\) segmentation mask in red and concentration mask in black circle. Details about matching the resolution, ortho-correction, and transformation are discussed in supplementary materials. The patch annotations are verified by experts visiting the physical location of emission the same day [2]. Most of the regions in California are verified by physical visits by California Air Resource Board [2,3]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.698, + 0.229, + 0.712 + ], + "angle": 0, + "content": "4.2. MHS Statistics" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.717, + 0.468, + 0.731 + ], + "angle": 0, + "content": "MHS statistics and properties are summarized in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.733, + 0.468, + 0.793 + ], + "angle": 0, + "content": "Annotations: MHS provides both segmentation masks and concentration maps which enable development of deep learning algorithms than can produce both \\(\\mathrm{CH}_4\\) plume location and concentration predictions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.794, + 0.468, + 0.854 + ], + "angle": 0, + "content": "Diversity: MHS dataset includes AVARIS-NG flightlines spanning 8 years (2015 - 2022) from six states in the U.S.: California, Nevada, New Mexico, Colorado, Texas, and Virginia." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Data Split: We divide MHS dataset into train/test splits of \\(80 - 20\\%\\) with overlapping time periods and locations. Our dataset covers 6 states. Each state has sub-regions/locations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "(e.g. Permian basin) that are covered by multiple nonoverlapping flightlines (\\(25k \\times 1.5k \\times 432\\) pixels). These flightlines are split into train and test sets. In each set, we create patches (\\(256 \\times 256 \\times 432\\) pixels) from the corresponding flightlines. From the patches/tiles, we take all positives patches (methane (\\(\\mathrm{CH}_4\\)) and randomly sample equal number of negative (\\(\\mathrm{no - CH_4}\\)) patches. This is done for both train and test sets separately to balance the data and we refer to Section 6.2 for detailed ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.252, + 0.708, + 0.269 + ], + "angle": 0, + "content": "5. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.441 + ], + "angle": 0, + "content": "Evaluation Metrics: Following the evaluation protocol of H-mrcnn [29] we report our performance in mean intersection-over-union (mIOU). Here, mIOU indicates the overlap between the predicted and the ground truth \\(\\mathrm{CH}_4\\) plume masks. ED represents the accuracy in plume core prediction. Additionally, as first stage of our two stage training procedure contains bounding box prediction, we also report our performance in predicting plume bounding boxes in terms of mean Average Precision (mAP) which tells us the effectiveness of MethaneMapper in eliminating the false positives in plume prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Data Pre-Processing: Each input hyperspectral image is approximately of size \\(25000 \\times 1500 \\times 432\\) taking up memory space of \\(55 - 60\\mathrm{GB}\\). We create tiles of each image in spatial domain, each tile is of size \\(256 \\times 256 \\times 432\\) [29] with an overlap of 128. The \\(\\mathrm{CH}_4\\) plume is available in very few pixels in the whole image, \\(90\\%\\) of the tiles are negative samples (no methane, just ground terrain). We can not use the whole hyperspectral image because of GPU memory limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details: The band-selectors module takes 432-channels hyperspectral image as input, the RGB band selector picks 60 channel from \\(400nm - 700nm\\) wavelength range and creates a 3-channel RGB image, the SWIR band selector picks 100 channel from wavelength range \\(2000nm - 2500nm\\). These input images are passed to two ResNet-50 [16] feature extractor backbones. The backbone networks are initialized with DETR [4] trained on COCO dataset [30] and input layer initialized randomly [15]. The transformer encoder-decoder and our query refiner have 6 layers and 8 heads. We initialized the transformer encoder-decoder with weights extracted and stripped from DETR [4] model. The dimension of transformer architecture is 256 and number of queries is 100. The SFG module takes in all 432-channels hyperspectral image and generates 1-channel output map of same spatial dimension as input. The feature extractor in SFG is ResNet-50 [16] initialized with DETR [4] trained on COCO dataset [30]. The decoder output embeddings are of size 512. The feature pyramid network in mask prediction module has 3 layers. More details are mentioned in supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17614" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.306 + ], + "angle": 0, + "content": "
MethodsBack boneSFG F.Ext.#paramsmAPmIOU
JPL-CH4-detection-v1.0 Dataset
1Hu et. alR-50-75M0.260.48
2H-mrcnnR-50-353M0.530.86
3MMR-50R-5080M0.630.91
MHS (Ours) Dataset
4SpectralFormerR-50-84M0.330.41
5UPSNet (stuff)R-50-69M0.320.38
6UPSNet (stuff + things)R-50-69M0.290.35
7DETRR-18*33M0.370.56
8DETRR-50*59M0.440.59
10R-18Linear Layer39M0.450.60
11MMR-18R-1844M0.520.63
12R-50R-5080M0.590.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.316, + 0.472, + 0.456 + ], + "angle": 0, + "content": "Table 2. Comparison with baselines. “-” represent Not Applicable and “*” represent no SFG module and a random query used for transformer decoder. The top section shows performance on JPL-CH\\(_4\\) dataset [39]. MethaneMapper achieves better results than heavily tuned H-mrcnn with \\(\\sim 5\\times\\) fewer parameters. The overall detection accuracy is higher on this dataset because the type of ground terrain is uniform across all flightlines. In MHS dataset, MM outperforms multiple baselines as shown in rows 4-12. MM accuracy is lower in MHS than JPL-CH\\(_4\\) dataset because MHS dataset has more variety of ground terrain spreading over 6 states" + }, + { + "type": "table", + "bbox": [ + 0.16, + 0.469, + 0.388, + 0.56 + ], + "angle": 0, + "content": "
MethodsmAPmIOU
LogReg [5]-0.05
SVM [36]-0.29
PCA + LogReg-0.06
PCA + SVM-0.31
MM (R-50)0.630.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.571, + 0.47, + 0.626 + ], + "angle": 0, + "content": "Table 3. Comparison with classical machine learning methods. \" - \" represent Not Available. The classical ML methods are not suited for the \\( {\\mathrm{{CH}}}_{4} \\) detection task. MethaneMapper outperforms all methods on JPL dataset [39]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.628, + 0.165, + 0.643 + ], + "angle": 0, + "content": "6. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.709 + ], + "angle": 0, + "content": "In this section we will discuss and validate all the design choices for MethaneMapper (MM) with ablations. We show that MM achieves state-of-the-art results in overall performance compared all other methods shown in Tables 2 & 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.308, + 0.734 + ], + "angle": 0, + "content": "6.1. Performance comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Deep Learning methods: We trained MM with ResNet-50 [16] backbone on the same dataset that H-mrcnn [29] (JPL-CH4-detection-V1.0 [39]) was trained on for fair comparison. To align with H-mrcnn we used the same split and input image size. The MM model with 80M parameters trained for 250 epochs outperforms by significant margin the H-mrcnn model with 352M parameters. Results are summarized in Table 2 that includes the performance of MM on the new larger MHS dataset. We note that though the code for H-mrcnn is available, many of the modules are deprecated and can not be reproduced. The 'Backbone' col" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "unn represents backbones used for feature extraction from input image,'SFG F.Ext.' represents the feature extractor in SFG module in MethaneMapper. We observed (qualitatively) that H-mrcnn fails to detect small \\(\\mathrm{CH}_4\\) plumes with concentration lower than \\(100\\mathrm{kg / hr}\\) while MM detects those." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.303 + ], + "angle": 0, + "content": "We did evaluation by implementing 3 baseline models [4, 17, 45] shown rows 4-8 of Table 2. These methods were not designed for \\(\\mathrm{CH}_4\\) detection task, therefore we needed to modify their input channel size. The poor performance of these methods may be attributed to the weak signal of interest in a high dimensional data, high number of confusers, and limited annotated data. Additionally, the only hyperspectral baseline method SpectralFormer [17] has low efficiency due its pixel-wise training scheme." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.893, + 0.544 + ], + "angle": 0, + "content": "Classical ML methods: We trained and tested multiple existing machine learning based approaches that are used for methane detection, performance shown in Table 3. Logistic regression (LogReg) [5] and multinomial logistic regression (MLR) [23] failed to produce any meaningful detection with \\(90\\%\\) false positive detections. We also trained a Support Vector Machine (SVM) [36,37] based classifier, it performed slightly better than LR and MLR methods with an IOU of \\(21\\%\\). SVMs are prone to false positives detections same as Gaussian Mixture Models [36]. We observed that all traditional methods are not suited for the task of \\(\\mathrm{CH}_4\\) detection. We also tested reducing the dimension using principal component analysis (PCA) or just taking bands which shows maximum \\(\\mathrm{CH}_4\\) absorption. In the later case, the traditional methods performed better than using all 432 bands, this backs our idea of just using bands from SWIR region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Qualitative results. Fig. 5 shows comparison of MM's mask and bounding box prediction with ground truth mask on different ground terrains. The Leakages are from different type of sources such as, oil refinery, pipeline and storage tank. MM makes correct predictions in varying scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.665, + 0.643 + ], + "angle": 0, + "content": "6.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.709 + ], + "angle": 0, + "content": "We did the experiments for ablation on MHS dataset with ResNet-50 as backbone and validate the design choices. One parameter is changed for each ablation and others kept at best settings. More ablations in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Spectral Feature Generator Module: In Table 2 lower section, we show the effectiveness of our SFG module for the query refiner block. Our baseline is standard implementation of DETR [4] for segmentation task represented by row-1 and row-2 of Tab. 2 lower section. Using \\(\\mathrm{CH}_4\\) candidates feature from SFG improves the bounding box detection performance by \\(0.14\\mathrm{mAP}\\) and mask prediction by \\(0.09\\mathrm{mIOU}\\). This demonstrate that guiding queries with \\(\\mathrm{CH}_4\\) candidates feature generated by SFG produces better embeddings as compared to random queries." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Along with this, we explored the provision of \\(\\mathrm{CH}_4\\) candidates feature at 2 places, (i) at input level concatenating" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.099, + 0.468, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.301, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Figure 4. Comparison of SLF with traditional filter in SFG module. White pixels represent methane and black no-methane. Red boundary represents ground-truth plume mask. SLF module generates better \\( \\mathrm{CH}_4 \\) candidates" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.468, + 0.474 + ], + "angle": 0, + "content": "it with \\( f_{comb} \\); and (ii) as input to query refiner. We see an improvement of 0.09 mAP and 0.08 mIOU when SFG module output is passed to query refiner. We hypothesize that this is because on concatenating with input features, the CH\\(_4\\) candidates feature information gets lost, while as cross-attention with queries reduces the search space for decoder and generate better embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.469, + 0.534 + ], + "angle": 0, + "content": "We also experimented with different types of feature extractors for SFG module, and observed that a Resnet18 or Resnet50 [16] is more effective than a 2 linear layer feature extractor as shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.469, + 0.702 + ], + "angle": 0, + "content": "Spectral Linear Filter: We experimented with SLF for computing covariance \\((Cov)\\) using different subsets of columns in the input hyperspectral image. We observed that the SLF is most effective when covariance is computed class-wise based on land cover. Class-wise \\(Cov\\) ensures that the radiance absorption by ground terrain is same for all the pixels while computing \\(\\mathrm{CH}_4\\) enhancement. As can be seen in Fig. 4, SLF amplifies \\(\\mathrm{CH}_4\\) candidate detection and reduces false positives. SLF leads to a \\(0.03\\mathrm{mAP}\\) improved in detection compared to traditional filters. The prediction from MM is shown row-1 of Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.853 + ], + "angle": 0, + "content": "Geographic generalization: To assess the geographical generalization capabilities of MM, we trained it on MHS data from all states except California and tested it on flightlines from California. We observed a slight drop of 0.04 mAP in detections. However, when trained on all data except Virginia, we noticed a significant drop of \\(0.09\\mathrm{mAP}\\) in detections. We attribute this to the fact that the land cover in Virginia is dense and moist vegetation, has a lower solar reflectance compared to the arid regions of California, Texas, and Nevada." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Temporal generalization: Testing MM on 2015 after training on data from 2016-2022 showed no performance drop." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Unbalanced test set: MM's performance dropped by 0.05" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.089, + 0.702, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.203, + 0.702, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.315, + 0.703, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.089, + 0.85, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.203, + 0.851, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.315, + 0.851, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.43, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Figure 5. Sample ground truths and predictions on MHS dataset. We show robustness of MethaneMapper predictions on different kind of ground terrain, rows 1 and 3 shows leakage at a refinery, row 2 shows leakage from pipeline in agricultural land, row 4 shows leakage from storage tank with concrete background." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.577 + ], + "angle": 0, + "content": "mAP on an unbalanced test set with only \\(10\\%\\) positive samples \\(\\mathrm{(CH_4)}\\) and \\(90\\%\\) negative samples (no- \\(\\mathrm{CH}_4\\)). This highlights the challenges in \\(\\mathrm{CH}_4\\) detection. Future work will address this issue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.593, + 0.618, + 0.608 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.782 + ], + "angle": 0, + "content": "This paper presents MethaneMapper - a hyperspectral Transformer for methane plume detection. It utilize spectral and spatial correlations using a spectral feature generator and a query refiner, to accurately delineate the \\(\\mathrm{CH}_4\\) plumes. Additionally, we curated a large-scale dataset for the task, a first of its kind, which will be made available to all researchers. The proposed MethaneMapper significantly improves upon the current methods in terms of detection and localization accuracy, as our extensive experiments demonstrate. Future work will extend the model to global monitoring [27] using multispectral satellite imaging data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.803, + 0.679, + 0.82 + ], + "angle": 0, + "content": "8. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research is partially supported by the following grants: NSF award SI2-SSI #1664172 and the US Army Research Laboratory (ARL) under agreement number W911NF2020157. We also thank Aditya Ramakrishnan in helping with the initial data cleaning process." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17616" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.473, + 0.213 + ], + "angle": 0, + "content": "[1] Alana K. Ayasse, Andrew K. Thorpe, Dar A. Roberts, Christopher C. Funk, Philip E. Dennison, Christian Frankenberg, Andrea Steffke, and Andrew D. Aubrey. Evaluating the effects of surface properties on methane retrievals using a synthetic airborne visible/infrared imaging spectrometer next generation (aviris-ng) image. Remote Sensing of Environment, 215:386-397, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.243 + ], + "angle": 0, + "content": "[2] CALIFORNIA AIR RESOURCE BOARD. Green house gas inventory by california air resource board, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] INC CARBON MAPPER. Carbon mapper methane emission exploration, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[5] Qi Cheng, Pramod K Varshney, and Manoj K Arora. Logistic regression for feature selection and soft classification of remote sensing data. IEEE Geoscience and Remote Sensing Letters, 3(4):491-494, 2006. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[6] E. Knapp David, Heckler Joseph, Seely Megs, and P. Asner Gregory. Global airborne observatory visible to infrared imaging spectrometer report, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[7] C Frankenberg, U Platt, and T Wagner. Iterative maximum a posteriori (imap)-doas for retrieval of strongly absorbing trace gases: Model studies for ch 4 and co 2 retrieval from near infrared spectra of sciamachy onboard envisat. Atmospheric Chemistry and Physics Discussions, 4(5):6067-6106, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[8] Christian Frankenberg, U Platt, and T Wagner. Iterative maximum a posteriori (imap)-doas for retrieval of strongly absorbing trace gases: Model studies for ch 4 and co 2 retrieval from near infrared spectra of sciamachy onboard envisat. Atmospheric Chemistry and Physics, 5(1):9-22, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[9] Christian Frankenberg, Andrew K Thorpe, David R Thompson, Glynn Hulley, Eric Adam Kort, Nick Vance, Jakob Borchardt, Thomas Krings, Konstantin Gerilowski, Colm Sweeney, et al. Airborne methane remote measurements reveal heavy-tail flux distribution in four corners region. Proceedings of the national academy of sciences, 113(35):9734-9739, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[10] Christopher C Funk, James Theiler, Dar A Roberts, and Christoph C Borel. Clustering to improve matched filter detection of weak gas plumes in hyperspectral thermal imagery. IEEE transactions on geoscience and remote sensing, 39(7):1410-1420, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[11] Bo-Cai Gao. Ndwi—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote sensing of environment, 58(3):257-266, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[12] Utsav B Gewali, Sildomar T Monteiro, and Eli Saber. Machine learning based hyperspectral image analysis: a survey. arXiv preprint arXiv:1802.08701, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] IE Gordon, LS Rothman, RJ Hargreaves, R Hashemi, EV Karlovets, FM Skinner, EK Conway, C Hill, RV Kochanov, Y Tan, et al. The hitran2020 molecular spectroscopic" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "database. Journal of quantitative spectroscopy and radiative transfer, 277:107949, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[17] Danfeng Hong, Zhu Han, Jing Yao, Lianru Gao, Bing Zhang, Antonio Plaza, and Jocelyn Chanussot. Spectralformer: Rethinking hyperspectral image classification with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-15, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.797, + 0.377 + ], + "angle": 0, + "content": "[18] Paris IEA. Methane from oil & gas, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[19] ASM Iftekhar, Hao Chen, Kaustav Kundu, Xinyu Li, Joseph Tighe, and Davide Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5353-5363, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[20] ASM Iftekhar, Satish Kumar, R Austin McEver, Suya You, and BS Manjunath. Gtnet: Guided transformer network for detecting human-object interactions. arXiv preprint arXiv:2108.00596, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[21] California Institute of Technology Jet Propulsion Laboratory. Airborne visible infrared imaging spectrometer - next generation (aviris-ng) overview, 2009. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[22] Siraput Jongaramrungruang, Andrew K Thorpe, Georgios Matheou, and Christian Frankenberg. Methanet—an ai-driven approach to quantifying methane point-source emission from high-resolution 2-d plume imagery. Remote Sensing of Environment, 269:112809, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[23] Mahdi Khodadadzadeh, Jun Li, Antonio Plaza, and José M Bioucas-Dias. A subspace-based multinomial logistic regression for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 11(12):2105-2109, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] Stefanie Kirschke, Philippe Bousquet, Philippe Ciais, Marielle Saunois, Josep G Canadell, Edward J Dlugokencky, Peter Bergamaschi, Daniel Bergmann, Donald R Blake, Lori Bruhwiler, et al. Three decades of global methane sources and sinks. Nature geoscience, 6(10):813-823, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[25] Satish Kumar, ASM Iftekhar, Michael Goebel, Tom Bullock, Mary H MacLean, Michael B Miller, Tyler Santander, Barry Giesbrecht, Scott T Grafton, and BS Manjunath. Stressnet: detecting stress in thermal videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 999-1009, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Satish Kumar, ASM Iftekhar, Ekta Prashnani, and BS Manjunath. Locl: Learning object-attribute composition using localization. arXiv preprint arXiv:2210.03780, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17617" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Satish Kumar, William Kingwill, Orbio Earth, Rozanne Mouton, Wojciech Adamczyk, Robert Huppertz, and Evan Sherwin. Guided transformer network for detecting methane emissions in sentinel-2 satellite imagery. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.202 + ], + "angle": 0, + "content": "[28] Satish Kumar, Rui Kou, Henry Hill, Jake Lempges, Eric Qian, and Vikram Jayaram. In-situ water quality monitoring in oil and gas operations. arXiv preprint arXiv:2301.08800, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.285 + ], + "angle": 0, + "content": "[29] Satish Kumar, Carlos Torres, Oytun Ulutan, Alana Ayasse, Dar Roberts, and BS Manjunath. Deep remote sensing methods for methane detection in overhead hyperspectral imagery. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1776-1785, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.288, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[31] META-STANFORD. Methane emissions technology alliance (meta), stanford natural gas initiative, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.387, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[32] MGISGeography. Ndvi (normalized difference vegetation index), 1979. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.414, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[33] Gunnar Myhre, Drew Shindell, and Julia Pongratz. Anthropogenic and natural radiative forcing. 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.443, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[34] Darius Phiri, Matamyo Simwanda, Serajis Salekin, Vincent R Nyirenda, Yuji Murayama, and Manjula Ranagalage. Sentinel-2 data for land cover/use mapping: a review. Remote Sensing, 12(14):2291, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[35] Dar A Roberts, Eliza S Bradley, Ross Cheung, Ira Leifer, Philip E Dennison, and Jack S Margolis. Mapping methane emissions from a marine geological seep source using imaging spectrometry. Remote Sensing of Environment, 114(3):592-606, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.568, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[36] CA Shah, PK Varshney, and MK Arora. Ica mixture model algorithm for unsupervised classification of remote sensing imagery. International Journal of Remote Sensing, 28(8):1711-1731, 2007. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.624, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[37] David MJ Tax and Robert PW Duin. Support vector data description. Machine learning, 54(1):45-66, 2004. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.652, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[38] DR Thompson, I Leifer, H Bovensmann, M Eastwood, M Fladeland, C Frankenberg, K Gerilowski, RO Green, S Kratwurst, T Krings, et al. Real-time remote detection and measurement for airborne imaging spectroscopy: a case study with methane. Atmospheric Measurement Techniques, 8(10):4383-4397, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.735, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[39] David R Thompson, Anuj Karpatne, Imme Ebert-Uphoff, Christian Frankenberg, Andrew K Thorpe, Brian D Bue, and Robert O Green. Isgeo dataset jpl-ch4-detection-2017-v1. 0: A benchmark for methane source detection from imaging spectrometer data. 2017. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.805, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[40] Andrew K Thorpe, Christian Frankenberg, David R Thompson, Riley M Duren, Andrew D Aubrey, Brian D Bue, Robert O Green, Konstantin Gerilowski, Thomas Krings, Jakob Borchardt, et al. Airborne doas retrievals of methane, carbon dioxide, and water vapor concentrations at high spatial resolution: application to aviris-ng. Atmospheric Measurement Techniques, 10(10):3833-3850, 2017. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[41] Andrew K Thorpe, Dar A Roberts, Eliza S Bradley, Christopher C Funk, Philip E Dennison, and Ira Leifer. High resolution mapping of methane emissions from marine and terrestrial sources using a cluster-tuned matched filter technique and imaging spectrometry. Remote Sensing of Environment, 134:305-318, 2013. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[42] George Turin. An introduction to matched filters. IRE transactions on Information theory, 6(3):311-329, 1960. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[43] Oytun Ulutan, ASM Iftekhar, and Bangalore S Manjunath. Vsgnet: Spatial attention network for detecting human object interactions using graph convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13617-13626, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[44] Foreign Agriculture Service US Dept. of Agriculture. Normalized difference vegetation index (ndvi), 1969. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.893, + 0.374 + ], + "angle": 0, + "content": "[45] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8826, 2019. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17618" + } + ] +] \ No newline at end of file diff --git a/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_origin.pdf b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..00553a6953293cd3effaf963e58907e3e107a287 --- /dev/null +++ b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/380575b8-d276-40f4-954b-81a84c333f47_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad0dda1b50e323fae8c470530dd51d985fce10f301daa2160130b007df3eadb4 +size 4426186 diff --git a/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/full.md b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9f0d5eab7b308ed3601a7eec76ce7d926a0c280 --- /dev/null +++ b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/full.md @@ -0,0 +1,305 @@ +# MethaneMapper: Spectral Absorption aware Hyperspectral Transformer for Methane Detection + +Satish Kumar + +satishkumar@ucsb.edu + +Ivan Arevalo + +ifa@ucsb.edu + +ASM Iftekhar + +iftekhar@ucsb.edu + +B S Manjunath + +manj@ucsb.edu + +Department of Electrical and Computer Engineering +University of California Santa Barbara + +# Abstract + +Methane $(\mathsf{CH}_4)$ is the chief contributor to global climate change. Recent Airborne Visible-Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) has been very useful in quantitative mapping of methane emissions. Existing methods for analyzing this data are sensitive to local terrain conditions, often require manual inspection from domain experts, prone to significant error and hence are not scalable. To address these challenges, we propose a novel end-to-end spectral absorption wavelength aware transformer network, MethaneMapper, to detect and quantify the emissions. MethaneMapper introduces two novel modules that help to locate the most relevant methane plume regions in the spectral domain and uses them to localize these accurately. Thorough evaluation shows that MethaneMapper achieves 0.63 mAP in detection and reduces the model size (by $5\times$ ) compared to the current state of the art. In addition, we also introduce a large-scale dataset of methane plume segmentation mask for over 1200 AVIRIS-NG flight lines from 2015-2022. It contains over 4000 methane plume sites. Our dataset will provide researchers the opportunity to develop and advance new methods for tackling this challenging green-house gas detection problem with significant broader social impact. Dataset and source code link1. + +# 1. Introduction + +We consider the problem of detecting and localizing methane $\left(\mathrm{CH}_4\right)$ plumes from hyperspectral imaging data. Detecting and localizing potential $\mathrm{CH}_4$ hot spots is a necessary first step in combating global warming due to greenhouse gas emissions. Methane gas is estimated to contribute $20\%$ of global warming induced by greenhouse gasses [24] with a Global Warming Potential (GWP) 86 times higher than carbon dioxide $(\mathrm{CO}_2)$ in a 20 year period [33]. To + +put into perspective, the amount of environmental damage that $\mathrm{CO}_{2}$ can do in 100 years, $\mathrm{CH}_4$ can do in 1.2 years. Hence it is critical to monitor and curb the $\mathrm{CH}_4$ emissions. While $\mathrm{CH}_4$ emission has many sources, of particular interest are those from oil and natural gas industries. According to the United States Environmental Protection Agency report, $\mathrm{CH}_4$ emissions from these industries accounts to 84 million tons per year [18]. These $\mathrm{CH}_4$ emissions emanate from specific locations, mainly from pipeline leakages, storage tank leak or leakage from oil extraction point. + +Current efforts to detect these sources mostly depend on aerial imagery. The Jet Propulsion Laboratory (JPL) has conducted thousands of aerial surveys in the last decade to collect data using an airborne sensor AVIRIS-NG [21]. Several methods have been proposed to detect potential emission sites from such imagery, for example, see [8,9,35,38, 40, 41]. However, these methods are in general very sensitive to background context and land-cover types, resulting in a large number of false positives that often require significant domain expert time to correct the detections. The primary reason is that these pixel-based methods are solely dependent on spectral correlations for detection. Spatial information can be very effective in reducing these false positives as $\mathrm{CH}_4$ plumes exhibit a plume-like structure morphology. There has been recent efforts in utilizing spatial correlation using deep learning methods [22, 29], however, these works don't leverage spectral properties to filter out confusers. For example, methane has similar spectral properties as white-painted commercial roofs or paved surfaces such as airport asphalts [1]. This paper presents a novel deep-network based solution to minimize the effects of such confusers in accurately localizing methane plumes. + +Our proposed approach, referred to as the MethaneMapper (MM), adapts the DETR [4], a transformer model that combines the spectral and spatial correlations in the imaging data to generate a map of potential methane $\mathrm{(CH_4)}$ plume candidates. These candidates reduce the search space for a hyperspectral decoder to detect $\mathrm{CH}_4$ plumes and re + +move potential confusers. MM is a light-weight end-to-end single-stage $\mathrm{CH}_4$ detector and introduces two novel modules: a Spectral Feature Generator and a Query Refiner. The former generates spectral features from a linear filter that maximizes the $\mathrm{CH}_4$ -to-noise ratio in the presence of additive background noise, while the latter integrates these features for decoding. + +A major bottle neck for development of $\mathrm{CH}_4$ detection methods is the limited availability of public training data. To address this, another significant contribution of this research is the introduction of a new Methane Hot Spots (MHS) dataset, largest of its kind available for computer vision researchers. MHS is curated by systematically collecting information from different publicly available datasets (airborne sensor [6], Non-profits [3,31] and satellites [34]) and generating the annotations as described in Section 4.1.1. This curated dataset contains methane segmentation masks for over 1200 AVIRIS-NG flight lines from years 2015 to 2022. Each flight line contains anywhere from $3 - 4\mathrm{CH}_4$ plume sites for a total of 4000 in the MHS dataset. + +Our contributions can be summarized as follows: + +1. We introduce a novel single-stage end-to-end approach for methane plume detection using a hyperspectral transformer. The two modules, Spectral Feature Generator and Query Refiner, work together to improve upon the traditional transformer design and enable localization of potential methane hot spots in the hyperspectral images using a Spectral-Aware Linear Filter and refine the query representation for better decoding. +2. A new Spectral Linear Filter (SLF) improves upon traditional linear filters by strategically picking correlated pixels in spectral domain to better whiten background distribution and amplify methane signal. +3. A new benchmark dataset, MHS, provides the largest $(\sim 35\times)$ publicly available dataset of annotated AVIRIS-NG flight lines from years 2015-2022. + +# 2. Related Works + +Our work is at the intersection of hyperspectral data for $\mathrm{CH}_4$ detection, deterministic linear filtering methods for spectral features and encoder-decoder based transformer. A review of the pertinent related works is given below. + +There are several recent papers on detecting methane plumes from the airborne imaging spectrometer AVIRIS-NG [21]. This includes the Iterative Maximum a Posterior Differential Optical Absorption Spectroscopy algorithm (IMAP-DOAS) [7, 8] and matched filters [9, 35, 38, 40, 41]. IMAP-DOAS requires data from 2 hyperspectral sensors, one airborne and another on ground, hence not very practical for most application scenarios. Matched-filter based methods use background statistics to normalize + +![](images/0c642682fdf22bacd4de13a996f1a9c7a413a775a8368849b65ce61b639bdef5.jpg) +Figure 1. Depiction of data collection process. Each flightline is $\sim 300$ km long. An array of 598 sensors records data at $1.5\mathrm{m}$ pixel spatial resolution. All flightlines are ortho-corrected. Each data-cube is of dimension $\sim 25000\times \sim 1500\times 432$ + +the spectral signals and match with the $\mathrm{CH}_4$ spectral signature at every spatial location (pixel-wise). This process, however, is sensitive to surface albedo and land cover with spectral absorption similar to $\mathrm{CH}_4$ , leading to spurious detections. Domain experts must then manually inspect each flight line to identify and delineate real $\mathrm{CH}_4$ plumes [39]. To suppress the effect of false positives due to variability of elements on ground, Christopher et.al. [10] and Thorbe et. al. [41] introduced cluster-tuned matched filter. It involves clustering the pixels with similar spectral properties using k-means clustering. Both IMAP-DOAS and all versions of matched filters are heavily prone to false positives as the information is processed pixel-wise. + +Machine learning approaches have been used for target detection, including $\mathrm{CH}_4$ identification, in hyperspectral imagery [12,17,28,36,37]. Similar to matched-filtering, these methods do not take into account the influence of confusers on the $\mathrm{CH}_4$ spectral signature and have similar issues concerning false positives. Recently introduced deep learning based H-mrcnn model [29] focus on capturing spatial correlation. H-mrcnn is an ensemble of mask-rcnncite networks processing blocks of hyperspectral data. This block processing in the spectral domain is inefficient and often results in overall poor performance. Methanet [22] is a more recent work focusing on estimating methane concentration from matched-filter data. In this regard, our proposed MethaneMapper uses both spectral and spatial correlation to accurately delineates $\mathrm{CH}_4$ plumes. + +Datasets: The only dataset publicly available with annotation for $\mathrm{CH}_4$ plume detection is JPL- $\mathrm{CH}_4$ -detection2017-V1.0 dataset [39]. It contains only 46 AVIRIS-NG [21] flight lines in the US Four-Corners region. Deep learning architectures require a large number of annotated samples, and for this reason we introduce the new MHS dataset with over 1200 annotated flightlines and $\sim 4000$ plume sites. + +![](images/7da038c7b5b519847a23b51c3895625349fa028d309a2347e15dc08561707b6b.jpg) +Figure 2. Overview of MethaneMapper (MM) architecture. Given a hyperspectral image, our RGB $(400\mathrm{nm} - 700\mathrm{nm})$ and SWIR $(2000\mathrm{nm} - 2500\mathrm{nm})$ band-pass filters passes a subset of channels in desired wavelength range and feed them to CNN backbones (ResNet) to extract features. These features are concatenated and fed to Transformer Encoder. Parallelly, our Spectral Feature Generator (SFG) modules take in all channels of input image and generate methane candidates features. Next these candidates are sent to Query Refiner (QR) to refine queries. Then these queries decoded using encoded feature from Transformer Encoder. Finally each decoded query is used to predict a plume mask via Mask Prediction and, bounding box and class via FFNs (Feed Forward Network). + +# 3. MethaneMapper (MM) Architecture + +# 3.1. Data Overview + +AVIRIS-NG hyperspectral imaging sensors capture spectral radiance values from $\mathrm{N}_0$ ( $\mathrm{N}_0 = 432$ ) channels corresponding to wavelengths ranging from $400nm - 2500nm$ as shown in Fig. 1. The complete hyperspectral image is represented as $\mathbf{x} \in \mathbb{R}^{H_0 \times W_0 \times N_0}$ where $H_0, W_0$ are the height & width, respectively, and $N_0 = 432$ is number of channels. This hyperspectral data includes a very weak signature of $\mathrm{CH}_4$ around $2100 - 2400nm$ , conflated with radiations from the surrounding land cover and background clutter. A single flight-line could be over a couple miles long (about $25\mathrm{K}$ pixels in one of the dimensions), with an array of sensors recording the data at $1.5\mathrm{m/pixel}$ resolution. The images are orthorectified before processing. + +# 3.2. Technical Overview + +Referring to Fig.2, MM contains the following main components: (i) 2 CNN backbones to extract a compact feature representation of the spectral regions of interest from the hyperspectral image, (ii) a Spectral Feature Generator (SFG), and (iii) a Query Refiner (QR) in between an encoder-decoder pair (inspired by GTNet [20], SSRT [19]). The hyperspectral image is first processed through two separate band-pass filters to select the channels in visible $(400 - 700nm)$ and short-wave infrared (SWIR)(2000 - 2500nm) wavelength regions, and are then passed through CNN backbones. Output of these backbones are concatenated together and then encoded using a transformer encoder. + +The SFG (Sec. 3.4) takes in all channels of the hyperspectral image and process them through a spectral linear filter. The SFG exploits the spectral correlation to generate methane candidates feature maps and passes them to QR. + +The QR (Sec. 3.5) uses these methane candidates to refine the learnable queries. Our hyperspectral decoder takes the encoded features from the encoder and refined queries from QR to generate the embeddings. The mask-prediction layer processes these embeddings along with the feature pyramid from the backbone layers to generate the final methane-plume segmentation prediction. + +These individual blocks are discussed in more detail below. + +# 3.3. Bandpass filtering for the Encoder + +The HSI is processed by two parallel band-pass filters; a visible wavelength $(400 - 700nm)$ (RGB) and a shortwave infrared wavelength $(2000 - 2500nm)$ (SWIR) bandpass filter. The RGB filter results in a 3 channel output corresponding to the normal red, green, and blue wavelengths. The SWIR generates channels, approximately $5\mathrm{nm}$ apart. The filtered outputs are $\mathbf{x}_{rgb} \in \mathbb{R}^{H_0 \times W_0 \times 3}$ and $\mathbf{x}_{swir} \in \mathbb{R}^{H_0 \times W_0 \times 100}$ . Using $\mathbf{x}_{rgb}$ and $\mathbf{x}_{swir}$ , two conventional CNN backbones (e.g. ResNet-50 [16, 25]) generate two feature maps respectively of size $\in \mathbb{R}^{H \times W \times N}$ . Here $H = \frac{H_0}{32}$ , $W = \frac{W_0}{32}$ and $N = 2048$ typically. We concatenate these feature maps along channel dimension and project through a $1 \times 1$ convolution layer to retain channel dimension of $N$ . The resulting output is $f_{comb} \in \mathbb{R}^{H \times W \times N}$ . + +Following the standard architecture of transformer encoder from previous works [4, 19, 20, 26, 43], we reduce the channel dimension of $f_{comb}$ using $1 \times 1$ convolution to $f_z \in \mathbb{R}^{H \times W \times d}$ and supplement position information by adding a fixed positional embedding $p \in \mathbb{R}^{H \times W \times d}$ . The encoder consists of a stack of multi-head self-attention modules and feed-forward networks (FFN). The encoded feature map is $f_e \in \mathbb{R}^{H \times W \times d}$ : + +$$ +f _ {e} = \operatorname {E n c o d e r} \left(f _ {z}, p\right) \tag {1} +$$ + +# 3.4. Spectral Feature Generator (SFG) + +In parallel, the input hyperspectral image is processed by the SFG module to generate methane candidates feature map $f_{mc}$ , providing the QR module with spatial information to help the network delineate the methane plumes. + +The SFG consist of a spectral linear filter (SLF) and a Feature Extractor (e.g.ResNet-50 [16]). The most common linear filtering approach for detecting $\mathrm{CH}_4$ is to take each pixel from the input hyperspectral image $\{\mathbf{x}_{ij} \mid \mathbf{x}_{ij} \in \mathbb{R}^{1 \times 1 \times N_0}\}_{i,j=1}^{H_0, W_0}$ and project it onto a $\mathrm{CH}_4$ spectral absorption signature vector of same size [13]. This is to reduce the interference from ground terrain and amplify the $\mathrm{CH}_4$ visibility in that pixel. Accurately modeling SLF is critical given that it is designed to reduce ground terrain interference. To model SLF we use the most common approach to matched filtering from information theory [42]. + +Spectral Linear Filter (SLF): The design of SLF is dependent on the spectral absorption pattern of $\mathrm{CH}_4$ gas [13] and distribution of ground terrain. Since our signal of interest, $\mathrm{CH}_4$ , is very weak, traditional methods of linear filtering [40, 41] are not effective. The conventional methods to whiten the ground terrain noise includes calculating the covariance $(\mathbf{Cov} \in \mathbb{R}^{N_0 \times N_0})$ of background by selecting a set of 10-15 adjacent columns $\{\mathbf{x}_i \mid \mathbf{x}_i \in \mathbb{R}^{1 \times H_0 \times N_0}\}_{i=1}^{W_0}$ . However, in a given flight-line, the terrain changes frequently, from water bodies to bare soil, vegetation, buildings and other urban structures. Therefore single approximation of the covariance can not provide correct estimate of $\mathrm{CH}_4$ and a localized context-based whitening will be more effective. To address this problem, we took a very simple and effective approach of doing land cover classification and segmentation [11, 32, 44], and then compute covariance per class from the land cover. More details in supplementary materials. This improves the quality of methane candidates in presence of confusers (materials with similar spectral absorption patterns as $\mathrm{CH}_4$ ) and also in cases where $\mathrm{CH}_4$ concentration is low. The final SLF design with per class covariance is: + +$$ +\mathbf {S L F} \left(\mathbf {x} _ {i j}\right) = \frac {\left(\mathbf {x} _ {i j} - \mu_ {k}\right) ^ {T} \mathbf {C o v} _ {k} ^ {- 1} t}{\sqrt {t ^ {T} \mathbf {C o v} _ {k} ^ {- 1}} t} \forall (i, j) \in \text {c l a s s} k \tag {2} +$$ + +where $t$ represents the spectral absorption pattern [13] of $\mathrm{CH}_4$ gas, and $\mathbf{Cov}_k$ , $\mu_k$ are the covariance and mean of $k^{th}$ class respectively. $\mathbf{x}_{ij}$ represents the pixel in input hyperspectral image at $(i,j)$ index in $k^{th}$ class. This operation generates a 2-D spatial $\mathrm{CH}_4$ candidates map of size $\mathbb{R}^{H_0\times W_0}$ . Next this $\mathrm{CH}_4$ candidates map is fed to a Feature Extractor to generate $\mathrm{CH}_4$ candidates feature map $f_{mc}$ . Details of the land cover segmentation/classification and complete SLF derivation are in the Supplementary materials. + +$$ +f _ {m c} = \operatorname {F e a t u r e E x t r a c t o r} (\mathbf {S L F} (\mathbf {x} _ {i j}) \forall i, j) \tag {3} +$$ + +# 3.5. Query Refiner (QR) + +Next the methane candidate feature map $f_{mc} \in \mathbb{R}^{H \times W \times d}$ is fed to the QR module along with a set of 100 learnable queries $Q \in \mathbb{R}^{100 \times d}$ . The $f_{mc}$ refines the learnable queries via cross-attention mechanism. This operation provides a narrow search space for the queries. The QR module follows a transformer decoder-like architecture inspired from [19,20]. The randomly initialized queries $Q \in \mathbb{R}^{100 \times d}$ are first passed through a self-attention layer to attend to themselves. Next, these queries attend to our methane candidates feature map $f_{mc}$ from SFG module through a cross-attention layer. The methane candidates feature map serves as key-values pairs in our attention architecture. The output of QR is $Q_{ref}$ . + +$$ +Q _ {r e f} = \mathbf {Q R} \left(f _ {m c}, Q\right) \tag {4} +$$ + +# 3.6. Hyperspectral Decoder + +The $Q_{ref}$ is fed to the decoder module along with encoder output $f_{e}$ to generate output embeddings. Our hyperspectral decoder follows the standard architecture with a minor difference. There are no self-attention layers, just stack of multi-headed cross attention layers. The refined queries are transformed into output embeddings $E_{out} \in \mathbb{R}^{100 \times d}$ . + +$$ +E _ {\text {o u t}} = \operatorname {D e c o d e r} \left(f _ {e}, p, Q _ {\text {r e f}}\right) \tag {5} +$$ + +# 3.7. Box and Mask Prediction + +The decoder output embeddings $(E_{out})$ are fed to two Feed Forward Network (FFNs) and a Mask prediction layer. The outputs of the FFNs are the bounding boxes covering each $\mathrm{CH}_4$ plume and a confidence score corresponding to each box. The mask-prediction module follows the standard segmentation head of DETR [4]. It computes multi-head attention scores of each embedding over the $f_{e}$ (Eq. 1), generating a low-resolution heatmap for each embedding. To make the final prediction a Feature Pyramid Network [14] like structure is used. Each heatmap is designed to capture one methane plume. A simple thresholding is used to merge the heatmaps as final segmentation mask. + +$$ +m a s k = \operatorname {M a s k} _ {-} \operatorname {p r e d} \left(E _ {\text {o u t}}, f _ {e}, f _ {\text {c o m b}}\right) \tag {6} +$$ + +# 3.8. Training and Inference + +We train MethaneMapper in two stages; first we train bounding box detection corresponding to each $\mathrm{CH}_4$ plume, and second by freezing the box detection network and training only the mask prediction module. We also trained both box and mask prediction modules end-to-end and achieved similar performance. We use a similar two-stage loss strategy for training MethaneMapper as that used in DETR [4]: + +![](images/4656e40c8c344ccdb4e448e5e781ba0cd9fac1831c406ca2c2e680ff5c98466a.jpg) +Figure 3. Sample images from MHS dataset. The colormap in black circle shows concentration maps corresponding to the plume mask shown in red. We are showing different types of leakage sources and land cover types. For better visualization, we plotted the binary mask on color image created using visible bands of hyperspectral image. + +first stage is the bipartite matching between the predictions and the ground truths both in bounding box and mask prediction, and then second stage is loss calculation for the matched pairs. The bipartite matching employs the Hungarian algorithm [4] to find the optimal matching between the predictions and the ground truths. After this matching, every prediction is associated with a ground truth. Next, we calculate the $l_{1}$ and $GIoU$ loss on both box and mask predictions and cross entropy loss for class prediction [4]. + +Inference: The inference pipeline is similar to training pipeline and can be implemented using approximately 50 lines of code. During inference, we first filter the detections with confidence levels below $50\%$ and a per-pixel max to determine which pixels are predicted to belong to a CH4 plume. + +# 4. Methane Hot Spots (MHS) dataset + +Another significant contribution of this work is a large scale curated MHS dataset. It contains the AVIRIS-NG spectral data with wavelength ranging from $380nm$ to $2510nm$ , a $5nm$ sampling [21], and capturing 432 channels per pixel. The images from the flight-line are orthorec + +
DatasetMHS (Ours) DatasetJPL-CH4 detection-V1.0 [39]
# plume sites3961161
# flightlines118546
# point source3675114
# diffused source28657
Time period2015 - 2022(8 years)2015(1 year)
Segmentation MaskYesYes
Bonding boxYesNo
Concentration mapYesNo
Number of Regions61
+ +Table 1. Statistics shows MHS dataset comparison with JPL-CH4-detection-V1.0 [39] dataset. Each flightline have multiple large and small plume sites. Each flightline have at least 4 plume sites. The Point Source represents high concentration $(300kg/hr)$ to leakage from sources like pipeline leak, storage tanks, oil and gas refineries. Diffused Source represent low concentration leakages from sources like biomass degradation in landfills. Our dataset is covers more diverse type of terrain over 6 states. + +tified and of size $\sim 23K\times \sim 1.5K\times 432$ . The only currently publicly-available dataset with methane plume segmentation masks is the JPL-CH4-detection-V1.0 [39] dataset released by JPL-NASA in 2017. The MHS dataset has approximately 4000 plume sites corresponding to approximately 1200 AVIRIS-NG flightlines as shown in Table 1. MHS also has higher diversity data with flight lines spanning from 2015-2022 and covering terrain from 6 states– California, Nevada, New Mexico, Colorado, Midland Texas, and Virginia. + +Data Pruning: We selected AVIRIS-NG flight lines over varying regions as it covers a wide variety of $\mathrm{CH}_4$ plume sources, such as leaks in oil and gas refineries, oil and gas extraction points, natural seeps, leaking underground storage tank, coal mines, dairy farms, landfill sites, and pipeline leaks. Along with varying emission sources, we selected regions with different types of ground terrains like, bare soil, rocks, mountains, light vegetation, water bodies and dense vegetation as shown with few samples in Fig. 3. Different types of ground terrain exhibit widely varying albedo and thus have a major impact on the quality of $\mathrm{CH}_4$ detections as shown in Fig. 4. Given this, training models with diverse ground terrain data leads to a more robust model. + +# 4.1. Concentration map and Segmentation mask + +Concentration map is provided in the form of a matrix of spatial dimensions same as the flightline $(\sim 23k\times \sim 1.5k\times 1)$ . There is one concentration map per flight-line (orthorectified). It shows methane concentration in parts-per-million (ppm) per-pixel on the ground. Pixel-regions with no methane presence are set to zero. + +Segmentation mask provided in the format of a "png" image file with three channels and of the same spatial dimension as the corresponding flight line ( $\sim 23k \times \sim 1.5k \times 3$ ). The segmentation mask is obtained from the concentration mask file by setting all pixel values above zero to represent methane plumes. We manually annotated Point Source and Diffused Source based on the type of ground terrain and concentration of methane gas. Following the benchmark dataset [39], three channels are used to color code Point Source (Red) and Diffused Source (Green). The distinction of Point Source and Diffused Source is derived from the JPL-CH4-detection-V1.0 benchmark dataset [39]. Our annotation style is also consistent with the JPL-CH4-detection-V1.0 benchmark dataset [39], so that both datasets can be merged seamlessly. + +# 4.1.1 Constructing Concentration map + +Concentration maps are generated by mapping expert-annotated methane-plume concentration maps to the orthocorrected AVIRIS-NG flightlines. These methane plume annotations are systematically collected from a nonprofit [3] entity. They provide concentration masks of methane emissions in $150 \times 150$ size patches along with location information from different sources (airborne sensors [6], satellites [34]). In order to map these patches from different sources to the AVIRIS-NG flight-lines, we use the pixel coordinate locations provided for both the annotations and flight-lines. We use this information to create a homography transformation to map each pixel to its corresponding location in the flight-line. Fig. 3 shows a sample of varying types of terrains with $\mathrm{CH}_4$ segmentation mask in red and concentration mask in black circle. Details about matching the resolution, ortho-correction, and transformation are discussed in supplementary materials. The patch annotations are verified by experts visiting the physical location of emission the same day [2]. Most of the regions in California are verified by physical visits by California Air Resource Board [2,3]. + +# 4.2. MHS Statistics + +MHS statistics and properties are summarized in Table 1. + +Annotations: MHS provides both segmentation masks and concentration maps which enable development of deep learning algorithms than can produce both $\mathrm{CH}_4$ plume location and concentration predictions. + +Diversity: MHS dataset includes AVARIS-NG flightlines spanning 8 years (2015 - 2022) from six states in the U.S.: California, Nevada, New Mexico, Colorado, Texas, and Virginia. + +Data Split: We divide MHS dataset into train/test splits of $80 - 20\%$ with overlapping time periods and locations. Our dataset covers 6 states. Each state has sub-regions/locations + +(e.g. Permian basin) that are covered by multiple nonoverlapping flightlines ( $25k \times 1.5k \times 432$ pixels). These flightlines are split into train and test sets. In each set, we create patches ( $256 \times 256 \times 432$ pixels) from the corresponding flightlines. From the patches/tiles, we take all positives patches (methane ( $\mathrm{CH}_4$ ) and randomly sample equal number of negative ( $\mathrm{no - CH_4}$ ) patches. This is done for both train and test sets separately to balance the data and we refer to Section 6.2 for detailed ablation studies. + +# 5. Experimental settings + +Evaluation Metrics: Following the evaluation protocol of H-mrcnn [29] we report our performance in mean intersection-over-union (mIOU). Here, mIOU indicates the overlap between the predicted and the ground truth $\mathrm{CH}_4$ plume masks. ED represents the accuracy in plume core prediction. Additionally, as first stage of our two stage training procedure contains bounding box prediction, we also report our performance in predicting plume bounding boxes in terms of mean Average Precision (mAP) which tells us the effectiveness of MethaneMapper in eliminating the false positives in plume prediction. + +Data Pre-Processing: Each input hyperspectral image is approximately of size $25000 \times 1500 \times 432$ taking up memory space of $55 - 60\mathrm{GB}$ . We create tiles of each image in spatial domain, each tile is of size $256 \times 256 \times 432$ [29] with an overlap of 128. The $\mathrm{CH}_4$ plume is available in very few pixels in the whole image, $90\%$ of the tiles are negative samples (no methane, just ground terrain). We can not use the whole hyperspectral image because of GPU memory limitations + +Implementation Details: The band-selectors module takes 432-channels hyperspectral image as input, the RGB band selector picks 60 channel from $400nm - 700nm$ wavelength range and creates a 3-channel RGB image, the SWIR band selector picks 100 channel from wavelength range $2000nm - 2500nm$ . These input images are passed to two ResNet-50 [16] feature extractor backbones. The backbone networks are initialized with DETR [4] trained on COCO dataset [30] and input layer initialized randomly [15]. The transformer encoder-decoder and our query refiner have 6 layers and 8 heads. We initialized the transformer encoder-decoder with weights extracted and stripped from DETR [4] model. The dimension of transformer architecture is 256 and number of queries is 100. The SFG module takes in all 432-channels hyperspectral image and generates 1-channel output map of same spatial dimension as input. The feature extractor in SFG is ResNet-50 [16] initialized with DETR [4] trained on COCO dataset [30]. The decoder output embeddings are of size 512. The feature pyramid network in mask prediction module has 3 layers. More details are mentioned in supplementary materials. + +
MethodsBack boneSFG F.Ext.#paramsmAPmIOU
JPL-CH4-detection-v1.0 Dataset
1Hu et. alR-50-75M0.260.48
2H-mrcnnR-50-353M0.530.86
3MMR-50R-5080M0.630.91
MHS (Ours) Dataset
4SpectralFormerR-50-84M0.330.41
5UPSNet (stuff)R-50-69M0.320.38
6UPSNet (stuff + things)R-50-69M0.290.35
7DETRR-18*33M0.370.56
8DETRR-50*59M0.440.59
10R-18Linear Layer39M0.450.60
11MMR-18R-1844M0.520.63
12R-50R-5080M0.590.68
+ +Table 2. Comparison with baselines. “-” represent Not Applicable and “*” represent no SFG module and a random query used for transformer decoder. The top section shows performance on JPL-CH $_4$ dataset [39]. MethaneMapper achieves better results than heavily tuned H-mrcnn with $\sim 5\times$ fewer parameters. The overall detection accuracy is higher on this dataset because the type of ground terrain is uniform across all flightlines. In MHS dataset, MM outperforms multiple baselines as shown in rows 4-12. MM accuracy is lower in MHS than JPL-CH $_4$ dataset because MHS dataset has more variety of ground terrain spreading over 6 states + +
MethodsmAPmIOU
LogReg [5]-0.05
SVM [36]-0.29
PCA + LogReg-0.06
PCA + SVM-0.31
MM (R-50)0.630.91
+ +Table 3. Comparison with classical machine learning methods. " - " represent Not Available. The classical ML methods are not suited for the ${\mathrm{{CH}}}_{4}$ detection task. MethaneMapper outperforms all methods on JPL dataset [39] + +# 6. Results + +In this section we will discuss and validate all the design choices for MethaneMapper (MM) with ablations. We show that MM achieves state-of-the-art results in overall performance compared all other methods shown in Tables 2 & 3. + +# 6.1. Performance comparison + +Deep Learning methods: We trained MM with ResNet-50 [16] backbone on the same dataset that H-mrcnn [29] (JPL-CH4-detection-V1.0 [39]) was trained on for fair comparison. To align with H-mrcnn we used the same split and input image size. The MM model with 80M parameters trained for 250 epochs outperforms by significant margin the H-mrcnn model with 352M parameters. Results are summarized in Table 2 that includes the performance of MM on the new larger MHS dataset. We note that though the code for H-mrcnn is available, many of the modules are deprecated and can not be reproduced. The 'Backbone' col + +unn represents backbones used for feature extraction from input image,'SFG F.Ext.' represents the feature extractor in SFG module in MethaneMapper. We observed (qualitatively) that H-mrcnn fails to detect small $\mathrm{CH}_4$ plumes with concentration lower than $100\mathrm{kg / hr}$ while MM detects those. + +We did evaluation by implementing 3 baseline models [4, 17, 45] shown rows 4-8 of Table 2. These methods were not designed for $\mathrm{CH}_4$ detection task, therefore we needed to modify their input channel size. The poor performance of these methods may be attributed to the weak signal of interest in a high dimensional data, high number of confusers, and limited annotated data. Additionally, the only hyperspectral baseline method SpectralFormer [17] has low efficiency due its pixel-wise training scheme. + +Classical ML methods: We trained and tested multiple existing machine learning based approaches that are used for methane detection, performance shown in Table 3. Logistic regression (LogReg) [5] and multinomial logistic regression (MLR) [23] failed to produce any meaningful detection with $90\%$ false positive detections. We also trained a Support Vector Machine (SVM) [36,37] based classifier, it performed slightly better than LR and MLR methods with an IOU of $21\%$ . SVMs are prone to false positives detections same as Gaussian Mixture Models [36]. We observed that all traditional methods are not suited for the task of $\mathrm{CH}_4$ detection. We also tested reducing the dimension using principal component analysis (PCA) or just taking bands which shows maximum $\mathrm{CH}_4$ absorption. In the later case, the traditional methods performed better than using all 432 bands, this backs our idea of just using bands from SWIR region. + +Qualitative results. Fig. 5 shows comparison of MM's mask and bounding box prediction with ground truth mask on different ground terrains. The Leakages are from different type of sources such as, oil refinery, pipeline and storage tank. MM makes correct predictions in varying scenarios. + +# 6.2. Ablation Studies + +We did the experiments for ablation on MHS dataset with ResNet-50 as backbone and validate the design choices. One parameter is changed for each ablation and others kept at best settings. More ablations in Supplementary. + +Spectral Feature Generator Module: In Table 2 lower section, we show the effectiveness of our SFG module for the query refiner block. Our baseline is standard implementation of DETR [4] for segmentation task represented by row-1 and row-2 of Tab. 2 lower section. Using $\mathrm{CH}_4$ candidates feature from SFG improves the bounding box detection performance by $0.14\mathrm{mAP}$ and mask prediction by $0.09\mathrm{mIOU}$ . This demonstrate that guiding queries with $\mathrm{CH}_4$ candidates feature generated by SFG produces better embeddings as compared to random queries. + +Along with this, we explored the provision of $\mathrm{CH}_4$ candidates feature at 2 places, (i) at input level concatenating + +![](images/7e3a6acfd3326c9d821617f2496dffeb73040da5c91197ffc6bac4229893a22b.jpg) +Figure 4. Comparison of SLF with traditional filter in SFG module. White pixels represent methane and black no-methane. Red boundary represents ground-truth plume mask. SLF module generates better $\mathrm{CH}_4$ candidates + +it with $f_{comb}$ ; and (ii) as input to query refiner. We see an improvement of 0.09 mAP and 0.08 mIOU when SFG module output is passed to query refiner. We hypothesize that this is because on concatenating with input features, the CH $_4$ candidates feature information gets lost, while as cross-attention with queries reduces the search space for decoder and generate better embeddings. + +We also experimented with different types of feature extractors for SFG module, and observed that a Resnet18 or Resnet50 [16] is more effective than a 2 linear layer feature extractor as shown in Table 2. + +Spectral Linear Filter: We experimented with SLF for computing covariance $(Cov)$ using different subsets of columns in the input hyperspectral image. We observed that the SLF is most effective when covariance is computed class-wise based on land cover. Class-wise $Cov$ ensures that the radiance absorption by ground terrain is same for all the pixels while computing $\mathrm{CH}_4$ enhancement. As can be seen in Fig. 4, SLF amplifies $\mathrm{CH}_4$ candidate detection and reduces false positives. SLF leads to a $0.03\mathrm{mAP}$ improved in detection compared to traditional filters. The prediction from MM is shown row-1 of Fig. 5. + +Geographic generalization: To assess the geographical generalization capabilities of MM, we trained it on MHS data from all states except California and tested it on flightlines from California. We observed a slight drop of 0.04 mAP in detections. However, when trained on all data except Virginia, we noticed a significant drop of $0.09\mathrm{mAP}$ in detections. We attribute this to the fact that the land cover in Virginia is dense and moist vegetation, has a lower solar reflectance compared to the arid regions of California, Texas, and Nevada. + +Temporal generalization: Testing MM on 2015 after training on data from 2016-2022 showed no performance drop. + +Unbalanced test set: MM's performance dropped by 0.05 + +![](images/f1a144124c273a7dfdba175c468825c0bd032af4ff156e04861e464c5dd07c01.jpg) + +![](images/5b28eeb09c44832b36a8595c354a7ecc2adc7c44c12b1b03c02745bdfe0b4660.jpg) + +![](images/ea65ccdbb0634334d83662aa8a54192394c8358967e1abafa58806ad2476bcc0.jpg) +Figure 5. Sample ground truths and predictions on MHS dataset. We show robustness of MethaneMapper predictions on different kind of ground terrain, rows 1 and 3 shows leakage at a refinery, row 2 shows leakage from pipeline in agricultural land, row 4 shows leakage from storage tank with concrete background. + +![](images/5eb15fe50361ce53547376d135c4c5a870c1277cfa21d169e86dbdbe62478522.jpg) + +![](images/016440e284b5f226b4af73e29fba63d590340f4cf634ea27a3a6f49c2749ac93.jpg) + +![](images/963cd45d6c3f7a64a247ca55e350dcfa086d013c4c8d3b5309aca1d649ee5a53.jpg) + +mAP on an unbalanced test set with only $10\%$ positive samples $\mathrm{(CH_4)}$ and $90\%$ negative samples (no- $\mathrm{CH}_4$ ). This highlights the challenges in $\mathrm{CH}_4$ detection. Future work will address this issue. + +# 7. Conclusion + +This paper presents MethaneMapper - a hyperspectral Transformer for methane plume detection. It utilize spectral and spatial correlations using a spectral feature generator and a query refiner, to accurately delineate the $\mathrm{CH}_4$ plumes. Additionally, we curated a large-scale dataset for the task, a first of its kind, which will be made available to all researchers. The proposed MethaneMapper significantly improves upon the current methods in terms of detection and localization accuracy, as our extensive experiments demonstrate. Future work will extend the model to global monitoring [27] using multispectral satellite imaging data. + +# 8. Acknowledgments + +This research is partially supported by the following grants: NSF award SI2-SSI #1664172 and the US Army Research Laboratory (ARL) under agreement number W911NF2020157. We also thank Aditya Ramakrishnan in helping with the initial data cleaning process. + +# References + +[1] Alana K. Ayasse, Andrew K. Thorpe, Dar A. Roberts, Christopher C. Funk, Philip E. Dennison, Christian Frankenberg, Andrea Steffke, and Andrew D. Aubrey. Evaluating the effects of surface properties on methane retrievals using a synthetic airborne visible/infrared imaging spectrometer next generation (aviris-ng) image. Remote Sensing of Environment, 215:386-397, 2018. 1 +[2] CALIFORNIA AIR RESOURCE BOARD. Green house gas inventory by california air resource board, 2022. 6 +[3] INC CARBON MAPPER. Carbon mapper methane emission exploration, 2022. 2, 6 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3, 4, 5, 6, 7 +[5] Qi Cheng, Pramod K Varshney, and Manoj K Arora. Logistic regression for feature selection and soft classification of remote sensing data. IEEE Geoscience and Remote Sensing Letters, 3(4):491-494, 2006. 7 +[6] E. Knapp David, Heckler Joseph, Seely Megs, and P. Asner Gregory. Global airborne observatory visible to infrared imaging spectrometer report, 2020. 2, 6 +[7] C Frankenberg, U Platt, and T Wagner. Iterative maximum a posteriori (imap)-doas for retrieval of strongly absorbing trace gases: Model studies for ch 4 and co 2 retrieval from near infrared spectra of sciamachy onboard envisat. Atmospheric Chemistry and Physics Discussions, 4(5):6067-6106, 2004. 2 +[8] Christian Frankenberg, U Platt, and T Wagner. Iterative maximum a posteriori (imap)-doas for retrieval of strongly absorbing trace gases: Model studies for ch 4 and co 2 retrieval from near infrared spectra of sciamachy onboard envisat. Atmospheric Chemistry and Physics, 5(1):9-22, 2005. 1, 2 +[9] Christian Frankenberg, Andrew K Thorpe, David R Thompson, Glynn Hulley, Eric Adam Kort, Nick Vance, Jakob Borchardt, Thomas Krings, Konstantin Gerilowski, Colm Sweeney, et al. Airborne methane remote measurements reveal heavy-tail flux distribution in four corners region. Proceedings of the national academy of sciences, 113(35):9734-9739, 2016. 1, 2 +[10] Christopher C Funk, James Theiler, Dar A Roberts, and Christoph C Borel. Clustering to improve matched filter detection of weak gas plumes in hyperspectral thermal imagery. IEEE transactions on geoscience and remote sensing, 39(7):1410-1420, 2001. 2 +[11] Bo-Cai Gao. Ndwi—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote sensing of environment, 58(3):257-266, 1996. 4 +[12] Utsav B Gewali, Sildomar T Monteiro, and Eli Saber. Machine learning based hyperspectral image analysis: a survey. arXiv preprint arXiv:1802.08701, 2018. 2 +[13] IE Gordon, LS Rothman, RJ Hargreaves, R Hashemi, EV Karlovets, FM Skinner, EK Conway, C Hill, RV Kochanov, Y Tan, et al. The hitran2020 molecular spectroscopic + +database. Journal of quantitative spectroscopy and radiative transfer, 277:107949, 2022. 4 +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4, 6, 7, 8 +[17] Danfeng Hong, Zhu Han, Jing Yao, Lianru Gao, Bing Zhang, Antonio Plaza, and Jocelyn Chanussot. Spectralformer: Rethinking hyperspectral image classification with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-15, 2021. 2, 7 +[18] Paris IEA. Methane from oil & gas, 2020. 1 +[19] ASM Iftekhar, Hao Chen, Kaustav Kundu, Xinyu Li, Joseph Tighe, and Davide Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5353-5363, 2022. 3, 4 +[20] ASM Iftekhar, Satish Kumar, R Austin McEver, Suya You, and BS Manjunath. Gtnet: Guided transformer network for detecting human-object interactions. arXiv preprint arXiv:2108.00596, 2021. 3, 4 +[21] California Institute of Technology Jet Propulsion Laboratory. Airborne visible infrared imaging spectrometer - next generation (aviris-ng) overview, 2009. 1, 2, 5 +[22] Siraput Jongaramrungruang, Andrew K Thorpe, Georgios Matheou, and Christian Frankenberg. Methanet—an ai-driven approach to quantifying methane point-source emission from high-resolution 2-d plume imagery. Remote Sensing of Environment, 269:112809, 2022. 1, 2 +[23] Mahdi Khodadadzadeh, Jun Li, Antonio Plaza, and José M Bioucas-Dias. A subspace-based multinomial logistic regression for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 11(12):2105-2109, 2014. 7 +[24] Stefanie Kirschke, Philippe Bousquet, Philippe Ciais, Marielle Saunois, Josep G Canadell, Edward J Dlugokencky, Peter Bergamaschi, Daniel Bergmann, Donald R Blake, Lori Bruhwiler, et al. Three decades of global methane sources and sinks. Nature geoscience, 6(10):813-823, 2013. 1 +[25] Satish Kumar, ASM Iftekhar, Michael Goebel, Tom Bullock, Mary H MacLean, Michael B Miller, Tyler Santander, Barry Giesbrecht, Scott T Grafton, and BS Manjunath. Stressnet: detecting stress in thermal videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 999-1009, 2021. 3 +[26] Satish Kumar, ASM Iftekhar, Ekta Prashnani, and BS Manjunath. Locl: Learning object-attribute composition using localization. arXiv preprint arXiv:2210.03780, 2022. 3 + +[27] Satish Kumar, William Kingwill, Orbio Earth, Rozanne Mouton, Wojciech Adamczyk, Robert Huppertz, and Evan Sherwin. Guided transformer network for detecting methane emissions in sentinel-2 satellite imagery. 8 +[28] Satish Kumar, Rui Kou, Henry Hill, Jake Lempges, Eric Qian, and Vikram Jayaram. In-situ water quality monitoring in oil and gas operations. arXiv preprint arXiv:2301.08800, 2023. 2 +[29] Satish Kumar, Carlos Torres, Oytun Ulutan, Alana Ayasse, Dar Roberts, and BS Manjunath. Deep remote sensing methods for methane detection in overhead hyperspectral imagery. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1776-1785, 2020. 1, 2, 6, 7 +[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6 +[31] META-STANFORD. Methane emissions technology alliance (meta), stanford natural gas initiative, 2022. 2 +[32] MGISGeography. Ndvi (normalized difference vegetation index), 1979. 4 +[33] Gunnar Myhre, Drew Shindell, and Julia Pongratz. Anthropogenic and natural radiative forcing. 2014. 1 +[34] Darius Phiri, Matamyo Simwanda, Serajis Salekin, Vincent R Nyirenda, Yuji Murayama, and Manjula Ranagalage. Sentinel-2 data for land cover/use mapping: a review. Remote Sensing, 12(14):2291, 2020. 2, 6 +[35] Dar A Roberts, Eliza S Bradley, Ross Cheung, Ira Leifer, Philip E Dennison, and Jack S Margolis. Mapping methane emissions from a marine geological seep source using imaging spectrometry. Remote Sensing of Environment, 114(3):592-606, 2010. 1, 2 +[36] CA Shah, PK Varshney, and MK Arora. Ica mixture model algorithm for unsupervised classification of remote sensing imagery. International Journal of Remote Sensing, 28(8):1711-1731, 2007. 2, 7 +[37] David MJ Tax and Robert PW Duin. Support vector data description. Machine learning, 54(1):45-66, 2004. 2, 7 +[38] DR Thompson, I Leifer, H Bovensmann, M Eastwood, M Fladeland, C Frankenberg, K Gerilowski, RO Green, S Kratwurst, T Krings, et al. Real-time remote detection and measurement for airborne imaging spectroscopy: a case study with methane. Atmospheric Measurement Techniques, 8(10):4383-4397, 2015. 1, 2 +[39] David R Thompson, Anuj Karpatne, Imme Ebert-Uphoff, Christian Frankenberg, Andrew K Thorpe, Brian D Bue, and Robert O Green. Isgeo dataset jpl-ch4-detection-2017-v1. 0: A benchmark for methane source detection from imaging spectrometer data. 2017. 2, 5, 6, 7 +[40] Andrew K Thorpe, Christian Frankenberg, David R Thompson, Riley M Duren, Andrew D Aubrey, Brian D Bue, Robert O Green, Konstantin Gerilowski, Thomas Krings, Jakob Borchardt, et al. Airborne doas retrievals of methane, carbon dioxide, and water vapor concentrations at high spatial resolution: application to aviris-ng. Atmospheric Measurement Techniques, 10(10):3833-3850, 2017. 1, 2, 4 + +[41] Andrew K Thorpe, Dar A Roberts, Eliza S Bradley, Christopher C Funk, Philip E Dennison, and Ira Leifer. High resolution mapping of methane emissions from marine and terrestrial sources using a cluster-tuned matched filter technique and imaging spectrometry. Remote Sensing of Environment, 134:305-318, 2013. 1, 2, 4 +[42] George Turin. An introduction to matched filters. IRE transactions on Information theory, 6(3):311-329, 1960. 4 +[43] Oytun Ulutan, ASM Iftekhar, and Bangalore S Manjunath. Vsgnet: Spatial attention network for detecting human object interactions using graph convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13617-13626, 2020. 3 +[44] Foreign Agriculture Service US Dept. of Agriculture. Normalized difference vegetation index (ndvi), 1969. 4 +[45] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8826, 2019. 7 \ No newline at end of file diff --git a/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/images.zip b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1917c98808c8e8c3b1a64aae8cf15efd53185db9 --- /dev/null +++ b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23f41d7639947d14967d8b408a54e95ef48179c7c1cede036dec6992c5aa8326 +size 485195 diff --git a/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/layout.json b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..19b942e1f14a0e945f8f327d4c3fe3c9836c19a4 --- /dev/null +++ b/2023/MethaneMapper_ Spectral Absorption Aware Hyperspectral Transformer for Methane Detection/layout.json @@ -0,0 +1,9653 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "text", + "content": "MethaneMapper: Spectral Absorption aware Hyperspectral Transformer for Methane Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 162, + 173, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 162, + 173, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 173, + 175 + ], + "type": "text", + "content": "Satish Kumar" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 178, + 194, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 178, + 194, + 189 + ], + "spans": [ + { + "bbox": [ + 84, + 178, + 194, + 189 + ], + "type": "text", + "content": "satishkumar@ucsb.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 162, + 286, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 162, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 222, + 162, + 286, + 175 + ], + "type": "text", + "content": "Ivan Arevalo" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 221, + 178, + 287, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 178, + 287, + 188 + ], + "spans": [ + { + "bbox": [ + 221, + 178, + 287, + 188 + ], + "type": "text", + "content": "ifa@ucsb.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 162, + 396, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 162, + 396, + 175 + ], + "spans": [ + { + "bbox": [ + 326, + 162, + 396, + 175 + ], + "type": "text", + "content": "ASM Iftekhar" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 178, + 408, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 178, + 408, + 188 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 408, + 188 + ], + "type": "text", + "content": "iftekhar@ucsb.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 435, + 162, + 509, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 435, + 162, + 509, + 175 + ], + "spans": [ + { + "bbox": [ + 435, + 162, + 509, + 175 + ], + "type": "text", + "content": "B S Manjunath" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 435, + 178, + 508, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 435, + 178, + 508, + 189 + ], + "spans": [ + { + "bbox": [ + 435, + 178, + 508, + 189 + ], + "type": "text", + "content": "manj@ucsb.edu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 170, + 195, + 424, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 195, + 424, + 224 + ], + "spans": [ + { + "bbox": [ + 170, + 195, + 424, + 224 + ], + "type": "text", + "content": "Department of Electrical and Computer Engineering \nUniversity of California Santa Barbara" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 276, + 290, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 290, + 552 + ], + "type": "text", + "content": "Methane " + }, + { + "bbox": [ + 46, + 276, + 290, + 552 + ], + "type": "inline_equation", + "content": "(\\mathsf{CH}_4)" + }, + { + "bbox": [ + 46, + 276, + 290, + 552 + ], + "type": "text", + "content": " is the chief contributor to global climate change. Recent Airborne Visible-Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) has been very useful in quantitative mapping of methane emissions. Existing methods for analyzing this data are sensitive to local terrain conditions, often require manual inspection from domain experts, prone to significant error and hence are not scalable. To address these challenges, we propose a novel end-to-end spectral absorption wavelength aware transformer network, MethaneMapper, to detect and quantify the emissions. MethaneMapper introduces two novel modules that help to locate the most relevant methane plume regions in the spectral domain and uses them to localize these accurately. Thorough evaluation shows that MethaneMapper achieves 0.63 mAP in detection and reduces the model size (by " + }, + { + "bbox": [ + 46, + 276, + 290, + 552 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 46, + 276, + 290, + 552 + ], + "type": "text", + "content": ") compared to the current state of the art. In addition, we also introduce a large-scale dataset of methane plume segmentation mask for over 1200 AVIRIS-NG flight lines from 2015-2022. It contains over 4000 methane plume sites. Our dataset will provide researchers the opportunity to develop and advance new methods for tackling this challenging green-house gas detection problem with significant broader social impact. Dataset and source code link1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": "We consider the problem of detecting and localizing methane " + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\left(\\mathrm{CH}_4\\right)" + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": " plumes from hyperspectral imaging data. Detecting and localizing potential " + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": " hot spots is a necessary first step in combating global warming due to greenhouse gas emissions. Methane gas is estimated to contribute " + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": " of global warming induced by greenhouse gasses [24] with a Global Warming Potential (GWP) 86 times higher than carbon dioxide " + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "(\\mathrm{CO}_2)" + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": " in a 20 year period [33]. To" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "text", + "content": "put into perspective, the amount of environmental damage that " + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "text", + "content": " can do in 100 years, " + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "text", + "content": " can do in 1.2 years. Hence it is critical to monitor and curb the " + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "text", + "content": " emissions. While " + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "text", + "content": " emission has many sources, of particular interest are those from oil and natural gas industries. According to the United States Environmental Protection Agency report, " + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "text", + "content": " emissions from these industries accounts to 84 million tons per year [18]. These " + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 251, + 547, + 373 + ], + "type": "text", + "content": " emissions emanate from specific locations, mainly from pipeline leakages, storage tank leak or leakage from oil extraction point." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 375, + 548, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 548, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 548, + 639 + ], + "type": "text", + "content": "Current efforts to detect these sources mostly depend on aerial imagery. The Jet Propulsion Laboratory (JPL) has conducted thousands of aerial surveys in the last decade to collect data using an airborne sensor AVIRIS-NG [21]. Several methods have been proposed to detect potential emission sites from such imagery, for example, see [8,9,35,38, 40, 41]. However, these methods are in general very sensitive to background context and land-cover types, resulting in a large number of false positives that often require significant domain expert time to correct the detections. The primary reason is that these pixel-based methods are solely dependent on spectral correlations for detection. Spatial information can be very effective in reducing these false positives as " + }, + { + "bbox": [ + 304, + 375, + 548, + 639 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 375, + 548, + 639 + ], + "type": "text", + "content": " plumes exhibit a plume-like structure morphology. There has been recent efforts in utilizing spatial correlation using deep learning methods [22, 29], however, these works don't leverage spectral properties to filter out confusers. For example, methane has similar spectral properties as white-painted commercial roofs or paved surfaces such as airport asphalts [1]. This paper presents a novel deep-network based solution to minimize the effects of such confusers in accurately localizing methane plumes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Our proposed approach, referred to as the MethaneMapper (MM), adapts the DETR [4], a transformer model that combines the spectral and spatial correlations in the imaging data to generate a map of potential methane " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{(CH_4)}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " plume candidates. These candidates reduce the search space for a hyperspectral decoder to detect " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " plumes and re" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "type": "text", + "content": "1https://github.com/UCSB-VRL/MethaneMapper-Spectral-Absorption-aware-Hyperspectral-Transformer-for-Methane-Detection" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17609" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "move potential confusers. MM is a light-weight end-to-end single-stage " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " detector and introduces two novel modules: a Spectral Feature Generator and a Query Refiner. The former generates spectral features from a linear filter that maximizes the " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "-to-noise ratio in the presence of additive background noise, while the latter integrates these features for decoding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 289, + 312 + ], + "type": "text", + "content": "A major bottle neck for development of " + }, + { + "bbox": [ + 46, + 156, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 156, + 289, + 312 + ], + "type": "text", + "content": " detection methods is the limited availability of public training data. To address this, another significant contribution of this research is the introduction of a new Methane Hot Spots (MHS) dataset, largest of its kind available for computer vision researchers. MHS is curated by systematically collecting information from different publicly available datasets (airborne sensor [6], Non-profits [3,31] and satellites [34]) and generating the annotations as described in Section 4.1.1. This curated dataset contains methane segmentation masks for over 1200 AVIRIS-NG flight lines from years 2015 to 2022. Each flight line contains anywhere from " + }, + { + "bbox": [ + 46, + 156, + 289, + 312 + ], + "type": "inline_equation", + "content": "3 - 4\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 156, + 289, + 312 + ], + "type": "text", + "content": " plume sites for a total of 4000 in the MHS dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 312, + 258, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 312, + 258, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 258, + 323 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 329, + 287, + 522 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 329, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 287, + 426 + ], + "type": "text", + "content": "1. We introduce a novel single-stage end-to-end approach for methane plume detection using a hyperspectral transformer. The two modules, Spectral Feature Generator and Query Refiner, work together to improve upon the traditional transformer design and enable localization of potential methane hot spots in the hyperspectral images using a Spectral-Aware Linear Filter and refine the query representation for better decoding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 431, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 431, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 54, + 431, + 287, + 480 + ], + "type": "text", + "content": "2. A new Spectral Linear Filter (SLF) improves upon traditional linear filters by strategically picking correlated pixels in spectral domain to better whiten background distribution and amplify methane signal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 487, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 287, + 522 + ], + "type": "text", + "content": "3. A new benchmark dataset, MHS, provides the largest " + }, + { + "bbox": [ + 54, + 487, + 287, + 522 + ], + "type": "inline_equation", + "content": "(\\sim 35\\times)" + }, + { + "bbox": [ + 54, + 487, + 287, + 522 + ], + "type": "text", + "content": " publicly available dataset of annotated AVIRIS-NG flight lines from years 2015-2022." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 537, + 138, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 138, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 138, + 551 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 605 + ], + "type": "text", + "content": "Our work is at the intersection of hyperspectral data for " + }, + { + "bbox": [ + 46, + 558, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 558, + 287, + 605 + ], + "type": "text", + "content": " detection, deterministic linear filtering methods for spectral features and encoder-decoder based transformer. A review of the pertinent related works is given below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "There are several recent papers on detecting methane plumes from the airborne imaging spectrometer AVIRIS-NG [21]. This includes the Iterative Maximum a Posterior Differential Optical Absorption Spectroscopy algorithm (IMAP-DOAS) [7, 8] and matched filters [9, 35, 38, 40, 41]. IMAP-DOAS requires data from 2 hyperspectral sensors, one airborne and another on ground, hence not very practical for most application scenarios. Matched-filter based methods use background statistics to normalize" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 541, + 209 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 541, + 209 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 541, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 541, + 209 + ], + "type": "image", + "image_path": "0c642682fdf22bacd4de13a996f1a9c7a413a775a8368849b65ce61b639bdef5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 214, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 214, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 214, + 547, + 258 + ], + "type": "text", + "content": "Figure 1. Depiction of data collection process. Each flightline is " + }, + { + "bbox": [ + 305, + 214, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\sim 300" + }, + { + "bbox": [ + 305, + 214, + 547, + 258 + ], + "type": "text", + "content": " km long. An array of 598 sensors records data at " + }, + { + "bbox": [ + 305, + 214, + 547, + 258 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 305, + 214, + 547, + 258 + ], + "type": "text", + "content": " pixel spatial resolution. All flightlines are ortho-corrected. Each data-cube is of dimension " + }, + { + "bbox": [ + 305, + 214, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\sim 25000\\times \\sim 1500\\times 432" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "type": "text", + "content": "the spectral signals and match with the " + }, + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "type": "text", + "content": " spectral signature at every spatial location (pixel-wise). This process, however, is sensitive to surface albedo and land cover with spectral absorption similar to " + }, + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "type": "text", + "content": ", leading to spurious detections. Domain experts must then manually inspect each flight line to identify and delineate real " + }, + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 287, + 547, + 443 + ], + "type": "text", + "content": " plumes [39]. To suppress the effect of false positives due to variability of elements on ground, Christopher et.al. [10] and Thorbe et. al. [41] introduced cluster-tuned matched filter. It involves clustering the pixels with similar spectral properties using k-means clustering. Both IMAP-DOAS and all versions of matched filters are heavily prone to false positives as the information is processed pixel-wise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "type": "text", + "content": "Machine learning approaches have been used for target detection, including " + }, + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "type": "text", + "content": " identification, in hyperspectral imagery [12,17,28,36,37]. Similar to matched-filtering, these methods do not take into account the influence of confusers on the " + }, + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "type": "text", + "content": " spectral signature and have similar issues concerning false positives. Recently introduced deep learning based H-mrcnn model [29] focus on capturing spatial correlation. H-mrcnn is an ensemble of mask-rcnncite networks processing blocks of hyperspectral data. This block processing in the spectral domain is inefficient and often results in overall poor performance. Methanet [22] is a more recent work focusing on estimating methane concentration from matched-filter data. In this regard, our proposed MethaneMapper uses both spectral and spatial correlation to accurately delineates " + }, + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 446, + 547, + 626 + ], + "type": "text", + "content": " plumes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Datasets: The only dataset publicly available with annotation for " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " plume detection is JPL- " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "-detection2017-V1.0 dataset [39]. It contains only 46 AVIRIS-NG [21] flight lines in the US Four-Corners region. Deep learning architectures require a large number of annotated samples, and for this reason we introduce the new MHS dataset with over 1200 annotated flightlines and " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\sim 4000" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " plume sites." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17610" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 558, + 198 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 558, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 558, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 558, + 198 + ], + "type": "image", + "image_path": "7da038c7b5b519847a23b51c3895625349fa028d309a2347e15dc08561707b6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 207, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 45, + 207, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 45, + 207, + 547, + 275 + ], + "type": "text", + "content": "Figure 2. Overview of MethaneMapper (MM) architecture. Given a hyperspectral image, our RGB " + }, + { + "bbox": [ + 45, + 207, + 547, + 275 + ], + "type": "inline_equation", + "content": "(400\\mathrm{nm} - 700\\mathrm{nm})" + }, + { + "bbox": [ + 45, + 207, + 547, + 275 + ], + "type": "text", + "content": " and SWIR " + }, + { + "bbox": [ + 45, + 207, + 547, + 275 + ], + "type": "inline_equation", + "content": "(2000\\mathrm{nm} - 2500\\mathrm{nm})" + }, + { + "bbox": [ + 45, + 207, + 547, + 275 + ], + "type": "text", + "content": " band-pass filters passes a subset of channels in desired wavelength range and feed them to CNN backbones (ResNet) to extract features. These features are concatenated and fed to Transformer Encoder. Parallelly, our Spectral Feature Generator (SFG) modules take in all channels of input image and generate methane candidates features. Next these candidates are sent to Query Refiner (QR) to refine queries. Then these queries decoded using encoded feature from Transformer Encoder. Finally each decoded query is used to predict a plume mask via Mask Prediction and, bounding box and class via FFNs (Feed Forward Network)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 251, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 251, + 303 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 251, + 303 + ], + "type": "text", + "content": "3. MethaneMapper (MM) Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 308, + 141, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 141, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 141, + 320 + ], + "type": "text", + "content": "3.1. Data Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": "AVIRIS-NG hyperspectral imaging sensors capture spectral radiance values from " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_0" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_0 = 432" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": ") channels corresponding to wavelengths ranging from " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "400nm - 2500nm" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": " as shown in Fig. 1. The complete hyperspectral image is represented as " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H_0 \\times W_0 \\times N_0}" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "H_0, W_0" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": " are the height & width, respectively, and " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "N_0 = 432" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": " is number of channels. This hyperspectral data includes a very weak signature of " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": " around " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "2100 - 2400nm" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": ", conflated with radiations from the surrounding land cover and background clutter. A single flight-line could be over a couple miles long (about " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "25\\mathrm{K}" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": " pixels in one of the dimensions), with an array of sensors recording the data at " + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m/pixel}" + }, + { + "bbox": [ + 45, + 327, + 290, + 483 + ], + "type": "text", + "content": " resolution. The images are orthorectified before processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 491, + 162, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 162, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 162, + 503 + ], + "type": "text", + "content": "3.2. Technical Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 510, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 510, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 45, + 510, + 287, + 664 + ], + "type": "text", + "content": "Referring to Fig.2, MM contains the following main components: (i) 2 CNN backbones to extract a compact feature representation of the spectral regions of interest from the hyperspectral image, (ii) a Spectral Feature Generator (SFG), and (iii) a Query Refiner (QR) in between an encoder-decoder pair (inspired by GTNet [20], SSRT [19]). The hyperspectral image is first processed through two separate band-pass filters to select the channels in visible " + }, + { + "bbox": [ + 45, + 510, + 287, + 664 + ], + "type": "inline_equation", + "content": "(400 - 700nm)" + }, + { + "bbox": [ + 45, + 510, + 287, + 664 + ], + "type": "text", + "content": " and short-wave infrared (SWIR)(2000 - 2500nm) wavelength regions, and are then passed through CNN backbones. Output of these backbones are concatenated together and then encoded using a transformer encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "The SFG (Sec. 3.4) takes in all channels of the hyperspectral image and process them through a spectral linear filter. The SFG exploits the spectral correlation to generate methane candidates feature maps and passes them to QR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 289, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 373 + ], + "type": "text", + "content": "The QR (Sec. 3.5) uses these methane candidates to refine the learnable queries. Our hyperspectral decoder takes the encoded features from the encoder and refined queries from QR to generate the embeddings. The mask-prediction layer processes these embeddings along with the feature pyramid from the backbone layers to generate the final methane-plume segmentation prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 375, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 375, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 545, + 387 + ], + "type": "text", + "content": "These individual blocks are discussed in more detail below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 401, + 491, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 491, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 491, + 415 + ], + "type": "text", + "content": "3.3. Bandpass filtering for the Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": "The HSI is processed by two parallel band-pass filters; a visible wavelength " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "(400 - 700nm)" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": " (RGB) and a shortwave infrared wavelength " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "(2000 - 2500nm)" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": " (SWIR) bandpass filter. The RGB filter results in a 3 channel output corresponding to the normal red, green, and blue wavelengths. The SWIR generates channels, approximately " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "5\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": " apart. The filtered outputs are " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{rgb} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{swir} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 100}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{rgb}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{swir}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": ", two conventional CNN backbones (e.g. ResNet-50 [16, 25]) generate two feature maps respectively of size " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{H \\times W \\times N}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "H = \\frac{H_0}{32}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "W = \\frac{W_0}{32}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "N = 2048" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": " typically. We concatenate these feature maps along channel dimension and project through a " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": " convolution layer to retain channel dimension of " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": ". The resulting output is " + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "inline_equation", + "content": "f_{comb} \\in \\mathbb{R}^{H \\times W \\times N}" + }, + { + "bbox": [ + 304, + 422, + 545, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": "Following the standard architecture of transformer encoder from previous works [4, 19, 20, 26, 43], we reduce the channel dimension of " + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "inline_equation", + "content": "f_{comb}" + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": " convolution to " + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "inline_equation", + "content": "f_z \\in \\mathbb{R}^{H \\times W \\times d}" + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": " and supplement position information by adding a fixed positional embedding " + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{H \\times W \\times d}" + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": ". The encoder consists of a stack of multi-head self-attention modules and feed-forward networks (FFN). The encoded feature map is " + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "inline_equation", + "content": "f_e \\in \\mathbb{R}^{H \\times W \\times d}" + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 383, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "f _ {e} = \\operatorname {E n c o d e r} \\left(f _ {z}, p\\right) \\tag {1}", + "image_path": "5b79528691ce122ec4f3b08e361105caf5e7d173f7f747c7d88c8bbdc7bfc57c.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17611" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 230, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 230, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 230, + 85 + ], + "type": "text", + "content": "3.4. Spectral Feature Generator (SFG)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "content": "In parallel, the input hyperspectral image is processed by the SFG module to generate methane candidates feature map " + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "inline_equation", + "content": "f_{mc}" + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "content": ", providing the QR module with spatial information to help the network delineate the methane plumes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "text", + "content": "The SFG consist of a spectral linear filter (SLF) and a Feature Extractor (e.g.ResNet-50 [16]). The most common linear filtering approach for detecting " + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "text", + "content": " is to take each pixel from the input hyperspectral image " + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{ij} \\mid \\mathbf{x}_{ij} \\in \\mathbb{R}^{1 \\times 1 \\times N_0}\\}_{i,j=1}^{H_0, W_0}" + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "text", + "content": " and project it onto a " + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "text", + "content": " spectral absorption signature vector of same size [13]. This is to reduce the interference from ground terrain and amplify the " + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 140, + 287, + 272 + ], + "type": "text", + "content": " visibility in that pixel. Accurately modeling SLF is critical given that it is designed to reduce ground terrain interference. To model SLF we use the most common approach to matched filtering from information theory [42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "content": "Spectral Linear Filter (SLF): The design of SLF is dependent on the spectral absorption pattern of " + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "content": " gas [13] and distribution of ground terrain. Since our signal of interest, " + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "content": ", is very weak, traditional methods of linear filtering [40, 41] are not effective. The conventional methods to whiten the ground terrain noise includes calculating the covariance " + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "inline_equation", + "content": "(\\mathbf{Cov} \\in \\mathbb{R}^{N_0 \\times N_0})" + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "content": " of background by selecting a set of 10-15 adjacent columns " + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i \\mid \\mathbf{x}_i \\in \\mathbb{R}^{1 \\times H_0 \\times N_0}\\}_{i=1}^{W_0}" + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "content": ". However, in a given flight-line, the terrain changes frequently, from water bodies to bare soil, vegetation, buildings and other urban structures. Therefore single approximation of the covariance can not provide correct estimate of " + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "content": " and a localized context-based whitening will be more effective. To address this problem, we took a very simple and effective approach of doing land cover classification and segmentation [11, 32, 44], and then compute covariance per class from the land cover. More details in supplementary materials. This improves the quality of methane candidates in presence of confusers (materials with similar spectral absorption patterns as " + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "content": ") and also in cases where " + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 289, + 287, + 553 + ], + "type": "text", + "content": " concentration is low. The final SLF design with per class covariance is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 561, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 561, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 561, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\mathbf {S L F} \\left(\\mathbf {x} _ {i j}\\right) = \\frac {\\left(\\mathbf {x} _ {i j} - \\mu_ {k}\\right) ^ {T} \\mathbf {C o v} _ {k} ^ {- 1} t}{\\sqrt {t ^ {T} \\mathbf {C o v} _ {k} ^ {- 1}} t} \\forall (i, j) \\in \\text {c l a s s} k \\tag {2}", + "image_path": "39ad7d4fa86112a043b768a654120ff8246c911712a8cc7cd13032db6a9fea9d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " represents the spectral absorption pattern [13] of " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " gas, and " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Cov}_k" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " are the covariance and mean of " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " class respectively. " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ij}" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " represents the pixel in input hyperspectral image at " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " index in " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " class. This operation generates a 2-D spatial " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " candidates map of size " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H_0\\times W_0}" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": ". Next this " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " candidates map is fed to a Feature Extractor to generate " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": " candidates feature map " + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{mc}" + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": ". Details of the land cover segmentation/classification and complete SLF derivation are in the Supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 339, + 83, + 545, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 83, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 339, + 83, + 545, + 97 + ], + "type": "interline_equation", + "content": "f _ {m c} = \\operatorname {F e a t u r e E x t r a c t o r} (\\mathbf {S L F} (\\mathbf {x} _ {i j}) \\forall i, j) \\tag {3}", + "image_path": "34827a04cc155a88988b3da5d7a9e932fc425b5ffbc3141011dc163d9557b05a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 102, + 422, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 102, + 422, + 115 + ], + "spans": [ + { + "bbox": [ + 306, + 102, + 422, + 115 + ], + "type": "text", + "content": "3.5. Query Refiner (QR)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "text", + "content": "Next the methane candidate feature map " + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "inline_equation", + "content": "f_{mc} \\in \\mathbb{R}^{H \\times W \\times d}" + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "text", + "content": " is fed to the QR module along with a set of 100 learnable queries " + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{100 \\times d}" + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "inline_equation", + "content": "f_{mc}" + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "text", + "content": " refines the learnable queries via cross-attention mechanism. This operation provides a narrow search space for the queries. The QR module follows a transformer decoder-like architecture inspired from [19,20]. The randomly initialized queries " + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{100 \\times d}" + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "text", + "content": " are first passed through a self-attention layer to attend to themselves. Next, these queries attend to our methane candidates feature map " + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "inline_equation", + "content": "f_{mc}" + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "text", + "content": " from SFG module through a cross-attention layer. The methane candidates feature map serves as key-values pairs in our attention architecture. The output of QR is " + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "inline_equation", + "content": "Q_{ref}" + }, + { + "bbox": [ + 305, + 121, + 545, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 285, + 545, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 285, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 381, + 285, + 545, + 298 + ], + "type": "interline_equation", + "content": "Q _ {r e f} = \\mathbf {Q R} \\left(f _ {m c}, Q\\right) \\tag {4}", + "image_path": "a7ebcb1350adc73454fe8693958b2454c354e6024a11330731fe5f5d1d5930fd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 304, + 432, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 432, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 432, + 317 + ], + "type": "text", + "content": "3.6. Hyperspectral Decoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "Q_{ref}" + }, + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "type": "text", + "content": " is fed to the decoder module along with encoder output " + }, + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "type": "text", + "content": " to generate output embeddings. Our hyperspectral decoder follows the standard architecture with a minor difference. There are no self-attention layers, just stack of multi-headed cross attention layers. The refined queries are transformed into output embeddings " + }, + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "E_{out} \\in \\mathbb{R}^{100 \\times d}" + }, + { + "bbox": [ + 305, + 323, + 545, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 403, + 545, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 403, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 365, + 403, + 545, + 416 + ], + "type": "interline_equation", + "content": "E _ {\\text {o u t}} = \\operatorname {D e c o d e r} \\left(f _ {e}, p, Q _ {\\text {r e f}}\\right) \\tag {5}", + "image_path": "f28c63a8e3b1cd0f16d7ecaacd73385693531ca244daad4b5e361511ea9d5d53.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 422, + 448, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 448, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 448, + 434 + ], + "type": "text", + "content": "3.7. Box and Mask Prediction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "type": "text", + "content": "The decoder output embeddings " + }, + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "type": "inline_equation", + "content": "(E_{out})" + }, + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "type": "text", + "content": " are fed to two Feed Forward Network (FFNs) and a Mask prediction layer. The outputs of the FFNs are the bounding boxes covering each " + }, + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "type": "text", + "content": " plume and a confidence score corresponding to each box. The mask-prediction module follows the standard segmentation head of DETR [4]. It computes multi-head attention scores of each embedding over the " + }, + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 305, + 440, + 545, + 584 + ], + "type": "text", + "content": " (Eq. 1), generating a low-resolution heatmap for each embedding. To make the final prediction a Feature Pyramid Network [14] like structure is used. Each heatmap is designed to capture one methane plume. A simple thresholding is used to merge the heatmaps as final segmentation mask." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 349, + 592, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 592, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 349, + 592, + 545, + 605 + ], + "type": "interline_equation", + "content": "m a s k = \\operatorname {M a s k} _ {-} \\operatorname {p r e d} \\left(E _ {\\text {o u t}}, f _ {e}, f _ {\\text {c o m b}}\\right) \\tag {6}", + "image_path": "799d82914083e9a96683954f9646c8e771c0fe5058a12bef261973ce12781df7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 611, + 436, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 436, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 436, + 624 + ], + "type": "text", + "content": "3.8. Training and Inference" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": "We train MethaneMapper in two stages; first we train bounding box detection corresponding to each " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " plume, and second by freezing the box detection network and training only the mask prediction module. We also trained both box and mask prediction modules end-to-end and achieved similar performance. We use a similar two-stage loss strategy for training MethaneMapper as that used in DETR [4]:" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17612" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 272, + 355 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 272, + 355 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 272, + 355 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 272, + 355 + ], + "type": "image", + "image_path": "4656e40c8c344ccdb4e448e5e781ba0cd9fac1831c406ca2c2e680ff5c98466a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 362, + 288, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 362, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 288, + 429 + ], + "type": "text", + "content": "Figure 3. Sample images from MHS dataset. The colormap in black circle shows concentration maps corresponding to the plume mask shown in red. We are showing different types of leakage sources and land cover types. For better visualization, we plotted the binary mask on color image created using visible bands of hyperspectral image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 441, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 548 + ], + "type": "text", + "content": "first stage is the bipartite matching between the predictions and the ground truths both in bounding box and mask prediction, and then second stage is loss calculation for the matched pairs. The bipartite matching employs the Hungarian algorithm [4] to find the optimal matching between the predictions and the ground truths. After this matching, every prediction is associated with a ground truth. Next, we calculate the " + }, + { + "bbox": [ + 46, + 441, + 287, + 548 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 46, + 441, + 287, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 441, + 287, + 548 + ], + "type": "inline_equation", + "content": "GIoU" + }, + { + "bbox": [ + 46, + 441, + 287, + 548 + ], + "type": "text", + "content": " loss on both box and mask predictions and cross entropy loss for class prediction [4]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "type": "text", + "content": "Inference: The inference pipeline is similar to training pipeline and can be implemented using approximately 50 lines of code. During inference, we first filter the detections with confidence levels below " + }, + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "type": "text", + "content": " and a per-pixel max to determine which pixels are predicted to belong to a CH4 plume." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 633, + 239, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 239, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 239, + 647 + ], + "type": "text", + "content": "4. Methane Hot Spots (MHS) dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Another significant contribution of this work is a large scale curated MHS dataset. It contains the AVIRIS-NG spectral data with wavelength ranging from " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "380nm" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "2510nm" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": ", a " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "5nm" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " sampling [21], and capturing 432 channels per pixel. The images from the flight-line are orthorec" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 558, + 217 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 558, + 217 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 558, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 558, + 217 + ], + "type": "table", + "html": "
DatasetMHS (Ours) DatasetJPL-CH4 detection-V1.0 [39]
# plume sites3961161
# flightlines118546
# point source3675114
# diffused source28657
Time period2015 - 2022(8 years)2015(1 year)
Segmentation MaskYesYes
Bonding boxYesNo
Concentration mapYesNo
Number of Regions61
", + "image_path": "cb9970bd25ff7d3f5a997f06b12daef31ba86ca955f9ebbe8c96a731e90769d9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 224, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 314 + ], + "type": "text", + "content": "Table 1. Statistics shows MHS dataset comparison with JPL-CH4-detection-V1.0 [39] dataset. Each flightline have multiple large and small plume sites. Each flightline have at least 4 plume sites. The Point Source represents high concentration " + }, + { + "bbox": [ + 304, + 224, + 547, + 314 + ], + "type": "inline_equation", + "content": "(300kg/hr)" + }, + { + "bbox": [ + 304, + 224, + 547, + 314 + ], + "type": "text", + "content": " to leakage from sources like pipeline leak, storage tanks, oil and gas refineries. Diffused Source represent low concentration leakages from sources like biomass degradation in landfills. Our dataset is covers more diverse type of terrain over 6 states." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "type": "text", + "content": "tified and of size " + }, + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "type": "inline_equation", + "content": "\\sim 23K\\times \\sim 1.5K\\times 432" + }, + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "type": "text", + "content": ". The only currently publicly-available dataset with methane plume segmentation masks is the JPL-CH4-detection-V1.0 [39] dataset released by JPL-NASA in 2017. The MHS dataset has approximately 4000 plume sites corresponding to approximately 1200 AVIRIS-NG flightlines as shown in Table 1. MHS also has higher diversity data with flight lines spanning from 2015-2022 and covering terrain from 6 states– California, Nevada, New Mexico, Colorado, Midland Texas, and Virginia." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 464, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 619 + ], + "type": "text", + "content": "Data Pruning: We selected AVIRIS-NG flight lines over varying regions as it covers a wide variety of " + }, + { + "bbox": [ + 304, + 464, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 464, + 545, + 619 + ], + "type": "text", + "content": " plume sources, such as leaks in oil and gas refineries, oil and gas extraction points, natural seeps, leaking underground storage tank, coal mines, dairy farms, landfill sites, and pipeline leaks. Along with varying emission sources, we selected regions with different types of ground terrains like, bare soil, rocks, mountains, light vegetation, water bodies and dense vegetation as shown with few samples in Fig. 3. Different types of ground terrain exhibit widely varying albedo and thus have a major impact on the quality of " + }, + { + "bbox": [ + 304, + 464, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 464, + 545, + 619 + ], + "type": "text", + "content": " detections as shown in Fig. 4. Given this, training models with diverse ground terrain data leads to a more robust model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 628, + 534, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 628, + 534, + 640 + ], + "spans": [ + { + "bbox": [ + 305, + 628, + 534, + 640 + ], + "type": "text", + "content": "4.1. Concentration map and Segmentation mask" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Concentration map is provided in the form of a matrix of spatial dimensions same as the flightline " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "(\\sim 23k\\times \\sim 1.5k\\times 1)" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": ". There is one concentration map per flight-line (orthorectified). It shows methane concentration in parts-per-million (ppm) per-pixel on the ground. Pixel-regions with no methane presence are set to zero." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17613" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "Segmentation mask provided in the format of a \"png\" image file with three channels and of the same spatial dimension as the corresponding flight line (" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "\\sim 23k \\times \\sim 1.5k \\times 3" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "). The segmentation mask is obtained from the concentration mask file by setting all pixel values above zero to represent methane plumes. We manually annotated Point Source and Diffused Source based on the type of ground terrain and concentration of methane gas. Following the benchmark dataset [39], three channels are used to color code Point Source (Red) and Diffused Source (Green). The distinction of Point Source and Diffused Source is derived from the JPL-CH4-detection-V1.0 benchmark dataset [39]. Our annotation style is also consistent with the JPL-CH4-detection-V1.0 benchmark dataset [39], so that both datasets can be merged seamlessly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 271, + 222, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 222, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 222, + 283 + ], + "type": "text", + "content": "4.1.1 Constructing Concentration map" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 543 + ], + "type": "text", + "content": "Concentration maps are generated by mapping expert-annotated methane-plume concentration maps to the orthocorrected AVIRIS-NG flightlines. These methane plume annotations are systematically collected from a nonprofit [3] entity. They provide concentration masks of methane emissions in " + }, + { + "bbox": [ + 46, + 291, + 289, + 543 + ], + "type": "inline_equation", + "content": "150 \\times 150" + }, + { + "bbox": [ + 46, + 291, + 289, + 543 + ], + "type": "text", + "content": " size patches along with location information from different sources (airborne sensors [6], satellites [34]). In order to map these patches from different sources to the AVIRIS-NG flight-lines, we use the pixel coordinate locations provided for both the annotations and flight-lines. We use this information to create a homography transformation to map each pixel to its corresponding location in the flight-line. Fig. 3 shows a sample of varying types of terrains with " + }, + { + "bbox": [ + 46, + 291, + 289, + 543 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 291, + 289, + 543 + ], + "type": "text", + "content": " segmentation mask in red and concentration mask in black circle. Details about matching the resolution, ortho-correction, and transformation are discussed in supplementary materials. The patch annotations are verified by experts visiting the physical location of emission the same day [2]. Most of the regions in California are verified by physical visits by California Air Resource Board [2,3]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 552, + 140, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 140, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 140, + 563 + ], + "type": "text", + "content": "4.2. MHS Statistics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 567, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 567, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 59, + 567, + 286, + 578 + ], + "type": "text", + "content": "MHS statistics and properties are summarized in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 580, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 286, + 628 + ], + "type": "text", + "content": "Annotations: MHS provides both segmentation masks and concentration maps which enable development of deep learning algorithms than can produce both " + }, + { + "bbox": [ + 47, + 580, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 580, + 286, + 628 + ], + "type": "text", + "content": " plume location and concentration predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 628, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 286, + 676 + ], + "type": "text", + "content": "Diversity: MHS dataset includes AVARIS-NG flightlines spanning 8 years (2015 - 2022) from six states in the U.S.: California, Nevada, New Mexico, Colorado, Texas, and Virginia." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "Data Split: We divide MHS dataset into train/test splits of " + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "inline_equation", + "content": "80 - 20\\%" + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": " with overlapping time periods and locations. Our dataset covers 6 states. Each state has sub-regions/locations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "(e.g. Permian basin) that are covered by multiple nonoverlapping flightlines (" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "25k \\times 1.5k \\times 432" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " pixels). These flightlines are split into train and test sets. In each set, we create patches (" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 432" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " pixels) from the corresponding flightlines. From the patches/tiles, we take all positives patches (methane (" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ") and randomly sample equal number of negative (" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{no - CH_4}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ") patches. This is done for both train and test sets separately to balance the data and we refer to Section 6.2 for detailed ablation studies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 199, + 433, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 199, + 433, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 433, + 213 + ], + "type": "text", + "content": "5. Experimental settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 217, + 545, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 349 + ], + "type": "text", + "content": "Evaluation Metrics: Following the evaluation protocol of H-mrcnn [29] we report our performance in mean intersection-over-union (mIOU). Here, mIOU indicates the overlap between the predicted and the ground truth " + }, + { + "bbox": [ + 304, + 217, + 545, + 349 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 217, + 545, + 349 + ], + "type": "text", + "content": " plume masks. ED represents the accuracy in plume core prediction. Additionally, as first stage of our two stage training procedure contains bounding box prediction, we also report our performance in predicting plume bounding boxes in terms of mean Average Precision (mAP) which tells us the effectiveness of MethaneMapper in eliminating the false positives in plume prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "text", + "content": "Data Pre-Processing: Each input hyperspectral image is approximately of size " + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "inline_equation", + "content": "25000 \\times 1500 \\times 432" + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "text", + "content": " taking up memory space of " + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "inline_equation", + "content": "55 - 60\\mathrm{GB}" + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "text", + "content": ". We create tiles of each image in spatial domain, each tile is of size " + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 432" + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "text", + "content": " [29] with an overlap of 128. The " + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "text", + "content": " plume is available in very few pixels in the whole image, " + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 352, + 545, + 458 + ], + "type": "text", + "content": " of the tiles are negative samples (no methane, just ground terrain). We can not use the whole hyperspectral image because of GPU memory limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details: The band-selectors module takes 432-channels hyperspectral image as input, the RGB band selector picks 60 channel from " + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "inline_equation", + "content": "400nm - 700nm" + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": " wavelength range and creates a 3-channel RGB image, the SWIR band selector picks 100 channel from wavelength range " + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "inline_equation", + "content": "2000nm - 2500nm" + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": ". These input images are passed to two ResNet-50 [16] feature extractor backbones. The backbone networks are initialized with DETR [4] trained on COCO dataset [30] and input layer initialized randomly [15]. The transformer encoder-decoder and our query refiner have 6 layers and 8 heads. We initialized the transformer encoder-decoder with weights extracted and stripped from DETR [4] model. The dimension of transformer architecture is 256 and number of queries is 100. The SFG module takes in all 432-channels hyperspectral image and generates 1-channel output map of same spatial dimension as input. The feature extractor in SFG is ResNet-50 [16] initialized with DETR [4] trained on COCO dataset [30]. The decoder output embeddings are of size 512. The feature pyramid network in mask prediction module has 3 layers. More details are mentioned in supplementary materials." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17614" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 242 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 242 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 242 + ], + "type": "table", + "html": "
MethodsBack boneSFG F.Ext.#paramsmAPmIOU
JPL-CH4-detection-v1.0 Dataset
1Hu et. alR-50-75M0.260.48
2H-mrcnnR-50-353M0.530.86
3MMR-50R-5080M0.630.91
MHS (Ours) Dataset
4SpectralFormerR-50-84M0.330.41
5UPSNet (stuff)R-50-69M0.320.38
6UPSNet (stuff + things)R-50-69M0.290.35
7DETRR-18*33M0.370.56
8DETRR-50*59M0.440.59
10R-18Linear Layer39M0.450.60
11MMR-18R-1844M0.520.63
12R-50R-5080M0.590.68
", + "image_path": "52149fbf023a2cb80b0a30fe5950268d9cf30a453cca08208897783f9fd917bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 97, + 371, + 237, + 443 + ], + "blocks": [ + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "type": "text", + "content": "Table 2. Comparison with baselines. “-” represent Not Applicable and “*” represent no SFG module and a random query used for transformer decoder. The top section shows performance on JPL-CH" + }, + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "type": "inline_equation", + "content": "_4" + }, + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "type": "text", + "content": " dataset [39]. MethaneMapper achieves better results than heavily tuned H-mrcnn with " + }, + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\sim 5\\times" + }, + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "type": "text", + "content": " fewer parameters. The overall detection accuracy is higher on this dataset because the type of ground terrain is uniform across all flightlines. In MHS dataset, MM outperforms multiple baselines as shown in rows 4-12. MM accuracy is lower in MHS than JPL-CH" + }, + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "type": "inline_equation", + "content": "_4" + }, + { + "bbox": [ + 46, + 250, + 288, + 361 + ], + "type": "text", + "content": " dataset because MHS dataset has more variety of ground terrain spreading over 6 states" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 371, + 237, + 443 + ], + "lines": [ + { + "bbox": [ + 97, + 371, + 237, + 443 + ], + "spans": [ + { + "bbox": [ + 97, + 371, + 237, + 443 + ], + "type": "table", + "html": "
MethodsmAPmIOU
LogReg [5]-0.05
SVM [36]-0.29
PCA + LogReg-0.06
PCA + SVM-0.31
MM (R-50)0.630.91
", + "image_path": "f2fda5f90154d35c0c45fb96478c1bea80411dd2222522969ef0a8c75b17042e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 452, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 495 + ], + "type": "text", + "content": "Table 3. Comparison with classical machine learning methods. \" - \" represent Not Available. The classical ML methods are not suited for the " + }, + { + "bbox": [ + 46, + 452, + 287, + 495 + ], + "type": "inline_equation", + "content": "{\\mathrm{{CH}}}_{4}" + }, + { + "bbox": [ + 46, + 452, + 287, + 495 + ], + "type": "text", + "content": " detection task. MethaneMapper outperforms all methods on JPL dataset [39]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 497, + 100, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 100, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 100, + 509 + ], + "type": "text", + "content": "6. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 512, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 561 + ], + "type": "text", + "content": "In this section we will discuss and validate all the design choices for MethaneMapper (MM) with ablations. We show that MM achieves state-of-the-art results in overall performance compared all other methods shown in Tables 2 & 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 567, + 188, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 188, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 188, + 581 + ], + "type": "text", + "content": "6.1. Performance comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Deep Learning methods: We trained MM with ResNet-50 [16] backbone on the same dataset that H-mrcnn [29] (JPL-CH4-detection-V1.0 [39]) was trained on for fair comparison. To align with H-mrcnn we used the same split and input image size. The MM model with 80M parameters trained for 250 epochs outperforms by significant margin the H-mrcnn model with 352M parameters. Results are summarized in Table 2 that includes the performance of MM on the new larger MHS dataset. We note that though the code for H-mrcnn is available, many of the modules are deprecated and can not be reproduced. The 'Backbone' col" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "unn represents backbones used for feature extraction from input image,'SFG F.Ext.' represents the feature extractor in SFG module in MethaneMapper. We observed (qualitatively) that H-mrcnn fails to detect small " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " plumes with concentration lower than " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "100\\mathrm{kg / hr}" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " while MM detects those." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 239 + ], + "type": "text", + "content": "We did evaluation by implementing 3 baseline models [4, 17, 45] shown rows 4-8 of Table 2. These methods were not designed for " + }, + { + "bbox": [ + 304, + 133, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 133, + 545, + 239 + ], + "type": "text", + "content": " detection task, therefore we needed to modify their input channel size. The poor performance of these methods may be attributed to the weak signal of interest in a high dimensional data, high number of confusers, and limited annotated data. Additionally, the only hyperspectral baseline method SpectralFormer [17] has low efficiency due its pixel-wise training scheme." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "text", + "content": "Classical ML methods: We trained and tested multiple existing machine learning based approaches that are used for methane detection, performance shown in Table 3. Logistic regression (LogReg) [5] and multinomial logistic regression (MLR) [23] failed to produce any meaningful detection with " + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "text", + "content": " false positive detections. We also trained a Support Vector Machine (SVM) [36,37] based classifier, it performed slightly better than LR and MLR methods with an IOU of " + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "text", + "content": ". SVMs are prone to false positives detections same as Gaussian Mixture Models [36]. We observed that all traditional methods are not suited for the task of " + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "text", + "content": " detection. We also tested reducing the dimension using principal component analysis (PCA) or just taking bands which shows maximum " + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 240, + 546, + 430 + ], + "type": "text", + "content": " absorption. In the later case, the traditional methods performed better than using all 432 bands, this backs our idea of just using bands from SWIR region." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 431, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 491 + ], + "type": "text", + "content": "Qualitative results. Fig. 5 shows comparison of MM's mask and bounding box prediction with ground truth mask on different ground terrains. The Leakages are from different type of sources such as, oil refinery, pipeline and storage tank. MM makes correct predictions in varying scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 498, + 406, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 406, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 406, + 509 + ], + "type": "text", + "content": "6.2. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 512, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 561 + ], + "type": "text", + "content": "We did the experiments for ablation on MHS dataset with ResNet-50 as backbone and validate the design choices. One parameter is changed for each ablation and others kept at best settings. More ablations in Supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "Spectral Feature Generator Module: In Table 2 lower section, we show the effectiveness of our SFG module for the query refiner block. Our baseline is standard implementation of DETR [4] for segmentation task represented by row-1 and row-2 of Tab. 2 lower section. Using " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " candidates feature from SFG improves the bounding box detection performance by " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "0.14\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " and mask prediction by " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "0.09\\mathrm{mIOU}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": ". This demonstrate that guiding queries with " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " candidates feature generated by SFG produces better embeddings as compared to random queries." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Along with this, we explored the provision of " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " candidates feature at 2 places, (i) at input level concatenating" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17615" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 78, + 286, + 236 + ], + "blocks": [ + { + "bbox": [ + 49, + 78, + 286, + 236 + ], + "lines": [ + { + "bbox": [ + 49, + 78, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 49, + 78, + 286, + 236 + ], + "type": "image", + "image_path": "7e3a6acfd3326c9d821617f2496dffeb73040da5c91197ffc6bac4229893a22b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 238, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 238, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 287, + 281 + ], + "type": "text", + "content": "Figure 4. Comparison of SLF with traditional filter in SFG module. White pixels represent methane and black no-methane. Red boundary represents ground-truth plume mask. SLF module generates better " + }, + { + "bbox": [ + 47, + 238, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 47, + 238, + 287, + 281 + ], + "type": "text", + "content": " candidates" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 286, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 286, + 375 + ], + "type": "text", + "content": "it with " + }, + { + "bbox": [ + 46, + 291, + 286, + 375 + ], + "type": "inline_equation", + "content": "f_{comb}" + }, + { + "bbox": [ + 46, + 291, + 286, + 375 + ], + "type": "text", + "content": "; and (ii) as input to query refiner. We see an improvement of 0.09 mAP and 0.08 mIOU when SFG module output is passed to query refiner. We hypothesize that this is because on concatenating with input features, the CH" + }, + { + "bbox": [ + 46, + 291, + 286, + 375 + ], + "type": "inline_equation", + "content": "_4" + }, + { + "bbox": [ + 46, + 291, + 286, + 375 + ], + "type": "text", + "content": " candidates feature information gets lost, while as cross-attention with queries reduces the search space for decoder and generate better embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 422 + ], + "type": "text", + "content": "We also experimented with different types of feature extractors for SFG module, and observed that a Resnet18 or Resnet50 [16] is more effective than a 2 linear layer feature extractor as shown in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "content": "Spectral Linear Filter: We experimented with SLF for computing covariance " + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "inline_equation", + "content": "(Cov)" + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "content": " using different subsets of columns in the input hyperspectral image. We observed that the SLF is most effective when covariance is computed class-wise based on land cover. Class-wise " + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "inline_equation", + "content": "Cov" + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "content": " ensures that the radiance absorption by ground terrain is same for all the pixels while computing " + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "content": " enhancement. As can be seen in Fig. 4, SLF amplifies " + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "content": " candidate detection and reduces false positives. SLF leads to a " + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "inline_equation", + "content": "0.03\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 424, + 287, + 555 + ], + "type": "text", + "content": " improved in detection compared to traditional filters. The prediction from MM is shown row-1 of Fig. 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 556, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 675 + ], + "type": "text", + "content": "Geographic generalization: To assess the geographical generalization capabilities of MM, we trained it on MHS data from all states except California and tested it on flightlines from California. We observed a slight drop of 0.04 mAP in detections. However, when trained on all data except Virginia, we noticed a significant drop of " + }, + { + "bbox": [ + 46, + 556, + 287, + 675 + ], + "type": "inline_equation", + "content": "0.09\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 556, + 287, + 675 + ], + "type": "text", + "content": " in detections. We attribute this to the fact that the land cover in Virginia is dense and moist vegetation, has a lower solar reflectance compared to the arid regions of California, Texas, and Nevada." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 700 + ], + "type": "text", + "content": "Temporal generalization: Testing MM on 2015 after training on data from 2016-2022 showed no performance drop." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "Unbalanced test set: MM's performance dropped by 0.05" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 341, + 70, + 429, + 159 + ], + "blocks": [ + { + "bbox": [ + 341, + 70, + 429, + 159 + ], + "lines": [ + { + "bbox": [ + 341, + 70, + 429, + 159 + ], + "spans": [ + { + "bbox": [ + 341, + 70, + 429, + 159 + ], + "type": "image", + "image_path": "f1a144124c273a7dfdba175c468825c0bd032af4ff156e04861e464c5dd07c01.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 341, + 160, + 429, + 248 + ], + "blocks": [ + { + "bbox": [ + 341, + 160, + 429, + 248 + ], + "lines": [ + { + "bbox": [ + 341, + 160, + 429, + 248 + ], + "spans": [ + { + "bbox": [ + 341, + 160, + 429, + 248 + ], + "type": "image", + "image_path": "5b28eeb09c44832b36a8595c354a7ecc2adc7c44c12b1b03c02745bdfe0b4660.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 341, + 249, + 430, + 338 + ], + "blocks": [ + { + "bbox": [ + 341, + 249, + 430, + 338 + ], + "lines": [ + { + "bbox": [ + 341, + 249, + 430, + 338 + ], + "spans": [ + { + "bbox": [ + 341, + 249, + 430, + 338 + ], + "type": "image", + "image_path": "ea65ccdbb0634334d83662aa8a54192394c8358967e1abafa58806ad2476bcc0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 340, + 545, + 394 + ], + "lines": [ + { + "bbox": [ + 305, + 340, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 545, + 394 + ], + "type": "text", + "content": "Figure 5. Sample ground truths and predictions on MHS dataset. We show robustness of MethaneMapper predictions on different kind of ground terrain, rows 1 and 3 shows leakage at a refinery, row 2 shows leakage from pipeline in agricultural land, row 4 shows leakage from storage tank with concrete background." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 431, + 70, + 520, + 159 + ], + "blocks": [ + { + "bbox": [ + 431, + 70, + 520, + 159 + ], + "lines": [ + { + "bbox": [ + 431, + 70, + 520, + 159 + ], + "spans": [ + { + "bbox": [ + 431, + 70, + 520, + 159 + ], + "type": "image", + "image_path": "5eb15fe50361ce53547376d135c4c5a870c1277cfa21d169e86dbdbe62478522.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 431, + 160, + 520, + 248 + ], + "blocks": [ + { + "bbox": [ + 431, + 160, + 520, + 248 + ], + "lines": [ + { + "bbox": [ + 431, + 160, + 520, + 248 + ], + "spans": [ + { + "bbox": [ + 431, + 160, + 520, + 248 + ], + "type": "image", + "image_path": "016440e284b5f226b4af73e29fba63d590340f4cf634ea27a3a6f49c2749ac93.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 431, + 249, + 520, + 338 + ], + "blocks": [ + { + "bbox": [ + 431, + 249, + 520, + 338 + ], + "lines": [ + { + "bbox": [ + 431, + 249, + 520, + 338 + ], + "spans": [ + { + "bbox": [ + 431, + 249, + 520, + 338 + ], + "type": "image", + "image_path": "963cd45d6c3f7a64a247ca55e350dcfa086d013c4c8d3b5309aca1d649ee5a53.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "text", + "content": "mAP on an unbalanced test set with only " + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "text", + "content": " positive samples " + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{(CH_4)}" + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "text", + "content": " negative samples (no- " + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "text", + "content": "). This highlights the challenges in " + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 305, + 410, + 545, + 456 + ], + "type": "text", + "content": " detection. Future work will address this issue." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 469, + 378, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 378, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 378, + 481 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 487, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 619 + ], + "type": "text", + "content": "This paper presents MethaneMapper - a hyperspectral Transformer for methane plume detection. It utilize spectral and spatial correlations using a spectral feature generator and a query refiner, to accurately delineate the " + }, + { + "bbox": [ + 304, + 487, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{CH}_4" + }, + { + "bbox": [ + 304, + 487, + 545, + 619 + ], + "type": "text", + "content": " plumes. Additionally, we curated a large-scale dataset for the task, a first of its kind, which will be made available to all researchers. The proposed MethaneMapper significantly improves upon the current methods in terms of detection and localization accuracy, as our extensive experiments demonstrate. Future work will extend the model to global monitoring [27] using multispectral satellite imaging data." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 635, + 415, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 415, + 649 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 415, + 649 + ], + "type": "text", + "content": "8. Acknowledgments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "This research is partially supported by the following grants: NSF award SI2-SSI #1664172 and the US Army Research Laboratory (ARL) under agreement number W911NF2020157. We also thank Aditya Ramakrishnan in helping with the initial data cleaning process." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17616" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 168 + ], + "type": "text", + "content": "[1] Alana K. Ayasse, Andrew K. Thorpe, Dar A. Roberts, Christopher C. Funk, Philip E. Dennison, Christian Frankenberg, Andrea Steffke, and Andrew D. Aubrey. Evaluating the effects of surface properties on methane retrievals using a synthetic airborne visible/infrared imaging spectrometer next generation (aviris-ng) image. Remote Sensing of Environment, 215:386-397, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 170, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 192 + ], + "type": "text", + "content": "[2] CALIFORNIA AIR RESOURCE BOARD. Green house gas inventory by california air resource board, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 215 + ], + "type": "text", + "content": "[3] INC CARBON MAPPER. Carbon mapper methane emission exploration, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 271 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 317 + ], + "type": "text", + "content": "[5] Qi Cheng, Pramod K Varshney, and Manoj K Arora. Logistic regression for feature selection and soft classification of remote sensing data. IEEE Geoscience and Remote Sensing Letters, 3(4):491-494, 2006. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "type": "text", + "content": "[6] E. Knapp David, Heckler Joseph, Seely Megs, and P. Asner Gregory. Global airborne observatory visible to infrared imaging spectrometer report, 2020. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 353, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 418 + ], + "type": "text", + "content": "[7] C Frankenberg, U Platt, and T Wagner. Iterative maximum a posteriori (imap)-doas for retrieval of strongly absorbing trace gases: Model studies for ch 4 and co 2 retrieval from near infrared spectra of sciamachy onboard envisat. Atmospheric Chemistry and Physics Discussions, 4(5):6067-6106, 2004. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "type": "text", + "content": "[8] Christian Frankenberg, U Platt, and T Wagner. Iterative maximum a posteriori (imap)-doas for retrieval of strongly absorbing trace gases: Model studies for ch 4 and co 2 retrieval from near infrared spectra of sciamachy onboard envisat. Atmospheric Chemistry and Physics, 5(1):9-22, 2005. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 287, + 553 + ], + "type": "text", + "content": "[9] Christian Frankenberg, Andrew K Thorpe, David R Thompson, Glynn Hulley, Eric Adam Kort, Nick Vance, Jakob Borchardt, Thomas Krings, Konstantin Gerilowski, Colm Sweeney, et al. Airborne methane remote measurements reveal heavy-tail flux distribution in four corners region. Proceedings of the national academy of sciences, 113(35):9734-9739, 2016. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "text", + "content": "[10] Christopher C Funk, James Theiler, Dar A Roberts, and Christoph C Borel. Clustering to improve matched filter detection of weak gas plumes in hyperspectral thermal imagery. IEEE transactions on geoscience and remote sensing, 39(7):1410-1420, 2001. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "type": "text", + "content": "[11] Bo-Cai Gao. Ndwi—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote sensing of environment, 58(3):257-266, 1996. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[12] Utsav B Gewali, Sildomar T Monteiro, and Eli Saber. Machine learning based hyperspectral image analysis: a survey. arXiv preprint arXiv:1802.08701, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[13] IE Gordon, LS Rothman, RJ Hargreaves, R Hashemi, EV Karlovets, FM Skinner, EK Conway, C Hill, RV Kochanov, Y Tan, et al. The hitran2020 molecular spectroscopic" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "database. Journal of quantitative spectroscopy and radiative transfer, 277:107949, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 130, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 184 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "text", + "content": "[17] Danfeng Hong, Zhu Han, Jing Yao, Lianru Gao, Bing Zhang, Antonio Plaza, and Jocelyn Chanussot. Spectralformer: Rethinking hyperspectral image classification with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-15, 2021. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 487, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 487, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 487, + 298 + ], + "type": "text", + "content": "[18] Paris IEA. Methane from oil & gas, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 300, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 364 + ], + "type": "text", + "content": "[19] ASM Iftekhar, Hao Chen, Kaustav Kundu, Xinyu Li, Joseph Tighe, and Davide Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5353-5363, 2022. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[20] ASM Iftekhar, Satish Kumar, R Austin McEver, Suya You, and BS Manjunath. Gtnet: Guided transformer network for detecting human-object interactions. arXiv preprint arXiv:2108.00596, 2021. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 444 + ], + "type": "text", + "content": "[21] California Institute of Technology Jet Propulsion Laboratory. Airborne visible infrared imaging spectrometer - next generation (aviris-ng) overview, 2009. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "text", + "content": "[22] Siraput Jongaramrungruang, Andrew K Thorpe, Georgios Matheou, and Christian Frankenberg. Methanet—an ai-driven approach to quantifying methane point-source emission from high-resolution 2-d plume imagery. Remote Sensing of Environment, 269:112809, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[23] Mahdi Khodadadzadeh, Jun Li, Antonio Plaza, and José M Bioucas-Dias. A subspace-based multinomial logistic regression for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 11(12):2105-2109, 2014. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[24] Stefanie Kirschke, Philippe Bousquet, Philippe Ciais, Marielle Saunois, Josep G Canadell, Edward J Dlugokencky, Peter Bergamaschi, Daniel Bergmann, Donald R Blake, Lori Bruhwiler, et al. Three decades of global methane sources and sinks. Nature geoscience, 6(10):813-823, 2013. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "type": "text", + "content": "[25] Satish Kumar, ASM Iftekhar, Michael Goebel, Tom Bullock, Mary H MacLean, Michael B Miller, Tyler Santander, Barry Giesbrecht, Scott T Grafton, and BS Manjunath. Stressnet: detecting stress in thermal videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 999-1009, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "text", + "content": "[26] Satish Kumar, ASM Iftekhar, Ekta Prashnani, and BS Manjunath. Locl: Learning object-attribute composition using localization. arXiv preprint arXiv:2210.03780, 2022. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17617" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Satish Kumar, William Kingwill, Orbio Earth, Rozanne Mouton, Wojciech Adamczyk, Robert Huppertz, and Evan Sherwin. Guided transformer network for detecting methane emissions in sentinel-2 satellite imagery. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 159 + ], + "type": "text", + "content": "[28] Satish Kumar, Rui Kou, Henry Hill, Jake Lempges, Eric Qian, and Vikram Jayaram. In-situ water quality monitoring in oil and gas operations. arXiv preprint arXiv:2301.08800, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 225 + ], + "type": "text", + "content": "[29] Satish Kumar, Carlos Torres, Oytun Ulutan, Alana Ayasse, Dar Roberts, and BS Manjunath. Deep remote sensing methods for methane detection in overhead hyperspectral imagery. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1776-1785, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 49, + 228, + 287, + 281 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 284, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 287, + 304 + ], + "type": "text", + "content": "[31] META-STANFORD. Methane emissions technology alliance (meta), stanford natural gas initiative, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 306, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 306, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 49, + 306, + 287, + 326 + ], + "type": "text", + "content": "[32] MGISGeography. Ndvi (normalized difference vegetation index), 1979. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 327, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 287, + 348 + ], + "type": "text", + "content": "[33] Gunnar Myhre, Drew Shindell, and Julia Pongratz. Anthropogenic and natural radiative forcing. 2014. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "type": "text", + "content": "[34] Darius Phiri, Matamyo Simwanda, Serajis Salekin, Vincent R Nyirenda, Yuji Murayama, and Manjula Ranagalage. Sentinel-2 data for land cover/use mapping: a review. Remote Sensing, 12(14):2291, 2020. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 394, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 287, + 447 + ], + "type": "text", + "content": "[35] Dar A Roberts, Eliza S Bradley, Ross Cheung, Ira Leifer, Philip E Dennison, and Jack S Margolis. Mapping methane emissions from a marine geological seep source using imaging spectrometry. Remote Sensing of Environment, 114(3):592-606, 2010. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "type": "text", + "content": "[36] CA Shah, PK Varshney, and MK Arora. Ica mixture model algorithm for unsupervised classification of remote sensing imagery. International Journal of Remote Sensing, 28(8):1711-1731, 2007. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 494, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 494, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 494, + 287, + 514 + ], + "type": "text", + "content": "[37] David MJ Tax and Robert PW Duin. Support vector data description. Machine learning, 54(1):45-66, 2004. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 516, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 516, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 49, + 516, + 287, + 579 + ], + "type": "text", + "content": "[38] DR Thompson, I Leifer, H Bovensmann, M Eastwood, M Fladeland, C Frankenberg, K Gerilowski, RO Green, S Kratwurst, T Krings, et al. Real-time remote detection and measurement for airborne imaging spectroscopy: a case study with methane. Atmospheric Measurement Techniques, 8(10):4383-4397, 2015. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 582, + 287, + 635 + ], + "type": "text", + "content": "[39] David R Thompson, Anuj Karpatne, Imme Ebert-Uphoff, Christian Frankenberg, Andrew K Thorpe, Brian D Bue, and Robert O Green. Isgeo dataset jpl-ch4-detection-2017-v1. 0: A benchmark for methane source detection from imaging spectrometer data. 2017. 2, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 637, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 637, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 637, + 287, + 712 + ], + "type": "text", + "content": "[40] Andrew K Thorpe, Christian Frankenberg, David R Thompson, Riley M Duren, Andrew D Aubrey, Brian D Bue, Robert O Green, Konstantin Gerilowski, Thomas Krings, Jakob Borchardt, et al. Airborne doas retrievals of methane, carbon dioxide, and water vapor concentrations at high spatial resolution: application to aviris-ng. Atmospheric Measurement Techniques, 10(10):3833-3850, 2017. 1, 2, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 296 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[41] Andrew K Thorpe, Dar A Roberts, Eliza S Bradley, Christopher C Funk, Philip E Dennison, and Ira Leifer. High resolution mapping of methane emissions from marine and terrestrial sources using a cluster-tuned matched filter technique and imaging spectrometry. Remote Sensing of Environment, 134:305-318, 2013. 1, 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 161 + ], + "type": "text", + "content": "[42] George Turin. An introduction to matched filters. IRE transactions on Information theory, 6(3):311-329, 1960. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[43] Oytun Ulutan, ASM Iftekhar, and Bangalore S Manjunath. Vsgnet: Spatial attention network for detecting human object interactions using graph convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13617-13626, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 239 + ], + "type": "text", + "content": "[44] Foreign Agriculture Service US Dept. of Agriculture. Normalized difference vegetation index (ndvi), 1969. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 242, + 546, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 546, + 296 + ], + "type": "text", + "content": "[45] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8826, 2019. 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17618" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_content_list.json b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9276dda3428ee7cc5ebe192fc1976efffa740bfb --- /dev/null +++ b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_content_list.json @@ -0,0 +1,1742 @@ +[ + { + "type": "text", + "text": "Micron-BERT: BERT-based Facial Micro-Expression Recognition", + "text_level": 1, + "bbox": [ + 148, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuan-Bac Nguyen1, Chi Nhan Duong2, Xin Li3, Susan Gauch1, Han-Seok Seo4, Khoa Luu1", + "bbox": [ + 120, + 162, + 844, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ CVIU Lab, University of Arkansas, USA $^{2}$ Concordia University, Canada", + "bbox": [ + 179, + 180, + 790, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 West Virginia University, USA 4 Dep. of Food Science, University of Arkansas, USA", + "bbox": [ + 132, + 198, + 836, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ {xnguyen, sgauch, hanseok, khoaluu}@uark.edu, $^{2}$ {dcnhan@ieee.org}, $^{3}$ xin.li@mail.wvu.edu", + "bbox": [ + 101, + 217, + 859, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Micro-expression recognition is one of the most challenging topics in affective computing. It aims to recognize tiny facial movements difficult for humans to perceive in a brief period, i.e., 0.25 to 0.5 seconds. Recent advances in pre-training deep Bidirectional Transformers (BERT) have significantly improved self-supervised learning tasks in computer vision. However, the standard BERT in vision problems is designed to learn only from full images or videos, and the architecture cannot accurately detect details of facial micro-expressions. This paper presents Micron-BERT ( $\\mu$ -BERT), a novel approach to facial micro-expression recognition. The proposed method can automatically capture these movements in an unsupervised manner based on two key ideas. First, we employ Diagonal Micro-Attention (DMA) to detect tiny differences between two frames. Second, we introduce a new Patch of Interest (PoI) module to localize and highlight micro-expression interest regions and simultaneously reduce noisy backgrounds and distractions. By incorporating these components into an end-to-end deep network, the proposed $\\mu$ -BERT significantly outperforms all previous work in various micro-expression tasks. $\\mu$ -BERT can be trained on a large-scale unlabeled dataset, i.e., up to 8 million images, and achieves high accuracy on new unseen facial micro-expression datasets. Empirical experiments show $\\mu$ -BERT consistently outperforms state-of-the-art performance on four micro-expression benchmarks, including SAMM, CASME II, SMIC, and CASME3, by significant margins. Code will be available at https://github.com/uark-cviu/Micron-BERT", + "bbox": [ + 73, + 301, + 473, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 767, + 209, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Facial expressions are a complex mixture of conscious reactions directed toward given stimuli. They involve experiential, behavioral, and physiological elements. Because they are crucial to understanding human reactions, this topic has been widely studied in various application domains [5]. In general, facial expression problems can be classified into two main categories, macro-expression, and", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad7894b0dc5aadb9f880d68016148b89febed11c0d74cd3d7b49092d87acef88.jpg", + "image_caption": [ + "Figure 1. Given two frames from a high-speed video, the proposed $\\mu$ -BERT method can localize and highlight the regions of micromovements. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 511, + 268, + 880, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "micro-expression. The main differences between the two are facial expression intensities, and duration [2]. In particular, macro-expressions happen spontaneously, cover large movement areas in a given face, e.g., mouth, eyes, cheeks, and typically last from 0.5 to 4 seconds. Humans can usually recognize these expressions. By contrast, micro-expressions are involuntary occurrences, have low intensity, and last between 5 milliseconds and half a second. Indeed, micro-expressions are challenging to identify and are mostly detectable only by experts. Micro-expression understanding is essential in numerous applications, primarily lie detection, which is crucial in criminal analysis.", + "bbox": [ + 496, + 521, + 892, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Micro-expression identification requires both semantics and micro-movement analysis. Since they are difficult to observe through human eyes, a high-speed camera, usually with 200 frames per second (FPS) [6, 15, 51], is typically used to capture the required video frames. Previous work [11] tried to understand this micro information using MagNet [29] to amplify small motions between two frames, e.g., onset and apex frames. However, these methods still have limitations in terms of accuracy and robustness. In summary, the contributions of this work are four-fold:", + "bbox": [ + 496, + 703, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- A novel Facial Micro-expression Recognition (MER) via Pre-training of Deep Bidirectional Transformers approach (Micron-BERT or $\\mu$ -BERT) is presented to", + "bbox": [ + 517, + 854, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1482", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tackle the problem in a self-supervised learning manner. The proposed method aims to identify and localize micro-movements in faces accurately.", + "bbox": [ + 107, + 90, + 467, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- As detecting the tiny moment changes in faces is an essential input to the MER module, a new Diagonal Micro Attention (DMA) mechanism is proposed to precisely identify small movements in faces between two consecutive video frames.", + "- A new Patch of Interest (POI) module is introduced to efficiently spot facial regions containing the microexpressions. Far apart from prior methods, it is trained in an unsupervised manner without using any facial labels, such as facial bounding boxes or landmarks.", + "- The proposed $\\mu$ -BERT framework is designed in a self-supervised learning manner and trained in an end-to-end deep network. Indeed, it consistently achieves State-of-the-Art (SOTA) results in various standard micro-expression benchmarks, including CASME II [50], CASME3 [14], SAMM [6] and SMIC [15]. It achieves high recognition accuracy on new unseen subjects of various gender, age, and ethnicity." + ], + "bbox": [ + 96, + 142, + 467, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 446, + 217, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generally, prior studies in micro-expression can be divided into two categories, including micro-expression spotting (MES) and micro-expression recognition.", + "bbox": [ + 76, + 472, + 467, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Micro-Expression Spotting (MES). The goal of MES is to determine the specific instant during which a micro-expression occurs. Li et al. [16] adopted a spatial-channel attention network to detect micro-expression action units. Tran et al. [39] attempted to standardize with the SMIC-E database and an evaluation protocol. MESNet [43] introduced a CNN-based approach with a $(2 + 1)\\mathrm{D}$ convolutional network, a clip proposal, and a classifier.", + "bbox": [ + 76, + 518, + 467, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Micro-Expression Recognition (MER). The goal of MER tasks is to classify the facial micro-expressions in a video. Ling et al. [11] present a new way of learning facial graph representations, allowing these small movements to be seen. Kumar and Bhanu [31] exploited connections between landmark points and their optical flow patch and achieved improvements over state-of-the-art (SOTA) methods for both the CASME II and SAMM. Liu et al. [21] presented a new method using transfer learning achieved an accuracy of $84.27\\%$ on a composite of three datasets. Wang et al. [45] presented an Eulerian-motion magnification-based approach that highlights these small movements.", + "bbox": [ + 76, + 640, + 467, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other Work. Other research, while not necessarily on MES or MER, is relevant to our approach. An advance in video motion magnification is shown in [29], outperforming the SOTA methods in multiple areas. This learning-based model can extract filters from data directly rather than rely on ones designed by hand, like the state-of-the-art method.", + "bbox": [ + 76, + 821, + 467, + 911 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. BERT Revisited", + "text_level": 1, + "bbox": [ + 500, + 89, + 658, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. BERT in Vision Problems", + "text_level": 1, + "bbox": [ + 500, + 114, + 732, + 130 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers and deep learning have significantly improved results for many tasks in computer vision [1, 7, 9, 22, 23, 26, 27, 30, 40, 41]. Worth mentioning is Vision Transformer (ViT) [7], one of the first research efforts at the intersection of Transformers and computer vision. Unlike the traditional CNN network, ViT splits an image into a sequence of patches and applies the Transformers-based framework directly. Inspired by the success of BERT in Natural Language Processing (NLP), Bidirectional Encoder representation from Image Transformers (BEiT) [1] is presented as a self-supervised learning framework in computer vision. In particular, image patches are tokenized using DALL-E [32] to the visual tokens. These tokens are then randomly masked before feeding into the transformer backbone. The training objective is to recover the original visual tokens from the corrupted patches. These methods [1, 38] have marked a remarkable improvement compared to supervised learning methods by leveraging large-scale unlabelled datasets, e.g., ImageNet-1K, ImagNet-21K [33], to discover semantic information.", + "bbox": [ + 496, + 138, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2. Limitations of BERT in Vision Problems", + "text_level": 1, + "bbox": [ + 500, + 454, + 844, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One limitation of using BERT in vision problems is the tokenization step. In the NLP field, a token has precisely one word mapped into it. In vision problems, however, many possible images or patches can share the same token as long as they have the same content. Therefore, designing a BERT model to mask a token and train a prediction model in the missing contexts in computer vision is more challenging than NLP. In addition, the tokenizer, i.e., DALL-E [32], is not robust enough to map similar contexts to a token. It yields noise in the tokenization process and affects the overall training performance. He et al., [9] presented a Masked Auto Encoder (MAE) that utilizes the BERT framework. Instead of tokenizing images, it eliminates patches of an image via a random masking strategy and reconstructs the context of these masked patches to the original content. Although this method can avoid using the tokenizer, it only considers the context inside an image. Thus, it does not apply to micro-expression, which requires understanding semantic information from consecutive video frames. In this paper, $\\mu$ -BERT is presented to address these limitations.", + "bbox": [ + 496, + 479, + 890, + 781 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4. The Proposed $\\mu$ -BERT Approach", + "text_level": 1, + "bbox": [ + 498, + 797, + 802, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\mu$ -BERT is designed to model micro-changes of facial texture across temporal dimensions, which is hard to observe by unaided human eyes via a reconstruction process. The proposed $\\mu$ -BERT architecture, shown in Figure 2, consists of five main blocks: a $\\mu$ -Encoder, Patch", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1483", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a96aa8f2d5c59122ab599131db679ec754bd921f014ce2a01b7bf27611f9fce.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed $\\mu$ -BERT approach to facial micro-expression recognition." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 851, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of Interest (PoI), Blockwise Swapping, Diagonal Micro Attention (DMA), and a $\\mu$ -Decoder. Given input images $I_{t}$ and $I_{t + \\delta}$ , the role of the $\\mu$ -Encoder is to represent $I_{t}$ and $I_{t + \\delta}$ into latent vectors. Then, Patch of Interest (PoI) constrains $\\mu$ -BERT to look into facial regions containing microexpressions rather than unrelated regions such as the background. Blockwise Swapping and Diagonal Micro Attention (DMA) allow the model to focus on facial regions that primarily consist of micro differences between frames. Finally, $\\mu$ -Decoder reconstructs the output signal back to the determined one. Compared to prior works, $\\mu$ -BERT can adaptively focus on changes in facial regions while ignoring the ones in the background and effectively recognizes micro-expressions even when face movements occur. Moreover, $\\mu$ -BERT can also alleviate the dependency on the accuracy of alignment approaches in pre-processing step.", + "bbox": [ + 75, + 316, + 473, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Non-overlapping Patches Representation", + "text_level": 1, + "bbox": [ + 76, + 566, + 426, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In $\\mu$ -BERT, an input frame $I_{t} \\in \\mathbb{R}^{H \\times W \\times C}$ is divided into a set of several non-overlapping patches $\\mathcal{P}_t$ as Eqn. (1).", + "bbox": [ + 76, + 589, + 468, + 622 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {t} = \\left\\{p _ {t} ^ {i} \\right\\} _ {i = 0} ^ {N _ {p} - 1} \\quad \\left| \\mathcal {P} _ {t} \\right| = H W / \\left(p s ^ {2}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 632, + 468, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H, W, C$ are the height, width, and number of channels, respectively. Each patch $p_t^i$ has a resolution of $ps \\times ps$ . In our experiments, $H = W = 224$ , $C = 3$ , and $ps = 8$ .", + "bbox": [ + 76, + 655, + 468, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. $\\mu$ -Encoder", + "text_level": 1, + "bbox": [ + 76, + 710, + 192, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each patch $p_i \\in \\mathcal{P}_t$ is linearly projected into a latent vector of dimension $d$ denoted as $\\mathbf{z}_t^i \\in \\mathbb{R}^{1 \\times d}$ , with additive fixed positional encoding [42]. Then, an image $I_t$ can be represented as in Eqn. (2).", + "bbox": [ + 76, + 733, + 468, + 794 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {t} = \\operatorname {c o n c a t} \\left[ \\mathbf {z} _ {t} ^ {0}, \\mathbf {z} _ {t} ^ {1}, \\dots \\mathbf {z} _ {t} ^ {N _ {p} - 1} \\right] \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 803, + 468, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} ^ {i} = \\alpha \\left(p _ {t} ^ {i}\\right) + \\mathbf {e} (i)\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 830, + 258, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\mathbf{e}$ are the projection embedding network and positional embedding, respectively. Let $\\mu$ -Encoder, denoted as $\\mathcal{E}$ , be a stack of continuous blocks. Each block consists", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1aa3c221ced05c8342dc33606299d97ff1be419bb0c70f3957dd271cdb0d336f.jpg", + "image_caption": [ + "Figure 3. Builing block of Encoder and Decoder. Each block includes Multi-Head Attention (MHA) and Layer Normalization." + ], + "image_footnote": [], + "bbox": [ + 573, + 311, + 820, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of alternating layers of Multi Head Attention (MHA) and Multi-Layer Perceptron (MLP), as illustrated in Figure 3. The Layer Norm (LN) is employed to the input signal before feeding to MHA and MLP layers, as in Eqn. (3).", + "bbox": [ + 498, + 436, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} _ {l} = \\mathbf {x} _ {l - 1} + \\operatorname {M H A} (\\operatorname {L N} (\\mathbf {x} _ {l - 1}))\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 503, + 792, + 518 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {l} = \\mathbf {x} ^ {\\prime} _ {l} + \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\mathbf {x} ^ {\\prime} _ {l}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 522, + 890, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {0} = \\mathbf {Z} _ {t}, 1 \\leq l \\leq L _ {e}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 541, + 733, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_{e}$ is the number of blocks in $\\mathcal{E}$ . Given $\\mathbf{Z}_t$ , the output latent vector $\\mathbf{P}_{\\mathbf{t}}$ is represented as in Eqn. (4).", + "bbox": [ + 500, + 561, + 888, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {t} = \\mathcal {E} \\left(\\mathbf {Z} _ {t}\\right) \\quad \\mathbf {P} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 597, + 890, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.3. $\\mu$ -Decoder", + "text_level": 1, + "bbox": [ + 500, + 621, + 616, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed auto-encoder is designed symmetrically. It means that the decoder part denoted as $\\mathcal{D}$ , has a similar architecture to the encoder $\\mathcal{E}$ . Given a latent vector $\\mathbf{P}_t$ , the decoded signal $\\mathbf{Q}_t$ is represented as in Eqn. (5).", + "bbox": [ + 498, + 645, + 890, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {t} = \\mathcal {D} \\left(\\mathbf {P} _ {t}\\right) \\quad \\mathbf {Q} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 710, + 890, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We add one more Linear layer to interpolate $\\mathbf{Q}_t$ to an intermediate signal $\\mathbf{y}_t$ before reshaping it into the image size.", + "bbox": [ + 498, + 734, + 890, + 766 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\xrightarrow {\\text {l i n e a r}} \\mathbf {y} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times p s \\times p s \\times C} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 786, + 890, + 805 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times p s \\times p s \\times C} \\xrightarrow {\\text {r e s h a p e}} \\mathbf {y} _ {t} ^ {\\prime} \\in \\mathbb {R} ^ {H \\times W \\times C}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 810, + 833, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.4. Blockwise Swapping", + "text_level": 1, + "bbox": [ + 500, + 838, + 694, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given two frames $I_{t}$ and $I_{t + \\delta}$ , we realize the fact that:", + "bbox": [ + 517, + 861, + 877, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {\\delta \\rightarrow 0} s \\left(p _ {t} ^ {i}, p _ {t + \\delta} ^ {i}\\right) = 1 \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 882, + 890, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1484", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6ec2580c36031dbf28f7accbd1bfed8be715f7615dafd87088460c7d1285de8b.jpg", + "image_caption": [ + "Figure 4. Blockwise Swapping. For each triplet, we present the $I_{t}$ (left), the $I_{t + \\delta}$ (middle) and the $I_{t / s}$ (right). The yellow blocks in $I_{t / s}$ represent swapped patches from $I_{t + \\delta}$ that are randomly swapped into $I_{t}$ . Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 439, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_t^i$ is the $i^{th}$ -patch at frame $t$ . $s$ denotes a function to measure the similarity between $p_t^i$ and $p_{t + \\delta}^i$ where a higher score indicates higher similarity and $0 \\leq s(p_t^i, p_{t + \\delta}^i) \\leq 1$ . Given a patch correlation as in Eqn. (7), we propose a Blockwise Swapping mechanism to (1) firstly randomly swap two corresponding blocks $p_t^i$ and $p_{t + \\delta}^i$ between two frames to create a swapped image $I_{t / s}$ , and then (2) enforce the model to spot these changes and reconstruct $I_t$ from $I_{t / s}$ . By doing so, the model is further strengthened in recognizing and restoring the swapped patches. As a result, the learned model can be enhanced by the capability to notice small differences between frames. Moreover, as shown in Eqn. (7), shorter time $\\delta$ causing larger similarity between $I_t$ from $I_{t / s}$ can further help to enhance the robustness on spotting these differences. The detail of this strategy is described in Algorithm 1 and Figure 4.", + "bbox": [ + 75, + 295, + 473, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.5. Diagonal Micro Attention (DMA)", + "text_level": 1, + "bbox": [ + 76, + 550, + 372, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a result of Blockwise Swapping, the image patches $\\mathcal{P}_{t / s}$ from $I_{t / s}$ consists of two types, i.e. $p_{t / s}^{j}$ from $\\mathcal{P}_t$ of $I_{t}$ and $p_{t / s}^{i}$ from $\\mathcal{P}_{t + \\delta}$ of $I_{t + \\delta}$ . Then, the next stage is to learn how to reconstruct $\\mathcal{P}_t$ from $\\mathcal{P}_{t / s}$ . Since $p_{t / s}^{i}$ includes all changes between $I_{t}$ and $I_{t / s}$ , more emphasis is placed on $p_{t / s}^{i}$ during reconstruction process. Theoretically, the ground truth of the index of $p_{t / s}^{i}$ in $\\mathcal{P}_{t / s}$ can be utilized to enforce the model focusing on these swapped patches. However, adopting this information may reduce the learning capability to spot these microchanges. Therefore, a novel attention mechanism named Diagonal MicroAttention (DMA) is presented to enforce the network automatically focusing on swapped patches $p_{t / s}^{i}$ and equip it with the ability to precisely spot and identify all changes between images. Notice that these changes may include patches in the background. The following section introduces a solution to constrain the learned network focusing on only meaningful facial regions. The details of DMA are presented in Figure 5. Formally, we construct an attention map $\\hat{A}$ between $\\mathcal{P}_{t + \\delta}$ and $\\mathcal{P}_{t / s}$ where the diag $(\\hat{A})$ illustrates correlations between two corresponding patches $p_{t + \\delta}^{i}$", + "bbox": [ + 75, + 575, + 472, + 904 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c2f42b4bb6f5535879a081ec2c47812701308136e2a1b03b3795f2d919f6d784.jpg", + "image_caption": [ + "Figure 5. Diagonal Micro-Attention (DMA) module. Diagonal values from the attention map between $\\mathbf{P}_{t / s}$ and $\\mathbf{P}_{t + \\delta}$ are used to rank the importance of each patch in the swapped image." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 882, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $p_{t / s}^{i}$ . From the observation that $\\hat{A}(i, i) > \\hat{A}(j, j)$ for all $p_{t / s}^{i} \\in \\mathcal{P}_{t + \\delta}$ and $p_{t / s}^{j} \\in \\mathcal{P}_{t}$ , $\\text{diag}(\\hat{A})$ can be effectively adopted as weights indicating important features. Full operations of DMA are presented in Eqn. (8) and Eqn. (9).", + "bbox": [ + 496, + 261, + 893, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {A} = \\operatorname {s o f t m a x} \\left(Q \\left(\\mathbf {P} _ {t + \\delta}\\right) \\otimes K \\left(\\mathbf {P} _ {t / s}\\right) ^ {T}\\right), \\sum_ {j = 0} ^ {N _ {p}} \\hat {A} (i, j) = 1 \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 349, + 892, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {d m a} = \\operatorname {d i a g} (\\hat {A}) \\times V \\left(\\mathbf {P} _ {t / s}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 415, + 892, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\times$ denotes the Element-wise multiplication operator.", + "bbox": [ + 500, + 443, + 890, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.6. Patch of Interest (POI)", + "text_level": 1, + "bbox": [ + 500, + 468, + 714, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Section 4.5, Diagonal Micro-Attention has been introduced to weigh the importance of swapped patches automatically. These swapped patches are randomly produced via Blockwise Swapping, as in Algorithm 1. In theory, the ideal case is when all swapped patches are located within the facial region only so that the deep network can learn the micro-movements from the facial parts solely and not be distracted by the background. In practice, however, we", + "bbox": [ + 496, + 489, + 893, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Blockwise Swapping", + "text_level": 1, + "bbox": [ + 501, + 636, + 732, + 652 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Input: $\\mathcal{P}_t, \\mathcal{P}_{t + \\delta}$ image patches ( $N_p = h \\times w$ ); $r_s$ : swapping ratio (default: 0.5); min_bs: minimum block size (default: 16); min_ar: minimum aspect ratio (default: 0.3)", + "Output: Swapped image patches $\\mathcal{P}_{t / s}$", + "$\\mathcal{P}_{t / s}\\gets \\mathcal{P}_t;N_p\\gets |\\mathcal{P}_t|$", + "$c\\gets 0$", + "while $c\\leq r_s\\times N_p$ do", + "$bs\\gets \\mathsf{rnd}(min\\_ bs,r_s\\times N_p - c)$", + "ar←rnd(min_ar,1/min_ar)", + "$m,n\\gets \\sqrt{bs}\\cdot ar,\\sqrt{bs / ar}$", + "$p,q\\gets \\mathsf{rnd}(0,h - m),\\mathsf{rnd}(0,w - n)$", + "$\\forall i \\in [p, p + m), j \\in [q, q + n)$ :", + "$k\\gets i\\times w + j$", + "$\\mathcal{P}_{t / s}(k)\\gets \\mathcal{P}_{t + \\delta}(k)$", + "$c\\gets c + m\\times n$", + "end while", + "return $\\mathcal{P}_{t / s}$" + ], + "bbox": [ + 509, + 655, + 893, + 891 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1485", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f5a4c6c365a07c269c3222fdd0d4939373702d463c4f899b3eee7a144e90857.jpg", + "image_caption": [ + "Figure 6. Patch of Interest (POI) module. The $\\mathbf{P}_t$ and $\\mathbf{P}_{t + \\delta /c}$ are sequence of patch features of $I_{t + \\delta}$ and its random cropped version. $p_{t + \\delta}^{CT}$ and $p_{t + \\delta /c}^{CT}$ are their corresponding contextual features." + ], + "image_footnote": [], + "bbox": [ + 169, + 88, + 377, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can only identify which parts are selected in the Blockwise Swapping algorithm if the facial regions are available. Thus, the Patch of Interest (POI) is introduced to automatically explore the salient regions and ignore the background patches in an image. Apart from prior methods, the proposed POI leverages the characteristic of self-attention and can be achieved through self-learning without facial labels, such as facial bounding boxes or segmentation masks. The idea of the POI module is illustrated in Figure 6. Thanks to POI, a capability of automatically focusing on facial regions is further equipped to the learned model, making it more robust against facial movements.", + "bbox": [ + 75, + 267, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The POI relies on the contextual agreement between the frame $I_{t + \\delta}$ and $\\mathrm{Crop}(I_{t + \\delta})$ . Motivated by the BERT framework, we add a Contextual Token $z^{CT}$ to the beginning of the sequence of patches, as in Eqn. (2), to learn the contextual information in the image. The deeper this token passes through the Transformer blocks, the more information is accumulated from $z_{t}^{i} \\in \\mathcal{P}_{t}$ . As a result, $z^{CT}$ becomes a placeholder to store the information extracted from other patches in the sequence and present the contextual information of the image. Let $p_{t + \\delta}^{CT}$ and $p_{t + \\delta / c}^{CT}$ be the contextual features of frame $I_{t + \\delta}$ and its cropped version $\\mathrm{Crop}(I_{t + \\delta})$ respectively. The agreement loss is then defined as in Eqn. (10).", + "bbox": [ + 75, + 449, + 470, + 631 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a g g} = H \\left(p _ {t + \\delta} ^ {C T}, p _ {t + \\delta / c} ^ {C T}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 643, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H$ is the function that enforces $p_{t + \\delta}^{CT}$ to be similar to $p_{t + \\delta / crop}^{CT}$ so that the model can discover the salient patches. The POI can be extracted from the attention map $A$ at the last attention layer of encoder $\\mathcal{E}$ . In particular, we measure:", + "bbox": [ + 76, + 667, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {t + \\delta} = A [ 0, \\cdot ] = \\left[ s _ {t + \\delta} ^ {0}, s _ {t + \\delta} ^ {1}, \\dots s _ {t + \\delta} ^ {N _ {p} - 1} \\right] \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 753, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sum_{i=0}^{N_p-1} s_{t+\\delta}^i = 1$ . The higher the score $s_{t+\\delta}^i$ , the richer the patch contains contextual information. Now, Eqn. (9) can be reformulated as in Eqn. (12).", + "bbox": [ + 76, + 789, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {W} = \\operatorname {d i a g} (\\hat {A}) \\times S _ {t + \\delta} \\\\ \\mathbf {P} _ {d m a} = \\mathbf {W} \\times V (\\mathbf {P} _ {t / s}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 849, + 351, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.7. Loss Functions", + "text_level": 1, + "bbox": [ + 500, + 90, + 651, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed $\\mu$ -BERT deep network is optimized using the proposed loss function as in Eqn. (13).", + "bbox": [ + 500, + 114, + 890, + 143 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\gamma \\times \\mathcal {L} _ {r} + \\beta \\times \\mathcal {L} _ {a g g} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 155, + 890, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ and $\\beta$ are the weights for each loss.", + "bbox": [ + 500, + 181, + 792, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction Loss. The output of the decoder $\\mathbf{y}'_t$ is reconstructed to the original image $I_t$ using the Mean Square Error (MSE) function.", + "bbox": [ + 500, + 196, + 890, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\operatorname {M S E} \\left(\\mathbf {y} ^ {\\prime} _ {t}, I _ {t}\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 251, + 890, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contextual Agreement Loss. MSE is also used to enforce the similarity of contextual features of $I_{t + \\delta / crop}$ and $I_{t + \\delta}$", + "bbox": [ + 500, + 277, + 890, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a g g} = \\operatorname {M S E} \\left(p _ {t + \\delta} ^ {C T}, p _ {t + \\delta / c r o p} ^ {C T}\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 318, + 890, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 349, + 705, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets and Protocols", + "text_level": 1, + "bbox": [ + 500, + 375, + 710, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CASME II [50]. With a 200 fps sampling rate and a facial resolution of $280 \\times 340$ , CASME II provides 247 microexpression samples from 26 subjects of the same ethnicity. Labels include apex frames, action units, and emotions.", + "bbox": [ + 498, + 397, + 890, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SAMM [6]. Also, using a 200 fps frame rate and a facial resolution of $400 \\times 400$ , SAMM consists of 159 samples from 32 participants and 13 ethnicities. The samples all have emotions, apex frames, and action unit labels.", + "bbox": [ + 498, + 458, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SMIC [15]. SIMC is made up of 164 samples. Lacking apex frame and action unit labels, the samples span 16 participants of 3 ethnicities. The recordings are taken with a resolution of $640 \\times 480$ at 100 fps.", + "bbox": [ + 498, + 518, + 890, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CASME3 [14]. Officially known as $\\mathrm{CAS(ME)^3}$ provides 1,109 labeled micro-expressions and 3,490 labeled macro-expressions. This dataset has roughly 80 hours of footage with a resolution of $1280 \\times 720$ .", + "bbox": [ + 498, + 579, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Micro-Expression Self-Training", + "text_level": 1, + "bbox": [ + 500, + 650, + 779, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use all raw frames from CASME3 for self-training except frames of test set. It is important to note that we do not use labels or meta information such as onset, offset, and apex index frames nor labeled emotions. In total, we construct an unlabelled dataset of 8M frames. The images are resized to $224 \\times 224$ . Then, each image is divided into patches of $8 \\times 8$ , yielding $N_{p} = 784$ patches. The temporal index $\\delta$ is selected randomly between a lower bound of 5 and an upper bound of 11, experimentally. The swapping ratio $r_{s}$ is selected as $50\\%$ of the number of patches being swapped from $I_{t + \\delta}$ to $I_{t}$ . Each patch is projected to a latent space of $d = 512$ dimensions before being fed into the encoder and decoder. For the encoder and decoder, we keep the same $d$ for all vectors and similar configurations, i.e., $L_{e} = L_{d} = 4$ . $\\mu$ -BERT is implemented easily in", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1486", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5ff512598e1c0412f6804a0aafa191e4d7017b0422cd50e8be31feec7e01c1ad.jpg", + "image_caption": [ + "Figure 7. We demonstrate how $\\mu$ -BERT perceives the tiny differences between two frames. The first two rows are onset and apex input frames. The third and fourth rows are the results of RAFT and MagNet, respectively. The rest of the rows are our $\\mu$ -BERT results." + ], + "image_footnote": [], + "bbox": [ + 179, + 88, + 787, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pytorch framework and trained by $32 \\times$ A100 GPUs (40G each). The learning rate is set to 0.0001 initially and then reduced to zero gradually under ConsineLinear [24] policy. The batch size is set to 64/GPU. The model is optimized by AdamW [25] for 100 epochs. The training is completed within three days.", + "bbox": [ + 75, + 500, + 473, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Micro-Expression Recognition", + "text_level": 1, + "bbox": [ + 76, + 601, + 346, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We leverage the pretrained $\\mu$ -BERT as an initial weight and take the encoder $\\mathcal{E}$ and DMA module of $\\mu$ -BERT as the MER backbone. The input of MER is the onset and apex frames which correspond to $I_{t}$ and $I_{t + \\delta}$ respectively. In Eqn. (8), $\\mathbf{P}_{dma}$ are the features representing the micro changes and movements between onset and apex frames. They can be effectively adopted for recognizing microexpressions. We adopt the standard metrics and protocols of MER2019 challenge [34] with the unweighted F1 score $UF1 = \\frac{1}{C} \\sum_{i=0}^{C-1} \\frac{2 \\times TP_i}{TP_i + FP_i + FN_i}$ and accuracy $UAR = \\frac{1}{C} \\sum_{i=0}^{C-1} \\frac{TP_i}{N_i}$ , where $C$ is the number of MEs, $N_i$ is the total number of $i^{th}$ ME in the dataset. Leave-one-out cross-validation (LOOCV) scheme is used for evaluation.", + "bbox": [ + 75, + 618, + 470, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Results", + "text_level": 1, + "bbox": [ + 76, + 830, + 171, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed $\\mu$ -BERT shows a significant improvement over prior methods and baselines, as shown in Table 1 on the CASME3. Tested using 3, 4, and 7 emotion", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classes, $\\mu$ -BERT achieves double-digit gains over the compared methods in each category. In the case of 3 emotion classes, $\\mu$ -BERT achieved a $56.04\\%$ UF1 score and $61.25\\%$ UAR, compared to RCN-A's [49] $39.28\\%$ UF1 and $38.93\\%$ UAR. For 4 emotion classes, $\\mu$ -BERT outperforms Baseline (+Depth) [14] $47.18\\%$ to $30.01\\%$ for UF1 and $49.13\\%$ to $29.82\\%$ for UAR. Large gains over Baseline (+Depth) [14] are seen in the case of 7 emotion classes, where $\\mu$ -BERT attains UF1 and UAR scores of $32.64\\%$ and $32.54\\%$ respectively, compared to $17.73\\%$ and $18.29\\%$ for the baseline.", + "bbox": [ + 496, + 500, + 893, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 details results for CASMEII. $\\mu$ -BERT shows improvements over all other methods. For three categories, it achieves a UF1 of $90.34\\%$ and UAR of $89.14\\%$ , representing $3.37\\%$ and $0.86\\%$ increases over the prior leading method (OFF-ApexNet [8]), respectively. Similar improvement is seen in five categories: a $4.83\\%$ over TSCNN [35] in terms of UF1 and a $0.89\\%$ increase over SMA-STN [19] for UAR. Similarly, $\\mu$ -BERT performs competitively with other methods on the SAMM as seen in Table 3. Using 5 emotion classes, $\\mu$ -BERT outperforms MinMaNet [47] by a large margin in terms of UF1 ( $83.86\\%$ vs $76.40\\%$ ) and UAR ( $84.75\\%$ vs $76.70\\%$ ), respectively. The performance of $\\mu$ -BERT on SMIC is compared against several others in Table 4. $\\mu$ -BERT outperforms others with a $7.5\\%$ increase in UF1 to $85.5\\%$ and a $3.97\\%$ boost in UAR to $83.84\\%$ .", + "bbox": [ + 496, + 655, + 893, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the composite dataset, $\\mu$ -BERT again outperforms", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1487", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4a1225ad7e1e8674e3c73f7883ada7726d6e5ba1b809682d6dce3f63ad32f9a3.jpg", + "table_caption": [ + "Table 1. MER on the CASME3 dataset." + ], + "table_footnote": [], + "table_body": "
Method# ClassesUF1 (%)UAR(%)
FR [54]334.9334.13
STSTNet [18]337.9537.92
RCN-A [49]339.2838.93
μ-BERT (ours)356.0461.25
Baseline [14]429.1529.10
Baseline (+Depth) [14]430.0129.82
μ-BERT (ours)447.1849.13
Baseline [14]717.5918.01
Baseline(+Depth) [14]717.7318.29
μ-BERT (ours)732.6432.54
", + "bbox": [ + 81, + 109, + 465, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/21638b10395015dc26f1fcd4f70bcc0fe6643e17a7b8a4758e7f3cf63dc05b86.jpg", + "table_caption": [ + "Table 2. MER on CASME II dataset." + ], + "table_footnote": [], + "table_body": "
Method# ClassesUF1 (%)UAR (%)
LR-GACNN [10]570.9081.30
AMAN [46]571.0075.40
Graph-TCN [13]572.4673.98
DSTAN [44]573.0075.00
GEME [28]573.5475.20
MiMaNet [47]575.9079.90
SMA-STN [19]579.4682.59
TSCNN [35]580.7080.97
μ-BERT (ours)585.5383.48
STSTNet [17]383.8286.86
OFF-ApexNet [8]386.9788.28
MAE [9]388.0387.28
μ-BERT (ours)390.3489.14
", + "bbox": [ + 81, + 347, + 465, + 628 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3692436b6c7b26b4bb8bf88d868b90e38603b86a40f0bd16e7fb91d5bcd01d89.jpg", + "table_caption": [ + "Table 3. MER on SAMM dataset." + ], + "table_footnote": [], + "table_body": "
Method# ClassesUF1 (%)UAR (%)
AMAN [46]567.0068.85
SMA-STN [19]570.3377.20
GRAPH-AU [12]570.4574.26
MTMNet [48]573.6074.10
MiMaNet [47]576.4076.70
MAE [9]580.4088.98
μ-BERT (ours)583.8684.75
", + "bbox": [ + 81, + 652, + 465, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "other methods (Table 5). Attaining a UF1 score of $89.03\\%$ and UAR of $88.42\\%$ , $\\mu$ -BERT realizes $0.73\\%$ , and $0.82\\%$ gains over previous best MiMaNet [47], respectively. Table 6 shows the impact of DMA and POI on CASME3. Our", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/203cb647620b6430cc3ba488647a5ae3a897daffaccf8cd719977c534048807b.jpg", + "table_caption": [ + "Table 4. MER on SMIC dataset." + ], + "table_footnote": [], + "table_body": "
Method# ClassesUF1 (%)UAR (%)
DIKD [36]371.0076.06
TSCNN [35]372.3672.74
MTMNet [48]374.4076.00
AMAN [46]377.0079.87
MiMaNet [47]377.8078.60
DSTAN [44]378.0077.00
MAE [9]381.8680.82
μ-BERT (ours)385.5083.84
", + "bbox": [ + 504, + 109, + 888, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method gives more modest gains of approximately $2\\%$ in both metrics. A greater improvement is seen with DMA, where UF1 and UAR increase by another $2 - 4\\%$ . Significant improvement from $\\mu$ -BERT is seen when adopting both modules, with a UF1 of $32.64\\%$ and UAR of $32.54\\%$ , representing roughly $10\\%$ gains over previous methods.", + "bbox": [ + 496, + 309, + 892, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. How $\\mu$ -BERT perceives micro-movements", + "text_level": 1, + "bbox": [ + 498, + 409, + 856, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To understand the micro-movements between two frames, the onset and apex frames are inputs for $\\mu$ -BERT. These frames represent the moments that the microexpression starts and is observed. We measure $\\mathrm{diag}(\\hat{A})$ (Subsection 4.5) and $\\mathbf{S}_{t + \\delta}$ (Eqn (11)) values to identify which regions contain small movements between two frames. Comparisons of $\\mu$ -BERT with RAFT [37], i.e., optical flow-based method and MagNet [29] are also conducted as in Fig. 7. The third and fourth columns in Fig 7 show the results of RAFT [37], and MagNet [29] on spotting the micro-movements, respectively. While RAFT is an optical flow-based method, MagNet amplifies small differences between the two frames. These methods are sensitive to the environment (e.g., lighting, illuminations). Thus, noises in the background still exist in their outputs. In addition, neither RAFT nor MagNet understand semantic information in the frame and distinguish changes inside facial or background regions. Meanwhile, $\\mu$ -BERT shows its advantages in perceiving micro-movements via distinguishing the facial regions and spotting the micro-expressions. In particular, the attention map in the fifth column, in Fig. 7 illustrates the micro-differences between onset and apex frames. The higher contrast represents the higher chance of small movements in these regions. With the POI module, $\\mu$ -BERT can automatically figure out the informative patches and ignore the background ones. Then, with DMA module, $\\mu$ -BERT, can detect and localize which corresponding patches/regions contain tiny movements. As shown in the seventh column, attention maps represent the most salient regions in the image. By empowering DMA and POI, $\\mu$ -BERT effectively identifies micro-movements within facial", + "bbox": [ + 496, + 431, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1488", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/636ec6cf3ddf6b60351bc0856382b98cdb7a00de07c5b1050728d1a270db543a.jpg", + "table_caption": [ + "Table 5. MER on the Composite dataset (MECG2019)." + ], + "table_footnote": [], + "table_body": "
Method# ClassesUF1 (%)UAR(%)
Dual-Inception [55]373.2272.78
FR [53]378.3878.32
NMER [20]378.8578.24
GRAPH-AU [12]379.1479.33
ICE-GAN [52]384.5084.10
BDCNN [3]385.0985.00
moment [48]386.4085.70
MiMaNet [47]388.3087.60
MAE [9]388.5087.40
μ-BERT (ours)389.0388.42
", + "bbox": [ + 81, + 109, + 465, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "regions, as demonstrated in the last column.", + "bbox": [ + 76, + 342, + 367, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 364, + 238, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section compares $\\mu$ -BERT against other self-supervised learning (SSL) methods on the MER task. CASME3 is used for experiments since it has many unlabeled images to demonstrate the power of SSL methods. We also analyze the essential contributions of Diagonal Micro-Attention (DMA) and Patch of Interest (POI) modules. Finally, we illustrate the robustness of $\\mu$ -BERT pretrained on CASME3 on unseen datasets and domains.", + "bbox": [ + 75, + 385, + 467, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparisons with self-supervised learning methods. We utilize the encoder and decoder parts of $\\mu$ -BERT (without DMA and POI) to train previous SSL methods (MoCo V3 [4], BEIT [1], MAE [9]) and then continue learning the MER task on the large-scale database CASME3 [14]. Overall results are shown in Table 6. It is expected that ViT-S achieves the lowest performance for UF1 and UAR as ImageNet and Micro-Expression are two different domains. Three self-supervised methods (MoCo V3, BEIT, and MAE) got better results when they were pretrained on CASME before fine-tuning to the recognition task. Compared to ViT-S, these SSL methods gain remarkable performance. Especially, MAE [9] achieves $3.5\\%$ and $2\\%$ up on UF1 and UAR compared to ViT-S, respectively.", + "bbox": [ + 75, + 507, + 467, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The role of Blockwise Swapping. Our basic setup of $\\mu$ -BERT (denoted as MB1) is employed to train in an SSL manner. It is noted that only Blockwise Swapping is involved, and it does not contain either DMA or POI. Compared to MAE, MB1 outperforms MAE by $2\\%$ in both UF1 and UAR, approximately. The reasons are: (1) Blockwise Swapping enforces the model to learn local context features inside an image, i.e., $I_{t}$ , and (2) It helps the network to figure out micro-disparities between two frames $I_{t}$ and $I_{t + \\delta}$ .", + "bbox": [ + 75, + 718, + 467, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The role of DMA. This module is the guide to tell the network where to look and which patches to focus. By doing so, the $\\mu$ -BERT gets more robust knowledge of micro", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5957ac4458c140478495df38a81b31ae81bc4a3d3bf07b76e8fbe681e5f5a6dd.jpg", + "table_caption": [ + "Table 6. MER performance on CASME3 by different self-supervised methods and various settings of $\\mu$ -BERT" + ], + "table_footnote": [], + "table_body": "
MethodPre-trainDMAPOIUF1UAR
ViT-S [7]ImageNetXX20.3418.76
MoCo V3 - R50 [4]CASME3XX19.1217.36
MoCo V3 - R101 [4]CASME3XX20.1418.52
MoCo V3 [4]CASME3XX22.1319.34
BEIT [1]CASME3XX23.5419.89
MAE [9]CASME3XX23.8620.87
μ-BERT (MB1)CASME3XX25.2722.96
μ-BERT (MB2)CASME3✓X27.3526.18
μ-BERT (MB3)CASME3✓✓32.6432.54
", + "bbox": [ + 504, + 122, + 890, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "movements between two frames. For this reason, the network (denoted as MB2) achieves $2\\%$ on UF1 and a significant $4\\%$ gain on UAR compared to MB1.", + "bbox": [ + 496, + 276, + 890, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The role of POI. Since MB1 are sensitive to background noise, the micro-disparities features $\\mathbf{P}_{dma}$ might contain unwanted features coming from the background. The POI is designed as a filter that only lets the typical interesting patches belonging to the subject go through and preserves the micro-movement features only. The improvements of up to $6\\%$ compared to MB2 demonstrate the important role of POI in $\\mu$ -BERT for micro-expression tasks. Qualitative results demonstrated in Supplementary Material can further emphasize the advantages of POI in assisting the network to be robust against facial movements.", + "bbox": [ + 496, + 321, + 890, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions and Discussions", + "text_level": 1, + "bbox": [ + 498, + 493, + 764, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike a few concurrent research on micro-expression, we move forward and study how to explore BERT pretraining for this problem. In our proposed $\\mu$ -BERT, we presented a novel Diagonal Micro Attention (DMA) to learn the micro-movements of the subject across frames. The Patch of Interest (POI) module is proposed to guide the network, focusing on the most salient parts, i.e., facial regions, and ignoring the noisy sensitivities from the background. Empowered by the simple design of $\\mu$ -BERT, SOTA performance on micro-expression recognition tasks is achieved in four benchmark datasets. Our perspective will inspire more future study efforts in this direction.", + "bbox": [ + 496, + 515, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We demonstrated the efficiency of the POI module in removing noise in the background, which is sensitive to lighting and illumination. However, suppose any facial parts, e.g., the forehead, are affected by lighting conditions while there are no movements. In that case, these noisy factors might also be included as micro-difference features. The robustness with different lighting conditions will be left as our future works.", + "bbox": [ + 496, + 696, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work is supported by Arkansas Biosciences Institute (ABI) Grant, NSF WVAR-CRESH and NSF Data Science, Data Analytics that are Robust and Trusted (DART). We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1489", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. 2021. 2, 8", + "[2] Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu. Video-based facial micro-expression analysis: A survey of datasets, features and algorithms. IEEE transactions on pattern analysis and machine intelligence, 2021. 1", + "[3] Bin Chen, Kun-Hong Liu, Yong Xu, Qing-Qiang Wu, and Jun-Feng Yao. Block division convolutional network with implicit deep features augmentation for micro-expression recognition. IEEE Transactions on Multimedia, 2022. 8", + "[4] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 8", + "[5] Ciprian Adrian Coraneanu, Marc Oliu Simón, Jeffrey F. Cohn, and Sergio Escalera Guerrero. Survey on rgb, 3d, thermal, and multimodal approaches for facial expression recognition: History, trends, and affect-related applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(8):1548-1568, 2016. 1", + "[6] Adrian K. Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap. Samm: A spontaneous micro-facial movement dataset. IEEE Transactions on Affective Computing, 9(1):116-129, 2018. 1, 2, 5", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 8", + "[8] Y.S. Gan, S.T. Liong, W.C. Yau, Y.C. Huang, and L.K. Tan. Off-apexnet on micro-expression recognition system. Signal Processing: Image Communication, 74:129-139, 2019. 6, 7", + "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 7, 8", + "[10] Ankith Jain Rakesh Kumar and Bir Bhanu. Micro-expression classification based on landmark relations with graph attention convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1511–1520, June 2021. 7", + "[11] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression recognition based on facial graph representation learning and facial action unit fusion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571-1580, 2021. 1, 2", + "[12] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression recognition based on facial graph representation learning and facial action unit fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1571-1580, June 2021. 7, 8" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ling Lei, Jianfeng Li, Tong Chen, and Shigang Li. A novel Graph-TCN with a graph structured representation for micro-expression recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2237–2245, 2020. 7", + "[14] Jingting Li, Zizhao Dong, Shaoyuan Lu, Su-Jing Wang, Wen-Jing Yan, Yinhuan Ma, Ye Liu, Changbing Huang, and Xiaolan Fu. Cas(me)sup3/sup: A third generation facial spontaneous micro-expression database with depth information and high ecological validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2, 5, 6, 7, 8", + "[15] Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietikainen. A spontaneous micro-expression database: Inducement, collection and baseline. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6, 2013. 1, 2, 5", + "[16] Yante Li, Xiaohua Huang, and Guoying Zhao. Microexpression action unit detection with spatial and channel attention. Neurocomputing, 436:221-231, 2021. 2", + "[17] S.T. Liong, Y.S. Gan, J. See, H. Khor, and Y. Huang. Shallow triple stream three-dimensional cnn (ststnet) for microexpression recognition. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-5. IEEE, 2019. 7", + "[18] Sze-Teng Liong, Yee Siang Gan, John See, Huai-Qian Khor, and Yen-Chang Huang. Shallow triple stream three-dimensional cnn (ststnet) for micro-expression recognition. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-5. IEEE, 2019. 7", + "[19] Jiateng Liu, Wenming Zheng, and Yuan Zong. SMA-STN: Segmented movement-attending spatiotemporal network for micro-expression recognition. arXiv preprint arXiv:2010.09342, 2020.6.7", + "[20] Yuchi Liu, Heming Du, Liang Zheng, and Tom Gedeon. A neural micro-expression recognizer. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-4. IEEE, 2019. 8", + "[21] Yanju Liu, Yange Li, Xinhai Yi, Zuojin Hu, Huiyu Zhang, and Yanzhong Liu. Lightweight vit model for microexpression recognition enhanced by transfer learning. Frontiers in Neurorobotics, 16, 2022. 2", + "[22] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 2", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2", + "[24] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1490", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[26] Xuan-Bac Nguyen, Duc Toan Bui, Chi Nhan Duong, Tien D Bui, and Khoa Luu. Clusformer: A transformer based clustering approach to unsupervised large-scale face and visual landmark recognition. 2", + "[27] Xuan-Bac Nguyen, Guee Sang Lee, Soo Hyung Kim, and Hyung Jeong Yang. Self-supervised learning based on spatial awareness for medical image analysis. IEEE Access, 8:162973-162981, 2020. 2", + "[28] Xuan Nie, Madhumita A Takalkar, Mengyang Duan, Haimin Zhang, and Min Xu. GEME: Dual-stream multi-task gender-based micro-expression recognition. Neurocomputing, 427:13-28, 2021. 7", + "[29] Tae-Hyun Oh, Ronnachai Jaroensri, Changil Kim, Mohamed Elgharib, Fredo Durand, William T. Freeman, and Wojciech Matusik. Learning-based video motion magnification, 2018. 1, 2, 7", + "[30] Kha Gia Quach, Ngan Le, Chi Nhan Duong, Ibsa Jalata, Kaushik Roy, and Khoa Luu. Non-volume preserving-based fusion to group-level emotion recognition on crowd videos. Pattern Recognition, 128:108646, 2022. 2", + "[31] Ankith Jain Rakesh Kumar and Bir Bhanu. Micro-expression classification based on landmark relations with graph attention convolutional network. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1511-1520, 2021. 2", + "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[33] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 2", + "[34] John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su-Jing Wang. Megc 2019 – the second facial microexpressions grand challenge. In 2019 14th IEEE International Conference on Automatic Face Gesture Recognition (FG 2019), pages 1–5, 2019. 6", + "[35] B. Song, K. Li, Y. Zong, J. Zhu, W. Zheng, J. Shi, and L. Zhao. Recognizing spontaneous micro-expression using a three-stream convolutional neural network. IEEE Access, 7:184537-184551, 2019. 6, 7", + "[36] Bo Sun, Siming Cao, Dongliang Li, Jun He, and Lejun Yu. Dynamic micro-expression recognition using knowledge distillation. IEEE Transactions on Affective Computing, 2020. 7", + "[37] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 7", + "[38] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers amp; distillation through attention. In International Conference on Machine Learning, volume 139, pages 10347-10357, July 2021. 2" + ], + "bbox": [ + 78, + 90, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Thuong-Khanh Tran, Quang-Nhat Vo, Xiaopeng Hong, Xiaobai Li, and Guoying Zhao. Micro-expression spotting: A new benchmark. Neurocomputing, 443:356-368, 2021. 2", + "[40] Thanh-Dat Truong, Quoc-Huy Bui, Chi Nhan Duong, Han-Seok Seo, Son Lam Phung, Xin Li, and Khoa Luu. Direcformer: A directed attention in transformer approach to robust action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20030-20040, June 2022. 2", + "[41] Thanh-Dat Truong, Chi Nhan Duong, The De Vu, Hoang Anh Pham, Bhiksha Raj, Ngan Le, and Khoa Luu. The right to talk: An audio-visual transformer approach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1105–1114, October 2021. 2", + "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[43] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale micro-expression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2", + "[44] Yan Wang, Yikun Huang, Can Liu, Xiaoying Gu, Dandan Yang, Shuopeng Wang, and Bo Zhang. Micro expression recognition via dual-stream spatiotemporal attention network. Journal of Healthcare Engineering, 2021, 2021. 7", + "[45] Yandan Wang, John See, Yee-Hui Oh, Raphael C.-W. Phan, Yogachandran Rahulamathavan, Huo-Chong Ling, Su-Wei Tan, and Xujie Li. Effective recognition of facial microexpressions with video motion magnification. Multimedia Tools and Applications, 76(20):21665-21690, 2016. 2", + "[46] Mengting Wei, Wenming Zheng, Yuan Zong, Xingxun Jiang, Cheng Lu, and Jiateng Liu. A novel micro-expression recognition approach using attention-based magnification-adaptive networks. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2420-2424. IEEE, 2022. 7", + "[47] Bin Xia and Shangfei Wang. Micro-expression recognition enhanced by macro-expression from spatial-temporal domain. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 1186-1193, 2021. 6, 7, 8", + "[48] Bin Xia, Weikang Wang, Shangfei Wang, and Enhong Chen. Learning from macro-expression: a micro-expression recognition framework. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2936-2944, 2020. 7, 8", + "[49] Zhaoqiang Xia, Wei Peng, Huai-Qian Khor, Xiaoyi Feng, and Guoying Zhao. Revealing the invisible with model and data shrinking for composite-database micro-expression recognition. IEEE Transactions on Image Processing, 29:8590-8605, 2020. 6, 7", + "[50] Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu. Casme ii: An improved spontaneous micro-expression database and the baseline evaluation. PLOS ONE, 9(1):1-8, 01 2014. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1491", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Wen-Jing Yan, Qi Wu, Yong-Jin Liu, Su-Jing Wang, and Xiaolan Fu. Casme database: A dataset of spontaneous microexpressions collected from neutralized faces. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-7, 2013. 1", + "[52] Jianhui Yu, Chaoyi Zhang, Yang Song, and Weidong Cai. ICE-GAN: Identity-aware and capsule-enhanced gan for micro-expression recognition and synthesis. arXiv preprint arXiv:2005.04370, 2020.8", + "[53] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition. arXiv preprint arXiv:2101.04838, 2021. 8", + "[54] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition. Pattern Recognition, 122:108275, 2022. 7", + "[55] L. Zhou, Q. Mao, and L. Xue. Dual-inception network for cross-database micro-expression recognition. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-5. IEEE, 2019. 8" + ], + "bbox": [ + 78, + 90, + 468, + 387 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1492", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_model.json b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..df9d6c4ce898f3de2f20c589033e5d2b770b7a4e --- /dev/null +++ b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "Micron-BERT: BERT-based Facial Micro-Expression Recognition" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.163, + 0.845, + 0.181 + ], + "angle": 0, + "content": "Xuan-Bac Nguyen1, Chi Nhan Duong2, Xin Li3, Susan Gauch1, Han-Seok Seo4, Khoa Luu1" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.181, + 0.792, + 0.198 + ], + "angle": 0, + "content": "\\(^{1}\\) CVIU Lab, University of Arkansas, USA \\(^{2}\\) Concordia University, Canada" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.199, + 0.837, + 0.216 + ], + "angle": 0, + "content": "3 West Virginia University, USA 4 Dep. of Food Science, University of Arkansas, USA" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.218, + 0.86, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\) {xnguyen, sgauch, hanseok, khoaluu}@uark.edu, \\(^{2}\\) {dcnhan@ieee.org}, \\(^{3}\\) xin.li@mail.wvu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.302, + 0.474, + 0.755 + ], + "angle": 0, + "content": "Micro-expression recognition is one of the most challenging topics in affective computing. It aims to recognize tiny facial movements difficult for humans to perceive in a brief period, i.e., 0.25 to 0.5 seconds. Recent advances in pre-training deep Bidirectional Transformers (BERT) have significantly improved self-supervised learning tasks in computer vision. However, the standard BERT in vision problems is designed to learn only from full images or videos, and the architecture cannot accurately detect details of facial micro-expressions. This paper presents Micron-BERT (\\(\\mu\\)-BERT), a novel approach to facial micro-expression recognition. The proposed method can automatically capture these movements in an unsupervised manner based on two key ideas. First, we employ Diagonal Micro-Attention (DMA) to detect tiny differences between two frames. Second, we introduce a new Patch of Interest (PoI) module to localize and highlight micro-expression interest regions and simultaneously reduce noisy backgrounds and distractions. By incorporating these components into an end-to-end deep network, the proposed \\(\\mu\\)-BERT significantly outperforms all previous work in various micro-expression tasks. \\(\\mu\\)-BERT can be trained on a large-scale unlabeled dataset, i.e., up to 8 million images, and achieves high accuracy on new unseen facial micro-expression datasets. Empirical experiments show \\(\\mu\\)-BERT consistently outperforms state-of-the-art performance on four micro-expression benchmarks, including SAMM, CASME II, SMIC, and CASME3, by significant margins. Code will be available at https://github.com/uark-cviu/Micron-BERT" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.768, + 0.21, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Facial expressions are a complex mixture of conscious reactions directed toward given stimuli. They involve experiential, behavioral, and physiological elements. Because they are crucial to understanding human reactions, this topic has been widely studied in various application domains [5]. In general, facial expression problems can be classified into two main categories, macro-expression, and" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.269, + 0.882, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.473, + 0.895, + 0.515 + ], + "angle": 0, + "content": "Figure 1. Given two frames from a high-speed video, the proposed \\(\\mu\\)-BERT method can localize and highlight the regions of micromovements. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.893, + 0.703 + ], + "angle": 0, + "content": "micro-expression. The main differences between the two are facial expression intensities, and duration [2]. In particular, macro-expressions happen spontaneously, cover large movement areas in a given face, e.g., mouth, eyes, cheeks, and typically last from 0.5 to 4 seconds. Humans can usually recognize these expressions. By contrast, micro-expressions are involuntary occurrences, have low intensity, and last between 5 milliseconds and half a second. Indeed, micro-expressions are challenging to identify and are mostly detectable only by experts. Micro-expression understanding is essential in numerous applications, primarily lie detection, which is crucial in criminal analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Micro-expression identification requires both semantics and micro-movement analysis. Since they are difficult to observe through human eyes, a high-speed camera, usually with 200 frames per second (FPS) [6, 15, 51], is typically used to capture the required video frames. Previous work [11] tried to understand this micro information using MagNet [29] to amplify small motions between two frames, e.g., onset and apex frames. However, these methods still have limitations in terms of accuracy and robustness. In summary, the contributions of this work are four-fold:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.856, + 0.895, + 0.902 + ], + "angle": 0, + "content": "- A novel Facial Micro-expression Recognition (MER) via Pre-training of Deep Bidirectional Transformers approach (Micron-BERT or \\(\\mu\\)-BERT) is presented to" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1482" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.136 + ], + "angle": 0, + "content": "tackle the problem in a self-supervised learning manner. The proposed method aims to identify and localize micro-movements in faces accurately." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.143, + 0.468, + 0.216 + ], + "angle": 0, + "content": "- As detecting the tiny moment changes in faces is an essential input to the MER module, a new Diagonal Micro Attention (DMA) mechanism is proposed to precisely identify small movements in faces between two consecutive video frames." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.23, + 0.468, + 0.305 + ], + "angle": 0, + "content": "- A new Patch of Interest (POI) module is introduced to efficiently spot facial regions containing the microexpressions. Far apart from prior methods, it is trained in an unsupervised manner without using any facial labels, such as facial bounding boxes or landmarks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.311, + 0.468, + 0.431 + ], + "angle": 0, + "content": "- The proposed \\(\\mu\\)-BERT framework is designed in a self-supervised learning manner and trained in an end-to-end deep network. Indeed, it consistently achieves State-of-the-Art (SOTA) results in various standard micro-expression benchmarks, including CASME II [50], CASME3 [14], SAMM [6] and SMIC [15]. It achieves high recognition accuracy on new unseen subjects of various gender, age, and ethnicity." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.143, + 0.468, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.447, + 0.218, + 0.462 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.473, + 0.468, + 0.517 + ], + "angle": 0, + "content": "Generally, prior studies in micro-expression can be divided into two categories, including micro-expression spotting (MES) and micro-expression recognition." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.519, + 0.468, + 0.639 + ], + "angle": 0, + "content": "Micro-Expression Spotting (MES). The goal of MES is to determine the specific instant during which a micro-expression occurs. Li et al. [16] adopted a spatial-channel attention network to detect micro-expression action units. Tran et al. [39] attempted to standardize with the SMIC-E database and an evaluation protocol. MESNet [43] introduced a CNN-based approach with a \\((2 + 1)\\mathrm{D}\\) convolutional network, a clip proposal, and a classifier." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.641, + 0.468, + 0.82 + ], + "angle": 0, + "content": "Micro-Expression Recognition (MER). The goal of MER tasks is to classify the facial micro-expressions in a video. Ling et al. [11] present a new way of learning facial graph representations, allowing these small movements to be seen. Kumar and Bhanu [31] exploited connections between landmark points and their optical flow patch and achieved improvements over state-of-the-art (SOTA) methods for both the CASME II and SAMM. Liu et al. [21] presented a new method using transfer learning achieved an accuracy of \\(84.27\\%\\) on a composite of three datasets. Wang et al. [45] presented an Eulerian-motion magnification-based approach that highlights these small movements." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.822, + 0.468, + 0.912 + ], + "angle": 0, + "content": "Other Work. Other research, while not necessarily on MES or MER, is relevant to our approach. An advance in video motion magnification is shown in [29], outperforming the SOTA methods in multiple areas. This learning-based model can extract filters from data directly rather than rely on ones designed by hand, like the state-of-the-art method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.659, + 0.105 + ], + "angle": 0, + "content": "3. BERT Revisited" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.116, + 0.733, + 0.131 + ], + "angle": 0, + "content": "3.1. BERT in Vision Problems" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Transformers and deep learning have significantly improved results for many tasks in computer vision [1, 7, 9, 22, 23, 26, 27, 30, 40, 41]. Worth mentioning is Vision Transformer (ViT) [7], one of the first research efforts at the intersection of Transformers and computer vision. Unlike the traditional CNN network, ViT splits an image into a sequence of patches and applies the Transformers-based framework directly. Inspired by the success of BERT in Natural Language Processing (NLP), Bidirectional Encoder representation from Image Transformers (BEiT) [1] is presented as a self-supervised learning framework in computer vision. In particular, image patches are tokenized using DALL-E [32] to the visual tokens. These tokens are then randomly masked before feeding into the transformer backbone. The training objective is to recover the original visual tokens from the corrupted patches. These methods [1, 38] have marked a remarkable improvement compared to supervised learning methods by leveraging large-scale unlabelled datasets, e.g., ImageNet-1K, ImagNet-21K [33], to discover semantic information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.845, + 0.47 + ], + "angle": 0, + "content": "3.2. Limitations of BERT in Vision Problems" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.782 + ], + "angle": 0, + "content": "One limitation of using BERT in vision problems is the tokenization step. In the NLP field, a token has precisely one word mapped into it. In vision problems, however, many possible images or patches can share the same token as long as they have the same content. Therefore, designing a BERT model to mask a token and train a prediction model in the missing contexts in computer vision is more challenging than NLP. In addition, the tokenizer, i.e., DALL-E [32], is not robust enough to map similar contexts to a token. It yields noise in the tokenization process and affects the overall training performance. He et al., [9] presented a Masked Auto Encoder (MAE) that utilizes the BERT framework. Instead of tokenizing images, it eliminates patches of an image via a random masking strategy and reconstructs the context of these masked patches to the original content. Although this method can avoid using the tokenizer, it only considers the context inside an image. Thus, it does not apply to micro-expression, which requires understanding semantic information from consecutive video frames. In this paper, \\(\\mu\\)-BERT is presented to address these limitations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.799, + 0.803, + 0.816 + ], + "angle": 0, + "content": "4. The Proposed \\(\\mu\\)-BERT Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\(\\mu\\)-BERT is designed to model micro-changes of facial texture across temporal dimensions, which is hard to observe by unaided human eyes via a reconstruction process. The proposed \\(\\mu\\)-BERT architecture, shown in Figure 2, consists of five main blocks: a \\(\\mu\\)-Encoder, Patch" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1483" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.852, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.196, + 0.29, + 0.772, + 0.305 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed \\(\\mu\\)-BERT approach to facial micro-expression recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.474, + 0.56 + ], + "angle": 0, + "content": "of Interest (PoI), Blockwise Swapping, Diagonal Micro Attention (DMA), and a \\(\\mu\\)-Decoder. Given input images \\(I_{t}\\) and \\(I_{t + \\delta}\\), the role of the \\(\\mu\\)-Encoder is to represent \\(I_{t}\\) and \\(I_{t + \\delta}\\) into latent vectors. Then, Patch of Interest (PoI) constrains \\(\\mu\\)-BERT to look into facial regions containing microexpressions rather than unrelated regions such as the background. Blockwise Swapping and Diagonal Micro Attention (DMA) allow the model to focus on facial regions that primarily consist of micro differences between frames. Finally, \\(\\mu\\)-Decoder reconstructs the output signal back to the determined one. Compared to prior works, \\(\\mu\\)-BERT can adaptively focus on changes in facial regions while ignoring the ones in the background and effectively recognizes micro-expressions even when face movements occur. Moreover, \\(\\mu\\)-BERT can also alleviate the dependency on the accuracy of alignment approaches in pre-processing step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.427, + 0.584 + ], + "angle": 0, + "content": "4.1. Non-overlapping Patches Representation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.59, + 0.47, + 0.623 + ], + "angle": 0, + "content": "In \\(\\mu\\)-BERT, an input frame \\(I_{t} \\in \\mathbb{R}^{H \\times W \\times C}\\) is divided into a set of several non-overlapping patches \\(\\mathcal{P}_t\\) as Eqn. (1)." + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.633, + 0.469, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {t} = \\left\\{p _ {t} ^ {i} \\right\\} _ {i = 0} ^ {N _ {p} - 1} \\quad \\left| \\mathcal {P} _ {t} \\right| = H W / \\left(p s ^ {2}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.656, + 0.469, + 0.702 + ], + "angle": 0, + "content": "where \\( H, W, C \\) are the height, width, and number of channels, respectively. Each patch \\( p_t^i \\) has a resolution of \\( ps \\times ps \\). In our experiments, \\( H = W = 224 \\), \\( C = 3 \\), and \\( ps = 8 \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.711, + 0.194, + 0.726 + ], + "angle": 0, + "content": "4.2. \\(\\mu\\)-Encoder" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.734, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Each patch \\( p_i \\in \\mathcal{P}_t \\) is linearly projected into a latent vector of dimension \\( d \\) denoted as \\( \\mathbf{z}_t^i \\in \\mathbb{R}^{1 \\times d} \\), with additive fixed positional encoding [42]. Then, an image \\( I_t \\) can be represented as in Eqn. (2)." + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.804, + 0.469, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {t} = \\operatorname {c o n c a t} \\left[ \\mathbf {z} _ {t} ^ {0}, \\mathbf {z} _ {t} ^ {1}, \\dots \\mathbf {z} _ {t} ^ {N _ {p} - 1} \\right] \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.831, + 0.259, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} ^ {i} = \\alpha \\left(p _ {t} ^ {i}\\right) + \\mathbf {e} (i)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\mathbf{e}\\) are the projection embedding network and positional embedding, respectively. Let \\(\\mu\\)-Encoder, denoted as \\(\\mathcal{E}\\), be a stack of continuous blocks. Each block consists" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.313, + 0.821, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Figure 3. Builing block of Encoder and Decoder. Each block includes Multi-Head Attention (MHA) and Layer Normalization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.5 + ], + "angle": 0, + "content": "of alternating layers of Multi Head Attention (MHA) and Multi-Layer Perceptron (MLP), as illustrated in Figure 3. The Layer Norm (LN) is employed to the input signal before feeding to MHA and MLP layers, as in Eqn. (3)." + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.505, + 0.793, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} _ {l} = \\mathbf {x} _ {l - 1} + \\operatorname {M H A} (\\operatorname {L N} (\\mathbf {x} _ {l - 1}))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.523, + 0.891, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {l} = \\mathbf {x} ^ {\\prime} _ {l} + \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\mathbf {x} ^ {\\prime} _ {l}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.542, + 0.735, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {0} = \\mathbf {Z} _ {t}, 1 \\leq l \\leq L _ {e}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.563, + 0.89, + 0.593 + ], + "angle": 0, + "content": "where \\(L_{e}\\) is the number of blocks in \\(\\mathcal{E}\\). Given \\(\\mathbf{Z}_t\\), the output latent vector \\(\\mathbf{P}_{\\mathbf{t}}\\) is represented as in Eqn. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.598, + 0.891, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {t} = \\mathcal {E} \\left(\\mathbf {Z} _ {t}\\right) \\quad \\mathbf {P} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.622, + 0.617, + 0.637 + ], + "angle": 0, + "content": "4.3. \\(\\mu\\)-Decoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.646, + 0.892, + 0.707 + ], + "angle": 0, + "content": "The proposed auto-encoder is designed symmetrically. It means that the decoder part denoted as \\(\\mathcal{D}\\), has a similar architecture to the encoder \\(\\mathcal{E}\\). Given a latent vector \\(\\mathbf{P}_t\\), the decoded signal \\(\\mathbf{Q}_t\\) is represented as in Eqn. (5)." + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.711, + 0.891, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {t} = \\mathcal {D} \\left(\\mathbf {P} _ {t}\\right) \\quad \\mathbf {Q} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.891, + 0.767 + ], + "angle": 0, + "content": "We add one more Linear layer to interpolate \\(\\mathbf{Q}_t\\) to an intermediate signal \\(\\mathbf{y}_t\\) before reshaping it into the image size." + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.787, + 0.891, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\xrightarrow {\\text {l i n e a r}} \\mathbf {y} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times p s \\times p s \\times C} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.811, + 0.834, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times p s \\times p s \\times C} \\xrightarrow {\\text {r e s h a p e}} \\mathbf {y} _ {t} ^ {\\prime} \\in \\mathbb {R} ^ {H \\times W \\times C}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.839, + 0.695, + 0.855 + ], + "angle": 0, + "content": "4.4. Blockwise Swapping" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.862, + 0.879, + 0.877 + ], + "angle": 0, + "content": "Given two frames \\(I_{t}\\) and \\(I_{t + \\delta}\\), we realize the fact that:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.883, + 0.891, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\lim _ {\\delta \\rightarrow 0} s \\left(p _ {t} ^ {i}, p _ {t + \\delta} ^ {i}\\right) = 1 \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1484" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.088, + 0.44, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.226, + 0.473, + 0.285 + ], + "angle": 0, + "content": "Figure 4. Blockwise Swapping. For each triplet, we present the \\( I_{t} \\) (left), the \\( I_{t + \\delta} \\) (middle) and the \\( I_{t / s} \\) (right). The yellow blocks in \\( I_{t / s} \\) represent swapped patches from \\( I_{t + \\delta} \\) that are randomly swapped into \\( I_{t} \\). Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.474, + 0.542 + ], + "angle": 0, + "content": "where \\( p_t^i \\) is the \\( i^{th} \\)-patch at frame \\( t \\). \\( s \\) denotes a function to measure the similarity between \\( p_t^i \\) and \\( p_{t + \\delta}^i \\) where a higher score indicates higher similarity and \\( 0 \\leq s(p_t^i, p_{t + \\delta}^i) \\leq 1 \\). Given a patch correlation as in Eqn. (7), we propose a Blockwise Swapping mechanism to (1) firstly randomly swap two corresponding blocks \\( p_t^i \\) and \\( p_{t + \\delta}^i \\) between two frames to create a swapped image \\( I_{t / s} \\), and then (2) enforce the model to spot these changes and reconstruct \\( I_t \\) from \\( I_{t / s} \\). By doing so, the model is further strengthened in recognizing and restoring the swapped patches. As a result, the learned model can be enhanced by the capability to notice small differences between frames. Moreover, as shown in Eqn. (7), shorter time \\( \\delta \\) causing larger similarity between \\( I_t \\) from \\( I_{t / s} \\) can further help to enhance the robustness on spotting these differences. The detail of this strategy is described in Algorithm 1 and Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.373, + 0.567 + ], + "angle": 0, + "content": "4.5. Diagonal Micro Attention (DMA)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.473, + 0.905 + ], + "angle": 0, + "content": "As a result of Blockwise Swapping, the image patches \\(\\mathcal{P}_{t / s}\\) from \\(I_{t / s}\\) consists of two types, i.e. \\(p_{t / s}^{j}\\) from \\(\\mathcal{P}_t\\) of \\(I_{t}\\) and \\(p_{t / s}^{i}\\) from \\(\\mathcal{P}_{t + \\delta}\\) of \\(I_{t + \\delta}\\). Then, the next stage is to learn how to reconstruct \\(\\mathcal{P}_t\\) from \\(\\mathcal{P}_{t / s}\\). Since \\(p_{t / s}^{i}\\) includes all changes between \\(I_{t}\\) and \\(I_{t / s}\\), more emphasis is placed on \\(p_{t / s}^{i}\\) during reconstruction process. Theoretically, the ground truth of the index of \\(p_{t / s}^{i}\\) in \\(\\mathcal{P}_{t / s}\\) can be utilized to enforce the model focusing on these swapped patches. However, adopting this information may reduce the learning capability to spot these microchanges. Therefore, a novel attention mechanism named Diagonal MicroAttention (DMA) is presented to enforce the network automatically focusing on swapped patches \\(p_{t / s}^{i}\\) and equip it with the ability to precisely spot and identify all changes between images. Notice that these changes may include patches in the background. The following section introduces a solution to constrain the learned network focusing on only meaningful facial regions. The details of DMA are presented in Figure 5. Formally, we construct an attention map \\(\\hat{A}\\) between \\(\\mathcal{P}_{t + \\delta}\\) and \\(\\mathcal{P}_{t / s}\\) where the diag \\((\\hat{A})\\) illustrates correlations between two corresponding patches \\(p_{t + \\delta}^{i}\\)" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.883, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.209, + 0.895, + 0.253 + ], + "angle": 0, + "content": "Figure 5. Diagonal Micro-Attention (DMA) module. Diagonal values from the attention map between \\(\\mathbf{P}_{t / s}\\) and \\(\\mathbf{P}_{t + \\delta}\\) are used to rank the importance of each patch in the swapped image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.895, + 0.331 + ], + "angle": 0, + "content": "and \\( p_{t / s}^{i} \\). From the observation that \\( \\hat{A}(i, i) > \\hat{A}(j, j) \\) for all \\( p_{t / s}^{i} \\in \\mathcal{P}_{t + \\delta} \\) and \\( p_{t / s}^{j} \\in \\mathcal{P}_{t} \\), \\( \\text{diag}(\\hat{A}) \\) can be effectively adopted as weights indicating important features. Full operations of DMA are presented in Eqn. (8) and Eqn. (9)." + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.35, + 0.893, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\hat {A} = \\operatorname {s o f t m a x} \\left(Q \\left(\\mathbf {P} _ {t + \\delta}\\right) \\otimes K \\left(\\mathbf {P} _ {t / s}\\right) ^ {T}\\right), \\sum_ {j = 0} ^ {N _ {p}} \\hat {A} (i, j) = 1 \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.416, + 0.893, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {d m a} = \\operatorname {d i a g} (\\hat {A}) \\times V \\left(\\mathbf {P} _ {t / s}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.444, + 0.892, + 0.46 + ], + "angle": 0, + "content": "where \\(\\times\\) denotes the Element-wise multiplication operator." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.715, + 0.485 + ], + "angle": 0, + "content": "4.6. Patch of Interest (POI)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.614 + ], + "angle": 0, + "content": "In Section 4.5, Diagonal Micro-Attention has been introduced to weigh the importance of swapped patches automatically. These swapped patches are randomly produced via Blockwise Swapping, as in Algorithm 1. In theory, the ideal case is when all swapped patches are located within the facial region only so that the deep network can learn the micro-movements from the facial parts solely and not be distracted by the background. In practice, however, we" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.637, + 0.733, + 0.654 + ], + "angle": 0, + "content": "Algorithm 1 Blockwise Swapping" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.656, + 0.894, + 0.698 + ], + "angle": 0, + "content": "Input: \\(\\mathcal{P}_t, \\mathcal{P}_{t + \\delta}\\) image patches (\\(N_p = h \\times w\\)); \\(r_s\\): swapping ratio (default: 0.5); min_bs: minimum block size (default: 16); min_ar: minimum aspect ratio (default: 0.3)" + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.699, + 0.747, + 0.712 + ], + "angle": 0, + "content": "Output: Swapped image patches \\(\\mathcal{P}_{t / s}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.713, + 0.662, + 0.727 + ], + "angle": 0, + "content": "\\(\\mathcal{P}_{t / s}\\gets \\mathcal{P}_t;N_p\\gets |\\mathcal{P}_t|\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.558, + 0.738 + ], + "angle": 0, + "content": "\\(c\\gets 0\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.74, + 0.654, + 0.754 + ], + "angle": 0, + "content": "while \\(c\\leq r_s\\times N_p\\) do" + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.755, + 0.733, + 0.768 + ], + "angle": 0, + "content": "\\(bs\\gets \\mathsf{rnd}(min\\_ bs,r_s\\times N_p - c)\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.769, + 0.724, + 0.782 + ], + "angle": 0, + "content": "ar←rnd(min_ar,1/min_ar)" + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.783, + 0.699, + 0.796 + ], + "angle": 0, + "content": "\\(m,n\\gets \\sqrt{bs}\\cdot ar,\\sqrt{bs / ar}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.797, + 0.76, + 0.81 + ], + "angle": 0, + "content": "\\(p,q\\gets \\mathsf{rnd}(0,h - m),\\mathsf{rnd}(0,w - n)\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.811, + 0.731, + 0.824 + ], + "angle": 0, + "content": "\\(\\forall i \\in [p, p + m), j \\in [q, q + n)\\) :" + }, + { + "type": "ref_text", + "bbox": [ + 0.564, + 0.825, + 0.657, + 0.837 + ], + "angle": 0, + "content": "\\(k\\gets i\\times w + j\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.564, + 0.838, + 0.687, + 0.851 + ], + "angle": 0, + "content": "\\(\\mathcal{P}_{t / s}(k)\\gets \\mathcal{P}_{t + \\delta}(k)\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.853, + 0.632, + 0.863 + ], + "angle": 0, + "content": "\\(c\\gets c + m\\times n\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.865, + 0.58, + 0.877 + ], + "angle": 0, + "content": "end while" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.879, + 0.591, + 0.892 + ], + "angle": 0, + "content": "return \\(\\mathcal{P}_{t / s}\\)" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.656, + 0.894, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1485" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.089, + 0.379, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.215, + 0.47, + 0.259 + ], + "angle": 0, + "content": "Figure 6. Patch of Interest (POI) module. The \\(\\mathbf{P}_t\\) and \\(\\mathbf{P}_{t + \\delta /c}\\) are sequence of patch features of \\(I_{t + \\delta}\\) and its random cropped version. \\(p_{t + \\delta}^{CT}\\) and \\(p_{t + \\delta /c}^{CT}\\) are their corresponding contextual features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.47, + 0.448 + ], + "angle": 0, + "content": "can only identify which parts are selected in the Blockwise Swapping algorithm if the facial regions are available. Thus, the Patch of Interest (POI) is introduced to automatically explore the salient regions and ignore the background patches in an image. Apart from prior methods, the proposed POI leverages the characteristic of self-attention and can be achieved through self-learning without facial labels, such as facial bounding boxes or segmentation masks. The idea of the POI module is illustrated in Figure 6. Thanks to POI, a capability of automatically focusing on facial regions is further equipped to the learned model, making it more robust against facial movements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.471, + 0.632 + ], + "angle": 0, + "content": "The POI relies on the contextual agreement between the frame \\( I_{t + \\delta} \\) and \\( \\mathrm{Crop}(I_{t + \\delta}) \\). Motivated by the BERT framework, we add a Contextual Token \\( z^{CT} \\) to the beginning of the sequence of patches, as in Eqn. (2), to learn the contextual information in the image. The deeper this token passes through the Transformer blocks, the more information is accumulated from \\( z_{t}^{i} \\in \\mathcal{P}_{t} \\). As a result, \\( z^{CT} \\) becomes a placeholder to store the information extracted from other patches in the sequence and present the contextual information of the image. Let \\( p_{t + \\delta}^{CT} \\) and \\( p_{t + \\delta / c}^{CT} \\) be the contextual features of frame \\( I_{t + \\delta} \\) and its cropped version \\( \\mathrm{Crop}(I_{t + \\delta}) \\) respectively. The agreement loss is then defined as in Eqn. (10)." + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.645, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a g g} = H \\left(p _ {t + \\delta} ^ {C T}, p _ {t + \\delta / c} ^ {C T}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.668, + 0.469, + 0.731 + ], + "angle": 0, + "content": "where \\(H\\) is the function that enforces \\(p_{t + \\delta}^{CT}\\) to be similar to \\(p_{t + \\delta / crop}^{CT}\\) so that the model can discover the salient patches. The POI can be extracted from the attention map \\(A\\) at the last attention layer of encoder \\(\\mathcal{E}\\). In particular, we measure:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.755, + 0.469, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {t + \\delta} = A [ 0, \\cdot ] = \\left[ s _ {t + \\delta} ^ {0}, s _ {t + \\delta} ^ {1}, \\dots s _ {t + \\delta} ^ {N _ {p} - 1} \\right] \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.79, + 0.469, + 0.839 + ], + "angle": 0, + "content": "where \\(\\sum_{i=0}^{N_p-1} s_{t+\\delta}^i = 1\\). The higher the score \\(s_{t+\\delta}^i\\), the richer the patch contains contextual information. Now, Eqn. (9) can be reformulated as in Eqn. (12)." + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.851, + 0.352, + 0.884 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {W} = \\operatorname {d i a g} (\\hat {A}) \\times S _ {t + \\delta} \\\\ \\mathbf {P} _ {d m a} = \\mathbf {W} \\times V (\\mathbf {P} _ {t / s}) \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.653, + 0.105 + ], + "angle": 0, + "content": "4.7. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.115, + 0.892, + 0.145 + ], + "angle": 0, + "content": "The proposed \\(\\mu\\)-BERT deep network is optimized using the proposed loss function as in Eqn. (13)." + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.156, + 0.892, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\gamma \\times \\mathcal {L} _ {r} + \\beta \\times \\mathcal {L} _ {a g g} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.182, + 0.794, + 0.196 + ], + "angle": 0, + "content": "where \\(\\gamma\\) and \\(\\beta\\) are the weights for each loss." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.197, + 0.892, + 0.241 + ], + "angle": 0, + "content": "Reconstruction Loss. The output of the decoder \\(\\mathbf{y}'_t\\) is reconstructed to the original image \\(I_t\\) using the Mean Square Error (MSE) function." + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.252, + 0.892, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\operatorname {M S E} \\left(\\mathbf {y} ^ {\\prime} _ {t}, I _ {t}\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.278, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Contextual Agreement Loss. MSE is also used to enforce the similarity of contextual features of \\( I_{t + \\delta / crop} \\) and \\( I_{t + \\delta} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.319, + 0.892, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a g g} = \\operatorname {M S E} \\left(p _ {t + \\delta} ^ {C T}, p _ {t + \\delta / c r o p} ^ {C T}\\right) \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.35, + 0.706, + 0.367 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.712, + 0.39 + ], + "angle": 0, + "content": "5.1. Datasets and Protocols" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.459 + ], + "angle": 0, + "content": "CASME II [50]. With a 200 fps sampling rate and a facial resolution of \\(280 \\times 340\\), CASME II provides 247 microexpression samples from 26 subjects of the same ethnicity. Labels include apex frames, action units, and emotions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.519 + ], + "angle": 0, + "content": "SAMM [6]. Also, using a 200 fps frame rate and a facial resolution of \\(400 \\times 400\\), SAMM consists of 159 samples from 32 participants and 13 ethnicities. The samples all have emotions, apex frames, and action unit labels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.892, + 0.58 + ], + "angle": 0, + "content": "SMIC [15]. SIMC is made up of 164 samples. Lacking apex frame and action unit labels, the samples span 16 participants of 3 ethnicities. The recordings are taken with a resolution of \\(640 \\times 480\\) at 100 fps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.58, + 0.892, + 0.64 + ], + "angle": 0, + "content": "CASME3 [14]. Officially known as \\(\\mathrm{CAS(ME)^3}\\) provides 1,109 labeled micro-expressions and 3,490 labeled macro-expressions. This dataset has roughly 80 hours of footage with a resolution of \\(1280 \\times 720\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.78, + 0.668 + ], + "angle": 0, + "content": "5.2. Micro-Expression Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We use all raw frames from CASME3 for self-training except frames of test set. It is important to note that we do not use labels or meta information such as onset, offset, and apex index frames nor labeled emotions. In total, we construct an unlabelled dataset of 8M frames. The images are resized to \\(224 \\times 224\\). Then, each image is divided into patches of \\(8 \\times 8\\), yielding \\(N_{p} = 784\\) patches. The temporal index \\(\\delta\\) is selected randomly between a lower bound of 5 and an upper bound of 11, experimentally. The swapping ratio \\(r_{s}\\) is selected as \\(50\\%\\) of the number of patches being swapped from \\(I_{t + \\delta}\\) to \\(I_{t}\\). Each patch is projected to a latent space of \\(d = 512\\) dimensions before being fed into the encoder and decoder. For the encoder and decoder, we keep the same \\(d\\) for all vectors and similar configurations, i.e., \\(L_{e} = L_{d} = 4\\). \\(\\mu\\)-BERT is implemented easily in" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1486" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.18, + 0.089, + 0.788, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.464, + 0.895, + 0.494 + ], + "angle": 0, + "content": "Figure 7. We demonstrate how \\(\\mu\\)-BERT perceives the tiny differences between two frames. The first two rows are onset and apex input frames. The third and fourth rows are the results of RAFT and MagNet, respectively. The rest of the rows are our \\(\\mu\\)-BERT results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.475, + 0.593 + ], + "angle": 0, + "content": "Pytorch framework and trained by \\(32 \\times\\) A100 GPUs (40G each). The learning rate is set to 0.0001 initially and then reduced to zero gradually under ConsineLinear [24] policy. The batch size is set to 64/GPU. The model is optimized by AdamW [25] for 100 epochs. The training is completed within three days." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.602, + 0.348, + 0.618 + ], + "angle": 0, + "content": "5.3. Micro-Expression Recognition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.471, + 0.822 + ], + "angle": 0, + "content": "We leverage the pretrained \\(\\mu\\)-BERT as an initial weight and take the encoder \\(\\mathcal{E}\\) and DMA module of \\(\\mu\\)-BERT as the MER backbone. The input of MER is the onset and apex frames which correspond to \\(I_{t}\\) and \\(I_{t + \\delta}\\) respectively. In Eqn. (8), \\(\\mathbf{P}_{dma}\\) are the features representing the micro changes and movements between onset and apex frames. They can be effectively adopted for recognizing microexpressions. We adopt the standard metrics and protocols of MER2019 challenge [34] with the unweighted F1 score \\(UF1 = \\frac{1}{C} \\sum_{i=0}^{C-1} \\frac{2 \\times TP_i}{TP_i + FP_i + FN_i}\\) and accuracy \\(UAR = \\frac{1}{C} \\sum_{i=0}^{C-1} \\frac{TP_i}{N_i}\\), where \\(C\\) is the number of MEs, \\(N_i\\) is the total number of \\(i^{th}\\) ME in the dataset. Leave-one-out cross-validation (LOOCV) scheme is used for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.172, + 0.847 + ], + "angle": 0, + "content": "5.4. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our proposed \\(\\mu\\)-BERT shows a significant improvement over prior methods and baselines, as shown in Table 1 on the CASME3. Tested using 3, 4, and 7 emotion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.895, + 0.652 + ], + "angle": 0, + "content": "classes, \\(\\mu\\)-BERT achieves double-digit gains over the compared methods in each category. In the case of 3 emotion classes, \\(\\mu\\)-BERT achieved a \\(56.04\\%\\) UF1 score and \\(61.25\\%\\) UAR, compared to RCN-A's [49] \\(39.28\\%\\) UF1 and \\(38.93\\%\\) UAR. For 4 emotion classes, \\(\\mu\\)-BERT outperforms Baseline (+Depth) [14] \\(47.18\\%\\) to \\(30.01\\%\\) for UF1 and \\(49.13\\%\\) to \\(29.82\\%\\) for UAR. Large gains over Baseline (+Depth) [14] are seen in the case of 7 emotion classes, where \\(\\mu\\)-BERT attains UF1 and UAR scores of \\(32.64\\%\\) and \\(32.54\\%\\) respectively, compared to \\(17.73\\%\\) and \\(18.29\\%\\) for the baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.895, + 0.882 + ], + "angle": 0, + "content": "Table 2 details results for CASMEII. \\(\\mu\\)-BERT shows improvements over all other methods. For three categories, it achieves a UF1 of \\(90.34\\%\\) and UAR of \\(89.14\\%\\), representing \\(3.37\\%\\) and \\(0.86\\%\\) increases over the prior leading method (OFF-ApexNet [8]), respectively. Similar improvement is seen in five categories: a \\(4.83\\%\\) over TSCNN [35] in terms of UF1 and a \\(0.89\\%\\) increase over SMA-STN [19] for UAR. Similarly, \\(\\mu\\)-BERT performs competitively with other methods on the SAMM as seen in Table 3. Using 5 emotion classes, \\(\\mu\\)-BERT outperforms MinMaNet [47] by a large margin in terms of UF1 (\\(83.86\\%\\) vs \\(76.40\\%\\)) and UAR (\\(84.75\\%\\) vs \\(76.70\\%\\)), respectively. The performance of \\(\\mu\\)-BERT on SMIC is compared against several others in Table 4. \\(\\mu\\)-BERT outperforms others with a \\(7.5\\%\\) increase in UF1 to \\(85.5\\%\\) and a \\(3.97\\%\\) boost in UAR to \\(83.84\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "On the composite dataset, \\(\\mu\\)-BERT again outperforms" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1487" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.09, + 0.393, + 0.103 + ], + "angle": 0, + "content": "Table 1. MER on the CASME3 dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.11, + 0.467, + 0.32 + ], + "angle": 0, + "content": "
Method# ClassesUF1 (%)UAR(%)
FR [54]334.9334.13
STSTNet [18]337.9537.92
RCN-A [49]339.2838.93
μ-BERT (ours)356.0461.25
Baseline [14]429.1529.10
Baseline (+Depth) [14]430.0129.82
μ-BERT (ours)447.1849.13
Baseline [14]717.5918.01
Baseline(+Depth) [14]717.7318.29
μ-BERT (ours)732.6432.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.328, + 0.384, + 0.341 + ], + "angle": 0, + "content": "Table 2. MER on CASME II dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.348, + 0.466, + 0.63 + ], + "angle": 0, + "content": "
Method# ClassesUF1 (%)UAR (%)
LR-GACNN [10]570.9081.30
AMAN [46]571.0075.40
Graph-TCN [13]572.4673.98
DSTAN [44]573.0075.00
GEME [28]573.5475.20
MiMaNet [47]575.9079.90
SMA-STN [19]579.4682.59
TSCNN [35]580.7080.97
μ-BERT (ours)585.5383.48
STSTNet [17]383.8286.86
OFF-ApexNet [8]386.9788.28
MAE [9]388.0387.28
μ-BERT (ours)390.3489.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.172, + 0.632, + 0.375, + 0.646 + ], + "angle": 0, + "content": "Table 3. MER on SAMM dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.653, + 0.466, + 0.82 + ], + "angle": 0, + "content": "
Method# ClassesUF1 (%)UAR (%)
AMAN [46]567.0068.85
SMA-STN [19]570.3377.20
GRAPH-AU [12]570.4574.26
MTMNet [48]573.6074.10
MiMaNet [47]576.4076.70
MAE [9]580.4088.98
μ-BERT (ours)583.8684.75
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "other methods (Table 5). Attaining a UF1 score of \\(89.03\\%\\) and UAR of \\(88.42\\%\\), \\(\\mu\\)-BERT realizes \\(0.73\\%\\), and \\(0.82\\%\\) gains over previous best MiMaNet [47], respectively. Table 6 shows the impact of DMA and POI on CASME3. Our" + }, + { + "type": "table_caption", + "bbox": [ + 0.599, + 0.09, + 0.794, + 0.103 + ], + "angle": 0, + "content": "Table 4. MER on SMIC dataset." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.11, + 0.89, + 0.3 + ], + "angle": 0, + "content": "
Method# ClassesUF1 (%)UAR (%)
DIKD [36]371.0076.06
TSCNN [35]372.3672.74
MTMNet [48]374.4076.00
AMAN [46]377.0079.87
MiMaNet [47]377.8078.60
DSTAN [44]378.0077.00
MAE [9]381.8680.82
μ-BERT (ours)385.5083.84
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.893, + 0.402 + ], + "angle": 0, + "content": "method gives more modest gains of approximately \\(2\\%\\) in both metrics. A greater improvement is seen with DMA, where UF1 and UAR increase by another \\(2 - 4\\%\\). Significant improvement from \\(\\mu\\)-BERT is seen when adopting both modules, with a UF1 of \\(32.64\\%\\) and UAR of \\(32.54\\%\\), representing roughly \\(10\\%\\) gains over previous methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.41, + 0.857, + 0.426 + ], + "angle": 0, + "content": "5.5. How \\(\\mu\\)-BERT perceives micro-movements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To understand the micro-movements between two frames, the onset and apex frames are inputs for \\(\\mu\\)-BERT. These frames represent the moments that the microexpression starts and is observed. We measure \\(\\mathrm{diag}(\\hat{A})\\) (Subsection 4.5) and \\(\\mathbf{S}_{t + \\delta}\\) (Eqn (11)) values to identify which regions contain small movements between two frames. Comparisons of \\(\\mu\\)-BERT with RAFT [37], i.e., optical flow-based method and MagNet [29] are also conducted as in Fig. 7. The third and fourth columns in Fig 7 show the results of RAFT [37], and MagNet [29] on spotting the micro-movements, respectively. While RAFT is an optical flow-based method, MagNet amplifies small differences between the two frames. These methods are sensitive to the environment (e.g., lighting, illuminations). Thus, noises in the background still exist in their outputs. In addition, neither RAFT nor MagNet understand semantic information in the frame and distinguish changes inside facial or background regions. Meanwhile, \\(\\mu\\)-BERT shows its advantages in perceiving micro-movements via distinguishing the facial regions and spotting the micro-expressions. In particular, the attention map in the fifth column, in Fig. 7 illustrates the micro-differences between onset and apex frames. The higher contrast represents the higher chance of small movements in these regions. With the POI module, \\(\\mu\\)-BERT can automatically figure out the informative patches and ignore the background ones. Then, with DMA module, \\(\\mu\\)-BERT, can detect and localize which corresponding patches/regions contain tiny movements. As shown in the seventh column, attention maps represent the most salient regions in the image. By empowering DMA and POI, \\(\\mu\\)-BERT effectively identifies micro-movements within facial" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1488" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.09, + 0.437, + 0.104 + ], + "angle": 0, + "content": "Table 5. MER on the Composite dataset (MECG2019)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.111, + 0.467, + 0.328 + ], + "angle": 0, + "content": "
Method# ClassesUF1 (%)UAR(%)
Dual-Inception [55]373.2272.78
FR [53]378.3878.32
NMER [20]378.8578.24
GRAPH-AU [12]379.1479.33
ICE-GAN [52]384.5084.10
BDCNN [3]385.0985.00
moment [48]386.4085.70
MiMaNet [47]388.3087.60
MAE [9]388.5087.40
μ-BERT (ours)389.0388.42
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.343, + 0.368, + 0.358 + ], + "angle": 0, + "content": "regions, as demonstrated in the last column." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.365, + 0.24, + 0.38 + ], + "angle": 0, + "content": "5.6. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.468, + 0.506 + ], + "angle": 0, + "content": "This section compares \\(\\mu\\)-BERT against other self-supervised learning (SSL) methods on the MER task. CASME3 is used for experiments since it has many unlabeled images to demonstrate the power of SSL methods. We also analyze the essential contributions of Diagonal Micro-Attention (DMA) and Patch of Interest (POI) modules. Finally, we illustrate the robustness of \\(\\mu\\)-BERT pretrained on CASME3 on unseen datasets and domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Comparisons with self-supervised learning methods. We utilize the encoder and decoder parts of \\(\\mu\\)-BERT (without DMA and POI) to train previous SSL methods (MoCo V3 [4], BEIT [1], MAE [9]) and then continue learning the MER task on the large-scale database CASME3 [14]. Overall results are shown in Table 6. It is expected that ViT-S achieves the lowest performance for UF1 and UAR as ImageNet and Micro-Expression are two different domains. Three self-supervised methods (MoCo V3, BEIT, and MAE) got better results when they were pretrained on CASME before fine-tuning to the recognition task. Compared to ViT-S, these SSL methods gain remarkable performance. Especially, MAE [9] achieves \\(3.5\\%\\) and \\(2\\%\\) up on UF1 and UAR compared to ViT-S, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.468, + 0.856 + ], + "angle": 0, + "content": "The role of Blockwise Swapping. Our basic setup of \\(\\mu\\)-BERT (denoted as MB1) is employed to train in an SSL manner. It is noted that only Blockwise Swapping is involved, and it does not contain either DMA or POI. Compared to MAE, MB1 outperforms MAE by \\(2\\%\\) in both UF1 and UAR, approximately. The reasons are: (1) Blockwise Swapping enforces the model to learn local context features inside an image, i.e., \\(I_{t}\\), and (2) It helps the network to figure out micro-disparities between two frames \\(I_{t}\\) and \\(I_{t + \\delta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The role of DMA. This module is the guide to tell the network where to look and which patches to focus. By doing so, the \\(\\mu\\)-BERT gets more robust knowledge of micro" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 6. MER performance on CASME3 by different self-supervised methods and various settings of \\( \\mu \\) -BERT" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.123, + 0.892, + 0.264 + ], + "angle": 0, + "content": "
MethodPre-trainDMAPOIUF1UAR
ViT-S [7]ImageNetXX20.3418.76
MoCo V3 - R50 [4]CASME3XX19.1217.36
MoCo V3 - R101 [4]CASME3XX20.1418.52
MoCo V3 [4]CASME3XX22.1319.34
BEIT [1]CASME3XX23.5419.89
MAE [9]CASME3XX23.8620.87
μ-BERT (MB1)CASME3XX25.2722.96
μ-BERT (MB2)CASME3✓X27.3526.18
μ-BERT (MB3)CASME3✓✓32.6432.54
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.322 + ], + "angle": 0, + "content": "movements between two frames. For this reason, the network (denoted as MB2) achieves \\(2\\%\\) on UF1 and a significant \\(4\\%\\) gain on UAR compared to MB1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.489 + ], + "angle": 0, + "content": "The role of POI. Since MB1 are sensitive to background noise, the micro-disparities features \\(\\mathbf{P}_{dma}\\) might contain unwanted features coming from the background. The POI is designed as a filter that only lets the typical interesting patches belonging to the subject go through and preserves the micro-movement features only. The improvements of up to \\(6\\%\\) compared to MB2 demonstrate the important role of POI in \\(\\mu\\)-BERT for micro-expression tasks. Qualitative results demonstrated in Supplementary Material can further emphasize the advantages of POI in assisting the network to be robust against facial movements." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.494, + 0.765, + 0.51 + ], + "angle": 0, + "content": "6. Conclusions and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Unlike a few concurrent research on micro-expression, we move forward and study how to explore BERT pretraining for this problem. In our proposed \\(\\mu\\)-BERT, we presented a novel Diagonal Micro Attention (DMA) to learn the micro-movements of the subject across frames. The Patch of Interest (POI) module is proposed to guide the network, focusing on the most salient parts, i.e., facial regions, and ignoring the noisy sensitivities from the background. Empowered by the simple design of \\(\\mu\\)-BERT, SOTA performance on micro-expression recognition tasks is achieved in four benchmark datasets. Our perspective will inspire more future study efforts in this direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Limitations. We demonstrated the efficiency of the POI module in removing noise in the background, which is sensitive to lighting and illumination. However, suppose any facial parts, e.g., the forehead, are affected by lighting conditions while there are no movements. In that case, these noisy factors might also be included as micro-difference features. The robustness with different lighting conditions will be left as our future works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work is supported by Arkansas Biosciences Institute (ABI) Grant, NSF WVAR-CRESH and NSF Data Science, Data Analytics that are Robust and Trusted (DART). We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1489" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.469, + 0.215 + ], + "angle": 0, + "content": "[2] Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu. Video-based facial micro-expression analysis: A survey of datasets, features and algorithms. IEEE transactions on pattern analysis and machine intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[3] Bin Chen, Kun-Hong Liu, Yong Xu, Qing-Qiang Wu, and Jun-Feng Yao. Block division convolutional network with implicit deep features augmentation for micro-expression recognition. IEEE Transactions on Multimedia, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[4] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[5] Ciprian Adrian Coraneanu, Marc Oliu Simón, Jeffrey F. Cohn, and Sergio Escalera Guerrero. Survey on rgb, 3d, thermal, and multimodal approaches for facial expression recognition: History, trends, and affect-related applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(8):1548-1568, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[6] Adrian K. Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap. Samm: A spontaneous micro-facial movement dataset. IEEE Transactions on Affective Computing, 9(1):116-129, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[8] Y.S. Gan, S.T. Liong, W.C. Yau, Y.C. Huang, and L.K. Tan. Off-apexnet on micro-expression recognition system. Signal Processing: Image Communication, 74:129-139, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[10] Ankith Jain Rakesh Kumar and Bir Bhanu. Micro-expression classification based on landmark relations with graph attention convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1511–1520, June 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[11] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression recognition based on facial graph representation learning and facial action unit fusion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571-1580, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[12] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression recognition based on facial graph representation learning and facial action unit fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1571-1580, June 2021. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Ling Lei, Jianfeng Li, Tong Chen, and Shigang Li. A novel Graph-TCN with a graph structured representation for micro-expression recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2237–2245, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[14] Jingting Li, Zizhao Dong, Shaoyuan Lu, Su-Jing Wang, Wen-Jing Yan, Yinhuan Ma, Ye Liu, Changbing Huang, and Xiaolan Fu. Cas(me)sup3/sup: A third generation facial spontaneous micro-expression database with depth information and high ecological validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[15] Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietikainen. A spontaneous micro-expression database: Inducement, collection and baseline. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6, 2013. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[16] Yante Li, Xiaohua Huang, and Guoying Zhao. Microexpression action unit detection with spatial and channel attention. Neurocomputing, 436:221-231, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[17] S.T. Liong, Y.S. Gan, J. See, H. Khor, and Y. Huang. Shallow triple stream three-dimensional cnn (ststnet) for microexpression recognition. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-5. IEEE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[18] Sze-Teng Liong, Yee Siang Gan, John See, Huai-Qian Khor, and Yen-Chang Huang. Shallow triple stream three-dimensional cnn (ststnet) for micro-expression recognition. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-5. IEEE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[19] Jiateng Liu, Wenming Zheng, and Yuan Zong. SMA-STN: Segmented movement-attending spatiotemporal network for micro-expression recognition. arXiv preprint arXiv:2010.09342, 2020.6.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[20] Yuchi Liu, Heming Du, Liang Zheng, and Tom Gedeon. A neural micro-expression recognizer. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-4. IEEE, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[21] Yanju Liu, Yange Li, Xinhai Yi, Zuojin Hu, Huiyu Zhang, and Yanzhong Liu. Lightweight vit model for microexpression recognition enhanced by transfer learning. Frontiers in Neurorobotics, 16, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[22] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1490" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.123, + 0.468, + 0.178 + ], + "angle": 0, + "content": "[26] Xuan-Bac Nguyen, Duc Toan Bui, Chi Nhan Duong, Tien D Bui, and Khoa Luu. Clusformer: A transformer based clustering approach to unsupervised large-scale face and visual landmark recognition. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.181, + 0.468, + 0.235 + ], + "angle": 0, + "content": "[27] Xuan-Bac Nguyen, Guee Sang Lee, Soo Hyung Kim, and Hyung Jeong Yang. Self-supervised learning based on spatial awareness for medical image analysis. IEEE Access, 8:162973-162981, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.239, + 0.468, + 0.293 + ], + "angle": 0, + "content": "[28] Xuan Nie, Madhumita A Takalkar, Mengyang Duan, Haimin Zhang, and Min Xu. GEME: Dual-stream multi-task gender-based micro-expression recognition. Neurocomputing, 427:13-28, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.297, + 0.468, + 0.352 + ], + "angle": 0, + "content": "[29] Tae-Hyun Oh, Ronnachai Jaroensri, Changil Kim, Mohamed Elgharib, Fredo Durand, William T. Freeman, and Wojciech Matusik. Learning-based video motion magnification, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.355, + 0.468, + 0.41 + ], + "angle": 0, + "content": "[30] Kha Gia Quach, Ngan Le, Chi Nhan Duong, Ibsa Jalata, Kaushik Roy, and Khoa Luu. Non-volume preserving-based fusion to group-level emotion recognition on crowd videos. Pattern Recognition, 128:108646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.413, + 0.468, + 0.481 + ], + "angle": 0, + "content": "[31] Ankith Jain Rakesh Kumar and Bir Bhanu. Micro-expression classification based on landmark relations with graph attention convolutional network. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1511-1520, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.484, + 0.468, + 0.552 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.556, + 0.468, + 0.598 + ], + "angle": 0, + "content": "[33] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.601, + 0.468, + 0.669 + ], + "angle": 0, + "content": "[34] John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su-Jing Wang. Megc 2019 – the second facial microexpressions grand challenge. In 2019 14th IEEE International Conference on Automatic Face Gesture Recognition (FG 2019), pages 1–5, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.672, + 0.468, + 0.727 + ], + "angle": 0, + "content": "[35] B. Song, K. Li, Y. Zong, J. Zhu, W. Zheng, J. Shi, and L. Zhao. Recognizing spontaneous micro-expression using a three-stream convolutional neural network. IEEE Access, 7:184537-184551, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.73, + 0.468, + 0.784 + ], + "angle": 0, + "content": "[36] Bo Sun, Siming Cao, Dongliang Li, Jun He, and Lejun Yu. Dynamic micro-expression recognition using knowledge distillation. IEEE Transactions on Affective Computing, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.787, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[37] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[38] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers amp; distillation through attention. In International Conference on Machine Learning, volume 139, pages 10347-10357, July 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[39] Thuong-Khanh Tran, Quang-Nhat Vo, Xiaopeng Hong, Xiaobai Li, and Guoying Zhao. Micro-expression spotting: A new benchmark. Neurocomputing, 443:356-368, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[40] Thanh-Dat Truong, Quoc-Huy Bui, Chi Nhan Duong, Han-Seok Seo, Son Lam Phung, Xin Li, and Khoa Luu. Direcformer: A directed attention in transformer approach to robust action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20030-20040, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[41] Thanh-Dat Truong, Chi Nhan Duong, The De Vu, Hoang Anh Pham, Bhiksha Raj, Ngan Le, and Khoa Luu. The right to talk: An audio-visual transformer approach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1105–1114, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[43] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale micro-expression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[44] Yan Wang, Yikun Huang, Can Liu, Xiaoying Gu, Dandan Yang, Shuopeng Wang, and Bo Zhang. Micro expression recognition via dual-stream spatiotemporal attention network. Journal of Healthcare Engineering, 2021, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[45] Yandan Wang, John See, Yee-Hui Oh, Raphael C.-W. Phan, Yogachandran Rahulamathavan, Huo-Chong Ling, Su-Wei Tan, and Xujie Li. Effective recognition of facial microexpressions with video motion magnification. Multimedia Tools and Applications, 76(20):21665-21690, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.549, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[46] Mengting Wei, Wenming Zheng, Yuan Zong, Xingxun Jiang, Cheng Lu, and Jiateng Liu. A novel micro-expression recognition approach using attention-based magnification-adaptive networks. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2420-2424. IEEE, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[47] Bin Xia and Shangfei Wang. Micro-expression recognition enhanced by macro-expression from spatial-temporal domain. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 1186-1193, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[48] Bin Xia, Weikang Wang, Shangfei Wang, and Enhong Chen. Learning from macro-expression: a micro-expression recognition framework. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2936-2944, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[49] Zhaoqiang Xia, Wei Peng, Huai-Qian Khor, Xiaoyi Feng, and Guoying Zhao. Revealing the invisible with model and data shrinking for composite-database micro-expression recognition. IEEE Transactions on Image Processing, 29:8590-8605, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu. Casme ii: An improved spontaneous micro-expression database and the baseline evaluation. PLOS ONE, 9(1):1-8, 01 2014. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1491" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[51] Wen-Jing Yan, Qi Wu, Yong-Jin Liu, Su-Jing Wang, and Xiaolan Fu. Casme database: A dataset of spontaneous microexpressions collected from neutralized faces. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-7, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[52] Jianhui Yu, Chaoyi Zhang, Yang Song, and Weidong Cai. ICE-GAN: Identity-aware and capsule-enhanced gan for micro-expression recognition and synthesis. arXiv preprint arXiv:2005.04370, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[53] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition. arXiv preprint arXiv:2101.04838, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.331 + ], + "angle": 0, + "content": "[54] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition. Pattern Recognition, 122:108275, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[55] L. Zhou, Q. Mao, and L. Xue. Dual-inception network for cross-database micro-expression recognition. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-5. IEEE, 2019. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1492" + } + ] +] \ No newline at end of file diff --git a/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_origin.pdf b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cb06912c9d1ce52a36eac0a0d63dddedc63109cf --- /dev/null +++ b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/16369e30-f89d-439b-a563-ebb4fc8aacfa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e278210ddcbae9c1bb369a4a18b2b1b746afd6e1baae5228294186a9751ab7f8 +size 8940038 diff --git a/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/full.md b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f04116056d5df725bed5ded09e1a8d064c81e202 --- /dev/null +++ b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/full.md @@ -0,0 +1,378 @@ +# Micron-BERT: BERT-based Facial Micro-Expression Recognition + +Xuan-Bac Nguyen1, Chi Nhan Duong2, Xin Li3, Susan Gauch1, Han-Seok Seo4, Khoa Luu1 + +$^{1}$ CVIU Lab, University of Arkansas, USA $^{2}$ Concordia University, Canada + +3 West Virginia University, USA 4 Dep. of Food Science, University of Arkansas, USA + +$^{1}$ {xnguyen, sgauch, hanseok, khoaluu}@uark.edu, $^{2}$ {dcnhan@ieee.org}, $^{3}$ xin.li@mail.wvu.edu + +# Abstract + +Micro-expression recognition is one of the most challenging topics in affective computing. It aims to recognize tiny facial movements difficult for humans to perceive in a brief period, i.e., 0.25 to 0.5 seconds. Recent advances in pre-training deep Bidirectional Transformers (BERT) have significantly improved self-supervised learning tasks in computer vision. However, the standard BERT in vision problems is designed to learn only from full images or videos, and the architecture cannot accurately detect details of facial micro-expressions. This paper presents Micron-BERT ( $\mu$ -BERT), a novel approach to facial micro-expression recognition. The proposed method can automatically capture these movements in an unsupervised manner based on two key ideas. First, we employ Diagonal Micro-Attention (DMA) to detect tiny differences between two frames. Second, we introduce a new Patch of Interest (PoI) module to localize and highlight micro-expression interest regions and simultaneously reduce noisy backgrounds and distractions. By incorporating these components into an end-to-end deep network, the proposed $\mu$ -BERT significantly outperforms all previous work in various micro-expression tasks. $\mu$ -BERT can be trained on a large-scale unlabeled dataset, i.e., up to 8 million images, and achieves high accuracy on new unseen facial micro-expression datasets. Empirical experiments show $\mu$ -BERT consistently outperforms state-of-the-art performance on four micro-expression benchmarks, including SAMM, CASME II, SMIC, and CASME3, by significant margins. Code will be available at https://github.com/uark-cviu/Micron-BERT + +# 1. Introduction + +Facial expressions are a complex mixture of conscious reactions directed toward given stimuli. They involve experiential, behavioral, and physiological elements. Because they are crucial to understanding human reactions, this topic has been widely studied in various application domains [5]. In general, facial expression problems can be classified into two main categories, macro-expression, and + +![](images/ad7894b0dc5aadb9f880d68016148b89febed11c0d74cd3d7b49092d87acef88.jpg) +Figure 1. Given two frames from a high-speed video, the proposed $\mu$ -BERT method can localize and highlight the regions of micromovements. Best viewed in color. + +micro-expression. The main differences between the two are facial expression intensities, and duration [2]. In particular, macro-expressions happen spontaneously, cover large movement areas in a given face, e.g., mouth, eyes, cheeks, and typically last from 0.5 to 4 seconds. Humans can usually recognize these expressions. By contrast, micro-expressions are involuntary occurrences, have low intensity, and last between 5 milliseconds and half a second. Indeed, micro-expressions are challenging to identify and are mostly detectable only by experts. Micro-expression understanding is essential in numerous applications, primarily lie detection, which is crucial in criminal analysis. + +Micro-expression identification requires both semantics and micro-movement analysis. Since they are difficult to observe through human eyes, a high-speed camera, usually with 200 frames per second (FPS) [6, 15, 51], is typically used to capture the required video frames. Previous work [11] tried to understand this micro information using MagNet [29] to amplify small motions between two frames, e.g., onset and apex frames. However, these methods still have limitations in terms of accuracy and robustness. In summary, the contributions of this work are four-fold: + +- A novel Facial Micro-expression Recognition (MER) via Pre-training of Deep Bidirectional Transformers approach (Micron-BERT or $\mu$ -BERT) is presented to + +tackle the problem in a self-supervised learning manner. The proposed method aims to identify and localize micro-movements in faces accurately. + +- As detecting the tiny moment changes in faces is an essential input to the MER module, a new Diagonal Micro Attention (DMA) mechanism is proposed to precisely identify small movements in faces between two consecutive video frames. +- A new Patch of Interest (POI) module is introduced to efficiently spot facial regions containing the microexpressions. Far apart from prior methods, it is trained in an unsupervised manner without using any facial labels, such as facial bounding boxes or landmarks. +- The proposed $\mu$ -BERT framework is designed in a self-supervised learning manner and trained in an end-to-end deep network. Indeed, it consistently achieves State-of-the-Art (SOTA) results in various standard micro-expression benchmarks, including CASME II [50], CASME3 [14], SAMM [6] and SMIC [15]. It achieves high recognition accuracy on new unseen subjects of various gender, age, and ethnicity. + +# 2. Related Work + +Generally, prior studies in micro-expression can be divided into two categories, including micro-expression spotting (MES) and micro-expression recognition. + +Micro-Expression Spotting (MES). The goal of MES is to determine the specific instant during which a micro-expression occurs. Li et al. [16] adopted a spatial-channel attention network to detect micro-expression action units. Tran et al. [39] attempted to standardize with the SMIC-E database and an evaluation protocol. MESNet [43] introduced a CNN-based approach with a $(2 + 1)\mathrm{D}$ convolutional network, a clip proposal, and a classifier. + +Micro-Expression Recognition (MER). The goal of MER tasks is to classify the facial micro-expressions in a video. Ling et al. [11] present a new way of learning facial graph representations, allowing these small movements to be seen. Kumar and Bhanu [31] exploited connections between landmark points and their optical flow patch and achieved improvements over state-of-the-art (SOTA) methods for both the CASME II and SAMM. Liu et al. [21] presented a new method using transfer learning achieved an accuracy of $84.27\%$ on a composite of three datasets. Wang et al. [45] presented an Eulerian-motion magnification-based approach that highlights these small movements. + +Other Work. Other research, while not necessarily on MES or MER, is relevant to our approach. An advance in video motion magnification is shown in [29], outperforming the SOTA methods in multiple areas. This learning-based model can extract filters from data directly rather than rely on ones designed by hand, like the state-of-the-art method. + +# 3. BERT Revisited + +# 3.1. BERT in Vision Problems + +Transformers and deep learning have significantly improved results for many tasks in computer vision [1, 7, 9, 22, 23, 26, 27, 30, 40, 41]. Worth mentioning is Vision Transformer (ViT) [7], one of the first research efforts at the intersection of Transformers and computer vision. Unlike the traditional CNN network, ViT splits an image into a sequence of patches and applies the Transformers-based framework directly. Inspired by the success of BERT in Natural Language Processing (NLP), Bidirectional Encoder representation from Image Transformers (BEiT) [1] is presented as a self-supervised learning framework in computer vision. In particular, image patches are tokenized using DALL-E [32] to the visual tokens. These tokens are then randomly masked before feeding into the transformer backbone. The training objective is to recover the original visual tokens from the corrupted patches. These methods [1, 38] have marked a remarkable improvement compared to supervised learning methods by leveraging large-scale unlabelled datasets, e.g., ImageNet-1K, ImagNet-21K [33], to discover semantic information. + +# 3.2. Limitations of BERT in Vision Problems + +One limitation of using BERT in vision problems is the tokenization step. In the NLP field, a token has precisely one word mapped into it. In vision problems, however, many possible images or patches can share the same token as long as they have the same content. Therefore, designing a BERT model to mask a token and train a prediction model in the missing contexts in computer vision is more challenging than NLP. In addition, the tokenizer, i.e., DALL-E [32], is not robust enough to map similar contexts to a token. It yields noise in the tokenization process and affects the overall training performance. He et al., [9] presented a Masked Auto Encoder (MAE) that utilizes the BERT framework. Instead of tokenizing images, it eliminates patches of an image via a random masking strategy and reconstructs the context of these masked patches to the original content. Although this method can avoid using the tokenizer, it only considers the context inside an image. Thus, it does not apply to micro-expression, which requires understanding semantic information from consecutive video frames. In this paper, $\mu$ -BERT is presented to address these limitations. + +# 4. The Proposed $\mu$ -BERT Approach + +$\mu$ -BERT is designed to model micro-changes of facial texture across temporal dimensions, which is hard to observe by unaided human eyes via a reconstruction process. The proposed $\mu$ -BERT architecture, shown in Figure 2, consists of five main blocks: a $\mu$ -Encoder, Patch + +![](images/2a96aa8f2d5c59122ab599131db679ec754bd921f014ce2a01b7bf27611f9fce.jpg) +Figure 2. An overview of the proposed $\mu$ -BERT approach to facial micro-expression recognition. + +of Interest (PoI), Blockwise Swapping, Diagonal Micro Attention (DMA), and a $\mu$ -Decoder. Given input images $I_{t}$ and $I_{t + \delta}$ , the role of the $\mu$ -Encoder is to represent $I_{t}$ and $I_{t + \delta}$ into latent vectors. Then, Patch of Interest (PoI) constrains $\mu$ -BERT to look into facial regions containing microexpressions rather than unrelated regions such as the background. Blockwise Swapping and Diagonal Micro Attention (DMA) allow the model to focus on facial regions that primarily consist of micro differences between frames. Finally, $\mu$ -Decoder reconstructs the output signal back to the determined one. Compared to prior works, $\mu$ -BERT can adaptively focus on changes in facial regions while ignoring the ones in the background and effectively recognizes micro-expressions even when face movements occur. Moreover, $\mu$ -BERT can also alleviate the dependency on the accuracy of alignment approaches in pre-processing step. + +# 4.1. Non-overlapping Patches Representation + +In $\mu$ -BERT, an input frame $I_{t} \in \mathbb{R}^{H \times W \times C}$ is divided into a set of several non-overlapping patches $\mathcal{P}_t$ as Eqn. (1). + +$$ +\mathcal {P} _ {t} = \left\{p _ {t} ^ {i} \right\} _ {i = 0} ^ {N _ {p} - 1} \quad \left| \mathcal {P} _ {t} \right| = H W / \left(p s ^ {2}\right) \tag {1} +$$ + +where $H, W, C$ are the height, width, and number of channels, respectively. Each patch $p_t^i$ has a resolution of $ps \times ps$ . In our experiments, $H = W = 224$ , $C = 3$ , and $ps = 8$ . + +# 4.2. $\mu$ -Encoder + +Each patch $p_i \in \mathcal{P}_t$ is linearly projected into a latent vector of dimension $d$ denoted as $\mathbf{z}_t^i \in \mathbb{R}^{1 \times d}$ , with additive fixed positional encoding [42]. Then, an image $I_t$ can be represented as in Eqn. (2). + +$$ +\mathbf {Z} _ {t} = \operatorname {c o n c a t} \left[ \mathbf {z} _ {t} ^ {0}, \mathbf {z} _ {t} ^ {1}, \dots \mathbf {z} _ {t} ^ {N _ {p} - 1} \right] \in \mathbb {R} ^ {N _ {p} \times d} \tag {2} +$$ + +$$ +\mathbf {z} _ {t} ^ {i} = \alpha \left(p _ {t} ^ {i}\right) + \mathbf {e} (i) +$$ + +where $\alpha$ and $\mathbf{e}$ are the projection embedding network and positional embedding, respectively. Let $\mu$ -Encoder, denoted as $\mathcal{E}$ , be a stack of continuous blocks. Each block consists + +![](images/1aa3c221ced05c8342dc33606299d97ff1be419bb0c70f3957dd271cdb0d336f.jpg) +Figure 3. Builing block of Encoder and Decoder. Each block includes Multi-Head Attention (MHA) and Layer Normalization. + +of alternating layers of Multi Head Attention (MHA) and Multi-Layer Perceptron (MLP), as illustrated in Figure 3. The Layer Norm (LN) is employed to the input signal before feeding to MHA and MLP layers, as in Eqn. (3). + +$$ +\mathbf {x} ^ {\prime} _ {l} = \mathbf {x} _ {l - 1} + \operatorname {M H A} (\operatorname {L N} (\mathbf {x} _ {l - 1})) +$$ + +$$ +\mathbf {x} _ {l} = \mathbf {x} ^ {\prime} _ {l} + \operatorname {M L P} \left(\operatorname {L N} \left(\mathbf {x} ^ {\prime} _ {l}\right)\right) \tag {3} +$$ + +$$ +\mathbf {x} _ {0} = \mathbf {Z} _ {t}, 1 \leq l \leq L _ {e} +$$ + +where $L_{e}$ is the number of blocks in $\mathcal{E}$ . Given $\mathbf{Z}_t$ , the output latent vector $\mathbf{P}_{\mathbf{t}}$ is represented as in Eqn. (4). + +$$ +\mathbf {P} _ {t} = \mathcal {E} \left(\mathbf {Z} _ {t}\right) \quad \mathbf {P} _ {t} \in \mathbb {R} ^ {N _ {p} \times d} \tag {4} +$$ + +# 4.3. $\mu$ -Decoder + +The proposed auto-encoder is designed symmetrically. It means that the decoder part denoted as $\mathcal{D}$ , has a similar architecture to the encoder $\mathcal{E}$ . Given a latent vector $\mathbf{P}_t$ , the decoded signal $\mathbf{Q}_t$ is represented as in Eqn. (5). + +$$ +\mathbf {Q} _ {t} = \mathcal {D} \left(\mathbf {P} _ {t}\right) \quad \mathbf {Q} _ {t} \in \mathbb {R} ^ {N _ {p} \times d} \tag {5} +$$ + +We add one more Linear layer to interpolate $\mathbf{Q}_t$ to an intermediate signal $\mathbf{y}_t$ before reshaping it into the image size. + +$$ +\mathbf {Q} _ {t} \in \mathbb {R} ^ {N _ {p} \times d} \xrightarrow {\text {l i n e a r}} \mathbf {y} _ {t} \in \mathbb {R} ^ {N _ {p} \times p s \times p s \times C} \tag {6} +$$ + +$$ +\mathbf {y} _ {t} \in \mathbb {R} ^ {N _ {p} \times p s \times p s \times C} \xrightarrow {\text {r e s h a p e}} \mathbf {y} _ {t} ^ {\prime} \in \mathbb {R} ^ {H \times W \times C} +$$ + +# 4.4. Blockwise Swapping + +Given two frames $I_{t}$ and $I_{t + \delta}$ , we realize the fact that: + +$$ +\lim _ {\delta \rightarrow 0} s \left(p _ {t} ^ {i}, p _ {t + \delta} ^ {i}\right) = 1 \tag {7} +$$ + +![](images/6ec2580c36031dbf28f7accbd1bfed8be715f7615dafd87088460c7d1285de8b.jpg) +Figure 4. Blockwise Swapping. For each triplet, we present the $I_{t}$ (left), the $I_{t + \delta}$ (middle) and the $I_{t / s}$ (right). The yellow blocks in $I_{t / s}$ represent swapped patches from $I_{t + \delta}$ that are randomly swapped into $I_{t}$ . Best viewed in color. + +where $p_t^i$ is the $i^{th}$ -patch at frame $t$ . $s$ denotes a function to measure the similarity between $p_t^i$ and $p_{t + \delta}^i$ where a higher score indicates higher similarity and $0 \leq s(p_t^i, p_{t + \delta}^i) \leq 1$ . Given a patch correlation as in Eqn. (7), we propose a Blockwise Swapping mechanism to (1) firstly randomly swap two corresponding blocks $p_t^i$ and $p_{t + \delta}^i$ between two frames to create a swapped image $I_{t / s}$ , and then (2) enforce the model to spot these changes and reconstruct $I_t$ from $I_{t / s}$ . By doing so, the model is further strengthened in recognizing and restoring the swapped patches. As a result, the learned model can be enhanced by the capability to notice small differences between frames. Moreover, as shown in Eqn. (7), shorter time $\delta$ causing larger similarity between $I_t$ from $I_{t / s}$ can further help to enhance the robustness on spotting these differences. The detail of this strategy is described in Algorithm 1 and Figure 4. + +# 4.5. Diagonal Micro Attention (DMA) + +As a result of Blockwise Swapping, the image patches $\mathcal{P}_{t / s}$ from $I_{t / s}$ consists of two types, i.e. $p_{t / s}^{j}$ from $\mathcal{P}_t$ of $I_{t}$ and $p_{t / s}^{i}$ from $\mathcal{P}_{t + \delta}$ of $I_{t + \delta}$ . Then, the next stage is to learn how to reconstruct $\mathcal{P}_t$ from $\mathcal{P}_{t / s}$ . Since $p_{t / s}^{i}$ includes all changes between $I_{t}$ and $I_{t / s}$ , more emphasis is placed on $p_{t / s}^{i}$ during reconstruction process. Theoretically, the ground truth of the index of $p_{t / s}^{i}$ in $\mathcal{P}_{t / s}$ can be utilized to enforce the model focusing on these swapped patches. However, adopting this information may reduce the learning capability to spot these microchanges. Therefore, a novel attention mechanism named Diagonal MicroAttention (DMA) is presented to enforce the network automatically focusing on swapped patches $p_{t / s}^{i}$ and equip it with the ability to precisely spot and identify all changes between images. Notice that these changes may include patches in the background. The following section introduces a solution to constrain the learned network focusing on only meaningful facial regions. The details of DMA are presented in Figure 5. Formally, we construct an attention map $\hat{A}$ between $\mathcal{P}_{t + \delta}$ and $\mathcal{P}_{t / s}$ where the diag $(\hat{A})$ illustrates correlations between two corresponding patches $p_{t + \delta}^{i}$ + +![](images/c2f42b4bb6f5535879a081ec2c47812701308136e2a1b03b3795f2d919f6d784.jpg) +Figure 5. Diagonal Micro-Attention (DMA) module. Diagonal values from the attention map between $\mathbf{P}_{t / s}$ and $\mathbf{P}_{t + \delta}$ are used to rank the importance of each patch in the swapped image. + +and $p_{t / s}^{i}$ . From the observation that $\hat{A}(i, i) > \hat{A}(j, j)$ for all $p_{t / s}^{i} \in \mathcal{P}_{t + \delta}$ and $p_{t / s}^{j} \in \mathcal{P}_{t}$ , $\text{diag}(\hat{A})$ can be effectively adopted as weights indicating important features. Full operations of DMA are presented in Eqn. (8) and Eqn. (9). + +$$ +\hat {A} = \operatorname {s o f t m a x} \left(Q \left(\mathbf {P} _ {t + \delta}\right) \otimes K \left(\mathbf {P} _ {t / s}\right) ^ {T}\right), \sum_ {j = 0} ^ {N _ {p}} \hat {A} (i, j) = 1 \tag {8} +$$ + +$$ +\mathbf {P} _ {d m a} = \operatorname {d i a g} (\hat {A}) \times V \left(\mathbf {P} _ {t / s}\right) \tag {9} +$$ + +where $\times$ denotes the Element-wise multiplication operator. + +# 4.6. Patch of Interest (POI) + +In Section 4.5, Diagonal Micro-Attention has been introduced to weigh the importance of swapped patches automatically. These swapped patches are randomly produced via Blockwise Swapping, as in Algorithm 1. In theory, the ideal case is when all swapped patches are located within the facial region only so that the deep network can learn the micro-movements from the facial parts solely and not be distracted by the background. In practice, however, we + +# Algorithm 1 Blockwise Swapping + +Input: $\mathcal{P}_t, \mathcal{P}_{t + \delta}$ image patches ( $N_p = h \times w$ ); $r_s$ : swapping ratio (default: 0.5); min_bs: minimum block size (default: 16); min_ar: minimum aspect ratio (default: 0.3) +Output: Swapped image patches $\mathcal{P}_{t / s}$ +$\mathcal{P}_{t / s}\gets \mathcal{P}_t;N_p\gets |\mathcal{P}_t|$ +$c\gets 0$ +while $c\leq r_s\times N_p$ do +$bs\gets \mathsf{rnd}(min\_ bs,r_s\times N_p - c)$ +ar←rnd(min_ar,1/min_ar) +$m,n\gets \sqrt{bs}\cdot ar,\sqrt{bs / ar}$ +$p,q\gets \mathsf{rnd}(0,h - m),\mathsf{rnd}(0,w - n)$ +$\forall i \in [p, p + m), j \in [q, q + n)$ : +$k\gets i\times w + j$ +$\mathcal{P}_{t / s}(k)\gets \mathcal{P}_{t + \delta}(k)$ +$c\gets c + m\times n$ +end while +return $\mathcal{P}_{t / s}$ + +![](images/8f5a4c6c365a07c269c3222fdd0d4939373702d463c4f899b3eee7a144e90857.jpg) +Figure 6. Patch of Interest (POI) module. The $\mathbf{P}_t$ and $\mathbf{P}_{t + \delta /c}$ are sequence of patch features of $I_{t + \delta}$ and its random cropped version. $p_{t + \delta}^{CT}$ and $p_{t + \delta /c}^{CT}$ are their corresponding contextual features. + +can only identify which parts are selected in the Blockwise Swapping algorithm if the facial regions are available. Thus, the Patch of Interest (POI) is introduced to automatically explore the salient regions and ignore the background patches in an image. Apart from prior methods, the proposed POI leverages the characteristic of self-attention and can be achieved through self-learning without facial labels, such as facial bounding boxes or segmentation masks. The idea of the POI module is illustrated in Figure 6. Thanks to POI, a capability of automatically focusing on facial regions is further equipped to the learned model, making it more robust against facial movements. + +The POI relies on the contextual agreement between the frame $I_{t + \delta}$ and $\mathrm{Crop}(I_{t + \delta})$ . Motivated by the BERT framework, we add a Contextual Token $z^{CT}$ to the beginning of the sequence of patches, as in Eqn. (2), to learn the contextual information in the image. The deeper this token passes through the Transformer blocks, the more information is accumulated from $z_{t}^{i} \in \mathcal{P}_{t}$ . As a result, $z^{CT}$ becomes a placeholder to store the information extracted from other patches in the sequence and present the contextual information of the image. Let $p_{t + \delta}^{CT}$ and $p_{t + \delta / c}^{CT}$ be the contextual features of frame $I_{t + \delta}$ and its cropped version $\mathrm{Crop}(I_{t + \delta})$ respectively. The agreement loss is then defined as in Eqn. (10). + +$$ +\mathcal {L} _ {a g g} = H \left(p _ {t + \delta} ^ {C T}, p _ {t + \delta / c} ^ {C T}\right) \tag {10} +$$ + +where $H$ is the function that enforces $p_{t + \delta}^{CT}$ to be similar to $p_{t + \delta / crop}^{CT}$ so that the model can discover the salient patches. The POI can be extracted from the attention map $A$ at the last attention layer of encoder $\mathcal{E}$ . In particular, we measure: + +$$ +\mathbf {S} _ {t + \delta} = A [ 0, \cdot ] = \left[ s _ {t + \delta} ^ {0}, s _ {t + \delta} ^ {1}, \dots s _ {t + \delta} ^ {N _ {p} - 1} \right] \tag {11} +$$ + +where $\sum_{i=0}^{N_p-1} s_{t+\delta}^i = 1$ . The higher the score $s_{t+\delta}^i$ , the richer the patch contains contextual information. Now, Eqn. (9) can be reformulated as in Eqn. (12). + +$$ +\begin{array}{l} \mathrm {W} = \operatorname {d i a g} (\hat {A}) \times S _ {t + \delta} \\ \mathbf {P} _ {d m a} = \mathbf {W} \times V (\mathbf {P} _ {t / s}) \\ \end{array} +$$ + +# 4.7. Loss Functions + +The proposed $\mu$ -BERT deep network is optimized using the proposed loss function as in Eqn. (13). + +$$ +\mathcal {L} = \gamma \times \mathcal {L} _ {r} + \beta \times \mathcal {L} _ {a g g} \tag {13} +$$ + +where $\gamma$ and $\beta$ are the weights for each loss. + +Reconstruction Loss. The output of the decoder $\mathbf{y}'_t$ is reconstructed to the original image $I_t$ using the Mean Square Error (MSE) function. + +$$ +\mathcal {L} _ {r} = \operatorname {M S E} \left(\mathbf {y} ^ {\prime} _ {t}, I _ {t}\right) \tag {14} +$$ + +Contextual Agreement Loss. MSE is also used to enforce the similarity of contextual features of $I_{t + \delta / crop}$ and $I_{t + \delta}$ + +$$ +\mathcal {L} _ {a g g} = \operatorname {M S E} \left(p _ {t + \delta} ^ {C T}, p _ {t + \delta / c r o p} ^ {C T}\right) \tag {15} +$$ + +# 5. Experimental Results + +# 5.1. Datasets and Protocols + +CASME II [50]. With a 200 fps sampling rate and a facial resolution of $280 \times 340$ , CASME II provides 247 microexpression samples from 26 subjects of the same ethnicity. Labels include apex frames, action units, and emotions. + +SAMM [6]. Also, using a 200 fps frame rate and a facial resolution of $400 \times 400$ , SAMM consists of 159 samples from 32 participants and 13 ethnicities. The samples all have emotions, apex frames, and action unit labels. + +SMIC [15]. SIMC is made up of 164 samples. Lacking apex frame and action unit labels, the samples span 16 participants of 3 ethnicities. The recordings are taken with a resolution of $640 \times 480$ at 100 fps. + +CASME3 [14]. Officially known as $\mathrm{CAS(ME)^3}$ provides 1,109 labeled micro-expressions and 3,490 labeled macro-expressions. This dataset has roughly 80 hours of footage with a resolution of $1280 \times 720$ . + +# 5.2. Micro-Expression Self-Training + +We use all raw frames from CASME3 for self-training except frames of test set. It is important to note that we do not use labels or meta information such as onset, offset, and apex index frames nor labeled emotions. In total, we construct an unlabelled dataset of 8M frames. The images are resized to $224 \times 224$ . Then, each image is divided into patches of $8 \times 8$ , yielding $N_{p} = 784$ patches. The temporal index $\delta$ is selected randomly between a lower bound of 5 and an upper bound of 11, experimentally. The swapping ratio $r_{s}$ is selected as $50\%$ of the number of patches being swapped from $I_{t + \delta}$ to $I_{t}$ . Each patch is projected to a latent space of $d = 512$ dimensions before being fed into the encoder and decoder. For the encoder and decoder, we keep the same $d$ for all vectors and similar configurations, i.e., $L_{e} = L_{d} = 4$ . $\mu$ -BERT is implemented easily in + +![](images/5ff512598e1c0412f6804a0aafa191e4d7017b0422cd50e8be31feec7e01c1ad.jpg) +Figure 7. We demonstrate how $\mu$ -BERT perceives the tiny differences between two frames. The first two rows are onset and apex input frames. The third and fourth rows are the results of RAFT and MagNet, respectively. The rest of the rows are our $\mu$ -BERT results. + +Pytorch framework and trained by $32 \times$ A100 GPUs (40G each). The learning rate is set to 0.0001 initially and then reduced to zero gradually under ConsineLinear [24] policy. The batch size is set to 64/GPU. The model is optimized by AdamW [25] for 100 epochs. The training is completed within three days. + +# 5.3. Micro-Expression Recognition + +We leverage the pretrained $\mu$ -BERT as an initial weight and take the encoder $\mathcal{E}$ and DMA module of $\mu$ -BERT as the MER backbone. The input of MER is the onset and apex frames which correspond to $I_{t}$ and $I_{t + \delta}$ respectively. In Eqn. (8), $\mathbf{P}_{dma}$ are the features representing the micro changes and movements between onset and apex frames. They can be effectively adopted for recognizing microexpressions. We adopt the standard metrics and protocols of MER2019 challenge [34] with the unweighted F1 score $UF1 = \frac{1}{C} \sum_{i=0}^{C-1} \frac{2 \times TP_i}{TP_i + FP_i + FN_i}$ and accuracy $UAR = \frac{1}{C} \sum_{i=0}^{C-1} \frac{TP_i}{N_i}$ , where $C$ is the number of MEs, $N_i$ is the total number of $i^{th}$ ME in the dataset. Leave-one-out cross-validation (LOOCV) scheme is used for evaluation. + +# 5.4. Results + +Our proposed $\mu$ -BERT shows a significant improvement over prior methods and baselines, as shown in Table 1 on the CASME3. Tested using 3, 4, and 7 emotion + +classes, $\mu$ -BERT achieves double-digit gains over the compared methods in each category. In the case of 3 emotion classes, $\mu$ -BERT achieved a $56.04\%$ UF1 score and $61.25\%$ UAR, compared to RCN-A's [49] $39.28\%$ UF1 and $38.93\%$ UAR. For 4 emotion classes, $\mu$ -BERT outperforms Baseline (+Depth) [14] $47.18\%$ to $30.01\%$ for UF1 and $49.13\%$ to $29.82\%$ for UAR. Large gains over Baseline (+Depth) [14] are seen in the case of 7 emotion classes, where $\mu$ -BERT attains UF1 and UAR scores of $32.64\%$ and $32.54\%$ respectively, compared to $17.73\%$ and $18.29\%$ for the baseline. + +Table 2 details results for CASMEII. $\mu$ -BERT shows improvements over all other methods. For three categories, it achieves a UF1 of $90.34\%$ and UAR of $89.14\%$ , representing $3.37\%$ and $0.86\%$ increases over the prior leading method (OFF-ApexNet [8]), respectively. Similar improvement is seen in five categories: a $4.83\%$ over TSCNN [35] in terms of UF1 and a $0.89\%$ increase over SMA-STN [19] for UAR. Similarly, $\mu$ -BERT performs competitively with other methods on the SAMM as seen in Table 3. Using 5 emotion classes, $\mu$ -BERT outperforms MinMaNet [47] by a large margin in terms of UF1 ( $83.86\%$ vs $76.40\%$ ) and UAR ( $84.75\%$ vs $76.70\%$ ), respectively. The performance of $\mu$ -BERT on SMIC is compared against several others in Table 4. $\mu$ -BERT outperforms others with a $7.5\%$ increase in UF1 to $85.5\%$ and a $3.97\%$ boost in UAR to $83.84\%$ . + +On the composite dataset, $\mu$ -BERT again outperforms + +Table 1. MER on the CASME3 dataset. + +
Method# ClassesUF1 (%)UAR(%)
FR [54]334.9334.13
STSTNet [18]337.9537.92
RCN-A [49]339.2838.93
μ-BERT (ours)356.0461.25
Baseline [14]429.1529.10
Baseline (+Depth) [14]430.0129.82
μ-BERT (ours)447.1849.13
Baseline [14]717.5918.01
Baseline(+Depth) [14]717.7318.29
μ-BERT (ours)732.6432.54
+ +Table 2. MER on CASME II dataset. + +
Method# ClassesUF1 (%)UAR (%)
LR-GACNN [10]570.9081.30
AMAN [46]571.0075.40
Graph-TCN [13]572.4673.98
DSTAN [44]573.0075.00
GEME [28]573.5475.20
MiMaNet [47]575.9079.90
SMA-STN [19]579.4682.59
TSCNN [35]580.7080.97
μ-BERT (ours)585.5383.48
STSTNet [17]383.8286.86
OFF-ApexNet [8]386.9788.28
MAE [9]388.0387.28
μ-BERT (ours)390.3489.14
+ +Table 3. MER on SAMM dataset. + +
Method# ClassesUF1 (%)UAR (%)
AMAN [46]567.0068.85
SMA-STN [19]570.3377.20
GRAPH-AU [12]570.4574.26
MTMNet [48]573.6074.10
MiMaNet [47]576.4076.70
MAE [9]580.4088.98
μ-BERT (ours)583.8684.75
+ +other methods (Table 5). Attaining a UF1 score of $89.03\%$ and UAR of $88.42\%$ , $\mu$ -BERT realizes $0.73\%$ , and $0.82\%$ gains over previous best MiMaNet [47], respectively. Table 6 shows the impact of DMA and POI on CASME3. Our + +Table 4. MER on SMIC dataset. + +
Method# ClassesUF1 (%)UAR (%)
DIKD [36]371.0076.06
TSCNN [35]372.3672.74
MTMNet [48]374.4076.00
AMAN [46]377.0079.87
MiMaNet [47]377.8078.60
DSTAN [44]378.0077.00
MAE [9]381.8680.82
μ-BERT (ours)385.5083.84
+ +method gives more modest gains of approximately $2\%$ in both metrics. A greater improvement is seen with DMA, where UF1 and UAR increase by another $2 - 4\%$ . Significant improvement from $\mu$ -BERT is seen when adopting both modules, with a UF1 of $32.64\%$ and UAR of $32.54\%$ , representing roughly $10\%$ gains over previous methods. + +# 5.5. How $\mu$ -BERT perceives micro-movements + +To understand the micro-movements between two frames, the onset and apex frames are inputs for $\mu$ -BERT. These frames represent the moments that the microexpression starts and is observed. We measure $\mathrm{diag}(\hat{A})$ (Subsection 4.5) and $\mathbf{S}_{t + \delta}$ (Eqn (11)) values to identify which regions contain small movements between two frames. Comparisons of $\mu$ -BERT with RAFT [37], i.e., optical flow-based method and MagNet [29] are also conducted as in Fig. 7. The third and fourth columns in Fig 7 show the results of RAFT [37], and MagNet [29] on spotting the micro-movements, respectively. While RAFT is an optical flow-based method, MagNet amplifies small differences between the two frames. These methods are sensitive to the environment (e.g., lighting, illuminations). Thus, noises in the background still exist in their outputs. In addition, neither RAFT nor MagNet understand semantic information in the frame and distinguish changes inside facial or background regions. Meanwhile, $\mu$ -BERT shows its advantages in perceiving micro-movements via distinguishing the facial regions and spotting the micro-expressions. In particular, the attention map in the fifth column, in Fig. 7 illustrates the micro-differences between onset and apex frames. The higher contrast represents the higher chance of small movements in these regions. With the POI module, $\mu$ -BERT can automatically figure out the informative patches and ignore the background ones. Then, with DMA module, $\mu$ -BERT, can detect and localize which corresponding patches/regions contain tiny movements. As shown in the seventh column, attention maps represent the most salient regions in the image. By empowering DMA and POI, $\mu$ -BERT effectively identifies micro-movements within facial + +Table 5. MER on the Composite dataset (MECG2019). + +
Method# ClassesUF1 (%)UAR(%)
Dual-Inception [55]373.2272.78
FR [53]378.3878.32
NMER [20]378.8578.24
GRAPH-AU [12]379.1479.33
ICE-GAN [52]384.5084.10
BDCNN [3]385.0985.00
moment [48]386.4085.70
MiMaNet [47]388.3087.60
MAE [9]388.5087.40
μ-BERT (ours)389.0388.42
+ +regions, as demonstrated in the last column. + +# 5.6. Ablation studies + +This section compares $\mu$ -BERT against other self-supervised learning (SSL) methods on the MER task. CASME3 is used for experiments since it has many unlabeled images to demonstrate the power of SSL methods. We also analyze the essential contributions of Diagonal Micro-Attention (DMA) and Patch of Interest (POI) modules. Finally, we illustrate the robustness of $\mu$ -BERT pretrained on CASME3 on unseen datasets and domains. + +Comparisons with self-supervised learning methods. We utilize the encoder and decoder parts of $\mu$ -BERT (without DMA and POI) to train previous SSL methods (MoCo V3 [4], BEIT [1], MAE [9]) and then continue learning the MER task on the large-scale database CASME3 [14]. Overall results are shown in Table 6. It is expected that ViT-S achieves the lowest performance for UF1 and UAR as ImageNet and Micro-Expression are two different domains. Three self-supervised methods (MoCo V3, BEIT, and MAE) got better results when they were pretrained on CASME before fine-tuning to the recognition task. Compared to ViT-S, these SSL methods gain remarkable performance. Especially, MAE [9] achieves $3.5\%$ and $2\%$ up on UF1 and UAR compared to ViT-S, respectively. + +The role of Blockwise Swapping. Our basic setup of $\mu$ -BERT (denoted as MB1) is employed to train in an SSL manner. It is noted that only Blockwise Swapping is involved, and it does not contain either DMA or POI. Compared to MAE, MB1 outperforms MAE by $2\%$ in both UF1 and UAR, approximately. The reasons are: (1) Blockwise Swapping enforces the model to learn local context features inside an image, i.e., $I_{t}$ , and (2) It helps the network to figure out micro-disparities between two frames $I_{t}$ and $I_{t + \delta}$ . + +The role of DMA. This module is the guide to tell the network where to look and which patches to focus. By doing so, the $\mu$ -BERT gets more robust knowledge of micro + +Table 6. MER performance on CASME3 by different self-supervised methods and various settings of $\mu$ -BERT + +
MethodPre-trainDMAPOIUF1UAR
ViT-S [7]ImageNetXX20.3418.76
MoCo V3 - R50 [4]CASME3XX19.1217.36
MoCo V3 - R101 [4]CASME3XX20.1418.52
MoCo V3 [4]CASME3XX22.1319.34
BEIT [1]CASME3XX23.5419.89
MAE [9]CASME3XX23.8620.87
μ-BERT (MB1)CASME3XX25.2722.96
μ-BERT (MB2)CASME3✓X27.3526.18
μ-BERT (MB3)CASME3✓✓32.6432.54
+ +movements between two frames. For this reason, the network (denoted as MB2) achieves $2\%$ on UF1 and a significant $4\%$ gain on UAR compared to MB1. + +The role of POI. Since MB1 are sensitive to background noise, the micro-disparities features $\mathbf{P}_{dma}$ might contain unwanted features coming from the background. The POI is designed as a filter that only lets the typical interesting patches belonging to the subject go through and preserves the micro-movement features only. The improvements of up to $6\%$ compared to MB2 demonstrate the important role of POI in $\mu$ -BERT for micro-expression tasks. Qualitative results demonstrated in Supplementary Material can further emphasize the advantages of POI in assisting the network to be robust against facial movements. + +# 6. Conclusions and Discussions + +Unlike a few concurrent research on micro-expression, we move forward and study how to explore BERT pretraining for this problem. In our proposed $\mu$ -BERT, we presented a novel Diagonal Micro Attention (DMA) to learn the micro-movements of the subject across frames. The Patch of Interest (POI) module is proposed to guide the network, focusing on the most salient parts, i.e., facial regions, and ignoring the noisy sensitivities from the background. Empowered by the simple design of $\mu$ -BERT, SOTA performance on micro-expression recognition tasks is achieved in four benchmark datasets. Our perspective will inspire more future study efforts in this direction. + +Limitations. We demonstrated the efficiency of the POI module in removing noise in the background, which is sensitive to lighting and illumination. However, suppose any facial parts, e.g., the forehead, are affected by lighting conditions while there are no movements. In that case, these noisy factors might also be included as micro-difference features. The robustness with different lighting conditions will be left as our future works. + +Acknowledgement This work is supported by Arkansas Biosciences Institute (ABI) Grant, NSF WVAR-CRESH and NSF Data Science, Data Analytics that are Robust and Trusted (DART). We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs. + +# References + +[1] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. 2021. 2, 8 +[2] Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu. Video-based facial micro-expression analysis: A survey of datasets, features and algorithms. IEEE transactions on pattern analysis and machine intelligence, 2021. 1 +[3] Bin Chen, Kun-Hong Liu, Yong Xu, Qing-Qiang Wu, and Jun-Feng Yao. Block division convolutional network with implicit deep features augmentation for micro-expression recognition. IEEE Transactions on Multimedia, 2022. 8 +[4] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 8 +[5] Ciprian Adrian Coraneanu, Marc Oliu Simón, Jeffrey F. Cohn, and Sergio Escalera Guerrero. Survey on rgb, 3d, thermal, and multimodal approaches for facial expression recognition: History, trends, and affect-related applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(8):1548-1568, 2016. 1 +[6] Adrian K. Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap. Samm: A spontaneous micro-facial movement dataset. IEEE Transactions on Affective Computing, 9(1):116-129, 2018. 1, 2, 5 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 8 +[8] Y.S. Gan, S.T. Liong, W.C. Yau, Y.C. Huang, and L.K. Tan. Off-apexnet on micro-expression recognition system. Signal Processing: Image Communication, 74:129-139, 2019. 6, 7 +[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 7, 8 +[10] Ankith Jain Rakesh Kumar and Bir Bhanu. Micro-expression classification based on landmark relations with graph attention convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1511–1520, June 2021. 7 +[11] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression recognition based on facial graph representation learning and facial action unit fusion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571-1580, 2021. 1, 2 +[12] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression recognition based on facial graph representation learning and facial action unit fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1571-1580, June 2021. 7, 8 + +[13] Ling Lei, Jianfeng Li, Tong Chen, and Shigang Li. A novel Graph-TCN with a graph structured representation for micro-expression recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2237–2245, 2020. 7 +[14] Jingting Li, Zizhao Dong, Shaoyuan Lu, Su-Jing Wang, Wen-Jing Yan, Yinhuan Ma, Ye Liu, Changbing Huang, and Xiaolan Fu. Cas(me)sup3/sup: A third generation facial spontaneous micro-expression database with depth information and high ecological validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2, 5, 6, 7, 8 +[15] Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietikainen. A spontaneous micro-expression database: Inducement, collection and baseline. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6, 2013. 1, 2, 5 +[16] Yante Li, Xiaohua Huang, and Guoying Zhao. Microexpression action unit detection with spatial and channel attention. Neurocomputing, 436:221-231, 2021. 2 +[17] S.T. Liong, Y.S. Gan, J. See, H. Khor, and Y. Huang. Shallow triple stream three-dimensional cnn (ststnet) for microexpression recognition. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-5. IEEE, 2019. 7 +[18] Sze-Teng Liong, Yee Siang Gan, John See, Huai-Qian Khor, and Yen-Chang Huang. Shallow triple stream three-dimensional cnn (ststnet) for micro-expression recognition. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-5. IEEE, 2019. 7 +[19] Jiateng Liu, Wenming Zheng, and Yuan Zong. SMA-STN: Segmented movement-attending spatiotemporal network for micro-expression recognition. arXiv preprint arXiv:2010.09342, 2020.6.7 +[20] Yuchi Liu, Heming Du, Liang Zheng, and Tom Gedeon. A neural micro-expression recognizer. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-4. IEEE, 2019. 8 +[21] Yanju Liu, Yange Li, Xinhai Yi, Zuojin Hu, Huiyu Zhang, and Yanzhong Liu. Lightweight vit model for microexpression recognition enhanced by transfer learning. Frontiers in Neurorobotics, 16, 2022. 2 +[22] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 2 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2 +[24] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 + +[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[26] Xuan-Bac Nguyen, Duc Toan Bui, Chi Nhan Duong, Tien D Bui, and Khoa Luu. Clusformer: A transformer based clustering approach to unsupervised large-scale face and visual landmark recognition. 2 +[27] Xuan-Bac Nguyen, Guee Sang Lee, Soo Hyung Kim, and Hyung Jeong Yang. Self-supervised learning based on spatial awareness for medical image analysis. IEEE Access, 8:162973-162981, 2020. 2 +[28] Xuan Nie, Madhumita A Takalkar, Mengyang Duan, Haimin Zhang, and Min Xu. GEME: Dual-stream multi-task gender-based micro-expression recognition. Neurocomputing, 427:13-28, 2021. 7 +[29] Tae-Hyun Oh, Ronnachai Jaroensri, Changil Kim, Mohamed Elgharib, Fredo Durand, William T. Freeman, and Wojciech Matusik. Learning-based video motion magnification, 2018. 1, 2, 7 +[30] Kha Gia Quach, Ngan Le, Chi Nhan Duong, Ibsa Jalata, Kaushik Roy, and Khoa Luu. Non-volume preserving-based fusion to group-level emotion recognition on crowd videos. Pattern Recognition, 128:108646, 2022. 2 +[31] Ankith Jain Rakesh Kumar and Bir Bhanu. Micro-expression classification based on landmark relations with graph attention convolutional network. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1511-1520, 2021. 2 +[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[33] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 2 +[34] John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su-Jing Wang. Megc 2019 – the second facial microexpressions grand challenge. In 2019 14th IEEE International Conference on Automatic Face Gesture Recognition (FG 2019), pages 1–5, 2019. 6 +[35] B. Song, K. Li, Y. Zong, J. Zhu, W. Zheng, J. Shi, and L. Zhao. Recognizing spontaneous micro-expression using a three-stream convolutional neural network. IEEE Access, 7:184537-184551, 2019. 6, 7 +[36] Bo Sun, Siming Cao, Dongliang Li, Jun He, and Lejun Yu. Dynamic micro-expression recognition using knowledge distillation. IEEE Transactions on Affective Computing, 2020. 7 +[37] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 7 +[38] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers amp; distillation through attention. In International Conference on Machine Learning, volume 139, pages 10347-10357, July 2021. 2 + +[39] Thuong-Khanh Tran, Quang-Nhat Vo, Xiaopeng Hong, Xiaobai Li, and Guoying Zhao. Micro-expression spotting: A new benchmark. Neurocomputing, 443:356-368, 2021. 2 +[40] Thanh-Dat Truong, Quoc-Huy Bui, Chi Nhan Duong, Han-Seok Seo, Son Lam Phung, Xin Li, and Khoa Luu. Direcformer: A directed attention in transformer approach to robust action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20030-20040, June 2022. 2 +[41] Thanh-Dat Truong, Chi Nhan Duong, The De Vu, Hoang Anh Pham, Bhiksha Raj, Ngan Le, and Khoa Luu. The right to talk: An audio-visual transformer approach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1105–1114, October 2021. 2 +[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[43] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale micro-expression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2 +[44] Yan Wang, Yikun Huang, Can Liu, Xiaoying Gu, Dandan Yang, Shuopeng Wang, and Bo Zhang. Micro expression recognition via dual-stream spatiotemporal attention network. Journal of Healthcare Engineering, 2021, 2021. 7 +[45] Yandan Wang, John See, Yee-Hui Oh, Raphael C.-W. Phan, Yogachandran Rahulamathavan, Huo-Chong Ling, Su-Wei Tan, and Xujie Li. Effective recognition of facial microexpressions with video motion magnification. Multimedia Tools and Applications, 76(20):21665-21690, 2016. 2 +[46] Mengting Wei, Wenming Zheng, Yuan Zong, Xingxun Jiang, Cheng Lu, and Jiateng Liu. A novel micro-expression recognition approach using attention-based magnification-adaptive networks. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2420-2424. IEEE, 2022. 7 +[47] Bin Xia and Shangfei Wang. Micro-expression recognition enhanced by macro-expression from spatial-temporal domain. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 1186-1193, 2021. 6, 7, 8 +[48] Bin Xia, Weikang Wang, Shangfei Wang, and Enhong Chen. Learning from macro-expression: a micro-expression recognition framework. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2936-2944, 2020. 7, 8 +[49] Zhaoqiang Xia, Wei Peng, Huai-Qian Khor, Xiaoyi Feng, and Guoying Zhao. Revealing the invisible with model and data shrinking for composite-database micro-expression recognition. IEEE Transactions on Image Processing, 29:8590-8605, 2020. 6, 7 +[50] Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu. Casme ii: An improved spontaneous micro-expression database and the baseline evaluation. PLOS ONE, 9(1):1-8, 01 2014. 2, 5 + +[51] Wen-Jing Yan, Qi Wu, Yong-Jin Liu, Su-Jing Wang, and Xiaolan Fu. Casme database: A dataset of spontaneous microexpressions collected from neutralized faces. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-7, 2013. 1 +[52] Jianhui Yu, Chaoyi Zhang, Yang Song, and Weidong Cai. ICE-GAN: Identity-aware and capsule-enhanced gan for micro-expression recognition and synthesis. arXiv preprint arXiv:2005.04370, 2020.8 +[53] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition. arXiv preprint arXiv:2101.04838, 2021. 8 +[54] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition. Pattern Recognition, 122:108275, 2022. 7 +[55] L. Zhou, Q. Mao, and L. Xue. Dual-inception network for cross-database micro-expression recognition. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-5. IEEE, 2019. 8 \ No newline at end of file diff --git a/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/images.zip b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ff2fbf537652f60d846bd580ab993e9d216fe95b --- /dev/null +++ b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:549b1b6354082e83d8a81cdb0f566d65992e9b1da918a1ab5ef69f4bb382bd1e +size 617706 diff --git a/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/layout.json b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4b384455199e2a1e8ad0e2f94f7367c224e5af16 --- /dev/null +++ b/2023/Micron-BERT_ BERT-Based Facial Micro-Expression Recognition/layout.json @@ -0,0 +1,12666 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 499, + 121 + ], + "type": "text", + "content": "Micron-BERT: BERT-based Facial Micro-Expression Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 129, + 517, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 129, + 517, + 143 + ], + "spans": [ + { + "bbox": [ + 74, + 129, + 517, + 143 + ], + "type": "text", + "content": "Xuan-Bac Nguyen1, Chi Nhan Duong2, Xin Li3, Susan Gauch1, Han-Seok Seo4, Khoa Luu1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 143, + 484, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 143, + 484, + 156 + ], + "spans": [ + { + "bbox": [ + 110, + 143, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 143, + 484, + 156 + ], + "type": "text", + "content": " CVIU Lab, University of Arkansas, USA " + }, + { + "bbox": [ + 110, + 143, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 143, + 484, + 156 + ], + "type": "text", + "content": " Concordia University, Canada" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 157, + 512, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 157, + 512, + 171 + ], + "spans": [ + { + "bbox": [ + 81, + 157, + 512, + 171 + ], + "type": "text", + "content": "3 West Virginia University, USA 4 Dep. of Food Science, University of Arkansas, USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 172, + 526, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 526, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 526, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 172, + 526, + 186 + ], + "type": "text", + "content": " {xnguyen, sgauch, hanseok, khoaluu}@uark.edu, " + }, + { + "bbox": [ + 62, + 172, + 526, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 172, + 526, + 186 + ], + "type": "text", + "content": " {dcnhan@ieee.org}, " + }, + { + "bbox": [ + 62, + 172, + 526, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 172, + 526, + 186 + ], + "type": "text", + "content": " xin.li@mail.wvu.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "text", + "content": "Micro-expression recognition is one of the most challenging topics in affective computing. It aims to recognize tiny facial movements difficult for humans to perceive in a brief period, i.e., 0.25 to 0.5 seconds. Recent advances in pre-training deep Bidirectional Transformers (BERT) have significantly improved self-supervised learning tasks in computer vision. However, the standard BERT in vision problems is designed to learn only from full images or videos, and the architecture cannot accurately detect details of facial micro-expressions. This paper presents Micron-BERT (" + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "text", + "content": "-BERT), a novel approach to facial micro-expression recognition. The proposed method can automatically capture these movements in an unsupervised manner based on two key ideas. First, we employ Diagonal Micro-Attention (DMA) to detect tiny differences between two frames. Second, we introduce a new Patch of Interest (PoI) module to localize and highlight micro-expression interest regions and simultaneously reduce noisy backgrounds and distractions. By incorporating these components into an end-to-end deep network, the proposed " + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "text", + "content": "-BERT significantly outperforms all previous work in various micro-expression tasks. " + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "text", + "content": "-BERT can be trained on a large-scale unlabeled dataset, i.e., up to 8 million images, and achieves high accuracy on new unseen facial micro-expression datasets. Empirical experiments show " + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 45, + 239, + 290, + 597 + ], + "type": "text", + "content": "-BERT consistently outperforms state-of-the-art performance on four micro-expression benchmarks, including SAMM, CASME II, SMIC, and CASME3, by significant margins. Code will be available at https://github.com/uark-cviu/Micron-BERT" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 608, + 128, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 128, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 128, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Facial expressions are a complex mixture of conscious reactions directed toward given stimuli. They involve experiential, behavioral, and physiological elements. Because they are crucial to understanding human reactions, this topic has been widely studied in various application domains [5]. In general, facial expression problems can be classified into two main categories, macro-expression, and" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 213, + 539, + 366 + ], + "blocks": [ + { + "bbox": [ + 313, + 213, + 539, + 366 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 539, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 539, + 366 + ], + "type": "image", + "image_path": "ad7894b0dc5aadb9f880d68016148b89febed11c0d74cd3d7b49092d87acef88.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 374, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 374, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 547, + 407 + ], + "type": "text", + "content": "Figure 1. Given two frames from a high-speed video, the proposed " + }, + { + "bbox": [ + 305, + 374, + 547, + 407 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 374, + 547, + 407 + ], + "type": "text", + "content": "-BERT method can localize and highlight the regions of micromovements. Best viewed in color." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 413, + 546, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 546, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 546, + 556 + ], + "type": "text", + "content": "micro-expression. The main differences between the two are facial expression intensities, and duration [2]. In particular, macro-expressions happen spontaneously, cover large movement areas in a given face, e.g., mouth, eyes, cheeks, and typically last from 0.5 to 4 seconds. Humans can usually recognize these expressions. By contrast, micro-expressions are involuntary occurrences, have low intensity, and last between 5 milliseconds and half a second. Indeed, micro-expressions are challenging to identify and are mostly detectable only by experts. Micro-expression understanding is essential in numerous applications, primarily lie detection, which is crucial in criminal analysis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 676 + ], + "type": "text", + "content": "Micro-expression identification requires both semantics and micro-movement analysis. Since they are difficult to observe through human eyes, a high-speed camera, usually with 200 frames per second (FPS) [6, 15, 51], is typically used to capture the required video frames. Previous work [11] tried to understand this micro information using MagNet [29] to amplify small motions between two frames, e.g., onset and apex frames. However, these methods still have limitations in terms of accuracy and robustness. In summary, the contributions of this work are four-fold:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 677, + 547, + 714 + ], + "type": "text", + "content": "- A novel Facial Micro-expression Recognition (MER) via Pre-training of Deep Bidirectional Transformers approach (Micron-BERT or " + }, + { + "bbox": [ + 317, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 317, + 677, + 547, + 714 + ], + "type": "text", + "content": "-BERT) is presented to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1482" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "type": "text", + "content": "tackle the problem in a self-supervised learning manner. The proposed method aims to identify and localize micro-movements in faces accurately." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 113, + 286, + 341 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 59, + 113, + 286, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 113, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 113, + 286, + 171 + ], + "type": "text", + "content": "- As detecting the tiny moment changes in faces is an essential input to the MER module, a new Diagonal Micro Attention (DMA) mechanism is proposed to precisely identify small movements in faces between two consecutive video frames." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 182, + 286, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 182, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 59, + 182, + 286, + 241 + ], + "type": "text", + "content": "- A new Patch of Interest (POI) module is introduced to efficiently spot facial regions containing the microexpressions. Far apart from prior methods, it is trained in an unsupervised manner without using any facial labels, such as facial bounding boxes or landmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 246, + 286, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 246, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 59, + 246, + 286, + 341 + ], + "type": "text", + "content": "- The proposed " + }, + { + "bbox": [ + 59, + 246, + 286, + 341 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 59, + 246, + 286, + 341 + ], + "type": "text", + "content": "-BERT framework is designed in a self-supervised learning manner and trained in an end-to-end deep network. Indeed, it consistently achieves State-of-the-Art (SOTA) results in various standard micro-expression benchmarks, including CASME II [50], CASME3 [14], SAMM [6] and SMIC [15]. It achieves high recognition accuracy on new unseen subjects of various gender, age, and ethnicity." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 354, + 133, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 133, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 133, + 365 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 374, + 286, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 286, + 409 + ], + "type": "text", + "content": "Generally, prior studies in micro-expression can be divided into two categories, including micro-expression spotting (MES) and micro-expression recognition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 411, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 286, + 506 + ], + "type": "text", + "content": "Micro-Expression Spotting (MES). The goal of MES is to determine the specific instant during which a micro-expression occurs. Li et al. [16] adopted a spatial-channel attention network to detect micro-expression action units. Tran et al. [39] attempted to standardize with the SMIC-E database and an evaluation protocol. MESNet [43] introduced a CNN-based approach with a " + }, + { + "bbox": [ + 47, + 411, + 286, + 506 + ], + "type": "inline_equation", + "content": "(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 47, + 411, + 286, + 506 + ], + "type": "text", + "content": " convolutional network, a clip proposal, and a classifier." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 507, + 286, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 286, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 286, + 649 + ], + "type": "text", + "content": "Micro-Expression Recognition (MER). The goal of MER tasks is to classify the facial micro-expressions in a video. Ling et al. [11] present a new way of learning facial graph representations, allowing these small movements to be seen. Kumar and Bhanu [31] exploited connections between landmark points and their optical flow patch and achieved improvements over state-of-the-art (SOTA) methods for both the CASME II and SAMM. Liu et al. [21] presented a new method using transfer learning achieved an accuracy of " + }, + { + "bbox": [ + 47, + 507, + 286, + 649 + ], + "type": "inline_equation", + "content": "84.27\\%" + }, + { + "bbox": [ + 47, + 507, + 286, + 649 + ], + "type": "text", + "content": " on a composite of three datasets. Wang et al. [45] presented an Eulerian-motion magnification-based approach that highlights these small movements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 651, + 286, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 286, + 722 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 286, + 722 + ], + "type": "text", + "content": "Other Work. Other research, while not necessarily on MES or MER, is relevant to our approach. An advance in video motion magnification is shown in [29], outperforming the SOTA methods in multiple areas. This learning-based model can extract filters from data directly rather than rely on ones designed by hand, like the state-of-the-art method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 71, + 403, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 403, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 403, + 83 + ], + "type": "text", + "content": "3. BERT Revisited" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 91, + 448, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 448, + 103 + ], + "type": "text", + "content": "3.1. BERT in Vision Problems" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 110, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 350 + ], + "type": "text", + "content": "Transformers and deep learning have significantly improved results for many tasks in computer vision [1, 7, 9, 22, 23, 26, 27, 30, 40, 41]. Worth mentioning is Vision Transformer (ViT) [7], one of the first research efforts at the intersection of Transformers and computer vision. Unlike the traditional CNN network, ViT splits an image into a sequence of patches and applies the Transformers-based framework directly. Inspired by the success of BERT in Natural Language Processing (NLP), Bidirectional Encoder representation from Image Transformers (BEiT) [1] is presented as a self-supervised learning framework in computer vision. In particular, image patches are tokenized using DALL-E [32] to the visual tokens. These tokens are then randomly masked before feeding into the transformer backbone. The training objective is to recover the original visual tokens from the corrupted patches. These methods [1, 38] have marked a remarkable improvement compared to supervised learning methods by leveraging large-scale unlabelled datasets, e.g., ImageNet-1K, ImagNet-21K [33], to discover semantic information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 360, + 517, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 517, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 517, + 372 + ], + "type": "text", + "content": "3.2. Limitations of BERT in Vision Problems" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": "One limitation of using BERT in vision problems is the tokenization step. In the NLP field, a token has precisely one word mapped into it. In vision problems, however, many possible images or patches can share the same token as long as they have the same content. Therefore, designing a BERT model to mask a token and train a prediction model in the missing contexts in computer vision is more challenging than NLP. In addition, the tokenizer, i.e., DALL-E [32], is not robust enough to map similar contexts to a token. It yields noise in the tokenization process and affects the overall training performance. He et al., [9] presented a Masked Auto Encoder (MAE) that utilizes the BERT framework. Instead of tokenizing images, it eliminates patches of an image via a random masking strategy and reconstructs the context of these masked patches to the original content. Although this method can avoid using the tokenizer, it only considers the context inside an image. Thus, it does not apply to micro-expression, which requires understanding semantic information from consecutive video frames. In this paper, " + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 380, + 545, + 619 + ], + "type": "text", + "content": "-BERT is presented to address these limitations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 632, + 491, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 632, + 491, + 646 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 491, + 646 + ], + "type": "text", + "content": "4. The Proposed " + }, + { + "bbox": [ + 305, + 632, + 491, + 646 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 632, + 491, + 646 + ], + "type": "text", + "content": "-BERT Approach" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "-BERT is designed to model micro-changes of facial texture across temporal dimensions, which is hard to observe by unaided human eyes via a reconstruction process. The proposed " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "-BERT architecture, shown in Figure 2, consists of five main blocks: a " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "-Encoder, Patch" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1483" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 521, + 224 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 521, + 224 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 521, + 224 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 521, + 224 + ], + "type": "image", + "image_path": "2a96aa8f2d5c59122ab599131db679ec754bd921f014ce2a01b7bf27611f9fce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 229, + 472, + 241 + ], + "lines": [ + { + "bbox": [ + 119, + 229, + 472, + 241 + ], + "spans": [ + { + "bbox": [ + 119, + 229, + 472, + 241 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed " + }, + { + "bbox": [ + 119, + 229, + 472, + 241 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 119, + 229, + 472, + 241 + ], + "type": "text", + "content": "-BERT approach to facial micro-expression recognition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": "of Interest (PoI), Blockwise Swapping, Diagonal Micro Attention (DMA), and a " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": "-Decoder. Given input images " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": ", the role of the " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": "-Encoder is to represent " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": " into latent vectors. Then, Patch of Interest (PoI) constrains " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": "-BERT to look into facial regions containing microexpressions rather than unrelated regions such as the background. Blockwise Swapping and Diagonal Micro Attention (DMA) allow the model to focus on facial regions that primarily consist of micro differences between frames. Finally, " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": "-Decoder reconstructs the output signal back to the determined one. Compared to prior works, " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": "-BERT can adaptively focus on changes in facial regions while ignoring the ones in the background and effectively recognizes micro-expressions even when face movements occur. Moreover, " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": "-BERT can also alleviate the dependency on the accuracy of alignment approaches in pre-processing step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 261, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 261, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 261, + 462 + ], + "type": "text", + "content": "4.1. Non-overlapping Patches Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "type": "text", + "content": "In " + }, + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "type": "text", + "content": "-BERT, an input frame " + }, + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "type": "inline_equation", + "content": "I_{t} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "type": "text", + "content": " is divided into a set of several non-overlapping patches " + }, + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 47, + 467, + 287, + 493 + ], + "type": "text", + "content": " as Eqn. (1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 501, + 287, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 501, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 91, + 501, + 287, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {t} = \\left\\{p _ {t} ^ {i} \\right\\} _ {i = 0} ^ {N _ {p} - 1} \\quad \\left| \\mathcal {P} _ {t} \\right| = H W / \\left(p s ^ {2}\\right) \\tag {1}", + "image_path": "99b667d21fa5fddc36f1ba2d0758ba5839fb0fb01f23b05d95830dc397d3e711.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "inline_equation", + "content": "H, W, C" + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "content": " are the height, width, and number of channels, respectively. Each patch " + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "inline_equation", + "content": "p_t^i" + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "content": " has a resolution of " + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "inline_equation", + "content": "ps \\times ps" + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "content": ". In our experiments, " + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "inline_equation", + "content": "H = W = 224" + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "inline_equation", + "content": "C = 3" + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "inline_equation", + "content": "ps = 8" + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 118, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 118, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 118, + 574 + ], + "type": "text", + "content": "4.2. " + }, + { + "bbox": [ + 47, + 563, + 118, + 574 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 563, + 118, + 574 + ], + "type": "text", + "content": "-Encoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "text", + "content": "Each patch " + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathcal{P}_t" + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "text", + "content": " is linearly projected into a latent vector of dimension " + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^i \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "text", + "content": ", with additive fixed positional encoding [42]. Then, an image " + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 47, + 581, + 287, + 629 + ], + "type": "text", + "content": " can be represented as in Eqn. (2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 636, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 636, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 85, + 636, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {t} = \\operatorname {c o n c a t} \\left[ \\mathbf {z} _ {t} ^ {0}, \\mathbf {z} _ {t} ^ {1}, \\dots \\mathbf {z} _ {t} ^ {N _ {p} - 1} \\right] \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {2}", + "image_path": "87af88f76628ad8b5adaf2c4181427e19bc782a899e09f7b66438ec0bf8b6035.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 658, + 158, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 658, + 158, + 670 + ], + "spans": [ + { + "bbox": [ + 88, + 658, + 158, + 670 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} ^ {i} = \\alpha \\left(p _ {t} ^ {i}\\right) + \\mathbf {e} (i)", + "image_path": "534e4949b29cf5911cd4f15ccce836b0124bb7dafd89abc2743a7e69cc33c6d1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " are the projection embedding network and positional embedding, respectively. Let " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "-Encoder, denoted as " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": ", be a stack of continuous blocks. Each block consists" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 351, + 247, + 502, + 309 + ], + "blocks": [ + { + "bbox": [ + 351, + 247, + 502, + 309 + ], + "lines": [ + { + "bbox": [ + 351, + 247, + 502, + 309 + ], + "spans": [ + { + "bbox": [ + 351, + 247, + 502, + 309 + ], + "type": "image", + "image_path": "1aa3c221ced05c8342dc33606299d97ff1be419bb0c70f3957dd271cdb0d336f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 316, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 338 + ], + "type": "text", + "content": "Figure 3. Builing block of Encoder and Decoder. Each block includes Multi-Head Attention (MHA) and Layer Normalization." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 346, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 396 + ], + "type": "text", + "content": "of alternating layers of Multi Head Attention (MHA) and Multi-Layer Perceptron (MLP), as illustrated in Figure 3. The Layer Norm (LN) is employed to the input signal before feeding to MHA and MLP layers, as in Eqn. (3)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 399, + 485, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 399, + 485, + 411 + ], + "spans": [ + { + "bbox": [ + 365, + 399, + 485, + 411 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} _ {l} = \\mathbf {x} _ {l - 1} + \\operatorname {M H A} (\\operatorname {L N} (\\mathbf {x} _ {l - 1}))", + "image_path": "3d3ca43b4f1db732cc0e1c6443689bd07e9a9ba4203532f7bd9cefedca5a0ae7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 368, + 414, + 545, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 414, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 368, + 414, + 545, + 426 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {l} = \\mathbf {x} ^ {\\prime} _ {l} + \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\mathbf {x} ^ {\\prime} _ {l}\\right)\\right) \\tag {3}", + "image_path": "66b6d6a8a4b59a256aeba232c68f80660e9c7066eb1642741159e2a3627f5337.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 429, + 449, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 429, + 449, + 440 + ], + "spans": [ + { + "bbox": [ + 367, + 429, + 449, + 440 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {0} = \\mathbf {Z} _ {t}, 1 \\leq l \\leq L _ {e}", + "image_path": "a2cd7ee2fb8001ecf71e7de993b1db85b32ef3f63fd82a1ef2944cfdc26107c3.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "text", + "content": " is the number of blocks in " + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t" + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "text", + "content": ", the output latent vector " + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}}" + }, + { + "bbox": [ + 306, + 445, + 544, + 469 + ], + "type": "text", + "content": " is represented as in Eqn. (4)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 473, + 545, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 473, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 367, + 473, + 545, + 487 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {t} = \\mathcal {E} \\left(\\mathbf {Z} _ {t}\\right) \\quad \\mathbf {P} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {4}", + "image_path": "2799378a33685e034754ffc78aa43896e62e505aa2810b2e2e52bf23905ae17b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 492, + 377, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 492, + 377, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 377, + 504 + ], + "type": "text", + "content": "4.3. " + }, + { + "bbox": [ + 306, + 492, + 377, + 504 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 306, + 492, + 377, + 504 + ], + "type": "text", + "content": "-Decoder" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "text", + "content": "The proposed auto-encoder is designed symmetrically. It means that the decoder part denoted as " + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "text", + "content": ", has a similar architecture to the encoder " + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "text", + "content": ". Given a latent vector " + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t" + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "text", + "content": ", the decoded signal " + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 305, + 511, + 545, + 559 + ], + "type": "text", + "content": " is represented as in Eqn. (5)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 365, + 563, + 545, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 563, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 365, + 563, + 545, + 576 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {t} = \\mathcal {D} \\left(\\mathbf {P} _ {t}\\right) \\quad \\mathbf {Q} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\tag {5}", + "image_path": "f65340e2ace31a1b8fbd9f77dbeb960fbb72d9046b888cefcae0bcf4a65822b5.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "text", + "content": "We add one more Linear layer to interpolate " + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "text", + "content": " to an intermediate signal " + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t" + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "text", + "content": " before reshaping it into the image size." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 357, + 623, + 545, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 623, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 357, + 623, + 545, + 638 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times d} \\xrightarrow {\\text {l i n e a r}} \\mathbf {y} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times p s \\times p s \\times C} \\tag {6}", + "image_path": "3843d8c3495ecb05cc1ced5041572d05f959d0697700f9fa6f654d685a8d64dc.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 329, + 642, + 510, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 642, + 510, + 658 + ], + "spans": [ + { + "bbox": [ + 329, + 642, + 510, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {t} \\in \\mathbb {R} ^ {N _ {p} \\times p s \\times p s \\times C} \\xrightarrow {\\text {r e s h a p e}} \\mathbf {y} _ {t} ^ {\\prime} \\in \\mathbb {R} ^ {H \\times W \\times C}", + "image_path": "01b4086db72c2e29f25f8a73d41187e79c1a020459f67e81581fb68aa9475ac7.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 664, + 425, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 425, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 425, + 677 + ], + "type": "text", + "content": "4.4. Blockwise Swapping" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 682, + 537, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 682, + 537, + 694 + ], + "spans": [ + { + "bbox": [ + 317, + 682, + 537, + 694 + ], + "type": "text", + "content": "Given two frames " + }, + { + "bbox": [ + 317, + 682, + 537, + 694 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 317, + 682, + 537, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 682, + 537, + 694 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 317, + 682, + 537, + 694 + ], + "type": "text", + "content": ", we realize the fact that:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 387, + 699, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 699, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 387, + 699, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\lim _ {\\delta \\rightarrow 0} s \\left(p _ {t} ^ {i}, p _ {t + \\delta} ^ {i}\\right) = 1 \\tag {7}", + "image_path": "90d74d976c4332994ecdfd9fdd96aab2d505b11e10921d6d8e7a9d251a278b2d.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1484" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 269, + 173 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 269, + 173 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 269, + 173 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 269, + 173 + ], + "type": "image", + "image_path": "6ec2580c36031dbf28f7accbd1bfed8be715f7615dafd87088460c7d1285de8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "text", + "content": "Figure 4. Blockwise Swapping. For each triplet, we present the " + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "text", + "content": " (left), the " + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "text", + "content": " (middle) and the " + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "inline_equation", + "content": "I_{t / s}" + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "text", + "content": " (right). The yellow blocks in " + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "inline_equation", + "content": "I_{t / s}" + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "text", + "content": " represent swapped patches from " + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "text", + "content": " that are randomly swapped into " + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 47, + 178, + 289, + 225 + ], + "type": "text", + "content": ". Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "p_t^i" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": "-patch at frame " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " denotes a function to measure the similarity between " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "p_t^i" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "p_{t + \\delta}^i" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " where a higher score indicates higher similarity and " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "0 \\leq s(p_t^i, p_{t + \\delta}^i) \\leq 1" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": ". Given a patch correlation as in Eqn. (7), we propose a Blockwise Swapping mechanism to (1) firstly randomly swap two corresponding blocks " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "p_t^i" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "p_{t + \\delta}^i" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " between two frames to create a swapped image " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "I_{t / s}" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": ", and then (2) enforce the model to spot these changes and reconstruct " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "I_{t / s}" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": ". By doing so, the model is further strengthened in recognizing and restoring the swapped patches. As a result, the learned model can be enhanced by the capability to notice small differences between frames. Moreover, as shown in Eqn. (7), shorter time " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " causing larger similarity between " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "inline_equation", + "content": "I_{t / s}" + }, + { + "bbox": [ + 46, + 234, + 290, + 429 + ], + "type": "text", + "content": " can further help to enhance the robustness on spotting these differences. The detail of this strategy is described in Algorithm 1 and Figure 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 436, + 228, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 228, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 228, + 449 + ], + "type": "text", + "content": "4.5. Diagonal Micro Attention (DMA)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": "As a result of Blockwise Swapping, the image patches " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t / s}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "I_{t / s}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " consists of two types, i.e. " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "p_{t / s}^{j}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "p_{t / s}^{i}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t + \\delta}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": ". Then, the next stage is to learn how to reconstruct " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t / s}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "p_{t / s}^{i}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " includes all changes between " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "I_{t / s}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": ", more emphasis is placed on " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "p_{t / s}^{i}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " during reconstruction process. Theoretically, the ground truth of the index of " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "p_{t / s}^{i}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t / s}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " can be utilized to enforce the model focusing on these swapped patches. However, adopting this information may reduce the learning capability to spot these microchanges. Therefore, a novel attention mechanism named Diagonal MicroAttention (DMA) is presented to enforce the network automatically focusing on swapped patches " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "p_{t / s}^{i}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " and equip it with the ability to precisely spot and identify all changes between images. Notice that these changes may include patches in the background. The following section introduces a solution to constrain the learned network focusing on only meaningful facial regions. The details of DMA are presented in Figure 5. Formally, we construct an attention map " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\hat{A}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t + \\delta}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t / s}" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " where the diag " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "(\\hat{A})" + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "text", + "content": " illustrates correlations between two corresponding patches " + }, + { + "bbox": [ + 46, + 456, + 289, + 716 + ], + "type": "inline_equation", + "content": "p_{t + \\delta}^{i}" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 540, + 158 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 540, + 158 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 540, + 158 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 540, + 158 + ], + "type": "image", + "image_path": "c2f42b4bb6f5535879a081ec2c47812701308136e2a1b03b3795f2d919f6d784.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 165, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 305, + 165, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 547, + 200 + ], + "type": "text", + "content": "Figure 5. Diagonal Micro-Attention (DMA) module. Diagonal values from the attention map between " + }, + { + "bbox": [ + 305, + 165, + 547, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{t / s}" + }, + { + "bbox": [ + 305, + 165, + 547, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 165, + 547, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{t + \\delta}" + }, + { + "bbox": [ + 305, + 165, + 547, + 200 + ], + "type": "text", + "content": " are used to rank the importance of each patch in the swapped image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "inline_equation", + "content": "p_{t / s}^{i}" + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "text", + "content": ". From the observation that " + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\hat{A}(i, i) > \\hat{A}(j, j)" + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "inline_equation", + "content": "p_{t / s}^{i} \\in \\mathcal{P}_{t + \\delta}" + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "inline_equation", + "content": "p_{t / s}^{j} \\in \\mathcal{P}_{t}" + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\text{diag}(\\hat{A})" + }, + { + "bbox": [ + 304, + 207, + 547, + 262 + ], + "type": "text", + "content": " can be effectively adopted as weights indicating important features. Full operations of DMA are presented in Eqn. (8) and Eqn. (9)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 277, + 546, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 546, + 309 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 546, + 309 + ], + "type": "interline_equation", + "content": "\\hat {A} = \\operatorname {s o f t m a x} \\left(Q \\left(\\mathbf {P} _ {t + \\delta}\\right) \\otimes K \\left(\\mathbf {P} _ {t / s}\\right) ^ {T}\\right), \\sum_ {j = 0} ^ {N _ {p}} \\hat {A} (i, j) = 1 \\tag {8}", + "image_path": "79b042008b32e8370786c0c6607bdfa62ee1382b8efaf0b3fd1c06b0f75a5401.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 329, + 546, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 329, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 370, + 329, + 546, + 342 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {d m a} = \\operatorname {d i a g} (\\hat {A}) \\times V \\left(\\mathbf {P} _ {t / s}\\right) \\tag {9}", + "image_path": "c9a79eb1b3b05f08a76eb9bd11eb376b354d8776c3bbea61f558f5153f1715c2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 351, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 545, + 364 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 351, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 306, + 351, + 545, + 364 + ], + "type": "text", + "content": " denotes the Element-wise multiplication operator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 371, + 437, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 437, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 437, + 384 + ], + "type": "text", + "content": "4.6. Patch of Interest (POI)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 388, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 486 + ], + "type": "text", + "content": "In Section 4.5, Diagonal Micro-Attention has been introduced to weigh the importance of swapped patches automatically. These swapped patches are randomly produced via Blockwise Swapping, as in Algorithm 1. In theory, the ideal case is when all swapped patches are located within the facial region only so that the deep network can learn the micro-movements from the facial parts solely and not be distracted by the background. In practice, however, we" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 504, + 448, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 448, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 448, + 517 + ], + "type": "text", + "content": "Algorithm 1 Blockwise Swapping" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 519, + 547, + 706 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t, \\mathcal{P}_{t + \\delta}" + }, + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "type": "text", + "content": " image patches (" + }, + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "type": "inline_equation", + "content": "N_p = h \\times w" + }, + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "type": "text", + "content": "); " + }, + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 312, + 519, + 547, + 552 + ], + "type": "text", + "content": ": swapping ratio (default: 0.5); min_bs: minimum block size (default: 16); min_ar: minimum aspect ratio (default: 0.3)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 553, + 457, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 457, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 457, + 563 + ], + "type": "text", + "content": "Output: Swapped image patches " + }, + { + "bbox": [ + 313, + 553, + 457, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t / s}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 564, + 405, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 564, + 405, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 564, + 405, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t / s}\\gets \\mathcal{P}_t;N_p\\gets |\\mathcal{P}_t|" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 576, + 341, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 341, + 584 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 341, + 584 + ], + "type": "inline_equation", + "content": "c\\gets 0" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 586, + 400, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 586, + 400, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 586, + 400, + 597 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 316, + 586, + 400, + 597 + ], + "type": "inline_equation", + "content": "c\\leq r_s\\times N_p" + }, + { + "bbox": [ + 316, + 586, + 400, + 597 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 597, + 448, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 597, + 448, + 608 + ], + "spans": [ + { + "bbox": [ + 326, + 597, + 448, + 608 + ], + "type": "inline_equation", + "content": "bs\\gets \\mathsf{rnd}(min\\_ bs,r_s\\times N_p - c)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 609, + 443, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 609, + 443, + 619 + ], + "spans": [ + { + "bbox": [ + 326, + 609, + 443, + 619 + ], + "type": "text", + "content": "ar←rnd(min_ar,1/min_ar)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 326, + 620, + 427, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 620, + 427, + 630 + ], + "spans": [ + { + "bbox": [ + 326, + 620, + 427, + 630 + ], + "type": "inline_equation", + "content": "m,n\\gets \\sqrt{bs}\\cdot ar,\\sqrt{bs / ar}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 326, + 631, + 465, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 631, + 465, + 641 + ], + "spans": [ + { + "bbox": [ + 326, + 631, + 465, + 641 + ], + "type": "inline_equation", + "content": "p,q\\gets \\mathsf{rnd}(0,h - m),\\mathsf{rnd}(0,w - n)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 326, + 642, + 447, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 642, + 447, + 652 + ], + "spans": [ + { + "bbox": [ + 326, + 642, + 447, + 652 + ], + "type": "inline_equation", + "content": "\\forall i \\in [p, p + m), j \\in [q, q + n)" + }, + { + "bbox": [ + 326, + 642, + 447, + 652 + ], + "type": "text", + "content": " :" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 345, + 653, + 402, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 653, + 402, + 662 + ], + "spans": [ + { + "bbox": [ + 345, + 653, + 402, + 662 + ], + "type": "inline_equation", + "content": "k\\gets i\\times w + j" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 345, + 663, + 420, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 663, + 420, + 673 + ], + "spans": [ + { + "bbox": [ + 345, + 663, + 420, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t / s}(k)\\gets \\mathcal{P}_{t + \\delta}(k)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 326, + 675, + 386, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 675, + 386, + 683 + ], + "spans": [ + { + "bbox": [ + 326, + 675, + 386, + 683 + ], + "type": "inline_equation", + "content": "c\\gets c + m\\times n" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 685, + 354, + 694 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 685, + 354, + 694 + ], + "spans": [ + { + "bbox": [ + 316, + 685, + 354, + 694 + ], + "type": "text", + "content": "end while" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 696, + 361, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 696, + 361, + 706 + ], + "spans": [ + { + "bbox": [ + 316, + 696, + 361, + 706 + ], + "type": "text", + "content": "return " + }, + { + "bbox": [ + 316, + 696, + 361, + 706 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t / s}" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1485" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 70, + 231, + 162 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 231, + 162 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 231, + 162 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 231, + 162 + ], + "type": "image", + "image_path": "8f5a4c6c365a07c269c3222fdd0d4939373702d463c4f899b3eee7a144e90857.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "text", + "content": "Figure 6. Patch of Interest (POI) module. The " + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t" + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{t + \\delta /c}" + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "text", + "content": " are sequence of patch features of " + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "text", + "content": " and its random cropped version. " + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "inline_equation", + "content": "p_{t + \\delta}^{CT}" + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "inline_equation", + "content": "p_{t + \\delta /c}^{CT}" + }, + { + "bbox": [ + 47, + 170, + 287, + 205 + ], + "type": "text", + "content": " are their corresponding contextual features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 212, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 287, + 354 + ], + "type": "text", + "content": "can only identify which parts are selected in the Blockwise Swapping algorithm if the facial regions are available. Thus, the Patch of Interest (POI) is introduced to automatically explore the salient regions and ignore the background patches in an image. Apart from prior methods, the proposed POI leverages the characteristic of self-attention and can be achieved through self-learning without facial labels, such as facial bounding boxes or segmentation masks. The idea of the POI module is illustrated in Figure 6. Thanks to POI, a capability of automatically focusing on facial regions is further equipped to the learned model, making it more robust against facial movements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": "The POI relies on the contextual agreement between the frame " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{Crop}(I_{t + \\delta})" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": ". Motivated by the BERT framework, we add a Contextual Token " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "z^{CT}" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": " to the beginning of the sequence of patches, as in Eqn. (2), to learn the contextual information in the image. The deeper this token passes through the Transformer blocks, the more information is accumulated from " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "z_{t}^{i} \\in \\mathcal{P}_{t}" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": ". As a result, " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "z^{CT}" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": " becomes a placeholder to store the information extracted from other patches in the sequence and present the contextual information of the image. Let " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "p_{t + \\delta}^{CT}" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "p_{t + \\delta / c}^{CT}" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": " be the contextual features of frame " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": " and its cropped version " + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{Crop}(I_{t + \\delta})" + }, + { + "bbox": [ + 46, + 356, + 288, + 500 + ], + "type": "text", + "content": " respectively. The agreement loss is then defined as in Eqn. (10)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 510, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 510, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 119, + 510, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a g g} = H \\left(p _ {t + \\delta} ^ {C T}, p _ {t + \\delta / c} ^ {C T}\\right) \\tag {10}", + "image_path": "16509718900946d0df4c45d24072356c67f2a26617dc99f65747fa9d0ac298a8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "text", + "content": " is the function that enforces " + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "inline_equation", + "content": "p_{t + \\delta}^{CT}" + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "text", + "content": " to be similar to " + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "inline_equation", + "content": "p_{t + \\delta / crop}^{CT}" + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "text", + "content": " so that the model can discover the salient patches. The POI can be extracted from the attention map " + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "text", + "content": " at the last attention layer of encoder " + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 529, + 287, + 578 + ], + "type": "text", + "content": ". In particular, we measure:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 597, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 597, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 87, + 597, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {t + \\delta} = A [ 0, \\cdot ] = \\left[ s _ {t + \\delta} ^ {0}, s _ {t + \\delta} ^ {1}, \\dots s _ {t + \\delta} ^ {N _ {p} - 1} \\right] \\tag {11}", + "image_path": "132aaf30bf24e5cadfdab8c4aecea7a7bafd61f7e261291a3c951bb471fffd10.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 625, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 287, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 625, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\sum_{i=0}^{N_p-1} s_{t+\\delta}^i = 1" + }, + { + "bbox": [ + 47, + 625, + 287, + 664 + ], + "type": "text", + "content": ". The higher the score " + }, + { + "bbox": [ + 47, + 625, + 287, + 664 + ], + "type": "inline_equation", + "content": "s_{t+\\delta}^i" + }, + { + "bbox": [ + 47, + 625, + 287, + 664 + ], + "type": "text", + "content": ", the richer the patch contains contextual information. Now, Eqn. (9) can be reformulated as in Eqn. (12)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 673, + 215, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 673, + 215, + 700 + ], + "spans": [ + { + "bbox": [ + 119, + 673, + 215, + 700 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {W} = \\operatorname {d i a g} (\\hat {A}) \\times S _ {t + \\delta} \\\\ \\mathbf {P} _ {d m a} = \\mathbf {W} \\times V (\\mathbf {P} _ {t / s}) \\\\ \\end{array}", + "image_path": "6b7ee4304dae87b75220f0dad3c2b93c59bc433b1524582d9e5dac5843094f14.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 399, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 399, + 83 + ], + "type": "text", + "content": "4.7. Loss Functions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 91, + 545, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 545, + 114 + ], + "type": "text", + "content": "The proposed " + }, + { + "bbox": [ + 306, + 91, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 306, + 91, + 545, + 114 + ], + "type": "text", + "content": "-BERT deep network is optimized using the proposed loss function as in Eqn. (13)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 123, + 545, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 123, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 378, + 123, + 545, + 135 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\gamma \\times \\mathcal {L} _ {r} + \\beta \\times \\mathcal {L} _ {a g g} \\tag {13}", + "image_path": "c73d7e3d86f93d0f3761b6af80a43fc98301cf854847c89bfce0e8c3aa02057c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 144, + 485, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 485, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 485, + 155 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 144, + 485, + 155 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 144, + 485, + 155 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 144, + 485, + 155 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 144, + 485, + 155 + ], + "type": "text", + "content": " are the weights for each loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 156, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 545, + 190 + ], + "type": "text", + "content": "Reconstruction Loss. The output of the decoder " + }, + { + "bbox": [ + 306, + 156, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{y}'_t" + }, + { + "bbox": [ + 306, + 156, + 545, + 190 + ], + "type": "text", + "content": " is reconstructed to the original image " + }, + { + "bbox": [ + 306, + 156, + 545, + 190 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 306, + 156, + 545, + 190 + ], + "type": "text", + "content": " using the Mean Square Error (MSE) function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 199, + 545, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 199, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 388, + 199, + 545, + 212 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\operatorname {M S E} \\left(\\mathbf {y} ^ {\\prime} _ {t}, I _ {t}\\right) \\tag {14}", + "image_path": "c793d8d3bb54506c8320f933934248897232757d428aa0fd73f47a9e77df0913.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 220, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 545, + 245 + ], + "type": "text", + "content": "Contextual Agreement Loss. MSE is also used to enforce the similarity of contextual features of " + }, + { + "bbox": [ + 306, + 220, + 545, + 245 + ], + "type": "inline_equation", + "content": "I_{t + \\delta / crop}" + }, + { + "bbox": [ + 306, + 220, + 545, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 220, + 545, + 245 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 252, + 545, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 252, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 367, + 252, + 545, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a g g} = \\operatorname {M S E} \\left(p _ {t + \\delta} ^ {C T}, p _ {t + \\delta / c r o p} ^ {C T}\\right) \\tag {15}", + "image_path": "ed7c1a4cf91aeecd7bb834d945ab6d587326063362f5fab5c40b85686c68804a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 277, + 432, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 432, + 290 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 432, + 290 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 297, + 435, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 435, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 435, + 308 + ], + "type": "text", + "content": "5.1. Datasets and Protocols" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 315, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 363 + ], + "type": "text", + "content": "CASME II [50]. With a 200 fps sampling rate and a facial resolution of " + }, + { + "bbox": [ + 305, + 315, + 545, + 363 + ], + "type": "inline_equation", + "content": "280 \\times 340" + }, + { + "bbox": [ + 305, + 315, + 545, + 363 + ], + "type": "text", + "content": ", CASME II provides 247 microexpression samples from 26 subjects of the same ethnicity. Labels include apex frames, action units, and emotions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 363, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 411 + ], + "type": "text", + "content": "SAMM [6]. Also, using a 200 fps frame rate and a facial resolution of " + }, + { + "bbox": [ + 305, + 363, + 545, + 411 + ], + "type": "inline_equation", + "content": "400 \\times 400" + }, + { + "bbox": [ + 305, + 363, + 545, + 411 + ], + "type": "text", + "content": ", SAMM consists of 159 samples from 32 participants and 13 ethnicities. The samples all have emotions, apex frames, and action unit labels." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 411, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 459 + ], + "type": "text", + "content": "SMIC [15]. SIMC is made up of 164 samples. Lacking apex frame and action unit labels, the samples span 16 participants of 3 ethnicities. The recordings are taken with a resolution of " + }, + { + "bbox": [ + 305, + 411, + 545, + 459 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 305, + 411, + 545, + 459 + ], + "type": "text", + "content": " at 100 fps." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 459, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 459, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 459, + 545, + 506 + ], + "type": "text", + "content": "CASME3 [14]. Officially known as " + }, + { + "bbox": [ + 305, + 459, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{CAS(ME)^3}" + }, + { + "bbox": [ + 305, + 459, + 545, + 506 + ], + "type": "text", + "content": " provides 1,109 labeled micro-expressions and 3,490 labeled macro-expressions. This dataset has roughly 80 hours of footage with a resolution of " + }, + { + "bbox": [ + 305, + 459, + 545, + 506 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 305, + 459, + 545, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 515, + 477, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 477, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 477, + 529 + ], + "type": "text", + "content": "5.2. Micro-Expression Self-Training" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "We use all raw frames from CASME3 for self-training except frames of test set. It is important to note that we do not use labels or meta information such as onset, offset, and apex index frames nor labeled emotions. In total, we construct an unlabelled dataset of 8M frames. The images are resized to " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ". Then, each image is divided into patches of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{p} = 784" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " patches. The temporal index " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " is selected randomly between a lower bound of 5 and an upper bound of 11, experimentally. The swapping ratio " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "r_{s}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " is selected as " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " of the number of patches being swapped from " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ". Each patch is projected to a latent space of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "d = 512" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " dimensions before being fed into the encoder and decoder. For the encoder and decoder, we keep the same " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " for all vectors and similar configurations, i.e., " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{e} = L_{d} = 4" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "-BERT is implemented easily in" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1486" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 70, + 482, + 358 + ], + "blocks": [ + { + "bbox": [ + 110, + 70, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 110, + 70, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 482, + 358 + ], + "type": "image", + "image_path": "5ff512598e1c0412f6804a0aafa191e4d7017b0422cd50e8be31feec7e01c1ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 367, + 547, + 391 + ], + "lines": [ + { + "bbox": [ + 46, + 367, + 547, + 391 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 547, + 391 + ], + "type": "text", + "content": "Figure 7. We demonstrate how " + }, + { + "bbox": [ + 46, + 367, + 547, + 391 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 367, + 547, + 391 + ], + "type": "text", + "content": "-BERT perceives the tiny differences between two frames. The first two rows are onset and apex input frames. The third and fourth rows are the results of RAFT and MagNet, respectively. The rest of the rows are our " + }, + { + "bbox": [ + 46, + 367, + 547, + 391 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 367, + 547, + 391 + ], + "type": "text", + "content": "-BERT results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 396, + 290, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 290, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 290, + 469 + ], + "type": "text", + "content": "Pytorch framework and trained by " + }, + { + "bbox": [ + 46, + 396, + 290, + 469 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 46, + 396, + 290, + 469 + ], + "type": "text", + "content": " A100 GPUs (40G each). The learning rate is set to 0.0001 initially and then reduced to zero gradually under ConsineLinear [24] policy. The batch size is set to 64/GPU. The model is optimized by AdamW [25] for 100 epochs. The training is completed within three days." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 476, + 212, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 212, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 212, + 489 + ], + "type": "text", + "content": "5.3. Micro-Expression Recognition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": "We leverage the pretrained " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": "-BERT as an initial weight and take the encoder " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": " and DMA module of " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": "-BERT as the MER backbone. The input of MER is the onset and apex frames which correspond to " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": " respectively. In Eqn. (8), " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{dma}" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": " are the features representing the micro changes and movements between onset and apex frames. They can be effectively adopted for recognizing microexpressions. We adopt the standard metrics and protocols of MER2019 challenge [34] with the unweighted F1 score " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "UF1 = \\frac{1}{C} \\sum_{i=0}^{C-1} \\frac{2 \\times TP_i}{TP_i + FP_i + FN_i}" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": " and accuracy " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "UAR = \\frac{1}{C} \\sum_{i=0}^{C-1} \\frac{TP_i}{N_i}" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": " is the number of MEs, " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "N_i" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": " is the total number of " + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 490, + 288, + 651 + ], + "type": "text", + "content": " ME in the dataset. Leave-one-out cross-validation (LOOCV) scheme is used for evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 105, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 105, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 105, + 670 + ], + "type": "text", + "content": "5.4. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Our proposed " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "-BERT shows a significant improvement over prior methods and baselines, as shown in Table 1 on the CASME3. Tested using 3, 4, and 7 emotion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": "classes, " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": "-BERT achieves double-digit gains over the compared methods in each category. In the case of 3 emotion classes, " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": "-BERT achieved a " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "56.04\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " UF1 score and " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "61.25\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " UAR, compared to RCN-A's [49] " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "39.28\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " UF1 and " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "38.93\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " UAR. For 4 emotion classes, " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": "-BERT outperforms Baseline (+Depth) [14] " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "47.18\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "30.01\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " for UF1 and " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "49.13\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "29.82\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " for UAR. Large gains over Baseline (+Depth) [14] are seen in the case of 7 emotion classes, where " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": "-BERT attains UF1 and UAR scores of " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "32.64\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "32.54\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " respectively, compared to " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "17.73\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "inline_equation", + "content": "18.29\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 516 + ], + "type": "text", + "content": " for the baseline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": "Table 2 details results for CASMEII. " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": "-BERT shows improvements over all other methods. For three categories, it achieves a UF1 of " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "90.34\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " and UAR of " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "89.14\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": ", representing " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "3.37\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "0.86\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " increases over the prior leading method (OFF-ApexNet [8]), respectively. Similar improvement is seen in five categories: a " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "4.83\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " over TSCNN [35] in terms of UF1 and a " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "0.89\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " increase over SMA-STN [19] for UAR. Similarly, " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": "-BERT performs competitively with other methods on the SAMM as seen in Table 3. Using 5 emotion classes, " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": "-BERT outperforms MinMaNet [47] by a large margin in terms of UF1 (" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "83.86\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "76.40\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": ") and UAR (" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "84.75\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "76.70\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": "), respectively. The performance of " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": "-BERT on SMIC is compared against several others in Table 4. " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": "-BERT outperforms others with a " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " increase in UF1 to " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "85.5\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "3.97\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": " boost in UAR to " + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "inline_equation", + "content": "83.84\\%" + }, + { + "bbox": [ + 304, + 519, + 547, + 698 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "On the composite dataset, " + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "-BERT again outperforms" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1487" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 87, + 285, + 253 + ], + "blocks": [ + { + "bbox": [ + 94, + 71, + 240, + 81 + ], + "lines": [ + { + "bbox": [ + 94, + 71, + 240, + 81 + ], + "spans": [ + { + "bbox": [ + 94, + 71, + 240, + 81 + ], + "type": "text", + "content": "Table 1. MER on the CASME3 dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 87, + 285, + 253 + ], + "lines": [ + { + "bbox": [ + 50, + 87, + 285, + 253 + ], + "spans": [ + { + "bbox": [ + 50, + 87, + 285, + 253 + ], + "type": "table", + "html": "
Method# ClassesUF1 (%)UAR(%)
FR [54]334.9334.13
STSTNet [18]337.9537.92
RCN-A [49]339.2838.93
μ-BERT (ours)356.0461.25
Baseline [14]429.1529.10
Baseline (+Depth) [14]430.0129.82
μ-BERT (ours)447.1849.13
Baseline [14]717.5918.01
Baseline(+Depth) [14]717.7318.29
μ-BERT (ours)732.6432.54
", + "image_path": "4a1225ad7e1e8674e3c73f7883ada7726d6e5ba1b809682d6dce3f63ad32f9a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 275, + 285, + 498 + ], + "blocks": [ + { + "bbox": [ + 99, + 259, + 235, + 270 + ], + "lines": [ + { + "bbox": [ + 99, + 259, + 235, + 270 + ], + "spans": [ + { + "bbox": [ + 99, + 259, + 235, + 270 + ], + "type": "text", + "content": "Table 2. MER on CASME II dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 275, + 285, + 498 + ], + "lines": [ + { + "bbox": [ + 50, + 275, + 285, + 498 + ], + "spans": [ + { + "bbox": [ + 50, + 275, + 285, + 498 + ], + "type": "table", + "html": "
Method# ClassesUF1 (%)UAR (%)
LR-GACNN [10]570.9081.30
AMAN [46]571.0075.40
Graph-TCN [13]572.4673.98
DSTAN [44]573.0075.00
GEME [28]573.5475.20
MiMaNet [47]575.9079.90
SMA-STN [19]579.4682.59
TSCNN [35]580.7080.97
μ-BERT (ours)585.5383.48
STSTNet [17]383.8286.86
OFF-ApexNet [8]386.9788.28
MAE [9]388.0387.28
μ-BERT (ours)390.3489.14
", + "image_path": "21638b10395015dc26f1fcd4f70bcc0fe6643e17a7b8a4758e7f3cf63dc05b86.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 517, + 285, + 649 + ], + "blocks": [ + { + "bbox": [ + 105, + 500, + 229, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 229, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 229, + 511 + ], + "type": "text", + "content": "Table 3. MER on SAMM dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 517, + 285, + 649 + ], + "lines": [ + { + "bbox": [ + 50, + 517, + 285, + 649 + ], + "spans": [ + { + "bbox": [ + 50, + 517, + 285, + 649 + ], + "type": "table", + "html": "
Method# ClassesUF1 (%)UAR (%)
AMAN [46]567.0068.85
SMA-STN [19]570.3377.20
GRAPH-AU [12]570.4574.26
MTMNet [48]573.6074.10
MiMaNet [47]576.4076.70
MAE [9]580.4088.98
μ-BERT (ours)583.8684.75
", + "image_path": "3692436b6c7b26b4bb8bf88d868b90e38603b86a40f0bd16e7fb91d5bcd01d89.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "other methods (Table 5). Attaining a UF1 score of " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "89.03\\%" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " and UAR of " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "88.42\\%" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "-BERT realizes " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.73\\%" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.82\\%" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " gains over previous best MiMaNet [47], respectively. Table 6 shows the impact of DMA and POI on CASME3. Our" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 87, + 544, + 237 + ], + "blocks": [ + { + "bbox": [ + 366, + 71, + 485, + 81 + ], + "lines": [ + { + "bbox": [ + 366, + 71, + 485, + 81 + ], + "spans": [ + { + "bbox": [ + 366, + 71, + 485, + 81 + ], + "type": "text", + "content": "Table 4. MER on SMIC dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 87, + 544, + 237 + ], + "lines": [ + { + "bbox": [ + 309, + 87, + 544, + 237 + ], + "spans": [ + { + "bbox": [ + 309, + 87, + 544, + 237 + ], + "type": "table", + "html": "
Method# ClassesUF1 (%)UAR (%)
DIKD [36]371.0076.06
TSCNN [35]372.3672.74
MTMNet [48]374.4076.00
AMAN [46]377.0079.87
MiMaNet [47]377.8078.60
DSTAN [44]378.0077.00
MAE [9]381.8680.82
μ-BERT (ours)385.5083.84
", + "image_path": "203cb647620b6430cc3ba488647a5ae3a897daffaccf8cd719977c534048807b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "text", + "content": "method gives more modest gains of approximately " + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "text", + "content": " in both metrics. A greater improvement is seen with DMA, where UF1 and UAR increase by another " + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "inline_equation", + "content": "2 - 4\\%" + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "text", + "content": ". Significant improvement from " + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "text", + "content": "-BERT is seen when adopting both modules, with a UF1 of " + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "inline_equation", + "content": "32.64\\%" + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "text", + "content": " and UAR of " + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "inline_equation", + "content": "32.54\\%" + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "text", + "content": ", representing roughly " + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 245, + 546, + 318 + ], + "type": "text", + "content": " gains over previous methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 324, + 524, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 524, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 524, + 337 + ], + "type": "text", + "content": "5.5. How " + }, + { + "bbox": [ + 305, + 324, + 524, + 337 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 324, + 524, + 337 + ], + "type": "text", + "content": "-BERT perceives micro-movements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": "To understand the micro-movements between two frames, the onset and apex frames are inputs for " + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": "-BERT. These frames represent the moments that the microexpression starts and is observed. We measure " + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\hat{A})" + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": " (Subsection 4.5) and " + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{t + \\delta}" + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": " (Eqn (11)) values to identify which regions contain small movements between two frames. Comparisons of " + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": "-BERT with RAFT [37], i.e., optical flow-based method and MagNet [29] are also conducted as in Fig. 7. The third and fourth columns in Fig 7 show the results of RAFT [37], and MagNet [29] on spotting the micro-movements, respectively. While RAFT is an optical flow-based method, MagNet amplifies small differences between the two frames. These methods are sensitive to the environment (e.g., lighting, illuminations). Thus, noises in the background still exist in their outputs. In addition, neither RAFT nor MagNet understand semantic information in the frame and distinguish changes inside facial or background regions. Meanwhile, " + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": "-BERT shows its advantages in perceiving micro-movements via distinguishing the facial regions and spotting the micro-expressions. In particular, the attention map in the fifth column, in Fig. 7 illustrates the micro-differences between onset and apex frames. The higher contrast represents the higher chance of small movements in these regions. With the POI module, " + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": "-BERT can automatically figure out the informative patches and ignore the background ones. Then, with DMA module, " + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": "-BERT, can detect and localize which corresponding patches/regions contain tiny movements. As shown in the seventh column, attention maps represent the most salient regions in the image. By empowering DMA and POI, " + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 342, + 546, + 713 + ], + "type": "text", + "content": "-BERT effectively identifies micro-movements within facial" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1488" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 87, + 285, + 259 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 267, + 82 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 267, + 82 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 267, + 82 + ], + "type": "text", + "content": "Table 5. MER on the Composite dataset (MECG2019)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 87, + 285, + 259 + ], + "lines": [ + { + "bbox": [ + 50, + 87, + 285, + 259 + ], + "spans": [ + { + "bbox": [ + 50, + 87, + 285, + 259 + ], + "type": "table", + "html": "
Method# ClassesUF1 (%)UAR(%)
Dual-Inception [55]373.2272.78
FR [53]378.3878.32
NMER [20]378.8578.24
GRAPH-AU [12]379.1479.33
ICE-GAN [52]384.5084.10
BDCNN [3]385.0985.00
moment [48]386.4085.70
MiMaNet [47]388.3087.60
MAE [9]388.5087.40
μ-BERT (ours)389.0388.42
", + "image_path": "636ec6cf3ddf6b60351bc0856382b98cdb7a00de07c5b1050728d1a270db543a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 271, + 225, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 225, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 225, + 283 + ], + "type": "text", + "content": "regions, as demonstrated in the last column." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 289, + 146, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 146, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 146, + 300 + ], + "type": "text", + "content": "5.6. Ablation studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "type": "text", + "content": "This section compares " + }, + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "type": "text", + "content": "-BERT against other self-supervised learning (SSL) methods on the MER task. CASME3 is used for experiments since it has many unlabeled images to demonstrate the power of SSL methods. We also analyze the essential contributions of Diagonal Micro-Attention (DMA) and Patch of Interest (POI) modules. Finally, we illustrate the robustness of " + }, + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "type": "text", + "content": "-BERT pretrained on CASME3 on unseen datasets and domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "type": "text", + "content": "Comparisons with self-supervised learning methods. We utilize the encoder and decoder parts of " + }, + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "type": "text", + "content": "-BERT (without DMA and POI) to train previous SSL methods (MoCo V3 [4], BEIT [1], MAE [9]) and then continue learning the MER task on the large-scale database CASME3 [14]. Overall results are shown in Table 6. It is expected that ViT-S achieves the lowest performance for UF1 and UAR as ImageNet and Micro-Expression are two different domains. Three self-supervised methods (MoCo V3, BEIT, and MAE) got better results when they were pretrained on CASME before fine-tuning to the recognition task. Compared to ViT-S, these SSL methods gain remarkable performance. Especially, MAE [9] achieves " + }, + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 402, + 286, + 568 + ], + "type": "text", + "content": " up on UF1 and UAR compared to ViT-S, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "text", + "content": "The role of Blockwise Swapping. Our basic setup of " + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "text", + "content": "-BERT (denoted as MB1) is employed to train in an SSL manner. It is noted that only Blockwise Swapping is involved, and it does not contain either DMA or POI. Compared to MAE, MB1 outperforms MAE by " + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "text", + "content": " in both UF1 and UAR, approximately. The reasons are: (1) Blockwise Swapping enforces the model to learn local context features inside an image, i.e., " + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "text", + "content": ", and (2) It helps the network to figure out micro-disparities between two frames " + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "inline_equation", + "content": "I_{t + \\delta}" + }, + { + "bbox": [ + 46, + 569, + 286, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "The role of DMA. This module is the guide to tell the network where to look and which patches to focus. By doing so, the " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "-BERT gets more robust knowledge of micro" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 97, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "type": "text", + "content": "Table 6. MER performance on CASME3 by different self-supervised methods and various settings of " + }, + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "type": "text", + "content": " -BERT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 97, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 309, + 97, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 309, + 97, + 545, + 209 + ], + "type": "table", + "html": "
MethodPre-trainDMAPOIUF1UAR
ViT-S [7]ImageNetXX20.3418.76
MoCo V3 - R50 [4]CASME3XX19.1217.36
MoCo V3 - R101 [4]CASME3XX20.1418.52
MoCo V3 [4]CASME3XX22.1319.34
BEIT [1]CASME3XX23.5419.89
MAE [9]CASME3XX23.8620.87
μ-BERT (MB1)CASME3XX25.2722.96
μ-BERT (MB2)CASME3✓X27.3526.18
μ-BERT (MB3)CASME3✓✓32.6432.54
", + "image_path": "5957ac4458c140478495df38a81b31ae81bc4a3d3bf07b76e8fbe681e5f5a6dd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "type": "text", + "content": "movements between two frames. For this reason, the network (denoted as MB2) achieves " + }, + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "type": "text", + "content": " on UF1 and a significant " + }, + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "type": "text", + "content": " gain on UAR compared to MB1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "content": "The role of POI. Since MB1 are sensitive to background noise, the micro-disparities features " + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{dma}" + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "content": " might contain unwanted features coming from the background. The POI is designed as a filter that only lets the typical interesting patches belonging to the subject go through and preserves the micro-movement features only. The improvements of up to " + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "content": " compared to MB2 demonstrate the important role of POI in " + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "content": "-BERT for micro-expression tasks. Qualitative results demonstrated in Supplementary Material can further emphasize the advantages of POI in assisting the network to be robust against facial movements." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 391, + 468, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 391, + 468, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 468, + 403 + ], + "type": "text", + "content": "6. Conclusions and Discussions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "type": "text", + "content": "Unlike a few concurrent research on micro-expression, we move forward and study how to explore BERT pretraining for this problem. In our proposed " + }, + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "type": "text", + "content": "-BERT, we presented a novel Diagonal Micro Attention (DMA) to learn the micro-movements of the subject across frames. The Patch of Interest (POI) module is proposed to guide the network, focusing on the most salient parts, i.e., facial regions, and ignoring the noisy sensitivities from the background. Empowered by the simple design of " + }, + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "type": "text", + "content": "-BERT, SOTA performance on micro-expression recognition tasks is achieved in four benchmark datasets. Our perspective will inspire more future study efforts in this direction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "content": "Limitations. We demonstrated the efficiency of the POI module in removing noise in the background, which is sensitive to lighting and illumination. However, suppose any facial parts, e.g., the forehead, are affected by lighting conditions while there are no movements. In that case, these noisy factors might also be included as micro-difference features. The robustness with different lighting conditions will be left as our future works." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement This work is supported by Arkansas Biosciences Institute (ABI) Grant, NSF WVAR-CRESH and NSF Data Science, Data Analytics that are Robust and Trusted (DART). We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1489" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. 2021. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 287, + 170 + ], + "type": "text", + "content": "[2] Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu. Video-based facial micro-expression analysis: A survey of datasets, features and algorithms. IEEE transactions on pattern analysis and machine intelligence, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "text", + "content": "[3] Bin Chen, Kun-Hong Liu, Yong Xu, Qing-Qiang Wu, and Jun-Feng Yao. Block division convolutional network with implicit deep features augmentation for micro-expression recognition. IEEE Transactions on Multimedia, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "type": "text", + "content": "[4] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 328 + ], + "type": "text", + "content": "[5] Ciprian Adrian Coraneanu, Marc Oliu Simón, Jeffrey F. Cohn, and Sergio Escalera Guerrero. Survey on rgb, 3d, thermal, and multimodal approaches for facial expression recognition: History, trends, and affect-related applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(8):1548-1568, 2016. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "type": "text", + "content": "[6] Adrian K. Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap. Samm: A spontaneous micro-facial movement dataset. IEEE Transactions on Affective Computing, 9(1):116-129, 2018. 1, 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 441 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 443, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 476 + ], + "type": "text", + "content": "[8] Y.S. Gan, S.T. Liong, W.C. Yau, Y.C. Huang, and L.K. Tan. Off-apexnet on micro-expression recognition system. Signal Processing: Image Communication, 74:129-139, 2019. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 478, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 287, + 532 + ], + "type": "text", + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 590 + ], + "type": "text", + "content": "[10] Ankith Jain Rakesh Kumar and Bir Bhanu. Micro-expression classification based on landmark relations with graph attention convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1511–1520, June 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "type": "text", + "content": "[11] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression recognition based on facial graph representation learning and facial action unit fusion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571-1580, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[12] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression recognition based on facial graph representation learning and facial action unit fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1571-1580, June 2021. 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Ling Lei, Jianfeng Li, Tong Chen, and Shigang Li. A novel Graph-TCN with a graph structured representation for micro-expression recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2237–2245, 2020. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 205 + ], + "type": "text", + "content": "[14] Jingting Li, Zizhao Dong, Shaoyuan Lu, Su-Jing Wang, Wen-Jing Yan, Yinhuan Ma, Ye Liu, Changbing Huang, and Xiaolan Fu. Cas(me)sup3/sup: A third generation facial spontaneous micro-expression database with depth information and high ecological validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 207, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 271 + ], + "type": "text", + "content": "[15] Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietikainen. A spontaneous micro-expression database: Inducement, collection and baseline. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6, 2013. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "type": "text", + "content": "[16] Yante Li, Xiaohua Huang, and Guoying Zhao. Microexpression action unit detection with spatial and channel attention. Neurocomputing, 436:221-231, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "text", + "content": "[17] S.T. Liong, Y.S. Gan, J. See, H. Khor, and Y. Huang. Shallow triple stream three-dimensional cnn (ststnet) for microexpression recognition. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-5. IEEE, 2019. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "type": "text", + "content": "[18] Sze-Teng Liong, Yee Siang Gan, John See, Huai-Qian Khor, and Yen-Chang Huang. Shallow triple stream three-dimensional cnn (ststnet) for micro-expression recognition. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-5. IEEE, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "text", + "content": "[19] Jiateng Liu, Wenming Zheng, and Yuan Zong. SMA-STN: Segmented movement-attending spatiotemporal network for micro-expression recognition. arXiv preprint arXiv:2010.09342, 2020.6.7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "text", + "content": "[20] Yuchi Liu, Heming Du, Liang Zheng, and Tom Gedeon. A neural micro-expression recognizer. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-4. IEEE, 2019. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[21] Yanju Liu, Yange Li, Xinhai Yi, Zuojin Hu, Huiyu Zhang, and Yanzhong Liu. Lightweight vit model for microexpression recognition enhanced by transfer learning. Frontiers in Neurorobotics, 16, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "type": "text", + "content": "[22] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1490" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 97, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 97, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 97, + 286, + 140 + ], + "type": "text", + "content": "[26] Xuan-Bac Nguyen, Duc Toan Bui, Chi Nhan Duong, Tien D Bui, and Khoa Luu. Clusformer: A transformer based clustering approach to unsupervised large-scale face and visual landmark recognition. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 143, + 286, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 143, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 143, + 286, + 186 + ], + "type": "text", + "content": "[27] Xuan-Bac Nguyen, Guee Sang Lee, Soo Hyung Kim, and Hyung Jeong Yang. Self-supervised learning based on spatial awareness for medical image analysis. IEEE Access, 8:162973-162981, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 189, + 286, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 189, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 286, + 232 + ], + "type": "text", + "content": "[28] Xuan Nie, Madhumita A Takalkar, Mengyang Duan, Haimin Zhang, and Min Xu. GEME: Dual-stream multi-task gender-based micro-expression recognition. Neurocomputing, 427:13-28, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 235, + 286, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 235, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 49, + 235, + 286, + 278 + ], + "type": "text", + "content": "[29] Tae-Hyun Oh, Ronnachai Jaroensri, Changil Kim, Mohamed Elgharib, Fredo Durand, William T. Freeman, and Wojciech Matusik. Learning-based video motion magnification, 2018. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 281, + 286, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 281, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 49, + 281, + 286, + 324 + ], + "type": "text", + "content": "[30] Kha Gia Quach, Ngan Le, Chi Nhan Duong, Ibsa Jalata, Kaushik Roy, and Khoa Luu. Non-volume preserving-based fusion to group-level emotion recognition on crowd videos. Pattern Recognition, 128:108646, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 327, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 286, + 380 + ], + "type": "text", + "content": "[31] Ankith Jain Rakesh Kumar and Bir Bhanu. Micro-expression classification based on landmark relations with graph attention convolutional network. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1511-1520, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 383, + 286, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 383, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 49, + 383, + 286, + 437 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 440, + 286, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 440, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 49, + 440, + 286, + 473 + ], + "type": "text", + "content": "[33] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 475, + 286, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 475, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 475, + 286, + 529 + ], + "type": "text", + "content": "[34] John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su-Jing Wang. Megc 2019 – the second facial microexpressions grand challenge. In 2019 14th IEEE International Conference on Automatic Face Gesture Recognition (FG 2019), pages 1–5, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 532, + 286, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 532, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 49, + 532, + 286, + 575 + ], + "type": "text", + "content": "[35] B. Song, K. Li, Y. Zong, J. Zhu, W. Zheng, J. Shi, and L. Zhao. Recognizing spontaneous micro-expression using a three-stream convolutional neural network. IEEE Access, 7:184537-184551, 2019. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "type": "text", + "content": "[36] Bo Sun, Siming Cao, Dongliang Li, Jun He, and Lejun Yu. Dynamic micro-expression recognition using knowledge distillation. IEEE Transactions on Affective Computing, 2020. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 623, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 623, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 623, + 286, + 656 + ], + "type": "text", + "content": "[37] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 658, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 286, + 713 + ], + "type": "text", + "content": "[38] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers amp; distillation through attention. In International Conference on Machine Learning, volume 139, pages 10347-10357, July 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[39] Thuong-Khanh Tran, Quang-Nhat Vo, Xiaopeng Hong, Xiaobai Li, and Guoying Zhao. Micro-expression spotting: A new benchmark. Neurocomputing, 443:356-368, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 173 + ], + "type": "text", + "content": "[40] Thanh-Dat Truong, Quoc-Huy Bui, Chi Nhan Duong, Han-Seok Seo, Son Lam Phung, Xin Li, and Khoa Luu. Direcformer: A directed attention in transformer approach to robust action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20030-20040, June 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "text", + "content": "[41] Thanh-Dat Truong, Chi Nhan Duong, The De Vu, Hoang Anh Pham, Bhiksha Raj, Ngan Le, and Khoa Luu. The right to talk: An audio-visual transformer approach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1105–1114, October 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "type": "text", + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "text", + "content": "[43] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale micro-expression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "text", + "content": "[44] Yan Wang, Yikun Huang, Can Liu, Xiaoying Gu, Dandan Yang, Shuopeng Wang, and Bo Zhang. Micro expression recognition via dual-stream spatiotemporal attention network. Journal of Healthcare Engineering, 2021, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "type": "text", + "content": "[45] Yandan Wang, John See, Yee-Hui Oh, Raphael C.-W. Phan, Yogachandran Rahulamathavan, Huo-Chong Ling, Su-Wei Tan, and Xujie Li. Effective recognition of facial microexpressions with video motion magnification. Multimedia Tools and Applications, 76(20):21665-21690, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 434, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 499 + ], + "type": "text", + "content": "[46] Mengting Wei, Wenming Zheng, Yuan Zong, Xingxun Jiang, Cheng Lu, and Jiateng Liu. A novel micro-expression recognition approach using attention-based magnification-adaptive networks. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2420-2424. IEEE, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[47] Bin Xia and Shangfei Wang. Micro-expression recognition enhanced by macro-expression from spatial-temporal domain. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 1186-1193, 2021. 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "type": "text", + "content": "[48] Bin Xia, Weikang Wang, Shangfei Wang, and Enhong Chen. Learning from macro-expression: a micro-expression recognition framework. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2936-2944, 2020. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "text", + "content": "[49] Zhaoqiang Xia, Wei Peng, Huai-Qian Khor, Xiaoyi Feng, and Guoying Zhao. Revealing the invisible with model and data shrinking for composite-database micro-expression recognition. IEEE Transactions on Image Processing, 29:8590-8605, 2020. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[50] Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu. Casme ii: An improved spontaneous micro-expression database and the baseline evaluation. PLOS ONE, 9(1):1-8, 01 2014. 2, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1491" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 307 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[51] Wen-Jing Yan, Qi Wu, Yong-Jin Liu, Su-Jing Wang, and Xiaolan Fu. Casme database: A dataset of spontaneous microexpressions collected from neutralized faces. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-7, 2013. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[52] Jianhui Yu, Chaoyi Zhang, Yang Song, and Weidong Cai. ICE-GAN: Identity-aware and capsule-enhanced gan for micro-expression recognition and synthesis. arXiv preprint arXiv:2005.04370, 2020.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[53] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition. arXiv preprint arXiv:2101.04838, 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[54] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition. Pattern Recognition, 122:108275, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "text", + "content": "[55] L. Zhou, Q. Mao, and L. Xue. Dual-inception network for cross-database micro-expression recognition. In 2019 14th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-5. IEEE, 2019. 8" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1492" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file